work_2cco4fotnjcvhnezvtxivsrouq	----	HugeDomains.com 
work_273jmyp63ba7zlvmlvgrkozvla	----	PII: 0196-8904(94)00075-B P e r g a m o n Energy Convers. Mgmt Vol. 36, No. 4, pp. 257-261, 1995 Copyright © 1995 Elsevier Science Ltd 0196-8904(94)00075-1 Printed in Great Britain. All rights reserved 0196-8904/95 $9.50 + 0.00 A N E X P E R T S Y S T E M A P P R O A C H T O T H E U N I T C O M M I T M E N T P R O B L E M D . P . K O T H A R P a n d A I J A Z A H M A D 2 1Centre for Energy Studies, Indian Institute of Technology, Delhi, Hauz Khas, New Delhi 110 016 and 2Electrical Engineering Department, Regional Engineering College, Srinagar, Kashmir, India (Received 26 March 1994; received for publication 8 December 1994) Abstract--An expert system plays a key role in the better conservation and management of electrical energy in a power station having a number of units that have to be committed for a given load. This paper presents a hybrid expert system dynamic programming approach to the unit commitment problem. Here, the scheduling output of the usual dynamic programming is enhanced by supplementing it with the rule based expert system. The proposed system limits the number of constraints and also checks the possible constraint violations in the generated schedule. The expert system communicates with the operator in a friendly manner, and hence, the various program parameters can be adjusted to have an optimal, operationally acceptable schedule. Unit commitment Dynamic programming Expert system 1. I N T R O D U C T I O N I n o r d e r to select the m o s t e c o n o m i c a l schedule o f units in a p o w e r system, different p r o g r a m a l g o r i t h m s h a v e been developed in the past. These algorithms are b a s e d o n m a t h e m a t i c a l p r o g r a m m i n g m e t h o d s , such as d y n a m i c p r o g r a m m i n g ( D P ) [ 1 , 7], L a g r a n g i a n relaxation [2], b r a n c h a n d b o u n d etc. Experience with unit c o m m i t m e n t ( U C ) using a D P technique has s h o w n that, in o r d e r to o b t a i n a g o o d schedule, o p e r a t o r experience should be included in setting u p the c o n t r o l p a r a m e t e r s . This results in tuning the heuristic d a t a a n d input p a r a m e t e r s to p r o d u c e a lower cost a n d o p e r a t i o n a l l y acceptable schedule. A h y b r i d Artificial N e u r a l N e t w o r k ( A N N ) - D P a p p r o a c h [3], h a s also been used. This p a p e r p r o p o s e s a h y b r i d expert system (ES) d y n a m i c p r o g r a m m i n g a p p r o a c h to unit c o m m i t m e n t . H e r e , the scheduling o u t p u t o f the usual D P is enhanced b y s u p p l e m e n t i n g it with the rule b a s e d ES. This ES c o m b i n e s the knowledge o f the U C p r o g r a m m e r a n d a set o f rules. T h e m a i n feature o f the ES, thus developed using T u r b o P r o l o g V2.0, includes m a k i n g possible the use o f the p r o g r a m b y a n o n - e x p e r t user t h r o u g h its ability to guide the user t h r o u g h the i m p o r t a n t decision m a k i n g processes in the p r o b l e m set u p period. T h e system guides the user in adjusting the m a r g i n a l constraints, thereby ensuring a feasible solution within e c o n o m i c a l time limits. T h e m e t h o d h a s been applied to a 10 unit system, wherein the schedule has been o b t a i n e d b y the ES to ensure a feasible solution. I t provides a flexible p r o g r a m m e structure t h a t is easily a d a p t a b l e to changes in the system o p e r a t i o n . 2. E X P E R T S Y S T E M D E V E L O P M E N T I n the past, the ES a p p r o a c h for U C has been applied b y M o k h t a r i e t al. [5]. H e r e , they h a v e developed a n ES to analyze a n d i m p r o v e the D P b a s e d U C . T h e ES tool selected is b a s e d on the inference engine o f the well k n o w n M y c i n ES a n d has been developed for the c o m b u s t i o n turbine cycling p r o b l e m . O u a n g a n d S h a h i d e p o u r [4] h a v e p r o p o s e d a U C ES which applies the E A S E + N E X P E R T shell to o b t a i n a c o m m i t m e n t schedule b y using heuristic rules. A user friendly interface a n d graphic utilities h a v e also been set up. ECM 36/*--D 257 258 KOTHARI and AHMAD: THE UNIT COMMITMENT PROBLEM In the work presented here, a rule based ES approach [6] for handling the UC problem is based on the application of an ES to supplement the DP technique so that a more appropriate solution is achieved. 2.1. Main features of the ES development The main features of the ES developed here are: (i) It combines the knowledge of the UC programmer with the rule base and, hence, can lead an inexperienced operator to a better unit schedule. (ii) The knowledge base (KB)[9, 10], which is the data or knowledge used to make decisions, consists of two parts, i.e. the working memory and the rule base. The KB, containing the knowledge of both schedulers and mathematical programmers, is represented by if/then rule statements. (iii) Each rule represents a piece of knowledge relevant to the problem and the rules can be added, removed and modified in the knowledge base conveniently. (iv) Production rules are close to natural language and, therefore, easy to understand. This enables the user to communicate with the ES in a friendly manner in order to adjust various programme parameters, so that an optimal and acceptable solution is approached. (v) It can give the steps that lead to the conclusions and user's queries about the main phase of the problem. The user can confirm or correct the conclusion by examining the explanations given by the inference engine. The ES developed functions in the following main capacities: (i) Preprocessor of the DP results. (ii) KB and reasoning. (iii) Postprocessor of the result. (iv) User consultant. 2.2. Preprocessor of the DP results There are many constraints in the UC problem which are too difficult to include in the DP algorithm because it increases the programm execution time exponentially. In order to adjust the input data properly, the operator must have a thorough knowledge of the UC programme. The constraints, like operational and complex, are accommodated in the ES as described below. 2.2.1. Inclusion of operational constraints. The operational constraints are included by the user who is guided by the preprocessor, and this is accomplished through an interactive question and answer session during which the user will enter data and answer questions concerning the constraints, such as spinning reserve requirements, unit minimum up and down time, must run status, unit maintenance schedules, unit minimum and maximum generating limits etc. The information regarding these constraints is requested by the system and has to be supplied by the user. Take the case of the constraint "must-run-status". Here, it is assumed that, during any given hour, one or more units in any group may be designated as "must run", i.e. those units for which there will never be any doubt about the need to commit them either because of outstanding efficiency or exceptional capacity. There may also be certain units which will be known, in advance, to be unavailable due to scheduled or forced outage. Such units are designated as "must out". All such information is supplied by the user. 2.2.2. Inclusion of complex constraints. The inclusion of complex constraints in the problem not only makes the problem difficult but increases the program execution time also. This is particularly true of constraints like "unit minimum up and down time", which are time dependent. Here, it is assumed that such a constraint is not violated frequently, so that it can be relaxed and handled external to the UC program itself by adjusting the input data which is fed to it by the preprocessor. Detections for constraint violations can be made by including rules in the knowledge. Also, constraints like fuel constrained units, crew constraints and pollution constraints are present in the actual system operation, and these can be generalized to form production rules that would be added to the knowledge base. Here, in the present study, although it has been tried to include as many constraints as possible, yet these constraints could not be included to avoid the complexity of the problem. KOTHARI and AHMAD: THE UNIT COMMITMENT PROBLEM 259 2.3. Knowledge base and reasoning All the knowledge in solving the U C problem is presented by either a property list or a production rule, i.e. facts and rules. The property list has a form o f object-attribute-value. F o r instance, if the m a x i m u m capacity o f unit-1 is 500 MW, this knowledge is stored in the knowledge base by a triple: (Unit-l, m a x i m u m capacity, 500). Similarly, if the value o f cost parameter A o f unit 10 is Rs. 2500, this knowledge is stored as: (Unit-10, cost A, 2500). Other facts and system constraints are also represented in the same manner. The value portion o f the triple can be update if required. A n o t h e r form o f knowledge is a production rule. A rule expresses the relationship between facts and has a form o f (if clause then clause). Such rule based representation allows the expert system to approach a problem in a way similar to a h u m a n expert. An example o f knowledge representation is given below: I f the unit start up cost is low, A n d the efficiency is high, A n d the unit cost parameter values are low, Then the unit has 'must-run-status'. Such a rule is represented in Turbo Prolog as follows: Hypothesis (must-run-status), condition (start-up-cost-low), condition (efficiency high), condition (cost-parameter-values-low). The inference engine examines the first predicate o f the rule. If the start up cost is low, the engine goes to check the second predicate, if the efficiency is high, then it goes to check the third predicate, and it the cost parameter values are low, then the engine concludes that the unit has must-run-status. The facts for all units, like m i n i m u m and m a x i m u m generating limits, cost curve parameters, start up times etc. and the load curve for the period are stored in the d a t a base. This corresponds to the prolog's static database. The preceding rule a b o u t must-run-status, for example, is a part o f the rule base/static database. 2.3.1. Production rules in the knowledge base. In order to have a fast solution, a production rule based and priority list based heuristics are implemented. The rules, here used, use the priority list as the starting point with additional heuristics leading to the optimal commitment decisions. The decisions in the ES developed here are made according to the following set o f rules: Rule 1: I f a particular unit which is on the top o f the priority list o f available uncommitted units but has a higher m i n i m u m up-time (say, 4 h) than o f the other units (say 1 h) and the spinning reserve requirement is insufficient only for say 1 h, then it is more economical to commit the second unit for 1 h rather than the first unit for at least the m i n i m u m up-time o f 4 h. Rule 2: If, at a n y stage, there is an outage associated with the normally committed units, then commit the next most economical unit. Rule 3: Assign the value o f m a x i m u m tolerable insecurity level (MTIL) to be 0.000445. If, due to some forced outage, this value o f M T I L is not satisfied, then commit the units to the next nearest higher value o f M T I L . Rule 4: I f the load is much less t h a n expected, then the " m u s t r u n " unit (if suggested by the user) should be one o f the committed units in such a case. Rule 5: I f a unit is committed, decommitted and committed again and the " o f f " status in between is close or equal to the m i n i m u m down time o f the unit, commit the unit continuously at the required output. Rule 6: At any instant, if some committed units o f smaller capacity can be replaced in an " o f f " state, then do such replacement, preserving the spinning reserve constraint. 260 KOTHARI and AHMAD: THE UNIT COMMITMENT PROBLEM Rule 7: All the c o m m i t t e d units must operate at least at their m i n i m u m o u t p u t conditions. Rule 8: I f the load d e m a n d is not met at any instant, first let the m o st efficient unit generate m o r e power until either its m a x i m u m o u t p u t is reached o r the load d e m a n d is fulfilled. Rule 9" At any stage, if the sum o f the m a x i m u m generating capacity o f all c o m m i t t e d units is greater t h a n the load demand, then d e c o m m i t only that unit which has the lowest efficiency. 2.4. Postprocessor T h e function o f the postprocessor is to guide the user in analyzing the schedule an d also m a k e suggestions to the user a b o u t adjustments to p r o g r a m p aram et ers to i m p ro v e the overall schedule cost. T h e parameters adjusted here include m i n i m u m up an d d o w n time, unit initial conditions a n d priority ordering. A n o t h e r function o f the ES postprocessor is to detect constraint violations, which is d o n e by including rules in the knowledge base. These rules can deduce logically a violation condition f r o m the schedule o u t p u t and, subsequently, advise the user an d r e c o m m e n d corrections for the problem. We have seen that m i n i m u m u p and d o w n times are particularly difficult to model and c a n n o t be i n c o r p o r a t e d directly in a D P routine. So, to detect such violations, a rule can be introduced in the knowledge base. F o r example, one o f such rules is: Check constraint (up-time-violated): condition (unit on), condition (unit off), condition (dur-less-than-1 h). This rule states that, if the unit is committed, then d e c o m m i t t e d an d the d u r a t i o n between o n an d off is less than 1 h, then the unit up time is violated. T h e condition " u p - t i m e - v i o l a t e d " will n o t be a fact in the database. T h e r e are a n u m b e r o f such rules which have been used in the p r o g r a m m e for checking constraint violations. 2.5. User interface A natural language permits the user to c o m m u n i c a t e with the ES during the consultation process using a h u m a n language. T h e main feature o f the user interface, here, is t h at it works as an e x p l a n a t o r y system which permits the user to answer a question with why. T h e system then responds with any i n f o r m a t i o n regarding the c o m m i t m e n t o f a particular unit at a particular time and so helps the user to u n d e r s t a n d the reasoning process. A n y question raised by the o p e r a t o r concerning the result is answered by the ES, an d i f the results seemed to be unsatisfactory, the K B was modified and the entire process r e p e a t e d until satisfactory g o o d results were obtained. 3. S A M P L E S Y S T E M The rule based ES developed here is applied to analyze an d improve the D P results o f the sample system o f Ref. [8]. F r o m various consultation sessions with the ES, it was seen t h at the system is quite helpful in improving the results. A n u m b e r o f U C runs were m a d e to test the capability and friendliness o f the ES. T h e p r o g r a m has been ru n o n an I B M - P C - A T . Th e consultation sessions provide almost all i n f o r m a t i o n a b o u t the various U C schedules given at different values o f M T I L . T h e system also gives an explanation a b o u t an y decisions t ak en by it. It was seen that, in o rd er to increase the effectiveness o f the system, a considerable database is required. As the rules included here in the K B have been obtained mainly f r o m a t h o r o u g h survey o f the literature an d are based o n heuristics, therefore, it is expected that, if the rules o b t ai n ed f r o m field experts are included here, the system will obviously pave the way for handling the practical U C problem. The sample system consists o f 10 thermal generating units with different generating capacities and start up times. T h e time span is 24 h. It was pointed out by the ES n o t to put unit 2 o ff fo r 1 h an d then restart it again. It suggested that the unit must be kept on line f r o m the f o u r t h to the sixth hour. F u r t h e r , it suggested t h at the value o f M T I L at which the earlier schedule was c o m p u t e d by D P can be increased so t h at the overall schedule will be m o r e economical. T h e final U C schedule is depicted in Table 1. KOTHARI and AHMAD: THE UNIT COMMITMENT PROBLEM 261 Table 1 Load Unit status Hour (MW) l 2 3 4 5 6 7 8 9 10 1 2000 0 1 0 2 1980 0 1 0 3 1940 0 1 0 4 1900 0 1 1 5 1840 0 0 0 6 1870 0 1 0 7 1820 0 0 0 8 1700 0 0 0 9 1510 0 0 0 I0 1410 0 0 0 11 1320 0 0 0 12 1260 0 0 0 13 1200 0 0 0 14 1160 0 0 0 15 1140 0 0 0 16 1160 0 0 0 17 1260 0 0 0 18 1380 0 0 1 19 1560 0 0 1 20 1700 0 0 1 21 1820 0 0 1 22 1900 0 1 1 23 1950 0 I 1 24 1990 0 1 1 I 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 I 1 1 1 1 1 1 1 I 0 1 0 0 1 0 0 1 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 1 0 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 When one compares the schedule obtained here to that obtained purely by D P [8], there is not much variation except the points discussed above, but the main feature is the capability o f the ES to approach the operator in a friendly manner. A better solution may be achieved if the knowledge o f a good number of field experts is included in the database. 4 . C O N C L U S I O N An expert system and dynamic programming hybrid has been presented for solving the unit commitment problem. The dynamic programming solution o f a sample system of 10 units was supplemented by the expert system, resulting in better management o f the units and, hence, better energy conservation. R E F E R E N C E S 1. P. G. Lowery, I E E E Trans. Power Apparatus Systems PAS-85, (1966). 2. Slobodan Ruzic and Nikola Rajakovic, I E E E Trans. Power Systems 6, No. 1 (1991). 3. Z. Ouang and S. M. Shahidepour, 1EEE Trans. Power Systems 6, No. 3 (1991). 4. Z. Ouang and S. M. Shahidepour, Electric Power System Res. 20, No. 3 (1990). 5. Sasan Mokhtari, Jagjit Singh and Bruce Wollenberg, I E E E Trans. Power Systems 3, No. 1 (1988). 6. S. K. Tong et al., I E E E Trans. Power Systems 6, No. 3 (1991). 7. C. K. Pang and H. C. Chen, I E E E Trans. Power Systems PAS-95, No. 4 (1976). 8. A. K. Ayub and A. D. Patton, I E E E Trans. Power Systems PAS-90, (1971). 9. T. Sakaguchi et al., I E E E Trans. Power Systems PAS-102, No. 2 (1983). 10. Tim Taylor et al., I E E E Trans. Power Systems 4, No. 1 (1989). 11. I. J. Nagrath and D. P. Kothari, Power System Engineering. McGraw-Hill, New Delhi (1994). 
work_27mb4lyoerh5zhj32bxqmm23su	----	SOME EXPERT SYSTEM NEED COMMON SENSE John McCarthy Computer Science Department Stanford University Stanford, CA 94305 jmc@cs.stanford.edu http://www-formal.stanford.edu/jmc/ 1984 Abstract An expert system is a computer program intended to embody the knowl- edge and ability of an expert in a certain domain. The ideas behind them and several examples have been described in other lectures in this symposium. Their performance in their specialized domains are often very impressive. Nevertheless, hardly any of them have certain common sense knowledge and ability possessed by any non-feeble-minded human. This lack makes them “brittle”. By this is meant that they are difficult to extend beyond the scope originally contemplated by their designers, and they usually don’t recognize their own limitations. Many important applications will require common sense abilities. The object of this lecture is to describe common sense abili- ties and the problems that require them. Common sense facts and methods are only very partially understood to- day, and extending this understanding is the key problem facing artificial intelligence. This isn’t exactly a new point of view. I have been advocating “Com- puter Programs with Common Sense”since I wrote a paper with that title in 1 1958. Studying common sense capability has sometimes been popular and sometimes unpopular among AI researchers. At present it’s popular, per- haps because new AI knowledge offers new hope of progress. Certainly AI researchers today know a lot more about what common sense is than I knew in 1958 — or in 1969 when I wrote another paper on the subject. However, expressing common sense knowledge in formal terms has proved very difficult, and the number of scientists working in the area is still far too small. One of the best known expert systems is MYCIN (Shortliffe 1976; Davis, Buchanan and Shortliffe 1977), a program for advising physicians on treating bacterial infections of the blood and meningitis. It does reasonably well without common sense, provided the user has common sense and understands the program’s limitations. MYCIN conducts a question and answer dialog. After asking basic facts about the patient such as name, sex and age, MYCIN asks about suspected bacterial organisms, suspected sites of infection, the presence of specific symptoms (e.g. fever, headache) relevant to diagnosis, the outcome of labo- ratory tests, and some others. It then recommends a certain course of antibi- otics. While the dialog is in English, MYCIN avoids having to understand freely written English by controlling the dialog. It outputs sentences, but the user types only single words or standard phrases. Its major innovations over many previous expert systems were that it uses measures of uncertainty (not probabilities) for its diagnoses and the fact that it is prepared to explain its reasoning to the physician, so he can decide whether to accept it. Our discussion of MYCIN begins with its ontology. The ontology of a program is the set of entities that its variables range over. Essentially this is what it can have information about. MYCIN’s ontology includes bacteria, symptoms, tests, possible sites of in- fection, antibiotics and treatments. Doctors, hospitals, illness and death are absent. Even patients are not really part of the ontology, although MYCIN asks for many facts about the specific patient. This is because patients aren’t values of variables, and MYCIN never compares the infections of two differ- ent patients. It would therefore be difficult to modify MYCIN to learn from its experience. MYCIN’s program, written in a general scheme called EMYCIN, is a so- called production system. A production system is a collection of rules, each of which has two parts — a pattern part and an action part. When a rule is activated, MYCIN tests whether the pattern part matches the database. If so this results in the variables in the pattern being matched to whatever 2 entities are required for the match of the database. If not the pattern fails and MYCIN tries another. If the match is successful, then MYCIN performs the action part of the pattern using the values of the variables determined by the pattern part. The whole process of questioning and recommending is built up out of productions. The production formalism turned out to be suitable for representing a large amount of information about the diagnosis and treatment of bacterial infections. When MYCIN is used in its intended manner it scores better than medical students or interns or practicing physicians and on a par with experts in bacterial diseases when the latter are asked to perform in the same way. However, MYCIN has not been put into production use, and the reasons given by experts in the area varied when I asked whether it would be appropriate to sell MYCIN cassettes to doctors wanting to put it on their micro-computers. Some said it would be ok if there were a means of keeping MYCIN’s database current with new discoveries in the field, i.e. with new tests, new theories, new diagnoses and new antibiotics. For example, MYCIN would have to be told about Legionnaire’s disease and the associated Legionnella bacteria which became understood only after MYCIN was finished. (MYCIN is very stubborn about new bacteria, and simply replies “unrecognized response”.) Others say that MYCIN is not even close to usable except experimen- tally, because it doesn’t know its own limitations. I suppose this is partly a question of whether the doctor using MYCIN is trusted to understand the documentation about its limitations. Programmers always develop the idea that the users of their programs are idiots, so the opinion that doctors aren’t smart enough not to be misled by MYCIN’s limitations may be at least partly a consequence of this ideology. An example of MYCIN not knowing its limitations can be excited by telling MYCIN that the patient has Cholerae Vibrio in his intestines. MYCIN will cheerfully recommend two weeks of tetracycline and nothing else. Pre- sumably this would indeed kill the bacteria, but most likely the patient will be dead of cholera long before that. However, the physician will presumably know that the diarrhea has to be treated and look elsewhere for how to do it. On the other hand it may be really true that some measure of common sense is required for usefulness even in this narrow domain. We’ll list some areas of common sense knowledge and reasoning ability and also apply the criteria to MYCIN and other hypothetical programs operating in MYCIN’s domain. 3 1 WHAT IS COMMON SENSE? Understanding common sense capability is now a hot area of research in artificial intelligence, but there is not yet any consensus. We will try to divide common sense capability into common sense knowledge and common sense reasoning, but even this cannot be made firm. Namely, what one man builds as a reasoning method into his program, another can express as a fact using a richer ontology. However, the latter can have problems in handling in a good way the generality he has introduced. 2 COMMON SENSE KNOWLEDGE We shall discuss various areas of common sense knowledge. 1. The most salient common sense knowledge concerns situations that change in time as a result of events. The most important events are actions, and for a program to plan intelligently, it must be able to determine the effects of its own actions. Consider the MYCIN domain as an example. The situation with which MYCIN deals includes the doctor, the patient and the illness. Since MYCIN’s actions are advice to the doctor, full planning would have to include infor- mation about the effects of MYCIN’s output on what the doctor will do. Since MYCIN doesn’t know about the doctor, it might plan the effects of the course of treatment on the patient. However, it doesn’t do this either. Its rules give the recommended treatment as a function of the information elicited about the patient, but MYCIN makes no prognosis of the effects of the treatment. Of course, the doctors who provided the information built into MYCIN considered the effects of the treatments. Ignoring prognosis is possible because of the specific narrow domain in which MYCIN operates. Suppose, for example, a certain antibiotic had the precondition for its usefulness that the patient not have a fever. Then MYCIN might have to make a plan for getting rid of the patient’s fever and verifying that it was gone as a part of the plan for using the antibiotic. In other domains, expert systems and other AI programs have to make plans, but MYCIN doesn’t. Perhaps if I knew more about bacterial diseases, I would conclude that their treatment sometimes really does require planning and that lack of planning ability limits MYCIN’s utility. The fact that MYCIN doesn’t give a prognosis is certainly a limitation. 4 For example, MYCIN cannot be asked on behalf of the patient or the admin- istration of the hospital when the patient is likely to be ready to go home. The doctor who uses MYCIN must do that part of the work himself. More- over, MYCIN cannot answer a question about a hypothetical treatment, e.g. “What will happen if I give this patient penicillin?” or even “What bad things might happen if I give this patient penicillin?”. 2. Various formalisms are used in artificial intelligence for representing facts about the effects of actions and other events. However, all systems that I know about give the effects of an event in a situation by describing a new situation that results from the event. This is often enough, but it doesn’t cover the important case of concurrent events and actions. For example, if a patient has cholera, while the antibiotic is killing the cholera bacteria, the damage to his intestines is causing loss of fluids that are likely to be fatal. Inventing a formalism that will conveniently express people’s common sense knowledge about concurrent events is a major unsolved problem of AI. 3. The world is extended in space and is occupied by objects that change their positions and are sometimes created and destroyed. The common sense facts about this are difficult to express but are probably not important in the MYCIN example. A major difficulty is in handling the kind of partial knowledge people ordinarily have. I can see part of the front of a person in the audience, and my idea of his shape uses this information to approximate his total shape. Thus I don’t expect him to stick out two feet in back even though I can’t see that he doesn’t. However, my idea of the shape of his back is less definite than that of the parts I can see. 4. The ability to represent and use knowledge about knowledge is often required for intelligent behavior. What airline flights there are to Singapore is recorded in the issue of the International Airline Guide current for the proposed flight day. Travel agents know how to book airline flights and can compute what they cost. An advanced MYCIN might need to reason that Dr. Smith knows about cholera, because he is a specialist in tropical medicine. 5. A program that must co-operate or compete with people or other pro- grams must be able to represent information about their knowledge, beliefs, goals, likes and dislikes, intentions and abilities. An advanced MYCIN might need to know that a patient won’t take a bad tasting medicine unless he is convinced of its necessity. 6. Common sense includes much knowledge whose domain overlaps that of the exact sciences but differs from it epistemologically. For example, if I spill the glass of water on the podium, everyone knows that the glass will 5 break and the water will spill. Everyone knows that this will take a fraction of a second and that the water will not splash even ten feet. However, this information is not obtained by using the formula for a falling body or the Navier-Stokes equations governing fluid flow. We don’t have the input data for the equations, most of us don’t know them, and we couldn’t integrate them fast enough to decide whether to jump out of the way. This common sense physics is contiguous with scientific physics. In fact scientific physics is imbedded in common sense physics, because it is common sense physics that tells us what the equation s = 0.5gt2 means. If MYCIN were extended to be a robot physician it would have to know common sense physics and maybe also some scientific physics. It is doubtful that the facts of the common sense world can be represented adequately by production rules. Consider the fact that when two objects collide they often make a noise. This fact can be used to make a noise, to avoid making a noise, to explain a noise or to explain the absence of a noise. It can also be used in specific situations involving a noise but also to understand general phenomena, e.g. should an intruder step on the gravel, the dog will hear it and bark. A production rule embodies a fact only as part of a specific procedure. Typically they match facts about specific objects, e.g. a specific bacterium, against a general rule and get a new fact about those objects. Much present AI research concerns how to represent facts in ways that permit them to be used for a wide variety of purposes. 3 COMMON SENSE REASONING Our ability to use common sense knowledge depends on being able to do common sense reasoning. Much artificial intelligence inference is not designed to use directly the rules of inference of any of the well known systems of mathematical logic. There is often no clear separation in the program between determining what inferences are correct and the strategy for finding the inferences required to solve the problem at hand. Nevertheless, the logical system usually corre- sponds to a subset of first order logic. Systems provide for inferring a fact about one or two particular objects from other facts about these objects and a general rule containing variables. Most expert systems, including MYCIN, never infer general statements, i.e. quantified formulas. 6 Human reasoning also involves obtaining facts by observation of the world, and computer programs also do this. Robert Filman did an interesting thesis on observation in a chess world where many facts that could be obtained by deduction are in fact obtained by observation. MYCIN’s doesn’t require this, but our hypothetical robot physician would have to draw conclusions from a patient’s appearance, and computer vision is not ready for it. An important new development in AI (since the middle 1970s) is the formalization of nonmonotonic reasoning. Deductive reasoning in mathematical logic has the following property — called monotonicity by analogy with similar mathematical concepts. Sup- pose we have a set of assumptions from which follow certain conclusions. Now suppose we add additional assumptions. There may be some new con- clusions, but every sentence that was a deductive consequence of the original hypotheses is still a consequence of the enlarged set. Ordinary human reasoning does not share this monotonicity property. If you know that I have a car, you may conclude that it is a good idea to ask me for a ride. If you then learn that my car is being fixed (which does not contradict what you knew before), you no longer conclude that you can get a ride. If you now learn that the car will be out in half an hour you reverse yourself again. Several artificial intelligence researchers, for example Marvin Minsky (1974) have pointed out that intelligent computer programs will have to reason non- monotonically. Some concluded that therefore logic is not an appropriate formalism. However, it has turned out that deduction in mathematical logic can be supplemented by additional modes of nonmonotonic reasoning, which are just as formal as deduction and just as susceptible to mathematical study and computer implementation. Formalized nonmonotonic reasoning turns out to give certain rules of conjecture rather than rules of inference — their conclu- sion are appropriate, but may be disconfirmed when more facts are obtained. One such method is circumscription, described in (McCarthy 1980). A mathematical description of circumscription is beyond the scope of this lecture, but the general idea is straightforward. We have a property applicable to objects or a relation applicable to pairs or triplets, etc. of objects. This property or relation is constrained by some sentences taken as assumptions, but there is still some freedom left. Circumscription further constrains the property or relation by requiring it to be true of a minimal set of objects. 7 As an example, consider representing the facts about whether an object can fly in a database of common sense knowledge. We could try to provide axioms that will determine whether each kind of object can fly, but this would make the database very large. Circumscription allows us to express the assumption that only those objects can fly for which there is a positive statement about it. Thus there will be positive statements that birds and airplanes can fly and no statement that camels can fly. Since we don’t include negative statements in the database, we could provide for flying camels, if there were any, by adding statements without removing existing statements. This much is often done by a simpler method — the closed world assumption discussed by Raymond Reiter. However, we also have exceptions to the general statement that birds can fly. For example, penguins, ostriches and birds with certain feathers removed can’t fly. Moreover, more exceptions may be found and even exceptions to the exceptions. Circumscription allows us to make the known exceptions and to provide for additional exceptions to be added later — again without changing existing statements. Nonmonotonic reasoning also seems to be involved in human communi- cation. Suppose I hire you to build me a bird cage, and you build it without a top, and I refuse to pay on the grounds that my bird might fly away. A judge will side with me. On the other hand suppose you build it with a top, and I refuse to pay full price on the grounds that my bird is a penguin, and the top is a waste. Unless I told you that my bird couldn’t fly, the judge will side with you. We can therefore regard it as a communication convention that if a bird can fly the fact need not be mentioned, but if the bird can’t fly and it is relevant, then the fact must be mentioned. References Davis, Randall; Buchanan, Bruce; and Shortliffe, Edward (1977). Production Rules as a Representation for a Knowledge-Based Consultation Program, Artificial Intelligence, Volume 8, Number 1, February. McCarthy, John (1960). Programs with Common Sense, Proceedings of the Teddington Conference on the Mechanization of Thought Processes, London: Her Majesty’s Stationery Office. (Reprinted in this volume, pp. 000–000). McCarthy, John and Patrick Hayes (1969). Some Philosophical Problems from the Standpoint of Artificial Intelligence, in B. Meltzer and D. Michie (eds), Machine Intelligence 4, Edinburgh University. (Reprinted in B. L. 8 Webber and N. J. Nilsson (eds.), Readings in Artificial Intelligence, Tioga, 1981, pp. 431–450; also in M. J. Ginsberg (ed.), Readings in Nonmonotonic Reasoning, Morgan Kaufmann, 1987, pp. 26–45; also in this volume, pp. 000–000.) McCarthy, John (1980). Circumscription — A Form of Nonmonotonic Rea- soning, Artificial Intelligence, Volume 13, Numbers 1,2. (Reprinted in B. L. Webber and N. J. Nilsson (eds.), Readings in Artificial Intelligence, Tioga, 1981, pp. 466–472; also in M. J. Ginsberg (ed.), Readings in Nonmonotonic Reasoning, Morgan Kaufmann, 1987, pp. 145–152; also in this volume, pp. 000–000.) Minsky, Marvin (1974). A Framework for Representing Knowledge, M.I.T. AI Memo 252. Shortliffe, Edward H. (1976). Computer-Based Medical Consultations: MYCIN, American Elsevier, New York, NY. ANSWERS TO QUESTIONS DISCUSSION OF THE PAPER QUESTION: You said the programs need common sense, but that’s like saying, If I could fly I wouldn’t have to pay Eastern Airliness $44 to haul me up here from Washington. So if the programs indeed need common sense, how do we go about it? Isn’t that the point of the argument? DR. MCCARTHY: I could have made this a defensive talk about artificial intelligence, but I chose to emphasize the problems that have been identified rather than the progress that has been made in solving them. Let me remind you that I have argued that the need for common sense is not a truism. Many useful things can be done without it, e.g. MYCIN and also chess programs. QUESTION: There seemed to be a strong element in your talk about common sense, and even humans developing it, emphasizing an experiential compo- nent — particularly when you were giving your example of dropping a glass of water. I’m wondering whether the development of these programs is going to take similar amounts of time. Are you going to have to have them go through the sets of experiences and be evaluated? Is there work going on in terms of speeding up the process or is it going to take 20 years for a program from the time you’ve put in its initial state to work up to where it has a decent amount of common sense? 9 DR. MCCARTHY: Consider your 20 years. If anyone had known in 1963 how to make a program learn from its experience to do what a human does after 20 years, they might have done it, and it might be pretty smart by now. Already in 1958 there had been work on programs that learn from experi- ence. However, all they could learn was to set optimal values of numerical parameters in the program, and they were quite limited in their ability to do that. Arthur Samuel’s checker program learned optimal values for its parameters, but the problem was that certain kinds of desired behavior did not correspond to any setting of the parameters, because it depended on the recognition of a certain kind of strategic situation. Thus the first prerequisite for a program to be able to learn something is that it be able to represent internally the desired modification of behavior. Simple changes in behavior must have simple representations. Turing’s universality theory convinces us that arbitrary behaviors can be represented, but they don’t tell us how to represent them in such a way that a small change in behavior is a small change in representation. Present methods of changing programs amount to education by brain surgery. QUESTION: I would ask you a question about programs needing common sense in a slightly different way, and I want to use the MYCIN program as an example. There are three actors there — the program, the physician, and the pa- tient. Taking as a criterion the safety of the patient, I submit that you need at least two of these three actors to have common sense. For example if (and sometimes this is the case) one only were sufficient, it would have to be the patient because if the program didn’t use common sense and the physician didn’t use common sense, the patient would have to have common sense and just leave. But usually, if the program had common sense built in and the physician had common sense but the patient didn’t, it really might not matter because the patient would do what he or she wants to do anyway. Let me take another possibility. If only the program has common sense and neither the physician nor the patient has common sense, then in the long run the program also will not use the common sense. What I want to say is that these issues of common sense must be looked at in this kind of frame of reference. DR. MCCARTHY: In the use of MYCIN, the physician is supposed to supply the common sense. The question is whether the program must also have 10 common sense, and I would say that the answer is not clear in the MYCIN case. Purely computational programs don’t require common sense, and none of the present chess programs have any. On the other hand, it seems clear that many other kinds of programs require common sense to be useful at all. /@steam.stanford.edu:/u/ftp/jmc/someneed.tex: begun 1996 May 12, latexed 1996 May 12 at 1:24 p.m. 11 
work_2eewznxmxbgbfbsg2ousndspha	----	PII: S0957-4174(96)00050-4 Pergamon Expert Systems With Applications, Vol. 11, No. 3, pp. 351-360, 1996 Copyright © 1996 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0957--4174/96 $15.00+0.00 PII: S0957-4174(96)00050-4 Self-Integrating Knowledge-Based Brain Tumor Diagnostic System CHING-HUNG WANG'~ Institute of Computer and Information Science, National Chiao-Tung University, Hsin-Chu 30050, Taiwan, R.O.C. TZUNG-PEI HONG Department of Information Management, Kaohsiung Polytechnic Institute, Kaohsiung 84008, Taiwan, R.O.C. SHIAN-SHYONG TSENG Institute of Computer and Information Science, National Chiao-Tung University, Hsin-Chu 30050, Taiwan, R.O.C. Abstract--In this paper, we present a self-integrating knowledge-based expert system f o r brain tumor diagnosis. The system we propose comprises knowledge building, knowledge inference and knowledge refinement. During knowlege building, an automatic knowledge-integration process, based on Darwin's theory o f natural selection, integrates knowledge derived from knowledge-acquisition tools and machine- learning methods to construct an initial knowledge base, thus eliminating a major bottleneck in developing a brain tumor diagnostic system. During the knowledge inference process, an inference engine exploits rules in the knowledge base to help diagnosticians determine brain tumor etiologies according to computer tomography pictures. And, a simple knowledge refinement method is proposed to modify the existing knowledge base during inference, which dramatically improves the accuracy o f the derived rules. The performance o f the brain tumor diagnostic system has been evaluated on actual brain tumor cases. Copyright © 1996 Elsevier Science Ltd 1. I N T R O D U C T I O N RECENTLY, EXPERT SYSTEMS h a v e been successfully applied to many fields and have shown excellent performance. Expert systems provide sound expertise in the form o f diagnosis, instruction, prediction, consulta- tion and so on. They can also be used as training tools to help new personnel interpret data and monitor observa- tions (Waterman, 1986). Developing a successful expert system requires, however, effectively integrating knowl- edge from a variety o f sources, such as that from domain experts, historical documentary evidence, or current records, to construct a complete, consistent and unambi- guous knowledge base (Baral, 1991; Gragun, 1987). For large-scale expert systems that generally cannot rely on a single knowledge source, the use o f multiple knowledge "~ Author for correspondence. A l s o Directorate General o f Tele- communication Laboratories, Ministry o f Transportation and Communications, Chung-Li, Taiwan 32617, Taiwan, R.O.C. inputs from many knowledge sources is especially important to ensure comprehensive coverage. Thus, integrating multiple knowledge sources plays a critical role in building successful expert systems. In this paper, we present a brain tumor diagnostic system that can integrate multiple knowledge sources to quickly build a prototype knowledge base. This prototype knowledge base then adapts itself according to inference results from the expert system, consequently improving the accuracy o f the rules it derives. The brain tumor diagnostic system (BTDS) consists o f three main functional units: knowledge building, knowl- edge inference and knowledge refinement (Wang & Tseng, 1995). The knowledge-building unit includes three modules: machine learning, knowledge acquisition and knowledge integration. The machine-learning mod- ule maintains a variety o f machine-learning strategies (Cendowska, 1987; Michalski, 1980; Mitchell, 1982; Quinlan, 1986) to induce knowledge from actual instances. The knowledge-acquisition module maintains 351 352 Ching Hung Wang et al. different knowledge-acquisition tools that allow knowl- edge engineers to acquire domain knowledge from various experts (Kelly, 1955; Hwang & Tseng, 1990). The knowledge-integration module uses evolutionary theory to automatically integrate knowledge from multiple sources (which may be derived by knowledge- acquisition tools or machine-learning methods) into the initial knowledge base. The inference unit helps diag- nosticians determine brain tumor etiologies according to computer axial tomography pictures. The knowledge- refinement unit uses a proposed knowledge-refinement method to modify the existing knowledge base during the inference process. The remainder o f this paper is organized as follows. The problem domain is introduced in Section 2. The architecture of the brain tumor diagnostic system is presented in Section 3. A knowlege-building unit is proposed in Section 4. A knowledge-inference unit is introduced in Section 5 . A knowledge-refinement method is proposed in Section 6. The implementation of the brain tumor diagnostic system is presented in Section 7. Conclusions are given in Section 8. 2. T H E P R O B L E M D O M A I N The field of brain tumor diagnosis is quite interesting and full of challenge since the brain is very complex and many causes of brain tumors are still unclear (Wills, 1982). Computer tomography (CT) is generally con- sidered the most reliable diagnostic technique for locating and characterizing brain tumors. Nearly all intracranial lesions are detected using CT. The usual examination involves scanning the neurocranium in a series of parallel transverse "slices". The head is bent forward so that the sectional plane lies at an angle of 12 ° to the orbitomeatal lines (Fig. 1). Each slice is 8 nun thick, so that 8-15 slices are usually sufficient to visualize the intracranial structures to be examined. A patient with a meningiomal tumor is shown in Fig. 2. Normally, several stages are necessary for doctors to FIGURE 1. Positions of six standard CT scans. FIGURE 2. An example of a CT picture. diagnose brain tumors. First, CT pictures of a patient's brain are analyzed and compared to determine the location and the density of the lesion. Next, the CT pictures are further analyzed to obtain data on calcifica- tion, degree of edema, shape of edema, degree of enhancement, type of enhancement, general appearance, size of mass, mass effect and bone change. After that, some possible brain tumors could be concluded. The brain tumor diagnosis is still difficult for inexperi- enced doctors due to the inherent complexity of brain tumors. Thus, combining multiple knowledge sources including knowledge from domain experts, historical documentary information and current records of actual instances, to develop a successful brain tumor diagnostic system is very important. From data supplied by Veterans' General Hospital (VGH) in Taipei, Taiwan, 12 parameters presently used in describing pictures derived by computerized axial tomography (CAT) scanning are shown in Table 1. One of six possible classes of brain tumors including pituitary adenoma, meningioma, medulloblastoma, glio- blastoma, astrocytoma and anaplastic protoplasmic astrocytoma (which are frequently found in Taiwan), must be identified. 348 actual cases of brain tumors from Veterans' General Hospital were used to evaluate the proposed system's performance. Table 2 shows an actual case expressed in terms of 12 features derived by computerized axial tomography (CAT) scanning, and a pathology report. 3. S Y S T E M A R C H I T E C T U R E The brain tumor diagnostic system proposed here consists of three main units: knowledge building, knowledge inference and knowledge refinement. These three units respectively generate, use and alter the rules in the knowledge base (Fig. 3). The knowledge building unit includes three modules: machine learning, knowledge acquisition and knowledge integration. The machine-learning module maintains a variety of learning methods (Michalski, 1980; Mitchell, 1982; Quinlan, 1986; Cendrowska, 1987) to induce various knowledge sources from different instance sets. Brain Tumor Diagnostic System 353 The knowledge-acquisition module maintains various knowledge acquisition tools (Kelly, 1955; Hwang & Tseng, 1990) that allow domain experts to input knowledge. Knowledge might then be directly obtained by various human experts using different knowledge- acquisition tools, or derived from different machine-learning methods. The knowledge-integration module rapidly combines multiple knowledge derived by the machine-learning module or the knowlege-acquisi- tion module to build a prototype knowledge base. The knowledge-integration approach is an adaptive search method, thus eliminating a major difficulty in knowledge integration. The knowledge inference component includes several modules: user interface, working memory, inference engine and explanation facility. The user interface helps users communicate easily with the expert system. The working memory stores facts that willbe used during the course of a consultation. The inference engine generates new facts based on the rules and facts currently known. The explanation facility, when requested, explains the system's reasoning to the user. A knowledge base integrated from multiple knowl- edge sources is often only a prototype, with TABLE 1 Twelve Brain Tumor Attributes and their Possible Values 1. LOCATION: (1) Brain parenchyma a. frontal b. temporal c. parietal d. occipital e. thalamus f. basal ganglia g. corpus callosum (2) Interior surface of brain a. frontal horn b. body of lateral ventricle c. atrium d. occipital horn e. temporal horn f. third ventricle (anterior) g. posterior third ventricle h. pineal region (3) Brain surface (excluding skull base, vault) Convexity: a. frontal b. temporal c. parietal d. occipital Parasagitah e. frontal f. parietal g. occipital h. flax i. tentorium (4) Skull vault (5) Anterior skull base (6) Middle cranial fossa (excluding sella) a. clivus b. sphenoid ridge c. parasagital skull base (7) Sellar (8) Sellar and suprasellar (9) Suprasellar (including tuberculum sellar) (10) ParaseUar (11) Cerebellopontine angle, ambiens cisterna (12) Brain stem (13) Fourth ventricle (14] Cerebellum (a. hemisphere b. vermis) (15) Cerebellar surface (extra-axial) (16) Cisterna magna (extra-axial) 2. PRECONTRAST: (1) Low (2)Iso (3) High (4) Mixed (5)With fat density (6)With air density 3. CALCIFICATION: (1) No (2) Marginal (3) Vascular-like (4) Lumpy, solid, punctate 4. EDENA: (1) No (2) < = 2 cm (3) <=1/2 hemisphere (4) >1/2 hemisphere 5. S H A P E E D E M A : (1) No (2) Smooth, regular (3) Digital, irregular 6. DEGREE_ENHANCEMENT: (1) No enhancement (2) Less than vessel :(3) Same as vessel (4) More than vessel 7. APPEARANCE_ENHANCEMENT: (1) Homogeneous (2) Thin regular marginal (3) Moderate regular marginal (4) Thick regular marginal (5) Gyrus- like (6) Grossly irregular (7) Mural nodule (8) Homogeneous with lucency inside (9) Thick irregular marginal 8. GENERAL. APPEARANCE: (1) Grossly cystic with fluid inside but no mural nodule (2) Cystic with mural nodule (3) Solid with small cyst/cysts (4) Solid with necrosis (5) Solid without necrosis or cyst (6) Mass with hemorrhage (7) Infiltrative lesion (8) Gyrus-like involvement (9) Leptomeningeal lesion 9. BONE__CHANGE: (1) No bony change (2) Sellar enlargement (3) Internal auditory meatus enlargement (4) Bony sclerosis (5) Bony erosion (6) Bony destruction 10. SIZE (cm): 11. MASS_EFFECT: (1) No mass effect, infiltrative type (2) With mass effect (3) Ipsilateral enlargement of ambiens cisterna 12. HYDROCEPHALUS: (1) No hydrocephalus, no previous shunting (2) Yes (3) No, but shunted previously 354 Ching Hung Wang et aL TABLE 2 A Case for Brain Tumor Diagnosis Feature Feature value Feature Feature value (1) Location sellarandsuprasellar (2) Precontrast high (3) Calcification marginal (4) Edema no (5) Shape edema smooth and regular (6) Size 1.2 cm (7) Enhancement degree (8) Enhancement appearance (9) General appearance (10) Bone change (11) Mass effect (12) Hydrocephalus less than vessel homogeneous with lucency solid with small cyst/cysts sellar enlargement with mass effect no hydrocephalus Pathology: pituitary adenoma unsatisfactory classification accuracy. Therefore, the prototype knowlege base must be refined. The knowl- edge-refinement unit automatically modifies the knowledge base according to results derived from the inference engine. The refinement algorithm also adopts an adaptive search method to alter the rules in the knowledge base. In the following sections, we will concentrate on the knowledge-building unit and the knowledge-refinement unit since the knowledge-inference engine is similar to other widely used types. 4. KNOWLEDGE-BUILDING UNIT Knowledge acquisition and machine learning are cur- rently two major techniques for acquiring knowledge from experts and data respectively. These two tech- niques, however, have their own limitations as Gaines pointed out (Gaines, 1989). Knowledge derived from machine-learning methods is quite dependent on the training data used, which easily makes the induced knowledge incomplete. Knowledge acquired from experts is often biased toward the experts' opinions which can easily make the derived knowledge subjective. In order to effectively construct a complete, consistent and objective knowledge base for brain tumor diagnosis, we were concerned with acquiring knowlege by integra- tion of the two techniques. Our aim was to construct an integrated brain-tumor diagnostic knowledge base from several individual knowledge sources. The knowledge-building unit can help knowledge engineers effectively acquire and integrate knowledge from various types of sources. In the following subsec- tion, we introduce the knowledge acquision, machine learning and knowlege integration functions. 4.1. K n o w l e d g e - A c q u i s i t i o n Module Recently, much study has been devoted to eliciting different types of knowledge by interviewing experts. Various knowledge-acquisition tools have been success- fully developed. In order to help BTDS easily acquire knowlege from various doctors, we included some commonly-used knowledge-acquisition tools in the Historic Documents i K n o w l e d g e A c q u i s i t i o n M a c h i n e I . ~ n i n , r i / ~ ~ ! M o d u l e M o d u l e ~ ~ o w l e d g e Ymowledge:'~ ~[ Acq~idUon I ~Inputs I II T°°l. / I ~,.~ Knowledge ~ _ ~ 'Mem,°d 111 . ~ _ _ ~ J i > / " I - I I n t e g r a t i o n L I " I i : - m ~ • ! IIAcq~a~aon II ~ t ~ I ILearninzl I i / / ~ t o w l e d g e ~ ! " [ x p e r t s Kn... o w t e d g e I n f e r e n c e . . ; . . . . . . . . . . . . . . . . . . . . . . . . . . . ~ ...... r FIGURE 3. Structure of the brain tumor diagnostic system. Brain Tumor Diagnostic System 355 knowledge-acquisition module. The knowledge-acquisi- tion module provides good flexibility and new knowledge-acquisition tools can be easily added to it. Experts can thus, depending on their preferences, choose the tools for knowledge input. The knowledge-acquisi- tion module has a knowledge-acquisition-tool manager that provides a user-friendly interface for operating various knowledge-acquisition toolsl The manager con- trols each knowledge-acquisition tool by invoking the services as required. Experts can thus easily apply any knowlege-acquisition tools to input their domain knowl- edge. Presently, two knowledge-acquisition tools, the Repertory Grid (Kelly, 1955) and EMCUD (Hwang & Tseng, 1990), are associated with the module. These tools meet the two general requirements described below. (1) knowledge-acquisition tools must be domain- independent; (2) the knowledge derived from tools must be easily translatable into the form of rules. A brief description of these knowledge-acquisition tools is given as follows. 4.1.1. Knowledge-Acquisition Tool: Repertory Grid. Operation of the repertory grid (Kelly, 1955) by a single expert can be briefly described as follows: Step 1. Elicit all the elements from the expert. At least two elements are needed to carry out the following procedure. Assume that five ele- ments, El, E2, E3, E4 and Es, are provided by the expert; we place them across the top of a grid. Step 2. Elicit constructs (traits and their opposites) from the expert. Each time three elements are chosen, ask for a construct to distinguish one element from the other two. The constructs obtained are listed down the side of the grid. Step 3. Rate all of the entries (elements, constructs) in the grid. Assume the traits C1, C2, C3, C4 and their opposites C'1, C~, C~, C~, have been given by the experts. As an example, the following repertory grid may be constructed. El E2 Ea E4 E5 C1 5 1 5 1 1 C~ C2 4 4 4 1 4 C~ C3 1 4 5 1 4 C~ C4 1 1 1 5 1 C~ Step 4. Generate production rules from the grids. 4.1.2. Knowledge-Acquisition Tool: EMCUD. EMCUD (Embedded Meanings Capturing and Uncertainty Decid- ing) (Hwang & Tseng, 1990) is a table-based knowledge acquisition method that can capture embedded meanings in given rules, and guide experts to decide certainty factors. The EMCUD strategy is briefly described as follows: Step 1. Apply some repertory grid-oriented method to derive the initial knowledge. Step 2. Construct an Attribute-Ordering Table that records the importance of each attribute to each object. Step 3. Elicit embedded meanings from the original rules and Attribute-Ordering Table. Generate embedded rules for each original rule. Step 4. Construct the constraint list to flag unwanted rules. Step 5. Guide experts to decide certainty factors of the embedded rules. 4.2. Machine-Learning Module Machine learning is another alternative for acquiring knowledge from training dam. Recently, several expert systems have been created that use machine-learning methods to generate rules from data (Gray 1990). In order to help knowledge engineers easily acquire knowl- edge from various sources, we include some commonly-used machine-learning tools in the machine- learning module. Knowledge engineers can, depending on training data representation, choose suitable tools for knowledge induction. Each machine-learning tool has a data store to hold the derived knowledge. If the derived knowledge is not expressed in the form of rules, it is then translated into the form of rules. Presently, four machine- learning tools, including Version Space (Mitchell, 1982), 1133 (Quinlan, 1986), PRISM (Cendrowska, 1987) and AQR (Michalski, 1980), are associated with the module. A brief description of these knowledge-acquisition tools is given as follows. 4.2.1. Machine-Learning Tool: Version Space. The Ver- sion Space learning strategy is mainly used for learning from training instances with only two classes: positive and negative (Mitchell, 1982). It attempts to induce concepts that include all positive training instances and exclude all negative training distances. The term "ver- sion space" is used to represent all legal hypotheses describable within a given concept-description language and consistent with all observed training instances. The term "consistenf' means that each hypothesis includes all given positive training instances and excludes all given negative ones. A version space can then be represented by two sets of hypotheses: set S and dual set G, defined a s " S={sls is a hypothesis consistent with observed instances. No other hypothesis exists that is both more specific than s and also consistent 356 Ching Hung Wang et al. with all observed instances }; G={glg is a hypothesis consistent with observed instances. No other hypothesis exists which is both more general than g and also consistent with all observed instances}. Sets S and G, together, precisely delimit a version space in which each hypothesis is both more general than some hypothesis in S and more specific than some hypothesis in G. When a new positive training instance is presented, set S is generalized to include this training instance; when a new negative training instance is presented, set G is specialized to exclude this training instance. When the Version Space is used to learn concepts from the multiple classes training set, one class is taken to be positive and all other classes are taken to be negative. 4.2.2. Machine-Learning Tool." ID3. In 1983, Quinlan proposed the ID3 learning algorithm that tries to form a decision tree from a set o f training instances (Quinlan, 1986). ID3 uses the heuristics of minimizing "entropy" in determining which attribute should be selected next in the decision tree. If Attribute A has m values (i.e. A1,A2 . . . . . Am) and the training set having attribute value Ai can be partitioned into n[ positive training instances and n,.- negative training instances, then the entropy o f choosing A as the next attribute is calculated according to the following formula: log2 n.+n+n7 ni- E = - n[ ~ ' n [ log2 n;" +n----~-" i = 0 Among all the feasible attributes, the one that entails the least entropy will be chosen as the next attribute. The same procedure is repeated until each terminal node in the decision tree contains only training instances with the same class. 4.2.3. Machine-Learning Tool: PRISM. The PRISM learning algorithm maximizes information gain instead of minimizing entropy in inducing modular rules (Cen- drowska, 1987). Attribute-valued pairs (selectors), in terms o f information theory, can be thought of as discrete messages. Given a message i, the amount o f information- gain about an event is defined as: r probability of event after i is received ] I(i)=l°g2[probability of event before i is received]" A selector (message) that provides more information- gain is then chosen to describe a class with a higher priority. The task o f the PRISM learning algorithm is to find the selector c~ x that contributes the most information- gain about a specified classificaton tS,, that is, for which 1(6,1%) is maximum. The major difference between PRISM and ID3 is that PRISM concentrates on finding only relevant attribute-value pairs, while ID3 is con- cerned with finding only the attribute that is, the most relevant overall, even though some values o f that attribute may be irrelevant. 4.2.4. Machine-Learning Tool: AQR. AQR is an induc- tion algorithm for generating a set o f classification rules (Michalski, 1980). When building decision rules, AQR performs a heuristic search through the hypothesis space to find the rules that account for all positive examples and no negative examples. AQR processes the training examples in stages; each stage generates a single rule, and then removes the examples it covers from the training set. This step is repeated until enough rules have been found to cover all the examples in the chosen class. 4.3. Knowledge-Integration Module The knowledge-integration module exploits all the available knowledge in the knowledge-acquisition mod- ule and the machine-learning module to construct a system with good performance. Some benefits o f integra- ing multiple knowledge sources in developing an expert system are described below (Medsker, 1995). (1) Knowledge acquired from different sources has good validity; (2) Domain knowledge is better understood from consensus among different knowledge sources; (3) Integrated knowledge can deal with more complex problems; (4) Knowledge integration may improve the per- formance o f the knowledge base. Since opinions o f different domain experts are differ- ent, the knowledge derived from each expert will be different, too. A similar problem also arises when separate knowledge sets are generated by individual learning methods. These various knowledge sets must be merged into a comprehensive knowledge base for the system to perform well. However, incompleteness, redundancy and inconsistency often arise. Removing them in knowledge integration is thus very important in developing a good brain tumour diagnostic system. The knowledge-integration module uses the genetic algorithm (Holland, 1 9 7 5 ) a s its integration engine to effectively integrate knowledge from multiple sources and rapidly construct a knowledge base. Here, we assume that all knowledge derived from the knowledge- acquisition and machine-learning modules are represented by rules since almost all knowledge derived by knowledge-acquisition tools or induced by machine- learning methods may easily be translated into or Brain Tumor Diagnostic System Doctors Historic Documents Actual Instances I~_°wlutodge IKnowledge Ac, quisilion [ [ Machine Learning I It Knowledge F.ncodmg I Integration I Knowledge Integration Module I l ' [ Knowledge Decoding \ I FIGURE 4. The flow chart of knowledge integration. 357 represented by rules. The flow chart for knowledge input and knowledge integration is shown in Fig. 4. In the knowledge-input stage, knowledge is acquired from various experts or induced from different training sets, and is represented as rule sets. In the knowledge-integration stage, each rule set is encoded into a bit string. The knowledge- integration module maintains a population of possible rule sets (bit strings) and uses the genetic algorithm to automatically search for the best integrated rule set to use as the knowledge base (Liao, 1995). The knowledge integration consists of three steps: encoding, integration and decoding. The encoding step transforms each rule set into a bit-string. The integration step chooses bit-string rule sets for "mating", gradually creating good offspring. The offspring then undergo recursive "evolution" until an optimal or a nearly optimal individual is found (Fig. 5). The decoding step then transforms the optimal or nearly optimal offspring into the form of rules. Since rule sets generated from different knowledge sources may vary in size and rule-set sizes may not be known beforehand, using an appropriate data structure to encode rule sets is therefore very important. In our system, variable-length bit strings are used to represent rule sets (De Jong, 1988). An example is given below. Example. Assume that two classes {C1, C2} in rule set RS are to be distinguished using three features {F1, F2, F3}. Assume Feature FI has three possible values {fu,f~z, f13} Feature F 2 has four possible values {fEl,fz2,fE3,f24}, and Feature F3 had three possible values {f3~,f3:,f33}. Also assume that the rule set RS has only two rules: R1: If (F1 =fl2) and (F2=f21) then Class is CI; R2: If (FI =f11) and (/73 =f32) then Class is C2. After encoding, the above rules are respectively repre- Initial populatlor Generation 0 Generation k P- Chromosocae ! ] - ~ -------~ C-'hromosome 2 V - - ~ ------ql Chr°m°s°me 3 [ - - ~ - - - - - - - b C,13romosomel x Knowledge encoding ~l'omo~m© cln'omosom© i Chromosome Genetic Operators Cla'°m°~°me2 i The best one Chromosome 3 ~ ............... - - - . ~ Chr°mos°me 3 ~ [ Chromosome / - - ~ R u l e Base [ i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ! Knowledge integration Knowledge decod/~g FIGURE 5. The knowledge-integration procedure. 358 Ching Hung Wang et al. sented as follows: Fi F2 F 3 Class RI 010 1000 111 10 /?2 100 1111 010 01 Finally, rule set R S is encoded into a chromosome: RS R/ n2 / X / X 010100011110 100111101001 \ ..__/\ .. / T T R~le-head points Four genetic operators, dynamic crossover, mutation, fusion and fission, are applied to the rule-set population during knowlege integration (Liao, 1995). The dynamic crossover operator takes two parent chromosomes and swaps parts o f their genetic information to produce offspring chromosomes. Unlike the conventional cross- over operator, the dynamic crossover operator selects crossover points that need not be at the same point- positions on both parent chromosomes: instead, the crossover points are at the positions the same distance from rule-head points. The mutation operator randomly changes some elements in a selected rule set to help the integration process escape from local-optimum "traps". The fusion operator checks and eliminates rule redun- dancy and subsumption relationships using an "OR" operation. If a string resulting from an "'OR" operation on two rules is the same as one o f the two rules, then a redundancy or subsumption relationship exists between the two rules. The fission operator selects the "closest" near-miss (Winston, 1992) rule to eliminate misclassi- fications and contradictions. In order to evaluate the fitness of an integrated rule set, an evaluation function is defined. The evaluation func- tion considers two factors: accuracy and complexity. Here, "complexity" is evaluated by the ratio of rule- increase in the integrated rule set, and "accuracy" is evaluated by the degree to which the integrated rule set can correctly classify test instances. Accuracy and complexity are then combined to represent the fitness value o f the rule set. The evaluation results are then fed back to the genetic algorithm to control how the solution space is searched to promote the quality o f rule sets. 5. KNOWLEDGE-INFERENCE UNIT Using the knowledge-integration approach proposed above, an integrated set o f rules can be formed from multiple knowledge sources. These rules comprise a knowledge base for brain tumor diagnosis. Some rules in the knowledge base are described below: rule 1: I F Appearance_of_Enhancement = "Homo- geneous" and Location = "Brain Paren- chyma, temporal" T H E N Pathology is Meningioma rule 2: I F Appearance_of Enhancement = "Moder- ate regular marginal" and Location = "Brain parenchyma, temporal" T H E N Pathology is Astrocytoma rule 3: I F Edema < = "1/2 hemisphere" and Appearance_of Enhancement= "Mural nod- ule" and Location= "Brain parenchyma, temporal" T H E N Pathology is Anaplastic Protoplasmic Astrocytoma rule 4: I F Appearance o f Enhancement= "Homo- geneus with lucency inside" and Location = "Brain parenchyma, temporal" T H E N Pathology is Glioblastoma rule 5: I F Appearance_of_Enhancement = "Moder- ate regular marginal" and Location = "'Brain parenchyma, parietal" T H E N Pathology is Anaplastic Protoplasmic Astrocytoma rule6: I F Appearance_of_Enhancement="Homo- geneous with lucency inside" and Location = "Brain parenchyma, parietal" T H E N Pathology is Glioblastoma rule 7: I F Precontrast = "Iso" and Location- = "Brain parenchyma, occipital" T H E N Pathology is Meningioma rule 8: IF Bone_Change = "Sellar enlargement" and Location = "Sellar and suprasellar" T H E N Pathology is Pituitary A d e n o m a rule 9: I F B o n e C h a n g e = "Bony erosion'" and Location = "Sellar and suprasellar" T H E N Pathology is Meningioma rule 10: I F Precontrast= "Low" and Appearance_of_ Enhancement= "Grossly irregular" and Location = "Cerebellum, vermis" T H E N Pathology is Astrocytoma rule 11: I F Precontrast = "High" and Appearance_ o f Enhancement= "Grossly irregular" and Location = "Cerebellum, vermis" T H E N Pathology is Medulloblastoma rule 12: I F Appearance_of_Enhancement= "Homo- geneous with lucency inside '" and Location = "Cerebellum, vermis" T H E N Pathology is Medulloblastoma In the diagnostic process, BTDS can assist doctors in determining brain tumor etiologies according to the features extracted from computer tomography pictures. Doctors first inspect the patient's symptoms and input the symptoms as facts into the diagnostic system. The inference engine then searches for diagnostic rules that B r a i n T u m o r D i a g n o s t i c S y s t e m Input Event Encode (4) I Chromosome 1 I • • • N e w p o p u l a t i o n B e s t C h r o m o s o m e + E v e n t s I ~ m . , [ Chromosome m I (7)~ Refine Inference ~ Current (1) ~--~Knowledge BasoJ I Genetic Adaptive Search ~ ~ New (Knowledge Integration) Knowledge Bas©j-' L . . . . . . . . Append Encode and Append FIGURE 6. The knowledge refinement process. 359 match the patient's symptoms, and suggests a pathol- ogy. 6. K N O W L E D G E - R E F I N E M E N T U N I T A knowledge base integrated from multiple knowledge sources is often only a prototype, with an unsatisfactory classification accuracy. During the inference process, rules in a knowledge base must be refined to improve the effectiveness o f the knowledge-base system. In this section, a knowledge-refinement scheme is proposed to refine rules during the inference process. The knowledge-refinement unit uses the knowledge- integration procedure as the basis for refining knowledge. A flow chart for the refinement process is shown in Fig. 6. During inference, an input event wrongly classified by the current knowledge base is appended to the set o f test instances. It is also encoded as a bit string and appended to the current best rule set. The new test set, including the wrongly-classified element, is then presented to the genetic adaptive search algorithm to evaluate rule sets for a new population. The refinement process works until the exception events can be correctly classified by the knowledge base, making the new knowledge base more accurate than the old one. 7. I M P L E M E N T A T I O N The brain tumor diagnostic system was implemented in C language on a SUN SPARC/2 workstation. Ten initial knowledge items (rule sets) were obtained from different groups of experts using the knowledge-acquisition module, or derived from historical documents or current records o f actual instances via machine-learning meth- ods. The knowedge-integration module automatically integrated the ten initial rule sets into a comprehensive knowledge base. 348 real brain tumour cases were used to evaluate the performance of the knowledge base. After 2000 execution generations o f the genetic algorithm, an accuracy rate o f 91.42% was obtained, with 92 rules in the resulting knowledge base. The knowledge base must be continuously refined to i m p r o v e the accuracy if rnisclassification occurs. These rules were then refined during the process o f inference. Finally, an accuracy o f 95.58% was achieved, with 103 rules in the resulting knowledge base. 8. C O N C L U S I O N S This paper presents the design o f a self-integrating knowledge-based brain tumor diagnostic system. The brain tumor diagnostic system proposed consists o f three main units: knowledge building, knowledge inference, and knowledge refinement. Genetic techniques are also shown here to be good tools for knowledge integration and knowledge refinement. The system was successfully implemented on a Sun/SPARC 2 workstation. 348 real brain tumor cases were used to evaluate the performance o f the brain tumor diagnostic system, with a classifica- tion accuracy higher than 95%. We may then conclude that the brain tumor diagnostic system is a successful medical system. Acknowledgements---The authors would like to thank Dr M. M. H. Tseng and Dr O. Y Guo of VGH, for their advice about brain tumors. R E F E R E N C E S Baral, C., Kraus, S. & Minker, J. (1991). Combining multiple knowledge bases. IEEE Transactions on Knowledge and Data Engineering, 3, 208-220. Cendrowska, J. (1987). PRISM: An algorithm for inducing modular rules. International Journal o f Man-Machine Studies, 27, 349-370. De Jong, K. A. (1988). Learning with genetic algorithm: An overview. Machine Learning, 3, 121-138. 3 6 0 Gaines, B. R. (1989). Integration issues in knowledge supports systems. International Journal of Man-Machine Studies, 26, 497-515. Gragun, B. J. & Studel, H. J. (1987). A decision-table based processor for checking completeness and consistency in rule-base expert- systems. International Journal o f Man-Machine Studies, 26, 633-648. Gray, N. A. B. (1990). Capturing knowledge through top-down induction of decision trees. 1EEE Expert, 41-50. Holland, J. H. (1975). Adaptation in natural and artificial systems. Ann Arbor, MI: University of Michigan Press. Hwang, G. J. & Tseng, S. S. (1990). EMCUD: A knowledge acquisition method which captures embedded meanings under uncertainty. International Journal of Man-Machine Studies, 33, 431-451. Kelly, G. A. (1955). The psychology of personal constructs. New York: Norton. Liao, C. M. (1995). Using genetic algorithm technique for integrating multiple rule-sets. M.Sc. thesis, NCTU, Hsinchu, Taiwan. Medsker, L., Tan, M. & Turban, E. (1995). Knowledge acquisition from multiple experts: Problems and issues. Expert Systems With Ching Hung Wang et aL Applications, 9, 35-40. Michalski, R. S. & Chilausky, R. (1980). Learning by being told and learning from examples: An experimental comparison of the two methods of knowledge acquisition in the context of developing an expert system for soybean disease diagnosis. International Journal o f Policy Analysis and Information Systems, 4, 125-160. Mitchell, T. M, (1982). Generalization as search. Artificial Intelligence, 18, 203-226. Quinlan, J. (1986). Induction of decision tree. Machine Learning, 1, 81-106. Wang, C. H., Tseng, S. S. & Hong, T. P. (1995). Design of a self- adaptive brain tumor diagnostic system. Journal of Information Science and Engineering, 11, 275-294. Waterman, D. (1986). A guide to expert systems. Reading, MA: Addison-Wesley. Wills, K., Teather, D. & Innocent, P. (1982). An expert system for medical diagnosis of brain tumors. International Journal of Man- Machine Studies, 16, 341-349. Winston, P. H. (1992). Artificial intelligence (3rd edn). Reading,/VIA: Addison-Wesley. 
work_27g256ct6zf4fiivjunlac54um	----	pf-mim1 Toward Normative Expert Systems: Part I The Pathfinder Project David E. Heckerman Departments of Computer Science and Pathology University of Southern California HMR 204, 2025 Zonal Ave Los Angeles, CA 90033 <dheck@sumex-aim.stanford.edu> Eric J. Horvitz Palo Alto Laboratory Rockwell International Science Center 444 High Street Palo Alto, California 94301 <horvitz@sumex-aim.stanford.edu> Bharat N. Nathwani Department of Pathology University of Southern California HMR 204, 2025 Zonal Ave Los Angeles, CA 90033 To appear in Methods of Information in Medicine, 1992 1 Abstract Pathfinder is an expert system that assists surgical pathologists with the diagnosis of lymph-node diseases. The program is one of a growing number of normative expert systems that use probability and decision theory to acquire, represent, manipulate, and explain uncertain medical knowledge. In this article, we describe Pathfinder and our research in uncertain-reasoning paradigms that was stimulated by the development of the program. We discuss limitations with early decision-theoretic methods for reason- ing under uncertainty and our initial attempts to use non-decision-theoretic methods. Then, we describe experimental and theoretical results that directed us to return to reasoning methods based in probability and decision theory. Keywords: expert systems, decision making, diagnosis, probability theory, decision theory, artificial intelligence, pathology 2 1 Introduction Decision-theoretic or normative expert systems have the potential to provide better decision support than do traditional expert systems in problem areas or domains where the accurate management of uncertainty is important. This potential for improvement arises because people, including experts, make mistakes when they make decisions under uncertainty. That is, people often deviate from the rules of decision theory, which provides a set of compelling principles or desiderata for how people should behave when reasoning or making decisions under uncertainty. Decision theory includes the rules of probability and the principle that a person should always choose the alternative that maximizes his expected utility. Traditional expert systems provide decision support by mimicking the recommendations of experts. They do so by managing uncertainty with non-decision-theoretic methods. Such systems are valuable, because they provide important information to a nonexpert who is confronted with a confusing decision, and because they offer reminders to users who may be stressed or fatigued. Nonetheless, they tend to duplicate the errors made by experts. In contrast, normative expert systems use decision theory to manage uncertainty. The word “normative” comes from decision analysts and cognitive psychologists who emphasize the importance of distinguishing between normative behavior, which is what we do when we follow the desiderata of decision theory, and descriptive behavior, which is what we do when unaided by these desiderata. By encoding expert knowledge in a decision-theoretic framework, we can reduce errors in reasoning, and thereby build expert systems that offer recommendations of higher quality. In this article, we describe Pathfinder, a normative expert system that assists surgical pathologists with the diagnosis of lymph-node diseases [1, 2]. The Pathfinder project began in 1983 as a joint project among researchers at Stanford University (David Heckerman, Eric Horvitz, and Larry Fagan) and the University of Southern California (Bharat Nathwani—the primary pathology expert—and Keung-Chi Ng) [3]. Currently, a commercial derivative of Pathfinder, called Intellipath, is being used by practicing pathologists and by pathologists in training as an educational tool [4]. Also in this article, we discuss the importance of the proper management of uncertainty for diagnosis of lymph-node diseases; and we discuss our research in uncertain-reasoning paradigms that was stimulated by the development of Pathfinder. In particular, we examine practical limitations with early decision-theoretic methods for reasoning under uncertainty and our initial attempts to overcome these limitations through the use of non-decision- theoretic reasoning paradigms. Then, we describe experiments with these non-decision- theoretic approaches as well as theoretical analyses that directed us to return to a method- ology based in probability and decision theory. In the companion to this article, we describe the decision-theoretic representations that we developed to make practical the construction of a normative version of Pathfinder. 2 Diagnosis in Surgical Pathology and Pathfinder Surgical pathologists perform diagnosis primarily by examining sections of tissue microscop- ically. Sometimes, pathologists also incorporate clinical, radiology, and laboratory informa- 3 tion, and examine tissue with expensive tests derived from immunology, microbiology, and cell-kinetics research. Based on this information, the pathologist provides a diagnosis to the surgeons and oncologists who participate in the patient’s treatment. That is, the pathologist tells these physicians, “the patient has disease x.” The well-being of patients depends strongly on the accuracy of the pathologist’s diag- nosis. In the case of lymph-node diagnosis, for example, let us suppose that the patient has Hodgkin’s disease, a malignant disease, but that the pathologist makes a diagnosis of mononucleosis, a benign disease that can resemble Hodgkin’s disease. In this situation, the patient’s chance of death is significantly greater than it would have been had the diagno- sis been correct, because he does not receive immediate treatment for his malignancy. In contrast, let us suppose that the patient has mononucleosis, and that the pathologist makes a diagnosis of Hodgkin’s disease. In this case, the patient likely will undergo expensive, painful, and debilitating treatment, to be “cured,” only because he never had the malignant disease in the first place. A general pathologist performs diagnosis on tissue sections from all parts of the body. When a general pathologist has difficulty with diagnosis, he frequently refers the case to a subspecialist, who has expertise in the diagnosis of a particular tissue type. This referral process usually incurs both a delay in diagnosis and an extra cost. Sometimes, the delay in diagnosis is unacceptable, and the pathologist cannot refer the case to a subspecialist. For example, surgeons often rely on pathologists for the timely diagnosis of disease in frozen tissue taken from patients under anesthesia [5, 6]. The subspecialty of lymph-node diagnosis is one of the most difficult areas in surgical pathology [7, 8, 9, 10]. For example, one multisite oncology study analyzed almost 9000 cases of malignant lymphoma. The study found that although experts show agreement with one another, the diagnoses rendered by general pathologists for certain diseases had to be changed by expert lymph-node pathologists in as many as 65 percent of the cases [10]. Our goal in building Pathfinder is to close the wide gap between the quality of lymph- node diagnoses made by general pathologists and those made by subspecialists. We hope to increase the accuracy of in-house pathology diagnoses, to reduce the frequency of referrals, and to assist pathologists with intraoperative diagnosis when there is insufficient time for expert consultation. Pathologists have difficulty with diagnosis for two reasons. First, they may misrecognize or fail to recognize microscopic features. Second, they may combine evidence inaccurately to form a diagnosis. The second problem arises because the pathologist must consider an enormous number of features and diseases, and because the relationships among diseases and features are uncertain. Most of Pathfinder research has concentrated on the evidence- combination problem. That is, we have worked to develop an expert system that can help pathologists cope with the many uncertain relationships in the diagnosis of lymph-node pathology. Indeed, Pathfinder reasons about more than 60 diseases that can invade the lymph node (25 benign diseases, 9 Hodgkin’s lymphomas, 18 non-Hodgkin’s lymphomas, and 10 metastatic diseases), using more than 130 microscopic, clinical, laboratory, immunologic, and molecular-biologic features. Similarly, in this article, we concentrate on the problem of managing uncertainty in large domains. Nevertheless, as we mention in Section 9, we also have addressed the feature-recognition problem. 4 3 A Pathfinder Dialog In rendering a diagnosis, a pathologist (1) identifies and quantifies features; (2) constructs a differential diagnosis, a set of diseases consistent with the observations; and (3) decides what additional features to evaluate and what costly tests to employ to narrow the differential diagnosis. He repeats these steps until he has observed all useful features. This procedure is called the hypothetico-deductive approach [11, 12, 13, 14]. Cognitive psychologists have found that physicians frequently employ this approach in performing clinical diagnosis [12, 14]. Pathfinder uses this same method, summarized in Figure 1, to assist pathologists with their task of diagnosis. Associated with each feature are two or more mutually exclusive and exhaustive instances. For example, the feature NECROSIS is associated with the instances ABSENT, PRESENT, and PROMINENT. The Pathfinder system allows a user to report instances for one or more salient features of a lymph-node section. Given these feature–instance pairs, the system displays a differential diagnosis ordered by likelihood of diseases. In response to a query from the user, Pathfinder recommends a set of features that are the most cost effective for narrowing the differential diagnosis. The pathologist can answer one or more of the recommended questions. This process continues until the differential diagnosis is a single disease, there are no additional tests or questions, or a pathologist determines that the informational benefits are not worth the costs of further observations or tests. The operation of the latest version of Pathfinder is illustrated by the set of screen photos in Figure 2. Figure 2(a) shows the initial Pathfinder screen. The FEATURE CATEGORY window displays the categories of features that are known to the system, the OBSERVED FEATURES window displays feature–instance pairs that will be observed by the pathologist, and the DIFFERENTIAL DIAGNOSIS window displays the list of possible diseases and their probabilities. The probabilities in Figure 2(a) are the prior probabilities of disease—the probabilities for disease given only that a patient’s lymph node has been removed and is being examined. If the user selects (double-clicks) the feature category SPHERICAL FEATURES, then Pathfinder displays a list of features for that category. To enter a particular feature, the user double-clicks on that feature, and then selects one of the mutually exclusive and ex- haustive instances for that feature. For example, Figure 2(b) shows what happens when the user selects the feature F % AREA (percent area of the lymph-node section that is occupied by follicles). In the figure, a third window appears that lists the instances for this feature: NA (not applicable), 1–10%, 11–50%, 51–75%, 76–90%, and >90%. Figure 2(c) shows the result of selecting the last instance for this feature. In particular, the feature–instance F % AREA: >90% appears in the middle column, and the differential diagnosis is revised, based on this observation. As we mentioned, the user can continue to enter any number of features of his own selection. Figure 2(d) shows the Pathfinder screen after the user has reported that follicles are in a back-to-back arrangement and show prominent polarity. Alternatively, the user can ask the program to recommend additional features for observation. Figure 2(e) shows that the most cost-effective feature to evaluate, given the current differential diagnosis, is monocytoid cells. If the user observes that monocytoid cells are prominent, then we obtain the differential diagnosis in Figure 2(f). In this case, the four features in the middle column have narrowed the differential diagnosis to a single disease: the early phase of AIDS. 5 Evidence-Gathering Decisions Salient Features Diagnosis Continue? Differential Diagnosis Figure 1: Hypothetico-deductive reasoning in Pathfinder. First, the pathologist reports instances of salient features to the system. The system then constructs a differential diagnosis—a list of hypotheses that are consistent with the observations, and an assignment of likelihood to each such hypothesis. Next, the system analyzes the current differential di- agnosis to identify the most useful features for the pathologist to observe. The process cycles until the differential diagnosis is narrowed to a single disease, there are no additional tests or questions, or the pathologist determines that the informational benefits are not worth the costs of further observations or tests. 6 (a) (b) (c) (d) (e) (f) Figure 2: A Pathfinder consultation. (a) Initially, Pathfinder displays (from left to right) the categories of features, an empty window that will contain feature-instance pairs reported by the user, and the prior probabilities of disease. (b) Double-clicking on the category SPHERICAL STRUCTURES and then on the feature F % AREA, the pathologist prepares to report to Pathfinder the percent area occupied by follicles. (c) Double-clicking on the instance >90%, the pathologist reports that more than 90% of the lymph-node is occupied by follicles. In response, the program produces a differential diagnosis in the right-hand window. (d) The pathologist now reports that follicles are in a back-to-back arrangement and show polarity. Pathfinder revises the differential diagnosis. (e) The pathologist has asked Pathfinder to display features that are useful for narrowing the differential diagnosis. The program displays the four most cost-effective features for the user to observe next. The most useful feature is monocytoid cells. (f) The user now reports that monocytoid cells are prominent. Pathfinder determines that only a single disease—AIDS EARLY (the early phase of AIDS)—is consistent with the four observations. (Adapted with permission from D. Heckerman, Probabilistic Similarity Networks, MIT Press, Cambridge, MA, 1991.)7 Figure 3: A graphical justification for the recommendation of MONOCYTOID CELLS. For each instance of the feature, the length and direction of a bar reflects the change in the probability of AIDS EARLY relative to that of FLORID FOLLIC HYPERP, given the observation of that feature–instance pair. The justification also includes the monetary cost of observing the feature. (Taken with permission from D. Heckerman, Probabilistic Similarity Networks, MIT Press, Cambridge, MA, 1991.) Pathfinder explains graphically its recommendations for additional observations. A bitmap of Pathfinder’s graphical justification of the diagnostic utility of the feature mono- cytoid cells is displayed in Figure 3. In this explanation, Pathfinder displays the change in probability of the two most likely hypotheses given the observation of each possible instance of the feature. The graph indicates that if monocytoid cells are absent, then the probability of FLORID FOLLIC HYPERP relative to that of AIDS EARLY increases slightly; if monocytoid cells are present or prominent, then the probability of FLORID FOLLIC HYPERP relative to that of AIDS EARLY decreases greatly; and if monocytoid cells show confluence, then the probability of FLORID FOLLIC HYPERP relative to that of AIDS EARLY remains unchanged. By glancing at this graph, we can see that this feature is useful for discriminating these two diseases. The window also displays the monetary cost of evaluating the feature, which is negligible in this case. 8 4 Decision-Theoretic Computations in Pathfinder Both an early version and the latest version of Pathfinder employ decision-theoretic compu- tations to assist pathologists with diagnosis. In this section, we examine these computations. In particular, we examine how Pathfinder (1) uses probabilistic inference to generate a dif- ferential diagnosis, (2) uses decision theory to recommend a diagnosis, and (3) uses decision theory to recommend features for observation. First, however, let us discuss some funda- mentals of probability and decision theory. 4.1 Probability Theory Probability theory has roots, more than 3 centuries ago, in the work of Bernoulli, Laplace, Fermat, and Pascal [15]. The theory describes how to infer the probability of one event from the probability of related events. The prevalent conception of the probability of some event x is that it is a measure of the frequency with which x occurs, when we repeat many times an experiment that has x as a possible outcome. A more general notion, however, is that the probability of x represents the degree of belief held by a person that the event x will occur in a single experiment. If a person assigns a probability of 1 to x, then he believes with certainty that x will occur. If he assigns a probability of 0 to x, then he believes with certainty that x will not happen. If he assigns a probability of between 0 and 1 to x, then he is to some degree unsure about whether or not x will occur. The interpretation of a probability as a frequency in a series of repeated experiments traditionally is referred to as the objective or frequentist interpretation. In contrast, the interpretation of a probability as a degree of belief is called the subjective or Bayesian in- terpretation, in honor of the Reverend Thomas Bayes, a scientist from the mid-1700s who helped to pioneer the theory of probabilistic inference [16, 15]. Both interpretations follow the same set of mathematical rules. In the Bayesian interpretation, a probability or belief will always depend on the state of knowledge of the person who provides that probability. For example, if we were to give someone a coin, he would likely assign a probability of 1/2 to the event that the coin would show heads on the next toss. If, however, we convinced that person that the coin was weighted in favor of heads, he would assign a higher probability to the event. Thus, we write the probability of x as p (x|ξ), which is read as the probability of x given ξ. The symbol ξ represents the state of knowledge or background knowledge of the person who provides the probability. The conception of probability as a measure of personal belief is central to research on the use of probability and decision theory for representing and reasoning with expert knowl- edge in computer-based reasoning systems. There is usually no alternative to acquiring from experts the bulk of probabilistic information used in an expert system. For example, there are more than 14 thousand probabilities in the latest version of Pathfinder; and some of these probabilities are on the order of 10−6. Thus, performing the experiments necessary to determine objective probabilities for Pathfinder would entail much time and great ex- pense. Fortunately, when experimental data is available, the Bayesian approach provides a mechanism for expert systems to update their probabilities, given this data [17, 18, 19]. 9 4.2 Decision Theory and Utility Assessment Decision theory extends the Bayesian interpretation of probability theory, and prescribes how a decision maker should choose among a set of alternatives or actions, given his utility or pref- erence for each possible outcome and his belief that each outcome will occur. In particular, decision theory includes the rules of probability theory and the maximum-expected-utility (MEU) principle, which states that a decision maker should always choose the alternative that maximizes his expected utility [20]. Utility assessment is nontrivial and is the subject of many debates. In this article, we mention only a few important points concerning utility assessment for Pathfinder. The interested reader should consult more general discussions by Keeney and Raiffa [21], McNeil et al. [22], and Howard [23]. For each disease pair (dj, dk) in Pathfinder, we encode the utility u(dj, dk), which sum- marizes the preferences of the decision maker for the situation in which a patient has disease dj, but is diagnosed as having disease dk. Factors that influence such preferences include the length of the patient’s expected life, the pain associated with treatment and with the disease itself, the psychological trauma to the patient and his family, and the monetary cost associated with treatment and with disability. An important consideration in the assessment of these (and any other) utilities is: Who is the decision maker? From our perspective, a pathologist is only a provider of information. Thus, the u(dj, dk) in the utility model of a computer-based diagnostic system should reflect the patient’s preferences. For example, consider the situation where a pathologist believes, after reviewing a case, that the probability of the benign infection mononucleosis is 0.9, and that the probability of Hodgkin’s disease is 0.1. Should the patient be treated for Hodgkin’s disease now, or should he wait for more definitive diagnostic signs to develop? As we discussed, delaying treatment of Hodgkin’s disease decreases the chances of long-term survival if the patient has this condition. On the other hand, the treatment for Hodgkin’s disease is highly invasive. The decision about therapy will depend on how the patient feels about the alternative outcomes. Different patients may have dramatically different preferences. As we discuss in Sections 4.4 and 4.5, differences in patient preferences can in principle affect recommendations made by an expert system for diagnosis. Thus, utility assessment poses a fundamental problem to any researcher who wants to develop such expert systems. Specifically, whenever a patient case is processed by an expert system, the system or a decision analyst should assess the utilities of that patient and provide these utilities to the system. Such utility assessment would be extremely time consuming and expensive. As we see in Sections 4.4 and 4.5, however, only Pathfinder’s diagnostic recommendations and not its recommendations for evidence gathering are sensitive to patient utilities. Thus, by allowing Pathfinder to make only evidence-gathering recommendations, we render the program’s recommendations insensitive to patient utilities. We can therefore encode in Pathfinder the utilities u(dj, dk) from one representative patient. To construct Pathfinder’s utility model, we assessed the utilities of Bharat Nathwani, the primary Pathfinder expert. We found it relatively easy to assess his utilities, because he was familiar with the ramifications of many specific correct and incorrect diagnoses. Another important consideration in utility assessment is the wide range of severities as- 10 sociated with outcomes. For example, if a patient has a viral infection and is incorrectly diagnosed as having cat-scratch disease—a disease caused by an organism that is killed with antibiotics—the consequences are not severe. In fact, the only non-negligible consequence is that the patient will take antibiotics unnecessarily for several weeks. If a patient has Hodgkin’s disease and is incorrectly diagnosed as having mononucleosis, however, the con- sequences are often lethal. It is important for us to measure preferences across such a wide range, because sometimes we must balance a large chance of a small loss with a small chance of a large loss. For example, even though the probability that a patient has syphilis is small—say, 0.001— treatment with antibiotics may be appropriate, because the patient may prefer the harmful effects of antibiotics to the small chance of the harmful effects of untreated disease. Early attempts to assess preferences for both minor and major outcomes in the same unit of measurement were fraught with paradoxes. For example, in a linear willingness-to-pay approach, a decision maker might be asked, “How much would you have to be paid to accept a one in ten-thousand chance of death?” If the decision maker answered, say, $1000, then the approach would dictate absurdly that he would be willing to be killed for $10 million. Howard (1980) constructed an approach that avoids many of the paradoxes of earlier models. Like several of its predecessors, the model determines what an individual is willing to pay to avoid a given chance of death, and what he is willing to be paid to assume a given chance of death. Also, like many of its predecessors, Howard’s model shows that, for small risks of death (typically, p < 0.001), the amount someone is willing to pay to avoid, or is willing to be paid to assume, such a risk is linear in p. That is, for small risks of death, an individual acts as would an expected-value decision maker with a finite value of life, called the small-risk value of life. For significant risks of death, however, the model deviates strongly from linearity. For example, the model shows that there is a maximum probability of death, beyond which an individual will accept no amount of money to risk that chance of death. Most people find this result to be intuitive.1 To use this model, we first determined Bharat Nathwani’s small-risk value of life. When asked what dollar amount he would be willing to pay to avoid chances of death ranging from 1 in 20 to 1 in 1000, he was consistent with the linear model to within a factor of 2, with a median small-risk value of life equal to $20 million (in 1988 dollars). To make the application of the model more convenient, we used Howard’s definition of a micromort: one–in–1-million chance of death [24]. In these units, the Pathfinder expert’s small-risk value of life was $20 per micromort.2 Given this small-risk value of life, we could then measure his preferences for major and minor outcomes in a common unit: 1 minus the probability of immediate, painless death that he was willing to accept to avoid a given outcome and to be once again healthy. In particular, we assessed his preferences for minor outcomes with willingness-to-pay questions, such as “How much would you be willing to pay to avoid taking antibiotics for two weeks?” We then translated these answers, via the linearity result of Howard’s model, to units of probability 1The result makes several assumptions, such as the decision maker is not suicidal and is not concerned about how his legacy will affect other people. 2In general, the micromort is a useful unit of measurement, because it helps to emphasize that the linear relationship between risk of death and willingness to pay holds for only small probabilities of death. 11 of death. For example, an answer of $100 translated to a utility of 1 − 5 micromorts = 1 − 0.000005 = 0.999995 We assessed his preferences for major outcomes directly in units of probability of death. For example, he imagined that he had—say—Hodgkin’s disease, and that he had been misdiag- nosed as having mononucleosis. He then imagined that there was a magic pill that would rid him of this disease with probability 1 − p, but would kill him, immediately and painlessly, with probability p. He then provided the value of p that made him indifferent between his current situation and the situation in which he takes the pill. The utility of this outcome is 1 − p. 4.3 Construction of a Differential Diagnosis First, let us examine the problem of differential diagnosis in general. Let m and n denote the number of diseases and features in a medical domain, respectively. Also, let d1, d2, . . . , dm denote the disease entities. For the moment, let us suppose that each disease dj may be present or absent. Let Dk denote some instance of diseases. That is, Dk denotes some assignment of present or absent to each of the diseases d1, d2, . . . , dm. Further, let f1, f2, . . . , fn denote the features in the domain, and let ij denote the observed instance for the jth feature. Now imagine that a user of a probabilistic expert system for this domain has observed instances for q features. To simplify the notation, let us renumber the n features so that the user has observed instances for the first q features. Typically, the user will want to know the probability of each disease instance, given the observations f1i1, f2i2, . . . , fqiq. This quantity for disease instance Dk is known as the posterior probability of Dk, and is denoted p(Dk|f1i1, f2i2, . . . , fqiq, ξ) (1) Thus, the number of probabilities of interest is exponential both in the number of observed features and in the number of diseases. In principle, an expert could assess directly these posterior probabilities. Aside from the intractable nature of this task, however, most physicians are more comfortable assessing probabilities in the opposite direction. That is, they are more comfortable assessing the probabilities that the set of observations f1i1, f2i2, . . . , fqiq will appear given a particular disease instance Dk, denoted p(f1i1, f2i2, . . . , fqiq|Dk, ξ) (2) Using Bayes’ theorem, the expert system can compute from these probabilities and the prior probability of disease instances p(Dk|ξ) the desired posterior probabilities p(Dk|f1i1, f2i2, . . . , fqiq, ξ) = p(f1i1, f2i2, . . . , fqiq|Dk, ξ) p(Dk|ξ) P Dl p(f1i1, f2i2, . . . , fqiq|Dl, ξ) p(Dl|ξ) (3) where the sum over Dl runs over all disease instances. Unfortunately, this approach to the problem is also intractable, because the number of probabilities of the form 12 p(f1i1, f2i2, . . . , fqiq|Dk, ξ) is exponential both in the number of diseases and in the number of features. To manage the complexity of the general case, researchers who built the first proba- bilistic expert systems made two assumptions. First, they supposed that all findings were conditionally independent, given any disease instance. That is, they assumed that, if the true disease state of the patient was known, then the likelihood of seeing any observation fkik did not depend on observations made about any other features. Thus, p(fjij |Dk, f1i1, . . . , fj−1ij−1, fj+1ij+1, . . . , fqiq, ξ) = p(fjij |Dk, ξ) (4) Given this assumption, it follows from the rules of probability [25] that p(f1i1, f2i2, . . . , fqiq|Dk, ξ) = p(f1i1|Dk, ξ) p(f2i2|Dk, ξ) · · · p(fqiq|Dk, ξ) (5) Second, these researchers supposed that the traditional disease entities were mutually ex- clusive and exhaustive. That is, they assumed that each disease instance corresponded to a situation where only one disease was present. Given these two assumptions, the expert system can compute the posterior probabilities of disease from the tractable computation p(dk|f1i1, f2i2, . . . , fqiq, ξ) = p(f1i1|dj, ξ) p(f2i2|dk, ξ) · · · p(fqiq|dk, ξ) p(dk|ξ) P dl p(f1i1|dl, ξ) p(f2i2|dl, ξ) · · · p(fqiq|dl, ξ) p(dl|ξ) (6) where dk represents the disease instance in which only disease dk is present. Thus, only the conditional probabilities p(fjij |dk, ξ) and the prior probabilities p(dk|ξ) are required for the computation. We call any model that employs these two assumptions a simple-Bayes model. Ledley and Lusted proposed this model for medical diagnosis in 1959 [26]. In the domain of lymph-node pathology, the assumption that diseases are mutually ex- clusive is appropriate, because co-occurring diseases almost always appear in different lymph nodes or in different regions of the same lymph node. Also, the large scope of Pathfinder makes reasonable the assumption that the set of diseases is exhaustive. The assumption of global conditional independence, however, is inaccurate. For example, given certain diseases, finding that follicles are abundant in the tissue section increases greatly the chances that sinuses in the interfollicular areas will be partially or completely destroyed. Nonetheless, to simplify our task, we used the simple-Bayes model to construct the first probabilistic version of Pathfinder. Later, after developing several graphical representation languages that we describe in the companion to this article, we encoded successfully the conditional noninde- pendence or conditional dependence among the features in the domain. We shall return to this discussion in Section 8. 4.4 Recommendation of a Diagnosis As we mentioned, a diagnosis is a statement of the form: “The patient has disease x.” Sometimes, as we saw in the patient case in Section 3, the posterior probability of one disease will equal 1 and the posterior probability of all other diseases will equal 0. In this case, making a diagnosis is not a decision. Rather, the diagnosis is a consequence of the rules of logic. In most cases, however, observations usually do not narrow the differential 13 diagnosis to a single disease. In these situations, making a diagnosis is a decision: an irrevocable allocation of resources under uncertainty. Using the MEU principle, the system can determine a diagnosis from the probabilities of disease and the utilities u(dj, dk). Let φ denote the set of feature–instance pairs f1i1, f2i2, . . . , fqiq that we have observed thus far. First, for each diagnosis dk, the system computes eu(dk|φ), the expected utility of that diagnosis given observations φ, using the formula eu(dk|φ) = X dj p(dj |φ) u(dj, dk) (7) To complete the determination, the system selects the optimal diagnosis, denoted dx(φ), using the equation dx(φ) = argmaxdk [eu(dk|φ)] (8) where the function argmaxdk [·] returns the disease that maximizes its argument. We do not allow Pathfinder to recommend diagnoses, because we have observed that such recommendations are somewhat sensitive to the utility model. That is, when we change the utilities in the model from values appropriate for one patient to values appropriate for another patient, the program’s recommendations can change significantly. By preventing Pathfinder from recommending diagnoses, we hope to encourage a change in the way pathologists and care-providing physicians communicate. In the short term, we hope that pathologists will begin to express clearly—in the language of probability—uncertainty associated with their observations. In the long term, we hope that each physician who is associated with the care of a patient—including the primary physician, the pathologist, the radiologist, the surgeon, the oncologist, and the radiotherapist—and the patient himself will communicate in decision- theoretic terms to determine the best treatment for that patient. Such communication could take place via a shared decision model embodied in an expanded version of Pathfinder. 4.5 Recommendation of Features to Narrow a Differential Diag- nosis Let us now consider how an expert system can use decision theory to recommend features for observation to narrow a differential diagnosis. First, the system enumerates all possible observation strategies. An example of an observation strategy is Observe f3. If f3 is present, then observe f2; otherwise, make no further obser- vations and make the diagnosis. If f3 and f2 are present, then observe f7, and make the diagnosis. If f3 is present and f2 is absent, then make the diagnosis. Next, the system computes the decision maker’s expected utility of all strategies, including the strategy in which the user observes no additional features. Finally, the system chooses the strategy that maximizes the decision maker’s expected utility. In practice, this approach is unfeasible, because there are more than 2n strategies for n unobserved features. To make computations tractable, both the old and new versions of Pathfinder employ the myopic approximation, introduced by Gorry and Barnett in 1968 [27]. In this approximation, a system identifies the best single feature to observe, by maximizing 14 the expected utility of the decision maker under the assumption that a diagnosis will be made after the user observes only one feature. Once the user observes the feature, the system repeats the myopic analysis, and may recommend additional features for observation. Let us examine formally the computation in Pathfinder. First, the system computes eu(dx(φ)|φ), the expected utility of the optimal diagnosis when the user observes no addi- tional features. From Equations 7 and 8, we have eu(dx(φ)|φ) = X dj p(dj |φ) u(dj, dx(φ)) (9) Now the system imagines that the user observes an additional feature fnew. Let φ 0 denote the union of the original set of observations and the observation for fnew. Pathfinder now identifies the optimal diagnosis, given the new set of observations: dx(φ 0 ) = argmaxdk   X dj p(dj |φ 0 ) u(dj, dk)   (10) The expected utility of this diagnosis, denoted eu(dx(φ 0 )|φ0), is given by eu(dx(φ 0 )|φ 0 ) = X dj p(dj |φ 0 ) u(dj, dx(φ 0 )) (11) In contrast, the expected utility of the original diagnosis, given observations φ 0 , is given by eu(dx(φ)|φ 0 ) = X dj p(dj |φ 0 ) u(dj, dx(φ)) (12) The quantity eu(dx(φ)|φ0) is never greater than the measure eu(dx(φ0)|φ0), because, by definition, the diagnosis dx(φ 0 ) maximizes expected utility, given the observations φ 0 . The system now computes the value of information of observing fnew, denoted vi(fnew|φ), which is the difference between eu(dx(φ 0 )|φ0) and eu(dx(φ)|φ0) averaged over the instances inew of the feature fnew. 3 That is, vi(fnew|φ) = X inew p(φ 0 |φ) h eu(dx(φ 0 )|φ 0 ) − eu(dx(φ)|φ 0 ) i (13) The value of information of observing fnew represents the largest amount that the decision maker would be willing to pay to observe fnew. This quantity is always greater than or equal to 0. Next, the system computes the net value of information of observing fnew, denoted nvi(fnew|φ), by subtracting the cost4 of observing fnew from the value of information of observing fnew. That is, nvi(fnew|φ) = vi(fnew|φ) − cost(fnew). (14) Finally, if there is at least one feature that has a net value of information greater than 0, Pathfinder recommends the feature for observation that has the highest net value of 3This definition and the definition of net value of information are appropriate for expected-value decision makers. Howard discusses the general case [28, 29]. 4We measured costs in dollars and then converted these costs to units of probability via Howard’s model. 15 information. Otherwise, the system suggests that the user should gather no additional evidence and make a diagnosis. In principle, the myopic approximation could affect the diagnostic accuracy of an expert system. For example, suppose that two features remain unobserved. In this case, the net value of information for the feature pair could exceed 0, and thus the user should observe the features. A value-of-information analysis on each feature alone, however, may indicate that neither feature is cost effective for observation. Consequently, the user would fail to observe these features, and thereby possibly make an incorrect diagnosis. Nonetheless, there is evidence that the myopic approximation does not often cause this problem in practice. For example, Gorry and Barnett have demonstrated that the approximation does not dimin- ish significantly the diagnostic accuracy of their program that assists physicians with the diagnosis of congenital heart disease [27]. In addition, although we have not yet conducted a similar experiment with Pathfinder, our expert almost always has been impressed by the questions generated by the myopic approximation. As we mentioned in previous sections, Pathfinder’s diagnostic recommendations are sen- sitive to the utility model, and we therefore do not allow Pathfinder to make such recom- mendations. Fortunately, however, we have found in an informal study that Pathfinder’s recommendations for evidence gathering are insensitive to the model. In particular, we have found that Pathfinder’s recommendations often are similar to those made by a second version of the program in which u(dj, dk) is equal to 1 when both dj and dk are benign diseases, u(dj, dk) is equal to 1 when both dj and dk are malignant diseases, and u(dj, dk) is equal to 0 otherwise. Consequently, we allow Pathfinder to make recommendations for evidence gathering. The observation that recommendations for evidence gathering are less sensitive to the utility model than are diagnostic recommendations may be due to the fact that more factors contribute to the computation of net value of information than to the computation of the diagnosis. That is, only the probabilities of diseases and the utilities u(dj, dk) contribute to the computation of the diagnosis, whereas these factors, the information content of a feature, and the cost of a feature contribute to the computation of net value of information. 5 Alternative Reasoning Methodologies The first medical expert systems employed computations based in probability theory. In particular, throughout the 1960s and early 1970s, medical expert systems used the simple- Bayes model to construct differential diagnoses. These systems included Warner’s system for the diagnosis of heart disease [30], Gorry’s program for the management of acute renal failure, and deDombal’s system for the diagnosis of acute abdominal pain [31]. Evaluations of most of these early systems showed that the programs performed well. In fact, the diagnoses rendered by several of them were more accurate than were those made by experienced physicians [31]. Nonetheless, in the early 1970s, researchers began to criticize these systems. They noted that the domains of these programs were small and did not reflect realistic clinical situations. Furthermore, researchers argued that errors due to the erroneous assumptions of the simple-Bayes model would become unacceptable as the domains of these systems were expanded [32, 33]. One group of investigators showed that 16 the diagnostic accuracy of an expert system based on the simple-Bayes model deteriorated significantly as the number of features in the system increased. These investigators traced the degradation in performance to violations of the conditional-independence assumptions in the simple-Bayes model [34]. Another group of researchers showed that the assumption of global conditional independence could be unrealistic in small domains as well [35]. In the early 1970s, perceptions of the inadequacy of the early decision-theoretic systems led to the development of alternative methods for reasoning under uncertainty [36, 37, 32]. Many of these developments occurred in the field of Artificial Intelligence in Medicine. Some of the alternative methods were ad hoc mechanisms, designed as custom-tailored techniques for particular domains or systems. These approaches included the MYCIN certainty-factor model and the QMR scoring scheme. Other methods were developed as alternative theoret- ical formalisms, such as the Dempster–Shafer theory of evidence and fuzzy decision theory. In the first year of Pathfinder research, we appreciated the limitations of the simple-Bayes model, and believed that the use of this model would significantly impair the performance of Pathfinder. Consequently, we examined several alternative reasoning methodologies. In this section, we introduce these approaches. In the following two sections, we describe our empirical and theoretical analyses of these methodologies in the context of Pathfinder. A well-known ad hoc method for managing uncertainty is the certainty-factor (CF) model [38]. Shortliffe and Buchanan designed the model to augment the rule-based approach to reasoning for MYCIN, a program for the diagnosis and treatment of bacteremias and meningitis [33]. In using the model, an expert attaches a certainty factor to each if–then rule. The certainty factor represents the expert’s change of the belief in the consequent of the rule, given the antecedent of the rule. In particular, a CF between 0 and 1 means that the expert’s belief in a consequent increases if the antecedent is true, whereas a CF between -1 and 0 means that the expert’s belief decreases. In a rule base, the consequent of one rule may serve as the antecedent of another rule. In addition, two or more rules may share the same antecedent or consequent. As a result, a rule base forms an inference network: a directed graph in which an arc from proposition a to proposition b corresponds to the rule “if a then b.” The CF model prescribes a method for propagating certainty factors through such a network. That is, given an observation of an antecedent in the network, we can use CF-model formulas to compute the effective certainty factor for any consequent in the network that is a descendent of that antecedent. Although the CF model was designed for MYCIN, the model has found many applications in other domains. Today, the model is the most popular method for managing uncertainty in rule-based systems. Quick Medical Reference or QMR (formerly Internist-1) uses another ad hoc method for managing uncertainty [39, 40]. The QMR project, now in its eighteenth year, assists internists with the diagnosis of more than 600 diseases, through the consideration of approximately 4000 manifestations or features of disease. In QMR, each feature has two instances; in particular, a feature is either absent or present. More important, each disease can be either absent or present. Thus, QMR can address cases in which more than one disease is present. The ad hoc scoring scheme employs two measures to represent the degree of association between a feature and a disease: an evoking strength and a frequency. The evoking strength for a given feature–disease pair represents the degree to which the presence of the disease causes that feature to be present. The frequency for a given feature–disease pair represents how often that feature is present in patients who have the disease. In addition, the scheme 17 represents the import of each feature, which is inversely proportional to the likelihood that an insignificant disease (such as the common cold) can cause the feature to be present. Given an assignment of present and absent to a subset of features, QMR uses evoking strengths, frequencies, and imports to assign a score to each disease. QMR then displays diseases in order of descending score. Like early decision-theoretic systems, QMR uses a hypothetico- deductive approach to reasoning. In particular, the system contains ad hoc algorithms for generating useful recommendations for additional evidence gathering based on the current differential diagnosis and on evoking strengths, frequencies, and imports. A more theoretical alternative to probabilistic reasoning was developed by Dempster and extended by Shafer [41, 42]. The approach, now called the Dempster–Shafer theory of evidence, was motivated by theoretical objections to the decision-theoretic approach [43]. Nonetheless, many artificial-intelligence (AI) researchers adopted special cases of the ap- proach to avoid the perceived computational intractability of decision theory [44, 45]. Cur- rently, the theory has many interpretations [42, 46, 47, 48]. One of the most popular inter- pretations is that given in Shafer’s original text. In this interpretation, an expert assesses the degree of support that a piece of evidence lends to hypotheses in the frame of discern- ment: a set of mutually exclusive and exhaustive hypotheses. He does so for every piece of evidence that may be observed. A combination rule can then be used to compute the degree of support that multiple pieces of evidence lend to hypotheses in the frame of discernment. In the interpretation, an expert assesses degrees of support for a single piece of evidence by constructing a basic probability assignment over the frame. That is, he assigns a mass, rang- ing from 0 to 1, to each subset of hypotheses in the frame. The mass for a particular piece of evidence and a subset of hypotheses represents the degree of support that the evidence lends to the subset. Like the CF model and the QMR scoring scheme, the Dempster–Shafer theory manipulates measures of change in belief. In Section 7.1, we examine the relationship among these methodologies. Another theoretical approach for managing uncertainty is fuzzy decision theory [49]. The theory addresses the presence of ambiguous terms such as “large” and “tall” in the specification of decision problems. Fuzzy decision theorists do not object to the use of probability theory or decision theory when events are defined precisely. They argue, however, that it is desirable to reason in situations where there is imprecision in the definition of events in addition to uncertainty about their occurrence. An example of a fuzzy decision problem is as follows: An urn contains many balls of various sizes, of which several are large. To draw a ball, you must pay a small sum of money. If you draw a large ball, however, you will win a valuable prize. Should you draw the ball? 6 Empirical Study of Reasoning Methods During the first year of Pathfinder research, we experimented with the reasoning method- ologies described in the previous section. In this section, we examine the results of those experiments. 18 6.1 Rule-Based Reasoning The first version of Pathfinder was a rule-based system that employed propositional logic for reasoning. After informally evaluating the system, we discovered two related problems. First, the program did not take into account the uncertainty associated with the relationships between observations and diseases. This deficiency of the program became apparent to us almost immediately. Indeed, as we have mentioned, proper management of uncertainty is crucial to accurate diagnosis in the domain of lymph-node pathology. We might have addressed this concern with the use of the CF model. We found, however, another problem with the rule-based approach, which forced us to abandon the methodology. In particular, our expert was frustrated by the system, because it asked many questions that were irrelevant to discriminating the diseases on the current differential diagnosis. This behavior was a result of the fact that the rule-based methodology generated recommendations for additional observations based on a fixed traversal through the rule base. As a result of our informal evaluations, we searched for a more flexible approach to the overall control of diagnostic reasoning. We discovered literature describing the hypothetico- deductive approach and systems such as QMR and Gorry’s diagnostic program that im- plemented the approach. We decided to construct a new version of Pathfinder modeled after QMR. Nonetheless, we were not satisfied with the scoring scheme of QMR because it had no theoretical foundation; and we searched for a more principled method for managing uncertainty. 6.2 Fuzzy Reasoning We considered fuzzy decision theory as a possible reasoning methodology for Pathfinder, but quickly rejected its use. We did so, because we found that neither general pathologists nor experts in hematopathology agreed on the meanings of fuzzy descriptions of feature– instance pairs such as “mild capsule thickening,” “rare Sternberg-Reed cells,” or “prominent necrosis.” For example, one expert stated that Sternberg-Reed cells were “rare” when there were one to five of these cells in any 4-square-centimeter section of a lymph node. Another expert stated that these cells were “rare” when there were one to ten of these cells in any 4-square-centimeter section. We did not believe that fuzzy decision theory—a scheme devised by researchers who were unfamiliar with the domain of lymph-node pathology—nor any other mechanism would provide meaningful inferences, if we continued to employ these fuzzy feature-instance descriptions. Instead, we asked the four hematopathology experts—Drs. Costan Berard, Jerome Burke, Ronald Dorfman, and Bharat Nathwani—to clarify the meanings of the descrip- tions that they were using. Although, as we have just discussed, the experts’ interpretations did not coincide initially, the experts did not find it difficult to construct unambiguous inter- pretations for each feature instance. The experts handled disagreements in a manner similar to that used by coauthors of a manuscript who are faced with a disagreement. That is, when their initial interpretations of a feature instance did not coincide, each pathologist put forth an argument for the merits of his interpretation of that feature instance. Then, in most cases, the four experts accepted unanimously one interpretation. When the experts could not agree, the primary author of the system (Bharat Nathwani) selected the interpretation. 19 6.3 The Dempster–Shafer Theory of Evidence Next, we examined the Dempster–Shafer theory of evidence. We discovered that the theory could be interpreted as a methodology for combining measures of change in belief, as could the CF model and QMR scoring scheme. We were attracted to the methodology, however, because it appeared to be a more principled approach to uncertainty management. Conse- quently, we constructed the second version of Pathfinder, using the Dempster–Shafer theory. In particular, we implemented a special case of the theory described by Barnett [44]. In this approach, the expert assigned nonzero masses only to (1) singleton subsets of the frame of discernment and (2) the entire frame of discernment. We refer to this simplified approach as the Dempster–Shafer–Barnett model. We then evaluated informally this version of Pathfinder by allowing the expert to exercise the system with real and imaginary cases. The expert was satisfied with the diagnostic accuracy of the system. At this time, probability theory was low on our list as a method for combining evi- dence to build a differential diagnosis, because of the limitations of the simple-Bayes model. Nevertheless, we were interested in experimenting with probabilistic reasoning, given the pio- neering work of Ledley, Lusted, Gorry, Barnett, and others. We re-examined the measures of uncertainty that we had assessed from our expert, and realized that these measures could be interpreted in probabilistic terms. We implemented the simple-Bayes model in Pathfinder, without assessing additional measures of uncertainty.5 We then compared the performance of the Dempster–Shafer–Barnett and simple-Bayes versions of Pathfinder. Without informing the expert, we switched the scoring scheme of Pathfinder from the Dempster–Shafer–Barnett approach to the simple-Bayes model. To our surprise, after running several cases with the probabilistic scheme, the expert exclaimed excitedly that the diagnostic accuracy of the program had improved significantly. Several years later, in a formal study, we compared the diagnostic accuracy of the Dempster–Shafer–Barnett, simple-Bayes, and CF models in the domain of lymph-node pathol- ogy. We verified our informal observation that the simple-Bayes model provided greater diagnostic accuracy (i.e., greater agreement with the expert) than did the Dempster–Shafer– Barnett model. We also found that the simple-Bayes model provided greater diagnostic accuracy than did the CF model [50]. 7 Theoretical Study of Reasoning Methods Surprised with the dominance of the simple-Bayes model over the alternative methods, we investigated the relationship between probability theory and alternative reasoning strategies over the next two years. We also studied the theoretical justifications for probabilistic and decision-theoretic reasoning. 5We assumed that the prior probability of each disease was equal. 20 7.1 Probabilistic Interpretations of Alternative Reasoning Meth- ods Heckerman and other researchers examined the relationship of the CF, QMR, and Dempster– Shafer–Barnett models with a simple probabilistic model for manipulating measures of change in belief called the odds–likelihood updating scheme. To understand the probabilistic model, let us suppose that we have a single disease d that can be true (d+) or false (d−). Further, suppose that we have n features f1, . . . , fn, where each feature can be present or absent. Let us apply the simple-Bayes model to this situation. That is, let us assume that all features are conditionally independent, given d+ and given d−. Thus, as in Section 4.3, we can use Bayes’ theorem to compute the posterior probability of d+; we obtain p(d+|f1, . . . , fn, ξ) = p(f1|d+, ξ) · · · p(fn|d+, ξ) p(d+|ξ) p(f1|d+, ξ) · · · p(fn|d+, ξ) p(d+|ξ) + p(f1|d−, ξ) · · · p(fn|d−, ξ) p(d−|ξ) (15) where any fi can be present or absent. In addition, we can apply Bayes’ theorem to compute the posterior probability of d−; we get p(d−|f1, . . . , fn, ξ) = p(f1|d−, ξ) · · · p(fn|d−, ξ) p(d−|ξ) p(f1|d+, ξ) · · · p(fn|d+, ξ) p(d+|ξ) + p(f1|d−, ξ) · · · p(fn|d−, ξ) p(d−|ξ) (16) When we divide Equation 15 by Equation 16, we obtain p(d+|f1, . . . , fn, ξ) p(d−|f1, . . . , fn, ξ) = p(f1|d+, ξ) p(f1|d−, ξ) · · · p(fn|d+, ξ) p(fn|d−, ξ) p(d+|ξ) p(d−|ξ) (17) We can rewrite Equation 17 as O(d+|f1, . . . , fn, ξ) = λ(f1, d+|ξ) · · · λ(fn, d+|ξ) O(d+|ξ) (18) where O(d+|ξ) = p(d+|ξ)p(d−|ξ) and O(d+|f1, . . . , fn, ξ) = p(d+|f1,...,fn,ξ) p(d−|f1,...,fn,ξ) (19) are the prior and posterior odds of d+, respectively, and λ(fi, d+|ξ) = p(f1|d+, ξ) p(f1|d−, ξ) (20) is the likelihood ratio for d+, given fi. Equation 18 is the odds–likelihood updating scheme. Heckerman showed that we can interpret the certainty factor for the rule “if fi then d+,” denoted CF (fi → d+|ξ) as a monotonically increasing function of the likelihood ratio λ(fi, d+|ξ). In particular, he showed that, if we make the identification CF (fi → d+|ξ) =    λ(fi,d+|ξ)−1 λ(fi,d+|ξ) λ(fi, d+|ξ) ≥ 1 λ(fi, d+|ξ) − 1 λ(fi, d+|ξ) < 1 (21) then the odds–likelihood updating scheme is identical to the formula in the CF model for combining certainty factors that is applied when a set of rules share the same consequent. 21 In addition, Heckerman showed that with the identification in Equation 21, the remaining formulas in the CF model are a close approximation to the rules of probability [51]. Grosof then showed that the Dempster–Shafer–Barnett model was isomorphic to the odds–likelihood updating scheme [52] via a different transformation of the likelihood ratio. In addition, Heckerman and Miller demonstrated that QMR’s ad hoc scoring scheme was isomorphic to the odds–likelihood updating scheme [53].6 These theoretical results helped us to understand the dominance of the simple-Bayes model over the nonprobabilistic alternatives. In particular, the other approaches did not avoid the assumptions of conditional independence of the simple-Bayes model; they merely obscured the assumptions. In fact, these approaches assumed that evidence was conditionally independent, given each disease and given the negation of each disease. When there are more than two mutually exclusive and exhaustive diseases in a domain, theses conditional independence assumptions are stronger than are the assumptions in the simple-Bayes model [58, 51]. We can understand the limitations of the alternative scoring schemes at a more intuitive level. Let us consider rule-based inference, in particular. The first rule-based inference schemes used the rules of logic. As a result, these schemes enjoyed a property known as modularity. That is, given the logical rule “if a then b,” and given that a is true, we can assert that b is true no matter how we established that a is true, and no matter what else we know to be true. For example, given the rule if l1 and l2 are parallel lines then l1 and l2 do not intersect we can assert that l1 and l2 do not intersect once we know that l1 and l2 are parallel lines. This assertion satisfies the property of modularity: The assertion depends on neither how we came to know that l1 and l2 are parallel, nor what else we know. The CF model is an extension of the rules of logic that imposes this same principle of modularity on inferences. For example, given the rule if PERITONITIS then APPENDICITIS, CF = 0.7 and given that a patient has peritonitis, the CF model allows us to increase the likelihood that the patient has appendicitis by the amount corresponding to a CF of 0.7, no matter how we establish that peritonitis is present. Given the correspondences described in the first part of this section, we see that the odds–likelihood updating scheme, the QMR scoring scheme, and the Dempster–Shafer–Barnett model also incorporate the property of modularity. We shall refer to these methods collectively as modular belief updating schemes. Unfortunately, these schemes in reality do not satisfy the property of modularity. Contin- uing our example, suppose the patient has vaginal bleeding. This fact increases the likelihood that she has a ruptured ectopic pregnancy, and thus increases the likelihood that she has peritonitis. The chances that the patient has an appendicitis decreases, however, because the presence of a ruptured ectopic pregnancy can account for the presence of peritonitis. Overall, we have that the likelihood of peritonitis increases, whereas the likelihood of appendicitis decreases. The modular rule linking peritonitis with appendicitis is inconsistent with these relationships. 6This work led to the construction of a probabilistic version of QMR, called QMR-DT [54, 55, 56, 57]. 22 In general, logical relationships represent complete models of interaction. In contrast, uncertain relationships encode invisible interactions. We summarize these hidden interac- tions with numerical measures, such as a certainty factor or likelihood ratio. In the process of such a summarization, we lose information about the detailed categorical interaction. There- fore, when we try to combine uncertain information, unexpected (nonmodular) interactions may occur. We should not expect that any modular belief updating scheme will be able to handle such subtle interactions. 7.2 A Practical Problem with Nonprobabilistic Methods In continuing to explore the difference between probabilistic and nonprobabilistic alterna- tives, we encountered a practical limitation associated with the use of modular belief updat- ing schemes. Specifically, these schemes require that we assess the strength of an uncertain relationship in the direction in which it is used. That is, we must specify the change in belief of an unobservable hypothesis, given an observable piece of evidence. Unfortunately, experts often are more comfortable quantifying uncertain relationships in the direction opposite to that in which they are used [59]. In particular, Kahneman and Tversky have shown that people usually are more comfort- able when they assess the likelihood of an effect given a cause rather than when they assess the likelihood of a cause given an effect. For example, expert physicians prefer to assess the likelihood of a finding, given a disease, rather than the likelihood (or belief update) of a disease, given a finding [60]. Henrion attributes this phenomenon to the nature of causal- ity. In particular, he notes that a predictive probability (the likelihood of a finding, given a disease) reflects a stable property of that disease. In contrast, a diagnostic probability (the likelihood of a disease, given a finding) depends on the incidence rates of that disease and of other diseases that may cause the finding. Thus, predictive probabilities are a more useful and parsimonious way to represent uncertain relationships—at least in medical domains (see [61], pages 252–3). The developers of QMR make a similar observation [39]. Unfortunately, effects are usually the observable pieces of evidence, and causes are the sought-after hypotheses. Thus, in using a modular belief updating scheme, we force experts to provide judgments of uncertainty in a direction that is more cognitively challenging. We thereby promote errors in assessment. In contrast, when we use probability theory to manage uncertainty, we can assess the strength of an uncertain relationship in one direction, and then reverse the relationship using Bayes’ theorem, when the need arises. 7.3 Compelling Principles for Uncertainty Management and De- cision Making After identifying limitations of non-decision-theoretic approaches for uncertain reasoning, we explored theoretical advantages of the decision-theoretic approach. Perhaps the most significant advantage we discovered was the fact that the rules of probability and the MEU principle follow from compelling principles, and that people often violate these principles when unaided by decision-theoretic systems. For example, Ramsey and deFinetti showed that anyone who does not follow the rules of probability theory would be willing to accept a “Dutch book”: a combination of bets 23 leading to a guaranteed loss of money under any circumstances [62, 63, 64]. In contrast, Cox developed a set of desiderata about fundamental properties of a measure of belief unre- lated to betting behavior that also imply the rules of probability [65, 25]. In addition, von Neumann and Morgenstern constructed a remarkable proof of the MEU principle [20]. They developed five compelling principles or desiderata that every decision maker should follow, and demonstrated that these desiderata imply the MEU principle. One desideratum is the principle of transitivity, which states that if a decision maker prefers outcome A to outcome B, and prefers outcome B to outcome C, then he must prefer outcome A to outcome C. To see that this desideratum is compelling, let us suppose that a decision maker’s preferences are not transitive. In particular, suppose he prefers A to B, B to C, and C to A. Because he prefers C to A, he should be willing to exchange A for C and a small payment. Similarly, this person should be willing to exchange C for B, and B for A. Thus, we can extract payments from him, and yet leave him with the same outcome. Repeating this procedure, called a money pump, we can extract an arbitrarily large payment from this person. Decision theorists and cognitive psychologists have devised justifications for each of von Neumann and Morgenstern’s desiderata [20, 66, 67]. Psychological studies have demonstrated that, in the real world, human decision makers exhibit a set of stereotypical deviations or biases from the desiderata of decision theory [68, 69]. From the perspective of decision theory, we can view these deviations as mistakes. For example, people sometimes have nontransitive preferences, and thus are vulnerable to a money pump. Also, physicians often forget to consider the prior probability of disease when making a diagnosis. As we mentioned in the introduction, we use the adjectives “normative” and “descriptive” to emphasize the differences between decision making consistent with the rules of decision theory and decision making in the real world. In medicine, descriptive errors in decision making are particularly dangerous, given the high stakes associated with some decisions. Such errors easily may lead to needless expen- ditures, pain and suffering, or loss of life. Fortunately, normative expert systems can help physicians to avoid such errors. For example, systems that encode explicitly the prior proba- bilities of diseases will likely improve the decisions made by physicians who would otherwise forget to consider this important information. Nonetheless, several researchers have argued that we should employ descriptive rather than normative methods for decision making in expert systems. In particular, investiga- tors have stated that decision-theoretic methods lack the expressiveness needed to encode expertise or to describe intelligent behavior [70, 37, 71, 43]. Indeed, several of these inves- tigators have argued that the CF model, fuzzy decision theory, and the Dempster–Shafer theory of evidence may be more appropriate than is decision theory for uncertainty man- agement, because these non-decision-theoretic methods are possibly more descriptive than is decision theory [71, 43, 72]. To our knowledge, however, there are no psychological studies that support these assertions. In fact, our experiment comparing the diagnostic accuracy of the simple-Bayes, CF, and Dempster–Shafer–Barnett models (see Section 6.3) demonstrated that, among these approaches, the simple-Bayes model is the most descriptive method for managing uncertainty. 24 8 A Return to Decision Theory: Practical Consider- ations Once we understood the limitations of non-decision-theoretic methods and the theoretical benefits of decision theory for managing uncertainty, we still faced the practical limitations of probability and decision theory. In particular, we were concerned that it would be unfeasible to relax the conditional independence assumptions of the simple-Bayes model for Pathfinder. Nonetheless, we speculated that adding the most salient conditional dependencies would not lead to a combinatorial explosion. Indeed, we conjectured that experts themselves could not appreciate all the subtle conditional dependencies that may exist among the features in large medical domains. To manage the complexity of their domain, these experts must be reasoning under many assumptions of conditional independence. We believed that, if we could capture these assumptions made by the experts, then we could produce normative expert systems that perform at least as well as do experts. Our work to achieve these goals in the domain of lymph-node pathology was successful. Over the next three years, we developed graphical knowledge representations that allowed us to capture the important conditional dependencies in this domain in a reasonable amount of time. In a formal evaluation, we demonstrated that the diagnostic accuracy of the new version of Pathfinder was at least as good as that of the Pathfinder expert [73]. In the companion to this article, we describe these new knowledge representations in detail. 9 Pathfinder in Clinical Practice Several years after the Pathfinder project was initiated, we re-engineered the program on MS-DOS computers, and made the system available commercially. The system, named Intellipath, consists of a normative expert system for lymph-node diagnosis and a set of supportive informational tools including an analog videodisc library of images, text infor- mation on diseases and microscopic features, references to the literature, and an automated report-generator. In 1988, the American Society of Clinical Pathologists began selling the Intellipath pro- gram to practicing pathologists and pathologists in training in North America. The program was a commercial success, and we constructed similar systems for other types of human tis- sue. Currently, approximately 200 pathologists are using the program, and systems for breast, bone, larynx, skin, small intestine, stomach, thymus, and urinary bladder pathology and for lung and thyroid cytology are available. In Section 2, we mentioned that pathologists have difficulty with diagnosis because they are unable to combine evidence accurately and because they misrecognize or fail to recognize microscopic features. Most of Pathfinder research addresses the first problem; the videodisc component of Intellipath, however, addresses the second problem. In particular, when a user has trouble recognizing a feature, he can ask Intellipath to display images of that feature. These images illustrate both typical and atypical presentations of the feature. Recently, we have integrated the latest version of the Pathfinder expert system with the Intellipath platform, including the videodisc for lymph-node pathology. We will evaluate this program, called Pathfinder II, in clinical trials funded by the National Cancer Institute. 25 In these studies, we shall determine whether pathologists who use the system perform better than do those pathologists who do not have access to the system. We shall quantify separately improvements in pathologists’ ability to combine evidence and improvements in their ability to recognize features. In addition, we shall determine whether pathologists can recognize situations in which they need assistance from an expert system or human expert. 10 Conclusions We are not alone in the investigation of probabilistic and decision-theoretic methods for uncertain reasoning in medical expert systems. Several other recent research projects have explored normative reasoning in medical expert systems, including the Nestor system for diagnosis of endocrinology disorders [74], the Glasgow Dyspepsia expert system for assisting in gastroenterology diagnosis [75], the Neurex system for diagnosis of neurological disorders [76], the Medas system for assisting physicians in emergency medicine [77], the Munin sys- tem for diagnosis of muscular disorders [78], and the Sleep Consultant system for diagnosis of sleep disorders [79]. Furthermore, over the last five years, a dedicated community of re- searchers addressing the management of uncertainty in reasoning systems has evolved. The proceedings of the main conference of this group of researchers, the Conference on Uncer- tainty in Artificial Intelligence, is a collection of the latest theoretical and empirical research on this topic [80, 81, 82, 83, 84]. We believe that the development of normative expert systems will lead to improvements in the capture and delivery of expert knowledge. This view is supported by our successes with Pathfinder. We hope that our experiences will inspire other medical-informatics investigators to develop normative expert systems for medicine. Acknowledgments We thank Greg Cooper, Lawrence Fagan, Thomas Lincoln, Randolph Miller, Keung-Chi Ng, Ramesh Patil, Edward Shortliffe, and Clive Taylor for advice and guidance on Pathfinder research. We also thank Costan Berard, Jerome Burke, and Ronald Dorfman for serving with Bharat Nathwani as domain experts in lymph-node pathology. Henri Suermondt, Mark Fischinger, Marty Chavez, and especially Keung-Chi Ng assisted with programming and data management. Lyn Dupré, Greg Cooper, and Keung-Chi Ng provided useful comments on this manuscript. Research on the Pathfinder project has been supported by the National Cancer Insti- tute under Grant RO1CA51729-01A, and by the National Library of Medicine under Grant RO1LM04529. Computational support has been provided in part by the SUMEX-AIM re- source under National Institutes of Health Grant LM05208. References [1] D.E. Heckerman, E.J. Horvitz, and B.N. Nathwani. Update on the Pathfinder project. In Proceedings of the Thirteenth Symposium on Computer Applications in Medical Care, 26 Washington, DC, pages 203–207. IEEE Computer Society Press, Silver Spring, MD, November 1989. [2] D.E. Heckerman, E.J. Horvitz, and B.N. Nathwani. The Pathfinder project. Technical Report KSL-90-08, Medical Computer Science Group, Section on Medical Informatics, Stanford University, Stanford, CA, February 1990. [3] D.E. Heckerman, E.J. Horvitz, and B.N. Nathwani. Pathfinder research directions. Technical Report KSL-89-64, Medical Computer Science Group, Section on Medical Informatics, Stanford University, Stanford, CA, October 1985. [4] B.N. Nathwani, D.E. Heckerman, E.J. Horvitz, and T.L. Lincoln. Integrated expert systems and videodisc in surgical pathology: An overview. Human Pathology, 21:11–27, 1990. [5] R.J. Rosai and L.V. Ackerman. The pathology of tumors: Part II. diagnostic techniques. Cancer, 29:22–39, 1979. [6] H.S. Levin. Operating room consultation by the pathologist. Urology Clinics of North America, 12:549–56, 1985. [7] G.E. Byrne. Rappaport classification of non-Hodgkin’s lymphoma: Histologic features and clinical significance. Cancer Treatment Reports, 61:935–944, 1977. [8] C.A. Coltman, R.A. Gams, J.H. Glick, and R.D. Jenkin. Lymphoma. In B. Hoogstraten, editor, Cancer Research Impact of the Cooperative Groups, pages 39–84. Masson Pub- lishing, 1980. [9] S.E. Jones, J.J. Butler, G.E. Byrne, C.A. Coltman, and T.E. Moon. Histopathologic review of lymphoma cases from the Southwest Oncology Group. Cancer, 39:1071–1076, 1977. [10] H. Kim, R.J. Zelman, M.A. Fox, J.M. Bennett, C.W. Berard, J.J. Butler, G.E. Byrne, R.F. Dorfman, R.J. Hartsock, R.J. Lukes, R.B. Mann, R.S. Neiman, J.W. Rebuck, W.W. Sheehan, D. Variakojis, J.F. Wilson, and H. Rappaport. Pathology panel for lymphoma clinical studies: A comprehensive analysis of cases accumulated since its inception. Journal of the National Cancer Institute, 68:43–67, 1982. [11] F.D. Bartlett. Thinking. Basic Books, New York, 1958. [12] A.S. Elstein, M.J. Loupe, and J.G. Erdman. An experimental study of medical diag- nostic thinking. Journal of Structural Learning, 2:45–53, 1971. [13] A.S. Elstein. Clinical judgment: Psychological research and medical practice. Science, 194:696–700, 1976. [14] A.S. Elstein, L.S. Shulman, and S.A. Sprafka. Medical Problem Solving: An Analysis of Clinical Reasoning. Harvard University Press, Cambridge, MA, 1978. [15] I. Hacking. The Emergence of Probability. Cambridge University Press, New York, 1975. 27 [16] T. Bayes. An essay towards solving a problem in the doctrine of chances. Biometrika, 46:293–298, 1958. Reprint of original work of 1763. [17] R.A. Howard. Decision analysis: Perspectives on inference, decision, and experimenta- tion. Proceedings of the IEEE, 58:632–643, 1970. [18] J. Pearl. How to do with probabilities what people say you can’t. In Proceedings of the Second Conference on Artificial Intelligence Applications, Miami Beach, FL, pages 6–12. IEEE Computer Society Press, Silver Spring, MD, December 1985. [19] D.J. Spiegelhalter. Probabilistic reasoning in predictive expert systems. In L.N. Kanal and J.F. Lemmer, editors, Uncertainty in Artificial Intelligence, pages 47–67, New York, 1986. North-Holland. [20] J. von Neumann and O. Morgenstern. Theory of Games and Economic Behavior. Prince- ton University Press, Princeton, NJ, 1947. [21] R.L. Keeney and H. Raiffa. Decisions with Multiple Objectives: Preferences and Value Tradeoffs. Wiley and Sons, New York, 1976. [22] B. J. McNeil, S. G. Pauker, H. C. Sox, and A. Tversky. On the elicitation of preferences for alternative therapies. New England Journal of Medicine, 306:1259–62, 1982. [23] R.A. Howard. On making life and death decisions. In R.C. Schwing and W.A. Albers, Jr., editors, Societal Risk Assessment, pages 89–113. Plenum Publishing, New York, 1980. [24] R.A. Howard. Microrisks for medical decision analysis. International Journal of Tech- nology Assessment in Health Care, 5:357–370, 1989. [25] M. Tribus. Rational Descriptions, Decisions, and Designs. Pergamon Press, New York, 1969. [26] R.S. Ledley and L.B. Lusted. Reasoning foundations of medical diagnosis. Science, 130:9–21, 1959. [27] G.A. Gorry and G.O. Barnett. Experience with a model of sequential diagnosis. Com- puters and Biomedical Research, 1:490–507, 1968. [28] R.A. Howard. Value of information lotteries. IEEE Transactions of Systems Science and Cybernetics, SSC-3(1):54–60, 1967. [29] R.A. Howard. Decision Analysis, EES231 Course Notes. Department of Engineering– Economic Systems, Stanford University, Stanford, CA, 1985. [30] H.R. Warner, A.F. Toronto, L.G. Veasy, and R. Stephenson. A mathematical approach to medical diagnosis: Application to congenital heart disease. Journal of the American Medical Association, 177:177–183, 1961. 28 [31] F.T. de Dombal, D.J. Leaper, J.R. Staniland, A.P. McCann, and J.C. Horrocks. Computer-aided diagnosis of acute abdominal pain. British Medical Journal, 2:9–13, 1972. [32] G.A. Gorry. Computer-assisted clinical decision making. Methods of Information in Medicine, 12:45–51, 1973. [33] E.H. Shortliffe. Computer-based Medical Consultations: MYCIN. North-Holland, New York, 1976. [34] D.G. Fryback. Bayes’ theorem and conditional nonindependence of data in medical diagnosis. Computers and Biomedical Research, 11:423–434, 1978. [35] M.J. Norusis and J.A. Jacquez. Diagnosis 1: Symptom nonindependence in mathemat- ical models for diagnosis. Computers and Biomedical Research, 8:156–172, 1975. [36] P. Szolovits. Artificial intelligence in medicine. In P. Szolovits, editor, Artificial Intelli- gence In Medicine, pages 1–19. Westview Press, Boulder, CO, 1982. [37] R. Davis. Consultation, knowledge acquisition, and instruction. In P. Szolovits, editor, Artificial Intelligence In Medicine, pages 57–78. Westview Press, Boulder, CO, 1982. [38] E.H. Shortliffe and B.G. Buchanan. A model of inexact reasoning in medicine. Mathe- matical Biosciences, 23:351–379, 1975. [39] R.A. Miller, E.P. Pople, and J.D. Myers. INTERNIST-1: An experimental computer- based diagnostic consultant for general internal medicine. New England Journal of Medicine, 307:476–486, 1982. [40] R.A. Miller, M.A. McNeil, S.M. Challinor, F.E. Masarie, and J.D. Myers. The INTERNIST-1/Quick Medical Reference project: Status report. Western Journal of Medicine, 145:816–822, 1986. [41] A.P. Dempster. Upper and lower probabilities induced by a multivalued mapping. Ann. Math. Statistics, 38:325–339, 1967. [42] G. Shafer. A Mathematical Theory of Evidence. Princeton University Press, Princeton, NJ, 1976. [43] G. Shafer. Savage revisited. Statistical Science, 1:463–501, 1986. [44] J. A. Barnett. Computational methods for a mathematical theory of evidence. In Proceedings of the Seventh International Joint Conference on Artificial Intelligence, Vancouver, BC, pages 868–875. International Joint Conference on Artificial Intelligence, August 1981. [45] J. Gordon and E.H. Shortliffe. A method for managing evidential reasoning in a hier- archical hypothesis space. Artificial Intelligence, 26:323–357, 1985. 29 [46] G. Shafer. Probability judgment in artificial intelligence. In Proceedings of the Workshop on Uncertainty and Probability in Artificial Intelligence, Los Angeles, CA, pages 91–98. Association for Uncertainty in Artificial Intelligence, Mountain View, CA, August 1985. Also in Kanal, L. and Lemmer, J., editors, Uncertainty in Artificial Intelligence, pages 127–136. North-Holland, New York, 1986. [47] R. Hummel and M. Landy. Evidence as opinions of experts. In Proceedings of the Second Workshop on Uncertainty in Artificial Intelligence, Philadelphia, PA, pages 135–143. Association for Uncertainty in Artificial Intelligence, Mountain View, CA, August 1986. Also in Kanal, L. and Lemmer, J., editors, Uncertainty in Artificial Intelligence 2, pages 43–53. North-Holland, New York, 1988. [48] P. Smets. Constructing the pignistic probability function in a context of uncertainty. In Proceedings of the Fifth Workshop on Uncertainty in Artificial Intelligence, Windsor, ON, pages 319–326. Association for Uncertainty in Artificial Intelligence, Mountain View, CA, August 1989. Also in Henrion, M., Shachter, R., Kanal, L., and Lemmer, J., editors, Uncertainty in Artificial Intelligence 5, pages 29–39. North-Holland, New York, 1990. [49] L. Zadeh. Is probability theory sufficient for dealing with uncertainty in AI: A negative view. In L.N Kanal. and J.F. Lemmer, editors, Uncertainty in Artificial Intelligence, pages 103–116. North-Holland, New York, 1986. [50] D.E. Heckerman. An empirical comparison of three inference methods. In Proceedings of the Fourth Workshop on Uncertainty in Artificial Intelligence, Minneapolis, MN, pages 158–169. Association for Uncertainty in Artificial Intelligence, Mountain View, CA, August 1988. Also in Shachter, R., Levitt, T., Kanal, L., and Lemmer, J., editors, Uncertainty in Artificial Intelligence 4, pages 283–302. North-Holland, New York, 1990. [51] D.E. Heckerman. Probabilistic interpretations for MYCIN’s certainty factors. In Pro- ceedings of the Workshop on Uncertainty and Probability in Artificial Intelligence, Los Angeles, CA, pages 9–20. Association for Uncertainty in Artificial Intelligence, Moun- tain View, CA, August 1985. Also in Kanal, L. and Lemmer, J., editors, Uncertainty in Artificial Intelligence, pages 167–196. North-Holland, New York, 1986. [52] B. Grosof. Evidential confirmation as transformed probability. In Proceedings of the Workshop on Uncertainty and Probability in Artificial Intelligence, Los Angeles, CA, pages 185–192. Association for Uncertainty in Artificial Intelligence, Mountain View, CA, August 1985. Also in Kanal, L. and Lemmer, J., editors, Uncertainty in Artificial Intelligence, pages 153–166. North-Holland, New York, 1986. [53] D.E. Heckerman and R.A. Miller. Towards a better understanding of the INTERNIST- 1 knowledge base. In Proceedings of Medinfo, Washington, DC, pages 27–31. North- Holland, New York, October 1986. [54] M. Henrion. Towards efficient probabilistic inference in multiply connected belief net- works. In R.M. Oliver and J.Q. Smith, editors, Influence Diagrams, Belief Nets, and Decision Analysis, chapter 17. Wiley and Sons, New York, 1990. 30 [55] D.E. Heckerman. A tractable algorithm for diagnosing multiple diseases. In Proceedings of the Fifth Workshop on Uncertainty in Artificial Intelligence, Windsor, ON, pages 174–181. Association for Uncertainty in Artificial Intelligence, Mountain View, CA, August 1989. Also in Henrion, M., Shachter, R., Kanal, L., and Lemmer, J., editors, Uncertainty in Artificial Intelligence 5, pages 163–171. North-Holland, New York, 1990. [56] M. Shwe, B. Middleton, D. Heckerman, M. Henrion, E. Horvitz, H. Lehmann, and G. Cooper. Probabilistic diagnosis using a reformulation of the INTERNIST-1/QMR knowledge base: Part I. the probabilistic model and inference algorithms. Methods in Information and Medicine, 30:241–250, 1991. [57] B. Middleton, M. Shwe, D. Heckerman, M. Henrion, E. Horvitz, H. Lehmann, and G. Cooper. Probabilistic diagnosis using a reformulation of the INTERNIST-1/QMR knowledge base: Part II. evaluation of diagnostic performance. Methods in Information and Medicine, 30:256–267, 1991. [58] R. Johnson. Independence and Bayesian updating methods. In Proceedings of the Workshop on Uncertainty and Probability in Artificial Intelligence, Los Angeles, CA, pages 28–30. Association for Uncertainty in Artificial Intelligence, Mountain View, CA, August 1985. Also in Kanal, L. and Lemmer J., editors, Uncertainty in Artificial Intel- ligence, pages 197–201. North-Holland, New York, 1986. [59] R.D. Shachter and D.E. Heckerman. Thinking backward for knowledge acquisition. AI Magazine, 8:55–63, 1987. [60] A. Tversky and D. Kahneman. Causal schemata in judgments under uncertainty. In D. Kahneman, P. Slovic, and A. Tversky, editors, Judgement Under Uncertainty: Heuristics and Biases. Cambridge University Press, New York, 1982. [61] E.J. Horvitz, J.S. Breese, and M. Henrion. Decision theory in expert systems and artificial intelligence. International Journal of Approximate Reasoning, 2:247–302, 1988. [62] F.P. Ramsey. Truth and probability. In R.B. Braithwaite, editor, The Foundations of Mathematics and other Logical Essays. Humanities Press, London, 1931. Reprinted in Kyburg and Smokler, 1964. [63] B. de Finetti. La prévision: See lois logiques, ses sources subjectives. Annales de l’Institut Henri Poincaré, 7:1–68, 1937. Translated in Kyburg and Smokler, 1964. [64] H.E. Kyburg and H.E. Smokler. Studies in Subjective Probability. Wiley and Sons, New York, 1964. [65] R. Cox. Probability, frequency and reasonable expectation. American Journal of Physics, 14:1–13, 1946. [66] L.J. Savage. The Foundations of Statistics. Dover, New York, 1954. [67] R.A. Howard. Risk preference. In R.A. Howard and J.E. Matheson, editors, Readings on the Principles and Applications of Decision Analysis, volume II, pages 629–663. Strategic Decisions Group, Menlo Park, CA, 1970. 31 [68] A. Tversky and D. Kahneman. Judgment under uncertainty: Heuristics and biases. Science, 185:1124–1131, 1974. [69] D. Kahneman, P. Slovic, and A. Tversky, editors. Judgment Under Uncertainty: Heuris- tics and Biases. Cambridge University Press, New York, 1982. [70] G.A. Gorry, J.P. Kassirer, A. Essig, and W.B. Schwartz. Decision analysis as the basis for computer-aided management of acute renal failure. American Journal of Medicine, 55:473–484, 1973. [71] B.G. Buchanan and E.H. Shortliffe, editors. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Addison–Wesley, Reading, MA, 1984. [72] L. Zadeh. Personal communication. 1986. [73] D. Heckerman and B. Nathwani. An evaluation of the diagnostic accuracy of Pathfinder. Computers and Biomedical Research, In press. [74] G.F. Cooper. NESTOR: A Computer-based Medical Diagnostic Aid that Integrates Causal and Probabilistic Knowledge. PhD thesis, Medical Computer Science Group, Stanford University, Stanford, CA, November 1984. Report HPP-84-48. [75] D.J. Spiegelhalter and R.P. Knill-Jones. Statistical and knowledge-based approaches to clinical decision support systems, with an application in gastroenterology. Journal of the Royal Statistical Society, 147:35–77, 1984. [76] J.A. Reggia and B.T. Perricone. Answer justification in medical decision support sys- tems based on Bayesian classification. Computers in Biology and Medicine, 15:161–167, 1985. [77] M. Ben-Bassat, V.K. Carlson, V.K. Puri, M.D. Davenport, J.A. Schriver, M.M. Latif, R. Smith, E.H. Lipnick, and M.H. Weil. Pattern-based interactive diagnosis of multiple disorders: The MEDAS system. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2:148–160, 1980. [78] S. Andreassen, M. Woldbye, B. Falck, and S.K. Andersen. MUNIN: A causal probabilis- tic network for interpretation of electromyographic findings. In Proceedings of the Tenth International Joint Conference on Artificial Intelligence, Milan, Italy, pages 366–372. Morgan Kaufmann, San Mateo, CA, August 1987. [79] G. Nino-Murcia and M. Shwe. An expert system for diagnosis of sleep disorders. In M. Chase, R. Lydic, and C. O’Connor, editors, Sleep Research, volume 20, page 433. Brain Information Service, Los Angeles, CA, 1991. [80] L. Kanal and J. Lemmer, editors. Uncertainty in Artificial Intelligence. North-Holland, New York, 1986. [81] J. Lemmer and L. Kanal, editors. Uncertainty in Artificial Intelligence 2. North-Holland, New York, 1988. 32 [82] L. Kanal, T. Levitt, and J. Lemmer, editors. Uncertainty in Artificial Intelligence 3. North-Holland, New York, 1989. [83] R. Shachter, T. Levitt, L. Kanal, and J. Lemmer, editors. Uncertainty in Artificial Intelligence 4. North-Holland, New York, 1990. [84] M. Henrion, R. Shachter, L. Kanal, and J. Lemmer, editors. Uncertainty in Artificial Intelligence 5. North-Holland, New York, 1990. 33 
work_2f5sazhjizewblsxvxrtlowiyu	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 221003657 Params is empty 221003657 exception Params is empty 2021/04/06-03:05:50 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221003657 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:50 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_2jjz24x67jeijo4fjg7ba55s2q	----	A Web-enabled hybrid approach to strategic marketing planning: Group Delphi+a Web-based expert system A Web-enabled hybrid approach to strategic marketing planning: Group DelphiCa Web-based expert system Shuliang Li* Westminster Business School, University of Westminster, 35 Marylebone Road, London NW1 5LS, UK Abstract A Web-enabled hybrid approach to strategic marketing planning is established in this paper. The proposed approach combines the group Delphi technique with a Web-based expert system, called WebStra (developed by the author), to support some key stages of the strategic marketing planning process. The Web-enabled approach is based upon client–server architecture that enables the sharing and delivery of computerised planning models and knowledge via the Internet, intranets or extranets, which allows widespread access by authorised users around the clock, across the world or throughout the company. In order to assess the overall value of the proposed approach, case-based evaluation work has been undertaken. Evaluation findings indicate that the approach is effective and efficient in terms of overcoming time and geographical barriers, saving decision-making time, coupling analysis with human judgment, helping improve decision-making quality, etc. q 2005 Elsevier Ltd. All rights reserved. Keywords: Strategic planning; Marketing strategy; Decision support system; Web-based expert system; Group Delphi; World Wide Web 1. Introduction The globalisation, the complexity and the dynamics of the business environments present real challenges to strategic marketing planners in the 21st century. The needs for appropriate techniques and technologies in support of strategic marketing planning have never been so great. Over the past years, attempts have been made by researchers to develop computer-based systems to support the process of strategic marketing planning. Some related typical work in this field may be found in Belardo, Duchessi, and Coleman (1994), Carlsson, Walden, and Kokkonen (1996), Cavusgil and Evirgen (1997), Levy and Yoon (1995) and McDonald and Wilson (1990), Mitri et al. (1999), Li (2000), and Li and Davies (2001), etc. The relevant systems developed in the past, however, are mainly restricted to assist with individual users using standalone PC-based programs, which may limit users’ access to computerised models and support systems. Moreover, program distribution is a serious problem for many types of expert systems, because most knowledge bases must be updated regularly (Eriksson, 1996). In addition, the critical importance of capturing and combining managerial judg- ment, especially groups of decision-makers’ judgment and intuition, with computer-based support has not been high- lighted adequately in previous research in this field. The World Wide Web is emerging as an increasingly important platform that can reduce the technological barriers and make it easier for users in different geographi- cal locations to access the decision support models and tools (Shim et al., 2002). The widespread use of World Wide Web and the Internet provide an opportunity for enabling computer-based decision support widely available. It is also argued that the Internet can complement traditional ways of competing (Porter, 2001). Appropriate Internet applications can improve communications with actual and potential customers, suppliers and collaborators; and can be a powerful promotion and sales tool (Hamill, 1997). Internet technologies and e-commerce applications can be used strategically for creating competitive advantage in the global markets. Devising Internet marketing strategies and associated e-commerce strategies as an organic part of business strategies is something that strategic marketing plans today should not neglect. The purpose of this study is to establish a Web-enabled approach that combines the advantages of a Web-based Expert Systems with Applications 29 (2005) 393–400 www.elsevier.com/locate/eswa 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.04.018 * Tel.: C44 20 7911 5000x3429; fax: C44 20 7911 5839. E-mail address: lish@wmin.ac.uk http://www.elsevier.com/locate/eswa expert system with the benefits of the group Delphi technique and links strategic marketing planning with Internet and e-commerce strategy formulation. The paper is structured as follows. The opening section outlines the logical process, the main functional components and judgmental ingredients of the Web-enabled approach. This is followed by discussions on the evaluation of the overall value of the approach. The final section offers some general conclusions and discussions. 2. A Web-enabled hybrid approach to strategic marketing planning Strategic marketing planning presents challenges to the experience, knowledge, judgment and intuition of individual managers. A group of managers from different functional departments can bring a variety of perspec- tives, knowledge, judgement and intuition to the planning process (Bass, 1983; Minkes, 1987). Eden (1990) points out that those who have the power to act must be integrally involved in the process of strategy develop- ment. Porter (1987) argues that strategic planning should employ multifunctional planning teams. Beveridge, Gear, and Minkes (1997) note that strategic decisions usually require consensus and commitment to a course of actions among groups of individuals because of differing perspectives, interests, and functional biases. This type of group decision-making can involve the resolution of conflict. Turban and Aronson (2001) note that decision- making in groups has the following benefits: groups are better than individuals at catching errors; groups are better than individuals at understanding problems; a group has more information or knowledge; and synergy may be produced. The group Delphi technique (Webler, Levine, Rakel, & Renn, 1991) provides a structured communication process in which a group of participants input, discuss and defend their judgment and intuition concerning existing knowl- edge and information. Firstly, views are discussed openly in the group Delphi. There is direct and immediate feedback. Any ambiguities are clarified immediately. Secondly, discussions or debate provides an internal check for consistency in accepted points of view (Webler et al., 1991). The group Delphi technique can be applied to resolve conflicting views and build consensus. It can also be used as a technique for reducing uncertainty surrounding the managerial inputs to the decision-making process. It is evident that the appropriate use of computer-based support systems can improve the process of marketing planning (Li, 2000; Wilson & McDonald, 1994). Keen and Scott Morton (1978) point out that computer-based decision support implies the use of computers to: assist decision- makers in their decision processes; support, rather than replace, managerial judgment. Computerised models are consistent and unbiased, but rigid, while managers are inconsistent and often lack analytical skills but are flexible in adapting to changing environments (Blattberg & Hoch, 1990; Li, Kinman, Duan, & Edwards, 2000) and have knowledge about their products and markets (Mintzberg, 1994a–c). It is argued that “the capacity of the human mind for formulating and solving complex problems is very small compared with the size of the problems whose solution is required for objective rational behaviour in the real world— or even for a reasonable approximation to such objective rationality” (Simon, 1957). The key ingredient is, therefore, the integration of the decision-makers into the model (McIntyre, 1982). It is also found that marketing planning is typically the shared collective responsibility of managers in many large companies (Li et al., 2000). Thus, combining computerised models with a group of decision-makers’ judgment and intuition would lead to better strategic decisions The World Wide Web provides a useful platform for developing, sharing and delivering decision support tools. The primary Web tools are Web servers using Hypertext Transfer Protocol (HTTP) containing Web pages created with Hypertext Markup Language (HTML), JavaScript, etc. accessed by client PCs running client software known as a Web browser (Shim et al., 2002). The driving forces for developing a Web-enabled hybrid approach towards strategic planning are: to publish marketing strategy expertise and guidelines on the Web; to transport the decision support models and tools over the Internet, extranets or intranets; to provide intelligent support round the clock, around the world or throughout the company; to link the formulation of marketing strategies with the development of e-com- merce strategies; and to combine Web-based intelligent support with group judgment and intuition. In particular, the Web-enabled approach aims at supporting the following three key stages of the strategic marketing planning process: † Assessing marketing environments and performing strategic analysis; † Producing strategic portfolio summary and generating strategic recommendations/options; † Selecting marketing strategies and related Internet/e- commerce strategies. The proposed approach is illustrated in Fig. 1. The Web-enabled approach is based on client–server architecture that allows the sharing and delivery of computerised intelligent decision support models via the Internet, intranets or extranets, which enables widespread access by any authorised user. The client can be an Internet-connected computer with a suitable Web browser such as Microsoft Internet Explorer. Users interact with the system using the normal Web access techniques of S. Li / Expert Systems with Applications 29 (2005) 393–400394 https://isiarticles.com/article/967 
work_23cetdgvhrdpvdevpor2azsreu	----	Ontology matching: A literature review Expert Systems with Applications 42 (2015) 949–971 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Ontology matching: A literature review http://dx.doi.org/10.1016/j.eswa.2014.08.032 0957-4174/� 2014 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. E-mail addresses: locerdeira@uvigo.es (L. Otero-Cerdeira), franjrm@uvigo.es (F.J. Rodríguez-Martínez), alma@uvigo.es (A. Gómez-Rodríguez). Lorena Otero-Cerdeira ⇑, Francisco J. Rodríguez-Martínez, Alma Gómez-Rodríguez LIA2 Research Group, Computer Science Department, University of Vigo, Spain a r t i c l e i n f o a b s t r a c t Article history: Available online 30 August 2014 Keywords: Ontology matching Literature review Classification framework User survey The amount of research papers published nowadays related to ontology matching is remarkable and we believe that reflects the growing interest of the research community. However, for new practitioners that approach the field, this amount of information might seem overwhelming. Therefore, the purpose of this work is to help in guiding new practitioners get a general idea on the state of the field and to determine possible research lines. To do so, we first perform a literature review of the field in the last decade by means of an online search. The articles retrieved are sorted using a classification framework that we propose, and the differ- ent categories are revised and analyzed. The information in this review is extended and supported by the results obtained by a survey that we have designed and conducted among the practitioners. � 2014 Elsevier Ltd. All rights reserved. 1. Introduction Ontology matching is a complex process that helps in reducing the semantic gap between different overlapping representations of the same domain. The existence of such different representations obeys to the natural human instinct to have different perspectives and hence to model problems differently. When these domains are represented using ontologies, the solution typically involves the use of ontology matching techniques. Ontologies and ontology matching techniques are an increasing trend as ontologies provide probably the most interesting opportu- nity to encode meaning of information. The last decades have born witness to a period of extensive research in this field. Nowadays, far from dying down, the activity seems to be increasing and new publications, where the ontology matching field is addressed, are continuously being released. This reflects the global interest in ontology matching which we have studied by means of an analytical review of the literature so far. Other works and publications have successfully presented the state-of-the-art in the field such as, Euzenat (2004), Shvaiko and Euzenat (2013) and Kalfoglou and Schorlemmer (2003b), although our purpose is quite different. We aim at retrieving articles related to ontology matching that have been published in the last decade, to classify and identify research lines relevant for ontology match- ing. We also aim at providing a reference framework for the inte- gration and classification of such articles. Therefore practitioners approaching the field for the first time would be aware of the dif- ferent types of publications regarding the field to better choose those that better fits their needs, they would gain knowledge about the main issues where the researchers have been working and what are the main trends and challenges still to be addressed in the next years. To this end, the remainder of the paper is organized as follows. In Section 2 a methodology to extract the articles is presented. Sec- tion 3 presents general statistical results of the retrieved publica- tions. Then, Section 4 illustrates the classification framework proposed and describes each one of the categories defined. In Sec- tion 5 we describe the limitations of the literature review and sug- gest a practitioner-oriented survey to support the results of the review. Such review is detailed in Section 6, and its limitations in Section 7. Finally, in Section 8 we present our discussion, conclud- ing remarks and directions for future work. 2. Procedures To retrieve the articles for this literature review, several well- known online databases were queried to obtain articles related to the ontology matching field. As result over 1600 articles were obtained which were filtered to narrow down the selection to the final 694 articles that are included in this review. This screen- ing allowed dismissing 58.09% of the initially retrieved articles. Although this is a high percentage of the articles it is worth notic- ing that the original search was broadly defined so an important number of false positives among the initial results were already expected. The screening of the articles was manually done, and http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.08.032&domain=pdf http://dx.doi.org/10.1016/j.eswa.2014.08.032 mailto:locerdeira@uvigo.es mailto:franjrm@uvigo.es mailto:alma@uvigo.es http://dx.doi.org/10.1016/j.eswa.2014.08.032 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 950 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 took over 3 months to complete it due to the several iterations and the amount of articles reviewed. This review covers journal articles and conference proceedings published within the last decade. Other publication forms such as books, newspapers, doctoral dissertations, posters, etc., were not considered as researchers use mainly journals and conference papers to obtain and spread knowledge, thus these types of publi- cations encompass the majority of the research papers published about this subject. The procedure followed to identify and filter the papers is reflected in Fig. 1. First, different online databases were queried using a combination of the following search strings: ontology matching, ontology mapping and ontology alignment. These dat- abases were: IEEEXplore Digital Library (IEEXplore, 2013), Science Direct (Direct, 2013), ACM Digital Library (ACM, 2013) and Scopus (Scopus, 2013). In addition to these databases, the articles published in the Ontology Alignment Evaluation Initiative (OAEI) (OAEI, 2013) were also included as this initiative is considered as the most prominent one regarding the evaluation of different matching systems and it helps practitioners improve their works on matching techniques. Sorting these data sources by amount of articles we have: Sco- pus (626), ACM Digital Library (383), Science Direct (267), OAEI (254) and IEEEXplore Digital Library (126), making this way a total of 1656 articles initially obtained. The next step was to dismiss those articles that were dupli- cated, i.e, that have been obtained through two or more data sources. When this situation arose, the criterion followed was to dismiss the articles belonging to the data source with a higher number of articles. By doing so, 335 articles were removed. Next, the 1321 remaining articles were analyzed considering their keywords and abstracts. Those whose keywords did not include specific mentions to the ontology matching field or whose abstracts did not introduce a research regarding this field were excluded. Of all the criteria considered, this was the one that pro- duced the sharpest cut down in the amount of articles. Addition- ally, while reviewing the keywords and abstracts, also the papers that corresponded to a poster publication were dismissed. Finally, the 795 articles remaining were carefully reviewed to dismiss those that did not consider ontology matching as their core Fig. 1. Procedure followed to retrieve th part. By applying this criterion another 101 articles were excluded, therefore leaving the 694 articles that are included in this litera- ture review. 3. Statistical results In this section some statistical information about the articles is presented and discussed. Articles were analyzed regarding publica- tion year and database from which they were obtained. 3.1. Articles sorted by publication year Fig. 2 represents the progression of the number of articles with respect to their publication years. The measurements and values that shape this progression are shown in Table 1. In this figure we can observe that the amount of published articles steadily increases from 2003 to 2012, where it peaks. The sharpest rise was found between 2005 and 2006 where the percentage of pub- lished articles rose from 3.75% to 8.36%. Between 2012 and 2013 we observe a pronounced decrease in the amount of published articles, although as this review covers only the first semester of 2013, it is highly likely that the amount of published articles by the end of the year would follow the increasing trend of the previ- ous years. This increasing pattern reflects the global interest of the research community in the ontology matching field. 3.2. Articles sorted by data source To retrieve the articles for this literature review, a total of five different data sources were used. Among these, four are well- known online databases that were queried to obtain the articles. The other source was the Ontology Alignment Evaluation Initiative site, where all the publications related to this initiative are published. The classification of the retrieved articles by data source is shown in Table 2 and graphically depicted in Fig. 3. These data state that Scopus provides the highest amount of articles to the total (41.79%, 290 articles) probably because it includes a wider variety of source journals. On the other hand, ScienceDirect pro- e articles for the literature review. Fig. 2. Articles sorted by publication year. Table 1 Articles with respect to publication year. Publication year Number of articles Percentage over total (%) 2003 4 0.58 2004 17 2.45 2005 26 3.75 2006 58 8.36 2007 75 10.81 2008 84 12.10 2009 83 11.96 2010 88 12.68 2011 100 14.41 2012 107 15.42 2013 52 7.49 Total 694 Table 2 Articles with respect to data source. Data source Number of articles Percentage over total (%) ACM Digital Library 72 10.37 IEEExplore Digital Library 114 16.43 Ontology Alignment Evaluation Initiative 174 25.07 ScienceDirect 44 6.34 Scopus 290 41.79 Total 694 Fig. 3. Articles sorted by data source. Fig. 4. Classification Framework. L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 951 vided significantly less articles than any of the other data sources analyzed (6.34%, 44 articles). The remaining data sources consid- ered respectively provide the 10.37% (72 articles, ACM Digital Library), 16.43% (114, IEEExplore Digital Library) and 25.07% (174 articles, OAEI). During the first filtering step (see Fig. 1) after initially retrieving the articles from the different data sources, they were processed in order to remove those duplicate results. In this situation the article was always removed from the data source that had more articles as the impact in the overall results for each database would be less significative than in the case of dismissing those from the data sources with less articles. 4. Classification Relying on the analysis of the articles selected for the literature review, we have defined an abstract framework that helps classify- ing them. This framework, depicted in Fig. 4, shows six different types of articles and it is in line with the general outline of the main issues in ontology matching proposed by Euzenat and Shvaiko in Euzenat and Shvaiko (2007, 2013). The different catego- ries identified cover the most prominent fields of interest in ontol- ogy matching. � Reviews. This category includes the publications devoted to sur- veying and reviewing the field of ontology matching. It also includes those articles focused on detailing the state-of-the- art as well as the future challenges in this field. � Matching techniques. This category covers the publications focused on different similarity measures, matching strategies and methodologies, that can be used in the matching systems. � Matching systems. This category includes those articles intro- ducing new matching systems and algorithms, and also those detailing enhancements to existing ones. � Processing frameworks. This category comprehends the articles that delve into the different uses of the alignments, i.e, those operations that can be performed from alignments, such as ontology merging, reasoning or mediation. � Practical applications. This category covers those articles that describe matching solutions applied to a real-life problem. � Evaluation. This category covers the articles describing different available approaches to evaluate the matching systems, as well as the different existing benchmarks and the most relevant per- formance measures. The results of classifying the articles within the framework defined are summarized in Table 3. According to our results the Table 3 General results of the classification. Category Number of articles Percentage (%) Reviews 46 6.63 Matching techniques 85 12.25 Matching systems 302 43.52 Processing frameworks 147 21.18 Practical applications 76 10.95 Evaluation 38 5.48 Total 694 952 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 greater efforts have been focused on developing matching systems and frameworks that exploit the alignments, while the evaluation of such systems as well as the review of the field and the develop- ment of applied solutions have not been object of such dedication. In Table 4 the results of the classification are shown sorted by year, identifying the amount of articles that match into each cate- gory and its yearly percentage. These results are graphically com- pared in Fig. 5 where the evolution can be more clearly distinguished. The amount of articles into Reviews showed an slight but upward trend within the first five years of the time frame consid- ered, peaking in 2008. Ever since, the trend in the amount of reviews per year has remained almost constant. The evolution of the Matching Techniques did not show any significant change until 2007 when the amount of articles rose from three to thirteen. After this sharp increase, the values remained constant for the next year and then they fell to six in 2009. In 2010 there was another signif- icant increase, followed by yet another fall in the amount of arti- cles. Apparently the periods of intense work in this category are followed by others where the amount of publications is signifi- cantly lower. Regarding the Matching Systems, the amount of published arti- cles shows an upward trend in the considered period, since, as sta- ted before, the values for 2013 can not be considered definitive as the articles for this review were retrieved in the first semester of the year. Regardless of the general upward trend, the amount of articles in this category reached some local peaks in 2006, 2009 and 2012, followed by periods of less activity. Anyway, it is impor- tant to notice that the periods of lower activity in this category cor- respond with periods of higher activity in matching techniques and vice versa. It is highly likely that the same researchers that define new matching systems are the ones that had suggested new matching techniques to be used within them, which also validates the process of construction of a matching system. The evolution in the number of Processing Frameworks was increasingly constant until 2007 when it reached a local peak, just to show a slight downward trend for two years, before starting a continuos rise that took the amount of published articles regarding this category to its top value in 2012. Regarding the Practical applications, the amount of articles shows in general an upward trend. It reaches a local peak in 2007 with 8 articles. In 2008 it shows a slight decrease. However, since 2010, the number of articles continues to rise every year. It is worth noticing that even when the data for 2013 does not cover the whole year, the amount of articles devoted to practical applica- tions was already a 66.6% of the articles published in 2012 in the same category. Finally, the publications related to Evaluation show an evolution pattern really similar to the one detected for the reviews, not show- ing any significant behavior. After providing some general outline of the evolution of the dif- ferent categories over the years, in the following sections a deeper analysis of each one of them is included, considering inner classifi- cations for each one of these general categories. 4.1. ‘Reviews’ category Within this category the publications have been further sorted according to their scope, namely, the publications were identified as being of either general purpose or specific purpose, as depicted in Fig. 6. More than half of the 46 articles included in this category, 26, were considered general purpose reviews as they offer some insight into the ontology matching field without specifically emphasizing on any subject. Among these, there are for instance, surveys (Falconer, Noy, & Storey, 2007; Shvaiko & Euzenat, 2005; Thayasivam, Chaudhari, & Doshi, 2012; Zhu, 2012), state-of-the- art articles (Droge, 2010; Gal & Shvaiko, 2008; Ngo, Bellahsene, & Todorov, 2013) and publications unveiling the future challenges of the field (Kotis & Lanzenberger, 2008; Shvaiko & Euzenat, 2013). In turn, the remaining 20 articles show a more limited scope within the field. When analyzing these articles we have stated that they were devoted either to delving into a very specific area within the ontology matching field or to studying the feasibility of apply- ing ontology matching to a certain domain. Some of the specific fields within ontology matching that were addressed cover topics such as (i) matching across different lan- guages (Fu, Brennan, & O’Sullivan, 2009), (ii) instance-based match- ing (Castano, Ferrara, Lorusso, & Montanelli, 2008), which is one among the different types of techniques to perform ontology matching. Other articles developed the subject of (iii) external sources for ontology matching (Fugazza & Vaccari, 2011; Lin & Sandkuhl, 2008a), these techniques take advantage of auxiliary or external resources in order to find matchings to terms based on lin- guistic relations between them such as synonymy or hyponymy. These external resources are usually lexicons or thesauri. On the other hand, among the domains where ontology match- ing could be used, we have identified articles on domains such as geography (Tomaszewski & Holden, 2012), medicine (Wennerberg, 2009) or agriculture (Lauser et al., 2008). 4.2. ‘Matching Techniques’ category Ontology matching techniques propose different approaches for the matching that are implemented in ontology matching algorithms. When building an ontology matching system, differ- ent algorithms are usually used, exploiting therefore different ontology matching techniques. In this category two different types of articles have been identified. Some articles are devoted to describing new or enhanced similarity measures and to analyz- ing the building blocks of the ontology matching algorithms, while others make use of such artifacts to define matching strategies or methodologies. In total there are 85 articles in this category, where 57:65% (49 articles) belong to the first group of basic matching techniques and 42:35% (36 articles) belong to complex matching techniques. The different matching techniques have been subject of study in the latest years. For the purpose of this review, to sort the (i) basic matching techniques we have followed the classification proposed by Euzenat and Shvaiko (Euzenat & Shvaiko, 2013), depicted in Fig. 7, since to the best our knowledge is the most complete one and reflects most of the other previous classifications. This classifi- cation is an evolution of another previously proposed by the same authors in Euzenat and Shvaiko (2007). This classification can be followed top-down and therefore focusing on the interpretation that the different techniques offer to the input information, but also bottom-up, focusing on the type of the input that the matching techniques use. Despite the fol- lowed approach both meet at the concrete techniques tier. Following the top-down interpretation, the matching tech- niques can be classified in a first level as: Table 4 Results of the classification. Year Category Number of articles Total Annual percentage 2003 Reviews 1 4 25.00% Matching techniques – – Matching systems 1 25.00% Processing frameworks 2 50.00% Practical applications – – Evaluation 0 – 2004 Reviews 1 17 5.88% Matching techniques 5 29.41% Matching systems 6 35.29% Processing frameworks 3 17.65% Practical applications – – Evaluation 2 11.76% 2005 Reviews 4 26 15.38% Matching techniques 3 11.54% Matching systems 6 23.08% Processing frameworks 9 34.62% Practical applications 1 3.85% Evaluation 3 11.54% 2006 Reviews 2 58 3.45% Matching techniques 3 5.17% Matching systems 30 51.72% Processing frameworks 18 31.03% Practical applications 4 6.90% Evaluation 1 1.72% 2007 Reviews 4 75 5.33% Matching techniques 13 17.33% Matching systems 27 36.00% Processing frameworks 19 25.33% Practical applications 8 10.67% Evaluation 4 5.33% 2008 Reviews 8 84 9.52% Matching techniques 13 15.48% Matching systems 34 40.48% Processing frameworks 16 19.05% Practical applications 5 5.95% Evaluation 8 9.52% 2009 Reviews 5 83 6.02% Matching techniques 6 7.23% Matching systems 46 55.42% Processing frameworks 11 13.25% Practical applications 10 12.05% Evaluation 5 6.02% 2010 Reviews 3 88 3.41% Matching techniques 18 20.45% Matching systems 37 42.05% Processing frameworks 16 18.18% Practical applications 10 11.36% Evaluation 4 4.55% 2011 Reviews 4 100 4.00% Matching techniques 9 9.00% Matching systems 47 47.00% Processing frameworks 21 21.00% Practical applications 13 13.00% Evaluation 6 6.00% 2012 Reviews 6 107 5.61% Matching techniques 9 8.41% Matching systems 52 48.60% Processing frameworks 23 21.50% Practical applications 15 14.02% Evaluation 2 1.87% 2013 Reviews 8 52 15.38% Matching techniques 6 11.54% Matching systems 16 30.77% Processing frameworks 9 17.31% Practical applications 10 19.23% Evaluation 3 5.77% L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 953 Fig. 5. Evolution of the different categories. Fig. 6. Specific types of articles within ‘Reviews’ category. 954 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 – Element-level matchers: these techniques obtain the corre- spondences by considering the entities in the ontologies in isolation, therefore ignoring that they are part of the struc- ture of the ontology. – Structure-level matchers: these techniques obtain the corre- spondences by analyzing how the entities fit in the structure of the ontology. In the second level of the classification, the techniques can be further classified as: – Syntactic: these techniques limit their input interpretation to the instructions stated in their corresponding algorithms. – Semantic: these techniques use some formal semantics to interpret their input and justify their results. If reading the classification bottom-up, the elementary match- ing techniques can be initially divided in two categories deter- mined by the origin of the information considered for the matching process: – Content-based: these techniques focus on the internal infor- mation coming from the ontologies to be matched. – Context-based: these techniques consider for the matching, external information that may come from relations between ontologies or other external resources (context). In the second level of the classification, both categories are fur- ther refined. The content-based category is further divided into four new groups, depending on the input that the techniques use: – Terminological: these methods consider their inputs as strings. – Structural: these methods are based on the structure of the entities (classes, individuals, relations) found in the ontology. – Extensional: these methods compute the correspondences by analyzing the set of instances of the classes (extension). – Semantic: these techniques need some semantic interpreta- tion of the input and usually use a reasoner to deduce the correspondences. In the second level of the classification for the context-based cat- egory, the techniques can be also further classified as syntactic or semantic techniques. The next level in any of both classifications already corresponds to the specific techniques. Following the different paths in this classification tree, several techniques may be reached. These cate- gories were used to further sort the 49 articles belonging to basic matching techniques. The classification of these articles was partic- ularly hard since most of them were not devoted to a single tech- nique but to several, so in the following we provide an example of articles that match into some categories but it is worth noticing that many of them may also be included in another one. � Formal Resource-based: these techniques use formal resources to support the matching process, such as upper level ontologies, domain-specific ontologies or the recorded alignments of previ- Fig. 7. Matching techniques classification. Extracted from the book ‘Ontology Matching’ (Euzenat & Shvaiko, 2013). L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 955 ously matched ontologies (alignment reuse). Examples of such techniques are Scharffe, Zamazal, and Fensel (2013) and Mascardi, Locoro, and Rosso (2010). � Informal Resource-based: these techniques, as those in the previ- ous category, also exploit an external resource, but in this case the external resources are informal ones. This group of tech- niques deduce relations between ontologies using the relation between the ontologies and such informal resources. An exam- ple of such category could not be found among the results of the review. � String-based: these techniques are based on the similarity of the strings that represent the names and descriptions of the entities in the ontologies. There are several string distance metrics that can be used in these methods Levenshtein, Jaccard, Jaro-Winkler, Euclidean, TFIDF, etc., (Cohen, Ravikumar, & Fienberg, 2003). Such techniques are present, for instance in the work of Akbari, Fathian, and Badie (2009). � Language-based: these techniques rely on Natural Language Pro- cessing, as these do not consider names as simply strings but words in some natural language. Techniques in this category are, for example, tokenisation, lemmatisation or stopword elimi- nation, some of which are applied by Shah and Syeda-Mahmood in Shah and Syeda-Mahmood (2004). This category also consid- ers those techniques that take advantage from external resources to find similarities between terms, using for instance, lexicons, dictionaries or thesauri. In He, Yang, and Huang (2011) for instance, the WordNet (WordNet, 2013) database is used as the external resource. � Constraint-based: these techniques consider criteria regarding the internal structure of the entities, such as the domain and range of the properties or the types of the attributes, to calculate the similarity between them. It is common to use these tech- niques in combination with others as in the work by Glückstad (2010). � Graph-based: these techniques consider the ontologies to match as labelled graphs, or even trees, and treat the ontology match- ing problem as a graph homomorphism problem. An example of these techniques can be found in the paper from Joslyn, Paulson, and White (2009). This category also considers those techniques that exploit as external resources, repositories where ontologies and their fragments, together with certain similarity measures are stored. A proposal in this line can be found in the paper from Aleksovski, Ten Kate, and Van Harmelen (2008). � Taxonomy-based: these techniques can be seen as a particular case of the previous ones which only consider the specialization relation. Examples were these techniques were applied could not be found in the articles belonging to basic matching tech- niques. However, an example of its application can be found in the work by Warin and Volk (2004). � Instance-based: these techniques exploit the extension of the classes in the ontologies, i.e., the individuals, with the intuition that if the individuals are alike, then the classes they belong to should also be similar. These techniques can use set-theoretic principles but also more elaborated statistical techniques. In this category we can classify the work by Loia, Fenza, De Maio and Salerno presented in Loia, Fenza, De Maio, and Salerno (2013). � Model-based: these techniques exploit the semantic interpreta- tion linked to the input ontologies. An example of this category are the description logics reasoning techniques, which are applied in the work published by Sánchez-Ruiz, Ontañón, González- Calero, and Plaza (2011). As mentioned at the beginning of this section, the techniques we have just presented, are the building blocks upon which (ii) complex matching techniques are built. This category includes meth- odologies that propose different ways to tackle the matching prob- lem but from a higher point of view. Examples of this can be found in Cohen et al. (2003) where Cohen, Ravikumar and Fienberg pro- pose the partitioning of the ontologies before starting the matching process, in Dargham and Fares (2008) where the Dargham and Fares present their methodology which takes as basis some well- known algorithms or in Acampora, Loia, Salerno, and Vitiello (2012) where Acampora, Loia, Salerno and Vitiello propose the use of a memetic algorithm to perform the alignment between two ontologies. 956 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 Other matching techniques also included in this category are those that can not be considered basic or that take advantage of other aspects of building a matching solution that are not directly related to computing the alignments. There are for instance articles devoted to presenting (i) different ways of aggregating the results from differ- ent similarity measures (Lai et al., 2010; Lin & Sandkuhl, 2007; Tian & Guo, 2010), others that (ii) combine the results of different matchers (Liu et al., 2012). There are others that also include techniques inher- ited from other fields such as (iii) learning methods (Rubiolo, Caliusco, Stegmayer, Coronel, & Fabrizi, 2012; Todorov, Geibel, & Kühnberger, 2010), probabilistic methods (Calı̀, Lukasiewicz, Predoiu, & Stuckenschmidt, 2008; Spiliopoulos, Vouros, & Karkaletsis, 2010) or those that consider the user’s involvement (Lin & Sandkuhl, 2008b) in the matching process. Despite the various types of matching techniques that have been presented in this section, both basic and complex ones, we are certain that many others will continue to arise. Some of these may offer a revisited version of previously exiting techniques, but more likely new types of techniques will be defined specially for those categories not fully explored yet. The main challenges these techniques face is their efficiency (Kotis & Lanzenberger, 2008; Shvaiko & Euzenat, 2013). Most of them describe approaches to matching ontologies which at a theoretical level may obtain a positive outcome. However, depending on the type of application where these techniques will be implemented, not all techniques will be equally valid. For instance, if considering a dynamic appli- cation, the amount of resources employed in terms of memory and time consumption should be kept to a minimum. 4.3. ‘Matching Systems’ category This category contains the articles focused on detailing new matching algorithms and systems as well as enhancements, mod- ifications or different approaches to previously defined ones. Some of these systems are well-known in the research community as they have participated for several years in the Ontology Alignment Evaluation Initiative. Although the purpose of this review is not to compile an exhaustive list of all the existing systems, it is worth mentioning some of the most relevant ones, such as: � AgreementMaker (Cruz, Antonelli, & Stroe, 2009a) is a schema and ontology matching system. It allows a high customization of the matching process, including several matching methods to be run on inputs with different levels of granularity, also allowing to define the amount of user participation and the for- mats that the input ontologies as well as the results of the align- ment may be stored in. This system has a high level of maturity because from 2007 there is a continuous flow of publications describing its foundations and enhancements: Sunna and Cruz (2007), Cruz, Antonelli, Stroe, Keles, and Maduko (2008), Cruz, Antonelli, and Stroe (2009b, 2009c), Cruz et al. (2010), Pesquita, Stroe, Cruz, and Couto (2010), Cross et al. (2011), Cruz, Stroe, Pesquita, Couto, and Cross (2011) and Cruz et al. (2011). � Anchor-Flood (Hanif & Aono, 2009) is an algorithm for ontology matching that starts out of an initial anchor, i.e, a pair of alike concepts between the ontologies. From this anchor using neigh- borhood concepts, new anchors are identified and the algorithm continues. Anchor-Flood was developed from 2008 to 2009, having in this period a total of 3 reported papers, (Hanif & Aono, 2008a, 2008b, 2009). This system has been tested in two different campaigns of OAEI. � AOAS (Zhang & Bodenreider, 2007) is an ontology matching sys- tem specifically devoted to aligning anatomical ontologies. It takes as input OWL ontologies and identifies 1:1, 1:n and n:m alignments. It is a hybrid approach that uses both direct tech- niques, such as lexical and structural ones, and indirect tech- niques, that consist in the identification of correspondences by means of a reference ontology. AOAS is a really specific sys- tem which, in the period from 2003 to 2013, only accounts for one publication, (Zhang & Bodenreider, 2007). � AROMA (David, 2011) finds equivalence and subsumption rela- tions between classes and properties of two different taxono- mies. It is defined as an hybrid, extensional and asymmetric approach that lays its foundations on the association rule para- digm and statistical measures. AROMA’s developers have been constantly working on it since 2006. There has been at least one paper published every year detailing this system and its evolution for the past 7 years. (David, Guillet, & Briand, 2006; David, 2007, 2008b, 2008a, 2011). This system has taken part in several editions of the OAEI. � ASCO (Le, Dieng-Kuntz, & Gandon, 2004) exploits all the infor- mation available from the entities in the ontologies, names, labels, descriptions, information about the structure, etc, to compute two types of similarities, a linguistic and a structural one, which are lately combined. The development of ASCO started in 2004, however it was discontinued until 2007 when it was reprised. In this 3 year-span, the publications describing it are: Le et al. (2004) and Thanh Le and Dieng-Kuntz (2007). � ASE (Kotis, Katasonov, & Leino, 2012a) is an automated ontology alignment tool based on AUTOMSv2 that computes equivalence and subsumption relations between two input ontologies. This system was released in 2012 and tested that year’s edition of the OAEI. So far, we have not found any other publication apart from Kotis et al. (2012a), describing it. However it is possible that other enhancements may be developed as this system was already based on a previous system by the same authors for which there was also a significant interval between the first release and the subsequent updates. � ASMOV (Behkamal, Naghibzadeh, & Moghadam, 2010) is an algorithm that derives an alignment from the lexical and struc- tural information of two input ontologies by computing a sim- ilarity measure between them. This algorithm also includes a step of semantic verification where the alignments are checked so that the final output does not contain semantic inconsisten- cies. The greatest efforts in maintaining ASMOV were concen- trated in 2008 when authors published Jean-Mary, Shironoshita, and Kabuka (2008) and Jean-Mary and Kabuka (2008), however, this system was constantly maintained from 2007 to 2010. In this period the following articles describe its features and performance: Jean-Mary and Kabuka (2007), Jean-Mary, Shironoshita, and Kabuka (2009, 2010) and Behkamal et al. (2010). Some of these articles detail the results obtained by ASMOV in the different editions it took part in. � AUTOMSv2 (Kotis, Katasonov, & Leino, 2012b) is an automated ontology matching tool that was build as an evolution of the previous tool AUTOMS (Kotis, Valarakos, & Vouros, 2006) which was enhanced with more alignment methods and synthesizing approaches as well as with multilingual support. Articles describing AUTOMSv2 were published in 2008 and 2012. This system shows one of the highest intervals of inactivity of the studied ones. Due to the short time interval from the last article, new articles describing AUTOMSv2 participation in OAEI or new versions could be published. � Coincidence-Based Weighting (Qazvinian, Abolhassani, (Hossein), & Hariri, 2008) uses an evolutionary approach for ontology matching. It takes as input OWL ontologies, which are pro- cessed as graphs, and a similarity matrix between the concepts of the ontologies, obtained from a string distance measure. It includes a genetic algorithm to iteratively refine the mappings. L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 957 This system was developed for 2 years, since 2007 to 2008, as reflect the articles describing it, (Haeri, Abolhassani, Qazvinian, & Hariri, 2007; Qazvinian et al., 2008). � CIDER (Gracia, Bernad, & Mena, 2011) is an ontology matching system that extracts the ontological context of the compared terms by using synonyms, hyponyms, domains, etc., and then enriches it by means of some lightweight inference rules. This system was developed using the Alignment API (David, Euzenat, Scharffe, & dos Santos, 2011). This system has been developed and maintained since 2008 to 2013. From the first paper describing it (Gracia & Mena, 2008) until the next one (Gracia et al., 2011), 3 years passed. Then again, publications regarding this system, were interrupted until last year with (Gracia & Asooja, 2013). This kind of systems developed over several years are usually the product of a deep effort of the developers to correct the issues detected in previous versions, and account for the time-span existing between the publication of the different articles. � CODI (Huber, Sztyler, Nößner, & Meilicke, 2011) uses some lex- ical similarity measures combined with schema information to output the alignments between concepts, properties and indi- viduals. It is based on the syntax and semantics of Markov logic, and turns every matching problem into an optimization prob- lem. According to the data in our review, CODI was developed and maintained from 2010 to 2011. In these years it took part in the corresponding OAEI contests (Huber et al., 2011; Noessner & Niepert, 2010). � COMA (Maßmann, Raunich, Aumüller, Arnold, & Rahm, 2011), COMA++ (Nasir & Noor, 2010) are ontology matching systems that have been developed over the last decade. They are highly evolved and customizable systems that support the combina- tion of different matching algorithms. These are highly evolved systems since they have been maintained and updated until 2011. In this period 5 articles were found, devoted to describing these systems, (Aumüller, Do, Massmann, & Rahm, 2005; Aumueller, Do, Massmann, & Rahm, 2005; Do & Rahm, 2002; Do, 2006; Engmann & Maßmann, 2007; Massmann, Engmann, & Rahm, 2006; Maßmann et al., 2011; Nasir & Noor, 2010). � DSSim (Nagy, Vargas-Vera, & Stolarski, 2009) is an ontology matching system which combines the similarity values pro- vided by both syntactic and semantic similarity algorithms to then refine the correctness of the outputs by means of a belief function. DSSim was developed from 2006 to 2009. In these 4 years, there was at least an article published every year detail- ing the system, its continuous enhancements, and its results in OAEI (Nagy, Vargas-Vera, & Motta, 2006, 2007; Nagy, Vargas- Vera, Stolarski, & Motta, 2008; Nagy et al., 2009). � Eff2Match (Chua & Kim, 2010) is an ontology matching tool that follows a process of anchor generation and expansion. This sys- tem is particularly focused on achieving an efficient perfor- mance and therefore it includes several techniques to reduce the amount of possible candidates to avoid unnecessary com- parisons. Considering the data obtained from our literature review, Eff2Match only has a related publication that was released in 2010 that presents its results for OAEI’10. � FalconAO (Hu & Qu, 2008) obtains the alignment between the input ontologies by internally running two algorithms, a lin- guistic one (LMO) (Zhang, Hu, & Qu, 2011) as a first step, to then use the alignments provided as an external output for the graph matching algorithm (GMO) (Hu, Jian, Qu, & Wang, 2005) subse- quently run. This system was continuously developed from 2005 to 2008 (Hu, Cheng, Zheng, Zhong, & Qu, 2006; Hu et al., 2007; Hu & Qu, 2008; Jian, Hu, Cheng, & Qu, 2005). In 2010, a new publication was released (Hu, Chen, Cheng, & Qu, 2010) with the results of this system in the OAEI. � FBEM (Stoermer & Rassadko, 2009a) is a matching system that is mainly focused on instance matching, this approach considers not only the similarity of entity features as keys and values, but also the fact that some features are more relevant for iden- tifying an entity than others. According to the data obtained in this review, this system is described in just 2 publications released in 2009 and 2010 respectively, (Stoermer & Rassadko, 2009b; Stoermer, Rassadko, & Vaidya, 2010). � FuzzyAlign (Fernández, Velasco, Marsa-Maestre, & Lopez- Carmona, 2012) is a fuzzy, rule-based ontology matching sys- tem that outputs the alignments between two input ontologies by exploiting the lexical and semantical information of the enti- ties’ names and the inner structure of the ontologies. This sys- tem was developed from 2009 to 2012, however, it was not a constant maintenance, since in this period only 2 articles account for its updates and enhancements, (Fernández, Velasco, & López-Carmona, 2009; Fernández et al., 2012). � GeRoMeSuite (Quix, Gal, Sagi, & Kensche, 2010) allows the matching of models represented in different languages, for instance XML Schemas with OWL ontologies. Besides it is a cus- tomizable system that includes several matching algorithms that may be combined according to different ways of aggrega- tion and filtering. GeRoMeSuite is a mature system that has been developed from 2007 to 2010 (Kensche, Quix, Li, & Li, 2007; Quix, Geisler, Kensche, & Li, 2008, 2009; Quix et al., 2010). How- ever, from 2010 to 2013 no new improvements were publicly released. It took part in OAEI editions from 2008 to 2010. � GLUE (Doan, Madhavan, Domingos, & Halevy, 2004) is a semi- automatic ontology matching system that uses a set of com- bined machine learning techniques to output the alignment between the taxonomies of two input ontologies. This system only has a publication released in 2004, which made us believe that its development has been discontinued. � GOMMA (Hartung, Kolb, Groß, & Rahm, 2013) uses several matchers to evaluate both the lexical and structural similarity of the entities from the input ontologies. It uses some enhanced comparison techniques to compute parallel string matching on graphical processing units. GOMMA was first released in 2011, and it has continued to be maintained up to date. The maturity level of this system is significative, as at least one publication regarding its results has been released every year since it was first developed (Groß, Hartung, Kirsten, & Rahm, 2012; Hartung et al., 2013; Kirsten, Gross, Hartung, & Rahm, 2011). This system has been tested so far in a edition of the OAEI, (Groß et al., 2012). � HCONE (Kotis & Vouros, 2004) is an approach to ontology merg- ing that uses WordNet (WordNet, 2013), a lexical database, as an external resource to obtain possible interpretations of the concepts being matched. HCONE was an stable system main- tained from 2004 to 2006 (Kotis & Vouros, 2004; Kotis, Vouros, & Stergiou, 2006; Vouros & Kotis, 2005). After 2006 we could not retrieve any publication were it was used or modified. � Hertuda (Hertling, 2012) is a very simple string matcher that separately handles the alignment of classes and properties, and select among the possible ones those reaching certain pre-established thresholds. Hertuda is a relatively young sys- tem which was released in 2012. Within the limits of this liter- ature review only two articles were found that described its overall behavior and results in OAEI: Hertling (2012) and Grau et al. (2013). � HotMatch (Dang et al., 2012a) combines several matching strat- egies, that exploit both the lexical and structural information, to obtain the alignments between the ontologies. Some filters are included to remove the false-positive mappings from the final 958 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 output. As happens with Hertuda, HotMatch is also a young sys- tem developed from 2012 up to date (Dang et al., 2012b; Grau et al., 2013), hence, new versions and enhancements are to be expected. This system, has also taken part in OAEI’12. � HMatch (Castano, Ferrara, & Messa, 2006) is an ontology match- ing system that linearly combines a linguistic affinity value and a contextual affinity one to compute the similarity of concept names and contexts. Internally it codifies the ontologies as graphs. HMatch was developed from 2006 to 2009 (Castano et al., 2006; Castano, Ferrara, Lorusso, & Montanelli, 2007; Castano et al., 2008). According to the articles retrieved for this review, HMatch only took part in the edition of 2006 of the OAEI. � IF-MAP (Kalfoglou & Schorlemmer, 2003a) is a matching system that lies its foundations on the mathematical theory of semantic information flow. To match two input ontologies it uses a refer- ence ontology used as common reference. In the time-span con- sidered in this literature review, only one article was found devoted to this system. Such publication was released in 2003, it is possible that there are other articles that had pub- lished before that, which would fall outside the scope of this review, anyhow, considering the time elapsed, we are prone to believing that this system has been discontinued. � iMatch (Albagli, Ben-Eliyahu-Zohary, & Shimony, 2012) is a probabilistic ontology matching system based on Markov net- works. It takes OWL ontologies as input and outputs 1:1 align- ments. The matching is tackled as a graph matching problem where the initial similarity between the nodes is provided, for instance, by the users. iMatch was first released in 2009 (Albagli, Ben-Eliyahu-Zohary, & Shimony, 2009) and then revis- ited in 2012 (Albagli et al., 2012). � KOSImap (Reul & Pan, 2010) uses description logic reasoning to firstly obtain implicit background knowledge for every entity in the ontologies, then build a similarity matrix out of the three types of similarities computed for the identified pairs of enti- ties, and finally dismiss those mappings considered false-posi- tives. KOSImap was only maintained for 2 years, since 2009, when it took part in the OAEI, (Reul & Pan, 2009) to 2010 (Reul & Pan, 2010). � LDOA (Kachroudi, Moussa, Zghal, & Yahia, 2011) combines some well-known terminological and structural similarity measures, but it also exploits an external resource by using Linked Data which provides additional information to the entities being matched. In the interval considered in this literature review, we have only retrieved one article devoted to describing this system, in 2011, describing its behavior in the OAEI. As it is rel- atively contemporary, new enhancements and publications are still to be expected. � Lily (Wang, 2011) combines different matching strategies to adapt itself to the problem being tackled at each moment, gen- eric ontology matching (GOM) for the normal-sized ontologies and large scale ontology matching (LOM) for more demanding matching tasks. It also includes a mapping debugging function used to improve the alignment results and to dismiss the faulty ones. Lily was first released in 2007 (Wang & Xu, 2007) and con- tinued to take part in the OAEI contests until 2011 (Wang & Xu, 2008; Wang & Xu, 2009; Wang, 2011). We consider that the reliability of this system has been enough proven as show the different results obtained in the OAEI. � LogMap (Jiménez-Ruiz & Cuenca Grau, 2011) is an ontology matching iterative process that starting with a set on anchor mappings obtained from lexical comparison, alternatively com- putes mapping repair and mapping discovery steps. To discover the new anchors structural information is also exploited. This system has a high level of maturity as in the last 3 years there have been at least 6 publication describing its performance and results in the OAEI contests (Jiménez-Ruiz & Cuenca Grau, 2011; Jiménez-Ruiz, Grau, & Zhou, 2011; Jiménez-Ruiz, Morant, & Grau, 2011; Jiménez-Ruiz, Grau, & Horrocks, 2012; Jiménez-Ruiz, Meilicke, Grau, & Horrocks, 2013). In this period LogMap’s developers have already implemented a light version of LogMap, LogMaplt. We are prone to believing that developers of this system will continue to improve it and include further functionalities and versions. � MaasMatch (Schadd & Roos, 2012a) computes a similarity cube between the concepts in the ontologies which is the result of aggregating a syntactic, a structural, a lexical and a virtual docu- ment similarity. An extraction algorithm is run to dismiss the faulty alignments from the final output. This system has been constantly updated since it was first released in 2011, its reliabil- ity and usefulness have been tested over the years by its partici- pation in the OAEI (Schadd & Roos, 2011, 2012a, 2012b, 2013). � MapPSO (Bock, Dänschel, & Stumpp, 2011) applies the particle swam optimization technique (PSO) to compute the alignment between two input ontologies. The MapEVO (Bock et al., 2011) system, developed by the same authors, relies on the use of evo- lutionary programming, another variant of population-based optimization algorithms. MapPSO has been described in at least 5 different publications between 2008 and 2011. Among those systems revised, MapPSO is one of those with the higher amount of publications (Bock & Hettenhausen, 2008, 2010; Bock, Liu, & Hettenhausen, 2009; Bock, 2010; Bock, Lenk, & Dänschel, 2010; Bock et al., 2011). These publications include the partici- pation of MapPSO in editions of OAEI from 2008 to 2011. � MapSSS (Cheatham, 2011) computes subsequently three types of metrics, syntactic, semantic and structural, and any positive result from any of them is included as a positive solution, and then it explores the neighborhood of the newly matched pair for new possible matches. Instead of defining a filtering system which would dismiss possible pair after being selected, this sys- tem works the other way round only selecting those nodes that match to only another node, and therefore not risking the pos- sibility of choosing a wrong solution. This system was released in 2011, and, ever since it took part in the annual contest of the OAEI. This system has been therefore significantly tested and evaluated (Cheatham, 2011; Cheatham & Hitzler, 2013). � MEDLEY (Hassen, 2012) is an ontology alignment system that uses lexical and structural methods to compute the alignment between classes, properties and instances. It also uses an exter- nal dictionary to tackle the problem of having concepts expressed in different natural languages. This system was described in 2012 in Hassen (2012) where the results of its par- ticipation in that year’s OAEI are summarized. Other publica- tions, as well as its participation in new OAEI editions, are to be expected because this system quite recent. � MoTo (Fanizzi, d’Amato, & Esposito, 2011) takes OWL ontologies as input and obtains equivalence relations between concepts. It initially uses several matchers whose results are combined by means of a metalearner. The alignments obtained are sorted, discarding those invalid ones. The remaining are divided into certain and uncertain. For the uncertain ones a validation pro- cess is started aiming at recovering them. In this literature review we have found publications describing this system both in 2010 (Esposito, Fanizzi, & d’Amato, 2010) and 2011 (Fanizzi et al., 2011), but so far, no other publication has been released. � OACAS (Zghal, Kachroudi, Yahia, & Nguifo, 2011) is an algorithm to align OWL-DL ontologies. It firstly transforms the ontologies into graphs and then it combines and aggregates different sim- ilarity measures. At each moment the most suitable similarity measure is applied according to the type of the entities being matched. It also exploits the neighboring relations of the enti- ties. The first article describing OACAS was published in 2009 (Zghal, Kachroudi, Yahia, & Nguifo, 2009), then until 2011 no L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 959 new articles were retrieved (Zghal et al., 2011). Therefore, con- sidering the two-year span between both, it is possible that new articles will be published within this or the next year. This sys- tem was publicly tested in the OAEI’11. � OLA (Kengue, Euzenat, & Valtchev, 2007) performs the alignment between two graph-represented ontologies and offers some extended features to manipulate the output alignment. This sys- tem was developed between 2004 (Euzenat & Valtchev, 2004) and 2005 (Euzenat, Guégan, & Valtchev, 2005), later, it was revisited in 2007 (Kengue et al., 2007). In this interval, it took part in the OAEI of 2005 and 2007. However, so far, it has been 6 years where no additional articles regarding OLA were retrieve by means of the queries run for this literature review. � oMap (Straccia & Troncy, 2005c) automatically aligns two OWL ontologies by using the prediction of different classifiers, such as terminological, machine learning-based, or some based on the structure and semantics of the OWL axioms. This system was deeply revised in 2005 (Straccia & Troncy, 2005a; Straccia & Troncy, 2005c; Straccia & Troncy, 2005b) when it was released and then in 2006 (Straccia & Troncy, 2006). In 2005 it took part in the OAEI, however, ever since, no new articles describing oMap have been published. � OMEN (Mitra, Noy, & Jaiswal, 2005) uses a Bayesian Network to improve the results of the alignment process by deriving unde- tected correspondences and rejecting existing false positives. By using probabilistic methods it enhances existing ontology map- pings by deriving missed matches and invalidating existing false matches. This system was described in just one paper in 2005 (Mitra et al., 2005). � OntoDNA (Kiu & Lee, 2007) is an automated and scalable system that uses hybrid unsupervised clustering techniques, which include Formal Concept Analysis (FCA) (Formica, 2006), Self- Organizing Map (SOM) and K-Means clustering, in addition to a Levenshtein edit distance as lexical measurement. OntoDNA was first described in 2006 (Kiu & Lee, 2006) and then revisited in 2007 to present its results in the OAEI (Kiu & Lee, 2007). � ontoMATCH (Lu, 2010) exploits both semantic and structural information to compute the alignments. It is internally divided into five components, a preprocessor, three individual matchers (Element Matcher, Relationship Matcher and Property Matcher), a combiner and a selector. This system was described in one paper in 2010 (Lu, 2010). � OPTIMA (Thayasivam et al., 2012) uses lexical information from the concepts to generate a seed alignment. Then it iteratively searches the space of candidate alignments following the tech- nique of expectation–maximization until convergence. This sys- tem was presented in 2011 taking part at the OAEI (Thayasivam & Doshi, 2011). It also took part in the following year’s contest (Thayasivam et al., 2012). � OWL-CM (Yaghlane & Laamari, 2007) uses different matchers to compute the alignments, whose results are then combined. It also includes some belief functions into the alignment process to improve the computed results. In this literature review we only retrieved one article in 2007 (Yaghlane & Laamari, 2007). � PRIOR+ (Mao & Peng, 2007) is an ontology matching system that lays its foundations on propagation theory, information retrie- val and artificial intelligence. It profits the linguistic and struc- tural information of the ontologies to match and measures the profile similarity of different elements in a vector space model. This system took part in the editions of 2006 (Mao & Peng, 2006) and 2007 (Mao & Peng, 2007) of the OAEI where its results and performance are described. � QOM (Ehrig & Staab, 2004) is a variation of the NOM algorithm devoted to improving the efficiency of the system. Some basic matchers are used whose results are refined by means of a sig- moid function, to be lately aggregated and sifted to output the final alignment. This system was developed in the first years of the interval considered in this literature review in 2 papers (Ehrig & Staab, 2004; Ehrig & Sure, 2004). � RiMOM (Wang et al., 2010) uses three different matching strat- egies, name-based, metadata-based and instance-based, whose results are then filtered and combined. A similarity propagation procedure is iteratively run until no more candidate mappings are discovered and the system converges. Within the limits of this review, RiMOM is the system that accounts for the highest number of individual publications describing it, 9. These articles span from 2004 to 2013. In this period, there have been periods of interruption in the flow of publications, however we believe they account for the development of a new version of the sys- tem (Li, Li, Zhang, & Tang, 2006; Li, Zhong, Li, & Tang, 2007; Li, Tang, Li, & Luo, 2009; Tang, Liang, Li, & Wang, 2004; Tang et al., 2006; Wang et al., 2010; Zhang, Zhong, Li, & Tang, 2008; Zhang, Zhong, Shi, Li, & Tang, 2009). RiMOM has taken part in several editions of the OAEI, therefore it has been tested and evaluated over several years. � SAMBO (Lambrix, Tan, & Liu, 2008) contains several matchers that exploit different features of the ontologies, and it is the user the one who decides to use one or several. If several are chosen the combination of the results by each of them are com- puted by means of a weighted sum. Results are then filtered according to some thresholds and presented to the user as sug- gested alignments to be confirmed. This system has been main- tained from 2005 to 2008. In this period, at least 5 different articles were published describing its performance and its results in the OAEI (Lambrix & Tan, 2005, 2006; Lambrix et al., 2008; Tan, Jakoniene, Lambrix, Aberg, & Shahmehri, 2006; Tan & Lambrix, 2007). � SEMA (Spiliopoulos, Valarakos, Vouros, & Karkaletsis, 2007) combines six different matching methods whose running sequence is pre-established and where each method takes as input the results of the previous methods. This procedure is iteratively applied until no new mappings are discovered. The matchers used a are lexical matcher, a latent features matcher, a vector space model matcher, an instance based matcher a structural based matcher and a property based matcher. SEMA was described in 2 different articles in 2007 (Spiliopoulos et al., 2007), one of them presenting its results in the OAEI con- test (Spiliopoulos et al., 2007). � SERIMI (Araújo, de Vries, & Schwabe, 2011c) is a matching sys- tem developed for instance matching, which is mainly divided into two phases, a selection one and a disambiguation one. Dur- ing these phases information retrieval strategies and string matching techniques are applied. This system was described in 4 articles between 2011 and 2012 (Araújo, Hidders, Schwabe, & de Vries, 2011a, 2011b; Araújo, Tran, DeVries, Hidders, & Schwabe, 2012). In 2011 it also took part in the OAEI (Araújo et al., 2011c). � ServOMap (Ba & Diallo, 2013) is a large scale ontology matching system that also supports multilingual terminologies. It uses an Ontology Server (ServO) and takes advantage of Information Retrieval techniques to compute the similarity between the entities in the input ontologies. This system was recently devel- oped, in 2012 (Ba & Diallo, 2012a, 2013; Diallo & Ba, 2012). However it has already taken part in the OAEI (Ba & Diallo, 2012b). This points out that developers are actively working on the maintenance of this system, therefore new improve- ments are to be expected, in addition to those already released in 2013 (Diallo & Kammoun, 2013; Kammoun & Diallo, 2013). � SIGMa (Lacoste-Julien et al., 2013) is a knowledge base iterative propagation alignment algorithm that uses both the structural information from the relationship graph as well as some simi- larity measures between entity properties. SIGMa is also one 960 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 of the systems that can be considered as young since the first publication where it is described is from 2012 (Lacoste-Julien et al., 2012). This system is being actively maintained, and a new publication was released in 2013 (Lacoste-Julien et al., 2013). � S-Match (Giunchiglia, Autayeu, & Pane, 2012) is an open source semantic matching framework that transforms input tree-like structures such as catalogs, conceptual models, etc., into light- weight ontologies to then determine the semantic correspon- dences between them. It contains the implementation of several semantic matching algorithms, each one suitable for dif- ferent purposes. The amount of publications describing S-Match is one of the highest among those considered for this review, 7. This system has been maintained since 2003, however it was not a steady process, as the last publications have intervals between them of at least 2 years (Giunchiglia, Shvaiko, & Yatskevich, 2004, 2005a, 2005b; Giunchiglia, Yatskevich, & Shvaiko, 2007; Giunchiglia et al., 2012; Shvaiko, Giunchiglia, & Yatskevich, 2009). � SOBOM (Xu, Wang, Cheng, & Zang, 2010a) is an ontology match- ing algorithm that uses as start point a set of anchors provided by a lexical matcher. It also uses the Semantic Inductive Similar- ity Flooding algorithm to compute the similarity between the concepts of the sub-ontologies obtained from the anchors. SOBOM has been described in publications ranging from 2008 to 2012 (Xu, Tao, Zang, & Wang, 2008; Xu et al., 2010a; Xu, Wang, Cheng, & Zang, 2010b; Xu, Wang, & Liu, 2012), including those of its participation in several editions of OEAI. In the two- year period from 2010 to 2012, no publications were retrieved regarding this system. As the last article dates from 2012, new updates and enhancements are still to be expected. � SODA (Zghal, Yahia, Nguifo, & Slimani, 2007) uses linguistic and structural similarity measures to compute the alignment between two OWL-DL ontologies which are firstly transformed into graphs. This graphs undergo two successive phases, first a linguistic similarity comparison and then a structural similarity comparison, after which the semantic similarity of the graphs is obtained. This algorithm outputs the correspondences between the entities together with their similarity measure values. So far, only one article describing SODA was found (Zghal et al., 2007), published in 2007. � TaxoMap (Hamdi, Safar, Niraula, & Reynaud, 2010) provides an alignment for two OWL ontologies by exploiting the informa- tion in the labels of the concepts and the subsumption links that connect those concepts in the hierarchy. TaxoMap has been maintained since it was released in 2007 up to 2010. In this interval, 6 articles were published describing the system and its participation in the OAEI (Hamdi, Zargayouna, Safar, & Reynaud, 2008; Hamdi, Safar, Niraula, & Reynaud, 2009; Hamdi et al., 2010; Safar & Reynaud, 2009; Safar, 2007; Zargayouna, Safar, & Reynaud, 2007). � TOAST (Jachnik, Szwabe, Misiorek, & Walkowiak, 2012) is an ontology matching system based on statistical relational learn- ing. This system needs a train set from which to learn the semantics equivalence relation on the basis of partial matches. TOAST is another system, considered young, as all the publications describing it were found for 2012 (Szwabe, Misiorek, & Walkowiak, 2012). However, this system has already taken part in the OAEI (Jachnik et al., 2012). If the devel- opers continue with this trend, new articles on this system are to be expected. � WeSeE-Match (Paulheim, 2012) exploits the idea of using infor- mation available on the web to match the ontologies which would supposedly be the procedure followed by a human trying to manually match some terms without being an expert in the domain of the matched terms. Therefore this approach uses a web search engine to retrieve documents relevant to the con- cepts to match and compare the results obtained, the more sim- ilar the search results, the higher the concepts’ similarity value. This system accounts for few publications, however these report the results obtained by the system in the OAEI of 2012 and 2013 (Paulheim, 2012; Paulheim & Hertling, 2013). � WikiMatch (Hertling & Paulheim, 2012a) exploits the use of Wikipedia’s search engine to obtain documents related to the concepts being matched. Since there is no duplicity in the titles names of the articles in Wikipedia for the same language, the algorithm compares the sets of retrieved titles to obtain the sim- ilarity between the two concepts. As happens with WeSeE- Match, WikiMatch has been only described so far in two articles, although these are quite recent. One of them presenting its over- all behavior (Hertling & Paulheim, 2012a) and the other report- ing its results for the OAEI’12 (Hertling & Paulheim, 2012b). � X-SOM (Curino, Orsi, & Tanca, 2007b) combines the similarity maps output by different matching algorithms by means of a neural network and uses logical reasoning and heuristics to enhance the quality of the mappings. This system was devel- oped between 2007 and 2010, adding up a total of 4 articles (Curino, Orsi, & Tanca, 2007a; Curino et al., 2007b; Merlin, Sorjamaa, Maillet, & Lendasse, 2009; Merlin, Sorjamaa, Maillet, & Lendasse, 2010), including those that describe its par- ticipation in OAEI’07. � YAM++ (Ngo & Bellahsene, 2012a) uses machine learning tech- niques to discover the mappings between entities in two ontol- ogies, even if these are not expressed in the same natural language. It uses matchers at element and structural level. At element level the similarity is computed by some terminologi- cal metrics which can be combined by machine learning based combination methods. At structural level the ontologies are transformed into graphs and considering the results of the ter- minological metrics as the starting points, a similarity flooding algorithm propagation is run. This system has a high level of maturity as it has been continuously evolving since 2009 up to 2013 (Duchateau, Coletta, Bellahsene, & Miller, 2009b, 2009a; Ngo, Bellahsene, & Coletta, 2011; Ngo & Bellahsene, 2012a, 2012b, 2013). In this period, at least 6 articles describing its behavior and overall results in the different editions of the OAEI have been published. Hence its validity and maturity is well proven. Further considering the evolution and maturity degree of these systems, in Table 5 the amount of articles published regarding each one of the presented systems is shown. These values show a vary- ing level of development in the different systems, as the amount of publications ranges from 1 to 9. In Fig. 8 these results are disaggre- gated by year. As it suggests, systems are devoted on average 2:6 years of work, which are usually consecutive. However, as hap- pens, for instance, with ASCO the work was interrupted and resumed 3 years later. In other systems, as RiMOM, this discontin- uation in the amount of publications accounts for the development of a new version. 4.4. ’Processing Frameworks’ category This category covers two types of publications, articles devoted to researching the processing and exploiting of the ontology align- ments (25.85%) and also those that describe some enhanced align- ment frameworks and alignment formats (74.10%). Among those articles devoted to processing and exploiting the alignments, the most common topics were related to (i) ontology merging (Kim, Kim, & Chung, 2011), i.e., integrating two ontologies from different sources into a single new one with the information from both of them, (ii) ontology transformation (Šváb-Zamazal, Table 5 Amount of articles yearly devoted to each system. System Articles AgreementMaker 9 Anchor-Flood 3 AOAS 1 AROMA 8 ASCO 2 ASE 1 ASMOV 5 AUTOMSv2 2 CBW 2 CIDER 3 CODI 2 COMA 5 DSSim 4 Eff2Match 1 FalconAO 5 FBEM 2 FuzzyAlign 2 GeRoMeSuite 4 GLUE 1 GOMMA 3 HCONE 3 Hertuda 2 HotMatch 2 Hmatch 3 IF-MAP 1 iMatch 2 KOSImap 2 LDOA 1 Lily 4 LogMap 6 MaasMatch 3 MapPSO 5 MapSSS 3 MEDLEY 1 MoTo 2 OACAS 2 OLA 3 oMap 4 OMEN 1 OntoDNA 2 ontoMATCH 1 OPTIMA 2 OWL-CM 1 PRIOR+ 2 QOM 2 RiMOM 9 SAMBO 5 SEMA 2 SERIMI 4 ServOMap 5 SIGMa 2 S-Match 7 SOBOM 4 SODA 1 TaxoMap 6 TOAST 2 WeSeE-Match 2 WikiMatch 2 X-SOM 4 YAM++ 6 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 961 Svátek, & Iannone, 2010), that implies expressing an ontology with respect to another one, (iii) reasoning (Zhang, Lin, Huang, & Wu, 2011), that involves using the correspondences between the ontol- ogies as rules for reasoning with them, and (iv) alignment argumen- tation (Trojahn, Quaresma, & Vieira, 2012), that is a way of explaining the alignments by providing arguments to support or dismiss them. Despite the remarkable variety of topics in the articles from this sub-category, they represent a small percentage of the processing frameworks category, which in their majority were focused mainly on defining and developing alignment frameworks such as (Noy & Musen, 2003), where the process does not finish with the align- ment but other actions are also available for the user, like manip- ulating the alignments or performing some of the procedures previously mentioned. 4.5. ‘Practical Applications’ category The articles in this category present articles where ontology matching has been applied to a real-life problem. Within this cat- egory we found articles devoted to different subjects such as (i) semantic web and web services (Di Martino, 2009) where most pub- lications presented ways of using ontology matching for service discovery or service composition, (ii) P2P systems (Atencia, Euzenat, Pirrò, & Rousset, 2011), where ontology matching was used as a way to reduce the semantical heterogeneity between the queries the users pose to the system and the documents stored, therefore improving the accuracy of the returned results. Other fields worth mentioning are (iii) learning systems (Arch-int & Arch-int, 2013), that focus the use of ontology matching tech- niques either on narrowing down the distance between the user’s and the stored documents or as a way to ease the knowledge share and reuse among users, and last, (iv) multi-agent systems (Mascardi, Ancona, Bordini, & Ricci, 2011), where the use of ontology match- ing has always been related to guaranteeing that the different agents in a communication process could be actually able to inter- act and achieve the common goals. In spite of the growing tendency in the development of practical applications, in general lines, it does not reach the 30% of the matching systems implemented each year, as Fig. 9 reflects. This situation is quite remarkable as it suggests that only a slight part of the matching systems developed have a practical application in real-life projects. To clarify this situation we have conducted a survey among ontology matching practitioners, where we asked them mainly about the future challenges of the field and its appli- cation in real-life projects. The description and results of this sur- vey are further detailed in Section 6. 4.6. ‘Evaluation’ category Out of the articles for this literature review, 38 are devoted to evaluating the performance of the matching systems. We can split these articles into two categories regarding the scope of the arti- cles. There are 14 articles (36.84%) focused on studying the perfor- mance measures and on proposing different alternatives to evaluate the matching systems. We have included such articles in a category named elementary approaches. The remaining 24 articles (63.16%) delve into evaluation meth- ods were different existing platforms, systems or benchmarks to evaluate the matching systems are explored. Regarding the (i) elementary approaches several articles explore alternatives to the well-known information retrieval measures of precision and recall which are used in this field to evaluate respec- tively the correctness and completeness of the matching systems. Examples of such publications are the works by Paulheim, Hertling and Ritze (Paulheim, Hertling, & Ritze, 2013), Niu, Wang, Wu, Qi and Yu (Niu, Wang, Wu, Qi, & Yu, 2011) or Euzenat (Euzenat, 2007). Additionally in this category we have also included those papers that describe a new evaluation method or approach such as the ones by Ferrara, Nikolov, Noessner and Scharffe (Ferrara, Nikolov, Noessner, & Scharffe, 2013) and by Tordai, van Ossenbrug- gen, Schreiber and Wielinga (Tordai, van Ossenbruggen, Schreiber, & Wielinga, 2011). These measures and approaches are usually included as part of the systems proposed to evaluate the matching systems which are included in the (ii) evaluation methods category. In this category we have included those papers delving into the different existing Fig. 8. Evolution of the systems over the years. 962 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 platforms, systems and benchmarks used for evaluation, as well as the results of evaluating the most widespread systems against these benchmarks or in these platforms. Regarding the data sets or benchmarks for evaluation, the most well-known and used are those developed for the Ontology Align- ment Evaluation Initiative which has been taken as a reference since 2004 (Euzenat, Stuckenschmidt, & Yatskevich, 2005) and that has been evolving over the years (Rosoiu, dos Santos, & Euzenat, 2011). This initiative contains various tracks with different data sets which evaluate several features of the tested systems. The benchmark test (Euzenat, Ros�oiu, & Trojahn, 2013) is built around a seed ontology and many variations of it, and its purpose is to provide a stable and detailed picture of the contesting algo- rithms. These tests are organized into simple tests, where the objec- tive is to compare the original ontology with itself, a random one and a generalization, systematic tests, where the original ontology is to be compared with others where some modifications have been included, such as removing names, translating into other languages, flattening or expanding the hierarchy, etc., and finally, real-life ontologies. The anatomy track evaluates the matching sys- tems with the task of matching two large ontologies, the Adult Mouse Anatomy and part of the NCI Thesaurus which describes the human anatomy. The conference track contains different ontol- ogies from the conference organization domain. The interest of this track lies in the fact that these ontologies have been independently defined. The MultiFarm track aims at testing the ability of the sys- tems to deal with multilingualism. The library track is a real-world task to match two thesaurus, the STW and the TheSoz, both used in libraries for indexation and retrieval. The interactive track tests the results obtained by the systems when the user is somehow involved. The Large BioMed track consists of finding alignments between the Foundational Model of Anatomy (FMA, 2013), SNOMED Clinical Terms (SNOMED, 2013), and the National Cancer Institute Thesaurus (NCI, 2013). Finally, the Instance Matching track focus its efforts on instance matching systems and techniques. Fig. 9. Evolution of matching systems vs. practical applications. L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 963 Besides the benchmarks, in this category we have also included some actual systems used to evaluate the matching systems such as Wrigley, García-Castro, and Nixon (2012) and Tyl and Loufek (2009). In this section we have presented our classification framework and further detailed the results of classifying the retrieved articles with this classification framework. In the following section we ana- lyze the limitations to this review. 5. Limitations of the Literature Review A literature review in the field of ontology matching is a very demanding task firstly due to the amount of background knowl- edge necessary to properly sort the identified articles, and secondly by the extent of the subject itself and the number of fields where the research on ontology matching is used. The articles studied for this review were retrieved by querying online databases with different expressions regarding the ontology matching field. In spite of the high amount of articles retrieved, over 1600, it is possible that many others have not been recovered as we are dealing with a very wide field of knowledge. Other databases besides those queried for this review, could have also been used to raise the amount of retrieved articles and broaden the scope of the review, however, the databases used are considered as the most relevant ones among the practitioners. In addition only those articles in english were finally included even though some publications were written in other languages, as we considered english as the predominant language regarding the research community. In spite of the limitations previously described, this paper makes a brief review of the ontology matching field between 2000 and 2013. The articles written in this period were also sorted according to a classification framework which has allowed us to identify the different topics and problems that researchers were tackling for the last decade. Nevertheless this has also brought up several questions regarding the current research interests of researchers and practitioners, mainly whether they have continued to research on the same topics or not, and to check, for instance, if they have changed topic or even field. Another main issue that we have detected is the fact that, in the last decade lots of different matching systems and techniques have been developed, however, we could not state their use in real-life applications. In order to clarify these doubts we have designed and per- formed a survey among the researchers. Its structure as well as the results of this survey are detailed in Section 6. 6. Trends in Ontology Matching: Practitioner-oriented Survey We conducted a survey to clarify those concerns emerged from our literature review. Such concerns were mainly related to the current state of the research on ontology matching and its applica- tion in real-life projects. 6.1. Participants and Survey design The participants in the survey were selected among those tak- ing part at the OAEI contests. In a two month period, from Decem- ber 2013 to February 2014, they were individually contacted by email and presented with the questionnaire shown in Table 6. Even though the participants were directly contacted by email, their identities and responses were strictly confidential and only avail- able to the team conducting the survey. Out of the 288 experts con- tacted, we received 46 replies. The survey was designed with 8 short open-ended questions. Although we have initially considered to define some of these questions as multiple-choice, we discarded that idea as we did not want to influence at any degree in the answers provided by the participants. These questions can be classified into three groups. Questions 1 to 3 are background questions, questions 4; 5 and 7 are research field questions and questions 6 and 8 are future challenges questions. The background questions were included in the questionnaire to assess the suitability of the participants and contextualize the answers they may provide. The research field questions were designed to gain knowledge about the current fields and topics that have become more attractive to the research community, and finally future challenges questions were designed to identify according to practitioners’ point of view the main challenges that are still to be addressed and the potential expansion fields for ontology matching. 6.2. Survey results Out of the 46 answers that we received, only 5 declined the par- ticipation in the survey, one answered it partially and 13 have not been researching on ontology matching for a while. Out of this researchers that have stopped working in the field, some of them have recently stopped and they answered the questionnaire any- how, as their contribution was still relevant while others suggested a more appropriate contact within their groups to redirect the requests, half of which were answered back. To sum up, the initial Table 6 Questionnaire used for the survey. Number Question Type 1. How long have you been researching in Ontology Matching? Background 2. What are your main purposes to do it? Background 3. How many research papers have you written on topics related to Ontology Matching? Background 4. Within the Ontology Matching field, in which particular topic are you currently working on? Research field 5. From your point of view, which are the main fields where the research on Ontology Matching is currently being applied? Research field 6. According to your expertise, which are the main challenges that are still to be addressed? Future challenges 7. Will you continue to research in Ontology Matching?Why? Research field 8. In which fields do you believe that Ontology Matching could also be used? Future challenges 964 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 amount of replies was cut down to 33 actual answers with profit- able information. 6.2.1. Background questions The background of the participants in the research is quite broad and varied, and it includes different types of researchers in the field. From a temporal point of view, there are those who have started in the field more than a decade ago but also those who have recently started to work in this field, specifically values range from 1 to 14 years. Most of the practitioners (78.12%) who answered this questionnaire have been researching in the field for over 5 years, being the average number of years working in the field is slightly superior to 7. In Fig. 10, the number of years researching in the field is shown in relation to the number of researchers. Moreover some researchers have directly tackled the ontology matching problem by focussing on very specific topics such as assessing the impact of using different similarity metrics in different ontology matching tasks, aligning large ontologies or improving ontol- ogy matching by using reasoning, while others arrived to ontology matching as a support tool for matters such as data integration, semantic interoperability or telecommunications systems interoperability. Anyhow the suitability of the participants is more than satisfac- tory as they account on their own for above 460 publications of dif- ferent types linked to this subject. 6.2.2. Research field questions The research field questions were included in the questionnaire to learn about the fields where respondents are working, as well as, the fields where they think ontology matching could also be applied. Out of the answers sent by the participants, we could determine that a high percentage of them (63%) is working in any of these four topics: instance matching, user involvement in the matching process, data interlinking and discovery of different types of correspondences, not 1:1 equivalence relations. The rest of the respondents mentioned other topics such as parallel ontology matching, large-scale ontology matching, ontology matching negotia- tion and mapping reuse, which are more specific. However, besides these main research topic, most respondents included similarity metrics and combinations of methods to improve the coverage of ontology matching as a way to enhance or support their other research interests. Regarding the question about the fields where ontology match- ing is being applied, the consensus shown was noticeable. This question provided mainly two types of answers. From a practical point of view, respondents agree that the medical and life science domain is the one that is using ontology matching the most. Other researchers offered a more theoretical type of answers, mostly mentioning data integration and interoperability as the fields where ontology matching is being applied. We have found really meaningful that several researchers pointed out that nowadays the use of ontology matching tech- niques is reserved to spot cases and that the research at this time seems merely foundational. However, they also agree that ontol- ogy matching can be applied in any field where there are two par- ties that need to communicate and that employ potentially different protocols, being this way the list of use cases potentially long. Finally, when questioned about whether they would continue to research on this field, the majority, 63.64%, confirmed they would follow with present or related research lines claiming as reasons for instance, that there are still plenty of challenges to address and that the development of new domains will sparkle new matching problems. On the contrary, 30.30% of the respondents sta- ted that they would, if not yet, change subjects. Among the reasons to do so, some mentioned they have moved to other related fields such as linking open information systems or knowledge transforma- tion, while others definitely quit the field claiming the lack of use- fulness for real applications or the little incentive coming from the application side. A small percentage, 6.06% were still considering whether to change subject or not. 6.2.3. Future challenges questions These group of questions were included in the questionnaire to gain knowledge about how practitioners see the future evolution of the field and the main challenges still to be addressed. These answers provided are really useful as they identify a variety of challenges to address and quote several new fields where matching techniques could also be used and hence they could be used to guide the research lines adopted by different research groups. Regarding the main challenges still to be addressed in the ontol- ogy matching field, most respondents agree on the need to auto- matically discover complex relations, instead of 1:1, to correctly align large ontologies and to focus on applying automatically created mappings to practical applications. Other topics that arose in the responses were not supported by so many respondents but they point out anyway challenges that need to be addressed, such as: 1. Automated acquisition of reference alignment for evaluating large scale matching systems. 2. Creating large datasets to asses matching algorithms. 3. Define good tools that are easy to use for non-experts. 4. Develop high quality and fast intelligent combinations of string-based and new semantic-similarity measures. 5. Holistic ontology matching. 6. How to effectively complement automatic computation with human validation. 7. How to minimize involvement of users when turning matches into mapping. 8. Human readable explanations for matches. 9. Improving the mapping process through semi-automatic machine learning. 10. Integration of domain knowledge into alignment techniques. 11. Learning what metrics to choose in which scenario. 12. Precision and Recall of automatic methods. 13. Scalability and parallelization of the matching. Fig. 10. Number of researchers in relation with the number of years working in the field. L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 965 14. Semantic mapping. The answers provided by the practitioners for this question point out challenges that are inline with those highlighted by Euz- enat and Shvaiko in Shvaiko and Euzenat (2013) and Euzenat and Shvaiko (2013). Table 7 presents a comparison of those challenges outlined by Euzenat and Shvaiko and those mentioned by respondents to our survey. As this table states, most of the challenges issued by prac- titioners have been also considered in those by Euzenat and Shvaiko, however there are certain mismatches worth noticing. As previously stated, most researchers mainly agree on 3 chal- lenges, identified in Table 7 with ‘⁄’, however, just one of these can be classified into the categories by Euzenat and Shvaiko. It is worth noticing that one of this challenges is the application of automatically created mappings to practical applications, which supports our working hypothesis to start this survey, that the application of ontology matching techniques in real-life projects is still a field to be further developed. Some practitioners, in addition to pointing out the challenges, also used the answer to this question to mention certain situations that they consider a mistake, as the fact that most approaches to ontology matching focus on lexicographic and structural informa- tion while language is more complex than that and hence achiev- ing a perfect precision and recall is impossible in real life applications. Other aspect they complained about was benchmarks habitually used to test the matching systems (OAEI benchmarks). They claim that even if they are really useful, their main drawback is that there are yet too many artificial datasets and tasks in it. Other practitioners took their remarks a step forward claiming that science culture does not reward creating and maintaining one tool and instead everyone creates a prototype for a paper and then abandons it. Finally, regarding the fields where ontology matching could also be applied we have obtain two types of answers. Some practitio- ners gave a fuzzy answer mentioning that matching techniques could be used practically anywhere where is no standard for informa- tion exchange and where the domains are open to adopt ontology approaches or in broad sense in any information related field. On the contrary, others actually mentioned fields to apply these techniques, such as: bioinformatics, information systems, e-com- merce, web services, intrusion detection systems, cultural heritage, library science, government, education, banking, personal and social data management, law, etc. Most agree that the fact that in these fields ontologies and ontology matching techniques are not already in use is due to a lack of information regarding the potential benefits. 7. Limitations of the Survey There are, of course, limitations to this survey, the foremost being the sampling size and the population. Although we feel that our 33 final responses offered a wide variety of useful remarks and points of view, it is true that the sample is still quite small and hence our analysis may be biased. Besides, in an effort to prevent the questions from influencing the answers and to obtain as much information as possible, we have defined the questionnaire with 8 open-ended questions. This fact, possibly together with the way some questions were posed, led us to obtaining answers that, Table 7 Comparison of future challenges. Challenges identified by Euzenat & Shvaiko Challenges mentioned by practitioners LARGE-SCALE AND EFFICIENT MATCHING Automated acquisition of reference alignment for evaluating large scale matching systems Creating large datasets to asses matching algorithms Scalability and parallelization of the matching ⁄ Correctly align large ontologies MATCHING WITH BACKGROUND KNOWLEDGE Integration of domain knowledge into alignment techniques MATCHER SELECTION, COMBINATION AND TUNING Develop high quality and fast intelligent combinations of string-based and new semantic-similarity measures Learning what metrics to choose in which scenario USER INVOLVEMENT How to effectively complement automatic computation with human validation How to minimize involvement of users when turning matches into mapping EXPLANATION OF MATCHING RESULTS Human readable explanations for matches UNCERTAINTY IN ONTOLOGY MATCHING – ALIGNMENT MANAGEMENT – Other challenges that do not fit in previous categories Improving the mapping process through semi-automatic machine learning Precision and Recall of automatic methods Semantic mapping Define good tools that are easy to use for non-experts Holistic ontology matching ⁄ Automatically discover complex relations, instead of 1:1 ⁄ Focus on applying automatically created mappings to practical applications 966 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 however interesting, did not exactly match what we expected from them. Also, the participants targeted were obtained from the par- ticipants at the OAEI contests, most of which are academically ori- ented, therefore our survey may be biased towards academical researchers rather than a balance between academical and indus- trial researchers. Finally, this survey was the answer to some con- cerns that arose while conducting the literature review, and we consider the results here as a first analysis. Our intention is to revi- sit these answers looking for deeper connections between the answers, to address questions such as: Is there any relation between how long a researcher has been working in a subject and the amount of publications?, Which type of researchers are more prone to quitting?, etc. 8. Conclusions In this paper, we have achieved a twofold goal. Initially we per- formed a literature review of the ontology matching field, whose results led to the definition and development of the survey lately conducted. To address the task of performing a literature review of the ontology matching field, we have defined a classification frame- work which helped in structuring our review by providing a com- prehensive model to sort the different types of publications. This review was based on an online search of ontology matching related papers from 2003 to the first semester of 2013. The initial amount of articles obtained, over 1600, was reduced by filtering them according to their topics, keywords, abstracts and content. With the articles left after the several trimming iterations, we have initially performed a statistical evaluation and analysis. Later, we have sorted the articles following the framework and then we have analyzed each one of the categories in the framework, evalu- ating the different types of articles and topics treated. While performing this deeper analysis of the articles and their topics, some concerns arose regarding the actual research interests of the practitioners as we detected a high amount of papers related to theoretical solutions and approaches while the number of applied ones was significantly lower. The approach chosen to clar- ify these concerns was to ask openly to the research community, by means of a practitioner-oriented survey. The purpose of such survey was to gain knowledge about the current state of the ontology matching field and the application of such techniques to real-life environments. We have noticed that most researchers share the same concerns about the practical application of the ontology matching techniques, and the problem of having too many theoretical solutions but few applied ones. However, due to the nature of the survey, with open-ended ques- tions, there is more information that we have not reflected in this work and which we plan on analyzing and exploiting in the future. By means of this work we have provided a general overview of the ontology matching field in the last decade. It can be used as a starting point for new practitioners to get a general idea but also, to help in deciding on research lines, hopefully by tackling some of the challenges highlighted in the survey. References Acampora, G., Loia, V., Salerno, S., & Vitiello, A. (2012). A hybrid evolutionary approach for solving the ontology alignment problem. International Journal of Intelligent Systems, 27, 189–216. ACM. (2013). ACM Digital Library. URL: <http://dl.acm.org/>. Akbari, I., Fathian, M., & Badie, K. (2009). An improved mlma+ and its application in ontology matching. In Innovative technologies in intelligent systems and industrial applications, 2009. CITISIA 2009 (pp. 56–60). Albagli, S., Ben-Eliyahu-Zohary, R., & Shimony, S. E. (2009). Markov network based ontology matching. In IJCAI (pp. 1884–1889). Albagli, S., Ben-Eliyahu-Zohary, R., & Shimony, S. E. (2012). Markov network based ontology matching. Journal of Computer and System Sciences, 78, 105–118. Aleksovski, Z., Ten Kate, W., & Van Harmelen, F. (2008). Using multiple ontologies as background knowledge in ontology matching. In Proceedings of the 1st international workshop on collective semantics: collective intelligence and the semantic web CISWeb’08 (pp. 35–49). Araújo, S., Hidders, J., Schwabe, D., & de Vries, A. P. (2011a). Serimi – resource description similarity, rdf instance matching and interlinking. CoRR abs/ 1107.1104. URL: <http://dblp.uni-trier.de/db/journals/corr/ corr1107.html#abs-1107-1104>. Araújo, S., Hidders, J., Schwabe, D., & de Vries, A. P. (2011b). Serimi – resource description similarity, rdf instance matching and interlinking. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2011.html#AraujoHSV11>. Araújo, S., Tran, D., DeVries, A., Hidders, J., & Schwabe, D. (2012). Serimi: Class-based disambiguation for effective instance matching over heterogeneous web data. In Z. G. Ives, & Y. Velegrakis (Eds.), WebDB (pp. 25–30). URL: <http://dblp.uni- trier.de/db/conf/webdb/webdb2012.html#AraujoTDHS12>. Araújo, S., de Vries, A.P., & Schwabe, D., (2011c). SERIMI results for OAEI 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR- WS.org. (pp. 212–219). Arch-int, N., & Arch-int, S. (2013). Semantic ontology mapping for interoperability of learning resource systems using a rule-based reasoning approach. Expert Systems with Applications, 40, 7428–7443. Atencia, M., Euzenat, J., Pirrò, G., & Rousset, M. C. (2011). Alignment-based trust for resource finding in semantic p2p networks. In Proceedings of the 10th http://refhub.elsevier.com/S0957-4174(14)00514-4/h0005 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0005 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0005 http://dl.acm.org/ http://refhub.elsevier.com/S0957-4174(14)00514-4/h0025 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0025 http://dblp.uni-trier.de/db/journals/corr/corr1107.html#abs-1107-1104 http://dblp.uni-trier.de/db/journals/corr/corr1107.html#abs-1107-1104 http://dblp.uni-trier.de/db/conf/semweb/om2011.html#AraujoHSV11 http://dblp.uni-trier.de/db/conf/webdb/webdb2012.html#AraujoTDHS12 http://dblp.uni-trier.de/db/conf/webdb/webdb2012.html#AraujoTDHS12 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0055 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0055 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0055 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0060 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0060 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 967 international conference on the semantic web – Volume Part I (pp. 51–66). Berlin, Heidelberg: Springer-Verlag. Aumüller, D., Do, H. H., Massmann, S., & Rahm, E. (2005). Schema and ontology matching with coma++. In F. Özcan (Ed.), SIGMOD conference (pp. 906–908). ACM. URL:<http:// dblp.uni-trier.de/db/conf/sigmod/sigmod2005.html#AumuellerDMR05>. Ba, M., & Diallo, G. (2012a). Knowledge reposiory as entity similarity computing enabler. In SITIS (pp. 975–981). Ba, M., & Diallo, G. (2012b). Servomap and servomap-lt results for oaei 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/ om2012.html#BaD12>. Ba, M., & Diallo, G. (2013). Large-scale biomedical ontology matching with ServOMap. IRBM, 34, 56–59. Behkamal, B., Naghibzadeh, M., & Moghadam, R. (2010). Using pattern detection techniques and refactoring to improve the performance of ASMOV. In 5th International symposium on telecommunications (IST) (pp. 979–984). Bock, J. (2010). Mappso results for oaei 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL: <http:// dblp.uni-trier.de/db/conf/semweb/om2010.html#Bock10>. Bock, J., Dänschel, C., & Stumpp, M. (2011). MapPSO and MapEVO results for OAEI 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 179–183). Bock, J., & Hettenhausen, J. (2008). Mappso results for oaei 2008. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2008.html#BockH08>. Bock, J., & Hettenhausen, J. (2010). Discrete particle swarm optimisation for ontology alignment. Information Sciences [in Press]. doi:http://dx.doi.org/ 10.1016/j.ins.2010.08.013. Bock, J., Lenk, A., & Dänschel, C., 2010. Ontology Alignment in the cloud. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. Cruz (Eds.), Proceedings of the 5th international workshop on ontology matching (OM-2010), CEUR workshop proceedings. <http://ceur-ws.org> (pp. 73–84). Bock, J., Liu, P., & Hettenhausen, J. (2009). Mappso results for oaei 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/ om2009.html#BockLH08>. Calı̀, A., Lukasiewicz, T., Predoiu, L., & Stuckenschmidt, H. (2008). Tightly integrated probabilistic description logic programs for representing ontology mappings. In Proceedings of the 5th international conference on Foundations of information and knowledge systems (pp. 178–198). Berlin, Heidelberg: Springer-Verlag. Castano, S., Ferrara, A., Lorusso, D., & Montanelli, S. (2007). The hmatch 2.0 suite for ontology matchmaking. In G. Semeraro, E. D. Sciascio, C. Morbidoni, & H. Stoermer (Eds.), SWAP, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/ swap/swap2007.html#CastanoFLM07>. Castano, S., Ferrara, A., Lorusso, D., & Montanelli, S. (2008). On the ontology instance matching problem. In 19th International workshop on database and expert systems application, 2008. DEXA’08 (pp. 180–184). Castano, S., Ferrara, A., & Messa, G. (2006). Results of the HMatch ontology matchmaker in OAEI 2006. In P. Shvaiko, J. Euzenat, N. F. Noy, H. Stuckenschmidt, V. R. Benjamins, & M. Uschold (Eds.), Ontology matching, CEUR-WS.org. (pp. 134–143). Cheatham, M. (2011). MapSSS results for OAEI 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 171–179). Cheatham, M., & Hitzler, P. (2013). Stringsauto and mapsss results for oaei 2013. In OM (pp. 146–152). Chua, W. W. K., & Kim, J. J. (2010). Eff2Match results for OAEI 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I.F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 150–157). Cohen, W. W., Ravikumar, P. D., & Fienberg, S. E. (2003). A Comparison of string distance metrics for name-matching tasks. In IIWeb (pp. 73–78). Cross, V., Stroe, C., Hu, X., Silwal, P., Panahiazar, M., Cruz, I. F., Parikh, P., & Sheth, A. P. (2011). Aligning the parasite experiment ontology and the ontology for biomedical investigations using agreementmaker. In O. Bodenreider, M. E. Martone, & A. Ruttenberg (Eds.), ICBO, CEUR-WS.org. URL: <http://dblp.uni- trier.de/db/conf/icbo/icbo2011.html#CrossSHSPCPS11>. Cruz, I. F., Antonelli, F. P., & Stroe, C. (2009a). AgreementMaker: Efficient matching for large real-world schemas and ontologies. In Proc. VLDB Endow. 2 (pp. 1586– 1589). Cruz, I. F., Antonelli, F. P., & Stroe, C. (2009b). Agreementmaker: Efficient matching for large real-world schemas and ontologies. PVLDB 2 (pp. 1586–1589). URL: <http://dblp.uni-trier.de/db/journals/pvldb/pvldb2.html#CruzAS09>. Cruz, I. F., Antonelli, F. P., & Stroe, C. (2009c). Efficient selection of mappings and automatic quality-driven combination of matching methods. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. Noy, & A. Rosenthal (Eds.), Proceedings of the 4th international workshop on ontology matching, CEUR workshop proceedings (pp. 49–60). URL: <http://ceur-ws.org>. Cruz, I. F., Antonelli, F. P., Stroe, C., Keles, U. C., & Maduko, A. (2008). Using agreementmaker to align ontologies for oaei 2009: Overview, results, and outlook. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/ semweb/om2009.html#CruzASKM08>. Cruz, I. F., Stroe, C., Caci, M., Caimi, F., Palmonari, M., Antonelli, F. P., & Keles, U. C. (2010). Using agreementmaker to align ontologies for oaei 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/ om2010.html#CruzSCCPAK10>. Cruz, I. F., Stroe, C., Caimi, F., Fabiani, A., Pesquita, C., Couto, F. M., & Palmonari, M. (2011). Using agreementmaker to align ontologies for oaei 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2011.html#CruzSCFPCP11>. Cruz, I. F., Stroe, C., Pesquita, C., Couto, F.M., & Cross, V. (2011). Biomedical ontology matching using the agreementmaker system. In O. Bodenreider, M. E. Martone, & A. Ruttenberg (Eds.), ICBO, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/ conf/icbo/icbo2011.html#CruzSPCC11>. Curino, C., Orsi, G., & Tanca, L. (2007a). X-som: A flexible ontology mapper, In Proc. of SWAE (DEXA Workshops) (pp. 424–428). Curino, C., Orsi, G., & Tanca, L. (2007b). X-SOM Results for OAEI 2007. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. (pp. 276–285). Dang, T. T., Gabriel, A., Hertling, S., Roskosch, P., Wlotzka, M., Zilke, J. R., Janssen, F., & Paulheim, H. (2012a). HotMatch results for OEAI 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR- WS.org. (pp. 145–151). Dang, T. T., Gabriel, A., Hertling, S., Roskosch, P., Wlotzka, M., Zilke, J. R., Janssen, F., & Paulheim, H. (2012b). Hotmatch results for oeai 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR- WS.org. URL:<http://dblp.uni-trier.de/db/conf/semweb/ om2012.html#DangGHRWZJP12>. Dargham, J., & Fares, E. (2008). A hybrid approach for ontology mapping. In Proceedings of the 2008 international conference on semantic web and web services, SWWS 2008 (pp. 343–349). David, J. (2007). AROMA: A method for the discovery of alignments between ontologies from association rules. Thèse d’informatique. Université de Nantes. Nantes (FR). URL: <http://tel.archives-ouvertes.fr/tel-00200040/en/>. David, J. (2008a). Aroma results for oaei 2008. David, J. (2008b). Aroma results for oaei 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR- WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2009.html# David08a>. David, J. (2011). AROMA results for OAEI 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 122–125). David, J., Euzenat, J., Scharffe, F., & dos Santos, C. T. (2011). The alignment api 4.0. Semantic Web 2, 3–10. David, J., Guillet, F., & Briand, H. (2006). Matching directories and owl ontologies with aroma. In P. Yu, V. Tsotras, E. Fox, & B. Liu (Eds.), Proc. 15th ACM international conference on information and knowledge management (CIKM), Arlington (VA US) (pp. 830–831). US: ACM. Di Martino, B. (2009). Semantic web services discovery based on structural ontology matching. International Journal of Web and Grid Services, 5, 46–65. Diallo, G. & Ba, M. (2012). Effective Method for Large Scale Ontology Matching. Semantic WebApplications and Tools for Life Sciences (SWAT4LS). CEUR Workshop Proceedings, 952. Diallo, G., & Kammoun, A. (2013). Towards learning based strategy for improving the recall of the servomap matching system. In A. Paschke, A. Burger, P. Romano, M. S. Marshall, & A. Splendiani (Eds.), SWAT4LS, CEUR-WS.org. URL:<http:// dblp.uni-trier.de/db/conf/swat4ls/swat4ls2013.html#DialloK13>. Direct, S. (2013). Science Direct. URL: <http://www.sciencedirect.com/>. Do, H., & Rahm, E. (2002). COMA – a system for flexible combination of schema matching approaches. In Proceedings of the 28th VLDB conference, Hong Kong, China. URL: <citeseer.nj.nec.com/do02coma.html>. Do, H. H. (2006). Schema matching and mapping-based data integration: Architecture, approaches and evaluation. Saarbrücken: Vdm Verlag Dr. Müller. Doan, A., Madhavan, J., Domingos, P., & Halevy, A. (2004). Ontology matching: A machine learning approach. In Handbook on ontologies in information systems (pp. 385–403). Springer-Verlag. Droge, E. (2010). Guidelines on ontology matching. Information-Wissenschaft und Praxis, 61, 143–147. Duchateau, F., Coletta, R., Bellahsene, Z., & Miller, R. J. (2009a). (not) yet another matcher. In CIKM (pp. 1537–1540). Duchateau, F., Coletta, R., Bellahsene, Z., & Miller, R. J. (2009b). Yam: A schema matcher factory. In CIKM (pp. 2079–2080). Ehrig, M., & Staab, S. (2004). QOM – quick ontology mapping. In Proc. 3rd international semantic web conference (ISWC04) (pp. 683–697). Springer. Ehrig, M., & Sure, Y. (2004). Ontology mapping – an integrated approach. In ESWS (pp. 76–91). Engmann, D., & Maßmann, S. (2007). Instance matching with coma++. In M. Jarke, T. Seidl, C. Quix, D. Kensche, S. Conrad, & E. Rahm, et al. (Eds.), BTW workshops (pp. 28–37). Aachen: Verlagshaus Mainz. URL:<http://dblp.uni-trier.de/db/conf/ btw/btw2007w.html#EngmannM07>. Esposito, F., Fanizzi, N., & d’Amato, C. (2010). Recovering uncertain mappings through structural validation and aggregation with the moto system. In SAC (pp. 1428–1432). Euzenat, J. (2004). State of the art on ontology alignment. Knowledge Web, 2. Euzenat, J. (2007). Semantic precision and recall for ontology alignment evaluation. In IJCAI international joint conference on artificial intelligence (pp. 348–353). Euzenat, J., Guégan, P., & Valtchev, P. (2005). Ola in the oaei 2005 alignment contest. In Integrating Ontologies. Euzenat, J., Ros�oiu, M. E., & Trojahn, C. (2013). Ontology matching benchmarks: Generation, stability, and discriminability. Journal of Web Semantics, 21, 30–48. Euzenat, J., & Shvaiko, P. (2007). Ontology matching. Berlin, New York: Springer. Euzenat, J., & Shvaiko, P. (2013). Ontology matching (2nd ed.). Heidelberg (DE): Springer-Verlag. URL:<http://book.ontologymatching.org>. http://refhub.elsevier.com/S0957-4174(14)00514-4/h0060 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0060 http://dblp.uni-trier.de/db/conf/sigmod/sigmod2005.html#AumuellerDMR05 http://dblp.uni-trier.de/db/conf/sigmod/sigmod2005.html#AumuellerDMR05 http://dblp.uni-trier.de/db/conf/semweb/om2012.html#BaD12 http://dblp.uni-trier.de/db/conf/semweb/om2012.html#BaD12 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0085 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0085 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#Bock10 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#Bock10 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#BockH08 http://dx.doi.org/10.1016/j.ins.2010.08.013 http://dx.doi.org/10.1016/j.ins.2010.08.013 http://ceur-ws.org http://dblp.uni-trier.de/db/conf/semweb/om2009.html#BockLH08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#BockLH08 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0125 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0125 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0125 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0125 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0125 http://dblp.uni-trier.de/db/conf/swap/swap2007.html#CastanoFLM07 http://dblp.uni-trier.de/db/conf/swap/swap2007.html#CastanoFLM07 http://dblp.uni-trier.de/db/conf/icbo/icbo2011.html#CrossSHSPCPS11 http://dblp.uni-trier.de/db/conf/icbo/icbo2011.html#CrossSHSPCPS11 http://dblp.uni-trier.de/db/journals/pvldb/pvldb2.html#CruzAS09 http://ceur-ws.org http://dblp.uni-trier.de/db/conf/semweb/om2009.html#CruzASKM08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#CruzASKM08 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#CruzSCCPAK10 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#CruzSCCPAK10 http://dblp.uni-trier.de/db/conf/semweb/om2011.html#CruzSCFPCP11 http://dblp.uni-trier.de/db/conf/icbo/icbo2011.html#CruzSPCC11 http://dblp.uni-trier.de/db/conf/icbo/icbo2011.html#CruzSPCC11 http://dblp.uni-trier.de/db/conf/semweb/om2012.html#DangGHRWZJP12 http://dblp.uni-trier.de/db/conf/semweb/om2012.html#DangGHRWZJP12 http://tel.archives-ouvertes.fr/tel-00200040/en/ http://dblp.uni-trier.de/db/conf/semweb/om2009.html#David08a http://dblp.uni-trier.de/db/conf/semweb/om2009.html#David08a http://refhub.elsevier.com/S0957-4174(14)00514-4/h0255 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0255 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0255 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0255 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0260 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0260 http://dblp.uni-trier.de/db/conf/swat4ls/swat4ls2013.html#DialloK13 http://dblp.uni-trier.de/db/conf/swat4ls/swat4ls2013.html#DialloK13 http://www.sciencedirect.com/ http://citeseer.nj.nec.com/do02coma.html http://refhub.elsevier.com/S0957-4174(14)00514-4/h0285 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0285 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0290 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0290 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0290 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0295 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0295 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0310 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0310 http://dblp.uni-trier.de/db/conf/btw/btw2007w.html#EngmannM07 http://dblp.uni-trier.de/db/conf/btw/btw2007w.html#EngmannM07 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0330 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0345 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0345 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0345 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0350 http://book.ontologymatching.org 968 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 Euzenat, J., Stuckenschmidt, H., & Yatskevich, M. (2005). Introduction to the ontology alignment evaluation 2005. In Proceedings of the K-Cap 2005 workshop on integrating ontologies, Banff (Canada) (pp. 97–102). Euzenat, J., & Valtchev, P. (2004). Similarity-based ontology alignment in OWL-Lite. In Proc. of the 15th european conference on artificial intelligence (ECAI) – Valencia (ES) (pp. 333–337). Falconer, S., Noy, N., & Storey, M. A. (2007). Ontology mapping – a user survey. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), Proceedings of the workshop on ontology matching (OM2007) at ISWC/ASWC2007, Busan, South Korea (pp. 113– 125). Fanizzi, N., d’Amato, C., & Esposito, F. (2011). Composite ontology matching with uncertain mappings recovery. ACM SIGAPP Applied Computing Review, 11, 17–29. Fernández, S., Velasco, J., Marsa-Maestre, I., & Lopez-Carmona, M. (2012). FuzzyAlign: A fuzzy method for ontology alignment. In: KEOD 2012 – proceedings of the international conference on knowledge engineering and ontology development (pp. 98–107). Fernández, S., Velasco, J. R., & López-Carmona, M. A. (2009). A fuzzy rule-based system for ontology mapping. In PRIMA (pp. 500–507). Ferrara, A., Nikolov, A., Noessner, J., & Scharffe, F. (2013). Evaluation of instance matching tools: The experience of {OAEI}. Web Semantics: Science, Services and Agents on the World Wide Web 21, 49–60. Special Issue on Evaluation of Semantic Technologies. FMA (2013). Foundational Model of Anatomy. URL: <http:// sig.biostr.washington.edu/projects/fm/>. Formica, A. (2006). Ontology-based concept similarity in formal concept analysis. Information Sciences, 176, 2624–2641. Fu, B., Brennan, R., & O’Sullivan, D. (2009). Multilingual ontology mapping: Challenges and a proposed framework. In Adaptive and emergent behaviour and complex systems – proceedings of the 23rd convention of the society for the study of artificial intelligence and simulation of behaviour, AISB 2009 (pp. 33–35). Fugazza, C., & Vaccari, L. (2011). Coupling human- and machine-driven mapping of SKOS thesauri. International Journal of Metadata, Semantics and Ontologies, 6, 155–165. Gal, A., & Shvaiko, P. (2008). Advances in ontology matching. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNCS, (Vol. 4891, pp. 176–198). Giunchiglia, F., Autayeu, A., & Pane, J. (2012). S-Match: An open source framework for matching lightweight ontologies. Semantic Web, 3, 307–317. Giunchiglia, F., Shvaiko, P., & Yatskevich, M. (2004). S-match: An algorithm and an implementation of semantic matching. In Proc. of the first european semantic web symposium – ESWS (pp. 61–75). Giunchiglia, F., Shvaiko, P., & Yatskevich, M. (2005a). S-match: An algorithm and an implementation of semantic matching. In Semantic Interoperability and Integration. Giunchiglia, F., Shvaiko, P., & Yatskevich, M. (2005b). Semantic schema matching. In OTM conferences (1) (pp. 347–365). Giunchiglia, F., Yatskevich, M., & Shvaiko, P. (2007). Semantic matching: Algorithms and implementation. Journal on Data Semantics, 9, 1–38. Glückstad, F. (2010). Terminological ontology and cognitive processes in translation. In Proceedings of the 24th Pacific Asia conference on language, information and computation (pp. 629–636). Gracia, J., & Asooja, K. (2013). Monolingual and cross-lingual ontology matching with cider-cl: Evaluation report for oaei 2013. In OM (pp. 109–116). Gracia, J., Bernad, J., & Mena, E. (2011). Ontology matching with cider: Evaluation report for oaei 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I.F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 1–8). Gracia, J., & Mena, E. (2008). Ontology matching with cider: Evaluation report for the oaei 2008. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL:<http://dblp.uni-trier.de/db/conf/semweb/ om2008.html#GraciaM08>. Grau, B. C., Dragisic, Z., Eckert, K., Euzenat, J., Ferrara, A., Granada, R., Ivanova, V., Jiménez-Ruiz, E., Kempf, A. O., Lambrix, P., Nikolov, A., Paulheim, H., Ritze, D., Scharffe, F., Shvaiko, P., Trojahn, C., & Zamazal, O. (2013). Results of the ontology alignment evaluation initiative 2013. In: ISWC workshop on ontology matching. Groß, A., Hartung, M., Kirsten, T., & Rahm, E. (2012). Gomma results for oaei 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/ conf/semweb/om2012.html#GrossHKR12>. Haeri, S. H., Abolhassani, H., Qazvinian, V., & Hariri, B. B. (2007). Coincidence-based scoring of mappings in ontology alignment. JACIII, 11, 803–816. URL: <http:// dblp.uni-trier.de/db/journals/jaciii/jaciii11.html#HosseinAQH07>. Hamdi, F., Safar, B., Niraula, N. B., & Reynaud, C. (2009). Taxomap in the oaei 2009 alignment contest. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni- trier.de/db/conf/semweb/om2009.html#HamdiSNR08>. Hamdi, F., Safar, B., Niraula, N. B., & Reynaud, C. (2010). TaxoMap alignment and refinement modules: results for OAEI 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 212–219). Hamdi, F., Zargayouna, H., Safar, B., & Reynaud, C. (2008). Taxomap in the oaei 2008 alignment contest. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL:<http://dblp.uni-trier.de/db/conf/semweb/ om2008.html#HamdiZSR08>. Hanif, M. S., & Aono, M. (2008a). Alignment results of anchor-flood algorithm for oaei-2008. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL:<http://dblp.uni-trier.de/db/conf/semweb/ om2008.html#HanifA08>. Hanif, M. S., & Aono, M. (2008b). Anchor-flood: Results for oaei 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL:<http://dblp.uni-trier.de/db/conf/semweb/ om2009.html#HanifA08a>. Hanif, M. S., & Aono, M. (2009). Anchor-flood: Results for OAEI 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. (pp. 127–134). Hartung, M., Kolb, L., Groß, A., & Rahm, E. (2013). Optimizing similarity computations for ontology matching – experiences from GOMMA. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNBI (Vol. 7970, pp. 81–89). Hassen, W. (2012). MEDLEY results for OAEI 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N.F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR- WS.org. (pp. 168–172). He, W., Yang, X., & Huang, D. (2011). A hybrid approach for measuring semantic similarity between ontologies based on wordnet. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNAI (Vol. 7091, pp. 68–78). Hertling, S. (2012). Hertuda results for OAEI 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR- WS.org. (pp. 141–144). Hertling, S., & Paulheim, H. (2012a). WikiMatch – using Wikipedia for ontology matching. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. (pp. 220–225). Hertling, S., & Paulheim, H. (2012b). Wikimatch results for oeai 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL:<http://dblp.uni-trier.de/db/conf/semweb/ om2012.html#HertlingP12a>. Hu, W., Chen, J., Cheng, G., & Qu, Y. (2010). ObjectCoref and Falcon-AO: Results for OAEI 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL:<http://dblp.uni-trier.de/db/conf/ semweb/om2010.html#HuCCQ10>. Hu, W., Cheng, G., Zheng, D., Zhong, X., & Qu, Y. (2006). The results of falcon-ao in the oaei 2006 campaign. In P. Shvaiko, J. Euzenat, N. F. Noy, H. Stuckenschmidt, V. R. Benjamins, & M. Uschold (Eds.), Ontology matching, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2006.html#HuCZZQ06>. Hu, W., Jian, N., Qu, Y., & Wang, Y. (2005). GMO: A Graph Matching for Ontologies. In B. Ashpole, M. Ehrig, J. Euzenat, & H. Stuckenschmidt (Eds.), Integrating ontologies, CEUR-WS.org. (pp. 41–48). Hu, W., & Qu, Y. (2008). Falcon-AO: A practical ontology matching system. Web Semantics: Science, Services and Agents on the World Wide Web, 6, 237–239. Hu, W., Zhao, Y., Li, D., Cheng, G., Wu, H., & Qu, Y. (2007). Falcon-ao: Results for oaei 2007. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2007.html#HuZLCWQ07>. Huber, J., Sztyler, T., Nößner, J., & Meilicke, C. (2011). CODI: combinatorial optimization for data integration: Results for OAEI 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 134–141). IEEXplore. (2013). IEEEXplore Digital Library. URL: <http://ieeexplore.ieee.org/ Xplore/home.jsp>. Jachnik, A., Szwabe, A., Misiorek, P., & Walkowiak, P. (2012). TOAST results for OAEI 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N.F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. (pp. 205–212). Jean-Mary, Y. R., & Kabuka, M. R. (2007). Asmov results for oaei 2007. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. URL: <http:// dblp.uni-trier.de/db/conf/semweb/om2007.html#Jean-MaryK07>. Jean-Mary, Y. R., & Kabuka, M. R. (2008). Asmov: Results for oaei 2008. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2008.html#Jean-MaryK08>. Jean-Mary, Y. R., Shironoshita, E. P., & Kabuka, M. R. (2008). Asmov: Results for oaei 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/ semweb/om2009.html#Jean-MarySK08>. Jean-Mary, Y. R., Shironoshita, E. P., & Kabuka, M. R. (2009). Ontology matching with semantic verification. Web Semantics, 7, 235–251. Jean-Mary, Y. R., Shironoshita, E. P., & Kabuka, M. R. (2010). Asmov: Results for oaei 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL:<http://dblp.uni-trier.de/db/conf/semweb/ om2010.html#Jean-MarySK10> Jian, N., Hu, W., Cheng, G., & Qu, Y. (2005). Falconao: Aligning ontologies with falcon. In B. Ashpole, M. Ehrig, J. Euzenat, & H. Stuckenschmidt (Eds.), Integrating Ontologies, CEUR-WS.org. URL:<http://dblp.uni-trier.de/db/conf/kcap/ kcap2005w.html#JianHCQ05>. Jiménez-Ruiz, E., & Cuenca Grau, B. (2011). LogMap: Logic-based and scalable ontology matching. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNCS, (Vol. 7031, pp. 273–288). Jiménez-Ruiz, E., Grau, B. C., & Horrocks, I. (2012). Logmap and logmaplt results for oaei 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/ conf/semweb/om2012.html#Jimenez-RuizGH12>. Jiménez-Ruiz, E., Grau, B. C., & Zhou, Y. (2011). Logmap 2.0: Towards logic-based, scalable and interactive ontology matching. In A. Paschke, A. Burger, P. Romando, M. S. Marshall, & A. Splendiani (Eds.), SWAT4LS (pp. 45–46). ACM. http://refhub.elsevier.com/S0957-4174(14)00514-4/h0375 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0375 http://sig.biostr.washington.edu/projects/fm/ http://sig.biostr.washington.edu/projects/fm/ http://refhub.elsevier.com/S0957-4174(14)00514-4/h0400 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0400 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0410 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0410 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0410 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0420 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0420 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0440 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0440 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#GraciaM08 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#GraciaM08 http://dblp.uni-trier.de/db/conf/semweb/om2012.html#GrossHKR12 http://dblp.uni-trier.de/db/conf/semweb/om2012.html#GrossHKR12 http://dblp.uni-trier.de/db/journals/jaciii/jaciii11.html#HosseinAQH07 http://dblp.uni-trier.de/db/journals/jaciii/jaciii11.html#HosseinAQH07 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#HamdiSNR08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#HamdiSNR08 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#HamdiZSR08 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#HamdiZSR08 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#HanifA08 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#HanifA08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#HanifA08a http://dblp.uni-trier.de/db/conf/semweb/om2009.html#HanifA08a http://dblp.uni-trier.de/db/conf/semweb/om2012.html#HertlingP12a http://dblp.uni-trier.de/db/conf/semweb/om2012.html#HertlingP12a http://dblp.uni-trier.de/db/conf/semweb/om2010.html#HuCCQ10 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#HuCCQ10 http://dblp.uni-trier.de/db/conf/semweb/om2006.html#HuCZZQ06 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0555 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0555 http://dblp.uni-trier.de/db/conf/semweb/om2007.html#HuZLCWQ07 http://ieeexplore.ieee.org/Xplore/home.jsp http://ieeexplore.ieee.org/Xplore/home.jsp http://dblp.uni-trier.de/db/conf/semweb/om2007.html#Jean-MaryK07 http://dblp.uni-trier.de/db/conf/semweb/om2007.html#Jean-MaryK07 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#Jean-MaryK08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#Jean-MarySK08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#Jean-MarySK08 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0595 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0595 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#Jean-MarySK10 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#Jean-MarySK10 http://dblp.uni-trier.de/db/conf/kcap/kcap2005w.html#JianHCQ05 http://dblp.uni-trier.de/db/conf/kcap/kcap2005w.html#JianHCQ05 http://dblp.uni-trier.de/db/conf/semweb/om2012.html#Jimenez-RuizGH12 http://dblp.uni-trier.de/db/conf/semweb/om2012.html#Jimenez-RuizGH12 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 969 URL:<http://dblp.uni-trier.de/db/conf/swat4ls/swat4ls2011.html#Jimenez- RuizGZ11>.. Jiménez-Ruiz, E., Meilicke, C., Grau, B. C., & Horrocks, I. (2013). Evaluating mapping repair systems with large biomedical ontologies. In 26th International workshop on description logics (pp. 246–257). Jiménez-Ruiz, E., Morant, A., & Grau, B. C. (2011). Logmap results for oaei 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR- WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2011.html#Jimenez- RuizMG11>. Joslyn, C., Paulson, P., & White, A. M. (2009). Measuring the structural preservation of semantic hierarchy alignment. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. (pp. 1–12). Kachroudi, M., Moussa, E. B., Zghal, S., & Yahia, S. B. (2011). LDOA results for OAEI 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 148–156). Kalfoglou, Y., & Schorlemmer, M. (2003a). IF-Map: An ontology-mapping method based on information-flow theory. Journal of Data Semantics, 98–127. Kalfoglou, Y., & Schorlemmer, M. (2003b). Ontology mapping: The state of the art. Science, 2, 3. Kammoun, A., & Diallo, G. (2013). Servomap results for oaei 2013. In OM (pp. 169– 176). Kengue, J. F. D., Euzenat, J., & Valtchev, P. (2007). OLA in the OAEI 2007 evaluation contest. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR- WS.org. (pp. 188–195). Kensche, D., Quix, C., Li, X., & Li, Y. (2007). Geromesuite: A system for holistic generic model management. In C. Koch, J. Gehrke, M. N. Garofalakis, D. Srivastava, K. Aberer, A. Deshpande, D. Florescu, C. Y. Chan, V. Ganti, C. C. Kanne, W. Klas, & E. J. Neuhold (Eds.), VLDB (pp. 1322–1325). ACM. URL: <http:// dblp.uni-trier.de/db/conf/vldb/vldb2007.html#KenscheQLL07>.. Kim, J., Kim, P., & Chung, H. (2011). Ontology construction using online ontologies based on selection, mapping and merging. International Journal of Web and Grid Services, 7, 170–189. Kirsten, T., Gross, A., Hartung, M., & Rahm, E. (2011). Gomma: A component-based infrastructure for managing and analyzing life science ontologies and their evolution. Journal of Biomedical Semantics, 2, 6. Kiu, C. C., & Lee, C. S. (2006). Ontology mapping and merging through ontodna for learning object reusability. Educational Technology and Society, 9, 27–42. URL: <http://dblp.uni-trier.de/db/journals/ets/ets9.html#KiuL06>. Kiu, C. C., & Lee, C. S. (2007). OntoDNA: Ontology alignment results for OAEI 2007. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. (pp. 196–205). Kotis, K., Katasonov, A., & Leino, J. (2012a). ASE results for OAEI 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. (pp. 116–123). Kotis, K., Katasonov, A., & Leino, J. (2012b). AUTOMSv2 results for OAEI 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. (pp. 124–132). Kotis, K., & Lanzenberger, M. (2008). Ontology matching: Current status, dilemmas and future challenges. In International conference on complex, intelligent and software intensive systems, 2008. CISIS 2008 (pp. 924–927). Kotis, K., Valarakos, A. G., & Vouros, G. A. (2006). Automs: Automated ontology mapping through synthesis of methods. In P. Shvaiko, J. Euzenat, N. F. Noy, H. Stuckenschmidt, V. R. Benjamins, & M. Uschold (Eds.), Ontology matching, CEUR- WS.org. Kotis, K., & Vouros, G. A. (2004). The HCONE approach to ontology merging. In Proceedings of the 21st international symposium on computer architecture (pp. 137–151). IEEE Computer Society Press. Kotis, K., Vouros, G. A., & Stergiou, K. (2006). Towards automatic merging of domain ontologies: The hcone-merge approach. Journal of Web Semantics, 4, 60–79. URL: <http://dblp.uni-trier.de/db/journals/ws/ws4.html#KotisVS06>. Lacoste-Julien, S., Palla, K., Davies, A., Kasneci, G., Graepel, T., & Ghahramani, Z. (2012). Sigma: Simple greedy matching for aligning large knowledge bases. CoRR abs/1207.4525. Lacoste-Julien, S., Palla, K., Davies, A., Kasneci, G., Graepel, T., & Ghahramani, Z. (2013). SIGMa: simple greedy matching for aligning large knowledge bases. In Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 572–580). New York, NY, USA: ACM. http:// dx.doi.org/10.1145/2487575.2487592. Lai, J., Wang, Y., Zhang, R., Gu, X., Yu, T., & Li, J., 2010. Aggregating multiple ontology similarity based on iowa operator. In 2010 2nd International workshop on database technology and applications (DBTA) (pp. 1–4). Lambrix, P., & Tan, H. (2005). A framework for aligning ontologies. In F. Fages & S. Soliman (Eds.), PPSWR (pp. 17–31). Springer. URL: <http://dblp.uni-trier.de/db/ conf/ppswr/ppswr2005.html#LambrixT05>. Lambrix, P., & Tan, H. (2006). Sambo – a system for aligning and merging biomedical ontologies. Journal of Web Semantics, 4, 196–206. Lambrix, P., Tan, H., & Liu, Q. (2008). SAMBO and SAMBOdtf Results for the Ontology Alignment Evaluation Initiative 2008. Proceedings of the 3rd International Workshop on Ontology Matching (OM-2008). CEUR Workshop Proceedings, 431,190–198. Lauser, B., Johannsen, G., Caracciolo, C., Van Hage, W., Keizer, J., & Mayr, P. (2008). Comparing human and automatic thesaurus mapping approaches in the agricultural domain. In Proceedings of the international conference on dublin core and metadata applications (pp. 43–53). Le, B. T., Dieng-Kuntz, R., & Gandon, F. (2004). On ontology matching problems – for building a corporate semantic web in a multi-communities organization. In ICEIS (4) (pp. 236–243). Li, J., Tang, J., Li, Y., & Luo, Q. (2009). RiMOM: A dynamic multistrategy ontology alignment framework. IEEE Transactions on Knowledge and Data Engineering, 21, 1218–1232. doi: http://dx.doi.org/10.1109/TKDE.2008.202. Li, Y., Li, J., Zhang, D., & Tang, J. (2006). Result of ontology alignment with rimom at oaei’06. In P. Shvaiko, J. Euzenat, N. F. Noy, H. Stuckenschmidt, V. R. Benjamins, & M. Uschold (Eds.), Ontology matching, CEUR-WS.org. URL: <http://dblp.uni- trier.de/db/conf/semweb/om2006.html#LiLZT06> Li, Y., Zhong, Q., Li, J., & Tang, J. (2007). Result of ontology alignment with rimom at oaei’07. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2007.html#LiZLT07>. Lin, F., & Sandkuhl, K. (2007). Polygon-based similarity aggregation for ontology matching. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNCS (Vol. 4743, pp. 255– 264). Lin, F., & Sandkuhl, K. (2008a). A survey of exploiting WordNet in ontology matching. IFIP International Federation for Information Processing, 276, 341–350. Lin, F., & Sandkuhl, K. (2008b). User-based constraint strategy in ontology matching. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNCS, (Vol. 5061, pp. 687–696). Liu, B., Cao, S. G., Cao, D. F., Li, Q. C., Liu, H. T., & Shi, S. N. (2012). An ontology based semantic heterogeneity measurement framework for optimization in distributed data mining. In 2012 International conference on machine learning and cybernetics (ICMLC) (pp. 118–123). Loia, V., Fenza, G., De Maio, C., & Salerno, S. (2013). Hybrid methodologies to foster ontology-based knowledge management platform. In 2013 IEEE symposium on intelligent agent (IA) (pp. 36–43). Lu, Z. (2010). ontoMATCH: A probabilistic architecture for ontology matching. In Conference on information sciences and interaction sciences (ICIS), 2010 3rd international (pp. 174–180). Mao, M., & Peng, Y. (2006). Prior system: Results for oaei 2006. In Ontology matching. Mao, M., & Peng, Y. (2007). The PRIOR+: Results for OAEI campaign 2007. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. (pp. 219– 226). Mascardi, V., Ancona, D., Bordini, R., & Ricci, A. (2011). Cool-agentspeak: Enhancing agentspeak-dl agents with plan exchange and ontology services. In International conference on web intelligence and intelligent agent technology (WI-IAT), 2011 IEEE/WIC/ACM (pp. 109–116). Mascardi, V., Locoro, A., & Rosso, P. (2010). Automatic ontology matching via upper ontologies: A systematic evaluation. IEEE Transactions on Knowledge and Data Engineering, 22, 609–623. Massmann, S., Engmann, D., & Rahm, E. (2006). Coma++: Results for the ontology alignment contest oaei 2006. In P. Shvaiko, J. Euzenat, N. F. Noy, H. Stuckenschmidt, V. R. Benjamins, & M. Uschold (Eds.), Ontology matching, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/ om2006.html#MassmannER06>. Maßmann, S., Raunich, S., Aumüller, D., Arnold, P., & Rahm, E. (2011). Evolution of the COMA match system. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 49–60). Merlin, P., Sorjamaa, A., Maillet, B., & Lendasse, A. (2009). X-som and l-som: a nested approach for missing value imputation. In ESANN. URL: <http://dblp.uni- trier.de/db/conf/esann/esann2009.html#MerlinSML09>. Merlin, P., Sorjamaa, A., Maillet, B., & Lendasse, A. (2010). X-som and l-som: A double classification approach for missing value imputation. Neurocomputing, 73, 1103–1108. URL: <http://dblp.uni-trier.de/db/journals/ijon/ ijon73.html#MerlinSML10>. Mitra, P., Noy, N. F., & Jaiswal, A. R. (2005). Ontology mapping discovery with uncertainty. In Fourth international conference on the semantic web (ISWC-2005) (pp. 537–547). Nagy, M., Vargas-Vera, M., & Motta, E. (2006). Dssim-ontology mapping with uncertainty. In P. Shvaiko, J. Euzenat, N. F. Noy, H. Stuckenschmidt, V. R. Benjamins, & M. Uschold (Eds.), Ontology matching, CEUR-WS.org. URL:<http:// dblp.uni-trier.de/db/conf/semweb/om2006.html#NagyVM06>. Nagy, M., Vargas-Vera, M., & Motta, E. (2007). Dssim – managing uncertainty on the semantic web. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR- WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2007.html#Nagy VM07a>. Nagy, M., Vargas-Vera, M., & Stolarski, P. (2009). DSSim results for OAEI 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. (pp. 160–169). Nagy, M., Vargas-Vera, M., Stolarski, P., & Motta, E. (2008). Dssim results for oaei 2008. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2008.html# NagyVSM08>. Nasir, S., & Noor, N. (2010). Analysing the effectiveness of COMA++ on the mapping between traditional Malay textile (TMT) knowledge model and CIDOC CRM. In Proceedings 2010 international symposium on information technology – visual informatics, ITSim’10 1 (pp. 1–6). NCI (2013). National Cancer Institute Thesaurus. URL: <http://ncit.nci.nih.gov/>. Ngo, D., & Bellahsene, Z. (2012a). YAM++: A multi-strategy based approach for ontology matching task. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNAI (Vol. 7603, pp. 421–425). Ngo, D., & Bellahsene, Z. (2012b). Yam++ results for oaei 2012. In OM. Ngo, D., & Bellahsene, Z. (2013). Yam++ results for oaei 2013. In OM (pp. 211–218). Ngo, D., Bellahsene, Z., & Coletta, R. (2011). Yam++ results for oaei 2011. In OM. http://dblp.uni-trier.de/db/conf/swat4ls/swat4ls2011.html#Jimenez-RuizGZ11 http://dblp.uni-trier.de/db/conf/swat4ls/swat4ls2011.html#Jimenez-RuizGZ11 http://dblp.uni-trier.de/db/conf/semweb/om2011.html#Jimenez-RuizMG11 http://dblp.uni-trier.de/db/conf/semweb/om2011.html#Jimenez-RuizMG11 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0645 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0645 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0650 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0650 http://dblp.uni-trier.de/db/conf/vldb/vldb2007.html#KenscheQLL07 http://dblp.uni-trier.de/db/conf/vldb/vldb2007.html#KenscheQLL07 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0670 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0670 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0670 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0675 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0675 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0675 http://dblp.uni-trier.de/db/journals/ets/ets9.html#KiuL06 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0710 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0710 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0710 http://dblp.uni-trier.de/db/journals/ws/ws4.html#KotisVS06 http://dx.doi.org/10.1145/2487575.2487592 http://dx.doi.org/10.1145/2487575.2487592 http://dblp.uni-trier.de/db/conf/ppswr/ppswr2005.html#LambrixT05 http://dblp.uni-trier.de/db/conf/ppswr/ppswr2005.html#LambrixT05 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0740 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0740 http://dx.doi.org/10.1109/TKDE.2008.202 http://dblp.uni-trier.de/db/conf/semweb/om2006.html#LiLZT06 http://dblp.uni-trier.de/db/conf/semweb/om2006.html#LiLZT06 http://dblp.uni-trier.de/db/conf/semweb/om2007.html#LiZLT07 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0780 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0780 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0780 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0820 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0820 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0820 http://dblp.uni-trier.de/db/conf/semweb/om2006.html#MassmannER06 http://dblp.uni-trier.de/db/conf/semweb/om2006.html#MassmannER06 http://dblp.uni-trier.de/db/conf/esann/esann2009.html#MerlinSML09 http://dblp.uni-trier.de/db/conf/esann/esann2009.html#MerlinSML09 http://dblp.uni-trier.de/db/journals/ijon/ijon73.html#MerlinSML10 http://dblp.uni-trier.de/db/journals/ijon/ijon73.html#MerlinSML10 http://dblp.uni-trier.de/db/conf/semweb/om2006.html#NagyVM06 http://dblp.uni-trier.de/db/conf/semweb/om2006.html#NagyVM06 http://dblp.uni-trier.de/db/conf/semweb/om2007.html#NagyVM07a http://dblp.uni-trier.de/db/conf/semweb/om2007.html#NagyVM07a http://dblp.uni-trier.de/db/conf/semweb/om2008.html#NagyVSM08 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#NagyVSM08 http://ncit.nci.nih.gov/ 970 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 Ngo, D., Bellahsene, Z., & Todorov, K. (2013). Opening the black box of ontology matching. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNCS (Vol. 7882, pp. 16– 30). Niu, X., Wang, H., Wu, G., Qi, G., & Yu, Y. (2011). Evaluating the stability and credibility of ontology matching methods. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNCS (Vol. 6643, pp. 275–289). Noessner, J., & Niepert, M. (2010). Codi: Combinatorial optimization for data integration: Results for oaei 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL: <http:// dblp.uni-trier.de/db/conf/semweb/om2010.html#NoessnerN10> Noy, N. F., & Musen, M. A. (2003). The {PROMPT} suite: Interactive tools for ontology merging and mapping. International Journal of Human-Computer Studies, 59, 983–1024. OAEI (2013). Ontology Alignment Evaluation Initiative. URL: <http:// oaei.ontologymatching.org/>. Paulheim, H. (2012). WeSeE-Match results for OEAI 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR- WS.org. (pp. 213–219). Paulheim, H., & Hertling, S. (2013). Wesee-match results for oaei 2013. In OM (pp. 197–202). Paulheim, H., Hertling, S., & Ritze, D. (2013). Towards evaluating interactive ontology matching tools. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNCS (Vol. 7882, pp. 31–45). Pesquita, C., Stroe, C., Cruz, I. F., & Couto, F. M. (2010). Blooms on agreementmaker: results for oaei 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/ conf/semweb/om2010.html#PesquitaSCC10>. Qazvinian, V., Abolhassani, H., (Hossein), S. H. H., & Hariri, B. B. (2008). Evolutionary coincidence-based ontology mapping extraction. Expert Systems, 25, 221–236. Quix, C., Gal, A., Sagi, T., & Kensche, D. (2010). An integrated matching system GeRoMeSuite and SMB: results for OAEI 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 166–171). Quix, C., Geisler, S., Kensche, D., & Li, X. (2008). Results of GeRoMeSuite for OAEI 2008. Proceedings of the 3rd International Workshop on Ontology Matching (OM-2008). CEUR Workshop Proceedings, 431, 160–166. Quix, C., Geisler, S., Kensche, D., & Li, X. (2009). Results of geromesuite for oaei 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/ semweb/om2009.html#QuixGKL08a>. Reul, Q., & Pan, J. (2010). KOSImap: Use of description logic reasoning to align heterogeneous ontologies. In CEUR workshop proceedings 573 (pp. 489–500). Reul, Q., & Pan, J. Z. (2009). Kosimap: Ontology alignments results for oaei 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/ semweb/om2009.html#ReulP08>. Rosoiu, M. E., dos Santos, C. T., & Euzenat, J. (2011). Ontology matching benchmarks: generation and evaluation. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 1–12). Rubiolo, M., Caliusco, M., Stegmayer, G., Coronel, M., & Fabrizi, M. G. (2012). Knowledge discovery through ontology matching: An approach based on an artificial neural network model. Information Sciences, 194, 107–119. Intelligent Knowledge-Based Models and Methodologies for Complex Information Systems. Safar, B. (2007). Exploiting wordnet as background knowledge. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), Proceedings of the workshop on ontology matching (OM2007) at ISWC/ASWC2007, Busan, South Korea. Safar, B., & Reynaud, C. (2009). Alignement d’ontologies basé sur des ressources complémentaires illustration sur le système taxomap. Technique et Science Informatiques, 28, 1211–1232. URL: <http://dblp.uni-trier.de/db/journals/tsi/ tsi28.html#SafarR09>. Sánchez-Ruiz, A., Ontañón, S., González-Calero, P., & Plaza, E. (2011). Measuring similarity in description logics using refinement operators. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNAI (Vol. 6880, pp. 289–303). Schadd, F. C., & Roos, N. (2011). Maasmatch results for oaei 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2011.html#SchaddR11>. Schadd, F. C., & Roos, N. (2012a). MaasMatch results for OAEI 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. (pp. 160–167). Schadd, F. C., & Roos, N. (2012b). Maasmatch results for oaei 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/ om2012.html#SchaddR12a>. Schadd, F. C., & Roos, N. (2013). Summary of the maasmatch participation in the oaei-2013 campaign. In OM (pp. 139–145). Scharffe, F., Zamazal, O., & Fensel, D. (2013). Ontology alignment design patterns. Knowledge and Information Systems, 1–28. Scopus (2013). Scopus. URL: <http://www.scopus.com/>. Shah, G., & Syeda-Mahmood, T. (2004). Searching databases for sematically-related schemas. In Proceedings of the 27th annual international ACM SIGIR conference on research and development in information retrieval (pp. 504–505). New York, NY, USA: ACM. Shvaiko, P., & Euzenat, J. (2005). A survey of schema-based matching approaches. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNCS (Vol. 3730, pp. 146–171). Cited By (since 1996)243. Shvaiko, P., & Euzenat, J. (2013). Ontology matching: State of the art and future challenges. IEEE Transactions on Knowledge and Data Engineering, 25, 158–176. Shvaiko, P., Giunchiglia, F., & Yatskevich, M. (2009). Semantic matching with s- match. In R. D. Virgilio, F. Giunchiglia, & L. Tanca (Eds.), Semantic Web Information Management (pp. 183–202). Springer. URL: <http://dblp.uni- trier.de/db/books/collections/Virgilio2009.html#ShvaikoGY09>. SNOMED. (2013). SNOMED Clinical Terms. URL: <http://www.ihtsdo.org/ index.php?id=545>. Spiliopoulos, V., Valarakos, A.G., Vouros, G. A., & Karkaletsis, V. (2007). SEMA: Results for the ontology alignment contest OAEI 2007. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. (pp. 244–254). Spiliopoulos, V., Vouros, G. A., & Karkaletsis, V. (2010). On the discovery of subsumption relations for the alignment of ontologies. Web Semantics: Science, Services and Agents on the World Wide Web, 8, 69–88. Stoermer, H., & Rassadko, N. (2009a). Results of OKKAM feature based entity matching algorithm for instance matching contest of OAEI 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. (pp. 200–207). Stoermer, H., & Rassadko, N. (2009b). Results of okkam feature based entity matching algorithm for instance matching contest of oaei 2009. In OM. Stoermer, H., Rassadko, N., & Vaidya, N. (2010). Feature-based entity matching: The fbem model, implementation, evaluation. In CAiSE (pp. 180–193). Straccia, U., & Troncy, R. (2005a). omap: An implemented framework for automatically aligning owl ontologies. In SWAP. Straccia, U., & Troncy, R. (2005b). oMAP: Combining classifiers for aligning OWL ontologies. In Proceedings of the 6th international conference on web information systems engineering (WISE 2005), New York City, USA. (pp. 133–147). URL: <http://www.cwi.nl/media/publications/wise05troncy.pdf>. Straccia, U., & Troncy, R. (2005c). oMAP: Results of the ontology alignment contest. In Proceedings of the K-Cap workshop on integrating ontologies (IntOnt 2005), Banff, Canada (pp. 92–96). Straccia, U., & Troncy, R. (2006). Towards distributed information retrieval in the semantic web: Query reformulation using the omap framework. In ESWC (pp. 378–392). Sunna, W., & Cruz, I. F. (2007). Using the agreementmaker to align ontologies for the oaei campaign 2007. In Shvaiko, P., Euzenat, J., Giunchiglia, F., & He, B. (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/ om2007.html#SunnaC07>. Šváb-Zamazal, O., Svátek, V., & Iannone, L. (2010). Pattern-based ontology transformation service exploiting oppl and owl-api. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNAI (Vol. 6317, pp. 105–119). Szwabe, A., Misiorek, P., & Walkowiak, P. (2012). Tensor-based relational learning for ontology matching. In KES (pp. 509–518). Tan, H., Jakoniene, V., Lambrix, P., Aberg, J., & Shahmehri, N. (2006). Alignment of biomedical ontologies using life science literature. In E. G. Bremer, J. Hakenberg, E. H. Han, D. P. Berrar, & W. Dubitzky (Eds.), KDLL (pp. 1–17). Springer. URL: <http://dblp.uni-trier.de/db/conf/pakdd/kdll2006.html#TanJLAS06>. Tan, H., & Lambrix, P. (2007). Sambo results for the ontology alignment evaluation initiative 2007. In OM. Tang, J., Li, J., Liang, B., Hunag, X., Li, Y., & Wang, K. (2006). Using Bayesian decision for ontology mapping. Web Semantics: Science, Services and Agents on the World Wide Web, 4, 243–262. http://dx.doi.org/10.1016/j.websem.2006.06.001. Tang, J., Liang, B. Y., Li, J., & Wang, K. (2004). Risk minimization based ontology mapping. In C. H. Chi & K. Y. Lam (Eds.), Proceedings of the advanced workshop on content computing (pp. 469–480). Berlin, Heidelberg: Springer. http://dx.doi.org/ 10.1007/978-3-540-30483-8_58. Thanh Le, B., & Dieng-Kuntz, R. (2007). A graph-based algorithm for alignment of owl ontologies. In Proceedings of the IEEE/WIC/ACM international conference on web intelligence (pp. 466–469). Washington, DC, USA: IEEE Computer Society. http:// dx.doi.org/10.1109/WI.2007.10. URL:<http://dx.doi.org/10.1109/WI.2007.10>. Thayasivam, U., Chaudhari, T., & Doshi, P., 2012. Optima+ results for OAEI 2012. In P. Shvaiko, J. Euzenat, A. Kementsietsidis, M. Mao, N. F. Noy, & H. Stuckenschmidt (Eds.), OM, CEUR-WS.org. (pp. 181–188). Thayasivam, U., & Doshi, P. (2011). Optima results for oaei 2011. In OM. Tian, X., & Guo, Y. (2010). A cosine theorem based algorithm for similarity aggregation of ontologies. In 2010 2nd International conference on signal processing systems (ICSPS) (pp. V2-16–V2-19). Todorov, K., Geibel, P., & Kühnberger, K. U. (2010). Mining concept similarities for heterogeneous ontologies. Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) LNAI (Vol. 6171, pp. 86–100). Tomaszewski, B., & Holden, E. (2012). The geographic information science and technology and information technology bodies of knowledge: An ontological alignment. In Proceedings of the 13th annual conference on Information technology education (pp. 195–200). New York, NY, USA: ACM. Tordai, A., van Ossenbruggen, J., Schreiber, G., & Wielinga, B. (2011). Let’s agree to disagree: On the evaluation of vocabulary alignment. In Proceedings of the sixth international conference on knowledge capture (pp. 65–72). New York, NY, USA: ACM. http://dblp.uni-trier.de/db/conf/semweb/om2010.html#NoessnerN10 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#NoessnerN10 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0915 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0915 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0915 http://oaei.ontologymatching.org/ http://oaei.ontologymatching.org/ http://dblp.uni-trier.de/db/conf/semweb/om2010.html#PesquitaSCC10 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#PesquitaSCC10 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0945 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0945 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0945 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0955 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0955 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0955 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#QuixGKL08a http://dblp.uni-trier.de/db/conf/semweb/om2009.html#QuixGKL08a http://dblp.uni-trier.de/db/conf/semweb/om2009.html#ReulP08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#ReulP08 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0980 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0980 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0980 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0980 http://refhub.elsevier.com/S0957-4174(14)00514-4/h0980 http://dblp.uni-trier.de/db/journals/tsi/tsi28.html#SafarR09 http://dblp.uni-trier.de/db/journals/tsi/tsi28.html#SafarR09 http://dblp.uni-trier.de/db/conf/semweb/om2011.html#SchaddR11 http://dblp.uni-trier.de/db/conf/semweb/om2012.html#SchaddR12a http://dblp.uni-trier.de/db/conf/semweb/om2012.html#SchaddR12a http://refhub.elsevier.com/S0957-4174(14)00514-4/h1020 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1020 http://www.scopus.com/ http://refhub.elsevier.com/S0957-4174(14)00514-4/h1030 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1030 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1030 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1030 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1040 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1040 http://dblp.uni-trier.de/db/books/collections/Virgilio2009.html#ShvaikoGY09 http://dblp.uni-trier.de/db/books/collections/Virgilio2009.html#ShvaikoGY09 http://www.ihtsdo.org/index.php?id=545 http://www.ihtsdo.org/index.php?id=545 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1065 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1065 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1065 http://www.cwi.nl/media/publications/wise05troncy.pdf http://dblp.uni-trier.de/db/conf/semweb/om2007.html#SunnaC07 http://dblp.uni-trier.de/db/conf/semweb/om2007.html#SunnaC07 http://dblp.uni-trier.de/db/conf/pakdd/kdll2006.html#TanJLAS06 http://dx.doi.org/10.1016/j.websem.2006.06.001 http://dx.doi.org/10.1007/978-3-540-30483-8_58 http://dx.doi.org/10.1007/978-3-540-30483-8_58 http://dx.doi.org/10.1109/WI.2007.10 http://dx.doi.org/10.1109/WI.2007.10 http://dx.doi.org/10.1109/WI.2007.10 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1165 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1165 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1165 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1165 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1170 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1170 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1170 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1170 L. Otero-Cerdeira et al. / Expert Systems with Applications 42 (2015) 949–971 971 Trojahn, C., Quaresma, P., & Vieira, R. (2012). Exploiting majority acceptable arguments for ontology matching. International Journal of Artificial Intelligence, 8, 1–19. Tyl, P., & Loufek, J. (2009). Comp – comparing ontology matching plug-in. In Fifth international conference on next generation web services practices, 2009. NWESP’09 (pp. 44–49). Vouros, G. A., & Kotis, K. (2005). Extending hcone-merge by approximating the intended meaning of ontology concepts iteratively. In A. Gómez-Pérez & J. Euzenat (Eds.), ESWC (pp. 198–210). Springer. URL: <http://dblp.uni-trier.de/db/ conf/esws/eswc2005.html#VourosK05>. Wang, P. (2011). Lily results on SEALS platform for OAEI 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 156–162). Wang, P., & Xu, B. (2007). Lily: The results for the ontology alignment contest oaei 2007. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. URL:<http://dblp.uni-trier.de/db/conf/semweb/om2007.html#WangX07>. Wang, P., & Xu, B. (2008). Lily: Ontology alignment results for oaei 2008. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & H. Stuckenschmidt (Eds.), OM, CEUR- WS.org. URL:<http://dblp.uni-trier.de/db/conf/semweb/om2008.html#Wang X08>. Wang, P., & Xu, B. (2009). Lily: Ontology alignment results for oaei 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/ om2009.html#WangX08a>. Wang, Z., Zhang, X., Hou, L., Zhao, Y., Li, J., Qi, Y., & Tang, J. (2010). RiMOM results for OAEI 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.) OM, CEUR-WS.org. (pp. 195–202). Warin, M., & Volk, H.M. (2004). Using WordNet and semantic similarity to disambiguate an ontology. Technical Report. STOCKHOLMS UNIVERSITET Institutionen för lingvistik. Wennerberg, P. (2009). Aligning medical domain ontologies for clinical query extraction. In Proceedings of the 12th conference of the european chapter of the association for computational linguistics: student research workshop, Association for Computational Linguistics, Stroudsburg, PA, USA (pp. 79–87). WordNet (2013). WordNet. URL: <http://wordnet.princeton.edu/>. Wrigley, S. N., García-Castro, R., & Nixon, L. (2012). Semantic evaluation at large scale (seals). In Proceedings of the 21st international conference companion on World Wide Web (pp. 299–302). New York, NY, USA: ACM. Xu, P., Tao, H., Zang, T., & Wang, Y. (2008). Alignment results of sobom for oaei 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/ semweb/om2009.html#XuTZW08>. Xu, P., Wang, Y., Cheng, L., & Zang, T. (2010a). Alignment results of SOBOM for OAEI 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 203–2011). Xu, P., Wang, Y., Cheng, L., & Zang, T. (2010b). Alignment results of sobom for oaei 2010. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/ om2010.html#XuWCZ10>. Xu, P., Wang, Y., & Liu, B. (2012). A differentor-based adaptive ontology-matching approach. Journal of Information Science, 38, 459–475. Yaghlane, B. B., & Laamari, N. (2007). OWL-CM: OWL combining matcher based on belief functions theory. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. (pp. 206–218). Zargayouna, H., Safar, B., & Reynaud, C. (2007). Taxomap in the oaei 2007 alignment contest. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2007.html# ZargayounaSR07>. Zghal, S., Kachroudi, M., Yahia, S. B., & Nguifo, E. M. (2009). Oacas – ontologies alignment using composition and aggregation of similarities. In J. L. G. Dietz (Ed.), KEOD (pp. 233–238). INSTICC Press. URL:<http://dblp.uni-trier.de/db/ conf/ic3k/keod2009.html#ZghalKYN09>.. Zghal, S., Kachroudi, M., Yahia, S. B., & Nguifo, E. M. (2011). OACAS: Results for OAEI 2011. In P. Shvaiko, J. Euzenat, T. Heath, C. Quix, M. Mao, & I. F. Cruz (Eds.), OM, CEUR-WS.org. (pp. 190–196). Zghal, S., Yahia, S. B., Nguifo, E. M., & Slimani, Y. (2007). SODA: An OWL-DL based ontology matching system. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. (pp. 261–267). Zhang, H., Hu, W., & Qu, Y. (2011). Constructing virtual documents for ontology matching using mapreduce. In J. Z. Pan, H. Chen, H. G. Kim, J. Li, Z. Wu, & I. Horrocks, et al. (Eds.), JIST (pp. 48–63). Springer. Zhang, J., Lin, P., Huang, P., & Wu, G. (2011). Research on semantic web service composition based on ontology reasoning and matching. Communications in Computer and Information Science CCIS, 243, 450–457. Zhang, S., & Bodenreider, O. (2007). Hybrid alignment strategy for anatomical ontologies: Results of the 2007 ontology alignment contest. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & B. He (Eds.), OM, CEUR-WS.org. (pp. 1–11). Zhang, X., Zhong, Q., Li, J., & Tang, J. (2008). Rimom results for oaei 2008. In P. Shvaiko, J. Euzenat, F. Giunchiglia, & H. Stuckenschmidt (Eds.), OM, CEUR- WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/om2008.html#Zhang ZLT08>. Zhang, X., Zhong, Q., Shi, F., Li, J., & Tang, J. (2009). Rimom results for oaei 2009. In P. Shvaiko, J. Euzenat, F. Giunchiglia, H. Stuckenschmidt, N. F. Noy, & A. Rosenthal (Eds.), OM, CEUR-WS.org. URL: <http://dblp.uni-trier.de/db/conf/semweb/ om2009.html#ZhangZSLT08>. Zhu, J. (2012). Survey on ontology mapping. Physics Procedia, 24(Part C), 1857–1862. http://refhub.elsevier.com/S0957-4174(14)00514-4/h1175 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1175 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1175 http://dblp.uni-trier.de/db/conf/esws/eswc2005.html#VourosK05 http://dblp.uni-trier.de/db/conf/esws/eswc2005.html#VourosK05 http://dblp.uni-trier.de/db/conf/semweb/om2007.html#WangX07 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#WangX08 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#WangX08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#WangX08a http://dblp.uni-trier.de/db/conf/semweb/om2009.html#WangX08a http://wordnet.princeton.edu/ http://refhub.elsevier.com/S0957-4174(14)00514-4/h1230 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1230 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1230 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#XuTZW08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#XuTZW08 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#XuWCZ10 http://dblp.uni-trier.de/db/conf/semweb/om2010.html#XuWCZ10 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1250 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1250 http://dblp.uni-trier.de/db/conf/semweb/om2007.html#ZargayounaSR07 http://dblp.uni-trier.de/db/conf/semweb/om2007.html#ZargayounaSR07 http://dblp.uni-trier.de/db/conf/ic3k/keod2009.html#ZghalKYN09 http://dblp.uni-trier.de/db/conf/ic3k/keod2009.html#ZghalKYN09 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1280 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1280 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1280 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1285 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1285 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1285 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#ZhangZLT08 http://dblp.uni-trier.de/db/conf/semweb/om2008.html#ZhangZLT08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#ZhangZSLT08 http://dblp.uni-trier.de/db/conf/semweb/om2009.html#ZhangZSLT08 http://refhub.elsevier.com/S0957-4174(14)00514-4/h1305 Ontology matching: A literature review 1 Introduction 2 Procedures 3 Statistical results 3.1 Articles sorted by publication year 3.2 Articles sorted by data source 4 Classification 4.1 ‘Reviews’ category 4.2 ‘Matching Techniques’ category 4.3 ‘Matching Systems’ category 4.4 ’Processing Frameworks’ category 4.5 ‘Practical Applications’ category 4.6 ‘Evaluation’ category 5 Limitations of the Literature Review 6 Trends in Ontology Matching: Practitioner-oriented Survey 6.1 Participants and Survey design 6.2 Survey results 6.2.1 Background questions 6.2.2 Research field questions 6.2.3 Future challenges questions 7 Limitations of the Survey 8 Conclusions References 
work_2jzz4ykqmfbbxhw6m4crviv56i	----	Intelligent Programm Support for Dynamic Integrated Expert Systems Construction doi: 10.1016/j.procs.2016.07.426 Intelligent Programm Support for Dynamic Integrated Expert Systems Construction Galina V. Rybina1, Victor M. Rybin1, Yury M. Blokhin1, and Sergey S. Parondzhanov1 National Research Nuclear University MEPhI (Moscow Engineering Physics Institute), Kashirskoye shosse, 31, Moscow, 115409, Russian Federation galina@ailab.mephi.ru Abstract The problems of intellectualization in the development process of integrated expert systems basing on the the problem-oriented methodology and the AT-TECHNOLOGY workbench are considered. The automation of dynamic integrated expert systems construciton is in focus. The intellgient programm environment and its basic components, including standard design procedure are reviewed. The detailed description of procedure for dynamic integrated expert system construction is given. The examples of applied integrated expert system prototypes developed with described procedure are listed. Keywords: dynamic integrated expert system, problem-oriented methodology, typical design procedure, intelligent programm environment, intelligent planner, AT-TECHNOLOGY workbench 1 Introduction Trends towards integration of research in different fields of artificial intelligence had most clearly manifested at the turn of the XX and the XXI centuries and made it necessary to combine se- mantically different objects, models, methodologies, concepts and technologies. As a result, new intelligent system architectures were emerged, in particular integrated expert systems (IES), i.e. systems with scalabale architecture and expansible functions [8, 9, 10, 11, 17, 3, 2]. Problem-oriented methodology for a wide class (static, dynamic, tutoring and others) of applied IES construction was developed by Rybina G.V. [8] from Department of Cybernetics National Research Nuclear University MEPhI. Problem-oriented methodology [8] is actively used and constantly develops: about 30 applied prototypes IES for diagnosis, control, de- sign, planning and tutoring problems were constructed (detailed description can be found in [8, 9, 10, 11]). The core idea is based on the conceptual modeling of the IES architecture at all levels of the integration processes in the IES and focusing on the modeling specific types of unformalised tasks that are relevant to the technology of traditional expert systems. In the lab- oratory ”Intelligent Systems and Technologies” of Department of Cybernetics in MEPhI several Procedia Computer Science Volume 88, 2016, Pages 205–210 7th Annual International Conference on Biologically Inspired Cognitive Architectures, BICA 2016 Selection and peer-review under responsibility of the Scientific Programme Committee of BICA 2016 c© The Authors. Published by Elsevier B.V. 205 http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2016.07.426&domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2016.07.426&domain=pdf generations of software tools for the automated support of problem-oriented methodology were created [8, 9, 10, 11, 13, 15, 16, 14]. It is called AT-TECHNOLOGY workbench. A large part of the problems is linked to the high complexity phases of design and imple- mentation of the IES, as showed by practical experience of creating a series of static, dynamic, and tutoring IES through the use of problem-oriented methodology and AT-TECHNOLOGY workbench. Problem domain and human factor provide a significant impact on modelling pro- cess. Designing of dynamic IES architecture is the most difficult phase, because in the dynamic IES important place is given to the integration of methods and means of temporal information acquisition, presentation and processing with the methods and means of the outside world simulation in real time. This leads to expansion of architecture (for example, intelligent control systems [16, 5, 7, 4]), built on the concept of dynamic IES relevant subsystems adequately reflecting all the processes and laws of functioning simulated systems [6, 16], as an integral phase of building dynamic IES. Therefore, the need to develop intelligent software environment [8, 9] for the further devel- opment of problem-oriented methodology and AT-TECHNOLOGY workbench with the aim of creating intelligent technology to build specific classes of IES has become urgent. Now let us review basic features of problem-oriented methodology for IES construction, which is described in details in [8]. 2 Features of Problem-Oriented Methodology for Inte- grated Expert Systems Construction In the context of solving the modern IES construction problems (in particular for the manage- ment of complex discrete systems) problem-oriented methodology has the following properties [8, 9, 10, 11]: a powerful combination method of acquiring knowledge that supports the au- tomated process of acquiring knowledge from the sources of knowledge of different typology (experts, database, text) is used to gain knowledge; generalized knowledge representation lan- guage designed for building models of problem areas in dynamic IES allows to represent tempo- ral knowledge, based on a modified interval Allen logic [1] and time control logic, together with the basic knowledge, including those containing knowledge with uncertainty, imprecision and vagueness; supports the use of various output means (universal AT-SOLVER and a specialized temporal solver designed for dynamic tasks); in the context of enhanced functionality and princi- ples of the components IES deep integration provides the possibility of implementing simulation techniques for modeling the external environment and how to interact with them; the high effi- ciency of the a large number of applied IES development, including dynamic areas of concern; instrumentally supported by a modern software such as workbench (AT-TECHNOLOGY). Significant place in the framework of the problem-oriented methodology for IES construct- ing (basic points are reflected in [8]) is given to the methods and means of intelligent software support for the development processes, which form general concept of ”intellectual environ- ment”. Complete formal description of the intellectual environment model and methods of the individual components implementation is presented in [8], so here only a brief description of the model in the form of quaternion is presented. MAT = 〈KB, K, P, TI〉 (1) KB is a technological knowledge base (KB) on the composition of the project, and typical design solutions used in development of IES. K = {Ki}, i = 1..m - set of current contexts Ki, Intelligent Programm Support Rybina, Rybin, Blokhin and Parondzhanov 206 consisting of a set of objects from the KB, editing or implementing on the current control step. P a special program - an intelligent planner that manages the development and IES testing process. TI = {TIi}, i = 1..n - many tools TIi, applied at various stages of IES development. A component of the KB is a declarative basis of intellectual support for the development of IES, acting as data storage in a given environment and defined as KB = 〈WKB, CKB, PKB〉 (2) WKB is a KB containing knowledge of the standard design procedures (SDP), describing the sequences and methods of using various tools to create applied IES and a sequence of steps for creating IES. CKB - is KB comprising knowledge about the use of SDP and reusable components (RUC), including fragments of previously created IES prototypes. PKB (optional) - is a KB containing specific knowledge used at various stages of creating IES prototype for solving problems that require innovative approaches. The current context Ki is represented as set of Ki = 〈KD, KP〉. KD here is a declarative context for storing static declarative information about the structure of the project, the knowl- edge engineer and the current user. KP is a procedural context, which includes objects clearly affecting the further planner steps (LC system phase, currently edited or executable object, the current target, the current executor, the global development plan, etc.). The main procedural (operational) component is intelligent planner. This model generally describes it. P = 〈SK, AF, Pa, Pb, I, GP〉 (3) SK here is the state of the current context, in which the planner was activated. AF = {AFi}, i = 1..k is a set of functional modules AFi, a part of planner. Pa is a selection procedure for the current target based on the global development plan. Pb is a selection procedure for the best executive function module from the list of possible candidates. I - procedures to ensure the interface with the corresponding components of the AT-TECHNOLOGY workbench; GP - operating procedures for the IES global development plan. Each reusable component (RUC), involved in the development of an IES prototype, is represented by tuple: RUC = 〈N, Arg, F, PINT, FN〉 (4) N in this model is the name of the component, by which it is registered in the workbench. Arg = {Argi}, i = 1..l - set of arguments containing current project database subtree serving the input parameters for the functions from the set. F = {Fi}, i = 1..s - a variety of methods (RUC interfaces) for this component at the implementation level. PINT - a set of other kinds of RUC interfaces, used by the methods of the RUC. FN = {FNi}, i = 1..v - set of functions names performed by this RUC. Any SDP can be represented as tuple SDP = 〈C, L, T〉 (5) , where C - is the set of conditions under which the SDP can be implemented; L - script implementation described in the describing internal language actions of the SDP ; T - set of parameters initialized by intelligent planner at SDP inclusion in the development plan of a IES prototype. Today there are already implemented and used following SDPs: ”Tutoring web-IES con- struction”, ”Distributed Knowledge Acquisition” and others. And currently researching SDP ”Dynamic IES construction” is described below. Intelligent Programm Support Rybina, Rybin, Blokhin and Parondzhanov 207 3 Description of Typical Design Procedure for Dynamic Integrated Expert Systems Construction Described above SDP model is specified as SDPD = 〈CD, LD, TD〉 for dynamic IES construc- tion, where CD - is a set of conditions to initialize SDPD realization; LD - scenario of dynamic IES construction; TD - a set of parameters initialized by intelligent planner when SDPD is included into dynamic IES construction plan. Conditions set CD. Following set of conditions must be satisfied to include SDPD into development plan: • following lifecycle stage is system requirements analysis; • in dynamic IES architecture model (as a hierarchy of EDFD) there is an element, describ- ing simulation model presence (here, complex technical systems are simulated); • there is at least one EDFD element, connected with solving of non-formalized problem. Scenario LD. There are following SDPD scenario stages: 1. System equirment analisys stage. Following actions are performed: automated knowledge acquisition based on temporal knowledge acquisition method [8], which per- forms direct acquisition process of knowledge, containing temporal references; simulation model development with help of specialized visual editor. 2. Design stage. Here are performed following actions: forming of reasoning tools (a synergy of universal AT-SOLVER and Temporal Solver is currently supported); conversa- tions of previously obtained knowledge field into knowledge base described in generalized knowledge representation language; explanation development; IES core components con- figuration. 3. Implementation stage. During the final stage of dynamic IES prototype construction following steps are performed: simulation visual representation development; development of components for communication with user with language for dialogue scenario descrip- tion;integration with external systems (DB, applied software modules, etc.); aggregate integration of IES components. Parameters set TD. In case of including SDPD into development plan there are two parameters included into current context: one identifies currently executed SDP, the other one - current scenario step. Now, using [12], let us review features of dynamic IES prototyping with SDPD, which functions are supported by a set of operational RUC in each IES construction stage. Let us assume, that knowledge engineer already built dynamic IES architecture model as a EDFD hierarchy. With help of AT-SOLVER in automated mode initial architecture layout of IES prototype is formed as it is shown in Fig. 1. Global plan is also generated with intelligent planner core. Detailed development plans generation is performed iteratively, and following selection and execution of the most priority task is performed with following modification of architecture layout. The process is looped, until all tasks are not done. Detailed plan is regenerated in it is necessary. Because it is usually assumed, that IES construction planning problem actions are performed with deterministic result, so deterministic planning approach is used as base [12]. It is performed Intelligent Programm Support Rybina, Rybin, Blokhin and Parondzhanov 208 Technological KB Architecture model in the form of EDFD hierarchy Initial architecture layout General plan of IES prototyping Detailed plan generation Planning task execution Synthesis of a new architecture layout fragment Current IES project RUC RUC RUC SDP SDP SDP Universal AT-SOLVER PDDL planner Knowledge Engineer Figure 1: Dynamic IES construction planning scheme in project state space, formed by sets of possible values of project parameters. Built EDFD hierarchy is analyzed with intelligent planner in following manner. Elements which have to be implemented are recognized, for example simulation model, DB fragments, non-formalized tasks, etc. These elements represent big tasks and precedence relation between them is built, in particular with use of PDDL-planners. Initial architecture layout generating and global IES development plan generating is per- formed once after EDFD hierarchy is build. Then detail development plan is generated iter- atively with help of intelligent planner after each architecture layout fragment adding. De- tailed plan is visualized with visualization component, and knowledge engineer can initialize specific plan task execution. Each plan task is performed with use of appropriate RUC of AT-TECHNOLOGY workbench. Samples of another complex SDPs, for example, connected with tutoring IES development are described in [8, 9, 10, 11]. The difficulties of the tutoring IES development technology are caused by supporting two different work modes DesignTime, oriented to work with teachers (course/discipline ontology creating processes, different typed training im-pacts creating, etc.) and RunTime, for working with students (current student model building processes, including psychological model, etc.). 4 Conclusion Currently, an experimental software study of the current version of the intelligent planner during educational IES prototyping on various courses is carried out. This study is carried out in particular, for the collective development of IES prototypes with limited resources. Research and development related to the use of described SDP ”Dynamic IES construciton” with intelligent software environment for two dynamic IES prototypes (”Management of med- Intelligent Programm Support Rybina, Rybin, Blokhin and Parondzhanov 209 ical forces and resources for major traffic accidents” and ”Resource management for satellite communications system between regional centers”) and tutoring IES prototypes for different courses/disciplines were succesfully performed. 4.1 Acknowledgments This work was supported by the Russian Foundation for Basic Research under grant no. 15-01- 04696 and Competitiveness Growth Program of the Federal Autonomous Educational Institu- tion of Higher Professional Education National Research Nuclear University MEPhI (Moscow Engineering Physics Institute). References [1] James Allen. Maintaining knowledge about temporal intervals. Communications of the ACM, 26(11):832–843, 1983. [2] B.E. Fedunov. Intelligent systems for the core of anthropocentric objects and its modeling. pages 394–400, 2012. [3] Crina Grosan and Ajith Abraham. Intelligent Systems - A Modern Approach, volume 17 of Intel- ligent Systems Reference Library. Springer, 2011. [4] I. M. Makarov, V. M. Lokhin, S. V. Manko, and M. P. Romanov. Artificial intelligence and intelligent control systems. Nauka, Moscow, 2006. [5] G. S. Osipov. Artificial intelligence methods. FIZMATLIT, Moscow, 2011. [6] M. Pidd. Computer simulation in Management Science. Wiley, 5th ed. Chichester, 2004. [7] V.M. Rybin and S.S. Parondzhanov. Using of dynamic integrated expert systems for intelligent control. International Journal of Applied Engineering Research, 10(13):33202–33205, 2015. [8] G. V. Rybina. Theory and technology of construction of integrated expert systems. Monography. Nauchtehlitizdat, Moscow, 2008. [9] G. V. Rybina. Intelligent systems: from A to Z. Monography series in 3 books. Vol. 1. Knowledge- based systems. Integrated expert systems. Nauchtehlitizdat, Moscow, 2014. [10] G. V. Rybina. Intelligent systems: from A to Z. Monography series in 3 books. Vol. 2. Intelligent dialogue systems. Dynamic intelligent systems. Nauchtehlitizdat, Moscow, 2015. [11] G. V. Rybina. Intelligent systems: from A to Z. Monography series in 3 books. Vol. 3. Problem- oriented intelligent systems. Tools for intelligent system developing. Dynamic intelligent systems. Nauchtehlitizdat, Moscow, 2015. [12] G. V. Rybina and Y. M. Blokhin. Modern automated planning methods and tools and their use for control of process of integrated expert systems construction. Artificial intelligence and decision making, (1):75–93, 2015. [13] G. V. Rybina and A. O. Deineko. Distributed knowledge acquisition for the automatic construction of integrated expert systems. Scientific and Technical Information Processing, 38:1–7, 2011. [14] G.V. Rybina and Y.M. Blokhin. Methods and means of intellectual planning: Implementation of the management of process control in the construction of an integrated expert system. Scientific and Technical Information Processing, 42(6):432–447, 2015. [15] G.V. Rybina and A.V. Mozgachev. The use of temporal inferences in dynamic integrated expert systems. Scientific and Technical Information Processing, 41(6):390–399, 2014. [16] G.V. Rybina, V.M. Rybin, S.S. Parondzhanov, and S.T.H. Aung. Some aspects of simulation application in dynamic integrated expert systems. Life Science Journal, 11(SPEC. ISSUE 8):144– 149, 2014. [17] A. R. Tyler. Expert systems research trends. Nova Science Publishers, New York, 2007. Intelligent Programm Support Rybina, Rybin, Blokhin and Parondzhanov 210 
work_2kjbpm2btjfo5fuxi67ouhrveu	----	paper.dvi Applying KADS to KADS� knowledge based guidance for knowledge engineering John K�C� Kingston AIAI�TR���� January ���� This paper was published in the journal �Expert Systems The International Journal of Knowledge Engineering � ��� �� February ���� Arti cial Intelligence Applications Institute University of Edinburgh �� South Bridge Edinburgh EH� �HN United Kingdom c� The University of Edinburgh� ����� Abstract The KADS methodology �Schreiber et al� ����� �Tansley � Hayball� ����� and its successor� CommonKADS �Wielinga et al� ��� � have proved to be very useful approaches for modelling the various transformations involved be� tween eliciting knowledge from an expert and encoding this knowledge in a computer program� These transformations are represented in a series of mod� els� While it is widely agreed that these methods are excellent approaches from a theoretical viewpoint� the documentation provided concentrates on de ning what models should be produced� with only general guidance on how the models should be produced� This has the advantage of making KADS and CommonKADS widely applicable� but it also means that considerable training and experience is required to become pro cient in them� This paper reviews three projects� which investigated the feasibility of producing speci c guidance for certain decisions which are required when using KADS or CommonKADS to develop a knowledge based system� Guid� ance was produced for the identi cation of the generic task addressed by a knowledge based system� for the selection of appropriate AI techniques for implementing the analysed knowledge� and for selecting a suitable tool for implementing the system� Each set of guidance was encoded in its own knowledge based system� which was itself developed with the assistance of KADS or CommonKADS� These projects therefore both studied and applied KADS and CommonKADS in order to produce knowledge based guidance for knowledge engineers� The projects showed that it was feasible to produce heuristic guidance which could be understood� applied� and occasionally overridden by knowl� edge engineers� The guidance provides reasonably experienced knowledge engineers with a framework for making the key decisions required by Com� monKADS� in the same way that CommonKADS provides knowledge engi� neers with a framework for representing knowledge� The projects also pro� duced some new insights about CommonKADS domain modelling and about the process of task identi cation� � Introduction The KADS methodology �Schreiber et al� ����� �Tansley � Hayball� ����� and its successor� CommonKADS �Wielinga et al� ������ are collections of structured meth� ods for building knowledge based systems� analagous to methods such as SSADM for software engineering� The development of these methods was funded by the European Community�s ESPRIT programme between ���� and ����� KADS and CommonKADS view the construction of KBS as a modelling activity� and so these methods require a number of models to be constructed which represent di�erent views on problem solving behaviour� in its organisational and application context� CommonKADS recommends the construction of six models � � a model of the organisational function and structure� � a model of the tasks required to perform a particular operation� � a model of the capabilities required of the agents who perform that operation� � a model of the communication required between agents during the operation� � a model of the expertise required to perform the operation� which is divided into three sub�levels � models of declarative knowledge about the domain� � models of the inference processes required during problem solving� � an ordering of the inference processes� � a model of the design of a KBS to perform all or part of this operation� For more details on these models� see �deHoog et al� ������ Experience has shown that the models recommended by KADS and Com� monKADS provide an excellent basis for representing the various transformations required between eliciting knowledge from an expert and encoding it in a com� puter program� In addition� KADS and CommonKADS provide various libraries of generic models which have proved very useful to knowledge engineers �see �Kingston� ������ for example�� However� experience has also shown that the task of developing these models is non�trivial� There are a number of decisions to be taken at each stage in the modelling process� some of which have a major impact on the models� or on the implemented system� These decisions include �� Deciding the approach which should be taken to modelling should models be produced bottom�up from acquired knowledge� top�down by instantiating generic problem�solving methods which CommonKADS provides� or by an intermediate approach which selects a generic model and then modi es it in the light of domain knowledge� �� Selecting models from a library� The library of generic inference structures is indexed according to the type of task which is being tackled so the knowledge engineer must determine the most appropriate task type for the current task� �� Deciding whether the design should preserve the structure of the expertise model� �� Deciding which knowledge representations and inference techniques should be used within the design� � �� Deciding on the most appropriate tool for implementing this knowledge based system� Much of the CommonKADS documentation concentrates on de ning what should be done in order to produce models� while only specifying how it should be done in general terms� This has the advantage that it speci es the content of mod� els without enforcing an approach on practitioners� thus making CommonKADS widely applicable and compatible with many di�erent approaches to knowledge engineering� however� by the same token� it requires knowledge engineers to be fairly experienced at making good knowledge engineering decisions before they can make full use of KADS or CommonKADS� The CommonKADS project has pro� vided guidance on making some of the decisions outlined above �e�g� guidance on top�down�bottom�up approaches to model construction �Wielinga� ����� and model instantiation �L�ockenho� � Valente� ������� however� many of the organisa� tions which have recognised the value of KADS and CommonKADS have had to obtain training and accept a long learning curve for each of their sta� who wants to become pro cient in the KADS approach� An alternative approach to providing extensive training and experience for all sta� would be to provide speci c guidance on developing KADS models� thus pro� viding in�house KADS guidance based on the company�s own knowledge engineering experiences� The purpose of this paper is to review three projects which investi� gated the feasibility of producing guidance for some of the important decisions listed above� The tasks for which guidance was produced were � identifying the task type� � selecting appropriate AI techniques at the design stage�� � choosing a suitable implementation tool� These projects were all performed by students in the Department of Arti cial Intelligence� University of Edinburgh as part of an M�Sc course in Intelligent Knowl� edge Based Systems� The students were supervised by sta� from the Department of Arti cial Intelligence� and by members of the Knowledge Engineering Methods group at AIAI� the supervisors also acted as experts from whom knowledge was elicited� It is a requirement that all students on Edinburgh�s M�Sc course produce func� tioning software as part of their project� which meant that the students were re� quired to acquire knowledge� analyse the knowledge� and to implement this knowl� edge in a knowledge�based system� It therefore seemed sensible for the students to use KADS to help them in this process� since practical experience of using �Guidelines on whether a design should preserve the structure of an expertise model are cur� rently being considered� � KADS ought to help them in producing useful guidance� The main body of this paper therefore shows how KADS was used to model the tasks involved in making KADS�related decisions in these three projects� � Identifying task types The rst project looked at the task of identifying the most appropriate task type for a particular problem �Krueger� ������ ��� The task The student who took on this project was set the task of developing a technique for distinguishing and classifying di�erent expert tasks� with a focus on the tax� onomy of expert tasks developed by the KADS methodology �see Figure ��� This taxonomy de nes the contents of the library of generic inference structures� there is a generic inference structure associated with �almost� every leaf node in the tax� onomy� The selection of a suitable generic inference structure therefore boils down to the identi cation of the most appropriate task type from this taxonomy� SYSTEM MODIFICATION SYSTEM SYNTHESIS SYSTEM ANALYSIS Prediction Identifying Repairing Controlling Remedying Modelling Transformation Design Planning Configuration Refinement design Transformational design Monitoring Classification Prediction of values Prediction of behaviour Maintaining Assessment Diagnosis Simple classification Multiple fault diagnosis Single fault diagnosis Localisation Causal Tracing Systematic Diagnosis Heuristic Classification Multiple stream refinemt design Single stream refinemt design Exploration- based design Hierarchical design Figure �� Taxonomy of task types � ��� The project Given that the student has been presented with a problem to solve using knowledge� based techniques� the rst stage of KADS analysis is to identify the type of task which is carried out in order to select a generic inference structure� Fortunately� the obvious �chicken and egg situation which arose here was circumvented by reading an early KADS report �Breuker�� ������ which suggests that the appropriate task type here is assessment � that is� assessing how well each of the task types matches the task under consideration� and then selecting the one which matches most closely� The student therefore designed and implemented a KBS which used an assessment approach to the identi cation of task types� KADS� generic inference structure for assessment tasks �Figure �� recommends that assessment is carried out by abstracting �i�e� generalising� key features of a particular problem case description� and specifying the preferred value of these features from a system model which represents the �ideal world � The features of the case are then matched against the preferred features to determine the degree of acceptability of the case� For this task� the �problem case is an expert task� and the �ideal world is �one of� a set of generic inference structures� The generic inference structure for assessment tasks was therefore adapted and instantiated in the ways described below the resulting inference structure is shown in Figure �� Case description System model Norms Decision class Abstract case description match abstract specify Figure �� Generic inference structure for assessment tasks In order to instantiate this generic structure to problem in hand� the student was required to make some alterations to the generic inference structure � The key features of an expert task were obtained by asking the user� rather than by abstracting from a detailed description� The abstraction inference was therefore replaced by an obtain step�� �In CommonKADS� obtain is considerd to be a transfer task rather than an inference step� Transfer tasks are represented by rounded rectangles� � � The set of inference structures goes through a two�stage speci cation the rst stage determines the features which must be asked of the user� and the second stage speci es parameters of these features which can be compared against the user�s replies� � A �feedback from the di�erences discovered to the set of inference structures under consideration is explicitly recorded in the problem�speci c inference structure� user’s knowledge about expert task differences parameters expected (with weightings) features of structures under consideration set of inference structures under consideration user-supplied parameters refine compare specify-2 specify-1 obtain Figure �� Instantiated inference structure for assessing inference structures The resulting system� known as SEXTANT�� asks users to identify inputs of � Selection of EXpert TAsks by Nature of the Task � a task� outputs of a task� and knowledge roles which exist in the problem solving process� it then matches this information against the corresponding attributes of each generic inference structure in order to attach a likelihood weighting to each inference structure� As weightings increase or decrease� some inference structures are removed from consideration until only one �or a few� are left� The remaining inference structure�s� are then recommended to the user� As the project progressed� it became obvious that an alternative technique for identifying task types could be developed� based on the taxonomy of task types shown in Figure �� By starting at the topmost node in the taxonomy� it is possible to traverse the taxonomy by asking a single question at each node to decide which is the most appropriate subcategory for the current problem� For example� at the topmost level� the following question could be asked � Does the task involve �� Establishing unknown properties or behaviour of an object within the domain� �� Composing a new structural description of a possible object within the domain� �� A combination of the above� If the rst answer is chosen� then the most appropriate subcategory is System Analysis tasks� if the second� then System Synthesis� if the last� then System Mod� i�cation is the most appropriate task type� By de ning suitable questions for each non�leaf node in the taxonomy� it becomes possible to identify task types simply by answering a series of questions� E�ectively� the taxonomy is transformed into a decision tree� and the identi cation of task types becomes a comparatively simple classi cation task� The nal version of the SEXTANT system incorporates both the �decision tree approach and the �assessment approach� Novice users of KADS can progress through the decision tree� however� if they are unable to answer questions in the decision tree� the assessment approach takes over� using the remaining candidates from the decision tree as the set of inference structures on which assessment is performed� Experienced users of KADS should be able to specify a relatively small set of task types at the outset� and so will only need to use the assessment part of the SEXTANT system in order to support them in their nal decision� The use of SEXTANT also forces knowledge engineers to consider the nature and the dependencies of each inference step carefully� which should provide assistance in the task of instantiating the chosen generic inference structure to the actual task being performed� SEXTANT was implemented in CLIPS ���� and is therefore capable of running on a range of hardware� � ��� Results This project demonstrated that it was feasible to provide guidance on task selec� tion� using either an assessment�based approach or a relatively simple decision tree approach� The creation of appropriate questions for the decision tree required con� siderable thought about the nature of the choices being made at each choice point� The questions which were generated were reasonably consistent in their format� which suggests that the tasks in the library form a coherent set� It is not clear that the tasks form a complete set of all possible expert tasks� however� and even those tasks which are in the library do not always have an associated generic inference structure� It is likely that more work is needed on these tasks types � particularly system synthesis and system modi cation tasks � in order to produce a complete set of knowledge based tasks �cf� �Tansley � Hayball� ����� �Kingston� ������� Since this project was performed� the CommonKADS project has rede ned generic inference structures� by simplifying the basic structures in the library and providing extensive guidance on con guring the generic structures to the require� ments and features of a particular task �L�ockenho� � Valente� ������ This alter� ation has greatly increased the scope for de ning many expert tasks� as variations of a smaller number of �basic tasks � It has also reduced the utility of the �as� sessment approach in SEXTANT� because the di�erences between the simpli ed inference structures in the library now require less detailed analysis� and because the con guration process requires users to think deeply about the nature of the inferences performed in the task� However� SEXTANT�s �decision tree approach is still useful� indeed� it could usefully be extended to incorporate guidance on con� guring inference structures� since this guidance consists of a set of questions which are used to recommend particular alterations to a basic inference structure� These �con guring questions could therefore be considered to be extensions to the lowest levels of the decision tree� The �decision tree component of SEXTANT has been used on other KADS� related projects� including the projects described later in this paper� � Selecting appropriate AI techniques The second project was intended to help knowledge engineers select appropriate knowledge representation and inference techniques when constructing a KBS design �MacNee� ������ ��� The task Once a task type has been identi ed� the acquired knowledge can be analysed by building the KADS expertise model� This consists of � � an inference structure� con gured and instantiated to represent the reasoning processes which take place in the task under consideration� � a set of domain models� identifying key concepts in the domain and the rela� tionships between them� � a task structure� enforcing an ordering on the inference steps�� The resulting expertise model must then be transformed into a program speci� cation� this is the task of the KADS design phase� The approach recommended for the CommonKADS design phase prescribes a ��stage design process� and includes suggested approaches for modelling the rst two phases �van de Velde et al� ����� � Application design typically consists of a conceptual decomposition of the expertise model into a number of functional units and�or data objects� � Architecture design requires decisions on whether to use rules� objects� or other representational techniques� for di�erent parts of the application design� � Platform design matches these chosen techniques with the facilities and environment o�ered by the chosen programming tool� with consequent alter� ations to the architecture design and�or the programming tool� This approach has been used successfully on some projects� and appears to represent a su�ciently detailed breakdown of the design process�� however� unless a strongly prescriptive top�down approach to modelling is being used� there is little speci c guidance on the design process� The aim of the project described in this section was to elicit and acquire some guidelines for the production of an architectural design� The approach used was based on the probing questions approach� developed at Rome Air Force Base� USA �Kline � Dolins� ����� and further developed at AIAI �Inder et al� ������ in which a KBS designer is asked a number of questions about the analysed knowledge� with the answers being used to produce some recommendations of suitable representa� tional techniques� The general format of the questions is if a certain feature exists in the analysed knowledge then consider using a particular knowledge representation� or imple� mentation technique� The designer is asked whether the if condition is true if it is� the recommen� dation supplied by the latter part of the �probing question is added to the set of recommendations� An example of a probing question might be �See �Wielinga et al� ���� for a fuller description of the expertise model� �These � phases approximately correspond to the stages of functional decomposition� be� havioural design and physical design speci ed in the original KADS project� � if the problem�solving task is such that a pre�enumerated set of so� lutions can be established �as distinct from the type of task in which solutions are constructed as a result of the satisfaction of constraints� then use goal�driven reasoning weighting� � else use data�driven reasoning weighting� � It can be seen that probing questions are e�ectively heuristics for knowledge engineers� encoded in a rule�based format� The student�s task was to acquire prob� ing questions �either by eliciting knowledge from experienced knowledge engineers� or by reading the available literature�� to collate the resulting collection of heuris� tic rules into a structure of some kind� and then to implement a knowledge based system to run these rules� ��� The project� acquisition and analysis The student chose to concentrate his e�orts on eliciting knowledge from experts� Three experts were used� each of whom was interviewed �or asked to make extensive comments on documents sent by electronic mail� on several occasions� Knowledge elicitation techniques used included introductory interviews� card sorting� and tri� adic comparison of actual KBS systems� One of the key results of knowledge elicitation was the identi cation of a number of functional requirements and design features for knowledge base systems� Many functional requirements relate to the KBS� need to model the domain objects and their relationships� and the need to model the inferences which are made about these domain objects� These requirements may lead to various needs for example� there may be requirements for certain knowledge representations� for the ability to represent uncertainty in knowledge� and for the ability to generate and examine large numbers of solutions� Design features are knowledge representation and in� ference techniques used to satisfy the functional requirements� Examples include data�driven reasoning� rules� semantic networks� and blackboard architectures� There is also a sizeable group of functional requirements� which are concerned with the capabilities of the KBS environment and with producing a smooth inter� action between the user and the KBS� These requirements may specify a need for dialogue with and explanations to the user� a need for a high level of computational e�ciency in the program� or a need to consider how �or whether� to present the user with large numbers of possible solutions� It was therefore decided that these envirnomental features would also be considered as design features� The categorisation of functional requirements and design features can be seen in Figure �� Note that �thoughts of God is intended to be a catch�all category for ill�de ned subjective in uences� such as design for elegance� or parsimonious design� �� expert knowledge Knowledge of KBS environment and user/KBS interaction (0) ‘THOUGHTS OF GOD’ (1a) KNOWLEDGE STRUCTURE -- (a): domain objects and relationships (1b) KNOWLEDGE STRUCTURE -- (b): inferences and generic task (2) UNCERTAINTY IN KNOWLEDGE (3) SOLUTIONS (4) DATA (5) DIALOGUE AND EXPLANATION (6) COMPUTATIONAL EFFICIENCY (7) DEVELOPMENT: Analysed construction, maintenance, and CATEGORIES OF DESIGN FEATURE expansion (8) SAFETY CRITICALITY (A) DEPTH-OF-REASONING AND ARCHITECTURE -- shallow, model-based, blackboard (B) KNOWLEDGE REPRESENTATION STRUCTURES -- rules, frames, OOP, networks, ... (C) INFERENCE TYPES -- goal-driven, data-driven (D) CONTROL OF FLOW OF INFERENCE: search strategy, constraints on search, meta-control (E) HANDLING OF UNCERTAINTY IN KNOWLEDGE AND OF INCOMPLETENESS AND INACCURACY OF DATA (F) USER INTERFACE: KBS input/output (G) KNOWLEDGE ACQUISITION CATEGORIES OF FUNCTIONAL REQUIREMENT Figure �� Categories of functional requirements and design features The next stage was to create a KADS expertise model to represent all the ac� quired knowledge from the di�erent viewpoints speci ed by KADS� In this project� the domain level of the model of expertise was developed rst� it was decided that the domain level consisted of the various functional requirements and design fea� tures� Once the domain level had been completed� the inference structure was developed� It was decided that the task type being performed was heuristic classi� �cation� The generic inference structure for heuristic classi cation and the instan� tiated inference structure which forms part of the expertise model can be seen in Figure �� It can be seen that the level of abstraction of the probing questions varies considerably� the conditions might be based on general knowledge of the operation of KBS systems� or on speci c knowledge of the task under consideration� and the recommendations might be to use a particular design feature� or to use one of a category of design features� Finally� a task structure was developed� which speci ed that the KBS should continue to ask questions until all possible relevant questions had been asked� in order to ensure that all functional requirements are considered� �� DESIGN FEATURE MATCH ANALYSED EXPERT KNOWLEDGE KNOWLEDGE OF KBS ENVIRONMENT CLASS FEATURE DESIGN REFINE FUNCTIONAL REQUIREMENTS Inference structure -- adapted and instantiated AND INTERACTION MATCH SOLUTION ABSTRACTION SOLUTION ABSTRACTION PROBLEM ABSTRACT PROBLEM REFINE Inference structure -- generic: heuristic classification Figure �� Generic and instantiated inference structures for a heuristic classi cation task ��� The project� design As the design is a relatively late stage in the development of the KBS� the student was able to apply the elicited probing questions to his own design problem� The re� sults of this exercise can be seen in an appendix to the project thesis �MacNee� ������ in which the nal implemented system was �retrospectively� applied to itself� The main resulting recommendations are summarised in the following table �� Design feature Recommendation Shallow reasoning Strong Rules Strong Goal driven reasoning Moderate Depth rst search Moderate Truth maintenance Moderate Certainty factors Moderate !Canned� text for explanations Moderate Data driven reasoning Weak Model�based reasoning Strong negative It is hardly surprising that a knowledge base which is based around heuristic probing questions should elicit strong recommendations for shallow reasoning and for the use of rules� The justi cation for some of the other recommendations is less obvious� this was noted during the project� and it was therefore decided that the implemented system would need to be able to justify its reasoning to the user in a clear and coherent manner�� An example of an explanation� based on the table above� is that the recommendation for depth� rst search arose because of the need to ask questions of the user� and because the student stated that the natural ow of dialogue was to ask detailed questions about one subject before asking general questions about another subject� Depth� rst search supports this mode of reasoning� and therefore makes it easy to ask questions in a natural order� ��� The project� platform design � implementation Having performed architectural design� the nal stage of design must be performed the matching up of the recommended design features with the chosen programming tool� The decisions required at this stage are described in more detail in section �� but for now� it is su�cient to note that the the probing questions should be consid� ered as a starting point for tool selection� not as a prescription� For this project� a choice of two programming tools was available CLIPS and Prolog� The table above indicates a stronger recommendation for goal�driven reasoning than for data� driven reasoning� which would favour Prolog� however� bearing in mind that the probing questions are heuristics� the factors contributing to this recommendation were examined in more detail� It turned out that the recommendation for goal� driven reasoning was based on a combination of three weak contributing factors� and one of these factors is actually not applicable in this case� because the system is designed to make the KBS ask all possible questions� instead of stopping when a single solution has been found� The recommendation for goal�driven reasoning was therefore downgraded to !weak�� �This decision accounts for the recommendation for �canned� text in the above table� �� CLIPS and Prolog are both equally strong in most of the other recommended de� sign features �although CLIPS does provide its own facilities for truth maintenance� which Prolog does not�� the conclusion is that� from the viewpoint of providing ad� equate design features� CLIPS and Prolog are both equally recommended for this project� The choice of tool was therefore heavily in uenced by other factors� such as the student�s previous experience� The tool chosen was CLIPS� a largely rule�based tool whose primary reasoning mechanism is depth� rst forward chaining� CLIPS also provides facilities for truth maintenance and certainty factors� Using CLIPS� the PDQ system �which� in this context� stands for �Probing Design Questions � was implemented in approximately � weeks� ��� Results PDQ is a workable system which produces recommendations which are helpful� although heuristic� as the above example concerning goal�driven reasoning shows� The questions require the knowledge engineer to have a good understanding of the problem� and some preliminary ideas about possible designs� for example� one question asks if the problem �can be subdivided into �� or more distinct modules � This suggests that PDQ is most approriate for knowledge engineers who have some experience of building knowledge based systems� Some further work has been done on the set of probing questions since the PDQ system was completed� in an attempt to separate those questions which produce abstract recommendations from those which recommend more speci c design fea� tures� This process has lead to the transferral of a small set of probing questions to the rst stage of the design process �application design�� because some design features �such as blackboard architectures or constraint�based programming� have such a profound e�ect on the architecture of a program that they must be con� sidered as alternative ways of performing the initial decomposition of the analysed knowledge� The probing questions are therefore able to provide some guidance on whether to perform a structure�preserving design� or whether to make use of an es� tablished AI paradigm� This decision is another of the key decisions for knowledge engineers listed in the introduction to this paper� It is possible that the �prob� ing questions technique could be extended to provide extensive support for this important decision� � Choosing a suitable implementation tool The nal project described in this paper aimed to produce guidance on the se� lection of a suitable shell or toolkit for implementing a knowledge based system �Robertson� ������ �� ��� The task The two projects described above have shown that knowledge�based guidance can be provided for certain areas of KADS modelling� However� the knowledge engineer�s decision making does not end when the �probing questions have been answered and the KADS design has been completed� as section ��� suggests� the choice of a programming tool requires careful consideration� taking into account both the recommendations of the probing questions and other factors external to the knowledge representation requirements� The requirements for this project were that the student should identify factors important to the selection of a KBS building tool� and develop a program which would recommend appropriate tools for a project� Given that the PDQ system had already been developed� it seemed sensible to use PDQ�s recommendations as a starting point for the identi cation of important factors� The student also referred to a number of books on knowledge engineering �e�g� �Price� ������ which gave their own advice on tool selection� ��� The project� acquisition Knowledge acquisition for this project was initially performed by reading various textbooks� and examining the output of the PDQ system� The results of this work were compiled� and represented graphically using a simple node and arc represen� tation� These diagrams� and associated text� were used as input to a knowledge elicitation interview with an experienced knowledge engineer� This interview pro� duced further useful information� which was used to alter and extend the diagrams� Finally� the student was directed to investigate a set of expert system building tools which was representative of all the categories which he had identi ed� The major result of the knowledge acquisition was two classi cations of KBS building tools � The rst classi cation divided tools into shells� toolkits and AI languages� These categories were further subdivided� shells were divided into those which evaluate rules by pattern matching �e�g� Xi Plus� OPS�� early versions of CLIPS� and those which are e�ectively �rule networks �e�g� Crystal��� Toolkits were subdivided into top�range toolkits �such as ART and KEE� and mid�range toolkits �such as Nexpert Object� ProKappa and Kappa PC�� Languages were not subdivided� since there are comparatively few popular languages for implementing knowledge based systems� � The second classi cation was based on the historical background of the tools� tools were classi ed as belonging to the �ART Camp �e�g� ART�IM� CLIPS� �This distinction was rst de ned in �Inder et al� ��� � �� ECLIPSE� or the �KEE Camp �including ProKappa and Kappa PC�� Tools in the same !camp� have similar features to each other� since many of them are derived from similar programming philosophies or tools� for example� all tools in the ART camp o�er e�cient forward chaining rules based on an implementation of the RETE algorithm� but are comparatively weak on backward chaining� because the algorithm used o�ers almost no support for backward chaining� whereas all tools in the KEE camp have good support for object�oriented programming� but comparatively ine�cient rules�� The other main conclusion from the knowledge acquisition phase �largely drawn from �Rothenberg� ������ was that the importance of particular features of a pro� gramming tool is dependent on the phase of the project� For example� at the very early stages of the project� rapid prototyping might be used to support knowledge acquisition� to investigate the feature of the tool� or to help the knowledge engineer learn to use the tool� at this stage� the training and debugging features of the tool are very important� However� when the nal system is delivered� the training and debugging tools are of little or no importance� whereas the speed and e�ciency of the execution of the program become very important� The student identi ed ve phases of program development �exploration� prototyping� development� elding and operation� and seven features of a tool �ease of use� e�ciency� extendability� exibility� portability� reliability and support� which are more or less desirable at various phases� For further details of the proposed relationship between phases of development and tool features� see �Rothenberg� ����� or �Robertson� ������ ��� The project� analysis � � � Domain level analysis The domain level of the expertise model comprised the classi cations identi ed in the knowledge acquisition phase� plus a selection of factors which in uence the choice of tool �such as the cost of the tool� the required platform for development� and the experience of programmers with the tool�� It also de ned a prototype !frame�� with a number of attributes �or� in the terminology of CommonKADS� a concept� with a number of properties� for de ning tools� This frame proved remarkably di�cult to de ne� because of di�culties in distinguishing properties of tools from tool�related concepts� For example� there was considerable discussion on whether the frame should have �forward chaining as a property �with values such as rete� procedural or none� or whether it should have �reasoning types as a multiple�valued property� with forward chaining being one of the values of �See �Mettrey� ���� for some benchmarking tests� Note� however� that the vendors of KAPPA� PC claim considerable performance improvements in newer versions of KAPPA�PC� which have appeared since Mettrey�s study was done� �� this property� This discussion led to an important obaservation concerning KADS� which is described in section ���� � � � Inference level analysis When the inference level is considered� it can be seen that this project is similar in structure to the PDQ project PDQ identi ed functional requirements which had to be mapped to design features� whereas this project uses various design features and other requirements to select a suitable tool� or class of tool� However� while the task of PDQ was to identify all relevant design features� the task required when selecting a tool is to select the best tool for the task� This requires comparison of a set of possible tools against all relevant factors� and assessment of how well each tool matches the overall set of criteria� The most appropriate task type in the KADS taxonomy is therefore assessment� rather than heuristic classi cation� By the time this project was carried out� some guidance had been published by members of the CommonKADS consortium �L�ockenho� � Valente� ����� on the con guration of inference structures to match particular assessment tasks� Starting with a minimal generic inference structure �Figure ��� this guidance consists of a set of questions which are asked of the user in order to decide which of the inference steps relevant to assessment tasks are actually performed in this task� The results of these questions are used to devise a problem�speci c version of the inference structure �e�g� Figure ��� which is then instantiated to the problem in hand by changing the generic labels on the knowledge roles into problem�speci c labels� The nal result �Figure �� forms the inference level of the model of expertise� A worked example of the con guration of an inference structure can be found in �Kingston� ������ System ModelCase Description Decision Class Match Cases Figure � Minimal generic inference structure for assessment tasks �� Abstract Case Description Measurement System Decision Classes Measure Case Set of Norms Specify Set of Norms Specify Conflict Resolution Conflict Resolution Criteria Resolve Conflicts Decision ClassDecision Class Resolve Conflicts Conflict Resolution Criteria Specify Conflict Resolution Specify Set of Norms Set of NormsMeasure Case Decision Classes Measurement System Abstract Case Description Abstract Case Description Measurement System Decision Classes Measure Case Set of Norms Specify Set of Norms Specify Conflict Resolution Conflict Resolution Criteria Resolve Conflicts Decision ClassDecision Class Resolve Conflicts Conflict Resolution Criteria Specify Conflict Resolution Specify Set of Norms Set of NormsMeasure Case Decision Classes Measurement System Abstract Case Description Figure �� Con gured inference structure for selecting a KBS tool �� Desired Product Features Available Products Measure Case Specify Set of Norms Available Product Features Chosen Products Conflict Resolution Criteria Specify Conflict Resolution Resolve Conflicts Ideal Product Figure �� Instantiated inference structure for selecting a KBS tool It can be seen that the task of selecting a KBS tool requires comparison of the features of available tools against the features desired for a particular project� this produces a shortlist of suggested tools� which is then further re ned to choose the ideal tool for the job� ��� The project� design and implementation The con gured inference structure contributed greatly to the quick and accurate development of an expertise model� Once the model was complete� design was per� formed� with the assistance of the PDQ system� PDQ produced very strong recom� mendations for using both rules �to represent the applicability of di�erent factors� �� and frames �to represent individual tools�� it also produced a strong recommen� dation for goal�driven reasoning� and a moderate recommendation for data�driven reasoning� At this point� the student was o�ered the choice of two tools CLIPS ��� or Sic� stus Prolog� The choice was restricted to two tools because these tools were easily available� and because the student had some familiarity with both these tools� the latter factor was important because of the tight timescales of the project� The lim� ited choice simpli ed the decision process for reasons described below� Prolog was chosen as the most appropriate tool� The system was therefore implemented using Prolog� using a goal�driven approach which investigated several di�erent categories of tool features one by one� The details of fteen di�erent KBS tools were also implemented� using a predicate called �frame which was de ned to allow the rep� resentation of object�like structures� The features of each tool were then matched against the features required by the user� the relative importance of each feature in the overall assessment was scaled according to the phase of development for which the tool would be used� and according to an importance value entered by the user� The nal list of tools� ordered by their overall score� was then displayed to the user� In order to test the system� it was retrospectively applied to the choice between CLIPS and Prolog for the student�s own project� The results of the consultation showed that Sicstus Prolog and CLIPS were among the more highly favoured tools� although they were not at the top of the list� however� most of the tools which were preferred were considerably more expensive� The main factors which led to Prolog being preferred were the recommendation for goal�driven reasoning from the probing questions� the problems in integrating rules and objects in version � of CLIPS� � and the student�s greater degree of familiarity with Prolog� which led the system to believe that features which were unavailable in Prolog �such as frames� could easily be programmed� It therefore seems that the choice of Prolog was the best decision from the available options� ��� Results This system can be considered to be a sophisticated prototype� Its reasoning ca� pabilities are wide�ranging �over many di�erent tool features and aspects of KBS development� and its algorithm for assessing tools provides results which closely resemble the opinions of expert knowledge engineers� in short� the heart of the sys� tem is su�ciently good to be commercially usable� It might nd favour as a tool for organisations to assess the adequacy of their existing tools for a proposed project� rather than to recommend a particular too� for a project which is already under Version � of CLIPS included a full object�oriented programming system� but these objects could not be pattern matched in the conditions of rules� Version � of CLIPS has introduced this facility� �� way� However� its knowledge base needs to be extended to include more tools� and it would bene t from improvements to its user interface and e�ciency� The use of KADS modelling and the probing questions has led to a knowledge base architecture which is well supported by documentation and justi cation more importantly it can be extended easily� because new tools can be added to the knowledge base simply by de ning a new !frame�� Ease of maintenance is a vital feature for a system which provides expertise about a domain which changes as rapidly as this one� � Discussion The three projects described above have produced knowledge based systems which provide guidance on particular decisions which users of CommonKADS have to make� The SEXTANT system assists knowledge engineers in the early stages of knowledge analysis� when they are selecting a suitable generic inference structure from CommonKADS� library of inference structures� PDQ provides suggestions for the architectural design of a knowledge based system� and the tool selection system provides guidance on matching the recommendeddesign with available shells or toolkits� The guidance which has been produced seems to be useful PDQ in particular has been used by later generations of students� and on commercial projects� It is important to realise that these systems are not intended to take over the decision�making role of a knowledge engineer� The guidance which these systems provide is heuristic� in some cases� careful analysis might suggest that the guidance should not be followed �see section ��� for an example�� The guidance also requires an understanding of knowledge engineering terminology� as well as an in�depth understanding of the problem to be solved� in order to make the best use of the advice� The key bene t of the guidance is in assisting knowledge engineers by providing a framework for performing the required analysis� just as CommonKADS provides a framework for representing knowledge� It does this by identifying the questions which need to be asked in order to make a particular decision� as well as providing suggested answers to those questions and �in some cases� justi cation for those answers� The projects also produced new insights about the modelling techniques rec� ommended by KADS and CommonKADS� The students found that using KADS helped them to think clearly about their knowledge bases and to identify speci c areas where problems occurred� However� KADS and CommonKADS are intended to assist rather than to replace the judgement of a knowledge engineer� the students all discovered that the KADS approach did not provide a �magic wand solution to all the problems which they encountered� For example� the SEXTANT project originally considered the choice of an appropriate task type to be an assessment �� task� but it was later discovered that a simpler approach� based on a decision tree� could be used to accomplish most of the job� Another example can be seen in the domain modelling for the tool selection project� where the student had di�culty in deciding whether �forward chaining should be a property or a value of a property� The crux of the problem was that neither approach could be ruled incorrect at a theoretical level� the choice between them depended on purely pragmatic con� straints� such as the number of possible values of the property� and the number of possible attributes of the concept� In other words� the de nition of concepts and properties in a domain appears to be context�dependent� This observation has far�reaching implications for CommonKADS domain modelling� in that it suggests that there is more than one !correct� model of any chosen domain of knowledge� There is therefore scope for further research on CommonKADS� identifying di�er� ent styles of domain modelling and the circumstances in which di�erent approaches are most appropriate� In terms of a single project� the choices of classi cations at the domain level are unlikely to a�ect the functionality of the nal system very much� although they may a�ect the ease of achieving that functionality� � Summary In summary� CommonKADS provides a declarative framework for representing knowledge� which can be used to promote clearer thinking and structuring of a knowledge base� CommonKADS is most appropriate for knowledge engineers who have enough experience to be able to make their own decisions about knowledge analysis or design in those exceptional circumstances when CommonKADS� rec� ommendations prove to be unhelpful� CommonKADS itself is su�ciently complex that knowledge engineers require experience with and�or guidance on the KADS approach to use it to its full extent� The projects described in this paper provide guidance for some of the key decisions which must be made when using Com� monKADS� This guidance is aimed at those people who would be capable of using CommonKADS it o�ers heuristic advice� which is normally good advice but may need to be overridden in exceptional circumstances� The results of these projects suggest that it is feasible to produce knowledge� based guidance for users of KADS or CommonKADS� as well as producing some new insights about domain modelling and task selection� It is hoped that the results of the projects above can be amalgamated with other projects which are producing guidance of CommonKADS �for tasks such as con guring inference structures �L�ockenho� � Valente� ����� and converting CommonKADS� Concep� tual Modelling Language to its Formal Modelling Language �Aben� ������� in order to make CommonKADS a more powerful tool for structuring knowledge engineer� ing� �� Acknowledgements I would like to thank Dr Dave Robertson and Dr Mandy Haggith of the Department of Arti cial Intelligence� University of Edinburgh� Ian Filby of AIAI� and Dr Robert Inder of the Human Communications Research Centre� University of Edinburgh for providing the students with supervision and expertise� I would also like to thank Dr Robertson� Dr Haggith and Robert Rae of AIAI for their comments on drafts of this paper� References �Aben� ����� Aben� M� ������� Formal methods in Knowl� edge Engineering� Unpublished Ph�D� thesis� SWI� University of Amsterdam� The relevant chapter is also available as CommonKADS report KADS� II�T����WP�UvA��������� �Breuker�� ����� Breuker�� J� A� ������� Model�driven Knowledge Ac� quisition� University of Amsterdam and STL� ES� PRIT project ����� Deliverable A�� �deHoog et al� ����� de Hoog� R�� Martil� R�� Wielinga� B�� Taylor� R�� Bright� C� and van de Velde� W� ������� The Common KADS model set� ESPRIT Project P���� KADS�II KADS�II�M��DM���b�UvA����� ���� University of Amsterdam and others� �Inder et al� ����� Inder� R�� � Aylett� R�� Bental� D�� Lydiard� T� and Rae� R� �October ������ Study on the Evaluation of Expert Systems Tools for Ground Segment Infras� tructure Final Report� Technical Report AIAI�TR� ��� AIAI� �Kingston� ����� Kingston� J�K�C� ������� Re�engineering IMPRESS and X�MATE using CommonKADS� In Expert Sys� tems ��� British Computer Society� Cambridge Uni� versity Press� Also available as AIAI�TR����� �Kingston� ����� Kingston� J�K�C� ������� Design by Exploration A Proposed CommonKADS Inference Structure� Sub� mitted to �Knowledge Acquisition�� �� �Kline � Dolins� ����� Kline� P�J� and Dolins� S�B� ������� Designing ex� pert systems � a guide to selecting implementation techniques� Wiley� �Krueger� ����� Krueger� A� �Sept ������ Classi cation of Expert Tasks the SEXTANT system� Unpublished M�Sc� thesis� Dept of Arti cial Intelligence� University of Edinburgh� �L�ockenho� � Valente� ����� L�ockenho�� C� and Valente� A� �March ������ A Li� brary of Assessment Modelling Components� In Pro� ceedings of �rd European KADS User Group Meeting� pages �������� Siemens� Munich� �MacNee� ����� MacNee� C� �Sept ������ PDQ A knowledge�based system to help knowledge�based system designers to select knowledge representation and inference tech� niques� Unpublished M�Sc� thesis� Dept of Arti cial Intelligence� University of Edinburgh� �Mettrey� ����� Mettrey� W� �February ������ Expert systems and tools Myths and realities� IEEE Expert� �Price� ����� Price� C� ������� Knowledge Engineering Toolkits� Ellis Horwood� This book concentrates on available toolkits� and on how to go about selecting an appro� priate toolkit� While some of the tools it mentions are now somewhat dated� the principles of tool se� lection are still valid� It also gives clear examples of implementation� �Robertson� ����� Robertson� S� �Sept ������ A KBS to advise on se� lection of KBS tools� Unpublished M�Sc� thesis� Dept of Arti cial Intelligence� University of Edinburgh� �Rothenberg� ����� Rothenberg� J� ������� Expert System Tool Evalu� ation� In Guida� G� and Tasso� C�� �eds��� Topics in Expert System Design� North�Holland� �Schreiber et al� ����� Schreiber� A� Th�� Wielinga� B� J� and Breuker� J� A�� �eds��� ������� KADS� A Principled Approach to Knowledge�Based System Development� Academic Press� London� �� �Tansley � Hayball� ����� Tansley� D�S�W� and Hayball� C�C� ������� Knowledge�Based Systems Analysis and Design� A KADS Developers Handbook� Prentice Hall� �van de Velde et al� ����� van de Velde� W�� Duursma� C� and Schreiber� G� ������� Design model and process� Unpublished KADS�II�M��MM�VUB� Vrije Universiteit Brussel� �Wielinga� ����� Wielinga� B� �October ������ Expertise Model Model De nition Document� CommonKADS Project Report� University of Amsterdam� KADS� II�M��UvA��������� �Wielinga et al� ����� Wielinga� B�� Van de Velde� W�� Schreiber� G� and Akkermans� H� ������� The KADS Knowledge Mod� elling Approach� In Proceedings of the Japanese Knowledge Acquisition Workshop JKAW�� �� �Wielinga et al� ����� Wielinga� B�� van de Velde� W�� Schreiber� G� and Akkermans� H� �Jun ������ Expertise Model Def� inition Document� CommonKADS Project Report� University of Amsterdam� �� 
work_2l47wd4umbgptlckvrzotnbl3q	----	1 Transforming Standalone Expert Systems into a Community of Cooperating Agents N. R. Jennings, L. Z. Varga, Dept. of Electronic Engineering, Queen Mary and Westfield College, University of London, Mile End Road, London E1 4NS, UK. email: n.r.jennings@qmw.ac.uk l.varga@qmw.ac.uk R. P. Aarnts, Volmac Nederland B. V., Daltonlaan 300, 3584 BJ Utrecht, The Netherlands. email: rob_aarnts@eurokom.ie J. Fuchs1 and P. Skarek PS Division, CERN, 1211 Geneve 23, Switzerland. email:joachim@wgs.estec.esa.nl (Fuchs); pskpsk@cernvm.cern.ch (Skarek) ABSTRACT Distributed Artificial Intelligence (DAI) systems in which multiple problem solving agents cooperate to achieve a common objective is a rapidly emerging and promising technology. However, as yet, there have been relatively few reported cases of such systems being employed to tackle real-world problems in realistic domains. One of the reasons for this is that DAI researchers have given virtually no consideration to the process of incorporating pre-existing systems into a community of cooperating agents. Yet reuse is a primary consideration for any organisation with a large software base. To redress the balance, this paper reports on an experiment undertaken at the CERN laboratories in which two pre-existing and standalone expert systems for diagnosing faults in a particle accelerator were transformed into a community of cooperating agents. The experiences and insights gained during this process provide a valuable first step towards satisfying the needs of potential users of DAI technology - identifying the types of changes required for cooperative problem solving, quantifying the effort involved in transforming standalone systems to ones suitable for cooperation and highlighting the benefits of a cooperating systems approach in a realistic industrial application. KEYWORDS: Distributed Artificial Intelligence, Multi-Agent Systems, Cooperating Expert Systems, Fault Diagnosis, Particle Accelerator Control 1. Now working at esa/estec, WGS, Keplerlaan 1, NL-2200 AS Noordwijk, The Netherlands 2 INTRODUCTION As computer hardware and software becomes increasingly powerful, so applications which used to be considered beyond the scope of automation come into reach. To cope with these increased demands, software systems are becoming correspondingly larger and more complex. However the problems encountered in building these large systems are not simply scaled up versions of those faced when constructing small ones. Since the late 1960’s, when the “software crisis” was first noted, it has been realised that large systems require radically different techniques and methods. One paradigm for overcoming the complexity barrier is to build systems of smaller more manageable components which can communicate and cooperate1,2,3. Such a Distributed Artificial Intelligence (DAI) approach has several potential advantages. Firstly, divide and conquer has long been championed as a means of constructing large systems because it limits the scope of each processor. The reduced size of the input domain means the complexity of the computation is lower, thus enabling the components to be simpler and more reliable. Secondly, decomposition aids problem conceptualisation; many tasks appear difficult because of their sheer size. Other benefits include reusability of problem solving components, greater robustness in the case of component failure, speed up due to parallel execution, enhanced problem solving due to the combination of multiple paradigms and sources of information and finally increased system modularity1,3. In DAI systems, individual problem solving entities are called agents; agents are grouped together to form communities which cooperate to achieve the goals of the individuals and of the system as a whole. It is assumed that each agent is capable of a range of useful problem solving activities in its own right, has its own aims and objectives and can communicate with others. The ability to solve some problems alone (coarse granularity4) distinguishes components of DAI systems from the components of neural systems in which individual nodes have very simple states (either on or off) and only by combining many thousands of them can problem solving expertise be recognised. 3 Agents in a community usually have problem solving expertise which is related, but distinct, and which frequently has to be combined to solve problems. Such joint work is needed because of the dependencies between agents’ actions, the necessity to meet global constraints and the fact that often no one individual has sufficient competence to solve the entire problem alone. There are two main causes of such interdependence (adapted from Davis and Smith5). Firstly, when problem partitioning yields components which cannot be solved in isolation. In speech recognition, for example, it is possible to segment an utterance and work on each component in isolation, but the amount of progress which can be made on each segment is limited. Allowing the sharing of hypotheses is a far more effective approach6. Secondly, even if subproblems are solvable in isolation, it may be impossible to synthesize their results because the solutions are incompatible or because they violate global constraints. For instance when constructing a house, many subproblems are highly interdependent (eg determining the size and location of rooms, wiring, plumbing, etc.). Each is solvable independently, but conflicts which arise when the solutions are collected are likely to be so severe that no amount of work can make them compatible. It is also unlikely that global constraints (eg total cost less than £70,000) would be satisfied. In such cases, compatible solutions can only be developed by having interaction and agreements between the agents during problem solving. It is the need for significant amounts of cooperation to achieve tasks and the relative autonomy of the agents to determine their own activities which distinguishes DAI from more conventional distributed systems work. Despite these potential advantages, there are still relatively few DAI applications working on realistic problems in real world domains7. One of the reasons for this is the mismatch between the needs of organisations who require solutions to their problems and the research objectives of the DAI community. Typically organisations which have problems amenable to a cooperating systems approach already possess computer systems in which they have invested substantial amounts of time and money. Naturally enough they want the return on this investment to be maximised, meaning the systems should be utilised until they become obsolete or significantly better alternatives become available. However, most work in DAI assumes that the agents have been 4 purpose built using tools and techniques designed solely for cooperative problem solving. Virtually no consideration is given to the process of incorporating pre-existing systems into a cooperating community - yet this must be a central concern if DAI is to leave the laboratory and progress to real applications. To help redress the balance, this paper reports on an experiment carried out at the CERN laboratory, under the auspices of the ARCHON project8,9, in which two standalone and pre- existing expert systems for diagnosing faults in a particle accelerator were transformed into a community of cooperating agents. The problems faced during this process and the insights which emerged are recounted as an important first step towards tackling the larger issue of providing a methodology for describing how pre-existing systems can be incorporated into a cooperating community. The framework used to facilitate the cooperative problem solving was GRATE10,11 which is described more fully in the following section. This paper also indicates the types of social interaction which can be expected between large, coarse grained agents - an important indication of the appropriateness of theoretical research into techniques for cooperation and coordination for real-size industrial applications. Finally because DAI (and AI in general) techniques are so rarely applied to real applications, this experiment offers valuable insights into the problems and constraints encountered during this process - such issues often fail to emerge when idealised problems or simulated environments are used12. GRATE: A GENERAL FRAMEWORK FOR COOPERATIVE PROBLEM SOLVING GRATE (Generic Rules and Agent model Testbed Environment) is a general framework for constructing communities of cooperating agents for industrial applications. GRATE agents have two major components - a cooperation and control layer and a domain level system (see figure 1). The domain level system solves problems such as detecting and diagnosing faults, proposing remedial activities and checking the validity of operator actions. These problems are expressed as tasks - atomic units of processing when viewed from the cooperation and control layer. The cooperation layer is a meta-level controller which operates on the domain level system; its 5 objective is to ensure that the agent’s activities are coordinated with those of others within the community. It decides which tasks should be performed locally, determines when social activity is appropriate, receives requests for cooperation from other community members, and so on. GRATE’s clear delineation of domain problem solving and knowledge related to cooperation and control has several advantages. Firstly, it increases software reusability in that the cooperation layer can be deployed in multiple applications without having to disentangle the knowledge used to guide social activity from that used to solve domain level problems. Secondly, the domain and cooperation layers can be developed independently provided that they respect the interface Acquaintance Models Self Model SITUATION ASSESSMENT MODULE COOPERATION MODULE Information Store CONTROL DATA Task 1 Task 2 Task n Inter-agent communication CONTROL MODULE Communication Manager Domain Level System Cooperation & Control Layer INTERFACE Figure 1: Detailed GRATE Agent Architecture 6 definition. This is especially important with respect to pre-existing software because it places very few restrictions or constraints on the types of system which can be incorporated. Thus the domain level systems may be written on different hardware and software platforms, use different knowledge representation formalisms and have different control regimes13. By providing a standard interface to the domain level system, all of the underlying heterogeneity can be masked; enabling the cooperation layer to be used in a wide range of applications. The division also simplifies the domain level system because it can continue to act on the basis of local information; it need not be concerned with the activities of other agents because any influence they have on its behaviour will be exerted through the cooperation layer. The disadvantage of the delineation, with respect to pre-existing systems, is that the control regime and some interface interpretation commands may need to be written to enable the cooperation layer to exact the appropriate control - this issue is discussed further in the section describing the implementation details. Also for systems which are purpose built for cooperative problem solving in a particular environment, this architecture may involve creating an artificial barrier in its local control regime if the interface definition is too rigid. In GRATE communities control is completely distributed, there is no hierarchy and all agents are equal. Agents have a degree of autonomy in generating new activities and in deciding which tasks to perform next. A global controller would have been the easiest way of ensuring the cooperating community acted in a coherent way; with its knowledge of the goals, actions and interactions of all community members the controller could have ensured: that misleading and distracting information was not spread, that multiple agents did not compete for unshareable resources simultaneously, that agents did not unwittingly undo the results of each others activities and that the same actions were not carried out redundantly. However because of the complexity of most industrial applications this approach was deemed inappropriate because: • Bandwidth limitations make it impossible for agents to be constantly informed of all developments in the system5. 7 • A local perspective simplifies conceptualisation and implementation. Problems become more complex if agents have to monitor the activities of all the others - the theory of bounded rationality14. • The controller would become a severe communication and computational bottleneck and would cause the whole system to collapse if it failed15. GRATE’s cooperation and control layer has three main problem solving modules. Each module is implemented as a separate forward-chaining production system with its own inference engine and local working memory. Communication between the modules is via message passing. The rules built into each module are generic, they are applicable for controlling activities in a broad class of industrial applications. Some credence to this claim is given by the fact that GRATE has been applied to the domain of electricity transportation management10 and detecting overloads in a telecommunication network16, as well as the problem of diagnosing faults in a particle accelerator which is reported here. A more comprehensive description of GRATE can be found in reference 17. The control module is GRATE’s interface to the domain level system and is responsible for managing all interactions with it. This interaction is controlled through the following set of primitives: • From GRATE to the domain level system: <START task inputs> <STOP task> <SUSPEND task> <KILL task> <REACTIVATE task> <MORE-INFO-AVAILABLE task info> • From the domain level system to GRATE: <DIRECTIVE-FAILED directive> <RESULTS-PRODUCED task> <DIRECTIVE-EXECUTED directive> <FINISHED task> 8 The situation assessment module makes decisions which affect both of the other modules. It decides which activities should be performed locally and which should be delegated, which requests for cooperation should be honoured, how requests should be realised, what actions should be taken as a result of freshly arriving information and so on. It issues instructions to, and receives feedback from, the other modules. Typical requests to the cooperation module include “get information X” and “send out information Y to interested acquaintances”. Requests to the control module are of the form “stop task T1” and “start task T2”. The cooperation module is responsible for managing the agent’s social activities. The need for such activity is detected by the situation assessment module, but its realisation is left to this module. Three primary objectives related to the agent’s role in a social problem solving context are supported. Firstly, the cooperation module has to establish new social interactions (eg find an agent capable of supplying a desired piece of information). Secondly, the module has to maintain cooperative activity once it has been established, tracking its progress until successful completion (eg sending out relevant intermediate and final results to interested agents). Finally the module has to respond to cooperative initiations from other agents. The cooperation layer’s other components provide support for the activities of the main problem solving modules. The information store provides a repository for all the data which the underlying domain level system has generated or which has been received as a result of interaction with others. Acquaintance and self models are representations of other agents and of the local domain level system respectively. They describe the agent’s current problem solving state, the tasks it is able to solve, the goals it is working towards and so on - a fuller description is given in reference 10 and an illustration of their use in the CERN experiment is given in the next section. The agent models and the information store contain all the domain dependent data needed at the cooperation and control layer - thus enabling the rules of the problem solving modules to be application independent. Agents communicate with one another via a message passing paradigm. This form of 9 communication has several advantages over a shared memory approach (such as a blackboard18,19). Firstly, message passing has well understood semantics and offers a more abstract means of communication20. No hidden interactions can occur; so there is greater comprehensibility, reliability and control over access rights. Secondly, message passing makes fewer assumptions about system architecture. Finally, shared memory systems do not easily scale up. If only a single blackboard exists then it becomes a severe bottleneck and if several exist the semantics revert to message passing21. DIAGNOSING FAULTS IN A PARTICLE ACCELERATOR This section describes the diagnosis problem of accelerator operation, details two pre-existing expert systems used for this task which are running at the CERN laboratories and outlines the potential benefits of cooperation in this application. Particle Accelerator Operation The CERN Proton Synchrotron (PS) accelerators are one of the world’s most sophisticated high energy research tools. The PS complex is at the heart of CERN’s experimental facilities acting as an injector for the larger accelerators - the Super Proton Synchrotron and the huge Large Electron Positron rings. In the PS, nuclear particles are focused into particle beams, accelerated, and directed through several linear and ring accelerators using electromagnetic fields. These beams are then used in the physicists’ experiments. Different experiments require different types of beam, the variations are provided by the accelerator’s different operational modes. The PS complex is controlled by a team of operators who maintain the beam performance and the operational modes of the accelerators. Accelerator operation is a task demanding technical competence, experience, diagnostic skill and judgement. The operator has to handle information coming from control system modules, accelerator components and the acceleration process itself. Observations are made via measurement devices and range from simple status displays of specific accelerator components to complicated application programs and graphical information. The 10 operator can change setpoint control values for the different components directly or he can use application programs which present the beam properties on a higher conceptual level. Different sections of the accelerator are controlled from different consoles in the same control room. In the present system if there is some suspicion about a problem in an overlapping area, operators communicate by talking to their colleagues on other consoles. The operator’s workload is constantly increasing as new accelerators and different operational modes are added to the system. To cope with this increase, automated tools are becoming an integral part of the control process. CERN has already equipped their control room with several supporting tools - ranging from simple ones that display status information to high level software based on Expert System technology22,23,24,25,26. This experiment on cooperative problem solving concentrated on two of the high level tools, BEDES and CODES, which employ expert systems technology for diagnosing faults in the accelerator. BEDES and CODES The main goal of BEDES (BEam Diagnostic Expert System) is to diagnose operational faults at the beam level for the PS Booster injection part of the PS complex. Operational faults occur, for example, if the intensity of the particle beam falls below a certain level or if the beam deviates considerably from its ideal trajectory. Such problems can be caused by the incorrect setting of a control parameter (e.g. wrong timing is set for a switching magnet), a breakdown in a controller (e.g. the switching magnet is not working), or an error in the control system (e.g. a module that controls the switching magnet is down). BEDES can diagnose the first two types of fault if the underlying control system is still working correctly. If such faults are detected, BEDES tries to recover from them by resetting the correct control value or by optimizing a control value respectively. CODES (COntrol Diagnostic Expert System) has a similar control structure to BEDES, the main difference is in the domain knowledge. Whilst BEDES works on the beam level, CODES 11 operates on the level of the accelerator’s control system. This control system consists of thousands of hardware and software modules in a large computer network and is used by the operator to ensure that the accelerator process works successfully. A fault in the control system usually manifests itself in terms of a deviation of the particle beam and so the problem will be picked up by BEDES. In such cases BEDES and CODES could work together to determine the source of the fault. BEDES could help to detect whether the problem is caused by wrong parameter settings or a breakdown of a controller, while CODES could determine whether the fault is caused by the control system itself. Preparing for Cooperative Problem Solving When the particle accelerator is running, the operator receives a myriad of information from which he has to make an overall judgement of the situation and take reasonable decisions. As humans are resource-bounded, the operator is only able to use a limited number of tools and correlate very few pieces of information in real time. Therefore assistance is required in producing a consistent interpretation of the information. At present BEDES and CODES report information independently to the operator who then has to translate it to a common domain, determine whether it is consistent and decide upon the appropriate course of action. If the tools were capable of exchanging information directly, they would be capable of producing a consistent view automatically, leaving the operator to concentrate on the more cognitive activities (eg interpreting and acting upon diverse sources of information, tweaking the system to enhance performance, etc.) for which he is better suited. From here stems the motivation for introducing cooperation between the expert systems27,28. BEDES and CODES have knowledge about the diagnosis of the same particle accelerator from different perspectives. In certain situations, faults can be identified or even recovered from using only one of the systems, but in many instances contributions from both of them are needed. By exchanging intermediate and final results, the expert systems are able to focus each other’s problem solving activity on promising areas and draw each other away from unprofitable avenues 12 of reasoning. Initially cooperation between BEDES and CODES was studied by means of several paper exercises. Later practical exercises started and the expert systems were enhanced with an application-specific cooperation software which sent hypotheses directly from BEDES to CODES29. These preliminary studies identified some key design issues which needed to be addressed, these include: how are local actions performed in one expert system?, what is the common language of the expert systems?, how does one expert system model the other expert system? and how does one expert system model itself? Each of these questions are addressed in turn before the GRATE experiment is described in detail. Local Actions The main unit of reasoning for both BEDES and CODES is the hypothesis. Hypotheses are stored on an agenda; the status of the agenda determines whether the expert system is active or just idling. At the beginning of each inference cycle the agenda is rearranged (see figure 2), which means that hypotheses are realigned according to their priority and any which have become obsolete are removed. The first hypothesis is then taken from the agenda, it is evaluated and possibly more detailed descendant hypotheses are created and injected into the agenda. Evaluation requires that data describing the current status is gathered from the accelerator’s control system (e.g. currently valid control values); reasoning about this data is then undertaken, the outcome of which is a change in state of the hypothesis being evaluated (e.g. from unconfirmed to confirmed). If the evaluated hypothesis is confirmed, the fault has been found and diagnosis stops - the results are then reported to the operator. 13 From this structure it is apparent that the natural unit of local activity is the basic inference cycle and that local action can be controlled through operations on the agenda (eg to stop diagnosis the agenda should be cleared and to focus on a promising hypothesis it should be moved to the beginning of the agenda). Using this strategy local and cooperative actions can be kept separate and well organised, cooperative features can be added to the expert systems without significantly modifying their existing reasoning mechanisms. Common Language As both BEDES and CODES represent their hypotheses using a similar structure, it was decided to use this as the basis of a common language between the two agents. Hypotheses are assertions together with accompanying knowledge about how to prove them. The assertion is about an element or parameter of the accelerator which might be in an incorrect state or could go wrong in the near future. The related knowledge is composed of the necessary inference steps to prove the assertion and has two parts: procedural steps including data acquisition (eg read values from the Agenda Empty? Retrieve Data from Control System Take First Hypothesis Wait Re-Arrange Agenda Create and Inject Hypotheses Evaluate Hypothesis ACTIVE IDLE Yes No Recover from Error Report ResultConfirmed NOT.Confirmed Figure 2: Expert System Control Loop 14 control system, filter uncertainty and discrepancies and compute the derived parameter) and declarative rules operating on the structural description of the diagnosed system stored in the knowledge base of the expert system (eg is the value close enough to “ideal”?, if not then create derived hypotheses for those parts that can cause the deviation). Hypotheses are implemented as frames. Although the structure of the hypotheses are the same for both expert systems, the contents of the slots are different. A suspected-entity slot describes the element which may be at fault. A state-of-entity slot provides detailed data about the state of a suspected entity - including information such as the element is in fault, is operating out of specification or is operating normally. During the evaluation cycle, this slot is used to indicate the progress of the fault finding. The state-of-hypothesis slot expresses the state of the hypothesis itself. Possible values include: NOT.EVALUATED The hypothesis is newly created and not yet evaluated. NOT.CONFIRMED The hypothesis could not be confirmed but no attempt has been made to deny it. CONFIRMED The hypothesis was confirmed but no recovery attempt has been made. The state of a hypothesis is important for cooperation, because it contains information on the current phase of the diagnosis process. For example if a hypothesis is confirmed by one agent, then the fault has been found and the other system should stop trying to locate it. Another example is that if a hypothesis is evaluated by an expert system because of a request from an acquaintance and it cannot be confirmed, then the originator should be informed since it affects its local problem solving behaviour. A rating slot indicates the priority of a hypothesis and is used to order items on the agenda. During evaluation the expert system might create new hypotheses of a more detailed level - resulting in a tree structure (see figure 3). BEDES and CODES are incapable of understanding each others hypotheses directly because they refer to different domains - BEDES to sub-systems and 15 elements, CODES to knobchains, modules and details. However there is a level of commonality, in that translation can be performed between element and knobchain level hypotheses. This process involves changing the value of some slots of the hypothesis and loading structural data into the knowledge base. For example if BEDES suspects that something is wrong with a controller element (the element is the suspected-entity of a BEDES hypothesis), then CODES cannot directly use this. CODES has to map the element to that set of control system elements (knobchain) which operates this controller element. It also has to load into its knowledge base the structure of the knobchain. The structural knowledge is physically stored in a centralised and separate database for ease of maintenance and because of its sheer size. However for the purposes of this experiment, the database was regarded as part of CODES’s domain level system. So that the agents are not unnecessarily distracted by extraneous hypotheses which they cannot understand (eg CODES G1 S1 S2 E1 E4E3E2 K1 K2 K6K3 K4 K5 M1 M2 M3 M4 M7M6 D1 D2 BEDES CODES general sub-system element knobchain module detail There is a correspondance between E1-4 and K1-4. K5 and K6 have been generated independently by CODES Level of Commonality Figure 3: Hierarchy of Hypotheses 16 cannot use those of a general or sub-system level), agents represent the types of hypotheses that their acquaintances can process in their agent models (see the following section for more details). This knowledge is then used to guide hypothesis interchange. The advantage of using the hypothesis as a common language is that it involves a minimal translation overhead. Also it is close to the language used by the domain level systems which reduces the amount of modification required in the pre-existing systems. The disadvantage is that any new agents which may be added to the community at a later stage must also be able to represent and understand knowledge in this particular format. A better approach in terms of extensibility would be to construct a domain independent interlingua in which assumptions about the knowledge representation commitments are stated explicitly - see the work on the knowledge interchange format30,31 for a more comprehensive discussion of this issue. Benefits of a Distributed AI Approach The benefits of a DAI approach in this particular domain include: 1) As the accelerator and its control system consist of huge numbers of elements, corresponding to more than 10,000 setpoint control values, it would be extremely difficult to maintain and develop a centralised knowledge base for the whole process. Decomposing the problem into smaller modules results in smaller subproblems which are much easier to tackle. A modularised approach also fits more naturally into the existing organisational structure - knowledge of different domains (located in different divisions or groups) can be kept separate, but can be combined by cooperative problem solving at runtime. 2) The overall system will be open. 32 New agents covering different aspects of the particle accelerator process can be added when they are developed without having to alter the application’s existing conceptual model. This is important because new accelerators are built and added during the lifetime of the accelerator complex, also new operational modes may be developed. 17 3) The computing power of several workstations connected together through a network can be utilised. Thus the agents can work in parallel and produce results faster by sharing the workload. 4) Some of the drudgery and non-cognitive aspects of the operator’s job are removed, leaving greater time for the higher level tasks which cannot be automated using the currently available technology. THE GRATE CERN EXPERIMENT This section describes how cooperative fault detection can be carried out using the methods and tools discussed above. A typical cooperative scenario involving BEDES and CODES is outlined, before the steps involved in transforming the stand-alone systems into a community of cooperating agents being controlled by GRATE are expanded upon. A Cooperative Scenario In the implemented scenario, the main form of cooperation manifests itself in terms of the intelligent sharing of information between the two agents. This information is used to indicate changes in the status of the particle accelerator and to direct agents’ problem solving by sharing intermediate and final results. There are three distinct phases to controlling the particle accelerator. Firstly, there is normal operating conditions in which no fault has been detected. In this phase BEDES monitors the accelerator system to identify possible discrepancies. Monitoring involves continually comparing measurable system properties (such as the particle beam’s intensity, efficiency and trajectory) with their archived “ideal” values. If there is a significant discrepancy, then there is a possible fault in the accelerator. When a possible fault is detected, BEDES carries out a preliminary diagnosis phase and produces a list of hypotheses about suspected subsystem components. Once BEDES has produced a list of hypotheses to explain the accelerator fault, the second 18 phase of verifying the cause of the problem begins. The fault may have occurred as the result of a problem at the beam level or a fault with the control system. Therefore as well as starting to verify the cause of the fault, BEDES also informs CODES that there is probably a faulty element in the accelerator. When CODES receives this notification, it starts a diagnostic process to determine whether the problem lies within the control system. At this stage BEDES and CODES share the common goal of trying to locate the accelerator’s fault; they are looking at different aspects of the problem but their work is related by the fact that the hypotheses of CODES are further specialisations of BEDES’s element level hypotheses. As the two agents proceed with their diagnoses, various possibilities for cooperation exist based on the exchange of information about hypotheses. When such information is received, the recipient will undertake one of the following courses of action. A practical illustration of each case follows. 1) take no action if the information is not relevant to its current problem solving context. 2) use the information to deflect the focus of its problem solving activity away from an unprofitable area. 3) use the information to concentrate its problem solving activity on a promising area. Case 1 As a result of its evaluation, BEDES creates some new element level hypotheses. These hypotheses are sent to CODES, which translates them, before adding them to its agenda. As they are new hypotheses (status NOT.EVALUATED) they are merely added to the list of things to do, they do not affect the focus of CODES’s current problem solving. Case 2 BEDES evaluates an element level hypothesis (denoted by H). As H is at the element level it 19 would have already been sent to CODES when it was created (see Case 1). However as a consequence of its evaluation task BEDES has produced more information about H. This additional information is either that H is NOT.CONFIRMED or that H is CONFIRMED (see Case 3 for the latter situation). In the former case, when CODES receives the NOT.CONFIRMED message it takes one of the following actions depending on its problem solving context: a) CODES has not started working on H yet. Since BEDES was unable to confirm H, the chance that CODES will find a fault with the derivatives of H has been lowered. Knowing this, CODES’s rating of H will be reduced meaning that other more likely hypotheses will be dealt with first. b) CODES has started work on H or its derivatives. The probability that one of the hypotheses of the derived tree can be confirmed has been decreased. If there are other high level hypotheses in its agenda, then CODES continues with those. CODES drops its attention on the hypothesis tree of H and will, after the next rearrange agenda, continue with a new tree. c) CODES has already finished the evaluation of H and all the hypotheses derived from it. If CODES was also unable to confirm any hypothesis in the tree of H, then this is further confirmation of BEDES’s result and can be used to increase the operator’s level of confidence in the information. However if CODES did find a fault, then there is a conflict. In this instance resolution is straightforward; since CODES works at a lower level than BEDES its results are assumed to be more reliable. The user is thus presented with the result of CODES and a short note about the conflict. Case 3 If BEDES can confirm H, then the problem solving effort of CODES now switches to concentrate on the hypothesis tree of H and its derivatives. If CODES also detects a fault in this tree then the user’s confidence in the diagnosis will be heightened. If it does not find a fault then 20 there is a conflict and the operator is informed. The above cases describe situations in which information supplied by BEDES is used to direct the problem solving of CODES. However there is also a valuable flow of information in the opposite direction. The exchange of hypotheses from CODES to BEDES works differently because BEDES cannot translate nor understand the hypotheses of CODES directly. So knobchain level hypotheses created by CODES (eg K5 and K6 in figure 2) are of no direct relevance to BEDES. BEDES is only interested in results related to the hypotheses which it has previously sent to CODES (K1-K4 in figure 2). There are two situations in which CODES should send a result to BEDES: a) CODES was able to confirm a hypothesis. Since BEDES cannot understand the hypothesis itself a back-translation is needed. This involves moving up the tree to find the root node from which all the other hypotheses were derived and then sending this hypothesis back to BEDES. Thus if CODES finds a fault in the detail level (say D1 in figure 2) then its root hypothesis (K3) should be translated to (E3) and sent back to BEDES. b) CODES could not confirm any of the hypotheses derived from those sent by BEDES. This only occurs if all hypothesis from the tree are evaluated and none of them could be confirmed. The result sent back (after translation) will be the original hypothesis of BEDES with the status NOT.CONFIRMED. In both of these cases, BEDES tries to integrate the information received into its own problem solving activity. The situations which might occur now are quite similar to those of Cases 1-3 in that the attention of BEDES can be drawn or dropped to a certain hypothesis depending on the status of the information supplied by CODES. If BEDES was ahead of CODES then the results will be compared; in the case of a conflict it is assumed that the agent which found a fault is more reliable. The cooperative fault finding phase comes to an end if all hypotheses are evaluated and none 21 of them could be confirmed (a transient fault or a false alarm) or a hypothesis has been confirmed. In either case the recovery phase will begin. This phase is outside the scope of this experiment. Integrating Pre-Existing Expert Systems Converting the standalone versions of BEDES and CODES into a community of cooperating agents being controlled by GRATE required three main activities to be carried out. Firstly, some adaptations to the control of BEDES and CODES were required and the domain level tasks had to be defined. Secondly, the interface between the expert systems and GRATE had to be constructed. Finally, the acquaintance and self models of the agents needed to be populated - this process includes specifying the recipes2 which control agent activity, enumerating the tasks which the domain level system can perform and representing the information which other agents would benefit from receiving. Expert System Adaptations As figure 2 illustrates, the initial control of both BEDES and CODES was a non-interruptible loop. However for the purpose of controlling local problem solving from an upper layer, such a coarse granularity was inappropriate. To utilise the benefits of cooperation, GRATE has to be capable of influencing the rating of hypotheses and of injecting new items into the agenda. Therefore the control cycle needed to be split into more manageable components. As the original coding of the control loop was carried out in a modular fashion it was relatively straightforward to decouple the cycling routines from the actual functionality which it drove. Having identified the control regime, the next step is to determine the domain level system tasks. When performing this analysis the overriding objective was to minimise the amount of change required to the structure of the pre-existing systems, whilst still permitting the benefits of interaction to take place. Inspection of the existing control loop suggested there should be six tasks 2. Recipes are sequences of actions known by an agent for achieving a particular objective33. 22 - one for each node of the graph. However a deeper examination of the system structure revealed that “retrieve data” and “evaluate hypothesis” are virtually indivisible because the former is deeply embedded within the latter. Also “create and inject” is intimately related to the evaluation process and also could not easily be separated. Selection of hypothesis from the agenda was regarded as the initialisation phase for evaluation, therefore it was decided to leave it hidden in the intelligent system. Thus the control loop was collapsed into two basic tasks - evaluating hypotheses and rearranging the agenda. • REARRANGE-AGENDA: re-arranges the agenda so that the highest priority tasks are near the beginning. Also removes any superfluous hypotheses. • EVALUATE-HYPO: takes the first hypothesis from the agenda and evaluates it. This level of control was considered appropriate for two reasons. Firstly because of the way the pre-existing systems were implemented; any other decomposition would have required significant modifications to the existing structure. Secondly, from the perspective of exploiting information gleaned from other agents, the advantages of a finer level control would have been negligible. As a consequence of this reconceptualisation it is apparent that some of the control which resided originally in the expert system has migrated up into GRATE’s cooperation and control layer. However not all of the control has been moved, a significant amount of lower level control remains with the domain level system. In this application GRATE exerts control over the domain level system through its agenda. Therefore some of the manipulation functions which existed within the domain level system needed to be made available at the cooperation and control layer if the benefits of information received from acquaintances is to be exploited. These functions include: • INJECT-HYPO: inject a hypothesis into the agenda • DELETE-HYPO: remove a hypothesis from the agenda 23 • GET-AGENDA: return the current contents of the agenda • CHANGE-RATING: modify the rating slot of a hypothesis in the agenda The responsibility for ensuring that these commands are executed in a coherent manner resides with the control of the domain level system. Thus if GRATE decides that the rating of a hypothesis should be modified, it issues the “change-rating” command to its domain level system. Once received this directive will not be acted upon immediately since, for example, the expert system may be in the middle of performing an evaluation. Only when it comes to its rearrange agenda task will the modification actually take place. From a design perspective, it is important that such domain specific control remains within the expert systems. Exporting it to the upper level would require GRATE’s control module to be at least as sophisticated as the control of the domain level system and would also mean that it was different for each and every application. Maintaining a clean separation of concerns allows GRATE’s control module to be simpler and more generic. Interfacing GRATE and the Domain Level Systems BEDES and CODES run on separate workstations and are implemented in KEE making use of SUN Common LISP; GRATE is written in Allegro Common LISP. Because of incompatibilities between the different pieces of software, and also for efficiency reasons, it was decided to run GRATE on a third workstation. Thus all the agents’ cooperation and control layers ran on one machine, this machine being different from the ones which were executing the domain level systems. To allow an agent’s control module to interact with its domain level system a communication package was utilised. This package established bidirectional communication between SUN Common LISP on one workstation and Allegro Common LISP on another. It was based on standard UNIX tools such as sockets and TCP/IP and had been developed as part of the application specific cooperation software used for preliminary experimental work. GRATE would issue commands such as: START(EVALUATE-HYPO). This directive would be picked up by the communication package which would send the message onto the workstation running the 24 appropriate domain level system. In addition to interacting with the domain level system, the cooperation and control layer needs to carry out reasoning about received and generated information. To facilitate this, some domain dependent functions needed to be written for inclusion into GRATE’s recipes. In this experiment these functions were primarily related to providing an interpretation of the common language (i.e. the hypotheses) and of presenting output to the operator. Examples of such functions include: • (has-slot-value <hypo> <slot> <value>) boolean function which verifies if a specified slot of a hypothesis contains a certain value • (has-equal-slots <hypo-1> <hypo-2> <slot>) boolean function which verifies if the slots of the two hypotheses contain the same value • (find-related-hypos <hypo-list> <hypo>) returns all members of the hypothesis list which have the same value in the SUSPECTED.ENTITY slot as the specified hypothesis • (confirm <hypo>) displays message to the operator that a hypothesis has been confirmed by both agents Instantiating the Agent Models When building a GRATE application a significant proportion of the knowledge required to control cooperative problem solving is built into the system. For these experiments, no additions were needed to the generic rule set. Thus each agent had exactly the same rules in its cooperation and control layer and the application builder was only concerned with the domain-dependent 25 features of GRATE (i.e. the agent models). Firstly, the self models need to be instantiated. This involves describing the tasks which the domain level system is able to perform - including the name, the inputs it must receive in order to execute and the results which are produced. In this experiment the self models contained descriptions of the following tasks: rearrange-agenda, evaluate-hypo, inject-hypo, delete-hypo, get-agenda and change-rating. Two sample descriptions are given below: TASK NAME: EVALUATE-HYPO MANDATORY INPUTS:(HYPO) RESULTS PRODUCED:(STATUS NEW-HYPOS) TASK NAME: CHANGE-RATING MANDATORY INPUTS: (HYPOS RATING-CHANGE) RESULTS PRODUCED: NIL Tasks are grouped together into recipes. Recipes have trigger conditions which indicate when they should be activated, a body which describes the actions to be performed and a description of the results produced. The recipe which encodes the basic control loop for the agent’s fault verification phase is shown in figure 43. This recipe is triggered when the accelerator monitoring phase detects a problem; it loops continuously until the cause of the fault has been ascertained whereupon a recovery mode recipe is invoked. RECIPE NAME:(VERIFY-CAUSE-OF-FAULT) TRIGGER: (ENTER-FAULT-FINDING-MODE) ACTIONS:( (START (REARRANGE-AGENDA (> FIRST-HYPO)) 3. “>” means unbound variable and “<” indicates a bound variable. Thus the evaluate-hypo task takes one input (called first-hypo) and produces two outputs (respectively named status and new-hypos). 26 (START (EVALUATE-HYPO (< FIRST-HYPO) (> STATUS) (> NEW-HYPOS))) (LOOP-UNTIL (FAULT-CONFIRMED (< STATUS)))) RESULTS: (NEW-HYPOS) Figure 4: Basic Control Loop for Fault Verification Phase Figure 5 illustrates a recipe which describes how CODES utilises information about hypotheses received from BEDES. In particular it highlights the way in which information about the state of hypotheses can be used to draw or deflect CODES’s attention from a particular branch of the search space. It encodes cases two and three of the cooperative scenarios highlighted earlier; note to simplify the example, cases of conflict are not dealt with. The recipe is triggered when CODES receives a hypothesis which BEDES has evaluated (status CONFIRMED or NOT.CONFIRMED). According to cooperative scenario case 1, CODES will already have received information about the hypothesis when it was first generated (status NOT.EVALUATED). To carry out the necessary reasoning, CODES has to identify those hypotheses in its agenda which are related to the one just received from BEDES. This matching process is carried out by the recipe’s first two actions GET-AGENDA and FIND-RELATED- HYPOS. The remaining recipe actions are conditional upon CODES’s current problem solving context. The first condition tests whether BEDES has also verified a hypothesis which CODES has already confirmed (CONFIRMED.HYPO). If this is the case, then the level of confidence in the diagnosis is increased and the operator should be informed. The second conditional action draws the attention of CODES to the hypotheses related to the one confirmed by BEDES - this is achieved by increasing the rating of the related hypotheses by a value of 30. The final action deflects CODES’s attention away from a hypothesis tree which appears less promising. 27 RECIPE NAME:(USE-EVALUATED-HYPO-INFORMATION) TRIGGER: (AND(info-available HYPO) (not (has-slot-value HYPO STATE.OF.HYPO NOT.EVALUATED)) ACTIONS: ( (start(GET-AGENDA (> FULL-AGENDA))) (start (FIND-RELATED-HYPOS (< FULL-AGENDA)(< HYPO)(> RELATED-HYPOS))) (start-if (and (has-slot-value HYPO STATE.OF.HYPO CONFIRMED) (has-equal-slot HYPO CONFIRMED.HYPO SUSPECTED.ENTITY)) (CONFIRM (< HYPO))) (start-if (has-slot-value HYPO STAT.OF.HYPO CONFIRMED) (CHANGE-RATING (< RELATED-HYPOS) 30)) (start-if (has-slot-value HYPO STAT.OF.HYPO NOT.CONFIRMED) (CHANGE-RATING (< RELATED-HYPOS) -20))) RESULTS: NIL Figure 5: CODES recipe for exploiting information received from BEDES Once the self models have been completed the acquaintance models need to be populated. In the example cooperative scenarios the most important feature to model about another agent is the information which it is known to be interested in. For example BEDES’s model of CODES contains the information that CODES would benefit from receiving any newly generated hypotheses which are at the element level (cooperative scenario case 1), any element level hypotheses that it has been unable to confirm (cooperative scenario case 2), or any element level hypotheses that it has been able to confirm (cooperative scenario case 3). 28 INTERESTS: (..(HYPO (AND (AT-ELEMENT-LEVEL HYPO) (HAS-SLOT-VALUE HYPO STAT.OF.HYPO NOT.EVALUATED))) (HYPO (AND (AT-ELEMENT-LEVEL HYPO) (HAS-SLOT-VALUE HYPO STAT.OF.HYPO CONFIRMED))) (HYPO (AND (AT-ELEMENT-LEVEL HYPO) (HAS-SLOT-VALUE HYPO STAT.OF.HYPO NOT.CONFIRMED)))...) RESULTS AND EXPERIENCES BEDES and CODES were successfully transformed from standalone expert systems to a community of cooperating agents under the control of GRATE. This transformation was achieved with minimal modifications to the pre-existing expert systems and with no augmentation to GRATE’s generic knowledge. The cooperating system was tested using a special development mode of the accelerator in real time. As the accelerator operates in a time sharing fashion it was possible to deliberately introduce faults into the system in the test mode without disturbing the other modes which were serving real physicists experiments. The results of this experiment highlighted some shortcomings in the design decision to map the BEDES and CODES expert systems directly into agents. This proved to be a less than optimal choice because of the large amount and diverse range of processing carried out in each system. As both systems were originally conceived as standalone pieces of software they contain a vast array of functionality which does not logically belong together - including monitoring, data acquisition, fault diagnosis and recovery. Also because of their sheer size, the expert systems were becoming unwieldy in their own right. Introducing such systems into a cooperating community merely exacerbated these problems. 29 Using the cooperation metaphor it is possible to divide the systems into a number of simpler and logically separate agents which could work on dedicated areas of the problem. A new design was proposed in which the functionality contained in BEDES and CODES was split into seven agents34. This new approach offers greater system modularity and allows the benefits of parallelism and interaction to be exploited to an even greater degree. Firstly the data acquisition and treatment functionalities were separated out. It became apparent that the reasoning process in the different agents was heavily reliant on the correctness of the data. For this reason, the treatment of acquisitions is likely to become increasingly sophisticated in the future and so a dedicated agent is warranted. An additional advantage is that it is easier to provide treated data from a single source rather than having to do separate acquisition and treatment in each and every agent. The next decision was to remove the user interface functionality from the individual systems and provide homogeneous presentation of data through a specialised agent. This provides the operator with one entry point to the entire agent community. It also allows him to be presented with high level information about process parameters which can be obtained through interaction with the acquisition agent. This is not possible in the existing control system and so the operator has to rely on raw data from the process. BEDES and CODES were originally conceived as diagnostic expert systems; it was some time before their recovery facilities were incorporated. Because of their add-on nature, it was decided to separate the recovery actions from the diagnostic part. Finally, BEDES reasons on two conceptual levels: on the high level of beam parameters (which must be deduced from raw data rather than acquired directly) and on the direct level of the equipment (raw data). In the modified design these two levels are mapped into separate agents. This is beneficial because it frees the high level reasoning from the shackles imposed by the strict mechanism of the hypothesis verification process. 30 CONCLUSIONS The stated aim of this work was to take two standalone and pre-existing expert systems and construct from them a community of cooperating agents. This was achieved by using the GRATE system to control the cooperative activity and required only slight modifications to the expert systems. The cooperating community worked together to diagnose faults which occurred in the real particle accelerator process. Most reported systems work with highly idealised problem solvers and simplified domains. This naturally makes experimentation easier, but is dangerous in that the assumptions which have to be made may hide important issues which need to be addressed if the technology is to be used in realistic environments eventually. This experiment used real expert systems working on a real world problem. By adopting this approach the experiment provided many useful insights into the fundamental problem of incorporating pre-existing systems in a community of cooperating agents. When undertaking this activity, the structure of the pre-existing system needs to be analysed to ensure it is open enough for the cooperation and control layer to exploit the information gleaned during interactions with fellow community members. Necessary modifications may include defining a finer granularity of control, making previously hidden control functions explicit or developing completely new functions. From this experiment it is clear that some of the basic control which previously resided in the domain level system needs to be moved into the cooperation and control layer. However lower level control which is more application specific is best left at the domain level. When deciding the separation of concerns there is a tradeoff between the amount of restructuring of the domain level system and the desired granularity of control. A detailed analysis of the existing system must be undertaken so that a balance can be struck which avoids a significant reformulation system but which still allows the benefits of interaction to be realised. A second important consideration which this experiment uncovered is the relationship between 31 pre-existing systems and agents - it is not always best to adopt a simple one to one mapping. Often standalone systems contain a number of logically disparate tasks which can be split into separate agents. This decomposition typically allows greater parallelism to be exploited in problem solving and makes better use of the cooperating systems metaphor. The types of interaction exhibited by BEDES and CODES are typical of a broad class of problem solving called Functionally Accurate, Cooperative (FA/C)35. In the FA/C paradigm agents asynchronously exchange partial results about the intermediate state of their processing to ensure the community arrives at a consistent interpretation of the whole problem. From a traditional computer science perspective, this type of cooperative problem solving can be viewed as a form of distributed search which has multiple loci of control36. In the CERN experiment, the partial results are hypotheses and as further evidence emerges (eg hypotheses become CONFIRMED or NOT.CONFIRMED) so the problem solving behaviour of the community develops accordingly. The types of cooperation exhibited in this experiment are similar in nature to other industrial applications which have been studied within the context of the ARCHON project (eg electricity transport management37 and management of electricity distribution38) which suggests that theoretical research into the FA/C paradigm has a practical use. This experiment also provides further support for two of Lesser’s observations about the FA/ C paradigm36. Firstly, that in comparison to a standalone version, an FA/C agent is more complex. This can be seen by the refashioning of and additions to the expert systems’ control regimes. Secondly it is observed that effective control of cooperative problem solving requires local control decisions to be influenced by the state of problem solving in other agents. In this experiment, the behaviour of other agents is monitored by the cooperation and control layer and influence is exerted through modifications of the agent’s agenda. The implemented cooperation schemes are relatively straightforward, but there is scope for greater sophistication which may enhance performance still further. At present when BEDES sends CODES its initial block of hypotheses, CODES starts processing them in the order in which they 32 arrive. This means both agents start trying to verify the fault from the same position in the search space. As the hypotheses are unrated at this point, this focuses the communities efforts on an uneccesarily small portion of the accelerator. It would be better if the two agents worked on different areas of the search space while the information about the faults is limited, then only when further information becomes available would they focus their joint efforts on promising areas. This could be realised by a relatively straightforward approach in which CODES starts working on hypotheses from the end of the list or in a more elegant manner by a form of negotiation39,40 in which agents decide upon the portion of the problem space on which they will concentrate their initial efforts. This division of labour is possible because both systems are usually capable of detecting the same fault. If, however, neither of them is able to find the fault in their part of the search space then only at this stage should they start trying to work on areas which the other agent has already processed. A final enhancement to the application would be to incorporate a more sophisticated mechanism for resolving conflicting opinions between the agents. For example rather than simply assuming the result produced by CODES is correct, it would be better if some conflict resolution expertise could be used to determine the source of the conflict and provide a means of resolving it41. Such a mechanism would enable better quality information to be presented to the operator. ACKNOWLEDGMENTS The work described in this paper has been carried out in the ESPRIT II project ARCHON (P2256) whose partners are: Atlas Elektronik, JRC Ispra, Framentec-Cognitech, Labein, Queen Mary and Westfield College, IRIDIA, Iberdrola, EA Technology, Amber, Technical University of Athens, University of Amsterdam, Volmac, CERN and University of Porto. REFERENCES 1 Bond, A. H. and Gasser, L. (eds), Readings in Distributed Artificial Intelligence, Morgan Kaufmann (1988). 33 2 Gasser, L. and Huhns, M. N., (eds), Distributed Artificial Intelligence Volume II, Pitman Publishing (1989). 3 Huhns, M. N. (ed) Distributed Artificial Intelligence, Pitman Publishing (1988). 4 Sridharan, N. S. 1986 Workshop on Distributed AI, AI Magazine, Fall, 75-85 (1987). 5 Davis, R. and Smith, R. G. Negotiation as a Metaphor for Distributed Problem Solving, Artificial Intelligence, 20, 63-109 (1983). 6 Lesser, V. R. and Erman, L. D. An Experiment in Distributed Interpretation, IEEE Trans. on Computers, 29(12), 1144-1163 (1980). 7 Jennings, N. R. and Wittig, T. ARCHON: Theory and Practice, in Distributed Artificial Intelligence: Theory and Praxis, (eds L.Gasser and N.M.Avouris), 179-195, Kluwer Academic Press (1992). 8 Wittig, T. ARCHON: An Architecture for Multi-Agent Systems, Ellis Horwood (1992). 9 Jennings, N. R. Cooperation in Industrial Systems, Proc. ESPRIT Conference, Brussels, Belgium, 253-263 (1991). 10 Jennings, N. R. Mamdani, E. H. Laresgoiti, I. Perez, J. and Corera, J. GRATE: A General Framework for Cooperative Problem Solving, Journal of Intelligent Systems Engineering, 1 (2) 102-114 (1992). 11 Jennings, N. R. Using GRATE to Build Cooperating Agents for Industrial Control, Proc. IFAC/ IFIP/IMACS International Symposium on Artificial Intelligence in Real Time Control, 691- 696, Delft, The Netherlands (1992). 12 Cohen, P. R. A Survey of the Eighth National Conference on Artificial Intelligence: Pulling Together or Pulling Apart, AI Magazine, 12 (1), 16-41 (1991). 34 13 Roda, C. and Jennings, N. R. The Impact of Heterogeneity on Cooperating Agents, Proc. AAAI Workshop on Cooperation among Heterogeneous Intelligent Systems, Anaheim, Los Angeles, USA (1991). 14 Simon, H. A. Models of Man, New-York, Wiley (1957). 15 Lesser, V. R., and Corkill, D. D, Distributed Problem Solving, Encyclopedia of Artificial Intelligence (Ed S.C.Shapiro), 245-251, John Wiley and Sons (1987). 16 Whitney, C. Cooperating Intelligent Agents: A Study of GRATE, BT Report MAIN-WP1008, BTRL Martlesham Heath, Ipswich, UK (1992). 17 Jennings, N. R. Joint Intentions as a Model of Multi-Agent Cooperation in Complex Dynamic Environments, Ph.D. Thesis, Dept. Electronic Engineering, Queen Mary and Westfield College (1992). 18 Engelmore, R. and Morgan, T. (eds) Blackboard Systems, Addison Wesley (1988). 19 Hayes-Roth, B. The Blackboard Architecture: A General Framework for Problem Solving?, Stanford Heuristic Programming Project, HPP-83-30, Stanford University (1983). 20 Hewitt, C. E. and Kornfield, W. A. Message Passing Semantics, SIGART Newsletter, 48 (1980). 21 Hewitt, C. E. and Lieberman, H. Design Issues in Parallel Architectures for Artificial Intelligence, Proc. of IEEE Computer Society International Conference, 418-423 (1984). 22 Malandain, E. Pasinelli, S. and Skarek, P. A Fault Diagnostic Expert System Prototype for the CERN PS, Europhysics Conference on Control Systems for Experimental Physics, Villars-sur- Ollon, Switzerland (1987). 23 Skarek, P. Malandain, E. Pasinelli, S. and Alarcon, I. A Fault Diagnosis Expert System for 35 CERN Using KEE, SEAS (SHARE European Association) Spring Meeting, Davos, Switzerland (1988). 24 Malandain, E. Pasinelli, S. and Skarek,P. Knowledge Engineering Methods for Accelerator Operation, European Particle Accelerator Conference, Rome, Italy (1988). 25 Malandain, E. and Skarek, P. Linking a Prototype Expert System to an Oracle Database, IASTED, International Conference on Expert Systems, Theory and Applications, Zurich, Switzerland (1989). 26 Malandain, E. An Expert System in the Accelerator Domain, International Workshop on Software Engineering, Artificial Intelligence and Expert Systems for High Energy Nuclear Physics, Lyon Villeurbanne, France (1990). 27 Fuchs, J. Skarek, P. Varga, L. and Wildner-Malandain,E. Integration of Generalized KB- Systems in Process Control and Diagnosis, Invited paper for the SEAS conference, Lausanne, Switzerland (1991) 28 Fuchs, J. Skarek, P. Varga, L. and Wildner-Malandain,E. Distributed Cooperative Architecture for Accelerator Operation, 2nd International Workshop on Software Engineering, Artificial Intelligence and Expert Systems for High Energy and Nuclear Physics, L’Agelonde, La- Londe-les-Maures, France (1992). 29 Varga, L. Cooperation Between the Two Diagnostic Expert Systems BEDES and CODES, CERN Technical Report, PS/CO/WP 91-02, (1991) 30 Neches, R. Fikes, R. Finin, T. Gruber, T. Patil, R. Senator, T. and Swartout, W. R. Enabling Technology for Knowledge Sharing, AI Magazine, Fall, 36-56 (1991). 31 Ginsberg, M. L. Knowledge Interchange Format: The KIF of Death, AI Magazine, Fall, 57-63 (1991). 36 32 Hewitt, C. E. The Challenge of Open Systems, BYTE, 10 (4), 223-244 (1985). 33 Pollack, M. E. Plans as Complex Mental Attitudes, in Intentions in Communication (Eds P.R.Cohen, J.Morgan and M.E.Pollack), 77-105, MIT Press (1990). 34 Fuchs, J. Skarek, P. Varga, L. and Wildner-Malandain, E., (1992), Distributed Cooperative Architecture for Accelerator Operation, ARCHON Technical Report 26, CERN, Geneva (1992). 35 Lesser, V. R. and Corkill, D. D. Functionally Accurate, Cooperative Distributed Systems, IEEE Trans. on Systems, Man and Cybernetics, 11 (1), 81-96 (1981). 36 Lesser, V. R. A Retrospective View of FA/C Distributed Problem Solving, IEEE Trans. on Systems, Man and Cybernetics, 21 (6), 1347-1362 (1991). 37 Aarnts, R. P. Corera, J. Perez, J. Gureghian, D. and Jennings, N. R. Examples of Cooperative Situations and their Implementation, Vleermuis Journal of Software Research, 3 (4), 74- 81 (1991). 38 Cockburn, D. Varga, L. Z. And Jennings, N. R. Cooperating Intelligent Systems for Electricity Distribution, Proc. Expert Systems 1992 (Applications Track) , Cambridge, UK (1992). 39 Laasri, B. Laasri, H. and Lesser, V. R. An Analysis of Negotiation and its Role in Cooperative Distributed Problem Solving, Proc. Second Generation Expert Systems Conference, Avignon, France (1991) 40 Conry, S. E. Kuwabara, K. Lesser, V. R. and Meyer, R. A. Multi-Stage Negotiation for Distributed Constraint Satisfaction, IEEE Trans. on Systems, Man and Cybernetics, 21 (6), 1462-1477 (1991) 41 Klein, M. Supporting Conflict Resolution in Cooperative Design Systems, IEEE Trans. on Systems, Man and Cybernetics, 21 (6), 1379-1390 (1991). 37 FIGURE LEGENDS 1) Detailed GRATE Agent Architecture 2) Expert System Control Loop 3) Hierarchy of Hypotheses 4) Basic Control Loop for Fault Verification Phase 5) CODES recipe for exploiting information received from BEDES 
work_2ntpv4mpyre3xdhmwt3t2mqfve	----	Annals of Emerging Technologies in Computing (AETiC) Vol. 5, No. 1, 2021 Obada Alhabashneh, "Fuzzy-based Adaptive Framework for Module Advising Expert System”, Annals of Emerging Technologies in Computing (AETiC), Print ISSN: 2516-0281, Online ISSN: 2516-029X, pp. 13-27, Vol. 5, No. 1, 1st January 2021, Published by International Association of Educators and Researchers (IAER), DOI: 10.33166/AETiC.2021.01.002, Available: http://aetic.theiaer.org/archive/v5/v5n1/p2.html. Research Article Fuzzy-based Adaptive Framework for Module Advising Expert System Obada Alhabashneh Mutah University, Karak, Jordan o.alhabashneh@mutah.edu.jo Received: 3rd November 2020; Accepted: 22nd December 2020; Published: 1st January 2021 Abstract: In the enrolment process, selecting the right module and lecturer is very important for students. The wrong choice may put them in a situation where they may fail the module. This could lead to a more complicated situation, such as receiving an academic warning, being de-graded, as well as withdrawn from the program or the university. However, module advising is time-consuming and requires knowledge of the university legislation, program requirements, modules available, lecturers, modules, and the student's case. Therefore, the creation of effective and efficient systems and tools to support the process is highly needed. This paper discusses the development of a fuzzy-based framework for the expert recommender system for module advising. The proposed framework builds three main spaces which are: student-space (SS), module- space (MS), and lecturer-space (LS). These spaces are used to estimate the risk level associated with each student, module, and lecturer. The framework then associates each abnormal student case in the students’ grade history with the estimated risk level in the SS, MS, and LS involved in that particular case. The fuzzy- based association-rule learning is then used to extract the dominant rules that classify the consequent situation for each eligible module if it is to be taken by the student for a specific semester. The proposed framework was developed and tested using real-life university data which included student enrollment records and student grade records. A five-fold cross-validation process was used for testing and validating the classifying accuracy of the fuzzy rule base. The fuzzy rule base achieved a 92% accuracy level in classifying the risk level for enrolling on a specific module for a specific student case. However, the average classifying accuracy achieved was 89.2% which is acceptable for this problem domain as it involves human behavior modeling and decision making. Keywords: Intelligent Academic Advisor; Module Adviser; Expert System; Fuzzy Logic; Fuzzy Rule-based. 1. Introduction Every academic semester, students enroll in different modules offered by their universities. Enrolment might sound like a simple process, but it involves many complications and problems for students particularly in these universities where the credit hour system is adopted. According to this system, although there are some constraints such as the academic plan, the students still have the freedom to select the sections to enroll in based on their preferences including the module, timeslot, and lecturer of the offered sections in the timetable. However, the lack of knowledge about the modules and the lecturers, coupled with the contradictory advice they receive from their colleagues and the complexity of the academic plan of their programs may affect the students' ability to select the right choice. This creates a need for professional advice as taking a wrong choice may lead to unwanted situations such as failing the module, or more seriously, academic warning, program withdrawal or even leaving the university [1-3]. To address this problem, universities usually have academic advisors to help the students in selecting their modules and sections for the upcoming semester based on different factors including http://aetic.theiaer.org/ http://aetic.theiaer.org/ http://www.theiaer.org/index.htm http://aetic.theiaer.org/archive/v5/v5n1/p2.html mailto:o.alhabashneh@mutah.edu.jo AETiC 2021, Vol. 5, No. 1 14 www.aetic.theiaer.org the student’s cumulative Grade Point Average (GPA), the available modules, timeslots, module’s lecturer, and the academic plan of the student. However, module advising is a challenging task as it is complicated, time-consuming, and needs qualified and knowledgeable human resources. This urges the need for developing intelligent systems for module-advising to aid the students in selecting the right modules at the right time to ensure a smooth passage for students throughout their academic plans to finish their degrees with fewer problems. Intelligent Recommender Systems (RSs) are widely used in different application areas including online shopping, movies, social networking, and others. However, they are less common in education and this area is still under development [1-3]. RSs can provide the students with personalized recommendations on suitable modules based on the student’s academic history. Such systems can also incorporate previous student's experiences to provide an improved recommendation. Furthermore, they adapt to the changes in the data and give the recommendation based on that. The need for RSs and their requirements in the education sector has been discussed in [4-6]. More recently, the use of RSs in education has received more attention and their potential has been proven [7, 8]. This paper discusses the development of a fuzzy-based framework to be used in recommender systems for module advising. The proposed framework builds the recommendation based on creating three main spaces which are: student-space (SS), module-space (MS), and lecturer-space (LS). These spaces are used to estimate the risk level associated with each student, module, and lecturer. The framework then associates each abnormal student case in the students’ grade history with the estimated risk level in the SS, MS, and LS involved in that particular case. Fuzzy-based association- rule learning [9,10] is then used to extract and summarize a fuzzy rule base. The fuzzy rule base is used to predict the risk level associated with the combination of a specific student, module, and lecture and build the recommendation to students based on that. The proposed approach provides a novel mechanism to estimate the consequent risk associated with the student selection of a specific module that is taught by a specific lecturer. The risk estimation is based on creating and analyzing the three space elements. 2. Related Work 2.1. Course/Module Advising in Higher Education Academic advising is vital for a good academic experience of students as it positively impacts their success and retention [11, 12]. However, it requires specific knowledge of the student's situation, history, program’s academic plan, modules, and lecturers [12]. Gordon et al. [13], defined academic advising as “situations in which an institutional representative gives insight or direction to a college student about an academic, social, or personal matter”. Module advising is one of the main tasks of the academic advisor, which refers to the process by which the academic advisor supports the student in selecting the right modules in which to enroll. Having the importance and the complexity of module advising, different researchers have argued that there is a need for developing intelligent systems to support this task [2, 4, 7, 14]. Developing such systems aims to minimize the demand for human advisors and gives them more time to focus on other important advising tasks including career development and financial issues. [12-16]. A novel approach for module long term plaining called Interactive Decision Support for Course Planning (IDiSC+) was proposed by Mohamed [12]. The approach used optimization techniques to support both the student and the advisor in building a recommended long-term academic plan (towards graduation) to be followed by the student. Laghari [14] developed an Automated Course Advising System (ACAS) for module planning. The system distributes the modules of the academic plan on different semesters based on the history of other students. There has been significant research effort in applying information technology for module advising. Roushan et al. [15] introduced an Internet-based approach to support the module-advising process which integrated the process of advising with the enrolment taking into account the constraints of the program plan and the university policy. However, the system did not replace the human academic-advisor but rather facilitated the advisor role by providing an automated tool for AETiC 2021, Vol. 5, No. 1 15 www.aetic.theiaer.org communication and information exchange. A decision-support system was proposed to support academic advisors in preparing a pre-enrolment plan for the students and assist in the offering of appropriate modules for the upcoming semester [16]. Al-Ghamdi et al.[17] developed an advisor expert system (PAS) for postgraduate students. The proposed system was designed to assist them in the selection of the most relevant modules without referring to a human module-advisor. A web- based advising system [18] was proposed which supports three types of users, including students, advisors, and heads of department to make sure that a complete picture is available for the students. Mattei et al.[19] developed a decision-theory advising tool to enhance the advisor-student relationship. The tool allowed students to browse the module offerings, possible future scenarios, and their outcomes. Shatnawi et al. [20] used the enrolment and marking history from similar cases and applied an association rule-based system to give general recommendations when selecting the modules on which to enroll. 2.2. Content-based Different researchers have used a content-based academic recommendation system for module selection. For example, Lin et al.[21] utilized a multi-agent approach and ontology to provide a dynamic and personalized recommended module list. The multi-agent approach, which included various agents, used a preference-driven planning algorithm supported by the ontology to build the recommendations. Darimola et al.[4] integrated Case-Based Reasoning (CBR) with Rule-Based Reasoning (RBR) techniques to provide an intelligent approach for module advising. The approach also used historical data to build a list of recommended modules for the following semester. 2.3. Collaborative filtering Collaborative filtering was also used for academic advising. Huang et al. [22] proposed a score- based prediction approach for course recommendation. The approach used a user domain collaborative filtering to create the recommended course list. Courses were clustered based on the student feedback in [23] and the resulted clusters were combined with fuzzy based rule association technique to create the course recommendation. Nafea et al. [24] developed a learning-style-based collaborative-filtering approach for module recommendations which utilized different metrics to identify similar profiles including k-means, cosine, and person correlation. Chang et al. [25] proposed a user-based collaborative-filtering approach to predict student grades. Mortenson et al. [26] introduced a collaborative filtering approach for module selection which utilizes an artificial immune mechanism for the prediction. Bydžovská [27] investigated the effect of student and module features on the enrolment patterns and designed a collaborative-filtering-based system to predict the module grade. Yao et al. [28] attempted to increase the fairness of the module recommendation by addressing the biased-recommendation problem against minority groups of students. They developed four different fairness metrics that can be optimized using the learning objectives 2.4. Knowledge-based Different knowledge-based recommendation approaches have been proposed for module selection. Xu et al. [29] developed a knowledge-based approach to offer a personalized module sequence to the new students. This approach utilized a dynamic learning algorithm that learns from the performance of other students on a specific module. Koutrika [30] argued that recommendation methods should not be 'hard-wired', but it should be flexible. In that sense, a new paradigm for the recommendation was introduced in which a recommendation approach can be defined declaratively as a high-level parameterized workflow comprising traditional relational operators and new operators that generate or combine recommendations. Keston et al. [31] utilized semantic web expert system technologies to build a knowledge base that is used by an intelligent web-based application to provide the required recommendations. Engin et al. [32] developed an expert rule-based system for module recommendation. The rules were captured from the real advisors and then injected into a rule-base using Oracle Policy Automation (OPA). AETiC 2021, Vol. 5, No. 1 16 www.aetic.theiaer.org Hashemi and Blondin [33] included several factors to be taken into consideration when recommending modules for students such as the frequency of the module offering, balancing the module load, and shortening the graduation path. All these factors and others were included in a rule base which was used for recommending modules. Ayman [34] proposed an expert system for module selection which included both prescriptive and developmental advising models and utilized object- oriented databases for data and rule representation. Abdullah Al-Ghamdi et al. [17] proposed a rule- based advising system. The system was designed for postgraduate students and built recommendations to support the students in the selection of modules related to the topic of their research thesis. Nambia and Dutta [35] introduced a dynamic and flexible rule-based advising system in which the rules were separated from the execution to enable the student to try different scenarios by updating the XML file where the rules are stored. Nguyen et al. [36] proposed a knowledge-based framework that utilized a learning data-warehouse for discovering patterns in the student behavior including module selection and achievements. These patterns were then used to make the recommendations. 2.5. Hybrid Other researchers proposed Hybrid approaches in which more than different perspectives were integrated. Daramola. et al. [4] designed an intelligent expert system for module advising which integrated rule-based reasoning (RBR) with Case-Based Reasoning (CBR) using the academic history of the students. Sobecki [37] applied Ant Colony Optimization (ACO) to provide an efficient module advisor system. The system predicted the final grade of the students in a module, based on a domain- specific representation integrated with the ACO. Abdulwahhab [38] integrated Genetic Algorithms (GAs) and a Decision Tree for short-term module-scheduling. 2.6. Fuzzy Based A few researchers have used fuzzy logic to develop module advising systems. Goodarzi and Rafe [39] developed a fuzzy-based expert system for student advising. The proposed system was a web-based module that can be integrated with the university portal. The module fuzzifies the business rules and the GPA of the students to advise them on which modules should be taken in the following term. Adak [40] used Fuzzy techniques to recommend elective modules to students. It analyzed transcripts of students to extract fuzzy rules to relate the module to the students who intend to take them. Baloul and Williams [41] used the Order of Preference by Similarity to Ideal Solution (TOPSIS) technique to develop a fuzzy model for probation students to minimize the risk of taking the wrong modules in the early stage of their study. 3. The proposed approach This paper discusses the development of a fuzzy-based framework of an expert system for module advising. The proposed framework assumes that recommendations are based on three main elements which are: student-space (SS), module-space (MS), and lecturer-space (LS). SS contains the current student status which includes their accumulative average (if there is any academic warning), the closest abnormal academic status and to which extent it's close to that status. It also contains the knowledge domains of the academic program and the progress of the students in each domain. MS contains the average module mark for the last five years and the knowledge domains that the module belongs to. LS contains the average of the marks for the lecturer for all modules and the average of the lecturer's marks for each module. The three spaces and the result of the final calculation are then combined in a matrix called the case-space (CS). Fuzzy based association-rule learning is then used to extract the dominant rules to classify the consequent case for each eligible module if it was taken by the student for a specific semester. The fuzzy logic is used to handle the uncertainty involved in modeling the human decision and to provide a transparent and interpretable mechanism module taking risk estimation. AETiC 2021, Vol. 5, No. 1 17 www.aetic.theiaer.org The main purpose of the proposed framework is to estimate the consequent risk level (Low, Medium, and High) taking a specific module. The risk level is assigned based on a list of unwanted cases associated with the student failing or not progressing in the module. These cases include a GPA decrease, moving down from one GPA category to a lower one (degrading), receiving an academic warning, withdrawn from the academic program, withdrawn from the university, and graduation pending. In addition to the risk-level estimation, the framework provides the student with a justification (interpretation) for the estimated risk-level based on the student's situation, the targeted module, and the lecturer. The framework was developed based on real-life university data which included a historical record of enrolment associated, student's marks, cumulative GPA, module offerings, academic plans for the programs, and finally the knowledge domains for both the programs and the modules. The proposed approach is depicted in Figure 1 Figure 1: The proposed framework Fig. 2 shows the use case diagram of the proposed system. Figure 2: System use case diagram AETiC 2021, Vol. 5, No. 1 18 www.aetic.theiaer.org The proposed approach consists of six main steps: 3.1. Step-1: Creating the Three Spaces and Abnormal Case Matrix In this step, the student-space (SS), module-space (MS), lecturer-space (LS), and abnormal-case matrix are created as follows: 3.1.1. Step-1.A: Creating the three Spaces The three spaces SS, MS, and LS are created as shown in Fig. 3Error! Reference source not found.. The spaces are extracted from the university database. Figure 3: The three spaces and abnormal case matrix Student space: this space contains information about the student as shown in Table 1. Table 1: Student space Element Description 1. Student Id A unique identification number for the student 2. Cumulative GPA The accumulative average of the student out of 100 3. Student Average Mark for the Program Knowledge Domains This is a matrix that includes the knowledge domains of the student academic program and the average mark of the student for the modules which belong to each domain. 4. Potential Risk Situation This refers to the approximate abnormal student situation for the current situation of the student. (i.e., the current abnormal situation of the student boundary between two GPA categories, such as “very good” and “satisfactory”). Each abnormal-situation type is given a code and a percent of 100 to indicate the level of risk which is given to the code. 5. Current Student Situation This refers to the student's current abnormal situation. i.e., If the current abnormal situation of the student boundary between two GPA categories, such as “very good” and “satisfactory”, or the student's GPA has dropped. Each abnormal-situation type is given a code and a percent of 100 to indicate the level of risk which is given to the code. Module Space: this space contains information about the module as shown in Table 2. Table 2: Module space Element Description 1. Module Id A unique identification number of the Module 2. General Average of Grades The average grades for the students in the Module 3. Average grade for each program A matrix includes the academic programs which contain the module and the average module grade for each program 4. Average grade for each Lecturer who taught the module A matrix which includes the lecturers who taught the module and the average grade for the module given by each Lecturer 5. Fail Rate The percentage of students who failed the module AETiC 2021, Vol. 5, No. 1 19 www.aetic.theiaer.org Lecturer Space: This space contains information about the module as shown in Table 3. Table 3: Lecturer space Element Description 1. Lecturer Id A unique identification number of the lecturer 2. General Average of Grades The average grade of the lecturer 3. Fail Rate The percentage of students who failed the modules taught by the lecturer. 3.1.2. Step-1.B: Creating Abnormal Case Matrix In this sub-step, the abnormal-case matrix is created by selecting the student id, module id, lecturer id, and the student abnormal situation for abnormal cases in the student enrolment records from the university database. The abnormal case is usually indicated by a flag or a symbol in the database as shown in Figure 4. Figure 4: Abnormal-case matrix with case mapping 3.2. Step-2: Case Risk Analysis and Calculation For each problematic case in the dataset, a risk weight is calculated for the three spaces: SS, MS, and LS and the associated abnormal situation. The product of this step is shown in Fig. 5. The risk was calculated for the three spaces using equations developed based on common sense and the opinion of field experts such as the academic registry members, lectures, and head of departments. The results than have been presented to the experts and evaluated by them. Figure 5: Case-Risk Analysis and Calculation Risk estimation of student-space: 𝑺𝑹(𝑺𝑺) = (𝟏𝟎𝟎 − ( 𝑮𝑷𝑨(𝑺𝑰𝑫) + 𝑲𝑫𝑨(𝑴𝑰𝑫,𝑺𝑰𝑫) 𝟐 ) + 𝑷𝑨𝑪 + 𝑪𝑨𝑪)/𝟑 Equation 1: Risk estimation of student-space Where SR is the risk estimation of the student space in relation to the current case. It takes the student space as input and returns a percentage value out of 100. SS is the student space. GPA is the student cumulative GPA, SID is the student Id, KDA is the student average mark of the knowledge domain to which the module belongs. PAC is the scaled weight (out 100) of the potential abnormal situation. CAC is the scaled weight (out 100) of the current abnormal situation of the student. Risk estimation of module-space: AETiC 2021, Vol. 5, No. 1 20 www.aetic.theiaer.org 𝑴𝑹(𝑴𝑺) = (𝟏𝟎𝟎 − 𝑴𝑨𝑮(𝑴𝑰𝑫) + 𝑴𝑷𝑨(𝑴𝑰𝑫,𝑷𝑰𝑫) + 𝑴𝑳𝑨(𝑴𝑰𝑫,𝑳𝑰𝑫) 𝟑 + 𝑴𝑭𝑹(𝑴𝑰𝑫))/𝟐 Equation 2: Risk estimation of module-space Where MR is the risk estimation of the module in relation to the current student situation. It takes the Module Space (MS) as an input and returns a percentage out of 100, MS is the module space. MAG is the general average mark of the module. MID is the Module identification number; MPA is the average module mark for the student's program. PID is the identification number of the student's academic program. LID is the lecturer identification number. CAI is the average mark of the Module for the Lecturer. MFR is the Module fail rate which is a percentage of the students who failed the Module. Risk estimation of lecturer-pace: 𝑳𝑹(𝑳𝑺) = ((𝟏𝟎𝟎 − 𝑳𝑨𝑮(𝑳𝑰𝑫)) + 𝑳𝑭𝑹(𝑳𝑰𝑫))/𝟐 Equation 3: Risk Estimation of Lecturer Space Where LR is the risk estimation that comes from the lecturer side out 100, LS is the lecturer-space. LAG is the general average mark of the lecturer. LID is the lecturer-identification number. LFR is the lecturer fail rate which is a percentage of the students who failed the modules taught by the lecturer. Situation risk (SR) estimation: The resulting situation types were provided to a set of experts (i.e., academic-registry staff, advisors, and heads of department) and they were asked to assign a specific weight which reflects the risk level. The risk level takes a value between 1 and 100. The net risk for each situation type is calculated by taking the average of the weights given by the experts. 3.3. Step-3: Linguistic Labelling In this step, the final risk weights of the three spaces and the resulting situation are given linguistic labels of H-high, M-medium, and L-low. These linguistic labels are generated using Type- 1 and based on predefined fuzzy sets based on the Mendel Wang method [10]. In the proposed approach, the member function of the fuzzy sets uses a triangle shape as shown in Figure 6 and is based on three parameters {a, b, c} as shown in Equation 4. 𝑇𝑟𝑖𝑎𝑛𝑔𝑙𝑒(𝑥;𝑎;𝑏;𝑐) = { 0, 𝑥 ≤ 𝑎. 𝑥 − 𝑎 𝑏 − 𝑎 , 𝑎 ≤ 𝑥 ≤ 𝑏. 𝑐 − 𝑥 𝑐 − 𝑏 , 𝑏 ≤ 𝑥 ≤ 𝑐. 0, 𝑐 ≤ 𝑥. Equation 4: Triangle fuzzy set Where the parameters {a, b, c} (with a<b<c) determine the x coordinates of the three corners of the underlying triangular MF as shown in Fig.r (6). Figure 6: Fuzzy member function for input and output variables The outcome of this step is a linguistic label of H-high, M-medium, and L-low for each of the three spaces and the resulting situation associated with each abnormal case in the student record as shown in Figure 7. AETiC 2021, Vol. 5, No. 1 21 www.aetic.theiaer.org Figure 7: Case risk linguistic labels 3.4. Step-4: Fuzzy Rule-Base Creation To create the fuzzy rule base, in this step and for each abnormal case, the linguistic labels resulting from the previous step are associated with each other in a form of antecedents and consequents (If-> Then) where the linguistic labels of the three spaces are the antecedents and the linguistic label of the resulting situation is the consequent as shown in Figure 8. SS MS LS Result ↓ ↓ ↓ ↓ H H H → H Antecedents Consequent Figure 8: Rule base creating 3.5. Step-5: Fuzzy Rule-base Compression and Validation The rule base extracted in the previous step could contain many rules based on the size of the dataset. It may have a large amount of repetition (i.e., Repeated rules) and contradiction. The contradiction here means rules that have the same antecedent with different consequents such as shown in the following example in Table 4: Table 4: Contradictory rule patterns R(SS) (MS IS Result ↓ ↓ ↓ ↓ H H H → H H H H → L H H H → M To address these issues a five-fold cross-validation [42] process was used to train the rule base to summarize it to the most dominant unique patterns. In this validation method the dataset is divided into 5 equal-size sets (D1, .., D5). For each fold the following steps are applied: 1. One of the subsets Dn is selected as a testing set Tn and the other subsets are groups in one training set D^. 2. The rule compression(summarization) technique [43] is applied to the training set to produce a summarized rule-base 3. The summarized rule-base is applied to the test set to predict the risk level. 4. The predicted risk levels are compared with the actual risk levels from the dataset, to identify the accuracy of the rule-base in predicting the risk level The compression technique: This technique uses two measures for rule quality, which are “generality” and “reliability” and are used to identify rule patterns with maximum firing strength. Generality measures the number of instances in the extracted rule base which supports each rule pattern. Reliability measures the confidence level of each rule pattern [43]. AETiC 2021, Vol. 5, No. 1 22 www.aetic.theiaer.org In the proposed approach, generality is calculated using scaled fuzzy support, and the reliability is calculated by multiplying the scaled fuzzy support by the firing strength of the rule pattern. The support of the rule pattern refers to the number of the rules in the rule-base that the pattern represents. The “confidence” refers to the strength of a specific rule pattern against the contradictory patterns (i.e. because other rule patterns have the same antecedent and a different consequent) [43]. Fuzzy support is used to identify the unique rule-patterns with their occurrence in the extracted rule base. The fuzzy support can be scaled for each unique rule-pattern using the total number of instances in the rule-base which have the same consequent. Equation 5 shows how the scaled fuzzy support for a unique rule pattern is calculated: 𝑠𝑐𝐹𝑢𝑧𝑧𝑆𝑢𝑝(𝑅𝑃𝑙) = 𝑵𝑹𝑷𝒍 𝑵𝑹𝑷𝒍 + 𝑵𝑹�̂�𝒍 Equation 5: Fuzzy support Where l =1 to M, l is the index of the rule pattern, RPi is a unique rule pattern i.e. (H, H, H →M), 𝑵𝑹𝑷𝒍 is the number of instances in the extracted rule base supporting the rule pattern RPi. 𝑵𝑹�̂�𝒍is the number of instances in the extracted rule base which support other patterns with the same consequent i.e. ({M, H, H →M}, {M, M, H →M},…). The “confidence” is a measure of the uniqueness of the pattern as it indicates its strength against the contradictory patterns, which are the other patterns having the same antecedents but a different consequent. Equation 6 shows how confidence is calculated 𝑠𝑐𝐶𝑜𝑛𝑓(𝑅𝑃𝑙) = 𝑠𝑐𝐹𝑢𝑧𝑧𝑆𝑢𝑝(𝑹𝑷𝒍) 𝐶𝑜𝑹𝑷𝒍 Equation 6: Confidence Where 𝐶𝑜𝑹𝑷𝒍 is the number of instances in the extracted rule base which supports the contradictory rule- patterns with 𝑅𝑃𝑙. The final scaled-weight of the rule pattern is calculated as the product of the fuzzy support and the fuzzy confidence as shown in Equation 7. 𝑠𝑐𝑊𝑖 = 𝑠𝑐𝐹𝑢𝑧𝑧𝑆𝑢𝑝 × 𝑠𝑐𝐶𝑜𝑛𝑓 Equation 7: Final scaled weight Each of the generated unique rule patterns is assigned the scaled fuzzy weight measure scWi as follows: Table 5: The scaled fuzzy weight of the unique rule-patterns SS MS LS Result scWi ↓ ↓ ↓ ↓ H H H → H 0.35 H H H → L 0.02 H H H → M 0.12 The scaled fuzzy weight of the unique rule patterns is then used to select the rules with the highest weights among the contradictory-patterns set. The result of the compression process is a summarized rule-base that contains dominants and consistent rule patterns. The resulting rule base will be used later to estimate the risk level considering a specific module, lecturer, and student. 3.6. Step-6, Compressed Rule Base Selection The accuracy levels of the five compressed rule-bases resulting from the previous step are compared and the compressed rule-base with the highest accuracy level was selected to be included in the system. 4. Experiment and Results 4.1. Dataset A real-life university data was used as a dataset for this paper. The dataset included student- enrolment records, student-mark records, module records, academic program records, and records for lecturers. The dataset consisted of the records of 5000 students who faced problems during their AETiC 2021, Vol. 5, No. 1 23 www.aetic.theiaer.org studies. These problems included academic warnings, program withdrawals, program changes, and cumulative average grade down-gradings. 4.2. Creating the Three Spaces (SS, MS, and LS) The university uses an Oracle database for the student information system (SIS), which was used to create a view for each of the three space types. A view was also created to include the risk estimation for each space associated with the risk estimation provided by the expert. 4.3. Creating the Fuzzy Rule-Base To create the fuzzy rule base, MatLab Fuzzy Logic Tool Kit was used to create the fuzzy sets, which were then used to create the linguistic labels for the data extracted from the three spaces. 4.4. Fuzzy Rule-Base Training and Compression A five-fold cross-validation technique was used to train the rule-base, then, for each fold, the fuzzy rule-base was divided into two different subsets, which were training and testing. The 80%- 20% rule was used to identify the size of each set. The compression technique discussed in section (3) was applied on the training set to produce the compressed rule-base. The resulting compressed rule- base was applied to the test set to predict the risk level. The predicted risk-levels were compared with the actual risk-levels provided by the experts to determine the accuracy of the rule-base in predicting the risk level as shown in Table 6. Table 6: Rule base classifying accuracy sample Passed? Predicted Actual Situation Type L_Risk L_ID M_risk M_ID S_Risk S_ID 1 L L A M 210508 H 1002 M ****112 0 M H D L 210125 H 2511 H ****210 1 H H E H 030410 M 1002 H ****525 1 H H D H 120215 L 5147 L ****402 1 M M B M 130514 H 1002 M ****111 1 M M F M 010816 L 2142 M ****237 1 H H D H 070516 M 2151 H ****252 0 M H D L 220589 H 2101 H ****332 The accuracy level of the test results for each fold is shown in Table 7. As shown in the table the proposed approach achieved 92% classifying accuracy and 89.2 as an average classifying accuracy. The comparison between the proposed approach and the other approaches is not a straightforward task as the main focus of the proposed approach differs from those of the others. The proposed approach aims to address the risk estimation while some others focused on long-term planning or supported the process of the academic advisory as a whole. However, the results can still be compared to indicate the performance of the proposed approach. For example, the multi-agent system proposed in [22] achieved only 60% user satisfaction with its effectiveness. Besides, the user satisfaction with the intelligent advising system proposed in [4] was 77.8% which indicates at the performance proposed system in this paper is acceptable. Table 7: 5-Fold accuracy test results Fold Number of Rules Accuracy 1 24 88% 2 25 91% 3 27 92% 4 23 86% 5 23 89% Average 89.2% AETiC 2021, Vol. 5, No. 1 24 www.aetic.theiaer.org 4.5. Best Compressed Rule Base Selection The accuracy levels of the five compressed rules bases resulting from the previous step were compared and the compressed rule-base with the highest accuracy level was selected to be included in the system as shown in Table 8. Table 8: Best rule set I SS MS IS Result scWi ↓ ↓ ↓ ↓ 1 H H H → H 0.352 2 H H M → H 0.311 3 H H L → M 0.273 4 H M H → H 0.218 5 H M M → H 0.346 6 H M L → M 0.438 7 H L H → H 0.517 8 H L M → M 0.593 9 M L L → L 0.433 10 M H H → H 0.214 11 M H M → M 0.511 12 M H L → M 0.162 13 M M H → M 0.169 14 M M M → M 0.283 15 M M L → M 0.364 16 M L H → M 0.407 17 M L M → M 0.502 18 L L L → L 0.584 19 L H H → H 0.224 20 L H M → M 0.436 21 L H L → M 0.365 22 L M H → M 0.156 23 L M M → M 0.132 24 L M L → L 0.214 25 L L H → L 0.375 26 L L M → L 0.325 27 L L L → L 0.154 5. Conclusion This paper has presented a fuzzy-logic based approach of an expert system for module advising. The approach creates three spaces which are student-space, module-space, and lecturer space, and associate them with the abnormal situation type for each abnormal case. Each space and the abnormal situation type is then given a linguistic label using fuzzy sets. The linguistic labels are then associated with each other for each case to generate a fuzzy rule. The fuzzy rules for all cases are combined in one rule-base and then compressed to extract those rules with the highest firing strength. The fuzzy logic was used to handle the uncertainty implied in the human judgment of the student case as well as to provide a transparent and interpretable mechanism for predicting the risk level of enrolling a student on a module. The approach was developed using real-life university data and achieved an acceptable level of accuracy of 92% which is expected to improve as more data is captured and used to train the rule base. Although the achieved accuracy might sound like it needs a bit of enhancement but having that the machine learning approach used for training the rule base the accuracy is expected to enhance as more data instances are included. Also, the accuracy level acceptance depends on the problem and context, especially with cases in which are molding human decisions or behavior. Although there are different approaches have been proposed in this area, this paper introduces a novel mechanism that creates three spaces to estimate the risk of the student situation associated with a specific module. The three spaces which are namely: Students, Lecturer, and Module provide a multi-angle view on the student case and makes the estimation more realistic. Also, applying the fuzzy logic provides a means to handle the uncertainty included in the human decision-making regarding module selection. AETiC 2021, Vol. 5, No. 1 25 www.aetic.theiaer.org The fuzzy rule base also provides a transparent mechanism to make the recommendation which means it doesn’t provide the risk level only but also the justification of that recommendation based on the risk level of the three spaces. References [1]. Lam, S. S. and Choi, S. P. M. (2013). Implementing an efficient preference-based academic advising system. International Journal of Applied Management Science, 5(4), pp 297–321, DOI: 10.1504/IJAMS.2013.057110. [2]. Almutawah, K. A. (2014). A decision support system for academic advisors. International Journal of Business Information Systems, 16(2), pp 177. DOI 10.1504/IJBIS.2014.062837. [3]. Daramola, O., Emebo, O., Afolabi, I. and Ayo, C. (2014). Implementation of an Intelligent Course Advisory Expert System Cased-Based Course Advisory Expert System. In (IJARAI) International Journal of Advanced Research in Artificial Intelligence, 3(5), Available: www.ijarai.thesai.org [4]. Bendakir, N. and Aïmeur, E. (2006). Using association rules for course recommendation. AAAI Workshop - Technical Report, vol. WS-06-05, pp 31–40, Available: https://www.aaai.org/Papers/Workshops/2006/WS- 06-05/WS06-05-005.pdf [5]. O’Mahony, M. P. and Smyth, B. (2007). A recommender system for on-line course enrolment: An initial study. RecSys’07: Proceedings of the 2007 ACM Conference on Recommender Systems, pp 133–136. DOI: 10.1145/1297231.1297254. [6]. Sandvig, J. and Burke, R. (2005). Aacorn: A CBR recommender for academic advising. Technical Report TR05- 015, Available: http://facweb.cs.depaul.edu/research/techreports/TR05-015.doc. [7]. Ajanovski, V. V. (2017). Guided Exploration of the Domain Space of Study Programs. In 4th Joint Workshop on Interfaces and Human Decision Making for Recommender Systems (IntRS), Available: http://ceur- ws.org/Vol-1884/paper8.pdf. [8]. Young, P. (2017). A Recommender System for Personalized Exploration of Majors, Minors, and Concentrations *, RecSys 2017 Poster Proceedings 27 (2017), Available: http://ceur-ws.org/Vol- 1905/recsys2017_poster12.pdf. [9]. Sharma, A. and Tiwari, N. (2013). Design and Analysis of Fuzzy based Association Rule Mining. International Journal of Computer Applications and Information Technology, 3(I), pp 12, Available: https://www.ijcait.com/IJCAIT/31/314.pdf. [10]. Wang, L.-X. X. and Mendel, J. M. (1992). Generating Fuzzy Rules by Learning from Examples. IEEE Transactions on Systems, Man and Cybernetics, 22(6), pp 1414–1427. DOI: 10.1109/21.199466. [11]. Young-Jones, A. D., Burt, T. D., Dixon, S. and Hawthorne, M. J. (2012). Academic Advising: Does it Really Impact Student Success? In Quality Assurance in Education. vol. 21, Available: www.emeraldinsight.com. [12]. Mohamed, A. (2015). A decision support model for long-term course planning. Decision Support Systems, vol. 74, pp 33–45. DOI: 10.1016/j.dss.2015.03.002. [13]. Harding, B. (2008). Students with specific advising needs. In V. N. Gordon, W. R. Habley and T. J. Grites (Eds.), Academic advising: A comprehensive handbook (2nd ed.), pp. 189–203, Jossey-Bass. [14]. Laghari, M. S. (2014). Automated Course Advising System. International Journal of Machine Learning and Computing, pp 47–51. DOI: 10.7763/ijmlc.2014.v4.384. [15]. Roushan, T., Chaki, D., Hasdak, O., Chowdhury, M. S., Rasel, A. A., Rahman, M. A. and Arif, H. (2014). University course advising: Overcoming the challenges using decision support system. 16th Int’l Conf. Computer and Information Technology, ICCIT 2013, pp 13–18. DOI: 10.1109/ICCITechn.2014.6997355. [16]. Talal Al-Nory, M. (2012). Simple decision support tool for university academic advising. Proceedings of 2012 International Symposium on Information Technologies in Medicine and Education, ITME 2012, vol. 1, pp 53–57. DOI: 10.1109/ITiME.2012.6291245. [17]. Abdullah, A.-G., Sumaia, A.-G., Fadel, A., AL-Ruhaili, F. and Thamary, A.-A. (2012). An Expert System for Advising Postgraduate Students. International Journal of Computer Science and Information Technology, 3(3), pp 4529–4532. [18]. Albalooshi, F. and Shatnawi, S. (2010). HE-Advisor: A Multidisciplinary Web-Based Higher Education Advisory System. In Global Journal of Computer Science and Technology, 10(7), pp. 37-49. [19]. Mattei, N., Dodson, T., Guerin, J. T., Goldsmith, J. and Mazur, J. M. (2014). Lessons Learned from Development of a Software Tool to Support Academic Advising. In Proceedings of the 2014 Zone 1 Conference of the American Society for Engineering Education - “Engineering Education: Industry Involvement and Interdisciplinary Trends”, ASEE Zone 1 2014. IEEE Computer Society. DOI: 10.1109/ASEEZone1.2014.6820659. [20]. Shatnawi, R., Althebyan, Q., Ghalib, B. and Al-Maolegi, M. (2014). Building A Smart Academic Advising System Using Association Rule Mining. arXiv, Available: http://arxiv.org/abs/1407.1807 AETiC 2021, Vol. 5, No. 1 26 www.aetic.theiaer.org [21]. Lin, F., Dunwei, S. L., Frank, W., Kinshuk, Z. and Mcgreal, R. (2007). e-Advisor: A Multi-agent System for Academic Advising, Available: http://io.acad.athabascau.ca/~oscarl/pub/ABSHL2007.pdf [22]. Huang, L., Wang, C. D., Chao, H. Y., Lai, J. H. and Yu, P. S. (2019). A Score Prediction Approach for Optional Course Recommendation via Cross-User-Domain Collaborative Filtering. IEEE Access, vol. 7, pp 19550– 19563, DOI: 10.1109/ACCESS.2019.2897979. [23]. Asadi, S. and Shokrollahi, Z. (2019). Developing a Course Recommender by Combining Clustering and Fuzzy Association Rules. Journal of AI and Data Mining, 7(2), pp. 249–262. DOI: 10.22044/jadm.2018.6260.1739. [24]. Nafea, S. M., Siewe, F. and He, Y. (2019). A Novel Algorithm for Course Learning Object Recommendation Based on Student Learning Styles. Proceedings of 2019 International Conference on Innovative Trends in Computer Engineering, ITCE 2019, pp 192–201, DOI: 10.1109/ITCE.2019.8646355. [25]. Chang, P. C., Lin, C. H. and Chen, M. H. (2016). A hybrid course recommendation system by integrating collaborative filtering and artificial immune systems. Algorithms, 9(3), DOI: 10.3390/a9030047. [26]. Newell, C. and Miller, L. (2013). Design and evaluation of a client-side recommender system. RecSys 2013 - Proceedings of the 7th ACM Conference on Recommender Systems, pp 473–474. DOI: 10.1145/2507157.2508220. [27]. Bydžovská, H. (2015). Are collaborative filtering methods suitable for student performance prediction? Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 9273, pp. 425–430. DOI: 10.1007/978-3-319-23485-4_42. [28]. Yao, S. and Huang, B. (2017). Beyond Parity: Fairness Objectives for Collaborative Filtering. Advances in Neural Information Processing Systems, vol. 2017-Decem, pp. 2922–2931, http://arxiv.org/abs/1705.08804. [29]. Xu, J., Xing, T. and van der Schaar, M. (2015). Personalized Course Sequence Recommendations. IEEE Transactions on Signal Processing, 64(20), pp. 5340 - 5352, DOI: 10.1109/TSP.2016.2595495. [30]. Koutrika, G., Bercovitz, B. and Garcia-Molina, H. (2009). FlexRecs: Expressing and combining flexible recommendations. SIGMOD-PODS’09 - Proceedings of the International Conference on Management of Data and 28th Symposium on Principles of Database Systems, pp 745–757. DOI: 10.1145/1559845.1559923. [31]. Keston, L. and Goodridge, W. (2015). AdviseMe: An Intelligent Web-Based Application for Academic Advising. International Journal of Advanced Computer Science and Applications, 6(8). DOI: 10.14569/ijacsa.2015.060831. [32]. Engin, G., Aksoyer, B., Avdagic, M., Bozanli, D., Hanay, U., Maden, D. and Ertek, G. (2014). Rule-based expert systems for supporting university students. Procedia Computer Science, vol. 31, pp 22–31. DOI: 10.1016/j.procs.2014.05.241. [33]. Hashemi, R. R. and Blondin, J. (2010). SASSY: A petri net based student-driven advising support system. ITNG2010 - 7th International Conference on Information Technology: New Generations, pp 150–155. DOI: 10.1109/ITNG.2010.57. [34]. Ayman, M. (2011). A Prototype Student Advising Expert System Supported with an Object-Oriented Database. International Journal of Advanced Computer Science and Applications, 1(3). DOI: 10.14569/specialissue.2011.010316. [35]. Nambiar, A. N. and Dutta, A. K. (2010). Expert system for student advising using JESS. ICEIT 2010 - 2010 International Conference on Educational and Information Technology, Proceedings, vol 1. DOI: 10.1109/ICEIT.2010.5607701. [36]. Thanh Binh, N., Thi Anh Duong, H., Hieu, T., Duc Nhuan, N. and Hong Son, N. (2008). An integrated approach for an academic advising system in adaptive credit-based learning environment. VNU Journal of Science, Natural Sciences and Technology, vol. 24, pp 110–121, Avaiable: https://repository.vnu.edu.vn/bitstream/11126/4690/3/TC_02468.pdf. [37]. Sobecki, J. and Tomczak, J. M. (2010). Student courses recommendation using ant colony optimization. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 5991 LNAI(PART 2), pp 124–133, DOI: 10.1007/978-3-642-12101-2_14. [38]. Abdulwahhab, R. S., Al Makhmari, H. S. and Al Battashi, S. N. (2015, March 12). An educational web application for academic advising. 2015 IEEE 8th GCC Conference and Exhibition, GCCCE 2015, DOI: 10.1109/IEEEGCC.2015.7060084. [39]. Goodarzi, M. H. and Rafe, V. (2012). Educational Advisor System Implemented by Web-Based Fuzzy Expert Systems. Journal of Software Engineering and Applications, 05(07), pp 500–507, DOI: 10.4236/jsea.2012.57058. [40]. Adak, M. F., Yumusak, N. and Taskin, H. (2016, April 28). An elective course suggestion system developed in computer engineering department using fuzzy logic. 2016 International Conference on Industrial Informatics and Computer Systems, CIICS 2016. DOI: 10.1109/ICCSII.2016.7462394. AETiC 2021, Vol. 5, No. 1 27 www.aetic.theiaer.org [41]. Baloul, F. M. and Williams, P. (2013). Fuzzy academic advising system for on probation students in colleges of applied sciences - OMN. Proceedings - 2013 International Conference on Computer, Electrical and Electronics Engineering: “Research Makes a Difference”, ICCEEE 2013, pp. 372–377, DOI: 10.1109/ICCEEE.2013.6633965. [42]. Moss, H. B., Leslie, D. S. and Rayson, P. (2018). Using J-K fold Cross Validation to Reduce Variance When Tuning NLP Models. arXiv, vol. 1, pp 2978–2989, Available: http://arxiv.org/abs/1806.07139. [43]. Wu, D., Mendel, J. M. and Joo, J. (2010). Linguistic summarization using IF-THEN rules. 2010 IEEE World Congress on Computational Intelligence, WCCI 2010, DOI: 10.1109/FUZZY.2010.5584500. © 2020 by the author(s). Published by Annals of Emerging Technologies in Computing (AETiC), under the terms and conditions of the Creative Commons Attribution (CC BY) license which can be accessed at http://creativecommons.org/licenses/by/4.0. 
work_2o6cjbbw2bdj7nsiozoi6ob4ku	----	A novel expert system for objective masticatory efficiency assessment RESEARCH ARTICLE A novel expert system for objective masticatory efficiency assessment Gustavo Vaccaro 1☯*, José Ignacio Peláez2,3☯, José Antonio Gil-Montoya4☯ 1 International Postgraduate School, School of Dentistry, Granada University, Granada, Spain, 2 Department of Languages and Computer Sciences, University of Malaga, Malaga, Spain, 3 Prometeo Project, National Secretary of Higher Education, Science, Technology and Innovation (SENESCYT), University of Guayaquil, Guayaquil, Ecuador, 4 Gerodontology Department, School of Dentistry, Granada University, Granada, Spain ☯ These authors contributed equally to this work. * fabianvaccaro@correo.ugr.es Abstract Most of the tools and diagnosis models of Masticatory Efficiency (ME) are not well docu- mented or severely limited to simple image processing approaches. This study presents a novel expert system for ME assessment based on automatic recognition of mixture patterns of masticated two-coloured chewing gums using a combination of computational intelligence and image processing techniques. The hypotheses tested were that the proposed system could accurately relate specimens to the number of chewing cycles, and that it could identify differences between the mixture patterns of edentulous individuals prior and after complete denture treatment. This study enrolled 80 fully-dentate adults (41 females and 39 males, 25 ± 5 years of age) as the reference population; and 40 edentulous adults (21 females and 19 males, 72 ± 8.9 years of age) for the testing group. The system was calibrated using the fea- tures extracted from 400 samples covering 0, 10, 15, and 20 chewing cycles. The calibrated system was used to automatically analyse and classify a set of 160 specimens retrieved from individuals in the testing group in two appointments. The ME was then computed as the predicted number of chewing strokes that a healthy reference individual would need to achieve a similar degree of mixture measured against the real number of cycles applied to the specimen. The trained classifier obtained a Mathews Correlation Coefficient score of 0.97. ME measurements showed almost perfect agreement considering pre- and post-treat- ment appointments separately (κ� 0.95). Wilcoxon signed-rank test showed that a com- plete denture treatment for edentulous patients elicited a statistically significant increase in the ME measurements (Z = -2.31, p < 0.01). We conclude that the proposed expert system proved able and reliable to accurately identify patterns in mixture and provided useful ME measurements. PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 1 / 20 a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Vaccaro G, Peláez JI, Gil-Montoya JA (2018) A novel expert system for objective masticatory efficiency assessment. PLoS ONE 13 (1): e0190386. https://doi.org/10.1371/journal. pone.0190386 Editor: Marco Magalhaes, University of Toronto, CANADA Received: August 13, 2016 Accepted: December 13, 2017 Published: January 31, 2018 Copyright: © 2018 Vaccaro et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All source code files and MEPAT dataset will be available at the GitHub repository at: https://github.com/fabianvaccaro/ perceptodent. Funding: This work was funded by the Secretaria Nacional de Educación, Ciencia y Teconologı́a (SENESCYT) of the Government of Ecuador, with budget allocation No. 0099-SPP, http://www. educacionsuperior.gob.ec. Competing interests: The authors have declared that no competing interests exist. https://doi.org/10.1371/journal.pone.0190386 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0190386&domain=pdf&date_stamp=2018-01-31 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0190386&domain=pdf&date_stamp=2018-01-31 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0190386&domain=pdf&date_stamp=2018-01-31 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0190386&domain=pdf&date_stamp=2018-01-31 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0190386&domain=pdf&date_stamp=2018-01-31 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0190386&domain=pdf&date_stamp=2018-01-31 https://doi.org/10.1371/journal.pone.0190386 https://doi.org/10.1371/journal.pone.0190386 http://creativecommons.org/licenses/by/4.0/ https://github.com/fabianvaccaro/perceptodent https://github.com/fabianvaccaro/perceptodent http://www.educacionsuperior.gob.ec http://www.educacionsuperior.gob.ec Introduction Objective evaluation of the masticatory function Health care services for the elderly and physically disabled population are ever-increasing chal- lenges where practitioners are required to evaluate the functional impairments of individuals in faster and more accurate ways while using less invasive methods. One approach to this mat- ter is the objective evaluation of the human mastication, which is a complex biomechanical process that involves coordinated movements of the jaw, tongue, lips, and cheek; and one of the main functions of the stomatognathic system. [1]. Objective mastication assessment can be performed in two ways: firstly, by quantifying the changes that the food has suffered during mastication, i.e. the Masticatory Performance; and secondly, by calculating the number of chewing strokes that would be required to achieve a certain degree of food degradation, i.e. the Masticatory Efficiency [2]. The Masticatory Perfor- mance (MP) has been defined as: a measure of the comminution of food attainable under stan- dardized testing conditions [3]; and is considered an objective indicator of oral functional capabilities, widely used to measure the impact of dental treatments, to assess levels of disabil- ity and orofacial functional impairments following stroke [4,5], and has also been associated with malnutrition risk [6]. On the other hand, the Masticatory Efficiency (ME) has been origi- nally defined as the number of extra chewing strokes needed by the patient to achieve the same pulverization as the standard person [7]; however, the strict measurement of the ME is pres- ently in disuse, mainly because patients with impaired mastication would need to masticate for very large periods of time. Furthermore, it is important to notice that several studies used the terms MP and ME interchangeably while referring exclusively to the MP. Current MP assessment techniques are based on the objective quantification of the degra- dation of a test-food subjected to mastication. The degradation level is determined by measur- ing a property (colour, weight, median particle size, chemical concentration, etc.) of a piece of natural or artificial food (e.g. Optosil/Optocal TM , peanuts, ham, chewing gums, paraffin blocks, carrots, jelly gums, etc.), where the property is prone to changes related to the number of chewing strokes. The fastest and easiest routine for objective MP assessment is the mixture quantification of a two-coloured cohesive specimen subjected to mastication [8–10]. In a mixing test a test-food specimen is formed by two differently-coloured layers of chewing gum or paraffin stacked together. Previous studies suggest that there are similarities among the visual characteristics of chewing gums masticated for the same number of chewing strokes when considering young and healthy human subjects [11]. These similarities have allowed experts to subjectively iden- tify the mixture using comparison tables. An example set of masticated specimens for 3, 9, 15, and 25 chewing cycles is shown in Fig 1; where it is possible to notice that the red and white layers are mixed, to some extent, in a regular fashion. The amount of mixture reached with each chewing cycle would depend on the masticatory function of the individual and on the structural characteristics of the specimen such as the size, thickness, density, hardness, viscos- ity, and tinctures used for colouring. Several studies have proposed simple digital image analysis approaches for mixture quanti- fication that are more precise than visual inspection techniques, and modern mixing tests cur- rently focus on these kinds of procedures [8]. The first attempt to measure the mixture of food using digital image analysis employed several custom-made algorithms, but these were not fully described, hence not possible to replicate [12]. Later on, the magic wand tool of the Adobe Photoshop Elements1 software was used to select and count the pixels corresponding to the regions of the masticated bolus that were not mixed [8]. The “unmixed fraction” of the specimen was manually calculated using a Microsoft Excel1 spreadsheet (Microsoft Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 2 / 20 https://doi.org/10.1371/journal.pone.0190386 Corporation, One Microsoft Way, Redmond, WA, USA). The measurements provided by this procedure were highly affected by the user, meaning high variability of the results; also, it was time-consuming and difficult for clinical settings. These difficulties promoted the creation of a specialized tool called ViewGum (dHAL Soft- ware. Kifissia, Greece, www.dhal.com), which has been receiving special attention from the dental and medical communities because of its ease of usage [9,13]. ViewGum uses the Bai and Sapiro segmentation algorithm to isolate the bolus from the background of the image, and measures the Circular Standard Deviation of the Hue channel (SDHue) in the HSI colour space; however, SDHue measurements may not to be suitable for white-coloured chewing gums [9,11], thus limiting the range of potential test-foods. In a different work, the Wolfram Mathematica1 software (Wolfram Research, Champaign, IL, USA) was used to compute the custom visual feature “DiffPix” for mixing quantification [10]; nevertheless, neither the seg- mentation, nor filtering, nor the feature extraction procedure itself were clearly exposed, so the DiffPix computation approach is not available for reproduction. The problematics of masticatory performance assessment The MP quantification techniques based on digital image processing retain various weak- nesses. Firstly, most of the algorithms or tools employed for this task are not well-documented Fig 1. Example set of masticated chewing gums. https://doi.org/10.1371/journal.pone.0190386.g001 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 3 / 20 http://www.dhal.com/ https://doi.org/10.1371/journal.pone.0190386.g001 https://doi.org/10.1371/journal.pone.0190386 [8,9,11–13]. Also, there is neither consensus about which properties of the test-food should be monitored, nor scales for the mixture. The mixture can be quantified by numerous features of the resultant bolus [9,11,14]; however, existing MP measurement approaches rely on a single feature as the MP indicator, thus leading to high variability in the measurements and severely limiting the assessment methodology to a narrow set of test-foods [11]. Furthermore, MP assessment techniques require specialized training, equipment, and considerable amounts of time. For all these reasons, a comprehensive and integral reference framework for assessing the mixture in two-coloured chewing gums is needed. Proposed approach and purpose of this work This study explores the possibility to accurately measure the level of mixture in two-coloured chewing gums subjected to mastication within a comprehensive reference scale (0 to 100%). Consequently, important questions arise: what aspect of the masticated bolus should be con- sidered as the mixture indicator? And, from other point of view: how do specimens mixed by 0%, 25%, 50% (and so on) look like? These are not easy questions, as there are limitless ways to describe a digital image; and on the other hand, mastication is an erratic process, thus samples mixed under similar conditions would surely present noticeable differences. To overcome these difficulties, this study redefines the MP and ME under the premise that it is possible to identify patterns in the visual characteristics of masticated two-coloured chew- ing gum specimens when considering a young and healthy reference population. Firstly, we have redefined the MP for mixing tests as “the set of measurements that characterize the state of a sample subjected to a given number of chewing strokes”, thus extending the original defi- nition to include more than one feature. On the other hand, the MP of a given specimen would not suffice to achieve a diagnosis of the masticatory function of the patient, because a reference dataset is needed for comparison. Therefore, we have also redefined the ME for mix- ing tests as “the equivalent number of chewing strokes that an individual from a healthy refer- ence population would need to achieve a similar degree of mixture under controlled experimental conditions measured against the known number of chewing strokes applied to the sample”. The key aspect of this new ME definition is the calculation of the number of chewing strokes that would produce a similar MP outcome for the reference population; there- fore, the challenge of evaluating the masticatory function of an individual can be summarized as a classification problem where the MP (observed state) of a masticated specimen is classified into a category represented by the number of chewing strokes needed by the reference popula- tion. Mathematically, the ME can be represented as: ME ¼ P T ð1Þ where ME � 0, P is the number of chewing strokes needed for an individual from a healthy reference population (P2 N,), and T is the known number of chewing strokes applied to the sample (T2 N, T > 0). Consequently, a specimen with a ME of 0 implies a total absence of mixture, while a ME of 1 implies a normal level of mixture, i.e. comparable to what a reference healthy person would achieve. Furthermore, an ME > 1 implies that the diagnosed individual chews better than the average healthy person. The ME can also be expressed as a percentage for easier interpretation and easily associated with linguistic tags as shown in Table 1. The issue of evaluating multiple visual characteristics and performing an accurate classifica- tion of the sample can be efficiently solved by the application of computational intelligence techniques for pattern identification and automatic classification. Therefore, the aim of this paper is to present and validate a novel expert system for mixture patterns recognition in Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 4 / 20 https://doi.org/10.1371/journal.pone.0190386 young and healthy individuals, and objective Masticatory Performance efficiency assessment in masticatory-compromised individuals. The hypotheses tested in this work are: 1. The proposed system can accurately classify masticated two-coloured chewing gum speci- mens into the corresponding group represented by the number of chewing cycles applied. 2. The proposed system is able to identify differences between the patterns of mixture of eden- tulous individuals prior and after treatment with complete removable dentures. Materials and methods Knowledgebase The proposed expert system involves the identification of mixture patterns in two-coloured chewing gums; these patterns are the basis of a classification procedure to compute the P score (see Eq 1) of new specimens obtained from masticatory-compromised individuals. However, it is important to notice that the structural characteristics of a chewing gum brand are related to its visual characteristics; besides, a brand’s availability may not be the same worldwide. To overcome these problems the proposed system comprises two main components, as shown in Fig 2: a calibration stage oriented to identify patterns in the mixture of a reference population for a selected chewing gum brand; and a diagnosis stage oriented determine the ME of a patient using the same brand of chewing gums. These are related by an auxiliary component called Masticatory Efficiency and Performance Assessment Technique (MEPAT) which con- tains the information generated from the calibration in order to accurately perform mastica- tory assessment tests. Calibration The calibration stage aims to identify patterns in the visual characteristics of masticated two- coloured chewing gums by analysing a broad distribution of reference samples. This process can be resource-intensive and time-consuming because it requires the participation of various reference individuals and the analysis of multiple samples per participant. In this work, the cal- ibration stage was performed in the Faculty of Dentistry of the University of Guayaquil, Ecua- dor; and was orchestrated by four trained clinicians. The calibration process comprises the test-food selection, reference population selection, sample retrieval, digitization, segmentation, feature extraction, feature selection, machine learning, and classifier validation steps; which are described in the following sections. Test-food selection. The test-food considered for this approach was a chewing gum wafer composed of two differently coloured layers. These colours were different enough to permit an adequate interpretation of the level of mixture with a mere visual inspection. Previous studies have proposed red-blue[8], green-blue [13], red-white [11], among other colour combinations. Table 1. Example of linguistic tags associated to Masticatory Efficiency levels. Linguistic tag Masticatory Efficiency level Totally impaired 0% Impeded 25% Limited 50% Adequate 75% Normal 100% Better than the norm Greater than 100% https://doi.org/10.1371/journal.pone.0190386.t001 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 5 / 20 https://doi.org/10.1371/journal.pone.0190386.t001 https://doi.org/10.1371/journal.pone.0190386 In this work, the selected test-food was composed of two flavours of Trident1 chewing gums: watermelon (red dye) and spearmint (white dye); which are commercially available in Ecuador at the time of the experiment. The chewing gum strips came in the form of individually wrapped strips measuring 2.5 × 9 × 38 mm. Each specimen was formed by manually stacking the two pieces of chewing gum. Reference population selection. A total of eighty volunteers (N1 = 80) were recruited for the reference group (G1): 41 females aged 25 ± 4.2 years; and 39 males aged 25 ± 5.8 years. Sub- jects were students from the Faculty of Dentistry of the University of Guayaquil, Ecuador. The inclusion criteria were being 18 to 35 years old, having at least 28 natural teeth, Angle Class I occlusion, and a DMFT score of 2 or less. Exclusion criteria were TMJ dysfunction symptoms, orofacial pain, bruxism, tooth wear, and the presence of fixed or removable orthodontic appli- ances. Written informed consent was obtained from all participants. Formal approval through the Ethical Committee for Human and Animal Experimentation of the University of Guaya- quil was obtained for this experiment. Sample retrieval. An operator instructed the subjects in G1 to masticate five specimens by 0, 5, 10, 15, 20 chewing cycles correspondently, and silently counted the number of cycles. Sub- jects rested for 30–60 seconds between mastication sessions to prevent fatigue. The subjects expelled the resultant boluses and the operator located each bolus between two sheets of trans- parent film intended for document lamination, and immediately flattened them to a 1mm thick wafer using a wheel-driven screw press. The flattening step is important because this assembly provides resistance to manipulation and is ideal digitization [9]. A total of 400 sam- ples were collected this way. Fig 2. Proposed mastication assessment solution model. https://doi.org/10.1371/journal.pone.0190386.g002 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 6 / 20 https://doi.org/10.1371/journal.pone.0190386.g002 https://doi.org/10.1371/journal.pone.0190386 Regarding the selection of the set of chewing cycles, previous studies indicate that this list must include 20 chewing strokes as the mean human mastication duration; and no more than 50 chewing strokes because of fatigue of the masticatory muscles [15]. On the other hand, 0 chewing strokes represents the “pre-mastication state” of the specimen, and it is obtained by retrieving the food specimen right after the subject introduces it into the mouth. This is impor- tant because saliva may play an important role during the image analysis process. Nevertheless, the task of selecting the adequate number of chewing cycles is currently a point of discussion among scholars. Digitization. All specimens were individually scanned on both sides using a Canoscan Lide 2201 flatbed scanner (300 dpi, standard calibration parameters for colour images). A flatbed scanner was chosen against digital photography because empirical experimentation showed that scanned images offered better image quality and even illumination. However, empirical tests during the calibration stage exhibited that digital photography under controlled conditions may also provide adequate images. Segmentation. The resultant digital images were segmented to isolate the area of the bolus against the background. Specimens often had irregular shapes, vague boundaries, and very heterogeneous coloration; therefore, these conditions made precise automatic segmenta- tion a nontrivial task. Active contour modelling have been used previously, but required addi- tional effort for the clinicians [13]; so a more general approach was needed. Watershed algorithms can serve to transform the brightness levels of the image into a set of clusters or “water pools” by identifying relevant markers on the image and filling their surroundings [16]; in this regard, improvements of the Watershed algorithm has been proposed in the literature which are specially tailored for complex medical imaging [17]. On the other hand, Watershed algorithms can become highly complex for colour image segmentation and usually identify many clusters per image, hence they often require post-processing methods. Another approach for this matter involves iterative shrinkage methods for the selection of a specific region in the image, which has been successfully used for complex medical imaging segmentation such as magnetic resonance imaging [18]; however, these are better suited for sparse level images with diffuse contours and may require supervised intervention to indicate the desired search region. In this regard, the proposed approach implemented a fully-automated colour-based seg- mentation algorithm, constructed upon the combination of Mean Shift [19,20], distance map, and K-Means classification algorithms [21]. Further details about this segmentation approach are detailed in S1 Appendix. This custom-made segmentation algorithm was used to divide the image into two regions: the bolus located in the centre of the image, which is considered as the Region of Interest (ROI); and the Background, surrounding the bolus. An example of the application of the proposed region classification process considering Mean Shift (MS) + dis- tance map (DM) + K-Means (KM) is shown in Fig 3, where it is possible to notice an improve- ment in the classification of regions, compared to the usage of KM, MS + KM, and DM + KM. Feature extraction and MP calculation. The proposed expert system approach considers the whole observed state of a specimen, i.e. the MP of the sample. Mathematically, the MP can be described as: MP ¼ðf 1 ; f 2 . . . ; fkÞ ð2Þ where fi represents the i th observed feature, and k represents the total number of features. It is important to notice that if just one feature is considered then it may lack of sufficient informa- tion about the sample; consequently, the MP comprises multiple features at once. Additionally, these features can be obtained from different extraction methodologies that are already Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 7 / 20 https://doi.org/10.1371/journal.pone.0190386 acknowledged by clinicians, e.g.: pixel counts, histogram analysis, etc. [10,13]. Under this premise, a custom set of feature extraction models was applied over the ROI of each pair of images corresponding to the same specimen. Feature extraction models included the mean and the absolute variance of the colour pixels; and the absolute variance, skewness, energy, entropy, and highest peaks of the colour histograms. These were computed for each compo- nent of the RGB, CIE-L�u�v� [22], HSI, and Normalized RGB colour spaces separately (see S2 Appendix). Additionally, the Circular Variance of the Hue channel of the HSI colour space was computed [9,13]. Consequently, the feature extraction stage considered a total of 121 fea- tures obtained from different image processing methodologies and colour space models (10 methods × 12 channels + CVOH). Heuristic extraction models such as Binary Particle Swarm Optimization or advanced texture characterization like Wavelet-based methods were not con- sidered for this experiment because of their high computational times compared to simple pixel and histogram feature extraction models; nevertheless, further improvement of the pro- posed solution should make extensive use of these kind of advanced characterization approaches [23,24]. The selection of the colour spaces was based on empirical experience. In this case the origi- nal images are represented in the RGB colour space, which is used for fast representation of 256 shades of Red, Green, and Blue. The CIE-L�u�v� is an easily computable transformation of the CIE XYZ (Tristimulus) colour space [25], extensively used in graphical computing. The HSI (Hue, Saturation, and Intensity) colour space is a cylindrical-coordinate representation of the points in the RGB with applications in computer vision [24], and previous studies suggest it is useful for characterizing the mixture of chewing-gums [9]. Finally, the Normalized RGB colour space is obtained from the RGB by a normalization procedure [26]; the influx of the Fig 3. Comparison of different automatic segmentation methods applied over chewing-gums identification. The goal of the segmentation process was to discriminate the chewing gum bolus in the centre of the image against the background. Three segmentations methods were employed: Mean Shift (MS), Distance Map (DM) and K-Means (KM). https://doi.org/10.1371/journal.pone.0190386.g003 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 8 / 20 https://doi.org/10.1371/journal.pone.0190386.g003 https://doi.org/10.1371/journal.pone.0190386 brightness is diminished by the normalization, so it is less susceptible to changes related to the light-source. Feature selection. The large number of features increases the chances of accurately char- acterizing a sample; however, it hiders the effectiveness of pattern recognition processes that will be applied in further stages. It is possible to discard some of the features by computing a relevancy score (q) associated to each extracted feature, computed as: qi ¼ jrðFiÞjþ gðFiÞ n 2 ! ! 2 0 B B B B B B B B B @ 1 C C C C C C C C C A ð3Þ where Fi represents the set of measurements obtained from extraction of the i th feature, ρ(Fi) is the coefficient of correlation with the number of chewing cycles calculated with Spearman’s rho, γ(Fi) is the amount of statistically different pairs of chewing strokes that the i th feature can discriminate, and n is the amount of different numbers of chewing strokes (n = | C |). The value γ(Fi) is calculated with ANOVA, considering the number of chewing strokes as the fixed factor, corrected with post hoc Bonferroni, and considering that: 0 � gðFiÞ� n 2 ! ð4Þ A feature can be discarded if the associated q score is lower than 0.5. The above-mentioned fea- ture selection process may serve to reduce the size of the Artificial Neural Network model that will be used for pattern recognition in this experiment. Additional dimensionality reduction can be achieved by applying principal component analysis (PCA) to the set of non-discarded features by converting the set of possibly correlated features into a set linearly uncorrelated principal components [27]. Nonetheless, previous experimentation showed that dimensional- ity reduction using PCA may not necessarily improve pattern recognition performance of this solution. In this regard, PCA results for the non-discarded features are shown in the Results section for comparison purposes. Machine learning. The information extracted from each specimen (S) during the calibra- tion phase was summarized as a 2-touple consisting of the MP of the specimen and the number of chewing strokes (t), such that: S ¼ðMP; tÞ ð5Þ Then, the specimens were grouped as: Si ¼fSjS ¼ðMP;xÞ^ x ¼ tig ð6Þ where Si represents the set of MPs of all specimens that were masticated by ti chewing strokes. The proposed methodology used an artificial neural networks (ANN) algorithm for pattern identification [28–31]. ANNs are rough mathematical models of biological neurons where electrical signals are represented as numerical values; they mimic some features of biological brains, especially the ability to learn, i.e. to acquire the ability of processing information in cer- tain patterns [32]. The chosen ANN architecture was the Multilayer Perceptron (MLP) [33]. The MLP maps a set of inputs to a set of desired outputs through multiple layers of nodes; Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 9 / 20 https://doi.org/10.1371/journal.pone.0190386 where each node is an artificial neuron. The basic MLP structure consists of an Input layer, an Output layer, and any number of Hidden layers [24,34–36]. In our case, the MLP structure consisted of k nodes in the Input layer (where k is the num- ber of features extracted from the specimens), 1 node in the Output layer (binary: true or false), and a variable number h of nodes in the Hidden Layer (k/3 � h� k). A set of MLPs was constructed for each Si with the task of determining if the sample was masticated by pi chewing strokes and considering different numbers of hidden neurons. The best number of neurons h in the Hidden layer was computed by sequentially increasing by 1 the value of h from k/3 to k, and executing 10 trainings per h value. Inputs and outputs were obtained by breaking the information (i.e., the S) of each sample in two: the MP was considered the input, and the p as the output. The entire data set was ran- domly divided in three groups for each training execution: 40% for the Training Group (TG), 30% for the Validation Group (VG), and 30% for the Testing Group (SG). Each MLP was trained using the data in the TG, and using the VG as a reference to stop overfitting [37]. Classifier validation. The proposed system fed each trained network with the MP inputs from the SG group, and compared the predicted outputs (the P score) against the known num- ber of chewing strokes. Then, the system computed the Matthews Correlation Coefficient (MCC) to assess the performance of a trained network for each execution [38]. The MCC is a measurement of the quality of binary classification that considers true and false positives and negatives, and can be regarded as a correlation coefficient between the observed and predicted classifications, where MCC = 1 represents a perfect prediction, MCC = 0 represents a random equivalent prediction, and MCC = -1 represents a complete disagreement between prediction and observations. The MCC was computed as follows: MCC ¼ TP � TN � FP � FN ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðTP þ FPÞ�ðTP þ FNÞ�ðTN þ FPÞ�ðTN þ FNÞ p ð7Þ where TP represents the number of true positives, TN represents the number of true negatives, FP represents the FP represents the number of False Positives, and FN represents the number of False Negatives. The system considered an MPL as suitable if its MCC was over 0.95, and choose the MLP with the highest MCC. The calibration stage would be considered unsuccess- ful if no suitable MLPs are found after iterating through all the possible h values. The system assembled the set of suitable MPLs for each pi as a unique classifier in the form of a binary cas- cade: new samples were classified for S1, then for S2, and so on. The system discarded any sam- ple that could not be classified into any Si group. The MEPAT The information obtained from the calibration stage includes details about the Test-Food, the features that best characterize it, relevant data about the experimental procedure, and the trained classifier. To perform diagnostic analyses over new specimens a clinician operator must follow the same sample retrieval procedure, utilize the same feature extraction methods, and feed the trained classifier with information about a new sample formatted in the right way. This paper proposes a new standardized representation of the information, procedures, and tools required to perform Masticatory Efficiency and performance diagnoses named Mas- ticatory Efficiently and Performance Assessment Technique (MEPAT). With this, the pro- posed solution explores the possibility to standardize the information resulting from a calibration execution following these objectives: 1. To be written in a comprehensive markup language. Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 10 / 20 https://doi.org/10.1371/journal.pone.0190386 2. To contain relevant information about the calibration experiment. 3. To contain all the information needed to execute a diagnosis over a new specimen. 4. To be portable between different users and devices. 5. To provide scalability for new feature extraction procedures. 6. To be open-source. The MEPAT follows a custom-made XML structure (see S1 File), which is software-inde- pendent, and can be easily stored and transferred. The MEPAT can be graphically represented as shown in Fig 4; and mathematically consists of a tuple: MEPAT ¼hTF; ES;CH;CLS;OP;PERi ð8Þ where TF represents the information about the test-food, ES represents the Experimental Set- tings, CH represents the set of selected features that characterize the test-food, CLS represents the trained classifier, OP represents the information about the Operator that orchestrated the calibration stage, and PER represents the performance of the overall MEPAT. On the other hand, additional information such as a Unique Identifier (UID), creation date, and upload date, may be used for synchronization purposes. Additional information about the MEPAT components can be found in S3 Appendix. The resultant information obtained from the cali- bration process was included in a MEPAT structure and used in further stages of this work. Diagnosis The ultimate purpose of the calibration stage was to provide the necessary information to per- form an adequate diagnosis of the masticatory function of an individual. To do so, the system employed a MEPAT to execute a single diagnosis mixing-test, which is defined as a calibrated methodology for assessing the masticatory function of an individual using a single specimen of a two-coloured chewing gum. Fig 4. Graphical representation of the MEPAT XML schema. https://doi.org/10.1371/journal.pone.0190386.g004 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 11 / 20 https://doi.org/10.1371/journal.pone.0190386.g004 https://doi.org/10.1371/journal.pone.0190386 MEPAT selection. This experiment considered the MEPAT that was created during the previous calibration execution because it contained all the information necessary to perform single diagnosis mixing-tests with the available resources at the University of Guayaquil in Ecuador. Clinical procedure and sample retrieval. A total of 40 volunteers (N2 = 40) were recruited for the testing group (G2): 21 females aged 73 ± 8.7 years; and 19 males aged 71 ± 9.1 years. Subjects were patients from the dental prosthetics clinic of the Faculty of Dentistry of the University of Guayaquil, Ecuador. The inclusion criteria were being older than 60 years old and complete edentulism. Exclusion criteria were TMJ dysfunction symptoms, orofacial pain, and severe cognitive impairment. Written informed consent was obtained from each subject after a full explanation of the research project. The Ethical Committee for Human and Animal Experimentation of the University of Guayaquil provided a formal approval for this experiment. Four samples were obtained from each patient: first, patients received two consecutive tests without wearing dental prosthesis; then, patients received the next two tests 30 days after com- plete removable dentures were fitted during a follow-up appointment. The overall procedure for sample retrieval for was conducted as follows: 1. An operator provided the patient with a specimen of the selected test-food. 2. The operator instructed the patient to masticate the test-food specimen by 20 chewing strokes on the preferred chewing side (notice that the largest number of chewing cycles dur- ing the calibration stage was 20). 3. The operator monitored the patient while silently counting the number of chewing strokes. 4. The masticated specimen was retrieved when 20 chewing cycles were achieved. 5. The operator put the specimen between two sheets of transparent film intended for docu- ment lamination. 6. The operator flattened the specimen to a 1mm thick wafer using a calibrated press. 7. The pressed wafer was scanned for both sides. Digital image analysis. The system segmented all the digital images obtained from the samples using the MS + DM + KM procedure (see S1 Appendix); then, the system extracted a set of features following the instructions stored in the CH component of the MEPAT. Classification. The proposed system extracted a set of features from each sample follow- ing the instructions in the CLS component of the MEPAT. Then, the classifier categorised each sample in one of the classes related to a number of chewing strokes. This classification provided the number of chewing strokes that a healthy reference individual would need to achieve a similar degree of mixture, i.e. the P score of the sample. Masticatory Efficiency quantification. The system computed the ME correspondent to each sample using Eq 1, considering T = 20 and the P score calculated in the Classification step. For instance, if a sample scored a P of 15 then the corresponding ME would be: ME = (15/20) × 100% = 75%. Statistical analysis Statistical analyses were performed on MATLAB 2015a (The MathWorks Inc., MA, USA) using the Statistic Toolbox. First, the Mathews Correlation Coefficient [38] was used for vali- dating the pattern identification performance for G1. Secondly, the inter-rater agreement Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 12 / 20 https://doi.org/10.1371/journal.pone.0190386 (consensus) of consecutive ME measurements of the G2 was verified using the Cohen’s Kappa statistic, considering the initial and follow-up appointments separately [39]. Thirdly, the differ- ences between the ME measured prior and after treatment with complete removable prosthe- ses (first appointment and follow-up appointment respectively) were evaluated using the Wilcoxon signed-rank test, considering the highest ME value from the two measurements. The performance of the proposed ANN classification approach was compared against a sin- gle-feature classification methodology. The goal was to verify the practicality of implementing a complex classification procedure instead of computing the ME directly from single MP mea- surements obtained from traditional methods. To do so we implemented a binary cascade to determine the closeness of each sample to the appropriate number of chewing strokes by com- puting its standard score (z-score) as the number of standard deviations by which the MP mea- surement is above the mean. The z-score has been used previously by Halazonetis et al. [13] to diagnose the state of the masticatory function by comparing the MP (CVOH, see S2 Appendix) measurements of a given sample against a known MP distribution. Mathematically, the z- score was computed as: zi;T ¼ mpi � mpT sT ð9Þ Where zi,T is the z-score of the i-th sample computed for T chewing strokes, mpi is the MP measured from a single feature, mpT is the mean MP value computed from the Training Group, and σT is the standard deviation. Then, samples from the Testing Group were pre-clas- sified as part to the T group if |zi,T| � 0.25; hence, the core binary classification performance per T group was computed using the MCC score. Finally, the samples were classified in the T group that provided the lowest absolute z-score; i.e., where the MP value was closest to the mean. Results Calibration stage This study retrieved a total of 400 specimens during the calibration stage (800 digital images). The complete calibration process required a combined total of 156.05 hours, which corre- sponded to 9 sessions of sample retrieval (~4 hours per session with 4 operators), and one ses- sion of image processing and classifier training (3.86 hours) which was performed on an Intel Core i7-5930K PC with 32GB of RAM. The Table 2 provides detailed information about the time required for the calibration phase; in this regard, the mastication, resting and digitization, and image processing steps accounted for 64.5%, 33%, and 2.5% of the calibration execution time respectively. A total of 35 features with a q score above 0.5 were selected as good mixture characterizers (calculated with Eq 3). On every case, the Kolmogorov-Smirnov test confirmed the normality Table 2. Distribution of time shares of the calibration stage for 400 samples. Execution time per sample in seconds: average (std. dev.) Total execution time in hours Time share Mastication 90.6 (19.1) 100.7 64.5% Digitization a 52.9 (4.2) 51.5 33.0% Image processing 3.5 (1.1) 3.9 2.5% Total 178.6 (5.8) 156.1 100.0% a Also includes the resting time https://doi.org/10.1371/journal.pone.0190386.t002 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 13 / 20 https://doi.org/10.1371/journal.pone.0190386.t002 https://doi.org/10.1371/journal.pone.0190386 of the distribution per chewing stroke (p < 0.05); significant differences among mixing states were confirmed with ANOVA corrected with post hoc Bonferroni (p < 0.05); and showed moderate-to-high correlation with the number of chewing strokes (| ρ | > 0.5; p < 0.05). The PCA executed over the set of selected features indicated that 3 principal components (PC1 to PC3) explained 91.301%, 4.488%, and 4.211% of the variance respectively. The Table 3 lists the selected features along their corresponding q scores and PCA factors rotated with Promax rotation method. Classifiers trained with all the 35 selected features performed slightly better than those trained solely with the first 3 principal components after 200 iterations, although no significant differences were found between training groups (p = 0.412). The classifier that showed the best Table 3. List of features selected as mixture state characterizers, showing the model, the colour space channel, the relevancy q score, and the PCA factors for the first 3 principal components. Feature extraction model Channel q PC1 PC2 PC3 A. Variance of the histogram H 0.841 0.99 -0.02 -0.04 Entropy of the histogram H 0.827 0.01 0.00 0.00 Value of the 1st peak of the histogram u 0.825 0.00 0.35 0.00 Value of the 1st peak of the histogram B 0.820 0.00 0.00 0.08 A. Variance of the pixels G 0.805 0.00 0.00 0.00 Value of the 1st peak of the histogram S 0.802 0.00 0.00 0.00 Value of the 1st peak of the histogram Rn 0.800 0.00 0.00 0.00 A. Variance of the pixels u 0.799 0.00 0.00 0.00 A. Variance of the pixels S 0.795 0.11 0.01 0.19 A. Variance of the pixels Rn 0.786 0.00 0.00 0.00 A. Variance of the pixels Gn 0.783 0.70 0.26 0.10 A. Variance of the pixels L 0.776 0.00 0.00 0.00 Value of the 1st peak of the histogram G 0.774 0.00 0.00 0.00 Value of the 1st peak of the histogram Gn 0.772 0.01 -0.05 0.25 A. Variance of the pixels B 0.772 0.01 -0.06 0.27 Skewness of the histogram H 0.769 0.00 -0.01 0.28 Mean of the pixels H 0.706 0.00 0.00 0.00 Value of the 1st peak of the histogram L 0.705 0.00 0.00 0.00 Value of the 1st peak of the histogram Bn 0.692 0.00 0.00 0.00 A. Variance of the pixels Bn 0.671 -0.01 -0.05 0.61 Entropy of the histogram R 0.657 0.00 0.00 0.00 Entropy of the histogram I 0.634 0.51 0.00 0.00 A. Variance of the histogram R 0.625 -0.02 -0.03 0.41 Entropy of the histogram u 0.625 -0.02 -0.08 0.40 Entropy of the histogram G 0.624 -0.03 -0.05 0.46 A. Variance of the histogram B 0.623 0.00 0.00 0.00 A. Variance of the histogram G 0.604 0.00 0.00 0.00 Entropy of the histogram S 0.600 0.00 0.00 0.00 A. Variance of the histogram I 0.599 -0.03 0.23 0.32 Entropy of the histogram L 0.598 0.00 0.00 0.00 Entropy of the histogram B 0.595 0.00 0.00 0.00 A. Variance of the histogram S 0.590 -0.02 0.93 -0.10 A. Variance of the histogram L 0.576 0.00 0.00 0.00 Entropy of the histogram Gn 0.571 0.00 0.01 0.00 Entropy of the histogram Rn 0.558 0.00 0.00 0.74 https://doi.org/10.1371/journal.pone.0190386.t003 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 14 / 20 https://doi.org/10.1371/journal.pone.0190386.t003 https://doi.org/10.1371/journal.pone.0190386 overall performance was trained using all the features listed in Table 3, with 16 neurons in the hidden layer (h = 16), and was obtained after 110 iterations. Further performance details of the resultant classifier are listed in Table 4. On the other hand, the core performances of single-feature classifiers are detailed in S4 Appendix. The name of each features is expressed using the corresponding Mixture Feature Code (MFC, see S3 Appendix). The MCC score was computed per group to better visualize the ability of the classifier to differentiate between numbers of chewing strokes. Finally, the global classification score (including T = 0, . . . 20) is presented. Diagnosis stage The complete diagnosis stage required a combined total of 6.23 hours, which corresponded to two sessions of sample retrieval (~2.6 hours per session with 4 operators), and two sessions of image processing and automatic classification (~ 12 minutes); with an average execution time per patient of 4.5 minutes. Some operators reported difficulties while counting the chewing cycles; in those cases, the operator asked an assistant for help and the final number of chewing cycles was determined by agreement. The Cohen’s Kappa statistic showed that repeated measurements of ME for the G2 group showed almost perfect agreement, considering pre- and post-treatment appointments sepa- rately (κ� 0.95). Furthermore, a Wilcoxon signed-rank test showed that a complete denture treatment for edentulous patients elicited a statistically significant increase in the ME measure- ments of the individuals (Z = -2.31, p < 0.01). Additionally, the Absolute Variance of the His- togram of the Hue channel (VhH) was evaluated separately for comparison reasons [11]; in this regard, complete denture treatment for edentulous patients elicited a statistically signifi- cant increase in VhH measurements (p < 0.001). The mean, median, and standard deviations of ME and VhH measurements are listed in Table 5 along with the corresponding linguistic tag associated to the ME level (see Table 1). Table 4. Performance details derived from the confusion matrix of the trained best trained classifier obtained from the calibration stage. Confusion matrix component Score Sensitivity 0.98 Specificity 0.99 Accuracy 0.99 MCC 0.97 https://doi.org/10.1371/journal.pone.0190386.t004 Table 5. Masticatory Efficiency (ME) and absolute variance of the histogram of the Hue (VhH) of edentulous individuals measured prior and after treatment with complete dentures. Statistic ME ME level tag VhH Prior treatment Mean 0.26 Impeded 10.26×106 Median 0.25 Impeded 9.56×106 Standard deviation 0.22 - 1897.01 Mode 0.50 Limited 10.11×106 After treatment Mean 0.71 Adequate 24.05×106 Median 0.75 Adequate 23.49×106 Standard deviation 0.23 - 3314.78 Mode 0.75 Adequate 23.99×106 https://doi.org/10.1371/journal.pone.0190386.t005 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 15 / 20 https://doi.org/10.1371/journal.pone.0190386.t004 https://doi.org/10.1371/journal.pone.0190386.t005 https://doi.org/10.1371/journal.pone.0190386 Discussion This paper introduced new definitions for ME and MP. The differences between ME and MP were clearly distinguished, and the MP was used as a component for the calculation of the ME. Additionally, a reference scale for the ME is presented for the first time. The calibration stage required most of the experimental time and resources, as it involved a complex (yet easier than traditional) clinical execution. On the other hand, the average execu- tion time needed to obtain a full ME diagnosis from a patient was 5 minutes, which is consid- erably fast for a clinical setting. The disparity between calibration and diagnosis stages was predictable, as the calibration stage required more samples and more complex computational processing steps. The visual features selected in this experiment as characterizers of the mixture were com- puted using various deterministic algorithms applied over a broad range of colour channels. Two feature reduction approaches were followed, first, a relevance-based method selected a total of 35 bare features, implying that the mixture may be assessed by more than a single fea- ture, and that some features are better than others at characterizing the mixture. The variance of the histogram of the Hue channel of the HSI colour space provided the best overall rele- vancy, while the entropy of the histogram of the normalized-red channel of the normalized RGB colour space provided the lowest acceptable relevancy. On the other hand, a PCA approach selected 3 PCs that accounted for 99.9% of the variance, thus suggesting that most of the features provided little-to-none new variance information to the model. For both feature selection approaches pattern identification and classifier training showed very high performance scores, with an MCC = 0.97 for the best case using the 35 bare features as inputs. This suggest that the proposed expert system was reliable at identifying mixture pat- terns for the selected test-food. On the other hand, results at S4 Appendix show that single- feature classifiers performed poorly in comparison to ANN-based classifiers for this purpose. The best single-feature classifier employed the EhH feature as the MP indicator with global MCC = 0.321. However, it is interesting to notice that many single-feature classifiers provided good core classification performance when tasked to identify samples masticated for 20 chew- ing strokes (T = 20); in this case Nhu, NhB, NhGn, Ehu, EhGn, VhGn, P2H, NhRn, NhL, NhG, EhH, V1H, NhS, NhI, Nhv, VhB scored an MCC � 0.5, where the Nhu obtained the highest core classification performance score with MCC = 0.660. Diagnosis stage results showed that ME of edentulous individuals were significantly higher after receiving treatment with complete dentures (p < 0.01), implying an increase in the masti- cation process outcome from “impeded” to “adequate” (see Table 1). The proposed methodology can be improved in future works by strengthening the feature extraction and selection processes; as one of the key factors when computing the ME via pattern recognition is the quality of the MP indicators and the amount of relevant informa- tion that these provide to the model. Therefore, more sophisticated mixture quantification approaches such as texture analysis and wavelet transformations may significantly improve the performance of the proposed system. Also, better feature selection approaches are may help to reduce the necessity of large datasets of samples during the training stage. It is possible to sustain that efforts to standardize MP evaluation in a worldwide scale may not be viable, as there are significant differences in digitization equipment and test-food avail- ability among countries. Differences in the digitization equipment may produce undesirable effects on the classifier outcome, although this phenomenon has not been profoundly studied, and the proposed calibration model can be easily adapted to handle different kinds of digitiza- tion devices. Additionally, choosing the right test-food was a crucial task, as it greatly influ- ences the outcome of the masticatory assessment [14]. A very useful list of specifications for Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 16 / 20 https://doi.org/10.1371/journal.pone.0190386 specimens aimed to use for a two-colour-mixing ability test was presented by Schimmel, et al (2015) [9]. Regarding this, we recommend that specimens must have two different colours that would mix when chewed, be sugar-free, be easy to chew, must not have a hard coating, and should not stick to artificial dentures. Specially made test-foods for masticatory performance assessment made by Lotte1 (Lotte Co., Ltd., Tokyo, Japan) may be acquired in Japan and Korea, but they are not available worldwide. Nonetheless, we consider plausible to find suitable specimens for mixing-tests most of the time. Further applications of the proposed expert system may be executed in two different scenar- ios: the calibration stage should be performed within a research context with more resources, and the diagnosis stage may be performed in a clinical practice context, linked by the sharing of experimental information in the form of a MEPAT. Shortcomings of the proposed method This study considered only one colour combination and brand of chewing gums, hence further studies may include other colour combinations and brands available in other regions and countries. In addition, the study sample for the diagnosis stage should be extended to include more masticatory-compromising pathologies and treatments. In our case, treatment with complete oral dentures was chosen because it represents a large portion of the daily routine of dental practice in Ecuador, and Masticatory Efficiency evaluation provides useful diagnostic information and relevant data for treatment enhancing. It is important to notice that the proposed system involved algorithms designed to reduce the influx of the operator during the image processing steps; therefore, subsequent analysis of the same set of images will always provide the same feature measurements. Nevertheless, pat- tern identification involved an aleatory component, which was required to ensure the robust- ness of the classifier. This means that subsequent executions of the machine learning step would provide different results each time. In this experiment, the classifier validation step diminished the randomness of the calibration execution by requiring a high MCC score and the selection of the classifier with the best performance. The proposed methodology comprises only simple pixel and histogram feature extraction models, so certain chewing gum colour combinations can affect the quality of the features. This phenomenon has been registered by Halazonetis et al. where the Circular Variance of the Hue channel performed poorly with colour combinations that included similar HUE values although easily differentiable by direct observation [13]. In this regard, the present study employed red and white chewing gum samples; so, the mixture of these two colours produced pink shades that may have affected the performance of HUE-based features. Conclusions Within the limitations of this study we conclude that the proposed expert system proved able and reliable to accurately identify patterns in mixture and subsequently classify masticated two-coloured chewing gum specimens into the corresponding group represented by the num- ber of chewing cycles applied considering a healthy reference population. Furthermore, the expert system proved able to identify differences in the ME of edentulous individuals with and without total prosthesis. Finally, we propose the inclusion of the newly presented ME defini- tion and reference scale in further research studies in this field. Supporting information S1 Appendix. Border-preserving region segmentation and classification. (DOCX) Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 17 / 20 http://www.plosone.org/article/fetchSingleRepresentation.action?uri=info:doi/10.1371/journal.pone.0190386.s001 https://doi.org/10.1371/journal.pone.0190386 S2 Appendix. Feature extraction models. (DOCX) S3 Appendix. Detailed information about the MEPAT construction. (DOCX) S4 Appendix. Core classification performance of single-feature classifiers. (DOCX) S1 File. XML schema for the MEPAT. (XML) Acknowledgments We express our gratitude to the volunteers who participated in this study. Our deepest appreci- ation goes to the Prometeo Project of the Secretarı́a de Educación Superior, Ciencia, Tecnolo- gı́a e Innovación (SENESCYT) of the Government of Ecuador for supporting this research work. We would also like to show our gratitude to the directive board of the Faculty of Den- tistry of the University of Guayaquil for providing us with many of the necessary resources for this study. Author Contributions Conceptualization: Gustavo Vaccaro, José Antonio Gil-Montoya. Data curation: José Ignacio Peláez, José Antonio Gil-Montoya. Formal analysis: Gustavo Vaccaro, José Ignacio Peláez. Funding acquisition: José Ignacio Peláez. Investigation: Gustavo Vaccaro, José Antonio Gil-Montoya. Methodology: Gustavo Vaccaro. Project administration: José Ignacio Peláez. Resources: José Ignacio Peláez. Software: Gustavo Vaccaro, José Ignacio Peláez. Supervision: José Ignacio Peláez, José Antonio Gil-Montoya. Validation: José Antonio Gil-Montoya. Visualization: Gustavo Vaccaro. Writing – original draft: Gustavo Vaccaro. Writing – review & editing: José Ignacio Peláez, José Antonio Gil-Montoya. References 1. Ohira A, Ono Y, Yano N, Takagi Y. The effect of chewing exercise in preschool children on maximum bite force and masticatory performance. Int J Paediatr Dent. 2012; 22: 146–53. https://doi.org/10.1111/ j.1365-263X.2011.01162.x PMID: 21781200 2. Bates J, Stafford G, Harrison A. Masticatory function—a review of the literature. III. Masticatory perfor- mance and efficiency. J Oral Rehabil. 1976; 3: 57–67. Recuperado: http://onlinelibrary.wiley.com/doi/ 10.1111/j.1365-2842.1976.tb00929.x/full PMID: 772184 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 18 / 20 http://www.plosone.org/article/fetchSingleRepresentation.action?uri=info:doi/10.1371/journal.pone.0190386.s002 http://www.plosone.org/article/fetchSingleRepresentation.action?uri=info:doi/10.1371/journal.pone.0190386.s003 http://www.plosone.org/article/fetchSingleRepresentation.action?uri=info:doi/10.1371/journal.pone.0190386.s004 http://www.plosone.org/article/fetchSingleRepresentation.action?uri=info:doi/10.1371/journal.pone.0190386.s005 https://doi.org/10.1111/j.1365-263X.2011.01162.x https://doi.org/10.1111/j.1365-263X.2011.01162.x http://www.ncbi.nlm.nih.gov/pubmed/21781200 http://onlinelibrary.wiley.com/doi/10.1111/j.1365-2842.1976.tb00929.x/full http://onlinelibrary.wiley.com/doi/10.1111/j.1365-2842.1976.tb00929.x/full http://www.ncbi.nlm.nih.gov/pubmed/772184 https://doi.org/10.1371/journal.pone.0190386 3. The Glossary of Prosthodontic Terms. J Prosthet Dent. Elsevier; 2005; 94: 10–92. https://doi.org/10. 1016/j.prosdent.2005.03.013 PMID: 16080238 4. Dai R, Lam OLT, Lo ECM, Li LSW, Wen Y, McGrath C. Orofacial functional impairments among patients following stroke: a systematic review. Oral Dis. 2014; https://doi.org/10.1111/odi.12274 PMID: 25041135 5. Yamashita S, Hatch JP, Rugh JD. Does chewing performance depend upon a specific masticatory pat- tern? J Oral Rehabil. 1999; 26: 547–53. Recuperado: http://www.ncbi.nlm.nih.gov/pubmed/10445472 PMID: 10445472 6. Gil-Montoya JA, Subirá C, Ramón JM, González-Moles MA. Oral health-related quality of life and nutri- tional status. J Public Health Dent. 2008; 68: 88–93. https://doi.org/10.1111/j.1752-7325.2007.00082.x PMID: 18248335 7. Manly RS, Braley LC. Masticatory Performance and Efficiency. J Dent Res. 1950; 29: 448–462. https:// doi.org/10.1177/00220345500290040701 PMID: 15436916 8. Schimmel M, Christou P, Herrmann F, Müller F. A two-colour chewing gum test for masticatory effi- ciency: development of different assessment methods. J Oral Rehabil. 2007; 34: 671–8. https://doi.org/ 10.1111/j.1365-2842.2007.01773.x PMID: 17716266 9. Schimmel M, Christou P, Miyazaki H, Halazonetis D, Herrmann FR, Müller F. A novel colourimetric technique to assess chewing function using two-coloured specimens: validation and application. J Dent. 2015; 43: 955–964. https://doi.org/10.1016/j.jdent.2015.06.003 PMID: 26111925 10. Weijenberg RAF, Scherder EJA, Visscher CM, Gorissen T, Yoshida E, Lobbezoo F. Two-colour chew- ing gum mixing ability: digitalisation and spatial heterogeneity analysis. J Oral Rehabil. 2013; 40: 737– 43. https://doi.org/10.1111/joor.12090 PMID: 23927753 11. Vaccaro G, Pelaez JI, Gil JA. Choosing the best image processing method for masticatory performance assessment when using two-coloured specimens. J Oral Rehabil. 2016; 43: 496–504. https://doi.org/ 10.1111/joor.12392 PMID: 26968333 12. Prinz JF. Quantitative evaluation of the effect of bolus size and number of chewing strokes on the intra- oral mixing of a two-colour chewing gum. J Oral Rehabil. 1999; 26: 243–7. PMID: 10194734 13. Halazonetis DJ, Schimmel M, Antonarakis GS, Christou P. Novel software for quantitative evaluation and graphical representation of masticatory efficiency. J Oral Rehabil. 2013; 40: 329–35. https://doi.org/ 10.1111/joor.12043 PMID: 23452188 14. van der Bilt A, Mojet J, Tekamp FA, Abbink JH. Comparing masticatory performance and mixing ability. J Oral Rehabil. 2010; 37: 79–84. https://doi.org/10.1111/j.1365-2842.2009.02040.x PMID: 19968766 15. van der Bilt A. Assessment of mastication with implications for oral rehabilitation: a review. J Oral Reha- bil. 2011; 38: 754–80. https://doi.org/10.1111/j.1365-2842.2010.02197.x PMID: 21241351 16. Beucher S, Lantuéjoul C. Use of Watersheds in Contour Detection. International workshop on image processing, real-time edge and motion detection. 1979. 17. Lu S, Wang S, Zhang Y. A note on the marker-based watershed method for X-ray image segmentation. Comput Methods Programs Biomed. 2017; 141: 1–2. https://doi.org/10.1016/j.cmpb.2017.01.014 PMID: 28241959 18. Zhang Y, Dong Z, Phillips P, Wang S, Ji G, Yang J. Exponential Wavelet Iterative Shrinkage Threshold- ing Algorithm for compressed sensing magnetic resonance imaging. Inf Sci (Ny). 2015; 322: 115–132. https://doi.org/10.1016/j.ins.2015.06.017 19. Comaniciu D, Meer P. Mean shift: a robust approach toward feature space analysis. IEEE Trans Pattern Anal Mach Intell. IEEE; 2002; 24: 603–619. https://doi.org/10.1109/34.1000236 20. Christoudias CM, Georgescu B, Meer P. Synergism in low level vision. Object recognition supported by user interaction for service robots. IEEE Comput. Soc; 2002. pp. 150–155. https://doi.org/10.1109/ ICPR.2002.1047421 21. MacQueen J. Some methods for classification and analysis of multivariate observations. Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Statistics. The Regents of the University of California; 1967. 22. Commission Internationale de L’Eclairage. Publication No. CIE 15.2. Colorimetry. 2nd ed. Vienna, Aus- tria: Central Bureau of the CIE.; 1986; 23. Zhang Y, Wang S, Phillips P, Ji G. Binary PSO with mutation operator for feature selection using deci- sion tree applied to spam detection. Knowledge-Based Syst. 2014; 64: 22–31. https://doi.org/10.1016/j. knosys.2014.03.015 24. Veredas F, Mesa H, Morente L. Binary tissue classification on wound images with neural networks and bayesian classifiers. IEEE Trans Med Imaging. IEEE; 2010; 29: 410–27. https://doi.org/10.1109/TMI. 2009.2033595 PMID: 19825516 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 19 / 20 https://doi.org/10.1016/j.prosdent.2005.03.013 https://doi.org/10.1016/j.prosdent.2005.03.013 http://www.ncbi.nlm.nih.gov/pubmed/16080238 https://doi.org/10.1111/odi.12274 http://www.ncbi.nlm.nih.gov/pubmed/25041135 http://www.ncbi.nlm.nih.gov/pubmed/10445472 http://www.ncbi.nlm.nih.gov/pubmed/10445472 https://doi.org/10.1111/j.1752-7325.2007.00082.x http://www.ncbi.nlm.nih.gov/pubmed/18248335 https://doi.org/10.1177/00220345500290040701 https://doi.org/10.1177/00220345500290040701 http://www.ncbi.nlm.nih.gov/pubmed/15436916 https://doi.org/10.1111/j.1365-2842.2007.01773.x https://doi.org/10.1111/j.1365-2842.2007.01773.x http://www.ncbi.nlm.nih.gov/pubmed/17716266 https://doi.org/10.1016/j.jdent.2015.06.003 http://www.ncbi.nlm.nih.gov/pubmed/26111925 https://doi.org/10.1111/joor.12090 http://www.ncbi.nlm.nih.gov/pubmed/23927753 https://doi.org/10.1111/joor.12392 https://doi.org/10.1111/joor.12392 http://www.ncbi.nlm.nih.gov/pubmed/26968333 http://www.ncbi.nlm.nih.gov/pubmed/10194734 https://doi.org/10.1111/joor.12043 https://doi.org/10.1111/joor.12043 http://www.ncbi.nlm.nih.gov/pubmed/23452188 https://doi.org/10.1111/j.1365-2842.2009.02040.x http://www.ncbi.nlm.nih.gov/pubmed/19968766 https://doi.org/10.1111/j.1365-2842.2010.02197.x http://www.ncbi.nlm.nih.gov/pubmed/21241351 https://doi.org/10.1016/j.cmpb.2017.01.014 http://www.ncbi.nlm.nih.gov/pubmed/28241959 https://doi.org/10.1016/j.ins.2015.06.017 https://doi.org/10.1109/34.1000236 https://doi.org/10.1109/ICPR.2002.1047421 https://doi.org/10.1109/ICPR.2002.1047421 https://doi.org/10.1016/j.knosys.2014.03.015 https://doi.org/10.1016/j.knosys.2014.03.015 https://doi.org/10.1109/TMI.2009.2033595 https://doi.org/10.1109/TMI.2009.2033595 http://www.ncbi.nlm.nih.gov/pubmed/19825516 https://doi.org/10.1371/journal.pone.0190386 25. Judd DB. Hue Saturation and Lightness of Surface Colors with Chromatic Illumination. J Opt Soc Am. Optical Society of America; 1940; 30: 2. https://doi.org/10.1364/JOSA.30.000002 26. Vezhnevets V, Sazonov V, Andreeva A. A Survey on Pixel-Based Skin Color Detection Techniques. Graphicon-2003. Moscow, Russia; 2003. pp. 85–92. 27. Jolliffe I. Principal Component Analysis. New York: Springer-Verlag; 2002. https://doi.org/10.1007/ b98835 28. Haykin S. Neural Networks: A Comprehensive Foundation. Prentice Hall PTR; 1998; 29. Ripley B.D. Pattern Recognition and Neural Networks. Oxford, UK .: Cambridge University Press; 1996. 30. Zhang Y, Sun Y, Phillips P, Liu G, Zhou X, Wang S. A Multilayer Perceptron Based Smart Pathological Brain Detection System by Fractional Fourier Entropy. J Med Syst. Springer US; 2016; 40: 173. https:// doi.org/10.1007/s10916-016-0525-2 PMID: 27250502 31. Wang S-H, Zhang Y, Li Y-J, Jia W-J, Liu F-Y, Yang M-M, et al. Single slice based detection for Alzhei- mer’s disease via wavelet entropy and multilayer perceptron trained by biogeography-based optimiza- tion. Multimed Tools Appl. Springer US; 2016; 1–25. https://doi.org/10.1007/s11042-016-4222-4 32. Suzuki K, editor. Artificial Neural Networks—Architectures and Applications. InTech; 2013. https://doi. org/10.5772/3409 33. Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-propagating errors. Nature. 1986; 323: 533–536. https://doi.org/10.1038/323533a0 34. Čepek M, Šnorek M, Chudáček V. Ecg signal classification using game neural network and its compari- son to other classifiers. Artificial Neural Networks-ICANN 2008. Springer; 2008. pp. 768–777. 35. Callejón AM, Casado AM, Fernández MA, Peláez JI. A System of Insolvency Prediction for industrial companies using a financial alternative model with neural networks. Int J Comput Intell Syst. Taylor & Francis Group; 2013; 6: 29–37. https://doi.org/10.1080/18756891.2013.754167 36. Vaccaro G, Pelaez JI. Dental tissue classification using computational intelligence and digital image analysis. Biodental Engineering III—Proceedings of the 3rd International Conference on Biodental Engi- neering, BIODENTAL 2014. Taylor and Francis—Balkema; 2014. pp. 221–226. https://doi.org/10. 1201/b17071 37. Lawrence S, Giles CL, Tsoi AC. Lessons in neural network training: Overfitting may be harder than expected. Proceedings of the Fourteenth National Conference on Artificial Intelligence, AAAI-97. Menlo Park, California: AAAI Press; 1997. pp. 540–545. 38. Matthews BW. Comparison of the predicted and observed secondary structure of T4 phage lysozyme. Biochim Biophys Acta—Protein Struct. 1975; 405: 442–451. https://doi.org/10.1016/0005-2795(75) 90109-9 39. Cohen J. A Coefficient of Agreement for Nominal Scales. Educ Psychol Meas. 1960; 20: 37–46. https:// doi.org/10.1177/001316446002000104 Objective masticatory efficiency assessment PLOS ONE | https://doi.org/10.1371/journal.pone.0190386 January 31, 2018 20 / 20 https://doi.org/10.1364/JOSA.30.000002 https://doi.org/10.1007/b98835 https://doi.org/10.1007/b98835 https://doi.org/10.1007/s10916-016-0525-2 https://doi.org/10.1007/s10916-016-0525-2 http://www.ncbi.nlm.nih.gov/pubmed/27250502 https://doi.org/10.1007/s11042-016-4222-4 https://doi.org/10.5772/3409 https://doi.org/10.5772/3409 https://doi.org/10.1038/323533a0 https://doi.org/10.1080/18756891.2013.754167 https://doi.org/10.1201/b17071 https://doi.org/10.1201/b17071 https://doi.org/10.1016/0005-2795(75)90109-9 https://doi.org/10.1016/0005-2795(75)90109-9 https://doi.org/10.1177/001316446002000104 https://doi.org/10.1177/001316446002000104 https://doi.org/10.1371/journal.pone.0190386 
work_2pw4vkdn4jh4dohqkeifcbjleq	----	Label Ranking Forests Cláudio Rebelo de Sá1,3 Carlos Soares2,3 Arno Knobbe1 Paulo Cortez4 1 LIACS, Universiteit Leiden, Netherlands 2 Faculdade de Engenharia, Universidade do Porto, Portugal 3 INESCTEC Porto, Porto, Portugal 4 ALGORITMI Centre, Department of Information Systems, University of Minho, Portugal Abstract The problem of Label Ranking is receiving increasing attention from several research communities. The algorithms that have been developed/adapted to treat rankings of a fixed set of labels as the tar- get object, include several different types of decision trees (DT). One DT-based algorithm, which has been very successful in other tasks but which has not been adapted for label ranking is the Random Forests (RF) algorithm. RFs are an ensemble learning method that combines different trees obtained using different randomization techniques. In this work, we propose an ensemble of decision trees for Label Ranking, based on Random Forests, which we refer to as Label Ranking Forests (LRF). Two different algorithms that learn DT for label ranking are used to obtain the trees. We then compare and discuss the results of LRF with standalone decision tree approaches. The results indicate that the method is highly competitive. 1 Introduction Label Ranking (LR) is an increasingly popular topic in the machine learn- ing literature (Ribeiro et al., 2012; de Sá et al., 2011; Cheng & Hüllermeier, 2011; Cheng et al., 2012; Vembu & Gärtner, 2010). LR studies a problem of learning a mapping from instances to rankings over a finite number of prede- fined labels. It can be considered a natural generalization of the conventional 1 classification problem, where the goal is to predict a single label instead of a ranking of all the labels (Cheng et al., 2009). Some application of Label Ranking approaches are (Hüllermeier et al., 2008): Meta-learning (Brazdil & Soares, 1999), where we try to predict a ranking of a set of algorithms according to the best expected accuracy on a given dataset; Microarray analysis (Hüllermeier et al., 2008) to find patterns in genes from Yeast on five different micro-array experiments (spo, heat, dtt, cold and diau); Image categorization (Fürnkranz et al., 2008) of landscape pictures from several categories (beach, sunset, field, fall foliage, mountain, urban). There are two main approaches to the problem of LR: methods that transform the ranking problem into multiple binary problems and methods that were developed or adapted to treat the rankings as target objects, with- out any transformation. An example of the former is the ranking by pair- wise comparisons (Hüllermeier et al., 2008). Examples of algorithms that were adapted to deal with rankings as the target objects include decision trees (Todorovski et al., 2002; Cheng et al., 2009), naive Bayes (Aiguzhinov et al., 2010) and k -Nearest Neighbor (Brazdil et al., 2003; Cheng et al., 2009). Some of the latter adaptations are based on statistical distribution of rank- ings (e.g., (Cheng et al., 2010)) while others are based on ranking distance measures (e.g., (Todorovski et al., 2002; de Sá et al., 2011)). Tree-based models have been used in classification (Quinlan, 1986), re- gression (Breiman et al., 1984) and also label ranking (Todorovski et al., 2002; Cheng et al., 2009; de Sá et al., 2015) tasks. These methods are popular for a number of reasons, including how they can clearly express information about the problem, because their structure is relatively easy to interpret even for people without a background in learning algorithms. In classification, combining the predictive power of an ensemble of trees often comes with significant accuracy improvements (Breiman, 2001). One of the earliest examples of ensemble methods is bagging (a contraction of bootstrap-aggregating) (Breiman, 1996). In bagging, an ensemble of trees is generated and each one is learned on a random selection of examples from the training set. A popular ensemble method is Random Forests (Breiman, 2001) which combines different randomization techniques. Considering the success of Random Forests in terms of improved accu- racy for classification and regression problems (Biau, 2012), some approaches have been proposed to deal with different targets, such as bipartite rankings (Clémençon et al., 2013). Label Ranking Forests should also be seen as a potential robust approach for LR. Adapting RF to Label Ranking can be a straightforward process once you have adapted decision trees. In this work, we propose an approach of ensemble learners which we 2 refer to as Label Ranking Forests (LRF). The proposed method is a natural adaptation of Random Forests for LR, combining the task-independent RF algorithm with the traditional algorithm for top-down induction of decision trees adapted for label ranking. The available adaptations of decision tree algorithms for LR include Label Ranking Trees (LRT) (Cheng et al., 2009), Ranking Trees (Rebelo et al., 2008) and Entropy-based Ranking Trees (de Sá et al., 2015). Considering that the set of trees, in most cases, predict distinct rankings, one should also take into account ranking aggregation methods. This paper extends previous work (de Sá et al., 2015), in which we pro- posed a new version of decision trees for LR, called the Entropy-based Rank- ing Trees and empirically compared them to existing approaches. The main contribution in this paper is the new Label Ranking Forests algorithm, which is an adaptation of the RF ensemble method, using Entropy-based Ranking Trees as the base level algorithm. The results indicate that LRF are compet- itive with state of the art methods and improve the accuracy of standalone decision trees. An additional contribution is an extension of the original experimental study on Entropy-based Ranking Trees, by analyzing model complexity. 2 Label Ranking In this section, we start by formalizing the problem of label ranking (Sec- tion 2.1) and then we discuss the adaptation of the decision trees algorithm for label ranking (Section 2.2) and one such adaptation, Entropy Ranking Trees (Section 2.3). 2.1 Formalization The Label Ranking (LR) task is similar to classification. In classification, given an instance x from the instance space X, the goal is to predict the label (or class) λ to which x belongs, from a pre-defined set L = {λ1, . . . ,λk}. In LR, the goal is to predict the ranking of the labels in L that is associated with x (Hüllermeier et al., 2008). A ranking can be represented as a total order over L defined on the permutation space Ω. A total order can be seen as a permutation π of the set {1, . . . ,k}, such that π(a) is the position of λa in π. As in classification, we do not assume the existence of a deterministic X → Ω mapping. Instead, every instance is associated with a probability distribution over Ω (Cheng et al., 2009). This means that, for each x ∈ X, there exists a probability distribution P(·|x) such that, for every π ∈ 3 Ω, P(π|x) is the probability that π is the ranking associated with x. The goal in LR is to learn the mapping X → Ω. The training data is a set of instances D = {〈xi,πi〉}, i = 1, . . . ,n, where xi is a vector containing the values x j i,j = 1, . . . ,m of m independent variables describing instance i and πi is the corresponding target ranking. Given an instance xi with label ranking πi, and the ranking π̂i predicted by an LR model, we can evaluate the accuracy of the prediction with loss functions on Ω. Some of these measures are based in the number of discordant label pairs: D(π,π̂) = #{(a,b)|π(a) > π(b) ∧ π̂(a) < π̂(b)} If normalized to the interval [−1, 1], this function is equivalent to Kendall’s τ coefficient, which is a correlation measure where D(π,π) = 1 and D(π,π−1) = −1, where π−1 denotes the inverse order of π (e.g. π = (1, 2, 3, 4) and π−1 = (4, 3, 2, 1)). The accuracy of a model can be estimated by averaging this coefficient over a set of examples. Other correlation measures, such as Spearman’s rank correlation coefficient (Spearman, 1904), have also been used (Brazdil et al., 2003). Although we assume total orders, it may be the case that two labels are tied in the same rank (i.e. πi(a) = πi(b),a 6= b). In this case, a variation of Kendall’s τ, the tau− b (Agresti, 2010) can be used. 2.2 Ranking Trees Tree-based models have been used in classification (Quinlan, 1986), regres- sion (Breiman et al., 1984), and label ranking (Todorovski et al., 2002; Cheng et al., 2009; de Sá et al., 2015) tasks. These methods are popular for a num- ber of reasons, including how they can clearly express information about the problem, because their structure is relatively easy to interpret even for people without a background in learning algorithms. It is also possible to obtain information about the importance of the various attributes for the prediction depending on how close to the root they are used. The Top-Down Induction of Decision Trees (TDIDT) algorithm is com- monly used for induction of decision trees (Mitchell, 1997). It is a recursive partitioning algorithm that iteratively splits data into smaller subsets which are increasingly more homogeneous in terms of the target variable (Algo- rithm 1). A split is a test on one of the attributes that divides the dataset into two disjoint subsets. For instance, given a numerical attribute x2, a split could be x2 ≥ 5. Given a splitting criterion that represents the gain in purity obtained with a split, the algorithm chooses the split that optimizes 4 its value in each iteration. In its simplest form, the TDIDT algorithm only stops when the nodes are pure, i.e., when the value of the target attribute is the same for all examples in the node. This usually causes the algorithm to overfit, i.e., to generate models that capture the noise in the data, as well as the regularities that are of general usefulness. One approach to address this problem is to introduce a stopping criterion in the algorithm that tests whether the best split is significantly improving the quality of the model. If not, the algorithm stops and returns a leaf node. The algorithm is executed recursively for the subsets of the data obtained based on the best split until the stopping criterion is met. A leaf node is represented by a value of the target attribute generated by a rule that solves potential conflicts in the set of training examples that are in the node. That value is the prediction that will be made for new examples that fall into that node. In classification, the prediction rule is usually the most frequent class among the training examples. Algorithm 1 TDIDT algorithm Input: Dataset D BestSplit = Test of the attributes that optimizes the SPLITTING CRI- TERION if STOPPING CRITERION == TRUE then Determine leaf prediction based on the target values in D Return a leaf node with the corresponding LEAF PREDICTION else LeftSubtree = TDIDT(D¬BestSplit) RightSubtree = TDIDT(DBestSplit) end if The adaptation of this algorithm for label ranking involves an appropriate choice of the splitting criterion, stopping criterion and the prediction rule (Algorithm 1). Splitting Criterion The splitting criterion is a measure that quantifies the quality of a given partition of the data. It is usually applied to all the possible splits of the data that can be made with tests on the values of individual attributes. In Ranking Trees (RT) the goal is to obtain leaf nodes that contain ex- amples with target rankings as similar between themselves as possible. To assess the similarity between the rankings of a set of training examples, the mean correlation between them is calculated using Kendall, Spearman or any other ranking correlation coefficient. The quality of the split is given by the 5 Table 1: Illustration of the splitting criterion Attribute Condition=true Condition=false values rank corr. values rank corr. x1 a 0.3 {b,c} -0.2 b 0.2 {a,c} 0.1 c 0.5 {a,b} 0.2 x2 < 5 -0.1 ≥ 5 0.1 weighted mean correlation of the values obtained for the subsets, where the weight is given by the number of examples in each subset. For simplicity, if we ignore the weights, the splitting criterion of ranking trees is illustrated both for nominal and numerical attributes in Table 1. The nominal attribute x1 has three values (a, b and c). Therefore, three binary splits are possible. For the numerical attribute x2, a split can be made in between every pair of consecutive values. In this case, the best split is x1 = c, with a mean correlation of 0.5, in comparison to a mean correlation of 0.2 for the remaining, i.e., the training examples for which x1 = {a,b}. Stopping Criterion The stopping criterion is used to determine if it is worthwhile to make a split or if there is a significant risk of overfit- ting (Mitchell, 1997). A split should only be made if the similarity between examples in the subsets increases substantially. Let Sparent be the similarity between the examples in the parent node and Ssplit the weighted mean sim- ilarity in the subsets obtained with the best split. The stopping criterion is defined as follows (Rebelo et al., 2008): (1 + Sparent) ≥ γ(1 + Ssplit) (1) Note that the relevance of the increase in similarity is controlled by the γ parameter. A γ ≥ 1 does not ensure increased purity of child nodes. On the other hand, small γ values require splits with very large increase in purity, which means that the algorithm will stop the recursion early. Prediction Rule The prediction rule is a method to generate a prediction from the (possibly conflicting) target values of the training examples in a leaf node. In LR, the aggregation of rankings is not so straightforward as in other tasks (e.g. classification or regression) and is known as the ranking aggregation problem (Yasutake et al., 2012). It is a classical problem in social choice literature (de Borda, 1781) but also in information retrieval 6 Table 2: Illustration of the prediction rule λ1 λ2 λ3 λ4 π1 1 3 2 4 π2 2 1 4 3 π 1.5 2 3 3.5 π̂ 1 2 3 4 tasks (Dwork et al., 2001). A consensus ranking minimizes the distance to all rankings (Kemeny & Snell, 1972). A simple approach, which we adopted in this work, is to compute the average ranking (Brazdil et al., 2003) of the predictions. It is calculated by averaging the rank for each label λj, π (j) = ∑ i πi (j) /n. The predicted ranking π̂ is the ranking π of the labels λj obtained based on the average ranks π (j). Table 2 illustrates the prediction rule used in this work. 2.3 Entropy Ranking Trees Recently, we proposed an alternative approach to decision trees for ranking data, the Entropy-based Ranking Trees (ERT) (de Sá et al., 2015). ERT uses an adaptation of Information Gain (IG) (de Sá et al., 2016) to as- sess the splitting points and the Minimum Description Length Principle Cut (MDLPC) (Fayyad & Irani, 1993) as the stopping criterion. To explain this method, we start by presenting the IG for rankings measure and then the adapted splitting and stopping criteria. Decision trees for classification, such as ID3 (Quinlan, 1986), use Infor- mation Gain (IG) as a splitting criterion to determine the best split points. IG is a statistical property that measures the gain in entropy, between the prior and actual state (Mitchell, 1997). In this case, we measure it in terms of the distribution of the target variable, before and after the split. In other words, considering a set S of size nS, since entropy, H, is a measure of disor- der, IG is basically how much uncertainty in S is eliminated after splitting on a numerical attribute xa: IG (xa,T; S) = H (S) − |S1| nS H (S1) − |S2| nS H (S2) where |S1| and |S2| are the number of instances on the left side (S1) and the number of instances on the right side (S2), respectively, of the cut point T in attribute xa. 7 In cases where S is a set of rankings, we can use the entropy for rankings (de Sá et al., 2016) which is defined as: Hranking (S) = K∑ i=1 P (πi,S) log (P (πi,S)) log ( kt (S) ) (2) where P (πi,S) is the proportion of rankings equal to πi in S, K is the number of distinct rankings in S and kt (S) is the average normalized Kendall τ (Kendall & Gibbons, 1970) distance in the subset S: kt (S) = ∑K i=1 ∑n j=1 τ(πi,πj)+1 2 K ×nS . As in Section 2.2, the leaves of the tree should not be forced to be pure. Instead, a stopping criterion should be used to avoid overfitting and be ro- bust to noise in rankings. Given an entropy measure, the adaptation of the splitting and stopping criteria comes in a natural way. As shown in (de Sá et al., 2016), the MDLPC Criterion can be used as a splitting criterion with the adapted version of entropy Hranking. This entropy measure also works with partial orders, however, in this work, we only use total orders. 3 Random Forests Random Forests (RF) (Breiman, 2001) are an ensemble method originally proposed for classification and regression problems. It essentially consists of the generation of multiple decision trees obtained using different randomiza- tion techniques. The set of predictions made by each of these trees is then aggregated to obtain the prediction of the ensemble. The RF algorithm is related to another popular ensemble method by the same author, Bagging (Breiman, 1996), which stands for bootstrap-aggregating. This is an ensemble method that takes a predefined number s of samples (without replacement) from the training data to construct s models. Given a new example, s predictions are generated, which are then aggregated, usu- ally with average or mode, to obtain a combined prediction. RF can be regarded as an extension of bagging. Given a forest size s and a training dataset D, a set of bootstrap samples, {D′1 . . . ,D′s} is generated by sampling with repetition from D. A decision tree is learned from each {D′1 . . . ,D′s}, which is grown in a slightly different way from the original. At each node, only a random subset of the m features can be used for splitting. In classification, the number of random features used in each split is usually√ m and in regression log2 m. This results in what is usually referred to as 8 random trees. As in bagging, each of the s random trees makes predictions on the test data, which are then combined using a suitable aggregation method. One of the reasons for the popularity of RF lays in the fact that they have few parameters to tune and can be applied to various tasks (Scornet et al., 2014). They require a simple implementation and even with small sample sizes it usually gives accurate results. Moreover, considering that it uses s independent learners, it can be parallelized. One of the reasons that makes RF a popular approach is that it is possible to take advantage of the algorithm to assess variable importance (Genuer et al., 2010). 3.1 Label Ranking Forests Considering the success of Random Forests in terms of improved accuracy for classification and regression problems, some approaches have been proposed to deal with different targets, such as bipartite rankings (Clémençon et al., 2013). Label Ranking Forests should also be seen as a potential robust approach for LR. Adapting RF to Label Ranking can be a straightforward process once you have adapted decision trees. Thus, we propose a new ensemble LR algorithm, the Label Ranking Forests based on Random Forests. With this approach, we expect to in- crease the accuracy of Label Ranking tree methods. In classification and regression, the aggregation of predictions is done in a simple way, mode and mean, respectively. However, as discussed in Section 2.2, the aggregation of rankings is not so straightforward. Like in Ranking Trees, we use the average ranking (Brazdil et al., 2003) to aggregate the predictions. Given the similarity of the LR task to classification, the number of random subset features we use in each split is √ m, the same value that is used in RF for classification. When the algorithm is not able to find a good split on any of the √ m selected features for the root node, it looks for a split on all the m features instead. This prevents the random feature selection mechanism from gener- ating empty trees. 4 Empirical Study In this section we describe the empirical study to investigate the performance of LRF and the tree methods used at the base level. We start by describ- ing the experimental setup (Section 4.1), then the results of the base-level 9 algorithms (Section 4.2) and finally the results of the new algorithm (Sec- tion 4.3). 4.1 Experimental setup The experiments are carried out on datasets from the KEBI Data Repository at the Philipps University of Marburg (Cheng et al., 2009) that are typically used in LR research (Table 3). They are based on classification and regression datasets, obtained using two different transformation methods: A) the target ranking is a permutation of the classes of the original target attribute, derived from the probabilities generated by a Naive Bayes classifier; B) the target ranking is derived for each example from the order of the values of a set of numerical variables, which are then no longer used as independent variables. A few basic statistics of the datasets used in our experiments are presented in Table 3. Although these are somewhat artificial datasets, they are quite useful as benchmarks for LR algorithms. A simple measure of the diversity of the target rankings is the Unique Ranking’s Proportion, Uπ. Uπ is the proportion of distinct target rankings for a given dataset (Table 3). As a practical example, the iris dataset has 5 distinct rankings for 150 instances, which yields Uπ = 5 150 ≈ 3%. This means that all the 150 rankings are duplicates of these 5. The code for all the experiments presented in this paper has been written in R (R Development Core Team, 2010).1 The generalization performance of the LR methods was estimated using a methodology that has been used previously for this purpose (Hüllermeier et al., 2008). The evaluation measure is Kendall’s τ and the performance of the methods was estimated using ten-fold cross-validation. 4.2 Results with Label Ranking Trees We evaluate the two variants of ranking trees described earlier: ranking trees (RT) and entropy ranking trees (ERT) (Sections 2.2 and 2.3). The RT algorithm has a parameter γ, that can affect the accuracy of the model. Based on previous results, we use γ = 0.98 for RT (de Sá et al., 2015). Table 4 presents the results obtained by the two decision tree approaches, RT and ERT, in comparison to the results for Label Ranking Trees (LRT), that are reproduced from the original paper (Cheng et al., 2009). We note that we have no information about the depth of the trees obtained with the latter and thus such information is omitted in Table 4. Even though LRT 1The code is available at https://github.com/rebelosa/labelrankingforests. 10 Table 3: Summary of the KEBI datasets Datasets type #examples #labels #attributes Uπ autorship A 841 4 70 2% bodyfat B 252 7 7 94% calhousing B 20,640 4 4 0.1% cpu-small B 8,192 5 6 1% elevators B 16,599 9 9 1% fried B 40,769 5 9 0.3% glass A 214 6 9 14% housing B 506 6 6 22% iris A 150 3 4 3% pendigits A 10,992 10 16 19% segment A 2310 7 18 6% stock B 950 5 5 5% vehicle A 846 4 18 2% vowel A 528 11 10 56% wine A 178 3 13 3% wisconsin B 194 16 16 100% performs best in most of the cases presented, both RT and ERT are also competitive methods. Figure 1 shows how much smaller ERT trees are, in general. By generating smaller trees, ERT provides more interpretable models when compared with RT. An exception is the calhousing dataset, where ERT generates larger trees. However, in this case, the increase in size is justified by a reasonable increase of accuracy (Table 4). To compare different ranking methods, we use a combination of Fried- man’s test and Dunn’s Multiple Comparison Procedure (Neave & Worthing- ton, 1992), which has been used before for this purpose (Brazdil et al., 2003). First we run the Friedman’s test to check whether the results are different or not, with the following hypotheses: H0 : The distributions of Kendall’s τ are equal H1 : The distributions of Kendall’s τ are not equal Using the Friedman test (implemented in the stats package (R Development Core Team, 2010)) we obtained a p-value < 1%, which shows strong evi- dence against H0. This means that there is a high probability that the three methods have different performance. 11 Table 4: Results obtained for Ranking Trees on KEBI datasets (the mean accuracy is represented in terms of Kendall’s tau, τ; the best mean accuracy values are in bold) RT ERT LRT mean mean mean accuracy depth accuracy depth accuracy authorship .883 8.0 .889 4.0 .882 bodyfat .111 11.9 .182 2.7 .117 calhousing .182 1.0 .291 11.6 .324 cpu-small .458 17.2 .437 6.1 .447 elevators .746 18.9 .757 7.9 .760 fried .797 20.2 .774 13.2 .890 glass .871 8.2 .854 3.0 .883 housing .794 12.9 .704 3.4 .797 iris .963 4.3 .853 2.0 .947 pendigits .871 14.0 .838 5.9 .935 segment .929 12.0 .901 5.0 .949 stock .897 10.8 .859 5.0 .895 vehicle .817 11.0 .787 4.1 .827 vowel .833 12.5 .598 3.6 .794 wine .905 4.0 .906 2.0 .882 wisconsin .334 10.0 .337 2.3 .343 average .712 11.1 .685 5.1 .730 Thus, we tested which of the three methods are different from one an- other with the Dunns Multiple Comparison Procedure (Neave & Worthing- ton, 1992). Using the R package dunn.test (Dinno, 2015), we tested the following hypotheses for each pair of methods a and b: H0 The distributions of Kendall’s τ for a and b are equal H1 The distributions of Kendall’s τ for a and b are not equal Table 5 indicates that there is no statistical evidence that the methods are different. The statistical tests confirm our observation that, although LRT generally obtains better results than RT and ERT, the latter approaches are competitive. 12 Figure 1: Comparison of the average depth of the trees obtained with RT (blue) and ERT (red) on KEBI datasets Table 5: Dunn’s test results (p-values) RT ERT LRT RT 0.22 0.37 ERT 0.22 0.13 LRT 0.37 0.13 13 4.3 Results with Label Ranking Forests We generated forests with 100 trees and aggregated the predicted rankings with the average ranking method (Brazdil et al., 2003). Table 6 presents the results obtained by the Label Ranking Forests using RT and ERT, referred to as LRF-RT and LRF-ERT, respectively. The average depth of the trees for LRF-RT is, for most cases, smaller than that of the tree obtained with the RT algorithm, while the accuracy is better. On average, for each 0.019 increase in accuracy there was a decrease of 1.8 in the average depth of the trees. One exception is the elevators dataset, with suffered a significant decrease in accuracy by using the LRF method. The comparison between ERT and LRF-ERT leads to different observa- tions. The average depth of the trees increases when using LRF. This can be explained by the fact that the measure of entropy for rankings used in ERT is very robust to noise in rankings (de Sá et al., 2016). Hence, it requires a larger amount of dissimilarity in a set of rankings to find a partition. As noted in Section 4.3 (Figure 1) the depth of the trees is much smaller with ERT than with RT. An additional observation is that the result obtained using LRF with ERT yielded a significant reduction in accuracy. In Figure 2, we can observe how much the accuracy increases/decreases with LRF when compared to the corresponding base-level trees alone. In the vast majority of datasets, there is some improvement in accuracy. The only exception is the elevators dataset, as mentioned above. Using the same statistical tests as before (Section 4.2), we compare LRF- RT and LRF-ERT with the RT, ERT and LRT methods. With the Fried- man’s test we got a p-value < 1%, which shows strong evidence against H0. Then, now that we know that there are some differences between the 2 meth- ods we will test which are different from one another with the Dunns Mul- tiple Comparison Procedure. Since we got a p-value around 25%, between LRF-RT and the LRF-ERT, we cannot conclude that there is no statistical evidence that the methods are different. On the pairwise comparisons of the methods Table 8, we measure how many time each method wins, in terms of accuracy. In this analysis, we conclude that Label Ranking Forests using RT give the best results, proving the effectiveness of the approach. On the other hand, even though LRF-ERT shows some improvement in terms of accuracy relatively to ERT, it did not behave much better than RT or LRT (Table 8). Again, this might be caused by the fact that the measure of entropy for rankings used in ERT is very robust to noise. For this reason, the depth of trees in LRF-ERT is, on average, 70% the depth of trees in LRF-RT. While this can be an advantage in terms of Label Ranking Trees, 14 Table 6: Results obtained for Label Ranking Forests on KEBI datasets, using two different label ranking trees, RT and ERT (the mean accuracy is represented in terms of Kendall’s tau, τ; the best mean accuracy values are in bold) LRF-RT LRF-ERT mean mean accuracy depth accuracy depth authorship .912 8.3 .906 7.7 bodyfat .212 10.6 .211 5.3 calhousing .185 1.0 .294 8.3 cpu-small .469 13.9 .471 7.8 elevators .605 10.0 .721 9.5 fried .887 15.5 .841 14.5 glass .874 6.0 .849 2.7 housing .780 10.9 .699 3.7 iris .973 4.9 .933 2.3 segment .930 10.8 .917 5.2 stock .892 9.9 .869 5.5 vehicle .850 10.0 .849 9.4 vowel .844 11.5 .701 4.9 wine .932 4.3 .925 2.8 wisconsin .460 8.8 .429 3.7 average .720 9.1 .708 6.2 Table 7: Dunn’s test for all the methods (p-values) RT ERT LRT LRF-RT LRF-ERT RT 0.23 0.34 0.31 0.44 ERT 0.23 0.13 0.11 0.28 LRT 0.34 0.13 0.46 0.29 LRF-RT 0.31 0.11 0.46 0.25 LRF-ERT 0.44 0.28 0.29 0.25 15 Figure 2: Accuracy gained/lost per dataset for using the ensemble method LRF, instead of standalone decision trees RT (blue) and ERT (red) on KEBI datasets Table 8: Pairwise comparisons of the methods in terms of win statistics. RT ERT LRT LRF-RT LRF-ERT Total (Rank) RT 9 6 3 7 25 (4) ERT 6 3 2 4 15 (5) LRT 9 12 7 9 37 (2) LRF-RT 12 13 8 13 46 (1) LRF-ERT 8 11 6 2 27 (3) 16 in Label Ranking Forests it is less relevant because it is hard to interpret 100 trees per dataset. 5 Conclusions In this work, we propose an ensemble of decision tree methods for Label Ranking, called Label Ranking Forests (LRF). The method is tested with two different base-level methods Ranking Trees (RT) and Entropy-based Ranking Trees (ERT). We present an empirical evaluation using well known datasets in this field. We also extend the analysis from previous work for tree-based methods, RT and ERT, and compare with the state of the art Label Ranking Trees (LRT) approach. The analysis on the decision trees shows that both RT and ERT are valid and competitive approaches. While RT usually gives better accuracy, on the other hand, ERT generates trees with much smaller depth (around 50% less, in comparison to RT). Our results were also compared with the published results for Label Ranking Trees (LRT) (Cheng et al., 2009). LRT has in general better accuracy than RT and ERT, however, statistical tests showed that none of the methods is significantly different. This means that both RT and ERT are competitive approaches, and, since they are distance-based methods, we can also say that this kind of approaches is worth pursuing. The two ensemble approaches, LRF-RT and LRF-ERT, used the base ranking tree models RT and ERT, respectively. Similarly to the application of Random Forests to other tasks, there was a general increase in accuracy when compared to the corresponding base-level methods. The results confirm that both LRF-RT and LRF-ERT are highly competitive LR methods. LRF- RT, in particular, stands out as a clear winner in terms of accuracy. As future work, we might improve the comparison with LRT method (Cheng et al., 2009), by implementing it and testing it both as learning algorithm and as the base-level method for Label Ranking Forests. Also, LRF can po- tentially produce similar benefits as the Random Forest method, in terms of feature selection or input variable importance measurement, when applied to LR datasets. Finally, the experiments in this paper were carried out on a set of standard benchmark datasets, which represent artificial LR problems. We plan to apply these approaches on real world datasets e.g. related with user preferences (Kamishima, 2003). 17 Acknowledgments This work is financed by the ERDF - European Regional Development Fund through the Operational Programme for Competitiveness and Internationali- sation - COMPETE 2020 Programme within project POCI-01-0145-FEDER- 006961, and by National Funds through the FC - Fundação para a Ciência e a Tecnologia (Portuguese Foundation for Science and Technology) as part of project UID/EEA/50014/2013. References Agresti, A. (2010), Analysis of ordinal categorical data, Vol. 656, John Wiley & Sons. Aiguzhinov, A., Soares, C. & Serra, A. P. (2010), A similarity-based adap- tation of naive bayes for label ranking: Application to the metalearning problem of algorithm recommendation, in ‘Discovery Science - 13th In- ternational Conference, DS 2010, Canberra, Australia, October 6-8, 2010. Proceedings’, pp. 16–26. Biau, G. (2012), ‘Analysis of a random forests model’, Journal of Machine Learning Research 13, 1063–1095. Brazdil, P. & Soares, C. (1999), Exploiting Past Experience in Ranking Clas- sifiers, in H. Bacelar-Nicolau, F. C. Nicolau & J. Janssen, eds, ‘Applied Stochastic Models and Data Analysis’, Instituto Nacional de Estat́ıstica, pp. 299–304. Brazdil, P., Soares, C. & da Costa, J. P. (2003), ‘Ranking learning algo- rithms: Using IBL and meta-learning on accuracy and time results’, Ma- chine Learning 50(3), 251–277. Breiman, L. (1996), ‘Bagging predictors’, Machine Learning 24(2), 123–140. Breiman, L. (2001), ‘Random forests’, Machine Learning 45(1), 5–32. Breiman, L., Friedman, J. H., Olshen, R. A. & Stone, C. J. (1984), Classifi- cation and Regression Trees, Wadsworth. Cheng, W. & Hüllermeier, E. (2011), ‘Label ranking with abstention: Pre- dicting partial orders by thresholding probability distributions (extended abstract)’, Computing Research Repository, CoRR. 18 Cheng, W., Dembczynski, K. & Hüllermeier, E. (2010), Label ranking meth- ods based on the plackett-luce model, in ‘ICML’, pp. 215–222. Cheng, W., Huhn, J. C. & Hüllermeier, E. (2009), Decision tree and instance- based learning for label ranking, in ‘Proceedings of the 26th Annual Inter- national Conference on Machine Learning, ICML 2009, Montreal, Quebec, Canada, June 14-18, 2009’, pp. 161–168. Cheng, W., Hüllermeier, E., Waegeman, W. & Welker, V. (2012), Label ranking with partial abstention based on thresholded probabilistic models, in ‘Advances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems 2012. Proceedings of a meeting held December 3-6, 2012, Lake Tahoe, Nevada, United States.’, pp. 2510–2518. Clémençon, S., Depecker, M. & Vayatis, N. (2013), ‘Ranking forests’, Journal of Machine Learning Research 14(1), 39–73. de Borda, J. C. (1781), ‘Mémoire sur les élections au scrutin’. de Sá, C. R., Rebelo, C., Soares, C. & Knobbe, A. J. (2015), Distance- based decision tree algorithms for label ranking, in ‘Progress in Artificial Intelligence - 17th Portuguese Conference on Artificial Intelligence, EPIA 2015, Coimbra, Portugal, September 8-11, 2015. Proceedings’, pp. 525– 534. de Sá, C. R., Soares, C. & Knobbe, A. J. (2016), ‘Entropy-based discretiza- tion methods for ranking data’, Inf. Sci. 329, 921–936. de Sá, C. R., Soares, C., Jorge, A. M., Azevedo, P. J. & da Costa, J. P. (2011), Mining association rules for label ranking, in ‘Advances in Knowledge Dis- covery and Data Mining - 15th Pacific-Asia Conference, PAKDD 2011, Shenzhen, China, May 24-27, 2011, Proceedings, Part II’, pp. 432–443. Dinno, A. (2015), dunn.test: Dunn’s Test of Multiple Comparisons Using Rank Sums. R package version 1.2.3. Dwork, C., Kumar, R., Naor, M. & Sivakumar, D. (2001), Rank aggregation methods for the web, in ‘Proceedings of the Tenth International World Wide Web Conference, WWW 10, Hong Kong, China, May 1-5, 2001’, pp. 613–622. Fayyad, U. M. & Irani, K. B. (1993), Multi-interval discretization of continuous-valued attributes for classification learning, in ‘Proceedings 19 of the 13th International Joint Conference on Artificial Intelligence. Chambéry, France, August 28 - September 3, 1993’, pp. 1022–1029. Fürnkranz, J., Hüllermeier, E., Loza Menćıa, E. & Brinker, K. (2008), ‘Multilabel classification via calibrated label ranking’, Machine Learning 73(2), 133–153. Genuer, R., Poggi, J. & Tuleau-Malot, C. (2010), ‘Variable selection using random forests’, Pattern Recognition Letters 31(14), 2225–2236. Hüllermeier, E., Fürnkranz, J., Cheng, W. & Brinker, K. (2008), ‘Label ranking by learning pairwise preferences’, Artificial Intelligence 172(16- 17), 1897–1916. Kamishima, T. (2003), Nantonac collaborative filtering: recommendation based on order responses, in ‘Proceedings of the Ninth ACM SIGKDD In- ternational Conference on Knowledge Discovery and Data Mining, Wash- ington, DC, USA, August 24 - 27, 2003’, pp. 583–588. Kemeny, J. & Snell, J. (1972), Mathematical Models in the Social Sciences, MIT Press. Kendall, M. & Gibbons, J. (1970), Rank correlation methods, Griffin London. Mitchell, T. (1997), Machine Learning, McGraw-Hill. Neave, H. & Worthington, P. (1992), Distribution-free Tests, Routledge. Quinlan, J. R. (1986), ‘Induction of decision trees’, Machine Learning 1(1), 81–106. R Development Core Team (2010), R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0. Rebelo, C., Soares, C. & Costa, J. (2008), Empirical Evaluation of Rank- ing Trees on Some Metalearning Problems, in J. Chomicki, V. Conitzer, U. Junker & P. Perny, eds, ‘Proceedings 4th AAAI Multidisciplinary Work- shop on Advances in Preference Handling’. Ribeiro, G., Duivesteijn, W., Soares, C. & Knobbe, A. J. (2012), Multilayer perceptron for label ranking, in ‘Artificial Neural Networks and Machine Learning - ICANN 2012 - 22nd International Conference on Artificial Neu- ral Networks, Lausanne, Switzerland, September 11-14, 2012, Proceedings, Part II’, pp. 25–32. 20 Scornet, E., Biau, G. & Vert, J.-P. (2014), ‘Consistency of random forests’, ArXiv e-prints. Spearman, C. (1904), ‘The proof and measurement of association between two things’, American Journal of Psychology 15, 72–101. Todorovski, L., Blockeel, H. & Džeroski, S. (2002), Ranking with Predictive Clustering Trees, in T. Elomaa, H. Mannila & H. Toivonen, eds, ‘Proc. of the 13th European Conf. on Machine Learning’, number 2430 in ‘LNAI’, Springer-Verlag, pp. 444–455. Vembu, S. & Gärtner, T. (2010), Label ranking algorithms: A survey, in J. Fürnkranz & E. Hüllermeier, eds, ‘Preference Learning’, Springer- Verlag, pp. 45–64. Yasutake, S., Hatano, K., Takimoto, E. & Takeda, M. (2012), Online rank ag- gregation, in ‘Proceedings of the 4th Asian Conference on Machine Learn- ing, ACML 2012, Singapore, Singapore, November 4-6, 2012’, pp. 539–553. 21 
work_2s7yi5h4orcibame2ndlhum7lu	----	Multi-faceted Assessment of Trademark Similarity This is an Open Access document downloaded from ORCA, Cardiff University's institutional repository: http://orca.cf.ac.uk/93588/ This is the author’s version of a work that was submitted to / accepted for publication. Citation for final published version: Setchi, Rossitza 2016. Multi-faceted Assessment of Trademark Similarity. Expert Systems with Applications 65 , pp. 16-27. 10.1016/j.eswa.2016.08.028 file Publishers page: http://dx.doi.org/10.1016/j.eswa.2016.08.028 <http://dx.doi.org/10.1016/j.eswa.2016.08.028> Please note: Changes made as a result of publishing processes such as copy-editing, formatting and page numbers may not be reflected in this version. For the definitive version of this publication, please refer to the published source. You are advised to consult the publisher’s version if you wish to cite this paper. This version is being made available in accordance with publisher policies. See http://orca.cf.ac.uk/policies.html for usage policies. Copyright and moral rights for publications made available in ORCA are retained by the copyright holders. Accepted Manuscript Multi-faceted Assessment of Trademark Similarity Rossitza Setchi , Fatahiyah Mohd Anuar PII: S0957-4174(16)30421-3 DOI: 10.1016/j.eswa.2016.08.028 Reference: ESWA 10821 To appear in: Expert Systems With Applications Received date: 22 December 2015 Revised date: 3 August 2016 Accepted date: 4 August 2016 Please cite this article as: Rossitza Setchi , Fatahiyah Mohd Anuar , Multi-faceted Assessment of Trademark Similarity, Expert Systems With Applications (2016), doi: 10.1016/j.eswa.2016.08.028 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. http://dx.doi.org/10.1016/j.eswa.2016.08.028 http://dx.doi.org/10.1016/j.eswa.2016.08.028 ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Highlights  A novel method for the assessment of trademark similarity is proposed.  The method blends together visual, semantic and phonetic similarity.  It produces an aggregated score based on the individual assessments.  Evaluation using information retrieval measures and human judgment. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Multi-faceted Assessment of Trademark Similarity Rossitza Setchi 1 and Fatahiyah Mohd Anuar 1,2 , 1 School of Engineering, Cardiff University, 14-17 The Parade, Cardiff CF24 3AA, UK E-mail: setchi@cf.ac.uk 2 Faculty of Engineering, Multimedia University, Cyberjaya, 63100, Malaysia E-mail: fatahiyah@mmu.edu.my Abstract—Trademarks are intellectual property assets with potentially high reputational value. Their infringement may lead to lost revenue, lower profits and damages to brand reputation. A test normally conducted to check whether a trademark is highly likely to infringe other existing, already registered, trademarks is called a likelihood of confusion test. One of the most influential factors in this test is establishing similarity in appearance, meaning or sound. However, even though the trademark registration process suggests a multi-faceted similarity assessment, relevant research in expert systems mainly focuses on computing individual aspects of similarity between trademarks. Therefore, this paper contributes to the knowledge in this field by proposing a method, which, similar to the way people perceive trademarks, blends together the three fundamental aspects of trademark similarity and produces an aggregated score based on the individual visual, semantic and phonetic assessments. In particular, semantic similarity is a new aspect, which has not been considered by other researchers in approaches aimed at providing decision support in trademark similarity assessment. Another specific scientific contribution of this paper is the innovative integration, using a fuzzy engine, of three independent assessments, which collectively provide a more balanced and human-centered view on potential infringement problems. In addition, the paper introduces the concept of degree of similarity since the line between similar and dissimilar trademarks is not always easy to define especially when dealing with blending three very different assessments. The work described in the paper is evaluated using a database comprising 1,400 trademarks compiled from a collection of real legal cases of trademark disputes. The evaluation involved two experiments. The first ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T experiment employed information retrieval measures to test the classification accuracy of the proposed method while the second used human collective opinion to examine correlations between the trademark scoring/rating and the ranking of the proposed method, and human judgment. In the first experiment, the proposed method improved the F-score, precision and accuracy of classification by 12.5%, 35% and 8.3%, respectively, against the best score computed using individual similarity. In the second experiment, the proposed method produced a perfect positive Spearman rank correlation score of 1.00 in the ranking task and a pairwise Pearson correlation score of 0.92 in the rating task. The test of significance conducted on both scores rejected the null hypotheses of the experiment and showed that both scores correlated well with collective human judgment. The combined overall assessment could add value to existing support systems and be beneficial for both trademark examiners and trademark applicants. The method could be further used in addressing recent cyberspace phenomena related to trademark infringement such as customer hijacking and cybersquatting. Keywords—Trademark assessment, trademark infringement, trademark retrieval, degree of similarity, fuzzy aggregation, semantic similarity, phonetic similarity, visual similarity. 1. Introduction Trademarks are valuable intellectual property (IP) assets that identify the commercial source or origin of products or services. They are visual signs in the form of logos or brand names that allow goods or services to be easily recognized and distinguished by consumers. Similar to other intangible company assets, trademarks can be subject to legal protection. Trademark registration through an IP office provides legal protection for companies and individuals on registered marks in the jurisdiction(s) that the registration office covers. It therefore provides legal certainty and underpins the right of the trademark owner. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Trademark infringement is a form of IP crime that may lead to lost revenue, lower profits and additional costs, such as the legal fees necessary to enforce a trademark. In addition, trademark infringement is time-consuming when enforcing rights and, perhaps more importantly, can lead to severe damage of brand reputation. Recent statistics show that trademark infringement has become a serious economic and legal issue. For example, the United States International Trade Commission, as reported by the Chairman of the Joint Economic Committee, stated that the number of investigated infringement cases rose from the year 2010 to 2011 by 23.2%. A total of 3,400 trademark infringement cases were filed in the US District Courts in 2012, which excluded a presumably even larger number of cases where settlements were reached prior to the filing of cases (Scott, 2013). Some of the reported cases involve new cybercrime phenomena such as customer hijacking and cybersquatting (Scott, 2013). In another investigation conducted by the US International Trade Commission in 2011, the average annual increase of trademark litigation cases concerning US-based companies from 2002–2011 was 39.8% (US International Trade Commission, 2011). Despite these alarming trademark infringement statistics, the number of newly registered trademarks together with the existing trademarks used in the market continues to grow (Office for Harmonization in the Internal Market [OHIM], 2012; Dodell, 2013). This trend, which has been observed worldwide, has recently created administrative problems for many trademark registration offices as the registration process has become more complex and lengthy. The trademark registration process includes a trademark similarity examination (OHIM, 2014), which requires a multi-faceted similarity assessment. One of the steps involved is making sure that the trademark to be registered is not similar to any trademark that has already been registered, as the registration of trademarks that are found to be identical or similar to any existing trademarks and provide identical or similar goods or services may potentially be opposed, as indicated in section 5 of the Trade Marks Act 1994 (Trade Marks Act, 1994). This is important in order to avoid infringements and protect the rights of existing registered trademarks. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T The current practice of examining trademark similarity generally involves a search to retrieve relevant trademarks from a very large trademark database on the basis of a specific type of similarity. For example, the Industrial Property Automation System (IPAS), a support system developed by the World Industrial Property Organisation (WIPO), provides three trademark search options, namely a bibliography search based on the filing date and registration number, a phonetic search based on phonetic rules and common prefixes and suffixes, and a logo search based on the Vienna classification code for figurative trademarks (WIPO, 2014). The research in this paper is motivated by the guidelines in the trademark examination manual, which require overall similarity assessment. From a theoretical point of view, the paper contributes to the body of knowledge in the area of intelligent human- centered decision support and in particular the use of fuzzy logic and semantics in complex evaluations and assessments related to infringement and the likelihood of confusion. Previous research has addressed some of these aspects to a certain degree. For example, the need to consider many facets or aspects in complex evaluations has been recognized by a number of researchers working in various domains. Many of them employ fuzzy logic, which is a particularly suitable reasoning technique in domains where the selection of the best alternative is highly complex and the judgement is based on subjective perceptions (Mardani et al., 2015). For example, a knowledge evaluation method aimed at estimating the quality of knowledge and its market value uses fuzzy logic to aggregate several aspects including knowledge complexity, marketable value, and the reputation of the knowledge supplier (Chen, 2011). Fuzzy numbers are also used to calculate the value of a patent and the chance of mitigation (Agliardi and Agliardi, 2011), which similar to quality of knowledge in the above example, are also parameters very difficult to measure objectively. Semantics and fuzzy logic are employed in group decision making (Gupta and Mohanty, 2016), consensus building (Li et al., 2017), opinion mining (Martínez-Cruz et al., 2016) and knowledge management (Li et al., 2011). ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T This paper offers an original approach to the problem of trademark infringement, which is based on multi-facet assessment and verified through human judgement. The proposed computational method for assessing trademark similarity employs multi-faceted evaluation of the three main aspects of trademark similarity: visual, semantic and phonetic. In particular, semantic similarity is a new aspect which has not been considered in any previous approaches aimed at developing decision support systems for trademark similarity assessment. Therefore, the specific scientific contribution of this paper is the innovative integration, using a fuzzy engine, of three independent assessments, which collectively provide a more balanced view on potential infringement problems. The combined overall assessment could add value to existing support systems and be beneficial for both trademark examiners and trademark applicants. The rest of the paper is organized as follows: The next section provides an overview of existing trademark search systems and briefly discusses fuzzy logic, the inference concept employed in this research. The proposed computational method is introduced in Section 3. Section 4 describes the experimental setup and presents the results. A discussion is provided in Section 5. Section 6 concludes the study. 2. Related Work This section reviews related work in the scope of this study. It consists of two subsections. The first subsection reviews existing trademark search systems, and the second subsection briefly discusses the concept of fuzzy inference, which inspired the development of the proposed method for the multi-faceted assessment of trademark similarity. 2.1 Existing Trademark Search Systems Table 1 shows examples of trademarks with different types of similarity: visual, semantic and phonetic. The trademark pair NEXT and NEST possess some degree of visual similarity due to the total number of letters and the number of identical letters used. In addition, although NEXT is a figurative trademark, its style/font is similar to the typeface font ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T of the trademark NEST, which contributes to the visual similarity between them. The second pair, MAGIC TIMES and MAGIC HOUR, are semantically similar due to the identical word that they share and the lexical relation between the non-identical words in the trademark text. The last pair, i.e. SVIZZEROTALER and SWISS TALER, are phonetically similar because although these trademarks are spelled differently, their pronunciation is similar. <insert Table 1 here> Many trademarks share more than one type of similarity; however, despite the existing variety in the types of similarity, most of the research in this area is focused on retrieving trademarks based on their visual similarity using low-level features. Examples of such systems include TRADEMARK (Kato et al., 1990), STAR (Wu et al., 1996) and ARTISAN (Eakins et al., 1996), which have been widely referred to by many researchers. TRADEMARK uses graphical descriptor vectors derived from shape features while STAR employs a traditional content-based image retrieval (CBIR) framework together with a set of shape-based descriptors, including Fourier descriptors, gray-level projection and moment invariants. In addition, it utilizes the spatial layout of the images although this has been found to be extremely challenging. ARTISAN also utilizes shape-based feature descriptors but includes Gestalt-based principles to retrieve abstract geometric trademark designs. These three studies have inspired further research in trademark image retrieval focused on the visual similarity aspect of trademarks. For example, Kim and Kim (1998) employed a moment-based shape descriptor and analyzed the distribution model of 90 moment orders for all the images in their database. A closed contour shape descriptor using angle code strings was developed by Peng and Chen (1998). Jain and Vailaya (1998) proposed the use of the edge direction histogram and improved the descriptor so that it became scale and rotation invariant. Other research includes a comparative study of several common shape-based descriptors for trademark similarity comparison (Eakins et al., 2003) and a compositional shape descriptor that combines several shape descriptors (Hong & Jiang, 2008; Wei et al., 2009). ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Despite the amount of work undertaken, visual similarity assessment is mainly limited to trademarks with figurative marks or logos. Notwithstanding, statistics of registered trademarks in five European countries have shown that only 30% of all trademarks employ logos as their proprietary marks (Schietse et al., 2007). The trademark similarity of the remaining 70% of registered trademarks is still insufficiently researched. For example, despite the recent advances in computational semantics, the existing trademark search systems that focus on text are primarily built around keyword-based retrieval or approximate string matching. Such systems return trademarks that match parts or entire words in query text. In Europe, OHIM recently launched a search system that allows trademark applicants and third parties to search for trademarks in different languages (OHIM, 2013). The system also provides an advanced search option that offers three search types: word prefix, full phrase and exact match. In the United Kingdom, the UK Intellectual Property Office (IPO) offers similar search options with an additional option that looks for similar query strings (UK IPO, 2013). The IPO search system utilises an approximate string-matching technique, which looks for fairly similar patterns in strings, together with several predefined criteria including word length and the number of similar and dissimilar letters shared by the words. Despite their usefulness, the comparison mechanism employed in such systems limits their effectiveness as it does not cover all similarity aspects that are normally assessed during the trademark examination process. Advances in computational semantics provide an opportunity to overcome the limitations of traditional text-based retrieval by exploring semantic similarity. In the context of trademark similarity examination and analysis, it allows the comparison of trademarks based on their semantic similarity derived using external knowledge sources such as lexical ontologies. From the point of view of knowledge engineering, a lexical ontology is a framework that specifies the underlying structure and lexical relationships for knowledge representation and the organization of lexical information (Storey et al., 1998). On the other hand, advances in computational linguistics and genealogy provide a mechanism to ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T compare trademarks based on their phonetic similarity (Convington, 1998; Kondrak, 2003; Pfeifer et al., 1996; Philips, 2000). This includes computational linguistics studies of similarities between cognates, i.e. words from different languages that share the same linguistic origin and etymology, and name-matching applications in genealogy, which retrieve similar names despite spelling variations. This research promotes the view that the existing work on visual similarity can be extended using the recent advances in semantic retrieval, computational linguistics and computational genealogy. This approach is consistent with the requirement for holistic assessment outlined in the OHIM trademark manual (OHIM, 2014). Trademark comparison based on visual, semantic and phonetic similarity, individually, has been the paramount focus of the present authors’ previous work (Anuar et al., 2013; 2014; 2016). The main contribution of this paper is that it extends previous approaches by providing a consolidated holistic assessment process. In addition, the paper introduces the concept of degree of similarity since the line between similar and dissimilar is not always easy to define. 2.1 Fuzzy Logic Studies on information retrieval of music and artist recommendations (McFee & Lanckriet, 2009; Zhang et al., 2009) compute multi-faceted similarity based on low-level features and subjective criteria. Fuzzy logic has not yet been applied to multi-faceted similarity assessment but has been used in many applications that require human reasoning and decision-making. Examples include control systems in the engineering domain, doctor– patient decision-making in the medical domain, and risk analysis in e-commerce (Abou & Saleh, 2011; Fazzolari et al., 2013; Ngai & Wat, 2005). Furthermore, the concept of fuzzy logic has long been recognized in legal studies (Cook, 2001; Kosko, 1994), which is an important consideration in the area of IP rights protection. This paper promotes the use of fuzzy logic to compute the degree of similarity between trademarks due to its natural modelling capability that can mimic the very complex system underlying the human mind. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T The concept of fuzzy logic was first introduced as a mathematical tool for dealing with uncertainty (Zadeh, 1965). From the point of view of set theory, the concept of fuzzy logic is an extension of the crisp set concept in which every preposition must be either ‘true’ or ‘false’ or in a range of values. Instead, fuzzy logic asserts that every preposition can simultaneously have a certain degree of a membership function of the ‘true’ or ‘false’ class. An inference system based on fuzzy logic uses fuzzy set operations and properties for reasoning and consists of a fuzzy rule base. A fuzzy rule generally has two components, the IF component, i.e. the antecedent, which describes a condition, and the THEN component, i.e. the consequent, which describes a conclusion. It follows the format: IF <antecedent>, THEN <consequent> (1) In the context of a human-oriented process that requires approximate human reasoning or decision-making based on experiences and insights, a human inference system tends to use verbal variables to create verbal rules in a form similar to Eq. 1. Since the terms and variables used in human inference systems are normally fuzzy rather than precise, a fuzzy inference system is highly applicable in such applications. Verbal terms and variables can therefore be expressed mathematically as membership degrees and membership functions with symbolic verbal phrases rather than numeric values. Indirectly, this provides a systematic mechanism to utilize the uncertain and imprecise information used in human judgment. The implementation of the fuzzy inference approach in various applications commonly involves two inference models, i.e. the Mamdani inference model, which is based on a fuzzy relational model, and the Takagi–Sugeno inference model (Akgun, 2012). Both models employ slightly different approaches in the output aggregation process in that Mamdani uses defuzzification and Takagi–Sugeno employs weighted average to compute the crisp output. An alternative approach is the Tsukamoto model, which represents the consequent of the fuzzy rules with monotonical membership functions (Jang et al., 1997). A ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T more recent approach is the inference model based on a combination of adaptive neural networks and fuzzy logic (Leng et al., 2009). The Mamdani inference model is employed in this paper due to its intuitive and linguistic model applicability, which makes it very suitable for human-oriented applications. 3. Trademark Degree-of-Similarity Aggregation Method This section introduces the proposed method and highlights the main steps involved in it. The method was based on a systematic analysis of 1,400 trademarks extracted from real dispute cases. This analysis revealed that the trademark cases in the collection were either real words/phrases such as ‘MAGIC HOUR’, out-of-vocabulary words/phrases such as ‘SVIZZEROTALER’ or a combination of both. In addition, the analysis also showed that in cases involving only out-of-vocabulary words, only visual and phonetic assessments were performed since such words do not carry any lexical meaning. The four different types of trademarks defined in OHIM (2014), namely word mark, figurative word mark, purely figurative mark and purely figurative mark with figurative word mark (Fig. 1), require different processing techniques and analytical approaches, hence the development of a method that facilitates the similarity comparison of both real words and out-of-vocabulary words. <insert Fig. 1 here> The conceptual model of the proposed system (Fig. 2) comprises four main modules. Three of these modules assess trademarks in terms of their visual, semantic and phonetic similarity while the fourth module, the fuzzy inference engine, aggregates the final score based on the three individual assessments. Each module has its individual functional requirements and uses a different approach to achieve its predefined function. For example, the visual similarity module employs visual descriptors based on the shape features of the individual letters included in the trademarks. Fig. 3 shows the flowchart of the proposed method and the four individual steps involved: (i) individual assessments, (ii) fuzzification, (iii) inference and (iv) defuzzification. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T <insert Fig. 2 here> <insert Fig. 3 here> 3.1 Step 1: Assessment of Visual, Semantic and Phonetic Similarity This step involves the assessment of the three main aspects of similarity. The visual similarity assessment of purely figurative trademarks such as logos is computed using an advanced algorithm (Anuar et al., 2013) that employs global and local shape features, i.e. Zernike moment and an edge-gradient co-occurrence matrix, represented as vectors. The similarity between the trademarks is then computed using normalized Euclidean distances between their corresponding vectors. The same approach, combined with the string algorithm (Navarro, 2001) is used to compute the visual similarity of trademarks with word marks and figurative word marks (Table 2). Unlike approximate string matching that uses binary values in the letter-to-letter comparison, such as ‘1’ and ‘I’ in the example shown in Fig. 4, the algorithm developed in this paper computes the visual similarity between letters using their shape descriptors. This provides a mechanism that differentiates between different letters and numbers that look similar, such as ‘1’ and ‘I’, and less similar letters and numbers, such as ‘1’ and ‘X’. Table 3 shows that the proposed algorithm exhibits better discriminating power compared to approximate string matching. <insert Table 2 here> <insert Fig. 4 here> <insert Table 3 here> The trademark semantic similarity assessment is based on a similarity computation model (Anuar et al., in press), which utilizes a lexical ontology, i.e. WordNet, as an external knowledge source. WordNet is a large electronic lexical database of the English language that is freely available and was developed based on psycholinguistic theories that model human semantic organization. It has been extended to over 30 different languages, ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T including Dutch, Spanish, German, Basque and Arabic (Abouenour et al., 2013; Fernandez- Montraveta et al., 2008; Gonzalo et al., 1999; Hinrichs et al., 2013; Pociello et al., 2011). The computation of semantic trademark similarity uses two sets of features to represent each trademark: the token feature set and the synonyms feature set. The token feature set consists of a set of words included in the trademark. For example, the token feature set for the trademark ‘Red Bull’ is (red, bull). The synonym feature set on the other hand comprises synonyms, direct hypernyms, i.e. words that are more general in meaning in the taxonomic hierarchy, and direct hyponyms, i.e. words that are instances of their corresponding trademark tokens. The similarity score is computed using a combination of Tversky’s contrast model of similarity (Tversky, 1977), which considers the number of shared features, together with the edge-based word similarity score between the tokens, derived using lexical ontology, i.e. WordNet. Finally, the phonetic similarity assessment computes trademark similarity based on the phonological features of the phonemes in the trademark text combined with typographic mapping and a token rearrangement process (Anuar et al., 2014). The algorithm represents the phonemes in a word string as vectors with phonetic features where each vector consists of 10 binary main features and two multi-valued features extracted from the phonological properties of human speech production (Kondrak, 2003). The algorithm differentiates between more similar phoneme pairs, such as ‘m’ and ‘n’, and less similar phoneme pairs, such as ‘m’ and ‘p’. In addition, the algorithm converts special characters or symbols in the trademark text to their corresponding meaning. For example, the ampersand symbol ‘&’ is substituted by ‘and’. This conversion allows typographic symbols to be processed in the way regular words appearing in trademarks are handled. 3.2 Step 2: Fuzzification The fuzzification step is the process of mapping the crisp values of the input variables to fuzzy sets. Three input variables corresponding to the visual, semantic and phonetic assessments are fuzzified in this step using five triangular-based membership ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T functions, as defined in Eq. 2. These functions were employed in this study because of their simplicity and good performance, which have been proven theoretically (Barua et al., 2014) and used in various engineering and non-engineering applications (Gañán et al., 2012; Kaur & Kaur, 2012; Ngai & Wat, 2005). Moreover, these functions have recently been used in a court case decision-making study that included traffic violations and crime cases (Sabahi & Akbarzadeh-T, 2014). A graphical representation of the input membership functions is shown in Fig. 5. 175.0, 25.0 75.0 75.0,0 )( 1,0 175.0, 25.0 1 75.05.0, 25.0 5.0 5.0,0 )( 75.0,0 75.05.0, 25.0 75.0 5.025.0, 25.0 25.0 25.0,0 )( 5.0,0 5.025.0, 25.0 5.0 25.00, 25.0 0,0 )( 25.0,0 25.00, 25.0 25.0 )( 5 43 21                    x x x xf x x x x x x xf x x x x x x xf x x x x x x xf x x x xf (2) <insert Fig. 5 here> 3.3 Step 3: Inference This step uses the Mamdani fuzzy inference model, a well-known inference model used in various fuzzy logic-based applications (Abou & Saleh, 2011; Akgun et al., 2012; Chatzichristofis et al., 2012). A set of fuzzy rules was first developed based on the OHIM trademark examination manual (OHIM, 2014) and an empirical study of 1,400 trademarks involved in dispute cases. The rules are expressed in tabular form using five two- dimensional fuzzy associative matrices, which correspond to a total of 125 rules. Fig. 6 shows the five associative matrices of the developed rules. Five input and output conditions are associated with each rule: very low (VL), low (L), medium (M), high (H), and very high (VH). Each cell in the associative matrices corresponds to the output condition triggered by ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T the rules associated with the condition of the input variables. For example, the verbal rule corresponding to the first cell of matrix (c) in Fig. 6 is translated as ‘IF the phonetic score IS M (medium) and the semantic score IS VL (very low) and the visual score IS VL (very low), THEN the output score IS L (low)’. The output membership functions that correspond to the five output conditions also consist of five triangular-based functions, as in Eq. 3. A graphical representation of these functions is shown in Fig. 7. <insert Fig. 6 here> <insert Fig. 7 here> f 1 (x) = 0.2 - x 0.3 , 0 £ x £ 0.2 0, x ³ 0.2 f 2 (x) = 0, x ³ 0 x + 0.1 0.3 , 0 £ x £ 0.2 0.5 - x 0.3 , 0.2 £ x £ 0.5 0, x ³ 0.5 f 3 (x) = 0, x £ 0.25 x - 0.2 0.3 , 0.2 £ x £ 0.5 0.8 - x 0.25 , 0.5 £ x £ 0.8 0, x ³ 0.75 f 4 (x) = 0, x £ 0.5 x - 0.5 0.3 , 0.5 £ x £ 0.8 1.1- x 0.3 , 0.8 £ x £ 1 f 5 (x) = 0, x £ 0.8 x - 0.8 0.3 , 0.8 £ x £ 1 (3) The aggregation of the compositional output involves a fuzzy operation between the fuzzified input and the fuzzy relations established by the rules. It is derived using the implication–aggregation (min–max) method (Akgun et al., 2012): m 0 = max(min(m i 1 (k),m i 2 (k),m i 3 (k))) (4) where m i 1 ,m 1 2 ,m i 3 are the mapping of the first, second and third inputs from the crisp set to the fuzzy set, i.e. the visual, semantic and phonetic similarity scores, respectively, and k is the k-th IF–THEN preposition, or the fuzzy rule. 3.4 Step 4: Defuzzification ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T This step uses the centroid or centre of mass defuzzification method to quantify the compositional output from the fuzzy set to the real output that corresponds to the degree-of- similarity value. It computes the centroid under the curve resulting from the compositional operation performed during the inference step. The centroid computation is given by the following equation: centroid = f(x) ×x dxò f(x) dxò (5) where f(x) is the membership function associated with the compositional output. Fig. 8 shows an illustrative example of the proposed aggregation process for the trademark pair SKYPINE and SKYLINE. Their degree of similarity was computed as 0.798. <insert Fig. 8 here> 4. Experimental Setup and Results This section describes the two experiments performed in this study and the evaluation method used to conduct them. The first experiment evaluated the proposed method from a computational point of view using information retrieval measures. The second experiment was designed to capture human perception, i.e. the way people view similarity in trademarks. 4.1 Experiment 1 The main objective of the first experiment was to test the classification performance of the proposed method when differentiating between possible cases of infringement. The developed method was compared to the traditional approach of considering the individual aspects of similarity. The experiment employed information retrieval measures such as F- score, precision score and accuracy. The scores were derived from the classification confusion matrix shown in Table 4, where TP, FP, FN and TN refer to true positive, false positive, false negative and true negative, respectively. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T <insert Table 4 here> A collection of real court cases comprising 1,400 trademarks (Schweizer, 2013) was analysed and used to create a database. An excerpt from a court case report for two disputed trademarks, AURA and AUREA, is shown below. It provides the conclusion and rationale of the experts investigating this particular case. Based on such findings, the database was then split into two groups, i.e. with degree of similarity that may or may not lead to confusion as judged by the experts. On the visual level, the trademarks have a strong similarity in the sense that the length of the verbal elements is almost identical (AURA/AUREA), i.e. four against five letters. Only the vowel ‘E’ of the contested trademark differs from the four letters of ‘AURA’ trademark. The overall visual impression is therefore very similar. Aurally, the signs are also very similar. The vowel ‘E’ can be easily used. The overall phonetic impression is also very similar. Although that there is no semantic similarity, the risk of misperception on trademarks does exist due to high visual and phonetic similarity. The fact that the opponent has an additional letter ‘E’ does not change the overall similarity finding. In view of that, the similarity of the trademarks is therefore recognized. For evaluation purposes, a repeated holdout evaluation procedure was performed in which the database was divided into two random disjoint training (50%) and testing (50%) sets. The training set was used to obtain a threshold score to classify the dataset employed in this experiment. Pairwise degrees of similarity scores between the trademark pairs in the training set were first computed using the proposed method. A histogram-based thresholding algorithm (Nobuyuki, 1979) was then used to estimate the threshold value of the computed degree-of-similarity scores by exhaustive searches for a value that minimized the intra-class variance of the binary classes. The threshold value obtained from the training set was then used to classify the data in the testing set. This procedure was repeated 1,000 times and in each repetition the F-score, precision and accuracy were computed using Eqs. 6–8: F - score = 2TP TP +FP + TP +FN (6) precision = TP TP +FP (7) ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T accuracy = TP + TN Total Data (8) where TP, TN, FP and FN are the true positive, true negative, false positive and false negative trademarks, respectively, as classified by the binary classification performed in this experiment, and Total Data (calculated as 700) is the total number of trademark pairs in the database. The average scores were then used to evaluate the overall performance of the proposed method. The procedure was repeated using the scores from the individual assessments of visual, semantic and phonetic similarity. Table 5 shows the classification results obtained using the three individual similarity assessments and the proposed method. <insert Table 5 here> 4.2 Experiment 2 The main objective of the second experiment was to prove the following two hypotheses: 1. The similarity ranking of the trademark pairs produced by the proposed method correlates with human collective judgment. 2. The similarity rating of each trademark pair produced by the proposed method correlates with human collective judgment. Two significance tests were performed using the Spearman rank correlation score and the Pearson pairwise correlation score to statistically prove these hypotheses and reject the null hypotheses of this experiment. The Spearman rank correlation score, which takes values in the range of -1 to 1 (both -1 and 1 being the negative and positive perfect correlations, respectively, and 0 indicating no correlation), is a measure of statistical dependence between two ranked variables. The score indicates how strong the relationship between the ranked variable can be and is described using a monotonic function. The Pearson pairwise correlation score on the other hand measures the strength of a linear association between two variables. The Pearson correlation attempts to draw a line of best fit through the values ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T of two variables; the score itself describes the dispersion of the data points from the line of best fit. The Pearson correlation score has the same value range as the Spearman rank correlation score. As it involved human judgment, this experiment used a crowdsourcing platform for evaluation purposes. Crowdsourcing is an open call task recently introduced in information retrieval studies and has been proven to produce fast and reliable results in a cost-effective way (Corney, 2010; Fadzli & Setchi, 2012; Snow et al., 2008). This task, commonly known as a human intelligence task (HIT), is a small portion of an even larger task distributed among a large group of workers without any apparent contact. A total of 25 trademarks were randomly selected from the database used in Experiment 1 as a query set in Experiment 2. The trademark similarity assessment system developed in this study was then used to rank the set of trademarks returned from each query from the highest degree-of-similarity (ds) score to the lowest. Three trademarks with high (ds > 3.5), medium (2.0 < ds ≤ 3.5) and low (ds ≤ 2.0) distribution scores were selected from the retrieved set and used in the crowdsourcing task. Table 6 shows the 25 queries used in this experiment together with the three retrieved results classified by the proposed method as having high, medium and low similarity, respectively. Fig. 9 shows one of the HITs used in the experiment. <insert Table 6 here> <insert Fig. 9 here> In each HIT, the workers were presented with three different trademarks and asked to score their similarity with the query trademark using a scale from 1 to 5 (1 being the least similar and 5 being the most similar). Each query was evaluated by 20 different workers, which resulted in a total of 500 HITs. The selection of the HIT workers was based on two criteria: the number and acceptance rate of their previously completed assignments. The first criterion required the workers to have completed at least 1,000 HITs. The acceptance ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T rate of the previously completed HITs was set to 95%, indicating the approval level of the work done as evidenced by their HITs requestors. These two criteria were introduced to ensure the quality of the collected feedback. Next, the average similarity scores for each query given by the workers were computed and compared with the normalized similarity score produced by the proposed method (Table 7). The similarity scores (Fig. 10) were used to compute the Spearman rank correlation score and the Pearson pairwise correlation score shown in Table 8. <insert Table 7 here> <insert Fig. 10 here> <insert Table 8 here> 5. Discussion The first experiment verified the classification performance of the proposed multi- faceted method, which aggregated a similarity score based on all three similarity aspects (see Table 5). The method produced an F-score of 0.911, which translated into respective improvements of 15.2%, 150% and 12.5% compared to the F-scores produced using visual, semantic and phonetic similarity individually. Among these three similarity aspects, phonetic similarity produced the best F-score (0.810) while semantic similarity showed the worst performance in terms of F-score (0.364). The proposed method also surpassed the three individual similarity aspects in terms of precision. With a precision score of 0.924, it improved the individual performance of the visual, semantic and phonetic similarity assessments by 35%, 312% and 35.4%, respectively. Similar improvements were demonstrated in terms of accuracy. The proposed method produced an accuracy score of 0.910 compared to the accuracy produced using visual, semantic and phonetic similarity (0.819, 0.610 and 0.840, respectively), which resulted in improvements of 11%, 49% and 8.3%, respectively. Overall, the results from the first experiment clearly show that the proposed degree-of-similarity aggregation method has the best classification performance ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T compared to assessments based on individual similarity aspects. Moreover, this approach is well aligned to the recommended trademark examination procedure, which requires trademarks to be examined in a holistic way. The second experiment was designed to investigate the performance of the proposed method in comparison with human collective judgment. Two correlation measures, the Spearman rank correlation and the Pearson pairwise correlation, were used to statistically prove the hypotheses. The proposed method obtained a perfect Spearman rank score of 1 and a Pearson pairwise correlation score of 0.92. A statistical significance test performed on both correlation scores rejected the null hypotheses of the experiment and indirectly proved that the degree-of-similarity scores produced by the proposed method correlated well with human collective judgment on trademark overall similarity. This strong correlation can be also observed in the scatter plot shown in Fig. 10, which displays a concentration of almost all points along the best-fit line (the straight black line on the graph). 5. Conclusions A support system to assess the overall degree of similarity between trademarks is essential for trademark protection so the work presented in this paper was motivated by the need to help prevent trademark infringement by identifying existing similarities between trademarks. This paper contributes to the body of knowledge in this area by the development of a method that measures the degree of similarity between trademarks on the basis of all three aspects of similarity: visual, semantic and phonetic. The method uses fuzzy logic to aggregate the overall assessment, which provides a more balanced and human-centered view on potential infringement problems. In addition, the paper introduces the concept of degree of similarity since the line between similar and dissimilar trademarks is not always easy to define especially when dealing with blending three very different assessments. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T One of the strengths of the proposed method is its rigorous evaluation using a large, purpose-built collection of real legal cases of trademark disputes. Moreover, the experiments performed in this study examined the performance of the proposed method from two points of view. First, the relative performance of the method was investigated from an information retrieval perspective in terms of classification performance. Using a crowdsourcing platform, the second experiment investigated the performance of the method relative to human judgment. The results of the experiments confirmed that there is a significant improvement in trademark similarity assessment when all similarity aspects are carefully considered. The results also showed that the proposed method demonstrates a statistically significant correlation against human collective judgment. Therefore, the experiments convincingly validated both original hypotheses outlined in this study. In conclusion, the proposed system can provide a support mechanism in the trademark similarity analysis performed by trademark examiners during trademark registration. Moreover, the method for assessing the trademark similarity could be extended to address recent cyberspace phenomena such as consumer hijacking and cybersquatting. A particular limitation of the proposed work is its focus on only one aspect of the concept of likelihood of confusion, i.e. computing the similarity between trademarks. In reality, there are several other factors influencing the perceptions of the consumers. Such factors include strength of the registered trademarks, proximity of the channels of trade, product relatedness and consumer traits (sophistication and care). Such a study, which is currently underway, requires a multi-disciplinary approach, which involves experts from business studies, marketing, psychology and engineering. Acknowledgements—The authors wish to acknowledge the help of Christopher Harrison, Peter Evans, William Morell and Rich Corken from the UK Intellectual Property Office in finalizing some of the ideas behind this research. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T References Abou, A., and Saleh, E. (2011). A Fuzzy decision support system for management of breast cancer. International Journal of Advanced Computer Science and Applications, 2(3), 34-40. Abouenour, L., Bouzoubaa, K., & Rosso, P. (2013). On the evaluation and improvement of Arabic WordNet coverage and usability. Language Resources and Evaluation, 47(3), 891- 917. Agliardi, E. and Agliardi, R. (2011). An Application of Fuzzy Methods to Evaluate a Patent under the Chance of Litigation. Expert Systems with Applications, 10, 13143-13148. Akgun, A., Sezer, E. A., Nefeslioglu, H. A., Gokceoglu, C., & Pradhan, B. (2012). An easy- to-use MATLAB program (MamLand) for the assessment of landslide susceptibility using a Mamdani fuzzy algorithm. Computers and Geosciences, 38(1), 23-34. Anuar, F. M., Setchi, R., and Lai, Y. K. (2013). Trademark image retrieval using an integrated shape descriptor. Expert Systems with Applications, 40, 105-121. Anuar, F. M., Setchi, R., and Lai, Y. K. (2014). Trademark retrieval based on phonetic similarity. IEEE International Conference of Systems, Man and Cybernetics, San Diego, USA, 1642-1647. Anuar, F. M., Setchi, R., and Lai, Y. K. (2016). Semantic retrieval of trademarks based on conceptual similarity. IEEE Transaction of Systems, Man and Cybernetics: System, vol. 46(2), pp. 220-233. Barua, A., Mudunuri, L.S., and Kosheleva, O. (2014). Why trapezoidal and triangular membership functions work so well: Towards a theoretical explanation. Journal of Uncertain Systems, 8(3), 164-168. Chatzichristofis, S. A., Zagoris, K., Boutalis, Y., and Arampatzis, A. (2012). A fuzzy rank- based late fusion method for image retrieval. In: Schoeffmann, K., Merialdo, B., Hauptmann, ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T A., Ngo, C-W., Andreopoulos, Y., Breiteneder, C. (Eds). Lecture Notes in Computer Science, 7131, 463-472. Berlin Heidelberg: Springer. Chen, T.-Y. (2011). Value Ontology-Based Multi-Aspect Intellectual Asset Valuation Method for Decision-Making Support in k-Commerce, Expert Systems with Applications, 38(5), 5471-5485. Cook, B. B. (2001). Fuzzy logic and judicial decision making. Judicature, 85(2), 70-100. Corney, J. R., Torres-, C., Jagadeesan, A. P., Yan, X. T., Regli, W. C., and Medellin, H. (2010). Putting the crowd to work in a knowledge-based factory. Advanced Engineering Informatics, 24(3), 243-250. Covington, M. A. (1998). Alignment of multiple languages for historical comparison. International Conference on Computational Linguistics, Montreal, Canada (pp. 275-280). Dodell, L. (2013) The trademark problem: Casualty insurance’s dirty little secret. Retrieved from: http://www.carriermanagement.com (accessed 14 Dec 2015). Eakins, J. P., Riley, K. J., and Edwards, J. D. (2003). Shape feature matching for trademark image retrieval. International Conference of Image and Video Retrieval, Urbana Champaign, USA (pp. 28–38). Eakins, J. P., Shields, K., and Boardman, J. (1996). ARTISAN – A shape retrieval system based on boundary family indexing. Storage and Retrieval for Still Image and Video Databases, iv, 2670, 17–28. Fadzli, S. A., and Setchi, R. (2012). Concept-based indexing of annotated images using semantic DNA. Engineering Applications of Artificial Intelligence, 25(8), 1644-1655. Fazzolari, M., Alcala, R., Nojima, Y., Ishibuchi, H., and Herrera, F. (2013). A review of the application of multi objective evolutionary fuzzy systems: Current status and further directions. IEEE Transactions on Fuzzy Systems, 21(1), 45-65. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Fernandez-Montraveta, A., Vazquez, G., and Fellbaum, C. (2008). The Spanish version of WordNet 3.0. In: Text Resources and Lexical Knowledge, 8, 175-182. Berlin and New York: Mouton de Gruyter. Gañán, C., Muñoz, J. L., Esparza, O., Mata, J., and Alins, J. (2012). Risk-based decision making for public key infrastructures using fuzzy logic. International Journal of Innovative Computing, Information and Control, 8(11), 7925-7942. Gonzalo, J., Verdejo, F. and Chugur, I. (1999). Using EuroWordNet in a concept-based approach to cross-language text retrieval. Applied Artificial Intelligence. 13(7), 647-678. Gupta, M. and Mohanty, B. K. (2016). An Algorithmic Approach to Group Decision Making Problems under Fuzzy and Dynamic Environment. Expert Systems with Applications, 55, 118-132. Hinrichs, E., Henrich, V., and Barkey, R. (2013). Using part-whole relations for automatic deduction of compound-internal relations in GermaNet. Language Resources and Evaluation, 47(3), 839-858. Hong, Z., and Jiang, Q. (2008). Hybrid content-based trademark retrieval using region and contour features. Advanced Information Networking and Applications Workshop, Okinawa, Japan (pp. 1163–1168). Jain, A. K., and Vailaya, A. (1998). Shape-based retrieval: A case study with trademark image databases. Pattern Recognition, 31(9), 1369–1390. Jang, J. S. R., Sun, C. T., and Mizutani, E. (1997). Fuzzy inference systems. Neuro-Fuzzy and Soft Computing: A Computational Approach to Learning and Machine Intelligence (pp. 73-91). Upper Saddle River, NJ: Prentice Hall. Kato, T., Fujimura, K., and Shimogaki, H. (1990). TRADEMARK. Multimedia image database system with intelligent human interface. Systems and Computers in Japan, 21(11), 33–46. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Kaur, A., and Kaur, A. (2012). Comparison of Mamdani-type and Sugeno-type fuzzy inference systems for air conditioning system. International Journal of Soft Computing and Engineering, 2, 323-325. Kim, Y. S., and Kim, W. Y. (1998). Content-based trademark retrieval system using a visually salient feature. Image and Vision Computing, 16(12–13), 931–939. Kondrak, G. (2003). Phonetic alignment and similarity. Computers and the Humanities, 37(3), 273-291. Kosko, B. (1994). Fuzzy Thinking: The New Science of Fuzzy Logic. New York: Hyperion. Leng, G., Zeng, X. J., and Keane, J. A. (2009). A hybrid learning algorithm with a similarity- based pruning strategy for self-adaptive neuro-fuzzy systems. Applied Soft Computing, 9(4), 1354-1366. Li, C.-C., Liu, L., and Li, C.-B. (2017). Personalized Individual Semantics In Computing With Words for Supporting Linguistic Group Decision Making. An Application on Consensus Reaching. Information Fusion, 33, 29-40. Li, M., Liu, L., and Li, C.-B. (2011). An approach to expert recommendation based on fuzzy linguistic method and fuzzy text classification in knowledge management systems. Expert Systems with Applications, 38, 8586-8596. Mardani, A., Jusoh, A. and Zavadskas, E. K. (2015). Fuzzy Multiple Criteria Decision- Making Techniques and Applications – Two Decades Review from 1994 to 2014. Expert Systems with Applications, 42(8), 4126-4148. McFee, B., and Lanckriet, G. R. (2009). Heterogeneous Embedding for Subjective Artist Similarity. International Society for Music Information Retrieval Conference, Kobe, Japan (pp. 513-518). Navarro, G. (2001). A guided tour to approximate string matching. ACM Computing Survey, 33(1), 31-88. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Ngai, E. W. T., and Wat, F. K. T. (2005). Fuzzy decision support system for risk analysis in e-commerce development. Decision Support Systems, 40(2), 235-255. Nobuyuki, O. (1979). A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man and Cybernetics, 9(1), 62-66. OHIM (2012). Annual report 2012. Retrieved from: https://oami.europa.eu (accessed 10 Dec 2013). OHIM (2013). Trademark Search. Available at: https://oami.europa.eu. (accessed 10 Dec 2013). OHIM (2014). Guidelines for examination in the office for harmonization in the internal market on community trade marks, part c opposition, section 2 identity and likelihood of confusion, chapter 3 comparison of signs. Retrieved from: https://oami.europa.eu (accessed 1 Feb 2014). Peng, H. L., and Chen, S. Y. (1997). Trademark shape recognition using closed contours. Pattern Recognition Letters, 18(8), 791–803. Pfeifer, U., Poersch, T., and Fuhr, N. (1996). Retrieval effectiveness of proper name search methods. Information Processing and Management, 32(6), 667-679. Philips, L. (2000). The double metaphone search algorithm. C/C++ Users Journal, 18(6), 38- 43. Pociello, E., Agirre, E., and Aldezabal, I. (2011). Methodology and construction of the Basque WordNet. Language Resources and Evaluation, 45(2), 121-142. Sabahi, F., and Akbarzadeh-T, M.R. (2014). Introducing validity in fuzzy probability for judicial decision-making. International Journal of Approximate Reasoning, 55(6), 1383-1403. Schietse, J., Eakins, J. P., and Veltkamp, R. C. (2007). Practice and challenges in trademark image retrieval. International Conference on Image and Video Retrieval, Amsterdam, The Netherlands (pp. 518–524). ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Schweizer, M. (2013). Trade Marks Court Cases Database. Retrieved from: http://decisions.ch (accessed 10 Dec 2013). Scott, C. D. (2013). Trademark strategy in the internet age: Customer hijacking and the doctrine of initial interest confusion. Journal of Retailing, 89(2), 176-189. Storey, V. C., Dey, D., Ullrich, H., and Sundaresan, S. (1998). An ontology-based expert system for database design. Data and Knowledge Engineering, 28(1), (31-46). Trade Marks Act 1994. (1994). Retrieved from: http://www.legislation.gov.uk/ (accessed 24 Nov 2014). Tversky, A. (1977). Features of similarity. Psychology Review, 84, 327–352. Snow, R., O'Connor, B., Jurafsky, D., and Ng, A. Y. (2008). Cheap and fast-but is it good? Evaluating non-expert annotations for natural language tasks. Conference on Empirical Methods in Natural Language Processing, Honolulu, Hawaii (pp. 254-263). USITC (2011). China: Effects of intellectual property infringement and indigenous innovation policies on the U.S. economy. Retrieved from: http://www.usitc.gov/ (accessed 20 Dec 2013). UKIPO (2013). Trademark Search. Available at: http://www.ipo.gov.uk. (accessed 10 Dec 2013). Wei, C. H., Li, Y., Chau, W. Y., and Li, C. T. (2009). Trademark image retrieval using synthetic features for describing global shape and interior structure. Pattern Recognition, 42(3), 386–394. WIPO (2014). WIPO IPAS Functional and Technical Overview. Retrieved from: https://www3.wipo.int/confluence/display/wipoimd/WIPO+IPAS+Functional+and+Technical+ Overview (accessed 24 Nov 2014). Wu, J. K., Lam, C. P., Mehtre, B. M., Gao, Y. J., and Narasimhalu, A. D. (1996). Content- based retrieval for trademark registration. Multimedia Tools and Applications, 3(3), 245–267. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8, 338-353. Zhang, B., Shen, J., Xiang, Q., and Wang, Y. (2009). CompositeMap: a novel framework for music similarity measure. 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, Boston, MA, USA (pp. 403-410). ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Table 1 Different type of trademark similarity Trademark 1 Trademark 2 Similarity Aspect NEST Visual MAGIC TIMES MAGIC HOUR Conceptual SVIZZEROTALER SWISS TALER Phonetic Table 2 Pseudocode of the visual similarity computation employed in the proposed algorithm Pseudocode: /*comment*/ 1: /* This part of the code is performed for the visual similarity score computation for trademarks with text*/ 2: define Qt and Dt as the query and trademark from the database 3: compute Aq and Ad as new strings that produce optimal alignment between Qt and Dt 4: define score as the letter-to-letter visual similarity matrix between Qt and Dt; 5: define m=maximum(length(Aq), length(Ad)); 6: for i=0 until m 7: if Aq(i)=Null || Ad(i)==Null 8: score(i)=0; 9: else 10: score(i)=compute visual similarity score between Aq(i) and Ad(i) 11: end 12: define total_score= sum(score)/m); Table 3 Degree-of-similarity computed using approximate string matching and visual similarity Approximate String Matching Visual Similarity 1NDEX :: INDEX 0.80 0.923 1NDEX :: XNDEX 0.80 0.861 ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Table 4 Confusion matrix employed for the computation of the F-score, precision score, and accuracy score. Actual Class Predicted Class Positive Negative Positive TP FN Negative FP TN Table 5 F-score, precision, and accuracy computed using visual, conceptual, and phonetic similarity and the proposed method. Visual Similarity Conceptual Similarity Phonetic Similarity Proposed Method F-score 0.791 0.364 0.810 0.911 Precision 0.683 0.224 0.682 0.924 Accuracy 0.819 0.610 0.840 0.910 ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Table 6 List of 25 queries and their corresponding results used in this experiment Queries Result 1 Result 2 Result 3 WEBIATOR WebFOCUS autoscout24 FRUIT TIGER LION FRUIT SMOOTH FRUIT RED BULL GSTAR XSTAR sakira SVIZZEROTALER SWISS TALER SEVIKAR SCHNEIDER NEST Nexans SKYLINE SKY ROOM PREVISA BONITA SWEETLAND HEIDI LAND AMORA AMORE AXARA ARTOR RIMOSTIL Rivotril REBOVIR REFODERM CYRA CYREL ara adria GLOBRIX Globix ZYLORIC GRILON Lifestyle Living Style LIFE TEX SNOW LIFE WOOD STONE MOONSTONE WILTON SwissTron NATURE ELLA NATURESSA MARQUELA ecopower ECOPOWER HARRY POTTER TRIX TREAC TREAKOL SANTHERA SANZEZA SALFIRA sunirse MUROLINO MURINO MONARI MATTERHORN MAGIC TIMES MAGIC HOUR Maritimer MATCH WORLD RED BULL FLYING BULL Feel'n LEARN SEE'N LEARN FEEL GOOD FIGUREHEAD bonvita BONAVITA Botoceutical FMH FNH FTG MR ACTIVIA ACTEVA ADWISTA ACCET ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Table 7 Similarity scores obtained from the hit assignments and the proposed trademark degree-of-similarity aggregation method. No QUERIES Human Interactive Task Rating Proposed Method Result 1 Result 2 Result 3 Result 1 Result 2 Result 3 1 webautor 3.40 2.35 1.00 4.98 2.99 1.90 2 FRUIT TIGER 3.45 2.05 1.20 3.94 2.17 1.75 3 GSTAR 4.05 2.45 1.00 4.23 2.82 1.86 4 SVIZZEROTALER 3.70 2.10 1.15 3.84 2.77 1.82 5 NEXT 4.00 2.80 1.10 4.29 2.86 1.79 6 SKYPINE 4.20 2.65 1.60 3.99 2.84 1.93 7 Prevista 4.70 3.20 1.35 4.17 2.68 1.96 8 SWEETLAND 3.70 2.10 1.20 3.94 2.85 2.00 9 AMORA 4.50 2.35 1.85 4.28 2.67 1.05 10 RIMOSTRIL 3.95 2.30 1.65 4.04 2.22 1.76 11 CYRA 3.75 2.25 1.45 3.94 2.68 1.83 12 GLOBRIX 4.75 1.60 1.40 4.14 2.14 1.84 13 Lifestyle 4.25 2.35 1.50 3.98 2.43 1.82 14 WOOD STONE 3.60 1.70 1.45 4.32 2.30 1.91 15 NUTELLA 3.65 2.20 1.40 3.74 2.96 2.00 16 ecopower 4.45 2.80 1.10 5.00 2.96 0.87 17 TWIX 4.00 1.70 1.20 3.98 2.48 1.94 18 SANTHERA 3.20 2.05 1.15 3.86 2.96 1.96 19 MUROLINO 4.50 3.35 1.65 3.97 2.59 1.85 20 MAGIC TIMES 3.70 2.15 1.50 3.78 2.82 1.88 21 RED BULL 3.90 3.00 1.75 3.85 3.33 1.98 22 Feel'n LEARN 4.00 2.55 1.30 3.95 3.28 1.85 23 bonvita 4.90 2.65 1.55 4.20 2.69 1.85 24 FMH 4.40 2.75 1.40 4.43 2.07 1.57 25 ACTIVIA 4.25 2.00 1.65 4.20 2.22 1.98 Average 4.04 2.38 1.38 4.12 2.67 1.80 Table 8 Spearman rank correlation and Pearson pairwise correlation Spearman Rank Correlation Pearson Pairwise Correlation 1.00 (p<0.05) 0.92 (p<0.0001) ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Fig. 1 Different types of trademarks (OHIM, 2014) Fig. 2 Conceptual model of the proposed method for multi-faceted assessment of trademarks. Trademark Conceptual Comparison Algorithm Inference Engine Module Semantic Technology Set Similarity Theory CBIR Technology Shape Features Conceptual Module Phoneme Analysis Phonological Features Phonetic Module Fuzzy Logic Trademark Similarity Aggregation Score Algorithm Trademark Phonetic Comparison Algorithm Visual Module Trademark Visual Comparison Algorithm FORTIS Word mark Figurative word mark Purely figurative mark Purely figurative mark with figurative word mark ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Phonetic SimilarityConceptual SimilarityVisual Similarity Fuzzification Inference Defuzzification Input Individual Assessments Ranked Output Fig. 3 Flow chart of the proposed method. I N D E X 1 N D E X .615 1 1 1 1 Shape-based Similarity Comparison Algorithm Similarity =4.615 / 5 = 0.923 Fig. 4 Illustrative example of the visual similarity score computation employed in the proposed method. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 very low low medium high very high 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fig. 5 Input membership functions used. PHONETIC VL CONCEPTUAL VL L M H VH V I S U A L VL VL VL L H VH L VL L L H VH M L L M H VH H H H H H VH VH VH VH VH VH VH PHONETIC L CONCEPTUAL VL L M H VH V I S U A L VL VL VL L H VH L VL L L H VH M L M M H VH H H H H H VH VH VH VH VH VH VH PHONETIC M CONCEPTUAL VL L M H VH V I S U A L VL L M M H VH L M M M H VH M M M M H VH H H H H H VH VH VH VH VH VH VH PHONETIC H CONCEPTUAL VL L M H VH V I S U A L VL H H H H VH L H H H H VH M H H H H VH H H H H H VH VH VH VH VH VH VH PHONETIC VH CONCEPTUAL VL L M H VH V I S U A L VL H H VH VH VH L H H VH VH VH M VH VH VH VH VH H VH VH VH VH VH VH VH VH VH VH VH (a) (b) (c) (d) (e) Fig. 6 Associative matrices used for rule derivation in the inference process. very low low medium high very high 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Fig. 7 Output membership functions utilized in the inference process. ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Fig. 8 Illustrative example of the proposed aggregation method for the trademark pair Skypine and SKYLINE. HIT Preview Trade Marks: Degree of Similarity Scoring Trademarks may seem similar because they look similar (i.e. have visual similarity), have similar meaning (conceptual similarity) or sound similar (phonetic similarity). This task examines the degree of similarity between different trademarks. Based on the above explanation, please rank the following trademarks on the scale of 1 to 5, 5 being the most similar to the query and 1 being the least similar to it. Query trade mark: 1. TRIX 2. TREAC 3. TREAKOL 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 Submit Fig. 9 An example task used in Experiment 2. Phonetic Similarity Module Visual Similarity Module Conceptual Similarity Module 0.62 0.42 0.81 87 88 ฀ ฀ 92 93 ฀ ฀ 112 113 ฀ ฀ 117 118 Implication Operation Method (min) :: SKYLINE 0.798 87. IF visual is medium AND conceptual is low AND phonetic is high THEN output is high 88. IF visual is medium AND conceptual is medium AND phonetic is high THEN output is high 92. IF visual is high AND conceptual is low AND phonetic is high THEN output is high 93. IF visual is high AND conceptual is medium AND phonetic is high THEN output is high 112. IF visual is medium AND conceptual is low AND phonetic is very high THEN output is very high 113. IF visual is medium AND conceptual is medium AND phonetic is very high THEN output is very high 117. IF visual is high AND conceptual is low AND phonetic is very high THEN output is very high 118. IF visual is high AND conceptual is medium AND phonetic is very high THEN output is very high Centroid Defuzzification Method In fe re n c e E n g in e M o d u le Aggregation Operation (max) ACCEPTED MANUSCRIPT A C C E P T E D M A N U S C R IP T Fig. 10 Similarity scores obtained in Experiment 2. 0.00 1.00 2.00 3.00 4.00 5.00 6.00 0.00 1.00 2.00 3.00 4.00 5.00 6.00 P ro p o s e d A lg o ri th m D e g re e o f S im il a ri ty S c o re HIT Similarity Score 
work_2vcf2a424rfptcuffnkojfx2hm	----	Sem-Fit: A semantic based expert system to provide recommendations in the tourism domain to provide recommendations Sem-Fit: A semantic based expert system in the tourism domain Ángel García-Crespo 1, José Luis López-Cuadrado 2, Ricardo Colomo-Palacios ⇑, Israel González-Carrasco 2, Belén Ruiz-Mezcua 3 Computer Science Department, Universidad Carlos III de Madrid, Av. Universidad 30, Leganés 28911, Madrid, Spain a r t i c l e i n f o a b s t r a c t Keywords: Semantic technologies Semantic labeling Fuzzy logic Recommender systems Hotels The hotel industry is one of the leading stakeholders in the tourism sector. In order to reduce the trav eler’s cost of seeking accommodations, enforce the return ratio efficiency of guest rooms and enhance total operating performance, evaluating and selecting a suitable hotel location has become one of the most critical issues for the hotel industry. In this scenario, recommender services are increasingly emerg ing which employ intelligent agents and artificial intelligence to ‘‘cut through’’ unlimited information and obtain personalized solutions. Taking this assumption into account, this paper presents Sem Fit, a seman tic hotel recommendation expert system, based on the consumer’s experience about recommendation provided by the system. Sem Fit uses the consumer’s experience point of view in order to apply fuzzy logic techniques to relating customer and hotel characteristics, represented by means of domain ontolo gies and affect grids. After receiving a recommendation, the customer provides a valuation about the rec ommendation generated by the system. Based on these valuations, the rules of the system are updated in 1. Introduction Tourism is one of the most powerful With roughly 11% of the world’s total emplo is often presented as the first global indust the first tourist continent (Longhi, 2007). Th nization (2006) predicts that by 2020, tou world will increase by over 200%. Because tion intensive business, there are opportun order to adjust the new recommendations to past user experiences. To test the validity of Sem Fit, the validation accomplished includes the interaction of the customer with the system and then the results are compared with the expert recommendation for each customer profile. Moreover, the values of preci sion and recall and F1 have been calculated, based on three points of view, to measure the degree of rel evance of the recommendations of the fuzzy system, showing that the system recommendations are on the same level as an expert in the domain. industries worldwide. yment or GDP, tourism ry, and Europe is by far e World Tourism Orga rist arrivals around the tourism is an informa merce (Cao & Schniederjans, 2006). Not in vain, to be able to com pete in an increasingly competitive and globalized market, companies need to be provided with new strategies that allow them to confront successfully the environment challenges (Acosta & Febles, 2010). Developments in search engines, carrying capacity and speed of networks, have influenced the number of travelers around the world that use technologies for planning and experi encing their travels (Buhalis & Law, 2008). Thus, according to ities to apply informa Kenteris, Gavalas, and Economou (2009), the convergence of IT tion technology (IT) to support tourism and tourists (Watson, Akselsen, Monod, & Pitt, 2004), and, pursuing these opportunities, and communications technologies and the rapid evolution of the Internet have been some of the most influential factors in tourism the tourism industry is leading eCommerce applications (Werthner & Ricci, 2004) and one of the fastest growing segments of eCom ⇑ Corresponding author. Tel.: +34 91 624 59 58; fax: +34 91 624 9129. E-mail addresses: angel.garcia@uc3m.es (Á. García-Crespo), joseluis.lopez. cuadrado@uc3m.es (J.L. López-Cuadrado), ricardo.colomo@uc3m.es (R. Colomo- Palacios), Israel.gonzalez@uc3m.es (I. González-Carrasco), belen.ruiz@uc3m.es (B. Ruiz-Mezcua). 1 Tel.: +34 91 624 94 17; fax: +34 91 624 9129. 2 Tel.: +34 91 624 91 17; fax: +34 91 624 9129. 3 Tel.: +34 91 624 99 68; fax: +34 91 624 9129. 1 that have changed travelers’ behavior. Indeed the Internet is cur rently the primary source of tourist destination information for travelers (Chiu, Yueh, Leung, & Hung, 2009). Buhalis (1998) pointed out that the use of the Internet in the tourism industry provides access to a large number of people, as well as offering the opportu nity to develop closer relationships with customers. ICTs have radically changed the efficiency and effectiveness of tourism organisations, the way that businesses are conducted in the marketplace, as well as how consumers interact with organisations (Buhalis, 2003). And in a highly competitive environment, mailto:angel.garcia@uc3m.es mailto:joseluis.lopez.cuadrado@uc3m.es mailto:joseluis.lopez.cuadrado@uc3m.es mailto:ricardo.colomo@uc3m.es mailto:Israel.gonzalez@uc3m.es mailto:belen.ruiz@uc3m.es referencia bibliográfica. Published in:Expert Systems with Applications, (15 September 2011), 38(10), 13310–13319 according to Luo, Feng, and Cai (2004), tourists who searched on the Internet tended to spend more at their destinations as com pared to those who consult other information sources. The hotel industry is one of the leading stakeholders in the tour ism sector. In order to reduce travelers’ cost of seeking accommo dations, enforce the return ratio efficiency of guest rooms and enhance total operating performance, evaluating and selecting a suitable hotel location has become one of the most critical issues for the hotel industry (Chou, Hsu, & Chen, 2008). Service quality in hotels can be regarded as a composite measure of various attri butes, including tangible attributes but also intangible/subjective attributes (safety, quietness, . . .) (Benítez, Martín, & Román, 2007). Consumers’ judgments towards a service depend basically on the strength of their beliefs or expectations about various fea tures or attributes associated with the service and the weight of attributes (Engel, Blackwell, & Miniard, 1995). And, since online travelers are enthusiastic about meeting other travelers who have similar attitudes, interests and way of life (Wang, Yu, & Fesenma ier, 2002), there is a strong connection between travelers’ judg ments and hotel recommendations. In this scenario, recommender services are increasingly emerging that employ intelligent agents and artificial intelligence to ‘‘cut through’’ unlim ited information and obtain personalized solutions (Ricci & Werth ner, 2006). Recommender systems are commonly defined as applications that e commerce sites exploit to suggest products and provide consumers with information to facilitate their deci sion making processes (Niininen, Buhalis, & March, 2007). How to deliver relevant information to both potential and existing cus tomers has become an important task for the hospitality industry (Xian, Kim, Hu, & Fesenmaier, 2007), which, in many cases, is based on artificial intelligence techniques (Schiaffino & Amandi, 2009). Destination recommendation systems are mostly fed with subjec tive information provided by the tourism industry itself (Goosen, Meeuwsen, Franke, & Kuyper, 2009), but most of the relevant infor mation for recommendation should come from customers. Indeed, according to Klein (1998), travel products in general are arguably ‘‘experience goods’’ in that full information on certain attributes cannot be known without direct experience. Taking this assumption into account, this paper presents Sem Fit, a semantic hotel recommendation expert system, based on the consumer’s experience about the recommendation provided by the system. The proposed expert system uses the consumer’s experience point of view in order to apply fuzzy logic techniques to relating customer and hotel characteristics. Hotel characteris tics are represented by means of a domain ontology. After receiving a recommendation, the customer provides a valuation about the recommendation generated by the system. Based on the customer’s valuations, the rules of the system are updated in order to adjust the new recommendations to the past user experiences. The paper consists of five sections and is structured as follows. Section 2 reviews the relevant literature. Section 3 discusses the main features of Sem Fit including the conceptual model, algorith mic and architecture. Section 4 describes the evaluation of the tool performed including a description of the sample, the method, re sults and discussion. Finally, the paper ends with a discussion of re search findings, limitations and concluding remarks. 2. Background It is a widely recognized fact that information and decision making have become the foundation for the world economy (Wang, 2008). Among many enterprise assets, knowledge is trea ted as a critical driving force for attaining enterprise performance goals because knowledge facilitates better business decision 2 making in a timely fashion (Han & Park, 2009). And due to the importance of tourism, recommender systems and decision sup port systems for tourism have been a field of study since the very beginnings of artificial intelligence. A recommender system can provide a set of solutions that best fit the user, depending on different factors concerning the user, the objective or the context where it is applied. Such systems can re duce search efforts (Liang, Lai, & Ku, 2006) and provide valuable information to assist consumers’ decision making process (Ricci, 2002) in order to solve the problem of information overload (Kuo, Chen, & Liang, 2009). Adomavicius and Tuzhilin (2005) pro vide a survey of recommender systems as well as describe various limitations of current recommendation methods, and discuss pos sible extensions that can improve recommendation capabilities and make recommender systems applicable to an even broader range of applications. Due to the importance of tourism, many efforts had been de voted to recommender systems for tourism (e.g. Castillo et al., 2008; Loh, Lorenzi, Saldana, & Licthnow, 2004; Ricci & Nguyen, 2007; Wallace, Maglogiannis, Karpouzis, Kormentzas, & Kollias, 2003), often based on artificial intelligence techniques, as was pre dicted in the early nineties by Crouch (1991): intelligent agents (e.g. Aciar, Serarols Tarres, Royo Vela, & De la Rosa i Esteva, 2007; Schiaffino & Amandi, 2009), fuzzy approaches (e.g. Lenar & Sobecki, 2007; Ngai & Wat, 2003), Bayesian networks (e.g. Huang & Bian, 2009; Jiang, Shang, & Liu, 2009), to cite just a few. Nor is the field of using semantics in tourism new. Fodor and Werthner (2004) presented Harmonise, a project that deals with business integration in tourism using ontologies for mediation. The SATINE project by Dogac et al. (2004) describes how to deploy semantically enriched travel Web services and how to exploit semantics through Web service registries. Niemann, Mochol, and Tolksdorf (2008) propose how to enhance hotel search with Semantic Web Technologies. Jakkilinki, Georgievski, and Sharda (2007) proposed an ontology based e Tourism Planner AuSTO that enables users to create an itinerary in one single application by this intelligent tool that builds on semantic web technologies. The LA_DMS project (Kanellopoulos, 2008) provides semantic based information for tourism destinations by combining the P2P para digm with semantic web technologies. In the specific field of using semantics to provide better infor mation for tourists there have been relevant and recent efforts in the literature. For example, García Crespo et al. (2009) proposed a semantically enriched recommendation platform for tourists on route, later expanded to Destination Management Organizations (García Crespo, Colomo Palacios, Gómez Berbís, Chamizo, & Rive ra, 2010a). Lee, Chang, and Wang (2009) used ontologies to provide recommendation in the context of the city of Tainan (Taiwan). Fi nally, and in the most similar contribution to the literature, Huang and Bian (2009) integrated Bayesian networks and semantics to provide personalized recommendations for tourist attractions over the Internet. According to Huang and Bian (2009), there are two challenges in developing a system for personalized recommendations for tourism. One is the integration of heterogeneous online travel information. The other is the semantic matching of tourist attractions with travelers’ preferences. Taking into account that all information will be given to the system by the users using this means, the challenge of Sem Fit will be the semantic match ing between attractions and preferences. To do so, Sem Fit will use the fuzzy logic paradigm in order to express the relationship between the hotel characteristics and the customer preferences. The Sem Fit’s fuzzy engine will recalculate this fuzzy relation ships based on the customer feeling about previous recommen dations, allowing the automatic adaptation of the recommender system to the customers’ preferences. Young people 10 20 30 40 Fig. 2. Membership function for the fuzzy set ‘‘young people’’. 3. Sem-Fit: fundamentals and internals Usually experts and customers express their knowledge and preferences in the form of imprecise terms such as ‘‘young’’ or ‘‘near’’. The fuzzy logic paradigm allows the representation of these imprecise terms in order to implement intelligent systems. Based on this paradigm, Sem Fit allows the representation of fuzzy vari ables in order to describe hotels and customers. Later on, Sem Fit uses these variables to perform recommendations. This section is structured as follows. First, an explanation on the fuzzy logic par adigm is given. Secondly, the process of capturing fuzzy knowledge in hotel recommendation is depicted. Later on, the recommenda tion process is defined and explained. Fourth, the customer feeling capturing procedure is portrayed. Finally, Sem Fit architecture and implementation is shown. 3.1. Fuzzy logic paradigm The fuzzy sets theory provides a framework for the representa tion of the uncertainty of many aspects of human knowledge (Zadeh, 1965). Classic sets theory establishes that the elements of the universe may belong to a set or may not belong to that set. Then, given the set of odd numbers: U f1; 2; 3; 4; 5; . . .g; we can affirm that ‘‘3’’ belongs to such a set. We can also affirm that the number ‘‘21’’ belongs to the set of numbers greater than 7, but number 3 does not belong to such a set. The membership function for a given set can be depicted as shown in Fig. 1. When the element belongs to the set, the function takes the value 1, and when the ele ment does not belong to the set, the function takes the value 0. For a given element, fuzzy sets theory proposes the use of inter mediate degrees of membership to a set. In this way, if we consider the set of young people we can consider that a person who is 15 years old belongs to such a set with a degree of 1 (belongs), a person who is 30 years old belongs in some degree (for example 0.7) and a person who is 80 years old does not belong to this set (degree of 0). We can represent this membership function as de picted in Fig. 2. Fuzzy sets provide a way for defining cases in which the mem bership to a given set is relative or the membership function is not defined at all, allowing the representation of imprecision or uncer tainty. Imprecision and uncertainty modeling by means of fuzzy sets allows solving problems which cannot be solved by means 1 0 1 2 3 4 5 6 7 8 9 10 11 … Fig. 1. Membership function for the set of numbers greater than 7. 3 of classic techniques. Some such domains are: classification prob lems, pattern recognition, signal processing, knowledge based sys tems or temporal reasoning. Using this fuzzy theory, the purpose of Sem Fit is to offer hotel recommendations based on expert criterion. This expert criterion will be represented by means of fuzzy sets based on the affect grid (Russell, Weiss, & Mendelsohn, 1989). However, the customer pref erences may be different from the expert point of view or may change as time goes by. For this reason it is necessary to update the fuzzy rules which represent the expert criterion using the cus tomers’ feedback. 3.2. Capturing the fuzzy knowledge about hotel recommendation Fig. 3 depicts the main elements in our proposal about captur ing knowledge in hotel recommendations: 1. The semantic descriptions of the hotels. 2. The fuzzy relationship between the characteristics of the cus tomers and the characteristics of the hotels. 3. The customer feeling about the recommendations. The first step of the recommendation process consists of repre sentation of knowledge about how the hotels are selected. This knowledge is expressed using fuzzy sets. In a first stage, the fuzzy sets are defined based on the expert knowledge. An expert defines by means of fuzzy sets the characteristics of the hotels and the characteristics of the customers. Some fuzzy sets defined are: lux ury hotel, relaxing trip or young people. After defining the fuzzy terms used by the expert to describe hotels and customers, the membership function is defined for each fuzzy set. This membership function allows for the transformation of the customer characteristics and the hotels characteristics into fuzzy values. Studies of emotion and affect in human beings have an estab lished history which originates in philosophy. As a result of this tradition, and using their own work as a basis (Russell, 1980), Russell et al. (1989) proposed a measure of affect which had a pro found impact on the field of social psychology. They termed the measure the affect grid, a scale designed as a quick means of assessing affect along the dimensions of pleasure displeasure and arousal sleepiness on a 1 9 scale. The affect grid may prove to be the instrument of choice when subjects are called on to make affective judgments in rapid succession or to make a large number of judgments, especially when those judgments are to be aggre gated (1989). Using this scale, researchers can collect emotion rat ings from stress and tension to calm, relaxation or serenity, and Fig. 3. System design. from ecstasy, excitement and joy to depression, melancholy, sad ness, and gloom. According to the studies of these authors, the affect grid is potentially suitable for any study that requires judgments about affect of either a descriptive or a subjective kind. Based on the cus tomer characteristics, the fuzzy system will estimate the custom ers affect grid for each hotel characteristic. Then, the fuzzy sets for the representation of the customer characteristics are related with the concepts of the hotel ontology by means of the affect grid. In this way, the results of the fuzzy recommender can be translated into concepts of the hotel ontology in order to obtain the recommendation. The measure of the affect grid will be translated to linguistic tags related to the hotel characteristics by mean of semantic annotations based on the rules defined by the expert options. Such tags are ade quate, very adequate, not adequate or never recommend. Each tag has a value related in order to rate each hotel based on the amount of positive or negative tags obtained from the fuzzy reasoning. Affect grid was previously employed by authors in previous works that combined semantic technologies with emotions (García Crespo, Colomo Palacios, Gómez Berbís, & García Sánchez, 2010b). 3.3. Recommendation process The main steps of the recommendation process are depicted in Fig. 4. The recommendation is primarily based on the customer char acteristics. The steps of the reasoning process are: 1. Obtaining the customer characteristics. The customer provides the information about his personal situation, preferences, etc. The information required is established based on the expert cri terion as mentioned in Section 3.1. 2. Fuzzification. The values provided by the customer are con verted into the fuzzy sets defined by the expert to describe the customer characteristics. The fuzzy value for each variable is obtained by means of the membership function related to each fuzzy set. 4 3. Once the customer profile has been fuzzified, the fuzzy rules are evaluated obtaining the set of fuzzy values for the hotel charac teristics, as well as its tag for defining the level of adequation based on the affect grid (Fig. 5). 4. The results obtained are defuzzified in order to obtain a set of concrete characteristics for the hotels. 5. With the defuzzified results, hotels with the obtained charac teristics are retrieved based on the hotels ontology and the annotations. 6. Next the fuzzy recommendation system will calculate the weight of each hotel based on the values of the suitability obtained from the decision matrix. The total weight can be configured in two modes, called ‘‘nor mal mode’’ and ‘‘sensible mode’’. The ‘‘normal’’ mode calculates the weight of the hotel as the summitry of all the suitabilities: Pproduct Xn i 1 Xm j 1 W ij; where Wij is the defuzzified value of the association between the customer characteristic i and the hotel characteristic j. The ‘‘sensible’’ mode calculates the weight of the hotel as the product of all the suitabilities. Pproduct Yn i 1 Ym j 1 W ij: The sensible mode allows greater differences between hotels, pro viding more differentiation between the recommendations. 7. The hotel which obtains the highest weight is the final recom mendation, called ‘‘Star recommendation’’. Also a set of alterna tive recommendations, called ‘‘suggested hotels’’ will be shown. These steps constitute the primary recommendation process. However, the expert criterion may not be adequate, because the customer tendencies can change. For this reason, the fuzzy rela tionships between the customer characteristics and the character istics of the hotels must be dynamically recalculated based on the customer feeling. Fig. 4. Recommendation process. Fig. 5. Affect grid for hotel characteristics. 3.4. Capturing the customer feeling As mentioned, the knowledge about the relationships between customer characteristics and hotel characteristics represented by means of the affect grid (Russell et al., 1989) has to be updated. There are two possibilities to update such values. On the one hand, the expert can update the values of the affect grid and, on the other hand, these values can be automatically updated by means of heu ristics. The first option is the classic approach in which an expert updates the knowledge after studying the customer tendencies. The second option allows the automatic adaption of the recom mender system based on the information acquired during the rec ommendation process. For this option it is necessary to take into account the level of agreement of the customer with the solution provided in order to make future recommendations. Such recom 5 mendations have an effect in two ways: on the one hand they al low the correction of the initial criterion of the expert because the negative valuation by the customers may imply the need to redefine the initial affect grid. On the other hand, the feedback of the customer may be stored in order to adjust the criterion of the expert to the customer preferences. For this reason, we propose the automatic reconfiguration of the fuzzy relationships using customer feedback. After receiving a rec ommendation the customer will rate his level of pleasure displea sure about the characteristics of the recommendation received. The objective of this question is to obtain the first impression of the customer about the recommendation. The fact that the rating of the customer is about his feeling about the recommendation in stead of his real experience is important. This is because the rec ommendation could be correct but the real experience of the customer could be negative, due to external factors (bad weather, personal problems, illness, etc.). The customer feelings are stored in the knowledge base. When the number of customer evaluations is greater than one hundred, the fuzzy system will evaluate the overall customer feelings based on fuzzy meta rules defined in order to determine when it will be necessary to update the affect grids. If the overall customer feeling is negative, then the affect grids related to the negative valued rec ommendations will be adjusted according to such meta rules. The updating process will be repeated every 100 customer valuations. 3.5. System architecture and implementation Fig. 6 depicts a three layer scheme that represents the recom mendation system architecture. The first layer corresponds to the user interface. There are two different elements in the user interface layer. On the one hand, the administrator GUI allows the administrator to define the fuzzy rules that represent knowledge about hotel recommen dation. The administrator GUI consists of a web based applica tion to allow the definition of the fuzzy sets as well as the membership function for each fuzzy set. This application also al lows the easy matching between customer characteristics and hotel characteristics by means of the affect grid. On the other hand, the GUI recommendation allows the customers to obtain intelligent recommendations about hotels. The GUI recommen dation consists of a web questionnaire in which the customer answers a set of questions proposed by the expert. Such ques tions determine the customer profile in order to determine the most suitable hotel. After the questions have been answered, the fuzzy engine evaluates the rules, and the most suitable hotel Fig. 6. System a 6 is shown to the customer. Optionally, a list of less suitable hotels can be displayed. After the recommendation, the customer pro vides his degree of pleasure displeasure about the recommenda tion received. This information will be used by the fuzzy engine for reconfiguring the fuzzy relationships between the customer characteristics and the hotel characteristics. The second layer represents the business logic. This layer con tains the fuzzy engine and the semantic engine. The fuzzy engine will evaluate the fuzzy rules defined by the expert in order to determine the most suitable hotel characteristics for a given cus tomer profile based on the affect grid. As mentioned, the fuzzy knowledge is defined by means of the GUI administrator. Once the hotel characteristics have been determined, the fuzzy engine retrieves the hotels with these characteristics by means of the semantic engine. The semantic engine, based on the Jena Frame work (Reynolds, 2006), will manage the hotels ontology and will provide the set of hotels with the characteristics determined by the fuzzy engine. Besides the recommendations, the fuzzy engine will recalculate the fuzzy affect grid based on the customer feed back. The fuzzy recalculator has been implemented as a daemon in the web server. Such a daemon monitors customer feedback, recalculates the fuzzy relationships between characteristics when the customer evaluation is significant and updates the affect grids according to such customer feedback. Finally, the persistence layer stores the knowledge about the hotel recommendation. As mentioned, on the one hand, the hotel ontology defines the relevant characteristics of each hotel. The concepts of the hotel ontology (Fig. 7) describe the category of each hotel based on stars, the room characteristics, the special equip ment of the hotel, etc. This ontology also describes special activi ties related to the hotel based on the season. All this information rchitecture. Fig. 7. Partial view of the hotels ontology. is used to describe the hotels available in the system. The hotels ontology has been defined using the Ontology Web Language (OWL) (Bechhofer et al., 2004). The OWL language has three vari ants: OWL Lite, OWL DL and OWL Full. OWL Lite provides a small set of features, while OWL DL is more expressive than OWL Lite providing decidability based on description logics. OWL Full allows full expressivity but decidability is not warranted. For this reason, we have used OWL DL for the ontology definition. The storage and ontology reasoning has been developed based on the Jena frame work. Besides the ontology specific storage, a database stores the information about the fuzzy sets and the fuzzy relationships with the ontology. On the other hand, the customer characteristics and their relationships with the hotel characteristics are repre sented by means of fuzzy sets and fuzzy relationships by means of affect grids and are stored in a database. Table 1 Comparison between system and expert recommendations. Precision Recall F1 Star recommendation vs. expert recommendation 0.58 0.58 0.58 Overall system suggestions vs. expert recommendation 0.19 0.96 0.32 4. Evaluation The subsequent section describes the empirical evaluation of the project. The final aim of this study is to work out if Sem Fit serves as a valid recommendation system in a controlled environment. 4.1. Research design Once the system has been developed and trained, it is necessary to prove the validity of our proposal; we must prove that the star recommendations provided by the system are accepted by the cus tomers. In case of disagreement with the star recommendation, it is of interest to know if the customer found a valid alternative in the list of suggested hotels. Additionally, given a set of customer profiles we have evaluated the similarity between the recommen dations provided by the fuzzy system and the recommendations provided by four experts in hotel recommendations. By means of this experiment we have evaluated the accuracy of the fuzzy sys tem comparing its recommendations to the recommendations of human experts. In case of disagreement between the expert crite ria and the fuzzy system’s star recommendation, it is of interest to know if the expert recommendation is included in the list of sug gested hotels. Finally, comparison between the expert recommen dation and the selection of the customer has been performed in order to validate the expert criterion. Precision, recall and F1 will be used in order to measure the de gree of relevance of the recommendations of the system. This tech nique has been employed based on two points of view: on the one hand, it is used to measure customer agreement with the recommendation received and, on the other hand, to measure the agreement between the system results and the expert recommen dations. As mentioned, this technique has been also employed to measure customer agreement with the expert’s recommendation. These measures had been used before to measure recommenda tions in semantic systems (e.g. García Crespo, Colomo Palacios, Gómez Berbís, & Ruiz Mezcua, 2010c, García Crespo, Rodríguez, Mencke, Gómez Berbís, & Colomo Palacios, 2010d, Paniagua Martín, García Crespo, Colomo Palacios, & Ruiz Mezcua, 2011). Precision, recall and F1 measures are defined as follows: 7 precision CorrectHotelsFound TotalHotelsFound ; recall CorrectHotelsFound TotalCorrectHotels ; F1 2 � precision � recall precision þ recall : The correct hotels will be determined by the expert criteria or the customer depending on the test. The following subsections describe the experimentation carried out and the results obtained. 4.2. Sample In order to determine the validity of the proposed approach, the designed evaluation was carried out. A set of 50 students (37 male and 13 female) in their final year of the Computer Science degree program at Universidad Carlos III de Madrid accessed the system in order to obtain a recommendation for their spring break. The system was trained with information on 10 hotels in Mallorca (Spain) based on the knowledge of an expert from a travel agency. With the results obtained from the system, each student selected the start preference or, in case of disagreement, one of the alterna tive recommendations provided or an empty alternative. Four experts in the travel domain recommend a hotel for each student in order to compare the recommendation of the system with the expert criteria. Recommendation generated by the fuzzy system was also compared with the expert recommendations in order to measure the validity of the star and alternative recommendations. 4.3. Results The experimentation was carried out in two separated tests that will be analyzed in the next subsections. 4.3.1. Test 1. Comparison between the system results and the expert recommendation Table 1 summarizes the precision and recall and F1 values ob tained by means of the comparison between the expert recommen dations and the results provided by the fuzzy system. This comparison has been divided into two stages. In the first stage, we have compared the star recommendation of the system with the expert recommendation. The star recommendation consists of a single recommendation (the top rated hotel for the customer characteristics) provided by the fuzzy system for each customer, and the expert recommendation is also a single recommendation suggested by the expert for each customer. The precision of the fuzzy system is 0.58. The value of recall is the same as precision be cause the number of returned categories is the same as the number of valid categories. We consider this value a promising result be cause more than 50% of the recommendations of the system are the same as the expert, and there is only one possible result. On the other hand, the fuzzy system provides a set of four alter natives for the star recommendation. We have compared the over all results of the system (five results, star recommendation and four suggested hotels) with the suggestion of the expert. In this scenario we have obtained a precision of 0.192, less than the 0.58 obtained for the star recommendation. It is logical conse quence because the number of returned values is greater (5 for each customer vs. only one star recommendation) and the number of correct values (recommendations of the expert) is the same. However the recall value achieves the value of 0.96. It means that in 96% of the cases the fuzzy system includes the recommendation of the expert in the set of results returned (either as star or sug gested recommendation). It is a good result because the system is capable of providing an expert recommendation in the set of suggestions in almost all of the cases. 4.3.2. Test 2. Comparison between the customer selection and the suggestions of both system and expert Test 1 compared the results of the system with the suggestions of the expert. But, are the expert suggestions a good basis? In order to answer this question, Test 2 compares the recommendations of the expert with the selection of the customer. Table 2 shows the precision, recall and F1 values obtained in this test. The precision value is 0.76, and the recall value is the same. It means that the 76% of expert recommendations are accepted by the customer. The second stage of the Test 2 consisted of the comparison be tween the system result and the final customer selection. In this stage we consider separately, on the one hand, only the star recom mendation and, on the other hand, the star recommendation plus the alternative suggestions. Precision and recall and F1 values for test 2 are presented in Table 3. The first row shows the comparison between the star recommendation and the customer selection. In this case, the precision value is 0.48. It means that the star recommendation of the system is the customer selection in the 48% of the cases. The value of recall is the same because the valid values are the same as the amount of values considered. The second row of Table 3 presents the comparison between the overall results of the system (star recommendation plus 4 sugges tions) and the final customer selection. As in Test 1, the precision value is smaller because the number of suggestions is greater, but the recall value is 0.96. It means that in 96% of the cases, the system offers a star recommendation or an alternative suggestion that satisfies the customer. 4.4. Discussion After the interaction of the customer with the system, the re sults were compared with the expert recommendation for each customer profile. We found that 24 customers found the star rec Table 2 Comparison between expert recommendation and customer selection. Precision Recall F1 Expert recommendation vs. customer selection 0.76 0.76 0.76 Table 3 Comparison between system recommendation and customer selection. Precision Recall F1 Star recommendation vs. customer selection 0.48 0.48 0.48 Overall system suggestions vs. customer selection 0.19 0.96 0.32 8 ommendation provided by the fuzzy system suitable and 48 cus tomers found a recommendation included in the star and suggested hotels suitable. Only two customers did not find a suit able recommendation. On the other hand, we found that 29 star recommendations coincide with the expert recommendation, and 48 expert recommendations were included in the sum of the star recommendations with the suggested recommendations. We can conclude that the sum of suggested and star recommendations provides an accurate set of recommendations in which the cus tomer can find a hotel. The system recommendations fit with the expert recommendation taking into account star and suggested ho tels. Finally the expert recommendation coincides with the cus tomer selection in 38 cases. However, the sum of star and suggested hotels of the fuzzy systems obtain better results than the expert recommendation. As shown in the previous subsection, the values of precision and recall and F1 have been calculated based on three points of view. Test 1 measured the similarity between the system recommenda tions and the expert recommendations. If we only take into ac count the star recommendation, both precision and recall values for the system recommendation in this scenario are 0.58. It means that 58% of the system star recommendations coincide with the ex pert recommendation. As mentioned, it is an acceptable margin be cause there is only one valid value (the expert recommendation which is highly subjective) and the system only offers one star rec ommendation. However, if we include in the study the four alter native suggestions of the system we can see that the precision decreases because the number of categories found is greater and the correct value is only one, but the recall value is 0.96. It is an excellent result because in 96% of the cases the system offers the expert recommendation. It means that the system recommenda tions are on the same level as an expert in the domain and the cus tomer can express his feelings in the same way that he does with the expert. Test 2 studies, on the one hand, the results of the expert vs. cus tomer selection. In this case, the precision and recall values ob tained are 0.76. We can see that the expert obtains better results with only one recommendation. This is natural, because the expert can take into account more parameters such as the corporal expression of the customer, non verbal behavior) in order to pro vide a recommendation. It is a good value, because the expert pro vides 3 good recommendations out of 4 recommendations. It also means that the knowledge represented in the fuzzy system is accu rate. On the other hand, Test 2 compared the star recommendation of the system with the customer selection. In this case, the preci sion value was 0.48. It means that the star recommendation is ac cepted by the customer in 48% of the cases. However, when we add the additional suggestions of the system, the precision is smaller because the number of suggestions is greater and the number of correct values is only one (the customer selection), but the recall is 0.96. It is an excellent result because it means that the customer finds a suitable recommendation in 96% of the cases. It is a high rate of positive recommendations considering that the selection of a hotel is a highly subjective decision. Cao and Li (2007) obtain an average recall of 83.82% for recommendations of 15 laptops in all 138 laptops for seven different customers; however in Cao and Li’s study the customer can select more than one recommen dation. Zenebe and Norcio (2009) present a comparative study of several techniques for recommending systems obtaining a recall of 38% for the fuzzy set theoretical method in a highly subjective domain such as movies selection. Zenebe and Norcio’s experiment includes a large number of customers and movies, and it is natural that the recall value was smaller. In our case of study the number of customers and products are smaller. Based on the obtained results, the proposed system is able to estimate customers’ feelings based on their profile and is capable of offering a set of recommendations that will be accepted most of the time. Future research may include a larger number of hotels and customers in order to measure the scalability of the proposed system. 5. Conclusions and future work We have presented Sem Fit, a semantic hotel recommendation expert system, based on consumers’ experience about the recom mendation provided by the system. The proposed expert system uses the consumers’ experience point of view in order to apply fuz zy logic techniques to relate customers and hotels characteristics, represented by means of domain ontologies and affect grids. After receiving a recommendation, the customer provides a valuation about the recommendation generated by the system. Based on the customers’ valuations, the rules of the system are updated in order to adjust the new recommendations to the past user experi ences. The validation accomplished shows that the sum of star and suggested hotels of the fuzzy systems obtains better results than the expert recommendation. Moreover, the values of precision and recall and F1 reveal that the Sem Fit recommendations are on the same level as an expert in the domain and the customer will be able to express his feelings in the same way that he does with the expert. As an extension of this paper, the knowledge base could be im proved including more products, such as suitable destinies or asso ciated services. Furthermore, other ways of collecting data on customer feeling about a recommendation with greater accuracy and clarity could be studied. Acknowledgements This work is supported by the Spanish Ministry of Industry, Tourism, and Commerce under the EUREKA project SITIO (TSI 020400 2009 148), SONAR2 (TSI 020100 2008 665) and GO2 (TSI 020400 2009 127). References Aciar, S. V., Serarols-Tarres, C., Royo-Vela, M., & De la Rosa i Esteva, J. L. (2007). Increasing effectiveness in e-commerce: Recommendations applying intelligent agents. International Journal of Business and Systems Research, 1(1), 81–97. Acosta, Z., & Febles, J. (2010). The organizacional management as instrument to overcome the resistance to the innovative process: An application in the canary company. International Journal of Human Capital and Information Technology Professionals, 1(2), 49–64. Adomavicius, G., & Tuzhilin, A. (2005). Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17(6), 734–749. Bechhofer, S., van Harmelen, F., Hendler, J., Horrocks, I., McGuinness, D. L., Patel- Schneider, P. F., et al. (2004). OWLWeb ontology language reference. http:// www.w3.org/TR/owl-ref/. Benítez, J. M., Martín, J. C., & Román, C. (2007). Using fuzzy number for measuring quality of service in the hotel industry. Tourism Management, 28(2), 544–555. Buhalis, D. (1998). Strategic use of information technologies in the tourism industry. Tourism Management, 19(5), 409–421. Buhalis, D. (2003). eTourism: Information technology for strategic tourism management. London: Pearson. Buhalis, D., & Law, R. (2008). Progress in information technology and tourism management: 20 years on and 10 years after the Internet—The state of eTourism research. Tourism Management, 29(4), 609–623. Cao, Q., & Schniederjans, M. J. (2006). Agent-mediated architecture for reputation- based electronic tourism systems: A neural network approach. Information & Management, 43(5), 598–606. Cao, Y., & Li, Y. (2007). An intelligent fuzzy-based recommendation system for consumer electronic products. Expert Systems with Applications, 33(1), 230–240. Castillo, L., Armengol, E., Onaindía, E., Sebastiá, L., González-Boticario, J., Rodríguez, A., et al. (2008). SAMAP: An user-oriented adaptive system for planning tourist visits. Expert Systems with Applications, 34(2), 1318–1332. Chiu, D. K. W., Yueh, Y. T. F., Leung, H. F., & Hung, P. C. K. (2009). Towards ubiquitous tourist service coordination and process integration: A collaborative travel agent system architecture with semantic web services. Information Systems Frontiers, 11(3), 241–256. 9 Chou, T. Y., Hsu, C. L., & Chen, M. C. (2008). A fuzzy multi-decision model for international tourist hotels location selection. International Journal of Hospitality Management, 27(2), 293–301. Crouch, G. I. (1991). Expert computer systems in tourism: Emerging possibilities. Journal of Travel Research, 29(3), 3–10. Dogac, A., Kabak, Y., Laleci, G., Sinir, S., Yildiz, A., Kirbas, S., et al. (2004). Semantically enriched Web services for the travel industry. ACM Sigmod Record, 33(3), 21–27. Engel, J. F., Blackwell, R. D., & Miniard, P. W. (1995). Consumer behavior. Forth Worth, TX: The Dryden Press. Fodor, O., & Werthner, H. (2004). Harmonise: A step toward an interoperable e- tourism marketplace. International Journal of Electronic Commerce, 9(2), 11–39. García-Crespo, A., Chamizo, J., Rivera, I., Mencke, M., Colomo-Palacios, R., & Gómez- Berbís, J. M. (2009). SPETA: Social pervasive e-Tourism advisor. Telematics and Informatics, 26(3), 306–315. García-Crespo, A., Colomo-Palacios, R., Gómez-Berbís, J. M., Chamizo, J., & Rivera, I. (2010a). Intelligent decision-support systems for e-tourism: Using SPETA II as a knowledge management platform for DMOs and e-tourism service providers. International Journal of Decision Support System Technology, 2(1), 35–47. García-Crespo, A., Colomo-Palacios, R., Gómez-Berbís, J. M., & García-Sánchez, F. (2010b). SOLAR: Social link advanced recommendation system. Future Generation Computer Systems, 26(3), 374–380. García-Crespo, A., Colomo-Palacios, R., Gómez-Berbís, J. M., & Ruiz-Mezcua, B. (2010c). SEMO: A framework for customer social networks analysis based on semantics. Journal of Information Technology, 25(2), 178–188. García-Crespo, A., Rodríguez, A., Mencke, M., Gómez-Berbís, J. M., & Colomo- Palacios, R. (2010d). ODDIN: Ontology-driven differential diagnosis based on logical inference and probabilistic refinements. Expert Systems with Applications, 37(3), 2621–2628. Goosen, M., Meeuwsen, H., Franke, J., & Kuyper, M. (2009). My Ideal tourism destination: Personalized destination recommendation system combining individual preferences and GIS data. Information Technology & Tourism, 11(1), 17–30. Han, K. H., & Park, J. W. (2009). Process-centered knowledge model and enterprise ontology for the development of knowledge management system. Expert Systems with Applications, 36(4), 7441–7447. Huang, Y., & Bian, L. (2009). A Bayesian network and analytic hierarchy process based personalized recommendations for tourist attractions over the Internet. Expert Systems with Applications, 36(1), 933–943. Jakkilinki, R., Georgievski, M., & Sharda, N. (2007). Connecting destinations with an ontology-based e-tourism planner. In M. Sigala, L. Mich, & J. Murphy (Eds.), Information and communication technologies in tourism (pp. 21–32). Berlin, Germany: Springer Wien. Jiang, Y., Shang, J., & Liu, Y. (2009). Maximizing customer satisfaction through an online recommendation system: A novel associative classification model. Decision Support Systems, 48(3), 470–479. Kanellopoulos, D. N. (2008). An ontology-based system for intelligent matching of travellers’ needs for group package tours. International Journal of Digital Culture and Electronic Tourism, 1(1), 76–99. Kenteris, M., Gavalas, D., & Economou, D. (2009). An innovative mobile electronic tourist guide application. Personal and Ubiquitous Computing, 13(2), 103–118. Klein, L. R. (1998). Evaluating the potential of interactive media through a new lens: search versus experience goods. Journal of Business Research, 41(3), 195–203. Kuo, M. H., Chen, L. C., & Liang, C. W. (2009). Building and evaluating a location- based service recommendation system with a preference adjustment mechanism. Expert Systems with Applications, 36(2 Part 2), 3543–3554. Lee, C. S., Chang, Y. C., & Wang, M. H. (2009). Ontological recommendation multi- agent for Tainan City travel. Expert Systems with Applications, 36(1), 6740–6753. Lenar, M., & Sobecki, J. (2007). Using recommendation to improve negotiations in agent-based systems. Journal of Universal Computer Science, 13(2), 267–286. Liang, T. P., Lai, H. J., & Ku, Y. C. (2006). Personalized content recommendation and user satisfaction: Theoretical synthesis and empirical findings. Journal of Management Information Systems, 23(3), 45–70. Loh, S., Lorenzi, F., Saldana, R., & Licthnow, D. (2004). A tourism recommender system based on collaboration and text analysis. Information Technology & Tourism, 6(3), 157–165. Longhi, C. (2007). Usages of the Internet and e-tourism. Towards a new economy of tourism. In Proceedings of the second international conference advances in tourism economics, Portugal. Luo, M., Feng, R., & Cai, L. A. (2004). Information search behavior and tourist characteristics: The internet vis-à-vis other information sources. Journal of Travel & Tourism Marketing, 17(2-3), 15–25. Ngai, E. W. T., & Wat, F. K. T. (2003). Design and development of a fuzzy expert system for hotel selection. Omega, 31(4), 275–286. Niemann, M., Mochol, M., & Tolksdorf, R. (2008). Enhancing hotel search with semantic Web technologies. Journal of Theoretical and Applied Electronic Commerce Research, 3(2), 82–96. Niininen, O., Buhalis, D., & March, R. (2007). Customer empowerment in tourism through consumer centric marketing (CCM). Qualitative Market Research: An International Journal, 10(3), 265–281. Paniagua-Martín, F., García-Crespo, Á., Colomo-Palacios, R., & Ruiz-Mezcua, B. (2011). SSAAMAR: Semantic annotation architecture for accessible multimedia resources. IEEE Multimedia, 18(2), 16–25. Reynolds, D. (2006). Jena rules. In Proceedings of the 2006 Jena user conference. Ricci, F. (2002). Travel recommender systems. IEEE Intelligent Systems, 17(6), 55–57. http://www.w3.org/TR/owl-ref/ http://www.w3.org/TR/owl-ref/ Ricci, F., & Nguyen, Q. N. (2007). Acquiring and revising preferences in a critique- based mobile recommender system. IEEE Intelligent Systems, 22(3), 22–29. Ricci, F., & Werthner, H. (2006). Recommender systems. International Journal of Electronic Commerce, 11(2), 5–9. Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology, 39, 1161–1178. Russell, J. A., Weiss, A., & Mendelsohn, G. A. (1989). The affect grid: A single-item scale of pleasure and arousal. Personality and Social Psychology, 57(3), 493–502. Schiaffino, S., & Amandi, A. (2009). Building an expert travel agent as a software agent. Expert Systems with Applications, 36(2), 1291–1299. Wallace, M., Maglogiannis, I., Karpouzis, K., Kormentzas, G., & Kollias, S. (2003). Intelligent one-stop-shop travel recommendations using an adaptive neural network and clustering of history. Information Technology & Tourism, 6, 181–193. Wang, J. (2008). Improving decision-making practices through information filtering. International Journal of Information and Decision Sciences, 1(1), 1–4. 10 Wang, Y., Yu, Q., & Fesenmaier, R. D. (2002). Defining the virtual tourist community: Implications for tourism marketing. Tourism Management, 23(4), 407–417. Watson, R. T., Akselsen, S., Monod, E., & Pitt, L. F. (2004). The open tourism consortium: laying the foundations for the future of tourism. European Management Journal, 22(3), 315–326. Werthner, H., & Ricci, F. (2004). E-Commerce and tourism. Communications of the ACM, 47(12), 101–105. World Tourism Organization (2006). Tourism market trends 2005: World overview & tourism topics. Madrid, Spain: World Tourism Organization. Xian, Z., Kim, S. E., Hu, C., & Fesenmaier, D. R. (2007). Language representation of restaurants: Implications for developing online recommender systems. Hospitality Management, 26(4), 1005–1018. Zadeh, L. (1965). Fuzzy sets. Information & Control, 8(3), 338–353. Zenebe, A., & Norcio, A. F. (2009). Representation, similarity measures and aggregation methods using fuzzy sets for content-based recommender systems. Fuzzy Sets and Systems, 160(1), 76–94. 
work_2wc5hvqsirhb3f4jngklcwbfky	----	International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-2, December, 2019 4622 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: B5116129219/2019©BEIESP DOI: 10.35940/ijeat.B5116.129219 ABSTRACT---Lung Cancer is the second most recurrent cancer in both men and women and which is the leading cause of cancer death worldwide. The American cancer Society (ACS) in US estimates nearly 228,150 new cases of lung cancer and 142,670 deaths from lung cancer for the year 2019. This paper proposes to build an ontology based expert system to diagnose Lung Cancer Disease and to identify the stage of Lung Cancer. Ontology is defined as a specification of conceptualization and describes knowledge about any domain in the form of concepts and relationships among them. It is a framework for representing shareable and reusable knowledge across a domain. The advantage of using ontology for knowledge representation of a particular domain is they are machine readable. We designed a System named OBESLC (Ontology Based Expert System for Lung Cancer) for lung cancer diagnosis, in that to construct an ontology we make use of Ontology Web Language (OWL) and Resource Description Framework (RDF) .The design of this system depends on knowledge about patient’s symptoms and the state of lung nodules to build knowledge base of Lung Cancer Disease. We verified our ontology OBESLC by querying it using SPARQL query language, a popular query language for extracting required information from Semantic web. We validate our ontology by developing reasoning rules using semantic Web Rule Language (SWRL).To provide the user interface, we implemented our approach in java using Jena API and Eclipse Editor. Keywords: Semantic Web, Ontology, Lung Cancer, RDF, OWL, SWRL, SPARQL. I. INTRODUCTION Lung cancer is the most common type of cancer and constitutes 24% of all cancer related deaths. About 13% of all new cancers are lung cancers and in US One in 16 people will be diagnosed with lung Cancer. Lung cancer occurs due to the uncontrolled growth of abnormal cells present in lungs which causes growth of tumors or lesions that reduces breathing ability of a person. The important identified reasons behind lung cancer disease are Smoking, Exposure to asbestos, Radon , some hazardous chemicals , exposure to continuous air pollution and Genetic factors etc. Many people with lung cancer were detected in advanced stages but not at early stage so the mortality rate is more in lung cancer patients. It is recommended that people who have family histology of Lung cancer and smokers must undergo tests like LCDT etc., periodically to identify lesions or nodules in the lungs and must consult physicians or general practitioners for diagnosing the disease. [8,9] Revised Manuscript Received on 2 December, 2019. J.Sirisha , Research Scholar,Krishna University, Machilipatnam,Andhra Pradesh, India.( Email: siri.jagannadham@gmail.com) Dr. M. Babu Reddy, Head(i/c), Department of Computer Science, Krishna University, Machilipatnam, Andhra Pradesh, India (email m_babureddy@yahoo.com_ In general an expert system is an intelligent system that supports decision making capabilities and solves complex problems through reasoning. Expert systems are Artificial intelligence based computational tools that consist of components like knowledge base and inference engine. Knowledge base consists of facts and rules related to a particular field or domain and Inference engine deduces new facts. Similarly in diagnosing any disease, A Medical expert system is a computer program consisting of knowledge regarding medical domain that provides accurate information about disease diagnosis there by it deduces prognosis and treatment plans etc. In our work we have designed an expert system which can consider the symptoms of patients and nodule size to detect the lung cancer patients in early stage thereby we can improve the perpetuity of a lung cancer patient.[1,2,7] Ontology is a main component of semantic web that refers to the science of describing different kinds of entities and relations among entities in a particular domain. The semantic web Technologies like RDF(Resourse Description Format),OWL(Ontology Web language) enables the machine to understand the knowledge stored in an ontology and do the complex work involved in searching , sharing and merging the information on the web. Building an Ontology based Expert system in medical domain is very much needed in present days because medical knowledge is increasingly more composite and uncontrollable.[13]. In this paper we are concentrating on expert system in medical domain and use of ontology as a computational aid by applying semantic aspects for generating rules to diagnose and identify the stage of disease. Here we developed a knowledge base related to lung cancer i.e., lung cancer ontology so that the people will know the details regarding this disease as their initial medical assistance .In general for detecting any type of cancer the physicians or medical experts depends on image analysis of lung nodules where images can be analyzed by radiologists using the scan reports and size of the nodule or nodules in the image which can be obtained through various types of scans like Computer Tomography(CT) and Magnetic Resonance Imaging(MRI).In our work, by taking the size of the nodule from image reports and some important symptoms into consideration our lung cancer ontology expert system (OBESLC) detects lung cancer patients and also identify the stage of patient using staging system by formulizing rules using SWRL language. An Ontology Based Expert System for Lung Cancer : OBESLC J.Sirisha, M. Babu Reddy An Ontology Based Expert System for Lung Cancer : OBESLC 4623 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: B5116129219/2019©BEIESP DOI: 10.35940/ijeat.B5116.129219 The following diagram depicts the concepts ,data properties and individuals created for OBESLC system using protégé tool Fig III.b Data Properties in OBESLC Fig III.c Concepts in OBESLC Fig III.d Individuals created for the class LungCancer Phase 3: Rule generation using Semantic Web Rule Language(SWRL) The process of generating rules is very much important for reasoning the developed ontology using “Semantic Web Rule language” (SWRL). In our lung cancer ontology reasoning makes the process of diagnosis and detecting the Stage of patient very effectively and efficiently. Here we developed many rules for identifying whether the patient has Lung Cancer or not, If so detecting the stage of Lung Cancer etc. To generate rules „SWRL tab‟ must be incorporated through SWRL plug-in in Protégé Tool. [26] In the OBESLC system First Rule is used to diagnose whether the patient has Lung Cancer or not by considering the symptoms of patient and nodule size in the lung. Here as the disease was identified mainly based on symptoms in the initial diagnosis process we should consider symptoms and the reasons for those symptoms also. So we have divided the Rule1 into four parts as Rule1a,Rule1b, Rule1c, Rule1d where Rule1 will be completed with a combination of Rule1d and one among Rule1a,Rule1b ,Rule1c.The second rule is used to identify the stage of lung cancer according to TNM system, The third rule categorizes the lesion as benign or malignant which is combination of rules to identify the lesion category according to its diameter. The Fourth rule is also combination of several rules allows us to suggest the treatment according the stage of lung cancer patient. Examples of some rules generated for OBESLC System: Rule1 - > Lung Cancer Detection Symptom checking to identify Lung cancer patient Rule1a : hasWeightLoss(?x, true) ^ hasSymOfCPorPCorCWB(?x, true) ^ hashabitOfEitherSorA(?x, true) ^ Patient(?x) -> probabilityOfHaving(?x, LungDesease) Rule1b : HasFamilyHistoryofCancer(?x, true) ^ hasWeightLoss(?x, true) ^ hasSymOfCPorPCorCWB(?x, true) ^ Patient(?x) -> probabilityOfHaving(?x, LungDesease) Rule 1c: hasWeightLoss(?x, true) ^ hasExposureOfAorRorCAP(?x, true) ^ hasSymOfCPorPCorCWB(?x, true) ^ Patient(?x) -> probabilityOfHaving(?x, LungDesease) To Diagnose Lung Cancer Patient Rule1d: Patient(?p) ^ probabilityOfHaving(?p, LungDesease) ^ Nodule(?n) ^ hasNoduleSize(?n, ?s) ^ swrlb:greaterThan(?s, 2) -> hasDesease(?p, LC) Rule2 -> To Identify Stage of Lung cancer hasSpreadBV(?x, false) ^ hasT(?p, "T1") ^ hasN(?p, "N0") ^ hasSpreadOP(?p, false) ^ hasSpreadLN(?p, false) ^ hasM(?p, "M0") ^ LungCancer(?p) -> classifiedInto(?p, StageIA) International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-2, December, 2019 4624 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: B5116129219/2019©BEIESP DOI: 10.35940/ijeat.B5116.129219 Rule3 -> To categorize the lesion as benignant or malignant Patient(?p)^Nodule(?x)^foundIn(?p,?x)^hasDiameter(?x,? s)^swrlb:greaterThan(?s,8) ^swrlb:lessThan(?s,15)->hasCategory(?x,LRc4A) Rule 4 -> Required Treatment according to the Lung Cancer Stage LungCancer(?x) ^ classifiedInto(?x, StageIA) -> treatedBy(?x, Surgery) III. IMPLEMENTATION& RESULTS To implement our proposed system we use different types of programming languages and computing software like Protégé Editor is used construct the medical ontology of OBESLC system and to update RDF (Resourse Description Framework) files. we make use of Jena API which is a programming toolkit that depends on java programming language to interact with our OBESLC system. Eclipse IDE was used to develop graphical demonstrators for the purpose of user interaction . SWRL language is used for reasoning the ontology by generating rules and to query the developed ontology we make use of SPARQL. This type of implemented system is needed by clinicians, General practioners and medical students so that they can have initial knowledge base of lung cancer disease and can easily diagnose the Lung cancer disease ,Stage identification and suggest Treatment plans to the patients as a part of their initial screening process. Sample rule execution of OBESLC system in Protégé Tool Executing a rule in protégé involves the steps like i) Exporting OWL axioms into rule engine ii) Execution of rule using rule engine iii)Translate the inferred axioms into OWL model. If we perform all these actions successfully we can able to impose our rules into the ontology and draw inferences according to our proposed rules Fig IV.a Rule1d execution in protégé using rule engine and transferring it into OWL Knowledge Fig IV.b Rule2 execution in protégé using rule engine and transferring it into OWL Knowledge Fig IV.c Rule4 execution in protégé using rule engine and transferring it into OWL Knowledge Validating OBESLC system by Sample SWRL Rule extraction Using developed demonstrators in Jena with Eclipse editor: After executing and trasfering any rule into OWL Model, we can extract the same using user interface demonstrators by providing communication between Owl file generated by Protégé Tool and Jena API with Eclipse editor. Here we have shown results of SWRL rules in the developed interfaces for Lung cancer Treatment (d,e,f) ,Lung cancer detection(g) and Lung cancer stage identification(h) respectively. (d) (e) An Ontology Based Expert System for Lung Cancer : OBESLC 4625 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: B5116129219/2019©BEIESP DOI: 10.35940/ijeat.B5116.129219 (f) (g) (h) Fig IV (d) Semantic representation of concepts before Rule execution e) Semantic representation of concepts after Rule execution f)The Result of the rule for treatment of lung cancer in developed interface (g) The Result of the rule for detecting lung cancer patient in developed interface (h) The Result of the rule for lung cancer stage identification in developed interface. Verifying the OBESLC System using SPARQL : To validate the developed ontology we make use of SPARQL query language which interrogate the ontology and produces the results . SPARQL stands for SPARQL Protocol And RDF Query Language.Querying and extracting information from the knowledge base is an important task in semantic web through which users and applications can interact with data in the ontologies. Here we have presented some sample SPARQL queries through which we can test our rules of OBESLC system by communicating with the developed user interface demonstrators. Query for Treatment : PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax- ns#> PREFIX owl: <http://www.w3.org/2002/07/owl#> PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> PREFIX xsd: <http://www.w3.org/2001/XMLSchema#> PREFIX lc: <http://www.semanticweb.org/venulanka/ontologies/2019/7/ lcntology#> SELECT ?Treatment ?CancerStage WHERE { ?CancerStage lc:classifiedInto ?stage. ?CancerStage lc:treatedBy ?Treatment } Query for identifying Stage of Lung Cancer: PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax- ns#> PREFIX owl: <http://www.w3.org/2002/07/owl#> PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> PREFIX xsd: <http://www.w3.org/2001/XMLSchema#> PREFIX lc: <http://www.semanticweb.org/venulanka/ontologies/2019 /7/lcntology#> SELECT ?LCancer ?Stage WHERE {?LCancer lc:hasT ?t. ?LCancer lc:hasN ?n. ?LCancer lc:hasM ?m. ?LCancer lc:hasSpreadBV ?x. ?LCancer lc:hasSpreadLN ?y. ?LCancer lc:hasSpreadOP ?z. ?LCancer lc:classifiedInto ?Stage } V. CONCLUSION Developing an expert system using ontology in the field of medical domain produces better results with less complexity. In this paper we have developed OBESLC System (Ontology Based Expert System for Lung Cancer) using state of art semantic web technologies to diagnose, to identify lung cancer stage and to provide treatment plan according the stage of patient. We have successfully implemented(includes verification and validation) the OBESLC system using SPARQL queries ,SWRL rules and used the Apache Jena API along with Eclipse Editor to extract details from ontology and presented the query results to the user with the help of graphical demonstrators. International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-2, December, 2019 4626 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: B5116129219/2019©BEIESP DOI: 10.35940/ijeat.B5116.129219 This system is very much help to the General Practioners and clinicians to assess the disease according to the patient symptoms along with scan(CT,MRI,PET etc)reports. OBESLC incorporates knowledge base of Lung cancer disease through which anybody who are not aware of this disease and especially medical students can have an idea regarding the reasons behind lung cancer attack, Types, stages of lung cancer and treatment plans available. This system contains probabilistic rules to diagnose the disease. In future we can enrich the system by incorporating a machine learning technique to learn from results of this system and use these results for further diagnosing. REFERENCES 1. Rawte, Vipula, and Bidisha Roy. "OBESTDD: Ontology based expert system for thyroid disease diagnosis." 2015 International Conference on Nascent Technologies in the Engineering Field (ICNTE). IEEE, 2015. 2. Al-Hamadani, Baydaa Taha, and Raad Fadhil Alwan. "An ontology- based expert system for general practitioners to diagnose cardiovascular diseases." Advances in Computational Sciences and Technology 8.1 (2015): 53-65. 3. NLCA: The National Clinical Lung Cancer Audit (LUCADA)Data Manual (2010), available from: http://www.hscic.gov.uk/lung 4. Sesen, M. Berkan, et al. "Lung cancer assistant: a hybrid clinical decision support application for lung cancer care." Journal of The Royal Society Interface 11.98 (2014): 20140534. 5. Sesen, M. Berkan, et al. "Lung Cancer Assistant: An ontology-driven, online decision support prototype for lung cancer treatment selection." OWL: Experiences and Directions Workshop (OWLED). 2012. 6. Klar, R., and A. Zaiss. "Medical expert systems: design and applications in pulmonary medicine." Lung 168.1 (1990): 1201-1209. 7. https://en.wikipedia.org/wiki/Expert_system 8. https://lungevity.org/for-supporters-advocates/lung-cancer-statistics 9. https://www.cancer.org/content/cancer/en/cancer/lung- cancer/about/key-statistics.html 10. https://www.mayoclinic.org/diseases-conditions/lung- cancer/symptoms-causes/syc-20374620 11. Thomas, Robert F. "The Benefits of expert systems in health care. practical experiences from CATEG05-ES." AIME 89. Springer, Berlin, Heidelberg, 1989. 93-97. 12. Brusa, Graciela, Ma Laura Caliusco, and Omar Chiotti. "A process for building a domain ontology: an experience in developing a government budgetary ontology." Proceedings of the second Australasian workshop on Advances in ontologies-Volume 72. Australian Computer Society, Inc., 2006. ------sparql 13. Noy, Natalya F., and Deborah L. McGuinness. "Ontology development 101: A guide to creating your first ontology." (2001). 14. Ganapathy, Gopinath, and S. Sagayaraj. "To generate the ontology from java source code OWL creation." (2011). 15. Tudorache, Tania, et al. "Supporting collaborative ontology development in Protégé." International Semantic Web Conference. Springer, Berlin, Heidelberg, 2008. 16. Fensel, Dieter, et al. "OIL: An ontology infrastructure for the semantic web." IEEE intelligent systems 16.2 (2001): 38-45. 17. Berners-Lee, Tim, James Hendler, and Ora Lassila. "The semantic web." Scientific american 284.5 (2001): 28-37. 18. Biswas, Dipanwita, et al. "Disease diagnosis system." International Journal of Computer Science & Informatics 1.2 (2011): 48-51. 19. Dehariya, Ashish, et al. "An effective approach for medical diagnosis preceded by artificial neural network ensemble." 2011 3rd International Conference on Electronics Computer Technology. Vol. 1. IEEE, 2011. 20. Wang, Hsien-Tseng, and Abdullah Uz Tansel. "Composite ontology- based medical diagnosis decision support system framework." Communications of the IIMA 13.2 (2013): 4 21. https://radiopaedia.org/articles/lung-rads 22. https://www.acr.org/Clinical-Resources/Reporting-and-Data- Systems/Lung-Rads 23. https://www.cancerresearchuk.org/about-cancer/lung-cancer/stages- types-grades/tnm-staging 24. The 8th lung cancer TNM classification and clinical staging system: review of the changes and clinical implications 25. https://radiologyassistant.nl/chest/lung-cancer-tnm-8th-edition 26. Liu, C-H., et al. "Ontology-based context representation and reasoning using owl and swrl." 2010 8th Annual Communication Networks and Services Research Conference. IEEE, 2010. 27. Sirisha, J., and M. Babu Reddy. "A Prototype Model for Developing Cancer Ontology in Medical Domain." (2017). 28. Knublauch, Holger, et al. "The Protégé OWL plugin: An open development environment for semantic web applications." International Semantic Web Conference. Vol. 3298. 2004. 29. Fernández-López, Mariano, Asunción Gómez-Pérez, and Natalia Juristo. "Methontology: from ontological art towards ontological engineering." (1997). 30. Harrison's Principles of Internal Medicine, 20e by J. Larry Jameson, Anthony S. Fauci, Dennis L. Kasper, Stephen L. Hauser, Dan L. Longo, Joseph Loscalzo ABOUT AUTHORS J.Sirisha, received her B.Tech degree in Computer science and Information Technology from Jawaharlal Nehru Technological University,Hyderabad and M.Tech degree in Computer Science and Engineering from Acharya Nagarjuna University,Guntur. She is working as Asst Professor in the Department of Information Technology, Prasad V. Potluri Siddhartha Institute of Technology, Vijayawada, Andhra Pradesh, India. She has 15 years of teaching experience and currently pursuing Ph.D from Krishna University, Machilipatnam.She hasPublished 14 research papers in various reputed Journals Her areas of interests include Data Mining, Semantic web Mining, and Social Networking. M.BabuReddy, received his Ph.D in Computer Science from Acharya Nagarjuna University,Guntur and M.Tech in Computer Science from GGIST,Patiala . He is presently working as Head(i/c) in the Department of Computer Science, Krishna University, Machilipatnam, Andhra Pradesh, India. He has 20 years of teaching experience .He has Published 62 research papers in various reputed Journals His areas of interests include Software Engineering, Artificial Intelligence, Machine Learning, Data Mining, Semantic web Mining 
work_2wnvavslprfthlo5orvk5l53du	----	Elsevier Editorial System(tm) for Expert Systems With Applications Manuscript Draft Manuscript Number: Title: An Ontology for Managing Network Services Quality Article Type: Full Length Article Keywords: Ontology; Multiservice IP Networks; Network Service Management; SLA/SLS; Semantics Corresponding Author: Prof. Paulo Carvalho, Ph.D. Corresponding Author's Institution: University of Minho First Author: Carlos Rodrigues, MSc Order of Authors: Carlos Rodrigues, MSc; Solange Rito Lima, Ph.D.; Luis Alvarez Sabucedo, Ph.D.; Paulo Carvalho, Ph.D. Highlights "An Ontology for Managing Network Services Quality" (Full length article) > Managing Internet services and resources urges for a formal and systematic approach. > Semantic description of network services fosters interoperability and self- management. > We propose an ontological model for multiservice IP networks. > Dynamic negotiation and auditing of network service contracts are improved. > An API is provided to allow service management platforms to access ontological contents. Highlights An Ontology for Managing Network Services Quality Carlos Rodrigues Department of Informatics, University of Minho Solange Rito Lima Department of Informatics, University of Minho Luis M. Álvarez Sabucedo Telematic Engineering Department, University of Vigo Paulo Carvalho Department of Informatics, University of Minho Preprint submitted to Expert Systems with Applications April 7, 2011 *Manuscript Click here to view linked References http://ees.elsevier.com/eswa/viewRCResults.aspx?pdf=1&docID=12406&rev=0&fileID=118483&msid={83070562-258D-42A8-8D62-36F27A54CEA1} An Ontology for Managing Network Services Quality Carlos Rodrigues Department of Informatics, University of Minho Solange Rito Lima Department of Informatics, University of Minho Luis M. Álvarez Sabucedo Telematic Engineering Department, University of Vigo Paulo Carvalho Department of Informatics, University of Minho Abstract The evolution of IP networks to a service-oriented paradigm poses new challenges to service providers regarding the management and auditing of network services. The road toward ubiquity, heterogeneity and virtualization of network services and resources urges for a formal and systematic approach to network management tasks. In this context, the semantic characterization and modeling of services provided to users assumes an essential role in fostering autonomic service management, service negotiation and configuration. The semantic and formal description of services and resources is also relevant to assist paradigms such as cloud computing, where a large diversity of resources have to be described and managed in a highly dynamic way. This paper is centered on the definition of an ontology for multiservice IP networks which intends to address multiple service management goals, namely: (i) to foster client and service provider interoperability; (ii) to man- age network service contracts, facilitating the dynamic negotiation between clients and ISPs; (iii) to access and query SLA/SLSs data on a individual or aggregated basis to assist service provisioning in the network; and (iv) to sus- tain service monitoring and auditing. In order to take full advantage of the proposed semantic model, a service model API is provided to allow service Preprint submitted to Expert Systems with Applications April 7, 2011 management platforms to access the ontological contents. This ontological development also takes advantage of SWRL to discover new knowledge, en- riching the possibilities of systems described using this support. Keywords: Ontology, Multiservice IP Networks, Network Service Management, SLA/SLS, Semantics 1. Introduction The evolution of the Internet as a convergent communication infrastruc- ture supporting a wide variety of applications and services poses new chal- lenges and needs to network management, which has to be more focused on managing services instead of network equipment. This approach requires the capability of viewing the network as a large distributed system, offering an encompassing set of services to users. Commonly, the type of service, its Quality of Service (QoS) requirements and other technical and administrative issues are settled between customers and Internet Service Providers (ISPs) through the establishment of Service Level Agreements (SLA). The technological component of this agreement is defined through Service Level Specifications (SLS). SLSs provide a valuable guidance to service deployment on network infrastructures and monitoring of contracts’ compliance. Attending to the ever growing number of home and business customers, contracted services and network heterogeneity, the implementation and management of network services are very demanding tasks for ISPs. Besides the inherent complexity, this process may lead to inefficient policy implementation and poor resource management. In fact, under the current variety of services offered, e.g. IP telephony, 3-play or 4-play solutions, the interaction between service providers and end customers is rigid and rather limitative regarding service negotiation and auditing tasks. For instance, from a user point-of-view, the possibility of a short-term upgrade on his access bandwidth to the Internet or a tight quality control of the subscribed service would be of undeniable relevance. From a service provider perspective, providing this sort of facilities, would clearly improve the level of service being offered, increasing competitiveness and resource management efficiency. These aspects are impelling ISPs to pursue autonomic solutions for service negotiation, configuration and management. Although several proposals exist in the literature toward achieving dy- namic service negotiation and management (D’Arienzo et al., 2004; Sarangan 3 and Chen, 2006; Cheng et al., 2008; Zaheer et al., 2010), the lack of a strong formal ground in addressing these tasks is evident and overcoming it is essen- tial (Atkinson and Floyd, 2004). A formal specification of network services management semantics is required as the building blocks to create reasoning mechanisms to allow developing self-managed ISPs. By using a knowledge based formal framework and an inference engine capable of reasoning over concepts, relations and changes of state, it is possible to create a more flexible and robust ground for specifying and implementing autonomic and adaptive management tasks. As a contribution in this context, this work proposes an ontology speci- fication in the domain of multiservice networks, which formally specifies the contractual and technical contents of SLAs, the network service management processes and their orchestration, promoting service autonomic management and configuration. This model provides support for a Service Management Platform that facilitates client and service provider interoperability, service contracts management including service data querying by the provider and, at some levels, by the client. This is enabled through a developed Service- Model API, which allows the applicational use of the proposed ontology. The multiservice network semantic model is developed in Web Ontology Language (OWL), assisted by the Protégé-OWL tool. The use of Seman- tic Web technologies enhances service management modeling expansiveness and reusability. This paper is structured as follows: research work on ontologies related to service definition and QoS is debated in Section 2; the developed model and each of its modules are presented in Section 3; the way semantics are applied based on the developed API is discussed in Section 4; examples of practical usage of the proposed model are provided in Section 5; the conclusions and future work are included in Section 6. 2. Related work Ontologies are being commonly used to bring semantics to the World- Wide Web (WWW). The WWW Consortium (W3C) developed the Resource Description Framework (RDF) (Brickley and Guha, 1999), a language for encoding knowledge on Web pages to make it understandable to electronic agents searching for information. The Defense Advanced Research Projects Agency (DARPA), in conjunction with the W3C, developed DARPA Agent Markup Language (DAML) by extending RDF with more expressive con- 4 structs aimed at facilitating agent interaction on the Web (Hendler and McGuinness, 2000). More recently, the W3C Web Ontology Working Group developed OWL (Web Ontology Language) (Bechhofer et al., 2004) based on description logic, maintaining as much compatibility as possible with the existing languages, including RDF and DAML. In the context of this work, several research studies focusing on ontologies for network services support and QoS are found within the research commu- nity. QoSOnt (Dobson et al., 2005) is an OWL ontology that centers on com- parative QoS metrics and requirements definition. For the purpose of us- ability and extensibility, the ontology is divided into different layers: the base layer, the domain usage layer, the attribute layer and the units layer. This structure allows replacing layers according to user needs. Although this ontology supplies the correct semantics for matchmaking, this was never demonstrated due to datatype limitations in OWL 1.0. To overcome this problem, a pure XML based solution was used, losing all of the virtues of OWL (Dobson and Sanchez-Macian, 2006). DAML-QoS (Zhou et al., 2004) is a QoS metrics ontology for Web Services (WS) developed in DAML+ OIL, with the aim of integrating the DAML-S framework (which evolved to OWL-S). The ontology is divided in three lay- ers: QoSProfile Layer, QoS Property Definition Layer and QoS Metrics Layer. The applicational scenario is defined in the QoSProfile, where customer in- quiring, QoS advertisement and templates’ definition can take place. The QoS properties domain rules are defined in the QoS Property Layer. The range of domain properties classes are defined in the QoS Metrics Layer. In (Zhou et al., 2005) a new Service Level Objective (SLO) concept, met- rics’ monitoring and statistical calculation semantics are presented. Through comparing SLOs, it is possible to infer that the initial WS performance ob- jectives are being met. For matchmaking bound restrictions, it is used the cardinality constraint of the ontology. The use of cardinality to express bounds upon QoS metrics is a misuse of ontology construction (Dobson and Sanchez-Macian, 2006). The second problem pointed to this model is the inexistent relation between the metrics definition and what they measure (Dobson and Sanchez-Macian, 2006) (Prudencio et al., 2009). MOQ (Kim et al., 2007) is another proposal of a QoS semantics model for WS, but it is not exactly an ontology. It only specifies axioms and does not present a taxonomy structure or a dictionary of concepts. This proposal covers in depth the concepts of composite requirements and service trace- 5 ability. It is divided into requirements ontology, measurement ontology and traceability ontology. The use of axioms allows requirement matching, re- quirement complexity identification, requirements compliance (through the establishment of conformance points) and service activities traceability. MonONTO (Moraes et al., 2008) ontology aims at creating a knowledge base to support a client recommendation system. The ontology serves as a support to a decision recommendation tool by providing high-level infor- mation to the user about the compliance of the network facing the service level demands. This process is primarily accomplished through the match- making of NetworkCharacteristics against ServiceCharacteristics individuals. These individuals are essentially concepts of QoS metrics. Some of the Net- workCharacteristics individuals relates to MeasurementTool individuals for monitoring tools conceptualization. This ontology was designed using OWL and SWRL (Semantic Web Rule Language). In (Alípio et al., 2006), it is proposed an ontology which aims at the au- tomation of network services management and mapping of services’ require- ments into the network. The ontology is viewed from three perspectives: the network service classification, the service level specification, the deploy- ment of network services. The network service categorization and the service level specification concepts follow (Babiarz et al., 2006) and Tequila (Goderis et al., 2003) guidelines. This ontology was developed in Flora-2 (based on F-Logic and Transaction Logic frameworks). Although being a more power- ful language, F-Logic lacks the interoperability and reusability of Semantic Web languages such as OWL. A group of generic ontologies to provide a framework for building SLAs is presented in (Green, 2006). The Unit Ontology contains all the comparable elements on SLA, with the intention of supporting the creation of any type of measurable unit. It also allows the definition of unit supported comparators and the creation of comparison operations. The other examples of avail- able ontologies are: the Temporal Ontology for temporal occurrences such as events and intervals; The Network Units Ontology for units related to telecommunications networks; and the SLA Ontology for basic SLA specifi- cation. Therefore, rather than a QoS ontology, it is proposed a set of reusable ontologies for providing support for other QoS semantic model implementa- tions. In (Royer et al., 2008), it is proposed an SLA ontology covering essentially Authentication, Authoring and Accounting (AAA) issues. It is applied a Profile-based solution, where the authorization for service use is dictated by 6 the user profile. A set of user profiles of a Customer entity are mapped into a set of SLAs. A user profile can be constrained to a defined SLS. For scalability reasons, users with the same user profile are settled into groups. There is also the notion of differentiated SLS and quantitative or qualitative type of metrics. Expressed through OWL, this ontology deepens the concepts of QoS Control Admission, which are not addressed by most of other QoS ontologies. The OWL developed ontology NetQoSOnt (Prudencio et al., 2009) in- tends mainly to be the support of a reasoning tool for service requirements matchmaking. It promotes the definition of SLSs containing quality param- eters that belong to the following levels: the Quality of Experience, the Quality in the Application Level, the Quality in the Network Level and the Quality in the Link Level. NetQoSOnt presents the concept of Layer and a separate module is created for each Layer defined. The layered structure of NetQoSOnt emulates the TCP/IP stack. A Base Module provides the skele- ton for layer creation. For QoS specification units, it is used the Measurement Units Ontology (MUO), a units specialized ontology. In the proposals discussed above, the lack of an unified and encompass- ing approach for semantic modeling of services and corresponding contracts in a multiservice environment is clear. In fact, most of the proposals are more focused on specification of network services metrics than on integrated service management. As mentioned, the focus is mainly on aspects such as: (i) the specification and characteristics of metrics; (ii) the process of met- ric compression and matchmaking; and (iii) description of services through metrics. In the present work, a holistic model for modeling multiservice networks is provided paying special attention to the characterization and auditing of services quality. This ontology focuses on service contracts to assist network services’ implementation by specifying how the defined contract elements are deployed in the network infrastructure, a feature not considered in the re- viewed works. Thus, the present ontology model provides a service contract description involving not only metrics, but other relevant entities to service management and network deployment, providing closer relations between classes of service and service contracts, and between service contracts and network infrastructure. Its modular structure leaves room to model expan- sion and integration with other proposals. 7 3. Multiservice Network Ontology The proposed model is divided in two main modules: the service man- agement module and the network module. As illustrated in Figure 1, these modules are organized as a layered structure where the upper layer has a dependency relation with the lower layer. This structure mimics real life where the management component is, indeed, above the physical network. This formal representation of a network is expressed in formal terms using the support of OWL, following the principles from Methontology (Fernández- López, M. et al., 1997). [Figure 1 about here.] The network module, as stated above, acts as the base layer. It includes concepts of network node, network interface and network equipment config- uration elements related to the implementation of contracted services in the network. The management module covers the domain network service man- agement related to service contracts, including service monitoring rules. This module uses several elements of the network module. Services are categorized by relating them to a type of SLS (Morand et al., 2005; Diaconescu et al., 2003; Goderis et al., 2003). According to recommendations from (Babiarz et al., 2006), ITU Y.1541 and 3GPP standards, current service types include: real-time services, multimedia services, data services, and default traffic ser- vice. Another important component of the proposed service model regards to multiservice monitoring (see Figure 1). This implies the definition of the main monitoring issues to include in the multiservice ontology to assist the auditing of Internet services both from an ISP and customer perspective. To service providers, it will also allow a tight control of services, network resources and related configuration procedures. On top of the Multiservice Ontology, a complete ServiceModel API offers to a Service Management Platform the access to the ontological contents. Without detailing the construction of the ontology at this point, it is relevant to highlight the identification of competence questions. These are the first and the last step in this methodology and fulfill the need to establish the requirements and the outcomes of the ontology itself, i.e., which questions the ontology will be able to answer. In the present case, the definition of an ontology for multiservice IP networks intends to address multiple service- oriented goals. Possible competence questions include: 8 (i) from a customer perspective: Which type of service packs are available for subscription at present? Which is the available bandwidth for a particular service from a specific access point? Is my contracted service being delivery within the negotiated QoS? (ii) from an ISP perspective: At an aggregate level, which is the allo- cated bandwidth for a particular service type? Which are the negotiated parameters per SLS? Which are the configuration parameters on each inter- face of edge network node and the available bandwidth per interface? Which services are supported between specific ingress and egress interfaces? Are the QoE/QoS requirements of a particular service being accomplished? On which points of the network are occurring QoS violations? In the description of modules provided in the sections below, a top-down approach will be followed to allow a broad view of the multiservice ontology. 3.1. Management Module The management module is where service contracts or SLAs are defined and managed. The first concept is the Client which identifies the customer part of the contract and stores all client information. A client is related to at least one SLA which represents a service contract. An SLA can have more than one SLS. The SLS structure, illustrated in Figure 2, follows the recommendations in (Morand et al., 2005; Diaconescu et al., 2003; Goderis et al., 2003), and is briefly described below. [Figure 2 about here.] • SLS Identification: This field identifies the SLS for management pur- poses, being used by both provider and customer. It is composed of a unique SLS id parameter and a Service id parameter, allowing to identify multiple SLSs within the same service. • Scope: The scope specifies the domain boundaries over which the ser- vice will be provided and managed, and where policies specified in a service contract are applied. Normally, SLSs are associated with uni- directional flows between at least one network entry point and at least one exit point. To cover bidirectionality, more than one SLS is associ- ated with a service. The entry points and the exit points are expressed through ingress and egress interfaces, respectively (see Section 3.2). At least two Interfaces (ingress and egress) instances must be specified. 9 The interface identification must be unique and is not restricted to the IP address (the identification can be defined at other protocol layer). • Traffic Classifier: The Traffic Classifier specifies how the nego- tiated service flows are identified for differentiated service treatment. Following Diffserv terminology, multifield (MF) classification and be- havior aggregate (BA) classification are supported (see Section 3.2). Usually, BA classification takes place over previously marked traffic, e.g. in network core nodes or, in the case of SLSs, between ISPs. Two traffic classifiers can be specified, an ingress traffic classifier and op- tionally an egress one. The ingress/egress classifier is then applied to each ingress/egress interface within the scope of the SLS. • Traffic Conditioner: This field specifies the policies and mechanisms applied to traffic flows in order to guarantee traffic conformance with the traffic profiles previously specified. Traffic conditioning occurs after traffic classification, so there is always a relation between the traffic classifier and the traffic conditioner specified within a SLS. An unlimited number of TrafficConditioner instances can be spec- ified. As in the traffic classifier property, the conditioners are divided into ingress and egress depending on their role. The ingress/egress conditioner is articulated with the ingress/egress classifier on each in- terface defined in the SLS scope as an ingress/egress QoS policy. This property is not mandatory. • Performance Guarantees: The Performance Guarantee fields specify the guarantees of service quality and performance provided by the ISP. Four quality metrics are considered: delay, jitter, bandwidth and packet loss, expressed through instances of the Bandwidth, Delay, Jitter and PacketLoss Metric subclasses. The definition of at least one instance of these Metric subclasses is mandatory, except on the Default Service type of SLS. Whenever there is a performance guarantee specification, a traffic conditioning action must also be specified. Delay and jitter are usually specified by their maximum allowed value or by a pair consisting of a maximum upper bound and a quantile. Packet loss (edge-to-edge) is represented by the ratio between the packet loss detected at the egress node and the number of packets sent at ingress node. Instead of quantitative, quality and performance parameters can also be specified in a qualitative manner. 10 • Reliability: The Reliability is usually specified by the mean downtime (MDT) and by the maximum allowed time to repair (TTR). The no compliance of the negotiated parameters may result in a penalty for the ISP. • Service Schedule: The Service Schedule defines the time period of service availability associated with an SLS. While a start date is al- ways specified, an end date is only specified in case of a reservation, ReservedServiceSchedule, in which the client requests the service during a specific period of time. In the default case, StandardService- Schedule, only the service start date is specified, i.e., the contract must be explicitly terminated by the client. • Monitoring: Monitoring refers to SLS’ performance parameters mon- itoring and reporting. For that purpose, a measurement period, a reporting date and a threshold notification are specified. Other pa- rameters such as the maximum outage time, total number of outage occurrences, reporting rules and reporting destination may be speci- fied. • Type of Service: The type of service is described by the Service class. This class allows the definition of services offered by the ISP to cus- tomers from a business-oriented perspective. Offered services are de- scribed through a set of qualitative metrics. The mapping from a qual- itative service description to a quantitative service specification is as- sured by the ISP. The Service class allows to relate the SLS with a specific instance of service offering. It also helps establishing SLS tem- plates on an application level. Services can be offered as a package (e.g. triple or quadruple play services) through the ServicePack class. 3.2. Network Module At present, an ISP is represented as a cloud network, where only edge (ingress and egress) nodes are visible. The abstract representation of domain internal nodes and inherent internal service configuration mechanisms are left for future work. Therefore, instead of representing configuration elements at per-hop level, the model is focused on a per-domain level. In this module (see Figure 3) there are three key elements: [Figure 3 about here.] 11 • Node: The Node class represents a network node (on the current model, corresponds to a domain border node). It is related to a set of Interface class instances. • Interface: The Interface class represents ingress and egress points of the ISP domain. Specifically it allows the mapping of external network interfaces or entry/exit points of ISP border nodes. The interface sup- ports a two-way traffic flow. It is possible to attach layer 2 and layer 3 addresses to the interface concept in order to relate it to a real net- work interface in the ISP domain. Each interface has a total bandwidth capacity and a reserved bandwidth capacity specified dynamically for ingress traffic and egress traffic. For QoS purposes, it is possible to specify a set of QoS policies. In this case, a QoS policy is a relation between a traffic classifier instance and a set of traffic conditioner in- stances applied to traffic classified by the former. A QoS policy can be an ingress policy or an egress policy. The Interface class, as illustrated in Figure 3, is defined by an iden- tifier, link and network layer addresses and total bandwidth capacity both downstream and upstream. It includes two counters for ingress and egress reserved bandwidth of all contracted services applied to this interface. Each instance can be related to a set of QoS policies applied on incoming and outgoing traffic. A boolean value is also defined for interface state indication. [Figure 4 about here.] • Traffic Classifier: The TrafficClassifier class, as represented in Fig- ure 4, has two subclasses: MF and BA. The BA class instances, applied to previously marked traffic, only have one field, a relation with a Mark class instance. The Mark class contemplates all forms of aggregated traffic marking (such as DSCP, IPv6 FlowLabel, MPLS Exp, etc.). The MF class allows the definition of traffic classification rules with multiple fields. There are no constraints on the number of allowed fields and these are divided into: link, network and transport header fields. This means that several types of fields can be used: IPv4 and IPv6 addresses, IPv6 Flow Label, ATM VPI/VCI and MPLS Labels. The fields used in the classification rule are combined through a logic operator rep- resented through the LogicOperator class instances AND and OR. For 12 a more complex classifying rule definition, other TrafficClassifier class instances can be stated as fields, working as nested classification rules. [Figure 5 about here.] • Traffic Conditioner: The traffic conditioner is designed to measure traf- fic flows against a predetermined traffic profile and, depending on the type of conditioner, take a predefined action based on that measure- ment. Traffic conditioning is important to ensure that traffic flows enter the ISP network in conformance with the established service profile. It is also an important policy for handling packets according to their con- formity level facing a certain traffic profile with the purpose of differen- tiating them in the network. According to their features, there are three TrafficConditioner subclasses (see Figure 5): the Marker, Policer and Shaper classes. The policer usually takes an immediate action on packets according to their compliance against predefined traffic pro- file. A Policer class instance must have a set of traffic measurement parameters and at least two levels of actions defined. Three different policers are defined in the current model. The TokenBucketPolicer represents a single rate policer with two level actions (for in profile and out of profile traffic). The SingleRateThreeColorMarker and TwoRateThreeColorMarker are examples of policers with three levels of conformance actions. The Shaper is the only conditioner subclass where no immediate action is taken on traffic flows. Instead, all pack- ets are buffered until traffic profile compliance is verified. The Marker class is a special type of conditioner which performs traffic marking and may be combined with other traffic conditioner elements. 3.3. Multiservice Monitoring Module As illustrated in Figure 1, the aim of the monitoring system to develop is twofold: (i) to monitor and control SLSs parameters in order to ensure that mea- sured values are in conformance with the negotiated service quality levels. This auditing purpose involves a prior characterization of each service re- quirements, monitoring parameters and corresponding metrics, and the defi- nition of appropriate measurement methodologies and tools to report multi- level QoS and performance metrics to users and system administrators; 13 (ii) to measure and control the usage of network resources. This includes the identification of network configuration aspects impacting on services per- formance, namely scheduling and queuing strategies on network nodes. In fact, monitoring network resources and triggering traffic control mechanisms accordingly will allow to maintain consistent quality levels for the supported services and the fulfillment of the negotiated SLSs. The monitoring process should provide measures reflecting the real status of services’ performance without introducing significant overhead or interfer- ing with network operation. Therefore, measurements have to be accurate, fast and carried out on a regular basis, while minimizing intrusion. For im- proving monitoring scalability, network services with identical QoS require- ments should be grouped and monitored as an aggregate, minimizing specific or customer dependent information. Another main concern of this task is to congregate users and ISPs perspec- tives regarding the description and control of services quality. This means that the perceived service quality for users (Quality of Experience - QoE), commonly expressed through subjective parameters, has to be identified and mapped into objective and quantifiable QoS parameters, able to be effec- tively controlled by network service providers. Therefore, the articulation of QoE and QoS, and the identification of appropriate measurement methodolo- gies for evaluating and controlling service quality levels in both perspectives (users and ISPs) is a main concern to cover in the present module. In this context, multiservice monitoring is expected to provide a clear identification and layering of all monitoring issues to include in the multi- service ontology to assist auditing and control of negotiated service levels through the proposed Service Management Platform. 3.4. The VoIP Service as Example As mentioned before, a service provider may describe each provided ser- vice through a set of qualitative metric values, which are then mapped to quantitative values to assist, for instance, configuration and service control. For example, a VoIP type of service can be described as: VoIP Service Bandwidth: Low_Bandwidth = at least 1 Mbps Delay: VeryLow_Delay = at most 100 ms Jitter: VeryLow_Jitter = at most 50 ms Packet Loss: VeryLow_Loss = at most 0.001 % of lost packets 14 This type of service description is used for SLS classification in accordance with the specified metrics. In other way, SLSs can be built based on the type of service description when required. An example of an SLS instance for the VoIP service is shown in Figure 6. [Figure 6 about here.] When the SLS instance is set, the TrafficClassifer and TrafficConditioner specified lead to QoS policy instances. A relation is then established between each QoS policy and network interfaces instances specified in the scope of the SLS (Figure 6). This policy information is useful for automating the deploy- ment of QoS mechanisms in the ISP network infrastructure. By establishing relations among all these entities, a change in one of them affects all other related entities. For example, a change in an SLS parameter is spread through all the corresponding SLS configurations in the network infrastructure. 4. Applying semantics This section discusses how the presented model is converted into an on- tological support. Thus, the characterization of the multiservice domain can be used in further software solutions ranging from web contents to complex software agents responsible for decision making. This ontology was devel- oped according to the basis proposed by Methontology (Fernández-López, M. et al., 1997). In this way, it is guaranteed its conformance with a set of methodological rules and the final product can be traced to its origin and reused in a simple and cost-effective manner. The proposed ontology provides the main concepts and properties re- quired to describe multiple services levels and corresponding quality in a network domain. For its implementation, according to the terminology pro- posed in Methontology, it was used Protégé to generate the OWL represen- tation. This representation uses not only classes and properties, but it also includes restrictions on the values of the previous ones. Therefore, it is en- sured the conformance of current contents and future pieces of information to the established parameters of the system. Besides per-class restrictions, a set of general rules are defined for es- tablishing new rule-based relations between individuals. These rules are ex- pressed using SWRL and they are applied to check information in order to 15 discover new possible instances and properties within the system. So far, there are defined rules for: • validation of interfaces capacity included on a contract scope; • compliance verification of monitored metrics in relation to service con- tract specifications; • changing interfaces network status; • qualitative classification of performance metrics; • classification of SLSs according to the type of service. For example, the following rule states that if all SLS performance metrics have qualitative values matching a definition of a type of service then the SLS specifies a service of that type. SLS(?sls)∧Service(?service) ∧ includesBandwidth(?sls, ?bandwidth) ∧ includesBandwidthQualV alue(?service, ?qualiBandwidth) ∧ includesDelay(?sls, ?delay) ∧ includesDelayQualV alue(?service, ?qualiDelay) ∧ includesJitter(?sls, ?jitter) ∧ includesJitterQualV alue(?service, ?qualiJitter) ∧ includesLossQualV alue(?service, ?qualiLoss) ∧ includesPacketLoss(?sls, ?loss) ∧ includesQualitativeV alue(?bandwidth, ?qualiBandwidth) ∧ includesQualitativeV alue(?delay, ?qualiDelay) ∧ includesQualitativeV alue(?jitter, ?qualiJitter) ∧ includesQualitativeV alue(?loss, ?qualiLoss) → definesSLS(?service, ?sls) Additional rules are defined for the above mentioned issues. Nevertheless, for other purposes, it is suggested to define rules at application level due to the complexity and limitations of SWRL (Zwaal et al., 2006) at using knowledge from different sources and involving advanced logical checks. On the top of the provided ontology, it is developed a complete software API. This API, referred as the ServiceModel API, is implemented following the diagram presented in Figure 7. 16 [Figure 7 about here.] The Jena Framework (McBride, 2002) plays a major role in the devel- oped software. It provides support for working with RDF and OWL based archives. The handling of OWL entities (classes, individuals and restrictions) is provided by the Jena Ontology API. Recall that the ontological content can be accessed from the local computer or from a remote server. The Pellet (Sirin et al., 2007) engine is the reasoner used due to its SWRL (Horrocks et al., 2004) support. Working on top of the Jena framework, the Jena beans API binds RDF resources (in this case, OWL classes) to Java beans simplifying the process of Java-to-RDF conversion. This feature enables users to work with individuals as Java objects. The persistence of the knowledge is guaranteed by the TDB (Owens et al., 2008) technology, which is clearly simpler and more efficient than the SDB solution (uses SQL databases for storing RDF datasets). However, the API integration of an SDB solution is not totally abandoned. The ServiceModel APl intents to assist future projects in several goals: (i) to foster client and service provider interoperability; (ii) to manage network service contracts, facilitating the dynamic negotiation between clients and ISPs; (iii) to access and query SLA/SLSs data on a individual or aggregated basis to assist service provisioning in the network; and (iv) to assist service monitoring and auditing. Therefore, this API, aimed to sustain further de- velopments, supports in a straight-forward manner for software developers the following features: • the insertion and removal of information on the Knowledge Base (cre- ating/destroying individuals); • the validation of the Knowledge Base information (classification and realization); • establishment of more complex rules based relations (not possible through SWRL); • Knowledge Base querying. This feature is implemented through SPARQL (Prud’hommeaux and Seaborne, 2008) and the ARQL Jena API; • Knowledge Base persistence. 17 5. Practical Application Once the semantic model for fully describing SLAs/SLSs is set, services on their top can be provided. Semantics is not, in general, an end by itself. On the contrary, its use is motivated by achieving further goals. In the case of this proposal, it is generated a framework to boost interoperability and advanced data mining features. Bearing that in mind, services were derived as proof-of- concept. Firstly, it is suggested a RDFa support to introduce annotations on xhtml descriptions about SLAs and SLSs. Afterwards, it is suggested the use of a semantic engine to recover information from a repository of information regarding the formers. RDFa (Birbeck and Adida, 2008) is a semantic technology that empowers users to include semantic annotations on XML (or xhtml) contents. These annotations are invisible for human user but easily recovered for software agents using GRDDL (Connolly, Editor). It is important to keep in mind that both technologies are official recommendations from W3C. Taking as a basis the provided OWL model for describing the system, annotations can be included in xhtml describing SLAs and SLSs. For the sake of clarity, it is included the following example: 1 <p xmlns:sla=" http: //owl . det . uvigo . es / s l s /"> 2 The provided connection under the i n t e r f a c e 3 <span property=" s l s : i d ">Service1</span> , 4 provides a t o t a l bandwidth of 5 <span property=" sls:includesTotalBandwidthCapacity "> 6 100 Mbps</span>. 7 </p> As shown in this xhtml snippet, a network interface and some of its properties are described. This definition of capacities of the Network Module can be directly recovered using GRDDL. The use case expected for this functionality is related to the web pages of ISPs. Service providers, when offering their services, can include this information into their web pages. Users will be able to recover this information through software agents on their behalf, and include it into a data repository for further decisions. Once these pieces of information are included in a Semantic Database, re- gardless of its origin, either from GRDDL extraction or from other sources, it is possible to get added-value services. Using SPARQL Queries (Prud’hommeaux 18 and Seaborne, 2008), for example, it is possible to locate SLAs/SLSs fulfill- ing specific properties the user is interested in. The only requirement is to identify the graph matching the desired properties and implement the corre- sponding SPARQL query. Authors successfully tested this feature by means of the API provided. It is actually rather simple to deploy a software tool that looks for, for instance, the cheapest SLA in the market or the one offering the fastest network access, among other features. 6. Conclusions and Future Work This paper has presented an innovative approach to the development of a semantic model in the domain of multiservice networks. This model for- mally specifies concepts related to service and SLS definition, network service management, configuration and auditing, creating the reasoning mechanisms to ground the development self-managed ISPs. Although being conceptually aligned with the differentiated service model, the solution is generic without being tied to a specific QoS paradigm. The usefulness of the present semantic service modeling has been pointed out for multiple applications in the context of multiservice management. In particular, aspects such as dynamic service negotiation between service provides and end customers, and auditing of Internet services being provided may be strongly improved as consequence of using the proposed ontology. Possibilities and features from this ontology are also presented to soft- ware developers by means of a ServiceModel API. The functionality within this library can be used for the above mentioned goals. Due to the modular schema for this software component, its inclusion on future projects consti- tutes a simple task that will provide a useful support in further developments. References M. D’Arienzo, A. Pescapè, G. Ventre, Dynamic Service Management in Het- erogeneous Networks, Journal of Network and System Management 12 (2). V. Sarangan, J. Chen, Comparative Study of Protocols for Dynamic Service Negotiation in Next Generation Internet, IEEE Communications Magazine 44 (3) (2006) 151–156. Y. Cheng, A. Leon-Garcia, I. Foster, Toward an Autonomic Service Man- agement Framework: A Holistic Vision of SOA, AON, and Autonomic 19 Computing, IEEE Communications Magazine 46 (5) (2008) 138–146, URL http://dx.doi.org/10.1109/MCOM.2008.4511662. F.-E. Zaheer, J. Xiao, R. Boutaba, Multi-provider Service Negotiation and Contracting in Network Virtualization, in: IEEE NOMS’10, 471–478, 2010. E. Atkinson, E. Floyd, IAB Concerns and Recommendations Regarding Internet Research and Evolution, RFC 3869, Internet Engineering Task Force, URL http://www.rfc-editor.org/rfc/rfc3869.txt, 2004. D. Brickley, R. Guha, Resource Description Framework (RDF) Schema Spec- ification, URL http://www.w3.org/TR/rdf-schema,W3C, 1999. J. Hendler, D. McGuinness, DARPA Agent Markup Language, IEEE Intel- ligent Systems 15 (6). S. Bechhofer, F. van Harmelen, J. Hendler, I. Horrocks, D. L. McGuinness, P. F. Patel-Schneider, , L. A. Stein, OWL Web Ontology Language Refer- ence, W3C, 2004. G. Dobson, R. Lock, I. Sommerville, QoSOnt: a QoS Ontology for Service- Centric Systems, in: EUROMICRO ’05: Proceedings of the 31st EUROMI- CRO Conference on Software Engineering and Advanced Applications, IEEE Computer Society, Washington, DC, USA, ISBN 0-7695-2431-1, 80– 87, URL http://dx.doi.org/10.1109/EUROMICRO.2005.49, 2005. G. Dobson, A. Sanchez-Macian, Towards Unified QoS/SLA Ontologies, in: SCW ’06: Proceedings of the IEEE Services Computing Workshops, IEEE Computer Society, Washington, DC, USA, ISBN 0-7695-2681-0, 169–174, URL http://dx.doi.org/10.1109/SCW.2006.40, 2006. C. Zhou, L.-T. Chia, B.-S. Lee, DAML-QoS Ontology for Web Services, in: IEEE International Conference on Web Services, IEEE Computer So- ciety, Los Alamitos, CA, USA, ISBN 0-7695-2167-3, URL http://doi. ieeecomputersociety.org/10.1109/ICWS.2004.1314772, 2004. C. Zhou, L.-T. Chia, B.-S. Lee, QoS Measurement Issues with DAML-QoS Ontology, in: IEEE International Conference on E-Business Engineering, IEEE Computer Society, Los Alamitos, CA, USA, ISBN 0-7695-2430-3, 395–403, URL http://doi.ieeecomputersociety.org/10.1109/ICEBE. 2005.100, 2005. 20 A. C. Prudencio, R. Willrich, M. Diaz, S. Tazi, Quality of Service Spec- ifications: A Semantic Approach, IEEE International Symposium on Network Computing and Applications (2009) 219–226URL http://doi. ieeecomputersociety.org/10.1109/NCA.2009.36. H. M. Kim, A. Sengupta, J. Evermann, MOQ: Web services ontologies for QoS and general quality evaluations, Int. Journal of Metadata, Semantics and Ontologies 2 (3) (2007) 195–200, ISSN 1744-2621, URL http://dx. doi.org/10.1504/IJMSO.2007.017612. P. Moraes, L. Sampaio, J. Monteiro, M. Portnoi, MonONTO: A Domain Ontology for Network Monitoring and Recommendation for Advanced In- ternet Applications Users, in: Network Operations and Management Sym- posium Workshops, IEEE NOMS 2008, 116–123, URL http://dx.doi. org/10.1109/NOMSW.2007.21, 2008. P. Alípio, J. Neves, P. Carvalho, An Ontology for Network Services., in: International Conference on Computational Science (3), 240–243, URL http://dx.doi.org/10.1007/11758532_33, 2006. J. Babiarz, K. Chan, F. Baker, Configuration Guidelines for DiffServ Ser- vice Classes, RFC 4594 (Informational), URL http://www.ietf.org/ rfc/rfc4594.txt, 2006. D. Goderis, Y. T’joens, C. Jacquenet, G. Memenious, G. Pavlou, R. Egan, D. Griffin, P. Georgatsos, L. Georgiadis, P. V. Heuven, Service Level Spec- ification Semantics, Parameters, and Negotiation Requirements, Internet- Draft, drafttequila-sls-03.txt, 2003. L. Green, Service level agreements: an ontological approach, in: ICEC ’06: Proceedings of the 8th international conference on Electronic commerce, ACM, New York, NY, USA, ISBN 1-59593-392-1, 185–194, URL http: //doi.acm.org/10.1145/1151454.1151490, 2006. J. C. Royer, R. Willrich, M. Diaz, User Profile-Based Authorization Poli- cies for Network QoS Services, in: IEEE International Symposium on Network Computing and Applications, IEEE Computer Society, Los Alamitos, CA, USA, ISBN 978-0-7695-3192-2, 68–75, URL http://doi. ieeecomputersociety.org/10.1109/NCA.2008.39, 2008. 21 Fernández-López, M., Gómez-Pérez, A, Juristo, N., METHONTOLOGY: From Ontological Art Towards Ontological Engineering., Symposium on Ontological Art Towards Ontological Engineering of AAAI. (1997) 33–40. P. Morand, M. Boucadair, R. E. P. Levis, H. Asgari, D. Griffin, J. Griem, J. Spencer, P. Trimintzios, M. Howarth, N. Wang, P. Flegkas, K. Ho, S. Georgoulas, G. Pavlou, P. Georgatsos, T. Damilatis, Mescal D1.3 - Final Specification of Protocols and Algorithms for Inter-domain SLS Manage- ment and Traffic Engineering for QoS-based IP Service Delivery and their Test Requirements , Mescal Project IST-2001-37961, 2005. A. Diaconescu, S. Antonio, M. Esposito, S. Romano, M. Potts, Cadenus D2.3 - Resource Management in SLA Networks, Cadenus Project IST- 1999-11017, 2003. H. Zwaal, M. Hutschemaekers, M. Verheijen, Manipulating context informa- tion with SWRL, A-MUSE Deliverable D3.12, 2006. B. McBride, Jena: a Semantic Web Toolkit, IEEE Internet Computing 6 (6) (2002) 55–59, ISSN 1089-7801, URL http://dx.doi.org/10.1109/MIC. 2002.1067737. E. Sirin, B. Parsia, B. Grau, A. Kalyanpur, Y. Katz, Pellet: A practical OWL-DL reasoner, Journal of Web Semantics 5 (2) (2007) 51–53, ISSN 1570-8268. I. Horrocks, P. F. Patel-Schneider, H. Boley, S. Tabet, B. Grosof, M. Dean, SWRL: A Semantic Web Rule Language Combining OWL and RuleML, W3C Member Submission, W3C, URL http://www.w3.org/Submission/ SWRL, 2004. A. Owens, A. Seaborne, N. Gibbins, mc schraefel, Clustered TDB: A Clus- tered Triple Store for Jena, in: WWW 2009, URL http://eprints.ecs. soton.ac.uk/16974/, 2008. E. Prud’hommeaux, A. Seaborne, SPARQL Query Language for RDF, W3C Recommendation, W3C, URL http://www.w3.org/TR/2008/ REC-rdf-sparql-query-20080115/, 2008. M. Birbeck, B. Adida, RDFa Primer, W3C Note, W3C, URL http://www. w3.org/TR/2008/NOTE-xhtml-rdfa-primer-20081014/, 2008. 22 D. Connolly(Editor), Gleaning Resource Descriptions from Dialects of Lan- guages (GRDDL), W3C Recommendation, W3C, URL http://www.w3. org/TR/grddl, 2007. 23 List of Figures 1 Service model diagram . . . . . . . . . . . . . . . . . . . . . . 25 2 SLS class diagram . . . . . . . . . . . . . . . . . . . . . . . . . 26 3 Interface class diagram . . . . . . . . . . . . . . . . . . . . . . 27 4 Classifier class diagram . . . . . . . . . . . . . . . . . . . . . . 28 5 Conditioner class diagram . . . . . . . . . . . . . . . . . . . . 29 6 SLS example diagram . . . . . . . . . . . . . . . . . . . . . . . 30 7 ServiceModel API structure diagram . . . . . . . . . . . . . . 31 24 Figure 1: Service model diagram 25 -i d : s tr in g -n a m e : s tr in g -e m a il : s tr in g -c o u n tr y : s tr in g -s tr e e tA d d re s s : s tr in g -t e le p h o n e : i n t C li e n t -m e tr ic V a lu e : f lo a t -u n it : s tr in g -u n it C o n v e rs io n : f lo a t -b a s e U n it : s tr in g M e tr ic -i d : s tr in g S L A -i d : s tr in g S L S -s ta rt D a te S e rv ic e S c h e d u le -i n c lu d e s S L A 0 .. * -i n c lu d e d S L A B y 1 -i n c lu d e s S L S 0 .. * -includedSLSBy 1 -i n c lu d e s S c o p e 1 .. * 1 -includesPerformanceGuarantees 1 1 -i s M o n it o re d B y 0 .. * -m o n it o ri z e s S L S 1 -i n c lu d e s S e rv ic e S c h e d u le 1 1 -i d : s tr in g N e tw o rk M o d u le :: T ra ff ic C la s s if ie r -i d : s tr in g N e tw o rk M o d u le :: T ra ff ic C o n d it io n e r -i d : s tr in g -n a m e : s tr in g -i n c lu d e s T o ta lB a n d w id th C a p a c it y : f lo a t -i n c lu d e s R e s e rv e d In g re s s B a n d w id th C a p a c it y : f lo a t -i n c lu d e s R e s e rv e d E g re s s B a n d w id th C a p a c it y : f lo a t -i n c lu d e s L a y e r2 A d d re s s : s tr in g -i n c lu d e s L a y e r3 A d d re s s : s tr in g -i s A c ti v e : b o o l N e tw o rk M o d u le :: In te rf a c e -i n c lu d e s C la s s if ie r 1 .. 2 1 -i n c lu d e s C o n d it io n e r * * -i s In C o n fo rm it y : b o o l M o n it o rS L S -s e rv ic e N a m e : s tr in g S e rv ic e -d e fi n e s S L S 0 .. * 1 -n a m e : s tr in g S e rv ic e P a c k -i n c lu d e s S e rv ic e 1 .. * * -i n c lu d e s P e rf o rm a n c e G u a ra n te e s 11 Figure 2: SLS class diagram 26 -i d : s tr in g -n a m e : s tr in g -i n c lu d e s T o ta lB a n d w id th C a p a c it y : f lo a t -i n c lu d e s R e s e rv e d In g re s s B a n d w id th C a p a c it y : f lo a t -i n c lu d e s R e s e rv e d E g re s s B a n d w id th C a p a c it y : f lo a t -i n c lu d e s L a y e r2 A d d re s s : s tr in g -i n c lu d e s L a y e r3 A d d re s s : s tr in g -i s A c ti v e : b o o l In te rf a c e -i d : s tr in g -n a m e : s tr in g N o d e -i d : s tr in g P o li c y -i d : s tr in g T ra ff ic C la s s if ie r -i d : s tr in g T ra ff ic C o n d it io n e r -i n c lu d e s In te rf a c e 0 .. * -i n c lu d e d In te rf a c e B y 1 -i n c lu d e s In g re s s P o li c ie s 0 .. * 1 -i n c lu d e s E g re s s P o li c ie s 0 .. * 1 N e tw o rk -i n c lu d e s N o d e 0 .. * 1 -includesClassifier 0 .. * 1 -includesConditioner 0 .. * * Figure 3: Interface class diagram 27 Mark -id : string TrafficClassifier BA -includesLayer2AddressSpec : string -includesLayer3AddressSpec : string -includesLayer4AddressSpec : string -includesProtocolID : string MF -in c lu d e s C la s s ific a tio n M a rk 1 * LogicOperator -includesLogicOperator 1 1 -includesNestedClassifier 0..* 1 Figure 4: Classifier class diagram 28 A c ti o n M A R K D R O P -i d : s tr in g T ra ff ic C o n d it io n e r M a rk e r S h a p e r -i n c lu d e s C o n fo rm a n c e L e v e l1 1 -i n c lu d e s O u tP ro fi le L e v e l1 1 -i n c lu d e s C IR : f lo a t -i n c lu d e s C B S : f lo a t -i n c lu d e s E B S : f lo a t S in g le R a te T h re e C o lo rM a rk e r -i n c lu d e s C IR : f lo a t -i n c lu d e s C B S : f lo a t -i n c lu d e s P IR : f lo a t -i n c lu d e s P B S : f lo a t T w o R a te T h re e C o lo rM a rk e r -i n c lu d e s E x c e e d P ro fi le L e v e l1 1 -i n c lu d e s S R : f lo a t -i n c lu d e s B S : f lo a t T o k e n B u c k e tP o li c e r -i n c lu d e s B S : f lo a t T o k e n B u c k e tS h a p e r -i n c lu d e s S R : f lo a t L e a k y B u c k e t T R A N S M IT -i n c lu d e s E x c e e d P ro fi le L e v e l 1 1 P o li c e r -i n c lu d e s In P ro fi le L e v e l1 1 Figure 5: Conditioner class diagram 29 id : s tr in g = S L S 1S L S 1 : S L S id : s tr in g = N 1 I1 n a m e : s tr in g = e th 0 in c lu d e s T o ta lB a n d w id th C a p a c it y : f lo a t = 1 0 0 0 0 0 0 in c lu d e s R e s e rv e d In g re s s B a n d w id th C a p a c it y : f lo a t = 1 0 2 4 in c lu d e s R e s e rv e d E g re s s B a n d w id th C a p a c it y : f lo a t = 0 in c lu d e s L a y e r2 A d d re s s : s tr in g = 0 0 :0 0 :0 0 :0 0 :0 0 :0 1 in c lu d e s L a y e r3 A d d re s s : s tr in g = 1 0 .0 .0 .1 is A c ti v e : b o o l = t ru e N 1 I1 : N e tw o rk M o d u le :: In te rf a c e id : s tr in g = N 2 I1 n a m e : s tr in g = e th 0 in c lu d e s T o ta lB a n d w id th C a p a c it y : f lo a t = 1 0 0 0 0 0 0 in c lu d e s R e s e rv e d In g re s s B a n d w id th C a p a c it y : f lo a t = 0 in c lu d e s R e s e rv e d E g re s s B a n d w id th C a p a c it y : f lo a t = 1 0 2 4 in c lu d e s L a y e r2 A d d re s s : s tr in g = 0 0 :0 0 :0 0 :0 0 :0 0 :0 2 in c lu d e s L a y e r3 A d d re s s : s tr in g = 1 0 .1 .0 .1 is A c ti v e : b o o l = t ru e N 2 I1 : N e tw o rk M o d u le :: In te rf a c e id : s tr in g = C l1 in c lu d e s L a y e r2 A d d re s s S p e c : s tr in g in c lu d e s L a y e r3 A d d re s s S p e c : s tr in g = 1 0 .0 .0 .5 1 in c lu d e s L a y e r4 A d d re s s S p e c : s tr in g = 1 2 3 4 5 in c lu d e s P ro to c o lI D : s tr in g C l1 : N e tw o rk M o d u le :: M F id : s tr in g = T k 1 in c lu d e s S R : f lo a t = 1 0 2 4 in c lu d e s B S : f lo a t = 1 5 0 0 T k 1 : N e tw o rk M o d u le :: T o k e n B u c k e tP o li c e r m e tr ic V a lu e : f lo a t = 1 0 2 4 u n it : s tr in g = b p s u n it C o n v e rs io n : f lo a t = 1 b a s e U n it : s tr in g = b p s B d 1 : B a n d w id th m e tr ic V a lu e : f lo a t = 1 0 0 u n it : s tr in g = m s u n it C o n v e rs io n : f lo a t = 1 b a s e U n it : s tr in g = m s D l1 : D e la y m e tr ic V a lu e : f lo a t = 5 0 u n it : s tr in g = m s u n it C o n v e rs io n : f lo a t = 1 b a s e U n it : s tr in g = m s J t1 : J it te r m e tr ic V a lu e : f lo a t = 0 .0 0 1 u n it : s tr in g = % u n it C o n v e rs io n : f lo a t = 1 b a s e U n it : s tr in g = % P L 1 : L o s s m e tr ic V a lu e : f lo a t = 9 9 u n it : s tr in g = % u n it C o n v e rs io n : f lo a t = 1 b a s e U n it : s tr in g = % R l1 : R e li a b il it y s ta rt D a te = 0 1 /0 1 /1 0 S c h d 1 : S ta n d a rd S e rv ic e S c h e d u le in c lu d e s In g re s s In te rf a c e in c lu d e s E g re s s In te rf a c e in c lu d e s In g re s s C la s s if ie r in c lu d e s In g re s s C o n d it io n e r in c lu d e s P e rf o rm a n c e G u a ra n te e s in c lu d e s S e rv ic e S c h e d u le in c lu d e s R e li a b il it y T k 1 A 1 : N e tw o rk M o d u le :: M A R K D R O P : N e tw o rk M o d u le :: D R O P in c lu d e s In P ro fi le A c ti o n in c lu d e s O u tP ro fi le A c ti o n E F : N e tw o rk M o d u le :: D S C P in c lu d e s R e s u lt in g M a rk id : s tr in g = S L S 1 P S L S 1 P : N e tw o rk M o d u le :: P o li c y in c lu d e s C la s s if ie r in c lu d e s C o n d it io n e r in c lu d e s In g re s s P o li c y Figure 6: SLS example diagram 30 PelletJena Framework JenaBeans TDB ServiceModel API Figure 7: ServiceModel API structure diagram 31 
work_2xgnobbhpnbmhmgkqlirqfz66q	----	[PDF] Expert Systems in Clinical Microbiology | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1128/CMR.00061-10 Corpus ID: 2239702Expert Systems in Clinical Microbiology @article{Winstanley2011ExpertSI, title={Expert Systems in Clinical Microbiology}, author={T. Winstanley and P. Courvalin}, journal={Clinical Microbiology Reviews}, year={2011}, volume={24}, pages={515 - 556} } T. Winstanley, P. Courvalin Published 2011 Computer Science, Medicine Clinical Microbiology Reviews SUMMARY This review aims to discuss expert systems in general and how they may be used in medicine as a whole and clinical microbiology in particular (with the aid of interpretive reading). It considers rule-based systems, pattern-based systems, and data mining and introduces neural nets. A variety of noncommercial systems is described, and the central role played by the EUCAST is stressed. The need for expert rules in the environment of reset EUCAST breakpoints is also questioned. Commercial… Expand View on ASM cmr.asm.org Save to Library Create Alert Cite Launch Research Feed Share This Paper 77 CitationsHighly Influential Citations 3 Background Citations 19 Methods Citations 12 View All Tables and Topics from this paper table 2 table 3 cefpodoxime Cefoxitin Expert Systems CNS disorder Aminoglycosides cefepime Infections, Hospital Cephalosporins Carbapenems Pneumonia Amoxicillin Antibiotics Interpretation Process Paper hearing impairment Instrument - device Rule (guideline) Scientific Publication Hyperactive behavior Phenotype Confusion Version 77 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency EUCAST expert rules in antimicrobial susceptibility testing. R. Leclercq, R. Cantón, +11 authors G. Kahlmeter Medicine Clinical microbiology and infection : the official publication of the European Society of Clinical Microbiology and Infectious Diseases 2013 422 PDF View 2 excerpts, cites background Save Alert Research Feed Clinical Microbiology Informatics D. Rhoads, V. Sintchenko, C. Rauch, L. Pantanowitz Medicine Clinical Microbiology Reviews 2014 40 Highly Influenced PDF View 6 excerpts, cites background and methods Save Alert Research Feed Clinical Decision Support Tools for Microbiology Laboratory Testing R. Jackups Computer Science 2020 Save Alert Research Feed Expert system in medicine and its application on pulmonary diseases Evren Burşuk, S. Demirci, M. A. Korpinar Medicine 2016 1 Save Alert Research Feed Applications of Artificial Intelligence in Clinical Microbiology Diagnostic Testing K. Smith, H. Wang, +5 authors D. Rhoads Computer Science 2020 2 Save Alert Research Feed Recommendations of the Spanish Antibiogram Committee (COESANT) for selecting antimicrobial agents and concentrations for in vitro susceptibility studies using automated systems. R. Cantón, A. Oliver, +27 authors L. Martinez-Martinez Medicine Enfermedades infecciosas y microbiologia clinica 2019 1 PDF Save Alert Research Feed Antimicrobial Susceptibility Testing Systems J. Karlowsky, S. Richter Medicine 2015 7 Save Alert Research Feed Validation of Antibiotic Susceptibility Testing Guidelines in a Routine Clinical Microbiology Laboratory Exemplifies General Key Challenges in Setting Clinical Breakpoints M. Hombach, P. Courvalin, E. Böttger Medicine Antimicrobial Agents and Chemotherapy 2014 12 PDF Save Alert Research Feed The critical influence of the intermediate category on interpretation errors in revised EUCAST and CLSI antimicrobial susceptibility testing guidelines. M. Hombach, E. Böttger, M. Roos Medicine Clinical microbiology and infection : the official publication of the European Society of Clinical Microbiology and Infectious Diseases 2013 30 View 1 excerpt, cites background Save Alert Research Feed Digital microbiology A. Egli, J. Schrenzel, Gilbert GREUB Computer Science, Medicine Clinical Microbiology and Infection 2020 View 1 excerpt, cites background Save Alert Research Feed ... 1 2 3 4 5 ... References SHOWING 1-10 OF 349 REFERENCES SORT BYRelevance Most Influenced Papers Recency [Interpretative reading and quality control of an antibiotic sensitivity test using an expert system. Application to the API ATB system and Enterobacteriaceae]. M. Peyret, J. Flandrois, G. Carret, C. Pichat Computer Science, Medicine Pathologie-biologie 1989 5 View 1 excerpt Save Alert Research Feed A decision support system for microbiology quality control B. Jackson, J. Schwartzman, Deborah E. Zuaro, E. Shultz Computer Science, Medicine AMIA 1997 2 View 1 excerpt, references background Save Alert Research Feed A study of the usage of a decision-support system for infective endocarditis C. Ekdahl, D. Karlsson, O. Wigertz, U. Forsum Medicine Medical informatics and the Internet in medicine 2000 17 View 3 excerpts, references background and methods Save Alert Research Feed Comparison and Evaluation of Osiris and Sirscan 2000 Antimicrobial Susceptibility Systems in the Clinical Microbiology Laboratory A. Nijs, R. Cartuyvels, A. Mewis, V. Peeters, J. Rummens, K. Magerman Biology, Medicine Journal of Clinical Microbiology 2003 33 PDF View 1 excerpt, references methods Save Alert Research Feed [Expert systems and antibiotic sensitivity test]. J. Flandrois, G. Carret Medicine Annales de biologie clinique 1991 2 Save Alert Research Feed Decision support systems for antibiotic prescribing V. Sintchenko, E. Coiera, G. Gilbert Medicine Current opinion in infectious diseases 2008 61 Save Alert Research Feed Potential Impact of the VITEK 2 System and the Advanced Expert System on the Clinical Laboratory of a University-Based Hospital C. Sanders, M. Peyret, +5 authors W. Sanders Medicine Journal of Clinical Microbiology 2001 35 PDF View 2 excerpts, references background Save Alert Research Feed [Bacterio-expert: an integrated system for assisting in the validation of antibiotic sensitivity tests. Retrospective application in 4053 Staphylococcus]. B. Legras, M. Weber, J. Legras, J. Burdin, L. Feldmann Medicine Pathologie-biologie 1991 1 Save Alert Research Feed [An expert system for differentiation of viral meningitis]. N. Janić, J. Kern, S. Vuletic Medicine Lijecnicki vjesnik 1991 1 View 1 excerpt, references background Save Alert Research Feed [Evaluation of ceftazidime treatment in septicemia expert systems]. H. Portier, C. Beuscart Medicine Presse medicale 1988 1 View 1 excerpt, references background Save Alert Research Feed ... 1 2 3 4 5 ... Related Papers Abstract Tables and Topics 77 Citations 349 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_2xtf2gstifhyza6sahun3emrqi	----	Expert System for Computer-assisted Annotation of MS/MS Spectra*□S Nadin Neuhauser‡¶, Annette Michalski‡¶, Jürgen Cox‡, and Matthias Mann‡§ An important step in mass spectrometry (MS)-based pro- teomics is the identification of peptides by their fragment spectra. Regardless of the identification score achieved, almost all tandem-MS (MS/MS) spectra contain remaining peaks that are not assigned by the search engine. These peaks may be explainable by human experts but the scale of modern proteomics experiments makes this impracti- cal. In computer science, Expert Systems are a mature technology to implement a list of rules generated by in- terviews with practitioners. We here develop such an Ex- pert System, making use of literature knowledge as well as a large body of high mass accuracy and pure fragmen- tation spectra. Interestingly, we find that even with high mass accuracy data, rule sets can quickly become too complex, leading to over-annotation. Therefore we estab- lish a rigorous false discovery rate, calculated by random insertion of peaks from a large collection of other MS/MS spectra, and use it to develop an optimized knowledge base. This rule set correctly annotates almost all peaks of medium or high abundance. For high resolution HCD data, median intensity coverage of fragment peaks in MS/MS spectra increases from 58% by search engine annotation alone to 86%. The resulting annotation performance sur- passes a human expert, especially on complex spectra such as those of larger phosphorylated peptides. Our system is also applicable to high resolution collision-in- duced dissociation data. It is available both as a part of MaxQuant and via a webserver that only requires an MS/MS spectrum and the corresponding peptides se- quence, and which outputs publication quality, annotated MS/MS spectra (www.biochem.mpg.de/mann/tools/). It provides expert knowledge to beginners in the field of MS-based proteomics and helps advanced users to focus on unusual and possibly novel types of fragment ions. Molecular & Cellular Proteomics 11: 10.1074/mcp. M112.020271, 1500 –1509, 2012. In MS-based proteomics, peptides are matched to peptide sequences in databases using search engines (1–3). Statisti- cal criteria are established for accepted versus rejected pep- tide spectra matches based on the search engine score, and usually a 99% certainty is required for reported peptides. The search engines typically only take sequence specific back- bone fragmentation into account (i.e. a, b, and y ions) and some of their neutral losses. However, tandem mass spectra— especially of larger peptides— can be quite com- plex and contain a number of medium or even high abun- dance peptide fragments that are not annotated by the search engine result. This can result in uncertainty for the user— especially if only relatively few peaks are annotated— be- cause it may reflect an incorrect identification. However, the most common cause of unlabeled peaks is that another pep- tide was present in the precursor selection window and was cofragmented. This has variously been termed “chimeric spectra” (4 – 6), or the problem of low precursor ion fraction (PIF)1 (7). Such spectra may still be identifiable with high confidence. The Andromeda search engine in MaxQuant, for instance, attempts to identify a second peptide in such cases (8, 9). However, even “pure” spectra (those with a high PIF) often still contain many unassigned peaks. These can be caused by different fragment types, such as internal ions, single or combined neutral losses as well as immonium and other ion types in the low mass region. A mass spectrometric expert can assign many or all of these peaks, based on expert knowledge of fragmentation and manual calculation of frag- ment masses, resulting in a higher degree of confidence for the identification. However, there are more and more practi- tioners of proteomics without in depth training or experience in annotating MS/MS spectra and such annotation would in any case be prohibitive for hundreds of thousands of spectra. Furthermore, even human experts may wrongly annotate a given peak— especially with low mass accuracy tandem mass spectra— or fail to consider every possibility that could have resulted in this fragment mass. Given the desirability of annotating fragment peaks to the highest degree possible, we turned to “Expert Systems,” a well-established technology in computer science. Expert Sys- tems achieved prominence in the 1970s and 1980s and were meant to solve complex problems by reasoning about knowl- From the ‡Department of Proteomics and Signal Transduction, Max-Planck Institute of Biochemistry, Am Klopferspitz 18, D-82152 Martinsried, Germany Received May 5, 2012, and in revised form, July 19, 2012 Author’s Choice—Final version full access. Published, MCP Papers in Press, August 10, 2012, DOI 10.1074/mcp.M112.020271 1 The abbreviations used are: PIF, Precursor Intensity Fraction; FDR, False Discovery Rate; MS/MS, Tandem mass spectrometry; HCD, Higher Energy Collisional Dissociation; PEP, Posterior Error Probability; PDF, Portable Document Format; IM, immonium ion; SC, side chain fragment ion; Th, Thomson. Technological Innovation and Resources Author’s Choice © 2012 by The American Society for Biochemistry and Molecular Biology, Inc. This paper is available on line at http://www.mcponline.org 1500 Molecular & Cellular Proteomics 11.11 edge (10, 11). Interestingly, one of the first examples was developed by Nobel Prize winner Joshua Lederberg more than 40 years ago, and dealt with the interpretation of mass spectrometric data. The program’s name was Heuristic DENTRAL (12), and it was capable of interpreting the mass spectra of aliphatic ethers and their fragments. The hypothe- ses produced by the program described molecular structures that are plausible explanations of the data. To infer these explanations from the data, the program incorporated a the- ory of chemical stability that provided limiting constraints as well as heuristic rules. In general, the aim of an Expert System is to encode knowl- edge extracted from professionals in the field in question. This then powers a rule-based system that can be applied broadly and in an automated manner. A rule-based Expert System represents the information obtained from human specialists in the form of IF-THEN rules. These are used to perform oper- ations on input data to reach appropriate conclusion. A ge- neric Expert System is essentially a computer program that provides a framework for performing a large number of infer- ences in a predictable way, using forward or backward chains, backtracking, and other mechanisms (13). Therefore, in contrast to statistics based learning, the “expert program” does not know what it knows through the raw volume of facts in the computer’s memory. Instead, like a human expert, it relies on a reasoning-like process of applying an empirically derived set of rules to the data. Here we implemented an Expert System for the interpreta- tion for high mass accuracy tandem mass spectrometry data of peptides. It was developed in an iterative manner together with human experts on peptide fragmentation, using the pub- lished literature on fragmentation pathways as well as large data sets of higher-energy collisional dissociation (HCD) (14) and collision-induced dissociation (CID) based peptide iden- tifications. Our goal was to achieve an annotation perform- ance similar or better than experienced mass spectrometrists (15), thus making comprehensively annotated peptide spectra available in large scale proteomics. EXPERIMENTAL PROCEDURES The benchmark data set is from Michalski et al.2 Briefly, E. coli, yeast and HeLa proteomes were separated on 1D gel electrophoresis and in gel digested (16). Resulting peptides were analyzed by liquid chromatography (LC) MS/MS on a linear ion trap - Orbitrap instrument (LTQ Velos (17) or ELITE (18), Thermo Fisher Scientific). Peptides were fragmented by HCD (14) or by CID, but in either case fragments were transferred to the Orbitrap analyzer to obtain high resolution tandem mass spectra (7500 at m/z 400). We scanned tandem mass spectra already from m/z 80 to capture immonium ions as completely as possible. Data analysis was performed by MaxQuant using the An- dromeda search engine (8, 9). Maximum initial mass deviation for precursor peaks was 6 ppm and maximum deviation for fragment ions for both the search engine and for the Expert System was 20 ppm. MaxQuant preprocessed the spectra to be annotated by the Expert System in the same way as it does for the Andromeda search engine: Peaks were filtered to the 10 most abundant ones in a sliding 100 m/z window, de-isotoped and shifted to charge one where possible. From this data, sequence-spectra pairs were selected that had a certainty of identification of 99.99% PIF values (7) larger than 95% and that were sequence unique (more than 16,000 peptides). The Expert System was written in the programming language C#, using the Microsoft .NET framework version 3.5 and the Workflow. Activities library, which contains a rule engine to implement an Expert System (Microsoft Corporation, Redmond, WA). MaxQuant contains the Expert System as an integrated option in its Viewer—the component that allows visualization of raw and anno- tated MS data. MaxQuant can freely be downloaded from www. maxquant.org. It requires Microsoft .NET 3.5, which is either already installed with Microsoft Windows or can be installed as a free Win- dows update. In our group we have implemented the Expert System both on a Windows cluster and in a desktop version. Additionally, we provide an Expert System web server, which can be accessed at 2 Michalski, A., Neuhauser, N., Cox, J., and Mann, M., unpublished data. FIG. 1. Basic concept of the Expert System. A, An Expert System is constructed by interviewing an expert in the domain (here peptide fragmentation and the accumulated literature) and devising a set of rules with associated priority and dependence on each other. The knowledge base contains the rules whereas the rule engine is generic and applies the rules to the data. B, Data are automatically processed following the steps depicted. Expert System for Annotation of MS/MS Spectra Molecular & Cellular Proteomics 11.11 1501 www.biochem.mpg.de/mann/tools/. Although MaxQuant allows the Expert System annotation of arbitrary numbers of MS/MS spectra, the webserver is currently limited to the submission of one MS/MS spec- trum at a time. After upload of a list of peaks with m/z value and their intensities—together with the corresponding peptide sequence—the spectrum with all annotations is displayed. This can then be exported in different graphical formats. RESULTS AND DISCUSSION Construction of the Expert System—Human experts per- form a generic set of tasks when solving problems such as the interpretation of an MS/MS spectrum. These rules have to be codified in the Expert System, mainly in the form of a series of IF-THEN rules. Fig. 1 shows the major steps involved in build- ing and using the Expert System. It is important to acquire all relevant rules to interpret MS/MS spectra as comprehensively as possible. However, to avoid over-annotation leading to false positives (see below), the number of rules and their interactions should not become too large. This balance was struck by evaluating the performance of different set of rules on large data sets in conjunction with human experts. Rules were encoded in a table-like structure, where they could be activated, deactivated or modified. To create the knowledge base, the extent of interactions of the rules also had to be determined—for instance, which combination of neutral losses to allow. After iterative construction of the knowledge base, the rule engine then applied the encoded knowledge to MS/MS spectra and displayed the result to the user (Fig. 1A). The processing steps that are performed on the raw MS and MS/MS spectra are shown in Fig. 1B (see also EXPERIMENTAL PROCEDURES). Note that the workflow is entirely automated and that user interaction is possible but not required. Arbitrary numbers of annotated spectra of pep- tides of interest can be produced as interactive screen images or high resolution, printable PDF files. The Expert System is very fast, and 16,000 spectra can be annotated in less than four hours on a desktop system. The IF-THEN constraints of our Expert System can be di- vided into four major parts (Fig. 2). At first the Expert System calculates any specific backbone fragments (a, b, and y-ion series), the charged precursor ion, the immonium ions as well as side chain fragments in the low-mass region and places them into a queue. In the second part of the workflow every element in this queue is filtered with respect to the actual MS/MS spectrum. Even if there is a peak corresponding to a calculated item in the queue, it may still be filtered out (symbolized by missing annotations after the filter in Fig. 2). For instance, a b1 ion is only allowed in very restricted circumstances. In the third step, neutral losses and internal fragments for the filtered values are calculated and added to the queue. They are then subjected to the same filtering rules as in step 2. Step 3 is iterative, as several subsequent neutral losses may be allowed. In the fourth and last step each potential annotation is given a priority. If there is more than one possible annotation, the one with the highest priority is chosen (i.e. the one that trig- gered the rules with higher priority). However, in this case the Expert System provides a pop-up (or “tool-tip”) containing the other possibility when hovering the mouse over the peak. (This can still happen if the FDR is properly controlled and is then typically caused by two different chemical designations for the same ion; or by different ions with the same chemical composition, such as small internal fragments with different sequence but the same amino acids). Determining a False Discovery Rate for Peak Annotation— Use of a very high threshold for peptide identification (99.99%) ensured that virtually none of the peptides in our collection should be misidentified. However, when building FIG. 2. Work flow of the Expert System. ➀ From the database sequence of the peptide identified by the search engine, a list of possible fragment ions is created. ➁ Peaks from the measured spec- trum are compared with the possible fragments and preliminarily annotated if they pass the rules of the Expert System. ➂ Neutral losses and internal fragments are generated from the candidate, annotated peaks and exposed to the Expert System rules. ➃ Potential conflicts are resolved via the priority of the annotations and peaks are labeled. Note that possible internal fragment ‘CA’ is crossed out because the b2 ion has the higher priority. Expert System for Annotation of MS/MS Spectra 1502 Molecular & Cellular Proteomics 11.11 FIG. 3. Calculation of false discovery rate for peak annotations. A, The upper panels represent a large number of identified MS/MS spectra from which annotated peaks are drawn to form a large peak collection of possible fragment masses. From each identified spectrum in the data set, 10 random fragments are inserted and the number of annotations by the Expert System is counted. This process is repeated 500 times for each peptide. B, Median FDR as determined in A as a function of peptide length distinguished by the mass difference of fragment ion and theoretical mass. The FDR for peak annotation rises with peptide length and is strongly dependent on the mass difference. Box plot at the bottom shows that 50% of the peptides were between 12 and 18 amino acids long. The box plots on the right summarize the range of FDR values regardless of peptide length. C, Graph of the median FDR as a function of peptide length but separated by intensity classes of the false annotated fragment peaks. Most false positives come from the low abundant peaks (blue) rather than the medium (green) or high abundance fragment peaks (yellow). D, Same plot as above but differentiated by the fragment ion type of the false positives. Getting lower number of false positives from regular fragment annotations (blue), compared with internal fragment (green) and neutral loss annotations (yellow). Expert System for Annotation of MS/MS Spectra Molecular & Cellular Proteomics 11.11 1503 the Expert System, we noticed that it was still possible to over-interpret the MS/MS spectra. This was initially surprising to us because our large scale data set had good signal to noise and peaks was only candidates for annotation when their calculated mass was less than 20 ppm from the ob- served mass. The over-interpretation became apparent through conflicting annotations for the same peak, and was typically caused by a combination of rules, such as several neutral losses from major sequence specific backbone or internal ions. Be- cause conflicting or wrong annotations would undermine the entire rational for the Expert System, we devised a scheme to stringently control the false discovery rate for peak annotation. The false discovery rate (FDR) is meant to represent the percent probability that a fragment peak is annotated by FIG. 4. Example spectra before and after Expert System annotation. A, Based on the search engine result, 34% of the fragments by peak intensities and 24% by peak number are explained, whereas the Expert System almost completely annotates the spectrum (for further explanation see main text). Posterior Error Probability (PEP) a statistical expectation value for peptide identification in Andromeda. Apart from the large fraction of a-, b-, and y-ions (pale blue/dark blue/red) and ions with neutral losses (orange), one can find internal fragment ions (purple) and in the low mass region one immonium ion of Isoleucine (green) and a side chain loss from arginine (turquoise). B, Expert System annotation of a phosphorylated peptide. Apart from the internal ions, several phosphorylation-related fragment ions were found. The asterisk (*) denotes loss of H3O4P with a delta mass of 97.9768 from the phosphorylated fragment ion. Expert System for Annotation of MS/MS Spectra 1504 Molecular & Cellular Proteomics 11.11 chance because its mass fits one of the Expert System rules for the peptide sequence. To calculate a proper FDR, we therefore needed to provide a set of background peaks that would represent false positives when they are labeled by the Expert System. Producing realistic background peaks turned out to be far from trivial because they need to have possible masses that can in principle be generated from peptide se- quences and they need to be independent of the sequence of the peptide in question. The principle of our solution to this problem is shown in Fig. 3A. From the large data set under- lying this study, we collect the m/z values of all annotated peaks, except those coming from immonium or side chain ions. They were stored in a large peak collection of several million entries, together with the respective peptide se- quences and the relative intensity of the peak. For each spec- trum in which we wanted to determine the FDR, we then inserted a random set of 10 peaks from the collection, where after we checked if the sequence of the selected peaks was independent from the sequence of the current spectrum. If one of the inserted peaks overlapped with an existing peak, it was discarded. By definition these 10 peaks represent pos- sible peptide fragments and, because they are chosen ran- domly from millions of other peaks, they collectively represent a good approximation to a true background set. This would FIG. 4 —continued Expert System for Annotation of MS/MS Spectra Molecular & Cellular Proteomics 11.11 1505 not be the case for permutation of the sequence of the pre- cursor in question, for instance, because many of the frag- ment peaks in permutated sequences are identical. Whenever the Expert System annotated one of these peaks, it was counted as a false positive. To find the number of repeats necessary to obtain a stable FDR for this procedure, we chose a set of spectra and simulated a thousand times on each one. We found that the FDR was constant after 500 iterations. For the final FDR calculation, for each spectrum we added a different set of 10 random peaks from the collection and repeated this 500 times. This was then applied to each of the more than 16,000 pure (high PIF) spectra in the large scale data set. Beyond providing a solid FDR estimate for each rule set, this procedure also allowed us to identify the rules or rule combinations that were responsible for miss-annotation, i.e. the rules that falsely annotated the inserted peaks. These mostly turned out to be chains of subsequent neutral losses. In conjunction with detailed evaluation of the frequency of ion types, we iteratively designed an optimal rule set (supplemen- tal Table S1). For instance, neutral losses from a particular amino acid were allowed if they occurred in more than five percent of the fragment sequences that contained that amino acid. Likewise, of a set of about 42 possible neutral side chain losses, only six were sufficiently important to retain them in the Expert System. The Figs. 3B–3D show the results of the median FDR as a function of the peptide length based on this final rule set. The overall FDR—indicated in red—is the same in all plots and shows a clear growing trend in the number of false positives with the length of the peptides. For small peptides of 12 amino acids or less, the FDR was less than 2.1% and all peptides in the range investigated had a peak annotation FDR of less than 5%. With these settings, the annotations are correct in more than 97% of the cases for the vast majority of MS/MS spectra. The Expert System could of course be pruned to provide a lower FDR by narrowing the mass tolerance window; however, this would come at the ex- pense of discarding correct annotations. To explore the influ- ence of mass accuracy on potential false positive annotations, we repeated these calculations with required mass deviations no larger than 5 ppm or no larger than 10 ppm. As can be seen in Fig. 3B, this further reduced possible errors to less than 1%, or less than 0.3%, respectively. This highlights the value of high mass accuracy in unambiguously identifying fragment mass identity. Furthermore, peaks with a low signal to noise are more likely to be miss-annotated than more intense peaks. In Fig. 3C we sorted the peak intensity of the false positives into three intensity classes (Fig. 3C). The median FDR of peaks with high or medium abundance are only 0.1 or 0.5%. For low abundance peaks it is higher but still with a median of no more than 2.1%. Next we separately investigated the FDR as a function of peptide length for the different fragment ion types. As can be seen in Fig. 3C, regular ions and internal fragments contribute very little to overall false annotation (0.4 and 0.5%), whereas neutral loss ions are wrongly annotated in 1.8% of the case or even more. FIG. 5. Expert System performance on a large data set. Median sequence coverage by summed fragment ion intensity is plotted as a function of identification score. Statistics is based on more than 16,000 spectra. For every identification score, the Expert System adds a large proportion of explainable peaks. Box plot below the graph indicates that 50% of peptides in the set have an Andromeda score between 98 and 140. Box plots on the right indicate the range of values for the intensity coverage for standard and Expert System annotation. Expert System for Annotation of MS/MS Spectra 1506 Molecular & Cellular Proteomics 11.11 http://www.mcponline.org/cgi/content/full/M112.020271/DC1 http://www.mcponline.org/cgi/content/full/M112.020271/DC1 Performance of the Expert System—Fig. 4 shows an illus- trative example of an HCD fragmented peptide before and after Expert System evaluation. The peptide was identified with an Andromeda score of 136 and posterior error proba- bility (PEP) of 1.1E-21 (the corresponding Mascot score was 83). The spectrum features an uninterrupted b-ion series from b2 to b9 and an uninterrupted y-ion series from y1 to y12, together covering the entire peptide sequence. Despite this unambiguous identification, the peaks used by the search engine to identify the peptide only accounted for 35% of the summed intensity of the peaks in the fragmentation spectrum. Coverage by number of explained peaks was even lower at 24% (allowing up to 10 peaks per 100 Th in the measured spectrum see EXPERIMENTAL PROCEDURES). There is a series of high abundance, high m/z fragments as well as a large number of low abundance peaks in the low and me- dium m/z range that are unexplained by the search engine. After annotation by the Expert System, this situation changes entirely. The high m/z series is revealed to be a prominent loss of CH4SO from oxidized methionine. The low FIG. 6. Web interface for the Expert System. A, Text field to paste the spectrum in text format (m/z value; intensity in arbitrary units). B, Form to enter the peptide sequence, modifications and their positions. C, Detected backbone fragments and their neutral losses are indicated in the peptide logo. Scalable spectrum annotated by the Expert System. Note that neutral loss peaks are very small compared with the major backbone fragments. The spectrum can be downloaded with the desired resolution and in the desired graphical format. Expert System for Annotation of MS/MS Spectra Molecular & Cellular Proteomics 11.11 1507 mass ions are neutral losses, internal fragments and com- binations between them and they were unambiguously and correctly assigned. Altogether, the Expert System ac- counted for almost all prominent ions and explained a total of 88% of the ion current. Manual annotation of this spec- trum would have been possible but would have been very time consuming. Interpretation of phosphorylated peptides, especially large ones, is more difficult than that of unmodified peptides. Fur- thermore, accurate placement of the phosphorylation site can be challenging. We used literature knowledge (19, 20) and the results of a large-scale investigation into the fragmentation of phosphorylated peptides to derive suitable fragmentation rules for the Expert System. This led to an additional six rules, which were easily integrated, illustrating the extensibility of the Expert System. Fig. 4B depicts an example annotation of the relatively complex fragmentation spectra typical of phos- phorylated peptides. The large ion series from the low mass range to about mass 1000 is caused by an extensive and uninterrupted internal ion series starting from the proline in the second position of the peptide sequence. As these internal fragments contain several glutamines, they lead to additional water and ammonia losses. However, there are also newly annotated fragments resulting from neutral losses in addition to loss of the phosphorylation site. Moreover, the neutral loss of HPO3 is annotated. Large-scale Evaluation of the Performance of the Expert System—We used the population of 16,000 spectra with high PIF—identified with a false discovery rate of 0.01% by the search engine—and annotated them automatically using the Expert System. For each spectrum we calculated the intensity coverage obtained by the fragments used by the search en- gine and the fragments explained by the Expert System. Higher scoring fragmentation spectra would be expected to have a larger fraction of their ion current annotatable than lower scoring peptides. Fig. 5A shows a plot of the median of these values for all search engine scores. A total of 95% of these Andromeda scores are within a range of 96 to 138. Here the median intensity coverage by standard annotation varies from 55% at 96 to 64% at 138. The Expert System, in con- trast, annotated between 86 and 89% of the total ion current in the fragment spectra of the same peptides. This repre- sents an average increase of 28%. There was only a small percentage of peptides that were lower scoring than 96 and for these the increased annotation percentage of the Expert System was even larger (34%). Interestingly, even in very high scoring HCD fragment spectra there are still many peaks not directly annotated by the search engine. For these, the average increase of annotated ion current be- cause of the Expert System was still 23%. The rule set of the Expert System was derived from HCD data. However, HCD and CID appear to produce similar ion types, although with different abundances. We therefore tested if the derived rule set was also applicable to high resolution CID data. This was indeed the case, and a total of 85% of the ion current in high resolution CID spectra ex- plained by the Expert System, although in CID spectra a higher percentage (79%) of the peaks are already accounted for by standard ion types. Therefore we conclude that the Expert System can be used equally well for high resolution HCD and CID data although the benefits for CID are not as large as they are for HCD. Webserver for Expert System Annotation of Spectra—The Expert System is now part of the Viewer component of Max- Quant, which is freely available at www.maxquant.org. In this environment, the Expert System can annotate arbitrarily large data sets of identified peptides and visualize and export them in different graphical formats such as PDF. Additionally, we established a webserver to make the Expert System available to any proteomics scientist, regardless of the computational workflow that he or she is using. The webserver is located at http://www.biochem.mpg.de/mann/tools/and its graphical in- terface is shown in Fig. 6. The user needs to supply a mass spectrum in the form of an m/z and peak intensity list as well as the sequence of the identified peptide (Figs. 6A, 6B). Com- mon modifications and their position in the sequence can also be specified. The webserver then provides an annotation of the spectrum within the stated mass tolerance as shown in Fig. 6C. The graph is scalable to enable detailed study of complex fragmentation spectra. Mass deviations in ppm (cal- culated mass – measured mass) can also be depicted. This annotated spectrum can be downloaded in a number of graphical formats for use in publications. CONCLUSION AND OUTLOOK Here we have made use of Expert Systems—a well-known technology in computer science—to automatically but accu- rately interpret the fragmentation spectra of identified pep- tides. We have shown that the Expert System performs very well on high mass accuracy data, annotating the large major- ity of medium to high abundance peaks. For HCD spectra it explains on average 28% more of the peak intensities than the search engine results alone. We derived a rigorous false pos- itive rate, ensuing that less than 5% of peaks can be miss- annotated—this rate is even lower for spectra with at least median scores and fragment ion intensities of at least mod- erate abundance. The rule set was derived by iterative inter- pretation of large HCD data set but we show that the Expert System is equally applicable to high resolution CID spectra. We envision different uses for the Expert System: For be- ginners in MS-based proteomics, it enables efficient training in the interpretation of MS/MS spectra without requiring much input from a specialist. For advanced users, it allows focusing on unusual and potentially novel types of fragments. One caveat is that the Expert System currently cannot explain fragment peaks that belong to cofragmented precursors; a very common occurrence that we deliberately avoided here by selecting only pure MS/MS spectra. This limitation can be Expert System for Annotation of MS/MS Spectra 1508 Molecular & Cellular Proteomics 11.11 addressed if both precursors are identified and communi- cated to the Expert System. Such a feature might be partic- ularly useful for instruments that allow deliberate multiplexing of precursors, which leads to complex MS/MS spectra (21). The Expert System has been in routine use in our laboratory for a number of months. During this time we have found that it provides helpful confirmation of the identification of the peptide and the identity of the previously unlabeled fragment ions. This is particularly welcome in the case of complicated spectra of important peptides, such as the ones regulated in the biological function in question. Compared with a human expert, the principal advantages of the Expert System are its speed, its ability to check for all supplied rules in a consistent manner as well as its rigorously controlled false positive rate. Obviously, the Expert System is limited to the knowledge supplied whereas an experienced mass spectrometrist can go beyond these rules and discover the origin of novel fragmen- tation mechanisms. As we have shown here, Expert Systems can readily be applied to problems in computational proteomics. Given their relative ease of implementation, they may become useful in other areas in MS-based proteomics, too. Acknowledgments—We thank Forest White for critical comments on this manuscript. * This work was supported by funding from the European Union 7th Framework project PROSPECTS (Proteomics Specification in Time and Space, grant HEALTH-F4-2008-201645). □S This article contains supplemental Table S1. ¶ These authors contributed equally. § To whom correspondence should be addressed: Department of Proteomics and Signal Transduction, Max-Planck Institute of Bio- chemistry, Am Klopferspitz 18, D-82152 Martinsried, Germany. E-mail: mmann@biochem.mpg.de. REFERENCES 1. Steen, H., and Mann, M. (2004) The ABC’s (and XYZ’s) of peptide sequenc- ing. Nat. Rev. Mol. Cell Biol. 5, 699 –711 2. Nesvizhskii, A. I., Vitek, O., and Aebersold, R. (2007) Analysis and validation of proteomic data generated by tandem mass spectrometry. Nat. Meth- ods 4, 787–797 3. Granholm, V., and Käll, L. (2011) Quality assessments of peptide-spectrum matches in shotgun proteomics. Proteomics 11, 1086 –1093 4. Houel, S., Abernathy, R., Renganathan, K., Meyer-Arendt, K., Ahn, N. G., and Old, W. M. (2010) Quantifying the impact of chimera MS/MS spectra on peptide identification in large-scale proteomics studies. J. Proteome Res. 9, 4152– 4160 5. Zhang, N., Li, X. J., Ye, M., Pan, S., Schwikowski, B., and Aebersold, R. (2005) ProbIDtree: an automated software program capable of identify- ing multiple peptides from a single collision-induced dissociation spec- trum collected by a tandem mass spectrometer. Proteomics 5, 4096 – 4106 6. Bern, M., Finney, G., Hoopmann, M. R., Merrihew, G., Toth, M. J., and MacCoss, M. J. (2010) Deconvolution of mixture spectra from ion-trap data-independent-acquisition tandem mass spectrometry. Anal. Chem. 82, 833– 841 7. Michalski, A., Cox, J., and Mann, M. (2011) More than 100,000 detectable peptide species elute in single shotgun proteomics runs but the majority is inaccessible to data-dependent LC-MS/MS. J. Proteome Res. 10, 1785–1793 8. Cox, J., and Mann, M. (2008) MaxQuant enables high peptide identification rates, individualized p.p.b.-range mass accuracies and proteome-wide protein quantification. Nat. Biotechnol. 26, 1367–1372 9. Cox, J., Neuhauser, N., Michalski, A., Scheltema, R. A., Olsen, J. V., and Mann, M. (2011) Andromeda: a peptide search engine integrated into the MaxQuant environment. J. Proteome Res. 10, 1794 –1805 10. Giarratano, J. C., and Riley, G. (2005) Expert systems: principles and programming. PWS Pub. Co., Boston 11. Liao, S. H. (2005) Expert system methodologies and applications - a dec- ade review from 1995 to 2004. Expert Syst. Appl. 28, 93–103 12. Schroll, G., Duffield, A. M., Djerassi, C., Buchanan, B. G., Sutherland, G. L., Feigenbaum, E. A., and Lederberg, J. (1969) Applications of artificial intelligence for chemical inference. III. Aliphatic ethers diagnosed by their low-resolution mass spectra and nuclear magnetic resonance data. J. Am. Chem. Soc. 91, 7440 –7445 13. Russell, S. J., Norvig, P., and Davis, E. (2010) Artificial intelligence: a modern approach. Prentice Hall, Upper Saddle River, NJ 14. Olsen, J. V., Macek, B., Lange, O., Makarov, A., Horning, S., and Mann, M. (2007) Higher-energy C-trap dissociation for peptide modification anal- ysis. Nat. Methods 4, 709 –712 15. Bin, M., and Johnson, R. (2012) De novo sequencing and homology search- ing. Mol. Cell. Proteomics 11, O111.014902 16. Shevchenko, A., Tomas, H., Havlis, J., Olsen, J. V., and Mann, M. (2006) In-gel digestion for mass spectrometric characterization of proteins and proteomes. Nat. Protoc. 1, 2856 –2860 17. Olsen, J. V., Schwartz, J. C., Griep-Raming, J., Nielsen, M. L., Damoc, E., Denisov, E., Lange, O., Remes, P., Taylor, D., Splendore, M., Wouters, E. R., Senko, M., Makarov, A., Mann, M., and Horning, S. (2009) A dual pressure linear ion trap Orbitrap instrument with very high sequencing speed. Mol. Cell. Proteomics 8, 2759 –2769 18. Michalski, A., Damoc, E., Lange, O., Denisov, E., Nolting, D., Muller, M., Viner, R., Schwartz, J., Remes, P., Belford, M., Dunyach, J. J., Cox, J., Horning, S., Mann, M., and Makarov, A. (2012) Ultra high resolution linear ion trap Orbitrap mass spectrometer (Orbitrap Elite) facilitates top down LC MS/MS and versatile peptide fragmentation modes. Mol. Cell. Pro- teomics 11, 10.1074/mcp.O111.013698 19. Boersema, P. J., Mohammed, S., and Heck, A. J. (2009) Phosphopeptide fragmentation and analysis by mass spectrometry. J. Mass Spectrom. 44, 861– 878 20. Kelstrup, C. D., Hekmat, O., Francavilla, C., and Olsen, J. V. (2011) Pin- pointing phosphorylation sites: Quantitative filtering and a novel site- specific x-ion fragment. J. Proteome Res. 10, 2937–2948 21. Michalski, A., Damoc, E., Hauschild, J. P., Lange, O., Wieghaus, A., Ma- karov, A., Nagaraj, N., Cox, J., Mann, M., and Horning, S. (2011) Mass spectrometry-based proteomics using Q Exactive, a high-performance benchtop quadrupole Orbitrap mass spectrometer. Mol. Cell. Proteomics 10, 10.1074/mcp.M111.011015 Expert System for Annotation of MS/MS Spectra Molecular & Cellular Proteomics 11.11 1509 http://www.mcponline.org/cgi/content/full/M112.020271/DC1 
work_2y55lv27fnazbowsoz2jmkrwtu	----	Analysis of prepositions near and away from Fòrum de Recerca. Núm. 22/2017, p. 433-449 ISSN: 1139-5486. DOI: http://dx.doi.org/10.6035/ForumRecerca.2017.22.26 433 Analysis of prepositions: near and away from Frames of reference Nuria Flor Fabregat al081314@uji.es Nuria Flor Fabregat. Analysis of prepositions: near and away from. Frames of reference 434 I. Abstract Traditional strategies and procedures to learn a foreign language include the study of rules of grammar and doing exercises such as filling the gaps, repetition of words, drills, memorization of irregular verbs and sentences which may express usual expressions of everyday life. Even if the array of exercises is adequate, polysemy in prepositions causes difficulties in choosing the proper preposition conveying the meaning required by different contexts. Two prepositions of the horizontal axis (near and away from) are taken into consideration in this paper. Approaching the problem from the theory of polysemy and understanding, the use of these prepositions is explored along the dimensions of function, topology – which is the study of physical space–, and force dynamics – introduced in studies such as Navarro (1998)–, as well as the notion of frame of reference (Levinson, 2004). Then, the different senses and uses of these prepositions of the horizontal axis are systematized, explained and examples are used to illustrate the difficulties in learning a language and the doubts which students may have in some situations. Key words: horizontal directions, landmark, trajector, frames of reference, visual perception. II. Introduction Traditional approaches to the teaching of prepositions in English are often reduced to a series of rules and typical examples, but students often have doubts in applying these rules correctly in different contexts. Handbooks of foreign language usually present grammar through irregularities and expose students to mechanical exercises, and the memorization of phrases and paradigms. According to Langacker (2008), while norms are required to learn grammar, the importance of wonder and curiosity cannot be disregarded. Meanings may be elaborated and constructed with complex expressions such as phrases, clauses and sentences. Communication reflects the basic experience of moving, perceiving and acting on the world. Many students of English as a Foreign Language (EFL) do not find the use of prepositions to be an easy topic. A point in case is the connection between the English prepositions such as in, on and at and the Spanish preposition en. Since the English and Spanish systems overlap in the contents of space relationships in terms of prepositions like these, Spanish EFL students often find problems in Fòrum de Recerca. Núm. 22/2017, p. 433-449 ISSN: 1139-5486. DOI: http://dx.doi.org/10.6035/ForumRecerca.2017.22.26 435 learning their proper use in a context (Navarro, Campoy & Caballero 2001). In my view, as a student but especially recently as a teacher, I have noticed that students of a foreign language do not know which prepositions are the correct option when producing oral or written English (compositions, letters or emails). Although they know about the rules of grammar to follow when dealing with prepositions, they do not master the nature of prepositions, their main senses and contextual uses. Based on their own answers in an exercise on writing compositions, they use their intuition. Practice, therefore, seems the most effective way towards correction. This research focuses specifically on the horizontal directions, near and away from, and will explore their main uses and senses. A previous paper dealt with the kind of practice that students need to improve their understanding of the preposition on: should the learning of rules and drills prevail or should learning rely on a practical approach as a more natural way to learn space relationships by focusing on the nature and contextual uses of prepositions? Therefore, from the perspective of cognitive linguistics and as a continuation of the strand of research explored by other studies of the approach to prepositional polysemy developed by Navarro (1998), I will devise and test this learning approach as regards two prepositions pertaining to the visual point of view. More specifically, I will focus on two prepositions of horizontal directions, near and away from, including examples from dictionaries. The main objectives of my research together with a detailed description of the procedures followed to carry out this study can be found under the objectives subsection and the method section below. Before diving into these sections, and for the purpose of clarity, a brief description of the theoretical bases that support this approach will be provided. III. Theoretical background: frame of reference The notion of frames of reference may be understood as places, the location of objects which they occupy and as the places containing the objects. Aristotle mentioned an example, the conundrum in which the river is the frame of reference and a boat is moving with the river. In the cognitive paradigm, prepositions are considered as particles which relate two elements, these are called the trajector and the landmark, which are referred to as TR and LM in short (Langacker, 1987). The trajector is the most significant entity, it is usually situated before the preposition and it can be changed more easily from one place to another. However, the landmark is the Nuria Flor Fabregat. Analysis of prepositions: near and away from. Frames of reference 436 entity to which the trajector is related. It is situated after the preposition and it is the point of reference for the trajector. A variety of frames of reference can be used when reading some sentences. For instance: I lived near the school (Macmillan dictionary) or I will be away from home for two weeks (Merriam Webster dictionary). It is clear that the sense of distance depends on the landmark and the speaker. Sometimes there is an ambiguity because of the position of the object and the landmark. Nevertheless, a question must be asked by some speakers when position and spatial information are referring from a frame to another direction of movement. There are some distinctions of spatial frames of reference according to Levinson (2004): relative, absolute, regarding space as relations between objects, directions and relations between objects, deictic, intrinsic, which depends on the landmark or visual perception and as well as viewer-centred, object-centred and environment-centred which are centred on the speaker, the object or the environment respectively. When relative space is being considered, it refers to the egocentric coordinate system and when absolute space is being considered, it refers to non-egocentric systems (Kant (1991), as cited by Levinson (2004)). The distinction between egocentric and allocentric refers to the coordinate system centred within the subjective body frame and the second one centred within elsewhere, the geographic orientation which is not often specified. Then, it is related to body-centred and environment-centred frames of reference. As philosophers argued (Campbell 1993), the egocentric frame is joined with body-centred, a speaker and a body-schema in a spatial interaction. Another distinction is based on the theory of vision, in which the notions are viewer-centred and object-centred. As Levinson (2004) pointed out, this theory of vision is the process of the vision of a retinal image to the recognition of an object itself, that is from 2.5D sketch to a model of 3D as a structural description. This distinction is related to the linguistic distinction of deictic and intrinsic perspectives. Then, the deictic perspective would be the viewer- centred, while the intrinsic perspective would be the object-centred. Indeed, there are also notions of orientation, called orientation- bound and orientation-free. The first orientation refers to both absolute and relative frames, while the second one, which is orientation-free, refers to intrinsic frames. In fact, linguists have distinguished deictic and intrinsic frames of reference. There are three different interpretations of deictic and intrinsic in Table 2.I (Levinson 2004). The first one is speaker-centric and non-speaker centric (Levelt 1989). The second one is centred on any of the speech participants and non-centred (Levinson 1983). The third one refers to ternary and binary spatial relations (Levelt 1984, 1996). Fòrum de Recerca. Núm. 22/2017, p. 433-449 ISSN: 1139-5486. DOI: http://dx.doi.org/10.6035/ForumRecerca.2017.22.26 437 Carlson-Radvansky and Irwin (1993: 224, quoted in Levinson 2004: 32) explained frames for spatial relationships among objects. In a viewer-centred frame, objects are represented in a retinocentric, head-centric or body-centric coordinate system based on the perceiver’s perspective of the world. In an object- centred frame, objects are coded with respect to their intrinsic axes. In an environment- centred frame, objects are represented with respect to salient features of the environment [...]. In order to talk about space, vertical and horizontal coordinate axes must be oriented to one of these reference frames so that linguistic spatial terms such as «above» and «to the left of» can be assigned. Thus, the notions of deictic, intrinsic and extrinsic are related to the corresponding linguistic interpretations of viewer-centred, object-centred and environment-centred frames of reference. So, in a spatial representation system, the frame of reference is accepted for a coordination of perception and language. For instance, it is clear that egocentric refers to relative or viewer-centred, 2.5D sketch refers to deictic frame, intrinsic corresponds to object-centred or 3D model, absolute corresponds to environment-centred. Regarding frames of reference in a linguistic view, the notion absolute as the frame of reference is used by many languages such as fixed bearings (West, North). Otherwise some European languages would use the notion of relative or viewer-centred such as left. Essentially, the frames of reference absolute, relative and intrinsic are the main ones in order to describe the horizontal spatial directions. For instance, according to Levinson (2004), some sentences are viewed as a deictic frame: 1. The ball is in front of me. The ball is in front of the tree. (Levinson, 2004). (As speaker-centric.) Then, regarding the prepositions near and away from: 2. I lived near the school (Macmillan dictionary) or I will be away from home for two weeks (Merriam Webster dictionary). (As speaker-centric.) Other sentences are viewed as an intrinsic frame: 3. The ball is in front of the chair. (Levinson, 2004). (As non- speaker-centric, the chair.) 4. The ball is in front of you. (Levinson, 2004). (As non- speaker-centric, the addressee.) 5. The ball is to the right of the lamp, from your point of view. (Levinson, 2004). (As non-speaker-centric, the addressee.) Then, regarding prepositions near and away from: Nuria Flor Fabregat. Analysis of prepositions: near and away from. Frames of reference 438 6. They are near to solving the puzzle. Keep away from the stove – it’s very hot (Macmillan). (As non-speaker-centric, the addressee.) Coordinate systems are frames of reference and, in language, frames can be distinguished according to origins such as the speaker, the addressee, etc. Table. 1 Classification of frames of reference INTRINSIC ABSOLUTE RELATIVE Origin ≠ ego Origin ≠ ego Origin = ego Object-centred Intrinsic-perspective 3D Model Environment-centred Viewer-centred Deictic-perspective 2.5 D sketch Allocentric Egocentric Orientation-free Orientation-bound 2.1. Linguistic categories of prepositions and the three dimensions of perception In cognitive linguistics, there is some research about space semantics which is likely to arrange an order, despite the difficulty of the concepts, by means of the application of a system with radial networks, in which each prepositional or adverbial sense is located on a node in accordance with its centrality within the network. More literal (spatial or physical) meanings tend to be associated with most focuses in more central senses, from which abstract meanings result from the help of metaphoric and metonymic processes (Navarro, 1998). For example: the bottle is on the table. The bottle is the TR which can be moved easily and its resting side falls across the LM, in this case, the table, which can work as a supporting point for the TR. Another example: A group of students were standing near the entrance. The students are the TR which can be moved easily and its resting side falls across the LM, in this case, the entrance, which can work as a building for the TR. Thus, prepositions, in cognitive linguistics, are considered as «linguistic categories» themselves surrounding a series of elements such as meanings or senses arranged in the structure of a radial category, with a prototype and peripheral members; these are, therefore, polysemic elements with several meanings. Fòrum de Recerca. Núm. 22/2017, p. 433-449 ISSN: 1139-5486. DOI: http://dx.doi.org/10.6035/ForumRecerca.2017.22.26 439 Among the range of research on prepositions from a cognitive linguistic perspective, Navarro’s model (1998, 2006) provides a fully- fledged model, which is a model completely developed, for the semantic representation of prepositions whose senses are derived and arranged in terms of three semantic dimensions of perceptual space or aspects of construal. Therefore, three dimensions that can help in determining the spatial relationship established between the two entities (trajector and landmark) as mentioned above, in human conceptualization, specifically: 1. Topology: The visual perception of objects gives the speaker clues for establishing and conceptualizing topological relations like coincidence, contact, inclusion, proximity, and the like. 2. Force-dynamics: Human beings have experience of self-motion and object motion, which provides the clues for conceptualizing patterns of interaction in terms of dynamics. 3. Function: Human beings have experience of the effects of interaction, as well as the consequences of those effects for survival and well-being. (Navarro, 2006: 171). Regarding the specialisation of meaning, metonymic and metaphoric extensions can be shown in senses. These senses can be detailed by profiling certain aspects of the conceptual schema. In force-dynamic configuration senses, the interaction axis between trajector and landmark is seen as the central aspect of the relation. Though still present, other aspects like the topological relation of contiguity and the functional orientation remain in the background. The direction of the movement is also determined by the trajector’s functional front and by the landmark’s accessible zone as well. The context for the sense of search for contiguity, which is derived from the central force-dynamic senses, requires motion verbs and other dynamic expressions like come, fly, go, run, swing, make, jump, dive and so on. In topological configuration senses, the topological relation of contiguity is prevalent over force-dynamic or functional aspects. The sense of coincidence reflects a «coincidence» between the trajector and the landmark. Some words such as location, place or point are frequent in this sense. The trajector is something understood as if it is attached to a part of an entity and the landmark designates that part. Some words are used in this sense like beginning, top, bottom, middle, centre, head, edge and so on. In functional configuration senses, from the conceptual image schema the functional space can be described with the relation of functionality itself. In the background it is remained the force- Nuria Flor Fabregat. Analysis of prepositions: near and away from. Frames of reference 440 dynamic dimension and the topological relationship. Here places are designated for the landmarks where people usually do certain activities or they participate in certain events. Landmarks are thus often buildings or public spaces. Then, trajectors are people who control or use them in a certain way with relation to them. Trajectors may also be concepts which realise these activities that are carried out by those people. (Navarro, 1998) 2.2. Vandeloise’s spatial relations The object to be located is called the trajector (Langacker), and this author uses the corresponding reference point, the landmark. In this case, Vandeloise (1991) refers to the object to be located as the target and to the object of reference as the landmark. In well-formed utterances, the target always corresponds with the subject of the relation, and the landmark coincides with its object. The linguistic principle may be expressed as follows: • subject of spatial relation = target • object of spatial relation = landmark What are the characteristics of target and landmark? It should be pointed out that the position of the target constitutes new information, while the position of landmark states known information. Although the target is difficult or small to perceive, generally the landmark is large and easy to be distinguished. Also the target is mobile, while the landmark is immobile and stable. For instance, look at the falling star! Near the church tower or look at the church tower! Near the falling star. The falling star, momentary and brief, is drawing the speaker’s attention. It seems as it is the ideal target, while the church tower, immobile and immense, shows the characteristics of the usual landmark. In contrast the second sentence is uncommon. Another example: the bus stop is near the house or the house is near the bus stop. In this contrast it is seen the reason of the sense from an element that is not understood explicitly. The pedestrians’ path is between the house and the bus stop. Here the speaker is recognised with the landmark. The first sentence may be understood by imagining a principal path from the house to the bus stop, whereas the second sentence proposes the opposite course. Indeed both target and landmark are immobile in both examples, and the degree of being acceptable is justified by the speaker who may be moved along the path possibly. When describing the scene, the speaker may choose among several different strategies for situating the landmark. When a non- egocentric landmark is eliminated, there is one strategy between not introducing an egocentric landmark and expressing a landmark with the correspondent preposition. Thus, the landmark may be omitted when the speaker is sufficiently identified with the landmark (e.g., Fòrum de Recerca. Núm. 22/2017, p. 433-449 ISSN: 1139-5486. DOI: http://dx.doi.org/10.6035/ForumRecerca.2017.22.26 441 Saint-Cloud).The degree to which the speaker locates the landmark determines the situation in which this landmark is expressed or not. Then, the landmark refers to the fact of the virtual position of the speaker. Thus, the speaker could change a variety of points of view, and a conversation may continue as long as the addressee is able to understand the situation and the movement of the speaker in such a case. The unexpressed landmark often identifies the speaker’s position, but this position may be either real or virtual. Concerning the expressions near and away from, what the speaker has to carry out is to move to the place of the landmark itself. These expressions near and away from are described as directions of distance between target and landmark. In terms of distance, it is considered to a certain norm which depends on the movement to be approached of the target/landmark and to the landmark/target. The expression près de (near) is reduced to spatial and temporal domains, whereas the other expression proche (close to) de suggests proximity in every domain. Thus, the expression near would refer to smaller distance than away from. For instance: Jupiter is near Saturn. The electron is far from its nucleus. In the first sentence near is identified as a larger distance between target and landmark than far from in the second example. So, the principal characteristic of this distance as a norm relate to the accessibility of the target/landmark to the landmark/target. According to Vandeloise (1991), this norm of distance depends on the trajector and the relation with the landmark as well as the dimension of the landmark, the speed of the target, the speed of the landmark, the size of the speaker, the speed of the speaker, the size and the speed of the addressee, the facility of access and types of access too. Regarding the dimension of the landmark, infrequently the landmark is smaller than the target. The norm may increase in proportion to the landmark. The distance between the target Jupiter and the landmark may be greater in the following example, Jupiter is near the Milky Way than in the example of Jupiter is near Saturn. Indeed, the speed of the target is also relevant, since the norm may increase with the speed of the target when the target is moving towards the landmark. For instance: the tortoise is far from the lake or the antelope is far from the lake. In this case the lake is seen further from the antelope than from the tortoise. However, when the target is moving away from the landmark, the extent of the norm will diminish as speed decreases. In other words, when the speed of the target is gathering with the landmark in an easier or more difficult way, the normal distance increases or decreases, respectively. Nuria Flor Fabregat. Analysis of prepositions: near and away from. Frames of reference 442 Thus, the speed of the landmark is not seen as a common context when the landmark is mobile. However, there are some sentences in which the distance increases with the speed of the landmark. For instance: the man is far from the helicopter. Here the landmark is moving from the target. The fox is near the rabbit. Here it depends on the speed of the landmark. Another case is the size of the speaker, which could be near when is a father or far when is a child and it depends on the age. Moreover, the speed of the speaker varies when the speaker is driving a car or walking. Also the size and the speed of the addressee could vary when the addressee is a hiker and then it is near –for example a farm from a village–, or if this is lame. Finally, the facility of access may be different when the path is easy or difficult to access and the distance increases or diminishes. For example, the red house is far. Here the speaker is walking up and it is far. The yellow house is near. When the speaker is walking down and it is near in such case. There are types of access such as the visual access and the physical access. For example, if the speaker sees the mountain from the hotel window, it may be near, or if the speaker wants to hike there, it may be far. Another situation is when a sailboat may be far from the visual access of the eyes, but it may be near through binocular. There is a value of distance which is changed by the types of access of the speaker or target. The access to the meeting point and the factors of making it easier or not easier between the target and the landmark play an important role. The main factors are pointed out below. Table 2. Classification of accessibility Accessibility Relative speed of target and landmark Distance Type of access The temporal sense is also used with the preposition near, which indicates the spatial reference. For example, it is near Christmas. In the domains of color terminology the expression proche de (close to) is preferred over près de (near). An example of this sense: mauve is close to blue. Thus, isn’t it more direct to define the prepositions near and away from concerning the accessibility and inaccessibility? Then, the distance itself would be one of several possible factors which may affect access. Fòrum de Recerca. Núm. 22/2017, p. 433-449 ISSN: 1139-5486. DOI: http://dx.doi.org/10.6035/ForumRecerca.2017.22.26 443 IV. Research objectives The main objective of this research is to analyse and compare the two prepositions near and away from in order to systematise the main uses and illustrate them with some sentences. Several authors’ explanations and their background knowledge are detailed too. Some prepositions express where something is in physical relation to another thing: Example: There was a bird near the tree. There was a bird away from the tree. About these examples, the following can be said: • A bird is the trajector in the relationship expressed by the preposition. • The tree is the landmark of the preposition. • The preposition tells us where the trajector is in relation to the landmark. • Near / away from are prepositions of distance. Also, because both the subject and the landmark are concrete things, we can say that near/ away from are being used literally. Although there are less than one hundred English prepositions, they do not take ending positions, that is to say they are not usually written in the end of a sentence, and even if the structure of most prepositional phrases is usually easy, the use of English prepositions is complex. Then, most prepositions have more than one meaning, as described, for example, in the prepositional approaches of polysemy in Navarro’s model. As it is well known, English prepositions play an important role in sentences of many key notions, those pertaining to physical objects in their visual perceptions, arrangements, orientations and so forth. V. Methodology The methodology of this research is to look for examples of the prepositions near and away from on the following dictionaries to identify their main uses: • Cambridge dictionary • Merriam Webster • Macmillan • English Oxford dictionary After comparing the meaning of these prepositions, some examples are written to illustrate their uses. Then, each sentence and the preposition is analysed regarding the frames of reference, the three dimensions, the spatial relations and the two concepts which are trajector and landmark. Nuria Flor Fabregat. Analysis of prepositions: near and away from. Frames of reference 444 In order to understand the tree structures of sentences, some concepts of analysis will first be made clear. In traditional grammar and modern syntactic theory, the notion constituency is an essential construct. It is described by linguists as assigned fixed hierarchical structures which are considered as inverted «trees» metaphorically. Then, details vary and styles may change, but an example of the usual format is the nominal or noun phrase as a table near the door (noun phrase- article and noun, prepositional phrase- preposition, article and noun). Three kinds of information can be found in syntactic tree structures. Grammatical category (N, P, NP...), linear order (left to right on the page) and constituency (hierarchical grouping). Prime examples are the subject and object relations. Thus, a subject is designated as a nominal whose profile corresponds to the trajector of a profile relationship. An object is characterised as a landmark of a profile relationship. A complex activity must be called as the act of talking and something which people do rather than whatever they have must be viewed as a language. Some aspects of this activity are motor, perceptual and mental, which are established by the procedure in brain, that is to say in a wide sense a cognitive activity is talking. The knowledge of a language is a situation of controlling skills in distinct contexts. Some regions of the brain are involved in language and the processing activity constitutes linguistic units. Thus, these units are not separated or independent, but they overlap with other units or even add them as components. In general, units embody the rules and regularities of a language. Thus, linguists may consider rules in one of the three ways which are as constructive rules, as filters or as schemas. In cognitive grammar, rules take the form of schemas. These units are connected by relationships of categorization and they can form networks of any size. (Langacker, 2008) VI. Results and discussion According to Merriam Webster, Macmillan, Cambridge dictionary, and English Oxford dictionary near can be used in the following ways: 1) as a preposition (close to someone or something): A group of students were standing near the entrance. I lived near the school. I’ll write and let you know nearer the time (Macmillan). Beaches near the city (Merriam Webster). 2) as an adverb (not far away in distance): Come nearer, and I’ll tell you the whole story (Macmillan). Is there a restaurant near here? (Cambridge) Fòrum de Recerca. Núm. 22/2017, p. 433-449 ISSN: 1139-5486. DOI: http://dx.doi.org/10.6035/ForumRecerca.2017.22.26 445 3) as an adjective: I went into the nearest room. A climb in the mountains led to near disaster (Macmillan). In the near future (not far distant in time (Merriam Webster). 4) in the preposition phrase near to: Pull your chair nearer to the table (Macmillan). a) getting close to a particular state or situation: (near to) Julian was near to panic as he suddenly realized that he was trapped (Macmillan). People near to retirement need to know their pension funds are sufficient (Macmillan). b) near to doing something: They are near to solving the puzzle (Macmillan). According to Merriam Webster, Macmillan, English Oxford dictionary and Cambridge dictionary, away from can be used in the following ways: 1) Somewhere else: somewhere else, or to or in a different place, position, or situation: as an adverb. Ms Watson is away on holiday until the end of the week (Cambridge). Keep/Stay away from him (Cambridge). a) Distant: at a distance (of or from here): as an adverb How far away is the station? (Cambridge). The office is a half-hour drive away (Cambridge). We live five kilometres away from each other (Cambridge). Keep away from the stove – it’s very hot (Macmillan). b) far from people, places, or things, especially so that you feel separated from them. I will be away from home for two weeks (Merriam Webster). It’s nice to have a weekend away from London (Macmillan). She’s been away from her family for too long (Macmillan). 2) Away as an adjective: - (of a sports fixture) played at the opponents’ ground: Tomorrow night’s away game at Leicester (English Oxford dictionary). An away victory (English Oxford dictionary). 3) Relating to or denoting a sports team that is playing at the opponents’ ground: The away side scored first (English Oxford dictionary). Away fans chanted and cheered (English Oxford dictionary). In this section, some examples of the preposition near are described with the purpose of understanding the use of frames of reference: - A group of students were standing near the entrance (Macmillan). The students are the trajector (TR) which can be moved easily and the landmark (LM) is situated after the preposition and it is the Nuria Flor Fabregat. Analysis of prepositions: near and away from. Frames of reference 446 point of reference for the trajector, in this case, the entrance, which can work as a building for the TR. The frame of reference is intrinsic which corresponds to object- centred, since it depends on the accessibility from the trajector to the landmark. The dimensions are topology because of the proximity and function when there is an interaction. The spatial relations of Vandeloise are the type of access and the facility of access to this entrance as a building and the speed of the target (a group of students). - I lived near the school (Macmillan). The speaker is the TR and the subject and the school is the LM and the place where the speaker is situated. The frames of reference are absolute which corresponds to environment-centred and relative which corresponds to viewer-centred, since it depends on the type of access. The dimensions are topology because of the proximity and function when there is an interaction and accessibility. The spatial relations of Vandeloise are the type of access and the facility of access of the school and the speed of the target as a speaker. - They are near to solving the puzzle (Macmillan). The subject, which is they, is the TR and the puzzle is the LM as the object which they try to solve. This preposition means near to doing something. The Frames of reference are intrinsic which corresponds to object-centred and relative which corresponds to viewer-centred. The dimensions are topology (metaphoric sense) because of the proximity as a temporal situation and function when there is an interaction and dynamic when there is a movement. The spatial relations of Vandeloise are the speed of the target as a subject of this sentence to solve the puzzle and the dimension of the landmark (the puzzle). Some examples of the preposition away from are the following: - Keep away from the stove – it’s very hot (Macmillan). The viewer is the TR and the stove is the LM as the object to be away from. The frames of reference are intrinsic which corresponds to object-centred and relative which corresponds to viewer-centred. The dimensions are topology because of the distant and function when there is an interaction and dynamic when the viewer has to keep away. The spatial relations of Vandeloise are the speed of the speaker to keep away from the stove and the dimension of the landmark (the stove). - I will be away from home for two weeks (Merriam Webster). The speaker is the TR and the subject and the home is the LM as the building. The frames of reference are absolute which Fòrum de Recerca. Núm. 22/2017, p. 433-449 ISSN: 1139-5486. DOI: http://dx.doi.org/10.6035/ForumRecerca.2017.22.26 447 corresponds to environment-centred and relative which corresponds to viewer-centred. The dimensions are topology because of the distant and function when there is an interaction. The spatial relations of Vandeloise are the type of access and the facility of access and the space of time of two weeks. - It’s nice to have a weekend away from London (Macmillan). The speaker is the TR as an opinion of the speaker and the city itself is the LM where the speaker is situated. The frames of reference is absolute which corresponds to environment-centred, since it depends on the situation of the city on the map. The dimensions are topology because of the distance and function when there is an interaction. The spatial relations of Vandeloise are the facility of access and the type of access and the speed of the target as the subject of this sentence and as well as the dimension of the landmark (the city). VII. Conclusions In grammatical structure, prepositions relate two elements and these are considered as nouns. These elements are called the trajector and the landmark, in linguistics, which are referred to as TR and LM (Langacker, 1987). Briefly, the trajector is usually situated before the preposition, it is the most meaningful concept and it can be changed more easily from one place to another. However, the landmark is situated after the preposition and it is the concept to which the trajector is related. Then, this is the point of reference for the trajector. Regarding the English language as a foreign language, this subject is not an easy one for many students and they encounter great difficulty as they learn English prepositions in this field of knowledge and study. A case in point is the connection between the English prepositions near and away from and the Spanish prepositions cerca and lejos. The frames of reference absolute, relative and intrinsic are the main ones in order to describe the horizontal spatial directions which are the directions of the prepositions near and away from in this research. The following three terms are synthesised above (Levinson, 2004): • Absolute which corresponds to environment-centred • Relative which corresponds to viewer-centred. • Intrinsic which corresponds to object-centred In these spatial directions, the usual dimensions are topology when there is a proximity, contact or inclusion and function when there is an interaction and some effects between the trajector and the landmark. Nuria Flor Fabregat. Analysis of prepositions: near and away from. Frames of reference 448 The norm of distance, which Vandeloise (1991) explained, depends on the trajector and the relation with the landmark. These two terms (TR and LM) are determined by the dimension of the landmark, the speed of the target or the subject of the sentence, the speed of the landmark or the object of the sentence, the size of the speaker, the speed of the speaker, the size and the speed of the addressee, the facility of access and types of access regarding the building or the place where the speaker is situated. Therefore, some questions should be asked to consider: what are the two elements which prepositions relate?, how are the three dimensions called in spatial directions?, what are the main frames of reference in the horizontal directions? VIII. References Langacker, Ronald W. 2008. Cognitive Grammar. A basic introduction. Oxford: Oxford University Press. Levinson, Stephen C. 2004. Space in language and cognition. Explorations in cognitive diversity. Cambridge: Cambridge University Press. Navarro, Ignasi. 1998. A Cognitive Semantics Analysis of the Lexical Units AT, ON, and IN in English. Ph.D. dissertation. Castelló de la Plana: Publicacions de la Universitat Jaume I de Castelló. Navarro, Ignasi, Campoy, M. Carmen & Caballero, Rosario. 2001. «Thinking with English Prepositions and Adverbs». In Docència Universitària: Avanços Recents. Primera Jornada de Millora Educativa de la Universitat Jaume I, 243-252. Castelló: Publicacions de la Universitat Jaume I de Castelló. Vandeloise, Claude. 1991. Spatial prepositions. A case study from French. Chicago: University of Chicago Press. Web pages https://dictionary.cambridge.org/es/ https://www.merriam-webster.com/ https://www.macmillandictionary.com/ https://www.oxforddictionaries.com/ Images https://www.shutterstock.com/video/clip-17171377-stock-footage- lausanne-switzerland-circa-january-students-outside-college- building-university-students.html https://www.pinterest.es/pin/126452702008017071/ Fòrum de Recerca. Núm. 22/2017, p. 433-449 ISSN: 1139-5486. DOI: http://dx.doi.org/10.6035/ForumRecerca.2017.22.26 449 https://www.123rf.com/photo_11248579_illustration-of-a-cartoon- house-inside-rounded-landscape.html https://www.123rf.com/photo_12107090_illustration-of-kids- solving-a-giant-jigsaw-puzzle.html https://www.pinterest.es/explore/map-of-london-city/ 
work_2yanabvlyjeqnphthvu7dvtlvi	----	Fuzzy expert system for the detection of episodes of poor water quality through continuous measurement Cecilio Anguloa,∗, Joan Cabestanya, Pablo Rodŕıguezb, Montserrat Batlleb, Antonio Gonzálezb, Sergio de Camposb aUniversitat Politècnica de Catalunya, UPC-Barcelona Tech, Vilanova i la Geltrú, Spain bAdasa Sistemas S.A.U., Barcelona, Spain Abstract In order to prevent and reduce water pollution, promote a sustainable use, protect the environment and enhance the status of aquatic ecosystems, this article deals with the application of advanced mathematical techniques de- signed to aid in the management of records of different water quality moni- toring networks. These studies include the development of a software tool for decision support, based on the application of fuzzy logic techniques, which can indicate water quality episodes from the behaviour of variables mea- sured at continuous automatic water control networks. Using a few physical- chemical variables recorded continuously, the expert system is able to obtain water quality phenomena indicators, which can be associated, with a high probability of cause-effect relationship, with human pressure on the water environment, such as urban discharges or diffuse agricultural pollution. In this sense, at the proposed expert system, automatic water quality control networks complement manual sampling of official administrative networks and laboratory analysis, providing information related to specific events (dis- charges) or continuous processes (eutrophication, fish risk) which can hardly be detected by discrete sampling. Keywords: Water quality system, Automated measurement networks, Fuzzy logic, Fuzzy inference system, Guadiana river ∗Neàpolis Building, Rambla de l’Exposició 62, 08800 Vilanova i la Geltrú, Spain. Tel: +34 938967273; Fax: +34 938967200 Email address: cecilio.angulo@upc.edu (Cecilio Angulo) Preprint submitted to Expert Systems with Applications June 7, 2010 1. Introduction On October 23th, 2000, the “Directive 2000/60/EC of the European Par- liament and of the Council establishing a framework for the Community action in the field of water policy” or, in short, the EU Water Framework Directive (European Parliament, 2000) was finally adopted. The Directive is about the organization of water management, in order to prevent and re- duce pollution, promote sustainable water use, protect the environment and enhance the status of aquatic ecosystems. As a starting point in the practice of the policy in Spain, tasks for the characterization of anthropogenic impact on the waters were carried out, as well as the diagnosis of their condition based on monitored indicators. Application of the Directive led to the execution of an analysis of the char- acteristics of each river basin, and a study of the impact of human activity on the water. Based on the results of such initial analysis as well as determining fac- tors such as cost-benefit analysis, citizen participation and others, each river basin’s management unit should be able to develop a management plan and a program of measures. The objectives of this plan would prevent deteriora- tion, enhance and restore the status of surface water bodies, ensuring that they are in good ecological and chemical status, and reduce pollution due to discharges and emissions of hazardous substances. Moreover, the deployment of the stations of the Automated Information Water Quality System (SAICA, in Spanish) has allowed to obtain automated and continuous information from the state of rivers (Serramià, 2005). Auto- mated information is different from those usually collected through manual measurements. Unlike the latter, which are timely and accurate measures, the SAICA ones present the advantages of immediacy and continuity. How- ever, they involve the management of large quantities of records spread across multiple variables that are supposed to be less precise. This article presents work developed through the Spanish project ECOW- ATCH in the context of implementing the EU Water Framework Directive. It deals with the application of advanced mathematical techniques designed to aid in the management of records from different observational networks of water quality. Studies and developments performed in the project ECOW- ATCH include the development of a software tool for decision support, based on the application of fuzzy logic techniques, which can detect quality episodes from the behavior of continuous variables measured in SAICA networks. 2 The developed expert system is able to generate, in real time, a set of combined indicators of water quality based on: • data from SAICA automated networks • expert knowledge about episodes of poor water quality expressed as rules whose purpose is to detect two different types of events: • ad hoc and spurious events, like urban discharges or caused by a waste water treatment plant (WWTP), • long and cumulative over time events, like episodes of eutrophication and, as a novelty, fish risk, i.e. risky environmental conditions for fish communities The working assumption along this project is that expert definition of continuous indicators will provide information on the evolution of the water quality of a river basin allowing early detection of abnormal events. These indicators will anticipate and complete the results obtained with hand-made programmed sampling. In addition, this form of dual management: • will increase knowledge about the river basin, by summarizing in syn- thetic values the information contained in a set of complex variables • justifies the operation and maintenance costs of continuous automatic monitoring networks, currently used for warning and inspection pur- poses Automatic monitoring networks will supply an added value as knowledge generators about the environment and can help assess the effectiveness and appropriateness of the actions taking place on the basin or specific parts of the same: introducing a management system, construction of a water treatment infrastructure or others. Based on the working hypothesis and with the aim of combining expert knowledge with information collected by SAICA network, the methodology depicted in Figure 1 has been used as project development script. As shown therein, the generation of indicators by the expert system is based on the expression of empirical expert rules in the form of fuzzy logic. Such expert fuzzy rules are constructed from a set of examples that allow the formulation 3 of the phenomena by the experts. The acquisition of this knowledge through the use of fuzzy logic will allow to design an expert system able to generalize to new situations. Figure 1: Development process for setting-up fuzzy logic rules. 2. Selecting poor water quality episodes The Spanish side of the Guadiana river basin has been chosen as the study area for this work, due to the availability of information records collected by several automatic measurement stations belonging to the Guadiana Hydro- graphic Confederation’s SAICA network. Figure 2(a) shows the location of the basin and the selected stations are depicted in Figure 2(b). 2.1. Episodes and related variables The expert system to be developed will focus on the detection of three types of episodes: (i) discharges to the environment, (ii) eutrophication and, experimentally, (iii) fish risk. The determination of quality variables to be used as inputs for each kind of episode to assess and their combined impact has been completed with the collaboration of experts, who have characterized a series of remarkable events, which have been used to calibrate the developed methods. 4 (a) Guadiana River basin. (b) Stations in the Guadiana River. Figure 2: Location of selected stations in the Guadiana River. The five variables extracted from expert knowledge to characterize episodes have been: pH, conductivity, turbidity, ammonium concentration and dis- solved oxygen (DO) concentration. Relevant cases corresponding to the studied episodes should be selected for training and development of automated fuzzy inference systems. In the following, episodes characterizing each type of analyzed event are discussed, along with a detailed description in terms of the consulted experts. 2.2. Episodes of Urban or WWTP discharge The two episodes of water discharge shown below correspond to data from the Guadajira automatic station, located downstream of a WWTP. 2.2.1. Episode 1. Guadajira, January 2009 The values of the variables selected by the experts for the duration of the episode are shown in Figure 3. Two different events can be distinguished, one of them is starting on January 22th, and another at the beginning of the 27th, marked with two separate rectangles. An explanation for the first event is a possible malfunction in the process of nitrification - denitrification of the upstream wastewater treatment plant, which would justify the increase of ammonium, as well as in the phosphate removal, while there is a decrease in the dissolved oxygen concentration. For the second event, aeration might not work properly in the WWTP, so no removal of phosphorus and ammonium exists, while there is an increase in 5 Figure 3: Measured variables for Episode 1 (discharge). organic matter, which is reflected in the decrease of dissolved oxygen. Both cases are similar, the second one having a largest pollution load than the first, although it could also be a direct discharge. This kind of situation is not unusual in this particular WWTP. A planned action exists for this specific WWTP in order to enlarge it to provide sufficient treatment capabilities for current needs. The proposed expert system would serve, in this case, as a tool for assessing the effectiveness and appropriateness of the actions currently being undertaken at the WWTP. 2.2.2. Episode 2. Guadajira, September 2008 In Figure 4 the values of the variables during the analyzed event are depicted. It can be seen that the event starts in the evening of September 22th, as marked with a rectangle. The increased presence of ammonium indicates that the process of nitri- fication - denitrification has not been completed, therefore the cause is in the WWTP. Moreover, there has neither been degradation of organic mat- ter before the discharge, since values for the dissolved oxygen concentration become null and turbidity increases due to the presence of organic matter. 6 Figure 4: Measured variables for Episode 2 (discharge). This episode is related to a direct urban discharge, from either the WWTP itself, or from some of the olive-related industries located in the area that are not networked to the sewerage system which serves the WWTP, and that are in the phase of requirements for implementation of appropriate treatment systems. 2.3. Episode of Eutrophication A process of eutrophication is an increase in the trophic status of water caused by nutrient enrichment (Smith et al., 1999). In the limit, the cumula- tive process of eutrophication can lead to a biological impact as it comes to situations of partial (night) or permanent anoxia (no oxygen), which could end in itself with fish life. The selected episode to be used for the expert system is shown in the next section. 2.3.1. Episode 3. Benavides, March 2007 A process of natural photosynthesis is going on: daytime primary pro- ducers consume CO2 and produce oxygen, thus decreasing H + concentration, and therefore pH increases. At night CO2 concentration increases, increasing H+, and low pH. As the process continues, the possible presence of N and P nutrients (not measured) causes biomass growth and increases production 7 Figure 5: Measured variables for Episode 3 (eutrophication and fish risk). / consumption of oxygen, which is manifested in the amplitude of oscillation in the daily cycle of this element. While N and P values are in excess the process will increase. From an initial state of eutrophication, low flow does not favor dilution, and the increase of conductivity may indicate a concentration of salts and increased nutrient concentrations (nitrates, phosphates...) that are a limit- ing factor for algae and macrophytes, favoring the uncontrolled growth of biomass. The system is no longer sustainable as the amount of biomass to degrade exceeds the system’s capacity, hence oxygen concentration de- creases until eventual anoxia situations. Values for quality variables during the episode are shown in Figure 5. 2.4. Episode of Fish risk According to the experience of operation managers, during the advanced stage of eutrophication an increase in average pH is observed. In these cir- cumstances, if a discharge of ammonium in a basic medium exists, it is trans- formed into ammonia, a highly toxic product and lethal to fish fauna, which was already under pressure due to periods of partial anoxia. Information about fish risk is useful for operation managers as an aid to 8 decision making for preventive action. In entering a risk area, when there is a capacity to act, dilution can be forced by releasing water from reservoirs to help breaking the cycle of eutrophication. An episode from a water station on the River Guadiana has been selected and is shown in the next section. 2.4.1. Episode 3. Benavides, March 2007 Ammonium (NH+4 ) is in the water in steady equilibrium with ammonia (NH3), highly toxic (NH3 +H + ↔ NH+4 ). Faced with a discharge of ammo- nium in the river, a small increase in pH (H+ decrease) moves the equilibrium towards the formation of NH3. While ammonium can be easily assimilated, ammonia is lethal to living beings. The situation can be particularly danger- ous on this zone of the river, which consists on a wide and clear channel with shallow waters, when eutrophication occurs. Both contributing factors lead to a strong proliferation of phytoplankton, which during the daily cycle cause significant variations in pH. During an episode with high levels of ammonium that can not be neutralized by oxidation, NH3 lethal concentrations to the fish stock can be reached. In this episode, the average pH values grow up close to 9, so a small discharge of ammonium is dangerous to the fish fauna, and a large one, lethal. Figure 5 shows the value for quality variables during the episode. 3. Fuzzy inference systems developed for the expert system The set of examples of abnormal events to be used for the definition and development of the expert system have been established. The next stage in the methodology is the automation, by the working group, of the steps to generate the expert system (see Figure 1). A first step, which consumes a significant amount of time if it is not well automated, is the inspection and validation of the database provided by the SAICA network’s stations. In particular, measurement errors for out of range and non-present values in some variables should be considered. Standardized solutions have been adopted: absent values have been filled by linear regression and outliers have been erased and replaced with mean data from previous and consequent samples. Experts have been paying particular attention to some little samples difficult to fix automatically. 3.1. Description Fuzzy inference systems are a widespread method in the treatment of information for the generation of expert systems in monitoring the water 9 quality (Nasiri et al., 2007; Lermontov et al., 2009; Hatzikos et al., 2009). Roughly speaking, a fuzzy inference system is an algorithm able to convert a strategy or set of linguistic rules in an automated strategy. Applications can be designed with fuzzy logic so that systems and processes respond with greater knowledge to imprecision, which seeks to mimic human behavior. Our approach is focused on the automated generation of fuzzy rules from recorded data according to a pre-defined structure of rules or strategy de- termined by human experts. Once this strategy is provided, some questions must be answered for the generation of the automated inference system. In particular: 1. How are membership functions for both input and output variables set-up? 2. Which labels for the output variable are assigned for each rule? 3. How are linguistic labels determined for the defuzzified output? Before developing the automated system, effectiveness of using fuzzy in- ference systems was firstly validated by manually creating an expert system prototype, which was tested on the two episodes of urban discharge in Guada- jira’s SAICA station, in January 2009 and September 2008. As a result, for the first of the episodes, a set of 54 basic rules was designed. One example of them is following: 54. IF (IncConduct is Negative) and (IncOD is Zero) and (IncAmmonium is Negative) and (IncPhosphates is Negative) THEN (UrbanDischarge is Normal) For the second example, turbidity measures are further provided, so the initial knowledge base is completed with the information of whether or not there is increased turbidity. This resulted in 162 rules. One example is following: 1. IF (IncConduct is Positive) and (IncOD is Negative) and (IncTurbidity is Positive) and (IncAmmonium is Positive) and (mOD is Low) THEN (UrbanDischarge is Evidences) After positive testing of manually generated fuzzy inference systems for these two examples, it was time to generalize, i.e. to develop an automated procedure to automatically obtain fuzzy rules according to provided data. 10 3.2. General Application The continuous data obtained through automated networks in the Guadi- ana river Hydrographic Confederation (CHG) have allowed characterization of different episodes in its basin, which experts have sought to associate the possible cause of. As an instance, in Figure 3 the expert has been able to find out two episodes of increased ammonium matching a decrease in oxy- gen, a slight decrease in pH and a slight increase in conductivity. In turn, a discharge of organic origin has been assigned as a possible cause. This type of information is quite important to perform a systematic study of water quality, so it is very important to have the support of technical experts who can discern this kind of behavior and match it with real expe- riences. The creation of an expert system capable of collecting, objectifying and quantifying this experience, can be made more transparent through an inference system based on fuzzy rules. The fuzzy inference systems should allow the identification of combined indices of automatic settings that are representative of specific events or indicators of possible trends in the evolu- tion of the pressures. Figure 6 shows the two types of application processes that have been followed for the development and calibration of the inference system, distinguishing between an automatic formulation of rules once the principles of phenomena are known, and a manual formulation for the risk fish phenomenon due to its special nature. 3.2.1. Expert strategy formulation As a result of the analysis of the information gathered by experts, a strategy formulation has been provided for the three types of events to be detected. The resulting algorithm for the detection of discharge episodes is the concatenation of three fuzzy inference systems, resulting in a value between 0 and 1 drawn from the linguistic labels “Normal”, “Evidence of discharge”, “Slight discharge” and “Serious discharge” about the output variable (Figure 7). Features were selected among incremental and mean values of the original variables, according to expert advice. A point to be studied later is how to best combine the outputs of the three inference systems, either by nesting them or by weighting the outputs. For the detection of eutrophication phenomena, the resulting algorithm is composed of a single fuzzy inference system, resulting in a value between 0 and 1 extracted from labels “Normal”, “Natural photosynthesis”, “Slight eutrophication” and “Advanced eutrophication” in the output. A diagram of 11 Figure 6: Algorithm calibration process. Figure 7: Fuzzy inference system strategy for the ‘discharge indicator’. 12 Figure 8: Strategy for the ‘eutrophication indicator’. the obtained algorithm is shown in Figure 8. In this case, since eutrophication is a process lasting for several days and depending on the day/night cycle, features were designed using a long temporal window. Finally, the algorithm for detecting fish risk situations consists of a single fuzzy inference system, resulting in a value between 0 and 1 drawn from the values of membership functions on linguistic labels “Normal”, “Possible risk”, “Moderate risk”, “High risk” and “Extreme risk” in the output. In this case, membership functions for the input variables have been manually designed. Representations of the membership functions, as designed by experts, have been depicted in Figure 9. 3.2.2. Automated design Once strategies have been provided by the experts, membership functions for the input and output variables and labels for the output in the rules must be calculated in an automated way from the SAICA network data. Later, they should be tuned or scaled so that the response is adjusted to the expert opinion on the known cases. Input variables’ membership functions. For the input variables, membership functions have been designed according to the following constraints: • three membership functions for each input variable 13 Figure 9: Fuzzy algorithm strategy for the ‘fish risk indicator’. • the universe of discourse is defined by the database ranges and symme- try • membership functions are gaussian functions centered on both bounds of the range and its middle • gaussian functions’ width is a compromise between data quartiles and normalization to the unity As an illustrative example, membership functions for the ‘DO increment’ variable are shown in Figure 10. For the ‘fish risk indicator’ case, the expert-designed membership func- tions for the input variables (depicted in Figure 9) have been translated to those shown in Figure 11. Output variable’s membership functions. For the output variable, member- ship functions have been designed according to the following constraints: • usual equally distributed triangular membership functions have been employed with the universe of discourse in the range [0, 1], for the final output variable 14 Figure 10: Membership functions for the variable ‘DO increment’. • for the concatenated fuzzy inference systems in the ‘discharge indi- cator’, the same distribution has been chosen with four membership functions Output variable’s labels for each rule. Labels for the output variable for each rule are also automatically associated by considering the strength of the input variables in the rule’s precedent, and its either direct or reverse relationship with the rule’s consequent. An easy linear formulation has been employed to assign a label to the output variable. Linguistic labels for the defuzzified output. Tresholds used to assign linguis- tic labels to the indicators have been calculated after the results on known episodes have been analyzed. Output normalization to the range [0, 1] could have been performed. However we are interested on generalization to other basins and rivers, hence it was determined not to scale it. Output normaliza- tion has been delayed for future studies on data coming from further Spanish river basins. 4. Experimentation and result analysis A MATLAB/GUI (Graphical User Interface) based tool has been devel- oped as a fuzzy expert system for the detection of episodes of poor water 15 (a) Fuzzy inference system for the ‘fish risk indicator’. (b) Mean pH. (c) Mean DO. (d) Mean Ammonium. Figure 11: Strategy and input membership functions for the ‘Fish risk indicator’. quality (discharge, eutrophication and fish risk) through continuous mea- surement. The prototype has been evaluated on the SAICA network data provided by the river Guadiana Hydrographic Confederation and obtained results have been analyzed. An example of the GUI-based user-friendly in- terface is illustrated in Figure 12. The indicator of urban or WWTP discharge (‘discharge indicator’) has been specifically calibrated for the Guadajira station and for episodes 1 and 2. It should be adjusted if implemented in other locations. This process could be performed with a set of selected episodes or with a long series of data. The ‘eutrophication indicator’ has been tuned for episode 3, and it has been directly applied to other stations (in the same basin or a different one) without ulterior adjustement. Finally, the ‘fish risk indicator’ is an absolute indicator, and no tuning process was performed in any case. Consequently, it can be directly applied to other stations or basins. 16 Figure 12: MATLAB/GUI prototype’s example. Below are the results of the implementation of the algorithms for calcu- lating event indicators. 4.1. Urban or WWTP discharge Figure 13 (top) shows the results for the ‘discharge indicator’ for Episode 1. It can be noted that the first event has been characterized as “Evidence of discharge”, while the second has been labeled as “Serious discharge”. The indicator has identified yet another possible event at the beginning of the 24th day, mainly due to an increase in ammonium concentration and a lowering of dissolved oxygen, but not important enough to become an “Evidence of discharge”. Figure 13 (bottom) shows the results for the ‘discharge indicator’ for Episode 2. In this case the indicator has a base value slightly below the threshold of an “Evidence of discharge”, so sometimes it is occasionally over- passed. The event on the afternoon of the 22nd day is properly detected and characterized as “Serious discharge”. And just at the beginning of the 24th day there is another event that is detected at the edge of being considered “Slight discharge”. An inspection of the original data does not show that it is a new event, but just a spike in long event conditions detected in the first place. The results suggest the use of the ‘discharge indicator’ thresholds shown in Table 1 to characterize detected events more easily. 17 Figure 13: ‘Discharge indicator’ results for Episodes 1 and 2. Discharge “Normal” “Evidence” “Slight” “Serious” Range < 0.4 0.4 − 0.6 0.6 − 0.8 > 0.8 Table 1: Thresholds for the ‘discharge indicator’. 18 Figure 14: ‘Eutrophication indicator’ and ‘Fish risk indicator’ results for Episode 3. Eutrophia “Normal” “Photosynthesis” “Slight” “Advanced” Range < 0.2 0.2 − 0.3 0.3 − 0.4 > 0.4 Table 2: Thresholds for the ‘eutrophication indicator’. In measuring stations located downstream of discharges, at most “Ev- idence of discharge” values were only detected. This makes sense because when water flowed into the main channel of the Guadiana river, dilution occurred and decreased the effects of the discharge. 4.2. Eutrophication The ‘eutrophication indicator’ for Episode 3 enters a state of “Natural photosynthesis” from the 9th day and gradually grows up into “Slight eu- trophication” in the middle of the study period. At this moment, the state moves rapidly to “Advanced eutrophication”, where it remains until the 30th day. Since then decreases progressively, as shown in Figure 14 (top). The results in the ‘eutrophication indicator’ suggest using the thresholds defined in Table 2 to characterize the detected events. It can be noted that the obtained values for this indicator have been lower 19 Fish risk “Normal” “Possible” “Moderate” “High” “Extreme” Range < 0.2 0.2 − 0.3 0.3 − 0.4 0.4 − 05 > 0.5 Table 3: Thresholds for the ‘fish risk indicator’. than in the case of a discharge, so that the thresholds should be proportion- ally lower, or should be scaled to obtain uniformity in the system. 4.3. Fish risk Figure 14 (bottom) shows the results of the ‘fish risk indicator’ for Episode 3. The indicator enters a situation of concern about the middle of the study period, and the obtained values are growing up beyond the threshold of “Moderate risk”. However, the detected discharge of ammonium is not sig- nificant enough to generate a dangerous situation. From the 25th day the risk values decrease to reach a “Normal” situation in the middle of the 29th. As it can be observed, the ‘eutrophication indicator’ draws a gradual pro- cess of degradation of the quality conditions in the river reaching the “Ad- vanced eutrophication” state. Under these conditions the pH is slightly basic which is manifested in a “Moderate risk” label for the ‘risk fish indicator’. This represents only a risk situation, but the conditions necessary to trigger a situation of danger to fish life exist if there is an increased concentration of ammonium, and this could occur at any time by an urban discharge. It has been observed in other case studies that in a situation of “Slight eutrophication” there has been a major discharge of ammonium, but the ‘fish risk indicator’ has not reached high-risk values. The reason for this phenomenon is that, although there is a high concentration of ammonium, environmental conditions with normal pH do not facilitate the formation of ammonia. Results obtained for the ‘fish risk indicator’ suggest using the thresholds defined in Table 3 to characterize the detected events. The developed ‘fish risk indicator’ is a novel helpful tool for better en- vironmental management, as it provides objective criteria for risk mitiga- tion actions, such as forcing the dilution of nutrients to break the cycle of eutrophication by providing flow from reservoirs, thereby avoiding possible river wildlife mortality. 20 5. Conclusions From the methodological point of view, algorithms based on fuzzy logic that allow the detection and identification of episodes of urban or WWTP dis- charge, eutrophication and fish risk have been developed. These algorithms have been obtained from the extraction of the experience and knowledge of experts in this field, as well for the characterization of the phenomena as for the selection of variables to use. The presented work demonstrates the ability of fuzzy-logic-based methods to synthesize complex information, in- terpretable only by a few experts, and translate it into more understandable indicators for environmental managers and the general public. The developed expert system is based on data from measuring stations belonging to SAICA networks. Using a few physical-chemical variables con- tinuously recorded, the expert system is able to obtain indicators of water quality phenomena that may be associated, with a high probability of cause- effect relationship, to human pressure on the water environment, as would be urban discharges or agricultural diffuse pollution. In this sense, SAICA networks for continuous water quality measurements complement adminis- trative official networks of manual sampling and analytical laboratory, by providing information about episodic events (discharges) or continuous pro- cesses (eutrophication) which can hardly be detected by discrete sampling. The next steps to perform are the extension of the methods of calcula- tion of indicators to more phenomena of water quality, particularly follow-up downstream episodes, and application in other basins with different dynamic behavior. Acknowledgments The SAICA network data have been provided by the Guadiana Hydro- graphic Confederation, which has also contributed its experience and knowl- edge of the basin needed to characterize the episodes. The work has been developed in the context of the project ECOWATCH, led by Adasa Sistemas S.A.U., which has received a grant (ref. 022/SGTB/2007/6.1) from “En- vironmental projects of scientific research, technological development and innovation” of the Spanish Ministry of Environment, Rural and Marine En- vironment under the National Program of Environmental Science and Tech- nology. 21 References European Parliament, C., December 2000. Directive 2000/60/ec of the eu- ropean parliament and of the council of 23 october 2000 establishing a framework for community action in the field of water policy. Official Jour- nal L 327, 1–73. Hatzikos, E., Hätönen, J., Bassiliades, N., Vlahavas, I., Fournou, E., 2009. Applying adaptive prediction to sea-water quality measurements. Expert Systems with Applications 36 (3, Part 2), 6773 – 6779. Lermontov, A., Yokoyama, L., Lermontov, M., Machado, M. A. S., 2009. River quality analysis using fuzzy water quality index: Ribeira do iguape river watershed, brazil. Ecological Indicators 9 (6), 1188 – 1197. Nasiri, F., Maqsood, I., Huang, G., Fuller, N., 2007. Water quality index: A fuzzy river-pollution decision support expert system. Journal of Water Resources Planning and Management 133 (2), 95–105. Serramià, A., 2005. Information systems and water quality in Spain. Inge- nieŕıa Qúımica 420, 155–158, in Spanish. Smith, V. H., Tilman, G. D., Nekola, J. C., 1999. Eutrophication: impacts of excess nutrient inputs on freshwater, marine, and terrestrial ecosystems. Environmental Pollution 100 (1-3), 179 – 196. 22 
work_2yfpqgwysjgvtfekdhm5l462zm	----	A Fuzzy Expert System Architecture For Data And Event Stream Processing HAL Id: cea-01803819 https://hal-cea.archives-ouvertes.fr/cea-01803819 Submitted on 7 Jan 2019 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. A fuzzy expert system architecture for data and event stream processing Jean-Philippe Poli, L. Boudet To cite this version: Jean-Philippe Poli, L. Boudet. A fuzzy expert system architecture for data and event stream pro- cessing. Fuzzy Sets and Systems, Elsevier, 2018, 343 (SI), pp.20-34. �10.1016/j.fss.2017.10.005�. �cea- 01803819� https://hal-cea.archives-ouvertes.fr/cea-01803819 https://hal.archives-ouvertes.fr A Fuzzy Expert System Architecture For Data And Event Stream Processing Jean-Philippe Poli, Laurence Boudet CEA, LIST, Data Analysis and System Intelligence Laboratory, 91191 Gif-sur-Yvette cedex, France. Abstract The Internet of Things was born from the proliferation of connected objects and is known as the third era of information technology. It results in the availability of a huge amount of continuously acquired data which need to be processed to be more valuable. This leads to a real paradigm shift: instead of processing fixed data like classical databases or files, the new algorithms have to deal with data streams which bring their own set of requirements. Researchers address new challenges in the way of storing, querying and processing those data which are always in motion. In many decision making scenarios, fuzzy expert systems have been useful to deduce a more conceptual knowledge from data. With the emergence of the Internet of Things and the growing presence of cloud-based architectures, it is necessary to improve fuzzy expert systems to support higher level operators, large rule bases and an abundant flow of inputs. In this paper, we introduce a modular fuzzy expert system which takes data or event streams in input and which outputs decisions on the fly. Its architecture relies on both a graph-based representation of the rule base and the cooperation of four customizable modules. Stress tests regarding the number of rules have been carried out to characterize its efficiency. Keywords: Fuzzy expert system, complex event processing, data stream ∗Corresponding author Email address: jean-philippe.poli@cea.fr (Jean-Philippe Poli, Laurence Boudet) Preprint submitted to Fuzzy Sets and Systems October 11, 2017 processing, rule base representation, policies 1. Introduction The emergence of connected objects and mobile devices gave birth to the Internet of Things and is leading towards a continuous data acquisition from different devices and sensors. Before this third era of information technology, the data were stored in data warehouse, queried at once and manipulated by5 algorithms as a whole. With such data in motion, the use cases have changed: for instance, new database management paradigms are introduced, special efforts are made on data compression to avoid networks overload, and supervised or unsupervised learning algorithms are rethought. Cugola and Margara [1] define the Information Flow Processing (IFP) do-10 main as the domain of tools capable of processing information as it flows. Usu- ally, the flow is coming from multiple sources and processed to extract relevant knowledge. They distinguish two subdomains:Complex Event Processing (CEP) and Data Stream Processing (DSP). Algorithms for processing such flows have to be fast and incremental [2] and are evaluated regarding two criteria: the15 number of passes over the stream (which must be as close as possible to 1) and the size of the workspace in memory [3]. On the one hand, DSP consists in processing data flows and in producing a new data flow as output. The Federal Standard defines a data stream as a “sequence of digitally encoded signals used to represent information in transmis-20 sion”. More formally, data stream can be defined [2, 3, 4] as a sequence of data items x1, . . . , xi, . . . , xn such that the items are read once in increasing order of the indexes i. A lot of tools have been introduced to process data streams. For instance, traditional database management systems, which work on persistent data, are replaced by data stream management systems whose queries run con-25 tinuously: anytime new data arrive, the result of a query is updated [5, 6]. Other researches mainly include data streams mining with either clustering methods [7] or neural networks [8]. In [9, 10], the authors are revealing the open chal- 2 lenges which must be addressed in the domain of data stream mining, including privacy issues, developing a methodology for stream preprocessing, developing30 online monitoring systems and balancing resources. On the other hand, CEP differs by the type of data items it considers: an item is a notification of event [11]. CEP are also associated with velocity: it aims at managing thousands of events per second [12], for instance, up to 125000 events for a financial software [13]. In this domain, processing mainly consists35 in filtering, gathering and combining those events to build a higher level infor- mation [14], to raise alerts or trigger processes. It makes use of the relationships which exist between events: indeed, events may be related in various ways (by cause, by timing, or by membership) [11]. Nowaday, DSP and CEP became usual considerations in many real world40 applications. We can cite for example: system monitoring and fault detection, home automation, security and finance [1]. However, in both data and event streams, the information may be incomplete and imprecise by nature [15]. For instance, sensors may be out of order or inaccurate, and data may be noisy. Fuzzy logic [16] has been specifically designed to mathematically represent un-45 certainty and vagueness and is a popular tool for dealing with imprecision in many real world problems. Taking advantage of fuzzy logic, fuzzy expert sys- tems allow to easily represent human knowledge about data and phenomena and have been successfully applied to many domains [17, 18]. Fuzzy expert systems often come with a higher computational cost compared50 with boolean logic expert systems. For instance, fuzzy inference needs to assess the whole rule base to compute the outputs, whereas in boolean expert systems, a subset of the rules are applied one by one to produce the inference. Whereas fuzzy inference involves simple operations and simple functions to lower the computational cost, it has been showed that fuzzy rule bases need to be more55 complicated if only piecewise-linear functions (e.g. trapezoids...) are used in- stead of non-linear membership functions (e.g. sigmoids...) [19]. Consequently, expensive functions in terms of computation are needed to assess the aggrega- tion and the defuzzification [20]. In addition, in real-world applications, it is 3 possible to have very large rule bases which require a great amount of processor60 time [21]. Moreover, to describe the relations between the data or the events, more sophisticated operators are needed for temporal [22, 23], spatial [24, 25] or even spatio-temporal [26] reasoning. These operators imply a higher com- putational cost. In addition, traditional fuzzy expert systems compute output values only when input values have changed. This is not compliant with event65 streams whose events are potentially arriving in an irregular manner: in such a case, expressions may change before the next event arrival (see section 3.3). Our work aims at developing a fuzzy expert system to process information flows, handling the imprecision brought by noisy data, sensors or network prob- lems with fuzzy logic. The motivation of our work is to provide an efficient fuzzy70 expert system in operational contexts. To enable human experts to author more complex rules, our system is able to efficiently assess complex fuzzy relations [23]. To ensure it can interface easily with the various information systems of our partners, we chose to avoid specific architectures (like GPU) and to develop a software for data and event stream processing on regular CPU platforms. Fi-75 nally, in industrial applications, the efficiency is important not only because there must be a lot of rules, but also because the rules can be applied to a huge number of objects or events per second. The paper is structured as follows: section 2 presents the related work about large rule base handling. Section 3 describes the architecture of our fuzzy expert80 system. Section 4 presents the implementation, and then the protocol and the results of experiments on both data and event streams. Finally, section 5 points out the conclusions. 2. Related work To address real-world applications, fuzzy expert systems have to be able to85 process large rule bases very fast, and thus face the well-known combinatorial explosion problem. Indeed, in case of a conjunctive combination of the terms of all the inputs (also known as grid fuzzy rule structure), the number of rules 4 grows exponentially with respect to the number of inputs. For instance, there are pn possible rules for n inputs with p terms each.90 Combs and Andrews [27] tried to solve this problem by proposing a rule construction schema based on the union of single-antecedent rules, called Union Rule Configuration. In this case, the number of rules evolves only linearly in function of the number of inputs. Mendel and Liang [28] contested this approach stating that the two rule bases may not be equivalent and suggest to inquire95 whether the replacement makes sense. Large rule bases are rarely given by human experts and are rather automat- ically inducted from datasets. Thus, automatic rule induction also faces the curse of dimensionality. To fix this issue, one can generate a partial rule base which does not cover the whole input space. Approaches based on input space100 partitioning by k-d trees [29] or quadtrees [30] result in nonuniform overlapping membership functions which are difficult to label with an understandable lin- guistic term. Jin [31] takes advantage of the conclusion drawn in [32] : optimal rules cover the extrema. The generated rule base is then checked for redundancy and potential inconsistency and optimized by a genetic algorithm. Hence, Jin’s105 algorithm generates 27 rules from a training set of 20000 examples described by 11 inputs. In [33], authors present S-FRULER a genetic fuzzy system for regression problem capable of learning rules from big datasets. First, a multi- granularity fuzzy discretization is applied on the whole dataset, which is then split into several partitions. Each partition is then treated as an independent110 problem, which allows distribution. The processing consists in selecting vari- ables randomly and then using genetic algorithms to induce rules. The rules are then combined into a unique rule base. Authors claim they obtain simple rule bases in terms of number of rules while maintaining a good precision. How- ever, the different steps do not guarantee to have an optimal rule base since115 random variable selection, genetic algorithms and rule combination can add a lot of biases. The interpretability of the rule base is not the only reason to decrease the number of rules: applying on information streams needs a fast processing of 5 the rules to make decisions on the fly. Fuzzy controllers have been introduced120 to overcome these drawbacks and are able to process inputs in real-time [34]. More recently, papers address the acceleration of fuzzy computation either with dedicated hardware [35] or with the help of Graphics Processing Units (GPU) [36]. However, to our experience, fuzzy expert softwares which run on classic CPU platforms are more convenient for many reasons. Firstly, they are easier to125 interface with an existing system than electronic chipsets. Then, DSP and CEP both rely on software intensive architectures. Moreover, in terms of scalability, it is possible to use from a single core of a machine to several machines and it can all be done transparently for the user ; for instance, it can take advantage of the virtualization as in cloud-based services.130 The next section describes a suitable architecture for general-purpose fuzzy expert systems in which the problem of the rule base size in terms of computa- tion speed is addressed by eliminating redundancy in the assessment. We also distribute the different roles like gathering inputs, computing and announcing results in different modules to be able to process information streams.135 3. Architecture description Fuzzy expert systems can infer regarding two main paradigms: • Mamdani type, in which rule conclusions are fuzzy set; • Takagi-Sugeno type, in which rule conclusions are a function of the inputs, i.e. a crisp value.140 Whatever the type of inference, when a group of inputs change at a time t, all the rules containing at least one of those inputs have to be reevaluated. In information streams, inputs may change several times per second, or rules must be applied on thousands of incoming events per second ; the evaluation of the whole rule base may thus need a huge computation time. The distribution of145 the rule base is not always a good solution: for instance, monitoring of vehicles 6 or of patients need to process the same rule base over different input streams. However, the evaluation for each vehicle or each patient can be concurrent. In this article, we introduce an architecture which tends to avoid the system saturation. In the remainder, without loss of generality, the examples will be150 given for a Mamdani type inference system. 3.1. Architecture overview Figure 1 presents the overview of the proposed architecture. The modularity is ensured by a separation of the tasks and a customization provided by the use of policies. A policy is a set of parameters which customize the behavior of each155 module. The combination of the behaviors of all the modules enable to address a lot of applications and issues : regular or irregular data rate, delay before inference, etc. The architecture is composed of several modules : • the active input queue gathers and groups the inputs by timestamps, • the scheduler monitors the system (via the operating system) and to160 decide which inputs group has to be processed, • the evaluator is in charge of the evaluation of the rules, • the output change broadcaster informs the user about outputs changes. The different modules help avoiding a system overload (for instance, the ac- tive input queue selects the inputs which should be treated) or user overfeeding165 (for instance, the output change broadcaster displays only the relevant infor- mation). We first introduce how we optimize the rule base representation by common subexpression elimination and the concept of expiration of expressions. We then describe each module of the architecture and give some examples of policies.170 3.2. Rule base representation The rule base in-memory model plays a major role in the efficiency of the fuzzy expert system. Expressions are usually modeled with a tree [37], as in 7 Active input queue Scheduler Evaluator Output change broadcaster Rule base Input stream Output stream observes injects uses calls notifies notifies notifies Figure 1: Architecture overview. Figure 2(a). However, some expressions can be included in several rules or other expressions: thus, in a tree representation, it is difficult to check the175 redundancy of such expressions, and it is necessary to compute them several times when a group of inputs changed. This problem is known as common subexpression elimination (CSE). To address the CSE problem in our architecture, we chose to represent each expression by a unique node: thus, the rule base is not represented by a tree180 anymore but by a graph (figure 2(b)). More precisely, we use an acyclic directed graph to avoid loops during the evaluation. In the graph, an edge A −→ B means that if the value of the node A changes, it affects the node B and B has to be evaluated again. In this case, B is called a direct successor of A. In the graph we are considering, a node may have several direct successors. A node185 can represent fuzzy expressions (including fuzzy propositions) or rules, and we consider particular nodes for defuzzification and aggregation. Thus, the changes propagate from input nodes to output nodes. The propagation stops if there are no changes during the evaluation of the current node. The propagation is achieved as particular breadth-first traversal of the graph.190 However, for a fuzzy relation of cardinality n, it is necessary to assess its n 8 Input X Input Y X is A Y is B Input Z Z is C (X is A AND Y is B) OR NOT Z is C IF (X is A AND Y is B) OR NOT Z is C THEN O is D1 (X is A AND Y is B) OR Z is C (X is A AND Y is B) OR Z is C THEN IF O is D2 Aggregation output O Defuzzification output O X is A AND Y is B NOT Z is C Z is C Input Z Input X Input Y X is A Y is B X is A AND Y is B (a) Tree-based representation Input X Input Y X is A Y is B Input Z Z is C NOT Z is C (X is A AND Y is B) OR NOT Z is C IF (X is A AND Y is B) OR NOT Z is C THEN O is D1 (X is A AND Y is B) OR Z is C (X is A AND Y is B) OR Z is C THEN IF O is D2 Aggregation output O Defuzzification output O X is A AND Y is B (0)(0) (0) (1)(1)(1) (2) (2) (3)(3) (4) (4) (5) (6) (b) Graph-based representation Figure 2: Representations of a base of two Mamdani-type rules. predecessors before its own evaluation, otherwise it would be evaluated n times, and worst, at a certain time, its value would be inconsistent. To avoid this effect, we added a priority information to the nodes. Before starting the fuzzy inference engine, the graph is traversed and a recursive function priority : Node →195 integer is applied. Let N be the current node to be treated, the function priority is defined as follow: • if N is an input node, then priority(N) = 0, • otherwise, let si be the direct successors of N: priority(N) = maxi(priority(si)) + 1.200 Let X, Y and Z be three input linguistic variables, and A, B, C a term from respectively X, Y , Z. Let D1 and D2 be two terms of an output linguistic variable O. Then, the rule base is composed of two rules: • IF (X is A AND Y is B) OR NOT Z is C THEN O is D1, • IF (X is A AND Y is B) OR Z is C THEN O is D2.205 In Figure 2(b), numbers in brackets represent the evaluation priority of each node; the three inputs are at the bottom of the figure and have a null priority, 9 which means they need to be evaluated first. We will develop in section 3.6 the use of the priority during evaluation. To the best of our knowledge, current fuzzy expert system does not imple-210 ment CSE. This is due to the fact that they only use classical fuzzy logic opera- tors which are really fast to compute. For instance, t-norms and t-conorms use arithmetic operators and binary comparisons (Zadeh’s operator min and max), whose complexity is O(1), whereas most of temporal operators complexity is in O(l) where l is a number of samples used [23] and most of spatial operators are215 at least in O(n × m) where n × m is the dimension of the image [24]. We can easily assess the number of nodes in the two representations of the rule base. Let Nt be the number of nodes in the tree-based one and Ng in the graph-based one : Nt(n, p) = n︸︷︷︸ inputs + 1︸︷︷︸ aggregation + 1︸︷︷︸ defuzzification + pn︸︷︷︸ rules ( p︸︷︷︸ propositions + (n − 1)︸ ︷︷ ︸ conjonctions + 1︸︷︷︸ implication ) (1) Ng(n, p) = n︸︷︷︸ inputs + n × p︸ ︷︷ ︸ propositions + n∑ i=2 pi ︸ ︷︷ ︸ conjonctions + pn︸︷︷︸ implications + 1︸︷︷︸ aggregation + 1︸︷︷︸ defuzzification (2) Using Landau notations, these equations show the number of nodes is asymp-220 totically O((p + n) · pn) for the tree-based rule base, and O(pn) for the graph- based one. 3.3. Expiration Among the sophisticated relations we have implemented, temporal operators [23] and those which derived from them need a special attention when applied on225 event streams. The particularity of event streams is that the system is noticed of events irregularly. For instance, let us consider the fact “the temperature was 10 too hot on Monday from 2 am to 3 am”. The system has received two events: at 2am, a temperature high enough to activate the fuzzy proposition “the tem- perature is too hot”, and at 3am, a lower temperature such as “the temperature230 is too hot” is false. Now, we consider the temporal operator “occurrence” from [22] which indicates that a phenomenon has occurred on a certain scope in the past: for instance, it can express that “the temperature was too hot during the last 24 hours”. Until the next Tuesday 3 am, the degree of truth of this occurrence is strictly greater than 0. After 24 hours, its degree of truth equals235 0, whereas the system inputs have not changed since Monday 3 am. Classical fuzzy expert systems cannot perform this trick since they need that inputs change to compute the outputs. We thus introduce in our system the notion of expiration. Some expressions in the rule base are represented by special nodes in the rule base graph, which are marked as “expirable”. After being240 evaluated, the expirable nodes signal to the scheduler they must be evaluated again after a certain delay (see section 3.5). If an input which is connected to an expirable node changed before this delay, the expiration is simply postponed by the scheduler. Thus, expirable components must provide an expiration frequency and a set245 of criteria to stop the expiration. The expiration frequency is a parameter which depends on the application and the set of criteria depend only on the definition of the operator. For instance, in home health patient monitoring applications, expressions usually expire every day for symptoms persistence, because doctors cannot have alerts that change too frequently, whereas in control applications,250 the expiration rate is usually less than 1 second to have a smooth behavior. More details can be found in [23]. 3.4. Active input queue Sensor networks are a particular case of information stream. Some sensors measure (data stream) and some others detect (event stream), but they usu-255 ally work in an asynchronous way. Moreover, some delays can appear in such networks. The active input queue is thus in charge of several steps before the 11 engine can process them. Firstly, it listens to the information stream to fetch the interesting values it contains. Then, it groups the input values by timestamp and enqueue. Finally,260 it signals the scheduler that a new group has been enqueued. Different policies can be conceived for this component. For instance, in some applications, it is necessary to wait for delayed sensors or delayed network packets before signaling the scheduler. Conversely, it can ignore delays and late arrivals, and thus filter these data. It may also be seen as a firewall which265 protects the scheduler from irrelevant inputs. 3.5. Scheduler The scheduler has an important role to play to limit the delay between the arrival of the data and the decision making. When a new input set is announced, it decides, regarding its own policy, whether it is important to compute it im-270 mediately, later or not at all. In the simplest configuration, the scheduler just fetches the first element in the active input queue, asks the evaluator to assess this group of inputs and gives the results to the broadcaster. With the use of policies, his behavior can be more sophisticated. For instance, one particular configuration can monitor the275 system to determine how busy the CPU cores are and to decide whether a group of inputs can be skipped. Moreover, the scheduler implements the expiration. All the expirable components of the rule base whose evaluation has changed are placed in another queue, waiting to expire. Another configuration may consist in evaluating on different processor cores280 of the machine. Each core receives a sub-part of the input set. A simple al- gorithm based on the graph representation of the rule base is used to separate independent inputs on different sub-parts: this may simply be achieved by find- ing connected components of graph with well-known algorithms of graph theory [38].285 12 3.6. Evaluator The evaluator is the component which evaluates the different expressions and rules in the rule base. For a set of inputs, it gives a particular set of outputs. It also takes advantage of the rule base representation to perform the computation only when necessary.290 To compute the different nodes of the graph representing the rule base, the evaluator traverses the graph in a certain order. To ensure the right order, we use a priority queue Q. The priority queue Q places the nodes with the lowest priority at the front of the queue and can contain each node at most once. Algorithm 1 presents the general evaluation algorithm. The algorithm takes295 four parameters: I, a dictionary which maps each input which has changed to its value, E, a set of nodes which has expired and must be evaluated again, M, a dictionary which maps each node of the graph to its value, and finally G, the rule base graph. Eventually, I or E can be exclusively empty: when I is empty, the procedure is called just to evaluate the expired nodes, whereas when E is300 empty, it is called to evaluate the consequences of the input change. The assess function takes the node to evaluate and M : it fetches the operands values in M and applies the operator, then stores the value of the current node in M and finally returns false if the value of the current node has not changed, true otherwise.305 In figure 2(b), the priority queue ensures the node “(X is A AND Y is B) OR Z is C” is evaluated at the right time. It ensures that if several paths lead to the same node N, all nodes on the paths are assessed before N. In fuzzy logic, different functions can be used for operators (conjunction, disjunction, negation, implication, aggregation, defuzzification) evaluation. The310 policies of the evaluator indicate which version of the operators must be used. 3.7. Output change broadcast The broadcaster is also an important module because it is in charge of build- ing the output stream. The last step is indeed to inform on the fly the calling 13 Algorithm 1 Evaluation of the rule base graph Inputs: � I: dictionary mapping nodes and values representing inputs which have just changed � E: set of expired nodes � M: dictionary mapping nodes and values resulting of the previous evaluation � G: rule base graph /* Initializes the priority queue Q */ Q ← ∅ /* Adds changed inputs into Q and update memory */ for all pair 〈node, value〉 in I do Q ← priority enqueue(Q, node, priority(node)) M ← M ⋃ 〈node, value〉 end for /* Adds expired nodes into Q */ for all node n in E do Q ← priority enqueue(Q, n, priority(n)) end for /* Rule base graph browsing */ while Q �= ∅ do current ← priority first(Q) Q ← priority dequeue(Q) if assess(current, M) then for all n such as n ∈ successors(current, G) do Q ← priority enqueue(Q, n, priority(n)) end for end if end while return M 14 system or the user that some outputs have changed. The policies are used to315 determine when and how the outputs have to be broadcast. For instance, the changes can be gathered and sent at regular time intervals or only outputs which have changed are broadcast with their new values. In a more verbose style, the changes can be sent with a trace of the activated rules. It may gather information from the graph and the evaluation of its node to320 build justifications (to explain why the decision has been made). The next section introduces implementation considerations and shows some experiments we achieved to characterize the performances of the software on both data and event stream processing. 4. Implementation and experiments325 The software has been developed in C# as an API (Application Program- ming Interface). This language is not compiled as C++, but it offers a good compromise between efficiency and the ease of interfacing with other systems (for instance by webservices) and runs on different platforms without recompil- ing.330 The succession of the nodes evaluation has been implemented without any optimization regarding the current fuzzy inference methods : the software per- forms all the steps needed in Takagi-Sugeno or Mamdani type inferences. We wanted to be as general as possible to be able to take into account future or experimental inference methods. Moreover, all calculations involved in the dif-335 ferent inferences (integrals, curves intersection, ...) are processed analytically, in opposition with, for instance, Matlab which represents a membership function by a given number of points. In terms of memory usage, for now, all the rules are kept in memory at all time. We also have to store the values of the different nodes of the rule base.340 To implement efficiently expiration, which affects only some temporal operators [23], the values of expirable expression operands are stored in memory regarding a temporal scope to allow partial recalculation. Indeed, it is necessary to find 15 a good compromise between efficiency, scalability and flexibility regarding our industrial partners and our needs to experiment new operators and inference345 methods. Implementation is still a work in progress to improve performances. For instance, from the results in [39], we improved the efficiency of the hash function and some data structures and divided by up to 170 the computation time of large rule bases. From the results in [40], we also improved memory usage in order350 to load larger rule bases. Without loss of generality, we have then experimented the system in two different ways: either with a data stream or an event stream. On one hand, to test the ability of the system to process data streams, we measured the time to process rule bases, varying the number of inputs and terms to increase the355 number of rules at each test. On the other hand, we used an event simulator to test the event stream processing capabilities, which is used to benchmark CEP softwares. It is of course possible to address both event streams and data streams with the same rule base. In the first series of experiments, all the inputs are changing at each time360 whereas in the second series of experiments, few inputs change simultaneously to reveal the potential of our system in real world applications of stream processing. These tests have been processed on only one core of an Intel Xeon X5650 at 2.67GHz on a Windows server and 42GB of RAM to process several tests in parallel without interfering.365 4.1. Data stream experiment This experiment aims at comparing the performances of the evaluation of different rule bases regarding two different policies of the evaluator module : • full recalculation mode : all the expressions and nodes are reassessed each time an input changes,370 • partial recalculation mode : last values of the nodes are kept in memory and are reassessed only when needed, 16 and different rule base representations: • tree-based representation (without CSE), • graph-based representation (with CSE).375 The graph based representation with partial recalculation is the default be- havior of our system. Its modularity, through the use of policies, allows to easily switch between the different modes. The full recalculation mode is obtained by clearing M before each call of the evaluation procedure (algorithm 1) whereas the partial recalculation mode stores in memory the dictionary M.380 4.1.1. Protocol These experiments have been carried out on artificial rule bases and data sets whose generation is described hereafter. Let {vi}1≤i≤n be n input linguistic variables, each defined by p terms T 1i , . . . , T p i . Let w be an unique output linguistic variable whose terms are W1, ..., WK. Those input variables combine385 into rules by the full conjunctive combination principle : IF v1 is T l1 1 and . . . and vn is T ln n THEN w is Wk where T lii refers to a term of vi with 1 ≤ li ≤ p and k = ∑n i=1 li − n + 1. Thus, for a given couple (n, p), there are pn possible combinations of those inputs (i.e. rules) and w has K = n(p − 1) + 1 terms.390 For the sake of simplicity, the terms T lii of each variable vi are defined by triangular membership functions on the domain [0, p + 1]. By construction, the support of each term T lii is [li − 1; li + 1] and its kernel is {li}. The same construction is used for the terms Wk of w. Figure 3 shows an example of a linguistic variable characterized by 3 terms.395 Each input variable vi receives a data stream of 20 values, which have been generated following an uniform distribution U([0, p + 1]). The architecture has been configured as follows: the active input queue is set in DSP mode, i.e. it waits to receive a value for each input. The scheduler evaluates this group as soon as possible, then the new value of the output is400 17 0 1 2 3 4 0 0,5 1 Variable 1 Term 1 Term 2 Term 3 Figure 3: Linguistic variable with 3 terms defined on the domain [0, 4]. broadcasted. This is the most simple configuration of these modules. The dif- ferent modes of evaluation of the architecture have been obtained by configuring the policy of the evaluator : in one case, it uses its memory functionality; in the other case, it has to compute all the values of the nodes again. The same input data streams have been used for both cases.405 By making both the number of inputs n and the number of terms p vary from 2 to 10, we are able to assess the performance of the architecture on large rule bases and to draw some conclusions. Due to the computational cost, the largest configuration was obtained with 7 input variables and 9 linguistic terms with the graph-based representation (4782969 rules), whereas the tree-based410 representation allows to reach 9 input variables and 4 terms (262144 rules). Even if these are not realistic cases, it is useful to benchmark the proposed system. 4.1.2. Results In this section, we first compare the average number of nodes being reevalu-415 ated in each mode and then compare the average evaluation time of the different rule bases. The averages are computed over the 20 values of the data stream to decrease the possible biases. Figure 4 represents the number of nodes to be evaluated in each configuration and figure 5 represents the computation times regarding the number of rules.420 These figures show the results for the full and partial recalculation modes and 18 for the tree-based and graph-based representations of the rule base in log-scale. We can see that for the graph-based rule bases, both the computation time and the number of nodes are linear regarding the number of rules whereas for tree- based rule bases, it is weakly exponential. Point clouds in figure 5 confirm the425 intuition: storing the value of each node allows to stop propagating the changes, and strongly decreases the number of nodes to evaluate. For a rule base with 6 input variables and 8 terms (262144 rules), 1572920 nodes may be evaluated in the full recalculation mode with tree-based rule base and 561784 with a graph, whereas in the partial one, only 35248 nodes in average are evaluated for trees430 and 275 for graphs. 1,E+00 1,E+01 1,E+02 1,E+03 1,E+04 1,E+05 1,E+06 1,E+07 1,E+08 1,E+00 1,E+02 1,E+04 1,E+06 Partial recalculation mode (Graph) Full recalculation mode (Graph) Partial recalculation mode (Tree) Full recalculation mode (Tree) Number of rules Av er ag e nu m be r o f r ee va lu at ed n od es Figure 4: Average number of nodes to be evaluated for different rule base sizes in the all four modes (log-scale for both axes). In particular, figure 5 shows the duration of the evaluation of the rule bases in full and partial modes with graph-based rule bases. With the same data streams as before, to evaluate 262144 rules, full recalculation mode needs almost 2s whereas the partial one needs only 300ms, i.e. the latter one is more than 6435 times faster than the former one on this rule base structure. 19 1,E-05 1,E-04 1,E-03 1,E-02 1,E-01 1,E+00 1,E+01 1,E+02 1,E+03 1,E+00 1,E+01 1,E+02 1,E+03 1,E+04 1,E+05 1,E+06 1,E+07 Partial recalculation mode (Graph) Full recalculation mode (Graph) Partial recalculation mode (Tree) Full recalculation mode (Tree) Number of rules Av er ag e co m pu ta tio n tim e (s ec ) Figure 5: Average computation time in seconds for different rule base sizes in all four modes (log-scale for both axes). 4.1.3. Discusssion The drastic reduction of the number of nodes to be evaluated can be ex- plained by a theoretical analysis. The fuzzy partitions used to create the terms of the linguistic variables explain why, at each time, for each variable, at most 2440 terms out of p are activated. Thus, at most Ng(n, 2) nodes (equ. 2) have to be evaluated: for n = 6, at most 208 nodes will be activated. But a large number of them are null because of the conjunctive combination of the inputs. Now, to count the number of needed reevaluations, we should consider the worst case : all the active elementary propositions become null, and the same number of445 propositions get a non-null value. This gives 2×Ng(n, 2) as a pessimistic upper bound of the number of nodes that need to be reevaluated. It seems that saved computational time is not as high as we could expect considering the saved computations shown just before. Indeed, saved compu- tations correspond to the evaluation of a null value by a quite simple function450 (mainly either by the membership function evaluation or by a conjunctive com- bination of two expressions) and its affectation to nodes that were already null. 20 We will see in section 4.2 that partial recalculation mode is really helpful in event stream processing. These tests are good stress tests because all the inputs change at the same455 time. For rule bases of conventional sizes, for instance 343 rules, the engine needs less than 0.5ms in the partial recalculation mode. Thus, we can handle inputs which change more than 2000 times per second on only one core. 4.2. Event stream experiment Streaming events poses some challenges to the design of a benchmark. As460 described in [41], the authors introduce “Linear Road”, a benchmark for event stream processing systems which matches with many requirements of such bench- marks. For instance, we cite the two most important requirements in our case: • the generated data must have a semantic validity, • the benchmark must be verifiable (even if the streams may vary depending465 on the moment they are generated). The linear road benchmark is inspired from the problem of variable tolling (also known as congestion pricing), i.e. the computation of tolls that vary according to different factors such as congestion levels and accident proximity [42]. The input data are generated by the MIT Traffic Simulator (MITSIM) [43].470 The benchmark consists in different challenges for stream data management systems, from historical queries to tolls assessment. In our case, we have only implemented one challenge out five: the detection of car accidents. The other challenges are not appropriate for a fuzzy expert system and need the capability to query historical data or an aggregated view of the expressway.475 4.2.1. Previous work In the following experiment, we used the “strictly persists” temporal oper- ator described in [23] and which evaluates how much a phenomenon persists during a given scope. In our case, the scope is fuzzy since some moments in 21 -5 -4,5 -4 -3,5 -3 -2,5 -2 -1,5 -1 -0,5 0 0 0,5 1 Scope last 2 minutes (a) Membership function for the fuzzy scope “the last 2 minutes” 0 10 20 30 40 50 0 0,5 1 Distance Very short (b) Membership function of a very short dis- tance Figure 6: Parameters used for the linear road benchmark. the past are considered as more important. Figure 6(a) shows the member-480 ship function used to define the scope: it considers the values from 2.5 minutes to 2 minutes before the present as more and more important, and moments from 2 minutes to the present as important. This fuzzy scope allows testing a representative case for the temporal operator in terms of computational cost. 4.2.2. Protocol485 We generated data with the MITSIM simulator [43] in a flat file which rep- resents a simulation of 6 minutes and 40 seconds, involving 5953 cars. We then select only a given number of cars inside this simulation. The data consist in cars positions, emitted every 30 seconds on a bidirectional expressway. Each direction is composed of 3 travel lanes and one exit and one entrance ramps.490 The position is given by the direction, the lane number and an integer that represents a 1-D coordinate on a specific lane. In the original benchmark, a car accident is detected if at least two cars report the same position for at least four times. More details can be found in [41] and [43]. We developed a software to play the simulation in real time and another one to check if the accident495 notifications are true regarding the simulation. To use the expiration, we send the position of a car only if it changes from its previous position. To match with a fuzzy problem, we changed the benchmark rules. Firstly, we used a fuzzy definition of the distance, to take into account the GPS inaccuracy 22 and the size of the car: figure 6(b) shows a narrow gaussian membership function500 which defines a short distance. Then, we decided to use the “strictly persists” temporal operator, described in section 4.2.1, and to develop two new nodes for the rule graph: • a distance node which simply computes the absolute value of the difference between two inputs values,505 • a “same lane” crisp node, which is boolean and indicates if two cars are on the same lane, same direction. For each unique pair of distinct cars (cari, carj), we wrote the two following rules : • IF (cari is on the same lane as carj AND distance between cari and carj510 IS very short) STRICTLY PERSISTS during at least the 2 last minutes THEN Accident(i, j) IS true, • IF NOT ( (cari is on the same lane as carj AND distance between cari and carj IS very short) STRICTLY PERSISTS during at least the 2 last minutes) THEN Accident(i, j) IS false.515 For n cars on the expressway, we thus have n2 − n pairs of rules. When one car emits a position, it affects 2 × (n − 1) rules which may be reevaluated. In particular, for the “strictly persists” operator, the different values of operand are kept in memory during the last 2.5min as a compressed signal : i.e., as we receive at most a position every 30s for each car, the signal for each pair of cars520 contains at most 10 values. However, we do not store the past values of the car positions. During this test, we used the following configuration : • Input queue groups inputs by timestamps with no delay, • Scheduler processes on one core,525 • Evaluator accepts temporal operators and use Zadeh norm and conorm, 23 Number Simultaneous events Number Response Total computation of cars min avg max of rules time (ms) time (s) 50 1 4.7 9 4900 9.38 1.8 100 2 5.6 9 19800 19.22 5.8 250 2 8 20 124500 317.57 5.8 400 2 12.4 24 319200 806.51 21.9 500 2 15.2 24 499000 836.56 40.1 750 2 21.7 37 1123500 847.53 40.5 1000 2 27.2 43 1998000 862.6 41.5 1100 2 29.2 48 2417800 912.37 41.9 1200 2 31 50 2877600 880.11 41.4 1300 2 32.5 53 3377400 808.65 37.2 Table 1: Results of the partial linear road benchmark performed on a scenario of 6’40” • Broadcaster indicates the value of an output only when it changes, • The persistence expires every ten seconds. When an output changes from false to true, we record the time and the value to indicate that an accident occurs. Note that in the case more than two cars530 are involved in the accident, more than one output will change: for instance, if an accident occurred between car1, car2 and car3, we will be noticed that an accident happened between car1 and car2, car1 and car3, car2 and car3 (thus, 3 notifications for the same accident). 4.2.3. Results535 Table 1 shows the results of the car accident detection in the linear road benchmark. We iterate the tests with different number of cars, from 50 to 1300. The table characterizes the number of simultaneous inputs that change during the simulation by the minimum, the average and the maximum values. It then shows the number of involved rules, the response time, i.e. the average delay to540 evaluate the rule base in milliseconds, and finally the total time of computations during the simulation. For instance, with 1300 cars, in average 32.5 cars report their positions simultaneously. 808.65ms were necessary to tell if an accident happened or not. 24 The total computation time lasts only 37.2s over 401s of simulation.Thus, the545 system was evaluating the rules during less than 10% of the total duration of the simulation. The results show that, thanks to partial recalculation mode, the number of rules does not impact so much on the response time. Naturally, the latter is more impacted by the number of simultaneous events.550 4.2.4. Discussion In this benchmark, regarding the current state of the software, we were lim- ited by two main factors. First, the number of simultaneous events is always low because of the nature of the simulation. Indeed, to the best of our compre- hension, to increase the number of simultaneous events in MITSIM, we have to555 increase the number of cars. However, the number of rules grows too fast regard- ing the number of cars. This leads to the second limitation : the construction of such large rule bases. We have not been able to deal with more than 1300 cars because of the time needed to create the rule base. Indeed, in our architecture, to apply common subexpression elimination on rule bases, we need to check if a560 subexpression has been created before. Even with hashing functions, this step implies a certain amount of computations. Despite this cost, we have to remind that without CSE, we could not evaluate such large rule bases. To better handle the linear road benchmark, systems may consider dynamic rule bases (when a new car appears on the express way, a new set of rules is565 added, and they are removed whenever the car disappears) or to filter the rules to evaluate (e.g. the rules concerning two cars are created only if the two cars are close enough). The goal of this experiment was to study the behavior of the system on large rule bases for event streams. In the case we would want to address this570 benchmark fully, we should use a multi-core scheduler and higher level rules, like “FOR ALL unique pair of cars (x, y), IF ... THEN ...”: we will thus have only two rules in memory for all the pair of cars, while keeping the performances of the inference engine. 25 5. Conclusion575 In this paper, we have presented a modular architecture for a fuzzy expert system designed to handle information streams (data streams or event streams). The architecture relies on two aspects. Firstly, the graph representation of the rule base indicates the dependency between inputs, expressions, rules and outputs. More generally, it indicates what must be computed and in which580 order. Secondly, the use of four cooperating modules permits to filter and to decide when it is possible to process a set of inputs. The introduction of policies in the four modules allows to customize their behaviors regarding the addressed projects or issues. Moreover, the flexibility of the rule base representation has been shown by the addition of two ad-hoc types of node in the graph (“distance”585 node and “same lane and direction” node). The described architecture has been implemented and used in several in- dustrial projects in different domains: home automation, decision making in industry and home care services. All projects needed to process either data stream or event stream, sometimes both of them at the same time.590 Uncertainty and imprecision are real-world challenges, but others emerge. The different experiments that have been presented in this paper show some limitations of the current system: the usage of only one core of the processor and the necessity to load a potentially huge number of rules. Considering CEP and several thousands of inputs per second, we should parallelize the computations.595 We should also consider higher level rules that could be applied efficiently to a lot of inputs while keeping the performances of the inference engine. Moreover, users need more fuzzy relations to be able to describe their scenarios or to characterize what they want to extract from the streams. Finally, online rule base optimization will allow users to sketch first rules and then let the system600 evolve. 26 References [1] G. Cugola, A. Margara, Processing flows of information: From data stream to complex event processing, ACM Comput. Surv. 44 (3) (2012) 15:1–15:62. [2] J. A. Silva, E. R. Faria, R. C. Barros, E. R. Hruschka, A. C. P. L. F. d.605 Carvalho, J. a. Gama, Data stream clustering: A survey, ACM Comput. Surv. 46 (1) (2013) 13:1–13:31. [3] M. R. Henzinger, P. Raghavan, S. Rajagopalan, Computing on data streams, in: J. M. Abello, J. S. Vitter (Eds.), External Memory Algorithms, American Mathematical Society, Boston, MA, USA, 1999, Ch. Computing610 on Data Streams, pp. 107–118. [4] S. Guha, A. Meyerson, N. Mishra, R. Motwani, L. O’Callaghan, Clustering data streams: Theory and practice, IEEE Trans. on Knowl. and Data Eng. 15 (3) (2003) 515–528. [5] A. Arasu, B. Babcock, S. Babu, J. Cieslewicz, M. Datar, K. Ito, R. Mot-615 wani, U. Srivastava, J. Widom, Stream: The stanford data stream man- agement system, Technical Report 2004-20, Stanford InfoLab (2004). [6] S. Chandrasekaran, O. Cooper, A. Deshpande, M. J. Franklin, J. M. Heller- stein, W. Hong, S. Krishnamurthy, S. R. Madden, F. Reiss, M. A. Shah, Telegraphcq: Continuous dataflow processing, in: Proceedings of the 2003620 ACM SIGMOD International Conference on Management of Data, SIG- MOD ’03, ACM, New York, NY, USA, 2003, pp. 668–668. [7] M. Khalilian, M. Norwati, Data stream clustering: Challenges and issues, Proceedings of International Multi Conference of Engineers and Computer Scientists (2010) 566–569.625 [8] D. Cardoso, M. De Gregorio, P. Lima, J. Gama, F. França, A weight- less neural network-based approach for stream data clustering, in: H. Yin, J. Costa, G. Barreto (Eds.), Intelligent Data Engineering and Automated 27 Learning - IDEAL 2012, Vol. 7435 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2012, pp. 328–335.630 [9] G. Krempl, I. Žliobaite, D. Brzeziński, E. Hüllermeier, M. Last, V. Lemaire, T. Noack, A. Shaker, S. Sievi, M. Spiliopoulou, J. Stefanowski, Open chal- lenges for data stream mining research, SIGKDD Explor. Newsl. 16 (1) (2014) 1–10. [10] S. Muthukrishnan, Data streams: Algorithms and applications, in: Pro-635 ceedings of the Fourteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’03, Society for Industrial and Applied Mathematics, Philadelphia, PA, USA, 2003, pp. 413–413. [11] D. C. Luckham, The Power of Events: An Introduction to Complex Event Processing in Distributed Enterprise Systems, Addison-Wesley Longman640 Publishing Co., Inc., Boston, MA, USA, 2001. [12] E. Wu, Y. Diao, S. Rizvi, High-performance complex event processing over streams, in: Proceedings of the 2006 ACM SIGMOD International Con- ference on Management of Data, ACM, New York, NY, USA, 2006, pp. 407–418.645 [13] M. Stonebraker, U. Çetintemel, S. Zdonik, The 8 requirements of real-time stream processing, SIGMOD Rec. 34 (4) (2005) 42–47. [14] E. Alevizos, A. Skarlatidis, A. Artikis, G. Paliouras, Complex event process- ing under uncertainty: A short survey, in: P. M. Fischer, G. Alonso, M. Are- nas, F. Geerts (Eds.), Proceedings of the Workshops of the EDBT/ICDT650 2015 Joint Conference, Vol. 1330 of CEUR Workshop Proceedings, CEUR- WS.org, 2015, pp. 97–103. [15] A. Artikis, C. Baber, P. Bizarro, C. Canudas-de Wit, O. Etzion, F. Fournier, P. Goulart, A. Howes, J. Lygeros, G. Paliouras, A. Schuster, I. Sharfman, Scalable proactive event-driven decision making, Technology655 and Society Magazine, IEEE 33 (3) (2014) 35–41. 28 [16] L. Zadeh, Fuzzy sets, Information and Control 8 (3) (1965) 338 – 353. [17] J. M. Garibaldi, Do Smart Adaptive Systems Exist? Best Practice for Se- lection and Combination of Intelligent Methods, Springer, 2005, Ch. Fuzzy expert systems, pp. 105–132.660 [18] W. Siler, J. Buckley, Fuzzy expert systems and fuzzy reasoning, Wiley- Interscience, 2005. [19] K. Basterretxea, J. M. Tarela, del Campo I, G. Bosque, An experimental study on nonlinear function computation for neural/fuzzy hardware design, IEEE Transactions on Neural Networks 18 (1) (2007) 266–283.665 [20] A. Laurent, M. J. Lesot (Eds.), Scalable Fuzzy Algorithms for Data Man- agement and Analysis: Methods and Design, Information Science Refer- ence, 2010. [21] G. Acampora, V. Loia, Fuzzy control interoperability and scalability for adaptive domotic framework, IEEE Transactions on Industrial Informatics670 1 (2) (2005) 97–111. [22] P. Cariñena, A. Bugaŕın, M. Mucientes, S. Barro, A language for expressing fuzzy temporal rules, Mathware and Soft Computing 7 (2-3) (2000) 213– 227. [23] J.-P. Poli, L. Boudet, D. Mercier, Online temporal reasoning for event and675 data streams processing, IEEE Conference on Fuzzy Systems, FUZZ-IEEE (2016) 2257–2264. [24] I. Bloch, Fuzzy spatial relationships for image processing and interpreta- tion: a review, Image and Vision Computing 23 (2) (2005) 89 – 110. [25] S. Schockaert, M. D. Cock, E. Kerre, Reasoning about fuzzy temporal and680 spatial information from the web, World Scientific, 2010. 29 [26] J.-M. Le Yaouanc, J.-P. Poli, A fuzzy spatio-temporal-based approach for activity recognition, in: Advances in Conceptual Modeling, Vol. 7518 of Lecture Notes in Computer Science, 2012, pp. 314–323. [27] W. E. Combs, J. E. Andrews, Combinatorial rule explosion eliminated by685 a fuzzy rule configuration, Trans. Fuz Sys. 6 (1) (1998) 1–11. [28] J. Mendel, Q. Liang, Comments on ”combinatorial rule explosion elimi- nated by a fuzzy rule configuration”, IEEE Transactions on Fuzzy Systems 7 (3) (1999) 369–373. [29] M. Sugeno, K. Tanaka, Successive identification of a fuzzy model and its690 applications to prediction of a complex system, Fuzzy sets and systems 42 (3) (1991) 315–334. [30] C.-T. Sun, Rule-base structure identification in an adaptive-network-based inference systems, in: IEEE Transaction on Fuzzy Systems, Vol. 2, 1994, pp. 64–73.695 [31] Y. Jin, Fuzzy modeling of high-dimensional systems: Complexity reduction and interpretability improvement, Trans. Fuz Sys. 8 (2) (2000) 212–221. [32] B. Kosko, Optimal fuzzy rules cover extrema, International Journal of In- telligent Systems 10 (2) (1995) 249–255. [33] I. Rodŕıguez-Fdez, M. Mucientes, A. Bugaŕın, S-FRULER: Scalable fuzzy700 rule learning through evolution for regression, Knowledge-Based Systems 110 (2016) 255–266. [34] L. Reznik, Fuzzy Controllers Handbook, Newnes, 1997. [35] K. Basterretxea, I. Del Campo, Scalable Fuzzy Algorithms for Data Man- agement and Analysis: Methods and Design, Information Science Refer-705 ence, 2010, Ch. Electronic Hardware for Fuzzy Computation, pp. 1–30. 30 [36] N. Harvey, R. H. Luke, J. M. Keller, D. Anderson, Speedup of fuzzy logic through stream processing on graphics processing units, IEEE Congress on Evolutionary Computation (2008) 3809–3815. [37] B. R. Preiss, Data Structures and Algorithms with Object-Oriented Design710 Patterns in Java, Worldwide Series in Computer Science, Wiley, 2000. [38] J. Hopcroft, R. Tarjan, Algorithm 447: Efficient algorithms for graph ma- nipulation, Commun. ACM 16 (6) (1973) 372–378. [39] J.-P. Poli, L. Boudet, Une architecture moderne de système expert flou pour le traitement des flux d’information, Rencontres Francophones sur la715 Logique Floue et ses Applications, 2015. [40] J.-P. Poli, L. Boudet, A modular fuzzy expert system architecture for data and event streams processing, in: IPMU, 2016, pp. 717–728. [41] A. Arasu, M. Cherniack, E. Galvez, D. Maier, A. S. Maskey, E. Ryvkina, M. Stonebraker, R. Tibbetts, Linear road: A stream data management720 benchmark, in: Proceedings of the Thirtieth International Conference on Very Large Data Bases - Volume 30, VLDB ’04, VLDB Endowment, 2004, pp. 480–491. [42] A. D. Palma, R. Lindsey, Traffic congestion pricing methods and technolo- gies, Tech. rep., Ecole Polytechnique (2009).725 [43] Q. Yang, H. N. Koutsopoulos, A microscopic traffic simulator for evaluation of dynamic traffic management systems, Transportation Research Part C: Emerging Technologies 4 (3) (1996) 113 – 129. 31 
work_2z3fjxhcrfhefbagjvcl4iyogm	----	Neural network ensemble operators for time series forecasting Kourentzes, N. , Barrow, D. K. and Crone, S. F. Author post-print (accepted) deposited in CURVE January 2016 Original citation & hyperlink: Kourentzes, N. , Barrow, D. K. and Crone, S. F. (2014) Neural network ensemble operators for time series forecasting. Expert Systems with Applications, volume 41 (9): 4235–4244 http://dx.doi.org/10.1016/j.eswa.2013.12.011 DOI 10.1016/j.eswa.2013.12.011 ISBN 978-86-7892-739-3 Publisher: Elsevier Copyright © and Moral Rights are retained by the author(s) and/ or other copyright owners. A copy can be downloaded for personal non-commercial research or study, without prior permission or charge. This item cannot be reproduced or quoted extensively from without first obtaining permission in writing from the copyright holder(s). The content must not be changed in any way or sold commercially in any format or medium without the formal permission of the copyright holders. This document is the author’s post-print version, incorporating any revisions agreed during the peer-review process. Some differences between the published version and this version may remain and you are advised to consult the published version if you wish to cite from it. CURVE is the Institutional Repository for Coventry University http://curve.coventry.ac.uk/open http://dx.doi.org/10.1016/j.eswa.2013.12.011 http://curve.coventry.ac.uk/open Neural Network Ensemble Operators for Time Series Forecasting Nikolaos Kourentzesa,∗, Devon K. Barrowa, Sven F. Cronea aLancaster University Management School Department of Management Science, Lancaster, LA1 4YX, UK Abstract The combination of forecasts resulting from an ensemble of neural networks has been shown to outperform the use of a single “best” network model. This is supported by an extensive body of literature, which shows that combining generally leads to improvements in forecasting accuracy and robustness, and that using the mean operator often outperforms more complex methods of combining forecasts. This paper proposes a mode ensemble operator based on kernel density estimation, which unlike the mean operator is insensitive to outliers and deviations from normality, and unlike the median operator does not require symmetric distributions. The three operators are compared empirically and the proposed mode ensemble operator is found to produce the most accurate forecasts, followed by the median, while the mean has relatively poor performance. The findings suggest that the mode operator should be considered as an alternative to the mean and median operators in forecasting applications. Experiments indicate that mode ensembles are useful in automating neural network models across a large number of time series, overcoming issues of uncertainty associated with data sampling, the stochasticity of neural network training and the distribution of the forecasts. Keywords: Time Series, Forecasting, Ensembles, Combination, Mode Estimation, Kernel Density Estimation, Neural Networks, Mean, Median ∗Correspondance: N Kourentzes, Department of Management Science, Lancaster Uni- versity Management School, Lancaster, Lancashire, LA1 4YX, UK. Tel.: +44-1524-592911 Email address: n.kourentzes@lancaster.ac.uk (Nikolaos Kourentzes) Preprint submitted to Expert Systems with Applications December 5, 2013 1. Introduction With the continuing increase in computing power and availability of data, there has been a growing interest in the use artificial Neural Networks (NNs) for forecasting purposes. NNs are typically used as ensembles of several net- work models to deal with sampling and modelling uncertainties that may otherwise impair their forecasting accuracy and robustness. Ensembles com- bine forecasts from the different models that comprise them. This paper proposes a new fundamental ensemble operator for neural networks that is based on estimating the mode of the forecast distribution, which has appeal- ing properties compared to established alternatives. Although the use of ensembles is nowadays accepted as the norm in fore- casting with NNs (Crone et al., 2011), their performance is a function of how the individual forecasts are combined (Stock and Watson, 2004). Improve- ments in the ensemble combination operators have direct impact on the re- sulting forecasting accuracy and the decision making that forecasts support. This has implications for multiple forecasting applications where NN ensem- bles have been used. Some examples include diverse forecasting applications such as: economic modelling and policy making (McAdam and McNelis, 2005; Inoue and Kilian, 2008), financial and commodities trading (Zhang and Berardi, 2001; Chen and Leung, 2004; Versace et al., 2004; Bodyanskiy and Popov, 2006; Yu et al., 2008), fast-moving consumer goods (Trapero et al., 2012), tourism (Pattie and Snyder, 1996), electricity load (Hippert et al., 2001; Taylor and Buizza, 2002), temperature and weather (Roebber et al., 2007; Langella et al., 2010), river flood (Campolo et al., 1999) and hydrolog- ical modelling (Dawson and Wilby, 2001), climate (Fildes and Kourentzes, 2011), and ecology (Araújo and New, 2007) to name a few. Zhang et al. (1998) lists multiple other forecasting applications where they have been em- ployed successfully. As NN ensembles are fundamental for producing accurate NN forecasts for these various applications; hence, improvements in the construction of the ensembles are important. In this paper, the performance of the proposed mode operator is investigated together with the two existing fundamental ensemble operators: the mean and the median. Two different datasets, in- cluding 3,443 real time series, are used to empirically evaluate the different operators. Furthermore, ensembles of both training initialisations and sam- pling (bagging) are used to investigate the performance of the operators. The proposed operator is found to be superior to established alternatives. 2 Moreover, the robustness and good performance of the median operator is validated. The findings provide useful insights for the application of NNs in large scale forecasting systems, where robustness and accuracy of the fore- casts are equally desirable. The rest of the paper is organised as follows: section 2 discusses the benefits of NN ensembles and the limitations of the established ensemble operators. Section 3 introduces multilayer perceptrons that will be used for this paper and section 4 discusses the three fundamental ensemble operators and presents the proposed method for mode ensembles. Sections 5 and 6 discuss the experimental design and the results respectively, followed by a discussion of the findings in section 7. 2. Forecasting with neural networks Over the last two decades there has been substantial research in the use of NNs for forecasting problems, with multiple successful applications (Zhang et al., 1998). Adya and Collopy (1998) found that NNs outperformed estab- lished statistical benchmarks in 73% of the papers reviewed. NNs are flexible nonlinear data driven models that have attractive properties for forecasting. They have been proven to be universal approximators (Hornik et al., 1989; Hornik, 1991), being able to fit to any underlying data generating process. NNs have been empirically shown to be able to forecast both linear (Zhang, 2001) and nonlinear (Zhang et al., 2001) time series of different forms. Their attractive properties have led to the rise of several types of NNs and appli- cations in the literature (for examples see Connor et al., 1994; Zhang et al., 1998; Efendigil et al., 2009; Khashei and Bijari, 2010). While NNs powerful approximation capabilities and self-adaptive data driven modelling approach allow them great flexibility in modelling time series data, it also complicates substantially model specification and the es- timation of their parameters. Direct optimisation through conventional min- imisation of error is not possible under the multilayer architecture of NNs and the back-propagation learning algorithm has been proposed to solve this problem (Rumelhart et al., 1986), later discussed in the context of time series by Werbos (1990). Several complex training (optimisation) algorithms have appeared in the literature, which may nevertheless be stuck in local optima (Hagan et al., 1996; Haykin, 2009). To alleviate this problem, training of the networks may be initialised several times and the best network model selected according to some fitting criteria. However, this may still lead to 3 suboptimal selection of parameters depending on the fitting criterion, result- ing in loss of predictive power in the out-of-sample set (Hansen and Salamon, 1990). Another challenge in the parameter estimation of NNs is due to the uncertainty associated with the training sample. Breiman (1996b) in his work on instability and stabilization in model selection showed that subset selection methods in regression, including artificial neural networks, are un- stable methods. Given a data set and a collection of models, a method is defined as unstable if a small change in the data results in large changes in the set of models. These issues pose a series of challenges in selecting the most appropriate model for practical applications and currently no universal guidelines exist on how best to do this. In dealing with the first, the NN literature has strongly argued, with supporting empirical evidence, that instead of selecting a single NN that may be susceptible to poor initial values (or model setup), it is preferable to consider a combination of different NN models (Hansen and Salamon, 1990; Zhang and Berardi, 2001; Versace et al., 2004; Barrow et al., 2010; Crone et al., 2011; Ben Taieb et al., 2012). Naftaly et al. (1997) showed that ensembles across NN training initialisations of the same model can improve accuracy while removing the need for identifying and choosing the best training initialisation. This has been verified numerous times in the literature (for example see Zhang and Berardi, 2001). These ensembles aim at reducing the parameter uncertainty due to the stochasticity of the training of the networks. Instead of relying on a single network that may be stuck to a local minima during its training, with poor forecasting performance, a combination of several networks is used. In the case of uncertainty about the training data, Breiman (1996a) proposed Bagging (Bootstrap aggregation and combination) for generating ensembles. The basic idea behind bagging is to train a model on permutations of the original sample and then combine the resulting models. The resulting ensemble is robust to small changes in the sample, alleviating this type of uncertainty. Recent research has lead to a series of studies involving the application of the Bagging algorithm for forecasting purposes with positive results in many application areas (Inoue and Kilian, 2008; Lee and Yang, 2006; Chen and Ren, 2009; Hillebrand and Medeiros, 2010; Langella et al., 2010). Apart from improving accuracy, using ensembles also avoids the problem of identifying and choosing the best trained network. In either case, neural network ensembles created from multiple initial- isations or from the application of the Bagging algorithm, require the use 4 of an ensemble combination operator. The forecast combination literature provides insights on how to best do this. Bates and Granger (1969) were amongst the first to show significant gains in forecasting accuracy through model combination. Newbold and Granger (1974) showed that a linear com- bination of univariate forecasts often outperformed individual models, while Ming Shi et al. (1999) provided similar evidence for nonlinear combinations. Makridakis and Winkler (1983) using simple averages concluded that the forecasting accuracy of the combined forecast improved, while the variabil- ity of accuracy among different combinations decreased as the number of methods in the average increased. The well known M competitions provided support to these results; model combination through averages improves ac- curacy (Makridakis et al., 1982; Makridakis and Hibon, 2000). Elliott and Timmermann (2004) showed that the good performance of equally weighted model averages is connected to the mean squared error loss function and under varying conditions optimally weighted averages can lead to better ac- curacy. Agnew (1985) found good accuracy of the median as an operator to combine forecasts. Stock and Watson (2004) considered simple averages, me- dians and trimmed averages of forecast, finding the average to be the most accurate, although one would expect the more robust median or trimmed mean to perform better. On the other hand, McNees (1992) found no sig- nificant differences between the performance of the mean and the median. Kourentzes et al. (2013) showed that combining models fitted on data sam- pled at different frequencies can achieve better forecasting accuracy at all short, medium and long term forecast horizons and found small differences in using either the mean or the median. There is a growing consensus that model combination has advantages over selecting a single model not only in terms of accuracy and error variability, but also simplifying model building and selection, and therefore the forecast- ing process as a whole. Nonetheless, the question of how to best combine different models has not been resolved. In the literature there are many dif- ferent ensemble methods, often based on the fundamental operators of mean and median, in an unweighted or weighted fashion. Barrow et al. (2010) ar- gued that the distribution of the forecasts involved in the calculation of the ensemble prediction may include outliers that may harm the performance of mean-based ensemble forecasts. Therefore, they proposed removing such elements from the ensemble, demonstrating improved performance. Jose and Winkler (2008) using a similar argument advocated the use of trimmed and winsorised means. On the other hand, median based ensembles, are more 5 robust to outliers and such special treatment may be unnecessary. However, the median, as a measure of central tendency is not robust to deviations from symmetric distributions. The median will merely calculate the middle value that separates the higher half from the lower half of the dataset, which is not guaranteed to describe well the location of the distribution of the forecasts that are used to construct the ensemble. Taking a different perspective, ensembles provide an estimate of where most forecasts tend to be. Mean and median are merely measures of the cen- tral tendency of the forecast distribution. In the case of normal distribution these coincide. Outliers and deviations from normality harm the quality of the estimation. An apparent alternative, that in theory is free of this prob- lem, is the mode. This measure of central tendency has been overlooked in the combination literature because of its inherent difficulty in estimating it for unknown distributions. This paper exploits the properties of the mode to propose a new fundamental ensemble operator. In the following sections this operator is introduced and evaluated against established alternatives. 3. Multilayer Perceptrons The most commonly used form of NNs for forecasting is the feedforward Multilayer Perceptron. The one-step ahead forecast ŷt+1 is computed using inputs that are lagged observations of the time series or other explanatory variables. I denotes the number of inputs pi of the NN. Their functional form is: ŷt+1 = β0 + H ∑ h=1 βhg ( γ0i + I ∑ i=1 γhipi ) . (1) In eq. (1), w = (β, γ) are the network weights with β = [β1, . . . , βH], γ = [γ11, . . . , γHI] for the output and the hidden layers respectively. The β0 and γ0i are the biases of each neuron, which for each neuron act similarly to the intercept in a regression. H is the number of hidden nodes in the network and g(·) is a non-linear transfer function, which is usually either the sigmoid logistic or the hyperbolic tangent function. NNs can model interactions between inputs, if any. The outputs of the hidden nodes are connected to an output node that produces the forecast. The output node is often linear as in eq. (1). 6 w 1 w 2 M S E −10 −5 0 5 10 −10 −5 0 5 10 0 0.2 0.4 0.6 0.8 1 Figure 1: Contour plot of the error surface of a neural network. The initial (⊕) and ending (•) weights for six different training initialisations are marked. In the time series forecasting context, neural networks can be perceived as equivalent to nonlinear autoregressive models (Connor et al., 1994). Lags of the time series, potentially together with lagged observations of explanatory variables, are used as inputs to the network. During training pairs of input vectors and targets are presented to the network. The network output is compared to the target and the resulting error is used to update the network weights. NN training is a complex nonlinear optimisation problem and the network can often get trapped in local minima of the error surface. In order to avoid poor quality results, training should be initialised several times with different random starting weights and biases to explore the error surface more fully. Figure 1 provides an example of an error surface of a very simple NN. The example network is tasked to model a time series with a simple autoregressive input and is of the form ŷt+1 = g (w2g (w1yt−1)), where g(·) is the hyperbolic tangent and w1 and w2 its weights. Six different training initialisations, with their respective final weights, are shown. Observe that minor differences in the starting weights can result in different estimates, even for such a simple model. In order to counter this uncertainty an ensemble of all trained networks can be used. As discussed before, this approach has been shown to be superior to choosing a single set of estimated weights. Note that the objective of training is not to identify the global optimum. 7 This would result in the model over-fitting to the training sample and would then generalise poorly to unseen data (Bishop, 1996), in particular given their powerful approximation capabilities (Hornik, 1991). Furthermore, as new data become available, the prior global optimum may no longer be an optimum. In general, as the fitting sample changes, with the availability of new information, so do the final weights of the trained networks, even if the initial values of the network weights were kept constant. This sampling induced uncertainty can again be countered by using ensembles of models, following the concept of bagging. 4. Ensemble operators Let ŷmt be a forecast from model m for period t, where m = 1, . . . , M and M the number of available forecasts to be combined in an ensemble forecast ỹt. In this section the construction of ỹt using the mean, median and the proposed mode operators is discussed. To apply any of these operators reliably a unimodal distribution is assumed. 4.1. Mean ensemble The mean is one of the most commonly used measures of central tendency and can be weighted or unweighted. Let wm be the weight for the forecasts from model m. Conventionally 0 ≤ wm ≤ 1 and ∑M m=1 wm = 1. The ensemble forecast for period t is calculated as: ỹMeant = M ∑ m=1 wmymt. (2) If all wm = M −1 the resulting combination is unweighted. The properties of the mean are well known, as well as its limitations. The mean is sensitive to outliers and unreliable for skewed distributions. To avoid some of its prob- lems one might use a winsorised or truncuted mean (Jose and Winkler, 2008). In this case the mean behaves more closely to the median. For distributions with finite variance, which is true for sets of forecasts, the maximum distance between the mean and the median is one standard deviation (Mallows, 1991). 8 4.2. Median ensemble Similarly the median can be unweighted or weighted, although the latter is rarely used. The median ensemble ỹMediant is simply calculated sorting wmymt and picking the middle value if M is odd or the mean of the two middle values otherwise. Although the median is more robust than the mean it still suffers with non-symmetric distributions. 4.3. Mode Ensemble The mode is defined as the most frequent value in a set of data. The mode is insensitive to outliers, in contrast to the mean and median. There is no formula to calculate the mode of an unknown distribution for continuous variables. There are two common ways to calculate it: either by discretising the data and identifying the most frequent bin, or by kernel density esti- mation. In this work the second approach is preferred in order to avoid the discretisation of the data. Furthermore, kernel density estimation lends itself well to the continuous-valued nature of forecasts. Kernel density estimation is a non-parametric way to estimate the proba- bility density function of a random variable, in this case the forecasts. Given forecasts of a distribution with unknown density f, we can approximate its shape using the kernel density estimator f̂th(x) = (Mh) −1 M ∑ m=1 K ( x − ŷmt h ) , (3) where K(·) is a function with the property ∫ K(x)dx = 1 that is called kernel and h > 0 is its bandwidth. The kernel is often chosen to be a unimodal symmetric density function, therefore making f̂h(x) a density function itself, which is often, for computational reasons, the Gaussian kernel φ(x): φh(x) = 1√ 2πh e − x 2 2h2 . (4) Figure 2 shows an example of the calculation of kernel density. A kernel with bandwidth h is fitted around each observation and the resulting sum approximates the density function of the sample. A number of alternative kernel functions have been proposed in the litera- ture, however the choice of kernel has been found to have minimal impact on the outcome for most cases (Wand and Jones, 1995). The bandwidth of the kernel h controls the amount of smoothing. A high bandwidth results in more 9 D e n si ty Observations Figure 2: Example calculation of kernel density estimation. smoothing. Therefore, the choice of h is crucial, as either under-smoothing or over-smoothing will provide misleading estimation of the density f (Sil- verman, 1981). The approximation by Silverman (1998) is often used in practice h = ( 4σ̂5 3M ) 1 5 , (5) where σ̂ is the standard deviation of the sample of the forecasts. This approx- imation is often adequate for Gaussian kernels. Botev et al. (2010) propose a methodology to automatically select the bandwidth that is free from the arbitrary normal reference rules used by existing methods. This is preferred in the calculation of the mode ensemble as the resulting bandwidth h allows fast convergence and good performance of the ensemble, as it is discussed in the results. The value x that corresponds to the maximum density approximates the mode of the true underlying distribution for a set of forecasts, which is also the value of the mode ensemble ŷModet+h . This is true as long as the estimated distribution is unimodal. Although the probability of facing non-unimodal distributions when dealing with forecasts is low, the following heuristic is proposed to resolve such cases. Since there is no preference between the modes, the one closer to the previous (forecasted or actual) value is retained as the mode. This results in smooth trace forecasts. Eq. (3) results in unweigthed ŷModet+h . It is trivial to introduce wm individual weights for each model. For kernel density estimation to adequately reveal the underlying density a relatively large number of observations are required. A small number of 10 observations will lead to a bad approximation. This is illustrated in figure 3. It shows the mean, median, mode ensembles as well as the selected “best” model forecast, as selected using a validation sample for four different forecast horizons. Furthermore, the estimated kerned density for each horizon is plotted. It is apparent by comparing 3a and 3b that the kernel density estimation using only 10 models is very poor. While in 3a the shape of the distribution is successfully approximated, in 3b there are not enough forecasts to identify the underlying shape of the distribution of the forecasts. Furthermore, in figure 3a it is easy to see that neither the mean, median or the “best” model are close to the most probable value of the forecast distribution. The mode ensemble offers an intuitive way of identifying where forecasts from different models converge and provide a robust forecast, independent of distributional assumptions. 5. Empirical Evaluation 5.1. Datasets To empirically evaluate the performance of the mean, median and the proposed mode ensemble for NNs, two large datasets of real monthly time series are used. The first dataset comes from Federal Reserve Economic Data (FRED) of St. Luis.1 From the complete dataset 3,000 monthly time series that contain 108 or more observations (9 years) were sampled. Long time series were preferred to allow for adequate training, validation and test sets. The second dataset comes from the UK Office for National Statistics and contains 443 monthly retail sales time series.2 Again, only time series with 108 or more observations were retained for the empirical evaluation. A summary of the characteristics of the time series in each dataset is provided in table 1. To identify the presence of trend in a time series the cox- stuart test was employed on a 12-period centred moving average fitted to each time series. The test was performed on the centred moving average to smooth any effects from irregularities and seasonality. To identify the presence of seasonality, seasonal indices were calculated for the de-trended time series and then these were tested for significant deviations from each other by means of a Friedman test. This procedure, based on non-parametric tests, is robust, 1The dataset can be accessed at http://research.stlouisfed.org/fred2/. 2The dataset can be accessed at http://www.ons.gov.uk/ons/rel/rsi/ retail-sales/january-2012/tsd-retail-sales.html. 11 (a) 100 models (b) 10 models Figure 3: Example of the distribution of NN forecasts of different number of models, as estimated by Gaussian kernel density estimation, for the first four steps ahead. The forecasts by model selection, mean, median and mode ensembles are provided. 12 however different tests may provide slightly different percentages to those in table 1. Table 1: Dataset characteristics. Series Length Series Patterns Dataset Series Min Mean Max Level Trend Season Trend-Season FRED 3000 111 327 1124 5.37% 40.70% 5.80% 48.13% Retail 443 179 270 289 15.12% 48.98% 1.81% 34.09% The last 18 observations from each time series are withheld as test set. The prior 18 observations are used as validation set to accommodate NNs training. 5.2. Experimental Design A number of NN ensemble models are fitted to each time series. Two are based on mean, two on median and two on mode ensembles. Hereafter, these are named NN-Mean, NN-Median and NN-Mode respectively. All combina- tion operators are applied in their unweighted version, as the objective is to test their fundamental performance. In each pair of ensembles, the first is a training ensemble, combining multiple training initialisations and the sec- ond is based on bagging, as described by Kunsch (1989). This moving block bootstrap samples the original time series while preserving the temporal and spatial covariance structure, as well as the serial correlation of the time series data. By assessing the operators using different types of ensembles we aim to assess the consistency of their performance. Furthermore, different sizes of ensembles are evaluated, from 10 members up to 100 members, with steps of 10. Results for single NN models, based on selecting the best one, are not provided as there is compeling evidence in the literature that ensembles are superior (for example see Zhang and Berardi, 2001; Barrow et al., 2010). This was validated in our experiments as well. The individual neural networks have identical setup. Following the sug- gestions of the literature, if trend is identified in a time series it is removed through first differencing (Zhang and Qi, 2005). The time series is then li- nearly scaled between -0.5 and 0.5 to facilitate the NN training. The inputs are identified through means of stepwise regression, which has been shown to perform well for identifying univariate input lags for NNs (Crone and 13 Kourentzes, 2010; Kourentzes and Crone, 2010). All networks use the hy- perbolic tangent transfer function for the hidden nodes and a linear output node. The number of hidden nodes was identified experimentally for each time series. Up to 60 hidden nodes were evaluated for each time series and the number of hidden nodes that minimised the validation Mean Squared Error (MSE) was chosen. Each network was trained using the Levenberg-Marquardt (LM) algo- rithm. The algorithm requires setting a scalar µLM and its increase and decrease steps. When the scalar is zero, the LM algorithm becomes just Newton’s method, using the approximate Hessian matrix. On the other hand, when µLM is large, it becomes gradient descent with a small step size. Newton’s method is more accurate and faster near an error minimum, so the aim is to shift toward Newton’s method as quickly as possible. If a step would increase the fitting error then µLM is increased. Here µLM = 10 −3, with an increase factor of µinc = 10 and a decrease factor of µdec = 10 −1. For a detailed description of the algorithm and its parameters see Hagan et al. (1996). MSE was used as the training cost function. The maximum training epochs are set to 1000. The training can stop earlier if µLM becomes equal or greater than µmax = 10 10. The MSE error at the validation set is tracked while training. If the error increases consequently for 50 epochs then training is stopped. The weights that give the lowest validation error are selected at the end of each training. This is common practice in the literature and helps to achieve good out-of-sample performance, since it avoids over-fitting to the training sample (Haykin, 2009). Following the suggestions of the forecasting literature (Adya and Col- lopy, 1998) two statistical benchmarks are used in this study, namely the naive forecast (random walk) and exponential smoothing. This is done to assess the accuracy gains of using NNs against established simpler statistical methods. The Naive requires no parameterisation or setup, hence is used as a baseline that any more complex model should outperform. The appro- priate exponential smoothing model is selected for each time series, depend- ing on the presence of trend and/or seasonality using Akaike’s Information Criterion. Model parameters are identified by optimising the log-likelihood function (Hyndman et al., 2002, 2008). Exponential smoothing was selected as a benchmark based on its widely demonstrated forecasting accuracy and robustness (Makridakis and Hibon, 2000; Gardner, 2006) and will be named ETS in this work. The use of these benchmarks can help establish the rel- ative performance of the NN models. In total, eight forecasting models are 14 fitted to each time series, six NNs and two statistical benchmarks. Rolling trace forecasts of 12 months are produced using each model. The rolling origin evaluation enables collecting a large sample of forecasts and their errors, while being robust to irregular forecast origins and outliers, thus providing reliable error measurements. Based on the long test set, 7 trace forecasts from t+1 up to t+12 months are collected for each time series. The reader is referred to Tashman (2000) for a detailed description of the evaluation scheme and its advantages. The forecasting accuracy is assessed using the Mean Absolute Scaled Er- ror (MASE). This is preferred due to its favourable statistical properties. MASE is calculated for each trace forecast as: MASE = m−1 ∑m j=1 |yj − ŷj| (n − 1)−1 ∑n r=2 |yr − yr−1| , (6) where yj and ŷj are the actual and forecasted value for j = 1, . . . , m out- of-sample observations. The denominator is the mean absolute error of the random walk in the fitting sample of n observations and is used to scale the error. MASE, being a scaled error, permits summarising model performance across time series of different scale and units, which mean squared or absolute errors cannot do, and is less biased from errors like the mean absolute per- centage error and its symmetric equivalent. Another advantage of this error is that it is very improbable that the denominator is zero, therefore making it easy to calculate in several scenarios and robust to time series with several values equal or close to zero (Hyndman and Koehler, 2006). Note that the Retail dataset contains several time series that do not permit the calculation of conventional percentage errors, due to zero values in the denominator. To summarise the results across the time series of each dataset the mean and median MASE across all series are calculated. 6. Results Table 2 presents the results for the FRED time series. Numbers in brack- ets refer to median MASE, while the rest to mean MASE. The table provides results for ensembles from 10 to 100 members. The results for bagging and training ensembles are presented separately to assess the impact of the ensem- ble type on the different ensemble operators. In each row the best performing method according to mean and median MASE is highlighted in boldface. 15 Table 2: Mean (Median) MASE for FRED dataset. Ensemble Size NN-Mean NN-Median NN-Mode Naive ETS Bagging 10 1.06 (0.66) 0.92 (0.64) 1.30 (0.77) 1.11 (0.87) 3.43 (0.62) 20 1.06 (0.65) 0.90 (0.63) 0.94 (0.65) 1.11 (0.87) 3.43 (0.62) 30 1.02 (0.65) 0.89 (0.62) 0.89 (0.62) 1.11 (0.87) 3.43 (0.62) 40 1.04 (0.65) 0.88 (0.62) 0.88 (0.61) 1.11 (0.87) 3.43 (0.62) 50 1.03 (0.64) 0.88 (0.62) 0.88 (0.61) 1.11 (0.87) 3.43 (0.62) 60 1.03 (0.64) 0.89 (0.62) 0.88 (0.61) 1.11 (0.87) 3.43 (0.62) 70 1.04 (0.65) 0.88 (0.62) 0.87 (0.61) 1.11 (0.87) 3.43 (0.62) 80 1.03 (0.65) 0.88 (0.62) 0.87 (0.61) 1.11 (0.87) 3.43 (0.62) 90 1.01 (0.65) 0.88 (0.61) 0.87 (0.61) 1.11 (0.87) 3.43 (0.62) 100 1.01 (0.65) 0.88 (0.61) 0.87 (0.61) 1.11 (0.87) 3.43 (0.62) Training ensemble 10 1.05 (0.64) 0.95 (0.62) 1.17 (0.70) 1.11 (0.87) 3.43 (0.62) 20 1.03 (0.65) 0.93 (0.62) 0.95 (0.64) 1.11 (0.87) 3.43 (0.62) 30 1.01 (0.64) 0.91 (0.62) 0.90 (0.62) 1.11 (0.87) 3.43 (0.62) 40 1.02 (0.64) 0.91 (0.62) 0.90 (0.61) 1.11 (0.87) 3.43 (0.62) 50 1.02 (0.64) 0.92 (0.62) 0.89 (0.61) 1.11 (0.87) 3.43 (0.62) 60 1.01 (0.64) 0.91 (0.62) 0.89 (0.62) 1.11 (0.87) 3.43 (0.62) 70 1.01 (0.64) 0.91 (0.61) 0.89 (0.61) 1.11 (0.87) 3.43 (0.62) 80 1.01 (0.64) 0.91 (0.62) 0.88 (0.61) 1.11 (0.87) 3.43 (0.62) 90 1.01 (0.64) 0.91 (0.61) 0.88 (0.61) 1.11 (0.87) 3.43 (0.62) 100 1.01 (0.64) 0.91 (0.62) 0.88 (0.61) 1.11 (0.87) 3.43 (0.62) Overall, the difference between the mean and median MASE results in- dicates that there are several difficult time series, particularly affecting the less robust mean MASE. Focusing on the bagging results, all NN-Mean, NN- Median and NN-Mode are more accurate than the benchmarks when con- sidering mean MASE. Furthermore, as the ensembles increase in size their accuracy improves. In particular, for NN-Mode after there are 30 or more members the forecasts are very accurate. This was to be expected since the kernel density estimation becomes reliable once there is an adequate num- ber of observations, as discussed in section 4. For ensembles of 70 or more members NN-Mode provides consistently the best accuracy, closely followed by NN-Median. Note that achieving large numbers of ensemble members is trivial with NNs, as this merely implies that more training initialisations or bootstrapped samples are used. Therefore, the requirement of the mode op- erator for 30 or more ensemble members is not a limiting factor. In contrast, 16 NN-Mean underperforms to the extent that ETS is more accurate for median MASE. This is an interesting finding, given how common is the mean oper- ator for ensembles in the literature. The more robust behaviour of median and the in-sensitive to outliers nature of the mode result in more accurate ensemble forecasts. Looking at mean MASE, all NNs behave more robust than ETS, the latter being severely affected by outliers. The results of the training ensembles are very similar. Again, as the number of members in the ensemble increases NN-Mode performs better and is the most accurate model for 40 or more ensemble members. NN-Median ranks second with small differences, while NN-Mean is substantially worse. Comparing the results between bagging and training ensembles we can see that the former is marginally more accurate for NN-Median and NN-Mode when mean MASE is considered. However, the same is not true for NN-Mean, indicating that the robustness and performance of this ensemble operator is affected by the type of ensemble. Table 3 presents the results for the Retail dataset. Its structure is the same as in table 2. The differences between mean and median MASE are smaller than the FRED results, showing that the time series in this dataset are better behaved. Considering the bagging results, NN-Median consistently outperforms the statistical benchmarks, while the same is true for NN-Mode, once there is an adequate number of members in the ensemble (again 30 or more). NN-Mode is the most accurate model with the lowest mean and median MASE. This is followed closely by NN-Median. On the other hand, NN-Mean often fails to outperform the benchmark ETS, although it is always better than the Naive. Looking at the accuracy of the training ensembles NN-Mode is overall more accurate for mean MASE, NN-Median is the most accurate for median MASE. Although all NN models outperform the Naive benchmark, the dif- ferences between either NN-Mode or NN-Median and ETS are very small. NN-Mean is worse than ETS in terms of mean MASE, while occasionally it is marginally better in terms of median MASE. Comparing accuracies between bagging and training ensembles there are differences in favour of the former when looking at NN-Median and NN-Mode, while the accuracy for NN-Mean is almost identical for both types of ensembles. Across both datasets NN-Mode and NN-Median are the most accurate models. NN-Mode seems to perform better when the size of ensemble is large enough. NN-Median has slightly lower accuracy. While large ensem- bles benefit NN-Median, it can perform well for small ensembles too. Both 17 Table 3: Mean (Median) MASE for Retail dataset. Ensemble Size NN-Mean NN-Median NN-Mode Naive ETS Bagging 10 1.33 (0.96) 1.11 (0.93) 1.44 (1.10) 1.45 (1.29) 1.12 (0.97) 20 1.37 (0.97) 1.10 (0.94) 1.14 (0.94) 1.45 (1.29) 1.12 (0.97) 30 1.29 (0.96) 1.10 (0.91) 1.09 (0.92) 1.45 (1.29) 1.12 (0.97) 40 1.31 (0.97) 1.10 (0.91) 1.09 (0.90) 1.45 (1.29) 1.12 (0.97) 50 1.30 (0.97) 1.10 (0.92) 1.09 (0.89) 1.45 (1.29) 1.12 (0.97) 60 1.26 (0.96) 1.09 (0.91) 1.09 (0.89) 1.45 (1.29) 1.12 (0.97) 70 1.26 (0.96) 1.09 (0.91) 1.09 (0.90) 1.45 (1.29) 1.12 (0.97) 80 1.26 (0.98) 1.09 (0.90) 1.08 (0.88) 1.45 (1.29) 1.12 (0.97) 90 1.27 (0.97) 1.09 (0.90) 1.09 (0.87) 1.45 (1.29) 1.12 (0.97) 100 1.27 (0.95) 1.09 (0.91) 1.09 (0.88) 1.45 (1.29) 1.12 (0.97) Training ensemble 10 1.34 (0.97) 1.14 (0.91) 1.27 (0.97) 1.45 (1.29) 1.12 (0.97) 20 1.33 (0.95) 1.14 (0.91) 1.14 (0.91) 1.45 (1.29) 1.12 (0.97) 30 1.31 (0.96) 1.13 (0.89) 1.11 (0.90) 1.45 (1.29) 1.12 (0.97) 40 1.28 (0.95) 1.12 (0.91) 1.11 (0.90) 1.45 (1.29) 1.12 (0.97) 50 1.28 (0.96) 1.12 (0.90) 1.11 (0.91) 1.45 (1.29) 1.12 (0.97) 60 1.29 (0.96) 1.12 (0.89) 1.11 (0.91) 1.45 (1.29) 1.12 (0.97) 70 1.29 (0.95) 1.13 (0.90) 1.11 (0.90) 1.45 (1.29) 1.12 (0.97) 80 1.29 (0.95) 1.13 (0.90) 1.11 (0.90) 1.45 (1.29) 1.12 (0.97) 90 1.28 (0.95) 1.13 (0.89) 1.12 (0.90) 1.45 (1.29) 1.12 (0.97) 100 1.28 (0.96) 1.13 (0.89) 1.12 (0.90) 1.45 (1.29) 1.12 (0.97) these models are on average more accurate than the statistical benchmarks ETS and Naive. On the other hand, NN-Mean provides mixed results. In both datasets it outperforms Naive, but not always ETS. It is substantially outperformed by both NN-Mode and NN-Median. 7. Discussion The value of ensembles for NNs has been argued theoretically and demon- strated empirically. The combination of the models has often involved some type of mean operator. The empirical evaluation in this paper found that the less commonly used median operator and the proposed mode operator are more accurate and thus preferable. The size of the ensemble was found to be important for the accuracy of all operators. Both mode and median, for the two datasets investigated here, seemed on average to converge for ensem- 18 bles of 60 or more members, with any additional members offering minimal changes in the forecasting performance. In particular, the mode, due to its reliance on kernel density estimation, required at least 30 members. However, after that point it was found to be on average the most accurate ensemble op- erator. This is illustrated in figures 4 and 5 that present the mean MASE for different number of ensemble members for the FRED and the Retail datasets respectively. The results for the different type of ensembles have been pooled together, since they had only small differences. Note that there is little evi- dence that the mean ensembles had converged even with 100 members. Even larger ensembles were not calculated due to the substantial computational resources required, especially when the objective is to forecast a large num- ber of time series, which is common in supply chain and retailing forecasting applications. 10 20 30 40 50 60 70 80 90 100 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2 1.25 Ensemble members m e a n M A S E FRED dataset Mean Median Mode Figure 4: Mean MASE for different number of ensemble members for the FRED dataset. In order to assess whether these differences are significant or not, we employ the testing methodology suggested by Koning et al. (2005) that is appropriate for comparing forecasts from multiple models. The comparison is done across all different ensemble sizes to highlight if an operator is consis- tently statistically different. First, a Freidman test is used to assess whether the accuracy of any model is significantly different from the rest. Subse- quently, the MCB test is used to reveal the exact ordering of the different operators, and any significant differences between them. For both datasets the mode operator was significantly better than the median, which in turn was significantly different than the mean, at 5% significance level. At this point, it is useful to comment on the associated computational 19 10 20 30 40 50 60 70 80 90 100 1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 Ensemble members m e a n M A S E Retail dataset Mean Median Mode Figure 5: Mean MASE for different number of ensemble members for the Retail dataset cost of the NN ensembles. The main cost comes from training the networks. Therefore, the more ensemble members that need to be trained, the less scalable forecasting with NNs becomes, and the operator that achieves good forecasting performance with the least amount of members is preferable. Table 4 provides an overview of the average time required for forecasting across all series, for each dataset. As a different number of hidden nodes are used for each time series, the complexity of NN training changes, requiring different amount of time. To keep the presentation of the values simple, we summarise the training time over different series into the reported average time. The ensemble size that gave the minimum error for each operator in figures 4 and 5 is used as reference for the comparison. The average time in seconds, as well as the percentage difference over the time needed for the mode ensembles, are provided in the table. The networks were trained in parallel on an i7-3930K CPU clocked at 4.5 Ghz with 12 logical cores. The mode operator needed the least number of ensemble members, re- quiring from 25% up to 200% less time than the mean or median operators across both datasets. Therefore, apart from the significant gains in forecast- ing accuracy, the proposed ensemble operator required the least computa- tional resources. In particular for the retailing dataset, the run-time was more than halved. In figures 4 and 5 it is clear that similar performance is achieved for a large range of ensemble sizes for the median operator. This allows exchanging marginal differences in accuracy for smaller run-times, thus improving its scalability as well. On the other hand, this is not the case with the mean 20 Table 4: Average computational time comparison. Ensemble operator FRED Retail Ensemble Mean Difference Ensemble Mean Difference size time size time Mean 100 9.74 secs +25.0% 70 5.33 secs +133.3% Median 100 9.74 secs +25.0% 90 6.85 secs +200.0% Mode 80 7.79 secs - 30 2.28 secs - operator, the accuracy of which improves with bigger ensembles. In the experiments, two types of ensembles were considered, bagging and training ensembles. Each one tackles a different type of parameter uncer- tainty. We examined whether the performance of the operators was affected by the type of ensemble. Again median and mode had very similar perfor- mance, favouring bagging. For the mean this behaviour was not evident. We attribute this different behaviour to the sensitivity of the mean to ex- treme values, which both median and mode are designed to avoid, albeit with different success. 8. Conclusions This paper evaluates different fundamental ensemble operators. The well known mean and the less commonly used median were compared, together with a proposed mode operator that is based on kernel density estimation. All these three different operators attempt to describe the location of the distribution of the forecasts of the members of an ensemble. However, they deal with outlying extreme values differently, with the mean being the most sensitive and the mode the least. Furthermore, distributional asymmetries can affect both the mean and the median, while the mode is immune. The findings in this paper suggest that both median and mode are very useful operators as they provided better accuracy than mean ensembles con- sistently across both datasets. The mode was found to be the most accurate, followed by the median. Based on this finding, we recommend investigating the use of the mode and median operators further in ensembles research and applications, which have been largely overlooked in the literature that has mainly focused on the mean. Furthermore, this work demonstrated that mode ensembles can robustly 21 and accurately forecast automatically a large number of time series with neural networks, while the commonly used mean ensembles were often out- performed by exponential smoothing forecasts. Moreover, mean ensembles required a very large number of members, which neither mode or median needed, with apparent implications for computational costs. In particular, the mode operator was found to require the least computation resources, due to the relatively small number of ensemble members that needed to be trained. We have already mentioned a number of applications that can benefit from improved NN ensemble forecasts, ranging from economic and business forecasting to climate modelling. Most of these applications are characterised by forecasting a few, yet important, time series. The improved scalability of mode ensembles over the commonly used mean ensembles allows applying NNs to areas that routinely require large scale automatic forecasting, which can benefit from the nonlinear modelling capabilities of NNs. One such ex- ample is retailing, where one has to forecast a large number of products, the sales of which are affected by multiple factor that interact in a nonlin- ear fashion, such as pricing, promotional and temperature effects. The im- proved scalability of mode ensembles, compounded with the ever increasing computing capabilities provides opportunities for novel important forecasting applications of NN ensembles. This paper found significant savings in com- puting time from the proposed operator, which over the complete number of time series accounts for several hours of computations. Such reduction will also help using NN ensembles in high frequency forecasting cycles, where the computational speed has been a limiting factor. Future work should explore these potentials. The empirical evaluation, in this paper, focused on the unweighted version of all these operators, trying to assess their fundamental properties. Although their differences are attributed to their robustness to extreme values, future research should extend this work to weighted versions of the operators. This will allow considering their use on further ensemble types, such as boosting. References Adya, M., Collopy, F., 1998. How effective are neural networks at forecasting and prediction? a review and evaluation. Journal of Forecasting 17 (5-6), 481–495. 22 Agnew, C. E., 1985. Bayesian consensus forecasts of macroeconomic vari- ables. Journal of Forecasting 4 (4), 363–376. Araújo, M. B., New, M., 2007. Ensemble forecasting of species distributions. Trends in ecology & evolution 22 (1), 42–47. Barrow, D., Crone, S., Kourentzes, N., 2010. An evaluation of neural net- work ensembles and model selection for time series prediction. In: Neural Networks (IJCNN), The 2010 International Joint Conference on. IEEE, pp. 1–8. Bates, J. M., Granger, C. W. J., 1969. The combination of forecasts. Oper- ational Research Quarterly 20 (4), 451–468. Ben Taieb, S., Bontempi, G., Atiya, A. F., Sorjamaa, A., 2012. A review and comparison of strategies for multi-step ahead time series forecasting based on the NN5 forecasting competition. Expert Systems with Applications 39 (8), 7067–7083. Bishop, C. M., 1996. Neural Networks for Pattern Recognition, 1st Edition. Oxford University Press, USA. Bodyanskiy, Y., Popov, S., 2006. Neural network approach to forecasting of quasiperiodic financial time series. European Journal of Operational Research 175 (3), 1357–1366. Botev, Z. I., Grotowski, J. F., Kroese, D. P., 2010. Kernel density estimation via diffusion. The Annals of Statistics 38 (5), 2916–2957. Breiman, L., 1996a. Bagging predictors. Machine learning 24 (2), 123–140. Breiman, L., 1996b. Heuristics of instability and stabilization in model selec- tion. The Annals of Statistics 24 (6), 2350–2383. Campolo, M., Andreussi, P., Soldati, A., 1999. River flood forecasting with a neural network model. Water resources research 35 (4), 1191–1197. Chen, A.-S., Leung, M. T., 2004. Regression neural network for error correc- tion in foreign exchange forecasting and trading. Computers & Operations Research 31 (7), 1049–1068. 23 Chen, T., Ren, J., 2009. Bagging for gaussian process regression. Neurocom- puting 72 (7), 1605–1610. Connor, J. T., Martin, R. D., Atlas, L. E., 1994. Recurrent neural networks and robust time series prediction. IEEE Transactions on Neural Networks 5, 240–254. Crone, S. F., Hibon, M., Nikolopoulos, K., 2011. Advances in forecasting with neural networks? Empirical evidence from the NN3 competition on time series prediction. International Journal of Forecasting 27 (3), 635–660. Crone, S. F., Kourentzes, N., 2010. Feature selection for time series prediction - a combined filter and wrapper approach for neural networks. Neurocom- puting 73 (10-12), 1923–1936. Dawson, C., Wilby, R., 2001. Hydrological modelling using artificial neural networks. Progress in physical Geography 25 (1), 80–108. Efendigil, T., Önüt, S., Kahraman, C., 2009. A decision support system for demand forecasting with artificial neural networks and neuro-fuzzy models: A comparative analysis. Expert Systems with Applications 36 (3), 6697– 6707. Elliott, G., Timmermann, A., 2004. Optimal forecast combinations under general loss functions and forecast error distributions. Journal of Econo- metrics 122 (1), 47–79. Fildes, R., Kourentzes, N., 2011. Validation and forecasting accuracy in mod- els of climate change. International Journal of Forecasting 27 (4), 968–995. Gardner, E. S., 2006. Exponential smoothing: The state of the art - part ii. International Journal of Forecasting 22 (4), 637–666. Hagan, M. T., Demuth, H. B., Beale, M. H., 1996. Neural Network Design. MA: PWS Publishing, Boston. Hansen, L. K., Salamon, P., 1990. Neural network ensembles. IEEE Trans- actions Pattern Analysis and Machine Intelligence 12 (10), 993–1001. Haykin, S., 2009. Neural Networks and Learning Machines. Pearson Educa- tion, Inc. 24 Hillebrand, E., Medeiros, M. C., 2010. The benefits of bagging for forecast models of realized volatility. Econometric Reviews 29 (5-6), 571–593. Hippert, H. S., Pedreira, C. E., Souza, R. C., 2001. Neural networks for short- term load forecasting: A review and evaluation. Power Systems, IEEE Transactions on 16 (1), 44–55. Hornik, K., 1991. Approximation capabilities of multilayer feedforward net- works. Neural Networks 4 (2), 251–257. Hornik, K., Stinchcombe, M., White, H., 1989. Multilayer feedforward net- works are universal approximators. Neural Networks 2 (5), 359–366. Hyndman, R. J., Koehler, A. B., 2006. Another look at measures of forecast accuracy. International Journal of Forecasting 22, 679–688. Hyndman, R. J., Koehler, A. B., Ord, J. K., Snyder, R. D., 2008. Forecasting with Exponential Smoothing: The State Space Approach. Springer. Hyndman, R. J., Koehler, A. B., Snyder, R. D., Grose, S., 2002. A state space framework for automatic forecasting using exponential smoothing methods. International Journal of Forecasting 18 (3), 439–454. Inoue, A., Kilian, L., 2008. How useful is bagging in forecasting economic time series? a case study of us consumer price inflation. Journal of the American Statistical Association 103 (482), 511–522. Jose, V. R. R., Winkler, R. L., 2008. Simple robust averages of forecasts: Some empirical results. International Journal of Forecasting 24, 163–169. Khashei, M., Bijari, M., 2010. An artificial neural network (p, d, q) model for timeseries forecasting. Expert Systems with Applications 37 (1), 479–489. Koning, A. J., Franses, P. H., Hibon, M., Stekler, H. O., 2005. The M3 competition: Statistical tests of the results. International Journal of Fore- casting 21 (3), 397 – 409. Kourentzes, N., Crone, S. F., 2010. Frequency independent automatic input variable selection for neural networks for forecasting. In: Neural Networks (IJCNN), The 2010 International Joint Conference on. IEEE, pp. 1–8. 25 Kourentzes, N., Petropoulos, F., Trapero, J. R., 2013. Improving forecasting by estimating time series structural components across multiple frequen- cies. International Journal of Forecasting. Kunsch, H. R., 1989. The jackknife and the bootstrap for general stationary observations. The Annals of Statistics 17 (3), 1217–1241. Langella, G., Basile, A., Bonfante, A., Terribile, F., 2010. High-resolution space–time rainfall analysis using integrated ann inference systems. Journal of Hydrology 387 (3), 328–342. Lee, T.-H., Yang, Y., 2006. Bagging binary and quantile predictors for time series. Journal of Econometrics 135 (1), 465–497. Makridakis, S., Andersen, A., Carbone, R., Fildes, R., Hibon, M., Lewandowski, R., Newton, J., Parzen, E., Winkler, R., 1982. The ac- curacy of extrapolation (time series) methods: Results of a forecasting competition. Journal of Forecasting 1 (2), 111–153. Makridakis, S., Hibon, M., 2000. The M3-competition: results, conclusions and implications. International Journal of Forecasting 16 (4), 451–476. Makridakis, S., Winkler, R. L., 1983. Averages of forecasts: Some empirical results. Management Science 29 (9), 987–996. Mallows, C., 1991. Another comment on ocinneide. The American Statisti- cian 45, 257. McAdam, P., McNelis, P., 2005. Forecasting inflation with thick models and neural networks. Economic Modelling 22 (5), 848–867. McNees, S. K., 1992. The uses and abuses of ‘consensus’ forecasts. Journal of Forecasting 11, 703–711. Ming Shi, S., Da Xu, L., Liu, B., 1999. Improving the accuracy of nonlinear combined forecasting using neural networks. Expert Systems with Appli- cations 16 (1), 49–54. Naftaly, U., Intrator, N., Horn, D., 1997. Optimal ensemble averaging of neural networks. Network: Computation in Neural Systems 8, 283–296. 26 Newbold, P., Granger, C. W. J., 1974. Experience with forecasting univariate time series and the combination of forecasts. Journal of the Royal Statis- tical Society. Series A (General) 137 (2), 131–165. Pattie, D. C., Snyder, J., 1996. Using a neural network to forecast visitor behavior. Annals of Tourism Research 23 (1), 151–164. Roebber, P. J., Butt, M. R., Reinke, S. J., Grafenauer, T. J., 2007. Real-time forecasting of snowfall using a neural network. Weather and forecasting 22 (3), 676–684. Rumelhart, D. E., Hinton, G. E., Williams, R. J., 1986. Learning internal rep- resentations by error propagation. In: Rumelhart, D. E., McClelland, J. L. (Eds.), Parallel distributed processing: explorations in the microstructure of cognition. Vol. 1. MIT Press, Cambridge, MA, USA, Ch. Learning in- ternal representations by error propagation, pp. 318–362. Silverman, B., 1998. Density Estimation for Statistics and Data Analysis. Chapman & Hall/CRC, London. Silverman, B. W., 1981. Using kernel density estimates to investigate multi- modality. Journal of the Royal Statistical Society. Series B (Methodologi- cal) 43 (1), 97–99. Stock, J. H., Watson, M. W., 2004. Combination forecasts of output growth in a seven-country data set. Journal of Forecasting 23 (6), 405–430. Tashman, L. J., 2000. Out-of-sample tests of forecasting accuracy: an ana- lysis and review. International Journal of Forecasting 16 (4), 437–450. Taylor, J. W., Buizza, R., 2002. Neural network load forecasting with weather ensemble predictions. Power Systems, IEEE Transactions on 17 (3), 626– 632. Trapero, J. R., Kourentzes, N., Fildes, R., 2012. Impact of information ex- change on supplier forecasting performance. Omega 40 (6), 738–747. Versace, M., Bhatt, R., Hinds, O., Shiffer, M., 2004. Predicting the exchange traded fund DIA with a combination of genetic algorithms and neural networks. Expert systems with applications 27 (3), 417–425. 27 Wand, M. P., Jones, M. C., 1995. Kernel Smoothing. Chapman & Hall. Werbos, P. J., 1990. Backpropagation through time - what it does and how to do it. Proceedings of the IEEE 78 (10), 1550–1560. Yu, L., Wang, S., Lai, K. K., 2008. Forecasting crude oil price with an EMD- based neural network ensemble learning paradigm. Energy Economics 30 (5), 2623–2635. Zhang, G., Patuwo, B. E., Hu, M. Y., 1998. Forecasting with artificial neural networks: The state of the art. International Journal of Forecasting 14 (1), 35–62. Zhang, G. P., 2001. An investigation of neural networks for linear time-series forecasting. Computers and Operations Research 28 (12), 1183–1202. Zhang, G. P., Berardi, V. L., 2001. Time series forecasting with neural net- work ensembles: an application for exchange rate prediction. Journal of the Operational Research Society 52, 652–664. Zhang, G. P., Patuwo, B. E., Hu, M. Y., 2001. A simulation study of artifi- cial neural networks for nonlinear time-series forecasting. Computers and Operations Research 28 (4), 381–396. Zhang, G. P., Qi, M., 2005. Neural network forecasting for seasonal and trend time series. European Journal of Operational Research 160 (2), 501–514. 28 Neural cover Neural 
work_2zlvlsgvvbe5tke7aiw2ffir4u	----	Refinement of the HEPAR Expert System: Tools and Techniques∗ Peter Lucas Department of Computer Science Utrecht University Padualaan 14 3584 CH Utrecht, The Netherlands June 26, 2013 Abstract Methods and tools for the static and dynamic verification and validation of software sys- tems are common place in the field of software engineering. In the field of expert systems, where it is more difficult to ensure that a system meets the specifications and expecta- tions than in traditional software engineering, such tools are generally not available. In this paper, the need for more support of the development process by method and tools is illustrated by the approach taken in building the HEPAR system, a rule-based expert system that can be used as a supportive tool in the diagnosis of disorders of the liver and biliary tract. At a certain stage in the development of this system an incremental devel- opment methodology has been adopted, in which implementation of parts of the expert system was followed by dynamic validation. For this purpose, a collection of software tools were implemented as extensions to a rule-based expert-system shell. These tools provide valuable information about the effects of modification of the HEPAR knowledge base, and indicate places in the knowledge base for refinement. It is believed that similar software tools may prove helpful in the development of other expert systems as well. Keywords. Knowledge engineering; validation of expert systems; refinement of expert systems; evaluation of expert systems; expert-system validation tools. 1 Introduction In the field of software engineering it is generally recognized that the implementation of large software systems must be supported by methods and tools for their verification and valida- tion [19]. Often a distinction is made between static and dynamic approaches to verification and validation. Typical examples of static methods are program code inspection and meth- ods for proving program correctness. The application of static methods does not require the program to be executed. In contrast, dynamic verification and validation methods involve execution of the program and examining its output when it is presented with certain input. As has been pointed out repeatedly by many researchers, dynamic methods can be used only to demonstrate the presence of errors in a program, and never to demonstrate their absence. ∗Published in: Artificial Intelligence in Medicine, 6(2), 175–188, 1994. 1 Despite this fundamental limitation, software engineers consider dynamic methods indispen- sible aids in the software-development process, and supporting software tools are invariably included in programming environments. It is therefore ironical that in the development of expert systems, where it is much more difficult to ensure that the system meets its specifica- tions and expectations than in software engineering, tools that aid in the dynamic verification and validation of the system are not generally available. In this paper we discuss several static and dynamic methods, including some software tools, which were applied during the development of the HEPAR system, a rule-based expert system in the field of hepatology that offers support in the diagnosis of disorders of the liver and biliary tract. For a description of this system and of two successive studies of its performance, the reader is referred to [10, 11, 12]. Taking the development of the HEPAR system as a real-life example, we describe where in the development process the need for static and dynamic verification and validation methods and tools arose. This experience provides further evidence that in the development process of expert systems more support by methods and tools tailored to verification and validation is needed than is presently offered. The structure of the paper is as follows. In Section 2, we review the design and imple- mentation of the HEPAR system and analyse the problems encountered in the design stage. Several static verification and validation methods were employed at this stage of the project. In Section 3, the subject of knowledge-base refinement is related to expert-system validation using dynamic techniques. In Section 4, we describe some simple software tools which were used for the purpose of dynamic validation during the refinement of the HEPAR system. 2 Development of the HEPAR system 2.1 Knowledge acquisition and design It is now well-recognized that the acquisition of domain knowledge in the process of building an expert system is a difficult task [6]. In recent years, many methodologies have therefore been proposed, providing systematic methods to be followed in building an expert system. Examples of such methodologies are KADS [3] and KEATS [15, 16]. Some of these method- ologies include a set of software tools which help the knowledge engineer in building a specific application, mainly by assisting in the analysis of the problem domain. Some assistance by software tools may be provided in the design of the expert system as well. Examples of such tools are Shelley [2] and Acquist [15, 16]. Most methodologies place considerable emphasis on the process of gathering domain knowledge to be incorporated into the expert system, and on the development of conceptual models of the domain, being the result of the analysis of the knowledge collected. Although the HEPAR system was actually designed long before such methodologies came into play, the development of the system was initially carried out in a structured way, mainly following the top-down design methodology from software engineering. The knowledge con- cerning diagnosis in liver and biliary disease incorporated into the HEPAR system was derived from the experience of a specialist in internal medicine and hepatology and from the medical literature. The analysis of the problem of diagnosis of disorders of the liver and biliary tract indicated that the following aspect were important in this domain: • Expert hepatologists follow a clear and unambiguous strategy in diagnosis. The early classification of a patient’s disorder into rather general categories, such as whether or not 2 the disorder is biliary-obstructive in nature, is used for the selection of supplementary tests to reduce the number of alternative diagnoses to be considered. • Early in the diagnostic process only a limited amount of patient data is available, mainly obtained from medical history and physical examination. Still, a hepatologist is often capable of coming up with a working diagnosis of the patient’s disorder. In the design of the HEPAR system, the strategy followed by the hepatologist has been taken as the point of departure for problem decomposition of the diagnostic process, by distinguishing several subtasks [11]. The requirement that the ultimate system ought to be able to assist the clinician in the initial assessment of the patient, in whom only a limited amount of data is available, as well as in the assessment of the patient in whom more specific test results are known, proved to be extremely difficult. In general, to explicitly deal with data not available in the patient would yield an exponential number of combinations of conditions on known and unknown patient findings to be taken into account. Although the hepatologist involved in the project was able to reduce the number of useful combinations considerably, we felt quite uncertain with respect to the suitability of this knowledge for classifying actual patient cases. Note that current knowledge-acquisition methodologies do not offer much help in solving this problem. In most popular methodologies, the design process is essentially viewed as the process of abstraction from reality. Our problem was that we required some form of experimental feedback in refining the expert system to accommodate to reality. Likewise, only limited attention has been given to tools that provide information about the diagnostic quality of the advice produced by the expert system. 2.2 Implementation and experimental feedback Implementation of the HEPAR system was carried out using the EMYCIN-like expert-system shell DELFI-2 [9]. The advice produced by the system is in the form of a collection of conclusions: 1. Whether the patient has a hepatocellular of biliary-obstructive disorder. 2. Whether the features found in the patient indicate benign or malignant disease. 3. A differential diagnosis consisting of a collection of specific disorders, selected out of a set of about 80 disorders, each explaining some of the findings observed in the patient. In HEPAR, the first two conclusions are intermediate and the last is a final conclusion. After completing a considerable portion of the knowledge base, it was decided to carry out some experiments with the system using data from real patients to investigate whether the system was able to meet our expectations. It appeared that the system was unable to come up with an acceptable advice in many cases. An analysis of the results of this initial experiment yielded the following reasons for the disappointing performance: • Many rules were formulated too rigorously, such that these rules almost never applied in a patient with the given disease. • Many rules were defined without explicitly mentioning the medical context in which they should hold. These rules frequently succeeded in patients for whom they had not been designed. 3 if same(patient,complaint,abdominal pain) and ⇒notsame(patient,complaint,fever) and same(patient,signs,hepatomegaly) and same(pain,character,continuous) and same(ultrasound liver,parenchyma,multiple cysts) and same(patient,nature-disorder,benign) then conclude(patient,diagnosis,polycystic disease) with CF = 1.00 fi Figure 1: Weakly formulated production rule. if (same(complab,duration,chronic) or same(patient,type-disorder,biliary obstructive)) and same(patient,sex,female) ⇒ same(patient,sex,male) and same(serol,mitochondrial Ab,yes) then conclude(patient,diagnosis,primary biliary cirrhosis) with CF = 0.80 ⇒ with CF = 0.60 fi Figure 2: Rigorously formulated production rule. These problems may actually be taken as an indication of the knowledge clinicians draw upon in medical practice. Firstly, the knowledge of the clinician is partly based on the descriptions given in medical textbooks, in which there is little place for the description of atypical disease patterns, and partly on experience in the management of specific disorders. Rules which have been formulated too rigorously tend to describe the typical picture of the disease, and may assume the availability of an unrealistic amount of data for the patient. Secondly, the clinician has considerable experience with disorders frequently observed in clinical practice. However, these disorders carry a clinical context which the clinician may not be able to make explicit. Formalizing such knowledge may yield rules with a wider application than intended. As an example, consider the production rule depicted in Figure 1. In this HEPAR rule we use the object–attribute–value representation and certainty factors as employed in the DELFI-2 system [9, 13]. This rule was originally formulated without the second condition (indicated by the right arrow); for the modified rule to be applicable, fever must be absent in the patient. So, in its original form it is an example of a too weakly formulated production rule. This missing condition caused the original rule to interact with Caroli’s disease, infected liver cysts and cystic liver metastases. 4 The production rule shown in Figure 2 is an example of a too rigorously formulated rule, since in the form left from the right arrows, the rule is only applicable for female patients. Typically, a patient having primary biliary cirrhosis is a female, but the disorder is not limited to the female sex. A new production rule was therefore added to the knowledge base in which the expressions specified right from the arrow replaced the expressions left from the arrow. As said above, clinicians often have to base their early decisions on incomplete clinical evidence; likewise, expert systems must be able to deal with incomplete evidence as well. In designing and implementing the HEPAR system we have tried to explicitly handle incomplete patient data in the following two ways: 1. By distinguishing several different conceptual levels in the diagnostic problem-solving process; 2. By explicitly incorporating knowledge about unknown diagnostic test results for a pa- tient into the knowledge base. To deal with the first source of incompleteness of information in the HEPAR system, rules were drafted covering only the symptoms and signs of a disorder obtained early in the diagnostic process, whereas other rules were drafted only covering the results of supplementary tests obtained later in the diagnostic process. In this way a more or less layered structure of the knowledge base was obtained. The second source of incompleteness was dealt with by inspecting rules for conditions on data not always available in the patient. Some of these rules were used as a basis for new rules containing conditions concerning unknown data. Mainly static verification and validation methods were employed at this stage; the early experiments were only carried out to validate our design rationale. 3 Knowledge base refinement and validation 3.1 Refinement parameters The process of refining an expert system may be viewed as the iterative process of validating, extending and adapting its knowledge base. Here, we are concerned with dynamic validation. Since the extension and adaptation are based on the results of the validation, it is clearly important to decide on the parameters used for validating an expert system. The point of departure for expert system validation is to consider a medical diagnostic expert system as a computer program that tries to construct a model of a given patient which it compares with prestored descriptions of disease patterns in its knowledge base. Under ideal circum- stances, validation of a diagnostic expert system should therefore not only pertain to the advice produced by the expert system, but also to the assumptions made by the reasoning process on which the conclusions are based. It is, for instance, important to know whether it is possible for an expert system to arrive at conclusions which are correct, but based on incorrect assumptions. So, the conclusions of a medical diagnostic expert system should not be interpreted as unique answers, but as judgements of the patient’s status [8]. As a con- sequence, not unique answers but judgements should be validated. This view of validation is also able to cope with situations in which the expert system’s conclusions are incorrect, but nevertheless acceptable in the light of the data available. This approach to validation is particularly valuable when applied in refining an expert system; it is less appropriate in the final validation of an expert system [7]. 5 3.2 Refinement by dynamic validation For the purpose of the refinement of the HEPAR knowledge base, we have investigated several approaches to system validation. A diagnostic expert system like HEPAR may be validated against: 1. Patient cases with known clinical diagnosis; 2. The conclusions of some other, but similar decision support system; 3. The judgement of human experts in the field. In all three cases, there is a need for a test, or a combination of tests, that may be taken as a ‘gold standard’ for the comparison, although this is not as crucial as in the final validation of an expert system, because inspection of the knowledge base may provide additional information. Examples of tests that are suitable as a gold standard in the diagnosis of disorders of the liver and biliary tract are the histological examination of liver biopsies, ERCP and surgical exploration. There is not a single test available in hepatology that may be employed as a gold standard in the entire domain, because diagnosis of hepatocellular disorders differs considerably from diagnosis of biliary obstruction. Initially in the refinement, we have used part of the data from more than 1000 patients obtained from the Danish COMIK group as a source for comparison. Originally, these data have been used in the development of the Copenhagen Pocket Chart, a paper chart based on the statistical technique of logistic regression, which may assist the clinician in the early assessment of a patient with jaundice [14]. Unfortunately, this database was of limited value because only 23 disease categories were distinguished in this database, whereas in HEPAR more than 80 disorders of the liver and biliary tract are distinguished, and not all data required for HEPAR to derive final conclusions were included in the database. Therefore, the database was mainly useful for getting insight into the extent to which HEPAR was capable of dealing with incomplete patient data. During the further development of the system, a database with patient data from the Leiden University Hospital was put together, which was applied as the main device for the refinement of the knowledge base. Comparison of an expert system with some similar decision-support system is seldom straightforward, because of differences in required input and produced output among the sys- tems. For the refinement of the HEPAR system, we have included production rules that map diagnostic conclusions of HEPAR to the possible diagnostic conclusions of the Copenhagen Pocket Chart. The results produced by the Copenhagen Pocker Chart could thus be used as a simple means for rapidly finding patient case which deserved further study with regard to the conclusions produced by HEPAR. The hepatologist involved in the project has studied the reasoning process of HEPAR for a considerable number of patients. Validation of the reasoning process of HEPAR turned out to be very time-consuming, and as a consequence we have not been able to involve other hepatologists. However, on a limited scale we have profited from discussion with other hepatologists in examing the results produced by HEPAR. Taking the conclusions concerning the final diagnosis produced by the HEPAR system as a point of departure for refining the system was difficult, because the system’s advice consists of more than one conclusion. This problem is known as the multiple response problem [18]. Consider for example the situation that the expert system did produce the correct answer with highest ranking, as well as a conclusion with lower ranking which is however totally 6 unacceptable to the clinician. Restricting the validation only to the single conclusion with the highest ranking will give a distorted account of the actual performance of the system. These problems cannot be solved by only collecting information concerning the number of conclusions generated by the system and the ranking of the clinical diagnosis. Furthermore, consider the situation in which the conclusion with highest ranking is incorrect, but where the differential diagnosis as a whole is acceptable to the clinician. Taking only the conclusion with highest ranking into account will then give an inadequate impression of the system’s capabilities. However, due to the layered approach to diagnostic problem solving modelled in HEPAR, it is not only possible to compare specific disorders as a diagnosis with the clinical diagnosis confirmed in the patient, but also to check whether the patient’s disorder has been classified into the right diagnostic category (e.g. hepatocellular disorder). This layered approach makes it less likely that the differential diagnosis produced by the system is as a whole unacceptable to the clinician. Most of the information obtained by the dynamic validation of the HEPAR knowledge base was automatically compiled by a collection of simple software tools which are discussed in the next section. Without the availability of these tools, dynamic validation would have been too time-consuming for being practically feasible. 4 Tools for knowledge base refinement 4.1 Testing tools in software engineering Some of the software tools that have been developed for the dynamic validation of the HEPAR system, and applied in the refinement and the performance studies of the system, have been inspired by software tools commonly available in programming environments. Test-data generators, programs that systematically produce test data to be used as input to the program to be tested, are one type of program that may be useful in the dynamic validation of expert systems. However, for realistic testing, data from real-life cases are often indispensible. Another tool is the dynamic analyser, also known as the execution flow summarizer. A dynamic analyser adds instrumentation statements to a computer program in order to collect information on how many times a statement is executed. A display part of the dynamic analyser prints a summarizing execution report [19]. A tool with similar usage as the dynamic analyser is the call-graph profiler [5]. 4.2 A tool for performance measurement The environment of software tools developed for the dynamic validation of HEPAR consists of a non-interactive batch version of the expert-system shell DELFI-2 which is able to use a database of patient cases as its input. This system produces a report containing the results for each individual patient. The report together with a file containing information about the final clinical diagnoses and the two intermediate conclusions concerning the patient is then processed by a program which produces a table summarizing the results. Figure 3 shows the overall structure of the validation environment. The tools collect information with regard to the number of correct, incorrect and unclassified patient cases concerning: 1. the type of hepatobiliary disease (hepatocellular or biliary-obstructive); 7 HEPAR knowledge base Database of patient cases Batch version of DELFI-2 Report Statistics program Final clinical diagnosis of patients Summarizing tables Rule-application analyser Rule-application report/graph Figure 3: Environment for dynamic validation. 2. the nature of the disorder (benign or malignant); 3. the final diagnosis. With regard to the final diagnosis, the system computes the average number of conclusions, whether the clinical diagnosis occurs as the conclusion ranked first, or among the list of alternatives generated. An example of such a table, produced after refinement of the HEPAR system using the software tools and a database with data from 82 patients from Leiden University Hospital, is reproduced in Table 1, which shows the results for the patients after refinement. The reader should note that the system does not reach 100% correctness, for reasons discussed in Section 3.1. To obtain some insight into the capabilities of the system in handling incomplete data, the batch version of the expert-system shell can select part of the patient data from the database. For example, the system is capable of selecting only data obtained from history and physical examination. An example of a table in which the results for incomplete data are produced using this environment is shown in Table 2. As can be seen in Table 2, the number of correct final diagnoses decreased considerably when the patient data entered into the system were more incomplete. However, the percentage of incorrectly classified cases did not increase Table 1: Diagnostic results of HEPAR for a population of 82 patients with hepatobiliary disease. Conclusion Correct Incorrect Unclassified Total n (%) n (%) n (%) n (%) Type of hepato- 74 (90) 4 (5) 4 (5) 82 (100) biliary disorde Benign/malignant 78 (95) 4 (5) 0 (0) 82 (100) nature of disorder Final diagnosis 71 (86) 8 (10) 3 (4) 82 (100) Clinical diagnosis 76 (93) 3 (4) 3 (4) 82 (100) among conclusions 8 Table 2: Assessment of the effects of incompleteness of information on the diagnostic conclusions of the system, for a database of 82 patients with hepatobiliary disease. Conclusion Correct (%) Incorrect (%) Unclassified (%) A B C A B C A B C Type of hepato- 90 90 50 5 5 0 5 5 50 biliary derangement Benign or malignant 95 95 94 5 5 6 0 0 0 nature of disorder Final diagnosis 86 49 24 10 9 10 4 43 66 A: All available data presented to system. B: Only data concerning symptoms, signs, haematology and bloodchemistry (no data from ultrasound or serology presented). C: Only data from medical interview and physical examination. significantly, only the percentage of unclassified cases did. Tables such as presented here were used by the hepatologist as an indication of the effects of changes to the knowledge base. An accompanying textual report provided information for each individual patient, and gave also information about the rules applied in deriving the final conclusions. This information served as a point of departure for a more in-depth study of the reasoning behaviour of the system. 4.3 Dynamic analysis of the HEPAR knowledge base In the previous section, we have discussed how the study of the results of the HEPAR system for individual patient cases, has been employed for refining the diagnostic quality of HEPAR. A second source of information that has been used for the refinement of the system, was the contents of the HEPAR knowledge base itself, by studying the overall behaviour of the system when provided with a complete database with patient data. These tools bear some resemblance with a dynamic analyser as described in Section 4.1. In order to obtain infor- mation concerning the frequency of rule application over a given database of patient cases, the testing environment discussed in Section 4.2 includes a collection of tools which uses the report produced by the batch version of the expert system for a database of patient cases. The following results are produced by these programs: • An enumeration of all production rules used, with for each rule information about how often it has been used for a given database; • An overview of the frequency distribution of the rule application, both in textual and in graphical form. Figure 4 which was automatically produced by the environment, contains the results after refinement of the HEPAR knowledge base for the 82 patient cases from Leiden University Hospital. Most production rules (72 of about 500 rules contained in HEPAR) were applied only once. The accompanying textual form, which is reproduced in Figure 5, shows that from the rules that were applied several times, those with highest frequency were applied to conclude about the intermediate hypotheses. Only few, one to three, production rules were applied several times to reach a final conclusion. Because we have tried to obtain a rulebase 9 4 8 12 16 20 24 28 32 36 40 44 0 16 32 48 64 80 absolute frequency number of rules Figure 4: Rule-application bar graph for 82 patients. in which at most a few rules will succeed for a given case, the report does not provide results concerning failed rules. The reports were studied by the hepatologist involved in the project as another source for the refinement of HEPAR. 5 Discussion Recent knowledge-engineering methodologies place considerable emphasis on the development of conceptual models. A suitable conceptual model may be of real help in designing and im- plementing an expert system, as was also observed in the development of the HEPAR system, where a diagnostic problem-solving strategy was used as the basis for problem decomposition and structuring of the knowledge base. However, in many fields of medicine, the development of an expert system is only possible with sufficient experimental feedback, for which software tools are required. Some other software tools supporting the building of rule-based expert systems have been developed in the past. Teiresias was an experimental tool that assisted in the refinement of rule-based expert systems by interacting with the user in the analysis of the conclusions concerning single cases, applying meta-knowledge about the rulebase [4]. Although such an analysis is certainly useful, an approach as embodied by Teiresias does not give information about how well the system performs over a database of cases. Seek is a system that automatically suggests generalizations and specialization of production rules, based on the analysis of the success and failure of rules on processing case data [17]. This system is more in the line of our approach. The broadness of the domain of hepatology, and the amount of patient data incorporated in the HEPAR system, suggest that automatic tech- niques as provided by Seek are as yet not powerful enough to be applied for refining a system like HEPAR. The parameters used as a point of departure for the refinement of the HEPAR knowledge base are only a few of the many that are possible. Another elegant example of 10 FREQUENCY #RULES RULE NUMBER 1 72 161, 500, 510, 660, 700, 720, 790, 800, 860, 900, 930, 950, 980, 1100, 1140, 1150, 1330, 1340, 1370, 1410, 1480, 1540, 1610, 1630, 1710, 1730, 1820, 1980, 2000, 2010, 2030, 2050, 2140, 2240, 2250, 2380, 2530, 2680, 2690, 2750, 3011, 3020, 3022, 3090, 3100, 3101, 3120, 3160, 3171, 3192, 3220, 3340, 3370, 3425, 3430, 3450, 3460, 3470, 3520, 3530, 3610, 3780, 3820, 3870, 3902, 3930, 3960, 4000, 4111, 4122, 4162, 4260 2 26 110, 120, 130, 440, 460, 470, 740, 830, 1090, 1130, 1490, 1940, 1950, 2130, 2180, 2370, 3010, 3041, 3110, 3390, 3410, 3427, 3428, 3440, 3920, 3940 3 30 111, 112, 131, 132, 140, 190, 230, 240, 560, 580, 600, 620, 1220, 1900, 1970, 2060, 2080, 2110, 2170, 2220, 2280, 2310, 2720, 3190, 3350, 3970, 4230, 4240, 4250, 4252 4 7 260, 341, 890, 1580, 2160, 3080, 4320 5 8 1550, 1560, 2120, 2350, 3001, 3060, 3102, 3420 6 7 160, 300, 2330, 3050, 3070, 3900, 3903 7 11 90, 100, 332, 1210, 1570, 2200, 2320, 2340, 2970, 3000, 3910 8 4 270, 290, 340, 350 9 2 150, 180 10 3 250, 4080, 4310 12 2 540, 3103 18 2 220, 4020 22 1 370 26 1 5010 27 1 5000 30 1 5030 36 1 711 45 1 5020 Figure 5: Textual rule application form. knowledge-base refinement has been proposed by Adlassnig and Scheithauer in the context of the CADIAG-2/PANCREAS system [1]. They have used ROC curves for the optimal adjust- ment of the internal classification threshold in this expert system. The technique is, however, not applicable in the HEPAR system, because here failure of classification is the result of logical falsification, and not of the failure of reaching an internal threshold. Most current expert-system shells and expert-system builder tools do not offer facilities supporting the refinement of an expert system by dynamic validation. In the development of the HEPAR system, we have therefore developed a collection of simple software tools that provide useful information for the refinement of the system. These tools have also been used in two successive final validation studies of HEPAR [10, 12]. Although there are many ways in which these tools can be improved, these validation studies would not have been possible without the assistance of these tools. In our opinion, future software tools for building expert systems should offer a wider range of facilities for the detailed analysis, verification and validation of an expert system than is currently provided. References [1] K.P. Adlassnig and W. Scheithauer, Performance evaluation of medical expert systems using ROC curves, Computers and Biomedical Research 22 (1989) 297-313. [2] A. Anjewierden, J. Wielemaker, C. Toussaint, Shelley – Computer aided knowledge engineering, in: B. Wielinga, J. Boose, B. Gaines, G. Schreiber and M. van Someren, eds., Current Trends in Knowledge Acquisition (IOS Press, Amsterdam, 1990) 41-59. 11 [3] J. Breuker and B. Wielinga, Models of expertise in knowledge acquisition, in: G. Guida and C. Tasso, eds., Topics in Expert Systems Design (North-Holland, Amsterdam, 1989) 265-295. [4] R. Davis and D.B. Lenat, Knowledge-based Systems in Artificial Intelligence (McGraw- Hill, New York, 1982). [5] S.L. Graham, P.B. Kessler and M.K. McKusick, Gprof: a call graph execution profiler, in: Proceedings of the SIGPLAN’82 Symposium on Compiler Construction, SIGPLAN Notices 17 (1982) 120-126. [6] G. Guida and C. Tasso, Building expert systems: from life cycle to development method- ology, in: G. Guida and C. Tasso, eds., Topics in Expert Systems Design: methodologies and tools (North-Holland, Amsterdam, 1989) 3-24. [7] J. Hilden and J.D.F. Habbema, Evaluation of clinical decision aids – more to think about, Medical Informatics 15 (1990) 275-284. [8] C.A. Kulikowski and S.M. Weis, Representation of expert knowledge for consultation: the CASNET and EXPERT projects, in: P. Szolovits, ed., Artificial Intelligence in Medicine (Westview Press, Boulder, 1982) 21-56. [9] P.J.F. Lucas, Knowledge Representation and Inference in Rule-based Systems. Centre for Mathematics and Computer Science, Report CS-R8613, Amsterdam, 1986. [10] P.J.F. Lucas, R.W. Segaar, A.R. Janssens, HEPAR: an expert system for the diagnosis of disorders of the liver and biliary tract, Liver 9 (1989) 266-275. [11] P.J.F. Lucas, A.R. Janssens, Development and validation of HEPAR, an expert system for the diagnosis of disorders of the liver and biliary tract, Journal of Medical Informatics 16 (1991) 259-270. [12] P.J.F. Lucas, A.R. Janssens, Second evaluation of HEPAR, an expert system for the diagnosis of disorders of the liver and biliary tract, Liver 11 (1991) 340-346. [13] P.J.F. Lucas, L.C. van der Gaag, Principles of Expert Systems (Addison-Wesley, Wok- ingham, 1991). [14] P. Matzen, A. Malchow-Møller, J. Hilden, C. Thomsen, L.B. Svendsen, J. Gammelgaard, E. Juhl, Differential diagnosis of jaundice: a pocket diagnostic chart, Liver 4 (1984) 360- 71. [15] E. Motta, T. Rajan and M. Eisenstadt, A methodology and tool for knowledge acqusition in KEATS-2, in: G. Guida and C. Tasso, eds., Topics in Expert Systems Design (North- Holland, Amsterdam, 1989) 297-322. [16] E. Motta, T. Rajan, J. Domingue and M. Eisenstadt, Methodological foundation of KEATS, the knowledge engineer’s assistant, in: B. Wielinga, J. Boose, B. Gaines, G. Schreiber and M. van Someren, eds., Current Trends in Knowledge Acquisition (IOS Press, Amsterdam, 1990) 257-275. [17] P.G. Politakis, Emperical Analysis for Expert Systems (Pitman, London, 1985). 12 [18] R.E. Shannon, Systems Simulation: the art and science (Prentice-Hall, Englewood Cliffs, New Jersey, 1975). [19] I. Sommerville, Software Engineering (Addison-Wesley, Wokingham, 1992). 13 
work_345bu7ezubahfezuj3appo2jgy	----	4th Global Business Research Congress (GBRC - 2018), Vol.7-p.47-51 Kizmaz, Deveci, Kilic _____________________________________________________________________________________________________ DOI: 10.17261/Pressacademia.2018.854 47 PressAcademia Procedia AN EXPERT SYSTEM APPROACH FOR THE INTERNAL AUDIT OF ISO 9001: 2015 DOI: 10.17261/Pressacademia.2018.854 PAP- V.7-2018(7)-p.47-51 Emine Kizmaz1, Esma Deveci2, Huseyin Selcuk Kilic3 1Marmara University, Goztepe Yerleskesi 34722 Kadıköy, Istanbul, Turkey. eminekizmaz@marun.edu.tr, ORCID: 0000-0002-0009-6175 2Marmara University, Goztepe Yerleskesi 34722 Kadıköy, Istanbul, Turkey. esmadeveci@marmara.edu.tr, ORCID: 0000-0002-6601-9922 3Marmara University, Goztepe Yerleskesi 34722 Kadıköy, Istanbul, Turkey. huseyin.kilic@marmara.edu.tr, ORCID: 0000-0003-3356-0162 To cite this document Kizmaz, E., Deveci, E., Kilic, H. S. (2018) An expert system approach for the internal audit of ISO 9001: 2015. PressAcademia Procedia (PAP), V.7, p.47-51. Permemant link to this document: http://doi.org/10.17261/Pressacademia.2018.854 Copyright: Published by PressAcademia and limited licenced re-use rights only. ABSTRACT Purpose- This research proposes an expert system approach for the internal audit of ISO 9001: 2015. Methodology- The International Organization for Standardization (ISO) 9001 Certificate is one of the main indicators of a good organization. One of the important steps of providing ISO 9001 quality management system is the internal audit. Depending on its importance, this study aims to facilitate the internal audit process with respect to ISO 9001: 2015 accreditation via Expert System approach. Expert systems are knowledge based systems which can be as successful as human experts in the solution of complicated problems by utilizing the expertise and knowledge of experts. However, rule based approach is used in the proposed expert system. Findings- The internal audit evaluations and scoring reports regarding ISO 9001: 2015 directives are obtained as a result of the expert system. Conclusion- The study provides an expert system for internal audit of ISO 9001: 2015 which is essential for all the organizations. With the proposed expert system, the organizations will be able to operate the internal audit process in a systematic way without the need of an expert. Moreover, the deficiencies that do not comply with ISO requirements can be easily identified. Keywords: Expert system, internal audit, ISO 9001: 2015 JEL Codes: L15, L50, O31 1. INTRODUCTION Expert systems are computer programs with intelligence and knowledge that resemble field experts in the solution of problems in a particular complex area. Initial work on expert systems began in the late 1950s, and nowadays many areas are being used, including medicine, geology, mathematics, chemistry, computer technology, management and military. The fact that the speed of change in technology and science is high today has brought the needs of experts in various fields to the highest level. The intense competition environment requires the right information, fast and at the lowest possible cost. The fact that the number of experts working on quality is low, the time it takes for quality specialists to grow up and the cost of operating these people is costly, has led to the idea of developing expert systems in this field. The objective of this study is to suggest an expert system approach for facilitating the internal audit process with respect to ISO 9001: 2015 QMS. ISO 9001 QMS Standardization is one of the standards of quality assurance system that has been widely accepted today. Implementation and continuity of ISO 9001 QMS require expertise for all organizations which are producing products and services. The continuity of the quality system established in the organizations must be ensured. It is important that internal audits are carried out to ensure that the ISO 9001 structure is applied at this stage and the sustainability of the ISO 9001 organization is ensured. In addition, these internal audits determine the deficiencies in the system, but there is no study to measure the audited ISO 9001 QMS sensitivity and system success. ISO 9001: 2008 QMS Standards are updated to ISO 9001:2015 QMS Standards. The firms must adapt their existing QMSs according to the standards of ISO 9001: 2015 until September 15, 2018 (URL 1). In this study, an expert system approach for measuring the success of the ISO 9001 QMS items and the system in general is proposed. With this expert system approach, internal audits can be performed more quickly and the system's performance can be measured. The remainder of this study is organized as follows: “Section 2” includes the literature review. “Section 3” provides the data and methodology. Findings are provided in “Section 4” and finally “Section 5” contains the conclusion with the references following. mailto:pinaryildiran@marun.edu.tr mailto:huseyin.kilic@marmara.edu.tr 4th Global Business Research Congress (GBRC - 2018), Vol.7-p.47-51 Kizmaz, Deveci, Kilic _____________________________________________________________________________________________________ DOI: 10.17261/Pressacademia.2018.854 48 PressAcademia Procedia 2. LITERATURE REVIEW Literature review of this study is based on two parts. The first one is Quality Management System (QMS) and the other one is Expert Systems (ES). Hence, the related studies about these two main topics are as below. 2.1. Quality Management System (QMS) The concept of quality has been used since ancient times, especially after the industrial revolution, with the increase of mechanization, the production style has shifted from workshop type production to factory production and mass production. With the change of many inhabitants living, the quality conceptually emerged in the 19th century (Montgomery, 2009). Many developments happened related with quality. Juran published Quality Control Handbook (1951) and separated the quality into two components: first one is quality of design and the other one is quality of conformance (Reeves and Bednar, 1994). Deming and Juran went to Japan to provide quality-related lectures. Deming pointed out that many problems arising from production originated from the process, and that this could be controlled using statistical methods. Juran set up a managerial approach to quality control and focused on project- based teamwork and client satisfaction. Crosby revealed the concept of “zero error” (Günaydın, 2001). Moreover, at the end of the 1980s, the automotive industry started to apply Statistical Process Control (SPC). The Malcolm Baldrige National Quality Award (1986) for American companies and the European Quality Foundation (EFQM) quality award for European companies began in 1992 (Boran, 2000). Regarding the quality management systems, International Organization for Standardization (ISO) is the world's grand developer of quality standards. ISO, was set up in Geneva, Switzerland in 1947. The International Standards which ISO develop are very beneficial. This is because it conduces to making the development, manufacturing and supply of products and services more effective, confident and cleaner (Shouman et al., 2009). Gotzamani and Tsiotras (2001) emphasize that ISO 9000 standards enable companies to design and utilize an efficient and active quality system with the consideration of continuous improvement and adaptation. Arditi and Gunaydın (1997) point out that the ISO 9000 standards are composed of two basic concepts including quality management and quality assurance. However, besides implementing ISO standards to the company, the internal audit is also important. Internal audit is a function that controls and analyzes the processes and procedures within the organization, suggests improvement, performs risk analysis, and serves all the interest groups both inside and outside the company (Aslan and Özçelik, 2012). According to Türedi (2012), total quality management and internal auditing are two mutually reinforcing management elements that progress through innovations in achieving the company's goals and they are in interaction and help each other to develop. Hence, depending on its importance, it is focused on internal audit in this study. 2.2. Expert Systems (ES) Expert systems are problem-solver or decision-maker software packets that could achieve a stage of performance commonly in attenuated problem areas. Expertise is transferred to the computer from experts. This transmitted information is kept in the computer. Then users operate the computer for particular suggestions as required. The expert system inquiries about cases and could make inferences and attain to a particular result. Afterwards, such as a human advisor, it informs one who is not a specialist and declares, if required, the logic underlying the suggestion (Aronson et al., 2005). Initial work on the expert system began towards the end of 1950s and the leading system in this field is DENDRAL. This system was initiated in 1965 by E. Feigenbaum and colleagues at Stanford University in the USA to provide assistance in the identification of the structure of an organic compound by mass spectrogram and raw chemical formula (Altuntaş and Çelik, 1998). Expert system methodologies are based on eleven parts including knowledge-based systems, rule-based systems, fuzzy expert systems, neural networks, case-based reasoning, object- oriented methodology, intelligent agent systems, system architecture, modeling, database methodology and finally ontology (Liao, 2005). Regarding the structure of the expert systems, it can be stated that there is not a standard structure and the main reason why expert systems cannot have a standard structure is that each structure is more suitable for another application than the other. In general, although expert system structures differ from each other, the components that make up the expert system structure can be described as knowledge base, working memory, inference engine, explanatory system, knowledge acquisition facility and user interface (Fidan, 1994). 3. DATA AND METHODOLOGY The aim of this study is primarily to develop an expert system that can be used by all companies. This expert system is a program established within the scope of ISO 9001: 2015 QMS, to use in internal audits and to measure achievement. With this program, the deficiencies of the QMS used can be identified. The accuracy of the output of this program depends on the truthfulness of the answers to the questions to be asked and their objectivity. For this expert system, some rules were developed and all the rules were built on each part of the QMS. The weight points and the scoring structure were created and, as a result, the scoring structure was combined with the expert system. Thus, an expert system structure consisting of knowledge base, working memory, extraction engine, explanatory system, information acquisition facility and user interface was established and a system supporting the companies was established. Knowledge acquisition facility is based on ISO 9001:2015 QMS Standards in proposed expert system approach. The knowledge base names and rule numbers are shown in Table 1 as follows. 4th Global Business Research Congress (GBRC - 2018), Vol.7-p.47-51 Kizmaz, Deveci, Kilic _____________________________________________________________________________________________________ DOI: 10.17261/Pressacademia.2018.854 49 PressAcademia Procedia Table 1: Knowledge Bases and Number of Rules These rules constitute the knowledge base of the expert system. The questions created by these rules can be answered in three ways as “Yes, No or Partially”. These responses are retrieved by the inference engine for use and sent to the working memory. Later the inference engine carries out the processes necessary for interrogating these knowledge bases by searching through the forward chained control strategy. The inference engine here, in order to initiate a rule, searches the case list for cases that satisfy the condition part of the rule, and performs the result part of that rule if it finds these cases. The designed system consists of two parts which are intertwined; The Internal Audit Part (1) and The Success Measurement (Scoring) Part (2). ISO 9001 QMSs, expert systems and internal audit related literature search and consulted experts have played an important role to identify the problem. The questions are prepared by the information obtained from experts and the accreditation of ISO 9001: 2015. Depending on the answers received from the user, the question flow can change. In internal audit part, there is a knowledge base that belongs to 4 processes of organizations. Each knowledge base is divided into the questions determined within the scope of ISO 9001: 2015 standards. A number of different questions are asked to the user regarding the auditing of the system under each process heading. The important point here is that the expert system determines the questions that the user will ask in the next step, according to the answer given by the user. In other words, the “No” answer given by the user to some questions may skip the next question and lead to other questions so it makes the expert system more complex. Depending on the problem type, the user is offered a choice of two kinds of answers; Yes / No or Yes / Partially / No. The approach controlled by the four process expert systems does not have the same designation. At this point, each Weight Score (WS) of processes is determined by information from experts and by literature review. The weight scores of these operations are given in Table 2 below. Table 2: Knowledge Bases and Weight Scores Here, each process is determined according to the importance of the Weight Score (WS) in the process and the Quality Management Standards. The Process Score (PS) for each process is calculated as follows. Process Total Score (Process TS) = 10 * Y (Yes) + 5 * P (Partially) + 0 * N (No) Process Maximum Score (Process MS) = number of questions answered * 10 Process Success Percentage (Process SP) = Process TS * 100 / Process MS Process Score (The value of the WS that corresponds to the Process' success) = Process WS* Process SP / 100 There are explanations in the software section of the program with respect to scoring information (section where constants and variables are assigned), functions, rules, score calculations, and module display sections found in this expert system. These explanations provide information about the processes that ensure the program to run and provide its results. This program is prepared using JAVA programming language. The implementation of the proposed expert system approach under the ISO 9001 QMS in an enterprise is accomplished. The questions that is asked by the proposed expert system during the audit are directed to the relevant people and the responses received are entered into the expert system. Knowledge Base Name I SO 9001 Quality System Submission subject to Audit Number of Rules Process 1 Mannagement Responsibility Process 24 Process 2 Quality Management System Process 26 Process 3 Production (New Product) Process 38 Process 4 Purchasing Process 20 Knowledge Base Name I SO 9001 Quality System Submission subject to Audit Weight Score Process 1 Mannagement Responsibility Process 30 Process 2 Quality Management System Process 30 Process 3 Production (New Product) Process 24 Process 4 Purchasing Process 16 TOTAL SCORE 100 4th Global Business Research Congress (GBRC - 2018), Vol.7-p.47-51 Kizmaz, Deveci, Kilic _____________________________________________________________________________________________________ DOI: 10.17261/Pressacademia.2018.854 50 PressAcademia Procedia 4. FINDINGS After the application the expert system in a company which is in the plastic sector, the related scores are obtained. Considering the 70 % score as the base for success. There are not the processes with success percentage below 70 %. However, since, system final score is 90.5 % so, it can be said that the internal audit of the company is successful and passes the required score threshold. Table 3: Summary Display of Implementation Results and Success Measurement of the Company 5. CONCLUSION Expert systems are knowledge based systems with software and hardware, which can be as successful as human experts solving complicated problems by using expertise and knowledge of experts. Expert systems have the benefits such as using time efficiently, reducing costs, increasing productivity, and reducing errors. The ISO 9001: 2015 standard anticipates continuous improvement, and also internal audit is one of the main processes that must be used for continuous improvement. In this study, the expert system approach is combined with the scoring technique to ensure that the quality systems of the companies are checked according to the ISO 9001: 2015 standards before the quality control and the necessary improvements are made. The proposed expert system is based on rules and knowledge. For this reason, questions are prepared with the information obtained from ISO and quality management experts and question lists of these processes were determined. Since the new ISO 9001: 2015 accreditation is based on the process, the four main processes of the organizations are discussed. These processes have been prepared by extensive research and consultation of experts. Management Responsibility Process, Quality Management System Process, Production (New Product) Process, Purchasing Process are processes that are integrated into the system. The proposed system can carry out the internal audit process without the need for the experts. This reduces the cost. Also, companies often try to keep the internal audit process as short as possible because they do not want to waste their workforce. Moreover, the system performance can be measured together with the scoring system. Hence, this system provides time saving, increases productivity and continuity. In the further studies, the expert system approach proposed in this study can be improved or modified depending on technological developments and developments in the ISO QMS processes. Accordingly, more compatible systems can be used instead of the JAVA programming language depending on the developments that may occur in the software system. REFERENCES Altuntaş, E., Çelik, T. (1998). Yapay zekanın tarihçesi. Otak Yayıncılık. Arditi, D., Gunaydin, H. M. (1997). Total quality management in the construction process. International Journal of Project Management, 15(4), 235-243. Aronson, J. E., Liang, T. P., Turban, E., (2005). Decision support systems and intelligent systems. Pearson Prentice-Hall. Aslan, S., Özçelik, H. (2012). İç denetim ve toplam kalite yönetimi ilişkisi. Uluslararası Yönetim İktisat ve İşletme Dergisi, 5(10), 109-119. Boran, S. (2000). Toplam kalite yönetimi. Fidan, S. (1994). Endüstri mühendisliğinde uzman sistemler ve proje yönetim yazilimi seçiminde bir uzman sistem yaklaşimi. Doctoral Dissertation Gotzamani, K. D., Tsiotras, G. D. (2001). An empirical study of the ISO 9000 standards’ contribution towards total quality management. International Journal of Operations & Production Management, 21(10), 1326-1342. 4th Global Business Research Congress (GBRC - 2018), Vol.7-p.47-51 Kizmaz, Deveci, Kilic _____________________________________________________________________________________________________ DOI: 10.17261/Pressacademia.2018.854 51 PressAcademia Procedia Günaydın, H. M. (2001). Toplam kalite yönetimi. Mimarlar Odası İzmir Şubesi. Liao, S. H., (2005). Expert system methodologies and applications—a decade review from 1995 to 2004. Expert systems with applications, 28(1), 93-103. Montgomery, D. C. (2009). Introduction to statistical quality control. John Wiley & Sons (New York). Reeves, C. A., & Bednar, D. A. (1994). Defining quality: alternatives and implications. Academy of management Review, 19(3), 419-445. Shouman, M., Eldrandaly, K., & Tantawy, A. (2009). Software quality assurance models and expert systems. Türedi, S. (2012). İç kontrol sistemi ve toplam kalite yönetimi ilişkisi. Uluslararası Alanya İşletme Fakültesi Dergisi, 4(1). URL 1: https://www.iso.org/iso-9001-revision.html (11.12.2017) https://www.iso.org/iso-9001-revision.html 
work_34b345ljjbbzpisupoqauu7yee	----	Expert Systems and the Emergence of Teledesign Pr e- Pr int Expert Systems and the Emergence of Teledesign Anthony Crabbe, Dept of Design, Nottingham Trent University. anthony.crabbe@ntu.ac.uk Preprint of article published in Design Studies 25(4): 415-423 ABSTRACT This paper considers the extent to which the amateur use of expert systems for home design challenges traditional views of the design process. The issues are examined in the context of competing definitions of design. The emergence of a design process characterised as "teledesign" is then considered, wherein retailers provide a CAD/CAM service to consumers, allowing the latter to use expert systems to modify template designs and get products fabricated to their own specifications. Such a system may be seen to empower consumers as designers, rather than just selectors of products and would differ considerably from established paradigms of design, manufacture and consumption, such as that given by Baudrillard. Keywords: Expert systems, design theory, architectural design, design process, design models. 1. Introduction An "expert system" is a computer program which performs many functions normally done by human experts. Expert systems generally comprise of an interface between the user and a knowledge system that is built out of the experiences of expert practitioners in the given field. Generally, the interface comprises the communication hardware and software needed to enable the user to interact with the computer, and it may also include other peripherals such as sensing or manufacturing equipment. The knowledge system comprises of two main parts. The first is the "knowledge base", which stores expert human knowledge on a subject in a formal and hierarchical way, containing definitions like "A = list 7", and rule-based statements, such as "if A and B, then C, not D". The second part is the "inference engine", a program which determines the consequences of a user action or command by handling each rather like a query, searching for the closest match between requested actions and pre-set rules. By such means (usually invisible to the user), the engine finds and applies the closest "expert" solution for each user action. The characteristics of expert systems may be observed in recent word processor applications, where the program goes even further than performing the "craft" role of the typesetter and plays a "professional" role in for instance, editing the grammar of the author’s original text input. mailto:anthony.crabbe@ntu.ac.uk Pr e- Pr int Other examples of expert systems are medical diagnosis programs and drafting of legal contracts. In design software, professional expertise is often incorporated within a straightforward drafting application, such as 3D Studio, which allows the mechanical behaviour of the model structure to be analysed by application of engineering formulae, that provide a simulation known as "finite element analysis". The use of such computer applications is now pervasive in the professional design world, and the expertise is often disguised, being embedded as just another program function, or a command on the menu bar. The development of new expert design applications raises three major issues concerning the future nature of the design profession. The first concerns the extent to which expert systems are evolving either as assistants or as competitors to the professional. The second issue concerns the extent to which use of expert systems may alter notions of what the design process involves. The third concerns the ways in which the retailer may come to exploit the potential of expert systems as a customer service. These issues will now be examined by general reference to the medical and design professions and by specific reference to design of the home. 2. Expert systems in medicine Medicine is a profession where expert systems are developing vigorously. Whole teams of doctors are developing expert databases for their own consulting areas, others are aimed at a much broader healthcare market, providing doctors, auxiliaries and even patients with such things as a preliminary diagnosis. 5GL Doctor is an example of the latter, being aimed at markets in the developing world, where access to a doctor may be difficult (see Figure 1). Accordingly, the system may be consulted at differing levels of expertise. At present, the advised role of the expert system in medicine is that of an assistant in practice management, finding data and second opinion. The main reservation within the profession seems to concern misuse of the assistant, for instance allowing it to make decisions in place of the professionals who properly carry the responsibility and indemnification. Pr e- Pr int Figure 1 There appear to be mixed feelings about the future effect of expert medical systems, particularly in light of their effect on the current roles played by professionals., The expert application can draw on a potentially limitless databank of case history and unlike the doctor, can give an immediate quantitative calculation of the probability that its diagnosis is accurate. With the right peripherals, the application might allow someone with lesser qualifications to conduct many examinations. In time, such developments could limit the general practitioner's role much more to surgical procedures, or to matters of strategy, such as practice management. To the optimists, this would represent only a change in role due to technological innovation, to the pessimists, a threat to job status and tenure. 3. Expert systems and design practices Expert systems are also beginning to affect the role played by professional designers. Computer technology encourages many more businesses to design and print their own materials, web pages, CDs and so forth. The graphic basis of design activity makes it particularly suitable for the functioning of home computers and thus ripe for further market penetration by the software industry. Following the boom in the home Do-It- Yourself market, there are now many cheap DIY programs offering the general public the opportunity to design much of their home environment from interiors and gardens to the house itself. Furthermore, influential retailers like Homebase, are actively encouraging this trend by creating their own software and catalogues of digital images that allow the customer to review products in a simulation of their own home and make their own Pr e- Pr int design choices. As witnessed by the popularity of the suffix "Pro" on many of these software applications, the marketing message being sent to home computer users is that computer technology will empower them to rapidly assume the creative role of professionals, in fields from which the untrained and unequipped would normally be excluded. Yet, these same applications are also used by professionals and could be seen to enhance their existing advantages over amateurs. So it is not easy to judge whether this computer empowerment is set to take work away from design professionals (as word processing did from printers), democratise the activity of designing (in an arts & crafts kind of sense), or merely maintain the creative gap between the professional and the DIY enthusiast. These questions oblige us to consider further the kinds of functions performed by those called "professional" designers, before examining how far they may be taken over by non-professionals. To characterise, even summarily, the activities undertaken by professional designers, is much harder than it would be for law or medicine. The design profession embraces a wide range of specialists, from those requiring chartered and indemnified status such as architects and engineers, to those who may simply designate themselves "designers", such as fashion and interior designers. Even well into the "Designer '80’s" the terms "commercial" and "applied" art were generally used to describe what most today would call design. Looking at design education in Britain, as recently as the 1970's, the study of design history was just a branch of art history. The teaching of design in higher education embraces a number of differing approaches, ranging from the arts and crafts like approach of the traditional art college, to the engineering and management emphasis of the traditional university. As reviewed by John Walker, recent commentators have sought to define design in conflicting ways, some in terms of arts and crafts tradition, some in terms of a common activity pertaining to manufacturing industries and yet others in terms of a generalised problem solving activity, not even restricted to professionals. To these views should also be added an even more pervasive one, that design amounts to a process of "valorising" utilitarian objects (in the sense of "adding" some hitherto unrecognised value to them). British versions of this idea of valorisation are to be found in ideologies as opposite as those found in the Arts and Crafts movement of the 19th Century and the Thatcher government of the 1980’s. At face value, the Thatcher government view appeared to be that design status (most obviously evidenced in "designer label" goods) added to the price that could be asked for them, and hence the profit gained from them. However, Thatcherites also seemed to have seen value addition as another instrument to be used in their bid to create a "market culture", where tensions between business and citizenry could be reconciled through a shared value system of consumer expectations and rights. In linking the role of design to social well being, Thatcherism continued to follow the precedents set by the more left-wing approach of the Pr e- Pr int Arts and Crafts movement. This influential international movement chose to valorise craft products by reference to the self-actualisation and personal fulfillment attained by those following craft processes. Such attainments were seen by the likes of William Morris to provide the bases for developing a harmonious social system, such as that imagined in News from Nowhere, where the profit motive became the least important driver of the production process. By comparison, the contemporary Italian version of valorisation seems to lay more stress on the values added by a more wilful, individualistic form of creativity, commonly associated with the activities of the artist. Such a view clearly encourages us to regard the outcomes of design as creations, rather than discoveries and to appreciate the individual and personal qualities of designed products, rather than their objective necessity and reproducibility. This bewildering array of perspectives suggests a fundamental dichotomy between design on the one hand as a description for a set of activities common to anyone involved in making and on the other hand, as a description for professional activities found in only certain types of commercial production. This makes it considerably harder to determine what role expert systems have to play in design, compared say to Medicine, and hence to predict what impact they may have on those practising design professionally. An instructive approach then, is to focus on just one example, the design of the home. The example is useful because it covers a multitude of products designed by professionals with very different status, ranging from the architect to the gardener. It is also useful because we have historical precedents for the homeowner taking over the functions of the professional. 4. Expert systems for DIY design of the home In addition to the store software mentioned earlier, there are now a number of inexpensive home computer software applications for house and garden design. 3D Dream House from Data Becker is typical. A simple command interface allows the novice to design either a new house, or simulate their existing one with a high degree of accuracy (see Figures 2-4). Expert knowledge is built into the programming so that correct parameters are automatically set for the following: plan drawing, perspective rendering, daylight simulation, wall and ceiling thickness, stair and roof geometry, window and door placement. For the interior, the consumer can specify from a wide range of colours, textures as well as a catalogue of furnishing items. Since these items are included as scaleable bitmaps, it would only be a very short step to manufacturers providing their own catalogues of such items for the consumer to try out in their simulated home. Pr e- Pr int Figure 2 Figure 3 Pr e- Pr int Figure 4 It would appear to be only the limitations of home computer performance and program cost that prevent inclusion of more detail such as plumbing and services, specific building regulations and materials costings. In other words, we may assume that we will shortly see expert programs which permit the homeowner to quite easily generate the complete design of a home that could be validated by regulatory authorities, without any professional input. Architects may retort that most of the results would lack the creativity, judgement and aesthetic sensitivity, which they alone could bring to the scheme and that just as with in- house graphics, there would be a proliferation of ill-conceived and unoriginal designs. This would be to emphasise a) the necessity of emotional and affective qualities in designing and b) the lack of such qualities in the mechanical operations of current computers. However, this view does not negate the great practical assistance dumb machinery can provide to human users. Following the example of graphic design, it is notable that applications like Photoshop allow a level of image manipulation and exploration that far exceeds the capacity of any designer to match using hand techniques. So expert systems present many, often unforeseen, possibilities for human review. In which case it could be argued that the primary difference remaining between the professional and DIY user of expert software, is the extent to which each has learnt to choose and act "well" from the solutions offered by the machine. As the argument shifts into consideration of design merit, we see how important the definitions of design mentioned earlier become in assessing the impact of DIY design. If, for instance, we elect to define design primarily in terms of a kind of problem solving Pr e- Pr int activity - as medical diagnosis might be chararcterised - then the only thing that really seems to separate professional from non-professional design is whether it occurs in the commercial or domestic arenas. However, doctors would be quick to affirm that proper diagnosis still depends on affects in the trained mind, like intuition and inspiration, to trigger innovative lines of enquiry, without which the science of diagnosis would never advance. This may explain the consensus in the medical profession that the expert system can never act as anything more than a passive assistant. Such a line of argument helps demonstrate the difficulty in choosing between views of design, either as problem solving, or as a problem solving/valorising activity. Those who subscribe to the former view can always argue that problem solving is not reducible simply to a set of formulaic inputs and outputs - the creative dimension is still essential and is based in human, rather than mechanical processes. Subscribers to the latter view may argue that the problem to be solved in design is broader still: it must address cultural and aesthetic needs and so the general notion of "function" in design must encompass both the utilitarian and aesthetic. Hence, work judged as failing to address this broader notion of function could fulfill some, but not all the requirements necessary to be counted as design "proper". These considerations make it easier to see the traditional strength of the valorising view, which helps professionals to protect their status by recourse to critical consensus. Whereas the amateur is usually designing for personal need, the professional has a vested interest in contributing through practice and commentary to the establishment of a recognisable discipline with its own rules and critical canon. Thus, the critical pitch seems very much tilted in favour of those who have established the critical canons. Yet, the significance of the non-professional contribution has often been promoted by professional champions. Historical examples are the rehabilitation of vernacular architecture by the likes of William Morris, and the artisan's aesthetic sense by Léger. The common factor linking these disparate examples is that they have come, by whatever means, to be seen as exemplars, both of products and creative processes. The primary limitation of the valorising view lies in the extent to which it naturally tends to exclude the more engineering based kind of design, so essential to technically sophisticated and mass manufactured products. Evidence of this kind of exclusion can be found in many production processes. For instance, the blueprints of many public buildings emerge from a loop where architects periodically review and amend their conceptualisations with consulting engineers. However, as the celebrity of architects like Norman Foster and Richard Rogers testifies, the credit for the cultural and aesthetic values of the final buildings is usually attributed to the creative/artistic vision of one outstanding architect, however much that individual may point to the effort of a multi- disciplinary team. Pr e- Pr int A most striking feature of current expert systems for design is that they provide no form of guidance on aesthetic principles, such as proportion, composition and colour. Unlike their historical antecedents, the hand and the pattern book, expert systems offer the novice no advice or means of exploring the different methods of composition and construction that have achieved exemplary status. It is then hard to imagine that an expert system would ever have an effect comparable to Palladio's Four Books on Architecture, which facilitated an entire amateur architectural movement amongst the landed classes of Britain and the United States in the 18th Century. In these books, the practical guidance about architectural elements presupposes that, for instance, the "proper" dimensions are those which have classic beauty as well as constructional or utilitarian function. Such an outlook remains embedded in the heart of modernist architectural practice, as in Le Corbusier's handbook for proportioning, The Modulor. Since theories of proportion can be presented by means of mathematical argument, it would not be hard to imagine the incorporation of a system like Modulor into a computer application like 3D Dream Home. On the other hand the justification for choosing the given mathematical method is clearly less objective. We have only to look at the present day schisms in the architectural profession between so-called "late" and "post" modernists to see how the incorporation of utilities like a proportioning system into the software would draw the DIY user even further into concerns that hitherto, had been largely professional. Palladianism seems to have emerged as a tendency exactly because the handbook Four Books on Architecture gave a wider audience access to design theory, especially the more intellectually able amateurs, who like Lord Burlington, might otherwise have continued to focus on their other interests, such as music. By comparison, current expert design systems seem utilitarian and much less useful as educational assistants, both because their knowledge base excludes cultural forms, and because they share the generic limitation, noted by some educationalists, of hiding their program arguments from the user, in the interests of functionality. Some architectural academics seek to overcome such limitations by exploiting the programs’ abilities to generate entire catalogues of possible solutions for the geometric relationships found in differing architectural types. By such means they hope to discover the "shape grammars" of architectural types like the Palladian Villa, from which they believe they could deduce rules which would rapidly lead the user to better design solutions. But the degree to which the "better" geometric solutions would be better architectural ones, is open to serious objections. As noted earlier, design problems usually require the designer to address the users’ cultural expectations, to some extent. So the assessment criteria for design success may be qualitative as well as quantitative. Public hostility in 1980’s Britain to the 1960’s high-rise dwelling type showed that however successful this type was in providing structural solutions for housing that were more cost and space efficient, the solutions were still examples of unsuccessful architecture, with many being demolished inside a fraction of their structural lifespan. It then remains to be seen what impact shape grammars will have on the expert systems Pr e- Pr int used by professional and amateur architects. The relative advantage of handbooks over expert systems is that they are created by experts whose interests reach beyond the performance of their own immediate tasks and out to the practice of others. This indicates a desire to lead the discipline, not just follow it. It is unlikely the same is true of authors of expert design systems, since they are in the main, anonymous consultants operating within a computer programming team. Seen from a market perspective, expert design systems look like another instance of the computer software industry trying to take market share off an established profession, by leading consumer demand away from professional decision making to the computer dependent kind. The professional architect may then feel that expert systems offer little threat and are unlikely to enable the amateur to produce work of comparable standard, any more than Four Books on Architecture enabled Burleigh’s Chiswick House to rival the success of it’s inspiration, Palladio’s Villa Rotonda. On the other hand, unlike handbooks, expert systems give their users the technical skills to rapidly generate schemes and make their own discoveries, free of the kind of prescription or burden of respect imposed by the handbook. It is to be hoped that this new functionality will, as in the shape grammar project, lead to new creative insights amongst practitioners of all kinds. It is to be feared that culturally empty expert systems will give the DIY enthusiast a limited understanding and false sense of confidence, which will lead to a proliferation of uninspiring design jobs. 5. Teledesign If only a minority of consumers have the time and inclination to use expert systems in place of professionals, expert systems still offer one more challenge to the professional designer, especially in the field of home furnishing. This is a more significant challenge since it involves both the marketing behaviour of suppliers and the shopping habits of the consumer. Tucked away in the files of 3D Dream Home is a folder of images labelled Ikea. This contains scaleable bitmap images of furnishings designed by that company. It is then easy to envisage the following scenario. There are shops and DIY stores that already offer a service whereby the shopper can provide details of say, their kitchen, to a sales assistant who can then use an easy to learn application, similar to 3D Dream Home, to make a virtual model of the kitchen. Guided by the shopper’s preferences, the assistant can drop in the images of various product ranges of kitchen fittings and indeed cost each ensemble to the customer’s satisfaction. The purchase of the final choice can then be fed into the supplier’s order system, which may in turn be linked to the manufacturer’s stock and production systems, giving both companies the benefit of "just in time" manufacture and sale. There are also after sales Pr e- Pr int benefits for manufacturer and retailer, in cases where the purchase is made electronically, rather than with cash. The companies can use the purchase transaction data to profile both their customers and each other in order to target their future marketing, exactly as is done in other stores, such as supermarkets. Going one step further, it is notable that the computer generated images of products can be easily re-dimensioned, re-coloured, indeed re-formulated, as can any images in a computer application. If the manufacturer constructs the images in a computer application that relates to his/her own design software, then these images, far from being bitmaps, could be the designs themselves that the customer modifies. Since the designer’s original software can have inbuilt expert functions, such as finite element analysis and costing, the customer’s modifications can be automatically appraised and relevant production parameters set. Because the design files can be generators for a CAD/CAM manufacturing system, the opportunity exists in the case of many home design products, for the customer’s final choice to be sent directly to the manufacturing line for fabrication as a bespoke item. Such a process could be characterised as "teledesign", a process where the consumer uses communications technology to execute the final design, remote from the design studio (represented in Figure 5). Indeed, technology for clothing design and manufacture is fast moving towards this point, with companies like Assyst and Lectra offering store-based or online software that allows customers to enter both their measurements and preferences into existing garment design templates. The customer may even design their own textile prints and the selected fabrics can then be marked and cut ready for sewing. Figure 5 Pr e- Pr int In this respect, expert systems seem set to have an even more profound impact upon the design profession than on medicine or law. Firstly, many designers may find themselves more engaged in the design of templates, analogous to those already found in presentation software like Powerpoint, where users may take a pre-designed graphic slide and from that adjust every element to their own needs. The final product is then created to the user’s design, following the designer’s earlier cues and hints – a system quite unlike Baudrillard’s famous characterisation of products as serial reproductions of the designers originating model. This suggests secondly, that teledesign may make the professional designer more contingent and less integral to the manufacturer’s needs. Thirdly, it suggests that the designer’s solutions will be less sacrosanct, the designer will be less of an arbiter of what is available to customers and more a facilitator of customer choice. Since few design specialisms involve indemnification, the consumer would be free to ignore professional advice in a way less likely found in medicine or law. Pictured in evolutionary terms, it would take a very strong initial design by a professional to survive without mutating into quite different forms driven by consumers’ desires. Such a production system could also lead to a different kind of valorisation of products, where the purchase of durable goods would no longer be just a matter of selection, but one of creation. There might even be parallels with the vision of News from Nowhere, where a technology (ironically, an industrial one) creates a production system that gives ordinary folk the power to create their own material and cultural environment, manifested in a sociable plurality of individual creations. However, this is only presuming as did Morris, that the creative impulse is common to most people, that they do not want to delegate creative responsibility to others (especially status figures) and thereby valorise their goods by association with a cultural milieu other than their own. Consider that Morris himself, the son of a Victorian brewery owner, aspired in thought and deed to the world views of both medieval knights and artisans. Yet, if e-commerce realises even part of its much vaunted potential, then it can be seen that expert systems would be essential tools, and teledesign even more prevalent. For instance, the consumer might store on his/her home computer, a virtual model of his home, created by expert software. This model would be linked to retailers’ catalogues, so that the consumer can review new items in situ, then customise and order his chosen design, all at his home computer. In such a future, do not be surprised to receive an unsolicited e-mail that reads "Valued Customer, our records show it is three years since you last re-furbished your kitchen. As you will see from the attached video file, our leading designers here at Ikeo have re-fitted your kitchen, using our latest Smørsgabord range, enhanced by a range of new paint finishes from…" Pr e- Pr int On the other hand, don’t be surprised to discover that it was an expert system which automatically loaded the new product ranges into the kitchen design file and the customer who took credit for the re-design. REFERENCES 1. Romiszoski, A. Artificial intelligence and expert systems in education: progress, promise and problems. Australian journal of Educational Technology, 3(1), 6-24 (1987). 2. Hillson SD. Connelly DP. Liu Y. The effects of computer-assisted electrocardiographic interpretation on physicians' diagnostic decisions. Medical Decision Making. 15(2):107-12 (1995, Apr-Jun). 3. Gardner RM. Lundsgaarde HP. Evaluation of user acceptance of a clinical expert system. Journal of the American Medical Informatics Association. 1(6):428-38 (1994, Nov-Dec). 4. Ridderikhoff J. van Herk E. A diagnostic support system in general practice: is it feasible? International Journal of Medical Informatics. 45(3):133-43 (1997, Jul). 5. Lovell NH. Celler BG. Implementation of a clinical workstation for general practice. Medinfo. 8 Pt 1:777 (1995). 6. Walker, JA. Design history and the history of design, pp28-32, Pluto Press, London (1990) 7. Jervis, S. The Penguin dictionary of design and designers, Introduction, Penguin, London (1984) 8. Bayley, S. In good shape: Style in Industrial Products 1900-60, Design Council, London (1979) 9. Papanek, V. Design for the real world: human ecology and social change. 2nd ed., Thames and Hudson (1985) 10. Pearman, H. Cautious optimism at Number Ten, Design , London, (March 1987) p459. 11. Morris, W. News from Nowhere and other writings, (1891), ed. Wilmer C. Penguin Books, London (1993) 12. Branzi, A. We are the primitives (1985), in Design Discourse: History, Theory, Criticisim. Ed. Margolin, V. University of Chicago Press, Chicago (1989). 13. Radice, B. Ettore Sottsass: a critical biography. Ch 3, Thames & Hudson, London (1993) 14. Léger, F. The machine aesthetic: The manufactured object, the artisan and the artist, 1924, reprinted in Benton T & Benton C. Form and function, Crosby Lockwood Staples, London (1975) 15. See e.g. Pawley, M. Norman Foster: a global architecture, Thames & Hudson, London (1999) 16. Palladio, A. The four books of architecture 1570, Trans. Ware I., Constable, London (1965) 17. Le Corbusier. The modulor. Trans de Francia P. & Bostock A. Faber, London (1954) 18. Op cit. note 1. 19. Stiny G, Mitchell W J, The Palladian Grammar Environment and Planning B: Planning and Design 5 5-18 (1978). 20. Madrazo L, The Concept of Type in Architecture, An Inquiry into the Nature of Architectural Form, Ph.D. Dissertation, University of Zurich, ETH Zentrum No. 11115 (1995) Epilogue. 21. http://www.assyst-intl.com and http://www.lectra.com 22. Baudrillard, J. The system of objects (1968), reprinted in Design after modernism: beyond the object ed. Thackera J. Thames and Hudson, London (1988) 
work_35b465fwlnexvbuebingsjly4y	----	doi:10.1016/j.eswa.2007.10.042 Available online at www.sciencedirect.com ARTICLE IN PRESS www.elsevier.com/locate/eswa Expert Systems with Applications xxx (2007) xxx–xxx Expert Systems with Applications Imbalanced text classification: A term weighting approach Ying Liu a,*, Han Tong Loh b, Aixin Sun c a Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong SAR, China b Department of Mechanical Engineering, National University of Singapore, 9 Engineering Drive 1, Singapore 117576, Singapore c School of Computer Engineering, Nanyang Technological University, Nanyang Avenue, Singapore 639798, Singapore Abstract The natural distribution of textual data used in text classification is often imbalanced. Categories with fewer examples are under-rep- resented and their classifiers often perform far below satisfactory. We tackle this problem using a simple probability based term weight- ing scheme to better distinguish documents in minor categories. This new scheme directly utilizes two critical information ratios, i.e. relevance indicators. Such relevance indicators are nicely supported by probability estimates which embody the category membership. Our experimental study using both Support Vector Machines and Naı̈ve Bayes classifiers and extensive comparison with other classic weighting schemes over two benchmarking data sets, including Reuters-21578, shows significant improvement for minor categories, while the performance for major categories are not jeopardized. Our approach has suggested a simple and effective solution to boost the per- formance of text classification over skewed data sets. � 2007 Elsevier Ltd. All rights reserved. Keywords: Text classification; Imbalanced data; Term weighting scheme 1. Introduction 1.1. Motivation Learning from imbalanced data has emerged as a new challenge to the machine learning (ML), data mining (DM) and text mining (TM) communities. Two recent workshops in 2000 (Japkowicz, 2000) and 2003 (Chawla, Japkowicz, & Kolcz, 2003) at AAAI and ICML confer- ences, respectively and a special issue in ACM SIGKDD explorations (Chawla, Japkowicz, & Kolcz, 2004) were dedicated to this topic. It has been witnessing growing interest and attention among researchers and practitioners seeking solutions in handling imbalanced data. An excel- lent review of the state-of-the-art is given by Weiss (2004). The data imbalance problem often occurs in classifica- tion and clustering scenarios when a portion of the classes possesses many more examples than others. As pointed out 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.10.042 * Corresponding author. Tel.: +852 34003782. E-mail address: mfyliu@polyu.edu.hk (Y. Liu). Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 by Chawla et al. (2004) when standard classification algo- rithms are applied to such skewed data, they tend to be overwhelmed by the major categories and ignore the minor ones. There are two main reasons why the uneven cases happen. One is due to the intrinsic nature of such events, e.g. credit fraud, cancer detection, network intrusion, and earthquake prediction (Chawla et al., 2004). These are rare events presented as a unique category but only occupy a very small portion of the entire example space. The other reason is due to the expense of collecting learning examples and legal or privacy reasons. In our previous study of building a manufacturing centered technical paper corpus (Liu & Loh, 2007), due to the costly efforts demanded for human labeling and diverse interests in the papers, we ended up naturally with a skewed collection. Automatic text classification (TC) has recently witnessed a booming interest, due to the increased availability of doc- uments in digital form and the ensuing need to organize them (Sebastiani, 2002). In TC tasks, given that most test collections are composed of documents belonging to multi- ple classes, the performance is usually reported in terms of micro-averaged and macro-averaged scores (Sebastiani, sification: A term weighting approach, Expert Systems with Ap- mailto:mfyliu@polyu.edu.hk 2 Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS 2002; Yang & Liu, 1999). Macro-averaging gives equal weights to the scores generated from each individual cate- gory. In comparison, micro-averaging tends to be domi- nated by the categories with more positive training instances. Due to the fact that many of these test corpora used in TC are either naturally skewed or artificially imbal- anced especially in the binary and so called ‘‘one-against- all” settings, classifiers often perform far less than satisfac- torily for minor categories (Lewis, Yang, Rose, & Li, 2004; Sebastiani, 2002; Yang & Liu, 1999). Therefore, micro- averaging mostly yields much better results than macro- averaging does. 1.2. Related work There have been several strategies in handling imbal- anced data sets in TC. Here, we only focus on the approaches adopted in TC and group them based on their primary intent. The first approach is based on sampling strategy. Yang (1996) has tested two sampling methods, i.e. proportion-enforced sampling and completeness-driven sampling. Her empirical study using the ExpNet system shows that a global sampling strategy which favors com- mon categories over rare categories is critical for the suc- cess of TC based on a statistical learning approach. Without such a global control, the global optimal perfor- mance will be compromised and the learning efficiency can be substantially decreased. Nickerson, Japkowicz, and Milios (2001) provide a guided sampling approach based on a clustering algorithm called Principal Direction Divisive Partitioning to deal with the between-class imbal- ance problem. It has shown improvement over existing methods of equalizing class imbalances, especially when there is a large between-class imbalance together with severe imbalance in the relative densities of the subcompo- nents of each class. Liu’s recent efforts (Liu, 2004) in testing different sampling strategies, i.e. under-sampling and over- sampling, and several classification algorithms, i.e. Naı̈ve Bayes, k-Nearest Neighbors (kNN) and Support Vector Machines (SVMs), improve the understanding of interac- tions among sampling method, classifier and performance measurement. The second major effort emphasizes cost sensitive learn- ing (Dietterich, Margineantu, Provost, & Turney, 2000; Elkan, 2001; Weiss & Provost, 2003). In many real scenar- ios like risk management and medical diagnosis, making wrong decisions are usually associated with very different costs. A wrong prediction of the nonexistence of cancer, i.e. false negative, may lead to death, while the wrong pre- diction of cancer existence, i.e. false positive, only results in unnecessary anxiety and extra medical tests. In view of this, assigning different cost factors to false negatives and false positives will lead to better performance with respect to positive (rare) classes (Chawla et al., 2004). Brank, Grobel- nik, Milic-Frayling, and Mladenic (2003) have reported their work on cost sensitive learning using SVMs on TC. They obtain better results with methods that directly mod- Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 ify the score threshold. They further propose a method based on the conditional class distributions for SVM scores that works well when only very few training examples are available. The recognition based approach, i.e. one-class learning, has provided another class of solutions (Japkowicz, Myers, & Gluck, 1995). One-class learning aims to create the deci- sion model based on the examples of the target category alone, which is different from the typical discriminative approach, i.e. the two classes setting. Manevitz and Yousef (2002) have applied one-class SVMs on TC. Raskutti and Kowalczyk (2004) claim that one-class learning is particu- larly helpful when data are extremely skewed and com- posed of many irrelevant features and very high dimensionality. Feature selection is often considered an important step in reducing the high dimensionality of the feature space in TC and many other problems in image processing and bioinformatics. However, its unique contribution in identi- fying the most salient features to boost the performance of minor categories has not been stressed until some recent work (Mladenic & Grobelnik, 1999). Yang and Pedersen (1997) has given a detailed evaluation of several feature selection schemes. We noted the marked difference between micro-averaged and macro-averaged values due to the poor performances over rare categories. Forman (2003) has done a very comprehensive study of various schemes for TC on a wide range of commonly used test corpora. He has recommended the best pair among different combina- tions of selection schemes and evaluation measures. The recent efforts from Zheng, Wu, and Srihari (2004) advance the understanding of feature selection in TC. They show the merits and great potential of explicitly combining posi- tive and negative features in a nearly optimal fashion according to the imbalanced data. Some recent work simply adapting existing machine learning techniques and not even directly targeting the issue of class imbalance have shown great potential with respect to the data imbalance problem. Castillo and Ser- rano (2004) and Fan, Yu, and Wang (2004) have reported the success using an ensemble approach, e.g. voting and boosting, to handle skewed data distribution. Challenged by real industry data with a huge number of records and an extremely skewed data distribution, Fan’s work shows that the ensemble approach is capable of improving the performance on rare classes. In their approaches, a set of weak classifiers using various learning algorithms are built up over minor categories. The final decision is reached based on the combination of outcomes from different clas- sifiers. Another promising approach which receives less attention falls into the category of semi-supervised learning or weakly supervised learning (Blum & Mitchell, 1998; Ghani, 2002; Goldman & Zhou, 2000; Lewis & Gale, 1994; Liu, Dai, Li, Lee, & Yu, 2003; Nigam, 2001; Yu, Zhai, & Han, 2003; Zelikovitz & Hirsh, 2000). The basic idea is to identify more positive examples from a large amount of unknown data. These approaches are especially sification: A term weighting approach, Expert Systems with Ap- Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx 3 ARTICLE IN PRESS viable when unlabeled data are steadily available. The last effort attacking the imbalance problem uses parameter tun- ing in kNNs (Baoli, Qin, & Shiwen, 2004). The authors expect to set k dynamically according to the data distribu- tion, in which a large k is granted given a minor category. In this paper, we tackle the data imbalance problem in text classification from a different angle. We present a new approach assigning better weights to the features from minor categories. After a brief review of the classic term weighting scheme, e.g. tfidf, in Section 2 and inspired by the analysis of various feature selection methods in Section 3, we introduce a simple probability based term weighting scheme which directly utilizes two critical information ratios, i.e. relevance indicators, in Section 4. These rele- vance indicators are nicely supported by the probability estimates which embody the category membership. The setup of experimental study is explained in Section 5. We carry out the evaluation and comparison of our new scheme with many other different weighting forms over two skewed data sets. We report the experimental findings and discuss their performance in Section 6. Section 7 con- cludes as well as highlights some future work. 2. Term weighting scheme Text classification (TC) is such a task to categorize doc- uments into predefined thematic categories. In particular, it aims to find the mapping n, from a set of documents D: {d1, . . . , di} to a set of thematic categories C: {C1, . . . , Cj}, i.e. n : D ? C. In its current practice, which is dominated by supervised learning, the construction of a text classifier is often conducted in two main phases (Debole & Sebas- tiani, 2003; Sebastiani, 2002): � Document indexing – the creation of numeric represen- tations of documents: – Term selection – to select a subset of terms from all terms occurring in the collection to represent the doc- uments in a better way, either to faster computing or to achieve better effectiveness in classification. – Term weighting – to assign a numeric value to each term to weight its contribution which helps a docu- ment stand out from others. � Classifier induction – the building of a classifier by learn- ing from the numeric representations of documents. In information retrieval and machine learning, term weighting has long been formulated in a form as term fre- quency times inverse documents frequency, i.e. tfidf (Baeza- Yates & Ribeiro-Neto, 1999; Salton & Buckley, 1988; Sal- ton & McGill, 1983; van-Rijsbergen, 1979). The more pop- ular ‘‘ltc” form (Baeza-Yates & Ribeiro-Neto, 1999; Salton & Buckley, 1988; Salton & McGill, 1983) is given by tfidfðti; d jÞ¼ tfðti; d jÞ� log N NðtiÞ � � ð1Þ Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 and its normalized version is wi;j ¼ tfidfðti; d jÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPjT j k¼1tfidfðtk; d jÞ 2 q ; ð2Þ where N and jTj denote the total number of documents and unique terms contained in the collection, respectively, and N(ti) represents the number of documents in the collection in which term ti occurs at least once, and tfðti; d jÞ¼ 1 þ logðnðti; d jÞÞ; if nðti; d jÞ > 0; 0; otherwise; � ð3Þ where n(ti, dj) is the number of times that term ti occurs in document dj. In practice, the summation in Eq. (2) is only concerned about the terms occurred in document dj. The significance of the classic term weighting schemes in Eqs. (1) and (2) is that they have embodied three funda- mental assumptions of term frequency distribution in a col- lection of documents (Debole & Sebastiani, 2003; Sebastiani, 2002). These assumptions are: � Rare terms are no less important than frequent terms – idf assumption. � Multiple appearances of a term in a document are no less important than single appearance – tf assumption. � For the same quantity of term matching, long docu- ments are no less important than short documents – nor- malization assumption. Because of these, the ‘‘ltc” and its normalized form have been extensively studied by many researchers and show their good performance over a number of different data sets (Sebastiani, 2002). Therefore, they have become the default choice in TC. 3. Inspiration from feature selection Feature selection serves as a key procedure to reduce the dimensionality of input data space to save computational cost. It has been integrated as a default step for many learning algorithms, like artificial neuron network, k-near- est neighbors, decision tree, etc. In the research community of machine learning, the computation constraints imposed by the high dimension of the input data space and the rich- ness of information available to maximally identify each individual object is a well known tradeoff. The ability of feature selection to capture the salient information by selecting the most important attributes, and thus making the computing tasks tractable has been shown in informa- tion retrieval and machine learning research (Forman, 2003; Ng, Goh, & Low, 1997; Ruiz & Srinivasan, 2002; Yang & Pedersen, 1997). Furthermore, feature selection is also beneficial since it tends to reduce the over-fitting problem, in which the trained objects are tuned to fit very well the data upon which they have been built, but per- forms poorly when applied to unseen data (Sebastiani, 2002). sification: A term weighting approach, Expert Systems with Ap- Table 1 Several feature selection methods, and their functions Feature selection method Mathematical form Information gain Pðtk; ciÞ log Pðtk;ciÞPðtkÞ�PðciÞþ Pð�tk; ciÞ log Pð�tk;ciÞ Pð�tkÞ�PðciÞ Mutual information log Pðtk;ciÞ PðtkÞPðciÞ Chi-square N�½Pðtk;ciÞ�Pð�tk;�ciÞ�Pðtk;�ciÞ�Pð�tk;ciÞ�2 PðtkÞ�Pð�tkÞ�PðciÞ�Pð�ciÞ Odds ratio log PðtkjciÞ�ð1�Pðtkj�ciÞÞ ð1�PðtkjciÞÞ�Pðtkj�ciÞ tk denotes a term; ci stands for a category; P(tk, ci) denotes the probability of documents from category ci where term tk occurs at least once; Pðtk;�ciÞ denotes the probability of documents not from category ci where term tk occurs at least once; Pð�tk; ciÞ denotes the probability of documents from category ci where term tk does not occur; Pð�tk; �ciÞ denotes the probability of documents not from category ci where term tk does not occur. Table 3 Feature selection methods and their formations as represented by information elements in Table 2 Method Mathematical form represented by information elements Information gain �AþCN log AþC N þ A N log A AþB � � þ CN log C CþD � � Mutual information log(AN/(A + B)(A + C)) Chi-square N(AD � BC)2/(A + C)(B + D)(A + B)(C + D) Odds ratio log(AD/BC) 4 Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS In TC, several feature selection methods have been intensively studied to distill the important terms while still keeping the dimension small. Table 1 shows the main func- tions of several popular feature selection methods. These methods are evolved either from the information theory or from the linear algebra literature (Sebastiani, 2002; Yang & Pedersen, 1997). Basically, there are two distinct ways to rank and assess the features, i.e. globally and locally. Global feature selec- tion aims to select features which are good across all cate- gories. Local feature selection intends to differentiate those terms that are more distinguishable for certain categories only. The sense of either ‘global’ or ‘local’ does not have much effect on the selection of method itself, but it does affect the performance of classifiers built upon different cat- egories. In TC, the main purpose is to address whether doc- ument belongs to a specific category. Obviously, we prefer the salient features which are unique from one category to another, i.e. a ‘local’ approach. Ideally, the salient feature set from one category does not have any items overlapping with those from other categories. If this cannot be avoided, then how to better present them has become an issue. While many previous works have shown the relative strengths and merits of these methods (Forman, 2003; Ng et al., 1997; Ruiz & Srinivasan, 2002; Sebastiani, 2002; Yang & Pedersen, 1997), our experience with feature selec- tion over a number of standard or ad hoc data sets shows the performance of such methods can be highly dependent on the data. This is partly due to the lack of understanding Table 2 Fundamental information elements used for feature selection in text classification ci �ci tk A B �tk C D A denotes the number of documents belonging to category ci where the term tk occurs at least once; B denotes the number of documents not belonging to category ci where the term tk occurs at least once; C denotes the number of documents belonging to category ci where the term tk does not occur; D denotes the number of documents not belonging to category ci where the term tk does not occur. Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 of different data sets in a quantitative way, and it needs fur- ther research. From our previous study of all feature selec- tion methods and what has been reported in the literature (Yang & Pedersen, 1997), we noted when these methods are applied to text classification for term selection purpose, they are basically utilizing four fundamental information elements shown in Table 2. These four information elements have been used to esti- mate the probability listed in Table 1. Table 3 shows the functions in Table 1 as presented by these four information elements A, B, C and D. 4. A probability based term weighting scheme 4.1. Revisit of tfidf As stated before, while many researchers believe that term weighting schemes in the form as tfidf representing those three aforementioned assumptions, we understand tfidf in a much simpler manner, i.e. � Local weight – the tf term, either normalized or not, specifies the weight of tk within a specific document, which is basically estimated based on the frequency or relative frequency of tk within this document. � Global weight – the idf term, either normalized or not, defines the contribution of tk to a specific document in a global sense. If we temporarily ignore how tfidf is defined, and focus on the core problem, i.e. whether this document is from this category, we realize that a set of terms is needed to rep- resent the documents effectively and a reference framework is required to make the comparison possible. As previous research shows that tf is very important (Leopold & Kin- dermann, 2002; Salton & Buckley, 1988; Sebastiani, 2002) and using tf alone can already achieve good performance, we retain the tf term. Now, let us consider idf, i.e. the glo- bal weighting of tk. The conjecture is that if term selection can effectively dif- ferentiate a set of terms tk out of all terms t to represent cat- egory ci, then it is desirable to transform that difference into some sort of numeric values for further processing. Our approach is to replace the idf term with the value that reflects the term’s strength of representing a specific cate- gory. Since this procedure is performed jointly with the cat- sification: A term weighting approach, Expert Systems with Ap- Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx 5 ARTICLE IN PRESS egory membership, this basically implies that the weights of tk are category specific. Therefore, the only problem left is how to compute such values. 4.2. Probability based term weights We decide to compute those term values using the most direct information, e.g. A, B and C, and combine them in a sensible way which is different from existing feature selec- tion measures. From Table 2, two important ratios which directly indicate terms’ relevance with respect to a specific category are noted, i.e. A/B and A/C: � A/B: if term tk is highly relevant to category ci only, which basically indicates that tk is a good feature to rep- resent category ci, then the value of A/B tends to be higher. � A/C: given two terms tk, tl and a category ci, the term with a higher value of A/C, will be the better feature to represent ci, since a larger portion of it occurs with category ci. In the following of this paper, we name A/B and A/C relevance indicators since these two ratios immediately indicate the term’s strength in representing a category. In fact, these two indicators are nicely supported by probabil- ity estimates. For instance, A/B can be extended as (A/N)/ (B/N), where N is the total number of documents, A/N is the probability estimate of documents from category ci where term tk occurs at least once and B/N is the probabil- ity estimate of documents not from category ci where term tk occurs at least once. In this manner, A/B can be inter- preted as a relevance indicator of term tk with respect to category ci. Surely, the higher the ratio, the more important the term tk is related to category ci. A similar analysis can be made with respect to A/C. The ratio reflects the expec- tation that a term is deemed as more relevant if it occurs in the larger portion of documents from category ci than other terms. Since the computing of both A/B and A/C has its intrin- sic connection with the probability estimates of category membership, we propose a new term weighting factor which utilizes the aforementioned two relevance indicators to replace idf in the classic tfidf weighting scheme. Consid- ering the probability foundation of A/B and A/C, the most immediate choice is to take the product of these two ratios. Therefore, the proposed weighting scheme is formulated as tf � log 1 þ A B A C � � : ð4Þ 5. Experiment setup Two data sets were tested in our experiment, i.e. MCV1 and Reuters-21578. MCV1 is an archive of 1434 English language manufacturing related engineering papers which we gathered by the courtesy of the Society of Manufactur- Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 ing Engineers (SME). It combines all engineering technical papers published by SME from year 1998 to year 2000. All documents were manually classified (Liu & Loh, 2007). There are a total of 18 major categories in MCV1. Fig. 1 gives the class distribution in MCV1. Reuters-21578 is a widely used benchmarking collection (Sebastiani, 2002). We followed Sun’s approach (Sun, Lim, Ng, & Srivastava, 2004) in generating the category infor- mation. Fig. 2 gives the class distribution of the Reuters data set used in our experiment. Unlike Sun et al. (2004), we did not randomly sample negative examples from cate- gories not belonging to any of the categories in our data set, instead we treated examples not from the target cate- gory in our data set as negatives. We compared our probability based term weighting scheme with a number of other well established weighting schemes, e.g. TFIDF, ‘ltc’ and normalized ‘ltc’, on MCV1 and Reuters-21578. We also carried out the bench- marking experiments between our conjecture and many other feature selection methods, e.g. chi-square (ChiS), cor- relation coefficient (CC), odds ratio (OddsR), and informa- tion gain (IG), by replacing the idf term with the feature selection value in the classic tfidf weighting scheme. There- fore, schemes are largely formulated in a form as tf � (fea- ture value) (TFFV). Table 4 shows all eight weighting schemes tested in our experiments and their mathematic formations. Please note that basically the majority of TFFV schemes are composed of two items, i.e. the normal- ized term frequency, tf(ti, dj)/max[tf(dj)], and the term’s feature value, e.g. N(AD � BC)2/(A + C)(B + D)(A + B)(C + D), in the chi-square scheme, where tf(ti, dj) is the frequency of term ti in the document dj and max[tf(dj)] is the maximum frequency of a term in the document dj. The only different ones are TFIDF weighting, ‘ltc’ form and the normalized ‘ltc’ form as specified in Table 4. Two popular classification algorithms were tested, i.e. Complement Naı̈ve Bayes (CompNB) (Rennie, Shih, Teevan, & Karger, 2003), and Support Vector Machine (SVM) (Vapnik, 1999). The CompNB has been recently reported that it can significantly improve the performance of Naı̈ve Bayes over a number of well known data sets, including Reuters-21578 and 20 Newsgroups. Various correction steps are adopted in CompNB, e.g. data trans- formation, better handling of word occurrence dependen- cies and so on. In our experiments, we borrowed the package implemented in Weka 3.5.3 Developer version (Witten & Frank, 2005). For SVM, we chose the well known implementation SVMLight (Joachims, 1998, 2001). Linear kernel has been adopted, since previous work has shown its effectiveness in TC (Dumais & Chen, 2000; Joachims, 1998). As for the performance measurement, precision, recall and their harmonic combination, i.e. the F1-value, were calculated (Baeza-Yates & Ribeiro-Neto, 1999; van-Rijsbergen, 1979). Performance was assessed based on fivefold cross validation. Since we are very con- cerned about the performance of every category, we report the overall performance in macro-averaged manner, i.e. sification: A term weighting approach, Expert Systems with Ap- Fig. 2. Class distribution in Reuters-21578. Fig. 1. Class distribution in MCV1. 6 Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS macro-average F1, to avoid the bias for minor categories in imbalanced data associated with micro-averaged scores (Sebastiani, 2002; Yang & Liu, 1999). Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 Major standard text preprocessing steps were applied in our experiments, including tokenization, stop word and punctuation removal, and stemming. However, feature sification: A term weighting approach, Expert Systems with Ap- Table 4 All eight weighting schemes tested in the experiments and their mathematic formations, where the normalized term frequency ntf is defined as tf(ti, dj)/ max[tf(dj)] Weighting scheme Name Mathematical formations tf � chi-square ChiS ntf � N(AD � BC)2/(A + C)(B + D)(A + B)(C + D) tf � correlation coef. CC ntf � ½ ffiffiffiffi N p ðAD � BCÞ= ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðA þ CÞðB þ DÞðA þ BÞðC þ DÞ p � tf � odds ratio OddsR ntf � log(AD/BC) tf � info gain IG ntf � ðAN log AN ðAþBÞðAþCÞþ C N log CN ðCþDÞðAþCÞÞ TFIDF TFIDF ntf � logð NNðtiÞÞ tfidf � ltc ltc tfðti; d jÞ � logð NNðtiÞÞ Normalized ltc nltc tfidfltcffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP tfidf 2ltc p Probability based Prob. ntf � log 1 þ AB A C � Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx 7 ARTICLE IN PRESS selection was skipped and all terms left after stop word and punctuation removal and stemming were kept as features. 6. Experimental results and discussion 6.1. Overall performance Fig. 3 shows the overall performance of eight weighting schemes tested over MCV1 and Reuters-21578 using SVM and CompNB. They are reported in terms of macro-aver- aged F1-values. Our first observation is that all TFFV weighting schemes, e.g. tf � chi-square, tf � information gain and our probability based one, outperform classic ones, i.e. TFIDF, ‘ltc’, and normalized ‘ltc’ schemes. The TFIDF’s perfor- mance on Reuters-21578 is in line with the literature (Sun Fig. 3. The macro-averaged F1-values of eight weighting schemes teste Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 et al., 2004; Yang & Liu, 1999). This has demonstrated the overall effectiveness of TFFV based schemes. In gen- eral, the performance patterns of eight weighting schemes on MCV1 and Reuters-21578 using two classification algo- rithms match very well. For example, our probability based term weighting scheme always take the lead in all eight schemes including the TFFV ones, and the normalized ‘ltc’ performs always the worst. When compared to TFIDF, the prevailing choice for term weighting in TC, our weighting strategy improves the overall performance from 6% to more than 12%, shown in Table 5. We also observe that when our scheme is adopted, CompNB has delivered a result which is very close to the best one that SVM can achieve using TFIDF scheme in Reuters-21578. This has demonstrated the great potential of using Comp- NB as a state-of-the-art classifier. d over MCV1 and Reuters-21578 using both SVM and CompNB. sification: A term weighting approach, Expert Systems with Ap- Table 5 Macro-averaged F1-values of TFIDF and probability based term weights on MCV1 and Reuters-21578 Classifier MCV1 21578 TFIDF Prob. TFIDF Prob. SVM 0.6729 0.7553 0.8381 0.8918 CompNB 0.4517 0.5653 0.6940 0.8120 8 Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS Among the three global based classic weighting schemes, i.e. TFIDF, ‘ltc’, and normalized ‘ltc’ form, none of them can generate comparable results over either MCV1 or Reu- ters-21578. A close look into their performance reveals that classifiers built for minor categories, e.g. composite manu- facturing, electronic manufacturing and others in MCV1 or rice, natgas, cocoa and others in Reuters-21578, do not produce satisfactory results. As a result, this has largely affected the overall performance negatively. Among all TFFVs, surprisingly, odds ratio does not perform as expected, since in literature odds ratio is mentioned as one of the leading feature selection methods for TC (Ruiz & Srinivasan, 2002; Sebastiani, 2002). This implies that it is always worthwhile to reassess the strength of a term selection method for a new data set, even if it tends to per- form well. 6.2. Gains for minor categories As shown in Figs. 1 and 2, both MCV1 and Reuters- 21578 are skewed data sets. While MCV1 possesses 18 cat- egories with one major category occupying up to 25% of Fig. 4. F1 scores of TFIDF and the probability based term weight Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 the whole population of supporting documents, there are six categories where each owns only 1% of MCV1, and other 11 categories falling below the average, i.e. 5.5%, if MCV1 is evenly distributed. The same case also happens to Reuters-21578 data set. While it has 13 categories, grain and crude, the two major categories, share around half of the population, there are eight categories in total whose shares falling below the average. Previous literature did not report successful stories over these minor categories (Sebastiani, 2002; Sun et al., 2004; Yang & Liu, 1999). Note, this imbalance situation is even worse when the training examples are arranged in the so called ‘‘one- against-all” setting for the induction of classifiers, i.e. examples from the target category (a minor category in this case) are considered as positive while examples from the rest categories are all deemed as negative. Nevertheless, given the nature of TC is to answer whether the document belongs to this particular category or not, the ‘‘one-against- all” setting is still the prevailing approach in TC, owing much to the fact that it dramatically reduces the number of classifiers to be induced. Since the proposed probability based weighting scheme is the best in the benchmarking test over both MCV1 and Reuters-21578, we intend to examine why this is the case. Therefore, we plot its performances in detail against TFIDF in Figs. 4 and 5, respectively. This is largely because TFIDF is the best among the three classic approaches as well as the default choice for TC in its cur- rent research and application (Sebastiani, 2002). A close examination of Figs. 4 and 5 shows that the probability based scheme produces much better results ing scheme tested over MCV1 using both SVM and CompNB. sification: A term weighting approach, Expert Systems with Ap- Fig. 5. F1 scores of TFIDF and the probability based term weighting scheme tested over Reuters-21578 using both SVM and CompNB. Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx 9 ARTICLE IN PRESS over minor categories in both MCV1 and Reuters-21578, regardless of classifiers used. For all minor categories shown in both figures, we observed a sharp increase of per- formance occurs when the system’s weighting method switch from TFIDF to the probability one. Table 6 reveals more insights with respect to the system performance. In general, we observe that using the proba- bility based term weighting scheme can greatly enhance the systems’ recalls. Although it falls slightly below TFIDF in terms of precision using SVM, it still improves the systems’ precisions in CompNB, far superior to those TFIDF can deliver. For SVM, while the averaged precision of TFIDF in MCV1 is 0.8355 which is about 5% higher than the prob- ability’s, the averaged recall of TFIDF is 0.6006 only, far less than the probability’s 0.7443. The case with Reuters- 21578 is even more impressive. While the averaged preci- sion of TFIDF is 0.8982 which is only 1.8% higher than the other, the averaged recall of probability based scheme reaches 0.9080, in contrast to TFIDF’s 0.7935. Overall, the probability based weighting scheme surpasses TFIDF in terms of F1-values over both data sets. Table 6 Macro-averaged precision and recall of TFIDF and probability based term weights on MCV1 and Reuters-21578 Data Classifier precision recall TFIDF Prob. TFIDF Prob. MCV1 SVM 0.8355 0.7857 0.6006 0.7443 CompNB 0.4342 0.6765 0.4788 0.5739 21578 SVM 0.8982 0.8803 0.7935 0.9080 CompNB 0.5671 0.7418 0.9678 0.9128 Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 6.3. Significance test To determine whether the performance improvement gained by the probability based scheme and other TFFVs over these two imbalanced data sets are significant, we per- formed the macro-sign test (S-test) and macro-t-test (T- test) on the paired F1-values. As pointed out by Yang and Liu (1999), on the one hand, the S-test may be more robust in reducing the influence of outliers, but at the risk of being insensitive or not sufficiently sensitive in perfor- mance comparison because it ignores the absolute differ- ence between F1-values; on the other hand, the T-test is sensitive to the absolute values but could be overly sensitive when F1-values are highly unstable, e.g. for the minor ategories. Therefore, we adopt both tests here to give a comprehensive understanding of the performance improvement. Since for both data sets, TFIDF performs better than the other two classic approaches, we choose it as the represen- tative of its peers. For both the S-test and T-test, we actually conduct two sets of tests over two data sets, respectively. One is to test all TFFV schemes, including the probability one, against TFIDF and the other one is to test the proba- bility scheme against others. While the first aims to assess the goodness of schemes in the form of TFFVs, the second intends to test whether the probability based scheme does generate better results. Table 7 summarizes p-values in the S-test for TFFV schemes against TFIDF and the probabil- ity one against others over two data sets. We consider two F1-values to be the same if their difference is not more than sification: A term weighting approach, Expert Systems with Ap- Table 8 t-Values of pairwise T-test on MCV1 and Reuters-21578, where a = 0.001 Test TFIDF ChiS CC OddsR IG Prob. MCV1 SVM, t-critical = 3.354 XX vs. TFIDF – 1.963E+01 2.151E+01 8.343E+00 2.588E+01 3.017E+01 Prob. vs. XX 3.017E+01 1.347E+01 1.069E+01 2.400E+01 6.571E+00 – CompNB, t-critical = 3.354 XX vs. TFIDF – 2.135E+01 2.049E+01 4.597E+00 2.419E+01 3.127E+01 Prob. vs. XX 3.127E+01 9.468E+00 1.043E+01 2.649E+01 8.192E+00 – Reuters-21578 SVM, t-critical = 3.467 XX vs. TFIDF – 2.516E+01 1.957E+01 1.692E+00 2.435E+01 2.889E+01 Prob. vs. XX 2.889E+01 3.682E+00 8.587E+00 2.262E+01 3.993E+00 – CompNB, t-critical = 3.467 XX vs. TFIDF – 2.157E+01 1.946E+01 5.318E+00 2.130E+01 3.064E+01 Prob. vs. XX 3.064E+01 4.926E+00 9.127E+00 2.167E+01 6.128E+00 – Table 7 p-Values of pairwise S-test on MCV1 and Reuters-21578, where two F1-values are the same if their difference is not more than 0.01 Test TFIDF ChiS CC OddsR IG Prob. MCV1 SVM XX vs. TFIDF – 4.813E�02 2.090E�03 5.923E�02 1.544E�02 6.561E�04 Prob. vs. XX 6.561E�04 6.363E�03 3.841E�02 6.561E�04 1.051E�01 – CompNB XX vs. TFIDF – 4.813E�02 4.813E�02 2.403E�01 4.813E�02 2.452E�02 Prob. vs. XX 2.452E�02 4.813E�02 2.452E�02 6.363E�03 1.544E�02 – Reuters-21578 SVM XX vs. TFIDF – 3.174E�03 3.174E�03 1.938E�01 5.859E�03 3.174E�03 Prob. vs. XX 3.174E�03 2.744E�01 1.938E�01 1.929E�02 3.872E�01 – CompNB XX vs. TFIDF – 3.271E�02 3.271E�02 1.133E�01 7.300E�02 3.271E�02 Prob. vs. XX 3.271E�02 1.133E�01 1.133E�01 5.859E�03 1.334E�01 – 10 Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS 0.01, i.e. 1%. Table 8 summarizes t-values of T-test for the identical comparison settings, where a is 0.001. From the results we can summarize the strength of dif- ferent schemes. Consider the merits evaluated based on TFFVs against TFIDF, TFFVs have shown that they are the better approach in handling imbalanced data. Among various TFFVs, our proposed scheme claims the leading performance tested in both MCV1 and Reuters-21578, regardless of the classifier used. However, the approach based on the odds ratio is not much superior to TFIDF. With respect to the evaluation based on the merits of the probability scheme against others, it is not surprising to see that the new scheme still takes the lead. It manages to perform better than other approaches in TFFV, e.g. infor- mation gain and chi-square, where the absolute difference of F1-values is considered. Finally, the results of informa- tion gain, chi-square and correlation coefficient shown in our tests are compatible with those in literature (Forman, 2003; Yang & Pedersen, 1997). In general, the more minor Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 categories the data set possesses, the better overall perfor- mance can be elevated if the probability based weighting scheme is chosen. 7. Conclusion and future work Handling of imbalanced data in TC has become an emerging challenge. In this paper, we introduce a new weighting paradigm which is generally formulated as tf � (feature value) (TFFV) to replace the classic TFIDF based approaches. We propose a probability based term weighting scheme, which directly makes use of two critical information ratios, as a new way to compute the term’s weight. These two ratios are deemed to possess the most salient information reflecting the term’s strength in associ- ating a category. Their computation does not impose any extra cost compared to the conventional feature selection methods. Our experimental study and extensive compari- sons based on two imbalanced data sets, MCV1 and Reu- sification: A term weighting approach, Expert Systems with Ap- Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx 11 ARTICLE IN PRESS ters-21578, show the merits of TFFV based approaches and their ability to handle imbalanced data. Among the various TFFVs, our probability based scheme offers the best overall performance in both data sets regardless of classifier used. Our approach has suggested an effective solution to improve the performance of imbalanced TC. Start from the work reported in this paper, there are a few immediate tasks awaiting us. Since the probability scheme is derived from the understanding of feature selec- tion, the A/B � A/C itself can also be considered as a new feature selection method that reflects the relevance of terms with respect to different thematic categories. It is interesting to further explore its joint application with other algo- rithms in TC. As for the slight decrease of precision noted, we intend to remedy the situation by switching the linear kernel with a string kernel in SVM (Lodhi, Saunders, Shawe-Taylor, Cristianini, & Watkins, 2002). Another challenge we are facing is to handle the situation where the critical information needed, e.g. A and B, cannot be easily secured, i.e. in text clustering. One potential direction is to infer these critical values from a small collection of labeled data and then test how robust these values or this probability approach could be, what strategies we can pro- pose to accommodate the variation of term occurrence in the unlabeled documents, and how to modify the critical values accordingly. The whole idea falls into the emerging paradigm of semi-supervised learning. We will report our study when the results become more solid. Acknowledgement The work described in this paper was partially sup- ported by a grant from the Research Grants Council of the Hong Kong Polytechnic University, Hong Kong Spe- cial Administrative Region, China (Project No. G-YF59). References Baeza-Yates, R., & Ribeiro-Neto, B. (1999). Modern information retrieval. Boston, MA, USA: Addison-Wesley. Baoli, L., Qin, L., & Shiwen, Y. (2004). An adaptive k-nearest neighbor text categorization strategy. ACM Transactions on Asian Language Information Processing (TALIP), 3(4), 215–226. Blum, A., & Mitchell, T. (1998). Combining labeled and unlabeled data with co-training. In COLT: Proceedings of the workshop on computa- tional learning theory (pp. 92–100). Brank, J., Grobelnik, M., Milic-Frayling, N., & Mladenic, D. (2003). Training text classifiers with SVM on very few positive examples. MSR- TR-2003-34. Castillo, M. D. D., & Serrano, J. I. (2004). A multistrategy approach for digital text categorization from imbalanced documents. ACM SIG- KDD Explorations Newsletter, 6(1) [Special issue on learning from imbalanced datasets]. Chawla, N., Japkowicz, N., & Kolcz, A. (Eds.). (2003). Proceedings of the ICML’2003 workshop on learning from imbalanced data sets. Chawla, N., Japkowicz, N., & Kolcz, A. (Eds.). (2004). ACM SIGKDD Explorations Newsletter, 6(1) [Special issue on learning from imbal- anced data sets]. Debole, F., & Sebastiani, F. (2003). Supervised term weighting for automated text categorization. In Proceedings of the 2003 ACM Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 symposium on applied computing (pp. 784–788). Melbourne, Florida, USA. Dietterich, T., Margineantu, D., Provost, F., & Turney, P. (Eds.). (2000). In Proceedings of the ICML’2000 workshop on cost-sensitive learning. Dumais, S., & Chen, H. (2000). Hierarchical classification of Web content. In Proceedings of the 23rd annual international ACM SIGIR conference on research and development in information retrieval (SIGIR2000) (pp. 256–263). Athens, Greece. Elkan, C. (2001). The foundations of cost-sensitive learning. In Proceed- ings of the 17th international joint conference on artificial intelligence (IJCAI’01) (pp. 973–978). Fan, W., Yu, P. S., & Wang, H. (2004). Mining extremely skewed trading anomalies. In Advances in database technology – EDBT 2004: Ninth international conference on extending database technology (pp. 801– 810). Heraklion Crete, Greece. Forman, G. (2003). An extensive empirical study of feature selection metrics for text classification. The Journal of Machine Learning Research, 3, 1289–1305 [Special issue on variable and feature selection]. Ghani, R. (2002). Combining labeled and unlabeled data for multiclass text categorization. In International conference on machine learning (ICML 2002), Sydney, Australia. Goldman, S., & Zhou, Y. (2000). Enhancing supervised learning with unlabeled data. In Proceedings of 17th international conference on machine learning (pp. 327–334). San Francisco, California, USA. Japkowicz, N. (Ed.). (2000). Proceedings of the AAAI’2000 workshop on learning from imbalanced data sets, AAAI Tech Report WS-00-05, AAAI. Japkowicz, N., Myers, C., & Gluck, M. A. (1995). A novelty detection approach to classification. In Proceedings of the 14th international joint conference on artificial intelligence (IJCAI-95) (pp. 518–523). Joachims, T. (1998). Text categorization with support vector machines: Learning with many relevant features. Machine learning: ECML-98, 10th European conference on machine learning (pp. 137–142). Berlin, Germany. Joachims, T. (2001). A statistical learning model of text classification with support vector machines. In Proceedings of the 24th annual interna- tional ACM SIGIR conference on research and development in information retrieval (pp. 128–136). New Orleans, Louisiana, United States. Leopold, E., & Kindermann, J. (2002). Text categorization with support vector machines – How to represent texts in input space. Machine Learning, 46(1–3), 423–444. Lewis, D. D., & Gale, W. A. (1994). A sequential algorithm for training text classifiers. In Proceedings of {SIGIR}-94, 17th ACM international conference on research and development in information retrieval (pp. 3– 12). Dublin, Ireland. Lewis, D. D., Yang, Y., Rose, T. G., & Li, F. (2004). RCV1: A new benchmark collection for text categorization research. Journal of Machine Learning Research, 5, 361–397. Liu, A. Y. C. (2004). The effect of oversampling and undersampling on classifying imbalanced text datasets. Masters thesis, University of Texas at Austin. Liu, Y., & Loh, H. T. (2007). Corpus building for corporate knowledge discovery and management: A case study of manufacturing. In Proceedings of the 11th international conference on knowledge-based and intelligent information and engineering systems, KES’07, Lecture notes in artificial intelligence, LNAI, Vol. 4692 (pp. 542–550). Vietri sul Mare, Italy. Liu, B., Dai, Y., Li, X., Lee, W. S., & Yu, P. (2003). Building text classifiers using positive and unlabeled examples. In Proceedings of the third IEEE international conference on data mining (ICDM’03), Melbourne, Florida. Lodhi, H., Saunders, C., Shawe-Taylor, J., Cristianini, N., & Watkins, C. (2002). Text classification using string kernels. The Journal of Machine Learning Research, 2, 419–444. Manevitz, L. M., & Yousef, M. (2002). One-class SVMS for document classification. The Journal of Machine Learning Research, 2, 139–154. sification: A term weighting approach, Expert Systems with Ap- 12 Y. Liu et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS Mladenic, D., & Grobelnik, M. (1999). Feature selection for unbalanced class distribution and Naive Bayes. In Proceedings of the 16th international conference on machine learning, ICML’99 (pp. 258–267). Ng, H. T., Goh, W. B., & Low, K. L. (1997). Feature selection, perception learning, and a usability case study for text categorization. In ACM SIGIR forum, Proceedings of the 20th annual international ACM SIGIR conference on research and development in information retrieval (pp. 67– 73). Philadelphia, Pennsylvania, United States. Nickerson, A., Japkowicz, N., & Milios, E. (2001). Using unsupervised learning to guide re-sampling in imbalanced data sets. In Proceedings of the eighth international workshop on AI and statistics (pp. 261– 265). Nigam, K. P. (2001). Using unlabeled data to improve text classification. PhD thesis, Carnegie Mellon University. Raskutti, B., & Kowalczyk, A. (2004). Extreme re-balancing for SVMs: A case study. ACM SIGKDD Explorations Newsletter, 6(1), 60–69 [Special issue on learning from imbalanced datasets]. Rennie, J. D. M., Shih, L., Teevan, J., & Karger, D. R. (2003). Tackling the poor assumptions of Naive Bayes text classifiers. In Proceedings of the 20th international conference on machine learning (pp. 616–623). Washington, DC, USA. Ruiz, M. E., & Srinivasan, P. (2002). Hierarchical text categorization using neural networks. Information Retrieval, 5(1), 87–118. Salton, G., & Buckley, C. (1988). Term weighting approaches in automatic text retrieval. Information Processing and Management, 24(5), 513– 523. Salton, G., & McGill, M. J. (1983). Introduction to modern information retrieval. New York, USA: McGraw-Hill. Sebastiani, F. (2002). Machine learning in automated text categorization. ACM Computing Surveys (CSUR), 34(1), 1–47. Sun, A., Lim, E.-P., Ng, W.-K., & Srivastava, J. (2004). Blocking reduction strategies in hierarchical text classification. IEEE Transac- tions on Knowledge and Data Engineering (TKDE), 16(10), 1305– 1308. Please cite this article in press as: Liu, Y. et al., Imbalanced text clas plications (2007), doi:10.1016/j.eswa.2007.10.042 van-Rijsbergen, C. J. (1979). Information retrieval (2nd ed.). London, UK: Butterworths. Vapnik, V. N. (1999). The nature of statistical learning theory (2nd ed.). New York: Springer-Verlag. Weiss, G. M. (2004). Mining with rarity: A unifying framework. ACM SIGKDD Explorations Newsletter, 6(1), 7–19 [Special issue on learning from imbalanced datasets]. Weiss, G. M., & Provost, F. (2003). Learning when training data are costly: The effect of class distribution on tree induction. Journal of Artificial Intelligence Research, 19, 315–354. Witten, I. H., & Frank, E. (2005). Data mining: Practical machine learning tools and techniques (2nd ed.). San Francisco, CA, USA: Morgan Kaufman. Yang, Y. (1996). Sampling strategies and learning efficiency in text categorization. In Proceedings of the AAAI spring symposium on machine learning in information access (pp. 88–95). Yang, Y., & Liu, X. (1999). A re-examination of text categorization methods. In Proceedings of the 22nd annual international ACM SIGIR conference on research and development in information retrieval (pp. 42– 49). Berkeley, California, United States. Yang, Y., & Pedersen, J. O. (1997). A Comparative study on feature selection in text categorization. In Proceedings of ICML-97, 14th international conference on machine learning (pp. 412–420). Yu, H., Zhai, C., & Han, J. (2003). Text classification from positive and unlabeled documents. In Proceedings of the 12th international confer- ence on information and knowledge management (CIKM 2003) (pp. 232–239). New Orleans, LA, USA. Zelikovitz, S., & Hirsh, H. (2000). Improving short text classification using unlabeled background knowledge. In Proceedings of the 17th interna- tional conference on machine learning (ICML2000). Zheng, Z., Wu, X., & Srihari, R. (2004). Feature selection for text categorization on imbalanced data. ACM SIGKDD Explorations Newsletter, 6(1), 80–89 [Special issue on learning from imbalanced datasets]. sification: A term weighting approach, Expert Systems with Ap- Imbalanced text classification: A term weighting approach Introduction Motivation Related work Term weighting scheme Inspiration from feature selection A probability based term weighting scheme Revisit of tfidf Probability based term weights Experiment setup Experimental results and discussion Overall performance Gains for minor categories Significance test Conclusion and future work Acknowledgement References 
work_35cydgiokfe3bhmwz2zsowhzwi	----	untitled Supporting activity modelling from activity traces Olivier L. Georgeon,1 Alain Mille,1 Thierry Bellet,2 Benoit Mathern1 and Frank E. Ritter3 (1) Université de Lyon, 86 Rue Pasteur 69007 Lyon, France Email: olivier.georgeon@liris.cnrs.fr (2) Institut Français des Sciences et Technologies des Transports, de l’Aménagement et des Réseaux, 25, Avenue François Mitterrand, 69500 Bron, France (3) The Pennsylvania State University, University Park, PA 16802, USA Abstract: We present a new method and tool for activity modelling through qualitative sequential data analysis. In particular, we address the question of constructing a symbolic abstract representation of an activity from an activity trace. We use knowledge engineering techniques to help the analyst build an ontology of the activity, that is, a set of symbols and hierarchical semantics that supports the construction of activity models. The ontology construction is pragmatic, evolutionist and driven by the analyst in accordance with their modelling goals and their research questions. Our tool helps the analyst define transformation rules to process the raw trace into abstract traces based on the ontology. The analyst visualizes the abstract traces and iteratively tests the ontology, the transformation rules and the visualization format to confirm the models of activity. With this tool and this method, we found innovative ways to represent a car-driving activity at different levels of abstraction from activity traces collected from an instrumented vehicle. As examples, we report two new strategies of lane changing on motorways that we have found and modelled with this approach. Keywords: sequence mining, timeline analysis, activity trace, knowledge-based system, activity modelling 1. Introduction We introduce here new principles based on knowledge engineering techniques for designing systems to help analysts create models of activity from activity traces. We illustrate these princi- ples with a software tool that we have imple- mented, and with an example modelling analysis that we have performed using this tool. By activity trace we mean a set of multiple streams of quantitative or symbolic data that record (at least partially) an activity performed by a subject. The analysts may be psychologists seeking to build theories of the subject’s cogni- tion, ergonomists seeking to design better user interfaces, analysts seeking to predict the sub- ject’s behaviour in specific conditions, trainers seeking to improve training techniques or even the subjects themselves seeking to improve their understanding of their actions. In each case, the created models of activity constitute micro- theories proposed by the analysts to describe, explain and try to predict how the subject per- forms the activity. The principles and the tool that we introduce here address three needs for helping analysts construct models of activity from activity traces. The first need is for helping the analyst learn previously unknown aspects or details of the subject’s activity from the activity traces. The DOI: 10.1111/j.1468-0394.2011.00584.x Article _____________________________ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 c! 2011 Blackwell Publishing Ltd Expert Systems 1 E X S Y 5 8 4 B Dispatch: 11.2.11 Journal: EXSY CE: Sandeep Journal Name Manuscript No. Author Received: No. of pages: 15 PE: Chris/Satish EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M mailto:olivier.georgeon@liris.cnrs.fr mailto:olivier.georgeon@liris.cnrs.fr mailto:olivier.georgeon@liris.cnrs.fr mailto:olivier.georgeon@liris.cnrs.fr second need is for helping the analyst con- struct meaningful symbolic representations of interesting aspects of the activity. These repre- sentations, associated with the explanations proposed by the analyst, constitute the models of activity. The third need is for helping the analyst test and support the created models of activity with regard to the activity trace. Activity traces have also been called protocols (Ericsson & Simon, 1993) or simply sequential data (Sanderson & Fisher, 1994). We prefer the term activity trace because this term conveys the idea that the trace is intended to be interpreted by somebody (designated here as the analyst). We think of a trace as a footprint that helps who sees it understand what happened. Our activity traces yet differ from mere footprints in that they are not accidentally produced or unprocessed but they rather result from the analyst’s choices and set-up. Many software tools have been implemented for activity trace analysis. A recent review (Hil- bert & Redmiles, 2000) notes 40 of them. These tools cannot autonomously generate a compre- hensive explanation of human behaviour but they interactively support analysis. This analysis consists of identifying, categorising, labelling and transforming pieces of data and informa- tion in the activity trace. We summarize this process by the notion of abstraction. The ana- lysts use their expertise and knowledge to for- mulate a whole set of tiny hypotheses and choices concerning how to collect the data, how to filter it, how to cluster and label it and how to display and to report it so that it responds to the analysis purpose. Although most of the existing tools acknowl- edge the central role of the analysts and the importance of their knowledge and expertise in the analysing process, these tools still lack knowledge representation mechanisms to sup- port the management of the analysts’ knowl- edge. For instance, MacShapa (Sanderson et al., 1994) does help analysts label and cluster the behavioural data. It also acknowledges the usage of these labels as symbols to describe the activity. It, however, does not help the analyst formulate and manage the symbolic inferences she can make from these symbols. We formulated the hypothesis that aspects of knowledge engineering can help design software systems that address the three needs identified above. We use ontology management facilities and rule engines to capture the hypotheses and choices made by the analyst. When interactively used by the analyst or a group of analysts, these facilities help formalize the way analysts find interesting symbolic patterns and infer models from them. Once this knowledge is formalized, the system uses it to automatically compute new representations of the activity from the activity traces. The system also helps analysts organize and store the different concepts and rules that summarize different studies and help capitalize on these studies. To explain our principles and demonstrate the system and its design, we have organized this paper as follows: Section 2 presents the principles of activity trace modelling, based on a pragmatic and evolutionist approach. Section 3 presents the prototype software tool that we have implemented from these principles, its technical features, its architecture and its user interface. Section 4 presents an example study in which we have used this tool to create models of lane change on motorways from activity traces generated with an instrumented car. The meth- od is then summarized in the conclusion. 2. Modelling activity traces The notion of activity traces is widely used in the human behaviour literature, and we cannot attribute its origin to a specific author. Only more specific-related notions can be identified, such as pattern languages, as reviewed by Dear- den and Finlay (2006), or grammar representa- tions (Olson et al., 1994). Despite the wide usage of the term activity trace, we could not find a definition of it, which led us to propose the following definition: An activity trace is a meaningful inscription, from the viewpoint of an analyst, of the flow of what has happened, from the viewpoint of a subject. With this definition, we want to highlight that an activity trace always implies two viewpoints, 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 2 Expert Systems c! 2011 Blackwell Publishing Ltd EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M situated in two different moments. It implies the subject’s viewpoint, when he or she was performing the activity, and the analyst’s view- point, when he or she is analysing the activity trace. Indeed, an activity trace cannot be an inscription of all that happened (if that had any sense), because an activity only concerns what relates to the subject’s perspective, goals and intentions. Thus, inevitably, the analyst has to make assumptions about what is meaningful to the subject when the analyst sets up the tracing mechanism. In addition, the activity trace de- pends on what activity aspects interest the analyst and what makes sense to her according to her previous knowledge and to her analysis goals. Because an activity trace depends on the ana- lyst’s knowledge and assumptions, it can only be modelled in an iterative way, each iteration pro- ducing new knowledge leading to new hypotheses for the next iteration. Ericsson and Simon (1993) described this iterative nature of analysing human behaviour: ‘In designing our data-gathering schemes, we make minimal essential theoretical commitments, then try to use the data to test stronger theories’ (p. 274). Moreover, none of the iterations can produce knowledge that could be proven to be true in an absolute sense, but only knowledge that is more efficient and useful with regard to the analyst’s goals and that is more convincing to the analyst’s community than the knowledge from the previous iteration. More broadly, this conception of knowledge relates to a pragmatic epistemology (James, 1907) and an evolutionist epistemology (Popper, 1972). These pragmatic and evolutionist aspects are crucial when defining a methodology and a tool for activity trace modelling. By fully acknowl- edging these aspects, we have designed a tool that facilitates and accelerates the evolutionist model- ling process. The tool helps formulate a series of micro-hypotheses of possibly useful symbols, pos- sibly useful transformation rules to transform the low-level data into higher-level data and possibly useful representations of the activity trace based on the micro-hypotheses. If the obtained repre- sentation does not help the analyst understand the activity better, then she rejects these micro- hypotheses; if it helps, then she keeps them. The tool is designed to shorten this formula- tion=usage=validation-or-rejection loop. This process leads to the construction of a set of micro-hypotheses that are validated by the ana- lyst. This set constitutes a formalization of the analyst’s knowledge about how to understand the activity. The tool stores this knowledge, helps the analyst keep track of it and helps the analysts’ community discuss and question it. The next section details how the tool reaches this goal by helping the analyst define sequences at the right level of abstraction and simulta- neously identify interesting subsequences, define them precisely and query the whole trace in search for their occurrences. 2.1. Collecting a symbolic trace A raw activity trace can be made of any kind of data describing a subject’s activity flow and intended for an analyst’s usage. In a broad sense, it can range from video or audio record- ing to computer logs. The only common point is that the data are temporally organized, meaning that each data piece is associated with a time- stamp referring to a common time base. The first abstraction step consists of converting these raw traces into sequences of symbols. We refer to this step as the discretization of the raw trace into a symbolic trace. The symbols in the sym- bolic trace have to be meaningful to the analyst, and they are chosen on a pragmatic and evolu- tionist basis, in compliance with a pragmatic and evolutionist epistemology introduced above (introduction of Section 2). The discretization process can be manual, semi-automatic or auto- matic. The definition of the symbols may evolve in parallel with the implementation of the dis- cretization process. This is because the later interpretation of the symbols may differ from the meaning initially intended by the analyst when she specifies the discretization algorithm. For example, in a study of car driving, we have used the classical mathematical curve ana- lysis method to generate symbols of interest from numerical values of the vehicle speed, the steering wheel angle and the pedal positions. In this case, the symbols correspond to threshold 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 c! 2011 Blackwell Publishing Ltd Expert Systems 3 EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M crossing, local extremum and inflexion points. Figure 1 illustrates this discretization process. In Figure 1, the curves represent the vehicle speed in km=h and the brake pedal position in percentage of range. Symbols of interest are shown on these curves as circles (threshold crossings, inflexion points, local extremums). The symbols are merged into the symbolic trace that is represented at the bottom of Figure 1. The figure also shows a derivative value as an example property of interest associated with a symbol generated by a brake pedal threshold crossing. The analyst specifies the way to gen- erate these symbols so that they correspond to meaningful events that describe the activity. In this example, the threshold crossing indicates the extent of the braking action and the deriva- tive value indicates the abruptness of this action. Notably, while creating these symbols, the ana- lyst claims the existence of the events that these symbols represent. By so doing, the analyst defines an ontology of the activity. Our experience has taught us that the system must maintain a connection between the raw trace and the symbolic trace, and provide parallel displays of both of them. The analyst needs to tune many parameters of the discretiza- tion algorithms, like the threshold values or noise filters. The analyst validates the chosen symbols and algorithms by comparing the symbolic trace to the raw trace and ensuring that the symbolic trace represents what is hap- pening. While she defines and validates these symbols, the analyst also supports her claim that the events represented by these symbols ‘exist’. This support arises because the method to generate these symbols from the recorded data is formally specified and explained by the analyst. In this example, after we fully specified the discretization algorithm in accordance to our specific modelling goals, the discretization algorithm could then compute the symbols fully automatically. 2.2. Modelling the symbolic trace At the symbolic level, analysts most often want to focus on relations between events. Indeed, events are not meaningful by themselves, but they become meaningful in the context where they relate to each other (Sanderson & Fisher, 1994). Examples of such relations include a ‘sequence following within a certain period of time’, ‘co-occurrence within a certain period of time’ and ‘causality with regard to a certain explanative theory’. Building and understanding these relations between events is a part of the analysing process. By definition, a set of ele- ments connected through relations is a graph. Therefore, we model the symbolic traces with a graph structure. More precisely, our trace graph structure has two parts: a sequence and an ontology. The sequence is a part of the graph that is made of event instances and of relation instances between event instances. The ontology is a part of the graph that is made of event classes and of relations between event classes. Figure 2 illus- trates this graph structure with a simplified example taken from the car-driving study. The sequence is represented in the bottom part of Figure 2 and the ontology in the top part. In the sequence graph, event instances are represented as circles and triangles. Relation instances between event instances are repre- sented as solid arrows between these circles and triangles. Example properties of event instances are represented with grey dashed arrows point- ing to their value at the bottom of the figure (e.g. the duration of an eye movement and accelera- tion value associated with an inflexion point of the speed curve). 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Time Trace Speed Brake Local extremum Threshold Inflections Figure 1: Discretization of analogical traces. 4 Expert Systems c! 2011 Blackwell Publishing Ltd EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M Our tool displays the sequence in a similar form as shown in Figure 2; the exact form is shown in Figure 4. This display uses two axes: the time axis and the ‘abstraction’ axis. That is, the events’ time-code attributes determine their x coordinates and the analyst specifies their y coordinates when she configures these event’s class in the ontology. The analyst can use the y coordinate to express different meanings; our recommendation is to use it to express an idea of abstraction level related to a specific analysis. In this example, the lowest level (circles) represents the events obtained from the discretization process; the intermediary level represents events that describe the activity in usual driving terms: accelerate, glance, turn signal on=off (blinker); and the higher level represents events describing lane-change behaviour: indica- tor of intention to change lane and index of lane change. This example expresses the analyst’s assumption that the conjunction of an accelera- tion and a glance to the left rear mirror can generate an indicator of the driver’s intention to change lane (L.C. Indicator). In the ontology graph, the nodes represent event classes and the edges represent the relation ‘sub-class of’ (dashed black arrows). The analyst defines the ontology during the modelling pro- cess. For example, the ‘Collected events’ class includes all the event classes that come from the discretization process. The ‘Driving descriptors’ class gathers the intermediary event classes that describe the activity in usual driving terms. The ‘Lane change descriptors’ class gathers the most abstract event classes describing lane changes. The dotted grey arrows in the figure represent the relation ‘type of’ going from event classes defined in the ontology to event instances in the sequence. Again, like most ontologies, this ontology is made by the analyst on a pragmatic basis. It is likely that two analysts will create two different ontologies. While the software cannot demon- strate that one is better than the other, it does help the analysts formalize and discuss them. This discussion leads to the construction of a language for describing the activity that repre- sents an agreement about the terms that can be used to describe the activity. In addition, the ontology also supports the analysts’ agreement about how the trace should be visualized, be- cause the visualization properties of the symbols are stored in the ontology. This trace formalism enables the analyst to conduct a hierarchical analysis of the activity. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Ontology Sequence Event Collected events Driving descriptors Lane Change descriptors Eye track L.C. IndexSpeed Accelerate Glance L.C. IndicatorBlinker Abstraction Time Eye left duration: 0.2s Speed acceleration: 1.1ms–2 Superclass of Type of Inference Properties Figure 2: Activity trace modelled in a knowledge engineering system. c! 2011 Blackwell Publishing Ltd Expert Systems 5 EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M Event instances form a hierarchy where higher-level events are temporal patterns of lower-level events. The ontology also defines a hierarchy because some event classes are sub-classes of others. Notably, these two hierarchies are different because lower- level event instances do not necessarily belong to a sub-class of the class of higher-level event instances. 3. System implementation We have implemented a prototype system based on an assemblage of open-source knowledge engineering tools: an ontology editor, an infer- ence engine, visualization facilities and docu- mentation facilities. This system is named ABSTRACT (Analysis of Behavior and Situa- tion for menTal Representation Assessment and Cognitive acTivity modelling). Figure 3 illus- trates this assemblage. The system can be split into three levels: a lower level, at the bottom of the figure, which is the collection system; a core level, in the centre of the figure, which is the symbolic trace system itself; and a higher level, on the top of the figure, which is a documentation level. 3.1. The collection system The collection system integrates tools to help the analyst prepare the symbolic trace. We call these tools collection agents. Collection agents may be automatic when specified once by the analyst or may require the analyst’s intervention. Auto- matic collection agents can be tools for prepro- cessing sensor data or computer logs, as in the example of Section 2.1. Semi-automatic collec- tion agents can be tools for helping the analyst take notes, record interviews or transcript video data. As noted, this discretization cannot be done blindly, but must be driven by the analyst. Hence, this level requires visualization facilities. We use Microsoft Excel with specific Visual Basic Application (VBA) macros as a visualiza- tion tool for the collection system. In this visualization, each event of the symbolic trace 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Collection System Visualization system Collection agents Collected trace Videos Notes, Interviews Logs Sensors Symbolic-Trace-System Visualization system Transformation engine Ontology editor Modeled Trace Ontology Inferences Style- sheets Documentation System Transfor- mations Models Explanations Figure 3: The architecture of the ABSTRACT activity analysis system. 6 Expert Systems c! 2011 Blackwell Publishing Ltd EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M is displayed as a line in the spreadsheet. The lines are coloured according to the event’s type and the event’s properties are organized in different columns. We have implemented a spe- cific video player and analogical data player that triggers VBA macros that automatically scrolls down the spreadsheet in synchronization. Some of these facilities are also available in commer- cial quantitative data analysis tools such as MacShapa (Sanderson et al., 1994) and NVivo. These facilities allow the analyst to check and validate or reject the symbolic trace, that is, refine the discretization algorithm and its di- verse parameters until she gets a satisfying and meaningful symbolic trace including appropri- ate properties of interest. 3.2. The symbolic trace system The symbolic trace system is the knowledge engineering system itself. At this level, the traces are modelled as described in Section 2.2. In addition, they are associated with a set of inference rules and a set of style sheets. A style sheet is a specification for displaying the trace on the screen. It specifies how semantic proper- ties of the events that are defined in the ontology should be converted into visualization proper- ties, such as shape and position. Style sheets also implement particular time scales, and particular filters to display only the interesting aspect of the trace for a particular analysis. They corre- spond to different ways of looking at the trace according to different modelling goals. The inference rules produce inferred symbols from patterns of previously existing symbols. The principle is to query the graph in the search for sub-graphs that match certain patterns, and to attach new nodes and arcs to the matching sub-graphs. These new nodes represent the in- ferred symbols and the new arcs represent the inference relations. The usage of this inference mechanism is further described in Section 3.5. Technically, the sequence part of the activity traces is encoded as resource description frame- work (RDF) graphs. We choose RDF because it is the most widely used specification for graph encoding. We use XML as a serialization of RDF to store sequences, because XML makes RDF graphs easy to share with other applica- tions. The ontology is encoded as RDF schema (RDFS), because RDFS is the simplest ontology language based on RDF. We use Protégé as a graphical ontology editor. That is, an installa- tion of Protégé is embedded in our tool, and the analyst uses it to define the ontology of his traces. The graphical displays of our traces are encoded under the scalable vector graphic (SVG) specification. Because Firefox natively supports SVG, we use it as a visualization tool, and we have implemented most of the tool as a web application in PHP. We use extensible stylesheet language (XSL) as a transformation language for transforming RDF traces into their SVG graphical representation. We use SPARQL as a query language for graphs, as we will explain in Section 3.5. 3.3. The documentation system Analysts using our symbolic trace system ex- pressed the need for a higher system layer provid- ing a way to both index and attach documentation to episodes of interest. We implemented this by associating the symbolic trace system with a Wiki. As we have made the choice of implementing ABSTRACT as a web application, analysts can reference each episode of interest by their URL, and easily paste this URL into a Wiki page. Moreover, some new Wiki implementations, like Semantic MediaWiki,1 include semantic facilities. We are still investigating how these semantic facilities can be used to merge the ontology editor with the documentation system into a single semantic documentation system. 3.4. System usage The user interface is accessible as a web page in any browser that supports SVG, such as Firefox. This interface is illustrated in Figure 4. It has four tabs: the Open tab that allows the analyst to select a trace in a list; the Info tab that displays general information about the selected trace, such as its creation date and its version, 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 1http://semantic-mediawiki.org c! 2011 Blackwell Publishing Ltd Expert Systems 7 EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M http://semantic-mediawiki.org and the management of stylesheets; the View tab that displays graphical visualizations of the trace; and the Edit tab that allows the analyst to write queries to transform the trace. The interface also provides a link to the ontology editor, Protégé. The View tab shown in Figure 4 provides the following functionalities (noted with numbered boxes): 1. Unique ID of the analysis, including aspects of the trace file and analyses done. 2. Selection of different visualization style sheets in drop-down lists. Different visuali- zations can be displayed simultaneously on the screen and their time code is synchro- nized. 3. Time code: this value corresponds to the cursor position in the visualization modules (vertical red line). The analyst can enter a time code and click Go to to focus on it. 4. Visualization sample with a time span of 10 s. The analyst can scroll the trace left and right with the mouse. This visualization example corresponds to the simplified de- scription given in Figure 2. 5. Visualization sample of an entire trace (20 min), with only the high-level symbols displayed. These trace examples come from our car-driving study and are further ex- plained in Section 4. 6. The analyst can show the symbols’ proper- ties by clicking on the symbols. 7. The system can synchronize with a video player. When this box is checked, the system gives the time-code control to the video player and automatically follows it. 3.5. The transformation mechanism The analyst uses the Edit tab to write queries that infer higher-level symbols from patterns of lower-level symbols. For instance, Table 1 illus- trates a query to infer the Lane change indicator symbol shown in Figure 2, which indicates that a 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Figure 4: ABSTRACT user interface. 8 Expert Systems c! 2011 Blackwell Publishing Ltd EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M lane change is about to happen. In this example, the analyst wants to test the hypothesis that this indicator can be inferred from a conjunction of an accelerate event with an acceleration value Z1 m=s2, followed by a glance event pointing to the left mirror (generated by an eye-tracker), both within 1 s of each other. The graph elements, either from the sequence or the ontology, are handled as triples [node, ed- ge, node]. A query consists of a selection clause (WHERE) and a CONSTRUCT clause. The selection clause specifies a pattern of triples that should match the graph, and the CONSTRUCT clause specifies a pattern of triples that should be added to the graph wherever a pattern matches the selection clause. In addition, match- ing patterns can be restricted by a FILTER clause. The syntax shown in Table 1 has been simplified for clarity.2 In this query, ?r1, ?r2, ?d1, ?d2 and ?v1 represent variables. Each of them must match the same graph element each time they appear in the query. The sequence function tests that the time codes ?d1 and ?d2 occur in order and within the parameter of 1 s of each other. In our implementation, the analyst has to know SPARQL to specify queries on the trace. To make it simpler, however, we have imple- mented a template mechanism that prepares skeletons of queries. We have also added some customized functions in our implementation of SPARQL, such as the sequence function pre- sented above. These functions facilitate the specification of queries that compare time codes of events, and make it easier to specify temporal constraints. In so doing, we are implementing semantics of time, for example, the semantics of the relation of co-occurrence and of sequential ordering. In the future, we plan to add options to let the analyst specify these queries from a graphical interface based on the visualization of the trace. The Edit tab allows the analyst to visualize the resulting trace in a similar way as the View tab, but it also allows her to reject the trace if she is not satisfied by the result. This feature helps the analyst refine the query in search for the best symbols and inference rules she can get to describe the activity from the actual data. Our system returns the number of times the pattern has matched in the trace. This number indicates the number of new symbols added. The system also provides an export function to a text file that can be imported into other tools like Microsoft Excel for further statistical computa- tions. Queries are saved as independent files and the tool helps the user reference them. The database of queries associated with the ontology constitutes a representation of the analyst’s understanding about how to make sense of the activity trace. 4. Example activity model We report here an example activity modelling analysis taken from a car-driving study (Hen- ning et al., 2007). Another example analysis – in a study of non-state political violence – is reported by Georgeon et al. (2010). Figure 5 shows a 10 s section of a car-driving activity trace focusing on a lane change on a motorway. The legend is presented in Figure 6. In Figure 5, the ‘Button’ is an index signal from the experimenter recorded during the experiment, the start thinking event comes from 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Table 1: Simplified inference query to infer a lane change CONSTRUCT (?r1, infer, Indicator_Symbol) (?r2, infer, Indicator_Symbol) (Indicator_Symbol, type, Lane_Change_Indicator) WHERE (?r1, type, Accelerate) (?r2, type, Left_Mirror_Glance) (?r1, time-code, ?d1) (?r2, tine-code, ?d2) (?r2, Acceleration_Value, ?v1) FILTER (?v1 > 1) (sequence(?d1,?d2,1)) # in order, within 1 second 2The complete SPARQL syntax can be found in the SPARQL documentation, http://www.w3.org/TR/rdf- sparql-query/ c! 2011 Blackwell Publishing Ltd Expert Systems 9 EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M http://www.w3.org/TR/rdf-sparql-query/ http://www.w3.org/TR/rdf-sparql-query/ a verbal signal given by the driver in a video- based post-experiment interview, and the lane crossing event is the moment when the left front wheel crosses the lane, manually encoded from the video. The representation of the driv- ing episode given in Figure 5 is then automati- cally generated with the inference rules defined by the analyst. Using the trace querying facilities of AB- STRACT on this driving data set, we could identify two categories of lane changes that we explain by the performance of two strategies (Figures 7 and 8). In these descriptions, the lower part is a representative trace episode from our database, while the upper part is drawn by hand as an abstract description of the strategies. This upper part also shows the car trajectory in the lanes, respecting the scale ratio length=width: about 300m of 4m wide lanes. The strategy displayed in Figure 7 is character- ized by beginning in a situation where the subject is impeded by a slow vehicle. In this case, the subject starts accelerating [1] almost at the same time as he looks at his left mirror [2]. Then, if there is no vehicle coming from behind, he starts looking at the left lane [3], he switches his blinker on [4] and he performs the lane change [5]. In this situation, the acceleration associated with a glance to the left mirror appears as a good predictor of the lane change. It occurs more than 1 s before the subject switches the blinker on. In the situation of Figure 8, no slow vehicle impedes the subject and he performs the lane change ‘on the fly’. In this case, we can find no behavioural sign of his intention to change lane before the blinker is switched on [1]. Nevertheless, the blinker appears to be a sufficient predictor in that case, because it is switched in anticipation of the lane change, several seconds before the man- oeuvre: looking to the left lane [2], looking to the left mirror [3] and starting steering [4]. In parallel to searching and identifying these categories of situations and strategies, we define symbols to represent them and inference rules to generate these symbols. Finally, we have named the first strategy Lane_change_delayed and represented it with white triangles, and the second strategy Lane_change_anticipated represented 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Figure 5: Example of lane change on a motorway (screenshot with text labels added at the top). First level of abstraction Second level of abstraction Backwards Leftwards Rightwards Frontwards Stable Relation of inference Steering Speed Eye Blinker Gas Brake Obstacle Time Figure 6: Legend of car driving symbolic trace. 10 Expert Systems c! 2011 Blackwell Publishing Ltd EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M with white squares. Figure 9 shows that these two types of event occur four times in a representative 20-min motorway ride of a subject. Figure 9 displays only the symbols that are useful to see the lane changes: the index button pressed by the experimenter (blue circles), the blinker (orange triangles up and down), the accel- erations (ochre triangles to the right), the left mirror glances (grey triangles to the left), main junctions on motorways given by the GPS posi- tion (green square and triangles on the bottom of the display) and the lane change category symbols (white triangles and squares). In this example, one lane change (marked by the verti- cal cursor line in the figure) was not categorized. The uncategorized cases would require further study to understand their specificity. Once the analyst has finished her analysis, she can export the abstract traces into a spreadsheet to com- pute and report the statistics of the occurrences of events of interest. Despite the extensive existing studies of car driving (e.g. Groeger, 2000) and on lane change manoeuvres (e.g. Salvucci & Liu, 2002), we could find no representation of the driving activity that could compare to ours in terms of comprehen- siveness of the data and capacity to support higher-level understanding. In the case of lane changes, this innovative description helped us discover different strategies that have not been reported in the literature before. In so doing, this study shows how our principles and tool have addressed the needs set out in the introduction: our prototype tool helped us generate compre- hensive symbolic representations of the activity at an appropriate abstraction level to discover pre- viously unknown knowledge about the activity. It also helped us explain, report and back up our models with the collected traces. 5. Related work In this section, we situate our work in relation to two research areas: the area of qualitative data 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Figure 8: Lane change anticipated without acceleration (Lane_change_anticipated). Figure 7: Lane change with acceleration (Lane_change_delayed). c! 2011 Blackwell Publishing Ltd Expert Systems 11 EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M analysis and the area of trace-based reasoning (TBR). In the area of qualitative data analysis, many software tools address the need for supporting the transcription of raw data into sequences of encoded events, for example, Dismal (Ritter & Larkin, 1994; Ritter & Wood, 2005), NVivo, INTERACT, InfoScope, MORAE, MarShapa (Sanderson et al., 1994) and MacVisSTA. As such, these tools relate to our collection system as described in Section 3.1. Among these tools, MacVisSTA (Rose et al., 2004) particularly relates to this aspect of our work in that it supports merging multi-modal data into a common timeline. At the symbolic level – also called the transcript level in some studies – tools like Theme (Magnus- son, 2000) support the automatic discovery of temporal patterns based on statistical properties. We consider our tool complementary to these tools because our tool helps find the symbolic pattern based on the meaning they have to the analyst rather than on their statistical properties. In the car-driving example, our symbols of interest are not particularly frequent or infrequent nor do they obey pre-assumed statistical laws. HyperRESEARCH appears to be the only qualitative data analysis tool that supports the validation of hypothetic theories through rule- based expert system techniques (Hesse-Biber et al., 2001). We find HyperRESEARCH’s un- derlying principles for theory building very similar to ours. It, however, does not focus on temporal semantics and does not offer symbolic timeline visualization facilities to support activ- ity modelling. Its rule engine also does not exploit elaborated hierarchical semantics de- fined in an ontology, as opposed to our solution based on SPARQL and RDFSs. The other related research area, TBR (Cordier et al., 2009), comes from the domain of knowledge representation, and more precisely from case- based reasoning (CBR) (Aamodt & Plaza, 1994). CBR consists of helping users solve new problems by adapting solutions that have helped them solve previous problems. TBR extends CBR to retrieve useful cases as episodes from the stream of an activity trace (Mille, 2006). Our work relates to TBR because both address the question of model- ling and representing a stream of activity for future usage. In particular, we pull lessons from the work of Settouti et al. (2009) that implemented a trace-based system to support the management and the transformation of traces. TBR has been used to implement companions that provide assis- tance to the user based on previous usage (e.g. Cram et al., 2008) or to support reflexive learning by providing the users with a dynamic display of his passed activity (Ollagnier-Beldame, 2006). As opposed to these works, our work helps a user (the analyst) understand the activity of another user (the subject). In so doing, our work brings princi- ples of qualitative data analysis to TBR and brings TBR techniques to address problems of quantita- tive data analysis. 6. Conclusion We have defined the principles of a methodology and a software tool to help an analyst create models of activity from activity traces in an iterative and interactive fashion. These principles relate to the notion of abductive reasoning in that they consist of helping analysts form, orga- nize and test micro-hypotheses to explain and represent the activity. Specifically, they help 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 Figure 9: Categorization of lane changes between Lyon and airport (20 min long). 12 Expert Systems c! 2011 Blackwell Publishing Ltd EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M investigate what concepts and semantics describe the activity the best and represent these concepts in an ontology and apply this semantics through a rule engine. These principles are summarized in Figure 10. In Figure 10, the collected raw data are repre- sented as curves along the activity axis. The first analysis step consists of collecting this raw trace. The second step consists of producing a symbolic trace (symbols represented by circles) through the discretization of the raw trace. The third step consists of modelling the symbolic trace by infer- ring more abstract symbols (represented as squares and triangles) and organizing these sym- bols in an ontology (hierarchy of white rectangles). The fourth step consists of producing explained models of activity (round-angled rectangles) that are backed up by the abstract trace. During the modelling process, the analyst formalizes her un- derstanding of the activity in the form of transfor- mation rules, ontology and documentations that are stored in the system, which allows capitalizing on the analyst’s knowledge across studies. To illustrate these principles, we have imple- mented a prototype software tool through an assemblage of open-source knowledge engineer- ing software modules. With this tool, we have modelled car-driving activity traces collected with an instrumented vehicle. This analysis allowed us to identify and describe two strate- gies of lane change on motorways. This example shows that our knowledge engineering approach of activity modelling from activity traces offers answers to the needs for tools to help analysts understand better an observed activity, create models of this activity, report and back up these models with the observational data. The abstract activity traces that we have con- structed constitute a model of the car driver in their own, in that the analysts can use our tool to query these traces to answer new questions they may have about the driving activity. Our future developments will consist of simulating the activ- ity, for instance, in the field of car driving, generating realistic driving behaviour in a driving simulator, based on our abstract activity traces. Acknowledgements We would like to acknowledge support for this project from the European Commission through the HumanIST Network of Excellence. Partial support for this report was provided by ONR (contracts N00014-06-1-0164 and N00014-08-1- 0481). We also thank Jean–Marc Trémeaux for his participation in the software implementation, and Matthias Henning for his contribution to the car-driving experiment. We appreciate the com- ments on this report from Jonathan Morgan and Ryan Kaulakis. References AAMODT, A. and E. PLAZA (1994) Case-based reasoning: foundational issues, methodological variations, and system approaches, AI Communications, 7, 39–59. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 1. Raw trace 2. Symbolic trace 3. Modeled trace 4. Activity models Activity Analysis Figure 10: Activity modelling from activity traces c! 2011 Blackwell Publishing Ltd Expert Systems 13 EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M CORDIER, A., B. MASCRET and A. MILLE (2009) Extend- ing case-based reasoning with traces, Grand Chal- lenges for Reasoning from Experiences, Workshop at IJCAI, Pasadena, CA, pp. 23–32Q1 . CRAM, D., B. FUCHS, Y. PRIÉ and A. MILLE (2008) An approach to user-centric context-aware assistance based on interaction traces, Modeling and Reasoning in Context, Human Centered Processes, Delft, The NetherlandsQ2 . DEARDEN, A. and J. FINLAY (2006) Pattern languages in HCI: a critical review, Human–Computer Interac- tion, 21, 49–102. ERICSSON, K.A. and H.A. SIMON (1993) Protocol Analysis: Verbal Reports as Data, Cambridge, MA: MIT Press. GEORGEON, O., J. MORGAN, J. HORGAN and K. BRAD- DOCK (2010) Process modeling for the study of non- state political violence, 19th Annual Conference on Behavior Representation in Modeling Simulation, Charleston, NC, Brims Society, pp. 240–247Q3 . GROEGER, J. (2000) Understanding Driving, Hove, UK: Psychology Press. HENNING, M.J., O. GEORGEON and J.F. KREMS (2007) The quality of behavioral and environmental indica- tors used to infer the intention to change lanes, 4th International Driving Symposium on Human Factors in Driver Assessment, Stevenson, Washington, USA, pp. 231–237Q4 . HESSE-BIBER, S., P. DUPUIS and S. KINDER (2001) Testing hypotheses on qualitative data: the use of HyperRESEARCH computer-assisted software, Social Science Computer Review, 18, 320–328. HILBERT, D.M. and D.F. REDMILES (2000) Extracting usability information from user interface events, ACM Computing Surveys (CSUR), 32, 384–421. JAMES, W. (1907) PragmatismQ5 . MAGNUSSON, M.S. (2000) Discovering hidden time patterns in behavior: T-Patterns and their detection, Behavior Research Methods, Instruments and Com- puters, 32, 93–110. MILLE, A. (2006) From case-based reasoning to traces- based reasoning, Annual Reviews in Control, 30, 223–232. OLLAGNIER-BELDAME, M. (2006) Traces d’interactions et processus cognitifs en activité conjointe: Le cas d’une co-rédaction médiée par un artefact numérique, CNRS – LIRIS, Lyon, Université Lumière Lyon 2. OLSON, G.M., J.D. HERBSLEB and H.H. REUTER (1994) Characterizing the sequential structure of interactive behaviors through statistical and grammatical tech- niques, Human–Computer Interaction, 9, 427–472. POPPER, K. (1972) Objective Knowledge, Oxford, UK: Oxford University Press. RITTER, F.E. and J.H. LARKIN (1994) Developing pro- cess models as summaries of HCI action sequences, Human–Computer Interaction, 9, 345–383. RITTER, F.E. and A.B. WOOD (2005) Dismal: a spread- sheet for sequential data analysis and HCI experi- mentation, Behavior Research Methods, Instruments, and Computers, 37, 71–81. ROSE, R.T., F. QUEK and Y. SHI (2004) MacVisSTA: a system for multimodal analysis, International Con- ference on Multimodal Interfaces, State College, PA, pp. 259–264. SALVUCCI, D.D. and A. LIU (2002) The time course of a lane change Q6: driver control and eye-movement beha- vior, Transportation Research, F, 123–132. SANDERSON, P.M. and C.A. FISHER (1994) Exploratory sequential data analysis: foundations, Human–Com- puter Interaction, 9, 251–317. SANDERSON, P.M., M.D. MCNEESE and B.S. ZAFF (1994) Handling complex real-word data with two cognitive engineering tools: COGENT and MacSHAPA, Be- havior Research Methods, Instruments, and Compu- ters, 26, 117–124. SETTOUTI, L.S., Y. PRIÉ, J.-C. MARTY and A. MILLE (2009) A trace-based system for technology- enhanced learning systems personalisation, the 9th IEEE International Conference on Advanced Learn- ing Technologies, Riga, pp. 93–97 Q7. The authors Olivier L. Georgeon Olivier L. Georgeon is a research associate in the LIRIS department at the Claude Bernard Uni- versity (Lyon, France). He previously had a 12- year industrial experience as a software engineer, developer and project manager in the domain of automatization of industrial processes. He re- ceived a PhD in cognitive psychology from the Université Lumière (Lyon, France) in 2008. His research interests are in learning through activity. His work both produces practical applications and addresses epistemological questions concern- ing learning from and about an activity. Alain Mille Alain Mille has been a professor at the Claude Bernard University (Lyon, France) since 2000. He is a scientific director in the LIRIS depart- ment (UMR 5205 CNRS), and leads the research group Supporting Interactions and Learning through Experience. After a first career as a computer engineer and a project manager (Hospices Civils de Lyon), Alain Mille was 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 14 Expert Systems c! 2011 Blackwell Publishing Ltd EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M asked to build a computer science department in a college of engineering (CPE-Lyon). In this latter position, he created a research team on case-based reasoning. The main topics were decision helping and complex tasks assistance. Alain Mille is on several national and interna- tional conference programme committees, and on the editorial board of Intellectica (a cognitive sciences journal). He is specifically interested in dynamic modelling of knowledge during com- puter-mediated activities. Thierry Bellet Thierry Bellet has been a researcher at the Ergonomics and Cognitive Sciences Laboratory of IFSTTAR (Institut Français des Sciences et Technologies des Transports, de l’Aménage- ment et des Réseaux) since 1999. His main areas of research concern human activity analysis and cognitive processes modelling (e.g. situational awareness, decision making, cognitive schemas). His research has produced computational simu- lations of mental activities of the car driver (e.g. COSMODRIVE: COgnitive Simulation MOdel of the DRIVEr). Benoit Mathern Benoit Mathern is a PhD student in computer science. He first worked as a computer engineer at the Ergonomics and Cognitive Sciences Labora- tory of IFSTTAR (Institut Français des Sciences et Technologies des Transports, de l’Aménage- ment et des Réseaux), and then he started his PhD work in collaboration with the LIRIS department of the Claude Bernard University (Lyon, France). His main research interests focus on knowledge engineering, knowledge discovery and human– machine interaction. His work is applied to cog- nitive science in the field of transportation. Frank E. Ritter Frank E. Ritter is one of the founding faculty of the College of IST, an interdisciplinary aca- demic unit at Penn State, to study how people process information using technology. Frank Ritter’s current research is in the development, application and methodology of cognitive mod- els, particularly as applied to interface design, predicting the effect of behavioural moderators and understanding learning. He edits the Oxford Series on Cognitive Models and Archi- tectures, is associate editor of Cognitive Systems Research, editorial board member of Human Factors, and the Journal of Educational Psychol- ogy and technical programme co-chair for the BRIMS 2009, 2010 and 2011 conferences and the associated special issues in Computational and Mathematical Organizational Theory. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 c! 2011 Blackwell Publishing Ltd Expert Systems 15 EXSY 584 (B W U K E X SY 5 84 W eb pd f: =0 2/ 11 /2 01 1 04 :4 6: 26 5 51 35 5 B yt es 1 5 PA G E S n op er at or =) 2 /1 1/ 20 11 4 :4 6: 31 P M Author Query Form _______________________________________________________ _______________________________________________________ Dear Author, During the copy-editing of your paper, the following queries arose. Please respond to these by marking up your proofs with the necessary changes/additions. Please write your answers clearly on the query sheet if there is insufficient space on the page proofs. If returning the proof by fax do not write too close to the paper's edge. Please remember that illegible mark-ups may delay publication. Journal EXSY Article 584 Query No. Description Author Response Q1 AUTHOR: Please provide date of workshop for reference Cordier et al. (2009). Q2 AUTHOR: Please provide further details for reference Cram et al. (2008). Q3 AUTHOR: Please provide date of conference for reference Georgeon et al. (2010). Q4 AUTHOR: Please provide the date of symposium for reference Henning et al. (2007). Q5 AUTHOR: Please update the reference James (1907) with further details. Q6 AUTHOR: Please provide volume number for reference Salvucci and Liu (2002). Q7 AUTHOR: Please provide the date of conference for reference Settouti et al. (2009). Olivier July 11 2009 Olivier March 21-24 2010 Olivier July 9-12, 2007 Olivier Unmarked définie par Olivier Olivier Transportation Research part F, 5 (5 is the volume number) Olivier July 15-17 2009 Olivier Cram, D., Fuchs, B., Prié, Y. and Mille, A. (2008, 8-12 June 2008). An approach to User-Centric Context-Aware Assistance based on Interaction Traces. Fifth International Workshop on Modeling and Reasoning in Context., Delft, The Netherlands, TELECOM Bretagne, pp. 89-101. Olivier JAMES, W. (1907) Pragmatism. Re-edition: Pragmatism (Philosophical Classics) Publisher: Dover Publications (1995) New York ISBN-10: 0486282708 128 pages Olivier Please also change the citation in page 3 to (James, 1907/1995) if you think that fits the journal's standards. Page 1 of 3 USING E-ANNOTATION TOOLS FOR ELECTRONIC PROOF CORRECTION Required Software Adobe Acrobat Professional or Acrobat Reader (version 7.0 or above) is required to e-annotate PDFs. Acrobat 8 Reader is a free download: http://www.adobe.com/products/acrobat/readstep2.html Once you have Acrobat Reader 8 on your PC and open the proof, you will see the Commenting Toolbar (if it does not appear automatically go to Tools>Commenting>Commenting Toolbar). The Commenting Toolbar looks like this: If you experience problems annotating files in Adobe Acrobat Reader 9 then you may need to change a preference setting in order to edit. In the “Documents” category under “Edit – Preferences”, please select the category ‘Documents’ and change the setting “PDF/A mode:” to “Never”. Note Tool — For making notes at specific points in the text Marks a point on the paper where a note or question needs to be addressed. Replacement text tool — For deleting one word/section of text and replacing it Strikes red line through text and opens up a replacement text box. Cross out text tool — For deleting text when there is nothing to replace selection Strikes through text in a red line. How to use it: 1. Right click into area of either inserted text or relevance to note 2. Select Add Note and a yellow speech bubble symbol and text box will appear 3. Type comment into the text box 4. Click the X in the top right hand corner of the note box to close. How to use it: 1. Select cursor from toolbar 2. Highlight word or sentence 3. Right click 4. Select Replace Text (Comment) option 5. Type replacement text in blue box 6. Click outside of the blue box to close How to use it: 1. Select cursor from toolbar 2. Highlight word or sentence 3. Right click 4. Select Cross Out Text http://www.adobe.com/products/acrobat/readstep2.html Page 2 of 3 Approved tool — For approving a proof and that no corrections at all are required. Highlight tool — For highlighting selection that should be changed to bold or italic. Highlights text in yellow and opens up a text box. Attach File Tool — For inserting large amounts of text or replacement figures as a files. Inserts symbol and speech bubble where a file has been inserted. Pencil tool — For circling parts of figures or making freeform marks Creates freeform shapes with a pencil tool. Particularly with graphics within the proof it may be useful to use the Drawing Markups toolbar. These tools allow you to draw circles, lines and comment on these marks. How to use it: 1. Click on the Stamp Tool in the toolbar 2. Select the Approved rubber stamp from the ‘standard business’ selection 3. Click on the text where you want to rubber stamp to appear (usually first page) How to use it: 1. Select Highlighter Tool from the commenting toolbar 2. Highlight the desired text 3. Add a note detailing the required change How to use it: 1. Select Tools > Drawing Markups > Pencil Tool 2. Draw with the cursor 3. Multiple pieces of pencil annotation can be grouped together 4. Once finished, move the cursor over the shape until an arrowhead appears and right click 5. Select Open Pop-Up Note and type in a details of required change 6. Click the X in the top right hand corner of the note box to close. How to use it: 1. Click on paperclip icon in the commenting toolbar 2. Click where you want to insert the attachment 3. Select the saved file from your PC/network 4. Select appearance of icon (paperclip, graph, attachment or tag) and close 
work_35yj7gknozerbjznmviwkbytwi	----	PII: 0003-2670(93)80016-E Analydca Chimica Acta, 284 (1993) 131-136 Elsevier Science Publishers B.V., Amsterdam 131 Expert system for the interpretation of infrared spectra G.N. Andreev, O.K. Argirov and P.N. Penchev Department of Chemistry, University of Pkmiiv, 4MW-Plovdiv (Bulgaria) (Received 7th December 1992; revised manuscript received 30th March 1993) Abstract An expert system for the interpretation of infrared spectra EXPIRS was created. The main features of EXPIRS are: hierarchical organization of the characteristic groups, realized by frames; registration of the multiple use of spectral bands; taking into account the solvent absorption and the chemical inconsistencies; documenting the interpretation course and providing explanations on request. The ten most important heuristics used by an expert for interpretation of infrared spectra were formulated and some of them were tested with EXPIRS. Keywords: Infrared spectrometry; Expert systems; Frames; Heuristics; Organic compounds Computer-assisted interpretation of infrared (IR) spectra has drawn the attention of scientists for more than a decade. Several different ap- proaches have been applied, including the utiliza- tion of correlation tables [l-S], symbolic 16-81 and fuzzy [9] logic, expert systems based on rules [lo-14 and refs. cited therein] and table-driven procedure [15], frames [16] and, recently, neural networks [17-191. Most of the authors have for- mulated their results in the terms of probability predictions for the characteristic groups in the studied compound. These systems produce re- sults in the form of tables: functional groups vs. probability. The final decision for the presence of a given substructure was left to the user. In the general case, however, the user is not a specialist in IR spectroscopy. An alternative approach is to base the inter- pretation rules on classical logic, leading to deci- sions clearly formulated by an expert. This can be Correspondence to: G.N. Andreev, Department of Chemistry, University of Plovdiv, 4000-Plovdiv (Bulgaria). achieved using the heuristic knowledge of a hu- man expert in conjunction with the positive char- acteristics of computers. The present work deals with the formulation of the principle heuristics used by an expert to interpret IR spectra and the implementation of some of them in an expert system. BASIC CONCEPTS OF THE EXPERT Until now, we have not found in the literature clear formulated heuristics used by an expert for the interpretation of vibrational spectra. Our ex- perience in structural elucidation of organic com- pounds by IR spectroscopy and the practice with EXPert in InfraRed Spectroscopy (EXPIRS) have led to the following ten most important heuris- tics: (1) Taking into consideration the preliminary information about the studied sample. (2) Correction of the spectral band intensities, when some of the bands are “obviously too inten- sive”. 0003~2670/93/$06.00 8 1993 - Elsevier Science Publishers B.V. All rights reserved 132 G.N. Andreev et aL /Anal. Chim. Acta 284 (1993) 131-136 (3) Comparison between the parameters of the spectrum bands (position, intensity, width) and characteristic group data. (4) Discussion of the alternatives for the func- tional group combinations, possibly existing in the compound at hand, satisfying the spectral data. (5) Discussion of the alternatives for the ex- planation of spectrum bands’ origin: normal vi- brations, overtones, combinations, Fermi reso- nances. (6) Excluding from consideration the bands used during the interpretation of every altema- tive, i.e. single use of each band by discussion of every alternative. (7) Taking into consideration the band shapes as well as the whole spectrum or any of its sectors. (8) Taking into consideration the absorption of moisture and carbon dioxide. (9) Taking into consideration spectrum regis- tration conditions: physical condition (influence upon characteristic group intervals and band in- tensities), solute-solvent interactions, blocking of spectrum intervals caused by solvent absorption, sample concentration (hydrogen bonds), sample thickness, pressure, etc. (10) Taking into consideration non-spectral reasons: chemical inconsistencies of the func- tional groups simultaneously predicted in a given alternative; chemical interaction between the sol- vent used and the predicted functional groups. It should be noted that this enumeration is not in order of importance because the neglection of any heuristic can lead to wrong conclusions. Fur- thermore, such an order should not be accepted as and algorithm, because some of the heuristics must be repeatedly applied in the course of inter- pretation, depending on the nature of the studied molecule. The third heuristic is the only heuristic used in all expert systems developed for the interpreta- tion of IR spectra. Most of these also use some of the other heuristics described above. However, we have not found any utilization of heuristics 5, 7, 8 and 10 whereas 2 and 6 have been used but not in the same manner as an expert would use them. The band of each characteristic group has a different relative intensity depending on the other groups included in the same molecule. For exarn- ple, y(CH) modes are very strong in the IR spectra of aromatic hydrocarbons but they often appear medium to strong in the spectra of aro- matic carbonyl compounds. The common solution for this problem of the system for automated interpretation is enlargement of the intensity in- terval in the knowledge base. However, this auto- matically brings one to the so-called “hyperpre- diction”. On the other hand, the expert detecting the presence of the “very strong” v(C0) modes “expands” the intensity of the other spectrum bands instead of extending the intensity intervals in his mind. The application of this principle in the expert system will avoid the hyperprediction. Heuristic 2 takes into account the latter ap- proach. The results obtained from EXPIRS testing pointed out that the most significant reduction of the hyperprediction can be achieved by the appli- cation of heuristic 6. EQUIPMENT AND MATERL4L.5 EXPIRS was developed on an IBM-PC com- patible computer with 640 kbyte RAM and the program was written in PASCAL (the system is available on request). 200 spectra from the Sadtler Fourier transform (FT)_IR library were used to test the system as well as spectra registered in our laboratory on a Perkin Elmer 1750 FT-IR spec- trometer with a resolution of 2 cm-‘. More than 70 characteristic groups containing the elements C, H, 0 and N were incorporated in the system and the appropriate rules for their interpretation were programmed. Program description The expert system is based on the concept of “characteristic group intervals”. [al. There are three levels in EXPIRS (Fig. 1). The first two levels involve the process of interpretation; the third one (documentation) is designed to register the interpretation course and the number of spectrum peaks used during the work of the program. G.N. Andreev et al. /Anal. Chim. Acta 284 (1993) 131436 133 UUCUMENTATION \...I . . . . . f . . . . . . . . . . . . . . . . . . . . . t . . . I . . . . . 1 I Fig. 1. Flow chart of EXPIRS. . . . . DATA - The data for the characteristic intervals, sol- vent absorption and chemical inconsistency were realized as arrays of the appropriate variables. Analysis of the literature, combined with our experience, reveals that the groups are best orga- nized hierarchically. Such an approach avoids re- peating the interpretation of the spectral inter- vals. For example, the primary XI-I,-OH, sec- ondary > CH-OH and tertiary X-OH alcohols have a common interval at 3606-3200 cm- ’ due to the stretching vibration of the hydroxyl group v(OH). The triple verification of this interval can be avoided if the characteristic OH group, which determines this interval, is used as a “parent group”. In this case, the rules for the different alcohols (primary, secondary and tertiary) must take into account not only the entire spectrum, but also the status of the parent group OH. Such an approach was utilized in the description of other characteristic groups, including alkanes, alkenes, amines, aldehydes, etc. This type of group organization corresponds to the chemist’s knowl- edge that the primary, secondary and tertiary alcohols are special cases of the concept “alcohol”. In other words such an organization reflects the different levels of abstraction of the chemical structure in the chemist’s mind. We have used frames to realize the hierarchi- cal organization of the characteristic groups in our expert system (Fig. 2). The principal frame in r+g:f$$g Fig. 2. Hierarchical organization of the frames in EXPIRS. the program is the frame “Characteristic group”; two specifications of this frame are given in Fig. 3 for the groups sp3C-H and CH,. When the system checks for the presence of the methylene group CT-I, it needs data for its parent group sp3C-H. If the status of the parent group has yet to be discovered, the procedure starts with its frame, etc. There are no limitations to the depth of such parent group checks. This makes the system independent on the group or- der in the data base and on the user’s range of demand, which differentiate EXPIRS from the approaches utilized elsewhere [ 10,16,21]. One can readily see (Fig. 3) that the parent group and its “generics” belong to the same frame and the hierarchy mentioned above is de- Characteristic group: Name : String for the screen: Conclusion for presence: Parent group: Name : co”clusion for presence Characteristic intervals: Interval No 1: Characteristic Arcmp: Name : String for the screen: ca”clusio” for presence: Parent group: Name: ‘Concl”.io” for preaance Characteristic iararva1s: Interval rio I: SPJCH spx-Ii Received by a procedure based on data for the characteristic intervals, spectrum, solvent absorption. Absent Absent 3000-2800 cm-l weak to strong Cb CHZ Received by a procedure based on data for the parent grou tic i.t.w.P& characteris- spectrum, solvent absorption. sp3cli Received 1480-1430 cm-l weak to * trong Fig. 3. Two entities from the frame “characteristic group”: sp3C-H and CH,. 134 termined not as frame hierarchy, but by means of the connection “parent group” between the dif- ferent specifications of the same frame “char- acteristic group”. Operation of the system The interaction between the user and the sys- tem is determined by the appropriate menus. There are three main options: INTERPRETA- TION, CONCLUSIONS, SAVING THE RE- SULTS. In the option INTERPRETATION the user: - puts in the name of the spectrum file; inter- action mode can be selected; - defines the intensity margins for strong, medium and weak peaks; - puts in the solvent or the disturbing media for the measured sample; - gives preliminary information (based on analysis or the origin of the sample) for the absence of some groups as well as the kind of the groups of interest. The option CONCLUSIONS provides infor- mation on the groups whose presence in the studied molecule is positive, negative or uncer- tain. The user can request explanation about the arguments on which the conclusion is based. Ad- ditional information on positively predicted groups which are not chemically coincident is available. In the option SAVING THE RESULTS user can save on disk or print the results of interpretation. RESULTS AND DISCUSSION the the We have obtained excellent to satisfactory agreement between the predicted and existing characteristic groups by testing the developed expert system with the IR spectra, measured in our laboratory or included in the Sadtler FT-IR library. The following examples of EXPIRS’ work il- lustrate the importance of some heuristics de- scribed above. G.h? Andreev et al. /Anal. Chim. Acta 284 (1993) 131-136 The interpretation of the spectrum of DL-2- methylbutanoic acid [CH,CH,CH(CH3)COOH] gives a report for the presence of the follow- ing characteristic groups: sp3C-H, CH,, CH,, COOH, alkyl-COOH. As seen EXPIRS identi- fies all three functional groups of the studied substance. It further specifies that the carboxylic group is aliphatic. The presence of C-H with sp3-hybridized carbon atom is also mentioned. Along with the correct predictions we also received some incorrect answers by the spectral interpretation. The latter can be divided into two different types: (1) negative statement for a char- acteristic group existing in the studied molecule; (2) positive statement for a characteristic group that is absent in the studied compound. We found that one could eliminate the first kind of incorrect answers by improving the data base. The errors of the second type, which appear more frequently (hyperpredictions), could not be avoided by the same manner. The latter are con- nected with the multiple use of the same bands in the course of interpretation. Our program was supplied with a counter in order to register the multiplicity of band usage. The hyperprediction will be illustrated by the interpretation of the spectrum of isopropylben- zene. EXPIRS reveals the existence of the follow- ing characteristic groups: sp3C-H, CH,, CH,, i-Pr, =C-H, C=C, cis-CH=CH, aryl, Ph-, o-aryl. This interpretation is based on the spectrum bands given in the following format: wavenum- ber(bandwidth) - multiplicity of band’s usage: 3083(m) - 1, 3064(m) - 1, 3029(s) - 1, 2963(s) - 1, 2929(s) - 1, 2889(s) - 1, 2873(s) - 1, 2802(m) - 1,1949(m) - 0,1872(w) - 0,1804(m) - 1,1744(w) - 0, 1665(w) - 1, 160%) - 2, 1533(w) - 0, 1496(m) - 0, 1465(s) - 1, 1451(s) - 1, 1386(m) - 3,1364(m) - 3,1320(m) - 1,1300(m) - 1,1279(m) - 1, 1208(w) - 0, 1102(m) - 0, 1079(m) - 1, 1048(m) - 2,1027(m) - 2,921(m) - 0,905(m) - 0, 777(w) - 0, 761(s) - 4, 697(s) - 4, 531(m) - 1, where s = strong, m = medium and w = weak. The multiple use of the 1605, 761 and 697 cm-’ bands is the reason for the C=C, c&CH=CH and o-aryl groups hyperprediction. Obviously, one could eliminate such incorrect predictions by tak- G.N. Andreev et al. /Ad Chim. Acta 284 (1993) 131-136 ing into account the multiple use of each spectral band. The ability of EXPIRS to take into considera- tion the influence of the solvent or dispersion agent used will be illustrated by the interpreta- tion of the spectrum of o-glucose measured in nujol mull (wavenumber/ relative intensity band- width): 3408/0.810, 3306/0.833br, 2925/0.905, 2855/ 0.828, 1459/0.720, 1376/0.662, 1340/ 0.630, 1296/0.568, 1224/O-568, 1203/0.570, 1149/ 0.695, 1111/0.751, 1078/0.611, 1050/ 0.753, 1024/ 0.824, 996/ 0.798, 916/ 0.597, 838/ 0.592, 775/0.559, 723/0.531, 614/0.709. The following message appears after the interpre- tation: (1) There are spectral data for PRESENCE of the following groups: OH, RCH,-OH, R&H- OH, R&-OH, Ar-OH, R-O-R, R-0-Ar, > N-H; (2) The presence of the following groups is UNCERTAIN: sp3C-H, CH,, (CH,),, CH,, i-Pr, t-Bu, Ar-NH. The second report deals with the groups whose characteristic intervals overlap with those of the dispersion media (or solvent used). An important feature of EXPIRS is that the user can follow the logic of the interpretation upon request. If the user wants to provide an explanation for any of the above results, he can use the option EXPLANATION. He goes to the option CONCLUSIONS and designate the corre- spoding group of interest. For example, the fol- lowing messages will appear for the groups R&H-OH and CH,: RXH-OH . . . . . . . . . . . . . . . . . . . . . . . . . . . REPORT: The conclusion for presence of OH is POSITIVE Is there a strong band between 1120 and 1080? YES: 1111 (s) There are spectral evidences for pres- ence of R&H-OH Press any key to continue.. . CI=................................. REPORT: The conclusion for presence of sp3C- H is UNCERTAIN 135 TOTAL OVERLAPPING WITH THE SOLVENT BANDS The presence of the CH2 is UNCER- TAIN Press any key to continue.. . This possibility makes the expert system appro- priate for educational purposes as well. The authors wish to thank the Bulgarian Na- tional Fund of Research at the Ministry of Edu- cation and Science for partial financial support through Grant No. X-124/91. REFERENCES 1 T. Visser and J.H. van der Maas, Anal. Chim. Acta, 122 (1980) 363; Anal. Chim. Acta, 133 (1981) 451. 2 M. Farkas, J. Markos, P. Szepesvaty, I. Bartha, G. Szalon- tai and Z. Simon, Anal. Chim. Acta, 133 (1981) 19. 3 G. Szalontai, Z. Simon, Z. Csapo, M. Farkas and Gy. Pfeifer, Anal. Chim. Acta, 133 (1981) 31. 4 B. Debska, J. Duliban, B. Guzowska-Swider and Z. Hippe, Anal. Chim. Acta, 133 (1981) 303. 5 W.-R. Leupold, C. Domingo, W. Niggemann and B. Schrader. Freszenius’ Z. Anal. Chem. 303 (1980) 337. 6 7 8 9 10 11 12 13 14 15 16 17 L.A. Gribov and M.E. Elyashberg, J. Mol. Struct., 5 (1970) 179; J. Mol. Struct., 50 (19781351. L.A. Gribov, M.E. Elyashberg and L.A. Moscovkina, J. Mol. Struct., 9 (1971) 357. K. Funatsu, Y. Susuta and S. Sasaki, Anal. Chim. Acta, 220 (1989) 155. T. Blaffert, Anal. Chim. Acta, 161 (1984) 135. H.B. Woodruff and G.M. Smith, Anal. Chem., 52 (1980) 2321. S.A. Tomellini, D.D. Saperstein, J.M. Stevenson, G.M. Smith, H.B. Woodruff and P.F. Seelig, Anal. Chem., 53 (1981) 2367. S.A. Tomellini, R.A. Hartwick and H.B. Woodruff, Appl. Spectrosc., 39 (1985) 330. B.J. Wythoff, C.F. Buck and S.A. Tomellini, Anal. Chim. Acta, 217 (1989) 203. H.J. Luinge, Vib. Spectrosc., 1 (1990) 3. M.O. Trulson and M.E. Munk, Anal. Chem., 55 (1983) 2137. P. Edwards and P.B. Ayscough, Chemom. Intell. Lab. Syst., 5 (1988) 81. E.W. Robb and M.E. Munk, Mikrochim. Acta (Wien) I, (19901 131. 18 M.E. Munk, M.S. Madison and E.W. Robb, Mikrochim. Acta (Wien) II, (1991) 505. 19 K.J. Fessenden and L. Gyorgyi, J. Chem. Sot. Perkin Trans 2, (1991) 1755. 136 G.N. Andreev et aL /Ad Chim. Acta 284 (1993) 131-136 20 N.B. Colthup, L.H. Daly and S.E. Wiberlay, Introduction to Infrared and Raman Spectroscopy, Academic Press, New York, 1975; L.J. Bellamy, The Infrared Spectra of Complex Molecules, Methuen, London, 1964, K. Nakan- ishi, Infrared Absorption Spectroscopy, Holden-Day, San Francisco, CA, and Nankodo, Tokyo, 1962. 21 H.J. Luinge and J.H. van der Maas, Anal. Chim. Acta, 223 (1989) 135. 
work_3665qv37y5dqbnw4huvnhsg4mu	----	1052-2 Expert System Aid in Differentiating Among Paroxysmal Supraventricular Tachycardias lACC February 1995 ABSTRACfS l8lA were found to result in a variability below 5% in parameters measured di- rectly from the average waveform, and up to 10% in those obtained from the time derivative. Subsequently, the feasibility of an automated version of the algorithm, based on objective operator independent criteria, was evalu- ated, and parameters obtained were found to be in excellent agreement with those obtained using manual approach. In summary, this algorithm provides a fast. easy and objective method for noise reduction in acoustic quantifica- tion signals. This algorithm may improve the on-line noninvasive assessment of systolic and diastolic LV function. Computer Aided Instructions: II Tuesday, March 21,1995,9:00 a.m.-12:30 p.m. Ernest N. Morial Convention Center, Hall B Results of Reduced Antlthrombotic Therapy Following Intra Coronary Stenting Tuesday, March 21, 1995, 10:30 a.m.-Noon Ernest N, Morial Convention Center, La Louisiane A 10:30 1741-1 1 Clinical Experience with Heparin-Coated Stents - The Benestent II Pilot Phase 1 H~kan Emanuelsson, Patrick W. Serruys, Jorge Belardi, Hans Bonnier, Antonio Colombo, Jean Fajadet, Jean-Jacques Goy, Guy Heyndrickx, Peter de Jaegere, Victor Legrand, Cad as Macaya, Pierre Materne, Wolfgang Rutsch, Ulrich Sigwart, Harry Suryapranata, enestent Study Group. Division of Cardiology, Sahlgrenska Hospital, G6teborg, Sweden; Thoraxcenter, Erasmus Univ. Rotterdam, The Netherlands Krzysztof P. Wr6blewski, Zhi Yun Tian. Children's Hospital of Philadelphia, Philadelphia, PA Congenital heart disease (CHD) is the most common congenital malforma- tion with an incidence rate of 0.7%. Many pregnant woman (in some coun- tries all) are offered an ultrasound scan at around 18 weeks of pregnancy. The scan incorporates a detailed anatomical survey of the fetus and if in- cludes at least a four chamber view, it is an excellent opportunity to detect congenital heart defects. Therefore a Windows based multimedia computer program for the instruction of Fetal Echo for ultrasound technologists, stu- dents. residents, fellows and primary care physicians has been developed. The application includes a step by step tutorial to instruct the user in reading Fetal Echo data, a browsing library of definitions of CHD's, graphics images and digitized echocardiograms, and the Expert System for automatic diagno- sis. The pictures, images and descriptions are linked to a database and can be viewed by searching for the defect name or symptoms. This provides a powerful instructional tool in this very difficult area of medical education. The minimum hardware requirements to run this program are: IBMIPC or compatible computer running MS Windows version 3.1 with 80386SX pro- cessor,4 MB RAM, 30 MB free disk space, multimedia with dual speed CD ROM, and a VGA card capable to display at least 256 colors. Steven Georgeson. Somerset Medical Center, Somerville, NJ The differentiation between atrioventricular reciprocating tachycardia via an accessory pathway (AVRT). AV nodal reentrant tachycardia (AVNRT) and atrial tachycardia (AT) in paroxysmal supraventricular tachycardia may be helpful in guiding pharmacologic therapy and in identifying patients for radiofre- quency catheter ablation. To help in this differentiation, an expert system was developed using a commercially-available expert system shell (EXSl'S). A simplified version of this system was designed to run on a palm-top com- puter. Both programs use the MS-DOS operating system. The expert system is rule-based and assigns probability values to the goal states through the process of backward chaining. The goal states for this expert system were AVRT, AVNRT, and AT. The user is queried for the presence of various abnor- malities on the presenting EKG (P wave location, ORS alternans, pseudo r wave in lead V1, pseudo S wave in the inferior leads), comparison with pre- vious EKG's (presence of pre-excitation) and the effect of vagal manueuvers or adenosine infusion on the tachycardia. From published data, each abnor- mality is assigned a probability based on the positive predictive value of the abnormality for AVRT, AVNRT and AT. By combining the positive predictive values, the expert system assigns a final probability for AVRT, AVNRT and AT. This expert system may be useful as a diagnostic tool and a teaching aid in the differentiation between AVRT, AVNRT and AT in paroxysmal supraven- tricular tachycardia. 10:45 Full Antiplatelet Therapy without Anticoagulation After Coronary Stenting The purpose of the Benestent II Pilot Phase was to explore the safety of reducing antithrombotic therapy in conjunction with implantation of heparin- coated stents. The study consists of three phases, where resumption of hep- arin therapy after stent implantation was progressively delayed in a stepped care approach. Material and Methods. Palmaz-Schatz stents with heparin coating were implanted in 51 patients (88% male) with stable angina pectoris. Heparin treatment was withheld 6 hours following removal of the sheath introducer from the femoral artery. The mean age was 59 years, 10% had a previous myocardial infarction, diabetes was prevalent in 8%, hypertension in 24% and 59% were current or previous smokers. Target lesion was located in the left anterior descending artery in 51 %, left circumflex in 8% and right coronary artery in 41 %. TIM I flow I-II was present in 14% and TIMIIlI in 86%. The mean pre-procedural minimal lumen diameter (MLD) was 1.10 mm and reference diameter 3.16 mm. Results. Stent implantation was successful in all patients and following the procedure, mean MLD increased to 2.77 mm and percent diameter stenosis was reduced from 65% to 19%. The maximum balloon size was 3.45 ± 0.40 mm. Post-stent dilatation was performed in 40 patients (80%). High pressure (> 12 atm) was used in 22 of these cases (54%). Two patients needed a sec- ond stent· one for an occlusive distal dissection and one for a distal lesion. There we;e no major complications, i.e. death, myocardial infarction, urgent CABG, re-PTCA or cerebrovascular accident. Peripheral vascular complica- tions required vascular surgery in 3 patients (5.9%) and blood transfusion in one (2%). Conclusions. In this pilot study implantation of heparinized stents was as- sociated with a 100% success rate, absence of serious complications and a moderate incidence of vascular complications. Further reduction of an- tithrombotic treatment may be feasible. Jean-Marc Lablanche, Gilles Grollier, Nicolas Danchin, Jean-Louis Bonnet, Eric Van Belle, Eugene Me Fadden, Michel E. Bertrand. Universib'esofUlle, Caen, Nancy and Marseille, France Subacute thrombosis remains a major limitation of coronary stenting. In addi- tion, local complications related to the intensive anticoagulation that is com- monly employed are frequent. We prospectively studied a regime of intensive anti platelet therapy (as- pirin 200 mg daily begun before percutaneous transluminal coronary angio- plasty, ticlodipine 500 mg daily for 3 months begun just after angioplasty) with periprocedural dextran infusion continued for a period at the discretion of the local investigator. To date, 98 patients (85 men, 13 women) undergo- ing 102 procedures involving 125 stents have been enrolled; 71 patients had one stent implanted; 19 patients had 2 stents implanted, and 2 patients had 3 stents implanted. Symptoms before coronary angioplasty included effort angina (23%), angina at rest (6%), unstable angina (23%), recent myocardial infarction (27%). The indications for stent implantation were occlusive dis- section 36%, dissection without occlusion 32%, suboptimal result or elec- tive implantation 32%. The stented site was the LAD (37%). RCA (35%). LCx (17%), and vein graft (11 %). Stent types were Wiktor (70%), Palmaz-Schatz (23%). Gianturco-Roubin (7%). The diameter of the stents implanted was 2.5 mm (2%), 3.0 mm (32%), 3.5 mm (44%), and 4.0 mm (22%). There were 2 deaths; 1 patient died from cardiogenic shock that was present before stent implantation 1 patient commited suicide; 4 patients developed q-wave AMI, 1 was already complete at the time of stent implant. 3 occurred In the hours after stent implant for occlusive dissection; 8 patients had non-O AMI; 3 pa- tients had CABG, 1 at 5 hours post stent for reocclusion (non-O wave AMI). 2 were performed electively without sequelae for an unsatisfactory angio- graphic result but without ongoing ischemia. Blood transfusion for peripro- Multimedia Instructional System for Fetal Echo Expert System Aid in Differentiating Among Paroxysmal Supraventricular Tachycardias 11052-1 1 1 1052-21 
work_37btiicl5fhs5lb2yfxzlmafnu	----	Local-Shapelets for Fast Classification of Spectrographic Measurements Local-Shapelets for Fast Classification of Spectrographic Measurements Daniel Gordona,c,∗, Danny Hendlera, Aryeh Kontorovicha, Lior Rokachb,c aDepartment of Computer Science, Ben-Gurion University of The Negev Be’er Sheva 84105, Israel bDepartment of Information Systems Engineering, Ben-Gurion University of The Negev Be’er Sheva 84105, Israel cTelekom Innovation Laboratories, Ben-Gurion University of The Negev Be’er Sheva 84105, Israel Abstract Spectroscopy is widely used in the food industry as a time-efficient alternative to chemical test- ing. Lightning-monitoring systems also employ spectroscopic measurements. The latter appli- cation is important as it can help predict the occurrence of severe storms, such as tornadoes. The shapelet based classification method is particularly well-suited for spectroscopic data sets. This technique for classifying time series extracts patterns unique to each class. A signif- icant downside of this approach is the time required to build the classification tree. In addition, for high throughput applications the classification time of long time series is inhibitive. Although some progress has been made in terms of reducing the time complexity of building shapelet based models, the problem of reducing classification time has remained an open challenge. We address this challenge by introducing local-shapelets. This variant of the shapelet method restricts the search for a match between shapelets and time series to the vicinity of the location from which each shapelet was extracted. This significantly reduces the time required to examine each shapelet during both the learning and classification phases. Classification based on local- shapelets is well-suited for spectroscopic data sets as these are typically very tightly aligned. Our experimental results on such data sets demonstrate that the new approach reduces learning and classification time by two orders of magnitude while retaining the accuracy of regular (non- local) shapelets-based classification. In addition, we provide some theoretical justification for local-shapelets. Keywords: Spectrography, time series, classification, shapelets, local Research highlights • We present an algorithm for classifying spectrographic measurements. • The concept of locality is introduced into an established time series algorithm. • A technique for estimating a tolerance parameter is presented. ∗Corresponding author. Tel: +972 (0)86428782; fax: +972 (0)86477650; Email addresses: gordonda@cs.bgu.ac.il (Daniel Gordon), hendlerd@cs.bgu.ac.il (Danny Hendler), karyeh@cs.bgu.ac.il (Aryeh Kontorovich), liorrk@bgu.ac.il (Lior Rokach) Preprint submitted to Expert Systems with Applications November 13, 2014 • Learning and classification times are reduced by two orders of magnitude. • Accuracy levels are retained. 1. Introduction Spectroscopy is a field devoted to the study and characterization of physical systems by measuring the electromagnetic frequencies they absorb or emit (Herrmann and Onkelinx, 1986). Items differing in their chemical composition or molecular bonds absorb or emit light at different wavelengths leaving a different spectroscopic fingerprint thus enabling differentiation between them. For example, Al-Jowder et al. (2002) used mid-infrared spectroscopy to detect meat adul- teration by comparing the spectra of adulterated meat with that of unadulterated meat. A study by Briandet et al. (1996) discriminated between two different types of coffee beans (Arabica and Robusta) using mid-infrared spectroscopy. Other methods for distinguishing between different types of food exist, which are based on wet chemical analysis (Bicchi et al., 1993; Lumley, 1996; Sharma et al., 1994). The advantages of spectroscopy over wet chemical analysis are in its sim- plicity (Briandet et al., 1996) and speed of response. Spectroscopic measurements are also generated by systems monitoring lightning (Eads et al., 2002). This application is important as relative percentages of different types of lightning can indicate the outbreak of severe storms, such as tornadoes. In addition to laboratory research, spectroscopic equipment is starting to be mass produced for every day use allowing anyone to analyze their surroundings with the aid of spectroscopic measurements (SCIO, 2014). The measurements are uploaded to a cloud service where they are analyzed and then the results of the analysis are made available. The service is cloud based, requiring algorithms with high throughput to enable a quick response to high volumes of queries by users. The outcome of the spectroscopic analysis of a physical system is a vector in which each index represents a frequency and each value is the measured intensity of that frequency. The representation of spectroscopic measures and time series are identical (Ye and Keogh, 2011a), as the only explicit data are the measurements and the meaning of each measurement is defined by its location in the vector. This equivalence allows the application of time series classifica- tion methods to the field of spectroscopy. A previous experimental study (Hills et al., 2013) showed that the shapelet based classification method is particularly suited for data sets from the field of spectroscopy, as it achieved a higher accuracy than other machine-learning classification methods. Recently, Ye and Keogh (2011a) introduced the shapelets approach for classifying time se- ries. A shapelet is a subsequence extracted from one of the time series in the data set. The shapelet is chosen by its ability to distinguish between time series from different classes. A test time series is classified based on its distance from the shapelet. In the case of multiple shapelets, these form the nodes of a classification tree. The intuition behind this approach is that the pattern best separating the classes is not necessarily an entire time series. Rather, a certain subsequence may best describe the discriminating pattern. Ye and Keogh’s algorithm considers all possible subsequences in the training set in order to identify those shapelets that yield the optimal split. Through the rest of this paper, we will refer to this algorithm as the YK-algorithm. Two key advantages of classification with shapelets are the accuracy and interpretability of the induced classification model, as it supplies information on the patterns characteristic of the different classes (Ye and Keogh, 2011a). A significant downside of this approach is the time re- quired for building the classification tree (Hills et al., 2013; Mueen et al., 2011; Rakthanmanon and Keogh, 2 0 50 100 150 200 250 Time series index A b so rb a n ce (a) Coffee data set 0 200 400 600 800 1000 Time series index A b so rb a n ce (b) Wheat data set Figure 1: Examples of two data sets (coffee and wheat) from the field of spectroscopy. The time series of each class are vertically separated and in different colors. As shown, the examples of each class are tightly aligned, i.e., similar patterns are exhibited at similar locations along the x-axis. 2013). The search for the best shapelet requires examining all subsequences of all lengths from all the time series in the training set, and for each shapelet calculating its distance to each time series at all possible locations. Even for small data sets, this process has a time scale of days, and for large data sets, the time scale becomes one of years. Hence, the original implementation on commonly available hardware is only practical on the smallest of data sets. Additionally, for high-throughput applications, the classification time may be prohibitively expensive for long time series. This is because at each node of the tree, all possible matches between the node’s shapelet and the time series to be classified need to be examined. 1.1. Our Contributions Our goal was to reduce both learning and classification time without impairing the accuracy of the induced model by exploiting a feature common to spectroscopic data – the localized nature of information in the time series. In the YK-algorithm, no importance is attributed to the location from which the shapelet was extracted. Hence, the best match between a shapelet and a time series is searched for anywhere in the time series. We observed that for many data sets from the field of spectroscopy, time series from the same class show similar behavior patterns at similar locations along the frequency axis. Fig. 1 presents examples of two data sets which strongly support this insight. Based on this insight, we propose a new property as part of the definition of a shapelet, derived from the location in the time series from which the shapelet was extracted. This property limits the scope of the search for the best match of a shapelet to a time series to the vicinity of the location from which the shapelet was extracted. The assumption of locality is justified as spectroscopic measurements of items with similar properties should have very similar spectroscopic fingerprints, especially in areas characteristic of a specimen which are not expected to be contaminated. Our current implementation assumes that all time series are of equal length. 3 Although the time series are generally aligned, some allowance for misalignment is neces- sary. We therefore introduce a method for learning the misalignment characteristic of a data set. We evaluate our approach on data sets from the field of spectroscopy, and show that local- shapelets can reduce learning and classification time (especially for data sets with long time series) by over two orders of magnitude without impairing accuracy. For reproducibility, we have made all our code available online (local shapelets, 2014). The rest of the article is organized as follows: First we present basic definitions required for understanding the article and shortly describe the YK-algorithm in Sect. 2. Then we present related work (Sect. 3), followed by a description of local-shapelets and our proposed method for determining the range to examine (Sect. 4). Next we present our experimental evaluation (Sect. 5) followed by a brief statistical analysis, which provides a theoretical justification for our local-shapelets approach (Sect. 6). Finally, we summarize our results and present additional research directions to pursue (Sect. 7). 2. Background Here we present a number of definitions necessary for the proper understanding of this article and a short description of the original shapelet algorithm as it is the basis of our work. 2.1. Definitions Definition 1. A time series T of length m is a series of m consecutive equally spaced measure- ments: T = t0, t1, ..., tm−1. Definition 2. A subsequence S of length k extracted from time series T of length m at index i such that k ≤ m is a series of k consecutive measurements: S = ti, ti+1, ..., ti+k−1. Definition 3. The Euclidean distance between two time series T,R of length m is: dE (T,R) = √√√ m−1∑ i=0 (ti − ri) 2 . Definition 4. The Euclidean distance between a time series T of length m and a subsequence S of length k such that k ≤ m is: dE (T,S ) = min i dE (T [i : i + k − 1],S ) i∈[0,m−k]. This is the minimal distance between S and all subsequences of length k in T . Definition 5. Given a data set D, with c classes and n examples, each class i with ni examples, the entropy is: Ent(D) =− c−1∑ i=0 ni n log ni n . 4 Intuitively, a data set’s entropy is a measure of its class-homogeneity. Larger homogeneity corresponds to smaller entropy values. Specifically, the smallest entropy value (0) corresponds to a data set in which all members belong to the same class. Definition 6. Given a data set D with n examples, split into two subsets D1 and D2, containing n1 and n2 examples respectively, such that D1 ∪ D2 = D and D1 ∩ D2 = ∅ the information gain (IG) is: IG(D, D1, D2) = Ent(D) − (n1 n Ent(D1) + n2 n Ent(D2) ) . Intuitively, information gain is a measure of the class-homogeneity induced by a split of data set D. The larger the information gain, the larger the decrease in entropy of the split data set w.r.t. D, in turn implying better class-homogeneity. As we will soon see, each shapelet induces a data set split and its effectiveness is measured by the split’s information gain. 2.2. The YK-Algorithm For completeness, we briefly present the YK-algorithm as presented in Ye and Keogh (2011a). First, we present the algorithm for two classes; we then extend the description to a multi-class data set. Let D be a dataset with two classes and n time series. The YK-algorithm examines all possi- ble subsequences of every length (from a minimal length, usually 3, to a maximal length which is usually the length of the shortest time series) from every time series. For each subsequence S , the distance to each time series is calculated, as defined in Definition 4. Then, the time series are ordered by their distance from S . Using this induced order, the average distance of every two adjacent time series to S is calculated. We will refer to this average distance as the splitting distance. Each of the n splitting distances defines two subsets, one containing all time series with a distance to S smaller than or equal to the splitting distance, and the other containing all time series with a distance to S greater than the splitting distance. For every possible split into two subsets, the information gain is calculated (see Definition 6). If the current information gain is better than the best so far, the shapelet is kept along with the corresponding splitting distance. Tie breaking is done by keeping the shapelet which induces a larger average distance between the two subsets which is referred to as the margin. After checking all possible subsequences, the best shapelet and the corresponding splitting distance are returned. This method can be easily extended to a multi-class problem by building a tree, with a shapelet and splitting distance in each node. A new node receives one of the two subsets created by the shapelet found by the node above it, and learns the best shapelet and splitting distance for this subset of time series. The stopping criteria for this recursive algorithm is that all the time series in the subset be of one class. Two important implementation issues are that all distance calculations are computed after local normalization (Goldin and Kanellakis, 1995) and that the margin is normalized, by dividing it by the length of the subsequence. Classification of a time series T is accomplished by traveling down the tree. At each node the distance of the shapelet S associated with the node to T is calculated. The node decides to which of its child nodes T should be directed, depending on whether its distance from S is smaller or greater than the splitting distance. When T reaches a leaf, it is assigned the class associated with this leaf. 5 2.2.1. Time Complexity of the YK-Algorithm Let m denote the length of a time series and let us assume all time series are of equal length. Assuming all shapelet lengths from 3 to m are examined, the number of different shapelets to examine in a single time series is ∑m i=3 i = O(m 2). Let n denote the number of time series in data set D. The number of shapelets to examine in the entire data set is O(nm2). When searching for the minimal distance between a shapelet S of length k and a time series T , the distance of S to all subsequences of length k in T needs to be calculated (see Definition 4). The time complexity of this operation is O(m2). Calculating the distance of S to all time series requires O(nm2) calculations. The total number of calculations for all shapelets and time series is O(n2m4) which explains the formidable time required for learning a model even for small data sets. 3. Related Work The time complexity of the YK-algorithm is formidable (see Sect. 2.2.1) leading to a large number of attempts to reduce it. As we will show in this section, none of the previous approaches utilized the location from which the shapelet was extracted to reduce the time required to learn a model. In addition, most of these approaches do not reduce the time required to classify a time series. The first attempt to reduce the time complexity of the YK-algorithm was introduced in the paper first presenting shapelets (Ye and Keogh, 2011a). Two optimizations were suggested. The first optimizes the distance calculation of a shapelet to a time series. The distance calculation is terminated if it exceeds the minimum distance found so far between the current shapelet and time series. This optimization was coined early-abandon. The second optimization (named entropy- pruning) checks whether the most optimistic IG possible, given the distances of a shapelet to time series already computed, can be better than the best IG found so far. If the IG cannot be improved, the shapelet under examination is discarded. As pointed out by Lines et al. (2012) this optimization requires testing O(2c) different possibilities (c is the number of classes in the data set), which can greatly reduce the effectiveness of this optimization when the number of classes is large. Later, Mueen et al. (2011) introduced additional optimizations. One optimization manages to compute the distance of a shapelet to a time series in constant time by precomputing necessary statistics. This optimization manages to reduce the time complexity to O(n2m3). A major down- side is that for each two time series, a matrix of size m2 needs to be maintained. This leads to a total space complexity of O((nm)2) which is untenable for large data sets (Gordon et al., 2015; Rakthanmanon and Keogh, 2013). A second optimization discards shapelets similar to shapelets that were already discarded. A disadvantage of this approach is the large time overhead when applied to data sets with a large number of classes. Two recent attempts managed to dramatically reduce the time complexity for learning a shapelet based model. The first method (Rakthanmanon and Keogh, 2013) quickly picks out a small number of shapelets from each shapelet length which seem able to effectively divide the data set into its classes. Then only this subset of shapelets are fully analyzed. This approach manages to reduce the time complexity to O(nm3) but requires a considerable amount of space to accommodate this reduction in time complexity. We will refer to this solution as the hashing- algorithm. A second approach (Gordon et al., 2015), named SALSA-R, randomly samples a constant number of shapelets (10,000) which are examined, reducing the time complexity to O(10,000 × nm2) with no excess space requirements. 6 As shown, none of the aforementioned optimizations utilize the location from which the shapelet was extracted to reduce the time complexity of the learning process. In addition none of them significantly reduces classification time. Xing et al. (2011) introduced an optimization for fast classification of streaming time series with the aid of shapelets. Their main idea was to prefer shapelets which appear early in a time series over shapelets which appear later. This ensures that time series can be classified quickly once initial measurements have arrived with no need to wait for further measurements. They coined this new type of shapelets as local-shapelets. Although we both name our shapelets similarly, the concepts differ. The motivation of Xing et al. was to classify streaming time series as early on as possible while ours is to optimize learning and classification time of non-streaming data sets. Also, the implementations differ. Xing et al. did not preserve the location from which the shapelet was extracted. Conversely, with our method, the location from which the shapelet was extracted is exploited with no limitation on the location from which to extract a shapelet. 4. Local-Shapelets In the material that follows, the basic idea of local-shapelets is described as well as modifi- cations which make it useful in practice. Definition 7. A shapelet is a tuple <~S ,d>. ~S is a series of consecutive measurements extracted from one of the time series in the data set and d is a cutoff distance. Time series with a distance to S smaller or equal to d traverse one side of the tree while time series with a distance to S greater than d traverse the other side of the tree. Definition 8. A local-shapelet is a tuple <~S ,d, i>. ~S and d are as in Definition 7. i is the location from which the local-shapelet was extracted. Unlike a shapelet, a local-shapelet contains information regarding the location from which it was extracted, which is utilized when calculating the distance of the local-shapelet to a time series. Definition 9. The Euclidean distance between a time series T of length m and a local-shapelet <~S ,d, i> of length k such that k ≤ m given x which defines a range around i is: dE (T,S ) = min j dE (T [ j : j + k − 1],S ) j∈[max(0,i−x),min(i+x,m−k)]. The distance between T and a local-shapelet <~S ,d, i> adds a constraint on the subsequences of T to which the distance of subsequence ~S is calculated. Instead of calculating the distance of ~S to all subsequences of length k in T , the distance of ~S is calculated only to subsequences in the vicinity of the location i from which ~S was extracted. This vicinity is defined by the constant x. Thus, the time required for calculating the distance of a shapelet to a time series is reduced. 4.1. The Tolerance Range A naı̈ve implementation of local-shapelets is to calculate the distance of a shapelet only to the single subsequence in the time series at the exact location i from which it was extracted. This approach may have detrimental impact on the learning process as exemplified in Fig. 2 which presents two time series from the same class of the data set Lightning7. As is clearly shown, the characteristic spike may appear at slightly different locations. Restricting the distance calculation 7 Time series index 0 100 200 300 A b so rb a n ce (a) Example from class 2 Time series index 0 100 200 300 A b so rb a n ce (b) Example from class 3 Figure 2: Two examples from different classes of the Lightning7 data set illustrating the need for a tolerance range. Without a tolerance range the spike characteristic of each class cannot be utilized as its location is different in different time series. of a shapelet to the exact location from which it was extracted would cause characteristic patterns to be overlooked. To accommodate this issue some tolerance, which we will refer to as the radius, needs to be added to the index i, such that the distance of a local-shapelet to a time series T is the minimum distance to all subsequences starting in the range [i − radius, i + radius] (see definition 9). We will refer to this range as the tolerance range. In food spectroscopy there are many sources of noise, such as differences in the residual water content of the freeze-dried samples (Briandet et al., 1996). This noise may cause distortions in the pattern generated but will not cause a shift in the frequencies emitted, therefore we set the tolerance range to 0. In lightning spectroscopy, interferences such as frequency-dependent dispersion induced by the ionosphere (Moore et al., 1995) may lead to a shift in the frequencies measured requiring the introduction of a tolerance range greater than 0. We present a method for computing the value of the tolerance required, as described in Pro- cedure 1. This is achieved by experimentally observing the tolerance required for a small number of subsequences. Procedure 1 splits each time series into ten equal consecutive and disjoint sub- sequences (line 1). From each class (C) and for each subsequence location, a single subsequence (subseq) is extracted from a randomly selected time series (rand-ts) (lines 2-6). Then the distance of subseq to all time series in C is calculated using the global method for calculating distances (see Definition 4). The location of the best match of subseq to each time series is recorded in locations (line 8). Before utilizing the information available in the list of locations, it is necessary to filter out values which are obviously non-characteristic of the data set (line 9). A major motivation is that a larger tolerance leads to an increase in learning and classification time as more distance calculations are required for each shapelet. Filtering out locations which are atypical should not impair accuracy significantly. For example, let us suppose we receive the following list of locations: 1,30,30,31,32,34,37,38,40,40,41,81. It is quite clear that the values 1,81 are outliers 8 Procedure 1 Algorithm for computing the tolerance characteristic of a data set Compute-Tolerance(D) {Input is a data set} 1: start-indices ← indices, s.t. time series will be split into 10 subsequences which are as equal in length as possible 2: for each class C in D do 3: for each start-index i in start-indices do 4: subseq-length ← length of subsequence 5: rand-ts ← randomly selected time series from C 6: subseq ← rand-ts[i : i+subseq-length-1] 7: for each time series t in class C do 8: locations ← append(location-of-best-distance(t, subseq)) 9: locations ← outlier-filter(locations) {Using IQR-filtering} 10: Radiusc,i ← max(locations) - min(locations) 11: Radiusc ← min(Radiusc,i) 12: tolerance ← max(Radiusc) 13: return tolerance and should be filtered out. Without filtering, the size of the range is 81, while after filtering, the size of the range is only 12. Our method for outlier filtering is based on the interquartile range (IQR). Using this method, the first (Q1) and third (Q3) quartiles are calculated and the IQR is calculated as I QR = Q3 − Q1. All values greater than Q3 +3× I QR or smaller than Q1 −3× I QR are filtered out. Although it is customary to use 1.5 as the multiplication factor, we chose a multiplication factor of 3 so as to not filter out too many values which may lead to overfitting. Our tuning of the multiplication factor to a value of 3 is a heuristic as it cannot guarantee total avoidance of overfitting, because other parameters such as the model complexity also play a part. Three advantages of IQR filtering are that it is simple, that it is a-parametric and that it does not automatically drop extreme values if they are similar to the rest. Once we have filtered out the outliers, the characteristic radius as reflected by this subse- quence of class C is calculated and recorded (line 10). After a radius for each of the subse- quences of a class has been calculated, the minimum of all these radii is selected to represent the locality of the class (line 11). We chose the minimal radius as this leads to the smallest number of distance calculations of a shapelet to a time series, which promises the best possible speedup in runtime. The last stage is the selection of a single radius as the tolerance for the data set. We chose the maximum radius from all classes (line 12) as the tolerance must accommodate the most loosely localized class. Otherwise, for some of the classes, the ultimate matches may reside outside of the recommended range and will not be examined. The time complexity of this phase is negligible as only 10 subsequences per class are exam- ined and the distance of each subsequence is calculated only to time series of its class. 4.2. Random Selection of Shapelets For completeness, we present the procedure used by SALSA-R for randomly selecting shapelets in Procedure 2. We chose a distribution similar to the uniform distribution but simpler to imple- ment. Procedure 2 describes the method for randomly selecting the next shapelet to examine. First (line 2), the time series from which to extract the shapelet is chosen. Then (lines 3-4), the 9 Procedure 2 Algorithm for randomly selecting shapelets extract-shapelet(D,min-sh-length) {Input is a data set and the minimum length of a shapelet} 1: num-ts ← number-of-time-series-in-data-set(D) 2: ts-index ← random-selection(0,num-ts) 3: highest-index ← times-series-length - min-sh-length + 1 4: sh-index ← random-selection(0,highest-index) 5: longest-possible-shapelet ← times-series-length - sh-index +1 6: sh-length ← random-selection(min-sh-length,longest-possible-shapelet) 7: sh ← extract-shapelet(D,ts-index,sh-index,sh-length) 8: return sh index in the time series from which to extract the shapelet is randomly generated. The range of possible indices is between 0 and the last index from which the shortest possible shapelet can be extracted. Last (lines 5-6), the length of the shapelet is randomly selected. The upper limit on the length of the shapelet (longest-possible-shapelet) is calculated based on the location from which the shapelet is to be extracted (line 5). The function random-selection(a,b) randomly selects an integer from the range [a,b-1] with a uniform distribution. 5. Experimental Results Our goal is to show that for data sets from the field of spectroscopy, the usage of local- shapelets reduces the time complexity during both training and classification phases in com- parison with global-shapelets (i.e., non-local shapelets) without degrading accuracy. First, we present the data sets from the field of spectroscopy with which we evaluated local-shapelets. Then, we establish the utility of our method (Procedure 1) for calculating the tolerance range. In the next phase of our evaluation we compare local-shapelets vs. global-shapelets within the YK- algorithm. In the last phase, we re-implement SALSA-R to utilize the locality of shapelets. We compare accuracy and run-time with the original implementation of SALSA-R and the hashing- algorithm. We ran all experiments on an Intel Xeon E5620 computer comprising two 2.40GHz processors with 24GB of RAM and with a 64-bit version of Windows Server 2008 R2 as the operating system. 5.1. Description of Data Sets Our experiments were conducted on a collection of 6 data sets from the field of spectroscopy, available online (Ye and Keogh, 2011b; Keogh et al., 2014). The collection contains four data sets from the field of food spectroscopy (Beef, Coffee, OliveOil, Wheat) and two from the field of lightning spectroscopy (Lightning2, Lightning7). Table 1 contains information on the number of examples in the training and test sets, the number of classes, the length of the time series and the tolerance range used. All data sets were already split into train and test sets. We preserved the original division to train and test sets to allow easy reproduction of our results, as well as a fair comparison with other published results. As the test sets are very small, our initial measurements of classification times were inaccu- rate due to minor overheads which dampened the effect of locality on the outcome and due to the inaccuracy of computer time measurements at small time scales. We solved this by enlarging each test set to a size of 1GB. This was done by duplicating examples. To ensure that the ac- curacy obtained would be identical to that on the original data set, we duplicated each example 10 an equal number of times. As the classifier is deterministic, the classification of an example will always be identical no matter how many times it appears. Therefore, the proportion of correct classifications out of all classification examples will not change and the accuracy will remain the same. Table 1: Description of the data sets dataset train set test set num. time series tolerance size size classes length Beef 30 252,510 5 470 0 Coffee 28 383,264 2 286 0 Lighting2 60 106,140 2 637 0 Lighting7 70 209,510 7 319 78 OliveOil 30 208,770 4 570 0 Wheat 49 115,434 7 1,050 0 5.1.1. Food Spectrographs Beef. The beef data set (Al-Jowder et al., 2002) contains the spectral absorbance of one type of beef cut (silverside). One class contains the spectral absorbance of the beef cut without any contaminates. Each of the other four classes contains the spectral absorbance of the beef cut contaminated with a different type of offal (kidney, liver, heart and tripe) which is cheaper than the declared beef cut. Coffee. The coffee data set (Briandet et al., 1996) contains the spectral absorbance of instant coffee of two different types of coffee beans Arabica and Robusta. Coffee from Arabica beans is more highly estimated as it has a finer and more pronounced taste. Approximately 90% of world coffee production is from Arabica and another 9% is from Robusta. As the price of Arabica is higher than that of Robusta, it is important to be able to distinguish between them even after the long process that is required to produce instant coffee. OliveOil. Olive oil samples from four different European countries (Greece, Italy, Portugal and Spain) were collected (Tapp et al., 2003). The classification task is to be able to discern the country of origin using the spectrograph of the olive oil sample. Wheat. This data set (Ye and Keogh, 2011a) consists of spectrographs of wheat grown during the years 1998-2005. There are a number of different types of wheat in the data set but the class was assigned based only on the year in which the wheat was grown and not on the type of wheat. 5.1.2. Lightning Spectrographs Data on frequencies emitted during lightning events were collected and then a Fourier trans- form was applied to produce spectrographs (Eads et al., 2002). The lightning events were cat- egorized into 7 different classes differing in the charge of the lightning (positive or negative), whether the event was gradual or abrupt and whether the event was intra-cloud or from cloud to ground. The original authors of this data set (Eads et al., 2002) note that there is a large inter- class variation and intra-class similarity. The data set Lightning7 contains examples of all 7 classes, while Lightning2 is a simpler binary problem of distinguishing between cloud to ground events and intra-cloud events. 11 5.2. Utility of Tolerance Range Calculation The last column of Table 1 presents the tolerance range used during our experiments. For the four data sets of food spectroscopy, the value of the tolerance range was not calculated using Procedure 1; rather we set it to 0 based on prior knowledge in this domain. A comparison of the tolerance range based on prior knowledge with those recommended by Procedure 1 shows large agreement. For three data sets (Coffee, OliveOil and wheat) values are identical. For the Beef data set, although the calculated tolerance range was not exactly 0, it was very close with a value of 2. Our procedure also succeeds when a tolerance range greater than 0 is required. When applied to the Lightning7 data set for which it is clear a tolerance range larger than 0 is required, as shown in Fig. 2, the calculated tolerance range is 78. These findings show that our method for predicting the tolerance range manages to successfully estimate the required range. 5.3. Local YK-Algorithm Here we show that local-shapelets reduce the time consumption without impairing accuracy. We compare local and global shapelets within the YK-algorithm framework which is the initial implementation of the shapelet algorithm which examines all possible shapelets. Due to the large time and space requirements of the YK-algorithm we could not collect results for the Lightning2 data set. For these experiments, we used the original code used by Ye and Keogh (2011a). As pointed out by Hills et al. (2013), the entropy-pruning optimization (see Sec. 3) has an overhead which grows exponentially with the increase in the number of classes in the data set. We encountered this experimentally with the wheat data set which has 7 classes. The original implementation did not finish examining all shapelets for the first node of the tree after 5 days, while the same implementation without the entropy-pruning optimization finished learning the whole tree in this period of time. Therefore we conducted the experiments using the original code without the entropy-pruning optimization. Results of our experiments are presented in Table 2. Each two columns compare results of global-shapelets vs. local-shapelets. The first comparison is the accuracy achieved, the second is the time required to learn a model and the third is the time required to classify the test set. In all measures, the local approach outperforms the global approach. The average improvement in accuracy is 8%, the average speedup during the learning phase is 9.5 and during the classification phase it is 80. Table 2: Global YK-algorithm vs. Local YK-algorithm data set accuracy (%) learning time (sec) classification time (sec) global local global local global local Beef 46.67 56.67 24,307 1,364 76.03 1.10 Coffee 96.43 92.86 1,466 407 36.79 1.54 Lighting7 43.84 53.42 98,636 67,660 109.75 59.58 OliveOil 33.33 50.00 17,448 2,287 46.85 1.01 Wheat 57.71 65.43 433,630 25,221 158.36 0.59 5.4. Local-SALSA-R In this set of experiments we re-implemented SALSA-R to use local-shapelets instead of global-shapelets. The number of shapelets randomly selected was set to 10,000 which was found 12 to be an optimal number by Gordon et al. (2015). We compared local-SALSA-R with global- SALSA-R and the hashing-algorithm. We repeated our experiments thirty times as each of the three methods includes an element of randomness. Table 3 presents a comparison of the average accuracy of local-SALSA-R with global- SALSA-R and the hashing-algorithm. We applied a Friedman test (Friedman, 1937) to test if there is a significant difference between accuracy attained by the different methods. The p-value was 0.85, which clearly shows that there is no significant difference in the accuracy achieved by any of the methods. Table 3: Comparison of accuracy of local-SALSA-R with that of global-SALSA-R and the hashing-algorithm data set global-SALSA-R hashing-algorithm local-SALSA-R Beef 51.78 50.78 62.11 Coffee 94.17 92.74 97.02 Lighting2 67.98 67.22 66.17 Lighting7 58.45 61.62 59.13 OliveOil 73.11 73.33 73.33 Wheat 66.72 69.94 66.46 A comparison of average learning and classification times is presented in Table 4. A Fried- man test on the learning and classification times shows that there is a significant difference be- tween the methods during both the learning (p-value = 0.0057) and classification phases (p-value = 0.030). A one sided Wilcoxon-test (Wilcoxon, 1945) affirms our claim that local-shapelets are significantly faster than global-shapelets during the learning phase (p-value = 0.016 for both global-SALSA-R and the hashing-algorithm) and the classification phase (p-value = 0.016 for global-SALSA-R and p-value = 0.031 for the hashing-algorithm). On average, local-SALSA-R reduces the time required to learn a model by a factor of 120 and 440 in comparison with global- SALSA-R and the hashing-algorithm, respectively. During classification, local-SALSA-R is 180 times faster than global-SALSA-R and 110 times faster than the hashing-algorithm, on average. Table 4: Comparison of learning and classification time of local-SALSA-R with that of global-SALSA-R and the hashing-algorithm data set learning time (sec) classification time (sec) global- hashing- local- global- hashing- local- SALSA-R algorithm SALSA-R SALSA-R algorithm SALSA-R Beef 44.98 169.77 0.47 87.46 104.87 0.51 Coffee 16.15 12.81 0.26 31.61 34.06 0.47 Lighting2 164.09 539.11 1.00 144.78 86.10 0.85 Lighting7 56.51 205.17 35.39 152.61 95.97 111.30 OliveOil 59.39 114.37 0.50 65.52 45.27 0.51 Wheat 315.19 1693.48 1.17 231.46 78.71 0.42 Analytically, it is easy to argue that the longer the time series, the greater the time saved by using local-shapelets, as the number of distance calculations avoided increases. We confirmed this experimentally as can be seen in Fig. 3. The figure shows the ratio between the time required by global-SALSA-R and local-SALSA-R as a function of the length of the time series for all data sets for which the tolerance range used was 0 (all data sets apart from Lightning7). This 13 figure clearly confirms that the ratio increases with the length of the time series. We chose to compare local-SALSA-R with global-SALSA-R and not with the hashing-algorithm as their implementation apart from the aspect of locality is identical, allowing isolation of locality as the only parameter influencing the outcome. Time series length T im e s p e e d u p 0 200 400 600 800 1000 0 1 0 0 3 0 0 5 0 0 Learning time Classification time Figure 3: Each point is the ratio of the time required by global-SALSA-R and local-SALSA-R. One plot is for learning times and the second for classification times. As can be seen the ratio increases with the length of the time series. 6. Statistical analysis The approach proposed in this paper was mainly motivated by algorithmic considerations: restricting the search to a small subset of the possible shapelet locations significantly speeds up both training and classification. In this section, we will argue that as a by-product, our approach offers statistical advantages as well. By restricting the number of features, we are constraining the complexity of the hypothesis class. As we show below, hypothesis classes of low complexity require fewer training examples to attain a certain accuracy level. The argument is made precise in the language of learning theory. A learner faced with a classification task is trying to learn a function g : X → {−1,1}, where X is the instance space (in our case, it is the set of all possible time series). The learner gets to observe example-label pairs (Xi,Yi) ∈ X× {−1,1} generated iid from some unknown distribution P over X× {−1,1}. This corresponds to the intuition that the training labels may be noisy, and indeed, there may be no “correct” classifier g : X → {−1,1} that achieves perfect accuracy. Although there are universal approximators capable of fitting arbitrary labeled samples, if unconstrained they will necessarily overfit (Devroye et al., 1996). Hence, when choosing the learning model, its richness (i.e., hypothesis complexity) must be taken into account. 14 The learner’s n observed labeled examples (Xi,Yi) constitute the training set, based on which it will produce a hypothesis g : X→{−1,1}. We will denote by H the collection of all admissible hypotheses (formally, H ⊂ 2X) and associate with every h ∈ H two key quantities: its sample (or training) error, êrr(h) = 1 n n∑ i=1 1{h(Xi ),Yi} and generalization error, err(h) = E[1{h(Xi ),Yi}] = P(h(X) , Y ). In words, êrr(h) is the relative fraction of mistakes that h makes on the training set while err(h) is the probability that h makes a mistake on a freshly drawn (X,Y ) pair — crucially, drawn from the same distribution used to generate the training set. Note that while the typical goal is to guarantee a small generalization error, the latter quantity cannot be computed without knowledge of the sampling distribution. Instead, the readily computable êrr(h) may be used (under certain conditions, detailed below) as a proxy for err(h). In this setting, the learner’s task is twofold: (i) algorithmic: efficiently find an h ∈ H for which êrr(h) is small, and (ii) statistical: guarantee that, with high probability, err(h) will not be much greater than êrr(h), regardless of which h ∈H the learner chooses. The foregoing sections were devoted to the algorithmic aspects, and we shall focus on the statistical one here. In the case of finite H, a particularly simple connection exists between err(h) and êrr(h): Theorem 1 (Mohri et al. (2012)). Suppose that |H|<∞ and the learner observes a training set consisting of n examples. Then, for any δ > 0, we have that err(h) ≤ êrr(h) + √ log |H|+ log(1/δ) 2n (1) holds with probability at least 1 −δ, uniformly over all h ∈H. For simplicity, let us consider the case where H = 2X (i.e., H consists of all possible binary functions). In this case, |H| = 2|X|, and hence even a modest reduction of the instance space — by reducing the feature set, for example — can have a noticeable effect on the second term in the right-hand side of (1), and hence yield a faster convergence rate. The basic features used in this paper are distances from a shapelet to a time series. It is precisely this feature set that gets reduced when our algorithm considers only a subset of the possible locations. This observation provides a statistical justification to our local-shapelets approach, in addition to the algorithmic speedup. 7. Conclusions The objective of our investigation was to utilize the localization of characteristic patterns in spectrographic measurements using the shapelet algorithm, which has previously been found to be suited (Hills et al., 2013) for this domain. Our adaption to the shapelet algorithm reduces the number of distance calculations and thus shortens the time required to train a model and classify examples. 15 As pointed out by Ye and Keogh (2011a), one important advantage of the shapelet approach is its interpretability, i.e., the process extracts informative subsequences representative of each class. The chosen shapelets provide insights into patterns characteristic of each class. One such example can be seen in Fig. 4. The shapelet, shown as a dashed red line, is overlaid on each of the two time series at the location from which it was extracted and represents a pattern characteristic of class 2. In addition to identifying the discriminative pattern, local-shapelets also identify the discriminative frequencies. We compared the frequencies found to be discriminative by the shapelet with those found to be discriminative by Briandet et al. (1996) and found that they coincide. Time series index A b so rb a n ce 160 180 200 220 240 260 Class 1 Class 2 Figure 4: Interpretability of local-shapelets. Two time series from the coffee data set are presented in black. Each time series is from a different class. Only part of each time series is presented so as to focus on the important details. Overlaid on each of the time series is the shapelet selected by local-SALSA-R as most discriminative appearing as a dashed red line. As can be seen the shapelet represents time series from class 2 (Robusta coffee beans). The algorithm we presented searches for matches of a shapelet on a time series only in the vicinity of the location from which the shapelet was originally extracted. We proposed an algo- rithm (Procedure 1) for calculating the exact vicinity based on properties of each class in the data set. Our main result is that local-shapelets can indeed speedup both the learning and classifica- tion methods 100-fold when using SALSA-R without impairing accuracy. We also show that our estimation of the vicinity to examine is quite accurate. This research can be extended in many ways. It may be interesting to explore possible trade- offs between speedup and accuracy as a function of the vicinity to examine. Another research direction is the application of local-shapelets to domains other than spectroscopy which may also require the adaption of our algorithm for time series of different lengths. 16 References R. Herrmann, C. Onkelinx, Quantities and units in clinical chemistry: Nebulizer and flame properties in flame emission and absorption spectrometry (Recommendations 1986), Pure and Applied Chemistry 58 (12) (1986) 1737–1742. O. Al-Jowder, E. K. Kemsley, R. H. Wilson, Detection of adulteration in cooked meat products by mid-infrared spec- troscopy, Journal of Agricultural and Food Chemistry 50 (6) (2002) 1325–1329. R. Briandet, E. K. Kemsley, R. H. Wilson, Discrimination of Arabica and Robusta in instant coffee by Fourier transform infrared spectroscopy and chemometrics, Journal of Agricultural and Food Chemistry 44 (1) (1996) 170–174. C. P. Bicchi, A. E. Binello, M. M. Legovich, G. M. Pellegrino, A. C. Vanni, Characterization of roasted coffee by S- HSGC and HPLC-UV and principal component analysis, Journal of Agricultural and Food Chemistry 41 (12) (1993) 2324–2328. I. Lumley, Authenticity of meat and meat products, in: Food Authentication, Springer, 108–139, 1996. N. Sharma, A. Srivastava, J. Gill, D. Joshi, Differentiation of meat from food animals by enzyme assay, Food Control 5 (4) (1994) 219–221. D. R. Eads, D. Hill, S. Davis, S. J. Perkins, J. Ma, R. B. Porter, J. P. Theiler, Genetic algorithms and support vector machines for time series classification, in: International Symposium on Optical Science and Technology, International Society for Optics and Photonics, 74–85, 2002. SCIO, http://www.consumerphysics.com/myscio/, 2014. L. Ye, E. Keogh, Time series shapelets: a novel technique that allows accurate, interpretable and fast classification, Data Mining and Knowledge Discovery (2011a) 1–34. J. Hills, J. Lines, E. Baranauskas, J. Mapp, A. Bagnall, Classification of time series by shapelet transformation, Data Mining and Knowledge Discovery (2013) 1–31. A. Mueen, E. Keogh, N. Young, Logical-shapelets: an expressive primitive for time series classification, in: Proceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ACM, 1154–1162, 2011. T. Rakthanmanon, E. Keogh, Fast Shapelets: A Scalable Algorithm for Discovering Time Series Shapelets, in: Proceed- ings of the Thirteenth SIAM Conference on Data Mining (SDM), SIAM, 668–676, 2013. local shapelets, ftp://www.ise.bgu.ac.il/, 2014. D. Goldin, P. Kanellakis, On similarity queries for time-series data: Constraint specification and implementation, in: Principles and Practice of Constraint Programming CP’95, Springer, 137–153, 1995. J. Lines, L. M. Davis, J. Hills, A. Bagnall, A shapelet transform for time series classification, in: Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ACM, 289–297, 2012. D. Gordon, D. Hendler, L. Rokach, (in press) Fast and Space-Efficient Shapelets-Based Time-Series Classification, Intelligent Data Analysis 19 (5). Z. Xing, J. Pei, S. Y. Philip, K. Wang, Extracting Interpretable Features for Early Classification on Time Series., in: Eleventh SIAM International Conference on Data Mining (SDM), SIAM, 247–258, 2011. K. R. Moore, P. C. Blain, S. D. Briles, R. G. Jones, Classification of RF transients in space using digital signal processing and neural network techniques, in: SPIE’s 1995 Symposium on OE/Aerospace Sensing and Dual Use Photonics, International Society for Optics and Photonics, 995–1006, 1995. L. Ye, E. Keogh, shapelet data sets, http://alumni.cs.ucr.edu/~lexiangy/shapelet.html, 2011b. E. Keogh, Q. Zhu, B. Hu, H. Y., X. Xi, L. Wei, C. A. Ratanamahatana, The UCR Time Series Classification/Clustering Homepage, www.cs.ucr.edu/~eamonn/time_series_data/, 2014. H. S. Tapp, M. Defernez, E. K. Kemsley, FTIR spectroscopy and multivariate analysis can distinguish the geographic origin of extra virgin olive oils, Journal of Agricultural and Food Chemistry 51 (21) (2003) 6110–6115. M. Friedman, The use of ranks to avoid the assumption of normality implicit in the analysis of variance, Journal of the American Statistical Association 32 (200) (1937) 675–701. F. Wilcoxon, Individual comparisons by ranking methods, Biometrics Bulletin 1 (6) (1945) 80–83. L. Devroye, L. Györfi, G. Lugosi, A probabilistic theory of pattern recognition, vol. 31 of Applications of Mathematics (New York), Springer-Verlag, New York, ISBN 0-387-94618-7, 1996. M. Mohri, A. Rostamizadeh, A. Talwalkar, Foundations of machine learning, MIT press, 2012. 17 http://www.consumerphysics.com/myscio/ ftp://www.ise.bgu.ac.il/ http://alumni.cs.ucr.edu/~lexiangy/shapelet.html www.cs.ucr.edu/~eamonn/time_series_data/ Introduction Our Contributions Background Definitions The YK-Algorithm Time Complexity of the YK-Algorithm Related Work Local-Shapelets The Tolerance Range Random Selection of Shapelets Experimental Results Description of Data Sets Food Spectrographs Lightning Spectrographs Utility of Tolerance Range Calculation Local YK-Algorithm Local-SALSA-R Statistical analysis Conclusions 
work_37p7nlfftrhp3cya4xhtjt7c7y	----	An expert system for diagnostics and estimation of steam turbine components' condition K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) 70-81 © 2020 WIT Press, www.witpress.com ISSN: 2056-3272 (paper format), ISSN: 2056-3280 (online), http://www.witpress.com/journals DOI: 10.2495/EQ-V5-N1-70-81 AN EXPERT SYSTEM FOR DIAGNOSTICS AND ESTIMATION OF STEAM TURBINE COMPONENTS’ CONDITION KONSTANTIN E. ARONSON, BORIS E. MURMANSKY, ILIA B. MURMANSKII & YURI M. BRODOV Ural Federal University named after the First President of Russia B.N. Yeltsin, Ekaterinburg, Russia ABSTRACT This article describes an expert system of probability type for diagnostics and state estimation of steam turbine technological subsystems’ components. The expert system is based on Bayes’ theorem and permits one to troubleshoot the equipment components, using expert experience, when there is a lack of baseline information on the indicators of turbine operation. Within a unified approach, the expert system solves the problems of diagnosing the flow steam path of the turbine, bearings, thermal expan- sion system, regulatory system, condensing unit, and the systems of regenerative feed-water and hot water heating. The knowledge base of the expert system for turbine unit rotors and bearings contains a description of 34 defects and 104 related diagnostic features that cause a change in its vibration state. The knowledge base for the condensing unit contains 12 hypotheses and 15 pieces of evidence (indica- tions); the procedures are also designated for 20 state parameters’ estimation. Similar knowledge bases containing the diagnostic features and fault hypotheses are formulated for other technological subsys- tems of a turbine unit. With the necessary initial information available, a number of problems can be solved within the expert system for various technological subsystems of steam turbine unit: for steam flow path, it is the correlation and regression analysis of multifactor relationship between the vibration and the regime parameters; for thermal expansion system, it is the evaluation of force acting on the longitudinal keys depending on the temperature state of the turbine cylinder; for condensing unit, it is the evaluation of separate effect of the heat exchange surface contamination and of the presence of air in condenser steam space on condenser thermal efficiency performance, as well as the evaluation of term for condenser cleaning and for tube system replacement. With the lack of initial information, the expert system formulates a diagnosis and calculates the probability of faults’ origin. Keywords: diagnostic, diagnostic features, expert system, evidence, faults, hypotheses, steam turbine. 1 INTRODUCTION At present, the research on the state parameters of turbine unit equipment, forecasting of their changes and determination of their residual life has become widespread [1–3]. In world prac- tice, technical condition monitoring and equipment diagnostics are carried out by the centers operating at the manufacturing plant or in a large operating organization. At the same time, the working methods of these centers at the enterprises have been developed individually for many years and are a commercial secret. Steam and gas turbine manufacturers in the territory of Russia currently do not have such centers [4, 5]. Some works of certain enterprises are known concerning individual areas of diagnosis [6–8]. The lack of common approaches to the development of diagnostic tasks for various elements of equipment makes it difficult to implement them at thermal power plants. Expert systems (ES) seem to be the most ambitious approach for many diagnostic tasks. This methodology is connected with a huge volume of factors of steam turbines operating. ES are very advantageous for steam turbine unit (STU) diagnosis. These systems are designed to solve problems that are difficult to formulate. The ES is based on Bayes’ theorem and permits one to troubleshoot the equipment components when there is a lack of baseline information on the indicators of turbine unit operation. The system also employs the experi- ence of experts. The inaccuracy and lack of initial information are taken into account by K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) 71 probabilistic methods using Bayes’ formula. The value of each piece of evidence is deter- mined by the Naylor method [9–12]. 2 METhODOLOGY An ES comprises a knowledge base and an information processing algorithm. The knowl- edge base contains information about STU failures as fault hypotheses and a table of evidence. The a priori probability of fault hypotheses and the evidence value are determined by the experts. The ES analyzes the evidence. If information is missing, the system receives it from a user or out of a database. The user sets the values of pieces of evidence and the ES calcu- lates a posteriori probabilities of hypotheses and forms a conclusion about the cause of failure. The system then makes recommendations to the staff about how to eliminate the malfunction. According to Bayes’ theorem [9], a posteriori probability of the hypothesis is calculated by the formula: P(h/E) P(E/h) P(h) P(E) = ⋅ , (1) where Р(Е/Н) is the probability of evidence Е, if the hypotheses h is true; Р(Н) stands for a priori probability of the hypotheses h; and P(E) stands for the probability of evidence E. P(E) is determined by the formula of total probability: P(E) = P(E/h)×P(h) + P(E/ h )×P( h ), (2) where Р(Е/ h ) is the probability of evidence Е, if the hypotheses h is false: Р( h ) = 1 – Р(Н). From (1) and (2), a posteriori probability of the hypothesis is calculated: P h E P E h P h P E h P h P E h P h ( / ) ( / ) ( ) ( / ) ( ) ( / ) ( ) = ⋅ ⋅ + ⋅ . (3) The ES comprises the knowledge bases designed for turbine flow part, for turbine bear- ings, for system of thermal expansions, for automatic regulatory system, for condensing unit, and for the systems of regenerative feed-water heating and hot water heating. For diagnostics, knowledge bases for various defects were collected. They were divided according to the tur- bine subsystem: vibration diagnostics, control system, heat expansion system, auxiliary equipment, etc. Each subsystem contains its own list of defects and diagnostic features. For example, the knowledge bases of the ES for vibration diagnostics of turbine rotors, bearings and other components include a description of 34 defects and 104 diagnostic features. These defects cause a change of vibration state of the turbine unit. All defects are divided into two groups: defects that occur during the turbine operation and defects added during turbine mounting and repair. The defects of turbine unit mounting and repair are identified in the analysis of start–stop actions. The defects of operation are revealed in the analysis of turbine component vibration, vibration changes or the relationship of vibration to the turbine operation. A number of connections are used to diagnose the system of steam distribution and turbine regulatory system (see Table 1). On the basis of these relationships the parameters of Table 2 are calculated. 72 K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) When diagnosing turbine regulatory systems, the ES also makes use of the dynamic response of the actuators – servomotors and slide valves. An ES consists of three main interconnected components: a knowledge base, an output mechanism and a database that provides diagnostic solutions, similar to how experts do it. The database is designed to store initial and intermediate data. It contains all the necessary data for diagnosis, obtained by measurements and observations. The knowledge base in the ES is intended to store long-term data describing the area under consideration and the rules that, when applied to the initial data, lead to the solution of the problem. Expert knowledge is presented in the knowledge base in the form of rules [11, 13]. The output mechanism interprets the rules and uses the facts of the knowledge base to solve the problems posed. It makes a diagnosis based on the information contained in the database. Available ES can be divided into three types: logical ES, cause–effect ES and ‘intelligent’ ones. Logical models are presented as tables of faults, which list the signs of possible faults (usually the parameters that exceed beyond the set point). In a broader sense, the emphasis is only set on the principle of tolerance control of parameters, while the description of malfunc- tions can be quite complicated. Cause–effect ES (PSM) are models that reflect the interrelations between the processes and conditions of the diagnostic object in case of malfunction occurrence and development (or the qualitative relationship between the malfunction causes and their consequences). PSM can have a hierarchical structure, which shows at each level in a graphical form the relation- ship of various faults leading to more serious malfunctions. Table 1: Parametric relationships generated in the diagnosis of the turbine automatic regula- tory (TAR) system. № Name 1 Turbine steam flow rate – setting of high-pressure part servomotor 2 Turbine steam flow rate – turbine capacity 3 Settings of high (medium) pressure control valves – settings of high (medium) pressure part servomotor 4 Steam pressure beyond the valves of high (medium) pressure part – settings of high (medium) pressure part servomotor 5 Force of high (medium, low) pressure part servomotor – settings of high (medium, low) pressure part servomotor 7 Force of medium (low) pressure part servomotor – steam pressure in a chamber of process or heating steam extraction 8 Steam pressure in control stage (or first stage) chamber – settings of high-pressure part servomotor 9 Steam pressure in control stage (or first stage) chamber – turbine steam flow rate 10 Settings of steam stop valve autoactuators – oil pressure above the autoactuators’ slide valves 11 Settings of steam stop valve autoactuators – oil pressure below the autoactuators’ slide valves K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) 73 A characteristic feature of ‘intelligent’ ES is, first of all, the form of diagnostic knowledge presentation. As for the fundamental capabilities of these systems, they do not differ signifi- cantly from the capabilities of ES of other types, with the exception of fuzzy logic [14]. The latter approach gives great opportunities in the analysis of complex systems because it allows one to operate with subjective assessments of an expert or user and to employ incomplete or inaccurate information. The program shell of a probabilistic ES includes a universal information processing pro- gram based on the Bayes’ formula taking into account the ‘price’ of each piece of evidence according to the Naylor method [11]. The specific filling of the knowledge base, that is, the formation of its content and the designation of a priori probabilities for the hypotheses and the ‘price’ of evidence, is carried out, as a rule, by the method of expert evaluations employ- ing the specialists dealing with thermo-mechanical equipment of thermal power plants, thus taking into account the specific features of the equipment at a particular station. To diagnose steam turbine subsystem parts, the ES employs various approaches, such as: • for turbine flow part – a correlation and regression analysis of the multi-factor relationship between vibration and regime parameters; • for thermal expansion system – the evaluation of forces acting on longitudinal keys under different temperatures of the left and right sides of turbine cylinder; • for condensing unit – the estimation of the effect of cooling surface fouling as well as of air content in condenser steam chamber on condenser efficiency, optimization of con- denser cleaning period, justification of periods of condenser tube system replacement, etc. Table 2: Main features for diagnostics and adjustment of the TAR system. № Feature name Electrohydraulic system of regulation and protection characteristic, determined from the recorded dependencies 1 Static characteristic of rotation speed control (RS) Degree of frequency response variations (FR) Degree of frequency regulation insensitivity Areas and values of the local non-uniformity of frequency regulation 2 Cam-operated steam distribu- tion device (CSD) and nozzle unit performances CSD technical condition Quality of the control valve setting adjustment Nozzle unit condition Flow part condition (salt fouling) 3 Force margins of regulatory system servomotors Identification of unstable operation areas of auto- matic regulatory system Regulation units lock detection Turbine regime optimization 4 Performances of steam stop valve autoactuators Technical condition of steam stop (safety) valve autoactuators Insensitivity of steam stop (safety) valve autoactua- tors 74 K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) The algorithms of the data analysis for defects causing the vibration are presented below. The indications of these defects are divided as follows: • boundary group where the defect indications are determined by a measured parameter drop outside the normalized limits; • factorial group where the defect indications are determined by an occurrence of a previ- ously unobserved factor; • correlation group where the defect indications are determined by a connection between the vibration and technological parameters. Boundary indications are determined by a measured parameter drop outside the permissible limits that is beyond the zone of partial or full serviceability. As a rule, the boundaries of these zones are designated in the regulations. Factor indications are characterized by a qualitative change in vibration parameters, for example, by the rise of rotational component of vibration in vertical or transverse direction or vibration components with frequency of 2w, 3w, 4w and so on, by abrupt increase in high-fre- quency harmonics, by the emergence of new frequencies in the turbine unit vibration spectrum: frequencies from 500 to 2000 hz point to leakage in regulatory system, whereas frequencies from 1000 to 1050 hz point to backlash in control valves. For the correlation indications, a change is estimated in the coefficient of correlation between the vibration characteristics and technological process parameters. 3 APPLICATION The ES functioning can be described by taking a steam turbine condensing unit as an exam- ple [15, 16]. A knowledge base for the condensing unit contains more than 30 hypotheses and 25 pieces of evidence (or indications); estimation procedures for 20 parameters of state are also specified. Tables 3 and 4 show a sample from the knowledge base. A preliminary list of malfunction hypotheses is set up on the basis of performed inves- tigations, statistical processing of data on equipment damage and literature data. After expert examination, the final list of hypotheses is filled in the knowledge base. The evi- dence list is being set up during the condenser unit operation. The measurement circuit, the results of the condenser unit tests, the operation regimes and maintenance logs are analyzed. Then a priori probabilities of the hypotheses and evidence probabilities are entered in the table of probabilities. In addition, evidence probabilities are entered to detect the fault (to confirm the hypothesis) and not to detect the fault (reject the hypothesis). These probabilities are specified for each evidence. In the first case, the evidence probability is denoted by the superscript (+), and in the second case, by the superscript (–). Table 5 shows an example of probability table for four hypotheses and three evidences. Malfunction evidence listed in Table 5: 1. high water heating; 2. Temperature difference between steam and cooling water outlet exceeds the norm; 3. Condenser pressure exceeds the norm. For hypotheses, the values of the maximal and minimal probability are also evaluated. This permits one to form a justified diagnosis for limiting values of probabilities. K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) 75 Table 3: hypotheses of condenser unit (CU) equipment malfunction. CU equipment Malfunction hypotheses Condenser Tube plates fouling Overpressure in the drain pipe Deterioration of siphon rarefaction Cooling surface fouling on the steam side Cooling surface fouling on the water side Elevated quantity of induced air Incomplete opening of the drain valve Elevated hydraulic resistance in the pressure line Condensate flooding on lower tube rows Level regulator fault Cooling water suction in the steam space Air suction between the condenser and condensate removal pump Water leakage from water ejector into condenser Improper organization of various streams discharge into condenser Steam jet ejector Steam grates or working nozzle clogging Inadequate flow rate of the full-flow condensate entering the ejector cooler Cooler heat transfer surface fouling on the water side Cooler heat transfer surface fouling on the steam side Steam–air mixture recirculation through one of the ejector stages Leaks in the partitions separating the coolers Airmeter or the exhaust pipe clogging high temperature of the full-flow condensate Dead air zones occurence in the drain pipe Leakage in ejector cooler Reduced heat exchange surface of the ejector cooler Circulation path (geodesic height) Changes in the hydraulic regime of water reservoir Skid of coarse gratings with aquatic vegetation and debris Significant fouling of rotating grates due to late cleaning or to wash- ing device malfunction Accumulation of air released by heating water Incomplete opening of the drain valve Water level lowering in the admission chamber 76 K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) During the ES function, it analyzes the equipment operation parameters and then, if there is lack of information, it asks the user a series of questions to make the information more precise. Figure 1 shows the form for evidence processing. The users click the buttons ‘Yes’ or ‘No’ and so indicate the degree of confidence in their answer to the question. The program corrects the a posteriori probabilities of the hypotheses taking into consideration the degree of Table 4: Evidence of CU equipment malfunction. CU equipment Malfunction evidence Condenser high water heating high steam pressure at the condenser inlet Low flow rate of cooling water high cooling water pressure at the condenser inlet Increased temperature difference between steam and cooling water outlet Oxygen presence in the full-flow condensate Condensate overcooling (tc – ts) < 0 high pressure in the cooling water drain pipes Low cooling water pressure at the condenser inlet Low pressure in the pressure line of the circulation pump high condensate hardness Exhaust steam pressure is above the standard (low vacuum) high hydraulic resistance of the condeser high condensate level in the condenser Steam jet ejector Pressure pulsations of steam–air mixture at the ejector inlet and discharge Low pressure in the pipeline upstream of the ejector high pressure of the working steam upstream of the ejector Increased working steam flow rate The sharp increase of the pressure in the suction chamber of the ejector when the exhaust air flow is within the range cor- responding to the design part of ejector performance A number of ejector cooler tubes are gagged high back pressure beyond the last stage of the ejector Water ejection from the exhaust Inlet pressure pulsations at the second and third ejector stages high pressure at the ejector suction Circulation path Unsatisfactory performance of the drain water siphon K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) 77 confidence in the user answers. The number of such questions corresponds to the number of evidence in the database of the ES leaving the evidence already processed out. 4 EXAMPLE As an example, it is here suggested to consider the diagnostics of ejector malfunction. It is known that ejector is an auxiliary equipment with a very high influence on the turbine. Ejec- tor diagnostics is performed upon request. Most parameters (signs) are defined manually in case of usual lack of measurements. The other point is a difficulty of automation; some parameters such as knocking, cooler overflowing and others are qualitative, not quantitative. The ejector state is defined in the operation process. For better diagnostics, there is need to provide ejector tests. Also some of the defects could be found only when dissembling the ejector. The initial data for this task are presented in Table 6. In the ES, the defects (malfunctions) that can occur in the ejector, as well as a priori prob- abilities of these defects (P(h)), are entered as listed in Table 7. Table 5: A sample of the probability table. № Hypothesis A priori proba- bilities Evidence probability for detection (non-detec- tion) of failure (hypothesis) 1+ 1– 2+ 2– 3+ 3– 1 Condenser tube plates fouling 0.5 0.8 0.2 0.005 0.005 0.5 0.005 2 Overpressure in the drain pipe 0.5 0.8 0.2 0.05 0.05 0.5 0.05 3 Deterioration of siphon rarefaction 0.6 0.1 0.05 0.05 0.1 0.05 0.2 4 Elevated quantity of induced air 0.7 0.05 0.05 0.7 0.05 0.7 0.05 Figure 1: Evidence processing. 78 K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) Each malfunction of the ejector has a list of signs (evidence) that correspond to it; this list is given in Table 8. As an example, the probabilities of evidence (malfunction signs) in the presence of a spe- cific defect (P (E: h)) and in the absence of the defect (P (E: not h)) are presented in Tables 9 and 10, respectively. 5 CONCLUSIONS An ES of probability type is presented for diagnostics and state estimation of components of steam turbine technological subsystems. The knowledge base is made up for rotors, bearings, turbine automatic control and protection system and for other components of the turbine unit, condensing unit equipment and other technological subsystems. The ES permits one to diagnose the condition of various subsystems and components of the turbine unit, to troubleshoot the equipment components and to formulate recommenda- tions about the ways and terms of defect elimination and of reducing the risk of their Table 6: Initial data for diagnostics. № Name Designation Units Notes 1 Condensate temperature at condensate pump input tc °С 2 Condensate temperature at ejector output t2c °С Manual input 3 Ejector motivating steam pressure Рm kg/cm 2 Manual input (indicator) for ejectors A and B 4 Dry air ejector flow rate Ga kg/h Determined by the ejector test results Table 7: List of defects. № Defect name A priori proba- bility, P(H) 1 Low motivating steam consumption 0.1 2 Inadequate flow rate of the main condensate 0.15 3 high temperature of the main condensate 0.15 4 Reduced heat transfer surface (high tube contamination or a large number of tubes plugged) 0.05 5 Leakage of partitions in the steam space 0.15 6 Tube system leaks 0.15 7 Drainage clogged 0.25 8 Flow path wear 0.1 K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) 79 development. To do this, the system employs the experience of experts. Information from the ES can be used to adjust the turbine operation regimes and to optimize the amount and timing of equipment repair. Table 8: List of signs. № Evidence (sign) name Designation Evidence evaluation algorithm 1 high heating of the main condensate in the ejector Δtej t2c – tc > 5°C 2 Low motivating steam pres- sure in the ejector Рmf Рm < 13 3 Water hammering in ejector housing h — 4 Ejector steaming Est A steam–air mixture column is visible from the ejector exhaust 5 Ejector capacity is low Gaj Ga < 80 kg/h 6 Ejector stage flooding Z When tapping the ejector housing, a dull sound is heard (the housing is filled with main condensate) 7 high pressure at the inlet of the ejector first stage at zero air flow Pas Pas0 > 1.5 kPа Table 9: Probabilities of evidence (malfunction signs) in the presence of a specific defect (Р(E:h)). № of malfunction hypothesis 1 2 3 4 5 6 7 8№ of evidence 1 — 0.9 — — 0.1 0.05 — — 2 0.9 — — — — — 0.1 — … … … … … … … … … 7 0.9 0.5 0.9 0.1 — — — 0.95 Table 10: Probabilities of evidence (malfunction signs) in the absence of a specific defect (Р(E:not h)). № of evidence 1 2 3 4 5 6 7 8 1 — 0.077 — — 0.22 0.23 — — 2 0.12 — — 0.00 0.00 0.00 0.23 — … … … … … … … … … 7 0.46 0.50 0.43 0.60 — — — 0.39 80 K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) NOMENCLATURE CSD – cam-operated steam distribution device CU – condensing unit ES – expert system STU – steam turbine unit AV – adjusting valve TAR – turbine automatic regulatory system RS – rotation speed FR – frequency response tc, ts – condensate temperature, saturation temperature Рm – pressure of primary steam Ga – «dry» air flow rate Δtej – heating of the main condensate Gaj – ejector capacity with ‘dry’ air Pas – pressure at the inlet of the ejector first stage at zero air flow h – water hammering in ejector housing Est – ejector steaming Z – ejector stage flooding REFERENCES [1] Mikhailov, V.E., Khomenok, L.А., Sudakov, A.V. & Obukhov, S.G., On complex di- agnostics and examination of the state of equipment of thermal power plants and hydro- power plants. Reliability and Safety of Energy, 2(9), pp. 9–14, 2010. [2] Andryushin, A.V., Polushkina, E.N. & Shnyrov, E.Yu., Development of the mainte- nance service system in TGK and OGK after the completion of industry restructuring processes. Thermal Engineering, 1, pp. 69–73, 2010. [3] Aronson, K.E., Brodov, Yu.M. & Novoselov, V.B., Development of a technical condi- tion monitoring system for a cogenerating steam turbine equipment. Thermal Engineer- ing, 12, pp. 65–68, 2012. [4] GOST 20911-89. Technical Diagnostics: Basic Terms and Definitions, Publishing house of Standards: M., 14 p., 1990. [5] Uryev, E.V., Fundamentals of Reliability and Technical Diagnostics of Turbomachines, USTU: Ekaterinburg, 71 p., 1996. [6] hayet, S.I., Aronson, K.E., Brodov, Yu.M. & Shempelev, A.G., Development and test- ing of system elements for status monitoring and diagnostics of a steam turbine con- denser. Thermal Engineering, 7, pp. 67–69, 2003. [7] Kovalev, N.A., Algorithms development for the operation and recognition of defects for automatic system of vibration diagnostics. CKTI Proceedings, 19, pp. 27–33. [8] Mirzabekov, A.M., haimov, V.A. & Khrabrov, V.P., Optical diagnostic system for erosion damage of the input edges of steam turbine blades. Thermal Engineering, 4, pp. 52–56, 1991. [9] Panov, E.V. (ed.), Artificial Intellect: Directory, Radio & Communication: М., 1, 461 pp., 1990. [10] Bashlykov, А.А., Expert system architecture to back up decision-making processes in fault diagnosing of thermal power station heat exchanging equipment (SPRINT). Col- lected volume: An Expansion of Intellectual Abilities of ACS, ed. А.А. Bashlykov, Ener- goatomizdat: М., pp. 5–8, 1989. K. E. Aronson, et al., Int. J. of Energy Prod. & Mgmt., Vol. 5, No. 1 (2020) 81 [11] Naylor, C., Build your own Expert System, Energoatomizdat: М., 286 pp., 1991. [12] Brooking, А., Johns, P., Forsyth, R. & Cox, F., In: Expert Systems: Principles and Case Studies, ed. R. Forsyth. Radio & Communication: М., 191 pp., 1987. [13] Jackson, P., An Introduction to Expert Systems, Williams Publishers: М., 624 pp., 2001. [14] Perminov, I.A., Orlick, V.G. & Gordinsky, A.A., Flow path state diagnostics for steam turbines of high capacity with the use of power plant computational complexes. CKTI Transactions, 273, pp. 58–61, 1992. [15] Brodov, Yu.M., Aronson, K.E. & Nierenstein, M.A., The concept of diagnostics system for condensing steam turbine unit. Thermal Engineering, 7, pp. 34–38, 1997. [16] Khaet, S.I., Aronson, K.E., Brodov, Yu, M. & Shempelev, A.G., Development and test- ing of the monitoring system elements for condition control and diagnostic of steam turbine condenser. Thermal Engineering, 7, pp. 67–69, 2003. 
work_3ahleve2qbbjjgsmjg73bbl33a	----	Chapter 1: Introduction Page 1 of 27 Performance of supply chain collaboration - A simulation study Abstract In the past few decades several supply chain management initiatives such as Vendor Managed Inventory, Continuous Replenishment and Collaborative Planning Forecasting and Replenishment (CPFR) have been proposed in literature to improve the performance of supply chains. But, identifying the benefits of collaboration is still a big challenge for many supply chains. Confusion around the optimum number of partners, investment in collaboration and duration of partnership are some of the barriers of healthy collaborative arrangements. To evolve competitive supply chain collaboration (SCC), all SC processes need to be assessed from time to time for evaluating the performance. In a growing field, performance measurement is highly indispensable in order to make continuous improvement; in a new field, it is equally important to check the performance to test conduciveness of SCC. In this research, collaborative performance measurement will act as a testing tool to identify conducive environment to collaborate, by the way of pinpointing areas requiring improvements before initializing collaboration. We use actual industrial data and simulation to help managerial decision-making on the number of collaborating partners, the level of investments and the involvement in supply chain processes. This approach will help the supply chains to obtain maximum benefit of collaborative relationships. The use of simulation for understanding the performance of SCC is relatively a new approach and this can be used by companies that are interested in collaboration without having to invest a huge sum of money in establishing the actual collaboration. Key words: supply chain collaboration, simulation, performance measurement, CPFR 1. Introduction Supply chain management (SCM) organizes and manages the whole process of activities of supply network from suppliers through manufacturers, retailers/wholesales till end users (Christopher, 1998). Traditionally, supply chain (SC) was designed with more focus on movement of materials rather than information flow. Due to ever increasing competition in businesses, many SCs have taken some twists from traditional way of functioning, from time to time, to adapt to the situation. Existing literature describes the SCM of the 21st century as an integrative value adding process of planning and controlling of materials and information between the supplier and the end user in order to increase customer satisfaction by reduced cost and improved services (Cooper et al., 1997). In today’s competitive unpredicted business world, cost reduction and good customer services are not stand-alone effort of any single SC member. As success of any product lies in customers' response to that product, it is important for businesses to achieve customer satisfaction by having efficient and Page 2 of 27 effective SCs. This may be possible through collaboration among SC partners. Hence, it is important to coordinate SC activities to streamline planning, production and replenishment (Ramanathan, 2012a). Market demand and changing nature of end-users can create more opportunities for SC players. At the same time, to be viable in a competitive market, all SC members need to be innovative and productive (Lee, 2002). As operating alone in a tight competition seem to be no longer beneficial for SCs, the importance of partnership has been adopted in various stages of many SCs (Samros, 2007). In the past, several SCM practices such as Vendor Managed Inventory (VMI), Efficient Consumer Response (ECR), Continuous Replenishment (CR), and Electronic Data Interchange (EDI) have been suggested in the literature to increase benefits of SCs. VMI technique was developed in the mid 1980’s, in which customer’s inventory policy and replenishment process were managed by the manufacturer or supplier. However, SC visibility was not predominately powerful in VMI to avoid bullwhip effect (Barratt and Oliveira, 2001). Forecast driven VMI and integration of CR with EDI was used to reduce the information distortion in VMI. ECR developed in 1992, was based on the concept of value adding by all partners in the supply chain. Both VMI and EDI together with ECR tried to create more responsive supply chain with broader visibility of information across the whole SC. Ever increasing SC demands have led to the invention of Collaborative Planning Forecasting and Replenishment (CPFR), another supply chain management tool incorporating planning, forecasting and replenishment under a single framework (Fliedner, 2003). CPFR, a second generation ECR (Seifert, 2003) aims to be responsive to consumer demand. It was introduced as a pilot project between Wal-Mart and Warner-Lambert in mid- nineties. According to VICS (2002), CPFR is a new collaborative business perspective that combines the intelligence of multiple trading partners in the planning and fulfilment of customers demand by linking sales and marketing best practices. Collaboration among SC members is a topic of interest for many researchers and practitioners (Barratt and Oliveira, 2001; Danese, 2007; Nyaga et al., 2011; Ramanathan, 2012a). Simatupang and Sridharan (2004) evolved four profiles for supply chain collaboration (SCC), namely efficient, synergistic, underrating and prospective collaboration. They proposed decision synchronization, incentive alignment and information sharing as three performance indices. In an attempt to maximize benefits of SCs, all SC members share information (data sharing) and collectively forecast the demand for products to have effective replenishment process (Aviv, 2007; Gavirneni et al, 1999). SCC activities help to improve the performance of involved members in a structured framework with the aim of maximizing profit through improved logistical services (Stank et al., 2001). However, majority of the articles in the literature have not highlighted important factors of good SCC practice. In this paper, we will be analysing the environments conducive to initiate SCC such as CPFR. The focus of this research is to identify the Page 3 of 27 suitable environments to collaborate in SCs. Revealing the actual benefits of SC collaboration with certain number of partners with specific level of investments for a specified period will help to make decision on implementing SCC at various levels. This is further explained through evidence from the existing literature in the next section. The rest of the paper is organised as follows: Section 2 will briefly explain the existing literature on SCC. Section 3 will describe research methodology used in this research. Section 4 explains the development of performance measurement of supply chain collaboration. Section 5 will discuss the results and analysis of simulation. Finally, Section 6 will conclude the paper with key findings, managerial implications, limitations and future work. 2. Supply chain collaboration for performance improvement: A Literature review SCM is being practiced by many businesses around the globe and hence it has a great wealth of literature from time of evolution of business processes. But, SCC is a relatively new research area and the literature is growing at a tremendous pace. Various advantages and disadvantages have been revealed by academics and practitioners. This section discusses some of the advantages and barriers of SCC. On realizing the importance of collaborative efforts in SCs, many researchers have developed theoretical and mathematical models to improve the structure and functionality of SCs. 2.1. Advantages of SC collaboration In the field of SCM, there is an overlap in the meaning of cooperation, coordination, collaboration, joint action plan and partnership, representing more or less the same concept (Yu et al., 2001; Corsten and Felde 2005). However, CPFR is specifically defined as a web-based attempt (Fliedner, 2003) or internet tool to coordinate the various supply chain activities such as forecasting, production and purchasing in SCs to improve the visibility of consumer demand (Barratt and Oliveira, 2001), to reduce any variance between supply and demand (Steermann, 2003). Caridi et al. (2005) viewed CPFR as a process of correcting, adjusting, proposing prices and quantities to reach an agreement on common unique forecast that can be used by buyers and sellers. VICS (2002) claimed that CPFR would help cost savings and gain competitive advantage. Several case studies have been reported in literatures that have examined the impact of collaboration (see www.ecch.com and ECR Europe, 2002). In SCCs, through joint planning and decision making, the understanding of the replenishment process is becoming clearer (Barratt and Oliveira, 2001). For example, Wal-Mart’s initiative of creating profile on purchase pattern of customers, namely ‘personality traits’, has helped to increase visibility of demand throughout the value chain (Mclvor et al., 2003). Information exchange and demand forecast http://www.ecch.com/ Page 4 of 27 based on sales data helped ‘Sport Obermeyer’ to improve forecast accuracy during demand uncertainty (Fisher 1997). In recent years, many academics and practitioners have suggested using collaborative arrangement to improve SC performance. Ramanathan and Muyldermans (2011) used structural equation models to identify underlying demand factors of soft drink sales in collaborative supply chains. They suggested using those factors for demand forecasting. Cheung et al. (2012) used actionable quantitative information from a number of upstream and downstream partners in developing knowledge-based system in supply chains. They have used simulation experiments to test SC models. Ramanathan and Gunasekaran (2013), Nyaga et al., (2011) and several other researchers insisted the importance of transparent information sharing, joint efforts and investments to improve trust and commitments in SCCs. Any SC can improve visibility using five important factors namely responsiveness, planning, shared targets, trust and common forecast (Barratt and Oliveira 2001). Real benefit of information sharing among SC partners lies in its effective and efficient use (Lee et al., 2000; Raghunathan, 2001) and it is also supported by proper use of Information Technology (IT) (Sanders and Premus, 2005; Cachon and Fisher, 2000). From the cases of Wal-mart and P&G, it is understandable that the use of various IT platforms is based on the scale of operations. 2.2. Barriers of SC collaboration Barriers of SC collaboration can be broadly classified under two categories: organisational and operational. Smaros (2007) argued that lack of internal integration (organisational barrier) would be a great obstacle for manufacturers to efficiently use demand and forecast information (operational barrier). Sometimes behavioural issues within organisation may also lead to failure of collaborative relationships. Fliedner (2003) considered lack of trust, lack of internal forecast, and fear of collusion as three main obstacles to implement collaboration. Boddy et al. (1998) identified six underlying barriers for partnering: insufficient focus on the long term, improper definition of cost and benefit, over reliance on relations, conflicts on priority, underestimating the scale of change and turbulence surrounding partnering. Use of technology and levels of information exchange in SCs have been discussed in the literature as both the advantage and the disadvantage (Cadilhon and Fearne, 2005; Sanders and Premus 2005; Samros, 2007). Occasionally, even a basic level of information exchange will yield potential benefits to businesses. For example, Metro Cash & Carry Vietnam is a German-owned business to business grocery wholesaler successfully engaged in collaboration with a disarming degree of simplicity. The company shares information among SC partners using telephone calls and fax machine without much sophisticated IT (Cadilhon and Fearne, 2005). The case of Metro Cash & Carry clarifies that free access Page 5 of 27 to available data is imminent in SCCs for planning and forecasting. But technology may not be a barrier for the success of collaboration (Cadilhon and Fearne, 2005; Smaros, 2007). This argument on technology totally disagrees with the basic concept of CPFR, which is a web-based attempt to coordinate the various activities among supply chain partners (Fliedner, 2003). Though information sharing and the role of IT were commonly accepted as significant phenomena in SCC (Sanders and Premus 2005), the use of technology is not argued widely as a necessary condition for collaboration; this is mainly because the technology used in CPFR varies widely across different CPFR cases (Danese, 2007). Also, due to availability of wider variety of technology and tools, proper technology selection becomes a complicated task for collaborating partners. To handle this issue, Caridi et al., (2005) proposed a new ‘learning model’ to incorporate intelligent agents to CPFR to measure performance of SCs at different collaborative environments. Barriers of partnering could be avoided through supplier training programme (Smith, 2006) and identifying opportunities to increase scope (Boddy et al., 1998). Continuous efforts of academics and practitioners to improve SCC have helped creating many models of SCs. 2.3. Models in SC collaboration In general, the nature of complexity is instrumental in the development of models at various levels of SCCs. Also due to increase in SC dependencies, SCC requires different combination of tasks and resources (Simatupang and Sridharan, 2004). For instance, CPFR business model is based on experiences of practitioners and strategies of their business development process (Ireland and Crum, 2005). Though, the basic structure of CPFR model has been accepted by many practitioners, it is also commonly agreed by many that some value addition to the existing model, depending on the industry implementing CPFR, will make SCCs responsive to market changes (Smith 2006; Chung and Leung 2005). Theoretical model developed by Corsten and Felde (2005) is related to the impact of trust (Humphreys et al., 2001), dependence, supplier collaboration on innovation, purchase cost reduction and financial performance. They established that supplier collaboration and the level of trust have positive impact on innovation and success of SCs. In literature, many conceptual frameworks are designed to explain the organizational and functional aspects of SCC whereas mathematical or simulation models are focussing mainly on the performance evaluation. Examples of SC models, suggested in the literature after the development of CPFR framework (mid-nineties), are given in Table 1. Aviv (2001) compared the effect of collaboration in two different set-up: one with centralized information and another with decentralized information. Based on uncertainty measure he concluded that diversified forecasting capabilities can improve the benefits of collaborative forecasting; in other words forecasting accuracy is strongly dependent on the collaborative strength. Page 6 of 27 Lee et al., (2000) developed a model to verify value of demand information sharing especially when demands are correlated significantly over a period of time. In a counter argument, Ragunathan (2001) emphasised the importance on effective use of available internal information for forecasting in comparison to investing on inter-organizational information system for information sharing in the case of non-stationary demand. Only a few studies exist in the literature on the performance analysis of SCC using simulation. Kim and Oh (2005) used system dynamics model to identify the performance of collaborative SCs in three different scenarios: manufacturer dominated SCs, supplier dominated SCs and balanced decision making. The authors identified that the balanced SCC will yield high benefits. Angerhofer and Angelides (2006) created a system dynamics model to evaluate the performance of supply chain management. The impact of six constituents - stakeholders, topology, levels of collaboration, enabling technology, business strategy and processes, were tested on SCs to measure the performance. Chang, et al., (2007) introduced an idea of augmented CPFR (A-CPFR) as an improvement to existing CPFR model with access to market information through application service provider. The authors tested its forecast accuracy through a simulation model. In a recent paper, Ramanathan (2012a) used AHP model to compare performance of two companies based on use of SC information. The author concluded that the companies using frequent information exchange among SCs can be benefited with continuous improvement in planning and forecasting. Table 1: Some existing models in SCC Author Type of model Key concept Simulation models Cheung et al. (2012) Knowledge-based model The model helps to formulate long-term successful SC partnerships. Chan and Zhang (2011) Collaborative transportation management The model helps to identify the potential benefits of collaboration in transportation. Chang et al. (2007) Verification of forecast accuracy (Augmented CPFR), with application of service provider, will have access to market information and hence can improve forecast accuracy and achieve considerable reduction of inventory. Angerhofer and Angelides (2006) Performance measurement The model helps to identify the areas need improvement by measuring the performance of the supply chain Kim and Oh (2005) Performance measurement The model tests impact of different decision making process in collaborative supply chain performance. Fu and Piplani (2004) Evaluation of supply-side collaboration Supply-side collaboration can improve the distributor’s performance. Optimisation and mathematical models Sinha et al. (2011) Optimisation model The model helps to improve the performance of petroleum supply chain. Page 7 of 27 Author Type of model Key concept Aviv (2001) Mathematical model for forecasting Products with shorter lead time have more benefit from supply chain collaboration. Aviv (2007) Mathematical model for forecasting Dominance or power of partnership, agility of the supply chain and internal service rate affect the benefits of collaborative forecasting. Aviv (2002) Mathematical model for joint forecasting and replenishment Auto-regressive demand process can decrease the demand uncertainty in VMI and CFAR (Collaborative Forecasting and Replenishment) programmes. Chen and Chen, (2005) Mathematical model for joint replenishment Developed four decision making models to determine optimal inventory replenishment and production policies in a supply chain considering three-level inventory system in a two echelon supply chain; Model also included major and minor set-up cost for manufacturers, and major transportation and minor processing cost for the retailer. Raghunathan, (2001); Lee et al. (2000) Mathematical model Inventory reduction and cost reduction can be achieved with efficient use of information sharing (Lee et al, 2000) and there is no need to invest in inter- organizational systems for information sharing if order history is available (Raghunathan, 2001). Mishra and Shah (2009) Structural equation model New product development will benefit from collaborative effort of supplier and customer, and cross functional involvement. Nagya et al. (2011) Structural equation model Impact of collaborative efforts in overall satisfaction Ramanathan and Muyldermans (2010;2011) Structural equation model Impact of demand information in collaborative forecasting Ramanathan and Gunasekaran (2013) Structural equation model Impact of SC collaboration in success of long term partnership Ramanathan (2012a) AHP model Role of SC information in company’s decision making Other models Shafiei et al (2012) Multi-enterprise collaborative decision support system The model helps decision makers to explore various options of solutions under what-if scenarios. Singh and Power (2009) Structural Equation Model Firm performance will increase if both supplier and customer are involved in collaborative relationship. Kwon et al. (2007) Multi-agent model The model helps to provide flexible solutions to address SC uncertainties. Caridi et al. (2005) Multi-agent model Mutli-agent system can be used to automate and optimise supply chain collaboration. Chung and Leung (2005) An improvement to CPFR model Inclusion of ‘Engineering change management’ increases the responsiveness to market changes. Simatupang and Sridharan ( 2004) Collaborative performance system Collaborative enablers are directly linked with collaborative performance metrics. Four types of collaboration identified: Efficient, underrating, prospective and synergistic. Stank et al. (2001) Logistical service performance model Collaboration with external supply chain partners along with internal support will improve logistical services. McCarthy and Golicic (2002) Collaborative Increased revenues and earnings are possible with Page 8 of 27 Author Type of model Key concept forecasting model SCCs. Lambert and Pohlen (2001) Conceptual model Developed a framework with following seven steps: supply chain mapping, identifying value addition process, identifying the effect of relationship on profitability, realign supply chain processes accordingly, measure individual performance, compare value with supply chain objectives, replicate steps at each link in the supply chain 2.4. Performance measurement of SC collaborations Models in SC collaborations are mainly classified under two categories: performance measurement models and decision making frameworks. Some models are supported with mathematical/empirical evidence (Angerhofer and Angelides, 2006; Kim and Oh 2005; Forslund and Jonsson 2007), and other models are purely conceptual in nature (Chen and Paulraj, 2004; Simatupang and Sridharan, 2004). In general, these two types of models are interrelated to each other in their way of functioning with respect to cause and effect. For example, performance measurement will lead to decision making process and decisions will lead to improve future performance. The main purpose of measuring the performance of SC network is to identify the problems in order to improve the SC efficiency and also to identify the conduciveness of collaboration. Many researchers conducted a detailed study on performance measurement of SC network based on cost and service level (Lee and Padmanabhan, 1997). But in SCCs, communication technologies such as information exchange and proper use of data are of high importance to the success of collaboration (Danese, 2007). Hence, measuring the proper use of technology and information are also becoming important in SCCs. Some researchers developed theoretical frameworks to measure the performance using balanced score card with many performance perspective measures (Chen and Paulraj, 2004). But a very few researchers initiated benchmarking of SCs (Simatupang and Sridharan, 2004; Ramanathan el al., 2011). Evidences from the literature confirm that key measures for evaluating SC performance include cost, quality and responsiveness. In recent literature, forecast accuracy is also used as an indicator of proper use of information in SCCs (Ramanathan and Muyldermans, 2010). Meanwhile, lack of information exchange will result in greater variability of demand forecast for upstream SC members (Yu et al., 2001), which is the clear indication of SC problem. Chen and Paulraj (2004) tried to create a conceptual framework to understand problems and opportunities associated with SC management. As there are many dimensions for SCCs, the performance measurement is also becoming a complicated process. Verifying whether the environment is conducive to SCC will help the companies to Page 9 of 27 identify the areas to be modified before implementation. This was partly answered from the findings of Aviv (2001) and Smaros (2007). Aviv’s (2001) confirmed that the products with short lead time could achieve better forecast accuracy compared to the products with long lead time (Smaors, 2006). Danese (2007) through several case studies across SC networks such as manufacturers, customers and suppliers, identified that different levels of collaboration exist in SCs and the benefits attached to each level will differ. Based on the analysis of these case studies, Danese (2007) classified the degree of collaboration as low, medium or high. Ramanathan (2012a) compared two case companies performance on demand planning and forecasting and suggested three different levels of collaborations in SCs, namely preparatory level, progressive level and futuristic level. However, not many articles have discussed the benefits of SCC in terms of the number of partners, investment and duration of partnerships. Most of the studies discussed above have confirmed the role of supply chain partners and their involvement in SC performance and profit. However, there is no specific study that discusses in detail the role of investment, the number of partners or the duration in collaborative partnerships. To fill this gap, in this paper we use the combination of all these three elements in SC collaboration. In order to find environment conducive for SCC, based on the literature and the actual practices in SCs, we propose in this study that the degree of collaboration will depend on factors namely the investment on collaborating technology and partnerships, the number of collaborating partners and the duration of collaboration. We attempt to develop a performance model for SCC using a well-known methodology called simulation in the following sections. 3. Research methodology Performance evaluation of SCC is a complex task and research on this topic is still in its infancy. We make an attempt to quantify the benefits of SCCs through the factors discussed above. The choice of methodology is most important to identify the correct solution to a particular research problem (Yin, 1989). Case study based simulation is being used in this research. Case study research will be beneficial to understand the role of above specified five factors in performance of SCC. Basic information such as duration of collaboration, the level of investments and the number of partners from the case companies will be simulated to create similar scenario. For this purpose, we have chosen two case companies from the packaging industry. In this paper, we have used simulation to identify the performance of SC collaboration based on the factors of SCC. To initialize the process of simulation, basic mathematical approach is used as outlined above in Section 3. All the measures are converted in terms of ratio to avoid using mixed units. Generally, rhw ratio of input to output is described as a performance indicator. Simulation will support analysing collaborative performance on supply chains for changing degrees over the collaborating period. Page 10 of 27 This what-if analysis will be instrumental in decision making on implementation of collaborative supply chain (Angerhofer and Angelides, 2006). The advantage of using simulation is that an existing or proposed system can be designed using what if analysis in order to optimize the benefits by identifying the pitfalls in the system. Some researchers attempted to use system dynamics simulation for what-if analysis (Kim and Oh, 2005; Angerhofer and Angelides, 2006; Chang et al., 2007). In this research the purpose of what-if analysis is to identify the conducive environment to implement SCCs. Schematic projection (see Figure 2) of research methodology can further simplify the understanding of SC performance. This research intends to establish links among all the coordinating factors of collaboration. Creating links with different modules will in turn be powerful to identify a weaker node which needs improvement. Traditionally, performance of supply chain is measured through demand amplification (Angerhofer and Angelides, 2006) and value additions in each node of supply chain. But in case of collaborative SCs, the value addition is not an independent activity and hence composite performance indicator is used to measure performance of collaborative supply chain. If SC handles product returns then the performance should include inventory management and disposition of the returned goods. We have considered five important factors of SC collaboration for our further analysis; namely, degree of SC collaboration, business objectives – operational and financial, information sharing and SC processes. We have categorised the SC performance as financial and non-financial. Non-financial performance of SC is measured through operational business objectives, SC processes and information sharing (see Figure 2). Page 11 of 27 Figure 2: Schematic projection of methodology 4. Development of performance measures for supply chain collaboration Though SC is a widely researched area, it needs a strong framework (Chen and Paulraj, 2004) for development of more systematic principles that will help SCs to develop against all odds and barriers. In recent business world, many companies collaborate for different purposes such as logistics, cost reduction and business expansion. Such SCC necessitates some value addition to business objectives along with the original SC operations models (ECR Europe, 2002). Also information sharing is critical in modern SCs to meet fluctuating demand (Ramanathan, 2012a & b). In the literature, degree of collaboration is not linked with performance of SCs in an effective way (Danese, 2007; Larsen et al., 2003; Ramanathan, 2012a). In this research based on the literature and actual SC practices in recent businesses, we consider five important factors of collaboration namely business objectives - financial and operational, supply chain processes, information sharing and degree of collaboration. 4.1. Business Objectives – Financial (BOF) Now-a-days, many businesses are striving to maximize profit by improving the quality of products and services to the end users by lowering the cost. Many leading companies such as Wal-Mart and Procter & Gamble use SCC to achieve this objective. VICS (2002) claims that CPFR will help cost Define: Degree = f(N,L,T); here L = f(I) ; IS = f(FA) BOO = f(CU,LT); BOF = f(Rv, HC, SC); Pr = f(Ap,P,R,D) Define performance in terms of above defined metrics For Dg = ‘1 to x’ period Calculate ISDg, BOODg, BOFDg, PrDg Collaborative performance (non-financial) = BOODg + PrDg +ISDg Collaborative performance (financial) = BOFDg N - Number of collaborating partners L - Level of collaboration I – Investment on collaboration T – Time (duration) of collaboration IS - Information Sharing FA - Forecasting Accuracy BOO- Bus.Obj. Operational CU - Capacity Utilization LT- Lead Time Rv-Revenue HC – Holding Cost SC – Stock out Cost Pr – SC processes BOF – Bus Obj.Financial Ap-Adherence to plan Ad-Adherence to delivery plan P– No. produced R – No. returned D- No. Delivered (No. sold to wholesaler/Retailer) Analyse the performance at various degrees to identify the conduciveness Define variables N, L, T, I, FA, CU, LT, Rv, HC, SC, Ap, P, R, D Page 12 of 27 savings and gain competitive advantage. Commonly SC collaboration is initiated among various SC members to meet customers’ needs, to improve product availability, to increase business performance, to increase sales, to achieve reduced cost, to increase revenues and earnings, to improve forecast accuracy, to increase visibility of demand (McCarthy and Golicic, 2002; Cooke, 2002; Ireland and Crum, 2005; Ramanathan et al., 2012b). Cost savings such as minimizing the logistics cost can possibly be one of the most important drivers of collaborations (Corsten and Felde, 2005; Chen and Chen 2005). Chen and Chen (2005) developed a mathematical model for joint replenishment in the process of reducing cost. For example, Ace Hardware’s CPFR pilot project earned a positive result in forecast accuracy from 80 to 90 percent and product costs dropped from 7 to 2.5 percent (Cooke 2002). In many cases the SC collaboration proved to be a promising tool to increase business performance, sales, revenues and earnings (McCarthy and Golicic, 2002; Cooke, 2002). In our research, sales revenue and costs involved in production will be used to quantify financial business objectives. In general, cost involves fixed cost and variable costs such as production cost, stock out or holding cost. Other hidden variable costs are not included for the purpose of calculations.             T j BOF 0 cost Variablecost Fixed cost holdingor Stockout - cost) productionUnit produced (No. price salesUnit returns) No.ofsales of (No.           T j 0 jjjj cost Total OCPC]P[SP])R[(D Here D – No. delivered (i.e., sold to retailer) R- No. returned I – Current inventory SP- Selling Price PC- Production Cost P-No. Produced OC- Other Cost (Holding cost or stock-out cost) Variable cost = Production cost + Holding cost or stock-out cost          DI SC,R)I(D SC D I HC,D)-R(I HC OC HC- holding cost SC- stock-out cost Stock-out cost or penalty cost is usually calculated for retailers but not for manufacturers (Aviv, 2007). Based on our interview with the case companies, we assume that manufacturers will also incur Page 13 of 27 penalty cost for not completing production on time to facilitate on time delivery; this is similar to stock- out cost of retailer. 4.2. Business Objectives – Operational (BOO) Customer retention is becoming a great challenge in current competitive business market. Improved business performance through SCC can help to attract and retain customers (Matchette and Seikel, 2004). Customer loyalty can also be built by effective SC activities. For example, making stock of right products available at right time in proper location of retail stores will help to attract and retain customers. This can be achieved through a wider cooperation from all SC members. For instance, efficient capacity utilization can help reducing production time (Aviv, 2007). Customer loyalty can also be achieved if SC activities include customer service such as accepting and handling product returns (Dowlatshahi, 2000). From the literature, we have considered three important factors namely number of product returns, product lead time, and capacity utilization (production capacity) to measure the business objectives (Aviv, 2007; Dowlatshahi, 2000). Capacity utilization P μPμP CU n,nn,n 22 )( PPC )PC.(      where nnP , ≠ μ (Aviv, 2007) assumed capacity utilization as the product of cost of production and square of the difference between production batch size for period n and average production size) Assume if nnP , = μ , Capacity utilization is 100% PC—Product cost P-Number of items produced nn P , - Production batch size suggested for the next n periods at the beginning of period n. μ -Average production size Reduced Return rate, D R RR  1 R- Number of returns D – Number delivered Adherence to production plan will reflect in the reduction of product lead time or production time (Aviv, 2007). Adherence to production plan (AP): Page 14 of 27 jnn jn P PP AP   , , 1 Here, Production plan         T j Tj0 , ,1, ,  Tnn jnnjnn jn P PP PP n - Current period and Tn 1 ; jn PP , - production plan at period ‘n’ for period ‘j’ Hence, operational business objectives can be quantified as follows: BOO = %100 APRRCU =                  jnn jnn,n P PP D R P μP , , 2 11 )( 4.3. Supply chain processes Supply chain operations reference model (SCOR) classified processes as plan, source, make, deliver and return. Based on type of products and market value, length or degree of collaboration will differ (Ramanathan, 2012b). Products with long production cycle time takes more time to reach the market, while product with short production cycle time takes less time. Though collaboration in SCs can help to sell all products with variable lead time, products with shorter lead time have more benefit in SC collaboration (Aviv, 2001). In this research, we assume that the availability of raw material (source) is not difficult and accordingly, we consider four processes namely plan, produce, replenish and return. In SCCs planning stage will include forecasting as its integral part and hence forecasting is not treated as a separate process. SCs with activities of product returns need to check the inventory level and to arrange a proper disposition for the product returns (Dowlatshahi, 2000). In this case, performance of collaborative processes is a collective measure of cost function of adherence to plan and cost of inventory. Production plan         T j Tj0 , ,1, ,  Tnn jnnjnn jn P PP PP Adherence to plan cost function cost Production ).( )( 2 , 0 , jn T j jAP nAP PPC PPC    (based on Aviv, 2007). Product returns will increase the level of current inventory. In SCs with product returns, inventory holding cost can be quantified as follows: D I HC,D)-R(I HC  Here D – No. delivered (i.e., sold to retailer) R- No. returned Page 15 of 27 I – Current inventory HC- Unit holding Cost Performance of collaborative SC processes can be calculated as 4.4. Degree of collaboration Previous case study research by Danese (2007) identified different levels of collaboration such as basic communication, limited collaboration and full collaboration. Larsen et al., (2003) and ECR Europe (2002) categorized the depth and level of collaboration into three different forms such as basic collaboration, developing collaboration and advanced collaboration. Whereas, Simatupang and Sridharan, (2004) categorized the level of collaborative practices into low and high collaborations. In general it is agreed that various levels and practice of collaboration can yield benefits across the whole SC. In our research, degree of collaboration is measured in terms of number of collaborating partners (can be two echelon or multi-echelon SC), duration of collaboration and level of involvement. In this research, level of involvement is defined as the involvement of top management in terms of investment on technology and people in SCC activities. In every SCC, active participation of each SC partner can help to enhance the overall performance (Lambert and Pohlen, 2001). Cooke (2002) identified the need to change corporate culture as a pre-requirement of collaboration. Long-term SCCs can change attitude of workers. Normally, level of involvement of top management in SCC will be reflected in their investment on collaborating technology and training (Ramanathan et al., 2011). Based on the literature, we define the degree of SCC in terms of number of collaborating partners, total number of years, and investment on collaborative effort. tInvolvemen of Level businessin Duration years ingCollaborat memberschain supply ofNumber partners SCC ofNumber Degree  Here ‘level’ will be identified from the case company and percentage value will be assumed based on the collaborative operations (activities) in proportion to total activities. Level of Involvement =       yearper y technologand gon trainin investment Total yearper y) technologand ng(on traini investment iveCollaborat 4.5. Information sharing                 cost Variable cost Production ).( 0 2 , 0 , HCR-D)(IPPC Average T j jjn T j jAP Page 16 of 27 In recent competitive market, a great deal of business is relying on SC information and proper use of data. SCC can contribute to improve information sharing among SC partners (Yu et al, 2001). According to VICS (2002) accelerated information sharing among SC partners will increase the reliability of the order generation. Li and Wang (2007) asserted that the benefit of information sharing is depending on two factors: one is the context and the other is the proper use of information. Optimizing the supply chain will be possible through collaboration (Horvath, 2001) and information sharing (Horvath 2001, Yu et al. 2001). Information sharing among SC partners will help improve forecast accuracy and hence will help potential cost savings (Aviv, 2007;Byrne and Heavey, 2006). An exceptional level of service can also be achieved through integrated data and information (Kim 2006). Critical information sharing among SC partners varies widely depending on the industries involved (Smaros, 2007). Ovalle and Marquez (2003) summarized the types of information under three headings: product information, customer demand and transaction information, and inventory information. Yu et al. (2001) revealed that the centralized information sharing benefits manufacturers more than the retailers. Though information sharing and the role of IT were accepted as significant phenomenon in collaboration (Sanders and Premus 2005), the use of technology is not argued widely as a necessary condition for collaboration. This argument is evident from Smaros’ (2007) statement ‘collaboration technology is not a key obstacle for large scale collaborative forecasting’. In SCCs, product replenishment is a sub-process of forecasting (CPFR, 2002). Internal forecasting is the one which is generated by each collaborating partner based on the time series data and other exceptional factors (such as sales promotions) and market criteria. Collective forecasting is based on all the individual internal forecast figures which in turn facilitate order generation. Internal forecast accuracy will reflect in the collective forecast figure and help to reduce bullwhip effect (Aviv, 2001). In SCCs, the forecasting accuracy and forecast information quality can improve the profit proportion (Forslund and Jonsson, 2007). From the above literature, we understand that effectiveness of information sharing in SCC will be reflected in forecasting accuracy (FA) of product demands (Ramanathan and Muyldermans, 2010) and returns. Accordingly, we calculate FA as follows: Forecasting Accuracy (Sales) =               AD FDADabs )( 1 Forecasting Accuracy (Returns) =               AR FRARabs )( 1 Collectively, Forecasting Accuracy (FA) can be calculated as: Page 17 of 27 100 2 )()( 1 0                                     T j j jj j jj AR FRARabs AD FDADabs We assume that the demand follows the normal distribution. FD—Expected Demand; AD- Actual Demand; FR- Expected Returns; AR- Actual Returns; j- SCC period (here, 0 < j < 6) The underlying assumption is that the use of technology and information system helps to exchange real time information without any delay in information sharing. Hence, the point of sale data is available to manufacturer without any delay, i.e. accessibility is 100%. Standard Deviation (SD) of Forecast Error (FE) describes the spread of errors or uncertainty about an error which can be used for setting safety stock. Forecasting Error is calculated as follows: Absolute percentage of error (Sales) % = 100% AD FD)abs(AD        Absolute percentage of error (Returns) % = 100% AR FR)abs(AR        In this paper, we measure the degree of collaboration based on the level of involvement; the length of collaboration (period) and the number of partners. The impact of change in the degree of collaboration will be identified in forecast accuracy, business objectives and processes. The overall performance of SCC is calculated as the sum of individual performance in terms of BOO, BOF, forecasting accuracy, and processes at various degrees of collaboration. 5. Analysis and discussion of simulation results Improving overall performance, in terms of both quality and service, of SCs along with other business objectives such as maximising profit and minimising costs are the common underlying features of CPFR. But not many researchers have considered the impact of other underlying factors such as degree of collaboration, involvement of top management, information sharing, customer support, business objectives and SC processes. Magnitude of benefits on implementation of SCC often varies widely across different industries as substantial amount of investment and time are involved (Ramanathan et al., 2011). For example, products that are mainly manufactured to stock (such as detergents and shampoo) will have longer shelf life (Fisher, 1997) and hence SCs may not require high degree of collaboration. At the same time, fast moving technology products such as laptops and software need to be sold in a short span of Page 18 of 27 time in order to avoid obsolescence which requires higher degree of collaborative support from other SC members. For the purpose of this research we have contacted five different global companies from the packaging industry, who practice SCC. Three of them have collaboration with either upstream or downstream SC partners but not with both. Finally, we have considered two manufacturing companies who have been involved in collaboration for over six years with both upstream and downstream customers. SCC information selected for further analyses were mainly focussed on five factors as explained before. For each company, we have collected data of 10 collaborating partners and simulated the data using excel. Table 2 describes the sample data of one of the companies collaborating with different supply chain partners at various degrees. The first three columns of the table represent SC investment in collaboration (in US dollars), number of partners and length of collaboration. All the remaining columns have used the formula as described in Section 4. Table 2: Analysis of sample data Coll. investment Coll. partners Coll. period Degree BOO Information sharing Forecast accuracy SC processes 83500 3 3 0.01 0.02 0.77 65% 0.79 50000 10 3 0.02 0.99 0.92 96% 0.62 55000 4 4 0.04 0.89 0.95 91% 0.54 34000 10 4 0.01 0.00 0.91 96% 0.57 48500 11 4 0.01 0.03 0.87 91% 0.12 53500 7 5 0.04 0.01 0.91 91% 0.01 133000 8 5 0.05 0.01 0.93 87% 0.56 49000 4 5 0.02 0.99 0.76 69% 0.49 45000 3 6 0.01 0.90 0.79 65% 0.51 56000 6 6 0.02 0.00 0.97 95% 0.54 59000 12 7 0.02 0.99 0.93 99% 0.45 43590 7 7 0.03 0.01 0.94 97% 0.45 We have simulated 1000 instances of SCC based on the company’s data. The results indicate that the forecasts accuracy becomes stable over a period of time with the same number of collaborating partners. Figure 3 indicates the effect of the levels of collaboration on the performance of the company in terms of financial and non-financial objectives (SC processes and information sharing). SC partners collaborating for longer period of time have achieved increasing performance both financially and operationally. But it is not guaranteed that the company individual financial business objectives will be achieved consistently in case of high investments on collaboration. Also, the higher the number of collaborating partners does not mean proportionately the higher the level of performance. The Page 19 of 27 performance of the company shows a very slow but incremental effect against the level of collaboration in terms of number of partners (see Figure 3). Our interview with the case companies revealed that collaborating partners who are in the same business for a long term will bring success for all SC collaborating partners. This is possible mainly due to the sharing of knowledge and well established SC network. But “new members need to wait to reveal the actual benefit of collaboration. Huge investment in SCCs will not always help to reap the benefit quickly. Time is the key success factor in collaboration. Committed SC partners make our SCs really profitable and successful in terms of performance”. The results of the analysis suggest that companies do not need to investment on collaboration every year in order to yield high profit. Companies that believe in high investment on collaborative relationship without having effective SC operations will be difficult to survive in the competitive market. Even though new partnership is encouraged in competitive business scenario, it is vital for companies to continue the existing profitable partnership for a longer period of time to obtain consistent performance. Page 20 of 27 Figure 3: Effect of levels of collaboration on performance y = 28.97x + 155.67 -500 0 500 1000 0 0.2 0.4 0.6 0.8 1 1.2 B O F Level of collaboration - period Financial performance y = 0.1057x + 0.8232 0 1 2 3 0 0.5 1 1.5 B O O , S C p ro ce se s a n d in fo rm a ti o n s h a ri n g Level of collaboration - period Non-financial Performance y = 21.986x + 158.25 -500 0 500 1000 0 0.2 0.4 0.6 0.8 1 1.2 B O F Level of collaboration -investment Financial performance y = -0.244x + 0.9932 0 1 2 3 0 0.2 0.4 0.6 0.8 1 1.2 B O O , S C p ro ce se s a n d in fo rm a ti o n s h a ri n g Level of collaboration - investment Non-financial Performance y = 1.2282x + 165.7 -200 0 200 400 600 0 5 10 15 B O F Level of collaboration - partners Financial performance y = 0.1057x + 0.8232 0 1 2 3 0 0.2 0.4 0.6 0.8 1 1.2B O O , S C p ro ce se s a n d in fo rm a ti o n s h a ri n g Level of collaboration-partners Non-financial Performance Page 21 of 27 6. Managerial implications, conclusions and future research This paper addresses a recent relevant practical approach of SC collaboration in performance improvement. Understanding the important factors of SC collaboration and their impact on the potential benefits of SC can help the top management to understand the required degree of collaboration with upstream and downstream partners. One of the interesting managerial insights on fundamental principal of collaboration is that neither investment nor number of partners nor duration of collaboration, will independently contribute to improve the performance of SCs. This result helps to understand the importance of involvement of each SC partner. Increasing the number of partners in SCs will complicate the decision making and hence slow down the performance. However, human interactions in SCs can assist appropriate investment decisions in IT and collaborations to improve SC processes. Long-term collaborating partners can help yielding sustainable benefits to SCs. In general, the financial performance of a company is an indicator of success of operational performance. From the data analysis, we identified that the less involvement of top management in SC collaboration results in poorer overall performance. By measuring the performance, the top management of the company can decide whether to improve its investments in collaborative activities. Measuring the forecast accuracy can alert the managers the usefulness of available information and also can point out the need for accessible information and technology (Ramanathan and Muyldermans, 2010). Different supply chain partners collaborating for various purposes will have individual business objectives. Successful collaboration will help the businesses to be successful in achieving those set objectives. By measuring both financial and operational objectives any company can understand the current accomplishment of expected achievements. For example, in the given case company, the higher investment in collaboration has not shown more substantial benefit in terms of revenue. Hence, the company can try to improve other aspects of the current collaborative arrangement instead of investing further in the collaboration. On knowing the potential benefits of SC collaboration, SC partners can extend their partnership further to increase profit, to reduce lead time and to improve customers’ satisfaction. In this research, we have tested the SC collaboration with different levels of involvement and partnerships for certain period of time using simulation techniques. For different degrees of collaboration, the benefits of SC are found different. In real businesses, it is risky to experiment various degrees of collaboration as it can involve a huge amount of investment. Findings of this research suggest that the degree of collaboration should be revised on analysing the performance of the company (see Figure 4). The conduciveness of collaboration for any company depends on its flexibility in changing the degree of collaboration to achieve the business objectives. For example, if too many SC partners are involved in the collaboration, the partners with the highest Page 22 of 27 investment may have the power of dominance in planning and decision making; this may affect the smaller players in SCC arrangements (Aviv, 2007). In this case, the top management of the focal company can alter the degree of collaboration such as duration of collaboration, level of involvement and number of participating members to achieve required performance. Irrespective of the degree of collaboration, another performance measure namely, ‘forecast accuracy of the company’ will explicitly indicate the role of information exchange in the collaborative SC (Ramanathan and Muyldermans, 2010). Since the products with shorter lead time can normally benefit more from collaborative forecasting (Aviv, 2001), in this research we suggest extending the use of collaborative forecasting for products with medium or longer lead time. In poor forecast accuracy, top management can increase accessibility of information exchange. The company can also think of revamping the IT technology in order to improve the efficiency of information sharing. Achieving the predefined business objectives in terms of financial and operational activities will help the SC partners to sustain in the competitive business market. Performance measurement, in terms of financial and operational business objectives, indicates the conduciveness of the current SC collaboration. The collaborating company can adjust the degree of collaboration to match its business objectives. For example, SCC can be strengthened to increase profit by reducing the cost of operations. Similarly, SC collaboration can help reducing the product returns or help selling the returned products in secondary markets. Our research confirms that production lead time and capacity utilisation can also be improved with SC collaboration of suppliers’ for on-time delivery of raw materials for timely planned production (see Figure 4). To evolve efficient and effective competitive supply chain collaboration, all SC processes need to be assessed from time to time for evaluating the performance. In a growing field, performance measurement is highly indispensable in order to improve further. In a new field, it is equally important to monitor the performance to test the conduciveness. Our research has indicated the importance of identifying conducive environments for successful supply chain collaborations. We have based our simulation study using data from two companies from packaging industry. The same research can be extended further for different industries that have SC collaboration with many partners involving huge investment for long duration. This can help to draw a general conclusion on suggested level of investment and supply chain partnership, specific to each business sector. Page 23 of 27 Collaborative Supply Chain Processes Plan, Produce, Replenish and Return Overall performance Degree Information sharing &Forecasting Business Objectives (Financial) Duration Partners Level of Involvement Investment Technology People Top Management Involvement Business Objectives (Operational) Adjust the degree Increase frequency, accessibility & revamp technology In c re a se p ro fi t; re d u c e c o st Require more involvement Adhere to plan, control production and delivery time and frequent monitoring Reduce returns; reduce product lead time; improve capacity utilization Figure 4: Areas of improvement in SC collaboration Page 24 of 27 References Angerhofer, B.J. and Angelides, M.C. (2006). A model and a performance measurement system for collaborative supply chains. Decision Support Systems 42, 283-301. Aviv, Y. (2001). The effect of collaborative forecasting on supply chain performance. Management Science 47 (10), 1326-1343. Aviv, Y. (2002). Gaining benefits from joint forecasting and replenishment process: The Case of Auto-Correlated demand. Manufacturing & Service Operations Management 4 (1), 55-74. Aviv, Y. (2007). On the benefits of collaborative forecasting partnerships between retailers and manufacturers. Management Science 53 (5), 777-794. Barratt, M. and Oliveira, A. (2001). Exploring the experiences of collaborative planning initiatives. International Journal of Physical Distribution & Logistics Management 31 (4), 266- 289. Boddy, D., Cahill, C., Charles, M, Fraser-Kraus. M and MacBeth, D. (1998) Success and failure in implementing supply chain partnering: an empirical study. European Journal of Purchasing and Supply Management 2(2-3), 143-151. Byrne, P.J. and Heavey, C. (2006). The impact of information sharing and forecasting in capacitated industrial supply chains: A case study. International Journal of Production Economics 103 (1), 420-437. Cachon, G.P. and Fisher, M. (2000). Supply chain inventory management and the value of shared information. Management Science 46 (8), 1032-1048. Cadilhon, J. and Fearne, A. P. (2005). Lessons in Collaboration: A case study from Vietnam. Supply Chain Management Review 9 (4), 11-12. Caridi, M., Cigolini, R. and Marco, D.D. (2005). Improving supply-chain collaboration by linking intelligent agents to CPFR. International Journal of Production Research 43(20), 4191-4218. Chang, T., Fu, H., Lee, W., Lin, Y. and Hsueh, H. (2007). A study of an augmented CPFR model for the 3C retail industry. Supply chain management: An International Journal 12, 200-209. Chan, Felix.T.S. and Zhang, T. (2011). The impact of collaborative transportation management on supply chain performance: A simulation approach. Expert Systmes with Applications, 38 (3), 2319-2329. Chen, I.J. and Paulraj, A. (2004). Towards a theory of supply chain management: the constructs and measurements. Journal of Operations Management 22, 119-150. Chen, T. H. and Chen, J. M. (2005). Optimizing supply chain collaboration based on joint replenishment and channel coordination. Transportation Research Part E: Logistics and Transportation Review 41, 261-285. Cheung, C.F., Cheung, C.M. and Kwok, S.K. (2012). A knowledge-based customization system for supply chain integration. Expert Systems with Applications. 39 (4), 3906-3924. Christopher, M. (1998). Logistics and Supply Chain Management, second ed. Financial Times Press, Prentice Hall, Englewood Cliffs, NJ. Chung, W.C. and Leung, W.F. (2005). Collaborative planning, forecasting and replenishment: a case study in copper clad laminate industry. Production Planning & Control 16 (6), 563-574. Cooke, 2002; Page 25 of 27 Cooper, M.C., Lamber, P.M. and Pagh, J.D. (1997). Supply chain management: more than just a new name for logistics. International Journal of Logistics Management 8(1), 1-14. Corsten, D. and Felde, J. (2005). Exploring the performance effects of key supplier collaboration. International Journal of Physical Distribution & Logistics Management 35 (6), 445-461. Danese, P. (2007). Designing CPFR collaborations: insights from seven case studies. International Journal of Operations and Production Management 27 (2), 181-204. Dowlatshahi, S. (2000). Developing a theory of Reverse Logistics, Interfaces 30 (3), 143-155. ECR, Europe (2002). European CPFR Insights, ECR European facilitated by Accenture, Brussels. Fisher, M.L. (1997). What is the right supply chain for your product? Harvard Business Review, 75 (2), 105-116. Fliedner, G. (2003). CPFR: An emerging supply chain tool. Industrial Management + Data Systems 103 (1/2), 14-21. Forslund, H. and Jonsson, P. (2007). The impact of forecast information quality on supply chain performance. International Journal of Operations & Production Management 27 (1), 90-107. Fu, Y. and Piplani, R. (2004). Supply-side collaboration and its value in supply chains. European Journal of Operational Research 152, 281-288. Gavirneni, S., Kapuscinski, R. and Tayur, S. (1999). Value of information in capacitated supply chains. Management Science 45 (1), 16-24. Horvath, L. (2001). Collaboration: The key to value creation in supply chain management. Supply Chain Management 6 (5), 205. Humphreys, P. K., Shiu, W. K. and Chan, F. T. S. (2001). Collaborative buyer-supplier relationships in Hong Kong manufacturing firms. Supply Chain Management 6 (3/4), 152-162. Ireland, R.K. and Crum, C. (2005). Supply chain collaboration: How to implement CPFR and other best collaborative practices. J. Ross Publishing Inc.: Florida. Kim, B. and Oh, H. (2005). The impact of decision-making sharing between supplier and manufacturer on their collaboration performance. Supply Chain Management 10 (3/4), 223- 236. Kwon, O., Im, G.P. and Lee, K.C. (2007). MACE-SCM: A mult-agent and case based reasoning collaboration mechanism for supply chain managment under supply and demand uncertainties. Expert Systems with Applications, 33 (3), 690-705. Lambert, D.M. and Pohlen, T.L. (2001). Supply chain metrics. International Journal of Logistics Management 12 (1), 1-19. Larsen, T.S., Thenoe, C. and Andresen, C. (2003). Supply chain collaboration: Theoretical perspectives and empirical evidence. International Journal of Physical Distribution & Logistics Management 33 (6), 531-549. Lee, H. and Whang, S. (2001). Demand chain excellence: A tail of two retailers. Supply Chain Management Review 5 (2), 40-46 Lee, H. L. (2002). Aligning Supply Chain Strategies with Product Uncertainties. California Management Review 44 (3), 105-119. Lee, H.L., So, K.C. and Tang, C.S. (2000). The value of information sharing in a two-level supply chain. Management Science 46 (5), 626-643. Page 26 of 27 Lee., H.L. and Padmanabhan.,V. (1997). Information distortion in a supply chain: The bullwhip effect. Management Science 43, 546. Li, X. and Wang, Q. (2007). Coordination mechanisms of supply chain systems. European Journal of Operational Research 179 (1), 1-16. Matchette, J. and Seikel, A. (2004). How to win friends and influence supply chain partners. Logistics Today 45 (12), 40-42. McCarthy, T.M. and Golicic, S. L. (2002). Implementing collaborative forecasting to improve supply chain performance. International Journal of Physical Distribution & Logistics Management 32 (6), 431-454. Mishra, A.A. and Shah, R. (2009). In union lies strength: Collaborative competence in new product development and its performance effects. Journal of Operations Management 27 (4), 324-338. Nyaga, G.N., Whipple, J.M. and Lynch,D.F. (2010). Examiningsupplychainrelation-ships:do buyer andsupplier perspectives on collaborative relationships differ? Journal of Operations Management 28,101–114. Ovalle, O.R., Marquez, A.C. (2003). The effectiveness of using e-collaboration tools in the supply chain: An assessment study with system dynamics. Journal of Purchasing & Supply Management 9 (4), 151-163. Raghunathan, S. (2001). Information sharing in a supply chain: A note on its value when demand is non stationary. Management Science 47 (4), 605-610. Raghunathan, S. (2001). Information sharing in a supply chain: A note on its value when demand is non stationary. Management Science 47 (4), 605-610. Ramanathan, U and Muyldermans, L. (2011). Identifying the underlying structure of demand during promotions: A structural equation modelling approach. Expert Systems with Applications 38(5), 5544-5552. Ramanathan, U. (2012a). Aligning supply chain collaboration using Analytic Hierarchy Process. Omega - The International Journal of Management Science (forthcoming.doi:10.1016/j.omega.2012.03.001). Ramanathan, U. (2012b). Supply chain collaboration for improved forecast accuracy of promotional sales, International Journal of Operations and Production Management, 32 (6) (forthcoming). Ramanathan, U. and Muyldermans, L. (2010). Identifying demand factors for promotional planning and forecasting: A case of a soft drink company in the UK. International Journal of Production Economics, 128 (2), 538-545. Ramanathan, U., Gunasekaran, A. and Subramanian, N. (2011), Performance metrics for collaborative supply chain: A conceptual framework from case studies, Benchmarking: An International Journal, 18 (6), 856-872. Ramanathan. U. and Gunasekaran, A., (2013). Supply chain collaboration: Impact of success in long-term partnerships. International Journal of Production Economics. (forthcoming.http://dx.doi.org/10.1016/j.ijpe.2012.06.002). Sanders, N. R. and Premus, R. (2005). Modelling the relationship between firm IT capability collaboration and performance. Journal of Business Logistics 26 (1), 1-23. Page 27 of 27 Seifert, D. (2003). Collaborative Planning Forecasting and Replenishment: How to create a supply chain advantage. Saranac Lake NY USA: AMACOM. Shafiei, F., Sundaram, D. And Piramuthu, S., (2012). Muti-enterprise collaboratie decision support system. Expert Systems with Applications. 39 (9), 7637-7651. Sinha, A.K., Aditya, H.K., Tiwari,M.K. achd Chan, F.T.S. (2011). Agent oriented petroleum supplhy chain coordination:Co-evaluationary particle swarm optimisation based approach. Expert Systems with Applications, 38(5). 6132-6145. Simatupang, T.M. and Sridharan, R. (2004). Benchmarking supply chain collaboration: An empirical study. Benchmarking: An international Journal 11 (5), 484-503. Singh, P.F. and Power, D. (2009). The nature and effectiveness of collaboration between firms, their customers and suppliers: a supply chain perspective. Supply Chain Management: An International Journal 14 (3), 189-200. Smaros, J. (2007). Forecasting collaboration in the European grocery sector: Observations from a case study. Journal of Operations Management 25 (3), 702-716. Smith, L. (2006). West Marine:CPFR success story. Supply Chain Management Review 10 (2), 29-36. Stank, T.P., Keller, S. B. and Daugherty, P.J. (2001). Supply chain collaboration and logistical service performance. Journal of Business Logistics 22 (1), 29-48. Steermann, H. (2003). A practical look at CPFR: The Sears-Michelin experience. Supply Chain Management Review 7 (4), 46-53. VICS (2002). CPFR guidelines Voluntary Inter-industry Commerce Standards. Available at: www.cpfr.org (accessed January 2007). Yu, Z., Yan, H and Cheng, T.C.E. (2001) Benefits of information sharing with supply chain partnerships, Industrial Management & Data Systems,101(3), 114-119. Acknowledgement: Author would like to thank - two anonymous reviewers for their valuable comments and Dr Yongmei Bentley for her support in improving the paper. 
work_3akpcw3alvfzpbwlc3yzkadwpa	----	An Expert System for Quantification of Bradykinesia Based on Wearable Inertial Sensors sensors Article An Expert System for Quantification of Bradykinesia Based on Wearable Inertial Sensors Vladislava Bobić 1,2,*, Milica Djurić-Jovičić 2, Nataša Dragašević 3, Mirjana B. Popović 1,4, Vladimir S. Kostić 3 and Goran Kvaščev 1 1 University of Belgrade-School of Electrical Engineering, 11000 Belgrade, Serbia; mpo@etf.rs (M.B.P.); kvascev@etf.rs (G.K.) 2 Innovation Center, School of Electrical Engineering, University of Belgrade, 11000 Belgrade, Serbia; milica.djuric@etf.rs 3 Clinic of Neurology, School of Medicine, University of Belgrade, 11000 Belgrade, Serbia; ntdragasevic@gmail.com (N.D.); vladimir.s.kostic@gmail.com (V.S.K.) 4 Institute for Medical Research, University of Belgrade, 11000 Belgrade, Serbia * Correspondence: vladislava.bobic@ic.etf.rs; Tel.: +381-11-3218-455 Received: 3 April 2019; Accepted: 4 June 2019; Published: 11 June 2019 ���������� ������� Abstract: Wearable sensors and advanced algorithms can provide significant decision support for clinical practice. Currently, the motor symptoms of patients with neurological disorders are often visually observed and evaluated, which may result in rough and subjective quantification. Using small inertial wearable sensors, fine repetitive and clinically important movements can be captured and objectively evaluated. In this paper, a new methodology is designed for objective evaluation and automatic scoring of bradykinesia in repetitive finger-tapping movements for patients with idiopathic Parkinson’s disease and atypical parkinsonism. The methodology comprises several simple and repeatable signal-processing techniques that are applied for the extraction of important movement features. The decision support system consists of simple rules designed to match universally defined criteria that are evaluated in clinical practice. The accuracy of the system is calculated based on the reference scores provided by two neurologists. The proposed expert system achieved an accuracy of 88.16% for files on which neurologists agreed with their scores. The introduced system is simple, repeatable, easy to implement, and can provide good assistance in clinical practice, providing a detailed analysis of finger-tapping performance and decision support for symptom evaluation. Keywords: decision support system; wearable inertial sensors; finger-tapping; automatic scoring; Parkinson’s disease; atypical parkinsonism; UPDRS 1. Introduction Wearable sensors and advanced algorithms are increasingly being used for the development of new clinical support systems for more efficient diagnostics and the evaluation of symptom severity and disease progress in Parkinson’s disease (PD) [1]. This covers a wide range of applications assessing different symptoms, such as tremor, hypokinesia, rigidity, and bradykinesia. Bradykinesia is one of the main manifestations of PD. It is evidenced as slowness of body movements, especially in tasks that require fine motor control [2]. In clinical practice, bradykinesia (as well as other motor symptoms) are usually assessed using the Unified Parkinson’s Disease Rating Scale (UPDRS), in which the third part of the examination is dedicated to motor skill evaluation (UPDRS III) [3]. Bradykinesia is evaluated using repetitive hand and leg movements, such as finger-tapping, hand opening/closing, pronation/supination, and foot (or toe) tapping [2]. As a part of the examination, patients are requested to repeatedly perform specified movements, as fast and Sensors 2019, 19, 2644; doi:10.3390/s19112644 www.mdpi.com/journal/sensors http://www.mdpi.com/journal/sensors http://www.mdpi.com http://www.mdpi.com/1424-8220/19/11/2644?type=check_update&version=1 http://dx.doi.org/10.3390/s19112644 http://www.mdpi.com/journal/sensors Sensors 2019, 19, 2644 2 of 17 with the biggest amplitude as possible, during some short period of time, usually 10–15 s [4–7], or for some specified number of repetitions, e.g., 10 times [8–10]. These movements are evaluated based on specifically defined criteria, including speed, amplitude, amplitude decrement, and number of hesitations or freezes. The performance is rated with scores ranging from 0 to 4, in which the lowest values correspond to normal movements, and higher values are given for more severe bradykinesia expressed through significant amplitude losses, decreasing speed, or an increased number of hesitations/freezes. However, in clinical practice, this examination is usually performed visually, which may result in subjective evaluation and rough quantification. Since precise evaluation represents a very important part of the long-term monitoring of the disease’s progress and patients’ response to therapy, researchers have dedicated their effort and time to design new systems that can be used for the objective evaluation and automatic scoring of symptom severity. In the literature, different approaches are presented for the objective evaluation and quantification of PD motor symptom severity, including bradykinesia. The introduced methodologies differ in terms of applied instrumentation, analysed movements, measurement protocols, the size and composition of patients’ groups, and implemented signal processing and learning techniques. Some studies implement RGB or infrared camera systems for measuring clinically important repetitive hand and leg movements [5,11,12]. Although such systems can provide high-precision measurements, they have some limitations. They are expensive and require dedicated space for recording (they are bulky), which significantly limits their applicability in clinical settings [13]. Due to these limitations, wearable systems, such as smartphones [14], magnetic sensors [13], and inertial measurement units (IMUs) [7–9,15–20], are increasingly being applied for bradykinesia assessment. IMUs are small, lightweight, easy to mount, and do not require dedicated space for recording, which makes them more suitable for fast and reliable everyday clinical applications. In the literature, bradykinesia is assessed by analysing different repetitive movements, including finger-tapping [8,13,21], hand opening/closing [16,17], hand pronation/supination [9,10,22], and toe tapping [23], as well as by simultaneous analysis of different movements [12,20,24]. From finger-tapping (FT) accelerometer data, researchers extract different features, describing the frequency and biomechanical properties of the movements, and use them as input into the ordinal logistic regression model for prediction of UPDRS FT scores [8]. It is shown that scores can be predicted with high predictive power (the Goodman–Kruskal Gamma score is 0.961). Four parameters extracted from the FT gyro data were found to be statistically correlated with clinical scores (from r = 0.73 to r = −0.80) [7]. A similar approach is applied to a repetitive hand opening/closing task [16]. Signals are acquired with small IMUs and described by the dominant grasping frequency and mean angle, and fitted with the clinical UPDRS scores using a regression model [16]. It is shown that the predicted scores are highly correlated with the clinical scores (the determination coefficient is r2 = 0.99). A methodology that combines principal component analysis and multiple linear regression is applied to quantify bradykinesia severity in FT movements [13]. The method is applied to features that are extracted from the data recorded using magnetic sensors. It is shown that this approach can provide scores with a mean square error of 0.45 compared to the reference UPDRS FT scores. Another study presented a new approach that uses a motion capture system and dynamical features rather than standard spectral features for automatic scoring of the FT performance [5]. The results show strong and significant correlations with clinical scores. In order to describe bradykinesia in multi-joint upper limb movements, researchers have introduced new performance indexes that are correlated with UPDRS bradykinesia scores and implemented for differentiation between PD patients with and without bradykinesia [22]. A similar approach is designed for the evaluation of bradykinesia in walking and sit-to-stand tasks, in which novel performance indices are successfully used for differentiation between healthy subjects and PD patients, and ON and OFF states in patients [25]. A support vector machine (SVM) classifier applied to spectral and nonlinear features achieves high-accuracy results (accuracy, sensitivity, and specificity above 97%) for the prediction of UPDRS FT scores (0–3) [19]. However, the method is applied to Sensors 2019, 19, 2644 3 of 17 gyro signals recorded from healthy subjects who mimick the impaired movements of PD patients. In another study, SVM was successfully applied (error below 5%) for estimation of the severity of several symptoms (bradykinesia, tremor, and dyskinesia) in 12 PD patients using the features extracted from accelerometer data describing multiple upper and lower extremity movements [20]. SVM was also applied for estimation of bradykinesia severity in a study comprising 78 PD patients and 18 healthy subjects, who were instructed to perform hand opening/closing for 10 s [4]. It was shown that SVM can predict clinical scores with an accuracy of 95.349%. Decision trees, applied to features extracted from inertial signals, were also used for prediction of UPDRS scores for a pronation/supination task, showing a mean agreement of 0.48 with clinical ratings [10]. Although supervised machine learning algorithms provide prediction of clinical scores with high accuracy, the applied models are trained on a smaller dataset with subjectively defined data labels, which may cause subjectivity in the results as well. Because of that, some researchers have introduced different approaches to this topic. Decision rules can be designed to match exactly the criteria of the decision-making process and instructions applied in clinical practice. Fuzzy rules were applied for prediction of clinical scores using inertial data recorded during foot tapping [23] and hand pronation/supination movements in [9]. Their designed rules provide good results, with an accuracy of about 90%. Great Lake Technologies proposed a commercialized smartphone application, called Kinesia One, that provides clinical scores and subscores for different criteria for several bradykinesia tasks using a inertial sensor positioned on the index finger [26]. In this paper, we propose a new decision support system for the provision of clinical scores based on the use of inertial data describing finger-tapping movements. The proposed system uses novel metric and decision rules that are especially designed to capture and evaluate the relevant characteristics of the finger-tapping movement. The system provides very good results for data obtained from patients with idiopathic Parkinson’s disease but also with atypical parkinsonism. The output of the system comprises the kinematic features describing the finger-tapping performance, a graphical presentation of the recorded data with marked irregularities, and important changes in the signal and bradykinesia severity scores. 2. Materials and Methods 2.1. Measurement System The used system comprises two miniature (10 × 12 mm) and lightweight inertial sensors with three-dimensional (3D) gyroscopes L3G4200 (STMicroelectronics, Geneva, Switzerland) positioned over the fingernails of the thumb and index finger, as shown in Figure 1 [6]. Inertial sensors are connected to sensor-control units (SCUs). An SCU acquires and wirelessly transmits sensor data to a remote computer, where custom-made software controls data acquisition (developed in CVI 9.0, NI LabWindows, National Instruments, Austin, Texas, USA). Sensors 2019, 19, x FOR PEER REVIEW 3 of 17 severity of several symptoms (bradykinesia, tremor, and dyskinesia) in 12 PD patients using the features extracted from accelerometer data describing multiple upper and lower extremity movements [20]. SVM was also applied for estimation of bradykinesia severity in a study comprising 78 PD patients and 18 healthy subjects, who were instructed to perform hand opening/closing for 10 s [4]. It was shown that SVM can predict clinical scores with an accuracy of 95.349%. Decision trees, applied to features extracted from inertial signals, were also used for prediction of UPDRS scores for a pronation/supination task, showing a mean agreement of 0.48 with clinical ratings [10]. Although supervised machine learning algorithms provide prediction of clinical scores with high accuracy, the applied models are trained on a smaller dataset with subjectively defined data labels, which may cause subjectivity in the results as well. Because of that, some researchers have introduced different approaches to this topic. Decision rules can be designed to match exactly the criteria of the decision-making process and instructions applied in clinical practice. Fuzzy rules were applied for prediction of clinical scores using inertial data recorded during foot tapping [23] and hand pronation/supination movements in [9]. Their designed rules provide good results, with an accuracy of about 90%. Great Lake Technologies proposed a commercialized smartphone application, called Kinesia One, that provides clinical scores and subscores for different criteria for several bradykinesia tasks using a inertial sensor positioned on the index finger [26]. In this paper, we propose a new decision support system for the provision of clinical scores based on the use of inertial data describing finger-tapping movements. The proposed system uses novel metric and decision rules that are especially designed to capture and evaluate the relevant characteristics of the finger-tapping movement. The system provides very good results for data obtained from patients with idiopathic Parkinson’s disease but also with atypical parkinsonism. The output of the system comprises the kinematic features describing the finger-tapping performance, a graphical presentation of the recorded data with marked irregularities, and important changes in the signal and bradykinesia severity scores. 2. Materials and Methods 2.1. Measurement System The used system comprises two miniature (10 × 12 mm) and lightweight inertial sensors with three-dimensional (3D) gyroscopes L3G4200 (STMicroelectronics, Geneva, Switzerland) positioned over the fingernails of the thumb and index finger, as shown in Figure 1 [6]. Inertial sensors are connected to sensor-control units (SCUs). An SCU acquires and wirelessly transmits sensor data to a remote computer, where custom-made software controls data acquisition (developed in CVI 9.0, NI LabWindows, National Instruments, Austin, Texas, USA). Figure 1. Illustration of the inertial sensor system, with local coordinate systems of the thumb (𝑋 , 𝑌 , 𝑍 ) and index finger (𝑋 , 𝑌 , 𝑍 ) sensors. SCU - sensor-control unit. Figure 1. Illustration of the inertial sensor system, with local coordinate systems of the thumb (X1, Y1, Z1) and index finger (X2, Y2, Z2) sensors. SCU-sensor-control unit. Sensors 2019, 19, 2644 4 of 17 2.2. Subjects Fifty-six subjects were recruited for this study from the Clinic of Neurology, Clinical Centre of Serbia, Belgrade. The subjects included 13 patients (Gender: seven male/six female, Age: 62.23 ± 10.79 years) with idiopathic Parkinson’s disease (PD), 17 patients (Gender: five male/12 female, Age: 58.41 ± 6.41 years) with atypical parkinsonism multiple system atrophy (MSA), 14 patients (Gender: 11 male/three female, Age: 65.71 ± 9.33 years) with atypical parkinsonism progressive supranuclear palsy (PSP), and 12 healthy controls (HC) (Gender: four male/eight female, Age: 58.40 ± 7.78 years). The patients were tested during their “off” phase (after at least 12 h of treatment withdrawal, if possible). Descriptive statistics (average ± standard deviation and median) of the clinical data for each group of subjects are presented in Table 1, including the Hoehn and Yahr (H&Y) scale, total UPDRS, UPDRS-III (complete Motor examination scores), and scores given solely for the finger-tapping task by two neurologists, separately for the less- and more-affected hand. Table 1. Descriptive statistics of the subjects’ data. Group Statistics H&Y UPDRS Total UPDRS III FTN1 Score FTN2 Score Less AH More AH Less AH More AH PD Avg ± std 1.80 ± 0.79 42.60 ± 16.93 24.60 ± 9.07 1.67 ± 0.89 2.17 ± 0.94 1.75 ± 0.97 2.17 ± 0.94 Median 2 36 19.5 2 2 2 2 MSA Avg ± std 3.18 ± 0.75 77.73 ± 13.70 46.64 ± 9.08 2.31 ± 0.70 2.81 ± 0.54 2.38 ± 0.72 2.81 ± 0.54 Median 3 79 45 2 3 2.5 3 PSP Avg ± std 3.45 ± 0.93 74.45 ± 20.08 42.91 ± 13.14 2.17 ± 0.94 2.62 ± 0.77 2.08 ± 0.79 2.77 ± 0.73 Median 4 79 46 2.5 3 2 3 HC Avg ± std / / / 0.44 ± 0.63 0.50 ± 0.73 Median / / / 0 0 1 PD—Parkinson’s disease; MSA—Multiple system atrophy; PSP—Progressive supranuclear palsy; HC—Healthy controls; H&Y—Hoehn and Yahr scale; UPDRS–Unified Parkinson’s Disease Rating Scale; UPDRS III—Unified Parkinson’s Disease Rating Scale, Part III—Motor examination; FTN1—Finger-tapping score provided by the first neurologist; FTN2—Finger-tapping score provided by the second neurologist; Less AH—The less-affected hand, More AH—The more-affected hand. 2.3. Measurement Methodology During the recordings, the subjects were sitting in the chair with their arms bent and supported at the elbow and hands placed in front of them. They were instructed to perform the finger-tapping test by tapping their thumb and index finger as quickly and as widely as possible for 15 s. Although instructions provided in the UPDRS test state that patients should tap their fingers 10 times [3], in this study, longer recordings were acquired to ensure that sufficient data for analysis were available. In order to become accustomed to the instrumentation and measurement methodology, for each subject, several trials were recorded per hand, with one minute of rest between the trials. Each trial was also recorded with a video camera, which filmed the hand in a close-up view. The most representative recording (one for each hand) was used for further analysis. The recording was selected by the neurologists as the recording that fullfills the requirements of tapping duration and patients’ understanding of the given instructions. The testing of each subject was performed during one day at the Clinic of Neurology, Clinical Centre of Serbia, Belgrade. The examination was carried out in accordance with the ethical standards of the Declaration of Helsinki, and approved by the Ethical Committee of the School of Medicine, University of Belgrade. All of the participants provided informed consent prior to participation in the study. 2.4. Scoring by Neurologists The recorded video data were later examined and scored by two neurologists with more than 10 years of experience, based on their knowledge and experience and the instructions given in the UPDRS, Part III–Motor examination, task 3.4 Finger tapping. The neurologists were blinded to the subjects’ identity, since the video data show a close-up view of each subject’s hand. The scores were given for Sensors 2019, 19, 2644 5 of 17 each patient, separately for the left and right hand. The scores given by neurologists are provided in Table 1, in the last two columns. The scores given for the patients were provided separately for the less- and more-affected hand (averaged for all patients per group), whereas, in the case of healthy controls, the scores were averaged for both hands and all HC participants. 2.5. Data Processing and Analysis The gyro data were recorded with a sampling frequency fs = 200 Hz. Calibrated data were processed in Matlab 9.0 R2016a (MathWorks, Natick, MA, USA). The flowchart of the expert system for calculation of UPDRS finger-tapping scores is presented in Figure 2. The inputs to the expert system are angular velocities from the thumb ( → ω1) and index finger sensors ( → ω2). No pre-processing was performed on the input signals. Upon the sensor’s placement, the coordinate system of the thumb (X1, Y1, Z1) and the coordinate system of the index finger (X2, Y2, Z2) were rotated with respect to each other (Figure 1). The angular velocities were transformed and analyzed from the index-finger coordinate system. The relative angular velocity of the thumb with respect to the index finger was calculated → ωr = → ω1 − → ω2 [27]. The dominant component of the relative angular velocity ωrd was automatically selected and used as the input in further data processing and analysis [27]. It was shown that, in most cases, the dominant component of the relative angular velocity is about the Y2−axis of the index-finger coordinate system. In other cases (when this component is not dominant), the coordinate system of the index finger was rotated, so the new Y2−axis represents the dominant rotation. Sensors 2019, 19, x FOR PEER REVIEW 5 of 17 2.4. Scoring by Neurologists The recorded video data were later examined and scored by two neurologists with more than 10 years of experience, based on their knowledge and experience and the instructions given in the UPDRS, Part III–Motor examination, task 3.4 Finger tapping. The neurologists were blinded to the subjects’ identity, since the video data show a close-up view of each subject’s hand. The scores were given for each patient, separately for the left and right hand. The scores given by neurologists are provided in Table 1, in the last two columns. The scores given for the patients were provided separately for the less- and more-affected hand (averaged for all patients per group), whereas, in the case of healthy controls, the scores were averaged for both hands and all HC participants. 2.5. Data Processing and Analysis The gyro data were recorded with a sampling frequency 𝑓 = 200 𝐻𝑧. Calibrated data were processed in Matlab 9.0 R2016a (MathWorks, Natick, MA, USA). The flowchart of the expert system for calculation of UPDRS finger-tapping scores is presented in Figure 2. The inputs to the expert system are angular velocities from the thumb (𝜔⃗) and index finger sensors (𝜔 ⃗). No pre-processing was performed on the input signals. Upon the sensor’s placement, the coordinate system of the thumb (𝑋 , 𝑌 , 𝑍 ) and the coordinate system of the index finger (𝑋 , 𝑌 , 𝑍 ) were rotated with respect to each other (Figure 1). The angular velocities were transformed and analyzed from the index-finger coordinate system. The relative angular velocity of the thumb with respect to the index finger was calculated 𝜔⃗ = 𝜔 ⃗ − 𝜔 ⃗ [27]. The dominant component of the relative angular velocity 𝜔 was automatically selected and used as the input in further data processing and analysis [27]. It was shown that, in most cases, the dominant component of the relative angular velocity is about the 𝑌 −axis of the index-finger coordinate system. In other cases (when this component is not dominant), the coordinate system of the index finger was rotated, so the new 𝑌 −axis represents the dominant rotation. Further analysis was divided into one pre-processing block for segmentation to individual taps and four blocks that calculate features to describe criteria defined in the UPDRS test: tapping amplitude, amplitude decrement, hesitations and freezes, and tapping speed. The calculated features are then used as the input to the decision-support system. As the result, a complete analysis of the patient’s finger-tapping performance, including the finger-tapping score (0–4), is provided. Figure 2. Block diagram of the expert system for UPDRS finger-tapping score calculation. Figure 2. Block diagram of the expert system for UPDRS finger-tapping score calculation. Further analysis was divided into one pre-processing block for segmentation to individual taps and four blocks that calculate features to describe criteria defined in the UPDRS test: tapping amplitude, amplitude decrement, hesitations and freezes, and tapping speed. The calculated features are then used as the input to the decision-support system. As the result, a complete analysis of the patient’s finger-tapping performance, including the finger-tapping score (0–4), is provided. Sensors 2019, 19, 2644 6 of 17 2.5.1. Individual Taps In order to evaluate characteristics of the tapping performance for individual taps, segmentation of the dominant component of the relative angular velocity ωrd was performed. A moving-average filter was applied to the observed signal with a span equal to ( fS/ f0)/2, where f0 represents the basic tapping frequency extracted from the spectrum. The filtered signal was normalized to its maximum value. From the obtained sequence, areas above 0.1 and below −0.1 were identified, corresponding to regions where positive peaks and negative valleys are located, respectively. Local extrema were identified for each of the extracted regions. Positive peaks correspond to the maximal closing velocity (circles, Figure 3), whereas negative valleys represent moments when fingers achieve the maximal opening velocity (squares, Figure 3). The samples in which the smoothed angular velocity ωrd passes through a zero value for the first time were identified between each neighbouring maximum and minimum marker. These samples represent the moments when fingers are closed (“zero posture”). The sequence was complemented with the first and last sample. The finally obtained samples were identified as time markers for drift removal and segmentation on individual taps (crosses, Figure 3). Sensors 2019, 19, x FOR PEER REVIEW 6 of 17 2.5.1. Individual Taps In order to evaluate characteristics of the tapping performance for individual taps, segmentation of the dominant component of the relative angular velocity 𝜔 was performed. A moving-average filter was applied to the observed signal with a span equal to (𝑓 /𝑓 )/2, where 𝑓 represents the basic tapping frequency extracted from the spectrum. The filtered signal was normalized to its maximum value. From the obtained sequence, areas above 0.1 and below −0.1 were identified, corresponding to regions where positive peaks and negative valleys are located, respectively. Local extrema were identified for each of the extracted regions. Positive peaks correspond to the maximal closing velocity (circles, Figure 3), whereas negative valleys represent moments when fingers achieve the maximal opening velocity (squares, Figure 3). The samples in which the smoothed angular velocity 𝜔 passes through a zero value for the first time were identified between each neighbouring maximum and minimum marker. These samples represent the moments when fingers are closed (“zero posture”). The sequence was complemented with the first and last sample. The finally obtained samples were identified as time markers for drift removal and segmentation on individual taps (crosses, Figure 3). Figure 3. An example of the normalized dominant component of the relative angular velocity 𝜔 for one MSA patient (ID: MSA11) with extracted markers. 2.5.2. Amplitude One of the evaluation criteria is the tapping amplitude, which evaluates how widely subjects can tap their fingers. The finger-tapping amplitude is defined as the angle that fingers formed during repetitive tapping movements. The tapping angle was calculated by integrating the dominant component of the relative angular velocity 𝜔 [27]. The drift was removed by using a third-order polynomial fitted (approximation) through markers corresponding to moments where fingers are closed (i.e., the angle is equal to zero, red crosses in Figure 4). Upon the drift’s removal, the obtained angle sequence was segmented into individual taps using the same time markers. The highest aperture of the fingers (the biggest angle that fingers form) was found for each individual tap and is expressed in degrees 𝛼(𝑖) (°) (black circles in Figure 4, lower panel). The final parametric result was calculated as the average of the maximum angles calculated for each tap - 𝛼 (°). Figure 3. An example of the normalized dominant component of the relative angular velocity ωrd for one MSA patient (ID: MSA11) with extracted markers. 2.5.2. Amplitude One of the evaluation criteria is the tapping amplitude, which evaluates how widely subjects can tap their fingers. The finger-tapping amplitude is defined as the angle that fingers formed during repetitive tapping movements. The tapping angle was calculated by integrating the dominant component of the relative angular velocity ωrd [27]. The drift was removed by using a third-order polynomial fitted (approximation) through markers corresponding to moments where fingers are closed (i.e., the angle is equal to zero, red crosses in Figure 4). Upon the drift’s removal, the obtained angle sequence was segmented into individual taps using the same time markers. The highest aperture of the fingers (the biggest angle that fingers form) was found for each individual tap and is expressed in degrees α(i) (◦) (black circles in Figure 4, lower panel). The final parametric result was calculated as the average of the maximum angles calculated for each tap-αav (◦). Sensors 2019, 19, 2644 7 of 17 Sensors 2019, 19, x FOR PEER REVIEW 7 of 17 Figure 4. Upper panel: Angle estimation. The dashed grey line marks the drifted angle sequence, and the solid black line corresponds to the angle sequence after drift removal. Red crosses show “zero posture” markers, and the dotted red line presents the polynomial fit used for drift removal. Lower panel: Angle amplitude decrement. The solid grey line shows the angle sequence, whereas black circles mark the angle amplitudes (highest finger apertures) per tap. The dashed red line presents the threshold 𝑇𝐻 used for the detection of decreased amplitudes. The example is given for one MSA patient (ID: MSA11). 2.5.3. Amplitude Decrement Physicians evaluate the amplitude decrement according to the part of the tapping sequence at which the amplitude starts to decrease. In order to objectively quantify changes of the tapping amplitude, we observed tap-to-tap changes in the highest finger apertures calculated for individual taps 𝛼(𝑖). The angle amplitude of each individual tap was compared with the previously achieved maximum aperture of the fingers. The threshold 𝑇𝐻 = 75% of the value of the previous maximum finger aperture was selected as the optimum (as shown in Figure 4, lower panel). This threshold was heuristically determined through extensive analysis of the used signal database. Threshold values from 50% to 90% (with a step of 5%) were chosen and tested. The threshold of 75% provides the best results for the prediction of scores. It was shown that higher threshold values cause detection of very small amplitude changes, which can appear due to normal movement variability. Lower threshold values detect angle decrements later in the tapping sequence, with some delay compared to the first real significant decrement. Indices of all taps that satisfy this criterion for the amplitude decrease were extracted using the chosen threshold 𝑇𝐻 (for the example shown in the lower panel of Figure 4, all taps are below the threshold except for the first one, which is used as the reference for the calculation of the threshold). The indices of the first tap from the obtained sequence were selected as the final parametric result and marked with 𝑖 (for the example in Figure 4, that is the second tap and, therefore, 𝑖 = 2). 2.5.4. Hesitations and Freezes Hesitations and freezes are manifested as irregularities or breaks of the tapping rhythm that may occur in different moments of the tapping performance and represent an important part of the finger- tapping evaluation. The continuous wavelet transform (CWT) was applied for the detection and Figure 4. Upper panel: Angle estimation. The dashed grey line marks the drifted angle sequence, and the solid black line corresponds to the angle sequence after drift removal. Red crosses show “zero posture” markers, and the dotted red line presents the polynomial fit used for drift removal. Lower panel: Angle amplitude decrement. The solid grey line shows the angle sequence, whereas black circles mark the angle amplitudes (highest finger apertures) per tap. The dashed red line presents the threshold THα used for the detection of decreased amplitudes. The example is given for one MSA patient (ID: MSA11). 2.5.3. Amplitude Decrement Physicians evaluate the amplitude decrement according to the part of the tapping sequence at which the amplitude starts to decrease. In order to objectively quantify changes of the tapping amplitude, we observed tap-to-tap changes in the highest finger apertures calculated for individual taps α(i). The angle amplitude of each individual tap was compared with the previously achieved maximum aperture of the fingers. The threshold THα = 75% of the value of the previous maximum finger aperture was selected as the optimum (as shown in Figure 4, lower panel). This threshold was heuristically determined through extensive analysis of the used signal database. Threshold values from 50% to 90% (with a step of 5%) were chosen and tested. The threshold of 75% provides the best results for the prediction of scores. It was shown that higher threshold values cause detection of very small amplitude changes, which can appear due to normal movement variability. Lower threshold values detect angle decrements later in the tapping sequence, with some delay compared to the first real significant decrement. Indices of all taps that satisfy this criterion for the amplitude decrease were extracted using the chosen threshold THα (for the example shown in the lower panel of Figure 4, all taps are below the threshold except for the first one, which is used as the reference for the calculation of the threshold). The indices of the first tap from the obtained sequence were selected as the final parametric result and marked with idec (for the example in Figure 4, that is the second tap and, therefore, idec = 2). 2.5.4. Hesitations and Freezes Hesitations and freezes are manifested as irregularities or breaks of the tapping rhythm that may occur in different moments of the tapping performance and represent an important part of the finger-tapping evaluation. The continuous wavelet transform (CWT) was applied for the detection Sensors 2019, 19, 2644 8 of 17 and localization of disruptions of the tapping rhythmicity [28]. It is a time-frequency analysis method that is suitable for the analysis of transient changes and spikes in rhythmic behaviour [29]. CWT was applied on the dominant component of the relative angular velocity (ωrd). The CWT method based on the Fast Fourier transform algorithm was used, together with the mother wavelet function from the complex Morlet family (center frequency f 0 = 1 Hz and time-frequency resolution σ = 0.7). A matrix of complex CWT coefficients was obtained as a result. We introduced a cross-sectional area by summing the CWT coefficients perpendicular to the time axis. The obtained characteristic was normalized with respect to its maximum value and is expressed as a percentage (CSAT (%)). In this way, we obtain a characteristic that describes the temporal changes in the tapping activity [28]. An example of the CSAT characteristic for one patient is presented in Figure 5, lower panel. The samples were then divided according to two thresholds: TH50 = 50% of the average CSAT value, and TH25 = 25% of the average CSAT value (the dashed grey and dotted black vertical lines in Figure 5, respectively). Samples with values below TH50 and above TH25 threshold were considered to be parts of the hesitation sequences (Figure 5, dotted grey vertical lines, with an “H” mark), whereas samples with the smallest amplitude (below TH25) were considered to be parts of freezes (Figure 5, dotted grey vertical lines, with an “F” mark). If a hesitation sequence lasts three times longer than the subjects’ average tapping frequency, then it is considered to be a freeze sequence. In addition, very short sequences (shorter than one half of the subjects’ average tapping frequency) were discarded from the analysis. The parametric result comprises the number of hesitation sequences Hnum and the number of freeze sequences Fnum. Using the average CSAT value for thresholds ensures that the detection of irregularities is adapted to the intrinsic properties of each signal, considering the signal parts with significant losses in power (below 50% and 25% of the average) as irregularities. The values of the applied thresholds were verified through an extensive search of the database. All detected irregularities were confirmed by neurologists during their visual inspection of video recordings. Sensors 2019, 19, x FOR PEER REVIEW 8 of 17 localization of disruptions of the tapping rhythmicity [28]. It is a time-frequency analysis method that is suitable for the analysis of transient changes and spikes in rhythmic behaviour [29]. CWT was applied on the dominant component of the relative angular velocity (𝜔 ). The CWT method based on the Fast Fourier transform algorithm was used, together with the mother wavelet function from the complex Morlet family (center frequency f0 = 1 Hz and time-frequency resolution σ = 0.7). A matrix of complex CWT coefficients was obtained as a result. We introduced a cross-sectional area by summing the CWT coefficients perpendicular to the time axis. The obtained characteristic was normalized with respect to its maximum value and is expressed as a percentage (𝐶𝑆𝐴 (%)). In this way, we obtain a characteristic that describes the temporal changes in the tapping activity [28]. An example of the 𝐶𝑆𝐴 characteristic for one patient is presented in Figure 5, lower panel. The samples were then divided according to two thresholds: 𝑇𝐻 = 50% of the average 𝐶𝑆𝐴 value, and 𝑇𝐻 = 25% of the average 𝐶𝑆𝐴 value (the dashed grey and dotted black vertical lines in Figure 5, respectively). Samples with values below 𝑇𝐻 and above 𝑇𝐻 threshold were considered to be parts of the hesitation sequences (Figure 5, dotted grey vertical lines, with an “H” mark), whereas samples with the smallest amplitude (below 𝑇𝐻 ) were considered to be parts of freezes (Figure 5, dotted grey vertical lines, with an “F” mark). If a hesitation sequence lasts three times longer than the subjects’ average tapping frequency, then it is considered to be a freeze sequence. In addition, very short sequences (shorter than one half of the subjects’ average tapping frequency) were discarded from the analysis. The parametric result comprises the number of hesitation sequences 𝐻 and the number of freeze sequences 𝐹 . Using the average 𝐶𝑆𝐴 value for thresholds ensures that the detection of irregularities is adapted to the intrinsic properties of each signal, considering the signal parts with significant losses in power (below 50% and 25% of the average) as irregularities. The values of the applied thresholds were verified through an extensive search of the database. All detected irregularities were confirmed by neurologists during their visual inspection of video recordings. Figure 5. Calculation of hesitations and freezes: angular velocity 𝜔 (upper panel) and calculated 𝐶𝑆𝐴 characteristic (bottom panel). The solid grey horizontal line marks the average 𝐶𝑆𝐴 value. The dashed grey horizontal line corresponds to the upper threshold 𝑇𝐻 = 50% of the 𝐶𝑆𝐴 average value. The dotted black horizontal line shows the lower threshold 𝑇𝐻 = 25% of the 𝐶𝑆𝐴 - average value. Similarly, dotted grey vertical lines show areas that are classified as hesitations (an “H” mark) and freezes (an “F” mark). The example is given for one PSP patient (ID: PSP14). 2.5.5. Speed An important criterion for evaluation of bradykinesia in the finger-tapping task is the tapping speed. During the evaluation, neurologists examine how fast subjects are tapping. If subjects tap faster, then during those 15 s of the tapping test they perform a larger number of taps, and vice versa. Figure 5. Calculation of hesitations and freezes: angular velocity ωrd (upper panel) and calculated CSAT characteristic (bottom panel). The solid grey horizontal line marks the average CSAT value. The dashed grey horizontal line corresponds to the upper threshold TH50 = 50% of the CSAT average value. The dotted black horizontal line shows the lower threshold TH25 = 25% of the CSAT- average value. Similarly, dotted grey vertical lines show areas that are classified as hesitations (an “H” mark) and freezes (an “F” mark). The example is given for one PSP patient (ID: PSP14). 2.5.5. Speed An important criterion for evaluation of bradykinesia in the finger-tapping task is the tapping speed. During the evaluation, neurologists examine how fast subjects are tapping. If subjects tap Sensors 2019, 19, 2644 9 of 17 faster, then during those 15 s of the tapping test they perform a larger number of taps, and vice versa. Although this can also be evaluated from the number of performed taps and their duration, by using the calculated matrix of CWT coefficients, the dominant tapping frequency can be found for each time sample. In this way, all changes of the tapping rhythm are assessed, detected, and included in the analysis. The vector of coefficients corresponding to one sample was extracted from the CWT matrix. From the obtained vector, the most prominent frequency was calculated as the frequency at which the coefficient with the highest value is located (as shown for the i-th sample in Figure 6). The procedure was repeated for all samples. In this way, the new frequency characteristic f (i) was obtained. The average value of the frequency characteristic f (i) was calculated and is marked as f (i)av (Hz). Sensors 2019, 19, x FOR PEER REVIEW 9 of 17 Although this can also be evaluated from the number of performed taps and their duration, by using the calculated matrix of CWT coefficients, the dominant tapping frequency can be found for each time sample. In this way, all changes of the tapping rhythm are assessed, detected, and included in the analysis. The vector of coefficients corresponding to one sample was extracted from the CWT matrix. From the obtained vector, the most prominent frequency was calculated as the frequency at which the coefficient with the highest value is located (as shown for the i-th sample in Figure 6). The procedure was repeated for all samples. In this way, the new frequency characteristic 𝑓 ( ) was obtained. The average value of the frequency characteristic 𝑓 ( ) was calculated and is marked as 𝑓 ( ) (Hz). Figure 6. Calculation of the frequency characteristic: scalogram of the obtained continuous wavelet transform (CWT) coefficients. The dashed black line marks the i-th sample. The CWT coefficients at the i-th sample are presented in the smaller upper panel. The red dashed line in the upper panel marks the frequency with the highest amplitude of the CWT coefficients for the i-th sample (referred to as 𝑓 ( )). The example is given for one PSP patient (ID: PSP14). 2.5.6. Decision Support System In the UPDRS motor scale, Part III–Motor examination, task 3.4 Finger tapping [3], instructions for bradykinesia evaluation are given as follows: 0. Normal: Regular rhythm, without hesitations or freezes. Fast movement, large amplitude, no amplitude decrement. 1. Slight: Any of the following: (a) the regular rhythm is broken with one or two interruptions or hesitations of the tapping movement; (b) slight slowing; (c) the amplitude decrements near the end of the 10 taps. 2. Mild: Any of the following: (a) three to five interruptions during tapping; (b) mild slowing; (c) the amplitude decrements midway in the 10-tap sequence. 3. Moderate: Any of the following: (a) over five interruptions during tapping or at least one freeze in ongoing movement; (b) moderate slowing; (c) the amplitude decrements starting after the first tap. 4. Severe: Cannot or can only barely perform the task due to slowing, interruptions, or decrements. Each hand is evaluated separately, in terms of speed, amplitude, hesitation and freezes, and decrementing amplitude. These criteria are described with the introduced features, which are then fed to the decision support system. The input feature set includes the average tapping angle 𝛼 , the average frequency 𝑓 ( ), the index of the first tap with a significant angle amplitude decrement 𝑖 , the number of hesitations 𝐻 , and the number of freezes 𝐹 . The rules are defined separately for Figure 6. Calculation of the frequency characteristic: scalogram of the obtained continuous wavelet transform (CWT) coefficients. The dashed black line marks the i-th sample. The CWT coefficients at the i-th sample are presented in the smaller upper panel. The red dashed line in the upper panel marks the frequency with the highest amplitude of the CWT coefficients for the i-th sample (referred to as fi (i)). The example is given for one PSP patient (ID: PSP14). 2.5.6. Decision Support System In the UPDRS motor scale, Part III–Motor examination, task 3.4 Finger tapping [3], instructions for bradykinesia evaluation are given as follows: 0 Normal: Regular rhythm, without hesitations or freezes. Fast movement, large amplitude, no amplitude decrement. 1 Slight: Any of the following: (a) the regular rhythm is broken with one or two interruptions or hesitations of the tapping movement; (b) slight slowing; (c) the amplitude decrements near the end of the 10 taps. 2 Mild: Any of the following: (a) three to five interruptions during tapping; (b) mild slowing; (c) the amplitude decrements midway in the 10-tap sequence. 3 Moderate: Any of the following: (a) over five interruptions during tapping or at least one freeze in ongoing movement; (b) moderate slowing; (c) the amplitude decrements starting after the first tap. 4 Severe: Cannot or can only barely perform the task due to slowing, interruptions, or decrements. Each hand is evaluated separately, in terms of speed, amplitude, hesitation and freezes, and decrementing amplitude. These criteria are described with the introduced features, which are then fed to the decision support system. The input feature set includes the average tapping angle αav, Sensors 2019, 19, 2644 10 of 17 the average frequency f (i)av , the index of the first tap with a significant angle amplitude decrement idec, the number of hesitations Hnum, and the number of freezes Fnum. The rules are defined separately for each feature to give the subscores for each criterion, which are afterwards used for the calculation of the final score. As indicated, the lowest score corresponds to “normal” movements. Therefore, the first step is to find values that could be considered to be the reference for normal movements. The defined methodology was initially applied to a signal database from the control group that included a subset of healthy controls with no signs of bradykinesia (scored with 0). During the examination of the video files, it was noticed that both patients and healthy controls performed the tapping task in two different ways. In the first group, the subjects tapped as widely as possible at the highest speed that allows for such a tapping. In the other group, the subjects tapped with smaller amplitudes, but at their fastest pace. Using the parameters describing the tapping speed and amplitude (αav and f (i) av , respectively), the selected healthy controls were divided into two clusters using the k-means algorithm. The coordinates of the cluster centers were used as the measure for discriminating the two types of tapping performance. From all three groups of patients, we randomly selected 50% percent of the files and assigned them to the testing group. By calculating the distance between the center of the clusters and the data (αav, f (i) av ) obtained from the testing group, each patient was assigned to one of the two defined clusters (C1, “wider and slower”; C2, “narrower and faster”). The scores provided by the neurologists are given as the final score and do not provide information about different aspects (characteristics) of the performance that were analyzed. Because of that, it was necessary to apply an unsupervised learning algorithm to analyze properties of the features and find a natural grouping among the data. Testing data corresponding to one of the parameters (αav and f (i) av ) and one of the clusters (C1 or C2) were additionally divided into four clusters (corresponding to scores 0–3) using the k-means algorithm. Although there are five scores in the UPDRS test, the data were divided into four clusters, since the highest score (corresponding to the worst performance) is assigned to patients that barely perform the task (the movement is affected by multiple types of disturbances simultaneously). The coordinates of the cluster centers (c1, c2, c3, c4) were used for calculation of decision boundaries: bi = ci + ci+1, 2 ; i = 1, 2, 3 (1) where ci and ci+1 represent the centers of two neighbouring clusters and bi represents the calculated boundary separating the two scores. The procedure was repeated for both clusters C1 and C2, and for both parameters αav and f (i) av , separately (1: C1 and αav, 2: C2 and αav, 3: C1 and f (i) av , 4: C2 and f (i) av ), resulting in four sets of boundaries, each with three values. For each analyzed file, decision boundaries for the αav and f (i) av features were selected from those four sets. If this coordinate pair (αav, f (i) av ) is closer to the center of the cluster C1 than to the center of the cluster C2, then the patient’s file was assigned to the cluster C1 and decision boundaries (bα1,2,3 and b f 1,2,3) for cluster C1 were selected, and vice versa. Decision boundaries for the remaining features (idec and Hnum, Fnum) were set to match the instructions and criteria given within the UPDRS scale (as mentioned above). The block scheme of the decision support system is presented in Figure 7. The first part of the decision-making process is divided into four blocks (each bordered with dashed black line). The inputs of these blocks are the calculated features: αav, f (i) av , idec, and Hnum, Fnum, respectively. For each feature, a subscore is calculated separately, based on the range within which the feature value is located. In this way, the four processing blocks result in four subscores: Sα, S f , Sdec, and SHF, respectively. If the subscore “3–Moderate” is obtained for at least three out of four features, then the final score SFT is set to be “4–Severe”. Otherwise, the final score SFT is selected as the maximum obtained subscore among the four subscores corresponding to the individual features. Sensors 2019, 19, 2644 11 of 17 Sensors 2019, 19, x FOR PEER REVIEW 11 of 17 Figure 7. The block scheme of the decision support system. The system is divided into four processing blocks (bordered with dashed black rectangles). The inputs to the blocks are the calculated features: 𝛼 , 𝑓 ( ) 𝑖 , and 𝐻 , 𝐹 , respectively. Each block implements rules and assigns a subscore for the input feature. The final score 𝑆 is decided based on the results obtained from all four blocks. 2.5.7. Statistical Analysis and Evaluation To find the agreement between the scores obtained from two neurologists (raters), Cohen’s kappa statistics for finding intra-rater reliability among categorical data were applied. The results obtained from the decision support algorithm were compared with the scores given by the neurologists. The performance was measured using the confusion matrix and the accuracy of the proposed method, expressed as the percent of equally assigned scores. Initially, results were evaluated for all recordings (Case I), and later for the recordings equally scored by both raters (Case II). 3. Results The intra-rater reliability was calculated with the Cohen’s Kappa statistic and it equals to κ = 0.79, showing some discrepancy between raters’ scores. This result is expected, since the scores are provided based on their visual and subjective estimation. Overall, 87 recordings obtained from 44 patients (PD: 26 recordings, MSA: 34, PSP: 27) were included in the analysis, as well as 24 recordings obtained from 12 healthy controls. The descriptive statistics for the introduced features are given in Table 2 for each group of subjects separately. Table 2. Descriptive statistics (average ± st.deviation) for each feature and group of subjects. Group 𝒇𝒂𝒗(𝒊) (𝐇𝐳) 𝜶𝒂𝒗 (°) 𝒊𝒅𝒆𝒄 (#) 𝑯𝒏𝒖𝒎 (#) 𝑭𝒏𝒖𝒎 (#) PD 2.04 ± 0.87 63.08 ± 8.54 5.00 ± 5.66 0–4 0 MSA 1.71 ± 1.26 56.27 ± 36.11 4.03 ± 4.74 0–7 0–2 PSP 2.37 ± 1.11 44.87 ± 31.74 5.62 ± 4.88 0–4 0–1 HC 3.32 ± 0.89 80.48 ± 26.55 11.00 ± 10.99 / / Figure 7. The block scheme of the decision support system. The system is divided into four processing blocks (bordered with dashed black rectangles). The inputs to the blocks are the calculated features: αav, f (i) av idec, and Hnum, Fnum, respectively. Each block implements rules and assigns a subscore for the input feature. The final score SFT is decided based on the results obtained from all four blocks. 2.5.7. Statistical Analysis and Evaluation To find the agreement between the scores obtained from two neurologists (raters), Cohen’s kappa statistics for finding intra-rater reliability among categorical data were applied. The results obtained from the decision support algorithm were compared with the scores given by the neurologists. The performance was measured using the confusion matrix and the accuracy of the proposed method, expressed as the percent of equally assigned scores. Initially, results were evaluated for all recordings (Case I), and later for the recordings equally scored by both raters (Case II). 3. Results The intra-rater reliability was calculated with the Cohen’s Kappa statistic and it equals to κ = 0.79, showing some discrepancy between raters’ scores. This result is expected, since the scores are provided based on their visual and subjective estimation. Overall, 87 recordings obtained from 44 patients (PD: 26 recordings, MSA: 34, PSP: 27) were included in the analysis, as well as 24 recordings obtained from 12 healthy controls. The descriptive statistics for the introduced features are given in Table 2 for each group of subjects separately. Table 2. Descriptive statistics (average ± st.deviation) for each feature and group of subjects. Group f(i)av (Hz) αav ( ◦) idec (#) Hnum (#) Fnum (#) PD 2.04 ± 0.87 63.08 ± 8.54 5.00 ± 5.66 0–4 0 MSA 1.71 ± 1.26 56.27 ± 36.11 4.03 ± 4.74 0–7 0–2 PSP 2.37 ± 1.11 44.87 ± 31.74 5.62 ± 4.88 0–4 0–1 HC 3.32 ± 0.89 80.48 ± 26.55 11.00 ± 10.99 / / The highest values for the αav and f (i) av parameters were obtained for HC. Among patients, the PD group achieved the biggest angle amplitude values (on average); however, their tapping frequency Sensors 2019, 19, 2644 12 of 17 was found to be lower (on average) compared to the PSP group. This discrepancy shows that we need to discriminate between the two types of movements and, consequently, use two sets of decision boundaries for these two features. The feature describing the angle decrement (idec) was found to be comparable among patients. Although some HC also show a decrease in the angle amplitude, this is observed later in the tapping sequence (usually after the 10th tap). PD patients did not experience any freezing during the performance, whereas the number of hesitations was found to be comparable among groups. Among HC, none of the subjects experienced either hesitation or a freeze. Figure 8 shows the obtained angle αav and frequency f (i) av features versus the calculated scores. The performance clusters are shown using the color- and shape-coded representation. It can be seen that the αav and f (i) av features decrease with higher scores, which is in line with the criteria that is observed within the UPDRS. In addition, it can be confirmed that files assigned to the cluster C1 are characterized by larger angle values and a lower tapping frequency, whereas the cluster C2 includes files with a lower angle amplitude and a larger tapping frequency. Sensors 2019, 19, x FOR PEER REVIEW 12 of 17 The highest values for the 𝛼 and 𝑓 ( ) parameters were obtained for HC. Among patients, the PD group achieved the biggest angle amplitude values (on average); however, their tapping frequency was found to be lower (on average) compared to the PSP group. This discrepancy shows that we need to discriminate between the two types of movements and, consequently, use two sets of decision boundaries for these two features. The feature describing the angle decrement (𝑖 ) was found to be comparable among patients. Although some HC also show a decrease in the angle amplitude, this is observed later in the tapping sequence (usually after the 10th tap). PD patients did not experience any freezing during the performance, whereas the number of hesitations was found to be comparable among groups. Among HC, none of the subjects experienced either hesitation or a freeze. Figure 8 shows the obtained angle 𝛼 and frequency 𝑓 ( ) features versus the calculated scores. The performance clusters are shown using the color- and shape-coded representation. It can be seen that the 𝛼 and 𝑓 ( ) features decrease with higher scores, which is in line with the criteria that is observed within the UPDRS. In addition, it can be confirmed that files assigned to the cluster 𝐶 are characterized by larger angle values and a lower tapping frequency, whereas the cluster 𝐶 includes files with a lower angle amplitude and a larger tapping frequency. (a) (b) Figure 8. Dependency of: (a) calculated scores and the 𝛼 feature; (b) calculated scores the and 𝑓 ( ) feature. Grey circles mark samples that are assigned to the 𝐶 cluster (“wider and slower” performance), whereas black crosses correspond to members of the cluster 𝐶 (“narrower and faster” performance). The results of the expert system are presented in Table 3, for each group separately, as well as the summary for all patients. The results from the left column were obtained using all the recordings (Case I—87 recordings, PD: 26, MSA: 34, PSP: 27) and averaged for two raters, whereas the right column shows results obtained using only the recordings equally scored by both raters (Case II—76 recordings, PD: 25, MSA: 29, PSP: 22). Results are also presented in Figure 9 by a confusion matrix. Table 3. Results of the decision support system for each group of patients separately and in total. The result is provided for two cases: when all recordings are included in the analysis (Case I) and when only recordings with the same score from both raters are included in the analysis (Case II). Group Case I Accuracy (%) Case II Accuracy (%) PD 82.69 ± 2.72 84.00 MSA 82.36 ± 8.32 89.65 PSP 83.76 ± 7.86 90.91 TOTAL 83.33 ± 6.50 88.16 1 PD—Parkinson’s disease; MSA—Multiple system atrophy; PSP—Progressive supranuclear palsy. Figure 8. Dependency of: (a) calculated scores and the αav feature; (b) calculated scores the and f (i) av feature. Grey circles mark samples that are assigned to the C1 cluster (“wider and slower” performance), whereas black crosses correspond to members of the cluster C2 (“narrower and faster” performance). The results of the expert system are presented in Table 3, for each group separately, as well as the summary for all patients. The results from the left column were obtained using all the recordings (Case I—87 recordings, PD: 26, MSA: 34, PSP: 27) and averaged for two raters, whereas the right column shows results obtained using only the recordings equally scored by both raters (Case II—76 recordings, PD: 25, MSA: 29, PSP: 22). Results are also presented in Figure 9 by a confusion matrix. Table 3. Results of the decision support system for each group of patients separately and in total. The result is provided for two cases: when all recordings are included in the analysis (Case I) and when only recordings with the same score from both raters are included in the analysis (Case II). Group Case I Accuracy (%) Case II Accuracy (%) PD 82.69 ± 2.72 84.00 MSA 82.36 ± 8.32 89.65 PSP 83.76 ± 7.86 90.91 TOTAL 83.33 ± 6.50 88.16 1 PD—Parkinson’s disease; MSA—Multiple system atrophy; PSP—Progressive supranuclear palsy. Sensors 2019, 19, 2644 13 of 17 Sensors 2019, 19, x FOR PEER REVIEW 13 of 17 (a) (b) Figure 9. Presentation of results using the confusion matrix. (a) Case I—The result obtained when all recordings are included. (b) Case II—The result obtained using only recordings on which both raters agreed. The cells on the diagonal of the confusion matrix show the overall success rate for each score (expressed as a percentage (%)), whereas the cells outside the diagonal show the error rate for the scores (expressed as a percentage (%)). When all recordings are included in the analysis, comparable results are obtained for all three groups of patients. It is shown that the decision support system provides results that agree with the scores of the neurologists with a good accuracy (above 80%). This result is improved when only recordings equally scored by both raters are included in the analysis, achieving nearly 90% matching between the system results and neurologists’ estimates. The scores evaluated by the proposed decision system and the scores given by neurologists (Figure 9) do not exceed a one score difference, except for one patient. In Figure 10, we present the results of the expert system. The example is given for two patients who were equally scored by both raters and our expert system (score 𝑆 = 3). Figure 10. The result of the expert system comprising a graphical representation with detected irregularities, calculated features, and the final score. The example is given for one MSA patient (ID: MSA11), right hand, and one PSP patient (ID: PSP14), right hand. Figure 9. Presentation of results using the confusion matrix. (a) Case I—The result obtained when all recordings are included. (b) Case II—The result obtained using only recordings on which both raters agreed. The cells on the diagonal of the confusion matrix show the overall success rate for each score (expressed as a percentage (%)), whereas the cells outside the diagonal show the error rate for the scores (expressed as a percentage (%)). When all recordings are included in the analysis, comparable results are obtained for all three groups of patients. It is shown that the decision support system provides results that agree with the scores of the neurologists with a good accuracy (above 80%). This result is improved when only recordings equally scored by both raters are included in the analysis, achieving nearly 90% matching between the system results and neurologists’ estimates. The scores evaluated by the proposed decision system and the scores given by neurologists (Figure 9) do not exceed a one score difference, except for one patient. In Figure 10, we present the results of the expert system. The example is given for two patients who were equally scored by both raters and our expert system (score SFT = 3). Sensors 2019, 19, x FOR PEER REVIEW 13 of 17 (a) (b) Figure 9. Presentation of results using the confusion matrix. (a) Case I—The result obtained when all recordings are included. (b) Case II—The result obtained using only recordings on which both raters agreed. The cells on the diagonal of the confusion matrix show the overall success rate for each score (expressed as a percentage (%)), whereas the cells outside the diagonal show the error rate for the scores (expressed as a percentage (%)). When all recordings are included in the analysis, comparable results are obtained for all three groups of patients. It is shown that the decision support system provides results that agree with the scores of the neurologists with a good accuracy (above 80%). This result is improved when only recordings equally scored by both raters are included in the analysis, achieving nearly 90% matching between the system results and neurologists’ estimates. The scores evaluated by the proposed decision system and the scores given by neurologists (Figure 9) do not exceed a one score difference, except for one patient. In Figure 10, we present the results of the expert system. The example is given for two patients who were equally scored by both raters and our expert system (score 𝑆 = 3). Figure 10. The result of the expert system comprising a graphical representation with detected irregularities, calculated features, and the final score. The example is given for one MSA patient (ID: MSA11), right hand, and one PSP patient (ID: PSP14), right hand. Figure 10. The result of the expert system comprising a graphical representation with detected irregularities, calculated features, and the final score. The example is given for one MSA patient (ID: MSA11), right hand, and one PSP patient (ID: PSP14), right hand. 4. Discussion In this paper, we introduced a new methodology that enables objective evaluation and quantification of the finger-tapping test that is usually used for bradykinesia assessment in patients with Parkinson’s disease. The system comprises two miniature and lightweight gyro sensors that Sensors 2019, 19, 2644 14 of 17 record the motion of fingers. The methodology for signal quantification is based on the use of simple, automatized, and repeatable signal-processing techniques. In this study, 15 s long finger-tapping sequences were recorded and analyzed. However, the methodology is applicable to other approaches as well (e.g., a 10-tap-long sequence). Patients with finger-tapping bradykinesia severity ranging from 0 to 4 were included in this study. Most of the studies in the literature include severity stages up to 3, indicating that patients with the highest severity cannot perform specified tasks at all. In this study, we included three patients that barely managed to perform the finger-tapping task with one of their hands. However, their performance was poor and affected by multiple performance disturbances. Therefore, they were evaluated with the highest bradykinesia severity score (4). In this way, the performance of the proposed system was examined for the entire range of severity stages. The parametric result comprises features that can be directly correlated with biomechanical properties of the movement and can, therefore, be used to assist physicians during the assessment of bradykinesia in repetitive finger-tapping. For such a purpose, we implemented the time-frequency method continuous wavelet transform, which allows us to evaluate temporal changes in the tapping frequency and to detect irregularities in rhythmic tapping behaviour, such as hesitation and freezes. Numerical integration and drift removal were used for estimation of tapping angle apertures that are calculated for each tap. Temporal changes in the angle apertures were found and described by the index of the first tap with a significant amplitude decrement (compared to previously achieved tapping angle amplitudes). The final feature set is used as the input into the decision support system. Although the literature suggests that machine learning algorithms can predict scores with a high degree of accuracy, labels used for learning are given by physicians, which may cause subjectivity in the obtained results. Typically, only a few dozen recordings are used for learning, which is not a sufficient number of training examples to obtain a clinically acceptable system. Therefore, the decision support system consists of simple rules with decision boundaries designed to match the UPDRS scoring criteria. The decision boundaries for the tapping frequency and angle amplitude are defined according to the feature values obtained from the healthy controls and testing group of patients using the clustering techniques. As shown in Figure 8, smaller angle amplitudes and frequencies correspond to higher scores, which is in line with the criteria that are defined within the UPDRS. The boundaries differ, and they are selected based on the type of movement: wider and slower (cluster C1, grey circles in Figure 8) or narrower and faster (cluster C2, black crosses in Figure 8). In addition, using the clustering techniques, the boundaries are not defined linearly or empirically, but solely based on the grouping of some randomly selected testing data. The decision boundaries for two other features, i.e., the criteria, are defined to match the rules defined in the UPDRS. The results of the decision support system (Table 3) demonstrate that the expert system has achieved an overall accuracy of 83.33 ± 6.50% (averaged for both raters), whereas this result is 82.69 ± 2.72% for PD, 82.36 ± 8.32% for MSA, and 83.76 ± 7.86 for PSP patients. By analyzing the recordings that were evaluated with the same score by both raters, the overall accuracy of the system is increased, achieving 88.16%, whereas this result is 84.00% for PD, 89.65% for MSA, and 90.91% for PSP patients. In the latter case, the decision support system provides wrong scores for only nine recordings (out of 76 recordings). The decision support system provides very good results, even for atypical parkinsonism, in which finger-tapping can be performed differently compared to the typical PD form [27]. By using a larger number of patients, fine tuning of the decision boundaries can be performed, providing even better results. The differences in Table 2 between the obtained parameters indicate that this methodology can be used as the basis for differential diagnostics of typical and atypical parkinsonism. As shown in the confusion matrices in Figure 9, the quantification errors are equal to a one score difference, except for one PD patient, in which our system provides results that are two scores lower than the score from both raters. The question that arises is whether a scale with such a small resolution and only four grades of performance is sufficient. Figure 10 shows the results of the expert system for two patients evaluated by the same score (SFT = 3). These two performances are different: the first Sensors 2019, 19, 2644 15 of 17 one has a lower tapping frequency and higher angle apertures, with a significant decrease in the angle amplitude after the first tap. The second performance has large variations in frequency and angle amplitudes, as well as four hesitations and one freeze. For the first patient, the resulting score is given based on the early amplitude decrease, whereas, for the second patient, the score is given based on the number of performance irregularities. In the UPDRS instructions, it is said that a score is to be given if any of the criteria (speed, amplitude, amplitude decrease, hesitations/freezes) are satisfied. If different or more than one criterion is satisfied for different patients, they cannot be compared. Due to that fact, some researchers have introduced continuous scoring of repetitive hand motions that are used for bradykinesia evaluation. Although they provide a more detailed scoring system, this evaluation does not correspond to standardized clinical scores and may be confusing to physicians. The proposed system provides a complete analysis of repetitive finger-tapping performance, objective measures of important biomechanical properties of the movement, and a graphical presentation of the recorded data with specific changes and irregularities marked in the data. The system can differentiate between different types of performances and provides decision support through automatically calculated scores and subscores for different criteria for the evaluation of movements. The scores are given using rules that are specifically designed to match the universal clinical criteria for evaluation of bradykinesia severity in repetitive finger-tapping movements. In addition, the system was tested on patients with different forms of parkinsonism and at different disease stages, and for the full range of symptom severity (from normal to severely impaired movements). The proposed expert system is detailed and objective, and, therefore, it can be used as a powerful support tool in clinical practice for the evaluation of symptom severity, monitoring the disease’s progress and a patient’s response to therapy, and comparisons with other patients. In the future, an intuitive graphical interface for a software application will be developed to provide a graphical presentation, numerical results for features, scores and subscores, and a statistical analysis. Future work will also include collaborations with other researchers and groups to obtain larger databases and augment the data for analysis in terms of included subjects and the number of tests used to assess bradykinesia and other motor symptoms. We also plan to develop a metric to be used for more efficient differential diagnostics of typical and atypical parkinsonism. Author Contributions: V.B. participated in data acquisition, development of the proposed method, processing the results, and writing the manuscript. M.D.-J. participated in the design of the study, data acquisition, development of the proposed method, and writing the manuscript. N.D. participated in the design of the study, organization and coordination of data acquisition, testing and validation of the method, and writing the manuscript. M.B.P. participated in the design of the study and writing and reviewing the manuscript. V.S.K. participated in the design of the study and writing and reviewing the manuscript. G.K. participated in the development of the proposed method and writing and reviewing the manuscript. Funding: This research is in part funded by the Serbian Ministry of Education, Science and Technological Development under Grant No. OI-175016. Acknowledgments: We would like to acknowledge the PhD student Minja Belić for assisting with recordings. Conflicts of Interest: The authors declare no conflicts of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results. References 1. Rovini, E.; Maremmani, C.; Cavallo, F. How wearable sensors can support parkinson’s disease diagnosis and treatment: A systematic review. Front. Neurosci. 2017, 11, 555. [CrossRef] [PubMed] 2. Jankovic, J. Parkinson’s disease: Clinical features and diagnosis. J. Neurol. Neurosurg. Psychiatry 2008, 79, 368–376. [CrossRef] [PubMed] 3. MDS UPDRS Rating Scale. Available online: https://www.movementdisorders.org/MDS-Files1/PDFs/Rating- Scales/MDS-UPDRS_English_FINAL.pdf (accessed on 31 January 2019). http://dx.doi.org/10.3389/fnins.2017.00555 http://www.ncbi.nlm.nih.gov/pubmed/29056899 http://dx.doi.org/10.1136/jnnp.2007.131045 http://www.ncbi.nlm.nih.gov/pubmed/18344392 https://www.movementdisorders.org/MDS-Files1/PDFs/Rating-Scales/MDS-UPDRS_English_FINAL.pdf https://www.movementdisorders.org/MDS-Files1/PDFs/Rating-Scales/MDS-UPDRS_English_FINAL.pdf Sensors 2019, 19, 2644 16 of 17 4. Lin, Z.; Xiong, Y.; Cai, G.; Dai, H.; Xia, X.; Tan, Y.; Lueth, T.C. Quantification of Parkinsonian Bradykinesia Based on Axis-Angle Representation and SVM Multiclass Classification Method. IEEE Access 2018, 6, 26895–26903. [CrossRef] 5. Lainscsek, C.; Rowat, P.; Schettino, L.; Lee, D.; Song, D.; Letellier, C.; Poizner, H. Finger tapping movements of Parkinson’s disease patients automatically rated using nonlinear delay differential equations. Chaos An Interdiscip. J. Nonlinear Sci. 2012, 22, 013119. [CrossRef] [PubMed] 6. Djuric-Jovicic, M.; Jovicic, N.; Radovanovic, S.; Jecmenica-Lukic, M.; Belic, M.; Popovic, M.; Kostic, V. Finger and foot tapping sensor system for objective motor assessment. Vojnosanit. Pregl. 2018. [CrossRef] 7. Kim, J.-W.; Lee, J.-H.; Kwon, Y.; Kim, C.-S.; Eom, G.-M.; Koh, S.-B.; Kwon, D.-Y.; Park, K.-W. Quantification of bradykinesia during clinical finger taps using a gyrosensor in patients with Parkinson’s disease. Med. Biol. Eng. Comput. 2011, 49, 365–371. [CrossRef] [PubMed] 8. Stamatakis, J.; Ambroise, J.; Crémers, J.; Sharei, H.; Delvaux, V.; Macq, B.; Garraux, G. Finger Tapping Clinimetric Score Prediction in Parkinson’s Disease Using Low-Cost Accelerometers. Comput. Intell. Neurosci. 2013, 2013, 1–13. [CrossRef] [PubMed] 9. Garza-Rodríguez, A.; Sánchez-Fernández, L.P.; Sánchez-Pérez, L.A.; Ornelas-Vences, C.; Ehrenberg-Inzunza, M. Pronation and supination analysis based on biomechanical signals from Parkinson’s disease patients. Artif. Intell. Med. 2018, 84, 7–22. [CrossRef] [PubMed] 10. Piro, N.; Piro, L.; Kassubek, J.; Blechschmidt-Trapp, R.; Piro, N.E.; Piro, L.K.; Kassubek, J.; Blechschmidt-Trapp, R.A. Analysis and Visualization of 3D Motion Data for UPDRS Rating of Patients with Parkinson’s Disease. Sensors 2016, 16, 930. [CrossRef] 11. Ferraris, C.; Nerino, R.; Chimienti, A.; Pettiti, G.; Cau, N.; Cimolin, V.; Azzaro, C.; Albani, G.; Priano, L.; Mauro, A.; et al. A Self-Managed System for Automated Assessment of UPDRS Upper Limb Tasks in Parkinson’s Disease. Sensors 2018, 18, 3523. [CrossRef] 12. Lee, W.L.; Sinclair, N.C.; Jones, M.; Tan, J.L.; Proud, E.L.; Peppard, R.; McDermott, H.J.; Perera, T. Objective evaluation of bradykinesia in Parkinson’s disease using an inexpensive marker-less motion tracking system. Physiol. Meas. 2019, 40, 014004. [CrossRef] [PubMed] 13. Sano, Y.; Kandori, A.; Shima, K.; Yamaguchi, Y.; Tsuji, T.; Noda, M.; Higashikawa, F.; Yokoe, M.; Sakoda, S. Quantifying Parkinson’s disease finger-tapping severity by extracting and synthesizing finger motion properties. Med. Biol. Eng. Comput. 2016, 54, 953–965. [CrossRef] [PubMed] 14. Arora, S.; Venkataraman, V.; Zhan, A.; Donohue, S.; Biglan, K.M.M.; Dorsey, E.R.R.; Little, M.A.A. Detecting and monitoring the symptoms of Parkinson’s disease using smartphones: A pilot study. Park. Relat. Disord. 2015, 21, 650–653. [CrossRef] [PubMed] 15. Van den Noort, J.C.; Verhagen, R.; van Dijk, K.J.; Veltink, P.H.; Vos, M.C.P.M.; de Bie, R.M.A.; Bour, L.J.; Heida, C.T. Quantification of Hand Motor Symptoms in Parkinson’s Disease: A Proof-of-Principle Study Using Inertial and Force Sensors. Ann. Biomed. Eng. 2017, 45, 2423–2436. [CrossRef] [PubMed] 16. Lin, Z.; Dai, H.; Xiong, Y.; Xia, X.; Horng, S.-J. Quantification assessment of bradykinesia in Parkinson’s disease based on a wearable device. In Proceedings of the 2017 39th IEEE Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Jeju Island, Korea, 11–15 July 2017; pp. 803–806. 17. Dai, H.; Lin, H.; Lueth, T.C. Quantitative assessment of parkinsonian bradykinesia based on an inertial measurement unit. Biomed. Eng. Online 2015, 14, 68. [CrossRef] [PubMed] 18. McKay, G.N.; Harrigan, T.P.; Brašić, J.R. A low-cost quantitative continuous measurement of movements in the extremities of people with Parkinson’s disease. MethodsX 2019, 6, 169–189. [CrossRef] [PubMed] 19. Alam, M.; Tabassum, T.; Munia, K.; Tavakolian, K. A Quantitative Assessment of Bradykinesia Using Inertial Measurement Unit Performance Measurement View project Signal-Image Processing View project. In Proceedings of the 2017 Design of Medical Devices Conference, Minneapolis, MN, USA, 10–13 April 2017. 20. Patel, S.; Lorincz, K.; Hughes, R.; Huggins, N.; Growdon, J.; Standaert, D.; Akay, M.; Dy, J.; Welsh, M.; Bonato, P. Monitoring motor fluctuations in patients with parkinsons disease using wearable sensors. IEEE Trans. Inf. Technol. Biomed. 2009, 13, 864–873. [CrossRef] 21. Djurić-Jovičić, M.; Petrović, I.; Ječmenica-Lukić, M.; Radovanović, S.; Dragašević-Mišković, N.; Belić, M.; Miler-Jerković, V.; Popović, M.B.; Kostić, V.S. Finger tapping analysis in patients with Parkinson’s disease and atypical parkinsonism. J. Clin. Neurosci. 2016, 30, 49–55. [CrossRef] http://dx.doi.org/10.1109/ACCESS.2018.2835463 http://dx.doi.org/10.1063/1.3683444 http://www.ncbi.nlm.nih.gov/pubmed/22462995 http://dx.doi.org/10.2298/VSP150502323D http://dx.doi.org/10.1007/s11517-010-0697-8 http://www.ncbi.nlm.nih.gov/pubmed/21052856 http://dx.doi.org/10.1155/2013/717853 http://www.ncbi.nlm.nih.gov/pubmed/23690760 http://dx.doi.org/10.1016/j.artmed.2017.10.001 http://www.ncbi.nlm.nih.gov/pubmed/29042162 http://dx.doi.org/10.3390/s16060930 http://dx.doi.org/10.3390/s18103523 http://dx.doi.org/10.1088/1361-6579/aafef2 http://www.ncbi.nlm.nih.gov/pubmed/30650391 http://dx.doi.org/10.1007/s11517-016-1467-z http://www.ncbi.nlm.nih.gov/pubmed/27032933 http://dx.doi.org/10.1016/j.parkreldis.2015.02.026 http://www.ncbi.nlm.nih.gov/pubmed/25819808 http://dx.doi.org/10.1007/s10439-017-1881-x http://www.ncbi.nlm.nih.gov/pubmed/28726022 http://dx.doi.org/10.1186/s12938-015-0067-8 http://www.ncbi.nlm.nih.gov/pubmed/26164814 http://dx.doi.org/10.1016/j.mex.2018.12.017 http://www.ncbi.nlm.nih.gov/pubmed/30733930 http://dx.doi.org/10.1109/TITB.2009.2033471 http://dx.doi.org/10.1016/j.jocn.2015.10.053 Sensors 2019, 19, 2644 17 of 17 22. Delrobaei, M.; Tran, S.; Gilmore, G.; McIsaac, K.; Jog, M. Characterization of multi-joint upper limb movements in a single task to assess bradykinesia. J. Neurol. Sci. 2016, 368, 337–342. [CrossRef] 23. Ornelas-Vences, C.; Sánchez-Fernández, L.P.; Sánchez-Pérez, L.A.; Martínez-Hernández, J.M. Computer model for leg agility quantification and assessment for Parkinson’s disease patients. Med. Biol. Eng. Comput. 2019, 57, 463–476. [CrossRef] 24. Mentzel, T.Q.; Lieverse, R.; Levens, A.; Mentzel, C.L.; Tenback, D.E.; Bakker, P.R.; Daanen, H.A.M.; van Harten, P.N. Reliability and validity of an instrument for the assessment of bradykinesia. Psychiatry Res. 2016, 238, 189–195. [CrossRef] [PubMed] 25. Memar, S.; Delrobaei, M.; Pieterman, M.; McIsaac, K.; Jog, M. Quantification of whole-body bradykinesia in Parkinson’s disease participants using multiple inertial sensors. J. Neurol. Sci. 2018, 387, 157–165. [CrossRef] [PubMed] 26. Kinesia ONETM. Available online: https://glneurotech.com/kinesia/products/kinesia-one/ (accessed on 13 May 2019). 27. Djurić-Jovičić, M.; Jovičić, N.; Roby-Brami, A.; Popović, M.; Kostić, V.; Djordjević, A.; Djurić-Jovičić, M.; Jovičić, N.S.; Roby-Brami, A.; Popović, M.B.; et al. Quantification of Finger-Tapping Angle Based on Wearable Sensors. Sensors 2017, 17, 203. [CrossRef] [PubMed] 28. Bobic, V.; Djuric-Jovicic, M.; Jarrasse, N.; Jecmenica-Lukic, M.; Petrovic, I.; Radovanovic, S.; Dragasevic, N.; Kostic, V. Spectral parameters for finger tapping quantification. Facta Univ.-Ser. Electron. Energ. 2017, 30, 585–597. [CrossRef] 29. Senhadji, L.; Wendling, F. Epileptic transient detection: Wavelets and time-frequency approaches. Neurophysiol. Clin. Neurophysiol. 2002, 32, 175–192. [CrossRef] © 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/). http://dx.doi.org/10.1016/j.jns.2016.07.056 http://dx.doi.org/10.1007/s11517-018-1894-0 http://dx.doi.org/10.1016/j.psychres.2016.02.011 http://www.ncbi.nlm.nih.gov/pubmed/27086232 http://dx.doi.org/10.1016/j.jns.2018.02.001 http://www.ncbi.nlm.nih.gov/pubmed/29571855 https://glneurotech.com/kinesia/products/kinesia-one/ http://dx.doi.org/10.3390/s17020203 http://www.ncbi.nlm.nih.gov/pubmed/28125051 http://dx.doi.org/10.2298/FUEE1704585B http://dx.doi.org/10.1016/S0987-7053(02)00304-0 http://creativecommons.org/ http://creativecommons.org/licenses/by/4.0/. Introduction Materials and Methods Measurement System Subjects Measurement Methodology Scoring by Neurologists Data Processing and Analysis Individual Taps Amplitude Amplitude Decrement Hesitations and Freezes Speed Decision Support System Statistical Analysis and Evaluation Results Discussion References 
work_3b2a2vjlanay3ksoid7xvudtrq	----	Piecewise linear value functions for multi-criteria decision-making Delft University of Technology Piecewise linear value functions for multi-criteria decision-making Rezaei, Jafar DOI 10.1016/j.eswa.2018.01.004 Publication date 2018 Document Version Final published version Published in Expert Systems with Applications Citation (APA) Rezaei, J. (2018). Piecewise linear value functions for multi-criteria decision-making. Expert Systems with Applications, 98, 43-56. https://doi.org/10.1016/j.eswa.2018.01.004 Important note To cite this publication, please use the final published version (if applicable). Please check the document version above. Copyright Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons. Takedown policy Please contact us and provide details if you believe this document breaches copyrights. We will remove access to the work immediately and investigate your claim. This work is downloaded from Delft University of Technology. For technical reasons the number of authors shown on this cover page is limited to a maximum of 10. https://doi.org/10.1016/j.eswa.2018.01.004 https://doi.org/10.1016/j.eswa.2018.01.004 Green Open Access added to TU Delft Institutional Repository ‘You share, we take care!’ – Taverne project https://www.openaccess.nl/en/you-share-we-take-care Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public. https://www.openaccess.nl/en/you-share-we-take-care Expert Systems With Applications 98 (2018) 43–56 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Piecewise linear value functions for multi-criteria decision-making Jafar Rezaei Faculty of Technology Policy and Management, Delft University of Technology, Delft, The Netherlands a r t i c l e i n f o Article history: Received 24 August 2017 Revised 1 January 2018 Accepted 2 January 2018 Available online 3 January 2018 Keywords: Multi-criteria decision-making MCDM Decision criteria Value function Monotonicity a b s t r a c t Multi-criteria decision-making (MCDM) concerns selecting, ranking or sorting a set of alternatives which are evaluated with respect to a number of criteria. There are several MCDM methods, the two core el- ements of which are (i) evaluating the performance of the alternatives with respect to the criteria, (ii) finding the importance (weight) of the criteria. There are several methods to find the weights of the cri- teria, however, when it comes to the alternative measures with respect to the criteria, usually the existing MCDM methods use simple monotonic linear value functions. Usually an increasing or decreasing linear function is assumed between a criterion level (over its entire range) and its value. This assumption, how- ever, might lead to improper results. This study proposes a family of piecewise value functions which can be used for different decision criteria for different decision problems. Several real-world examples from existing literature are provided to illustrate the applicability of the proposed value functions. A numerical example of supplier selection (including a comparison between simple monotonic linear value functions, piecewise linear value functions, and exponential value functions) shows how considering proper value functions could affect the final results of an MCDM problem. © 2018 Elsevier Ltd. All rights reserved. 1 d t i h p p f w a f l U w a l o t a c a U d t w d s t t t l ( c D o C o n h t o t i D h 0 . Introduction Decision theory is primarily concerned with identifying the best ecision. In many real-world situations the decision is to select he best alternative(s) from among a set of alternatives consider- ng a set of criteria. This subdivision of decision-making, which as gained enormous attention, due to its practical value, in the ast recent is called multi-criteria decision-making (MCDM). More recisely, MCDM concerns problems in which the decision-maker aces m alternatives ( a 1 , a 2 , …, a m ), which should be evaluated ith respect to n criteria ( c 1 , c 2 , …, c n ), in order to find the best lternative(s), rank or sort them. In most cases, an additive value unction is used to find the overall value of alternative i, U i , as fol- ows: i = n ∑ j=1 w j u i j , (1) here u ij is the value of alternative i with respect to criterion j , nd w j shows the importance (weight) of criterion j . In some prob- ems, the decision-maker is able to find u ij from external sources as bjective measures, in some other problems, u ij reflects a qualita- ive evaluation provided by the decision-maker(s), experts or users s subjective measures. Price of a car is an objective criterion while omfort of a car is a subjective one. For objective criteria, we usu- E-mail address: j.rezaei@tudelft.nl t [ ttps://doi.org/10.1016/j.eswa.2018.01.004 957-4174/© 2018 Elsevier Ltd. All rights reserved. lly use physical quantities, for instance, ‘International System of nits’ (SI), while for subjective criteria, we do not have such stan- ards, which is why we mostly use pairwise comparison, linguis- ic variables, or Likert scales in order to evaluate the alternatives ith regard to such criteria. In order to find the weights, w j , the ecision-maker might use different tools and methods, from the implest way, which is assigning weights to the criteria intuitively, o use simple methods like SMART (simple multi-attribute rating echnique) ( Edwards, 1977 ), to more structured methods like mul- iple attribute utility theory (MAUT) ( Keeney & Raiffa, 1976 ), ana- ytic hierarchy process (AHP) ( Saaty, 1977 ), and best worst method BWM) ( Rezaei, 2015, 2016 ). While these methods are usually alled ‘multi attribute utility and value theories’ ( Carrico, Hogan, yson, & Athanassopoulos, 1997 ), there is another class of meth- ds, called outranking methods, like ELECTRE (ELimination and hoice Expressing REality) family ( Roy, 1968 ), PROMETHEE meth- ds ( Brans, Mareschal, & Vincke, 1984 ) which do not necessarily eed the weights to select, rank or sort the alternatives. What, owever, is in common in these methods is the way they consider he nature of the criteria. That is to say, in the current literature, ne of the common assumptions about the criteria (most of the ime it is not explicitly mentioned in the literature), is monotonic- ty. efinition 1 ( Keeney & Raiffa, 1976 ). Let u represents a value func- ion for criterion X , then u is monotonically increasing if: x 1 > x 2 ] ⇔ [ u ( x 1 ) > u ( x 2 ) ] . (2) https://doi.org/10.1016/j.eswa.2018.01.004 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2018.01.004&domain=pdf mailto:j.rezaei@tudelft.nl https://doi.org/10.1016/j.eswa.2018.01.004 44 J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 V Fig. 1. Increasing value function. t a w b s v t c f 2 fi t s w d s s a d t 2 f r i u less efficient fuel energy. 1 It is worth-mentioning that the studies we discuss to support each value func- tion have some theoretical or practical support for the proposed value functions. It does not, however, mean that those studies have used these value functions in their analysis. Definition 2 ( Keeney & Raiffa, 1976 ). Let u represents a value func- tion for criterion X , then u is monotonically decreasing if: [ x 1 > x 2 ] ⇔ [ u ( x 1 ) < u ( x 2 ) ] . (3) A function which is not monotonic is called non-monotonic and may have different shapes. For instance, a value function with the first part increasing and the second part decreasing called non- monotonic, by splitting of which, we have two monotonic func- tions. This assumption – monotonicity – however, is an oversimpli- fication in some real-world decision-making problems. Another simplification is the use of simple linear functions over the en- tire range of a criterion. Considering the two assumptions (mono- tonicity, linearity), we usually see simple increasing and decreas- ing linear value functions for the decision criteria in MCDM prob- lems. The literature is full of such applications. For instance, many of the studies reviewed in the following review papers implic- itly adopt such assumptions: the MCDM applications in supplier selection ( Ho, Xu, & Dey, 2010 ), in infrastructure management ( Kabir, Sadiq, & Tesfamariam, 2014 ), in sustainable energy plan- ning ( Pohekar & Ramachandran, 2004 ), and in forest management and planning ( Ananda & Herath, 2009 ). While in some studies the use of monotonic and/or linear value function might be logical, their use in some other applications might be unfitting. For in- stance, Alanne, Salo, Saari, and Gustafsson (2007) , for evaluation of residential energy supply systems use monotonic-linear value functions for all the selected evaluation criteria including “global warming potential (kg CO 2 m −2 a −1 )”, and “acidification potential (kg SO 2 m −2 a −1 )”. Considering a monotonic-linear value function for such criteria implies that the decision-maker accepts any level of such harmful environmental criteria for an energy supply sys- tem. However, if the decision-maker does not accept some high levels of such criteria (which seems logical), a piecewise linear function might better represent the preferences of the decision- maker (see the decrease-level value function in the next section). Some authors have discussed nonlinear monotonic value functions (e.g., exponential value functions by Kirkwood, 1997; Pratt, 1964 ). Others use qualitative scoring to address the non- monotonicity ( Brugha, 20 0 0; Kakeneno & Brugha, 2017; O’Brien & Brugha, 2010 ). We can also find some forms of eliciting piecewise linear value function in Jacquet-Lagreze and Siskos (2001 ), and Stewart and Janssen (2013 ). Some other value function construc- tion or elicitation frameworks can be found in Herrera, Herrera- iedma, and Verdegay (1996 ), Lahdelma and Salminen (2012 ), Mustajoki and Hämäläinen (20 0 0 ), Stewart and Janssen (2013 ), and Yager (1988 ). Although in PROMETHEE we use different types of piecewise functions for pairwise comparisons ( Brans, Mareschal, & Vincke, 1984 ), the functions are not used to evaluate the decision criteria. So, despite some effort s in literature, there is no a library of some standard piecewise linear value functions which can be used in different methods like AHP or BWM. It is also important to note that while in many studies value functions are elicited ac- cording to the preference data we have from the decision-maker(s), in MCDM, usually we use the value function as a subjective input. This implies that, in MCDM methods (except a few methods, such as UTA), the value function is not elicited, but an approximation is used. This also suggests that the rich literature on determining and eliciting value functions is not actually helping MCDM methods in this area. In this paper, first, a number of piecewise linear value functions with different shapes are proposed to be considered for the decision criteria. It is then shown, with some real-world ex- amples, how such consideration might change the final results of a decision problem. A comparison between simple monotonic linear value functions, piecewise linear value functions, and exponential value functions is conducted, which shows the effectiveness of the proposed pricewise value functions. This is a significant contribu- ion to this field and it is expected to be widely used by MCDM pplications. In the next section, some piecewise linear value functions along ith some real-world examples are presented, which is followed y some remarks in Section 3 . In Section 4 , some numerical analy- es are used to show the applicability of considering the proposed alue functions in a decision problem. In Section 5 , the determina- ion of the value functions is discussed. In Section 6 , the paper is oncluded, some limitations of the study are discussed, and some uture research directions are proposed. . Piecewise linear value functions In this section, a number of piecewise value functions are de- ned for decision criteria. We provide some example cases from he existing literature or practical decision-making problems to upport 1 each value function. In all the following value functions e consider [ d l j , d u j ] as the defined domain for the criterion by the ecision-maker; x ij shows the performance of alternative i with re- pect to criterion j ; and u ij shows the value of alternative i with re- pect to criterion j . For instance, if a decision-maker wants to buy car considering price as one criterion, if all the alternatives the ecision-maker considers are between €17,0 0 0 and €25,0 0 0, then he criterion might be defined for this range [17,0 0 0, 25,0 0 0] . .1. Increasing Increasing value function is perhaps the most commonly used unction in MCDM applications. It basically shows that as the crite- ion level, x ij , increases, its value, u ij , increases as well. It is shown n Fig. 1 and formulated as follows: i j = ⎧ ⎨ ⎩ x i j − d l j d u j − d l j , d l j ≤ x i j ≤ d u j , 0 , otherwise . (4) For this function we can think of: • Product quality in supplier selection ( Xia & Wu, 2007 ). Con- sidering a set of suppliers, a buyer may always prefer a sup- plier with a higher product quality compared to a supplier with lower product quality. • Energy efficiency in alternative-fuel bus selection ( Tzeng, Lin, & Opricovic, 2005 ). Considering a set of buses, a bus with more efficient fuel energy might always be preferred to a bus with J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 45 Fig. 2. Decreasing value function. Fig. 3. V-shape value function. 2 i l u 2 c a i u r t Fig. 4. Inverted V-shape value function. 2 x a F u 2 i t l u .2. Decreasing Decreasing value function shows that as the criterion level, x ij , ncreases, its value, u ij , decreases. It is shown in Fig. 2 and formu- ated as follows: i j = ⎧ ⎨ ⎩ d u j − x i j d u j − d l j , d l j ≤ x i j ≤ d u j , 0 , otherwise . (5) For this function we can think of: • Product price in supplier selection ( Xia & Wu, 2007 ). Consid- ering a set of suppliers, a supplier with a lower product price might always be preferred to a supplier with higher product price. So, a higher product price has a lower value. • Maintenance cost in alternative-fuel bus selection ( Tzeng et al., 2005 ). Considering a set of buses, a bus with less maintenance cost might be preferred to a bus with higher maintenance cost. So, a higher maintenance cost is associated with a lower value. .3. V-shape V-shape value function shows that as the criterion level, x ij , in- reases up to a certain level, d m j , its value, u ij , decreases gradually, nd after that certain level, d m j , its value, u ij , increases gradually. It s shown in Fig. 3 and formulated as follows: i j = ⎧ ⎪ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎪ ⎩ d m j − x i j d m j − d l j , d l j ≤ x i j ≤ d m j , x i j − d m j d u j − d m j , d m j ≤ x i j ≤ d u j , 0 , otherwise . (6) For this function, we could not find many examples, and it may epresent a small number of very particular decision criteria. For his function we can think of: • Relative market share in selecting a firm for investment ( Wilson & Anell, 1999 ). Wilson and Anell (1999) found that for invest- ment decision-making, firms with low and high market share are more desirable to the investors. This implies that the value of a firm decreases while its market share increases up to a cer- tain level, d m , and after that its value increases again. • Firm size in R&D productivity ( Tsai & Wang, 2005 ). Tsai and Wang (2005) found that both small and large firms have higher R&D productivity compared to medium-sized firms. This is true for both high-tech and traditional industries. This means that the relationship between size and value (measured by R&D pro- ductivity) is V-shape with a minimum level of value assigned to a certain size of d m j . .4. Inverted V-shape Inverted V-shape value function shows that as the criterion level, ij , increases up to a certain level, d m j , its value, u ij , increases, and fter that certain level, d m j , its value, u ij , decreases. It is shown in ig. 4 and formulated as follows: i j = ⎧ ⎪ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎪ ⎩ x i j − d l j d m j − d l j , d l j ≤ x i j ≤ d m j , d u j − x i j d u j − d m j , d m j ≤ x i j ≤ d u j , 0 , otherwise . (7) For this function we can think of: • Commute time in selecting a job ( Redmond & Mokhtar- ian, 2001 ). For many people, the ideal commute, d m j , is larger than zero. This implies that commute times between zero and the optimal commute time, and between the optimal commute time and larger times, have lower value than the optimal com- mute time for such individuals. This suggests an inverted V- shape value function. • Cognitive proximity in innovation partner selection ( Nooteboom, 20 0 0 ). For a company there is an optimal cognitive distance to the partner they are working on innova- tion ( d m j ). This implies that any distance less than d m j or larger than d m j has less value. .5. Increase-level Increase-level value function shows that as the criterion level, x ij , ncreases up to a certain level, d m j , its value, u ij , increases, and after hat certain level, d m j , its value, u ij , will remain at the maximum evel. It is shown in Fig. 5 and formulated as follows: i j = ⎧ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎩ x i j − d l j d m j − d l j , d l j ≤ x i j ≤ d m j , 1 , d m j ≤ x i j ≤ d u j , 0 , otherwise . (8) For this function we can think of: 46 J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 Fig. 5. Increase-level value function. Fig. 6. Level-decrease value function. Fig. 7. Level-increase value function. Fig. 8. Decrease-level value function. 2 x i I u 2 x a I u • Fill rate in supplier selection ( Chae, 2009 ). Although a buyer prefers suppliers with higher fill rate, which implies that as the fill rate increases its value increases, the buyer might be indif- ferent to any increase after a certain level, d m j , as usually buy- ers pre-identify a desirable service level which is satisfied by a certain minimum level of supplier’s fill rate. • Diversity of restaurants in hotel location selection ( Chou, Hsu, & Chen, 2008 ). In order to find the best location for an international hotel, a decision-maker prefers locations with more divers restaurants. However, reaching a level, d m j , might fully satisfy a decision-maker implying that the decision-maker might not be sensitive to any increase after that certain level. 2.6. Level-decrease Level-decrease value function shows that as the criterion level, x ij , increases up to a certain level, d m j , its value, u ij , remains at max- imum level, and after that certain level, d m j , its value, u ij , decreases gradually. It is shown in Fig. 6 and formulated as follows: u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 1 , d l j ≤ x i j ≤ d m j , d u j − x i j d u j − d m j , d m j ≤ x i j ≤ d u j , 0 , otherwise . (9) For this function we can think of: • Distance in selecting a university ( Carrico et al., 1997 ). While a student prefers a closer university to a farther university, this preference might start after a certain distance, d m j , implying that any distance between [ d l j , d m j ] is optimal and indifferent for the student. • Lead time in supplier selection ( Çebi & Otay, 2016 ). Although a supplier with a shorter lead time is preferred, if the lead time is in a limit such that it does not negatively affect the com- pany’s production, the company might then be indifferent to that range. .7. Level-increase Level-increase value function shows that as the criterion level, ij , increases up to a certain level, d m j , its value, u ij , remains at min- mum level, and after that certain level, d m j , its value, u ij , increases. t is shown in Fig. 7 and formulated as follows: i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 0 , d l j ≤ x i j ≤ d m j , x i j − d m j d u j − d m j , d m j ≤ x i j ≤ d u j , 0 , otherwise . (10) For this function we can think of: • Level of trust in making a buyer–supplier relationship ( Ploetner & Ehret, 2006 ). Trust increases the level of partnership between a buyer and a supplier, however it is only effective after a cer- tain threshold, d m j . • Level of relational satisfaction in evaluating quality commu- nication in marriage ( Montgomery, 1981 ). Below a minimum level of relational satisfaction, d m j , quality communication can- not take place thus results in minimum value. .8. Decrease-level Decrease-level value function shows that as the criterion level, ij , increases up to a certain level, d m j , its value, u ij , decreases, and fter that certain level d m j , its value, u ij , will remain at minimum. t is shown in Fig. 8 and formulated as follows: i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ d m j − x i j d m j − d l j , d l j ≤ x i j ≤ d m j , 0 , d m j ≤ x i j ≤ d u j , 0 , otherwise . (11) For this function we can think of: • Carbon emission in transportation mode selection ( Hoen, Tan, Fransoo, & van Houtum, 2014 ). In selecting a transportation mode, the more the carbon emission by the mode, the less the value of that mode. A decision-maker, however might assign J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 47 Fig. 9. Increasing stepwise value function. 2 l a j I u w h t t s r v u w Fig. 10. Decreasing stepwise value function. 2 l a u s u w m c u w 3 s u R m e d i c b zero value to a mode with carbon emission higher than a cer- tain level, d m j . • Distance when selecting a school ( Frenette, 2004 ). It has been shown that the longer the distance to the school the less the preference to attend that school. It is also clear that, for some people, there is no value after a certain distance, d m j . .9. Increasing stepwise Increasing stepwise value function shows that as the criterion evel, x ij , increases up to a certain level, d m j , its value remains at certain level, u 0 , and after that certain level, d m j , its value, u ij , umps to a higher level (maximum) and remains at the maximum. t is shown in Fig. 9 and formulated as follows: i j = ⎧ ⎨ ⎩ u 0 , d l j ≤ x i j ≤ d m j , 1 , d m j ≤ x i j ≤ d u j , 0 , otherwise . (12) here 0 < u 0 < 1. For this function we can think of: • Suppliers capabilities in supplier segmentation ( Rezaei & Ortt, 2012 ). Suppliers of a company are evaluated based on their capabilities, and then segmented based on two levels (low and high) with respect to their capabilities. As such a supplier scored between d l j and d m j is considered as a low-level capa- bilities supplier while a supplier scored between d m j and d u j is considered as a high-level capabilities supplier. • Symmetry in selecting a close type of partnership ( Lambert, Emmelhainz, & Gardner, 1996 ). In order to have a successful relationship between supply chain partners, there should be some demographical similarities (for instance, in terms of brand image, productivity) between them. So, more symmetry means closer relationship. However, if we consider two levels of closeness, it is clear that for some level of symmetry the value of closeness remains the same. For the increasing stepwise value function, a criterion might ave more than one jump. For instance, if a decision-maker wants o consider three levels low, medium, and high when segmenting he suppliers with respect to their capabilities, then an increasing tepwise function with two jumps should be defined for this crite- ion. The following value function is a general increasing stepwise alue function with k jumps. i j = ⎧ ⎪ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎪ ⎩ u 0 , d l j ≤ x i j ≤ d m 1 j , u 1 , d m 1 j ≤ x i j ≤ d m 2 j , . . . 1 , d mk j ≤ x i j ≤ d u j , 0 , otherwise . (13) here 0 < u < u < … < 1. 0 1 .10. Decreasing stepwise Decreasing stepwise value function shows that as the criterion evel, x ij , increases up to a certain level, d m j , its value, u ij , remains t a the maximum level, and after that certain level, d m j , its value, ij , jumps down to a lower level, u 0 , and remains at that level. It is hown in Fig. 10 and formulated as follows: i j = ⎧ ⎨ ⎩ 1 , d l j ≤ x i j ≤ d m j , u 0 , d m j ≤ x i j ≤ d u j , 0 , otherwise . (14) here 0 < u 0 < 1. For this function we can think of: • Considering supply risk in portfolio modeling ( Kraljic, 1983 ). For a company, an item with a higher level of risk results in less value, however, due to portfolio modeling, there is no dif- ference between all levels of risk in the domain [ d l j , d m j ] . Simi- larly, all levels of risk in the domain [ d m j , d u j ] result in the same value. • Delay in logistics service provider selection ( Qi, 2015 ). Some companies consider stepwise value function for delay in de- livering the items by a logistics service provider, which means that the value of that provider decreases when delay increases, however it is constant within certain intervals. For the decreasing stepwise function, a criterion might have ore than one jump. The following value function is a general de- reasing stepwise function with k jumps. i j = ⎧ ⎪ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎪ ⎩ 1 , d l j ≤ x i j ≤ d m 1 j , u 1 , d m 1 j ≤ x i j ≤ d m 2 j , . . . u k , d mk j ≤ x i j ≤ d u j , 0 , otherwise . (15) here 0 < u k < … < u 1 < 1. . Some remarks on the value functions Here, a number of remarks are discussed, shedding light on ome aspects of the proposed value functions, which might be sed in real-world applications. emark 1. Shape and parameters of a value function is decision- aker-dependent, implying that (i) while a decision-maker consid- rs, for instance, a level-increasing function for the size of gar- en when buying a house, another decision-maker considers an ncreasing stepwise function, and (ii) while two decision-makers onsider increasing stepwise function for the size of garden when uying a house, the parameters they consider for their functions ( d l j , d m j , d u j ) might be different. 48 J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 Fig. 11. Increasing-level-decreasing value function. Table 1 Suppliers performance with respect to different decision criteria ( x ij ). Criteria Supplier Quality Price ( €/item) Trust CO 2 (g/item) Delivery (day) 1 85 27 4 10 0 0 3 2 90 28 2 1500 4 3 80 26 5 20 0 0 3 4 75 25 5 10 0 0 2 5 95 29 7 1700 3 6 99 30 6 20 0 0 1 w w v e h a d t a r m n t t t d m o U w s T w w T v 2 4 v e u p a s b fi w a e g 2 Please note that we report some weights for the criteria as the aim of the study is not the weighing part. Remark 2. A decision-maker might consider a hybrid value func- tion for a criterion. For instance, a criterion might be character- ized with an increasing-level-decreasing, which is a combination of increasing-level and level-decreasing. This function can also be considered as a special form of inverted V-shape function. It is shown in Fig. 11 and formulated as follows: u i j = ⎧ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ x i j − d l j d m j − d l j , d l j ≤ x i j ≤ d m 1 j , 1 , d m 1 ≤ x i j ≤ d m 2 , d u j − x i j d u j − d m j , d m 2 ≤ x i j ≤ d u j , 0 , otherwise . (16) For instance, a decision-maker has to select the best R&D part- ner among 10 partners. One of the criteria is distance and the com- pany gives less preference to very close or very distant partners, which are distributed in the range [10 km, 20 0 0 km]. The com- pany considers an optimal distance of [20 0 km, 50 0 km]. This im- plies that distance follows an ‘increasing-level-decreasing’ function for this decision-maker: [ d l j , d m 1 j , d m 2 j , d u j ] = [ 10 , 200 , 500 , 2000 ] . 4. Numerical and comparison analyses In this section, we show how to incorporate the proposed piecewise value functions into account when applying an MCDM method, and we show that the results might be different when we consider the proposed piecewise value functions. We consider an MCDM problem, where a buyer should select a supplier from among six qualified suppliers considering five crite- ria: quality which is measured by 1 −α, where α shows the lot-size average imperfect rate; price (euro) per item; trust, which is mea- sured by a Likert scale (1: very low to 7: very high); CO 2 (gram) per item; delivery (day), the amount of time which takes to deliver items from the supplier to the buyer (all the criteria are continuous except trust). Table 1 shows the performance of the six suppliers with respect to the five criteria. The buyer has used an elicitation method 2 to find the weights hich are as follows: ∗ quality = 0 . 20 , w ∗price = 0 . 30 , w ∗trust = 0 . 27 , w ∗CO 2 = 0 . 08 , w ∗ delivery = 0 . 15 . And we assume that the decision-maker considers piecewise alue functions for these criteria (see, Table 2 ). So, as can be seen from Table 2 , the decision-maker consid- rs a level-increase linear function for quality with the lowest and ighest values of 0 and 100, respectively. For the decision-maker ny number below 85 has no value at all. For criterion price, the ecision-maker gives the highest value to any price below 15 (al- hough in the existing set of suppliers there is no supplier with price within this range), after which the value decreases till it eaches to the maximum price of 30. For criterion trust which is easured using a Likert scale (1: very low to 7: very high), any umber less than 3 has no value for the decision-maker, while he value gradually increases between 3 and 7. For CO 2 emission, here is a decreasing value function from 0 to 1500 g per item, af- er which till 20 0 0 g, all the numbers have zero value. Finally, for elivery there is a simple decreasing function with minimum and aximum values of 0 and 5 days. By using the following equation, we can find the overall value f each supplier and then rank them to find the best supplier. i = n ∑ j=1 w j u i j (17) here, u ij is the value of the performance of supplier i with re- pect to criterion j (using the equations in Table 2 for the data in able 1 ), and w j is the weight of criterion j as follows: ∗ qual ity = 0 . 20 , w ∗price = 0 . 30 , w ∗trust = 0 . 27 , ∗ CO 2 = 0 . 08 , w ∗delivery = 0 . 15 . The value scores and the aggregated values are presented in able 3 (see also Fig. 13 for the final results). As can be seen from Table 3 , supplier 6 with the greatest overall alue of 0.51 is ranked as the first supplier. Suppliers 5, 4, 3, 1, and are ranked in the next places. .1. Comparing with the simple linear value functions In existing literature, considering the nature of the criteria, the alues are calculated, for instance, using the following simple lin- ar value function: i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ x i j − d l j d u j − d l j , if more x i j is more desirable ( such as quality ) , d u j − x i j d u j − d l j , if more x i j is less desirable ( such as price ) . (18) Eq. (18) is used to find the values of the criteria for each sup- lier using the data in Table 1 . Considering the criteria weights, nd u ij ( Eq. (18) for the data in Table 1 ) using Eq. (17) the value cores and also the aggregated overall score of each alternative can e calculated which are shown in Table 4 (see also Fig. 13 for the nal results). In Table 4 it is assumed that quality and trust are criteria for hich the higher the better, while for the other criteria (price, CO 2 , nd delivery), the lower the better. In fact, we consider simple lin- ar functions (increasing and decreasing respectively) for the two roups of criteria. J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 49 Table 2 Piecewise value functions. Shape Value function u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 0 , d l j ≤ x i j ≤ d m j , x i j − d m j d u j − d m j d m j ≤ x i j ≤ d u j , 0 , otherwise . = ⎧ ⎪ ⎨ ⎪ ⎩ 0 , 0 ≤ x i j ≤ 85 , x i j − 85 100 − 85 , 85 ≤ x i j ≤ 100 , 0 , otherwise . u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 1 , d l j ≤ x ≤ d m j , d u j − x i j d u j − d m j , d m j ≤ x ≤ d u j , 0 , otherwise . = ⎧ ⎪ ⎨ ⎪ ⎩ 1 , 13 ≤ x i j ≤ 15 , 30 − x i j 30 − 15 , 15 ≤ x i j ≤ 30 , 0 , otherwise . u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 0 , d l j ≤ x i j ≤ d m j , x i j − d m j d u j − d m j , d m j ≤ x i j ≤ d u j , 0 , otherwise . = ⎧ ⎪ ⎨ ⎪ ⎩ 0 , 1 ≤ x i j ≤ 3 , x i j − 3 7 − 3 , 3 ≤ x i j ≤ 7 , 0 , otherwise . u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ d m j − x i j d m j − d l j , d l j ≤ x i j ≤ d m j , 0 , d m j ≤ x i j ≤ d u j , 0 , otherwise . = ⎧ ⎪ ⎨ ⎪ ⎩ 1500 − x i j 1500 − 0 , 0 ≤ x i j ≤ 1500 , 0 , 1500 ≤ x i j ≤ 2000 , 0 , otherwise . u i j = ⎧ ⎨ ⎩ d u j − x i j d u j − d l j , d l j ≤ x i j ≤ d u j , 0 , otherwise . = { 5 − x i j 5 − 0 , 0 ≤ x i j ≤ 5 , 0 , otherwise . r ( a c i 3 t s v e t According to Table 4 , the best supplier is supplier 4, which is anked as the 3rd one considering the piecewise value functions Table 3 ). The ranking of the other suppliers is also different. So, s can be seen, such differences are associated to the way we cal- ulate the value of the criteria. If we look at the criterion trust, for nstance ( Table 3 ), we can see that only the numbers greater than can be used for compensating the other criteria. In other words, he values 1, 2 and 3 for this criterion have no selection power. No upplier can compensate its weakness in other criteria by having a alue between 1 and 3 for trust. However, such important issue is ntirely ignored in the simple way of determining the value func- ions which is very popular in existing studies. This consideration 50 J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 Table 3 Value scores, u ij , and the aggregated overall score considering the proposed piecewise value functions. Supplier Quality Price Trust CO 2 Delivery Aggregated value Rank 1 0.00 0.20 0.25 0.50 0.40 0.23 5 2 0.33 0.13 0.00 0.00 0.20 0.14 6 3 0.00 0.27 0.50 0.00 0.40 0.28 4 4 0.00 0.33 0.50 0.50 0.60 0.37 3 5 0.67 0.07 1.00 0.00 0.40 0.48 2 6 0.93 0.00 0.75 0.00 0.80 0.51 1 Table 4 Value scores, u ij , and the aggregated overall scores considering the simple linear value func- tions. Supplier Quality Price Trust CO 2 Delivery Aggregated value Rank 1 0.42 0.60 0.40 1.00 0.33 0.50 4 2 0.63 0.40 0.00 0.50 0.00 0.29 6 3 0.21 0.80 0.60 0.00 0.33 0.49 5 4 0.00 1.00 0.60 1.00 0.67 0.64 1 5 0.83 0.20 1.00 0.30 0.33 0.57 2 6 1.00 0.00 0.80 0.00 1.00 0.57 3 Fig. 12. Exponential value functions. s u f c F f v d M v ( f t 3 To see how these value functions are elicited considering the decision-maker is even of a higher importance for compensatory methods such as AHP and BWM. 4.2. Comparing with the exponential value functions Another important way to approximate the value functions in practice is the use of exponential value functions ( Kirkwood, 1997; Pratt, 1964 ). The exponential value functions can specifically be used when the preferences are monotonically increasing or de- creasing. Although this approach is not popular in MCDM do- main, and we were not able to find any application of these value functions particularly in MCDM field, we would like to compare our results to the results of applying these functions, which are, to some extent, close to some of our proposed piecewise value functions (such as level-increase, level-decrease, increase-level, and decrease-level). Using the same notations as before and consider- ing a shape parameter ρ which is called ‘risk tolerance’, a mono- tonically increasing exponential value function can be shown as follows: u i j = ⎧ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎩ 1 − exp [ − ( x i j − d l j ) /ρ ] 1 − exp [ − ( d u j − d l j ) /ρ ] , ρ � = Infinity x i j − d l j d u j − d l j , otherwise . (19) r A monotonically decreasing exponential value function can be hown as follows: i j = ⎧ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎩ 1 − exp [ − ( d u j − x i j ) /ρ ] 1 − exp [ − ( d u j − d l j ) /ρ ] , ρ � = Infinity d u j − x i j d u j − d l j , otherwise . (20) Fig. 12 shows the monotonically increasing exponential value unctions (for different values of ρ) (a), and the monotonically de- reasing exponential value functions (for different values of ρ) (b). Risk-averse decision-makers have ρ > 0 (hill-like functions in ig. 12 ), while risk-seeking decision-makers have ρ < 0 (bowl-like unctions in Fig. 12 ). ρ = Infinity (straight-line in Fig. 12 ) shows the alue for the risk neutral decision-makers. In fact, ρ = Infinity pro- uces the simple linear value functions which are very popular in CDM field. In order to do the comparison analysis, we use exponential alue functions for the criteria of the aforementioned example Table 1 ) to check the similarities and differences. To make a air comparison, we try to generate 3 the corresponding exponen- ial value functions of the piecewise value functions ( Table 2 ) as isk tolerance, refer to Kirkwood (1997) . J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 51 c t a t t p l ρ f s o p lose as possible. For quality, a monotonically increasing exponen- ial value function with negative ρ would be appropriate. For price monotonically decreasing exponential value function with a posi- ive ρ, for trust, a monotonically increasing exponential value func- ion with a negative ρ, for CO 2 , a monotonically decreasing ex- onential value function with a negative ρ, and, finally, for de- ivery a monotonically decreasing exponential value function with Table 5 Exponential value functions. Shape = Infinity would be suitable. Table 5 shows the functions, where unctions with different ρ′ s are shown and a more suitable one is hown in bold. Using the exponential value functions of Table 5 , for the data f Table 1 , we get the value scores and the aggregated values as resented in Table 6 (see also Fig. 13 for the final results). Function u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 1 − exp [ −( x i j − d l j ) /ρ] 1 − exp[ −( d u j − d l j ) /ρ] , ρ � = Infinity x i j − d l j d u j − d l j , otherwise . = ⎧ ⎨ ⎩ 1 − exp [ −( x i j − 0 ) /ρ] 1 − exp[ −( 100 − 0 ) /ρ] , ρ � = Infinity x i j − 0 100 − 0 , otherwise . u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 1 − exp [ −( d u j − x i j ) /ρ] 1 − exp[ −( d u j − d l j ) /ρ] , ρ � = Infinity d u j − x i j d u j − d l j , otherwise . = ⎧ ⎨ ⎩ 1 − exp [ −( 30 − x i j ) /ρ] 1 − exp[ −( 30 − 0 ) /ρ] , ρ � = Infinity 30 − x i j 30 − 0 , otherwise . u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 1 − exp [ −( x i j − d l j ) /ρ] 1 − exp[ −( d u j − d l j ) /ρ] , ρ � = Infinity x i j − d l j d u j − d l j , otherwise . = ⎧ ⎨ ⎩ 1 − exp [ −( x i j − 1 ) /ρ] 1 − exp[ −( 7 − 1 ) /ρ] , ρ � = Infinity x i j − 1 7 − 1 , otherwise . u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 1 − exp [ −( d u j − x i j ) /ρ] 1 − exp[ −( d u j − d l j ) /ρ] , ρ � = Infinity d u j − x i j d u j − d l j , otherwise . = ⎧ ⎨ ⎩ 1 − exp [ −( 20 0 0 − x i j ) /ρ] 1 − exp[ −( 20 0 0 − 1500 ) /ρ] , ρ � = Infinity 20 0 0 − x i j 20 0 0 − 0 , otherwise . ( continued on next page ) 52 J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 Table 5 ( continued ) Shape Function u i j = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ 1 − exp [ −( d u j − x i j ) /ρ] 1 − exp[ −( d u j − d l j ) /ρ] , Infinity d u j − x i j d u j − d l j , otherwise . = ⎧ ⎨ ⎩ 1 − exp [ −( 5 − x i j ) /ρ] 1 − exp[ −( 5 − 0 ) /ρ] , ρ � = Infinity 5 − x i j 5 − 0 , otherwise . Table 6 Value scores, u ij , and the aggregated overall scores considering the exponential value func- tions. Supplier Quality Price Trust CO 2 Delivery Aggregated value Rank 1 0.85 0.45 0.05 0.08 0.40 0.22 5 2 0.90 0.33 0.00 0.02 0.20 0.16 6 3 0.80 0.55 0.13 0.00 0.40 0.26 4 4 0.75 0.63 0.13 0.08 0.60 0.32 3 5 0.95 0.18 1.00 0.00 0.40 0.46 1 6 0.99 0.00 0.37 0.00 0.80 0.38 2 Fig. 13. Final results using three types of value functions. v o t p t l c c a As can be seen from Table 6 , supplier 5 with the greatest over- all value of 0.46 is ranked as the first supplier, which is different from what we get from the proposed piecewise value functions ( Table 3 ). While supplier 5 was ranked the 2nd based on our pro- posed value functions, using the exponential value functions, this supplier becomes number 2. Other suppliers (1, 2, 3, 4) have the same ranking based on the two approaches. The differences are ob- viously associated to the way we get the values of the criteria. We also checked some other close ρ values for the exponential value functions. There are some changes in the aggregated values, yet, the ranking is the same. As it can be seen, the results of the two approaches (piecewise alue functions and exponential functions are much closer to each ther than to the results of the regular simple linear value func- ions). Our observation is that the exponential value functions can lay a role close to a number of proposed value functions in his study such as increase-level, decrease-level, level-increase, and evel-decrease. In order to make an exponential value functions lose to one of the mentioned proposed value functions we should hoose ρ values close to zero. If we try to make the ρ as close s we perfectly make the “level” part of the criterion, then the J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 53 Fig. 14. Fitting an exponential value function to a level-decrease value function. o h s i e 1 c i t p O c t i c t i o t c v f o l r a t a t w e e ( e e p e a i 5 fi h l n a v m fi s 1 & t s e fl n r u m T p m f t m s t c t 6 m i i v ther part becomes very steep and not representative. On the other and, if we want to choose a ρ value which better represents the lope of the function for the “increase” or the “decrease” part, then t is impossible to cover the “level” part of the value function prop- rly. For example, let us consider the criterion price again. In Fig. 4 , it can be seen that, if we consider ρ = 1, the level part is fully overed, but the decrease part of the exponential value function s very much different from the decrease part of the linear func- ion. Even if we choose ρ = 3, which does not fully cover the level art of the proposed function, the decrease part is really different. n the other hand, we can find ρ = 8 as a close one to the de- rease part of the linear function, but this time it is not possible o cover the level part properly. So, although a very good approx- mation, the exponential value functions might not be suitable for ases in which a decision-maker has a clear value-indifference in- erval (a level part) for a criterion. However, these functions are ndeed suitable when the decision-maker has different preferences n the lower and on the upper parts of the criterion measure. From the figure we can also see that while we could make the wo piecewise and exponential value functions, to some degree, lose to each other, they are too different from the simple linear alue function. It is also clear that none of the simple linear value unctions or the exponential value functions can represent the V- r inverted V-shape value functions. As a general conclusion, we think that the proposed piecewise inear functions have two salient features: (i) simplicity; and (ii) epresentativeness. That is, it is easy to work with linear functions, nd it is easy for a practitioner to find a more representative func- ion from the proposed library of the pricewise value functions for particular criterion. The cut-off points can also be estimated by he decision-maker. The simple monotonic-linear value functions, hich are dominant in existing literature, are very simple. How- ver, they might not be representative in some cases. Finally, the xponential value functions might have a better representativeness compared to the simple monotonic-linear value functions). How- ver, they are not simple. Working with non-liner functions is not asy for practitioner, and, more importantly, it is very difficult for a ractitioner to estimate a value for ρ (the shape parameter of the xponential value functions), as it cannot be easily interpreted by practitioner (please note that we consider a value function as an nput for an MCDM problem in this study). s I t l . Determining the value functions One of the big challenges in real-world decision-making is to nd a proper value function for a decision criterion. This, per- aps, has been one of the main reasons why the use of simple inear value functions in multi-criteria decision-making is domi- ant. The linear value functions are easy for modeling purposes nd can, to some extent, represent the reality. More complicated alue functions, although might be closer to reality of the decision- aker’s preferences, are more difficult to be elicited and are dif- cult for modeling purposes. We refer the interested readers to ome existing procedures for identifying value functions ( Fishburn, 967; Keeney & Nair, 1976; Kirkwood, 1997; Pratt, 1964; Stewart Janssen, 2013 ). We think that the proposed value functions in his paper do not have the disadvantage of nonlinearity and at the ame time have the advantages of being closer to the real pref- rences of the decision-maker as they provide some diversity and exibility in modeling the functions. As we do not consider the onlinearity of the value functions we do not use the concept of isk tolerance in determining the value functions as it has been sed by others. We rather propose a simple procedure, which is ore practical. A decision analyst, could first show the value functions in able 2 to the decision-maker to see which one most suits the reference structure of the decision-maker. Once the decision- aker selects a particular value function, the other details of the unction, such as the lower bound, the upper bound and the hresholds can be determined. We should highlight again that in ost MCDM methods, the value function is not elicited. It is rather imply assumed to have a particular shape, and this is why we hink having a pre-specified set of standard value functions which an be used as subjective approximation of the real preferences of he decision-maker can make a significant impact on the results. . Conclusion, limitations and future research This study proposes a set of piecewise value functions for ulti-criteria decision-making (MCDM) problems. While the exist- ng applications of MCDM methods usually use two general simple ncreasing and decreasing linear value functions, this study pro- ides several real-world examples to support the applicability of ome other forms of value functions for the criteria used in MCDM. t is also explicated how, in some decision problems, a combina- ion of two or more value functions can be used for a particu- ar decision criterion. The proposed functions can be used for dif- 54 J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 Ç C E F F H J K K K K L M O P P P Q R R R R R S T ferent MCDM methods in different decision problems. A numeri- cal example of supplier selection problem (including a comparison between simple monotonic linear value functions, piecewise lin- ear value functions, and exponential value functions) showed how the use of the proposed value functions could affect the final re- sults. Considering these value functions could better represent the real preferences of the decision-maker. It can also help reduce the inappropriate compensations of the decision criteria, for instance, through using a level-increasing function which assigns zero value to any value of the criterion below a certain threshold. The pro- posed value functions are presented in a general form such that they can be tailor-made for a specific decision-maker. That is to say, not only it is possible for two different decision-makers to use two different value functions for a single criterion. It is also pos- sible to use different domain (e.g. min and max) values for that particular value function. Despite the advantages of the proposed value functions, they have some limitations. Although the proposed value functions con- sider some real-world features of the decision criteria, they are lin- ear which might be, to some degree, a simplification. We think that the bigger problem in existing literature of MCDM is mono- tonicity assumption and not linearity assumption. Nevertheless, more research needs to be conducted to empirically find the share of each. Furthermore, to formulate the decision criteria one should pay enough attention to check the real contribution of the decision criterion into the ultimate goal of the decision-making problem. For instance, if a criterion contributes to another criterion which has a real role in making the decision, one should exclude the ini- tial one. For a detailed discussion on this matter, interested read- ers are referred to Brugha (1998) . One interesting future direction would be to apply the proposed value functions in some real-world MCDM problems and compare their fitness to the other value func- tions. In this regard, finding a more systematic approach to deter- mine the value functions in practice would be also very interest- ing. It would be also interesting to study the cases in which there are more than one decision-maker. As different decision-makers may choose different value functions, different domains, and dif- ferent thresholds for a single criterion, proposing a way to find the final output of the MCDM problem for the group would be an interesting future research. Finally, finding a sensitivity anal- ysis for the proposed value functions is recommended. Consider- ing the studies of Bertsch and Fichtner (2016 ), Bertsch, Treitz, Gel- dermann, and Rentz (2007 ), Insua and French (1991 ), Wulf and Bertsch (2017 ) could give interesting ideas to make such sensitivity analysis framework. References Alanne, K. , Salo, A. , Saari, A. , & Gustafsson, S.-I. (2007). Multi-criteria evaluation of residential energy supply systems. Energy and Buildings, 39 , 1218–1226 . Ananda, J. , & Herath, G. (2009). A critical review of multi-criteria decision making methods with special reference to forest management and planning. Ecological Economics, 68 , 2535–2548 . Bertsch, V. , & Fichtner, W. (2016). A participatory multi-criteria approach for power generation and transmission planning. Annals of Operations Research, 245 , 177–207 . Bertsch, V. , Treitz, M. , Geldermann, J. , & Rentz, O. (2007). Sensitivity analyses in multi-attribute decision support for off-site nuclear emergency and recovery management. International Journal of Energy Sector Management, 1 , 342–365 . Brans, J. P. , Mareschal, B. , Vincke, P. , & Brans, J. P. (1984). PROMETHEE: A new fam- ily of outranking methods in multicriteria analysis. In Proceedings of the 1984 conference of the international federation of operational research societies (IFORS) (pp. 477–490). Amsterdam: North Holland. 84 . Brugha, C. M. (1998). Structuring and weighting criteria in multi criteria deci- sion making (MCDM). In Trends in multicriteria decision making (pp. 229–242). Springer . Brugha, C. M. (20 0 0). Relative measurement and the power function. European Jour- nal of Operational Research, 121 , 627–640 . Carrico, C. S. , Hogan, S. M. , Dyson, R. G. , & Athanassopoulos, A. D. (1997). Data envel- opment analysis and university selection. The Journal of the Operational Research Society, 48 , 1163–1177 . ebi, F. , & Otay, İ. (2016). A two-stage fuzzy approach for supplier evaluation and order allocation problem with quantity discounts and lead time. Information Sci- ences, 339 , 143–157 . Chae, B. (2009). Developing key performance indicators for supply chain: An indus- try perspective. Supply Chain Management: An International Journal, 14 , 422–428 . hou, T.-Y. , Hsu, C.-L. , & Chen, M.-C. (2008). A fuzzy multi-criteria decision model for international tourist hotels location selection. International Journal of Hospi- tality Management, 27 , 293–301 . dwards, W. (1977). How to use multiattribute utility measurement for social deci- sion making. IEEE Transactions on Systems, Man and Cybernetics, 7 , 326–340 . ishburn, P. C. (1967). Methods of estimating additive utilities. Management Science, 13 , 435–453 . renette, M. (2004). Access to college and university: Does distance to school mat- ter. Canadian Public Policy/Analyse de Politiques, 30 , 427–443 . Herrera, F. , Herrera-Viedma, E. , & Verdegay, J. (1996). Direct approach processes in group decision making using linguistic OWA operators. Fuzzy Sets and Systems, 79 , 175–190 . o, W. , Xu, X. , & Dey, P. K. (2010). Multi-criteria decision making approaches for supplier evaluation and selection: A literature review. European Journal of Oper- ational Research, 202 , 16–24 . Hoen, K. , Tan, T. , Fransoo, J. , & van Houtum, G. (2014). Effect of carbon emission reg- ulations on transport mode selection under stochastic demand. Flexible Services and Manufacturing Journal, 26 , 170–195 . Insua, D. R. , & French, S. (1991). A framework for sensitivity analysis in discrete multi-objective decision-making. European Journal of Operational Research, 54 , 176–190 . acquet-Lagreze, E. , & Siskos, Y. (2001). Preference disaggregation: 20 years of MCDA experience. European Journal of Operational Research, 130 , 233–245 . abir, G. , Sadiq, R. , & Tesfamariam, S. (2014). A review of multi-criteria decision– making methods for infrastructure management. Structure and Infrastructure En- gineering, 10 , 1176–1210 . akeneno, J. R. , & Brugha, C. M. (2017). Usability of nomology-based methodologies in supporting problem structuring across cultures: The case of participatory de- cision-making in Tanzania rural communities. Central European Journal of Oper- ations Research, 25 , 393–415 . Keeney, R. L., & Nair, K. (1976). Evaluating potential nuclear power plant sites in the Pacific Northwest using decision analysis, IIASA Professional Paper. IIASA, Laxenburg, Austria: PP-76-001. Keeney, R. L. , & Raiffa, H. (1976). Decisions with multiple objectives: preferences and value tradeoffs . USA: John Wiley & Sons, Inc . irkwood, C. W. (1997). Strategic decision making . California, USA: Wadsworth Pub- lishing Company . raljic, P. (1983). Purchasing must become supply management. Harvard Business Review, 61 , 109–117 . ahdelma, R. , & Salminen, P. (2012). The shape of the utility or value function in stochastic multicriteria acceptability analysis. OR Spectrum, 34 , 785–802 . Lambert, D. M. , Emmelhainz, M. A. , & Gardner, J. T. (1996). Developing and imple- menting supply chain partnerships. The International Journal of Logistics Manage- ment, 7 , 1–18 . Montgomery, B. M. (1981). The form and function of quality communication in mar- riage. Family Relations, 30 , 21–30 . ustajoki, J. , & Hämäläinen, R. P. (20 0 0). Web-HIPRE: Global decision support by value tree and AHP analysis. INFOR: Information Systems and Operational Re- search, 38 , 208–220 . Nooteboom, B. (20 0 0). Learning and innovation in organizations and economies . Ox- ford: University Press . ’Brien, D. B. , & Brugha, C. M. (2010). Adapting and refining in multi-criteria deci- sion-making. Journal of the Operational Research Society, 61 , 756–767 . loetner, O. , & Ehret, M. (2006). From relationships to partnerships—New forms of cooperation between buyer and seller. Industrial Marketing Management, 35 , 4–9 . ohekar, S. , & Ramachandran, M. (2004). Application of multi-criteria decision mak- ing to sustainable energy planning—a review. Renewable and Sustainable Energy Reviews, 8 , 365–381 . ratt, J. W. (1964). Risk aversion in the small and in the large. Econometrica, 32 , 122–136 . i, X. (2015). Disruption management for liner shipping. In Handbook of ocean con- tainer transport logistics (pp. 231–249). Springer . edmond, L. S. , & Mokhtarian, P. L. (2001). The positive utility of the commute: Modeling ideal commute time and relative desired commute amount. Trans- portation, 28 , 179–205 . ezaei, J. (2015). Best-worst multi-criteria decision-making method. Omega, 53 , 49–57 . ezaei, J. (2016). Best-worst multi-criteria decision-making method: Some proper- ties and a linear model. Omega, 64 , 126–130 . ezaei, J. , & Ortt, R. (2012). A multi-variable approach to supplier segmentation. In- ternational Journal of Production Research, 50 , 4593–4611 . oy, B. (1968). Classement et choix en présence de points de vue multiples (la méthode ELECTRE). RIRO, 2 , 57–75 . Saaty, T. L. (1977). A scaling method for priorities in hierarchical structures. Journal of Mathematical Psychology, 15 , 234–281 . tewart, T. J. , & Janssen, R. (2013). Integrated value function construction with appli- cation to impact assessments. International Transactions in Operational Research, 20 , 559–578 . sai, K.-H. , & Wang, J.-C. (2005). Does R&D performance decline with firm size?—A re-examination in terms of elasticity. Research Policy, 34 , 966–976 . http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0001 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0001 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0001 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0001 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0001 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0001 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0002 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0002 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0002 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0002 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0003 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0003 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0003 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0003 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0004 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0004 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0004 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0004 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0004 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0004 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0005 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0005 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0005 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0005 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0005 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0005 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0006 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0006 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0007 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0007 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0008 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0008 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0008 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0008 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0008 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0008 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0009 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0009 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0009 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0009 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0010 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0010 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0011 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0011 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0011 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0011 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0011 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0012 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0012 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0013 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0013 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0014 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0014 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0015 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0015 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0015 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0015 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0015 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0016 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0016 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0016 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0016 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0016 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0017 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0017 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0017 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0017 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0017 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0017 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0019 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0019 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0019 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0019 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0020 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0020 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0020 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0020 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0020 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0021 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0021 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0021 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0021 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0022 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0022 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0022 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0022 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0023 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0023 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0024 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0024 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0025 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0025 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0025 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0025 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0026 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0026 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0026 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0026 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0026 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0028 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0028 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0029 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0029 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0029 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0029 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0030 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0030 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0031 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0031 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0031 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0031 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0032 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0032 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0032 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0032 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0033 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0033 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0033 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0033 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0034 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0034 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0035 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0035 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0036 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0036 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0036 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0036 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0037 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0037 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0038 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0038 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0039 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0039 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0039 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0039 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0040 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0040 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0041 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0041 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0042 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0042 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0042 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0042 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0043 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0043 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0043 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0043 J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 55 T W W X Y zeng, G.-H. , Lin, C.-W. , & Opricovic, S. (2005). Multi-criteria analysis of alternative– fuel buses for public transportation. Energy Policy, 33 , 1373–1383 . ilson, T. L. , & Anell, B. I. (1999). Business service firms and market share. Journal of Small Business Strategy, 10 , 41–53 . ulf, D. , & Bertsch, V. (2017). A natural language generation approach to sup- port understanding and traceability of multi-dimensional preferential sensitivity analysis in multi-criteria decision making. Expert Systems with Applications, 83 , 131–144 . ia, W. , & Wu, Z. (2007). Supplier selection with multiple criteria in volume dis- count environments. Omega, 35 , 494–504 . ager, R. R. (1988). On ordered weighted averaging aggregation operators in multi- criteria decision making. IEEE Transactions on Systems, Man, and Cybernetics, 18 , 183–190 . http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0044 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0044 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0044 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0044 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0044 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0045 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0045 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0045 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0045 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0046 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0046 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0046 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0046 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0047 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0047 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0047 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0047 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0048 http://refhub.elsevier.com/S0957-4174(18)30004-6/sbref0048 56 J. Rezaei / Expert Systems With Applications 98 (2018) 43–56 t Delft University of Technology, the Netherlands, where he also obtained his Ph.D. One of ) analysis. He has published in various academic journals, including International Journal arketing Management, Applied Soft Computing, Applied Mathematical Modelling, IEEE Journal of Operational Research, Information Science, Omega, and Expert Systems with Jafar Rezaei is an associate professor of operations and supply chain management a his main research interests is in the area of multi-criteria decision-making (MCDM of Production Economics, International Journal of Production Research, Industrial M Transactions on Engineering Management, Journal of Cleaner Production, European Applications. Piecewise linear value functions for multi-criteria decision-making 1 Introduction 2 Piecewise linear value functions 2.1 Increasing 2.2 Decreasing 2.3 V-shape 2.4 Inverted V-shape 2.5 Increase-level 2.6 Level-decrease 2.7 Level-increase 2.8 Decrease-level 2.9 Increasing stepwise 2.10 Decreasing stepwise 3 Some remarks on the value functions 4 Numerical and comparison analyses 4.1. Comparing with the simple linear value functions&#13; 4.2. Comparing with the exponential value functions 5 Determining the value functions 6 Conclusion, limitations and future research References 
work_3cy3ym4zi5ht5g5z7rnaurcdty	----	UN CO RR EC TE D PR O O F Intuitionistic fuzzy rough sets: at the cross- roads of imperfect knowledge Chris Cornelis, Martine De Cock and Etienne E. Kerre Department of Applied Mathematics and Computer Science, Ghent University, Fuzziness and Uncertainty Modelling Research Unit, Krijgslaan 281 (S9), B-9000 Ghent, Belgium E-mail: {chris.cornelis, martine.decock, etienne.kerre}@rug.ac.be Abstract: Just like rough set theory, fuzzy set theory addresses the topic of dealing with imperfect knowledge. Recent investiga- tions have shown how both theories can be combined into a more flexible, more expressive framework for modelling and processing incomplete information in information systems. At the same time, intuitionistic fuzzy sets have been proposed as an attractive extension of fuzzy sets, enriching the latter with extra features to represent uncertainty (on top of vagueness). Unfortunately, the various tentative definitions of the concept of an ‘intuitionistic fuzzy rough set’ that were raised in their wake are a far cry from the original objectives of rough set theory. We intend to fill an obvious gap by introducing a new definition of intuitionistic fuzzy rough sets, as the most natural generalization of Pawlak’s original concept of rough sets. Keywords: rough set theory, intuitionistic fuzzy set theory, L-fuzzy set theory, incomplete information, lower and upper approximation 1. Introduction As a new trend in the attempts to combine the best of several worlds, very recently all kinds of suggestions for approaches merging rough set theory and intuitionistic fuzzy set theory have started to appear. The present evolution vividly reminds us of the origin of fuzzy rough set theory, as the (so far) happy marriage of fuzzy set theory and rough set theory. A remarkable difference, however, is that in the latter case a long engagement period with intense discussions concerning the relation between fuzzy set theory and rough set theory preceded the marriage, (and in a way is still going on, simultaneously with the research on the new hybrid theory). So far, this comparison stage seems very limited for the combination of intuitionistic fuzzy set theory and (fuzzy) rough set theory. As far as we know, only Çoker (1998) went into the matter by claiming that fuzzy rough sets are intuitionistic fuzzy sets (Chakra- barty et al., 1998; Samanta & Mondal, 2001; Jena & Ghosh, 2002; Rizvi et al., 2002), which appears to be shattering the dream of a new hybrid theory. On the other hand, there exist many views on the notion ‘rough set’ which can be grouped into two main streams. Several suggested options for fuzzification have led to an even greater number of views on the notion ‘fuzzy rough set’. Typically, under the same formal umbrella, they can be further generalized to the notion ‘L-fuzzy rough set’ where the membership degrees are taken from some suitable lattice L which is not necessarily the unit interval. On top of this, there exist semantically different interpretations of intuitionistic fuzzy set theory (which is a special kind of L-fuzzy set theory). Needless to say, when trying to compare and/or to combine rough set theory, fuzzy set theory and intuitionistic fuzzy set theory, one finds oneself at a complicated crossroads with an abundance of possible ways to proceed. The aim of this paper is to provide the reader with a road map. We do this by mapping out research results obtained so far in the literature, as well as by exploring by our- selves a very important road which was virgin territory until now. However, before we can prepare a hybrid theory, it is absolutely necessary to check the origin of all ingredients, for they can have an important influence on the flavour of the resulting product! For this reason we start the paper with a short overview of all set theoretical models involved (Section 2). Making such a study forces us into thinking about relationships between them. Staying within the scope of the paper, we will only focus on intuitionistic fuzzy set theory versus fuzzy rough set theory (Section 3). After a critical examination of the added value of a hybrid intuitionistic fuzzy rough set theory, in Section 4 we present an overview of existing approaches (all originated independently from the others). Finally we fill an obvious gap by a very natural generalization of Pawlak’s original concept of a rough set. Article_______________________________________ Journal: EXSY H Disk used ED: Ramesh: Pgn by: solly Article : 02005003 Pages: 10 Despatch Date: 8/8/2003 260 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES UN CO RR EC TE D PR O O F 2. Short overview of some set theoretical models 2.1. Rough set theory Rough sets, remarkably enough, are not a uniquely defined notion. A very accurate classification of various ap- proaches to definitions was given by Yao (1996). For our purposes, it is sufficient to distinguish between two main streams in the literature, denoted ‘Pawlak rough sets’ and ‘Iwinski rough sets’, respectively; both come with their own particular perception of the notion ‘rough set’. Pawlak rough sets The first stream was initiated by Pawlak (1982) who launched rough set theory as a framework for the construction of approximations of concepts when only incomplete information is available. The available information consists of a set A of examples (a subset of a universe X, X being a non-empty set of objects we want to say something about) of a concept C, and a relation R in X. R models ‘indiscernibility’ or ‘indistinguishability’ and therefore generally is a tolerance relation (i.e. a reflexive and symmetrical relation) and in most cases even an equivalence relation (i.e. a transitive tolerance relation). In this paper we will use R-foresets to denote equivalence classes. The R-foreset of an element y of X is the set Ry ¼ {x|(x, y)AR}. The couple (X, R) is called an approximation space. Rough set analysis makes statements about the membership of some element x of X to the concept C of which A is a set of examples, based on the indistinguishability between x and the elements of A. To arrive at such statements, A is approximated in two ways. The lower and upper approximations of A in (X, R) are respectively the following subsets of X: A ¼ fyjy 2 X and Ry � Ag �AA ¼ fyjy 2 X and Ry \ Ag The underlying meaning is that �AA is the set of elements possibly belonging to the concept C (weak membership), while A is the set of elements necessarily belonging to C (strong membership); for y belongs to A if all elements of X indistinguishable from y belong to A (hence there is no doubt that y also belongs to A), while y belongs to �AA as soon as an element of A is indistinguishable from y. If y belongs to the boundary region �AA=A, then there is doubt, because in this case y is at the same time indistinguishable from at least one element of A and at least one element of X that is not in A. A set A is called definable if A ¼ �AA (Radzikowska & Kerre, 2002). The existing variety in terminology on this concept of definability might indicate the high importance individual researchers attach to it. Thiele for instance uses the term ‘R-exact’ as a synonym for ‘definable’ (see Thiele, 1998). Pawlak (1982) defines a ‘composed set’ as a finite union of equivalence classes. In 1985 he generalizes this definition to any union of equivalence classes; it can be verified that this notion of composed set coincides with that of definable set. Still other terms to denote the same notion are mentioned in Iwinski (1987). A similar phenomenon occurs concerning the definition of the concept ‘rough set’: some authors say that a set A in X is a ‘rough set’ if A 6¼ �AA (see for example Komorowski et al., 1999). Hence they call a set ‘rough’ if the boundary region is not empty, i.e. if there is at least one object which cannot be classified with certainty as a member of the concept, nor as a member of its complement. Pawlak (1982), on the other hand, originally defined a rough set (A1, A2) as the class of all sets that have A1 and A2 respectively as lower and upper approximations. This convention is also adopted by Thiele (2001). Still others call (A1, A2) a rough set (in (X, R)) as soon as there is a set A in X such that A ¼ A1 and �AA ¼ A2 (see for example Radzikowska & Kerre, 2002); (A1, A2) is then also called the rough set of A. Without loss of generality of the discussion, we follow the latter approach. Rough set theory is strongly related to Gentilhomme’s (1968) flou set theory. A flou set in a universe X is a couple (A1, A2) satisfying A1DA2DX. The first coordinate A1 is called the certain area, the second coordinate A2 the area of maximal extension and A2}A1 the flou zone (or vague zone). The idea of dividing the universe into three areas is a very strong linkage between flou sets and rough sets. One could say that with the introduction of rough set theory an important next step forward was made, providing a semantically justifiable and computationally feasible way to construct the three areas involved, by approximating a set A by its lower and upper approximations. One can verify that for a reflexive relation R A � A which justifies the statement that all rough sets are flou sets. Interestingly enough, given a flou set (A1, A2) in X, it is not always possible to find an approximation space (X, R) such that (A1, A2) is a rough set in (X, R). To illustrate this, we rely on the fact that if the relation R is symmetrical then Að Þ � A � ðAÞ Example 1 Suppose that A2 contains one more element than A1, i.e. A2 ¼ A1,{x0}, x0 in X \A1. If (A1, A2) is the rough set of A then A1 ¼ A � A � A ¼ A2 Hence either A1 ¼ ACA2 or A1CA ¼ A2. In the first case A1 ¼ A implies A ¼ A. Hence ðAÞ ¼ �AA and, because of the symmetry of R, �AA DA, which contradicts AC �AA ¼ A2. In the second case a similar line of reasoning leads to the contradiction ADA. Iwinski rough sets The second stream in the literature was initiated by Iwinski (1987), who did not use an equivalence Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 261 EXSY : 02005003 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES UN CO RR EC TE D PR O O F relation as an initial building block to define the rough set concept. Instead he departed from a complete subalgebra B of the Boolean algebra P(X) of subsets of X and defined a rough set as a pair of sets (A, B) with A in B, B in B and ADB. At the mathematical level this approach coincides with the one outlined by Pawlak (B being the class of definable sets) (Iwinski, 1987). At the semantic level there is a significant difference: although Iwinski’s formulation provides an elegant mathematical model (Yao, 1996), the absence of the equivalence relation and the set of examples to be approximated makes it hard to interpret. Despite its short history, rough set theory has already been applied successfully in many knowledge-based systems. We refer to (Komorowski et al., (1999) for an extensive overview of achievements in the field. 2.2. L-fuzzy set theory From fuzzy setsy Fuzzy set theory was introduced by Zadeh (1945) as a framework for modelling the vagueness present in everyday life (and in particular, in natural language). A fuzzy set A in a universe X is characterized by an X-[0, 1] mapping, usually denoted by A or by mA, called the membership function. For all x in X, A(x)corresponds to the degree to which x belongs to A. While fundamental research on the theory is still very topical, the application of the fuzzy set theoretical representation of vague concepts also has many working applications nowadays. The rapid growth of the series Studies in Fuzziness and Soft Computing (Kacprzyk, 1992–2002), consisting of both monographs and edited volumes, is a striking illustration of the continuous evolution and the high number of achievements in this field. y to L-fuzzy sets The mapping of elements of the universe to the interval [0, 1], however, implies a crisp, linear ordering of these elements, making [0, 1]-valued fuzzy set theory inadequate to deal with incomparable information. From the beginning therefore some attention has been paid to other partially ordered sets as well. In 1967 Goguen formally introduced the notion of an L-fuzzy set with a membership function taking values in a lattice L. In this paper we assume that (L, rL) is a complete lattice with smallest element 0L and greatest element 1L. An L-fuzzy set A in a universe X is a mapping from X to L, again called the membership function. The L-fuzzy set A is said to be included in the L-fuzzy set B, usually denoted by ADB, if A(x)rL B(x) for all x in X. An L-fuzzy set R in X � X is called a binary fuzzy relation on X. For all x and y in X, R(x, y) expresses the degree to which x and y are related through R. For every y in X, the R-foreset of y is an L-fuzzy set in X, denoted as Ry and defined by Ry(x) ¼ R(x, y) for all x in X. Logical operators L-fuzzy set theoretical operations such as complement, intersection and union can be defined by means of suitable generalizations of the well-known connectives from Boolean logic. Negation, conjunction, disjunction and implication can be generalized respectively to negator, triangular norm, triangular conorm and implicator, all mappings taking values in L. More specifically, a negator in L is any decreasing L-L mapping N satisfying N(0L) ¼ 1L. It is called involutive if N(N(x)) ¼ x for all x in L. A triangular norm (t-norm for short) T in L is any increasing, commutative and associative L 2-L mapping satisfying T(1L, x) ¼ x, for all x in L. A triangular conorm (t-conorm for short) S in L is any increasing, commutative and associative L 2-L mapping satisfying S(0L, x) ¼ x, for all x in L. The N- complement of an L-fuzzy set A in X as well as the T- intersection and the S-union of L-fuzzy sets A and B in X are the L-fuzzy sets coN(A), A-T B and A,S B defined by coNðAÞðxÞ ¼ NðAðxÞÞ A \T BðxÞ ¼ TðAðxÞ; BðxÞÞ A [S BðxÞ ¼ SðAðxÞ; BðxÞÞ for all x in X. The dual of a t-conorm S in L with respect to a negator N in L is a t-norm T in L defined as T(x, y) ¼ N(S(N(x), N(y))). An implicator in L is any L2-L mapping I satisfying I(0L, 0L) ¼ 1L, I(1L, x) ¼ x for all x in L. Moreover we require I to be decreasing in its first, and increasing in its second, component. If S and N are respectively a t-conorm and a negator in L, then it is well known that the mapping IS,N defined by IS;Nðx; yÞ ¼ SðNðxÞ; yÞ is an implicator in L, usually called S-implicator (induced by S and N). Likewise, if T is a t-norm in L, the mapping IT defined by ITðx; yÞ ¼ supfljl 2 L and Tðx; lÞ �L yg is an implicator in L, usually called the residual implicator (of T). The partial mappings of a t-norm T in L are sup- morphisms if T sup i2I xi; y � � ¼ sup i2I Tðxi; yÞ for every family I of indexes. Examples It is easy to verify that the meet and the join operation on L are respectively a t-norm and a t-conorm on L. We denote them by TM and SM respectively. Also A-B is a shorter notation for A-TMB, while A,B corresponds to A,SMB. The [0, 1]-[0, 1] mapping Ns defined as Ns(x) ¼ 1 � x, for all x in [0, 1], is a negator on [0, 1], often called the standard negator. For a [0, 1]-fuzzy set A, coNsðAÞ is commonly denoted by co(A). Table 1 depicts the values of well-known t-norms and t-conorms on [0, 1], for all x and y in [0, 1]. The first column of Table 2 Q1 Q2 262 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 EXSY : 02005003 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES UN CO RR EC TE D PR O O F shows the values of the S-implicators on [0, 1] induced by the t-conorms of Table 1 and the standard negator Ns, while the second column lists the values of the corresponding residual implicators. Finally we mention that an L-fuzzy relation R on X is called an L-fuzzy T-equivalence relation if it is reflexive (R(x, x) ¼ 1L, for all x in X), symmetrical (R(x, y) ¼ R(y, x), for all x and y in X) and T-transitive (T(R(x, y), R(y, z))rR(x, z), for all x, y and z in X). When L ¼ {0, 1}, L-fuzzy set theory coincides with traditional set theory, in this context also called crisp set theory. {0, 1}-fuzzy sets, {0, 1}-fuzzy relations,y are usually also called crisp sets, crisp relations,y. When L ¼ [0, 1], fuzzy set theory in the sense of Zadeh is recovered. [0, 1]-fuzzy sets, [0, 1]-fuzzy relations,y are commonly called fuzzy sets, fuzzy relations,y. Furthermore it is customary to omit the indication ‘in [0, 1]’ when describing the logical operators, and hence to talk about negators, triangular norms etc. Yet another interesting choice for L gives rise to intuitionistic fuzzy set theory, as described below. 2.3. Intuitionistic fuzzy set theory A particularly interesting lattice of membership degrees leads to intuitionistic fuzzy set (IFS) theory (Atanassov, 1999). This theory basically defies the claim that, from the fact that an element x ‘belongs’ to a given degree (say mA(x)) to a fuzzy set A, it naturally follows that x should ‘not belong’ to A to the extent 1 � mA(x), an assertion implicit in the concept of a fuzzy set. On the contrary, IFSs assign to each element x of the universe both a degree of membership mA(x) and one of non-membership nA(x) such that mA(x) þ nA(x)r1, thus relaxing the enforced duality nA(x) ¼ 1 � mA(x) from fuzzy set theory. Obviously, when mA(x) þ nA(x) ¼ 1 for all elements of the universe, the traditional fuzzy set concept is recovered. By complementing the membership degree with a non- membership degree that expresses to what extent the element does not belong to the IFS, such that the sum of the degrees does not exceed 1, a whole spectrum of knowledge not accessible to fuzzy sets can be accessed. The applica- tions of this simple idea are manyfold indeed: it may be used to express positive as well as negative preferences; in a logical context, with a proposition a degree of truth and one of falsity may be associated; within databases, it can serve to evaluate the satisfaction as well as the violation of relational constraints. More generally, IFSs address the fundamental two-sidedness of knowledge, of positive versus negative information, and by not treating the two sides as exactly complementary (like fuzzy sets do), a margin of hesitation is created. This hesitation is quantified for each x in X by the number pA(x) ¼ 1 � mA(x) � nA(x). IFSs can be considered as special instances of L-fuzzy sets (Deschrijver et al., 2002). Let (L*,rL*) be the complete, bounded lattice defined by L� ¼ fðx1; x2Þ 2 ½0; 1�2jx1 þ x2 � 1g ðx1; x2Þ �L� ðy1; y2Þ , x1 � y1 and x2 � y2 The units of this lattice are denoted 0L* ¼ (0, 1) and 1L* ¼ (1, 0). For each element xAL*, by x1 and x2 we denote its first and second components, respectively. An IFS A in a universe X is a mapping from X to L*. For every xAX, the value mA(x) ¼ (A(x))1 is called the membership degree of x to A; the value nA(x) ¼ (A(x))2 is called the non-member- ship degree of x to A; and the value pA(x) is called the hesitation degree of x to A. Just as L*-fuzzy sets are called IFSs, L*-fuzzy relations are called IF relations. Logical operators The terms IF negator, IF t-norm, IF t- conorm and IF implicator are used to denote respectively a negator in L*, a t-norm in L*, a t-conorm in L* and an implicator in L*. A t-norm T on L* (respectively t-conorm S) is called t-representable (Deschrijver et al., 2002) if there exists a t-norm T and a t-conorm S on [0, 1] (respectively a t-conorm S0 and a t-norm T0 on [0, 1]) such that, for x ¼ (x1, x2), y ¼ (y1, y2)AL*, Tðx; yÞ ¼ ðTðx1; y1Þ; Sðx2; y2ÞÞ Sðx; yÞ ¼ ðS0ðx1; y1Þ; T0ðx2; y2ÞÞ T and S (respectively S0 and T0) are called the representants of T (respectively S). Finally, denoting the first projection mapping on L* by pr1, we recall from Deschrijver et al., (2002) that the [0, 1]-[0, 1] mapping N defined by N(a)¼ pr1N(a, 1� a) for all a in [0, 1] is an involutive negator on [0, 1], if N is an involutive negator on L*. N is called the negator induced by N. Furthermore N(x1, x2)¼ (N(1� x2),1� N(x1)), for all x in L*. Examples The standard IF negator is defined by Ns(x) ¼ (x2, x1), for all x in L*. The meet and the join Table 1: Triangular norms and conorms on [0, 1] t-norm t-conorm TM(x, y) ¼ min(x, y) SM(x, y) ¼ max(x, y) TP(x, y) ¼ x,y SP(x, y) ¼ x þ y � x,y TW(x, y) ¼ max(x þ y � 1, 0) SW(x, y) ¼ min(x þ y, 1) Table 2: S-implicators and residual implicators on [0, 1] S-implicator Residual implicator ISM;Nsðx; yÞ ¼ maxð1 � x; yÞ ITM ðx; yÞ ¼ 1 if x � y y else � ISP;Nsðx; yÞ ¼ 1 � x þ x; y ITP ðx; yÞ ¼ 1 if x � y y x else � ISW;Nsðx; yÞ ¼ minð1 � x þ y; 1Þ ITW ðx; yÞ ¼ minð1 � x þ y; 1Þ Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 263 EXSY : 02005003 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES UN CO RR EC TE D PR O O F operators on L* are respectively the IF t-norm TM and the IF t-conorm SM defined by TMðx; yÞ ¼ ðminðx1; y1Þ; maxðx2; y2ÞÞ SMðx; yÞ ¼ ðmaxðx1; y1Þ; minðx2; y2ÞÞ Combining TW and SW of Table 1 gives rise to the t- representable IF t-norm TW and IF t-conorm SW defined by TWðx; yÞ ¼ ðmaxð0; x1 þ y1 � 1Þ; minð1; x2 þ y2ÞÞ SWðx; yÞ ¼ ðminð1; x1 þ y1Þ; maxð0; x2 þ y2 � 1ÞÞ However, TL and SL are also possible extensions of TW and SW to IF theory: TLðx; yÞ ¼ ðmaxð0; x1 þ y1 � 1Þ; minð1; x2 þ 1 � y1; y2 þ 1 � x1ÞÞ SLðx; yÞ ¼ ðminð1; x1 þ 1 � y2; y1 þ 1 � x2Þ; maxð0; x2 þ y2 � 1ÞÞ They are not, however, t-representable. All of these IF t- conorms induce IF S-implicators ISM;Nsðx; yÞ ¼ðmaxðx2; y1Þ; minðx1; y2ÞÞ ISW;Nsðx; yÞ ¼ðminð1; x2 þ y1Þ; maxð0; x1 þ y2 � 1ÞÞ ISL; Nsðx; yÞ ¼ðminð1; y1 þ 1 � x1; x2 þ 1 � y2Þ; maxð0; y2 þ x1 � 1ÞÞ while the IF t-norms have residual IF implicators: ITMðx; yÞ ¼ 1L� if x1 � y1 and x2 � y2 ð1 � y2; y2Þ if x1 � y1 and x2oy2 ðy1; 0Þ if x14y1 and x2 � y2 ðy1; y2Þ if x14y1 and x2oy2 8>>< >>: ITWðx; yÞ ¼ ðminð1; 1 þ y1 � x1; 1 þ x2 � y2Þ; maxð0; y2 � x2ÞÞ ITL equals TsL;Ns. 2.4. Fuzzy rough set theory The two main streams in the perception of the notion ‘rough set’ both invoke generalizations to a ‘fuzzy rough set’ notion, intended to approximate a fuzzy set in a fuzzy approximation space, i.e. defined by a fuzzy relation. 1 Many people worked on the fuzzification of upper and lower approximations in the spirit of Pawlak (e.g. Nakamura, 1988; Dubois & Prade, 1990; Yao, 1997, 1998; Thiele, 1998; Radzikowska & Kerre, 2002). In doing so, the central focus moved from elements’ indistinguish- ability (with respect to their attribute values in an information system) to their similarity, again with respect to those attribute values: objects are categorized into classes with ‘soft’ boundaries based on their similarity to one another. A concrete advantage of such a scheme is that abrupt transitions between classes are replaced by gradual ones, allowing an element to belong (to varying degrees) to more than one class. An example at hand is an attribute in an information table that records ages: in order to restrict the number of equivalence classes, the classical rough set theory advises to discretize age values by a crisp partition of the universe, e.g. using intervals [0, 10], [10, 20],y. This does not always reflect our intuition, however: by imposing such harsh boundaries, a person who has just turned 11 will not be taken into account in the [0, 10] class, even when he is only at a minimal remove from full membership in that class. A general definition of fuzzy rough set, absorbing earlier suggestions in the same direction, was given by Radzi- kowska and Kerre (2002). Paraphrasing the following relations which hold in the crisp case y 2 A , ð8x 2 XÞððx; yÞ 2 R ) x 2 AÞ y 2 �AA , ð9x 2 XÞððx; yÞ 2 R and x 2 AÞ they define the lower and upper approximations of a fuzzy set A in X as the fuzzy sets A and �AA in X, constructed by means of an implicator I, a t–norm T and a fuzzy T- equivalence relation R in X: AðyÞ ¼ inf x2X IðRðx; yÞ; AðxÞÞ AðyÞ ¼ sup x2X TðRðx; yÞ; AðxÞÞ for all y in X. For an element y in X, its membership degree in the lower approximation of A is determined by looking at the elements resembling y (the foreset Ry) and by computing to what extent Ry is contained in A. Its membership degree in the upper approximation on the other hand is determined by the overlap between Ry and A. Technically, these operations amount to taking the super- direct image and the direct image respectively of A under the fuzzy relation R, (De Cock, 2002). A couple of fuzzy sets (A1, A2) is called a rough set in the approximation space (X, R, T, I) if there exists a fuzzy set A such that A ¼ A1 and �AA ¼ A2. A suggestion for fuzzification of the second stream view was launched by Nanda and Majumdar (1992). Bearing in mind Iwinski’s original view one would expect a fuzzy rough set to be a couple (A, B) of fuzzy sets A and B, both coming from some kind of algebra and such that ADB. However, Nanda and Majumdar chose to construct something that might be called ‘the fuzzy rough set of a rough set’. Starting from an Iwinski rough set, i.e. a couple (P, Q) of sets P and Q in X, a fuzzy rough set is a couple of fuzzy sets (A, B). A is a fuzzy set in P, while B is a fuzzy set in Q. Furthermore the first should be included in the second. When trying to express such a requirement one 1 Yao (1997) used ‘fuzzy rough set’ to denote the approximation of a crisp set in a fuzzy approximation space, whereas in his view a ‘rough fuzzy set’ gives an approximation of a fuzzy set in a crisp approximation space. Most authors, like us, use ‘fuzzy rough sets’ as a general term to cover ‘fuzzified rough set theory’. 264 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 EXSY : 02005003 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES UN CO RR EC TE D PR O O F already notices that it is not very convenient that both fuzzy sets are defined on different universes. This is why authors introduce their own ‘tricks’ for extending the universe (e.g. extending the universes of A and B to X (Chakrabarty et al., 1998; Çoker, 1998), extending the smaller universe (that of A) to the bigger one (that of B) (Samanta & Mondal, 2001)). In the literature we did not find any semantic motivation behind this kind of fuzzy rough sets, or any applications using them. Still for many authors this notion of fuzzy rough set seems to be an attractive starting point for the introduction of intuitionistic fuzzy rough sets (Chakrabarty et al., 1998; Samanta & Mondal, 2001; Jena & Ghosh, 2002). Before we go into that topic, however, we will focus on the links between fuzzy rough set theory and IFS theory. 3. Fuzzy rough set theory versus IFS theory Along with the lower and upper approximations, rough set theory provides two kinds of membership: if an element belongs to the lower (respectively upper) approximation of A, we are dealing with strong (respectively weak) member- ship of A (Pawlak, 1982). It is very natural to extend this idea to fuzzy rough set theory: the strong membership function of A is the membership function of the lower approximation of A, while the weak membership function of A is the membership function of the upper approxima- tion of A. In IFS theory we are also dealing with two kinds of functions, namely a membership function m and a non- membership function n such that m � coðnÞ ð1Þ Note that the strong membership function of the fuzzy rough set can be considered as the membership function m of an IFS, while the complement of the weak membership function of the fuzzy rough set can be used as the non- membership function n. This remark (Çoker, 1998) immediately reveals the strong link between rough set theory and IFS theory. One might even argue that IFS theory could really draw profit from fuzzy rough set theory, because the latter brings a means to construct the membership and non-membership functions of an IFS from a fuzzy set of examples and a fuzzy information relation (a fuzzy relation modelling similarity). It is a far greater challenge, however, to imagine what we could do if the input fed to a knowledge-based system carries information not only about the positive side but also about the negative side. How can we additionally benefit if the set of examples is IF, if the information relation is IF? To answer this question, in the next section we lift the study yet one level higher, by exploring different ways of introducing the concept ‘intuitionistic fuzzy rough set’. Keeping in mind what we learned in this section, we expect it to be ‘a couple of couples of functions’ (be it (strong or weak) membership functions, or non-membership func- tions). 4. Intuitionistic fuzzy rough sets A very natural way to extend concepts from fuzzy set theory to their generalizations in IFS theory is the replacement of [0, 1] by L* as the evaluation set for the membership degrees. � Doing so in Nanda and Majumdar’s view on fuzzy rough sets leads to Chakrabarty et al.’s (1998), approach to intuitionistic fuzzy rough sets; they construct an IF rough set (A, B) of a rough set (P, Q). A and B are both IFSs in X such that A � B; i:e: mA � mB and nA nB From this point of view the lower approximation A and the upper approximation B are both IFSs. In other words, the strong membership is itself characterized by a membership function mA and a non-membership function nA, while the membership function mB and the non-membership function nB together constitute the weak membership. In this way we can reflect hesitation on the strong and the weak membership. Jena and Ghosh (2002) reintroduce the same notion. � Samanta and Mondal (2001) also introduce this notion but they call it a rough IF set. Furthermore they also define their concept of IF rough set, looking at it from a different angle: in their approach an IF rough set is a couple (A, B) such that A and B are both fuzzy rough sets (in the sense of Nanda and Majumdar) and A is included in the complement of B, i.e. A � coðBÞ ð2Þ Even though we omit the details of the definition of complement and inclusion of fuzzy rough sets here (cf. (Samanta and Mondal (2001) for the full story), the reader can still compare (2) to (1) to see that A refers to membership of the IF rough set, while B corresponds to non-membership. � Obviously to Samanta and Mondal an intuitionistic fuzzy rough set (A, B) is a generalization of an IFS in which membership and non-membership functions are no longer fuzzy sets but fuzzy rough sets A and B. Note that for Chakrabarty et al. on the other hand an intuitionistic fuzzy rough set (A, B) is a generalization of a fuzzy rough set, in which upper and lower approximations are no longer fuzzy sets but IFSs A and B. � In contrast to the proposals above, the approach of Rizvi et al. (2002) is along the lines of Pawlak’s rough sets, explicitly indicating how lower and upper approx- imations of an IFS should be derived in an approxima- tion space. Rizvi et al. describe their proposal as ‘rough intuitionistic fuzzy set’, and in comparison with the Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 265 EXSY : 02005003 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES UN CO RR EC TE D PR O O F approaches above one could say that it is again about having a hesitation margin on the weak and strong memberships, i.e. on the upper and lower approxima- tions. The hesitation margin on the lower and upper approximations can be seen as a consequence of the hesitation margin of the original IFS to be approxi- mated. Remarkably enough, the lower and upper approximations themselves are not IFSs in X but IFSs in the class of equivalence classes of R! The question remains unanswered why the definition of Radzikowska and Kerre (2002) — which has existed already in a more specific form for more than a decade — did not undergo the natural transformation process towards intuitionistic fuzzy rough set theory until now. In the remainder of this paper it becomes clear that it in fact leads to a mathematically elegant and semantically interpretable concept whereas the above-mentioned pro- posals all suffer from various drawbacks making them less eligible for applications. Until now we have been using the notations A and �AA to denote the lower and the upper approximation of A (whatever A may be), tacitly assuming that the (fuzzy) relation as well as the t-norm and the implicator are used. This was done not only for ease of notation, but also to keep a uniform notation throughout all the different models. However, to maintain clearness in the remainder, we switch to a slightly different notation. Furthermore from now on we assume that T is an IF triangular norm, I an IF implicator and R an IF T- equivalence relation in X. Together they constitute the approximation space (X, R, T, I). We use A and B to denote IFSs in X. As usual we start by defining concepts of lower and upper approximation. Definition 1 The lower and upper approximations of A are respectively the IFSs RkIA and RmTA in X defined by R #I AðyÞ ¼ inf x2X IðRðx; yÞ; AðxÞÞ R "T AðyÞ ¼ sup x2X TðRðx; yÞ; AðxÞÞ for all y in X. We say that a couple of IFSs (A1, A2) is an intuitionistic fuzzy rough set in the approximation space (X, R,T, I) if there exists an IFS A such that RkIA ¼ A1 and RmTA ¼ A2. The lower approximation of A is included in A, while A is included in its upper approximation. Proposition 1 (De Cock, 2002) RkIADADRmTA. The following propositions describe how the lower and upper approximations behave with respect to a refinement of the IFS to be approximated, or a refinement of the IF relation that defines the approximation space. Proposition 2 (De Cock, 2002) If ADB then RkIAD RkIB and RmTADRmTB. Proposition 3 (De Cock, 2002) If R1 and R2 are IF T- equivalence relations such that R1DR2 then R1 #I A R2 #I A R1 "T A � R2 "T A We now define a concept of definability, similar to the one in rough set theory. Definition 2 A is called definable if and only if RkIA ¼ RmTA. In classical rough set theory, a set is definable if and only if it is a union of equivalence classes. This property no longer holds in intuitionistic fuzzy rough set theory. However, if we imply sufficient conditions on the IF t-norm T and the IF implicator I defining the approximation space, we can still establish a weakened theorem. For the proof we rely on the following two propositions. Proposition 4 (De Cock, 2002) If the partial mappings of T are sup-morphisms and I is the residual implicator of T then the following are equivalent (1) A ¼ RkIA (2) A ¼ RmTA Proposition 5 (De Cock, 2002) If the partial mappings of T are sup-morphisms and I is the residual implicator of T then R "T ðR "T AÞ ¼ R "T A R #I ðR #I AÞ ¼ R #I A Corollary 1 If the partial mappings of T are sup- morphisms and I is the residual implicator of T then R "T ðR #I AÞ ¼ R #I A R #I ðR "T AÞ ¼ R "T A Theorem 1 If the partial mappings of T are sup-morphisms and I is the residual implicator of T then any union of R- foresets is definable, i.e. ð9BÞ B � X and A ¼ [ z2B Rz ! implies R #I A ¼ R "T A Proof Due to the symmetry and the T-transitivity of R and the fact that the partial mappings of T are sup- morphisms, we obtain R "T AðyÞ ¼ sup x2X TðRðx; yÞ; AðxÞÞ ¼ sup x2X TðRðx; yÞ; sup z2B Rðz; xÞÞ ¼ sup z2B sup x2X TðRðx; yÞ; Rðz; xÞÞ � sup z2B Rðz; yÞ ¼AðyÞ 266 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 EXSY : 02005003 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES UN CO RR EC TE D PR O O F Proposition 1 now implies A ¼ RmTA. Combining this result with Proposition 4 we obtain the definability of A. The following example illustrates that the opposite of Theorem 1 does not generally hold, i.e. not every definable set corresponds to a union of R-foresets. Example 2 Let X ¼ {x0}, R(x0, x0) ¼ (1, 0), A(x0) ¼ (0.5, 0.5). Then R "T Aðx0Þ ¼ TðRðx0; x0Þ; Aðx0ÞÞ ¼ ð0:5; 0:5Þ ¼ Aðx0Þ That is, A is definable. Since the only R-foreset is Rx0 and Rx0(x0)>L* A(x0) we cannot rewrite A as a union of R- foresets. Under the same conditions implied on T and I as in Theorem 1, the SM-union and TM-intersection of two definable IFSs is definable. This is a corollary of the next proposition. Proposition 6 (De Cock, 2002) If the partial mappings of T are sup-morphisms and I is the residual implicator of T then R "T ðA [ BÞ ¼ R "T A [ R "T B R #I ðA \ BÞ ¼ R #I A \ R #I B with AB and A-B defined as in Section 2.2, i.e. A [ BðxÞ ¼ ðminðmAðxÞ; mBðxÞÞ; maxðnAðxÞ; nBðxÞÞÞ A \ BðxÞ ¼ ðmaxðmAðxÞ; mBðxÞÞ; minðnAðxÞ; nBðxÞÞÞ Proposition 7 Let T be a t-representable IF t-norm such that T ¼ (T, S) and such that S(x, y) ¼ 1 � T(1 � x, 1 � y), and let R1, R2 be two fuzzy T-equivalence relations such that R1(x, y)rR2(x, y), for all x and x in X. Then R defined by Rðx; yÞ ¼ ðR1ðx; yÞ; 1 � R2ðx; yÞÞ for x and y in X, is an IF T-equivalence relation. Proof Reflexivity and symmetry of R follow immediately from the corresponding properties of R1 and R2. To prove T-transitivity, let x, y and zAX. TðRðx; yÞ; Rðy; zÞÞ ¼ðTðR1ðx; yÞ; R1ðy; zÞÞ; Sð1 � R2ðx; yÞ; 1 � R2ðy; zÞÞÞ ¼ðTðR1ðx; yÞ; R1ðy; zÞÞ; 1 � TðR2ðx; yÞ; R2ðy; zÞÞÞ �L� ðR1ðx; zÞ; 1 � R2ðx; zÞÞ ¼Rðx; zÞ Example 3 Let X ¼ [0, 100], and let the fuzzy TW- equivalence relation Ec on X be defined by Ecðx; yÞ ¼ max 1 � jx � yj c ; 0 � � for all x and y in X, and with real parameter c>0. Obviously, if c1rc2 then Ec1ðx; yÞ � Ec2ðx; yÞ. By Proposi- tion 7, ðEc1; coðEc2ÞÞ is an IF TW-equivalence relation. Example 4 Figure 1 shows the membership function mA and the non-membership function nA of the IFS A in the universe X ¼ [0, 100]. Using the non-t-representable IF t- norm TL, its residual IF implicator ITL and the IF relation R defined by Rðx; yÞ ¼ ðE40ðx; yÞ; 1 � E40ðx; yÞÞ for all x and y in [0, 100] we computed the lower approximation of A( ¼ A1) as well as the upper approxima- tion of A( ¼ A2). They are both depicted in Figure 1. An IFS A is characterized by means of a membership function mA and a non-membership function nA. A natural question which arises is whether the lower approximation and the upper approximation of A could be defined in terms of the lower and the upper approximation of mA and nA (all within the proper approximation spaces of course). Generally such a ‘divide and conquer’ approach is every- thing but trivial in IFS theory, and sometimes even impossible. However, some conditions implied on the logical operators involved can allow for some results in this direction. Particularly attractive are the t-representable t- norms and t-conorms, and the S-implicators that can be associated with them. Lemma 1 For every family (ai, bi)iAI in L* sup i2I ðai; biÞ ¼ sup i2I ai; inf i2I bi � � inf i2I ðai; biÞ ¼ inf i2I ai; sup i2I bi � � Proposition 8 Let T be t-representable such that T ¼ (T, S), let N be an involutive negator on [0, 1], and let I Figure 1: A (solid lines), and the upper (broken lines) and lower (dotted lines) approximations of A. Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 267 EXSY : 02005003 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES UN CO RR EC TE D PR O O F be the S-implicator on [0, 1] induced by S and N. Then R "T A ¼ ðmR "T mA; ðcoNðnRÞÞ #I nAÞ Proof For all y in X R "T AðyÞ ¼ sup x2X TðRðx; yÞ; AðxÞÞ ¼ sup x2X ðTðmRðx; yÞ; mAðxÞÞ; SðnRðx; yÞ; nAðxÞÞÞ ¼ðsup x2X TðmRðx; yÞ; mAðxÞÞ; inf x2X IðNðnRðx; yÞÞ; nAðxÞÞÞ ¼ððmR "T mAÞðyÞ; ðcoNðnRÞ #I nAÞðyÞÞ Proposition 9 Let S be a t-representable IF t-conorm such that S ¼ (S, T), let N be an involutive IF negator, let I be the IF S-implicator induced by S and N, let N be the negator on [0, 1] induced by N and let I be the S-implicator induced by S and N. Then R #I A ¼ ððconRÞ #I mA; coðcoNmRÞ "T nAÞ Proof For all y in X R #I AðyÞ ¼ inf x2X IðRðx; yÞ; AðxÞÞ ¼ inf x2X SðNðRðx; yÞÞ; AðxÞÞ ¼ inf x2X SððNð1 � nRðx; yÞÞ; mAðxÞÞ; supx2XTð1 � NðmRðx; yÞÞ; nAðxÞÞÞ ¼ðinf x2X IðconRðx; yÞ; mAðxÞÞ; sup x2X TðcoðcoNðmRÞÞðx; yÞÞ; nAðxÞÞÞ ¼ððconR #I mAÞðyÞ; ðcoðcoNðmRÞÞ "T nAÞðyÞÞ Observe that in both propositions on the ‘fuzzy level’ the approximations are taken under the membership function mR, or something semantically very much related such as the N-complement of the non-membership function nR or once even the standard complement of the N-complement of mR. Presented in this way, the resulting formulae look quite complicated. For better understanding, assume for a moment that the negator N is the standard negator, and that the IF relation R is in reality a fuzzy relation, i.e. mR ¼ co(nR); then the formulae reduce to R "T A ¼ ðmR "T mA; mR #I nAÞ R #I A ¼ ðmR #I mA; mR "T nAÞ Apparently in this case the membership function of the upper approximation of A is the upper approximation of the membership function of A, while the non-membership function of the upper approximation of A is the lower approximation of the non-membership function of A. For the lower approximation of A the dual proposition holds. Example 5 Figure 2 shows the same IFS A we used in Example 4. However, to compute its lower approximation A1 and its upper approximation A2, this time we used the t- representable IF t-norm TW, its residual IF implicator and the IF relation R defined by Rðx; yÞ ¼ ðE30ðx; yÞ; E50ðx; yÞÞ for all x and y in [0,100]. 5. Conclusion Rough sets and IFSs both capture particular facets of the same notion—imprecision. In this paper, it was shown how they can be usefully combined into a single framework encapsulating the best of (so far) largely separate worlds. The link, on the syntactical level, between fuzzy rough sets and IFSs, identified by Çoker, has not proven much of an obstacle in this sense: indeed, by allowing each ingredient to retain its own distinguishing semantics, it was possible to create an end product which is both syntactically sound and semantically meaningful. We feel especially justified in our cause since we have exploited to the fullest a time- honoured adage: namely, that to every object there is a positive and a negative side which need to be addressed individually in order to come up with a true representation of that object. Future work will involve tailoring this general framework to the specific needs of everyday applications. Acknowledgement Chris Cornelis and Martine De Cock would like to thank the Fund for Scientific Research Flanders—FWO for funding the research reported in this paper. Figure 2: Upper (broken lines) and lower (dotted lines) approximations of A. 268 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 EXSY : 02005003 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES UN CO RR EC TE D PR O O F References ATANASSOV, K.T. (1999) Intuitionistic Fuzzy Sets, Heidelberg: Physica. CHAKRABARTY, K., T. GEDEON and L. KOCZY (1998) Intuitionistic fuzzy rough sets, in Proceedings of the Fourth Joint Conference on Information Sciences, 211–214. ÇOKER, D. (1998) Fuzzy rough sets are intuitionistic L-fuzzy sets, Fuzzy Sets and Systems, 96, 381–383. DE COCK, M. (2002) A thorough study of linguistic modifiers in fuzzy set theory, PhD thesis (in Dutch), Ghent University. DESCHRIJVER, G., C. CORNELIS and E.E. KERRE (2002) Intuitio- nistic fuzzy connectives revisited, in Proceedings of the 9th International Conference on Information Processing and Manage- ment of Uncertainty in Knowledge-based Systems, 1839–1844 DUBOIS, D. and H. PRADE (1990) Rough fuzzy sets and fuzzy rough sets, International Journal of General Systems, 17, 191–209. GENTILHOMME, Y. (1968) Les ensembles floues en linguistique, Cahiers de Linguistique Théorique, 5, 47–63. GOGUEN, J. (1967) L-fuzzy sets, Journal of Mathematical Analysis and Applications, 18, 145–174. IWINSKI, T.B. (1987) Algebraic approach to rough sets, Bulletin of the Polish Academy of Science and Mathematics, 35, 673–683. JENA, S.P. and S.K. GHOSH (2002) Intuitionistic fuzzy rough sets. Notes on Intuitionistic Fuzzy Sets 8, 1, 1–18. KACPRZYK, J. (ed.) (1992–2002) Studies in Fuzziness and Soft Computing, Heidelberg: Physica, Vols 1–100. KOMOROWSKI, J., Z. PAWLAK, L. POLKOWSKI and A. SKOWRON (1999) Rough sets: a tutorial, in Rough Fuzzy Hybridization: A New Trend in Decision-Making, S.K. Pal and A. Skowron (eds), Singapore: Springer. NAKAMURA, A. (1988) Fuzzy rough sets, Note on Multiple-Valued Logic in Japan, 9, 1–8. NANDA, S. and S. MAJUMDAR (1992) Fuzzy rough sets, Fuzzy Sets and Systems, 45, 157–160. PAWLAK, Z. (1982) Rough sets, International Journal of Computer and Information Sciences, 11 (5), 341–356. PAWLAK, Z. (1985) Rough sets and fuzzy sets, Fuzzy Sets and Systems, 17, 99–102. RADZIKOWSKA, A.M. and E.E. KERRE (2002) A comparative study of fuzzy rough sets, Fuzzy Sets and Systems, 126, 137–156. RIZVI, S., H.J. NAQVI and D. NADEEM (2002) Rough intuitionistic fuzzy sets, in Proceedings of the 6th Joint Conference on Information Sciences, H.J. Caulfield, S. Chen, H. Chen, R. Duro, V. Honavar, E.E. Kerre, M. Lu, M.G. Romay, T.K. Shih, D. Ventura, P.P. Wang and Y. Yang (eds), 101–104. SAMANTA, S.K. and T.K. MONDAL (2001) Intuitionistic fuzzy rough sets and rough intuitionistic fuzzy sets, Journal of Fuzzy Mathematics, 9 (3), 561–582. THIELE, H. (1998) Fuzzy rough sets versus rough fuzzy sets—an interpretation and a comparative study using concepts of modal logic, Technical Report ISSN 1433-3325, University of Dort- mund. THIELE, H. (2001) Generalizing the explicit concept of rough set on the basis of modal logic, in Computational Intelligence in Theory and Practice, B. Reusch and K.H. Temme (eds), Berlin: Physica. YAO, Y.Y. (1996) Two views of the theory of rough sets in finite universes, International Journal of Approximate Reasoning, 15 (4), 291–317. YAO, Y.Y. (1997) Combination of rough and fuzzy sets based on alpha-level sets, in Rough Sets and Data Mining: Analysis for Imprecise Data, T.Y. Lin and N. Cercone (eds), Baston, MA: Kluwer Academic, 301–321. YAO, Y.Y. (1998) A comparative study of fuzzy sets and rough sets, Information Sciences, 109, 227–242. ZADEH, L.A. (1965) Fuzzy sets, Information and Control, 8, 338–353. The authors Chris Cornelis Chris Cornelis received the BSc and MSc degrees in computer science from Ghent University, Belgium, in 1998 and 2000, respectively, and is currently working towards his PhD degree, sponsored by a grant from the National Science Foundation—Flanders. His research interests involve mathematical models of imprecision, including interval-valued fuzzy sets, intuitionistic fuzzy sets, L-fuzzy sets, rough sets and various combinations of these. He has applied these models in approximate reasoning, possibility theory and data mining. Martine De Cock Martine De Cock received an MSc degree in computer science and a teaching degree at Ghent University in 1998. In 2002 she obtained a PhD in computer science with a thesis on the representation of linguistic modifiers in fuzzy set theory. Currently she is working in the Fuzziness and Uncertainty Modelling Research Unit (Ghent University) as a postdoctoral researcher of the Fund of Scientific Research—Flanders (FWO). Her current research interests include rough set theory, L-fuzzy set theory, fuzzy association rules, and the representation of approximate equality. Etienne E. Kerre Etienne E. Kerre was born in Zele, Belgium, on 8 May 1945. He obtained his MSc degree in mathematics in 1967 and his PhD in mathematics in 1970 from Ghent University. Since 1984, he has been a lecturer, and since 1991 a full professor, at Ghent University. He is a referee for more than 30 international scientific journals, and also a member of the editorial board of international journals and conferences on fuzzy set theory. He has been an honorary chairman at various international conferences. In 1976, he founded the Fuzziness and Uncertainty Modelling Re- search Unit (Ghent University) and since then his research has been focused on the modeling of fuzziness and uncertainty, and has resulted in a great number of contributions in fuzzy set theory and its various general- izations, and in evidence theory. He has been particularly involved in the theories of fuzzy relational calculus and of fuzzy mathematical structures. Over the years he has also been a promotor of 16 PhDs on fuzzy set theory. His current research interests include fuzzy and intuitionistic fuzzy relations, fuzzy topology and fuzzy image processing. He has authored or co-authored 11 books, and more than 100 papers in international refereed journals. Q3 Q4 Q5 Q6 Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 269 EXSY : 02005003 BWUK EXSY 02005003.PDF 08-Aug-03 16:38 226131 Bytes 10 PAGES 
work_3czyinsxobhzjdakbts7whwj5q	----	Solid/liquid separation equipment simulation & design – an expert systems approach Browse Explore more content Tarleton 1.pdf (260.79 kB) Solid/liquid separation equipment simulation & design – an expert systems approach CiteDownload (260.79 kB)ShareEmbed journal contribution posted on 30.09.2009, 13:01 by Richard J. Wakeman, Steve Tarleton Published texts on solid/liquid separation technology have allowed a limited amount of knowledge to be available widely, but none put forward the rules of thumb in such a manner that they are readily assimilable by the non-expert. There are many unpublished techniques and approaches. The computer technologist might target solid/liquid separation as a technology ripe for the application of expert systems. It is argued here that expert systems on their own are inadequate in this subject area, but the most effective software utilises a well-chosen mix of algorithm, graphics, expert system, and interactive input from the engineer. In this paper some examples are given of both public and private knowledge and an example of the application of the combined approach to equipment selection demonstrates that efficient software can save time and enable decisions to be made rapidly. Equipment selection using the pC-SELECT software is demonstrated. Categories Chemical Engineering not elsewhere classified Keywords untagged History School Aeronautical, Automotive, Chemical and Materials Engineering Department Chemical Engineering Citation WAKEMAN, R.J. and TARLETON, E.S., 1991. Solid/liquid separation equipment simulation & design – an expert systems approach. Filtration & Separation, 28 (4), pp. 268-274 Publisher © Elsevier Version AM (Accepted Manuscript) Publication date 1991 Notes This article was published in the journal, Filtration & Separation [© Elsevier] and the definitive version is available at: http://dx.doi.org/10.1016/0015-1882(91)80118-O DOI https://doi.org/10.1016/0015-1882(91)80118-O ISSN 0015-1882 Language en Administrator link https://repository.lboro.ac.uk/account/articles/9242387 Licence CC BY-NC-ND 4.0 Exports Select an optionRefWorksBibTeXRef. managerEndnoteDataCiteNLMDC Categories Chemical Engineering not elsewhere classified Keywords untagged Licence CC BY-NC-ND 4.0 Exports Select an optionRefWorksBibTeXRef. managerEndnoteDataCiteNLMDC Hide footerAboutFeaturesToolsBlogAmbassadorsContactFAQPrivacy PolicyCookie PolicyT&CsAccessibility StatementDisclaimerSitemap figshare. credit for all your research. 
work_3f36lrvqwjdjdmkat65ktfoney	----	RRS1_2017.indd Romanian Statistical Review nr. 1 / 2017 3 Expert System and Heuristics Algorithm for Cloud Resource Scheduling Mamatha E (sricsrmax@gmail.com) Dept of Engineering Mathematics, GITAM University, Bangalore, India Sasritha S Dept of Engineering Mathematics, GITAM University, Bangalore, India CS Reddy School of Computing, SASTRA University, Thanjavur, India ABSTRACT Rule-based scheduling algorithms have been widely used on cloud comput- ing systems and there is still plenty of room to improve their performance. This paper proposes to develop an expert system to allocate resources in cloud by using Rule based Algorithm, thereby measuring the performance of the system by letting the sys- tem adapt new rules based on the feedback. Here performance of the action helps to make better allocation of the resources to improve quality of services, scalability and fl exibility. The performance measure is based on how the allocation of the resources is dynamically optimized and how the resources are utilized properly. It aims to maxi- mize the utilization of the resources. The data and resource are given to the algorithm which allocates the data to resources and an output is obtained based on the action occurred. Once the action is completed, the performance of every action is measured that contains how the resources are allocated and how effi ciently it worked. In addition to performance, resource allocation in cloud environment is also considered. Keywords: Cloud computing, Scheduling and Expert System, Heuristic Models JEL Classifi cation: C87 INTRODUCTION Cloud Computing, the long-held dream of computing industry, has the capability to change large IT industry, making software even more usable as a service and changing the way IT hardware is made and purchased. Developers with creative ideas for new Internet services no longer need the large capital costs in hardware to install their service or the human expense to operate it [1-5]. They need not be bothered about over-provisioning for a service whose attractiveness does not meet their predictions, thus wasting the Romanian Statistical Review nr. 1 / 20174 resources, or under-provisioning for one that becomes wildly accepted, thus missing potential customers and revenue. The prominent feature of cloud computing that it allows the consumption of services over internet with subscription based model. Based on the level of abstraction, various models for cloud computing like Infrastructure- as-a-Service (IaaS), Platform-as-a-Service (PaaS) and Software-as-a-Service (SaaS). This model of service consumption is extremely suitable for many workloads and cloud computing has become highly successful technology [5]. It allows its users to pay for what they use and also remove the upfront infrastructure cost. Cloud service providers receive resource requests from a number of users through the use of virtualization. It is very essential for cloud providers to operate very effi ciently in multiplexing at the scale to remain profi table. The increase in the use of cloud computing has risen to develop massive data centers with very large number of servers. The resource management at this scale is the concerned issue. Scheduling is responsible for arbitrate of resources and is at the center of resource management. The issue of effi ciency at this rate and developing model of consumption of cloud providers needs new approaches and techniques to be applied to the age old problem of scheduling. Virtual machine is the primary unit of scheduling in this model. In this study, we deal with problem of virtual machine scheduling over physical machines. We aim to understand and solve the various aspect of scheduling in cloud environments [6,7]. Specifi cally, we leverage different fi ne grained monitoring information to make better scheduling decisions and used learning based approach of scheduling in widely different environments. There is increasing concern over energy consumption by cloud data centers and cloud operators are focusing on energy savings through effective utilization of resources [8,9]. SLAs is also very important for them for the performance of applications running. We propose algorithms which try to minimize the energy consumption in the data center duly maintaining the SLA ensures. The algorithms try to use very less number of physical machines in the data center by dynamically rebalancing the physical machines based on their resource utilization. The algorithms also do an optimization of virtual machines on a physical machine, reducing SLA violations. For large scale distributed system autonomic management [10,11] is one of the most wanted features and even more important in dynamic infrastructures such as Clouds. This is self-managing such as self-healing, self-regulating, self-protecting, and self-improving. Good work in both academia and industry has been already carried out. Tsai [12] reported an overview of the early efforts in developing Metaheuristic scheduling for autonomic systems for storage management. Computing Grids have benefi ted Romanian Statistical Review nr. 1 / 2017 5 from the application of autonomic models for management of resources and the scheduling of applications [13,14,15]. Solutions for secure Cloud platforms have been proposed in the literature [16,17]. However, existing works are yet to address issues related to recognition of attacks against SaaS with the aim of exploiting elasticity. The basic idea of the proposed algorithm is to leverage the strengths of heuristic algorithms, such as swarm optimization [18], Scheduling Computer and Manufacturing Processes [19], Hadoop map- task scheduling [21,22] and ant colony optimization [20], by integrating them into a single algorithm. One of the important tasks for cloud provider is scheduling data packets for better achievement and effi cient resource pooling along with elasticity. In cloud computing scenarios scheduling algorithm becomes very signifi cant where the cloud service providers have to operate at very much competent to be competitive and take advantage at scale[23,24]. The wide acceptance of cloud computing means data centers with many more machines and the usage model is much different than traditional clusters, like hour boundaries, auction based prices are to name a few. Thus scheduling in cloud data center is more challenging than traditional cluster schedulers. Also, these data center run many different kinds of applications with varying expectations from infrastructure. Resource usage patterns in traditional data centers are have less variance in than the unpredictability faced by cloud data centers. PROBLEM ANALYSIS AND DEFINITION Since Rule based algorithms are straight forward and can be easily implemented, so there is lot of possibility to improve the performance of these algorithms especially in cloud environments. Conventional models and its corresponding algorithms are pretty good for small scale environments, however owing to the advent improvement of computer and related internet technological improvements; there is a huge range to enhancements in these algorithms to get better performance in large scale system. Extreme use of number of servers in the recent period has reduced the usage of traditional scheduling techniques. The resource management at this scale is the concerned as an issue. Scheduling is responsible for arbitrate of resources and is at the center of resource management. The issue of effi ciency at this rate and developing model of consumption of cloud providers wants new methodologies and techniques that to be extended to apply to presently available old algorithms of scheduling. The main objective of the paper is to develop and implement more effi cient scheduling algorithm suitable for cloud system. Since parallel and Romanian Statistical Review nr. 1 / 20176 distributed computing technologies was widely used to improve the diversity performance of computer systems, a number of models and thoughts have been anticipated for different approaches and congenital limitations in diverse eras. Whatever it may be the contemplation it is for; the way to competently use computer resources is a key research matter. Among all of them, scheduling is indispensable in the success of escalating the performance of the computing system. The wide acceptance of cloud computing means data centers with many more machines and the usage model is much different than traditional clusters, like hour boundaries, auction based prices are to name a few. Thus scheduling in cloud data center is more challenging than traditional cluster schedulers. PRESENT SYSTEM AND PROPOSED SCHEDULING ALGORITHM In this research paper, by the simple scheduling problems, we mean problems for which all the solutions can be checked in a reasonable time by using classical exhaustive algorithms running on modern computer systems. In comparison, with the large scale scheduling problems, like the problems for which not all the solutions can be examined in a reasonable time by using the same algorithms running on the same computer systems. These observations make it easy to understand that exhaustive algorithms will take a prohibitive amount of time to check all the candidate solutions for large scheduling problems because the number of candidate solutions is simply way too large to be checked in a reasonable time. As a result, researchers have paid their attention to the development of scheduling algorithms that are effi cient and effective, such as heuristics. Workfl ow is used with the automation of procedures where by fi les and data are passed between participants according to a defi ned set of rules to achieve an overall goal. A workfl ow management system is the one which manages and executes workfl ows on computing resources. Workfl ow Scheduling: It is a kind of global task scheduling as it focuses on mapping and managing the execution of inter-dependent tasks on shared resources that are not directly under its control. The authors classify and review hyper-heuristic approaches into the following four categories: based on the random choice of low level heuristics, greedy and puckish, meta heuristic-based, and those employing learning mechanisms to manage low level heuristics. The hyper heuristics can be used to operate at a higher level of abstraction. Meta heuristic techniques are expensive techniques that require knowledge in problem domain and heuristic technique. Hyper heuristic technique does not require problem Romanian Statistical Review nr. 1 / 2017 7 specifi c knowledge. In order to solve hard computational search problems the hyper heuristic techniques can be used. The hyper heuristic techniques can be operated on the search space of heuristics. Proposed System: The basic idea of the proposed algorithm is to use the diversity detection and improvement detection operators to balance the intensifi cation and diversifi cation in the search of the solutions during the convergence process. The proposed algorithm, called hyper-heuristic scheduling algorithm (HHSA).The parameters max and ni, where max denotes the maximum number of iterations the selected low-level heuristic algorithm is to be run; ni the number of iterations the solutions of the selected low- level heuristic algorithm are not improved. Line 2 reads in the tasks and jobs to be scheduled, i.e., the problem question. Line 3 initializes the population of solutions Z = fz1; z2;:::; zNg, where N is the population size. Online 4, a heuristic algorithm Hi is randomly selected from the candidate pool H = fH1;H2; :::;Hng. Hyper-heuristics are high level problem independent heuristics that work with any set of problem dependent heuristics and adaptively apply and combine them to solve a specifi c problem. This could be due to the fact that variants of differential evolution, which we mainly use as basic heuristics due to their competitive performance and simple confi guration, strongly depend on the population distribution. Hyper-heuristics might be regarded as a special form of genetic programming, the key intuition underlying research in this area is that, for a given type of problem, there are often a number of straightforward heuristics already in existence that can work well (but perhaps not optimally) for certain sorts of instances of that type of problem. Perhaps it is possible to combine those existing heuristics into some more elaborate algorithm that will work well across a range of problems. Cloud Computing denotes both the applications delivered as services, the hardware and systems software in the datacenters that provide those services. The services has been refered to as Software as a Service (SaaS). The datacenter hardware and software together combinedly called a Cloud. When a Cloud is of the kind pay-as-you-go manner to the public, then it is a Public Cloud; if the service is being sold then it is a Utility Computing. The term Private Cloud refers to internal datacenters of a business or other organizations, thart are not made available to the public. Thus, Cloud Computing is the combination of SaaS and Utility Computing, but does not comprise Private Clouds. Anyone can be users or providers of Software as a Service, and users or providers of Utility Computing. Romanian Statistical Review nr. 1 / 20178 Advantage:  Statistical multiplexing is a method that can be used to maximize utilization when compared to a private cloud.  Cloud computing can also offer services at a cost of a medium- sized datacenter and can even make a good profi t. Disadvantage:  The one important disadvantage is the chance of data loss.  User can leave the site permanently after facing a poor service; this negative feedback may result in a permanent loss of a portion of the revenue stream. WORKFLOW ENGINE: A Workfl ow engine is a software service that is used to provide the run-time environment in order to create, maintain and develop workfl ow instances. The representation of a workfl ow process is in a form which supports automated manipulation. Invoked Applications: Interfaces to support interaction with a variety of IT applications Workfl ow Client Applications: Interfaces to support interaction with the user interface. Administration and Monitoring: Interfaces to provide system monitoring and metric functions to facilitate the management of composite workfl ow application environments. It can be seen that scheduling is a functional module of a Workfl ow Engine, becoming a signifi cant part of workfl ow management systems. Workfl ow is concerned with the automation of procedures whereby fi les and data are passed between Participants according to a defi ned set of rules to achieve an overall goal. A workfl ow management system defi nes, maintains and develops workfl ows on computing resources. VIRTUAL GRID EXECUTION SYSTEM (VGES): vgES provides an uniform qualitative resource abstraction over grid and cloud systems. We apply vgES for scheduling a set of deadline sensitive weather forecasting workfl ows. Specifi cally, this paper reports on our experiences with: 1. Virtualized reservations for batch queue systems, 2. Coordinated usage of Tera Grid (batch queue),Amazon EC2 (cloud), our own clusters (batch queue) and Eucalyptus(cloud) resources, and Romanian Statistical Review nr. 1 / 2017 9 3. Fault tolerance through automated task replication. The combined effect of these techniques was to enable a new workfl ow planning method to balance the performance, reliability and the cost considerations. These results point toward improved resource selection and execution management support for a variety of e-Science applications over grids and cloud systems. This paper brings together many of the results from the VGrADS project demonstrating the effectiveness of virtual grids for scheduling LEAD workfl ows. In the process, it demonstrates a seamless merging of cloud and HPC resources in service of a scientifi c application. It also applies advanced scheduling techniques for both performance improvement and fault tolerance in a realistic context. LEAD has been run as a distributed application since its inception, but VGrADS methods have opened new capabilities for resource management and adaptation in its execution. This paper details the vgES implementation of virtual grids and their use in fault tolerant workfl ow planning of workfl ow sets with time and accuracy constraints. Our experiments show the effi ciency of the implementation and the effectiveness of the overall approach. Modules Description: A simple random method is used to select the low-level heuristic Hi from the candidate pool H. The diversity detection operator is used by HHSA to decide “when” to change the low-level heuristic algorithm Hi. This mechanism implies that the higher the temperature, the higher the opportunity to escape from the local search space to fi nd better solutions. The timer is fi xed at startup and end up mode of an application. We associate with such an event the workload dependent. Hyper heuristic Algorithm: Time-based consists in setting a timer that schedules the time instant when the Scheduled task has to be performed. The timer is fi xed at startup mode of an application. Hyper-heuristic algorithms can then maintain a high search diversity to increase the chance of fi nding better solutions at later iterations while not increasing the computation time. A time-based rejuvenation policy intends to identify the optimal time to rejuvenate with respect to one or more performance indices. The VMM does not degrade, and therefore, it is only necessary to keep memory of the age that was reached at the workload changing point. Romanian Statistical Review nr. 1 / 201710 Class Diagram to Scheduling Algorithm Sequence Diagram for Cloud Scheduling Algorithm Romanian Statistical Review nr. 1 / 2017 11 ALGORITHM DESCRIPTION: A hyper-heuristic is a heuristic search method that learns to automate, repeatedly by the incorporation of machine learning techniques, in which the process of selecting, combining, and generating or adapting a number of simpler heuristics (or components of such heuristics) to effi ciently solve the computational search problems. One of the main motivations for studying the hyper-heuristics is to build systems that which can handle the classes of problems rather than solving just a single problem. There might be multiple heuristics from which one can choose for solving the problem, and then the each heuristic has its own strength and also the weakness. The idea is to automatically devise algorithms by combining the strength and compensating for the weakness which are known heuristics. In the typical hyper-heuristic framework there consists of the high-level methodologies and a set of low-level heuristics (either constructive or perturbative heuristics). When a problem instance is given, the high-level method selects which low-level heuristic should be applied at any of the given time, based upon the current problem state, or the search stage. Romanian Statistical Review nr. 1 / 201712 Scheduling Flow Diagram Primitive model of Server Side Sample Code: using System; using System.Collections.Generic; using System.ComponentModel; using System.Data; using System.Drawing; using System.Linq; using System.Text; using System.Windows.Forms; using System.Data.SqlClient; namespace server { Romanian Statistical Review nr. 1 / 2017 13 public partial class Form2 : Form { SqlConnection cn = new SqlConnection(“Data Source=SRUJAN\\ SQLEXPRESS;Initial Catalog=schedule;Integrated Security=True”); SqlCommand cmd; SqlDataReader dr, dr1; public Form2() { InitializeComponent(); } private void Form2_Load(object sender, EventArgs e) { cn.Open(); } private void button6_Click(object sender, EventArgs e) { //+cal1(); //string sta = rsc + “N State”; //textBox6.Text = sta.ToString(); } private void button4_Click(object sender, EventArgs e) { } string t, t1, t2; private void button2_Click(object sender, EventArgs e) { SqlCommand cmd = new SqlCommand(“select * from vm1load”, cn); SqlDataReader dr = cmd.ExecuteReader(); while (dr.Read()) { t1 = dr[0].ToString(); } dr.Dispose(); cmd.Dispose(); if (t1 == null) { MessageBox.Show(“Start sharinf fi le”); } else Romanian Statistical Review nr. 1 / 201714 { textBox7.Text = t1.ToString(); } } private void button3_Click(object sender, EventArgs e) { cmd = new SqlCommand(“select * from vm2load”, cn); dr = cmd.ExecuteReader(); while (dr.Read()) { t2 = dr[0].ToString(); } dr.Dispose(); cmd.Dispose(); if (t2 == null) { MessageBox.Show(“Start sharinf fi le”); } else { textBox1.Text = t2.ToString(); } } private void button1_Click(object sender, EventArgs e) { //int v=Convert.ToInt32(textBox1.Text); int v1 = Convert.ToInt32(textBox7.Text); int v2 = Convert.ToInt32(textBox1.Text); if (v1 < 100) { label1.Visible = true; label1.Text = “ Failure Occured Unable to Load”; } if (v2 < 100) { label2.Visible = true; label2.Text = “Failure Occured Unable to Load”; } } private void pictureBox1_Click(object sender, EventArgs e) Romanian Statistical Review nr. 1 / 2017 15 { if (textBox7.Text == “” && textBox1.Text == “”) { MessageBox.Show(“Start any of Your Server And Proceed”); } else { Form2 f2 = new Form2(); f2.Show(); } } private void button4_Click_1(object sender, EventArgs e) { Form4 f4 = new Form4(); f4.Show(); this.Hide(); } private void label9_Click(object sender, EventArgs e) { } } } CONCLUSION The proposed algorithm uses two detection operators to automatically determine when to change the low level heuristic algorithm and a perturbation operator to fi ne tune the solutions obtained by each low-level algorithm to further improve the scheduling results in terms of make span. As the simulation results show, the proposed algorithm can not only provide better results than the traditional rule-based scheduling algorithms, it also outperforms the other heuristic scheduling algorithms, in solving the workfl ow scheduling and Hadoop map-task scheduling problems on cloud computing environments. With the incorporation of genetic programming into hyper-heuristic research, a new level of approaches are found that we have termed ‘heuristics to generate heuristics’. These approaches provide richer heuristic search spaces, and thus the freedom to create new methodologies for solving the underlying combinatorial problems. In addition, the simulation results show further that the proposed algorithm converges faster than the other heuristic algorithms evaluated in this Romanian Statistical Review nr. 1 / 201716 study for most of the datasets. In brief, the basic idea of the proposed hyper heuristic algorithm is to leverage the strengths of all the low level algorithms while not increasing the computation time, by running one and only one low-level algorithm at each iteration. This is fundamentally different from the so-called hybrid heuristic algorithm, which runs more than one low level algorithm for each iteration; thus requiring a much longer computation time. REFERENCES 1. P. Chr´etienne, E. G. Coffman, J. K. Lenstra, and Z. Liu, Eds., 1995, Scheduling Theory and Its Applications. John Wiely & Sons Ltd. 2. A. Allahverdi, C. T. Ng, T. C. E. Cheng, and M. Y. Kovalyov, 2008, “A survey of scheduling problems with setup times or costs,” European Journal of Operational Research, vol. 187, no. 3, pp. 985–1032. 3. M. Armbrust, A. Fox, R. Griffi th, A. D. Joseph, R. Katz, A. Konwinski, G. Lee, D. Patterson, A. Rabkin, I. Stoica, and M. Zaharia, “A view of cloud computing,” Communications of the ACM, vol. 53, no. 4, pp. 50–58, 2010. 4. Reddy, C. S., et al., 2016, “Obtaining Description for Simple Images using Surface Realization Techniques and Natural Language Processing” Indian Journal of Science and Technology 9.22. 5. I. Foster, Y. Zhao, I. Raicu, and S. Lu, 2008 , “Cloud computing and grid computing 360-degree compared” in Proceedings of the Grid Computing Environments Workshop, pp. 1–10. 6. Mamatha, E., C. S. Reddy and Ramakrishna Prasad,, 2012, “Mathematical Modeling of Markovian Queuing Network with Repairs, Breakdown and fi xed Buffer” i-Manager’s Journal on Software Engineering 6.3 (2012): 21. 7. Tsai, Chun-Wei, et al., 2014, “A hyper-heuristic scheduling algorithm for cloud” IEEE Transactions on Cloud Computing 2.2 (2014): 236-250. 8. X. Lei, X. Liao, T. Huang, H. Li, and C. Hu, 2013, “Outsourcing large matrix inversion computation to a public cloud” IEEE Transactions on Cloud Computing, vol. 1, no. 1, pp. 78–87 9. Mamatha, E., C. S. Reddy, and K. R. Prasad, 2016, “Antialiased Digital Pixel Plotting for Raster Scan Lines Using Area Evaluation”, Emerging Research in Computing, Information, Communication and Applications. Springer Singapore, 461-468 10. J. H. Holland, 1975, Adaptation in Natural and Artifi cial Systems: An Introductory Analysis with Applications to Biology, Control, and Artifi cial Intelligence. University of Michigan Press. 11. Mamatha et al., An Effi cient Line Clipping Algorithm in 2D Space, in press, International Arab Journal of Information Technology, 12. C. W. Tsai and J. Rodrigues, “Metaheuristic scheduling for cloud: A survey”, 2014, IEEE Systems Journal, vol. 8, no. 1, pp. 279–297. 13. Y. T. J. Leung, 2004, Handbook of Scheduling: Algorithms, Models and Performance Analysis. Chapman & Hall/CRC. 14. K. M. Elsayed and A. K. Khattab, 2006, “Channel-aware earliest deadline due fair scheduling for wireless multimedia networks” Wireless Personal Communications, vol. 38, no. 2, pp. 233–252. 15. Mamatha, et al., 2008, “Performance evaluation of homogeneous parallel processor system of Markov modeled queue” i-Manager’s Journal on Software Engineering 3.2: 58. 16. Z. Wu, X. Liu, Z. Ni, D. Yuan and Y. Yang, 2011, “A market-oriented hierarchical scheduling strategy in cloud workfl ow systems” The Journal of Supercomputing, pp. 1–38. 17. M. R. Garey, D. S. Johnson and R. Sethi, 1976, “The complexity of fl owshop and jobshop scheduling” Mathematics of Operations Research, vol. 1, no. 2, pp. 117–129. Romanian Statistical Review nr. 1 / 2017 17 18. J. Bła˙zewicz, K. H. Ecker, E. Pesch, G. Schmidt and J. We¸glarz, 2001, Scheduling Computer and Manufacturing Processes. Springer-Verlag New York, Inc. 19. D. Shi and T. Chen, 2013, “Optimal periodic scheduling of sensor networks: A branch and bound approach” Systems & Control Letters, vol. 62, no. 9, pp. 732–738. 20. M. Dorigo and L. M. Gambardella, 1997, “Ant colony system: A cooperative learning approach to the traveling salesman problem” IEEE Transactions on Evolutionary Computation, vol. 1, no. 1, pp. 53–66. 21. Apache Hadoop. [Online]. Available: http://hadoop.apache.org 22. Y. M. Huang and Z. H. Nie, “Cloud computing with Linux and Apache Hadoop” [Online]. Available: http://www.ibm.com/developerworks/aix/library/au-cloud apache. 23. Mamatha, E., C. S. Reddy and S. Krishna Anand, 2016, “Focal point computation and homogeneous geometrical transformation for linear curves” Perspectives in Science 24. M. Rahman, X. Li, and H. Palit, 2011, “Hybrid heuristic for scheduling data analytics workfl ow applications in hybrid cloud environment,” in Proceedings of the IEEE International Symposium on Parallel and Distributed Processing Workshops, pp. 966–974. RESULTS AND DIAGRAMS Romanian Statistical Review nr. 1 / 201718 
work_3f4o5vkwirca5doewxivlid55u	----	Human action recognition using an ensemble of body-part detectors Bhaskar Chakraborty, Andrew D. Bagdanov, Jordi Gonzàlez and Xavier Roca {bhaskar, bagdanov, poal, xavir}@cvc.uab.es Universitat Autònoma de Barcelona, Computer Vision Center Campus UAB Edifici O, 08193 Bellaterra, Spain May 3, 2010 Abstract This paper describes an approach to human action recognition based on the probabilistic optimization model of body parts using Hidden Markov Model (HMM). Our proposed method is able to distinguish between similar actions by only consid- ering the body parts having major contribution to the actions, for example, legs for walking, jogging and running; hands for boxing, waving and clapping. We apply HMMs to model the stochastic movement of the body-parts for action recognition. The HMM construction requires an ensemble of body-part detectors, followed by grouping of part detections to perform human identification. Three example-based body part detectors are trained to detect three components of the human body: the head, the legs and the arms. These detectors cope with viewpoint changes and self-occlusions through the use of ten sub-classifiers that detect body parts under a specific range of viewpoints. Each sub-classifier is a Support Vector Machine (SVM) trained on features selected for the discriminative power for each particular part/viewpoint combination. Grouping of these detections is then performed using a simple geometric constraint model which yields a viewpoint invariant human de- tector. We test our approach on the most commonly used action dataset, the KTH 1 dataset, and obtain promising result, which proves that with a simple and compact representation we can achieve robust recognition of human actions, compared to complex representation. 1 Introduction Visual recognition and understanding of human actions are among the most important research areas in Computer Vision (Moeslund, et al. 2006, Shipley & Zacks 2008, Wang, et al. 2003, Turaga, et al. 2008). Good solutions to these problems would yield huge potential for many applications such as the search and structuring of large video archives, video surveillance, human-computer interaction, gesture recognition and video editing. Human detection and action recognition are extremely challenging due to the non-rigid nature of humans in video sequences caused by changes in pose, changing illumination conditions and erratic non-linear motion. When viewed as a stochastic estimation problem, the critical issue in action recog- nition becomes the definition and computation of the likelihood, Pr(a|H,I), of action a given the human H in the image sequence I, since the human H is the main agent performing the action a. The difficulty in working with this likelihood is directly related to the complexity of the joint distribution Pr(H,I) over all possible human figures H in the image sequence I. Holistic approaches which attempt to model the entire human figure, generally, must resort to very sophisticated and complex models of this joint distribution, resulting in very demanding model estimation and optimization problems. Basic human actions can be represented using the local motion of individual body parts. Actions like walking, jogging, running, boxing and waving are systematic com- binations of the motion of different human body components. From this perspective, it can also be observed that not all body parts contribute equally to all action classes. For example, actions like walking, running and jogging are characterized mostly by the movement of the legs. Boxing, waving and hand clapping, on the other hand, mostly depend on the arms. Our approach is based on these observations and we define the action likelihood Pr(a|H,I) instead as Pr(a|B,I), where B is an ensemble of body parts 2 T ra in in g T ra in in g Labeled Body-parts Labeled Body-parts Legs Arms Head HOG Feature Extraction and Selection HOG Feature Extraction and Selection Head Arms Legs View-invariant Body-part Detectors View-invariant Body-part Detectors Head SVMs Arm SVMs Leg SVMs Gaussian Mixture Model (GMM) Gaussian Mixture Model (GMM) Arm Cluster Leg Cluster Walking HMM Running HMM Clapping HMM Action HMMs Action HMMs (a) Q u e ry Q u e ry Query Video Body-part Detector HOG Feature Extraction Feature Quantization Using GMM Classification by Action HMMs Label the Query Video (b) Figure 1: The proposed framework for action recognition based on probabilistic optimiza- tion model of body parts using Hidden Markov Models (a) Construction of body-part detectors and Action HMMs (b) Action recognition. which the human H is composed of. Moreover, the likelihood is further simplified by conditioning actions only on those body parts which contribute most to a particular ac- tion. Features from body parts B are used to model the action likelihood Pr(a|B,I), and optimizing this likelihood over all known actions yields a maximum likelihood estimate of the action in the image sequence. The ensemble of body part detectors is an important component of our method and we use SVM-based body part detectors over a range of viewpoints to build viewpoint- invariant body part detectors. We model human actions as a repeating chain of body- 3 part poses (Chakraborty, et al. 2008). Similar poses are grouped together by applying a Gaussian Mixture Model (GMM) on the features of the body parts in order to identify key-poses. These key-poses then serve as a vocabulary of hidden action-states for Hidden Markov Models (HMMs) that model the temporal-stochastic evolution of each action class. Figure 1 shows the overview of our proposed method. In the next section we review relevant work in the literature. Section 3 describes our probabilistic model for action recognition. In section 4 we describe the viewpoint- invariant part detectors used to define the part ensembles in an image sequence and the HMM-based approach to model the likelihood. Section 4.2 describes the HMM learning process. Experimental results are reported in section 5 and section 6 concludes the paper with some discussions and indications of future research directions. 2 Related work Several methods for learning and recognizing human actions directly from image mea- surements have been proposed in the literature (Black & Jepson 1996, Davis & Bobick 1997, Zelnik-Manor & Irani 2001, Chomat & Crowley 1999). These global feature based action recognition techniques mainly employ flow, interest points and silhouette features. These methods may work well under fixed, well-defined conditions, but are not generally applicable, especially for varying viewpoints. For example, accurately estimating the po- sitions of joints in different view-points is an extremely difficult problem and is computa- tionally expensive. Optical flow images are the least affected by appearance but are often too noisy due to inaccuracies in flow calculation. A common issue with interest point detectors is that the detected points are sometimes too few to sufficiently characterize human actions, and hence reduce recognition performance. This issue has been avoided in (Niebles, et al. 2008) by employing a separable linear filter (Dollár, et al. 2005), rather than space-time interest point detectors, to obtain motion features using a quadrature pair of 1D temporal Gabor filters. View-invariance in action recognition is addressed in (Yilmaz & Shah 2005, Rao, et al. 2002, Parameswaran & Chellappa 2006). In our 4 approach we overcome these problems by identifying human body parts under different view-points and explicitly modelling their influences on action interpretation. For action modelling and classification, many researchers use HMMs due to the sequential-stochastic nature of human actions. The HMMs and AdaBoost are used to recognize 3D human action, considering the joint position and pose angles (Fengjun & Ramkant 2006). In (Ahmad & Lee 2006) action recognition is performed using silhouette features and the Cartesian component of optical flow velocity. HMMs are then used to classify actions. Also (Mendoza & de la Blanca 2007), detect human actions using HMMs based on the contour histogram of full bodies. Again, all these methods use features from the whole body to model the HMMs. Human silhouette changes intensely under different view-points and the extracted features from the whole body do not represent the action for all views. In our approach we reduce the complexity of HMM learning by identifying the contributing body parts for an action. In this way action modelling becomes more explicit and can be generalized in different view points. Although conceptually appealing and promising, the merits of part-based models have not yet been widely recognized in action recognition. One goal of this work is to explore this area. Recent works on part-based event detection use hidden conditional random fields (Wang & Mori 2008), flow based shape feature from body parts (Ke, et al. 2005) and Histograms of Oriented Rectangles (HORs) (Ikizler & Duygulu 2009). These part-based approaches to action recognition employ complex learning techniques and the general concept of view invariance is missing. Although very promising, these approaches have several problems. Learning the parameters of a hidden conditional random field is a very complex procedure and requires a lot of training examples. Methods like HOR need to extract contours and filled silhouettes and the success of the method depends strongly on the quality of those silhouettes. Our approach, in contrast, is built upon example- based body part detectors which are simple to design and yet robust in functionality. In our method we combine the advantages of full body and part-based action recog- nition approaches. We use simple SVM-based body part detectors inspired by (Mohan, et al. 2001), but instead of using Haar-like features we use HOGs (Dalal & Triggs 2005). 5 The disadvantage of Haar-like features is that they are affected by human appearance, while HOG features are designed to extract shape information from the human body contour. Furthermore, we apply a feature selection technique to reduce the dimension- ality of the resulting SVMs. Those selected features from the body parts are then used to learn an HMM for each action class. We use Gaussian Mixture Models (GMM) to identify the group of similar body-parts, that have major contribution in an action, to obtain the action-states or key-poses. View-invariance is addressed by designing multiple sub-classifiers for each body-part, responsible for detecting the body-limbs in different view-points. 3 A probabilistic model for action recognition The task of human action recognition can be formulated as an optimization problem. Let A = {a1,a2, . . . ,an} denote a set of possible actions, where each ai denotes a specific action such as walking, running or hand clapping. We write the likelihood of a specific action ai in a given image sequence I with human H as: Pr(ai|H,I) for ai ∈ A. Given an instance of I with detected human H, a maximum likelihood estimation of the action being performed is: a∗ = argmax ai∈A Pr(ai|H,I). (1) Rather than holistically modelling the entire human H, we consider it to be an ensemble of detectable body parts: B = {B1,B2, . . . ,Bm}, where each Bi represents one of m-body parts such as the head, legs, arms or torso. The likelihood can now be expressed as: Pr(ai|H,I) = Pr(ai|{B1,B2, ...,Bm},I) (2) 6 Our model strengthen the fact that not all actions depend equally on all body parts. Actions like walking, jogging and running, for example, depend primarily on the legs. Actions like boxing, waving and clapping, on the other hand, depend primarily on the arms. To simplify the likelihood in (Equation 2), we define a dependence function over the set of all subsets of body parts: d(ai) : A −→P(B), where P(B) is the power set of B, or {c|c ⊆ B}. The set d(ai) determines the subset of body parts on which action ai most strongly depends. The likelihood is now approximated as: Pr(ai|H,I) ≈ Pr(ai|d(ai),I), (3) and the maximum likelihood estimate of the action given an ensemble of body parts and an image sequence is: a∗ = argmax ai∈A Pr(ai|d(ai),I). (4) The approximate likelihood in Equation 3 makes explicit the dependence of an action class on a subset of body parts. This approximation assumes that the action ai is inde- pendent of the body parts excluded from d(ai) and thus the likelihood can be computed by simply excluding the irrelevant parts rather than optimizing (or integrating) over them. In the following sections we describe how we model the approximate likelihood in Equation 3 using viewpoint invariant body-part detectors and HMMs over detected features of these body parts, and then from these arrive at the estimate of Equation 4 for action classification. 4 Action modelling by HMM Human action is a stochastic movement of an ensemble body parts. Moreover, for many actions it is possible to find the body parts that have a major contribution on it. We choose Hidden Markov Models for modelling Pr(ai|d(ai),I) [Equation 3] where d(ai) is 7 either Blegs or Barms. That is, we use d(ai) to indicate whether action ai depends mostly on the arms or on the legs. An HMM is a collection of finite states connected by transitions. Each state is char- acterized by two sets of probabilities: a transition probability, and either a discrete out- put probability distribution or a continuous output probability density function. These functions define the conditional probability of emitting each output symbol from a finite alphabet, conditioned on an unknown state. More formally, it is defined by: (1) A set of states S, with an initial state SI and a final state SF ; (2) The transition probability matrix, T = {tij}, where tij is the probability of taking the transition from state i to state j and (3) The output probability matrix R. For a discrete HMM, R = {rj(Ok)}, where Ok represents a discrete observation symbol. The initial state distribution is π = {πi}, and the complete parameter set of the HMM can be expressed as: λ = (T,R,π). (5) Here, for each action ai one discrete HMM is constructed using features from the con- tributing body parts: Blegs or Barms. We obtain the set of hidden states, S, using Gaussian Mixture Model (GMM) on the features from the detected body-parts. The transition probability matrix, T, is learnt and action classification is done after com- puting the output probability matrix, R, accordingly. Following sections describe the body-part detection and HMM learning. 4.1 Body-part detection Here, we use three body part detectors for the head, leg and arms respectively. Body part detection is based on sliding-window technique. Specific sized rectangular bounding boxes are used for each body part. These bounding boxes are slided over the image-frame, taking the section of the image-frame as an input to the head, leg and arm detectors. These inputs are then independently classified as either a respective body part or a non- body part. In order to prevent possible false positives, those detected components are combined into a proper geometrical configuration into another specific sized bounding 8 -1 0 1 -2 20 1 0 1 1 2 -1 0 00 -1 2 1 Feature Selection π/2 - π/2 Selected Blocks of HOG Gx Gy Training Image Gradient Image Dividing the Gradient Image into (8X8) Blocks Computing 6-bin HOG features Figure 2: Feature extraction and selection method from a training image. The training image is divided into several blocks and then HOG features are extracted. Finally standard deviation based feature selection method is applied to obtain feature vector for SVM. box as a full human. All these bounding boxes are obtained from the training samples. Furthermore, the image itself is processed at several scales. This allows the system to be scale invariant. 4.1.1 Features extraction and selection In our approach (Algorithm 1), labeled training images for each body part detector are divided into (8 × 8) blocks after applying Sobel mask on them. HOG features are extracted from those blocks and a 6 bin histogram is computed over the angle range (-π 2 , +π 2 ). So, this gives one 6-dim feature vector for each of those (8 × 8) pixel blocks. Next, we select the best feature vector group among all of them. This feature selection method is based on the minimization of the standard deviation (σ) of different dimensions of those feature vectors. 9 Algorithm 1 Feature extraction for body component SVM Require: Training images of body component. Ensure: Features for SVM. 1: for every training image do 2: Apply Sobel operator 3: Divide Sobel image into (8 × 8) blocks 4: for each blocks do 5: Compute 6 bin histogram of HOG features within the range (-π 2 , +π 2 ) 6: Keep it in feature array. 7: end for 8: Apply feature selection algorithm over the feature array. 9: end for 10: Selected features are used to learn SVM. Let there be N number of training image samples of size W × H, divided into n number of (8 × 8) blocks. For each of these (8 × 8) blocks we have 6-bin HOG features as, G = 〈g1,g2, . . . ,g6〉. We define the standard deviation, σij, of ith block and jth bin of HOG feature as: σij = ( 1 N ) N∑ t=1 (g(t)ij −µij) 2 (6) where i = 1, 2,. . . , n; j = 1, 2,. . . , 6; t = 1, 2, . . . , N; g(t)ij is defined as jth bin HOG feature of ith block of tth training image and µij is defined as the mean of jth gradient of ith block over all the training images. The values of σij, in Equation 6, are sorted and those 6dim feature vector packets are taken for which the σij is smaller than a predefined threshold. In our case we ran the experiment several times to obtain this threshold empirically. These selected features are used to train SVMs for the body part detectors. Figure 2 shows the feature extraction technique. 10 4.1.2 Geometric constraints on body-part detection The obtained outcome of those component detectors are combined based on the geo- metric constraint of full human body (Algorithm 2) to avoid the false positives obtained from the part detectors. We define the detected body-part component bounding boxes as, Head Bounding Box (RH), Leg Bounding Box (RL) and Arm Bounding Box (RA) obtained from body-part detectors and the full human bounding box (RF ). Algorithm 2 Geometric constraints on body-part detection Require: Detected body-part bounding boxes: RH, RL and RA; Width and height of full human bounding box: WRF and HRF . Ensure: Full human bounding box: RF , if geometric constraints are satisfied; NULL, otherwise. 1: (XCH ,YCH ) ← CENTROID(RH). 2: (XCL,YCL ) ← CENTROID(RL). 3: if (XCH − WRH 2 ) < XCL < (XCH + WRH 2 ) then 4: if YCL < HRF 2 then 5: Obtain RF using {(XCH − WRF 2 ), (YCH + HRH 2 ), WF , HF}. 6: (XCF ,YCF ) ← CENTROID(RF ). 7: (XCA,YCA ) ← CENTROID(RA). 8: if (XCF − WRF 2 ) < XCA < (XCF + WRF 2 ) then 9: if YCH > YCA > YCF then 10: Return(RF ). 11: end if 12: end if 13: end if 14: else 15: Return(NULL). 16: end if Removal of false positives are depicted in Figure 3. When the detectors are applied 11 (a) (b) (c) (d) Figure 3: Removal of false positives using geometric constraints of different component detectors. (a) original image (b) detection of head and legs with overlapping bounding boxes (c) after getting single detection window for head and leg including false positives (d) detection of head and leg after removal of false positives using geometric constraint. in the image there are usually many overlapping detected windows for a particular body component. We obtain a single bounding box from those overlapping detected windows in the following way. If two bounding boxes share more than 70% overlapping area they are merged to obtain one single box. In this way overlapping detection windows are converted into a single one. After that, the above geometric constraint is applied to remove possible false positives and we obtain an ensemble of body-part detectors resulting in a full human detection. 4.2 Construction of action HMMs For learning an HMM for a particular action, several sequences of frames are chosen and every such sequence is called action-cycle/cycle for that action. The frames that define one cycle depends on the training sequences and the action itself. For example, the number of frames in an action cycle varies when it is performed in a circular path. Let 12 Figure 4: Training dataset for each body-part detectors. It shows training samples for each sub-classifier. there be M frames inside one action cycle and in each of these frames the body parts are detected using component detectors. Let assume that in the kth frame we detect body part bounding boxes where there are n best 6dim feature vectors from the Section: 4.1.1 as {Gk1 , G k 2 , ..., G k n} where each Gki = 〈 gi1,g i 2, . . . ,g i 6 〉k . Then we compute mean over all these Gki ’s to get the features from the kth frame. So, we have 〈µ1,µ2, . . . ,µM〉 as a feature to construct the HMM where each of these µi’s can be expressed as: µi = ( 1 M ) M∑ i=1 〈 gi1,g i 2, . . . ,g i 6 〉 . (7) The significance of taking the mean is to get the general orientation of the body part which in turn signifies one pose or a series of similar poses in an action. We fit a Gaussian Mixture Model (GMM) using Expectation Maximization (EM) on these mean features to obtain key-pose alphabets for a particular action. These key-poses are the centre of each cluster obtained from the GMM. Once we obtain these key-pose alphabets, we assign every detected body pose in the frames of each of the action cycles using the nearest-neighbour approach to get a sequence of HMM states for them. These state sequences are used to learn the parameters of the HMM in Equation 5. In this way, we obtain the model to compute probability value in Equation 3. 13 Figure 5: Results of head, arm and leg detection as a validation process. Here detection of profile head and leg is shown. Arm detections are shown where both arms are visible and one is occluded. For an unknown action, the detected body poses in all frames are mapped into the state sequences in the similar way, but without dividing them into cycles, since in such cases the cycle information is not known. P(Ot|st = s), the probability of an observation sequence Ot given the state s at time t, is computed using both Forward and Baum- Welch (Ghahramani 2002, Rabiner 1989) algorithms for every action class HMM. The action class that gives the maximum of these probability values (Equation 4), is the class label of that unknown class. 5 Experiments We use three publicly available datasets for our experiments on body-part detection and action recognition. HumanEva dataset1: This dataset is introduced in (Sigal & Black 2006). We use this to create our human body-part component dataset. HumanEva dataset contains 7 calibrated video sequences (4 grayscale and 3 colour) that are synchronized with 3D body poses obtained from a motion capture system. The dataset contains 4 subjects performing 6 common actions (e.g. walking, jogging and gesturing etc.). KTH dataset2: This dataset is the most common dataset for the action recognition 1http://vision.cs.brown.edu/humaneva/ 2http://www.nada.kth.se/cvap/actions/ 14 introduced in (Schuldt, et al. 2004). It covers 25 subjects and four different recording conditions of the videos. There are six actions in this dataset: walking, running, jogging, boxing, clapping and waving. The resolution, (120×160), of this dataset is quite low. HERMES indoor dataset3: This dataset is described in (González, et al. 2009). In this HERMES sequence (2003 frames @ 15 fps, 1392 × 1040 pixels) there are three people in a room. They act in a discussion sequence sitting around a table, where bags are carried, left and picked from the floor, and bottles are carried, left and picked from the vending machine and from the table. In this discussion sequence several agents are involved in different simultaneous grouping, grouped and splitting events, while they are partially or completely occluded. 5.1 Training of the body part detectors Our system uses four head detectors, two leg detectors and four arm detectors. The four head detectors correspond to view angle ranges (π 4 , 3π 4 ), ( 3π 4 , 5π 4 ), ( 5π 4 , 7π 4 ) and ( 7π 4 , π 4 ). We chose this division so that the consecutive ranges have smooth transition of head poses. For arms, there are four classifiers corresponding to different arm positions, grouped in the same angular views of the head. Detecting arms is a difficult task since arms have more degrees of freedom compared to the other body parts. For each action we use four major arm poses and considering the pose symmetry the detection of other pose variation is achieved. To detect the legs, two sub-classifiers have been added: one representing front(rear) view and the other one for profile views of the legs. Figure 4 shows some training samples from our body-part dataset. In our method for head, leg and arms detection the bounding box sizes are fixed to (72×48), (184×108) and (124×64) pixels respectively. A (264×124) pixel bounding box is applied for full human. Since the test image is zoomed to various sizes, and in each zoomed image the components of those sizes are searched, the fixed component sizes do not affect scale invariance of human detection. To train each component detector, 10, 000 true positives and 20, 000 false positives selected from the HumanEva dataset are used. The 3http://iselab.cvc.uab.es/indoor-database 15 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 T u re P o si tiv e R a te False Positive Rate Head Detection Leg Detection Arm Detection 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 False Positive Rate T ru e P o si tiv e R a te Head Detection Leg Detection Arm Detection (a) HumanEva dataset (b) KTH dataset 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 False Positive Rate T ru e P o si tiv e R a te Head Detection Leg Detection Arm Detection (c) HERMES indoor sequence Figure 6: ROC curves for different body part detectors on various datasets. The false alarm rate is the number of false positive detections per window inspected. sizes of different body-part component bounding boxes are determined after learning the statistics of height and width of each component from all the sequences of HumanEva dataset. A tolerance of 10 pixels is also used along the height and width of the each bounding box. 5.2 Performance Evaluation of Body-part Detection We performed extensive experiments and quantitative evaluation of the proposed ap- proach of body part detection. We validate our part detector using HumanEva dataset. For a particular body-part detector we usually obtain different detection scores from the 16 Figure 7: Performance of body part detectors on the KTH dataset. Detection of head, arms and leg shown for profile poses. sub-classifiers that the detector is composed of. We chose the detection result based on the best detection score among themselves. Figure 5 shows the results of the compo- nent detectors in HumanEva dataset. These detections are the strong examples of the view invariant detection since in the HumanEva dataset the agents perform actions in a circular path and our component detector show robust performance in detecting them. To test the performance of body part detectors, KTH Dataset (Schuldt et al. 2004) and HERMES indoor sequence dataset (González et al. 2009) are used. The Receiver Operating Characteristic (ROC) curves are shown in the Figure 6 for three component detectors: head, legs and arms of different datasets and also for the detection of full human. These ROC curves are generated considering the overall performance of all the sub-classifiers of each group of the three classifiers. ROC analysis reveals that head detection and leg detection are quite accurate. Although there are false positives, but the geometric constraint eliminates them. For the arms, however, the detection rate is not very high since arm pose varies dramatically while performing actions. In the HumanEva and Hermes indoor datasets we obtain high detection rates for full human compared to the KTH dataset. In the KTH dataset, there are several image sequences where it is impossible to detect arms due to different clothes than the other two datasets. In such cases we are able to detect head and legs, and when they are in proper geometrical arrangement, the full human bounding box is constructed. There are some sequences where only legs are 17 detected due to low resolution. Figure 7 shows some examples of human detection on the KTH dataset. In those images the detected bounding boxes are found at different scales and they are drawn after rescaling them to 1:1 scale. For low resolution image sequences much information has been lost due to scaling and the Sobel mask hardly found important edges from the particular body parts. Our system works well in high resolution datasets like, the HumanEva (resolution 644×484) and the HERMES indoor sequences (resolution 1392×1040). It gives an average 97% recognition rate for walking, jogging and boxing actions. However, on low resolution datasets like the KTH (resolution 160 × 120) we achieve lower performance on the actions like jogging and hand clapping. In the KTH dataset there are many cases where the human is visible at the original resolution but when it is zoomed to 640 × 480 to detect the body-parts, the objects are blurred and detection suffers. Table 1: Comparison of mean classification accuracy on the KTH dataset. (Ke et al. 2005) 62.9% (Wong & Cipolla 2007) 71.16% (Schuldt et al. 2004) 71.72% Our Approach 79.2% (Dollár et al. 2005) 81.17% (Niebles et al. 2008) 81.5% 5.3 Action Recognition We train the walking, jogging and boxing HMMs on the HumanEva dataset and tested on KTH dataset. But for the actions running, boxing, hand waving and hand clapping we use KTH dataset for both training and testing. In these cases we take random sample of 50% of the video sequences as training and the rest as test. Table 1 shows a comparison of our recognition rate with other methods. The average action recognition rate obtained from our method is quite promising but fails to surpass (Dollár et al. 2005) and (Niebles et al. 2008). But, in both of these cases the training and test are done on 18 the KTH dataset. We, instead, use HumanEva as a training dataset for the actions, walking, boxing and jogging and test on the KTH dataset. Table 2 shows the confusion matrix of our approach on the KTH dataset. In the confusion matrix we can see that our approach, clearly, can able to distinguish the leg and arm related actions. We obtain 100% recognition rate for walking and boxing actions which is quite impressive. But the jogging and hand clapping have the poor recognition rate. For jogging, the major confusion occurs with walking (40%) and 23.1% running actions get confused as jogging. In the case of hand clapping we get the confusions with other two actions, the boxing and hand waving. These confusions occur due to the low resolution of the KTH dataset which affects the performance of the body-part detectors. Table 2: Confusion matrix of action recognition using our component-wise HMM ap- proach. KTH dataset are used. Walk Jog Run Box Wave Clap Walk 100.0 0.0 0.0 0.0 0.0 0.0 Jog 40.0 60.0 0.0 0.0 0.0 0.0 Run 0.0 23.1 76.9 0.0 0.0 0.0 Box 0.0 0.0 0.0 100.0 0.0 0.0 Wave 0.0 0.0 0.0 20.0 73.4 6.6 Clap 0.0 0.0 0.0 13.5 19.8 66.7 6 Discussion and Conclusion This work presents a novel approach for recognizing actions based on a probabilistic op- timization model of body limbs. It also includes a view-point invariant human detection is achieved using example-based body-part detectors. Stochastic changes of body components are modelled using HMM. This method is also able to distinguish very similar actions like walking, jogging and running (considering features from legs); boxing, hand waving and hand clapping (considering features from 19 hands). The leg movements, sometimes, are very similar in the actions like jogging and running. Also, in some cases there are problems of resolution and contrast which makes it is difficult to distinguish those actions. Actions, involving hand motion also suffer from similar problems, some part of hand waving action look similar to hand clapping, which in tern causes ambiguity. We observe that there are major confusions occurred in actions jogging and clapping. In the KTH dataset some sequences of jogging action resemblance with walking and running actions. Moreover, other state-of-the the art methods also suffer in the same way. The confusion between clapping and waving could be because of the arm detectors are difficult to design for every possible degree of freedom specially in those two actions there are a variety of arm pose changes. On the hand, arm detectors perform well for boxing since in this case the pose changes do not vary a lot. The advantage of our approach is two fold; first, the body part detection is robust except for resolution limited images. Second, the HMM-based action model is capable of recognizing actions even when the body part detectors fail on some frames of an action sequence. The performance can be improved in human detection by adding more training sam- ples and introducing more angular views. There is no good dataset for different body- part components, so building a component dataset is an important task. For action recognition, higher order HMM can be applied. Information of other body parts which have minor contribution in the action like, arms for the walking can be included in order to minimize the misclassification rate. References M. Ahmad & S.-W. Lee (2006). ‘HMM-based Human Action Recognition Using Multi- view Image Sequences’. In ICPR ’06: Proceedings of the 18th International Conference on Pattern Recognition, vol. 1, pp. 263–266, Washington, DC, USA. IEEE Computer Society. 20 M. Black & A. Jepson (1996). ‘Eigentracking: Robust matching and tracking of articu- lated objects using a view-based representation’. B. Chakraborty, et al. (2008). ‘View-Invariant Human Action Detection Using Component-Wise HMM of Body Parts’. In V Conference on Articulated Motion and Deformable Objects, pp. 208–217, Andratx, Mallorca, Spain. O. Chomat & J. L. Crowley (1999). ‘Probabilistic recognition of activity using local appearance’. In In Conference on Computer Vision and Pattern Recognition (CVPR ’99), Fort Collins, pp. 104–109. N. Dalal & B. Triggs (2005). ‘Histograms of Oriented Gradients for Human Detection’. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 01, pp. 886–893, Washington, DC, USA. IEEE Computer Society. J. W. Davis & A. F. Bobick (1997). ‘The Representation and Recognition of Action Using Temporal Templates’. In CVPR ’97: Proceedings of the IEEE Computer Vision and Pattern Recognition, pp. 928–934. P. Dollár, et al. (2005). ‘Behavior recognition via sparse spatio-temporal features’. pp. 65–72. L. Fengjun & N. Ramkant (2006). ‘Recognition and Segmentation of 3-D Human Action Using HMM and Multi-class AdaBoost’. In 9th European Conference on Computer Vision, pp. 359–372, Graz, Austria. Z. Ghahramani (2002). ‘An introduction to hidden Markov models and Bayesian net- works’ pp. 9–42. J. González, et al. (2009). ‘Understanding dynamic scenes based on human sequence evaluation’. Image and Vision Computing 27(10):1433–1444. N. Ikizler & P. Duygulu (2009). ‘Histogram of oriented rectangles: A new pose descriptor for human action recognition’. Image and Vision Computing 27(10):1515–1526. 21 Y. Ke, et al. (2005). ‘Efficient Visual Event Detection using Volumetric Features’. In IEEE International Conference on Computer Vision, vol. 1, pp. 166–173. M. Mendoza & N. P. de la Blanca (2007). ‘HMM-Based Action Recognition Using Con- tour Histograms’. In Iberian Conference on Pattern Recognition and Image Analysis, Part I, pp. 394–401, Girona, Spain. T. Moeslund, et al. (2006). ‘A Survey of Advances in Vision-Based Human Motion Capture and Analysis’. In Computer Vision and Image Understanding, vol. 8, pp. 231–268. A. Mohan, et al. (2001). ‘Example-based Object Detection in Images by Components’. IEEE Transaction on Pattern Analysis and Machine Intelligence 23(4):349–361. J. C. Niebles, et al. (2008). ‘Unsupervised Learning of Human Action Categories Using Spatial-Temporal Words’. International Journal of Computer Vision 79(3):299–318. V. Parameswaran & R. Chellappa (2006). ‘View Invariance for Human Action Recogni- tion’. International Journal of Computer Vision 66(1):83–101. L. R. Rabiner (1989). ‘A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition.’. Institute of Electrical and Electronics Engineers 2:257–286. C. Rao, et al. (2002). ‘View-Invariant Representation and Recognition of Actions’. International Journal of Computer Vision 50(2):203–226. C. Schuldt, et al. (2004). ‘Recognizing Human Actions: a Local SVM Approach.’. In International Conference on Pattern Recognition, pp. 32–36, Cambridge, UK. T. F. Shipley & J. M. Zacks (2008). In ”Understanding Events From Perception to Action”. Oxford University Press. L. Sigal & M. J. Black (2006). ‘HumanEva: Synchronized Video and Motion Capture Dataset for Evaluation of Articulated Human Motion’. In Technical Report CS-06-08, Brown University. 22 P. Turaga, et al. (2008). ‘Machine Recognition of Human Activities: A Survey’. Circuits and Systems for Video Technology, IEEE Transactions on 18(11):1473–1488. L. Wang, et al. (2003). ‘Recent developments in human motion analysis’. Pattern Recognition 36(3):585–601. Y. Wang & G. Mori (2008). ‘Learning a discriminative hidden part model for human action recognition’. In Advances in Neural Information Processing Systems, vol. 21, pp. 1721–1728. MIT Press. S. Wong & R. Cipolla (2007). ‘Extracting Spatiotemporal Interest Points using Global Information’. In 11th IEEE International Conference on Computer Vision, pp. 1–8. A. Yilmaz & M. Shah (2005). ‘Recognizing Human Actions in Videos Acquired by Uncalibrated Moving Cameras’. IEEE International Conference on Computer Vision 1:150–157. L. Zelnik-Manor & M. Irani (2001). ‘Event-based analysis in video’. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp. 123–130. 23 
work_3g2qgpmfbnbqxfqrqvxk4quel4	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy Effective synchronizing algorithms R. Kudłacik a, A. Roman b,⇑, H. Wagner b a IBM Poland, SWG Krakow Laboratory, Armii Krajowej 18, 30-150 Krakow, Poland b Institute of Computer Science, Jagiellonian University, Łojasiewicza 6, 30-348 Krakow, Poland a r t i c l e i n f o Keywords: Circuit testing Conformance testing Synchronizing sequences Synchronizing automata Reset word Synchronizing algorithm a b s t r a c t The notion of a synchronizing sequence plays an important role in the model-based testing of reactive systems, such as sequential circuits or communication protocols. The main problem in this approach is to find the shortest possible sequence which synchronizes the automaton being a model of the system under test. This can be done with a synchronizing algorithm. In this paper we analyze the synchronizing algorithms described in the literature, both exact (with exponential runtime) and greedy (polynomial). We investigate the implementation of the exact algorithm and show how this implementation can be optimized by use of some efficient data structures. We also propose a new greedy algorithm, which relies on some new heuristics. We compare our algorithms with the existing ones, with respect to both runtime and quality aspect. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction Synchronizing words (called also: synchronizing sequences, re- set sequences, reset words or recurrent words) play an important role in the model-based testing of reactive systems (Broy, Jonsson, Katoen, Leucker, & Pretschner, 2005). In presence, with advanced computer technology, systems are getting larger and more compli- cated, but also less reliable. Therefore, testing is indispensable part of system design and implementation. Finite automata are the most frequently used models that describe structure and behavior of the reactive systems, such as sequential circuits, certain types of programs, and, more recently, communication protocols (Fukada, Nakata, Kitamichi, Higashino, & Cavalli, 2001; Ponce, Csopaki, & Tarnay, 1994; Zhao, Liu, Guo, & Zhang, 2010). Because of its prac- tical importance and theoretical interest, the problem of testing fi- nite state machines has been studied in different areas and at various times. Originally, in 1950s and 1960s, the researchers work in this area was motivated by automata theory and sequential cir- cuit testing. The area seemed to have mostly died down, but in 1990s the problem was resurrected due to its applications to con- formance testing of communication protocols. The problem of conformance testing can be described as follows (Lee & Yannakakis, 1996). Let there be given a finite state machine MS which acts as the system specification and for which we know completely its internal structure. Let MI be another machine, which is the alleged implementation of the system and for which we can only observe its behavior. We want to test whether MI correctly implements or conforms to MS. Synchronizing words allow us to bring the machine into one state, no matter which state we currently are in. This helps much in designing effective test cases, e.g. for sequential circuits. In Pomeranz and Reddy (1998) authors show a class of faults for which a synchronizing word for the faulty circuit can be easily determined from the synchronizing word of the fault free circuit. They also consider circuits that have a reset mechanism, and show how reset can ensure that no single fault would cause the circuit to become unsynchronizable. In Hyunwoo, Somenzi, and Pixley (1993) a framework and algo- rithms for test generation based on the multiple observation time strategy are developed by taking advantage of synchronizing words. When a circuit is synchronizable, test generation can em- ploy the multiple observation time strategy and provide better fault coverage, while using the conventional tester operation mod- el. The authors investigate how a synchronizing word simplifies test generation. The central problem in the approach based on the synchroniz- ing words is to find the shortest one for a given automaton. As the problem is NP-hard (see Section 2), the polynomial algorithms cannot be optimal, that is they cannot find the shortest possible synchronizing words (unless P ¼NP, which is strongly believed to be false). In last years some efforts were made in the field of algorithmic approach for finding short synchronizing words (Deshmukh & Hawat, 1994). Pixley, Jeong, and Hachtel (1994) presented an efficient method based upon the universal alignment theorem 0957-4174/$ - see front matter � 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2012.04.079 ⇑ Corresponding author. E-mail addresses: rafal.kudlacik@gmail.com (R. Kudłacik), roman@ii.uj.edu.pl (A. Roman), hub.wag@gmail.com (H. Wagner). Expert Systems with Applications 39 (2012) 11746–11757 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy and binary decision diagrams to compute a synchronizing word. There are also Natarajan (1986) and Eppstein (1990) algorithms. The problem of synchronizing finite state automata has a long history. While its statement is simple (find a word that sends all states to one state), there are still some important questions to be answered. One of the most intriguing issues is the famous Černý Conjecture (Černý, Pirická, & Rosenauerová, 1971), which states that for any n-state synchronizing automaton there exists a syn- chronizing word of length at most (n � 1)2. Should the conjecture be true, this would be a strict upper bound, as there exist automata with minimal synchronizing words of length exactly (n � 1)2. The Černý Conjecture has profound theoretical significance (remaining one of the last ‘basic’ unanswered questions in the field of auto- mata theory, especially after the Road Coloring Problem has been recently solved by Trahtman (2009)). On the other hand, there are several practical applications of finding short reset sequences: part orienters (Natarajan, 1986), finding one’s location on a map/ graph (Kari, 2002), resetting biocomputers (Ananichev & Volkov, 2003), networking (determining a leader in a network) (Kari, 2002) and testing electronic circuits, mentioned above. Clearly, finding short reset words is important both for theoretical and practical reasons. The paper is organized as follows. In Section 2 we give the basic definitions on automata and synchronizing words. In Sec- tion 3 we introduce two auxiliary constructions which are com- monly used in the synchronizing algorithms. In Section 4 we present the well-known synchronizing algorithms, both exact and greedy. In Sections 5 and 6 we present our two main re- sults: application of efficient data structures to the exact syn- chronizing algorithm and a new, efficient heuristic algorithm. Both sections end with the experimental results and efficiency comparison to other algorithms. 2. Synchronizing words An alphabet is a nonempty, finite set. A word over some alphabet A is a sequence of letters from A. The length of a word w is the number of its letters and is denoted by jwj. By e we denote the empty word of length 0. If A is an alphabet, by A⁄ we denote the set all words over A. For example, if A = {a, b}, then A⁄ = {e, a, b, aa, ab, ba, bb, aaa, . . .}. The catenation of words is denoted by a dot: if u, v 2 A⁄, then u�v = uv. A finite state automaton is a triple A¼ðQ; A; dÞ, where Q is a finite set of states, A is an alphabet and d is a transition function, d:Q � A ? Q. Note that initial and terminal states are not marked – we are not interested in languages accepted by automata, but rather in automaton action itself. In the following, for the sake of simplicity, we will use the word ’automaton’ instead of ‘a finite state automa- ton’. The transition function can be extended to PðQÞ� A�, that is, to the sets of states and words over A. The same symbol d will be used to refer to the extended function d : PðQÞ� A� !PðQÞ. This makes no confusion: "P # Q, a 2 A, w 2 A⁄ dðP; eÞ¼ P; dðP; awÞ¼ [ p2P fdðdðp; aÞ; wÞg: A word w is called a synchronizing word for A¼ðQ; A; dÞ iff jd(Q, w)j = 1. We say that such a word synchronizes A. We also say that A¼ðQ; A; dÞ is synchronizing if there exists w 2 A⁄ that syn- chronizes it. If, for a given A, there is no shorter synchronizing word than w, the word w is called the shortest synchronizing word (SSW) for A. There are two main algorithmic problems in the synchronization theory: in the first one, given a synchronizing automaton A¼ðQ; A; dÞ we ask about SSW for A. In the second one we ask to find any synchronizing word, not necessarily the shortest (but, of course, the shortest word is found, the better), in a reasonable time. These problems can be restated in the form of the following decision problems: Problem FIND-SSW. Input: a synchronizing automaton A and k 2 N. Output: YES iff the shortest word synchronizing A has length k. Problem FIND-SW-OF-LENGTH-K Input: a synchronizing automaton A and k 2 N. Output: YES iff there exists a synchronizing word of length k for A. The decision problem FIND-MSW has been recently shown to be DP-complete (Olschewski & Ummels, 2010). The decision problem FIND-SW-OF-LENGTH-K is NP-complete (Eppstein, 1990). It is well-known that the length of SSW for an n-state synchronizing automaton is at most n 3�n 6 (Klyachko, Rystsov, & Spivak, 1987; Pin, 1983). The Černý conjecture states that this length can be bounded by (n � 1)2. Černý showed (Černý, 1964) that for each n P 1 there exists an automaton with SSW of length (n � 1)2, so the conjectured bound is tight. These automata are called the Černý automata. An n-state Černy automaton will be denoted by Cn. Černy automaton is defined over a two-element alphabet A = {a, b} and its transition function is as follows: 8q 2f0; . . . ; n � 1g dðq; xÞ¼ ðq þ 1Þmodn if x ¼ a q if x ¼ b ^ q – n � 1 0 if x ¼ b ^ q ¼ n � 1 8>< >: ð1Þ Černý automata are very important, as automata with jSSWj = (n � 1)2 are very rare. Only eight such automata are known that are not isomorphic with the Černy automata (Trahtman, 2006). 3. Auxiliary constructions In this section we describe two auxiliary constructions used throughout the paper. Let A¼ðQ; A; dÞ be a synchronizing autom- aton. A pair automaton for A is the automaton A2 ¼ðQ 0; A; d0Þ, where: Q 0 ¼ [ p;q2Q^p–q ffp; qgg[f0g; d0 : Q 0 � A ! Q 0; d0ðfp; qg; lÞ¼ fdðp; lÞ; dðq; lÞg ifdðp; lÞ – dðq; lÞ; 0 otherwise; ( d0ð0; lÞ¼ 0 8l 2 A: Let A¼ðQ; A; dÞ be an automat on. A sequence (q1, q2), (q2, q3), . . . , (ql, ql+1), q1, . . . , ql+1 2 Q, is called a path in A, if for each i = 1, . . . , l there exists ai 2 A, such that d(qi, ai) = qi+1. We will identify such a path with a word a1a2 . . . al (notice that if there is more than one letter transforming some qi into qi+1, then the path including (qi, qi+1) can be identified with more than one word). Pair automaton shows how the pairs of states behave when words are applied to the original automaton. If p, q 2 S # Q and w is a path leading from {p, q} to 0, it means that jd(S, w)j < jSj, where p, q 2 S. In such a situ- ation we say that pair {p, q} of states was synchronized by w. Pair R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 11747 Author's personal copy automaton is utilized in all heuristic algorithms, see Sections 4.3, 4.4 and 4.5. The next proposition is a straightforward, but very important fact, utilized in all heuristic algorithms. Proposition 1. A word w 2 A⁄ synchronizes A2 iff w synchronizes A. Proposition 1 implies the following necessary and sufficient condition for A to be synchronized: Proposition 2. A is synchronizing iff each pair of its states is synchronizing. The problem of finding SSW can be restated as a problem of path-searching in a so-called power-set automaton (or power automaton for short) of A. A power-set automaton for A¼ðQ; A; dÞ is an automaton PðAÞ¼ ð2Q ; A; DÞ, where: 2Q ¼fP � Qgnf;g; D : 2Q � A ! 2Q ; Dðq; lÞ¼ [ s2q fdðs; lÞg 8q 2 2Q ; l 2 A: Like for d, we can extend D to 2Q � A⁄. Let PðAÞ¼ ð2Q ; A; DÞ be a power automaton of A¼ðQ; A; dÞ. State Q 2 2Q will be called a start state of PðAÞ. The size of the power automaton is exponential in the size of the original automaton. There are 2jQj � 1 states and jAj(2jQj� 1) edges. States of the power automaton represent subsets of states of the input automaton and are labeled by the corresponding subsets. We will only consider a subautomaton of the power automaton which is reachable from the start state. Specifically, when we say that the power automaton is small, we mean that the reachable subautomaton is small. Černy automata Cn are the interesting examples here, as all states of their power automata are reachable from the start state. We will sometimes refer to the ‘‘size of state s’’. This means that s is a state of the power automaton and it represents an jsj-element subset s of Q. Since edges in pair automaton represent transitions between subsets of states, the power automaton can be thought of as a way of expressing the global behavior of the input automaton when certain letter (or word) is applied. In contrast, pair-automa- ton describes local behavior only (it shows how the pairs of states are transformed). Proposition 3. The sets of synchronizing words in A and PðAÞ coincide. It means that A is synchronizing iff PðAÞ is synchronizing. It is clear that in a power automaton a path leading from Q 2 2Q to any state F 2 2Q, such that jFj = 1, represents a synchronizing word for A. Also, the shortest such path determines the shortest synchronizing word for A. So the entire problem can be rephrased as a basic graph problem. This is convenient as the single-source path-searching algorithms (exact or otherwise) have been extensively studied. Also, augmenting the generic path-searching methods with knowledge specific to this problem may give some interesting results. 4. Synchronizing algorithms In this section we describe 5 synchronizing algorithms, that is, the algorithms that find a synchronizing word for a given automa- ton. Two of them (EXACT and SEMIGROUP) are exponential ones that always find the shortest synchronizing words. Three others (NATARA- JAN, EPPSTEIN and SYNCHROP) are heuristic algorithms working in poly- nomial time, so they are faster, but they find not necessarily the shortest synchronizing words. In the following sections we assume jQj = n. 4.1. Exact exponential algorithm There are two well-known algorithms finding the shortest syn- chronizing words. Due to the fact that this problem is NP-hard, their runtime complexity is exponential in the size of the input automaton, which limits their use. The standard exact algorithm is a simple breadth-first-search in the power automaton. The runtime is X(2n) in the worst case. Standard implementation requires X(2nn) space. Due to these discouraging fact this algorithm is often disregarded in the literature. Algorithm 1 EXACT ALGORITHMðAÞ 1: Input: an automaton A¼ðQ; A; dÞ 2: Output: SSW of A (if exists) 3: queue Q empty 4: push Q into queue Q 5: mark Q as visited 6: while Q is not empty 7: S popðQÞ 8: foreach a 2 A 9: T d(S, a) 10: if jTj = 1 11: return reversed path from T to Q 12: if T is not visited 13: push T into Q 14: mark T as visited 15: return A is not synchronizing 4.2. Semigroup algorithm Another algorithm (which is typically more memory-efficient) was described in Trahtman (2006) and uses a notion of syntactic semigroup. Let A¼ðQ; A; dÞ be an automaton. Alphabet letters (and also words over A) represent functions Q ? Q, so if f is a func- tion from Q to Q and w 2 A⁄, by f�w we denote the composition of two functions: f and a function corresponding to w. Syntactic semi- group for A is constructed as follows: process all words over A⁄ in the lexicographic order. If a processed word defines a new function f: Q ? Q, add f to list L. The procedure is stopped in two cases: (1) when "f 2 L "a 2 A f�a 2 L, that is, when no new function can be de- fined; (2) when a constant function (mapping all elements into one element) is found. The word corresponding to the constant func- tion is SSW, as words were processed in the lexicographic order. The semigroup algorithm does not require a costly power automa- ton construction phase, but its standard implementation is terribly inefficient in the worst case. So it is slightly better than the power automaton algorithm, but the above fact limits its use only to small automata. The algorithm’s runtime complexity is O(jAjn � s2) with O(n � s) space required (Trahtman, 2006), where s is the size of the syn- tactic semigroup S. Syntactic semigroup size can be as big as nn, but since only a subset of S (containing only words no longer than SSW) is considered, the average runtime is usually much lower. The semigroup algorithm is used in a well-known synchroniza- tion package TESTAS. Its worst-time complexity can be drastically reduced and we show it in Section 5. 4.3. Natarajan algorithm One of the first heuristic algorithms for finding short synchro- nizing words was provided by Natarajan (1986). The algorithm is shown in Listing 2. 11748 R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 Author's personal copy Algorithm 2 NATARAJANðAÞ 1: Input: synchronizing automaton A¼ðQ; A; dÞ 2: Output: synchronizing word for A 3: Q {1, 2, . . . , n}; s e 4: while jQj > 1: 5: choose two states p, q 2 Q 6: w the shortest path from {p, q} to 0 7: Q d(Q, w) 8: s s.w 9: return s A loop in line 4. is performed O(n) times. The shortest path (line 6.) can be found in O(jAjn2). Transformation in line 7. is done in O(n3), because jQj = O(n) and jwj = O(n2). Hence, the total complex- ity is O(jAjn3 + n4). 4.4. Eppstein and cycle algorithms Eppstein proposed a modification of Natarajan’s algorithm. The modification is based on a preprocessing in which for each pair of states we compute the first letter of the shortest word synchroniz- ing these states. Eppstein has shown that this preprocessing allows us to reduce the complexity to O(n3 + jAjn2). CYCLE is a slight modification of EPPSTEIN. In CYCLE, when a pair of states in synchronized into some state q, it is required that in the next step q must be one of the elements in the chosen pair. CYCLE works optimally for Černý automata, that is, always returns SSW. 4.5. SYNCHROP algorithm SYNCHROP (and its modified version, SYNCHROPL) algorithm (Roman, 2009) is, in comparison to NATARAJAN, a ‘one-step-ahead’ procedure – we do not choose arbitrary pair of states as in line 5. of NATARAJAN. Let w(p, q) be the shortest word synchronizing {p, q}. For each {p, q} we check how the set of states in the pair automaton will be transformed if we apply w(p, q) to all states we currently are in. Each transformation is rated in terms of some heuristically de- fined cost function. We choose the pair with the lowest cost func- tion. The remaining part of the algorithm is exactly the same as in NATARAJAN. In its original version, SYNCHROP algorithm does not use the preprocessing introduced in EPPSTEIN. Therefore, its complexity is O(n5 + jAjn2). The detailed description and discussion on SYNCHROP properties and complexity is given in Section 6. 5. Optimizing exponential algorithms In this section we deal with the exact synchronizing algorithms. We show how the selection of the efficient data structures affects on the time complexity. Let us consider the basic version of the algorithm shown in Listing 1. While the algorithm looks very simple, its performance greatly depends on the data-structures used. The following aspects of the algorithm must be considered: 1. transition function computation, 2. state representation, 3. queue implementation, 4. visited states’ set implementation, 5. predecessor tree implementation (required, if the actual SSW must be returned rather than its length). Judging by the complexity (given in Trahtman (2006)) of the SEMIGROUP algorithm implementation in TESTAS, checking if t was previously visited (Listing 1, line 12.) is assumed to be performed in H(nm), where m is the number of elements visited. This step can easily be done in time nlog m using the standard tree-based dictionaries. So the worst case runtime complexity can easily be re- duced from H(jAjns2) to H(jAjnslog s). Another simple optimization is possible. Note that the original algorithm generates sequences of states of size n (and not sets). We can treat these elements as sets without losing any valuable information. This should also speed the process up: semigroup size s can reach the value of nn (and reaches 22n for Cn ), while there are at most 2n subsets of the considered set. Finally, for small n, a trick can be used: sets can be mapped to integers. This technique will be described in more details later. An ordinary array can be used to check if a set was previously added to the visited sets (actually, similar trick can be applied when sequences of size n are considered: radix n rather than 2 must be used). This allows us to skip the logarithmic part in H(jAjnslog s) yielding H(jAjns) (assuming that n is small). Of course, one could argue that there is no point in analyzing asymp- totic complexity with bounded values of input size. In such cases this notation should be understood as a way to express the order of complexity. This makes a really big difference. For example, TESTAS (which apparently uses this algorithm) cannot handle calculating SSW of Cn for n P 16. After these simple optimizations, all automata of size up to 26 (or more) should be handled easily (space complexity be- comes a bigger problem in case of slowly-synchronizing auto- mata). More detailed comparison will be shown later. Interestingly, after performing these optimizations, the algo- rithm is very similar to the power-automaton-based approach. In fact, we will try to merge the best features of these two approaches. 5.1. Power automaton traversal We would like to focus now on the algorithm that utilizes the concept of the power-automaton: a breadth-first-search is per- formed, beginning with the start state (set of all states of the input automaton). When a singleton state is found, the SSW has been found and the computation can be terminated. In the effect, typi- cally only a subset of states of the power automaton is utilized. This is an important fact that will enable us to examine larger automata. 5.1.1. On-line transition function computation The power automaton transition function can be calculated on- line (i.e. whenever it is required). In order to compute a single power automaton transition, one to n transition functions of the original automaton must be calculated. This is reasonable and is, in fact, a standard approach used. Typically the resulting runtime complexity is O(jAjn2nM(n)), where M(n) is the cost of performing one mapping of state into an associated value. For small n it can be assumed to be O(jAjn2n) (this is a slight abuse of the O notation, as this algorithm is limited to small n; in fact we are not interested in asymptotic behavior of this function). 5.1.2. Off-line transition function computation It is possible to generate all transitions in the power automaton in an amortized constant time. This approach leads us to H(jAj2n) complexity of calculating SSW. Currently, memory requirements make it usable only for n < 30, so we will only investigate this case. Power automaton’s states S1 . . . S2n�1 are generated in Gray’s code order, which can be done in amortized constant time. This way, consecutive states differ on exactly one position (i.e jSi � Si+1j = 1, where � denotes a symmetric-difference operation). Power automaton transition for certain l 2 A can be expressed as DðSi; lÞ¼ dðSi \ Si�1; lÞ[ DðSi � Si�1; lÞ; ð2Þ R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 11749 Author's personal copy where jSi \ Si�1j ¼ jSij� 1; jSi � Si�1j ¼ 1: If Si is the ith generated state, d(Si \ Si�1,l) can be determined in con- stant time. This can be done by storing the count of each element of the current transition’s function value. This information can be up- dated in constant time. Some additional pre-computation is necessary, so that logarithm of a number can be computed in constant time. This is required to map a power automaton’s state (representing the sym- metric difference of consecutive states) into input automaton’s state. Since jSi � Si�1j = 1, Si � Si�1 is encoded as an integer I being a power of two. So log(I) denotes the state’s number. Listing 3 shows how to calculate this value quickly. Finally, the entire algo- rithm shown in Listing 4 can be run in a constant time. Algorithm 3 FAST_LOG2(n) 1: Input: integer n = 2k 2: Output: log2n = k 3: r right (n)// right half, r < 216 4: if r > 0: 5: return log2[r]// precomputed value 6: l left (n) shr 16 7: return 16 + log2[l]// assuming 32 bit integers are used Algorithm 4 FAST POWER AUTOMATON TRANSITION (Si,Si+1, prev_trans, image_count, l) 1: Input: States Si, Si+1; transition function for Si; count of each element in the transition function’s image; letter l 2 A for which the transition function is calculated 2: Output: power automaton transition for a given state and letter 3: ex Si XOR Si+1// bitwise symmetric difference 4: in Si AND next// bitwise and 5: ret prev_trans 6: change fast_log2(ex)// changed state’s number 7: change_to delta (change, l) 8: if in == prev:// change was added 9: image_count[change]+ = 1 10: ret = ret OR change_to 11: else// change was removed 12: image_count[change_to]� = 1 13: if image_count[change_to] == 0 14: ret = ret XOR change_to 15: return ret A full graph representing the power automaton is constructed in memory. This approach is good for calculating SSW of slowly-synchroniz- ing automata. It takes exactly H(jAj2n) integer operations (that is, never takes less, unlike other implementations). It is not suitable for larger (n > 30) automata due to memory requirements. It is also not recommended for random and fast-synchronizing automata as they tend to have small power automata. This should be a good choice when the inexact algorithm (run prior to the exact one) shows that SSW may be long. 5.1.3. Mapping states to arbitrary information In the algorithm a set of visited states must be maintained. We will consider a more general solution: mapping states into arbi- trary values (when a Boolean value is used, this approach is equiv- alent to defining a subset by its characteristic function). This method will be useful for storing predecessor tree. 5.1.4. Dense mapping When we can afford keeping the entire mapping in memory, the situation is rather simple. To each set can be assigned an integer that will uniquely identify this set. There exists a convenient corre- spondence r: r : 2f1;...;ng ! N; r�1 : f0; . . . ; 2n � 1g! 2f1;...;ng; rðSÞ¼ X i2S 2i�1; r�1ðIÞ¼ fi 6 n : bit ði � 1Þ of I is set to 1g: In other words, subsets of a fixed set can be represented as its characteristic vector. This binary vector can be encoded as an inte- ger. It is a common technique. All relevant operations can be performed quickly. Hence, an ordinary array (of size 2n) can be used to map power automaton’s states into arbitrary data (for example, an information if a given set was visited earlier during the search). Since the array must fit into memory, n must be small. 5.1.5. Sparse mapping In case where only a subset (of unknown size) must be mapped, sparse data structures should be used. Such structures include tree-based dictionaries (like various BSTs) and hash tables. Trees require a strict weak ordering on elements (that is a proper ’ < ’ predicate must be supplied). Note that x ¼ y () :ðx < yÞ^:ðy < xÞ. Hash tables require an equality predicate (=). Also, it is required that hash value h can be calculated for each element. Tree-based dictionaries typically guarantee that insertions and retrievals perform H(log2m) key comparisons, where m is the number of stored keys. Hence, the complexity of performing key comparisons must be included in the estimation of the total com- plexity (this fact seems to be often omitted). Hash tables promise insertions and retrievals in O(1) time on average. Once more, the complexity of comparing keys and calcu- lating hashes must be taken into account. Using a trie data structure is also an option, but no efficient implementations seem to be available. An optimized implementa- tion based on a compact two-array approach would be suitable for our needs. There exist standard implementations of tree-based and hash- based dictionaries and we will not delve into further details here. We used std<map from STL library (which implements red–black trees) and boost<unordered_map from the boost library (Björn, 2005). 5.1.6. State-set representation for the sparse mapping Since only a small subautomaton of the power automaton must typically be stored, n can be much bigger (>80). In effect, power automaton’s states can no longer be encoded into a single integer value. An alternative representation must be devised. Essentially, we need to represent subsets of {1, 2, . . . , n}. We will now investi- gate various data-structures that can be used for this purpose. Let a data structure S represent a set S # {1, 2, . . . , n}. By itera- tion over S we mean enumerating elements of S, preferably in sorted order. So if S represents the set {1, 2, 3}, iterating over S yields elements 1, 2, 3, and signalizes that no more elements are present. 5.1.7. Tree-based sets Using the standard set structures based on AVL or red–black trees (like std<set from the C++standard library) as state repre- sentation severely decreases performance. Also, memory footprint is unacceptable. 11750 R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 Author's personal copy 5.1.8. Sorted arrays Arrays of fixed size n (or dynamically sized vectors) can be used as long as the sorted order (or any other canonical order that en- ables consistent comparisons required by the dictionary data- structures) is preserved. This solution also has some drawbacks, as comparing two states requires comparing as many as n integers. Furthermore, each power automaton transition requires sorting the resulting set (array) of states. Also taking h(state) requires com- bining hashes of all values (n/2 on average). 5.1.9. Fast bit-vector It turns out that an efficient, yet simple, data structure can be used to store individual states of power-automaton. It will be a mix of three described approaches. It encodes states into several integer values (using a modification of the described technique). Contained states can be iterated in sorted order (like in case of std<set) that is preserved with no additional overhead (like ex- plicit sorting in case of using arrays). Finally, it uses a plain array to store the values. The most important improvement is that efficient comparisons and hash calculations can be performed. Normally, individual val- ues of the contained sets are compared. Here the data-structure is treated as a sequence of bits (or integers). This way, the actual ele- ments do not have to be decoded. The raw integers representing subsets of values are compared. This is a vast improvement, since as many as 32 (or 64 depending on the CPU) values are compared in parallel. Hashes are calculated similarly. This structure is also much more compact: only 1 bit per state is needed. An array requires at least 8 bits (typically 16). The memory overhead of the standard set data-structures is much greater (at least 70 bits). As an improvement for very sparse sets, iteration can be modified: rather than checking bits one by one, bit-masks of certain size (16, 8) can be applied to cull large empty chunks. This is an extension of a basic approach that skips empty (zero) blocks. The only problem with this solution is that the iteration re- quires amortized H(n) time in the worst case (typically amortized time is constant). Creation of the data-structure takes n/32 time (memory must be cleared), so improving this element might be beneficial. This solution can be extended by storing an additional unsorted array of values. It would speed the iteration process up, but memory consumption would be increased by an unacceptable factor (each stored value would require at least 9 bits instead of 1). 5.1.10. Bit-vector indexed with an implicit binary tree Iteration speed can be increased, only slightly affecting the memory footprint. Exploiting the fact that a fixed universe of keys is considered, an index (represented by an implicit tree) can be built over a bit-vector. The main advantage over standard (explicit) trees is that the pointers to children are implicit (that is, they are not stored in memory). It can be implemented similarly to a bin- ary-heap: an array (or bit-vector) A is used, elements A[2i] and A[2i + 1] represent children of node i (nodes are numbered from 1). Node A b i2c � � is the parent of i and A biþ12 c � � is the parent of the right-sibling (actually this represents sibling’s or grand-sibling’s parent; the important thing is that i + 1 is on the same level as i (unless i is the rightmost node)). Each node contains one bit of information, answering the fol- lowing question: is my sub-tree empty? Specifically, if the root node’s sub-tree is empty, it means the whole tree is empty. If leaf p is non-empty, it means that the value p � n is present in the set. Note that 32 (or 64) nodes can be encoded in one integer. For simplicity’s sake we will assume that n is a power of 2. There are log(n) + 1 levels in the tree. Number of nodes at level L equals 2L, so the total number of nodes is 2n. Note that this struc- ture can be accessed just as a plain bit-vector (values A[n] . . . A[2n � 1] directly represent contained elements). Hence, checking if x 2 S takes O(1) time. Fig. 1 shows the structure for n = 8. This structure can repre- sent subsets of {0, . . . ,n � 1}. The figure shows the state of the structure after inserting elements 1 and 6. Gray color represents marked nodes (that is, nodes with value 1). Note that respective nodes 9 and 14 are gray. Also, the paths from these nodes to root are gray. Nodes 8, . . . , 15 directly represent the elements in a current set. Adding an element (we are not interested in removal in our applications, so this operation will not be described) requires updating all ancestors up to the root. Adding an element to an empty structure takes logn time. Next time an element is added, when a non-empty node is found, a path from this node to the root is already updated, so the overhead is smaller. It seems that the cost of populating the structure with n distinct values is nlogn. Noticing that each node (including internal nodes) will be marked as non-empty only once, yields a more precise result, 2n. Let us investigate a more general case: any number of elements is added. Clearly, the amortized cost of adding new elements is non-decreasing in the number of elements in the set. We have already shown that when all elements are added, the cost is linear. This value cannot be greater for a smaller number of elements, since the cost must be positive. The amor- tized complexity of inserting one element is logarithmic in the worst case, but the total complexity (that is, the sum of costs of all operations) is never worse than linear. Of course, this is nothing new – the same thing can be achieved with an ordinary bit-vector. The complexity of iteration is more interesting. Iteration uses the function shown in Listing 5. Algorithm 5 UPPERBOUND(tree, n, elem) 1: Input: n-element set, its subset represented by tree and an element elem from the set 2: Output: smallest element next present in the set such that next P elem 3: node n + elem 4: if (tree[node]) 5: return elem 6: prev_node node 7: while ((not tree[node]) and (node != root)) 8: prev_node node 9: node parent (node) 10: return down (tree, n, prev_node, node) Fig. 1. Exemplary bit-vector indexed with a tree. R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 11751 Author's personal copy Algorithm 6 DOWN(tree, n, from, node) 1: if (leaf (node) and (not tree[node])) 2: return node-n// found 3: if (leaf (node) and tree[node]) 4: return-1// found nothing 5: if tree[left (node)] and left (node) > from 6: return down (tree, n, node, left (node)) 7: return down (tree, n, node, right (node)) The UPPERBOUND goes up to the root searching for the non-empty nodes. When a node is found, we go down, following the leftmost path (leading to leaves with indices greater than the start node) of non-empty nodes. Clearly, at most 2log n nodes are visited. This happens, for example, when only one value is present. The entire structure can be iterated using the function ITERATE (see Listing 7). Algorithm 7 ITERATE (tree, n) 1: Input: structure S 2: Output: next elements of S 3: elem = �1 4: while(elem != �1) 5: elem = UPPERBOUND(tree, n, elem + 1) 1: yield elem // the next element was found Let us investigate the amortized complexity of iterating over a full set. This is trivial, since UPPERBOUND function returns immedi- ately in line 5. Hence, the amortized complexity is constant when all elements are present. We already mentioned that when one ele- ment is present, the amortized complexity (as well as total) is 2log n. Instead of going up to a parent we can visit our grand-sibling’s parent without skipping any valuable nodes (sometimes we could visit the grand-sibling itself if it is non-empty, but this happens only once per UPPERBOUND call and lowers performance in the worst case). The remaining part of the algorithm is unchanged: if the cur- rent node is non-empty (marked), we start moving down. Other- wise, we continue moving right-up. This small change is important, as it enables us to perform more detailed complexity analysis. Note that the cost (counted in the number of visited nodes) between two leaves p, q is bounded by 3log(jp � qj). The search process can be divided into two phases. The search starts from node p. 1. Right-up jumps are performed until a non-empty node is found. Successive right-up jumps cull sub-trees of expo- nentially growing sizes: 1, 2, 4 etc. Since there are jp � qj empty leaves between the considered leaves, it is enough to perform logjp � qj jumps. 2. Now, the left-most non-empty path is followed. For each performed up or right-up jump a down-move must be performed (there were at most logjp � qj such jumps). Each down-move requires checking two nodes. It can be seen that at most 3log(jp � qj) nodes will be accessed. Let us now investigate the worst case. Assuming there are at least two elements (and one of them equals 0 for simplicity), they are iterated left-to-right: yi is the number of elements checked dur- ing step i (that is, during the ith call to UPPERBOUND). Let us put xi = yi+1 � yi "i < n. Now, the cost of iterating through all elements equals P i xi. Note that P i yi 6 n, but we will assume P iyi ¼ n, which is the worst case. The total cost is maxxð P i logðxiÞÞ¼ maxxðlogðPi xiÞÞ. From loga- rithm’s monotonicity we only have to maximize the product, which happens when xi = xj "i, j. This leads to function a n/a, where a is a parameter. Differentiating shows that it is maximized for a = e. Taking into account that we operate on positive integer values, we finally obtain that the cost is bounded by Oðlog 3 n 3 � � Þ¼ O n3 log 3 � � ¼ OðnÞ. The worst case amortized complexity has been improved from H(n) to H(logn), while worst case total complexity is still O(n). Memory consumption is increased twice (two bits per value are required). 5.1.11. Other data-structures Van Emde Boas Tree is an extension of the indexing tree concept described above. This heavy, recursive data-structure (each node contains a smaller VEB-Tree that is used as an index) enables iter- ation of s in H(jsjlog logn), so it guarantees H(log logn) amortized complexity. This structure is complex, so it would be an improve- ment only for much bigger n. It is of no use in our applications. 5.1.12. Partial power-set automaton The following technique can be used to reduce the amount of computations involved in calculating power automaton’s transi- tions. This, as far as we know, novel technique is most useful for small n. Let us define a partial power automaton of A for X � Q as PXðAÞ¼ ð2Q ; A; DjXÞ. In other words, the transition function domain is restricted to X. Note that the co-domain does not change. It is obvious that S # 2Q ^ [ S ¼ Q )PðAÞ¼ [ x2S PxðAÞ: From a mathematical standpoint it is a trivial tautology, but it turns out that it can be useful for our purposes. Let us consider a simple case. Put Q = {1, . . . , 32}, X = {1, . . . , 16}, Y = {17, . . . , 32}. Then jPXðAÞj¼ jPYðAÞj¼ 216 � 1. Clearly, values of DX and DY can be precomputed in 32 � 28 ( n ffiffiffiffiffi 2n p in general). Later, they can be used to construct the transition function of the entire power automaton by taking a union of transitions of the partial power automaton. Assuming that the union operation is faster then per- forming the transition in the original automaton, a speedup will occur. This is a low-level optimization, but experiments show that it can boost overall performance by a factor of 10. It is indeed possible to perform a fast set-union operation. When the represented universe is fixed (in this case: {1, . . . , 32}), a set can be represented as its characteristic vector, en- coded by an integer. Set-union is done by bit-wise OR operation, so the union of two states of size 32 is performed in one CPU instruction. This approach can be generalized. Let m (dividing n, for simplic- ity) be the maximum value such that m2n/m transitions can be pre- computed. Let us assume that our machine’s word size is 32 bits. There are d n32e words necessary to store one transition value. We need m values, so we need to bit-or m values, which takes mn32 oper- ations. Normally we would need to perform exactly n transitions in the original automaton. For m < 32 this technique should be faster. When m = 2, n = 32, exactly two results must be united. This can be done using one bit-or operation. In the result three integer oper- ations must be performed (instead of n/2 on average). The prepro- cessing phase takes 2 � 216 operations. 5.2. Possible further improvements As it was noted in Volkov (2008), an average length of SSW should be quite low (comparing to the conjectured (n � 1)2). According to the considerations of Volkov (referring to the paper 11752 R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 Author's personal copy of Higgins (1988)), the expected SSW length is O(n). More precisely, it can be proved that a randomly chosen n-state autom- aton with a sufficiently large alphabet is synchronizing with prob- ability 1 as n goes to infinity and the length of its SSW does not exceed 2n. Therefore, statistically, only a small part of the power automaton must be visited in the search process. Recently, Skvort- sov and Tipikin (2011) provided experimentally the average length of SSW for a random n-state automaton over binary alphabet – it is O(n0.55). Regarding to the above facts, the described algorithm is not only exact, but also turns out to be quite efficient in an average case. Also, for small n, such an algorithm can be more effective than some polynomial-time solutions. One more improvement can be introduced. It aims at reducing the runtime complexity for certain types of automata (such as Černý automaton) whose SSWs are of the form wkv, w, v 2 A⁄. A similar heuristics was described in Trahtman (2006) to enhance the greedy algorithm. For each word wa such that jd(Q, w)j > jd(Q, wa)j, w 2 A⁄, a 2 A, the powers of this word are checked. It may turn out that (wa)k is a synchronizing word (just as in case of the Černý automaton). Actually, many slowly synchronizing automata (i.e. with long SSW) fall into this category. Note that by using this heuristics only a small fraction of the power automaton is visited, even though the length of SSW is quadratic and (normally) vast number of states would have to be traversed. Experimental data show that employing this heuristics for the Černý automata reduces the number of visited states to Hð1:5nÞ Hð ffiffiffiffiffi 2n p Þ. 5.3. Performance comparison Table 1 shows how the described optimizations affected perfor- mance. The results are compared with the popular synchronization tool, TESTAS (v. 1.0). Fig. 2 shows the results graphically. The measurements clearly show that the optimizations are very effective. Version 4 (using raw arrays of integers and partial power automaton concept), while significantly faster, cannot be used for larger automata. Other solutions, while slower, scale better and can be used for much larger automata, as long as the reachable power automaton size fits in memory. Algorithm 3 (using hash tables and the described fast bit-vec- tor) should be used for medium size automata. It can be more effective than algorithm 4 for smaller automata if the power automaton is relatively small. Approach 4 is superior for small (n < 30), slowly-synchronizing automata. Note that all of these approaches are faster than the algorithm used in TESTAS, due to the described complexity improvements. Fig. 2 shows performance for growing n. It must be noted that the memory conservation plays a major role here. In case of slowly synchronizing automata the compact structures enable cache-friendly (therefore fast) computations for small n, while for larger n they are necessary for the algorithm to work, due to memory requirements. Performance of algorithm 3 supports this conclusion. Algorithm 3 should be as fast as standard implementations using power automaton. Unlike them, it can handle larger automata (n > 30) as long as the reachable power automaton is small. Computations were performed with the use of a laptop manu- factured in 2004 (Athlon XP-M 2800+(@2.13 GHz), 768 MB of DDR RAM (@133 MHz)). 6. New greedy algorithm In this section we introduce a new greedy algorithm. It will be based on SYNCHROP, which itself is based on EPPSTEIN. Both algorithms are focused on choosing a pair of states, which should be synchro- nized in the next step and applying the proper word to the whole automaton. We can reformulate this task equivalently in terms of the pair automaton states (both EPPSTEIN and SYNCHROP use this struc- ture) as choosing the state, which should be transformed to a sin- gleton 0 state. From now on all procedures described below are regarded to the pair automaton A0 ¼ ðQ 0; A; d0Þ of the original input automaton A¼ðQ; A; dÞ. The choice of state depends on the set P of states that we are actually in, that is, the set P = d0(Q0, w), where w is the catenation of the words found in all previous steps. Such set will be called the active states set. The choice is based on the evaluation of the ac- tive states arrangement in the pair automaton. In SYNCHROP the eval- uation is based on the following heuristics: Heuresis 1. Let us define the distance d(p) of state p = {s1, s2} to the singleton state 0 as dðpÞ¼ min w2R� ðjwj : dðs1; wÞ¼ dðs2; wÞÞ: ð3Þ Let w be a synchronizing word for some state q. The bigger is the difference between d(p) and d(d(p, w)), the more profitable is the selection of w in the next algorithm step, because the distance of d(p, w) to the singleton state is smaller. The idea enclosed in the above heuristics is utilized in SYNCHROP in the following way: we compute the differences between states p and d(p, w), where w is the shortest synchronizing word for the pair q, as Dqðp; wÞ¼ dðdðp; wÞÞ� dðpÞ if p – q; 0 if p ¼ q: ð4Þ Table 1 Runtime (in seconds) for different implementations and different sizes of input Černý automata. No algorithm version C12 C14 C18 0 TESTAS: SSW < 1.0 s 18 s Time-out 1 Set of int 0.34 s 3.3 s 60 s 2 hash_set of vector of int 0.07 s 0.34 s 9.7 s 3 hash_set of fast_bitset 0.01 s 0.07 s 2.0 s 4 Int array 0.0 s 0.01 s 0.2 s Fig. 2. Graphical representation of performance for Cn. R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 11753 Author's personal copy We compute Dq(p, w) for all active states except the singleton state. Let X be the set of all active states in the pair automaton. We define U1ðwÞ¼ X p2X Dqðp; wÞ: ð5Þ Having U1(w) for all the shortest words that synchronize pairs of states we choose the one with the smallest U1 and apply it to the automaton. The modified version of SYNCHROP, SYNCHROPL, uses U2 function – a modification of U1 – which takes into account the length of the word: U2ðwÞ¼ X p2X Dqðp; wÞþ jwj ¼ U1ðwÞþ jwj: ð6Þ Thanks to this penalty component shorter words are preferable. This is a good solution in case where there are two or more candi- date words with the same U1 value. Let us consider the complexity of SYNCHROP. The preprocessing part is the same as in EPPSTEIN and can be done in O(n3 + jAjn2). The new part is the choice of s1 and s2, which are to be synchro- nized. This is done O(n) times. To compute U1(w) we have to pro- cess all active states in order to compute D((s1, s2), w) values, which are O(n2). The U1 value has to be computed for all pairs of states of the input automaton (O(n2)). Hence, the total complexity of the SYNCHROP is O(n3 + jAjn2 + n(n2 � n2)) = O(n5 + jAjn2). 6.1. FASTSYNCHRO – a better SYNCHROP Comparing to NATARAJAN and EPPSTEIN, SYNCHROP and SYNCHROPL give good results, but have high complexity. In this section we present a modification of SYNCHROP with improved complexity and still pre- serving the quality. This algorithm will be called FASTSYNCHRO. The first modification regarding to SYNCHROP is the way that synchronizing word is created. Instead of computing U for the words synchronizing pairs of states of A, we will compute it for all letters from A. This modified function will be denoted by U3. Then, we will choose the letter that minimizes U3. This letter will be denoted by s. Let X be the set of active states in the pair automaton, A – the alphabet and d(p) – the distance of p to the singleton state. We define U3ðlÞ¼ X p2X ðdðd0ðp; lÞÞ� dðpÞÞ; p 2 X; l 2 A; s ¼ arg min l2A ðU3ðlÞÞ: Letter s is applied to all active states and added at the end of the currently found word, increasing its length by 1. The drawback of this solution is that it does not guarantee us to finally find the synchronizing word – we do not know if the number of active states will decrease after some number of steps. Therefore, we will use U3 only when it improves the arrangement of all active states in the pair automaton (that is, when U3 < 0). If U3 P 0, we will use U2 function for finding the word that guarantees the synchronization, hence necessarily decreases the number of active states. However, we introduce one restriction. In SYNCHROP the greatest impact on the complexity has the computation of U for all pairs of states and all the shortest words that synchronize these pairs. This can be done in O(n4). We will reduce it to O(n3) by reducing the number of processed words from square to linear order of magnitude by choosing only n shortest words synchro- nizing the pairs (if there is less than n words, we choose all of them). Such a choice, inspired by EPPSTEIN, is simple in realization and gives better results in average than other choices of words that were checked by us. The above modifications are given in Listing 8. Algorithm 8 FASTSYNCHRO(A) 1: Input: automaton A¼ðQ; A; dÞ 2: Output: synchronizing word w (if exists) 3: w e 4: A0 pair automaton of A 5: if A is not synchronizing return null 6: perform Eppstein preprocessing// see Section 4.4 7: X Q// X is the set of active states 8: count 0// a counter 9: while (jXj > 1) 10: a arg minl2A{U3(l)} 11: if (U3(a) < 0 AND count++ < jQj2) 12: w w.a; X d(X, a) 13: else 14: compute U2 for minfjQj; jXj2�jXj 2 g shortest words synchronizing the active states (denote Y for this set of words) 15: v argminy2Y{jyj} 16: w w.v; X d(X, v) 17: return w Let n = jQj. The complexity of line 4. is O(jAjn2). In line 5. we check if an automaton is synchronizing. This requires the use of BFS on a pair automaton with reverse transitions. This takes O(jAjn2). Eppstein preprocessing in line 6. takes O(jAjn2 + n3). Now consider the instructions in the while loop. The cost of line 10. is the cost of computing U3 for all letters – O(jAjn2). Application of the letter or word to all active states (lines 12. and 16.) is O(n), thanks to the Eppstein preprocessing. The cost of line 14. is O(n3), thanks to the restriction on the number of processed words. Now it remains to compute the number of while loop calls. Each application of v to the set of active states reduces their number at least by 1, so lines 14. – 16. will be executed at most n � 1 times. Alternatively, line 12. may be executed, but it does not necessarily reduce the number of active states. Therefore, to keep a tight rein on the total complexity, we set the restriction on the number of line 12. executions. In tests we have noticed that setting the limit to n2 had no influence on the algorithm results. Summarizing: the total cost of lines 4. – 6. is O(jAjn2 + n3), the cost of the instructions inside the while loop is O(n � n3) + O(n2 � jAjn2). This gives us the following theorem. Theorem 1. FASTSYNCHRO works in O(jAjn4) time complexity. 6.2. Experiments and comparison In this section we present the results of the experiments on heuristic algorithms. We focus on the efficiency (the running time) and the quality (the length of the synchronizing word found). We also make some remarks on FASTSYNCHRO algorithm. We tested five heuristic algorithms: EPPSTEIN, CYCLE, SYNCHROP, SYNCHROPL and FASTSYNCHRO. 6.2.1. Efficiency The efficiency has a great impact on the algorithms usability. In this subsection we present the efficiency comparison for EPPSTEIN, NATARAJAN, SYNCHROP and FASTSYNCHRO. Algorithms were tested for n = {10, 20, . . . , 300}. For each n one hundred random automata were generated such that 8a 2 A 8p; q 2 Q Prðdðp; aÞ¼ qÞ¼ 1n. If the generated automaton was not synchronizing, the procedure 11754 R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 Author's personal copy was repeated. Tests were performed for jAj 2 {2,10} (Figs. 3 and 4). We have also performed tests for Černý automata (Fig. 5). It is clearly seen from Fig. 3 that SYNCHROP, due to its high com- plexity, is worse than other algorithms. The runtimes for other algorithms are comparable. The results for jAj = 10 does not differ much from these with jAj = 2. Running times are of course higher, but still the differences between NATARAJAN, EPPSTEIN and FASTSYNCHRO are small. Tests on Černý automata were performed to check how the algorithms work for automata with long synchronizing words. We can see noticeable decrease of efficiency in case of FASTSYNCHRO. EPPSTEIN and NATARAJAN work much faster in this case. 6.2.2. Quality In order to compare the algorithms in the widest possible con- text, the tests were performed for three different quality measures. If the number of all automata for a given number of states and alphabet size was reasonable, the algorithms were tested on all such automata. If the number of such automata was too big, we re- duced the tests to a subset of all possible random automata. All algorithms were run on the same set of automata. The first quality measure is the mean difference between the length of the word found by the algorithm and the SSW length. De- note by ALGðAÞ the word returned by algorithm ALG for automaton A. Let X be the set of all automata that were given as the input to ALG. Formally, we can define the first measure as M1ðXÞ¼ P A2XðjALGðAÞj� jSSWðAÞjÞ jXj : ð7Þ Table 2 shows that CYCLE and EPPSTEIN, despite their fast speed, does not give a good results in terms of M1. SYNCHROP and SYNCHROPL, based on Heuristics 1, are much better. FASTSYNCHRO is comparable to them and sometimes outperforms them (n = 5, 6, 10). The second quality measure, M2 is the ratio of cases in which ALG found the SSW to all cases: M2ðXÞ¼ P A2X½jALGðAÞj¼ jSSWðAÞj jXj ; ð8Þ where [expr] = 1 if expr is true and 0 otherwise. Notice how the quality decreases with increasing the alphabet size from 2 to 10. The ordering of the algorithms is the same as in case of M1 measure. To test the algorithms for automata with larger number of states, we need a measure which does not involve computing the Fig. 3. Efficiency for automata with jAj = 2. Fig. 4. Efficiency for automata with jAj = 10. Fig. 5. Efficiency for Černý automata. Table 2 Quality of algorithms in terms of M1, M2 and M3. C, sc Ep, SP, SPL, FS correspond to CYCLE, EPPSTEIN, SYNCHROP, SYNCHROPL and FASTSYNCHRO algorithms. n,jAj C EP SP SPL FS jXj Measure M1 3, 2 0.20 0.16 0 0 0 all automata 4, 2 0.36 0.33 0.06 0.03 0.04 All automata 4, 3 0.45 0.42 0.08 0.05 0.05 All automata 5, 2 0.64 0.55 0.21 0.17 0.13 All automata 6, 2 0.91 0.77 0.38 0.34 0.24 All automata 10, 2 1.97 1.63 1.00 0.96 0.78 105 random automata 10, 10 1.78 1.77 0.72 0.69 0.54 105 random automata 20, 2 4.18 3.49 2.18 2.07 2.01 105 random automata 20, 10 3.45 3.31 1.61 1.54 1.54 105 random automata Measure M2 3, 2 0.80 0.84 1 1 1 All automata 4, 2 0.71 0.72 0.94 0.97 0.97 All automata 4, 3 0.64 0.65 0.93 0.96 0.95 All automata 5, 2 0.57 0.60 0.85 0.87 0.89 All automata 6, 2 0.47 0.51 0.76 0.77 0.82 All automata 10, 2 0.25 0.28 0.49 0.50 0.57 105 random automata 10, 10 0.12 0.12 0.41 0.43 0.54 105 random automata 20, 2 0.08 0.10 0.23 0.24 0.28 105 random automata 20, 10 0.02 0.02 0.13 0.14 0.16 105 random automata Measure M3 50, 2 26.21 24.44 21.75 21.53 21.93 104 random automata 50, 10 16.32 15.49 12.84 12.71 13.00 104 random automata 100, 2 40.75 37.53 33.16 32.84 33.95 103 random automata 100, 10 25.30 23.41 19.84 19.61 20.78 103 random automata R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 11755 Author's personal copy SSW length. Therefore, as M3 we took the mean length of the syn- chronizing words found by a given ALG. The use of this measure is meaningful only in relative comparison of two or more algorithms. M3ðXÞ¼ P A2XjALGðAÞj jXj : ð9Þ FASTSYNCHRO, in terms of M3, gives a little worse results than SYNCHROP and SYNCHROPL, however these results are still much better than those of EPPSTEIN and CYCLE. 6.3. Analysis of FASTSYNCHRO behavior In this section we investigate the behavior of FASTSYNCHRO. The analysis will allow us to explain the decrease of FASTSYNCHRO effi- ciency shown in Fig. 5. We will check what is the impact of differ- ent algorithm’s parts on the process of building the synchronizing word. Recall that in FASTSYNCHRO a synchronizing word can be ob- tained in two ways: first one is choosing the letter a 2 A and apply it to the set of active states. The other is to use U2 to find the pair of states and transform the set of active states by the word synchro- nizing this pair. We will refer to these two ways as to the first and the second part of the algorithm. We have performed an experiment for random automata. By g1 (resp. g2) we denote the number of executions of the first (resp. second) part of the algorithm. By k we denote the mean length of the synchronizing word found by FASTSYNCHRO and by k⁄ the esti- mated value of the SSW length for random automata over binary alphabet (Skvortsov & Tipikin, 2011). Notice that each execution of the first part of the algorithm cor- responds to the generation of exactly one letter added to the syn- chronizing word that is constructed. The value k � g1 expresses the number of letters added in result of the second part execution. From Table 3 we can see that with the increase of the number of states the fraction of the second part execution also grows. The ratio k�g1g2 is the mean length of the word added during the second part execution. This value remains relatively small and grows slightly with the increase of the number of states. When jAj is increased to 10, we can see that the influence of the second part decreases. Despite that its frequency is almost the same as in case jAj = 2, the mean length of the word added by each execution is 2–3 times shorter. The same experiment was performed for Černý automata (Table 4). These experiments explain why the runtime depends so much on the length of the synchronizing word found by the algorithm: the number of executions of the algorithm’s first part increases significantly. Also, the words found by the execution of the algo- rithm’s second part are not short anymore. Fortunately, the auto- mata with long SSWs are very rare, so the case of Černý automata is exceptional. 7. Conclusions We presented some efficient data structures for exact (expo- nential) synchronizing algorithm. Their application into the well-known algorithm that uses a power-set automaton makes the algorithm more effective than existing implementations. We also presented a new, greedy synchronizing algorithm and we compared it with some previously known greedy algorithms. Experiments show that our FASTSYNCHRO algorithm in general works better (that is, finds shorter synchronizing words) and usually works in a comparable time or faster than other methods. For lar- ger automata FASTSYNCHRO works twice as long as EPPSTEIN, but it finds much shorter synchronizing words. When one wants to find a synchronizing word, two factors have to be considered: the quality (the length of the synchronizing word) and the time. If time is a key issue, the optimal choice would be the EPPSTEIN algorithm. But if the quality is much more important (and this is usually the case in industrial testing of electronic cir- cuits, when one has to apply the same synchronizing word to thou- sands or millions copies of circuits), the best choice is to use our new FASTSYNCHRO algorithm. References Ananichev, D. S., & Volkov, M. V. (2003). Synchronizing monotonic automata. Lecture Notes in Computer Science, 2710, 111–121. Björn, K. (2005). Beyond the C++ standard library: An introduction to boost. Addison- Wesley. Broy, M., Jonsson, B., Katoen, J.-P., Leucker, M., & Pretschner, A. (2005). Model-based testing of reactive systems model-based testing of reactive systems. Advanced lectures. Lecture Notes in Computer Science, 3072. Černý, J. (1964). Poznámka k. homogénnym experimentom s konecnymi automatmi. Matematicko-fyzikálny Časopis Slovenskej Akadémie Vied, 14, 208–215. Černý, J., Pirická, A., & Rosenauerová, B. (1971). On directable automata. Kybernetika, 7(4), 289–298. Deshmukh, R. G., & Hawat, G. N. (1994). An algorithm to determine shortest length distinguishing, homing, and synchronizing sequences for sequential machines. In Proc. Southcon 94 conference (pp. 496–501). Eppstein, D. (1990). Reset sequences for monotonic automata. SIAM Journal of Computing, 19(3), 500–510. Fukada, A., Nakata, A., Kitamichi, J., Higashino, T. & Cavalli, A. R. (2001). A conformance testing method for communication protocols modeled as concurrent dfsms. In ICOIN (pp. 155–162). Higgins, P. M. (1988). The range order of a product of i transformations from a finite full transformation semigroup. Semigroup Forum, 37, 31–36. Hyunwoo, C., Somenzi, F., & Pixley, C. (1993). Multiple observation time single reference test generation using synchronizing sequences. In Proc. IEEE European conf. on design automaton (pp. 494–498). Kari, J. (2002). Synchronization and stability of finite automata. Journal of Universal Computer Science, 8(2), 270–277. Klyachko, A. A., Rystsov, I. K., & Spivak, M. A. (1987). An extremal combinatorial problem associated with the bound of the length of a synchronizing word in an automaton. Cybernetics and Systems Analysis, 23(2). translated from Kibernetika, No 2, 1987, pp. 16–20, 25. Lee, D., & Yannakakis, M. (1996). Principles and methods of testing finite state machines – a survey. In Proceedings of the IEEE (Vol. 84, pp. 1090–1123). Natarajan, B. K. (1986). An algorithmic approach to the automated design of part orienters. In Proc. IEEE symposium on foundations of computer science (pp. 132– 142). Olschewski, J., & Ummels, M. (2010). The complexity of finding reset words in finite automata. Lecture Notes in Computer Science, 6281, 568–579. Pin, J.-E. (1983). On two combinatorial problems arising from automata theory. Annals of Discrete Mathematics, 17, 535–548. Table 3 Behavior of FASTSYNCHRO for random automata. n,jAj g1 g2 k k⁄ k � g1 k�g1 g2 jXj 10, 2 6.52 0.44 7.52 6.92 0.99 2.28 10 000 20, 2 10.11 0.77 12.26 10.13 2.15 2.80 10 000 50, 2 16.12 1.63 22.07 16.77 5.95 3.65 10 000 100, 2 21.69 2.83 34.07 24.55 12.38 4.38 1 000 200, 2 27.38 4.56 51.81 35.94 24.42 5.35 1 000 10, 10 4.01 0.02 4.04 n/a 0.02 1.00 10 000 20, 10 6.68 0.21 6.91 n/a 0.22 1.07 10 000 50, 10 12.00 0.76 13.02 n/a 1.01 1.34 10 000 100, 10 17.60 1.92 20.88 n/a 3.28 1.71 1 000 200, 10 24.32 4.16 32.72 n/a 8.40 2.02 1 000 Table 4 Behavior of FASTSYNCHRO for Černý automata. n g1 g2 k k � g1 k�g1 g2 10 24 3 81 57 19 20 67 4 361 294 73 50 211 5 2401 2190 438 100 520 6 9801 9281 1546 200 1238 7 39601 38363 5480 11756 R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 Author's personal copy Pixley, C., Jeong, S.-W., & Hachtel, G. D. (1994). Exact calculation of synchronizing sequences based on binary decision diagrams. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 13(8), 1024–1034. Pomeranz, I., & Reddy, S. M. (1998). On synchronizing sequences and test sequence partitioning. In Proc. 16th IEEE VLSI test symposium (pp.158–167). Ponce, A.M., Csopaki, G., & Tarnay, K. (1994). Formal specification of conformance testing documents for communication protocols. In 5th IEEE international symposium on personal, indoor and mobile radio communications (Vol. 4, pp. 1167–1172). Roman, A. (2009). Synchronizing finite automata with short reset words. Applied Mathematics and Computation, 209(1), 125–136. Skvortsov, E., & Tipikin, E. (2011). Experimental study of the shortest reset word of random automata. Lecture Notes in Computer Science, 6807, 290–298. Trahtman, A. N. (2006). An efficient algorithm finds noticeable trends and examples concerning the cerny conjecture. Lecture Notes in Computer Science, 4162, 789–800. Trahtman, A. N. (2009). The road coloring problem. Israel Journal of Mathematics, 1(172), 51–60. Volkov, M. V. (2008). Synchronizing automata and the cerny conjecture. Lecture Notes in Computer Science, 5196, 11–27. Zhao, Y., Liu, Y., Guo, X., & Zhang, C. (2010). Conformance testing for is-is protocol based on e-lotos. In IEEE int. conf. on information theory and information security (pp. 54–57). R. Kudłacik et al. / Expert Systems with Applications 39 (2012) 11746–11757 11757 
work_3gmfusebffdg3m6qdqgyme5u7e	----	Mining the data from a hyperheuristic approach using associative classification This is a repository copy of Mining the data from a hyperheuristic approach using associative classification. White Rose Research Online URL for this paper: http://eprints.whiterose.ac.uk/75051/ Version: Submitted Version Article: Thabtah, Fadi A. and Cowling, Peter I. orcid.org/0000-0003-1310-6683 (2008) Mining the data from a hyperheuristic approach using associative classification. Expert systems with applications. pp. 1093-1101. ISSN 0957-4174 https://doi.org/10.1016/j.eswa.2006.12.018 eprints@whiterose.ac.uk https://eprints.whiterose.ac.uk/ Reuse Items deposited in White Rose Research Online are protected by copyright, with all rights reserved unless indicated otherwise. They may be downloaded and/or printed for private study, or other acts as permitted by national copyright laws. The publisher or other rights holders may allow further reproduction and re-use of the full text version. This is indicated by the licence information on the White Rose Research Online record for the item. Takedown If you consider content in White Rose Research Online to be in breach of UK law, please notify us by emailing eprints@whiterose.ac.uk including the URL of the record and the reason for the withdrawal request. Mining the data from a hyperheuristic approach using associative classification Fadi Thabtah a,*, Peter Cowling b a Department of Computing and Engineering, University of Huddersfield, Huddersfield, UK b MOSAIC Research Centre, University of Bradford, Bradford, UK Abstract Associative classification is a promising classification approach that utilises association rule mining to construct accurate classification models. In this paper, we investigate the potential of associative classifiers as well as other traditional classifiers such as decision trees and rule inducers in solutions (data sets) produced by a general-purpose optimisation heuristic called the hyperheuristic for a personnel scheduling problem. The hyperheuristic requires us to decide which of several simpler search neighbourhoods to apply at each step while constructing a solutions. After experimenting 16 different solution generated by a hyperheuristic called Peckish using different classifi- cation approaches, the results indicated that associative classification approach is the most applicable approach to such kind of problems with reference to accuracy. Particularly, associative classification algorithms such as CBA, MCAR and MMAC were able to predict the selection of low-level heuristics from the data sets more accurately than C4.5, RIPPER and PART algorithms, respectively. � 2007 Elsevier Ltd. All rights reserved. Keywords: Associative classification; Classification; Data mining; Hyperheuristic; Scheduling 1. Introduction Heuristic and metaheuristic approaches have been applied widely in personnel-scheduling problems (Blum & Roli, 2003; Cowling, Kendall, & Han, 2002). A metaheuris- tic could be defined as a recursive process which directs a simpler local search method by using different concepts for exploring and exploiting the search space in order to achieve good enough solutions (Blum & Roli, 2003). There are several different metaheuristic strategies for solving scheduling and optimisation problems such as local search, tabu search, simulated annealing and variable neighbour- hood search. Hamiez and Hao (2001) have used a tabu search-based method to solve the sport league scheduling problem (SLSP). Their implementation of the enhanced tabu search algorithm was able to schedule a timetable for up to 40 teams and its performance in term of the CPU time was excellent if compared with previous algorithms such as that of (McAloon, TretKoff, & Wetzel, 1997) that had been used for solving the same problem. Aicklen and Dowsland (2000) have used Genetic algorithms to deal with a nurse rostering problem in major UK hospi- tals, and Hansen and Mladenovic (1997) have showed that variable neighbourhood search is an effective approach for solving optimisation problems in which it generates good or sometimes near-optimal solutions in a moderate time. Cowling, Kendall, and Soubeiga (2000) and Cowling et al. (2002) argued that metaheuristic and heuristic approaches tend to be knowledge rich and require exten- sive experience in the problem domain and the selected heuristic techniques, and therefore they are expensive in term of their implementation. A new general framework to deal with large and complex optimisation and schedul- ing problems, called a hyperheuristic, has been proposed 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.12.018 * Corresponding author. E-mail addresses: F.Thabtah@hud.ac.uk (F. Thabtah), P.I.Cowling@- bradford.ac.uk (P. Cowling). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 34 (2008) 1093–1101 Expert Systems with Applications by Cowling et al. (2000). It tends to robustly find good solutions for large and complex scheduling problems and has been proven to be effective in many experiments (Cowling & Chakhlevitch, 2003; Cowling et al., 2002). A hyperheuristic approach can be described as a super- visor, which controls the choice of which local search neighbourhood to choose while constructing a solution/ schedule. A local search neighbour, also known as a low- level heuristic, is a rule or a simple method that generally yields a small change in the schedule. Often these low-level heuristics are based on normal methods of constructing a schedule such as adding an event, deleting an event or swapping two events. Fig. 1 represents the general hype- rheuristic framework that at each iteration, selects and applies the low-level heuristic that has the largest improve- ment on the objective function, i.e. LLH5 in the figure shown below. The arrows going from and to the hyper- heuristyic in Fig. 1 represent the selected low-level heuris- tics improvement values on the objective function obtained after trying them by the hyperheuristic. The training scheduling problem that we consider in this paper is a complex optimisation problem for a large finan- cial service company (Cowling et al., 2002). It involves a number of events, trainers, and locations to be scheduled over a period of time. The task is to create a timetable of courses distributed over several locations in a specific period of time using a known number of trainers. A more detailed description of the problem is presented in the next section. In this paper, our aim is to determine an applicable data mining technique to the problem of deciding which low-level heuristic to apply in a given situation, using infor- mation about heuristic performance derived earlier. In particular, we would like to answer questions like: which learning algorithm can derive knowledge that could direct the search in order to produce good solutions? To achieve our goal, we compare three associative classi- fication techniques CBA (Liu, Hsu, & Ma, 1998), MCAR (Thabtah, Cowling, & Peng, 2005), MMAC (Thabtah, Cowling, & Peng, 2004) and two popular traditional classification techniques (PART (Frank & Witten, 1998) and RIPPER (Cohen, 1995)) on data sets generated using a hybrid hyperhueristic called Peckish (Cowling & Chakhlevitch, 2003), for the trainer scheduling problem. We analyse data from several solutions of the Peckish hyperheuristic that combines greedy (best first) and ran- dom approaches. We identify that the learning task involves classifying low-level heuristics in terms whether they improved the objective function in old solutions in order to produce useful rules. These rules then will be used to decide the class attribute ‘‘low-level heuristic’’ while con- structing new solutions. We use the classification algo- rithms mentioned above to learn the rules. The training scheduling problem and different hyperheu- ristic approaches utilised to solve it are discussed in Section 2. Section 3 is devoted to the applicability of data mining classification algorithms to predict the behaviour of low-level heuristics used by the Peckish hyperheuristic. Data sets, their features and experimental results are presented in Section 4 and finally conclusions are given in Section 5. 2. The training scheduling problem and hyperheuristics A much simpler version of the training scheduling prob- lem has been solved in Cowling et al. (2002) using a Hyper- Genetic algorithm. A larger and more complex problem, which has been described in Cowling and Chakhlevitch (2003) is summarised in this section. It involves a number of events, trainers, and locations to be scheduled over a period of time. The task is to create a timetable of geo- graphically distributed courses over a period of time using different trainers, and the aim is to maximise the total pri- ority of courses and to minimise the amount of travel for each trainer. The problem is associated with a large num- ber of constraints such as: • Each event is to be scheduled at one location from the available number of locations. • Each event must start within a specified time period. • Each event can occur at most once. • Each event to be delivered by competent trainers from the available trainers. • Each location has a limited number of rooms and rooms have different capacities and capabilities. The data used to build the solutions of the training scheduling problem is real data provided by a financial firm where training is given by 50 trainers over a period of 3 months in 16 different locations. Further, there are about 200 events to be scheduled and 95 different low-level heuris- tics that can be used to build each solution. However, solu- tions given to us by the authors of Cowling and Chakhlevitch (2003) have been constructed using only 10 low-level heuristics, where each low-level heuristic repre- sents local search neighbourhoods. For example, selecting a location with the lowest possible travel penalty for a par- ticular trainer to deliver a course as early as possible corre- sponds to a low-level heuristic. In solving the trainer scheduling problem, three hype- rheuristic approaches, random, greedy and hybrid, have been used. All of these approaches aim to manage the choice of which low-level heuristics to choose during the process of building the solution. The random approach LLH1 LLH2 LLH5LLH3 LLH4 LLH6 Hyperheurisic Fig. 1. Hypreheuristic general framework. 1094 F. Thabtah, P. Cowling / Expert Systems with Applications 34 (2008) 1093–1101 selects a low-level heuristic at random from the available ones in the problem. At each choice point in the search space, commonly all low-level heuristics have an equal chance to be selected. On the other hand, the greedy approach selects the low-level heuristic that yields the big- gest improvement in the objective function. If none of the available ones improve the objective function then the algorithm will be trapped in a local optimum. The hybrid approach is named Peckish and consists of a combination of greedy and random approaches. It builds a solution by selecting a single low-level heuristic during each iteration in the search in order to apply. The choice is based on the low-level heuristic that has the largest improvement on the objective function in the problem (if one exists). In the case that none of the available low-level heuristics improve upon the objective function value, then the choice is random. In this paper, we choose a low-level heuristic from a can- didate list of good low-level heuristics. By changing the length of this candidate list and considering how it is merged, we can trade off the degree of greediness and ran- domness in the Peckish hyperheuristic. As a result, several different solutions produced by the Peckish hyperheuristic are investigated. We analysed the strategy used by the Peckish hyperheu- ristic to construct a solution and observed that all available low-level heuristics in the problem must be tested in order to record their effect on the objective function at each iter- ation and apply only a single one. Data mining could pro- vide a much quicker prediction of effective low-level heuristics at each iteration. In the next section, we investi- gate some of the popular data mining techniques for learn- ing the sets of low-level heuristics that improve the objective function and have been applied by the Peckish hyperhueristic. 3. Data mining for the selection of low-level heuristics Since we are aiming to use knowledge derived from old solutions of the problem, data mining seems an appropri- ate technique to extract that knowledge. The next task is to identify which data mining method is applicable to extract knowledge from solutions generated by the Peck- ish hyperheuristic. As mentioned earlier, the Peckish hype- rheuristic usually selects and applies the low-level heuristic that leads to the largest improvement on the objective function and this is the class we want to find. In other words, we can learn rules that predict the performance of low-level heuristics in some solution runs and use these rules to forecast which low-level heuristics the hyperheu- ristic should choose in other runs. Since we are predicting a particular attribute (low-level heuristic), as a result, supervised learning approaches such as classification are appropriate. There are many classification approaches for extracting knowledge from data that have been studied in the litera- ture Cendrowska (1987), Quinlan (1993) and Cohen (1995). Three common approaches, divide-and-conquer (Quinlan, 1987), rule induction (Cohen, 1995; Furnkranz & Widmer, 1994) and associative classification (Li, Han, & Pei, 2001; Liu et al., 1998) have been selected for our base comparison. Further, five classification techniques related to such approaches have been compared, which are PART (Frank & Witten, 1998), RIPPER (Cohen, 1995), CBA (Liu et al., 1998), MCAR (Thabtah et al., 2005) and MMAC (Thabtah et al., 2004). Our choice of these methods is based on the different schemes they use in learning rules from data sets. In the next subsection, we briefly survey these algorithms. 3.1. Associative classification Associative classification techniques employ association rule discovery methods to find the rules. This approach was introduced in 1997 by Ali, Manganaris, and Srikant (1997) to produce rules for describing relationships between attri- bute values and the class attribute and not for prediction, which is the ultimate goal for classification. In 1998, asso- ciative classification has been successfully employed to build classification models (classifiers) by Liu et al. (1998) and later attracted many researchers, e.g. (Yin & Han, 2003), from data mining and machine learning communi- ties. In this subsection we survey associative classification techniques used in this paper to generate rules from the hyperheuristic data. 3.1.1. Classification based on association (CBA) The idea of using association rule mining in classifica- tion problems was first introduced in Liu et al. (1998), in which an algorithm called CBA is proposed, which oper- ates in three main steps. Firstly, if the intended data set contains any real or integer attributes, it is disctretised using multi-interval discretisation method of Fayyad and Irani (1993). Secondly, the Apriori candidate generation step (Agrawal & Srikant, 1994) is adopted to find the potential rules. Apriori candidate generation method necessitates multiple passes, where the potential rules found in the previous pass are used for the generation of potential rules in the current pass. This repetitive scans requires high CPU time and main memory. Once all poten- tial rules are produced, the subset that leads to the least error rate against the training data set is selected to from the classifier. The selection of such subset is accomplished using the database coverage heuristic, which ensures that every rule in the classifier must cover correctly at least one training data object. 3.1.2. MCAR: multi-class classification based on association rule A recently developed associative classification algo- rithm called MCAR (Thabtah et al., 2005) employs tid- list intersections to quickly find the rule. This algorithm consists of two main phases: rules generation and a clas- sifier builder. In the first phase, the training data set is F. Thabtah, P. Cowling / Expert Systems with Applications 34 (2008) 1093–1101 1095 scanned once to discover the potential rules of size one, and then MCAR intersects the potential rules tid-lists of size one to find potential rules of size two and so forth. This rules discovery method does no require passing over the training data multiple times. In the second phase, rules created are used to build a classifier by considering their effectiveness on the training data set. Potential rules that cover certain number of training objects will be kept in the final classifier. Finally, MCAR adds upon previous rule ranking approaches in associative classification, which are based on (confidence, support, rule length) by looking at the class distribution frequencies in the training data and prefers rules that are associated with dominant classes. Experimental results showed that MCAR rule ranking method reduces rule random selection during the process of ranking the rules especially for dense clas- sification data. 3.1.3. Multi-class, multi-label associative classification (MMAC) The MMAC algorithm consists of three steps: rules gen- eration, recursive learning and classification. It passes over the training data in the first step to discover and generate a complete set of rules. Training instances that are associated with the produced rules are discarded. In the second step, MMAC proceeds to discover more rules that pass user pre- defined thresholds denoted by minimum-support and mini- mum-confidence from the remaining unclassified instances, until no further potential rules can be found. Finally, rule sets derived during each iteration are merged to form a glo- bal multi-label classifier that then is tested against test data. The distinguishing feature of MMAC is its ability of gener- ating rules with multiple classes from data sets where each of their data objects is associated with just a single class. This provides decision makers with useful knowledge dis- carded by other current associative classification algorithms. 3.2. Traditional classification approaches 3.2.1. C4.5 C4.5 algorithm was created by Quinlan (1993) as a deci- sion tree method for extracting rules from a data set. C4.5 is an extension of the ID3 algorithm (Quinlan, 1979), which accounts for missing values, continuous attributes and pruning of decision trees. As for the ID3 algorithm, C4.5 uses information gain to select the root attribute. The algo- rithm selects a root attribute from the ones available in the training data set. C4.5 makes the selection of the root based on the most informative attribute and the process of selecting an attribute is repeated recursively at the so- called child nodes of the root, excluding the attributes that have been chosen before, until the remaining training data objects can not be split any more (Quinlan, 1979). At that point, a decision tree is outputted where each node corre- sponds to an attribute and each arc to a possible value of that attribute. Each path from the root node to any give leaf in the tree corresponds to a rule. One of the major extensions of the ID3 algorithm that C4.5 proposed is that of pruning. Two known pruning methods used by C4.5 to simplify the decision trees constructed are sub-tree replace- ment and pessimistic error estimation (Witten & Frank, 2000). 3.2.2. Repeated incremental pruning to produce error reduction algorithm (RIPPER) RIPPER is a rule induction algorithm that has been developed in 1995 by Cohen (1995). It builds the rules set as follows: The training data set is divided into two sets, a pruning set and a growing set. RIPPER constructs the classifier using these two sets by repeatedly inserting rules starting from an empty rule set. The rule-growing algorithm starts with an empty rule, and heuristically adds one condi- tion at a time until the rule has no error rate on the growing set. In fact, RIPPER is a refined version of an earlier developed algorithm called Incremental Reduced Error Pruning (IREP) (Furnkranz & Widmer, 1994) that adds a post pruning heuristic on the rules. This heuristic has been applied to the classifier produced by IREP as an optimisation phase, aiming to simplify the rule set fea- tures. For each rule ri in the rule set, two alternative rules are built; the replacement of ri and the revision of ri. The replacement of ri is created by growing an empty rule r 0 i and then pruning it in order to reduce the error rate of the rules set including r0 i on the pruning data set. The revi- sion of ri is constructed similarly except that the revision rule is built heuristically by adding one condition at a time to the original ri rather than to an empty rule. Then the three rules are examined on the pruning data to select the rule with the least error rate. The integration of IREP and the optimisation procedure forms the RIPPER algorithm. 3.2.3. PART Unlike the C4.5 and RIPPER techniques that operate in two phases, the PART algorithm generates rules one at a time by avoiding extensive pruning (Frank & Witten, 1998). The C4.5 algorithm employs a divide-and-conquer approach, and the RIPPER algorithm uses rule induction approach to derive the rules. PART combines both approaches to find and generate rules. It adopts rule induc- tion approach to generate a set of rules and uses divide- and-conquer to build partial decision trees. The way PART builds and prunes a partial decision tree is similar to that of C4.5, but PART avoids constructing a complete decision tree and builds partial decision trees. PART differs from RIPPER in the way rules are created, where in PART, each rule corresponds to the leaf with the largest coverage in the partial decision tree. On the other hand, RIPPER builds the rule in a greedy fashion, starting from an empty rule, it adds conditions, until the rule has no error rate and the process is repeated. Missing values and pruning tech- niques are treated in the same way as C4.5. 1096 F. Thabtah, P. Cowling / Expert Systems with Applications 34 (2008) 1093–1101 4. Data and experimental results 4.1. Data sets and their features Data from 16 different solutions produced by Peckish hyperheuristic for the training scheduling problem were provided by the authors of Cowling and Chakhlevitch (2003). Each solution represents 500 iterations of applied low-level heuristics and is given in a text file. Twelve of the data files each represents only a single solution, whereas, each of the remaining four files represents ten combined solutions. Ten different low-level heuristics (LLH 1, LLH 2, LLH 20, LLH 27, LLH 37, LLH 43, LLH 47, LLH 58, LLH 68, LLH 74) have been used to produce each solution. Each file consists of 15 different attributes and 5000 instances. One iteration in a single solu- tion is shown in Table 1 where the bold row indicates that low-level heuristic number 74 was applied by the Peckish hyperheuristic because it has the largest improvement on objective function. In Table 1, column LLH represents the low-level tested, ESM and RSM columns stand for event and resource selection methods, respectively which specify how an event is scheduled. SE indicates the selected event number to be scheduled. UE reflects whether another event is a conflict with currently scheduled event, i.e. they share the same trainer, location, or timeslot. EID corresponds to the event identification number and R stands for rescheduled, which means, if swapping unscheduled event from the schedule with the selected event is possible, then, it is possible to reschedule back the removed event from the schedule. OP, NP, OPE and NPE correspond to old priority, new priority, old penalty and new penalty, respectively. The new priority and new penalty values represent the total pri- ority and penalty of the schedule after applying a low-level heuristic. The difference between the new priority and new penalty values gives the value of the objective function. IMP stands for amount of improvement and represents the change in value on the objective function after a low- level heuristic has been applied. CPU is the time in search for the low-level heuristic to be applied and SC reflects whether or not the current schedule has been changed, i.e. whether the current low-level heuristic had any effect at all, (‘1’ changed or ‘0’ not changed). Finally AP column indicates whether or not the current low-level heuristic has been applied (‘1’ applied or ‘0’ not applied) by the hyperheuristic. After analysing the data in each file, we identified that six attributes have some correlation to the class attribute (LLH), which are (OP, NP, OPE, NPE, IMP, AP). How- ever, we are interested to learn rules that represent useful sequence of applied low-level heuristics at different itera- tions and lead to improvement on the objective function. Therefore, solutions generated by the Peckish hyperheuris- tic have been filtered to retain certain iterations where improvements have occurred upon the objective function. Furthermore, a PL/SQL program has been designed and implemented to generate a new structure of each solution in order to enable the extraction of rules that represent the sequence of applied low-level heuristics. In other words, each training instance in the new structure should contain the low-level heuristics that improved the objective func- tion in the current iteration along with others applied in the previous iterations. Specifically, in each solution run and for each iteration, we record low-level heuristics applied in the previous three iterations along with the ones that improved the objective function in the current iteration. Table 2 represents part of a solution run (data) gener- ated in the new structure after applying the PL/SQL pro- gram on the initial solution features, where columns LLH_3, LLH_2 and LLH_1 represent the low-level heu- ristics applied at the previous three iterations. Column LLH represents the current low-level heuristic that improved the objective function and column Imp repre- sents the improvement on the objective function value. Finally column Apply represents whether or not the selected low-level heuristic has been applied by the hype- rheuristic. As shown in Table 2, data generated by the hyperheuristic have multiple labels, since there could be more than one low-level heuristic that improve the objec- tive function at any give iteration. Hence, each training instance in the scheduling data may associate with more than one class. Table 1 One iteration of the Peckish hyperheuristic for the training scheduling problem LLH ESM RSM SE UE EID R OP NP OPE NPE IMP CPU SC AP 1 0 1 1 0 �1 �1 72,999 72,999 830 830 0 0.03 0 0 2 0 2 128 0 �1 �1 72,999 72,999 830 830 0 0.02 0 0 20 4 0 12 0 �1 61 72,999 72,999 830 830 0 0.02 0 0 27 0 2 36 �1 �1 �1 72,999 72,999 830 830 0 0 0 0 37 1 2 70 �1 �1 58 72,999 72,999 830 830 0 0 0 0 43 1 8 142 �1 �1 �1 72,999 72,999 830 830 0 0 0 0 47 2 2 22 �1 �1 �1 72,999 72,999 830 830 0 0 0 0 58 3 3 18 �1 �1 �1 72,999 72,999 830 830 0 0 0 0 68 4 3 25 �1 �1 �1 72,999 72,999 830 830 0 0 0 0 74 4 9 57 �1 �1 �1 72,999 73,099 830 830 100 0 1 1 F. Thabtah, P. Cowling / Expert Systems with Applications 34 (2008) 1093–1101 1097 4.2. Experimental results In this section, we describe the experiments to evaluate classification accuracy and rules features produced by dif- ferent rule learning algorithms from the optimisation data sets. We have performed a number of experiments using ten-fold cross validation on 16 different data files derived by the Peckish hyperheuristics for the trainer scheduling problem. The sizes of the data files after applying the PL/ SQL program vary. There are 12 data files, each contains 182–750 training instances, where each file represents only one single run of the Peckish hyperheuristic. The remaining four data files contain 1500–2400 training instances and represent ten combined solutions. Two popular traditional classification algorithms (PART, C4.5) and three associa- tive classification techniques (CBA, MCAR, MMAC) have been compared in terms of accuracy. The experiments of PART and C4.5 have been conducted using Weka software system (WEKA, 2000) and CBA experiments have been performed using a VC++ version provided by the authors of CBA (1998). Finally MCAR and MMAC algorithms were implemented in Java under Windows XP on 1.7 Ghz, 256 RAM machine. The relative prediction accuracy, which corresponds to the difference of the classification accuracy of CBA, PART and C4.5 algorithms with respect to (MCAR, MMAC) are shown in Figs. 2–4, respectively. Fig. 2 for instance signifies the difference of the accuracy between CBA classifiers rela- tive to those derived by (MCAR, MMAC) on the 16 optimisation data sets. The relative prediction accuracy figures shown are computed using the formulae ðAccuracyMCAR�AccuracyCBAÞ AccuracyCBA and ðAccuracyMMAC�AccuracyCBAÞ AccuracyCBA for CBA. The same sort of formulae has been used for PART and C4.5 algorithms, respectively. We used a minsupp of 2% and a minconf of 30% in the experiments of CBA, MCAR and MMAC algorithms. The label-weight evaluation mea- sure (Thabtah et al., 2004) has been used to calculate the accuracy for MMAC algorithm in the figures. -40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00% 50.00% 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Data sets A c c u ra c y ( % ) d if fe re n c e M CA R M M A C CB A Fig. 2. Difference of accuracy between CBA and (MCAR, MMAC). -40.00% -20.00% 0.00% 20.00% 40.00% 60.00% 80.00% 100.00% 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Data sets A c c u ra c y ( % ) D if fe re n c e M CA R M M A C P A RT Fig. 3. Difference of accuracy between PART and (MCAR, MMAC). Table 2 Sample of optimisation data Iteration LLH 3 LLH 2 LLH 1 LLH Imp Apply 1 58 68 58 20 100 1 1 58 68 58 43 25 0 2 20 27 43 74 50 1 2 20 27 43 2 10 0 2 20 27 43 58 8 0 3 43 27 58 68 875 1 4 37 20 2 37 1055 1 4 37 20 2 2 950 0 4 37 20 2 74 69 0 4 37 20 2 58 9 0 1098 F. Thabtah, P. Cowling / Expert Systems with Applications 34 (2008) 1093–1101 Label-weight evaluation method assigns each class in a multi-label rule a value based on its number of occurrences with rule’s consequent (left-hand-side) in the training data. To clarify, a training object may belong to several classes where each one associated with it by a number of occur- rences in the training data set. Each class can be assigned a weight according to how many times that class has been associated with the training object. Thus, unlike the error- rate method (Witten & Frank, 2000), which considers only one class for each rule in computing the correct predictions, label-weight gives a value for each possible class in a rule according to its frequency in the training data. This gives the top ranked class in a rule the highest weight and not all the weight as error-rate method does. The accuracy results shown in the graphs indicated that associative classification algorithms outperformed the other learning techniques over the majority of test instances. Particularly, CBA, MCAR and MMAC outperformed the other learning algorithms on 5, 6, 7 and 5 benchmark problems, respectively. The won-loss-tied record of MCAR against CBA, C4.5 and PART are 11-5-0, 10-6-0 and 11-5-0, respectively. The MMAC won-loss-tied records against CBA, C4.5 and PART are, 10-6-0, 9-7-0 and 11-5-0, res- pectively. These figures show that associative classifica- tion approach is able to produce more accurate classifiers than decision trees and rule induction approaches, res- pectively. It should be noted that CBA, PART and C4.5 algorithms outperformed (MCAR, MMAC) on data set number 14 in Figs. 2–4, respectively. After analysing the data in this par- ticular set, it turned out that classes in this set are not evenly distributed. For example, class LLH2 and LLH20 were fre- quently applied by the hyperheuristic in this data set, whereas, classes LLH27, LLH37, LLH43 and LLH58 were rarely been used by the hyperheuristic. This is compound by the limited number of training instances in this particular data set (182 training data objects). Analysis of the classifiers produced revealed consistency in the accuracy of both PART and C4.5 because the differ- ence on average in the accuracy between them in all exper- iments is less than 1.6 %. This supports research works conducted in Frank and Witten (1998), where they show that despite the simplicity of PART, it generates rules as accurately as C4.5 and RIPPER. Also C4.5 and PART algorithms showed consistency in the number of rules produced. Analysis of the rules features generated from the hyper- huristic data has been carried out. Fig. 5 shows the number of rules extracted from nine data sets, categorised by the number of classes. MMAC is able to extract rules that are associated with up to four classes for this data. This is one of the principle reasons for improving the accuracy within applications. Fig. 5 also demonstrates that the majority of rules created from each solution are associated with one or two class labels. It turns out that this reflects accurately the nature of the hyperheuristic data, since dur- ing each iteration, normally only one or two low-level heu- ristics improve on the objective function in the scheduling problem. Thus, each training instance usually corresponds to just one or two classes. The additional classes discovered by MMAC algorithm from the real data represent useful knowledge discarded by CBA, PART and RIPPER algorithms. The fact that -60.00% -40.00% -20.00% 0.00% 20.00% 40.00% 60.00% 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Data sets A c c u ra c y ( % ) d if fe re n c e M CA R M M A C C4.5 Fig. 4. Difference of accuracy between C4.5 and (MCAR, MMAC). 0 4 8 12 16 20 24 28 32 36 40 1 2 3 4 5 6 7 8 9 Data set N u m b e r o f R u le s 1-label 2-labels 3-labels 4-labels Fig. 5. Distribution of the rules with regards to their labels. F. Thabtah, P. Cowling / Expert Systems with Applications 34 (2008) 1093–1101 1099 MMAC is able to extract rules with multiple classes enables domain experts to benefit from this additional useful infor- mation. In addition, these multi-label rules can contribute in the prediction step, possibly improving upon the classi- fication accuracy. The numbers of rules produced by C4.5, PART, CBA and MCAR algorithms are listed in Table 3. Since MMAC produces rules with multiple classes, we did not record their numbers of rules generated for fair comparison. The values in Table 3 show that C4.5 always generates more rules than PART, CBA and MCAR. This contradicts some earlier results reported in Liu et al. (1998) and Thabtah et al. (2005) on classifier sizes obtained against UCI data collec- tion (Merz & Murphy, 1996), which show that associative classification approaches like CBA and MCAR normally generate more rules than decision trees. For this reason, we performed extensive analysis on the classifiers derived by C4.5 from the optimisation data sets. After analysing the decision trees constructed by C4.5 algorithm at each iteration, we observed that many rules are generated, which do not cover any training data. We found out that the reason for these many useless rules appear to be the attribute that C4.5 splits the training instances on, if that attribute has many distinct values and only few of these values appear in the training data, then a rule for each branch will be generated, and hence only some of these branches cover training instances. The rest will represent rules that cover not even a single training instance. In other words, when the training data set consists of attributes that have several distinct values and a split occurs, the expected number of rules to be derived by C4.5 can be large. Since data sets used to derive the results in Table 3 contain four attributes where each one of them has 10 different values (low-level heuris- tics) and some of these low-level heuristics are never result in solution improvement for the hyperheuristic, this explains the large numbers of rules derived by C4.5. 5. Conclusions In this paper, we have studied data sets produced from a complex personnel scheduling problem, called the training scheduling problem. These data sets represent solutions generated by a general hybrid approach, called the Peckish hyperheuristic, which is a robust and general-purpose opti- misation heuristic that requires us to decide which of sev- eral simpler low-level heuristic techniques to apply at each step while building the schedule. Our study focused on analysing the behaviour of low-level heuristics that were selected by the hyperheuristic and improved upon the qual- ity of the current solution in order to extract useful rules. These rules can be used later to quickly predict the appro- priate low-level heuristics to call next. For this purpose, we have compared five data mining classification algorithms, (PART, C4.5, CBA, MCAR, MMAC) on 16 different solu- tions produced by Peckish hyperheuristic. The experimental tests showed a better performance for associative classification techniques (MCAR, MMAC, CBA) over decision trees (C4.5), rule induction (RIPPER) and PART algorithm with reference to the accuracy of pre- dicting the appropriate set of low-level heuristics. Since the MMAC algorithm was able to produce rules with multiple classes, including very useful information that the hype- rheuristic can use in forecasting the behaviour of low-level heuristic while constructing a new solution. It is the most applicable data mining approach, which can be used to pre- dict low-level heuristic performance within the Peckish hyperheuristic. Furthermore, C4.5 generated more rules than the other rule learning algorithms since useless rules were extracted by the C4.5 algorithm, which have not even a single repre- sentation in the training data. The reason of these useless rules turn out to be the training data attributes, in which when some of these attributes are associated with many distinct values and only a subset of these values have suf- ficient representation in the training data, there will be valid rules for this subset and the rest will represent useless rules. References Agrawal, R., & Srikant, R. (1994). Fast algorithms for mining association rule. In Proceedings of the 20th international conference on very large data bases (pp. 487–499). Aicklen, U., & Dowsland, K. (2000). Exploiting problem structure in a genetic approach to a nurse rostering problem. Journal of Scheduling, 3, 139–153. Ali, K., Manganaris, S., & Srikant, R. (1997). Partial classification using association rules. In D. Heckerman, H. Mannila, D. Pregibon, & R. Uthurusamy (Eds.), Proceedings of the third international conference on knowledge discovery and data mining (pp. 115–118). Blum, C., & Roli, A. (2003). Metaheuristics in combinatorial optimisa- tion: overview and conceptual comparison. ACM Computing Surveys (CSUR) (pp. 268–308). CBA (1998). <http://www.comp.nus.edu.sg/~dm2/p_download.html>. Cendrowska, J. (1987). PRISM: an algorithm for inducing modular rules. International Journal of Man-Machine Studies, 27(4), 349–370. Table 3 Number of rules of C4.5, PART, CBA and MCAR on the optimisation data sets Data C4.5 PART CBA MCAR 1 11 3 3 14 2 55 16 18 32 3 46 19 12 28 4 82 17 11 33 5 46 17 10 31 6 19 16 23 31 7 91 69 1 20 8 100 51 3 9 9 163 71 2 12 10 163 145 1 21 11 46 21 10 22 12 64 19 22 37 13 46 19 6 33 14 64 24 23 42 15 55 23 16 31 16 64 56 5 18 1100 F. Thabtah, P. Cowling / Expert Systems with Applications 34 (2008) 1093–1101 Cohen, W. (1995). Fast effective rule induction. In Proceedings of the 12th international conference on machine learning (pp. 115–123). CA: Morgan Kaufmann. Cowling, P., & Chakhlevitch, K. (2003). Hyperheuristics for managing a large collection of low level heuristics to schedule personnel. In Proceedings of IEEE conference on evolutionary computation (pp. 1214– 1221). Canberra, Australia. Cowling, P., Kendall, G., & Han, L. (2002) An investigation of a hyperheuristic genetic algorithm applied to a trainer scheduling problem. In Proceedings of congress on evolutionary computation (CEC 2002) (pp. 1185–1190). Hilton Hawaiian Village Hotel, Hono- lulu, Hawaii, May 12–17, ISBN 0-7803-7282-4. Cowling, P., Kendall, G., & Soubeiga, E. (2000). A hyperheuristic approach to scheduling a sales summit. In Proceedings of the third international conference of practice and theory of automated timetabling (PATAT 2000). LNCS (vol. 2079, pp. 176–190). Springer. Fayyad, U., & Irani, K. (1993). Multi—interval discretisation of contin- ues-valued attributes for classification learning. In Proceedings ofI JCAI (pp. 1022–1027). Frank, E., & Witten, I. (1998). Generating accurate rule sets without global optimisation. In Proceedings of the fifteenth international conference on machine learning (pp. 144–151). Madison, Wisconsin: Morgan Kaufmann. Furnkranz, J., & Widmer, G. (1994). Incremental reduced error pruning. In Machine learning: Proceedings of the 11th annual conference. New Brunswick, New Jersey: Morgan Kaufmann. Hamiez, J., & Hao, J. (2001). Solving the sports league scheduling problem with Tabu search. Lecture notes in artificial intelligence (vol. 2148, pp. 24–36). Springer. Hansen, P., & Mladenovic, N. (1997). Variable neighbourhood search. Computer and Operations Research, 1097–1100. Li, W., Han, J., & Pei, J. (2001). CMAR: accurate and efficient classification based on multiple-class association rule. In Proceedings of the ICDM’01 (pp. 369–376). CA: San Jose. Liu, B., Hsu, W., & Ma, Y. (1998). Integrating classification and association rule mining. In Proceedings of the KDD (pp. 80–86). New York, NY. McAloon, K., TretKoff, C., & Wetzel, G., 1997. Sports league scheduling. In Proceedings of the third ILOG optimisation suite international users conference. Paris, France, 1997. Merz, C., & Murphy, P. (1996). UCI repository of machine learning databases. Irvine, CA: University of California, Department of Information and Computer Science. Quinlan, J. (1979). Discovering rules from large collections of examples: a case study. In D. Michie (Ed.), Expert systems in the micro-electronic age (pp. 168–201). Edinburgh: Edinburgh University Press. Quinlan, J. (1987). Generating production rules from decision trees. In Proceedings of the 10th international joint conferences on artificial intelligence (pp. 304–307). CA: Morgan Kaufmann. Quinlan, J. (1993). C4.5: Programs for machine learning. San Mateo, CA: Morgan Kaufman. Thabtah, F., Cowling, P., & Peng, Y. (2004). MMAC: a new multi-class, multi-label associative classification approach. In Proceedings of the fourth IEEE international conference on data mining (ICDM ’04) (pp. 217–224) Brighton, UK. (Nominated for the Best paper award). Thabtah, F., Cowling, P., & Peng, Y. (2005). MCAR: multi-class classification based on association rule approach. In Proceeding of the 3rd IEEE international conference on computer systems and applications (pp. 1–7). Cairo, Egypt. WEKA (2000): Data Mining Software in Java: <http://www.cs.wai- kato.ac.nz/ml/weka>. Witten, I., & Frank, E. (2000). Data mining: practical machine learning tools and techniques association rules. In Proceedings of the 3rd KDD conference (pp. 283–286). Yin, X., & Han, J. (2003) CPAR: classification based on predictive association rule. In Proceedings of the SDM (pp. 369–376). San Francisco, CA. F. Thabtah, P. Cowling / Expert Systems with Applications 34 (2008) 1093–1101 1101 Mining the data from a hyperheuristic approach using associative classification Introduction The training scheduling problem and hyperheuristics Data mining for the selection of low-level heuristics Associative classification Classification based on association (CBA) MCAR: multi-class classification based on association rule Multi-class, multi-label associative classification (MMAC) Traditional classification approaches C4.5 Repeated incremental pruning to produce error reduction algorithm (RIPPER) PART Data and experimental results Data sets and their features Experimental results Conclusions References 
work_3hopjr7buneu3mpkb67fz4jd6a	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 221017040 Params is empty 221017040 exception Params is empty 2021/04/06-03:06:06 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221017040 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:06:06 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_3ikurlx7ifgmperterkvnzglka	----	Intelligent modeling of e-business maturity Intelligent modeling of e-business maturity George Xirogiannis a,*, Michael Glykas b a University of Piraeus, Department of Informatics, 80, Karaoli & Dimitriou St., 185 34 Piraeus, Athens, Greece b University of Aegean, Department of Financial and Management Engineering, 31, Fostini Street, 82 100 Chios, Greece Abstract E-business has a significant impact on managers and academics. Despite the rhetoric surrounding e-business strategy formulation mechanisms, which support reasoning of the effect of strategic change activities to the maturity of the e-business models, are still emerg- ing. This paper describes an attempt to build and operate such a reasoning mechanism as a novel supplement to e-business strategy for- mulation exercises. This new approach proposes the utilization of the fuzzy causal characteristics of Fuzzy Cognitive Maps (FCMs) as the underlying methodology in order to generate a hierarchical and dynamic network of interconnected maturity indicators. By using FCMs, this research aims at simulating complex strategic models with imprecise relationships while quantifying the impact of strategic changes to the overall e-business efficiency. This research establishes generic adaptive domains – maps in order to implement the inte- gration of hierarchical FCMs into e-business strategy formulation activities. Finally, this paper discusses experiments with the proposed mechanism and comments on its usability. � 2006 Elsevier Ltd. All rights reserved. Keywords: Fuzzy cognitive maps; E-business modeling; Strategy planning; Decision support 1. Introduction Today, there is an increasing demand for a strategic- level assessment of e-business capabilities that can be assembled and analyzed rapidly at low cost and without significant intrusion into the subject enterprises. The bene- fits from completing such an exercise are quite straightfor- ward, for instance, identification of significant strengths and weaknesses, establishment of a rationale for action, a reference point for measuring future progress, etc. This paper proposes a novel supplement to strategic- level maturity assessment methodologies based on fuzzy cognitive maps (FCMs). This decision aid mechanism pro- poses a new approach to supplement the current status analysis and objectives composition phases of typical e- business strategy formulation projects, by supporting ‘‘intelligent’’ modeling of e-business maturity and ‘‘intelli- gent’’ reasoning of the anticipated impact of e-business strategic change initiatives. The proposed mechanism uti- lizes the fuzzy causal characteristics of FCMs as a new modeling technique to develop a causal representation of dynamic e-business maturity domains. This research pro- poses a holistic set of adaptive domains in order to generate a hierarchical network of interconnected e-business matu- rity indicators. The proposed mechanism aims at simulat- ing the operational efficiency of complex hierarchical strategy models with imprecise relationships while quan- tifying the impact of strategic alignment to the overall e-business efficiency. Also, this paper proposes an updated FCM algorithm to model effectively the hierarchical and distributed nature of e-business maturity. This application of FCMs in modeling the maturity of e-business is considered to be novel. Moreover, it is the belief of this paper that the fuzzy reasoning capabilities enhance considerably the usefulness of the proposed mech- anism while reducing the effort to identify precise maturity measurements. The proposed model has both theoretical and practical benefits. Given the demand for effective strategic positioning of e-business initiatives, such a suc- cinct mechanism of conveying the essential dynamics of 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.01.042 * Corresponding author. E-mail addresses: georgex@unipi.gr (G. Xirogiannis), mglikas@ aegean.gr (M. Glykas). www.elsevier.com/locate/eswa Expert Systems with Applications 32 (2007) 687–702 Expert Systems with Applications mailto:georgex@unipi.gr mailto:mglikas@ aegean.gr mailto:mglikas@ aegean.gr e-business fundamental principles is believed to be useful for anyone contemplating or undertaking an e-business strategy formulation exercise. Primarily, the proposed model targets the principle beneficiaries and stakeholders of strategy formulation projects (enterprise top administra- tion, strategic decision makers, internal auditors, etc) assisting them to reason effectively about the status of e-business maturity metrics, given the (actual or hypothet- ical) implementation of a set of strategic changes. Never- theless, the explanatory nature of the mechanism can prove to be useful in a wider educational setting. This paper consists of five sections. Section 2 presents a short literature overview, Section 3 presents an overview of the FCM based system, while Section 4 discusses the new approach to e-business maturity modeling based on FCMs. Finally, Section 5 concludes this paper and briefly discuses future research activities. 2. Literature overview 2.1. E-business drivers E-business offers promise to apply web and other elec- tronic channel technologies to enable fully the integration of end-to-end processes. It involves both core and support business aspects, it focuses on information sharing effi- ciency, not just financial transactions. E-business primary objective is business improvement through: • Deployment of new technologies in the value chain. • Connection of the value chains between enterprises (B2B) and between enterprises and consumers (B2C) in order to improve service, exploit alternative dis- tribution/communication channels and support cost reduction due to the associated value chain optimi- zation. • Increase of the speed of information processing (mainly at real-time) and responsiveness by utilizing common information sources (both external and internal). E-business has a significant impact on every business function. Integrated information technology causes a shift in the value chain of the enterprise. It causes a considerable deflation of prices due to radical cost reductions, annihila- tion of profit margins, disintermediation of companies and industries due to the transparent product/service delivery to the end customer, increase in cross selling volumes and so forth. On the other hand, no industry is immune to intense competition due to chain reactions that affect all electronic network partners (Palmer, 2002). This may cause a higher level of uncertainty of future business prospects, but it is only fair to say that adaptive risk management may reduce such pitfalls. Also, the current enterprise valu- ation can be radical altered by this new business environ- ment therefore enterprises must reconsider their core competencies and strategies to maintain their competitive advantages. The new economy associated with e-business has broken down many of the traditional barriers. The fundamental shift in focus from optimizing the efficiency of individual enterprises to optimizing the efficiency of a network of enterprises for competitive advantage is a considerable challenge (Chung, Yam, Chan, & Potter, 2005). E-business activities now operate across an extended network of digi- tally connected partners to enable demand/capacity/price optimizations while offering self-service client relationships at multiple channels with a significant communication speed. It is the view of this paper that while e-business solution providers promise financial prosperity and sales volumes, case studies clearly indicate that awareness, targeted strate- gic planning and holistic organizational alignment are the key success factors for managing business in the digital age. Understanding the speed and scope of e-business impacts while generating the essential momentum forms the basis for setting realistic strategic priorities, mapping out a go-forward plan while evaluating the critical factors for e-business success. Effective service/product delivery through electronic channels requires efficient process con- trol and management of measurable targets, in order to maintain the necessary range of organizational buy-in, to manage risk and assure accountability. 2.2. Relevant research in business modeling 2.2.1. Modeling traditional business activities Enterprises usually employ modeling techniques to drive re-design activities and communicate the impact of internal change. Such modeling techniques may loosely fit within the area of decision support systems (Carlsson & Turban, 2002; Shim et al., 2002; Sprague & Watson, 1986). Several modeling approaches can be brought to bear on the task of supplementing business modeling activities. In particular the field of knowledge-based systems (Harmon & King, 1985; Metaxiotis, Psarras, & Samouilidis, 2003) could ful- fill the desire for more accurate predictive business model- ing tools. The research presented by Lin, Yang, and Pai (2002) proposed generic structures with no formal reason- ing capabilities to model traditional business processes, which could represent a business process in various con- cerns and multiple layers of abstraction. The research presented by Burgess (1998) modeled busi- ness process models with system dynamics to support the feasibility stage of business process re-engineering (BPR). Similarly, research (Burgess, 1998) modeled the interaction between competitive capabilities of quality and cost during total quality management (TQM) initiatives (Burgess, 1996). This model did not decompose hierarchical relation- ships, nor did it allow the connection of the sub-models. Finally, the model required formal definition of causal rela- tionships (e.g. functions), which posed a significant over- head in supplementing the business modeling exercise. The research presented by Crowe, Fong, Bauman, and Zayas-Castro (2002) reported the development of a tool 688 G. Xirogiannis, M. Glykas / Expert Systems with Applications 32 (2007) 687–702 https://isiarticles.com/article/3737 
work_3jqspwm2ifdpjfftgomw57gcc4	----	Identifying objects by touch: An &#x201C;expert system&#x201D; Perception & Psychophysics 1985. 37 (4). 299-302 Identifying objects by touch: An "expert system" ROBERTA L. KLATZKY University of California, Santa Barbara, California and SUSAN J. LEDERMAN and VICTORIA A. METZGER Queen's University, Kingston, Ontario, Canada How good are we at recognizing objects by touch? Intuition may suggest that the haptic system is a poor recognition device, and previous research with nonsense shapes and tangible-graphics displays supports this opinion. We argue that the recognition capabilities of touch are best as- sessed with three-dimensional, familiar objects. The present study provides a baseline measure of recognition under those circumstances, and it indicates that haptic object recognition can be both rapid and accurate. Many of you may remember a children's game in which the goal is to identify, by touch alone, several common objects plucked from a bag or a tray. The joy of playing lies in the discovery that success is possible. For those of us who have played the game and experienced that suc- cess, it may not be surprising that objects can be identi- fied through tactual exploration. However, the present report documents that this is no trivial observation. In a systematic study of haptic object identification, it was found that people are highly accurate and quite fast at iden- tifying a large number (100) of objects. We deem this observation surprising, because it has often been argued, both empirically and theoretically, that the haptic system (i.e., purposive touch, as defined in Gib- son, 1966) is inadequate for object identification, espe- cially when compared with vision. Two lines of research are used to support this claim. One compares haptic and visual identification of raised two-dimensional (and some- times fully three-dimensional) nonsense shapes (e.g., Bryant & Raz, 1975; Cashdan, 1968; Rock & Victor, 1964). The second uses tangible graphics displays-raised line drawings of objects, maps, graphs, etc.-in which raised symbols represent spatial and structural informa- tion (e.g., Lederman & Campbell, 1982; Lederman, Klatzky, & Barber, 1985; Magee & Kennedy, 1980). Such displays 'are intended to be read by hand rather than by eye. Although the research assesses the level of per- formance that can be achieved by touch alone, compari- sons with vision are nevertheless implicit. These studies demonstrate either that touch is ineffective for reading and This work was based on an honor psychology thesis conducted by the third author under the supervision of the second author at Queen's University, 1982. It was supported by Natural Sciences and Engineer- ing Research Council of Canada Grant A9854 to the second author. The first author acknowledges the support of the Center for Advanced Study in the Behavioral Sciences. Reprint requests may be sent to either the first or second author: R. Klatzky, Psychology-UCSB, Santa Barbara, CA 93106; or S. Lederman, Psychology, Queen's University, King- ston, Ontario, Canada K7L 3N6. identification or that touch is so dominated by vision that its contribution to pattern perception is minimal. However, we will suggest reasons why performance with arbitrary, unfamiliar three-dimensional or raised two- dimensional stimuli might underestimate the capacity for haptic object recognition. Further, we offer the present study as an "existence proof' that haptics can be highly effective for object identification. Our argument is simi- lar to one made by Reed, Durlach, and Braida (1982), who have been developing a sensory substitution system for the deaf. They suggest that the success of "Tadoma"-a method of understanding speech in which the "listener's" fingers and thumb(s) are placed on the cheeks, lips, and jaws of the speaker-is proof that the tactual system can process complex spatiotemporal pat- terns at rates close to those of auditory speech percep- tion. Similarly, we argue by demonstration that touch can perform a very common identification task with consider- able competence, despite its poor performance in other tasks (particularly those with less naturalistic stimuli). One reason for caution in generalizing results from studies with artificial objects or raised graphics displays to haptic performance as a whole is that such studies generally require pattern "apprehension" -obtaining in- formation about volumetric, topographical, and other at- tributes of the stimuli-as opposed to categorization. Even when artificial displays depict real objects and a categor- ical response is required, the stimuli generally fail to re- tain many of the properties of the objects themselves, such as temperature, size, or texture. The cues that these dis- plays do provide are usually dictated by the original visual master from which a raised replica was derived. It there- fore becomes necessary to determine the shape of the stimulus, perhaps even to form a visual image, in order to identify it. In contrast, real objects might be recognized on the basis of nonstructural cues. A kitchen sponge, for example, could be identified by its texture, without regard for its shape or size. As mentioned above, vision-touch comparisons have often been interpreted as evidence that haptic performance 299 Copyright 1985 Psychonomic Society, Inc. 300 KLATZKY, LEDERMAN, AND METZGER is poor. Several considerations, however, suggest that these comparisons are inappropriate to assess haptic ob- ject recognition. One concern is the degree of practice, which has been found to improve haptic discrimination performance (Gibson, 1966; Simons & Locher, 1979). A lack of familiarity with haptic identification of artifi- cial displays might be critical to its inferiority relative to vision. Another issue is whether the displays that have been used in previous research adequately allow for fun- damental differences between the visual and haptic sen- sory systems (Berla, 1982; Ikeda & Uchikawa, 1978; Lederman, 1979). For example, the resolving power of the fingertip is much less than that of the eye (Weinstein, 1968), and the tactile system may have inherent difficulty in representing stimulus orientation (Pick, Klein, & Pick, 1966). Both of these factors could influence haptic per- formance profoundly; yet stimulus construction cannot easily compensate for either. Consider that changing the size of a stimulus to accommodate the poor resolution of touch also changes the rate at which it can be explored and, accordingly, changes the temporal-integration and memory demands of the task (Berla, 1982). Similarly, us- ing explicit orientation cues may aid tactual apprehension (Lederman, 1979; Lederman & Campbell, 1982), but may also add irrelevant information that interferes with the per- ception of structural properties of the display. The foregoing suggests that in order to assess haptic object identification, one should avoid artificial objects or two-dimensional displays and, instead, use real objects. The cues that real objects provide are ecologically deter- mined, rather than based on a visual replica. Haptic manipulation of objects is commonplace and therefore familiar. Real objects maintain in full scale the attributes that contribute to haptic identification, and their proper orientation is determined by such intrinsic characteristics as principal axes, flat surfaces, and center of gravity. Thus, objects seem ideally suited to recognition through haptic exploration. Some extant studies suggest, in agreement with this reasoning, that haptic identification of common objects is quite accurate (Bigelow, 1981; Hoop, 1971; Schiff & Dytell, 1972; Simpkins, 1979). However, those investi- gations do not provide a general assessment of haptic iden- tification capabilities, because of various limitations- using very young subjects, a small sample of objects, or a task other than identification per se. The present study directly assessed adults' haptic identification of hand-size common objects that were readily identifiable through vi- sion. Its goal was to provide baseline measures of speed and accuracy. METHOD The subjects were students at Queen's University, 20-23 years of age. Three females and 2 males took part in a visual identifica- tion task, and 10 females and 10 males participated in a haptic iden- tification task. The stimuli were 100 common objects, of a size that could be held within the hands (see Appendix). None made a noise, either functionally or in the course of manual exploration, and none had a clear identifiable odor. Forty-one of the stimuli were selected from a list of objects (Snodgrass & Vanderwart, 1980) as having high (>70%) name agreement when they were presented pictorially; the remainder were selected by the experimenters as being unam- biguously identifiable by name. The objects were roughly classifi- able as personal articles, articles for entertainment, foods, cloth- ing, tools, kitchen supplies, office supplies, and household articles. There were from 8 to 23 items in each class. The visual identification task served as a pretest for verifying the nameability of the stimuli by sight. In addition to the 100 stimuli described above, an additional 10 were included, but were discarded when they were incorrectly named by a subject. A name was con- sidered correct if it was commonly applied to objects of the given type, was not commonly applied to distinctly different objects, and was not the name of a relatively abstract category. (Thus, for ex- ample, "thread" or "spool" would be acceptable for that object, but "dowel" or "sewing implement" would not.) Each object was placed, in turn, on a table before the subject, who was asked to name it. All of the stimuli ultimately used were named correctly by all five subjects. Although reaction times were not measured, they were generally brief, on the order of a second. For the haptic identification task, each subject sat at a table that had been padded with a towel to reduce noise. He or she was blind- folded and wore headphones through which white noise was deli- vered, in order to mask inadvertent noise from the exploration. The subject's chin was placed in a chinrest clamped to the table, and he or she made responses into a microphone that triggered a voice key for purposes of timing responses. The subject was free to use both hands and to pick up the objects while exploring. On each trial, the experimenter set an object on the padded table and turned on a tape recorder. (The order of objects was determined randomly for each subject.) She then tapped the subject's hand, to indicate that the object was available for exploration. When the sub- ject first touched the object, the experimenter pressed a finger switch, starting a timer that terminated with the subject's vocal response. The subject's task was to identify each object as quickly and ac- curately as possible, or, if he or she could not do so, to say, "I don't know." In addition, following vocalization of a name, the subject was asked to describe the properties that had been used to identify the object. RESULTS The principal dependent variables were reaction time (between the first tactile contact with the object and vocali- zation) and errors, either misnaming or omission ("1- don't-know" responses). Misnaming errors were further categorized as superordinate (giving the name of a higher order category, e.g., "vegetable" for a pumpkin), cate- gorically related (e.g., "sock" instead of "sweater"), corrected superordinate or related (following an initial su- perordinate or categorically related name with a correct response, e.g. "clothing ... sweater"), and categorically unrelated (e.g., "rock" for a potato). Of the 2,000 responses, only 83 (4.2 %) were errors. Four errors were omissions. There were 22 superordinate errors, 29 categorically related, 14 corrected superor- dinate or related, and 14 unrelated. Males and females did not differ in their error rates, by t test. Analysis of the reaction times for correct responses in- dicated that the model response latency was 1-2 sec and that 68 % of responses occurred within 3 sec of contact. Only 6% of responses took longer than 5 sec of contact. These data differed little by gender of subject. When mean response latencies were computed for each stimulus, all but two items had mean values of 5 sec or less. The devi- ant items were rice (mean latency = 6.6 sec) and T-shirt (mean latency = 10.6 sec). These items also accounted for relatively high proportions of the errors (6 per object). Another type of data was subjects' phenomenological reports of the object properties that had led to their iden- tifications. These were divided into 16 categories, derived from an initial sampling of the data, which distinguished among reports of a distinctive component and of the size, shape, texture, temperature, and function of the whole object or a component. Two independent raters assigned the reports to the 16 categories, with 85 % agreement. An average of 1.9 comments was made about each item. Global shape (e.g., of a whistle), global texture (e.g., of sandpaper), and presence of a distinct component (e.g., cap on a pen) were predominantly mentioned as the basis for identification (on 46%, 36%, and 35% of trials, respectively), with component texture (16%), global size (15%), and component shape (7%) next most often reported. DISCUSSION The principal finding from this study is that haptic iden- tification of a wide range of objects can be remarkably fast and accurate. Given the present scoring system, 96% of the naming responses were correct. With a more lenient system, allowing superordinate category names and related categorical responses, and permitting false starts, the accuracy rate would be 99 %. Moreover, 94 % of cor- rect (under strict scoring) names were given within a 5-sec interval. This performance is all the more remarkable when it is compared with haptic reading and identifica- tion performance observed in past research. Difficulties with raised drawings are informally confirmed in our laboratories as well; frequently, people are unable to iden- tify even simple outlines after 2-3 min of exploration! The observed levels of haptic identification are also con- siderably better than the level of identification of com- mon odors (Cain, 1979, 1982; Desor & Beauchamp, 1974; Engen & Ross, 1973)-generally estimated at about 40%-50%. An important factor underlying the relatively low level of identification by odor is label availability; thus, performance is subject to considerable improvement with practice (Cain, 1979; Desor & Beauchamp, 1974). One might argue, then, that by equating stimuli for the natural-language frequency of their labels, identification rates for touch and olfaction would be comparable. However, even after accounting for difficulties in associa- tive retrieval of names, a residual error in odor identifi- cation remains, apparently reflecting discrimination failures imposed by the sensory system itself (Cain, 1979). Odor identifications also have considerably longer laten- cies than the responses in the present experiment. In the Desor and Beauchamp study, unpracticed subjects took approximately 10 sec to give correct responses and highly HAPTIC OBJECT IDENTIFICATION 301 practiced subjects still averaged over 2.5 sec to respond. The rapid and accurate responses by unpracticed subjects in the present experiment suggest that the superiority of haptic over olfactory identification cannot be explained solely by superior accessibility to labels. Earlier it was suggested that studies with arbitrary con- figurations or two-dimensional simulations might underes- timate the capacity for haptic object identification, because these types of stimuli deprive the haptic system of some of its most effective cues. The present data provide sup- port for this suggestion. Most bases for identification that are emphasized in the phenomenological reports of these subjects-global shape, texture, and size cues-are not readily available with simulated objects. Raised-line draw- ings might appear to provide shape information, but the usual nature of that information is a projection to the reti- nal plane, not the tangible, grasped shape of haptic ex- ploration. Texture cues in raised drawings are normally minimal and usually arbitrarily assigned to the concepts they are intended to represent. Moreover, often only rela- tive size can be portrayed. Also relevant to this argument is a study (Krantz, 1972) that used factor analysis to assess the basis for haptic ob- ject identification. It identified five factors-amount of exertion needed to explore (related to compliance), rough- ness, size, temperature, and sharpness-none of which is adequately represented by raised-line drawing tech- niques in current use. Although we have argued that the haptic system is well equipped to identify familiar objects, we do not mean to claim that the perception of form through touch is gener- ally accurate and efficient. It is one thing to assign an ob- ject to a known category and quite another to apprehend its structural characteristics. Studies using artificial ob- jects and graphics displays appear well suited to assess haptic perception in this latter sense-and they assess it . as poor. In contrast, past experience with objects-visual as well as tactual-might enable categorization on the basis of minimal cues, without full apprehension of surface and shape. The present study indicates that, in this latter task, the haptic system can be very competent indeed, more competent than might be expected from intuition or ex- trapolation from experiments on haptic perception of form. We see the results of the current study as a neces- sary first (phenomenological) step towards developing a model of haptic object identification, one component of which may be a representation that is readily accessed by both haptics and vision (as suggested by Garbin & Bern- stein, 1984; Gibson, 1966). The present work also has implications for applied con- cerns. One application that motivated this research was the design of effective tangible graphics for the visually handicapped. Our study suggests that effective graphic aids should eschew simple mimicry of two-dimensional visual displays. Instead, they might incorporate the three- dimensionality of real objects and retain critical proper- ties such as texture. A rather different area in which research on haptic object identification might be applied 302 KLATZKY, LEDERMAN, AND METZGER is the development of intelligent robotic systems that use tactile sensing (Harmon, 1982). By understanding the mediators between haptic exploration and identification, we may discern levels of information representation that are appropriate for robots to simulate human behavior, thus rendering them capable of highly accurate identifi- cation performance. REFERENCES BERLA, E. P. (1982). Haptic perception of tangible graphic displays. In W. Schiff & E. Foulke (Eds.), Tactual perceptions: A sourcebook (pp. 364-386). Cambridge: Cambridge University Press. BIGELOW, A. E. (1981). Children's tactile identification of miniaturized common objects. Developmental Psychology, 17,111-114. BRYANT, P., & RAZ, I. (1975). Visual and tactual perception of shape by young children. Developmental Psychology, 11, 525-526. CAIN, W. S. (1979). To know with the nose: Keys to odor identifica- tion. Science, 203, 467-469. CAIN, W. S. (1982). Odor identification by males and females: Predic- tions vs performance. Chemical Senses, 7, 129-142. CASHDAN, S. (1968). Visual and haptic form discrimination under con- ditions of successive stimulation. Journal of Experimental Psychol- ogy Monograph, 76(2, Pt. I). DESOR, J. A., & BEAUCHAMP, G. K. (1974). The human capacity to transmit olfactory information. Perception & Psychophysics, 16, 551-556. ENGEN, L. T., & Ross, B. (1973). Long-term memory of odors with and without verbal descriptions. Journal ofExperimental Psychology, 100, 221-227. GARBIN, C. P., & BERSTEIN, I. H. (1984). Visual and haptic percep- tion of three-dimensional solid forms. Perception & Psychophysics, 36, 104-110. GIBSON, J. J. (1966). The senses considered as perceptual systems. Boston: Houghton Mifflin. HARMON, L. (1982). Automated tactile sensing. Robotics Research, 1, 3-32. Hoop, N. H. (1971). Haptic perception in preschool children. Ameri- can Journal of Occupational Therapy, 25, 340-344. IKEDA, M., & UCHIKAWA, K. (1978). Integrating time for visual pat- tern perception and a comparison with the tactile model. Vision Research, 18, 1565-1571. KRANTZ, M. (1972). Haptic recognition of objects in children. Journal of Genetic Psychology, 120,121-133. LEDERMAN, S. J. (1979). Tactual mapping from a psychologist's per- spective. Bulletin ofthe Association of Canadian Map Libraries, 32, 21-25. LEDERMAN, S., & CAMPBELL, J. (1982). Tangible graphs for the blind. Human Factors, 24, 85-100. LEDERMAN, S., KLATZKY, R. L., & BARBER, P. (1985). Spatial and movement-based heuristics for encoding pattern information through touch. Journal of Experimental Psychology: General, 114, 33-49. MAGEE, L. E., & KENNEDY, J. M. (1980). Exploring pictures tactu- ally. Nature, 283, 287-288. PICK, H. L., KLEIN, R. E., & PICK, A. D. (1966). Visual and tactual identification of form orientation. Journal ofExperimental Child Psy- chology, 4, 391-397. REED, C., DURLACH, N., & BRAIDA, L. (1982). Research on tactile com- munication of speech: A review. American Speech-Language-Hearing Association Monographs, No. 20, 1-23. ROCK, I., & VICTOR, J. (1964). Vision and touch: An experimentally created conflict between the two senses. Science, 143, 594-596. SCHIFF, W., & DYTELL, R. S. (1972). Deafand hearing children's per- formance on a tactual perception battery. Perceptual and Motor Skills, 35, 683-706. SIMONS, R. W., & LOCHER, P. J. (1979). Role of extended perceptual experience upon haptic perception of nonrepresentational shapes. Per- ceptual and Motor Skills, 48, 987-991. SIMPKINS, K. E. (1979). Tactual discrimination of household objects. Journal of Visual Impairment and Blindness, 73, 86-92. SNODGRASS, J. G., & VANDERWART, M. (1980). A standardized set of 260pictures: Norms for name agreement, image agreement. familiar- ity. and visual complexity. Journal ofExperimental Psychology: Hu- man Learning and Memory, 6, 174-215. WEINSTEIN, S. (1968). Intensive and extensive aspects of tactile sensi- tivity as a function of body pan, sex, and laterality. In D. R. Ken- shalo (Ed.), The skin senses (pp. 195-218). Springfield, II.: Thomas. APPENDIX Stimulus Objects Category Articles Personal comb, emery board, glasses, hair dryer, ring, swab, toothbrush, wallet, watch Entertainment balloon, baseball bat, baseball glove, birdie, crayons, golf ball, playing cards, record, tennis racket, whistle Clothing belt, boot, mitten, scarf, shoelace, sock, sweater, T-shirt, tie Foods carrot, cracker, egg, onion, potato, pump- kin, rice, tea bag Office Supplies binder, book, bow, calculator, clipboard, envelope, eraser, notebook, paper clip, paper pad, pen, pencil, pencil sharpener, ruler, sta- pler, tape Tools clamp, hammer, paintbrush, sandpaper, scis- sors, screw, screwdriver, twine, wrench Household ash tray, bandage, button, candle, clothespin. dust pan, electric cord, flashlight, flower pot, hook, key, light bulb, match book, padlock, rubber band, safety pin, scrub brush, sponge, thread, toilet paper, toothpicks, umbrella, watering can Kitchen Supplies baby bottle, bottle opener, bowl, fork, fun- nel, glass, kettle, knife, ladle, muffin pan, mug, plate, pot, spatula, strainer, wooden spoon (Manuscript received November 21. 1984: revision accepted for publication January 20, 1985.) 
work_3klorapqwnd2pmpmvo24rrzmta	----	Recommender system based on pairwise association rules Expert Systems With Applications 115 (2019) 535–542 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Recommender system based on pairwise association rules Timur Osadchiy a , ∗, Ivan Poliakov a , Patrick Olivier a , Maisie Rowland b , Emma Foster b a Open Lab, School of Computing, Newcastle University, Newcastle upon Tyne, United Kingdom b Institute of Health and Society, Newcastle University, Newcastle upon Tyne, United Kingdom a r t i c l e i n f o Article history: Received 4 April 2018 Revised 9 July 2018 Accepted 10 July 2018 Available online 21 August 2018 Keywords: Association rules Cold-start problem Data mining Ontologies Recommender systems a b s t r a c t Recommender systems based on methods such as collaborative and content-based filtering rely on ex- tensive user profiles and item descriptors as well as on an extensive history of user preferences. Such methods face a number of challenges; including the cold-start problem in systems characterized by ir- regular usage, privacy concerns, and contexts where the range of indicators representing user interests is limited. We describe a recommender algorithm that builds a model of collective preferences indepen- dently of personal user interests and does not require a complex system of ratings. The performance of the algorithm is analyzed on a large transactional data set generated by a real-world dietary intake recall system. © 2018 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license. ( http://creativecommons.org/licenses/by/4.0/ ) 1 a t l 2 d b u p E f s u e a c r t s U N k m F i s p p m a i u fi i c w n M i b m ( c & h 0 . Introduction Recommender systems aim to identify consumer preferences nd accurately suggest relevant items (e.g. products, services, con- ent). They are used in various application domains, including on- ine retail, tourism and entertainment ( Covington, Adams, & Sargin, 016; Linden, Smith, & York, 2003 ). Widely adopted recommen- ation techniques often utilize collaborative filtering or content- ased recommendation methods ( Pazzani & Billsus, 2007 ). Collaborative filtering produces recommendations based on ser preference models that are generated from explicit and/or im- licit characteristics and metrics corresponding to user interests. xplicit indicators normally imply users assigning ratings to items; or example, to products viewed in, or purchased from, an online tore. Examples of implicit indicators include the amount of time sers spend interacting with content (e.g. watching a video) or lev- ls of interaction (e.g. scroll offset of a web page containing an rticle they are reading). Items that are positively rated or pur- hased by consumers with similar preference models are used as ecommendations for target users. User similarity can be expressed hrough correlations in purchasing history or ratings given to the ame products, which can be further amplified with demograph- ∗ Corresponding author at: Open Lab, School of Computing, Newcastle University, rban Sciences Building, 1 Science Square, Science Central, Newcastle upon Tyne, E4 5TG, United Kingdom. E-mail addresses: t.osadchiy@newcastle.ac.uk (T. Osadchiy), ivan.polia ov@newcastle.ac.uk (I. Poliakov), patrick.olivier@newcastle.ac.uk (P. Olivier), aisie.rowland@newcastle.ac.uk (M. Rowland), emma.foster@newcastle.ac.uk (E. oster). u u c m o a t ttps://doi.org/10.1016/j.eswa.2018.07.077 957-4174/© 2018 The Authors. Published by Elsevier Ltd. This is an open access article u cs (e.g. age, gender, occupation). Content-based filtering identifies imilarities between items based on a set of their descriptors (e.g. urpose of an item, author, artist, keywords). Items similar to those ositively rated or purchased by the target user, are used as recom- endations. Collaborative and content-based filtering recommender systems re therefore heavily dependent on extensive user- or item- profile nformation and are most effective when there is a rich history of ser preferences or behavior. Sparse data sets and lean user pro- les typically result in low-quality recommendations or an inabil- ty to produce recommendations at all. This is referred to as the old-start problem, where new users are added into the system ith empty behavior profiles or new items are added that have ot been reviewed or rated by anyone ( Shaw, Xu, & Geva, 2010 ). any solutions to the cold-start problem have been considered, ncluding hybrid methods that combine collaborative and content- ased filtering ( Schein, Popescul, Ungar, & Pennock, 2002 ), and ethods that aim to predict user preferences from demographics Lika, Kolomvatsos, & Hadjiefthymiades, 2014 ), or knowledge of so- ial relationships ( Carrer-Neto, Hernández-Alcaraz, Valencia-García, García-Sánchez, 2012 ). There are, however, a number of application contexts where sers interact anonymously; for example, online shops where an nregistered user browses and adds products to their basket to heck out later. In other contexts, such as email recipient recom- endation ( Roth et al., 2010 ), applications requiring high levels f privacy, or those where individual interactions with a system re necessarily infrequent (e.g. dietary surveys Bradley et al., 2016 ) here are no features and ratings to exploit, and the construction nder the CC BY license. ( http://creativecommons.org/licenses/by/4.0/ ) https://doi.org/10.1016/j.eswa.2018.07.077 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2018.07.077&domain=pdf http://creativecommons.org/licenses/by/4.0/ mailto:t.osadchiy@newcastle.ac.uk mailto:ivan.poliakov@newcastle.ac.uk mailto:patrick.olivier@newcastle.ac.uk mailto:maisie.rowland@newcastle.ac.uk mailto:emma.foster@newcastle.ac.uk https://doi.org/10.1016/j.eswa.2018.07.077 http://creativecommons.org/licenses/by/4.0/ 536 T. Osadchiy et al. / Expert Systems With Applications 115 (2019) 535–542 o g g c l t t I a t a m fi a b t t u t D c fi l A B C o w t t a t q p e o 3 3 v d b p t w s i d i d D l m t 4 c t d f t of personal behavior and preference models is not possible. To ad- dress these challenges we describe a recommender algorithm that is independent of any personal user model and does not require a complex system of ratings. Based on a set of observed items se- lected by a user, the algorithm produces a set of items ranked by confidence of their being observed next. In designing the under- lying algorithm, we review existing methods that aim to address similar tasks, adapt them to meet the constraints of the application context that is our primary concern (dietary surveys), and propose a novel alternative. The performance of three methods is compared through the task of recommending omitted foods in a real world dietary recall system. 2. Related work While various approaches have been proposed to address the cold-start problem in recommender systems, the majority of these rely on knowledge of content ( Popescul, Pennock, & Lawrence, 2001 ) and users ( Lika et al., 2014 ), including social relationships between users ( Carrer-Neto et al., 2012 ), whereas our concern is with contexts where such information is not available. Shaw et al. addressed the cold-start problem in recommender systems by us- ing a data analysis technique, which is applied to large data sets for discovering items that frequently appear together in a single transaction. This technique is known as association rules ( Agrawal, Imielinski, & Swami, 1993; Shaw et al., 2010 ). Each association rule normally consists of a set of antecedent items that lead to a con- sequent item with a certain confidence. Pazzani and Billsus con- sider the list of topics of books users voted for as transactions, which allowed them to extract association rules for topics that fre- quently appeared together as part of a user’s interests ( Pazzani & Billsus, 2007 ). To expand preferences for each user, the algo- rithm then generates all possible combinations of topics for every book the user voted for and filters association rules, where the an- tecedent part of the rule matches one of the combinations. The consequent list of topics is added to the preferences of that user. In combination with a domain ontology association rules can be effectively employed for extracting, understanding and for- malizing new knowledge ( Ruiz, Foguem, & Grabot, 2014; Sene, Kamsu-Foguem, & Rumeau, 2018 ). However, association rules have to be adapted for recommendation tasks since they are primar- ily designed to be used as exploratory tools ( Rudin, Letham, Salleb-Aouissi, Kogan, & Madigan, 2011 ) to discover previously un- known relations that need to be analyzed for their interesting- ness ( Atkinson, Figueroa, & Pérez, 2013 ). As we will probably want to provide more specificity, and recommend the exact titles of the books instead of generic categories, this potentially leads to a vast number of mined association rules and matching all pos- sible combinations of the observed items may not result in rules being found. Furthermore, a consequent item may appear in mul- tiple matching rules, meaning that a function must be introduced that aggregates the confidences of found rules into a single score for the consequent item. Finally, only the associations with a sup- port (i.e. how often a rule holds as true across the data set) higher than a defined threshold are normally extracted ( Li & Deng, 2007 ). The produced list of rules is supposed to be of a reasonable size, to allow manual examination. In a recommender system, even as- sociations with low frequencies could still be relevant, if other rel- evant rules with higher confidence are not found. This requires the extraction of as many rules as possible, making the mining process a computationally expensive task ( Zheng, Kohavi, & Mason, 2001 ). Roth et al. (2010) introduced a method for building implicit so- cial graphs based on histories of interaction between users and es- timations of their affinity and applied it to the problem of email recipient recommendation. Based on a set of email addresses se- lected by a user (the seed group) the algorithm extracts all groups f contacts with whom the user has ever exchanged emails. Here a roup of contacts is a set of email contacts that were observed to- ether in a recipient list. For each of the contacts in each group, ex- luding members of the seed group, an interaction score is calcu- ated based on the volume of messages exchanged with that group, he recency of those messages, and the number of intersections of he seed group with the group of contacts that is being considered. nteraction scores of contacts that are present in multiple groups re aggregated into a single interaction score, which is then used o rank the set of email recommendations. The implicit social graph is a promising alternative to mining ssociation rules. It instead measures the confidence of recom- ended items based solely on observed transactions that are pre- ltered by intersections with given items. However, the method lso estimates the relevance of a group of recommended emails ased on the strength of social interaction of the target user with hat group, which is not a meaningful metric for applications hat do not assume communication (or other interaction) between sers. Association rules and implicit social graphs are data en- ities that represent item-to-group relationships. However, uMouchel and Pregibon (2001) suggested that a more effi- ient approach to discovering “interesting” associations is to first nd pairs of items that frequently appear together and then ana- yze larger sized item sets that contain those pairs. For example, if BC appears in a data set with a certain frequency, then pairs AB, C and AC would be at least as frequent as the triplet. Raeder and hawla (2011) effectively analyzed associations through a graph f individual items connected to each other with edges that are eighted by the frequency of two items appearing together. Items hat have stronger relationships with each other are compared o other items form clusters, which are then targeted for further nalysis. Similar to the implicit social graph this method avoids he need to mine all possible association rules, but without re- uiring any additional indicators of relevance except for the item airs frequencies. The relevance of produced recommendations is ffectively inf erred from the likelihood of their appearing with the bserved items. . Associated food recommender algorithm .1. Intake24 We introduce a new recommendation algorithm that was de- eloped for Intake24, a system for conducting 24-h multiple-pass ietary recall surveys ( Bradley et al., 2016 ). Intake24 is designed to e a cost-effective alternative to interviewer-led 24-h recalls and rovides respondents with a web-based interface through which hey enter their dietary intake for the previous day. Respondents ill likely only ever use the system if they are a part of a dietary tudy and only for a short period of time. Within a survey, a respondent typically records their dietary ntake for the previous day on three separate occasions. A single ay normally consists of four to seven meals (e.g. breakfast, morn- ng snack, lunch etc.) which include a selection of foods, drinks, esserts, condiments, and such (referred to generically as foods). uring the first step of a recall session, a respondent reports a ist of names of foods consumed during each of the previous day’s eals in a free-text format. For each text entry, the system re- urns a list of relevant foods selected from a taxonomy of around 800 foods, organized in a tree-based, multi-level structure. Spe- ific foods are terminal nodes of this taxonomy and are linked to heir nutrient values and portion size estimation methods. Respon- ents select one food from the returned list to add to their meal; or example: Coca-Cola (not diet); Beef, stir fried (meat only); Toma- oes, tinned; Basmati rice; Onions, fried; Chilli powder; Kidney beans . T. Osadchiy et al. / Expert Systems With Applications 115 (2019) 535–542 537 t 4 a o I t m s o d a t c i w t d m r s t n f t d p a p 3 a s o a r o a a o e t p t s s s o o 3 r s E q w s i r m t e l h d I n r s d d s t p 3 R s c p f t i s s i p m t ( ( i s t b i c d t c Completing an accurate recall requires respondents to be able o identify foods they ate from a database that covers more than 800 foods; for example, there are more than 30 types of bread lone. Thus, one of the key features in determining the usability f a dietary recall system is its presentation of food search results. f respondents are not able to readily identify items from a list re- urned in response to their textual description of the food, they are ore likely to select foods perceived as the closest match or even kip reporting the intake of that food. In other words, the relevance f search results, in terms of prioritizing them appropriately, may irectly affect the accuracy of dietary recall through level of effort nd time required to select the correct foods and report intake. The main application for the recommender algorithm is to au- omate the extraction of questions about foods that are commonly onsumed together (associated foods). In Intake24, this feature is mplemented as a link between an antecedent food (e.g. toast, hite bread ) and the consequent associated food category (e.g. but- er/margarine/oils ) along with a question that is asked if a respon- ent selects the antecedent food (e.g. Did you have any butter or argarine on your toast? ). Such food associations prompts are cur- ently hand-crafted by trained nutritional experts, which for thou- ands of foods is inevitably a time consuming process that is prone o omissions. Eating habits depend on region, culture, diet, and a umber of other factors, which requires defining new associations or every context in which a system is deployed. Furthermore, over ime new foods and recipes emerge and dietary trends change. In- eed, existing rules are often curated based only on personal ex- erience or previous research, and no published study has evalu- ted their appropriateness or explored alternative data driven ap- roaches to extracting such associations. .2. Generic procedure Our approach assumes that the patterns in eating behaviors of n observed population; that is, the respondents who took part in urveys conducted in a given country, has some relevance to those f an individual in that population. The recommender algorithm ssumes no prior knowledge about an individual except their cur- ently selected food items. The algorithm is trained on a large set f observed meals and produces a model of the eating behavior of given population, where a meal is a group of uniquely identifi- ble foods (e.g. vanilla ice cream, pear juice) reported to be eaten n a single occasion. Each individual food can be recorded as being aten only once during a meal. During the recommendation step, he resulting model accepts a set of foods, which we refer to as in- ut foods IF , and returns a set of recommended foods RF mapped o likelihoods of being reported along with IF (recommendation cores). IF are excluded from recommendations. In the following ection we discuss three possible implementations that were con- idered for the recommender algorithm. Along with the description f our methods we include examples of generated models and rec- mmendations for a sample transaction data set. .3. Association rules We introduce a recommender algorithm based on association ules (AR) that generates a model of eating behavior from a data et of meals (in the training step) in the form of association rules. ach rule consists of a set of antecedent foods and a single conse- uent food, together with the confidence that the consequent food ill be present in a meal given the antecedent foods that were ob- erved. The procedure for retrieving association rules is described n Agrawal et al. (1993) . The AR algorithm makes predictions from stored association ules with antecedent part antc similar to IF , and produces recom- endations from the consequent parts of the rules. To do so, AR akes association rules that have a consequent food that is differ- nt from any of IF and the antecedent foods antc that include at east one of IF ( Algorithm 1 ). The algorithm calculates the likeli- Algorithm 1: Recommendations based on association rules. function Recommend input : AM, association rules based model IF , foods selected by a respondent returns : RF , list of food recommendations 1 RF ← ∅ ; 2 foreach rule r l ∈ AM & r l.consequent / ∈ IF : 3 f ← rl.consequent; 4 if ∃ a f : a f ∈ rl.antecedent & a f ∈ IF : 5 if f / ∈ RF : 6 RF [ f ] ← 0 7 ant c ← rl.antecedent ; 8 c ← rl.con f idence ; 9 intr ← size ({ a f : a f ∈ antc & a f ∈ IF } ) ; 10 ms ← intr 2 / (size (antc) ∗ size (IF )) ; 11 RF [ f ] ← RF [ f ] + c ∗ ms 12 return RF ood of a recommended food f to be selected next as the confi- ence of the rule c multiplied by the similarity between antc and F (i.e. match score ms ). The match score ms is calculated as the umber of foods that appear both in IF and antc (i.e. intersections) aised to the power of two and divided by the size of IF and the ize of antc . We introduce the match score so that the recommen- ations from the rules with antc that are more similar to IF pro- uce recommendations that will appear higher. We then sum the cores for every f as its single recommendation score RF [ f ]. Recommendations produced by AR applied to the example ransaction data set { abcd, ade, de, ab } and given items { ab } are rovided in Table 1 . .4. Transactional item confidence We adapted the implicit social graph method described in oth et al. (2010) for our food recommendation task, which re- ulted in a recommender algorithm based on transactional item onfidence (TIC). One key difference to our food recommendation roblem is that the original email recipient recommendation task or which the implicit social graph was developed assumed two ypes of relationships between items in a data set (outgoing and ncoming emails). Our data set assumes only one type of relation- hip, which is the co-occurrence of foods in a meal. For that rea- on, TIC produces recommendations based on similarity of histor- cally observed transactions to IF and the frequency of foods ap- earing in those transactions. During the training step, TIC converts all reported meals to a ap of unique meals (or transactions) TM , so that there are no wo transactions of the same length containing the same foods Algorithm 2 ). For every food f in a transaction m , the confidence conditional probability) is calculated as TM [ m, f ] of f being present n m given that the rest of the foods from m were observed. To do o, the algorithm counts the number cf of reported meals that con- ain all the foods from m , excluding f , and divides it by the num- er cm of reported meals containing all of the foods from m . This s similar to the confidence measured in AR, but in this case we alculate it only for the full-sized meals that were observed in the ata set M , and not for all possible combinations of foods within hose meals. At the recommend step, the algorithm retrieves all transactions ontaining any of the input foods IF ( Algorithm 3 ). Within each of 538 T. Osadchiy et al. / Expert Systems With Applications 115 (2019) 535–542 Table 1 Recommender algorithm based on association rules applied to the example data set. Model based on AR Filtered rules Recommendations 1. a ⇒ b 0.67, d 0.67, c 0.33, e 0.33 1. a ⇒ d 0.67, c 0.33, e 0.33 d: 2.50 2. b ⇒ a 1.00, c 0.50, d 0.50 2. b ⇒ c 0.50, d 0.50 c: 1.96 3. c ⇒ a 1.00, b 1.00, d 1.00 6. a, b ⇒ c 0.50, d 0.50 e: 0.29 4. d ⇒ a 0.67, e 0.67, b 0.33, c 0.33 7. a , c ⇒ d 1.00 5. e ⇒ d 1.00, a 0.50 8. a , d ⇒ c 0.50, e 0.50 6. a, b ⇒ c 0.50, d 0.50 9. a , e ⇒ d 1.00 7. a, c ⇒ b 1.00, d 1.00 10. b , c ⇒ d 1.00 8. a, d ⇒ b 0.50, c 0.50, e 0.50 11. b , d ⇒ c 1.00 9. a, e ⇒ d 1.00 14. a, b , c ⇒ d 1.00 10. b, c ⇒ a 1.00, d 1.00 15. a, b , d ⇒ c 1.00 11. b, d ⇒ a 1.00, c 1.00 12. c, d ⇒ a 1.00, b 1.00 13. d, e ⇒ a 0.50 14. a, b, c ⇒ d 1.00 15. a, b, d ⇒ c 1.00 16. a, c, d ⇒ b 1.00 17. b, c, d ⇒ a 1.00 Algorithm 2: Training the model based on transactional item confidence. function Train input : M, data set of all meals returns : T M, map of unique meals with confidence for every food 1 T M ← ∅ ; 2 foreach meal m ∈ M : 3 if m / ∈ T M : 4 T M[ m ] ← ∅ 5 cm ← size ({ m 1 : m 1 ∈ M & m ∈ m 1 } ) ; 6 foreach food f ∈ m : 7 m 2 ← { f 1 : f 1 ∈ m & f 1 � = f } ; 8 c f ← size ({ m 3 : m 3 ∈ M & m 2 ∈ m 3 } ) ; 9 T M[ m, f ] ← c f /cm 10 return TM Algorithm 3: Recommendations based on transactional item confidence. function Recommend input : T M, map of unique meals with confidence for every food IF , foods selected by a respondent returns : RF , list of food recommendations 1 RF ← ∅ ; 2 foreach meal m ∈ T M : 3 if ∃ f 1 : f 1 ∈ m & f 1 ∈ IF : 4 foreach food f ∈ m & f / ∈ IF : 5 if f / ∈ RF : 6 RF [ f ] ← 0 7 con f ← m [ f ] ; 8 inter ← size ({ f 2 : f 2 ∈ m & f 2 ∈ IF } ) ; 9 RF [ f ] ← RF [ f ] + inter ∗ con f 10 return RF t v 3 d o a w P t a t C t a c t t c i the retrieved transactions m , foods f that are not included in IF are mapped to a score that is calculated as the number of intersections of m with IF (i.e. similarity) multiplied by the food’s confidence TM [ m, f ]. Multiple scores for f measured from different transactions are summed into a final recommendation score RF [ f ]. Recommendations produced by the TIC applied to an example ransaction data set { abcd, ade, de, ab } given items { ab } are pro- ided below ( Table 2 ). .5. Pairwise association rules Unlike the previous two algorithms, which produce recommen- ations from association rules and transactions similar to currently bserved IF , the recommender algorithm based on pairwise associ- tion rules (PAR) recommends foods that are likely to be observed ith any of IF in pairs. During the training stage ( Algorithm 4 ), AR for every observed food f counts the number OD [ f ] of meals Algorithm 4: Training the model based on pairwise associa- tion rules. function Train input : M, data set of all meals returns : P M, pairwise association rules 1 OD ← ∅ , food occurrences; 2 CD ← ∅ , food co-occurrences; 3 foreach meal m ∈ M : 4 foreach food f ∈ m : 5 if f / ∈ OD & f / ∈ CD : 6 OD [ f ] ← 0 ; 7 CD [ f ] ← ∅ ; 8 OD [ f ] ← OD [ f ] + 1 ; 9 foreach food f 1 ∈ m & f 1 � = f : 10 if f 1 / ∈ CD [ f ] : 11 CD [ f, f 1] ← 0 12 CD [ f, f 1] ← CD [ f, f 1] + 1 13 P M ← [ OD, CD ] ; 14 return PM hat contain that food. For every observed pair of foods { f, f 1}, it lso counts the number CD [ f, f 1] of reported meals that contain hat pair. At the recommend step ( Algorithm 5 ), PAR retrieves pairs D [ inf ], where one food inf is observed in IF . For every pair { inf, f }, he algorithm calculates the conditional probability p , of f being in meal, given that inf was observed as the number of meals that ontain that pair CD [ inf, f ], divided by the number of meals OD [ inf ] hat contain only inf . For example, if item A appeared 10 times in he data set and co-occurred with item B only 2 times, then the onditional probability that item B will occur the next time the A s present is 0.2. Multiple probabilities retrieved for f from differ- T. Osadchiy et al. / Expert Systems With Applications 115 (2019) 535–542 539 Table 2 Recommender algorithm based on transactional confidence applied to the example data set. Model based on TIC Filtered rules Recommendations 1. a 1.00, b 1.00, c 1.00, d 1.00 1. a 1.0, b 1.00, c 1.00, d 1.00 d: 3.00 2. d 1.00, a 0.50, e 0.50 2. d 1.00, a 0.50, e 0.50 c: 2.00 3. d 1.00, e 0.67 e: 0.50 4. a 1.00, b 0.67 Algorithm 5: Recommendations based on pairwise association rules. function Recommend input : P M, pairwise association rules IF , foods selected by a respondent returns : RF , list of food recommendations 1 RF ← ∅ ; 2 P ← ∅ , conditional probabilities of foods; 3 W ← ∅ , conditional probability weights; 4 OD ← P M[ OD ] , food occurrences ; 5 CD ← P M[ CD ] , food co-occurrences ; 6 foreach input food in f ∈ IF : 7 foreach food f ∈ CD [ in f ] & f / ∈ IF : 8 if f / ∈ P & f / ∈ W : 9 P [ f ] ← ∅ ; 10 W [ f ] ← ∅ 11 p ← CD [ in f, f ] / OD [ in f ] ; 12 P [ f ] ← P [ f ] + { p} ; 13 W [ f ] ← W [ f ] + { OD [ in f ] } 14 foreach food f ∈ P : 15 RF [ f ] ← sum (P [ f ]) ∗ sum (W [ f ]) 16 return RF e P i d t m g s a f c p R a 1 C w s t v 4 p p W f i s T R p d m a c o r r t W e p w e a l ( D 1 t a ( a s m fi t ( e b A c a n i Z p i a s fi o t ( 5 5 d nt associations are summed into its single recommendation score [ f ]. As demonstrated in Roth et al. (2010) the number of times tems are observed together is an important relevance metric. In- eed, if we simply aggregate the probabilities derived from mul- iple associations, we lose information as to whether a recom- ended food has ever been observed with all IF . For example, iven two input items C and D , the aggregation may produce two cores R CD (A ) = 0 . 5 , where item A appeared with both C and D , nd R C (B ) = 0 . 5 , where item B appeared only with item C . There- ore A should receive a higher score. Likewise we take into ac- ount the frequency of an input food inf that matched a retrieved air. For example, we may have two equal scores, R C (A ) = 0 . 5 and D (B ) = 0 . 5 , where A and B historically appeared only with items C nd D respectively; but C appeared 10 times and item D appeared 00 times, which implies that the recommendation produced by should have a higher score. For these reasons, the algorithm eights aggregated probabilities P [ f ] by multiplying them by the ummed frequency of inf . Recommendations produced by PAR applied to the example ransaction data set { abcd, ade, de, ab } given items { ab } are pro- ided in Table 3 . . Methodology We compare the three algorithms for 20 0 0 0 randomly sam- led meals, each containing no fewer than two foods, reported by articipants of various ages in the UK between 2014 and 2018. e also randomize the order of foods in each meal. We use k- old ( k = 10 ) cross validation to segment the data set into train- ng and testing sets ( Salzberg, 1997 ). On each step we use nine ubsets for training a model, leaving out one subset for testing. he testing procedure is similar to the procedure described in oth et al. (2010) : we sample a few foods from each meal (in- ut foods), leaving the rest (at least one food) to simulate respon- ents’ omitted foods to be guessed by the algorithm. Every trained odel makes predictions, starting from an input size of one food nd gradually incrementing it to five. In the course of the evaluation, we plot the precision-recall (PR) urves for every algorithm on every increment. For the purposes f the evaluation, we measure the recall as the percentage of cor- ect predictions out of the total number of foods selected by the espondent, and the precision as the percentage of correct predic- ions out of the total number of predictions made by the algorithm. e count predictions that were present in the set of foods actually ntered by the respondent (excluding input foods) as correct (true ositives). We analyze the quality of the top 15 recommendations, hich is a slightly larger size than viewed by most users ( Burges t al., 2005; Van Deursen & Van Dijk, 2009 ). As the measure of lgorithm ranking quality for every size of input foods we calcu- ate the mean value of Normalized Discounted Cumulative Gain nDCG) at rank 15 ( Burges et al., 2005 ) as nDC G 15 = DC G 15 /IDC G 15 . iscounted cumulative gain is measured as DCG 15 = ∑ 15 i =1 (2 r(i ) − ) / log (i + 1) , where r ( i ) is the relevance score of the i th food. As he relevance score, we use 0 for a wrong prediction and 1 for correct prediction. Thus, the Ideal Discounted Cumulative Gain IDCG) in our case is always 1, which is a single correct prediction s the first result. We then select the implementation that demon- trates the highest performance and apply it to the task of recom- ending foods omitted by respondents with a lower level of speci- city, and for ranking search results returned in response to their ext queries. For the implementation of AR we use the FP-growth algorithm frequent patterns algorithm) ( Li & Deng, 2007 ). FP-growth is an fficient and scalable association rules mining algorithm that is ased on building frequent-pattern tree structure. In contrast to priori-like algorithms that serve the same purpose, the FP-growth ompresses a large database into a much smaller data structure voiding costly repeated database scans and generation of a large umber of candidate sets. We use a parallel version of FP-growth mplemented in the Apache Spark framework ( Li, Wang, Zhang, hang, & Chang, 2008; Meng et al., 2016 ). As a parameter, this im- lementation accepts the minimum support for an item set to be dentified as frequent and the minimum confidence for the gener- ted association rules. To gather as many association rules as pos- ible we set both the minimum support and the minimum con- dence to the lowest value (3 × 10 4 ) that allows the completion f the mining process of our data set on our machine within a ime limit of 5 minutes. The evaluation is conducted on Mac Pro 2.9 GHz Intel Core i5, 16 GB). . Results .1. General performance As can be observed from the PR curves ( Figs. 1 and 2 ), PAR pro- uces the largest area under the curve, which increases with the 540 T. Osadchiy et al. / Expert Systems With Applications 115 (2019) 535–542 Table 3 Recommender algorithm based on pairwise association rules applied to the example data set. Model based on PAR Filtered rules Recommendations 1. a 3.0 ⇒ b 2.0, d 2.0, c 1.0, e 1.0 1. a 3.0 ⇒ d 2.0, c 1.0, e 1.0 d: 5.8 2. b 2.0 ⇒ a 2.0, c 1.0, d 1.0 2. b 2.0 ⇒ c 1.0, d 1.0 c: 4.2 3. c 1.0 ⇒ a 1.0, b 1.0, d 1.0 e: 1 4. d 3.0 ⇒ a 2.0, e 2.0, b 1.0, c 1.0 5. e 2.0 ⇒ d 2.0, a 1.0 Fig. 1. Precision-Recall curves for an input size of 2 foods. Fig. 2. Precision-Recall curves for an input size of 4 foods. Table 4 Mean training and recommendation times in mil- liseconds. Model Training Mean recommendation AR 3905.1 39.5 PAR 6904.9 2.5 TIC 93710.2 32.0 Fig. 3. The ratio of mean nDCG for the top 15 results to the number of input foods. Fig. 4. The ratio of recall for the 15 results to the number of input foods for pair- wise association rules with the first and the second levels of specificities and man- ually entered associated food prompts. 5 r f a f a i t i e r c c c size of input foods. PAR also demonstrates higher nDCG than TIC and AR for all input sizes ( Fig. 3 ). PAR is the second fastest algo- rithm to produce a model (after AR) but the fastest to produce a single set of recommendations ( Table 4 ). Based on this comparison, PAR is selected to be used for the implementation of the associ- ated foods recommender algorithm. At the same time, these results demonstrate that the quality of predictions produced by PAR is still relatively low. In the following experiments we aim to improve the performance of the algorithm by exploiting the context of the task it is used for. .2. Associated food questions To compare the efficacy of recommendations produced by the ecommender algorithm to the existing hand-coded associated ood questions we go through the same evaluation procedure as bove, except that on the recommend step a trained model returns ood categories instead of exact foods. In this case, true positives re considered to be foods selected by the respondent (excluding nput foods) that belong to one of the food categories predicted by he recommender algorithm. The taxonomy of foods implemented n Intake24 allows control of the specificity of the returned cat- gories. So, we demonstrate the performance of the algorithm in eturning the direct parent category of a food (first level, e.g. Flake ereals is the parent category for Choco flakes ) and a more generic ategory (second level, e.g. Breakfast cereals ) that is a parent of the ategory with the first-level specificity ( Fig. 4 ). Since the existing T. Osadchiy et al. / Expert Systems With Applications 115 (2019) 535–542 541 Table 5 Omitted foods captured with pairwise association rules but not with manually entered associated food prompts. Input foods First-level specificity Second-level specificity Chicken breast; Fanta; Instant potato Gravy Sauces, condiments, gravy and stock Bananas; Fruit and yoghurt smoothie; Semi skimmed milk Sugar Sugar, jams, marmalades, spreads and pates Blackcurrant squash (juice), e.g ribena; Heinz beans and sausages Brown bread toasted Brown, wholemeal and 50:50 bread Porridge, made with skimmed milk; Tea; White sugar Butter Butter/margarine/oils Tuna mayo sandwich; Volvic mineral water, still or fizzy Chocolate covered biscuits Sweet biscuits Bread sticks; Coffee Dips Pickles, olives, dips and dressings Cheese and tomato pizza (includes homemade); Raspberries Ice cream Ice cream & ice lollies Cheese sandwich; Tea Crisps and snacks Crisps, snacks and nuts Green Olives; Water Wine Alcohol Bottled mineral water; Chicken breast fillet; Chips, fried; Hot sauce Fizzy drinks Drinks Still energy drink, eg Lucozade Hydroactive, Gatorade, Powerade; Tuna in brine, tinned; White bread sliced Mayonnaise Sauces/condiments/gravy/stock a t r d f a w 7 l r o s I t l d o c d e f 5 s fi t t m 2 e e V w t t n t b a Fig. 5. The ratio of mean nDCG for the top 15 results to the number of input foods for the search results ranked based on pairwise association rules and FRC. u q p i s t a w o g 6 e w m i f a u p r p ssociated food questions do not store any relevance scores plot- ing their PR curves or assessing their nDCG is impossible. For that eason we compare the recall of the top 15 recommendations pro- uced by the algorithm to the recall of all hand-coded associated oods rules extracted for given input foods. In the simulation of respondents omitting foods hand-coded ssociation food rules recognize 8.3% of omitted foods at most, hereas the recommender algorithm’s peak recall is at 58.0% and 9.1% for the first and the second levels of specificity respectively. Table 5 includes examples of commonly forgotten foods estab- ished in the validation of Intake24 ( Bradley et al., 2016 ) but cor- ectly predicted by the recommender algorithm with two levels f specificity. At the time of writing this paper, none of these as- ociations were covered by hand-coded associated foods rules in ntake24. In addition to that, controlling the specificity of the re- urned recommendations allows us to address the cold-start prob- em, so that new foods that have not been reported by any respon- ents can still be captured by their categories. However, the names f some food categories predicted with the second-level specificity ould be perceived as too generic (e.g. ”Pickles, olives, dips and ressings”) and may require being assigned names that would be asier to understand by respondents when displayed in associated ood prompts. .3. Search ranking In response to a respondent’s text query, the existing Intake24 earch algorithm ranks foods based on two types of scores. The rst is the matching cost of the known food description against he query. The matching cost is based on several metrics, including he edit distance between matched words (the approximate string atching is performed using Levenshtein automata Burges et al., 005 ); phonetic similarity of words (using a pluggable phonetic ncoding algorithm that depends on the localization language, .g. Soundex or Metaphone for English Elmagarmid, Ipeirotis, & erykios, 2007 ); the relative ordering of words; the number of ords not matched; and so forth. The lower the matching cost, he better the food name matches the query. The second score is he likelihood of the food being selected, which is measured by the umber of times the food was previously reported. The results are hen sorted, first by decreasing food report count (FRC) and then y increasing matching cost. The evaluation of the associated foods recommender algorithm pplied to the task of ranking search results, follows the same eval- ation procedure, with some variations. In response to each text uery that was recorded into the Intake24 database for each re- orted food (excluding input foods), we retrieve a list of foods us- ng the existing search algorithm. We count foods selected by a re- pondent as true positives and the rest of the results as false nega- ives. We compare the mean nDCG produced by the existing search lgorithm and by the new search algorithm, where FRC is replaced ith PAR. As we can see from the figure below ( Fig. 5 ), PAR slightly utperforms FRC starting from an input size of two foods, with the ap gradually widening as the number of input foods increases. . Conclusions We aimed to address one of the key issues in automated di- tary assessment, which is unintentional under-reporting. To do so, e developed an associated foods recommender algorithm to re- ind respondents of omitted foods and improve the ranking qual- ty of search results returned in response to respondents’ free-text ood name queries. The algorithm, in contrast with collaborative nd content-based filtering approaches, is independent of personal ser profiles and does not require an extensive history of users’ references or a multitude of item descriptors. Instead, the algo- ithm uses transactions performed by respondents from a given opulation to build a collective model of preferences. 542 T. Osadchiy et al. / Expert Systems With Applications 115 (2019) 535–542 D E L L L L M P P R R R R S V Z We considered three approaches to the implementation of the recommender algorithm, based on an implicit social graph ( Roth et al., 2010 ), association rules ( Agrawal et al., 1993 ), and an- alyzing pairwise association rules ( DuMouchel & Pregibon, 2001 ). The evaluation, performed on a large data set of real dietary re- calls, has demonstrated that the implementation based on pair- wise association rules performs better for the defined task. By con- trolling the specificity of the produced recommendations within a reasonable level we achieved a recall of 79.1%. That is signifi- cantly higher than food associations hand-coded by trained nutri- tionists, the recall for which reached only 8.3%. Where a respon- dent filled in at least one food, the recommender algorithm im- proves the ranking of search results. The algorithm was evaluated on dietary recalls of respondents from the UK. As a future work we are planning to analyze how dietary specificities of different regions affect the accuracy of the recommender algorithm. Although the evaluation results described in this paper were produced by analyzing food contents of meals reported by respondents in Intake24, the described methods apply to any recommender tasks where selection of items by the target user can be observed (e.g. email recipients or tags recommenda- tions on community platforms). Supplementary material Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.eswa.2018.07.077 . References Agrawal, R., Imielinski, T., & Swami, A. (1993). Mining association rules between sets of items in large databases. In Proc. 1993 ACM SIGMOD international conference on management of data, SIGMOD’93: 22 (pp. 207–216). ACM. doi: 10.1145/170036. 170072 . Atkinson, J. , Figueroa, A. , & Pérez, C. (2013). A semantically-based lattice approach for assessing patterns in text mining tasks. Computación y Sistemas, 17 (4) . Bradley, J., Simpson, E., Poliakov, I., Matthews, J. N. S., Olivier, P., Adamson, A. J., & Foster, E. (2016). Comparison of INTAKE24 (an online 24-h dietary recall tool) with interviewer-led 24h recall in 11–24 year-old. Nutrients, 8 (6), 358. doi: 10. 3390/nu8060358 . Burges, C., Shaked, T., Renshaw, E., Lazier, A., Deeds, M., Hamilton, N., & Hullen- der, G. (2005). Learning to rank using gradient descent. In Proceedings of the 22nd international conference on machine learning (pp. 89–96). ACM. doi: 10.1145/ 1102351.1102363 . Carrer-Neto, W., Hernández-Alcaraz, M. L., Valencia-García, R., & García- Sánchez, F. (2012). Social knowledge-based recommender system. Application to the movies domain. Expert Systems with Applications, 39 (12), 10990–110 0 0. doi: 10.1016/j.eswa.2012.03.025 . Covington, P., Adams, J., & Sargin, E. (2016). Deep neural networks for youtube rec- ommendations. In Proceedings of the 10th ACM conference on recommender sys- tems, RecSys ’16 (pp. 191–198). ACM. doi: 10.1145/2959100.2959190 . uMouchel, W., & Pregibon, D. (2001). Empirical bayes screening for multi-item associations. In Proceedings of the 7th ACM SIGKDD international conference on knowledge discovery and data mining, KDD’01 (pp. 67–76). ACM. doi: 10.1145/ 502512.502526 . lmagarmid, A. K., Ipeirotis, P. G., & Verykios, V. S. (2007). Duplicate record detec- tion: A survey. IEEE Transactions on Knowledge and Data Engineering, 19 (1), 1–16. doi: 10.1109/TKDE.2007.250581 . i, C. R. J., & Deng, Z. H. (2007). Mining frequent ordered patterns without candi- date generation. In Proceedings 4th international conference on fuzzy systems and knowledge discovery, FSKD 2007: Vol. 1 (pp. 402–406). IEEE. doi: 10.1109/FSKD. 2007.402 . i, H., Wang, Y., Zhang, D., Zhang, M., & Chang, E. (2008). Pfp: Parallel fp-growth for query recommendation. In Proceedings of the 2008 ACM conference on recom- mender systems (pp. 107–114). ACM. doi: 10.1145/1454008.1454027 . ika, B., Kolomvatsos, K., & Hadjiefthymiades, S. (2014). Facing the cold start prob- lem in recommender systems. Expert Systems with Applications, 41 (4 PART 2), 2065–2073. doi: 10.1016/j.eswa.2013.09.005 . inden, G., Smith, B., & York, J. (2003). Amazon.com recommendations: Item-to- item collaborative filtering. IEEE Internet Computing, 7 (1), 76–80. doi: 10.1109/ MIC.2003.1167344 . eng, X., Bradley, J., Yavuz, B., Sparks, E., Venkataraman, S., Liu, D., et al. (2016). Ml- lib: Machine learning in apache spark. The Journal of Machine Learning Research, 17 (1), 1235–1241. doi: 10.1145/2882903.2912565 . azzani, M. J., & Billsus, D. (2007). Content-based recommendation sys- tems. In The adaptive web: Vol. 4321 (pp. 325–341). Springer. doi: 10.1007/ 978- 3- 540- 72079- 9 _ 10 . opescul, A. , Pennock, D. M. , & Lawrence, S. (2001). Probabilistic models for uni- fied collaborative and content-based recommendation in sparse-data environ- ments. In Proceedings of 17th conference on uncertainty in artificial intelligence (pp. 437–4 4 4). Morgan Kaufmann Publishers Inc . aeder, T., & Chawla, N. V. (2011). Market basket analysis with networks. Social Net- work Analysis and Mining, 1 (2), 97–113. doi: 10.1007/s13278- 010- 0 0 03- 7 . oth, M., Ben-David, A., Deutscher, D., Flysher, G., Horn, I., Leichtberg, A., et al. (2010). Suggesting friends using the implicit social graph. In Proceedings of 16th ACM SIGKDD international conference on knowledge discovery and data mining (pp. 233–242). ACM. doi: 10.1145/1835804.1835836 . udin, C. , Letham, B. , Salleb-Aouissi, A. , Kogan, E. , & Madigan, D. (2011). Sequential event prediction with association rules. In Proceedings of 24th annunal confer- ence on learning theory (pp. 615–634) . uiz, P. P. , Foguem, B. K. , & Grabot, B. (2014). Generating knowledge in maintenance from experience feedback. Knowledge-Based Systems, 68 , 4–20 . Salzberg, S. L. (1997). On comparing classifiers: Pitfalls to avoid and a recommended approach. Data Mining and Knowledge Discovery, 1 (3), 317–328 . Schein, A. I., Popescul, A., Ungar, L. H., & Pennock, D. M. (2002). Methods and met- rics for cold-start recommendations. In Proceedings of the 25th annual interna- tional ACM SIGIR conference on research and development in information retrieval, SIGIR ’02 (pp. 253–260). ACM. doi: 10.1145/564376.564421 . ene, A. , Kamsu-Foguem, B. , & Rumeau, P. (2018). Discovering frequent patterns for in-flight incidents. Cognitive Systems Research, 49 , 97–113 . Shaw, G., Xu, Y., & Geva, S. (2010). Using association rules to solve the cold- start problem in recommender systems. In Pacific-Asia conference on knowl- edge discovery and data mining: 6118 (pp. 340–347). Springer. doi: 10.1007/ 978- 3- 642- 13657- 3 _ 37 . an Deursen, A. J., & Van Dijk, J. A. (2009). Using the Internet: Skill related problems in users’ online behavior. Interacting with Computers, 21 (5–6), 393–402. doi: 10. 1016/j.intcom.20 09.06.0 05 . heng, Z., Kohavi, R., & Mason, L. (2001). Real world performance of association rule algorithms. In Proceedings of 7th ACM SIGKDD international conference on knowledge discovery and data mining, KDD’01 (pp. 401–406). ACM. doi: 10.1145/ 502512.502572 . https://doi.org/10.1016/j.eswa.2018.07.077 https://doi.org/10.1145/170036.170072 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0002 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0002 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0002 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0002 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0002 https://doi.org/10.3390/nu8060358 https://doi.org/10.1145/1102351.1102363 https://doi.org/10.1016/j.eswa.2012.03.025 https://doi.org/10.1145/2959100.2959190 https://doi.org/10.1145/502512.502526 https://doi.org/10.1109/TKDE.2007.250581 https://doi.org/10.1109/FSKD.2007.402 https://doi.org/10.1145/1454008.1454027 https://doi.org/10.1016/j.eswa.2013.09.005 https://doi.org/10.1109/MIC.2003.1167344 https://doi.org/10.1145/2882903.2912565 https://doi.org/10.1007/978-3-540-72079-9_10 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0015 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0015 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0015 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0015 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0015 https://doi.org/10.1007/s13278-010-0003-7 https://doi.org/10.1145/1835804.1835836 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0018 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0019 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0019 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0019 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0019 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0019 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0020 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0020 https://doi.org/10.1145/564376.564421 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0022 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0022 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0022 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0022 http://refhub.elsevier.com/S0957-4174(18)30441-X/sbref0022 https://doi.org/10.1007/978-3-642-13657-3_37 https://doi.org/10.1016/j.intcom.2009.06.005 https://doi.org/10.1145/502512.502572 Recommender system based on pairwise association rules 1 Introduction 2 Related work 3 Associated food recommender algorithm 3.1 Intake24 3.2 Generic procedure 3.3 Association rules 3.4 Transactional item confidence 3.5 Pairwise association rules 4 Methodology 5 Results 5.1 General performance 5.2 Associated food questions 5.3 Search ranking 6 Conclusions Supplementary material References 
work_3lzp6xtt5javxo757ggrue4gse	----	PII: S0957-4174(03)00066-6 An application of expert systems to botanical taxonomy W. Fajardo Contreras a , E. Gibaja Galindo a,*, A. Bailón Morillas b , P. Moral Lorenzo a a Universidad de Granada, E.T.S. de Ingenierı́a Informática, Departamento de Ciencias de la Computación e Inteligencia Artificial, C/Periodista Daniel Saucedo Aranda, 18071 Granada, Spain. b Universidad de Almerı́a, Escuela Politécnica Superior, Departamento de Lenguajes y Computación, Carretera Sacramento s/n, 04120 La Cañada de San Urbano, Almerı́a, Spain. Abstract The implementation of intelligent systems is not particularly widespread in the field of Botany and even less so on Internet. At present, we can currently only find hypertext documents or databases which store unprocessed information. The GREEN (Gymnosperms Remote Expert Executed Over Networks) System is presented as the application of Artificial Intelligence techniques to the problem of botanical identification. GREEN is an Expert System for the identification of Iberian Gymnosperms which allows online queries to be made. It can be consulted in: http://drimys.ugr.es/experto/index.html q 2003 Elsevier Ltd. All rights reserved. Keywords: Gymnosperms; Identification keys; Expert Systems; Artificial Intelligence; World Wide Web; Iberian Peninsula 1. Introduction Plant Taxonomy is a complex, meticulous science which allows taxa to be identified by retrieving information contained on them in a classification system. There are various ways which this identification may be carried out, although the one most commonly used employs dichotomic keys (a process which requires knowledge of botanical terminology and organography). As a result of the complex- ity of this process, botany-related activities are not particularly automated. In fact, the systems which exist are basically databases which store files on the specimens. Artificial Intelligence can offer a more productive approach to these systems by processing the information they contain in order to obtain knowledge which has not been stored explicitly in the database. Within the wealth and variety offered by the plant kingdom, the subject of scientific disclosure has been dealt with using Artificial Intelligence techniques with a specific study of the group of Gymnosperms (Gymnospermae) in the Iberian peninsula. This group was chosen due to the presence of important forest species which it contains. In addition, many of these offer resources or are cultivated as ornamental, which makes their identification useful for non botanical expert users. This has all given rise to GREEN (Gymnosperms Remote Expert Executed Over Networks), a pioneering system in the application of Artificial Intelligence techniques to the field of botany. GREEN is an online decision aid system, resulting in a much greater and faster diffusion of knowledge and a broader receptor spectrum. 2. Material and methods We have divided this study on GREEN into 5 parts: † A first part (Sections 2.1 and 2.2) in which we describe the structure of the system, and defines the main modules which comprise the system and the knowledge gathered. † A second part (Sections 2.3, 2.4, 2.5, 2.6) in which we develop the process for acquiring and validating the knowledge available on the problem domain until a knowledge base is finally obtained. In this part, the processing of imprecise information, common to this type of problem, is also discussed. † A third part (Section 2.7) is devoted to the reasoning process which the System uses. 0957-4174/03/$ - see front matter q 2003 Elsevier Ltd. All rights reserved. doi:10.1016/S0957-4174(03)00066-6 Expert Systems with Applications 25 (2003) 425–430 www.elsevier.com/locate/eswa * Corresponding author. Tel.: þ34-958240468; fax.: þ 34-958243317. E-mail addresses: gibaja@decsai.ugr.es (E.G. Galindo). http://www.elsevier.com/locate/eswa † A fourth part (Section 2.8) in which we discuss other important features of the System. † Finally, we finish (Section 3) with conclusions drawn directly from what has been presented in this article and from the bibliography used. 2.1. System structure The system structure is directly derived from the way in which botanical experts work. Dichotomic keys of the type IF – THEN are used for the classification and recognition of plant species. That is to say, that each key leads to either another key or a plant species. In this way, when a botanist wants to classify a particular species, it is possible to distinguish: † A source of knowledge comprising all the available information on each plant species in the form of dichotomic rules. † A process of the use of this knowledge in order to solve the particular problem (keys are searched until a particular species is identified). This description coincides perfectly with that of a Knowledge-Based System and more specifically with that of a rule-based Expert System (Luger and Stubblefield, 1993) with: a Knowledge Base which stores knowledge about the domain of the problem in the form of rules and an Inference Engine which extracts information from the Knowledge Base. In addition to the two essential modules described in the previous paragraph and reflecting the ideal structure of a Knowledge-Based System, the System has: † An uncertainty processing module fitting the nature and subjectivity of the observer. † A justifying module which will explain the results achieved to the System in a language close to the natural language. † We will also add user support modules. † A multimedia database to reference known species. † A glossary of scientific terms to make the System more accessible to users who are not botanical experts. Additionally to design and implement a server which will deal with user (or client) requests and send the results by Internet is needed. In Section 2.2 we outline the process for the design and implementation of the System, detailing the Artificial Intelligence techniques which have been applied. 2.2. Knowledge gathered by the system The first stage is to determine its application domain, that is, the type of knowledge the System will manage. As we have mentioned, the group of Gymnosperms has been chosen from which information is provided on 46 taxa present in the Iberian Peninsula (Castroviejo et al., 1986) both autochthonous and cultivated. In addition to the Knowledge Base, which has been optimized in order to obtain results in the queries, the System gathers information on the System domain in other formats and these are incorporated into a multimedia database which provides images and data about its distribution and ecology and a glossary of botanical terms which make the arduous task of species identification easier and more enjoyable. 2.3. Knowledge acquisition and elicitation The first problem when developing the System is that the information available on the problem domain does not have a structure which may be directly translated to a Computer System. The information is dispersed, incomplete; it is imprecise and unstructured. In order to be able to represent the knowledge in an appropriate way, a process of knowledge acquisition and elicitation is needed, and on which the final functions of the System depend to a large extent. In order to begin the acquisition and elicitation process, we begin with different keys (Blanca and Morales, 1991, Font Quer 1979, López González 1982, Garcı́a Rollán, 1983, Krüssmann, 1972). We gather and summarize their information, thereby producing a list of diagnostic characters (descriptors or attributes) at family, genus, species and subspecies level. This hierarchical organization of the information offers the advantage of multilevel answers so that, even with little information, some objective may be reached in the higher levels of the hierarchy. This has a simple explanation: Generally, in order to reach an objective in the higher levels of the hierarchy only a small amount of information is needed, which is also what is observed more easily. Heuristically, this leads us to suppose that the minimum amount of information which the user knows will be that which will allow inference in the highest levels. As information becomes known, the response will be refined until the lower and less general levels of the hierarchy are reached. The more information we have, the more we will know, nevertheless, results may generally be obtained with little information. All information has subsequently been compared by observing nature and consulting herbalist documents and experts. The most important taxonomical characters in Gymnos- perms have been divided into different groups: general aspect of the taxon, characteristics of the leaf, of the branches, of the shoots, monoecious or dioecious, charac- teristics of the fructification (cone and ‘berry’ cone), of the seeds, and ecology of the taxon. With these characters, decision tables have been compiled (Durkin, 1994), which gather the identifying diagnostic characters for each taxon ‘Table 1’. As it is not W.F. Contreras et al. / Expert Systems with Applications 25 (2003) 425–430426 advisable for these tables to have many empty cells, they have been filled in since many were not necessary when the taxon were identified using the traditional method. Although initially filling in a table of this type supposes a greater effort than using dichotomic keys directly, this investment is easily compensated for since these will enable us to apply Artificial Intelligence techniques in order to obtain keys which are different from the standard ones. Botany uses identification keys, whereas applied Artifi- cial Intelligence techniques determine the minimum set of diagnostic characters in order to recognize the different taxa. Artificial Intelligence allows us to find determining characters, which exclude others, and this enables quicker identification than that provided by the traditional method. 2.4. Obtaining the Knowledge Base A set of rules (represented in the Knowledge Base) is obtained automatically from the tables. For this, we apply Artificial Intelligence learning techniques (Machine Learn- ing), in particular we modify the ID3 algorithm proposed by Quinlan (Ignizio, 1991), so that it allows us to obtain more than one rule per objective. For this: † We use Occam’s razor criterion as a heuristic for ramification (simple explanations are preferable to more complex explanations) quantifying this criterion through the use of the concept of entropy. In this way, rules of minimum length are created which exclude irrelevant knowledge, since irrelevant descriptors will not be taken into account. † We obtain a Knowledge Base, the content of which is more complete than that of the dichotomic keys, since it contains all the consistent rules which may be obtained according to the selected descriptors in order to determine the objectives. The rules provide a structuring of the knowledge which the user can understand and which is similar to the dichotomic keys used by expert botanists. When the System presents its conclusions in the form of rules, the user understands the reasoning followed by GREEN perfectly and the user becomes familiar with the reasoning process followed by the human experts who have contributed their knowledge to the System (learning). 2.5. Treatment of uncertainty Information about the domain is based on what normally happens, but every rule has its exceptions. As it is usual for some data not to be known with absolute certainty and since expert knowledge is not always defined with complete certainty, errors of measurement may be committed. But this does not mean that the information that we have should be rejected as not only are experts able to work with uncertainty but good results can also be obtained regardless. Given this large amount of sources of uncertainty, GREEN incorporates a module to deal with uncertainty. Uncertainty is modeled using certainty factors (Shortlife & Buchanan, 1975) since it is a simple computational model which allows experts to estimate confidence in each hypothesis and in the conclusion, facilitating the expression of subjective certainty estimations. This model also enables knowledge to be represented easily in the form of rules and has successfully been used in many other systems. 2.6. Consistency reinforcer During the development of the Knowledge Base, inconsistencies may arise mainly due to errors during the knowledge acquisition and elicitation stage or during the design or implementation of the technique for automatically obtaining the rules. Another important note is that GREEN is capable of accommodating uncertainty which is why inconsistencies about the certainty of results cause an additional impact. Consequently, this makes it necessary for GREEN to incorporate a consistency reinforcer which systematically analyzes each of the rules in the Knowledge Base in order to be able to detect possible errors (Gonzalez and Dankel Table 1 Decision table Arrangement of the ‘berry’ cones Color of the ‘berry’ cone Pruinose ‘berry’ cone Size of the ‘berry’ cone No. of seeds in the ‘berry’ cone Juniperus communis subsp. communis Axillary Bluish-black Yes Between 0.6 and 1 cm 3 Juniperus communis subsp. hemisphaerica Axillary Bluish-black Yes Between 0.6 and 1 cm 3 Juniperus communis subsp. alpina Axillary Bluish-black Yes Between 0.6 and 1 cm 3 Juniperus oxycedrus subsp. oxycedrus Axillary Brown No Between 0.6 and 1 cm 1 – 3 Juniperus oxycedrus subsp. badia Axillary Brown No More than 1 cm 1 – 3 (· · ·) (· · ·) (· · ·) (· · ·) (· · ·) (· · ·) W.F. Contreras et al. / Expert Systems with Applications 25 (2003) 425–430 427 1993) which have been introduced during the design process thereby guaranteeing that the Knowledge Base has been correctly designed and implemented. 2.7. System reasoning The Inference Engine provides the control mechanism and knowledge inference (a process used in an expert System in order to derive new information from information known). It combines the input facts with the knowledge gathered in the Knowledge Base thereby responding to user queries. In order to design the Inference Motor, Ignizio’s BASELINE with forward chaining has been taken as a model (Ignizio, 1991). The Inference Engine incorporated into the System is quite a different module from the Knowledge Base. This differentiation is important since: 1. Knowledge may be represented more naturally. The knowledge model together with the inference process reflects the problem-solving mechanism followed by a human being better than a model which incrusts knowledge within the inference process. 2. The System designers can focus on capturing and organizing the knowledge common to the problem domain independently of its implementation. 3. It enables the content of Knowledge Base to be changed without the need to change the control System so that a) the Knowledge Base may be updated without changing the Inference Engine b) a single Inference Engine may be used to solve different problems. 2.8. Other characteristics As we have already mentioned, GREEN is extremely easy to use (see Fig. 1). The specimen descriptors are grouped into general categories (general appearance, leaf, branch, cone, etc.) with names which are familiar to all users. Within each category, users select the descriptor they know and enter a value for the degree of belief. The System has been provided with two methods for entering the query: basic and advanced. In the basic mode, the user has a set of options, so that the use of certainty Fig. 1. A screen shot for the user interface for introduction of data. Author: Eva Lucrecia Gibaja Galindo. W.F. Contreras et al. / Expert Systems with Applications 25 (2003) 425–430428 Fig. 2. A screen shot for the user interface for identification results. Author: Eva Lucrecia Gibaja Galindo. Fig. 3. A screen shot for the user interface for additional information. Author: Eva Lucrecia Gibaja Galindo. W.F. Contreras et al. / Expert Systems with Applications 25 (2003) 425–430 429 factors is clear. In the advanced mode, the user must manually enter the certainty value of the observation. After entering the data, the inference process is executed and the System gives the user a set of results ordered according to how well they fit the query and an outline of the reasoning followed in order to reach these conclusions. If the user wishes, it is possible to increase the information about the specimen by accessing the multimedia database. GREEN is specifically designed to work on Internet which is why interaction with the user is carried out using forms which send the data and the queries to a remote server. The entire transfer of information online has been minimized so as not to overload the server and in order to obtain a satisfactory System response time for the user. GREEN has been designed independently of the type of botanical database on which it is employed, so that it may be easily adapted in order to classify species other than Gymnosperms. Figs. 1 – 4. 3. Conclusions 1. Computing offers new advantages to the popularization of Botany, including the production of automatic keys or computer-generated keys, which will make it possible for non-experts to identify plants. 2. In this paper, an expert System is presented which will offer the user a new ‘interactive’ species identification method. 3. The GREEN System is a practical tool which may be used online and which will enable different taxa comprising the Iberian Gymnosperm flora to be recognized. References Blanca, G., & Morales, C. (1991). Flora del Parque Natural de la Sierra de Baza. Granada: Servicio de Publicaciones de la Universidad de Granada. Castroviejo, S., Laı́nz, M., López González, G., Montserrat, P., Muñoz Garmendia, F., Paiva, J., & Villar, L. (1986). Flora Ibérica. Plantas vasculares de la Penı́nsula Ibérica e Islas Baleares (Vol. 1). Lycopodiaceae-Papaveraceae, Madrid: Real Jardı́n Botánico. Durkin, J. (1994). Expert systems. Design and development. London: Prentice Hall International. Font Quer, P. (1979). Diccionario de Botánica. Barcelona: Labor. Garcı́a Rollán, M. (1983). Claves de la flora de España (Vol. I). Penı́nsula y Baleares, Madrid: Mundi-Prensa. Gonzalez, A. J., & Dankel, D. D. (1993). The Engineering of knowledge- based systems. Theory and practice. Englewood Cliffs, NJ: Prentice- Hall International. Ignizio, J. P. (1991). Introduction to expert systems. The development and implementation of rule-based expert systems. New York: McGraw-Hill. Krüssmann, G. (1972). Manual of cultivated conifers. Portland: Timber Press. López González, G. (1982). La Guı́a de Incafo de los árboles y arbustos de la Penı́nsula Ibérica. Madrid: INCAFO. Luger, G. F., & Stubblefield, W. A. (1993). Artificial intelligence. Structures and strategies for complex problem solving. The Benjamin/Cummings series in artificial intelligence, Redwood City: Benjamin/Cummings. Shortlife, E., & Buchanan, B. G. (1975). A model of inexact reasoning in medicine. Mathematical Biosciences, 23, 351 – 379. Fig. 4. Other screen shot for the user interface for additional information. Author: Eva Lucrecia Gibaja Galindo. W.F. Contreras et al. / Expert Systems with Applications 25 (2003) 425–430430 An application of expert systems to botanical taxonomy Introduction Material and methods System structure Knowledge gathered by the system Knowledge acquisition and elicitation Obtaining the Knowledge Base Treatment of uncertainty Consistency reinforcer System reasoning Other characteristics Conclusions References 
work_3rdd7ykfkveu5asatfqcf24ohq	----	Expert systems for knowledge management: crossing the chasm between information processing and sense making Expert systems for knowledge management: crossing the chasm between information processing and sense making Y. Malhotra* Florida Atlantic University, 818 N.W. 89th Avenue, Fort Lauderdale, FL 33324, USA Abstract Based on insights from research in information systems, information science, business strategy and organization science, this paper develops the bases for advancing the paradigm of AI and expert systems technologies to account for two related issues: (a) dynamic radical discontinuous change impacting organizational performance; and (b) human sense-making processes that can complement the machine learning capabilities for designing and implementing more effective knowledge management systems. q 2001 Elsevier Science Ltd. All rights reserved. Keywords: Expert systems; Arti®cial intelligence; Knowledge management; Information systems; Information science; Business strategy; Discontinuous change; Sense making; Information processing ªThere has been an over-concentration on Shannon's de®nition of information in terms of uncertainty (a very good de®nition for the original purposes) with little attempt to understand how MEANING directs a message in a network. This, combined with a concen- tration on end-points (equilibria) rather than proper- ties of the trajectory (move sequence) in games has lead to a very unsatisfactory treatment of the dynamics of organizations.º Ð John H. Holland (personal communication, June 21, 1995) 1 1. Introduction The narrative cited above as an observation by the noted psychologist and computer scientist John Holland was in response to my query to him regarding the possibility of using intelligent information technologies for devising self-adaptive organizations. As meaning seems to be a crucial construct in understanding how humans convert information into action [and consequently performance], it is evident that information-processing based ®elds of arti®- cial intelligence and expert systems could bene®t from understanding how humans translate information into mean- ings that guide their actions. In essence, this issue is relevant to the design of both human- and machine-based knowledge management systems. Most such systems had been tradi- tionally based on consensus and convergence-oriented information processing systems, often based on mathema- tical and computation models. Increasing radical discontin- uous change (cf. Huber & Glick, 1993; Nadler, Shaw, & Walton, 1995) that characterizes business environments of today and tomorrow, however, requires systems that are capable of multiple Ð complementary and contradictory Ð interpretations. Despite observations made by Churchman (1971) and Mason and Mitroff (1973), the paradigm of information systems, arti®cial intelligence (AI) and expert systems have yet to address the needs posed by wicked environ- ments that defy the logic of pre-determination, pre- diction and pre-speci®cation of information, control and performance systems (cf. Malhotra, 1997). Wicked Expert Systems with Applications 20 (2001) 7±16PERGAMON Expert Systems with Applications 0957-4174/01/$ - see front matter q 2001 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 7 - 4 1 7 4 ( 0 0 ) 0 0 0 4 5 - 2 www.elsevier.com/locate/eswa * Tel.: 11-954-916-1585. E-mail address: yogesh.malhotra@brint.com (Y. Malhotra). 1 Considering organizational adaptation for survival and competence as the key driver for most organizational information and knowledge processes (cf. Malhotra, 2000a,b,c), it seemed logical to develop the model of IT-enabled self-adaptive organizations based upon technologies that are often considered as a benchmark for self-adaptive behavior. In this context, genetic algorithms (also referred to as adaptive computation) offer the closest archetype for devising technology-enabled organizations that could possibly exhibit self-adaptive behavior given the dynamically chan- ging environment. By offering the basis for evolution of solutions to parti- cular problems, controlling the generation, variation, adaptation and selection of possible solutions using genetically based processes, it seemed probable that genetic algorithms could offer the basis for self-adaptive evolution of organizations. As solutions alter and combine, the worst ones are discarded and the better ones survive to go on and produce even better solutions. Thus, genetic algorithms breed programs that solve problems even when no person can fully understand their structure. business environments Ð characterized by radical discontinuous change Ð impose upon organizations the need for capabilities for developing multiple mean- ings or interpretations and continuously renewing those meanings given the changing dynamics of the environ- ment. Scholars in business strategy have advocated human and social processes such as `creative abrasion' and `creative con¯ict' (cf. Eisenhardt, Kahwajy, & Bourgeois, 1997; Leonard, 1997) for enabling the inter- pretive ¯exibility (Nonaka & Takeuchi, 1995) of the organization. It is also evident that there is an imperative need for relating the static notion of information captured in data- bases or processed through computing machinery to the dynamic notion of human sense making. More importantly, our current understanding of information as the [indirect] enabler of performance can immensely bene®t from unra- veling the intervening processes of human sense making that are more directly related to action (or inaction) and resulting performance outcomes (or lack thereof). Based upon a review of the current state of AI and expert systems research and practice in knowledge management, this article develops the bases for AI and expert systems researchers to develop knowledge management systems for addressing the above needs. Section 2 provides an over- view of the state-of-the-art expert systems research and practice issues related to knowledge management highlight- ing key relationships with the key theses of the article. Section 3 offers a more current understanding of knowledge management as it relates to organizational adaptability and sustainability by drawing upon information systems and business strategy research. Section 4 highlights the contrast between the computational model of information processing and human sense making while recognizing both as valid meaning making processes. Finally, sense-making bases of human action and performance are discussed in Section 6, followed by conclusions and recommendations for future research in Section 8. 2. State of related research and practice in AI and expert systems Faced with uncertain and unpredictable business environ- ments, organizations have been turning to AI and expert systems to develop knowledge management systems that can provide the bases for future sustainability and compe- tence. For instance, faced with competition and uncertainty in the ®nance industry, banks are using neural networks to make better sense of a plethora of data for functions such as asset management, trading, credit card fraud detection and portfolio management (Young, 1999). Similarly, insurance and underwriting industries are relying upon knowledge management and AI technologies to offer multiple channels for rapid response to customers (Rabkin & Tingley, 1999). Many such knowledge management implementations using AI and expert systems rely upon the meaning making and sense-making capabilities of AI and expert systems technol- ogies and humans using them. In recent years, there have been signi®cant advances in endowing inanimate objects with limited sense-making capabilities characteristic of self-adaptive behavior of humans. For instance, some proponents of `perceptual intel- ligence' (cf. Pentland, 2000) have suggested such capabil- ities derived from a computers' ability to isolate variables of interest by classifying any situation based on categorization heuristics for taking appropriate action. Their suggestion is that once a computer has the perceptual ability to know who, what, when, where and why, then the probabilistic rules derived by statistical learning methods are normally suf®- cient for the computer to determine a course of action. However, these models, though helpful for procedural deci- sion making, need to advance beyond the static, pre-speci- ®ed and pre-determined logic to account for dynamically changing environments that may require fundamental and radical rede®nition of underlying rules as well as the beha- vior of the actors. Similarly, research on `perceptual interfaces' has been trying to unravel how people experience information that computers deliver (cf. Reeves & Nass, 2000). This stream of research is based on the premise that human experience with information is caused by stimulation of the senses. While paying attention to the chemical senses (taste and olfaction), the cutaneous senses (skin and its receptors), vision and hearing, this research has yet to take into consideration the interpretive, meaning making and sense-making processes that occur at a more cerebral level. The personal constructivist theory discussed in this article could help better relate information to meaning and consequent beha- vior (or actions) in above cases. Simultaneously, the state-of-art research and practice in data mining, often described as ªknowledge discovery from databases,º ªadvanced data analysis,º and machine learn- ing, has been trying to decipher how computers might auto- matically learn from past experience to predict future outcomes (Mitchell, 1999). However, as explained later, current thinking in business strategy is imposing upon the organization the need to move beyond prediction of future to anticipation of surprise (Malhotra, 2000a,b). The most advanced machine learning capabilities Ð such as those of the most advanced chess-playing computer (cf. Camp- bell, 1999) Ð are still limited by pre-speci®ed, pre- determined de®nition of problems that are solved based on the pre-speci®ed rules of the game. Though interesting, such capabilities may have limited use in the emerging game of strategy that is being rede®ned as it is being played. In such game, all ªrules are up for grabsº even though computational machinery has yet to evolve to the stage of sensing changes that it has not been pre-programmed to sense and to re-evaluate the rules embedded in the logic devised by human programmers. In contrast to machine learning, humans are endowed with Y. Malhotra / Expert Systems with Applications 20 (2001) 7±168 https://isiarticles.com/article/11857 
work_3sxzzd3jjneznfcxhrsajjc32i	----	Data mining with various optimization methods Data mining with various optimization methods Vladimir Nedic a, Slobodan Cvetanovic b, Danijela Despotovic c, Milan Despotovic d,⇑, Sasa Babic e a Faculty of Phil. and Arts, University of Kragujevac, Jovana Cvijica bb, 34000 Kragujevac, Serbia b Faculty of Economics, University of Nis, Trg kralja Aleksandra Ujedinitelja 11, 18000 Nis, Serbia c Faculty of Economics, University of Kragujevac, Djure Pucara Starog 3, 34000 Kragujevac, Serbia d Faculty of Engineering, University of Kragujevac, Sestre Janjic 6, 34000 Kragujevac, Serbia e College of Applied Mechanical Engineering, Trstenik, Serbia a r t i c l e i n f o Keywords: Traffic noise Artificial intelligence Genetic algorithm Hooke and Jeeves Simulated annealing Particle swarm optimization Software a b s t r a c t Road traffic represents the main source of noise in urban environments that is proven to significantly affect human mental and physical health and labour productivity. Thus, in order to control noise sound level in urban areas, it is very important to develop methods for modelling the road traffic noise. As observed in the literature, the models that deal with this issue are mainly based on regression analysis, while other approaches are very rare. In this paper a novel approach for modelling traffic noise that is based on optimization is presented. Four optimization techniques were used in simulation in this work: genetic algorithms, Hooke and Jeeves algorithm, simulated annealing and particle swarm optimization. Two different scenarios are presented in this paper. In the first scenario the optimization methods use the whole measurement dataset to find the most suitable parameters, whereas in the second scenario optimized parameters were found using only some of the measurement data, while the rest of the data was used to evaluate the predictive capabilities of the model. The goodness of the model is evaluated by the coefficient of determination and other statistical parameters, and results show agreement of high extent between measured data and calculated values in both scenarios. In addition, the model was com- pared with classical statistical model, and superior capabilities of proposed model were demonstrated. The simulations were done using the originally developed user friendly software package. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction Road traffic noise along with the noise coming from railways and industries represents very important factor regarding environ- mental pollution in urban areas. The influence of traffic noise on human health has been studied on numerous occasions in recent years (Brink, 2011; Fyhri & Klboe, 2009; Pirrera, De Valck, & Cluydts, 2010) resulting that this kind of annoyance significantly affects both mental and physical health in many ways: causing anxiety, stress, hearing impediments, sleep disturbance, cardiovas- cular problems, etc. Thus, in order to control noise sound level in urban areas, it is very important to develop methods for prediction of the traffic noise. Due to the rapid development of means of transportation and road traffic, the influence of the traffic flow structure on the level of traffic noise is an important area of research. Through the monitoring of basic flow parameters and their trends it is possible to predict and monitor noise that appears in the certain part of the transport network. In this way, the effect of noise reduction can be achieved through different modes of traffic management, which is particularly important for human health and environmental improvement. The first traffic noise prediction (TNP) models date back to early 1950s. Since then large number of methods and models for traffic noise prediction has been developed. The critical reviews of the most used ones are given in Steele (2001) and Quartieri et al. (2009). Most of the TNP models that are presented in literature are based on linear regression analysis. The main limit of those models, as concluded in Quartieri et al. (2009) and Guarnaccia, Lenza, Mastorakis, and Quartieri (2011), is ‘‘that they do not take into account the intrinsic random nature of traffic flow, in the sense that they do not take care of how vehicles really run, consid- ering only how many they are’’. More advanced models involve artificial neural networks (ANN) (Cammarata, Cavalieri, & Fichera, 1995; Givargis & Karimi, 2010) and genetic algorithms (Gndogdu, Gkdad, & Yksel, 2005; Rahmani, Mousavi, & Kamali, 2011). ANN model that was used in Cammarata et al. (1995) has 3 inputs: equivalent number of vehicles, which was obtained by adding to the number of cars number of motorcycles multiplied by 3 and number of trucks multiplied by 6, the average height of the build- ings on the sides of the road, and the width of the road. In order to increase the number of inputs authors decomposed equivalent number of vehicles into the number of cars, the number of 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.12.025 ⇑ Corresponding author. Tel.: +381 69 844 9679. E-mail address: mdespotovic@kg.ac.rs (M. Despotovic). Expert Systems with Applications 41 (2014) 3993–3999 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2013.12.025&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.12.025 mailto:mdespotovic@kg.ac.rs http://dx.doi.org/10.1016/j.eswa.2013.12.025 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa motorcycles, and the number of trucks, and got the ANN model with 5 inputs. In terms of the parameters involved in the CoRTN (Calculation of Road Traffic Noise) model (Quartieri et al., 2009), which was initially developed in 1975 by the Transport and Road Research Laboratory and the Department of Transport of the Uni- ted Kingdom, the ANN model that was used in Givargis and Karimi (2010) has 5 input variables: the total hourly traffic flow, the percentage of heavy vehicles, the hourly mean traffic speed, the gradient of the road, and the angle of view. Authors tested the developed model on the data collected on Tehran’s roads, and found no significant differences between the outputs of the developed ANN and the calibrated CoRTN model. In Gndogdu et al. (2005) genetic algorithm was used to model the traffic noise in relation to traffic composition (vehicle per hour), the road gradi- ent and the ratio of building height to the road width. In Rahmani et al. (2011) the proposed model is a function of total equivalent traffic flow and equivalent traffic speed. In both papers the authors used MATLAB to find the optimized values of model parameters. In this paper an application of four optimization techniques for the prediction of traffic noise is presented. These techniques are: genetic algorithms, Hooke and Jeeves algorithm, simulated anneal- ing, and particle swarm optimization. The model that is proposed consists of five variables: the number of light motor vehicles, the number of medium trucks, the number of heavy trucks, the num- ber of buses and the average traffic flow speed. All optimized mod- els are tested on data measured on Serbian road using the originally developed user friendly software package. 2. Problem formulation The most suitable measure for depicting traffic noise emission is equivalent sound pressure level ðLeqÞ, which is expressed in units of dbA and corresponds to fictitious noise source emitting steady noise, which in specific period of time contains the same acoustic energy as the observed source with fluctuating noise. For a number of discrete measurements ðNÞ; Leq for time period T is expressed by following equation: Leq ¼ 10log10 1=T XN i¼1 10 Li 10 ! ð1Þ where Li is sound pressure level, which corresponds to i th measurement. In order to reduce the noise it is necessary to know functional relationship between the equivalent sound pressure level and influential parameters. Leq is correlated to numerous parameters, such as numbers and types of vehicles, their velocities, type of road surface, width and slope of the road, height of buildings facing the road, etc. As mentioned in the introduction, in this paper the following variables were considered: the number of light motor vehicles (LMV), the number of medium trucks (STV), the number of heavy trucks (TTV), the number of buses (BUS) and the average traffic flow speed (Vavg). A brief description of how these variables were measured is given in the following chapter. 3. Data sampling For traffic data measurement and for noise measurement on the road M5, automatic traffic counters QLTC-10C and sound level me- ter Bruel&Kajer type 2230 class 1 respectively were used. The equivalent sound pressure levels were measured for time period of 15 min. In order to include greater number of scenarios that might occur in urban environments, a total of 124 measurements of equivalent noise levels for time periods of 15 min were carried out. Measurements of Leq for time period of 15 min were performed at various times to include diversity of the traffic flow as much as possible. Simultaneously, variations in traffic flow, traffic speed and composition of traffic flow were measured. For that reasons the surveys at the same time also consist of the following param- eters: the number of light motor vehicles, the number of medium trucks, the number of heavy trucks, the number of buses, and the average traffic speed in the given time periods. Measurements were taken in accordance with recommenda- tions for road traffic noise measurement; microphone was mounted away from reflecting facades, at a height of 1.2 m above the ground level and 7.5 m away from central line of the road. Dur- ing the measurements it has been taken care that climate condi- tions are as similar as possible (no wind, no rain) in order to eliminate their influence. 4. Mathematical model and methods The equivalent sound pressure level is supposed to be modeled by the following equation: Leq ¼ N1 � log10ðLMVÞþ N2 � log10ðSTVÞþ N3 � log10ðTTVÞ þ N4 � log10ðBUSÞþ N5 � Vavg N6 þ N7 � log10ðVavgÞ ð2Þ where Niði ¼ 1 � 7Þ are coefficients. The problem transforms to find coefficients Ni , such that supposed model best fits experimental data. For that purpose genetic algorithms, Hooke and Jeeves algorithm, simulated annealing, and particle swarm optimization are used. These techniques are briefly described in following subchapters. 4.1. Genetic algorithms Genetic algorithms (Rao, 1996) are class of evolutionary algo- rithms that could be used for a large number of different applica- tion areas. The principle of genetic algorithms is based on Darwin’s theory of evolution, by which the fittest individuals have the best chances to survive. Genetic algorithms operate with a set of individuals (chromosomes) called population. The information Fig. 1. Flowchart of the Genetic algorithm workflow. 3994 V. Nedic et al. / Expert Systems with Applications 41 (2014) 3993–3999 https://isiarticles.com/article/22312 
work_3mu6jnosszdnhdltvokblqh2oy	----	Cloud computing service composition: A systematic literature review Expert Systems with Applications 41 (2014) 3809–3824 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Review Cloud computing service composition: A systematic literature review 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.12.017 ⇑ Corresponding author. E-mail addresses: amin.jula@gmail.com, aminjula@ftsm.ukm.my (A. Jula), elan@ftsm.ukm.my (E. Sundararajan), zalinda@ftsm.ukm.my (Z. Othman). Amin Jula a,⇑, Elankovan Sundararajan b, Zalinda Othman a a Data Mining and Optimization Research Group, Centre for Artificial Intelligence, Faculty of Information Science and Technology, Universiti Kebangsaan Malaysia, UKM Bangi, 43600 Selangor, Malaysia b Centre of Software Technology and Management, Faculty of Information Science and Technology, Universiti Kebangsaan Malaysia, UKM Bangi, 43600 Selangor, Malaysia a r t i c l e i n f o a b s t r a c t Keywords: Cloud computing service composition Systematic literature review Quality of service parameter QoS Research objectives Importance percentage of quality of service parameters The increasing tendency of network service users to use cloud computing encourages web service vendors to supply services that have different functional and nonfunctional (quality of service) features and provide them in a service pool. Based on supply and demand rules and because of the exuberant growth of the services that are offered, cloud service brokers face tough competition against each other in providing quality of service enhancements. Such competition leads to a difficult and complicated pro- cess to provide simple service selection and composition in supplying composite services in the cloud, which should be considered an NP-hard problem. How to select appropriate services from the service pool, overcome composition restrictions, determine the importance of different quality of service param- eters, focus on the dynamic characteristics of the problem, and address rapid changes in the properties of the services and network appear to be among the most important issues that must be investigated and addressed. In this paper, utilizing a systematic literature review, important questions that can be raised about the research performed in addressing the above-mentioned problem have been extracted and put forth. Then, by dividing the research into four main groups based on the problem-solving approaches and identifying the investigated quality of service parameters, intended objectives, and developing environ- ments, beneficial results and statistics are obtained that can contribute to future research. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction exposes a very large number of similar single services to different In recent years, the importance of affordable access to reliable high-performance hardware and software resources and avoiding maintenance costs and security concerns has encouraged large institution managers and stakeholders of information technology companies to migrate to cloud computing. The birth of giant trust- worthy clouds has led to a dramatic reduction in apprehension to- ward such an approach. There are two challenges to address from the standpoint of the significance of all of the needed service accessibilities and efficient allocation possibilities. First, anticipating all of the possible re- quired services is extremely difficult, particularly for software ser- vices. Designing and providing simple and single fundamental services by different service providers will be considered constitu- tive and constructive parts of complicated required services and can be utilized in encountering this problem. The second challenge is in selecting the optimum required single services, which are pro- vided by different service providers with different quality of ser- vice (QoS) attributes; an optimal combination for forming a complicated service must be composed. Addressing this challenge as an optimization problem is an NP-hard problem because it service providers in the cloud. Service composition is one of the best approaches that has been proposed by researchers and applied by cloud providers; this ap- proach can consider both of the mentioned challenges simulta- neously. Selecting appropriate services from a service pool, addressing service composition restrictions, determining the important QoS parameters, understanding the dynamic character- istics of the problem, and having rapid changes in the properties of the services and network are some important issues that must be addressed in this approach to assure the service users’ satisfaction. In the early 2000s and in the years before applications in cloud computing, service composition was introduced and investigated for web services (Kosuga et al., 2002; Milanovic & Malek, 2004; Schmid, Chart, Sifalakis, & Scott, 2002; Singh, 2001). Different arti- ficial and evolutionary algorithms (Ai & Tang, 2008; Canfora, Di Penta, Esposito, & Villani, 2005; Liao et al., 2011; Luo et al., 2011; Tang, Ai, & IEEE, 2010; Wang, 2009; Zhao, Ma, Wang, & Wen, 2012; Zhao et al., 2012) and classic algorithms (Gabrel, Manouvrier, Megdiche, Murat, & IEEE, 2012; Gao, Yang, Tang, Zhang, & Society, 2005; Gekas & Fasli, 2005; Liu, Li, & Wang, 2012; Liu, Wang, Shen, Luo, & Yan, 2012; Liu, Xiong, Zhao, Dong, & Yao, 2012; Liu, Zheng, Zhang, & Ren, 2011; Torkashvan & Haghighi, 2012) have been applied extensively to solve the problem. Design- ing workflows and frameworks for the composition of single ser- vices to achieve specific goals is another approach that has been http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2013.12.017&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.12.017 mailto:amin.jula@gmail.com mailto:aminjula@ftsm.ukm.my mailto:elan@ftsm.ukm.my mailto:zalinda@ftsm.ukm.my http://dx.doi.org/10.1016/j.eswa.2013.12.017 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 3810 A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 observed in the field (Chen, Li, & Cao, 2006; He, Yan, Jin, & Yang, 2008; Song, Dou, & Chen, 2011; Song, Dou, & Chen, 2011). Service composition techniques were first applied in cloud com- puting systems in 2009 (Kofler, ul Haq, & Schikuta, 2009; Zeng, Guo, Ou, & Han, 2009). Afterward, a substantial effort was made in this area. The number of ongoing studies in cloud computing service composition is rapidly increasing due to the increasing ten- dency of researchers within different areas of expertise to address the problem. A glimpse of reliable published works shows that researchers who are interested in this promising area face a tre- mendous number of novel ideas, mechanisms, frameworks, algo- rithms and approaches, and further expanding the scope of problems. Furthermore, several existing datasets, effective QoS parameters and implementation environments with different fea- tures and effects should be recognized. Hence, and due to the ab- sence of related surveys, a systematic review on cloud computing service composition is necessary and will help facilitate future re- searches. A systematic review in which the most important aspects of the accomplished researchers must be investigated, and useful information and statistics must be extracted. This paper provides a systematic literature review on state-of- the art approaches and techniques in addressing cloud computing service composition. The discussed advancements and develop- ments of this topic provide useful information to motivate further investigations in this area. Identifying the different objectives of performing cloud computing service composition studies and the reasonable and purposeful classifications of such divergent ap- proaches and mechanisms is a major achievement of the review. Furthermore, this study extracts all of the considered QoS param- eters, introduces the most significant and the least considered parameters and calculates the importance parameter percentages, aiming to eliminate barriers to future research efforts, such as pro- posing comprehensive and reliable mathematical models for calcu- lating composite service QoS values. The goals of this research also include declaring the appropriate, investigated QoS datasets, utiliz- ing software in generating problems for evaluating methods and discussing the most widely used implementation environments. Because the investigated subject is very extensive, it is impossi- ble to include all relevant topics. Hence, some related subjects do not fall within the research scope of this review but will be briefly mentioned. Providing network services requires common lan- guages and protocols and is closely related to service composition. However, services that provide details will not be considered in this work. Strong and independent studies are required to address research methodologies and experimental and statistical perfor- mance evaluation strategies. Describing the accurate significance of the parameters for real cloud customers is also beyond the scope of this report but can be achieved by conducting comprehensive studies, utilizing interviews and questionnaires and adopting appropriate statistical methods. The structure of this paper is organized as follows. After the introduction, the main aims of the paper and research questions will be defined and described in Section 2. Briefly, cloud computing and its characteristics will be explained in Section 3. Cloud com- puting service composition (CCSC) will be defined in Section 4, including its challenges and classified applied approaches. In Sec- tion 5, an extensive discussion on the objectives of the investigated research, their approaches, and utilized datasets is provided, and effective information and statistics are extracted for future re- search. The final sections of the paper contain the conclusions and references. 2. Survey goals and execution Survey goals and research questions are described in Section 2.1, and the statistics on published and presented papers in different journals and conferences are presented in Section 2.2. The authors of the present paper have profited from ‘‘Guidelines for performing Systematic Literature Reviews in Software Engineering’’ (Kitchenham & Charters, 2007) and have also used (Garousi & Zhi, 2013) to conduct and perform this research. 2.1. Survey goals and research questions The present research aims at collecting and investigating all of the credible and effective studies that have examined CCSC. More specifically, the extraction of salient features and methods of pa- pers will be considered, and their characteristics will be described. To achieve the above-mentioned goals and identify the methods that have been selected by researchers for their studies and result assessment methods, case studies are covered for which new methods are proposed and datasets and benchmarks are used. Most researchers have considered QoS parameters and proposed objective functions and user trends that are important in designing these functions. The following research questions (RQs) are raised. RQ 1. What are the main goals of the researches? RQ 2. What is the proposed approach and what are the methods used? How have the researchers conducted the research? RQ 3. What datasets or benchmarks are used and what case studies are considered? RQ 4. What evaluating procedures have been used to assess the results in each paper? RQ 5. What other research has been considered in each paper to compare the results? RQ 6. What QoS parameters are accounted for? RQ 7. How can the user requirements and tendencies be consid- ered in the applied objective function? 2.2. Publication statistics In this study, attempts have been made to examine all of the pa- pers that have been published on CCSC in particular and that in- clude novel methods or interesting ideas. To achieve this goal and to answer the research questions in Section 2.1, 34 papers that were published from 2009 to December 2013 were selected from different high-level refereed journals and prestigious international conferences and are considered in Section 4. In each study, if it was required to be familiar with some concepts and methods and to read further on the topic, other books and papers are proposed and referred to. The result of this effort is a comprehensive collec- tion of resources that can provide an acceptable level of concepts and information about the service composition problem in cloud computing and the different views of addressing this problem that are introduced in the literature. 3. Cloud computing 3.1. Cloud definition There are different definitions for cloud computing in the liter- ature, many of which do not cover all of the features of the cloud. In one attempt, Vaquero et al. attempted to extract a comprehensive definition using 22 different compliments (Vaquero, Rodero- Merino, Caceres, & Lindner, 2008). Efforts have been made to stan- dardize the definition of the cloud, in which we accept the cloud definition provided by the National Institute of Standards and Technology (NIST) (Peter Mell, 2011). The NIST cloud computing definition: ‘‘Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 3811 servers, storage, applications, and services) that can be rapidly provi- sioned and released with minimal management effort or service pro- vider interaction. This cloud model is composed of five essential characteristics, three service models, and four deployment models’’. The cloud cannot be considered to be a new concept or technol- ogy that arose in recent years; instead, its root can be found in what John McCarthy described as the ability to provide computa- tional resources as a ‘‘utility’’ (Hamdaqa & Tahvildari, 2012). Based on standard material presented by NIST, cloud computing is com- posed of five main characteristics, and two other characteristics are added based on the literature, with three spanning service models and four models of deployment, which will be described in some detail in the following sections (see Fig. 1). 3.2. Cloud computing characteristics On-demand self-service. A user can request one or more services whenever he needs them and can pay using a ‘‘pay-and-go’’ method without having to interact with humans using an online control panel. Broad network access. Resources and services that are located in different vendor areas in the cloud can be available from an extensive range of locations and can be provisioned Fig. 1. Cloud computing, characteristics, deployment m through standard mechanisms by inharmonious thin or thick clients. The terms ‘‘easy-to-access standardized mechanisms’’ and ‘‘global reach capability’’ are also used to refer to this characteristic (Hamdaqa & Tahvildari, 2012; Yakimenko et al., 2009). Resource pooling. Providing a collection of resources simulates the behavior of a single blended resource (Wischik, Handley, & Braun, 2008). In other words, the user does not have knowl- edge and does not need to know about the location of the pro- vided resources. This approach helps vendors to provide several different real or virtual resources in the cloud in a dynamic manner. Rapid elasticity. Fundamentally, elasticity is another name for scalability; elasticity means the ability to scale up (or scale down) resources whenever required. Users can request different ser- vices and resources as much as they need at any time. This char- acteristic is so admirable that Amazon, as a well-known cloud service vendor, has named one of its most popular and commonly used services the Elastic Compute Cloud (EC2). Measured service. Different aspects of the cloud should automat- ically be controlled, monitored, optimized, and reported at sev- eral abstract levels for the resources of both the vendors and consumers. odels, service pool, and types of services and users. 3812 A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 Multi-Tenacity. This concept is the fifth cloud characteristic that is suggested by the Cloud Security Alliance. Multi-tenacity means that it is essential to have models for policy-driven enforcement, segmentation, isolation, governance, service lev- els, and chargeback/billing for different consumer categories (Espadas et al., 2013). Auditability and certifiability. It is important for services to pre- pare logs and trails to make it possible to evaluate the degree to which regulations and policies are observed (Hamdaqa & Tahvildari, 2012). 3.3. Cloud computing service models Definition 1 (Service). A service is a mechanism that is capable of providing one or more functionalities, which it is possible to use in compliance with provider-defined restrictions and rules and through an interface (Ellinger, 2013). Definition 2 (Platform). A platform is a fundamental computer system that includes hardware equipment, operating systems, and, in some cases, application development tools and user inter- faces on which applications can be deployed and executed. Definition 3 (Infrastructure). Infrastructure refers to underlying physical components that are required for a system to perform its functionalities. In information systems, these components can contain processors, storage, network equipment, and, in some cases, database management systems and operating systems. Software as a Service (SaaS). A software or application that is executing on a vendor’s infrastructure is recognized as a service provided that the consumer has limited permission to access; the provision is through a thin client (e.g., a web browser) or a program interface for sending data and receiving results. The consumer is unaware of the application provider’s infra- structure and has limited authority to configure some settings. Platform as a Service (PaaS). In this service model, the service vendor provides moderate basic requisites, including the oper- ating system, network, and servers, and development tools to allow the consumer to develop acquired applications or soft- ware and manage their configuration settings. Infrastructure as a Service (IaaS). The consumer has developed the required applications and needs only a basic infrastructure. In such cases, processors, networks, and storage can be provided by vendors as services with consumer provisions. 3.4. Cloud computing deployment models Public cloud. This approach is the major model of cloud comput- ing; here, the cloud owner provides public services in the vast majority of cases on the Internet based on predefined rules, pol- icies, and a pricing model. Possessing a large number of wide- spread world resources enables providers to offer a consumer different choices to select appropriate resources while consider- ing the QoS. Private cloud. A private cloud is designed and established to pre- pare most of the benefits of a public cloud exclusively for an organization or institute. Setting up such a system because of the utilization of corporate firewalls can lead to decreased security concerns. Because the organization that implements a private cloud is responsible for all of the affairs of the system, facing abundant costs is the blind spot of establishing a private cloud. Community cloud. Based on their similar requirements, con- cerns, and policies, a number of organizations establish a com- munity and share cloud computing to be used by their community member’s consumers. A third-party service pro- vider or a series of community members can be responsible for providing the required infrastructure of the cloud comput- ing. Lowering costs and dividing expenses between community members along with supporting high security are the most important advantages of a community cloud (Dillon, Chen, & Chang, 2010). Hybrid cloud. A combination of two or more different public, pri- vate, or community clouds led to the creation of a different cloud model called hybrid cloud, in which constitutive infra- structures not only keep their specific properties but also require standardized or agreed functionalities to enable them to communicate with each other with respect to interoperabil- ity and portability on applications and data. 4. The cloud computing service composition problem (CCSC) 4.1. CCSC definition Fast development in the utilization of cloud computing leads to publishing more cloud services on the worldwide service pool. Because of the presence of complex and diverse services, a single simple service cannot satisfy the existing functional requirements for many real-world cases. To complete a complex service, it is essential to have a batch of atomic simple services that work with each other. Therefore, there is a strong need to embed a service composition (SC) system in cloud computing. The process of service introduction, requesting, and binding, as shown in Fig. 2, can be accounted for in such a way that service providers introduce their available services to the broker to expose them to user requests. However, users also send their service re- quests to the broker, who must select the best service or set of ser- vices on the basis of user requirements and tendencies; the broker wants providers to bind selected services to the users with respect to predefined rules and agreements. Increasing the number of available services causes an increase in the number of similar-function services for different servers. These similar services are located in different places and have dis- tinct values in terms of the QoS parameters. For this reason, SC ap- plies appropriate techniques to select an atomic service among the different similar services that are located on distinct servers to al- low the highest QoS to be achieved according to the end-user requirements and priorities. Because of intrinsic changes in cloud environments, available services, and end-user requirements, SC should be designed dynamically, with automated function capabilities. Therefore, selecting appropriate and optimal simple services to be combined together to provide composite complex services is one of the most important problems in service composition. The SC problem in cloud computing can be defined as determining what atomic simple services should be selected such that the ob- tained complex composite service satisfies both the functional and QoS requirements based on the end-user requirements. Be- cause of various and abundant effective parameters and a large number of simple services provided by many service providers in the cloud pool, CCSC is considered an NP-hard problem (Fei, Dongming, Yefa, & Zude, 2008). In this paper, it is assumed that every composed service (CS) in the cloud consists of n unique services (USs) and has p QoS param- eters. To terminate a CS, a combination of unique services act sequentially in an ordinal workflow (wf), as shown in Fig. 3. Fig. 2. Process of service introduction, requesting, and binding. A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 3813 Define qi (USj) as the value of QoS parameter i for unique service j. Accordingly, the QoS vector of unique service j is defined as Eq. (1): qðUSjÞ¼ ðq1ðUSjÞ; q2ðUSjÞ; . . . ; qpðUSjÞÞ ð1Þ However, if wfk is the workflow of CSk, then it is possible to define Qi (wfk) to obtain the value of QoS parameter i for workflow k. Then, the QoS vector of workflow k is described in Eq. (2): QðwfkÞ¼ ðQ 1ðwfkÞ; Q 2ðwfkÞ; . . . ; Q pðwfkÞÞ ð2Þ 4.2. CCSC challenges The dynamic nature of cloud environments involves occasional and consciously planned changes; these changes expose cloud computing to different challenges in the SC. The most remarkable challenges are the following: Dynamically contracting service providers. The pricing policy of most service providers is determined by service fees based on supply and demand. Thus, mechanisms for updating the table of available resource characteristics must be predicted (Gutierrez-Garcia & Sim, 2013). Addressing incomplete cloud resources. Optimal service selec- tion by a broker depends on the availability of complete and updated information on the services. Facing several changes in the service characteristics could result in the loss of some data (Gutierrez-Garcia & Sim, 2013; Yi & Blake, 2010). Describing and measuring QoS attributes of network services. A lack of consensus on the definition and description of network services’ QoS parameters among worldwide cloud service Fig. 3. Cloud service co vendors is still an important challenge for cloud developers. The absence of agreed-upon ways to measure network QoS is another problem that is not completely solved and that should be considered (Qiang, Yuhong, & Vasilakos, 2012). Interservice dependency/conflict. Dependency or conflicts that exist among two or more services leads to a complicated service composition problem. In real-world scenarios, encountering dependency and conflict among services is quite common and should be considered in SC (Strunk, 2010). Security. Designing and apprising security rules, policies, and instructions are among the basic responsibilities of cloud ser- vice vendors. However, a principled self-administered frame- work for supplying and provisioning services in which vendors’ security concerns and policies are observed must be provided (Takabi, Joshi, & Gail-Joon, 2010; Zissis & Lekkas, 2012). 4.3. Current approaches for CCSC The different approaches proposed in the literature can be divided into five distinct categories: including classic and graph- based algorithms (CGBAs), combinatorial algorithms (CAs), ma- chine-based approaches (MBAs), structures (STs), and frameworks (FWs). Different studies that belong to these categories are investi- gated in Sections 4.3.1 to 4.3.5. 4.3.1. Classic and graph-based algorithms CCSC is a problem in which there are many potential solutions, among which one or a limited number of solutions are optimal. Thus, CCSC is known as an optimization problem (Anselmi, Arda- gna, & Cremonesi, 2007; Yu & Lin, 2005). Some types of classic algorithms, such as backtracking and branch-and-bound, can be used to solve optimization problems. These algorithms can guaran- tee that an optimal solution will be found solely by taking expo- nential time complexity (Neapolitan & Naimipour, 2009). Thus, using classic algorithms to solve optimization problems is possible only with some modifications and improvements for decreasing the execution time. To achieve a feasible concrete workflow for CCSC with respect to the consumer QoS requirement, the problem is considered to be equivalent to a multi-dimensional multi-choice knapsack problem (MMKP) in which a parameter called happiness that is calculated based on QoS parameters is used as the utility (Kofler et al., 2009). To solve the MMKP and obtain the benefits of heterologous multi-processing environments (e.g., grid computing (Preve, 2011)), a parallel form of the branch-and-bound algorithm is proposed in which each node of the decision tree is an independent mposition process. 3814 A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 instance of a main routine that could be assigned to separated computational nodes. To evaluate the algorithm, researchers ran- domly generated four different size problems; they executed the problems in serial and parallel modes and compared the results against one another. It is obvious that the most important problem that the proposed algorithm faces is the exponential complexity execution time. The performance evaluation of the algorithm could also be completed and could be more reliable if the results were compared with similar algorithms’ results on real-world datasets. To rectify the high execution time problem in the method, a two- phase approach is also proposed in Kofler, Haq, and Schikuta (2010). In this approach, the first phase is similar to the previous phase, whereas for second phase, the previous existing solutions are typically reused for similar requests, and some changes are ap- plied to correspond with the new situation. After designing a new cloud resource data-saving method, a matching algorithm called SMA is applied in Zeng et al. (2009) to check whether the output parameters of a service and the input parameters of another service are compatible with each other. In this algorithm, the service matching problem is mapped to calcu- late the semantic similarity of the different input and output parameters of different services. The matching degrees of each pair of services that are stored in a table lead to the composition of a weighted directed graph in which finding all reachable paths for two nodes can yield all of the possible service composition meth- ods. Researchers proposed an improved Fast-EP algorithm called FastB+�EP to obtain all possible paths in a shorter amount of time. However, two-step graph building and searching leads to increased execution time and decreasing algorithm performance, especially in cases when there is an increase in the size of the problem and the number of required services. In view of all of the cloud participants (e.g., resource vendors and service consumers) as agents and simulating their activities by a colored petri net (Barzegar, Davoudpour, Meybodi, Sadeghian, & Tirandazian, 2011; Jensen, 1995), an agent-based algorithm for CCSC is introduced (Gutierrez-Garcia & Sim, 2010). In the proposed algorithm, the broker agent receives a consumer agent’s require- ments and attempts to find a service vendor agent for each single service. Thereafter, the contract net protocol (Smith, 1981) is used to select proper single services to compose the solution. Using the experimental results, it is proven that the proposed method can achieve solutions in linear time with respect to the number of ser- vice vendor agents and that it can consider parallel activities to ad- dress heterogeneous services. A two-phase service composition approach is proposed for dy- namic cloud environments in which the service performance changes in different transactions (Shangguang, Zibin, Qibo, Hua, & Fangchun, 2011). In the first phase, the cloud model (Li, Cheung, Shi, & Ng, 1998) is applied to change the qualitative value of the QoS parameters to their quantitative equivalent to calculate the uncertainty level. Thereafter, mixed integer programming (MIP) is utilized in the second phase to find appropriate services in which the binary decision vector is used to determine whether a service is selected. Another two-phase method in which MIP is also applied to focus on service performance fluctuations in the dynamic cloud environment by solving CCSC is developed in Zhu, Li, Luo, and Zheng (2012). To decrease the number of candidate single services, some appropriate services are selected based on the history of sin- gle services and using K-means in the first phase of the proposed method. In the second phase, MIP is used to select the best single services among the preliminary selected services. The authors proved that their two-phase approach can outperform HireSome (Lin, Dou, Luo, & Jinjun, 2011). Applying linear programming (LP) (Korte & Vygen, 2012; Vanderbei, 2008) for optimizing virtual machine resource alloca- tion in physical servers for video surveillance was proposed in Hossain, Hassan, Al Qurishi, and Alghamdi (2012) for the first time. Composing an optimal set of media services to prepare a composite service over virtual machine resource allocation was also consid- ered. To reach this goal, the authors mapped the virtual machine resource allocation problem to the multi-dimensional bin-packing problem and used LP to solve it. They also suggested the use of a customized best-fit decreasing method (Kou & Markowsky, 1977) to solve the problem, which considerably increases the probability of finding appropriate results compared to LP but cannot guarantee that an optimal solution will be obtained. To evaluate two pro- posed methods, randomly generated service composition problems were considered, and the results are compared to fractional knap- sack mapping and round-robin allocation execution results. For generating composite services with the lowest execution cost a two-phase method is applied in Liu et al. (2012). The first phase includes utilizing a state transition matrix for the analysis of the dynamic process of composite service execution. In this phase, each composite service status was modeled on a state tran- sition diagram, which is used to produce a state transition matrix. The execution cost of each composite service can be calculated using its state transition matrix. In phase two, Business Process Execution Language for Web Services (BPEL4WS) (Huang, Lo, Chao, & Younas, 2006) is used to find an optimal solution. Because the process of BPEL4WS, which is known as the distributed traffic model, is extremely time consuming, researchers divided it into three parts, each of which is executed by an independent computer. To consider the user preferences and different resource charac- teristics bundled together, Liu et al. proposed a three-layer hierar- chical structure; the first layer includes optimal service selection; the second layer is called the criterion layer and includes timeli- ness, stability, and security, based on which services are divided into three categories; and the third layer is designed for the repre- sentation of nine additional QoS parameters (Liu et al., 2012). Next, the phases of the proposed algorithm (SSUP) include generating and normalizing a user requirement matrix, generating a QoS weight specification using the analytic hierarchy process (AHP) (Jeonghwan, Rothrock, McDermott, & Barnes, 2010), generating and normalizing the service attribute matrix, and performing the service-demand similarity calculation. The experimental results indicate that SSUP could outperform AHP when solving the CCSC problem. Worm et al. successfully addressing two antithetical aims for the same provider, namely, the aims of revenue maximization and quality assurance (Worm, Zivkovic, van den Berg, & van der Mei, 2012). The authors based their method on three decision cri- teria: service availability at the decision instant, executing cost, and resting time to deadline. With respect to the response time and with an emphasis on service availability, dynamic program- ming (DP) is used to achieve the main goals of the study, in which it is necessary to save intermediate results to select the best service among all available services. In addition, in the case of a deadline, an arrival decision rule would select services with the lowest price for the remaining tasks. The proposed algorithm is executed for static and dynamic service composition but uses real-world data- sets and benchmarks, comparing the results with the previous re- search; calculation time and memory complexity are neglected. Zhou and Mao proposed a cloud-based semantics approach for the composition of web services utilizing a Bayesian decision (Zhou & Mao, 2012). The authors applied a Bayesian approach to antici- pate the semantics of a web service for which a discovery graph is generated based on a cloud to use for implementation. They also obtained relations that are based on the graph that could encoun- ter a Markov chain (Song, Geng, Kusiak, & Xu, 2011). In addition, using an equation for the cosine theorem (Xiangkun & Yi, 2010), it is possible to achieve a similarity of services, which helps to A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 3815 identify the cloud service interface through Web services and de- velop a composition system. The proposed method has been ap- plied to Amazon services and can be compared with an existing approach (Stewart, 2009). On the basis of variability modeling (Sinnema & Deelstra, 2007), cloud feature models (CFMs) are presented as mechanisms to explain cloud services and user requirements together to pre- pare a suitable ground for their cloud service selection process (CSSP) (Wittern, Kuhlenkamp, & Menzel, 2012). A CFM is repre- sented by a directed graph in which the nodes play service roles and the edges denote the relations between services. Further- more, the CSSP should satisfy user objectives and also respect the requirements. To satisfy these goals, the user must enter the objectives and requirements to start the process, and the CSSP uses decision support tools to generate a list of offered services based on the input model. Focusing on the dynamic characteris- tics of cloud computing and updating the graph throughout the execution could help the proposed method to enhance its capabilities. A two-step procedure to satisfy users’ QoS requirements is proposed (Huang, Liu, Yu, Duan, & Tanaka, 2013) in which in the first step, for each user’s request, the proposed method at- tempts to select single services that meet the first two types of user functional requirements and that eliminate the remaining requirements. Then, a virtual network of service providers of se- lected services is generated and modeled by a directed acyclic graph (DAG). In the second step, in the case of a single user’s QoS requirement, it is sufficient to apply an algorithm to deter- mine the shortest path in the DAG. However, for the two QoS parameters, the problem is changed to a multi-constrained path problem (Korkmaz & Krunz, 2001) when this constructing auxil- iary graph FPTAS (CAGF) is used. In the CAGF, an auxiliary graph is used in which the two-weight preliminary DAG is changed to a one-weight DAG (which can be solved as in the previous case) by considering the first QoS parameter as a weight and merging the second parameter into the connectivity among the expanded vertices. Researchers have also proposed a method for consider- ing more than two QoS parameters. The execution time and re- turned path weight (the final QoS) are two parameters that are considered when making a comparison with existing DAG-based algorithms to determine the performance of the method; however, generating several graphs in the algorithm is a time- consuming activity that increases the execution time. This inter- esting study could also be enriched by addressing known datasets. To obtain the best set of single services for service composition, three algorithms are proposed (Qi & Bouguettaya, 2013) for gener- ating a service skyline that can be considered a specific set of ser- vice vendors that other possible sets cannot dominate with respect to the QoS parameters (Yu & Bouguettaya, 2010). The basic idea of these algorithms is based on reducing the search space. The first algorithm is called OPA, and it examines all of the service execution plans, one for each phase, and saves the best found solution. The researchers also applied some improvements on OPA for decreas- ing the processing time and the consumed space. DPA is a second algorithm; it uses a tree structure and is based on the progressive examination of service execution plans that are sorted in ascending order based on their score and progressive results output. This pro- gressive method provides the algorithm pipeline capabilities. To overcome the problem of the repetition of nodes and thus over- head, researchers have proposed reducing the parent table data structure. For a third attempt, a linear approach is used to design a bottom-up algorithm (BUA) to address the expensive computa- tional costs of DPA for an increasing number of services. Evaluating the proposed algorithms by using different models of problems is another advantage of this study. HireSome-II is proposed (Dou, Zhang, Liu, & Chen, 2013) to im- prove the reliability of composition QoS values. With HireSome-II, the QoS history of services is investigated for evaluation instead of what the service providers claim. Thus, the context of service com- position is implemented using a two-layer tree structure in which the required service is located in the root, and the nodes are lo- cated in the second layer as candidate services. In addition, the K-means clustering algorithm can be used to categorize the history of each candidate service into two peer groups. Inserting these two peer groups into the tree will provide a three-layer tree, the Task- Service-History Record tree, which identifies the best performance history for candidate services. Based on the required accuracy, additional history layers can be inserted into the tree. Based on experimental results, the performance of the proposed method proved to be superior to the authors’ previous method, HireSome-I, especially in the case of facing abundant candidate services. Despite its strengths, with increasing required services, service providers and history volume, this method will face significant increases in execution time due to the use of K-means for categorizing the history of all candidates. In (Wu, Zhang, Zheng, Lou, & Wei, 2013) a model for QoS-satis- fied predictions in CCSC is proposed based on the hidden Markov model (HMM) (Li, Fang, & Xia, 2014; Zeng, Duan, & Wu, 2010). The proposed model was to ensure customer satisfaction of the composite service QoS, utilizing the basic form of HMM, in which QoS parameters are considered as a state-set, and distinct observa- tion-probability functions are each defined for one of the parame- ters. Due to the high computational complexity of the functions, researchers also applied a two part, forward–backward structure for reducing the complexity, in which each part includes three stages, initialization, recursion and end. To evaluate the proposed model, a real cloud computing simulation system was constructed that provides more than 100 different services. The model param- eters can be adjusted with a support vector machine-based algo- rithm. The obtained evaluation error rates were small but with respect to the complicated structure of the model, which will face many services in the real world. This model is predicted to be con- fronted with larger error rates and an ineligible execution time. A novel method for QoS mapping is proposed in which a set of three rules is used to map QoS specifications and guarantees to- gether in the cloud (Karim, Chen, & Miri, 2013). To overcome the complexity, the analytic hierarchy process (AHP) (Ishizaka & Labib, 2011; Jeonghwan et al., 2010) is utilized in which customer prefer- ences, possible solutions and ranking the services, are conditions, search space and the goals, respectively. AHP has also led to specify QoS weights to rank and select candidate services. To evaluate the method, a case study was described, and the obtained results were discussed. Because the AHP was constructed from a graph that in- cluded many-to-many relationships, increasing the number of re- quired services and search space led to increased method time complexity, which is unacceptable in real environments. Two novel cloud service ranking approaches, CloudRank1 and CloudRank2, are proposed in which the basic strategy ranked the services based on the prediction of their QoS (Zibin, Xinmiao, Yilei, Lyu, & Jianmin, 2013). These two algorithms are composed of three steps, including similarity computation, preference value compu- tation and ranking prediction, and are different with respect to preference value treatment. Because of the negative effects of equal preference value treatment in the accuracy of ranking pre- diction in the first algorithm, the confidence value of service QoS preference values is defined and used in the treatment process of the second algorithm to remove the effects, reaching more accu- racy. Researchers have also provided a new dataset called tpds 2012 that includes a response-time and throughput of 500 services provided by 300 users to evaluate the proposed algorithms. The evaluation results indicate that the algorithms outperformed 3816 A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 previous approaches; however, based on mathematical consider- ations, the time complexity function of the algorithms was O(n2- m) + O(n2m) + O(n2), where n and m are the number of services and users, respectively. Because the number of services is typically much greater than the number of users, apparent increases in the execution time of the algorithms may occur. 4.3.2. Combinatorial algorithms CCSC is an optimization problem, and a distinctive feature of this problem is being involved in very large search spaces to obtain optimal results. In addition, the importance of achieving an opti- mal solution in a shorter time frame forces researchers to seriously consider the use of combinatorial algorithms (Knuth, 2011; Abonyi et al., 2013). Thus, different efforts have been conducted using dif- ferent models of combinatorial algorithms to solve the CCSC prob- lem, and these efforts are investigated in the following studies. To simplify the composition of services in cloud computing without priority weights, an algebra of communication process (ACP) (Fokkink, 2000) based on a QoS value aggregation algorithm is applied (Zhang & Dou, 2010) in which an artificial neural net- work (Haykin, 1998) is used to focus on service consumer require- ments without predefined user priority parameters. A three-layer perceptron (Shrme, 2011) neural network is used in this study by employing back-propagation (Wang, Wang, Zhang, & Guo, 2011) and least-mean-square-error (Sayed, 2003) algorithms to learn the network and adjust the weights, which are initialized ran- domly. In the proposed neural network, the number of neurons for the input layer is equal to the number of QoS parameters con- sidered. The number of neurons in the second layer is also equal to the number of neurons in the first layer, and the third layer has one neuron for an output. During the training process, the users’ prior- ity information is entered into the neural network to realize suit- able values for the weights. Weight parameter efficiency is evaluated by applying the mean square error (Zhou & Bovik, 2009) as the objective function. In this research, datasets and benchmarks are not investigated, and the results are not compared. By dividing the QoS parameters into three groups, including ascending, descending, and equal QoS attributes, and by using sim- ple additive weighting to normalize the values of those parame- ters, a new model for calculating the QoS of composite services (Ye, Zhou, & Bouguettaya, 2011) is proposed. These authors also applied a genetic algorithm to solve the CCSC problem in which a roulette wheel selection algorithm is used to select chromosomes to execute a crossover operation. The proposed model uses the achievement of QoS services as the fitness value. The results ob- tained with the proposed method were compared with different existing algorithms, such as LP and culture algorithm (Reynolds, 1999) and has shown to be more efficient. To achieve the goal of the representation of automatic service compositions for dynamic cloud environments, in Jungmann and Kleinjohann (2012) the problem is modeled as a Markov decision process (Arapostathis, Borkar, Fernández-Gaucherand, Ghosh, & Marcus, 1993; Chang & Yuan, 2009). A set of all possible composi- tions of services and a set of all possible actions for changing com- positions by adding new valid services are generated, and the composite service provider is considered an agent. A reward func- tion is also used to learn the optimal policy and optimal composi- tion for a user request that utilizes utility values, such as previous user feedback and former execution information. The reward func- tion results are used by the agent for subsequent decisions and up- dated using reinforcement learning techniques (Kaelbling, Littman, & Moore, 1996; Liu & Zeng, 2009). A judgment about this research cannot be made due to a lack of appropriate performance evalua- tion and comparison of results. Because of the growing importance of networks in the QoS of cloud service composition, in Klein, Ishikawa, and Honiden (2012) an approach is suggest for considering the non-network and network QoS of services separately. To this end, they estimated the real network latency among the desired services and their applicants with low time complexity by proposing a network mod- el that allowed services that had a lower latency to be selected. Researchers also introduced a QoS equation to calculate the net- work QoS, latency, and transfer rate. In the last phase of the ap- proach, based on the genetic algorithm, a selection algorithm is designed to apply proposed models to generate composite services, and its results are compared with the Dijkstra algorithm (Neapolitan & Naimipour, 2009) and random selection. The results of this interesting study could be enriched by the use of real-world datasets. With respect to self-adaptivity (Denaro, Pezze, & Tosi, 2007) in the service provider system, an improved genetic algorithm is pro- posed (Ludwig, 2012) in which a clonal selection algorithm (de Castro & Von Zuben, 2002) is used instead of tournament selection to select individuals for the crossover and mutation operators in a more discernible manner. The explanation of the proposed algo- rithm and the experimental results in the original paper do not go into detail regarding the researchers’ work on self-adaptivity. In Yang, Mi, and Sun (2012) game theory is used to propose a service level agreement (SLA)-based service composition algo- rithm. In this research, SLA is defined as quadruple, including SLA main information, service vendors and consumer information, ser- vice type and parameters, and a set of responsibilities for service vendors and consumers. To establish an SLA, SC is intended to be a multiple dynamic game, called a bid game, in which service ven- dors and service consumers are players that aim to achieve their goals. In this competitive game, every consumer should promul- gate a price for each requested service based on effective parame- ters and other consumers’ proposed prices, and vendors can assign their services according to received proposed prices with respect to the requested level of the quality of services that are agreed upon and signed in the SLA. The reliability of this method is limited due to the lack of comparison with other techniques and the lack of comparison with real-world datasets. A parallel form of the chaos optimization algorithm (Jiang, Kwong, Chen, & Ysim, 2012) called FC-PACO-RM is proposed to solve the CCSC problem (Fei, Yuanjun, Lida, & Lin, 2013). The researchers attempted to dynamically modify the sequence length based on the evolutionary state of the solutions. They also utilized the roulette wheel selection algorithm before executing the chaos operator to eliminate improper randomly generated solutions and escape from their destructive consequences. Because one of the main goals of this study has been to reduce the execution time, parallelization of the proposed algorithm is also considered. To accomplish this goal, a full-connection topology has been selected because of its high searching capability and message passing inter- face (MPI) (Barney, 2012). Finally, a novel migration method called Way-Reflex Migration is introduced and applied to reduce commu- nication overhead of fully connected topologies. Compared with a genetic, chaos genetic and chaos optimization algorithm, the pro- posed method showed better results in view of the best solution fitness and execution time. In Wang, Sun, Zou, and Yang (2013) particle swarm optimiza- tion (PSO) with integer array coding is applied to achieve a fast method to solve the CCSC problem. To achieve this goal, the skyline operator with binary decision variables is used to eliminate impro- per services from the search space. Because it is necessary to con- trast all of the services pair-wise when using the skyline operator, its execution time is not acceptable for an increasing number of services. Thus, the researchers used offline skyline services to de- crease the response time (Borzsony, Kossmann, & Stocker, 2001). The QWS dataset (Al-Masri & Mahmoud, 2008) and Synthetic Generator (Wang, Sun, & Yang, 2010) are used to evaluate the A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 3817 algorithms. Comparing the results of the proposed algorithm with another service composition called GOA (Ardagna & Pernici, 2007) demonstrates that the proposed method achieves positive results. Simultaneous optimization of the response time and execution cost of the service composition process motivated Jula et al. to pro- pose Imperialist competitive-gravitational attraction search algo- rithm (ICGAS) as a new memetic algorithm (Moscato & Cotta, 2003) for solving the CCSC problem (Jula, Othman, & Sundararajan, 2013). Because they wanted to address the enormous number of services that were provided by several service vendors, they at- tempted to apply the advantages of the exploration capability of an imperialist competitive algorithm (ICA) (Atashpaz-Gargari & Lu- cas, 2007) and the significant exploitation ability of a gravitation search algorithm (GSA) (Jula & Naseri, 2012) simultaneously. In the proposed algorithm, to balance the process of execution of the two algorithms and to increase the performance, the number of population members of GSA is determined to be equal to 20% of the number of members of the ICA algorithms. Researchers also introduced a new mathematical model for calculating the QoS of the nominated composite services based on user-defined weights for different QoS parameters. Using a very large real-world dataset to evaluate the algorithm and to compare its results with the re- sults from executing a genetic algorithm, PSO and classic ICA have demonstrated suitable results for ICGAS. A negative selection algorithm, which is inspired from a nega- tive selection mechanism in biological immune recognition and can eliminate improper solutions through an execution process, is used in Zhao, Wen, and Li (2013). In the proposed algorithm, the solutions and candidate services are implemented in the form of vector and integer strings, respectively. There are two different applied models for calculating the local and global fitness. To com- pute the fitness of the solutions for a local search, users are respon- sible for defining their criteria with related weights and a set of constraints based on which quality vectors are generated for all of the services for the decision-making process. Global fitness is also computed using an equation that is designed based on local fitness. The performance capability of the proposed algorithm is proven by comparing its results with the results obtained from standard particle swarm intelligence and the clonal selection im- mune algorithm. Attempting to select the best services based on overall quality of services a model is designed in which customer feedback and the objective performance analysis results are considered as two inputs, and the output is the quality of services (Lie, Yan, & Orgun, 2013). Because customer feedback is composed of linguistic vari- ables, a mapping table is applied to change the inputs to fuzzy numbers. Then, two input values that were changed to fuzzy rat- ings and a fuzzy simple additive weighting system (Chou, Chang, & Shen, 2008) are utilized to obtain the model output. Researchers have also applied a filtering mechanism for removing misleading values that are given by unprofessional customers. The method was evaluated with a case study; however, the results have not been compared to other approaches. This method has also not been applied to select a set of services for assessing its effectiveness in service composition. Thus, this method cannot present proper effi- ciency in the case of facing many QoS parameters, customer feed- backs and requests. 4.3.3. Machine-based methods In Bao and Dou (2012) researchers designed finite state ma- chines (FSMs) (Koshy, 2004) to consider service correlations and rightful task executing sequences; each FSM is used to implement a group of services that have limitations on their invocation and executing sequence called CWS community and another FSM for a desired composite service called the target process. The proposed method is composed of two phases. In the first phase, a composi- tion service tree is created on the basis of CWS community and a target process. A pruning policy is also applied to eliminate impro- per paths of the tree and to reduce the processing time. The QoS of all paths in the tree are calculated in the second phase based on user requirements, and the path with the highest QoS is selected as the final solution. To reach this goal, a dichotomous simple addi- tive weighting technique (Liangzhao et al., 2004) is used in the first part, from which the scaling of paths is performed by dividing them into negative and positive criteria, and in the second part, the total QoS of all of the paths is calculated. Researchers could prove appropriate efficiency of their proposed methods by compar- ing its executing time with the executing time of the enumeration method (Gerede, Hull, Ibarra, & Su, 2004). 4.3.4. Structures Based on the B+-tree (Bayer & Unterauer, 1977), an index struc- ture has been designed by in Sundareswaran, Squicciarini, and Lin (2012) to simplify the process of information insertion and retrie- val for cloud service providers. In the proposed structure, different properties, such as the service type, security, QoS, and measure- ment and pricing units, have specific locations to be stored and considered. To increase the speed at which the information man- agement operators are executed and appropriate vendor queries can be found, service vendors with the same characteristics should be stored together in adjacent rows. Researchers also proposed a query algorithm based on a designed structure to search the pro- viders’ database for the best vendors; after finding k vendors close to the optimal for vendors with each desired service, a refinement procedure is designed to reduce the number of selected vendors and sort them according to their Hamming distances, which entails starting with the optimal and progressing in ascending order to facilitate the selection of better providers. The proposed method is compared with a brute-force search algorithm and has shown al- most 100 times better execution speed for solving the CCSC prob- lem with 10,000 service providers. 4.3.5. Frameworks Pham, Jamjoom, Jordan, and Shae (2010) proposed a new frame- work for service composition in which a composition agent is responsible for receiving the request and providing service man- agement. The agent analyzes each request and divides it into the required single services. Using a knowledge base, the service dependencies are identified, and a service recovery section extracts similar service information. A composition is successful provided that all of the required single services are available and are used to update the knowledge base. There is a packaging engine that generates a new software package by using existing composition together with the new composition and that registers it in a service catalog. Finally, a service delivery section utilizes service catalog information for service provisioning. Chunqing, Shixing, Guopeng, Bu Sung, and Singhal (2012) pro- posed a systematic framework for automatic service conflict detec- tion and supplying policies and user requirements. The first phase of the proposed framework is a conflict analysis section that in- cludes two sub-sections called the comparator and analyzer. The comparator checks the conformity of the policies and user require- ments based on their priorities, and the analyzer is responsible for uncovering the contradictions between the user requirements and their affiliate relations while applying Satisfiability Modulo Theo- ries (SMT) (Moura & Bjørner, 2011). The filter, allocator, and solver are three parts of the second phase of the framework, which is called the solution derivation. This section finds appropriate single services and eliminates policy-violating services. It determines a set of appropriate single services for each user’s needs and finally concludes with the best composite service with respect to policies 3818 A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 and requirements. It uses the backtracking algorithm (Neapolitan & Naimipour, 2009) for the assigned tasks. According to the definition of trust as a conceptual probability by which a composite service is expected to execute a task as well as expected by the user, a trust-based method and a framework are proposed by Xiaona, Bixin, Rui, Cuicui, and Shanshan (2012) to solve the CCSC problem. Applying the trust in the process of service composition is divided into three parts, including trustworthy in service selection, guaranteeing trust in the composition processes, and trustworthy in binding the generated plan. In the designed framework, the requirement analyzer classifies the requirements of the user into their different elements, including functional and non-functional requirements and expected input–output parame- ters. The service retriever restores information on the services from the resource pool using query requests. Inappropriate services are eliminated from the candidate list by a service filter, and the name and type of the remaining services are identified by an WSDL ana- lyzer. The clustering component, template generator, and binding optimizer are responsible for checking the services’ composability, checking the math interfaces, and evaluating the binding plan trust, respectively. CloudRecommender is a cloud-based three-layer structured service composition system that was proposed by Zhang, Ranjan, Nepal, Menzel, and Haller (2012). The first layer is a configuration management layer in which a cloud service ontology and cloud QoS ontology are located because of its two parts for uncovering services based on their functionality and QoS parameters; the ser- vices are mapped to a rational model and data structure. Applica- tion logic is the second layer, which is implemented to select single services in the form of SQL queries to include criteria, views, and stored procedures. The third layer is a widget that divides the user Table 1 Desired objectives in the investigated researches. Reference Approach RO1 RO2 Kofler et al. (2009) CGBA p Zeng et al. (2009) CGBA Gutierrez-Garcia and Sim (2010) CGBA Zhang and Dou (2010) CA p Pham et al. (2010) FW Wang et al. (2011) CGBA p Ye et al. (2011) CA Zhu et al. (2012) CGBA Hossain et al. (2012) CGBA Liu et al. (2012) CGBA p Liu et al. (2012) CGBA Worm et al. (2012) CGBA Zhou and Mao (2012) CGBA Wittern et al. (2012) CGBA p Jungmann and Kleinjohann (2012) CA p Ludwig (2012) CA Yang et al. (2012) CA Bao and Dou (2012) MBM p Sundareswaran et al. (2012) ST Chunqing et al. (2012) FW p Xiaona et al. (2012) FW p Zhang et al. (2012) FW Huang et al. (2013) CGBA p Qi and Bouguettaya (2013) CGBA Wang et al. (2013) CA Jula et al. (2013) CA p Zhao et al. (2013) CA p Dou et al. (2013) CGBA Wu et al. (2013) CGBA p Karim et al. (2013) CGBA p Zibin et al. (2013) CGBA Wu et al. (2013) FW p Lie et al. (2013) CA p p Fei Tao et al. (2013) CA CGBA, classic and graph-based algorithms; CA, combinatorial algorithms; MBM, machin interface into four objects, including the computing resources, storage resources, network resources, and recommendation. This layer is implemented using RESTful (Richardson, 2007) and several JavaScript frameworks. A novel framework is proposed for adaptive service selection in mobile cloud computing (Wu et al., 2013). The framework extracts the QoS preferences of the customer immediately after receiving a request. Then, based on the Euclidean distance, some of the nearest customer preference services are selected and suggested to the ser- vice adapter. Finally, the service adapter selects the best service among the suggested services to be assigned to the customer, with respect to device context matching and effectiveness of the service option. To reach the context matching service based on the input information, a fuzzy cognitive map model is also utilized in the ser- vice adapter module. The weaknesses of this method include that the proposed framework can only be used to select a single service; furthermore, this strategy has not been compared to other approaches. 5. Discussion 5.1. Objectives of the researches To respond to the first research question RQ1 and achieve a comprehensive view of the topic, it is essential to categorize the goals of the researches. Objective scrutiny of the considered papers guarantees the existence of nine categories, RO1 to RO9, in which each paper can be placed in one or more categories, as described in Table 1 and Fig. 4(a) and (b). Based on Table 1, the largest amount of researchers’ attention has been focused on RO3, and there is a large difference in terms of attention paid in the RO3 RO4 RO5 RO6 RO7 RO8 RO9 p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p e-based methods; ST, structures; FW, Frameworks. A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 3819 literature between RO3 and RO1. This difference may be because decreasing the time complexity of the algorithms is a priority of researchers, and user satisfaction is the second-most important objective for researchers. The objective categories of the research are defined as follows: 1. User requirement satisfaction (RO1). Strictly abiding by cus- tomers’ requirements and providing facilities for them to deter- mine or describe their needs would make them more satisfied. 2. Qualitative to quantitative transformation of QoS parameters (RO2). When accurately considering the qualitative QoS param- eters and applying them in decision making, it is important to transform them into quantitative values using reliable methods. 3. Algorithm improvements (RO3). There are many different heu- ristic and non-heuristic algorithms that are introduced by algo- rithm designers and applied to solve NP-hard problems. The following efforts to improve the current algorithms and specify them for CCSC to obtain the best solutions or reduce the execu- tion time is one of the more often investigated objectives in the literature. 4. Introducing data storage and indexing structures (RO4). Pos- sessing appropriate and well-defined data structures and dat- abases can play an important role in the design of an efficient algorithm. Using a suitable indexing method is also useful in increasing the search speed, especially when the number of cases is very high. 5. Self-adaptability, automaticity, increasing reliability and accu- racy, and quality assurance (RO5). Establishing automated and self-adaptable service brokers is unavoidable (Denaro et al., 2007) because of increasing complexity, the number of requests, the number of available services in the pool, their diversity, and the limitation of human abilities. One of the main factors that attracts customers and retains them in utilizing cloud computing is reliability. Because of the importance of providing a reliable and self-adaptable service composition organization, the most important part of a cloud is in its direct contact with customers; researchers must consider this direct contact more seriously than before. 6. Proposing an improvised QoS mathematical model (RO6). Calculating a QoS value for composite services requires a math- ematical model in which all aspects, parameters, user require- ments, and tendencies are investigated. To reach this goal, some researchers have attempted to present improved models that focus more on these objectives. Fig. 4. (a) Importance percentage of objective categories a 7. Revenue maximization (RO7). Encouraging service providers to expose their high-quality services depends on the ability to amass significant profits. Thus, revenue maximization can be noted as an important fundamental aim. 8. Optimization of the service discovery process (RO8). If the ser- vice composer policy is not to register services based on prede- fined requirements, then it must discover required available simple services in the network. It is critical to have the type of policy that uses optimal discovery methods. 9. Proposing new frameworks and structures (RO9). Reaching some basic goals, e.g., the definition of new roles, requires changes in the framework of the CCSC and in the structures. In some cases, to achieve a goal, there could be a need to design a new frame- work. This scenario has led some researchers to be encouraged to design new frameworks or to change the existing frameworks. 5.2. Applied approaches To address RQ2, all of the proposed or applied approaches are di- vided into five distinct categories, as mentioned in Section 4.3. These categories and their statistics are summarized in Table 2 and Fig. 5(a) and (b), from which it can be inferred that the highest and lowest percentages of usage can be seen in the classic and graph-based algo- rithms (52%) and combinatorial algorithms (24%), respectively. 5.3. Investigated datasets The obtained answers for the third question RQ3 are not prom- ising. Possessing different datasets each of them can support sev- eral QoS parameters, and the predefined composition problem are useful and unavoidable for evaluating the proposed approaches and comparing their results. Unfortunately, the number of datasets that are available to all and in the research domain is very low and is limited to three datasets, QWS (Al-Masri & Mahmoud, 2009), WS-DREAM (Zibin, Yilei, & Lyu, 2010) and tpds 2012 (Zibin et al., 2013), and an unknown randomly generated dataset RG (Shanggu- ang et al., 2011). Researchers have also used a synthetic generator, rarely. The datasets used in each study are listed in Table 2. 5.4. Significance of QoS parameters and their percentage Based on the literature review presented in this paper, the pri- ority and importance percentage of the QoS parameters and their nd (b) number of repetitions of objective categories. 3820 A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 relevant factors have not yet been studied comprehensively. Thus, an analysis of the considered QoS parameters in the observed pa- pers and their statistics is essential. In this section, we attempted to extract the importance of different QoS parameters and their percentage considering this field of research. As specified in Table 2 and to answer RQ6, according to the sig- nificance and priority, most researchers have accounted for differ- ent QoS parameters, although others have neglected them. Based on the abovementioned parameters and their frequency of occur- rence in the literature, it is important to obtain the most important and effective QoS parameters and their importance percentage. To reach this goal, using Eq. (3), the number of occurrences of a parameter has been counted separately and divided by the sum of the number of occurrences of all parameters. The importance per- centage of each parameter is obtained by multiplying its calculated value by 100. The result obtained is the ratio of the importance percentage of every parameter to all of the other parameters. The number of occurrences of all of the parameters and their impor- tance percentages are shown in Fig. 6(a) and (b). imp percentageðiÞ¼ occurr noðiÞ Pparam no j¼1 occurr noðjÞ ð3Þ Table 2 Utilized tools, datasets and QoS parameters. Reference Tools Dataset QoS c Kofler et al. (2009) Kepler Workflow tool, CORBA, C++ – Resp Zeng et al. (2009) Visual C++ 6, PostgreSQL 8.4. – Resp Gutierrez-Garcia and Sim (2010) Not mentioned – Not m Zhang and Dou (2010) Not mentioned – Not m Pham et al. (2010) Not mentioned – Not m Wang et al. (2011) Matlab 7.6 WS-DREAM, RG Resp Relia Ye et al. (2011) Not mentioned – Resp Zhu et al. (2012) Visual C#.NET – Not m Hossain et al. (2012) Not mentioned – Resp Liu et al. (2012) Not mentioned – Cost Liu et al. (2012) Not mentioned WS-DREAM Not m Worm et al. (2012) Not mentioned – Resp Zhou and Mao (2012) Not mentioned – Not m Wittern et al. (2012) Researchers built a tool based on eclipse modeling framework – Not m Jungmann and Kleinjohann (2012) Not mentioned – Not m Ludwig (2012) Java – Resp Yang et al. (2012) Not mentioned – Not m Bao and Dou (2012) Not mentioned – Resp Sundareswaran et al. (2012) C – Not m Chunqing et al. (2012) Java, Cauldron, zChaff solver – Cost Xiaona et al. (2012) Not mentioned Seekda Avail Zhang et al. (2012) JavaScript, RESTful – Not m Huang et al. (2013) Not mentioned – Not m Qi and Bouguettaya (2013) Java Using Synthetic Generator Resp Wang et al. (2013) Matlab 7.6, Lp-Solve 5.5 QWS, Synthetic Generator Not m Jula et al. (2013) Visual C#.NET WS-DREAM Resp Zhao et al. (2013) Not mentioned – Resp Dou et al. (2013) Not mentioned – Cost, Wu et al. (2013) Not mentioned – Not m Karim et al. (2013) Not mentioned – Cost, Relia Zibin et al. (2013) Planet-lab, Axis2 tpds 2012 Resp Wu et al. (2013) Not mentioned – Not m Lie et al. (2013) Not mentioned – Not m Fei Tao et al. (2013) Not mentioned – Resp Func where imp_percentage(i) is the importance percentage of QoS parameter i, occur_no(i) is the number of repetitions of QoS param- eter i in the investigated papers, and param_no is the number of ob- served QoS parameters in the literature. There could be another way to look at the importance percent- age of the QoS parameters. It is possible to consider this issue in view of the ratio of the frequency of occurrences of the parameters in the papers. To increase the total number of papers, it is possible to account for all of the papers, including those that did not men- tion any parameters or excluded them. To reach the results, it is sufficient to divide the number of occurrences of each parameter by the total number of papers. The obtained results are shown in Fig. 7(a) and (b) when including and excluding papers that do not consider the QoS parameters, respectively. 6. Conclusion and future works Showcasing pertinent achievements and findings with a com- prehensive review generally increases the research efforts and pa- pers in that particular scientific field. Proposing novel techniques onsidered parameters onse time (RT), Cost onse time (RT), Availability (Avail), Reliability (Reli) entioned entioned entioned onse time (RT), Cost, Throughput (Throu), Reputation (Reput), Availability (Avail), bility (Reli) onse time (RT), Cost, Availability (Avail), Reputation (Reput) entioned onse time (RT), Cost entioned onse time (RT), Cost, Availability (Avail) entioned entioned entioned onse time (RT), Cost, Availability (Avail), Reliability (Reli) entioned onse time (RT), Cost, Reliability (Reli), Availability (Avail), Reputation (Reput) entioned ability (Avail), Durability (Dur) entioned entioned onse time (RT), Cost entioned onse time (RT), Cost onse time (RT), Cost, Availability (Avail), Reliability (Reli) Latency, Reputation (Reput), entioned Response Time (RT), Security (Secur), Reputation (Reput), Availability (Avail), bility (Reli), Durability (Dur), Data Control (DC) onse Time (RT), Throughput (Thr) entioned entioned onse time (RT), Cost, Reliability Reli, Energy, Trust, Maintainability (Maint), tion similarity (FS) Fig. 5. (a) Approach categories and (b) number of papers in each category. Fig. 6. (a) Percentage of repetition of QoS parameters and (b) number of repetitions of QoS parameters. Fig. 7. (a) Percentage of repetition of QoS parameters in all investigated papers and (b) percentage of repetition of QoS parameters in all investigated papers, excluding those that did not mention any parameters. A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 3821 and approaches for addressing cloud computing service composi- tion from different aspects in addition to addressing the lack of comparable activities were the strong motivating factors for pre- paring this systematic literature review. This paper has provided a complete definition of the CCSC in combination with its associ- ated concepts and a comprehensive analysis of the different ap- plied algorithms, mechanisms, frameworks and techniques extracted from 34 authentic published papers, spanning 3822 A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 2009–2013. The achievements of this review shed light on the re- search grounds of CCSC for future studies. This investigation demonstrates that all cloud computing ser- vice composition innovations and improvements can be catego- rized into the following five groups: classic and graph-based algorithms, combinatorial algorithms, machine-based methods, structures and frameworks. The most widely applied category is the classic and graph-based algorithms group (52%), and the least- used categories are machine-based methods and structures (3% and 4%, respectively). The objectives of these reports can also be di- vided into 9 categories, among which algorithm improvements (RO3) and user requirement satisfaction (RO1) have attracted the most attention (36% and 25%, respectively), while revenue maximi- zation (RO7) has been the least important objective (2%). Counting the number of QoS parameter occurrences indicated that 14 different parameters have been considered in the literature, among which service cost and response time are the most repeated ones (24% and 22%, respectively). Calculating the importance per- centage of the QoS parameters also revealed that the importance percentages of two mentioned parameters after including and excluding papers that do not consider the QoS parameters were 44% and 41% and 88% and 82%, respectively. To evaluate the pro- posed approaches, 4 QoS datasets have been previously identified, WS-DREAMS2, QWS, tpds2012 and RG, the first three of which are real-world extracted, and the last of which is generated randomly. Synthetic generators and Seedka are also two applications that have applied for runtime generating QoS values. With respect to the findings in this paper, achieving certain goals is of utmost importance for the planning of future work. Aim- ing to prepare an identical, competitive environment for compar- ing the proposed algorithms and approaches, it is indispensable to provide a set of differently sized, standard problems and a com- prehensive QoS dataset. The dataset should include a great number of unique services and service providers and encompass cost, re- sponse time, availability, reliability and reputation as significant QoS parameters. Another essential research goal is to focus on designing comprehensive mathematical models for calculating the QoS values composite services that cover all of the involved parameters and their importance percentage. Proposing real-time algorithms that can obtain a few optimal composite services for the given requests represent a significant achievement in the field. Furthermore, the less considered objectives (e.g., RO2, RO6, RO7, and RO8) should be regarded. Finally, acknowledging the stunning growth in mobile computing, future research efforts should be di- rected towards this forward-looking area. References Ai, L. F., & Tang, M. L. (2008). A penalty-based genetic algorithm for QoS-aware web service composition with inter-service dependencies and conflicts. New York: IEEE. Al-Masri, E., & Mahmoud, Q. H. (2008). Investigating web services on the world wide web. In Proceedings of the 17th international conference on world wide web (pp. 795–804). Beijing, China: ACM. Al-Masri, E., & Mahmoud, Q. H. (2009). Discovering the best web service: A neural network-based solution. In systems, man and cybernetics, 2009. SMC 2009. IEEE international conference on (pp. 4250–4255). Anselmi, J., Ardagna, D., & Cremonesi, P. (2007). A QoS-based selection approach of autonomic grid services. In Proceedings of the 2007 workshop on service-oriented computing performance. Aspects, issues, and approaches (pp. 1–8). Monterey, California, USA: ACM. Arapostathis, A., Borkar, V. S., Fernández-Gaucherand, E., Ghosh, M. K., & Marcus, S. I. (1993). Discrete-time controlled Markov processes with average cost criterion: a survey. SIAM Journal on Control and Optimization, 31, 282–344. Ardagna, D., & Pernici, B. (2007). Adaptive service composition in flexible processes. IEEE Transactions on Software Engineering, 33, 369–384. Atashpaz-Gargari, E., & Lucas, C. (2007). Imperialist competitive algorithm: An algorithm for optimization inspired by imperialistic competition. In Evolutionary Computation, 2007. CEC 2007. IEEE Congress on (pp. 4661–4667). Bao, H. H., & Dou, W. C. (2012). A QoS-aware service selection method for cloud service composition. In 2012 IEEE 26th international parallel and distributed processing symposium workshops & Phd Forum (pp. 2254–2261). New York: IEEE. Barney, B. (2012). Message Passing Interface (MPI). In Lawrence Livermore National Laboratory (Vol. 2013). Barzegar, S., Davoudpour, M., Meybodi, M. R., Sadeghian, A., & Tirandazian, M. (2011). Formalized learning automata with adaptive fuzzy coloured petri net; an application specific to managing traffic signals. Scientia Iranica, 18, 554–565. Bayer, R., & Unterauer, K. (1977). Prefix B-trees. ACM Transaction on Database Systems, 2, 11–26. Borzsony, S., Kossmann, D., & Stocker, K. (2001). The Skyline operator. In Data Engineering, 2001. Proceedings. 17th international conference on (pp. 421–430). Canfora, G., Di Penta, M., Esposito, R., & Villani, M. L. (2005). An approach for QoS- aware service composition based on genetic algorithms. New York: Assoc Computing Machinery. Chang, W.-L., & Yuan, S.-T. (2009). A Markov-based collaborative pricing system for information goods bundling. Expert Systems with Applications, 36, 1660–1674. Chen, L., Li, M. L., & Cao, J. (2006). ECA rule-based workflow modeling and implementation for service composition. IEICE Transactions on Information and Systems, E89D, 624–630. Chou, S.-Y., Chang, Y.-H., & Shen, C.-Y. (2008). A fuzzy simple additive weighting system under group decision-making for facility location selection with objective/subjective attributes. European Journal of Operational Research, 189, 132–145. Chunqing, C., Shixing, Y., Guopeng, Z., Bu Sung, L., & Singhal, S. (2012). A Systematic Framework Enabling Automatic Conflict Detection and Explanation in Cloud Service Selection for Enterprises. In Cloud Computing (CLOUD), 2012 IEEE 5th International Conference on (pp. 883–890). de Castro, L. N., & Von Zuben, F. J. (2002). Learning and optimization using the clonal selection principle. IEEE Transactions on Evolutionary Computation, 6, 239–251. Denaro, G., Pezze, M., & Tosi, D. (2007). Designing self-adaptive service-oriented applications. In Autonomic Computing, 2007. ICAC ‘07. Fourth International Conference on (pp. 16–16). Dillon, T., Chen, W., & Chang, E. (2010). Cloud computing: issues and challenges. In Advanced Information Networking and Applications (AINA), 2010 24th IEEE International Conference on (pp. 27–33). Dou, W., Zhang, X., Liu, J., & Chen, J. (2013). HireSome-II: Towards privacy-aware cross-cloud service composition for big data applications. IEEE Transactions on Parallel and Distributed Systems, 1. Ellinger, R. S. (2013). Governance in SOA Patterns (white paper). In The Northrop Grumman Corporation for Consideration by OASIS SOA Reference Architecture Team (pp. 1–11). Espadas, J., Molina, A., Jiménez, G., Molina, M., Ramírez, R., & Concha, D. (2013). A tenant-based resource allocation model for scaling software-as-a-service applications over cloud computing infrastructures. Future Generation Computer Systems, 29, 273–286. Fei, T., Dongming, Z., Yefa, H., & Zude, Z. (2008). Resource service composition and its optimal-selection based on particle swarm optimization in manufacturing grid system. IEEE Transactions on Industrial Informatics, 4, 315–327. Fei, T., Yuanjun, L., Lida, X., & Lin, Z. (2013). FC-PACO-RM: a parallel method for service composition optimal-selection in cloud manufacturing system. IEEE Transactions on Industrial Informatics, 9, 2023–2033. Fokkink, W. (2000). Introduction to Process Algebra. Springer-Verlag New York, Inc.. Gabrel, V., Manouvrier, M., Megdiche, I., & Murat, C.IEEE. (2012). A new 0-1 linear program for QoS and transactional-aware web service composition. New York: IEEE. Gao, A. Q., Yang, D. Q., Tang, S. W., Zhang, M., & Society, I. C. (2005). Web service composition using integer programming-based models. Los Alamitos: IEEE Computer Soc. Garousi, V., & Zhi, J. (2013). A survey of software testing practices in Canada. Journal of Systems and Software, 86, 1354–1376. Gekas, J., & Fasli, M. (2005). Automatic Web service composition based on graph network analysis metrics. In R. Meersman, Z. Tari, M. S. Hacid, J. Mylopoulos, B. Pernici, O. Babaoglu, H. A. Jacobsen, J. Loyall, M. Kifer, & S. Spaccapietra (Eds.). On the Move to Meaningful Internet Systems 2005: Coopis, Doa, and Odbase, Pt 2, Proceedings (3761, pp. 1571–1587). Berlin: Springer-Verlag Berlin. Gerede, E., Hull, R., Ibarra, O. H., & Su, J. (2004). Automated composition of e- services: lookaheads. In Proceedings of the 2nd International Conference on Service Oriented Computing (pp. 252–262). New York, NY, USA: ACM. Gutierrez-Garcia, J. O., & Sim, K. (2013). Agent-based cloud service composition. Applied Intelligence, 38, 436–464. Gutierrez-Garcia, J. O., & Sim, K. M. (2010). Agent-based service composition in cloud computing. In T. H. Kim, S. S. Yau, O. Gervasi, B. H. Kang, A. Stoica, & D. Slezak (Eds.). Grid and Distributed Computing, Control and Automation (121, pp. 1–10). Berlin: Springer-Verlag Berlin. Hamdaqa, M., & Tahvildari, L. (2012). Cloud Computing Uncovered: A Research Landscape. In H. Ali & M. Atif (Eds.), Advances in Computers (86, pp. 41–85). Elsevier. Haykin, S. (1998). Neural Networks: A Comprehensive Foundation. Prentice Hall PTR. He, Q., Yan, J., Jin, H., & Yang, Y. (2008). Adaptation of web service composition based on workflow patterns. In A. Bouguettaya, I. Krueger, & T. Margaria (Eds.). Service-Oriented Computing – ICSOC 2008, Proceedings (5364, pp. 22–37). Berlin: Springer-Verlag Berlin. Hossain, M. S., Hassan, M. M., Al Qurishi, M., & Alghamdi, A. (2012). Resource Allocation for Service Composition in Cloud-based Video Surveillance Platform. New York: IEEE. Huang, C.-L., Lo, C.-C., Chao, K.-M., & Younas, M. (2006). Reaching consensus: a moderated fuzzy web services discovery method. Information and Software Technology, 48, 410–423. http://refhub.elsevier.com/S0957-4174(13)00992-5/h0005 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0005 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0010 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0010 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0010 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0655 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0655 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0655 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0655 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0660 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0660 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0660 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0030 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0030 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0040 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0040 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0040 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0050 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0050 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0050 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0055 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0055 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0065 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0065 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0065 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0070 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0070 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0075 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0075 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0075 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0090 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0090 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0665 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0665 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0665 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0115 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0115 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0115 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0115 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0120 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0120 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0120 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0125 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0125 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0125 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0670 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0135 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0135 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0135 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0140 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0140 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0140 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0145 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0145 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0675 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0675 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0675 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0675 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0675 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0155 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0155 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0155 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0160 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0160 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0165 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0165 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0165 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0165 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0790 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0790 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0790 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0175 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0180 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0180 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0180 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0180 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0185 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0185 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0185 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0190 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0190 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0190 A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 3823 Huang, J., Liu, Y., Yu, R., Duan, Q., & Tanaka, Y. (2013). Modeling and algorithms for QoS-aware service composition in virtualization-based cloud computing. IEICE Transactions on Communications, E96.B, 10–19. Ishizaka, A., & Labib, A. (2011). Review of the main developments in the analytic hierarchy process. Expert Systems with Applications, 38, 14336–14345. Jensen, K. (1995). Coloured Petri Nets: Basic Concepts, Analysis Methods and Practical Use (2). Springer-Verlag. Jeonghwan, J., Rothrock, L., McDermott, P. L., & Barnes, M. (2010). Using the analytic hierarchy process to examine judgment consistency in a complex multiattribute task. IEEE Transactions on Systems, Man, and Cybernetics Part A: Systems and Humans, 40, 1105–1115. Jiang, H., Kwong, C. K., Chen, Z., & Ysim, Y. C. (2012). Chaos particle swarm optimization and T-S fuzzy modeling approaches to constrained predictive control. Expert Systems with Applications, 39, 194–201. Jula, A., & Naseri, N. K. (2012). A hybrid genetic algorithm-gravitational attraction search algorithm (HYGAGA) to solve grid task scheduling problem. In In International Conference on Soft Computing and its Applications (ICSCA’2012) (pp. 158–162). Planetary Scientific Research Center (PSRC). Jula, A., Othman, Z., & Sundararajan, E. (2013). A hybrid imperialist competitive- gravitational attraction search algorithm to optimize cloud service composition. In Memetic Computing (MC), 2013 IEEE Workshop on (pp. 37–43). Jungmann, A., & Kleinjohann, B. (2012). Towards the Application of Reinforcement Learning Techniques for Quality-Based Service Selection in Automated Service Composition. In Services Computing (SCC), 2012 IEEE Ninth International Conference on (pp. 701–702). Kaelbling, L. P., Littman, M. L., & Moore, A. W. (1996). Reinforcement learning: a survey. Journal of Artificial Intelligence Research, 4, 237–285. Karim, R., Chen, D., & Miri, A. (2013). An end-to-end QoS mapping approach for cloud service selection. In Services (SERVICES), 203 IEEE Ninth World Congress on (pp. 341–348). Kitchenham, B., & Charters, S. (2007). Guidelines for performing Systematic Literature Reviews in Software Engineering. (2.3 ed., pp. 1–65). Keele University and Durham University. Klein, A., Ishikawa, F., & Honiden, S. (2012). Towards network-aware service composition in the cloud. In the 21st international conference on World Wide Web (pp. 959-968). Lyon, France: ACM. Knuth, D. E. (2011). The Art of Computer Programming. Volume. 4A: Combinatorial Algorithms, Part 1, Pearson Education India. Kofler, K., Haq, I. U., & Schikuta, E. (2010). User-centric, heuristic optimization of service composition in clouds. LNCS, 6271, 405–417. Kofler, K., ul Haq, I., & Schikuta, E. (2009). A parallel branch and bound algorithm for workflow QoS optimization. In Parallel Processing, 2009. ICPP ‘09. International Conference on (pp. 478–485). Korkmaz, T., & Krunz, M. (2001). Multi-constrained optimal path selection. In INFOCOM 2001. Twentieth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings (2, pp. 834–843). IEEE. Korte, B., & Vygen, J. (2012). Linear Programming (21). Berlin Heidelberg: Combinatorial Optimization Springer, pp. 51-71. Koshy, T. (2004). Chapter 11 – formal languages and finite-state machines. In Discrete Mathematics With Applications (pp. 733–802). Burlington: Academic Press. Kosuga, M., Kirimoto, N., Yamazaki, T., Nakanishi, T., Masuzaki, M., & Hasuike, K. (2002). A multimedia service composition scheme for ubiquitous networks. Journal of Network and Computer Applications, 25, 279–293. Kou, L. T., & Markowsky, G. (1977). Multidimensional bin packing algorithms. IBM Journal of Research and Development, 21, 443–448. Abonyi, J., Akerkar, R., Alavi, A. H., Arango, C., Aydogdu, I., Brest, J., et al. (2013). List of Contributors. In X.-S. Yang, Z. Cui, R. Xiao, A. H. Gandomi & M. Karamanoglu (Eds.), Swarm Intelligence and Bio-inspired Computation (pp. xv-xviii). Oxford: Elsevier. Li, D., Cheung, D., Shi, X., & Ng, V. (1998). Uncertainty reasoning based on cloud models in controllers. Computers & Mathematics with Applications, 35, 99–123. Li, Z., Fang, H., & Xia, L. (2014). Increasing mapping based hidden Markov model for dynamic process monitoring and diagnosis. Expert Systems with Applications, 41, 744–751. Liangzhao, Z., Benatallah, B., Ngu, A. H. H., Dumas, M., Kalagnanam, J., & Chang, H. (2004). QoS-aware middleware for web services composition. IEEE Transactions on Software Engineering, 30, 311–327. Liao, J. X., Liu, Y., Zhu, X. M., Xu, T., & Wang, J. Y.IEEE. (2011). Niching particle swarm optimization algorithm for service composition. In 2011 IEEE Global Telecommunications Conference. New York: IEEE. Lie, Q., Yan, W., & Orgun, M. A. (2013). Cloud service selection based on the aggregation of user feedback and quantitative performance assessment. In Services Computing (SCC), 2013 IEEE International Conference on (pp. 152–159). Lin, W., Dou, W., Luo, X., & Jinjun, C. (2011). A history record-based service optimization method for QoS-aware service composition. In Web Services (ICWS), 2011 IEEE International Conference on (pp. 666–673). Liu, F., & Zeng, G. (2009). Study of genetic algorithm with reinforcement learning to solve the TSP. Expert Systems with Applications, 36, 6995–7001. Liu, H., Zheng, Z. B., Zhang, W. M., & Ren, K. J. (2011). A global graph-based approach for transaction and QoS-aware service composition. KSII Transactions on Internet and Information Systems, 5, 1252–1273. Liu, M., Wang, M. R., Shen, W. M., Luo, N., & Yan, J. W. (2012). A quality of service (QoS)-aware execution plan selection approach for a service composition process. Future Generation Computer Systems-the International Journal of Grid Computing and Escience, 28, 1080–1089. Liu, S., Xiong, G., Zhao, H., Dong, X., & Yao, J. (2012). Service composition execution optimization based on state transition matrix for cloud computing. In Intelligent Control and Automation (WCICA), 2012 10th World Congress on (pp. 4126-4131). Liu, Y., Li, M., & Wang, Q. (2012). A novel user-preference-driven service selection strategy in cloud computing. International Journal of Advancements in Computing Technology, 4, 414–421. Ludwig, S. A. (2012). Clonal selection based genetic algorithm for workflow service selection. In Evolutionary Computation (CEC), 2012 IEEE Congress on (pp. 1–7). Luo, Y. S., Qi, Y., Hou, D., Shen, L. F., Chen, Y., & Zhong, X. (2011). A novel heuristic algorithm for QoS-aware end-to-end service composition. Computer Communications, 34, 1137–1144. Milanovic, N., & Malek, M. (2004). Current solutions for web service composition. IEEE Internet Computing, 8, 51–59. Moscato, P., & Cotta, C. (2003). A Gentle Introduction to Memetic Algorithms. In F. Glover & G. Kochenberger (Eds.), Handbook of Metaheuristics (57, pp. 105–144). US: Springer. Moura, L. D., & Bjørner, N. (2011). Satisfiability modulo theories: introduction and applications. Communication ACM, 54, 69–77. Neapolitan, R., & Naimipour, K. (2009). Foundations of Algorithms (4). Jones and Bartlett Publishers, Inc (December 28, 2009). Peter Mell, T. G. (2011). The NIST Definition of Cloud Computing. In N. I. o. S. a. Technology (Ed.): U.S. Department of Commerce. Pham, T. V., Jamjoom, H., Jordan, K., & Shae, Z.-Y. (2010). A service composition framework for market-oriented high performance computing cloud. In Proceedings of the 19th ACM International Symposium on High Performance Distributed Computing (pp. 284–287). Chicago, Illinois: ACM. Preve, N. P. (2011). Grid Computing: Towards a Global Interconnected Infrastructure. Springer. Qi, Y., & Bouguettaya, A. (2013). Efficient service skyline computation for composite service selection. IEEE Transactions on Knowledge and Data Engineering, 25, 776–789. Qiang, D., Yuhong, Y., & Vasilakos, A. V. (2012). A survey on service-oriented network virtualization toward convergence of networking and cloud computing. IEEE Transactions on Network and Service Management, 9, 373–392. Reynolds, R. G. (1999). Cultural Algorithms: Theory and Applications. In C. David, D. Marco, G. Fred, D. Dipankar, M. Pablo, P. Riccardo, & V. P. Kenneth (Eds.), New Ideas in Optimization (pp. 367–378). UK: McGraw-Hill Ltd.. Richardson, L. (2007). RESTful Web Services (1). O’Reilly Media. Sayed, A. H. (2003). Fundamentals of Adaptive Filtering. Wiley. Schmid, S., Chart, T., Sifalakis, M., & Scott, A. (2002). Flexible, Dynamic, and Scalable Service Composition for Active Routers. In Proceedings of the IFIP-TC6 4th International Working Conference on Active Networks (pp. 253–266). Springer- Verlag. Shangguang, W., Zibin, Z., Qibo, S., Hua, Z., & Fangchun, Y. (2011). Cloud model for service selection. In Computer Communications Workshops (INFOCOM WKSHPS), 2011 IEEE Conference on (pp. 666–671). Shrme, A. E. (2011). Hybrid intelligent technique for automatic communication signals recognition using bees algorithm and MLP neural networks based on the efficient features. Expert Systems with Applications, 38, 6000–6006. Singh, M. P. (2001). Physics of service composition. IEEE Internet Computing, 5, 6–7. Sinnema, M., & Deelstra, S. (2007). Classifying variability modeling techniques. Information and Software Technology, 49, 717–739. Smith, R. G. (1981). Correction to the contract net protocol: high-level communication and control in a distributed problem solver. IEEE Transactions on Computers, C-30, 372. Song, X., Dou, W., & Chen, J. (2011). A workflow framework for intelligent service composition. Future Generation Computer Systems, 27, 627–636. Song, X. D., Dou, W. C., & Chen, J. J. (2011). A workflow framework for intelligent service composition. Future Generation Computer Systems-the International Journal of Grid Computing and Escience, 27, 627–636. Song, Z., Geng, X., Kusiak, A., & Xu, C. (2011). Mining Markov chain transition matrix from wind speed time series data. Expert Systems with Applications, 38, 10229–10239. Stewart, B. J. (2009). Leveraging Amazon Web Services for Enterprise Application Integration (1). IBM Corporation, developerWorks, pp. 1–41. Strunk, A. (2010). QoS-aware service composition: a survey. In Web Services (ECOWS), 2010 IEEE 8th European Conference on (pp. 67–74). Sundareswaran, S., Squicciarini, A., & Lin, D. (2012). A Brokerage-Based Approach for Cloud Service Selection. In Cloud Computing (CLOUD), 2012 IEEE 5th International Conference on (pp. 558–565). Takabi, H., Joshi, J. B. D., & Gail-Joon, A. (2010). Security and privacy challenges in cloud computing environments. IEEE Security and Privacy, 8, 24–31. Tang, M. L., & Ai, L. F.IEEE. (2010). A Hybrid Genetic Algorithm for the Optimal Constrained Web Service Selection Problem in Web Service Composition. New York: IEEE. Torkashvan, M., & Haghighi, H. (2012). A Greedy Approach for Service Composition. New York: IEEE. Vanderbei, R. J. (2008). Linear Programming: Foundations and Extensions. Springer London, Limited. Vaquero, L. M., Rodero-Merino, L., Caceres, J., & Lindner, M. (2008). A break in the clouds: towards a cloud definition. SIGCOMM Compututer Communication Review, 39, 50–55. Wang, J.-Z., Wang, J.-J., Zhang, Z.-G., & Guo, S.-P. (2011). Forecasting stock indices with back propagation neural network. Expert Systems with Applications, 38, 14346–14355. http://refhub.elsevier.com/S0957-4174(13)00992-5/h0685 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0685 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0685 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0200 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0200 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0690 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0690 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0210 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0210 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0210 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0210 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0215 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0215 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0215 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0695 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0695 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0695 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0695 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0235 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0235 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0700 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0700 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0705 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0705 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0705 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0710 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0710 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0280 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0280 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0280 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0285 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0285 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0285 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0290 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0290 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0300 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0300 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0305 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0305 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0305 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0310 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0310 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0310 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0720 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0720 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0720 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0330 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0330 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0335 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0335 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0335 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0340 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0340 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0340 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0340 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0350 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0350 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0350 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0360 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0360 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0360 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0365 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0365 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0725 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0725 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0725 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0810 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0810 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0735 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0735 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0390 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0390 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0390 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0390 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0740 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0740 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0400 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0400 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0400 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0405 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0405 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0405 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0745 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0745 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0745 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0750 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0420 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0755 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0755 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0755 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0755 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0435 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0435 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0435 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0440 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0445 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0445 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0760 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0760 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0760 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0455 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0455 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0460 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0460 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0460 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0465 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0465 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0465 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0765 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0765 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0485 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0485 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0770 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0770 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0770 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0495 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0495 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0775 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0775 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0505 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0505 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0505 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0510 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0510 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0510 3824 A. Jula et al. / Expert Systems with Applications 41 (2014) 3809–3824 Wang, S., Sun, Q., & Yang, F. (2010). Towards web service selection based on QoS estimation. Internatinal Journal of Web Grid Services, 6, 424–443. Wang, S. G., Sun, Q. B., Zou, H., & Yang, F. C. (2013). Particle swarm optimization with skyline operator for fast cloud-based web service composition. Mobile Networks & Applications, 18, 116–121. Wang, Y. W. (2009). Application of Chaos Ant Colony Algorithm in Web Service composition based on QoS. Los Alamitos: IEEE Computer Soc. Wischik, D., Handley, M., & Braun, M. B. (2008). The resource pooling principle. SIGCOMM Computer Communication Review, 38, 47–52. Wittern, E., Kuhlenkamp, J., & Menzel, M. (2012). Cloud service selection based on variability modeling. LNCS, 7636, 127–141. Worm, D., Zivkovic, M., van den Berg, H., & van der Mei, R. (2012). Revenue maximization with quality assurance for composite web services. In Service- Oriented Computing and Applications (SOCA), 2012 5th IEEE International Conference on (pp. 1–9). Wu, Q., Zhang, M., Zheng, R., Lou, Y., & Wei, W. (2013). A QoS-satisfied prediction model for cloud-service composition based on a hidden Markov model. Mathematical Problems in Engineering, 2013, 7. Xiangkun, T., & Yi, G. (2010). A cosine theorem based algorithm for similarity aggregation of ontologies. In Signal Processing Systems (ICSPS), 2010 2nd International Conference on Vol. 2 (pp. V2–16-V12-19). Xiaona, W., Bixin, L., Rui, S., Cuicui, L., & Shanshan, Q. (2012). Trust-Based Service Composition and Optimization. In Software Engineering Conference (APSEC), 2012 19th Asia-Pacific Vol. 1 (pp. 67–72). Yakimenko, O. A., Slegers, N. J., Bourakov, E. A., Hewgley, C. W., Bordetsky, A. B., Jensen, R. P., Robinson, A. B., Malone, J. R., & Heidt, P. E. (2009). Mobile system for precise aero delivery with global reach network capability. In Control and Automation, 2009. ICCA 2009, IEEE International Conference on (pp. 1394–1398). Yang, Y., Mi, Z., & Sun, J. (2012). Game theory based IaaS services composition in cloud computing environment. Advances in Information Sciences and Service Sciences, 4, 238–246. Ye, Z., Zhou, X., & Bouguettaya, A. (2011). Genetic algorithm based QoS-aware service compositions in cloud computing. In J. Yu, M. Kim, & R. Unland (Eds.). Database Systems for Advanced Applications (6588, pp. 321–334). Berlin Heidelberg: Springer. Yi, W., & Blake, M. B. (2010). Service-oriented computing and cloud computing: challenges and opportunities. IEEE Internet Computing, 14, 72–75. Yu, Q., & Bouguettaya, A. (2010). Computing service skyline from uncertain QoWS. IEEE Transactions on Services Computing, 3, 16–29. Yu, T., & Lin, K.-J. (2005). Service selection algorithms for composing complex services with multiple qos constraints. In Proceedings of the Third International Conference on Service-Oriented Computing (pp. 130–143). Amsterdam, The Netherlands: Springer-Verlag. Zeng, C., Guo, X. A., Ou, W. J., & Han, D. (2009). Cloud computing service composition and search based on semantic. In M. G. Jaatun, G. Zhao, & C. Rong (Eds.). Cloud Computing, Proceedings (5931, pp. 290–300). Berlin: Springer-Verlag Berlin. Zeng, J., Duan, J., & Wu, C. (2010). A new distance measure for hidden Markov models. Expert Systems with Applications, 37, 1550–1555. Zhang, M., Ranjan, R., Nepal, S., Menzel, M., & Haller, A. (2012). A declarative recommender system for cloud infrastructure services selection. In Proceedings of the 9th International Conference on Economics of Grids, Clouds, Systems, and Services (pp. 102–113). Berlin, Germany: Springer-Verlag. Zhang, X. Y., & Dou, W. C. (2010). Preference-aware QoS evaluation for cloud web service composition based on artificial neural networks. In F. L. Wang, Z. G. Gong, X. F. Luo, & J. S. Lei (Eds.). Web Information Systems and Mining (6318, pp. 410–417). Berlin: Springer-Verlag Berlin. Zhao, S. S., Ma, L., Wang, L., & Wen, Z. P. (2012). An improved ant colony optimization algorithm for QoS-aware dynamic web service composition. Los Alamitos: IEEE Computer Soc.. Zhao, X., Wen, Z., & Li, X. (2013). QoS-aware web service selection with negative selection algorithm. Knowledge and Information Systems, 1–25. Zhao, X. C., Song, B. Q., Huang, P. Y., Wen, Z. C., Weng, J. L., & Fan, Y. (2012). An improved discrete immune optimization algorithm based on PSO for QoS- driven web service composition. Applied Soft Computing, 12, 2208–2216. Zhou, W., & Bovik, A. C. (2009). Mean squared error: love it or leave it? A new look at signal fidelity measures. IEEE Signal Processing Magazine, 26, 98–117. Zhou, X., & Mao, F. (2012). A Semantics Web Service Composition Approach Based on Cloud Computing. In Computational and Information Sciences (ICCIS), 2012 Fourth International Conference on (pp. 807–810). Zhu, Y., Li, W., Luo, J., & Zheng, X. (2012). A novel two-phase approach for QoS- aware service composition based on history records. In Service-Oriented Computing and Applications (SOCA), 2012 5th IEEE International Conference on (pp. 1–8). Zibin, Z., Xinmiao, W., Yilei, Z., Lyu, M. R., & Jianmin, W. (2013). QoS ranking prediction for cloud services. IEEE Transactions on Parallel and Distributed Systems, 24, 1213–1222. Zibin, Z., Yilei, Z., & Lyu, M. R. (2010). Distributed QoS Evaluation for Real-World Web Services. In Web Services (ICWS), 2010 IEEE International Conference on (pp. 83–90). Zissis, D., & Lekkas, D. (2012). Addressing cloud computing security issues. Future Generation Computer Systems, 28, 583–592. http://refhub.elsevier.com/S0957-4174(13)00992-5/h0515 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0515 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0520 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0520 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0520 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0525 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0525 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0530 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0530 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0780 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0780 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0545 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0545 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0545 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0565 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0565 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0565 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0785 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0785 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0785 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0785 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0575 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0575 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0580 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0580 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0585 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0585 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0585 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0585 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0590 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0590 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0590 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0595 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0595 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0600 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0600 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0600 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0600 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0605 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0605 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0605 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0605 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0610 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0610 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0610 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0615 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0615 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0620 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0620 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0620 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0625 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0625 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0640 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0640 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0640 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0650 http://refhub.elsevier.com/S0957-4174(13)00992-5/h0650 Cloud computing service composition: A systematic literature review 1 Introduction 2 Survey goals and execution 2.1 Survey goals and research questions 2.2 Publication statistics 3 Cloud computing 3.1 Cloud definition 3.2 Cloud computing characteristics 3.3 Cloud computing service models 3.4 Cloud computing deployment models 4 The cloud computing service composition problem (CCSC) 4.1 CCSC definition 4.2 CCSC challenges 4.3 Current approaches for CCSC 4.3.1 Classic and graph-based algorithms 4.3.2 Combinatorial algorithms 4.3.3 Machine-based methods 4.3.4 Structures 4.3.5 Frameworks 5 Discussion 5.1 Objectives of the researches 5.2 Applied approaches 5.3 Investigated datasets 5.4 Significance of QoS parameters and their percentage 6 Conclusion and future works References 
work_3rmubmxsfbbjlk6w2bf2tmt7nm	----	Article Reference Studying the Emotion-Antecedent Appraisal Process: An Expert System Approach SCHERER, Klaus R. Abstract The surprising convergence between independently developed appraisal theories of emotion elicitation and differentiation is briefly reviewed. It is argued that three problems are responsible for the lack of more widespread acceptance of such theories: (1) the criticism of excessive cognitivism raised by psychologists working on affective phenomena; (2) the lack of process orientation in linking appraisal to the complex unfolding of emotion episodes over time; and (3) the lack of consensus on the number and types of appraisal criteria between theorists in this domain. Although readers are referred to recent theoretical discussions and evidence from the neurosciences with respect to the first two issues, an empirical study using computerised experimentation is reported with respect to the third issue. Data obtained with an expert system based on Scherer's (1984a) “stimulus evaluation check” predictions show the feasibility of this approach in determining the number and types of appraisal criteria needed to explain emotion differentiation. It is suggested to use computer modelling and experimentation as a powerful tool [...] SCHERER, Klaus R. Studying the Emotion-Antecedent Appraisal Process: An Expert System Approach. Cognition and Emotion, 1993, vol. 7, no. 3/4, p. 325-355 DOI : 10.1080/02699939308409192 Available at: http://archive-ouverte.unige.ch/unige:102025 Disclaimer: layout of this document may differ from the published version. 1 / 1 http://archive-ouverte.unige.ch/unige:102025 Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=pcem20 Download by: [Université de Genève] Date: 08 January 2018, At: 02:28 Cognition and Emotion ISSN: 0269-9931 (Print) 1464-0600 (Online) Journal homepage: http://www.tandfonline.com/loi/pcem20 Studying the emotion-antecedent appraisal process: An expert system approach Klaus R. Scherer To cite this article: Klaus R. Scherer (1993) Studying the emotion-antecedent appraisal process: An expert system approach, Cognition and Emotion, 7:3-4, 325-355, DOI: 10.1080/02699939308409192 To link to this article: https://doi.org/10.1080/02699939308409192 Published online: 07 Jan 2008. Submit your article to this journal Article views: 377 View related articles Citing articles: 143 View citing articles http://www.tandfonline.com/action/journalInformation?journalCode=pcem20 http://www.tandfonline.com/loi/pcem20 http://www.tandfonline.com/action/showCitFormats?doi=10.1080/02699939308409192 https://doi.org/10.1080/02699939308409192 http://www.tandfonline.com/action/authorSubmission?journalCode=pcem20&show=instructions http://www.tandfonline.com/action/authorSubmission?journalCode=pcem20&show=instructions http://www.tandfonline.com/doi/mlt/10.1080/02699939308409192 http://www.tandfonline.com/doi/mlt/10.1080/02699939308409192 http://www.tandfonline.com/doi/citedby/10.1080/02699939308409192#tabModule http://www.tandfonline.com/doi/citedby/10.1080/02699939308409192#tabModule COGNITION AND EMOTION. 1993, 7(3/4), 325-355 Studying the Emotion-Antecedent Appraisal Process: An Expert System Approach Klaus R. Scherer University of Geneva, Switzerland The surprising convergence between independently developed appraisal theories of emotion elicitation and differentiation is briefly reviewed. It is argued that three problems are responsible for the lack of more widespread acceptance of such theories: (1) the criticism of excessive cognitivism raised by psychologists working on affective phenomena; (2) the lack of process orientation in linking appraisal to the complex unfolding of emotion episodes over time; and (3) the lack of consensus on the number and types of appraisal criteria between theorists in this domain. Although readers are referred to recent theoretical discussions and evidence from the neurosciences with respect to the first two issues, an empirical study using computerised experimentation is reported with respect to the third issue. Data obtained with an expert system based on Scherer’s (1984a) “stimulus evaluation check” predictions show the feasibility of this approach in determining the n u m b e r and types of appraisal criteria needed to explain emotion differentia- tion. It is suggested to use computer modelling and experimentation as a powerful tool to further theoretical development and collect pertinent data on the emotion-antecedent appraisal process. INTRODUCTION The notion that emotions are elicited and differentiated via appraisal of situations or events as centrally important to a person has a venerable history. The idea can be traced from t h e writings of early philosophers such as Aristotle, Descartes, and Spinoza to theoretical suggestions by pioneering emotion psychologists such as Stumpf (see Reisenzein & Schonpflug, 1992). In the 196Os, Arnold (1960) and Lazarus (1968) had Requests for reprints should be sent to Klaus R . Scherer, Department of Psychology, University of Geneva, 9, Rte de Drize, Carouge, CH-1227 Geneva, Switzerland. This paper was specifically prepared for the special issue of Cognition and Emotion on Appraisal and Beyond. The author gratefully acknowledges important contributions and suggestions by George Chwclos, Nico Frijda, Keith Oatley, Ursula Scherer. and two anonymous reviewers. Q 1993 Lawrence Erlbaum Associates Limited D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 326 SCHERER explicitly formulated theories incorporating rudimentary appraisal criteria in an effort to explain the emotional consequences of being faced with a particular event. At the beginning of the 1980s a number of psychologists independently proposed detailed and comprehensive sets of appraisal criteria to explain the elicitation and differentiation of the emotions (De Rivera, 1977; Frijda, 1986; Johnson-Laird & Oatley, 1989; Mees, 1985; Ortony, Clore, & Collins, 1988; Roseman, 1984, 1991; Scherer, 1981, 1982,1983, 1984a,b, 1986; Smith & Ellsworth, 1985,1987; Solomon, 1976; Weiner, 1982) and engaged in empirical research to demonstrate the validity of these hypothetical suggestions (Ellsworth & Smith, 1988; Frijda, 1987; Frijda, Kuipers, & ter Schure, 1989; Gehm & Scherer, 1988; Manstead & Tetlock, 1989; Reisenzein & Hofrnann, 1990; Roseman, 1984, 1991; Roseman, Spindel, & Jose, 1990; Smith & Ellsworth, 1985, 1987; Tesser, 1990; Weiner, 1986). In a comparative review of such “appraisal theories of emotion differentiation” Scherer (1988) attempted to show the extraordinary degree of convergence of the different theoretical sugges- tions, especially with respect to the central criteria postulated in the different approaches (see Table 1, reproduced from Scherer, 1988). This convergence is all the more surprising since the theorists concerned come from widely different traditions in psychology and philosophy. The impres- sion that appraisal theories of emotion differentiation have generated a highly cumulative body of research has been confirmed in more recent reviews as well as in some comparative empirical studies (Lazarus & Smith, 1988; Manstead & Tetlock, 1989; Reisenzein & Hofmann, 1990; Roseman, et al. 1990; Scherer, 1988). It seems reasonable to take such theoretical and empirical convergence as an indication of the plausibility and validity of appraisal theories, particularly in the light of the absence of rival theories that could reasonably claim to explain emotion differentiation by alternative con- ceptual frameworks. Yet, appraisal theories currently face three major challenges which seem to prevent more widespread acceptance of this explanatory framework: (1) the reproach of excessive cognitivism; (2) the lack of process orientation; and (3) the lack of consensus on the number and types of appraisal criteria. 1. The Reproach of Excessive Cognitivism Appraisal theorists are often accused of excessive cognitivism by psycholo- gists dealing with a wide variety of different affective phenomena. Critics question the likelihood that elaborate cognitive evaluations are performed during the few milliseconds that seem sufficient to bring about a n emotion episode. It is further suggested that affective arousal can be triggered without any evaluative processing at all (Zajonc, 1980). The “cognition- emotion controversy” (Lazarus, 1984a,b; LeDoux, 1987, 1989; Leventhal D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 T A B L E 1 C o n v er g en ce o f S et s of A pp ra is al C ri te ri a as S u g g es te d b y D if fe re nt A pp ra is al T h eo ri st s S ch er er F ri jd a O rt o n yl C lo re R o se rn an S rn ith lE IL F w o rt h S o lo m o n W ei n er N o ve lt y S ud de nn es s Fa m il ia ri ty Pr ed ic ta bi li ty In tr in si c p le as an tn es s G o a l si g n if ic an ce C on ce rn r el ev an ce O ut co m e pr ob ab il it y E xp ec ta ti on C on du ci ve ne ss U rg en cy C o p in g p o te n ti al C au se : A ge nt C au se : M ot iv e C on tr ol Po w er A dj us tm en t C o m p at ib ili ty s ta n d ar d s E xt er na l In te rn al C ha ng e Fa m ili ar ity V al en ce Fo ca lit y C er ta in ty Pr es en ce O pe nl C lo se d U rg en cy In te nu se lf -O th er M od if ia bi lit y C on tr ol la bi li ty V al ue r el ev an ce U ne xp ec te dn es s A pp ea li ng ne ss L ik el ih oo d Pr os pe ct r ea li sa ti on D es ir ab il it y Pr ox im it y A ge nc y B la m ew or th in es s A tt en ti on Pl ea sa nt ne ss A pp IA ve M ot iv es S co pe F oc us Pr ob ab il it y C er ta in ty M ot iv e co ns is te nc y G oa ll P at h ob st ac le E va lu at io n A nt ic ip at ed e ff or t A ge nc y P ow er A ge nc y A ge nc y L eg it im ac y R es po ns ib il it y L oc us o f ca us al it y St ab il it y C on tr ol la bi li ty P ow er C on tr ol la bi li ty A pp , A pp ro ac h; A ve , av oi da nc e. R ep ro du ce d fr om S ch er er ( 19 88 . p . 92 ) w N 4 D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 328 SCHERER & Scherer, 1987; Zajonc, 1980, 1984; Zajonc & Markus, 1984) is centrally concerned with this issue. The crux of the matter, however, is the definition of cognition, a term which has not gained in precision by becoming increasingly fashionable. Although the formulations used by some theorists may suggest that appraisal is viewed as a conscious, and consequently exclusively cortically based process, other theorists in this tradition have insisted early o n that the cognitivistic connotations of the terms “appraisal” or “evaluation” do nor preclude that a substantial part of these processes occur in an unconscious fashion, mediated via sub- cortical, e.g. limbic system, structures (Scherer, 1984a,b). Leventhal and Scherer (1987) have pointed out that evaluation can occur at the sensori- motor, schematic, or conceptual levels, respectively, and that, rather than discussing the cognition issue on an abstract level, one should determine the precise nature of the information-processing involved. LeDoux (1989), from a neuropsychological point of view, has likewise advocated to address the issue of the nature of emotion-antecedent information-processing and its underlying neural pathways rather than getting sidetracked by the issue of the definition of cognition: “The process involved in stimulus evaluation could, if one chose, be called cognitive processes. The meaning of the stimulus is not given in physical characteris- tics of the stimulus but instead is determined by computations performed by the brain. As computation is the benchmark of the cognitive, the computation of affective significance could be considered a cognitive process” (LeDoux, 1989, p. 271). LeDoux and his coworkers have in fact empirically demonstrated the existence of subcortical stimulus evaluation patterns for affect eliciting situations in rats (LeDoux, 1987, 1989; LeDoux, Farb, & Rugiero, 1990). The empirical demonstration of such patterns in humans is hardly to be expected at present because most current research on emotion-antecedent appraisal in human subjects uses self-report of emotional experiences (necessarily involving higher centres of the brain). Subjects are generally asked to recall or infer the nature of their event or situation appraisal, often with the help of rating scales constructed on the basis of the theoretically assumed appraisal dimensions. Clearly, verbally reported appraisal patterns are mediated via conscious, almost exclusively cortically controlled information-processing, and are thus easy targets for charges of excessive cognitivism. They are also subject to the criticism that such recall or inference illustrates social representations of emotions rather than reflecting the actual emotion-eliciting process. Given the difficulty of settling these issues empirically, Scherer (1993) has suggested to look toward potential contributions from the neurosciences to better understand the nature of the appraisal process. The author denotes a number of possibilities of empirically studying controversial D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 329 questions related to the appraisal notion with the help of modern neuroscience technology, such as electroencephalographic signal analysis and imaging techniques, and adopting neuropsychologically oriented experimental designs as well as case studies of neurologically impaired patients. Such procedures might help to overcome one of the most serious limitations of current empirical research on emotion-antecedent appraisal: The reliance on respondents’ verbal reports of recalled or inferred situation evaluations. 2. The Lack of Process Orientation The second problem mentioned earlier, lack of a process orientation in many appraisal theories, is responsible for the frequently encountered opinion that appraisal theories basically provide a semantic grid for the comprehension of the use of emotion terms or labels, and are thus limited to structural analyses or explications of semantic fields of emotion terms. This impression is due partly to the explicit semantic orientation of some ‘of the models that have been proposed (Ortony et al., 1988), and partly to the use of verbal labels in all theories to identify the emotional states that are seen to be elicited and differentiated by the appraisal process. It is certainly one of the legitimate applications of appraisal theories to identify the nature of the emotion-antecedent appraisal process that determines which verbal label will be chosen to communicate the nature of the emotion episode. However, appraisal theories need to go beyond semantics and attempt to specify the true nature of the emotion-antecedent appraisal process. This process might result in an emotional state that the person concerned is unable or unwilling to label with one of the standard emotion terms that are currently used in emotion research. Scherer (1984a) has argued that the stimulus or event evaluation process can elicit as many different emotional states as there are distinguishable outcomes of the appraisal process. This suggestion clearly contradicts the notion that there are a very limited number of “basic” or “fundamental” discrete emotions (Ekman, 1984, 1992; Izard, 1977; Tomkins, 1984). In order to allow systematic discussion of this issue, it is necessary to agree on a consensual definition of emotion that helps to explicate the boundaries between different emotional states and their components (see Scherer, 1993). A further requirement for advancing in the debate on this issue is the specification of the micro-generic process of appraisal and reaction. Although many emotion theories give the impression that emotions are static states that can be conveniently labelled with a single term, there can be little doubt that we need to talk about emotion epkodes that are characterised by continuously occurring changes in the underlying appraisal and reaction processes (see Folkman & Lazarus, 1985; Frijda, D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 330 SCHERER 1986; Scherer, 1984a,b). In consequence, it is not sufficient to specify a pattern of appraisal results that is supposed to explain a static emotion as indexed by a label. The nature of the appraisal process and the immediate effects of the evaluation results on the other components of emotion (such as subjective feeling, physiological responses, motor expression, and action tendencies) need to be explored. Unfortunately, most of the appraisal theorists have so far devoted only very limited attention to the process underlying the evaluation of situations, events, or actions. An exception to this general pattern is the component process theory suggested by Scherer (1984a,b, 1986, 1988), which postulates that the appraisal criteria (stimulus evaluation checks, abbreviated as SECs) pro- posed occur in an invariant sequence (in the order shown in Table 2). The sequence notion, which is based on phylogenetic, ontogenetic, and micro- genetic (logical) considerations, cannot be discussed in detail in the present context. Generally speaking, it is assumed that the appraisal process is conrrunrfy operative with evaluations being continuously performed to update the organism’s information on an event or situation (including the current needs or goals of the organism and the possibility to act on these). In consequence, the sequential stimulus evaluation checks are expected to occur in very rapid succession (similar to a rotating radar antenna updating the reflection patterns on the screen). This continuous operation can explain the sudden changes that can occur during emotion episodes and which are often based on re-evaluations of the event or of one’s coping potential (cf. Lazarus’, 1968, “secondary appraisal”; see Scherer, 1984a,b, for further details on the hypothesised sequential processing). Many different objections have been raised against this sequence notion. Quite a few of these can be refuted on logical grounds or on the basis of recent insights into the neural bases of information-processing, particularly with respect to neural networks (see Scherer, 1993, for a detailed dis- cussion). However, empirical research is needed to demonstrate the feasibility of the sequence hypothesis and to encourage further work in this direction. Unfortunately, our dependence on verbal report of recalled or inferred appraisal processes does not lend itself to the study of the sequence hypothesis. It is likely that the different steps of the evaluation process occur extremely rapidly and are not generally represented in awareness. Any reconstruction of these processes is likely to miss the temporal dynamics of the process. In the future, neuroscience technology might allow us to monitor such rapidly occumng evaluation sequences directly. Also, i t seems feasible to develop sophisticated research designs making use of latency time measures in carefully designed stimulus presentation modes to shed some light on these time-critical processes (see D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 331 Scherer, 1993, for concrete suggestions on adopting appropriate paradigms from the cognitive neurosciences). Unfortunately, such studies might well be slow in the making. 3. The Lack or Consensus on the Number and Types of Appraisal Criteria The third problem concerns the issue of how many and precisely which evaluation or appraisal dimensions are necessary to account for the degree of emotion differentiation that can be empirically demonstrated. Although, as mentioned earlier, there is much convergence in this field, authors do differ with respect to the number and definition of appraisal dimensions that are proposed. A few recent studies have attempted to compare different appraisal theories and to empirically determine how many dimensions are needed and which dimensions seem to account for most of the variance (Manstead & Tetlock, 1989; Mauro, Sato, & Tucker, 1992; Reisenzein & Hofmann, 1990; Roseman et al., 1990). All of these studies are limited to post hoc evaluation of how well the dimensions studied explain differentiation between the emotions reported by the subjects. In other words, the same group of subjects provides both the emotion and the appraisal information and statistical analysis is limited to identifying the shared variance. Needless to say, the results cannot be generalised beyond the respective set of emotions and dimensions studied. Even though such information is eminently useful for the further develop- ment of appraisal theories, it seems desirable to develop a model that emphasises the prediction of emotional states on the basis of a minimal set of necessary and sufficient dimensions or criteria of appraisal. The empirical study to be reported in this paper suggests such a predictive approach. Based on Scherer’s component process model of emotion (1984a,b, 1986, 1988), an expert system on emotion differentia- tion that contains such a minimal set of evaluation criteria is presented and submitted to a first empirical test. As shown earlier, the question of how many and which appraisal criteria are minimally needed to explain emotion differentiation is one of the central issues in research on emotion-antecedent appraisal. It is argued here that one can work towards settling the issue by constructing, and continuously refining, an expert system that attempts to diagnose the nature of an emotional experience based exclusively on information about the results of the stimulus or event evaluation processes that have elicited the emotion. The knowledge base of the expert system would contain a limited set of evaluation or appraisal criteria together with theoretically defined (and empirically updated) predictions about which pattern of evaluation results is likely to produce a particular emotion out of a limited D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 T A B L E 2 P at te rn s o f S ti m ul us E va lu at io n C he ck s (S E C ) P re di ct ed to D if fe re nt ia te 1 4 M aj or E m ot io ns ~ E N JI H A P E L A IJ O Y D IS P ID IS G C O N IS C O SA D ID E J D E SP A IR A N X IW O R N ov el ty S ud de nn es s Fa m il ia ri ty Pr ed ic ta bi li ty In tr in si c pl ea sa nt ne ss G oa l s ig ni fic an ce C on ce rn r el ev an ce O ut co m e pr ob ab il it y E xp ec ta ti on C on du ci ve ne ss U rg en cy C op in g po te nt ia l C au se : A ge nt C au se : M ot iv e C on tr ol Po w er A dj us tm en t C om pa tib ili ty s ta nd ar ds E xt er na l In te rn al lo w op en m ed iu m hi gh op en v hi gh co ns on an t co nd uc iv e v lo w op en in te nt op en op en hi gh op en op en hi /m ed op en lo w op en se lf lr el a v hi gh op en v co nd uc iv e lo w op en ch al in t op en op en m ed iu m op en op en op en lo w lo w v lo w bo dy v hi gh op en op en m ed iu m op en op en op en op en op en op en op en op en op en op en op en re ld or de r hi gh op en op en lo w ot he r in te nt hi gh lo w hi gh v lo w v lo w lo w lo w op en op en op en v hi gh op en ob st ru ct lo w op en ch al ne g v lo w v lo w m ed iu m op en op en hi gh v lo w lo w op en op en v hi gh di ss on an t ob st ru ct hi gh o th h at ch al ne g v lo w v lo w v lo w op en op en lo w op en op en op en bo d yl se lf m ed iu m op en ob st ru ct m ed iu m ot w na t op en op en lo w m ed iu m op en op en D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 F E A R IR R IC O A R A G E IH O A B O R II N D S H A M E G U IL T P R ID E N ov el ty S ud de nn es s Fa m il ia ri ty Pr ed ic ta bi li ty in tr in si c pl ea sa nt ne ss G oa l s ig ni fi ca nc e C on ce rn r el ev an ce O ut co m e pr ob ab il it y E xp ec ta ti on C on du ci ve ne ss U rg en cy C op in g po te nt ia l C au se : A ge nt C au se : M ot iv e C on tr ol P ow er A dj us tm en t C om pa ti bi li ty st an da rd s E xt er na l In te rn al hi gh op en lo w lo w bo dy hi gh di ss on an t ob st ru ct v hi gh ot hl na t op en op en v lo w lo w op en oD en lo w op en m ed iu m op en or de r v hi gh op en ob st ru ct m ed iu m op en in t/n eg hi gh m ed iu m hi gh lo w lo w hi gh lo w lo w op en or de r v hi gh di ss on an t ob st ru ct hi gh ot he r in te nt hi gh hi gh hi gh lo w lo w v lo w hi gh v hi gh op en bo dy v hi gh co ns on an t op en lo w op en op en m ed iu m m ed iu m hi gh op en op en lo w op en op en op en se lf v hi gh op en op en hi gh se lf in t/ ne g op en op en m ed iu m op en v lo w op en op en op en op en re ld or de r v hi gh op en hi gh m ed iu m se lf in te nt op en op en m ed iu m v lo w v lo w op en op en op en op en se lf v hi gh op en hi gh lo w se lf in te nt op en op en hi gh hi gh v hi gh _ _ _ _ _ ~ ~ ~ ~ A bb re vi ar io ns : E N JI H A P , en jo ym en th ap pi ne ss ; E L A II O Y , el at io nl jo y; D IS P ID IS G , di sp le as ur el di sg us t; C O N IS C O , co nt em pt ls co rn ; SA D ID E I, s ad ne sd de je ct io n; I R R IC O A , ir ri ta ti od co ld a ng er ; R A G E IH O A , ra ge ho t an ge r; B O R II N D , bo re do di nd if fe re nc e; v , ve ry ; re la , re la ti on sh ip s; n or . n at ur e; c ha , c ha nc e; n eg , n eg li ge nc e; i nr o r in te nt , i nt en ti on ; 0t h o r ot he r. o th er p er so n( s) . R ep ro du ce d fr om S ch er er ( 19 88 , p . 11 2) . w w 0 D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 334 SCHERER set of possibilities. At present, this system is limited to predicting the verbal labels given to the emotions experienced and to obtain the required information about appraisal processes by requesting verbal report of recalled or inferred evaluation results. As shown earlier, this is a highly imperfect approach to study the dynamic appraisal and reaction processes involved in emotional episodes, many of which do not require involvement of consciousness or language-or may not even be accessible to them. However, even an approximative approach to a predictive model seems useful at our present state of knowledge. METHOD Designing the Expert System The aim was to develop a computer program that would allow a user to enter information on a situation in which a strong emotion had been experienced and have the program predict or diagnose the nature of that emotional state (as represented by a verbal label).’ Using TurboPascal3.0, a program called GENESE (Geneva Expert System on Emotions) was developed.* In contrast to expert systems based on IF-THEN rules the present system is of the type that employs algorithms determining the relative similarity between input vectors and prototypical category vectors representing the knowledge base. In the present case t h e “knowledge base” consists of a set of vectors (one for each emotion) which contain quantified predictions relative to the typical stimulus evaluation check outcomes for specific emotions. These vectors have been derived from the prediction tables published by the author in earlier work (Scherer, 1984a,b, 1986, 1988). The most recent set of predictions is shown in Table 2 (reproduced from Scherer, 1988). Concretely, then, for each of the specific emotions contained in the expert system, a vector of numbers (which represent the predicted results of selected stimulus evaluation checks for the respective emotions) con- stitutes the prototypical pattern which will be used to classify user- generated input vectors. The input vector for a target emotion to be classified (which is determined by the user’s choice of a recalled emotional experience he or she wants to have diagnosed) is determined by the computer asking the 15 questions listed in Table 3 and requiring the user to answer with the help of predefined answer categories. Each of these ’ A similar approach was independently developed by Frijda and Swagerman (1987). * The prototype of the system was written in 1987 by Philippe N a r k 1 and Roland Bapst based on specifications by the author who has continuously modified the program since. D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 335 questions corresponds to a particular stimulus evaluation check or sub- check. The numbers representing the predicted prototypical answer alter- natives for each question constitute the entries for the stimulus evaluation checks into the prediction vector for the respective emotion. These prediction vectors are shown in the second row of the vector matrices for the 14 emotions in Table 5. It should be noted that although the prediction vectors have been derived from earlier prediction tables, not all of the stimulus evaluation subchecks listed in Table 2 have been included in the quantified prediction vectors of the GENESE expert system. The need for a selection of what seemed to be the most important and differentiating checks was imposed by the necessity to curtail the number of questions posed to the user. Furthermore, for some subchecks, e.g. agent of causation, several ques- tions had to be asked to obtain the required quantitative information. Table 3 shows the correspondence between the stimulus evaluation checks or subchecks and the specific questions. It should be noted further that the prediction vectors (as contained in the system and shown in Table 5) are based on but do not necessarily correspond exactly to the earlier prediction tables (e.g. Table 2). The author considers theory development a dynamic process. Consequently, predictions change and evolve over time. For example, the prediction vectors in Table 5 show some changes over earlier hypothesising. In particular, an attempt has been made to reduce the number of “open” or “not pertinent” predictions (see Table 2)- particularly in the case of shame and guilt-as these reduce the discriminative power of the vectors in the expert system. The present version of GENESE contains prediction vectors for the 14 emotions listed in Table 2. The choice of these 14 emotions was determined by the arguments advanced in Scherer (1986) advocating to distinguish between more quiet and more aroused varieties of some of the major emotions, e.g. imtatiodcold anger vs. ragehot anger. The input vector, as based on the user’s answers to the 15 questions, is systematically compared to the 14 predicted emotion vectors, using Euclidian distance measures. The distance indices obtained in this fashion are then adjusted on the basis of theoretical considerations concerning the need to weight particular combinations of input values. The following adjustments of the distance indices are used in the present version of the expert system: - 0.3 for shame and guilt if the causal agent is “self’ + 0.8 for all positive emotions if the event is evaluated as unpleasant + 0.3 for contempt except in cases in which another person is the causal and hindering goal attainment agent and the act is highly immoral (- 0.6) D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 336 SCHERER - 0.5 for sadness and desperation if the event happened in the past and - 0.5 for imtation and anger if power and adjustment are high - 0.3 for joy, desperation, fear, and anger if the intensity of the emotion if power and adjustment are low is rated above 4 on a 6-point scale The nature of the adjustments and the size of these increments or decrements of the distance value computed on the basis of the comparison of the input vector with the prediction vectors are based on rules of thumb and are subject to change in future versions of the system. TABLE 3 Questions Posed by t h e Expert System a n d their Correspondence to t h e Stimulus Evaluation Checks ( S E W 1. Did the situation that elicited your emotion happen very suddenly or abruptly? (0) not pertinent (1) not at all (2) a little (3) moderately (4) strongly ( 5 ) extremely 2. Did the situation concern an event o r an action that had happened in the past, that had just happened or that was to be expected for the future? [see text] (0) not pertinent (1) the event had happened a long time ago (2) it happened in the recent past (4) it was to be expected for the near future 3. This type of event, independent of your personal evaluation, would it be generally considered as pleasant or unpleasant? [SEC?-MTRINSIC PLEASANTNESS] (0) not pertinent (1) very unpleasant ( 2 ) rather unpleasant (3) indifferent (4) rather pleasant 4. Was the event relevant for your general well-being, for urgent needs you felt, or for specific goals or plans you were pursuing at the time? [SEC>RELEVANCE] (0) not pertinent (1) not at all ( 2 ) a little (3) moderately (4) strongly ( 5 ) extremely 5 . Did you expect the event and its consequences before the situation actually happened? [ SEC3-EXPECTATION) (0) not pertinent (4) a little ( 5 ) strongly 6. Did the event help or hinder you in satisfying your needs, in pursuing your plans or in attaining your goals? [SECICONDUCTVENESS] (0) not pertinent (4) it helped a little 7. Did you feel that action on your part w a s urgently required to cope with the event and its consequences? [SECI-URGENCY] (0) not wrtinent (1) not at all (2) a little (3) moderately (4) strongly ( 5 ) extremely [SECl-NOVELTY] (3) it had just happened at that moment ( 5 ) it was to be expected in the long run ( 5 ) very pleasant (1) never in my life (2) not really (3) I did not exclude it (1) it hindered a lot (2) it hindered a little (3) it had no effect ( 5 ) it helped a lot (Continued) D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 337 TABLE 3 (Con tinued) 8. Was t h e event caused by your own action-n o t h e r words, were you partially or fully responsible for what happened? [ S E C M A U S A T I O N ] (0) not pertinent ( I ) not at all (2) a little, but unintentionally (3) somewhat, but I was unaware of the consequences (4) quite responsible, 1 knew what I was doing (5) fully responsible, I absolutely wanted t o d o what I did 9. W a s t h e event caused by o n e or several o t h e r p e r s o n s i n o t h e r words, were o t h e r people partially or fully responsible for what happened? [ S E C M A U S A T I O N ] (0) not pertinent (2) a little, but unintentionally (3) somewhat, but helshdthey were unaware of the consequences (4) quite responsible, he/she/they knew what they were doing (5) fully responsible, h e k h d t h e y absolutely wanted t o d o what they did 10. Was t h e event mainly d u e to chance? [ S E C M A U S A T I O N ] (0) not pertinent (3) somewhat, but human action contributed to it 1 1 . Can t h e occurrence and the consequences of this type of event generally be controlled or modified by human action? (SEC4-CONTROLI (0) not pertinent (1) not at all (2) a little (3) moderately (4) strongly (5) extremely 12. Did you feel that you had enough power t o cope with the e v e n t 4 . e . being able t o influence what was happening or t o modify the consequences? [SEC&POWER] (0) not pertinent ( I ) not at all (2) a little (3) moderately (4) strongly ( 5 ) e x m m e l y 13. Did you feel that, after having used all your means of intervention, you could live with the situation a n d adapt to the consequences? [SEC&ADIUSTMENT] (0) not pertinent (1) not at all (2) with much difficulty (3) somewhat (4) quite easily ( 5 ) without any problem a t all 14. Would the large majority of people consider what happened t o be quite in accordance with social norms and morally acceptable? [SECS-NORM C O M P A T I B I L I T Y ] (0) not pertinent ( 1 ) certainly not (2) not really (3) probably (4) most likely (5) certainly 15. If you were personally responsible for what happened, did your action correspond to your self-image? [SECS-SELF COMPATIBILITY] (0) not pertinent ( I was not responsible) (1) not at all (2) not really (3) somewhat (4) strongly (5) extremely well (1) not at all (1) not at all (2) a little, but human action was t h e decisive factor (4) strongly ( 5 ) exclusively The emotion with the smallest overall distance measure is suggested to the user as diagnosis of the experienced emotional state. If t h e user does not accept the diagnosis as valid, the emotion with the vector that shows the second smallest distance is proposed as a second guess. If the user rejects this one also, he or she is prompted to provide the correct response in the list of the 14 emotions contained in the standard version of the system. If the user identifies one of these 14 emotions as correct, the D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 338 SCHERER respective prediction vector is changed in the direction of the empirical input vector (using an adaptable weighting function) to establish an empirically updated prediction matrix for this particular user. The user can indicate that none of the 14 emotion labels proposed corresponds to the real emotion that was felt. He or she then has the possibility to enter a freely chosen verbal label for that particular state. This label, together with the input vector, is then added to the personalised prediction matrix. In this manner, an unlimited number of new emotions can be added to the personalised knowledge base of a user. It should be noted that the personalised knowledge base for a particular user no longer represents pure theoretical predictions because the prediction vectors have been adapted to fit the empirical input. After prolonged usage by a particular user, the prediction matrix may actually represent a true empirical knowledge base-at least for that particular user. The system stores all the information provided by the user in two separate data files, one which contains the complete protocol of the session and one that contains the personalised vector matrix. Procedure The system h a s been designed in such a fashion that it does not require any intervention by an experimenter. Users are expected to start the program and follow instructions on the screen which should be self- explanatory. I n the following, a brief summary of the typical procedure is given. Following a title page and the entry of a user code that permits repeated access and establishes a personalised data base, the user is requested to remember a situation that has produced a strong emotional response: Please recall a situation in which you experienced a strong emotional feeling. The emotion might have been elicited by an event that happened t o you or by the consequences of your own behaviour. This might have happened recently or quite some time ago. I will ask you a certain number of questions concerning this situation and will then attempt to diagnose t h e emotion you felt at that time. Before continuing, please recall the situation as best as you can and t r y to reconstruct the details of what happened. The program then pauses until the subject confirms to now recall a situation very vividly by pressing a key. He o r she is t h e n asked to type a brief description of the situation on the keyboard. To ensure anonymity and privacy, the text typed is not shown on the screen. Then, the 15 D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 339 questions shown in Table 3 are presented consecutively. The questions are always presented in the order given in Table 3 because the underlying theory predicts that this is the natural micro-genetic appraisal sequence. It is hypothesised that following the original sequence of the appraisal in assessing the different checks may help the subject to recall t h e appraisal process faster and with fewer errors. The subject is then asked to enter the intensity with which the emotion was felt on a &point scale from very weak to extremely strong, as well as hidher age group and gender. Then, the subject is presented with the following message: I have now completed a first diagnosis of the affective state elicited by the situation you described and I am about to present you with a label that I consider to be a good description of the emotion you experienced. Please remember that, at t h e time, you may not have been conscious of all aspects of your emotional experience. Therefore, it is quite possible that the verbal label you normally use to describe your feelings in that situation does not exactly correspond to the term I will suggest i n my diagnosis. If that is the case, please consider the possibility that the diagnosis which I suggest might reflect some part of what you felt i n the situation-possibly without realising it. I GUIEWI EXPERT SYSTen OH DWT"T'OM irritationicold anger sadncssidcprcssion - displeasureidisgust conteqt/scorn - anxictyiuorry - indiffcrcnce/boredor - cibarrassrcntisharc - despcrat ionlgr icf - hot angcriragc - guilt feelings feariterror joyielat ion happinessifcel good prideijubi lation The shorter I line in the follouing graphic display, the m e appropriate should be the respective label as a description of your feeling state. FIG. 1 . different emotion concepts. Feedback screen showing relative distances of input vector to predicted vectors for D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 340 SCHERER The system then presents the first diagnosis or suggested hypothesis and asks the user to indicate whether it is correct or not. If the user enters “incorrect”, a second diagnosis is presented. If that is again incorrect, the user is presented with the list of 14 emotions and asked to indicate the correct emotion. Following either of these three cases the user is presented with feedback on the diagnostic process in form of a graph showing the relative distances of the various emotion concepts to the situation described (see Fig. 1). If the subject indicates that none of the 14 emotion labels in t h e list describes the felt affect accurately, he or she is given the opportunity to enter a new concept: Aha, something new!? Do you really want to teach me a new emotion? I t will change your personal knowledge base! Your decision-(y = yedn = no followed by the new concept] After each of the four possible options: (1) first diagnosis correct; (2) second diagnosis correct; (3) correct emotion identified in list; (4) new concept entered, the user is given the possibility to enter a new situation or to exit. Ad mi n istra t i on This expert system was used in a number of pilot studies, using English, German, and French versions. The French version was used for a first major study of the accuracy of the expert system in diagnosing emotional states on the basis of theoretically predicted appraisal patterns. The program was inserted in a batch-file environment that allows automatic administration. After each user completes a session, the system returns to a title screen inviting potential users to test the power of the system to diagnose emotional states. To avoid the possibility that users would start a session and leave in the middle, a time limit for the responses to each screen was set. If the time limit is exceeded the system returns automatically to the title page. A personal computer (Olivetti M240) on which this batch-file system had been installed, was placed in the exhibition of the University of Geneva at the 1990 Geneva book fair (Salon du Livre). This is a large international bookshow with exhibitors and visitors from different countries, mostly French speaking. Posters positioned around the PC invited passers by to test the GENESE emotion expert system. During three days of the exhibition, 201 persons used the system entering generally one, but sometimes two or three situations. I n addition, 35 first D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 341 year students in psychology at the University of Geneva (in their first 2 months of study) used the system as part of a course exercise (also in a completely automatised fashion). In all, 236 persons entered the data for a total of 282 emotional situations in this manner. Data Analysis A major concern for the analysis is the possibility that some users may have entered nonsensical information while just playing with or trying to mislead the system. However, there were very few cases where the text entered suggested that this was the case. In some cases no text was entered and it was difficult to decide whether the input vector constituted a serious trial or not. To avoid biasing the data by subjective judgement of the “seriousness” of t h e entry it was decided to retain all situations, assuming that nonsensical entries should work against finding accurate diagnoses and thus lead to a conservative estimate of the power of the system. Data were excluded only in the following, clearly discernible cases: In some situations there was virtually no variability in the responses to t h e questions, e.g. a user responding with 1 to all questions. Fifteen situations in which 13 or more of the answers had the same numerical value were excluded. In 14 cases of the total of 282 situations entered, users neither judged any of the diagnoses as correct nor identified any of the 14 emotion labels suggested as the correct response. In these cases new concepts were entered. Because the number of such cases was small, and because in some cases strange concepts like “le spleen total” were entered, it was decided to exclude these cases from analysis. After having excluded these cases, a total of 253 situations were analysed with respect to the number of hits and misses and the correlation between the predicted and the empirically obtained appraisal profiles for the 14 emotions studied. RESULTS Tables 4 and 5 show the major results of the analyses. In Table 4, the first column contains the total number of situations that were entered for each of the 14 emotions (using the final indication of a correct diagnosis or the user correction as a criterion). Column 2 shows in how many of these cases the first diagnosis was correct, and Col. 3 in how many cases the second diagnosis was correct. Column 4 shows the total number of misses. However, some of the latter cases can be considered as “dubious misses” as the input vectors not only deviate strongly from the predicted vectors but also from the empirically obtained mean vectors for each of the emotion (as shown in Table 5). It is highly probable, then, that the D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 342 SCHERER TABLE 4 Results of Expert System Runs for 253 Emotion Situations Total 1st 2nd Total Dubious Marked Profle Emotion Situat Hit Hit M i v u M i r r u % Correct Diff Correl Happiness/Feel good 27 21 1 5 2 88.0 3 0.76 JoyElation 3 4 3 0 3 1 1 100.0 3 0.63 DispleasurJDisgust 10 4 3 3 1 n . 8 1 0.60 ContemptFhrn 3 3 0 0 0 100.0 2 0.51 SadnesslDepression 34 19 3 12 3 71 .O 2 0.24 DesperatiodGrief 58 49 8 1 1 100.0 2 0.75 Anxie ty/Wony 19 2 0 17 5 14.3 4 0.13 Fearmerror 19 2 2 15 4 26.7 10 0.61 ImtatiodCold anger 11 3 4 4 1 70.0 3 0.30 Hot angermage 19 10 4 5 1 n.7 6 0.74 IndifferencJBoredom 3 1 0 2 0 33.3 3 0.36 EmbarrassmentIShame 4 1 0 3 1 33.3 4 0.23 PnddJubilation 9 6 0 3 1 75.0 6 0.75 Guilt feelings 3 1 0 2 1 50.0 9 -0.10 Totals 253 152 2.8 73 22 n . 9 Notrs: To& S h u t , total number of situations clearly categoriscd by respondents; 1st Hit, number correctly recognised on first attempt; 2nd Hit, number correctly recognised on second attempt; Total misses, number misscd as shown by user correction; Dubww Misses, cases in which the deviation of the input profile from the mean empirical profile exceeded half a standard deviation; % Correct, percentage of total hits (first plus second) on the basis of total number of situations minus dubious misses; Murked Difl, marked difference between predicted and empirically obtained vectors; Projile Correl, Pcarson r between the mean empirical input profile and the theoretically specified prediction profile over N = 15 questions (0 in prediction vector treated as missing observation). appraisal information was not provided in the correct manner. Twenty-two situations were considered dubious because the absolute value of the sum of the differences (deviations) obtained by deducting t h e individual values for each question from the mean value-Row 1 in Table 5 exceeded the value corresponding to a standard deviation for all difference scores. Column 5 shows the number of these “dubious misses” per emotion. Column 6 shows the percentage of correct diagnoses (excluding the “dubious misses” which are considered to be the result of incorrect input). Table 5 lists, for each of the 14 emotions, the mean input vector (Row l), the theoretically predicted SEC vector as represented in the knowledge base (Row 2), t h e difference between the two (Row 3), and t h e standard deviations of the empirical values in the input vector (Row 4). This table allows to compare t h e theoretically predicted SEC vectors in the know- ledge base with the empirically obtained input vectors. Thus it permits to determine the stimulus evaluation checks for which the empirical values greatly differ from the predicted value and for which, in consequence, a D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 343 revision of the prediction might be required (if there is reason to believe that the input value does not represent artifacts or errors). For this decision, the standard deviations of the empirical values are useful: High discordance of the values entered for a particular stimulus evaluation check can be taken to indicate that there may not be a standard appraisal pattern (or that the respondents did n o t understand the question). The data in Table 5 generate a large number of interesting issues to be explored. A detailed discussion of these points would exceed the space available in this paper. However, some general trends can be inferred by doing a rough analysis of the size of the difference between theoretical and empirical patterns across stimulus evaluation checks and across emotion^.^ Counting the number of cases in which a difference score exceeds the value of 1 (absolute) for different emotions yields an indication of where the predicted patterns deviate most strongly from the empirically obtained patterns (these values are shown in Col. 7, Table 4). Another way of evaluating the fit between theoretical and empirical patterns is to correlate the two vectors for each emotion. Column 8 in Table 4 shows the mean Pearson correlation coefficients (over respondents) between the predicted profile o r vector in the knowledge base, and the empirically obtained input vectors for each of the emotions across the 15 appraisal criteria (as shown in Table 5 ) . At first observation, although not directly pertinent to the questions outlined earlier, concerns the relative frequency of the different emotions which were presented to the expert system (&I. 1, Table 4). The categories mentioned most frequently are sadneddepression and desperatiodgrief, both of which are closely linked to some kind of permanent loss. Positive emotions, happinesdfeel good and joy/elation are also mentioned rela- tively frequently. Anxietylworry and feadterror, both related to apprehen- sion about impending dangers, are in third position with respect to frequency. Anger states (imtatiodcold anger and hot angedrage) are mentioned the least frequently of the four major fundamental emotion types. The remaining emotions are all relatively low in occurrence. The most important question concerns the accuracy of the expert system in diagnosing the emotional state descriptions entered by the users. Column 6, Table 4 shows the percentage of correct diagnoses o n either the first or second guess. The data in Cols 2 and 3 show that first hits are generally much more frequent (84.4% of all correct diagnoses) than second guesses. The accuracy percentage in Col. 6 is based on a comparison of all It should be noted that the differences should not be interpreted in cases in which the theoretical prediction is h o t pertinent-as the difference score is not interpretable. Also, as explained in the description of the expert system design, the quantitative prediction vectors are based on but not identical to the patterns in the published prediction tables. D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 T A B L E 5 E m p ir ic a ll y O b ta in e d a n d T h e o re ti c a ll y P re d ic te d A p p ra is a l V ec to rs f o r 14 E m o ti o n s E m ot io n N ov T im e Pl ea R el e E xp ec C o n d u U rg en E go C O th C C ha C C on 1 P ow er A dj uc E x1 In t E nj oy J O Y D is gu s C on te m S ad D es pe r 2. 59 2. 00 -0 .5 9 1. 18 3. 29 4. 00 0. 71 1. 30 3. 10 3. 00 -0 .1 0 1. 68 2. 33 2. 00 -0 .3 3 1. 11 3. 06 2. 00 -1 .0 6 1. 24 4. 12 4. 00 -0 .1 2 0. 88 2. 19 4. 41 2. 00 4. 00 -0 .1 9 -0 .4 1 0. 86 0. 70 2. 97 4. 26 3. 00 3. 00 0. 03 -1 .2 6 0. 93 0. 99 2. 30 1. 70 3. 00 1. 00 0. 70 -0 .7 0 1. 44 0. 98 1. 00 1. 67 2. 00 2. 00 1. 00 0. 33 1. 33 0. 89 2. 21 1. 85 2. 00 2. 00 -0 .2 1 0. 15 1. 15 0. 97 2. 74 1. 79 3. 00 2. 00 0. 26 0. 21 0. 91 1. 05 2. 70 3. 59 3. 00 4. 00 0. 30 0. 41 1. 57 0. 88 3. 76 2. 85 4. 00 1. 00 0. 24 -1 .8 5 1. 30 1. 16 1. 60 1. 90 2. 00 2. 00 0. 40 0 .1 0 0. 92 0. 90 1. 67 2. 33 2. 00 2. 00 0. 33 -0 .3 3 0. 89 1. 11 2. 44 2. 65 4. 00 2. 00 1. 56 -0 .6 5 1. 58 1. 39 2. 90 2. 17 5. 00 1. 00 2. 10 -1 .1 7 1. 98 1. 07 2. % 4. 00 1. 04 0. 87 3. 97 5. 00 1. 03 0. 98 2. 40 3. 00 0. 60 I . 20 1 . 00 3. 00 2. 00 0. 67 2. 38 2. 00 -0 .3 8 0. 94 2. 00 1 . 00 -1 .0 0 0. 93 1. 52 3. 11 1. 00 2. 00 -0 .5 2 -1 .1 1 0. 95 1. 38 2. 62 3. 12 2. 00 3. 00 -0 .6 2 -0 .1 2 1. 39 1. 24 2. 40 2. 00 3. 00 1. 00 0. 60 -1 .0 0 1. 60 1. 20 2. 67 2. 00 2. 00 1. 00 -0 .6 7 -1 .0 0 1. 11 0. 67 2. 26 2. 15 2. 00 2. 00 -0 .2 6 -0 .1 5 1. 27 1. 07 3. 34 1. 95 3. 00 1. 00 -0 .3 4 -0 .9 5 1. 51 0. 98 3. 41 3. 00 1. 29 2. 97 3. 00 0. 03 1. 44 2. 70 3. 00 0. 30 - I . 30 4. 00 5. 00 1 .O O 0. 67 2. 62 3. 00 0. 38 1. 46 2. 53 2. 00 -0 .5 3 1. 42 -0 .4 1 - 2. 19 2. 00 -0 .1 9 1. 06 2. 71 3. 00 0. 29 1. 24 3. 30 3. 00 -0 .3 0 1. 10 1 . OO 1 . 00 0. 00 0. 00 2. 21 3. 00 0. 79 1. 25 2. 38 3. 00 0. 62 1. 53 3. 22 0. 00 1. 55 3. 32 0. 00 1. 40 2. 30 3. 00 0. 70 1. 36 2. 33 3. 00 0. 67 1. 78 3. 12 3. 00 -0 .1 2 1. 59 2. 52 3. 00 0. 48 1. 41 2. 37 0. 00 1. 20 2. 59 0. 00 1. 26 1. 70 2. 00 0. 30 0. 84 I . 33 4. 00 2. 67 0. 89 1. 71 0. 00 0. 91 1. 50 0. 00 0. 72 . 3. 33 5. 00 1. 67 1. 21 3. 53 4. 00 0. 47 1. 14 2. 20 4. 00 1. 80 1. 36 3. 67 4. 00 0. 33 0. 89 2. 26 2. 00 -0 .2 6 0. 87 1. 69 1 .O O -0 .6 9 0. 54 3. 85 2. 93 0. 00 0. 00 1. 09 1. 44 3. 97 2. 88 0. 00 0. 00 0. 93 1. 56 2. 70 1. 50 3. 00 0. 00 0. 30 1. 56 1. 60 3. 00 1. 00 2. 00 0. 00 0. 67 1. 33 2. 26 1. 56 0. 00 0. 00 1. 17 1. 12 1. 90 1. 47 0. 00 0. 00 1. 27 1. 27 -1 .0 0 . D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 E m ot io n N ov T im e P le o R el e E xp ec C on du U rg en E go C O ih C C ho C C on i P ow er A dj u r Ex 1 In r A n x i 2. 42 2. 74 1. 26 3. 05 3. 16 3. 00 2. 79 2. 32 2. 11 2. 16 2. 95 2. 42 2. 68 3. 05 1. 68 2. 00 5. 00 2. 00 3. 00 2. 00 2. 00 3. 00 1. 00 4. 00 3. 00 3. 00 2. 00 3. 00 0. 00 0. 00 1. 43 1. 10 0. 49 1. 42 1. 11 0. 74 1. 51 1. 28 1. 39 1. 25 1. 64 1. 07 0. 58 1. 41 1. 35 F ea r 4. 32 2. 79 1. 42 2. 79 2. 05 2. 11 3. 53 2. 11 2. 11 2. 37 3. 42 2. 26 1. 47 1. 00 1. 58 5. 00 4. 00 2. 00 5. 00 1. 00 1. 00 5. 00 1. 00 4. 00 4. 00 2. 00 1. 00 2. 00 0. 00 0. 00 0. 68 1. 21 0. 58 2. 21 -1 .0 5 -1 .1 1 1. 47 -1 .1 1 1. 89 1. 63 -1 .4 2 -1 .2 6 0. 53 0. 94 1. 05 0. 71 1. 80 0. 82 0. 88 1. 44 0. 99 1. 20 1. 48 1. 30 1. 36 0. 97 0. 84 1. 51 lr ri t 3. 27 2. 64 2. 18 2. 73 3. 45 2. 36 3. 09 2. 55 3. 36 1. 55 3. 55 2. 45 2. 91 3. 00 2. 27 3. 00 2. 00 3. 00 3. 00 2. 00 2. 00 3. 00 2. 00 4. 00 3. 00 4. 00 3. 00 4. 00 2. 00 2. 00 -0 .2 7 -0 .6 4 0. 82 0. 27 -1 .4 5 -0 .3 6 -0 .0 9 -0 .5 5 0. 64 1. 45 0. 45 0. 55 1. 09 -1 .0 0 -0 .2 7 1. 21 1. 19 1. 02 1. 70 0. 96 0. 76 1. 02 0. 96 0. 76 0. 69 1. 21 1. 14 0. 84 0. 91 1. 16 A ng er 2. 63 2. 05 1. 32 2. 79 2. 11 2. 21 3. 26 1. 53 2. 89 1. 63 4. 21 2. 68 2. 42 1. 74 2. 11 4. 00 3. 00 2. 00 4. 00 1. 00 1. 00 4. 00 1. 00 4. 00 2. 00 4. 00 4. 00 3. 00 1. 00 0. 00 1. 37 0. 95 0. 68 1. 21 -1 .1 1 -1 .2 1 0. 74 -0 .5 3 1. 11 0. 37 -0 .2 1 1. 32 0. 58 -0 .7 4 1. 37 1. 10 0. 60 1. 51 1. 08 1. 13 1. 09 0. 93 1. 39 1. 15 1. 00 1. 17 0. 88 1. 01 1. 29 ln di ff 3. 67 3. 67 2. 33 2. 67 3. 67 2. 33 1. 67 0. 67 2. 67 1. 67 3. 67 1. 67 3. 67 2. 33 2. 33 1. 00 3. 00 3. 00 2. 00 5. 00 3. 00 1. 00 0. 00 0. 00 0. 00 3. 00 3. 00 4. 00 0. 00 0. 00 0. 44 0. 44 0. 89 1. 56 1. 78 0. 44 0. 89 0. 44 1. 56 0. 44 0. 44 0. 44 1. 78 1. 56 1. 11 S h am e 4. 50 2. 75 1. 50 2. 00 1. 50 2. 50 1. 25 4. 00 2. 75 1. 50 3. 50 2. 50 2. 50 3. 00 3. 75 2. 00 3. 00 2. 00 4. 00 2. 00 2. 00 2. 00 4. 00 2. 00 1. 00 2. 00 2. 00 3. 00 2. 00 2. 00 -2 .5 0 0. 25 0. 50 2. 00 0. 50 -0 .5 0 0. 75 0. 00 -0 .7 5 -0 .5 0 -1 .5 0 -0 .5 0 0. 50 -1 .0 0 -1 .7 5 0. 50 0. 38 0. 50 2. 00 0. 50 1. 50 0. 38 1. 00 1. 25 0. 75 1. 50 1. 50 1. 25 1. 00 0. 75 -0 .4 2 2. 26 0. 74 -0 .0 5 -1 .1 6 -1 .0 0 0. 21 -1 .3 2 1. 89 0. 84 0. 05 -0 .4 2 0. 32 . -2 .6 7 -0 .6 7 0. 67 -0 .6 7 1. 33 0. 67 -0 .6 7 -0 .6 7 1. 33 0. 33 w R (C on ti nu ed ) D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 TA B LE 5 (C on ti nu ed ) ~ ~ ~ ~ E m o ti o n N o v T im e P le a R el e E xp ec C o n d u U rg en E g o C O rh C C h o C C on 1 P ow er A d ju s E x1 In r ~ ~ ~ ~ G ui lt 3. 33 3. 33 1. 67 0. 33 1. 33 1. 67 3. 33 3. 33 2. 33 1. 00 3. 00 2. 00 1. 67 3. 67 2. 33 2. 00 2. 00 3. 00 4. 00 2. 00 3. 00 3. 00 4. 00 1. 00 1. 00 2. 00 2. 00 3. 00 2. 00 1. 00 -1 .3 3 -1 .3 3 1. 33 3. 67 0. 67 1. 33 -0 .3 3 ,0 67 -1 .3 3 0. 00 -1 .0 0 0. 00 1. 33 -1 .6 7 -1 .3 3 1. 56 0. 44 0. 44 0. 44 0. 44 1. 11 1. 56 0 .4 4 1. 78 0. 00 2. 00 1. 33 1. 11 1. 11 1. 11 Pr id e 1. 11 1. 67 4. 44 3. 67 3. 67 2. 78 2. 00 3. 44 1. 89 1. 33 4. 00 2. 89 3. 11 3. 56 3. 33 2. 00 3. 00 4. 00 4. 00 4. 00 4. 00 2. 00 5. 00 2. 00 2. 00 4. 00 4. 00 5. 00 4. 00 5. 00 0. 89 1. 33 -0 .4 4 0. 33 0. 33 1. 22 0. 00 1. 56 0. 11 0. 67 0. 00 1. 11 1. 89 0. 44 1. 67 0. 62 1. 26 0. 99 1. 56 1. 19 2. 02 1. 33 1. 85 1. 85 0. 81 1. 11 1. 70 1. 65 1. 38 1. 56 N o re : Fi rs t ro w : M ea n em pi ri ca l in pu t ve ct or . S ec on d ro w : T he or et ic al p re di ct ed v ec to r. T hi rd r ow : D if fe re nc e be tw ee n em pi ri ca l an d th eo re ti ca l ve ct or s. F ou rt h ro w : S ta nd ar d de vi at io n fo r va lu es in e m pi ri ca l v ec to r. si gn if ie s ca se s in w hi ch d if fe re nc e sc or e is n ot m ea ni ng fu l be ca us e pr ed ic ti on i s “0 -n ot pe rt in en t o r op en ”. A b b re vi ar io m : N ov , n ov el ty ; T im e, w he n di d th e ev en t ha pp en ; P le a, in tr in si c pl ea sa nt ne ss ; R el e, re le va nc e; E xp ec , e xp ec ta ti on ; C o n d u . g oa l co nd uc iv en es s; U rg en , u rg en cy ; E g o C , s el f as c au sa l a ge nt ; O th C , o th er (s ) as c au sa l a ge nt ; C h o C , c ha nc e as c au sa l a ge nt ; C o n r, c on tr ol ; P o w er , Po w er t o co pe ; A d ju s, c ap ac it y to a dj us t; E xr , ex te rn al ; In r, i nt er na l. D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 347 hits with true misses (excluding the dubious misses because there is a very high probability that the information on the appraisal criteria was entered incorrectly). As shown in the row for the totals, the overall percentage of hits is 77.9% (180 first and second hits compared to 51 true misses). However, averaging the accuracy percentages in Col. 6 across all emotion categories yields a mean accuracy percentage of only 65.5%. This difference between the total accuracy percentage and the average percentage across the different categories is due to marked differences in the number of situa- tions per category. Because the accuracy percentage is rather low in some of the categories containing a small number of cases, the average percentage drops. It is difficult to decide whether this lower accuracy is due to the small number of cases or to greater difficulties in predicting the respective categories. One has to assume that the true accuracy of the present version of GENESE lies somewhere between 65% and 80%. In view of the fact that with 14 emotion alternatives one would expect 7.14% accuracy if the system operated on chance level, this result seems quite respectable. Closer inspection of the accuracy percentages for the individual emo- tions shows that the average across the emotion categories is reduced by very low percentages for anxietylwony and feadterror, o n the one hand, and indifferencehoredom, embarrassmentlshame, and guilt feelings, on the other. With respect to the latter group, it is difficult to evaluate the lack of precision in the diagnoses, because o n l y very few cases are involved and the results may not be very stable. However, it is possible that the low performance for indifferencehoredom is due to the fact that the SEC profile for this state is not highly differentiated across the different stimulus evaluation checks (see Table 4). Shame and guilt are among the most complex human emotions and the current prediction profiles might well be too simplistic to differentiate these emotions. The comparatively low correlations between predicted and actually obtained profiles (shown in Table 4 , Col. 8) suggest important divergences between prediction and empirical means. In consequence, it is not too surprising to find low accuracy for these emotions. I n contrast, the abnormally low accuracy percentages for anxiety and fear are quite unexpected. One possible explanation is the rapidity with which fear situations tend to change-particularly due to the occurrence of events that eliminate the danger or due to a re-evaluation of an event or stimulus as less dangerous. Because of the low accuracy in the anxiety and fear cases, the individual data files and particularly the input profiles were closely scrutinised. This qualitative analysis showed that i n many cases subjects entered appraisal results from both the danger anticipation and the resolution part of the emotion process. A concrete example may demonstrate this phenomenon: A man between 41 and 60 years of age describes a situation in which his daughter leaned D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 348 SCHERER over a burning candle in such a way that her hair started to catch fire. The input vector constituted by the answers to the SEC-based questions and the predicted fear vector are reproduced below (see Table 3 for the exact text of the questions corresponding to the vector entries): Question 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Criterion nov tern pie rei exp con urg ego 0 t h cha c o n pow adj ext int Input 5.0 3.0 1.0 0.0 1.0 0.0 5.0 1.0 1.0 5.0 4.0 4.0 4.0 0.0 0.0 Prediction 5.0 4.0 2.0 5.0 1.0 1.0 5.0 1.0 4.0 4.0 2.0 1.0 2.0 0.0 0.0 Difference 0 -1 -1 * 0 0 0 -3 1 2 3 2 * The comparison between the input and prediction vectors shows that for the first 10 questions there is rather good correspondence (the difference for “other responsibility”-question %is due to the cause of the event being exclusively seen in chance factors, which is of course possible in the present case). However, the answers concerning coping potential (control, power, and adjustment) are clearly related to a phase in the continuous appraisal process in which the danger has already passed (e.g. the flames having been extinguished) and the situation is under control. Otherwise it would be difficult to understand that “very intensive” fear results in spite of the strong ability to control and master the event (an input of 4- “strongly” for both control and power) and it being “quite easy” (4) to adjust to the cosequences of the situation. In this situation, fear was probably quickly followed by relief after realising that no serious con- sequences had ensued. Yet, the total situation was stored and referred to under the most prominent and distinctive label-in this case fear. Given that the appraisal results reported by the subject are likely to come partly from the fear phase and partly from the relief phase of this emotion episode, it is not surprising that the expert system does not correctly diagnose the target emotion-in this case fear. Many other similar examples for the anxiety and fear cases could be listed. This probably reflects a tendency of the subjects to respond with respect to the total situation which may be characterised by a rapid change i n the type of emotion-especially in the case of fear which has been empirically shown to be of very brief duration (Frijda, Mesquita, Sonnemans, & van Goozen, 1991; Scherer & Wallbott, submitted; Scherer, Wallbott, Matsumoto, & Kudoh, 1988; Scherer, Wallbott, & Summerfield, 1986; Wallbott & Scherer, 1986). In consequence, some of the lack of accuracy may well be due to the respondents’ tendency to report appraisals from different phases of an emotion situation rather than responding to all SEC appraisal questions with respect to a singular and well-defined slice of time. In addition, further refinement of the prediction profiles is required to improve the predictions D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 349 for fear and anxiety. The correlation between predicted and obtained profiles is comparatively low for fear and particularly anxiety which would account for low accuracy. It is possible, then, that the theoretically predicted profiles for fear and anxiety are quite unrepresentative of reality and require major changes. Alternatively, it is possible that anxiety/wony situations, in particular, are very variable in their appraisal patterns which would imply that no clear-cut prototype profile can be defined. I n this case, i t could be one or more central criteria which determine the special nature of this emotion (one or all of which might be missing from the list of stimulus evaluation checks). Thus, the differentiation might hinge on one or more very central criteria which may not be contained in the list of stimulus evaluation checks or imperfectly measured by the questions. This might imply the need for a revision in the theoretical underpinnings of GENESE, i.e. t h e list of stimulus evaluation checks. At present, t h e second question in the system (see Table 3) requires the subject to indicate whether the emotion-inducing event happened in the past, is about to happen, or is likely to happen in the future. This is not based on a particular stimulus evaluation check but is part of the facets of situa- tions which are part of component process theory (see Scherer, 1984a). This particular facet was added to the prediction profile precisely because of the need to differentiate anxiety and fear, which imply threats of negative outcomes in the future, from other negative emotions. However, it may be necessary to go beyond the straightforward timing issue and include dimensions such as certainty (Frijda, 1987; Roseman, 1984, 1991; Smith & Ellsworth, 1985, 1987; see also Reisenzein & Hofmann, 1990). Although “outcome probability” was added to the prediction table in Scherer (1988) this check was not implemented as a question in the expert system (due to the reasons given above in the description of the expert system design). The present results could be interpreted to show that this appraisal dimension might be a major discriminating factor for fear and anxiety and thus needs to be added to the prediction vectors in the expert system. These considerations demonstrate one of the major uses of the GENESE expert system, providing impetus and direction for theory development. The comparison between predicted and actual appraisal, as well as the precision of diagnosis, should help to identify the points where emotion-specific appraisals are badly represented in t h e theoretical predictions or where appropriate appraisal criteria are still lacking. CONCLUSIONS This paper illustrates how an empirical expert system approach to the study of emotion-antecedent appraisal can go beyond the established paradigms of obtaining correlational evidence between self-report of verbally labelled D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 3% SCHERER emotional experiences and inferred appraisal dimensions. In particular, the study examined the feasibility of using an expert system to empirically test the author’s predictions on emotion differentiation as based on a limited number of stimulus evaluation checks. The results of a first major study reported here demonstrated an accuracy of posr hoc diagnosis that substantially exceeds chance for many of the emotions studied and that lends support to the specific appraisal theory suggested by the author. The present results might well underestimate the actual capacity of the system (and the support for the SEC predictions) as there is some evidence for incorrect input by some users. One particular problem is the reporting of appraisal results from different points in time during the emotional episode which may reflect different emotions (e.g. fear and relief). Because most real-life emotion episodes seem to consist of rapid sequences of changing emotional states (see also Scherer & Tannenbaum, 1986). it is necessary to make the requirement that appraisal reports need to be focused on one clearly defined point or time slice in the emotion episode more apparent to the users of the expert system. One possibility would be to ask the users to segment the recalled emotion episode into several clearly distinguishable segments and to report the appraisal process separately for each of these segments. In this case, GENESE could attempt to diagnose a sequence of emotions rather than an overall state. One of the major sources for possible errors in reporting the recalled appraisal results is the wording of the questions. For example, even the use of the word “consequence” might have the effect of focusing the respondent’s attention on the aftermath of the emotion episode rather than the crucial period of appraisal at the onset of an emotion-eliciting event. This would obviously lead to a reporting of appraisal results from totally different time periods in the emotion episode (and thus render an accurate expert system diagnostic impossible). This problem is one that the expert system approach shares with all other research paradigms in appraisal research that attempt to elicit verbal report of appraisal processes via questionnaires or interviewing. The process of appraisal is clearly non- verbal and probably occurs largely outside of awareness. Thus, the attempt to obtain a verbal report of many fine details from recall of a process that generally occurs in a split second is obviously fraught with many dangers. A particular problem is the conceptualisation of some of the major appraisal dimensions. In the process of developing GENESE it became clear t h a t many subjects had great difficulty in understanding the concept of goal conduciveness (even in the simple formulation used i n question 6, Table 3). In the further development of GENESE much attention will have to be paid to this problem. Providing copious HELP screens that the respondent can call up to get more information about a particular question D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 351 are part of such efforts to avoid noise in the data that is due to incorrect responding to the questions. The discussion of the results has attempted to show how the expert system approach yields precise suggestions as to where the theoretical assumptions need sharpening or modification. For example, the low accuracy for anxiety and fear clearly indicated the need to add a dimension or check likely to capture the future orientation of the respective appraisals. In consequence, the check of “outcome probability” (or certainty) has been added to the revised version of GENESE (and is gwen strong weight). First informal observation of some trial runs seem to show that the accuracy of GENESE in diagnosing these emotions has improved quite dramatically. Further studies like the one described here will be necessary to fine tune the prediction vectors with respect to the question of how many and which specific types of appraisal dimensions are required, and how they should be weighted, to satisfactorily diagnose the emotional states reported by the users of the system (see also Frijda & Swagerman, 1987). Obviously, the expert system approach could provide a principled way of comparing rival appraisal theories and bring about further convergence. The requirement for using this approach in critical experiments opposing different theories is that pertinent questions for the hypothesised appraisal dimensions or criteria can be formulated and that explicit, quantified predictions for an overlapping set of emotion concepts are made by each of the respective theories. In principle, these requirements could be met by most, if not all, of the appraisal theories reviewed in the introduction. Although the present version of GENESE is based on the determination of Euclidian distance in a vector space, it is certainly feasible to implement a configurational, rule-based algorithm if that were to be preferable for a comparison between theories. The automatic computer-based administration of GENESE allows for easy and economical administration of the procedure to large numbers of subjects, providing a high degree of anonymity. In consequence, the system seems to be well suited to collect large sets of data that would allow to base predictions at least in part on stable empirical patterns. Although some scholars in this area seem convinced that theoretical predictions should be made totally independently of empirical evidence, the present author believes that theory development and refinement must occur in a constant interactive process with empirical data collection. Thus, the predictions made on the basis of the stimulus evaluation check notion of the component process model (Scherer, 1984a,b, 1986, 1988) will change as a result of continuous empirical research. Concretely, the empirically found input patterns (as aggregated over many respondents) for the emotions reported in the study above, in so far as errors in answering the D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 352 SCHERER questions can be excluded, will be used to modify the theoretical prediction vectors in the standard version of GENESE. The development and use of the expert system as a tool for refining theory has just started and new ways of making use of the information provided by the system are being explored. Although at present only the experienced intensity of the emotion is to be judged, future versions of the system will contain additional questions on the duration of the emotion episode and expressive and psychophysiological responding. This should allow to examine the relationship between appraisal patterns on the one hand and specific response patterning on the other. One possible use of this procedure might be to determine to what extent theoretical predictions only work in situations characterised by particular response profiles. Such refinement may also help to study the issue of pure vs. blended emotions. More generally, GENESE might allow an empirical access to the issue of whether there are basic or fundamental emotions and how many there are of these. Users can specify new emotion concepts not contained in the “knowledge base” if neither of two attempts at a diagnosis have provided a satisfactory classification. The name given to this state is then associated with the input vector provided for the respective appraisal. One can determine, once a large number of such added emotion concepts has been obtained, which states (as defined by highly similar appraisal vectors) reoccur very frequently and ought to be added to the basic version of the system. In addition, these data allow to study the labelling used for specific appraisal patterns in a more inductive fashion. GENESE also allows us to study individual differences in the appraisal process. Because information about age and gender is obtained, it will be possible, given a large number of respondents, to investigate the effect of these variables on the appraisal patterns reported for specific emotional experiences. More background information could be obtained to refine this kind of analysis. Even more importantly, as the system is able to learn, i.e. modify the appraisal vectors on the basis of the empirical input (see section on the design of GENESE earlier), it is also possible to determine user-specific emotion appraisal patterns. For example, a group of users could be asked to use the system repeatedly over a period of some months, entering each week some of the major emotions that occurred. It would then be possible to compare the resulting matrices, having been adjusted to the empirical appraisal pattern input for each situation in order to find interindividual differences in emotion-antecedent appraisal. This might provide interesting insights into the issue of habitual emotionality and may even lead to a better understanding of moods or affective disturbance. Manuscript received 23 March 1992 Revised manuscript received 15 August 1992 D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 353 REFERENCES Arnold, M.B. (1960). Emotion and personality. Vols 1 and 2. New York: Columbia University Press. De Rivera, J . (1977). A structural theory of the emotions. Psychological Issues, I0 (4), Monograph 40. Ekman, P. (1984). Expression and the nature of emotion. In K.R. Scherer & P. Ekman (Eds), Approaches to emotion. Hillsdale. NJ: Lawrence Erlbaum Associates Inc. pp. 319-344. Ekman, P. (1992). An argument for basic emotions. Cognition and Emotion. 6 , 169-200. Ellsworth, P.C. & Smith, C . A . (1988). From appraisal to emotion: Differences among unpleasant feelings. Motivation and Emotion, 12, 271-302. Folkman, S. & Lazarus, R.S. (1985). If it changes i t must be a process: Study of emotion and coping during three stages of a college examination. Journal of Personality and Social Psychology, 4 8 , 150-170. Frijda, N. (1986). The emoriom. Cambridge University Press. Frijda, N . H . (1987). Emotion, cognitive structure, and action tendency. Cognition and Emorion, I , 115-143. Frijda, N . H . , Kuipers, P., & ter Schure, E . (1989). Relation among emotion, appraisal, and emotional action readiness. Journal of Personality and Social Psychology, 5 7 , 212-228. Frijda, N. & Swageman, J. (1987). Can computers feel? Theory and design of an emotional system. Cognition and Emotion, I , 235-258. Frijda, N . H . , Mesquita, B., Sonnemans, J . , & van Goozen, S . (1991). The duration of affective phenomena, or emotions, sentiments and passions. K . Strongman (Ed.), International review of emotion and motivation. New York: Wiley, pp. 187-u5. Gehm, Th. & Scherer, K.R. (1988) Relating situation evaluation to emotion differentiation: Nonmetric analysis of cross-cultural questionnaire data. In K.R. Scherer (Ed.), Facers of emotion: Recent research. Hillsdale, NJ: Lawrence Erlbaum Associates Inc, pp. 61-78. Izard, C . E . (1977). Human emotions. New York: Plenum. Johnson-Laird, P.N. & Oatley, K. (1989). The language of emotions: An analysis of a semantic field. Cognition and Emotion, 3 , 81-123. Lazarus, R.S. (1968). Emotions and adaptation: Conceptual and empirical relations. In W.J. Arnold (Ed.), Nebraska Symposium on Motivation, Vol. 16. Lincoln, NE: University of Nebraska Press, pp. 175-270. Lazarus, R.S. (1984a). Thoughts on the relations between emotion and cognition. In K.R. Scherer & P. Ekman (Eds), Approaches ro emotion. Hillsdale, NJ: Lawrence Erlbaum Associates Inc, pp. 247-257. Lazarus, R.S. (1984b). On the primacy of cognition. American Psychologisr. 39, 124-129. Lazarus, R.S. & Smith, C . A . (1988). Knowledge and appraisal in the cognition-emotion relationship. Cognition and Emotion, 2 , 281-300. LeDoux, J.E. (1987). Emotion. In F. Plum & V. Mountcastle (Eds), Handbook of physiology. N e r v o u system, Vol. 5 . Higher functions. Washington, DC: American Physiological Society, pp. 419459. LeDoux, J . E . (1989). Cognitive-emotional interactions in the brain. Cognition and Emotion. LeDoux, J . E . . Farb, C.. & Rugiero, D . A . (1990). Topographic organization of neurons in the accoustic thalamus that project to the amygdala. The Journal of Neuroscience, 10, 1 043- 1 054. Leventhal, H. & Scherer. K.R. (1987). The relationship of emotion and cognition: A functional approach to a semantic controversy. Cognition and Emorion. I , 3-28. Manstead, A.S.R. & Tetlock. P.E. (1989). Cognitive appraisals and emotional experience: Further evidence. Cognrrion and Emotion, 3 , 225-240. 3, 267-289. D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 3% SCHERER Mauro, R.. Sato, K., & Tucker, J. (1992). The role of appraisal in human emotions: A cross- cultural study. Journal of Personality and Social Psychology, 62, 301-317. Mees, U . (1985). Was meinen wir, wenn wir von Gefiihlen E d e n ? Zur psychologirhen Textur von Emotionswortern. Sprache und Kognition, I , 2-20, Ortony, A., Clore, G. L., & Collins, A. (1988) The cognitive structure of emotiom. Cambridge University Press. Reiscnrein, R . & H o h a n n , T. (1990). An investigation of dimensions of cognitive appraisal in emotion using the repertory grid technique. Motivation and Emotion, 14. 1-26. Reisenzein. R. & Schdnpflug, W. (1992). Stumpf's cognitivecvaluative theory of emotion. American Psychologist, 47, 34-45. Roseman, I.J. (1984). Cognitive determinants of emotion: A structural theory. In P. Shaver (Ed.), Review of personality and social psychology, Vol. 5. Emotions, relatiomhips, and health. Beverly Hills, CA: Sage, pp. 11-36. Roseman, I.J. (1991). Appraisal determinants of discrete emotions. Cognirion and Emotion, Roscman. I.J., Spindel, M.S., & Jose, P. E. (1990). Appraisal of emotioncliciting events: Testing a theory of discrete emotions. Journal of Personaliry and Social Psychology, 59, 899-915. Scherer, K.R. (1981). Wider die Vernachllssigung der Emotion in der Psychologic. In W. Michaelis (Ed.), Bericht iiber den 32. Kongrefl der Deutrchen Gesellschaji ftir Psychologie in Ziirich 1980. Gottingen: Hogrefe, pp. 304-317. Scherer, K.R. (1982). Emotion as a process: Function, origin, and regulation. Social Science Information, 21, 555-570. Scherer, K.R. (1983). Rolegomina zu einer Taxonomie affektiver Zustande: Ein Komponenten- RozeBModell. In G. U e r (Ed.), Berichr cibrr dm 33. Kongrefl der Deutschm Geselkchafi fir Psychologic in Mainz. Gottingen: Hogrefe, pp. 415-223. Scherer, K.R. (1984a). On the nature and function of emotion: A component process approach. In K.R. Scherer & P. Ekman (Eds), Approaches to emotion. Hillsdale, NJ: Lawrence Erlbaum Associates Inc, pp. 293-317. Scherer, K.R. (1984b). Emotion as a multicomponent process: A model and some cross- cultural data. In P. Shaver (Ed.), Review of Personality and Social Psychology, Vol. 5 . Beverly Hills, CA: Sage, pp. 37-63. Scherer, K.R. (1986). Vocal affect expression: A review and a model for future research. Psychological Bulletin. 99, 143-165. Scherer, K.R. (1988). Criteria for emotion-antecedent appraisal: A review. In V. Hamilton, G.H. Bower, & N.H. Frijda (Eds), Cognitive perspectives on emotion and motivation. Dordrecht: Nijhoff. pp. 89-126. Scherer, K.R. (1993). Neuroscience projections to current debates in emotion psychology. Cognition and Emotion, 7. 1-41. Scherer, K.R. & Tannenbaum, P.H. (1986). Emotional experiences in everyday life: A survey approach. Motivation and Emotion, 10, 295-314. Scherer, K.R. & Wallbott, H.G. (Submitted). Evidence for universality and cultural variation of diflerential emotion response panerning. University of Geneva. Scherer, K.R., Wallbott, H . G . . Matsumoto, D., & Kudoh, T. (1988). Emotional experience in cultural context: A comparison between Europe, Japan, and the USA. In K.R. Scherer (Ed.), Facets of emotion: Recent research. Hillsdale, NJ: Lawrence Erlbaum Associates Inc, pp. 5-30. Scherer, K.R., Wallbott. H.G., & Summerfield, A.B. (Eds) (1986). Experiencing emotion: A crosscultural study. Cambridge University Press. Smith, C.A. & Ellsworth, P.C. (1985). Patterns of cognitive appraisal in emotion. Jownal of Personality and Social Psychology, 48, 81H38. 5 , 161-200. D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 STUDYING EMOTION-ANTECEDENT APPRAISAL 355 Smith, C . A . & Ellsworth, P.C. (1987). Patterns of appraisal and emotion related to taking an exam. Journal of Personality and Social Psychology, S2, 4 7 5 4 3 8 . Solomon, R . C . (1976). The passions. The myth and nature of human emotion. Garden City, NY: Doubleday. Tesser, A. (1990). Smith and Ellsworth's appraisal model of emotion: A replication. extension, and test. Personality and Social Psychology Bulletin, 16, 210-223. Tomkins. S . S . (1984). Affect theory. I n K.R. Scherer & P. Ekman (Eds), Approaches 10 emotion. Hillsdale, NJ: Lawrence Erlbaum Associates Inc, pp. 163-196. Wallbott, H. & Scherer, K.R. (1986). How universal and specific is emotional experience? Social Science I n f o m t i o n . 26, 763-795. Weiner, B. (1982). The emotional consequences of causal attributions. In M.S. Clark & S.T. Fiske (Eds), Affect and cognition. The 7th Annual Camegic Symposium on Cognition. Hillsdale, NJ: Lawrence Erlbaum Associates Inc, pp. 185-209. Weiner, B . (1986). An amibutional theory of motivation and emorion. New York: Springer. Zajonc, R . B . (1980). Thinking and feeling: Preferences need no inferences. American Zajonc, R.B. (1984). On primacy of affect. In K.R. Scherer & P. Ekman (Eds), Approaches 10 emotion. Hillsdale. NJ: Lawrence Erlbaum Associates Inc, pp. 259-270. Zajonc, R.B. & Markus, H . (1984). Affect and cognition: the hard interface. In C . E . Izard, J. Kagan, & R.B. Zajonc (Eds), Emotions, cognition, and behavior. Cambridge University Press, pp. 73-102. Psychologict, 35, 151-175. D ow nl oa de d by [ U ni ve rs it é de G en èv e] a t 02 :2 8 08 J an ua ry 2 01 8 
work_3srazpdo6zbtxm7kkztjahnhlq	----	New artificial life model for image enhancement Alex F. de Araujo¹, Christos E. Constantinou² and João Manuel R. S. Tavares¹ 1 Instituto de Engenharia Mecânica e Gestão Industrial, Faculdade de Engenharia, Universidade do Porto, R. Dr Roberto Frias s/n, 4200-465 - Porto, PORTUGAL. 2 Department of Urology, School of Medicine Stanford University, Stanford, CA, USA. Corresponding author: Prof. João Manuel R. S. Tavares Faculdade de Engenharia da Universidade do Porto (FEUP) Departamento de Engenharia Mecânica (DEMec) Rua Dr. Roberto Frias, s/n, 4200-465 PORTO - PORTUGAL Tel: +315 22 5081487, Fax: +315 22 5081445 Email: tavares@fe.up.pt, Url: www.fe.up.pt/~tavares 1 http://www.fe.up.pt/%7Etavares New artificial life model for image enhancement Abstract In this work, a method to enhance images based on a new artificial life model is presented. The model is inspired on the behaviour of a herbivore organism, when this organism is in a certain environment and selects its food. This organism travels through the image iteratively, selecting the more suitable food and eating parts of it in each iteration. The path that the organism travels through in the image is defined by a priori knowledge about the environment and how it should move in it. Here, we modeled the control and perception centers of the organism, as well as the simulation of its actions and effects on the environment. To demonstrate the efficiency of our method quantitative and qualitative results of the enhancement of synthetic and real images with low contrast and different levels of noise are presented. Obtained results confirm the ability of the new artificial life model for improving the contrast of the objects in the input images. Keywords: Image Processing; Image Enhancement; Artificial Intelligence; Artificial Life Model. 1. Introduction The acquisition, transmission and compression process of the images can damage them, making it difficult, for example, to locate and extract information of the represented objects. Image enhancement techniques have been developed to soften the effect of this damage. These techniques aim to improve the quality of the images and the contrast between the represented objects, highlighting their most significant features and improving the visual perception of relevant features. Moreover, they allow the images to be represented more appropriately for further analysis by computational methods. There are many factors that can contribute to the loss of contrast, such as 2 loss of focus, noise presence, reflexes, shadows and insufficient illumination of the environment during the image acquisition process [Kao et al., 2010]. The enhancement of degraded images can be improved using different techniques that can be divided into techniques based on spatial domain, the ones that manipulate directly the pixels in the damaged image, and the techniques based on the frequency domain, that work with the image frequency information, generally obtained by Fourier and Wavelets transforms [Hanmandlu et al., 2003, Trifas et al., 2006]. The image enhancement techniques applied on the spatial domain are widely used and try to manipulate directly the image pixels, performing computational operations on these pixels based on their values. For example, considering an image in gray scale, where the intensities can have values from 0 (zero) to 255, we verify that the intensities of the original image use only a reduced range of these values. So, the enhancement methods on the spatial domain try to redistribute the intensities in a way that they can use a wider range of values, improving the perception of the different objects in the image. Some of the traditional techniques applied on the spatial domain are based in the histogram equalization, normalization, quadratic enhancement, square rooted enhancement and logarithmic enhancement [Gonzalez and Woods, 2006]. The enhancement in the frequency domain has its techniques based on the convolution theorem and can be interpreted as an enhancement extension in the spatial domain, with its operations applied to the image frequencies, usually obtained by the application of a transform, as the Fourier Transform. After the frequencies manipulation, an inverse transform is applied so that the image can be represented again in the spatial domain. This kind of enhancement method is generally obtained through filtering techniques, as the high-pass filter [Gonzalez and Woods, 2006]. Recent studies have been adopting models based on artificial life to process tasks of the computational image processing and analysis obtaining effective results. Examples include brain segmentation in magnetic resonance images [McInerney et al., 2002, Farag et al., 2007, Feng et al., 2008, Prasad et al., 2011a, Prasad et al., 2011b], and features extraction and classification from medical images [Srinivasan and Sundaram, 2013]. The use of these artificial life models, better 3 known in computational image processing, tries to make a deformable “living” model provided with a "primitive brain" capable of making decisions in search of the best results in the segmentation process. The objective of this paper is to present a new method based on artificial life to image enhancement based on the modelling of an organism, in particular, its control and perception centers. On the contrary of what is common in image segmentation processes using artificial life models, our model is not inspired in the organism’s shape, that is, the geometry of the organism’s body, but in its behavior when it is located in a certain environment and performs the food selection process. With the proposed enhancement method, it is intended to enhance images with low contrast and images affected by noise interference, highlighting the transitions between presented objects, and at the same time avoiding enhancing the noise that affects the original image quality. The results of the proposed enhancement method using the developed model allow us to conclude that this method is promising, being capable of considerably improving the visual perception and the quality of the affected images. Another contribution of this work is the innovating utilization of artificial life models as enhancement image techniques. This paper is organized as follows: in the next section, a review about existing artificial life models is presented. In section 3, the proposed model is described. Experimental results are presented and discussed in section 4, following the obtained conclusions and suggestions for future work. 2. Review about methods based on Artificial Life The studies related with computational image processing and analysis try to develop methods capable of manipulating complex images with different features, in an efficient, robust and reliable way. Various computational methods to different tasks can be found in the literature, such as for noise removal [Chen et al., 2010, Cao et al., 2011, Zhang et al., 2010a], enhancement [Cheng and Xu, 2000, Cheng et al., 2003, Lai et al., 2012, Cho and Bui, 2014], segmentation [Xue et al., 2003, 4 Nomura et al., 2005, Ma et al., 2010a], tracking [Pinho and Tavares, 2009, Nguyen and Ranganath, 2012], registration [Oliveira et al., 2011, Oliveira et al., 2010] and recognition [Papa et al., 2012]. Besides the diverse tasks that can be made by these methods, many techniques have been used, as Genetic Algorithms [Wang and Tan, 2011, Hashemi et al., 2010, Yi Kim and Jung, 2005], Artificial Neural Networks [Petersen et al., 2002, Shkvarko et al., 2011], Active Contours [Ghita and Whelan, 2010, Ma et al., 2010b], Region Growing [Fan et al., 2005, Peter et al., 2008], models based on differential equations and finite elements [Chao and Tsai, 2010, Yu et al., 2010]. A research field based on artificial life has emerged in the computer graphics area for modeling animation behavior and the evolution of plants and animals [Kim and Cho, 2006]. This investigation area has also formed the basis for new researches in computational image processing and analysis [McInerney et al., 2002, Farag et al., 2007, Feng et al., 2008, Prasad et al., 2011a, Prasad et al., 2011b, Osuna-Enciso et al., 2013, Horng, 2011], forming the methods called image processing methods based on artificial life models. The artificial life models address the diverse biological processes that characterize the live species in the attempt to overcome the problems found in the image processing, such as the complexity of the represented objects and low contrast between them and the rest of the represented scene. The majority of the artificial models used in this area apply techniques based in physical and geometrical principles, aiming to make the used deformable models [Ma et al., 2010b] more flexible and capable of performing a better image exploitation, using a priori knowledge of the area associated with the images, and analyzing better the neighborhood of the active model used to control the deformation in a more adequate and robust way. Making an analogy with the artificial life systems, the geometry of the deformable model is generally considered as being an "organism" (a worm), that keeps changing according to the available priori knowledge, associated with the information that its sensorial organs return while the organism changes its shape and moves through the image (or environment). Despite being the most common utilization of the artificial models in image processing, there are many biological processes that can be used as basis for the same 5 models, such as growing, natural selection, evolution, locomotion and learning [Kim and Cho, 2006, McInerney et al., 2002]. Another important issue is that the models based on artificial animal life are more relevant to the image processing techniques, because they present bodies with motor and sensorial organs, and mainly a brain with learning (cognitive system), perception and behavioral centers [McInerney et al., 2002, Horng, 2011]. The motor center coordinates the muscular actions to perform specific actions, as locomotion and the control of the sensorial center. This last one is used so that the artificial organism acquires information about what happens around it and sends this information to the cognitive center, which is going to decide which actions must be performed to obtain best results according to each situation. The perception center is the part of the brain that allows the organism to have its sensors trained, being formed by mechanisms that allow it to acquire and understand sensorial information about the environment that surrounds it. This behavior is useful to perceive the modifications that occur during the processing, since the individual actions can have significant changes in the environment. The learning center allows the individual to learn how to control all his organs and his behavior through practice. The bahavioral center has the routines with the actions that the individual must perform, considering their perception. For the creation of an artificial life model, the work-tasks pyramid shown in Figure 1 is usually followed. (Insert Figure 1 about here) Observing the pyramid in Figure 1, from bottom-up, there is the organism’s geometric modeling in its base. In this layer, the organism’s type is defined, as well as its appearance and geometric morphology. In the second layer, there is the physical modeling where the biomechanical principles are modeled for simulation and formation of the biological tissues, such as muscles. The next layer incorporates the motor control routines, which are responsible for stimulating the muscle actuators 6 to allow the locomotion movements. In the fourth layer, there is the behaviour and perception modeling, with an important part, for example, in detecting and navigating between obstacles. And on the top of the pyramid, the cognitive layer, which is responsible for controlling the acquired knowledge by the organism during the learning process, as well as the planning of the actions to be performed with some level of intelligence. 3. Proposed Model The computational image processing based on artificial life models uses an analogy between biological processes from certain organisms and the desired operations to do on the images being processed. One of the objectives of these models is to automate the applied computational methods, making them more robust, efficient and automated to deal with the existing variations in the involved image set. The organisms used in the models have features that deserve highlighting: they are endowed with a prior knowledge about some features of the environment, namely from the scene, in which they are found; they have sensors that operate in making decisions about the actions to be performed during processing; they have a set of actuators that in association with a knowledge system, allows the organism to adapt to the environment according to the collected data from the sensors. Some models based on artificial life, such as the ones based on worm, called deformable organisms, have been successfully used in the image segmentation; as in, for example, [McInerney et al., 2002, Farag et al., 2007, Prasad et al., 2011b]. Usually these models have an analogy to worms, considering their great flexibility and their subsequent facility to deformation. Thus, each organism has a "skeleton" defined to describe its initial shape. When the organism is placed in an image, it starts a search for a compatible region with the geometry of its skeleton and with the a priori knowledge that it has about the input scene. When a compatible object is found, a set of sensors and 7 actuators start to work, deforming the worm’s "skeleton" evolving this “skeleton” in order to fully enclose the object’s region in the input image. Although this technique of artificial intelligence has been wide used for image segmentation with considerably success, it has not been so common for image enhancement. Hence, we extend it use to enhance the contrast of images by proposing a new artificial life model. The adopted model is not based on analogy with the organism’s body shape, but on the behaviour of a herbivore organism when it is in a specific environment and performs the process of selection of its food. The organism’s body shape was not considered in this approach because what matters here is the effect that the behaviour of this organism produces in the environment, and the shape of its body has no direct influence on the operations considered in this work. In Figure 2, there is the flow diagram of the proposed method, with the steps of the method described in detail in the following sections. (Insert Figure 2 about here) 3.1. Model Description The low contrast images problem can be reduced using computational methods for image enhancement, making the difference between different objects more enhanced, allowing us to obtain better results in the later stages of image processing and analysis, such as segmentation and features extraction. Considering a grayscale image, it is desired that the pixels with low and high intensities are represented so that their intensities are as distinct as possible. Based on this principle, a good enhancement method in the spatial domain is the one able to verify all pixels of an original image, and recalculate their intensities taking into account the neighbours so that the intensity differences between the pixels belonging to neighbouring and distinct regions is increased. Thus, the model based on artificial life used in the proposed image enhancement method was inspired by the behaviour of the herbivorous organisms when they are in a specific environment and perform 8 the selection process of its food. To this end, it was considered a natural pasture as environment. These considerations have resulted from the observation that while feeding on a natural pasture, these organisms make a selection of the food, setting a "priority order" of the food to be eaten first. If we consider an area composed by a single type of food with different heights, there is a tendency that the smaller food will be eaten first. Usually this happens because the smaller parts are smoother and more nutritious than the bigger ones. Furthermore, the organism is not fixed in a single point until it eats all of the food available there. In general, it is always moving in the environment while feeding itself, guided by its cognitive system using the information collected about the food around it. Thus, the difference between the small and big foods present in a given environment tends to become more evident, at the same way as it is desired in the image enhancement operation. In the built model, the organism moves iteratively on the image until a threshold for the height of the food is reached. 3.2. Modeling the control center The control center has been modeled from the movement operations of the organism, and reduces the height of the food as a result of the eating process performed by organism. Two control operations are defined in this step: moving and reducing the height ("eat"). The movement consists in moving the organism from one point to another one on the image, avoiding passing through the same point more than one time at the same iteration. In each iteration, all pixels of the image are visited and their intensities are recalculated. To simulate how the organism feeds, the intensity of each point is reduced by considering the following assumptions: usually the organism eats a bit of food that is in the region where it is, but it does not always eat all the food at the same time; the smaller foods are fresher and more nutritious, and they have priority to be consumed. These assumptions are part of the a priori knowledge set that the organism has about their actions and the environment. So, considering a local analysis on the image, the smaller its current value, the bigger 9 the reduction of the intensity value of each pixel; i.e., in a very bright pixel a small reduction occurs while in darker pixels a greater reduction, in proportion to its size, is performed in each iteration. To model this behavior, we adopted the following rational function: ,)( 2 kx kx xf + = (1) where 255...,,1,0=x is the possible intensity values for pixels in the 8-bits images, and k is a positive constant used to control how much of the food height should be reduced at each iteration. The higher the value of k, the faster the food is reduced and less time is spent to reach the stop criterion. However, very high values for k cause similar regions to merge quickly, and distinct objects may be wrongly joined. On the other hand, if its value is too low, the growth of the food can be greater than the amount that is eaten, so the different objects represented in the image can be joined to the image foreground. As the intensities of the pixels in the input images range from 0 to 255, k was defined as equal to the maximum possible value, i.e., 255=k . As such, if the intensity of a pixel is 0 (zero), the reduction is also 0 (zero), but if the intensity is 1 (one), then the reduction is almost equal to the total height associated to the pixel. This behavior is repeated for all pixels with low intensity, and for pixels with high values, the reduction is smaller. It should be noted that, in the images considered in this work, the pixels belonging to the regions of interest are almost black. For this reason, it was considered that the best “food” (i.e., the darkest pixels) has the maximum possible height, which was experimentally considered as equal to 10% of the maximum available height. The function of Equation 1 generates a curve which shows a more significant loss for the pixels with smaller intensity, while those with bigger intensities are less affected, Figure 3. Thus, the darkest pixels tend to darken faster than the lighter ones, increasing the difference between them and, consequently, improving the enhancement of the image associated area. (Insert Figure 3 about here) 10 3.3. Modeling the perception center The perception center was modeled according to the a priori knowledge of the organism's behaviour, considering the information about the neighbourhood acquired by the organism. The perception center will act primarily to receive the height information of the neighboring points of the organism; i.e., the intensity level, and which points have been visited in a given iteration. This information is important to choose the way for the organism to go. To determine how to move in the environment, all 8 (eight) pixels neighboring the pixel occupied by the organism are visited, and at the same time that the intensity of the visited pixel is reduced, the organism stores the coordinates of the pixel with smaller intensity, which is added to its path at the end of the visit to the neighboring pixels. In addition, the visited pixels are marked as visited, being released for further visits after the end of the current iteration. This operation allows the organism to walk in a coordinated way in the image, following the darker regions. However, it may happen that the organism is isolated in a pixel whose all the neighbors have already been visited. In this case, the organism carries out a search for the nearest pixel that has not yet been visited and continues processing. This search for the nearest pixel ensures that all pixels on the image are visited, and have their values updated at each iteration. 3.4. Effects of the organism’s actions in the environment The organism walking process in the environment following the points with best food generates a secondary effect, which is the damage to the bigger foods during the movement process, since the organism moves over them. This effect is much common because the big foods are less flexible and are frequently broken when they are pressed, and the broken parts are usually discarded because when food is damaged, it loses its nutritive capability, and it will hardly be consumed by the organism. 11 Another interesting effect that happens with the adopted model is a consequence of the time that the organism needs to eat all the food available. Under the considered conditions in the used analogy, many days are usually needed for this process to be performed. So, as the food is a natural herb, the pieces not yet fully eaten tend to grow as time goes by. However, very big foods and even those small ones tend to be broken rather than grow, and the high ones break due to their low flexibility, as already previously explained, and the small ones do too because they are young and have low resistance. Thus, only those foods with middle height usually grow effectively. So to simulate these effects the function of Equation 2 was adopted, because this one represents an undulating curve, ranging from the approximate range of ],[ 21 yy− , where 1y and 2y belong to the set of real numbers, as can be seen from the curve of Figure 4. In that curve, the values 1y and 2y are the smallest and the biggest values along the Y-axis, respectively. This curve allows the performance of an approximate simulation of the degrading operations of the big and small foods and the growth of the intermediate foods according to the course of time: i4 2 sin 3 180 4 f( ) . 180 i i x x x π π π π   +   = (2) (Insert Figure 4 about here) 3.5 Stopping Criterion In the implemented method, we tried to integrate an automatic stopping criterion based on a threshold defined from the height of the food present in the environment. For this purpose, while the organism moves in the environment, the heights of the foods are summed and compared with the sum of the heights in the input image, and the processing is finished as soon as at least 30% of the 12 height of all foods has been eaten. This threshold was defined experimentally for the images to be studied. 4. Results and Discussion To test the efficiency of the proposed model, it was necessary to take images where it is possible to quantitatively measure the loss of contrast in the output image when compared with the ideal image, as well as the restoration rate obtained from the application of the proposed enhancement method. For this purpose, 8 images were used for testing, where 4 grayscale images were synthetically created and the other 4 were real ones, including the traditional images "Lena" and "Cameramen". After that, each image was damaged twice, resulting in 16 grayscale images with changed contrast. Both reduction contrast processes were applied over the original images, and in each of them a Gaussian noise was added with intensity equal to 0.3, and in one of these reduction contrast processes a circular median filter was applied with radius equal to 2 for blurring the image before applying the noise. The results of the image enhancement obtained from the images with the contrast affected were visually and statistically analysed based on the analysis of the indexes returned by PSNR (Peak Signal Noise Ratio), SSIM (Structutal Similarity) and Detail Variance and Background variance (DV-BV) metrics. The PSNR and SSIM indexes were calculated comparing the original image and the images resulting from the controlled degradation process. In addition, we analyzed the profile of a random line of the images, and compared the proposed method with other enhancement methods traditionally used in the spatial domain: normalization, histogram equalization, square enhancement, square root enhancement and logarithmic enhancement methods [Gonzalez and Woods, 2006]. 4.1. Adopted evaluation metrics 13 Generally, the comparison between two images is a natural task for the human visual system, but the realization of this same task for a computer system is more complex and not so natural. However, there are many studies attempting to provide techniques able to compare images, including techniques to statistically compare the performance of image processing methods [Wang et al., 2004, Ramponi et al., 1996]. 4.1.1 Analysis based on error The indexes of comparison based on error try to estimate the perceived error between the processed image and the original image to quantify the image degradation. The techniques of this category have as their main disadvantage the fact that they can fail in cases where translations in the images happen. For example, similar images where one of the objects has been displaced can be classified as distinct by this process. Moreover, these indexes perform the comparison based on the variation of the intensities of the pixels in the images, and images with different types of distortion can have similar indexes [Wang et al., 2004]. Despite of that, the indexes based on error are often used to compare the quality of image enhancement [Hashemi et al., 2010, Shkvarko et al., 2011, Ghita and Whelan, 2010] and image smoothing [Chen et al., 2010, Jin and Yang, 2011] methods, due to their simplicity and the fact that these images are usually affected by a few displacements during the computational processing. Some of the most known indexes based on error are the PSNR (Peak Signal Noise Ratio), RMS (Root Mean Square), MSE (Mean Squared Error) and ReErr (Relative Error) [Wang et al., 2002]. However, the most commonly used to analyze the performance of the restoration and enhancement methods is the PSNR, which attempts to calculate the relationship between the highest possible force strength of a signal (in the case of the image it is the highest intensity value) and its strength affected by noise [Dash et al., 2011, Yang et al., 2009]. In this case, the PSNR is represented in 14 function to a logarithmic scale on the base 10 (decibel), because some signals have a very high value. The PSNR can be calculated from the MSE, which can be defined as: [ ]∑ ∑ −= − = − = 1 0 21 1 ),(),( 1 m i n j r jiIjiInm MSE , (3) where m and n are the dimensions of the input image, I is the original image, and rI is the affected image by processing or by any artifact. From the MSE index, the PSNR can be calculated as follows:       = MSE MAX PSNR I 2 10log10 , (4) where IMAX is the maximum value that a pixel can be. To the grayscale images represented by 8 bits, for example, 255=IMAX . During the interpretation of the PSNR index, the higher its value, the more similar are the two compared images. It is important to note that for identical images, the MSE index value will be zero, and thus the PSNR is undefined. 4.1.2 Analysis based on structural information This approach attempts to note the changes in the structural information of the image to quantify the occurred degradation. The analysis of the structural information assumes that the human vision system is adapted for extracting structural information from what is seen, searching for changes in these structures to detect changes and consequently, possible degradation generated by some process [Wang et al., 2002]. The SSIM index (Structutal Similarity) is an index of this class most often used to analyze the quality of computational methods for image processing [Zhang et al., 2010b, Chen et al., 2010]. The SSIM index was proposed by Wang and colleagues [Wang et al., 2004] in an attempt to avoid that the images with very different visual qualities have the same index, as can happen in the 15 indexes based on error. This index is calculated based on three components of comparison: luminance, contrast and image structure. The first parameter is calculated from the average intensity of the signal, in this case the input image. The contrast is calculated from the standard deviation, and the structure parameter is computed from the normalized image using the standard deviation of the same image. So, the SSIM index can be obtained using the equation: ,),(.),(.),(),( γβα yxsyxcyxlyxSSIM = (5) where, 0>α , 0>β and 0>γ are constant parameters used to define the weight of each component in the final index. The component l refers to the luminance, c to the contrast and s to the structure. The three components are relatively independent, and changes in one of them do not affect the other ones. In [Wang et al., 2004] a more detailed analysis of each component is presented, showing in detail how they are calculated. The SSIM value shows an index for each pixel of the image and, to make its use more practical, it is common to use a mean SSIM index, also called MSSIM, which is calculated from the average of the elements obtained from SSIM. For equal images, this index is equal to 1 (one positive), being reduced as the images differ, reaching the value -1 (negative one) for two images exactly opposite (one image is the negative of the other one). 4.1.3 Detail Variance and Background Variance values The Detail Variance (DV) and Background variance (BV) values are used to give indications about the level of enhancement in an image, without necessarily using the other image as a benchmark. These values are calculated from the local variance of the n neighbors of all image pixels, creating a matrix of variances in a first step. Afterwards, each pixel is classified into two classes: the variance of each pixel is compared to a threshold, and if it is greater than the threshold value, the pixel is classified as belonging to the image foreground, otherwise the pixel belongs to the image background. 16 After classifying all pixels of the image, the mean variance of the pixels belonging to two classes are calculated, and respectively called DV and BV. To check the level of enhancement applied to the image, these two values are analyzed as follows: if the DV index of the processed image increases when it is compared to the DV index of the input image, while the BV value suffers a little change, it is considered that the image was efficiently enhanced. In this work, we adopted a neighborhood of n n× neighbors, and the threshold was calculated using the Otsu method [Osuna- Enciso et al., 2013]. 4.2. Experimental Tests In Figures 5, 6 and 7, the results obtained by applying the square root enhancement method are presented (c); quadratic enhancement method (d); logarithmic enhancement method (e); normalization method (f); histogram equalization method (g); and proposed enhancement method (h), on the input image with reduced contrast (b). In each figure, the image (a) represents the original image and the images together with the symbol (-) represent the difference between the processed images and their respective original images, for example, the image (b-) is the difference between the images (b) and (a), (c-) is the difference between the images (c) and (a), and so on. Note in these figures that the proposed method (figures 5h, 6h and 7h) allowed to enhance properly the objects represented in the input images (figures 5b, 6b and 7b, respectively), restoring the transition details between such objects, as can be seen with more accuracy in image (h) of Figure 5. In addition, our method allowed to maintain the transitions in the areas with linear transitions of the gray levels, as noted in image (h) of Figure 6. Analyzing the difference between the processed images (b-, c-, d-, and f-) and their respective original images, it is possible to note that the proposed method returned images more similar to their respective original images (figures 5a, 6a and 7a). Note that the difference between two identical images generates an image in black (image 17 a-), i.e., the darker the image resulting from the difference, the more the processed image is similar to the original image, indicating a better result. (Insert Figure 5 about here) (Insert Figure 6 about here) (Insert Figure 7 about here) In Figure 8, the results of the same enhancement methods applied in 4 real images are presented, including images "Lena" and "Cameramen" with contrast affected by the addition of synthetic Gaussian noise. The respective PSNR and SSIM indexes for the 4 images are presented in Tables 1 and 2, in lines 13, 14, 15 and 16, respectively. It is possible to note from the analysis of the indexes and the images of Figure 8 that the proposed method obtained promising results, being able to restore the contrast in the input images. (Insert Figure 8 about here) To complement the comparison of the results, a quantitative analysis of the results was made using the PSNR and SSIM indexes. These indexes were obtained from the comparison between the original images and their respective images resulting from the application of the enhancement methods, Tables 1 and 2. In these tables, the values obtained for the synthetic images shown in Figs. 5–7 are indicated in bold. 18 (Insert Table 1 about here) (Insert Table 2 about here) The PSNR index allows the performance of an analysis based on error, showing the ability of our enhancement method to restore the information intensity of the pixels of the degraded image. The PSNR indexes for different methods of enhancement tabulated in Table 1, and plotted in Figure 9 allow us to check the best performance of the proposed method. From the presented graph, it is possible to clearly note that the results of the new method were better for all images from the test set, since the PSNR index had the highest values for this method. (Insert Figure 9 about here) The SSIM index shows the ability of the processing method to preserve the structural information of the processed image. The graph of Figure 10 has been plotted from the data presented in Table 2, and together with the analysis of Figure 9, it allows us to conclude that our method was able to enhance the input images, preserving the structural information of the processed images better than the other methods presented in the literature, as indicated by the highest values for the SSIM index. (Insert Figure 10 about here) To complement the analysis of the proposed enhancement method, the profile of a line of the original images, affected and resulting from the application of the method was extracted and plotted in the graphs of Figure 11. In this figure, the profiles of a random row of images of the images of 19 Figures 5, 6 and 7, respectively, are presented. The graph of the line profile for each image has three components: the blue component for the line profile in the original image; the red line for the affected image; and the green one for the image result of the proposed method. The extraction of the line profile showed that the image affected had the line profile very irregular due to interference caused by damage in the original image with the addition of noise. In addition, the line profile of the image presented in Figure 5 shows the smoothed transition due to the blurring operation. Through the analysis of these graphs, it is clear that the method has greatly approximated the input image to the original image, reducing the smoothing transitions between objects and also reducing the interference caused by noise. This behavior was observed for almost all images studied. (Insert Figure 11 about here) Another objective analysis that can be made about the results of the image enhancement methods is through the DV and BV values. In Table 3 the DV and BV values for the 4 images of Figure 8 (Cup, Lena, Cameramen and Parrot) are presented. In the second column of the table, the values of the original images are indicated in bold as are the values that should be approximated after the application of the enhancement methods under comparison. It is possible to note from these data that the proposed enhancement method obtained the best results for three of the four analyzed images (Cup, Lena and Cameramen), returning images with high DV values and BV values close to the respective values in the original image. (Insert Table 3 about here) To better visualize the results of this analysis, the points classified as belonging to detail value (DV) of the four images of Figure 8 were marked in black color and the images are presented in Figure 20 12. Note that the detected details from the images processed by the proposed method (h) are closer to the detail detected from the original images (a) indicating that the proposed method performed a more efficient processing than the other methods tested. (Insert Figure 12 about here) In Figure 13, we have the result of the proposed method and the enhancement methods studied in this paper applied in real images. It is observed that the resulting images of the enhancement proposed method allowed a more efficient segmentation of the images, due to the enhancement of the transitions between the inner and the outer regions of the objects. It is noted from the edges extracted by the Chan-Vese method [Chan and Vese, 2001]. In the ultrasound image of the pelvic cavity, it is noted that a greater part of the bladder was detected in the image enhanced by the proposed method (image h). In image "Cameramen", it is also realized that the edges extracted from the image enhanced by the proposed method is more regular, and the objects are better involved. In the image of the skin lesion, an edge with less discontinuity was obtained, even though the lesion was in an area affected by the strong presence of hair. This work presents a pioneer application of an artificial life model for image enhancement. The results obtained using both synthetic and real images exposed that one advantage of the proposed model is a superior enhancement of the image transitions, making the structures presented from the image background more distinguishable, and the higher preservation of the structural information presented. Another advantage of the model developed is that it is based on the behavior of the organism mimicked instead of being based on the organism morphological characteristics, as happens with the deformable models commonly used in image segmentation. Although the evident advantages, the presented model can still be improved, for example, by integrating more 21 sophisticated cognitive and sensorial centers that will increase the efficiency and robustness of the enhancement process. (Insert Figure 13 about here) Being an iterative method, the proposed method requires a higher computational time than the other methods discussed, Table 4. In that table, the times required to enhance the images "Lena" and "Cameramen" are indicated, both images are in grayscale and with size equal to 256x256 pixels. The tests were performed using a core of an Intel Core i5 processor at 3.2 GHz, 8 GB of RAM and 64-bits bus. The time for each method was obtained by averaging 150 executions, in order to increase the accuracy of the indicated values. (Insert Table 4 about here) 5. Conclusions The proposed method for image enhancement in the spatial domain is based in a new artificial life model, which processes the original image using operations based on the pixels’ values in order to highlight the intensity differences between neighbour regions. In comparison to other well known enhancement methods presented in the literature, the suggested method showed very promising experimental results. A quantitative and qualitative analysis of experimental results concluded that the proposed method is able to improve the transitions between the objects presented in degraded images, and got better results than the traditional enhancement methods adopted and compared in this work. In addition, tests on real images, such as ultrasound of the pelvic cavity, "Cameramen" and the image of a skin 22 lesion, showed that the proposed method is capable of enhancing the transitions between the structures present in these images, increasing, for example, the efficiency of the segmentation methods. The analysis of the results obtained by the proposed model confirmed that, as in several applications of image segmentation, the artificial life models can accomplish acceptable results in applications of image enhancement, suggesting these models also for the field of computational image processing. The effective image enhancement makes the image segmentation step easier and more accurate as it was demonstrated by using the Chan-Vese segmentation method on enhanced images. Additionally, the characterization and classification of structures from images can be more robust, once these tasks are highly dependent on the quality of the image segmentation step. Moreover, the experimental results obtained using the proposed image enhancement method allow us to conclude that the new artificial life model is very interesting for ultrasound imaging, once the ultrasound images are commonly affected by artifacts during the acquisition process, and the transitions between the structures presented are attenuated. Hence, the method proposed can effectively contribute for a more successfull ultrasound image processing and analysis by emphasizing the transitions between the structures presented and reducing the artifacts. In future studies, we intend to integrate the modeling of the cognitive center to the organism of the proposed model, giving it the ability to make smarter and elaborated decisions, as well as add some other control parameters of the environment, such as its relief, which can influence in the way that the organism moves in the enviroment. We also intend to extend the proposed method to improve the enhancement of image sequences using the relationship between consecutive frames. Furthermore, we want to apply techniques of high performance computing to reduce the computational time of the proposed method, such as parallel computing and processing using multiresolution. Acknowledgments: 23 The first author would like to tanks Fundação para a Ciência e Tecnologia (FCT), in Portugal, for his PhD grant with reference SFRH/BD/61983/2009. This work was partially done in the scope of the project “A novel framework for Supervised Mobile Assessment and Risk Triage of Skin lesions via Non-invasive Screening”, with reference PTDC/BBB-BMD/3088/2012, financially supported by Fundação para a Ciência e a Tecnologia (FCT) in Portugal. References [Cao et al., 2011] Cao, Y., Luo, Y., and Yang, S. (2011). Image denoising based on hierarchical markov random field. Pattern Recognition Letters, 32(2):368–374. [Chan and Vese, 2001] Chan, T. and Vese, L. (2001). Active contours without edges. IEEE Transactions on Image Processing, 10(2):266 –277. [Chao and Tsai, 2010] Chao, S.-M. and Tsai, D.-M. (2010). An improved anisotropic diffusion model for detail- and edge-preserving smoothing. Pattern Recognition Letters, 31(13):2012–2023. [Chen et al., 2010] Chen, Q., Sun, Q., and Xia, D. (2010). Homogeneity similarity based image denoising. Journal of Pattern Recognition, 43(12):4089–4100. [Cheng and Xu, 2000] Cheng, H. and Xu, H. (2000). A novel fuzzy logic approach to contrast enhancement. Pattern Recognition, 33(5):809–819. [Cheng et al., 2003] Cheng, H. D., Xue, M., and Shi, X. J. (2003). Contrast enhancement based on a novel homogeneity measurement. Pattern Recognition, 36(11):2687–2697. [Cho and Bui, 2014] Cho, D. and Bui, T. D. (2014). Fast image enhancement in compressed wavelet domain. Signal Processing, 98(0):295 – 307. 24 [Dash et al., 2011] Dash, R., Sa, P. K., and Majhi, B. (2011). Restoration of images corrupted with blur and impulse noise. In Proceedings the International Conference on Communication, Computing & Security, volume 1, pages 377–382. [Fan et al., 2005] Fan, J., Zeng, G., Body, M., and Hacid, M.-S. (2005). Seeded region growing: an extensive and comparative study. Pattern Recognition Letters, 26(8):1139–1156. [Farag et al., 2007] Farag, A. A., Suri, J. S., Micheli-Tzanakou, E., Hamarneh, G., and McIntosh, C. (2007). Deformable organisms for medical image analysis. In Deformable Models, Topics in Biomedical Engineering. International Book Series, pages 387–443. Springer New York. [Feng et al., 2008] Feng, J., Wang, X., and Luo, S. (2008). Medical imaging and informatics. chapter A Worm Model Based on Artificial Life for Automatic Segmentation of Medical Images, pages 35–43. Springer-Verlag. [Ghita and Whelan, 2010] Ghita, O. and Whelan, P. F. (2010). A new gvf-based image enhancement formulation for use in the presence of mixed noise. Pattern Recognition, 43(8):2646– 2658. [Gonzalez and Woods, 2006] Gonzalez, R. C. and Woods, R. E. (2006). Digital Image Processing (3rd Edition). Prentice-Hall, Inc., Upper Saddle River, NJ, USA. [Hanmandlu et al., 2003] Hanmandlu, M., Jha, D., and Sharma, R. (2003). Color image enhancement by fuzzy intensification. Pattern Recognition Letters, 24(1-3):81–87. [Hashemi et al., 2010] Hashemi, S., Kiani, S., Noroozi, N., and Moghaddam, M. E. (2010). An image contrast enhancement method based on genetic algorithm. Pattern Recognition Letters, 31(13):1816–1824. [Horng, 2011] Horng, M.-H. (2011). Multilevel thresholding selection based on the artificial bee colony algorithm for image segmentation. Expert Systems with Applications, 38(11):13785 – 13791. 25 [Jin and Yang, 2011] Jin, Z. and Yang, X. (2011). A variational model to remove the multiplicative noise in ultrasound images. Journal of Mathematical Imaging and Vision, 39(1):62–74. [Kao et al., 2010] Kao, W.-C., Hsu, M.-C., and Yang, Y.-Y. (2010). Local contrast enhancement and adaptive feature extraction for illumination-invariant face recognition. Pattern Recognition, 43(5):1736–1747. [Kim and Cho, 2006] Kim, K.-J. and Cho, S.-B. (2006). A comprehensive overview of the applications of artificial life. Artificial Life, 12(1):153–182. [Lai et al., 2012] Lai, Y.-R., Chung, K.-L., Lin, G.-Y., and Chen, C.-H. (2012). Gaussian mixture modeling of histograms for contrast enhancement. Expert Systems with Applications, 39(8):6720 – 6728. [Ma et al., 2010a] Ma, Z., Jorge, R. N. M., and ao Manuel R. S. Tavares, J. (2010a). A shape guided c-v model to segment the levator ani muscle in axial magnetic resonance images. Medical Engineering & Physics, 32(7):766–774. [Ma et al., 2010b] Ma, Z., Tavares, J. M. R., Jorge, R. N., and Mascarenhas, T. (2010b). A review of algorithms for medical image segmentation and their applications to the female pelvic cavity. Computer Methods in Biomechanics and Biomedical Engineering, 13(2):235–246. [McInerney et al., 2002] McInerney, T., Hamarneh, G., Shenton, M., and Terzopoulos, D. (2002). Deformable organisms for automatic medical image analysis. Medical Image Analysis, 6(3):251–266. [Nguyen and Ranganath, 2012] Nguyen, T. D. and Ranganath, S. (2012). Facial expressions in american sign language: Tracking and recognition. Pattern Recognition, 45(5):1877–1891. [Nomura et al., 2005] Nomura, S., Yamanaka, K., Katai, O., Kawakami, H., and Shiose, T. (2005). A novel adaptive morphological approach for degraded character image segmentation. Pattern Recognition, 38(11):1961–1975. 26 [Oliveira et al., 2010] Oliveira, F. P., Pataky, T. C., and Tavares, J. M. R. (2010). Registration of pedobarographic image data in the frequency domain. Computer Methods in Biomechanics and Biomedical Engineering, 3(6):731–740. [Oliveira et al., 2011] Oliveira, F. P. M., Sousa, A., Santos, R., and Tavares, J. M. R. S. (2011). Spatio-temporal alignment of pedobarographic image sequences. Medical and Biological Engineering and Computing, 49(7):843–850. [Osuna-Enciso et al., 2013] Osuna-Enciso, V., Cuevas, E., and Sossa, H. (2013). A comparison of nature inspired algorithms for multi-threshold imagesegmentation. Expert Systems with Applications, 40(4):1213 – 1219. [Papa et al., 2012] Papa, J. a. P., Falcão, A. X., de Albuquerque, V. H. C., and Tavares, J. a. M. R. S. (2012). Efficient supervised optimum-path forest classification for large datasets. Pattern Recognition, 45(1):512–520. [Peter et al., 2008] Peter, Z., Bousson, V., Bergot, C., and Peyrin, F. (2008). A constrained region growing approach based on watershed for the segmentation of low contrast structures in bone micro-ct images. Pattern Recognition, 41(7):2358–2368. [Petersen et al., 2002] Petersen, E. M., Deridder, D., and Handels, H. (2002). Image processing with neural networks - a review. Pattern Recognition, 35(10):2279–2301. [Pinho and Tavares, 2009] Pinho, R. R. and Tavares, J. M. R. (2009). Tracking features in image sequences with kalman filtering, global optimization, mahalanobis distance and a management model. Computer Modeling in Engineering & Sciences, 46(1):55–75. [Prasad et al., 2011a] Prasad, G., Joshi, A. A., Feng, A., Barysheva, M., Mcmahon, K. L., De Zubicaray, G. I., Martin, N. G., Wright, M. J., Toga, A. W., Terzopoulos, D., and Thompson, P. M. (2011a). Deformable Organisms and Error Learning for Brain Segmentation. In Pennec, X., Joshi, S., and Nielsen, M., editors, Proceedings of the Third International Workshop on 27 Mathematical Foundations of Computational Anatomy - Geometrical and Statistical Methods for Modelling Biological Shape Variability, pages 135–147. [Prasad et al., 2011b] Prasad, G., Joshi, A. A., Thompson, P. M., Toga, A. W., Shattuck, D. W., and Terzopoulos, D. (2011b). Skull-stripping with deformable organisms. In ISBI’11, pages 1662–1665. [Ramponi et al., 1996] Ramponi, G., Strobel, N. K., Mitra, S. K., and Yu, T.-H. (1996). Nonlinear unsharp masking methods for image contrast enhancement. Journal of Electronic Imaging, 5(3):353–367. [Shkvarko et al., 2011] Shkvarko, Y., Atoche, A. C., and Torres-Roman, D. (2011). Near real time enhancement of geospatial imagery via systolic implementation of neural network-adapted convex regularization techniques. Pattern Recognition Letters, 32(16):2197–2205. [Srinivasan and Sundaram, 2013] Srinivasan, A. and Sundaram, S. (2013). Applications of deformable models for in-dopth analysis and feature extraction from medical images–a review. Pattern Recognition and Image Analysis, 23(2):296–318. [Terzopoulos, 1999] Terzopoulos, D. (1999). Artificial life for computer graphics. Communications of the ACM, 42(8):32–42. [Trifas et al., 2006] Trifas, M. A., Tyler, J. M., and Pianykh, O. S. (2006). Applying multiresolution methods to medical image enhancement. In Proceedings of the 44th annual Southeast regional conference, ACM-SE 44, pages 254–259, New York, NY, USA. ACM. [Wang and Tan, 2011] Wang, J. and Tan, Y. (2011). Morphological image enhancement procedure design by using genetic programming. In Proceedings of the 13th annual conference on Genetic and evolutionary computation, GECCO ’11, pages 1435–1442, New York, NY, USA. ACM. [Wang et al., 2002] Wang, Z., Bovik, A. C., and Lu, L. (2002). Why is image quality assessment so difficult? In Proceedings of IEEE International Conference on Acoustic, Speech, and Signal Processing, volume 4. 28 [Wang et al., 2004] Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. (2004). Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600–612. [Xue et al., 2003] Xue, J.-H., Pizurica, A., Philips, W., Kerre, E., Van De Walle, R., and Lemahieu, I. (2003). An integrated method of adaptive enhancement for unsupervised segmentation of mri brain images. Pattern Recognition Letters, 24(15):2549–2560. [Yang et al., 2009] Yang, J., Wang, Y., Xu, W., and Dai, Q. (2009). Image and video denoising using adaptative dual-tree discrete wavelet packets. IEEE Transactions on Circuits and Systems for Video Technology, 19(5):642–655. [Yi Kim and Jung, 2005] Yi Kim, E. and Jung, K. (2005). Genetic algorithms for video segmentation. Pattern Recognition, 38(1):59–73. [Yu et al., 2010] Yu, J., Tan, J., and Wang, Y. (2010). Ultrasound speckle reduction by a susan-controlled anisotropic diffusion method. Pattern Recognition, 43(9):3083–3092. [Zhang et al., 2010a] Zhang, L., Dong, W., Zhang, D., and Shi, G. (2010a). Two-stage image denoising by principal component analysis with local pixel grouping. Pattern Recognition, 43(4):1531–1549. [Zhang et al., 2010b] Zhang, L., Dong, W., Zhang, D., and Shi, G. (2010b). Two-stage image denoising by principal component analysis with local pixel grouping. Journal of Pattern Recognition, 43(4):1531–1549. 29 FIGURE CAPTIONS Figure 1: Pyramid usually adopted to create an artificial life model (adapted from [Terzopoulos, 1999]). Figure 2: Flow diagram of the proposed image enhancement method. Figure 3: Graph for intensity reduction function for 255=k . Figure 4: Graph of food’s height loss due to the degradation caused by the organism and food’s high gain due to the growing of the food over time. Figure 5: Results of the enhancement methods on synthetic images I: (a) original image, (b) affected image with the addition of noise and blurring; images returned after the applying the methods (c) square root enhancement, (d) quadratic enhancement (e) logarithmic enhancement, (f) normalization, (g) histogram equalization and (h) using our method based on a new artificial life model. (The images (a-), (b-), (c-), (d-), (e-), (f-) (g-) and (h-) represent the difference between the respective processed and original images.) Figure 6: Results of the enhancement methods on syntetic images II: (a) original image, (b) affected image with the addition of noise and blurring; images returned after the applying the methods (c) square root enhancement, (d) quadratic enhancement (e) logarithmic enhancement, (f) normalization, (g) histogram equalization and (h) using our method based on a new artificial life model. (The images (a-), (b-), (c-), (d-), (e-), (f-) (g-) and (h-) represent the difference between the respective processed and original images.) 30 Figure 7: Results of the enhancement methods on synthetic images III: (a) original image, (b) affected image with the addition of noise and blurring; images returned after the applying the methods (c) square root enhancement, (d) quadratic enhancement (e) logarithmic enhancement, (f) normalization, (g) histogram equalization and (h) using our method based on a new artificial life model. (The images (a-), (b-), (c-), (d-), (e-), (f-) (g-) and (h-) represent the difference between the respective processed and original images.) Figure 8: Results of the enhancement methods on real images: (a) original image, (b) affected image with the addition of noise and blurring; images returned after the applying the methods (c) square root enhancement, (d) quadratic enhancement (e) logarithmic enhancement, (f) normalization, (g) histogram equalization and (h) using our method based on a new artificial life model. Figure 9: PSNR graph plotted from the data in Table 1. Figure 10: SSIM graph plotted from the data in Table 2. Figure 11: Profile of the 128th line (randomly chosen) of the original images, affected images, and image returned from our method applied on the images of Figures 5, 6 and 7. Figure 12: Representation of the points classified as belonging to the class DV marked in black: (a) original image, (b) damaged image with the addition of noise and blurring operations; images after application of the (c) square root enhancement method, (d) quadratic enhancement method, (e) logarithmic enhancement method, (f) normalization, (g) histogram equalization, and (h) using our method based on a new artificial life model. 31 Figure 13: Results of the enhancement methods on real images: the original image (a), original image segmented (b), segmentation of images returned by the square root enhancement method (c), quadratic enhancement method, (d) logarithmic enhancement method (e), normalization (f), histogram equalization (g), and the proposed enhancement method (h). 32 TABLE CAPTIONS Table 1: PSNR indexes of the proposed method and other methods for image enhancement presented in the literature. (The lines with values in bold are related to the images of Figs. 5-7.) Table 2: SSIM indexes of the proposed method and other methods for image enhancement presented in the literature. (The lines with values in bold are related to the images of Figs. 5-7.) Table 3: DV and BV values for the images of Figure 8. (The values indicated in bold are the ones that should be approximated.) Table 4: Computational times (in ms) to enhance the images "Lena" and "Cameramen". 33 TABLES Table 1 Image Affected Image Histogram Equalization Normalization Quadratic Enhancement Square Root Enhancement Logarithmic Enhancement Proposed Method 1 5.780 1.327 4.312 1.553 9.787 10.402 12.457 2 5.009 1.406 4.462 2.218 7.794 10.936 12.930 3 3.602 1.716 3.083 1.644 7.597 9.723 12.020 4 9.090 0.455 6.640 2.635 17.130 14.151 18.394 5 5.791 1.348 4.353 1.571 10.392 10.671 15.153 6 4.952 1.326 4.240 2.236 7.990 11.558 15.337 7 3.698 1.666 2.693 1.667 8.220 10.398 16.091 8 8.892 0.443 6.220 2.650 16.543 14.244 18.307 9 6.156 0.946 4.210 2.261 9.787 10.183 12.191 10 4.731 1.560 2.538 2.103 6.951 6.891 8.215 11 4.813 1.479 4.109 2.099 6.241 7.025 8.388 12 4.106 2.260 3.690 3.781 4.856 4.878 5.628 13 6.351 1.024 4.176 2.191 10.978 10.202 13.726 14 5.090 1.522 2.641 2.111 7.910 6.891 9.473 15 4.959 1.427 3.595 2.109 7.142 7.148 10.004 16 5.018 2.202 4.451 4.222 5.235 4.885 6.397 Table 2 Image Affected Image Histogram Equalization Normalization Quadratic Enhancement Square Root Enhancement Logarithmic Enhancement Proposed Method 1 0.147 0.078 0.103 0.061 0.597 0.578 0.728 2 0.265 0.059 0.234 0.113 0.559 0.597 0.747 3 0.134 0.091 0.127 0.120 0.435 0.533 0.722 4 0.448 0.011 0.338 0.113 0.884 0.828 0.917 5 0.169 0.092 0.127 0.078 0.631 0.601 0.844 6 0.272 0.060 0.240 0.118 0.610 0.637 0.843 7 0.162 0.100 0.136 0.129 0.536 0.598 0.874 8 0.440 0.012 0.324 0.127 0.871 0.829 0.912 9 0.308 0.039 0.199 0.033 0.654 0.666 0.765 10 0.133 0.025 0.073 0.060 0.376 0.412 0.536 11 0.195 0.007 0.126 0.012 0.418 0.482 0.560 12 0.100 0.022 0.116 0.135 0.168 0.208 0.268 13 0.337 0.070 0.197 0.044 0.700 0.667 0.814 14 0.185 0.025 0.094 0.071 0.439 0.412 0.603 15 0.224 0.024 0.117 0.026 0.482 0.498 0.650 16 0.179 0.028 0.190 0.212 0.211 0.210 0.337 34 Table 3 Image Original Image Affected Image Histogram Equalization Normalization Quadratic Enhancement Square Root Enhancement Logarithmic Enhancement Proposed Method DV BV DV BV DV BV DV BV DV BV DV BV DV BV DV BV Cup 3.18 0.05 1.41 0.24 5.37 0.30 1.82 0.25 2.21 0.29 1.06 0.15 0.67 0.04 4.48 0.03 Lena 2.81 0.07 1.34 0.22 5.21 0.26 2.39 0.18 2.62 0.18 0.82 0.18 0.54 0.04 2.65 0.06 Came- ramen 4.27 0.04 1.63 0.19 5.34 0.24 2.28 0.20 2.81 0.21 1.02 0.13 0.51 0.06 5.47 0.03 Parrot 2.51 0.15 1.40 0.15 6.29 0.22 1.97 0.13 2.66 0.14 0.92 0.13 0.71 0.03 3.06 0.06 Table 4 Image Histogram Equalization Normalization Quadratic Enhancement Square Root Enhancement Logarithmic Enhancement Proposed Method Lena 0.11 0.62 5.40 1.35 4.68 1460.05 Camer amen 0.11 0.52 5.41 1.35 4.68 2926.87 35 FIGURES Figure 1 Figure 2 36 Figure 3 Figure 4 37 Figure 5 38 Figure 6 39 Figure 7 40 Figure 8 Figure 9 Figure 10 41 Figure 11 42 Figure 12 43 Figure 13 44 New artificial life model for image enhancement Alex F. de Araujo¹, Christos E. Constantinou² and João Manuel R. S. Tavares¹ 
work_3tgwtllla5dxldjqob45z2m7se	----	w d r - ci d o 6 / c f g - , / STRUCTURAL DYNAMICS TEST SIMULATION AND OPTIMIZATION FOR AEROSPACE COMPONENTS S. E. Klenke and T. J. Baca Experimental Structural Dynamics Department Sandia National Laboratories Albuquerque, New Mexico 87185-0557 ABSTRACT: This paper initially describes an innovative approach to product realization called Knowledge Based Testing (KBT). This research program integrates test simulation and optimization software, rapid fabrication techniques and computational model validation to support a new experimentally-based design concept. This design concept implements well defined tests earlier in the design cycle enabling the realization of simulation and optimization software environment provides engineers with an essential tool needed to support this KBT approach. This software environment, called the Virtual Environment for Test Optimization (VETO), integrates analysis and test based models to support optimal structural dynamic test design. A goal in developing this software tool is to provide test and analysis engineers with a capability of mathematically simulating the complete structural dynamics test environment within a computer. A developed computational model of an aerospace component can be combined with analytical and/or experimentally derived models of typical structural dynamic test instrumentation within the VETO to determine an optimal test design. The VETO provides the user with a unique analysis and visualization environment to evaluate new and existing test methods in addition to simuIating specific experiments designed to maximize test based information needed to validate computational models. The results of both a modal and a vibration test design are presented for a reentry vehicle and a space truss structure. INTRODUCTION: This paper presents the Knowledge Based Testing (KBT) concept which incorporates aspects of design, analysis, and test with rapid prototyping to optimize test based product information from an experiment. The purpose of developing a KBT program is to utilize testing earlier in the design cycle. There are a number % ?- . < highly reliable aerospace components. A test i of research activities that will be discussed in this paper that help modify the conventional testing paradigm that normally tests a product at the end of development. One of these research areas being developed is the Virtual Environment for Test Optimization, VETO'. This innovative tesdanalysis tool reduces test instrumentation iterations, producing better tests through optimal test design. Communication between test and analysis engineers is enhanced early in the design cycle. Traditionally, the role of testing in the product realization process is limited to the end of the design cycle, after hardware manufacturing decisions have been made. As a result, data analysis and test requirements for a component are only considered when the hardware is scheduled for testing. Thus, the full benefit of the analysis in guiding the test is not realized. A goal in developing this software tool is to provide test and analysis organizations with a capability of mathematically simulating the complete test environment within a computer. Derived models of test equipment, instrumentation and hardware, called virtual instruments, can be combined within VETO to provide the user with a unique analysis and visualization capability to evaluate new and existing test methods. By providing engineers with a tool that allows them to optimize an experimental design within a computer environment, pre-test analysis can be performed using analytical models to rapidly evaluate components before manufacturing has occurred. The benefits of using this type of experimental design tool can be very extensive. The user can evaluate the use of different types of test instrumentation and equipment as well as investigating new testing techniques for system identification used to experimentally validate analytical models. A second research activity has focused on using the recently developed stereolithography to reduce the time between concept and product realization by evaluating plastic models to predict the structural behavior of actual metal parts. This TER Page 1 American Institute of Aeronautics and Astronautics DISTRiBUTION OF TW DOCUMENT If3 Portions of this document may be iIlegible in electronic image products. Images are produced from the best available original dolument. DISCLAIMER This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or use- fulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any spe- cific commercial product, process, or service by trade name, trademark, manufac- turer, or otherwise does not necessarily constitute or imply its endorsement, recom- mendation. or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. stereolithography process has provided the means of economically generating very exact plastic prototype parts from three dimensional solid models. This process rapidly produces prototype plastic parts with astounding accuracy including bolt holes, countersinks, fillets, cutouts, etc. These stereolithography plastic prototypes include all the detail and are nearly perfect replicas of the actual parts. They afford detail that cannot be included in finite element models without extremely expensive high order meshes of very small features which may be important to the structural performance of the design. These prototypes are currently used primarily to verify geometries, for interference checks, and for product visualization. The ability to perform mechanical tests on these plastic prototypes and infer actual metal part structural performance could significantly reduce design cycle times. Mechanical tests of interest include static loading, modal testing and vibration testing. These rapid prototype test results could also be used to validate analytical models early in the design process to provide predictive models for effective design iterations and tradeoff studies. These very exact plastic prototypes offer a new realm of opportunity for the use of plastic models to predict the structural performance of actual metal parts. A third area of ongoing research related to the KBT program is in computational model validation4. Improved finite element modeling techniques and the ever decreasing cost of computing have been a major contributor to the increased capability of computationally-based engineering simulations. The current trend toward increased reliance on simulations for design and performance evaluation is pervasive throughout the industry. However, within the KBT program it is also recognized that there is a need for increasingly sophisticated testing and physical simulation that is essential to the development and validation of engineering models. Thus, by combining the use of the VETO, to optimally design an experiment, with the use of rapid prototyping techniques for generating component parts, the processes of model updating and validation becomes much more efficient. These tools are critical to providing test based information needed to produce confidence in the predictive capabilities of computational models of aerospace components. KNOWLEDGE BASED TESTING (KBT) OVERVIEW: As was mentioned in the previous section, the goal of the KBT program is to position testing earlier in the design cycle. This new vision or role for testing is partly motivated by increased modeling capabilities as well as the recent increase in computational power. In this vision, the definition of testing will not simply mean to provide the "admiral's tests" but will be used as an underlying tool that is essential to model development and validation. A view of the KBT concept is shown in Figure 1 . This bubble chart depicts Rapid Product Realization Figure 1. Knowledge Based Testing Chart the important interactions between design (starting with requirements and specifications), analysis and test used to support rapid product realization. The initial step in the KBT program is the generation of a computer aided design model which represents the geometry of the component or component housing to be tested. This model is generally driven by certain requirements and specifications. This geometric model is then used to assist component visualization, to generate a computational model used for dynamic analysis, and to produce a rapid prototype component through a stereolithography or "fastcast" process* '. The developed computational model of the component is then combined with analytical andor experimentally derived models of the test instrumentation within the VETO software to determine an optimal test design. The VETO is then used to simulate the structural dynamics experiment in order to maximize the test based information gathered from the experiment. The next step in the KBT process is the performance of the experimental test or characterization test using the rapid prototype component given the VETO test design. The results of this experiment are then used to update and validate the computational model. This systematic approach to rapid prototype evaluation integrated with optimal test design software is an +Fastcast uses a stereolithography or laser sintered part as a mold for an automatic metal casting system. Page 2 American Institute of Aeronautics and Astronautics important part of the KBT concept that helps bridge the gap between the old testing paradigm and this new testing vision. Design and analysis methods are integrated to support test simulations. These computationally-based simulations provide predictions of component or system response to testing environments and the results of these simulations are then used to guide an actual characterization test. By completing the test design within the virtual environment, the user can effectively evaluate what test information is needed from the experiment to update and validate the computational model. After this validation process is complete and some level of confidence is established in the computational model, further optimization and simulation studies can be performed using the model before any manufacturing or building of the component is done. It should be noted that the KBT concept also includes aspects of testing necessary for product verification. The KBT definition of product verification includes certification testing, design margin testing6 and failure mode/ mechanism identification testing. Lon g term monitoring of component performance is the final type of KJ3T which involves health monitoring7. ROLE OF STRUCTURAL DYNAMIC TEST SIMULATION AND OPTIMIZATION IN KBT: The VETO software environment currently integrates analysis and test based models to support optimal structural dynamic test design. The structural dynamics testing environment was selected as the initial VETO environment for investigation into areas of design/ analysis/test interfaces, visualization, versatility and repeatability. This initial VETO effort has focused on assisting engineers to maximize the value and information gathered from these tests. A major objective of the VETO software development effort is flexibility. Because the virtual environment is a prototype software system, a primary concern for its design is that the code be easy to develop. To minimize this effort, existing software tools are used wherever possible, provided that the necessary functionality and flexibility are available. Another significant design objective is to provide a final software system that can be used by a variety of individuals who have not been involved in its development. As is described below, VETO integrates several commercial tools to meet these objectives. Currently, the VETO software tool has been developed to support a modal virtual test environment while development of a vibration virtual environment is under development. Many of the tools needed to - support simulations of these two structural dynamic test capabilities are similar, however, because of the various differences in these technologies, two distinct simulation approaches have been developed. The first approach, which was formulated to support modal test optimization, provided the user with an environment where numerical integration is performed on a system of state space models which described the modal test configuration. This method served very well in the support of simulations which did not include closed loop control. However, in our second approach an effort is being made to address simulations that do include control aspects, namely vibration tests. The vibration virtual environment will include hardware-in- the-loop (both control and data acquisition elements) in addition to instrument and equipment models to support the simulation. Further details on these approaches are described below. For both of the structural dynamic testing environments, the database, integration, utility and user interface functions are performed in the Vetomain module, Figure 2. This main interface provides the Figure 2. Vetomain Graphical Interface communication links between the two commercial software packages: AVS which is used for visualization and MATLAB which is used to perform modeling and time integration. The "File" option of the vetomain menu bar allows users to load finite element (FE) models and previously defined virtual test files into the VETO software. The setup for the test simulation is performed using Vetomain to construct parametric models of the instruments, and to formulate interconnections between the models. The user is able to interact with the virtual instrument to provide and view information on the devices needed in the simulation. This user interaction includes the selection of the number of desired response locations, the location of input excitation and the type of instruments needed to support the simulation. This interface module also provides numerous tools to assist the engineer in setting up and understanding how the various virtual instruments interact together to support a specific structural dynamic test simulation. These tools will help guide the engineer in the design of tests that will accurately identify all the desired modes of the structure, select the appropriate test instrumentation as well as identify the proper excitation levels for the test. Page 3 American Institute of Aeronautics and Astronautics MODAL TEST SIMULATIONS AND OPTIMIZATION: The SIMULINK Dynamic System Simulation Software provided by MATLAB is used as the environment to assemble and ultimately integrate . mathematical models of the modal test system. This same software controls the simulation processing. Dynamic response equations are integrated by SIMULINK to provide system output time histories. Within the VETO software, inputs such as type of device and interconnection of instrumentation models are combined to facilitate the rapid connection of various models (including models of test instrumentation, equipment and hardware) which comprise a given modal testing process. In order to achieve rapid set up of this virtual environment, models representing the instrumentation and test equipment need to be developed. These models consist of mathematical descriptions of the dynamic response of the instruments derived either theoretically or experimentally. Most of the instruments modeled to date have been modeled in the discrete state space domain. A number of system identification tools were used in MATLAB to generate the mathematical models. Development was based on an experimental frequency response function of the instrument or equipment. The models of the different types of instruments and equipment (transducers, ampwiers, filters, etc.) needed to represent a complete modal testing environment are located in a SIMULINK Virtual Test Equipment Library (VTELib). When preparing for a test simulation, the selection of the desired test instrumentation from the Vetomain is performed with the assistance of a MATLAB code which searches the VTELib for available instrument models. Optimal experimental design and simulation of the complete test environment is further facilitated by the VETO'S ability to include models of external inputs and electronic instrumentation noise. In addition, complex instrumentation models, such as the data acquisition system (Front End), are constructed by combining multiple submodels to simulate the dynamic response behavior of the hardware. As these models are added to the new simulation system, interconnecting lines are placed between the block models within SIMULINK. These lines represent the flow of signals in the actual modal test system and are specified using the "wire" instrument in Vetomain. Using these interconnecting lines, the input signals from the actuator devices (e.g. impact hammers) are fed to both the device under test (D JT) model or simulation of system excitation and to the "Front End" device for simulation of data acquisition. This DUT model is directly generated using the FE modal data given the user desired input and sensor locations for the simulation. The "Front End" device also receives the signals from the sensors that have been attached to the DUT to simulate structural system response to the actuator input. Both the simulated actuator and sensor signals are linked through amplifier and filtering blocks to represent preconditioning of the signals. The process of simulation begins when the user selects "Run" from the "Simulate" option on Vetomain. The data files which define the dynamics of the desired instrumentation are loaded into the test simulation system and the "Simulation Monitor" is created and displayed. This monitor allows the user to observe the estimated system response based on the numerical integration. The Simulation Monitor represents the data acquisition environment commonly used to gather data in a physical test and is a graphical interface through which the user interacts with the modal test simulation system. After collecting the desired amount of simulated data, the user can activate a window providing an interface to analysis routines for computing measures such .as frequency response functions, power spectral densities and coherence based on the simulated data. MODAL TEST APPLICATION: An aerospace component was selected as a test case for application in the VETO environment. The VETO software simulation tool was used to design an optimal experiment for this mock reentry vehicle, Figure 3. The Figure 3. Mock Re-entry Vehicle Model goal of performing this test design optimization was to observe the vibration modes of interest and to study the interaction of the support flanges with the reentry vehicle housing. The initial steps in the test design , Page 4 American Institute of Aeronautics and Astronautics were to select an appropriate set of instrumentation (including sensors and actuators) needed to perform a modal experiment within the VETO environment and to simulate responses on the aerospace component. A symmetric finite element model of the structure was loaded into the VETO environment for use in the modal test simulation. The test design was performed over a frequency band, up to 250 Hz, which included fifteen vibration modes of the reentry vehicle. The outcome of the VETO test design "Setup" was to excite the structure using an impact hammer and to measure acceleration responses on the reentry vehicle at 40 different locations in order to characterize the dynamic behavior of the component. Approximately half of the response locations were automatically selected using an analysis code to optimize the sensor locations. Some care was taken in utilizing this code to ensure that redundant or closely spaced response locations on the structure were not used in the simulation. Other instrumentation such as the signal conditioning amplifiers and the data acquisition system were also set up with the use of Vetomain in preparation for the test simulation. Data acquisition parameters for sampling, averaging and acquiring the desired analysis measurements were also selected for use in the post-simulation analysis. A number of "Pre-simulation" tools were used to determine the completeness of the test design. First, the effects of mass loading the component were calculated given the test design sensor set. Small accelerometers, Endevco 2250s, were selected in the test design in order to minimize the mass loading effects that might occur during the experimentation. This analysis predicted that very small changes in the frequencies of vibration (approximately 0.1 %) would be experienced if an experimental test was conducted based on the selection of small accelerometers in the test design. Second, a normal mode indicator function and a driving point frequency response function were viewed before conducting the test simulation in order to assess whether the selected sensor and actuator (selected impact location) set would accurately identify all the desired modes of interest on the aerospace component, Figures 4 and 5 . By using the normal mode indicator function, it was determined that a single input location at the nose of the reentry vehicle would not excite a1 of the modes of the structure. Therefore, additional excitation locations would need to be included in the test setup so that all the modes of vibration of the reentry vehicle could be observed. Finally, the Modal Assurance Criterion (MAC) was calculated for the test design to determine if the modes of vibration of the structure could easily be distinguished from one :::: 0.1 50 100 150 2w 250 30( Frequency Figure 4. Normal Mode Indicator Function another given the selected sensor set. Small values on the off-diagonal terms of this MAC matrix, Figure 6, Actuator uui-:i. l0"t 1 0' 1 02 10' Frequency Figure 5 . Drivin Point Frequency Response % unction Mod3 Assurance Criteria Mode# '' Mode # Figure 6 . Modal Assurance Criterion indicate the relative independence of the modes of vibration, thus facilitating correlation with analysis. With the complete test design within the VETO environment, a SIMULINK block model of the test environment is automatically generated to support the simulation of the modal test. Figure 7 shows a partial Page 5 American Institute of Aeronautics and Astronautics Figure 7. Modal Test Simulation Block Model block model of the SIMULINK environment. The next step in the modal test simulation is the numerical integration of the mathematical models within SIMULINK to estimate the system responses. Using the Simulation Monitor, these responses are observed for each set or frame of data to be collected, Figure 8. Once the data are gathered to support the desired measurement set, the test simulation within SIMULINK is concluded. A window which provides Figure 8. Simulation Monitor an interface to the post-simulation analysis routines is then used to download the data for measurement analysis. A number of analysis routines for computing desired measures such as frequency response functions, power spectral densities and coherences are available. The simulated data, which are based on the FE dynamic analysis, were used to generate frequency response functions, Figure 9. VIBRATION TEST SIMULATION AND OPTIMIZATION: As was mentioned earlier, the vibration virtual environment is currently under development. Some of the issues that must be addressed in this particular 50 100 150 200 2YJ 300 10-2 Frequsncy (Hz) Figure 9. Measured FRF using Simulated Data simulation environment that are not included in the modal virtual environment are: closed loop vibration control, detailed shakedamplifier models and shaker/ fixture/component interaction. Because of these issues as well as others, it was determined that a new simulation approach would be developed using some of the existing hardware that directly support vibration tests. These simulations, called hardware-in-the-loop simulations, would provide the user with the ability to combine actual vibration test hardware (such as vibration control and data acquisition systems) with instrumentation models to support vibration test design. Models of the external load (the shaker/ amplifier, interface block, fixturing, device under test, accelerometers and signal conditioning elements) will be combined into a single state space model using MATLAJ3 routines and then downloaded onto a real- time control processor. With the simulation model of the external load residing on the processor, the physical hardware is then connected to the processor for performing the vibration test simulation. The particular processor used in our simulation environment was developed in-house and has the capability of handling 16 inputs and 16 outputs with a total of 128 states. The sampling rate of the processor is based on the size or number of states of the model. Figure 10 shows a simple representation of this hardware-in-the-loop simulation environment. The virtual vibration environment will allow the user to evaluate the overall testability of a component or system. The test engineer will be able to observe the effects that different control parameters might have on the DUT without putting the physical hardware through an actual vibration test. An advantage of this environment is that it can help limit unnecessary vibration inputs to flight hardware. New and existing fixture designs can also be studied through the development of analytical models that can be integrated into this simulation environment. Testing Page 6 American Institute of Aeronautics and Astronautics r . I(eal-l'ime Codrol Processor Sun Workstation d MATLAB &FtT(P VibrationControl Conplter Figure 10. Vibration VETO Concept methodologies such as the number of control transducers, the location of control transducers and the control strategy or method can also be investigated within the vibration environment. These developments will help assist analysis, design and test engineers in maximizing the value of each vibration test. VIBRATION TEST APPLICATION: A large gamma truss structure was selected as the test case for the vibration test simulation environment*. Figure 11 shows a uicture of the actual truss hardware Figure 1 1. Gamma Truss that was used to demonstrate the hardware-in-the-loop simulation concept. It should be noted that due to the ongoing development of the vibration virtual environment, not all of the instrumentation models (namely the shakerhnplifier model) were included in this demonstration of the vibration simulation capability. Our goal at this point is simply to show the hardware-in-the-loop simulation process. The truss structure was designed with integrated sensors and actuators to provide a testbed for studying structural controls applications. Through numerous control studies, experimental data had been gathered in the form of input/output models for the truss. For use in the vibration test simulation, an eight inpuueight output experimental model was selected as the model that would represent the truss in the simulation. This state space model of the truss was combined with models of the sensors (accelerometers) and signal conditioning elements to form a single state space model of the vibration external load. This combined state space model was then downloaded onto the real- time control processor in order to support running the test simulation. An actual vibration test control and data acquisition system was then connected to the real- time processor for the simulation study. A single vibration drive signal was generated from an arbitrary shaped random spectrum (up to 200 Hz) and then used in the simulation to excite the external load model on the processor. The first output or response channel from the processor was feed back into the vibration control computer as the control channel. The drive signal was updated in order to match this shaped random spectrum for the chosen control channel. This drive signal was notching at 4 4 Hz due to a lightly damped mode at that frequency. Figure 12 shows the drive spectrum for this vibration simulation. By utilizing this vibration test simulation environment, the engineer is able to select desired control parameters such as control Degrees of Freedom, frequency resolution and time constants to successfully control a specified vibration test, Figure 13. This simulation example shows the strong advantage of having an environment to setup and customize vibration tests. im Figure 12. Vibration Drive Spectrum Page 7 American Institute of Aeronautics and Astronautics Using this tool, test engineers can easily change parameters within the simulation to optimize vibration tests. CONCLUSIONS: An important goal in developing the KBT program is to provide an environment in which designers and analysts are given earlier access to test based information in order to make intelligent design decisions. There are a number of significant research activities ongoing that directly support this new testing vision. A test simulation and optimization software tool called VETO is one of these activities that helps position testing earlier in the design cycle. Using this simulation software and rapid prototype parts, characterization tests can be performed in order to generate test data needed to update and validate computational models. Within the VETO software environment, engineers are able to investigate the testing of aerospace components, using computational models, prior to the existence of any hardware. A goal in developing this software tool is to provide test and analysis organizations with a capability of mathematically simulating the complete test environment within a computer. Applications of this test simulation environment have been shown for both modal and vibration capabilities essential to the development of aerospace components. ACKNOWLEDGEMENT: This work was supported by the United States Department of Energy under Contract DE-AC04- 94AL85000. REFERENCES: [ l ] Klenke, S., Reese, G., Schoof, L. and Shierling, C., "Modal Test Optimization Using VETO (Virtual Environment for Test Optimization)", Sandia Report SAND95259 1, January 1996. [2] Gregory, D. and Hansche, B., "Rapid Prototype and Test", Sandia Report, 1996. [3] Jacobs, P., "Rapid Prototyping & Manufacturing Fundamentals of Stereolithography", McGraw, 1992. [4] Dalton, E., Chambers, B., Bishnoi, K., Bateman, V., and Baca, T., "Test Validation of MANTA: A Code for the Practical Simulation of Shock in Complex Structures", Proceedings of the 65th Shock and Vibration Symposium, Vol. 1, San Diego, CA, October [5] Zanner, F., and Maguire, M., "FASTCAST, A Program to Remove Uncertainty from Investment Casting", Modem Casting, 8/93. 1994, pp. 66-75. [6] Baca, T., Bell, R., and Robbins, S., "Conservatism Implications of Shock Test Tailoring for Multiple Design Environments", Proceedings of the 58th Shock and Vibration Symposium, Vol. 1, October 1987, [7] James 111, G., "Development of Structural Health Monitoring Techniques Using Dynamic Testing", Sandia Report SAND96-0810, April 1996. [8] Lauffer, J., Allen, J., and Peterson, L., "Structural Dynamics Considerations for Structural Control", Proceedings of the 7th International Modal Analysis Conference, Vol 11, February 1989, pp. 1079-1086. pp.29-47. Page 8 American Institute of Aeronautics and Astronautics 
work_3tqvfeholvburhnwt43i3vso7i	----	Microsoft Word - IfacFin.doc TOWARDS SPECIALIZED EXPERT SYSTEM BUILDING TOOLS: A TOOL FOR BUILDING IRRIGATION EXPERT SYSTEMS Samhaa R. El-Beltagy, Gamal Al-Shorbagi, Hesham Hassan and Ahmed Rafea Central Lab for Agricultural Expert Systems Agricultural Research Center Ministry of Agriculture and Land Reclamation. Cairo, Egypt El-Nour St. P.O Box 438 Dokki,Giza, Egypt E-mail: {samhaa, gamal_sh, hesham, rafea}@.claes.sci.eg) Abstract: Expert system development is very often an expensive process which requires significant time and effort. This paper investigates the idea of building a specialized tool for a specific task in order to realize very rapid development times with very little effort on the part of the developer thus considerably reducing the overall development cost. The particular task that this paper addresses is that of irrigation scheduling. The work presented describes various aspects of a tool that was built for rapidly developing irrigation expert systems for vegetable crops in Egypt, and shows how greatly this tool simplifies the process of building such systems. Copyright © 2004 IFAC Keywords: Agriculture, Expert systems, Knowledge tools, Knowledge engineering, Modelling. 1. INTRODUCTION Despite the availability of a wide spectrum of expert system tools and shells, developing an expert system is still often a time consuming and expensive process. Further more, the development of a ‘good’ expert system is still also highly dependant on the skills of its knowledge engineers. Reducing expert system development costs in terms of development time and required skills continues to be the goal of many research bodies. When addressing this issue for expert system development in general, improvements in cost and effort are likely to come solely from enhancements in development shells. However, much more significant improvements can be achieved in domains where expert system development is a recurring activity for a set of recurring tasks. Building agricultural experts systems for a wide variety of crops and a number of tasks (diagnosis, treatment, irrigation, fertilization, etc) is an example of such domains. Instead of replicating the same effort every time an expert system is being developed, commonalities can be captured and re-used. Such re-use can be achieved by standardizing ontologies and by building specialized tools for specific tasks within a specific domain. Such tools would not only contain a problem solving model for some given task, but would also include all knowledge that can be re-used for this particular task. The goal of this paper is to present an example of such a tool built specifically for irrigation expert systems. Section 2 of this paper briefly reviews some related work, section 3 explains the goals of irrigation expert systems as well as the goal of the tool to be presented, section 4 describes the tool that has been built, and section 5 concludes this paper and presents some ideas for feature work. 2. RELATED WORK Knowledge system technology has been applied to a variety of agricultural problems since the early 1980s. Generally we can classify agricultural activities into: activities that are done on the farm prior to cultivation and activities, which are done during cultivation operations. The scope of this research concentrates on an activity type that is done during cultivation. Specifically, this work focuses on the irrigation scheduling activity. Many systems have been built to address the irrigation scheduling problem. These include for example the NEPER Wheat expert system (Kamel, et al., 1995). The NEPER Wheat expert was developed for handling most of the production management activities including irrigation for the wheat crop. The generic task knowledge system development methodology (Chandrasekaran, 1988) has been utilized to develop this system. The generic task methodology is one that has recognized the importance of task re-use for speeding up the process of building knowledge based system. The methodology has identified a number of problem solving building blocks that target specific problems. For example, the hierarchical classification problem solver can be used to build a diagnosis system while a routine design problem solver can be used to build a scheduling system. Other examples of irrigation expert systems include the barley crop management expert system (Boner I., 1992) which was designed to produce water and fertilizer recommendations to maximize yield., and the CALEX/Cotton expert system (Ostergard M.and Goodell P. 1992; Plant R. 1989) which was designed to link diverse production components including irrigation, pest management and agronomy. The CUPTEX (El- Dessouki, et al., 1993) and CITEX (Edrees S., 1997) systems for cucumber and citrus production management respectively cover most of the agricultural practices for cucumber crop under tunnel and citrus cultivation in open fields including irrigation and fertilization. References for other expert system in this domain may be found in (Rafea, 1998). However each of these systems was built from scratch to target a specific crop. The extent to which these system could make use of previously built ones was limited on identifying a suitable knowledge representation scheme, and/or a suitable problem solving model. While these do in fact reduce the effort involved in building an irrigation expert system, they do not capitalize on the fact that one generic model containing common knowledge, equations and a problem solving model, can be applied to all. Through this application the reduction in both time and effort can be much more significant. 3. IRRIGATION EXPERT SYSTEMS: A CLOSER LOOK The main goal of most irrigation expert systems is to produce an irrigation schedule for a particular crop in a particular farm. The output schedule is a plan of water quantities to be applied and the time of application according to the requirements of the plant and the affecting factors like the soil type, climate, source of water, etc. Determining a crop’s water requirements is no trivial task, but is one that involves many equations that in turn involve an extensive number of complex variables relative to the soil, water, climate, crop etc. This work capitalizes on the fact that even though the irrigation requirements for various crops may vary, a number of basic concepts and irrigation determination techniques are shared among all. By recognizing this fact, a tool was built to hide all the complexities of an irrigation’s system equations and knowledge from the developer, while highlighting any missing knowledge that can vary from one crop to another thus guiding the developer as to ‘what’ knowledge needs to be acquired. In a sense, the developer is simply offered an empty template for specifying the inputs and rules for determining or calculating the value of a property of some pre-identified concept. This greatly simplifies the knowledge acquisition task, and makes it possible for any person with some computer knowledge to easily build an irrigation expert system. In the next section we present this tool. 4. THE DEVELOPED IRRIGATION TOOL A typical irrigation expert system is made out of concepts, relations, and equations on top of which a task layer is built. The developed specialized irrigation tool was built on top of an expert system development tool called KSR. KSR was designed and implemented at the Central Laboratory for Agricultural expert systems in Egypt. In this work, we have aimed to identify and capture all knowledge that is related to the irrigation task regardless of the crop, as well as identify knowledge that may vary from one crop to another. The steps for developing the specialized irrigation building tool can be summarized as follows: 1. Identify the main tasks involved in the production of an irrigation schedule 2. Acquire and model knowledge related to each of these tasks (concepts, equations, relations, etc) 3. Identify concepts and relations that may vary from one crop to another 4. Develop a generic irrigation task layer 5. Build an interface that hides non-changing knowledge from the user as well as enables him/her from accessing modifiable knowledge and editing it. The result of following these steps was a tool capable of assisting developers in rapidly building an irrigation scheduling system for any vegetable crop in Egypt. 4.1 Tool features and description The tool enables the production of an irrigation schedule on a weekly basis provided that climate reference data is available. The irrigation model caters for flooding, or drip irrigation systems. The model also includes equations to cover the case of planting within an open field, high tunnels, or low tunnels. Since the method of determining the value of potential evotranspiration (et0), which is essential for determining a crop’s water requirements, differs across planting environments, two well known methods for calculating this value known respectively as Hargreaves (Hargreaves and Samani, 1982) and Pennman (Doorenbos and Pruitt, 1984), have been adopted. The Hargreaves model applies only to high tunnels, while the Pennman model is applicable to the other two planting environments. The tool currently has about ninety five built in concepts, thirty five relations, eleven tables, and fifty two equations as well as the irrigation scheduling problem solving model. 4.2 Using the tool To develop an irrigation expert system, all the developer has to do is to examine the major tasks, identify any missing knowledge and enter that into the tool. Within the developed tool, eight major tasks for generating an irrigation schedule have been identified. These are as follows: 1. Determine the Growth Stage of the Plant 2. Determine factors affecting the Plantation 3. Calculate the value of et0 (potential evotranspiration) 4. Calculate the value of eta (exhaustion of actual water) 5. Calculate the value of pawc (water available in soil) 6. Calculate the irrigation interval 7. Calculate water requirements 8. Calculate required irrigation units For each of these tasks, any relations or rules that may vary from crop to crop were identified. These have been termed ‘modifiable steps’ as they can be modified depending on the crop under consideration. For each modifiable step, the output is determined before hand, but the inputs that influence that output are left for the developer to specify. Each modifiable step is either represented by a set of rules constituting a relation, or a table. The expert system developer need only inspect the ‘modifiable steps’ associated with each of the eight major identified tasks and fill in the missing knowledge for each. These are clearly represented in the developed tool as shown in figure 1. After examining each step, the task of the developer is to acquire the knowledge which will accurately determine the required output of the step. So, the steps for developing an irrigation system can be summarized as follows: • Determine the factors that influence any of the output(s) of a ‘modifiable step’. • Understand how the variation of these factors can influence the output. • Through the developed tool, locate the concepts and properties representing the identified factors and use them to write the rules or to fill in the tables associated with each step, so as to reflect the acquired knowledge To further assist the developer in this task, each of the modifiable steps is described and the requirements from the developer are clearly laid out Figure 1: Irrigation Expert system developer interface for him/her. An example of the description of modifiable steps associated with two of the identified major tasks is shown in figure 2. In total the developer needs only modify five relations and four tables. While the developer can view any of the built in equations, he/she can not edit any of these. For modifying relations and tables, two specialized editors are provided. The relation editor is shown in figure 3. A separate interface is also provided for the entry of reference data (for the farm, soil, water, climate, etc.) by either the expert system developer or the end user. 4.3 Knowledge Representation For the irrigation task, mathematical dependency graphs were chosen to model the irrigation knowledge. The concept of dependency networks (DNs) was first introduced in the context of developing truth maintenance systems (Forbus and Kleer, 1993). A DNs is a directed graph consisting of nodes- representing variables or concepts- and links- representing the dependencies between these nodes. A mathematical dependency graph is a novel knowledge representation scheme suitable for representing mathematical knowledge in a declarative manner such that it can be traversed to calculate a quantity of a numeric object. A dependency graph consists of a set of nodes related to each other by directed arcs. An arc between a node A and a node B means that the value of node A depends on the value of node B or in other words, that there is a depend-on relation between these two nodes. The mathematical dependency between a node and any other nodes means that the value of this Figure 3: A snapshot of the relation editor �������� � � ���� �� ����� ������������� � ���� ���� � �������� ����� � ���� ���� ������� �� ��������������� �������������� ����� � ������� �������� ��������������� ���� ��� ������������ �� ��������� ��� ��� � ��� ������ ��������������������� ����� Steps associated with Task 1: Determine the Growth Stage of the Plant There is only one step associated with this task the name of which matches the name of the task itself “Determine Growth Stage”. The step is of type ‘Relation’ meaning that it is composed of a set of rules. The purpose of this step is to be able to determine the length of each of a crop’s growth stages in days. Example growth stages include: Initiation, Vegetative, Flowering, etc. Factors that can affect these, can include the crop’s variety and/or other factors depending on the crop itself. Steps associated with Task 2: Determine factors affecting the Plantation Step1: Calculate Expected Fruit Yield (efy):This modifiable step is represented by a table. In this table you should place any factor that can affect expected fruit yield, as a column. There is only one possible output (efy) the value of which is specified based on the variation of the affecting factors. Step2: Calculate Average Expected Fruit Yield:This step is also represented by a table, and steps similar to those described for efy should be followed to fill it. In this table the only possible output is that of the average expected fruit yield Step3: Determine Variety Characteristic Bases: This step is of type ‘Relation’ and its output is the value of the ‘variety factor’. The developer should thus aim to write the rules the determine the ‘variety factor’. Step4: Determine Optimal Number of Plants: This step is also of type ‘Relation’ and its output is the value of of the ‘optimal number of plants factor’ Figure 2: A sample of help provided for filling in modifiable steps node is computed through a mathematical function, which includes the other nodes as parameters of this mathematical function. This form of knowledge representation was chosen for the following reasons: 1. To avoid recalculations. 2. To enhance readability and maintainability. 3. To promote reusability. 4. To enable explanation. Figure 4 shows a simplified diagram for part of the dependency graph used in our tool. F in the diagram denotes a function while R denotes a relation. 5. CONCLUSIONS AND FUTURE WORK An irrigation scheduling expert system is a complex system that requires a lot of effort for building. A typical system will have a considerable number of concepts, relations, tables and equations built into it. By capturing re-usable knowledge across various irrigation systems, the required development time can be drastically reduced. This paper presented a tool that can achieve this for irrigation expert systems for vegetable crops in Egypt. Through this tool the process of developing an irrigation expert system is greatly expedited and the effort for developing such a system is greatly reduced. The developer does not need to concern himself/herself with the complexities of the built in model. All he/she has to do is focus on acquiring knowledge to determine the output a very small number of relations and tables. Of the thirty five relations built into the system, only 5 need be modified and of the eleven tables only 4, while none of the fifty two built in equations need be changed. The problem-solving model is also built in and hidden from the developer. This not only cuts down of the development time of an irrigation expert system, but also alleviates the need for a skilled knowledge engineer for developing a powerful system. By providing clear instructions for how to use the tool, almost any person with access to a domain expert can easily build an irrigation expert system. In fact, domain experts themselves can easily use this tool for research purposes. The tool was used for training non- computer specialists on developing irrigation expert systems as part of a one week workshop on expert system development in general. Only two days were designated for the purpose of training a developing a simple irrigation expert system. All workshop attendees were able to develop a working system within this limited time span. Future enhancements to the tool will extend it to cover horticultural crops. Work is also currently being carried out in order to augment the developed tool with explanation facilities which is very likely to transform the tool into a powerful education instrument. REFERENCES Boner I., Parente A. and Thompson K. (1992). Knowledge-Based systems for crop management. In: Fourth International Conference On Computers In Agriculture, Orlando, Florida. American Society of Agricultural Engineers. . Chandrasekaran,B. (1988). Generic Tasks as Building Blocks for Knowledge-Based Systems: The Diagnosis and Routine Design Examples. The Knowledge Engineering Review 3, 183-210. Figure 4: A simplified diagram of a section of the dependency graph employed current_month Pcf et0_pennman Pcf et0 Farm.value=high_tunnel Farm.value=open_field OR Low_tunnel Irrigation.method=daily Irrigation.method=weekly Pcf et0_hargerve F et0_har_c F et0_har_n R control_f F smooth_et0_hargerve Irrigation.method=daily Irrigation.method=weekly currentay -avg_tc -avg_rh -ash -msh -ra -next_month Not Shown Not Shown Doorenbos J. & Pruitt W.O. (1984. Crop WaterRequirements. Food and Agriculture Organization of the United Nations. Edrees S., El-Azhary E. Tawifik K. Rafea A (1997). An Expert Systems for Citrus Production Management. In: Proceedings of the Fifth International Conference on Artificial Intelligence Applications, Cairo, Egypt. El-Dessouki, A., S. El-Azhary, and S. Edrees (1993). CUPTEX: An Integrated Expert System For Crop Management Of Cucumber. In: Proceedings of ESADW-93, Cairo, Egypt. MOALR.. Forbus K. & Kleer J. (1993) Building Problem Solvers, pp 151-170. The MIT press, Cambridge, Massachusetts, London. Hargreaves,G.H. and Z.A.Samani, (1982). Estimating potential evapotranspiration. Journal of Irrigation and Drainage Division 108, 225- 230. Kamel,A., K.Schroeder, J.Sticklen, A.Rafea, A.Salah, U.Schulthess, R.Ward, and J.Ritchie, (1995). An Integrated Wheat Crop Management System Based on Generic Task Knowledge Based Systems and CERES Numerical Simulation. AI Applications 9. Ostergard M.and Goodell P. (1992). Delivering Expert Systems to Agriculture: Experiences with CALEX/Cotton. In: Fourth International Conference On Computers In Agriculture, Orlando, Florida. American Society of Agricultural Engineers. Plant R. (1989). An integrated expert decision support system for agricultural management. Agricultural Systems 29, 49-66. Rafea A. (1998) Agricultural Expert Systems. In: The handbook of applied expert systems (J.Liebowitz (Ed)). CRC Press , LLC,USA. 
work_3vo6ohjwjzcebnzidwoautzbu4	----	.. . Expert Sysrems will1 Applications 41 (2014) 7579-7595 Bivariate quality control using two-stage intelligent monitoring scheme CrossMark - Ibrahirn Masood a.'!', Adnan Hassan 'Faculty of Mechanical and Manufacturing Engineering, Universiti Tun Hussein Onn Malaysia, 86400 Pant Raja, Batu Pahat, Johor, Malaysia b ~ a c u l t y ofMechanical Engineering, Universiti Teltnologi Malaysia, 81310 UTM Sltudai, Johor, Malaysia A R T 1 C L E I N F O A B S T R A C T Article history: Available online 5 June 2014 -- Keywords: Balanced monitoring Bivariate quality control Statistical features Synergistic artificial neural network Two-stage monitoring In manufacturing industries, it is well known that process variation is a major source of poor quality products. As such, monitoring and diagnosis of variation is essential towards continuous quality improve- ment. This becomes more challenging when involving two correlated variables (bivariate), whereby selection of statistical process control (SPC) scheme becomes more critical. Nevertl~eless, the existing tra- ditional SPC schemes for bivariate quality control (BQC) were mainly designed for rapid detection of unnatural variation with limited capability in avoiding false alarm, that is, imbalanced monitoring per- formance. Another issue is the difficulty in identibing the source of unnatural variation, that is, lack of diagnosis, especially when dealing with small shifts. In this research, a scheme to address balanced mon- itoring and accurate diagnosis was investigated. Design consideration involved extensive simulation experiments to select input representation based on raw data and statistical features, artificial neural net- work recognizer design based on synergistic model, and monitoring-diagnosis approach based on two- stage technique. The study focused on bivariate process for cross correlation function, p = 0.1-0.9 and mean shifts, p = k0.75-3.00 standard deviations. The proposed two-stage intelligent monitoring scheme (2s-IMS) gave superior performance, namely, average run length, ARLl= 3.18-16.75 (for out-of-control process), ARLO = 335.01-543.93 (for in-control process) and recognition accuracy. RA = 89.5-98.5%. This scheme was validated in manufacturing of audio video device component. This research has provided a new perspective in realizing balanced monitoring and accurate diagnosis in BQC. 2014 Elsevier Ltd. All rights reserved. 1. Introduction In manufacturing industries, when quality feature of a product involves two correlated variables (bivariate), a n appropriate SPC charting scheme is necessary to monitor and diagnose these variables jointly. Specifically, process monitoring refers t o t h e iden- tification of process condition either in a statistically in-control or out-of-control, whereas process diagnosis refers t o t h e identifica- tion of t h e source variable(s) for out-of-control condition. In addressing this issue, the traditional SPC charting schemes for BQC such as x2 (Hotelling, 7947), multivariate cumulative sum (MCUSUM) (Crosier, 1988), and multivariate exponentially weighted moving average (MEWMA) (Lowly. Woodall, Champ, & Rigdon, 1992; Prabhu 8 Rungel-, 1997) are known t o be effective in monitoring aspect. Unfortunately, they a r e merely unable to provide diagnosis information, which is greatly useful for a quality practi- tioner in finding t h e root cause error and solution for corrective action. Since then, major researches have been focused on diagnosis * Corresponding author. Tel.: +607 4537700. E-mail addresses: ibraliim@~~Lli~n.edu.~ny (I. Masood), atlnan@fl~m.ul~ii.~ny (A. Hassan). URLs: l~~~p://www.~~~I~~n~edu~my (I. Masood), http://www.utm.my (A. Hassan). aspect. Shewhart-based control charts with Bonferroni-type control limits (Alt, 1985), principle component analysis (PCA) (Jacltson, 1991), multivariate profile charts (Fuchs & Ben,jamini, 1994), r2 decomposition (Mason, Tracy, & Young, 1995) and Minimax control chart (Sepulveda & Nachlas. 1997), among other, have been investi- gated for such purpose. Further discussions on this issue can be found in Lowry and Montgomery (1995), I<ourti and MacCregor (1996), Mason, Tracy, and Young (1997) and Bcrsimis, Psaraltis, ancl Panaretos (2007). In the related study. development in soft computing technology has motivated researchers t o explore the use of machine learning (ML) technology for automatically recognizing SPC chart patterns towards improving capability in monitoring and diagnosis. Identifi- cation ofthese patterns coupled with engineering Imowledge of the process would lead t o more specific diagnosis information. Expert systems (ES) (Chih 8 Kollier, 1994; Chih & Kollier, 1995), Fuzzy inference system (FIS) (Wang & Chen, 2001), artificial neural net- work (ANN), decision tree learning (DT) (Guh 8, Shiue, 2005), and support vector machine (SVM) (Cheng 8 Cheng, 2008) methods. among others, have been studied in designing the advanced SPC pattern recognition schemes. Extensive literature review revealed t h a t most of t h e proposed schemes were developed based on http://dx.cloi.o1~g/10.1 01G/j.eswa.2014.05.042 0957-417410 2014 Elsevier Ltd. All rights reserved. 7580 I. Masood, R HassanlExpert Systems with Applications 41 (2014) 7579-7595 research i n ANN models such as an integrated bivariate SPC-ANN (Chen & Wang, 2004; Nial<i & Abbasi, 2005: Yu, Xi, & Zhou, 2009), novelty detector ANN (Zorriassatine, 7'anuoclt, & O'Brien. 2003), modular ANN (Guh, 2007), ensemble ANN (Yu &Xi, 2009), multi- module-structure ANN (El-Miclany, El-Baz, & Abcl-Elwahecl, 2010), hybrid learning ANN (Salehi, Bahreininejad, & Nalthai, 201 I ) , an integrated ANN-SVM (Salehi, Kazemzadeh, 8 Salmasnia, 2012), and feature-based ANN (Masood & Hassan, 2013). The integrated bivariate SPC-ANN schemes combined the tradi- tional SPC chart(s) with an ANN model. The traditional SPC chart(s) role for monitoring the existence of unnatural variation in bivariate process, whereas an ANN model roles for diagnosing the sources of variation. I n that case, an ANN model is utilized only when neces- sary, that i s , when an out-of-control signal is triggered. Inversely, the other schemes such as novelty detector ANN consist of fully ANN or fully ML-based model for monitoring and diagnosing simul- taneously. In that case, an ANN model is continuously utilized, that is, for triggering out-of-control signal and then, for identifying the sources of variation. Further discussion on these schemes can be found in (Masood & Hassan, 2010; Hachicha & Ghorbel. 2012). 1.1. Problem situation and solution concept When dealing with monitoring and diagnosis of bivariate pro- cess variation in mean shifts, based on process monitoring view- point, an effective bivariate SPC scheme should be able to identify out-of-control condition as quicltly as possible at the shortest ARL1 (average run length for out-of-control process. ARL, - + 1). Concurrently, it should be able to maintain small false alarm a t the longest ARL (average run length for in-control process. ARLO > 200). Nevertheless, the existing traditional SPC schemes were mainly designed by focusing on rapid detection of out-of- control condition (ARL, * I ) but it has limited capability in avoid- ing false alarm (ARL < 200). Fig. 1 illustrates the concepts of imbalanced monitoring vs. balanced monitoring as the central theme for this investigation. Based o n diagnosis viewpoint, an effective bivariate SPC scheme should be able to identify the source variable(s) of out-of-control condition a s accurate as possible. Nevertheless, it is difficult to cor- rectly recognize when dealing with small shifts ( ~ 1 . 0 standard deviation). Chih and Rollier (1994). Chih and Rollier (1995), Zorriassatine, Tannock, and O'Brien (2003), Chen and Wang (2004) and Yu and Xi (2009). for examples, have reported less than 80% accuracy for diagnosing mean shifts at 1.0 standard deviation. Among others, only Guh (2007) and Yu et al. (2009) reported the satisfied results ( >90% accuracy). The imbalanced monitoring and lack of diagnosis capability as mentioned above need further investigation. In order to minimize erroneous decision making in BQC, it is essential to enhance the overall performance towards achieving balanced monitoring (rap- idly detect process variationlmean shifts with small false alarm as shown in Fig. 1) and accurate diagnosis (accurately identify the sources ofvariation/mean shifts). Additionally, the BQCapplications are still relevant in today's manufacturing industries. In solving this issue, a two-stage intelligent monitoring scheme (2s-IMS) was designed to deal with dynamic correlated data streams of bivariate process. This paper is organized as follows. Section 2 describes a modeling of bivariate process data streams and patterns. Section 3 presents the frameworlc and procedures of the 2s-IMS. Section 4 discusses the performance of the proposed scheme in comparison to the traditional SPC. Section 5 finally outlines some conclusions. 2. Modeling of bivariate process data streams and patterns A large amount of bivariate samples is required for evaluating the performance of the 2s-IMS. Ideally, such samples should be tapped from real world. Unfortunately, they are not economically available or too limited. As such, there is a need for modeling of synthetic samples based on Lehman (1977) mathematical model. Further discussion on data generator can be found in Masood and Hassan (201 3). In bivariate process, two variables are being monitored jointly. Let Xl-i=(Xl.I,. . .,XI-24) and X2-i=(X2.1.. . . . X Z . ~ ~ ) represent 24 observation samples for process variable 1 and process variable 2 respectively. Observation window for both variables start with sam- ples i = (1,. . . -24). It is dynamically followed by (i + I ) , (i + 2) and so on. When a process is in the state of statistically in-control, samples from both variables can be assumed as identically and indepen- dently distributed (i.i.d.) with zero mean ( p o = 0) and unity standard deviation (go = 1). Depending on process situation, the bivariate samples can be in low correlation ( p = 0.1 -0.3), moderate correla- tion ( p = 0.4-0.6) or high correlation ( p = 0.7-0.9). Data correlation ( p ) shows a measure of degree of linear relationship between the twovariables. Generally, this data relationship is difficult to be iden- tified using Shewhart control chart as shown in Fig. 2. On the other hand, it can be clearly indicated using scatter diagram. Low corre- lated samples yield a circular pattern (circular distributed scatter plot), moderate correlated samples yield a perfect ellipse pattern, whereas high correlated samples yield a slim ellipse pattern. Disturbance from assignable causes on the component variables (variable-1 only, variable-2 only, or both variables) is a major source of process variation. This occurrence could be identified by various causable patterns such as mean shifts (sudden shifts), trends, cyclic, systematic or mixture. In this research, investigation was focused on sudden shifts patterns (upward and downshift shifts) with positive correlation ( p > 0). Seven possible categories of bivariate patterns were considered in representing the bivariate process variation in mean shifts as follows: N (0,O): both variables XI-i and X2-i remain in-control. US (1,O): shifted upwards, while X2.i remains in-control. US (0.1): X2.i shifted upwards, while remains in-control. US ( 1 , l ) : both variables Xl.i and X2.i shifted upwards. DS (1,O): shifted downwards, while X2.i remains in-control. DS ( 0 , l ) : X2.i shifted downwards, while remains in-control. DS ( 1 , l ) : both variables X1.i and X2.i shifted downwards. Reference bivariate shift patterns based on mean shifts f3.00 standard deviations are summarized in Fig. 3. Their structures are unique to indicate the changes in process mean shifts and data correlation. The degree of mean shifts can be identified when the center position shifted away from zero point (0,O). 3. Two-stage intelligent monitoring scheme As noted in Section I , an integrated MSPC-ANN was combined in a single-stage monitoring scheme (direct monitoring-diagnosis) as proposed in Chen and Wang(2004). Niaki and Abbasi (2005). and YLI et al. (2009). The other schemes based on fully ANN-based models as proposed in Zorriassatine, Tannocli, and O'Bricn (2003), C u h (2007). Yuand Xi (2009) and El-Midany et al. (2010) also can be classified as a single-stage monitoring scheme. In this research, two-stage mon- itoring scheme was investigated by integrating the powerful of MEWMA control chart and Synergistic-ANN model for improving the monitoring-diagnosis performance. Framework and pseudo- code (algorithm) for the proposed scheme are summarized in Figs. 4 and 5 respectively. It should be noted that an initial setting as fol- lows needs to be performed before it can be put into application: Load the trained raw data-ANN recognizer into the system. I. Masood, A. Hassan/Expert Systems with Applications 41 (2014) 7579-7595 Y I S e n s ~ t ~ v ~ t v in mean s h ~ R detection Capablhtv In false alann avo~dance I/ Shorter ARLl represents fastel Longer ARLO repiesenis smaller 1 detect~on of plocess mean sh~fts false alaim I t b U i 1 Current state I Imbalanced n~onitoring: able to detect process mean shifts rapidly (ARL, =1 1) but has limited capability to avoid false alarm (ARLO 5 200) I Desired state (for this research) 1 I / Balanced monltonn. [reasonable fol cu~rent ~ r a c t l c e i able to detect process mean 1 s h ~ f t s rapldly (ARL, Z, 1) and nlalnta~n small false alarm (ARLO >> 200) % I I I Ideal state i 1 Perfect balanced: able to detect process mean shifts as soon as possible (ARLI = 1) 1 I . . 1 w~tllout tnggenng any false alarm (ARLO = m) 1 L ~ ~ Fig. 1. Current state and desired state towards balanced monitoring. Set t h e values of means (p01,po2) and standard deviations (nol.ao2) of bivariate in-control process (for variables and X2.i). These parameters can be obtained based on historical or preliminary samples. Perform in-process quality control inspection until 24 observa- tion samples (individual or subgroup) to begin the system. Recognition window size is set to 24 observation samples (for variables Xl-i and X2_i) since it provided sufficient training results and statistically acceptable to represent normal distribution. Preli- minary experiments suggested that a smaller window size (<24) gave lower training result due to insufficient pattern properties, while a larger window size (>24) does not increase the training result but burden the ANN training. Rational to integrate the MEWMA control chart and the Synergistic-ANN model are based on preliminary experiments. Generally, the MEWMA control chart is ltnown to be effective for detecting bivariate process mean shifts more rapidly compared to the x2 control chart. Furthermore, it is very sensitive when deal- ing with small shifts (G1.00 standard deviations). Unfortunately, based on one point out-of-control detection technique, it gave lim- ited capability to avoid false alarm (ARb 6 200). This becomes more critical when the variables are highly correlated. In the related study, pattern recognition scheme using a Synergistic- ANN model gave better capability in avoiding false alarm (ARL,, > 200). As such, it can be concluded that process identifica- tion based on recognition of process data stream patterns (Synergistic-ANN model) is more effective compared to detection of one point out-of-control (MEWMA control chart). Nevertheless, different techniques should have their respective advantages in terms of pointlpattern discrimination properties. In order to fur- ther improve the monitoring performance (ARLl =. l , ARLO >> 200), it is useful to combine both discrimination properties (MEWMA control chart and Synergistic-ANN recognizer) by approaching I. Masood, R HassanlExpert Systems with Applications 41 (2014) 7579-7595 Shewhart Control Chart Scatter Diagram -- Low Correlation Low correlation 3 1 2 2 1 1 I 1 Y -2 I ' I' zo 41 m KI 100 1 '0 20 40 60 $0 Inn Nvrnbor orsamplcs -3 -2 -1 0 1 2 Number uisamplcs XI Moderate Correction Moderatc correlation 3 I 2 ..' ZO 40 60 SO 100 3~ 20 40 60 RO l W - 4 -3 -2 -1 0 1 2 3 Nlmbcl oisdnlplcs Numher urnampler X 1 High Correlation High correlation 3 3 I I I I 2 2 3~ 20 40 60 80 100 ' 0 20 10 60 80 IMI .4 -3 -2 -1 0 1 2 3 Nunlbel ofratnplcs Number ulsamples x1 Fig. 2. Shewhart control charts and its respective scatter diagrams. two-stage monitoring and diagnosis. In the first stage monitoring, the MEWMA control chart is used for triggering bivariate process mean shifts based on 'one point out-of-control' as per usual. Once the shift is triggered, the Synergistic-ANN recognizer will perform second stage monitoring and diagnosis through recognition of pro- cess data stream patterns that contain one of several out-of-control points. This approach is suited for 'recognition only when neces- sary' concept, that is, it is unnecessary to perform recognition while the process lies within a statistically in-control state. Besides, recognition is only necessary for identifying patterns sus- pected to a statistically out-of-control state. Besides producing smaller false alarm, this approach will also reduce computational efforts and time consumes for pattern recognition operation. 3.1. M E W M A control chart The MEWMA control chart developed by Lowry ct al. ( 1992) is a logical extension of the univariate EWMA control chart. In the bivariate case, the MEWMA statistics can be defined as follows: [U:(EWMAI, - 1 1 ) ' + u j ( E W M A z i - p 2 j 2 - 2 u : , ( E W M A l ; - p , ) ( E W M A z i - @ , j ] n M E W M A , = (u:.: - u:,, Covariance matrix of MEWMA: h M E W M A The standardized samples (Zli, Zzi) with cross correlation func- tion ( p ) were used. Thus, a1 = a2 = 1 ; 012 = p. Notations L and i rep- resent the constant parameter and the number of samples. The starting value of EWMA (EWMAo) was set as zero to represent the process target (/A,,). The MEWMA statistic samples will be out-of-control if it exceeded the control limit (H). In this research, three sets of design parameters (A, H: 0.05, 7.35; 0.10, 8.64; 0.20, 9.65) as reported in Prabhu and Runger (1997) were investigated. 3.2. Synergistic-ANN model pattern recognizer Synergistic-ANN model as shown in Fig. (5 was developed for ( 1 ) pattern recognizer. It is a parallel combination between two I. Masood, A. H a s s a n l E x p e r t S y s t e m s w i t h Applications 41 (2014) 7579-7595 . . . . . . .:. . -: - -5.0 n -5.0 -2.5 0.0 2.5 5.0 -5.0 -2.5 0.0 2.5 5.0 XI (Down-Shift) X 1 (Down-Shift) - Fig. 3. S u m m a r y o f bivariare s h i f t patterns for p = 0.1, 0.5 and 0.9 individual ANNs that are: ( i ) raw data-based ANN, and (ii) statisti- recognizers can be combined using simple summation: 01, = cal features-ANN as shown in Fig. 7. X(ORD-~,OF-~), where i = (1,. . . .7) are the number of outputs. Final Let OR^ = (ORD-i,. . . , ORD.7) and OF = (OF.l,. . . , OF.7) represent decision ( 0 ,,,,,,) was determined based on the maximum value seven outputs from raw data-based ANN and statistical features- from the c o m b i n e d ~ ~ u t ~ u t s : ANN recognizers respectively. Outputs from these individual Osynergy = max(O/l, . . . ,017) (5) 1. Masood, R HassanlExpert S y s t e m s w i t h Applications 41 (2014) 7579-7595 Bivariate shift patte~nsforrnoderate data correlation (p = 0.5) .- Partially developed shift Fully developed shift -5.0 -2.5 0.0 2.5 5.0 X l (Up-Shift) -5.0 -2.5 0.0 2.5 5.0 X I (Up-Shift) -5.0 -2.5 0.0 2.5 5.0 X I (Nonnal) h - 4 3 2.5 - w z . . b 0.0 3 .- . 5 - - -. 2 -25 5- -5.0 -5.0 -5.0 -2.5 0.0 2.5 5.0 -5.0 -2.5 0.0 2.5 5.0 X1 (Up-Shift) XI (Up-Shift) XI (Normal) u X I (Down-Shift) krFzFiT X I (Down-Shift) Fig. 3 ( c o n t i n u e d ) Multilayer perceptrons (MLP) model trained with back-propa- 48 neurons, while statistical features input representation gation (BPN) algorithm was applied for the individual ANNs. This requires only 14 neurons. The output layer contains seven neu- model comprises an input layer, one or more hidden layer(s) and rons, which was determined according to the number of pattern an output layer. The size of input representation determines the categories. Based on preliminary experiments, one hidden layer number of input neurons. Raw data input representation requires with 26 neurons and 22 neurons were selected for raw I. Mflsood, A. HossanfExpert S y s t e m s w i t h Applications 41 (2014) 7579-7595 X I (Down-Sh~ft) XI ( Down-Shift ) - --- - L --- -- ! Fig. 3 ( c o n t i n u e d ) data-based ANN and statistical features-ANN. The experiments did not improve the training results but provided poorer results. revealed that initially, the training results improved in-line with These excess neurons could burden the network computationally, the increment in the number of neurons. Once the neurons reduces the network generalization capability and increases the exceeded the required numbers, further increment of the neurons training time. I. Masood. R HassanlExpert Systems with Applications 41 (2014) 7579-7595 i......... ............................. ......................... ! First stage MEWMA , monitol-ing / control chart Next Yes (out-of-control) 11 Troubleshooting I and I i renew setting / / i ................................................ : I; t I YCS (out-of-control) I I ; Identify the sources of mean shift I ................................................................................................................................... Fig. 4. Frameworlc for the 2s-IMS. 3.3. lnput representation lnput representation is a technique to represent input signal into ANN for achieving effective recognition. There are various approaches could be used to represent input signal. Raw data (standardized samples) is the basic approach (Zorriassatine, Tannock, & O'Bricn. 2003). Besides raw data, feature-based approach that involves extracted features from raw data is one of the successful technique in image processing (Br~~nzcll & Eriltsson, 2000: I<losgen Q Zytkow, 2002). This approach has also been applied in the area of univariate quality control (UQC), which is aim to improve accuracy for recognizing univariate control chart patterns (i.e., normal, upward shifts, upward trends, downward shifts, downward trend, and cyclic) by reducing network size, corn- putational efforts and training time of an ANN (Gauri & Chaltraborty, 2006: Gauri & Chaltraborty, 2008: Guh. 2010; Hassan, Nabi Baltsh, Shaharoun, 5 Jamaludin, 2003; Pham B Wani, 1997). Pham and Wani (1997) firstly investigated nine shape features. Then, it was improved by Cauri and Chaltraborty (2006) and Gauri and Chaltraborty (2008). In the related study, Hassan ct al. (2003) proposed six statistical features comprising of mean, standard deviation, skewness, mean-square value, autocorrelation, and last value of CUSUM. Guh (2010) proposed another six statis- tical features comprising of mean, standard deviation, sltewness, Iturtosis, slope, and Pearson correlation coefficient. More recently. Masood and Hassan (2013) proposed another set of statistical features for BQC. A few researchers have combined features and raw data in a serial form, i.e.. x2-statistics with raw data (Guh, 2007) and statis- tical features with raw data (Yu & XI, 2009; Yu et al., 2009), for strengthening pattern properties in BQC. Nevertheless, this approach has increased network size, computational effort, and training time. This becomes more difficult to implement in a complex case. The other approach in UQC is using multi-resolution wavelet analysis (MRWA) for denoising or filtering raw data through several decomposition levels without changing the network size (Al-Assaf, 2004; Assalch & Al-Assaf, 2005), concurrent pattern rec- ognition (Chcu, Lu, & Lam, 2007), and image processing (Wang, I ~ L I O , & Qi, 2007). The MRWA has played a crucial role for process I. Masood, A. HassanlExpert Systems with Applications 41 (2014) 7579-7595 Step 1: Invut samples (bivariatel. Window size = 24, starting observation samples are: XI., = (Xi. ,,.., and X 2 ~ ; = (%.I ,.. ., X2.24). It is followed by (ilh + I ) , (iIh + 2) and so on. Slep 2: Samples standardizalion. Rescale observation samples inlo a slandardized range wlthi~i [-3, +3]: ZI = (XI -~I,,~)/OO~ and Z2 = (X* -p,JJ/q,2 Input samples (original) and standardized samples can be represented graphically using Shewhart control charts and scatter diagram. Step 3: First stage monitoring (MEWMA coi~trol c h a ~ t ) : Standardized samples (ZJ. 2:) are converted into MEWMA statistics (?n,J;rrw,,). Decision rulc: If' T'MElr.orx < Upper control limit (11) Proceed to the next samples else Proceed to the second stage monitoring. md Stcp 4: Input revresentation. (i) Raw data-based input representation: ZI.PI, Z2~~.,, . . . , ZI.I.?<, Z2.P24 (ii) Statistical features input representation: - Fealure extraction: the slaiidardized samples (21, Z?) are converted inlo statistical featul-es, namely, the last value of exponentially weighted moving average (LEWMAJ with h = (0.25, 0.20, 0.15, 0.10), mean (11). inultiplication of meall with standard deviation (MSI)), and multiplication of mean with mean square value (MMSV). - Fou~tlieen extracted features are represented as follows: (LEWMAI,I,_pl, LEWMAo.zo-rl, LEWMAO.IS_PI, LEWMAUIU_PII IPI, M S D P I , MMSVrj, LEWMAO.IS_PI, LEWMAo,zo_r2, LEWMAo I S - ~ 2 . L E W A O I O - P ~ I prr, M S b , MMSV,,). Number of salnplcs i s dcnoted by P. Step 5 : Second staee monitorins and d~amlosis (svnereistic-ANN recoenizer): Decision mle: y Maximum output of ANN belongs to N(0,O): Process is "in-control"; proceed to the next samples else Process is "out-of-control"; identify the sourccs of lnean sliiits; perform diagnosis, troubleshooting; renew setting and return to Step 1 (out of scopes of this research). end Fig. 5. Pseudo-code for the 2s-IMS. I I ANN-based C o m b i n a t i o n of the ! ORD.I OKII-2 - , ta- O X I > . ~ Raw data- own.. Kaw a a b a s e d Statistical F e a t u r e s Features- OF.S I I Fig. 6. Synergistic-ANN model. controlling or monitoring. Insufficient denoising will distort recognizer for improving pattern discrimination capability. Raw waveforms and introduce errors. Inversely, excessive denoising data input representation consists of 48 data, i.e., 24 'consecutive will over-smooth the sharp features of underlying signals by standardized samples of bivariate process (Z1.P1l Z1.p2,. . . .Z24.P1. recognizing them as noise or outliers. Z24.P2). Statistical features input representation consists of last In this research, raw data and improved set of statistical features value of exponentially weighted moving average (LEWMA]) with were applied separately into training of the Synergistic-ANN A = [0.25,0.20,0.15,0.10], mean ( p ) , multiplication of mean with I. Masood, A HassanlExpert Systems w i t h Applicatioiu 41 (2014) 7579-7595 Fig. 7. Individual ANN recognizer. W L g r e r - L w e r L W (48 01 02 0 3 0 4 4 0 6 0 7 Raw data-ANN standard deviation (MSD), and multiplication of mean with mean square value (MMSV). Each bivariate pattern was represented by 14 data as follows: LEWMA0.25-~~, LEWM&.20-P1, LEWM&.15.Pl, LEWMAO.IO-PI, , & I , MSDPI. MMSVpi, LEwMAo.25-~2, LEWMAo.2o-p2. LEWMAO.IS-P~. LEWMAO.IO-PZ~ PPZ, MSDPZ, MMSVp2. LEWMAn features were talten based on observation win- dow = 24. The EWMA-statistics as derived using Eq. (6) incorpo- rates historical data in a form of weighted average of all past and current observation samples (Lucas cF Saccucci. 1990): W L W - L w ~ L w (14 0 1 02 0 3 0 4 05 0 6 0 7 Statistical feature-ANN Xi represents the original samples. In this study, the standardized samples (Zi) were used instead of Xi so that Eq. (6) becomes: where 0 < A ,< 1 is a constant parameter and i = [ I , 2 , . . . ,241 are the number of samples. The starting value of EWMA (EWMb) was set as zero to represent the process target (PO). Four value of constant parameter (A = 0.25,0.20,0.15,0.10) were selected based on a range within [0.05,0.40] recommended by Lucas and Saccurci (1990). Besides resulting longer ARLO, these parameters could influence the performance of EWMA control chart in detecting process mean shifts. Preliminary experiments suggested that the EWMA with small constant parameter ( A = 0.05) were more sensitive in identify- ing small shifts ( ~ 0 . 7 5 standard deviations), while the EWMA with large constant parameter (A = 0.40) were more sensitive in identify- ing large shifts (>2.00 standard deviations). The MSD and MMSV features were used to magnify the magnitude of mean shifts ( P ~ ~ P Z ) : where (,u1,p2), ( a l , 0 2 ) (,uf,p:) are the means, standard deviations and mean square value respectively. The mathematical expressions of mean and standard deviation are widely available in textbook on SPC. The mean square value feature can be derived as in Hassail ct al. (2003). Further discussion on selection of statistical features can be found in Masood ancl Hassan (2013). 3.4. Recognizer training and testing Partially developed shift patterns and dynamic patterns were applied into the ANN training and testing respectively since these approaches have been proven effective to suit for on-line process situation (Gi~h, 2007). Detail parameters for the training patterns are summarized in Tablcs 1 ancl 2. In order to achieve the best training result for overall pattern categories, the amount of training patterns were set as follows: (i) bivariate normal patterns = [I500 x (total combination of data correlation)] and (ii) bivariate shift patterns = 1100 x (total combi- nation of mean shifts) x (total combinations of data correlation)]. In order to improve discrimination capability between normal and shift patterns, a huge amount of N (0,O) patterns was applied into ANN training. The US ( 1 , l ) and DS ( 1 , l ) pattern categories also require a huge amount of training patterns since it contain a more complex combination of mean shifts compared to the other bivar- iate shifts pattern categories. Guh (2007) reported that the utilization of partially developed shift patterns in ANN training could provide the shorter ARLl results. In order to achieve the best ARLl results for this scheme, different percentage of partially developed shift patterns were utilized for different range of mean shifts as shown in Table 2. The starting points of sudden shifts (SS) were determined empiri- cally. The actual value of data correlation is dependent to the var- iability in the bivariate samples. The simulated values ( p = 0.1,0.3, 0.5, 0.7. 0.9) as shown in Table 1 could only be achieved when the process data streams are in fully normal pattern or in fully devel- oped shift pattern. Input representations were normalized to a compact range between [-1,+1]. The maximum and the minimum values for normalization were talcen from the overall data of train- ing patterns. Based on BPN algorithm. 'gradient decent with momentum and adaptive learning rate' (traingdx) was used for training the MLP model. The other training parameters setting were learning rate (0.05) learning rate increment (1.05). maximum number of epochs (500) and error goal (0.001), whereas the network performance was based on mean square error (MSE). Hyperbolic tangent func- tion was used for hidden layer, while sigmoid function was used for an output layer. The training session was stopped either when the number of training epochs was met or the required MSE has been reached. 4. Performance results and discussion The monitoring and diagnosis performances of 2s-IMS were evaluated based on average run lengths (ARLo,ARLl) and recogni- tion accuracy (RA) as summarized in Table 4. The ARLs results were also compared to the traditional multivariate statistical process control (MSPC) charting schemes such as X 2 (Hotelling, 1947). MCUSUM (Pignatiello & Kunger, lf)90), and MEWMA (Lowry et dl., 1992), as reported in the literature. In order to achieve balanced monitoring and accurate diagnosis, the proposed 2s-IMS should be able to achieve the target perfor- mances as follows: I. Masood. A. Hossan/Expert Systems with Applications 41 (2014) 7579-7595 Table 1 . . Parameters for t h e training patterns. Pattern category M e a n shift ( i n standard deviations) Data correlation ( p ) Amount of training patterns N ( 0 , o ) X I : 0.00 0.1, 0.3, 0.5, 0.7, 0.9 1500 x 5 = 7500 X2: 0.00 US (1,o) X l : 1.00, 1.25,. . ..3.00 100 x 9 x 5 = 4 5 0 0 X2: 0.00, 0.00,. . .,o.oo US ( 0 , l ) X2: 0.00, O . O O , . . ..O.OO l O O x 9 x 5 = 4 5 0 0 Xl: 1.00, 1.25,. . ..3.00 us (1.1) Xl: 1.00, 1.00, 1.25, 1.25,. ..,3.00 100 x 25 x 5 = 12.500 X2: 1.00, 1.25, 1.00, 1.25. . . ..3.00 DS (1.0) X l : -1.00, -1.25.. . ., -3.00 l O O x 9 x 5 = 4 5 0 0 X2: 0.00, o.oo,.. ..o.oo DS ( 0 , 1 ) X2: 0.00, 0.00 ,..., 0.00 lOOx9x5=4500 X l : -1.00, -1.25,. . . , -3.00 0s ( 1 , 1 ) Xl: -1.00, -1.00, -1.25, -1.25,. . .,-3.00 100 x 25 x 5 = 12,500 X 2 : -1.00, -1.25, -1.00, -1.25,. . ..-3.00 Table 2 Parameters for the partially developed shift patterns. Range of mean shifts ( i n standard deviations) A m o u n t of partially developed shift patterns Starting point of sudden shift (SS) Sample 9 t h Sample 1 3 t h Sample 1 7 t h Table 3 Summary of monitoring-diagnosis capabilities. Traditional MSPC 2s-IMS Effective in monitoring (to identify out-of-control signal) Limited to avoid false alarm (ARLO r 200) Unable to identify the sources of variation (mean shifts) Comparable to the traditional MSPC in monitoring aspect Capable to maintain smaller false alarm (ARLO >> 200) High accuracy in identifying the sources of variation (mean shifts) (i) ARI, >> 200 to maintain small false alarm in monitoring bivariate in-control process. (ii) Short ARLl (average ARLl 6 7.5 for shifts range k0.75-3.00 standard deviations) to rapidly detect bivariate process mean shifts. (iii) High RA (average RA 2 95% for shifts range 20.75-3.00 stan- dard deviations) to accurately identify the sources of mean shifts. 4.1. Monitoring performance In monitoring aspect. the ARb represents the average number of natural observation samples of in-control process before the first out-of-control process signal exist as a false alarm. In other word, the ARLO measures how long a SPC scheme could maintain an in- control process running without any false alarm. On the other hand, the ARLl represents the average number of unnatural obser- vation samples before it is truly identified as out-of-control process signal. In other word, the ARLl measures how fast a SPC scheme could detect process mean shifts. Further discussion on this mea- sure can be referred to Montgoluery (2009). Ideally, a SPC scheme should provide ARLO as long as possible in order to minimize cost for investigating the discrepancy and trou- bleshooting while the process still within control. Meanwhile, it should provide ARLl as short as possible in order to minimize cost for reworlts or waste materials. Since t h e false alarm cannot be eliminated, the ARLO >> 200 is considered as the de facto level for balanced monitoring. In this research, the ARLs results of 2s-IMS were simulated based on correctly classified patterns. Generally, it can be observed that the smaller the mean shifts, the longer the ARLl values. This trend support the conclusion that process mean shifts with smaller magnitudes would be more difficult to detect. Specifically, the 2s-IMS indicated rapid detection capability for large shifts (shifts = 3 a . ARL1 = 3.18-3.19) and moderate shifts (shifts = 2 a , ARLl = 4.76-4.78) with short ranges of ARL1. It was also capable to deal with smaller shifts (shifts=Ilo.0.75a], ARL1= 110.33-10.60,15.69-16.751). In comparison to the X2 charting scheme, detection capability as shown by 2s-IMS was faster for small and moderate shifts (shifts = 0.75-20). In comparison to the MCUSUM and the MEWMA, it was slightly comparable in rapid detection for large shifts (shifts = 2.50, ARL,: 2s-IMS = 3.80-3.81, MCUSUM = 2.91, MEWMA=3.51) and moderate shifts (shifts=1.5o. ARL,: 2s-IMS = 6.41-6.52, MCUSUM = 5.23; MEWMA = 6.12). Similar trend can also be found when dealing with smaller shifts (shifts = la, ARLI: 2s-IMS = 10.33-10.60, MCUSUM = 9.28; MEWMA = 10.20). Meanwhile, based on the range of ARLO results ( p = 0.1,0.5, 0.9; ARLo=335.01, 543.93, 477.45), the 2s-IMS was observed to be more effective in maintaining smaller false alarm compared to the traditional MSPC (ARb r 200). It should be noted that the results for medium and high correlation processes have exceeded 370 as shown in the Shewhart control chart (Nels011, 1985; Shewhart, 1931). Overall, it can be concluded that the proposed scheme indicated balanced monitoring performance. 4.2. Diagnosis performance In diagnosis aspect, the RA measures how accurate is a SPC scheme could identify the sources of mean shifts towards diagnos- ing the root cause error and conducting troubleshooting. Generally, it can be observed that the smaller the mean shifts, the lower the RA results. This trend supports the conclusion that diagnosis infor- mation for small process mean shifts (61.0 standard deviations) - 7590 I. Masood, A. HassanjExpert Systems with Applications 41 (2014) 7579-7595 Table 4 . Performance comparison between t h e 2s-IMS and t h e traditional MSPC. Pattern category Mean shifts Average run lengths Recognition accuracv 2s-IMS X2 UCL = 10.6 MCUSUM k = 0.50 h = 4.75 MEWMA I = 0.10 H = 8.66 2s-IMS XI X2 ARLO for p = 0.1. 0.5. 0.9 ARLO for p = 0.0 RA for p = 0.1, 0.5. 0.9 N (0,o) 0.00 0.00 335.01. 543.93. 477.45 200 (0.005) 203 (0.0049) 200 (0.005) N A ARL, for p = 0.1, 0.5, 0.9 us (1,O) 0.75 0.00 17.60. 18.34, 20.00 92.7. 90.4. 89.5 US ( 0 , l ) 0.00 0.75 16.20, 15.99, 16.21 92.9. 89.3. 90.6 US (1.1) 0.75 0.75 13.64. 13.28, 14.17 82.4. 94.8, 99.9 DS (1,O) -0.75 0.00 16.31, 16.43, 17.35 92.3, 89.2. 89.4 Ds ( 0 , l ) 0.00 -0.75 16.94, 17.44, 18.75 92.3. 87.8. 88.5 DS (1.1) -0.75 -0.75 13.46. 13.37. 14.03 84.1. 96.1. 99.9 Average 15.69, 15.81, 16.75 89.5, 91.3, 93.0 us (1,o) 1 .OO 0.00 11.52. 11.57, 11.70 42-0.976 9.28-0.892 95.3, 93.1, 94.4 us ( 0 , l ) 0.00 1 .OO 10.50, 10.22, 10.20 95.8. 93.5, 94.4 Us ( 1 , l ) 1 .OO 1.00 9.1 6. 9.09. 9.66 90.0, 96.5, 100 Ds (1.0) -1 .OO 0.00 10.99, 10.86, 11.06 95.3. 93.2. 92.3 0 s ( 0 , l ) 0.00 -1.00 11.08, 11.12, 11.36 93.8, 92.1, 92.6 Ds (1,1) -1.00 -1.00 9.15. 9.12. 9.63 89.5. 98.0. 100 Average 10.40, 10.33, 10.60 93.3. 94.4. 95.6 us (1,O) 1.50 0.00 7.02, 7.07, 7.03 15.8-0.937 5.23-0.809 97.4, 96.5, 97.1 us (0,1) 0.00 1.50 6.54, 6.33, 6.40 97.1. 96.5. 96.2 U s ( 1 , l ) 1.50 1.50 5.82, 5.73. 5.94 91.7. 97.9. 100 DS (1,O) -1.50 0.00 6.81, 6.81, 6.92 97.4, 96.3, 95.5 0 s ( 0 , l ) 0.00 -1.50 6.82, 6.80, 6.85 96.2. 95.8. 95.6 0s (1,1) -1.50 -1.50 5.81. 5.69. 5.98 93.2. 99.0. 100 Average 6.47, 6.41, 6.52 95.5. 97.0. 97.4 us (1.0) 2.00 0.00 5.23, 5.15, 5.19 97.8, 97.1, 97.6 Us (0.1) 0.00 2.00 4.80. 4.72, 4.70 97.7. 97.8. 97.1 Us (1,1) 2.00 2.00 4.36, 4.32, 4.39 91.6, 98.4, 100 DS (1,o) -2.00 0.00 5.04, 5.04. 5.02 96.8. 96.7. 96.6 0 s (0,1) 0.00 -2.00 4.97, 5.03, 4.98 96.5. 96.5. 95.6 DS(1,l) -2.00 -2.00 4 29, 4.27 4 33 93.7. 98.9. 100 Average 4.78. 4.76. 4.77 95.7, 97.6. 97.8 Us (1,O) 2.50 0.00 4.10, 4.14, 4.12 98.0, 98.4, 98.0 US (0.1) 0.00 2.50 3.83, 3.81, 3.81 97.3. 97.4, 97.0 US (1.1) 2.50 2.50 3.54, 3.49, 3.53 93.2, 98.4, 100 DS (1,O) -2.50 0.00 3.99. 3.96. 3.95 97.3. 97.3, 97.0 0 s ( 0 , l ) 0.00 -2.50 3.97. 4.02. 3.98 96.5. 96.6. 97.0 D S ( 1 , l ) -2.50 -2.5O 3.41.3.40. 3.46 94.9. 98.8. 100 Average 3.81, 3.80, 3.81 96.2. 97.8. 98.2 Us (1.0) 3.00 0.00 3.47, 3.46, 3.46 98.6, 98.3. 98.2 U s (0.1) 0.00 3.00 3.20, 3.20. 3.21 97.8. 97.8, 98.0 US ( 1 , l ) 3.00 3.00 2.98, 2.93, 2.98 93.8. 98.4, 100 DS (1,o) -3.00 0.00 3.31, 3.30. 3.27 98.0, 97.1. 97.6 0 s ( 0 , l ) 0.00 -3.00 3.33.3.32, 3.32 96.7, 97.1, 97.1 DS(1,l) -3.00 -3.0° 2.84. 2.85. 2.90 94.6. 99.1. 100 Average 3.19, 3.18. 3.19 96.6. 98.0, 98.5 Grand average +(0.75-3.00) 7.39, 7.38, 7.61 94.5. 96.0, 96.8 Note: Design parameters for MEWMA control c h a r t in 2s-IMS (1= 0.1, H = 8.64). would be more difficult to identify. Specifically, the 2s-IMS indicated accurate diagnosis capability for large shifts (shifts = 30, RA = 96.6-98.5%) and moderate shifts (shifts = 20, RA = 95.7- 97.8%) with high ranges of RA. Although the results were slightly Flange Internal diametc degraded, it is still effective to deal with smaller shifts (shifts = [lo,0.75o], RA= [93.3-95.6%,89.5-93.0%]). It should be 1' noted that the RA results for medium and highly correlated pro- /' cesses were higher compared to low correlation process, which is i effective for practical case. Since the traditional MSPC charting i 1, n schemes were unable to provide diagnosis information, diagnosis \, 1" capability as shown by 2s-IMS was absolutely capable in solving \~ such issue. Overall, it can be concluded that the proposed scheme '> 1. indicated accurate diagnosis performance. Table 3 summarizes the comparison of monitoring-diagnosis capabilities between the Groove & flange view Roller head 2s-IMS and the traditional MSPC. Fig. 8. Functional features of roller head. I. Masood, A. HassanjExpert Systems with Applications 41 (2014) 7579-7595 Extnision round bar Tuining to rough size Turning to size Honing inner diameters N~ckel electroplating Beal-ings assembly Fig. 9. Process plan for the manufacture of roller head. 5. Industrial case study . Broadly, the need for BQC could be found in manufacturing industries involved in the production of mating, rotational or mov- ing parts. Investigation for this study was focused on the manufac- turing o f audio video device (AVD) component, namely, roller head. This investigation was based on the author's working experience in manufacturing industry in Johor, Malaysia. In an AVD, the roller head functions to guide and control the movement path of a film tape. Inner diameters of roller head (ID1 and ID2) as shown in Fig. 8 a r e two dependent quality characteristics (bivariate) that need for joint monitoring-diagnosis. In current practice, such func- tional features are still widely monitored independently using Shewhart control charts. It is unsure why the MSPC was not imple- mented. Based on the author's point of view, it could be due to lack of motivation, ltnowledge and sltills to adapt new technology. The process plan for the manufacture of roller head can be illus- trated in Fig. 9. Initially, an aluminium extrusion round bar was turned t o rough size (rough cut machining). Then, it was turned to size (finish cut machining) to form functional features such as inner diameters, and groove and flange, among others. The machining of inner diameters was then extended into honing pro- cess to achieve tight tolerance for bearing assembly. Hard coated surface was also necessary. As such, the machined work-piece was electroplated using nickel alloy before assembly. rTool rRoller head <. ..-.---.-. ,./' Tool bluntness Loading error (Decrement in ID) (Increment in ID) Fig. 10. Process variation occurred in turning-to-size operation. automatically loaded into pneumatic chuck using a robotic system. Bluntness in the cutting tool will cause gradual decrement in both inner diameters (ID1 ,ID2) with positive cross correlation ( p > 0). In another situation. such inner diameters could be suddenly increased simultaneously and yields positive cross correlation ( p > 0) due to loading error. Based on two examples of bivariate process variation, industrial process samples were simulated into the 2S-IMS for validating its applicability in real world. The first case study involves tool bluntness. The mean ( p ) and standard deviation (o) of bivariate Bivariate process variation can be found in turning to size oper- in-control process were determined based on the first 24 samples ation d u e to tool bluntness and loading error as illustrated in (observations lst-24th). Tool bluntness begins between observa- Fig. 10. These disturbances will cause unnatural changes in t h e tion samples 41st-50th. Validation results are summarized in process data streams as shown in '1-able 5. The work piece is Table (5, whereby the determination of process status (monitoring) Table 5 Sources of variation in machining inner diameters. Stable process Process noise N (0,O) Tool bluntness DS ( 1 , l ) Loading error US (1.1) XI., ([Dl) 1~:: f l w w t . I-"_- -- N o r m a l _ N o r m a l " k M L-.. Down-Trend I -" Down-Trend h L Up-Shift Scatter diagram i.01-1 u XI ( Down-Trend ) I. Masood R HassanjExpert Systems with Applications 41 (2014) 7579-7595 - - 7592 Table 6 . - . Inspection results based on tool bluntness case. i Original samples Standardized samples Window range Monitoring-diagnosis decision Xi.1 (ID11 Xi-2 (ID2) 4 1 ((Dl) Zi-1 (ID2) 1 7.9420 7.9428 0.3393 1.0790 2 7.941 2 7.9420 -1.1414 -0.591 7 3 7.9412 7.941 6 -1.1414 -1.4271 4 7.9420 7.9428 0.3393 1.0790 5 7.9412 7.9420 -1.1414 -0.5917 6 7.9412 7.941 6 -1.1414 -1.4271 7 7.9420 7.9428 0.3393 1.0790 8 7.9424 7.9420 1.0797 -0.5917 9 7.941 6 7.9420 -0.401 0 -0.591 7 1 0 7.941 2 7.941 6 -1.1414 -1.4271 11 7.941 6 7.9424 -0.401 0 0.2437 1 2 7.9428 7.9432 1.8201 1.9144 1 3 7.9420 7.9424 0.3393 0.2437 1 4 7.941 6 7.9424 -0.4010 0.2437 1 5 7.9424 7.9428 1.0797 1.0790 16 7.9412 7.942 -1.1414 -0.5917 1 7 7.941 2 7.9416 -1.1414 -1.4271 1 8 7.9420 7.9424 0.3393 0.2437 1 9 7.9428 7.9428 1.8201 1.0790 20 7.9420 7.9424 0.3393 0.2437 2 1 7.941 2 7.9416 -1.1414 -1.4271 22 7.9424 7.9428 1.0797 1.0790 23 7.9424 7.9424 1.0797 0.2437 2 4 7.9420 7.9424 0.3393 0.2437 1-24 N 25 7.941 2 7.941 6 -1.1414 -1.4271 2-25 N 26 7.9424 7.9420 1.0797 -0.591 7 3-26 N 27 7.9424 7.9428 1.0797 1.0790 4-27 N 28 7.941 2 7.9420 -1.1414 -0.591 7 5-28 N 29 7.9420 7.9428 0.3393 1.0790 6-29 N 3 0 7.9420 7.9424 0.3393 0.2437 7-30 N 3 1 7.9412 7.9420 -1.1414 -0.5917 8-31 N 3 2 7.9420 7.9428 0.3393 1.0790 9-32 N 3 3 7.9428 7.9424 1.8201 0.2437 10-33 N 3 4 7.941 6 7.9424 -0.4010 0.2437 11 -34 N 3 5 7.9424 7.9432 1.0797 1.9144 12-35 N 3 6 7.9428 7.9424 1.8201 0.2437 13-36 N 3 7 7.941 6 7.9420 -0.401 0 -0.591 7 14-37 N 3 8 7.9420 7.9424 0.3393 0.2437 15-38 N 3 9 7.9424 7.9420 1.0797 -0.5917 16-39 N 40 7.941 6 7.9420 -0.4010 -0.5917 17-40 N 41 7.9408 7.9412 -1.8818 -2.2625 18-41 N 42 7.9408 7.9408 -1.8818 -3.0978 19-42 N 43 7.9404 7.9408 -2.6222 -3.0978 20-43 N 44 7.9404 7.9408 -2.6222 -3.0978 21-44 DS ( I 1 ) 4 5 7.9404 7.9404 -2.6222 -3.9332 22-45 DS ( I I ) 46 7.9400 7.9404 -3.3626 -3.9332 23-46 D5 (1 1 ) 4 7 7.9400 7.9400 -3.3626 -4.7686 24-47 DS (1 1 ) 48 7.9396 7.9400 -4.1029 -4.7686 25-48 DS (1 1) 4 9 7.9396 7.9396 -4.1 029 -5.6040 26-49 DS ( I I ) 50 7.9396 7.9396 -4.1 029 -5.6040 27-50 DS ( I 1 ) ( p l , p z ) = (7.9417,7.9422): (g,,02) = (4.6687 x 10-~,4.2495 x Note: Observation samples highlighted in bold (41st-50th) represent out-of-control process. Table 7 Outputs of the scheme for tool bluntness case. Reco~nition window (RW) 1-24 2-25 3-26 4-27 5-28 6-29 7-30 8-31 9-32 P Decision based on MEWMA control chart RW P Decision based on MEWMA control chart RW P N US (I 0 ) US ( 0 1 ) us (1 1 ) DS (1 0) DS ( 0 1 ) DS (1 1 ) Note: Bold value represents the maximum output of ANN that determines pattern category. I. Mosood. A. HonanfExpert Systems with Applications 41 (2014) 7579-7595 7593 and sources of variation (diagnosis) are based on output of the scheme as shown in Table 7. In t h e first 40 samples, this scheme was able to correctly recog- nize t h e bivariate process data streams as in-control patterns (N). In this case, it was effective to identify bivariate in-control process without triggering any false alarm. Bluntness of the cutting tool begins a t sample 41st, whereby this scheme was able to correctly recognize bivariate process data streams as Down-Shift patterns (DS (1 , I ) ) starting from sample 44th (at window range 21st-44th). In overall diagnosis aspect, this scheme was observed to be effective to identify the sources of variation in mean shifts without mistalte. The second case study involves loading error. Similar as in the first case study, the mean ( p ) and the standard deviation (o) of bivariate in-control process were computed based on the first 24 observation samples. Loading error exist between samples 40th- 50th. Validation results and related output of the scheme are sum- marized in Tables S a n d 9 respectively. Based on t h e first 39 samples, this scheme is effective to cor- rectly recognize the bivariate process data streams as in-control patterns (N). In this situation, the process was running smoothly without false alarm. Improper condition of pneumatic chuck and robotic arm causes loading error between samples 40th-50th. In this situation, this scheme was able to correctly recognize the bivariate process data streams as Up-Shift patterns (US (1 , l ) ) start- ing from sample 40th (at window range: 17th-40th). In overall diagnosis aspect, this scheme is capable to correctly identify the sources of variation in mean shifts without mistalte. Table 8 Inspection results based on loading error case. i Original sa~nples Standardized samples Window range Monitoring-diagnosis decision XILI ([Dl 1 x,.z (ID21 Z i ~ l (ID1 ) zi-i (ID21 1 7.941 6 7.9420 -0.2856 -0.5099 2 7.9412 7.9428 -1.1424 1.3727 3 7.9420 7.9424 0.5712 0.43 1 4 4 7.941 2 7.941 6 -1.1424 -1.4512 5 7.9420 7.9428 0.571 2 1.3727 6 7.941 2 7.9420 -1.1424 -0.5099 7 7.941 2 7.941 6 -1.1424 -1.4512 8 7.941 6 7.9424 -0.2856 0.4314 9 7.9424 7.9420 1.4279 -0.5099 10 7.941 6 7.9420 -0.2856 -0.5099 11 7.941 2 7.941 6 -1.1424 -1.4512 1 2 7.9424 7.9428 1.4279 1.3727 13 7.9420 7.9424 0.571 2 0.4314 14 7.941 6 7.9424 0 . 2 8 5 6 0.4314 15 7.9412 7.9416 -1.1424 -1.4512 16 7.9424 7.9428 1.4279 1.3727 17 7.941 6 7.9420 -0.2856 -0.5099 18 7.9420 7.9424 0.5712 0.4314 19 7.9412 7.9420 -1.1424 -0.5099 20 7.9424 7.9424 1.4279 0.4314 21 7.9420 7.9424 0.5712 0.4314 22 7.9420 7.9424 0.5712 0.43 1 4 23 7.941 2 7.9416 1 . 1 424 -1.4512 24 7.9424 7.9428 1.4279 1.3727 1-24 N 25 7.9420 7.9424 0.571 2 0.4314 2-25 N 26 7.9412 7.9416 -1.1424 -1.4512 3-26 N 27 7.9424 7.9420 1.4279 -0.5099 4-27 N 28 7.9424 7.9428 1.4279 1.3727 5-28 N 29 7.9412 7.9420 -1.1424 -0.5099 6-29 N 30 7.9420 7.9428 0.5712 1.3727 7-30 N 3 1 7.9428 7.9424 2.2847 0.4314 8-31 N 32 7.9420 7.9424 0.5712 0.43 1 4 9-32 N 33 7.941 2 7.9420 -1.1424 -0.5099 10-33 N 34 7.9420 7.9428 0.5712 1.3727 11-34 N 35 7.9428 7.9424 2.2847 0.43 1 4 12-35 N 36 7.941 6 7.9424 -0.2856 0.4314 13-36 N 37 7.9424 7.9428 1.4279 1.3727 14-37 N 3 8 7.9416 7.9420 -0.2856 -0.5099 15-38 N 39 7.9428 7.9424 2.2847 0.43 1 4 16-39 N 40 7.9428 7.9432 2.2847 2.3140 17-40 US (1 1 ) 41 7.9432 7.9428 3.1415 1.3727 18-41 US (1 1 ) 42 7.9436 7.9432 3.9982 2.3140 19-42 US (1 1 ) 43 7.9428 7.9432 2.2847 2.3140 20-43 US (1 1 ) 44 7.9432 7.9428 3.1415 1.3727 21 -44 US (1 1 ) 45 7.9436 7.9432 3.9982 2.3140 22-45 US (1 1 ) 46 7.9428 7.9432 2.2847 2.3140 23-46 US (1 1 ) 47 7.9432 7.9428 3.1415 1.3727 24-47 US (1 1 ) 48 7.9428 7.9436 2.2847 3.2553 25-48 US(1 I ) 49 7.9428 7.9432 2.2847 2.3140 26-49 US (1 1 ) 50 7.9436 7.9432 3.9982 2.3140 27-50 US(1 1 ) (/L,,/L~) = (7.9417,7.9422); ( u 1 . u 2 ) = (4.6687 x 10-4,4.2495 x Note: Observation samples highlighted in bold (40th-50th) represent out-of-control process. I. Masood, A. HassanlExpert Systems with Applications 41 (2014) 7579-7595 Table 9 . - Outputs of t h e scheme for loading error case. - - W i n d o w range (RW) 1-24 2-25 3-26 4-27 5-28 6-29 7-30 8-31 9-32 P 0.6896 0.6910 0.8333 0.7723 0.7733 0.7822 0.7733 0.7234 0.7407 Decision based o n MEWMA control c h a r t N N N N N N N N N RW 10-33 11-34 12-35 13-36 14-37 15-38 16-39 17-40 18-41 P 0.7924 0.7753 0.7202 0.6941 0.7075 0.7254 0.6693 0.6973 0.7040 N N N N N N N N 0.8220 0.4510 US (1 0) 0.5067 0.6528 US (01) 0.1 665 0.1370 US (1 I ) 0.9479 1.1855 DS ( 1 0) 0.0985 0.0719 DS (01) 0.1533 0.1453 DS (1 1) 0.1257 0.1129 RW 19-42 20-43 21-44 22-45 23-46 24-47 25-48 26-49 27-50 P 0.7546 0.7504 0.7561 0.781 5 0.7806 0.7486 0.7258 0.7228 0.6886 N 0.1775 0.1609 0.1053 0.0434 0.0403 0.0376 0.0313 0.0318 0.0200 US (1 0) 0.7116 0.4926 0.5886 0.6606 0.4464 0.5521 0.3923 0.2824 0.3099 US (01) 0.1184 0.1124 0.1317 0.1724 0.1631 0.1540 0.1762 0.1711 0.2213 US (1 1 ) 1.4012 1.6147 1.5717 1.5235 1.6681 1.5983 1.7006 1.7666 1.7163 DS (1 0) 0.0632 0.0852 0.0817 0.0766 0.0802 0.0970 0.1002 0.1142 0.1221 DS (0 1) 0.1537 0.1508 0.1350 0.1830 0.2023 0.1735 0.1754 0.2060 0.2468 DS (1 1) 0.1304 0.1144 0.0875 0.0636 0.0597 0.0384 0.0434 0.0493 0.0283 Note: Bold value represents t h e m a x i m u m o u t p u t of ANN t h a t determines p a t t e r n category. G . Conclusions This paper proposed two-stage monitoring approach in moni- toring and diagnosis of bivariate process variation in mean shifts. Based on the frameworlc of 2s-IMS that integrates the powerful of MEWMA control chart and Synergistic-ANN recognizer, it has resulted in a smaller false alarm ( A R b = 335.01-543.93), rapid shifts detection (ARL1 = 3.18-16.75), and accurate diagnosis capa- bility (RA = 89.5-98.5%) compared to the traditional SPC charting schemes for BQC. Since the monitoring and diagnosis performances were evaluated using modeling data, real industrial data were used for the purpose of validation. The case studies involved tool bluntness and loading error in machining operations, whereby the proposed scheme has shown a n effective monitoring capability in identifying the bivariate in-control process without any false alarm. The scheme also effective in diagnosis aspect, that is, in cor- rectly identifying the sources of mean shifts when process becomes out-of-control. Based on the promising results, the 2s-IMS could be a reference in realizing balanced monitoring and accurate diagnosis of bivariate process variation. In the future work, further investigation will be extended to other causable patterns such as trends and cyclic. Aclcnowledgements The authors would like to thank Universiti Tun Hussein Onn Malaysia (UTHM). Universiti Teltnologi Malaysia (UTM), and Ministry of Higher Education (MOHE) of Malaysia who sponsoring this work. References Al-Assdf, Y. (2004). Recognition of control c h a r t p a t t e r n s using ~nulti-resolution wavelets analysis and ncul-dl networlts. Conlputers. a r ~ d I~~dustrialEngir~ecr~~~g, 47, 17-29. AIL, F. B. (1985). Multivariate quality contl-ol. In N. L. Johnson 8 S. IKotz (Etls.). Encvclopedia of.stotistica1 mcierices (vol. 6). New Yol-It: Wiley. Assaleh. I<.. 4 AI-Assal, Y. (2005). Features extraction a n d analysls for cldss~fying cdusable patterns in c o n t ~ - o l rhal-ts, CoinputeIs aitd Industrial Engirzrering, 49, 168..181. Bersimis, S., Psaraliis, 5 . . 8 I'anaretos, 1. (2007). Multivariate ~ t a t i s l ~ c a l pi-ocess control c h a m : An ovel-view. @ruliLy and Rellabilrty Engineeririg I ~ ~ r e r r ~ a t i t r n a i , 23, 5 17-543. BI-iinzell, I-]., 8 Eriltssnn. J. (2000). Feature recluction for cla.ss~fication of r n u l t i d ~ m e n s ~ o n a l data. Porter-ri Recognition, 33(121, 1741 - - 1748. Clien, 2.. Lu. S., 8 Lam, S. (2007). A hybrid system Ibr SI'C concurrenl patlel-11 recognilion. /ldvclnced Enginc.eriilg Inloi-~iialics. 21, 303-3 10. Clien. L. H.. & Wang, T. Y. (2004). AI-Lificial neural networlts lo classify mcan shifts from niultivariate %%hart s ~ g n a l s . Computers olld 111dush.iol Engineerillg, 47, 195-~205. Clieng, C. S., & Cheng, 1-1. 1'. (2008). I d e n t i f y ~ n g the source of variance shifts in the rnultivai-iafe pl-occss using neural netwol-Its and s u p p o r l vector machines. Expert 5,y.stern.s wit11 Applico1ror1.s. 3.5, 198-206. Chih. W. H.. 4 Kollier, D. A. (1994). U ~ a g n o s i s charactcristics Sor hivariale p a t t e r n recogliltion s c h c m e in SPC. Iruernalionul Jouri~o! o j r2uolity u r ~ d Reliability Molitr.yrnelit, l I ( 1 ) . 53-66. Chih. W. I<., 8 Rollier. I). A. (1995). A ~ ~ i e t h o d o i o g y of pattern recogmtion s c h e m e s for t w o variables in SPC. Il~terr~ationol jou1-11a1 of Q~ialin, and I(eliability Mnnngenrent, 12(3), 86-107. Crosiel-, R. B. (1988). M u l t ~ v a r ~ a t e g e n e ~ a l ~ z a t i o n s of c u ~ n u l a t i v e s u m q w l ~ t y c.ontrol schemes. 7'ech11ometrics. 30(3), 291-303. El-Midany. I . T., El-Baz, M. A,. & Abtl-Elwahecl. M. S. (2010). A proposccl Irarneworlt for conlrol c h a r t p a t t c r n recognition in multivariale process using &I-tificial neural networlts. Expert Systems with Aljplicutio~is, 37, 1035- 1042. Fuchs. C.. 4 Benjamini, Y. (1994). M u l t ~ v a r i a t e profile chart? for s t a t ~ s t i c a l ~ l r o c e s s control. Technom~trics, 36(2), 182- 195. Gauri, S. I<., 8 Chaltraborty, 5 . (2006). Feature-l~ased recogiiltlon of control c h a r t patterns. Conlputen und lndlrstriul Erigir~eerin,y, 51. 726-742. Gauri. S. I<.. 4 Chakraborly, S. (2008). Improved recognition of control c h a r t p a l l c r n s using artificial neural netwol-Its. I ~ ~ t e n ~ n t i o r ~ c i l J o u n ~ u l ofAdvnnted Manlrfuucturirlg Tei:hnolom,, 36, 1191--1201. Guh, R. S. (2007). On-line identification and q u a n t i l i c a t ~ o n of mean shifts In b i v a r ~ a t e processes u s ~ n g a NN-bared approach. Quality and Reliabilrtj~ Engineering I ~ ~ t e r n o t i o n a l . 23, 367 -- 385. Cull, R. S . (2010). Simultaneous process mean and variance monitol-ing using artiiicial neul-al nctworlts. Cornptrters rind Ir~dlistrifll Er~girirering, 58, 739-753. Guh. K, S., 8 Shiue, Y. K. (2005). On-line idenlification of contl-ul c h a r t p a t t e r n s using selt:o~-ganizing approaches, lnternatioiiu! lour1101 of Production Kesenrch, 43(6). 1225--- 1254. H a c h ~ c h a , W., 4 Ghorbel. A. (2012). A survey of contl-ol-chart pattern-~'ecognitio~i I ~ t e r a t u r e (1991-2010) based oli a new conceptual clasrificat~on s c h e ~ n e . C(~nlputers o r ~ d Incll~striol Eli~@l~ee~-in,y, 63, 204-222. Hassan, A., Nabi Baksli, M. S., S h a h a r o u n , M. A,, K. Jamaludin, H. (2001). Impl-ovecl SPC c h a r t paltern recognition using statislical features. 111Li.i-norionoljouii~al o j Producrion Rescorch, 41(7). 1587-1 603. 1.lotelling. 1-3. H. (1947). Multival-late quality contl'ol. In C. Eisenhdrt, M. W . Hastay, 8 W. A. W a l l ~ s (Eds.), Teclli~iques ofstub.sticul ailnlysi.~. New York: McGraw-I-1111. Jacltson. J. E. (1391). A user's guide to principle compollerrt. NL'W Jel-spy: Wiley. IKlosgen, W.. 8 Zytltow, J. M. (2002). HaildbooI< of dotti minii~g arid knowledge discovery. Lontlon: OxFord University I'ress. I<ou~-ti. T., & MacGregol-, 1. F. (1996). Multivariate SI'C methods for process a n d product monicoring.Joun~nl oJQuolity Technoloa~, 28(4), 409-428. Lehrnan. K. S. (1977). Corilputrr .simulation arid rnodelii~g: An introd~rcrion. London: ILawrence Erlbal~m. ILowry. C. A,, & Montgomery, il. C. (1995). A revlew of multivariate c o ~ i t r o l clla~.ts. IIE T ~ ~ a i ~ s a c t i o n s . 27(6), 800---810. Lowl-y, C. A,, Woodall, W. H.. Champ, C. W.. & Rigtlon, S. E, (1992). MultivariaLe exponentially weiglitetl m o v l n g average control chart. I~rcl~ntr~nr~i-ii.5, 34(1), 46-53. I. Masood, A. HassanIExpert Systems with Applications 4 1 (2014) 7579-7595 Lucas, J. M., 8 Saccucci, M. S . (1990). Exponentially w e ~ g h t e d niovlng average control s c h e m e s : Properties a n d e n h a n c e m e n t s . Tr.cl~noineC~~cs, :2(1). 1--12. Mason, R. I..,Tracy, N. I)., & Yoimg, J . C. ( 1 9 9 5 ) . Decomposition o f T 2 for multivartate control c h a r t ~nterpl-etation.]ournul ofQ11uli1y Techi~olo,gv, 2 7 ( 2 ) , 109-1 19. Mason. K. L., Tracy. N. D.. 8 Y0ung.j. C. (1997). A practical a p p r o a c h fol- interpreting multivariate 'I" conii-ol c h a r t signals. jc~l~i-nol of' Q u a l ~ l y Techi~ology, 2 9 ( 4 ) , 396-406. Masood, I.. W Hassan, A. ( 2 0 1 0 ) . Issues i n d e v c l o p m c n t of artiliclal ncurdl network- b a s e d control c h a r t p a t t e r n recognition schemes. Eul.oj)enn Joul-nui of Scientifii: Reseu!-ch, 3 9 ( 3 ) , 336--355. Masood, I., 8 tlassan, A. ( 2 0 1 3 ) . I'attel-n I - e c o g n ~ t i o n for h ~ v a r i a t e process Iiiean shifts sing fearrure-based artificial neural nelwol-k. Inlei-~lutional Journul of'Advunced M n n u f n c t ~ ~ r i n g Technology, G6(9-12). 1201 -12 18. Montgomery, D. C. (2009). I ~ ~ t r o d u c l i o n Lo slolislicul qunlity conlral ( 6 t h etl.). New i e r s e y : John Wiley 8 S o n s lnc. Nelson, L. S. (19SS). I n t e r p r e t i n g Shewhal-t X-bar control chart. j o t l n ~ n l of Qgollty Jeclunology, 17(2). 1 1 4 - - - I 1 6 . N~alti, S. T. A., & Ahbasi, 8 . ( 2 0 0 5 ) . Fault d ~ a g ~ i o s i s In multivariate control charts u s i n g art~iicial neural networlts. Qt~olily U I I ~ Reliability E11,qi11eeri17gl11ter11(1tioiin/. 21. 825-540. P h a m . D. T., 81 W a n i , M. A. ( 1 9 9 7 ) . Featurc-based control c h a r t p a t t e r n recognilion. I n t e l - n a ~ i o n a l ~ o t r r n u l ofProdlictio11 Re.sccrrc11, 3517). 1875-1 890. l'ignatiello. J. J., 6. Kunger, 6 . C. ( 1 9 9 0 ) . Coniparisons of m ~ l l r i v a ~ - i a t e CUSllM ciiarts. joui-rial of Quuliry Jecir11o1og.y~ 2 2 ( 3 ) , 1 7 3 - 186. I'I-al~hu, 5. S., & Rungel-. C;. C. ( 1 9 9 7 ) . Designing a r n t ~ l t i v a r i a t e EWMA control chart. ]oul-no1 of Qrriity J e c h ~ ~ o l o & y , 29(1), S...lS . Salehi, M., Balireininejad, A,, W Nalthai. I. (201 I ) . On-line analysis of out-of-control signals ill ~ i i u i t i v a r i a t e m a n u f a c t u r i n g processes using a hybrid l e a r n i ~ i g - b a s e d model. Nf2tIr0~~m/~lltiIIg, 74(12---13). 2083-2095. Salehi, M.. I(azemzatleli, K. B.. X. Salmasnia, A. ( 2 0 1 2 ) . On-linc tlctection o l m e a n a n d varlallce shirt using iieural nelworlts alitl s u p p o r t vectol- mdcliine in multivariate processes. Applied Soft i b m p ~ r l i ~ ~ g , 12, 2973-2994. Sepulveda, A,, & Nachlds, J. A. ( 1 9 9 7 ) . A s i m ~ ~ l a t i o n approach to m u l l i v a ~ iate quality control. Cornp~rlen ortd I n d ~ l r t ~ - i a l Eligiiieei-lng, 33(1-2), 113-1 16. S h e w h a r t , W . A. ( 1 9 3 1 ) . The ecoiiomic control ofqunliry i n a ~ u ~ ~ o c r u r r d P I - o d ~ l r t s . N e w York: Van Nostrand. W a n g , T. Y.. & C h e ~ i , I.. II. ( 2 0 0 1 ) . Mean s h ~ f t s detectloll and classificatio~l in multivariate pl-ocess: A neul-di-fuzzy appl-oacli. ]ourno1 01 li~telIlge~?cc Muiiufuc~irri~~g, 33, 2 11 -22 1. W a n g , C. H.. I ( i ~ o , W., 8 Qi, H. ( 2 0 0 7 ) . An integrated approach fol-process r n o n i t o r i ~ i g using w a v e l e t analysis a~lcl c o m p e t i l ~ v c n e u r a l network. I n t ~ i ~ ~ i a t i o ~ t o l J o u r ~ i a i of I'rod~ictio!~ Re.senrcl~. 45. 2 2 7 244. Yu, J. N.. W Xi. L. F. t 2 0 0 9 ) . A n e r ~ r a l nehuork e n s e m l ~ l e - b a ~ e d m o d e l for on-ltne ~ i i o n i t o r l n g a n d d i a ~ n o y l r of out-of-control signals in multivariate manuCactu~-ing processes. Expert Syslems wit11 Applicu~roi~.s, 36. 909-921. Yu. 1. B., XI, L. F., 8 Zliou, X . J . (2009). IcientifLing s o u r c e ( s ) of out-of-con11-ol signals in mullivariate m a n u l a c l u r i n g processes using selective neural netwoi-lc ensemble. Engineer~ng Ap]~ficnlions o j ArtiJcinl intelligence, 22, 141 -1 52. Lorriassatinr, F., Tannock, J. D. T., & O'Rl-leu. C. ( 2 0 0 3 ) . llsinx novelty detection t o i r l e n t ~ f y ahnol-malities caused by meall shifts in bivariate processes. C o ~ n p ~ ~ t e r . ; n i ~ d h~dush-iol En.y~~~errin,y, 4 4 , 3 8 5 408. 
work_3vq4z57v65enhnkvtjrsqjz3ta	----	Data mining models for student careers Data mining models for student careers Renza Campagni, Donatella Merlini ⇑, Renzo Sprugnoli, Maria Cecilia Verri Dipartimento di Statistica, Informatica, Applicazioni Università degli Studi di Firenze, Viale Morgagni 65, 50134 Firenze, Italy a r t i c l e i n f o Article history: Available online 6 March 2015 Keywords: Data mining Educational data mining Student careers Clustering Frequent pattern analysis a b s t r a c t This paper presents a data mining methodology to analyze the careers of University graduated students. We present different approaches based on clustering and sequential patterns techniques in order to iden- tify strategies for improving the performance of students and the scheduling of exams. We introduce an ideal career as the career of an ideal student which has taken each examination just after the end of the corresponding course, without delays. We then compare the career of a generic student with the ideal one by using the different techniques just introduced. Finally, we apply the methodology to a real case study and interpret the results which underline that the more students follow the order given by the ideal career the more they get good performance in terms of graduation time and final grade. � 2015 Elsevier Ltd. All rights reserved. 1. Introduction The large supply of data stored in the computer systems of sev- eral companies, both public and private, has given a push in the direction of the development of new technologies for data manage- ment and analysis. Data mining techniques originate in this con- text, with the aim of discovering hidden and non-trivial relationships among information of various nature. This collection of techniques, used in different sectors, including the educational environment, comes from the traditional methods of data analysis and have the characteristic of being able to treat large amounts of data. In the field of education, educational data mining is a recent research area that explores and analyzes the information stored in student databases in order to understand and improve the per- formance of the student learning process. Data are analyzed by using statistical, machine learning and data mining algorithms, with the aim of resolving problems of educational research and improve the entire educational process. Recently there has been an increase in the use of educational software instruments and of databases containing students information, so we have large repositories of data reflecting how students learn. In addition, the use of Internet in education has created the context of e-learning or web-based education which continuously generates large amounts of data concerning the interactions between teaching and learning. Educational data mining tries to use all this information to better understand learners and learning, and to develop methodologies which, integrating the data with the the- ory, allow to improve the educational process. Educational data mining is a growing research area that involves researchers all over the world from different and related research areas and since 2008 an annual International Conference on Educational Data Mining has been established (http://www.educationaldatamining.org). Great efforts have been made in the direction of describing the state of the art of this research area and, in the recent years, several survey papers have been published on the subject (Baker, 2010, 2014; Baker & Yacef, 2009; Luan, 2002; Peña-Ayala, 2014; Romero & Ventura, 2010, 2013). As already observed, usually data mining techniques are applied to large data sets. In the context of education, however, we are often faced with data sets corresponding to small groups of stu- dents following the same curriculum. Referring to a university con- text, for example, even when a degree program is frequented by many students the data of interest correspond to relatively small data sets. The recent paper (Natek & Zwilling, 2014) focuses on the study of data mining techniques applied to small data sets con- cerning higher education institutions and concludes that the use of these techniques in real-life situations is useful and promising and can provide administrators with precious tools for decision. Over the years, several data mining models have been designed and implemented to analyze the performance of students. For example, in Delavari, Shirazi, and Beikzadeh (2004) and Delavari, Somnuk, and Beikzadeh (2008), a model is proposed which pre- sents the advantages of data mining technology in higher educa- tional systems; the authors give a sort of road map to assist the institutions to identify the ways to improve their processes. In Daimi and Miller (2009), the authors illustrate a classification http://dx.doi.org/10.1016/j.eswa.2015.02.052 0957-4174/� 2015 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. E-mail addresses: renza.campagni@unifi.it (R. Campagni), donatella.merlini@ unifi.it (D. Merlini), renzo.sprugnoli@unifi.it (R. Sprugnoli), mariacecilia.verri@unifi. it (M.C. Verri). Expert Systems with Applications 42 (2015) 5508–5521 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.02.052&domain=pdf http://www.educationaldatamining.org http://dx.doi.org/10.1016/j.eswa.2015.02.052 mailto:renza.campagni@unifi.it mailto:donatella.merlini@ unifi.it mailto:donatella.merlini@ unifi.it mailto:renzo.sprugnoli@unifi.it mailto:mariacecilia.verri@unifi.it mailto:mariacecilia.verri@unifi.it http://dx.doi.org/10.1016/j.eswa.2015.02.052 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa model to investigate the profile of students which most likely leave university without ending their career. In particular, they use some classification algorithms implemented in the WEKA system (Witten, Frank, & Hall, 2011). Recommendations of suitable courses for stu- dents are analyzed with different approaches in Bydzovska and Popelínský (2014) with the aim of predicting student success. In Damaševičius (2010) a framework is proposed for mining educa- tional data using association rules. More recently, Romero, Zafra, Luna, and Ventura (2013) proposes the application of association rule mining to improve quizzes and courses and (Saarela et al., 2014) applies frequent itemset mining and association rule learn- ing to students previously grouped by clustering techniques. In Guruler, Istanbullu, and Karahasan (2010), in order to explore the factors having impact on the success of university students, a system based on the decision tree classification technique is pre- sented. Clustering is used in Campagni, Merlini, and Verri (2014) for analyzing data concerning the evaluation of courses taken by students, linked to their results in the corresponding exams. The work presented in Dutt, Aghabozrgi, Ismail, and Mahroeian (2015) reviews different clustering algorithms applied to educa- tional data mining context while (Peña-Ayala, 2014) is an interest- ing review of recent educational data mining development whose contents are in turn analyzed by a data mining approach. As already observed, data mining techniques have also been applied in computer-based, e-learning and web-based educational systems (Bouchet, Harley, & Trevors, 2013; Bogarín, Romero, Cerezo, & Sánchez-Santillán, 2014; Castro, Vellido, Nebot, & Mugica, 2007; Hämäläinen, Laine, & Sutinen, 2006; Koedinger, Cunningham, Skogsholm, & Leber, 2008; Mostow & Beck, 2006; Merceron & Yacef, 2005; Romero, Romero, Luna, & Ventura, 2010; Romero, Ventura, & García, 2008; Romero, López, Luna, & Ventura, 2013). The existing literature about the use of data mining in educational systems is mainly concerned with techniques such as clustering, classification and association rules (Damaševičius, 2010; Tan, Steinbach, & Kumar, 2006; Witten et al., 2011; Wu & Kumar, 2009). An academic curriculum usually defines a specific learning pro- gram which puts some types of restrictions on how the students are required to take courses. These constraints typically describe a set of courses and a set of relationships between them. In the cur- rent practice, however, students have many degree of freedom, therefore helping students to choose courses, discovering patterns and key courses, planning future courses and refining curricula based on the feedback of students are important educational tasks, as recently pointed out in Aher and Lobo (2013), Kardan, Sadeghi, Ghidary, and Sani (2013), Méndez, Ochoa, and Chiluiza (2014) and Pechenizkiy, Trcka, Bra, and Toledo (2012). The present work fits into this context extending and unifying the results presented in Campagni, Merlini, and Sprugnoli (2012a, 2012b, 2012c). In par- ticular, we introduce the concept of ideal career, that is, the career of a graduated student who takes every examination just after the end of the corresponding course, without delay, and propose a data mining methodology, based on clustering and sequential pattern analysis, to study the student behavior by comparing student careers with the ideal one. Sequential pattern analysis has been used in the context of educational data mining mainly in com- puter-based environments. For example, Soundranayagam and Yacef (2010) explores the order in which students access e-learn- ing resources as they solve set assessment tasks, such as tests, assignments and exams and the links with students learning. A method to automatically detect collaborative patterns of student and tutor dialogue moves is illustrated in D’Mello, Olney, and Person (2010). Paper (Martinez, Yacef, Kay, Al-Qaraghuli, & Kharrufa, 2011) mines and clusters frequent patterns to compare distinct behaviors between low and high achievement groups around an interactive tabletop. A data mining methodology for identifying and comparing learning behaviors from students learn- ing interaction traces is presented in Kinnebrew, Loretz, and Biswas (2013); in particular, the paper proposes an algorithm that employs a novel combination of sequence mining techniques to identify differentially frequent patterns between groups of stu- dents. Paper (Guerra, Sahebi, Brusilovsky, & Lin, 2014) models and examines patterns of student behavior with parameterized exercise. A recent research which proceeds in a direction similar to ours is illustrated in Asif, Merceron, and Pathan (2014), where the progression of a student is analyzed by defining a tuple that shows how the results of a year stay the same, increase or decrease compared to first year. The preprocessing phase is the first step in any data mining pro- cess and allows us to transform the available data into a format suitable for the analysis. The importance of this task has been recently highlighted in Romero, Romero, and Ventura (2014). In Section 2, we illustrate the preprocessing phase necessary to orga- nize data for our analysis. A crucial aspect during this phase is the insertion in the database of the reference to the semester in which a course has been given by a teacher and the semester in which the student has taken the corresponding exam. This information allows us to define the ideal career together with the career of each graduate student. We represent a career as a trajectory of points in the plane. In particular, the ideal career is defined by a sequence of points sI ¼ðð0; e0Þ;ð1; e1Þ;ð2; e2Þ; � � � ;ðn; enÞ;ðn þ 1; enþ1ÞÞ, where ei is an exam identifier and i its position in the career. The position i ¼ 0 denotes the starting point of the career while i ¼ n þ 1 corresponds to the final examination given last by all students. This particular career, without loss of generality, can be repre- sented by the bisecting line of the first quadrant (green lines in Fig. 1). The career of a generic student J is then represented by a broken line, corresponding to the sequence of points tJ ¼ðð0; eJ 0Þ;ð1; eJ 1Þ;ð2; eJ 2Þ; � � � ;ðn; eJ nÞ;ðn þ 1; eJ nþ1ÞÞ, where eJ i is the identifier of the exam given by student J at time i (red lines in Fig. 1). We then compute the distance between a generic career tJ and sI in different ways, by using for example the Bubblesort distance, defined as the number of inversions in the permutation relative to tJ , or the area between the lines tJ and sI . Finally, we insert these values in the database. In Section 3, we analyse the preprocessed data with clustering and sequential pattern techniques. For what concerns the cluster analysis, the idea is to explore the database with the aim of understanding if there exists a relation between the distance from the ideal career and the success of stu- dents. This kind of analysis, accompanied by cluster validation, can highlight different groups of students characterized by similar dis- tances and behaviors and can give some suggestions to improve the organization of the laurea degree or to recommend precedence relations among courses. Sequential pattern analysis aims to find relationships between occurrences of sequential events, that is, to find if any specific order of the occurrences exists. In this paper we consider as events the exams taken by a student; the temporal information is the semester in which the exam has been taken or the delay with which it has been taken. We study an organization of the univer- sity which allows students to take an exam in different sessions after the end of the course, as in Italy. The temporal information allows us to see the career of a student as a sequence hs1s2 . . . smi where each element sj is a collection of one or more exams taken in the same semester or having the same delay. If we use the seme- ster as temporal information, m indicates the number of semesters in which a student takes exams; if we use the delay, m indicates the maximum number of delays, in semesters, with which a stu- dent takes one or more exams. By analyzing the sequential pat- terns, we can explain some behaviors which may seem R. Campagni et al. / Expert Systems with Applications 42 (2015) 5508–5521 5509 https://isiarticles.com/article/46055 
work_3vtrulz3czgv7pbhfl7yw5lxem	----	eswa.eps Expert Systems with Applications 36 (2009) 9719–9728 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Improving the interest operator for face recognition Yong Xu a,*, Lu Yao a, David Zhang b, Jing-Yu Yang c a Bio-Computing Research Center, Harbin Institute of Technology, Shenzhen Graduate School, Shenzhen, Guang dong 518055, China b Biometrics Research Center, The Hong Kong Polytechnic University, China c School of Computer Science and Engineering, Nanjing University of Science and Technology, China a r t i c l e i n f o Keywords: Face recognition Feature extraction Interest operator 0957-4174/$ - see front matter � 2009 Elsevier Ltd. A doi:10.1016/j.eswa.2009.02.032 * Corresponding author. Tel.: +86 752 26032458; fa E-mail address: laterfall2@yahoo.com.cn (Y. Xu). a b s t r a c t When the conventional interest operator is used as the feature extraction procedure of face recognition, it has the following two shortcomings: first, though the purpose of the conventional interest operator is to use the intensity variation between neighboring pixels to represent the image, it cannot obtain all vari- ation information between neighboring pixels. Second, under varying lighting conditions two images of the same face usually have different feature extraction results even though the face itself does not have obvious change. In this paper, we propose two new interest operators for face recognition, which are used to calculate the pixel intensity variation information of overlapping blocks produced from the original face image. The following two factors allow the new operators to perform better than the conventional interest operator: the first factor is that by taking the relative rather than absolute variation of the pixel intensity as the feature of an image block, the new operators can obtain robust block features. The second factor is that the scheme to partition an image into overlapping rather than non-overlapping blocks allows the proposed operators to produce more representation information for the face image. Experi- mental results show that the proposed operators offer significant accuracy improvement over the con- ventional interest operator. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction Since the interest operator was first proposed, it has been wildly used to represent image or target feature. For example, Moravec (1981) used an interest operator to compute the intensity vari- ances in the horizontal, vertical and both diagonal directions for each pixel point in an image and selected the minimum of these values as the variance of that point. Nasrabadi and Choo (1994) used the interest operator in the stereo vision correspondence. In Nasrabadi and Choo (1994), the interest operator was applied to all image blocks and the point having the local maximal variance in each local block was selected as the so-called interesting point of that block. Căleanu, Huang, Gui, Tiponu, and Maranescu (2007), Căleanu (2000), Căleanu (2001), Zhao, Huang, and Sun (2004) and Zhao (2007) used the interest operator to extract fea- tures from face images. The interest operator was also exploited for target recognition (Haber & Modersitzki, 2004). Generally speaking, the interest operator for recognition can be viewed as a feature extraction algorithm that calculates variation information in different directions of the image pixel intensity. Note that when the interest operator is applied to face recognition, the first step is to divide each image into the same number of blocks with a fixed ll rights reserved. x: +86 755 26032461. size. Then feature extraction is performed for each block by using the interest operator. Finally, one can treat the feature extraction results of all the blocks of a face image as the representation of this image and can classify face images using the representation and a classifier. It should be pointed out that complex imaging conditions such as varying pose, facial expression and lighting conditions may re- strict the ability of the interest operator. Actually, complex imaging conditions degrade performances of most of face recognition tech- niques (Adini, Moses, & Ullman, 1997). In addition, the human face is a deformable object and a face can produce different deforma- tions at different times, resulting in various facial expressions. Con- sequently, there may be much difference between a block of one image of a face and the same block of another image of the same face, if the two images are associated with different poses or facial expressions. Face recognition using the conventional interest oper- ator, therefore, has the following shortcoming: the feature extrac- tion results of different images of the same face might have much difference. Another shortcoming of the conventional interest oper- ator stems from the algorithm itself. Under varying lighting condi- tions the algorithm usually also produces different feature extraction results for the same face, even if the face itself does not have any change in pose and expression. This is because vary- ing lighting conditions produce different intensity values for the image pixel. For example, high intensity lighting usually produces mailto:laterfall2@yahoo.com.cn http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 9720 Y. Xu et al. / Expert Systems with Applications 36 (2009) 9719–9728 high intensity value for the pixel and consequently the variation between two neighboring pixels usually also appears to be high. On the other hand, low intensity lighting usually produces low intensity value for the pixel and consequently the variation be- tween two neighboring pixels is also low. As a result, when we compare the feature extraction results of two images that are ob- tained under the high and low intensity lighting conditions, a low similarity may be produced. The third shortcoming of the con- ventional interest operator is that it cannot obtain all variations be- tween neighboring pixels. Indeed, since the conventional interest operator is applied to only pixels within an image block, the con- ventional interest operator is not able to compute the variations between two neighboring pixels that are located in two blocks, respectively. This is illustrated by Fig. 1. In this paper, in order to overcome the shortcomings of the con- ventional interest operator, we propose to exploit new operators and the overlapping block partition scheme for face recognition. The rationale of the proposed approach is as follows: first, because the improved interest operator takes the relative rather than abso- lute variation of image pixel intensity as the features of the face image, this approach can produce more stable face features for the same face than the conventional interest operator, especially under varying lighting conditions. This is helpful for obtaining a satisfactory classification accuracy. Second, since in our approach different blocks overlap one another, our approach allows intensity variations between all pixels to be provided whereas the conven- tional interest operator cannot do so. Experimental results illus- trate that our approach can produce a higher accuracy in comparison with the conventional interest operator. Experiments also show that a linear feature extraction approach can be com- bined with the proposed interest operators to obtain further per- formance improvement. 2. The conventional interest operator As mentioned above, the conventional interest operator evalu- ates the variation of image pixels in the horizontal, vertical and both diagonal directions for each block of an image. The conven- tional interest operator can be described as follows: it calculates the mean l and the center variance r2 of a block using (1) and (2), respectively. It calculates r20; r 2 45; r 2 90 and r 2 135, which respec- tively stand for the intensity variations of block pixels in horizon- tal, diagonal 45�, vertical and diagonal 135� directions, using (3)– (6), respectively. Fig. 1. Illustration of the non-overlapping block partition scheme used for the conventional interest operator. This figure shows that an image is divided into nine non-overlapping blocks each consisting of a number of pixels. For a block, we denote the boundary pixels that adjoin another block by ‘�’ and we denote the other pixels by ‘‘�”. Note that for two ‘�’ pixels respectively located in two blocks, even if they are neighbors, the pixel intensity variation of the two pixels cannot be obtained by the conventional interest operator. u ¼ 1 P � Q XP x¼1 XQ y¼1 pðx; yÞ ð1Þ r2 ¼ 1 P � Q XP x¼1 XQ y¼1 ½pðx; yÞ� u�2 ð2Þ r20 ¼ 1 P � Q XP�1 x¼1 XQ y¼1 ½pðx þ 1; yÞ� pðx; yÞ�2 ð3Þ r245 ¼ 1 P � Q XP�1 x¼1 XQ�1 y¼1 ½pðx þ 1; yÞ� pðx; y þ 1Þ�2 ð4Þ r290 ¼ 1 P � Q XP x¼1 XQ�1 y¼1 ½pðx; y þ 1Þ� pðx; yÞ�2 ð5Þ r2135 ¼ 1 P � Q XP�1 x¼1 XQ�1 y¼1 ½pðx þ 1; y þ 1Þ� pðx; yÞ�2 ð6Þ Hereafter we suppose that the size of each block is P � Q and p(x, y)(1 6 x 6 P, 1 6 y 6 Q) represents the pixel intensity of the point (x, y) in a block. Fig. 2 shows the feature extraction results of a face image obtained using the conventional interest operator. From Fig. 2, we can clearly see that if two original images of the same face are obtained under varying conditions, their resultant images may have obvious difference. 3. Description of our approach Our approach consists of two components, an overlapping block partition scheme and an improved interest operator. The goal of our approach is to improve image presentation ability of the con- ventional interest operator. The basic idea for improving the inter- est operator is to take the relative variation of the gray intensity as features of the image block. Based on this idea, we develop two new interest operators as shown in the following. 3.1. Improved interest operator 1 Improved interest operator 1 is defined as follows: r2 ¼ 1 P � Q XP x¼1 XQ y¼1 ½pðx; yÞ� u�2 ðu þ c1Þ 2 ð7Þ r20 ¼ 1 P � Q XP�1 x¼1 XQ y¼1 ½pðx þ 1; yÞ� pðx; yÞ�2 ðu þ c1Þ 2 ð8Þ r245 ¼ 1 P � Q XP�1 x¼1 XQ�1 y¼1 ½pðx þ 1; yÞ� pðx; y þ 1Þ�2 ðu þ c1Þ 2 ð9Þ r290 ¼ 1 P � Q XP x¼1 XQ�1 y¼1 ½pðx; y þ 1Þ� pðx; yÞ�2 ðu þ c1Þ 2 ð10Þ r2135 ¼ 1 P � Q XP�1 x¼1 XQ�1 y¼1 ½pðx þ 1; y þ 1Þ� pðx; yÞ�2 ðu þ c1Þ 2 ð11Þ where u is also defined by (1) and c1 is a positive constant. Indeed, the new operator as defined in (7)–(11) produces the relative rather than absolute variation of the pixel intensity by dividing the result of the conventional interest operator by a quantity associated with the mean of the pixel intensity. The difference between improved interest operator 1 and the conventional interest operator is as fol- lows. As presented above, varying lighting condition is one typical factor that makes two corresponding pixels from two images of the same face have quite different intensities. As a result, under varying lighting condition, two corresponding image blocks from two images of the same face usually appear to have different inten- sity variations. The conventional interest operator, therefore, usu- Fig. 2. The original face and resultant images obtained using the conventional interest operator. Each row shows the face and resultant images. For example, (b), (c), (d), (e) and (f) respectively stand for the resultant images on r2; r20; r 2 45; r 2 90; r 2 135 of the face image shown in (a). The original face image is first divided into a number of non- overlapping 2 by 2 blocks and then the conventional interest operator is implemented for each image block. Y. Xu et al. / Expert Systems with Applications 36 (2009) 9719–9728 9721 ally does not perform well in producing stable features for the same face under varying lighting condition. However, by using the rela- tive variation of the pixel intensity as shown in (7)–(11) as the fea- ture of the face image, improved interest operator 1 is able to produce more stable features for the same face under varying light- ing condition. The use of constant c1 can prevent the operator from being unfeasible in the case where the mean of the pixel intensity is zero. 3.2. Improved interest operator 2 Improved interest operator 2 is designed as follows: r2 ¼ 1 P � Q XP x¼1 XQ y¼1 jpðx; yÞ� uj u þ c2 ð12Þ r20 ¼ 1 P � Q XP�1 x¼1 XQ y¼1 jpðx þ 1; yÞ� pðx; yÞj u þ c2 ð13Þ r245 ¼ 1 P � Q XP�1 x¼1 XQ�1 y¼1 jpðx þ 1; yÞ� pðx; y þ 1Þj u þ c2 ð14Þ r290 ¼ 1 P � Q XP x¼1 XQ�1 y¼1 jpðx; y þ 1Þ� pðx; yÞj u þ c2 ð15Þ r2135 ¼ 1 P � Q XP�1 x¼1 XQ�1 y¼1 jpðx þ 1; y þ 1Þ� pðx; yÞj u þ c2 ð16Þ where u is still defined by (1) and c2 is a positive constant. Differing from improved interest operator 1, the second improvement opera- tor takes as the feature of an image block the result of the conven- tional interest operator divided by the sum of the mean of the pixel intensity and c2. The use of c2 enables the operator to be workable in the case where the mean of the pixel intensity is zero. Improved interest operator 2 also has the similar advantages to improved interest operator 1. 3.3. Overlapping block partition scheme As mentioned above, though the conventional interest operator was developed to calculate the pixel intensity variation, not all var- iation information between neighboring pixels could be obtained. Indeed, the conventional interest operator cannot calculate the variations between two neighboring pixels respectively located in two blocks as shown in Fig. 1. In order to overcome this draw- back of the conventional interest operator, we propose to partition an image into a number of overlapping blocks and to extract fea- tures from each block by using either of the two interest operators respectively presented in Sections 3.1 and 3.2. Note that in our partition scheme two horizontally neighboring blocks overlap one another and so do two vertically neighboring blocks. Fig. 3 shows that how an original 6 � 6 image is partitioned into four overlapping image blocks each having the size of 4 � 4. Here each rectangle unit represents a pixel. The overlapping region is represented by the shadow, which shows that half of each block is overlapped by a neighboring block. Our approach implements improved interest operator 1 (or improved interest operator 2) for each image block produced by the overlapping block partition scheme. Advantages of our approach can be described as follows. First, because our approach takes the relative rather than absolute vari- ation of the pixel intensity as the feature of an image block, the fea- ture extraction result appears to be less affected by the lighting 2 2 Fig. 3. Illustration of the scheme to partition an image into overlapping image blocks. The overlapping region is represented by the shadow. Each image block is of size of 4 � 4. Half of every block is overlapped by a neighboring block. For example, half of the 4 � 4 image block located in the left top is overlapped by its right neighbor image block. For the same block, its down neighbor 4 � 4 block also overlaps half of its size. 9722 Y. Xu et al. / Expert Systems with Applications 36 (2009) 9719–9728 condition than the conventional interest operator. This helps our approach obtain more robust features than the conventional inter- est operator. Second, our approach allows information of the inten- sity variation of all neighboring pixels to be obtained due to the overlapping partition scheme, whereas the conventional interest operator cannot do so. Fig. 4 shows an original face image and the feature extraction result of this image obtained using improved interest operator 1. The overlapping block partition scheme can be shown more clearly by the following example: if the original image is repre- sented by an 80 � 80 matrix and we partition the original image Fig. 4. The original face image and the resultant images obtained using improved interes example, (b), (c), (d), (e) and (f) respectively stand for the resultant images on r2; r20; r 2 4 divided into overlapping 4 by 4 blocks and then improved interest operator 1 is implem image block. into a number of 4 � 4 overlapping blocks and half of each block is overlapped by a neighboring block, we will obtain 1521 overlap- ping image blocks. When we calculate each of the r2; r20; r 2 45; r 2 90; r2135 values for all the 1521 image blocks of the original image, the result that we obtain is a 1521-dimensional vector. The vector can be also shown as a 39 � 39 matrix as illustrated in Fig. 4. Since each block has its own five values r2; r20; r 2 45; r 2 90; r 2 135, the feature extraction result of the original image can be regarded as five 1521-dimensional vectors or five 39 � 39 matrices. Concatenating these one-dimensional vectors, we can obtain a 7605-dimensional vector. More generally, if the original image is an m � n matrix and we partition the original image into a number of s � t overlapping blocks and half of each block is overlapped by a neighboring block, (2m/s � 1) � (2n/t � 1) image blocks will be generated. As a result, for this original image, the feature extraction result produced by our approach will be a 5(2m/s � 1) � (2n/t � 1)-dimensional vector. Similar to the features obtained using the conventional interest operator, the features obtained using our approach are also usually very high-dimensional. However, as shown in Section 5, linear fea- ture extraction procedures such as 2DPCA (two-dimensional PCA) or 2DFLD (two-dimensional Fisher discriminant analysis) (Cho, Chang, Kim, & Lee, 2006; Kongsontana & Rangsanser, 2005; Mutelo, Khor, Woo, & Dlay, 2006; Nhat & Lee, 2005; Sanguansat, Asdornw- ised, Jitapunkul, & Marukatat, 2006; Wang, Wang, Zhang, & Feng, 2005; Xu, Zhang, Yang, & Yang, 2008; Yang, Zhang, Frangi, & Yang, 2004; Yang et al., 2004) can be used to transform the feature extraction results of the interest operator into lower-dimensional features. t operator 1 with c1 = 10. Each row shows a face image and the resultant images. For 5; r 2 90; r 2 135 of the face image shown in (a). Note that the original face image is first ented for each image block. Half of an image block is overlapped by a neighboring 0 2 4 6 8 10 12 14 16 18 20 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 2DPCA IO+2DPCA(2×2) IO+2DPCA(2×4) (a) 0 2 4 6 8 10 12 14 16 18 20 0.4 0.5 0.6 0.7 0.8 0.9 1 2DFLD IO+2DFLD(2×2) IO+2DFLD(2×4) (b) 110 120 130 140 150 160 170 180 190 200 0.74 0.75 0.76 0.77 0.78 0.79 0.8 0.81 PCA IO+PCA(2×2) IO+PCA(2×4) (c) Fig. 5. Face recognition results on the AR database obtained by different linear feature extraction procedures combined with or without the conventional interest operator. (a) Classification right rate associated with 2DPCA, (b) classification right rate associated with 2DFLD and (c) classification right rate associated with PCA. The horizontal coordinate represents the number of the transforming axes of the linear feature extraction procedure used for feature extraction and the vertical coordinate represents the classification accuracy. The horizontal and vertical coordinates of Figs. 6, 7, 9–11 are same as this figure. Table 1 Experimental result comparison of our approach and other approaches on the AR face database. Best accuracy(%) Mean of the accuracy(%) Size of the block 2DPCA 78.8 70.1 2DFLD 81.7 76.4 PCA 79.6 78.9 IO + 2DPCA 79.0 69.5 2 � 4 IO + 2DFLD 81.3 70.0 2 � 4 IO + PCA 78.8 76.0 2 � 4 OIIO1 + 2DPCA 81.7 75.7 2 � 4 OIIO1 + 2DFLD 84.2 78.0 2 � 4 OIIO1 + PCA 82.1 81.4 2 � 4 OIIO2 + 2DPCA 91.9 82.2 2 � 4 OIIO2 + 2DFLD 93.5 86.7 2 � 4 OIIO2 + PCA 91.9 90.1 2 � 4 Y. Xu et al. / Expert Systems with Applications 36 (2009) 9719–9728 9723 4. More discussion on improved and the conventional operators In this section we will compare improved interest operators with conventional interest operator and some other algorithms that exploit gradient information of the image. First, we analyze the relationship and difference between the conventional interest operator and the gradient operator. As presented above, the inter- est operator computes the pixel intensity variation in different directions. Gradient operators also evaluate the pixel intensity var- iation. The difference between the interest operator and the gradi- ent operator are as follows: a gradient operator produces a vector i.e. gradient vector that uses two components to indicate the gra- dient whereas the interest operator obtains only scalar values. The square-root of the squared sum of the two components of the gradient vector is usually used to denote the magnitude value of the gradient. The similarity between the interest operator and the gradient operator is as follows: for the conventional interest operator, r245 and r 2 135 as defined in (4) and (6) act as the mean of the squares of the two components of the Roberts gradient oper- ator (Gonzalez & Woods, 1987), respectively. Additionally, accord- ing to the definition of the gradient, r20 and r 2 90 as defined in (3) and (5) can also be respectively regarded as the mean of the squares of two components of a gradient operator. Consequently, we can conclude that the interest operator and the gradient oper- ator provide image gradient information in different ways. The interest operator aims at calculating ‘average’ gradient information in different directions of an image block, whereas the gradient operator provides gradient information of each image pixel in the form of a vector. Gradient operators and improved gradient opera- tors have been applied to recognition problems (Gao & Leung, 2002; Haber & Modersitzki, 2004; TakaÂcs, 1998; Wei & Lai, 2006) and image matching problems (Ando, 2000; Wolfson & Rigoutsos, 1997). Approaches exploiting image gradient information for recogni- tion can be classified into two classes. The first class of approaches exploits the image gradient information itself to perform recogni- tion. The approaches proposed in Haber and Modersitzki (2004) and Wei and Lai (2006) as well as the conventional interest oper- ator are examples of the first class. In Haber and Modersitzki (2004) and Wei and Lai (2006), ‘normalized gradient’ and ‘relative gradient’ were respectively used as image features. In Wei and Lai (2006), the magnitude of the conventional gradient of a pixel point was first divided by the sum of a constant and the maximum mag- nitude of the gradient in an image block. Then the division result was taken as ‘relative gradient’ of the pixel point. The second class of approaches firstly exploits the image gradient information to produce image edges and then employs image edges to perform recognition. The rationale of the second class of approaches is that edge information is a useful object representation feature (Gao & Leung, 2002). The approaches used in Gao and Leung (2002), Ta- 0 2 4 6 8 10 12 14 16 18 20 0.4 0.5 0.6 0.7 0.8 0.9 1 2DPCA IIO+2DPCA(2×2) IIO+2DPCA(2×4) OIIO(1/2)+2DPCA(2×2) OIIO(1/2)+2DPCA(2×4) (a) 0 2 4 6 8 10 12 14 16 18 20 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 2DFLD IIO+2DFLD(2×2) IIO+2DFLD(2×4) OIIO(1/2)+2DFLD(2×2) OIIO(1/2)+2DFLD(2×4) (b) 110 120 130 140 150 160 170 180 190 200 0.74 0.75 0.76 0.77 0.78 0.79 0.8 0.81 0.82 0.83 PCA IIO+PCA(2×2) IIO+PCA(2×4) OIIO+PCA(2×2) OIIO+PCA(2×4) (c) Fig. 6. Face recognition results on the AR database obtained by different linear feature extraction procedures combined with or without improved interest operator 1. (a) Classification right rate associated with 2DPCA, (b) classification right rate associated with 2DFLD and (c) classification right rate associated with PCA. 0 2 4 6 8 10 12 14 16 18 20 0.4 0.5 0.6 0.7 0.8 0.9 1 2DPCA IIO+2DPCA(2×2) IIO+2DPCA(2×4) OIIO(1/2)+2DPCA(2×2) OIIO(1/2)+2DPCA(2×4) (a) 0 2 4 6 8 10 12 14 16 18 20 0.4 0.5 0.6 0.7 0.8 0.9 1 2DFLD IIO+2DFLD(2×2) IIO+2DFLD(2×4) OIIO(1/2)+2DFLD(2×2) OIIO(1/2)+2DFLD(2×4) (b) 110 120 130 140 150 160 170 180 190 200 0.78 0.8 0.82 0.84 0.86 0.88 0.9 0.92 PCA IIO+PCA(2×2) IIO+PCA(2×4) OIIO(1/2)+PCA(2×2) OIIO(1/2)+PCA(2×4) (c) Fig. 7. Face recognition results on the AR database obtained by different linear feature extraction procedures combined with or without improved interest operator 2. (a): classification right rate associated with 2DPCA. (b): classification right rate associated with 2DFLD. (c): classification right rate associated with PCA. 9724 Y. Xu et al. / Expert Systems with Applications 36 (2009) 9719–9728 Fig. 8. Some face images of a same face. These images are obtained under different illumination conditions. (a) Face images obtained in the cases where the azimuth angle and the elevation angle of the light source with respective to the camera axis are small. (b) Face images obtained in the cases where the azimuth angle and the elevation angle of the light source with respective to the camera axis are quite large. 0 2 4 6 8 10 12 14 16 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 2DPCA IO+2DPCA(2×2) IO+2DPCA(2×4) (a) 0 2 4 6 8 10 12 14 16 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 2DFLD IO+2DFLD(2×2) IO+2DFLD(2×4) (b) 110 120 130 140 150 160 170 180 190 200 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 PCA IO+PCA(2×2) IO+PCA(2×4) (c) Fig. 9. Face recognition results on the YaleB database obtained by different linear feature extraction procedures combined with or without the conventional interest operator. (a): classification right rate associated with 2DPCA. (b): classification right rate associated with 2DFLD. (c): classification right rate associated with PCA. Y. Xu et al. / Expert Systems with Applications 36 (2009) 9719–9728 9725 kaÂcs (1998) and Wang et al. (1998) are three typical examples of the second class. In TakaÂcs (1998), the binary coding result of So- bel edge detection operator was used as face edge feature. In Gao and Leung (2002), the line edge map approach detected the edge feature and then classified faces using the edge feature. It was as- sumed that image edge detection using gradient operators was al- most not influenced by varying lighting condition (Gao & Leung, 2002); however, the approaches in Gao and Leung (2002), TakaÂcs (1998) and Wang et al. (1998) are all not able to obtain truly invari- ant facial features with respect to lighting condition. This is be- cause the pixel intensity value varies with the lighting condition and consequently the gradient value seems to be also variable with respect to the lighting condition. The interest operator appears to be also helpful for indicating salient image edge information in different directions as shown in Fig. 4. Actually, an interest operator obtains ‘average’ directional edge of an image block since it sums gradient information in a cer- tain direction. Improved interest operators have the following advantage: by improving the conventional interest operator, im- proved operators enable the obtained ‘average’ directional edge information of a face image block to be less variable with respect to the lighting condition. This is very beneficial to face recognition. In addition, since the average relative variation of the pixel inten- sity appears to in some extent be robust to facial expression and pose, improved interest operators may also produce more stable face feature than the conventional interest operator under the con- ditions of varying facial expression or pose. 5. Experimental result In this section we compare our approach and the conventional interest operator using the AR and YaleB face databases. We eval- uate the similarity between the feature extraction results of two face images by using sðx; yÞ¼ xT y jjxjj � jjyjj ; ð17Þ where x, y stand for two vectors respectively corresponding to the feature extraction results of the two face images. After the similar- ities between a testing sample and each of all training samples are computed, we select out the training sample that has the maximum similarity to the testing sample and classify the testing sample into the class that the selected training sample belongs to. 5.1. Experiments on the AR face database The AR face database includes more than 4000 face images showing faces with different facial expressions, in varying lighting 0 2 4 6 8 10 12 14 16 0.65 0.7 0.75 0.8 0.85 2DPCA IIO+2DPCA(2×2) IIO+2DPCA(2×4) OIIO(1/2)+2DPCA(2×2) OIIO(1/2)+2DPCA(2×4) (a) 0 2 4 6 8 10 12 14 16 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 2DFLD IIO+2DFLD(2×2) IIO+2DFLD(2×4) OIIO(1/2)+2DFLD(2×2) OIIO(1/2)+2DFLD(2×4) (b) 110 120 130 140 150 160 170 180 190 200 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 PCA IIO+PCA(2×2) IIO+PCA(2×4) OIIO(1/2)+PCA(2×2) OIIO(1/2)+PCA(2×4) (c) 9726 Y. Xu et al. / Expert Systems with Applications 36 (2009) 9719–9728 conditions and occluded in several ways (Yang et al., 2004)1. Each subject have 26 face images. For each subject, the filenames of the 26 images contain the numbers from 1 to 26, respectively. If the number contained in the filename is 1, we call the image the first image of the subject, and so on. We use the computer generates 13 random integers in the range of from 1 to 26. For every subject, we take the 13 images whose filenames contain the numbers from the random integer sequence as the training samples and consider the remaining samples as test samples. The generate random integer sequence is 2, 4, 7, 8, 10, 15, 16, 19, 21, 22, 24, 25, 26. Fig. 5 shows face recognition results on the AR database ob- tained by different linear feature extraction procedures combined with or without the conventional interest operator. Note that here- after by IO (Interest Operator) we denote the conventional interest operator. By IIO1 we denote improved interest operator 1 based on the non-overlapping partition scheme. By IIO2 we denote im- proved interest operator 2 based on the non-overlapping partition scheme. In addition, we use OIIO1 (Overlapped Improved Interest Operator 1) to represent improved interest operator 1 based on the overlapping partition scheme. We also use OIIO2 to represent improved interest operator 2 based on the overlapping partition scheme. If the interest operator is followed by a linear feature extraction procedure, we add some indicators to the denotation. For example, ‘IO + 2DPCA’ means that feature extraction is imple- mented by the conventional interest operator followed by 2DPCA. We use a bracket to show the size of each block or the overlapping region. For example, ‘IO + 2DPCA(2 � 2)’ means that the image is partitioned into a number of 2 � 2 non-overlapping image blocks; whereas ‘OIIO(1/2) + 2DPCA(2 � 2)’ means that the image is parti- tioned into a number of 2 � 2 overlapping image blocks and 1/2 re- gions of two neighboring blocks are overlapped each other. According to Fig. 5, the combination of PCA and the conventional interest operator is not able to obtain a higher accuracy than PCA. The combination of 2DPCA and the conventional interest operator may produce a higher or lower accuracy than 2DPCA. So does the combination of 2DFLD and the conventional interest operator. Table 1 shows experimental result comparison of our approach and other approaches on the AR face database. From Table 1 and Figs. 5–7, we can see that the combination of either of the two im- proved interest operators and 2DPCA or 2DFLD can produce a high- er accuracy than 2DPCA or 2DFLD. The combination of an improved interest operator and 2DPCA or 2DFLD also performs better than the combination of the conventional interest operator and 2DPCA or 2DFLD. Fox example, the combination of the conventional inter- est operator and 2DPCA obtains the mean of recognition accuracy of 69.5% and the best accuracy of 79.0%. When OIIO1 is combined with 2DPCA, the mean of the accuracy and the best accuracy are 75.7% and 81.7%, respectively. For the combination of OIIO2 and 2DPCA, the mean of the accuracy and the highest accuracy are 82.2% and 91.9%, respectively. This means that compared to the combination of the conventional interest operator and 2DPCA, the combination of OIIO2 and 2DPCA improves the mean accuracy and the highest accuracy 12.7% and 12.9%, respectively. 5.2. Experiments on the YaleB face database The images in the YaleB database2 are obtained with varying illuminations and unfixed poses and there exists a wide rang of illumination cases. To focus on face recognition with varying illu- minations, we select and use 45 face images with pose 00 of every person. We crop each of these images to obtain a 32 � 32 image Fig. 10. Face recognition results on the YaleB database obtained by different linear feature extraction procedures combined with or without improved interest operator 1. (a): classification right rate associated with 2DPCA. (b): classification right rate associated with 2DFLD. (c): classification right rate associated with PCA.1 h t t p :/ / c o b w e b . e cn . p u rd u e . e du / ~ a l e i x / a l e i x _ f a c e _ D B . h t m l ; h t t p : // c o b - web.ecn.purdue.edu/~aleix/ar.html. 2 http://cvc.yale.edu/projects/yalefacesB/yalefacesB.html. http://cobweb.ecn.purdue.edu/~aleix/aleix_face_DB.html http://cobweb.ecn.purdue.edu/~aleix/ar.html http://cobweb.ecn.purdue.edu/~aleix/ar.html http://cvc.yale.edu/projects/yalefacesB/yalefacesB.html 0 2 4 6 8 10 12 14 16 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 2DPCA IIO+2DPCA(2×2) IIO+2DPCA(2×4) OIIO(1/2)+2DPCA(2×2) OIIO(1/2)+2DPCA(2×4) (a) 0 2 4 6 8 10 12 14 16 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 2DFLD IIO+2DFLD(2×2) IIO+2DFLD(2×4) OIIO(1/2)+2DFLD(2×2) OIIO(1/2)+2DFLD(2×4) (b) 110 120 130 140 150 160 170 180 190 200 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 PCA IIO+PCA(2×2) IIO+PCA(2×4) OIIO(1/2)+PCA(2×2) OIIO(1/2)+PCA(2×4) (c) Fig. 11. Face recognition results on the YaleB database obtained by different linear feature extraction procedures combined with or without improved interest operator 2. (a): classification right rate associated with 2DPCA. (b): classification right rate associated with 2DFLD. (c): classification right rate associated with PCA. Table 2 Experimental result comparison of our approach and other approaches on the YaleB face database. Best accuracy(%) Mean of the accuracy(%) Size of the block 2DPCA 77.2 74.5 2DFLD 75.2 69.0 PCA 58.8 58.8 IO + 2DPCA 78.4 70.9 2 � 2 IO + 2DFLD 78.8 72.3 2 � 2 IO + PCA 81.6 79.7 2 � 2 OIIO1 + 2DPCA 82.8 75.5 2 � 4 OIIO1 + 2DFLD 92.8 84.1 2 � 2 OIIO1 + PCA 85.8 85.1 2 � 4 OIIO2 + 2DPCA 94.4 89.9 2 � 2 OIIO2 + 2DFLD 98.0 90.3 2 � 4 OIIO2 + PCA 95.6 92.1 2 � 2 Y. Xu et al. / Expert Systems with Applications 36 (2009) 9719–9728 9727 (Xu, Yang, Jin, & Zheng, 2006). We test our approach using the ob- tained images. Some face images of a same face are shown in Fig. 8. Fig. 9 shows face recognition results on the YaleB database ob- tained by different linear feature extraction procedures combined with or without the conventional interest operator. Fig. 10 shows face recognition results on the YaleB database obtained by differ- ent linear feature extraction procedures combined with or without improved interest operator 1. Fig. 11 shows face recognition results on the YaleB database obtained by different linear feature extrac- tion procedures combined with or without improved interest oper- ator 2. Table 2 shows experimental result comparison of our approach and other approaches on the YaleB face database. Table 2 and Figs. 9–11 show that the combination of improved interest operator 1 or 2 and a linear feature extraction procedure performs better than the combination of the conventional interest operator and the same linear feature extraction procedure. Fox example, the combination of the conventional interest operator and 2DFLD obtains the mean of recognition accuracy of 72.3% and the highest recognition accuracy of 78.8%. When OIIO1 is combined with 2DFLD, the mean of the accuracy and the highest accuracy are 84.1% and 92.8%, respectively. For the combination of OIIO2 and 2DFLD, the mean of the accuracy and the highest accuracy are 90.3% and 98.0%,, respectively. This means that compared to the combination of the conventional interest operator and 2DFLD, the combination of OIIO2 and 2DFLD improves the mean accuracy and the highest accuracy 18.0% and 19.2%, respectively. 5.3. More experimental analysis In this subsection, we further investigate our approach by ana- lyzing the similarity of different images of the same face for the AR face database. Analysis is performed for the first, second, seventh, eighth and eleventh face images of each subject. The used five images of one subject are shown in Fig. 12. When compared with image (a), images (b) has different expression, image (c) is ob- tained under different lighting condition, images (d) and (e) are two occluded face images. Note that after the conventional interest operator or the pro- posed interest algorithm is applied to all the five images of a face, we can also evaluate the similarities between the resultant images Fig. 12. Illustration of the five images of a face used for similarity analysis. Table 3 Means of the similarities between the resultant images of (a) and those of (b), (c), (d), (e). Each row shows the means of the similarities between the resultant images associated with an interest operator of image (a) and images (b), (c), (d), (e). IIO1 and IIO2 respectively denote improved interest operators 1 and 2 applied to the non- overlapping partition scheme. Similarity of (b) and (a) Similarity of (c) and (a) Similarity of (d) and (a) Similarity of (e) and (a) IO 0.916 0.861 0.848 0.816 IIO1 0.946 0.941 0.914 0.868 OIIO1 0.950 0.950 0.920 0.877 IIO2 0.966 0.959 0.924 0.901 OIIO2 0.968 0.964 0.938 0.911 9728 Y. Xu et al. / Expert Systems with Applications 36 (2009) 9719–9728 of image (a) and each of images (b), (c), (d), (e) by using s1ðx1; y1Þ¼ xT 1 y1 jjx1jj�jjy1jj , where x1, y1 represent the two one-dimen- sional vectors of the resultant images of the two original face images, respectively. Note that an interest operator produces five resultant images for each original image, so the corresponding one-dimensional vector used for similarity computation is ob- tained by treating the five resultant images of the original image as an integral image and by concatenating the rows or columns of the integral image. For each subject, we respectively compute the similarities be- tween the resultant images of image (a) and those of images (b), (c), (d), (e) produced by an interest operator. Then we calculate the means of the similarities based on all the subjects. The means obtained in the case where each image block is of the same size of 2 � 2 are shown in Table 3. In the overlapping block partition scheme, half of each block is overlapped by a neighboring block. The second to fifth columns of Table 3 respectively show the means of the similarities between the resultant images of image (a) and those of images (b), (c), (d), (e). From Table 3, we can see that, when improved interest opera- tors 1 and 2 are applied to non-overlapping image blocks, the ob- tained similarities between the resultant images are higher than the similarities of the resultant images obtained using the conven- tional interest operator. Moreover, OIIO1 and OIIO2 produce higher similarities respectively than IIO1 and IIO2. This implies that the combination of the improved interest operator and the overlapping block partition scheme is more helpful for obtaining high similar- ities than the combination of the improved interest operator and the non-overlapping partition scheme. Therefore, we can conclude that both the improved interest operator and the overlapping par- tition scheme are useful for resulting in higher similarities for the resultant images of image (a) and those of images (b), (c), (d), (e). This also means that our approach can effectively reduce the differ- ence between different images of the same face, which is very ben- eficial to face recognition. 6. Conclusion The new approach proposed in this paper differs from the fea- ture extraction approach using the conventional interest operator in the following two aspects. The first aspect is that the proposed approach takes the relative rather than absolute variation of the pixel intensity as the feature of an image block. This allows the proposed approach to obtain robust block features that are less af- fected by varying imaging conditions such as varying lighting and facial expression. The second aspect is that the proposed approach adopts the overlapping block partition scheme. This enables the proposed approach to fully reflect the pixel intensity variation information and to produce more representative feature extraction results. As a result, the proposed approach can do better in face recognition than the conventional interest operator. The analysis and experimental results illustrate the feasibility and the satisfactory performance of the proposed approach. Exper- imental results show that the use of the proposed approach can ob- tain more than 10 percent accuracy improvement. Acknowledgement We wish to thank 863 Program Project (No. 2006 AA01Z193), National Natural Science Foundation of China (Nos. 60602038, 60632050) and Natural Science Foundation of Guangdong prov- ince, China (No. 06300862) for supporting. References Adini, Y., Moses, Y., & Ullman, S. (1997). Face recognition: The problem of compensating for changes in illumination direction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(7), 721–732. Ando, S. (2000). Image field categorization and edge/corner detection from gradient covariance. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(2), 179–190. Cho, D. U., Chang, U. D., Kim, B. H., & Lee, S. H. (2006). 2D direct LDA algorithm for face recognition. In Proceedings of the fourth international conference on software engineering research, management and applications, WA, United States (pp. 245– 248). Căleanu, C. D. (2000). Facial recognition using committee of neural networks. In Proceedings of the fifth seminar on neural network applications in electrical engineering, NEUREL, Belgrade, Yugoslavia (pp. 97–100). Căleanu, C. D. (2001). Face recognition using parallel neural processing and interest operator method. Ph.D. Thesis, University POLITEHNICA Timisoara. Căleanu, C. D., Huang, D.-S., Gui, V., Tiponu, V., & Maranescu, V. (2007). Interest operator versus Gabor filtering for facial imagery classification. Pattern Recognition Letters, 28, 950–956. Gao, Y., & Leung, M. K. H. (2002). Face recognition using line edge map. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24, 764–779. Gonzalez, Rafael C., & Woods, Richard E. (1987). Digital image processing (2nd ed.). Boston, MA, USA: Addison-Wesley Longman Publishing Co., Inc.. Haber, E., & Modersitzki, J. (2004). Intensity gradient based registration and fusion of multi-modal images. Technical Report, Department of Mathematics & Computer Science, Emory University, Atlanta. Kongsontana, S. & Rangsanser, Y. (2005). Face recognition using 2DFLD algorithm. In Proceedings of the eighth international symposium on signal processing and its applications (vol. 2, pp. 675–67). Moravec, H. (1981). Robot rover visual navigation. Ann Arbor, MI: University of Michigan Research Press. Mutelo, R. M., Khor, L. C., Woo, W. L., & Dlay, S. S. (2006). A novel fisher discriminant for biometrics recognition: 2DPCA plus 2DFLD. ISCAS2006, IEEE, 4325–4328. Nasrabadi, N. M., & Choo, C. Y. (1994). Hopfield network for stereo vision correspondence. In M. Gupta & G. Knopf (Eds.). Neuro vision systems principles and applications (vol. 2, pp. 442–458). IEEE. Nhat, V. D. M., & Lee, S. (2005). Improvement on PCA and 2DPCA algorithms for face recognition. In Proceedings of the fourth international Conference on I Singapore Image and Video Retrieval (CIVR), Singapore (pp. 568–577). Sanguansat, P., Asdornwised, W., Jitapunkul, S., & Marukatat, S. (2006). Two- dimensional linear discriminant analysis of principle component vectors for face recognition. IEICE Transactions on Information and Systems, E89-D(7), 2164–2170. TakaÑcs, B. (1998). Comparing face images using the modified hausdorff distance. Pattern Recognition, 31, 1873–1881. Wang, L. C., Der, S. Z., & Nasrabadi, N. M. (1998). Automatic target recognition using a feature decomposition and data decomposition modular neural network. IEEE Transactions on Image Processing, 7, 1113–1121. Wang, L., Wang, X., Zhang, X., & Feng, J. (2005). The equivalence of two-dimensional PCA to line-based PCA. Pattern Recognition Letters, 26, 57–60. Wei, Shou-Der, & Lai, Shang-Hong (2006). Robust and efficient image alignment based on relative gradient matching. IEEE Transactions on Image Processing, 15, 2936–2943. Wolfson, H., & Rigoutsos, I. (1997). Geometric hashing: An overview. IEEE Computational Science & Engineering Magazine, 4(4), 10–21. Xu, Y., Yang, J.-Y., Jin, Z., Zheng, Y.-J. (2006). Local correlation classification and its application to face recognition across illumination. In The international conference of machine learning and cybernetics, Dalian (pp. 3277–3281). Xu, Y., Zhang, D., Yang, J., & Yang, J.-Y. (2008). An approach for directly extracting features from matrix data and its application in face recognition. Neurocomputing, 71, 1857–1865. Yang, J., Zhang, D., Frangi, A. F., & Yang, J. Y. (2004). Two dimensional PCA: A new approach to appearance-based face representation and recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24, 131–137. Zhao, T. (2007). Several approaches for feature extraction and selection for face recognition. M.S. Degree Thesis, University of Harbin Institute of Technology (in Chinese). Zhao, Z. Q., Huang, D. S., & Sun, B. Y. (2004). Human face recognition based on multi- features using neural networks committee. Pattern Recognition Letters, 25, 1351–1358. Improving the interest operator for face recognition Introduction The conventional interest operator Description of our approach Improved interest operator 1 Improved interest operator 2 Overlapping block partition scheme More discussion on improved and the conventional operators Experimental result Experiments on the AR face database Experiments on the YaleB face database More experimental analysis Conclusion Acknowledgement References 
work_3w545bukofdphbrifcp74hjary	----	NASA Technical Reports Server (NTRS) 
work_3x5tecw52rfrjgj4blfiqt3qne	----	.E &DflFH 940BI5-4 119 DEVELOPMENT OF AN EXPERT SYSTEM FOR TRANSPORTATION OF HAZARDOUS AND RADIOACTIVE MATEIUALS J. J. Ferrada R D. Michelhaugh R R Raw1 Oak Ridge National Laboratory* Oak Ridge, Tennessee 37831-6495 For presentation and publication at the "Spectrum '94: International Nuclear and Hazardous Waste Management meeting in Atlanta, Georgia August 14 - 18,1994 *Managed by Martin Marietta Energy Systems, Inc,, for the U. S. Department of Energy under contract DE-ACO5-84OR21400. r ', TO 13685 DEVELOPMENT OF AN EXPERT SYSTEM FOR TRANSPORTATION OF HAZARDOUS AND RADIOACTIVE MATERIALS J. J. Fayac& R. D. Mi&lhaugh R. R Raw1 Oak Kidgc Natwnal i..abarator)rs F. 0. Box 2008, MS-6495 Oak Riduc. Tennessee 3783 1 Oak R i d s Nauonsl Laboratory+ Oak Ridge National Laboratory P. 0. Box 2008, MS6495 Oak Ricfne. Tamessee 3783 1 P. 0. BOX 2008, MS-6495 oak Ridae, T~~ 3783 1 (615) 574-4998 (615) ~ 4 - 6 8 1 9 L INTRODUCTION IL DEVELOPMENT (615) 57744713 The iniaai prototype to dunonmate the feasibility of dcvtloph the .expat systan was based on a mwiufe for &-g and pdcagmg radioactive material. Later, devciopbd feasibility stage inciuded ( I ) andysia of comm&d software related to regulation ~ ~ c e s 9 . (2) lcnourlsd~c q u i d t i o n , and (3) d e v c l o p m t of the la#a subtask WM to (a) develop m 4 d s La cap- the howldgc of Ciifl-1 of kamprt&tim and p s d a p g and (b) nrUIyLt the fessibility of e p p d h g these diEcrcnt moduItJ as onrr frnal full program. Two i d h i d u d modula arc Usat. anc for esnspadng and pdcpgtng of radiosctivc materials and Motha for trmspedng and packaghg hazardous chemical matcrsls. The final product will integrate &me two m d r r into an o v d l hazardous and radioaclivc m m d s system. B ~nodulc ha! kludai hetar&= c h d o d s W ~ S c!4pat systam prototype. ne strattgy to develop the *Managed by Martin Mmena Energy Systans, Inc. for the U.S. Dcpawtnt of Energy under contract DE-ACOS -&OR2 1400. 1 May 20% 1'3'34 Q2: 59PM FRDM ec--A sect i o n TD 13685 P.03 . 2 Kmmkdge rcqubtthn. The dcvtinpmhlt of an expert system rn any developmcntsl stage requva R dose relahonhp between the developer and he subjcct expert. Awniingly, lhr knuwidge muvr flow from llle expert to the dcvclopcrl in a marina such that the latter CM visuake the "branches and Ilunk" of tiu tequucd knowledge "m." Typicaily, &e initirl prototype IS caoccmd with oniy a particular poruon of tho problcm and do03 uot providc tho fulI rangc of uttimatt solurions. 2 b w d an he value ofthis f r d ~ a snd will dctonnrnc the path for the following stages of the tc?gic diagram Fig. 1. R STAGE 1 ...... e....-.- STAGE 2 -.--___. ST4GE 3 ... - P. Q4 ....._....___.... . .......--/--. . DOES~JOTOUAUFY wherein n de-bescd system thu repme& the applicstion of regulations CM be implcmcntcd 3 c May =PI 1994 Q3:QBPM FROM ec-a s e c t . i o n Rrh@ ulf lhtst critaia sugpskd that V i d Basic WM the best cllwoLuLLcIt in wh& to rrwtr thc expert mcm. Although a Prolog-bastd codc had already been created during the proof-ofanccpt stage, nnd it was cicar that putung the Proiog v m o n tog& mqukcd littlo additional etfwt, rnulrimcdi. f e w w e n option was abandoned The solu~on fatmd for the pmtotyping stagc was to ltadslate the Pmlog cadc into Visual Basic code, which is a Windows application. diftlcdt to obtain using ody Prolag. Coqueady, this TO 13685 P.05 . 4 tqay 2 0 , 1 9 9 4 6 3 : 01PM FRUM ec-a s e c t ion matenal q d i f i u as any of thee PSNs. Ifnat. chnr the functiddcscr~puons p u p IS c k k d to sec if the fimcti00 a€ the matenal is h s t d lfthe nwmrl does not fit any of& p m ~ u PSNs, thcn thc bpprqlnrtcr h c u h m the ~ c r d hatard dass, n.o.s. IS d x t d If& matend Qcs not mctt the d&iuon of any of the DOT hazard clssecs it sull might be regulated 89 i.u&(al by it9 appemmg rn the HMT appendmx F d y . if the matsnai i9 c l d e d ns an EPA wRstc but doem't meet the dehitian of my DOT hazard dass or haznrdous substance, thcn the PSN af h&us waste, n.0.a is selected TO 13635 P. 06 IV. CCINCLUSION 5 F.Q7 Fig. 2. Chemlcal hazardow m&lcrirl hgk diagram. p ~ ~ p ~ s c d in HM+ 169A'. T t h watlc WM pcrfarmcd anticipation o f t h e pubIication of thc Final Ruie MM- 169A tbis ytat. REFERENCES 1. 2. 3. 4. 5. 6 R. Mic&lha* Portwrouth aiaermw Diffusion Plant. pasanal canmunicatim to 2.1. Ferrads. Oak Ridge National L h r a t q y , Oak Ridge. Tenncsscc, Fe?)nmy 1992. J J. Fnrsda V. M. Green. and R R Rawl, "FesrbdhyS- . AppiicatiCn ofthe Hazardous Material Regulations. O m - I 2 14 1, Oak Ridge Nationrl I.-, Oak Ridge, T c m ~ a c ~ , Sepl~mbs, 1992. HM-169A. Fadaal Register, Vd. 54, NO. 218, p. 47454. Novcrnh. 1987. TOTFtL P. 07 DISCLAIMER This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or use- fulness of any information, apparatus, product, or proccss disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any spc- cific commercial product, process, or service by trade name, trademark, manufac-. turer, or otherwise does not necessarily constitute or imply its endorsement, recom- mendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. 
work_3x6bngh4dnfylnzuyowf6pdo6u	----	Unknown Scptcmbcr 1982 Also numbered: HPP-82-3 Report No. STAN-CS-82-932 Expert Systems Research: Modeling the Medical Decision Making Process bY k&ad I-i. Shortliffc and Larvrcncc M. Pagan Departments of Medicine and Computer Science Stanford University Sl;mford, CA 94305 Technical Memo HPP-82-3 . EXPERT SYSTEMS RESEARCH: MODELING THE MEDICAL DECISION MAKING PROCESS* Edward H. Shortliffe, MD, PhD Lawrence M. Fagan, PhD:* Heuristic Programming Project Departments of Medicine and Computer Science Stanford University Stanford, California 94305 Presented at the Workshop on Integrated Approached to Patient Monitoring University of Florida at Gainesville 5-6 March 1982 *The Ventilator Manager project was supported by NIH Grant GM-24669, awarded to the Institute of Medical Science at Pacific Medical Center (San Francisco). Computing resources were provided by the SUMEX-AIM facility at Stanford University under NIH Grant RR-00785. Dr. Shortliffe is recipient of research career development award LM-00048 from the National Library of Medicine. **Current address: Phd to MD program, University of Miami School of Medicine, Miami, F l o r i d a 3 3 1 0 1 . Shortliffe and Fagan HPP-82-3 Int reduction During the quarter century since the birth of the branch of computer science known as artificial intelligence (Al), rnuh of the research has focused on developing symbolic models of human inference. In the last decade several related Al research themes have come together to form what is now known as “expert systems research.“’ In this paper we review Al and expert systems to acquaint the reader with the field and to suggest ways in which this research will eventually be applied to advanced medical monitoring. Knowledge Engineering Artificial intelligence has been described as “the study of ideas that enable computers to do the things that make human beings seem intelligent.“* An implicit assumption is that the computer should have the ability to reason symbolically (rather than by combining numbers statistically or by using other numerical manipulations that underlie conventional computer programs). Related assumptions are that intelligent programs should be able to acquire new knowledge and to apply it appropriately; they should also be able to manipulate and communicate ideas. Of particular relevance to the study of medical inference is Al research into the construction of systems for knowledge-based consultation. “Knowledge” is the key word and must be distinguished from “data.” Computers have long been used to store data, but isolated data points do not become knowledge until they have been analyzed and summarized. Accordingly, we suggest that there are at least four types of knowledge that should be distinguished from statistical data. These characterize the information that must be available to an expert system? 1. knowledge derived from data analysis (largely numerical or statistical); 2. judgmental or empirical subjective knowledge -- the kind that experts recognize is based on their own experience but which may be difficult toverify without complex and time- consuming studies; 3. common sense or scientific knowledge -- these kinds of knowledge are often simple facts (e.g., cars are for transportation, Gainesville is in Florida), but are symbolic in nature and must be known by an expert in a field; Shortliffe and Fagan HPP-82-3 4. strategic knowledge or “self-knowledge” -- this is the kind of knowledge that often distinguishes an expert from a’well-trained novice in a field, e.g., an anesthesiologist’s method of selecting an optimal anesthetic agent for a compromised patient or a cardiologist’s favorite technique for choosing the diagnostic tests by which to assess a patient with a new complaint of chest pain; thus, judgment combined with a “strategic plan” can be used to adapt a program to idiosyncratic situations. An expert system, then, is a computer program that contains and can apply specialized knowledge. It uses this knowledge to make suggestions to users who may not have the full range of expertise available to the system. The construction of such programs is known as ‘*knowledge engineering ‘14* ’ and typically requires close collaboration between human experts in a field and computer scientists familiar with expert systems research. In order to simulate an encounter between a non-expert and a human consultant, an expert system generally must contain all four kinds of knowledge outlined above-and, thus, is more than a conventional data retrieval system. An expert system functions as an interface between. the intended user of the system and the domain expert (or experts) who collaborated in its construction. Since the experts may not be generally available to all who would like their advice, the expert system becomes a surrogate. During the encounter between an expert and a person seeking advice, there are three types of information transfer: 1. the expert requests information about the case under consideration; 2. the expert offers a recommendation or conclusion based on the available data; and 3. if requested, the expert explains the basis for the final decisions. Conventional approaches to computer-based consultation typically address only the first two of these items. It is when a system designer wishes to construct programs that can explain the basis for their decisions, in terms that a physician can easily understand, that the techniques of knowledge engineering become particularly pertinent. -2- Shortliffe and Fagan HPP-82-3 Questions Asked By Physicians In our work designing and building medical expert systems at Stanford over the last decade, we have been guided by key questions that are frequently asked by physicians and that help to direct the design of the program:! . 1. Do I need this system? 2. Will it help without being dogmatic? 3. Does it justify its recommendations so that I can judge them for myself? 4. Is it fast and easy to use? 5. Is it designed to make me feel comfortable when I use it? The goal of most such systems is to answer all five questions in the affirmative. We will address these questions indirectly by discussing the major themes in expert systems research. An early expert system, known as MYCIN, is used to illustrate many pertinent principles. We also briefly describe a more recent program, the Ventilator Manager system, that applies knowledge engineering to monitoring patients in the intensive care unit. The MYCIN System The‘ MYCIN program was developed at Stanford University in the mid-1970s7 and remains a good example of issues in the design and construction of knowledge-based programs. MYCIN was - designed to advise physicians regarding the selection of antibiotics for patients with severe infections. When designing MYCIN, we were aware that the need for this kind of consultation system would not guarantee its acceptance by physicians. It was also important that it reach decisions comparable to those of infectious disease experts and that it be able to explain the basis for its decisions. Artificial intelligence techniques seemed to offer solutions to these design considerations! Shortliffe and Fagan ’ Issues in Knowledge Engineering HPP-82-3 Knowledge Representation We have described the four kinds of knowledge that an expert system requires to provide specialized consultation. An active area of Al research is developing new techniques to encode and manipulate such symbolic (non-numeric) knowledge. The method used by the MYCIN and Ventilator Manager programs is called production rules (Fig. 1). Other methodologies include frames, semantic networks,* and combinations of these. Most knowledge can be encoded in whatever formalism a designer wishes, but the representation technique must accomplish the following: 1. formulation of explanations that convey a line of reasoning that a user can understand and critique; 2. separation of the domain knowledge from the program itself; in this way, knowledge can be changed without affecting the computer program; 3. a level of performance such that an expert can easily identify and correct faults in the knowledge; and 4. interaction with a n inferential method such that the system reliably reaches excellent decisions. Drawing on past experience with production rules in a large system that reasoned about chemical -structure’, we used production rules9 to encode the knowledge of infectious disease therapy”. A production rule is a conditional statement that indicates the circumstances under which a-particular conclusion may be drawn (Fig. 1). MYCIN contains about 550 rules dealing with the diagnosis and treatment of bacteremia and meningitis. The rules are encoded in a stylized language that th? computer can interpret, and routines have been written to translate them into English for display to the user as shown in the figure. The strengths of the conclusions are indicated by numerical weights called certainty factors (CFs).” Most expert systems that deal with uncertain inferences have been forced to incorporate scoring systems to keep track of the weight of evidence favoring competing hypotheses. As is true in other complex domains, human actions and beliefs in medicine interact with Shortliffe and Fagan HPP-82-3 I f : 1 ) T h e i n f e c t i o n w h i c h r e q u i r e s t h e r a p y i s m e n i n g i t i s , and 2 ) T h e p a t i e n t h a s e v i d e n c e o f a s e r i o u s s k i n o r s o f t t i s s u e i n f e c t i o n , a n d 3 ) O r g a n i s m s w e r e n o t s e e n o n t h e s t a i n o f t h e c u l t u r e , and 4 ) T h e t y p e o f i n f e c t i o n i s b a c t e r i a l Then : T h e r e i s e v i d e n c e t h a t t h e o r g a n i s m ( o t h e r t h a n t h o s e s e e n o n c u l t u r e s o r s m e a r s ) w h i c h m i g h t b e c a u s i n g t h e i n f e c t i o n i s s t a p h y l o c o c c u s - c o a g - p o s (.75) o r s t r e p t o c o c c u s (5) Figure 1: A sample rule from the MYCIN system shows that the premise of the rule is a set of conditions (here the four clauses following the word “If:“) that are tested by the program. If the conjunction of the premise conditions is true, the conclusion or action of the rule (the !‘Then:” clause) is executed and the inference is appropriate. common sense and with precise knowledge in ways that defy simple delineation. The challenge at this stage in the development of expert systems is, therefore, to constrain the task realistically so that the program is tenable but still allows useful decisions for solving real problems. This has been demonstrated by Clancey in his use of MYCIN for tutoring medical students.‘* Rules that had been adequate for excellent performance in a consultation system were seriously deficient when they were used for teaching. l3 We have also found that MYCIN-like rules need substantial modification to deal- with inference in settings like medical monitoring where parameters change rapidly and analyses of temporal trends are crucial.‘4 Knowledge Acquisition As researchers gain experience in constructiing consultation systems, identifying and encoding expert knowledge has’been perceived as one of the most complex and arduous tasks encountered. Experts often have difficulty distilling their knowledge, and, because there are other major problems in identifying and encoding knowledge, the most appropriate domains for expert systems research at this time may therefore be fields in which the knowledge is already highly-structured and well- -5- Shortliffe and Fagan HPP-82-3 specified.*. In the case of MYCIN, we obtained rules regarding antibiotic selection and bacterial identification by talking with experts in infectious disease. We quickly learned that people express their knowledge best in the context of problem solving. Thus, we presented difficult cases to the experts, observed their performance as they gathered information they needed when deciding how to treat, and asked pertinent questions when the experts seemed to make leaps in logic or to ignore certain data. In this way, rules were obtained and then encoded using the formalism of production rules. One member of our group, intrigued with the idea of a’ program that could guide this kind of knowledge gathering, developed a program known as TEIRESIAS? It used the rules already known to MYCIN combined with learning strategies to help an expert identify missing or erroneous knowledge and to update MYCIN’s rules interactively. TEIRESIAS provides an example of one form of “machine learning’* (learning by being told, as opposed to learning by experience or by analogy). Models of Reasoning Once knowledge has been captured from experts, books, or other sources and once it has been encoded, the program must have an effective method to search through the knowledge base for relevant facts and to tie them together in ways that simulate human reasoning. There are two issues: lj control of the reasoning process, and 2) management of uncertainty. (Control of the Reasoning Process Entire books have been written on techniques for traversing a symbolic search space.17 The approach we used in MYCIN is known as goal-directed reasoning or backward-chaining. MYCIN’s rules are only loosely related to one another before a consultation begins, i.e., the system builder *One of our present research projects, ONCOCIN,” is a cancer chemotherapy consultation program. We chose this field for our current work precisely because the knowledge in the protocols used to guide the treatment of these patients is formalized, written down, and subject to rigorous analysis. Shortliffe and Fagan . HPP-82-3 does not tell the program explicitly how or when the rules should interact. As it considers a particular patient, MYCIN selects the relevant rules and chains them together. Two rules chain together if the conclusion of one helps determine the truth value of a condition in the premise of the other; thus, the resulting reasoning network is created dynamically1 MYCIN reasons backward from its goal of determining therapy for a patient. It starts by considering rules for therapy selection, but the premise of each of those rules in turn sets up new questions or subgoals. ‘These new goals then invoke new rules and a reasoning network is thereby developed. When the truth of a premise is best determined by asking the physician rather than by applying a rule, e.g., to determine the value of a laboratory test likely to be known by the doctor, a question is displayed. The physician responds and the program continues to select additional rules. Once information on the patient is entered, some rules will fail to apply. The rules that are invoked provide a chain of inference specific to the case under consideration. When reasoning is based on observations rather than goals, problem solving proceeds forward -- from data to conclusions -- so-called “data-driven” reasoning. Many expert systems use this alternative approach, and there are psychological data to suggest that “forward reasoning” more closely approximates the way human beings solve complex problems. As discussed below, the Ventilator Manager (VM) program uses the forward reasoning approach. A series of rules are - examined to determine the reliability of incoming data. In successive steps, VM considers the meaning of the measurements in the current context. Management of Uncertainty When individual inference steps are not certain, a level of complexity is added to the interpretation of conclusions reached by an intelligent program. Not only are conventions needed to assign weights to the rules, but also, when two or more pieces of evidence support the same conclusion, some mechanism is needed to determine the net strength of the hypothesis. Expert system researchers have found that formal probability theory is less than satisfactory for these -7- Shortliffe and Fagan HPP-82-3 purposes, although some have attempted’to adapt it to symbolic reasoning systems.‘8 This issue has been particularly relevant for MYCIN, where the knowledge expressed in a rule tends to include “suggestive“ or “strongly suggestive” evidence for a given conclusion. The need to combine pieces of evidence regarding a single hypothesis, but derived from a number of different rules, requires a numeric system to capture and represent an expert’s measure of belief regarding the inference stated in the rules. Conditional probabilities were not adequate for this purpose,” and we devised instead the system of “certainty factors” mentioned earlier. These numbers lie on a -1 to + 1 scale, with -1 indicating absolute disproof of a hypothesis, + 1 indicating its proof, and 0 indicating the absence of evidence for or against the hypothesis (or evidence equally weighted in both directions). We have described the model’s relationship to formal probability theory, and its methods for combining evidence from diverse sources (rules and user estimates).” Generation of Good Advice The ultimate test of the validity of the model of reasoning used in an expert system is its ability to reach accurate conclusions and thereby to give valuable advice. Our discussion of system evaluation below mentions some of the difficulties that arise in attempting to show that a program is performing at the level of an expert. There is a related controversy among expert systems researchers. Some do not believe it will be possible for computer-based consultants to function at the level of human experts until we better understand and model the inferences used by intelligent problem solvers. Others argue that the details of a reasoning model are not important as long as the system reaches good decisions and can explain its reasoning in understandable terms. However, psychological studies of human problem solving, e.g., Elstein’s work in the field of medicine,lg are being studied increasingly by knowledge engineers as they attempt to develop systems that are more robust. Explanation of Decisions As we have frequently stressed, a crucial aspect of the interaction between an expert and his client is the ability of the non-expert to iequest explanations regarding the recommendation received. -8. Shortliffe and Fagan HPP-82-3 Our group recently surveyed several hundred physicians and found that they cited explanation as a principal requirement for a clinically acceptable computer- based consultation? In fact, most physicians felt that explanation is more important than absolute accuracy since they felt the latter is unrealistic. If a program can explain the basis for its conclusions, users can examine the reasoning carefully and decide for themselves whether to follow that advice. We believe this is important for an expert system in medicine; consultation programs are tools for use by individuals who are highly trained and unwilling to turn over their decisions to a computer program, regardless of how well it has been validated. One of the great advantages of the rules used in MYCIN is the development of mechanisms to explain and justify system #performance. These capabilities also contribute greatly to MYCIN’s educational role.13 _ Since all questions asked by MYCIN are generated by a rule that is under consideration, and since all rules can be displayed in English, a rule is easily understood by a physician who wants to know why MYCIN has asked a particular question during a consultation. In addition, we have developed simple techniques for understanding free text questions entered by a physician who is obtaining a consultation.*’ MYCIN can thereby analyze a question and recover the rules used to make specific decisions during the interaction. Despite the power that rules provided for facilitating explanations in an expert system, there are still serious limitations to the MYCIN approach. Many of the problems with explanation parallel those with knowledge acquisition. For example, an expert whose thoughts jump directly from data to therapy provides little structure upon which an explanation can be based. This tendency to skip intermediate concepts prevents the -exposition of general principles of action. Clancey has discussed some of these issues in his thesis,13 and we are continuing research in this area? Validation and Evaluation As expert systems have begun to mature, techniques have become necessary to demonstrate that they can perform as an expert in the field. MYCIN has been evaluated now in three large studies and we have discovered that the design of validation experiments is itself an area of research interest; -9. Shortliffe and Fagan HPP-82-3 a number of specific points are pertinent. First, any evaluation is difficult because there is so much difference of opinion in medicine, even among experts. Hence, it is unclear how to select a standard by which to measure the system’s performance. Actual clinical outcome cannot be used because each patient is treated in only one way and because a poor outcome in a gravely ill patient cannot necessarily be ascribed to poor selection of therapy. Second, although MYCIN performed at or near expert level in almost all cases, the evaluating experts in an early study had serious reservations about the clinical utility of the program. It is difficult to assess how much of this opinion is due to inadequacies in the knowledge or design of the system and how much to bias against any computer-based consultation aid. In a subsequent study, we attempted to eliminate this bias from the study by having the evaluators unaware of which recommendations were from MYCIN and which from physicians. 23 In that setting, MYCIN’s recommendations were preferred uniformly or judged equivalent to those of five infectious disease experts. Eventually, other questions must be answered regarding MYCIN and systems like it. Each question requires its own evaluation. Is the system used ? If so, do the users follow the system’s advice? If so, does the user benefit from the encounter with the system? Will the system be cost- - effective? What are the legal implications in the use of, or failure to use, such systems? The answers to these questions are years away for most consultation systems but ultimately are just as important as whether the methodology leads to accurate and reliable advice. Generalization We mentioned earlier the importance of separating the knowledge in an expert system from the program that processes that knowledge and generates the advice. One reason for this separation is the ease with which information can be added or corrected. Many expert systems include “editors” that allow a knowledge engineer to modify the knowledge base without having to change the program -10. Shortliffe and Fagan in any way. HPP-82-3 There is a second advantage. One can imagine the development of interchangeable knowledge bases, each driven by the same program. Each knowledge base would have to be structured in accordance with the conventions required by the program, but once this were accomplished, a single program could generate advice in a number of different domains. This area of active research is known as “generalization” or the development of “system building tools.“24* 25 The key idea is to develop a general purpose program that can be used by knowledge engineers to build an expert system in a new domain. Such programs must define conventions for the representation of knowledge and for the inferential models. The builders of systems must in turn comply with these conventions. Consider, for example, the system building tool that we have constructed from MYCIN. By removing all knowledge of infectious diseases, i.e., all the rules, we were left with a set of programs that we call “Essential MYCIN” or EMYCIN.25 EMYCIN can in turn be used to build an expert system in a new domain so long as it is natural to structure the knowledge in terms of production rules. Additional code was added to EMYCIN to allow the system designer to produce a knowledge base quickly and accurately. Several consultation programs have been developed using EMYCIN, including consultants for other medical problems and a consultant for structural engineering design. An Expert System For Medical Monitoring The Ventilator Manager (VM) program is an experiment in expert system development that 1 builds on our experience with production rules in the MYCIN system. VM is designed to interpret on- line quantitative data in the intensive care unit (ICU). These data are used to manage post-operative mechanical ventilatory assistance. VM was strongly influenced by the MYCIN architecture outlined earlier, but the program was redesigned to allow for the description of events that change over time. -11. Shortliffe and Fagan HPP-82-3 VM is an extension of a physiologic monitoring system,** and is designed to perform five specialized tasks in the ICU: 1. to detect possible errors in measurement; 2. to recognize untoward events in the patient/machine system and suggest corrective action, 3. to summarize the patient’s physiologic status, 4. to suggest adjustments to therapy based on the patient’s status over time and long-term therapeutic goals, and 5. to maintain a set of case-specific expectations and goals for future evaluation by the program. The program interprets physiologic measurements over time and uses a model of intensive care therapies and clinical knowledge about the diagnostic implications of data. Interpretation of Dynamic Data . Most medical decision making programs, including MYCIN, have based their advice on the data available at one particular time. In actual practice, the clinician receives additional information from tests and observations over time and reevaluates the diagnosis and prognosis of the patient. Both the progression of the disease and the response to previous therapy are .important for assessing the patient’s situation. . Data are collected in different therapeutic situations, or “contexts.” In order to interpret the data properly, VM includes a model of the stages that a patient follows from ICU admission through the end of the critical monitoring phase. The correct interpretation of physiologic measurements depends on knowing which stage the patient is in. The goals for intensive care are also stated in terms of these clinical contexts. The program maintains descriptions of the current and ,optimal ventilator-y therapies for any given time. . . VM was developed as a collaborative research project between Stanford University and Pacific Medical Center (PMC) in San Francisco. It was tested with patient information acquired from a physiolo ic monitoring system implemented in the cardiac surgery ICU at PMC and developed by Dr. John Osborn and his colleagues. 96 -12. Shortliffe and Fagan HPP-82-3 I F : R e l a t i o n s a b o u t o n e o r m o r e p a r a m e t e r s h o l d THEN: 1 ) M a k e a c o n c l u s i o n . b a s e d o n t h e s e f a c t s : 2 ) M a k e a p p r o p r i a t e s u g g e s t i o n s t o . c l i n i c i a n s ; a n d 3 ) C r e a t e n e w e x p e c t a t i o n s a b o u t t h e f u t u r e v a l u e s o f p a r a m e t e r s . Figure 2: The structure of a prototypical rule from the VM system is similar to the production rules used in MYCIN, with both a premise and conclusion specified. The “conclusion” of a VM rule frequently indicates actions that the system should take, e.g., suggest a change ventilator setting, in addition to inferences that can be drawn. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ................ Knowledge Representation in VM Knowledge is represented in VM by production rules similar to those used in MYCIN (Fig. 2). In addition to the premise and conclusion (or action) associated with each rule, several other parameters are included: (1) the rule’s symbolic name, (2) the rule group, e.g., rules about instrument faults, (3) the main concept (definition) of the rule and (4) all of the therapeutic states in which it makes sense. Fig. 3 shows a sample rule for determining hemodynamic stability. The VM knowledge base includes rules to support five reasoning steps that recur whenever new data from the monitoring system become available. These are: - 1. rules to characterize measured data as reasonable or spurious; 2. rules to determine therapeutic state of the patient (viz., the mode of ventilation); 3. rules to adjust expectations of future values of measured variables when the patient’s state changes; 4. rules to check physiologic status, including cardiac rate, hemodynamics, ventilation, or oxygenation; and 5. rules to check compliance with long-term therapeutic goals. Each step is associated with a collection of rules, sorted by the type of conclusions made in the action -13. Shortliffe and Fagan ’ HPP-82-3 STATUS RULE: STABLE-HEMODYNAMICS D E F I N I T I O N : Defines stable hemodynamics based o n b l o o d p r e s s u r e s a n d h e a r t r a t e s . APPLIES to patients on VOLUME, CMV, A S S I S T , T-PIECE C O M M E N T : L o o k a t m e a n a r t e r i a l p r e s s u r e f o r c h a n g e s i n b l o o d p r e s s u r e a n d s y s t o l i c b l o o d p r e s s u r e f o r m a x i m u m p r e s s u r e s . IF: HEART RATE is ACCEPTABLE PULSE RATE does NOT CHANGE by 20 beats/minute in 1 5 m i n u t e s MEAN ARTERIAL PRESSURE is ACCEPTABLE MEAN ARTERIAL PRESSURE does NOT CHANGE by 15 t o r r i n 1 5 m i n u t e s SYSTOLIC BLOOD PRESSURE is ACCEPTABLE THEN : The HEMODYNAMICS are STABLE Figure 3: In a VM interpretation of a rule, the meaning of “ACCEPTABLE” varies with the clinical context, i.e., the type of ventilatory assistance. VOLUME, CMV, ASSIST & T-PIECE refer to the types of ventilation therapies for which VM has been given rule-based knowledge. portion of the rule -- e.g., all rules that determine the validity of the data. Between steps (3) and (4) above, a special algorithm is used to compare the validated measurements with the current expectations, thus determining whether a measurement can be classified as high or low. Symbolic Measurement Ranges Most of the rules symbolically represent the values measured, and the terms “acceptable” or “ideal” characterize the appropriate ranges. The meaning of acceptable will change as the patient moves from state to state, but the statement of the relation between the physiologic measurements remains constant. The use of symbolic statements, e.g., “heart rate is acceptable,” allows the exposition of common principles of physiologic interpretation in different contexts and minimizes the number of rules needed to describe the complexity of the diagnostic situation. Shortliffe and Fagan HPP-82-3 The meaning of the symbolic range is determined by rules that establish expectations about the value of measurti data. For example, when a patient is taken off the ventilator, the upper limit of acceptability for the expired carbon dioxide measurement is raised. The actual numeric calculation of “expired PC02 high” in the premise of any rule changes when the context switches (removal from ventilatory support), but the statement of the rules remains the same.. A sample rule that creates these expectations is shown in Fig. 4. INITIALIZING RULE: INITIALIZE-CMV D E F I N I T I O N : I n i t i a l i z e e x p e c t a t i o n s f o r p a t i e n t s o n c o n t r o l l e d m a n d a t o r y v e n t i l a t i o n ( C M V ) t h e r a p y A P P L I E S to al1 patients on CMV IF ONE OF: PATIENT TRANSITIONED FROM VOLUME TO CMV PATIENT TRANSITIONED FROM ASSIST TO CMV THEN EXPECT THE FOLLOWING: . [ a c c e p t a b l e r a n g e 1. v e r y [ ideal ] v e r y l o w low min max h i g h h i g h s-m s-w --- m-w w--- -B-B MEAN PRESSURE 60 75 80 95 110 120 HEART RATE 60 110 . . EXPIRED I;CO2 22 28 30 l 35 42 50 Figure 4: This portion of an initializing rule establishes initial expectations of acceptable and ideal ranges of variables. Not . all ranges are defined for each measurement. PC02 refers to the carbon dioxide in expired air measured at the mouth by one of the on-line monitoring devices. Interpreting Rules The VM rule interpreter is based on the MYCIN interpreter. The major changes are: (1) forward- chaining (data-driven) invocation of rules as opposed to backward-chaining, (2) checks to see that information acquired in a previous time frame is still valid for making conclusions, and (3) iteration -15- Shortliffe and Fagan HPP-82-3 through appropriate parts of the rule set each time new information is available. A data-driven approach is necessary to take advantage of the small set of measurement values available in each time frame. This means that the inference works forward from the available information as opposed to -working backward from a goal and posing questions to the user when no data are available. Because of the demanding nature of the ICU, the system must acquire and interpret data with minimal staff intervention. Therefore, since some facts about patients will not be known to the system, caveats are attached to the suggestions it prints out. Each of the five groups of rules (corresponding to the five reasoning steps mentioned above) is considered in order. Each rule is examined to determine whether it applies in the current context. For example, rules designed to evaluate “T-piece” breathing status are not examined when the T- piece is not in place. The premise of the rule is examined for validity, and the conclusions are recorded by the program along with expectations on the future ranges of measurement values. Suggestions to clinicians are also printed out. Often the examination of the premise requires the use of a value acquired earlier in time, e.g., the temperature, which is volunteered to the monitoring system at intervals. The reliability of the stored value is determined by evaluating either a time constant (for variables that predictably change over time) or a rule (for cases in which the assessment of a value’s reliability is dependent upon - context-specific information). Associated with each parameter in the system is a specific mechanism for determining its reliability over time. If a measurement is concluded to be spurious or outdated, then it-is treated as if it were unknown, requiring alternate methods for determining the status of the patient. Rule invocation is repeated each time that a new set of measurements is available (currently every 2 to 10 minutes). Identical conclusions made in contiguous time frames are represented by keeping track of the interval specified by the times of the first and last assertion. A list of these intervals summarizes the history of a particular conclusion. The evaluation of a premise such as “Patient hyperventilating for a -16- Shortliffe and Fagan ’ HPP-82-3 30-minute period within the last hour” is made by direct examination of the intervals stored along with conclusions, as opposed to looking at the original measurements. Expectations are associated with the appropriate measurement and are classified by duration and type, e.g., the upper limit of the * acceptable range. Expectations can persist for a fixed interval, e.g., “for twenty minutes starting in ten minutes,” or for the duration of one or more clinical situations, e.g., while the patient is on the ventilator. Comparison of Mycin and VM Design Gods MYCIN was designed to serve on the ward as an expert consultant for antimicrobial therapy selection. A typical interaction might take place after the patient has been diagnosed and preliminary cultures drawn but little microbiological data are available. In critical situations, a tentative decision about therapy must-often be made pending actual culture results. In return for assistance in making this decision, the clinician is asked to spend the small amount of time required to seek a consultation. The intensive care unit is quite different, however. Continuous surveillance and evaluation of the patient’s status is required. The problem is one of making therapeutic adjustments over a long period of time, many of which are minor, such as adjusting the respiratory rate on the ventilator. The main reasons for using VM are to monitor status or to investigate an unusual event. The program must therefore be able to interpret measurements with minimal human participation. When an - interaction does take place, e.g., when an unexpected event is noted by the program, it must be terse and concise. This difference in the timing and style of the man/machine interaction has considerable impact on system design. For example, the VM system must be able to do the following: 1. to reach effective decisions on the presumption that input from a clinician will be brief, 2. to use historical data to determine a clinical situation, 3. to provide advice at any point in the hospital course of the patient, 4. to follow-up on the outcomes of previous therapeutic decisions, and -17- Shortliffe and Fagan HPP-82-3 5. to summarize conclusions made’ over time. W’s environment thus differs from MYCIN’s in that natural language is an unlikely mode of communication. A consultation program should also be able to model the changing medical environment so that the program can interpret the available data in context. Areas like infectious disease require an assessment of clinical problems in a variety of changing clinical situations, e.g., “patients after positive or negative culture data are available,” “patients who are severely ill but lack culture results,” ” patients after partial or complete therapy,” or “patients with acquired superinfection.” It is also necessary for VM to contain knowledge that can be used to evaluate its therapeutic advice, just as a human consultant follows a case over a period of time. This is complicated by the fact that the user of the system may not follow the therap,y recommended. If the patient does not react as expected to the given therapy, then the program has to determine what alternate therapeutic steps may be required. During the implementation of the VM program, we observed many types of clinical behavior that represent a challenge to symbolic modeling. One such behavior is the unwillingness of clinicians to change therapies frequently. After a patient meets the criteria for switching from therapy A to ‘B, e.g., IMV to T-piece, clinicians tend to allow the patient’s status to drop below optimal criteria before returning to therapy A. This was represented in the knowledge base by pairs of therapy selection rules (A to B, B to A) with a grey zone between the two criteria, e.g., “acceptable” limits might be used to suggest going from therapy A to therapy B, whereas “very high” or “very low” limits would be used for going from B to A. If the same limit were used for going in each direction, a small fluctuation of one measurement near a cut-off value would provide very erratic therapy suggestions. A more robust approach would be to make decisions in such situations based upon how long the patient has been in the given state and in accordance with the previous therapy or therapies. -18- Shottliffe and Fagan HPP-82-3 A more detailed disc&ion of VM is included in a thesis based on the work,14 and, in another paper, we have described in greater detail the contrasts between MYCIN and this approach to monitoring.” Advantages of Al for Medical Monitoring Symbolic Models Provide a Context for Data Interpretation The MYCIN and VM systems are experiments in how best to build symbolic models for environments in which clinical data are gathered. Both models were designed to provide a context in which numeric data could be interpreted flexibly. The model of “ventilator-y mode” that we developed for VM is used to guide the system’s interpretation of parameters such as heart rate and mean arterial pressure. Similarly, in MYCIN, the clinical context is determined by physical findings and other “soft” data, e.g., MYCIN’s list of previous medications and their time of administration is used to determine how lab findings should be interpreted in the context of “partially treated meningitis.” Context is somewhat more problematic in VM because we have had to assume that only a small amount of current non-numeric information will be available to describe a patient. The context therefore must be deduced from some of the numerical information, which is, in turn, used to interpret the remaining data in the current or in subsequent time periods. For example, certain sudden changes in airway parameters can generally be interpreted as the context “suctioning the patient” a and the interpretation of cardiac and hemodynamic parameters are then keyed into this new clinical situation. This high level model of the patient’s current state can be used to help in the analysis of trends in measurements, e.g., to provide a basis for predicting difficulties, to allow for selective tracking of “problems” as opposed to measurements, and to provide a more clinical orientation for subsequent collection of data. Symbolic Models Can Simulate the Flexibility of Expert Reasoning Symbolic processing can merge mathematic analysis with judgment or “intelligence” similar to that of an expert. For example, the VM program often throws out data because the monitor is operating outside its limits of reliability or is improperly connected. When the trend concerning this -19. Shortliffe and Fagan HPP-82-3 time period is examined, the unreliable data are ignored, much as they would be by an experienced clinician observing the same trend information. Similarly, recognitio? that the clinical situation drastically changed in the immediate past can be used to tag as dubious any conclusions made by an e averaging function applied to current data. VM uses production rules to determine the utility of “recent” measurements based on how quickly the clinical situation is changing. Just as the operational assumptions for the monitors can be tested, so can any mathematical or statistical procedure. Intelligent selection of a mathematical technique to analyze data typically requires expert knowledge of statistics as well as of the clinical domain. The use of such knowledge when picking and applying statistical tests for the analysis of medical data bases has recently been examined by Blum? Symbolic Models Facilitate Statements Of Expectations The ultimate goal of computer-based monitoring systems is to predict problems before they happen rather than to “alarm” as difficulties arise. Providing a context is the first step towards reaching that goal. VM has a limited mechanism for adjusting the acceptable limits for one or more variables based on recent events interpreted by the program. Other Al programs also address this problem, e.g., programs for military signal analysis and speech understanding can set up “future expectations” and then perform some analysis when future expectations are not realized. Similarly, air traffic control programs can plan ahead in time, can detect conflicts, and can plan alternate traffic routes. This type of computational capability would be especially useful in determining whether the patient’s response to therapy is acceptable. Symbolic Models Provide an Ability to Select Relevant Data For Analysis Al allows for a selective evaluation of the measured data. Many approaches to the interpretation of data have concentrated on selecting the “most important variable” or on creating a “figure of merit index” that tries to summarize difficulties into one specific number. Our approach to representing clinical knowledge allows the physician to specify a variety of relations between variables, or between variables and other available information. As always, using knowledge about -20. Shortliffe and Fagan HPP-82-3 the current clinical context can be used to guide the evaluation of rules and, thereby, to draw attention to specific relations only when they are important. A single index can be replaced with specific subtopics, e.g., “hemodynamic status,*’ and with their relations to therapeutic goals, e.g., “the hemodynamics have not been stable long enough to suggest removing ventilatory support.” We wish to turn a large amount of data into useful diagnostic conclusions supported by logically referenced specific numerical data. Summary Several years of experience with MYCIN have led to an understanding of additional requirements for symbolic processing approaches to medical decision making. These include extending the knowledge base beyond the facts necessary for expert performance, providing a structure for a large number of production rules, and extending the inferential aids to include assistance throughout the patient’s clinical course. For decision aids in the intensive care unit, or in other equally dynamic situations, programs cannot depend on interaction with the clinical users. Furthermore, they must handle data that are changing over time and may be missing or spurious. They must also be able to track the patient’s status during the course of disease or in response to therapy. The VM program has suggested that knowledge engineering techniques, such as those developed for MYCIN, can be adapted to dynamic clinical settings such as monitoring patients in the in tensive care unit. -21* Shortliffe and Fagan HPP-82-3 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. l-l. 12. 13. 14. 15. References Buchanan, B.G., “Research on expert systems,” Report HPP-81-1, Heuristic Programming Project, Stanford University, February 1981, To appear in Machine intelligence 70, (D. Michie, ed.), 1962 Winston, P.H., Artificial Intelligence, Addison-Wesley, Reading, 1977. Shortliffe E.H. Buchanan B.G. Feigenbaum E.A., “Knowledge engineering for medical decision making: a review of computer-based clinical decision aids,” Proceedings of the I.EEE, I 1979, pp. 1207- 1224. Feigenbaum, E.A., “The art of artificial intelligence: themes and case studies in knowledge engineering,” Proceedings of the 5th International Joint Conference on Artificial Intelligence, Cambridge, Massachusetts, 1977, pp. 1014- 1029. Michie, D., “Knowledge engineering,” Cybernetics, Vol. 2, 1973, pp. 197-200. Shortliffe, E.H., “Consultation systems for physicians: the role of artificial intelligence techniques,” Proceedings Third National Conference of the Canadian Society for Computational Studies of Intelligence, Canadian Society for Computational Studies of Intelligence, Victoria, British Columbia, May 1980, Also in Readings in ArtificiaI intelligence (B. Webber and N. Nilsson, eds.) Menlo Park: Tioga Publishing Co., 1981. Shortliffe, E.H., Computer-based medical consultations: MYCIN, Elsevier/North Holland, New York, 1976. Buchanan B.G. Feigenbaum E.A., “ D E N D R A L a n d Meta-DENDRAL: their applications dimension,” Artificial Intelligence, Vol. 11, 1978, pp. 5-24. . Davis R. King J., “An overview of production systems,” in Machine Representation of Knowledge, E.W. Elcock and D. Michie, eds., Wiley and Sons, New York, 1976. Davis R. Buchanan B.G. Shortliffe E.H., “Production rules as a representation for a knowledge-based consultation program,” Artificial Intelligence, Vol. 8, 1977”, pp. 15-45. Shortliffe E.H. Buchanan B.G., “A model of inexact reasoning in medicine,” Math. Biosci., Vol. 23, 1975, pp. 351-379. Clancey, W.J., “Tutoring rules for guiding a case method dialogue,” Int. J. Man-Machine Studies, Vol. 11, 1979, pp. 2549. Clancey, W.J., “Transfer of rule-based expertise through a tutorial dialogue,” Memo STAN- CS-79-769, Computer Science Department, Stanford University, September 1979, Doctoral D i s s e r t a t i o n Fagan, L.M., VM: Representing time-dependent relations in a clinica/ setting, PhD dissertation, Heuristic Programming Project, Stanford University, 1980. Shortliffe E.H. Scott A.C. Bischoff M.B. et al., “ONCOCIN: An expert system for oncology protocol management,” Proceedings of the 7th International Joint Conference on Artificial l -22- Shortliffe and Fagan . HPP-82-3 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. Intelligence, IJCAI, Vancouver, British Columbia, August 1981. Davis, R., “Applications of meta-level knowledge to the construction, maintenance, and use of large knowledge bases,” Memo HPP-76-7, Heuristic Programming Project, Stanford University, Julv 1976, Doctoral DissertationI- Nilsson, N.J., Problem solving methods in artificia/ intelligence, McGraw-Hill Book Company, Inc., San Francisco, 1971. Duda R.O. Gaschnig J. Hart P.E. et al., “Development of the PROSPECTOR consultation system for mineral exploration,” Final report SRI Projects 5821 and 6415, SRI International, 1978. Elstein A.S. Shulman L.S. and Sprafka S.A., Medical problem solving: an analysis of clinical reasoning, Harvard University Press, Cambridge, 1978. Teach R.L. Shortliffe E.H., “An analysis of physician attitudes regarding computer-based consultation systems,” Comput. Biomed. Res., Vol. 14, 1980, pp. 542-558. Scott A.C. Clancey W.J. Davis R. Shortliffe E.H., “Explanation capabilities of knowledge-based production systems, ” Amer. J. Computational Linguistics, 1977, . Wallis J.W. Shortliffe E.H., “Explanatory power for medical expert systems: studies in the representation of causal relationships for clinical consultations”, Submitted for publication, December 1981. Yu V.L. Fagan L.M. Wraith S.M. et al., “Antimicrobial selection by a computer: a blinded evaluation by infectious disease experts,” JA MA, Vol. 242, 1979, pp. 1279-l 282. Nii H.P. Aiello N., “AGE (Attempt to Generalize): A knowledge-based program for building knowledge-based programs,” Proceedings of the 6th International Joint Conference on Artificial Intelligence, Tokyo, Japan, August 1979, pp. 645-655. van Melle, W., “A domain-independent system that aids in constructing knowledge-based consultation programs,” M e m o HPP-80-11, Heuristic Programming Project, Stanford University, June 1980. Osborn J. J. Beaumont J.O. Raison J.C.A. Abbott R.P., “Computation for quantitative on-line measurement in an intensive care ward,” in Computers in Biomedical Research, R.W. Stacey and B.D. Waxman, eds., Academic Press, New York, 1969, pp. 207-237. Fagan L.M. Shortliffe E.H. Buchanan B.G., “Computer-based medical decision making: from MYCIN to VM,” Automedica, Vol. 3, 1980, pp. 97-l 06. Blum, R.L., “Discovery, confirmation, and incorporation of causal relationships from a large time-oriented clinical database: the RX project,” in Computers in Biomedical Research, H.R. Warner, ed., Academic Press, New York, 1982, pp. 164-187. -23. 
work_3xoq4ohu2nczje4o7o4w7z2gcu	----	PDF hosted at the Radboud Repository of the Radboud University Nijmegen The following full text is a publisher's version. For additional information about this publication click this link. http://hdl.handle.net/2066/112336 Please be advised that this information was generated on 2021-04-06 and may be subject to change. http://hdl.handle.net/2066/112336 Journal of Chromatography, 550 (199 1) 257-266 Elsevier Science Publishers B.V., Amsterdam CHROMSYMP. 2094 Expert system for repeatability testing of high-performance liquid chromatographic methods M. MULHOLLAND*’ and N. WALKER Philips Scient&, Cambridge (UK) F. MARIS and H. HINDRIKS Organon International, Oss (Netherlands) L. BUYDENS University of Nijmegen. Nijmegen (Netherlands) T. BLAFFERT Philips Research Laboratories, Hamburg (Germany) and P. J. SCHOENMAKERS Philips Research Laboratories, Eindhoven (NetherIan&) ABSTRACT A repeatability test is performed as part of the validation of the precision of a high-performance liquid chromatographic method. The purpose of the test is to establish the random errors of the method with respect to various method features. It is usual to test the repeatability of the sample preparation and the injection procedure by a number of repetitions. The number of repetitions depends on the application of the method. For instance, a quality control method is expected to analyse up to and over 50 samples in a single run. The repeatability test of the injection process therefore needs to reflect this number, in order to identify fully potential errors. These errors could be due to the instrument error or to drifting factors such as temperature. The expert system described provides the necessary expertise to identify the required tests and interpret the results. It also contains a diagnosis module which can identify sources of unacceptable errors. The diagnosis module is linked to a reoptimization process which can modify the method to improve the resolution between a critical pair of peaks. This expert system was built as part of the ESCA project, which is a 3-year project investigating the application of expert systems to analytical chemistry. A repeatability test requires both heuristic and algorithmic knowledge and therefore provides a good challenge for expert system technology. The system is implemented in a multiple windows software envi- ronment and uses workstation hardware. This software has been evaluated in practical laboratory envi- ronments and some of the results and conclusions are described. The major advantage of an expert system implementation appears to be that it can give advice when problems occur. It also ensures that the statistics are used and interpreted correctly. This is not possible using conventional algorithmic software. INTRODUCTION High-performance liquid chromatography (HPLC) is usually applied as a quantitative technique, where the relative concentrations of components in a mixture ’ Present address: University of New South Wales, Sydney, Australia. 0021-9673/91/$03.50 0 1991 Elsevier Science Publishers B.V. 258 M. MULHOLLAND et al. are determined. Each sample application requires a unique combination of chemical and instrumental conditions. The development of a method is a complex series of choices and optimizations which consider features of both the chemistry of the sam- ple and its application. When developing a method, the analyst has certain expecta- tions of its quantitative performance. These include accuracy, sensitivity, specificity and precision. To ascertain whether the new method achieves the expected levels of performance requires a series of validation experiments. Each validation test can be defined in four stages: (1) The method characteristic is defined; this could be accuracy, precision or any other performance characteristic which is important to the application. (2) The actual test is defined; for precision this could be repeatability, reproduc- ibility of ruggedness. (3) An experimental procedure with which to test these characteristics is then selected and carried out. This includes the measurement of various method character- istics. (4) The final stage in a method validation test is to diagnose the results. This requires pass/fail criteria and a method for identifying and curing problems. It may sometimes require the reoptimization of the method if it fails to meet the specified performance levels. HPLC method validation is thus an integral part of process of method devel- opment. Links back to method optimization are often required. It is also necessary to be aware of the performance requirements during the method development process. For instance, a method developed on a delicate microbore column would not be suitable for multiple usage throughout busy quality control laboratories. This paper describes an expert system built to tackle these problems and create these links for a repeatability test. The system is built as part of a research project investigating the use of expert systems in analytical chemistry, ESCA. A small expert system was initially built as a test case for the expert system development tool Goldworks. It used a combination of spreadsheets and expert sys- tem facilities such as frames and rules [ 11. However, as the work of ESCA progressed it became clear that analysts did not just need stand-alone packages but communi- cation links throughout the method development process. It was therefore decided to build a system which could perform a repeatability test and also communicate with the method optimization process. Three expert systems have been built which tackled the various stages in the method development process: first guess [2], selectivity opti- mization [3] and optimization of the instrumentation [4,5]. Links to the latter system looked most promising for the purpose of re-optimizing methods which had failed a method validation test. However, links or other systems would eventually be required in order to form a complete picture [6]. STRUCTURE OF THE EXPERT SYSTEM The structure of the system is illustrated in Fig. 1. The system starts by consult- ing the system optimization module. This module optimizes the physical parameters such as flow-rate, column dimensions and detector flow cell. The aim of the optimiza- tion is to provide the fastest analysis time within the required resolution. In the stand-alone system the optimization limits consist of the hard physical limits on the EXPERT SYSTEM FOR TESTING OF HPLC METHODS 259 REPEATABILITY METHOD FEATURE Sample Preparation TEST DESIGN pj Single Point ( Concentration Range TEST MEASURE DIAGNOSIS Fig. 1. Overview of the repeatability test expert system. instrument capability, the available sample volume, with user-defined ranges for the resolution and signal-to-noise ratio of the analysis. The role of these limits changes when they are applied to optimizing a method which has predefined constraints on its repeatability. The limits need to specify ranges within which an optimized method should be repeatable. Some of these limits are shown in Table I. They are divided into three categories: the possible range, the reliable range and the robust range. The range is selected by examining the method application. Some examples are shown in Table I. When the required range is specified the system optimizes for the fastest analysis time. The results include the following information: (1) the selected column dimen- sions; (2) the selected detector. flow cell; (3) the value of the time constant; (4) the flow-rate and pressure drop of the instrument; (5) the projected analysis time, resolu- tion and signal-to-noise ratio for the analysis; and (6) the selected values for the TABLE I RANGE REQUIREMENTS FOR THE OPTIMIZATION Parameter Possible range Reliable range Robust range Flow-rate 0.01-10 ml/min 0.1-5 ml/min 0.3-3 ml/min Pressure 0400 bar S-250 bar 5200 bar Resolution 1 2 3 Signal-to-noise ratio 10 100 200 Column I.D. 0.25-25 mm 2-8mm 4-8 mm Examples: Sample No. <25 25 25 Lab. No. 1 >l >2 a The range of values within which the method is likely to be robust, i.e., not affected much by small changes. 260 M. MULHOLLAND et al. TABLE II AN EXAMPLE OF A PEAK TABLE Peak Retention time Peak height Peak area Asymmetry Plate count Relevant: No. (s) (mm) yes/no 19 91 145 180 20 70 100 20 1000 1000 1000 1000 23 000 yes 23 900 yes 21 700 yes 24 900 yes sample concentration and injection volume. The user can interact with these results by selecting from several options. When one is satisfied with the optimization, the repeatability test can begin. The method features to be tested are then defined as the sample preparation and the injection procedure. The analyst inputs a description of the HPLC method to- gether with information on the expected usage. With this information an experi- mental design can be recommended. Originally, the system recommended a number of repetitions of the sample preparation and the injection procedure at a single con- centration point. However, during the evaluation this was found to be inadequate. Therefore, designs were added to deal with multiple concentration levels. After the design has been recommended, the user carries out the experiments and collects data for retention times, peak areas, peak heights and concentration. These data are input to the system using peak tables; an example is shown in Table II. The user is only allowed to proceed to the next stage if all the peak tables are entered and all have the same number of peaks. The test measure is now made on these data. For each peak, the relative stan- dard deviation (R.S.D.) of each value is calculated [l]. These measurements are then used to diagnose the repeatability of the method. The diagnosis is performed in several stages: (1) The concentration variations are measured against a pass/fail criterion, usually 1%. If they pass then the diagnosis ends here and the method is concluded to be repeatable. However, if it fails it then proceeds to the next stage. (2) A Grubbs test for outliers is performed and the user is allowed to remove any outliers identified. (3) The results can be viewed graphically to determine any drifting conditions. (4) Potential problems are diagnosed and a cure is suggested. TABLE III CLASSIFICATION OF R.S.D. VALUES Classification Height R.S.D. (%) Area R.S.D. (%) Retention time R.S.D. (%) Small 0 0 0 Medium 1 1 0.5 Large 2 2 1 EXPERT SYSTEM FOR TESMNG OF HPLC METHODS 261 The R.S.D. values are used to identify possible problems. The R.S.D.s are first classified as small, medium or large according to Table III. The combined behaviour of these values can indicate specific problems. For instance, if multiple injections of the same sample gives rise to large variations in peak height and area, but not in the retention times, then this could be due to a variation in injection volume or possibly degradation of the sample. The diagnosis section of the repeatability test gives a list of possible problems with the method. It does this in four stages. (1) The classifications of small, medium and large are used to create lists of possible problems for each peak. Each problem within a list has a priority or weight. A value of zero means that the problem is not possible. Two lists are formed, one for sample preparation problems and the other for injection problems. (2) The individual lists of problems are combined into two overall lists by summing the priorities of each problem over the peaks. (3) The list of problems are then checked. Certain problems, although diag- nosed, can be discounted if the chromatographic method contains the correct fea- tures. For instance, if pH variation is diagnosed but the method is utilizing a buffer then this problem can be discounted. (4) The list of problems with injection and sample preparation are combined into one complete ordered list. The list of problems with injection is considered to be more important than those for sample preparation, so are given a higher priority. If a problem occurs in both lists then the higher of the two priorites is taken. Once the problem list has been established the system goes on to recommend corrective ac- tions. Each possible problem diagnosed has a list of actions that can be taken to try to solve the problem. These actions can be divided into two categories: actions that involve changing the method, such as the re-optimization of resolution, and actions that involve checking or maintaining the chromatograph. The expert system provides suggested actions to the user based on their priority score and checking and mainte- nance cures are always performed first. If a certain action does not solve the problem, this is measured by performing a reduced test, then the next highest priority is sug- gested. If the method cannot perform after any recommended checking or mainte- nance cures then some reoptimization is suggested. If the resolution between a pair of peaks falls below a predefined level throughout repeatability test, then the user is routed to the system optimization module. This causes the resolution to be increased to a larger value. Fig. 2 shows a graphical representation of the diagnosis process. The software is implemented using Pascal in a multiple windows environment. A fuller description of the software and hardware environment is given elsewhere. This system was evaluated using several pharmaceutical examples. The results for the analysis of ethinylestradiol are presented here. EXPERIMENTAL Sample preparation Samples are prepared by adding a formulated tablet containing ethinylestradiol to 2.0 ml of an internal standard solution of estradiol. This is then shaken and sonicated until the tablet is completely dispersed. The mixture is centrifuged and 10 ~1 of the supematant are injected onto the column. 262 M. MULHOLLAND et al. TEST MEASURE , RSD - CONCENTRATION N.. - RSD - RETENTION HEIGHTS L-.---J AREAS RE-OPTIMISE system Optimization . DRIFT DRIFT DIAGNOSIS I t I I LDIAGNOSIS FIX PROBLEM I I Fig. 2. Overview of the measurement and diagnosis modules l olvsrt 1 2 3 4 additlvm 1 2 3 buffmr PH l olvent mix umcb Ymm Ymm no no umsd no no no ummd no controlled no in-line Yma nmm pmrcmntagm FlcN 49 water 61 namm wmlght namm value cone Fig. 3. Chromatogram of ethinylestradiol (latest eluting peak) with estradiol as the internal standard. EXPERT SYSTEM FOR TESTING OF HPLC METHODS 263 Chromatographic conditions The sample is eluted from a 100 x 5 mm I.D. Novapak Cls column with a mobile phase of acetonitrile-methanol-phosphoric acid (60:40:0.05) at a flow-rate of 2 ml/min. The column temperature is ambient and the sample is detecting using a UV detector at 205 nm. RESULTS Fig. 3 shows as an example a chromatogram of ethinylestradiol together with estradiol. This method is input to the system using a number of tables reserved for this information, an example of the solvent description is shown in Fig. 4. The recom- mended experimental design involves a total of 98 injections. This includes the in- jection of standards to calibrate the instrument. Fig. 5 shows the conclusions obtained from the diagnosis of these results. The calculated values for the R.S.D. are shown for each of the sample preparation tests and the injection procedure test. For the injection procedure test no problems are found and all the R.S.D.s have values below the medium classification shown in Table III. However, for the sample preparation a small problem is observed in the variation of peak heights and areas for peak 2. The problem is diagnosed as due to either sample degradation or inadequate sample preparation. For this example the sample degradation is considered the priority problem. The suggested action is to ltt.ua 1 i RETENTION TIME (HINUTESJ Fig. 4. Example input window for the solvent description. A d ic yn oa ia o f th e ex p er im en ta l rs a u lt a sh ow th e fo ll ow in g p os si b le p ro b le m s w it h th e m et h od p ro b le m a m p l e in js ot io n p ro b le m a vi th 8 a m p le d eg ra d a ti on 0 .9 0 m a n p l e p e a k ra ts n ti o n ti re 0 0 1 0 I 1 6 6 0 2 2 0 .2 1 0 4 4 in je ct p ea k re te n ti o n ti re 0 0 I 0 0 2 4 2 6 9 ta m t p tm k h e ig h t p sa k a m a 0 0 O s4 1 2 1 7 0 I 6 6 3 7 4 1 I 1 6 7 0 7 i # 37 4 4 7 te m t p e a k h e ig h t p r 0 0 I 4 0 7 1 3 2 0 .1 1 2 7 7 0 .6 4 6 6 6 F ig . 5 . S cr ee n s re su lt in g fr om t h e d ia gn os is o f th e m et h od fo r et h in yl es tr a d io l. re d a vu lo p 8 a m p lo p re p a ra ti o n T h e sy a ta m su gg es ts th a t y ou re d ev el op th e sa m p le p re p a ra ti on , a n d th en op ti m iz e th e in st ru m en t co n d it io n s. T h is a ct io n h a 8 b e e n c h o se n to a tt e m p t to so lv e th e p ro b le m of E a c h o f th e p ro b le m s h a e a n u m b e r o f p os si b le co rr sc ti ve a & io n s a ss oc ia te d \ :i th it . T h e a c ti o n s, a rr a n g e d in o rd e r o f p re fe re n c e , a re li st e d h e re . T h e c h o se n a c ti o n is d es cr ib ed m or e fu ll y -e d ev a lo p sa m p le p re p a ra ti on II EXPERT SYSTEM FOR TESTING OF HPLC METHODS 265 The suggestion of a corrective action to do is made in the folloving way. 1. For each diagnosed problem, a list of poasibla actiona is generatad. 2. The lists of actions are combined and prioritised according to which subsystem the, action8 are to be done in. The order is User actions - highe8t System optimize actiono Edit method actions 3. Any actiona that have already been performed are removed from the list. Any action8 that are not deemed necssaary are also removed. No actions were rejected aa having been performed The following actiona were rejected on the grounda of not being needed Increame rsaolution limit. re-optlmiaei rejected because resolution was > 1.6 Fig. 6. Explanation provided for the suggested corrective actions. re-develop the sample preparation. Fig. 6 shows the explanation available for this corrective action. It is interesting that the evaluators felt that the measured performance is accept- able for this method, but were pleased to find the system suggesting actions that could further improve the method. This is a typical advantage of expert systems over con- ventional software packages. CONCLUSIONS The use of expert system technology to develop a repeatability test package is considered successful. The approach allows the use of both algorithmic and heuristic knowledge. Algorithmic knowledge, such as that required for R.S.D. calculation, can be implemented in a conventional equation, whereas heuristic knowlegde, such as that required for problem diagnosis, can be implemented as rules. The evaluation of this system proved extremely valuable as it resulted in several additions which enhanced the software considerably. The software was shown to give good advice and even suggested improvements to methods that were previously con- sidered acceptable. However, several additions could still be made to complete the repeatability test 266 M. MULHOLLAND et al. system. It could be integrated with a chromatography data station to improve the efficiency of data transfer within the software. To complete the picture the system would need to be linked with the remainder of the method development process. This would allow any problems identified to be cured by re-developing the method. REFERENCES 1 M. Mulholland, J. A. van Leeuwen and B. G. M. Vandeginste, Anal. Chim. Acta, 223 (1989) 183-192. 2 H. Hindriks, F. Maris, J. Vink, A. Pee&s, M. de Smet, D. L. Massart and L. Buydens, J. Chromatogr., 485 (1989) 255-265. 3 A. Peeters, L. Buydens, D. L. Massart and P. J. Schoemnakers, Chromatographia, 26 (1988) 101-109. 4 P. J. Schoenmakers, N. Dunand, A. Clelland, G. Musch and T. Blaffert, Chromatographia, 26 (1988) 374. 5 P. J. Schoenmakers and N. Dunand, J. Chromatogr., 485 (1989) 219-236. 6 P. Conti, H. Pityns, N. Vanden Driessche, M. de Smet, T. Hamoir, F. Maris, H. Hindriks, P. J. Schoen- makers and D. L. Massart, Chromatographia, submitted for publication. 
work_3xox3v7zbbak3ezlahwjewhqqi	----	Microsoft Word - expert_system.doc ASPECTS OF AN EXPERT SYSTEM FOR ON-LINE EOLIAN SITES DESIGN Octavian Căpăţână, Rareş Cazan, Mihaela Drăgan IPA - R&D Institute for Automation - Cluj Subsidiary Abstract: Our aim is to present the necessity and the objectives taken into account in the development of an ON LINE expert system, with free access, for the eolian sites designing process. The expert system is based on a set of rules and metric relations, and also on a Romanian patent. By eolian site design, we assume: 1. approximation of average and annual eolian potential, using terrain data, and 2. offering a blueprint for eolian plant project. This blueprint contains: potential technical solution, recommended installed power, optimal pillar height, approximation of capital costs, data about the providers for eolian plants, etc. Copyright © 2007 IFAC Keywords: Eolian energy, expert system, assisted design. 1. MOTIVATION 1.1 Eolian energy It is estimated that from the overall solar energy at the level of our planet, a percentage between 1 and 3 is transformed in wind, and a percentage between 0.01 and 0.06 is transformed in biomass (by photosynthesis). Production costs for one MWh, in USA, in 2006, produced based on wind, coal, or natural gas are approximately even, in case of production costs comprising all costs including investment absorption, loan interest and risk costs. In USA, these costs vary between 52 and 55 $/MWh (EIA). A comparison made in between costs for different clean energies, leads to the conclusion that among the future solutions, the eolian energy must be taken into consideration (see table 1). Table 1 Comparison in between costs for different clean energies Data source for table 1 BHA – British hydropower association BC Hydro – British Columbia – Canada BWEA – British wind energy association Kapa System Antena – Geothermal comparison Other (Ordean, et al., 2006) (*) A correction has to be done, if the cost for 1 installed W is 1.6 $, in eolian industry; we must emphasize the fact that the real cost for 1 W disponibile for utility, is in fact $5 $6.1 ≅ cF . Energy type Effic % Usage factor Cost./ invest. $/W Life time Cost prod/ KWh cUSD μhydro 70 >55% 4 50 4-9 hydro 90 2 wind <30% 1,6* 20-25 6-12 solar 15 6-24 15-25 60-200 geothrm 70 25 5-9 biogas 27 8 20 4-5 This viewpoint is confirmed by the dynamic registered by the eolian power growth: plus 25% in 2002 compared with 2001, plus 38% in 2003 compared with 2002, plus 41% in 2005 compared with 2004. 1.2 Eolian emplacement Generally speaking, eolian emplacement comprises three distinct phases: eolian region, eolian site layout and generator/ generators emplacement. The first phase is referring to the brutish emplacement, in a geographical area, limited by the coastal and hilly zones. Currently, there are wind maps (in Romania, in 1992, ICEMENERG has created a map of this kind, representing eolian potential), more or less general (ICEMENERG).These maps aren’t giving any other information beside the following rule: RULE 1: Wind generators are practical in the coastal and hilly regions. In its metric form, this rule suggests that wind generators are to be taken into consideration where the average wind speed is at least 4,5m/s (16km/s). Consequently, we are considering the eolian potential maps almost useless. The second phase consists in delimiting an area in which the eolian potential justifies an investment, or, inversely considered, in area X (property, chartered) an investment is worth it. The question being raised is if the eolian potential justifies an investment of Z funds in the considered area? The third phase is referring to the final emplacement, meaning that we already know the exact area (several ha) and we want to establish the foundation of the eolian generator pillar. This phase is conducted, usually, by measurements. A single measurement consists of data acquisition (intensity, direction associated with time and calendar) conducted in a period of at least one year. Such a measurement, by means of few data, may cost up to 100.000 euro. A measurement conducted in an area, does not imply the fact that the measurement was taken in the optimal point, and for an investment raising up to several millions of euro (1.6 euro/installed W) it must be found the optimal point. Regarding the reduction of the measurement points’ number or regarding the eolian potential assessment in a given area based on a measurement point, it exists at OSIM, a licence request - A00302/3 on May 2007. 2. ON-LINE ASSISTED DESIGN OF EOLIAN SITES An investor has an established area within the regions with eolian potential and would like a more precise assessment of the eolian potential inside his area; an area, with the sense given here, regardless of the relation with the site’s geography, may be estimated between 1 ha and several ha. Based on the cumulated knowledge, we develop an expert system which will assist an investor in assessing the eolian potential of an area. Fig. 1. Romanian eolian potential map elaborated by ICEMENERG As presented, the first phase is superfluous. The expert system, being on-line, in case of GSM coverage, the investor located in the desired area, may succeed the second phase (see figure 2), arriving in the third phase, the last one, in which it is established the exact eolian generator/generators pillar/pillars foundation/ foundations. 2.1 Approximation of eolian potential based on terrain data The expert system consists of an inference engine based on a set of rules and metric relations, some of the rules being listed below: RULE 2: The increase of wind speed with increasing height in a proportion of 1/7th root of altitude (wind profile power law) (Exception – the case of coastal areas). RULE 3: A wind turbine can extract at most 59% of the energy that would flow through the turbine cross section (Betz law) RULE 4: An eolian farm is more economical than several eolian isolated generators with the same installed power. RULE 5: A generator is emplaced in an area without obstacles in at least 250 m. RULE 6: The smallest distance between a residential area and an eolian turbine is 250 m. RULE 7: The turbine has to be emplaced at minimum 12 m above any obstacle located on a radius of 250 m. RULE 8: The cost for one installed W off shore is double than the cost for one installed W on shore. etc. The metric relations comprise the following ones: 1. shoreonshoreoff vv −− ⋅= 1.1 (1) 2. 35.0][ vSWP aervant ⋅⋅⋅= ρ (2) where: - Pvant – total wind power pass through swept area S - ρaer – air density [kg/m3] - S – swept area of rotor wings [m2] - v – wind velocity [m/s]. Fig. 2. Interactive ON LINE form of the solicited eolian terrain data 3. ][][ WPWP vantturbinaeol ⋅=− η (3) where: - η – turbine and gear efficiency gearturbine ηηη ⋅= (4) where: - 59.0=<turbineη - (Betz Law) - gearη - efficiency that counts the losses in the gear mechanism. 4. ][][ WPFWP turbinaeolovercmedie −⋅⋅= η (5) where: - Fc – capacity factor (<30%) - overη - efficiency that counts the losses in the short periods of time when the production has no utility load (and must be lost by dump load). Other metric relations, regarding costs assessment are, among others, the following: 5. The cost of one installed W on shore is in between 1.6$ - 2$. 6. The cost of one km of low voltage electrical network is 25.000 $. The set of rules and metric relations allow the approximation of 4 fundamental elements of an eolian project: i) cost approximation; ii) annual or multi-annual average eolian potential; iii) installed power; iv) generator’s pillar height. Generation of questions to which the potential investor should offer answers, is automatically adjusted – considering the anterior answers. In case of essential questions not being answered, the process is interrupted and the client is apprised that he should document himself previously, and that he must take on the responsibility for provided data without factual coverage. In cases in which the solicitant has the chance to be located in an area/situation in which exists, and could be indicated a general deformation grade related to the trees within the area, this exercise – ON-LINE assessment of eolian potential - may exclude, in a high proportion, direct measurements within the region. Fig. 3. Eolian site project blueprint Fig. 4. Eolian site project blueprint. The case of an individual small wind generator In cases in which the data provided to the expert system are poor or insufficient in order to apply the rules or metric relations, the next step is excluded. Based on the answers given by the solicitant, summarized in the interactive form, the inference engine based on twelve rules and seven metric sizing relations, will generate the blueprint of the eolian site project (see figure 3). In order to generate this blueprint, we used both the results from the interactive form, and, directly, the answers from the form. In certain cases, the expert system assumes a block diagram of the eolian generator and of the annexed installations (see figure 4). Also, by the blueprint of the eolian site, in some cases, it is estimated an investment cost, according as in other cases, to direct measurements being imposed in order to determine the eolian potential. 3. CONCLUSIONS In an analysis SWOT ENERO – Center for Promotion of Clean and Efficient Energy in Romania, the technical barriers are inventoried, being stated the following facts (Ţânţăreanu, 2002): “There is no precise knowledge of the wind potential since no substantial wind audit has been conducted in the country. Only a general analysis was attempted during the 1990’s. There are no specific Romanian norms and standards relating to wind energy”. The proposed expert system is intended to reduce and cover these shortcomings; moreover the ON LINE implementation method will communicate, by free access, the eolian potential of the location of each solicitant. REFERENCES BWEA report on onshore wind costs. On the web: http://www.bwea.com/pdf/briefings/target-2005- small.pdf EIA Energy Information Administration. International Energy Outlook. On the web: http://www.eia.doe.gov/oiaf/ieo/pdf/0484(2006). pdf ICEMENERG - Institutul de cercetări şi modernizări energetice. On the web: www.icemenerg.ro Ordean, M., D. Chiorean, I. Rogoz, C. Lehene, I. Stoian, E. Stâncel, (2006). SCADA Systems – Support for The Maintenance Management of Hydro Power Plants. In International Conference on Automation, Quality and Testing, Robotics, Tome 1, pp. 238-242, Cluj-Napoca. Ţânţăreanu, C. (2002) A Swot Analyses on the wind energy development in Romania On the web: http://www.enero.ro/ro/doc/Ebulletin4-01- ENERO_SWOT.pdf 
work_3yrlw3wgnjbahm23qkojfkuotu	----	Microsoft Word - _completo_ Clustering people according to their preference criteria.doc Clustering people according to their preference criteria Jorge Díez*, Juan José del Coz, Oscar Luaces, and Antonio Bahamonde Centro de Inteligencia Artificial. Universidad de Oviedo at Gijón, Campus de Viesques, E-33271 Gijón (Asturias), Spain {jdiez, juanjo, oluaces, antonio}@aic.uniovi.es Abstract. Learning preferences is a useful task in application fields such as collaborative filtering, information retrieval, adaptive assistants or analysis of sensory data provided by panels. SVMs, using preference judgments, can induce ranking functions that map objects into real numbers, in such a way that more preferable objects achieve higher values. In this paper we present a new algorithm to build clusters of people with closely related tastes, and hence people whose preference judgment sets can be merged in order to learn more reliable ranking functions. In some application fields, these clusters can be seen as market segments that demand different kinds of products. The method proposed starts representing people’s preferences in a metric space, where it is possible to define a kernel based similarity function; finally a clustering algorithm discovers significant groups with homogeneous tastes. The key point of our proposal is to use the ranking functions induced from the preference judgments of each person; we will show that those functions codify the criteria used by each person to decide her preferences. To illustrate the performance of our approach, we present two experimental cases. The first one deals with the collaborative filtering database EachMovie. The second database describes a real case of consumers of beef meat. Keywords: learning preferences, clustering, adaptive assistants, analysis of sensory data, market segmentation. * Corresponding author. Fax: +34 985 18 21 25. Phone: +34 985 18 25 88. E-mail address: jdiez@aic.uniovi.es 2 1. Introduction Supervised inductive learning deals with sets of training examples; these represent pairs of input and the attached outputs of a function that has to be found in a given family of hypotheses. The input is described by a set of attribute values, while the output is in fact another attribute of the examples called class; its type determines the approach and even the name of the learning task. Regression is used when the class is a continuous number, and categorical classification is employed when the class or output of training examples is one of a finite set of symbolic categories. In this paper we tackle a slightly different problem: learning people’s preferences for consumable products, or for system configurations, or for responding to information requests. Here the training material can be expressed as in regression problems: the description of each object is then followed by a number that assesses the degree of satisfaction. Alternatively, training examples can be represented by preference judgments: pairs of vectors (x (1) , x (2) ) where someone expresses the fact that he or she prefers x (1) to x (2) . In other words, training sets are samples of binary relations between objects. As pointed out in (Cohen et al., 1999; Dumais et al., 2003), obtaining preference information may be easier and more natural than obtaining the labels needed for a classification or regression approach. Moreover, this type of information is more accurate, since people tend to rate their preferences in a relative way, comparing objects with the other partners in the same batch. There is a kind of batch effect that often biases the ratings. Thus, an object presented in a batch surrounded by worse objects will probably obtain a higher rating than if it were presented together with better objects. There are algorithms in the literature able to learn preferences. Sometimes they aim to classify pairs of objects (x (1) , x (2) ), deciding whether x (1) is preferable to x (2) or not, as in (Branting and Broos, 1997; Cohen et al., 1999). Another approach consists in learning a real preference or ranking function f from the space of objects considered in such a way that f(x (1) ) > f(x (2) ) whenever x (1) is preferable to x (2) . This functional approach can start from a set of objects endowed with a (usually ordinal) rating, as in regression (Herbrich et al., 2000; Crammer and Singer, 2001; Shashua and Levin, 2002), or can stem from sets of preference judgments, as in (Tesauro, 1989; Utgoff and Clouse, 1991; Freund et al., 1998; Fiechter and Rogers, 2000; Joachims, 2002; Bahamonde et al. 2004). 3 In this paper we present a new algorithm that, given a family of preference judgment sets from a number of people, discovers groups with homogeneous tastes. The usefulness of this task is twofold. First, we can merge the sets of preference judgments of members of the same group, thus attaining more useful and reliable knowledge about peoples’ preferences; this is usually a critical point, since the available data from individuals is frequently quite limited. In second place, the discovery of clusters is important by itself. This is the case when we deal with data collected from panels of consumers; then the groups found may constitute significant market segments that demand different kinds of products. At the end of the paper, we will present an application to food industry. Using the data from a panel of beef meat consumers, we discuss the implications of the results returned by the algorithm for meat industries and breeders. A shorter version of this case study was presented in (Díez et al. 2005). To cluster anything it is necessary to define a similarity measure. The core idea of our proposal is that similarity can be defined between the ranking functions learned from the preference judgment set of each person; these functions codify the criteria used to decide the preferences. Since those ranking functions are induced by a classification Support Vector Machine (SVM) (Vapnik, 1998), the similarity measure is defined by means of a kernel method. In the next section, we review the literature related to explaining the usefulness of clusters of preference criteria in different areas of application. Then, we discuss the details of our proposal. Finally, the paper concludes with a report of the experiments conducted to evaluate the proposed algorithm. For this purpose we use EachMovie (McJones, 1997), a publicly available collaborative filtering database for movie ratings. Finally, we review the results obtained with the dataset of the panel of beef meat consumers. 2. How clusters can be useful The learning tasks involved in recommender systems (Resnick and Varian, 1997) can be considered as special cases of ordinal regression. Here users rate one kind of object and receive recommendations about 4 objects that they are likely to prefer. Such advice can be elaborated according to the relationship of the properties of the objects and the user’s past ratings; this is the content-based model (Basu et al., 1998; Pazzani, 1999). Or, on the other hand, in the model called collaborative or social filtering, the recommendations are induced from the user and other users’ ratings, formulating them as a learning task (Goldberg et al., 1992; Resnick et al., 1994; Shardanand and Maes, 1995). Within this context, as pointed out in (Hofmann and Puzicha, 1999), there is another fundamental problem in addition to the prediction of ratings: the discovery of meaningful groups or clusters of persons and objects able to explain the observed preferences by some smaller number of typical preference patterns. This point of view gives rise to the latent class model (Cheung et al., 2000). See also (Ungar and Foster, 1998). There are other application fields where clusters and preferences appear together as a desirable mixture. For instance, in (Joachims, 2002), Joachims presents an information retrieval system equipped with a ranking function learned from click-through data collected from user interaction with a www search engine. To improve his proposal, the author acknowledges the need to obtain feasible training data. This raises the question of the convenience of developing clustering algorithms to find homogeneous groups of users. An Adaptive Route Advisor is described in (Fiechter and Rogers, 2000); the system is able to recommend a route to lead users through a digitalized road map taking into account their preferences. An interesting extension discussed in the paper is to modify route recommendations depending on the time of the day or the purpose of the trip. The approach suggested the inclusion of an algorithm that clusters user preferences into contexts. 3. Sensory data Analysis The quality or acceptability of market products can be measured in a number of different dimensions. Sensory analysis is concerned with those aspects that are principally appreciated through sensory impressions. It is typically used by food industries and breeders to improve some production decisions. 5 An excellent survey of the use of this type of data in the food industry can be found in (Murray et al., 2001; Buck et al., 2001); for a Machine Learning perspective, see (Corney, 2002) and (Goyache et al., 2001; Del Coz et al. 2004; Luaces et al. 2004; Díez et al. 2005). From a conceptual point of view, sensory data include the assessment of products provided by two different kinds of panels. The first one is made up of a small group of expert, trained judges; these will describe each product by attribute-value pairs. Expert panelists are thus required to have enough sensory accuracy so as to discriminate between different and similar products; note that experts are not necessarily asked to rate the overall quality of products. This panel will play the role of a bundle of sophisticated sensors. So, experts’ assessments are used to describe the products in addition to other measures usually obtained by some chemical, biological or physical devices. The second kind of panel is made up of untrained consumers; these are asked to rate their degree of acceptance of the tested products on a scale. The aim of the analysis of these data is to be able to relate product descriptions with consumer preferences. Therefore, sensory studies of markets will start out from tables such as Table 1. Each row represents a product rated by a consumer in a given tasting session. The concept of session is important, since we can interpret consumer ratings relative to each session (Joachims, 2002; Bahamonde et al. 2004; Díez et al. 2004). In this way, we do not need to assume that a rating of “7” means the same thing to every consumer and in every session (Cohen et al., 1999). Additionally, when dealing with food products, we must realize that the size of the sample prevents panelist from testing all products. Hence, we cannot ask our panelist to spend long periods rating the whole set of food samples. Typically, each consumer only participates in one or a small number of tasting sessions, usually in the same day. Notice that tasting a large sample of food may be physically impossible, or the number of tests performed would damage the sensory capacity of consumers. On the other hand, expert descriptions are ratings in an ordinal scale of different aspects of products related to their taste, odor, color, etc. Here we must assume that a rating of “7” (in say, texture) means the same for a given expert in every product; though not necessarily for every expert. 6 Sometimes, to describe a kind of food products, it is possible to use categorical attributes instead of numerical ones. In these cases, it will be necessary the use of an adequate kernel in the algorithms that we are going to explain in the next section. From a practical point of view, the market is not interested in tastes of individual consumers, the purpose of marketing studies of sensorial data is to discover, if they exist, widespread ways to appreciate food products that can be considered as market segments. These segments can be seen as clusters of consumers with similar tastes. In this paper, we will show that the similarity of preference criteria of consumers can be computed in a high dimension space by means of a kernel-based method. 4. Clustering people according to their preferences In this section we are going to discuss the available options to build a method for clustering people using their preference criterion as the measure of similarity. For ease of reference, let S be a set of n objects from an input space X that will be rated by m people. Therefore, given a space of ordinal values, Scale, we have a family ri: Si → Scale, i = 1,…,m (1) of rankings, one for each person, that in general are partially defined with respect to S, since usually not everybody rates everything, that is, Si ⊆ S. To measure the similarity between the preferences of two people pi and pj, a first attempt consists of comparing the vectors (ri(x): x ∈ Si) and (rj(x): x ∈ Sj) of their ratings. To do this, we must realize that Si and Sj can have few common objects. In fact, in sensory data it is frecuent that Si ∩ Sj = Ø. Different tools have been employed to make comparisons when using these rating vectors for prediction tasks in collaborative filtering (see (Breese et al., 1998)); the most obvious and unsuccessful being Euclidean distance, used in nearest neighbor algorithms. Pearson’s correlation or the cosine of the vectors has been put forward to consider the possible differences in the scales used by different people. 7 However, these comparison techniques, devised for prediction purposes, are not easily extendable to clustering. To illustrate this point, let us consider one person pi with a coherent preference criterion, and let us divide pi’s rating vector in two parts: (ri(x): x ∈ Si1), (ri(x): x ∈ Si2), with Si1 ∩ Si2 = Ø and Si1 U Si2 = Si. (2) These two vectors would not have anything in common for any reasonable comparison measure. However, both vectors represent the same rating criterion, the criterion of pi; so pi rating both Si1 and Si2 must be included in the same cluster. However, it is not obvious how to represent what we have lazily called rating criterion of a person. The proposal presented in this paper is to represent preference criteria explicitly by means of a function able to generalize somehow the preferences provided by the ratings. The following sections will show how these functions can be defined and learned. Then we will describe the clustering method based on the similarity of these functions. But before going on it is necessary to reconsider the kind of data that we will use as training material to learn people’s preferences. If we have a vectorial way to describe the objects in S and a rating vector (r(x): x ∈ S), we can try to use regression to induce a function that maps object descriptions into ratings. However, this is not a faithful way of capturing people’s preferences; see (Herbrich et al., 2000) for a discussion of the differences between ordinal and continuous (metric or least squares) regression. Additionally, when consumer panels are involved, an important reason to discard any kind of regression is that ratings are relative orderings instead of absolute values, due to the so-called batch effect (Joachims, 2002; Díez et al., 2004). Therefore, when it is possible to consider explicitly the tasting session where the rates were made, we will take them into account. Thus, a more general setting will be used. Following (Herbrich et al., 2000; Joachims, 2002), in order to capture the ordinal nature of ratings, and the relative character of people’s preferences, for each consumer pi, we will separate the ratings given at each tasting session. Then we will create a preference judgment set PJi = {( x (1) , x (2) ) : x (1) , x (2) ∈ Si, ri(x (1) ) > ri(x (2) ), session(pi, x (1) ) = session(pi, x (2) )} (3) 8 suitable for our needs just considering all pairs (x (1) , x (2) ) of object in Si such that they were presented in the same session to consumer pi, and the rating of the first object was higher than the rating of the second. 4.1 Class separation and preferences Let us first consider a simplified case in this subsection before presenting the general learning framework. So, let us assume that the input space X is in fact a subset of real vectors; in symbols, X ⊂ R d . In order to learn our preference problems, we will try to find a real ranking (or preference) function f: R d � R that maximizes the probability of having f(x (1) ) > f(x (2) ) whenever x (1) is preferable to x (2) . If we now restrict the space of hypothesis to linear functions, each preference judgments set PJi gives rise to a set of constraints (x (1) , x (2) ) ∈ PJi � fi(x (1) ) > f(x (2) ) � fi(x (1) - x (2) ) > 0 � fi(x (2) - x (1) ) < 0 (4) Therefore, ranking functions can be learned using a binary classification algorithm able to discriminate the class of objects according to the sign returned, as happens with Support Vector Machines (SVM). Notice that the training set is Ti = {( x (1) - x (2) ; +1), (x (2) - x (1) ; -1): (x (1) , x (2) ) ∈ PJi}. (5) The function learned passes through the origin of coordinates, and it is then defined by ( ) ∑ = == d k kikii 1 ,f xwxwx (6) where wi is the weight vector of the consumer pi. The return of fi on an object representation x can be thought as the assessment of x in the sense that fi(x) will be used to predict preferences between x and other products. On the other hand, the weight vector wi is the director vector of a hyperplane ( )0, =xw i called the assessment hyperplane. From a geometrical point of view, the distance (in the direction of wi) from the hyperplane to each object is proportional to the value returned by fi. In fact (see Figure 1), 9 ( ) i i i i i w x w xw xxw ′ = ′ =′= f, );0,(distance (7) Of course, it would be difficult to find a linear function able to separate the positive from the negative differences of the training set Ti. In terms of preferences, not ever the assessment of an object will be proportional to the components of a vectorial description. It is not true that the more (or the less) the better; for instance, the amount of sugar or salt in our favorite foods usually has an optimal point and the increase or decrease from that equilibrium point is frequently rejected. Additionally, there may be situations where the objects considered have not a straightforward description by means of real vectors. To take into account all these issues, it is possible to use a kernel trick (Herbrich et al., 2000). To present this general framework, let us assume that the space X of objects can be mapped to a Hilbert space F using φ: X � F. Then, as we have just seen, to learn a ranking or preference function we only have to find a linear separation in F for the training set TiF = {(φ(x (1) ) - φ(x (2) ); +1), (φ(x (2) ) - φ(x (1) ); -1): (x (1) ,x (2) ) ∈ PJi }. (8) To avoid handling the components of product descriptions in F, the kernel trick helps us redefining Ti as Ti = {(x (1) , x (2) ; +1), (x (2) , x (1) ; -1): (x (1) ,x (2) ) ∈ PJi }. (9) Therefore, the input space that we need is in fact the product of X by itself, and it is mapped into the feature space F using Ψ: X x X � F , Ψ(x (1) ,x (2) ) = φ(x (1) ) - φ(x (2) ). (10) Hence, the associated kernel to this transformation is ( ) ( ) ( ) ( ) ( ) ),(),(),(),( )(),()(),()(),()(),( )()(,)()(,,,,;, (4)(2)(3)(2)(4)(1)(3)(1) (4)(2)(3)(2)(4)(1)(3)(1) (4)(3)(2)(1)(4)(3)(2)(1)(4)(3)(2)(1) xxxxxxxx xxxxxxxx xxxxxxxxxxxx KKKK +−−= +−−= −−=ΨΨ= φφφφφφφφ φφφφK (11) where K(x (.) ,x (.) ) is the kernel associated to the transformation φ of the individual objects. The kernel K so built is called the Herbrich’s Kernel (Herbrich et al., 2000) attached to K. 10 The separation function induced by a classification SVM from Ti with kernel K will be a function Fi: X x X � R of the form ( ) ( )( ) ( ) ( ) ( )( ) ( )( ) ( ) ( )( )∑∑ ∈∈ =−−= ii SVs ssss SVs ssssi yy 21)2()1(21)2()1(21 ,,,,, xxxxxxxxxxF Kαφφφφα (12) where SVi is a set of indexes for the support vectors, and ys is the class (+1 or -1) of the pairs (xs (1) , xs (2) ) of Ti. This function has an important property that is an immediate consequence of the kernel definition: Fi(x (1) ,0) > Fi(x (2) ,0) � Fi(x (1) , x (2) ) > 0 (13) Thus, we define the following assessment function fi: X � R, fi(x) = Fi(x,0). (14) It is trivial to see that fi is as coherent with the preference judgments of PJi as Fi is a separating function for Ti. Moreover, in the case of using as K the linear kernel, fi coincides with the function defined in (6). The expression of fi can be simplified given that it is possible to skip a constant term from Fi(x,0); in the linear case this constant is 0. Therefore, in practice, the general assessment function fi is given by ( ) ( ) ( ) ( ) ( ) ( )( )∑∑ ∈∈ −=−= ii SVs ssss SVs ssssi KKyy xxxxxxxx ,,,f )2()1()2()1( αφφφα (15) 4.2 Defining a similarity measure of people’s preferences Let us start with a motivating example assuming that there are 3 people that have been asked about their preferences about a kind of products. To simplify the presentation, let us suppose that the products studied can be represented by a single number. Think, for instance, in the proportion of milk to coffee in your café au lait. If the first person (p1) prefers coffees with a proportion of 0.20 of milk, her preference function could be a parabola with vertex in that point. Given that ranking or preference functions cross the origin of coordinates, the preference function of p1 has the form f 1(x) = a1 ∗ (-x 2 + x ∗ 2 ∗ 0.20) = < a1 ∗ (-1, 0.40), (x 2 , x)> (16) 11 where a1 is a positive constant. If the other two persons prefer the combinations around 0.21 and 0.40 respectively, the director vectors of the preference functions of each people are the following w1 = a1 ∗ (-1, 0.40); w2 = a2 ∗ (-1, 0.42); w3 = a3 ∗ (-1, 0.80). (17) Notice that the values of positive constants ai are not relevant if we want to use the functions fi to predict the preference of pi about a pair of possible combinations. What is really important is the relative weight of the components of wi. Thus, a way to measure the similarity of the tastes of these people is by the cosine of the director vectors of their preference functions ji ji jiji pp ww ww ww ⋅ ⋅ == ),cos(),(similarity (18) Using this definition, the similarity of p1, p2, p3 is similarity(p1,p2) = 0.9999; similarity(p1,p3) = 0.9570; similarity(p2,p3) = 0.9618. (19) In other words, as one could hope, using this approach the tastes of p1 and p2 are nearer from each other than the taste of p3. Notice that we have summarized the tastes of people in their preference function instead of using somehow the set of ratings assigned to a collection of products. Probably the products tasted were different for each person, what would make difficult to define any reliable similarity measure based on these ratings. Returning back to the general setting, let ((PJi, wi): i = 1,…, m) be a collection of m preference judgment sets PJi endowed with the director vector wi of their respective ranking functions. We define the similarity of their preference criterion by the cosine of equation (18). To make operative this definition, we need to arrange the equation (15) to make visible the director vector wi attached to a set of preference judgments PJi. In fact, the assessment function fi can be represented by a weight vector wi in the higher dimensional space of features such that fi(x) = < wi , φ(x)>, (20) 12 where ( )∑ ∈ −= iSVs )( s )( sssi )()(y 21 xxw φφα (21) Now, given that the definition of similarity uses scalar products instead of coordinates of weighting vectors, we can easily rewrite (18) in terms of the kernels used in the derivations explained in the previous subsection. The essential equality is: ( )∑ ∑ ∑ ∑ ∈ ∈ ∈ ∈ = −−= i j i j SVs SVl llsslsls SVs SVl llsslslsji yy yy )2()1()2()1( )2()1()2()1( ,,, )()(),()(, K xxxx xxxxww αα φφφφαα (22) 4.3 The clustering method Once we have defined a reasonable similarity measure for preference criteria, we proceed to look for clusters of consumers with homogeneous tastes. In principle, we could use any available clustering algorithm. However, we avoided those methods, like k-means, that require frequent recomputations of the centroids of each cluster. The reason is that the updating of (22) would result computationally expensive. Additionally, we need a mechanism able to estimate an appropriate number of clusters directly from the data, without any explicit manual intervention. Hence, we applied a nonparametric pairwise algorithm (Dubnov et al. 2002), although this is not the only possibility. The following paragraphs sketch a description of this algorithm as we used it in the experimental results reported in the last section. Let M = (Mij) be a square matrix where Mij stands for the similarity between data points i and j; in our case, data points are the vectorial representation of the preference criteria of consumers, and similarities are given by equation (18). In the following, M will be called the proximity matrix. Then, matrix M is transformed iteratively, following a two step procedure that converges to a two values matrix (1 and 0), yielding a bipartition of the data set into two clusters. Then, recursively, the partition mechanism is applied to each of the resulting clusters represented by their corresponding submatrices. To guarantee that 13 only meaningful splits take place, in (Dubnov et al. 2002) the authors provide a cross validation method that measures an index that can be read as a significance level; we will only accept splits whose level is above 0.90. The basic iterative transformation uses the following formulae to go from iteration t to t+1: ( ) ( )        +++ + ++ +++ + +=+ =+ ∑∑ k ikjk jk jk k jkik ik ikij ik ij ij tVtV tV tV tVtV tV tVtM ktM tM tV )1()1( )1( log)1( )1()1( )1( log)1( 2 1 )1( }:)(max{ )( )1( 2 1 2 1 (23) The first step normalizes the columns of the proximity matrix using the L∞ norm; then the proximities are re-estimated using the Jensen-Shannon divergence. The idea is to formalize that two preference criteria are close (after these two steps) if they were both similar and dissimilar to analogous sets of criteria before the transformation. 5. Experimental results To illustrate the performance of our proposals, we used two datasets. The first one is EachMovie, a dataset widely used in the field of collaborative filtering. The second dataset deals with ratings of beef meat consumers, and was previously used in (Díez et al. 2005). This is a dataset that was built in order to study the beef market preferences. 5.1 Clustering spectators according to their tastes about movies The database EachMovie contains ratings of 1,628 movies provided by 72,916 people. The scale used has 6 values {0, 0.2, 0.4, 0.6, 0.8, 1}, but less than only 2.4% of the possible values are filled. Thus, for instance, 11,651 people have not given any rating to any movie. To take into account explicitly the absence of ratings, that is so frequent in this collection, and as usual when dealing with EachMovie, we moved the scale of ratings to {-1.5, -1, -0.5, 0.5, 1, 1.5}, assuming zero as missing value. 14 As many authors do, in order to avoid the extreme sparsity of data, we considered a subset of EachMovie. We only considered movies with at least 1,000 ratings; there are 504 movies in such conditions. We likewise only took into consideration people who had submitted at least 275 ratings for these movies; this makes a total of 189 people. To accomplish our setting for clustering, we considered 100 people as our sample to study possible similarity of preference criteria; i.e. we set m = 100 (see equation (1)) and the remaining available ratings were used to define the input space X of the objects (movies) considered in our experiment. In other words, we represented each movie by a vector of 89 (= 189 – 100) ratings. The distributions of ratings in the description of the movies, that is in the attribute space, and those provided by our 100 spectators are quite similar, see Table 2. The main novelty in this setting with respect to typical collaborative filtering applications is that we use some consumers to describe the products. Notice that we only use “expert” consumers, those who evaluated a lot of movies, so in fact they form a kind of expert panel. For each spectator pi from the set of 100 people considered, we have a rating vector (ri(x): x ∈ Si), and we have to assemble a set of preference judgments PJi from it. Notice that |PJi| is of the order of O(|Si| 2 ). Here |Si|<=504, so to avoid too big PJi datasets, we proceed as follows. For each movie x (1) ∈ Si ranked by pi, 10 more movies ranked by pi with different ratings were randomly selected; then, if x (2) is one of such movies, we include (x (1) , x (2) ) in PJi if and only if ri(x (1) ) > ri(x (2) ), otherwise, we include (x (2) , x (1) ) in PJi. Notice that in EachMovie there is no indication about rating sessions, therefore we assumed only one session, and hence all ratings are comparable. Finally, we used two different kernels to appreciate their influence in the clusters obtained applying the method described in this paper. In Figure 2 we depict the tree of clusters obtained with linear kernel and polynomial kernel (degree=2). The overlapping between the clusters obtained by these two kernels is reported in Table 3. The set of all preference judgments, PJ = U(PJi: i = 1,...,m) 15 has a lot of disagreements. If, for each pair of movies, we choose the most frequent relative ordering in PJ, then the percentage of pairs disagreeing with that majority opinion is 25.92% (see Table 4). Notice that this percentage is the minimum training error for any algorithm trying to learn the preferences of the whole group of 100 spectators. The clustering algorithm, with both kernels, has discovered groups of almost 50 people, whose disagreements are lower than the global percentage of disagreements: 17.4% for the linear kernel (li_r), and 16.71% for the polynomial one (po_r). These clusters are quite the same; since 42 spectators are present in these majority groups for both kernels (see Table 3). However, the other two smaller groups have internal disagreements similar to that of the whole group in both kernels. The reason for this behavior may be that these groups are made up of very peculiar spectators whose preference functions are difficult to learn. The preference functions learned from the union of preference judgments of the members of each cluster have an estimation of misclassifications similar to the percentage of disagreements, especially in the case of the polynomial kernel, see Table 4. 5.2 Market segments in beef meat To illustrate our method with a real world application, we used a database described in (Sañudo et al. 2004; Díez et al. 2005). The data collects the sensory ratings of a panel of beef meat consumers about two aspects: tenderness, and acceptability. First we will describe the dataset, to conclude with a discussion of the implications of the results obtained by our method. 5.2.1 Beef meat dataset For this experience, 101 animals of 7 Spanish breeds were slaughtered to obtain two kinds of carcasses: lights, from animals with a live weight around 300–350 kg (light); and heavies, from animals at 530–560 kg. The set of animals was uniformly distributed by breeds and weights. Additionally, to test the influence of aging in consumers’ appreciation, each piece of meat was prepared with 3 aging periods, 1, 7, and 21 16 days. Notice that these settings give rise to 303 (101 animal with 3 different aging periods) samples of meat. On the other hand, the 7 breeds used constitute a wide representation of beef cattle. These breeds can be divided into four types: double muscled (DM, one breed), fast growth (FG, two breeds), dual purpose (DP, one breed), and unimproved rustic type (UR, three breeds). In Table 5, we show the average percentages of fat, muscle and bone for each breed. Notice that it is important to distinguish three kinds of fat: inter-muscular, subcutaneous, and intramuscular, that is included in the percentage of muscle. In the experimental results reported below, we used the carcass composition of each animal; the breed was only used to provide an explanation of the tastes of clusters. Additionally, each kind of meat was also described by a panel of 11 trained experts who rate 12 traits of products such as fibrosis, flavor, odor, etc.; we considered the average rate of each trait. The characterization of meat samples was completed with 6 physical features describing its texture. The panel was organized in a set of tasting sessions, where a group of potential consumers assessed 3 or 7 instances of the available beef kinds. Frequently, each consumer only participated in a small number (sometimes only one) of tasting sessions, usually in the same day. To apply the clustering method described in this paper, we first selected those people involved in our consumers’ panel whose ratings gave rise to more than 30 preference judgments; these yielded us to consider a set of 171 panelists (m=171) that tested from 9 to 14 samples of meat. Notice that this is just a subset of all available panelists since samples rated with the same rate did not produce preference judgment pairs. This individual treatment contrasts with the global one that we reported in (Luaces et al. 2004; Del Coz et al. 2004; Díez et al. 2004). In these prior works, we were interested in modeling the general opinion of consumers; so for each session, to summarize the opinions of consumers, we computed the mean of the ratings obtained by each piece of meat. Here, we will also use the polynomial kernel of degree 2 reported in those papers, since the nature of the learning problems is essentially the same. Since there are 303 samples of meat, then the opinions of our panelists can only be estimated inducing a preference or ranking function (see section 4.1) using PJi datasets obtained applying equation (3). Notice that only such functions can be used in order to compare the preferences of different consumers; in 17 general, two arbitrary consumers have not tested samples of the same animal prepared with the same aging. However, it is possible to compare the preference functions of any couple of consumers as vectors in a high dimension space following the kernel based method of section 4.2. The clustering algorithm returns the trees depicted in Figure 3. Split nodes achieved a confidence level of 91% for tenderness dataset, and 97% for acceptance. The leaves of these trees reached lower confidence levels, and therefore their splits were rejected. In the scores reported in Table 6, we observe that both in acceptance and tenderness there are percentages of disagreements over 20%, while the clusters found decrease this amount in 5 points. Using the set of all preference judgments of each cluster, the preference functions learned (with a polynomial kernel of degree 2) have an estimation of misclassification around 20%. 5.2.2 Implications of beef meat results In general, it is well known that meat qualities are mainly the result of a set of complex factors. In this study, we focus our attention on two of them: breed and aging period. Specifically, we are interested in knowing if there are different groups of people who prefer some breeds to others or who prefer long to short aging periods. In fact, these are the most relevant features in the acceptance and tenderness appreciation of beef meat by consumers according to (Sañudo et al. 2004; Luaces et al. 2004). To gain insight into the meaning of the preference criteria of each cluster, we used the ranking or preference functions to order the samples of meat; then we assessed 10 points to those samples included in the first decile, 9 to the second decile, and so on. Graphical representations of the average points obtained by each breed and each aging period are shown in Figure 4 (acceptance) and Figure 5 (tenderness); notice that the average score of all samples is 5.5. The results are quite the same if we use quartiles instead of deciles or any other division of the relative rankings of each cluster. In the acceptance dataset (Figure 4a), let us emphasize the opposite role played by Retinta and Asturiana breeds: they were first and last (or almost last) in each cluster alternatively. In (Luaces et al. 2004) we used Boolean attributes to include the breed in the description of each sample, and then Retinta and 18 Asturiana were found to be the most relevant Boolean features in order to explain consumer’s acceptance of meat. Additionally, these two breeds have significant differences in carcass composition (see Table 5). Notice that Asturiana breed is the only double muscled breed of the sample, and then it has the lowest values in percentages of subcutaneous and inter-muscular fat, and bone; while Retinta is the unimproved rustic breed with the highest percentages of fat and bone. Therefore, there are some reasons so as to assign opposite ratings to samples of these two breeds, although, in general, the final acceptance scorings rely on a complex set of features. On the other hand, analyzing aging periods in both clusters, (Figure 4b) we can see that people in the left cluster prefer a 7 days aging period meanwhile in the right cluster longer aging periods have better acceptance. In the tenderness dataset (Figure 5a), meat from Pirenaica and Retinta breeds are the tenderest for people in left cluster, however they are ranked in low positions in right cluster. We can say exactly the opposite of meat from Asturiana and Parda breeds. Again, Asturiana and Retinta breeds play opposite roles in each cluster. In this case, we cannot appreciate differences between clusters’ preferences about aging periods (Figure 5b): in both clusters, all breeds obtain higher tenderness scores as aging periods increase. This is not a surprisingly result, it is well documented that long aging periods give rise to more tender meat (Sañudo et al. 2004). 6. Conclusions We have presented a new algorithm to build clusters of preference criteria. Starting from a collection of preference judgments of different people, our algorithm discovers groups or clusters of people with closely related tastes; that is to say, people whose preference judgments can be merged in order to learn more reliable ranking functions able to express the preferences of the people involved. These clusters can be also interpreted as market segments. In building these clusters, the key insight is that ranking 19 functions, learned from the set of preference judgment of each person, codify the rationale or criteria for his or her preferences. We believe that the contributions of this paper fall mainly within the applications of preference learning. Thus, clustering is useful for strengthening applications such as collaborative filtering, information retrieval or adaptive assistants. We illustrated the usefulness of the method presenting the experimental results achieved with a collaborative filter dataset (EachMovie), and with a dataset that gathers ratings of beef meat consumers. In this case, the clusters represent market segments with differentiated requirements from this kind of food products. It is straightforward the extension of this case to other fields that handle sensory data provided by consumer panels. Acknowledgements The authors would like to thank: the Company Compaq for providing us with the EachMovie database (McJones, 1997); Thorsten Joachims for his SVM light (Joachims, 1998), and the authors of Spider (Weston et al., 2006), a MatLab toolbox that includes kernel based algorithms; these systems were used in the experiments reported in this paper. References Bahamonde, A.; Bayón, G. F.; Díez, J.; Quevedo, J. R.; Luaces, O.; del Coz, J. J.; Alonso, J.; Goyache, F. (2004). Feature subset selection for learning preferences: a case study. Proceedings of the 21st International Conference on Machine Learning, ICML 2004, Banff, Canada, pp. 49-56. Basu, C., Hirsh, H., and Cohen, W.W. (1998). Recommendation as classification: Using social and content-based information in recommendation. In Proc. AAAI. pp. 714-720. Branting, K.L. and Broos, P.S. (1997). Automated acquisition of user preferences. International Journal of Human- Computer Studies 46:55-77. 20 Breese, J.S., Heckerman, D., and Kadie, C. (1998). Empirical analysis of predictive algorithms for collaborative filtering. In Proc. Conference on Uncertainty in Artificial Intelligence. pp. 43-52. Buck, D., Wakeling, I., Greenhoff, K., and Hasted, A. (2001). Predicting paired preferences from sensory data. Food quality and preference 12:481-487. Cheung, W.K., Kwok, J.T., Law, M.H., and Tsui, K.C. (2000). Mining customer preference ratings for product recommendations using the support vector machine and the latent class model. In Proc. International Conference on Data Mining Methods and Databases for Engineering. pp. 601-610. Cohen, W.W., Shapire, R.E., and Singer, Y. (1999). Learning to order things. Journal of Artificial Intelligence Research 10:243-270. Corney, D. (2002). Designing food with bayesian belief networks. In Proc. International Conference on Adaptive Computing in Engineering Design and Manufacture. pp. 83-94. Crammer, K. and Singer, Y. (2001). Pranking with ranking. In Proc. Conference on Neural Information Processing Systems. pp. 641-647. Del Coz, J. J.; Bayón, G. F.; Díez, J.; Luaces, O.; Bahamonde, A.; Sañudo, C. (2004). Trait selection for assessing beef meat quality using non-linear SVM. Proceedings of the 18th Annual Conference on Neural Information Processing Systems (NIPS 2004). Vancouver, British Columbia, Canada, pp. 321-328. Díez, J.; Bayón, G.F.; Quevedo, J. R.; del Coz, J. J.; Luaces, O.; Alonso, J.; Bahamonde, A. (2004). Discovering relevancies in very difficult regression problems: applications to sensory data analysis. In Proceedings of the European Conference on Artificial Intelligence (ECAI ’04), Valencia, Spain, pp. 993-994. Díez, J.; del Coz, J. J.; Sañudo, C.; Albertí, P.; Bahamonde, A. (2005). A Kernel Based Method for Discovering Market Segments in Beef Meat. Proceedings of the 9th European Conference on Principles and Practice of Knowledge Discovery in Databases, ECML/PKDD'2005, Porto, Portugal, pp. 462-469. Dubnov, S.; El-Yaniv, R.; Gdalyahu, Y.; Schneidman, E.; Tishby, N.; Yona, G.. A New Nonparametric Pairwise Clustering Algorithm Based on Iterative Estimation of Distance Profiles. Machine Learning, 47 (2002) 35–61 Dumais, S., Bharat, K., Joachims, T., and Weigend, A. (2003). Workshop on implicit measures of user interests and preferences. In ACM SIGIR Conference, Toronto, Canada. Fiechter, C.N. and Rogers, S. (2000). Learning subjective functions with large margins. In Proc. International Conference on Machine Learning. pp. 287-294. 21 Freund, Y., Iyer, R., Schapire, R.E., and Singer, Y. (1998). An efficient boosting algorithm for combining preferences. In Proc. International Conference on Machine Learning. pp. 170-178. Goldberg, D., Nichols, D., Oki, B.M., and Terry, D. (1992). Using collaborative filtering to weave an information tapestry. Communications of the ACM 35 (12):61-70. Goyache, F., Bahamonde, A., Alonso, J., López, S., del Coz J.J., Quevedo, J.R., Ranilla, J., Luaces, O., Alvarez, I., Royo, L., and Díez, J. (2001). The usefulness of artificial intelligence techniques to assess subjective quality of products in the food industry. Trends in Food Science and Technology 12 (10):370-381. Herbrich, R.; Graepel, T.; and Obermayer, K.: Large margin rank boundaries for ordinal regression. In A. Smola, P. Bartlett, B. Scholkopf, and D. Schuurmans, editors, Advances in Large Margin Classifiers, 115–132. MIT Press, Cambridge, MA, (2000) Hofmann, T and Puzicha, J. (1999) Latent class models for collaborative filtering. In Proc. International Joint Conference on Artificial Intelligence. pp.688-693. Joachims, T. (1998). Making large-scale SVM learning practical. Advances in Kernel Methods - Support Vector Learning, MIT Press, Cambridge MA. Chapter 11. Joachims, T. (2002). Optimizing search engines using clickthrough data. In Proc. ACM Conference on Knowledge Discovery and Data Mining. Luaces, O.; Bayón, G.F.; Quevedo, J.R.; Díez, J.; del Coz, J.J.; Bahamonde, A. (2004). Analyzing sensory data using non-linear preference learning with feature subset selection. Proceedings of the 15th European Conference of Machine Learning, (ECML 04). Pisa, Italia, pp. 286-297. McJones, P. (1997). EachMovie collaborative filtering dataset. DEC (now Compaq) Systems Research Center. http://www.research.compaq.com/SRC/eachmovie/. Murray, J.M., Delahunty, C.M., and Baxter, I.A. (2001). Descriptive sensory analysis: Past, present and future. Food Research International 34:461-471. Pazzani, M.J. (1999). A framework for collaborative filtering, content-based and demographic filtering. Artificial Intelligence Review 13 (5-6):393-408. Resnick, P. and Varian, H. (1997). Introduction to special section on recommender systems. Communicatons of the ACM 40(3):56-58. 22 Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and Riedl, J. (1994). GroupLens: An open architecture for collaborative filtering of networks. In Proc. Conference on Computer Supported Cooperative Work. pp. 175-186. Sañudo, C.; Macie, E.S.; Olleta, J.L.; Villarroel, M.; Panea, B.; Albertí, P.. The effects of slaughter weight, breed type and ageing time on beef meat quality using two different tex-ture devices. Meat Science, 66 (2004), 925–932 Shardanand, U. and Maes, P. (1995). Social information filtering: Algorithms for automatic “Word of Mouth”. In Proc. Conference on Human Factors in Computing Systems. pp. 210-217. Shashua, A. and Levin, A. (2002). Ranking with large margin principle: Two approaches. In Proc. Neural Information and Processing Systems. Tesauro, G. (1989). Connectionist learning of expert preferences by comparison training. In Proc. Neural Information and Processing Systems. pp. 99-106. Ungar, L.H. and Foster, D.P. (1998). Clustering methods for collaborative filtering. In Proc. AAAI’98 Workshop on Recommendation Systems. pp. 112-125. Utgoff, J. P. and Clouse, J. (1991). Two kinds of training information for evaluation function learning. In Proc. AAAI'91. pp. 596-600. Vapnik, V. (1998). Statistical learning theory. John Wiley, New York. Weston, J.; Elisseeff, A.; BakIr, G.; Sinz, F. (2006) SPIDER: object-orientated machine learning library. http://www.kyb.tuebingen.mpg.de/bs/people/spider/ 23 Table 1. Sensory data collected from panels of experts and consumers. Each product is described by expert assessments (Attj) in addition to other (O-Atti) chemical, biological or physical analysis outputs Expert sensory appreciations Expert-1 Expert-k Other descriptive attributes Consumer preferences Att1 … Attm … Att1 … Attm O-Att1 … O-Attn Session Consumer Rating <num>…<num>…<num>…<num> <num> … <num> <i> <Id> <num> … … … … … … <num>…<num>…<num>…<num> <num> … <num> <i> <Id> <num> Table 2. Rating distribution in data sets used for the experiments In attribute space In the 100 spectators’ ratings Movies’ Rating Number Percentage Number Percentage 0 4,392 17.02% 6,665 19.32% 0.2 1,955 7.57% 2,651 7.68% 0.4 3,813 14.77% 4,801 13.92% 0.6 6,373 24.69% 8,193 23.75% 0.8 5,774 22.37% 7,471 21.66% 1 3,503 15.57% 4,716 13.67% Totals 25,810 34,497 Table 3. Number of common spectators in clusters obtained with the linear kernel and the polynomial kernel of degree 2 Polynomial of degree 2 po_l_l po_l_r po_r li_l_l 12 12 3 27 li_l_r 5 17 2 24 L in e a l li_r 7 0 42 49 24 29 47 24 Table 4. Number of disagreements (minimum possible training error) and classification error estimations in each cluster depending of the kernel used Kernel disagreements cluster disagreements classification errors linear 25.92% li_l_l 27.35% 32.71% li_l_r 25.78% 30.62% li_r 17.40% 19.18% polynomial 25.92% po_l_l 23.8% 24.68% po_l_r 30.36% 30.86% po_r 16.71% 17.00% Table 5. Carcass compositions of 7 Spanish beef breeds used in the experiment Breed Fat % Bone Muscle Intramuscular Name Type inter-muscular subcutaneous % % fat % Asturiana Valles DM 4.77 0.89 16.00 78.34 0.90 Avileña UR 13.17 3.53 19.25 64.05 2.28 Morucha UR 12.46 3.46 19.28 64.80 2.10 Parda Alpina DP 9.65 2.32 20.86 67.17 1.82 Pirenaica FG 9.02 3.01 17.33 70.63 1.48 Retinta UR 14.16 4.75 20.89 60.20 2.13 Rubia Gallega FG 5.73 1.20 16.56 76.52 1.12 Table 6. Number of disagreements (minimum possible training error) and classification error estimations in each cluster for acceptance and tenderness Dataset disagreements cluster |PJ| disagreements classification errors acceptance 21.41% left 1927 16.19% 19.20% right 2150 17.07% 21.12% tenderness 20.25% left 2487 15.96% 19.38% right 2432 15.21% 19.59% 25 Fig. 1. The assessment of each object is proportional to the distance of the vector that represents it to the (assessment) hyperplane perpendicular to wi. In the picture, x (1) is preferable to x (2) , since it is further Fig. 2. Trace of the clustering algorithm in the EachMovie dataset. The left hand side corresponds to the linear kernel while the right side represents the result obtained with the polynomial kernel of degree 2. In each node we report the number of spectators in bold, and the confidence level required to split the node (in parenthesis). The leaves are labelled for further reference 26 Fig. 3. Trace of the clustering algorithm in the beef meat dataset Fig. 4. Acceptance of beef meat. Average ranking scores for each breed and aging period 27 Fig. 5. Tenderness of beef meat. Average ranking scores for each breed and aging period 
work_3zj3tdg2jja53nmljlfxxfkeq4	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220995899 Params is empty 220995899 exception Params is empty 2021/04/06-03:05:40 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220995899 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:40 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_42y67v6mcvacfodxyqueyuewk4	----	15-SE512381-Zhao_TOMEJ Send Orders for Reprints to reprints@benthamscience.ae 892 The Open Mechanical Engineering Journal, 2014, 8, 892-898 1874-155X/14 2014 Bentham Open Open Access Modularized Cutting Tool Selection Expert System Chen Wang1, Wu Zhao,*1, Ling Chen2, Kai Zhang1 and Xin Guo1 1School of Manufacturing Science and Engineering, Sichuan University, Cheng Du 610065. China 2Division of Production and Materials Engineering, Lund University, Lund, 999027, Sweden Abstract: This paper is to present a rule-based cutting tool selecting expert system which has knowledge modules and rule bases. Besides, according to different process targets, the selection progress will apply corresponding constraints and rule modules. The logic of tool selection follows a decision-making procedure as an experienced engineers. The strategy of system is to guide the user through several standard steps: information input; feature recognition; selection of machining method; selection of tool material and type; calculation of process parameter and solving cutting problem. This system also has a modularized structure which allows adding new functions and new modules to expand knowledge base and data base. Modules involves in this system are composed of the user interface, knowledge acquisition facility, explanation facility, the knowledge base module, the inference engine and the database module. Keywords: Tool selection, expert system, modularization, process optimization. 1. INTRODUCTION In mechanical manufacture industry, process planning involves scheduling resources, such as machine tools, work piece, cutting tools, operation sequence, processing parameters and the choice of auxiliary functions [1]. During the last 30 years some process planning expert systems have been developed to select tools and cutting parameters for specific operations. In early stage, process engineers selected adequate cutting parameters with their experience. Obviously, this kind of parameters were selected intuitively considering safety factor. In order to computerize and achieve more accurate cutting process solutions, automation techniques were applied. Then, engineers tried to apply intelligent techniques to optimize of the cutting parameters. Those intelligent techniques includes several optimization methods, such as: genetic algorithms [2] (Cus & Milfelner 2005), neural networks [3] (Abellan, Romeros, Siller, Estrud, & Vila, 2008) and bio-inspired techniques [4] (Zain, A. M., Haron, H., & Sharif, S. 2010). However, these methods provide intrinsic mathematical nonlinear function, hidden in a "black box" giving solutions, on the other hand, it decreases the transparency of process. In the field of manufacturing process planning, expert system involves two steps. First of all, knowledge base system extracts knowledge or experience and then applies fuzzy logic to deduce solutions. As commonly discussed in scientific literature, an online expert system which selecting process information applies fuzzy logic as reasoning mechanism to acquire the knowledge of mechanical engineers [5] (Wong and Hamouda 2003). On the other hand, in order to calculate the cutting parameters and manufacturing route in cutting process, a new computer aided process planning is developed [6] (Vidal, A., Alberti, M., Ciurana, J., & Casadesus, M. 2005). Moreover, in order to acquire optimal process parameters subject to specific precision and surface quality, tool life expectancy and production time sunder different processing conditions, some theoretical and experimental work has been carried out by organizations in several countries [7-9] (Chien&Chou, 2001; Vivanco, Luis, Costa, & Ortiz 2004 and references therein). These expert systems mainly dealt with the geometrical matters and selection of cutting tools [10]. Most of them concentrated on specific components and realize part function. Some expert systems took the component material and geometric features into consideration, however, ignored the selecting process of the cutting tool type and material. The main aim of this research is to help designers and manufacturing planners to select optimal cutting tools and optimum cutting conditions to reduce cost and time, as well as to solve some technical problems in cutting process with the aid of innovation strategy. The proposed modularized muti-objective cutting process expert system implements the following features, such as component feature recognition; selecting tool material; selecting tool type and optimum cutting conditions and innovative design. 2 COMPONENTS OF CPES Expert system use a large number specific domain knowledge of experts to solve problem [11]. And essentially evolves two kind of terms. One type is constructed with specific kinds of programming languages and tools such as rule-based systems, frame-based systems, and programming languages such as Prolog and LISP [12]. Another type is more appropriate for expert system, because expert system can reasons problem as human expert [13]. This paper tries to propose a classical optimization method integrating innovation strategy, and the design of expert system is based Modularized Cutting Tool Selection Expert System The Open Mechanical Engineering Journal, 2014, Volume 8 893 on modularize method. This system also applied theoretical models and semantic method. The proposed expert system involves user interface, inference engine, explanation facility, knowledge acquisition facility, knowledgebase, and working memory and other executable programs. The structure of the proposed expert system is illustrated as Fig. (1). 2.1. Inference Engine The inference engine which manipulates the stored knowledge to reach solutions is one of the basic components of expert system. The inference engine could search knowledge base, and provide answers to user by applying rules to the solution of a particular problem [14]. The scanning process of knowledge will continue until these antecedents match the assertions in database, then user will get the deduced results. Besides, the interpreter provides explanations and can be updated by adding a new separate knowledge base module. 2.2. Modularized Knowledge Base The knowledge of cutting process expert system mainly comes from two sources: experience of domain experts and technical documents [15]. Owning an expandable structure is important for the expansibility of knowledge base, just like domain experts learns new knowledge and applies this knowledge in the future. An independent and modularized knowledge base is proposed, this makes it possible to update the knowledge base by simply input data and editing the knowledge information. The modularized knowledgebase mainly involves factual statements, frames, objects, cases, tables, IF-THEN rules and equations [16]. Knowledge representation technology and rules are the key components of knowledge base. Relevant information about component characteristics and cutting tools in machining process are included in the hierarchy shown as Fig. (2). Elements called ‘objects’ were presented as frames in hierarchy model. The framework consists of a set of slot which contains a description for the attribute of object. Characteristics of the components, cutting tools and machining process are also expressed as objects which can be classes. The objects relation of the expert system is composed together in a hierarchy building. System identify the characteristics of objects and then apply the appropriate rules to perform the task. Rules in each module will deduce a serious of corresponding conclusions according to correlated conditions. This procedure reflects the logic of knowledge base. What we want is a knowledge base with fine structure rather than a huge plant knowledge base which includes all the rules. The relevant knowledge in expert system is represented in the terms of the formal logic. The knowledge base contains a sequence of rules in a modular approach. Thus, expert system can be used in modularized way to adapting to different conditions. The proposed knowledge base here includes rules which can be divided into several modules. In each module, rules will be applied according to the proposed approach. These rules can be split into the modules shown as Table 1. 2.3. Database Database systems mainly involves two main aspects: data and mathematical model [17]. The cutting tool data about Fig. (1). Structure of the proposed system. 894 The Open Mechanical Engineering Journal, 2014, Volume 8 Wang et al. cutting tool materials, machine tools and work piece materials in catalogues are mainly obtained from the several cutting tool manufactures, large factories and handbooks. The recommended data are put into forms in the database. Mathematical model can be used to calculate the cutting parameters based on the database. The database of system consists of several independent groups of data, shown as Table 2, involving materials, cutting tools, tool materials, cutting parameters, processing method and equations, respectively. 3. SYSTEM LOGIC The cutting process expert system run under an interactive operation environment. Users can describe the characteristic of component with the information integrated into the database. In the process of the operating the software, expert system will guide user through these main steps, including: Inputting information, feature recognition, selecting machining process; selecting tool material; selecting tool type; optimizing cutting conditions and solving problems. The scenario of cutting process expert system is shown as Fig. (3). 3.1. Process Information Input and Feature Recognition The detail information about work piece is the foundation of reasoning process. Under the current technical level, characters identified by engineers have more efficient and Fig. (2). Hierarchy tree of system. Table 1. Knowledge modules. No. Module Function I Stability robustness and allowable Gives robustness to the system and presents the rough allowable input cutting parameters. II control of cutting forces Keeping the forces of the system under prescribed upper limit in spite of variations in system parameters. III Drives constraints Covers the spindle and feed drive motors constraints. IV Initial cutting parameter selection Gives initial computational input parameters subjected potential initial spindle speed candidates. V Cost function definition and rules to inference with it A novel multi-objective cost function to evaluate the performance of the system. VI Automatic feedback and monitoring Automatic feedback to the system. VII Expert human machine interface Monitor key parameters to better inter-actuation with expert engineers and operators. Types D rills Twist drill Flat drill Deep hole drill Reaming drill Center drill Hollow drill M ill cutters Face milling cutter Vertical milling Side milling cutter Angle cutter Saw blade milling T cutter E lectrodes Graphite Electrodes T urning tools Boring Threading Ext. Machining Param eters Cutting speed Tool life Metal removalrate M achining type Milling Drilling Turning EDM Working  piece M etals Alloy Steel ALuminum PLastics ... D im ensions Three high Geometrical feature Precision Roughness ... M aterial properties Hardness Toughness Machinability Thermal Properties ... Status Feature Holes Pockets Threading Surface and side machining Slots Tool  material U ltrahard PCBN CBN PCD C eram ics Al2O3 mix Si3N4-sialon C em ented carbide WC-C0 Coated carbides H SS M type T type Coated C erm ets Coated Uncoated Knowledge  base Modularized Cutting Tool Selection Expert System The Open Mechanical Engineering Journal, 2014, Volume 8 895 accurate compare to characters identified by computer programs. The identification of piece involves feature specification and dimensions, which need to be entered to system by user. This system provide standardized typical characteristics for user to select. In this stage, system also offer 3D model with typical geometric features for referring. For example, turning surfaces has some of the following characteristics such as face, diameter, arc, recess, grooving and taper, etc. And these characteristics are also be divided as two groups according to the importance, shown as Fig. (4). 3.3. Selecting Machining Process The above information about work piece enables expert system to select the appropriate cutting process and cutting material. The selection machining method need to consider the these factors such as the type, size and precision, surface finishing, dimensional tolerance of the feature. In this stage, system will select the processing methods including Fig. (4). Feature recognition. Primary features 1-Cylinder 2-Face 3-Taper 4-Concave 5-Convex 6-Chamfer 7-Radius Secondary features 1-Nicked radius 2-Groove 3-Recess 4-Thread 5-Axial recess Fig. (3). The scenario of cutting process expert system. 896 The Open Mechanical Engineering Journal, 2014, Volume 8 Wang et al. traditional processing methods such as drilling, turning, milling, etc. and special working technology such as Laser, EDM, etc. The relationship between the aspect ratio and the machining type is shown as Fig. (5). Taking the typical surface finishing process for example, once the feature of curved surface is identified and the parameter is defined, then the system will make judgment. Fig. (5). Corresponding processing method to features. 3.4. Selecting the Cutting Tool The tool selection is the progress of information determination about tool holder, insert, cutting conditions, and coolant. Cutting tools are mainly composed of two parts: tool holder and indexable insert. The selecting sequence is first to select tool holder and then suiting insert. System offer two approaches for selecting cutting tools, manual searching based on the database and automatic selection by system. Manual searching which has more reliability was performed by the user owing abundant of tacit knowledge. The required information for tool selection includes: tool holder (type, clamping system, hand of cut, point angle, size, etc.), insert (size, shape, nose radius, grade, etc.), processing conditions, and coolant. What’s more, system will recommend reference case and relevant parameter to user, according to machinability data, feature. The parameters of cutting tools are included as ISO code in table contain the information about the parameters, carbide grades and functions. Selecting tool are carried out through a series of IF-THEN structures to choose the suitable tool holder and insert from the tool library. The material of cutting tools relies on the material of component to be machined. So it is needed to select the material of component firstly, then CPES will deduce the corresponding tool material. There are three way for use to select component material in this system. One option is that user selects material from the databases stored in expert system. Another way is to browse and select the material from the Cambridge Engineering Selector (CES) linked to expert system. The last option allows users input the material by themselves depending on own experiment and knowledge. Besides, this system provides a serious of rules for selecting suitable tool material corresponding to the selected material of component. 3.5. Calculating Process Parameters The system offers extra function to calculate process parameters such as machining cost and time. The data put in by user can be processed through these equations and rules to calculate parameters. Mathematical stored as rules in knowledge modules are shown in Table 1. Taking the module V for example, this module is designed to calculate cost and includes serious of equations and rules. Inference engine can use this cost function to calculate the parameters. For example, the formula (1) can be applied CylindricalPocketsSlots Feature types Feature specification Surfaces EMD Laser Turning Drilling Milling ... Rules Parameter   Fig. (6). Selection of tool type. Modularized Cutting Tool Selection Expert System The Open Mechanical Engineering Journal, 2014, Volume 8 897 to calculate the tool cost of milling, which contains other subfunction, such as TOL: life of the tool; MRR: material remove rate; SURF: surface finish; ROS: robustness of the system. They are defined as following: J(TOL,MRR,TES,R,Ci(i=1,...4)) = c1 ⋅NF1 ⋅TOL + C2 ⋅NF2 ⋅MRR + C3 ⋅ NF3 SURF + C4 ⋅NF4 ⋅ROS (1) where ci is weighting factor, NF1 represents normalisation factor. TOL = Ktol ⋅V −α1 ⋅αdc −α2 ⋅ ft −α3 (2) where Ktol is model constant, V is cutting speed (m/min), αi is model parameter, adc represents axial depth of cut (mm)and ft represents feed per tooth (mm/tooth). MRR = adc ⋅r dc⋅fc (3) where rdc represents radial depth of cut (mm) and fc represents feed velocity (mm/s). SURF = Ksurf ⋅V β1 ⋅ fc β2 ⋅αdc β3 (4) where Ksurf is a model constant and, bi represents surface roughness. ROS = min sqrt αdc,ROS( ) 2 + SS,ROS( ) 2}{( ) (5) where Ss, ROS is the length of the spindle speed. NFi = Ji − Jmax Jmax − Jmiin (6) J = Ci ⋅NFi ⋅Ji j=1 4 ∑ (7) Besides, there are some more simple rules to estimate process time and cost. Pt = Vf Rmr (8) where Pt (min) is process time, Vf (mm3) is form feature volume, Rmr (mm3/min) is material removal rate. The equations and rules of estimated cutting time and feature cost are shown as below: Set Value ((Processing: RPMd), 256.4* Processing: SPd mmmin- 1/Features: Di mm); Set Value ((Processing: FRdmmin-1); Processing: frdmmrev-1* Processing: RPMd); Set Value ((Processing: CTD min., Features: De mm/ Machining: FRd mmmin-1); Set Value ((Processing: Cost M￥), Processing: CTM min *18); Set Value ((Processing: Cost D￥), Processing: CTD min*18); Set Value ((Processing: Cost T￥), Processing: CTT min*18); IF (Processing cost: drilling<Processing cost: milling) And (Processing cost: drilling<Processing cost: Turning) THEN {‘‘c:ma. wks’’.; Read cell (1, 1); Closefile (“c;ma.wks”); }; Machining Process is drilling. 3.6. Modify Parameters and Solving Cutting Problems In this section, a problem solving module is developed, which can provide two approaches for user to modify the cutting parameter. One way is to offer typical cutting problems for reference, another option is to guide the thinking of user in systematic innovation strategy to generate solutions. The typical problems have been classified into several groups which will be presented to the user according to the design phase. The typical problems mainly come from various tool manufacturers. In this approach, user can view the problem definition option, which displays the problem, possible cause and recommended remedy. Then user can modify the cutting data. Most cutting problems are caused by unsuitable cutting parameters. So, user needs to check the cutting data firstly with the ideal data calculated by system to find out whether the cutting parameters are correct. If the values of cutting data are match with the ideal data, user should to revise the cutting data. As shown in Fig. (6), the cutting parameter for turning of steel at medium velocity are evaluated, and because the parameter about hardness and tool life is lower than the ideal parameter calculated by system, a warning is given to user Fig. (7). Fig. (7). The logic of modify parameters. Yes No Check and modify the cutting data Evalution Cutting process solution Start Process type Input material Application type Machining conditions ... Definite Problem Compare with ideal values and modify the cutting parameters Choose references case Innovation tools Push domain knowledge 898 The Open Mechanical Engineering Journal, 2014, Volume 8 Wang et al. CONCLUSION In this paper, a modularized multi-objective cutting process expert system were developed. This system enables user to select cutting tools and define cutting parameters, with the goal of short time and low cost, etc. It offer a user- friendly interface to users as well as desired results. The modularized multi-objective cutting process expert system composed of several units: the user interface; knowledge acquisition facility; explanation facility; the knowledge base module; the inference engine and the database module. The prototype system was developed based on SQL and Java. System presents the knowledge in a frame format and can complete reasoning based on rules. The strategy of system is guiding user through several steps: information input; feature recognition; selection of machining method; selection of tool material and type; calculation of process parameters and solving cutting problem. Although the prototype was developed, there were insufficient with the ability such as recognition of geometric feature and insufficient of knowledge which need further research. CONFLICT OF INTEREST We confirm that there are no conflicts of interest associated with this publication and all the financial support for this work have been declared in acknowledgement. ACKNOWLEDGEMENTS This work was supported in part by NSFC (National Natural Science Foundation of China) under Grant No.51175357, and Sichuan Province Science and technology support program under Grant No. 2014GZ0114. Gratitude is also extended to the reviewers and the Editor for their valuable comments. REFERENCES [1] A. Oral, and M. C. Cakir, “Automated cutting tool selection and cutting tool sequence optimisation for rotational parts,” Robotics and Computer-Integrated Manufacturing, vol. 20, no. 2, pp. 127- 141, 2004. [2] F. Cus, and J. Balic, “Optimization of cutting process by GA approach,” Robotics and Computer-Integrated Manufacturing, vol. 19, no.1, pp. 113-121, 2003. [3] J. V. Abellan, F. Romero, H. R. Siller, A. Estruch, and C. Vila, “Adaptive control optimization of cutting parameters for high quality machining operations based on neural networks and search algorithms,” Advances in robotics, automation and control. I-Tech Education and Publishing, Austria, pp. 472-491, 2008. [4] A. M. Zain, H. Haron, and S. Sharif, “Application of GA to optimize cutting conditions for minimizing surface roughness in end milling machining process,” Expert Systems with Applications, vol. 37, no.6, pp. 4650-4659, 2010. [5] S. V. Wong, and A. M. S. Hamouda, “Machinability data representation with artificial neural network,” Journal of Materials Processing Technology, vol. 138, no.1, pp.538-544, 2003. [6] A. Vidal, M. Alberti, J. Ciurana, and M. Casadesus, “A decision support system for optimising the selection of parameters when planning milling operations,” International Journal of Machine Tools and Manufacture, vol. 45, no.2, pp. 201-210, 2005. [7] W. T. Chien, and C. Y. Chou, “The predictive model for machinability of 304 stainless steel,” Journal of materials processing technology, vol. 118, no.1, pp.442-447, 2001. [8] A. Iqbal, N. He, L. Li, and N. U. Dar, “A fuzzy expert system for optimizing parameters and predicting performance measures in hard-milling process,” Expert Systems with Applications, vol.32, pp. 1020-1027, 2007. [9] S. R. Maity, P. Chatterjee, and S. Chakraborty, “Cutting tool material selection using grey complex proportional assessment method,” Materials & Design, vol. 36, pp. 372-378, 2012. [10] Y. Altintas, “Manufacturing automation: metal cutting mechanics, machine tool vibrations, and CNC design,” Cambridge university press, 2012. [11] N. Mandal, B. Doloi, B. Mondal, and R. Das, “Optimization of flank wear using Zirconia Toughened Alumina (ZTA) cutting tool: Taguchi method and Regression analysis,” Measurement, vol. 44, no.10, pp. 2149-2155, 2011. [12] J. Balic, F. Cus, B. Vaupotic, M. Hijjawi, Z. Bandar, K. Crockett, and D. Mclean, “Intelligent automatic cutting-tool selections for turning operations,” In: International Conference on Artificial Intelligence and Machine Learning, Dubai, United Arab Emirates, AIML-11, vol.7, 2011. [13] A. R. Yildiz, “Cuckoo search algorithm for the selection of optimal machining parameters in milling operations,” The International Journal of Advanced Manufacturing Technology, vol. 64, no.1-4, pp.55-61, 2013. [14] M. Nalbant, H. Gökkaya, and G. Sur, “Application of Taguchi method in the optimization of cutting parameters for surface roughness in turning,” Materials & design, vol. 28, no.4, pp.1379- 1385, 2007. [15] K. V. Wong, and L. O. De, “Applications of nanofluids: current and future,” Advances in Mechanical Engineering, 2010. [16] A. R. Yildiz, “Cuckoo search algorithm for the selection of optimal machining parameters in milling operations,” The International Journal of Advanced Manufacturing Technology, vol. 64, no.1-4, pp. 55-61, 2013. [17] T. Boutros, and M. Liang, “Detection and diagnosis of bearing and cutting tool faults using hidden Markov models,” Mechanical Systems and Signal Processing, vol. 25, no. 6, pp. 2102-2124, 2011. Received: December 8, 2014 Revised: December 15, 2014 Accepted: December 16, 2014 © Wang et al.; Licensee Bentham Open. This is an open access article licensed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0/) which permits unrestricted, non-commercial use, distribution and reproduction in any medium, provided the work is properly cited. 
work_4274vtpvknednjpjobxit4chnm	----	IASINEW.dvi ��������� ��� �������������������� ��������� "!#���%$ �'&(��)�&����+* ,�)�� � *%-�.0/21�)3-�4�45/ 687:9;9=<?>�@:AB>;C2DFE <�E D�>;G H'IJCLK�MON�PJQRISE�M�E TVU W�XY2Z�[�\�Y2\O[ ] U ^_U `ba5c Z c \ a5dOe [ f Uhg�c�d�i�j0k lBUhg�j0c�m Y k [ noUhp \ XY2Z j0k Y2\O[ q8r�s�t�u�v�w�t x3y�zh{�|~}�|2����|����O{����5��{���y��=�~}'{�zh���%�O}��������O{ ����}3�%������������|2������{���{������ {�y���� � ���O{�z �'���O��|~}�����zh�����h}���� ���%�'�3��zh}��'���5{�zh{;}�|�|�� zh��}���z �'�~{����+�8��y����~��{�� ��y~}�|���������y��O{����%�O}��������O{S}���������z ���~��|2�'z �5���O�B�'���J�"y�z � ��z �B��y��������(� ��y~}�|�������{b��y����%�'���S��}�z �V��������� �O{8�%����|2�����Jz �5��������}'���' �z ���%�������~��������¡ �'z ���' ��~{�����z �5��������}'����}��~������|��h}��~}����'������������� ��¢�}����£������}�z � �O� �Jx3y�zh{ {�y���� �¤zh{�|~}�����zh�����h}���� �#{���z ���O�#���V��y������O��zh��}��b��zh}��'���5{�zh{����O}'{��'��z ���5{ ���������Bz ��{�}���z � z ���¥���b��������� z ���_��y���}'����z �'�#���¤��y��:z ���~�����5��zh}��¤}��~� }��'�'��}�¦�}���z �'��§�}�� � ��¦(zh}���z �'����}'�����'��{�� ¨ ©�ª�«�¬®­8¯�°R±;«®²�­�ª ³J´ ��µ ´ ���%�3¶��_�%�~ �����$· ��¸�� ���µ����~ ��%�· ´ �%µ����(���� 8�����º¹�����µ���$· ��¸¹����%�%$·� � �»¶£�%$ �· ���� ¼��·�� :��½����� � ¥µ�¾�µ� �����µ¥¹�� ¥�� ¥�%µB¹��� � ��� #µ����� ���$L ��¿$��%�(¼�� ��µ��%µ �� ���¹�������µ�* �2�������� "�� �µ�¹���µ��%�� 0���( �� ���µ£�( ��8��� ��%�%� ��$B¹������'¶8À �~1¥Á�Â����».Ã$��%�(¼�� ��µ��%µ���½����� � 8µ�¾�µ� ����ĵ ´ ���%�01J�%µ��(¹����R ��#$ �����3¶£�� ´ Ã� Å�Å�¾ ���~ ��� �����µJ����$¥ ������� � 0�� ��Æ Ã� Å�Å�¾B �����µ������%� ¼ * ¹�1VÁ�Â����_¶��� �Ç�µJ����$ �� J�������� � ����%�~ �¾V������$��� �������µ�*3ÈÉ�������~ ������¥ ´ �( J ´ � �������µ�)���µ�¶����%����µ� ´ �� Ã���� �µS���~ ��� ���$V 0 ����Ê��� �µ��%$ ���� S�%� 0�� � ���$B$�� ��%� ¼ ´ �:�%� 0�� ����������� �������µ�µ�)�����¾¸¹��:�������� � ����%�®* ³J´ �� 0 �������¶��� �Ç¥�(Ë��� ���$ ¹2¾Ì ´ �Í����µ�µ���¹��%�%�� �¾Î ´ ���� �¾¿¶���µ¥��µ���$L��µV�ι���µ��%µ� ��»���� � 0�� ��?µ���� ´ �����µ������%� ¼�µ�* �'1 ³J´ ������� ´ ��µ��%µ;�%µ;��� o���R ´ �S�����%�:�%$ ���£��½��� ���µ�µ���$:¶£�� ´ �%�:��µ����~ �������� ¶ ´ �%� ´ ����¾¥¹��:��µ���$#��µ����� ������%µ��8�� ��_������µ���Ï2� ���~ �* ³J´ �%µ£µ����~ �������� ÐÑ¥Ò�Ó�ÓOÔBÕ�Ö¸×�Ø%Ù(ÚÛ�ÜOÝ�Þ�Û�ÞOÝ�ß�à Ø%á�Ø�â�ã�ä Ü ä Þ ã�å Ð�Ý�ß�æ Ø0ç�ä�å�è'é�ê ß�ë Ø0ç�éOä�Õ Û ê Ý�ß8ì Ø0í Þ ÚÛ�Ü é�ê Û�ÞOÝ î / ï � Å�Å�¾¥��½����� � Jµ�¾�µ� ����ð 0�� �$��%�(¼�� ��µ��%µ �:��µ� S¹����(¹����� ��_¹��� �� �� ��$¥�%�¥ ´ �8��½����� � �ñ µJ���%��$¥�������� �$��%� ¼� ��_ ´ � 0���%���'¶£�%� ¼R���( � ��� ��®À ò ³J´ �� 0���( �� ���óô�� 3 ´ �����~ ��� �¾¥õö ��(Ç���µ£ ´ ��&(���%� �:÷�ø�* $�1 ³J´ ��µ�¾�µ� ����ù$ �����%µ�¶£�� ´ ����µ�µ���¹��%�%�� �¾�$��%µ� � ���¹�� �������µ¤¶ ´ �%� ´ ����¾V¹������hú ´ �� £�2������ ��%�������� �$��%µ��� ��� ���)������~&���½¸�� �����������&���)� Ã� Å�Å�¾V�� �� �� �* ï ��¼(ú � ���µ¥-V ��É/º$ �����%�� �µ������¸µ ´ �(����µR�� �$��%µ� � ���¹�� �������µ����%���'¶���$»¹2¾û ´ � µ�¾�µ� ����¥* ³J´ ��&(���%� �8÷ô��µ�µ���¼�� ��$B ����:�� ������%µ��£�� S ��� ´ ��������µ���Ï2� ���~ S�� ;�: ������ ����¾V¹����� 3��� ���� 3 ´ �8�(¹��'&��8 0�� ���µS¶ ´ �%������&(���%� �:÷�ü��( � ���� ´ ��$# ��_� Ã���� �����¾_�����~ ����%�V���V�������� � ����%�~ �¾�����&����+)������%��Ç��J ´ �8÷�ý2$��%µ� � ���¹�� �������µ ¶ ´ �%� ´ ����¾V¹��� Ã� Å�Å�¾B¹�� J����¶���¾�µJ���� � ����%�®* �51V!#��µ� ��� %ý~ ´ ���¸ �������µS��µ���$¸¹2¾V ´ �:µ ´ ���%�®�( ����� ò �� 3����$¸������¾¥�� �øR �¾2����* þ ������ ��� o �¶����2����¹��� �µ�.+ÿ����;1¤�%µ���µ�µ���¼�� ��$_ ���µ���� ´ �� ������£ ����� ���µ����~ �ú î�� �:* � ���$�� ��( ��®)� �* ��* ³ ����$ �� ���µ����®)�!Î* ����µ� ��%�®) þ * ������¹����2�®)���* � ����$��%���( �� �%� ¼# ´ �Vµ����B����������¾É$ ��¼� ����º.0 ´ �V���� � ����%�~ �¾û$ ��¼� ����V�( � ���� ´ ��$Ì ��· ´ � $��� ����� b�%�����%�%���( �������1o����$� ´ ��� ������µ�µ��� �¾�$ ��¼� ����R.0 ´ ������ � ����%�~ �¾_$ ��¼� ���� �( � ���� ´ ��$# ��� ´ �8 ���&��� �µ����%�����%�%���( �������1�* O1 ³J´ �RÇ�� �'¶£����$ ¼���¹���µ���������µ��%µ� �µ��� ��¸������������ �¹���µ�������$º�V ������:¹���µ���* þ ������������ ���½� �����$�µb ´ ��� �� ��������� =&(�( ��%�(¹����£�%�� ´ ��¶���¾_�� �����¾�¹��£���hú ´ �� 3�2������ ��%�������� 3� �� ;����$� ����� ���µ����~ �µo ´ �������� �.Ãó���õR1® ���¼��� ´ �� "¶£�� ´ ´ �:$ �������%�¸�� ¤���%���'¶��(¹�����&(���%� ��µJ 0�� �ó�* þ �2������ ��%�����o������������ �������ú ����%��µ3 ´ �J���'¶��� b����$�� ������ "�%�%���� �µo�� � ´ �J������&��� �µ����� �$��%µ������ �µ���¶ ´ �%��� �¸� ���2ý2�2������ ��%�����"������������ :�������� ��%µ���µ� ´ �_���� ��� ����( ���&��_µ��� 8¶ ´ �%� ´ ó ����%����µ����®* ¼21 ³J´ �V¶ ´ �����¸������������ ������Ì�� � ´ �%µ�µ ´ ���%������$¿����µ� ���¾û ´ �¥Ç�� �'¶£����$ ¼��Oý ¹���µ��:$ ��µ���¼��#�%µ£µ�� ������ � ���$¥¹2¾¥ ´ �:������� ��½ �����¸¹��� �¶������·���� ���$ ����� ��$ ������������ b����$� ´ �£�%$ ������½��� ���µ�µ���$�¶£�� ´ �%�R���� ������%µ��S�� ��8������µ���Ï2� ���~ �* ´ 1 ³J´ �8��½����� � £��µ���µJ ´ �:���( �� ����o�%��� ¼����(¼���¶ ´ ���¸��$��� ��%� ¼��_������������ ��� �R �������* �01 ³J´ ��µ� � ����� �� ����� ¤�_ Ã���� ��������� ��%µ���µJ ´ �Rµ� � ����� �� ����o$ ��µ��� ����� ������¥�� ¤� ������������ ����%��µ: ´ �¸µ��������#���%�������( ���$Ì ��ºµ� ��� ��¸ ´ �¥����µ�µ���¹��%�%�� �¾º$��%µ� � ��hú ¹�� ������®*¥Á�� ��%� ¼¥ ´ �B�%� 0�� ����������� �������µ�µ:�¸ Ã���� R�%µ�������µ��%$ �� ���$º ��#¹�� ò ������ �¾�ø¥�� b�� ´ ��µ�� �� � ���������&���$·¾��� 8�¥$��%µ� � ���¹�� ������®)��� ´ �� �¶£�%µ��R�� ��%µ ������µ��%$ �� ���$V ���¹�� ò Ã���%�%ø�* � 1 ³J´ �SÇ�� �'¶£����$ ¼��S¹���µ��£�%µ"��&��� �¾�����½ ��¹�������� ��*����� ´ ´ �S ������S�������(¼��� ����$º ´ �B������������ :�������(¼��� ����%���'¶£µ����%����µ� :���~¾ÉÇ��%��$·�� J� ´ ��� ¼���µ: �� ¹��B������ ��( ���$»�%�~ ��· ´ �VÇ�� �'¶£����$ ¼��V¹���µ��Vµ� � ����� �� ���* ³J´ ��µ��¥� ´ ��� ¼���µ ����¾#¹��R���� � 0�� �����$#���� ´ �� �¶£�� ´ �%�Í�B ������:�� 8������������ ��� �¶£�� ´ �%�# ´ � Ç�� �'¶£����$ ¼���¹���µ��� ���¼��( �$ ��$û��µ��¸¶ ´ ������* ³J´ ��µ� ´ ��Ç�� �'¶£����$ ¼��_¹���µ�� ����¾�¹��£��½� �����$ ��$_ 0�� �¶��( �$·.0¹2¾���$�$��%� ¼8 �������µ"¹���¾�����$_ ´ �£����$2ý2����&����01 �� J$ �'¶£�~¶��( �$º.0¹2¾V��$�$��%� ¼: �������µ����%������$V¹��� 0�� ��� ´ ��µ� ��( � J����&����01�* þ �Oú ������%��¾�)���� ��¶» ������������R¹�����$�$ ��$R�%�:���~¾����%������¶£�� ´ �%�8 ´ ��Ç�� �'¶£����$ ¼�� ¹���µ���¼�����¹����=µ� � ����� �� ���* Ç�1 ³J´ �%µJµ ´ ���%�=���'&��� �µ���¶£�%$ �8 ���� ¼����� " �����µ������%� ¼�µ�* ³ �_ ������ ´ ´ �%µJ¼������+) �¶��É�%�%� ��µ_.0 �¾2����µO18�� J �����µ������%� ¼�µ��( ��V��µ���$®*���� �_���( � :�� S ´ �B �������µ î�� ï � Å�Å�¾¥��½����� � Jµ�¾�µ� ����ð 0�� �$��%�(¼�� ��µ��%µ �%µ�$ ��µ���¼�� ��$º ��ͼ���� �� ��( ��¥$��%µ� � ���¹�� �������µ� ��Í ´ �B Ã���� �µR���� � ���µ�������$��%� ¼ ��R ´ �������������� �µ� ��� 0�� � ���$B ����%�_ ´ ��������µ���Ï2� ���~ S����$V��� �� ´ �� ����( � S�%µ $ ��µ���¼�� ��$V ��_����$��� 0¾_ ´ �8���� �����$ ¾B¼���� �� ��( ���$#$��%µ� � ���¹�� �������µ�* �01 ³J´ � µ�¾�µ� ���� �%µû�(¹����L ��ô����$ ���_ �¶�� � �� �������µû 0 ���Ï2� ���~ ���¾ ���� Ì�%� ´ �ô����$��%�����������%$®À ´ �ô�%����� ���~ ��%���¸ Ã���� ��� �µF����$ ´ �ô�(¼�¼� ���&(�( �ú �%� ¼��5���%����&��%�( ������B Ã���� ��� �µ�* �B1 ³J´ ���%� 0�� ������������ ¼��%� �������~ � ����=µ� � ��( ���¼�¾¥��µ���µJ�R&(�( ��%���~ ��� ;¹�����Ç2¶��( �$ � ´ ���%���%� ¼R ���� ´ ���%Ï2� ��* ��1 ³J´ ����µ��� ��%�~ ��� � Ã�����������(¹�����µ� ´ ����µ��� � ��R�%� ��� ¤Ç�� �'¶£����$ ¼��� 0 ����ö��� �ú µ��%$ ������$V ��_$��%µ����%��¾� ´ ���������� ���µ����� �µ�.Ã����µ� J¹����%� ¼R$��%�(¼�� ��µ���µO1b ���¼���ú ´ �� J¶£�� ´ ´ ���� S����µ�µ���¹��%�%�� �¾�����$¸� ������ú�µ�µ��� �¾º.Ã���� � ����%�~ �¾�1S$ ��¼� �����µ�* �21����º ´ �_$��%µ�Ç�); �������µ8�( ��B� �� :µ� ��� ���$û�%�º���û��µ�µ�����¹�����$º 0�� ��¥* ³J´ �� �� �( ��¥µ������( ��( ���������µ�¶ ´ �%� ´ µ� ��� ��V ´ �¥µ� � ����� �� �����$ ��µ��� ����� �������µ��� £ ´ � �� ������%µ���µ�)2����$��� = ´ ��������µ���Ï2� ���~ �µ�����$ ������µ¤¶ ´ �%� ´ �����~ ����%�B ´ �� ���½� $ ��µ��� ����� �������µ��� 3 ´ ���¥* ³J´ �Bµ ´ ���%�b$ �����%µ:¶£�� ´ µ���&��� �����$��%�� ��������( �����µ�*���� �B�� J ´ ��� �������� ��%µ���µ ´ ��µ��R ���½� �µ�¶ ´ �%� ´ �( �����µ���$Í���� ´ �� ���µ��� ������%µ���µ£�� 8��µ�������µ���Ï2� ���~ �µ�¶ ´ ��� �V ��������%µ� ��¸¹��R��$��� ���$®* þ �� ������%��¾�)=��$��� ��%� ¼V�V ���������������µ� ��¸���(Ç��_���%�3 ´ � � ������µ�µ��( �¾É������� ��½ ������µ�¹��� �¶������É ´ ��µ�������������µ�* ³J´ �� ��� 0�� ���)=¶ ´ ���Í ´ ����µ��� ��$��� �µ��:� ��¶L �������)2�� ��%µ�� �� �����¶���¾�µ�� ������µ�µ��( �¾� ����� ����( ���� ��¶ ���½� �µ� 0�� � ´ �( �������¹���������µ���)��%�¸�����~¾¸����µ���µ�)��� £�%µ£µ����B�������~ S ��Vµ��������� J��� ���� ����� ��� ���½� �µ 0 ����ð ´ ���(���� ����� ��%�( ���$��%�� ��������( �¾�* ! "$#&%'%)(&*+%¤¬�« ²�ª�«,%¤¬.-0/S±1% þ µb¶�� ´ ��&��������~ ������ ��$�¹��� 0�� ���)~ ´ �JÇ�� �'¶£����$ ¼��J¹���µ��£������µ��%µ� �µ¤�� ®�������������� ¹���µ�������$¸�R �������¹���µ���* ³J´ ����½����� � b�%�~ ��� � Ã�����J�%µ" ´ �( b����$�������¶ ´ �%� ´ ���%���'¶£µ3 ´ �J��½����� � " �����$�$®) $��%µ����%��¾�)b�� �����$��� 0¾û���~¾û�� � ´ �¥ �¶��º¹���µ���µ�*32� �¶��� �Ç�µ���µ��º�������(¼��� � 0�� ¶ ´ �%� ´ ´ �B����Ï2���%µ��� ������º�� J� ��¶öÇ�� �'¶£����$ ¼��B�%µ8��� �B�� J�� �µ�����µ� R�%������ � ����~ ���µ�Ç�µ�* î�î �:* � ���$�� ��( ��®)� �* ��* ³ ����$ �� ���µ����®)�!Î* ����µ� ��%�®) þ * ������¹����2�®)���* � ����$��%���( �� 4)576 8+9�:3;=<?>A@CB�<�@�>�:EDGFH<�9�:EI1J�DLKNMO:LP�QG:3RCS.;�: ,,TVUWTVU X�Y[Z�\�]_^[\�Z�`[a+bdcWe0Z ï ����ô������¼��%�����~�����%�~ o�� �&�����¶8)��������������� 3 ����� ���µ����~ �µ;�������� �.Ãó���õR1= ���¼��� ´ �� ¶£�� ´ ´ �B$ �������%�º�� J���%���'¶��(¹�����&(���%� ��µ8 0�� Ró�)3¶ ´ �� ��_õ �%µ8 ��� 0�� � ��%� ¼¸ ��Í� ���( � ��%�����%�( S���~ ��� �¾¥����$¸óô ��B�R���( � ��%�����%�( S 0���( �� ��8�� ; ´ �( J���~ ��� �¾�* ï ����ð ´ �8��$��� ��� �ñ µS�����%�~ ��� 3&�����¶8)��������������� £�%�����%��$ ��µ�À �~1 þ $ �������� ������¿¶ ´ �%� ´ �:��µ� B¹��·��µV�������� �� ´ ����µ���&��Í��µV����µ�µ���¹�����)� 0 ���ú Ï2� ���~ ���¾B���2��Ç��%� ¼��%��Ç����������������~ J���#ó ����$¸õ�* ¹�1 ³J´ �R���'¶��� �����$·� ������ ��%�%���� �µJ¶ ´ �%� ´ óð����¾# ������ ´ �%�Í����µ����� b�V�2��ú ���� ��%����� ������������ b�� ¤ ´ �£���� ��� ����( ���&��£µ��� ¤¶ ´ �%� ´ ó ������ �����¾:���_�%������µ�� �� 3��� ���2ý2�2������ ��%�����®������������ �* �'1 þ µ��� �����µb�� o�( � � ���¹�� ���µ�¶ ´ �%� ´ ���������&������( � ��%�����%�( �&(���%� ��µb¶ ´ ���B�:� ��¶ ��������%µ���$��� ���$¥�� £$�� ��%� ¼: ´ ���%� 0�� ����������� �������µ�µ�* 2��¿µ����� ������ -()S¶�� ´ ��&��Í��½����%���%� ��$̶ ´ �( _�º Ã���� B ����� ���µ����~ �µ� 0 ����? ´ � �����%�~ S�� ;&�����¶F�� ; ´ �%µJµ ´ ���%�+* ³J´ �� Ã���� �µ��( ����� 3 �¶��_ �¾2����µ�À"��� ���� S Ã���� �µ�����$¥�%� ��� � Ã���� �µ�* þ �¸��� ���� £ Ã���� ��%µ£�Bµ� � ����� �� ���¶ ´ �� ��8 ´ �: ���µ����� J�� ¤�B���%��µ�µ þ �� �Áö�%µ µ� ��� ���$®* þ �û�%� ��� � Ã���� R�%µ��#µ� � ����� �� ��_¶ ´ �%� ´ µ� ��� ���µ: ´ �_ ���µ����� ���� ´ ����&���$ �( 0 ��� 3�(¼�¼� ���¼��( ��%� ¼�µ���&��� ����~��� ���� = Ã���� �µ�*���¹2&�������µ���¾�)' ´ �b��� ���� = Ã���� �µ;µ� ��� �� ´ ����� ���� �$��%µ� � ���¹�� �������µJ�� ��'&��%$ ��$·¹2¾Í ´ ��µ��� �������µ ´ ��&��%� ¼¸�¥������µ���Ï2� ���~ ¶ ´ �%� ´ �����%�~ �µ� ��� ´ �8µ������8 ��( �¼��� ������������� �* î�f ï � Å�Å�¾¥��½����� � Jµ�¾�µ� ����ð 0�� �$��%�(¼�� ��µ��%µ ,,TVUWT�, X�Y[Z�g�hdijZNbdcWe0Z ³J´ �� �������¹���µ����������� ��%µ���µ� �������µSµ� � ����� �� ���$¸�%�~ ��� 0��� £���%��µ�µ���µ�* ,,TVUWT�,,TVUWT k&ijcWe0eml�T ���%��µ�µ þ �����~ ����%��µS �������µS��µJ¹������'¶8À �� onGp ����$_�'�� qn_r ����$_�'�� tsusus�����$_�'�� qnLv ´ ��� .+ÿ����;1xw ����� ´ �� ��������%µS ����� ���µ����~ ���$¸�%� ï ��¼ * � * ³J´ �»�� ������%µ���µ+nLy�zu{|pu�ususus0�}nLy�zu{~v �( ��¿��½��� ���µ�µ���$ ¹2¾ ´ � µ����~ ���������µ nGp0�}n_r��ususus=�}nLv�)£¶ ´ �%� ´ �%� ´ ���� B �� �� �( ��Í ��� 0�� � ��%� ¼Î ��» ´ �É������������ �µ¥���� �ú ���µ�������$��%� ¼: ��+�,������pu�ususus=���,������v�) ���µ������� ���&�����¾�* ³J´ �%µ£ ������R�%µ�������� ���� ���$É ��#�V������������ ������%����$ ò ��( �¼��� ������������� �ø¥¶ ´ �%� ´ �����~ ����%��µ£���¸�( �����¶ ´ �� ��8µ���&��� ����o���%��µ�µ þ �� £Á �������µS�������%$V¹��8������� ���� ���$®* ÈÉ�¥µ���¾Î ´ �( R���%�� ´ ��µ��B �������µV.Ã���%� ´ ��&��%� ¼·������µ���Ï2� ���~ �µR¶ ´ �%� ´ ���%� �� ´ �( J ��( �¼��� ������������� O1S�( ���������� ���� ���$¸ ´ ���� ¼ ´ ´ � ò �����%�¥¶���¾�ø�* þ µ�µ������%� ¼� ´ �( �{ �������µ��( ���������� ���� ���$¥ ��� ´ �%µb ��( �¼��� J ´ ���V ´ �� ���µ����� �� "����� ´ ������8��&(���%���( ������#�%µJ���%������$#�%�#���¸��� ���� £ Ã���� �����$¸ ´ �H�������= ���µ����� �%µV��¹� ����%� ��$ ¹2¾L������¹��%���%� ¼Ì���%�� ´ ��µ��É���( � ��%���� ���µ����� �µ�* ³J´ �N�������� ���µ����� ����� ���µ����~ �µJ ´ ��&(���%� �� ���������&���$¸¹2¾B ´ �8 ��( �¼��� £ Ã���� �* ,,TVUWT�,,T�,,T k&ijcWe0e �+T ���%��µ�µ�� �����~ ����%��µJ �������µS��µS¹������'¶Ê.õ���� ï ��¼ * î 1�À �� �nGp ����$_�'�� qn_r ����$_�'�� tsusus�����$_�'�� qnLv ´ ��� .+ÿ(1 ´ �8���� � ����%�~ �¾¥ ´ �( w �%µS����$�������$B¹2¾��W� î 4 �:* � ���$�� ��( ��®)� �* ��* ³ ����$ �� ���µ����®)�!Î* ����µ� ��%�®) þ * ������¹����2�®)���* � ����$��%���( �� ³J´ �%µ£ ������R�%µ�������� ���� ���$º ��¸�B ��( �¼��� ������������� �¶ ´ �%� ´ �����~ ����%��µ8���É�( ���� ¶ ´ �� ��8µ���&��� ����o���%��µ�µ�� �� ��L �������µS�������%$¥¹���������� ���� ���$®* ÈÉ��µ���¾B ´ �( S���%�� ´ ��µ��� �������µ�.Ã���%� ´ ��&��%� ¼R����µ���Ï2� ���~ �µS¶ ´ �%� ´ ���%�Ê ��� ´ � µ������8 ��( �¼��� ������������� O1S�( ���������� ���� ���$¸ ´ ���� ¼ ´ ´ �8µ���������$��( �¾V¶���¾�* ³J´ ��¼������b�� S�¸���%��µ�µ�� ��������%µ8� �� 8 ��#¼���� �� ��( ��V$��%µ� � ���¹�� �������µ£ 0�� � ´ � Ã���� �µ�¹�� 8 ��Í���� :� �����ºµ������_�� � ´ ����� ¶ ´ �%� ´º´ ��$Î���� �����$ ¾· ���������&���$Î� &(���%� ��� �%���� �$ �� b ��:�%���� �����µ��J�� �$ ���� �����µ��£ ´ ������ � ����%�~ �¾��� ® ´ �J ��( �¼��� � Ã���� .0 ´ �( ���������µJ ��B$ ���� �����µ����� ��%���� �����µ��8 ´ ���������� � ����%�~ �¾¥����&����=�� 3 ´ �8 ��( �¼��� Ã���� £����µ�µ���¹��%�%�� �¾�$��%µ� � ���¹�� �������1�* ,,TVUWT�,,T��.T k&ijcWe0e�kmT?���%��µ�µ��L�����~ ����%��µJ �������µS��µJ¹������'¶8À �� onGp ����$_�'�� qn,� ����$_�'�� �susus£����$_�'�� qnLv ´ ��� .+ÿ����;1 ó�����-��bó������ ����� ´ �� ��������%µS ����� ���µ����~ ���$¸�%� ï ��¼ * f * ³J´ �%µ� ��������%µS������� ���� ���$¸ ��� ´ �8 ��( �¼��� £ Ã���� £�%�V ´ �8µ������8¶���¾¥�����%��µ�µ�� ������8�%µ�* f�� ï � Å�Å�¾¥��½����� � Jµ�¾�µ� ����ð 0�� �$��%�(¼�� ��µ��%µ � ���%��Ç������%��µ�µ þ �� ��¿ �������µ�)2 ´ ��������µ���Ï2� ���~ ��� ;�R���%��µ�µJ�¿ �������$ �2��µ�� �� �����~ ����%�º�B����µ�µ���¹��%�%�� �¾¥$��%µ� � ���¹�� ������®* ³J´ �:¼������"�� b ´ �%µ� ������R�%µ� ��¸µ ´ �� 0 � ´ � µ���¼������������~ S���( � S�� 3 ´ ������µ�µ���¹��%�%�� �¾�$��%µ� � ���¹�� ������®* ,,TVUWT�,,T��GT k&ijcWe0e  |T ���%��µSÁô�����~ ����%��µJ �������µS��µS¹������'¶8À �� onGp ����$_�'�� qn,� ����$_�'�� �susus£����$_�'�� qnLv ´ ��� .+ÿ����;1 ó����¡��-���ó�-C����� ����� ´ �� ��������%µS ����� ���µ����~ ���$¸�%� ï ��¼ * 42* ³J´ ��������µ���Ï2� ���~ S�� 3µ���� ´ �R �������$ �2��µS� �� J�����~ ����%�¸�R����µ�µ���¹��%�%�� �¾�$��%µ� � ��hú ¹�� ������®*£����� ´ �� ��������%µ£������� ���� ���$# ��B ´ �� ��( �¼��� � Ã���� ��%��Ç����_���%��µ�µ þ �������) ´ �� ��� 0�� ��8 ��� ´ ���( ��������� � ���µ�������$��%� ¼R ��� ´ �8�����%�V¶���¾�* ³J´ ��¼������¤�� � ´ �%µ� ��������%µ�µ��%���%�%�( � ��¸ ´ �( 8�� S�¸���%��µ�µ:� �������)®¶£�� ´ ´ � $��hË��� ��������S ´ �( ¤� �'¶¿ ´ �J����µ�µ���¹��%�%�� �¾�$��%µ� � ���¹�� ������8�� � ´ �S ��( �¼��� b�%µ3��¹� ����%� ��$ ¹2¾ûµ ´ �� 0 ��%� ¼¸ ´ �¥µ���¼������������~ :���( � :�� ��·$��%µ� � ���¹�� ������º���� � ���µ�������$��%� ¼# ��É��� ��$�$��� ����������®µ���� ����8������������ �* 4)5j4 ¢)D[£�:3B�D[JCB�M7@C;�¤OD[JC;�D[J�<�9�:3>A@CMO:WRCS.;�:3;=<?>A@CB�<�@�>�: ���%��µ�µ þ ����$:ÁÌ �������µ ´ ��&��S�J¼���� �� ��( ��%� ¼���Ë����� 3�� ��'&��%$��%� ¼S����µ�µ���¹��%�%�� �¾�$��%µ� � ��hú ¹�� �������µS ��_ ´ �8 ��( �¼��� � Ã���� �µ� ��( � ��%� ¼����� ´ �� £ 0 ����ù ´ �8����µ�µ���¹��%�%�� �¾_$��%µ� � ���¹���ú ��������� ´ ��������µ���Ï2� ���~ S.Ã�%������µ����� ������%��µ�µ þ ������51��� ; 0 ���� ´ ��$��%µ� � ���¹�� ������ �� b���Í��$�$��� ����������oµ���� ����� Ã���� �.Ã�%�#����µ��:�� b�B���%��µ�µ�Á ������51�*������ ´ �������µ£�( �� ������� ���� ���$V ��: ´ �� ��( �¼��� S������������ � ´ ���� ¼ ´ ´ �������%��¶���¾B����$B���� � ���µ�������$ ��_ ´ ���� �µ� J�%�%� ���� 3 �����µ������%� ¼ * f - �:* � ���$�� ��( ��®)� �* ��* ³ ����$ �� ���µ����®)�!Î* ����µ� ��%�®) þ * ������¹����2�®)���* � ����$��%���( �� ���%��µ�µ � ����$º�F �������µ£���� 8� �����#���� �����$ ¾¸¼���� �� ��( ���$É$��%µ� � ���¹�� �������µS����$ �( ���������� ���� ���$¸ ��� ´ �� ��( �¼��� J ´ ���� ¼ ´ ´ �8µ���������$��( �¾B¶���¾�)����� � ���µ�������$��%� ¼ ��_ ´ �8µ���������$#�%�%� ���� 3 �����µ������%� ¼ * 2�� ����µ��º�� R�Ì����$��%������$��%�(¼�� ��µ��%µB Ã� Å�Å�¾L��½����� � ¥µ�¾�µ� ���� ¹����%�%$ � � ¹2¾ ��������µ��� � ´ �%µ�µ ´ ���%�+)o���%��µ�µ��ô �������µ8�( ��_�(¹����� ��·����$ ���" ´ �_���� ������Î�� S ´ � �(¼�¼� ���&(�( ������_�5���%����&��%�( ������# Ã���� ��� �µ�* ³J´ �%µ£µ ´ ���%�o�����É���%µ��B¹��:��µ���$# ��V¹����%�%$¸� �Í�� ���$��%�� ������2ý~�� ������~ ���$#��½����� � µ�¾�µ� �����µ£�%�V¶ ´ �%� ´ ���%��µ�µ��L�� £Á �������µ����%��¾V�������~ � ����= �������* 4)5j¥ 8+9�:3I1J�DLKNMO:LP�QG:�¦)RCS.;�:�£�S.J�S,QG:W> þ ������ �$��%� ¼� ��� ´ �8¶���¾¥ ´ �8Ç�� �'¶£����$ ¼��8¹���µ��8¶���µ£µ� � ����� �� ���$®)��� �µ£�������(¼���ú �����~ ��� ���¼� ���� �%µRµ� � ����� �� ���$»�%�~ ��É�Í ������_¹���µ��¥�������(¼��� �����$Ì�·������������ ¹���µ����������(¼��� �* ³J´ �V������������ _�������(¼��� _���%���'¶£µ: ´ �¥��½����� � R ��É��$��� �� ��¶ð������������ �µ�)� �� $��%µ����%��¾_�� J ��B����$��� 0¾� ´ ���¥* ���������� ����%� ¼� ´ �:����$��������( �������µS�'&��� 8�_������������ �)�¶��R�����~ ������¸ ´ �( � ´ � ��½����� � �����¾¥������ ��( ��:� ´ ��� ¼���µ£¹��� ´ �%�¥ ´ ��$ �������� ������V ���½� ��� 3 ´ �:������������ ����$Î�%�É ´ �B���� ��� ����( ���&��Bµ��� �* ³J´ ����½����� � ��:��µ� 8 ��(Ç��V���( ��B� �� � ��Í����$��� 0¾ ´ �Í���������%� ¼É�� � ´ �#������������ _¹���������µ��#�%�Ì ´ �%µ�����µ��Í���%�S ´ �Í������� ���� �������µ ¹��� �¶������· ´ ��µ��R �������µ£¶ ´ �%� ´ ��� 0�� � ��V ´ �%µ������������� 8����$# ´ �������������� ��� �µ����� �( ���$������(¼���$®* �������"��$��� ���$®)' �������µ®����¾�¹��"µ�� ¹�µ���Ï2� ���~ ���¾�����$�������$��� ~ ´ �¤��½����� � =���(Ç���µ µ���� ´ ��$ �����%µ������®* ³J´ �: ���½� 8����$_�'�� � ´ �R����µ�µ���¹��%�%�� �¾V$��%µ� � ���¹�� ������¸��µ�µ���¼�� ��$Í ��¸�B�� ������%µ�� ����¾V¹��8� ´ ��� ¼���$#����$ �� S �¶��B������µ� � ����%�~ �µ�À § ´ ���� ������%µ����:��µ� £� �� J����µ����� �µ��� ���&�������µS���������%� ¼�¨ § ´ ������ � ���µ�������$ �������£¹��� �¶������B ���½� �����$Bµ� � ����� �� ����:��µ� �¹��������%� ����hú � ��$®* ³J´ ����½����� � £����¾V���%µ��_$ ������ ��8���� ������%µ����� ���$�$¥��� ��¶F��� ��* ³J´ �·������µ���Ï2� ���~ ¸����¾L¹��·� ´ ��� ¼���$F�%�L ´ �ɵ������ɶ���¾ �»�� ������%µ��Í�%µ�) ¹�� �� �'¶8)������%��Ç��B ´ �¸� ´ ��� ¼���µ������� ��( ���$Ì���¿�É�� ������%µ���)b�� ��%µ����%���'¶���$Ì �� ������%�����# ´ �Í������µ���Ï2� ���~ _¹2¾Ì�Î������������ �����¾¿� ��¶ ��� ��* ³J´ �%µ������� ��( ������ �%µ f � ï � Å�Å�¾¥��½����� � Jµ�¾�µ� ����ð 0�� �$��%�(¼�� ��µ��%µ �����%����$V ���$��� ����� ��( ������¥����$V�� S�%µ���¹2&�������µ� ´ �( S 0 ����Æ� �'¶F���V ´ ��������µ���Ï2� ���~ ¶£�%�%�� ��� 0�� J ��_��� �� ´ �� ������������� �* È �� ´ �%�#�V���%��µ�µ8�F �������)� ´ �: ��( �¼��� 8������������ _.Ãó:1£����¾#¹��: ������%������$͹2¾ ��� �� ´ �� �* È �� ´ �%���£���%��µ�µoÁÌ �������)'���� ´ �� o ´ �b ��( �¼��� S.Ãó���1��� ; ´ �b��$�$��� �����������µ���� ���� ������������ :.Ãó�-'1�����¾V¹��� ������%������$®* ³J´ �����2���B�������~ �µ©�_����-���������ÿ����û����¾¥���%µ�� ¹��8����$�������$®* þ ������·����¾L¹��Í ��������� ��( �¾ ò $ ������ ���$�øÌ�%�L ´ �·¶���¾ �� ¥�%µB ��������� ��( �¾ $��%µ�������� ���� ���$¥ 0 ����ù�� �µS�� ���&�������µ� ��( �¼��� �* ª "$#&%ù©�ª&-0%¤¬.%bªR±1%¬«»ª&­�²�ª&% ¥)576 ®�@CMO:E¯�°,S.M7@�S_<�¤OD[J �.TVUWTVU ±Hg�Z�²E³je0Z�`dcWg=amZA´�cLi}h[c�a�³j]_^ µ �� �¶.·�.j¸L·�1b$ ��� �� ��� ´ ������µ�µ���¹��%�%�� �¾:$��%µ� � ���¹�� ��������� = ´ � ¹3�� ������%µ��J�� ;�� �������) ����$�¶®ü · .j¸L·�1b ´ ������µ�µ���¹��%�%�� �¾R$��%µ� � ���¹�� ��������� ; ´ �( S�%� ��� b Ã���� S¶ ´ �%� ´ �%µ� ��R¹�� ���������( ���$û¶£�� ´ ´ �%µ��� ������%µ���*� £�� ��+¸L·»º½¼d·�)"¶ ´ �� ��¾¼d·£�%µ8 ´ �V������&��� �µ�� �� �$��%µ������ �µ��:�� � ´ �������������� � ��� 0�� � ���$· ��¥¹2¾# ´ ���� ������%µ���* ³J´ �R$ ��¼� ������� ���( �� ´ �%� ¼_�%µ���½��� ���µ�µ���$¥¹2¾B �¶��B�2����¹��� �µS$ ����� ��$¥��µJ¹������'¶8À ¿HÀ ÿ0·ÂÁ ���5½ Ã=Ä ���%�É.j¶.·�.j¸L·�1���¶ ü· .j¸L·�1�1 .�-'1 Å z0��·¡Á ���%� Ã=Ä ���5½º.j¶.·�.j¸L·�1���-�Æt¶ ü· .j¸L·�1�1 .V��1 ¿HÀ ÿ0·3µ ´ �'¶£µS ���¶ ´ �( J��½� ����~ J ´ �8�� ������%µ����%µS����µ�µ���¹���������$ Å z0��·"µ ´ �'¶£µ ��B¶ ´ �( ���½� ����~ � ´ �:�� ������%µ����%µ£���� � ����%�#¶ ´ ���Í���������( ��%� ¼_�� £¶£�� ´ �_¼���&���� �%� ��� � Ã���� �* 2� o ´ ��µ��8&(���%� ��µJ$ ��� �� £µ��( ��%µ� 0¾ ���5½º. ¿HÀ ÿ0·Ç��-�Æ Å z0��·�1�Á - .VÈ�1 ´ ���¸ ´ ��¾B¶£�%�%��¹��8� �� ������%��Å���$®* ³J´ �����~ ��� ��S�� ������%µ������( � 3¶£�%�%�~¹��� ´ �� ��� 0�� ��J� ´ �( ����� ��� ���Å���$�¹2¾: �¶��8�2���Rú ¹��� �µCÉ ¿HÀ ÿS����$mÉ Å z0�5)~��������� ���$��������� �$��%� ¼� ��� ´ �S ������� �¾2����*CÉ ¿HÀ ÿS����$ É Å z0������¶���¾�µ£µ��( ��%µ� 0¾�À ���5½º.OÉ ¿HÀ ÿ���-�Æ3É Å z0��11Á - .0/21 f È �:* � ���$�� ��( ��®)� �* ��* ³ ����$ �� ���µ����®)�!Î* ����µ� ��%�®) þ * ������¹����2�®)���* � ����$��%���( �� �.TVUWT�, k&ijcWe0e l �~1�2� �É ¿HÀ ÿNÁ -V����$ÊÉ Å z0��ºö. � ��-�Ë£ ´ ���» ´ �V ������B�%µR������µ��%$ �� ���$û �� ¹�� ò ���� ¼ ´ ��¾�ø·µ��( ��%µÇ����$Ì����$û ´ �¥��&(���%���( ������» ���µ����� R�%µ:��¹� ����%� ��$»¹2¾ ��µ�µ���¼����%� ¼��_���( � ��%�����%�( £�������� � ����%�~ �¾#����&����o ��_ ´ �R������µ���Ï2� ���~ �$��%µ� � ��hú ¹�� ������¥�������� �$��%� ¼� ���É Å z0������$#ÿB. ï ��¼ *%- � );-�-();-0������$º-0È�1�* ¹�1+2� [É ¿HÀ ÿ»Ì -:.0 ´ ���_� ������µ�µ��( �¾+É Å z0��Á � 1[� ´ �£ ������J�%µ¤������µ��%$ �� ���$ ���¹�� ò ���� ¼ ´ ��¾�ø�� �� Jµ��( ��%µÇ����$®*12��¥ ´ �%µS����µ���À § �� ,�|Í � )� ´ �£ ������J���� � ���µ�������$��%� ¼� ��: ´ �£ ���&��� �µ����%�����%�%���( ��������%µ ��&(���%���( ���$®)� ´ �( £�%µ �� 1ÎLnGpS�� »ÎLn_r��� tsusus£�� �ÎLnLv_ ´ ���Î.j�1��ÿ(1�Î)w . ´ �� ��B ´ �V������� ���� ���&��¸�%�Î ´ �V�%���� ��%���" ������B�%µ:������µ��%$ �� ���$Î ��·¹�� þ ��Á81�* ³J´ �Í ���µ����� ���$ $��%µ� � ���¹�� ������¿�%µB��¹� ����%� ��$L¹2¾L�%�~&��� � ��%� ¼ f / ï � Å�Å�¾¥��½����� � Jµ�¾�µ� ����ð 0�� �$��%�(¼�� ��µ��%µ ´ ���%���� ��%���G¶GÏOÐ�v?Ñ7Ò�Ó�)® ´ ���É��µ�µ���¼����%� ¼V���É�������� � ����%�~ �¾Í����&����3 ��¸�� �������� �$��%� ¼� ���É ¿HÀ ÿ�����$�� . ï ��¼ *%-�/_����$º- � 1�¨ § �� ��3Á � ´ �R ���µ����� ���½��� ���µ�µ���µ��V Ã���%�o�������� � ����%�~ �¾Ì. ï ��¼ *%- � ����$ - î 1S���������%� ¼� ´ �( ����%�® ´ �8&(���%� ��µ£�%�¥ ´ ��������&��� �µ��8�� "$��%µ������ �µ�� �( �����Ï2�����%��¾B����µ�µ���¹����?¨ �'1+2� ,É ¿HÀ ÿ�Á -�����$mÉ Å z0��Á � ´ �S ���µ����� ;��½��� ���µ�µ���µ3�� Ã���%�2�������� � ����%�~ �¾ . ï ��¼ *%- � ����$É- î 1�* �.TVUWT�� k&ijcWe0eÔ� �~1�2� �É ¿HÀ ÿ Á -8����$~¶GÏOÐ�v?Ñ7Ò�ÓJ����$~¶ pÕ�ÖØ×ÇÙ �( ������������( ���¹�����) ´ �( £��������µ�À ���5½F���%�º.j¶GÏOÐ�v?Ñ7Ò�Ó0��¶ pÕ�ÖØ×ÇÙ 1�Á -�� ´ ���¥ ´ �8�������� � ����%�~ �¾¥����&����®�� 3 ´ �� ��( �¼��� ��%µS����$�������$¥��µS 0���%���'¶£µ�À f?� �:* � ���$�� ��( ��®)� �* ��* ³ ����$ �� ���µ����®)�!Î* ����µ� ��%�®) þ * ������¹����2�®)���* � ����$��%���( �� § �� )�+Í � )��� J�%µJ$ ���� �����µ���$.¨ § �� )�+Ì � )��� J�%µJ�%���� �����µ���$.¨ § �� )� Á � )��� J�%µJ� �� J����$�������$®* ¹�1+2� ÚÉ ¿HÀ ÿ�ÁÊ-:����$N¶GÏOÐ�v?Ñ7Ò�Ó�����$N¶ pÕ�ÖØ×ÇÙ �( ���� �� ����������( ���¹�����)� ´ ���Í ´ ��������� � ����%�~ �¾_����&������%µ�µ ´ �� 0 ���$B�%�~ ��� ´ ����������µ��� ���$��� ����� ������·.Ã�� [��Í � �%µ �%���� �����µ���$®)��� )�+Ì � �%µJ$ ���� �����µ���$�1�* �'1+2� �É ¿HÀ ÿ»Ì -� ´ �� ������ ´ ��µJ� ����Ë����� �)L¶ pÕ�ÖØ×ÇÙ ��������%���%� ¼R��������$�������$®* ï ��¼�� ��¥- f µ ´ �'¶£µ� ´ �� ��( �¼��� �$��%µ� � ���¹�� �������µ£¹��� 0�� �������$º�( 0 ��� � ´ �� ������ ��&(���%���( ������®*¤ÈÉ����µ�µ�������$NÉ ¿HÀ ÿ Á -(* �.TVUWT�� k&ijcWe0e�k �~1�2� �É ¿HÀ ÿ�Á -� ´ ���¸ ´ �8µ ´ �� 0 J�%µS���� � 0�� �����$¥�������� �$��%� ¼� �����-8����$����2) ����$¥��� ��¶ �������� � ����%�~ �¾¥����&����=�%µS��µ�µ���¼�� ��$¥ ��� ´ �� ��( �¼��� �* ¹�1+2� �É ¿HÀ ÿ~Ì -� ´ ���É ´ �Bµ ´ �� 0 8�%µ8� �� 8���� � 0�� �����$®)=¹�� ��¸� ��¶Ê�������� �ú ����%�~ �¾¥����&����=�%µSµ� ��%�%�®��µ�µ���¼�� ��$V ��� ´ �8 ��( �¼��� �* Û ]Wa=Z�À1�£�� ������%��¾�)2�%�_ ´ �%µb����µ���) ´ �� ������£���� � ���µ�������$��%� ¼8 ��: ´ �� ���&��� �µ�� �%�����%�%���( ������Vµ ´ �����%$ ´ ��&��8¹������¥��&(���%���( ���$®* ³J´ �( J��������µ�À �� ÎLnGpS�� �ÎLn_r��� tsusus£�� �ÎLnLv ´ ��� .j�1��ÿ(1ÜÎS.Ãó��¡��-���ó�������1 f?� ï � Å�Å�¾¥��½����� � Jµ�¾�µ� ����ð 0�� �$��%�(¼�� ��µ��%µ �� J ´ �¸�%���� ��%���b������� ���� ���&��¥¶���µ þ ��Á:* ³J´ �%µ�¶������%$»�%������¾· ´ �¸� ��¶Æ ��( �¼��� $��%µ� � ���¹�� ������Í ��#¹�����¹� ����%� ��$º¹2¾Éµ ´ �� 0 ��%� ¼V ´ �_µ���¼������������~ ����( � 8�� � ´ �_���%$ $��%µ� � ���¹�� �����������$8 ´ ���8¹2¾8�%�~&��� � ��%� ¼£����$8��µ�µ���¼����%� ¼J�£� ��¶û�������� � ����%�~ �¾8����&���� ��¥�� �*»�£��&��� � ´ ������µ�µ�)®¶�� ´ ��&����� ��� 0�� � ���$¸ ´ ���%�~ ��� ��� ��� ��( ������#��½��� ���µ�µ���$͹2¾ ¹�1�������µ��%$ �� ��%� ¼º ´ �É��µ������£���������%� ¼û�� 8 ´ �Í �������* ï ��¼�� ��»-�4ûµ ´ �'¶£µB ´ � ��( �¼��� £¹��� 0�� ��8����$¸�( 0 ��� £ ���������&(���%���( ������®)���µ�µ������%� ¼mÉ ¿HÀ ÿ�Á -(* �.TVUWTÞÝ k&ijcWe0eÔ  þ ���%��µ�µ"Á ��������%µ;��&(���%���( ���$Rµ��%���%�%�( ���¾� ��������%��µ�µb�º ������b¶£�� ´ ´ �S$��hË��� �������� ´ �( �� �'¶ß¶ Õ�ÖØ×ÇÙ �%µ"��¹� ����%� ��$�¹2¾Rµ ´ �� 0 ��%� ¼�.Ã�� �� ������µ�µ��( �¾�1; ´ �£µ���¼������������~ "���( � �� "���¸��$�$��� ����������®µ���� ����8$��%µ� � ���¹�� ������®* ¥)5j4 8+9�:Êà+D[J[<?>�D[M»¢d<?>�S_<?:WQ.á ³J´ ���%� 0�� ��������R��� ¼��%� ����µ���µ��B¹�����Ç2¶��( �$º� ´ ���%���%� ¼B&(�( ��%���~ � ��¥���� � 0�� �� ���%� ´ �� ���Ï2���� ���$Í������ ��( �������µ� 0�� ��%� 0�� � ��%� ¼_ ´ �&�������; ���µ����� �µ�.Ã�%�·��� 8����µ���)3$��hú �(¼�� ��µ���µO1�* ³ � �������������%�%µ ´ ´ �%µ# ���µ�Ç�):���%�� ´ �û ��� ����%�����8 Ã���� �µÌ.0�� ·������¾F� �� ��Oý µ��������� ���$»���( � ��� J ´ ���B18�( ��V$ ��µ���¼����( ���$û��µ:¼������%µ�)¤¶ ´ �%� ´ ¶£�%�%�"¹��_ Ã� � ´ �� ��&(���%���( ���$R��� ��¹2¾8��� ��* þ 0 ��� "µ��������� ��%� ¼��£¼������+)� ´ ���%� 0�� ��������b��� ¼��%� ��¹����%�%$�µ �8 � ���� ´ ��&��%� ¼8 ´ �( b¼���������µ��8 ����=* ³J´ �%µ" � �����������µ��%µ� �µ"�� ® �������µ"������$ ��$� 0 ���� ´ �R$��%µ�ÇV����$# Ã���� �µ�¼���� �� ��( ���$͹2¾V ´ �R���� � ���µ�������$��%� ¼������������� �µ�¶ ´ �%� ´ �( �� ���%µ��8������$ ��$R 0 ���� ´ �J$��%µ�Ç�* þ �� ������%��¾�)� ´ ��¹�����Ç2¶��( �$�������$��%� ¼�����$�� ´ ���%���%� ¼ f�î �:* � ���$�� ��( ��®)� �* ��* ³ ����$ �� ���µ����®)�!Î* ����µ� ��%�®) þ * ������¹����2�®)���* � ����$��%���( �� ���¸��� � ´ ����$¥����$¥ ´ �� �������µ���&(���%���( ������V���¸ ´ ���� ´ �� ´ ����$®)��( ��8µ������ ´ �'¶ �%�~ ��� ��%������$®* þ � ´ �¸¹���¼��%�����%� ¼ )¤�|�� �µ� ����( ´ �%µR¹����%�� Rµ� ��( � ��%� ¼º 0 ���� ´ �¸µ��������� ���$ ¼������b¹2¾·�#$ ���� ´ ý~¹�����Ç2¶��( �$Î������$��%� ¼V�� � �������µ�����$Î������������ �µ�)3���~ ��%�"�¥ ������ ¶ ´ �%� ´ �%����µ����º ´ �_¹���µ��Oý2����&������%µ������������~ ��� ���$®)3 ´ �( :�%µ��¥ �������¶ ´ �%� ´ ¼��� �µ �� �µ£�%� ��� S$��( ��� 0 ����ð��� �µ��%$ ��* þ 0 ��� £���~ ��� ��%� ¼� ´ �8 ���Ï2���� ���$¥$��( ��� ´ �%µS ������ �%µ���&(���%���( ���$®)� ´ ���V ´ ���� �������µ�µJ�����~ ��%�2� ��µS¹2¾B�( � ������� ��%� ¼� �����&(���%���( ��8 ´ � �%������$��%�( ��8� ��½� S �������*12� 3���%�®�� �µJ�%� ��� �$��( ����( ��8�� ������( ���$®) ´ �( £��������µJ���%� �� �µ��%� ��� � Ã���� �µ��( �� ò Ã���%�%ø�)" ´ �%µR ������V�%µ:��&(���%���( ���$.¨S�� £� �� �)b ´ �¸�%� 0�� �������� ��� ¼��%� ��¹����%�%$�µ���� �� ´ �� £���( ´ µ� ��( � ���$¸ 0 ����ð ´ ���� �µ� ò ������ �¾�ø_�%� ��� � Ã���� �* 2��· ´ �%µ�¶���¾�)o ´ ��¹�����Ç2¶��( �$Î������$��%� ¼V����$É ´ �� 0�� �¶��( �$·��&(���%���( ������É�( �� �%�~ ��� ��%������$¸���~ ��%�� ´ �»�������� ��( �¼��� R.0 ´ ��¼������01� ���������&���µ£��&(���%� ��* â "$#&%'ãåä_%¤¬ ©�ª�«,%¤¬.-0/S±1% ³J´ �8��µ��� £�%�~ ��� � Ã����� ´ ��µS �¶��B��� � �� J¼������%µ�À § ��¥�����(¹����R ´ ����µ��� 8 ��¥���~ ��� :���%�; ´ ���%� ��� �$��( ��¥ ���Ï2���� ���$͹2¾# ´ ��µ�� �������µ�¶ ´ �%� ´ �%�������¥ ´ �8¹���µ��8����&����V¨ § ��:$��%µ����%��¾R ´ ��������� ���µ����� �µ�)2�%����� �����µ���)� ´ ��$��%�(¼�� ��µ���µb 0������$� ���¹�� � �� �V.0 0�� �¶ ´ �%� ´ ´ ������µ�µ���¹��%�%�� �¾B$ ��¼� ���� ¿HÀ ÿ�Áö-:����$¸ ´ �R���� � ����%�~ �¾ $ ��¼� ���� Å z0��ºû. � ��-�Ë�* æ "$#&%'«t(&*ÔçØ/Jª&/�«�­£¬,èêé ­8¯�°ÔçØ% ³J´ �b��½����%�����( ��� �¾8����$������b���%���'¶£µo ´ ����µ��� ; ���Ç������R � �����Ç ´ �'¶û ´ ���%� 0�� �������� ��� ¼��%� � ´ ��µS ������ ´ ��$¸ ´ �»�������®���������%��µ������®* ë ìƲØä�±¤°&ä_ä�²�­�ªê/JªR¯ ±"­�ªR±�ç�°&ä�²�­�ª&ä ��¾Æ��µ��%� ¼ ´ �%µ»µ ´ ���%�+)V��� �F�����ƹ����%�%$Ê� �Æ��½����� � ̵�¾�µ� �����µ¿�����%����¾ô 0�� $��%�(¼�� ��µ��%µ�ý~�� ������~ ���$¸�(�����%�%���( �������µ�* f?f ï � Å�Å�¾¥��½����� � Jµ�¾�µ� ����ð 0�� �$��%�(¼�� ��µ��%µ 2� �µ���½����� � _�%�~ ��� � Ã�����#���%���'¶£µ� ´ �¸��½����� � � ��ι����%�%$»� � �º&��� �¾3����½ ��¹���� Ç�� �'¶£����$ ¼���¹���µ���������µ��%µ� ��%� ¼V�� � �������µ�����$É������������ �µ�* þ �%����µ� 8���~¾ÍÇ��%��$·�� ����$��������( �������µS�����¸¹��������� ��( ���$#�%�~ ��� ´ �%µS¹���µ���* ³J´ ���%� 0�� ������������ ¼��%� ���%µ��(¹����� ����� �������µ�µ� 0��� S���%��µ�µ���µ��� ;�� %ý~ ´ ���B �������µ�) ����µ� ¤�� � ´ ���ô�� ���µ����~ ���$��%�R��� ò �� �����$R������¾:�� �ø� 0�� ����( �*1���� ´ ´ �S�� ������%µ�� ���( � J����$V ´ �8������µ���Ï2� ���~ £����¾B¹��� Ã� Å�Å�¾�)�¹�� S���� � ����%�®*12��¥��$�$��� ������®)���µ�� �ú ������������¾É����$º�¸� ������µ�µ��� �¾É$ ��¼� ����_�( ��_��µ�µ���¼�� ��$É ��# ´ �_$��� ����� 8����$É ���&��� �µ�� �%�����%�%���( �������µ�)2 ����� ���µ����~ ��%� ¼R ´ �8���� � ����%�~ �¾V���2���B�������~ �µ£�( � ���� ´ ��$¸ ��� ´ �8$��hú ����� _����$Ì ´ �¥ ���&��� �µ��¸ �������)b ���µ������� ���&�����¾�* ³J´ �¥ Ã���� �µ_����¾»¹��V Ã� Å�Å�¾»����$ 0 ���Ï2� ���~ ���¾V�������� � ����%�®* ³J´ �%µJµ�¾�µ� ����¥)���&����·�� "�� £¶���µ��%���� ��%���%��¾B$ ��µ���¼�� ��$¸ ��m�� �$��%�(¼�� ��µ��%µS�� ���¹ ú ������µ�)�¶������%$:¹��Sµ���������µ�µ� Ã���%��¾8��µ���$R ���¹����%�%$:� �R�� ���$��%�� ������2ý~�� ������~ ���$:��½����� � µ�¾�µ� �����µ�* í %�-0%¤¬.%bªR±1%�ä î -�ËÔ�:* � ���$�� ��( ��®)£!Î* ����µ� ��%�®) þ * ������¹����2�®)���* � ����$��%���( ��®)���* ³ ����$ �� ���µ����®* ���û ´ � ³ ����( ������~ R�� � ������ � ����%�~ �¾û����$E2����� ������%µ������Î�%�û���ÌÂ"½����� � �2¾�µ� ���� 0�� 2��������� ������~ £Á��%�(¼�� ��µ��%µ�����$N�¤ ���$��%�� ������®) ï � Å�Å�¾·�2¾�µ� �����µ ����$ þ � ��������%���,2��~ ����%�%��¼���������) þ � þ $ &(��������$ ³ ��½� Ø���2��Ç�)L2��Wïµ��+)=-�4�4�-(* î �=Ë� �* ��* ³ ����$ �� ���µ����®)��:* � ���$�� ��( ��®)5!Î* ����µ� ��%�®)~��* � ����$��%���( ��®) þ * ������¹����2�®) :* �2 ���&2 ��+)�2O* ���� �� ����)�ð:* ������µ� ��%� ��µ����®)J��* ³ ����$ �� ���µ����®* ï � Å�Å�¾ Â"½����� � �2¾�µ� ���� 0�� ) £���( ��%� ¼ µ ��µ�µ)ñ£�%µ�Ç þ µ�µ���µ�µ��%� ¼ )?�¤ ���������$��%� ¼�µ=�� � ´ � ï ��� � ´ þ ���2������!#���� ��%� ¼��� G����������$��%����� ï � Å�Å�¾B�2¾�µ� �����µ þ µ�µ������%�( ������®)A2���Å�� Ç(��) ���(�����®)�!Í��¾ î ú f );-�4�4?�2* î È=Ë:Á:* Á�� ¹����%µ�)? �* �¤ ���$ ��* ³J´ ���� �����$ ��µd�;��µ�µ���¹��%�%�� ���µ�* þ �����%�%���( �������µo���%�J ���� ú ���µ����~ ��( ������:$ ��µo�����������%µ�µ���������µ=�����%� 0�� ����( ��%Ï2� ��)'!Í��µ����®)?�3�( ��%µ�)�-�4 f?f * î /�Ë:Á:* Á�� ¹����%µ�)� �* �¤ ���$ ��* ³J´ � ³ ����( ������~ S�� � ������ � ����%�~ �¾B�%�+ò8� �'¶£����$ ¼���ú ����µ���$É�2¾�µ� �����µ � µ��%� ¼ ï � Å�Å�¾Í�2�� �µ£����$��;��µ�µ���¹��%�%�� �¾ ³J´ ���� �¾�)W2��~ ��� ����5ú ����������o����� ��������� d2��~ ����%�%��¼����~ ��2¾�µ� ����¥) &����+* È2)�� � * �2)����������� 8-�4 f?f * î � ËR� * µ ��¹����%�%��¾�))ñ8* !Í�( � ��%��ú������������� ���)) �* �¤ ���$ ��* � µ��B�� ï � Å�Å�¾ µ ��¼��%�_�%�Î� ñ£������ú7����µ���$»�2¾�µ� ���� �%�t�;�� � ����%�%���ó ���������¼�¾�) µ ��� ¼��(¼���µ:�� ��2¾�µ� �������µ 2�� 0�� ����( ��%Ï2� ��µ�) ï ��&2 ����� �)=-�4 f�î * f 4 �:* � ���$�� ��( ��®)� �* ��* ³ ����$ �� ���µ����®)�!Î* ����µ� ��%�®) þ * ������¹����2�®)���* � ����$��%���( �� î � ËR� * ��* �����%µ�µ����®)� �* ï �( � ����~¾�)© �* �¤ ���$ ��*�Á������%�%� ¼º¶£�� ´ 2����� ������%µ������Ì����$ � ������ � ����%�~ �¾º�%�· ´ �_Â"½����� � :�2¾�µ� ����ÄÁ��%�(¹��� ��¾2Ø2Ø2O) µ ��� ¼��(¼���µ��� R�2¾�µ�ú �������µ�2�� 0�� ����( ��%Ï2� ��µ�) ï ��&2 ����� �)=-�4 f�î * î î Ë� �* ï �( � ����~¾�)1 �* �¤ ���$ ��)oÂJ* ÈL¾�µ�µ�* þ ���� ���½ �%���( �� ñJ����µ������%� ¼¸�%�É�~ñ£������ú ����µ���$LÂ"½����� � V�2¾�µ� ���� � µ��%� ¼t�;��µ�µ���¹��%�%�� �¾ ³J´ ���� �¾�À þ ����µ��·�2 ���$ ¾�) µ ��� ¼��(¼���µ£�� ��2¾�µ� �������µ�2�� 0�� ����( ��%Ï2� ��µ�) ï ��&2 ����� �)=-�4 f�î * î f Ë:Á:* Á�� ¹����%µ�)� �* �¤ ���$ ��*� �����$��%�%� ¼ � ������ � ����%�~ �¾��%�8Â"½����� � 3�2¾�µ� �����µ�À[�b�� �ú Ã���%�%µ�)�Á����B������ �����µ�)WñJ������$�����µ�) µ ��� ¼��(¼���µ£�� ��2¾�µ� �������µ©2�� 0�� ����( ��%Ï2� ��µ�) �£�'&�����¹� ���)3-�4 f?f * î 4=Ë:!Î* �;�� � ��������®)���* �;�� � ��������®*��¤ ������%µ�$®ñ ����$��������� � ��������%�%���%Ï2� ��)~!Í��µ�µ���� ô �Ú2�ÂJ)W�3�( ��%µ�)=-�4 î �2* õ[öO÷ ø ÷ ù�úGûù�ü=ý=þ7ùØþOýAÿ�����þO÷ ù��d÷ ����ø ù�÷ �� ��0ü���þ ����ÇýAÿ�� ÷ ��ù� Çø ù������}öO÷ �Aÿ � ü0þO÷ ù����_÷ ����ù��0ýAÿ��=ù���ü�ù��[þØûù�ü=÷ ��ùØþOý ! ��� Ç÷ "# �ü%$&�' �(=ö �)��� �þ�ÿ�*,+�+�$ - ���}öO÷ öOý=ö �./��þ - ��./��þ )�ùØöO÷ ���0�'�Ç÷ ����� 1��.2 !��ý�)�ù���÷ ù�� � ��ù�ü� �)43 5 þ7ù����6� - ù�7�V÷ ����(���ý0��ü�8 ÿ:9�9 � ÿ - ù�7�V÷�ÿ:;�;�<�<0ÿ� !��)�=ù���÷ ù ö Çø>8!?@*,<�A#<&+�B%*�A#;,C�D�$�AFE 4 � 
work_43nmngcsd5hjzjny3wjgodmwf4	----	IE Paper Template Informatica Economică vol. 14, no. 3/2010 199 Multicriterial Methods used in Expert Systems for Business Decision Making Cornelia MUNTEAN, Mihaela I. MUNTEAN West University of Timisoara, Romania cornelia.muntean@feaa.uvt.ro, mihaela.muntean@feaa.uvt.ro Many organizations are in front of most competitive economic environments, where, in order to survive, they must reduce costs all the time and adopt the most intelligent business strate- gies. In most decision making activities the manager has to decide which variant is the most advantageous, taking into account a multitude of criterions. Expert systems use the expert's knowledge and problem solving skills in a particular subject area throughout an organiza- tion, and can propose the optimal variant to be chosen. In this paper we have outlined the role of multicriterial methods in programming expert systems to decide in favor of the most eligible variant between a multitude of possibilities. We also made a case study and designed the prototype of an expert system for choosing the most profitable offer among many, in the prenegotiation stage, for a company, in order to organize the negotiation processes accor- dingly. In this respect, we tried to highlight the usefulness of multicriterial mathematical me- thods in three negotiation processes of a Romanian negotiation team with foreign negotiation teams for the acquisition of an equipment. Keywords: Business Decision Making, Expert Systems, Multicriterial Methods Introduction Today’s decisions are complex, combin- ing hard facts with experts’ intuition. Rapid business decision-making often requires col- laboration across time zones, organizations and cultural norms. In the field of business decision support, more and more recent research [1] has been concentrating on the human side of the per- son-technology relation in decision making. It has been shown in many works that busi- ness decision making environment is a unity of decision makers’ experience, beliefs and perceptions on one side, and decision support tools and techniques – on the other side. The information environment surrounding busi- ness activities and decisions is getting increa- singly complex due to growing volumes of information of potential relevance to certain business activities [2]. An Expert System (ES) is a knowledge-based computer program containing expert domain knowledge about objects, events, situations and courses of action, which emulates the process of human experts in the particular domain. For long term use, a knowledge base stores rules, facts and other knowledge struc- tures, much as a database stores data. When the ES is used, an inference engine processes the knowledge structures, bringing problem specific information into the system, and makes recommendations to the user based on the information and knowledge structures available [3]. After making a recommendation, users rou- tinely view the decision making logic used by the ES. Since the system remembers its logical chain of reasoning, a user may ask for an explanation of a recommendation and the system will display the factors it considered in providing a particular recommendation. This main attribute, the ability to explain rea- soning, enhances user confidence in the rec- ommendation and acceptance of the ES. Some expert systems are designed to take the place of human experts while others are de- signed to aid them. In negociation, as well as in most other cir- cumstances, people must take a decision from a multitude of possible decisions, in or- der to achieve a certain goal. It is perfectly normal for human reasoning to analyse and to compare the possibilities, in order to adopt that decision which permits the best fulfil- ment of the desired goal. Although we fre- quently use the term "optimal decision", in 1 200 Informatica Economică vol. 14, no. 3/2010 most situations this "optimality" is a very complex concept which can't be defined but by mean of a mathematical model. In our case study we will use three multicriterial methods, implemented in an expert system designed in Exsys Corvid. Corvid provides an object-oriented structure that makes it easy to build systems using me- thods and properties of variables, while not requiring the developer to change the way they think and describe their decision-making steps and logic. The result is a very flexible and powerful development environment that can easily be learned. We used Corvid for implementing our application presented in paragraph 3. 2 Multicriterial Methods. Business Deci- sion Making Theoretical Approaches Mathematical models appeared and were used in the process of decision making in business, particularly in negociation, quite from the necessity to sustain the logical rea- soning in negociation and to manage a great number of factors simultaneously. Furthermore, the applying of these mathe- matical methods grants the approach of some new qualitative problems, so that it is not at all surprising the fact that in negociation as well there are used more and more mathe- matical tools, techiques and models. The process of decision-making is defined by following elements [4]: the decision-maker, the assemblage of decision alternatives, the assemblage of decision criterions, the assem- blage of goals. The decision-maker is the person who must select the most advantageous variant from a multitude of possible ones, variant called the optimum choice. The assemblage of decision alternatives, V, is the assemblage of action possibilities at a given moment. The assemblage of decision criterions, C, is the assemblage of parameters which defines the process and in respect of which we have in view the comparison of alternatives. The decision criterions are characterized by a number of levels according to the different alternatives and/or status of unbiased condi- tions. All these levels can make up decision goals that are possible to achieve, from the point of view of that particular criterion. Decision models with an assemblage of crite- rions, called also multicriterial decision mod- els, could be multi-attribute decision models, which are presented below, or multi- objective decision models, which are subject of linear programming [5]. Multi-attribute decision models subsist in the determination of the optimum variant from a finite variant assemblage V={V1, V2, ...,Vm}, variants that are compared one with another in respect with numerical or non-numerical criterions belonging to a finite assemblage C={ C1, C2, ...,Cn}. Each criterion has a min- imum or maximum goal. For some multi-attribute decision problems, in which the matrix of consequences contains heterogeneous data, numerical or non- numerical, the homogenization of these data is done by the normalization procedure [6], which transforms the matrix of consequences in a matrix R=(rij)i=1,m; j=1,n} with elements in the interval [0,1]. In almost all multi-attribute decision prob- lems there is information regarding the im- portance of each criterion. This is generally expressed by the vector P={p1, p2, ..., pn} and indicates the level of importance given by the decision-maker to each criterion. rij =          ≤≤ ≤≤ criterionsfor a a criterionsfor a a ij ij mi ij mi ij min, min max, max 1 1 (1) Every multi-attribute decision problem could be expressed by a matrix A, called the matrix of consequences (Table 1), with elements aij indicating the evaluation (consequence) of Informatica Economică vol. 14, no. 3/2010 201 variant i, i=1, 2, ..., m (Vi), by criterion j, j=1, 2, ..., n, (Cj). Table 1. The matrix of consequences Ci Vj C1 ... Cn V1 a11 ... a1n ... ... ... ... Vm am1 ... amn P p1 ... pn Source: [5] Multi-attribute decision problems could be classified into three categories [7]: direct me- thods, indirect methods and methods which use a certain distance for the construction of hierarchies. Direct methods build a function defined on the assemblage of variants with real values and selects variants for which the function takes the greatest value. Indirect methods determine a hierarchy on the assemblage of variants based on an algo- rithm. Methods which use the distance select a va- riant which is the closest to the ideal solution. In our case study we will use a direct method (the method of simple additive weight) and a method which uses the distance (TOPSIS), methods that are presented below: The method of simple additive weight The method consists in defining the function f : V→R, given by: f(Vi) = mi p rp n j j n j ijj ,1, 1 1 = ∑ ∑ = = (2) where rij are the elements of the normalized matrix R, calculated with relation (1) and pj are the elements of the importance rates vec- tor P, given as last row in Table 1. The opti- mum variant will be that for which f(Vi) takes the maximum value. The TOPSIS method The TOPSIS method (Technique for Order Preference by Similarity to Ideal Solution) is based on the idea that the optimum variant must have the minimum distance to the ideal solution. The steps of the TOPSIS method are: • Step 1. We build the normalized matrix R=(rij), i=1,...,m, j=1,...,n; • Step 2. We build the weighted normalized matrix V=(vij), i=1,...,m, j=1,...,n, where vij = ∑ = n j j ijj p rp 1 (3) • Step 3. We calculate the ideal solution A and the ideal negative solution B, defined as: A= (a1, a2, ..., an), B= (b1, b2, ..., bn) (4) where: aj =     ≤≤ ≤≤ min,min max,max 1 1 isCcriteriontheifv isCcriteriontheifv jijmi jij mi (5) bj =     ≤≤ ≤≤ max,min min,max 1 1 isCcriteriontheifv isCcriteriontheifv jijmi jij mi (6) • Step 4. We calculate the distance between the solutions: Si = ( )∑ = − n j ij ajv 1 2 , i = 1, 2, … ,m; (7) Ti = ( )∑ = − n j jij bv 1 2 , i = 1, 2, ... , m; (8) • Step 5. We calculate the relative nearness from the ideal solution: Ci = ii i TS T + (9) • Step 6. We make a classification on the assemblage V according to the descending values of Ci obtained in step 5. 202 Informatica Economică vol. 14, no. 3/2010 3 Multicriterial Methods. Case Study The case study in this paper wants to mark out the role of mathematical methods imple- mented in an expert system for the stage of preparation in an international negotiation process. For that purpose we tried to highlight the usefulness of multicriterial methods in three business negotiation processes of a same Romanian negotiation team with three other international negotiation teams in order to purchase an industrial equipment. In order to achieve the chased goal we made a case study at S.C. Chimcomplex S.A., us- ing the two multicriterial methods presented in paragraph 2 (the method of simple additive weight and the TOPSIS method) for select- ing, in the stage of prenegotiation, the best offer and for organizing the negotiation processes thereafter. Chimcomplex will have to decide between three offers of three for- eign companies, taking into account eight se- lection criterions:  C1 : the account of the good that has to be purchased (million Euro);  C2 : requested advance money (%);  C3 : time period allowed for the payments (years);  C4 : payment staggering (month);  C5 : rate of interest (%);  C6 : time of delivery of the equipment (month);  C7 :guarantee period (years);  C8 : offer validity (month). The following three offers of three foreign companies will be approached further as po- tential variants of the Romanian company Chimcomplex S.A:  the offer of Vichem Company, from France – variant 1 (V1);  the offer of Michelis Company, from Germany – variant 2 (V2);  the offer of Itochu Corporation, from Ja- pan – variant 3 (V3) The Romanian company Chimcomplex S.A. confers to each invoked criterion a specific rate of importance on a scale from 1 to 10 (rate 10 for the most important criterion and 1 for the lowest importance criterion). So the importance rates are: For C1 : 10; for C2 : 9; for C3 : 8; for C4 : 6; for C5 : 6; for C6 : 5; for C7 : 4; for C8 : 3. In following table are presented the offers of the three companies according to the crite- rions invoked by Chimcomplex S.A and the importance rates given by the Romanian ne- gotiation team (Table 2). Table 2. The characteristics of the variants according to the criterions and the impor- tance rates for each of the criterions C1 Mil.Euro C2 % C3 years C4 Month C5 % C6 month C7 years C8 Month V1 5 10 7 6 7,5 1 1 2 V2 4,75 11 6 12 7 2 2 1 V3 5,25 9 5 5 6,5 1,5 1,5 1,5 P 10 9 8 6 6 5 4 3 For doing all calculations more quickly, we used the expert system generator Corvid, and put all entrance data into an input files, con- taining the characteristics of each variant in respect to each criterion and the importance rates of the eight criterions. Further we will apply, using Corvid va- riables, the two mathematical methods de- scribed in paragraph 2 (the method of simple additive weight and the TOPSIS method) for deriving the best variant between the three offers for the Romanian company. Finally we will compare the results obtained with the two methods. Considering that the matrix in Figure 1 con- tains heterogeneous data, there will be neces- sary a normalization procedure. This occurs by minimization for C1, C2, C5, C6 and by maximization for C3, C4, C7, C8. For the max- imum and minimum criterions we used rela- Informatica Economică vol. 14, no. 3/2010 203 tion (1) and calculated the elements of the normalized matrix R=(rij) with i= ;3,1 j= 8,1 , like in Figure1. Fig. 1. Determination of the normalized matrix in a Corvid logic block. Next, for applying the simple additive weight method, we calculate the values for the func- tions f(Vi) using relation (2), like in Figure 2. Fig. 2. Determination of fV1, fV2 and fV3 for classifying the variants with the simple addi- tive weight method We obtain following results: → fV1 = 0,85882 → fV2 = 0,85867 → fV3 = 0,80088 According to this method, the order of the variants is: V1 → V2 → V3, like in figure 3. It is easy to observe that the values for the variants 1 and 2 are very close (they differ only at the forth decimal), so we can con- clude that this method is unconvincing, and it can't help much in selecting the optimal va- riant. Next we apply the TOPSIS method. Here we also need the normalized matrix R=(rij), i=1,...,m, j=1,...,n and then we build the weighted normalized matrix V=(vij), i=1,...,m, j=1,...,n, using relation (3). Thereaf- ter we calculate the ideal solution A and the ideal negative solution B, like in Figure 4. 204 Informatica Economică vol. 14, no. 3/2010 Fig. 3: Results calculated by the Expert System with the simple additive weight method After calculating the distance between the so- lution and the ideal solution with relation (7) and the distance between the solution and the ideal negative solution with relation (8), we determine the relative nearness from the ideal solution Ci, with i= 3,1 using relation (9) and we obtain: →C1= 0.407 →C2= 0.706 →C3= 0.323. Fig. 4. Part of the logic block calculating the ideal solution A and the ideal negative solution B We make a classification on the assemblage V according to the descending values of Ci , and we obtain following order of variants: V2 → V1 → V3 , like in figure 5, a little different from the simple additive weight method, where the order was V1 → V2 → V3., but the function value for V1 was very close to that of V2. Fig. 5. Results calculated by the Expert System with the TOPSIS method 4 Conclusions For some business negotiations the stakes are much too high to be lost. That's why it is de- sired to select and to apply the most efficient strategy which could grant the winning of the negotiation. In this respect it is useful to call on mathematical methods for identifying the best variants for overcoming a deadlock by anticipating the movements of the partner. Sometimes there are situations in which the negotiators must choose from a multitude of variants, must make a hierarchy and select the optimal offer. In such cases the most ap- propriate tool are multicriterial methods. The purpose of this paper was to relieve the usefulness and importance of multicriterial Informatica Economică vol. 14, no. 3/2010 205 methods in an expert system used for the preparation stage of an international business negotiation process. Of course, there are also important cultural accents of the negotiators of different countries, which could also be in- troduced into the expert system program [8]. While applying two different methods. the method of additive weight and the TOPSIS method, the results were nearly the same, so the decision-maker could select the most profitable offer among many in the prenegot- iation stage, in order to organize the negotia- tion processes accordingly. The expert sys- tem used the input data as a text file and cal- culated, using Corvid variables, the functions that indicate the proposed order of variants, accordingly to each of the two different me- thods. Acknowledgments This work was supported by ANCS-CNMP, project number PNII–91-049/2007 References [1] B. Pradeep Kumar, J. Selvam, V. S. Mee- nakshi, K. Kanthi, A. L. Suseela and V. Lalith Kumar, Business Decision Mak- ing, Management and Information Tech- nology, Available at: http://www.acm.org/ubiquity/views/pf/v 8i08_lalith.pdf, 2007. [2] C. Muntean and I. Hauer, “Improving the management in organizations by using Expert Systems,” Simpozion Ştiinţific Internaţional „Managementul dezvoltării rurale durabile”, Timişoara, 2010. [3] C. Muntean, A. Butuza and O. Dobrican, Sisteme expert. Elemente de teorie şi aplicaţii, Timisoara: Ed. Mirton, 2007. [4] M. Gheorghiţă, Modelarea şi simularea proceselor economice, Bucureşti: Ed. ASE, 2001. [5] M. Andraşiu, A. Baciu, A. Pascu, E. Puşcaş and A. Taşnadi, Metode de deci- zii multicriteriale, Bucureşti: Ed. Tehnică, 1986. [6] M. Neamţu and D. Opriş, Jocuri econo- mice. Dinamică economică discretă. Aplicaţii, Timişoara: Ed. Mirton, 2008. [7] G. Ionescu, E. Cazan and A.L. Negruţ, Modelarea şi optimizarea deciziilor ma- nageriale, Cluj – Napoca: Ed. Dacia, 2000. [8] C. Vasiliu, Tehnici de negociere şi comunicare în afaceri, Bucureşti: Ed. ASE, 2003. Cornelia L. MUNTEAN has graduated the Faculty of Computer Science at the „Politehnica” University of Timişoara in 1986. She holds a PhD diploma in Engineering and is currently assistant professor at the department of Busi- ness Information Systems and Statistics at the West University of Timişoara. She joined the staff of the Faculty of Economics and Business Administra- tion of the West University of Timisoara in 1991 as a teaching assistant, then graduated in 1997 as a senior lecturer and in 2005 as an assistant professor. She is the author of 7 books and about 50 journal articles in the field of artificial intelligence and intelligent systems. Her work focuses on expert systems and business applications in de- cision support. Mihaela I. MUNTEAN has graduated from the Faculty of Computer Science, „Politehnica” University of Timisoara in 1986. Currently, professor Mihaela I. Muntean is the chair of the Business Information Systems and Sta- tistics Department at the West University of Timişoara and an IT independent consultant. She is interested in information technology and knowledge man- agement, her research activity results are published in over 70 papers in in- dexed reviews and conference proceedings. Now, she is teaching advances in database management systems, business intelligence, decision support systems, bringing im- portant contribution to the foundation of higher education in Economic Informatics within the Faculty of Economics and Business Administration Timişoara. http://www.biblioteca.ase.ro/catalog/detalii.php?c=2&q=curry&st=s&dela=0&ct=108788� http://www.biblioteca.ase.ro/catalog/detalii.php?c=2&q=curry&st=s&dela=0&ct=108788� 
work_43tkopwxz5giffwdwgogqhn3xq	----	PDF - An Automated FX Trading System Using Adaptive Reinforcement Learning - working paper ��������� � ���� ��� ������ ����������� ���������� �� ���������� ����� �� � ���� ������� ���� ���� ����� ��� �������� � ���� ����� � ���� � �� ! � " " " #$ ��#%��#�%#�&� � � � � � � � � �� ���������'( ����� ���������� � �� � ���� ���) � ���%��� ��*��� �� � � � � �� �� �� �� � ���� ��� ����� � � � � � � ���������� � � ������������������ ��� � �������� ������������� ��� ������ ������ ���������� ��� �� �� ���� � � ����������������������� �� �� ������� ������ ���� ������ ������ ��������� �� � ��� ��� ��� � ������ �������� ������ ��� � ���� �� �� ��� ����������� �� ����� ���� ��� � � � � ��� ���� �� ��� � �� �������� � �� � ������ ����� �� ������ � ���� ������� � � �! �" �# �� �������� �$ �%��� ���� � � & '�()*+,,-� � � � �! �" �# �� ������ � ������� ��.���������/ �������� �� ������������� ��� ������ ���� � ���������� ��� �� �� ��� 0� ���1�� � �� �����2 3�� ����� �����4� $ �%��� ���� � ������� ��.���������/ �������� �� ������������� ��� ������ ���� � ���������� ��� �� �� ��� 0� ���1������� ���2 3�� ����� �����4� � � � � � � � � � '������� ���������������� ��������������� 1� � / ��������5��� ���� ������� �� ������������� ��� ������ ����� � ��� ����� ��5������ � �� �� ���� 6+�(! 7 ��� 8 � � ��1�,(++9�:;,<-;�.�=1�,(++9�99>:,(� 0?� ���1���������?���� ��2 3�� ����� �����4� AN AUTOMATED FX TRADING SYSTEM USING ADAPTIVE REINFORCEMENT LEARNING M.A.H. DEMPSTER and V. LEEMANS Centre for Financial Research Judge Institute of Management University of Cambridge & Cambridge Systems Associates Limited {mahd2, vl227}@cam.ac.uk www-cfr.jims.cam.ac.uk 24th December 2004 Abstract This paper introduces adaptive reinforcement learning (ARL) as the basis for a fully automated trading system application. The system is designed to trade FX markets and relies on a layered structure consisting of a machine learning algorithm, a risk management overlay and a dynamic utility optimization layer. An existing machine-learning method called recurrent reinforcement learning (RRL) was chosen as the underlying algorithm for ARL. One of the strengths of our approach is that the dynamic optimization layer makes a fixed choice of model tuning parameters unnecessary. It also allows for a risk-return trade-off to be made by the user within the system. The trading system is able to make consistent gains out-of-sample while avoiding large draw-downs. 1 INTRODUCTION TO TRADING SYSTEMS An active subject of research among professionals as well as academics over the past decade has been the construction of systems that autonomously trade in securities or currencies: and a summary of earlier work can be found in [1]. In the quest for a trading algorithm, artificial intelligence methods have been employed to construct systems that are better than or at least as good as their human counterparts in timing trade entry and exit oppor- tunities. Often such a system takes as inputs past market prices (or some transformation 1 thereof) and attempts to return a trading signal indicating the preferred position to hold in a given currency relative to another (i.e. long or short). The majority of systems described in the literature aim to maximize trading profits or some risk-adjusted measure such as the Sharpe ratio. Many attempts have been made to come up with a consistently profitable system and inspiration has come from different fields ranging from fundamental analysis, econometric modelling of financial markets, to machine-learning [5, 8]. Few attempts were successful and those that seemed most promising often could not be used to trade actual markets because of associated practical disadvantages. Among others these included large draw- downs in profits and excessive switching behaviour resulting in very high transaction costs. Professional traders have generally regarded those automated systems as far too risky in comparison to the returns they were themselves able to deliver. Even if a trading model was shown to produce an acceptable risk-return profile on historical data there was no guarantee that the system would keep on working in the future. It would cease working precisely at the moment it became unable to adapt to the changing market conditions. This paper aims to deal with the above problems to obtain a usable, fully automated and intelligent trading system. To accomplish this a risk management layer and a dynamic optimization layer are added to a known machine-learning algorithm. The middle layer manages risk in an intelligent manner so that it protects past gains and avoids losses by re- straining or even shutting down trading activity in times of high uncertainty. The top layer dynamically optimizes the global trading performance in terms of a trader’s risk prefer- ences by automatically tuning the system’s hyper-parameters. While the machine-learning system is designed to learn from its past trading experiences, the optimization overlay is an attempt to adapt the evolutionary behaviour of the system and its perception of risk to the evolution of the market itself. In the past an automated trading system based on 2 superimposed artificial intelligence algorithms was proposed [5]. This research departs from a similar principle by developing a fully layered system where risk management, automatic parameter tuning and dynamic utility optimization are combined. (For an earlier attempt in this direction see [13].) The machine learning algorithm combined with the dynamic optimization is termed adaptive reinforcement learning. Section 2 of this paper briefly discusses the RRL machine-learning algorithm underlying the trading system. Section 3 looks at the different layers of the trading system individually. In 3.1 the modifications to the standard algorithm are set out and in 3.2 and 3.3 the risk management and optimization layers are explained. The impressive performance of the trading system is demonstrated in Section 4 and the final section concludes. 2 2 THE MACHINE-LEARNING ALGORITHM The basic model from which we depart was originally applied with some success to FX trading by Moody and Saffell [8] in 1999. This section briefly reviews the mathematical model specification and estimation procedure as explained in [9]. The trading model takes as inputs standardized past time series changes and outputs the preferred position (long or short) in one of the two currencies involved. Further, it is assumed that the model can be specified as a recurrent single layer neural network of the price of one currency in terms of another, i.e. Ft = sign( M∑ i=0 wi,trt−i + wM+1,tFt−1 + vt) (1) where Ft ∈ {−1, 1} is the position to take at time t, wt and vt are respectively the weight vector and threshold of the neural network at time t and rt := pt − pt−1 are the returns in terms of price changes of the FX time series. For simplicity of exposition we assume here that the discrete time index t refers to a basic data frequency. We also assume in this section only that the trading system always takes one position or another. For a more detailed discussion of the actual continuous time situation see [1-5]. At every trade exactly 1 unit of a certain currency is bought or sold and hence the profit at time T can be calculated (ignoring interest rates) as: PT = T∑ t=0 Rt (2) Rt := Ft−1rt − δ|Ft − Ft−1|, (3) where δ is the transaction cost per trade. For computational efficiency, [8, 9] choose the differential sharpe ratio as the risk- adjusted return measure to be optimized. It is derived by considering a moving average version of the classical sharpe ratio given by: Ŝ(t) := At Bt (4) At = At−1 + η(Rt − At−1) (5) Bt = Bt−1 + η(R 2 t − Bt−1). (6) When Ŝ(t) is expanded as a Taylor series in the adaptation parameter η, the first derivative term can be regarded as an instantaneous performance measure Dt := dŜ(t) dη |η=0 (7) = Bt−1∆At − 12 At−1∆Bt (Bt−1 − A2t−1) 3 2 . (8) 3 Because of its simplicity and tractability, a simple gradient ascent is used as optimization algorithm to gradually improve the model parameters by means of the recursion: wi,t = wi,t−1 + ρ∆wi,t. (9) ∆w can then be approximated for an on-line update by considering only the term that depends on the most recent trading return Rt, viz ∆wi,t = dDt dwi (10) ≈ dDt dRt {dRt dFt dFt dwi,t + dRt dFt−1 dFt−1 dwi,t−1 }. (11) Since the neural network is recurrent, an updating scheme similar to back-propagation through time can be applied with: dFt dwi,t ≈ ∂Ft ∂wi,t + ∂Ft ∂Ft−1 dFt−1 dwi,t−1 . (12) All terms and factors in (11) and (12) can be calculated from the previous expressions. Together, they form the set of equations necessary for updating the model weights. This machine-learning technique was termed recurrent reinforcement learning (RRL) [9]. As demonstrated by Gold [6], RRL lends itself perfectly to a rolling-window approach [3]. Here, the model is trained for a number of epochs ne on a training set of length Ltrain and then applied to a test set of length Ltest. Next, the training window is advanced by Ltest points and the whole procedure is repeated. The out-of-sample performance of the model is the sum of the performances on the individual non overlapping test sets. We found that the optimal values of Ltrain and Ltest on our FX (midpoint quote) datasets were around 2000 and 500 ticks respectively, which are the values chosen in [6]. The number of recursive training epochs ne was set to 10. The optimal values of these 3 parameters did not change significantly when different FX data sets were used. 3 THE TRADING SYSTEM The trading system discussed in this paper consists of 3 layers: the basic trading system, the risk and performance management layer, and the paramater optimization layer as shown in Figure 1. Each of these layers is discussed in more detail in the sequel, starting with the bottom-most layer. 3.1 Layer 1: Extensions to the machine-learning algorithm The core of the trading system discussed here is the RRL algorithm [8], discussed in the previous section. A number of changes to this basic algorithm were necessary however to 4 Figure 1: Schematic illustration of the automated trading system consisting of 3 main layers: the machine-learning algorithm, the risk and performance management layer and the dynamic optimization layer. The combination of the first and third layer is termed adaptive reinforcement learning (ARL). The trader’s risk aversion ν is an exogenous parameter to the system. Layer 3 optimizes the trailing stop-loss level x, the trading threshold y, the trading cost δ, the adaptation parameter η and the learning rate ρ. The auto-shut-down critical loss parameter z was fixed. 5 improve its performance and to make it suitable for a structural risk management approach. RRL was extended to take account of other inputs than past returns rt−i and the current position Ft. Experiments were conducted to include signals originating from 14 popular technical indicators as first suggested in [4]. Performance did not increase by adding in- dicators to current data, except when a low number of past returns M was fed into the system. The filtering of the price data timeseries offered by the technical indicators thus did not contribute to increased performance on current data. This indicates that the RRL algorithm is able to efficiently exploit the information in past returns timeseries. During the training phase, the transaction cost factor δ was not fixed to the actual bid-ask spread as in [6], but was instead left as a tuning parameter. Setting a higher cost factor, necessitates a higher expected raw profit before engaging in a trade. Consequently the parameter δ will influence the performance and risk profile of the resulting trading strategy. Large jumps in FX returns, as e.g. when central banks intervene, can introduce instabil- ity into the underlying RRL algorithm. To prevent the weights from taking unreasonably large values, all weights were rescaled by a factor f < 1 as soon as one of the weights hit a certain threshold value. This also prevents the weights from drifting slowly upwards without bound. This phenomenon is undesirable because the user may get incorrect model evaluations at some point because of the numerical instabilities associated with very large weights. We also improved the performance of the RRL trading model by implementing a bet- ter position updating scheme. The traditional update scheme sometimes causes indecisive switching between different positions from one tick to the next. This can generate huge transaction costs and is highly undesirable. This behaviour is reduced by recalculating the output Ft of the trading model in (1) twice. The first time is as before: after the new inputs have been received a new trading signal is generated. This trading signal is in fact based on the weights wi,t−1 that were calculated in the previous time period. Here a sec- ond evaluation of the model is added after the weights are updated for the current period. Since the weight updating is designed to improve the model at each step, it makes sense to recalculate the trading decision with the most up-to-date version of the parameters wi,t. This recalculation may result in a different trading signal than was previously output by the model. This final trading signal is used for effective decision making by the risk and performance control layer. 6 3.2 Layer 2: Risk and performance management One of the strengths of the layered structure is that the decision to actually enter a trade is separated from the trade recommendations made by the basic structure in Layer 1. While Layer 1 outputs the preferred position to hold in the model and thus idealized world, Layer 2 evaluates this trade recommendation by considering additional risk factors relating to the real world before taking a decision to trade. These risk factors impose requirements on real world trading systems that are hard to include in the trading model itself, but can more easily be treated in the risk management overlay structure. More than half a century ago J.M. Keynes, the famous economist noted that: ‘The markets can stay irrational longer than you can stay solvent.’ When a trade goes bad, a psychological tendency exists to keep the position open in the hope that the market will reverse itself and the trade will again turn profitable. This implies a risk-seeking attitude towards losses as opposed to risk-aversion with regard to profits. However, the market could very well not reverse itself and eventually force the close out of the position at a huge loss. This characteristic of human psychology needs to be avoided by a successful automated trading system. Layer 2 avoids this pitfall by building in a trailing stop-loss for every trade. A stop-loss is set and adjusted so that it is always x basis points under or above the best price ever reached during the life of the position. At this point x is an unknown parameter that will influence the performance. A numerical value for x will be obtained via the optimization in Layer 3. If a position is closed out before a trade exit signal is given by the system, the current market behaviour was clearly not expected by the RRL trading model. Otherwise it could have foreseen the situation and would not have suffered a loss. If the market behaves ‘irra- tionally’ according to the model, it is likely that its deviant behaviour will persist for some time. For this reason a cool-down period is included in Layer 2. After the position has been closed out the system needs to wait a while before trading can be resumed. The length of this optimal period of non-activity was determined experimentally as a fixed number of basic time intervals (here 1 minute). An important feature of the risk management layer is that allows the evaluation of the strength of the trading signal given by Layer 0. This is accomplished by looking at the unthresholded output generated by (1). For example, a very strong buy signal corresponds to an output close to +1. The performance management systems offers the possibility of validating a trading signal only when the output exceeds a specified non-zero threshold instead of just applying a sign-function. This threshold is a parameter y that is not fixed in advance but that can be set by the optimization in Layer 3 to influence the risk-return trade-off made by the trading system as a whole. 7 It is common knowledge in the trading community that automated trading systems often work for a certain while until they suddenly stop being profitable. At that point, the system must be shut down to avoid further losses and should be redesigned and thoroughly tested before making it operational again. The performance management layer makes sure that anomalous performance is detected in an early stage to protect the profits already earned. In that case, the system is automatically shut down and a warning is issued. The detection module employs a measure known as maximum draw-down. If at a certain point the draw-down in cumulative profits from its maximum proves larger than an amount z, the auto-shut-down procedure is initialized. This number z can be determined as a Value at Risk by using a fitted draw-down distribution, or it can be set more pragmatically to a fixed value. 3.3 Layer 3: Dynamic optimization of utility The trading system up to Layer 2 is able to trade autonomously and manage risk appro- priately. However, the system’s risk-return profile depends on a number of parameters δ, η, ρ, x, y and z discussed above. It is the task of Layer 3 to search for optimal values of these parameters so that some utility of the resulting risk-return profile is maximized. Therefore, it is key to find a suitable utility function that is dependent on the cumulative profit, on a measure of trading risk and on the supervising trader’s personal risk aversion. We define the risk measure Σ and utility function U as follows: Σ := ∑N i=0(Ri) 2I(Ri < 0)∑N i=0(Ri) 2I(Ri > 0) (13) U (R, Σ, ν) := a.(1 − ν).R − ν.Σ, (14) where the strategy’s raw return at time i is Ri := Wi − Wi−1, Wi is the wealth or cumula- tive profit at time i, R = WN N is the average profit per frequency interval, ν is the trader’s personal risk aversion, a is a constant and N is the number of out-of-sample time intervals. A trader will get a higher utility from a larger average profit R and a lower risk factor Σ. A higher risk aversion parameter ν induces the optimization to find a performance profile with lower risk. Often this also implies a lower return. The risk measure in (13) was constructed in such a way that it satisfies the following conditions: 1. Large negative strategy returns imply a higher risk value than several smaller ones that result in the same total negative return. This is because we want to avoid large sudden draw-downs that could, for example, lead to margin calls. 2. A larger total impact of losing trades (as measured by the cumulative sum of squares of the negative strategy returns ∑N i=0(Ri) 2I(Ri < 0)) versus the total impact of winning trades (as measured by the cumulative sum of squares of the positive strategy returns 8 ∑N i=0(Ri) 2I(Ri > 0)) implies a higher risk value even if the total profit remains the same. The risk measure Σ essentially implies that systems with monitonically increasing profits should be favoured over ones with whipsawing profits. 3. Comparing the total impact of losing versus winning trades is done in a relative way, i.e. by using a ratio, so that strategies that produce returns which only differ by a constant scaling factor result in the same risk value. Since the positive mean return would be multiplied by the scaling factor, the scaled up strategy would be preferred in a risk-return framework. 4. The risk measure Σ penalizes only negative strategy returns, and not positive returns smaller than the mean. Using a centered semivariance measure would perversely assign a higher risk value to a strategy with infrequent very profitable trades as these shift the mean upwards while leaving the lower part of the distribution unaltered. It should be noted that due to the non-standard from of our chosen risk measure Σ, we can not rely on earlier theoretical work by Ruszczyński et al. [11],[12] that determines bounds for the risk-aversion parameter in order for efficient strategies in terms of risk-return utility to be second order stochastically dominant. It is left to further research to extend to our risk measure Σ proofs of consistency between the volatility risk measure and second order stochastic dominance as in Ogryczak and Ruszczyński [10]. Our utility function depends on the average profit R and the risk measure Σ, which in turn are controlled by the values of 5 parameters: trading cost δ, adaptation parameter η, learning rate ρ, stop-loss parameter x and trading threshold y. Hence the parameter optimization problem becomes formally: max δ,η,ρ,x,y U (R, Σ : δ, η, ρ, x, y; ν). (15) Unfortunately, (15) cannot be solved analytically as we use an online updating scheme. Moreover, numerical optimization in a 5-dimensional space requires too much computing power and time. Therefore, we implement a one-at-a-time random search optimization in which each parameter is approximately optimized individually while keeping the others constant: max δ max η max ρ max x max y U (R, Σ : δ, η, ρ, x, y; ν). (16) The individual optimizations were performed by evaluating 15 random values distributed normally around the starting value and picking the best. This saves function evaluations and we do not have to estimate numerical derivatives of a utility function that is not smooth in the optimized parameters. Of course, other more sophisticated optimization schemes could be applied. One-at-a-time optimization only works if the objective function is sufficiently well- behaved. To make a preliminary qualitative assessment of the response behaviour of the 9 system when its 5 tuning parameters are adjusted, two numerical initialization vectors for these 5 parameters were chosen and the local performance behaviour monitored. Each of the 5 parameters were modified in turn while keeping the other parameters constant. Figures 5 and 6 show the local performance behaviour for one of the initialization vectors chosen. The other vector gave very similar results. In Figures 7 and 8 the associated per- formance profiles are plotted in risk-return space to illustrate the impact on performance. The main empirical observation is that performance measured in each of the dimensions is generally well-behaved when each of the 5 parameters lies within vague boundaries that can easily be read from the graphs. Allowing values outside these boundaries will generally increase the search time of the optimization. The length of the time for this top-layer optimization interval is set to a multiple of the training interval length Ltrain so that it is safe to superimpose it on the previous layers. It is sensible to choose a relatively lower frequency for this optimization as the parameters being optimized determine the risk management characteristics as well as the adaptive behaviour of the underlying learning algorithm. As these are both aspects of the system that typically span multiple time periods or even multiple trades, more time is required to see the impact of these hyper-parameters on trading performance. At the start, the 5 parameters δ, η, ρ, x and y are initialized with 5 randomly chosen values because there is no information available about what values result in good performance. The optimization then searches for values that result in optimal utility of performance over some past period of time. The ARL framework not only eliminates the problem of having to choose and fix hyper-parameters a priori, but also allows the trading system to dynamically adapt to changing market conditions and even to take into account changing risk preferences of the user over time. 4 APPLICATION TO FX TRADING This section reports the use of the ARL trading system to trade the foreign exchange (FX) markets based on historical data. For this purpose, the system was fed historical FX data and its out-of-sample performance on this data recorded and analyzed. A few more exper- iments were conducted to test the impact of the dynamic optimization of utility. Figure 2 shows the out-of-sample cumulative profits when the system was used to trade the Euro-Dollar currency pair1. Trading was only allowed during the active hours that cor- respond to sufficient liquidity: from 9 a.m. London time until 5 p.m. London time. During these hours, the average bid-ask spread in the inter-dealer FX market was lower than 2 pips. 1Historical FX data for Euro-Dollar was kindly supplied by HSBC Bank, London, for which we are grateful. The frequency is 1 minute and the data spans a period from January 2000 until January 2002. 10 As execution speeds and liquidity are usually very high on interdealer electronic trading platforms like EBS and Reuters3000, no further slippage was assumed. The ARL system was able to earn 5104 pips over 2 years, or more than 26% per annum. Given the absence of any serious draw-dawns in the profit graph, this is not a bad performance. Preliminary results on recent FX market data covering a period up to January 2004 indicates that the profit curve exhibits a diminishing slope. This could signify that markets are becoming more and more efficient and that it is becoming more difficult to make profits when only price information is fed into the system. 0 0.5 1 1.5 2 2.5 x 10 5 −1000 0 1000 2000 3000 4000 5000 6000 Time (01−2000 01−2002) A b so lu te P ro fit ( p ip s) 0 0.5 1 1.5 2 2.5 x 10 5 0.8 0.85 0.9 0.95 1 1.05 Time (01−2000 01−2002) E u ro − D o lla r E xc h a n g e R a te ( p ip s) (a) (b) Figure 2: (a) Out-of-sample cumulative profit measured in basis points for trading exactly one Euro against Dollar from January 2000 until January 2002. The trading system was dynamically optimized for a risk-aversion parameter of 0.5. The automated trading system earned a profit of 5104 pips or approximately 26% p.a., whereas a buy and hold strategy would have resulted in a 1636 pips loss or approximately an 8% p.a. loss. (b) The Euro-Dollar exchange rate over the period considered. To illustrate the benefits of our structural approach and the top-level utility optimiza- tion, a series of experiments was run to compare performance and utility of performance with and without dynamic optimization turned on. Table 1 shows the resulting perfor- mance statistics for traders with different risk aversions. From these experiments we see that as risk aversion increases, utility generally goes down because of heavier penalization of the risk factor. This makes the system less free to chose a utility maximizing strategy. When very low risk aversion parameters of 0 until 0.2 are used, the impact of risk in the utility function is insufficient to obtain a less risky strategy. When risk aversion is increased until 0.4, the cumulative profit and total number of trades go up, but the average profit made per trade decreases. The reason why the system is choosing less profitable trades 11 on average is most likely because it needs to select less risky trades. Fortunately at this point the system adopts a new strategy that takes on a larger number of trades so that the cumulative profit still ends up high, despite the average profit and the utility going down. When the risk aversion is further increased until 0.9, the system needs to select only the most profitable trades and thus the average profit increases again. The total number of trades drops and unfortunately so does the cumulative profit. ν WN Σ N R Direction U with U without (pips) (pips) Correct (%) layer 3 layer 3 0 4779 0.5334 2854 1.67 62 % 2.07 1.92 0.1 4717 0.5369 2852 1.65 62 % 1.79 1.67 0.2 4717 0.5369 2852 1.65 62 % 1.53 1.43 0.3 4663 0.5596 3004 1.55 62 % 1.25 1.18 0.4 5144 0.5225 3353 1.53 62 % 1.13 0.93 0.5 5104 0.4880 3337 1.53 62 % 0.86 0.69 0.6 5083 0.4962 3220 1.58 62 % 0.58 0.44 0.7 4780 0.4254 3129 1.53 61 % 0.32 0.19 0.8 4635 0.3901 2657 1.74 62 % 0.09 -0.05 0.9 4669 0.3761 2639 1.77 63 % -0.14 -0.30 Table 1: Out-of-sample heuristics on the trading simulations: Risk aversion ν, cumulative profit WN , risk σ, total number of trades N , average profit per trade R, percentage of trades in the right direction (after deduction of costs), utility with dynamic layer 3 optimization and the corresponding utility without layer 3 optimization turned on. While trading, the spread was fixed at 2 pips and the system was restricted to only trade during hours when the true spread was not higher than 2 pips. Figure 3 demonstrates that when realistic risk aversion parameters larger than 0.3 are used, the optimization layer makes sure that the resulting system becomes less and less risky as the risk aversion parameter is increased. The average net profit made per trade ranged from 1.53 pips up to 1.77 pips, depending on the trader’s risk aversion. To demonstrate further that the dynamical optimization performed in layer 3 adds to an increased trading performance, we compared the utilities of performance for different risk aversions with the corresponding utilities when no dynamic optimization was performed. In the second case, the values of the 5 hyper-parameters were set so that they gave optimal performance on the whole dataset. To assure that this static parameter setting is the best one possible, and thus provides a reliable benchmark, static optimal values were calculated based on the training as well as the test sets. It should be noted that the results with those static parameters are only meant as an upper boundary on performance without op- 12 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.4 0.5 0.6 Risk Aversion Parameter T ra d in g S tr a te g y ri sk Figure 3: Out-of-sample trading risk vs. risk aversion parameter. The resulting performance of the trading system becomes less risky as the risk aversion parameter in the layer 3 optimization is increased. For very low risk aversion parameters, the impact of risk in the utility function is insufficient to make the strategy less risky. timization and not as a reflection of real performance. Indeed, when optimizing over both training and test sets, knowledge of all future prices is assumed at any point in time. Hence, these static parameter results are better than what can be achieved in out-of-sample trad- ing. Nevertheless, they provide us with a best-case scenario for static parameter settings. Ideally the ARL system should be able to deliver an even better performance compared to this hard-to-beat benchmark. Figure 4 shows the utilities that were obtained for several risk aversions of the trading system with and without Layer 3 activated. Clearly the ARL system performs consistently better than the benchmark which illustrates that the addition of Layer 3 to the trading system contributes significantly to better performance. 5 CONCLUSIONS In this paper we have developed an automated trading system based on adaptive reinforce- ment learning. The parameters that govern the learning behaviour of the machine-learning algorithm and the risk management layer are dynamically optimized to maximize a trader’s utility. A risk-aversion parameter allows control of the system’s trade-off between strategy risk and return. ARL automatically tunes the hyper-parameters and thus eliminates the need for manual model calibration or fitting and hence greatly reduces the risk of data- snooping. The system performed very well on the high-frequency historical FX dataset 13 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 −0.5 0 0.5 1 1.5 2 2.5 Risk Aversion Parameter U til ity o f P e rf o rm a n ce Figure 4: Utility of out-of-sample trading performance vs. risk aversion parameter. The dotted line represents utility when Layer 3 is deactivated and the parameters are fixed to their optimal static values. The solid line represents utility when dynamic optimization in Layer 3 is activated. For every risk aversion setting, applying dynamic optimization was clearly beneficial. that was examined. A first possible avenue for future research is to examine whether the trading performance of this system can be further improved by feeding other information than price data into the system. In earlier work [2, 7] we have demonstrated that order book or order flow information is able to enhance the performance of automated trading systems. Secondly, the risk management layer could be extended to control several auto- mated FX trading systems that trade different currencies, in an attempt to emulate the actions of a human FX trader. 14 A P P E N D IX A 0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 2 0 − 1 0 0 00 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 T o ta l C u m u la tiv e P ro fit ( p ip s) 0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 2 0 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 N u m b e r o f T ra d e s 0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 2 0 − 4 − 2024 A ve ra g e P ro fit p e r T ra d e ( p ip s) 0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 2 0 1 0 2 0 3 0 4 0 5 0 6 0 7 0 P e rc e n ta g e o f T ra d e s P ro fit a b le ( % ) 0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 2 0 0 .3 0 .3 5 0 .4 0 .4 5 0 .5 0 .5 5 0 .6 S tr a te g y R is k F a ct o r 0 0 .0 5 0 .1 0 .1 5 0 .2 0 .2 5 0 .3 0 .3 5 − 5 0 00 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 T o ta l C u m u la tiv e P ro fit ( p ip s) 0 0 .0 5 0 .1 0 .1 5 0 .2 0 .2 5 0 .3 0 .3 5 0 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0 N u m b e r o f T ra d e s 0 0 .0 5 0 .1 0 .1 5 0 .2 0 .2 5 0 .3 0 .3 5 − 2 − 101234 A ve ra g e P ro fit p e r T ra d e ( p ip s) 0 0 .0 5 0 .1 0 .1 5 0 .2 0 .2 5 0 .3 0 .3 5 2 0 3 0 4 0 5 0 6 0 7 0 P e rc e n ta g e o f T ra d e s P ro fit a b le ( % ) 0 0 .0 5 0 .1 0 .1 5 0 .2 0 .2 5 0 .3 0 .3 5 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 S tr a te g y R is k F a ct o r 0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 1 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 3 0 0 0 T o ta l C u m u la tiv e P ro fit ( p ip s) 0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 1 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 N u m b e r o f T ra d e s 0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 1 0123456 A ve ra g e P ro fit p e r T ra d e ( p ip s) 0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 1 3 0 4 0 5 0 6 0 7 0 P e rc e n ta g e o f T ra d e s P ro fit a b le ( % ) 0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 1 0 .3 0 .3 5 0 .4 0 .4 5 0 .5 0 .5 5 0 .6 S tr a te g y R is k F a ct o r (a ) (b ) (c ) V ar yi n g th e tr ad in g co st δ . V ar yi n g th e ad ap ta ti on p ar am et er η . V ar yi n g th e le ar n in g ra te ρ . F ig u re 5: L oc al p er fo rm an ce b eh av io u r fo r d iff er en t va lu es of th e p ar am et er s. D ot te d li n es re p re se nt a sy st em fe d w it h 14 te ch n ic al in d ic at or s as ex tr a in p u ts . 15 − 2 0 0 − 1 8 0 − 1 6 0 − 1 4 0 − 1 2 0 − 1 0 0 − 8 0 − 6 0 − 4 0 − 2 0 0 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 3 0 0 0 T o ta l C u m u la tiv e P ro fit ( p ip s) − 2 0 0 − 1 8 0 − 1 6 0 − 1 4 0 − 1 2 0 − 1 0 0 − 8 0 − 6 0 − 4 0 − 2 0 0 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 6 0 0 0 N u m b e r o f T ra d e s − 2 0 0 − 1 8 0 − 1 6 0 − 1 4 0 − 1 2 0 − 1 0 0 − 8 0 − 6 0 − 4 0 − 2 0 0 01234 A ve ra g e P ro fit p e r T ra d e ( p ip s) − 2 0 0 − 1 8 0 − 1 6 0 − 1 4 0 − 1 2 0 − 1 0 0 − 8 0 − 6 0 − 4 0 − 2 0 0 0 2 0 4 0 6 0 8 0 P e rc e n ta g e o f T ra d e s P ro fit a b le ( % ) − 2 0 0 − 1 8 0 − 1 6 0 − 1 4 0 − 1 2 0 − 1 0 0 − 8 0 − 6 0 − 4 0 − 2 0 0 0 .4 0 .4 5 0 .5 0 .5 5 0 .6 0 .6 5 0 .7 S tr a te g y R is k F a ct o r 0 0 .5 1 1 .5 2 2 .5 3 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 T o ta l C u m u la tiv e P ro fit ( p ip s) 0 0 .5 1 1 .5 2 2 .5 3 0 2 0 0 0 4 0 0 0 6 0 0 0 8 0 0 0 N u m b e r o f T ra d e s 0 0 .5 1 1 .5 2 2 .5 3 01234 A ve ra g e P ro fit p e r T ra d e ( p ip s) 0 0 .5 1 1 .5 2 2 .5 3 3 0 4 0 5 0 6 0 7 0 P e rc e n ta g e o f T ra d e s P ro fit a b le ( % ) 0 0 .5 1 1 .5 2 2 .5 3 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 S tr a te g y R is k F a ct o r (d ) (e ) V ar yi n g th e st op -l os s d ep th x . V ar yi n g th e tr ad in g tr es h ol d y . F ig u re 6: L oc al p er fo rm an ce b eh av io u r fo r d iff er en t va lu es of th e p ar am et er s. D ot te d li n es re p re se nt a sy st em fe d w it h 14 te ch n ic al in d ic at or s as ex tr a in p u ts . 16 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 3 0 0 0 3 5 0 0 4 0 0 0 4 5 0 0 d a ta 1 P a ra m e te r s e tt in g s s ta rt in g p o in t s im u la ti o n O O 2 l o c a l ri s k − re tu rn o p ti m a 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 3 0 0 0 3 5 0 0 4 0 0 0 4 5 0 0 d a ta 7 P a ra m e te r s e tt in g s s ta rt in g p o in t s im u la ti o n O O 2 l o c a l ri s k − re tu rn o p ti m a 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 3 0 0 0 3 5 0 0 4 0 0 0 4 5 0 0 d a ta 9 P a ra m e te r s e tt in g s s ta rt in g p o in t s im u la ti o n O O 3 l o c a l ri s k − re tu rn o p ti m a O (a ) (b ) (c ) V ar yi n g th e tr ad in g co st δ . V ar yi n g th e ad ap ta ti on p ar am et er η . V ar yi n g th e le ar n in g ra te ρ . F ig u re 7: L oc al p er fo rm an ce b eh av io u r fo r d iff er en t va lu es of th e p ar am et er s, p lo tt ed in ri sk -r et u rn sp ac e. 17 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 3 0 0 0 3 5 0 0 4 0 0 0 4 5 0 0 d a ta 1 0 P a ra m e te r s e tt in g s s ta rt in g p o in t s im u la ti o n O 2 l o c a l ri s k − re tu rn o p ti m a O 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 3 0 0 0 3 5 0 0 4 0 0 0 4 5 0 0 d a ta 1 1 P a ra m e te r s e tt in g s s ta rt in g p o in t s im u la ti o n O 2 l o c a l ri s k − re tu rn o p ti m a O (d ) (e ) V ar yi n g th e st op -l os s d ep th x . V ar yi n g th e tr ad in g tr es h ol d y . F ig u re 8: L oc al p er fo rm an ce b eh av io u r fo r d iff er en t va lu es of th e p ar am et er s, p lo tt ed in ri sk -r et u rn sp ac e. 18 References [1] Austin M.P., Bates R.G., Dempster M.A.H., Leemans V. and Williams S.N. (2004) Adaptive Systems for Foreign Exchange Trading. Quantitative Finance 4 (4) 37-45. [2] Bates R.G., Dempster M.A.H. and Romahi Y.S. (2003). Evolutionary Reinforcement Learning in FX Order Book and Order Flow Analysis. Proceedings of the 2003 IEEE International Conference on Computational Intelligence for Financial Engineering, Hong Kong March 2003 355-362. [3] Dempster M.A.H. and Jones C.M. (2001). A real-time adaptive trading system using genetic programming. Quantitative Finance 1 397-413. [4] Dempster M.A.H., Payne T.W., Romahi Y.S. and Thompson G.W.P. (2001). Compu- tational learning techniques for intraday FX trading using popular technical indicators. Special issue on Computational Finance, IEEE Transactions on Neural Networks 12 744-754. [5] Dempster M.A.H. and Romahi Y.S. (2002). Intraday FX trading: An evolutionary re- inforcement learning approach. Intelligent Data Engineering and Automated Learning 2002, Lecture Notes in Computer Science, Springer-Verlag 2412 347-358. [6] Gold C. (2003). FX trading via recurrent reinforcement learning. IEEE International Conference on Computational Intelligence in Financial Engineering 2003. [7] Leemans V. (2003). Real time trading systems. MPhil dissertation, Centre for Finan- cial Research, Judge Institute of Management, University of Cambridge. [8] Moody J. and Saffell M. (1999). Minimising downside risk via stochastic dynamic programming. 6th International Conference Computational Finance 1999 Edited by Abu-Mostafa Y. S., LeBaron B., Lo A. W. and Weigend A. S., MIT Press, Cambridge, MA. 403-415. [9] Moody J. and Saffell M. (2001). Learning to trade via direct reinforcement. IEEE Transactions on Neural Networks 12 (4) 875-889. [10] Ogryczak W. and Ruszczyński A. (1999). From stochastic dominance to mean-risk models: Semideviations as risk measures. European Journal of Operations Research 116 33-50. [11] Ogryczak W. and Ruszczyński A. (2002). Dual stochastic dominance and related mean- risk models. SIAM Journal on Optimization 13 (1) 60-78. [12] Ruszczyński A. and Vanderbei R.J. (2003). Frontiers of stochastically nondominated portfolios. Econometrica 71 (4) 1287-1297. [13] Venturi S. (2003). Evolutionary Algorithms for Currency Trading. PhD thesis, Centre for Financial Research, Judge Institute of Management, University of Cambridge. 19 
work_45zfpdeutjfite47n32bqpobra	----	Untitled Document Formally analysing the concepts of domestic violence Jonas Poelmans1, Paul Elzinga3, Stijn Viaene1,2, Guido Dedene1,4 1K.U.Leuven, Faculty of Business and Economics, Naamsestraat 69, 3000 Leuven, Belgium 2Vlerick Leuven Gent Management School, Vlamingenstraat 83, 3000 Leuven, Belgium 3Amsterdam-Amstelland Police, James Wattstraat 84, 1000 CG Amsterdam, The Netherlands 4Universiteit van Amsterdam Business School, Roetersstraat 11 1018 WB Amsterdam, The Netherlands {Jonas.Poelmans, Stijn.Viaene, Guido.Dedene}@econ.kuleuven.be Paul.Elzinga@amsterdam.politie.nl Abstract. The types of police inquiries performed these days are incredibly diverse. Often data processing architectures are not suited to cope with this diversity since most of the case data is still stored as unstructured text. In this paper Formal Concept Analysis (FCA) is showcased for its exploratory data analysis capabilities in discovering domestic violence intelligence from a dataset of unstructured police reports filed with the regional police Amsterdam-Amstelland in the Netherlands. From this data analysis it is shown that FCA can be a powerful instrument to operationally improve policing practice. For one, it is shown that the definition of domestic violence employed by the police is not always as clear as it should be, making it hard to use it effectively for classification purposes. In addition, this paper presents newly discovered knowledge for automatically classifying certain cases as either domestic or non-domestic violence is. Moreover, it provides practical advice for detecting incorrect classifications performed by police officers. A final aspect to be discussed is the problems encountered because of the sometimes unstructured way of working of police officers. The added value of this paper resides in both using FCA for exploratory data analysis, as well as with the application of FCA for the detection of domestic violence. Keywords: Formal Concept Analysis (FCA), domestic violence, knowledge discovery in databases, text mining, exploratory data analysis, knowledge enrichment, concept discovery 1 Introduction Concept discovery is a relatively new approach for discovering knowledge from textual information [10]. At the core of the method is the visualization of the underlying concepts of the data by means of Formal Concept Analysis (FCA) lattices [8, 9] which are interpreted, analysed and discussed by do- main experts. FCA arose twenty-five years ago as a mathematical theory [14] and has over the years grown into a powerful framework for data analysis, data visualization [15], information retrieval and text mining [16, 17, 20]. In this paper FCA is for the first time used as an exploratory data analysis and knowledge enrichment technique for police data. Compared to traditional black-box data mining tech- niques, this human-centred approach has the advantage of actively engaging expert knowledge in the discovery process. The goal of Intelligence Led Policing (ILP) is to complement intuition led police actions with information coming from analyses on aggregated operational data, such as crime figures and criminal characteristics [25, 26, 39]. While over 80% of all information available to police organizations CORE Metadata, citation and similar papers at core.ac.uk Provided by Research Papers in Economics https://core.ac.uk/display/6259772?utm_source=pdf&utm_medium=banner&utm_campaign=pdf-decoration-v1 2 resides in textual form, analysis has to date been primarily focused on the structured portion of the available data. Though text mining has been identified as a promising area in the formal framework for crime data mining by Chen et al. [27], this work has hardly found its way into mainstream scientific literature. One of the notorious exceptions is the paper by Ananyan [28] in which historical police reports were analysed to identify hidden patterns. According to the Ministry of Justice of the Netherlands, 45% of the population once fell victim to non-incidental domestic violence and for 27% of the population, the incidents even occurred on a weekly or daily basis [22]. These gloomy statistics brought this topic to the centre of the political agenda and made it to one of the pivotal projects of the Balkenende administration when it took office in 20031 and the Amsterdam-Amstelland police in the Netherlands [32]. Sufficient insight into the nature of domestic violence, being able to swiftly recognise suspicious cases and label reports accordingly is of the utmost importance. However, in the past intensive audits of the police databases related to filed reports established that many reports tended to be wrongly labelled as domestic or as non-domestic violence cases. In this paper we shall demonstrate the effectiveness of concept discovery methods for distilling new knowledge from the unstructured text in police reports. FCA amongst others helped us to improve the definition, the understanding by police officers and the management of the notion of domestic vi- olence. Additionally, we aim at automating detection of domestic violence from the unstructured text in police reports. The very first steps taken in this direction are described in [37] and in [38] an independent research track pursued in parallel with the work presented in this paper based on Emergent Self Organizing Maps is described. Although the usage of FCA for browsing text col- lections has been suggested before by Cole et al. [18, 35], almost none of these papers have focused on how FCA can be used for knowledge enrichment and for discovering different types of knowledge in unstructured text. Neither has it been thoroughly discussed in the literature how FCA can be used to incrementally construct and refine a high-quality domain-specific thesaurus (which is a prerequisite for developing an effective information retrieval system). Moreover, only minor attention has been paid to the possibilities offered by FCA to incorporate prior knowledge in the knowledge discovery 1 http://www.regering.nl/Het_kabinet/Eerdere_kabinetten/Kabinet_Balkenende_II/Regeerakkoord#internelink4 3 process. Finally, some of the aspects of this paper have already been discussed in the literature in a fragmented way (e.g. information retrieval, knowledge browsing), but an integrated approach has nev- er been pursued. FCA is particularly suited for exploratory data analysis because of its human-centredness. Repre- sentations that expose the underlying conceptual structure of the information promote the creation of new knowledge. What makes FCA an especially appealing technique for knowledge discovery in da- tabases from a practitioner’s point of view is the compactness of its information representation and the minimal need for users to tune (hyper-) parameters to distill a useful, actionable picture of the mining exercise. Concepts are the elementary units of human reasoning and this notion of concept is central to FCA [23, 24]. The underlying structure of the information is considered to be a concept system and FCA concept lattices are used to visualize the concepts and their interrelationships. These visual repre- sentations support human actors in their information discovery and knowledge creation exercise. This paper is composed as follows. In section 2 we describe the current situation of the domestic violence reporting procedure and previous attempts to improve the situation. In section 3 we cover the essentials of FCA theory, introducing the pivotal FCA notions of concept and concept lattice and describing the process of FCA for knowledge discovery. Section 4 elaborates on the dataset used in our research, while section 5 focuses on how this dataset was analysed and discusses the results of the application of FCA for exploratory analysis of domestic violence cases using this dataset. In section 6 the results of the domain exploration are validated. Finally, section 7 presents a number of concluding remarks. 2 Domestic violence discovery According to the U.S. Office on Violence against Women, domestic violence is a “pattern of abusive behavior in any relationship that is used by one partner to gain or maintain power and control over another intimate partner” [1]. Domestic violence can take the form of physical violence, which includes biting, pushing, maltreating, stabbing or even killing the victim. Physical violence is often accompanied by mental or emotional abuse, which includes insults and verbal threats of physical violence towards the victim, the self or others, including children. Domestic violence occurs all over 4 the world, in various cultures [2] and affects people throughout society, irrespective of economic status [3]. 2.1 Current situation The XPol database – the database of the Amsterdam-Amstelland police – contains most of the documents with regard to criminal offences. Documents related to certain types of crime receive corresponding labels. It is of the utmost importance that a correct label is assigned to each of the filed police reports. First, there are some legal consequences. If the police judged an incident to be domestic violence, the public prosecutor can accuse the offender of committing a domestic violence crime. This is taken into account by the judge as an aggravating circumstance, often resulting in a more severe penalty. Second, police officers will be able to better assess new incidents between the perpetrator and the victim, resulting in a more effective way of tackling the problem. Finally, if a domestic violence label was incorrectly assigned to a case, this will result in a waste of the valuable time of the police officers assigned to the case. Immediately after the reporting of a crime, police officers are given the possibility to judge whether or not it is a domestic violence case. If they believe it is, they can indicate this by assigning the label “domestic violence” to the report. However, not all domestic violence cases are recognised as such by police officers. This may have several reasons, for example, because of a lack of training, a lack of prior experience or new types of domestic violence occurring. As a consequence, many documents are lacking the appropriate label, which put on the agenda the need for a more efficient and effective case triage software program to automatically filter out suspicious cases for in-depth, manual inspection and classification. The in-place case triage system has been configured to filter out these reports for in- depth manual inspection and classification, with the aim of substantially reducing the number of domestic violence cases that are not recognised as such. It retrieves suspicious cases that lack the label of domestic violence and sends them back to the data quality management team. At present, each case retrieved by the in-place case triage system is subjected to an in-depth manual inspection by one of the co-workers of the quality control department. If analysis reveals that a case was wrongly classified as non-domestic violence, it is sent back to the police officer responsible for the case, who is obliged to re-examine and reclassify the police report. It is obvious that this is a very time-consuming and, by 5 consequence, costly procedure. Given that it takes an individual at least five minutes to read and classify a case, it is clear that more accurate triage will result in major savings. Currently the triage is based on either one or both of the following two criteria being met. The first criterion is whether the perpetrator and the victim live at the same address. The second criterion is whether any or a combination of the following expressions appear in the case documents: “ex- boyfriend”, “ex-girlfriend”, “ex-husband”, “ex-wife”, “domestic”, “stalk”, “lived together”, “live together”, “son and scared”, “child and scared”, “child and threat”, “son and threat”, “daughter and threat” or “daughter and scared”. Fig. 1. Current domestic violence reporting procedure A summary of the current domestic violence reporting procedure is displayed in Figure 1. There are several problems associated with this process. First, recent audits have confirmed that many of the retrieved cases are wrongly selected for in-depth manual inspection. Going back to 2006, the system retrieved 1157 cases, 80% of which actually turned out to be non-domestic violence cases. For example, going back to 2007, the triage system retrieved 1091 of such cases in which the victim made 6 a statement to the police. Second, because of a lack of manpower the data management quality team was not able to analyse each retrieved police report. Third, audits of the police databases revealed that not all domestic violence cases lacking the appropriate label were retrieved by the case triage system. Fourth, no actions have yet been undertaken to address the issue of the filed reports that were wrongly classified as domestic violence. 2.2 Previous attempts to resolve situation Previous attempts have mainly focused on developing a machine learning classifier that automatically classified cases as domestic or as non-domestic violence. Previously developed systems were mainly multi-layer perceptions that were trained on a dataset consisting of cases that were labelled by police officers as domestic or as non-domestic violence. These systems did not provide any insight into the problem, since they are black-boxes and their performance was around 80% only [31]. As a consequence, these systems never made it into operational policing practice. We found that a critical error was that the developers never performed an in-depth exploration of the data. They overlooked the complexity of the notion of domestic violence, were unaware that different people have different visions about the nature and scope of it and did not pay attention to niche cases. Moreover, the correctness of the labels assigned to cases by police officers was never verified. We found that different police officers regularly assigned different labels to the same situation. Finally, the developers did not dispose of a high-quality domain-specific thesaurus that contained sufficient discriminant terms for accurately classifying cases. 3 FCA knowledge discovery process This section introduces the main ideas of FCA and how it was used during the knowledge discovery process. According to R.S. Brachman and T. Anand [29], much attention and effort has been focused on the development of data mining techniques, but only a minor effort has been devoted to the development of tools that support the analyst in the overall discovery task. They argue for a more human-centred approach. Human-centred KDD refers to the constitutive character of human interpretation for the discovery of knowledge, and stresses the complex, interactive process of KDD as 7 being led by human thought. In most real-world knowledge discovery applications, an indispensable part of the discovery process is that the analyst explores and sifts through the raw data to become familiar with it and to get a feel for what the data may cover. Often an explicit specification of what one is looking for only arises during an interactive process of data exploration, analysis and segmentation. R.S. Brachman et al. [30] introduce the notion of data archeology for KDD tasks in which a precise specification of the discovery strategy, the crucial questions and the basic goals of the task have to be elaborated during an unpredictable exploration of the data. Data archeology can be considered as a highly human-centred process of asking, exploring, analysing, interpreting and learning by interacting with the underlying database. Comprehensible support should be provided to the analyst during the KDD process. According to Brachman et al. [29] this should be embedded into a knowledge discovery support environment. How the process of human-centred KDD can be supported by Formal Concept Analysis (FCA) was for the first time investigated by Stumme et al. [12]. Smyth et al. [33] already stated that the algorithm designer and the scientist should be able to bring in prior knowledge so the data mining algorithm does not just rediscover what is already known. Moreover, the scientist should be able to “ get inside” and “ steer” the direction of the data mining algorithm. FCA fulfils these requirements. Starting from initial knowledge on the problem area, it provides the user with a visual display of the relevant concepts available in the dataset and their relationships. Additionally, the user can visually interact with the concept lattice and thereby steer the knowledge discovery process. What makes FCA into an especially appealing technique for knowledge discovery in databases is that it meets the important requirement stated by, amongst others, Fayyad et al. [34] that data mining should be primarily concerned with making it easy, convenient and practical to explore very large databases for organizations and users with vast amounts of data but without years of training as data analysts. FCA offers the user an intuitive visual display of different types of structures available in the dataset and guides the user in the exploration of the dataset. This end-user-friendly interface also makes the data mining more transparent to the user. When compared to other, more traditional, techniques such as associates rules, FCA has a larger explanatory power because of its underlying non-hierarchical structure [36]. While traditional 8 association rules are flat, FCA provides an order of significance, which makes its representation richer and more intuitive to use. 3.1 FCA essentials Formal Concept Analysis is a recent mathematical technique that can be used as an unsupervised clustering technique [11, 13]. Police reports containing terms from the same term clusters are grouped in concepts. The starting point of the analysis is a database table consisting of rows M (i.e. objects), columns F (i.e. attributes) and crosses T M F⊆ × (i.e. relationships between objects and attributes). The mathematical structure used to represent such a cross table is called a formal context (T, M, F). An example of a cross table is displayed in Table 1. In this table reports of domestic violence (i.e. the objects) are related (i.e. the crosses) to a number of terms (i.e. the attributes); here a report is related to a term if the report contains this term. The dataset in Table 1 is an excerpt of the one we used in our research. Given a formal context, FCA then derives all concepts from this context and orders them according to a subconcept-superconcept relation, which results in a line diagram (a.k.a. lattice). Table 1. Example of a formal context kicking dad hits me stabbing cursing scratching maltreating report 1 X X X report 2 X X X report 3 X X X X X report 4 X report 5 X X The notion of concept is central to FCA. The way FCA looks at concepts is in line with the international standard ISO 704, which formulates the following definition. A concept is considered to be a unit of thought constituted of two parts: its extension and its intension, [14, 16]. The extension consists of all objects belonging to the concept, while the intension comprises all attributes shared by those objects. Let us illustrate the notion of concept of a formal context using the data in Table 1. For a set of objects O M⊆ , the common features, written ( )Oσ , can be identified via the following formula: ( ) { | : ( , ) }A O f F o O o f Tσ= = ∈ ∀ ∈ ∈ 9 Take the attributes that describe report 5 in Table 1, for example. By collecting all reports of this context that share these attributes, we get to a set O M⊆ consisting of reports 2, 3 and 5. This set O of objects is closely connected to set A consisting of the attributes “ cursing” and “ scratching.” ( ) { | : ( , ) }O A i M f A i f Tτ= = ∈ ∀ ∈ ∈ That is, O is the set of all objects sharing all attributes of A, and A is the set of all attributes that are valid descriptions for all the objects contained in O. Each such pair (O, A) is called a formal concept (or concept) of the given context. The set ( )A Oσ= is called the intent, while ( )O Aτ= is called the extent of the concept (O, A). There is a natural hierarchical ordering relation between the concepts of a given context that is called the subconcept-superconcept relation. 1 1 2 2 1 2 2 1( , ) ( , ) ( )O A O A O O A A⊆ ⇔ ⊆ ⇔ ⊆ A concept d 1 1( , )O A= is called a subconcept of a concept e 2 2( , )O A= (or equivalently, e is called a superconcept of a concept d) if the extent of d is a subset of the extent of e (or equivalently, if the intent of d is a superset of the intent of e). For example, the concept with intent “ cursing” , “ scratching” and “ stabbing” is a subconcept of a concept with intent “ cursing” and “ scratching.” With reference to Table 1, the extent of the latter is composed of reports 2, 3 and 5, while the extent of the former is composed of reports 2 and 3. The set of all concepts of a formal context combined with the subconcept-superconcept relation defined for these concepts gives rise to the mathematical structure of a complete lattice, called the concept lattice of the context, which is made accessible to human reasoning by using the representation of a (labelled) line diagram. The line diagram in Figure 1, for example, is a compact representation of the concept lattice of the formal context abstracted from Table 1. The circles or nodes in this line diagram represent the formal concepts. It displays only concepts that describe objects and is therefore a subpart of the concept lattice. The shaded boxes (upward) linked to a node represent the attributes used to name the concept. The non-shaded boxes (downward) linked to a node represent the objects used to name the concept. The information contained in the formal context of Table 1 can be distilled from the line diagram in Figure 1 by applying the following reading rule: an object “ g” is 10 described by an attribute “ m” if and only if there is an ascending path from the node named by “ g” to the node named by “ m” . For example, report 5 is described by the attributes “ cursing” and “ scratching” . Fig. 2. Line diagram corresponding to the context from Table 1 Retrieving the extension of a formal concept from a line diagram such as the one in Figure 2 implies collecting all objects on all paths leading down from the corresponding node. In this example, the objects associated with the third concept in row 3 are reports 2 and 3. To retrieve the intension of a formal concept, one traces all paths leading up from the corresponding node in order to collect all attributes. In this example the third concept in row 3 is defined by the attributes “ stabbing” , “ cursing” and “ scratching” . The top and bottom concepts in the lattice are special: the top concept contains all objects in its extension, whereas the bottom concept contains all attributes in its intension. A concept is a subconcept of all concepts that can be reached by travelling upward. This concept will inherit all 11 attributes associated with these superconcepts. Note that the extension of the concept with attributes “ kicking” and “ dad hits me” is empty. This does not mean that there is no report that contains these attributes. However, it does mean that there is no report containing only these two attributes. 3.2 Human-centred knowledge discovery with FCA In contrast to most data mining algorithms, the discovery process using FCA is human-centred. It is definitely not a black-box that runs and optimises without intervention beyond specifying initial model choices and parameters. During the mining process two persons, an exploratory data analyst and a domain expert, were the driving force behind the exploration and collaborated intensively. There was a continuous process of iterating back and forth between the FCA lattices and the police reports. This knowledge discovery process is summarised in Figure 3. It is an abstract description of the methodology that is displayed here, but this process will be exemplified in the results section. Fig. 3. Abstract human-centered FCA knowledge discovery process The process of using FCA for exploratory data analysis consists basically of iteratively applying the following process. A lattice is constructed by the exploratory data analyst based on the domain expert’s prior knowledge of the problem area, the police reports contained in the dataset and the terms contained in the thesaurus. The lattice provides a reduced search space to the domain expert, who then visually inspects and analyses this lattice paying special attention to anomalies and counter-intuitive facts. The latter provide a clear guideline to the exploratory data analyst and the domain expert in order for them to pursue their data exploration. The obtained results, together with the relevant prior 12 knowledge of the domain expert, are then incorporated into the existing visual representation, resulting in a new lattice. The FCA lattice can be considered as a knowledge browser. Our contention is that it allows for an effective interaction between the human actors and the underlying information. The focus of the use of the FCA technique is on truly gaining incremental insight into the problem area by optimally incorporating prior domain knowledge in learning cycles. This insight encompasses an enrichment as well as a validation of the correctness and the practical usefulness of existing prior knowledge. Additionally, FCA is used to enrich and refine the domain-specific thesaurus. This thesaurus plays a key role in the incremental knowledge discovery germane to our research. FCA is also used to discover missing values and inconstancies from police reports. Finally, FCA is used to investigate some important, significant aspects of operational policing practice concerning domestic violence cases and to discover accurate and comprehensible classification rules. Each of these aspects of the process will be described in more detail in section 5, where we comment on the empirical analysis and results. 4 Dataset The dataset we report on in this paper consists of a selection of 4814 police reports describing a whole range of violent incidents from the year 2007. The domestic violence cases for that period are a subset of this dataset. The 1091 cases selected by the in-place case triage system for 2007 are a subset of this dataset too. This latter selection came about by, amongst other things, filtering from a larger set those police reports that did not contain the reporting of a crime by a victim, which is necessary for establishing domestic violence. This happens, for example, when a police officer is sent to an incident and later on writes a report in which he/she mentions his/her findings, while the victim has not made an official statement to the police. The follow-up reports referring to previous cases were also removed from the initial set of reports. Ultimately, this gave rise to a set of 4814 reports that were used as input for our investigation. From these reports, the person who reported the crime, the suspect, the persons involved in the crime, the witnesses, the project code and the statement made by the victim 13 to the police were extracted. Of the 4814 reports, 1657 were classified as domestic violence; the others were not. An example of a report is displayed in Figure 4. Title of incident Violent incident xxx Reporting date 26-11-2007 Project code Domestic violence against seniors (+55) Crime location Amsterdam Keizersgracht yyy Suspect (male) Suspect (18-45yr) zzz Address Amsterdam Keizersgracht yyy Involved (male) Involved (18-45yr) Neighbours Address Amsterdam Keizersgracht www Victim (female) Victim (older than 45yr) uuu Address Amsterdam Keizersgracht vvv Reporting of the crime Last night I was attacked by my husband. I was watching television in the living room when he suddenly attacked me with a knife. I fell on the floor. Then he tried to kick me in my stomach. I tried to escape through the back door while I was yelling for help. I ran to the neighbours for help. They called the emergency services. Meanwhile my son ran away. My leg was bleeding; my head was bouncing, etc. Fig. 4. Example police report The validation set consists of a selection of 4738 cases describing a whole range of violent incidents from the year 2006 where the victim made a statement to the police. Again, the follow-up reports were first removed. Of these 4738 cases 1734 were classified as domestic violence by police officers. In 2006 the in-place case triage system retrieved 1157 police reports containing a statement made by the victim that had to be manually classified by police officers. 318 were classified as domestic violence, while 839 were classified as non-domestic violence. In addition to the set of reports, we had an initial thesaurus – a collection of 123 domain-specific terms – at our disposal, which was obtained by performing frequency analyses on the set of police reports. The terms that occurred most often were retrieved and added to the initially empty thesaurus. Each police report was then searched for each of these terms. The result was a cross table in which a cross indicated that the corresponding police report contained the corresponding term. 14 5 Analyses and results In this section, we showcase the possibilities of FCA as a knowledge discovery and knowledge enrichment technique. The knowledge discovery process using FCA is summarised and displayed in Figure 5. Fig. 5. Detailed human-centered knowledge discovery process using FCA It is clear that the process displayed in Figure 5 contains an iterative learning loop. Initially, an FCA lattice is constructed based on expert prior knowledge, the terms contained in the thesaurus and the police reports contained in the dataset. Then, the FCA lattice is analysed by the exploratory data analyst and domain expert. Based on the results obtained through the analysis process, which is described in the subsequent paragraphs and demonstrated in detail in the next subsections, a new lattice can be constructed. The FCA lattices are used as an instrument to discover new case labelling rules and to enrich, test and refine expert prior knowledge. Furthermore, the FCA lattices are used to browse and annotate the collection of police reports and efficiently select representative reports for in-depth manual inspection. 15 The first major aspect of the process consists in searching these reports for new attributes that can be used to discriminate between the domestic and non-domestic violence reports or that may lead to an enrichment of existing domain knowledge. New referential terms were not acquired and selected using a term extractor, but they were obtained by carefully reading some representative reports and then selecting relevant terms as attributes. We built in the necessary validation mechanisms such as using synonym lists, spelling checking, etc. to ensure the completeness of the thesaurus. During the research the thesaurus was under constant evolution: when new terms and concepts were discovered, the terms were added to the thesaurus. Because of the large number of police reports in the dataset, it was not possible to visually analyse concept lattices containing more than 14 attributes. Therefore, terms with a similar semantic meaning or referring to the same domain concept were clustered by the domain experts. When these term clusters were used to create an FCA lattice, they were considered as attributes. This approach ensured that the thesaurus remained at all times a reflection of the already gained knowledge. The second major aspect of the process consists of verifying the correctness of the labels assigned by police officers to the selected cases and searching the reports for missing values and inconsistencies. This allowed for the discovery of faulty case labellings and situations that were often not recognised by police officers as domestic or as non-domestic violence. This information was used by the data quality management team to significantly improve the quality of the data contained in the police databases and to improve the way police officers handle domestic violence cases. The information was also useful for the domestic violence programme manager to improve the training of police officers. We also found some regularly occurring confusing situations that could not be uniquely classified as domestic or non-domestic violence based on the domestic violence definition. These situations were presented to the programme manager and were used to enrich, improve and refine the concept and definition of domestic violence. The third major aspect of the process consists in discovering accurate and comprehensible case labelling rules to automatically classify cases as domestic or as non-domestic violence. In the past this turned out to be impossible. We found that this was largely due to the incorrect labels assigned by police officers to cases, to the vagueness of the domestic violence definition and to the lack of a high- 16 quality thesaurus. We managed to resolve many of these problems using FCA, resulting in a set of highly accurate and comprehensible classification rules. All these different aspects of the process, which have only been briefly introduced so far, are discussed more extensively in the next sections. 5.1 Domain exploration starting from expert prior knowledge In this section it is illustrated how we used prior knowledge to start and guide the exploration of the data. We based our initial lattice on the domestic violence definition, by clustering the terms contained in the thesaurus into term clusters associated with one of the two components of the definition (i.e. prior knowledge incorporation). The definition of domestic violence employed by the police organiza- tion of the Netherlands is as follows: “ Domestic violence can be characterised as serious acts of vi- olence committed by someone in the domestic sphere of the victim. Violence includes all forms of physical assault. The domestic sphere includes all partners, ex-partners, family members, relatives and family friends of the victim. The notion of family friend includes persons that have a friendly relation- ship with the victim and (regularly) meet with the victim in his/her home [6].” We intended to verify whether a report can be classified as domestic violence by checking it for the occurrence of one or more terms related to each of the two components of the domestic violence defi- nition. That is, a case can be labelled as domestic violence if the following two conditions are fulfilled. First, a criminal offence has occurred. This may range from verbal threats over pushing and kicking to even killing the victim. To verify whether a criminal offence has occurred, the report is searched for terms such as “ hit” , “ stab” and “ kick” . These terms are grouped into the term cluster “ acts of vio- lence” . Second, a person in the domestic circle of the victim is involved in the crime. It should be noted that a report is always written from the point of view of the victim and not from the point of view of the officer. A victim always adds “ my” , “ your” , “ her” and “ his” when referring to the persons involved in the crime. Therefore, the report is searched for terms such as “ my dad” , “ my mom” and “ my son” . These terms are grouped into the term cluster “ family members” . The report is also searched for terms such as “ my ex-boyfriend” , “ my ex-husband” , and “ my ex-wife” . These terms are grouped into the term cluster “ ex-partners” . Furthermore, the report is searched for terms such as “ my nephew” , “ her uncle” , “ my aunt” , “ my step-father” and “ his step-daughter” . These terms are grouped 17 under the term cluster “ relatives.” Then the report is searched for terms such as “ family friend” and “ co-occupant” . These terms are grouped into the term cluster “ family friends” . Reports that were assigned the label “ domestic violence” have been classified as such by police of- ficers. The remaining reports were classified as non-domestic violence. This results in the lattice dis- played in Figure 6. Fig. 6. Initial lattice based on the police reports from 2007 18 From an initial inspection of the lattice in Figure 6 it quickly became clear that a lattice containing only term clusters based on the starting definition of domestic violence would not discriminate sufficiently between domestic and non-domestic violence reports (i.e. knowledge enrichment). Many non-domestic violence reports seemed to also contain terms attributed to one or more of the term clusters (i.e. prior knowledge validation). Still, some interesting findings emerged from this lattice and triggered further investigation. These findings are discussed in the next section. The lattice structure also made it possible for us to discover the most frequently occurring types of domestic violence cases for 2007. These are summarised in Table 2. Table 2. Most frequently occurring types of domestic violence in 2007 % of all domestic violence cases of 2007 “ Acts of violence” and “ family members” and “ partners” 25% “ Acts of violence” and “ family members” and “ partners” and “ ex-persons” 16% “ Acts of violence” and “ family members” and “ ex-persons” 15% “ Acts of violence” and “ family members” 10% “ Acts of violence” and “ family members” and “ partners” and “ relatives” 6% “ Acts of violence” and “ partners” 5% 5.2 Prior knowledge testing and referential term discovery In this section it is demonstrated how we used prior knowledge to guide the exploration of the data. In contrast to what the domain expert initially thought, not all cases labelled as domestic violence by police officers contained terms associated with the two components of the definition (i.e. prior knowledge testing). This led to the discovery of cases that were assigned a wrong label by police officers (i.e. detection of faulty case labellings), to new domain-specific terms that were lacking in the original thesaurus (i.e. referential term discovery) and to a labelling error that was regularly made by police officers (i.e. improvement of training of police officers). Table 3. Interesting observations from the lattice in Figure 6 Non-domestic violence Domestic violence No “ acts of violence” 67 42 No “ acts of violence” and one or more of the persons clusters 61 19 19 Only “ acts of violence” 879 64 As can be seen from Table 3, a total of 61 (i.e. 42 and 19) domestic violence cases did not contain a term from the “ acts of violence” term cluster. Of these 61 cases 19 contained a term from one of the clusters containing terms referring to a person in the domestic sphere of the victim. After in-depth manual inspection of these 19 cases, it turned out that they contained other violence terms, such as “ abduction” , “ strangle” and “ deprivation of liberty” , which were lacking in the initial thesaurus. The remaining 42 cases, on the other hand, turned out to be wrongly classified as domestic violence. Interestingly, some 28% (i.e. 879) of the non-domestic violence reports only contain terms from the “ acts of violence” cluster, while there are only 64 domestic violence reports in the dataset that share that characteristic. Manual inspection, again, revealed that more than two thirds of these reports were wrongly classified as domestic violence. For some unknown reason, police officers regularly seem to misclassify burglary, car theft, bicycle theft and street robbery cases as domestic violence. Therefore, terms such as “ street robbery” , burglary” and “ car theft” were combined into a new term cluster called “ burglary cases” . 5.3 Term clustering and concept discovery In this section it is shown how new domain-specific terms, discovered by careful analysis of police reports, and terms with a similar semantic meaning, proposed by the domain expert, were clustered together in term clusters (i.e. term clustering). These term clusters led to the discovery of new concepts that were lacking in the domain expert’s conception of the problem area (i.e. concept discovery). Additionally, two new term clusters based on prior knowledge were introduced (i.e. prior knowledge incorporation). These new term-clusters were used to construct the second FCA lattice. When browsing a sample of the remaining police reports, we spotted some interesting terms that led to the discovery of two new and important concepts that were lacking in the domain expert’s conception of the problem area. The reports contained terms such as “ I had a relationship with” , “ relational problems” and “ marriage problems” . These terms typically refer to the concept of a broken relationship, which is why they were brought together into the cluster “ relational problems” . A 20 distinction was made between a broken relationship and an ongoing relationship. Terms such as “ I have a relationship with” and “ live together” were brought together in the cluster “ in a relationship” . According to the literature, domestic violence is a phenomenon that mainly occurs inside the house [4, 5, 6, 21]. Therefore, an attribute called “ private locations” was introduced. This term cluster contained terms such as “ bathroom” , “ living room” and “ bedroom” . An attribute called “ public locations” was also introduced. To summarise, although the lattice in Figure 5 could not be used to effectively distinguish domestic violence reports from non-domestic violence reports, it could be used to detect cases that were wrongly classified as domestic violence. Also, it helped in discovering new attributes that turned out to be missing in the user’ s understanding of the problem area. The redefined lattice structure, taking into account the above analyses, is displayed in Figure 7. In order to keep the lattice comprehensible, the terms belonging to the clusters “ family members” , “ relatives” , “ partners” , “ ex-partners” and “ family friends” have been lumped into a cluster “ persons” . 21 Fig. 7. First refined lattice based on the police reports from 2007 5.4 Detecting faulty case labellings and confusing situations In this section, we demonstrate how FCA was used to detect faulty case labellings and situations that are confusing to police officers. This was used to improve the training of police officers, to enrich and refine the domestic violence definition and to improve the quality of the data contained in the police databases. We also discovered new referential terms and clustered them based on their semantic meaning, leading to a further enrichment of the thesaurus and the existing domain knowledge. 22 It should be clear from the lattice in Figure 7 that the terms contained in the cluster “ relational problems” tend to be associated with domestic violence cases. Some of the more interesting observations from this lattice are displayed in Table 4. Table 4. Results from the lattice in Figure 7 Non-domestic violence Domestic violence “ relational problems” 58 365 “ private locations” 1340 1365 “ public locations” 1015 505 Apparently, only 58 non-domestic violence reports contained one or more terms from the “ relational problems” cluster. Further investigation revealed that a startling 95% of these cases had been wrongly classified as non-domestic violence. Moreover, about 70% of these cases had in common that a third person made a statement to the police for someone else. For example, one case described a father who made a statement to the police about the sexual abuse of his daughter by her stepfather. This is a clear case of domestic violence. But since it was not the victim who made the statement to the police, the police officer did not recognise it as such. Analysis of the remaining 30% of these misclassified cases led to the discovery of a new and important concept that was initially lacking from the domain expert’ s understanding of domestic violence. Many of the reports turned out to contain terms such as “ I was attacked by the new boyfriend of my ex girlfriend” and “ I was maltreated by the new girlfriend of my ex boyfriend” . These terms were grouped into the cluster “ attack by new friend of ex-person” . Police officers and policy makers confirmed that this type of situation was to be seen as domestic violence, mainly because the perpetrator often aims at emotionally hurting the ex-partner. Consequently, the expectation was for the terms contained in this cluster to frequently occur in domestic violence reports. However, this turned out to be incorrect. It became clear from the investigation that this type of situation in general was very confusing to police officers. A quick scan revealed that more than 50% of police officers actually had trouble with this. The ensuing investigation and discussions with police officers and policy makers revealed that this situation needed to be addressed during the training of police officers. Several interesting cases like the previous one were picked up during the data exploration. All of them gave rise to a clearer insight into the nature of domestic violence. 23 5.5 Prior knowledge incorporation and testing In this section we demonstrate how expert prior knowledge was incorporated into the FCA knowledge discovery process. It is also made clear how we used FCA to verify the correctness and the practical usefulness of this prior knowledge. Most of the domestic violence cases under scrutiny (1365 cases or 82%) contained one or more terms from the “ private locations” term cluster. However, 1340 (42%) of the non-domestic violence cases also contained one or more terms from this same term cluster. In addition, a hypothesis that was formulated prior to the data exploration was that almost no domestic violence case was expected to have taken place on the street. Surprisingly, this hypothesis was proven incorrect by the data. In about one-fourth of the domestic violence cases there had been an incident at a public location. While scrutinising these police reports, we discovered that this was often the case when ex-partners were involved. It became apparent that it was not possible to distinguish domestic from non-domestic violence reports by means of the type of locations mentioned in the reports. Combining the clusters “ private locations” and “ public locations” with clusters such as “ family members” or “ ex-persons” , for example, did not yield the expected results in terms of discriminatory power. 5.6 Definition refinement: niche cases In this section we focus on how FCA was used to enrich and refine the operationally employed domestic violence definition. Using FCA, we discovered multiple niche cases, which were presented to the domestic violence programme manager. This resulted in an enrichment of the domain knowledge, a refinement of the domestic violence definition and an improvement of the training of police officers. We continued our knowledge discovery exercise in search of additional attributes to help us distinguish domestic violence from non-domestic violence reports. We noticed that in a large number of the domestic violence cases (416 cases or 28%) the perpetrator and the victim happened to live at the same address at the time the victim made their statement to the police. Most of these cases (379 cases or 91%) were classified as domestic violence. When studying the remaining 37 non-domestic violence cases more carefully, we found, much to our surprise, that the perpetrator and the victim often lived together in the same institution (e.g. a youth institution, a prison or a retirement home). It 24 turned out that of the 41 cases where the perpetrator and the victim lived in the same institution only 30 actually had been classified as cases of domestic violence. This finding brought about a lively discussion amongst the police officers of the Amsterdam police force. More importantly, it exposed the discord amongst police officers on how to classify such cases. We took note of all their reflections and presented them to the board members responsible for the domestic violence policy. After intensive debate the following classification guidelines, displayed in Table 5, were obtained. Table 5. Classification guidelines for incidents involving inhabitants of the same institution Perpetrator Victim Classification Caretaker Inhabitant Domestic violence Inhabitant Caretaker Non-domestic violence Inhabitant younger than 18y Inhabitant younger than 18y Domestic violence Inhabitant older than 18y Inhabitant older than 18y Non-domestic violence Inhabitant of prison older than 18y Inhabitant of prison older than 18y Individual evaluation Inhabitant older than 18y Inhabitant younger than 18y Domestic violence Inhabitant younger than 18y Inhabitant older than 18y Individual evaluation The presence or absence of a dependency relationship between the perpetrator and the victim was in the end the decisive factor for classifying a case as either domestic or as non-domestic violence. The non-domestic violence cases where the perpetrator and the victim lived at the same address and were not inhabitants of an institution turned out to be wrongly classified as non-domestic violence. Therefore, a new attribute called “ institution” was introduced. 5.7 Missing values detection In this section, we demonstrate how FCA was used to detect missing values and inconsistencies in police reports. We also show how we exposed inefficiencies in the overall domestic violence policy employed by the police, using FCA.. Another interesting finding that emerged in our search for novel and potentially interesting classification attributes was that some 34% of the reports (1623 cases) did not mention a suspect. According to the domestic violence definition (which specifies that the perpetrator must belong to the domestic circle of the victim), the offender has to be known in domestic violence cases. Naturally, we 25 had assumed that these reports described non-domestic violence cases. Nevertheless, when looking into these cases, we found that 181 of them turned out to describe domestic violence cases after all. Analysis revealed that this was a result of police officers’ rather haphazard ways of registering victims for these cases. Apparently, while some officers immediately registered a suspect at the moment the victim mentioned this person as a suspect, others preferred to first interrogate them before casting the label of suspect. In the latter cases, the person then would just be added to the list of persons who were said to be involved in or witnessed the crime. Because such lists included friends, family members or bystanders, they could potentially be very extensive and diverse, which is why suspects easily got lost in these lists. When we inquired about the proper policy regarding the labelling of suspects, we were told there simply was none. Our analysis made a strong case for the need of such a policy. In the end, the quick-win proposal that could be implemented to solve this issue involved a relatively simple change to the registration software: an additional data entry field would need to be introduced for police officers to register the persons that were mentioned by the victim as offenders. Classification of police reports can only be performed on the basis of comprehensible and correct rules that do not inflate the false negative rate, while minimizing the false positive rate. Automatically assigning the non-domestic violence label to a case that does not mention a suspect is thus unacceptable because of the high false negative rate. Nevertheless, we found out that some 44% of the reports (711 cases) that lacked a labelled suspect did contain a description of the actual suspect. Of these 711 cases, only 16 reports were classified as domestic violence. After studying these 16 reports, we discovered that the majority of them were wrongly classified as domestic violence. Classifying cases as non-domestic violence because they lack a labelled suspect and contain a description of the suspect was thus acceptable. All of this newly discovered knowledge can once again be added to the lattice in Figure 7. When we introduce the attributes “ same address” , “ no suspect” and “ description of suspect” to this lattice, this results in the refined lattice structure displayed in Figure 8. 26 Fig. 8. Second refined lattice based on the police reports from 2007 The lattice in Figure 8 proved to be of much more use for discriminating domestic from non-domestic violence reports. We summarised some of the most interesting findings embedded in that lattice structure in Table 6. Table 6. Results from the lattice in Figure 8 Non-domestic violence Domestic violence Acts of violence and same address 37 379 Acts of violence and no suspect and description of suspect 695 16 Acts of violence and no suspect 1442 181 27 5.8 Discovering accurate and comprehensible classification rules In this section we focus on how we used FCA to discover accurate and comprehensible classification rules. We also illustrate how FCA can play a key role in detecting faulty case labellings. While further exploring the domestic violence reports, it became apparent that in many cases the victim made statements such as “ I want to institute legal proceedings against my husband” and “ I want to institute legal proceedings against my brother” . These sentences were brought together into the cluster “ legal proceedings against domestic sphere” . Another type of phrasing that was regularly used by victims of domestic violence was, for example, “ the crime was committed by my dad” or “ the crime was committed by my ex-boyfriend” . These sentences were brought together into the cluster “ committed by domestic sphere” . Yet another type of wording that was also frequently used by a victim was phrases such as “ I was maltreated by my husband” and “ I was threatened by my ex- partner” . These sentences in turn were brought together into the cluster “ threatened by domestic sphere” . Finally, neighbourhood quarrels (non-domestic violence) often made reference to phrases such as “ I want to institute legal proceedings against my neighbour” and “ committed by the man next door” , so these sentences were combined into the cluster “ neighbours” . Thus, the lattice was further refined and the result is displayed in Figure 9, with some of the most interesting facts summarised in Table 7. Table 7. Results from the lattice displayed in Figure 9 Non-domestic violence Domestic violence “ legal proceedings against domestic sphere” 19 266 “ committed by domestic sphere” 5 81 “ threatened by domestic sphere” 4 98 “ neighbors" 67 5 After browsing the 19 non-domestic violence cases in which the victim used one or more terms from the “ legal proceedings against domestic sphere” cluster, it turned out that these reports should have been classified as domestic violence. The same observation was made when the 5 non-domestic violence reports containing a term from the “ committed by domestic sphere” cluster and the 4 non- domestic violence cases containing a term from the “ threatened by domestic sphere” cluster were analysed. In-depth investigation of the 5 domestic violence cases in which a term from the 28 “ neighbours” cluster occurred, showed that these reports should have been classified as non-domestic violence. Fig. 9. Third refined lattice based on the police reports from 2007 5.9 Operational validation In this section, we clarify how FCA was used for the validation of some aspects of operational policing practice. For some specific situations it was verified whether police officers disposed of sufficient knowledge about the problem area to recognise these cases as domestic violence. Some very important special domestic violence situations were considered, including incest and honour-related violence. For the first type of situation, reports were searched for terms such as “ incest” and “ sexual abuse by my father” . For the second type of situation, reports were searched for terms such as “ marriage of convenience” and “ marry off” . The resulting lattice after incorporating these special cases is displayed in Figure 10. Table 8 summarises the classification. Table 8. Results from the lattice displayed in Figure 10 Non-domestic violence Domestic violence “ incest” 7 8 “ honor-related violence” 2 18 29 Careful inquiry into these cases taught us that police officers regularly misclassified incest cases as non-domestic violence. On the other hand, even for insiders it was quite surprising to observe how almost all honour-related violent incidents ended up being correctly classified as domestic violence. The latter was probably attributable to the intensive sensitisation campaigns organised to inform police officers of this important societal problem. Fig. 10. Fourth refined lattice based on the police reports from 2007 6 Validation experiment In this section we elaborate on the run-time power of the distilled knowledge. We start by mapping the proposed lattice structures obtained during discovery of the 2007 police data on the police reports from 2006. We demonstrate that the findings obtained through in-depth analysis of the 2007 police data are also valid for the police reports from 2006. Then, we apply the discovered knowledge to automatically classify the output of the in-place case triage system. Finally, we demonstrate how the newly discovered knowledge was used to detect and reclassify filed reports that were incorrectly labelled by police officers. 30 For the classification rules discovered in section 5, we verified how many domestic and non- domestic violence reports correspond to each rule. The rules and these counts are represented in Table 8. For the first eight rules, the non-domestic violence cases turned out to be incorrect labellings performed by police officers. For rules 9 up to 13, the domestic violence cases turned out to be incorrect labellings performed by police officers. Using rule 14, we found that in 160 cases that were classified as domestic violence by police officers a formally labelled suspect was lacking. Table 8. Discovered knowledge applied to police reports from 2006 Non-domestic violence Domestic violence Domestic violence rules 1. “ legal proceedings against domestic sphere” 24 237 2. “ committed by domestic sphere” 9 101 3. “ threatened by domestic sphere " 11 106 4. “ incest” 0 3 5. “ attack by new friend of ex-person” 6 12 6. “ relational problems” 61 364 7. “ same address” and not in “ institution” 23 299 8. “ honor-related violence” 1 16 Non-domestic violence rules 9. “ burglary cases” 32 24 10. “ neighbors” 13 6 11. “ no suspect” and “ description of suspect” 504 15 12. no “ acts of violence” 30 38 13. “ acts of violence” and no “ persons” 865 94 Data quality check extra 14. “ no suspect” 1074 160 For classification, the protocol is as follows. When a case comes in for labelling, the first step consists in verifying whether one of the domestic violence rules is satisfied. If this is the case, the case is classified as domestic violence. If the “ no suspect” or one of the non-domestic violence rules turns out to be also satisfied, the case is sent to the data quality management team, because there probably is a data quality problem. Otherwise, it is verified whether one of the non-domestic violence rules is satisfied. If this is the case, the case is classified as non-domestic violence. Otherwise, the case is left unclassified. By applying the first thirteen rules in Table 8, 50% of the dataset of 2006 could be automatically correctly classified). A further validation encompassed the application of the discovered knowledge to automatically classify the output of the in-place case triage system. For example, going back to 2006, the system retrieved 1157 cases, 80% of which actually turned out to be non-domestic violence cases. It is to deal 31 with these shortcomings in the current system that the rules in Table 8 will prove to be extremely useful. Some 9% of the cases contained terms from the “ committed by domestic sphere” , “ threatened by domestic sphere” or “ legal proceedings against domestic sphere” clusters and could be automatically classified as domestic violence. About 10% of the cases contained one or more terms from the “ relational problems” cluster and could for that reason be automatically classified as domestic violence. A further 11% of the retrieved cases could be classified as domestic violence simply because the perpetrator and the victim lived at the same address, which was not an institution. About 18% of the retrieved cases did not mention a suspect. If the policies we proposed had been implemented, these could all have been classified as non-domestic violence. Some 5% of the cases lacked a formally designated suspect but contained a description of a suspect. These cases could be classified as non- domestic violence. Another 14% of the cases retrieved by the triage system in 2006 could immediately be classified as non-domestic violence. They all contained one or more terms from the “ acts of violence” cluster and none from the “ persons” cluster. In sum, 514 of the 1157 cases retrieved by the triage system in 2007 could be correctly classified in an automated way when making use of the newly discovered knowledge. These findings are displayed in Table 9. Table 9. % of the 2006 cases classified automatically % retrieved cases classified automatically Current situation 0% Applying first 13 discovered rules from Table 8 44% Adding data field for suspect mentioned by victim to police registration form 54% The proposal that could be implemented to solve this issue involves a rather small change to the triage software: incorporating the first thirteen rules from Table 8 into the existing triage model. As a result, about 44% of the retrieved cases will be automatically classified correctly. Moreover, if an additional data field for the suspect mentioned by the victim is added to the police registration form, the fourteenth rule of Table 8 can also be integrated into the triage model. This would result in an automatic and correct classification of about 54% of the retrieved cases. An additional result is that a large number of the filed reports that were wrongly classified can now be automatically detected and corrected, the results of which are displayed in Table 10. 32 Table 10. Number of filed reports that were incorrectly classified, but corrected by means of the 13 rules Non-domestic corrected to Domestic Domestic corrected to Non-domestic Total Year 2006 135 110 245 Year 2007 124 88 212 First quarter 2008 54 24 78 Using the newly discovered rules, many of these incorrectly classified police reports can be automatically detected and reclassified. For example, for the year 2007, we found 212 filed police reports that were incorrectly classified. 7 Conclusions Domestic violence is one of the top priorities of the Amsterdam-Amstelland police. When a victim makes a statement to the police, police officers are given the possibility to indicate whether it is a domestic violence case. Still, this has proven to be problematic. The use of FCA , however, can play a significant role in overcoming some of the hurdles encountered when dealing with domestic violence cases. This paper specifically showcased the possibilities of using FCA for knowledge discovery from police reports. The FCA lattices prove to be very useful as knowledge browsers. The construction of an initial lattice containing term clusters created by a domain expert on the basis of the domestic violence definition and the incremental refinement of this lattice was shown to provide a powerful framework for exploring unstructured data. First, it was shown that the domestic violence definition is too vague, making it hard to use it effectively for classification purposes. Moreover, the scope of terms such as ex-partners and violence, was nowhere communicated in the definition. Second, we exposed that there exists a considerable amount of confusion amongst police officers about the nature and scope of domestic violence. Regularly occurring domestic violence situations such as incest or an ex-boyfriend attacking the new boyfriend of a girl were often not recognised as such by police officers. Third, using FCA, we were able to discover some essential characteristics that discriminate domestic from non- domestic violence reports. These characteristics include phrasings, words and word combinations that typically occur in either domestic or non-domestic violence cases. This newly discovered knowledge was then used to automatically assign a label to the cases retrieved by the in-place case triage system. It turned out to be possible to automatically and correctly 33 classify about 44% of the cases that used to be set aside for manual inspection. Moreover, a large part of the filed reports that were incorrectly classified, could be automatically detected and reclassified. Acknowledgements The authors would like to thank the police of Amsterdam-Amstelland for granting them the liberty to conduct and publish this research. In particular, we are most grateful to Deputy Police Chief Reinder Doeleman and Police Chief Hans Schönfeld for their continued support. Jonas Poelmans is aspirant of the Fonds Voor Wetenschappelijk Onderzoek – Vlaanderen or Research Foundation – Flanders. References [1] Office on Violence against Women (2007) About Domestic Violence (http:www.usdoj.gov/ovw/domviolence.htm). Retrieved on 2007-10-22 [2] Watts, C., Timmerman, C. (2002) Violence against women: global scope and magnitude. The Lancet 359 (9313): pp.1232-1237. RMID 1155557 [3] Waits, K. (1985). The criminal Justice System’ s response to Battering: Understanding the problem, forging the solutions. Washington Law Review 60: pp. 267-330 [4] Vincent, J.P., Jouriles, E.N. (2000) Domestic violence. Guidelines for research-informed practice. Jessica Kingsley Publishers Londen and Philadelphia [5] Black, C.M. (1999) Domestic violence: Findings from a new British Crime Survey self-completion questionnaire. London: Home Office Research Study. [6] Keus, R., Kruijff, M.S. (2000) Huiselijk geweld, draaiboek voor de aanpak. Directie Preventie, Jeugd en Sanctiebeleid van de Nederlandse justitie. [7] Yevtushenko, S.A. (2000). System of data analysis “ Concept Explorer.” Proceedings of the 7th national conference or Artificial Intelligence. KII-2000. 127-134, Russia [8] Ganter, B., Wille, R. (1999), Formal Concept Analysis: Mathematical Foundations. Springer, Heidelberg. [9] Wille, R. (1982), Restructuring lattice theory: an approach based on hierarchies of concepts, I. Rival (ed.). Ordered sets. Reidel, Dordrecht-Boston, 445-470. [10] Poelmans, J., Elzinga, P., Viaene, S., Dedene, G. (2010), Formal Concept Analysis in knowledge discovery: a survey. Lecture Notes in Computer Science, 6208, 139-153, 18th international conference on conceptual structures (ICCS): from information to intelligence. 26 - 30 July, Kuching, Sarawak, Malaysia. Springer. [11] Wille, R. (2002), Why can concept lattices support knowledge discovery in databases?, Journal of Experimental & Theoretical Artificial Intelligence, 14: 2, 81-92. [12] Stumme, G., Wille, R., Wille, U. (1998), Conceptual Knowledge Discovery in Databases Using Formal Concept Analysis Methods, In: J.M. Zytkow, M. Quafofou (eds.): Principles of Data Mining and Knowledge Discovery, Proc. 2 nd European Symposium on PKDD ’ 98, LNAI 1510, Springer, Heidelberg, 1998, 450- 458. [13] Stumme, G. (2002) Efficient Data Mining Based on Formal Concept Analysis. Lecture Notes in Computer Science Vol. 2453, Springer, Heidelberg, 3-22 [14] Stumme, G. (2002), Formal Concept Analysis on its Way from Mathematics to Computer Science. Proc. 10th Intl. Conf. on Conceptual Structures (ICCS 2002). LNCS, Springer, Heidelberg 2002. [15] Priss, U. (2000), Lattice-based information Retrieval. Knowledge Organization, 27, 3, 132-142. [16] Godin, R., Gescei, J., Pichet, C. (1989), Design of browsing interface for information retrieval. In: N.J.Belkin, C.J. van Rijsbergen (Eds.), Proc. SIGIR ’ 89, 32-39. [17] Carpineto, C., Romano, G. (2005), Using concept lattices for text retrieval and mining. In Formal Concept Analysis-State of the Art, Proc. of the first International Conference on Formal Concept Analysis, Berlin, Springer. [18] Cole, R. , Eklund, P. (2001), Browsing Semi-structured Web Texts Using Formal Concept Analysis. In H. Delugach, G., Stumme (Eds.), Conceptual Structures: Broadening the Base, LNAI 2120, Berlin, Springer, 319-332. 34 [19] Eklund, P., Ducrou, J., Brawn, P. (2004), Concept Lattices for Information Visualization: Can Novice Read Line Diagrams? In P. Eklund (Ed.), Concept lattices: Second International Conference on Formal Concept Analysis, LNCS 2961, Berlin, Springer, 14-27. [20] Priss, U. (1997), A Graphical Interface for Document Retrieval Based on Formal Concept Analysis. In: E. Santos (Ed.), Proc. of the 8th Midwest Artificial Intelligence and Cognitive Science Conference. AAAI Technical Report CF-97-01, 66-70. [21] Beke, B.M.W.A., Bottenberg, M. (2003) De vele gezichten van huiselijk geweld. In opdracht van Programma Bureau Veilig / Gemeente Rotterdam. Uitgeverij SWP Amsterdam. [22] T. van Dijk, Huiselijk geweld, aard, omvang en hulpverlening (Ministerie van Justitie, Dienst Preventie, Jeugd-bescherming en Reclassering, oktober 1997). [23] Peirce, Ch. S. (1992), Reasoning and the logic of Things : The Cambridge Conferences lectures of 1898 edited by K. L. Ketner and H. Putman, Cambridge: Harvard University Press. [24] Arnauld, A., Nicole, P. (1985), la logique ou l’ Art de penser. Edition Gallimard . [25] Collier, P.M. (2006) Policing and the intelligent application of knowledge. Public money & management. Vol. 26, No. 2, pp. 109-116. [26] Collier, P.M., Edwards, J.S. and Shaw, D. (2004) Communicating knowledge about police performance. International Journal of Productivity & Performance Management. Vol. 53, No. 5, pp. 458-467 [27] Chen, H., Chung, W., Xu, J.J., Wang, G., Qin, Y., Chau, M. (2004) Crime data mining: a general frame- work and some examples. IEEE Computer, April 2004. [28] Ananyan, S. (2002) Crime Pattern Analysis Through Text Mining. Proceedings of the Tenth Americas Con- ference on Information Systems, New York, New York, August 2004. [29] Brachman, R., Anand, T. (1996) The process of knowledge discovery in databases: a human-centered ap- proach. In advances in knowledge discovery and data mining, ed. U. Fayyad, G. Piatetsky-Shapiro, P. Smyth and R. Uthurusamy. AAAI/MIT Press [30] Brachman, R.J., Selfridge, P.G., Terveen, L.G., Altman, B., Borgida, A., Halper, F., Kirk, T., Lazar, A., Mc Guinnes, D.L. and Resnick, L.A. (1993) Integrated support for data archaeology. International Journal of In telligent and Cooperative Information Systems, 2:159-185. [31] Raaijmakers, S.A., Kraaij, W., Dietz, J.B. (2007) Automatische detectie van huiselijk geweld in processen- verbaal. TNO-rapport 34293. [32] Politie Amsterdam-Amstelland (2008) http://www.politie-amsterdam-amstelland.nl/get.cfm?id=86, Retrieved on 2008-02-22. [33] Smyth, P., Pregibon, D., Faloutsos, C. (2002) Data-driven evolution of data mining algorithms. Communica tions of the ACM, Vol. 45, no. 8. [34] Fayyad, U., Uthurusamy, R. (2002) Evolving data mining into solutions for insights. Communications of the ACM, Vol. 45, no. 8 [35] Cole, R.J. (2000) The management and visualization of document collections using Formal Concept Analy- sis. Ph. D. Thesis, Griffith University. [36] Christopher, A. (1965) A city is not a tree. Architectural Forum, Vol 122, No 1, April 1965, pp 58-62 (Part I) and Vol 122, No 2, May 1965, pp 58-62 (Part II) [37] Poelmans, J., Elzinga, P., Viaene, S., Dedene, G. (2008). An exploration into the power of formal concept analysis for domestic violence analysis, Lecture Notes in Computer Science, 5077, 404 – 416, Advances in Data Mining. Applications and Theoretical Aspects, 8th Industrial Conference (ICDM), Leipzig, Germany, July 16-18, 2008, Springer. [38]Poelmans, J., Elzinga, P., Viaene, S., Van Hulle, M. & Dedene G. (2009). Gaining insight in domestic violence with emergent self organizing maps, Expert systems with applications, 36, (9), 11864 – 11874. [39] Viaene S., De Hertogh S., Lutin L., Maandag A., den Hengst S., Doeleman R. (2009). Intelligence-led policing at the Amsterdam-Amstelland police department: operationalized business intelligence with an enterprise ambition. Intelligent systems in accounting, finance and management. 16 (4) : 279 -292. 
work_46baihgfgfdi7ms3f2zcfabmu4	----	untitled The Knowledge Engineering Review, Vol. 26:1, 19–24. & Cambridge University Press, 2011 doi:10.1017/S0269888910000366 From expert systems to context-based intelligent assistant systems: a testimony P A T R I C K B R É Z I L L O N LIP6, University Pierre et Marie Curie - UPMC, 4 Place Jussieu, 75005, Paris, France e-mail: Patrick.Brezillon@lip6.fr Abstract This paper presents a personal interpretation of the evolution of artificial intelligence (AI) systems during these last 25 years. This evolution is presented along five generations of AI systems, namely expert systems, joint cognitive systems, intelligent systems, intelligent assistant systems, and the coming generation of context-based intelligent assistant systems. Our testimony relies on different real-world applications in different domains, especially for the French national power company, the subway companies in Paris and in Rio de Janeiro, in medicine, a platform for e-maintenance, road safety, and open sources. Our main claim is to underline that the next generation of AI systems (context-based intelligent assistant systems) requires a radically different consideration on context and its relations with the users, the task at hand, the situation, and the environment in which the task is accomplish by the user; the observation of users through their behaviors and not a profile library; a robust conceptual framework for modeling and managing context; and a computational tool for representing in a uniform way pieces of knowledge, of reasoning, and of contexts. 1 Introduction This paper is a testimony of 20 years of research in artificial intelligence (AI). My initial interest for AI presented several facets. First, coming from the domain of mathematical modeling and having faced the limits of this approach in real-world applications, I was curious to study how AI tools (especially expert systems) could be a modeling tool for some problems in which human reasoning played a key role. Second, AI belongs to cognitive sciences and has to deal with different disciplines such as cognitive psychology or neurosciences far from the engineering viewpoint. Third, the initial paradigm of separation of the representation of the knowledge from its use was a powerful alternative to the usual approach in software engineering, at least in the design stage. Fourth, AI has been (and is always) a battlefield where ‘hard’ science (i.e. theory-based science) was opposed to ‘soft’ science (i.e. cognition-based science). For example, the mathematical logic bloomed in a large variety of logics attempting each to deal with a specific aspect of cognition. Fifth, AI is a domain where the success of an AI system is measured by its disappearance in classical software. The first concept that plays an important role in AI is knowledge, and more specifically its nature. Education provides students with theoretical knowledge, when enterprises ask for their employees to have operational knowledge. For example, the Ohm’s law is learned through instantiations where R has an exact value (say, 3.0 V) when real resistances have a value with a limited accuracy that may play a crucial role in an electronic circuit. The transformation of theoretical knowledge into operational knowledge is a process of contextualization during which we adapt our knowledge to the task, the situation, and the local environment. Expertise is contextualized knowledge. A second point is the difference between data, information, and knowledge. Information is data with meaning that results from a process of interpretation. For example, a temperature of 248C (the data) will be interpreted differently by a person living in the Polar circle (248C is hot) and a person living near the equator (248C is cold). This is due to the fact that their mental models to interpret the data are different. Relevant information enriches the person’s knowledge (the person can establishes a link between the information and his mental model). A third point concerns procedures and practices. Organizations develop procedures that are collections of secure action sequences to address a given focus in any case. As a consequence, actors prefer to plan their action in real time rather than to rely on procedures for two main reasons. First, the procedure is never perfectly adapted to any specific situation and can lead to improper actions or sub-optimal incident resolution strategies. Second, a practice captures advantages and disadvantages of the situation at hand when a procedure tries to capture only the abstracted solution. Given a practical unit of reasoning: if P1;P2; . . . then C1 In a logical and theoretical reasoning (a procedure), the action leads to the conclusion C1 that is added to the base alone. A practical reasoning is an inductive probabilistic reasoning, the con- clusion cannot be detached from the premises, that is, take meaning out of premises: C1 is added to the base linked to P1, P2, etc. This means that knowledge is defined in a context of use that is lost in procedures because procedures are supposed to be applicable to a large class of problems when a practice is a contextualized version of the procedure in a specific context. As a side effect, this implicitly shows that there is not always a unique and optimal solution, but a set of best solutions in given contexts. It is a problem clearly identified in different disciplines: prescribed task versus effective task, task versus activity, logic of functioning versus logic of use. The difference between procedure and practice is a question of knowledge organization: parallel structures for procedures and sequential structures (and simplification of the domain knowledge from the user’s viewpoint) for practices. The role played by context between procedure and practice in real-world applications is, in my view, at the origin of successive mutations in AI that occurred during these past 20 years. 2 The time of expert systems The paradigm of separation of the representation of knowledge from its use, is born from the attack of ill-defined problems (expertise, heuristics, combinatorial, etc.) with no algorithmic solutions. This corresponds to a shift of the interest from procedural knowledge (algorithm, software, etc.) to declarative knowledge (fact base, knowledge base, rule engine), although expert systems face both types of knowledge. Knowledge was represented in flat knowledge bases, easy to update without a need to recompile the software because an inference engine made the use of the knowledge removed from the representation. The rule base frame is employed to represent rules constraining potential courses of action, which are treated as non-negotiable (‘facts of nature’, prescriptions from authorities). However, this was a source of problem because part of the knowledge comes often from technical domains where knowledge organization and structure were not represented explicitly. This leads to the introduction of some implicit knowledge on the use of the knowledge. A first option concerned the introduction of screening clauses. Figure 1 presents an example of a rule in MYCIN. Buchanan and Shortliffe (1984) proposed to add screening clauses to the condition part of rules, which were then activated only in some kind of context. For example, the 4th clause in the rule above (taken from MYCIN) constrains the activation of the rule, but is not a clause for identifying the infection. Clause 4 acts to restrict the rule application to the context of adults (the implicit reason in the rule is that tetracycline gives a violet color to teeth that is a problem for children). Another option was the introduction of rule packets as a structuring factor of rule bases. The reason was, say in equipment diagnosis as in Figure 2, that one does not need to consider the knowledge on a protective relay when one diagnoses a circuit breaker. 20 P . B R É Z I L L O N Here, it is the whole IF part that acts as the context of a circuit breaker. The explicit use of context in such a rule offers also two advantages: (1) the rule packet, although valid for all the circuit-breakers (about 20 in this application for the French national power company, Electricité de France), is triggered for the unique circuit breaker ‘$cb’ in the cut-off system ‘$cos’; and (2) the triggering of this rule allows to inhibit the test of other equipment pieces (Freeze ‘check_pw’ and ‘check_teac’) of the cut-off system. Three lessons learned at the end of this phase are: (1) the need of a user-centered approach, (2) need to deal explicitly with context, and (3) need of new methodologies different from classical engineering. 3 The time of joint cognitive systems The expression ‘joint cognitive system’ was coined to stress the fact that neither the system nor the user is able to solve the problem at hand alone (Woods, 1985). The emphasis was on the word ‘cognitive’. In a broad sense, this implies that the system and the user are able to understand each other (or more realistically make compatible their interpretations (Karsenty and Brézillon, 1995)) during problem solving. This entails strong requirements about the design of the system as: the sharing of the cognitive representations, an agreement about the contextual knowledge of the problem solving, the ability to follow each others reasoning, the capability of exchanging expla- nations, the capability to acquire incrementally new knowledge and learn new practices from the other, and a sharing of the interaction control. A key point here is the user and system com- plementarity to exploit their agent asymmetry. The complementarity implies that the system may contain either some knowledge (model or data) or functions (e.g. database exploitation) that really complements the user’s skills. However, joint cognitive systems focused on the task at hand, when some tasks such as knowledge acquisition, practice learning and explanation are parts of the task at hand and managed by the system as well as the user, because you cannot anticipate or ‘design away’ all the problems that might arise during user–system interaction. Thus, the system must know how to acquire knowledge incrementally when needed. This was not possible as long as context was not made explicit. IF 1. The infection which requires therapy is meningitis, 2. Only circumstantial evidence is available for this case, 3. The type of meningitis is bacterial, 4. The age of the patient is greater than 17 years old, and 5. The patient is an alcoholic, THEN There is evidence that the organisms, which might be causing the infection, are diplococcus-pneumoniae (.3) or e.coli (.2) Figure 1 Example of screening clause ! check_cb "Test of the circuit breaker in the control system $name" > if failure freeze check_pw, check_teac IF equipment_piece ($cos) := ($cb) , nature ($cb) := circuit_breaker . THEN call the rule packet 'Circuit-Breaker_Diagnosis($cos, $cb)' Figure 2 A rule calling a rule packet From expert systems to context-based intelligent assistant systems 21 The interest here is not on the probability of a branch but about the possibility to determine, as soon as possible, on which path the operator is, to determine what is the next action to undertake. In other words, each state of nature (a sequence of events is a state of nature) describes a state of the world, or using our words, a context for action. As many contextual elements may intervene in several scenarios (e.g. traffic activity, position of the next train), operators prefer to take them into account as soon as possible to get a general picture of the best path to choose. At this step, contextual knowledge is proceduralized and in the meantime, operators postpone actions in sequences like macro-actions. The proceduralized context corresponds to the needed knowledge for addressing the focus of the operators. The main objective is to eliminate event nodes and to use contextual information for the choice of the actions. Previously, users were considered as data gatherers with a total lack of understanding of the system behavior (i.e. not on the expert knowledge in the machine, but the way in which the system used this knowledge). The system relies on a logic of functioning often far from the logic of use of users. Thus, for completing or modifying knowledge in the system, users prefer to intervene superficially by adding new rules to control the firing of existing rules. As a consequence, the system rapidly becomes intractable. Humans usually function in an anticipative mode; operators explicitly check hypotheses to eliminate discovery or surprise. An anticipatory system would use knowledge (as a model itself and the relevant part of its environment) about future states to decide what action to take at the moment. An intelligent system should predict what will probably happen and preadapt itself for the occurrence of a crucial or time-critical event. An anticipatory capability supposes that the system has a simulation component. 4 The time of intelligent systems In Frederik Pohl’s (1969) science fiction novel, The Age of Pussyfoot, each person needs a mobile device—a ‘satisfactor’—to buy products, communicate with others, watch television, control the children, and even receive medications. In this future world, computing-centered humans cannot function in society without their satisfactors. This type of computing is now everywhere: mobile phones, intelligent houses, and soon communicating objects. However, such (new) systems pro- pose services, integrate few contextual cues (GPS, weather, etc.) and a user model often coming from statistical consideration. However, it is not the right way because the line of reasoning of the user does not deal with the line of reasoning of the system. An intelligent system must present different properties like: > Provide users with a first approximation of environmental trends and events, > Point out useful information implicit in large volumes of data to alert users to sudden changes, > Develop multiple scenarios and perspectives on a given line of action, > Attract user attention to existing and emerging strategic issues, > Support users in sharing and communicating their views and perspectives, > Guide user attention to specific data and its interpretation in relation to particular issues. With such qualities, an intelligent system must be more considered as an intelligent assistant system, like a secretary with her boss, solving tactical and operational problems and preparing elements to her boss for political and strategic problems. For example, context-aware applications can be assimilated to intelligent systems. Note that the role of the system with respect to the user is more precise than in joint cognitive systems. There is an evolution from the problem-solver view to the information-processing view. In the e-business arena, one speaks of the push model. 5 The time of intelligent assistant systems An intelligent assistant system (IAS) has two sides: the process side and the human side. An IAS has to understand: (a) the real-world process, (b) the tasks at hand, and (c) the operator’s 22 P . B R É Z I L L O N behavior. The IAS may learn (i.e. acquire knowledge) by observing the behavior of the operators and the process. This is an incremental knowledge acquisition process where knowledge is acquired when needed and, above all, in its context of use. Only a kernel of knowledge has to be elicited directly, mainly to select the right formal representation for the knowledge. Depending on the operator’s experience, this cooperation can take one of two forms: a waking state if there is no discrepancy between the operator’s actions and those of the system, and a cooperative state if a difference is noted relative to either the operator’s action or its possible consequences in the future. In the wake state, the IAS must follow the operators’ actions on the process, observe the corresponding effect on the process’ behavior and compare it to a simulated behavior. The IAS does not intervene, if there is no discrepancy between the observed and simulated behaviors of the process. When a problem is detected, the IAS first waits to allow time for the operator to react and, only after a while, alert the operator about the problem. The main reason is to not disturb the operator in a critical situation. If the operator does not see the problem, then, the system enters a cooperative state. However, more frequently, it is the operator that triggers the cooperative state. As the operator’s assistant, the system must automatically solve minor problems and prepare elements for complex problems that the operator will have to solve. It must also benefit from operator interaction to learn new practices and acquire new knowledge when it faces failure. For example, when the user overrides a decision for various reasons, the assistant will not understand this action if they have incompatible internal representations. If the same decision must be made in altered circumstances (i.e. in other contexts), the assistant must know when it is out of its domain and again ask the user to acquire the current context of use for the decision. In the e-business arena, one speaks of the pull model. Generally, there are different methods for executing a task, and the choice of a method relies heavily on contextual cues and operators’ experience. Clearly, an IAS must be designed and developed in a formalism providing a uniform representation of knowledge, reasoning, and context elements. Moreover, the formalism must allow incremental knowledge acquisition and practice learning. Thus, the human side begins to be addressed, but not yet the context side. 6 The time of context-based intelligent assistant systems As discussed above, the importance of context was clear since a long time. However, it is at the beginning of the 90s that a community appeared in AI with McCarthy (1993) as a leader. He formalized context as first class objects and introduced the basic relation ist(c, p) (the proposition p is true in the context c), the formula being always asserted in a context. As a consequence, McCarthy showed that: > context is always relative to another context, > contexts have an infinite dimension, > contexts cannot be described completely, > when several contexts occur in a discussion, there is a common context above all of them in which terms and predicates can be lifted. Now, a number of disciplines have a community of context, about context-aware applications, ubiquitous computing, pervasive computing, context-driven testing, etc. For our purpose in this paper, we retain that context is the key factor of intelligent assistant systems. Making context explicit allows us to use knowledge in its context of use, to capture variants in the reasoning (e.g. recording practices effectively developed by operators), to generate relevant explanations. After 15 years, there is a conceptual framework for context modeling and management and its implementation in a piece of software called Contextual Graphs. The working definition of context is that ‘context constrains a focus without intervening in it explicitly’. As a consequence, (1) context is relative to the focus, (2) as the focus evolves, its context evolves too, and (3) context is highly domain-dependent. Linking context to a focus, one can deduce that the focus divides From expert systems to context-based intelligent assistant systems 23 context into external knowledge and contextual knowledge. The latter constitutes a kind of tank where contextual elements are related to the focus in a flat way, whereas the former has nothing to do with the focus at the moment. The focus evolves because a new event occurs (e.g. an unpre- dicted event) or as a result of a decision made at the previous stage of the focus. For addressing the focus, the actor selects a subset of contextual knowledge that is organized, assembled, and structured in a proceduralized context that is used at the current focus. A following step concerns how to represent knowledge in context with the notion of ‘focus dressing’, and a three-layer model of context (a meta-model, a domain model, and instantiation). A practice representing the development of an actor’s reasoning as a kind of contextualization of a procedure, there is a formalism that allows us to represent in a uniform way, the elements of knowledge and the context in which they are to be used in a reasoning. The formalism called Contextual Graphs has been used in several real-world applications in different domains. Con- textual graphs are the building blocks of ‘experience bases’ on which an IAS can reason and accompany a user in the realization of his task. 7 Conclusion Evolution of AI systems comes from a move of the focus on the task alone (Expert System time), through tasks and user (Joint Cognitive System time), after to the user accomplish a task in a situation (Intelligent System time) to a holistic view on the user, the task, the situation, and the immediate environment with an emphasis on context. Thus, this view of AI systems is one that travel smoothly from an engineering viewpoint to a cognitive science viewpoint in which context plays a central role in the modeling and the representation of the knowledge and the reasoning. A first consequence is that it is not possible to reason in an abstract way: context is deeply embedded in the actor, the task at hand, the situation and the environment, and context is intimately related to a focus of attention. A second consequence is that most of the communities interested in context management now prefer to study context by a bottom-up approach instead of the previous top–down approach. However, it is always important to work in a coherent conceptual frame- work, with an operational definition of context, a modeling of context linked to other elements (focus, task, user, etc.), and to use this solid framework for developing relevant tools for analysis knowledge and reasoning in context. We do this with contextual graphs for representing users’ behaviors (i.e. practices instead of procedures), but there are other aspects to study, like a description of a focus or a situation in terms of contextual elements. Finally, the current approaches try to catch what humans are doing, but one will soon have to face some challenge like representing time in context and develop AI systems that are able to predict new behaviors. References Buchanan, B. G. & Shortliffe, E. H. 1984. Rule Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Addison-Wesley. Karsenty, L. & Brézillon, P. 1995. Cooperative problem solving and explanation. International Journal of Expert Systems With Applications 8(4), 445–462. McCarthy, J. 1993. Notes on formalizing context. In Proceedings of the 13th International Joint Conference on Artificial Intelligence 1, 555–560. Pohl, F. 1969. The Age of Pussyfoot. Ballantine. Woods, D. D. 1985. Cognitive technologies: the design of joint human-machine cognitive systems. AI Magazine 6(4), 86–92. 24 P . B R É Z I L L O N 
work_46wtlsghvvcntlousxuj3tstfm	----	PII: S0957-4174(02)00189-6 A recommendation mechanism for contextualized mobile advertising Soe-Tsyr Yuan*, Y.W. Tsao Department of Information Management, Fu-Jen University, 510 Chung Cheng Road, Hsinchuang, Taipei Hsien 24205, Taiwan, ROC Abstract Mobile advertising complements the Internet and interactive television advertising and makes it possible for advertisers to create tailor- made campaigns targeting users according to where they are, their needs of the moment and the devices they are using (i.e. contextualized mobile advertising). Therefore, it is necessary that a fully personalized mobile advertising infrastructure be made. In this paper, we present such a personalized contextualized mobile advertising infrastructure for the advertisement of commercial/non-commercial activities. We name this infrastructure MALCR, in which the primary ingredient is a recommendation mechanism that is supported by the following concepts: (1) minimize users’ inputs (a typical interaction metaphor for mobile devices) for implicit browsing behaviors to be best utilized; (2) implicit browsing behaviors are then analyzed with a view to understanding the users’ interests in the values of features of advertisements; (3) having understood the users’ interests, Mobile Ads relevant to a designated location are subsequently scored and ranked; (4) Top-N scored advertisements are recommended. The recommendation mechanism is novel in its combination of two-level Neural Network learning, Neural Network sensitivity analysis, and attribute-based filtering. This recommendation mechanism is also justified (by thorough evaluations) to show its ability in furnishing effective personalized contextualized mobile advertising. q 2003 Elsevier Science Ltd. All rights reserved. Keywords: Mobile commerce; Mobile advertising; Neural Network; Sensitivity; Analysis; Information filtering; Recommender systems 1. Introduction With the popularity of mobile devices (such as wireless phones, PDAs, vehicle-mounted devices, etc.), technologies and applications have increasingly been focusing on new sets of tasks, problems, and domains that appreciate the advent of wireless computing and wireless Internet. For instance, mobile commerce refers to any activities related to a (potential) commercial transaction conducted through communication networks that interface with mobile devices. Varshey identified a few fields of applications in mobile commerce (Varshey & Vetter, 2001), such as mobile financial applications, mobile advertising, mobile inventory management, product locating and shipping, mobile enter- tainment, etc. According to Ovum’s report (Interactive Advertising: New Revenue Streams for Fixed & Mobile Operators) (Nelson, 2000), mobile advertising will begin to reach critical mass between 2002 and 2003 and ultimately generate USD16.4 billion by 2005. However, Ovum warned that the quickest way to alienate users is to inundate them with messages. Mobile advertising must be carried out with a basic intention of offering something of value to the consumers. Mobile advertising can complement Internet and interactive television advertising and make it possible for advertisers to create tailor-made campaigns targeting users according to where they are, their needs of the moment and the device they are using (i.e. contextualized mobile advertising). Therefore, it is essential that a fully personalized mobile advertising infrastructure be made. In this paper, we present a personalized contextualized mobile advertising infrastructure for advertising the commercial/non-commercial activities. We name this infra- structure MALCR—an abbreviation for Mobile Advertising by Location-based Customized Recommendation. The concepts behind MALCR are depicted in Fig. 1. The advertisements (obtained from Mobile AD service provi- ders) represent the kind of information available in the environment. The users’ mobile devices serve as points of access to the environment. MALCR then supplies the application that provides a device-independent gateway to information in the environment (Mark, 1999). MALCR’s contributions are three-fold: 1. Furnish a new mobile advertising infrastructure that can unfold both modes of interactive advertising (pull and push) characterized by location-based (Tewari et al., 0957-4174/03/$ - see front matter q 2003 Elsevier Science Ltd. All rights reserved. doi:10.1016/S0957-4174(02)00189-6 Expert Systems with Applications 24 (2003) 399–414 www.elsevier.com/locate/eswa * Corresponding author. Tel./fax: þ886-2-2369-3220. E-mail address: yuans@tpts1.seed.net.tw (S.T. Yuan). http://www.elsevier.com/locate/eswa 2000) and customized recommendations (i.e. contextua- lized recommendations). We believe that such a method of advertising affects customers’ participation of adver- tised activities better than non-location-based/non-cus- tomized advertisements. For instance, an advertisement of sale promotion activity at a particular location would better drive the participation of a customer who is nearby and who is interested in the items for sale than advertisements that do not apply on both scores. 2. Provide a representation space (called vector-based representation space) that is suitable for both the representation of the features in advertised activities and the analysis of users’ interests. 3. Devise a recommendation mechanism that efficiently learns from users’ handset-screen-browsing implicit behaviors and captures users’ preferences in order to provide good recommendations (the evaluations are to be shown in Section 5). This mechanism is mainly a combination of two-level Neural Network learning, Neural Network sensitivity analysis, and attribute-based filtering. This paper is organized as follows: Section 2 provides the architecture of MALCR. Section 3 defines the vector-based representation space that is used in the recommendation mechanism that is presented in Section 4. Section 5 provides the evaluation results. Finally, a discussion and a conclusion are given in Sections 6 and 7, respectively. 2. The architecture of MALCR This section provides MALCR’s architecture (as shown in Fig. 2) and describes how the push and the pull modes of mobile advertising unfold (as shown in Fig. 3). Fig. 2 shows that the primary tasks involved in MALCR are representing Mobile Ads, learning users’ profiles, and providing recommendations by furnishing Mobile Ads that are most similar to a given user’s profile and relevant to users’ locations. In other words, there should be such a common representation space (to be described in Section 3) as to represent both Mobile Ads and users’ profiles before they can be compared. This concept was initially deployed in text filtering (Oard & Marchionini, 1996), but it was first exerted in the area of mobile advertising. By virtue of the limitations of mobile devices, it is better to learn users’ profiles (understanding of user’s needs) mostly from implicit browsing behaviors than to request users’ interests from direct keypad inputs. As a result, an approach for learning users’ profiles (represented in the common representation space) has to be devised and Section 4 will detail this mechanism. Having Mobile Ads and users’ profiles represented in the same space, what follows in the recommendation is Fig. 1. MALCR’s concepts. Fig. 2. The architecture of MALCR. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414400 the provision of a scoring mechanism (that will be detailed in Section 4). Advertising through MALCR proceeds in two ways—the pull mode (the dominating mode) and the push mode (being feasible provided that permission from users is granted). From Fig. 3, working with the function of the positioning gateway residing in wireless carriers (to acquire the location of a given mobile user), the pull mode unfolds simply by the wireless shipping of the requested recommendations obtained from MALCR (that compares with the given user’s profile the Mobile Ads relevant to the location where the user invokes this pull). On the other hand, when a user grants permission, SMS is used to ship the recommen- dations obtained from MALCR (that compares with the given user’s profile the Mobile Ads relevant to a location where the user last makes use of his/her mobile device so as to avoid the expensive location tracking of the user). The in depth version of MALCR’s architecture is shown in Fig. 4 that details the components required to operate the tasks of representing Mobile Ads (the components of Mobile AS Extractor and Mobile AD Database), learning users’ profiles (the components of Personalization Agent, User Stereotype KB, and User Profile database), providing recommendations (the component of Recommendation Function). Mobile AD Extractor transforms Mobile Ads with the vector-based representation and stores the transformed Ads Fig. 3. Pull/push modes of mobile advertising. Fig. 4. The detailed version of MALCR’s architecture. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414 401 in the Mobile Ad Database. Personalization Agent (a combination of two-level Neural Network learning and Neural Network sensitivity analysis) learns users’ profiles with the help of User Stereotype Knowledge Base (that is exerted to expedite the process of Neural Network learning) and stores the learned profiles into the User Profile database. The details of key components will be described in Section 4. 3. Vector-based representation space In this section, we identify the features of advertised activities and present a suitable representation space (called vector-based representation space) that can be used in representing advertised activity Mobile Ads and users’ profiles. One major feature of advertised activities is their categories, exemplified in the first row of Table 1, based on their characteristics. The categories are wholesale and retail, arts and entertainment, etc. For instance, the activity of a sale promotion in a mall falls in the category of wholesale and retail, while the activity of a live performance of a singer belongs in the category of arts and entertainment. The usefulness of advertised activities to users often also depends on the other features of the activities such as day, time, place, fee, performers, etc. Each feature has its prevailing values, such as weekdays and weekends in the day attribute, free and fee-based in the fee attribute, etc. It is often the case that an advertised activity spans multiple values of a feature, such as the activity of a sale promotion in a mall takes place on both weekdays and weekends. From the features of advertised activities described above, we believe a suitable representation for Mobile Ads should have the following traits: (1) a simple representation as Mobile Ads are short in lifespan and are updated frequently; (2) a representation allowing the encoding of the span of multiple values of multiple features; (3) a representation enabling simple comparison. As a result, the vector-based representation is chosen in this paper and is defined in the following two definitions (Definitions 1 and 2). Definition 1. A Mobile Ad is represented as follows: ðI1a1;I1a2;…;I1am1 ;I2a1;I2a2;…;I2am2 ;…Ina1;Ina2;…;Inamn Þ Iiaj [ {0;1}; 1 ,¼ i ,¼ n; 1 ,¼ j ,¼ mi where n is the total number of features characterizing advertised activities; mi; the number of possible values for the ith feature. Iiaj ¼ 1; if a given advertisement embodies the jth value of the designated ith feature 0 otherwise: Definition 2. A user profile is represented as follows: ðWI1a1; WI1a2; …; WI1am1 ; WI2a1; WI2a2; …; WI2am2 ; …; WIna1; WIna2; WInamn Þ 0 ,¼ WIiaj ,¼ 1; 1 ,¼ i ,¼ n; 1 ,¼ j ,¼ mi where n is the total number of features characterizing advertised activities; mi; the number of possible values for the ith feature; WIiaj; a numerical value (ranging from 0 to 1) indicating a user’s interest in the jth value of the designated ith feature. For example, a given advertisement is about a SOGO sale promotion that features as follows (using the feature set of Table 1): Wholesale and Retail, Weekdays and Week- ends, A time slot and B time slot, Indoors and Informal, Fee- Based, no performers. Accordingly, this advertisement is represented as the vector ð1; 0; 0; 1; 1; 1; 1; 0; 0; 1; 0; 1; 0; 0Þ as this advertisement is about a Wholesale and Retail activity and thus the values of I1a1; I1a2; …; I1am1 ðm1 ¼ 3Þ are 1, 0, 0, respectively (similar reasoning for other features). For a user profile example, take as an exemplar the user who is only interested in the activities of indoor promotion sales (such as sales in malls). Chances are the profile of this user looks like (0.52, 0, 0, 0, 0.05, 0, 0.1, 0, 0, 0.33, 0, 0, 0, 0) that shows none-zero preference values at the feature values corresponding to the user’s interest. The computing of magnitude of preference at the feature values will be detailed in Section 4. 4. The recommendation mechanism The concepts underlying the recommendation mechan- ism are four-fold: (1) minimize users’ inputs (a typical interaction metaphor for mobile devices) and thus implicit browsing behaviors are best utilized; (2) implicit browsing behaviors are then analyzed to the understanding of users’ interests to values of features of advertisements; (3) with the understanding of users’ interests, Mobile Ads relevant to Table 1 Features in commercial/non-commercial advertisements Attributes Attribute values Category Wholesale and retail, arts and entertainments, others Day Weekdays, weekend Time A time slot (17:00 pm before), B time slot (17:00 pm after) Place Outdoors, indoors and formal, indoors and informal Fee Free, fee-based Performer Top celebrities, others S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414402 a designated location are subsequently scored and ranked; (4) Top-N (Karypis, 2001) scored advertisements are recommended. After describing the vector space representation, what remains for the recommendation mechanism is the descrip- tion of the task of learning users’ profiles from implicit browsing behaviors and providing recommendations based on the understanding of users’ interests to values of features of advertisements. We detail the task of learning users’ profiles as follows: describe the objectives of User Stereotype KB and the contents in User Profile database (in Section 4.1) and present the functions of Personalization Agent (in Section 4.2). Subsequently, in Section 4.3 we delineate Recommendation Function that computes the recommendations from the learned user profiles and Mobile Ads (both of which are encoded in the vector space representation). Before going into the details of subsequent sections, we first depict the browsing interface (as shown in Fig. 5) that we presume at the mobile devices. This browsing interface reveals the types of implicit browsing behaviors that can be analyzed in the task learning users’ profiles. Fig. 5 shows how the main screen displays the top five recommendations of activities each of which can be further clicked for more details (three levels of details are under consideration for the analysis of users’ interests 1 ). At the bottom of the main screen, a general query based on the advertisement features is also furnished in case the top five recommendations are not favored by the user. Given the display design in Fig. 5, the implicit browsing behaviors, accordingly, can be of the variety of clicking order, clicking depth, and clicking count that are to be taken into account for understanding users’ interests. Clicking order means the order of clicking rendered on an item of recommendation. Clicking depth represents the number of levels of details clicked for an item of recommendation. Clicking count indicates the count of the requests for the complete details of an item of recommendation (i.e., the count of requests of the clicking depth of 3). 4.1. User stereotype KB and user profile database The objective of User Stereotype KB is to expedite the learning of the users’ interests in Personalized Agent (that exerts two-level Neural Network learning). The performance of a well-trained Neural Network has been recognized as being marvelous, but Neural Network learning does suffer from lengthy training and learning. However, a pre-training phase is capable of mitigating the problem of lengthy training (Yuan & Liu, 2000). User Stereotype KB stores a set of pre-trained user stereotype vectors (and the corresponding learned Neural Network weights) representing a variety of typical users’ interests, such as those who love activities of weekend arts and entertainment. Each pre-trained user stereotype vector is obtained by the following steps: (1) pre-train a Neural Network by feeding it with training examples each of which is composed of either a stereotype advertisement (an advertisement attracting the designated user stereotype) vector (please see Appendix A for details) and a score of 1 or an advertisement (that does not attract the designated stereotype user) vector and a score of 0; (2) perform Neural Network sensitivity analysis (to be described in Section 4.2) to the pre-trained Neural Network model and obtain a stereotype vector. In User Profile Database, a user may be associated with multiple user stereotype vectors, each of which is initially brought in from User Stereotype KB when the user is new to MALCR and invokes the general query (shown in Fig. 5) (or the user is not new but invokes a new general query) and thus identifies the applicable user stereotype vectors. Fig. 5. Presumed browsing interface at mobile devices. 1 We believe it is rare for a mobile user browse beyond three levels of details due to the size limitation of the mobile devices’ screens. However, if it is the case, for simplicity we take beyond three levels of details as detailed as three levels of details when analyzing users’ interests. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414 403 The pre-trained user stereotype vectors associated with a user in the User Profile Database will evolve and thus become customized user stereotype vectors corresponding to the knowledge learned from the user’s implicit browsing behaviors. This evolution will be described in Section 4.2. Accordingly, the user profile of a given user is defined as in Definition 3. Definition 3. If a user is associated with the following user stereotype vectors: User Stereotype 1 ðW11; W21; W31; …; Wn1Þ User Stereotype 2 ðW12; W22; W32; …; Wn2Þ .. . User Stereotype m ðW1j; W2j; W3j; …; WnjÞ Then User Profile of this user is defined as follows: User Profile ðW1; W2; W3; …; WnÞ where Wi ¼ Xm j¼1 WijRj m 0 ,¼ Wi ,¼ 1; 1 ,¼ i ,¼ n; 1 ,¼ j ,¼ m; 0 ,¼ Rj ,¼ 1 where n is the total number of feature values in the vector space representation; m, the total number of stereotype vectors associated with the user; Rj is the ratio of the reference of the jth stereotype vector to the total number of reference to all stereotype vectors. For instance, a user that is associated with two stereotype vectors, (0, 0.36, 0, 0, 0.31, 0.03, 0.06, 0, 0.11, 0, 0, 0.05, 0.08, 0) and (0.42, 0, 0, 0.35, 0, 0, 0.13, 0, 0, 0.04, 0.01, 0.05, 0, 0), each of which is referenced (due to the user’s browsing behaviors) 7 times and 3 times, respectively (R1and R2are 7/10 and 7/10), then this user’s User Profile is (0.126, 0.252, 0, 0.105, 0.217, 0.021, 0.081, 0, 0.077, 0.012, 0.003, 0.05, 0.056, 0). 4.2. Personalization Agent Personalization Agent aims to learn by Neural Networks to understand users’ interests to values of features of advertisements from users’ implicit browsing behaviors. Personalization Agent employs two-level Neural Network learning together with Neural Network sensitivity analysis to achieve this end. In this section, we will explain the rationale behind two-level Neural Network learning (Sec- tion 4.2.1), describe the main steps involved in Personaliza- tion Agent, and how Neural Network sensitivity analysis is accomplished (Section 4.2.3). 4.2.1. Rationale behind two-level Neural Network learning Neural Networks are often used with event triggers as inputs and event predictions as outputs and a set of event triggers and corresponding pre-known event predictions as training examples. We call this deployment of Neural Networks, one-level Neural Network learning. In our case of learning to understand a user’s interests, pre-known event predictions have to be explicitly furnished by the user (such as event triggers’ scores, ScoreU) for corresponding event triggers (Mobile AD representation) when one-level Neural Network learning is employed (as shown in Fig. 6(a)), resulting in the learned and recommended event predictions (ScoreR). However, because of the physical limitations of mobile devices (tiny keypads and screens, etc.), frequent requests of explicit inputs from users are not eligible. Therefore, an alternative way of deploying Neural Network learning has to be devised. Two-level Neural Network learning, accord- ingly (as shown in Fig. 6(b)), is such an appropriate deployment. Instead of requesting explicit inputs of ScoreU from users, two-level Neural Network learning exerts the first Fig. 6. one-level and two-level Neural Network learning shown in (a) and (b) respectively. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414404 Neural Network (User_Score Neural Network, abbreviated as USNN) to generate ScoreU from users’ implicit behaviors, such as clicking order (O), clicking depth (D), and clicking count (C), and applies the second Neural Network (Preference_Weight Neural Network, abbreviated as PWNN) to learn users’ interests. USNN is such a pre-trained Neural Network that generates reasonable ScoreU from the values of ðO; D; CÞ: For instance, an item of recommendation with ðO; D; CÞ being (1,3,2) should have a higher ScoreU than another item of recommendation with ðO; D; CÞ being (2, 1, 0). 4.2.2. The algorithm of Personalization Agent After giving the justifications of the use of two-level Neural Network learning, what remains is to present the complete flow of the steps as shown in Fig. 7 (i.e. the algorithm as shown below) of Personalization Agent. From Fig. 7, the procedures of Personalization Agent for understanding a user’s interests on values of features of advertisements are two-fold: 1. On the request of a new stereotype. Personalization Agent brings from User Stereotype KB the pre-trained user stereotype vector and the corresponding Neural Network weights into the User Profile. Compute the customized user stereotype vector by training PWNN (initialized by the Neural Network weights acquired from User Stereotype KB) with the training example that is composed of a Mobile Ad and the user’s ScoreU (predicted by using USNN with the user’s implicit browsing behaviors ðO; D; CÞ to the Mobile Ad), and performing a sensitivity analysis to PWNN. 2. On the use of existing stereotype 2 . Evolve the customized user stereotype vector by training PWNN with the training example that is composed of a Mobile Ad and the user’s ScoreU (predicted by using USNN with the user’s implicit browsing behaviors ðO; D; CÞ to the Mobile Ad), and performing Sensitivity Analysis to PWNN. Personalization_Agent (Stype, M_AD,O,D,C,) Stype is the pre-trained User Stereotype a user requests to add into his/her User Profile when using the query function looking for M_AD Mobile AD. ðO; D; CÞ are the parameters of user feedback when a user browses M_ADs, O represents order, D represents depth, C represents count. 1. Insert Stype into the User Profile if it is a new User Stereotype requested by the user. 2. Input ðO; D; CÞ to USNN (User_Score Neural Network) and get ScoreU as output. 3. Compose (M_AD, ScoreU) as a training example of a User Stereotype which the M_AD belongs to. 4. Use the User Stereotype’s Neural Network weights as the initial weights of PWNN (Preference_Weight Neural Network) if it is the case of a new User Stereotype requested by the user. 5. Train PWNN (that corresponds to the User Stereotype indicated in Step 3) with the training examples obtained from Step 3. 6. Use Sensitivity Analysis (SA_Function) to generate the attribute preference weights from the PWNN, normal- ize the sum to 1, and store them back to the User Profile. SA_Function ( ) 1. For i ¼ 1 to n do Scorei ˆ PWNNpretrainedðX1; X2; X3; …; XnÞ 1 ,¼ i ,¼ n Xj ¼ 1; j ¼ i Xj ¼ 0; j – i for i indicates each input attribute of PWNN, Xi is each input value, and Scorei is the output value of pre- trained PWNN. 2. Compute Scoresum Scoresum ¼ Xn i¼1 Scorei Fig. 7. Flow of steps in Personalization Agent. 2 Without loss of generality, each Mobile Ad corresponds to a User Stereotype, and thus any Mobile Ad browsed by a user contributes to the evolution of a User Stereotype vector (regardless of it is a new stereotype or an existing stereotype). S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414 405 3. Compute Wi Wi ¼ Scorei Scoresum where Wi is the preference weight of Xi in the User Stereotype. The use of multiple User Stereotypes (and thus PWNNs) captured in User Profile prevents the quality of Neural Network learning from downfall due to abrupt drastic change in user’s interests. In other words, the differentiation between Mobile Ads are to be well taken care of by their corresponding PWNNs and what needs to be done is to devise a way to integrate the customized user stereotype vectors obtained from PWNNs (this will be described in Section 4.3). 4.2.3. Sensitivity analysis In order to transform trained PWNNs (the understanding of a user’s interests) into the vector-based representation, it is necessary to obtain the understanding of a user’s interests in values of features in advertisements. Neural Network sensitivity analysis is often employed to this end (Frost & Karri, 1999; Han & Kamber, 2001) The structure of PWNNs we employ (as shown in Fig. 8) includes 14 input nodes ðX1; X2; X3; …; XnÞ (corresponding to the representation of Mobile Ads shown in Table 1), three hidden-layer nodes, one output node (corresponding to ScoreU). For a trained PWNN, the learned data are embedded in the weights of the PWNN. Sensitivity analysis aims to transform the black box of learned weights into a vector showing the user’s interests among values of features in advertisement. The concepts are as simple as follows: (1) among 14 input nodes (i.e. 14 values of features), assign 1 to one feature value and 0 to the remaining feature values and compute the corresponding output (Scorei); (2) repeat Step 1 to all of the input nodes; (3) sum those 14 output scores and obtain Scoresum; (4) compute the percentage of Scorei to Scoresum and obtain the understanding of the user’s relative preference to the designated feature value. 4.3. Recommendation Function With the learned understanding of users’ interests, Recommendation Function aims to provide a scoring mechanism that scores and ranks the Mobile Ads relevant to a designated location and then recommends the Top-N scored advertisements. The algorithms of Recommendation Function are then shown below. Recommendation_Function (P,M_ADs) P is the User Profile of a specific user. M_ADs are the candidate Mobile Advertisements relevant to the specific user location. 1. For each M_AD do ScoreR ˆ Xn i¼1 Xm j¼1 WIi aj Iiaj 1 ,¼ i ,¼ n; 1 ,¼ j ,¼ m Iiaj indicates the jth value of the ith feature in the M_AD WIiaj is the preference weight of the jth value of the ith feature in P 2. Rank the scores of M_ADs 3. Recommend Top-N M_ADs if in the Pull mode 4. Push Top-1 M_AD to the user if in the Push mode The main idea behind Recommendation Function is attribute-based filtering that is described as follows: with the User Profile (described in Definition 3), the score of a given Mobile Ad can be simply the inner product of the vectors of the Mobile Ad and the User Profile. In other words, a Mobile Ad (represented by a vector of 0/1 over the feature values) that is of interest to a user (that is, the value of 1 mostly occurs to the feature values the user prefers) would induce a high score as the User Profile (represented by a vector of preference weight over values of features in advertisements) also embodies high values in those preference weights corresponding to 1’s feature values in the Mobile Ad. Contrarily, a Mobile Ad that does not attract the user often has most of the 0 value go to the feature values that the user prefers (and thus corresponds to high preference weights in User Profile) and thus results in small value in score due to cancelling out of zero production. 5. Evaluation For the nature of recommendation mechanisms in general, evaluations rest on the quality of the recommen- dations (Ben Schafer et al., 1999; Sarwar et al., 2000). Therefore, this section aims to provide the evaluation results on recommendation quality. Fig. 8. PWNN structure. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414406 However, we need to first define the measurements we employ to perform the evaluation (Averaged ScoreU Growth, Instance Precision, Instance Recall, and Instance Fallout as described in Section 5.1), identify experimental user types (Extremely Focused, Extremely Scattered, Middle as described in Section 5.2), and then present the evaluation results for different cases. For example, the case where users will not change their interests and their experimental user types (Section 5.3); the case where users will change their interests but not their experimental user types (Section 5.4) and the case where users will change their interests but not their experimental user types (Section 5.5). In other words, we endeavor to furnish a thorough investigation of how MALCR’s recommendation mechanism responds to a variety of situations as completely as we can, in order to justify the contributions of MALCR. (The simulated recommendation system is then shown in Appendix B.) 5.1. Recommendation quality measurements This section defines and explains the four quality measurements (used in evaluating MALCR’s recommen- dation mechanism), Averaged ScoreU Growth, Instance Precision, Instance Recall, and Instance Fallout: † Averaged ScoreU Growth. MALCR’s objective is to recommend Mobile Ads that are best suited to the user. ScoreU is a score computed from a user’s implicit browsing behaviors ðO; D; CÞ rendered on a recommen- dation that in turn manifests how close this recommen- dation matches the user’s real interests. Therefore, average the ScoreU among the recommendations and attain an averaged ScoreUshowing how close the Top-N recommendations match the user’s interests. The nearer the averaged score is to the value of 1, the closer the Top- N recommendations match the user’s interest. Accord- ingly, the growth of the averaged ScoreUbetween the user’s feedback exhibits the velocity in attaining the accuracy of the recommendation mechanism. † Instance Precision, Instance Recall and Instance Fall- out. The measurements of Precision, Recall and Fallout have been widely used in the area of information filtering (Lanquillon, 1999) and recommendation systems. How- ever, the measurements we employ are slightly different from previous definitions of these measurements, and thus they are called Instance Precision, Instance Recall and Instance Fallout. The word ‘Instance’, we take to mean that we see Precision, Recall, and Fallout from the aspect of the microcosm. That is, Precision, Recall, and Fallout are measured by comparing the elements between a learned vector representation (i.e. a Mobile Ad recommendation) and a target vector representation (i.e. a given user’s preferred Mobile Ad). Table 2 then shows the microcosm view of element comparison between a learned vector representation and a target representation. In Table 2, a matched value of 1 in a pair of corresponding elements between learned representation and target representation is interpreted as Found as it represents the feature value (corresponding to the element pair) preferred by a user and also appears at the learned recommendation. Similarly, a matched value of 0 then is interpreted as Correctly Rejected as the feature value disliked by the user also disappears at the learned recommendation. On the other hand, an unmatched pair of corresponding elements with the values 0 in the learning representation and 1 in the target representation is interpreted as Missed as the feature value preferred by the user disappears at the learned representation. Similarly, an unmatched pair of corresponding elements with the values 1 in the learning representation and 0 in the target represen- tation is interpreted as False Alarm as the feature value disliked by the user nevertheless appears at the learned representation. Based on the interpretations shown in Table 1 for Found, Correctly Rejected, Missed, and False Alarm, it is logical to define the instance-based measurements as follows: Definition 4. Instance Precision Instance Precision ¼ Found/(Found þ False Alarm) Definition 5. Instance Recall Instance Recall ¼ Found/(Found þ Missed) Definition 6. Instance Fallout Instance Fallout ¼ False Alarm/(False Alarm þ Correctly Reject) For a given recommendation, Instance Precision, Instance Recall, and Instance Fallout measure the percen- tage of the accurate hit among recommended feature values, the percentage of the accurate hit among the user’s truly preferred feature values and the percentage of the inaccurate hit among the user’s disliked feature values. Therefore, the higher the values in Instance Precision and Instance Recall, Table 2 Instance-based measurements Target R1 representation Feature value is 1 Feature value is 0 Learned representation Feature value is 1 Found False alarm Feature value is 0 Missed Correctly rejected S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414 407 the better the quality of MALCR’s recommendations, however, the lower value in Instance Fallout, the better the recommendation quality. What follows are examples of the computation of these three measurements when the target representation is (10011111010100) and the learned representation is (10001111001100): † Instance Precision ¼ 6/6 þ 1 ¼ 0.86 (Found ¼ 6, False Alarm ¼ 1) † Instance Recall ¼ 6/6 þ 2 ¼ 0.75 (Found ¼ 6, Missed ¼ 2) † Instance Fallout ¼ 1/1 þ 6 ¼ 0.14 (False Alarm ¼ 1, Correctly Reject ¼ 6) These three measurements complement each other in the following ways: (1) when high Instance Precision (e.g. False Alarm ¼ 0) occurs but feature values still exist that are preferred by a user but not learned by the recommendation mechanism (i.e. Missed . 0), Instance Recall then can complement this void. In other words, high quality in recommendations is justified only by high values in both Instance Precision and Instance Recall. (2) Instance Fallout is able to differentiate the two cases where it is low in Instance Precision but one case is due to small Found (and thus small Fallout) and the other case is owing to large False Alarm (and thus large Fallout). 5.2. Experimental user types Without loss of generality, certain experimental user types (Extremely Focused, Extremely Scattered, Middle) are assumed in order to investigate how robust MALCR’s recommendation mechanism is in the following three sets of experiments discussed in this section. The three sets of experiments are deployed as follows: (1) for each set of experiments, there are three types of users, each of which is comprised of 50 users (each of which exercises MALCR 10 times (represented as Login1, Login2,…,Login10) after his/her MALCR’s first use (rep- resented as Login0); (2) record the measurements of ScoreU, Instance Precision, Instance Recall, and Instance Fallout that are produced from each use of MALCR. The following are then the descriptions of the three experimental user types that differentiate with each other mainly in the frequencies of the use of general query: † Extremely Focused (U1). This user type exemplifies the group of users whose interests are highly concentrated. Without loss of generality, this type of user is simulated as follows: (1) a general query (and thus a pre-trained User Stereotype) is randomly generated to emulate the first use of MALCR (Login0) by such a user; (2) subsequent recommendations from MALCR (i.e. recommendations obtained in Login0, Login2,…, Login10) are assumed to conform to this user’s interests. † Extremely Scattered (U2). This user type represents the group of users whose interests are spread over a wide variety of advertisements. Therefore, we simulate this type of user as follows: three general queries (and thus three pre-trained User Stereotypes) are generated via MALCR (the number 3 is heuristically chosen for manifesting the semantics of ‘scattered’ that is feasible with the use of mobile devices by the user). † Middle (U3). This user type indicates the group of users acting between the previous two extreme types of users. Therefore, this user type is simulated as follows: (1) two general queries are generated in each use of MALCR from Login1 to Login5; (2) subsequent recommendations from MALCR (i.e. recommendations obtained from Login6 to Login10) are assumed to conform to this user’s interests. 5.3. Stable User’s interests and experimental user type This set of experiments aims to investigate how the quality of MALCR varies with numerous uses of MALCR on the condition that a target representation (representing a user’s interests) is randomly assigned prior to Login0 and stays the same throughout Login1 – Login10. The evaluation results of ScoreUand those instance-based measurements are shown in Figs. 9 and 10 and they disclose the following observations: † The exhibition of 1 in ScoreU (shown in Fig. 9) for all of the three user types (U1, U2, U3) manifests that users’ interests are well captured after the general query invoked at Login0 regardless of the different number of user stereotypes subsequently employed by the three different user types. † Fig. 10 exhibits an averaged Instance Precision of 0.95, an averaged Instance Recall of 0.88, and an averaged Instance Fallout of 0.06 at Login10 (similarly for Login1 – Loing9). That is, a combination of the high values in Instance Precision and Instance Recall and the low values in Instance Fallout exactly indicate a high recommendation quality of MALCR. In other words, MALCR exhibits the competence of learning user’s interests effectively in this set of experiments. Fig. 9. Averaged ScoreU Growth for the case of stable user interests and stable user type. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414408 5.4. Unstable user’s interests but stable user type This set of experiments investigates how the quality of MALCR varies with numerous uses of MALCR when there is a change in the interests of a user who, nonetheless, will not change his/her user type. The experiment setting is then deployed as follows: (1) There are two randomly assigned target represen- tations; (2) the first one is generated prior to Login0 to represent the user’s initial interests; (3) the second one is generated at Login3 to signal the change of the user’s interests. Fig. 10. The measurements of Instance Precision (a), Instance Recall (b), and Instance Fallout (c) for the case of stable user interests and stable user type. Fig. 11. ScoreU results for the case of unstable user’s interests and stable user’s user type. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414 409 However, the change of a user’s interests is operated by the user in two ways: (1) explicit change in which the user drives the change by invoking a new general query (and thus bringing in a new pre-trained User Stereotype); (2) implicit change in which the user changes his/her interests with the browsing behaviors and no new pre-trained User Stereotype is brought in. Implicit change in the user’s interests is anticipated to give rise to more complexity in learning which consequently results in a case of explicit change. We would like to explore how robust MALCR can be in the case of implicit change. Another experiment variable employed in the experiment setting is if we weigh the most recent User Stereotype (stored in database of User Profile) more than the others due to the consideration of the time at which User Stereotypes are brought in (i.e. an intuition that the newly identified User Stereotype is supposed to reflect more the user’s contem- porary interests). For simplicity, we investigate this issue with the principle of 80/20 (i.e. the most recently employed User Stereotype is weighed 80% of importance in comparison with 20% of importance rendering on the other User Stereotypes. Combining the factors of explicit change/implicit change and yes/no weighing on the most current User Stereotype, there are four situations shown as below: † Implicit change and no weighing on the most current User Stereotype (L0) † Implicit change and yes weighing on the most current User Stereotype (L1) † Explicit change and no weighing on the most current User Stereotype (L3) † Explicit change and yes weighing on the most current User Stereotype (L4) In this set of experiments, for each user type (U1 – U3), we conduct the experiments for each of the situations (L0 – L4). As a result, there are plenty of figures (for details please Fig. 12. Instance Precision results for the case of unstable user’s interests and stable user’s user type. Fig. 13. Instance Recall results for the case of unstable user’s interests and stable user’s user type. Fig. 14. Instance Fallout results for the case of unstable user’s interests and stable user’s user type. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414410 see Taso and Yuan (2002)). In order to condense the results, we average over the user types and obtain only one set of the averaged results for L0 – L4. The evaluation results of ScoreUand those instance-based measurements for the four situations are then shown in Figs. 11 – 14 that disclose the following observations: † From Fig. 11, ScoreU in the situations of L0 and L1 (i.e. the cases of implicit change), in general, is inferior to that in the situations of L2 and L3 (i.e. the cases of explicit change). However, after three times of learning (Login3 – Login5), ScoreU in the cases of implicit change still performs as well as an accuracy of 0.8. On the other hand, this figure also shows that the factor of yes/no weighing on the most current User Stereotype does not significantly affect the accuracy rate. † From Fig. 11, ScoreU in the situation of L0 reaches as high as an accuracy of 0.97 after two times of learning (Login4 and Login5), but that of L1 oscillates at Login3 – Login5 even though the accuracy rate is still high. The rationale behind this oscillation can be explained as follows: the pre-trained User Stereotype (that is brought in by a user who explicitly invokes the general query) is quite different in nature from Login3’s randomly generated target representation (as described in the above experiment setting). † From Figs. 12 – 14, we know there is no significant difference between yes/no weighing on the most current User Stereotype. In the cases of explicit change, the situation of no weighing is even slightly better than the situation of yes weighing. † Instance Precision and Instance Recall in the cases of explicit change, in general, are superior to those in the cases of implicit change. However, they all exhibit quite a satisfactory level of quality in learning the target representations (about Instance Precision of 75 – 95% and Instance Recall of 50 – 75%). † Instance Fallout is very low for all situations (L0 – L4). † In summary, the quality of MALCR’s recommendations in this set of experiments, in general, is good because of satisfactory Precision Instance and Precision Recall and low Precision Fallout (regardless of the user types U1 – U3, the situations of implicit/explicit change and yes/no weighing on the most current User Stereotype). 5.5. Unstable experimental user type This set of experiments investigates how the quality of MALCR varies with numerous uses of MALCR when there is a change in the user type (from Extremely Focused to Extremely Scattered). The experiment setting is then deployed as follows: (1) the change in the user type occurs at Login4. That is, Login1 – Login3 take on the user type of Extremely Focused and Login4 – Login10 take on the user type of Extremely Scattered; (2) there are two subsets of experiments—one Fig. 15. ScoreU results for the case of unstable user type. Fig. 16. Instance Precision for the case of unstable user type. Fig. 17. Instance Recall for the case of unstable user type. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414 411 for the case at which there is no change in the user’s interests and the other for the case at which there is a change in the user’s interests (that takes place at Login5). The change in the user’s interests is exercised with the approach of explicit change and no weighing on the most current User Stereotype (that performs the best as discussed in Section 5.4) The evaluation results of ScoreUand the instance-based measurements for the four situations are then shown in Figs. 15 – 18 that disclose the following observations: † The case in which there is no change in the user’s interests always performs well (in terms of ScoreU and instance-based measurements) regardless of the user types employed. † At the case in which there is a change in the user’s interests at Login5, ScoreUis back to a level as good as 0.88 after two times of learning since Login5. Instance Precision and Instance Fallout exhibit MALCR’s high accuracy in learning the user’s preferred feature values, but Instance Recall shows the slow progress of learning the whole set of the user’s preferred feature values. However, ScoreUof 0.88 and high Instance Precision already adequately show MALCR’s good performance. 6. Discussion Although this paper presents a Mobile Ads’ recommen- dation mechanism that is justified with good evaluation results in terms of the quality of recommendations, there are still other ways of evaluating the advertising effects such as communicative effects and sales effects (Hsu, 2000). The effects of Internet-based advertising are often measured by such methods as the number of page visits, the number of visitors, etc. (Lee, 2001). For Mobile Ads, appropriate methods of evaluating the advertising effects are worthy of further investigation. For MALCR, we devise such a method represented as a measurement (defined in Definition 7) that will be evaluated in our future work. This method measures the advertising effect at the use of MALCR’s push mode to a user (i.e. a pushed Top-1 recommendation). The concepts behind this method are two-fold: (1) the Top-1 recommendation affects the user if this user responds by exerting MALCR; (2) this response is correlated with the Top-1 recommendation only when this response is made within a certain amount of time; (3) ScoreU rendered on the Top-1 recommendation is used to represent the magnitude of this advertising effect. Definition 7. Advertising effect measurement effect ¼ L £ 1 log T £ ScoreU where L (login) is 1 if a user exerts MALCR after receiving the Top-1 recommendation pushed by MALCR and 0, otherwise; T (time), the lapse of time between the push of the Top-1 recommendation by MALCR and the exertion of MALCR by a user; S (ScoreU) is the ScoreU rendered on the Top-1 recommendation when L ¼ 1: 7. Conclusion In this paper, we present a personalized, contextualized mobile advertising infrastructure (MALCR) for the adver- tisement of commercial/non-commercial activities. The contributions of MALCR are three-fold: (1) furnish a new mobile advertising infrastructure that can unfold both modes of interactive advertising (pull and push) characterized with location-based and customized recommendations (i.e. con- textualized recommendations); (2) provide a representation space (called vector-based representation space) that is suitable for both the representation of the features in advertised activities and the analysis of users’ interests; (3) devise a recommendation mechanism that efficiently learns implicit behaviors from users’ handset-screen-browsing and captures users’ preferences in order to provide good location sensitive recommendations. This mechanism is largely a combination of attribute-based filtering, two-level Neural Network learning, and Neural Network sensitivity analysis. The evaluation results exhibit good recommendation quality embodied in MALCR. This paper also devises a method for measuring MALCR’s advertising effect (that is to be investigated in our future work). Fig. 18. Instance Fallout for the case of unstable user type. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414412 Appendix A. The set of stereotype activities Table A1 Appendix B. MALCR’s simulated user interface Fig. A1. References Ben Schafer, J., Konstan, J., & Riedi, J. (1999). Recommender systems in e- commerce. Proceedings of the First ACM Conference on Electronic Commerce, 158 – 166. Frost, F., & Karri, V. (1999). Determining the influence of input parameters on BP neural network output error using sensitivity analysis. Third International Conference on Computational Intelligence and Multi- media Applications (ICCIMA’99), 45 – 49. Han, J., & Kamber, M. (2001). Data mining concepts and techniques. Morgan Kaufmann Publishers, pp. 303 – 311. Hsu, S. F (2000). A study on the types and properties of internet advertising effects. Master Thesis, Graduate Institute of Business Administration, National Taiwan University, Taiwan. Karypis, G (2000). Evaluation of item-based Top-N recommendation algorithms. Technical report CS-TR-00-46, Computer Science Depart- ment, University of Minnesota. Lanquillon, C. (1999). Information filtering in changing domains. Work- shop on Machine Learning for Information Filtering, International Joint Conference on Artificial Intelligence (IJCAI’99). Table A1 Features Category Day Time Place Fee Performers Example Whole sales and retail Arts and entertain ment Others Week day Week end A Slot B Slot Out doors Indoors and formal Indoors and informal Free Fee -based Top-100 performers Others Stereo type I1a1 I1a2 I1a3 I2a1 I2a2 I3a1 I3a2 I4a1 I4a2 I4a3 I5a1 I5a2 I6a1 I6a2 S01 1 0 0 1 1 1 1 1 0 1 0 1 0 0 Discounted sales S02 1 0 0 1 1 1 1 0 0 1 1 0 0 0 Member-based sales S03 1 0 0 0 1 1 0 1 0 0 1 1 1 0 Stars promoting sales S04 0 1 0 1 0 0 1 0 1 0 0 1 1 0 Weekday arts/ entertainment S05 0 1 0 0 1 1 1 0 1 0 0 1 1 0 Weekend arts/ entertainment S06 0 1 0 1 1 1 0 1 0 0 1 1 0 0 Local culture arts exhibition S07 0 1 0 1 1 1 0 0 1 0 0 1 0 0 Country culture arts exhibition S08 0 1 0 0 1 0 1 0 1 0 0 1 1 0 Concerts S09 0 1 0 0 1 1 1 1 0 0 1 0 1 1 Fee-to-charity performance S10 0 1 0 0 1 0 1 1 1 1 1 0 1 1 Singer contact /movie preview S11 0 0 1 0 1 1 0 0 1 0 0 1 1 0 Formal athletic competition S12 0 0 1 1 1 1 1 1 0 1 1 0 0 0 Athletic activities S13 0 0 1 1 1 1 0 1 0 0 1 1 0 0 Garden arts exhibition S14 0 0 1 0 1 1 0 1 0 0 1 1 0 0 Outskirts activities Fig. A1. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414 413 Lee, R. C (2001). Evaluation of internet-based advertisements, http:// magazines.sina.com.tw/brain/contents/300/300-004_1.html. Mark, W. (1999). Turning pervasive computing into mediated spaces. IBM Systems Journal, Pervasive Computing, 38(4). Nelson, R (2000). Interactive advertising: New revenue streams for fixed and mobile operators. An ovum report. Oard, D.W., & Marchionini, G (1996). A conceptual framework for text filtering. Technical report CS-TR3643, University of Maryland, College Park, MD. Sarwar, B., Karypis, G., Konstan, J., & Riedl, J. (2000). Analysis of recommendation algorithms for e-commerce. Proceedings of the Second ACM Conference on Electronic Commerce, 158 – 167. Taso, E., & Yuan, S.-T (2002). Location-based and customized mechanism for mobile advertising. Technical report, Information Management Department, Fu-Jen University, Taiwan. Tewari, G., Youll, J., & Maes, P. (2000). Personalized location-based brokering using an agent-based intermediary architecture. Proceedings of the 2000 International Conference on Electronic Commerce, Seoul, Korea. Varshey, U., & Vetter, R. (2001). A framework for the emerging mobile commerce applications. Proceedings of the 34th Annual Hawaii International Conference on System Sciences (HICSS-34). Yuan, S.-T., & Liu, A. (2000). Next-generation agent-enabled comparison shopping. International Journal of Expert Systems with Applications, 18(4), 283 – 297. S.-T. Yuan, Y.W. Tsao / Expert Systems with Applications 24 (2003) 399–414414 http://magazines.sina.com.tw/brain/contents/300/300-004_1.html http://magazines.sina.com.tw/brain/contents/300/300-004_1.html A recommendation mechanism for contextualized mobile advertising Introduction The architecture of MALCR Vector-based representation space The recommendation mechanism User stereotype KB and user profile database Personalization Agent Recommendation Function Evaluation Recommendation quality measurements Experimental user types Stable User’s interests and experimental user type Unstable user’s interests but stable user type Unstable experimental user type Discussion Conclusion The set of stereotype activities MALCR’s simulated user interface References 
work_4e2xp7dgsrfatbouk5aoymocqi	----	Response Modeling with Support Vector Regression Dongil Kim, Hyoung-joo Lee and Sungzoon Cho ∗ Department of Industrial Engineering, Seoul National University, San 56-1, Shillim-dong, Kwanak-gu, Seoul, 151-744, Korea Abstract Response modeling, which predicts whether each customer will respond or how much each customer will spend based on the database of customers, becomes a key factor of direct marketing. In previous researches, several classification approaches, include Support Vector Machines (SVM) and Neural Networks (NN), have been applied for response modeling. However, there are two drawbacks of conventional approaches: (1) response models only predict classification scores rather than predicting total amount of money spent, (2) too large training data. For the first drawback, we applied Support Vector Regression (SVR) for response modeling to predict total amount of money spent of each respondent. For the second drawback, we employed a pattern selection method designed for SVR. This paper provides experimental results of a direct marketing dataset in terms of model fit, training time complexity and profitability. Key words: Response Modeling, Direct Marketing, Support vector machines; Regression; Pattern selection ∗ Corresponding author. Tel: +82-2-883-4913, Fax: +82-2-883-4913 Email addresses: dikim01@snu.ac.kr (Dongil Kim), imhjlee@gmail.com Preprint submitted to Elsevier Science 26 May 2006 1 Introduction A response model, given a mailing campaign, predicts whether each customer will respond or how much each customer will spend based on the database of customers’ demographic information and/or purchase history. Marketers will send mails or catalogs to customers who are predicted to respond or to spend large amounts of money. A well-targeted mail increases the profit, while a mistargeted or unwanted mail not only increases the marketing cost but also may worsen the customer’s relationship to the firm (Gönül et al., 2000; Potharst et al., 2000). Various methods have been used for response modeling such as statistical techniques (Bentz and Merunka, 2000; Haughton and Oulabi, 1997; Ling and Li, 1998; Suh et al., 1999), machine learning techniques, (Wang et al., 2005; Chiu, 2002; Cheung et al., 2003; Shin and Cho, 2006; Viaene el al., 2001; Yu and Cho, 2005) and neural networks (NN) (Potharst et al., 2000; Bentz and Merunka, 2000; Zahavi and Levin, 1997). So far, a response model have been usually formulated as a binary classification problem because of its straightforwardness. The customers are divided into two classes, respondents and non-respondents. A classifier is constructed to predict whether a given customer will respond or not. From a modeling point of view, however, as pointed out in (KDD98 Cup, 1998) for the KDD-CUP-98 task, there is an inverse correlation between the likelihood to buy and the dollar amount to spend (Wang et al., 2005). This is because the more dollar amount is involved, the more cautious the customer is in making a purchase decision. In this sense, the likelihood to buy that is estimated by a classification model may not lead to more profit. Therefore, in addition to the classification approach, a regression model needs to be applied for response modeling to (Hyoung-joo Lee), zoon@snu.ac.kr (Sungzoon Cho). 2 predict total amount of money of each customer. Support Vector Machine (SVM) is known as the most spot-lighted algorithm with great generalization performances by employing Structural Risk Mini- mization (SRM) principle (Vapnik, 1995). Support Vector Regression (SVR), a regression version of SVM was developed to estimate regression functions (Druker et al., 1997). Like SVM, SVR is capable of solving non-linear prob- lems using kernel functions and successful in various domains (Druker et al., 1997; Müller et al., 1997). However, there was a difficulty to train SVR on real-world dataset. As the number of training patterns increases, SVR train- ing takes much longer with a time complexity of O(N 3) where N denotes the number of training patterns. So far, many algorithms such as Chunking, SMO, SVMlight and SOR have been proposed to reduce the training time. However, their training time complexity is still strongly related to the number of train- ing patterns (Platt, 1999). We take another direction called pattern selection which is focusing on reducing the number of training patterns. NPPS (Shin and Cho, 2003) is a proved method of pattern selection for SVM, but it’s only for SVM classifiers. Rather we use a pattern selection method based on ε-tube which is especially designed for SVR (Kim and Cho, 2006). In this paper, we applied SVR for response modeling to predict total amount of money spent of each respondent. As mentioned before, a single response model based on classification method can score the likely to respond of each customer. However, direct marketers would like to know not only respondents but also profitable customers who will spend more money than others. Hence, after predicting respondents by a classification response model, a regression model is needed to predict total amount of money spent of each respondent. As the classification model was not our concerned, we supposed there was a primitive ideal classifier which could find all respondents without False Posi- 3 Fig. 1. The ideal procedure of the paper. Especially, the paper is focused on the two dark boxes. tive (FP) errors. We made a subset consists of only but all respondents from original dataset. SVR model was applied to the subset of respondents. The ideal procedure of this paper is presented in Fig 1. We used the DMEF4 dataset from the Direct Marketing Educational Foundation (DMEF) which is collected from a catalog mailing task. The remaining of this paper is organized as follows. In Section 2, we introduce the concepts of SVR and provide the main idea of the pattern selection method with a simple toy example. In Section 3, we present details of DMEF datasets and parameters for experiment. In Section 4, experiment results are following. In section 5, we summarize the result and conclude the paper with a remark on limitations and future research directions. 2 Pattern Selection for Support Vector Regression 2.1 Support Vector Regression For a brief review of SVR, consider a regression function f(x) to be estimated with training patterns {(xi, yi)}, f (x) = w · x + b with w, x ∈ RN , b ∈ R (1) 4 where {(x1, y1), · · · , (xn, yn)} ∈ RN × R (2) SVR is moved around to include training patterns inside ε-insensitive tube (ε-tube). By the SRM principle, the generalization accuracy is optimized by the flatness of the regression function. Since the flatness is guaranteed on a small w, SVR is moved to minimize the norm, ‖w‖2. An optimization problem could be formulated with constraints where C, ε, and ξ, ξ∗ are trade-off cost between empirical error and the flatness, size of ε-tube and slack variables, respectively, for the following soft margin problem. Minimize 1 2 ‖w‖2 + C n∑ i=1 (ξi + ξ ∗ i ), (3) Subject to yi − w · xi − b ≤ ε + ξi, w · xi + b − yi ≤ ε + ξ∗i , ξi, ξ ∗ i ≥ 0, i = 1, · · · , n. Hence, SVR is trained by minimizing ‖w‖2 with including training patterns inside the ε-tube. With adding Lagrangian multipliers α and α∗, the QP prob- lem can be optimized as dual problem. The estimated regression function from SVR is following where ns is the number of support vectors: f (x) = ns∑ i=1 (α − α∗)K(xi, x) + b (4) 2.2 Pattern Selection for Support Vector Regression The training time complexity of SVR is O(N 3). If the number of training patterns increases, the training time increases more radically, i.e. in a cubic proportion. Marketing databases usually consists of over one millions of cus- tomers and hundreds of input variables. Hence, it takes too long time to train SVR directly to marketing dataset. We applied a pattern selection method from our previous research (Kim and Cho, 2006). 5 Fig. 2. (a) The regression function after training original dataset, and (b) the re- gression function after training ONLY patterns inside estimated ε–tube. SVR trains patterns based on ε-loss function foundation. SVR makes ε-tube on the training patterns. The patterns in ε-tube are not counted as error, and patterns out of ε-tube, i.e. Support Vectors (SVs), are used for training. In addition, SVR estimates the regression function as the center-line of ε-tube. Hence, if ε-tube can be estimated before training, we can find the regression function with only those patterns inside ε-tube (See Fig. 2) . However, remov- ing all patterns outside ε-tube could lead to reduction of ε-tube itself, thus it is desirable to keep some of “outside” patterns for training. Hence, we defined a “fitness” probability for each pattern based on its location with respect to ε-tube and then selected patterns stochastically. We made k bootstrap samples of size l (l < n) from original training pattern set (D). We trained an SVR with each bootstrap sample and obtained k SVR regression functions. Each regression function was used to see if a training pattern is located inside ε-tube. Each training pattern in D is located inside a minimum of zero ε-tubes to a maximum of k ε-tubes. Let mj denote the number of times that pattern j is found in-side an ε-tube. We use mj as the likelihood that pattern j is actually located inside the real ε-tube. Each mj is converted to a probability, pj as in Eq. (5). Since we want to select patterns 6 Fig. 3. (a) Original dataset and an SVR trained on it, (b) A bootstrap sample and an SVR trained on it, (c) Original dataset and ε-tube of (b)’s SVR, and (d) Selected patterns and an SVR trained on them inside ε-tube, pattern j is selected with a probability of pj, pj = mj∑n i=1 mj . (5) The procedure of the pattern selection method is presented in Fig. 3 with a simple toy example. The algorithm is presented in Fig. 4. 7 1. Initialize the number of bootstrap samples, k Initialize the number of patterns in each bootstrap sample, l Initialize the number of patterns to be selected, s 2. Make k bootstrap samples, Di, (i = 1, · · · , k), from the origi- nal dataset D by random sampling without replacement 3. Train SVR fi with Di, (i = 1, · · · , k) 4. Count the number of times mj that pattern j is found inside ε-tube of fi 5. Convert mj to pj according to Eq. (5) 6. Select s patterns stochastically from D without replacement based on pj 7. Train final SVR with s selected patterns Fig. 4. ε-tube based pattern selection algorithm 3 Dataset and Experimental Settings 3.1 Dataset: DMEF dataset We utilized DMEF4 dataset which contains 101,532 customers and 91 input variables. The response rate is 9.4% with 9,571 respondents and 91,961 non- respondents. We selected a subset based on the weighted dollars used from previous researches (Yu and Cho, 2005; Ha et al., 2005). We would like to design a regression model as a scoring model after picking up respondents with a primary classification model. Since a classification model was not our 8 Table 1 Input variables Name Formulation Description ORIGINAL VARIABLES Purseas Number of seasons with a purchase Falord LTD fall orders Ordtyr Number of orders this year Puryear Number of years with a purchase Sprord LTD spring orders Derived Variables DERIVED VARIABLES Recency Order days since 10/1992 Tran53 I(180 ≤ recency ≤ 270) Tran54 I(270 ≤ recency ≤ 366) Tran55 I(366 ≤ recency ≤ 730) Tran38 1/recency Comb2 ∑14 m=1 ProdGrpm Number of product groups purchased from this year Tran46 √ comb2 Tran42 log(1 + ordtyr × falord) Interaction between the number of orders Tran44 √ ordhist × sprord Interaction between LTD orders and LTD spring orders Tran25 1/(1+lorditm) Inverse of latest-season items interest in this research, we assumed that there was an ideal response model built with a classification algorithm that could pick all respondent without false acceptances. Hence, we selected a new dataset consists of customers only whose target dollars are positive, i.e. only respondents. The final selected dataset consists of 4,000 customers. For performance evaluation, the dataset was partitioned into training and 9 test sets. A half of customers were randomly assigned to the training set while the other half to the test set. Performance of a model shows a large varia- tion with regard to a specific data split (Malthouse, 2002). So ten different training/test splits were generated. All results are averaged of each exclusive training/tes sets. We are not interested in feature selection/extraction. Malt- house extracted 17 input variables for this dataset and Ha et al (Ha et al., 2005) used 15 out of them, removing two variables whose variations are neg- ligible. In this paper, these 15 variables were used as input variables as listed in Table 1. We formulated this dataset to a regression problem by setting the total amount of dollars as target variable. 3.2 Experimental Settings We made three different response models based on SVR in the experiments. All experimental results of SVR with pattern selection (SVR-PS) were compared with results of SVR with all data (SVR-100) and SVR with random sampling (SVR-Random). We set hyper-parameters of SVR. The hyper-parameters of SVR were deter- mined by cross–validation for SVR-100 with C × ε = {0.1, 1, 5, 10, 50, 100}× {0.01, 0.05, 0.07, 0.1, 0.5, 0.7, 1}. RBF kernel was used as a kernel function and the kernel parameter σ was fixed to 1.0 for all datasets. The parameters of the pattern selection method were set as follows. In this case, the number of boot- strap samples k was set to 10. The other parameter that controls the number of patterns in a bootstrap sample l was set to 25% of the number of patterns in dataset, n. The number of selected patterns was set to 10% to 90% of n. During the experiments, we used all data as normalized vectors. Root Mean Squared Error (RMSE) is used to estimate the model fits. We calculated the 10 averaged profits per one catalog mail to estimate the profitability of response models. Also we recorded training times. All stochastic results were the aver- aged values of 10-time repeats. 4 Experimental Results 4.1 Model Fit and Time Complexity Fig. 5 shows the experimental results. The solid lines present the result of SVR with all data (SVR-100). The lines with circles presents the results of SVR with pattern selection (SVR-PS) while the lines with boxes presents the results of SVR with random sampling (SVR-Random). Fig. 5 (a) is RMSEs which have real values (dollars). It shows SVR-PS is slightly less accurate than SVR-100, but is significantly more accurate than SVR-Random. As more patterns were selected, the training dataset became similar to the original dataset and the gaps between all SVRs became narrower. With 10% pattern selection, the gap of RMSEs between SVR-PS and SVR-Random is larger than 15 dollars. When the percentage of selected patterns is 40% of all, the gaps of RMSEs between SVR-PS and SVR-100 are less than 5 dollars. Fig. 5 (b) shows averaged training time of all SVRs. It takes 1200 seconds to train all data with SVR. However, with 10% of pattern selection, SVR-PS needed only 16% of training time of SVR-100. Although the percentage of selected patterns was up to 40%, which resulted only 5 dollars’ gap, SVR-PS needed only 25% of training time of SVR-100. SVR-PS is more efficient than SVR-100. 11 10 20 30 40 50 60 70 80 90 55 60 65 70 75 80 85 Percentage of patterns selected (%) R M S E SVR−100 SVR−PS SVR−Random (a) 10 20 30 40 50 60 70 80 90 0 200 400 600 800 1000 1200 1400 Percentage of patterns selected (%) T im e (s ec ) SVR−100 SVR−PS SVR−Random (b) Fig. 5. (a) RMSE resulted from experiments and (b) the training times 12 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 50 60 70 80 90 100 110 120 130 140 150 Mailing depth (%) A ve ra ge p ro fi t p er m ai l ($ ) SVR−100 SVR−PS SVR−Random 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 50 60 70 80 90 100 110 120 130 140 150 Mailing depth (%) A ve ra ge p ro fi t p er m ai l ($ ) SVR−100 SVR−PS SVR−Random (a) (b) 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 50 60 70 80 90 100 110 120 130 140 150 Mailing depth (%) A ve ra ge p ro fi t p er m ai l ($ ) SVR−100 SVR−PS SVR−Random 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 50 60 70 80 90 100 110 120 130 140 150 Mailing depth (%) A ve ra ge p ro fi t p er m ai l ($ ) SVR−100 SVR−PS SVR−Random (c) (d) Fig. 6. Average profit per one mail. (a) Results of 10% pattern selection and random sampling, from (b) to (d) are 30%, 50%, 70% respectively 4.2 Profitability Fig. 6 compares three response models of SVR in terms of profit. For each setting, the response model predicted the dollar amount to each test customer. Then, we sorted the test customers by descending order in terms of the dollar amount predicted by model. A mailing depth of Fig. 6 is an expected profit made in a promotion to each percentage of top-decile. As it shows, if we send catalog mails to all respondents, the averaged profit of each mail is roughly 13 60 dollars. However, if we want to select 10% of all respondents, we can make more profit based on SVR model than only use classification model. As mailing depths go larger, averaged profits per mail decrease. SVR-PS resulted between SVR-100 and SVR-Random. As we saw the training time complexity analysis in Fig. 5 (b), SVR-PS guaranteed efficient models with acceptable profitability and fast training speed. 5 Conclusions and Discussion We applied SVR for response modeling. We assumed that there was a pre- vious response model that could find all respondents, perfectly. With SVR, we estimated total amount of dollar spent for each customer to find more profitable respondents. As results, we could find high profit customers rather than just response rate. Also, to reduce the training time complexity, we used the pattern selection method. The pattern selection method made SVR be an efficient model and, at the same time, avoid worsening the accuracy. SVR-100 resulted the best and SVR-Random resulted the worst in terms of accuracy and profit, while SVR-Random resulted the best in terms of training speed. SVR-PS located a reasonable area in terms of efficiency which guarantees ac- curate like SVR-100 and fast like SVR-Random. SVR-PS needs only 16∼25% of training time of SVR-100, with acceptable accuracy loss. There are some limitations of this research. First, it was an early research of applying SVR for response modeling. We formulated the problem as a regres- sion form, however, we might have missed some key factors of data setting for regression formulation. With further researches, we can find more effec- tive regression formulation to be applied response modeling. Second, the re- sults of SVR-PS looked efficient, however, we couldn’t decide those were good 14 enough to apply real-world problems. The pattern selection method should be improved to be more accurate. Finally, various experiments including profit analysis should be followed in further study. We assumed there was a per- fect classification model, however, there will be a comparison with the real classification model in further study. References Bentz, Y., Merunka D., 2000. Neural networks and the multinomial logit for brand choice modeling: a hybrid approach, Journal of Forecasting 19, 177- 200. Cheung, K.-W., Kwok, J.T., Law, M.H., Tsui, K.-C., 2003.mining customer product ratings for personalized marketing, Decision Support Systems 35, 231-243. Chiu, C., 2002.a case-based customer classification approach for direct mar- keting, Expert Systems with Applications 22(2), 163-168. Drucker, H., Burges, C. J. C., Kaufman, L., Smola, A., Vapnik, V., 1997.Sup- port Vector Regression Machines, In: Mozer, M. C., Jordan, M. I., Petsche, T.(eds.): Advances in Neural In-formation Processing System 9, MIT Press, Cambridge, MA, 155-161. Gönül, F.F., Kim, B.D., Shi, M., 2000. Mailing smarter to catalog customer, Journal of Interactive Marketing 14(2), 2-16. Ha, K., Cho, S., MacLachlan, D., 2005.Response models based on bagging neural networks, Journal of Interactive Marketing 19(1), 17-30. Haughton, D., Oulabi, S., 1997.Direct marketing modeling with CART and CHAID, Journal of Direct Marketing 11(4), 42-52. KDD98, The KDD-CUP-98 result, 1998. http://www.kdnuggets.com/ meetings/kdd98/kdd-cup-98.html. 15 Kim, D., Cho, S., 2006.ε-tube based Pattern Selection for Support Vector Machines Lecture Notes in Artificial Intelligence 3918, 215-224. Ling, C.X., Li, C., 1998.Data mining for direct marketing: problems and so- lutions, Proceedings of ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining (KDD-98), New York, pp. 73-79. Malthouse, E.C., 2002.Performance-based variable selection for scoring mod- els. Journal of Interactive Marketing 16(4), 37-50. Müller K.-R., Smola A., Rätsch G., Schölkopf B., Kohlmorgen J., Vapnik V., 1997.Predicting time series with support vector machines. In: GerstnerW., Germond A., Hasler M., and Nicoud J.-D. (Eds.), Artificial Neural Networks ICANN97, Berlin. Lecture Notes in Computer Science 1327, 999-1004. Platt, J. C., 1999.Fast Training of Support Vector Machines Using Sequen- tial Minimal Optimization, Advanced in Kernel Methods; Support Vector Machines, MIT Press, Cambridge, MA, pp. 185-208. Potharst, R., Kaymak, U., Pijls W., 2000. Neural networks for target selection in direct marketing, Erasmus Research Institute of Management (ERIM), RSM Erasmus University, Research Paper ERS-2001-14-LIS, Available at http://ideas.repec.org/s/dgr/eureri.html. Shin, H., Cho, S., 2003.Fast Pattern Selection Algorithm for Support Vector Classifiers: Time Complexity Analysis, Lecture Notes in Computer Science 2690, 1008-1015. Shin, H., Cho, S., 2006.Response modeling with support vector machines, Expert Systems with Applications 30(4), 746-760. Suh, E.H., Noh, K.C., Suh, C.K., 1999.Customer listsegmentation using the combined response model, Expert Systems with Applications 17(2), 89-97. Vapnik, V., 1995. The Natural of Statistical Learning Theory, Springer, New York. Viaene, S., Baesens, B., Gestel, T., Suykens, J.A.K., 2001.Van den Poel, D., 16 Vanthienen, J., De Moor, B., Dedene, G., Knowledge discovery in a direct marketing case using least squares support vector machines, International Journal of Intelligent Systems 16, 1023-1036. Wang, K., Zhou, S., Yang, Q., Yeung, J.M.S., 2005. Mining customer value: from association rules to direct marketing, Data Mining and Knowledge Discovery 11, 57-79. Yu, E., Cho, S., 2005.Constructing response model using ensemble based on feature subset selection, Expert Systems with Applications, In Press, Available online at http://www.sciencedirect.com/science/journal/ 09574174. Zahavi, J., Levin, N, 1997.Applying neural computing to target marketing. Journal of Direct Marketing 11(4), 76-93. 17 
work_4fmqd4q7pfb7nclr5zelrvwsja	----	Material Recognition using Tactile Sensing Emmett Kerr, TM McGinnity and Sonya Colemana,b,a aIntelligent Systems Research Centre, University of Ulster, Magee Campus, Northern Ireland. e-mail: ep.kerr@ulster.ac.uk, sa.coleman@ulster.ac.uk bCollege of Science and Technology, Nottingham Trent University, UK. e-mail: martin.mcginnity@ntu.ac.uk Abstract Identification of the material from which an object is made is of significant value for effective robotic grasping and manipulation. Characteristics of the material can be retrieved using different sensory modalities: vision based, tactile based or sound based. Compressibility, surface texture and thermal properties can each be retrieved from physical contact with an object using tactile sensors. This paper presents a method for collecting data using a biomimetic fingertip in contact with various materials and then using these data to classify the materials both individually and into groups of their type. Following acquisition of data, principal component analysis (PCA) is used to extract features. These features are used to train seven different classifiers and hybrid structures of these classifiers for comparison. For all materials, the artificial systems were evaluated against each other, compared with human performance and were all found to outperform human participants’ average performance. These results highlighted the sensitive nature of the BioTAC sensors and pave the way for research that requires a sensitive and accurate approach such as vital signs monitoring using robotic systems. Keywords: Tactile Sensing, Signal Processing, Neural Networks, Artificial Intelligence, Classification, Material classification 1. Introduction Humans can quickly gain information about properties of an object or material simply by viewing them from different angles; for example they can estimate how it might feel to touch or how heavy it may be. This is due to our highly sophisticated visual capabilities and ability to adapt prior knowledge Preprint submitted to Expert Systems with Applications October 19, 2017 learned from similarly shaped and known objects. However, there are some properties which are difficult to detect by vision alone, for example thermal conductivity, compressibility or a determination from what material an ob- ject is made. In order to learn these characteristics and remember them for future interactions, physical manipulation of an object is required. Distin- guishing between objects and materials of different compressibility, tempera- ture and texture can be achieved by performing complex manipulation tasks such as squeezing or rubbing. These complex manipulation tasks can inher- ently be performed extremely well by humans due to our very sophisticated tactile perception. In order to produce the tactual perceptions required to learn about ob- ject properties, humans perform various types of movements when interacting with an object. Experimental psychologists have identified six general types of exploratory movements: pressure to determine hardness, static contact to determine thermal properties, lateral sliding movements to determine surface texture, enclosure to determine global shape and volume, lifting to deter- mine weight, and contour following to determine exact shape (Lederman and Klatzky, 1987). Humans can complete these exploratory movements very fast and rapidly evaluate the object leading to a possible identification. One of the main applications of an artificial tactile sensing system is expected to be in robotic applications. In order to retrieve the necessary tactual perceptions, a robot must explore an object in a similar manner to that of a human. However, in comparison to humans, these exploratory movements can be slow for a robot and typically involve the evaluation of a great volume of acquired data, thus being computationally expensive. It would be a useful skill for a robot if it could complete a preliminary evaluation of the object by quickly identifying the physical nature of the material (e.g. metal, wood, plastic etc.) through completing a small number of initial basic actions, thus removing the need for extensive manipulation. These initial basic actions could include the retrieval of thermal information upon initial contact with a material which may provide an indication of its physical nature, e.g. wood, metal or plastic. Furthermore the texture of a material can be identified by sliding a robotic hand or finger along the material, i.e. determining if it is rough or smooth. We use an artificial fingertip to acquire data relating to the thermal prop- erties and surface texture of materials, that are first analysed to initially identify to which group the material belongs (e.g. wood, metal, plastic etc.) and subsequently to determine the specific individual material (e.g. alu- 2 minium, copper, pine etc.). Building on the work presented in (Kerr et al., 2013, 2014a) and (Kerr et al., 2014b) the contribution of this work is the im- plementation of six further classifiers, hybrid configurations of the classifiers and a comparison of the performances of these classifiers. The classifiers implemented and compared are a one stage ANN (Kerr et al., 2014a), a two stage ANN (Kerr et al., 2014b), a one stage SVM, a two stage SVM, a hybrid ANN and SVM approach, GMM, LDA, NB, k -NN and MLP. Utili- sation of these techniques to classify materials from tactile data resulted in an increase in classification accuracy and system efficiency in comparison to previous work (Kerr et al., 2014b) resulting in an approach capable of a fast initial identification of the material group and/ or individual material. These classifiers are also evaluated against human performance when identifying the same test materials as reported in (Kerr et al., 2014b). Therefore, in this paper we propose an approach for material classification using a combination of two of the three key characteristics, namely surface texture (represented through vibration data) and thermal characteristics. Although one weakness of the approach is that it does not achieve accuracies as high as that of (Xu et al., 2013), the strengths include the fact that it is computationally efficient and it demonstrates high classification accuracies across 14 materials, some of which are very similar, unlike many methods presented in the literature where the test materials can be very different. A range of common classifiers together with novel combinations of classifiers are evaluated and proven to be accurate and efficient algorithms for material classification. Furthermore, the system is tested with previously unlearned materials in order to assess its robustness. The remainder of this paper is organised as follows: Section 2 presents an overview of related research in tactile sensing-based material classifica- tion. Section 3 describes the experimental set-up, including an overview of the equipment used and the learning algorithms used for the robot system. Section 4 presents an evaluation of the performance of the various artifi- cial systems developed and compares their performance to that of human subjects. Section 5 discusses the artificial system and human participants’ performance. Conclusions and plans for future work are given in Section 6. 2. Background and Related Research Three of the key characteristics that are critical in performing an effi- cient grasp on an object are texture, compressibility and thermal properties. 3 Such characteristics can be obtained from tactile sensing. Materials from dif- ferent classes may have similar compressibility; for example neither acrylic or medium density fibreboard (MDF) are very compressible but are from different classes of materials, i.e. plastic and wood. Drimus et al. (2011), de- veloped a sensor and presented a technique used to classify between ten rigid and deformable household objects. The objects were squeezed in a gripper fitted with force sensors and a k -NN classification algorithm was utilised to classify the different objects, based on their compressibility. Classification accuracy of up to 92% was achieved for classification between the ten objects ranging from a rubber ball to a plastic bottle. The algorithm struggled to distinguish between similarly compressible objects even though the objects were very different, for example a “bad orange” and an empty bottle were found to have similar compressibility and therefore the algorithm failed to classify these objects correctly. Another approach which uses images de- rived from collected tactile sensor data is that of Pezzementi et al. (2011), which utilises computer vision algorithms to allow tactile sensor readings, representing compressibility, to be viewed as images. The tactile array sen- sor explores the object images representing contact made on the array are produced. Six feature descriptors including the well known Scale Invariant Feature Transform (SIFT) (Lowe, 1999) and MR-8 (Varma and Zisserman, 2005) are used to extract features from these images. The Bag-of-Words (BoW) approach (Jurie and Triggs, 2005), commonly used in computer vi- sion methods, is then used to learn and classify the objects based on these features. Luo et al. (2015), also present an approach which utilises tactile images and a BoW approach to object recognition. An optimised BoW ap- proach and dictionary were produced by considering different outputs of the classic BoW with different dictionary sizes for different views of objects. The classification performance of this optimised approach was compared against the performance of the classic BoW model utilising 12 objects. It was found that the proposed approach outperformed the classic BoW model particularly for classification of objects with similar features. Exploiting other characteristics such as the object’s surface texture or its thermal properties may help to distinguish between very different ma- terials which have similar compressibility, leading to improved classification accuracy. Related literature reports methods that are capable of classifying materials using surface textile and achieving high performance rates, i.e. is the object rough or smooth? Fishel and Loeb (2012) used the BioTAC sensors with a Bayesian exploration method for the intelligent identification of tex- 4 tures. Classification from a set of 117 textures was achieved. This approach related the textures to human style descriptions frequently used in psy- chophysical literature exploring texture discrimination (i.e. sticky/slippery, rough/smooth). The focus was to find the most useful exploratory motions on the surface of the material, relating to six general types of exploratory movements that humans make when tactually exploring objects, as outlined by experimental psychologists Lederman and Klatzky (1987). Ho et al. (2012) assessed the ability of their sensors to classify between three different ma- terials (denim, a photo and tape) based on their surface texture. Classifi- cation accuracy of 91% was achieved between two very different materials, i.e. denim and a photo. However, the method struggled to distinguish ac- curately between two similar materials, namely the photo and tape. Jamali and Sammut (2010) propose a method which used an artificial silicon tactile sensor with Polyvinylidene Fluoride (PVDF) films to slide along materials and collect vibration data. A total of seven materials were tested includ- ing sponge, carpet, wood, two tiles of different roughness and two pieces of vinyl of different roughness. A Bayes classifier was trained with the Fourier coefficients of the sensor output data retrieved from the sliding motion over each material. It was shown that the authors’ classifier performed well for classification between dissimilar surface textures, however, their approach was unable to distinguish between two types of tiles, as the texture of the surfaces were similar. Jamali and Sammut (2011) extended this work to eval- uate multiple classifiers and added a majority voting algorithm. In a similar way to their previous work, data were retrieved from the fingertip making contact with test materials. The test materials included the same seven as in their previous work with the addition of a second type of carpet (differ- ent roughness), making eight material surfaces in total. Whilst the fingertip was being rubbed over each surface, different textures induced different in- tensities of vibration in the silicon fingertip. As in their previous work, the data from the fingertip were pre-processed and the Fourier coefficients of the sensor outputs were used as inputs to various classifiers that were imple- mented in the WEKA machine learning toolkit (Hall et al., 2009). It was found that the best performance of 95%±4% was achieved using the Naive Bayes tree (NBTree). The system struggled most when trying to classify between materials of similar roughness. The approach proposed by Chathu- ranga et al. (2013) is another example of a system which produced promising results but still struggled to classify between similar material surfaces. The authors developed a biomimetic fingertip and used it to discriminate between 5 seven different materials based on surface texture. Using ANNs as classifiers, accuracies of up to 85% were achieved. The thermal characteristics of a material is another modality that can be assessed and used to identify a material/object. Although one can generally tell the difference between materials that feel hot and cold, one would struggle to discriminate between similar materials on the basis of their actual temper- atures. What human fingers detect thermally is not the absolute temperature of the material alone, but the effect of thermal conductivity and diffusivity. This is due to the fact that the human finger is at a consistent temperature, approximately 34◦ Celsius, and therefore will be different from the ambient temperature of most objects encountered. The nerves of the finger detect the flow of heat to/from the source. This suggests that human temperature sens- ing may be emulated by thermal conduction sensing (Monkman and Taylor, 1993). An approach which analyses the thermal properties of various ma- terials is described in (Kerr et al., 2013). The materials are classified into initial groupings and further identified individually via the data gathered from test materials using the Syntouch R© BioTACTM sensor. PCA was ap- plied to identify relevant features from the static temperature and thermal conductivity data that were used to train an ANN. The system was compared with humans and was found to outperform human performance. A success rate of 73% was obtained by the system when classifying the material group (i.e. plastic, wood, metal etc.) and 61% was achieved when identifying the individual material. When similar experiments were conducted using hu- mans, results showed classification rates of 61% for group identification and 51% for individual material identification. However, even though the system achieves better performance than the humans, the classification rates are poor which is potentially caused by the fact that only thermal conductivity data are being used. Additionally, the system is computationally expensive due to the large number of input features necessary to accurately represent the data and the substantial number of neurons in the hidden layer. Bhat- tacharjee et al. (2015) present another approach that focusses on material classification. Their approach uses a tactile sensor that is capable of measur- ing solely thermal properties and they focus on classification within a short period of time, for instances where longer term contact is not viable. Three classification algorithms are considered, SVM+PCA, k -NN+PCA and Hid- den Markov Models (HMM). It was found that SVM+PCA performed best with classification rates of 84% and up to 98% with longer contact times across 11 materials with varying initial conditions and 3-fold cross valida- 6 tion. Although the classification rates achieved are impressive, the materials tested varied widely in thermal conductivity and the purpose of the sensor is solely to determine thermal properties; it therefore cannot measure material properties such as texture or compressibility, thus limiting the applications for which it can be used. Rather than classifying materials by analysing just one property, a com- bination of compressibility, texture and/ or thermal properties can be used to achieve higher rates of classification accuracy. Kerr et al. (2014a), used a combination of two of these properties, extending their work in (Kerr et al., 2013) by considering surface texture (vibration) as an additional modality alongside the thermal properties of a material. Additionally, the authors re- designed the system to improve efficiency by selecting a reduced number of principal components, reducing the number of neurons in the hidden layer and hence training the system faster than previously. The one stage ANN system in (Kerr et al., 2014a) obtained a classification rate of 83.81% for ma- terial group classification and 79.05% for individual material identification. Conversely, the humans achieved classification rates of 73.00% and 60.90% respectively. A two stage ANN approach to the same material classification experiment is presented in (Kerr et al., 2014b). The output from the network in the first stage is used as an input to the second stage, which is a set of networks, one for each material group. This second stage is used to identify each individual material, within a specified group. Despite not being as effec- tive as the one stage ANN method presented in (Kerr et al., 2014a), this two stage approach obtained a classification rate of 83.81% for material group identification and 70.48% for individual material identification. Bhattachar- jee et al. (2016) continued their earlier research and presented a method for distinguishing between contact with inanimate objects and humans, using a novel portable handheld device developed by the authors consisting of three tactile sensing modalities: a force sensor to detect contact, a heat-transfer sensor that is actively heated, and a small thermally-isolated temperature sensor (Wade et al., 2015). Data was collected from the arms of 10 human subjects from 3 different locations on the right arm and 80 objects consisting of 8 similar objects from 10 different bathrooms (Bhattacharjee et al., 2016). The effect of varying durations of contact made with the human subjects and the objects was also evaluated. A SVM was used to learn to classify between an object and human in the first instance and it achieved an aver- age classification accuracy of 98.75% when contact was held for 3.65 seconds, 93.13% for 1.0 second of contact and 82.13% for 0.5 seconds of contact. High 7 classification accuracies were also achieved when generalising to new contact locations within the same bathroom with an average of 92.14% for 3.65 sec- onds of contact, 91.43% for 1.0 second and 84.29% for 0.5 seconds. However, when generalising to new environments, i.e. similar objects in different bath- rooms, accuracies were lower achieving 84.00% for 3.65 seconds of contact, 71.00% for 1.0 second and 65.00% for 0.5 seconds. Like many other meth- ods it struggled to identify similar objects especially when the environment changes. Hoelscher et al. (2015) use a BioTAC fingertip sensor to collect several features from a range of 49 objects to develop a system capable of identifying material and recognising objects. The objects are made of mate- rials including plastic, metal, stone ceramic, paper etc. They collect data by making static and lateral contact on each material with the BioTAC sensor. The authors then compare seven different methods of extracting features from their processed data, including temporal data, PCA reduced dimensionality raw data, pressure features, electrode features, physically motivated features, temperature features and mean features. The authors compare two genera- tive (NB and Gaussian) and two discriminative (SVM and Random Forests) classifiers to evaluate identification of the materials and objects from the collected data. The best performing method consisted of an SVM classifier with dimensionality-reduced mean values of filtered data, achieving a classi- fication accuracy of material and object identification of over 97%. Xu et al. (2013) presented an algorithm which considered three key properties of the material to enable classification; compliance (compressibility), texture and thermal conductivity. To collect data for these three key properties a Bio- TAC fingertip was mounted on a Shadow Robot Hand. Based on a Bayesian exploration approach in the authors previous work (Fishel and Loeb, 2012) three exploratory movements were selected and used to explore ten test ma- terials. A classification accuracy of 99% using 10 materials was achieved with the only failure being that a damp sponge was classified as a feather. The approach in (Xu et al., 2013) may have obtained 99% classification, however this was while assessing 10 substantially different materials, varying from a brick to a light feather. Therefore, the materials have very different proper- ties allowing for more straight forward classification. Furthermore, all three tactile properties were used making the system both slow and computation- ally expensive. It may be concluded that using just one material property does not allow for desirable classification of materials and combining modalities can lead to high computation cost. The ideal scenario would be to design a system that 8 can perform classification with accuracy similar to that of (Xu et al., 2013), using similar materials, with reduced computational burden and hence run in real-time (or close to). An overview of the best performing machine learning methods for the application of material identification using tactile sensing can been seen in Table 1. It demonstrates that the biggest difficulty in classification of materials is between similar materials or materials of similar roughness. Although, this is to be expected, it is still an ongoing challenge within the field. 3. Methodology 3.1. Experimental set-up 3.1.1. Tactile Sensor The BioTAC fingertip is a tactile sensor that is shaped like a human fin- gertip and is filled with an incompressible liquid, giving it similar compliance to a human fingertip (Lin et al., 2009; Kerr et al., 2013). The sensor is ca- pable of detecting a full range of information: micro vibrations, forces and temperature, similar to a human finger. Figure 1 shows a cross section view of the BioTAC fingertip. Figure 1: Cross Section View of BioTAC Fingertip Tactile Sensor (Syntouch, 2013) There are various sensors on the fingertip that enable retrieval of the aforementioned sensory information. An array of 19 electrodes measures the force applied on the fingertip by reading the impedance between each electrode and four common excitation electrodes. As the impedance over the sensing electrode increases, the measured voltage decreases, demonstrating 9 Table 1: Table comparing tactile sensing-based material identification methods Research Group Machine Learning Method Sensor Type Accuracy achieved (%) Description Xu et al. (2013) Probability Matrices Multi-modal tactile sensor (BioTAC) 99.00 10 very different materials were used. Continuous exploration and three modalities required therefore high computational costs. Hoelscher et al. (2015) SVM Multi-modal tactile sensor (BioTAC) 97.00 Struggled to distinguish between similar materials. Required training from all materials in the group Drimus et al. (2011) KNN Single Mode piezoresistive tactile sensor array 92.0 Classification across just 10 objects. Struggled to identify between similar objects. Chathuranga et al. (2013) ANN Multi-modal tactile sensor 85.0 Classification across just 7 materials, 6 of which were fabrics. Ho et al. (2012) ANN Single Mode electro-conductive soft skin sensors 81.0 Reported to be a very computationally expensive approach. Jamali and Sammut (2011) NBTree Single mode silicon tactile sensor with PVDF films 95.0 Struggled to distinguish between materials of similar roughness. Bhattacharjee et al. (2015) SVM Single mode thermistor-based tactile sensor 98.0 Longer contact with materials required for maximum accuracy and test materials varied widely Bhattacharjee et al. (2016) SVM Multi-modal Hand-held tactile sensor 98.8 Classification accuracies achieved across already learned data, accuracy drops to 84.0% when tested against new data from the same materials. 10 contact in that area of the finger. Having a 19 electrode array allows the exact point of contact on the fingertip to be derived. Furthermore, it enables multiple points of contact to be derived at any instance. Inside the core of the fingertip there is a pressure sensor located in the rigid core. This sensor reads pressure changes in the conductive fluid located between the hard core and the elastomeric skin. The vibration sensor generated two values, a DC pressure signal (PDC) that is obtained using a low pass filter and an AC pressure signal (PAC) that is obtained using a band pass filter. The PAC data represent the vibration strength recorded when the BioTAC finger makes contact with materials or objects. Static temperature (TDC) and thermal flow (TAC) data can also be collected from using the BioTAC sensor as there is a thermistor at the tip of the finger. TAC data represent the rate that heat is transferring from the fingertip to an object. This is made possible by the thermistor which is located in the tip of the fingertip. Using the BioTAC fingertip, two actions were performed on test materi- als to classify them. The first action was a press action where the fingertip was pressed into the material at a constant force. The second action was a slide action where the fingertip was dragged across the surface of the ma- terial at a constant force. Both actions aim to replicate two of the actions outlined by experimental psychologists Lederman and Klatzky (1987), that a human completes when first investigating an unknown material, namely static contact to determine thermal properties and lateral sliding movements to determine surface texture. The press action enables collection of the TAC and TDC data whereas the slide action enables the collection of PDC, TDC, PAC and TAC; these data are used for classification in Section 3.2. 3.1.2. Materials Classified Consistent with previous work (Kerr et al., 2014a,b), we use 14 mate- rials with the slide and press actions. We have intentionally selected some materials to be similar to determine if the approach can readily classify the material type and then identify the specific material. The 14 materials are Acrylic (Ac), Rough Acrylic (AcR), Copper (Cr), Aluminium (Al), Rough Copper (CrR), Rough Aluminium (AlR), Redbrick (R), Glossy Cardboard (GC), Plain Cardboard (PC), Soft Foam (S), Carpet (Ct), Doormat (D), MDF (M) and Pine (P). Figure 2 presents pictures of each of the materials used. Analogous with (Kerr et al., 2014a,b), the materials used in the experi- ments were divided into six groups corresponding to the six different material 11 Figure 2: Samples of the 14 materials used in the experimental set-up. 12 Table 2: Table showing the individual materials and the groups to which they belong. Material Group Acrylic (Ac) Plastic (P) Rough Acrylic (AcR) Plastic (P) Copper (Cr) Metal (Me) Aluminium (Al) Metal (Me) Rough Copper (CrR) Metal (Me) Rough Aluminium (AlR) Metal (Me) Redbrick (R) Masonry (Ma) Glossy Cardboard (GC) Cardboard (C) Plain Cardboard (PC) Cardboard (C) Soft Foam (S) Fabrics (F) Carpet (Ct) Fabrics (F) Doormat (D) Fabrics (F) MDF (M) Wood (W) Pine (P) Wood (W) types. The individual materials and the groups to which they belong are shown in Table 2. The artificial system was then evaluated on its ability to classify each material individually. 3.2. Data Collection Analogous with work carried out in (Kerr et al., 2013; Xu et al., 2013), the BioTAC fingertip was powered on and left to rest for 15-20 minutes to allow it to reach its steady state temperature (approximately 31◦C, 10◦C above ambient). A rig including a motorised arm and turntable was designed and built for the experiment. Two stepper motors (one for the arm and one for the turntable) were connected to an Arduino UnoTM board and a Graphical User Interface (GUI) was developed using Python to control the motors. An image of the experimental rig and GUI can be seen in Figure 3(a) and (b) respectively. The fingertip was placed at the end of the motorised arm so that it could be moved towards or away from the material. Additionally, the material was placed on the motorised turntable to allow for the material to be moved below the finger to replicate a sliding action. To collect data corresponding to a thermal exploratory movement, the fingertip is pressed onto a material using a constant force of 2N, measured 13 (a) (b) Figure 3: a) Image showing the experimental rig; b) Screen shot of the developed GUI. 14 by an ATI Nano17 6-axis F/T Sensor. To allow time for the heat flow to stabilise, it is necessary to collect 20 seconds of data for the press action. For the slide action, the finger tip was pressed on the material using a constant force of 1.59N (also measured with the aforementioned force sensor), and moved along the material surface for approximately 5 cm. The BioTAC data are recorded at 4400Hz and in a sequence, with the PAC values and the individual electrode values alternating (i.e. PAC data are recorded at 2200Hz). All other values such as TDC (static temperature), TAC (thermal flow rate) and PDC (static vibration) are given at the end of each sequence of values. An example of this sequence can be seen in Figure 4. All values are sampled with 12-bit resolution. These values are extracted and individual datasets for PAC, PDC, TAC and TDC data are obtained using MATLAB. Figure 4: Diagram showing the Sequence of Data from the BioTac Fingertip 3.3. Pre-Processing To obtain high quality thermal conductivity (TAC and TDC) time series data, we use only the data collected after the contact force has reached its maximum, this was found to occur approximately 8 seconds after contact. The TAC and TDC datasets, collected during the press action, each con- tain 800 data points. Additionally a set of 220 data points for each of TAC, TDC, PAC and PDC is obtained during the slide action. In line with other work presented in the literature, initial evaluation of using three modalities (i.e. thermal properties, vibration data and impedance values to represent compressibility) was conducted. It was hypothesised that due to the vast 15 quantity of electrode data collected from the BioTAC fingertip, the process- ing time would increase greatly when identifying materials. Initial visual analysis of the electrode values (voltage related to impedance) and initial ex- perimentation was conducted to quantify the additional time it would take to include analysis of compressibility when classifying the materials in this experiment. As demonstrated in Figure 5, it is clear that all materials be- have relatively similar during the press action with the exception of softer and rougher materials such as the softfoam and doormat. However, these dif- ferences are also clearly identified when analysing the vibration data (Kerr et al., 2014a) and therefore it was felt that any additional benefit from in- cluding compressibility data in this experiment was minimal to none. In fact, preliminary results indicated a decrease in accuracy when the electrode data were considered as part of the algorithm. Furthermore, the computation time to include the compressibility data was analysed. It took approximately 6 times longer to train one fold of a SVM network when the vast compress- ibility data was included in comparison to when only thermal and vibration data are considered. Therefore, considering the decrease in accuracy in pre- liminary investigations with compressibility data being considered, together with the increase in computation time, we use only thermal and vibration data in the experiments in this paper. Initially we trained the classifiers with raw TAC, TDC, PAC and PDC data collected directly from the BioTAC sensor, however these resulted in poor classification accuracy such as 68% training accuracy for the ANN with 56% testing accuracy. Hence, to improve the characteristics of the data we applied Principal Component Analysis (PCA) to reduce the dimensionality of the inputs. We applied PCA using both Eigen-decomposition of the sample’s covariance and Singular Value Decomposition (SVD). With the sensor data, using the Eigen-decomposition based approach generated the best inputs for the classifiers, evidenced by the training and testing accuracies over multiple experiments. Hence, the presented approaches and results use PCA with Eigen-decomposition. In (Kerr et al., 2013), PCA was applied to a combined set of TAC and TDC data (1600 values), however in this paper we apply PCA to each individual modality (PAC, PDC, TAC, TDC) first. In this way, the principal components matrix for each material is formed from dif- ferent modalities tailored for either the press or slide action experiment. As described in Section 3.4, the principal components matrices are used to train the numerous classifiers and hybrid combinations of classifiers evaluated. 16 Figure 5: Graph showing Initial Analysis of Compressibility Data from the BioTac Fin- gertip 3.4. Classifiers The data collected from the fingertip during the press and slide action ex- periments were normalised and PCA using Eigen-decomposition was applied to extract features. The optimal number of principal components (PCs) for each dataset was determined empirically as three. Hence in total there are 18 inputs to a classifier (six modalities and three PCs per modality, 12-bit values) for each experiment. To obtain the classification accuracy of each classifier in each experiment, the classified test data form an output ma- trix which is compared with a matrix containing the actual material labels (6 classes for classification of materials into the groups and 14 classes for classification of the individual materials). Average classification accuracy is obtained using 5 fold cross validation. The MATLAB package (MATLAB, 2013) and R package (R Core Team, 2013) were used to implement the range of classifiers. Each approach used for classification is described in more detail in Section 3.4.1 to Section 3.4.10. 3.4.1. One Stage Support Vector Machine (SVM) Support Vector Machine is a useful and well established technique used in machine learning to learn and classify data. The main objective of a 17 SVM is to separate a set of inputs by identifying a line, or hyperplane in n-dimensional space. Each of the input points is known to belong to one of two classes. If it is possible for the algorithm to find a suitable hyperplane that will minimise future misclassification, then the data are easily classified based on what side of the hyperplane they reside. However, it is common that it is not possible to linearly separate the data. In this instance the SVM algorithm is required to map the input data to a higher dimensional space through the use of a kernel function (KF). Although there are many types of KFs, five are evaluated in this work and their performances compared. Within each KF there are various parameters that can be changed to optimise the performance of the classifier. One common parameter is the soft margin (sometimes known as the box constraint ), C. The purpose of the soft margin is to modify the SVM algorithm to tolerate training errors by intuitively tolerating a few outliers on the wrong side of the hyperplane. Deciding on the size of the margin requires parameter tuning and optimisation. The KFs evaluated for the classification of materials and their respective parameters that were optimised are: • Linear KF, Quadratic KF and Multilayer Perceptron (MLP) KF: Soft margin value (C) is optimised. • Polynomial KF: Two parameters are optimised in this KF; the soft mar- gin value (C) and the order of polynomial (p). Various combinations of values were evaluated in order to find the best performing combination of parameters. • Radial Basis Function (RBF) KF: Two parameters are optimised in this KF; the soft margin value (C) and the gamma parameter for the Gaussian RBF (γ). The Gamma parameter is a positive number rep- resenting the scaling factor of the RBF. Various combinations of values were evaluated in order to find the best performing combination of parameters. SVM is more commonly used as a binary classifier. However, binary classification is not suitable for the material classification task outlined in this work; the standard SVM algorithm has been adapted to enable multi- classification for material classification. Two common ways of utilising binary SVM classification techniques to allow for multi-class classification are by either having a sequence of numerous one-versus-one binary classifiers or by 18 having a multiple one-versus-all classifier. The approach utilised in this work is a series of one-versus-all classifiers in which the system is trained with each class classified against the samples of all the other classes, to develop a SVM multi-class classifier. Two separate experiments were conducted: one for the classification of the materials into their six respective groups and one for the classification of the fourteen individual materials. All KFs were evaluated and their parameters optimised for both experiments. For each optimised KF implemented and evaluated, three different methods for calculating the separating hyperplane were also implemented and evaluated. The three standard methods evalu- ated were Sequential Minimal Optimization (SMO), Quadratic Programming (QP) and Least-Squares (LS) method. Table 3 outlines the best performance achieved from each KF for material group classification, the best performing method for calculating the separating hyperplane and the respective opti- mised parameters. It can be seen that the two best performing KFs with a classification rate of 86.19% are the quadratic KF and the polynomial KF. Although the two KFs performed equally well in terms of classification, the polynomial KF was found to be marginally more efficient than the quadratic KF at achieving the highest classification rate. The quadratic KF took 0.49 seconds to train for one fold of data whereas the polynomial KF took 0.47 seconds. Therefore, the Polynomial KF with a soft-margin value of 0.01 and a polynomial order of 2 was used in the SVM implemented in this approach. Table 3: Table comparing the SVM KF classification accuracies. Kernel Function Best Performing Method Optimised Parameters Classification Accuracy Linear QP C = 1.00 79.524% Quadratic SMO C = 0.01 86.190% MLP LS C = 0.01 57.143% Polynomial SMO C = 0.01 p = 2.00 86.190% RBF QP C = 0.10 γ = 1.00 66.667% 19 The SVM system was used for both experiments with the only difference being the number of groups/ classes (g) that the SVM had to classify. For the first experiment, i.e. classifying the materials into their six groups, g = 6 and for the second experiment involving 14 individual materials, g = 14 as described in Section 3.1.2. 3.4.2. Two Stage SVM In contrast to having two separate experiments for the classification of the materials into their groups and the classification of individual materials, as described in Section 3.4.1, this section explains a two stage approach to material classification. This approach utilises the multi-class SVM classifica- tion algorithm explained in Section 3.4.1 to firstly classify the materials into their groups and then uses this information to further classify the materials individually using a mixture of binary and multi-class SVM classification in the second stage. The same SVM design as outlined in Section 3.4.1 is used in the first stage of this two stage approach. For the second stage, 6 SVM systems are trained, one for each material group, some of which have different numbers of outputs/ groups for the individual materials within them, requiring the need for binary classification in some cases and multi-class classification in others. These SVM systems are saved and later used for testing the data input at the second stage. Similar to the development of the SVM in the one stage approach, five different KFs with a range of parameters and hyperplane calculation methods were evaluated for the second stage SVM classification system. The three equally best performing KFs in the second stage were the Quadratic KF, the Polynomial KF and the RBF KF, with the most efficient proving to be the Polynomial KF. Therefore, the Polynomial KF was used in the second stage SVM. Upon completion of material group classification in the first stage, the results of the classification are used as the inputs to the second stage. A model of the two stage SVM approach can be seen in Figure 6. If a material was correctly classified into a specific group in the first stage then this activates the second SVM to classify the individual material within this material group. However, if the material group is incorrectly classified in the first stage then it is recorded as an incorrect classification in the final output matrix and is not classified in the second stage to improve efficiency. Likewise, if the material is classified incorrectly in the second stage in comparison with ground truth (i.e. it has been correctly classified 20 Figure 6: Diagram showing the 2-Stage SVM approach used for material classification 21 in its group but incorrectly classified individually), then it is recorded as an incorrect classification in the output matrix. 3.4.3. One Stage ANN We use a back propagation ANN as illustrated in Figure 7. During train- ing, the number of neurons in the hidden layer and the number of training epochs were varied and evaluated. This determined that the ideal structure for the ANN contained 75 neurons and required 1500 epochs for training. The network structure illustrated in Figure 7 was applied in both exper- iments differing numbers of outputs (n). In the first experiment, n = 6 as there are six material groups and in the second experiment n = 14 corre- sponding to fourteen individual test materials, as described in Section 3.1.2. 3.4.4. Two Stage ANN The two stage approach (Kerr et al., 2014b), used a series of back prop- agation ANNs, each with one hidden layer, where the output of the first ANN is the input of the second to allow for group classification first followed by individual classification. This differs from the one stage ANN approach outlined Section 3.4.3 as the two networks work in series rather than two individual networks, one classifying materials into groups and one classify- ing individual materials. The two stage approach structure can be seen in Figure 8. The ANNs used in the two stage approach are similar in structure to the ANN used in the one stage approach explained in Section 3.4.3. The first ANN is used to classify the materials into groups and the second stage of the ANN is comprised of six individual ANNs for each material group (i.e. plastic, metal, masonry, fabrics, paper and wood). These ANNs were all trained individually with their respective outputs relating to how many materials there are in each group, for example plastic has two outputs for the two plastic materials whereas metal had four outputs for the four metal materials in that group. After training these networks were saved and then used for classifying the resulting output from the first ANN. In order to avoid false positives being identified, a threshold was empirically identified and set for testing the materials within the group. If the output of the neuron fired for the material sample being tested is not greater than 0.3, then it was considered to be a failed classification. 22 Figure 7: Diagram showing the ANN used for material classification 23 Figure 8: Diagram showing the two stage ANN used for material classification 24 3.4.5. SVM and ANN Hybrid System A hybrid approach of the SVM classifier for material group classification and separate ANN classifiers for the individual material classification within each group has been developed. The structure of the hybrid system can be seen in Figure 9. Figure 9: Diagram showing the SVM+ANN hybrid approach used for material classifica- tion A SVM classifier, as described in Section 3.4.2, is used first to classify the materials into their material group (i.e. using the polynomial KF with a poly- nomial order value of 2 and a soft margin value of 0.01). Upon classification into their groups an output matrix is formed and the values in this matrix initiates the second stage of the hybrid approach with six ANNs for indi- vidual material classification within each of the material groups. The ANNs are constructed as described in Section 3.4.4 (i.e. with one hidden layer of 75 neurons and trained for 1500 epochs). As with the other two stage ap- 25 proaches outlined in this section, in order to improve efficiency only materials correctly classified into their group are classified in the second stage. 3.4.6. Gaussian Mixture Models (GMMs) GMM based on the Maximum Likelihood (ML) estimation using Expec- tation Maximisation (EM) was also used for classification. The Gaussian probability density function (PDF) is a bell shaped curve defined by two pa- rameters, mean and variance. The Gaussian distribution is commonly used for approximating a class model shape in a selected feature space. It is as- sumed that the class model is truly a model of one basic class, however if the actual PDF is multi-modal it fails. GMM is a mixture of several Gaussian distributions and is therefore able to represent different subclasses within a class. The resulting probability density function is defined as a weighted sum of the Gaussian distributions that make up the GMM. The PDF of the GMM can be used to calculate the likelihood of any new input data within each class and identify the class for which the PDF generates the maximum likelihood and therefore classify the data. The EM algorithm is an iterative method for calculating the maximum likelihood distribution in GMMs (Dempster et al., 1977). GMMs are calculated to represent each material (cluster centres are ini- tialised randomly and the k -means algorithm is used for calculating conver- gence). The EM algorithm is then used to calculate the maximum likelihood and the associated class label is selected. These classified labels are used to form an output matrix. As the cluster centres are initialised randomly 10 runs of each classification (i.e. material groups and individual materials) were completed and an average classification accuracy calculated. 3.4.7. Linear Discriminant Analysis (LDA) LDA finds a linear combination of features that characterises or separates two or more classes of objects or events by creating the linear combination which yields the largest mean differences between the classes (Martinez and Kak, 2001). It is a generalisation of Fisher’s linear discriminant and it is a widely used method in statistics, pattern recognition and machine learning. LDA can be used for binary or multi-class classification. In this work it is used for multi-class classification to identify materials individually and into their groups. 26 3.4.8. Naive Bayes Naive Bayes is a family of algorithms used for training classifiers capable of binary or multi-class classification. Rather than the classifier taking all features of a class into account collectively to describe the class, all naive Bayes classifiers assume that the value of a particular feature is completely independent of any other feature in the class meaning they can be trained very efficiently during a supervised learning task (Rish, 2001). In this work it is used firstly to classify materials individually and secondly classify them into their groups. 3.4.9. k-Nearest Neighbour (k-NN) k -NN is a non-parametric method which can be used for classification or regression. In this work is it used for multi-class classification of materials. When used for classification. an object is assigned to the class which is most common amongst its k nearest neighbours following a majority vote. The number of neighbours is defined by the user and is normally a small positive integer, for example if k =1 then the object is assigned to the class of the single nearest neighbour or if k =3 then the most popular class amongst the three closest neighbours is the class that the object will be assigned to. A range of k values was tested to identify the most accurate number of neighbours to take into account for maximum classification accuracy. The k value that proved to be most effective for classifying the materials individually and into their groups was k =5. 3.4.10. Multi-Layer Perceptron (MLP) Multi-layer Perceptron (MLP) is a feed forward ANN which uses back- propagation to train the network. A MLP consists of multiple layers of adaptive weights that are trained to map sets of input data onto a set of appropriate outputs (Bishop, 1995). A range of numbers of layers and neu- rons on each layer was evaluated in order to find the optimal structure for the MLP. The structure of the MLP used in this work to classify the mate- rials into their groups and individually was a network consisting of 10 layers containing 19 neurons each. 4. Results In order to evaluate the artificial system’s performance, analysis com- paring the algorithms was conducted. In addition, the artificial system was 27 evaluated against human performance. In this section, the results of the human evaluation experiments are presented. The initial analysis of the extracted data collected from the fingertip during the experiments is also presented. Furthermore, the performances of the aforementioned classifiers are presented, evaluated against each other and against human performance. 4.1. Human Evaluation To evaluate each artificial system, they are compared against human per- formance in identifying the the same set of 14 materials. The participants were evaluated on identifying the materials both individually and identify- ing which group each material belonged to. Initially they were evaluated by only being allowed to press into the materials with their finger and assess its thermal properties and secondly they were evaluated using both a press and slide action (to evaluate the surface texture of the material). The findings from these experiments can be found in (Kerr et al., 2013) and (Kerr et al., 2014a) respectively. 4.2. System Evaluation A comparison of the results obtained for all of the classification systems presented and the average classification accuracies of the human participants are shown in Table 4. As specified in Section 3.4, the results for the artificial system are calculated by computing the average of the classification rates using 5-fold cross validation. As shown in Table 4, SVM has outperformed the other approaches for both the classification of the materials into their groups and for classification of the individual materials. The two stage SVM has proven to be capable of maintaining the group classification accuracy for the classification of the individual materials in its second stage. As the SVM only has to use binary classification in three out of the six groups and the most materials to classify between in the other groups is four, then it was possible to achieve such classification accuracy in the second stage using SVM. The second best performing algorithm is the SVM and ANN hybrid approach. The SVM and ANN hybrid approach using the quadratic KF achieved 81.89% classification. Using the polynomial KF within the hybrid approach proved to be marginally more accurate, achieving 82.86% accu- racy. The Gaussian Mixture Models (GMM) approach was also a highly performing classifier with 86.14% and over 80% for material groups. How- ever, it can be seen in Table 4 that aside from the fact that they perform 28 Table 4: Table comparing the experimental results. Modalities Utilised Run Time (to train one fold) Material Group Individual Materials Human Participants - Thermal Only (Kerr et al., 2013) N/A 61.00% 50.90% Human Participants - Thermal and Vibration N/A 79.76% 69.64% One Stage ANN - Thermal Only (Kerr et al., 2013) 4.94secs 73.00% 60.90% One Stage ANN - Thermal and Vibration 5.47secs 83.81% 79.05% Two Stage ANN - Thermal and Vibration 7.66secs 83.81% 70.48% One Stage SVM - Thermal and Vibration 0.48secs 86.19% 72.38% Two Stage SVM - Thermal and Vibration 0.55secs 86.19% 86.19% Hybrid SVM and ANN - Thermal and Vibration 2.66secs 86.19% 82.86% GMM - Thermal and Vibration 13.35secs 86.14% 80.24% LDA - Thermal and Vibration 1.52secs 78.57% 80.95% NB - Thermal and Vibration 1.68secs 71.43% 80.95% k -NN - Thermal and Vibration 0.95secs 71.43% 64.29% MLP - Thermal and Vibration 5.45secs 71.43% 66.67% 29 with slightly less accuracy in comparison with the two stage SVM algorithm, like the other systems, the GMM system also takes significantly longer to train (taking 13.35secs to train for one fold of data) than the SVM system (taking 0.48secs to train for one fold of data). Interestingly the LDA and NB classifiers both performed better when classifying the materials individually rather than into their groups. This is contrary to what was expected as there are 14 outputs when classifying into individual materials compared with only six outputs for group classification. As a consequence, LDA and NB both performed marginally better than GMM for classifying the individual ma- terials with 80.95% accuracy however performed poorer than GMM for the classification of material groups with 78.57% accuracy for LDA and 71.43% accuracy for NB. k -NN was found to perform optimally when k = 5 with a fast training time of 0.95sec for one node, however in comparison with the other classifiers it is still lacking in accuracy for classifying materials individ- ually and into their groups. MLP was found to be relatively slow to train and was out performed by the majority of the other classifiers. The training times presented are based on training one fold of data for each classifier and the training was completed in Matlab on a desktop PC with a Intel R© Xeon R© CPU E5-16070 @ 3.00GHz 4 processor and 16Gb of RAM. It is demon- strated in Table 4 that the novel two stage SVM and hybrid SVM and ANN approaches have outperformed all standard approaches such as GMM, LDA or singular SVM and ANN approaches. These architectures have proven to be more accurate than “off-the-shelf” architectures with the novel two stage SVM system being the best performing and most efficient system for the task of material classification when compared with others. In order to assess robustness, completely unseen objects and materials were presented to the system. Seven previously completely unseen objects in total were presented namely, a plastic bottle, a ceramic cup, a glass cup, a metal cup, a metal fork, a plastic fork and a paper plate. Three datasets for each object were tested. There is no “any other material” option in the system so every material will be classified as one of the 6 materials outlined in Table 2. Interestingly, materials that the system had not previously been trained on, i.e. ceramic and glass, were identified as a material with similar properties. The glass was identified as plastic in the majority of cases and the ceramic was classed as metal in all cases. In some instances the system struggled to identify the materials on which the system had been trained. The paper plate was only correctly identified as paper in 1 out of the 3 instances. However, this may be due to the thin plastic film coating on 30 paper plates that makes them water retardant, as shown in Figure 10(a). This made it difficult to identify the plate as paper and indeed it was also incorrectly identified as plastic in both other instances. The plastic fork was also correctly identified as plastic in 1 of the three instances. This may be due to the fact that this specific metal fork was manufactured to be “metal like”, i.e. look like a metal fork as shown in Figure 10(b). This means that there is a fine film on the surface of the fork making it difficult to identify as plastic, similar to the difficulty faced with identifying the paper plate. Indeed, in both other cases the fork was identified as being metal. (a) (b) Figure 10: a) Image showing two of the previously unlearned objects; a) the paper plate; b) the “metal-like” plastic fork However, the metal cup was identified as the correct material in at least 2 of the 3 instances and both the plastic bottle and metal fork were correctly identified as their respective material in all 3 instances. This illustrates the 31 robustness of the system and with further training for more materials the system is capable of correctly identifying previously learned materials. 5. Discussion From the results of the human evaluation carried out in (Kerr et al., 2014a), it was found that the average accuracy was 79.76% when identifying the material groups and 69.64% when identifying individual materials. A detailed analysis of these findings can be seen in the confusion matrices in Figure 11(b) and 12(b). It can be seen from Table 4 that all of the artificial systems presented have outperformed the human participants (when perfor- mance is averaged across all test materials), in some cases by over 6% for material group classification and by over 16% for individual material classi- fication. This is due to the fact that is easier for a human to determine what group a material would belong to (i.e. identify that the material is definitely metal) than it is for them to identify a specific material within that group (i.e. is it copper or aluminium). It can be seen from the confusion matrix in Figure 11(a) and 11(b) that both the artificial system and the human partic- ipants successfully classified the soft foam in every instance and did not get it confused with any other material. It can also be seen from Figure 11(a) that where the artificial system classified the pine material correctly on 60% of the instances, the highest misclassification was with MDF, another ma- terial with similar properties and in the same material group. Likewise, for classifying MDF, the largest misclassification was with pine. This misclas- sification can also be witnessed as a common misclassification amongst the human participants, with 50% of the participants correctly classifying MDF but in almost 42% of the instances the human participants misclassified MDF as pine as shown in Figure 11(b). It is shown in Figure 11(a) that a common misclassification of the artificial system was a confusion between acrylic and glossy cardboard, with glossy cardboard being classified correctly for 40% of the instances however in just over 33% of the instances it was misclassified as another material with similar texture and thermal conductivity properties: acrylic. Interestingly, the identification of the rough materials was evidently chal- lenging for the human participants, with only 2 out of 12 being able to identify all three rough materials. For example, one of the most challenging materials for humans to identify was the rough copper, and it was commonly misclassified as redbrick, as shown in Figure 11(b). This is due to the fact 32 (a) (b) Figure 11: a) Confusion matrix showing the percentage accuracy of the artificial system (using 1 stage SVM) for classifying all the materials individually; b) Confusion matrix showing the percentage accuracy of the human participants for classifying all the materials individually, collated from the findings in (Kerr et al., 2014a). 33 that the redbrick is perceived as cold and rough by the human participants, in a similar manner to how metals are perceived. Therefore, this is one of the instances where the artificial system outperforms the human participants, as the system can analyse thermal conductivity much more accurately than the human participants and therefore not misclassify it as a masonry material. This is shown in Figure 11(a) where the artificial system classifies the rough copper in just over 93% of the instances with the only misclassification be- ing with another rough metal: aluminium. This common misclassification is also evident when classifying the materials into their groups, as shown in Figure 12(a) and 12(b). The artificial system outperforms the human par- ticipants when classifying materials in both metal and masonry groups. The majority of instances of misclassification for the human participants when classifying masonry objects is confusing it with a metal object, as both are perceived as cold. From the results of the human evaluation reported in (Kerr et al., 2014a), it was evident that the introduction of texture analysis by sliding their finger along the material has increased the accuracy of the human participants for classifying materials, particularly when identifying the material groups. It should be noted that due to the nature of some materials, humans can in- tuitively take other characteristics into account, for example compressibility (i.e. soft foam is much more compressible than the other materials). The artificial system was trained to analyse only the thermal properties and sur- face texture of the material and not compressibility. This may have given the human participants a slight advantage in identifying some materials. This is evident when comparing the accuracy of the artificial system for classifying fabrics as a material group (80%) with the accuracy of the human partici- pants (100%), as shown in Figure 12(a) and 12(b). The human participants have no misclassification of the fabrics when identifying their material group. It can be assumed that this is due to the fact that the human participants could identify the compressibility of the objects like soft foam and carpet whereas the artificial system was not trained to analysis this modality. How- ever, even though the artificial system did not consider a key modality for humans, namely compressibility, overall it still outperformed humans. 6. Conclusion and Future Work A range of classifiers and combinations of classifiers that were imple- mented, tuned and evaluated for the classification of materials into their 34 (a) (b) Figure 12: a) Confusion matrix showing the percentage accuracy of the artificial system (using 2 stage SVM) for classifying all the materials into groups; b) Confusion matrix showing the percentage accuracy of the human participants for classifying all the materials into groups, collated from the findings in (Kerr et al., 2014a). 35 respective groups and the classification of individual materials has been pre- sented. This classification was based on a combination of the thermal and surface texture properties of the materials. It is proven that a two stage SVM approach performs best and is the most efficient method in completing the material classification task. It was found that the artificial system’s average performance across all test materials exceeded that of human participants’ average performance, with the two stage SVM approach outperforming hu- mans by over 16%. However, as the system does require training it is still slower at classification than the human participants due to humans’ inherent sophisticated tactile sensing. There are two immediate future directions of this work. Firstly, the pre- sented tactile based methods for material classification could be used in col- laboration with vision based methods. If an estimate of the physical size of an object is determined using image processing, then the volume of the object can also be calculated. The tactile sensing methods presented in this paper can identify the material from which the object is made and armed with the volume of the object, an estimate of the object’s weight can be calculated. The incorporation of vision would further strengthen the identification of objects and materials. To train a robotic system on a vast array of objects and incorporate vision, a significant dataset of tactile data, static image and video data should be collected for numerous materials, objects, clothing and human skin. A deep learning structure could then be utilised for the robotic system to learn the vast dataset of materials and objects. This information could be used to help determine the effort and grasping method required to successfully lift previously unlearned objects. This would provide a useful skill to assistive robots in a home environment when learning the affordances of objects they have not previously encountered. Secondly, the success of the challenging task of material identification presented in this paper proved invaluable in providing familiarity with the use of the BioTAC fingertip and the analysis of the data it presents. It became apparent that the high sen- sitivity of the BioTAC sensor means that it may be capable of measuring biomedical parameters, namely human vital signs such as a human pulse. Therefore, this also directs the research in this paper towards the automated collection and analysis of human vital signs based on tactile perceptions. 36 References Bhattacharjee, T., Wade, J., Chitalia, Y., and Kemp, C. (2016). Data-driven thermal recognition of contact with people and objects. In 2016 IEEE Haptics Symposium, HAPTICS 2016, Philadelphia, PA, USA, April 8- 11, 2016, pages 297–304. Bhattacharjee, T., Wade, J., and Kemp, C. (2015). Material recognition from heat transfer given varying initial conditions and short-duration contact. In Proceedings of Robotics: Science and Systems, Rome, Italy. Bishop, C. (1995). Neural Networks for Pattern Recognition. Oxford Univer- sity Press. Chathuranga, D., Ho, V., and Hirai, S. (2013). Investigation of a biomimetic fingertip’s ability to discriminate fabrics based on surface textures. In Advanced Intelligent Mechatronics (AIM), 2013 IEEE/ASME Interna- tional Conference on, pages 1667–1674. Dempster, A., Laird, N., and Rubin, D. (1977). Maximum likelihood from incomplete data via the em algorithm. JOURNAL OF THE ROYAL STATISTICAL SOCIETY, SERIES B, 39(1):1–38. Drimus, A., Kootstra, G., Bilberg, A., and Kragic, D. (2011). Classification of rigid and deformable objects using a novel tactile sensor. In Advanced Robotics (ICAR), 2011 15th International Conference on, pages 427– 434. Fishel, J. and Loeb, G. (2012). Bayesian exploration for intelligent identifi- cation of textures. Frontiers in Neurorobotics, 6(4). Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., and Witten, I. (2009). The WEKA data mining software: an update. SIGKDD Explorations, 11(1):10–18. Ho, V., Araki, T., Makikawa, M., and Hirai, S. (2012). Experimental in- vestigation of surface identification ability of a low-profile fabric tactile sensor. In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ In- ternational Conference on, pages 4497–4504. 37 Hoelscher, J., Peters, J., and Hermans, T. (2015). Evaluation of tactile feature extraction for interactive object recognition. In Proceedings of 15th IEEE-RAS International Conference on Humanoid Robots, pages 310–317, Seoul. IEEE. Jamali, N. and Sammut, C. (2010). Material classification by tactile sensing using surface textures. In Robotics and Automation (ICRA), 2010 IEEE International Conference on, pages 2336–2341. Jamali, N. and Sammut, C. (2011). Majority voting: Material classification by tactile sensing using surface texture. Robotics, IEEE Transactions on, 27(3):508–521. Jurie, F. and Triggs, B. (2005). Creating efficient codebooks for visual recog- nition. In Computer Vision, 2005. ICCV 2005. Tenth IEEE Interna- tional Conference on, volume 1, pages 604–610. Kerr, E., McGinnity, T., and Coleman, S. (2013). Material classification based on thermal properties - a robot and human evaluation. In 2013 IEEE International Conference on Robotics and Biomimetics, December 12-14 2013, Shenzhen, China, pages 1048–1053. Kerr, E., McGinnity, T., and Coleman, S. (2014a). Material classification based on thermal and surface texture properties evaluated against hu- man performance. In 2014 IEEE International Conference on Control, Automation, Robotics and Vision, ICARCV 2014, Singapore, Dec 2014. Kerr, E., McGinnity, T., and Coleman, S. (2014b). Tactile approach to material classification - evaluated with human performance. In Irish Machine Vision and Image Processing 2014, Northern Ireland, pages 175–180. Lederman, S. and Klatzky, R. (1987). Hand movements: A window into haptic object recognition. Cognitive Psychology, 19(3):342–368. Lin, C., Erickson, T., Fishel, J., Wettels, N., and Loeb, G. (2009). Signal processing and fabrication of a biomimetic tactile sensor array with ther- mal, force and microvibration modalities. In Robotics and Biomimetics (ROBIO), 2009 IEEE International Conference on, pages 129–134. 38 Lowe, D. (1999). Object recognition from local scale-invariant features. In Computer Vision, 1999. The Proceedings of the Seventh IEEE Interna- tional Conference on, volume 2, pages 1150–1157. Luo, S., Liu, X., Althoefer, K., and Liu, H. (2015). Tactile object recog- nition with semi-supervised learning. In Liu, H., Kubota, N., Zhu, X., Dillmann, R., and Zhou, D., editors, Intelligent Robotics and Applica- tions, volume 9245 of Lecture Notes in Computer Science, pages 15–26. Springer International Publishing. Martinez, A. and Kak, A. (2001). Pca versus lda. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 23(2):228–233. MATLAB (2013). version 8.10.0 (R2013a). The MathWorks Inc., Natick, Massachusetts. Monkman, G. J. and Taylor, P. (1993). Thermal tactile sensing. Robotics and Automation, IEEE Transactions on, 9(3):313–318. Pezzementi, Z., Plaku, E., Reyda, C., and Hager, G. (2011). Tactile-object recognition from appearance information. Robotics, IEEE Transactions on, 27(3):473–487. R Core Team (2013). R: A Language and Environment for Statistical Com- puting. R Foundation for Statistical Computing, Vienna, Austria. Rish, I. (2001). An empirical study of the naive bayes classifier. Technical report. Syntouch (2013). The syntouch website. Varma, M. and Zisserman, A. (2005). A statistical approach to texture classification from single images. Int. J. Comput. Vision, 62(1-2):61–81. Wade, J., Bhattacharjee, T., and Kemp, C. (2015). A handheld device for the in situ acquisition of multimodal tactile sensing data. CoRR, abs/1511.03152. Xu, D., Loeb, G., and Fishel, J. (2013). Tactile identification of objects using bayesian exploration. In IEEE International Conference on Robotics and Automation (ICRA) 2013. 39 
work_4h5iqpocibetdhutakf4nxpl3q	----	NASA Technical Reports Server (NTRS) 
work_4hioebfyd5dapocfnau6ybwyha	----	Intelligent monitoring for people assistance and safety Editorial DOI: 10.1111/exsy.12016 Intelligent monitoring for people assistance and safety This expert systems special issue on ‘Intelligent Monitoring for People Assistance and Safety’ contains the revised and extended best papers dealing with different issues concerning people assistance and safety through intelligent monitoring and activity interpretation, presented at ‘IWINAC 2011: the fourth International Work-Conference on the Interplay between Natural and Artificial Computation’. People assistance and safety is a hot topic and of crucial importance in indoor environments such as homes, offices, hospitals and schools as well as in outdoor areas. Environ- ments are increasingly well equipped with multiple sensing technologies that can monitor simple and complex activities and behaviours (Gascueña and Fernández-Caballero, 2011). Intelligent monitoring implies not only the analysis of the data captured from the various sensors but also their interpretation from the detection of the presence of certain events or actions previously defined (Rivas, Martínez-Tomás and Fernández-Caballero, 2011). From a historical perspective, it is acknowledged that the evolution of monitoring systems has gone through three generations. In the first generation (1960–1980), closed- circuit television analogue systems were used, which consisted of several cameras connected to a series of monitors. These systems do not process information and require a human operator to be permanently concentrated on analysing the situations observed on the monitors. However, in the second generation (1990–2000), advances attained in digital video communication (e.g. digital compression, bandwidth reduction and robust transmission) were used to increase the efficiency of monitoring systems: closed-circuit television systems were combined with computer vision technology to process images automatically, in order to be proactive in the detection of alarm events during recording. These semi- automatic systems required a robust tracking and detection algorithm for behaviour analysis. Whereas these systems represented a clear improvement with respect to first generation systems by reducing the dependency on human operators to detect anomalous situations, their algorithms and techniques were responsible for triggering a high number of false positives. In the third generation (2000-today), a series of heteroge- neous sensors (e.g. fixed cameras, pan-tilt-zoom (PTZ) cameras, audio sensors and RFID tags (radio-frequency identification) will be geographically distributed throughout the scenario to be observed. From the image processing point of view, these systems are based on distributed proces- sing capabilities and the use of embedded signal processing devices to gain distributed scalability and robustness. The main problems that need to be solved in third generation systems are the integration of data obtained from different sensors, establishing a correspondence of the signals in time and space and coordinating and distributing the processing task and video communication. The papers The papers selected in this special issue cover different layers present in modern intelligent monitoring systems (Micheloni et al., 2010), namely, the sensor layer (sensors and networks) and the surveillance layer (feature extraction, recognition, tracking and event analysis). A surveillance layer is described in two papers covering the aspect of human action/activity recognition. The propo- sal on ‘Human Activity Recognition based on Kinematic Features’ (Hernández et al. 2014), introduces a novel approach to recognise human actions in 2D sequences. It is based on the real-time visual tracking of people and the extraction of simple features. The paper proposes three complementary modules for the classical monitoring problems associated to (a) tracking; (b) feature extraction; and (c) action recognition. Tracking is based on the hybridi- sation of a particle filter and a local search procedure. Feature extraction characterises the silhouette of the tracked person by dividing it into several horizontal rectangular boxes. Then, the system computes statistics on the temporal evolution of these rectangular boxes. Lastly, action recogni- tion uses these statistics into a support vector machine to classify the actions. Another paper, ‘Human Action Recogni- tion with Sparse Classification and Multiple View Learning’ (Cilla et al., 2014), employs multiple camera viewpoints in the recognition of human actions to increase the perfor- mance. The authors present a feature fusion approach to efficiently combine the 2D observations. Multiple view dimensionality reduction is employed to learn a common parametrisation of 2D action descriptors computed for each one of the available viewpoints. Canonical correlation analysis and their variants are employed to obtain such para- metrisations. A sparse sequence classifier based on L1 regularisation is proposed to avoid the problem of having to choose the proper number of dimensions of the common parametrisation. On the other hand, two papers are related to sensors and sensor networks. The first work, ‘Learning Routines Over Long-Term Sensor Data Using Topic Models’ (Castanedo et al., 2014) faces the issue of the large amount of data generated by sensor networks when used as infrastructure to create intelligent environments. In this paper, topic © 2013 Wiley Publishing Ltd Expert Systems, September 2014, Vol. 31, No. 4 343 models are employed to learn the latent structure and the dynamics of sensor network data in order, efficiently to analyse data with the aim of learning and discovering what is happening in the monitored environment. Lastly, the paper ‘Ambient Assisted Living System with Capacitive Occupancy Sensor’ (Fernandez-Luque et al., 2014) introduces an Ambi- ent Assisted Living system that allows inferring a potential dangerous action of an elderly person living alone at home. This inference is obtained by a specific sensitisation with sensor nodes and a reasoning layer embedded in a personal computer that learns of the users’ behaviour patterns and advices when the current one differs significantly of the normal patterns. In this paper, a force-capacitive transducer-based sensor is implemented and tested. This sensor is based on electromechanical films transducer, which is able to detect force variations in a quasi-passive way. Acknowledgements We would like to thank Jon G. Hall, the editor-in-chief of Expert Systems for his interest and ongoing help for this special issue. This work is partially supported by the Spanish Ministerio de Ciencia e Innovación under project TIN2010- 20845-C03. References CASTANEDO, F. LÓPEZ-DE-IPIÑA, D. AGHAJAN, H. K. KLEIHORST, R. (2014) Learning routines over long-term sensor data using topic models, Expert Systems, Special issue on “Intelligent Monitoring for People Assistance and Safety”, 31, 365–377. CILLA, R. PATRICIO, M. A. BERLANGA, A. MOLINA, J. M. (2014) Human action recognition with sparse classification and multiple view learning, Expert Systems, Special issue on “Intelligent Monitoring for People Assistance and Safety”, 31, 354–364. FERNANDEZ-LUQUE, F. J. MARTÍNEZ, F. L. DOMENECH, G. ZAPATA, J. RUIZ R. (2014) Ambient assisted living system with capacitive occupancy sensor, Expert Systems, Special issue on “Intelligent Monitoring for People Assistance and Safety”, 31, 378–388. GASCUEÑA, J. M. FERNÁNDEZ-CABALLERO, A. (2011) On the use of agent technology in intelligent, multisensory and distributed surveillance, The Knowledge Engineering Review, 26, 191–208. HERNÁNDEZ, J. CABIDO, R. MONTEMAYOR, A. S. PANTRIGO, J. J. (2014) Human activity recognition based on kinematic features, Expert Systems, Special issue on “Intelligent Monitoring for People Assistance and Safety”, 31, 345–353. MICHELONI, C. REMAGNINO, P. ENG, H. L. GENG, J. (2010) Intelli- gent monitoring of complex environments, IEEE Intelligent Systems, 25, 12–14. RIVAS, A. MARTINEZ-TOMÁS, R. FERNÁNDEZ-CABALLERO, A. (2011) Multiagent system for knowledge-based event recognition and composition, Expert Systems, 28, 488–501. Rafael Martínez-Tomás Departmento de Inteligencia Artificial UNED Spain Antonio Fernández-Caballero Departmento de Sistemas Informáticos Universidad de Castilla-La Mancha Albacete Spain José Manuel Ferrández Department of Electronics and Computer Technology Universidad Politécnica de Cartagena Spain © 2013 Wiley Publishing Ltd344 Expert Systems, September 2014, Vol. 31, No. 4 
work_4i55a2fyfnerfnvoanvfber6by	----	Computer systems to facilitating organizational learning: IT and organizational context Computer systems to facilitating organizational learning: IT and organizational context Shih-Wei Chou* Department of Information Management, National Kaohsiung First University of Science and Technology, 2 Juoyue Road, Nantz District, Kaohsiung 811, Taiwan, ROC Abstract A comprehensive model that delineates the interrelationships among computer systems, organizational context and organizational learning is absent. This study aims to fill this void. Unlike previous research, this study investigates the role of computer systems, i.e. organizational learning computer systems (OLCS), in facilitating organizational learning. In our framework, we argued that contextual variables mediated the impact of OLCS on organizational learning. In order to test the feasibility of this framework, we conducted an empirical study. This study employed a survey instrument, which contained data collected from 500 organizations in manufacturing, service industry, and academic institutions. A total of 165 usable responses were analyzed. The results indicate that OLCS have a positive impact on the organizational learning processes. Both ‘problem characteristic’ and ‘organizational culture’ moderate the influence of OLCS on organizational learning. The implications of the study are provided, and future research is suggested. q 2002 Elsevier Science Ltd. All rights reserved. Keywords: Organizational learning computer systems; Organizational context; Organizational learning 1. Background A comprehensive framework concerning organiz- ational learning has been proposed by Huber (1991). This framework identified four constructs, which are crucial to the effectiveness of organizational learning; they are knowledge acquisition, information distribution, information interpretation, and organizational memory. Since Huber’s comprehensive view of organizational learning theory was published, at least four researchers have referenced this framework in conducting related studies (Goodman & Darr, 1998; Hernes, 1999; Nonaka & Takeuchi, 1995; Robey, Boudreau, & Rose, 2000). However, other researchers criticized that Huber’s frame- work did not examine the role of computer systems for acquiring knowledge and enhancing the effect of organizational learning. They have proved that computer systems are changing many organizational processes including communication (Kiesler & Sproull, 1987), group decision making (Kiesler, Siegel, & McGuire, 1984), coordination (Rice & Shook, 1990), and colla- borative work (Kraut, Galegher, Fish, & Chalfonte, 1992). Despite previously mentioned criticism, there are relatively few field studies that examine the effect of computer systems on facilitating organizational learning (Constant, Sproull, & Kiesler, 1996; Goodman & Darr, 1998; Orlikowski, 1993a). Orlikowski (1993b) argued that ‘organizational context’, such as corporate strategies and structure and culture, is one of the critical factors that influence the adoption and using of IT. A similar concept was presented in Orlikowski (1993a), which reveals that a number of organizational elements, such as mental models (which affect how people understand and appreciate IT) and structural properties (reward systems and workplace norms), significantly influence the implementation and usage of IT. Although several researchers (Dutton & Dukerich, 1991; Henderson & Clark, 1990; Orlikowski, 1993a,b) have investigated the impact of organizational context on the applicability of IT in an organization, there are relatively few empirical studies that examine the role of organizational context, which serves as a moderator between IT and organizational learning. The purposes of this study are: (a) to examine 0957-4174/03/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 7 - 4 1 7 4 ( 0 2 ) 0 0 1 5 5 - 0 Expert Systems with Applications 24 (2003) 273–280 www.elsevier.com/locate/eswa * Tel.: þ886-7601-1000x4114; fax: 886-7601-1042. E-mail address: swchou@ccms.nkfust.edu.tw (S.-W. Chou). http://www.elsevier.com/locate/eswa the role of computer system in facilitating organizational learning; (b) to realize the impact of organizational context on the effect of adopting OLCS to facilitate organizational learning. 1.1. Conceptual framework Many definitions of organizational learning have been proposed. For example, Argyris (1993) emphasized that organizational learning is a process of detecting and correcting errors. Similarly, Fiol and Lyles (1985) claimed that organizational learning contains the procedures to improve actions through knowledge acquisition and cre- ation. Huber (1991) argued that organization learns through its process of information. He proposed an organizational learning framework, which contained four constructs: knowledge acquisition, information distribution, infor- mation interpretation, and organizational memory. Good- man and Darr (1998) employed Huber’s framework and argued that organizational level learning occurred when the problem-solution exchanges and consequences were com- municated and known by other organizational members (broadcasting ). As Brown and Duguid’s (1991) research indicated, in order to have significant learning and innovation in the informal communities-of-practice in which they work, it is necessary to have a mechanism for organizations to share their interpretation about the problem-solving exchanges and to update the organizational memory about their experiences (updates ). Walsh and Ungson’s (1991) research indicated that there was some form of organizational memory storing problem-solution exchanges and consequences. The basic processes that contribute to the occurrence, breadth, and depth of organizational learning depend on organizational memory (memory ) (Huber, 1991). Since the system with the aforementioned functions (broadcasting, updates, and memory) may have great impact on the organizational learning, we call this type of system ‘organizational learning computer system (OLCS)’. The organizational context also plays a critical role in affecting technology and learning processes (Goodhue & Thompson, 1995; Tyre & Orlikowski, 1994). Contextual variables can be viewed as increasing or decreasing the effect of OLCS on organizational learning processes. A wide variety of organizational variables have been pro- posed, such as rewards systems, performance measure, problem-solution characteristics (Goodman & Darr, 1998), corporate strategies, structure, culture (Goodman & Darr, 1998; Orlikowski, 1993b), trust (Scott, 2000), and manage- ment style (Nonaka & Takeuchi, 1995). A culture of collaboration and mutual trust should facilitate the role of OLCS and organizational learning. However, organizations with a conservative culture and bureaucratic structure should substantially reduce the propensity to exchange problems and solutions. Further, the nature and complexity of the task structure will influence the form of problems and solutions to be exchanged. Multiple attributes are needed to describe a complex problem and the environmental conditions surrounding it. Also, for complex problems, there are many possible solutions and applicable rules to implement those specific solutions. Therefore, it is usually more difficult to formulate and deliver complex solutions for those contributors. On the other hand, when the problem statement has few attributes and there is only one solution with few implementation rules, problem-solution exchanges should be easy. Given the unclear functions of OLCS in facilitating organizational learning, and the unspecified effect of mediating organizational context on the impact of OLCS on organizational learning, we therefore conducted our research accordingly. 2. Research methodology In order to explore the impact of OLCS and organiz- ational context on organizational learning, we developed the research framework in Fig. 1. There are two research questions. (a) Does OLCS play a role in facilitating organizational learning? (b) What types of organizational context may moderate the effect of OLCS on facilitating organizational learning? The research design of this research contains the following issues. 1. Purpose. The purpose of this research is ‘hypotheses testing’. In order to do so, we conducted a study that contained both correlational and causal study. The causal study was conducted to verify the relationship between OLCS and the organizational learning process. The other question that we wanted to investigate was the effect of organizational context. In our research, organizational context served as a moderator. Since contextual variables can be viewed as increasing or decreasing the costs inherent in the decision to contribute or to adopt knowledge, we examined the effect of the moderator between OLCS and the organizational learning process. Fig. 1. Research framework. S.-W. Chou / Expert Systems with Applications 24 (2003) 273–280274 https://isiarticles.com/article/3911 
work_4iiwswkqhnbwvipp45zsv6kbka	----	Analysing of existing customers to improve acquiring of new customers Weak Signal Identification with Semantic Web Mining Dirk Thorleuchter a,* , Dirk Van den Poel b a Fraunhofer INT, D-53879 Euskirchen, Appelsgarten 2, Germany, dirk.thorleuchter@int.fraunhofer.de b Ghent University, Faculty of Economics and Business Administration, B-9000 Gent, Tweekerkenstraat 2, Belgium, dirk.vandenpoel@ugent.be URL: http://www.crm.UGent.be _________________ * Corresponding author at: Fraunhofer INT, Appelsgarten 2, 53879 Euskirchen, Germany. Tel.: +49 2251 18305; fax: +49 2251 18 38 305 E-mail address: Dirk.Thorleuchter@int.fraunhofer.de (D. Thorleuchter). Abstract We investigate an automated identification of weak signals according to Ansoff to improve strategic planning and technological forecasting. Literature shows that weak signals can be found in the organization’s environment and that they appear in different contexts. We use internet information to represent organization’s environment and we select these websites that are related to a given hypothesis. In contrast to related research, a methodology is provided that uses latent semantic indexing (LSI) for the identification of weak signals. This improves existing knowledge based approaches because LSI considers the aspects of meaning and thus, it is able to identify similar textual patterns in different contexts. A new weak signal maximization approach is introduced that replaces the commonly used prediction modeling approach in LSI. It enables to calculate the largest number of relevant weak signals represented by singular value decomposition (SVD) dimensions. A case study identifies and analyses weak signals to predict trends in the field of on-site medical oxygen production. This supports the planning of research and development (R&D) for a medical oxygen supplier. As a result, it is shown that the proposed methodology enables organizations to identify weak signals from the internet for a given hypothesis. This helps strategic planners to react ahead of time. Key Words: Weak Signal, Ansoff, Latent semantic indexing, SVD, Web Mining. mailto:dirk.vandenpoel@ugent.be Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 2 1 Introduction A successful planning of research and development (R&D) requires an overview on current and future environmental conditions (Choi, Kim, & Park, 2007) to predict the arising of new technological approaches - the technology push - (Thorleuchter, 2008) and to predict changes in consumers’ needs - the market pull - (Thorleuchter, Van den Poel, & Prinzie, 2010d) by time. Literature introduces a concept of environmental scanning (Abebe, Angriawan, & Tran, 2010; Tabatabei, 2011) that enables this prediction by extracting and analyzing information from the environment especially to identify events, trends, and relationships (Choo & Auster, 1993). The concept of environmental scanning realizes a predictive view by applying a weak signal approach (Ansoff, 1975). A weak signal is an event or a development where an accurate estimation of its impact on a target (e.g. on organization’s R&D) cannot be given because a single weak signal probably appears by chance (Ansoff, 1982). However, identifying several weak signals from different sources aiming at a common target is probably a hind that this target will be impacted in future. Thus, environmental changes can be predicted in advance that show future problem areas and opportunities. This enables the use of weak signals as early warning system for strategic planning. As shown by Decker, Wagner, and Scholz (2005), the internet is a valuable information source for an environmental scanning and thus, for detecting weak signals. A website or a document itself is normally not a weak signal however; it might be that a website or a document contains a textual pattern that represents a weak signal (Uskali, 2005). Thus, a full text access to information in the internet is necessary to identify these weak signals. Performance reasons based on the large number of internet websites enforces a (semi-) automated approach e.g. web mining rather than a human based manual scanning (Gericke et al., 2009; Tabatabei, 2011). Especially for R&D planning, information about three areas has to be considered (Thorleuchter, Van den Poel, & Prinzie, 2010c): the science for new technological aspects (technology push), the users for new product ideas (market pull), and the industry for new product development aspects (the link between technology and market). Technological Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 3 research results are described in articles published in scientific journals, in conference proceedings, and in various scientific document repositories. In recent years, access to the full text of these articles using the internet becomes much easier because of the increased number of open access journals and articles available today. Further, some publishers (e.g. Elsevier) offer open archives that enable a full text access to articles after a specific embargo period of time. Additionally, some publishers allow manuscript posting where accepted manuscripts can be posted on authors’ personal or institutional websites. The Google Books initiative enables full text access to selected pages of conference proceedings published in books. This shows that in contrast to several years ago, the full text access to a large number of scientific articles is available today using the internet (Thorleuchter, Van den Poel, & Prinzie, 2010a). Information about new product development can be found on companies’ websites and in business magazines. Today, many magazines publish articles on their websites and thus, a full text access on this information is also available. Patents as representative for both, scientific results and industrial products are also published with full text in the internet (Thorleuchter, Van den Poel, & Prinzie, 2010b). Information about new product ideas from the users can be found in internet forums, blogs, micro blogs, review sites etc. The full text access to this information using the internet is possible, too. Overall, the planning of R&D can be supported by an environmental scanning and weak signals detection using the full text information in the internet today. The proposed methodology uses semantic text classification combined with an automated web mining approach for environmental scanning and weak signals detection. This is in contrast to related research, where knowledge structure based text classification approaches are used (Yoon, 2012). The use of semantic text classification considers the fact, that weak signals are formulized by different persons, in different languages, and in different contexts. It might be that two textual patterns representing weak signals are related to a specific topic even if they do not share a common word. This relationship can only be identified with semantic approaches that consider aspects of meaning rather than aspects of words (Thorleuchter & Van den Poel, 2013b). Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 4 A further contrast to related research is the use of a new weak signal maximization approach. Existing literature that investigate latent semantic indexing as well known semantic approach apply prediction modeling approaches to calculate a performance optimized number of singular value decomposition (SVD) dimensions (Thorleuchter & Van den Poel, 2012e). They use training and test set that consists of a well-balanced number of positive and negative examples (Thorleuchter & Van den Poel, 2013a). The creation of a training and test set is not applicable to weak signal identification because weak signals for a specific topic occur low frequently. The number of positive examples for a specific topic is not sufficient to create a well-balanced training and test set. Further, an evaluation of weak signals’ impacts can only be done considering the collection of all weak signals. Thus, a new weak signal maximization approach is proposed to identify the maximal number of weak signals for a specific topic to enable such an evaluation. Up to now, the applied practical approaches for weak signal identification using a wide scope environmental scanning have failed. High tech companies in Europe had problems realizing a weak signal detection and evaluation because of the high manual effort caused by the lack of environmental scanning tools and the low quality of the results (Schwarz, 2005). Existing successful practical approaches for weak signal are restricted to a small number of documents e.g. 50 selected web pages (Decker et al., 2005) or financial news articles of one Finish newspaper (Uskali, 2005). Thus, the proposed semi-automated methodology bridges these gaps by implementing a web mining based environmental scanning and semantic weak signal identification. This enables a wide scope for environmental scanning, a low manual effort for human experts, and an improved identification performance. In a case study, the proposed methodology is applied in the field of on-site medical oxygen production. R&D planners have provided a hypothesis concerning future developments. The methodology identifies relevant weak signals that are related to the given hypothesis. The weak signals do not verify or falsify the hypothesis; however they show that the hypothesis is in accordance to current trends extracted from the internet. This supports R&D planners by their decision making process. Overall, a methodology is proposed that enables a practical use of the weak signal concept considering a wide scope of information from the internet, aspects of meaning, and Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 5 performance aspects to reduce the manual effort. Trends and developments can be identified in advance and they are a valuable source for R&D planners to support their decision making. 2 Background 2.1 Using Internet for R&D environmental scanning The internet contains a huge amount of information and literature shows that the added value of this information outperforms the added value gained from using traditional information sources (D’Haen, Van den Poel, & Thorleuchter, 2012). Organizations use the internet in different ways e.g. for collecting and analyzing information from organization’s customers (Alallak, 2010) and from competitive organizations (Teo & Choo, 2001) to advance organization’s strategic planning (Purandre, 2008). Web mining approaches support organizations by information collecting because they offer an automated possibility to scan the internet for relevant information on websites (Kosala & Blockeel, 2000; Kobayashi & Takeda, 2000). They apply automated filtering algorithms to reduce the large number of websites identified by use of search engines. This is necessary to overcome performance restrictions because many retrieved and filtered results represent non-relevant information and thus, low precision values in information retrieval are obtained. Further, many relevant documents are not retrieved by the internet search engine. This leads to low recall values. In recent years, information about the R&D environment (science, industry, and consumer) is available and accessible in the internet as shown in the introduction chapter. This opened an opportunity to use the internet for R&D environmental scanning today. 2.2 Weak signals identification for R&D The concept of weak signals has been introduced as early warning system to advance strategic planning (Ansoff, 1975; Tabatabei, 2011). It enables a timely identification of future events or developments that are relevant for a decision maker (Kuosa, 2010). Furthermore future events and developments are named topics. Literature introduces many different definitions of weak signals and most of them describe weak signals as unstructured Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 6 information with low content value (Mendonça et al., 2004). In a first stage, the weak signals reflect aspects of a threat or an opportunity. Then, their information content increases more and more e.g. they also describe the origin of a threat or an opportunity. Finally, weak signals become strong signals in a second stage and they indicate possible actions in future (Holopainen & Toivonen, 2012). Examples for weak or strong signals are articles in newspapers describing a specific topic, changes in sentiments of experts concerning this topic, and trends in the jurisdiction with impact on this topic (Mendonça, Cardoso, & Caraça, 2012). Strong signals point to a concrete topic that will occur with medium to high probability. A large number of strong signals for a specific topic can be found in the internet. This is because the topic is mentioned and discussed widely on several websites, in news articles, in internet blogs etc. Strong signals are not of interest for strategic planning because they occur too late for considering in strategic decision makings and thus, they do not provide a preview on environmental changes (Yoon, 2012). In contrast to the high frequent occurrence of strong signals concerning a specific topic, weak signals occur low frequently. Further, they can be used for strategic decision making because they occur early enough. For a specific topic, a small number of weak signals can be seen and it is hard to identify them from the large amount of information in the internet. This is the reason why many implementations of weak signal identification approaches fail in practice (Schwarz, 2005). Further, the occurrence of one weak signal is not sufficient for a predictive view, however; the occurrence of several weak signals that aim to the same topic might give a hind for future changes (Hiltunen, 2008). Thus, several weak signals have to be identified concerning the same topic and used for a strategic decision making (Ilmola & Kuusi, 2006; Rossel, 2009; Tabatabei, 2011). Strategic R&D decisions are a subsection of strategic decisions. Descriptions of R&D topics are characterized by the occurrence of domain specific technical words (characteristic terms). This is in contrast to the colloquial language where the meaning of a specific term is often not clearly defined. Texts describing R&D topics also consist of an above-chance frequent co-occurrence of technical terms. This means that a specific technical term occurs more frequently together with a further term than it would be expected by chance. Both characteristics, the occurrence of characteristic terms and the frequent co-occurrences enable an easier identification of weak signals than it could be realized by using colloquial texts. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 7 2.3 Latent semantic indexing for weak signals identification With text classification, texts can be assigned to different classes. The classes have to be pre-defined in advance. This is done manually by human experts or automated by use of a set of training examples and machine based learning (Ko & Seo, 2009; Lin & Hong, 2011; Sudhamathy & Jothi Venkateswaran, 2012; Finzen, Kintz, & Kaufmann, 2012). Knowledge structure approaches are commonly used as instance-based learning algorithms for classification. Examples are the k nearest neighbor classification, simple probabilistic algorithms (e.g. naïve Bayes), decision tree models (e.g. C4.5), and support vector machine algorithms (Buckinx, Moons, Van den Poel, & Wets, 2004; Lee & Wang, 2012; Shi & Setchi, 2012). These approaches are already used for identification of weak signals theoretically (Tabatabei, 2011). However, they do not consider semantic aspects of the information. This is important because several weak signals for a specific topic have to be identified that are normally formulized by different persons. Thus, considering aspect of meaning is important to identify related weak signals. Further, some of the knowledge structure approaches do not consider dependencies of terms. Weak signals in R&D are characterized by the above- chance frequent co-occurrence of technical terms thus, it is also important to consider these dependencies. In contrast to knowledge structure approaches, semantic approaches are better suited to consider aspect of meaning and to calculate term dependencies. The calculation of these semantic relationships between terms based on computational eigenvector techniques from algebra (Jiang, Berry, Donato, Ostrouchov, & Grady, 1999; Luo, Chen, & Xiong, 2011). Terms that occur together in a textual pattern are considered as well as terms that might be in this textual pattern (Thorleuchter & Van den Poel, 2012c; Thorleuchter & Van den Poel, 2012d). Semantic indexing is normally applied using LSI. Text patterns standing behind several documents from a document collection are identified (Park, Kim, Choi, & Kim, 2012). These text patterns also enable a clustering of the documents. They consist of a list of semantically related terms and the meaning expressed by the set of these terms is stated in different documents from the collection (Christidis, Mentzas, & Apostolou, 2012; Tsai, 2012). Further, the impact of each document on each text pattern is calculated. This considers well Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 8 term dependencies (Thorleuchter, Van den Poel, & Prinzie, 2012). Thus, LSI they can be used to identify weak signals. Many modern approaches with better theoretical foundation and better performance than LSI have been introduced and applied in literature e.g. ‘Probabilistic Latent Semantic Indexing’ (Hofmann, 1999), ‘Latent Dirichlet Allocation’ (Blei, Ng, & Jordan, 2003; Ramirez, Brena, Magatti, & Stella, 2012), and ‘Non-Negative Matrix Factorization’ (Lee & Seung, 1999; Lee & Seung, 2001). In contrast to LSI, the improved approaches are of higher computational complexity than LSI. Applying them to a very large-scale document collection retrieved from the internet is difficult. Thus, LSI is used in the proposed approach to show the feasibility of this approach - in the knowledge that the performance can be improved by using the modern approaches instead of LSI. A very new and also very interesting approach is proposed by Ramírez and Brena (2012). Their query based topic modeling approach allows analyzing very large-scale collections with at least similar or even better performance than the above mentioned modern approaches by reducing computational complexity. As the case study (see Sect. 4) was already finished before we have been aware of this new approach, the use of this new approach might be interesting as an avenue of future research. As mentioned in Sect. 2.2, weak signals occur with low frequency that means they occur on a few numbers of websites in the internet. Using weak signals as classes for text classification fails because the classes cannot be defined in advance by human experts. Further, a machine based learning approach cannot be performed because it could not be guaranteed that the number of positive training examples is above a specific threshold to ensure statistical significance for the classification results. This excludes the use of knowledge structure or semantic classification approaches. To overcome these limitations, the classes have to be defined by a semantic clustering approach where the identified semantic textual patterns are evaluated to identify weak and strong signals in a textual collection. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 9 3 Methodology Fig. 1 shows the processing of the proposed methodology in different steps. The proposed methodology in Fig. 1 identifies semantic textual patterns from the internet and it analyzes them to identify weak and strong signals. It applies an environmental search as described in Sect. 2.1 by using the internet. It considers the characteristics of weak signals as described in Sec. 2.2 by applying a semantic clustering approach (see Sect. 2.3). The methodology starts with a hypothesis where an existing strategic decision problem is described. The weak signals are identified based on the topic described in this hypothesis. Thus, it is important that the hypothesis is formulized clearly and comprehensibly. The words that are used to formulize the hypothesis are considered for the next step, the creation of search queries. It is important that the search queries are created by human experts in high quality that many relevant documents are retrieved and that non-relevant documents are not Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 10 considered. The full text of the retrieved documents is crawled. Terms in the full text are compared to terms in the hypothesis to identify relevant sections within each document. These sections are used for further processing while the other sections are discarded. LSI is applied on the data and it creates a number of k different semantic textual patterns. They represent semantic aspects that occur on several different documents retrieved from the internet and thus, they can be used to represent strong and weak signals (see Sect. 2.2). The selection of k is based on a new weak signal maximization approach that leans on the weak signal characteristics from Sect. 2.2. This ensures that k is large enough to consider all weak signals and that k is small enough to discard semantic textual patterns that are not in accordance to the weak signal characteristics. The semantic textual patterns are split in weak signals and in strong signals. The relevance of each weak signal is analyzed manually concerning its impact on relevant terms. The development of a weak signal is calculated. As a result, the developments of weak signals are presented to the decision maker to improve strategic decision making. 3.1 Web Mining and Text Mining Search queries are created manually to represent the hypothesis. A single search query is often not suited to cover a topic. Thus, several search queries have to be created to cover all different aspects of the hypothesis. To ensure an environmental scanning, it has to be considered that websites are written in different languages. Thus, created search queries for a specific language have to be translated in different languages. This should also be done manually by human experts to ensure a higher quality than a search engine can offer by automated translation. The next steps are processed automatically. Each query is executed by an internet search engine and the search is restricted to the corresponding language from the query. The query results (the URLs of the websites) are used with a crawler that extracts the full text from all URLs. The retrieved results are stored as documents (one document per URL) in folders separated by the languages of the corresponding websites (Thorleuchter & Van den Poel, 2011b). Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 11 The full text from the retrieved documents are preprocessed and filtered to reduce complexity. In a first step, the raw test is cleaned from existing scripting code, images, and html-tags. Specific characters and punctuation are eliminated and typographical errors are corrected by using a dictionary from the corresponding language. Single words (terms) are identified by tokenization and case conversion is applied. In a second step, filtering methods are applied to reduce the number of terms for further processing. This includes stop word filtering (filtering of non-informative terms), via part-of-speech tagging (filtering specific syntactic category) up to stemming (reducing the number of terms with the same stem). Lemmatizing is not applied because existing practical methods are still error prone. The number of terms is reduced further by applying Zipf distribution (Zipf, 1949; Zeng et al., 2012). The preprocessed full text from the retrieved documents consists of texts in different lengths from several bytes up to several megabytes that depend on the corresponding websites. The text normally is split in several sentences. However, a website could reflect several different topics. Then probably the text in one sentence is relevant for the topic described in the hypothesis and text in other sections or sentences is not. Tokenization is applied on the full text of each retrieved document and the term unit is used as a sentence. This means the text is split in its different sentences and a text similarity measure is used to compare the terms from each sentence to the terms from the description of the hypothesis. For this, term vectors in vector space model have to be build and Jaccard’s coefficient can be used because it considers well the different lengths of term vectors (Thorleuchter & Van den Poel, 2011a). A similarity result value above a specific threshold shows that the corresponding sentence is relevant for the hypothesis and can be used for further processing. The other sentences are deleted. This reduces the lengths of the document and it ensures that the information used for latent semantic indexing is relevant. The documents are written in different languages. However, the processing of latent semantic indexing requires that documents are written is the same language. The translation of documents to a target language can be done automatically e.g. by use of Google translate application programming interface (API). It offers an automated translation of a document collection. The quality of the automated translations is low compared to the quality from a manual translation of a human expert. However, a manual translation of each document Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 12 leads to high efforts because of the large amount of documents. Further, a high quality translation is not necessary because the translated text is transformed to term vectors in the next step. Thus, grammar aspects are not of interest and it is sufficient to translate relevant terms one-to-one. This can be done with automated approaches in good quality, too. Term vectors in vector space model are created from the collection of all translated documents. For the components of the vectors, weighted frequencies are used. This is because literature shows that they outperform raw frequencies (Prinzie, & Van den Poel, 2007; Prinzie, & Van den Poel, 2006; Van den Poel, De Schamphelaere, & Wets, 2004). This is because weighted frequencies show the impact of a term on the collection of all documents (Sparck Jones, 1973). Large weights are assigned to terms that occur frequently in a very small number of documents and that do not or seldom occur in most of the documents from the collection (Salton & Buckley, 1988). The well-known terms weighting scheme proposed by Salton, Allan, & Buckley (1994) is used to calculate the weight wi,j for term i and for a document j by      m 1p 2 i 2 pji iji ji p dfntf dfntf w ))/(log( )/log( , , , (1) In formula 1, the number of documents in the collection is n, the number of components from the term vectors is m, and the number of documents that contain term i is dfi. Further, inverse document frequency as represented by log(n/dfi) and term frequency (tfi,j) are used (Chen, Chiu, & Chang, 2005). The divisor is a length normalization factor that considers different lengths of the documents. 3.2 Latent semantic indexing The created vectors of weighted frequencies are composed to build a term-by-document matrix A with rank r (r ≤ min(m,n)). The large number of components from the term vectors leads to a large dimensionality of the matrix. Further, many terms only occur in a small number of documents and thus, the corresponding term vector component in many documents is set to zero. In total, this leads to term vectors with many zero values in the components and also to a term-by-document matrix that also consists of many zero values. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 13 This fact is used by singular value decomposition to reduce the dimensionality of the matrix because it can be expected that the rank of the matrix is lower than its dimensionality. Singular value decomposition is a commonly used matrix factorization technique that is applied within LSI. A reduced matrix dimensionality leads to a summarizing of terms concerning aspects of meaning (Deerwester et al., 1990). Aspects of meaning are calculated based on the co-occurrences of terms in the documents. This enables to group semantically related terms together with high discriminatory power to other groups. The groups of terms represent semantic textual patterns (Thorleuchter & Van den Poel, 2012b). Singular value decomposition calculates singular values for each group and thus, for each semantic textual pattern. The singular values are sorted in descending order in a diagonal (r x r) matrix Σ. Singular value decomposition calculates two further matrices, too. The (m x r) matrix U shows the impact of terms on the semantic textual patterns and the (n x r) matrix V shows the impact of the documents on the semantic textual patterns. The calculation is shown in formula 2. A = U Σ V t (2) Matrix U and matrix V are used for calculating the weak signal maximization approach as introduced in Sect. 3.3. 3.3 Weak signal maximization approach for latent semantic indexing Literature shows how to reduce the dimensionality of the matrix from r to k with singular value decomposition (Chen, Chu, & Chen, 2010). The data is split in a test and training set both containing a specific percentage of positive examples (Thorleuchter, Herberz, & Van den Poel, 2012). The training set is used to build a LSI-subspace (Zhong & Li, 2010) and the test set is projected into this subspace to calculate the predictive performance. The performance on each rank-k model is measured using the area under the receiver operating characteristics curve, logistic regression, and n-fold cross validation (DeLong, DeLong, & Clarke-Pearson, 1988; Hanley & McNeil, 1982; Halpern et al., 1996; Migueis, Van den Poel, Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 14 Camanho, & Cunha, 2012; Van Erkel & Pattynama, 1998). As a result, an optimal value of k based on computational complexity and on predictive performance is selected. Existing approaches from literature cannot be used to identify an optimized number of semantic textual patterns representing weak signals from the internet. This is because weak signals occur low frequently as shown in Sect. 2.2. Thus, the percentage of positive examples in a randomly selected training and test set is very low and to obtain significant results, the training and test set have to be very large. The sets contain unseen documents retrieved by an internet search. For training and testing, these documents have to be evaluated concerning the occurrence of weak signals by human experts manually. This causes an unmanageable high manual effort. Our proposed weak signal maximization approach identifies the value of k where the number of weak signals represented by low frequently occurred semantic textual patterns with a strong relationship to the given hypothesis is maximized. Singular value decomposition calculates k semantic textual patterns from the collection of all retrieved internet documents. It is characteristic for this processing that a small number of k leads to a small number of semantic textual patterns that are impacted by a large number of documents. These patterns occur frequently in the collection of all documents and thus, they are not weak signals. Thus, using singular value decomposition with a very small number of k does not lead to the identification of any weak signals. A very large number of k leads to a small number of patterns impacted by a large number of documents and to a very large number of patterns impacted by a very small number of documents. As shown in Sect. 2.2., weak signals should occur at least more than once or twice otherwise they probably occur by chance. To identify a weak signal, the number of documents with impact on the pattern should be above a specific threshold. Thus, a very large number of k also leads to the identification of none weak signals in total. Weak signals are not only defined by the number of impacted documents however, they should be related to the given hypothesis, too. Thus, relevant terms in a semantic textual pattern as defined by a term impact on the pattern above a specific threshold are compared Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 15 to relevant terms from the given hypothesis by using text similarity measures. A similarity value above a specific threshold shows that a low frequent semantic textual pattern is also related to the given hypothesis and thus, it can be defined as weak signal. Patterns that are impacted by a large number of documents also are impacted by the large number of terms from the different documents. Thus, comparing these terms to the terms from the given hypothesis normally leads to a low similarity value. Furthermore, patterns that are impacted by a very small number of documents e.g. one or two are only impacted by a very small number of terms. Here, a low similarity value is also obtained by comparing them to terms from the given hypothesis. Thus, a maximal number of weak signals can be identified if k is not too small and not too large. Several rank-k models are created and the number of weak signals is calculated for each model. Comparing is done by use of a similarity measure e.g. Jaccard’s coefficient that considers well different lengths of input vectors because the term vectors created from the hypothesis normally is from different size than the term vectors created from the semantic textual patterns. 4 Case Study The case study applies the proposed methodology to support R&D planners of a company that offers medical oxygen to hospitals. This is normally done in two different ways: off- and on-site production. Off-site production means that the medical oxygen is produced by oxygen generators in the company, stored in high pressure gas cylinders, and transported to the hospitals. On-site production means that the company sells oxygen generators to hospitals and the oxygen is produced in the hospitals. The production of medical oxygen is based on two different methods that separate air in his components. The cryogenic air separation uses a low temperature rectification principle based on the fact that gases have different temperatures for changing aggregation states. Applying this method can produce a large quantity of oxygen with high oxygen purity of more than 99% that also can be used to obtain liquid oxygen. The non-cryogenic air separation uses the pressure swing adsorption (PSA) principle. It produces oxygen with a purity of about 93%. The R&D department of the http://www.dict.cc/englisch-deutsch/aggregation.html http://www.dict.cc/englisch-deutsch/states.html Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 16 company is responsible for the technical improvement of the oxygen generators (for both methods and for both, on- and off-site production). Medical oxygen for hospitals in Europe has to meet the requirements of the European Pharmacopoeia where the oxygen purity is an important point. In the past, a hospital management had to buy oxygen at a purity of at least 99% in Europe and in many non- European states. This excludes the use of pressure swing adsorption principle for on-site oxygen production. Since mid-2011, the European Pharmacopoeia has changed the requirements for European states. After transposing this to national law, it allows an oxygen purity of 93% for hospitals if the used oxygen generators are certificated according to ISO 13485:2003. Based on this legislation amendment, the R&D planners have stated a hypothesis: The use of PSA for medical oxygen on-site production (93% purity) in Europe will increase in future. This increase will be equally distributed in European states. New companies especially from the domain of machinery and plant engineering will become suppliers by offering PSA oxygen generators for hospitals in future. The aim of the case study is to identify weak signals that are in accordance or that are not in accordance to the hypothesis. 4.1 Web Mining and Text Mining Based on the given hypothesis, ten search queries are created in English language. Examples are ‘Medical +oxygen +high +purity’, ‘Oxygen +pressure +swing +adsorption +PSA’, and ‘Ultra +high +purity +oxygen +on +site +generation’. They describe the area of high purity medical oxygen and they enable the identification of all internet documents dealing with this topic. The search queries are translated in different languages: German (GE), French (FR), Polish (PO), Czech (CZ), and Romanian (RO) to cover different regions in Europe. The queries are executed automatically using Google API in mid-2012 one year after the European Pharmacopoeia has changed the percentage because the process of transposing Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 17 this to national law is time consuming. The hyperlinks of all query results are stored separately into the different languages. For each retrieved website, a crawler is used to extract the textual information and to store it in a document. As a result, 14.792 plain text documents are created automatically with a size of 213 megabytes in total. Text mining methods are applied as mentioned in Sect. 3.1. This identifies several non-relevant documents and it reduces documents’ sizes of relevant documents. Overall, 8375 documents with a size of 40 megabytes are obtained. 4.2 LSI with weak signal maximization Based on the data collected in Sect. 4.1, LSI is applied together with the proposed weak signal maximization approach. The thresholds of the approach are determined manually in a two-step process by human experts. Starting values are determined in a first step and the results from several rank-k models are evaluated in the second step. Based on this evaluation, some parameter values are adapted (first step) and several rank-k models based on the adapted values are created and evaluated again (second step). This two-step process repeats until the parameter values are optimized. As a result, the threshold for the impact of a document on a textual pattern as depicted in matrix V from -1 to 1 is determined to 0.4. The percentage of documents with impact greater than or equal to this threshold on a textual pattern to the number of documents in total is also determined to identify weak and strong signals: A value of 0% to 2% is representative for very low frequently occurred semantic textual pattern that might be occurred by change and can be discarded. Weak signals are identified from 2% on to 6% and strong signals are identified from 6% on to 10%. More than 10% means that the content of a semantic textual pattern can be found on more than on every tenth website. This information is normally well- known and thus, not relevant for forecasting. Matrix U shows the impact of a term on a semantic textual pattern from -1 to 1. The threshold for identifying relevant terms is determined to 0.4. Thus, a term vector for each semantic textual pattern is built on terms with an impact greater than or equal to 0.4. The term vectors are compared to the term vector from the hypothesis where the threshold for the result value is determined to 0.3 to identify patterns that are related to the hypothesis. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 18 4.3 Results An optimal value of k is identified with k=16. Thus, 16 semantic textual patterns are identified. Three patterns can be identified that are impacted by more than 10% of all documents. Eight patterns are identified where the percentage of the impacted documents is between 0% and 2%. Further, two strong signals and three weak signals have been identified. Comparing them to the hypothesis shows that the strong signals are not related to the hypothesis. This is because they deal about cryogenic air separation techniques and the transportation of high pressure gas cylinders. Two of the three weak signals are related to the hypothesis. The first weak signal with 4.2% impacted documents describes various aspects of PSA based oxygen generators producing 93% purity. This is in accordance to the first sentence of the hypothesis. The second weak signal based on a small number of impacted documents (2.3%) and it shows a new technological development that enables PSA based oxygen generators to produce medical oxygen with 99% purity in a multi-stage process. The increase of purity can be realized with low additional costs. Oxygen generators based on this technique can be used in Europe as well as in many non-European states and they are independent of future legislation amendments. This weak signal is not in accordance to the first sentence of the hypothesis because in future, it might be that 99% purity PSA generators are increasingly used for on-site medical oxygen production. Table 1: Number of documents with impact on the first and second weak signal in different languages compared to number of documents in total Weak signal EN CZ FR GE PL RO ∑ 1 Number of documents 182 9 66 68 3 24 352 Percentage 52% 2% 19% 19% 1% 7% 100% 2 Number of 118 3 47 13 1 11 193 Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 19 documents Percentage 61% 2% 24% 7% 0% 6% 100% Number of documents in total 3449 514 1133 1583 582 1114 8375 Percentage 41% 6% 14% 19% 7% 13% 100% The number of documents with impact on the first and second weak signal is sorted by language (see Table 1) e.g. 182 documents in English language (EN)) have an impact on the first weak signal. Further 52% of all documents with impact on the first weak signal are English language documents. They are compared to the distribution of all 8375 documents e.g. 41% of all documents are English language documents. Table 1 shows that more English and French websites are related to the first weak signal than it would be expected (increase from 41% to 52% and from 14% to 19%). In contrast to this, Romanian websites do not mention the weak signal as often as it would be expected (decrease from 13% to 7%). Table 1 also shows that on Czech and Polish websites the 93% purity PSA medical oxygen on-site production is seldom mentioned related to the number of Czech and Polish websites in total. Further, it shows that information about on-site medical oxygen generators with 99% purity often can be found on English and French websites. This is not in accordance to the second sentence of the hypothesis because the topic is not equally distributed in European states. The URLs of the documents with impact on the first weak signal are also evaluated concerning companies’ websites. As a result, the following list of companies is identified: ‘Linde, Air Liquide, Praxair, Messergroup, Pangas, Westfalen-ag, Basigas, DA- Energietechnik, Iga-gas, Airco-Druckluft, Cryotec, aircom24, Airtexx, IGS, Oxymat, Oxyplus, Oxair’. This list contains established medical oxygen suppliers as well as companies in the domain of machinery and plant engineering. This is in accordance to the third sentence of the hypothesis. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 20 4.4 Evaluation In the case study, the proposed approach identifies two weak signals each represented by a semantic textual pattern. The identification is based on the assignment of retrieved internet documents to the two corresponding semantic textual patterns by LSI / WSM. The performance of this assignment is evaluated based on precision and recall. For each of the two semantic textual patterns, the number of documents that are correctly assigned to the pattern is the true positive (TP). The number of documents incorrectly assigned to the pattern is the false positive (FP) and the number of documents that are incorrectly not assigned to the pattern (missing documents) is the false negative (FN). Precision is defined as TP / (TP + FP) and recall is defined as TP / (TP + FN). The evaluation is processed for documents in English and German language because the number of these documents is sufficient for a statistical evaluation. Table 2: Precision and recall for the assignment of English and German documents to the semantic textual patterns standing behind the first and second weak signal Language Number of Documents First / Second Weak signal TP FP FN Precision Recall EN 3449 First 153 29 69 84% 69% EN 3449 Second 96 22 33 81% 74% GE 1583 First 48 20 26 71% 65% GE 1583 Second 9 4 3 69% 75% Table 2 shows that the average value is 76% for precision at 70% recall. This outperforms the average value of the frequent baseline (3% precision at 70% recall). Further, the precision and recall values for German documents are smaller than that of English documents. An explanation for this is that the German documents are translated to the English language automatically. This reduces quality of the translated documents and thus, this also reduces the corresponding precision and recall values. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 21 5 Conclusion This work proposes a new methodology that enables the automated identification of weak signals for strategic forecasting. Weak signals are extracted from an organization’s environment as represented by internet information. Based on a given hypothesis about the future, related websites are identified and textual information is extracted. LSI with a new weak signal maximization approach is applied on this textual information to identify weak signals. The identified weak signals describe trends and future developments and it is analyzed whether they are in accordance to the given hypothesis or not. The websites standing behind an identified weak signal can be analyzed concerning language aspects to identify the spacial distribution of this weak signal. Further the website address also can be used to identify relevant organizations or companies related to the weak signal. This enables strategic planners to identify new trends, the spatial distribution of these trends, and the corresponding players (e.g. competitive organizations) ahead of time. Future work should focus on the optimization of the parameters by applying a parameter selection procedure. This enables an improved weak signal maximization approach. Further, the development of weak signals over time can be investigated by this methodology. For this, web mining has to retrieve the data at different points of time. Then, one probably could see that new weak signals occur, that existing weak signals disappear, or that existing weak signals become strong signals. Literature Abebe, M., Angriawan, A., & Tran, H. (2010). Chief executive external network ties and environmental scanning activities: An empirical examination. Strategic Management Review, 4(1), 30-43. Alallak, B. (2010). Evaluating the adoption and use of Internet-based marketing Information systems to improve marketing intelligence. International Journal of Marketing Studies, 2(2), 87-101. Ansoff, I. H. (1975). Managing strategic surprise by response to weak signals. California Management Review, 18(2), 21-33. Ansoff, I. H. (1984). Implanting strategic management. New Jersey: Prentice Hall Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet allocation. Journal of Machine Learning Research, 3(4–5), 993-1022. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 22 Buckinx, W., Moons, E., Van den Poel, D., & Wets, G.(2004). Customer-adapted coupon targeting using feature selection. Expert Systems with Applications, 26(4), 509-518. Chen, M. C., Chiu, A. L., & Chang, H. H. (2005). Mining changes in customer behavior in retail marketing. Expert System with Applications, 28(4), 773-781. Chen, M.-Y., Chu, H.-C., & Chen, Y.-M. (2010). Developing a semantic-enable information retrieval mechanism. Expert Systems with Applications, 37(1), 322-340. Choi, C., Kim, S., & Park, Y. (2007). A patent-based cross impact analysis for quantitative estimation of technological impact: the case of information and communication technology. Technological Forecasting and Social Change, 74(2007), 1296-1314. Choo, C. W., & Auster, E. (1993). Environmental scanning: Acquisition and use of information by managers. Annual Review of Information Science and Technology, 28, 279-314. Christidis, K., Mentzas, G., & Apostolou, D. (2012). Using latent topics to enhance search and recommendation in Enterprise Social Software. Expert Systems with Applications, 39(10), 9297-9307. Decker, R., Wagner, R., & Scholz, S. W. (2005). An internet-based approach to environmental scanning in marketing planning. Marketing Intelligence & Planning, 23(2), 189-200. Deerwester, S., Dumais, S., Furnas, G., Landauer, T., & Harshman, R. (1990). Indexing by latent semantic analysis. Journal of the American Society for Information Science, 41(6), 391-407. DeLong, E. R., DeLong, D. M., & Clarke-Pearson, D. L. (1988). Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics, 44(3), 837-845. D’Haen, J., Van den Poel, D., & Thorleuchter, D. (2013). Predicting customer profitability during acquisition: Finding the optimal combination of data source and data mining technique. Expert Systems with Applications, doi: 10.1016/j.eswa.2012.10.023. Finzen, J., Kintz, M., & Kaufmann, S. (2012). Aggregating web–based ideation platforms. International Journal of Technology Intelligence and Planning, 8(1), 32-46. Gericke, W., Thorleuchter, D., Weck, G., Reiländer, F., & Loß, D. (2009). Vertrauliche Verarbeitung staatlich eingestufter Information - die Informationstechnologie im Geheimschutz. Informatik Spektrum, 32(2), 102-109. Halpern, E. J., Albert, M., Krieger, A. M., Metz, C. E., & Maidment, A. D. (1996). Comparison of receiver operating characteristic curves on the basis of optimal operating points. Academic Radiology, 3(3), 245-253. Hanley, J.A., & McNeil, B.J. (1982). The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology, 143(1), 29-36. Hiltunen, E. (2008). The future sign and its three dimensions. Futures, 40(3), 247-260. Hofmann, T. (1999). Probabilistic Latent Semantic Indexing. In: Proceedings of the Twenty-Second Annual International SIGIR Conference on Research and Development in Information Retrieval (SIGIR-99). Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 23 Holopainen, M., & Toivonen, M. (2012). Weak signals: Ansoff today. Futures, 44(3), 198-205. Ilmola, L., & Kuusi, O. (2006). Filters of weak signals hinder foresight: Monitoring weak signals efficiently in corporate decision-making. Futures, 38(8), 908-924. Jiang, J., Berry, M. W., Donato, J. M., Ostrouchov, G., & Grady, N. W. (1999). Mining consumer product data via latent semantic indexing. Intelligent Data Analysis, 3(5), 377-398. Ko, Y., & Seo, J. (2009). Text classification from unlabeled documents with bootstrapping and feature projection techniques. Information Processing & Management, 45(1), 70-83. Kobayashi, M., & Takeda, K. (2000). Information retrieval on the web. Association for Computing Machinery, 32(2), 144. Kosala, R., & Blockeel, H. (2000). Web research: A survey. ACM SIGKDD Explorations Newsletter, 2(1). Kuosa, T. (2010). Futures signals sense-making framework (FSSF): A start-up tool to analyse and categorise weak signals, wild cards, drivers, trends, and other types of information. Futures, 42(1), 42- 48. Lee, C.H., & Wang S.H. (2012). An information fusion approach to integrate image annotation and text mining methods for geographic knowledge discovery. Expert Systems with Applications, 39(10), 8954- 8967. Lee, D. D., & Seung, H. S. (1999). Learning the parts of objects by non-negative matrix factorization. Nature 401(6755), 788-791. Lee, D. D., & Seung, H. S. (2001). Algorithms for Non-negative Matrix Factorization. In: Advances in Neural Information Processing Systems 13. Proceedings of the 2000 Conference. MIT Press. pp. 556- 562. Lin, M-H., & Hong, C-F. (2011). Opportunities for Crossing the Chasm between Early Adopters and the Early Majority through New Uses of Innovative Products. The Review of Socionetwork Strategies, 5(2), 27-42. Luo, Q., Chen, E., & Xiong, H. (2011). A semantic term weighting scheme for text categorization. Expert Systems with Applications, 38(10), 12708-12716. Mendonça, S., Cardoso, G., & Caraça, J. (2012). The strategic strength of weak signal analysis. Futures, 44(3), 218-228. Mendonça, S., Pina e Cunha, M., Kaivo-oja, J., & Ruff, F. (2004). Wild cards, weak signals and organisational improvisation. Futures, 36(2), 201-218. Migueis, V. L., Van den Poel, D., Camanho, A. S., & Cunha, J.F. (2012). Modeling partial customer churn: On the value of first product-category purchase sequences. Expert Systems with Applications, 39(12), 11250-11256. Park, D, H., Kim, H. K., Choi, I. Y., & Kim, J. K. (2012). A literature review and classification of recommender systems research. Expert Systems with Applications, 39(11), 10059-10072. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 24 Prinzie, A., & Van den Poel, D. (2006). Investigating purchasing-sequence patterns for financial services using Markov, MTD and MTDg models. European Journal of Operational Research, 170(3), 710-734. Prinzie, A., & Van den Poel, D. (2007). Predicting home-appliance acquisition sequences: Markov/Markov for Discrimination and survival analysis for modeling sequential information in NPTB models. Decision Support Systems, 44(1), 28-45. Purandre, P. (2008). Web mining: A key to improve business on web. IADIS European Conference Data Mining, (pp. 155-159). Ramírez, E. H., & Brena, R. F. (2012). Query Based Topic Modeling: An Information-Theoretic Framework for Semantic Analysis in Large-Scale Collections. In: Quantitative Semantics and Soft Computing Methods for the Web: Perspectives and Applications (pp. 69-95), Information Science Pub., USA. Ramirez, E. H., Brena, R. F., Magatti, D., & Stella, F. (2012). Topic model validation. Neurocomputing 76(1), 125-133. Rossel, P. (2009). Weak signals as a flexible framing space for enhanced management and decision- making. Technology Analysis & Strategic Management, 21(3), 291-305. Salton, G., Allan, J., & Buckley, C. (1994). Automatic structuring and retrieval of large text files. Communications of the ACM, 37(2), 97-108. Salton, G., & Buckley, C. (1988). Term-weighting approaches in automatic text retrieval. Information Processing & Management, 24(5), 513–523. Schwarz, J. O. (2005). Pitfalls in implementing a strategic early warning system. Future Studies, 7(4), 22-31. Shi, L., & Setchi, R. (2012). User-oriented ontology-based clustering of stored memories. Expert Systems with Applications, 39(10), 9730-9742. Sparck Jones, K. (1973). Index term weighting. Information Storage and Retrieval, 9(11), 619-633. Sudhamathy, G. & Jothi Venkateswaran, C. (2012). Fuzzy Temporal Clustering Approach for E- Commerce Websites. International Journal of Engineering and Technology, 4(3), 119-132. Tabatabei, N. (2011). Detecting Weak Signals by Internet-Based Environmental Scanning. Master Thesis, Waterloo University: Waterloo. Teo, T. S., & Choo, W. Y. (2001). Assessing the impact of using Internet for competitive intelligence. Information & Management, 39(1), 67-83. Tsai, H.H. (2012). Global data mining: An empirical study of current trends, future forecasts and technology diffusions. Expert Systems with Applications, 39(9), 8172-8181. Thorleuchter, D. (2008). Finding Technological Ideas and Inventions with Text Mining and Technique Philosophy. In C. Preisach, H. Burkhardt, L. Schmidt-Thieme, & R. Decker (Eds.), Data Analysis, Machine Learning, and Applications (pp. 413-420). Berlin: Springer. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 25 Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2010a). Mining Ideas from Textual Information. Expert Systems with Applications, 37(10), 7182-7188. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2010b). A compared R&D-based and patent-based cross impact analysis for identifying relationships between technologies. Technological Forecasting and Social Change, 77(7), 1037-1050. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2010c). Mining innovative ideas to support new product research and development. In H. Locarek-Junge, & C. Weihs (Eds.), Classification as a Tool for Research (pp. 587-594). Berlin: Springer. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2010d). Extracting consumers needs for new products - A web mining approach. In Proceedings WKDD 2010 (p. 441.). Los Alamitos: IEEE Computer Society. Thorleuchter, D., & Van den Poel, D. (2011a). Semantic Technology Classification - A Defence and Security Case Study. In Proc. Uncertainty Reasoning and Knowledge Engineering (pp. 36-39). New York: IEEE. Thorleuchter, D., & Van den Poel, D. (2011b). Companies Website Optimising concerning Consumer’s searching for new Products. In Proc. Uncertainty Reasoning and Knowledge Engineering (pp. 40-43). New York: IEEE. Thorleuchter, D., & Van den Poel, D. (2011c). High Granular Multi-Level-Security Model for Improved Usability. In: System Science, Engineering Design and Manufacturing Informatization 1 (pp. 191-194). New York: IEEE. Thorleuchter, D., Herberz, S., & Van den Poel, D. (2012). Mining Social Behavior Ideas of Przewalski Horses. Lecture Notes in Electrical Engineering, 121, 649-656. Thorleuchter, D., Schulze, J., & Van den Poel, D. (2012). Improved Emergency Management by Loosely Coupled Logistic System. Communications in Computer and Information Science, 318, 5-8. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2012). Analyzing existing customers’ websites to improve the customer acquisition process as well as the profitability prediction in B-to-B marketing. Expert Systems with Applications, 39(3), 2597-2605. Thorleuchter, D., & Van den Poel, D. (2012a). Extraction of Ideas from Microsystems Technology. Advances in Intelligent and Soft Computing, 168, 563-568. Thorleuchter, D., & Van den Poel, D. (2012b). Predicting e-commerce company success by mining the text of its publicly-accessible website. Expert Systems with Applications, 39(17), 13026-13034. Thorleuchter, D., & Van den Poel, D. (2012c). Using NMF for Analyzing War Logs. Communications in Computer and Information Science, 318, 73-76. Thorleuchter, D., & Van den Poel, D. (2012d). Using Webcrawling of Publicly-Available Websites to Assess E-Commerce Relationships. In SRII Global Conference 2012 (pp. 402-410). San Jose, CA, USA: IEEE. Weak Signal Identification with Semantic Web Mining _______________________________________________________________________ 26 Thorleuchter, D., & Van den Poel, D. (2012e). Improved Multilevel Security with Latent Semantic Indexing. Expert Systems with Applications, 39(18), 13462-13471. Thorleuchter, D., Weck, G., & Van den Poel, D. (2012a). Granular Deleting in Multi Level Security Models - an Electronic Engineering approach. Lecture Notes in Electrical Engineering, 1, 177, 609- 614. Thorleuchter, D., Weck, G., & Van den Poel, D. (2012b). Usability based Modeling for Advanced IT- Security - an Electronic Engineering approach. Lecture Notes in Electrical Engineering, 1, 177, 615- 619. Thorleuchter, D., & Van den Poel, D. (2013a). Technology classification with latent semantic indexing. Expert Systems with Applications, 40(5), 1786-1795. Thorleuchter, D., & Van den Poel, D. (2013b). Protecting Research and Technology from Espionage. Expert Systems with Applications, doi: 10.1016/j.eswa.2012.12.051. Uskali, T. (2005). Paying attention to weak signals: The key concept for innovation journalism. Innovation Journalism, 2(11), p. 19. Van den Poel, D., De Schamphelaere, J., & Wets, G (2004). Direct and indirect effects of retail promotions on sales and profits in the do-it-yourself market. Expert Systems with Applications, 27(1), 53-62. Van Erkel, A. R., & Pattynama, P. M. T. (1998). Receiver operating characteristic (ROC) analysis: Basic principles and applications in radiology. European Journal of Radiology, 27(2), 88-94. Yoon, J. (2012). Detecting weak signals for long-term business opportunities using text mining of Web news. Expert Systems with Applications, 39(16), 12543-12550. Zeng, J., Duan, J., Cao, W., & Wu C. (2012). Topics modeling based on selective Zipf distribution. Expert Systems with Applications, 39(7), 6541-6546. Zhong, J., & Li, X. (2010). Unified collaborative filtering model based on combination of latent features. Expert Systems with Applications, 37(8), 5666-5672. Zipf, G. K. (1949). Human Behaviour and the Principle of Least Effort. Cambridge: Addison-Wesley. 
work_4iou6ghsc5hgzi7fvogqcu63pm	----	Shell fitting space for classification Expert Systems with Applications 38 (2011) 4530–4539 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Shell fitting space for classification Mostafa Ghazizadeh Ahsaee ⇑, Hadi Sadoghi Yazdi, Mahmoud Naghibzadeh Department of Computer Engineering, Ferdowsi University of Mashhad, Mashhad, Iran a r t i c l e i n f o a b s t r a c t Keywords: Shell fitting space Distance based transformation space Fitting Classification 0957-4174/$ - see front matter � 2010 Published by doi:10.1016/j.eswa.2010.09.127 ⇑ Corresponding author. E-mail addresses: GhazizadehAhsaee.most@stu-m Ahsaee), h-sadoghi@um.ac.ir (H. Sadoghi Yazdi), Naghibzadeh). In this paper, a shell fitting space (SFS) is presented to map non-linearly separable data to linearly sep- arable ones. A linear or quadratic transformation maps data into a new space for better classification, if the transformation method is properly guessed. This new SFS space can be of high or low dimensionality, and the number of dimensions is generally low and it is equal to the number of classes. The SFS method is based on fitting a hyper-plane or shell to the learning data or enclosing them into a hyper-surface. In the proposed method, the hyper-planes, curves, or cortex become the axis of the new space. In the new space a linear support vector machine (SVM) multi-class classifier is applied to classify the learn data. � 2010 Published by Elsevier Ltd. 1. Introduction Classification is an important research area with a wide range of applications. Nonlinear discriminant functions (NDF) are useful in training a system to recognize specific patterns and now many applications are based on this method. Neural network and support vector machine are preeminent mathemat- ical tools of NDF. Support vector machines (Vapnik, 1995) are very popular and powerful in learning systems because of the utilization of kernel machine in linearization, providing good generalization properties, their ability to classify input patterns with minimized structural misclassification risk and finding acceptable separating hyper-plane between two classes in the feature space. The result of applying kernels allows the algorithm to fit the maximum-margin hyper-plane in the transformed feature space. The transformation may be non-linear and the transformed space may be high dimensional; thus though the classifier is a hyper-plane in the high-dimensional feature space it may be non-linear in the original input space. If the used kernel is a Gaussian radial basis function, the corresponding feature space is a Hilbert space of infinite dimension. Maximum margin clas- sifiers are well regularized, so the infinite dimension does not spoil the results. Of course kernel methods (KMs) (Abe, 2005; Huang, Kecman, & Kopriva, 2006; Scholkopf & Smola, 2002; Shawe-Taylor & Cris- tianini, 2004) map input space into a high dimensional feature (HDF) space that may be helpful in linearization. As we know some kernels were proposed for this purpose, namely polynomi- Elsevier Ltd. ail.um.ac.ir (M. Ghazizadeh naghibzadeh@um.ac.ir (M. als, Gaussians, and splines (Friedman, 1991), but these kernels do not guaranty linearization in HDF. This problem motivates us for presentation of new space in which patterns can be clas- sified by linear classifier, we name it shell fitting space (SFS) be- cause we use the concept of shell fitting for the creation of the new space. Support vector machines (SVM) and its variants (Abe, 2005; Lin & Wang, 2002; Sadoghi Yazdi, Effati, & Saberi, 2007; Wang, 2005; Wu, Jie-Chi, & Lee, 2007) is a particular instance of KMs. But it has some weaknesses as follows. Slow training (compared to neural network) due to computationally intensive solution to QP problem especially for large amounts of training data ) needs special algorithms. � The kernel to be used is not deterministic and it changes for each data set and finally a large feature space is produced (with many dimensions). � Slow classification for the trained SVM. � Generates complex solutions (normally > 60% of training points are used as support vectors), especially for large amounts of training data. � Difficult to incorporate prior knowledge. Our proposed approach is not to expand the original space into a new space with many dimensions in comparison with kernel methods. In SFS the mapping is done to an m-dimensional space; where m is the number of classes. The rest of this paper is as follows. Section 2 is devoted to the development of our method, Section 3 explains how the new method works on example datasets, Section 4 is devoted to the application of the method on real datasets. In Section 5 we test our method with a simple linear classifier on SFS data and compare its accuracy results with multi-class SVM results on input space data. Conclusions are made in Section 6. http://dx.doi.org/10.1016/j.eswa.2010.09.127 mailto:GhazizadehAhsaee.most@stu-mail.um.ac.ir mailto:h-sadoghi@um.ac.ir mailto:naghibzadeh@um.ac.ir http://dx.doi.org/10.1016/j.eswa.2010.09.127 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa M. Ghazizadeh Ahsaee et al. / Expert Systems with Applications 38 (2011) 4530–4539 4531 2. Shell Fitting Space formulation Definitions: fxij 2 ½x11; x21; . . . ; xK 1 1; x12; . . . ; xK 2 2; . . . ; x1j; x2j; . . . ; xkj j; . . .�; j ¼ 1; . . . ; mg is the ith sample with n dimensions of class j. Cj is the fitted curve, hyperplane, or surrounded cortex or (shell) to the set {(Xij, yj), i = 1, . . . , kj} of data, where yj is the jth label of the training data and kj shows the number of samples with label j. In general our mapping is done from a space with m patterns (classes) of data in n-dimensional space to m-dimensional space by the fol- lowing notation: u : X ¼fx1; x2; . . . ; xng2 Rn ! uðXÞ 2 Rm where / is a function that depends on distance of data to each pat- tern’s fitted curve, hyperplane or surrounded cortex or (shell). This transformation can be seen in Fig. 1, where X is a feature in the input space and D = {d1, d2, . . . , dm} is the feature space element and ci is the class i. 2.1. Dealing with hyperplane When a hyperplane is fitted to a set of data the distance of the point Xi from that plane is calculated with the following formula: Suppose that the fitted hyperplane is y in (1) y ¼ w1 x1i þ w2x2i þ���þ wnxni ð1Þ So the distance from y is obtained as (2) dðXi; CÞ¼ w1x1i þ w2x2i þ���þ wnxniffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi w21 þ w22 þ . . . þ w2n q ð2Þ For instance in y = w1x1 + w2x2 we have the shape shown in Fig. 2a. 1 1 2 2y w x w x= + 2 2 2 1 2211 ww xwxw = + + d (a) (b) Fig. 2. (a) Distance from a hyperplane an Fig. 1. RBF-like transformation space. 2.2. Dealing with a hypersphere When a hypersphere is fitted to a set of data it can be formu- lated as: ðx1 � a1Þ 2 þðx2 � a2Þ 2 þ���þðxn � anÞ 2 ¼ r2 ð3Þ the distance between Xi and C is calculated by (4) dðXi; CÞ¼ r � ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x21i þ x 2 2i þ���þ x 2 ni q���� ���� � � ð4Þ where r is as follows: r ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðx1 � a1Þ 2 þðx2 � a2Þ 2 þ�� �þðxn � anÞ 2 q ð5Þ For example, consider two points (x1, y1) and (x2, y2) in Fig. 2. The distances d1 and d2 are calculated as shown in Fig. 2b. 2.3. Dealing with a polynomial in two-dimensional space When a polynomial is fitted to a set of data, points obtained by minimizing Eq. (6) are examined to see which one corresponds to the actual global minimum. To do so, Eq. (6) is evaluated for each of these points and the one that gives a lower value is selected r ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðx � x1Þ 2 þðy � y1Þ 2 q ð6Þ where y is as follows: y ¼ w0 þ w1 x1 þ w2x2 þ���þ wnxn ð7Þ For example suppose: y ¼ w0 þ w1 x1 þ w2x2 þ���þ w6x6 then we have: rðxÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðx � x1Þ 2 þðw0 þ w1x þ w2 x2 þ���þ w6x6 � y1Þ 2 q ð8Þ The points that minimize r are calculated by setting the derivatives of r equal to zero. Generally, by setting the derivatives of r equal to zero we have simple formulas as (9) @r=@x ¼ x � x0 þðwnxn þ . . . þ w1 x þ w0 � y1Þ � ðnwn xn�1 þðn � 1Þwn�1 xn�2 þ���þ w1Þ¼ 0 ð9Þ Then for each answer xi we find r(xi) from (8) and at the end the minimum distance is: Min frðxiÞjxi 2fanswersgg 2 1 2 11 )y-()x-( yxrd +−= 2 2 2 22 )y-()x-( yxrd +−= d (b) distance from a hypersphere. 4532 M. Ghazizadeh Ahsaee et al. / Expert Systems with Applications 38 (2011) 4530–4539 2.4. Dealing with a line in n-dimensional space Generally an equation for a line in n-dimensional space can be written as follows: Lðx; y; . . . ; zÞ : w ¼ a1x þ b1 y ¼ a2 x þ b2 . . . z ¼ anx þ bn 8>>>< >>>: ð10Þ So the distance between a point (x0, y0, . . . , z0) and (x, y, . . . , z) on L can be calculated with the formula below: dðw; x; . . . ; zÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðw � w0Þ 2 þðx � x0Þ 2 þ�� �þðz � z0Þ 2 q ð11Þ Considering Eq. (10), the distance can be expressed as: dðxÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ða1 x þ b1 � w0Þ 2 þðx � x0Þ 2 þ�� �þðanx þ bn � z0Þ 2 q ð12Þ Once again, we like to minimize (12). Therefore, its derivatives are set to zero and the x values for which d(x) is minimized are calculated @d=@x ¼ða1x þ b1 � w0Þa1 þðx � x0Þþ �� �þðanx þ bn � z0Þan ¼ 0 ð13Þ Then for each solution xi, d(xi) is evaluated to find the actual global minimum distance, that is: Min fdðxiÞjxi 2fanswersgg 2.5. Dealing with a curve in n-dimensional space If the number of features is more than two, we can use neural network-based methods like adaptive neural-fuzzy inference sys- tem (ANFIS) for this purpose. Detailed explanation can be followed from Wu et al. (2007) and a brief description is given in the Appendix. Steps of transformation using ANFIS are: (a) Learn ANFIS the classes of data. (b) Evaluate new data by the system. (c) Calculate the distance with a subtraction operation; evalua- tion result is subtracted from the main value of data as dis- tance. You can see the outputs in Section 3. 2.6. Dealing with a voluminous classes in n-dimensional space 2.6.1. SVDD one-class classification Three general approaches are proposed to resolve the one-class classification problem (Tax, 2001). The most straightforward meth- od to obtain a one-class classifier is to estimate the density of the training data and to set a threshold on this density. Several distri- butions can be assumed, such as a Gaussian or a Poisson distribu- tion. The most popular density models are the Gaussian model, the mixture of Gaussians, and the Parzen density (Bishop, 1995; Par- zen, 1962). In the second method a closed boundary around the target set is optimized. K-centers, nearest neighbor method and support vector data description (SVDD) are examples of the bound- ary methods (Tax & Duin, 2004; Ypma et al., 1998). Reconstruction methods are another one-class classification method which have not been primarily constructed for one-class classification, but rather to model the data. By using prior knowledge about the data and making assumptions about the generating process, a model is chosen and fitted to the data. Some types of reconstruction meth- ods are: the k-means clustering, learning vector quantization, self- organizing maps, PCA, a mixture of PCAs, diabolo networks, and auto-encoder networks. To obtain data description for a group of N target objects in in- put space, we try to find a sphere with minimum volume which encloses all or most of these target objects. The problem of finding the minimum hyper-sphere represented by a center ‘‘a” and radius ‘‘R” can be formulated into: FðR; aÞ¼ R2 þ C X i ni ð14Þ s:t: kxi � ak 2 6 R2 þ ni; 8i; ni P 0 ð15Þ The parameter C gives the tradeoff between the volume of the description and the errors. The free parameters, a, R and ni, have to be optimized, taking the constraints (15) into account. Con- straints (15) can be incorporated into formula (14) by introducing Lagrange multipliers and constructing the Lagrangian: LðR; a; ai; ci; niÞ¼ R 2 þ C X i ni � X i aifR2 þ ni �ðkxik 2 � 2a:xi þkak 2Þg� X i cini ð16Þ with the Lagrange multipliers ai P 0 and ci P 0. Setting partial derivatives to 0 gives these constraints: @L @R ¼ 0 : X i ai ¼ 1 ð17Þ @L @a ¼ 0 : a ¼ P iai xiP iai ¼ X i ai xi ð18Þ @L @ni ¼ 0 : C � ai � ci ¼ 0 ð19Þ From the last equation ai = C � ci and because ai P 0, ci P 0, La- grange multipliers ci can be removed when we demand that 0 < ai < C ð20Þ Resubstituting (17)–(19) into (16) results in: L ¼ X i aiðxi � xiÞ� X i;j aiajðxi � xjÞ ð21Þ By definition, R2 is the distance from the center of the sphere ‘‘a” to (any of the support vectors on) the boundary. Support vectors which fall outside the description (ai = C) are excluded. Therefore: R2 ¼ðxk � xkÞ� 2 X i aiðxi � xkÞ� X i;j aiajðxi � xjÞ ð22Þ Because we are able to give an expression for the center of the hypersphere ‘‘a”, we can test if a new object ‘‘z” is accepted by the description. For that, the distance from the object ‘‘z” to the cen- ter of the hypersphere ‘‘a” has to be calculated. A test object z is ac- cepted when this distance is smaller than or equal to the radius: kz � ak2 ¼ðz � zÞ� 2 X i aiðz � xiÞ� X i;j aiajðxi � xjÞ6 R2 ð23Þ 2.6.2. Using SVDD for our distance based method Here we used one SVDD for each class of a dataset (for example 2 SVDD for 2 classes). Then to check if a data belongs to a class, its distance from cortex of each SVDD classes is calculated and this distance is used as a main criterion to classify the data. In other words, a data belongs to a class, if its distance to a voluminous class’ cortex is less than the other classes’ in kernel space. The dis- tance is calculated by (24). As stated above, like hypersphere-but in kernel space, the transformation is done M. Ghazizadeh Ahsaee et al. / Expert Systems with Applications 38 (2011) 4530–4539 4533 d ¼ðz � zÞ� 2 X i aiðz � xiÞ� X i;j aiajðxi � xjÞ� R2 ð24Þ Fig. 4. Fitting a line to each of data classes. 3. Experimental results First the proposed circle fitting space transformation method is now demonstrated using simple data as an illustrating exam- ple, and at the end, our method will be tested using some well known datasets. In the following discussion, we assume that the label for each train data is known, in advance. Here our method has been implemented and tested on MATLAB (Math- Work Inc.). 3.1. Operation on synthetic data We describe our approach by some examples. At first we intro- duce our synthetic data as shown in Fig. 3a–i. In Fig. 3a there are two classes of linear shape which are linearly separable. But in other figures (b–i) are not linearly separable and these figures are divided into three categories: (1) sphere-based Fig. 3. (a) Two linear classes; (b) circle classes; (c) half-circle classes; (d) half-circle class (h) two classes for ANFIS and (i) two classes for ANFIS in 3-D space. figures (b–f), (2) sinuous function (Fig. 3g): more complex classes which are not easily separable. (3) Fig. 3h and i: are used in learn- ing ANFIS. es; (e) half-circle classes; (f) half-circle classes interfere each other; (g) sin-function; Fig. 5. Transformation space of Fig. 4. 4534 M. Ghazizadeh Ahsaee et al. / Expert Systems with Applications 38 (2011) 4530–4539 Consider two data classes in Fig. 3a that are linear classes. Firstly a line is fitted to each class of data with least square error fitting method. The result is seen in Fig. 4. And then to map them to a new space we use a formula like the following to compute the distance of each data (x0, y0) to the lines. Here d(X0, l1) stands for the distance of point X0 = (x0, y0) from line d1 dðX0; l1Þ¼ ja1 x0 þ b1y0 þ c1jffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi a21 þ b 2 1 q dðX0; l2Þ¼ ja2 x0 þ b2y0 þ c2jffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi a22 þ b 2 2 q We use these distances to map data to a new 2-dimensional space like Fig. 5 which is linearly separable. Fig. 6. (a) Fitted spheres and Fig. 7. Fitting a half-circle t As the second example, consider Fig. 3b. These classes are examples of hypersphere but in a 2D-space. Curve fitting is done first and the distance between each circular curve and dataset ele- ments is calculated. The results are depicted in Fig. 6a and b. Hor- izontal axis shows the distance between each dataset element and the internal circle and the other axis is the distance between each dataset element and the external circle. 3.2. Transformation of circle classes Our proposed method is applied to a data series in 2-dimen- sional space (two classes with two features) and the results are shown in Figs. 7–11. We have tested our method on four kinds of half-circle (Fig. 3c–f). The two half-circles may even interfere each other (Fig. 3f). The results of curve fitting and calculating distance between each dataset element and each fitted data are depicted in Figs. 7–10. 3.3. Transformation of more complex classes Data distribution may seem so complex, but this method tries to separate these two classes of data (Fig. 11). 3.4. Transformation using ANFIS Other experimental results are those which use ANFIS to fit a curve (Figs. 12 and 13). Here we have used two classes in 2- and 3-dimensional space. The two ANFIS were used to fit curves to the data set. As shown in Fig. 12, the result is linearly separable. (b) Output of mapping. o each of data classes. Fig. 8. Fitting a half-circle to each of data classes. Fig. 10. Fitting a half-circle to each of data classes. Fig. 9. Fitting a half-circle to each of data classes. M. Ghazizadeh Ahsaee et al. / Expert Systems with Applications 38 (2011) 4530–4539 4535 3.5. Transformation using SVDD As stated before, SVDD is a one-class classification used in clas- sification of voluminous data. At this point, a dataset is shown in Fig. 14a and the result of our proposed method is depicted in Fig. 14b. 4. Experiments on Datasets In this section our proposed method is used to classify some datasets. These datasets are from UCI Machine Learning Reposi- tory. Datasets used here are breast-cancer-wisconsin dataset, Iris dataset, Ionosphere dataset, Transfusion dataset and heart dataset with 11, 5, 35, 15, 4, and 14 attributes with 2, 3, 2, 3, 2, and 2 classes of data in sequence. 4.1. Circle fitting on datasets As we said in previous sections, circle fitting is used in classifi- cation of like-circle data. We show here that our method is good to classify breast-cancer-wisconsin and Iris datasets and to some ex- tent to classify ionosphere dataset using circle fitting. Our method calculated the distance of each data in the dataset to circles fitted Fig. 11. Fitting a curve to each of data classes. Fig. 13. Curves have been fitted to data sets of two classes and the result of transformation has been shown in right figure. Fig. 12. Fitting a curve to a data set using ANFIS and transform them to a new linearly separable space. 4536 M. Ghazizadeh Ahsaee et al. / Expert Systems with Applications 38 (2011) 4530–4539 on data and the classes are identified. We can see the result in Figs. 15–17. We can see from Figs. 15 and 17 that data may have interfer- ence in their distribution or may not be well fitted to a hypersphere. 4.2. Using SVDD on dataset Here we used one SVDD for each class of iris dataset (3 SVDD for 3 classes). Then we calculated distances from cortex of each SVDD class. The result of classification is depicted in Fig. 18. For other datasets (breast-cancer-wisconsin dataset, Ionosphere dataset, Wine dataset, Transfusion dataset and heart dataset), the same process is followed and the result shows, Figs. 19–22, that SVDD is good to be used as a tool for our distance based classification. 5. Accuracy of our method in comparison with multi-class SVM To show the performance of our method, we first transformed the data of each dataset to the proposed shell fitting space and we then used a simple linear classification (Adaline) to classify the transformed data. In this stage, 50% of the data in each dataset was used for the training purpose. Kernel-based SVM classifier was used on the original data to compare the accuracy of our classifica- tion results. Amongst the available kernels the one with the best performance for each dataset was used in the training stage. Both systems were tested for each test data of each dataset. The compar- ison results of the mean values of 30 runs of both methods follow. It is worth mentioning that in each run of each dataset the data are randomly divided into training class and test class. Adaline classifier uses a W, weight vector, to separate classes with a hyper-plane where W is obtained from (25): W ¼ DXðXX0Þ�1 ð25Þ In (25) D is the vector of the desired label for each training data of the class of training data and X is the vector of training samples. We used rbf, linear, polynomial, . . . , kernels for SVM classifier and se- lected the best results for the comparison purpose. Accuracy (AC) is calculated for a dataset using the following definition: AC ¼ðnumber of true classificationsÞ =ðnumber of true and false classificationsÞ Or, AC ¼ number of true positives þ true negatives number of true positives þ true negatives þ false positives þ false negatives ð26Þ The comparison results are shown in Table 1. As you can see, in our method only the hyper-sphere fitting to Wine dataset and heart dataset has led to a lower accuracy. This shows that, if the fitting shape is not selected well, the result might not be desirable. 6. Conclusions In this paper we introduced a new transformation space which is easier than the other space transformations to classify input Fig. 20. Result of SVDDD on Ionosphere dataset. Fig. 19. Result of SVDDD on breast-cancer-wisconsin dataset. Fig. 17. Result of circle fitting on ionosphere dataset. Fig. 16. Result of circle fitting on Iris dataset. Fig. 18. Result of SVDD on Iris dataset. Fig. 14. (a) Input dataset and (b) result of classification using distance based method. Fig. 15. Result of circle fitting on breast-cancer-wisconsin dataset. M. Ghazizadeh Ahsaee et al. / Expert Systems with Applications 38 (2011) 4530–4539 4537 Fig. 21. Result of SVDDD on Wine dataset. Fig. 22. Result of SVDDD on Ionosphere dataset. Table 1 Comparison of our method accuracy with multi-class SVM (mean accuracy in 30 runs). Mean accuracy in 30 runs Our method (circle fitting) (%) Our method (using SVDD) (%) Multi-class SVM (%) Cancer 96.76 97.25 96.38 Heart 53.86 100 77.38 Wine 30.80 100 76.55 Iris 93.44 96.46 95.20 Ionosphere 72.27 99.72 85.89 4538 M. Ghazizadeh Ahsaee et al. / Expert Systems with Applications 38 (2011) 4530–4539 data. The main idea was to use distance of data to a line, curve, or hyperplane fitted to each class’ data or shells enclosing the classes. Fig. 23. ANFIS a The results show that this space transformation works well for col- lections of data. Appendix. Adaptive neuro-fuzzy inference system (ANFIS) architecture (Jang, 1993) A typical architecture of ANFIS is shown in Fig. 23 for modeling of function, in which a circle indicates a fixed node, and a square indicates an adaptive node. For simplicity, we consider two inputs x, y and one output z in the fuzzy inference system (FIS). The ANFIS used in this paper implements a first-order Sugeno fuzzy model. Among the many fuzzy inference systems, the Sugeno fuzzy model is the most widely used due to its high interpretability and compu- tational efficiency, and built-in optimal and adaptive techniques. For example for a first-order Sugeno fuzzy model, a common rule set with two fuzzy if-then rules can be expressed as (27) and (28) Rule 1 : If x is A1 and y is B1; then z1 ¼ p1x þ q1y þ r1 ð27Þ Rule 2 : If x is A2 and y is B2; then z2 ¼ p2x þ q2y þ r2 ð28Þ where Ai, Bi (i = 1, 2) are fuzzy sets in the antecedent, and pi, qi, ri (i = 1, 2) are the design parameters that are determined during the training process. As in Fig. 23, the ANFIS consists of five layers. Layer 1, every node i in this layer is an adaptive node with a node function like (29) O1i ¼ lAiðxÞ; i ¼ 1; 2 O1i ¼ lBiðyÞ; i ¼ 3; 4 ð29Þ where x, y are the input of node i, lAiðxÞ and lBiðyÞ and can adopt any fuzzy membership function (MF). In this paper, Gaussian MFs which are defined by (30) are used gaussianðx; c; rÞ¼ e� 1 2 x�c rð Þ 2 ð30Þ where c is center of Gaussian membership function and r is stan- dard deviation of this cluster. Layer 2, every node in the second layer represents the ring strength of a rule by multiplying the incoming signals and forward- ing the product as (31) O2i ¼ wi ¼ lAiðxÞlBiðyÞ; i ¼ 1 ð31Þ Layer 3, the ith node in this layer calculates the ratio of the ith rule’s ring strength to the sum of all rules’ rings strengths: O3i ¼ -i ¼ xi x1 þ x2 ; i ¼ 1; 2 ð32Þ where -i is referred to as the normalized ring strengths. rchitecture. M. Ghazizadeh Ahsaee et al. / Expert Systems with Applications 38 (2011) 4530–4539 4539 Layer 4, the node function in this layer is represented by (33). O4i ¼ ~xiZi ¼ ~xiðpix þ qi y þ riÞ; i ¼ 1; 2 ð33Þ where -i is the output of layer 3, {pi, qi, ri} and is the parameter set. Parameters in this layer are referred to as the consequent parameters. Layer 5, the single node in this layer computes the overall out- put as the summation of all incoming signals like (34) O51 ¼ X2 i�1 -i zi ¼ x1 z1 þ x2z2 x1 þ x2 ð34Þ It is seen from the ANFIS architecture that when the values of the premise parameters are fixed, the overall output can be expressed as a linear combination of the consequent parameters: z ¼ð~x1 xÞp1 þð~x1 yÞq1 þð~x1Þr1 þð~x2xÞp2ð~x2yÞq2 þð~x2Þr2 ð35Þ The hybrid learning algorithm (Jang, 1993) and (Jang, Sun, & Mizu- tani, 1997) combining the least square method and the back prop- agation (BP) algorithm can be used to solve this problem. This algorithm converges much faster since it reduces the dimension of the search space of the BP algorithm. During the learning process, the premise parameters in layer 1 and the consequent parameters in layer 4 are tuned until the desired response of the FIS is achieved. The hybrid learning algorithm has a two-step process. First, while holding the premise parameters fixed, the functional signals are propagated forward to layer 4, where the consequent parameters are identified by the least square method. Second, the consequent parameters are held fixed while the error signals, the derivative of the error measure with respect to each node output, are propagated from the output end to the input end, and the premise parameters are updated by the standard BP algorithm. References Abe, S. (2005). Support vector machines for pattern classification. Springer-Verlag London Limited. Abe, S. (2005). Advances in pattern recognition. Springer-Verlag London Limited. ISSN: 1617-7916. Bishop, C. (1995). Neural networks for pattern recognition. Walton Street, Oxford OX2 6DP: Oxford University Press. Friedman, J. H. (1991). Multivariate adaptive regression splines. Annals of Statistics, 19, 1–141. Huang, T., Kecman, V., & Kopriva, I. (2006). Kernel based algorithms for mining huge data sets. Berlin, Heidelberg: Springer-Verlag. UC Irvine Machine Learning Repository. <http://archive.ics.uci.edu/ml/> Jang, J.-S. R. (1993). ANFIS: Adaptive-network-based fuzzy inference system. IEEE Transactions on Systems, Man, and Cybernetics, 23(3), 665–685. Jang, J.-S. R., Sun, C. T., & Mizutani, E. (1997). Neuro–Fuzzy and Soft Computing: A computational approach to learning and machine intelligence. Englewood Cliffs, NJ: Prentice-Hall. Lin, C.-f., & Wang, S.-d. (2002). Fuzzy support vector machine. IEEE Transactions on Neural Networks, 13(2), 464–471. Parzen, E. (1962). On estimation of a probability density function and mode. Annals of Mathematical Statistics, 33, 1065–1076. Sadoghi Yazdi, H., Effati, S., & Saberi, Z. (2007). The probabilistic constraints in the support vector machine. Applied Mathematics and Computation, 194(2), 467–479. Scholkopf, B., & Smola, A. J. (2002). Learning with kernels. Massachusetts Institute of Technology. Shawe-Taylor, J., & Cristianini, N. (2004). Kernel methods for pattern analysis. Cambridge University Press. Tax, D. M. J. (2001). One-class classification: concept learning in the absence of counter- examples (Vol. 65). Netherlands: Technische Universiteit Delft. Tax, D. M. J., & Duin, R. P. W. (2004). Support vector data description. Machine learning (vol. 54). Kluwer Academic Publishers (pp. 45–66). Vapnik, V. (1995). The nature of statistical learning theory. New York: Springer- Verlag. Wang, L. (Ed.). (2005). Support Vector Machines: Theory and Applications. Berlin Heidelberg: Springer-Verlag. ISSN: 1434-9922. Wu, C., Jie-Chi, Y., Lee, Y. (2007). An approximate approach for training polynomial kernel SVMs in linear time. In 45th Annual meeting of the association for computational linguistics companion volume proceedings of the demo and poster sessions, June 2007 (pp. 65–68). Ypma, R.D. (1998). Support objects for domain approximation. In Proceedings of international conference on artificial neural networks (ICANN98). Skovde, Sweden. http://archive.ics.uci.edu/ml/ Shell fitting space for classification Introduction Shell Fitting Space formulation Dealing with hyperplane Dealing with a hypersphere Dealing with a polynomial in two-dimensional space Dealing with a line in n-dimensional space Dealing with a curve in n-dimensional space Dealing with a voluminous classes in n-dimensional space SVDD one-class classification Using SVDD for our distance based method Experimental results Operation on synthetic data Transformation of circle classes Transformation of more complex classes Transformation using ANFIS Transformation using SVDD Experiments on Datasets Circle fitting on datasets Using SVDD on dataset Accuracy of our method in comparison with multi-class SVM Conclusions Adaptive neuro-fuzzy inference system (ANFIS) architecture (Jang, 1993) References 
work_4ivtz7u4tbc4thdfcolusvzlzq	----	doi:10.1016/j.eswa.2003.12.018 A multiagent approach for diagnostic expert systems via the internet Khaled Shaalan a,*, Mona El-Badry b , Ahmed Rafea c a Department of Computer Science, Faculty of Computers and Information, Cairo University, 5 Tharwat Street, Orman, Giza 12613, Egypt b Central Laboratory for Agricultural Expert Systems (CLAES), P.O. Box 100, Dokki, Giza, Egypt c Department of Computer Science, American University, 113, Sharia Kasr El-Aini, P.O. Box 2511, 11511 Cairo, Egypt Abstract In recent years there has been considerable interest in the possibility of building complex problem solving systems as groups of co- operating experts. This has led us to develop a multiagent expert systems capable to run on servers that can support a large group of users (clients) who communicate with the system over the network. The system provides an architecture to coordinate the behavior of several specific agent types. Two types of agents are involved. One type works on the server computer and the other type works on the client computers. The society of agents in our system consists of expert systems agents (diagnosis agents, and a treatment agent) working on the server side, each of which contains an autonomous knowledge-based system. Typically, agents will have expertise in distinct but related domains. The whole system is capable of solving problems, which require the cumulative expertise of the agent community. Besides to the user interface agent who employs an intelligent data collector, so-called communication model in KADS, working on the client sides. We took the advantage of a successful pre-existing expert systems—developed at CLAES (Central Laboratory for Agricultural Expert Systems, Egypt)—for constructing an architecture of a community of cooperating agents. This paper describes our experience with decomposing the diagnosis expert systems into a multi-agent system. Experiments on a set of test cases from real agricultural expert systems were preformed. The expert systems agents are implemented in Knowledge Representation Object Language (KROL) and JAVA languages using KADS knowledge engineering methodology on the WWW platform. q 2003 Elsevier Ltd. All rights reserved. Keywords: Expert systems; Agents; WWW 1. Introduction As computer hardware and software become increas- ingly powerful, so do applications, which used to be considered beyond the scope of automation. To cope with increased demands, software systems are becoming correspondingly larger and more complex (Genesereth & Ketchpel, 1994; Jennings & Varga, 1993; Jennings & Wooldridge, 1995). In recent years, there has been considerable interest in the possibility of building complex problem solving systems as groups of co-operating experts. Distributed Artificial Intelligence (DAI) is the study of how such systems might be built. Historically, empirical research in DAI has focused on three main areas: blackboard architecture (Engelmore & Morgan, 1988; Nwana, 1996), systems based on negotiation (Smith, 1980), and multi-agent planning systems (Durfee and Lesser, 1988). More recently, co-operating expert systems have emerged as a research area of some importance (Nwana & Ndumu, 1999). Agent-based computing represents an exciting new synthesis both for Artificial Intelligence (AI) and, more generally, Computer Science. It has the potential to significantly improve the theory and the practice of modeling, designing, and implementing computer systems (Jennings, 2000). Since the 1980s, software agents and multi- agent systems have grown into what is now one of the most active areas of research and development activity in computing generally (Jennings & Wooldridge, 1998). There are many reasons for the current intensity of interest, but certainly one of the most important is that the concept of an agent as an autonomous system, capable of interacting with other agents in order to satisfy its design objectives, is a natural one for software designers (Wooldridge & Ciancar- ini, 2001). Just as we can understand many systems as being composed of essentially passive objects, which have state, 0957-4174/$ - see front matter q 2003 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2003.12.018 Expert Systems with Applications 27 (2004) 1–10 www.elsevier.com/locate/eswa * Corresponding author. Tel.: þ2-012-2223377; fax: þ2-02-761-7628. E-mail addresses: shaalan@mail.claes.sci.eg (K. Shaalan); mona@ mail.claes.sci.eg (M. El-Badry); rafea@aucegypt.edu (A. Rafea). http://www.elsevier.com/locate/eswa and upon which we can perform operations, so we can understand many others as being made up of interacting, semi-autonomous agents. There are several benefits of using the intelligent agent paradigm for software systems. Agents can provide a high level of abstraction for dealing with intelligent systems and distributed systems. The agent abstraction is a natural extension of object-oriented technology, encapsulating the agents knowledge within an active process and providing a standard interface for intercommunications. This leads to another benefit of agent-based systems, interoperability (Genesereth & Ketchpel, 1994). With a common agent communication Language, such as KQML (Finin, Fritzon, Mckay, & McEntire, 1994; Finin, Labrou, & Mayfield, 1997; Labrou & Finin, 1997), programs written as agents can communicate and cooperates with other such programs. In other words, agent-based system architecture provides a consistent interface for intelligent systems to interact with. Finally, agent systems often provide a high-level ‘human- like’ interface to take the GUI revolution one step further (Moore et al., 1997). Multi-agent systems are the best way to characterize or design distributed computing systems (Huhns & Stephens, 1999). Multi-agent systems are commonly intended as computational systems, where several (semi-) autonomous entities interact or work together to perform some tasks. There are several motivations for having multiple agent systems (Nwana, 1996). They include: † To solve problems that are too large for a centralized single agent to do due to resource limitations or the sheer risk of having one centralized system; † To allow for interconnecting and interoperation of multiple existing legacy systems such as expert systems and decision support systems; † To provide solutions to inherently distributed problems such as distributed sensor networks (Durfee and Rosenschein, 1994) or air-traffic control; † To enhance modularity (which reduces complexity), speed (due to parallelism), reliability (due to redun- dancy), flexibility (i.e. new tasks are composed more easily from the more modular organization) and reusability at the knowledge level (hence shareability of resources). The rationale for interconnecting computational agents and expert system is to enable them to cooperate in solving problems, to share expertise, to work in parallel on common problems, to be developed and implemented modularly, to be fault tolerant through redundancy, to represent multiple viewpoints and the knowledge of multiple experts, and to be reusable (Huhns & Stephens, 1999). Nevertheless, With the advent of the Internet, many researchers have been taking a closer look at distributed software systems. Recently, a large share of this research has focused on intelligent distributed systems, which have come to be known as multi-agent systems. As a result, new development methodologies specifically designed for multi-agent systems have been introduced (Iglesias, Garijo, & Gonzalez, 1998) and several tools are now available for building multiagent systems (Raphael & Deloach, 2000). At the Central Laboratory for Agricultural Expert System (CLAES), at the Agriculture Research Center of Ministry of Agriculture and Land Reclamation in Egypt, a number of successful agricultural expert systems were developed and deployed to a large number of users (Rafea, 1995, 1998; Rafea, El-Azhari, & Hassan, 1995; Rafea, El-Azhari, Ibrahim, Soliman, & Mahmoud, 1995; Rafea, Hassan, & Hazman, 2003; Rafea & Mahmoud, 2001; Rafea & Salah, 1994; Rafea, Warkentin, & Ruth, 1991, 1992). These applications are traditional standalone systems. In these expert systems, the knowledge acquired from multiple domain experts that belongs to different disciplines are implemented as an integrated system. The aim of this research is to harness the potential of the agent technology for constructing a community of cooperating agents capable of diagnosing disorders in the agriculture domain. By implementing expert systems as multi-agents that perform their tasks remotely, the expertise can be published on the Web. Expert systems running on the Internet can support a large group of users who communicate with the system over the network. This paper is organized as follows. Section 2 gives an overall structure of the proposed architecture of our system. Section 3 describes the user interface agent. Section 4 introduces the coordination agent. Section 5 presents the selected expert systems agents. Section 6 shows the ability of the system in producing diagnosis of disorders based on a set of input observations from real-world cases in the multi- agent environment. Section 7 concludes the paper. 2. Overview of the proposed system Our proposed system considers the two most common synergetic expert system applications—diagnosis and treat- ment. From the architectural point of view, the system provides a framework to coordinate the behavior of several specific agent types. The society of agents in our system consists of expert system agents (diagnosis agents, and a treatment agent) working on the server side, The diagnosis knowledge is distributed among several agents, each agent is an autonomous expert for a certain domain knowledge. Typically, agents will have expertise in distinct but related domains. The whole system is capable of solving problems, which require the cumulative expertise of the agent community. In addition to the user interface agent who employs an intelligent data collector working on the client sides. Those agents communicate by passing messages to the coordination agent. The clients communicate with the server through socket connection via the Internet. At the client side the user interface agent send requests for K. Shaalan et al. / Expert Systems with Applications 27 (2004) 1–102 diagnosis/treatment services to the server. Answers are sent back to clients. We have applied our system in the domain knowledge of diagnosing disorders of the grapes crop, see Fig. 1. A diagnosis agent is responsible for producing diagnosis of disorders based on a set of input observations. The treatment agent, who is responsible for finding a suitable recommendation to treat a certain disorder taking into consideration the possible causes of this disorder, provides a set of operations to be performed in order to treat the disorder and remove the causes. The treatment agent can only be consulted after a successful diagnosis session. In the early stage of this research, we have designed the system with only two expert agents, diagnosis and treatment. But later on, we discovered that the diagnosis agent combines different specialization that are well defined in the agricultural domain. Consequently, we improved our architecture by furtherly decomposing the diagnosis agent. In the agent community decomposition concerns the partitioning of the problem domain into agents. The key problem lies in deciding how the domain is to be partitioned. This decision can only be reached by the extensive consultation with the domain experts. In our case, we figured out that the grape diagnosis consists of six agents, namely: fungal agent, insect agent, nematode agent, snail agent, mites agent, and nutrition deficiency agent. The new architecture concerning diagnosis has the following advan- tages: † It tries to simulate what may happen in real world in a very fine-grained manner. Each agent corresponds to a domain expert specialization. † Partitioning of domain knowledge makes it easy to maintain knowledge, and to improve performance. In real life, the domain expert who diagnoses a certain disorder can also give the treatment method of this disorder. The treatment methods suggested by different domain experts may conflict with each other such that a decision is needed to decide which one is applied first, and so forth. So, we proposed in our architecture to separate the treatment process from the diagnosis process, and collectively kept into the treatment agent in order to resolve conflicts that may arise in the treatment application. Expert systems agents include a bilingual KQML translator module. The message transferred through the system takes the KQML format. Expert system agents are implemented in KROL (Shaalan, Rafea, & Rafea, 1998), while the user interface agent is implemented in Java. KQML message needs to be translated to/from Prolog. So we have designed a module which performs this bilingual translation. Similarly, another module for translating user interface messages is designed, which translates KQML to Java and the vice versa. 3. The user interface agent The user interface agent provides access to the expert systems agents. For portability, this program is a Java applet. When a user browses the WWW page of the system, this applet is downloaded. This is responsible for providing means through which the user could initiate problem- solving sessions, by activating the communication model, as well as handling the presentation of the expert system results. The communication model acts as an intelligent data collector. The user interface agent includes a Java-KQML bilingual translator. 3.1. The communication model Common-KADS project has addressed many issues related to the development of expert system. One of the most important issues is known as the communication Fig. 1. The society of agents. K. Shaalan et al. / Expert Systems with Applications 27 (2004) 1–10 3 model (Hoog et al., 1992). One of the best advantages of the communication model is that it completely separates the user interface model from the knowledge model. However, dividing an integrated expert system into a two-component architecture—the problem solving component and an inter- face or a front-end component—is not usually a straightfor- ward task since the control of the interface is usually managed by the problem-solving component itself. In what follows we show the importance of the communication model in our system. For instance, in many expert systems, the questions to be asked are determined by answers to previously asked questions. In this case several solutions are possible. The first and the simplest of these, is devising an application specific interface where the user is presented with all possible inputs. Needless to say, this approach will not meet the needs of any reasonably large application, and in addition, will confuse the user. Another approach entails keeping control embedded within the problem solving server application. In that case, the interface will be used to present the user with an input request upon receiving such a request from the server. While this approach might seem reasonable, it suffers from major limitations. First, it relies on heavy communication between the server application and the client front-end, so the user may have to wait for prolonged periods of time depending on the network traffic and bandwidth. In case of synchronous communication, it can engage the server in one connection for an indefinite amount of time, making it impossible for other users to make use of that same server. Although time-out operations could be implemented to avoid indefinite postponement, the application server will still not be fully utilized. In case of asynchronous communication the server will have to maintain extra knowledge such that data inputs could be mapped to application clients. This would be necessary in order to maintain data values that are consistent with its clients. If however, the interface component employed a communication model that had just enough knowledge about which inputs it should ask about and in which cases, then the client could use this knowledge to collect all needed inputs in an intelligent fashion. Next, send them in one batch to the server for processing. In this case the connection between the client and the server will only be open for the period of sending the inputs, processing them at the server and receiving the output. For most practical applications, this period is usually reasonably short. Meantime, if other clients need to service a request, the request will be placed in a wait queue where the waiting time will be short enough to make that wait transparent. However, care must be taken in selection of the queue size. Otherwise, if the number of the clients for the application grows, then wait times might also grow to unsatisfactory figures. The implementation of the communication model is inspired by the hierarchical classification model (HC). HC is a problem solving method identified by Chandrasekran (1988) for solving diagnostic types of problems as part of his Generic Task approach to expert system development. In HC, knowledge is represented as hypotheses hierarchically organized in a tree structure such that a general hypothesis is always above more specific ones in the tree. Using a control strategy known as ‘establish and refine’, hypothese are explored top down. If a hypothesis at the top level succeeds (establishes), its immediate descendants are required to be established themselves one by one. This process of attempting to establish the descendants is referred to as ‘refining’ the parent hypothesis. If, on the other hand, a hypothesis fails, then it is said to be ruled out and so are all the hypothese beneath it in the tree (El-Beltagy et al., 1995). In our model, knowledge components for which input is desired, are also organized in a hierarchical fashion. 4. The coordination agent This component acts as the interface between the user interface agent and other agents. It is responsible for establishing the communication link with the desired expert system agent based on the user request, through the socket connection. In the proposed agent society, TCP/IP is chosen as the low-level transport mechanism for agent-to-agent communication. In our implementation, a general-purpose client sockets class ESClient is defined that inherits from Socket class of the java.net package. Each time the user interface agent needs to communicate with any of the other agents, it sends to the coordination agent a message. The coordination agent in turns creates an instance of this class giving it both the host names of the target agent, and the port number on which the agent’s server socket resides. Host names and port numbers are passed in the first place to the coordination agent as parameters from the applet’s initiating HTML page. 5. The expert system agents Each expert system agent consists of two basic components, a mediator and the expert system itself. As indicated by its name, a mediator is the module responsible for sending and receiving messages through the network connection. This communication can be established through a server socket. All agent communication takes place through the coordination agent. The expert system consists of three sub-components: Domain knowledge: it contains the static knowledge of the domain, concepts, properties, and two types of relations. A concept is identified through its name. Concepts can have properties that are defined through their names and descriptions of the values that the property can take. The relations may be relations K. Shaalan et al. / Expert Systems with Applications 27 (2004) 1–104 between concepts or relations between property expressions. Inference knowledge: describes what inferences can be made on the basis of the knowledge in the domain knowledge. It describes what inference can made, but not how or when they are made. In KADS the selected knowledge engineering methodology, the following terms are used to denote the various aspects of a primitive inference: inference step, and role. An inference step performs an act that operates on some input data and has the capability of producing a new piece of information as its output. The elements on which inference steps operate or produce are called roles. A role can be an input role or an output role. It acts as a placeholder for domain objects. Domain objects can be linked to more than one role. Task knowledge: actually describes the steps of execut- ing the knowledge sources in the inference knowledge. In a previous work (Shaalan et al., 1998), we described the implementation of expert systems in KROL. In KROL, Webkrol is the library that handles all the network necessary function. 5.1. The diagnosis agent As mentioned above there are six diagnosis agents. The user interface agent is responsible for loading the diagnosis communication model in the client side through a WWW browser. The diagnosis session proceeds as follows, see Fig. 2. 1. The diagnosis communication model asks the user about all the observations and collects his responses. 2. The user interface agent calls the Java-KQML translator module to convert the collected observation into a KQML message format. 3. The user interface agent sends this message to the coordination agent. 4. The coordination agent initiates the socket connection with the server and broadcast this message to all diagnosis agents at the server site. 5. At the server side, for each agent, after receiving the diagnosis data, the Prolog-KQML translator translates the KQML message into a prolog term. This message is directly interpreted and executed by the diagnosis expert system. Fig. 2. Observation and disorder flow in the diagnosis session. K. Shaalan et al. / Expert Systems with Applications 27 (2004) 1–10 5 6. The results are converted again to a KQML format and sent back to the coordination agent. 7. The coordination agent collects all the answers and sends it to the user interface agent. 8. The user interface in turn formulates the output and displays it. 5.2. The treatment agent Treatment agent session can be conducted after a successful diagnosis session. The treatment sessions that correspond to this behavior is described as follows. After the diagnosis session reaches a successful disorders (i.e. at least one of the diagnosis agent find disorders). This information is sent to the coordination agent that transfers it to the user interface agent. If the user requests a treatment for these disorders, the user interface agent calls the treatment communication model and start collecting data related to treatment. Both the treatment input data and the diagnosis result (the confirmed disorders) will transfer to the coordination agent. The coordination agent in turn opens the connection with the treatment agent and mediates this information to it. The treatment agent in turn tries to get a treatment for these disorders and sends the results back to the coordination agent that will transfers it to the user interface agent that will display its contents through the WWW browser, see Fig. 3. 6. An example of utilization and testing To bring out the advantages of employing multi-agent technology and expert systems, it was necessary to look at some real test cases from their conception to realization. We Fig. 3. Disorder and treatment flow from a diagnosis session to a treatment session. K. Shaalan et al. / Expert Systems with Applications 27 (2004) 1–106 have generated test cases that trigger mostly knowledge of the six diagnosis agents. For each case a brief description of its objective is given as well as its related input/output screens in diagnosis session and in treatment session. The function of coordination agent is shown through screens at run time. It is recommended that the end user of our system use a java enabled browser (e.g. internet explorer). Throughout this section we show screens taken from running our system on the MS Internet Explorer. When the user connects to the server site through any available web browser, the front page shown in Fig. 4 appears. This page initiates the user interface agent and carries the necessary information about the address and the port numbers of all expert system agents. We have adopted a testing methodology based on manually generated test cases taken from the diagnosis and treatment documents of grapes crop. The objective is to cover all the agents, verify the proposed system by considering the number of expert systems agents that succeed in diagnosis: 0-agent succeed, 1-one agent succeed, and many agent succeed. The methodology also trace down the function of the coordination agent in each case. The main steps in our methodology are: 1. Walk through the diagnosis design document and manually generate random test cases. It is worth noting that the generated test cases are by no means exhaustive. The test case is generated by walking backward through the inference chain, starting from a target disorder(s) and randomly selecting values that conclude the selection reached at any given point in the inference chain. 2. For each disorder obtained from step 1, walk through the treatment design document and manually generate random test cases. This time the test cases are generated by walking through the inference chain to get the treatment operation(s). 3. Merge two or more test cases in order to test the behavior of the multi-agent environment. This will try to simulate real life situations, when it may happen that one or more disorders appear at the same time (this step is used for testing the possibilities of many-agent succeeded). 4. Run the generated test cases on the system. 5. Compare the results with the target disorder. For time and space constraints, we will show a test case that tries to diagnose and treat disorders in case of three diagnosis agents respond. 6.1. Verify three diagnosis agents: nutrition deficiency, mite, and snail Objective. The objective of test case is to conduct an experiment that demonstrates the capabilities of the system to diagnose and treat disorders in case of three diagnosis agents respond. In this case we choose disorders that belong to nutrition deficiency, mite, and snail expert agents. This test case covers the diagnosis and the treatment for the disorders due to iron deficiency, zinc deficiency, manganese deficiency, mite, and land snail disorders. The treatment of iron deficiency will be on date 26/05/2001, the material used is iron_chelate. The treatment of zinc deficiency will be on date 20/05/2001, the material used is zinc_sulphate. The treatment of manganese deficiency will be on date 23/05/2001, the material used is manganese_sulphate. The treatment of mite will be on date 01/06/2001, the material used is vertamic. The treatment of land snail will be on date 29/05/2001, the material used is iron_sulphate. 6.1.1. Diagnosis session Diagnosis input: Growth stage: vegetative Leaves color: yellow Leaves status: small, drop Color of spots: brown Position of leaves color: whole leaf Type of infection: new leaves Range of infection: adjacent spots Depth of water: medium Snail exist: yes Fig. 5. An example of input screens in the diagnosis session.Fig. 4. Front page screen. K. Shaalan et al. / Expert Systems with Applications 27 (2004) 1–10 7 An example of input screens is shown in Fig. 5. After all the required symptoms are collected, the user interface automatically sends them to the coordination agent. Which in turn will establish the connection and broadcast the collected data to all the diagnosis agents. The answers from those agents are sent back to the user interface agent. Diagnosis output: The output from the diagnosis agents of this case is shown in Fig. 6, which indicates that the three agents that are involved in the test case have found that the input symptoms match a certain disorder in their knowledge base. The user interface asks the user whether or not it may call the treatment agent. Assume we continue with the treatment agent. 6.1.2. Treatment session The user interface agent will call the communication model of treatment. Then the treatment communication model start asking about some information concerning treatment, such as the infection range, and the material used. After that the coordination agent starts the connection with the server and calls the treatment agent. Treatment input: Zinc used material: zinc sulphate. Treatment output: The result from treatment agent again returns back to the user interface agent through the coordination agent. The recommended treatment includes the date of the treat operation, the material used, the material quantity, the method if any, and the treatment advice, see Fig. 7. The right screen, is the detailed screen of the disorder shown on its left. 6.1.3. Coordination agent Fig. 8 shows a trace of the flow of control of the output from coordination agent to the diagnosis agents, and the input from each agent at run time. Each agent located in a server with server name listen to a port number. The input shows that there are three agents answer with reply (at port 1095, 1070, 1075) and all the other three agents answer with sorry. The coordination agent will then send to the treatment agent. 7. Conclusions In this paper we have described our experience with decomposing the diagnosis expert system into a multi-agent system. Our approach consists of a number of cooperating agents capable of solving the diagnosis and treatment problem in the agricultural domain. It has been applied to the grapes crop for providing services to a large group of users over the internet. The system can be geared towards any other related system or application. The widespread use of the internet and WWW provides an opportunity for developing ES that fit into the agent technology widely available. Fig. 6. Output in the diagnosis session. Fig. 7. An example of output screens in the treatment session. K. Shaalan et al. / Expert Systems with Applications 27 (2004) 1–108 The new architecture concerning diagnosis has the following advantages: † It tries to simulate what may happen in real world in a very fine-grained manner. Each agent corresponds to a domain expert specialization. † Partitioning of domain knowledge makes it easy to maintain, easy to understand, and improve performance. † The distribution of responsibilities among agents, achieves modularity and reduced complexity. † Re-usability at the knowledge level was achieved since the proposed agents are capable of providing their services to other expert systems as well as to other applications provided that these other applications use the same ontology and understand the same communi- cation language. The expert systems agents are implemented in KROL using KADS knowledge engineering methodology. The user interface agent and the coordination agent are implemented in JAVA. Experiments on a set of test cases from real agricultural expert system were performed. The results observed in our experiments were satisfactory and proved that the system is robust. The system can diagnose and treat 26 disorders and the domain ontology consists of 66 concepts and more than 100 rules. References Chandrasekaran, B. (1988). Generic tasks as building blocks for knowl- edge-based system: the diagnosis and routine design examples. The Knowledge Engineering Review, 3(3), 183 – 210. Durfee, E. H., & Lesser, V. R (1988). Using partail global plans to coordinate distributed problem solvers. In A. H. Bond L. Gasser, & Kaufmann, (Eds.), Readings in Distributed Artificial Intelligence (pp. 285 – 293). San Mateo, CA. Durfee, E. H., & Rosenschein, J. S (1994). Distributed problem solving and multi-agent systems: Comparisons and examples. In Proceedings of the 13th international DAI workshop. El-Beltagy, S., Rafea, A., Kamel, A., Sticklen, J., Schulthess, U., & Ward, R (1995). An expert system for wheat disorders diagnosis and treatment using a hierarchical classification problem solver. Second IFAC/IFIP/ EurAgEng workshop on artificial intelligence in agriculture, Wagenin- gen, Netherlands. New York: Pergamon Press. Engelmore, R. S., & Morgan, T. (Eds.), (1988). Blackboard systems. Reading, MA: Addison-Wesley. Finin T., Fritzon R., Mckay D., & McEntire R (1994). KOML as an agent communication language. In the proceeding of the third international conference on information and knowledge management. New York: ACM Press. Available at http://www.cs.umbc.edu/KQML/papers. Finin, T., Labrou, Y., & Mayfield, J. (1997). KQML as an agent communication language. In J. M. Bradshaw (Ed.), Software agents (pp. 291 – 316). Cambridge, MA: The MIT Press. Genesereth, M. R., & Ketchpel, S. P. (1994). Software Agents Communi- cations of the ACM, 37(7), 48 – 53. Hoog, R. de., Martil, R., Wielinga, B. J., Taylor, R., Bright, C., & Velde, V. D. W (1992). The CommonKADS Model Set. Deliverable ESPRIT project P5248, KADS-II/WP I-II/RR/UvA/018/4.0. Amsterdam: Uni- versiteit van Amsterdam. Huhns, M. N., & Stephens, L. (1999). Multiagent systems and societies of agents. In G. Weiss (Ed.), Multiagent systems. A modern approach to distributed artificial intelligence (pp. 79 – 120). Cambridge, MA: MIT Press. Iglesias, C., Garijo, M., & Gonzalez, J. (1998). A survey of agent-oriented methodologies. In J. P. Müller, M. P. Singh, & A. S. Rao (Eds.), Intelligent agents: Agents theories, architectures, and languages (Vol. 1555). Lecture Notes in Computer Science, Berlin: Springer. Jennings, N. R. (2000). On agent-based software engineering. Artificial Intelligence, 117, 277 – 296. Jennings, N. R., Varga, L. Z., Aarnts, R. P., Fuchs, J., & Skarek, P. (1993). Transforming standalone expert systems into a community of Fig. 8. Output/input to the coordination agent. K. Shaalan et al. / Expert Systems with Applications 27 (2004) 1–10 9 http://www.cs.umbc.edu/KQML/papers cooperating agents. Engineering Applications of Artificial Intelligence, 6(4), 317 – 331. Jennings, N. R., & Wooldridge, M. (1995). Applying agent technology. Applied Artificial Intelligence, 9(4), 357 – 369. Jennings, N. R., & Wooldridge, M (1998). Applications of intelligent agents. Chapter 1. Agent technology: foundations, applications, and markets. Berlin: Springer. Labrou Y., & Finin T (1997). A proposal for a new KQML specification. Online at http://www.cs.umbc.edu/kqml/papers/. Moore, R., Dowding, J., Bratt, H., Gawron, J. M., Gorfu, Y., & Cheyer, A (1997). CommandTalk: A spoken-language interface for battlefield simulations. In Proceedings of the fifth conference on applied natural language processing, Washington, DC (pp. 1 – 7). Association for Computational Linguistics. Nwana, H. S. (1996). Software agents: an overview. The Knowledge Engineering Review, 11(3), 205 – 244. Nwana, H. S., & Ndumu, D. T. (1999). A perspective on software agents research. The Knowledge Engineering Review, 14(2), 1 – 18. Rafea, A (1995). On integrating agricultural expert systems with databases and multimedia. In Proceedings of the first international conference on multiple objective decision support systems for land, water, and environmental management: concepts, approaches, and applications, Honolulu, Hawaii, USA. Rafea, A. (1998). Agriculture. In J. Liebwitz (Ed.), A chapter in the handbook of applied expert systems. Boca Raton, FL: CRC Press. Rafea, A., El-Azhari, S., & Hassan E (1995). Integrating multimedia with expert systems for crop production management. In Proceeding of the second IFAC/IFIP/EnrAgEng workshop on artificial intelligence in agriculture, Netherlands. Rafea, A., El-Azhari, S., Ibrahim, I., Soliman, E., & Mahmoud, M (1995). Experience with the development and deployment of expert systems in agriculture. In Proceeding of IAAI-95, Montreal, Canada. Rafea, A., Hassan, H., & Hazman, M. (2003). Automatic knowledge acquisition tool for irrigation and fertilization expert systems. Expert System with Applications (ESWA): An International Journal, (4), 49 – 57. Rafea, A., & Mahmoud, M (2001). The evaluation and impact of NEPER wheat expert system. Fourth international workshop on artificial intelligence in agriculture IFAC/CIGR, Budapest, Hungary. Rafea, A., & Salah, A. (1994). Guiding object-oriented design via the knowledge level architecture: the irrigated wheat testbed. Mathematical and Computer Modeling, 20(8), 1 – 16. Rafea, A., Warkentin, M., & Ruth, S (1991). An expert system for cucumber production in plastic tunnels. In Proceeding of the world congress on expert systems, Florida, Orlando, USA (pp. 909 – 916). Rafea, A., Warkentin, M., & Ruth, S. (1992). Knowledge engineering: Creating expert systems for crop production management in Egypt. In C. Mann, & S. Ruth (Eds.), Expert systems in developing countries: Practice and promise (pp. 89 – 103). Boulder, CO: Westview Press. Raphael, M. J., & Deloach, S. A (2000). A knowledge base for knowledge-based multiagent system construction. National aero- space and electronics conference (NAECON), Dayton, OH, October 10 – 12. Shaalan, K., Rafea, M., & Rafea, A. (1998). KROL: a knowledge representation object language on top of prolog. Expert System with Applications (ESWA): An International Journal, 15, 33 – 46. Smith, R. G. (1980). The contract net protocol. IEEE Transactions on Computers, C-29(12). Wooldridge, M., & Ciancarini, P. (2001). Agent-oriented software engineering: The state of the art. In P. Ciancarini, & M. Wooldridge (Eds.), Agent-oriented software engineering. Lecture Notes in AI Volume, Berlin: Springer. K. Shaalan et al. / Expert Systems with Applications 27 (2004) 1–1010 http://www.cs.umbc.edu/kqml/papers/ A multiagent approach for diagnostic expert systems via the internet Introduction Overview of the proposed system The user interface agent The communication model The coordination agent The expert system agents The diagnosis agent The treatment agent An example of utilization and testing Verify three diagnosis agents: nutrition deficiency, mite, and snail Conclusions References 
work_4jkqyhqrdfgr3irbnnknpv4lli	----	Parametric-based path generation for automated vehicles at roundabouts HAL Id: hal-01674532 https://hal.inria.fr/hal-01674532 Submitted on 31 Jan 2018 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Parametric-based path generation for automated vehicles at roundabouts David Gonzalez, Joshué Pérez, Vicente Milanés To cite this version: David Gonzalez, Joshué Pérez, Vicente Milanés. Parametric-based path generation for automated vehicles at roundabouts. Expert Systems with Applications, Elsevier, 2017, 71, pp.332 - 341. �10.1016/j.eswa.2016.11.023�. �hal-01674532� https://hal.inria.fr/hal-01674532 https://hal.archives-ouvertes.fr Parametric-Based Path Generation for Automated Vehicles at Roundabouts David Gonzáleza, Joshué Pérezb, Vicente Milanésc a Author is with the Robotics and Intelligent Transportation Systems (RITS) Team, Inria of Paris, 2 rue Simone Iff, 75012 Paris, France email: david.gonzalez@inria.fr b Author is on the Automotive Area / Industry and Transport Division, Tecnalia Research and Innovation , Parque Científico y Tecnológico de Bizkaia, Derio 700, 48160 Derio, Bizkaia, Spain email: joshue.perez@tecnalia.com c Author is with the Research Department, Renault SAS, 1 Avenue du Golf, 78280 Guyancourt, France email: vicente.milanes@renault.com Abstract Urban environments are becoming more and more complex because several factors as consecutive crossroads or lanes changes. These scenarios demand specific infrastructures—i.e. roundabouts, for improving traffic flow compared with traditional intersections. A roundabout removes timeouts associated with traffic lights at crossroads and trajectory conflicts among drivers. However, it is a challenging scenario for both humans and automated vehicles. This work presents a path planning method for automated vehicle driving at roundabouts. The proposed system achieves a G1 continuous path, minimizing curvature steps to increase smoothness, dividing the driving process in three stages: entrance maneuver, driving within the roundabout and exit maneuver. Parametric equations are generated to deal with automated roundabout driving. This approach allows a real time planning considering two-lane roundabouts, taking different exits. Tests in simulated environments and on our prototype platform—Cybercar—validate the system on real urban environments, showing the proper behavior of the system. Keywords: Roundabouts, automated vehicles, path planning, intelligent transportation systems. 1. Introduction More advanced automated capabilities are being deployed on modern vehicles. Embedded sensor-based driving assistance systems are now able to deal with several traffic situations faster and safer than human drivers do (Van Schijndel-de Nooij et al., 2011). Despite of these remarkable progresses, many efforts have still to be carried out by governments and manufactures for developing systems able to intelligently deal with all the different driving scenarios. Among all these scenarios, roundabouts remain as an unsolved challenge for the development of advanced safety systems. They replace traditional intersections removing the delays associated to traffic lights (Johnson and Hange, 2003), improving the traffic flow. The roundabout is a complex driving scenario, especially for semi-automated and automated vehicles in the oncoming years (Isebrands, 2009). Roundabouts are relatively recent on urban environments, making difficult for some drivers—e.g. elderly people— to know how to deal with them1. Figure 1(a) shows the proper circulation in function of the exit according to the French road circulation code at roundabouts 2,3. DARPA Challenges in the United States from 2004 to 2007 are among the most important automated vehicle competitions. Although roundabouts were out of the defined scenarios, several driving functions were shown, including traversing intersections (Ferguson et al., 2009). First demonstrations including driving through roundabouts were carried out in 2013. Specifically, the PROUD project demo from University of Parma (Italy) on 2013 (Broggi et al., 2013); Mercedes demonstration in Germany covering the so-called Bertha route (Ziegler et al., 2014); and the AUTOPIA urban and peri-urban automated driving demo (Spain) (Godoy et al., 2015). All these demonstrations show several capabilities for automated vehicles in 1http://www.roundaboutsofbritain.com/ 2http://www.securite-routiere.gouv.fr/ 3http://www.mt.public.lu/transports/circulation/code/ Preprint submitted to Expert Systems with Applications December 5, 2016 (a) Roundabout behavior taking different exits. (b) Definitions based on roundabout design in (Rice, 2010b). Figure 1: Roundabout behavior and definition. complex scenarios. However, although roundabouts were considered on these routes, they were modeled as standard intersections, without dedicated analysis in this scenario. In the last two decades, roundabouts have been increasingly considered as a good alternative to address inter- sections, especially in urban areas (Manage et al., 2003). They provide a number of safety, operational, and other benefits when compared to other types of intersections (Tracza and Chodura, 2012). A detailed definition of the ben- efits of roundabouts in urban environment–i.e. compact size, operational efficiency, traffic safety and calming, access management, aesthetics and environmental benefits, are presented in (Rodegerdts et al., 2010) and (Rice, 2010a). From the safety point of view, roundabouts have demonstrated positive effects worldwide. In United States, Robin- son et al., (Robinson et al., 2000) concluded that they reduce the crash problems at intersections, approximately by 30%, compared with signalized intersections. In Europe, De Brabander and Vereeck (De Brabander and Vereeck, 2007) showed some conclusions on a Flanders’ traffic study between 1994 and 2000, indicating that roundabouts can reduce by 39% the number of accidents. On the negative side, the driving rules at roundabouts are unknown for some drivers, causing an increment of 28% of vulnerable road user injuries compared with classical signalized junc- tions (De Brabander and Vereeck, 2007). This fact turns roundabouts into suitable scenarios for automated driving, maintaining traffic flow benefits and removing safety concerns. From the traffic flow point of view, there are several studies in the United States and China that show how some drivers do not have enough skills experiencing in roundabouts, generating traffic jams in the vicinities (Abaza and Hussein, 2009; Bai et al., 2009). Roundabouts with and without traffic-signal controls have been intensively studied and widely used (Xiaoguang et al., 2004). Despite of these efforts, roundabout applications have been only considered from the management point of view, providing reference speed control (Bai et al., 2009; Molinete et al., 2009; Yu and Qin, 2009). From a classification point of view, a roundabout-dedicated Federal Highway Administration (FHWA) Techni- cal Summary (Rodegerdts et al., 2010) divided them according to the size and number of lanes having: 1) Mini- roundabouts–a type of intersection that can be used at physically-constrained locations in place of stop-controlled or signalized intersections to help improve safety problems and reduce excessive delays at minor approaches (Robinson et al., 2000); 2) Single-lane; and 3) Multi-lane roundabout. This work is focused on the second and third types but results can be easily adopted in mini-roundabouts without any extra complexity. Path and trajectory planning research has been very active on the intelligent vehicles field (see (González et al., 2016; Paden et al., 2016) for state-of-the-art review) in last years. However, trajectory generation in roundabouts has not been further investigated, specially in real environments. Two main demonstrations have effectively tackled roundabouts: 1) Broggi et al. (Broggi et al., 2014) presented the PROUD test, involving different driving scenarios as highways, rural roads and urban roads (involving 6 roundabouts). The planning stage was composed of a database of precomputed clothoids, from which the best path was selected, targeting the center of the lane. The road layout was detected implementing vision algorithms, as well as obstacles and other road users; and 2) Ziegler et al. (Ziegler et al., 2014) developed a planner that was based on global waypoints passed through a numerical optimization approach to plan its way through urban environments and perform the Bertha-Benz memorial route, managing several roundabouts. 2 Stereo cameras detected the road layout, the obstacles and other road users. This paper presents a general approach using geometrical roundabout definitions in a digital map (center in 2- dimensional Cartesian coordinates, the radius, the number of lanes and the entrances and exits) to build a real-time reference path using parametric equations. The path is then given to the control stage where a lateral controller is able to follow the geometrical reference. Although, different control architectures for automated vehicles have been proposed in the literature, most of them consider perception, decision and control stages as the key ones (González et al., 2016). Our approach can be easily integrated in the decision and path planning modules, because its modularity permits an adaptability of the algorithm to any automated vehicle architecture. The system was firstly tested on simu- lation; then, tests on real roads were performed, providing a real-time trajectory planning generation for roundabouts. This work is based on previous research (see (González and Pérez, 2013) and (Perez Rastelli et al., 2014) for details), the main contributions are: • G1 continuous roundabout entrance and exit by using an improved generation of the parametric control points. • Path continuity, curvature smoothness and comfort constraints considerations for in-roundabout driving. • Real time path generation considering two-lane roundabouts, including lane-change and taking different exits. • The implementation of the algorithm in a real roundabout using a real electric platform. The remainder of the paper is structured as follows. Section II presents a review of the definitions and constraints for the roundabouts, and gives an overview of the approach adopted, based on parametric curves generation. The automated roundabout algorithm will be described in section III, based on the path generation and the curve generation process. The simulator, the prototype vehicle and the test environment used for the validation of this proposal will be highlighted in section IV. Some of the results will be compared with simulations and real driving tests and discussed in section V. Finally, section VI presents some concluding remarks and future works. 2. Roundabouts: definitions and constraints This section provides a comprehensive review about possible configurations on roundabouts, according to the size, the number of lanes and geometrical characteristics. Then, a detailed description about the considerations to take into account when it comes to autonomously drive through a roundabout is given. 2.1. Definitions A roundabout can be described as a composition and convergence of several intersections. According to the specific driving area, some parameters may vary as: number of entrances/exits, lanes, speed limit, geometric design or path alignment. A detailed review about the roundabout types is found in (Rice, 2010b). Our proposal faces the roundabout driving for the planning and control point of view, identifying three stages entrance, circulatory roadway and exit (see Fig. 1(b)) as follows: • Entrance: where the entry curve is generated. It covers the route from the straight stretch up to the beginning of the circular area. It can be divided in two sub-stages: the small radius entry and the tangent or large radius (see Figure 1(b)). • Circulatory roadway: is defined as the middle stage, where generally circular shapes are used (Rice, 2010b). It can be composed by one or multiple lanes, forcing lane-change maneuvers when entering/exiting the inner lane. • Exit: is the final stage and it allows the vehicle to leave the roundabout. This stage is complementary to the entrance, having the same sub-stages: first the tangent or large radius and then the small radius exit. Other parameters can significantly affect the automated roundabout path generation. Some of them are: 1) The central island, i.e. the raised zone defined in the center of the roundabout. It defines the minimum turning radius in the circulatory roadway; 2) Splitter islands, used in the entrance and exit of roundabouts, forcing the reduction of the entry speed; and 3) Channelized approach, describing dedicated exits in the roundabout to improve traffic flow. Other traffic rules (e.g. U.K.), are easily tunable due to the modular characteristics of parametric curves used in the approach (i.e. circle and Bézier equations). 3 Straight or curve driving Entrance trajectory calculation Circulatory roadway calculation Exit trajectory calculation Straight path Bezier curves Parametric circle equations Entrance & Exit Circulatory road P3 P2 P1 P0 P0 P1 P2 P3 1 1 2 2 3 3 Entrance is close to the exit No Yes Next exit No Yes Exit finished Enable dynamic trajectory generator Roundabout detected No Yes Figure 2: Roundabout phases and parametric path generation. 2.2. Constraints There are some constraints to the proposed method imposed by both, the physical limitations of the experimental platform and the infrastructure design. Below, the main characteristics of type of roundabouts considered in this work are: • The maximum turning radius in the roundabout—considering the inner lane—has to be equal or bigger to the minimum turning radius of the vehicle with 3 meter width per lane. For the experimental vehicle used in this work, the minimum internal turning radius is 7.0 meters. • The circulatory roadway is formed by two lanes and the roundabout comprises four exits. • Perpendicular and tangential roads to the large radius of the roundabout are used. • The interaction with additional vehicles—either automated or not—is out of the scope of this work. The reader is referred to (Milanés et al., 2012) and (Muffert et al., 2012) for possible approaches addressing intersection interaction. • The reference speed is defined in function of the current stretch in which the vehicle is. It is defined based on several parameters as the curvature and comfort level in the driving process (see (Perez Rastelli et al., 2014) for details). Based on Sections 2.1 and 2.2, the following details the design and development of a path planning algorithm for automated vehicles in roundabouts. This approach is included into the Inria control architecture for automated vehicles (González and Pérez, 2013). This algorithm is able to create a real time path for dealing with any complex scenario as roundabouts. Simulations and real experiments prove the efficiency of this proposal. 3. Automated Roundabout Algorithm The proposed algorithm is developed within a global architecture for automated vehicles previously developed (see (González and Pérez, 2013) for details) where modularity plays a key role. The main contributions of this work address decision and control stages. Moreover, the information coming from the perception stage is used to know in real time the vehicle location and the obstacles around. The decision stage uses this information, based on the defined global planner, to generate a local path. Finally, the control stage follows the real-time generated path, being able to autonomously drive through different intersections, specifically roundabouts. 4 In (Perez Rastelli et al., 2014), a description of the curve generation process used for intersections is presented. This process divides the interpolating curve to provide continuous paths for the vehicle to follow. The algorithm uses reference points, e.g. the middle points of crossroads and the distance to the sidewalks, the road width and vehicle kinematic to compute the different stretches. A geometrical approach that considers the joint points between straight and curve segments was proposed. This algorithm cannot be extrapolated to roundabouts because it presents singular constraints (see Sec. II for details). The main contribution of this work is an automated path planning algorithm, based on parametric equations, for smooth path generation and tracking in roundabouts. Our approach considers several parameters as the vehicle dimensions, speed limit, comfort accelerations and characteristic of the infrastructure (i.e. small radius entry curve, tangent radius and central island). Joints between different segments in the path are at least G1 (geometrically continuous). Next sections describe the implementation carried out in the decision stage, based on the path generation process in roundabouts and the control laws used for the tracking. 3.1. G1–continuous path generation To meet roundabouts special shape, the path planning was divided in three stages: entrance, circulatory roadway and exit (see Figure 2). The left part of this figure shows the flowchart of the proposed algorithm. The first and the last stages can be modeled as standard intersections, but considering the small radius entry curve and the tangent radius, based on parametric equations as in (González and Pérez, 2013). The junction points require special attention to avoid discontinuities on the path. The circulatory roadway can be modeled as in (Pérez et al., 2011), using circular parametric equations, but also considering lane changes. The definition of entrance and exit segments is based on Bézier curve generation. The polynomial is described as follow: B(t) = n∑ i=0 Pibi,n(t); t ∈ [0, 1] (1) where n is the degree of the polynomial, and Pi are the control points (see Figure 2), and bi,n is the Bernstein basis polynomial given by: bi,n(t) = ( n i ) (1 − t)n−iti (2) In this work, Bézier curves of third and fourth degree are implemented according to geometrical characteristics for the entrance and the exit of the roundabout. k(t) = B′(t) × B′′(t)∣∣∣∣∣∣B′(t)∣∣∣∣∣∣3 (3) Equation 3 presents the computation of the curvature for B(t), being B′(t) and B′′(t) the first and second derivative of the Bézier curve B(t).In Bézier curves t represents the parameter defined between [0,1]; an iterative process generates the curve at each entry or exit stage. Right part of Figure 2 shows a vehicle in the three stages proposed for the roundabout. Once the entrance of the roundabout is detected, the first parametric equation is generated up to the tangent line of the circle. In this stage, two conditions are evaluated: if the exit is close, then the second stage is omitted, jumping directly to final stage. The circulatory roadway is generated using the circle parametric generation, as in (Pérez et al., 2011). Many techniques, as splines (Labakhua et al., 2008), clothoids (Brezak and Petrovic, 2014) and polynomials (Piazzi et al., 2002), can be used for the path planning in this scenario. However, Bézier shows several advantages, as presented in previous works (Hand et al., 2010; González and Pérez, 2013). These are resumed as follow: • The control points P0 and Pn define the beginning and the end of the curve, respectively. • The tangent vector formed by −−−−→ P0 P1 defines the initial direction of the curve, as well as −−−−−→ Pn−1 Pn defines the ending direction at the end of the curve. • The control points define the convex hull area (the Bézier curve lies within). 5 P0P0 P1P1 P2P2 P3P3 P4P4 (a) Solution for the roundabouts. (b) Kinematic model of Ackerman. Figure 3: Roundabout entry solution and vehicle model. Depending on the roundabout, different Bézier degree polynomials are generated. First, 3rd degree Béziers are tested according to an optimization function that takes into account the curvature of the generated Bezier curve, and the curvature error in the joining points of two curves. If no curve meets the requirements (not complying with road/vehicle constraints), 4th degree Béziers are implemented in a similar process to find the best curve. The extra control point coming from the curve degree elevation (P2 point in Figure 3(a)) permits to pull the curve closer to the splitter island, introducing more flexibility in the path and enabling configurations unreachable with 3rd degree Béziers. Figure 3(a) shows an example of the path generation with a 4th degree Bézier, where point P2 is moved in-between the line generated from the curb to the splitter island (yellow squares), thus forcing the Bézier to stay within bounds (see the convex-hull in grey), while still pulling the curve towards the splitter island. The proposed algorithm sets the position of Bézier control points in a dynamic way, assuring a safety path gener- ation. It is a resolution complete planner, meaning that the optimality of the solution will depend on the discretized search of the control points position (if a solution exists in the current control points search resolution, the planner is able to find it). In case the curve does not exist, the planner notifies control and decision stages (requesting safety measurements and a new strategy, respectively). Algorithm 1 describes the path generation, where k is the curvature of the evaluated point in the curve at a given iteration (i.e. Equation 3 at a given t) and kmax is the maximum curvature value. kstart and kend are the curvature continuity errors at P0 and Pn respectively, these are computed to assess the continuity of the curve w.r.t. adjacent stretches, e.g. straight segments before the roundabout and curve segments inside the roundabout (see Fig. 3(a)). The parameter n is the Bézier degree. These data are stored in an dynamic array, allocating a reward if the value of kmax is lower than the vehicle’s maximum feasible k. The best option is thus the one with the smallest reward value (minimizing the continuity error between segments). 3.2. Curve generation process in roundabouts Figure 3(a) shows the characteristics of the path generation at roundabouts. In the entrance of the roundabout, the small radius entry curve, the tangent radius and the radius of the circulatory roadway are considered in the path generation. In this sense, the splitter island’s position limits the Bézier convex-hull (gray region in Figure 3(a)) which gives a secure internal band to the vehicle since the curve lies within the hull (Perez Rastelli et al., 2014). Points P3 and 6 Algorithm 1 Path generation in roundabouts 1: Start 2: Read algorithm properties 3: Get geometrical data of the route—radius, road width, etc. 4: for n=3 ; n<=4 ; n++ do 5: Generate all control points 6: while the curve evaluation process is not completed do 7: Evaluate kstart, kend and kmax from each curve 8: if kmax is ≤ vehicle’s maximum feasible k then 9: Assign reward as max (kstart, kend ) 10: Save control points and their associated reward 11: else discard the curve 12: end if 13: end while 14: if A curve has been found then 15: Break the For loop and provide the optimal curve (minimum value for: max (kstart, kend )) 16: end if 17: end for 18: if No curve was found then 19: Inform control stage for safety measurements 20: Inform decision stage requesting new path strategy 21: end if 22: Generate the best curve (iterate through t) 23: End P4 are set over the tangent line to the circulatory roadway to ensure soft curvature transitions in the path. The vector defined by −−−−→ P0 P1 describes the direction of the entry road in Figure 3(a), being −−−−→ P3 P4 the points defining the exit stretch when leaving the roundabout. The generation of the control points is as follow: • The yellow points in Figure 3(a) describe the road layout (I1 for the splitter island and I2 for the curb), con- straining the movement of point P2 and ensuring the convex-hull does not surpasses the splitter island (the point lies within the line formed by these yellow points). • The distance from the curve (green) to the curb and the splitter island (yellow squares) has to be at least half of the vehicle’s with. If not, the curve is discarded (avoiding going over the curb). • The maximum curvature in the Bézier curve must be less than the maximum physical curvature of the vehicle, defined in Equation 3. Figure 3(a) shows the area where the control points (from P0 to P4) are established. The position of the point P0 is calculated as in Equation 4 and P4 is represented by the Equation 8: P0 = Ri + L0 Ri−1 − Ri ‖Ri−1 − Ri‖ (4) P1 = Ri + L1 Ri−1 − Ri ‖Ri−1 − Ri‖ (5) P2 = I1 + L2 I1 − I2 ‖I1 − I2‖ (6) P3 = P4 + L3 −→ T (7) 7 (a) Roundabout and Cyberbus on real implementation. 225 230 235 240 245 2 4 6 8 10 12 14 16 18 20 22 O rd in a te A xi s [m ] Abscissa Axis [m] Vehicle behavior Path Planning Circulatory roadway (b) Adjusting the controller inside the roundabout. Figure 4: Real platform and controller adjustments. P4(x,y) = r cos(arctan( Ri−1y−Riy R1−1x−Rix ) + ϕ) + Rix r sin(arctan( Ri−1y−Riy Ri−1x−Rix ) + ϕ) + Riy  (8) where L0 and L1 is the linear separation between P0, P1 respectively from Ri. Ri and Ri−1 define the roundabout center and the previous intersection respectively. Please notice that L0 is always greater than L1 to avoid unexpected curve behaviors and both, L0 and L1, are always greater than the external radius of the roundabout. L2 describes the position of P2 inside the line −−→ I1 I2. Moreover, P3 is defined by L3 which is the distance from P4 to P3 in the direction of vector −→ T , which is a vector tangent to the roundabout at P4. Please also notice that if only 3rd degree curves are needed, then P2 is removed. Moreover, ϕ is defined by Equation 9: ϕ = L4 r (9) where L4 is the segment of arch formed between −−−−−→ Ri−1Ri and P4, and r is the radius of the roundabout (Circulatory roadway). The exit of the roundabout is managed as an entrance, using symmetrical conditions. The circulatory roadway segment of the roundabout is generated by parametric circle equations, as in (Pérez et al., 2011). Having this in mind, line 5 in Algorithm 1 generates all the curves to be evaluated. The control points are constrained to linear search spaces (defined by L0, L1, L2, L3 and L4). The curve generation is done iteratively, being the limits and resolution for each of the search spaces as follows: • For the external control points P0 and P4, L0 will be constraint at maximum 20m from the external radius and L4 at 20m. • For P1, L1 will be constraint at maximum L0 distance from the external radius (forcing P1 to always be further into the intersection than P0, avoiding undesired behaviors). For P3, L3 will be limited to Ln. • If P2 exists (4th degree Béziers), L2 will be constraint to ∣∣∣∣∣∣∣∣−−→I1 I2∣∣∣∣∣∣∣∣. • Outer control points related to L0 and L4 will have a resolution of 2m (10 iterations per control point). Inner control points length L1 and L3 will also be covered in 10 iterations, having variable resolution (since they are limited by L0 and L4 respectively. Finally, L2 is covered in 5 iterations, presenting also a variable resolution depending on ∣∣∣∣∣∣∣∣−−→I1 I2∣∣∣∣∣∣∣∣. 8 3.3. Kinematic Model and lateral control law Most of the roundabouts in urban scenarios are limited to 30 km/h (Rodegerdts et al., 2010). For this reason, a kinematic model is considered for the path generation algorithm. Vehicle’s characteristics have been used for motion model in real implementation, as in (Piazzi et al., 2002) and (Sotelo, 2003). This bicycle model is also considered for the control law. The differential equations, describing the movement in a Cartesian plane (see Figure 3(b)), are as follow: ẋ(t) = d x(t)dt = Vt cos(θ(t)) ẏ(t) = dy(t)dt = Vt sin(θ(t)) θ̇(t) = dθ(t)dt = Vt l tan(φ(t)) (10) where θ is the orientation angle with respect to plane x, y, φ is the steering angle of the front wheel, L is the wheel base and Vt is the longitudinal speed. The point (x, y) are defined with respect to the center of the rear axle of the vehicle. Due to the low speed at roundabouts, the centrifugal force is considered as despicable (no slide angles nor slip angles are considered); the wheel slipping and the forces transferred between wheels of the same axle track are approximated to zero. Thus, we assume that the maximum feasible curvature is when the turning angle of the wheel is maximum. The control law is based on classic control variables for automated vehicles; such as lateral error, angular error and the curvature (Naranjo et al., 2009). It is defined as follows: U(t) = α1k(t) + α2 Lerror + α3 Aerror (11) where k(t) is the curvature, Lerror is the lateral error and Aerror is the angular error. And α1, α2 and α3 are the controller gains. These errors are calculated w.r.t. the look-ahead control point (placed at the frontal bumper of the vehicle) and the reference path. The Lerror has a proportional effect for the control action, since it is associated to the reference error in y. Moreover, Aerror has a derivative influence ( dy dt ) based on the information of Equation 10. According to the control law, it is a Proportional Derivative (PD), where α2 reduces the lateral error and α3 avoids oscillations when the vehicle is inside the roundabout. This controller allows a faster and smooth lateral action on these kind of scenarios. Figure 4(b) shows the final tuning of the controller for the real roundabout. Several turns were performed at difference speeds (between 1.0 and 4.0 m/s). The maximum error valued is around 15 cm, which is low, compared to the SLAM error (Xie et al., 2010). Regarding the comfort while driving inside the roundabout, there is not os- cillation, and due to the speed range used in the experiment, the lateral acceleration never exceeds 1.0 m/s2—Fairy uncomfortable-based on the ISO 2631-1 Standard (ISO, 1997). 4. Simulator, Prototype vehicle and Test environment Modularity is the one most important criterion in the design of the proposed control algorithms. The software environment used for the implementation is RTMaps4, a real-time tool, designed for fast and robust implementation in multitasks and multisensor data. It interfaces both, simulations and real platforms, based on a multi-thread structure. Only the acquisition and actuators components are adapted, as in (Gonzalez et al., 2014), while all algorithms remains identical. 4.1. Simulator Pro-SiVIC is a software environment for the assessment of the robustness and reliability of on-board sensors. It has a multi-sensorial environment, considering real car parameters such as the inertia, steering wheel response or lateral acceleration with yaw angles and it is compatible with RTMaps. RITS control architecture is implemented on this simulator (Gonzalez et al., 2014). Experiments were first validated using Pro-SiVIC which allows implementing a virtual vehicle in urban scenario, including roundabouts. 4http://www.intempora.com/ 9 (a) Roundabout (b) Curvature (c) Derivative of the curvature Figure 5: Behaviour over the Roundabout using different path planning methods 10 4.2. Prototype vehicle The prototype used is an electric minibus system without mechanical driving actuators. It is capable of traveling autonomously from one point to another in a road network. It has been used in some European project demonstration, as CityMobil and CityMobil II (Bouraoui et al., 2011; Roldao et al., 2015). Simultaneous Localization and Mapping (SLAM) fused with an Inertial Measurement Unit (IMU) is used for the localization of the vehicle. The positioning algorithm is capable of estimate the vehicle location using real time information (see (Trehard et al., 2014) for details). Figure 4(a) shows the prototype vehicle. 4.3. Test environment Figure 4(a) shows test track used for the real experiments. The facilities are at INRIA-Rocquencourt, France. The main characteristics of this roundabout are: • The radius of the roundabout is 8.5 meters. • It has four entrances/exits, 3 meters wide each. • The roundabout has two lanes inside, 3 meters each. • The flow direction is counter-clockwise. 5. Experiments Experiments are focused on showing the path planning generation for the three phases on the roundabout. Different entrances and exits of the roundabout were carried out, including lane change inside the roundabout. The lateral acceleration, the curvature, its derivative and the lateral and angular errors are used for evaluating the proposed method. 5.1. Simulations Simulations were carried out with RTMaps and ProSivic (see section 4). Figure 5 shows a comparison using three different methods: 1) based on static path generation (blue line); 2) based on the introduction of road constraints for path generation (González and Pérez, 2013) (dotted red line); and 3) the algorithm proposed in this paper (black line). The road limits and the road verge are depicted in gray and dotted gray respectively. The first method was tested on curve stretches giving a proper behavior. It fixes the polynomial control points to a certain distance. One can appreciate how such generation causes that the generated path to cross over the sidewalk. The second method (dotted red line) modifies the control points location according to the road constraints, forcing the path to be into the road. This second approach causes a discontinuity on the curvature (see 5(b)) that is also reflected on its derivative 5(c). The last method automatically sets the control points based on Bézier properties and constraints explained in section 3, achieving a smoother path planning. Figures 5(b) and 5(c) show the curvature and their derivative per each experiment. The static path generation (blue line) presents fewer but more abrupt changes. The second algorithm (dotted red line) shows a curvature overshoot (more pronounced in its derivative), causing higher lateral accelerations and discomfort. Finally, the proposed method (black line) shows a better behavior for changes of the curvature, improving the comfort. Figure 6 shows the behavior of the proposed algorithm when taking different exits using both inner and outer lanes. For the scenario when the first exit is taken, the path only considers Bézier curves as in a standard intersection. This is because the exit is too close to the entrance, making the circulatory roadway negligible. The other two trajectories depict the scenario when the second exit is taken. Despite the external lane is used when taking the second exit, the use of the inner lane allows to show the curvature evolution. Figure 6(c) shows the curvature in the three scenarios shown in Figure 6(a). For the first case, the curvature is smoother and smaller than in the second and third cases. It is because the behavior is similar than a standard intersection and there are no inflection points in the curvature. However, the curvature is a little higher when taking the inner lane. Experiment 2 shows the curvature generation taking the external lane. The derivative change has a 11 −20 −10 0 10 20 30 40 50 0 10 20 30 40 50 60 O rd in a te A xi s [m ] Abscissa Axis [m] [1]−>Taking Exit 1 −External Lane− [2]−>Taking Exit 2 −External Lane− [3]−>Taking Exit 2 −Internal Lane− 1 3 2 (a) Planning in the Roundabout -Exit 1 and 2- −20 −10 0 10 20 30 40 50 0 10 20 30 40 50 60 O rd in a te A xi s [m ] Abscissa Axis [m] [4]−>Taking Exit 3 −External Lane− [5]−>Taking Exit 3 −Internal Lane− [6]−>Taking Exit 4 −External Lane− [7]−>Taking Exit 4 −Internal Lane− 5 4 6 7 (b) Planning in the Roundabout -Exit 3 and 4- 0 10 20 30 40 50 60 70 80 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 C u rv a tu re p ro fil e [ 1 /m ] Distance [m] [1]−>Taking Exit 1 −External Lane− [2]−>Taking Exit 2 −External Lane− [3]−>Taking Exit 2 −Internal Lane− (c) Curvature -Exit 1 and 2- 0 10 20 30 40 50 60 70 80 90 100 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2 C u rv a tu re p ro fil e [ 1 /m ] Distance [m] [4]−>Taking Exit 3 −External Lane− [5]−>Taking Exit 3 −Internal Lane− [6]−>Taking Exit 4 −External Lane− [7]−>Taking Exit 4 −Internal Lane− (d) Curvature -Exit 3 and 4- 0 10 20 30 40 50 60 70 80 −0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02 0.03 0.04 0.05 C u rv a tu re d e ri va tiv e [ 1 /m 2 ] Distance [m] [1]−>Taking Exit 1 −External Lane− [2]−>Taking Exit 2 −External Lane− [3]−>Taking Exit 2 −Internal Lane− (e) Derivative of the curvature -Exit 1 and 2- 0 10 20 30 40 50 60 70 80 90 100 −0.1 −0.08 −0.06 −0.04 −0.02 0 0.02 0.04 0.06 0.08 0.1 C u rv a tu re d e ri va tiv e [ 1 /m 2 ] Distance [m] [4]−>Taking Exit 3 −External Lane− [5]−>Taking Exit 3 −Internal Lane− [6]−>Taking Exit 4 −External Lane− [7]−>Taking Exit 4 −Internal Lane− (f) Derivative of the curvature -Exit 3 and 4- Figure 6: Experiments with different radii, lane changes and exits better behavior on the experiment 3 (see Fig. 6(e)). This shows how the comfort of the drivers is better in the third case, because the reduction of the lateral accelerations inside the roundabout. Experiments taking the third and fourth exits following the French Road Circulation Code are shown in Figures 6(b),6(d) and 6(f). Figure 6(b) shows the Cartesian position of the trajectories generated. Curvature and its derivative are shown in figure 6(d) and 6(f) respectively. At the beginning of Figure 6(d), the entrance of the roundabout is shown. When the outer lane is taken (see tests 4 and 6 in Figure 6(b)), the the vehicle steers more than when taking the inner lane (i.e. tests 5 and 7 in Figure 6(b)), causing a higher curvature in the entry and exit stages. In both cases, 12 the use of the inner lane has a better behavior. The covered distance is smaller (see Fig. 6(d)) and the derivative of the curvatures have smoother changes during the whole maneuver (see Fig. 6(f)). (a) Inria’s facilities (b) Inria’s facilities (zoom) 0 50 100 150 200 time [s] -0.6 -0.4 -0.2 0 0.2 0.4 C o n tr o l v a ri a b le s Control variables Lateral error Angular error Curvature (c) Lateral control variables 110 115 120 125 130 135 time [s] -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 C o n tr o l v a ri a b le s Control variables Lateral error Angular error Curvature (d) Lateral control variables (zoom) 0 50 100 150 200 time [s] -0.3 -0.2 -0.1 0 0.1 0.2 A cc e le ra tio n [ m /s ²] Onboard IMU Axis X Axis Z Axis Y (e) On-board IMU 105 110 115 120 125 130 135 time [s] -0.2 -0.15 -0.1 -0.05 0 0.05 A cc e le ra tio n [ m /s ²] Onboard IMU Axis X Axis Z Axis Y (f) On-board IMU (zoom) Figure 7: Real urban scenario: Roundabout and intersections. 13 5.2. Real tests The proposed path planning for roundabouts was implemented on a Cyberbus and tested at Inria’s facilities (see Figure 7(a)). The chosen route contains straight segments, urban intersections and a roundabout. This experiment was carried at 8km/h, which is a normal velocity for these kind of platforms (Cybernetic Transportation Systems—CTS— please see (González and Pérez, 2013)). It shows how the vehicle (black line) follows the reference generated by the path planning (in red). Figure 7(c) shows the behavior of the control variables, displaying the lateral and angular error and the curvature. The lateral error is low and it keeps within a low value of 25cm, see Figure 7(d) (keep in mind that the implemented SLAM system has an error of about 15cm). Figure 7(e) shows the evolution of the accelerations read from the on- board IMU. The continuous line shows the lateral acceleration, which increase when the vehicle enters the roundabout (see Figure 7(f)), but always keeping a comfortable driving (accelerations under 0.35m/s2 (ISO, 1997)). Figure 7(b) shows the real-time roundabout path generation and depicts the three stages: entrance, circulatory roadway and exit. The entrance stage differs from simulations in that the entry angle is tangent rather than perpen- dicular to the circulatory roadway. This angle information is taken from the digital global map. In Figure 7(d) the curvature (in grey) grows up to 1/r (from 113 to 118 seconds approximately ), where r is the radius of the roundabout (8.5 meters for the present case). The circulatory roadway keeps the curvature constant. The exit stage shows how the generation of the curve can handle curved exit paths. The curvature of the exit stage is depicted between seconds 123 to 130 of Figure 7(d). Please notice that it reaches the curvature value of the next curved segment and a G1 behavior depicted at the curve joint in Figure 7(b). The curve generation time for each intersection is around 2ms (milliseconds) for more than 2000 curves evaluated in a core duo Intel system (2.0GHz). 6. CONCLUSIONS This paper presents a new path-planning algorithm for roundabouts. It generates a real-time path considering the specific roundabout constraints and the vehicle dynamics at low speeds. The proposed method is based on parametric equations, showing good results in a real implementation on a Cyberbus. The proposed real-time path planning relies on road and vehicle constraints. Specifically, it considers the radius, central point of the roundabout and the exit to be taken on the roundabout. The path-planning generation was split in three stages: entrance, circulatory roadway and exit. Each part has a specific path generation based on parametric curves. Emergency situation can be handled through obstacle avoidance modules and a real time recalculation of the path (see (Gonzalez et al., 2014) for details). Different exits of the roundabout were considered based on the French Road Circulation Code showing a com- fortable performance, validating the proposed algorithm. Future work will be focused on the interaction with other vehicles on the roundabout. Acknowledgments Authors wants to express their gratitude to the FP7 EU AutoNet2030 project for its support in the development of this work. References Abaza, O. A., Hussein, Z. S., September 2009. Comparative analysis of multilane roundabout capacity "case study". In: IEEE 70th Vehicular Technology Conference Fall (VTC 2009-Fall). pp. 1–5. Bai, Y., Xue, K., Yang, X., May 2009. Block mechanism of left-turned flow at signal-controlled roundabout. In: WRI Global Congress on Intelligent Systems. Vol. 3. pp. 443–449. Bouraoui, L., Boussard, C., Charlot, F., Holguin, C., Nashashibi, F., Parent, M., Resende, P., 2011. An on-demand personal automated transport system: The citymobil demonstration in la rochelle. IEEE Intelligent Vehicles Symposium 4, 1086 – 1091. Brezak, M., Petrovic, I., April 2014. Real-time approximation of clothoids with bounded error for path planning applications. IEEE Transactions on Robotics 30 (2), 507–515. Broggi, A., Buzzoni, M., Debattisti, S., Grisleri, P., Laghi, M., Medici, P., Versari, P., Sept 2013. Extensive tests of autonomous driving technologies. IEEE Transactions on Intelligent Transportation Systems 14 (3), 1403–1415. Broggi, A., Cerri, P., Debattisti, S., Laghi, M. C., Medici, P., Panciroli, M., Prioletti, A., 2014. Proud-public road urban driverless test: Architecture and results. In: 2014 IEEE Intelligent Vehicles Symposium Proceedings. IEEE, pp. 648–654. 14 De Brabander, B., Vereeck, L., 2007. Safety effects of roundabouts in flanders: Signal type, speed limits and vulnerable road users. Accident Analysis and Prevention. Ferguson, D., Howard, T., Likhachev, M., 2009. Motion planning in urban environments. In: Buehler, M., Iagnemma, K., Singh, S. (Eds.), The DARPA Urban Challenge. Vol. 56 of Springer Tracts in Advanced Robotics. Springer Berlin Heidelberg, pp. 61–89. Godoy, J., Perez Rastelli, J., Onieva, E., Villagra, J., Milanés, V., Haber, R., 2015. A driverless vehicle demonstration on motorways and in urban environments. Transport. González, D., Pérez, J., 2013. Control architecture for cybernetic transportation systems in urban environments. In: IEEE Intelligent Vehicles Symposium. pp. 1119–1124. Gonzalez, D., Perez, J., Lattarulo, R., Milanes, V., Nashashibi, F., Oct 2014. Continuous curvature planning with obstacle avoidance capabilities in urban scenarios. In: IEEE 17th International Conference on Intelligent Transportation Systems (ITSC). pp. 1430–1435. González, D., Pérez, J., Milanés, V., Nashashibi, F., 2016. A review of motion planning techniques for automated vehicles. IEEE Transactions on Intelligent Transportation Systems 17 (4), 1135–1145. Hand, L., Yashiro, H., Nejad, H., Do, Q. H., Mita, S., June 2010. Bézier curve based path planning for autonomous vehicle in urban environment. IEEE Intelligent Vehicles Symposium 4, 1036 – 1042. Isebrands, H., 2009. Roundabouts and signals: Harmony even with increasing traffic volumes current guidance. ITE Journal. ISO, 1997. Mechanical vibration and shock-evaluation of human exposure to whole-body vibration-part 1: General requirements. Tech. Rep. ISO 2631-1, International Organization for Standardization. Johnson, M., Hange, W., 2003. Modern roundabout intersections: When to use them? a comparison with signalized intersections. Tech. rep., Institute of Transportation Engineers. Labakhua, L., Nunes, U., Rodrigues, R., Leite, F. S., 2008. Smooth trajectory planning for fully automated passengers vehicles: spline and clothoid based methods and its simulation. In: Informatics in Control Automation and Robotics. Springer, pp. 169–182. Manage, S., Nakamura, H., Suzuki, K., 2003. Performance an of roundabouts as an alternative for intersection control in japan. Journal of the Eastern Asia Society for Transportation Studies 5, 871–883. Milanés, V., Villagra, J., Godoy, J., Simó, J., Pérez, J., Onieva, E., 2012. An intelligent v2i-based traffic management system. IEEE Transactions on Intelligent Transportation Systems 13 (1), 49–58. Molinete, B., Bouraui, L., Naranjo, J., Kostense, H., Hendriks, J., Alonso, J., Lobrino, R., Isasi, L., 2009. Cybercars-2: Close communications for cooperation between cybercars. Tech. rep., Technical Report Project No IST-2004-0228062. Muffert, M., Milbich, T., Pfeiffer, D., Franke, U., June 2012. May i enter the roundabout? a time-to-contact computation based on stereo-vision. In: Intelligent Vehicles Symposium (IV), 2012 IEEE. pp. 565–570. Naranjo, J. E., Bouraoui, L., García, R., Parent, M., , Sotelo, M. A., March 2009. Interoperable control architecture for cybercars and dual-mode cars. In: IEEE Transactions on Intelligent Transportation Systems. Vol. 10. pp. 146 – 154. Paden, B., Cap, M., Yong, S. Z., Yershov, D., Frazzoli, E., 2016. A survey of motion planning and control techniques for self-driving urban vehicles. arXiv preprint arXiv:1604.07446. Pérez, J., Milanés, V., de Pedro, T., Vlacic, L., 2011. Autonomous driving manoeuvres in urban road traffic environment: a study on roundabouts. In: Proceedings of the 18th World Congress The International Federation of Automatic Control. pp. 13795–13800. Perez Rastelli, J., Lattarulo, R., Nashashibi, F., June 2014. Dynamic trajectory generation using continuous-curvature algorithms for door to door assistance vehicles. In: IEEE Intelligent Vehicles Symposium Proceedings. pp. 510–515. Piazzi, A., Lo Bianco, C. G., Bertozzi, M., Fascioli, A., Broggi, A., August 2002. Quintic g2-splines for the iterative steering of vision-based autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems 3, 27 – 36. Rice, E., 2010a. Mini-roundabouts. Tech. rep., U.S. Departmet of Trasportation Federal Highway Administration (FHWA-SA-10-007). Rice, E., 2010b. Roundabouts. Tech. rep., U.S. Departmet of Trasportation Federal Highway Administration (FHWA-SA-10-006). Robinson, B. W., Rodegerdts, L., Scarbrough, W., Kittelson, W., Troutbeck, R., Brilon, W., Bondzio, L., Courage, K., Kyte, M., Mason, J., Flannery, A., Myers, E., Bunker, J., G., J., 2000. Roundabouts: An informational guide. Tech. rep., Report FHWA-RD-00-067. FHWA, U.S. Department of Transportation. Rodegerdts, L., Bansen, J., Tiesler, C., Knudsen, J., Myers, E., 2010. Roundabouts: An informational guide. Tech. rep., National Cooperative Highway Research Program. Roldao, L., Perez Rastelli, J., González, D., Milanés, V., 2015. Description and technical specification of cybernetic transportation systems: an urban transportation concept -citymobil2 project approach-. In: IEEE ITSC. pp. 176–181. Sotelo, M. A., 2003. Lateral control strategy for autonomous steering of ackerman-like vehicles. Robotics and Autonomous Systems 45, 223 – 233. Tracza, M., Chodura, J., 2012. Performance and safety of roundabouts with traffic signals. In: SIIV 5th International Congress Sustainability of Road Infrastructures. pp. 789 – 800. Trehard, G., Alsayed, Z., Pollard, E., Bradai, B., Nashashibi, F., Sept 2014. Credibilist simultaneous localization and mapping with a lidar. In: IEEE/RSJ International Conference onIntelligent Robots and Systems (IROS 2014). pp. 2699–2706. Van Schijndel-de Nooij, W., Krosse, B., Van den, T., Maas, S., Van Nunen, E., Zwijnenberg, H., Schieben, A., Mosebach, H., Ford, N., McDonald, M., Jeffery, D., Piao, J., Sanchez, J., 2011. Definition of necessary vehicle and infrastructure systems for automated driving. Tech. rep., Study Report European Commission. Xiaoguang, Y., Xiugang, L., Kun, X., 2004. A new traffic-signal control for modern roundabouts method and application. IEEE Intelligent Trans- portation Systems Society 5, 282 – 287. Xie, J., Nashashibi, F., Parent, M., Favrot, O. G., December 2010. A real-time robust global localization for autonomous mobile robots in large environments. In: 11th International Conference on Control Automation Robotics & Vision (ICARCV). Singapore, pp. 1397 – 1402. Yu, L., Qin, C., Sept 2009. Real-time signal control method for multi-approach roundabouts. In: International Conference on Management and Service Science (MASS). pp. 1–4. 15 Ziegler, J., Bender, P., Schreiber, M., Lategahn, H., Strauss, T., Stiller, C., Dang, T., Franke, U., Appenrodt, N., Keller, C., Kaus, E., Herrtwich, R., Rabe, C., Pfeiffer, D., Lindner, F., Stein, F., Erbs, F., Enzweiler, M., Knoppel, C., Hipp, J., Haueis, M., Trepte, M., Brenk, C., Tamke, A., Ghanaat, M., Braun, M., Joos, A., Fritz, H., Mock, H., Hein, M., Zeeb, E., Summer 2014. Making bertha drive-an autonomous journey on a historic route. IEEE Intelligent Transportation Systems Magazine 6 (2), 8–20. 16 
work_4jkq3wam25a6lnvvdjedgrfube	----	ARTIFICIAL INTELLIGENCE 135 The Organization of Expert Systems* A Tutorial Mark Stefik, Jan Aikins, Robert Balzer, John Benoit, Lawrence Birnbaum, Frederick Hayes-Roth, Earl Sacerdoti** Xerox Palo Alto Research Center, Palo Alto, CA 94304, U.S.A. Recommended by Daniel G. Bobrow ABSTRACT This is a tutorial about the organization of expert problem -solving programs. We begin with a restricted class of problems that admits a very simple organization. To make this organization feasible it is required that the input data be static and reliable and that the solution space be small enough to search exhaustively. These assumptions are then relaxed, one at a time, in case study of ten more sophisticated organizational prescriptions. The first cases give techniques for dealing with unreliable data and time-varying data. Other cases show techniques for creating and reasoning with abstract solution spaces and using multiple lines of reasoning. The prescriptions are compared for their coverage and illustrated by examples from recent expert systems. 1. Introduction Twenty years ago, Newell [29] surveyed several organizational alternatives for problem solvers. He was concerned with how one should go about designing problem-solving systems. Many techniques have been developed in artificial intelligence (henceforth Al) research since then and many examples of expert systems have been built. Expert systems are problem-solving programs that * There is currently much interest and activity in expert systems both for research and applications. A forthcoming book edited by Hayes-Roth, Waterman, and Lenat [21]provides abroad introduction to thecreation andvalidation ofexpert systems for ageneralcomputerscience audience. An extended version of this tutorial, which introduces concepts and vocabulary for an audience without an AT background, will appear as achapter in the book. ** Additional affiliations: 3. Aikins, Hewlett-Packard, Palo Alto, CA; R. Balzer, USC/Information Sciences Institute, Marina del Rey, CA; J. Benoit, The MITRE Corporation, McLean, VA; L. Birnbauni, Yale University, New Haven, CT; F. Hayes-Roth, Teknowledge, Palo Alto, CA; E. Sacerdoti, Machine Intelligence Corp., Palo Alto, CA. Artificial Intelligence 18 (1982) 135-173 0004.3702/82/0000-0000/$02.75 © 1982 North-Holland 136 M. STEFIK ET AL. solve substantial problems generally conceded as being difficult and requiring expertise. They are called knowledge based because their performance depends critically on the use of facts and heuristics used by experts. Expert systems have been used as a vehicle for Al research under the rationale that they provide a forcing function for research in problem solving and a reality test. Recently some textbooks have appeared that organize principles of Al (e.g., [301) and give examples of advanced programming techniques (e.g., [6]). However, there is no guidebook for an expert systems designer to the issues and choices in designing a system. Furthermore, an unguided sampling of expert systems from the literature can be quite confusing. Examples are scattered in various journals, conference proceedings, and tech- nical reports. Systems with seemingly similar tasks sometimes have radically different organizations, and seemingly different tasks are sometimes performed with only minor variations on a single organization. The variations reflect the immaturity of the field in which most of the systems are experimental. From the diversity of experiments one would like to extract alternatives and prin- ciples to guide a designer. This tutorial is organized as follows: Section 2 is a catalog of some generic expert tasks. For each task there is a checklist of requirements and key problems associated with expert performance of the task. The purpose of this section is to extract a set of architecturally relevant issues that cover a variety of problem-solving tasks, Section 3 presents the substance of the organizational ideas. It addresses the issues from Section 2 and makes prescriptions. 2. A Characterization of Expert Tasks In this section we will consider several generic tasks that experts perform. Examining these tasks will help us to understand what makes expert reasoning difficult. The difficulties provide a guide to architectural relevance; they enable us to focus on issues that relate to critical steps in reasoning. Interpretation Interpretation is the analysis of data to determine their meaning. Example. Interpretation of mass spectrometer data [2]. In this case, the data are measurements of the masses of molecular fragments and interpretation means the determination of one or more chemical structures. Requirements. Find consistent and correct interpretations of the data. It is often important that analysis systems be rigorously complete, that is, that they consider the possible interpretations systematically and discard candidates only when there is enough evidence to rule them out. Key problems. Data are often noisy and errorful, that is, data values may be missing, erroneous, or extraneous. EXPERT SYSTEMS TUTORIAL 137 (1) This means that interpreters must cope with partial information. (2) For any given problem, the data may seem contradictory. The interpreter must be able to hypothesize which data are believable. (3) When the data are unreliable, the interpretation will also be unreliable. For credibility it is important to identify where information is uncertain or incomplete and where assumptions have been made. (4) Reasoning chains can be long and complicated. It is helpful to be able to explain how the interpretation is supported by the evidence. Diagnosis Diagnosis is the process of fault-finding in a system (or determination of a disease state in a living system) based on interpretation of potentially noisy data. Example. Diagnosis of infectious diseases [34]. Requirements. Requirements include those of interpretation. A diagnostician must understand the system organization (i.e., its anatomy) and the relation- ships and interactions between subsystems. Key problems. (1) Faults can sometimes be masked by the symptoms of other faults. Some diagnostic systems ignore this problem by making a single fault assumption. (2) Faults can be intermittent. A diagnostician sometimes has to stress a system in order to reveal faults. (3) Diagnostic equipment can itself fail. A diagnostician has to do his best with faulty sensors. (4) Some data about a system are inaccessible, expensive, or dangerous to retrieve. A diagnostician must decide which measurements to take. (5) The anatomy of natural systems such as the human body is not fully understood. A diagnostician may need to combine several (somewhat in- consistent) partial models. Monitoring Monitoring means to continuously interpret signals and to set off alarms when intervention is required. Example. Monitoring a patient using a mechanical breathing device after surgery [17]. Requirements. A monitoring system is a partial diagnostic system with the requirement that the recognition of alarm conditions be carried out in real time. For credibility, it must avoid false alarms. Key problems. What constitutes an alarm condition is often context-depen- dent. To account for this, monitoring systems have to vary signal expectations with time and situation. 138 M.STEFIKETAL. Prediction Prediction means to forecast the course of the future from a model of the past and present. Example. Predicting the effects of a change in economic policy. (Some plan- fling programs have a predictive component. There is currently an opportunity to develop expert prediction programs in a variety of areas.) Requirements. Prediction requires reasoning about time. Predictors must be able to refer to things that change over time and to e~ventsthat are ordered in time. They must have adequate models of the ways that various actions change the state of the modeled environment over time. Key problems. (1) Prediction requires the integration of incomplete infor- mation. When information is complete, prediction is not an Al problem (e.g., where will Jupiter be two years from next Thursday). (2) Predictions should account for multiple possible futures (hypothetical reasoning), and should indicate sensitivity to variations in the input data. (3) Predictors must be able to make use of diverse data, since indicators of the future can be found in many places. (4) The predictive theory may need to be contingent; the likelihood of distant futures may depend on nearer but unpredictable events. Planning A plan is a program of actions that can be carried out to achieve goals. Planning means to create plans. Example. Experiment planning in molecular genetics [36]. Requirements. A planner must construct a plan that achieves goals without consuming excessive resources or violating constraints. If goals conflict, a planner establishes priorities. If planning requirements or decision data are not fully known or change with time, then a planner must be flexible and oppor- tunistic. Since planning always involves a certain amount of prediction, it has the requirements of that task as well. Key problems. (1) Planning problems are sufficiently large and complicated that a planner does not immediately understand all of the consequences of his actions. This means that the planner must be able to act tentatively, so as to explore possible plans. (2) If the details are overwhelming, he must be able to focus on the most important considerations. (3) In large complex problems, there often are interactions between plans for different subgoals. A planner must attend to these relationships and cope with goal interactions. EX1~ERTSYSTEMS TUTORIAL 139 (4) Often the planning context is only approximately known, so that a planner must operate in the face of uncertainty. This requires preparing for contingencies. (5) If the plan is to be carried out by multiple actors, coordination (e.g. choreography) is required. Design Design is the making of specifications to create objects that satisfy particular requirements. Example. Designing a digital circuit. (This is an area of increased interest and activity in expert systems.) Requirements. Design has many of the same requirements as planning. Key problems. (1) In large problems, a designer cannot immediately assess the consequences of design decisions. He must be able to explore design pos- sibilities tentatively. (2) Constraints on a design come from many sources. Usually there is no comprehensive theory that integrates constraints with design choices. (3) In very large systems, a designer must cope with the system complexity by factoring the design into subproblems. He must also cope with interactions between the subproblems, since they are seldom independent. (4) When a design is large, it is easy to forget the reasons for some design decisions and hard to assess the impact of a change to part of a design. This suggests that a design system should record justifications for design decisions and be able to use these justifications to explain decisions later. This is especially apparent when subsystems are designed by different designers. (5) When designs are being modified, it is important to be able to reconsider the design possibilities. During redesign, designers need to be able to see the ‘big picture’ in order to escape from points in the design space that are only locally optimal. (6) Many design problems require reasoning about spatial relationships. Reasoning about distance, shapes, and contours demands considerable com- putational resources. We do not yet have good ways to reason approximately or qualitatively about shape and spatial relationships. Several issues appear repeatedly across this catalog of expert tasks: Large solution spaces. In interpretation problems like the mass spec- trometery example [2], some problems require millions of possible chemical structures to be considered. In planning and design tasks, the number of reasonable solutions is usually a very small fraction of a very large number of possible solutions. In each of these tasks, the size and characterization of the solution space is an important organizational parameter. Tentative reasoning. Many diagnostic procedures profitably employ assump- tions about the number of faults or about the reliability of sensors. Part way 140 M. STEFIK ET AL. through diagnosis, it may be discovered that these assumptions are unwar- ranted. This places a premimum on the ability to undo the effects of the assumptions. Similarly, in design and planning tasks it is often appropriate because of scale to employ simplifying assumptions (e.g., abstractions). In any given design, some of the assumptions will fail in a design, so there is an incentive to employ methods that facilitate the reworking of assumptions and trade-offs during iterations of the design process. Time-varying data. Patient monitoring and diagnosis tasks are concerned with situations that evolve over time—as diseases follow their natural course or as treatments are administered. Noisy data. Sensors often yield noisy data. This is a factor for any task involving reasoning from measurements such as interpretation, diagnostic, and monitoring tasks. The next section considers organizational prescriptions that deal with each of these issues. 3. Knowledge Engineering Prescriptions Feigenbaum [18] defines the activity of knowledge engineering as follows. “The knowledge engineer practices the art of bringing the principles and tools of Al research to bear on difficult applications problems requiring experts’ knowledge for their solution. The technical issues of acquiring this knowledge, representing it, andusing it appropriately to construct and explain lines-of-reasoning, are important problems in the design of knowledge-based systems The art of constructing intelligent agents is both part of and an extension of the programming art. It is the art of building complex computer programs that represent and reason with knowledge of the world.” [18, pp. 1014—1016] This section is intended as a prescriptive guide to building expert systems. To illustrate the strengths and limitations of organizational alternatives we will cite a number of contemporary systems. In presenting these examples we adopt a level of detail that is adequate for making the ideas clear yet avoiding the particulars of the task and the programming implementation. The reader seeking a more detailed discussion of implementation is encouraged to consult a textbook on Al programming (e.g., [6]). One of the most variable characteristics of expert systems is the way that they search for solutions. The choice of search method is affected by many characteristics of a domain, such as the size of the solution space, errors in the data, and the availability of abstractions. Inference is at the heart of a reasoning system and failure to organize it properly can result in problem- solvers that are hopelessly inefficient, naive, or unreliable. As a consequence of this, search is one of the most studied topics in artificial intelligence. EXPERT SYSTEMS TUTORIAL 141 Our pedagogical style is to start with a very restricted class of problems that admits a very simple search process. We will articulate the domain restrictions under which this organization is applicable, and thereby expose its limitations. Then we will relax the requirements on the task domain and introduce ameliorating techniques as architectural prescriptions. Fig. I shows a chart of 2 I 3 I 4 I No Evaluator for Partial Solution Fixed Order of Abstracted Steps 6 State~trigge red Expectations FIG. 1. Case 1 begins with a restricted class of problems that admits a very simple organization. These assumptions are relaxed oneat a time in the other cases. ~1 Requirements —p. Prescriptions .—--+ Small Solution Space Data Reliable & Fixed Reliable Knowledge Exhaustive Search Monotonic Reasoning Single Line of Reasoning I IUnreliable Dataor Knowledge Combining Evidence from Multiple Sources Probability Models Fuzzy Models Exact Models Time.Varying Data I Big, Factorable Solution Space Hierarchical Generate and Test 5 9 1 Representation Method Too Inefficient Tuned Data Structures Knowledge Compilation Cognitive Economy Single Line of Reasoning Too Weak Multiple Lines of Reasoning 10 Single Knowledge Source Too Weak Heterogenous Models Opportunistic Scheduling Variable-Width Search No Fixed Sequence of Subproblems Abstract Search Space 7 Subproblerns Interact Constraint Propagation Least Commitment 8 Efficient Guessing is Needed Belief Revision for Plausible Reasoning 142 M. STEFIK ETAL. the cases that we will consider. Each box in the figure corresponds to one of the cases and the numbering indicates the order in which the cases are discussed. The lines connecting the boxes organize the cases into a tree structure such that a sequence of cases along a branch corresponds to increas- ingly elaborate considerations of a basic idea. The first three branches consider the complications of unreliable data or knowledge, time-varying data, and a large search space. Any given problem may require combining ideas from any of these topics. The problem of a large search space is then considered along three major branches. The first branch (cases 5 through 8) considers organiza- tions for abstracting asearch space. The second branch focuses on methods for incomplete search. The third branch considers only ways to make the knowl- edge base itself more efficient. This breakdown is mainly pedagogical. Real systems may combine these ideas. 3.1. Case 1—Small search space with reliable knowledge and data Systems for complex tasks are generally more complicated than systems for simple tasks. In this section we will consider a very simple architecture which has been used for relatively simple applications. We begin by listing the requirements for task simplicity: (1) The data and knowledge should be reliable. (2) The data and knowledge should be static. (3) The space of possible solutions should be small. On the surface these requirements may seem quite mild. Indeed, there is a widely held belief among people who have not looked closely into problem solving that most problems satisfy these requirements. After all, for many problems the facts seem straightforward and there are not that many solutions to consider. Under closer examination, however, most real tasks fail to meet these requirements including the examples of expert tasks listed in Section 2. The first requirement is that the data and knowledge be reliable. Reliable data are not noisy or errorful. There can be no extraneous signals and no missed signals (e.g., due to sampling). In real applications few sources of data meet these requirements. In addition to data reliability, the knowledge must be reliable. Reliable knowledge is applicable without concern about consistency or correctness. Systematic application of reliable knowledge should not lead to false, approximate, or tentative conclusions. The main advantage of reliability for both data and knowledge is the monotonicity of the system. In the simplest architecture, the memory is a monotonic data base to which conclusions are added by the reasoning system as they are inferred. No provisions need to be made for retraction of facts pending new information. It is enough to develop a single line of reasoning; that is, there is no need to develop multiple arguments to support potential conclusions. If more than one inference rule is applicable at a given time, the order in which they are applied is unimportant. EXPERT SYSTEMS TUTORIAL 143 The second requirement is intended to avoid the problem of reasoning with time-dependent data. This means that the system need not be concerned with invalidating facts as time passes. The requirement that the search space should be small implies that no provisions need to be made to cope with the limitations of computational resources. There need be no concern about computationally efficient data structures or for avoiding combinatorial explosions. It doesn’t matter whether the search is for one solution or all possible solutions as long as the space is small. If the search is exhaustive, the maximum size of the search space depends on the time it takes to consider a single solution. A useful number to keep in mind for this maximum is ten factorial (10!). If 25 milliseconds are required to consider a solution, then 10! solutions can be considered sequen- tially during a full twenty-four hour day. This is often a practical upper limit for exhaustive search. The surprise is that the ceiling is so low. An organization for solving these problems has two main parts: a memory and an inference method. The simplest organization of the memory would be a list of inferred facts (i.e., beliefs). For many problems, the beliefs can be represented in the predicate calculus1 such as (On BlockI Block2), (NOT (On BIock2 Table-i)). Some systems attempt to optimize the storage format of the data. For example, in frame systems (see [3]) the indexing of facts is organized to make the most common access paths more efficient. Data which are used together are stored together in frames. In the following sections we explore some more sophisticated organizations that will enable us to relax these restrictions on the problems. 3.2. Case 2—Unreliable data or knowledge Experts sometimes make judgments under pressure of time. All the data may not be available; some of the data may be suspect; some of the knowledge for interpreting the data may be unreliable. These difficulties are part of the normal state of affairs in many interpretation and diagnostic tasks. The general problem of drawing inferences from uncertain or incomplete data has invited a variety of technical approaches. One of the earliest and simplest approaches to reasoning with uncertainty was incorporated in the MYCIN expert system [7, 34] for selecting antibiotic therapy for bacteremia. One of the requirements for MYCIN was to represent judgmental reasoning such as “A suggests B” or “C and D tend to rule out E”. To this end, MYCIN introduced a model of approximate implication using In our examples we use the prefix form of the notation because of its obvious similarity to list notations as in LISP. i~w M.STEFIKETAL. numbers called certainty factors to indicate the strength of a heuristic rule. The following is an example of a rule from MYCIN’S knowledge base. If (1) the infection is primary-bacteremia and (2) the site of the culture is one of the sterile sites and (3) the suspected portal of entry of the organism is the gastro-intestinal tract, then there is suggestive evidence (.7) that the identity of the organism is bacteroides. The number ‘.7’ in this rule indicates that the evidence is strongly indicative (0.7 out of 1) but not absolutely certain. Evidence confirming a hypothesis is collected separately from that which disconfirms it, and the ‘truth’ of the hypothesis at any time is the algebraic sum of the evidence. This admits the combination of evidence in favor and against the same hypothesis. The introduction of these numbers is a departure from the exactness of predicate calculus. In MYCIN, things are not just true or false; reasoning is inexact and that inexactness is numerically characterized in the rules by an expert physician. Facts about the world are represented as 4-tuples cor- responding to an atomic formula with a numeric truth value. For example, (IDENTITY ORGANISM-2 KLEBSIELLA .25) is interpreted as ‘The identify of organism-2 is Kiebsiella with certainty 0.25’. In predicate calculus, the rules of inference tell us how to,combine wffs and truth values. MYCIN has its own way to combine formulas. When the premise of a rule is evaluated, each predicate returns a number between 1 and —1 (—1 means ‘definitely false’). MYCIN’s version of AND performs a minimization of its arguments; OR performs a maximization of its arguments. This results in a numerical value between —1 and 1 for the premise of a rule. For rules whose premise values surpass an empirical threshold of 0.2, the rule’s conclusion is made with a certainty that is the product of the premise value and the certainty factor of the rule. These rules of combination can be shown to have certain properties—such as insensitivity to the order in which the rules are applied. MYCIN’S certainty factors are derived from probabilities but have some distinct differences (see [34]). A reasonable question about such approaches is whether they are un- necessarily ad hoc. A commonly voiced criticism is that MYCIN introduces its own formalism for reasoning with uncertainty when thereare thoroughly-studied probabilistic approaches available. For example, Bayes’ Rule could be used to calculate the probability (e.g., of a disease) in light of specified evidence, from the a priori probability of the disease and the conditional probabilities relating the observations to the diseases. The main difficulty with Bayes’ Rule is the large amount of data that are required to determine the conditional prob- abilities needed in the formula. This amount of data is so unwieldy that the EXPERT SYSTEMS TUTORIAL 145 conditional independence of observations is often assumed. It can be argued that such independence assumptions undermine the rigorous statistical model. A middle ground which replaces observations with subjective estimates of prior probabilities has been proposed by Duda, Hart, and Nilsson [14] and analyzed for its limitations by Pednault, Zucker, and Muresan [31]. Another approach to inexact reasoning that diverges from classical logic is fuzzy logic as discussed by Zadeh [40] and others. In fuzzy logic, a statement like ‘X is a large number’ is interpreted as having an imprecise denotation characterized by a fuzzy set. A fuzzy set is a set of values with corresponding possibility values as follows. Fuzzy Proposition: X is a large number. Corresponding fuzzy set: (XE [0,10], .1), (XE [10,1000], .2), ({X> 1000}, .7). The interpretation of the proposition ‘X is large’ is that ‘X might be less than 10’ with possibility 0.1, or between 10 and 1000 with possibility 0.2, and so on. The fuzzy values are intended to characterize an imprecise denotation of the proposition. Fuzzy logic deals with the laws of inference for fuzzy sets. Its utility in reasoning about unreliable data depends on the appropriateness of interpreting soft data (see Zadeh [41]) as fuzzy propositions. There is little agreement among Al researchers on the utility of these modified logics for intelligent systems, or even on their advantages for reasoning with incomplete data. The pseudo-probability and fuzzy approaches for reasoning with partial and unreliable data depart from the predicate calculus by introducing a notion of inexactness. Other approaches are possible. For example, the use of exact inference methods on unreliable data in an expert system is illustrated by the GAl program reported by Stefik [37]. GAl is a data interpretation system which copes with errorful data. It exploits the redundancy of experimental data in order to correct errors that it may contain. GAl infers DNA structures from segmentation data. GAl’S task is to assemble models of complete DNA structures given data about pieces (called segments) of the structures. The segment data are produced by chemical processes which break DNA apart in predictable ways. In a typical problem, several in- dependent breaking processes called digestions are performed and the resulting segments are measured. These digestions give independent measurements of the DNA molecule. For example, independent estimates of the molecular weight can be computed by summing the weights of the segments in any of the ‘complete’ digestions. (A digest is called complete if all of the molecules have been cleaved in all possible places by the enzymes.) 146 M. STEFIK ET AL. An example of a rule for correcting missing data is: If a segment appears in a complete digestion for an enzyme that fails to appear in the incomplete digestion for that enzyme, it may be added to the list of segments for the incomplete digestion. This rule is based on the observation that segments are easier to overlook in incomplete digestions than in complete digestions. This rule places more confidence in data from complete digestions than from incomplete ones. Other data correction rules incorporate more elaborate reasoning by modeling pre- dictable instrument error such as failure to resolve measurements which are too close together. Such rules enable GA! to look to the data for evidence of instrument failure. In summary, several methods for reasoning with unreliable data and knowl- edge have been proposed. The probability-related and possibility methods use modified logics to handle approximations. They use numerical measures for combining evidence. In contrast, data correction rules can reason with partial information without compromising the exactness of predicate calculus. All of these methods depend on the formalization of extra meta-knowledge in order to correct the data, take back assumptions, or combine evidence. The availability of this meta-knowledge is a critical factor in the viability of these approaches to particular applications. A special method for contending with both fallible knowledge and limited computational resources will be considered in Section 3.10. 3.3. Case 3—Time-varying data Some expert tasks involve reasoning about situations that change over time. One of the earliest approaches in Al to take this into account was the situational calculus introduced by McCarthy and Hayes for representing sequences of actions and their effects [25]. The central idea is to include ‘situations’ along with the other objects modeled in the domain. For example, the formula (On Block-I Table-2 Situation-2) could represent the fact that in Situation-2, Block-i is on Table-2. A key feature of this formulation is that situations are discrete. This discreteness reflects the intended use of this calculus in robot planning problems. A robot starts in an initial situation and performs a sequence of actions. After each action, the state of the robot’s world is modeled by another situation. In this representation, a situation variable can take situations as values. In some implementations, the actual situation variable is usually left implicit by index- ing the formulas according to situations. Actions in the situational calculus are represented by functions whose EXPERT SYSTEMS TUTORIAL 147 domains and ranges are situations. For each action, a set of frame axioms characterizes the set of assertions (i.e., ‘the frame’) that remain fixed while an action takes place within it. In a robot planning task, an example of an action would be the Move action for moving an object to a new location. A frame axiom for Move would be that all objects not explicitly moved are left in their original location. While many issues can be raised about this approach to representing changing situations, many Al systems have used it for a variety of tasks with only minor variations. Sometimes the changes of situation are signalled by time-varying data, rather than by the autonomous actions of a robot. An example of this is shown by the VM system reported by Fagan [16, 17]. VM (for Ventilator Manager) is a program that interprets the clinical significance of patient data from a physiological monitoring system. VM monitors the post-surgical progress of a patient requiring mechanical breathing assistance. A device called a mechanical ventilator provides breathing assistance for seriously ill patients. The type and settings of the ventilator are adjusted to match the patient’s needs. As the patient’s status improves, various adjustments and changes are made, such as replacing the mechanical ventilator with a ‘T-piece’ to supply oxygen to the patient. In VM’s application, a patient’s situation is affected by the progression of disease and the response to therapeutic interventions. For such applications, the model of clinical reasoning must account for information received from tests and observations over time. VM illustrates knowledge suitable for coping with time-varying data. This knowledge in VM is organized in terms of several kindsof rules: transition rules, initialization rules, status rules, and therapy rules. Periodically VM receives a new set of instrument measurements and then it reruns all of its rules. Transition rules are used to detect when the patient’s state has changed, e.g., when the patient starts to breathe on the T-piece. The following is an example of a transition rule. If (1) the current context is ‘Assist’ and (2) respiration rate has been stable for 20 minutes and (3) l/E ratio has been stable for 20 minutes, then the patient is on ‘CMV’ (controlled mandatory ventilation). This rule governs the transition between an ‘Assist’ context and a ‘CMV’ context. When the premise of a transition rule is satisfied in VM, a new ‘context’ is entered. These contexts correspond to specific states or situations. When a context is changed, VM uses initialization rules to update its in- formation for the new context (e.g., expectations and unacceptable limits for the measurements). These rules refer to the recent history of the patient to establish new expectations and information for the new context. Part of one 148 M. STEFIK ET AL. such rule follows: If (I) the patient transitioned from ‘Assist’ to ‘T-piece’ or (2) the patient transitioned from ‘CMV’ to ‘T-piece’ then expect the following: Very low Low {~-Acceptable——] f—Ideal—] Mm Max High Very high SYS 110 150 DIA 60 95 MAP 6() 75 110 120 Pulse rate 60 120 ECO2 22 28 30 40 45 50 v~’sreasoning about time is limited to adjacent time intervals. It is concer- ned only with the previous state and the next state. Its mechanisms for dealing with this are (I) state-triggered expectations and (2) rules for dynamic belief revision. The transition rules in VM govern changes of context. Data arrives periodically, but context is changed only when it is adequately supported by the evidence. The initialization rules are essentially like the frame axioms— establishing what changes and what stays the same in the new context. Once a context is set, the expectations are used to govern VM’s behavior until the context is changed again. Programs which need to reason about more distant events require more elaborate representations of events and time. For example, planning and prediction tasks require reasoning about possible futures. For these ap- plications, the situational calculus must be extended to allow for multiple possible futures with undetermined operations, unordered sets of possible future events, and the possible actions of uncontrolled multiple actors. While Al systems capable of such sorts of reasoning seem within reach, their construction is still a research enterprise. 3.4. Case 4—Large but factorable solution space In the restricted class of problems that we started with in Section 3.1, it was stipulated that: (1) The data and knowledge must be reliable. (2) The data must be static. (3) The search space must be small. We have already discussed some techniques for relaxing the first two requirements. In this section we will begin the consideration of techniques for coping with very large search spaces. EXPERT SYSTEMS TUTORIAL 149 In many data analysis tasks, it is not enough to find just one interpretation of the data. It is often desirable to find every interpretation that is consistent with the data. This conservative attitude is standard in high risk applications such as the analysis of poisonous substances or medical diagnosis. A systematic ap- proach would be to consider all possible cases, and to rule out those that are inconsistent with the data. This approach is called reasoning by elimination and has been familiar to philosophers for years, but it has often been regarded as impractical. The difficulty is that there is often no practical way to consider all of the possible solutions. The DENDRAL program [2] is probably the best known Al program that reasons by elimination (using generate and test). The key to making it work is to incorporate early pruning into the generate and test cycle. This section illustrates some of the characteristics of problems on which this approach will work, and gives examples of the kinds of knowledge that are needed. Since the problem area for the DENDRAL program is rather complicated for tutorial purposes, we will consider a simpler expert system, GAl [37], that was already mentioned in Section 3.2. Like DENDRAL, GAl is a data interpretation program that infers a complete molecular structure from measurements of molecular pieces. Fig. 2 shows a Complete A Digest: 5 5 Complete B Digest: 3 7 Complete A&B Digest: 1 2 3 4 FIG. 2. Enzyme A cleaves thecircular molecule at thepoints labeled A. Enzyme B cleaves it at the points labeled B. The table lists the fragments that would be observed under ideal digestion experiments. The Data 150 M. STEFIK ET AL. simple example of the kind of data that GA! would have about a molecule. The top part of the figure shows that a molecule is made up of segments of measureable length. The lines labeled A and B that cut across the circle indicate the sites where the molecule is cleaved by enzymes (named A and B, respectively). All of the molecules that GAl j5 concerned with are linear or circular. This means that all of the molecular segments are linear pieces that can he arranged end to end. When a sample of molecules is completely digested by an enzyme, pieces are released whose sizes can be measured. The goal is to infer the structure of the original molecule from the digest data. Sometimes more than one molecular structure is consistent with the available data. A primary task in problems like this is to create a workable generator of all of the possible solutions (i.e., all of the possible molecules). In GA!, the first step is to apply data correction rules as shown in Section 3.2. Then GAl determines an initial set of generator constraints—a set of segments and enzyme sites for building candidate molecules. The rules for deriving the list will not be elaborated here, but they make conservative use of molecular weight estimates and redundant data from several digests. The generation process then begins by combining these segments and sites, and testing whether the combinations fit the evidence. For example, the following lists (among others) correspond to the complete molecule in Fig. 2. (1 A 2 B 3 A 4 B), (2 B 3 A 4 B 1 A), (1 B 4 A 3 B 2 A). These equivalent representations can he generated by starting with any of the four segments in the picture of the molecule in Fig. 2, and reading off the sites and segments around the circle either clockwise or counterclockwise. This provides us with eight equivalent representations for the same molecule. A generator is said to be nonredundant if it produces exactly one of the equivalent representations of a solution (the canonical form) during the generation process. GAl does this by incorporating rules for pruning non- canonical structures during the generation process. An example of such a rule follows: If circular structures are being generated, only the smallest segment in the list of initial segments should he used for the first segment. The key to effective use of generate and test is to prune classes of in- consistent candidates as early as possible. For example, consider the following structures from the generation process for the sample problem: (1 B 2 B). GA! treats this as a description of all the molecules that match the pattern EXPERT SYSTEMS TUTORIAL 151 (1 B 2 B (Segment) (Site) (Segment) (Site)) where any of the remaining segments and sites may be filled in the template. It is easy to see that no molecule matching this template is consistent with the data in Fig. 2—because all such molecules would yield a segment of length 2 in the complete digest for B. Other pruning considerations are more global. When such pruning rules exist, the solution space is said to be factorable. GAl has been run on problems where the number of possible candidates is several billion. However, the pruning rules are so effective at eliminating classes of solutions that most problems require only a few seconds of computer time. After the generator is finished, usually twenty or thirty candidates make it through all of the pruning rules. Only one or two of these will be consistent with all of the data. In principle, it is possible to add more pruning rules to GA! so that only the consistent solutions remain. However, the rules needed to do this become increasingly complex and specialized. It becomes difficult to prove the correctness of such rules (so that solutions will not be missed) and to ensure that the rules are faithfully represented in the program. There is also a point of diminishing returns when each new specialized rule covers a smaller number of cases. In GA! this problem is addressed by applying a digestion process model to the final candidates and comparing its predictions with the observed data. A simple scoring function then penalizes candidates that predict extra or missing segments. Because the number of disagreements between the idealized digests of two different molecules diverges rapidly for small molecular differences, it was not necessary to tune the scoring function to recognize wrong solutions. In summary, generate-and-test is an appropriate method to consider when it is important to find all of the solutions to a problem. For the method to be workable, the generator must partition the solution space in ways that allow for early pruning. These criteria are often associated with data interpretation and diagnostic problems. 3.5. Case 5—No evaluator for partial solutions There are many problems involving large search spaces for which generate and test is a method of last resort. The most common difficulty is that no generator of solutions can be found for which early pruning is viable. Design and planning problems are of this nature. One usually cannot tell from a fragment of a plan or design whether that fragment is part of a complete solution; there is no reliable evaluator of partial solutions expressed as solution fragments. In this section we will consider the first of several approaches to problem solving without early pruning. These approaches have in common the idea of abstracting the search space but differ in their assumptions about the nature of that space. Abstraction emphasizes the important considerations of a problem and enables its partitioning into subprohlems. In the simplest case there is a 152 M. STEFIK ET AL. fixed partitioning in the abstract space which is appropriate for all of the problems in an application. This case is illustrated by the Ri program reported by McDermott [27]. RI’S area of expertise is the configuring of Digital Equipment Corporation’s VAX computer systems. Its input is a customer’s order and its output is a set of diagrams displaying the spatial relationships among the components on the order. This task includes a substantial element of design. In order to determine whether a customer’s order is satisfactory, Rl must determine a spatial configuration for the components and add any necessary components that are missing. The configuration task can be viewed as a hierarchy of subtasks with strong temporal interdependencies. RI partitions the configuration task into six ordered subtasks as follows. (1) Determine whether there is anything grossly wrong with the customer’s purchase order (e.g., mismatched items, major prerequisites missing). (2) Put the appropriate components in the cpu and cpu expansion cabinets. (3) Put boxes in the unibus expansion cabinet and put the appropriate components in those boxes. (4) Put panels in the unibus expansion cabinets. (5) Lay out the system on the floor. (6) Do the cabling. The actions within each subtask are highly variable; they depend on the particular combination of components in an order and on the way these components have been configured so far. Associated with each subtask in Rl is a set of rules for carrying out the subtask. An example of a rule for the third subtask follows: If the most current active context is assigning a power supply and a unibus adaptor has been put in a cabinet and the position it occupies in the cabinet (its nexus) is known and there is space available in the cabinet for a power supply for that nexus and there is an available power supply and there is no H7i01 regulator available, then add an H7i01 regulator to the order. Ri has about 800 rules about configuring VAX systems. Most of the rules are like the example above. They define situations in which some partial con- figuration should be extended in particular ways. These rules enable Rl to combine partial configurations to form an acceptable configuration. They indicate what components can (or must) be combined and what constraints EXPERT SYSTEMS TUTORIAL 153 must be satisfied in order for the combinations to be acceptable. They make use of a data base describing properties of about 400 VAX components. Other rules describe the temporal relationships between subtasks by determining their ordering. (These rules are analogous to the transition rules described for VM in Section 3.3 except that the rules monitor the state of RI’S problem solving instead of data from external sensors.) The approach that RI uses is called Match. It is one of Newell’s weak methods for search [28]. Match enables Ri to explore the space of possible configurations with the basic operations of creating the extending partial configurations. Match explores this space by starting in an initial state, going through intermediate states, and stopping in a final state without any back- tracking. Each state in the space is a partially instantiated configuration. Rl proceeds through its six major tasks in the same order for each problem; it never varies the order and it never backs up in any problem. The benefit of the abstraction space is that Ri needs to do very little search. The conditions that make Match viable are both its source of power and its weakness. The key requirement is that there can be no backtracking. This means that at any intermediate state, RI must be able to determine whether the state is on a solution path. This requires that there must exist a partial ordering on decisions for the task such that the consequences of applying an operator bear only on ‘later’ parts of the solution. It is interesting that Match is in fact insufficient for the complete task in Rl. The subtask of placing modules on the unibus is formulated essentially as a bin-packing problem—namely how to find an optimal sequence that fits within spatial and power constraints. No way of solving this problem without search is known. Consequently Rl uses a different method for this part of the problem. In summary, the use of abstractions should be considered for applications where there is a large search space but no method for early pruning. Ri is an example of a system which uses a fixed abstract solution. Within this frame- work, it uses the Match weak method to search for a solution. Whether Match is practical for an application depends on how difficult it is to order the intermediate states. 3.6. Case 6—No fixed partitioning of subproblems When every example problem in an application can be usefully partitioned into the same subproblems, then the organization described in the previous section should be considered. In applications with more variety to the problems, no fixed set of subproblems can provide a useful abstraction. For example, planning domains such as errand running (see [22]) require plans rich with structure. To be useful, abstractions must embody the variable structure of the plans. 154 M.STEFIKETAL. In this section we will consider an approach called top-down refinement that tailors an abstraction to fit each problem. The following aspects of the approach are important. (1) Abstractions for each problem are composed from terms (selected from a space of terms) to fit the structure of the problem. (2) During the problem-solving process, these concepts represent partial solutions that are combined and evaluated. (3) The concepts are assigned fixed and predetermined abstraction levels. (4) The problem solution proceeds top down, that is, from the most abstract to the most specific. (5) Solutions to the problem are completed at one level before moving down to the next more specific level. (6) Within each level, subproblems are solved in a problem-independent order. (This creates a partial ordering on the intermediate abstract states.) The best known example of a program using this approach is the ABSTRIPS program reported by Sacerdoti [33]. ABSTRIPS was an early robot planning program. It made plans for a robot to move objects (e.g., boxes) between rooms. A design goal for ABSTRIPS was to provide abstractions sufficiently different from the detailed ‘ground’ space to achieve a significant improve- ment in problem-solving efficiency, but sufficiently similar so that the mapping down from abstractions would not be time-consuming. This led to an interes- ting and simple approach for representing abstractions. Abstractions in ABSTRIPS are plans. They differ from ground level plans only in the level of detail used to specify the preconditions of operators. This level of detail is indicated by associating a number (termed a criticality value) with all of the literals used in preconditions. For example, Sacerdoti suggested the following criticality assignments in a robot planning domain: Type andColor 4 InRoom 3 Pluggedln and Unplugged 2 NextTo 1 In this example, the predicates for Type and Color of objects are given high criticalities, since the robot has no operator for changing them. These predi- cates together with the set of robot actions are combined to form plans for solving particular problems; the space of possible plans is the set of all of the plans that can be built up from these pieces. The most abstract plans are the ones that include only the higher criticality concepts. Planning in AB5TRIP5 starts by setting criticality to a maximum. In it, planning within each level proceeds backwards from goals. Preconditions whose criti- cality is below the current level are invisible to the planner, since it is presumed that they will be accounted for during a later pass. After a plan is completed at one level, the criticality level is decremented and planning is started on the EXPERT SYSTEMS TUTORIAL 155 lower level. The previous abstract version of the plan is used to guide the creation of the next level. For example, an early version of a plan may determine the route that the robot takes through the rooms. In more detailed versions, steps for opening and closing doors are included. In this way, the abstract plans converge to the specific plan. The sequence of abstract plans is created differently for each problem. ABSTRIPS was a great advance over its predecessor STRIPS, which lacked the hierarchical planning ability. Generally, when hierarchical and non-hierarchical approaches have been systematically compared, the former have dominated. ABSTRIPS was substantially more efficient than STRIPS, and the effect increased dramatically as longer plans were tried. Since then, many other hierarchical planning programs have been created. In most of the later programs, the abstraction concepts have simply been arranged in a hierarchy, without actually assigning them criticality numbers. In summary, the interesting feature of top-down refinement is the flexibility of the abstractions. Abstraction states are individually constructed to fit each problem in the domain. In contrast to Match, top-down refinement places only a partial ordering on the intermediate states of the problem-solver. Still, there are some important conditions about problem solving in the domain of ap- plication that aie inherent in the method. The basic assumption is that a small fixed amount of problem-solving knowledge about criticality levels and top- down generation is sufficient. Furthermore, it must be possible to assign a partial criticality ordering to the domain concepts. What is important for one problem must be important for all of the problems. The next section suggests some ways to relax these requirements. 3.7. Case 7—Interacting subproblems One basic difficulty with top-down refinement is the lack of feedback from the problem-solving process. It is presumed that the same kinds of decisions should be made at the same point (i.e., criticality level) for each problem in the domain. In this section, we will explore an approach that is based on a different principle for guiding the reasoning process called the least-commitment prin- ciple. The basic idea is that decisions should not be made arbitrarily or prematurely. They should be postponed until there is enough information. Reasoning based on the least-commitment principle requires the following abilities. (1) The ability to know when there is enough information to make a decision. (2) The ability to suspend problem-solving activity on a subproblem when the information is not available. (3) The ability to move between subproblems, restarting work as infor- mation becomes available. (4) The ability to combine information from different suhproblems. 156 M. STEFIK ET AL. Level 1 Paint the ceiling and paint the ladder. Level 2 Level 3 Get Paint H Get Iadder~~~pply_paint to ceiling._K______ Split J Join Get Paint I—J~~ply_paintto ladder. / Level 3 (after conflict resolution.) Get Paint H Get ladder. H Apply paint to ceiling._H Is~itI ______ ____ ____________ Paint J—~_~oi2j—___—{_Applypaint to ceiling. FIG. 3. Example of planning in NOAH. NOAH analyzes the interactions between the steps in order to assign them an ordering in time. In this example, the ‘painting the ladder’ is seen to be in conflict with using it. To complete both goals, NOAH decides to paint the ceiling first. Later processing will factor out common subplans like ‘getpaint’. Fig. 3 shows an example of this style of reasoning from the NOAH system reported by Sacerdoti [32]. NOAH was a robot planning system that used a least-commitment approach to assign a time-ordering to operators in a plan. Earlier planning programs inserted operators into a plan as they worked backwards from goals. In contrast, NOAH assigned the operators only a partial ordering and added specifications for a complete ordering of the operators only as required. In Fig. 3, NOAH starts with two subgoals: paint the ceiling and paint the ladder. Plans for the two subgoals are expanded in parallel and a conflict is found. If the ladder is painted first, it will be wet andwe won’t be able to paint the ceiling. In other words, the step to paint the ladder violates a precondition (that the ladder be usable) for the step to paint the ceiling. This interference EXPERT SYSTEMS TUTORIAL 157 between the subgoals provides NOAH with enough information to order the tasks. If it had arbitrarily ordered the steps and painted the ladder first, it would have had to plan around the wet ladder, perhaps waiting for it to dry. The resulting plan would not have been optimal. Another example of this idea was given in MOLGEN reported by Stefik [36]. MOLGEN is an expert system that used this style of reasoning for designing molecular genetics experiments. MOLGEN’S organization involved the following features: (1) Interactions between nearly independent subproblems are represented as constraints. (2) Interactions between subproblems are discovered via constraint propagation. (3) MOLGEN uses explicit problem-solving operators (as opposed to domain- specific operations) to reason with constraints. (4) MOLGEN alternates between least-commitment and heuristic strategies in problem-solving. In the least commitment strategy, MOLGEN makes a choice only when its available constraints sufficiently narrow its alternatives. Its problem-solving operators are capable of being suspended so that a decision could be post- poned. Constraint propagation is the mechanism for moving information between subproblems. It enables MOLGEN to exploit the synergy between decisions in different subproblems. In contrast with ABSTRIPS, strict backward expansion of plans within levels, MOLGEN expands plans opportunistically in response to the propagation of constraints. The fourth feature illustrates an interesting limitation of the least commit- ment principle. Every problem-solver has only partial knowledge about solving problems in a domain. With only the least-commitment principle, the solution process must come to a halt whenever there are choices to be made, but no compelling reason for deciding any of them. We call this situation a least- commitment deadlock. When MOLGEN recognizes this situation, it switches to its heuristic approach and makes a guess. In many cases, aguess will be workable, and the solution process can continue to completion. In other cases, a bad guess can lead to conflicts. The number of conflicts caused by (inaccurate) guessing is a measure of the incompleteness of the problem-solving knowledge. Conflicts can also arise from the least-commitment process in cases where the goals are fundamentally unattainable. In summary, the least-commitment principle coordinates decision-making with the availability of information and moves the focus of problem-solving activity among the available subproblems. The least-commitment principle is of no help when there are many options and no compelling reasons forchoices. In these cases, some form of plausible reasoning is necessary. In general, this approach uses more information to control the problem-solving process than the top-down refinement approach. 158 M. STEFIK ET AL. 3.8. Case 8—Guessing is needed Guessing or plausible reasoning is an inherent part of heuristic search. For example, the generator in a generate and test system guesses about partial solutions so that they can be tested. A more subtle example is a problem-solver based on top-down refinement, which implicitly assumes that it will be able to refine its higher abstractions to specific solutions. Some examples follow of generic situations in reasoning where guessing is important: (1) Many problem-solvers need to cope with incomplete knowledge and may be unable to determine the best choice at some stage in its problem solving. In such cases, a problem-solver is unable to finish without making a guess. Examples of this are assumptions introduced as a first step in hypothetical reasoning and assumptions introduced to break aleast-commitment deadlock as discussed in the previous section. (2) A search space may be quite dense in solutions. If solutions are plentiful and equally desirable and there is no need to get them all, guessing is efficient. (3) Sometimes there is an effective way to converge to solutions by sys- tematically improving approximations. (Top-down refinement is an example of this.) In cases where convergence is rapid, it may be appropriate to guess even when solutions are rare. The difficulty with guessing is in identifying wrong guesses and recovering from them efficiently. This section considers how plausible reasoning can benefit from particular architectural features. One of the best known systems with architectural provisions for guessing is the EL system for circuit analysis reported by Stallman and Sussman [35]. EL analyzes analog electrical circuits. It has two main methods that are described below—forward reasoning and the method of ‘assumed states’. The assumed states provide the examples of guessing. Forward reasoning with electrical laws is used to compute electrical parameters (e.g., voltage or current) at one node of a circuit from parameters at other nodes. EL uses only a few laws, such as Ohm’s Law which defines a linear relationship between voltage and current for a resistor, and Kirchhoff’s current law which states that the current flowing out of a node equals the current flowing into it. Much of EL’S power derives from two things: (1) the exhaustive application of these laws and (2) the ability to reason with these laws symbolically as shown in Fig. 4. This figure illustrates a circuit in which resistors are connected in a ladder arrangement. The analysis task is to determine the voltages and resistances at all of the nodes of the circuit. The interesting aspect of this is that symbolic reasoning about the circuit is much simpler than writing and solving equations for the series and parallel resistor network. Analysis begins with the intro- duction of a variable e to represent the unknown node voltage at the end of the ladder. This yields acurrent e/5 through resistorR6. Then by Kirchhoff’s Law we EXPERT SYSTEMS TUTORIAL 159 Ri 5 ohms Ri 5 ohms Before anal~J~ R3 5 ohms AfterAnaly.~ R3 5 ohms R5 5 ohms R5 5 ohms R6 ohms FIG. 4. Symbolic propagation of electrical parameters. Analysis begins by assigning the symbol e to the unknown voltage at theupper right corner of the ladder. Other values are derived by stepwise application of Ohm’s and Kirchhoff’s laws. have the same current through R5 which gives us avoltage 2e on the left of the resistor. This voltage across R4 allows us to compute the current through it using Ohm’s Law. The application of electrical laws in terms of symbolic unknowns continues until all of the voltages are defined in terms of e. At that time we have 8e = 10 volts, e = 5/4 volts. Sometimes circuit analysis requires the introduction of more than one variable to represent unknown circuit parameters. In general, the analysis evolts 160 M.STEFIKETAL. involves two main processes: i-step deductions and coincidence. The i-step deductions are direct applications (sometimes symbolic) of the electrical laws. A coincidence occurs when a 1-step deduction is made which assigns a value to a circuit parameter that already has a value (symbolic or numeric). At the time of the coincidence, it is often possible to solve the resulting equation for one variable in terms of the others. This allows EL to eliminate unknowns. The propagation method can be extended to any devices where the electrical laws are invertible and where the algebra required for the symbolic reasoning is tractable. Unfortunately, many simple and important electrical devices, such as inverters and transistors, are too complicated for this approach. For example, a diode is approximately represented by exponential equations. Electrical engineering has an approach for these devices called the method of assumed states. This is where guessing enters into EL’S problem solving. The method of assumed states uses a piecewise linear approximation for complicated devices. The method requires making an assumption about which linear region a device is operating in. EL has two possible states for diodes (on or off) and three states for transistors (active, cutoff, and saturated). Once a state is assumed, EL can use tractable linear expressions for the propagation analysis as before. After making an assumption, EL must check whether the assumed states are consistent with the voltages and currents predicted for the devices. Incorrect assumptions are detected by means of a contradiction, which is the event in which chosen assumptions are seen to be inconsistent. When this happens, then some of the assumptions need to be changed. Intelligent processing of con- tradictions involves determining which assumptions to revise. Implementing this idea in the problem-solver led to the following important architectural features: (i) queue-based control; (2) dependency-directed backtracking. The operators in EL that perform the propagation analysis are called demons. They are placed in queues and run sequentially by a scheduler. When demons run, they make assertions in the data base and then return to the scheduler. This data base activity causes other demons whose triggers match the assertions to be added to the queue. EL has three queues for DC analysis with different priorities. The lowest priority queue is used for device-state assumptions. These demons are given a low priority so that the immediate consequences of an assumption will be explored before more assumptions are made. The middle queue is used for most of the electrical laws. The high priority queue is used for demons that detect contradictions. These demons are given a high priority so that invalid assumptions will be detected before too much computational work is done. These demons trigger the dependency-directed backtracking. EL keeps dependency records of all of its deductions and assumptions. In El., EXPERT SYSTEMS TUTORIAL 161 an assertion is believed (or in) if it has well-founded support from atomic assumptions. An assertion without such support is said to be out. If an out assertion returns to favor, it is said to be unouted. Fig. 5 shows an example of this process in a data base. Al, Bl, and Cl are atomic data that are currently in. Suppose that Al and A2 are mutually exclusive device-state assumptions. The top of Fig. 5 shows which facts are in when Al is in. Arrows are used to indicate support. (Assertions following from A2 are shown in dotted lines to show that they are in the data base, but that they are out.) The bottom figure shows what happens if Al is outed and A2 is unouted. An important aspect of EL’S problem-solving is its ability to recover from tentative assumptions. The details of the implementation and knowledge will not be covered here, since there are many ways to approach this problem. The main points are: (1) In the event of a contradiction, EL needs to decide what to withdraw. It is (a) /——~fl / FIG. 5. Example of belief revision in EL. ~I~hcdark boxes arc in and the lighter ones are out. In (a), Al is in and so all of its consequences are also in. In (b), Al is out but A2 is in. (b) 162 M. STEFIK ET AL. not effective to simply withdraw all of the assumptions that are antecedants of the contradictory assertion. EL must decide which of the assumptions are most unlikely to change and this requires domain-specific knowledge. (2) EL must redo some of the propagation analysis. Sometimes it is possible to salvage some of the symbolic manipulation (e.g., variable elimination) that has been done. EL has special demons that decide carefully what to forget. (3) Contradictions are remembered so that choice combinations that are found to be inconsistent are not tried again. These ideas were the intellectual precursors to work on belief revision systems.2 Belief revision can he used for reasoning with assumptions or defaults. For a problem solver to revise its beliefs in response to new knowl- edge, it must reason about dependencies among its current set of beliefs. New beliefs can be the consequences of new information received or derived. A critical issue in this style of reasoning is well-founded support and there are some pitfalls for the unwary involving cyclical support structures. An important question is “what mechanisms should be used to resolve ambiguities when there are several possible revisions?” It is clear this choice needs to be controlled, but the details for making the decision remain to be worked out. In the examples above, we used knowledge about justifications to reason about choices. Doyle [13] has proposed a style of dialectical argumentation where the primary step is to argue about the kinds of support for beliefs. In such system the complexity of knowledge about belief revision would itself require a substantial knowledge base. Every approach depends critically on the kinds of dependency records that are created and saved. This work is at the frontier of current Al research. In summary, EL 15 an example of a program with organizational provisions for plausible reasoning. It uses symbolic forward reasoning for analyzing circuits. To analyze complicated devices EL has to assume linear operating regions. It uses dependency-directed backtracking so that it can recover from incorrect assumptions. It uses a priority-oriented queue to schedule tasks so that contradictions will be found quickly and so that the immediate con- sequences of assumptions will be considered before further assumptions are made. 3.9. Case 9—Single line of reasoning is too weak When we explain to someone how we solved a problem, we often invoke 20-20 hindsight and leave out the mistakes that we made along the way. Our explanation makes it appear that we followed a very direct and reasonable route from beginning to end. For developing intuitions about problem-solving behavior, this gives a misleading impression that problem-solving is the pursuit 2 A bibliography of recent papers on these ideas waspublished by Doyle and London [121. Basic algorithms for revisingbeliefs have been reported by Doyle [Ill and McAllester [261. EXPERT SYSTEMS TUTORIAL 163 of a ~ngle line of reasoning. Actually, there are important and somewhat subtle reasons for being able to use multiple lines of reasoning in problem solving and several of the systems described above gain power from this ability. These systems use multiple lines of reasoning for two major purposes as explained below: (1) To broaden the coverage of an incomplete search, or (2) to combine the strengths of separate models. The HEARSAY-Il system 115] provides the best example of the first purpose. (It is described in the next section.) In coping with the conflicting demands of searching a large space with limited computational resources, HEARSAY-Il performs a heuristic and incomplete search. In general, programs that have fallible evaluators can decrease the chances of discarding a good solution from weak evidence by carrying a limited number of solutions in parallel. A good example of combining the strengths of multiple models is given by the SYN program reported by Sussman, Steele, and de Kleer [10, 381. SYN is a program for circuit synthesis, that is, for determining values for components in electrical circuits. The EL program described in the previous section determined circuit parameters such as voltage given fully specified components in a circuit. SYN determines values for the components (e.g., the resistance of resistors) given the form of the circuit and some constraints on its behavior. SYN uses many of the propagation analysis ideas developed for El.. The interesting new organizational idea in SYN is the idea of slices or multiple views of a circuit. This corresponds to the idea of equivalent circuits in electrical engineering practice. A simple example of a slice is the idea that a voltage divider made from two resistors in series can be viewed as a single resistor; one slice of the circuit describes it as two resistors and another slice describes it as one. To analyze the voltage divider, SYN uses the second slice to compute the current through the divider. Then by reverting back to the first slice, SYN can compute the voltage at the midpoint. In general, the idea is to switch to equivalent representations of circuits to overcome blockages in the propagation of constraints. The power of slices is that they provide redundant paths for information to travel in propagation analysis. By exploiting electrically equivalent forms of circuits involving resistors, capacitors, and inductances, SYN is capable of analyzing rather complex circuits without extensive algebraic manipulation. The idea of slices is not limited to electrical circuits. For example, algebraic transformations of equations can be viewed as means for shifting perspectives. Sussman and Steele also give an example of understanding a mechanical watch by using structural and func- tional decompositions. In summary, slices are used to combine the strengths of different models. When they are combined with forward reasoning, they provide redundant paths for information to propagate. A problem-solver based on this idea must know how to create and combine multiple views. 164 M. STEFIK ET AL. 3.10. Case 10—Single source of knowledge is too weak An important adjunct to the use of multiple reasons in problem solving is the use of multiple sources of knowledge. In this section we will consider the HEARSAY-Il system, which coordinates diverse sources of knowledge using an opportunistic scheduler. HEARSAY-Il is a system for speech understanding reported by Erman et a]. [iSl. It recognizes spoken requests for information from a data base. Production of speech involves a series of transformations starting with the speaker’s intentions, through choice of semantic and syntactic structures, and ending with sound generation. To understand speech HEARSAY- !! must reverse this process. In HEARSAY-Il the knowledge for understanding speech is organized as a set of interacting modules (called knowledge sources or KSs) as shown by the arrows in Fig. 6. The KSs cooperate in searching a multi-level space of partial V — Verify .~-. FIG. 6. (Levels and knowledge sources in HERSAY-II). The knowledge sources areas follows: Semantics: generates interpretation for the information retrieval system. SEG: digitizes the signal, measures parameters, produces labeled segmentation. POM: creates syllable-class hypotheses from segments. MOW: creates word hypotheses from syllable classes. Word-Ct!: controls the number of hypotheses that MOW makes. Word-Seq: creates word-sequence hypotheses for potential phrases. Word-Seq-Ct!: controls the number of hypotheses that Word-Seq makes. Predict: predicts words that follow phrases. Verify: rates consistency between segment hypotheses and contiguous word-phrase pairs. Concat: creates a phrase hypothesis from a verified contiguous word-phrase pair. RPOL: rates the credibility of hypotheses. Levels Knowledge Sources Data Base Interface Semantics I Phrase IParse Concat~ Word Sequence IWor .d-Seq .4.— Word MO~ Word-Seq~.~j~ V RPOL Syllable POM .~i Segment SEG .4- — Parameter EXPERT SYSTEMS TUTORIAL 165 solutions. They extract acoustic parameters, classify acoustic segments into phonetic classes, recognize words, parse phrases, and generate and evaluate hypotheses about undetected words and syllables. The KSs communicate through a global data base called a blackboard with seven information levels as shown in the figure. These levels are HEARSAY-H’s heterogeneous abstraction spaces. The primary relationship between levels is compositional: word sequences are composed of words, words are composed of syllables, and so on. Entities on the blackboard are hypotheses. When KSs are activated, they create and modify hypotheses on the blackboard, record the evidential support between levels, and assign credibility ratings. For example, a sequence of acoustic segments can be evidence for identifying asyllable in a specific interval of the utterance; the identification of a word as an adjective can be evidence that the following word will be an adjective or noun. HEARSAY-II’S use of abstraction differs from the systems considered in cases 5 through 8. Those systems all use uniform abstraction spaces. The abstractions are uniform in that they use the same vocabulary as the final solutions and diper only in the amount of detail. For example in the planning systems, abstract plans have the same structure and vocabulary as final plans. In HEARSAY-Il, the diversity of knowledge needed to solve problems justifies the use of heterogeneous abstraction spaces. A computational system for understanding speech is caught between three conflicting requirements: a large space of possible messages to understand, variability in the signal, and the need to finish in a limited amount of time. The number of possible ideal messages is a function of the vocabulary, language constraints, and the semantics of the application. The number of actual messages that a system encounters is much larger than this because speech is affected by many sources of variability and noise. At the semantic level, errors correspond to peculiarities of conceptualization. At the syntactic level errors correspond to peculiarities of grammar. At the lexical and phonemic levels the variance is in word choice and articulation. Errors at the lower levels com- pound difficulties at the high levels. For example, the inability to distinguish between the four phrases till Bob rings, tell Bob rings, till Bob brings, tell Bob brings, may derive from ambiguities in the acoustic levels. HEARSAY-Il copes with this by getting the KSs at different levels to cooperate in the solution process. This has led to the following interesting architectural features: (I) HEARSAY-lI combines both top-down and bottom-up processing. (2) HEARSAY-LI reasons about resource allocation with a process called opportunistic scheduling. M.STEFIKETAL. An example of top-down processing is the reduction of a general sentential concept into alternate sentence forms, each sentence form into specific alter- native word sequences, specific words into alternative phonic sequences and so on until a best interpretation is identified. Bottom-up processing tries to synthesize interpretations from the data. For example, one might combine temporally adjacent word hypotheses into syntactic or conceptual units. In HEARSAY-Il, some KSs use top-down processing and other KSs use bottom-up processing. All KSs compete to he scheduled and HEARSAY-Il tries to choose the most promising KSs at any given moment using opportunistic scheduling. Oppor- tunistic scheduling combines the idea of least commitment with strategies for managing limited computational resources. The opportunistic scheduler adapts automatically to changing conditions of uncertainty by changing the breadth of search. The basic mechanism for this is the interaction between KS-assigned credibility ratings on hypotheses and scheduler-assigned priorities of pending KS activations. When hypotheses have been rated equally, KS activations for their extension are scheduled together. In this way, ambiguity between com- peting hypotheses causes HEARSAY-lI to search with more breadth, and to delay the choice among competing hypotheses until more information is brought to bear. HEARSAY-II’S approach to data interpretation differs from that of GAl discuss- ed in Section 3.4. Both programs contend with very large search spaces. Both programs need to have effective ways to rule out large classes of solutions. GAL does this with early pruning. In the absence of constraints, it would expand every solution in the space. HEARSAY-Il constructs a complete solution by extending and combining partial candidates. Because of its opportunistic sche- duler, it heuristically selects a limited number of partial candidates to pursue. To avoid missing solutions, HEARSAY-Il must not focus the search too narrowly on the most ‘promising’ subspaces. In summary, HEARSAY-Il provides an example of an architecture created to meet several conflicting requirements. Multiple levels provide the necessary abstractions for searching a large space. The levels are heterogeneous to match the diversity of the interpretation knowledge. Opportunistic scheduling com- bines the least-commitment idea with the ability to manage computational resources by varying the breadth of search and by combining top-down and bottom-up processing. 3.11. Case 11—General representation methods are too inefficient Research on expert systems has benefited from the simplicity of using uniform representation systems. However, as knowledge bases get larger, the efficiency penalty incurred by using declarative and uniform representations can become significant. Attention to these matters will become increasingly important in EXPERT SYSTEMS TUTORIAL 167 ambitiously conceived future expert systems with increasingly large knowledge bases. This section considers architectural approaches for tuning the performance of expert systems by making changes to the representation of knowledge. Three main ideas will be considered: (1) Use of specialized data structures; (2) Knowledge compilation; (3) Knowledge transformations for cognitive economy. The organization of data structures has consequences for the efficiency of retrieving information. The selection and creation of efficient data structures is a principal part of most computer science curriculums. Consequently, several of the programs discussed in the previous cases (e.g., DENDRAL, GAl, and HEARSAY- ii) use specialized data structures. In general, these data structures are design- ed so that facts that are used together are stored together and facts that are used frequently can be retrieved efficiently. Choice of data structure depends on assumptions about how the data will be used. A common assumption about special data structures is that they are complete with regard to specific relationships. From example, in GAl’S data structure for molecular hypotheses, molecular segments are connected if and only if they are linked in the data structure. This sort of assumption is commonplace in representations like maps, which are assumed to show all of the streets and street intersections. Representations whose structure in the medium is analogous to the structure of the world being modeled are some- times called analogical representations. A less understood issue is the use and selection of data structures in systems where many kinds of information are used. There is not much experience with systems that mix a variety of different representations. One step in this direction is to tag relations with information describing the chosen represen- tation so that specialized information can be accessed and manipulated using uniform mechanisms. Some ideas along this line appeared in Davis’ thesis [8], where schemata were used to describe some formatting and computational choices, but the work has not been extended to describe general dimensions of representation (see [4j) for use by a problem solver. A second important idea for knowledge bases is the idea of compiling knowledge. By compilation, we mean any process which transforms one re- presentation of knowledge into another representation which can be used more efficiently. This transformation process can include optimization as well as the tailoring of representations for particular instruction sets. Space does not permit a detailed discussion of techniques here, but some examples are listed below to suggest the breadth of the idea. Example I. Burton reported a System for taking ATN grammars and compiling them into executable code [51. The compiled grammars could be executed to 168 M. STEFIK ET AL. parse sentences much more rapidly than previous interpreter-based ap- proaches. Example 2. Production system languages have been studied and experimented with for several years (e.g., see [91). A basic difficulty with many production Systems is that large production system programs execute more slowly than small ones. The extra instructions do not need to execute to slow down the system; their mere presence interferes with the matching process that selects productions to run. An example of one such language is QPS2 reported by Forgy and McDermott [19]. Forgy conducted a study of ways to make such produc- tion Systems more efficient by compiling them into a network of ‘node programs’ [20]. The compiler exploits two properties of production systems: structural similarity and temporal redundancy. Structural similarity refers to the fact that many productions contain similar conditions; temporal redun- dancy refers to the fact that individual productions change only a few facts in the memory so that most of the data is unchanged from cycle to cycle. Forgy’s RETE matching process exploits this by looking only at the changes in the memory. Forgy’s analysis shows how several orders of magnitude of speed up can be achieved by compiling the productions and by making some simple changes to computing hardware. Example 3. Another system that compiles a knowledge base of production rules, EMYCIN, was reported by van Melle [39]. EMYCIN is not considered a ‘pure’ production system since it is not strictly data-driven; the order that the rules are tried in EMYCIN is controlled by the indexing of parameters. This means that EMYCIN’S interpreter does not repeatedly check elements in the working memory so that some of the optimizations used by Forgy would provide much less of an improvement for EMYCIN. EMYCIN’S rule compiler concentrates on eliminating redundancy in the testing of similar patterns in rules and compiles them into decision trees represented as LISP code. Example 4. The HARPY system for speech recognition reported by Lowerre [24] illustrates several issues about compilation. HARPY represents the knowledge for recognizing speech in a unified data structure (context-free production rules) which represents the set of all possible utterances in HARPY’S domain. This data structure represents essentially the same information that was used in HEARSAY- H except for the parameterization and segmentation information. HARPY’S knowledge compiler combines the syntax, lexical, and juncture knowledge into a single large transition network. First, it compiles the grammer into a word network, then it replaces each word with acopy of its pronunciation graph and inserts word-juncture rules at the word boundaries. In the final network, each path from a start node to an end node represents a sequence of segments for some sentence. With the knowledge in its compiled form, HARPY is capable of performing arapid search process that attempts to find the best match between EXPERT SYSTEMS TUTORIAL 169 the utterance and the set of interpretations. The major concerns about the extensibility of this idea are (1) that the highly stylized form of the input that HARPY can accept makes it difficult to add new knowledge and (2) the com- pilation is expensive for a large knowledge base (13 hours of PDP-l0 time for a thousand word, simple grammar). The promise of knowledge compilation ideas is to make it possible to use very general means for representing knowledge while an expert system is being built and debugged. Then a compiler can be applied to make the knowledge base efficient enough to compete with hand coding. Given a compiler, there is no need to sacrifice flexibility for efficiency: the knowledge base can be changed at any time and recompiled as needed. In addition, the compiler can be modified to re-represent the knowledge efficiently as hardware is changed, or as the trade offs of representation become better understood. The tech- niques for doing this are just beginning to be explored and will probably become increasingly important in the next few years. So far in our discussion of efficiency we have assumed that it is necessary for the designers of a knowledge base to anticipate how knowledge will be used and to arrange for it to be represented efficiently. Lenat, Hayes-Roth, and Klahr [23] coined the term cognitive economy to refer to systems which automatically improve their performance by changing representations, chang- ing access (e.g., caching), and compiling knowledge bases. Systems like this need to be able to predict how representations should be changed, perhaps by measurements on representative problems. The ideas of cognitive economy and knowledge compilation are more speculative than the other ideas we have considered and there are many theoretical and pragmatic issues to be resolved before they can be widely used. They are included here at the end in the hope that they will receive more attention in artificial intelligence research. 4. Summary Our pedagogical tour of cases began with the consideration of a very simple architecture. It required that problems in an application have a small search space and that data and knowledge be reliable and constant. In the second case we considered ways to cope with unreliable data or knowledge. Probabilistic, fuzzy, and exact methods were discussed. All of these methods are based on the idea of increasing reliability by combining evidence. Each method requires the use of meta-knowledge about how to combine evidence. The probabilistic (and pseudo-probabilistic) approaches use various a priori and conditional probability estimates; the fuzzy approaches use fuzzy set descriptions; the exact approaches use non-monotonic data correction rules. Errorful data and knowledge seem to be ubiquitous in real applications. In the HEARSAY-Il example of case 10, we saw how an opportunistic scheduler was used to cope with the conflicting requirements of errorful data, limited corn- 170 M. STEFIK ETAL. putational resources, and a large search space by varying the breadth of a heuristic search. In the third case we considered ways to work with time-varying data. We started with the situational calculus and then considered a program that used transition rules to trigger expectations in amonitoring task. More sophisticated ways to reason with time seem to require more research. The remaining cases dealt with ways to cope with large search spaces. We started with the hierarchical generate and test approach in the fourth case which requires a search space to be factorable in order to allow early pruning. This approach explores the space of solutions systematically and can be quite effective for returning all consistent solutions. We considered the issues of canonical forms and completeness in using this approach. Generate and test is a weak method, applicable only when early pruning is feasible. It requires the ability to generate candidates in away that large classes can be pruned from very sparse partial solutions. In many applications, solutions must be instantiated in substantial detail before they can be ruled out. Fortunately, it is not necessary in all applications to consider all possible solutions and to choose the best. The next few cases describe reasoning with abstractions to reduce the combinatorics without early pruning. The methods presented are usually applicable in satisficing tasks. In case five, we considered a form of reasoning based on fixed abstractions. This approach requires that all of the information needed for testing partial solutions be available (or be able to be generated) before a subproblem is generated. This requirement exposes the weakness of this approach: some problems cannot be solved from the available information without backtrack- ing. While the abstractions make the problem solver efficient, their use is too rigid for some applications. The next approach to abstraction is more flexible. Abstractions are com- posed from a set of concepts in a hierarchical space. This method is called top-down refinement. The simplest version of top-down refinement (case six) uses a fixed criticality ordering of the concepts and a fixed partial order for solving subproblems. Top-down refinement does not allow for variability in the readiness to make decisions. In the seventh case, we introduced the least commitment principle which says that decisions should be postponed until there is enough in- formation. This approach tends to exploit the synergy of interactions between subproblems. It requires the ability to suspend activity in subproblems, move information between them, and then restart them as information becomes available. In this case, the problem-solving knowledge is much richer than in the previous methods. It was suggested that principles like least commitment should be incorporated as part of meta-level problem solving. An inherent difficulty with pure least commitment approaches is the phenomenon of a least commitment deadlock. When a problem solver runs out EXPERT SYSTEMS TUTORIAL 171 of decisions that it knows it can make correctly, it must guess to con- tinue. We suggested that the amount of guessing is a measure of its missing knowledge, hut that knowledge bases will always he incomplete. This theme was continued in case eight where dependency-directed backtracking was discussed as an architecture to support efficient retraction of beliefs in plausible reasoning. Cases nine and ten illustrated: (1) the use of multiple lines of reasoning to enhance the power of a problem solver, (2) the use of heterogeneous abstrac- tion models to capture the variety of knowledge in some applications, and (3) the use of an opportunistic scheduler to use knowledge sources as soon as they become applicable (either top-down or bottom-up) and to control the breadth of search. Finally, we considered some methods for speeding up processing and in- formation retrieval: specialized data structures and knowledge compilers. These techniques do not attack the basic combinatorics of search, but they do reduce the cycle time of problem solvers and will become increasingly im- portant in future expert systems with large knowledge bases. In articulating these ideas about expert systems, we were forced to decide what was essential and important about expert systems. In some cases, other choices of expert Systems could have been made to make the same points. Our view is that the recent critical work in expert systems has focused on the mechanisms of problem-solving, and research in this area has been the most fruitful when it has been directed towards substantial applications. The ap- plications and the research reported here are both empirical: knowledge bases are tried, tested, and revised; the architectural research follows a similar, but slower cycle. An expert system is always a product of its time. System builders operate against a background of competing ideas and controversies and must also confront the limitations of their resources. To summarize experiments and create a simplified theory, we must necessarily step outside of this rich historical context. Inescapably, we are bound to our current vantage point. As this paper is written, there is a ferment of activity in expert system research. The theory of building intelligent systems is far from complete and the ideas expressed here are by no means universally accepted in the Al community. To wait for the ideas to settle and survive the test of history would preclude creating a timely guide to current thinking. ACKNOWLEDGMENT In August 1980, the National Science Foundation and the Defense Advance Research Projects Agency co-sponsored a workshop on expert systems in San Diego. The conference organizers were Rick Hayes-Roth, Don Waterman, and Douglas Lenat. The purpose of this workshop was to bring together researchers from many different institutions to write a definitive book [211 on expert systems. The participants were organized in groups corresponding to book chapters. 172 M. STEFIK ET AL. The ideas in this paper are the product of the ‘architecture’ group, of which Mark Stefik was the appointed chairman at the conference. The other members of the group were the other co-authors of this paper. During theconference we struggled intensively for three days to organize these ideas. The task of writing this tutorial fell to the chairman and was carried out over the next several months. Credit, however, belongs to all of the members of the group for actively and generously pulling together to define the scope of this tutorial, sharpen distinctions, organize the ideas, and later critique versions of the text. Thanks also to Daniel Bobrow, John Seeley Brown, Johan de Kleer, Richard Fikes, Adele Goldberg, Richard Lyon, John McDermott, Allen Newell, and Chris Tong for reading early versions of this text and providing many helpful suggestions. Thanks to Lynn Conway for encouraging this work, and to the Xerox Corporation for providing the intellectual and computing environments in which it could be done. REFERENCES I. Barr, A. and Feigenbaum, E.A., The handbook of artificial intelligence, Computer Science Department, Stanford University, Stanford, CA, 1980. 2. Buchanan, B.G. and Feigenbaum, E.A., DENDRAI. and Meta-DENDRAL: Their applications dimension, Artificial Intelligence 11(1978) 5—24. 3. Bobrow, D.G. and Winograd, T., An overview of KRL, a knowledge representation language, Cognitive Sci. 1(l) (1977) 3—46. 4. Bobrow, D.G., Dimensions of representation, in: D.G. Bobrow and A. Collins, Eds., Representation and Understanding (Academic Press, New York, 1975). 5. Burton, R.R., Semantic grammar: an engineering technique for constructing natural language understanding systems, Bolt Beranek and Newman Rept. No. 3453 (December 1976). 6. Charniak, E., Riesbeck C.K., and McDermott, DV., Artificial Intelligence Programming (Eribaum, Hilisdale, NJ, 1980). 7. Davis, R., Buchanan, B.G. and Shortliffe, E., Production rules as a representation for a knowledge-based consultation program, Artificial Intelligence 8 (1977) 15—45. 8. Davis, R., Applications of meta-level knowledge to the construction, maintenance and use of large knowledge bases, Ph.D. Thesis, Stanford University, AT Lab Memo AIM-283 (July 1976). 9. Davis, R. and King, J., An overview of production systems, in: E.W. Elcock and D. Michie, Eds., Machine Intelligence (Wiley, New York, 1976) 300—332. 10. de Kleer, J. and Sussman, G.J., Propagation of constraints applied to circuit synthesis, Circuit Theory AppI. 8 (1980) 127—144. 11. Doyle, J., A truth maintenance system, Artificial Intelligence 12 (1979) 231—272. 12. Doyle J. and London P., A selected descriptor-indexed bibliography to the literature on belief revision, Sigart Newsletter 71 (1980) 7—23. 13. Doyle, J., A model for deliberation, action, and introspection, Al TR 581, Artificial In- telligence Laboratory, MIT, Cambridge, MA (May 1980). 14. Duda, R.O., Hart, P.E. and Nilsson, N.J., Subjective Bayesian methods for rule-based inference systems, SRI International, Artificial Intelligence Center Technical Note 124 (Janu- ary 1976). iS. Erman L.D., Hayes-Roth, F., Lesser, V.R. and Reddy, DR., The HEARSAY-Il speech- understanding system: integrating knowledge to resolve uncertainty, ACM Comput. Surveys 12(2) (1980) 213—253. 16. Fagan, L.M., Kunz, J.C., Feigenbaum, E.A. and Osborn, J.J., Representation of dynamic clinical knowledge: measurement interpretation in the intensive care unit, Proc. Sixth Internat. Joint Conf. Artificial Intelligence, August 1979. 17. Fagan, J.M., vM: Representing time-dependent relations in a medical setting, Doctoral Dissertation, Computer Science Department, Stanford University, Stanford, CA, June 1980. EXPERT SYSTEMS TUTORIAL 173 18. Feigenbaum, E.A., The art of artificial intelligence: I. Themes and case studies of knowledge engineering, Proc. Fifth Internat. Joint Conf. Artificial Intelligence (1977) 1014—1029. 19. Forgy, C. and McDermott, J., OPS: A domain-independent production system, Proc. Fifth Internat. Joint C’onf. Artificial Intelligence (1977) 933—939. 20. Forgy, CA., On the efficient implementation of production systems, Doctoral Dissertation, Department of Computer Science, Carnegie-Mellon University, Pittsburgh, PA, February 1979. 21. Hayes-Roth, F., Waterman, D. and Lenat, D., Eds., Building Expert Systems (in preparation). 22. Hayes-Roth, B. and Hayes-Roth, F., A cognitive model of planning, cognitive Sci. 3 (1979) 275—310. 23. Lenat, D.B., Hayes-Roth, F. and Klahr, P., Cognitive economy, Computer Science Depart- ment, Stanford University, Heuristic Programming Project Rept. HPP-79-IS (June 1979). 24, Lowerre, B.T., The HARPY speech recognition system, Ph.D. Thesis, Computer Science Department, Carnegie-Mellon University, Pittsburgh, PA, 1976. 25. McCarthy, J., and Hayes, P.J., Some philosophical problems from the standpoint of artificial intelligence, in: B. Meltzer and D. Michie, Eds., Machine Intelligence 9 (Wiley, New York, 1979). 26. McAllester, D.A., An outlook on truth maintenance, MIT Al Memo No. 551, April 1980. 27. McDermott, J., Rt: A rule-based configurer of computer systems, Department of Computer Science, Carnegie-Mellon University, Rept. CMU-CS-80-1 19 (April 1980). 28. Newell, A., Artificial intelligence and theconcept of mind, in: R.C. Schank and KM. Colby, Eds., Computer Models of Thought and Language (Freeman, San Francisco, 1973). 29. Newell, A., Some problems of basic organization in problem-solving programs, in: M.C. Yovits, G.T. Jacobi and GD. Goldstein, Eds., Proc. Second Conf. on Self-Organizing Systems (Spartan Books, Washington DC, 1962). 30. Nilsson, N.J., Principles of Artificial Intelligence (Tioga, Palo Alto, 1980). 31. Pednault, E.P.D., Zucker, SW., and Muresan, L.V., On the independence assumption underlying subjective Bayesian updating, Artificial Intelligence 16 (1981) 213—222. 32. Sacerdoti, ED., A Structurefor Plans and Behavior (Elsevier, NewYork, 1977). 33. Sacerdoti, E.D., Planning in ahierarchy of abstraction spaces, Artificial Intelligence 5(2) (1974) 115—135. 34. Shortliffe, E.H., Computer-Based Medical Consultations: MYCIN (Elsevier, NewYork, 1976). 35. Stallman, R.M, and Sussman, G.J., Forward reasoning and dependency-directed backtracking in a system for computer-aided circuit analysis, Artificial Intelligence 9 (1977) 135—196. 36. Stefik, M.J., Planning with constraints, Artificial Intelligence 16(2) (1981) 111—140. 37. Stefik, M., Inferring DNA structures from segmentation data, Artificial Intelligence 11 (1978) 85—114. 38. Sussman, G.J. and Steele, G.L., CONSTRAiNTS—A language for expressing almost-hierarchical descriptions, Artificial Intelligence 14 (1980) 1—39. 39. van Melle, W., A domain-independent system that aids in constructing knowledge-based consultation programs, Doctoral Dissertation, Computer Science Department, Stanford Uni- versity, Rept. No. STAN-CS-80-820 (June 1980). 40. Zadeh, L.A., A theory of approximate reasoning, in: J.E. Hayes, D. Michie andLI. Mikulich, Eds., Machine Intelligence 9 (Wiley, NewYork, 1979). 41. Zadeh, L.A., Possibility theory and soft data analysis, University of California at Berkeley, Electronics Research Laboratory, Memo No. UCB/ERL M79/66 (August 1979). Received August 1981 
work_4l5buqbcubck7bnoj55k4ib7xu	----	NASA Technical Reports Server (NTRS) 
work_4lo326pejbdlzbws57oadv4af4	----	Privacy-preserving collaborative recommendations based on random perturbations Skip to main content Home Research Outputs People Faculties, Schools & Groups Research Areas Accountancy Africa Ageing Agricultural Management Agriculture Agriculture & Food Animal Care Any Arts Arts & Humanities Biomedical Sciences Biomedical/Medical Sciences/Health Biotechnology & Informatics Business Administration Cancer & Neurodegeneration Cell Signalling Chemical Engineering Chemistry Civil Law Computer Science Computing Social Sciences Criminal Law Education Energy Engineering Fisheries Food Food Analysis Food Management Fossil Fuels Gene Expression Health Higher Education Housing Humanities Infectious Diseases Journalism Education Land Management Law Management & Commerce Mathematics Mechanical Engineering Medical Science Other Miscellaneous Categories Peace Studies Physics Renewable Energy Sources Science & Technology Seafood Vegetables Research Centres/Groups Air Quality Management Resource Centre Applied Marketing Research Group Applied Statistics Group Bat Conservation Research Lab Big Data Enterprise and Artificial Intelligence Laboratory Bristol Bio-Energy Centre Bristol Centre for Economics and Finance Bristol Centre for Linguistics Bristol Economic Analysis Bristol Group for Water Research Bristol Inter-disciplinary Group for Education Research Bristol Leadership and Change Centre Bristol Robotics Laboratory Centre for Appearance Research Centre for Applied Legal Research Centre for Architecture and Built Environment Research Centre for Fine Print Research Centre for Health and Clinical Research Centre for Machine Vision Centre for Moving Image Research Centre for Public Health and Wellbeing Centre for Research in Biosciences Centre for Sustainable Planning and Environments Centre for Transport and Society Centre for Water, Communities and Resilience Collaborative Entrepreneurship Research Group Commercial Law Research Unit Computer Science Research Centre Creative Technologies Laboratory Data Research Access and Governance Network (DRAGoN) Digital Cultures Research Centre Document and Location Research Group Education Innovation Centre Engineering Modelling and Simulation Research Group Environmental Law and Sustainability Research Group Global Crime, Justice and Security Research Group Human Resources, Work and Employment Innovation, Operations Management and Supply Institute for Sustainability, Health and Environment Institute of Bio-Sensing Technology Mathematics and Statistics Research Group Moving Image Research Group Psychological Sciences Research Group Regional History Centre Research Group in Mathematics and its Applications Robotic Engineering and Computing for Healthcare - FET Science Communication Unit Social Justice Research Group Social Science Research Group Software Engineering Research Group Sustainable Economies Research Group (SERG) The WHO Collaborating Centre for Healthy Urban Environment Unconventional Computing Group Visual Culture Research Group Browse By Year By Author By Type About OAI Research Repository All Output Person Project Advanced Search Privacy-preserving collaborative recommendations based on random perturbations Polatidis, Nikolaos; Georgiadis, Christos K.; Pimenidis, Elias; Mouratidis, Haralambos Home Outputs Authors Nikolaos Polatidis Christos K. Georgiadis Elias Pimenidis Elias.Pimenidis@uwe.ac.uk Senior Lecturer in Computer Science Haralambos Mouratidis Abstract © 2016 Elsevier Ltd Collaborative recommender systems offer a solution to the information overload problem found in online environments such as e-commerce. The use of collaborative filtering, the most widely used recommendation method, gives rise to potential privacy issues. In addition, the user ratings utilized in collaborative filtering systems to recommend products or services must be protected. The purpose of this research is to provide a solution to the privacy concerns of collaborative filtering users, while maintaining high accuracy of recommendations. This paper proposes a multi-level privacy-preserving method for collaborative filtering systems by perturbing each rating before it is submitted to the server. The perturbation method is based on multiple levels and different ranges of random values for each level. Before the submission of each rating, the privacy level and the perturbation range are selected randomly from a fixed range of privacy levels. The proposed privacy method has been experimentally evaluated with the results showing that with a small decrease of utility, user privacy can be protected, while the proposed approach offers practical and effective results. Citation Georgiadis, C. K., Polatidis, N., Georgiadis, C., Pimenidis, E., & Mouratidis, H. (2017). Privacy-preserving collaborative recommendations based on random perturbations. Expert Systems with Applications, 71, 18-25. https://doi.org/10.1016/j.eswa.2016.11.018 Journal Article Type Article Acceptance Date Nov 15, 2016 Online Publication Date Nov 18, 2016 Publication Date Apr 1, 2017 Journal Expert Systems with Applications Print ISSN 0957-4174 Publisher Elsevier Peer Reviewed Peer Reviewed Volume 71 Pages 18-25 DOI https://doi.org/10.1016/j.eswa.2016.11.018 Keywords collaborative filtering, random perturbations, multi-level privacy, recommender systems Public URL https://uwe-repository.worktribe.com/output/889764 Publisher URL http://dx.doi.org/10.1016/j.eswa.2016.11.018 Additional Information Additional Information : The final publication is available at http://dx.doi.org/10.1016/j.eswa.2016.11.018. Files polatidis et al ESWA 2016 accepted.pdf (529 Kb) PDF Download Preview Licence http://creativecommons.org/licenses/by-nc-nd/4.0/ Organisation(s) FET Dept of Computer Sci & Creative Tech Research Centres/Groups Software Engineering Research Group You might also like A switching multi-level method for the long tail recommendation problem (2019) Journal Article An explanation-based approach for experiment reproducibility in recommender systems (2019) Journal Article Secure Social Media Spaces for Communities of Vulnerable People (2019) Conference Proceeding Hybrid Data Set Optimization in Recommender Systems Using Fuzzy T-Norms (2019) Conference Proceeding An approach for Web Service selection and dynamic composition based on Linked open Data (2018) Book Chapter Downloadable Citations HTML BIB RTF UWE Bristol Research Repository Powered by Worktribe | Accessibility About UWE Bristol Research Repository Administrator e-mail: repository@uwe.ac.uk This application uses the following open-source libraries: SheetJS Community Edition Apache License Version 2.0 (http://www.apache.org/licenses/) PDF.js Apache License Version 2.0 (http://www.apache.org/licenses/) Font Awesome SIL OFL 1.1 (http://scripts.sil.org/OFL) MIT License (http://opensource.org/licenses/mit-license.html) CC BY 3.0 ( http://creativecommons.org/licenses/by/3.0/) Powered by Worktribe © 2020 Advanced Search Just leave the fields blank that you don't want to search Repository ID Title all of any of Name Year Keywords all of any of Research Centres/Groups ACE Central ACE Dept of Art & Design ACE Dept of Arts & Cultural Industries ACE Dept of Creative & Cultural Industries ACE Dept of Education and Childhood ACE Dept of Film & Journalism ACE Technical Resources APD Central Academic Practice Directorate DIR Central DIR Directorate Projects DIR Planning & BI Directorate FAC Accommodation Services FAC Business Support FAC Catering & Bar Services FAC Central FAC Centre for Sport FAC Cleaning Services FAC Conference/ECC FAC Estates Operations FAC FT Facilities Technology FAC Health & Safety Unit FAC Logistics FAC Printing & Stationery FAC Security FAC Space Management & Design FAC Sustainability FAC Travel & Access FBL Bristol Business Engagement Centre FBL Central FBL Dept of Accounting Economics & Finance FBL Dept of Business & Management FBL Dept of Law FCM Central FCM Corporate Communications FCM Creative Strategy FCM Global Centre FCM International Recruitment & Admissions FCM Marketing FCM Student Journey Communications FCM TSU Temps FCM UK Recruitment & Admissions FET Central FET Dept of Architecture & Built Environ FET Dept of Computer Sci & Creative Tech FET Dept of Engineering Design & Mathematics FET Dept of Geography & Envrnmental Mgmt FET Technical Resources FIN Central FIN Commercial Services FIN Corporate Services FIN Faculty Finance FIN Financial Services FIN Procurement & Payments FIN Systems and Management Accounts FIN Treasury and Operations Facilities Faculty of Arts Creative Industries & Education Faculty of Business & Law Faculty of Environment & Technology Faculty of Health & Applied Sciences Finance Future Students Future Students, Comms and Marketing HAR Dept of Animal and Agriculture Science HAR Dept of Equine Science HAR Dept of Sport Science HAS Central HAS Dept of Allied Health Professions HAS Dept of Applied Sciences HAS Dept of Health & Social Sciences HAS Dept of Nursing & Midwifery HAS Technical Resources HRS Advice Hub HRS Business Partners HRS Central HRS Consultancy HRS Employee Relations & Reward HRS Equality & Diversity HRS HR Online Project HRS Learning & Development HRS Organisation & Learning Development HRS Payroll & Pensions HRS Resourcing HRS Systems & Information Hartpury (Associate Faculty) Human Resources IT Services ITS Applications Development & Testing ITS Central ITS Compliance & Security Team ITS Enterprise Architecture & Strategy ITS IT Operations ITS Strategic Business Engagement LCI Careers and Enterprise LCI Equality Diversity and Inclusivity LCI Library Services Library Careers and Inclusivity RBI Central RBI Research & Business Enterprise Service RBI Research & Development Research Business & Innovation SAS Administration & Advice SAS Casuals SAS Central SAS Policy Development & Student Experience SAS Student Data & Systems SAS Student Journey Programme SAS Student Support & Wellbeing SCM Central SCM Corporate Communications SCM Creative Strategy SCM Strategic Marketing SFS Admissions SFS Central SFS Global Centre SFS International SFS Recruitment & Outreach SPO Central Strategic Communications and Marketing Strategic Programmes Office Student and Academic Services University of the West of England Type Book Book Chapter Conference Proceeding Dataset Digital Artefact Exhibition / Performance Journal Article Other Patent Physical Artefact Presentation / Conference Report Thesis Working Paper Publication Status Submitted Accepted In Press Published Unpublished Journal or Publication Title all of any of Order the results By last modified (most recent first) By last modified (oldest first) By year (most recent first) By year (oldest first) By title Search Cancel 
work_4lud4hjutngljgcvc6bm2ewc2e	----	NASA Technical Reports Server (NTRS) 
work_4nanuojsqzczdmswysyxmurysu	----	International Journal of Scientific Research in Computer Science, Engineering and Information Technology CSEIT206285 | Accepted : 08 April 2020 | Published : 15 April 2020 | March-April-2020 [ 6 (2) : 345-350 ] International Journal of Scientific Research in Computer Science, Engineering and Information Technology © 2020 IJSRCSEIT | Volume 6 | Issue 2 | ISSN : 2456-3307 DOI : https://doi.org/10.32628/CSEIT206285 213 Designing an Expert System for Online Grocery Shopping Store Sayli Patil, Anisha Soni, Shivani Pawar, Prof. Sonal Chaudhari Department of Computer Engineering, Datta Meghe College of Engineering, Navi Mumbai, Maharashtra, India ABSTRACT In this digital era consumers prefer online shopping to save time. But in their busy schedules it can be time- consuming to choose items as per their requirements. The area of our project is to create a user friendly and easily available application for purchasing mechanism. This application will be easily available on the internet for users to download it and then install it. To start, offer only the most popular grocery items, like milk and bread, and as the business expands, so can your product line. The operating format for the app would be very easy to establish. Customers would simply log on to the app, select the items they wish to purchase, enter payment and shipping information and wait for their groceries. We propose the solution using an application designed on Android Operating System platform. We choose this platform base since Google’s Android is one of the foremost reasons of this mobile revolution. Keywords : Agrofresh, Online Grocery Store I. INTRODUCTION The term "Electronic commerce" (or e-Commerce) refers to the use of an electronic medium to carry out commercial transactions. Most of the time, it refers to the sale of products via Internet, but the term eCommerce also covers purchasing mechanisms via Internet. A client who purchases on the Internet is called a cyber consumer. E-Commerce is not only limited to online sales, but also covers: 1. Preparation of estimates online 2. Consulting of users 3.Access plan to point of sales 4.Quality assurance 5.Online payment MOTIVATION “If shopper continue to value innovation that offer convenience and time-saving solutions they will have less and less direct interaction with the retailers and brands. Gathering more customer data and using it to achieve long-term shopper loyalty will be critical for retailer and brands going forward. This will likely take the form of artificial intelligence using customer information to predict and fulfil shopper demand via automated delivery.” II. EXISTING SYSTEM The existing system is an automated system. But it was found to be inefficient in meeting the growing demands of population. • It is limited to a single system. • It is less user-friendly. http://ijsrcseit.com/ http://ijsrcseit.com/ http://ijsrcseit.com/ https://doi.org/10.32628/CSEIT206285 Volume 6, Issue 2, March-April-2020 | http://ijsrcseit.com Sayli Patil, et al Int J Sci Res CSE & IT, March-April-2020; 6 (2) : 345-350 346 • It is having lots of manual work (Manual system does not mean that you are working with pen and paper, it also include working on spread sheets and other simple software's). • The present system is very less secure. • It is unable to generate different kinds of report. • User must go to shop and order products. • It is difficult to identify the required product. • Description of the product obtained only on manually. • Accuracy not guaranteed. • Not in reach of distant users. III. PROPOSED SYSTEM In proposed system the customer need not go to the shop for buying the products. Customer can order the product he wishes to buy through the application in his Smartphone. As per the requirement provided by the user, immediate service is provided because the owner’s shop is nearby user’s residence. This system is purely based on client usage. The development of the new system contains the following activities, which try to automate the entire process keeping in view of the database integration approach. To debug the existing system, remove procedures those cause data redundancy, make navigational sequence proper. To provide information about audits on different level and also to reflect the current work status depending on organization/auditor or date. Required to build strong password mechanism. • User friendliness is provided in the application with various controls. • The system makes the overall project management much easier and flexible. • It can be accessed over the Internet. • Various classes have been used to provide file upload and mail features. • There is no risk of data mismanagement at any level while the project development is under process. • Report generation feature is provided using • Crystal Reports to generate different kinds of reports like bar graphs, pie charts and table type charts etc. • It provides high level of security using different protocols like https etc. IV. OBJECTIVE AND METHODOLOGY The scope of objective is very vast as it targets large no of people residing over the local area. The main objective for establishing an online presence are – 1. Promoting a service or product online. 2. Providing product support or customer service. A. GENERAL CONCEPT OF THE SYSTEM For grocery shopping, consumer preferences could be formed by combining three aspects, i.e., individual interest, product replenishment, and product promotion. Once the preferences can be precisely estimated, a corresponding recommendation technique is of high potential to be utilized in practical applications. http://www.ijsrcseit.com/ Volume 6, Issue 2, March-April-2020 | http://ijsrcseit.com Sayli Patil, et al Int J Sci Res CSE & IT, March-April-2020; 6 (2) : 345-350 347 Firstly, a bipartite network is constructed based on purchasing histories of all users to estimate the individual interest. Note that consumers are likely to accept product recommendations that are similar to what they have bought before. Also, it is observed that one is willing to accept recommendations from consumers of similar tastes. Thus, by introducing the random walk approach, how much a specific consumer likes a product can be generally estimated. Secondly, most daily necessities are consumable and are targets of grocery shopping. Therefore, consumers may buy the same product repeatedly. This is regarded as “product replenishment.” Note that consumable products normally are with a constant consumption rate. People then have to purchase them periodically. When something is going to be exhausted, the purchasing intent of a consumer becomes firm. In this work, we thus propose a statistical model to estimate this factor of product replenishment. Finally, a most common strategy to increase the sales of a product is promotion. People always like to do their purchases at a reduced price. Therefore, to consider the effect of product promotion in the recommendation system is necessary. Specifically, we model the degree of product promotion as well as the customer sensitivity of money saving to estimate the willingness for a consumer to buy a specific product. B. DATABASE MAPPING In simple words, data mapping is the process of mapping data fields from a source file to their related target fields. For example, in this the user will receive the otp then he will automatically generated to next page for their personal data filling and their they will be asked to enter their 'full name', 'address', 'landmark' are mapped to the relevant fields in a Delimited file, which is our database destination. Whereas depending on the number, schema, and primary keys and foreign keys of the relational databases data sources, database mappings can have a varying degree of complexity. Data Mapping Techniques: Although an essential step in any data management process, data mapping can be complex and time- consuming. The process of connecting data sources, building mappings for data transformation and integration, and validating the transformed data can take up significant developer resources, particularly when the entire process is done manually. Semi-Automated Data Mapping--in our application we have used Semi-Automated Data Mapping technique because the process involves identifying two data objects that are semantically related and then building mappings between them. Once schema mapping has been done, Java code is generated to achieve the required data conversion tasks. DATA MINING TOOL USED Open-Source: In this admin opted an Open-source mapping tools that provide us a low-cost alternative to on-premise data mapping solutions. These tools work better for small businesses with lower data volumes and simpler use-cases. Database Indexing Indexing is a way to optimize the performance of a database by minimizing the number of disk accesses required when a query is processed. It is a data structure technique which is used to quickly locate and access the data in a database.Indexes are created using a few database columns. The first column is the Search key that contains a copy of the primary key or candidate key of the table. These http://www.ijsrcseit.com/ Volume 6, Issue 2, March-April-2020 | http://ijsrcseit.com Sayli Patil, et al Int J Sci Res CSE & IT, March-April-2020; 6 (2) : 345-350 348 values are stored in sorted order so that the corresponding data can be accessed quickly. The data may or may not be stored in sorted order. The second column is the Data Reference or Pointer which contains a set of pointers holding the address of the disk block where that particular key value can be found. C. Recommendation System The proposed scheme can be divided into four parts, including information manager, database, analyzer, and recommender. The functionalities of each part is illustrated as follows. 1. Information manager: This part is to build consumer profiles such as name,email id, address, etc. On the other hand, when the consumer logs in to purchase products, the information manager begins to record the consumer behavior. 2. Database: This part is to store all on-line purchasing behavior and purchase histories of a consumer. Also, offers of all the products are stored. 3. Analyzer: This part is to analyze database contents for obtaining product similarities, individual interests, replenishment time intervals, promotion degree so as to estimate consumer preferences. Admin uses these consumer preferences to comprehensively evaluate the needs of a consumer. 4. Recommender: This part is to produce a recommendation list for a specific consumer according to the analysis result. The recommendation list shows the products which are most likely to be purchased. In general, this list includes not only replenishment goods but also products on promotion. V. PROTOTYPING OF PROPOSED RECOMMENDATION SCHEME In this work, admin develop a prototype of the proposed scheme to facilitate the empirical studies. When the consumer starts a shopping session, the shopping basket is empty. Consumer preferences have no value. Subsequently, the consumer put the milk into shopping basket. The proposed scheme computes preferences based on the current basket and the purchase histories. Thus, it can be seen that milk and toast are most frequently purchased together. The indexing has various attributes: Access Types: This refers to the type of access such as value-based search, range access, etc. Access Time: It refers to the time needed to find particular data element or set of elements. Insertion Time: It refers to the time taken to find the appropriate space and insert a new data. Deletion Time: Time taken to find an item and delete it as well as update the index structure. Space Overhead: It refers to the additional space required by the index. http://www.ijsrcseit.com/ Volume 6, Issue 2, March-April-2020 | http://ijsrcseit.com Sayli Patil, et al Int J Sci Res CSE & IT, March-April-2020; 6 (2) : 345-350 349 In this admin uses Non-clustered Indexing because a non-clustered index just tells the admin where the data lies, i.e. it gives admin a list of virtual pointers or references to the location where the data is actually stored. Data is not physically stored in the order of the index. Instead, data is present in leaf nodes. For eg. the contents page of a book. Each entry gives us the page number or location of the information stored. The actual data here(information on each page of the book) is not organized but admin has an ordered reference(contents page) to where the data points actually lie can have only dense ordering in the non- clustered index as sparse ordering is not possible because data is not physically organized accordingly. The contents page of the array. Each entry gives admin the location of the product stored . The actual data here is not organized but admin has an ordered reference(contents array)to where the data points actually lie. VI. CONCLUSION Due to fast moving lifestyle, online shopping has been growing drastically in India. With developed internet penetration, increasing adoption of devices like smartphones, tablets, and laptops, and access to the Internet and the shift in buying behavior among the consumers has contributed to the rapid growth of the online consumer base. The increase of online shopping has become a trendy way for consumers to shop over internet. Online grocery services meet a number of consumer needs including providing products for niche markets or helping the time starved consumer shop for the mundane weekly groceries. VII. REFERENCES [1]. He, Y. C., Wang, X. Z., He, Y. L., Zhao, S. L., & Li, W. B. (2016). Exact and approximate algorithms for discounted {0-1} knapsack problem. Information Sciences, 369, 634-647. [2]. Journal of Internet and e-Business Studies http://ibimapublishing.com/articles/JIEBS/2018/ 931248/ Vol. 2018 (2018), Article ID 931248, 13 pages DOI: 10.5171/2018. 931248 [3]. An Exploratory Study on Consumer Online Grocery shopping ttp://www.msruas.ac.in/pdf_files/Publications/ M CJournals/August2016/Paper2.pdfChandini A. V. and Nagendra Faculty of Management and Commerce, Ramaiah University of Applied Science,Banglore. [4]. Lv, J., Wang, X., Huang, M., Cheng, H., & Li, F. (2016). Solving 0-1 knapsack problem by greedy degree and expectation efficiency. Applied Soft Computing, 41, 94-103. [5]. Konig, D., Lohrey, M., & Zetzsche, G. (2016). Knapsack and subset sum problems in nilpotent, polycyclic, and co-contextfree groups. Algebra and Computer Science, 677, 138-153. [6]. Natarajan, T., Balasubramanian, S. A., & Kasilingam, D. L. (2017). Understanding the intention to use mobile shopping applications and its influence on price sensitivity. Journal of Retailing and Consumer Services, 37, 8-22. [7]. Shanbhag, S., Nair, S., Nai, N., & Shaikh, B. (2016). Intelligent Shopping Agent. [8]. Ingham, J., Cadieux, J., & Berrada, A. M. (2015). e-Shopping acceptance: A qualitative and meta- http://www.ijsrcseit.com/ Volume 6, Issue 2, March-April-2020 | http://ijsrcseit.com Sayli Patil, et al Int J Sci Res CSE & IT, March-April-2020; 6 (2) : 345-350 350 analytic review. Information & Management, 52(1), 44-60. Cite this article as : Sayli Patil, Anisha Soni, Shivani Pawar, Prof. Sonal Chaudhari, "Designing an Expert System for Online Grocery Shopping Store", International Journal of Scientific Research in Computer Science, Engineering and Information Technology (IJSRCSEIT), ISSN : 2456-3307, Volume 6 Issue 2, pp. 345-350, March- April 2020. Available at doi : https://doi.org/10.32628/CSEIT206285 Journal URL : http://ijsrcseit.com/CSEIT206285 http://www.ijsrcseit.com/ https://doi.org/10.32628/CSEIT206285 https://search.crossref.org/?q=10.32628/CSEIT206285 http://ijsrcseit.com/CSEIT206285 
work_4vomlmdu4rhxdouiiiukxac2fy	----	Multi-Robot Path Planning using Co-Evolutionary Genetic Programming Multi-Robot Path Planning using Co-Evolutionary Genetic Programming Rahul Kala School of Cybernetics, University of Reading, Reading, Berkshire, UK rkala001@gmail.com, Ph: +44 (0) 7466830600, http://rkala.99k.org/ Citation: R. Kala (2012) Multi-Robot Path Planning using Co-Evolutionary Genetic Programming, Expert Systems With Applications, 39(3): 3817-3831. Final Version Available At: http://www.sciencedirect.com/science/article/pii/S0957417411014138 Abstract: Motion planning for multiple mobile robots must ensure the optimality of the path of each and every robot, as well as overall path optimality, which requires cooperation amongst robots. The paper proposes a solution to the problem, considering different source and goal of each robot. Each robot uses a Grammar based Genetic Programming for figuring the optimal path in a maze-like map, while a master evolutionary algorithm caters to the needs of overall path optimality. Co-operation amongst the individual robots’ evolutionary algorithms ensures generation of overall optimal paths. The other feature of the algorithm includes local optimization using memory based lookup where optimal paths between various crosses in map are stored and regularly updated. Feature called wait for robot is used in place of conventionally used priority based techniques. Experiments are carried out with a number of maps, scenarios, and different robotic speeds. Experimental results confirm the usefulness of the algorithm in a variety of scenarios. Keywords: Path Planning, Motion Planning, Mobile Robotics, Genetic Programming, Grammatical Evolution, Co-operative Evolution, Multi-Robot Systems. 1. Introduction The problem of multi-robot motion planning deals with computation of paths of various robots such that each robot has an optimal or near optimal path, but the overall path of all the robots combined is optimal. This is a more complex task as compared to a single-robot motion planning, where the factor of coordination among the various robots is not applicable, and the single robot can use its own means to compute the path (Parker, Schneider, and Schultz, 2005). The problem of motion planning may be centralized or decentralized. In centralized planning all the robots are centrally planned by a planner, usually taking into account all the complex interactions that they may have. This results in the generation of a very complex configuration space, over which the search is to be performed. The decentralized planning, on the other hand, has an independent planner for every robot. Each robot is planned separately in its own configuration space, which makes the planning much simpler. Then efforts may be made to avoid the possibility of collision between the various robots. The centralized planning is more time consuming, but optimal as compared to decentralized planning (Arai and Ota, 1992; Sánchez-Ante and Latombe, 2002). In this paper we assume a simple problem modeling scenario. We have a maze like map of M x N grids. Each grid may be black or white, denoting the presence or absence of obstacle, which are all assumed to be stationary. The grid is however not the unit position for robots which can occupy partial grid positions on the map. The complete map consists of horizontal and vertical lanes, which cris-cross each other at grids known as crosses. The task is to move n robots R1, R2, … Rn. Each robot Ri has a start grid Si, which is its initial location; a goal grid Gi, which is the destination where it intends to arrive at the end of its journey; and a constant speed of motion Vi (0 < Vi ≤ 1) grids/unit time step. All robots are assumed to be of the same size of 1 x 1 grids. The planner needs to plan the path of all the robots Ri. Let the path of robot Ri be given by Pi(t), which denotes its position at time t. We know that Pi(0)=Si. Let the path generated be such that the robot Ri reaches the goal Gi at time Ti i.e. Pi(Ti)=Gi. We assume that after reaching the goal, the robot http://www.sciencedirect.com/science/article/pii/S0957417411014138 disappears from the goal, and does not obstruct other robots to pass by i.e. P i(Ti+1) = NIL. In this assumption we differ from the works over robot priorities by Oliver et al. (); Bennewitz, Burgard, and Thrun (2001, 2002); and Carpin and Pagello (2009). The planner needs to work such that the average time of travel for all the robots is as small as possible i.e. min(∑Ti)/n. At the same time the planner needs to ensure that no collision occurs. Hence Pi(t) ≠ Pj(t) ˅ 1 ≤ i, j ≤ n, i≠ j. Genetic Programming is extensively used evolutionary technique for problem solving. Here we represent the individual as a program and hence try to evolve it using the methodology of the evolutionary algorithms. The solution is obtained by executing the program so formed. Tree based representation of a program is a very common representation of individuals of genetic programming (J. R. Koza, 1992; Shukla, Tiwari, and Kala, 2010). The linear representation of the program of genetic programming gives rise to Grammatical Evolution. Here the individual is a sequence of integers. Every problem has its grammar, which represents the program syntax. The grammar is specified by Backus-Naur or BNF form. The generation of the phenotype solution from its genotype solution or sequence of integers takes place by selecting and firing rules specified in the grammar (O’Neil and Ryan, 2003, 2001). Our implementation of the Genetic Programming in this paper is similar to the general algorithm of grammatical evolution. The evolutionary algorithms are used for solving various kinds of problems. These algorithms however face problems when the dimensionality of the problem becomes very large. In such a case the optimization provided by these algorithms becomes very slow, with fairly large chance of the algorithm getting struck at some local optima (Kala, Shukla, and Tiwari, 2010a). Co-operative evolutionary algorithms or co- evolutionary algorithms break up the problem into a number of smaller problems that together solve the main problems. The smaller problems have smaller dimensionality, and are hence very easy to be optimized. The different sub-problems or sub-modules evolve in parallel, which ultimately results in generation of very good modules that unite with each other to solve the problem well. Cooperation is encouraged between modules, and a module is given high fitness not only because it results in generation of higher fitness solutions, but also because it enables other modules to achieve high fitness value (Potter and DeJong, 1994, 2000; García-Pedrajas, Hervás-Martínez, and Muñoz Pérez, 2003). In this paper we first present a co-evolutionary genetic programming based planning of multiple robots. Each of the individuals has its own optimization process that is based on the principles of Grammatical Evolution. These all try to generate and optimize the paths of the individual robots. However these need to consider the possibility of collision between the robots, for which the factor of co-operation comes into play and has been induced in the algorithm. A master genetic algorithm runs to select the best paths of the robots, out of all the paths available in the individual optimizations. The major problem with the assumed scenario is the variable speed. This puts forward a number of scenarios, where sub-optimality may be possible without careful planning. Consider a big corridor with two robots A and B marching towards it. Consider maximum velocity of A being higher i.e. V A >> VB. Consider that robot B is closer to the corridor entry, and would enter it before A. But if robot B enters the corridor first, robot A would have to follow it, and both robots would be able to walk with a maximum velocity of VB. It may be clearly seen that optimal strategy would be in case B waits for robot A to enter the corridor, and follows it inside the corridor. In such a case both robots would get a chance to travel by their maximum velocities. Hence we implement the concept of wait for robot, where a robot may be asked to wait so that some other robot may cross. Optimization by evolutionary algorithms may get very complex in problems having too many dimensions and multiple modalities. In such a case it is often useful to use a local search strategy based on predefined heuristics that can aid the evolutionary algorithms. This saves the algorithm from likely convergence into local optima. The local search algorithm carries forward the task to converge the algorithm into some local or global optima, while the evolutionary algorithm does the task of locating the best possible regions, where the global optima may lie. In such a case the combined efforts of heuristics and evolutionary algorithms results in better optimization (Kala,Shukla, Tiwari, 2009). The same problem of complex optimization is prevalent in the present algorithm. Here the evolutionary algorithm tries to optimize the individual robotic path. In such a case also we use a local search strategy which tries to modify the path to make it smaller in length. Longer paths are replaced by smaller paths. For carrying forward this replacement, an external memory is reserved that stores the smallest paths between all pairs of crosses. This table is regularly updated as more paths are found. The novelty of the paper is threefold. We first propose a co-evolutionary genetic programming model for solving the problem. This model uses principles of cooperative evolution to generate paths that are optimal in time, by mixing centralized as well as decentralized planning. Secondly, we propose the model and issues with multi-speed, map based (single-lane) multi-robot motion planning. This includes the use of the wait for robot feature incorporated into the evolutionary approach. Thirdly we propose use of external memory for local optimization of the paths. This memory is initialized, updated, and queried for reduction in paths of individual robots. The paper is organized as follows. In section 2 we present some of the related works into the field. In section 3 we present the co-evolutionary genetic programming model of the algorithm. Section 4 presents the concept of wait for robot. Section 5 presents the memory based local search used in the algorithm. The experimental results are given in section 6. Section 7 gives the conclusion remarks. 2. Related Works The entire paradigm of path planning of multi-robots involve a large set of issues that range from dynamic changes in the robotic plan, to validity of non-holonomic constants, and execution and memory time complexity. Oliver et al. (1999) give a general architecture of path planning in these scalable scenarios. They followed a decentralized approach of planning where every robot does its own independent path planning. The map is built centrally at start and all changes are continuously monitored and any needed correction is made. Possible collisions are resolved by the coordination module, that uses a priority based coordination with a higher priority robot allowed to make the collision prone move, while the other robot compromises or waits. Possible collisions are predicted be extrapolation of the motion of robots, and any possibility is marked by a check-point. One of the robots is allowed to move, while the path of other may be re-planned. Motion planning may be regarded as a part of the entire robotic architecture, for which again numerous models have been built. Asama, Matsumoto, and Ishida (1989) present a complete system for multi-robot planning, coordination, and control in their designed system ACTRESS (ACTor-based Robots and Equipments Synthetic System). In this system the authors demonstrate the mechanisms in which the different modules including sensors, processing, planning, manipulators, communication etc. come into play for intelligent and autonomous task accomplishment. The extension of their work can be found in (Asama et al. 1991) where the authors developed a multi-level path planning. The planning was different for static and dynamic cases. Special preventions were taken for deadlock avoidance that may take place many times in planning of multiple robots. The complete hierarchy of the system includes static path planning, use of local algorithm, use of mobile rules, task prioritization, deadlock avoidance by low-level deadlock solver, deadlock avoidance by high-level deadlock solver, and human operator expertise. Vendrell, Mellado, and Crespo (2001) presented a general framework for motion planning with multiple robots. They highlighted the need to plan, sense, and accordingly re-plan by every robot. The authors further present a hierarchical model of the various robot activities that range from high abstraction to low abstraction. The hierarchy includes mission, task, action, motions, trajectories, and robot order. A number of models have been made into the earlier works of the authors on single robots that revolve around evolutionary algorithms. In (Kala, Shukla and Tiwari, 2010b) a hierarchical implementation of a customized evolutionary algorithm was proposed. A number of evolutionary operators were defined to result in better convergence. The hierarchical nature enabled the robot to work in dynamic environment. The coarser evolutionary algorithm mainly looked at the static obstacles, and their optimality, the finer evolutionary algorithm tried to escape from all dynamic obstacles using a space-time graph map. In another work Kala, Shukla, and Tiwari (2010c) used evolutionary algorithm for generation of path in an incremental manner. Here the complexity of path or the number of turns that the path could have increased with time. The concept of momentum was floated which controlled the resolution with which the robot checked for the path feasibility and other metrics. The general approach used in our algorithm is of Evolutionary Algorithm, which is a member of the class of probabilistic algorithms, for which a bulk of research has been done. Kavaraki and Latombe (1997), Kavaraki et al. (2002) presented a method for multi-robot path planning called as the Probabilistic Road Map (PRM), which has since then been extensively modified and used for research. The algorithm consists of an offline phase and an online phase. In the offline phase the algorithm learns the map, and stores the summary about the paths between various points in a special data structure called as the road map. This step involves identification of a number of nodes that might represent difficult points in the configuration space, and are chosen by heuristics. The number of points is further increased by selection of points around these points. This selection may go on till required points are obtained, with the selection being weighted by a node’s difficulty or other selected heuristics. A local planning technique is used to swiftly compute the connectivity of these points to the other points. The collection of these vertices and edges makes the road map to use used for online planning. The online planner is query based which gets a query for a path and reacts to it by returning the shortest path based on the present map as well as the summarizations stored in the offline phase. This path may be further optimized by some local optimization technique, to increase the path optimality. Bohlin and Kavraki (2000) further extend the basic PRM to build a lazy PRM. The PRM does a large number of collision checks that usually results in a large wastage of time. In Lazy PRM the concept is to generate random nodes in the offline mode and try to connect each node to its neighbors for connectivity assessment. The feasibility of the paths is checked lazily in a coarser to finer manner. An initial lazy checking of the path feasibility reduces number of checks, at the same time giving an idea of the feasibility. If path is found to be feasible, it may be checked at a finer level as well. The in-feasible paths are enhanced by placement of the nodes. The PRM methods can be adapted to present a strong variance between central and de-central path planning with multiple robots. A comparison of centralized planning, decoupled planning with global coordination, and decoupled planning with pair wise coordination was done by Sachez-Ante and Latombe (2002). A large number of experiments of different number of robots with large degrees of freedom and scenarios were done. Experimental results confirmed that the requirement of robot dependency was a major factor behind the better performance of the coordination level. Clark (2005) used a probabilistic roadmap technique for solving the problem of motion planning of multiple robots. They used a single query based probabilistic roadmap planner. Here the author made a complete architecture comprising of software, communication, planning, and control where a robot could get the positional updates from all the other robots to get the complete roadmap and make decision. Here the author considered possibility of breaking of connectivity between two robots, resulting in generation of two unconnected sub-networks. The algorithm could track such changes and still make decisions of motion. Another probabilistic implementation can be seen in the work of Clark, Rock and Latombe (2003) who solved the problem of path planning for multiple-robot space system. Their planning algorithm generated random points or milestones and attempted to reach one milestone from the other, till the location was near to the goal. Heuristics were used for the selection of the milestone, whose neighborhood became the possible location of next milestone. Re-planning was an integral version of the algorithm, as the map could change in real time as per availability of information. Combined part results of a single decentralized planning system are used to serve as milestones of the centralized planner that checks for feasibility in combined motion. Svestka and Overmars (1998) present a probabilistic solution for solving the problem of multi-robot path planning. In this approach they define the paradigm of coordinated graphs, which allows only a collision free motion of the mobile robot. The robot is not allowed to collide with any of the moving robots. Their approach is partially central in nature, where the major task of planning lies with a central authority, on whose decisions the individual robots are made to move. The entire configuration space is made discrete with respect to the movement of the multiple robots. This the authors do by the concept of flat graphs, which takes into account the prospective motion of multiple robots. Observing the highly time consuming nature of the algorithm and the low scalability to the number of robots (large attributed to the central planning mechanism), the authors further extend their work by using multi-level graph clustering. Here the original graph is decomposed into a number of independent sub-graphs. This layered solution to the problem limits the number of nodes of the graph, which ultimately affects the time complexity of the algorithm in an exponential manner. The path hence generated by the algorithm is given some characteristics of non-central nature by introduction of mechanisms of smoothness of the path, for greater optimality. This is done by the application of a local planning algorithm for every robot in motion. Dongyong et al. (2006) used an evolutionary approach to solve the problem. Here the paths kept improving with time. The authors used a chaos based evolution, where the individuals were disturbed by some factors depending upon user set parameters. If the disturbed individual resulted in a better fitness, it was retained else the parent was retained. The crossover and mutation rates were adapted with time as per the fitness value. The authors optimized the navigation time, path length, and path smoothness. Kuffner and LaValle (2000) present an interesting application of another probabilistic algorithm called RRT. Here the RRT Trees start exploding both from the source as well as from the goal. The intersection of the two trees hence provides the path early. Another deviation from the conventional RRT Trees that the authors present is that the exploration of any node continues to take place in the same direction till the goal, an obstacle, or map wall is met. This results in heavy exploration at every step of the algorithm, and the general ideal goes a long way in enabling the RRT escape from a local minima. Wang and Wu (2005) make use of cooperative co-evolution in solving the problem of multi-robot path planning. This is primarily an evolutionary approach, where every robot has its own evolution. The evolution of all robots are shared, coordinated and communicated by a central mechanism. Every robot, in its evolution, enjoys a fitness that is a function of the length of path travelled and the constraints violated. The constraints refer to the individual robot constraints as well as the constraints to avoid collision with other robot. Slusny, Neruda and Vidnerova (2010) used evolutionary radial basis function networks and reinforcement learning for the task of localization and motion planning of a robot. The models were later analyzed by rule extraction technique, which converted the model to a rule based model for easy analysis and understanding. Kala, Shukla, and Tiwari (2010d) also used genetic algorithms for the optimization of a fuzzy inference system that was used for motion planning. This approach was performed at two levels. At the first level a probabilistic A* algorithm was used for making a rough sketch of the path, which was later on made detailed by the fuzzy based planner. In another work Kala, Shukla and Tiwari (2011) used evolutionary algorithms to make the rough sketch of the path, which was later traversed by the fuzzy planner. In this work again the concepts of incremental evolution were used. Another novel part of the algorithm was of variable speeds, which again has its inspiration from literature, with numerous related models. Kolushev and Bogdanov (1999) take into account the speeds of the moving robots while deciding their path in a graphical representation of the map or the configuration space. The robots may speed down or speed up below the maximum allowable velocity to avoid collisions. At any edge, the speed is also taken account, and correspondingly both the lowest path length plan and the shortest duration plan may be found out. A simple graph search algorithm may be used for planning. Rules are made to alter the velocity such that no collision occurs at the nodes as well as the edges of the graph. Priority based schemes have been extensively used for motion planning. Bennewitz, Burgard, and Thrun (2002) proposed a prioritized A* algorithm to coordinate the motion of multiple robots. Here the planning was done using a simple A* approach. Every robot had a priority, based on which the motion of the robots were carried out in case of possibilities of collision. The authors used a simple hill climbing approach to compute the best combination of priorities for the overall movement of robots. Priority is a major method for solving collisions in a multi-robot motion planning system. Bennewitz, Burgard, and Thrun (2001) consider the problem of finding the optimal priority scheme as the given problem. Their approach uses a combination of hill climbing and random search to set the correct priority scheme of the robots. Every priority scheme is judged by the optimality of paths generated in terms of length. At any step random exchange of priorities takes place. If these swaps result in better optimality, the change is retained, else rejected. The entire process is started multiple times from different configurations to escape from being trapped at local minima. We differ from these approaches by the fact that in our case the robots vanish after reaching their goals, and hence do not restrict other robot movements. The concept of priority here is implemented by the concept of wait for robot. 3. Cooperative Genetic Programming The basic working of the algorithm is a cooperative genetic programming planner. This planner operates in two levels: master and slave, similar to the architecture in (Moriarty, 1997; Moriarty and Miikkulainen, 1997; García-Pedrajas, Hervás-Martínez, and Muñoz Pérez, 2003). The slave genetic programming is basically a decentralized path planning for all the various robots in the system. Using this level the system tries to generate better and better paths for the individual motion of the various robots. The second level is the master level. Each robot in the slave has numerous genetic programming individual over which it tries to carry forward the optimization. The master is simply a genetic algorithm that tries to select the best combination of paths for the various robots, such that the overall path of all robots combined is optimal. In simple terms it may be regarded as the slaves optimize the individual robot paths, and the master optimizes the net work plan of all robots. However it needs to be noted that the slaves are conscious of cooperation amongst each other to generate collision free paths, and to help each other to optimize. Section 3.1 discusses the slave genetic programming. Section 3.2 presents the master genetic algorithm. The complete algorithm along with the interaction between the master and slave levels is done in section 3.3 3.1 Slave Genetic Programming The slave genetic programming operates at the level of individual robots for the generation of their paths, which when combined with the other paths of the other robots result in an overall optimal movement of the entire pool of robots. Each robot has its own instance of genetic programming which comprises of multiple individuals representing the potential path of movement of the robot from source and possibly finishing at the goal. Based on the principles of Grammatical Evolution (O’Neill and Ryan, 2001, 2003), we assume that the individual representation is in form of sequence of integers. In this manner we adopt a linear representation of the individual representing the path of the robot. Each integer denotes the movement of the robot. We know that once the robot is traveling on a straight path, it would continue to follow the same path, until it has got some point where it is possible to make a turn. The turns can only me made at some crossing, where multiple lanes meet each other. Whenever the robot meets a crossing, it needs to decide which way it has to follow. The number of ways possible may vary depending upon the type of crossing. The decision to which of the path is to be followed is decided by the individual, which uniquely determines the path of the robot. Before we discuss this conversion of the genotype or the sequence of integers to the phenotype or the robotic path, let us formally define some terms. The direction of motion of any robot Ri may be given by d={left, right, up, down} representing the four possible directions of motion. We assume the various paths to criss-cross each other only in multiples of right angles. A robot Ri traveling in direction d at a point (x,y) is said to be at a crossing if any path criss- crosses the path of traversal of the robot in direction d at point (x,y). It may be verified that the definition of cross does not depend upon direction d. Hence     otherwise intersects line horizontal and verticala i.e. crossing, a is y)(x, ),( false true yxCross (1) Each of the marked positions is a crossing as shown in figure 1. The robot is said to be struck at the location (x, y) while moving in direction d if it is not possible to move in the same direction. This may happen because of reaching of map limits, end of the road, or having some static obstacle ahead. Hence     otherwise y) (x,at whileddirection in movelonger nocan Robot ),,( false true dyxStruck (2) We are further interested in the number of possible moves (along with the actual moves) at every crossing, which would be used by robot for decision making. A robot would normally not like to traverse back into the same direction in which it was traveling since the traveled path would be wasteful. This however may be needed if the robot was struck. Let paths possible at any crossing (Cross(x, y)=true) at location (x,y) and direction d may hence be given by D(x,y,d), where D(x, y, d)  {left, right, up, down}. Let size(D) denote the total number of elements in the set or the total number of moves possible. Hence       truedyxstruckyxpt falsedyxstruckdyxpt dyxD ),,(),( ),,(~),( ),,( (3) Here pt(x,y) is a function that looks into the possibility of all the moves {left, right, up, down} and returns the set of possible moves. ~d is the direction of motion opposite to d. Figure 1: Various types of crosses. The figure depicts the numerous possible crosses that may be possible in a map where vertical and horizontal roads criss cross. The figure shows crosses with 4, 3, and 2 paths. Let the genetic individual for robot Ri (that is the i th instance of genetic programming) be given by I i <I i 1, I i 2, I i 3, … I i o>. We need to convert the genotype I i into the phenotype or the robotic path. Let ci be the pointer which points towards the integers in genotype and returns them whenever queried for decision making. We further define moveA(x, y, ci) as a function that returns the initial direction of motion of the robot while robot is at position (x, y) with pointer ci. Here the robot is allowed to take any possible way. Further let moveB(x, y, d, ci) be a function that returns the next direction of motion of the robot while at position (x, y) moving with direction d with pointer ci. Here the robot is not allowed to move back in direction it is coming from, unless it is struck. We need to first compute the initial direction of motion of the robot. We take a pointer ci initially set to the first integer (ci = I i 1). The robot is initially located at its source P(0)=Si. For the source we select the possible directions of movement of the robot. This is given by D=pt(x, y). Here (x, y) =Si. The number of possible moves are given by size(D). We select the k th move (d=Dk) in D where k is given by ci mod size(D) as the first move of the robot. Hence moveA(x, y, ci) = Dk, (4) D=pt(x, y), k = ci mod size(D) The pointer is updated to point to the next location (ci=I i 2). Now the robot is made to move in the same direction (d) till it reaches a crossing (Cross(x, y)=true) or it gets struck (Struck(x, y, d)=true). If either of these conditions is met, the robot has to make a decision regarding the next move. For this it first queries the total number of moves possible (D(x, y, d)) and then uses the integer pointed out by the pointer ci to Crosses decide which one of these moves is to be executed. Let the move be moveB(x, y, d, ci) which is given by k th move in D (d=Dk) where k is given by ci mod size(D). Hence moveB(x, y, ci) = Dk, (5) D=D(x, y, d), k = ci mod size(D) The pointer is further incremented to point to next integer. In this mechanism the algorithm keeps iterating with decision being made at every crossing and when the robot is struck. The motion of the robot may stop in either of two possible ways, either the robot reaches its goal, or the genotype is short of integers that is crossing after ci=I i n is to be processed. In case the stop is due to the termination of integer sequence, we reset the pointer (ci=I i 1) and allow the algorithm to proceed. This reset may however take place only once in the entire motion of the robot. In this manner the genotype may be finally converted into the phenotype. The complete algorithm is summarized in figure 2. Figure 2: Conversion of a genetic programming genotype to robotic path phenotype. The algorithm takes an individual which is a collection or integers. The robot is moved in its computed direction. Whenever any cross is encountered, the pointed integer is consulted to get the next direction of motion, and the pointer is updated. The algorithm terminates if goal is found or when the individual terminates twice. Individual Genotype Initialize the pointer ci and get corresponding initial direction of motion Initially place the robot at source While robot is not at goal Make robot move in next direction Robot at cross or struck? Update the pointer ci and get corresponding next direction of motion Pointer ci at last gene Reset ci Pointer ci never reset No Yes No Yes Return computed path End We study the conversion by an example. Let the map be as given in figure 3. The start position and goal of the robot is marked as shown in figure. Let the genotype be given by <2, 5, 3, 5, 2, 1, 3, 3, 3, 4, 5, 2, 2>. Initially the pointer is at 2. The possible moves are {down}. The algorithm selects the 2 mod1 i.e. the 0 th move which is down (Note that ordering starts from 0 and not 1). The robot marches down, until it reaches cross A. Here possible moves are {right, down}. The pointer is at 5. The robot selects 5 mod 2 i.e. 1 st move, which comes out to be down. The robot marches down till cross B. Here the possible move is {right}. Pointer is at 3. The move selected is 3 mod 1, i.e. right (0 th move). The robot comes to point C. Now possible moves are {right, down}. Pointer is at 2. The move selected is 2 mod 2, i.e. right (0 th move). Robot moves till cross D. Here possible moves are {up}, which is selected by pointer at 1.The robot travels up and reached cross E for which the possible move is only right. This is selected by pointer at 3. This makes the robot reach cross F with possible moves {right, down}. Pointer is at 2. The selected move is 3 mod 2 i.e. the first move or down. Robot moves in the same direction until it reaches cross G. The rest of the motion may be verified accordingly. At the end the robot will terminate its journey at goal, with some genes to spare. Figure 3: Sample map for conversion of genotype to phenotype. Figure is a sample map that shows the path used by a randomly generated robot for motion from source. The path traversed comes out to be Source → A → B→ C→ D→ E→ F→ G→ H→ I→ Goal The next important task to be carried out in the implementation of the slave genetic programming is the genetic operators. We use two genetic operators for the algorithm, crossover and mutation. Crossover is applied to the top elite individuals. The number of individuals that undergo crossover is given by cross x popSize, where cross is the crossover rate and popSize is the total number of individuals in the population pool. These individuals are paired in groups of 2 and crossover is applied to each paid. Scattered crossover technique is used for the generation of two children from every pair of parents. The generated children replace the two poorest fitness individuals in the population pool. The mutation operator is applied to rest of the individuals in the population pool. All these left individuals undergo change in their genes, where the change is given by a uniform mutation. Here the gene value is changed to a random value with a probability given by the mutation rate. The last part left in the implementation of the slave genetic programming is to have a fitness evaluation technique. We have already seen the mechanism of conversion of a genotype or an integer sequence into its phenotype or the robotic path. The fitness for any individual is measured by two factors denoting the individual and cooperative aspects of the overall path planning. The first factor for robot Ri (F i 1) measures the fitness of the individual path. Using this measure, an individual tries to generate shorter and shorter paths which may ultimately contribute towards the overall optimality of all the robots. This fitness is measured by the total length of the path. In case the robot did not reach the goal, a penalty is added proportional to the distance between the last point touched by the robot and the actual goal location. This may be given by A B C D E F Source G Goal H I F i 1= li + α || Pi - Gi || (6) Here li is the total length of path, Gi is the goal and Pi is the last point touched by the robot. α is the penalty constant canned as the penalty constant for non-reach ability of goal. By optimizing based on this objective it is highly likely that individual robots disregard the cooperation factor which is a major factor in co-evolution and enables different modules to develop characteristics to contribute well to the overall problem. This factor is served by the other factor of the fitness function (F i 2). In this factor we look into the possible collisions between the robot with other robot, or the factors by which this path results in sub-optimal paths of other robots. The path of this robot is a part of multiple paths of the other robots. These paths are generated by the master as we shall see later. We select the best e individuals of the master. Each of these individuals specifies the paths of each of the robots. We simply replace the path corresponding to robot Ri in all these e individuals by the current individual. We note the difference in the fitness value, which is a measure of F i 2. This may be hence given by      e j jIRj i XFitXFitF i i1 , 2 )()( (7) Here Xj is the j th best individual of the master. i i IRj X , denotes best j th individual of master with the path of robot Ri replaced by the current genotype I i . The net fitness may be given by F i =F i 1 + F i 2 (8) 3.2 Master Genetic Algorithm While the slave genetic programming does the task of generation of good paths for the individual robot, which have already taken the factor of cooperation into them, the task left is the evolution of the overall working strategy of the entire algorithm. This task is done by the master genetic algorithm, which decides how each and every robot is to move. The first task is the individual representation of the master genetic algorithm. The individual is a set of n pointers, each pointing to some individual of the slave genetic programming. Let these pointers or the genetic algorithm individual be given by <J1, J2, J3, … Jn>. Here any pointer Ji points to one of the individuals of the i th slave genetic programming (I i ) that represents the path of the i th robot. This may be easily understood from figure 4. The next task is the application of the genetic operators for the individual. Similar to the slave genetic programming, we use the genetic operators of crossover and mutation in this technique. The crossover mixes two parent individuals and generates two children. These two children replace the two weakest individuals of the genetic algorithm. A scattered crossover technique is followed in which every child randomly takes half of the pointers from first parent and the other half from the other parent. The mutation operator simply changes the genes of the individual. The gene is simply a pointer to some individual of the slave genetic programming. The mutation deletes an existing pointer and makes it point to some other individual in the genetic programming. Preference may be given to the newly created individuals of the genetic programming. The last task to be carried out is the fitness evaluation. The fitness evaluation of the master genetic algorithm simply returns the average traveling time of each of the robots. The traveling time of the robot is measured by the time that the robot takes to reach its goal. However the master genetic algorithm needs to ensure that all the paths are traversed without any kind of collision between the robots. For this a simultaneous simulation of all robots is done, knowing that the various robots have different speeds. It is further possible that some of the robots do not reach their goal. For both these aspects a penalty is added to the fitness function. The resultant fitness function of the master genetic algorithm hence becomes.   CGPTF n i iii .|||| 1    (9) Here Ti is the time the robot Ri reaches its goal Gi, Pi is the last point touched by the robot in its journey, α and β are penalty constants (β > α), C is the time first collision occurs (0 is no collision occurs). α is called as the penalty constant for non-reach ability of goal. β is called as the penalty constant for collision. Figure 4: Individual representation of master genetic algorithm. Every individual of the master genetic algorithm points to an individual of the slave genetic programming. Figure represents one individual of the master genetic algorithm The penalty constant for collision (β) is usually kept higher than the penalty constant for non-reach ability of goal (α). This is to encourage collision free movements. It would be better not to have collisions that not to have a robot reach its goal. The path may be feasible only when both these factors are zero, but characteristic positions of robots in the map may result in scenarios where it is not possible to have all the robots have a collision free journey to their goals. 3.3 Algorithm Outline We have already discussed the individual master and slave evolutionary algorithms. In this section we briefly give a combined picture of these two algorithms. The complete algorithm is presented in figure 5. It may be easily seen that the algorithm is iterative in nature, where both the evolutionary algorithms run in parallel. They hence share a common stopping criterion. A unit iteration of all the evolutionary algorithms is executed in parallel. This gives all of these a chance to evolve their population as per the changing scenarios and findings. It may be noted that all the various evolutionary algorithms do not run in isolation, but are rather dependent on each other. A boost in the fitness value of the master genetic algorithm by selection of some individual from some slave genetic programming may mean the slave genetic programming being re-directed to produce different kinds of individuals. Similarly if some new individual produced by one of the slave genetic programming results in improvement of the overall path being controlled by the master genetic programming, it may redirect the master to produce more individuals of the same kind. We have already discussed about the factor of cooperation between the various slave genetic algorithms. 4. Waiting for Robot The algorithm framework presented in section 3 is self-sufficient to carry planning of multiple robots, but it fails on a number of scenarios for which we introduce the concept of waiting for robot. Consider the case . Genetic Programming for Robot 1 Genetic Programming for Robot 2 Genetic Programming for Robot 3 Genetic Programming for Robot n Slave Genetic Programming Individuals Master Genetic Algorithm Individual where a slower robot is following a higher speed robot. This scenario was discussed earlier and was presented in figure 6 to ease the rest of the discussion. In this scenario the path of the two robots are shown. It would be better if the slower robot would have waited at point A for the faster robot to move. In such a case the slower robot would have followed the faster robot which naturally means a better moving strategy minimizing the collective motion time of all the robots. In reality with multiple criss-crosses and multiple robots there can be complex scenarios where one robot waiting for another robot may be fruitful. Figure 5: The basic evolutionary algorithm We informally state that wait may only be performed at the crossing (cross(x, y)=true); more precisely at a unit step from crossing, since we are to leave the crossing free for the moving robot. Waiting at any other location other than crossing may be equivalent to waiting at some crossing. Further the wait must always end with some robot crossing the point of crossing for which the robot is waiting. It is natural that there is no use waiting more than that, unless the robot is actually waiting for another robot after which it plans to move. This would be equivalent to the robot waiting for the second crossing robot with no necessary relation with the first crossing robot. It may be hence noted that the concept of waiting for robot at crossing is important and may further boost the performance of the model presented in section 3. We hence formally define the concept and modify the algorithm presented in section 3 as per its requirements. For all instances of genetic programming, initialize their individuals Initialize master genetic algorithm individuals While stopping criterion is not met For all instances of slave genetic programming Selection Crossover Mutation Selection in Genetic Algorithm Crossover in Genetic Algorithm Mutation in Genetic Algorithm Fitness Evaluation in Genetic Algorithm Fitness Evaluation End Return Best individual End We defined paths possible at any crossing (Cross(x, y)=true) at location (x,y) and direction d be given by D(x,y,d), where D(x, y, d)  {left, right, up, down}. Now the robot may be additionally asked to wait at some crossing and hence the option is added to D(x,y,d). The modified value is given by Dmod(x, y, d) = D(x, y, d) U {wait}. Here wait signifies that the robot is being asked to wait for a robot. size(Dmod) denotes the total number of elements in the set or the total number of moves possible. Consider the genotype pointer is pointing towards any location ci at any instance of time. Now the move made by the robot moveB(x, y, d, ci) is given by Figure 6: The problems with planning with multiple speeds. The figure shows two robots RA and RB moving with velocities VA and VB (VA >> VB). If RB enters the corridor first, RA will have to follow for the rest part of the journey with reduced speeds. Optimal path has RA entering before RB. moveB(x, y, ci) = Dk, (10) D=Dmod(x, y, d), k = ci mod size(Dmod) Similarly the first move of the robot is modified to work over Dmod in place of D. The restriction however here is that the first move cannot be a wait. This is because it is not necessary that the robot is initially located at a crossing, which is a requirement for the wait move. The first move moveA(x, y, cj) is hence given by moveA(x, y, ci) = Dk, (11) D=pt(x, y), k = ci mod size(Dmod) moveA(x, y, cj)={wait}  j<i. More precisely the robot has to wait just before the crossing and not on the crossing. In case it waits on the crossing, it would prohibit any robot to pass by including the robot for which it is waiting to pass by. Hence the decision towards whether the robot has to wait or not needs to be made before the robot enters into crossing. This may not be the immediately preceding step of crossing (as robots have fractional velocities) but the step after which the robot enters anywhere within the crossing region. We define the penultimate position of the robot (x,y) such that any step further from this in direction d would result in the robot entering into the crossing area. Let this be given by pen(x,y,d). Now as soon as any robot enters into pen(x, y, d), it queries the next crossing cross(x’, y’, d) and gets the move which it would be expected to make moveB(x’, y’, d). Here (x’, y’) is the location of the next crossing (in this query the pointer ci is not altered). In case this move comes to be wait (or moveB(x’, y’, d)={wait}), the robot does not move and waits at (x, y) and the pointer ci is incremented (or the wait move is already executed). Note that after the robot is allowed to move, it would reach the crossing and the next move would be made as per the new integer pointed by ci. In case the move does not come to be wait, the robot continues its motion. VB RA RB Goal Initial Locations VA Long Corridor A (x, y) (x’, y’) (x’’, y’’) Further when a robot is waiting, it must always wait for some robot that is guaranteed to cross from the crossing of wait. This ensures that we know exactly about when the robot may be allowed to resume its journey. Let (x, y) be the location of a robot Ri penultimate to crossing. Let (x’,y’) be the location of the crossing. Further let (x’’, y’’) be the position next to crossing (outside the crossing area) where Ri would be moving after wait is over. Since we know the complete genotype of motion of Ri, we can easily compute its entire path, excluding the waits which are dependent on the movement of the other robot and cannot be computed in isolation. We search for robots Rj (j ≠ i) which are presently moving and would be found at location (x’’, y’’) at some instance of time ahead but not after Ri reaches its goal if allowed to move without waits. Again it be noted that complete path of Rj is already known excluding the waits. If we get some Rj such that Rj is moving and Rj crosses (x’’, y’’) at some time ahead but not after Ri reaches its goal if allowed to move without waits, we state that the robot Ri will wait for Rj (Ri → Rj). The current state of Ri is labeled as waiting, and all it is not allowed to move further. Only after Rj crosses location (x’’, y’’), the robot Ri is allowed to move. In case no Rj (j ≠ i) satisfies the criterion, the wait of Ri is cancelled and its normal motion continues. It is important to put the check that Rj should be moving to escape from deadlock. If this condition is not checked it would be possible to have a situation Ri → Rj and Rj → Ri, in which case both Ri and Rj are not moving and waiting for each other. More generally it would be possible to have a scenario Ra → Rb → Rc, → Rd, … → Ra, which is a deadlock. In this manner we may be able to model multiple deadlock free scenarios with multiple robots waiting in any complex or simple manner. Hence by this introduction of wait in the problem modeling, we are able to generate more flexible movement strategy of the multiple robots. 5. Memory based Local Search The model of the algorithm discussed in section 4 may take a reasonably long time to carry out the task of optimization. It is a common technique to assist the evolutionary process by some heuristic means. In this algorithm we do this by storing a lookup table at a dedicated memory and use the same for optimization. The purpose of the lookup table is to store information about shortest paths between crosses. In case the path of any of the robots happens to be larger than the path stored in this lookup table, we may easily replace it by the smallest path already known. In such a case the robot would reach the goal in a much smaller path and if the modified path has no collisions, the overall optimality of the solution may be enhanced. The lookup table enables the robot to quickly attain the optimal path, while it may be reasonably away from the same. Another use of the lookup table is to enable cooperation amongst the robots. If one robot finds an optimal path between any pair of crosses, it must be able to share the same with the other robots, to enable them benefit from the finding. By making a lookup table if a robot finds the smallest path between any pair of crosses (as per the current exploration), it stores it into the lookup table which may be accessed by all the other robots if they need to go between the same pair of crosses in their journey. The lookup table or memory is supposed to store smallest paths between crosses. Suppose that there are v crosses (cross(x, y)=true)in the system. We select a total of η (η ≤ v) crosses randomly from the available v crosses that participate in the lookup table. η decides the size of the lookup table. The lookup table is a data structure that stores the smallest path between any of these crosses, where each cross is defined by the robot position (x, y) and direction d . Let the selected crosses be (V1, V2, …. Vη). The lookup table stores the minimal path length Val(Vi, Vj) between from cross Vi to cross Vj. It further stores the actual set of genotype integers that result in the generation of smallest path P(Vi, Vj) from cross Vi to cross Vj. One of the tasks associated is the initialization and update of the lookup table. Initially we set       ji ji ji VV VV VVVal 0 infinity ),( (12) P(Vi, Vj)=NIL (13) As we proceed with the algorithm, we get multiple paths on converting the genetic individual genotypes to corresponding phenotype paths. For the converted paths we check all pairs of crosses for possibility of shorter paths. Hence Val(Vi, Vj) = min{ length(P’Vi → Vj), Valold(Vi, Vj)}  Vi and Vj in generated path P, Vj comes after Vi (14) Here Valold(Vi, Vj) is the previous value and Val(Vi, Vj) is the updated value of the data structure. length(P’Vi → Vj) is the length of path between Vi and Vj in P’. Similarly the path needs to be modified. Hence we get       otherwisechange no ),()'(' ),( jioldVVVV ji VVValPlengthP VVP jiji (14) In this manner the lookup table is ready and keeps getting updated as newer paths are found. The last part left is to use the table as the local search of the individuals. For any path of the slave genetic programming we apply an additional operator reduce that attempts to shorten the path of the robot with a probability red. If the probability red is very low, there is almost no replacement and the factor of local search is almost lost. In case the factor is very high, it is likely that the individual slave genetic programming algorithms keep developing characteristics as per the characteristics in the lookup table, and the overall algorithm might get converged at some local minima. For any path in its genotype form which needs to be reduced, we randomly try to select pairs (Vi, Vj) in the path P’ of the robot, such that P’Vi → Vj > Val (Vi, Vj). For this path we replace the genotype integers from Vi to Vj by P(Vi, Vj). Suitable genes may be added at the two ends of the path. The previous path P’ (S → Vi-1 →Vi → Vj → Vj+1 → G) hence becomes (S → Vi-1 →P(Vi, Vj ) → Vj+1 → G) where S is the source and G is the last pointed touched by the robot. If no Vi, Vj satisfy the criterion, that is the path is optimized as per the lookup table, then no actions take place by this operator. The complete operation may be repeated over a number of times, as it is likely that the generated path may further be reduced by the lookup table. In this manner the application of this additional operator to the genetic programming individual may result in an enhanced optimization to the problem, considering the reasonably complex nature of the problem. This plays a major role in giving good results early. 6. Results The algorithm was tested on a simulation tool built in JAVA. The complete algorithm was coded in multiple modules which included the master genetic algorithm, slave genetic programming, and the operations over the individuals of these two evolutionary techniques. A simulation tool used the genotype of the evolutionary technique to simulate the complete system and generate the necessary phenotype paths. This was done separately for both the master and the slave evolutionary techniques. A JAVA Applet based GUI client was built that used multi-threading to show the optimal motion strategy of all the robots, as calculated by the algorithm. The maze-like map was given to the algorithm in a form of a BMP image. The paint utility was used for the generation of the map. Maze-like maps were used for all the testing experiments. The robots could only move in four directions. Hence in these maps it is not necessary that the optimal path lies close to the straight line joining the source and the goal, which would have been the case for robots which can move in multiple directions. If the motion is to be made from one corner of a square to its diagonally opposite corner, in these maps it makes no difference whether the journey is made close to the diagonal joining the corners, or using the vertical and horizontal edges. The distance of travel (which would come out to be equivalent to the Manhattan distance) would be the same. For the same reasons the map used for testing purposes had a number of turns such that the final optimal path between any source and goal is large and complex enough. The algorithm was first simulated for 2 robots. The intention was to check the path optimality of both these robots, as well as their cooperation to figure out a collision-free path. In order to ensure a complex overall path, the two robots were given diagonally opposite source and goals. Further to make collisions move likely, the source of the first robot was made the goal of the other, and vice versa. The first robot was initially located at (0, 2) and had to reach a goal (24, 23). The other robot was initially placed at (24, 23) and had to reach a goal of (0, 2). The speed of the two robots was 0.5 and 0.3. The simulation was carried out for a total of 20 iterations. The master genetic algorithm had a population size of 400. The mutation rate was fixed to 0.1. The top 40% individuals were used for the crossover operations. The rest individuals were generated by mutation operation. The genetic programming had a population size of 250 individuals. The mutation rate was 0.1. Here the top 20% individuals were used for crossover, the rest being generated by mutation. The individual had a size of 25 real numbers or genes. The penalty constant for not reaching goal was 100, and the penalty constant for collision was 1000. The algorithm, upon simulation generated the paths, which were displayed. The parameters and results are summarized in table 1. The motion of the robots along with time may be seen in video 1. The general motion of the robots is also given in figure 7 for ease of discussion. Table 1: Summary of situation and results for simulation with 2 robots S. No. Factor Robot 1 Robot 2 1. Source (0, 2) (24, 23) 2. Goal (24, 23) (0, 2) 3. Speed 0.5 0.3 4. Reached Goal Yes Yes 5. Collision No No 6. Time 142 383 7. Path Length 66 115 Figure 7(a): Path traced by robot 1 for simulation with 2 robots Figure 7(b): Path traced by robot 2 for simulation with 2 robots It can be easily seen that as per the set locations, the collisions between the robots was natural. In this case the robots made use of the wait feature and hence first robot waited for the second robot at point A (shown in figure 7). This enabled the two robots to move to their own goals. It may again be easily seen that wait can only be applied when the robots do not go in opposite directions. Suppose a robot is waiting at a cross, on the way where the other robot wants to go to. It is natural that there would be a collision, since the moving robot would collide with the waiting robot. Further we observe that the speeds of the robots play a big role in deciding the path. The paths for same set of parameters would turn out to be different if the Source Goal A Goal Source A velocities are changed. The velocities decide the possible points of collision, to which the colliding robots cooperate to find a collision free strategy. Table 2: Summary of situation and results for simulation with 3 robots S. No. Factor Robot 1 Robot 2 Robot 3 1. Source (0, 2) (24, 23) (23, 0) 2. Goal (24, 23) (0, 2) (1, 24) 3. Speed 0.4 0.7 0.8 4. Reached Goal Yes Yes Yes 5. Collision No No No 6. Time 166 167 71 7. Path Length 66 117 57 Figure 8(a): Path traced by robot 1 for simulation with 3 robots Figure 8(b): Path traced by robot 2 for simulation with 3 robots Figure 8(c): Path traced by robot 3 for simulation with 3 robots Source Goal Goal Source Goal Source The other execution was carried over the same map. To test the algorithm performance in higher number of robots, we added another robot to the simulation. So there were a total of 3 robots that tried to find their goal from the starting initial positions. The evolutionary parameters were kept same as the previous run. The source of the first robot was (0, 2), while its goal was (24, 23) and its speed was 0.4. The second robot had the source as (24, 23), goal as (0, 2), and speed 0.7. The third robot had source (23,0), goal (1, 24) and a speed 0.8. The simulation was carried out and the results were displayed using the GUI tool. The parameters and results are summarized in table 2. The motion of all the robots is given in video 2. The results are further drawn in summarized form in figure 8. Table 3: Summary of situation and results for simulation with 4 robots S. No. Factor Robot 1 Robot 2 Robot 3 Robot 4 1. Source (0, 2) (24, 21) (23, 0) (1, 24) 2. Goal (10, 24) (23, 0) (1, 24) (24, 21) 3. Speed 0.6 0.8 0.7 0.8 4. Reached Goal Yes Yes Yes Yes 5. Collision No No No No 6. Time 112 57 81 88 7. Path Length 67 46 57 71 Figure 9(a): Path traced by robot 1 for simulation with 4 robots Figure 9(b): Path traced by robot 2 for simulation with 4 robots Figure 9(c): Path traced by robot 3 for simulation with 4 robots Figure 9(d): Path traced by robot 4 for simulation with 4 robots Source Goal Goal Source Source Goal Goal Source Here as well we see that every robot tried to simultaneously cater to two needs, to find an optimal path, and to find a collision free path. It is evident that there are limited options or paths, and hence an optimal path may not be collision free and vice versa. An overall solution is hence difficult to obtain. It may further be seen that speed is a major factor. Consider the point A as shown in figure 9. If robot 1 traveling from top left corner was traveling faster, it would have resulted in some collision with robot 2 traveling from bottom right corner. This would have resulted in either of the robot waiting, or both of them force to recalculate their paths. We further test the scalability of the algorithm with more robots. The next simulation involved 4 robots. The initial locations of the four robots were (0, 2), (24, 21), (23, 0), and (1, 24). The goals were specified as (10, 24), (23, 0), (1, 24), and (24, 21). The speeds were set to be 0.6, 0.8, 0.7, and 0.8. The evolutionary parameters were kept same as the previous run. The result in this case is given in video 3, and the same is summarized in figure 9. The parameters and result statistics is given in table 3. It may be seen that though the modeling scenario was reasonably complex with multiple robots that were supposed to find their way out to the goal, the algorithm was able to generate these paths for all robots. All the robots could reach their goals in a reasonably optimal path. It may be possible for better paths to exist or the individual robots, but this may lead to collisions and hence needs to be avoided. The factor of optimization is the average running time of the robots, which is optimal in this case. It may be emphasized that every addition in robot means an immense increase in possibilities at the coordination level. Two robots may only need to ensure that they somehow do not collide, but multiple robots can interact in plentiful ways. We may avoid collision between two robots by some change in their paths, but the changed paths might result in more collisions with any of the other robots. Hence a very complex interaction amongst robots makes the problem difficult. The last experiment was done using 5 robots. The locations and the speeds of the robots were changed. The source of first robot was (2,0) and goal was (10, 24). The speed of this robot was 0.6. The second robot had a source of (24, 21) and goal of (23, 0). The speed of the robot was 0.8. The third robot was at (24, 3) and was supposed to move to (0, 19). The speed was kept as 0.7. The fourth robot had a source (0, 19) and goal (24, 21). The speed was 0.8. The fifth and the last robot was had a source (11, 0), goal (0, 14), and speed 0.2. The simulation was carried out and the final results were recorded. The movement of all the robots is shown in video 4. Figure 10 is the equivalent figure for the same simulation. The parameters are results are given in table 4. Table 4: Summary of situation and results for simulation with 5 robots S. No. Factor Robot 1 Robot 2 Robot 3 Robot 4 Robot 5 1. Source (2, 0) (24, 21) (24, 3) (0, 19) (11, 0) 2. Goal (10, 24) (23, 0) (0, 19) (24, 21) (0, 14) 3. Speed 0.6 0.8 0.7 0.8 0.2 4. Reached Goal Yes Yes Yes Yes Yes 5. Collision No No No No No 6. Time 115 69 81 83 149 7. Path Length 69 56 57 67 30 It may again be verified that despite heavy complexity, the algorithm could figure out a collision-free motion strategy for the motion of the robots. It needs to be again noted that motion of a robot by a path completely blocks the path. For the entire duration that the robot occupies that path, no robot would be able to move in any direction without collision. Hence the number of alternate paths plays a major role in the system. Too many robots with less number of alternative path might mean no path may be possible for some or the other robot that reaches the goal without collision. Either the robot may not reach goal at all, or it may reach the goal with collision. In some cases the robot might have to take too many unnecessary (and in some cases redundant) paths before moving on a path that takes it to goal. Hence there is a limitation of the map to the maximum number of robots that can move in. For the same reasons we do not experiment with larger number of robots. Hence we see that multiple robots may be optimally moved into the map. Figure 10(a): Path traced by robot 1 for simulation with 5 robots Figure 10(b): Path traced by robot 2 for simulation with 5 robots Figure 10(c): Path traced by robot 3 for simulation with 5 robots Figure 10(d): Path traced by robot 4 for simulation with 5 robots Figure 10(e): Path traced by robot 4 for simulation with 5 robots Source Goal Source Goal Source Goal Source Goal Source Goal The next scenario we generate is a very simple one. Here we make two robots move in opposite directions, so as to make them reach a common goal. In this scenario we aim to test the working of the wait feature of the algorithm. It is evident that it is not possible to move the robots in a manner that collision can be avoided without any robot waiting. The map has one crossing and one of the robots needs to wait so that the motion may terminate without collision. The simulation for this case is given in video 5 and figure 11. It may be seen that one of the robot waits for the other. This again proves the usefulness of the wait feature of the algorithm. Figure 11: Path traced by two robots using wait for robot feature Apart from the designed algorithm and the wait feature, the other module we designed was the local optimizations using a lookup table. We stated that this lookup table was important for a robot to use the findings of the previous generations, as well as the other robots. This enables swift motion towards the optima. We executed the same scenario with 4 robots without the lookup table. We observed that the results were highly sub-optimal. Many times robots were struck at some loops, where they returned back at a point they had visited earlier. The results with the execution of the program without lookup table are given in table 4. The paths of the various robots may be seen in figure 12. Video 6 shows the motion of the robots. This shows that the module of lookup table was effective and played a vital role in making the algorithm. We further study the execution time of the algorithm for the increase in number of robots. Based on our understanding of the algorithm it may be easily seen that the execution time should increase with increasing number of robots. This is because of the fact that for every increase in robot, a Genetic Programming instance is created and executed. Also the simulation plays a keen role in deciding the execution time. If the path to goal for all the robots is very simple, they would all attain their goals and the complete simulation would stop. On the other hand if there is a collision in the path of the robot, or it does not reach the goal till the end, there is a lot of computation that is performed. This is because for a very long time (until the genomic length is completely iterated) some or the other robot keeps wandering at the map. The increase in number of robots increases the possibility of collisions to a great extent and hence multiple paths need to be generated and tried. All this consumes a lot of time and execution time further increases. For the same reasons simulations having slower moving robots would take more computation time. We take the same map as discussed in the above discussions. We plot the execution time for the various numbers of robots. This is shown in figure 13. It can be easily seen that there is more than linear increase in computation time for increase in robots. This is because of the cooperation factor as discussed. We further extend the analysis to the stud y of the optimization of a single scenario. We study how the optimization time changes along with generations. Every generation has a different execution time. We may infer that the execution time depends upon the path of the robots. Every individual of both the genetic programming and genetic algorithm represents paths which need to be simulated multiple times. Longer paths would mean longer simulation time and hence the total time of execution of that generation would increase. As we move along with generations, paths start getting close to optimal solutions. The lengths of the paths reduce. More robots start to reach their goal, which means less wandering in the maps. Hence in general the execution time for higher generations is smaller than the execution time for the smaller generations. However the evolutionary operators may sometimes result in too fit or too unfit individuals. In case of the former, there is a sharp decrease in execution time, while the latter results in a sharp increase of the same. The sharp increase or decrease may be averaged out to some degree by the other individuals in the population pool. Figure 14 shows the graph for the execution time for various generations. The general trend has been plotted as a dashed trend line. The simulation of 4 robots was used. Figure 12(a): Path traced by robot 1 for simulation with 4 robots without memory based optimization Figure 12(b): Path traced by robot 2 for simulation with 4 robots without memory based optimization Figure 12(c): Path traced by robot 3 for simulation with 4 robots without memory based optimization Figure 12(d): Path traced by robot 4 for simulation with 4 robots without memory based optimization 7 Conclusions In this paper we attempted to solve the problem of multi-robot motion planning. The modeling scenario had a maze-like map where the different robots were initially located at distinct places and were given their own goals that they were supposed to reach. We further assumed that each robot moves with its own speed. The algorithm framework made use of co-evolutionary genetic programming. The task of planning was performed at two levels. At the first level a linear representation of Genetic Programming was used. The individual in this case consisted of instructions for movement whenever a cross is encountered. There was a Source Goal Goal Source Source Goal Goal Source different instance for every robot. The other level consisted of a genetic algorithm instance. This algorithm selected the individuals from the genetic programming and tried to generate a combination such that the overall path of all the robots combined is optimal. An individual of this level had pointers pointing out to genetic programming instances. The resultant algorithm could solve the proposed problem, but there were a number of scenarios that it could not solve. For this we included the feature of wait for robot, where a robot may be asked to wait before some crossing till a robot passed the same. In this manner robots were capable of escaping from reasonably complex scenarios in a cooperative manner so as to reach their own goals. Many collision prone scenarios could hence be controlled, making collision free planning possible. We further saw that this feature resulted in generation of optimal paths. This algorithm also solved the purpose, but the resulting problem became highly complex. This necessitated the need of some local search strategy, which we included into the algorithm in form of a lookup table. Here, as the algorithm proceeds, optimal paths between any two crossings are monitored and stored. The path of any individual may hence be modified, and it may be made to follow by the pre-computed optimal path, rather than the path as per its genes. This feature hence made it possible to generate optimal paths reasonably early in highly complex scenarios. Figure 13: Execution time v/s number of robots Figure 14: Execution time v/s generations The algorithm was developed as a JAVA software, which could read BMP images as maps. The various parameters of the evolutionary algorithm as well as the problem specifications could be defined as well. A number of executions of the algorithm were carried out with different number of robots. In all scenarios we saw that the algorithm could figure out a collision free path for all robots from the source to the goal. The paths seemed optimal in length and travel time. The different runs were done with different positions of the robots. This ensures that the algorithm generates optimal results to different modeling scenarios. We further did experimentation of the algorithm with the same scenarios without the memory based optimization. It was seen that the algorithm could not perform well. The resultant paths were not optimal. Further the effectiveness of the wait for robot feature was experimented by a simple scenario, and the resultant path had a robot to wait for the other robot. The experimental results prove that the algorithm could effectively solve the proposed problem as per the mentioned modeling scenarios in a mix of easy to complex scenarios. In all the cases presented, it was a common observation that the speed of the various robots played a big role in deciding their paths. The different speeds meant different points of likely collision, for which some alternative strategy had to be built. Further it was observed that the later generations of the algorithm were a lot quicker. This was due to the optimized nature of the individuals at those times. The increase in the number of robots implied an increase in the complexity which increased the execution time. The presented approach however has some limitations as well. The biggest limitation is the optimization time of the algorithm. This algorithm may hence not be usable for many real time scenarios. It may also be possible that the real world scenario may have dynamic maps, which the present approach does not consider. The other limitation of the algorithm is that the modeling scenario gets very complicated for the addition of large number of robots. In lust to have an optimal overall path with a central planning technique, we put a restriction on the maximum possible number of robots that may be computed in real time. In real world it may be possible to move any number of robots in some complex mechanism. This approach may not be able to evolve the same. The addition of some heuristics to the planning algorithm may help to further increase its scalability to such scenarios. These heuristics may further help to get the execution time of the algorithm down. All these problems may be worked over into the future. Further an exhaustive testing of the algorithm with different models, maps, and robot locations and velocities may be done. This would clearly speak out the positive and negative aspects of the algorithm. References T. Arai & J. Ota (1992) Motion Planning of multiple mobile robots, Proceedings of the 1992 IEEE/RSJ International Conference on Intelligent robots and Systems, pp 1761-1768. H. Asama, A. Matsumoto, and Y. Ishida (1989) Design of an Autonomous and Distributed Robot System: ACTRESS, IEEE/RSJ Intl. Workshop on Intelligent Robots and Systems, pp 283-290, Tsukuba, Japan H. Asama, K. Ozaki, H. Itakura, A. Matsumoto, Y. Ishida, & I. Endo (1991) Collision avoidance among multiple mobile robots based on rules and communication, IEEE/RSJ Intl. Workshop on Intelligent Robots and Systems, pp 1215-1220, Osaka, Japan M. Bennewitz, W. Burgard, & S. Thrun (2001) Optimizing schedules for prioritized path planning of multi- robot systems, Proceedings 2001 IEEE International Conference on Robotics and Automation, pp 271 – 276 M. Bennewitz, W. Burgard, & S. Thrun (2002) Finding and optimizing solvable priority schemes for decoupled path planning techniques for teams of mobile robots, Robotics and Autonomous Systems, 41: 89–99 R. Bohlin & L. E. Kavralki (2000) Path Planning using Lazy PRM, Proc. IEEE Intl. Conf. On Robotics and Automation, pp 521-528, San Francisco, CA S. Carpin & E. Pagello (2009) An experimental study of distributed robot coordination, Robotics and Autonomous Systems, 57: 129-133 C. M. Clark (2005) Probabilistic Road Map sampling strategies for multi-robot motion planning, Robotics and Autonomous Systems, 53: 244–264 C. M. Clark, S. M. Rock, J. C. Latombe (2003) Dynamic Networks for Motion Planning in Multi-Robot Space Systems, In: Intl. Symp. of Artificial Intelligence, Robotics and Automation in Space Y. Dongyong, C. Jinyin, N. Matsumoto, Y. Yamane (2006) Multi-robot Path Planning Based on Cooperative Co-evolution and Adaptive CGA, Proceedings of the IEEE/WIC/ACM International Conference on Intelligent Agent Technology, pp 547 – 550, Hong Kong N. García-Pedrajas, C. Hervás-Martínez, J. Muñoz Pérez (2003) COVNET: a cooperative coevolutionary model for evolving artificial neural networks, IEEE Transactions on Neural Networks 14 (3): 575– 596. R. Kala, A. Shukla, & R. Tiwari (2009) Comparative analysis of intelligent hybrid systems for detection of PIMA indian diabetes, Proceedings of the IEEE 2009 World Congress on Nature & Biologically Inspired Computing, pp 947 - 952, Coimbatote, India R. Kala, A. Shukla, & R. Tiwari (2010a) Clustering Based Hierarchical Genetic Algorithm for Complex Fitness Landscapes, International Journal of Intelligent Systems Technologies and Applications, 9(2): 185-205 R. Kala, A. Shukla, & R. Tiwari (2010b) Dynamic Environment Robot Path Planning using Hierarchical Evolutionary Algorithms, Cybernetics and Systems, 41(6): 435-454 R. Kala, A. Shukla, & R. Tiwari (2010c) Robotic Path Planning using Evolutionary Momentum based Exploration, Journal of Experimental and Theoretical Artificial Intelligence, Accepted In Press R. Kala, A. Shukla, & R. Tiwari (2010) Fusion of probabilistic A* algorithm and fuzzy inference system for robotic path planning, Artificial Intelligence Review, 33(4): 275-306 R. Kala, A. Shukla, & R. Tiwari (2011) Evolving Robotic Path with Genetically Optimized Fuzzy Planner. International Journal of Computational Vision and Robotics, Inderscience Publishers, Accepted In Press L. E. Kavraki, & J. C. Latombe (1997) Probabilistic roadmaps for robot path planning, In Practical Motion Planning in Robotics: Current Approaches and Future Directions, pp 33-53, John Wiley L. E. Kavraki, P. Svestka, J. C. Latombe, & M. E. Overmars (2002) Probabalistic roadmaps for path planning in high dimensional spaces, IEEE Trans. on Robotics and Automation, 12(4): 566 - 580 F. A. Kolushev & A. A. Bogdanov (2000) Multi-agent Optimal Path Planning for Mobile Robots in Environment with Obstacles, Proc. PSI'99, Lecture Notes in Computer Science 1755, pp. 503-510, Springer-Verlag Berlin, Heidelberg. J. R. Koza (1992) Genetic programming: on the programming of computers by means of natural selection. MIT Press, Cambridge, Mass. J. J. Kuffner & S. M. LaValle (2000) RRT-Connect: An Efficient Approach to Single-Query Path Planning, In Proc. 2000 IEEE Int’l Conf. on Robotics and Automation, Vol. 2, pp 995 – 1001, San Francisco, CA , USA D. E. Moriarty (1997) Symbiotic Evolution of Neural Networks in Sequential Decision Tasks. Ph.D. thesis, Department of Computer Science, University of Texas, Austin, Texas, 1997. D. E. Moriarty & R. Miikkulainen (1997) Forming neural networks through efficient and adaptive coevolution. Evolutionary Computation, 5(4):373–399. M. O’Neill, & C. Ryan (2001). Grammatical evolution. IEEE Transactions on Evolutionary Computation, 5(4): 349–358 M. O’Neill & C. Ryan (2003) Grammatical Evolution: Evolutionary Automatic Programming in a Arbitrary Language, volume 4 of Genetic programming. Kluwer Academic Publishers C. S. Oliver, M. Saptharishi, J. M. Dolan, A. Trebi-Ollennu, & P. K. Khosla (1999) Multi-Robot Path Planning by Predicting Structure in a Dynamic Environment, In Proceedings of the First IFAC Conference on Mechatronic Systems, Vol. II, pp 593–598 L. P. Parker, F. E. Schneider, & A. C. Schultz (2005) Multi-Robot Systems: From Swarms to Intelligent Automata Vol. 3, Springer-Verlag, New York. M. A. Potter, K. A. De Jong (2000) Cooperative coevolution: an architecture for evolving coadapted subcomponents, Evolutionary Computation 8 (1): 1–29. M. A. Potter, & K. A. DeJong (1994). A cooperative coevolutionary approach to function optimization. In Davidor, Y. and Schwefel, H.-P., editors, Proceedings of the Third Conference on Parallel Problem Solving from Nature, pp 249–257, Springer-Verlag, Berlin, Germany. M. A. Potter, K. A. De Jong (2000) Cooperative coevolution: an architecture for evolving coadapted subcomponents, Evolutionary Computation 8 (1): 1–29. G. Sánchez-Ante & J.C. Latombe (2002) Using a PRM Planner to Compare Centralized and Decoupled Planning for Multi-Robot Systems, Proc. IEEE Int. Conf. on Robotics and Autom., pp 2112 – 2119 A. Shukla, R. Tiwari, and R. Kala (2010) Towards Hybrid and Adaptive Computing, Springer-Verlag Berlin, Heidelbrg S. Slusny, R. Neruda, P. Vidnerová (2010) Comparison of behavior-based and planning techniques on the small robot maze exploration problem, Neural Networks, 23: 560-567 P. Svestka & M. H. Overmars (1998) Coordinated path planning for multiple robots, Robotics and Autonomous Systems, 23: 125-152 E. Vendrell, M. Mellado, and A. Crespo (2001) Robot planning and re-planning using decomposition, abstraction, deduction, and prediction, Engineering Applications of Artificial Intelligence, 14: 505– 518 M. Wang & T. Wu (2005) Cooperative co-evolution based distributed path planning of multiple mobile robots, Journal of Zhejiang University Science, 6A(7): 697-706 
work_4wbxt34ucbeb7ju7vmmvb5zvcy	----	doi:10.1016/j.eswa.2005.09.007 Ontology construction for information classification Sung-Shun Weng a,*, Hsine-Jen Tsai a , Shang-Chia Liu b , Cheng-Hsin Hsu a a Department of Information Management, Fu Jen Catholic University, 510 Chung-Cheng Road, Hsin-Chuang City, Taipei 242, Taiwan, ROC b Graduate Institute of Business Administration, Fu Jen Catholic University, 510 Chung-Cheng Road, Hsin-Chuang City, Taipei 242, Taiwan, ROC Abstract Following the advent of the Internet technology and the rapid growth of its applications, users have spent long periods of time browsing through the ocean of information found in the Internet. This time-consuming hunt, however, makes searching, retrieving, displaying, integrating and maintaining data such arduous tasks. One way to solve this problem is to study the concept behind the Semantic Web in accordance with the principles of ontology. Apart from facilitating the process of information search in the Semantic Web, ontology also provides a method that will enable computers to exchange, search and identify text information. But establishing the ontology necessitates a great deal of expert assistance; manually setting it up would entail a lot of time, not to mention that there are only a handful of experts available. For this reason, using automatic technology to construct the ontology is a subject worth pursuing. This research uses the theory of formal concept analysis to serve as the groundwork in assembling the different levels of ontological concepts in an automated fashion. An ontology diagram will be presented to show the correlation of concepts and their corresponding significance. Moreover, the experiments of this research select a collection of different concepts in an attempt to classify the relationships between documents and concepts. The objective is to develop an automated technology of ontology construction that will support the present information classification system, as well as to upgrade the ontological aspect of the Semantic Web. q 2005 Elsevier Ltd. All rights reserved. Keywords: Ontology; Semantic Web; Information classification; Formal concept analysis 1. Introduction Following the rapid widespread of computer networks, data handling had slowly grown dependent on computer networks and servers particularly in carrying out automated exchange tasks. In other words, data storage, data management, data transmission and even analysis rely on computer and network technology. At the same time, as a result of the vigorous developments and accessibility of the World Wide Web (WWW), a great quantity of information was suddenly made available to people. However, due to the enormousness of the data, users waste a lot of time browsing the Internet and searching for the information they need; it makes the tasks of searching, accessing, displaying, integrating and maintaining data more laborious. With the aim of solving this difficulty, Berners-Lee and Fischetti (1999) conceived the concept of the Semantic Web. Based on this concept, ontology and intelligent agents constitute the foundations of the Semantic Web. Ontology is made capable to ‘describe metadata’ in order to build one complete glossary that will clearly define the data 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.09.007 * Corresponding author. Tel.: C886 2 29052717; fax: C886 2 29052812. E-mail address: im1032@mails.fju.edu.tw (S.-S. Weng). found in the World Wide Web (WWW). The sharing and re- using of ontology enable users to communicate with computers by using accurate syntax and semantics. The powerful capabilities of the Semantic Web and its full development depend on the proper knowledge and handling of the agents. Aside from being able to fathom the user’s demands, intelligent agents must also be able to find similar or related resources from the Semantic Web as constructed by ontology, and furthermore, to build trust in the whole architecture of the Semantic Web. This can be achieved through the mutual collaboration of different agents to test, verify, and assure the reliability of information. Semantic Web is anything but a new kind of network; it is built within the existing network environment and provides a highly readable data without modifying or altering any of the contents. Simply put, a Semantic Web is an integration of numerous metadata. These data serve to describe documents, web pages and general concepts. It may be impossible for computers to understand the context of any document but with the help of the Semantic Web, exchanging, searching and recognizing the meaning of characters become possible. Therefore, the development and success of the Semantic Web will greatly depend on how fast and efficiently ontology is established. New websites may possibly consider creating its blueprint for ontology during its early developmental phase but Expert Systems with Applications 31 (2006) 1–12 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–122 because data or information can quickly become obsolete, the integration between the ontology and the website can be expected to gradually diminish. When this occurs, there is a need to build a whole new ontology. Furthermore, data found in current websites have completely changed or evolved other than those in the time when the Internet was just beginning. For this reason, the automatic construction of ontology as a stimulus to promote great advancement of the Semantic Web is a topic worthy of an in-depth analysis. These days, the study of ontology in the Semantic Web suggests the setting up of ontology that will represent Web data during the preliminary stages of constructing the website. The current information classification systems are likewise depen- dent on the classification framework defined by the experts. Since, these measures may encounter difficulties with the current information classification systems, this research aims to provide new theories that will help solve the above-mentioned problems. We believe that there are several difficulties in manually defining a classification framework. Overcoming these challenges would be most beneficial to the existing information classification systems and ontology in the Semantic Web. The following are brief explanations of these problems: 1.1. Lack of flexibility in the classification framework Whether it is the current information classification systems or ontology of the Semantic Web, modifying the existing classification framework is very difficult, if not impossible. Whether the website appears as a database or in a format of table of contents, it is imperative to adhere to the pre-defined classification framework in categorizing information. But with the rapid addition to or modifications in the data, it becomes clear that there is an apparent lack of flexibility in manually defining the classification framework. If we can base the classification framework on the data content and automatically generate the corresponding classification, then the tremendous inflow and/or modifications in the data will no longer be a problem. Although the current ontological construction technology can achieve a partially automated classification framework, still there pose several limitations. It is therefore, one of the serious aims of this research to make a significant breakthrough and achieve a fully automated classification framework. 1.2. Conceptual relationships and significance of resources The classification relationships manifested in the existing classification framework is generally absolute. Its incapability to show the corresponding significance of identical classifi- cation levels will result in the omission of vital information when searching for a general concept. Therefore, identifying the significance of various classifications and resources in a proper manner can increase the accuracy rate of the search function. On account of the above-mentioned grounds, this research intends to provide a automatic construction technology of ontology that will help solve the difficulties enumerated above. Solving these problems will not only upgrade the existing systems but will also support the construction of ontology in the Semantic Web. 2. Review of related literature Up until the present times, extensive and general construc- tion methods are not found within the domain of ontology learning. Although in other fields such as linguistics, information retrieval, machine learning, data mining and software engineering, etc. there are plenty of studies and related technologies that can apply the benefits of ontology learning. Maedche and Staab (2001) mentioned that ontology learning can be divided into four parts: extract, prune, refine, import or reuse. We will, for now, direct our attention to the extraction methods upon which we will base our research findings. There are four categories in the construction methods of ontology learning, these are: dictionary-based, text clustering, association rules, and knowledge base. These categories are further explained below: 2.1. Dictionary-based construction method Using the compilation concept of a traditional dictionary, the hierarchy of concepts is automatically formed. Traditional dictionaries present entries together with their synonyms, root words, etymology, etc. The definitions and relationships presented in the dictionary are used to determine the hierarchy relationships of concepts (Khan & Luo, 2002; Kietz, Maedche, & Volz, 2000; Tan, Han, & Elmasri, 2000). The dictionary-based construction method normally is the groundwork of other construction methods. The other three methods are somehow related to the dictionary-based construction method either in the preliminary construction phase or in the final pruning and verification stage. This is so because the dictionary-based method has its own limitations and will only be effective when paired with another kind of method. It is never used independently. The limitations are listed as follows: (1) The ontology formed using the dictionary-based method has a general description and is not at all domain specific. Only when it is combined with another method does it provide a more significant and valuable ontological framework. (2) The dictionary-based method is generally restricted to the volume size of the dictionary and can thus form domains having different scopes. Using this method alone will not only pose possible setbacks due to the quality of the dictionary, it will also prove incapable of adapting to the incessantly changing environment 2.2. Text-clustering-based construction method Using the text-clustering-based method to computerize the establishment of conceptual hierarchy is based on related terms S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–12 3 grouped together according to their synonyms. Every cluster is represented by a particular word or term that is believed to be more frequently used. Thus, repeating the exercise can derive the hierarchy of the terminologies. Right now, there are still several problems found in using this method which restrict its usability as explained below (Hotho, Maedche, & Staab, 2001): (1) Text clustering is generally regarded as an objective type of method that generates well-defined results considered optimum in certain aspects. However, this is the exact opposite of what goes on in reality. Different users have different requirements for clustering because they have varying opinions and perspectives in viewing a particular document. Thus, we ought to adopt individual standards so as to accommodate differing perspectives in accomplishing the task of text clustering. (2) Text clustering is normally required in high dimensional spaces to perform clustering computations since, every word or term is seen as an attribute of the entire document. However, experiments and mathematical analysis confirmed that clustering calculations in high dimensional spaces are inefficient because every data point possesses tendency of similar distances with other existing data points. (3) Text clustering by itself is ineffective unless it is combined with a specific domain. The common solution is to formulate clustering regulations which unfortunately produces another set of problems such that clustering regulations generate too many features. 2.3. Construction method based on association rules Using association rules to achieve an automatic construction of concept hierarchies is derived from the idea that association rules with stronger support, confidence and more extensive conceptual relationships can be placed on the upper level of ontology (Maedche & Staab, 2000). For example, we compute the support and confidence, respectively, on (region, accom- modation facilities) and (regions, hotels). If the result of support and confidence in (region, accommodation facilities) is clearly higher than (region, hotels), ‘accommodation facilities’ is placed higher than ‘hotels’. Using association rules as a basis for construction method still causes several restrictions which are further explained below (Maedche & Staab, 2000; Wei, Bressan, & Ooi, 2000): (1) Using association rules can generate combinations of different conceptual relationships, for example, the combination of (Person, Person, HIT) and (Person, Person, LOVE). However, if we treat them as two separate concepts, it will be difficult to arrive at the needed support. (2) When the document is composed of transactions required by the computation of association rules, different tactics of combination can result in different outcomes. For instance, there are 100 documents about ‘Hotel’ with contents giving detailed descriptions about room types and facilities. After natural language processing, it might result in 10,000 concepts. If a document becomes one trans- action, Hotel/Address and Room/Bed with support of 100% can be derived. However, the computation complex- ity can most likely be high. If every sentence in the document makes up one transaction, then the vital relationship between Hotel and Address may be inad- vertently omitted. 2.4. Construction method based on knowledge base Using the knowledge base as a basis for the construction method requires the prior construction of knowledge bases in related domains. The knowledge base must include basic rules and simple examples. When a user enters keywords to search for information, the rules in the knowledge base is used in order to filter data, while similar examples are displayed to make a possible comparison. When the required result is picked out, rules in the knowledge base again are used to establish related ontology as well as giving the summary and results. This type of method is different from the above-mentioned three methods since, the rules in the knowledge base can be regarded as a kind of ontological manifestation. The rules in the knowledge base are used to assemble related ontology (Alani et al., 2003). 2.5. Formal Concept Analysis Formal concept analysis (FCA) is a method for data analysis, knowledge representation and information manage- ment. It was proposed by Rudolf Wille in 1982 (Wille, 1982). In recent years, formal concept analysis has grown into an international research community with applications in many disciplines, such as linguistics, software engineering, psychol- ogy, medicine, AI, database, library science, ecology, information retrieval, etc. One of the distinctive features of formal concept analysis is its ability to generate graphic visualization from the structure of any given data set, more particularly in social science where it is almost impossible to perform a quantitative analysis. Formal concept analysis serves to augment formal analysis methods and to complement statistics and conceptual analysis. FCA also presents a lot of benefits to the field of information science. The exercise of FCA in mathematics can be used to explain classification systems. Formal classification system is capable of analyzing based on the consistence among relationships (Ganter & Wille, 1999). Stumme (2002) explains that FCA shifted emphasis to applications in computer science partly due to a merger with the conceptual graphs community (Sowa, 1984). An overview of the relationship between conceptual graphs and FCA is provided by Mineau, Stumme, and Wille (1999). 3. Research methodology This research uses ontology learning technology to construct conceptual maps of documents in order to provide an effective reference to users as they perform information searches. The main three problems are given as follows: S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–124 (1) How can document data produce the conceptual maps of ontology construction technology? (2) How can the degree of relationships among various concepts be expressed and studied? (3) How can users break through conceptual maps in order to perform quick and accurate searches? The document conceptual maps developed by ontology learning technology in this research are intended to provide a good reference for users when they perform information searches (Fig. 1). The entire system architecture in this study can be divided into three major subsystems and other relative components. These subsystems are enumerated and explained below: 3.1. Term parsing subsystem When documents of various data sources are entered, they must pass through different preprocessing methods in order for them to qualify in subsequent requirements. The five steps encountered in this phase are: (1) Elimination of document layout: as a result of having various data sources, document layout can also appear in various forms. Thus, the first step is to disregard all irrelevant information, such as: typesetting format, annotation and other additional information. The output of this phase is a data stream of characters. (2) Lexical analysis: lexical analysis is the process of transforming the data stream of characters into a data stream of terms (Baeza-Yates,& Ribeiro-Neto, 1999). English lexical analysis makes use of a space symbol or punctuation marks to convert data streams into a set of terms. (3) Elimination of stop words: in the second stage of the lexical analysis, we noticed that the most frequently used terms Fig. 1. System a normally do not have distinguishing or recognizing property. In fact, in one document, more than 80% of the terms are meaningless and are often filtered out during the analysis. The terms referred here usually are articles, prepositions, conjunctions, and other terms that do not constitute the main idea or concept of the document. Examples are a, as, and, etc. Eliminating these stop words not only saves memory space but also decreases complicated calculations. (4) Elimination of stemming words: different writers exercise differentwriting styles. For thisreason, the chances ofhaving slight variations in the context due to how a particular term was used are inevitable. Plurality, verbal nouns, and tenses can alter the basic form of the word. The standard form of the word or root word is used to replace its different forms. Say for instance the word ‘connect’, variations of this word are connecting, connection, connections, etc. Using the root word to replace all other variations of the same word can free up memory space and reduce complicated calculations. (5) Thesaurus: it is probable that different words can actually mean the same thing, thus, the thesaurus is used to disregard redundant terms (Baeza-Yates,& Ribeiro-Neto, 1999). 3.2. Ontology construction subsystem Upon changing the document content into a set of terms, the ontology construction subsystem adopts the ontology con- struction technology to produce document conceptual maps. This subsystem consists of two major components: 3.2.1. Establish a set of conceptual relationships and hierarchy among terms Here, we use the idea of formal concept analysis (Buchli, 2003; Ganter & Wille, 1999) to establish a set of conceptual relationships and hierarchy of terms. We believe that while the more general abstract concepts appear more often in rchitecture. S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–12 5 documents of some related subject, the technical concepts in more details appear less often. Generally speaking, there are three kinds of relationships that exist among concepts. These are independence, intersection and inheritance. In order to establish the concept relationships and hierarchy of the different terms, there are five steps to follow: Step 1: Produce the binary relation matrix between documents and terms In every document, a term that will best represent the main concept of the document must be obtained in the term retrieval subsystem. We do this by referring to the document set and the term set. If a term appears in a document, the corresponding entry of the matrix is labeled as ‘X’. From here we can generate the binary relation matrix between documents and terms. The initial starting point in using FCA is setting up a context (Buchli, 2003). A context is a triple: LZ(D,T,L). In this research, the context of the ontology is identified as L, the related document set of the ontology is represented by D, the related term set of the ontology is marked as T, and lastly, I is a binary relation between D and T: I4D!T. Step 2: Generate the concept set C If we let X be the partial set of D, and Y as the partial set of T, therefore, X4D, Y4T. The mappings: sðXÞ Z ft2Tjcd2X : ðt; dÞ2Ig; the common terms of X, and tðYÞ Z fd2Djct2Y : ðt; dÞ2Ig; the common documents of Y. Based on the above definitions, a concept is defined. A concept is a pair of sets: a set of documents and a set of terms (X,Y) such that: YZs(X) and XZt(Y). Therefore, a concept is a maximal collection of documents sharing common terms. Thus, taking concept c as an example, it means that the biggest document set that contains the common terms is in the maximal rectangle constituted by all the relationships I in the binary relation matrix. The set of all the concepts of c is represented by C. Step 3: Calculate hierarchy relationship of concepts The set of all the concepts of a given context forms a complete partial order. Thus, we define that a concept (X0,Y0) is a sub-concept of concept (X1,Y1), denoted by (X0,Y0)4(X1,Y1). In the event that the document set X1 of a term set Y1 is contained in the document set X2 of another term set Y2, denoted by X14X2, (X1,Y1) becomes the sub-concept of (X2, Y2), denoted by (X1,Y1)4(X2,Y2). For concept set C, it means c1(X1,Y1) becomes the sub-concept of c2(X2,Y2). Step 4: Generate the entire hierarchy of concepts It is possible for concept c to have various father concepts as well as sub-concepts. For this reason, computing various hierarchy relationships for different concepts is required in order to obtain the entire hierarchy of concepts. Each node in the hierarchy represents a concept. Given two elements (D1,T1) and (D2,T2) in the concept hierarchy, their supremum or join is defined as (Buchli, 2003): ðD1; T1ÞgðD2; T2Þ Z ðtðT1hT2Þ; T1hT2Þ: Let c1(X1,Y1) and c2(X2,Y2) be two concepts, the supremum of the two concepts is computed in order to determine their respective positions in the concept hierarchy. Step 5: Generate the inter-relationships of concepts After constructing the hierarchy relationships among concepts, we now identify the inter-relationships of concepts. Let c1(X1,Y1) and c2(X2,Y2) are two concepts, if Y13Y2 and Y23Y1, since the two concepts are partially contained by one another, it allows us to identify the inter-relationship between c1 and c2. 3.2.2. Calculate the degree of relevancy among concepts After establishing the relationships between concepts, we can begin to calculate the degree of relevancy among concepts which are not directly inherited. In this regard, let us examine the method of calculation formulated by Kang, Huh, Lee, & Kim (2000) in computing for the correlation between concepts. The formula and related variables are as follows: fjk Z relevancyðTj; TkÞ Z Pn iZ1 dijk Pn iZ1 dij !WeightingFactorðTkÞ (1) dijk Z tfijk !log10 N dfjk !wj � � (1.1) dij Z tfij !log10 N dfj !wj � � (1.2) WeightingFactorðTkÞ Z log10 N dfk log10N (1.3) Formula (1) describes the degree of relevancy between two terms. All degrees of relevancy have a corresponding direction. The significances computed by different terms as central points are different. For example, in formula (1), the central point of the calculation is Tj as the correlation between Tk and Tj is being established. Incidentally, formula (1) can be broken down into three other equations as seen in formulas (1.1), (1.2), and (1.3). Note that formulas (1.1) and (1.2) make use of the TF-IDF concept (Salton & McGill, 1983). In formula (1.1), dijk is decided by the frequency that both Tk and Tj appear simultaneously and the inverse document frequency. tfijk represents the frequency that Tj and Tk both appear in document i, dfjk represents the total document number that Tj and Tk appear together. Consequently, when both terms have higher relevancy, the frequency of Tk and Tj appearing in the same document should also be high and they should centralize in some specific documents. The same concept applies to formula (1.2). On the other hand, as seen in formula (1.3), WeightingFactor(Tk) corresponds to the specificity of Tk against the documents. As the term Tk becomes more general, the value of the WeightingFactor(Tk) decreases. The descrip- tion of variables in formula (1) is shown in Table 1. To further explain, let us use Fig. 2 as an example. Fig. 2 shows the frequency of the terms in every document. The result Fig. 3. Ontology conceptual map from Fig. 2. Table 1 Description of variables in formula (1) Variables Description Variables Description N Total number of keywords tfij Frequency of term j in document i tfijk Co-occurrence of term j,k in document I dfj Document fre- quency of term j wj Weight for inverse docu- ment frequency dfjk Document fre- quency of term j,k S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–126 calculated by formula (1) and the conceptual hierarchy generated by the formal concept analysis (FCA) constitute the ontology conceptual map in Fig. 3. In the same figure, the arrows with full lines serve to show the inheritance among concepts, while the dotted lines show the mutual relationships between two concepts. The number seen on the dotted lines is the significance of concepts, which follows a single direction. This is so because different concepts have their own related concepts and thus, have different significances. 3.3. Ontology management subsystem Ontology management subsystem has two major parts. First, for construction builders, the most important consideration is the absence of error in the hierarchy relationship among concepts rather than the correlation among concepts. As a result, it is necessaryfor builderstobeinvolvedanddetermine the validityof the ontological construction result. In the event that an error occurs in the hierarchy relationship or there was a failure in forming the necessary hierarchy relationship, the builder should examine the adequacy of data representation or other possible causes. From the standpoint of users, an error in the hierarchy will result in users’ misunderstanding about the concepts. When this happens, users will rather find another sources or methods for learning. Secondly, as far as the Semantic Web is concerned, Fig. 2. Matrix denoting frequencies of terms appearing in documents. erroneous ontology will lead to numerous errors when the intelligent agents perform an information search. In the blueprint of the Semantic Webs, intelligent agents search and examine different ontology, and then generate the final result. Thus, errors in the ontologywill correspondto errorsin the results provided by the intelligent agents. This happens because errors can be found if ontology is used by human while intelligent agents judge by the results so that the ontology with errors cannot be used by intelligent agents. With regard to the second part of this subsystem, it is the aim of this research to supply users with an effective searching interface that will facilitate their search. In addition, users can select a concept based on the conceptual map inFig.3.It willleadthemtouncovera relatedconcept,ortheycan simultaneously select several concepts to get the relationships among them as well as documents with relative significance. 4. Experiment design and results This research is conducted with the principal aim to upgrade the existing Internet applications. Data in the experiment come from Internet sources. The system proposed in this research is implemented in the Internet. Furthermore, it is seen from the system architecture in Fig. 1 that the system requires the use of some function library. For this reason, this research has chosen the Java language as the implementation language. 4.1. Experiment evaluation design The ultimate objective of the ontological construction technology in this research is to build a map for related ontological concepts that will help users in their search for relevant information. With the present ontological construction technology, errors could not be completely avoided in establishing the hierarchy relationship no matter with the techniques of dictionary-based, text clustering, association rules or the knowledge base. In this regard, we use the hierarchy relationship by comparing the concept nodes to Fig. 4. Diagram of a conceptual hierarchy. S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–12 7 obtain the accuracy rate of the entire ontology. Thus, in measuring the efficiency of the construction method, this research adopted the most commonly used measures in data mining, namely, precision and recall, for the general assessment (Han & Kamber, 2001). This is further illustrated in the following equations: Precision Z jfRelevantg&fRetrievedgj jfRetrievedgj (2) Recall Z jfRelevantg&fRetrievedgj jfRelevantgj (3) In measuring the conceptual hierarchy in ontology, precision refers to the accurate ratio of conceptual hierarchies that are automatically constructed, while recall refers to the accurate ratio of conceptual hierarchies that should be generated. Fig. 4 shows a diagram of a conceptual hierarchy. As seen in the diagram, the hierarchy has a total of eight concept nodes and eight conceptual relationships. Among them is an erroneously constructed conceptual relationship rep- resented by the full line. On the other hand, the dotted line shows conceptual relationship that should exist but was not automatically constructed. Thus, applying formulas (2) and (3) in measuring the conceptual hierarchy will yield a precision of ð8K1Þ=8Z87:5% and a recall of 7=ð8K1C1ÞZ87:5%. 4.2. Experiment 1 The usage scope of the experiment materials in experiment 1 is smaller while the contents are more identical in order to test the efficiency and accuracy of the construction method in this research. The experiment materials used were the recorded Table 2 Number of term set in different situations Original set of terms Set after usage of prop- erties of speech Number of term set 4468 865 Filtering rate (%) 100 19 dissertations found in the ‘Dissertation and thesis abstract system’ (http://datas.ncl.edu.tw/theabs/1/), provided the titles included the terms ‘data mining’. A total of 187 documents were collected. Wu, Day, and Hsu (2001) pointed out that subject words and keywords are usually composed of noun–verb and noun–noun terms. Thus, by the syntactical functions and morphological features of speech and the sifted speech base can filter out majority of the irrelevant terms. And due to the possibility that two or more terms can mean exactly the same thing, particularly proper nouns of foreign languages having multiple meanings and translations, it becomes necessary to establish a dictionary of synonyms that will facilitate the accurate translation of terms. Accomplishing this will certainly lead to a higher rate of efficiency. Finally, we come to the stop word list. Although by means of the properties of speech and synonyms can gather up most of the terms based on nouns, it is not the case that all nouns have the distinguishing meanings. For this reason, there is a need filter out the stop words in order to increase the efficiency rate. Table 2 shows the original set of terms collected from experiment 1, the set of terms after usage of properties of speech, synonyms, and stop words, as well as the final collection of terms and the filtering rate of the term set. If we take a closer look at the table, we will notice that the set after usage of properties of speech has the highest filtering rate. This is because, we only sieve out certain nouns and verbs to represent concepts. It also proves that describing terms and sentences have the highest number in any given document. The filtering rate of the synonyms and stop words may appear lower but its effect on the overall efficiency must not be overlooked. The final collection of terms is only 15% of the initial collection. Different numbers of term collection gave rise to different ontological content and efficiency levels. The amount of terms filtered out decides their capability to represent the ontology of data content. Too many conceptual nodes will generate noises, while too few conceptual nodes may not be adequate enough. This research stands by the fact that the hierarchy rate of ontology conceptual hierarchy can be used to denote the level of disorganization in data contents. Supposing a single conceptual node falls under the root node and there exists no other nodes below it, when such is the case, we believe that this particular node has a low correlation with other existing nodes and it is no longer considered as part of the conceptual hierarchy. It is commonly known as an independent node, see node C (filled with oblique lines) in Fig. 5. Thus, we have come to a definition of the hierarchy ratio as shown Set after usage of synonyms Set after usage of stop words Final collection of terms 764 676 676 17 15 15 http://datas.ncl.edu.tw/theabs/1/ Fig. 5. Diagram that shows hierarchy ratio and the independent node. S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–128 below: Hierarchy ratio Z 1K number of independent nodes total number of nodes (4) After obtaining the hierarchical ratio, the optimum number of term sets we obtained was 107. As a result, we combined the term sets with the FCA algorithm to produce the complete ontological framework. The ontological results from exper- iment 1 are shown in Table 3. Since, the term sets were filtered out, the number of documents decreased, from 187 to 184 in this experiment. On the other hand, the depth and breadth of the hierarchy reveal the range of the content included in the ontology. The wider the hierarchy expands, the more diversified and general the concepts are. While the deeper the hierarchy goes, the more detailed the contents become. The number of hierarchy relationships manifests the degree of complication of the nodes. This experiment had a total of 107 nodes but only 132 hierarchy relationships were produced. This means that the relationships among nodes are not at all complicated. The precision and recall of ontology determined by the experts are 84.1 and 81.1%, respectively. In Table 3, the numbers in the parentheses following the precision and recall represent the number of errors in the hierarchy relationships and un-constructed relationships. Based on the results of experiment 1, Fig. 6 is the result as the node with ‘Classification algorithm’ was chosen to become the inquiry output. From Fig. 6, we could see that the father node of ‘Classification algorithm’ is ‘Data mining’, while its sub-nodes are ‘Bayesian classification’, ‘Neural networks’, ‘Decision tree’, ‘Patent analysis’ and ‘Virus’. The five nodes connected to the node with ‘Classification algorithm’ as well as the papers related to ‘Classification algorithm’ and their respective significances are shown in Fig. 6. Therefore, we can represent the hierarchy relationships by using the relationships among nodes. Some other correlated nodes and Table 3 Ontology results of experiment 1 Experiment 1 Number of nodes Number of docu 107 184 Hierarchy ratio Number of hiera relationships 83.18% 132 documents are also used to convey the significance of resources. 4.3. Experiment 2 The primary objective of this experiment is to test the efficiency and accuracy of the data such that the experimental data contents are more general and diversified. The source of data is the timely news found in http://news.pchome.com.tw/. The primary subject is the news content divided into nine sections, namely: politics, society, finance and economics, science and technology, entertainment, sports, lifestyle, international news, and cross-strait. The news materials to be tested were divided into two groups. The first is under the ‘date’ unit. This group was tested for a total of seven days, from January 14 to January 20. Data were gathered every night after 8 pm. The second one belongs to the ‘category’ group and is composed of nine categories. Same length of time was spent in testing the materials of the same category. Each category consists of seven day’s worth of materials. The number and distribution of news materials are shown in Table 4 below. In each experiment, we were able to obtain the optimum collection of terms and performed the FCA algorithm to produce a complete ontological framework. Due to the filtering process done on the term sets, the number of documents decreased. Furthermore, we noticed that the number of ‘category’ nodes is fewer and there is a more reduction in the documents. The cause of this, as we have inferred, is the centralization of the range of ‘category’ data. The depth and breadth of the hierarchy show the scope range of the ontology. The wider the hierarchy extends, the more diversified and general the concept is. On the other hand, the deeper the hierarchy goes, the more detailed the content becomes. We also noticed that the ‘category’ group has a higher average depth and a smaller width than the ‘date’ group. This is just as we expected since, the scope of daily news encompasses everything, while the data in the ‘category’ group merely include their respective categories. The results of experiment 2 can be seen in Table 5, Figs. 7 and 8. We notice from the results that the ontology generated by the ‘category’ group showed a higher precision, particularly the ‘finance and economics’ category with 88.68%. The precision was at its lowest on January 17 with only 72.20%. Another observation worth noting is the big changes under recall. The lowest recall rate was 73.58%. This verifies that the method used in this research is indeed feasible and completely satisfies the requirements of dynamically generating classifi- cation frameworks. ments Depth of hierarchy Breadth of hierarchy 5 47 rchy Precision Recall 84.1% (21) 81.1% (25) http://news.pchome.com.tw/ Fig. 6. Result that the node with ‘Classification algorithm’ is chosen to become the inquiry output. Table 4 The number of news materials and their distribution 1/14 1/15 1/16 1/17 1/18 1/19 1/20 Total Politics 58 48 33 39 40 53 48 319 Society 66 45 34 32 33 49 44 303 Finance and economics 67 48 35 42 53 46 36 327 Science and technology 52 30 32 47 37 33 38 269 Entertainment 64 40 49 61 78 68 51 411 Sports 61 41 40 46 46 47 42 323 Lifestyle 55 51 41 40 39 50 49 325 International 47 42 37 45 35 32 35 273 Cross-strait 61 33 31 35 34 30 33 257 Total 531 378 332 387 395 408 376 2807 S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–12 9 Table 5 Results of experiment 2 Number of nodes Number of documents Depth of hierarchy Breadth of hierarchy Hierarchy ratio (%) Number of hierarchy relationships Precision Recall 1/14 139 424 2 58 79.14 178 75.8%(43) 80.78%(36) 1/15 167 323 2 61 84.43 220 78.6%(47) 81.36%(41) 1/16 118 362 2 55 73.73 120 80.8%(19) 77.50%(27) 1/17 277 405 4 81 88.45 435 72.2%(121) 88.74%(49) 1/18 204 424 2 54 88.73 304 75.7%(74) 87.50%(38) 1/19 154 439 2 64 76.62 170 77.6%(38) 80.00%(34) 1/20 107 319 2 73 89.16 234 81.62%(43) 83.76%(38) Politics 47 218 3 45 100 53 86.79%(7) 75.47%(13) Society 100 210 3 33 87.00 134 81.34%(25) 86.57%(18) Finance and Econ. 48 244 4 46 100 53 88.68%(6) 73.58%(14) Science and tech. 99 198 2 39 84.85 163 82.21%(29) 87.12%(21) Entertainment 57 340 3 34 100 67 86.57%(9) 86.57%(9) Sports 77 234 3 35 100 59 83.05%(10) 80.66%(12) Lifestyle 157 226 3 64 81.53 268 82.09%(48) 86.19%(37) International 91 206 3 41 80.22 82 80.27%(17) 82.93%(14) Cross-Strait 81 193 4 32 81.48 79 82.28%(14) 81.01%(15) S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–1210 5. Conclusion As a result of the extensive developments in the Internet, sharing knowledge with each other has finally become a reality. Unfortunately, it is for the same reason that we are facing an overflow of data and information. Nevertheless, the Semantic Web concept proposed by Berners-Lee and Fischetti (1999) paved the way to the formulation of possible and effective solutions. The most vital tools in searching for information and related resources in a Semantic Web are the ontology and intelligent agent. In the field of ontology, ontological framework is normally formed using manual or semi-automated methods requiring the expertise of developers and specialists. This is highly incompatible with the developments of World Wide Web as well as the new E-technology because it restricts the process of knowledge sharing. Consequently, this research adopted the formal concept analysis algorithm to study the automation of developing Fig. 7. Result of experiment 2 based on dates. ontological framework and with the hope of fully satisfying the requirements of such. Since, the Semantic Web technology is still in its research stage, it is still difficult to thoroughly assess the cost and efficiency of an automatic ontology development. In this research, analyzing the news material was an attempt to develop the ontology of the news website http://news.pchome. com.tw/. The objective is to apply it in the future Semantic Webs so that users can find the information they need at a much faster pace, as well as to recommend users with the most related content. The experiment materials used in this research consist of dissertation papers and news information. The precision and recall of the ontological framework generated by the different data were computed. The ontological framework tallied with the data content concept and is capable of solving problems associated with the flexibility of the classification framework and conceptual relationships. From the results of experiments 1 and 2, we knew that the methods used in this research complied with the conceptual hierarchy and the relationships in the Fig. 8. Result of experiment 2 based on news categories. http://news.pchome.com.tw/ http://news.pchome.com.tw/ Table 6 Different ontology construction methods and comparison of limitations Limitations of the method Construction method Dictionary-based Text clustering Association rules Knowledge base FCA Construction method depends on the specific data X X Construction method cannot be used independently X X Preprocessing of data can influence the construction method X Faces a difficult choice between effi- cient management and data loss a X X Output is incapable of being interpreted X a An enormous collection of terms can result in lower efficiency. Although reducing the number of term collection may raise up the efficiency level, it runs the risk of losing portions of the information. Thus, a choice must be made between the two factors. S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–12 11 hierarchy produced by the data content. Unlike in the existing ontology development technology, it does not require, for example, a dictionary with a specific domain for dictionary- based construction method and does not have the difficult problems encountered in text clustering construction method. Thus, the methods applied in this research have overcome the limitations of the existing methods with regard to automatically generating conceptual hierarchy. If we take a look again at the results of experiments 1 and 2, we will see that the methods used in the experiments were able to obtain a higher precision and recall rate particularly in smaller data scope and in data with more identical content. Furthermore, the methods produced favorable results in its experiment on the general news information. In connection with the ontology conceptual hierarchy, it was observed that smaller data scope or data with more identical content will produce deeper conceptual hierarchies with narrower breadths. When general data have shallower conceptual hierarchy, the breadth of the hierarchy is wider. This is because the content of general data are more dispersed so they tend to produce a more flat conceptual hierarchy, while the specific data content generate a more complete ontological conceptual hierarchy. All these prove that the methods adopted in this research are more suitable to data with smaller scopes. We believe that this research is able to make the following contributions: (1) It solves the limitations of the existing ontological construction for classification framework. At present, the ontological construction technology is divided into four kinds, namely: dictionary-based, text clustering, association rules, and knowledge bases. This research experiment attempted to work these limitations out and came out with comparative measures as shown in Table 6. This research made use of the formal concept analysis and combined it with the conceptual relationships to construct the ontological concept diagram. Table 6 shows how the ontological construction method used in this research was able to overcome and solve the limitations of existing methods. In addition, the results of experiment 1 and 2 proved that the methods were able to fully satisfy the requirements for the automatic construction of ontological classification framework. (2) It solves the problems of conceptual relationships and resources significance. The categorical relationships manifested by the present classification framework are usually absolute but are in no way capable of showing the relative significance of identical classification level. This problem can result in a fixed path when searching for a concept and can easily omit vital information. The methods proposed in this research can construct the relationships between various concepts and sort the significance among concepts. Due to the less flexibility brought about by manually defining the classification framework, the correlation between concepts becomes more difficult to identify. Furthermore, the significance of different document resources against concepts can be expressed by different combinations of concepts. The document significances obtained by different conceptual combinations will not be exactly the same, thus, helping the users to increase the precision rate of their searches and reduce the time spent in searching for information. Finally, in the present ontological construction methods, no extensive automated construction method exists that can absolutely fulfill the requirements of a Semantic Web. Above all is the enormous amount of information that had increased exponentially since, the beginning of the Internet technology. Therefore, the manual construction of ontology is definitely not the answer if we want to see the full development of Semantic Webs. What is needed is an automatic construction method that will solve the difficulties encountered in the present. According to our experiment results, the method we proposed can achieve more favorable results compared to other methods under similar data content with smaller scopes. Therefore, if we can integrate it with other ontological construction methods, say for instance the text clustering method, the general data contents can be segmented and grouped into similar contents together, and then proceed to apply the methods suggested in this research. We believe that S.-S. Weng et al. / Expert Systems with Applications 31 (2006) 1–1212 the system in this research can thus yield better results. Moreover, it will merit further investigation and analysis by future research studies. References Alani, H., Kim, S., Millard, D. E., Weal, M. J., Hall, W., Lewis, P. H., et al. (2003). Automatic ontology-based knowledge extraction from web documents. IEEE Intelligent Systems, 18(1), 14–21. Baeza-Yates, R., & Ribeiro-Neto, B. (1999). Modern information retrieval. New York: Addison-Wesley. Berners-Lee, T., & Fischetti, M. (1999). Weaving the web: The original design and ultimate destiny of the world wide web by its inventor. San Francisco, CA: HarperAudio. Buchli, F. (2003). Detecting software patterns using formal concept analysis. Bern, Switzerland: University of Bern. Ganter, B., & Wille, R. (1999). Formal concept analysis: Mathematical foundations. New York: Springer. Han, J., & Kamber, M. (2001). Data mining: Concepts and techniques. New York: Morgan Kaufmann. Hotho, A., Maedche, A., & Staab, S. (2001). Ontology-based text clustering Proceedings of the IJCAI-2001 workshop text learning: Beyond super- vision, Seattle. Kang, S. H., Huh, W., Lee, S., & Kim, Y. (2000). Automatic classification of WWW documents using a neural network Proceedings of international conference on production research, Bangkok. Khan, L., & Luo, F. (2002). Ontology construction for information selection Proceedings of the 14th IEEE international conference on tools with artificial intelligence, Washington, DC pp. 122–127. Kietz, J. U., Maedche, A., & Volz, R. (2000). A method for semi-automatic ontology acquisition from a corporate intranet Proceedings of the EKAW’2000 workshop on ontologies and texts. France: Juan-les-Pins. Maedche, A., & Staab, S. (2001). Ontology learning for the semantic web. IEEE Intelligent Systems, 16(2), 72–79. Maedche, A., & Staab, S. (2000). Discovering conceptual relations from text Proceedings of the 14th European conference on artificial intelligence, Berlin pp. 321–325. Mineau, G., Stumme, G., & Wille, R. (1999). Conceptual structures represented by conceptual graphs and formal concept analysis. In W. Tepfenhart, & W. Cyre, Conceptual structures: Standards and practices. Proceedings of the seventh international conference on conceptual structures (pp. 423–441). Berlin: Springer. Salton, G., & McGill, M. (1983). Introduction to modern information retrieval. New York, NY: McGraw Hill. Sowa, J. (1984). Conceptual structures: Information processing in mind and machine. Reading, MA: Addison-Wesley. Stumme, G. (2002). Formal concept analysis on its way from mathematics to computer science. In U. Priss, D. Corbett, & G. Angelova (Eds.), Conceptual structures: Integration and interfaces, 10th international conference on conceptual structures (pp. 2–19). Berlin: Springer. Tan, K. W., Han, H., & Elmasri, R. (2000). Web data cleansing and preparation for ontology extraction using WordNet Proceedings of the first international conference on web information systems engineering, Hong Kong, Vol. 2 pp. 11–18. Wei, J., Bressan, S., & Ooi, B. C. (2000). Mining term association rules for automatic global query expansion: Methodology and preliminary results Proceedings of the first international conference on web information systems engineering, Hong Kong, Vol. 1 pp. 366–373. Wille, R. (1982). Restructuring lattice theory: An approach based on hierarchies of concepts. In I. Rival (Ed.), Ordered sets (pp. 445–470). Dordrecht-Boston: Reidel. Wu, S. H., Day, M. Y., & Hsu, W. L. (2001). FAQ—centered organizational memory Proceedings of the IJCAI’2001 workshop on knowledge manage- ment and organizational memories, Seattle. Ontology construction for information classification Introduction Lack of flexibility in the classification framework Conceptual relationships and significance of resources Review of related literature Dictionary-based construction method Text-clustering-based construction method Construction method based on association rules Construction method based on knowledge base Formal Concept Analysis Research methodology Term parsing subsystem Ontology construction subsystem Ontology management subsystem Experiment design and results Experiment evaluation design Experiment 1 Experiment 2 Conclusion References 
work_4xw647c65jdk3bdu6iyboysxwi	----	A rule-based expert system for music perception Behavior Research Methods, Instruments, & Computers /988, 20 (2), 255-262 A rule-based expert system for music perception JACQUELINE A. JONES, BENJAMIN O. MILLER, and DON L. SCARBOROUGH Brooklyn College of the City University of New York, Brooklyn, New York We describe an expert system written in Pascal to simulate part of Lerdahl and Jackendoffs (1983) theory of music perception. This program, a production system based on a Hearsay II ar- chitecture (Nii, 1986), illustrates a number of techniques, including program modularity, com- plex data structures, and simulated parallelism using operating system concepts. People perceive structure in the music they hear. Con- sider the simple song "Row, Row, Row Your Boat." Asked to sing this melody, most people would, without thinking about it, stress certain syllables, breaking the melody up into sections corresponding to the following lines: ROW row row your boat GENT ly down the stream MER ri ly mer ri ly mer ri ly mer ri ly LIFE is but a dream, in which italicized syllables are stressed, capitalized syl- lables more than lowercase. The stresses are not random or idiosyncratic. The song divides into rhythmically simi- lar halves, each consisting of two lines, with two stressed notes per line, equally spaced in time. This pattern of stresses is the metric structure. The division into lines is the grouping structure, also an important part of music perception. Listeners also perceive the tonality of the piece-the key it is in. After listening to a little bit of "Row, Row," a listener knows what chord to expect to end the piece. These are the kinds of intuitions that Lerdahl and Jackendoff (1983) formalized in A Genera- tive Theory of Tonal Music (GTTM). The theory does this by means of four stages of analysis, each of which is em- bodied in a set of rules. This theory is, to our knowledge, the most comprehensive formal theory of music percep- tion published to date. We are building a computer simulation of GTTM to test the theory. The basic framework of the simulation is in place and currently consists of about 10,000 lines of code. Turbo Pascal was chosen for initial development of our simulation because of the many strengths of the language (see Jones & Harrow, 1986; Kaplan, 1985) and because it is a language that we know well. There are, however, problems with using Pascal for a large project with several programmers. The authors are listed alphabetically. All contributedequally to this work. This researchwas supportedin part by a National ScienceFoun- dation Graduate Fellowshipawarded to Miller, and by a PSC-CUNY Research Award to Jones. We thank Hollis Scarborough for help in preparing the figures. Requestsfor reprints should be sent to Jacque- line A. Jones, Department of Computer andInformation Science, Brook- lyn College, Brooklyn, NY 11210, or to Don Scarborough, Department of Psychology, Brooklyn College, Brooklyn, NY 11210. First, standard Pascal (and dialects such as Version 3.01 of Turbo Pascal that we have been using) provides limited support for modular development oflarge programs. All subprocedures are nested within the main procedure, and there is no separate compilation of subprocedures. Thus, if a line of code is modified, the entire program must be recompiled. Although compilation is fast with Turbo Pascal, the size of our program makes this time- consuming. More important, standard Pascal does not allow a sub- program to define static variables-that is, variables that are allocated permanent memory locations for the life of the program. Thus it is often necessary to use global variables-that is, variables declared in the main procedure-for this purpose. However, with global vari- ables, modifying a piece of code anywhere in the pro- gram may affect other aspects of the program in unantic- ipated ways. Also, name conflicts arise when several programmers are defining global variables for different subprograms. To minimize such problems, we adopted several strate- gies for program development. First, we borrowed a Mod- ula 2 approach (Wirth, 1985) to implementing the sys- tem. Each module of the program is split into two separate source files: a definition file (with a DEF file extension) and an implementation file (with an IMP file extension). The DEF file contains all the global variables whose values must be retained between calls to that module; these global variables take the place of static variables. The IMP file contains the actual implementation code for the mod- ule. To minimize name conflicts, we adopted a conven- tion of prefixing all true global type definitions and global variables with an underbar. The names of global variables that mimic static variables-that is, variables that exist only for the use of the module in which they are defined- are preceded by the initials of the name of the module in which they are defined. A small main source file defines the overall program structure. This file contains global type definitions and variable declarations that must be shared by all parts of the program. The main file then specifies the DEF files that are to be "included" in this file in a special step dur- ing compilation. Collectively, then, the main ftle and the DEF files specify the entire global environment of the pro- gram. The main file next specifies the IMP files as an 255 Copyright 1988 Psychonomic Society, Inc. 256 JONES, MILLER, AND SCARBOROUGH additional set of "include" files to be brought into the main file during compilation. Finally, the main file con- tains a short main procedure that calls an initialization subroutine and then passes control to the scheduling rou- tine described below. A RULE-BASED EXPERT SYSTEM Although we considered implementing GTIM as a neu- ral network (e.g., Grossberg, 1980; Rumelhart & McClel- land, 1986; Schneider, 1987), and hope to do so eventu- ally, there are several problems with the connectionist approach that led us to reject this approach for now. First, simulation of connectionist networks requires powerful computational resources for anything but the smallest net- works. Second, the high degree of interaction between system components in a neural network makes modular development difficult. Third, many properties of these systems are not well understood. Fourth, many aspects of GTIM are not well specified or articulated. To attempt to implement a complex and incompletely specified rule- based theory (GTIM) in terms of a complex, interactive, nonsymbolic neural network is to risk chaos. Thus, our first goal is to see whether the theory can be adequately implemented at the symbolic level at which it is cast. However, as noted below, some parts of the simulation are easily cast into a connectionist framework. We chose to develop this system as a rule-based expert system because (1) GTIM is expressed in terms of rules; (2) the technology for rule-based systems is developing rapidly, and there are many available techniques to call upon; and (3) a rule-based approach supports implemen- tation of a modular system. Although GTTM is expressed as a set of rules, many of the rules cannot be translated directly into simula- tion rules because, in general, the rules are not statements about how a piece of music should be analyzed, but rather are constraints that such an analysis should satisfy. A big hurdle in implementing GTTM lies in discovering psychologically plausible procedures that can produce analyses satisfying such constraints. We want the simu- lation to be plausible as a real-time implementation of the psychological processes underlying music perception. To this end, we chose to implement the theory as a rule-based production system (Newell & Simon, 1972; Winston, 1984), using a blackboard type of architecture (Nii, 1986). Blackboard systems have been popular in psychologi- cal modeling (e.g., McClelland & Rumelhart, 1981). A blackboard system has three basic components: (1) a global data structure called the blackboard; (2) a set of knowledge sources (KSs), each of which encapsulates knowledge about some aspect of the problem; and (3) a control structure that monitors and controls which KSs are active. Our current implementation of GTIM in terms of these components is described below. ARCIDTECTURE OF THE SYSTEM The Blackboard The blackboard stores all the input data about a problem as well as information accumulated as the analysis pro- ceeds. A blackboard is generally divided into several levels, each representing a different type of information. In our implementation, the lowest level of the blackboard contains data about the individual notes in the music. Other levels of the blackboard represent information about the analysis of the piece, including levels for grouping infor- mation, metric information, and tonality information, and levels based on higher level combinations of this infor- mation. In our implementation, information within each of these levels is stored in linked lists. Pointers between levels of the blackboard provide appropriate linking and synchronization between levels. Our blackboard data structure had to be suitable for representing musical input. This representation had to be compact, because musical notation is very dense, contain- ing thousands of notes, each supplying many different kinds of information. For this project, we assumed that the basic elements of a listener's experience include such information as pitch, note duration, dynamic level, and timbre. Thus, it seemed reasonable to build the simula- tion assuming this level of knowledge as a starting point, an assumption also made by Lerdahl and Jackendoff (1983). Input to the simulation program is a symbolic representation of such information, which forms the lowest level of the blackboard and represents what the listener hears. We needed to represent the time continuum of the music as well as the multiple events taking place at each time, with some notes continuing to sound as others begin and others end. We used a matrix in which each row represents a different instrument (or voice) and the sequence of notes it plays, and each column represents the notes played by all instruments at a specific time. A simple array representation would not be practical, given the limited memory of most personal computers, because it would be a "sparse" matrix (Tenenbaum & Augenstein, 1986), with far more storage occupied by blank information than by actual information. Suppose we wanted to represent piano music. At any moment, up to 10 notes may be played (more, if we consider sustained notes and cases where a-finger plays 2 notes), so we might use an array with 10 rows, representing 10 possible notes at any time, and with columns representing successive notes in time. However, usually fewer than 10 notes are actually played at any time. An example of this, taken from Schubert, is illustrated schematically in Figure la. The passage begins with an 8-note chord held for two time intervals, followed by a sequence of 14 single notes. This general pattern is repeated eight times in the music. If we used an array of 10 rows and 16 columns to store the music, then in each occurrence of this rhythmic pattern AN EXPERT SYSTEM FOR MUSIC PERCEPTION 257 we would have an array with 160 cells, but 130 would be empty. To avoid this problem, we used a sparse matrix im- plemented as a linked list. A linked list is a dynamic data structure that contains only those pieces of information that are necessary at any given time. An element of a linked list, known as a node, is a record that can contain several different kinds of data at the same time. An array exists in storage from the moment it is declared, whether it is used or not, but nodes in a linked list are allocated only as needed. The nodes are linked by pointers. Pointers-graphically represented as arrows-point to places where data are stored. One node can point to the next node or to the previous node, allowing us to create a list. Figure 1b shows the same Schubert example as a sparse matrix using linked lists. Here, the top line represents nodes that simply mark off successive events in the music. Each box that is linked to an event is itself a node that stores information about an actual note. In this represen- (a) ~O - 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 - ~- --- ---- --- 3 I 6 10 o o o o r-----f--+_~r__+_-._+_-_+_--l--+--_+_--_t_-+_~;-_+_-_+_-_+_-- o III II) '0 7 z ! 4 c.. CI 5 c 'il .Q 8 0 9 0 10 0 --t--f---I--.-+----+-- 1---- I-------+--+--f--.-+--l 2 34567 8 9 10 11 12 13 14 15 16 lime ~ (b) TIme -~ Events I c 0. CI c 'il .Q ~.... o z Figure 1. (a) Example of a "sparse" array with many empty cells. (b) A linked-list representation of (a). 258 JONES, MILLER, AND SCARBOROUGH tation, each occurrence of the Schubert pattern requires 46 nodes, compared with 160 entries in the array im- plementation. A further advantage of the linked-list representation is that it requires no prior commitment as to the number of notes that can be represented at any time or the number of notes per measure. To simplify computation, the blackboard is a single global data structure that integrates all four facets of the simulation: the score, the metric structure, the grouping structure, and the tonality analysis. The "backbone" of the blackboard data structure is called the notochord. This consists of a doubly linked list of nodes corresponding to successive events in the score. Each of these nodes, called event nodes, contains only one piece of informa- tion about the piece: the offset (time) from the beginning of the previous event to the beginning of the current event. An event node also stores pointers connecting it to nodes containing actual note information, as well as to the group- ing and metric representations. Note that the event nodes themselves contain no score information. This is reserved for the note nodes, each of which is a record of all the information about a single note in a score. In addition to this information about each note, a note node contains a set of pointers that link it to other notes and to the notochord. The information in note 1 of "Row, Row" is shown in Figure 2 in simplified schematic form. (Note that there is no next voice because this is monophonic input. Assume that the soprano sings the melody.) The Knowledge Sources The GTTM rules represent sources of knowledge about how a piece should be analyzed. Where possible, we im- plemented each GTTM rule as a separate KS. A KS can be thought of as an "if-then" production rule; that is, a KS is a rule that specifies an action (or consequent) that applies whenever certain conditions (the antecedents or event "if' parts) occur. An example of one of the simpler GTTM rules (Grouping Preference Rule 3d) is as follows: IF ntn2nJ14 is a 4-note sequence, and n, and n, differ in duration, and n, and n2 have equal durations, and n, and 14 have equal durations, THEN there is evidence of a grouping boundary be- tween n, and nJ' In the GTTM simulation, such a rule is implemented as a KS module containing two parts: (1) a pattern match- ing procedure that looks at the blackboard for the speci- fied antecedent pattern; and (2) a consequent procedure that makes appropriate changes to the blackboard. To create a uniform interface between different KSs, we adopted a limited form of "object oriented program- ming" (Jacky & Kalet, 1987). Each KS is treated as an "object" that controls certain data structures, and that contains a set of methods for operating on these data struc- tures. The KS operates on the data structure when it is sent a message specifying an operation to perform. We designed our KSs so that each responds to a limited set of messages (e.g., Initialize, Evaluate, Execute), thus facilitating a modular design that makes it fairly easy to add and replace KSs. The Control Structure The production rules (the KSs) operate under control of a scheduling procedure that monitors the activities of the KSs and the state of the blackboard. The scheduling mechanism reflects, in part, the fact that the simulation must be run on a serial von Neumann type of computer, where only a single KS can be active at anyone time. However, the scheduling mechanism can also represent changes in attention and focus as the analysis of the piece proceeds. Furthermore, when appropriately implemented, the scheduler can simulate parallel concurrent activity of node pre vi ous note for this voice ( ni 1) / voice = soprano pitch = C in 3rd octave duration = dotted quarter loudness = mf articulation = normal offset from p r e v. note a next voice ( ni 1) / next note (note 2) Figure 2. Schematic representation of the first note node for "Row, Row, Row Your Boat." AN EXPERT SYSTEM FOR MUSIC PERCEPTION 259 many KSs. The key to this aspect of our simulation lies in the window that specifies the portion of the blackboard that is "visible." A KS can examine only the portion of the blackboard within the window. The window represents the capacity of short-term memory, which applies to music as well as to verbal material (Dowling & Harwood, 1986). In contrast, Ler- dahl and Jackendoff(1983) offered analyses that assumed unlimited memory. Clearly, though, the perception of notes currently being heard is affected by how much a listener retains of the previous events. Therefore, our simulationis organized around the concept of a temporal window, so that in running the simulation we can vary parametrically the amount of prior input and analysis available at any point. Initially, the program reads a file of information about a piece of music and builds the notochord data structure with the linked note lists to represent the piece. An initial portion of the notochord is then marked off by the win- dow. The portion of the piece within the window represents what a listener first hears. This might be any- thing from 1 note to the entire piece, although currently we are working with a window size of 6 to 10 notes. The scheduling procedure maintains a task table, which is an array of records, one record per KS. The array en- try for a KS provides the scheduler with information about the status of the KS, such as whether it is currently run- nable and its priority. A KS can run if the conditions it requires are currently met. The scheduler selects a KS on the basis of which KSs are runnable and the priority blackboard assigned to each. It then copies the task table entry for that KS to a global variable, where it is accessible to the KS, and then sends the KS a message specifying the ac- tion to be performed (e.g., Modify blackboard). The KS can modify entries in its task table entry to tell the scheduler about changes to its status and/or the actions it has performed. An outline of the system is shown in Figure 3, which also shows the Message Manager mod- ule that controls all output messages sent by the scheduler and the KSs. This general framework for controlling execution lends itself to several scheduling algorithms. In round-robin scheduling, the schedulersimply goes down the list of en- tries in the task table, giving each runnable KS a chance to execute within the current window. Then the window is advanced (the new window may overlap the old and may exclude earlier events), and the scheduler calls each KS so the KS may evaluate its applicability to the new window. The schedulerthen calls the runnable KSs again. A more complex algorithm schedules KSs on the basis of priority. In this method, the scheduler maintains a list of runnable KSs in order of priority and always calls the KS with the highest priority. This scheduling procedure simulates parallel execution of the KSs in much the same way that multiuser operat- ing systems such as UNIX produce apparent parallel ex- ecutionof severaltasks: The scheduler, whichkeeps track of all the tasks in the system, selects a task and allocates a small amount of time to it, and then selects another task for execution whenthe first one has usedits allocated time, to flIes, prInter, console window 1 KSl KS2 I KS' I( ) KSj param~ters I KSn KS task table --@ InactIve knowledge 80urcetr Figure 3. Control Structure for the simulation. KS = Knowledge structure. 260 JONES, MILLER, AND SCARBOROUGH finished, or blocked itself because of some other limita- tion (Deitel, 1983). In our simulation, the scheduler does not directly control the amount of time that a KS has for execution. Rather, the size of the window determines how much data a KS has to work with, and this in turn deter- mines how long it will execute before running out of in- put data. With a small window, the effect is that all KSs appear to be executing concurrently and no KS can get very much ahead of or behind any other KS. Scheduling permits more complex control if priorities are changed dynamically during analysis. Changing pri- orities is one way in which the system can incorporate a top-down component in the analysis: On the basis of earlier events in the piece, the system can learn which KSs are probably most appropriate, and this can be reflected in the priorities. As we develop the simulation, we will explore more complicated scheduling procedures. For example, the order in which several KSs are allowed to modify the blackboard can have important effects if the blackboard information is "nonmonotonic" (Rich, 1983), that is, if blackboard information does not simply accumulate, but can be removed or revised. In this con- text, we may explore other variations in blackboard ar- chitectures (e.g., Hayes-Roth, 1985). Grouping Analysis Grouping preference rules and subrules identify possi- ble grouping boundaries in the score on the basis of such qualities as scale distance between notes, duration of time between notes, length of notes, and articulation of notes. Our prototype has KSs so far for two of the seven group- ing preference rules, which comprise about nine subrules. Each subrule routine determines whether it finds a group boundary. If a subrule determines that it has found a boundary between two notes, a grouping node is created and added to the list of boundaries produced by that rule. Different grouping preference rules identify different boundaries, with varying amounts of evidence in their favor. This state of affairs must eventually give way to a single well-formed hierarchy of boundaries based on GPR3d-note length GPR3c - artlculatlon GPR3a - pltdl GPR2b - attaok point GPR2a - Blur/relt Notochord the strength of the evidence. This distillation will be car- ried out by higher level rules operating on the data struc- ture(s) built by the grouping rules. Each grouping KS creates a separate linked list of nodes running parallel to the notochord, one node for each can- didate boundary it has found. Each grouping node is linked to the appropriate event in the notochord, as well as to the next grouping node created by the same KS, and to any other nodes produced by other grouping rules that have also identified a boundary at the same event. Link- ing all nodes supporting a boundary at a particular event makes it easy for a higher level rule to evaluate all the evidence for a particular boundary. This is illustrated in Figure 4, which shows the grouping analysis for "Row, Row." The boundary markers, attached to the note fol- lowing the boundary, show results that replicate what we obtained by applying the rules by hand. The appropriate set of weights applied to the rules will select the conjunc- tion of candidate boundaries between notes 11 and 12 (af- ter "stream") as the most important. This boundary di- vides the piece into two half-songs. The boundaries between notes 5 and 6 (after "boat") and between notes 23 and 24 (after the last "merrily") will be selected as the second most important. As noted in the introduction, these boundaries also correspond to our intuitive feeling that the music divides into lines at these points. Metric Analysis The metric rules of GTTM cannot be translated directly into algorithms. That is, whereas the grouping rules them- selves constitute parsing algorithms, the metric rules specify only what the metric hierarchy ought to look like when the analysis is finished. For this reason, we carry out the metric analysis in a single pass with a collection of routines that embody a single algorithm. The output of this algorithm conforms to the requirements ofGTTM, but the algorithm itself is not taken directly from the GTTM rules. Instead, our approach is an adaptation of the grid theory of Povel (1984; Povel & Essens, 1985). Briefly, grid theory works by trying to fit different-sized Row. row, row your boat. gen-tly down thestream-- M..- ~ 1)4 mer- rf- Iy. mer- rf- Iy. mtr- rf- Iy, Ute II but Q dreclm. Figure 4. Grouping analysis of "Row, Row, Row Your Boat." GPR = grouping preference rule. AN EXPERT SYSTEM FOR MUSIC PERCEPTION 261 J J. J. 11 ne half-song song ,-. J. J. J JJ -rJ J'J JJ. J '1mmmmJ JJ JJ................................................. Figure 5. Metric analysis of "Row, Row, Row Your Boat." The number of dots indicates the relative rhythmic salience of the notes. grids of equally spaced marks (grid ticks) to the musical events and then determining the degree of stress, or good- ness of fit, between each of the grids and the events. We have modified the procedures of grid theory in several ways. For instance, only those grids suggested by the in- tervals in the music are considered. Each grid that fits corresponds to a metric level in the GTTM hierarchy. Our use of the grid stress is not to identify the one best grid, but rather to eliminate from the hierarchy those levels that least fit the music, that is, levels not heard by a listener. The representation of metric structure is analogous to that outlined above for the score. Metric nodes contain information about the metric hierarchy at each beat, as well as pointers to other items and to the notochord. Here again, the linked-list structure saves space. If a metric hi- erarchy were represented by a two-dimensional array it would be quite sparse. By allocating a node for each beat rather than for each beat-by-level combination, the linked list of metric nodes wastes no space. The metric analysis of "Row, Row" is shown in Figure 5. Tonality Analysis Listeners can identify the key of a piece, and this can affect both meter and grouping. For example, strong beats in the metric structure generally fallon the important pitches of the key, as is clear in the "Row, Row" exam- ple, in which the accents generally coincide with the notes C, E, and G. On a piano keyboard, there are 12 black and white keys in each octave, representing the chromatic scale. In anyone piece of music, some of these notes are more important than others, reflecting the key in which the piece is written. Thus, "Row, Row," which is in the key of C major, contains 25 notes, but only 5 (C, D, E, F, and G) of the possible 12 chromatic notes occur. C occurs eight times, E occurs five times, and G occurs five times. Thus, the notes C, E, and G account for 18 of the 25 notes, reflecting the fact that C, E, and G form the C major chord, which is the fundamental (tonic) chord for the key of C major. Lerdahl and Jackendoff (1983) did not discuss how listeners extract the tonality of a piece, but the example just given suggests a simple algorithm: If we count the frequency of the notes of the chromatic scale, the most common notes will generally identify the key of the piece, as Simon (1968) has shown. Bharucha (1987) adapted this general idea to a network model, and we have followed a similar approach. In our implementation, each note of the chromatic scale is represented by a pitch node. These pitch nodes are connected to chordnodes; each chord node receives input from just three pitch nodes. Finally, sets of three chord nodes (representing the tonic, sulxlominant, and dominant chords of a particular key) are connected to key nodes. As the analysis of a piece progresses, pitch nodes are activated to the degree that their notes occur in the input. These pitch nodes in tum activate the chord nodes, which in tum activate key nodes. The activation level of each node is also affected by a decay parameter, so the activity of a node dies down if input to that node is not sustained, thus defining a different sort of temporal window. The most active key node at any point in the piece defines the perceived key at that point. This tonality extraction mechanism is implemented as a single KS that is triggered by the appearance of new notes in the window. The tonality KS posts the current best hypotheses about the tonality on the blackboard by creating tonality nodes, with each new input event lead- ing to the creation of a new tonality node when this KS is run. DISCUSSION Our research program addresses basic questions about music perception, but it also illustrates a number of programming techniques that should be of interest to psy- chologists working in many areas, especially in the study of speech perception, reading, and other types of tem- poral pattern perception. First, organizing a simulation in the manner of a multitasking operating system allows the simulation of parallel processing within the context of a highly modular program structure. This permits easy experimentation with a minimum of reprogramming. Sec- ond, the use of linked lists permits a data structure that efficiently stores information about temporal patterns in which the number of potential events far exceeds the num- ber of actual events. In effect, linked lists take the sparse- ness out of a sparse matrix. Third, an input window is essential for simulating processes in which stimulus pat- terns exceed the span of the psychological present, be- cause a program that has access to all the data is not a reasonable psychological model. Finally, the program- ming language must be equal to the task. We chose a powerful, widely known language that supports the com- plex data structures and sophisticated programming tech- niques that are essential in projects of this kind. 262 JONES, MILLER, AND SCARBOROUGH REFERENCES BHARUCHA, J. J. (1987). MUSACT: A connectionist model ofmusi- cal harmony. In Program of the NinthAnnualConference of the Cog- nitive Science Society (pp. 508-517). Hillsdale, NJ: Erlbaum. DEITEL, H. M. (1983). An introduction to operating systems. Reading, MA: Addison-Wesley. Dowuxo, W. J., &; HARWOOD, D. L. (1986). Music cognition. Orlando, FL: Academic Press. GROSSBERG, S. (1980). How does the brain build a cognitive code? Psy- chological Review, 87, I-51. HAYES-RoTH, B. (1985). A blackboard architecture for control. Artifi- cial Intelligence, 26, 251-321. JACKY, I., &; KALET, I. (1987). An object-oriented programming dis- cipline for standardPascal. Communications of theACM,30, 772-776. lONES, I. A., &; HARROW, K. (1986). Problem solving usingTurbo Pas- cal. Englewood Cliffs, NI: Prentice-Hall. KAPLAN, H. (1985). Design decisionsin a Pascal-basedoperant system. Behavior Research Methods, Instruments, & Computers, 17, 307-318. LERDAHL, F., &; JACKENDOFF, R. (1983). A generative theory of tonal music. Cambridge, MA: MIT Press. MCCLELLAND, I., &; RUMELHART, D. (1981). An interactive model of context effects in letter perception: Part I. An account of basic find- ings. Psychological Review, 88, 375-407. NEWELL, A., &; SIMON, H. A. (1972). Human problemsolving. Engle- wood Cliffs, NI: Prentice-Hall. NIl, H. (1986). Blackboardsystems: The blackboard model of problem solving and the evolution of blackboard architectures. Artificial In- telligence Magazine, 7(2), 38-53. POVEL, D.-I. (1984). A theoretical framework for rhythm perception. Psychalogical Research, 45, 315-337. POVEL, D.-I., &; EssENS, P. (1985). Perception of temporal patterns. Music Perception, 2, 411-440. RICH, E. (1983). Artificialintelligence. New York: McGraw-Hill. RUMELHART, D., &; MCCLELLAND, I. (1986). Parallel distributed processing: Explorations in the microstructure of cognition. Cam- bridge, MA: MIT Press. ScHNEIDER, W. (1987). Connectionism: Is it a paradigm shift for psy- chology?Behavior Research Methods, Instruments, & Computers, 19, 73-83. SIMON, H. A. (1968). Perception du pattern musical par AUDITEUR. Sciences de l'Art, V-2, 28-34. TENENBAUM, A., &; AUGENSTEIN, M. (1986). Data structures using Pas- cal. Englewood Cliffs, NI: Prentice-Hall. WINSTON, P. (1984). Artificialintelligence (2nd ed.). Reading, MA: Addison-Wesley. WIRTH, N. (1985). Programming in Modula-Z (3rd ed.). New York: Springer-Verlag. 
work_4y7xmob74fe2rije7yanywi3se	----	DOCUMENT RESUME ED 279 998 CS 008 691 AUTHOR Willson, Victor L. TITLE Methodological Limitations of the Application of Expert Systems Methodology in Reading. PUB DATE Dec 86 NOTE 21p.; Paper presented at the Annual Meeting of the National Reading Conference (36th, Austin, TX, December 2-6, 1986). PUB TYPE Information Analyses (070) -- Speeches/Conference Papers (150) EDRS PRICE MF01/PC01 Plus Postage. DESCRIPTORS Evaluation Problems; *Reading Research; *Research Methodology; *Research Problems; *Research Utilization; *Theory Practice Relationship ABSTRACT Methodological deficiencies inherent in expert-novice reading research make it impossible to draw inferences about curriculum change. First, comparisons of intact groups are often used as a basis for making causal inferences about how observed characteristics affect behaviors. While comparing different groups is not by itself a useless activity, progressing directly to training is premature at best. Second, the think-aloud protocol technique is often used for inferring a subject's cognitive structure of subject matter. This method is inappropriate because it assumes that the organization of this structure resides consciously in a person's mind and can be verbally reproduced. Third, retrospective methods have been employed to infer causality by selecting groups currently differing and discovering differences in their past on putative causal variables, which are then inferred to have caused the present differences. While this technique must be used in historical analyses, it becomes suspect when the inferences are used to speculate on implications for current practice. Finally, techniques employed in naturalistic inquiry often confuse a change in methodology with a change in the discipline being studied, and rely heavily on impressionistic, one-shot observation for many facts. (JD) **********************************************************************x Reproductions supplied by EDRS are the best that can be made from the original document. ********************************************************************** U.S. DEPARTMENT OF EDUCATION Office of Educational Research and Improvement EDUCATIONAL RESOURCES INFORMATION CENTER (ERIC) $ This document has been reproduced as received from the person or organization originating it. C Minor changes have been made to improve reproduction quality. Points of view or opinions stated in this dap- ment do not necessarily represent official OERI position or policy. Methodological limitations of the application of expert systems methodology in reading Victor L. Willson Texas A & M University "PERMISSION TO REPRODIJCE THIS MATERIAL HAS BEEN GRANTED BY Victor L. Willson TO THE EDUCATIONAL RESOURCES INFORMAl ION CENTER (ERIC)." Paper presented at the Annual Meeting of the National Reading Conference, Austin TX, December 1986 Methodological limitations of the application of expert systems methodology in reading The comparison between the expert and inexpert, whether in organizations or in humans, has become a major research technique in education in the last decade. In research on problem solving comparisons have been made between physicists and physics stu- dents. In reading research the basic comparison is between good and poor readers for much of the information-processing theory being advanced. In educational administration successful schools are compared with unsuccessful schools. Creative students are compared with normal students. In all of these areas a similar paradigm is being used: the expert system is compared on a number of attributes with a less or inexpert system. Some attributes are assumed to be causal (such as program differences or strategies employed) and others are assumed to be outcomes, such as achieve- ment or time to solution of a problem. Differences between the expert and inexpert on the causal variables are then assumed to be evidence for causation of the variables, and these salient variables are prcmoted as efficacious for remedying the deficien- cies of the inexpert. This paradigm is an important departure from the dominant experimental paradigm used in education, and this paper presents a critical examination of its methodological limitations. It is an intaresting situation that the expertise methodology has sprung from two quite different areas of research, cognitive psychology and curriculum behaviorism. Cogni- tive psychology has drawn from and has itself influenced artifi- 2 3 cial intelligence (AI) research in machine computing. As Chi, Feltovitch & Glaser (1982) has noted a shift took place in AI research from power strategies to knowledge strategies. The best knowledge strategies available for stuay were found in human beings, so that new computer programs were developed to imitate the way human experts organized and processed information, as could be determined in comparisons with inexpert humans. The effective schools movement, not directly influenced by artificial intelligence in any obvious waY, sought to study the school environment. One outcome of such study, conducted by anthropolo- gists, psychologists, and educational researchers, was that some schools (classrooms, teachers, administrators, etc.) were super- ior in performance to others. Comparisons between variables de- fined as input, or causal,' and output or dependent, led to pre- scriptions for change in inexpert schools to make them more like the best schools being observed. For example Clark & McCarthy (1983) reported on a cohort sequential type design in which volunteer New York City schools implemented a new program based on effective schools literature. Since most of the expert-novice comparisons were presented in the methodological trappings of experimental research (ANOVA statistical analysis and interpretation) there have been only one or two serious evaluations of the causal logic underlying the method and its basis for making causal inferences (Rowan, Bossert Dwyer,1983). The main thesis presented here is that all re- search applying this method suffers from internal validity flaws sufficiently serious to render it uninterpretable. Furthermore, 3 4 expertise method is incapable of supporting causal inference regarding change in the inexpert without true experimental re- search. Techniques employed The techniques being drawn upon in research on expertise include, but are not limited to, the following: comparisons between inact groups; think-aloud protocol; naturalistic inquiry, including ethnographic field method; and retrospective research. There are researchers who are employing experimental research as part of their research strategies, and their applications, specifically exempted from the criticism levelled here, will be mentioned as exemplars of appropriate or adequate research. Comparison between intact groups. This technique is widely used in expertise research. In problem-solving research Chase & Simon (1973) compared the ability of chess masters and novices to chunk board groupings. They found more elements in the chunks of mas- ters than in the novices. Simon & Simon (1978) found differences in the problem-solving behaviors of physicists and physics stu- dents using verbal protocols of the tasks each performed when solving novel problems. The effective schools movement has used comparisons between schools defined as outstanding or excellent and those defined as inferior or deficient to make programmatic decisions about how schools ought to be run. A well-constructed criticism of the effective schools reseat-ch methodology was made by Rowan, Bossert, & Dwyer (1983). Their specific points will be incorporated into this review; these points include difficulty 4 with causal ordering, instrumentation, limitations of generaliza- tion, and nonequivalent control group comparisons. The effective schools research of the 197U's was employed in examining reading at both the school and classroom level. Teacher effectiveness has been particularly emphasized (Rupley, Wise, & Logan, 1986). Brophy's (1973) work on process-product research with primary grade teachers is a widely cited example; later important studies include Medley (1977) and Rosenshine (1978); the latter study raised the problem of little experimental veri- fication for effectiveness research. The Stanford Program on Teaching Effectiveness (Crawford, Gage, Corno, Staybrook, Mitman, Schunk, Stallings, Baskin, Hanvey, Austin & Newman, 1978), and the First Grade Reading Group Study (Anderson, Evertson & Brophy, 1978) are experimental or quasi-experimental studies based on initial observations contrasting good and poor teachers (Rupley et al, 1986). It is important to note that in these studies curriculum recommendations were made after comparative interven- tion was made, not directly on the basis of the original compari- sons. Much of the recent research on reading from a cognitive perspective is based on comparisons between good and poor read- ers. For example, in a recent article by Underwood & Zola (1986) good and poor readers were compared on letter recognition span. In this study no differences were found and no particular instructional inferences were made. In other studies this has not been the case: McGee (1982) compared good and poor fifth grade readers and poor third grade readers, finding differences in 5 recall of text structure ordered from good to poor fifth graders to third graders. McGee concluded that young readers "benefit from following the top-level structure of text to guide reading and remembering passage information..." Even though a disclaimer below this quote suggests the need for more research on efective- ness of instruction, there is a clear message that the observed difference.a are caused by what good readers do, and that poor readers will be helped by some strategy based on the good read- ers' processes. While the study is itself limited because of the text reading level (third grade), it is part of a chain of re- search related to automaticity (Laberge & Samuels, 1974) which is itself based in part on these same good-poor reader differences. There is simply no basis for assuming that the poor readers can be made to perform like the good readers or that their processing will oecome automatic, or if automatic in the same way that the good readers' process is automatic. Another such study is due to Sannomiya (1984) in which poor third grade comprehenders were compared with good sixth grade comprehenders on text comprehen- sion under auditory or visual conditions. In this study both age and ability are confounded. Again, there is no evidence that the poor comprehenders can be made to look like good sixth graders, or that different modes of presentation will change reading performance in this direction. Intact groups are also used as the basis for inferring developmental change. For example, Baldwin & Coady (1978) com- pared fifth graders and college students on their use of punctua- tion as clues to meaning in isolated sentences. They found dif- ferences between the groups and inferred developmental differ- 6 ences in use of punctuation as clues to meaning. It is common to see studies that mix age and reading ability. Juel (1983) com- pared grade two, grade five, and upper division undergraduates; good and poor readers were identified for the elementary groups. In this study the word adult '4Eis used interchangeably with the college sample, the implication being that these readers are a norm for adult performance. This assumption is most definitely wrong, and an assumption that the elementary students are likely to or can become like these adults is unwarranted. Juel never- theless suggests that presenting children with practice words with similar letter combinations would help to develop versatil- ity in decoding. That may be true, but the comparisons made in her study do not support such conclusions. There are many other child-adult comparisons in recent literature in which the adults are high ability college students (McGee, 1982; Schwartz, 1980; Taylor, 1980), or secondary students (Fletcher, Satz, & Scholes, 1981). The use of intact groups has been repeatedly criticized in the educational research methodology literature from Campbell & Stanley (1963) onward with respect to the inference of causality for observed characteristics affecting behaviors. In the case of good versus poor readers, the inference is that what good readers do, poor readers can do, and that instruction directly oriented toward the discrepancy will remediate deficiencies in the poor readers. The good readers are the experts, and the poor readers the novices.' The critical assumption is that the good readers were themselves in the poor readers' state at some point. Often, 7 since the two groups are age matched, this is not true. The good readers were never like the poor readers. Consequently, the inference that the observed differences in condition can lead to training is by itself without basis. Similarly, developmental studies are susceptible to the same difficulty, especially when they involve elementary, secondary, and college populations. In the Baldwin & Coady (1978) study a comparison between fifth grade and college students is meaningless, because differences may be due to selection: if one were able to select the fifth graders who will eventually qo to college, would we still see the dif- ferences in use of punctuation clues? Even if we did find the differences, how cnmfortable would we be in ignoring any other differences that remain between the prospective college-bound fifth graders and the college students. Any variables upon which tne two groups differ become possible alternative causal vari- ables, and training in the absense of experimental demonstration is merely yuesswork. A similar problem exists for comparisons with secondary students when dropout rate becomes appreciable (after grade 10 ), or when students begin self-selecting into courses (grade 9). Differential maturation and history are other threats from the Campbell & Stanley list which are relevant. Finally, regression threats due to selection of extremes are not only omnipresent in good-poor comparisons, their effects should always be estimated statistically just to provide a comparison with the observed differences. Comparing different groups is not by itself a useless activity, but progressing directly to training is premature at best. Differences between good and poor readers, or between 8 developmentally different groups of readers, is useful for supplying clues or hints for more careFul investigation. The tendency to assume a causal shortcut, permitting the ignoring of experimentation, is unfortunate; while the technique may prove correct in a few instances, our experience in educational research with intact groups is lengthy enough to predict many erroneous conclusions and wasted resourc,_2s if the method is allowed to predominate. Similarly, research on developmental differences has largely opted for cross-sectional designs, not wishing to do the hard research implicit in true longitudinal study of development. In the good reader-poor reader research this is particularly telling, for we have little data on long term development of either group from a cognitive, information processing theoretic perspective. 'his is the causal ordering problem that Rowan et al (1983) pointed out; cross-sectional designs that substitute for longitudinal designs almost always have this difficulty. Think-aloud protocol. This technique has tieen used in the study of expert and novice organization of knowledge and was eloquently and favorably defended by Ericsson & Simon (1980) as a valid means to record information that humans are attendlng to in short-term memory. It was attacked by Phillips (1983) as an inappropriate technique to infer human's cognitive structure of subject matter. The core of Phillips' argument is that the external organization imposed in the learning required for a task may require a person to reproduce it verbally, but there is no 9 10 evidence that that organization resides internally in the per- son's mind. Similarly, the content and organization of a ques- tion, with perhaps the exception of free response, imposes an organization on the su'Jject's response that does not necessarily mirror the internal representation of the response. The use of think aloud method, while it occurs in a variety of research contexts, is a major technique in naturalistic or ethnographic research. A recent study by Nicholson (1984) in which 3600 minutes of inerviewing with junior high students was conducted is an example in point. This study will be examined in more detail below, but interview techniques in the comparison of experts and novices are likely to suffer from many difficulties. In reading it is particularly problematical because the researchers usually share the same culture (reading, education, etc.) as both the experts and the novices. This is usually a drawback for ethnogra- phers, who are attempting to view the culture with fresh eyes. in Nicholson's work the experts were teachers, and the similarity between researcher and expert was far greater than between re- searcher and novices (teenagers). The commonness of a shared language of educationese is quite troublesome for a researher in such conditions and the trustworthiness of such interviewing must be questioned; it is not that interviewing cannot be done well, it is that arEat care must be taken to support the evidence presented in such a context. Retrospective studies. In research on creativity the compari- son between creative and noncreative individuals has led to the formulation of programs to teach creativity (Van Tassel-Baska, 11 1986). Also, researchers on creativity have employed retrospec- tive methods to examine prior differences between more and less creative individuals and then to propose ch-nges in education which are expected to engender the same effects in young students as were observed in the creative adults. Segal, Busse & Mansfield (1981) compared retrospectively two groups of biologists, highly cited and nonhighly cited, using self-report survey technique. They found post-doctoral productivity to be related to pre doctoral productivity and high school science interest. As noted by example aove this research technique is used to infer causality by selecting groups currently differing and dis- covering differences in their past en putative causal variables, which are then inferred to have caused the present differences. This technique is apparently not used search, for a search over the last study, by Castagna (1982) in which a influential persons in western history and autobiographies. The implication changed these people and that some was very much in reading re- ten years found only one historical examination of was made using biographies is that decisive reading purposeful, some was not. Of course, historical analyses must use such methods; it is only if an implication for current practice is made that the analysis becomes suspect. Naturalistic inquiry. This body of techniques, attempting to become a method in educational research ,in Kaplan's (1964) sense, draws upon ethnographic research from cultural anthropology, but then leaves it in a philosophical sense. Recent 11 1 2 apologies by Harste (undated) and Weaver (198b) liken the use of naturalisatic inquiry to a paradigm shift, citing Kuhn's now dated and largely refuted work (1963). While this debate more properly belongs in a different critical paper, the use of the techniques in the expert-novice studies requires a small aside. The appeal to a paradigm shift has been misunderstood and mislaid to boot. The shift occurred in psychology in the late 1960s and is often tied to Neisser's (1967) resurrection of internal mental representational constructs, the shift being away from behaviorism. This paradigm shift has flowed into educational research rapidly and convincingly, predating the widespread interest in ethnographic techniques by a decade. The latter interest, it is presumed, was an outgrowth of the real paradigm shift. Paradigm shifts occur in disciplines W.en the prevailing theories are overturned by new, revolutionary ones, that nevertheless account for the facts and relationships previously learned. In paradigm shifts the old is not discarded, it is reinterpreted. There is no such change occurring in reading, notwithstanding the wishful thinking of Harste (undated). The mistake is in confusing a change in methodology with a change in the discipline. Methodologies cannot and never will drive disciplines to the extent that the naturalistic inquirers maintain that they do; recent arguments by Kuhn (1976) himself have backpedalled on the theory-ladenness argument of data. Cooke & Campbell (1979) attack the emphasis by philosophers of science on the Preeminence of theory, relegating,facts to an unwarranted secondary status. That is, facts ar'e observed by resrearchers working from different methodological perspectives. They must reconcile them; their methodologies become more suspect than the facts, which are interobserver confirmable. If the facts are not confirmable, then they cannot be admitted. This latter issue becomes the main problem for the naturalistic researchers, for they rely heavily on impressionistic, one-shot observation for many facts. Many researchers using this method deny intersubjective confirmability, but they abandon science for art. They are not wrong, they merely inquire in another domain. A number of naturalistic studies in reading have been published in the last few years. The study by Nicholson (1984) is the primary study I have encountered which purported to compare experts and novices. The study actually examined the structures of teenagers' understanding of' classroom material; teachers were apparently ignored, although there is an appeal to teachers as experts at athe end of the study. A small section on low achievers was also tacked on. The catchy title was misleading or there was a serious editing problem because there was no comparaison between experts and novices in this study. If there had been it would have told us nothing about how to change stu- dents' conceptions. This is a common problem in naturalistic studies. One gets whatever one happens to find in the setting. If there is nothing very interesting going on little of use will be brought out. Also, naturalistic studies are limited by what passes for actual practice, not by what is possible. It is quite possible that most of what will occur in education in the next century is being tested in laboratory schools, industrial set- tings, and nontraditional educational locations. The public 13 schools are likely to be the last places to find out about these changes, whether through experimental or noturalistic means. Naturalistic research on expert-novice differences in read- ing is limited by selection, ie. the choice of locations; hj history, the context of the location; by instrumentation, espec- ially changes in the observer/interviewer; and by te.voral limi- tations in when the study is conducted and for how long. It is not argued here that naturalistic inquiry is less appropriate than the quasi-experimental research described earlier. Neither is likely to be able to draw valid conclusions regarding curriculum change in the absense of careful experimental manipulation of variables. Summary This paper has sought to draw attention to methodological deficiencies inherent in expert-novice research with respect to drawing inferences about curriculum change. Much credit must be given to the reading research community for generally not leaping to conclusions from such literature, in comparison with some fields of psychology, engineering, and.science education. While some reading studies seem to overreach their conclusions, far more have used the observed differences to probe experimentally hypotheses generated by the observations. This approach cannot be faulted, even if one cannot resist challenging the original premise: that good readers can tell us anything about how poor readers ought to proceed. The methodological threats to internal validity of such research ventures should make us pause to consider if good-poor or expert-novice comparisons are really of 14 15 value: history, selection, instrumentation, maturation, and regression. While no study necessarily is damned due to possible internal invalidity thrats, the weight of methodological argument certainly should make us pause. Ex post facto methods, such as meta analysis, can never rectify the poor initial choice of field of explorat4ion. If we want to see how poor readers can be made into good readers we ought to find examples, or better yet, create examples, and then work to find out what is replicable. That is good science and good research. 15 16 References Anderson, L.M., Evertson, C.M. & Brophy, J.E. (1978). The first grade reading group study: Technical report of experimen- tal effects and process-outcome relationships, Report No. 4071. Austin:Research and Development Center for Teacher Education, University of Texas. Baldwin,R.S. & Coady, J.M. (1978). Psycholingusitic approaches to a theory of punctuation. Journal of Reading Behavior,10, 363-375. Brophy, J.E. (1973). Stability of teacher effectiveness. American Educational Research Journa1,10,245-252. Campbell,'D.T. & Stanley, J.C. (1963).Experimental and guasi- experimental designs for research on teaching. Chicago: Rand-McNally. Castagna, E. (1982).Caught in the act: the decisive readinl of some notable men and women and its influence on their actions and attitudes.Metuchen,NJ: The Scarecrow Press. Chase W.G. & Simon, H. A. (1973). Perception in chess. Cognitive Psychology,4, 55-37. Chi, M.T.H., Feltovitch, P.J. & Glaser, R. (1981). Categorization and representation of physics problems by experts and nov- ices.Cognitive Science,5,121-152. Clark, T.A. & McCarthy, D.P. (1983). School improvement in New York City: The evolution of a project. Educational Re- searcher 12,17-24. 16 Crawford, J., Gage, N., Corno, L., Staybrook, N., Mitman, A., Schunk, D., Stallings, J., Baskin, E., Hanvey, P., Austin, D., & Newman,. R. An experiment on teacher effectiveness and parent assisted instruction in the third gradet. Stanford, CA: Center for Educational Research at Stanford, Stanford University. Cook, T.D. & Campbell, D.T. (1979). Quasi-experimentation: _Design and analysis issues for field settings. Chicago:Rand- McNally. Ericsson, K.A. & Simon, H.A. (18U). Verbal reports as data. Review,87,215-251. Fletcher,J.M., Satz, P., & Scholes,R.J. (1981). Developmen- tal chulges in the linguistic performance correlates of reading achievement. Brain and Language,13,78-90. Harste, J.C. (undated). Portrait of'a new paradigm: Reading comprehension research. In A.Crismore (Ed.), Land- scapes:, A state-of-the-art assessment of reading compre- hension research 1974-1984. Final report, USDE-C-300- 83-0130. Bloomington,IN:Indiana University Language Edu- cation Departments. Joel, C. (1983). The development and use of mediated word identification. Reading Research Quarterly,18, 306-327. Kuhn, T.S. (1962). The structure Of scientific revolutions. Chicago: University of Chicago Press. 17 Laberge, D. & Samuels, S. (1974). Toward a theory of automatic information processing in reading. Cognitive Psychology,6, 293-323. McGee, L.M. (1982). Awareness of text structure: Effects on children's recall of expository text. Readina Research Quarterly,17,581-59U. Medley, D.M. (1977). Teacher competency and teacher effective- ness: A review of process:product research. Washington, D.C.: Aamerican Association of Colleges of Teacher Educa- tion. Neisser, U. (1967). Coanitive psychology. New York: Appleton- Century-Crofts. Nicholson,T. (1984). Experts and novices: A study of reading in the high school classroom. Reading Research Quarterly,19. 426-451. Phillips, D.C. (1983). On describing a student's cognitive struc- ture. Educational 12sysi., 59-74. Rowan, B., Bossert, S.T. & Dwyer, D.C. (1983). Research on effective schools: A cautionary note.Educational Researcher, 5,24-31. Rosenshine, B.V. (1978). Instructional principles in direct in- struction. Paper presented at the annual meeting of the American Educational Research Association, Toronto, March, 1978. 18 1 9 Rupley, W.H., Wise, B.S., & Logan, J.W. (186). Research in effective teaching: An overview of its development. In J.V. Hoffman (Ed.), Effective teaching of readino: research and practice. Newark,Del: International Reading Association. Sannomiya, M. (1984). Modality effects on text processing as a function of ability to comprehend. Perceptual_ and Motor Skills 58 379-382. Schwartz, R.M. (1980). Levels of processing: The strategic de- mands of reading cmprehension. Reading Research Quar- terly,15,433-450. Segal, S.M., Busse, T.V. & Mansfield, R.S. (1980). The relationship of scientific creativity in the biological sciences to predoctoral accomplishments and experience. American Educational Research Journal 17,491-502. Simon, D.P. & Simon, H.A. (1978). Individual differences in solving physics problems. In R.Siegler (Ed.), Children's thinking: What develops?. Hillsdale,NJ: Lawrence Erlbaum. Taylor, B.M. (1980).Children's memory for expository text after reading. Reading Research Quarterly,15,399-411. Underwood, N.R. & Zola, D. (1986). The span of letter recognition of good and poor readers. Reading Research Quarterly,21,6- 19. Van Tassel-Baska, J. (1986). Effective curriculum and instructional models for talented students. Gifted child Quarterly,30, 164-169. 19 2 0 Weaver, C. (1986). Parallels between new paradigms in science and in reading and literary theories: An essay review. Research in the teaching of English,19,298-316. 20 21 
work_4yughouq2bh3zcgpjvvfv7yykq	----	Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=tijr20 IETE Journal of Research ISSN: 0377-2063 (Print) 0974-780X (Online) Journal homepage: http://www.tandfonline.com/loi/tijr20 An Expert System for LD Steel Making P N Mishra, S K Kak & S C Srivastava To cite this article: P N Mishra, S K Kak & S C Srivastava (2001) An Expert System for LD Steel Making, IETE Journal of Research, 47:1-2, 85-89, DOI: 10.1080/03772063.2001.11416207 To link to this article: https://doi.org/10.1080/03772063.2001.11416207 Published online: 26 Mar 2015. Submit your article to this journal Article views: 7 View related articles http://www.tandfonline.com/action/journalInformation?journalCode=tijr20 http://www.tandfonline.com/loi/tijr20 http://www.tandfonline.com/action/showCitFormats?doi=10.1080/03772063.2001.11416207 https://doi.org/10.1080/03772063.2001.11416207 http://www.tandfonline.com/action/authorSubmission?journalCode=tijr20&show=instructions http://www.tandfonline.com/action/authorSubmission?journalCode=tijr20&show=instructions http://www.tandfonline.com/doi/mlt/10.1080/03772063.2001.11416207 http://www.tandfonline.com/doi/mlt/10.1080/03772063.2001.11416207 JETE Journal of Research Vol47, Nos 1&2. January-April 2001, pp 85-89 An Expert System for LD Steel Making P N MISHRA Instrumentation & Electronics, National Metallurgical Laboratory, Jamshedpur 831 007, India. S KKAK Electronics Engineering Department, Institute of Technology, Banaras Hindu University, Varanasi 221 005, India. AND S C SRIVASTAVA ANC Division, National Metallurgical Laboratory, Jamshedpur 831 007, India. LD process of stealmaking accounts for 70% of steel production throughout the world. Due to various constraints like the complexity of LD process; variation in quantity, quality and grade of input materials; varying operating conditions, random parameters and their values; the interventions of an expert and skilled operator is necessary to tackle the complex situation. To overcome this, there exists, a need of a system, which does not only possess the process control capabilities, but also emulates operator's expertise in terms of his knowledge and skill. Such a system has been developed named "Expert Steelmaker". The paper discusses the "Expert Steelmaker" which can presently be used in an advisory mode by the LD steelmaking operator. Indexing terms: Expert system, LD steelmaking, Process control, Charge calculations, Materials and energy balance. L D process of steelmaking accounts for 70% of steel production throughout the. world. Due to various constraints like the complexity ofLD process; variation in quantity, quality & grade of input materials; varying operating conditions, random parameters & their values; the interventions of an expert and skilled operator is necessary to tackle the complex situation. To overcome this, there exists a need of a system which does not only posses the process control capabilities as discussed else where [I] but also emulates operator's expertise in terms of his knowledge and skill. We have developed such a system, named "Expert Steelmaker". No expert system tools have been used for building our "Expert Steel maker" but, it has been developed using Prolog language only. A few other expert systems for LD steelmaking have been reported in the literature. One such system is LD-ES which was built using an expert system support tool named FACOMESHELL [2]. The ESHELL runs on UTILISP, a dialect of LISP. It uses fuzzy logic for some of its knowledge representation. Another expert system combined with a mathematical m,pdel for BOF process control has been reported by T Yoshida et al [3]. This system applies a nile based reasoning technique to improve the output of presently available mathematical models and uses fuzzy reasoning to estimate in furnace reactions. Further, using an expert system, a one man operation of blowing has been established [4]. Paper No 164-A; Copyright© 2001 by the JETE. Since engineering problem solving is not entirely symbolic and sophisticated numerical computation need to be performed to determine physical system parameters, a knowledge based system in that problem solving environment needs to be integrated with computational procedures for successful application. Sunderam et al [5] has proposed organisation of numeric computation using a task structure vocabulary and demonstrated its potential as a viable tool for integration of numerical methods with knowledge based systems. However, our prolog based ES does the computation as well as symbolic processing. The paper discusses the "Expert Steelmaker'' which can presently be used in an advisory mode by the LD steelmaking operator. TOWARDS BUILDING AN EXPERT COMPUTER SYSTEM The most important step in building such a system is knowledge acquisition followed by its representation for machine use and Interpretation. After building knowledge base, inference engine and user Interface are created so that inferences are drawn and made available to the outside user. The first step in acquiring the process knowledge is the identification of an expert or group of experts. It has been pointed out by Mittal and Dyn [6] that multiple expert consultation is beneficial than basing the system on the knowledge of a single expert; "further the expert should be 86 IETE JOURNAL OF RESEARCH, Vol 47, Nos I & 2, 2001 always a practtcmg one. We also arrived at the same conclusion independently and have, therefore, chosen multiple experts for the codification of knowledge to be used by our system. These experts have been working in industry, in the area of LD steelmaking for more than a decade. Also, these persons are considered as experts by their colleagues. Various expert operators of leading steel plants of India have been interviewed in various phases to acquire the operational knowledge; these, however; showed area of commonality and some regions of conflict in their eX;Jertise which had then to be resolved by crosschecking and observing the process. The knowledge thus acquired had to be suitability represented with the help of a high .level computer language. The evolution of certain high level languages for Artificial Intelligence has resulted in standard languages like LISP, ADA, PROLOG etc. IBM PC compatible turbo-Prolog has been used by us for the development of our expert system. It has been earlier demonstrated by Veronica Dahl [7] that logic programm- ing can be used as a representation of knowledge for both database knowledge as well as procedural knowledge. In the development process of this expert system we have used the knowledge available in the references [8,9] extensively in addition to the knowledge elicited from experts from iron and steel industry. Modular programm- ing has been used for realizing the expert system. An important feature introduced is a timer module that runs in background so as to remind the operator about different actions to be taken during a particular heat cycle. The aim and scope of "Expert Steelmaker" The aim of designing this Expert Steelmaker was to develop an adviser to an operator for complete heat cycle of an LD furnace. It has to achieve the following: (a) To calculate the amount of materials needed to make a desired quality and quantity of steel. These materials are hot metal, steelscrap, burnt lime. burnt dolomite, oxygen etc. (b) To optimize the process under various detined con- straints and (c) To guide the operator through the complete heat cycle specifying actions to be taken in response to feedback given interactively by the operators. Our approach to achieve these aims is through knowledge based reasoning, dynamic process modeling and interactive operator's response evaluation. Rules have been formulated on the basis of knowledge acquired from experts and existing literature on LD steelmaking. The declarative and procedural knowledge thus obtained have both been represented in Prolog. Blow control strategies for LD process control have widely been used throughout the world, where in, blowing control is dependent on the operator's observation of the process. If the heal becomes too wild and sloppy, appropriate action is taken on the lance height or oxygen blowing rate. Sometime materials additions are also done. We have adopted a similar strategy in our expert system. Various methods for automatic blowing control of BOF using computers have been employed since 1960. These control methods are classified into two types: (a) Static control in which the necessary amount of oxy- gen and that of the coolant are fixed before blowing by statistical or theoretical analysis of data on mass bal- ance and heat balance. (b) Semi dynamic control in which the carbon content in the molten steel is estimated and controlled to the aim value by analysis and information derived from the exhaust gas. However, the controllability of these methods is limited and not always satisfactory in achieving the composition and temperature at the end of blowing. A combination of static control and dy- namic control have been proposed earlier [I 0, 11). Optimal control for LD converter process have also been reported [12]. Even then operator is a necessary element in controlling the process. Our system with static control coupled with blow control strategy in the form of knowledge base (KB) evolves into an optimal dynamic control for the LD process of steelmaking. KNOWLEDGE BASED CHARGE CALCULATION The knowledge based charge calculation module is the heart of the Expert Steel maker. It accepts input parameters in the following syntactical format using standard notation of elements and compounds to be provided by the operator of the LD furnace. Composition of hot metal: AI %, C %, Fe%, Mn %, P %, S%, Si %, other %, Composition of stee·l scrap AI %, C %, Fe%, Mn %, P %, S%, Si%, other %. Composition of gray iron: AI %, C %, Fe%, Mn%, P %, S %, Si %, other %. Composition of silicon carbide: SiC %, Si02 %, A 120 3 %, other %. Composition of calcium carbide: CaC2 %, CaS %, CaO% other%. Composition of iron ore: CaO %, FeO %,Fe20 3 %, MgO %, MnO %, P20 5 %, Si02 %, Al 20 3 %, other%. P N MISRA et al : EXPERT SYSTEM FOR LD STEEL MAKING 87 Composition of ferro-silicon: SiC %, Fe %, other%. Composition of burnt lime: CaO %, MgO %, Si02, Fe20 3 %, A 120 3 %, other %, LOI %. Composition of burnt dolomite: CaO %, MgO %, Si02%, Fe20 3 %, A 120 3 %, other %, LOI. Composition of steel to be made C %, 0 %. Tonnage oxygen 0 2 %, by volume, N 2 plus Ar %, by volume. Element distribution (% of weight charged): % of C oxidized to CO, % of C oxidized to C02. % of total Mn to steel, % of total Mn to slag. % of total P to steel, % of total P to slag. % of total S to steel, % of total S to slag. % of total Fe oxidized to Fe20 2 fume. Composition of slag (weight ratios) - Ca0/Si02 , MgO/ Si02, Fe 20iSi02. The output information provided by the charge modules are: weight of hot metal, weight of steelscrap (or any two combination of mate- rials like hotmetal, steelscrap, gray iron, ferrosilicon, silicon carbide and calcium carbide), weight of burnt lime to be added, weight of burnt dolomite to be added, and weight of tonnage oxygen to be blown for a ton of steel of desired composition at the desired temperature. This module works in conjunction with the knowledge base such that if the desired composition of steel to be made is known in terms of grade of steel to be made, it automatically assumes the composition of steel to be made. Additionally, KB has a range of values for all input informations. Whenever, the operator exceeds the limit of the range for an input, while inputting the informations, the system prompts him to recheck. If the operator confirms, KB will update the range of the corresponding input information for future use, giving it learning capability. Otherwise it understands it as a mistake and asks the operator to re-enter the correct value. If the operator is ignorant about the particular input informa· tion's value, the knowledge module assume the value based on its latest knowledge of that information. The charge calculation module has a grade sub module which stores all the composition range for different grades of steel, which are most often produced by a ·particular steel plant. However, this sub module can be updated for any new type of grade Of steel to be produced by dynamically interacting with the knowledge module. GUIDANCE AND EXPLANATION FOR THE OPERATOR The expert system (ES) guides the operator of LD furnace through the complete process cycle after outputting the quantity of materials required for the desired heat. It is an interactive process, in the sense that the expert system advises the action to be taken by the operator, waits for his response and then goes to the next action. After estimated time of oxygen blow, it tells that heat is ready to be tapped. Basic explanation capabilities like answers to what and why type of questions have also been incorporated into the system to make it user friendly. An example is given below: Operator ES Operator ES What (softblow) Softblow means keep lance height at more than 2.5 m Why (softblow) Softblow because dc-dt is between 48-60. RESULTS AND DISCUSSION To validate and test the charge calculation module, material and energy balance for a few typical cases have been undertaken using data samples from ref [8]. The results are in agreement with the corresponding material· and energy values [8] and are given in Table I. Charge calculation also seem to be reasonable though they have not been practically verified. The response of ES to various input conditions have been evaluated and found satisfactory. CONCLUSION An advisory expert system has been developed for LD process steel making which undertakes charge calculation intelligently and guides interactively the operator of LD process for a complete cycle till the heat is ready to be tapped. The system does not use any ES shell and is implemented using Prolog language which has been used for representing declarative and procedural knowledge: timing and mathematical calculation. The system should be quite useful for the Steel Industry. Similar technique may be used in other advisory mode process control application where material and energy balance are important and operator interaction with the process is a necessity. By providing suitable hardware and software interface, the system can be used for on line process control application also. 88 IETE JOURNAL OF RESEARCH, Vol47, Nos I & 2, 2001 TABLE 1 Material balance and heat effect at selected temperatures for six cases Case- I Case-IV Values as per Values as per Values as per Values as per Ref. (8) our expert Ref. (8) our expert system system (kg) (kg) (kg) (kg) INPUT ITEMS INPUT ITEMS Hot metal 1000 1000 Steel Crap 1000 1000 Burnt lime 63 63 Burnt lime 0 0 Burnt dolomite 20 19.6 Burnt dolomite 0 0 99.5% Oxygen 86 85.6 99.5% Oxygen 3 3.36 TOTAL INPUT 1169 1168.2 TOTAL INPUT 1003 1003.36 OUTPUT ITEMS OUTPUT ITEMS Steel 912 912.1 Steel 981 981 Slag 141 140.49 Slag 13 13.3 Gas & Fume 116 115.56 Gas & Fume 9 8.69 TOTAL OUTPUT 1169 1168.15 TOTAL OUTPUT 1003 1002.99 Heat effect -131 MCal -132 MCal Heat effect 315 MCal 315 MCal Case- II Case-Y (kg) (kg) (kg) (kg) INPUT ITEMS INPUT ITEMS Gray Iron Scrap 1000 1000 Iron Ore 1000 1.000 Burnt lime 158 157.5 Burnt lime 149 149.1 Burnt dolomite 49 49 Burnt dolomite 33 33.38 99.5% Oxygen 92 91.92 99.5% Oxygen -265 -265.2 TOTAL INPUT 1299 1298.42 TOTAL INPUT 917 917.28 OUTPUT ITEMS OUTPUT ITEMS Steel 877 876.6 Steel 599 599.3 Slag 340 339.5 Slag 313 312.9 Gas & Fume 82 82.1 Gas & Fume 5 4.8 TOTAL OUTPUT 1299 1298.2 TOTAL OUTPUT 917 917 Heat effect 122 MCal 120.5 MCal Heat effect 148 MCal 148 MCal Case-III Case-VI (kg) (kg) (kg) (kg) INPUT ITEMS INPUT ITEMS Calcium. carbide 1000 1000 Silicon carbide 1000 1000 Burnt lime -937 -936.5 Burnt lime 5298 5297.9 Burnt dolomite 12 12.1 Burnt dolomite 1648 1648.5 99.5% Oxygen 655 656.19 99.5% Oxygen 1900 1900 TOTAL INPUT 730 731.79 TOTAL INPUT 9846 9846.4 OUTPUT ITEMS OUTPUT ITEMS Steel 14 14 Steel -2081 -2081 Slag -32 -31.6 Slag 11070 11072 Gas & Fume 748 748.3 Gas & Fume 857 857 TOTAL OUTPUT 730 730.7 TOTAL OUTPUT 9846 9848 Heat effect -232 MCal -230 MCal Heat effect -242 MCal -380 MCal P N MISRA et al : EXPERT SYSTEM FOR LD STEEL MAKING 89 REFERENCES I. S N Singh, P N Mishra & L N Das, PC based process con- trol instrumentation, CSIO Communication, India, vol 16, nos 1-4, pp 91-101, 1989. 2. Michitaka Kanemoto et al, An application of expert sys- tem to LD converter process, ISJJ International, vol 30, no 2, pp 128-135, 1990. 3. T Yoshida et al, A hybrid expert system combined with a mathematical model for BOF process control, Proc of IFAC workshop, Espoo, Finland, 26-28 August, 1991, (Oxford UK Pergamon, 1992) pp 51-56. 4. T Hatamaka et al, Development of blowing control sys- tem in BOF applying AI technol6gy, Proc of IECON, 91, Kobe. Japan. 28 October-November 1991. 5. A Sunderam, J K McDowell & M C Howley, Task struc- ture: a vocabulary for integrating numerical methods and knowledge based systems, Math Comput Simul (Nether- lands). vol 36, nos 4-6, pp 337-346, October 1994. 6. Mittal S & Dyn C L, Knowledge acquisition from multi- ple experts, AI Magazine, vol 6, pp 32-36, Summer, 1985. 7. Veronica Dahl, Logic programming as representation of Knowledge, IEEE Computer, vol 16, no I 0, pp I 06-111. October 1983. 8. R D Pehlke. W F Porter, R F Urban & J M Gaines (Ed). BOF steelmaking, (vol 4 Operation). The Iron and Steel Society of the American Institute of Mining, Metallurgi- cal, Petroleum Engineers, 1977. 9. Turbo prolog, Owners handbook 1.1, Borland Interna- tional, USA, 1986. 10. T Yamazaki et al, Development of the automatic blowing control system, Proc The International Oxygen Steelmaking Congress, Linz Austria, May 25-29, pp 509- 502, 1987. II. L Coutinho Neto et al, Static control at CSN's Steelmaking shop, Proc The International . Oxygen Steelmaking Congress, Linz, Austria, pp 569-575, May 25-29,. 1987. 12. Shigeru Hirohama et al, Application of optimal control theory to LD converter process control, Nippon Steel Technical Report No 27, October 1985. AUTHOR P N Mishra did his MSc (Tech) in Instrumentation from BITS, Pilani. After working as Instrumentation Engineer at BASL, Patratu (1976), Physicist at Research & Control Laboratory, Durgapur Steel Plant, SAIL (1977-1981), and Electronics Engineer at Tata Steel, Jamshedpur (1981-1988), he joined as Scientist at .National Metalurgical Laboratory, Jamshedpur, where he is presently Head of Instrumentation & Electronics Division. His present research interest is in instrumentation, expert system and XRF. * * * * * 
work_52lkxy2ltjhd3lqw7ty7jqfqce	----	A Prototype Student Advising Expert System Supported with an Object-Oriented Database (IJACSA) International Journal of Advanced Computer Science and Applications, Special Issue on Artificial Intelligence 100 | P a g e www.ijacsa.thesai.org A Prototype Student Advising Expert System Supported with an Object-Oriented Database M. Ayman Al Ahmar Deputy Dean, College of Information Technology Ajman University of Science and Technology (AUST) United Arab Emirates (UAE) Abstract— Using intelligent computer systems technology to support the academic advising process offers many advantages over the traditional student advising. The objective of this research is to develop a prototype student advising expert system that assists the students of Information Systems (IS) major in selecting their courses for each semester towards the academic degree. The system can also be used by academic advisors in their academic planning for students. The expert system is capable of advising students using prescriptive advising model and developmental advising model. The system is supported with an object-oriented database and provides a friendly graphical user interface. Academic advising cases tested using the system showed high matching (93%) between the automated advising provided by the expert system and the advising performed by human advisors. This proves that the developed prototype expert system is successful and promising. Keywords- academic advising; expert system; object-oriented database. I. INTRODUCTION Student academic advising is an essential task in educational institutions. Traditionally a university student plans the courses semester-by-semester towards a degree through lengthy meetings with the human academic advisor. Advising meetings are usually held during the beginning of each academic semester. Since student advising is a time-consuming effort, there is a need for computerization of some parts of the advising process. Utilizing a computerized advising system, students can save the software consultation results and can then meet with the human advisor for further consultation (if there is still a need for the traditional face-to-face meeting). This hopefully will save valuable time for academic advisors and for students. The objective of this research is to develop a prototype rule based Expert System (ES) for the academic advising of the students of the Information Systems (IS) Department, College of Information Technology, Ajman University of Science and Technology (AUST), UAE. The system is called "IS-Advisor" and helps students in course selection for each academic semester. Literature reveals many endeavors in the field of automating academic advising activities including the application of expert systems [1-9]. There can never be a 'global' expert student advising system applicable to all academic institutions and departments because of the existence of academic regulations and expert advising knowledge and reasoning specific to each academic unit. As an example to illustrate this concept, AUST regulations allow each student to register from three to six courses per semester. However, the accumulated advising experience and grade statistics related to the IS department show that students who can be advised to register six courses without difficulty are students with AGPA 3.00 or above (out of 4.00), whereas students with AGPA greater than 2.00 and less than 2.25 are better advised to register 4 courses only the next semester to give them a chance to increase their AGPA. The proposed ES (IS-Advisor) represents such specific advising knowledge and reasoning as rules in its knowledge base component and reasoning strategies in its inference engine component. Thus, the ES developed in this research is unique in its specific knowledge base and reasoning strategies and is intended to be of great help to the department of IS. Another particular feature of IS-Advisor is the Object-Oriented (OO) architecture of its database as will be addressed in subsequent sections. II. MODELS OF ACADEMIC ADVISING From the literature we select two models of academic advising adopted in the proposed expert system: Prescriptive advising model and developmental advising model. The prescriptive advising model is characterized by an advisor- student relationship in which students follow the prescriptive procedure of their advisors without assuming responsibility for decision making [10]. The developmental advising models rely on a shared responsibility between the student and the advisor in which the advisor directs the student to proper resources [11]. Literature studies show findings that support both models. For example, Fielstein L. in the research paper titled "Developmental versus prescriptive advising: Must it be one or the other?" stated that: "…intuitive students appeared to endorse the developmental approach to advising. On the other hand, the more 'thinking' students did not value a collaborative relationship and seemed more content with the criteria associated with prescriptive advising" [12]. In general advisors need to look at each student as an individual with individual characteristics. (IJACSA) International Journal of Advanced Computer Science and Applications, Special Issue on Artificial Intelligence 101 | P a g e www.ijacsa.thesai.org III. EXPERT SYSTEMS "Expert Systems (ES) are programs that attempt to emulate the behavior of human experts, usually confined to a specific field" [13]. Regarding the domain of academic advising, ES technology seems to be the most successful method of computerization because the dialogue between human advisor and the student can be conveniently emulated by the dialogue between the ES and the student, and the reasoning of the academic advisor can be successfully automated by the reasoning power of ES; particularly the rule-based ES. A rule based ES captures human knowledge using If-Then rules in a rule-based knowledge base. Academic advising process can be successfully modeled in computers as a rule-based expert system since most advising regulations are based on academic 'rules' such as "if you pass course A, then you can register course B" and so on. The proposed system in this research (IS- Advisor) is modeled as an ES with an object-oriented database, thus the main components of IS-Advisor are: The OO database, the rule-based knowledge base, the inference engine, and the user interface. The ES can explain its results by tracing the If- Then rules used to reach the conclusions through a component called the explanation subsystem. The following sections explain the details of the developed ES by explaining its components. The system is developed using Kappa-PC expert system shell [14]. Kappa-PC supports OO modeling which is adopted in this system since OO database allows each student and each course to be modeled as a single object. The ES knowledge was compiled from the university and college regulations, the long term experience of the author as an academic advisor and the Deputy Dean of the College of IT, and discussing the knowledge with students and advisors for their feedback. IV. THE DEVELOPED EXPERT SYSTEM A. The OO Database (OODB) An important objective in database design is to develop an efficient database structure so that data can be stored, accessed, and modified easily. Much of the work in creating an effective database is in the modeling. It is the application domain that determines how the database should be modeled in order to be successful. The nature of university subjects' and students' records (the domain of this research) reveals that the OO model is the most appropriate database modeling method. OO structure allows each course and each student to be constructed as a different object, and the database modeled as a collection of these objects. This structure gives more flexibility to each object to have whatever features (i.e. attributes or fields) required to identify it while maintaining the integrity of the whole system. The database of IS-Advisor consists of the main classes: Courses and Students. Fig. 1 presents a portion of the object hierarchy of IS-Advisor which is the Kappa-PC's graphical representation of the OO database structure. Each study plan course in the database includes the following data: Title, ID, plan semester number (1 to 8), number of pre- requisite courses, List of pre-requisite courses (if any), pre- requisite hours (Some courses have a specified number of hours as their pre-requisite), type of course (There are three types of courses: Compulsory courses, major elective courses, and university elective courses), keywords describing course contents (e.g. mathematics, programming, algorithm, management, marketing, etc.; these keywords are used to assist students in selecting courses based on their preferences as will be addressed later), course components (theory, lab, and/or tutorial), and course status (offered or not offered; note that fall -or odd- semester courses are offered in fall semester and spring -or even- semester courses are offered in spring semester). Each student object includes the following fields: ID, name, AGPA, passed compulsory courses, passed major elective courses, passed university elective courses, course grades semester-by-semester, earned credit hours, allowable courses, registered courses, course keyword preferences, and load preferences. Note that some data listed above are known and saved in the database (example: offered courses in a particular semester or AGPA of a student) and some data are inferred by the ES (example: lists of allowable and registered courses of a student). It is important to note that the proposed ES is intended to be used for course selection only, and based on courses selected by all students the timing of lectures will be determined manually by the timetabling committee in order to prevent the time conflict between courses. Thus the ES's recommended courses for students will be used as the input for the college timetabling committee. Therefore course timing is not a factor in the current version of the system and a component to automate the determination of lecture timings can be added to the system as a future work. Figure 1. The Object hierarchy of the OO databse of IS-Advisor. B. The Rule-Based Knowledge Base (RBKB) The rules of the rule base can be classified into two categories: Academic rules and student-preference rules. Academic rules are rules that are concerned with academic regulation like pre-requisites, the minimum and maximum number of courses that can be registered by a student (usually: minimum 3 courses and maximum 6 courses), etc. As an example of this rule category, consider the following rules written in English: Rule1: If: The student passed Programming I AND Programming II is offered Then: Add Programming II to the student's allowable courses list. Rule2: (IJACSA) International Journal of Advanced Computer Science and Applications, Special Issue on Artificial Intelligence 102 | P a g e www.ijacsa.thesai.org If: The student's passed hours are greater than or equal to 45AND Computer Ethics is offered Then: Add Computer Ethics to the student's allowable courses list. Rule3: If: The student passed Computer Applications AND The student's passed hours are greater than or equal to 40 AND Computer Networks I is offered Then: Add Computer Networks I to the student's allowable courses list. Rule4: If: The number of courses in the student's recommended courses list is less than 3 Then: Show the message: Students should register minimum 3 courses and maximum 6 courses. If your case is an exception, please contact your academic advisor. Student-preference rules are If-Then rules related to preferences input by the student like preferred courses and preferred number of courses that the student is willing to register in a particular semester. As an example of this rule category, consider the following rule: Rule5: If: The student's course preference keyword is Management Then: Mark all allowable courses having Management as a course keyword. C. The Inference Engine (IE) Kappa-PC ES shell supports both rule-based reasoning (forward- and backward-chaining) as well as the micro- managing of the reasoning using classical programming techniques (particularly list processing). IS-Advisor's inference engine uses both If-Then rules processing and list processing techniques. The overall reasoning procedure of the IS-Advisor is unique and different than other academic advising expert systems available in literature since it is based on the accumulated academic advising knowledge within the IS department at AUST. There are three main steps performed in the process of determining the recommended courses for a particular IS student. In Step 1 all courses that are offered and can be registered by the student are stored in a list called Allowable Courses. Step 2 performs the ranking process for the courses contained in Allowable Courses list. The courses are ranked in a descending order as following: (1) Courses that are pre- requisite for subsequent courses (have the highest priority), (2) Courses matching student preferences (in case preferences are given), (3) Courses officially in the current student's registration semester (fall or spring) according to the study plan, (4) Courses whose pre-requisites were passed in the previous semester (in order not to leave a long time gap between a course and its pre-requisite), and (5) Remaining 'equal' allowable courses (if any) are displayed to the user in order to rank them as preferred. The list resulted from this step is called Ordered Allowable Courses. Step 3 is the filtering step that generates the ordered list of Recommended Courses based on the contents of the list Ordered Allowable Courses. This step follows one of the two advising models: Perspective advising (option 'One-Step Advising' in Fig. 4) or developmental advising (option 'Student's Preferences' in Fig. 4). In 'One-Step Advising' option the list of Recommended Courses is generated as following: (a) Students with AGPA greater than or equal to 3.00 are given the courses ranked from 1 to 6 (from the Ordered Allowable Courses list). (b) Students with AGPA greater than 2.24 and less than 3.00 are given the courses ranked from 1 to 5. (c) Students with AGPA greater than or equal to 2.00 and less than 2.25 are given the courses ranked from 1 to 4. Note that if the remaining number of courses for a student towards graduation is less than the number of courses that can be suggested by the system, then the students is recommended to take the remaining courses only. In "Students' Preferences" option the student is asked to select the number of courses he/she is willing to register (3 to 6 courses) and course keyword preferences. Consequently the list Recommended Courses is prepared as explained in 'One- Step Advising' option above however here level 2 of ranking (courses matching student's preferences) is activated and the number of courses is equal to the number of courses selected by the student (if possible). In addition, more system messages are given here during the user-system interaction in order to guide the student to consider a 'more' suitable course selection. D. The User Interface and Sample Consultation Interactions between the users and the system are supported through a friendly graphical user interface running under Windows environment. Fig. 2 shows the main screen of the system where various options are displayed. The Button "Offered Courses" presents all currently offered courses and the button "Study Plan Structure" displays a semester-by- semester structure plan of the IS program. The user can enter the user manual and get more help on using the system by selecting the "Help" button, or exit the system by clicking "Exit". The main option here is "Academic Advising" from which the user is directed to a screen asking for the student ID number (Fig. 3). After a welcoming message and displaying the currently available student transcript data, the student is given two options as shown in Fig. 4. These two options work as explained while discussing the inference engine above. Fig. 5 shows the result of selecting the option One-Step Advising for a particular student whose AGPA is between 2.25 and 2.99. Note that the Recommended Courses list generated here is ranked by priority from 1 (highest priority) to 5 (lowest priority). The student can click here on the option "Explain!" and get the explanation screen shown in Fig. 6. This explanation screen (related to this 'consultation 1' example) gives the academic reasons for suggesting these courses and for ranking them this way. (IJACSA) International Journal of Advanced Computer Science and Applications, Special Issue on Artificial Intelligence 103 | P a g e www.ijacsa.thesai.org Figure 2. The main screen of IS-Advisor. Figure 3. Entering student ID number. Figure 4. Academic advising models. Figure 5. Recommended courses by the system. Figure 6. A Sample Explanation Screen (consultation 1). As a second consultation (consultation 2) example for a student with AGPA greater than 2.99, the option 'Student's Preferences" on the screen of Fig. 4 results in various query screens as shown in Fig. 7, and Fig. 8. Figure 7. Specifying the number of courses. Figure 8. Selecting course keyword preferences. The result of this consultation is displayed in Fig. 9 below. Figure 9. Recommended courses by the system. (IJACSA) International Journal of Advanced Computer Science and Applications, Special Issue on Artificial Intelligence 104 | P a g e www.ijacsa.thesai.org The options Explain! And More Allowable Courses give the screens of Fig. 10 and Fig. 11 respectivily. Note that for this particular sample advising consultation, the student's prefered number of courses is 4 (see Fig. 7 and Fig. 10). Since the student's AGPA is above 2.99 in this example, the student can register up to 6 courses, this fact is given to the student in Fig. 10 and clicking on 'More Allowable Courses' will suggest to the student the list of allowable courses (Fig. 11). In case the student agrees to increase the number of courses, he/she can increase his prefered number of courses in the screen of Fig. 7. Figure 10. A Sample Explanation Screen (consultation 2). Figure 11. A Sample List of More Allowable Courses. V. SYSTEM TESTING The system was tested by comparing its results with randomly selected actual (i.e. human advisor guided) student registration files for three academic semesters. The results of comparing 130 registration processes show that 93% of the actual cases matched with the system's recommendations. This number shows that the prototype system is successful and going in the correct direction. The reasons for the 7% unmatched results include:  Warned students: Warned students are students with AGPA less than 2.00 and the regulations require them to repeat some low grade courses to increase their AGPA. Such category of students is outside the scope of the current version of the prototype system. (Planned to be modeled in a modified version).  Exceptional cases: The current version of IS-Advisor follows strictly the registration rules; however, there are few exceptions which are performed under some special conditions in human-guided advising situations and not counted for in IS-Advisor. To give an example: Some students (because of official medical reports) request to register less than 3 courses in a semester. Such exceptional cases cannot be handled by the current version of the system, and such students are directed by the system to contact their academic advisors (see Fig. 7). VI. CONCLUSION AND FUTURE WORK In this research a prototype expert system with an object- oriented database for student academic advising has been designed and developed. By implementing prescriptive advising model and developmental advising model, the system provides the students and advisors with a useful tool for quick and easy course selection and evaluation of various alternatives. The system has a graphical user interface and simple menus; information is displayed in a way that is familiar for both advisors and students. The present state of the system was discussed and illustrated with sample consultations. The system is successful and efficient. System testing revealed that 93% of academic advising test cases show an agreement between the system advising in course selection and human advising. Enriching the system by adding more data and knowledge rules is a continuous process. Many parts of the system can be improved further and some issues deserve future work, among them: 1. Currently the system operates as a stand-alone system. It would be better to connect IS-Advisor with the university's student information system. This will automate the process of importing students' data. 2. Advising of students with exceptional cases and warned students is outside the scope of the current system. This feature can be added in future developments of the system. 3. The system can be improved so that it automates the determination of lecture timings for courses based on the courses recommended for all students so that time conflict between lectures is prevented. REFERENCES [1] T. Feghali, I. Zbib, and S. Hallal, "A web-based decision support tool for academic advising", Educational Technology & Society, Vol. 14, No. 1, pp. 82–94, 2011. [2] A. N. Nambiar and A. K. Dutta, "Expert system for student advising using JESS", International Conference on Educational and Information Technology (ICEIT), China, September 17-19, 2010. [3] F. Albalooshi and S. Shatnawi, "HE-Advisor: A multidisciplinary web- based higher education advisory system", Global Journal of Computer Science & Technology, Vol. 10, No. 7, pp. 37 -49, September 2010. [4] R. Zucker, "ViCurriAS: A curriculum visualization tool for faculty, advisors, and students", Journal of Computing Sciences in Colleges, Vol. 25, No. 2, pp. 138-145, December 2009. [5] D. Pokrajac and M. Rasamny, "Interactive virtual expert system for advising (InVEStA)", 36th Annual ASEE/IEEE Frontiers in Education Conference, San Diego, CA, USA, October 27-31, 2006. [6] K. Kowalski, "On-line advising with JavaScript rule-based system", Proceedings of Society for Information Technology & Teacher Education International Conference, Chesapeake, VA, USA, pp. 2922- 2927, 2004. [7] R. M. Siegfried, A. M. Wittenstein, and T. Sharma, "An automated advising system for course selection and scheduling", Journal of Computing Sciences in Colleges, Vol. 18, No. 3, pp. 17-25, February 2003. [8] A. M. Wittenstein and T. Sharma, "FROSH2: An expert system for freshman advisement", Proceeding of the National Conference on Undergraduate Research (NCUR), University of Wisconsin, Whitewater, Wisconsin, USA, April 25-27, 2002. (IJACSA) International Journal of Advanced Computer Science and Applications, Special Issue on Artificial Intelligence 105 | P a g e www.ijacsa.thesai.org [9] M. Patankar, "A rule-based expert system approach to academic advising", Innovations in Education and Training International, Vol. 35, No. 1, pp. 49-58, February 1998. [10] B. B. Crookston, "A developmental view of academic advising as teaching", Journal of College Student Personnel, Vol. 13, No. 1, pp. 12- 17, 1972. [11] C. M. Chando, "Predicting advising style preference from student characteristics", Doctoral dissertation, University of Memphis, UAS, 1997. [12] L. L. Fielstein, "Developmental versus prescriptive advising: Must it be one or the other?", The National Academic Advising Association (NACADA) Journal, Vol. 14, No. 2, pp. 76-79, 1994. [13] R. J. Schalkoff, "Intelligent systems: Principles, paradigms and pragmatics", Jones & Bartlett Publishers, 2009. [14] Intellicorp, Kappa-PC 2.4 ES shell manuals, Intellicorp, Inc., USA, 1997. AUTHORS PROFILE Dr. Ayman Al Ahmar is Assistant Professor and the Deputy Dean of the College of Information Technology, Ajman University of Science and Technology, UAE. He received his B.Sc. (1994), M.Sc. (1997), and Ph.D. (2001) degrees from Middle East Technical University (METU), Ankara, Turkey. His current research interests include Artificial Intelligence, Software Engineering, and Engineering Information Systems. He is a member of IEEE and IEEE Computer Society. 
work_52lz7we3xfbxronrfdabeljndu	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 221022341 Params is empty 221022341 exception Params is empty 2021/04/06-03:06:12 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221022341 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:06:12 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_53aiurken5bufdu5lhgrn6m7uy	----	Microsoft Word - 9 rad - vol 2 no 4.doc International Journal for Quality research UDK- 005 Original Scientific Paper (1.01) Vol.2, No. 4, 2008 309 Aleksandar Vujovic 1) Zdravko Krivokapic2) 1) Aleksandar Vujovic – Center for Quality - Faculty of Mechanical Engineering Podgorica 2) Zdravko Krivokapic – Center for Quality – Faculty of Mechanical Engineering Podgorica Defining Preventive Action For Improvement Business Process Performances By Using Expert System Abstract:This paper present own approach for improvement business process performances through improvement in area of quality management system by using expert system. In that way it is defining own approach of analogy between quality management system and human body in part of unwilling function. It was used a global approach for looking for arise of nonconformity or anomalies in quality management system. On identified places it was defined preventive action for improvement. Like the support it was realized an expert system which in the exit, depend of input data, give set of measurement for improvement. Key words: Expert system, improvement, preventive action 1. INTRODUCTION There are numerous studies that deal with researches of benefits and disadvantages in systems with implemented quality management system. Premises on insignificance of the system of quality regarding improvement of performance are based on allegations that by that system, procedures are over-emphasized through excessive care of implementation or non- coverage by procedures, and real quality is neglected [1,2,3]. However, most researches point to real benefits of ISO 9001 implementation, contrary to those who claim that the price of implementation and maintenance of QMS is bigger than profits realized by it [4,5]. There are negative premises in literature related with TQM model regarding influence on organizational performances, as it is the case with ISO 9001 model too. Such premises point to its inapplicability, and therefore, in this paper and in idea of association of ISO 9001 and the model of business excellence in direction of improvement, comparison with performances of organizations that are winners of the award for excellence as a measure of level of TQM implementation, was pointless. Therefore author has chosen to point here to pessimistic attitudes and to promote optimistic premises through review and analysis of literary sources related to that subject. [6,7] Include premises that TQM has no efficiency regarding organizational performances. This premises are accompanied by researchers that indicate how it is very hard or almost impossible to establish a relation between TQM and organizational values and believe that such a relation is unreal [8,9]. There are many studies that indicate how TQM model implemented into organizational management is not just effective but also efficient even in terms of financial results of the organization [10,11,12,13,14]. On such defined basis and assertions that present majority in literary sources, and point to positive influence of ISO 9001 and TQM on organizational performances even in the part of finance, an opportunity emerges that improvement in these models would directly contribute to improvement of organizational performances. This is one of directions that this work followed in its realization. Every inconsistency i.e. disagreement with 310 A. Vujović , Z. Krivokapić requests of ISO 9001 model brings results in weakening of quality management system performances and thereby organizational performances. Hereby conditions are created to identify inconsistencies with a term of error and apply the theory of learning based on errors or CBR- Case Based Reasoning in this paper. Through application of this theory or learning based on experiences of others, system is being developed to predict a possibility of error occurrence in identified areas of ISO 9001 and to present measures for preventive action in that direction as to indicate possibilities and places for improvement for organizations. Through established analogy with the human organism and division of its activity to willing and unwilling, an opportunity is created to act upon analogy with e.g. sportsman, and to perform a comparison between middle-class sportsmen (quality management system) and top-class sportsmen (system with the award for business excellence) according to these activities (ISO 9001 standard requests). Thereby it is possible to point to areas that are critical in view of performances and indicate priorities regarding improvements with the aim of achieving top performances. 2. ESTABLISHMENT OF THE ANALOGY BETWEEN THE QUALITY MANAGEMENT SYSTEM AND THE HUMAN ORGANISM In need to establish the analogy between the process modulated organizational structure and the human organism, so as to create the system that is independent from organizational functions and based only on the process model, following division of man functions was made [15,16]: • willing and • unwilling functions. Willing functions1 are those dependent on man’s profession and performed by man’s will. They are variable and dictated by a central control of the organism. For example, when a worker at the construction site lifts his hand, it is not the same as when a referee at the game lifts his hand and etc. Willing functions refer to functions of external motoric organs. Second category is made of unwilling or automatic functions and their use is given by their very existence. Those are functions that are same in all professions and all people (considering that they exist, i.e. that organism is healthy) and do not depend on the man will but are simply executed. For example, those are functions of secreting enzymes, hormones, heartbeats, and similar, that regulate organism functioning, and functions that cannot be controlled [17,18]. With such a ratio of functions in the organism, we can establish the analogy of the system with implemented quality management system. In order to meet requirements of this paper, only analogy in terms of unwilling functions has been considered. Analogy that regards unwilling functions goes in the direction of: • Identification of weak spots, namely organizational performances through acquisition of inconsistencies and isolation of critical areas by Pareto method, • Comparison with top organizational performances in accordance with assessment applied in the competition for the Oscar for quality award, • Analysis and conclusion on capacities and opportunities of particular area of ISO 9001 standard accordance to bussines exellence and top form, • Defining the preventive action depending of on number of nonconformities in area on one hand and degre of readeness on the other hand. 1 Term “functions” is used in medical terminology, although it is equally correct, to use a term “activities” in view of ISO 9000 standard terminology. For reasons of consistent referencing and use of theories from the field of medicine, the author has chosen to use the term functions. Vol.2, No. 4, 2008 311 Analogy established like this justifies the approach to work, which is, just as it is with the man, to monitor process performances in process modulated structure and based on, in medical terms, their condition and diagnostics of non-functionalities, to perform preventive actions in order to maintain the top form. All functions (willing and unwilling) exist in the organism i.e. they are carried out, of course on condition it is a healthy organism, better said organism that has all organs performing their given functions. Analogically with that, systems observed from the standpoint of this paper are certified quality management systems, which meet all requirements i.e. perform all functions provided by ISO 9001 standard. That is of course confirmed by the certificate relating mentioned referential. 3. APPLICATION OF ANALOGY AND DATA ANALYSIS Unique database consisting of 1009 inconsistencies, identified in over 350 organizations was created to meet demands of this paper, and in that way, a basis was found to carry out analysis with the aim of defining critical areas, specificities of an organization and guidelines for improvement in direction of defining preventive measures and decrease of corrective action that would be in the spirit of ISO 9001 standard also. If you consider that there are 500 certificates in Serbia and Montenegro in the part of the most competent certification bodies, then the number of 350 makes 70% out of total number, which indicates significance of the sample for analysis. Inconsistencies represent findings of external revisers who underwent demanding trainings mostly with foreign and competent trainers, which of course increases significance of data for analysis. These inconsistencies were stored in the database classified in relation to requests of standard and activities of the organization by application of realized DSS (Decision Support System) i.e. system for the support of decision-making. One of forms of that system is shown in Figure 1. Figure 1 According with analogy with unwilling function, we provide collecting of all nonconformity and his separation by module of ISO 9001 standard. Like the output on the DSS system is histogram view on nonconformities in decresing mode. On that way it is possible to apply Pareto method and separate critical requirement of ISO 9001 model or area which is critical according to nonconformities arise and that particularly for all module. That critical area is show in Table 1. 312 A. Vujović , Z. Krivokapić Table 1. Module name Critical areas Module 5 5.6 - management review 5.4 - planning Module 6 6.2 - human resource 6.3 - infrastructure Module 7 7.5 - producting and service 7.4 - purchase 7.6 - management by facilities for meusurement and attend Module 8 8.2.2 - internal audit 8.5 - improvement 8.2.1 - satisfaction of user 8.3 - management by nonconformity product On the other hand, by using methodology AHP (Analitic Hierrarhy Process) we are defining significante of particular requirement of standard based on they importance for organisational performance improvement [19]. Rezults are show in Table 2. Table 2. Requirement Kz Requirement Kz 823 0.083 83 0.036 821 0.082 52 0.035 85 0.069 81 0.033 84 0.067 822 0.032 55 0.066 73 0.026 824 0.064 61 0.022 56 0.056 72 0.022 53 0.054 42 0.019 71 0.04 76 0.018 41 0.038 62 0.008 51 0.038 74 0.008 75 0.038 63 0.005 54 0.037 64 0.005 Kz – Coefficient of significante Vol.2, No. 4, 2008 313 Now, based on presented results and on one side critical of area in part of nonconformity arise and on the other side depend of coefficient of significance for attain business excellence, we can talking or judge about “strength” of measurement for improvement which must be defining for that requirement. Therefore, for the higher level of critically and higher level of coefficient of significant, that must be making a greater or stronger preventive action for improvement this area. 4. IMPROVEMENT OF BUSSINES PROCESS PERFORMANCES BY USING EXPERT SYSTEM Like one of tools in area of artificiall inteligence, which could be using for making the knowledge is expert system. Expert szstem is deferent of other system in area of artificiall inteligence because they try to explicitly embody expertise and knowledge by using softvare [21]. Expert system are also marking like one of most comercial area and also according to number of project like one of most using tools in area of artificial inteligence [20, 22]. For example, it is predict that for the first part of 21. century, almost 75% of all document in low will be written by help of expert systems [23]. Or some prediction tell that expert system will be of vital significiant for measurement of quality of product and service [24]. Expert systems is very important and growing area in modern condition of business [25, 26], with special focus on economy in country which is on the high level of development. This research talking about growing trends and significiant and usefulness of making one expert systems. For needs in this papers, like the best solution for developing expert system, we choose the ACQUIRE tool. That is results of analises of all available tools for developing an expert systems [19]. Expert systems which is show in this paper, start from the array of forms which propose a number of answer and through which users input data in to the system (Figure 2). This forms is rotate one after the other dependence of answer and movement through decision tree of expert system Figure 2. 314 A. Vujović , Z. Krivokapić After number of forms, dependence of answer of users, on the exit this system give propose of preventive action for improvement (Figure 3) accordance with approach which was presented in poit 3 of this paper. Figure 3. This expert system give example of preventive action for module 5 and 8 of standard ISO 9001. That is because that module are most critical area. With this paper we would like just to confirm of our idea and approach and not to go in direction of comercial activities. For the defining preventive action according to presented approach, it was use the next sources of knowledge: • knowledge of one expert from area of management of quality and business process improvement, • experiance from eleven high level worl organization in area of quality and business process performances [27], • standard with quideliness for improvement bussines process performances [28], • best experiance from process of external audit of ISO 9001 system [29], • experiance and practice of organisation which was involved in competition for Oscar of Qualiy award [30], • theory and TQM principle which is present in [31], • experiance which is presented in [32] and which are show the way for attain a business excellence. Usefulness of idea for developing this expert system is prove by validation activities in one of most imprortant and prestige organisation in Montenegro in area of quality and business process performances about that Vol.2, No. 4, 2008 315 talking next rows of this paper. 5. ACTIVITY OF VALIDATION EXPERT SYSTEM In part of validation of expert system, it was be provided next activities: 1. cheking of funcionality of expert system in real condition, 2. cheking of expert system on few nonconformity which was find in real condition and analysis of results, 3. testing a potencial client and they satisfaction and during that period also it was loking for they sugesstion. According to: clear defining plan for testing, object in which will plan to provide testing and validation of softver (one of the most successful firm in Montenegro in area of medium service firm) and according to element from standard for evaluation softvare quality ISO/IEC 9126 - 1:2001, and based on [33], it was defining check list for evaluation of tehnical and ergonomical caracteristic of our expert system. Evalution was be on scale 1-10 by the expert from previously mentioned firm. Results is show on Table 3. Table 3. Category Average mark Average mark Total average mark Short comings of presented software 7.9 Benefits of new program solution 8.3 T ec hn ic al ch ar ac te ri st ic Influence this solution for job organisation 8.9 8.4 Global ergonom characteristic 8.7 E rg on om c ha ra ct er is ti c Adaptability of system 8.9 8.8 8.6 Softver is equaly pozitiv evaluated in term of technical like in term of ergonom characteristics. In that term, this solution have short time of reaction, it is compatabil with most using operativ system, it have great inteface with user, it have easy way for input data and good presentation on exit, installation is realy simple, softer is very competative etc. Based on evaluation and work with system in real condition, management in the firm are express sutisfaction with presented softvare and his possibilities and indicate to intention to buy software and be involve in future development of software. 316 A. Vujović , Z. Krivokapić 6. CONCLUSION In literature it is possible to find opposite attitudes concerning the influence of quality management system on bussines process performances. However, the most using attitudes talk about positive influence of quality management system on improvement of business process performances. It could be conclude that any improvement in part of quality management system, lead to improvement in business process performancess. To improvement in quality management system it could be possible to do throught analyses of nonconformities and defining preventive caracter measurements for needs for »ruggedization« of critical places. So far as it will use analogy with human body and his unwilling function then nonconformity could be grouped. Then it could be possible to find critical area for all firms undependently of his type or size. After that, by using, very faintly or almost not using, expert system in area of quality management, throught using methods and techniques for collecting and formalize knowledge, it could be possible to defining preventive action for improvement in area of quality management system and with that also improvement in area of improvement bussines process performances. REFERENCE [1] Seddon J., “In Pursuit of Quality: Case Against ISO 9000” 1997., (Cork: Oak Tree Press). [2] Cuff E., Sharrock W., Francis D., “Perspectives in Sociology”, 1979, (London: Routledge). [3] Wenger E., “Communities of Practice: Learning, Meaning, and Identity”, 1998, (Cambridge: Cambridge University Press). [4] Casadesus M., Gimenez G., Martõ Bronsoms R., “Tipologõas de empresas certifiicadas segun la normativa ISO 9000. Analisis de los resultados de un estudio empõrico”. paper presented at the IX Congreso Nacional de ACEDE, 1999. [5] Schenkel A., «Conceptualizing and Exploring the Organizational Effects of ISO 9000: Insights from the 0resund Bridge Project”, Total quality management, Vol 15, No. 8, 1155-1168, October 2004. [6] Eskildson L., «Improving the odds of TQMs success», Quality Progress, Vol 27, No 4, pp. 61-63, 1994. [7] Harary O., «Ten reason why TQM doesnt work», Management Review, Vol 86, No 1, pp.38-44. [8] Bergquist T., Ramsing K., «Measuring performance after meeting award criteria», Quality Progress, Vol 32, No 9, pp. 66-72, 1999. [9] Przasnyski Z., Tai L., «Stock market to Malcolm Baldrige National Quality Award announcement. Does quality pay?», Total quality Management, Vol 10, No 3, pp.391-200, 1999. [10] Schaffer R., Thompson H., “Successful Change Programs Begin with Results”, Harvard Business Review, September/October 1992, pp 80–89. [11] Opara Emmanuel Uzoma, “The Empirical Test of Total Quality Management: An Application of TQM at Chevron and Its Impact on Productivity”. Quality Management Journal. Vol 4 No 1, p 10, 1996. [12] Cherkasky Stanley M, “Total Quality for a Sustainable Overall Service Performance”, Quality, pp 4–7, 1992. [13] Simmons B., White A., “The Relationship between ISO 9000 and Business Performance: Does Registration Really Matter?” Journal of Management Issues, Fall, Vol 11, No 3, pp 330–343, 2000. [14] Lemark D., Reed R., «Commitment to Total Quality Management. Is there a relationship with firm performance?, Journal of Quality Management, Vol 2., No1, pp.67-86, 1997. [15] www.fakultet.fpzg.hr-neurologija [16] Belak L., Gaćina N., Radić T., «Tehnologija hrane», Visoka škola za turistički menadžment u Šibeniku, Šibenik 2005. Vol.2, No. 4, 2008 317 [17] Vilber K., «Bez granica-popularna psihologija», Babun, Beograd, 2002. [18] www.medicina.hr [19] Aleksandar Vujović, «Poboljšavanje performansi poslovnog sistema na bazi sistema menadžmenta primjenom alata vještačke inteligencije», Doktorska disertacija, Mašinski fakultet u Podgorici, januar, 2008. [20] Hossein Bidgoli, “Modern information system for managers“, Department of Management, California State University, Bekersfield, California, 1997. [21] Kosko B., Kosko B., “Knowledge Based Problem Solving“, Preceentice Hall, Englewood Cliffs, NJ, 1986. [22] Welbank M., “A review of knowledge acquisition techniques for expert systems’’, 2004 [23] Batchelor A., “Productivity + efficiency =profit”, Hong Kong Lawyer Magazine, July, 1995. [24] Godfrey B., “Trends in quality: Godfrey’s top ten list”, Journal of Business Strategy, p. 44, March/April, 1994. [25] Alavi M., “Managing organizational knowledge”, in Zmud, W.R. (Ed.), Framing the Domains of IT Management: Projecting the Future, Chapter 2, Pinnaflex Educational Resources, Cincinnati, OH, 2000. [26] Hansen T., Nohria N., Tierney T., “What’s your strategy for managing knowledge?”, Harvard Business Review, Vol. 77, No. 2, pp. 106 - 16, 1999. [27] Oakland J., “Total Organizational Excellence - Achieving world class performance”, Elsevier Butterworth - Heinemann, Oxford, 2005 [28] JUS ISO 9004:2001 - ”Sistem menadžmenta kvalitetom - Uputstva za poboljšavanje performansi”, Savezni zavod za standardizaciju, Beograd, 2001. [29] Tricker R., “Best Practice - ISO 9001:2000 - The Quality Management Process”, Van Haren Publishing, 2006. [30] FQCE, “Kurs za ocjenjivače po modelu nagrade kvalitet Oskar kvaliteta“, Beograd, 2004. [31] Perović M., “Menadžment - informatika - kvalitet“, CIM Centar, Mašinski fakultet u Kragujevcu, 2003. godina [32] Raković R., “Kvalitetom ka poslovnoj izvrsnosti”, Energoprojekt - Ingraf, Beograd, jun, 2006 [33] www.zpr.fer.hr Received: 03.09.2008 Accepted: 08.10.2008 Open for discussion: 1 Year 
work_57lmsnappjfstksrbmb66iisiu	----	doi:10.1016/j.eswa.2003.12.006 Intelligent agents for supporting construction procurement negotiation Ren-Jye Dzeng*, Yu-Chun Lin Department of Civil Engineering, National Chiao-Tung University, 1001 Ta-Hsieu Rd., Hsinchu 30050, Taiwan, ROC Abstract Negotiation is commonly required to reach a final contractual agreement in construction material procurement. However, even simple negotiations often result in suboptimal agreements, thus ‘leaving money on the table.’ An automated system that could evaluate bids, negotiate to finalize the bid, and value the individual characteristics of negotiating parties would be useful to both contractors and suppliers. This study examines common negotiable issues and options for construction material procurement, and presents an agent-based system, named C-Negotiators, that helps a contractor and suppliers to negotiate via the Internet. Genetic algorithm is used to find the most beneficial agreement for all parties, and web-based development is used to improve negotiation efficiency. Experiments also were conducted and demonstrated that C-Negotiators improved negotiation efficiency by saving negotiation time and cost, and improved negotiation effectiveness by suggesting a better agreement with higher joint payoff. Although the increase in payoff was smaller than expected, the improvement should increase for more complex negotiation problems involving more issues and options, or complicated preferences and for inexperienced negotiators. The application of the system is mainly limited by its symmetric optimization, while procurement negotiations in the construction industry are biased towards the contractor, and also by user comfort with their preferences and negotiations being monitored by the system. q 2003 Elsevier Ltd. All rights reserved. Keywords: Genetic algorithm; Intelligent agent; Procurement negotiation; E-commerce application 1. Introduction Most construction business processes rely heavily on traditional means of communication such as face-to-face meetings and the exchange of paper documents such as technical drawings, specifications, and site instructions. The construction industry has long recognized the need to increase the efficiency of these processes by exchanging large volumes of information quickly and cheaply (Deng, Li, Tam, Shen, & Love, 2001). As international competition continues to intensify, significant numbers of construction organizations are strategically investing heavily in infor- mation technology to gain competitive advantage (Betts, 1999). Various web-based collaboration platforms for project management and procurement have also been developed. One example is the PrimeContract developed by Primavera System Inc (Primavera System, 2003), which attempts to streamline intra-company and inter-company business processes, and supports project team communication, procurement, bid/auction, bid anal- ysis, and contracting. Construction material procurement is a key business where negotiation is commonly required to reach final contractual agreement. However, even simple negotiations often result in suboptimal agreements, thus ‘leaving money on the table’ (Raiffa, 1982). Based on Oliver (1996), while many factors lead negotiators to miss out on gains, falsely assuming fixed pies and the framing of the situation often cause parties to fail to reach mutually beneficial agreements. The challenge of negotiation arises partly from the fact that each side has private information about their own payoff function but is ignorant of the values and strategies of the other side. Exacerbating this situation is the negotiators’ incentive to misrepresent their preferences. For illustration, consider the following scenario: a general contractor has solicited several bid proposals from suppliers registered on a web-based construction procure- ment platform. Besides preferring a cheaper price, the contractor also prefers payment by usance check instead of cash to maintain cash flow level, and prefers delivery of materials in small consignments as required to avoid 0957-4174/$ - see front matter q 2003 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2003.12.006 Expert Systems with Applications 27 (2004) 107–119 www.elsevier.com/locate/eswa * Corresponding author. Tel.: þ886-35733529; fax: þ886-35745074. E-mail address: rjdzeng@mail.nctu.edu.tw (R.-J. Dzeng). http://www.elsevier.com/locate/eswa overloading site storage space. Suppliers with no previous business with a contractor are likely to attempt to show their competitiveness by offering the contractor considerable flexibility during negotiations. For example, one supplier, as a new comer, may be willing to offer large price discounts if procurement can be extended to include other related materials, as well as one-week notice for delivery, and 60-day usance check if future procurement is possible. Currently, such routine business transactions often involve some form of negotiation between the parties. Usually neither side knows the preferences of the other: the prospective contractor does not know the costs of the supplier, and the supplier does not know how much the buyer values each negotiable issue. The cost and time involved in negotiation mean that contractors must limit the number of prospective suppliers they negotiate with, and also the number of options included in negotiations. A cheap and efficient negotiation method would allow the exploration of more prospective suppliers and options. Clearly, an automated system that could evaluate bids, negotiate to finalize the bid, and value the characteristics of the negotiating parties would be useful to both contractors and suppliers. Bazerman (1994) classified negotiations into two cat- egories based on negotiator attitudes: distribution (claiming a share of the pie) and integration (enlarging the available pie). The distribution type of negotiation is a zero-sum game, i.e. a gain for one party is a loss for another. Such negotiations involve each party using the strategy of predicting the bottom line of the other and presenting an offer that maximizes their own benefit. Such negotiations generally result in low satisfaction level. On the other hand, the integration type of negotiations promotes cooperation among negotiators. Because each negotiator has different preferences regarding each negotiable issue and option, the strategy is not to attempt to win on all issues, but to identify those issues that the negotiators care about most and make tradeoffs accordingly. Such negotiations usually achieve a higher satisfaction level. Decision support research has focused on the design and development of tools for aiding negotiators in various domains, such as Genie (Harris, Kraus, Wilkenfield, & Blake, 1991) that stresses model visualization capabilities, NEGOPLAN (Krovi, Graesser, & Pracht, 1999) that generates if-then production rules, and GBML (Matwin et al., 1991) that identifies rules for improving negotiations. Software agents are also applied to facilitate negotiation. Software agents are computer programs with a certain degree of autonomy, which are continuously active and interact with other systems on behalf of users (Bradshaw, 1997). Snadholm and Lesser (1995) found that cooperative agents often exist and perform enterprise tasks such as production planning and meeting scheduling. Competitive agents do not give in during negotiations except in exchange for compensation because they only care about their benefit and do not consider mutual benefit. However, this competitive style prevents the disclosure of individual preferences and often compromises individual benefit. Several electronic commerce web sites, such as Auction Bot (2001), eBay’s Auction Web (2003), Kasbah, 2003, and OnSale (2003), also offer agents that help in on-line price negotiations. For example, Kasbah adopts distribution type of negotiations and allows users to define their own agents with buying strategies (i.e. anxious, cool-head, and frugal), selling strategies (i.e., anxious, cool-head, greedy, and initialization parameters (e.g. asking price, acceptable price, and deadline). T@T (2003), additionally to price, allows both buyer and sellers to customize their agents and negotiate on the issues of warranty, delivery time and method, service plan, return policy, and free bonus. No agents have been developed specifically for negotiating construction procurement between contractor and suppliers. This study examines common negotiable issues and options for construction material procurement, and develops an agent-based system that helps a contractor and suppliers to negotiate via the Internet. Genetic algorithm is used to find the most beneficial agreement for both parties, and web- based development helps improve the negotiation effi- ciency. Experiments were also conducted to assess the effectiveness and efficiency of the system compared with human negotiation. 2. Practice of negotiation on construction procurement The general contractor for a project needs to procure materials and equipment from suppliers to execute the project. The contractor also may need to subcontract part of the work to specialty traders because of considerations of technological specialties, resource availability, profit mar- gin, and risk diversification. After identifying procurement items, tenders from suppliers are invited and evaluated. The evaluation may lead directly to a final decision awarding the contract to the supplier offering the best deal without further negotiation. This situation frequently occurs when space in negotiable issues is limited because negation takes time and man-hours. However, especially for a valuable or complex contract, the procurement of some items still requires further negotiation with several prospective suppliers. Negotiation on issues such as price, terms of payment, and delivery may give the contractor business leverage. In practice, issues to be negotiated are determined at the beginning of negotiations, but new issues sometimes may arise during negotiations. The contractor proposes desired options for the negotiable issues, and the supplier proposes a price based on these options. The proposed price may be continuously lowered during the negotiations. The supplier may also request that terms be modified or include new issues to offset price decreases. The negotiation ends when both parties agree on the options and price. R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119108 3. Negotiable issues A survey and follow-up interview (Dzeng, Ho, & Lin, 2003) were conducted to identify key issues that may arise during construction material procurement negotiations. Common options used for each issue, and the preferences of both contractors and suppliers regarding these options were also studied. The survey was sent out to 90 contractors in Taiwan in 2003, and received 55 responses (response rate ¼ 61.11%), within which 50 responses were valid (usable response rate ¼ 55:6%). Key issues identified included: price, payment term, payment period, advance payment, resource provision, freightage, delivery, and opportunities for extended procurement, mass procurement, and future procurement. These issues may be classified into four categories based on range of options available. The first category is price, for which options lie on a continuous spectrum. The second category includes issues for which a limited number of commonly used options exist. For example, options for payment terms include: ‘cash’, ‘30-day check’, ‘45-day check’, and ‘60-day check’; for payment period options include ‘on delivery’, ‘on completion of mile- stones’, ‘on completion’, ‘monthly’, and ‘bi-weekly’; for advance payment options include ‘10’, ‘15’, ‘20’, ‘25’, and ‘30%’; for freightage options include ‘included’ and ‘excluded’, for delivery options include ‘single delivery’, ‘multiple deliveries’, and ‘on-call delivery’. The third category includes issues whose options are a list of items and quantities. For example, options for resource provision are a list of provided resources and quantities, and options for extended procurement opportu- nities are a list of additional procured items and their quantities. The fourth category includes issues for which options are quantity related. For example, options for mass procurement opportunity are the maximum quantities procurable; and options for future procurement opportunity are possible future procurement quantities. The implied procured item for these issues is the item being negotiated on. 4. Modeling negotiation preferences Negotiating strategies and the reaching of a satisfactory agreement are usually determined subjectively based on the experience and intuition of negotiators. Controlling the behaviors of artificial agents in such a humane way is difficult, and a systematic, quantitative model is necessary. This study uses the weighted payoff function (i.e. utility function) to represent the preferences of human negotiators regarding each option for each negotiable issue, and to model the aggression and concession in the negotiation. Negotiation can be viewed as a process of seeking an agreement point in a multidimensional space. Each dimen- sion corresponds to a negotiable issue, and can be discrete or real valued. Each issue may have several options. Each negotiating party values these options differently, and a multidimensional payoff function exists over the space of possible agreement points. Suppose n negotiable issues exist, where an offer x can be represented using an array one-dimensional matrix ½x1; x2; …; xn�; where xi denotes the chosen option for issue i: The payoff of a negotiator for a particular offer x can be represented as follows. UðxÞ ¼ Xn i¼1 WiUiðxiÞ ð1Þ UðxÞ : total payoff of a negotiator for the chosen set of options x; UiðxiÞ : issue payoff of a negotiator for the chosen option xi for issue i; Wi : weight of issue i for calculating negotiator payoff. Based on this concept, contractor and supplier payoff functions are discussed below for the above four issue categories. 4.1. Category I issues Figs. 1 and 2 illustrate conceptually typical payoff functions of contractors and suppliers for category I issue, which is price. Fig. 1 describes the preference of a typical contractor regarding price. As buyers, contractors have an acceptable price range ½Amin; Amax�; which they consider reasonable and are willing to accept. Additionally, con- tractors also have a desired price range ½Dmin; Dmax�; which falls within the acceptable price range. Starting from Amax (the highest acceptable price), contractor’s payoff increases with decreasing price. The payoff reaches the highest peak when the price reaches Dmax (the highest desired price). Fig. 1. Contractor’s payoff function for price. R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119 109 Further decrease in price has little effect on payoff increase until the price reaches Dmin (the lowest desired price). Initial contractor asking price in negotiations is within the desired prince range, and depends on various conditions such as competition among prospective suppliers and familiarity with the negotiating supplier. Price below Dmin decreases the payoff rather than increasing it because the contractor starts to see the price as unreasonable and thus doubts supplier credibility. The payoff continuously decreases with price until the price reaches Amin (the lowest acceptable price), any price below which is considered unacceptable to the contractor. Fig. 2 describes price preference of typical suppliers. Like a contractor, a supplier also has both an acceptable price range ½A0min; A 0 max� and a desired price range ½D0min; D 0 max�: However, unlike a contractor, supplier payoff increases with increasing price. Excluding the possibility of fraud on the contractor’s side, the supplier may have no apparent upper boundary for the price they are willing to accept ðA0max ¼ 1Þ: Once again, the desired price range falls within the acceptable price range. Fig. 3 shows both contractor and supplier payoff functions, where DD ¼ ½Dmin; D 0 max� is the maximum possible difference between the initial asking price of the contractor and the initial offering price of the supplier; i.e. space of starting points for the negotiation. DA ¼ ½Amin; Amax� represents the space of agreement points in the negotiation. 4.2. Category II issues Fig. 4 illustrates five typical payoff functions for category II issues, including options for mass procurement, payment term, payment period, advance payment, freightage, and delivery. The options for each of these issues can be enumerated and their quantities are limited. Therefore, the respective payoff functions are discrete rather than continu- ous, as for category I. Type I payoff function shows that negotiator payoff positively correlates with issue options. For example, the contractor generally prefers longer payment term, in order to delay the payment as long as possible, and thus contractor’s payoff for ‘60-day check’ is greater than that of ‘cash’. Type II is a variation of type I and shows that the negotiator payoff tends to positively correlate with most issue options, but remains constant for some intermediate options. For example, some contractors may be indifferent between ‘30- day check’ and ‘45-day check’ for payment term. Type III shows that negotiator payoff tends to negatively correlate with issue options. For example, a supplier may prefer shorter payment term, and thus may have a smaller payoff for ‘60-day check’ than for ‘cash’. Similarly, type IV is a variation of type III. For example, some suppliers may feel indifferent between ‘30-day check’ and ‘45-day check’ for payment term. Type V shows that the negotiator is indifferent among issue options. Generally, contractor payoff is of type I or II for issue payment period and delivery; of type III or IV for issue payment term, advance payment, and freightage; and of type V for issue mass procurement opportunity. On the other hand, supplier payoff is of type III or IV for issue payment period, delivery; and mass procurement opportunity; of type I or II for issue payment term, advance payment, and freightage; and of type V for issue. Nevertheless, many factors may affect both contractor and supplier payoff. For example, payment period options include ‘on delivery’, ‘on completion of milestones’, ‘on completion’, ‘monthly’, and ‘bi-weekly’. Normally, a contractor will have a payoff of type III for payment period (i.e. prefer ‘on completion’ to any other options) to delay the payment as long as possible, maintain high level of Fig. 3. Price negotiation space for contractor and suppliers. Fig. 2. Supplier’s payoff function for price. R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119110 reserved cash, and guarantee quality of work received. However, the contractor payoff may be of type IV or V if the payment involved is small or the work duration is short. Similarly, a supplier may have a payoff of type I for payment period (i.e. prefer ‘bi-weekly’ to any other options) to receive payment as early as possible and so maintain higher cash reserves and also reduce the risk of completing the job without receiving payment. When the payment involved is small or the work duration is short, a type II payoff may be typical. Under such a condition, if the application of payment involves lengthy and complicated paper work, a supplier may even choose ‘on completion’ rather than ‘bi-weekly’ (type III or IV). 4.3. Category III and IV issues Category III issues, such as resource provision and procurement extension, require two inputs: a list of items and their quantities. Meanwhile, category IV issues, such as mass procurement opportunity and future procurement opportu- nity, although requiring only quantity as an input, can be treated in the same way as category III because the implied procured item for these issues is that being negotiated on. This study uses estimated value as the unified measure- ment unit for procurement extension and resource pro- vision, and expected value as the unit for mass procurement opportunity and future procurement opportunity, where the unit is a monetary value in both cases. These issues have no individual payoff functions because the options of these issues generally are offered by the contractor as presump- tions, and cannot be changed by a supplier. Nevertheless, the monetary value of these issues may significantly influence supplier payoff for issues such as price. The influence is bigger than for a contractor because the options offered by the contractor normally use the capacity leeway of that contractor rather than squeezing capacity (i.e. offering additional procurement that was originally planned or additional resource capacity that is not used), but represent an opportunity for a supplier to gain extra contract value or achieve cost savings. 5. Modeling negotiation process for agents Krovi et al. (1999) proposed that the negotiation process might be captured by the interaction of three types of environments, namely: task, agent, and communications environments. Such environments of the proposed process model are described below. 5.1. Task environment The task involves artificial agents playing the roles of contractor and suppliers in a virtual supply chain. The buyer is the contractor, and the sellers are the suppliers providing materials to the contractor. All parties try to maximize individual payoff through negotiation. Moreover, all parties have conflicts of interest regarding one or more issues. For example, the contractor prefers low price while the suppliers prefer high price. Notably, different suppliers may prioritize the negotiable issues differently. The negotiation process is sequential for individual suppliers (i.e. making an offer and then waiting for a counter-offer), but may be parallel for the contractor negotiating with multiple suppliers (i.e. simultaneously making offers to multiple suppliers). Agents negotiate by exchanging offers and counter-offers for the negotiable issues. Each negotiation session is free of time constraints. 5.2. Agent environment Each negotiation session between a contractor and a supplier involves three artificial agents, namely the contractor agent, supplier agent, and coordinator agent. Human negotiators control the contractor and supplier agents by setting payoff functions for each negotiable issue. The payoff function of the contractor agent differs from that of the supplier agent, and neither side knows the payoff functions of the others. However, the coordinator agent knows the payoff functions of both sides, and tries to identify an agreement point that meets the satisfaction levels and maximizes the joint payoff of both sides. 5.3. Communication environment The contractor and supplier agents interact by exchan- ging information via the coordinator agent in the form of offers and counter-offers. Offers made by each party are communicated via messages. Each agent incorporates the actions of other agents into its negotiation plan. Messages and actions of the three agents are detailed in a subsequent section. Fig. 4. Typical payoff functions for category II issues. R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119 111 6. Genetic algorithm representation Oliver (1996) automated negotiations on electronic commerce in consumer products using genetic algorithm (GA), and found that the negotiation results equaled, and sometimes outperformed, those achieved by humans. Matwin et al. (1991) suggested that GA is suitable for problems with four characteristics. The first characteristic is that the problem can be described as a sequence of moves. Second, the sequence may converge towards a final state when all states are comparable regarding their contribution to a certain goal. Third, moves that contribute to the achievement of a given result can be identified. Finally, credit for good results can be attributed to contributing moves. Negotiation involves a sequence of offers and counter- offers. The sequence leads to either a final agreement point or a failure state. Each intermediate offer results in a payoff for the contractor and supplier, and is traced. The goal is to maximize the payoff sum, and the payoff of each offer can be calculated after determining the payoff for each negotiable option. Besides the present negotiation problem fitting the above characteristics, GA is also a heuristic search procedure that relies on very little domain knowl- edge. Because no research has been conducted on the negotiation strategies adopted in construction procurement, this study chose GA as a search algorithm for finding the most beneficial agreement point. Each negotiation offer is represented as a gene, so that GA can apply genetic operators such as mutation and crossover to create a population of offers and evolve those offers to find the most beneficial one(s). Gene cell representation comprises two parts: the first cell is a threshold for accepting an offer, and the remaining cells represent the options associated with each negotiable issue. Offer i; a string of cells, can be represented by array ½Ti; Oij�; where Ti denotes the threshold for offer i; and Oij represents a list of options for each negotiable issue j of offer i: Threshold Ti equals the payoff of offer Oij to the offer maker. The payoff of an offer i to a negotiator, representing negotiator satisfaction level with a particular offer, is defined as the weighted average of the payoff for each individual issue, as shown in formula (2). UðiÞ ¼ Xm j¼1 WjUjðOijÞ ð2Þ UjðOijÞ: individual payoff of option Oij on issue j; Wj : weight on issue j; m : number of negotiable issues Since the objective here is to find the offer that is most beneficial to both negotiating parties, joint payoff (the sum of the contractor’s payoff and the supplier’s payoff on an offer) is defined as the objective function. Because negotiating party attempts to maximize their individual payoff, the GA fitness function of the contractor or supplier is the individual payoff function. Additionally, a fitness improvement factor g (formula (3)) (Buckles & Petry, 1992) is used to determine whether the evolution has reached convergence. The evolution stops if the improvement factor of an evolved population is below the user-defined threshold. g ¼ðthe highest fitness score2the average fitness scoreÞ= ðthe average fitness scoreÞ ð3Þ Genetic operators include reproduction, crossover, and mutation. The reproduction operator enables individuals to be duplicated for possible inclusion in the next generation. Crossover in biological terms describes the blending of chromosomes from the parents to produce new chromo- somes in the offspring. Both crossover and mutation operations ignore the threshold cell. Section 7 describes the proposed computation model for negotiation that involves three collaborative agents based on the gene representation, fitness functions, and genetic operators described above. 7. Computational model Fig. 5 presents the flowchart of the proposed compu- tational model. The model involves three agents, namely contractor, coordinator, and supplier agents. Boxes with solid lines denote activities performed by agents, while boxes with dashed lines indicate those performed by humans. To avoid confusion below, italic fonts refer to agents, while normal fonts refer to humans. The contractor must initiate the agent with a set of negotiation criteria (i.e. negotiable issues and their allow- able options, weights and payoff functions) and GA settings (i.e. population size n; crossover rate c; mutation rate m; and threshold for fitness improvement factor g). Information on the negotiation criteria and GA settings is passed to Coordinator, but not to Supplier, except for the negotiable issues and allowable options, which are passed further to Supplier. The supplier must respond to this message by determining acceptable options, weight, and payoff function for each negotiable issue. The proposed model comprises two modes, namely the traditional and automatic modes. In the traditional mode, Contractor generates and passes request-for-quotations (RFQ) to Supplier based on the variety of combinations of allowable options determined by the contractor. Supplier presents each message to the supplier, and the supplier provides an appropriate price quotation for each RFQ. The whole process merely provides book-keeping and payoff calculation functions, and does not involve any intelligent decision-making. In the automatic mode, Coordinator coordinates Contractor and Supplier to reach an offer that R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119112 maximizes the sum of contractor and supplier payoff. The following discussion refers to the automatic mode. Contractor and Supplier have similar objectives; i.e. to generate an offer that is acceptable to its own criteria and has payoff higher than the offer proposed by the counter part through a continuously random selection process, and evolve the offer to find the best one through GA. Using Contractor as an example (Fig. 5), the first step is to generate n offers as its population. Each offer includes an option for each negotiable issue and a threshold T; which equals the corresponding payoff of the option based on the payoff function of the contractor. The threshold represents the satisfaction level of the contractor with the offer. The second step involves randomly selecting an offer from the population and submitting it to Coordinator, who at this time also receives an offer from Supplier. If the offer of Contractor provides Supplier with a payoff higher than that from the Supplier offer, and Supplier offer provides Contractor a payoff higher than the payoff of the Contractor offer, both offers are saved in their respective tentative pools. Otherwise, Coordinator passes the offer back to Contractor and Supplier and asks them to select another offer. This process continues until Coordinator has n Contractor offers and n Supplier offers. The total of 2n offers are used to calculate the fitness evolution improve- ment factor g: If g is smaller than the pre-determined threshold, the search has reached a convergence, and Coordinator presents the offer with the highest sum of payoff of both contractor and supplier as the final result. Otherwise, Coordinator requests both Contractor and Supplier to generate another generation of offer populations. Through generations of evolution, when g is below the threshold, most offers in the population already have fitness scores approaching the best offer. Thus, further evolution achieves insignificant improvement, and so evolution can stop. The activity ‘Perform GA to generate another n offers’ is a typical GA evolution process. The fitness is first calculated for each offer. Second, the tournament selection method is used to select offers for reproduction. In the tournament selection, a number of individuals are picked using roulette wheel selection, and the best of these are chosen for reproduction. When the corresponding two offers are selected, the next step is to apply user-defined parameters Fig. 5. Computation model for negotiation agents. R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119 113 to determine if their offspring should be generated through reproduction, crossover, or mutation. The fitness scores are calculated for the two newly generated descendants, and only that descendent with the higher score is kept. This process continues for n times and produces a new population of n offers. 8. System implementation The interface and reasoning engine of the proposed system, named C-Negotiators, are developed using Micro- soft Visual Basic 6.0. Meanwhile, the database for storing the contractor and supplier information was implemented using Microsoft Access 2000. The interface enables the contractor to choose negotiable issues and allowable options, and define their preferences (payoff) regarding each option (shown in Fig. 6), and GA parameters. It also allows the suppliers to define their preferences regarding each of the allowable options set by the contractor. The engine includes GA engine, three types of intelligent agents, namely Contractor, Coordinator, and Supplier, and message communication between agents. 9. Case study This study conducted experiments to assess the perform- ance of C-Negotiators on the joint payoff sum of suggested negotiation agreements, and achieving benefits in saving negotiation time and costs. First, three procurement items, namely pre-mixed concrete, rebar, and rebar assembly, of two projects A and B were selected for the experiments. Both projects involved an office-plant complex. Project A has five stories plus one underground story, while Project B has five stories plus two underground stories. Table 1 compares the two projects in terms of total floor area, total contract volume (with specific subcontract volume for pre- mixed concrete, rebar, and rebar assembly), and position and years of procurement experience of the participating professionals. Three experiments were conducted to assess the performance of C-Negotiators. Experiment I, termed traditional negotiation, represented a human negotiation process, the negotiation outcome of which was the actual agreement obtained from existing procurement contracts. C- Negotiator was not involved in this experiment. Experiment II, termed traditional Internet negotiation, represented a replayed human negotiation process conducted via the Internet without the involvement of artificial agents. C- Negotiator provided only the interface and book-keeping functionality in this experiment. Experiment III, termed agent-based negotiation, represented the negotiation process conducted via the Internet with full agent involvement. In this experiment, agents applied GA and negotiated via the Internet, and suggested final agreements. The parameter values used for GA were based on the suggestions proposed by DeJong, 1980), and included population size ¼ 50; Fig. 6. Initiation dialog for contractor. R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119114 crossover rate ¼ 0:7; and mutation rate ¼ 0:02: The threshold for the fitness improvement factor was set to 5%. A virtual project was also used to help participants familiarize themselves with the system before the exper- iments began. 9.1. Performance measures Performance measures for the experiments include individual payoffs, joint payoffs, negotiation costs, and negotiation time. Once the contractor and suppliers determined their payoff functions, the same set of functions was used to derive the individual and joint payoffs for all three sets of negotiation outcomes. Negotiation costs include man-hour, communications, and travel costs. The costs incurred during the pre- negotiation (e.g. searching for business opportunities, historic data collection, cost estimation, and document preparation) and post-negotiation phases (e.g. contract preparation and reproduction, and delivery of goods) were assumed to be identical for all experiments and thus were ignored. The following formulas were used to calculate man-hours, communications and travel costs, where C represents cost, Q is quantity, T denotes time, and R represents rate. Cman-hr ¼ ½Qmeeting p Qnegotiator p ðTmeeting þ TtravelÞ þ Tcommunication þ Tsystem� p Rman-hr ð4Þ Ccommunication ¼ Tcommunication p Rcommunication ð5Þ Ctravel ¼ Ctravel p Qmeeting p Qparticipants ð6Þ Negotiation time included communication time ðTcommunicationÞ and negotiation meeting time ðTmeetingÞ: For Experiments II and III, the time further included time for initiating negotiable issues and options, and payoff functions, and time for system execution. Participant contemplation time during the negotiation phase was counted as part of the initiation time. Time spent playing with the virtual project during the pre-experiment phase, and that spent on Q and A conversation between participants and experimenters was ignored. 9.2. Experimental results and discussions Table 2 lists the contractor negotiation outcomes for Experiments I, II, and III, with three suppliers (i.e. pre- mixed concrete, re-bar, and re-bar labor) for both projects. Each set of outcomes includes the agreed unit price and options for the negotiable issues, as well as individual and joint payoffs. The table also lists the percentage increase in joint payoff (shown by percentage numbers under each joint payoff) for Experiments II and III compared to Experiment I. Table 3 lists the estimated values for the variables used in the formulas (4 – 6) and the calculated negotiation cost and time for the contractor and suppliers. The percentage cost and time savings for Experiments II and III in comparison with Experiment I are also listed. The traditional Internet negotiation did not always produce an agreement with higher joint payoffs than the corresponding result in the traditional negotiation. Examples are Experiment II (Project A) for pre-mixed concrete (144.242 vs. 145.958) and rebar labor (145.200 vs. 146.733) as shown in Table 2. This is because C-Negotiator for the traditional Internet negotiation merely provides message communication and book-keeping services, and does not provide any guide on negotiation. However, as shown in Table 3, traditional Internet or agent-based negotiation always require less cost (from 83.8% less to 94.5% less) and time (from 95.3% less to 98% less) than traditional negotiation because communication via the Internet significantly reduces the costs and time required for travel and meetings. Agent-based negotiation always reached an agreement with higher joint payoff (from 1.5% more to 9.8% more) than the other two methods as shown in Table 2. This difference occurred because the human negotiators tried to reach a mutually acceptable agreement based on experience, while the agents tried to maximize the joint payoff through extensive search. Thus, agents are motivated in finding the best agreement. The agent-based negotiation was also less costly (from 84% less to 94.5% less) than the traditional negotiation, but was not always less costly than the traditional Internet negotiation. The agent-based negotiation Table 1 Projects used in experiments Project data Project A Project B Total floor area 23523 M 2 67576 M 2 Total contract volume $US 7.77 millions $US 22.79 millions Pre-mixed concrete $US 1.18 millions $US 2.82 millions Rebar $US 1.23 millions $US 2.99 millions Rebar assembly $US 0.44 millions $US 1.19 millions Participants’ background Contractor Deputy Section Manager (9 yr) Deputy Section Manager (9 yr) Pre-mixed concrete supplier Vice President (20 yr) Senior Manager (18 yr) Rebar supplier Owner (18 yr) President (21 yr) Rebar assembly labor supplier Owner (15 yr) Owner (16 yr) R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119 115 also required less time (from 95.3% less to 98% less) than the other two methods as shown in Table 3. The improvement in joint payoff was smaller than expected. This phenomenon occurred because the number of negotiable issues and options were limited, and human negotiators could reach good agreement based on years of experience. Nevertheless, the experiments also demon- strated that C-Negotiators occasionally might help negotia- tors ‘leave less money on the table’, achieving improvements of as much as 9.8% of payoff, as in the negotiation with re-bar supplier for Project B (Table 2). Fig. 7 illustrates the payoff points for negotiating pre- mixed concrete for Project A, which include the original agreement points (from actual contracts), and the best ten agreements of the first and last GA generations from agent- based negotiation. The distribution of points in Fig. 7 was typical, and resembled the negotiation outcomes of most procured items and projects in the experiments. The original point was above the 458 line, meaning the actual agreement favored the contractor. In other words, the tested procurement items were of an unbalanced market (buyer’s market), and the contractor had more negotiation power. The ten best agreements of the first GA generation might include points close to the original, points better than the original, and occasionally points approaching the best of the last generation. GA found a better point through generations of evolution. However, the small number of negotiable issues and options limited the improvement. Thus, rather than saying that GA improves the optimum through evolution, GA can be said to improve the population average through evolution. Fig. 8 displays a special set of negotiation payoff points. The set resembled the typical one (e.g. Fig. 7) in that the average joint payoff of the first GA generation was higher than the payoff of the actual agreement, and was below the payoff of the last GA generation. However, unlike the typical one, the agreement points of Fig. 8 were distributed around the area below the 458 line, indicating that the supplier had strong negotiating power. A follow-up investigation indicated that this result occurred because the supplier was an owner-designated supplier and the contractor thus had less bargaining power. 10. Application limitation Some limitations to the application of C-Negotiators exist. First, although the experiments demonstrated that the GA-suggested best agreement point produced higher joint payoff than the original agreement made by human, the improvement was smaller than expected, ranging from 1.5% (re-bar labor of Project A) to 9.8% (re-bar of Project B). One reason for the small improvement was that the number of negotiable issues and options of interest to negotiators for the tested procurement items was small, and the human Table 2 Negotiation outcomes of the experiments Item Negotiation outcomes Experiment (Project A) Experiment (Project B) I II III I II III Pre-mixed concrete Agreed options Unit price $57 $56 $58 $47 $46 $43 Payment term 60-day check 45-day check 60-day check 60-day check 45-day check 30-day check Payment period Monthly On-delivery Monthly Monthly Monthly On-delivery Delivery On-call Multiple Multiple On-call On-call On-call Freightage Included Included Included Included Included Included Payoffs Contractor 71.964 62.273 68.000 90.727 86.909 90.273 Supplier 73.994 81.969 79.531 63.600 68.400 67.500 Joint 145.958 144.242 147.531 154.327 155.309 157.773 Improved (%) 21.2 þ2.2 þ0.6 þ2.2 Re-bar Agreed options Unit price $293 $286 $277 $272 $263 $253 Payment term 60-day check 45-day check 30-day check 60-day check 45-day check 30-day check Payment period Monthly Monthly Monthly Monthly Monthly Monthly Delivery Multiple Multiple Multiple Single Multiple Multiple Freightage Included Included Included Included Included Excluded Payoffs Contractor 85.268 88.415 92.545 86.250 88.077 86.394 Supplier 68.603 70.846 72.500 64.357 71.286 79.000 Joint 153.871 159.261 165.045 150.607 159.363 165.394 Improved (%) þ3.5 þ7.3 þ5.8 þ9.8 Re-bar labor Agreed options Unit price $103 $110 $100 $108 $113 $108 Payment term Cash 30-day check 30-day check Cash 30-day check 30-day check Payment period Monthly Monthly Monthly Monthly Monthly On delivery Delivery Multiple On-call On-call Multiple On-call On-call Payoffs Contractor 72.000 78.000 87.000 73.000 82.000 85.500 Supplier 74.733 67.200 62.000 64.000 54.667 54.000 Joint 146.733 145.200 149.000 137.000 136.667 139.500 Improved (%) 21.0 þ1.5 20.2 þ1.8 R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119116 negotiators could reach near-optimum agreement based on years of negotiation experience. However, the improvement should increase with increasing number of negotiable issues or options. Also note that the GA-suggested best agreement point was always closer to 458 than the original because the GA sought to optimize the joint payoff rather than the individual payoff of the contractors or suppliers. Since the construction Table 3 Estimated cost and time for the experiments Estimated data and result Experiment (Project A) Experiment (Project B) I II III I II III Variables Contractor Rman-hr $9.38 $9.38 $9.38 $9.38 $9.38 $9.38 Qnegotiator 3 1 1 3 1 1 Qmeeting 3 1 1 4 1 1 Qoffer 4 7 N/A 4 7 N/A Supplier (concrete) Rman-hr $11.46 $11.46 $11.46 $12.50 $12.50 $12.50 Qnegotiator 2 1 1 3 1 1 Qmeeting 2 1 1 3 1 1 Qoffer 4 5 N/A 4 6 N/A Supplier (rebar) Rman2hr $13.02 $13.02 $13.02 $13.89 $13.89 $13.89 Qnegotiator 3 1 1 3 1 1 Qmeeting 3 1 1 4 1 1 Qoffer 3 5 N/A 3 6 N/A Supplier (labor) Rman-hr $14.58 $14.58 $14.58 $14.58 $14.58 $14.58 Qnegotiator 3 1 1 2 1 1 Qmeeting 3 1 1 3 1 1 Qoffer 3 4 N/A 3 5 N/A Cost (US$) Contractor $692 $66 $66 $861 $66 $66 Improved (%) þ90.5 þ90.5 þ92.3 þ92.3 Supplier (concrete) $470 $76 $75 $1038 $76 $79 Improved (%) þ83.8 þ84.0 þ92.7 þ92.4 Supplier (rebar) $1,100 $81 $81 $1499 $83 $83 Improved (%) þ92.6 þ92.6 þ94.5 þ94.5 Supplier (labor) $1,189 $85 $81 $793 $84 $83 Improved (%) þ92.9 þ93.2 þ89.4 þ89.5 Time (min) Contractor 1620 40 33 1620 40 33 Improved (%) þ97.5 þ98.0 þ97.5 þ98.0 Supplier (concrete) 1200 52 46 1200 48 56 Improved (%) þ95.7% þ96.2% þ96.0% þ95.3% Supplier (rebar) 1620 56 47 1620 55 57 Improved (%) þ96.5% þ97.1% þ96.6% þ96.5% Supplier (labor) 1620 55 38 1200 46 45 Improved (%) 96.6 þ97.7 96.2 96.3 Fig. 7. Typical negotiation points (for pre-mixed concrete of project B). R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119 117 procurement market is a buyer’s market, the contractor might not accept the GA-suggested best offer if the contractor payoff for the offer was below that of the original offer. The remedy to this problem presented here was to choose the GA-suggested 10 best offers, and present only those offers with a contractor payoff equal to or higher than that of the original. Negotiators found acceptable agree- ments from the presented ones in most cases. When none of the presented agreements was acceptable, the negotiators might adjust some terms to reflect their concerns. Another anticipated application limitation did not occur in the experiments. Users who were familiar with C-Negotia- tors and knew that the system made suggestions based on the payoff functions might attempt to manipulate the system by misrepresenting their preferences. Understanding that the system suggestion is merely a suggestion that helps reach agreement and does not necessarily become the agreement may discourage such manipulation behavior. Furthermore, some conservative users may be uncomfortable with the system knowing their preferences and observing their offers. Such distrust may discourage users from using C-Negotiators or other similar systems. Nevertheless, C-Negotiators is multidimensional instead of single dimensional, unlike the difficulty encountered in practice when applying game-theoretic models of bargaining. Moreover, C-Negotiators does not require common knowledge of negotiating parties. For example, a negotiating agent in C-Negotiators does not directly require any explicit knowledge of its counter-part agent such as payoff functions. 11. Conclusion Effective intelligent agents not only allow a contractor and suppliers to negotiate more efficiently through efficient exchange, but also offer a better agreement that benefits both parties more than one they could have reached. This work examines the negotiable issues and options for construction procurement. Contractor and supplier prefer- ences regarding the options also were compared. An agent- based system, C-Negotiators, also was developed to help the negotiation between contractor and supplier using the genetic algorithm. Experiments were also conducted and demonstrated that C-Negotiators improved negotiation efficiency by reducing negotiation time and cost, and improved effectiveness by suggesting a better agreement with higher joint payoff. Although the increased payoff was smaller than expected, we believe that the improvement will increase for more complex negotiation problems involving more issues and options, or complicated preferences and for inexperienced negotiators. System application is limited mainly by its symmetric (uni- directional) optimization, while the procurement nego- tiation in construction is biased towards the contractor, and also by level of user comfort with system monitoring of their preferences and negotiations. Acknowledgements The authors would like to thank the National Science Council, Taiwan, for financially supporting this work under Contract No. NSC-91-2211-E009-059. References Auction Bot (2001), http://auction.eecs.umich.edu/, last updated, 2001. Bazerman, M. H. (1994). Judgment in managerial decision making. New York: Wiley. Fig. 8. Special negotiation points (for pre-mixed concrete of project A). R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119118 http://auction.eecs.umich.edu/ Betts, M. (1999). Strategic management of IT in construction. Oxford: Blackwell. Buckles, B. P., & Petry, F. E. (1992). Genetic algorithms. Los Alamitos, CA: IEEE Computer Society Press. Bradshaw, J. M. (1997). Software agents. Menlo Park, CA: AAAI Press. Deng, Z. M., Li, H., Tam, C. M., Shen, Q. P., & Love, P. E. D. (2001). An application of internet-based project management system. Automation in Construction, 10, 239 – 246. DeJong, K. (1980). Adaptive systems design: A genetic approach. IEEE Transactions on Systems, Man and Cybernetics, 10, 566 – 574. Dzeng, R. J., Ho, C. L., Lin, & Y. C (2003). Mobil management on construction procurement and negotiation. Technical Report, Taiwan: National Science Council (in Chinese). eBay (2003). http://www.ebay.com/, last updated, 2003. Harris, M., Kraus, S., Wilkenfield, J., & Blake, E (1991). A decision support system for generalized negotiations. The 13th Annual Conference of the Cognitive Science Society, San Diego, CA, pp. 382 – 387. Kasbah (2003). http://kasbah.media.mit.edu/cgi-bin/KasbahLogin, last updated, 2003. Krovi, R., Graesser, A. C., & Pracht, W. E. (1999). Agent behaviors in virtual negotiation environments. IEEE Transactions on Systems, Man and Cybernetics, 29, 15 – 25. Matwin, S., Szapiro, T., & Haigh, K. (1991). Genetic algorithms approach to a negotiation support system. IEEE Transactions on Systems, Man and Cybernetics, 21(1), 102 – 114. Oliver, J. R. (1996). A machine-learning approach to automated negotiation and prospects for electronic commerce. Journal of Management Information Systems, 13(3), 83 – 112. OnSale (2003), http://www.onsale.com/, last updated, 2003. Primavera System (2003). http://www.primavera.com/products/ primecontract.html, last updated 2003. Raiffa, H. (1982). The art and science of negotiation. Cambridge, MA: Harvard University Press. Snadholm, T., & Lesser V. R (1995). On automated contracting in multi- enterprise manufacturing. Conference on Improving Manufacturing Performance in a Distributed Enterprise: Advanced Systems and Tools, Edinburgh, Scotland, pp. 33 – 42. T@T (2003). http://ecommerce.media.mit.edu/tete-a-tete/, last updated, 2003. R.-J. Dzeng, Y.-C. Lin / Expert Systems with Applications 27 (2004) 107–119 119 http://www.ebay.com/ http://kasbah.media.mit.edu/cgi-bin/KasbahLogin http://www.onsale.com/ http://www.primavera.com/products/primecontract.html http://www.primavera.com/products/primecontract.html http://ecommerce.media.mit.edu/tete-a-tete/ Intelligent agents for supporting construction procurement negotiation Introduction Practice of negotiation on construction procurement Negotiable issues Modeling negotiation preferences Category I issues Category II issues Category III and IV issues Modeling negotiation process for agents Task environment Agent environment Communication environment Genetic algorithm representation Computational model System implementation Case study Performance measures Experimental results and discussions Application limitation Conclusion Acknowledgements References 
work_57ykd43wibaffamofsstokmdy4	----	  Open Peer Review Any reports and responses or comments on the article can be found at the end of the article. SOFTWARE TOOL ARTICLE  ESforRPD2: Expert System for Rice Plant Disease Diagnosis [version 1; referees: 1 approved with reservations] Fahrul Agus ,     Muh. Ihsan , Dyna Marisa Khairina , Krishna Purnawan Candra 2 GIS and Environment Modelling Lab. CSIT, Mulawarman University, Samarinda, East Kalimantan, 75242, Indonesia Department of Agricultural Product Technology, Faculty of Agriculture, Mulawarman University, Samarinda, 75123, Indonesia Abstract One of the factors causing rice production disturbance in Indonesia is the lack of knowledge of farmers on early symptoms of rice plant diseases. These diseases are increasingly rampant because of the lack of experts. This study aimed to overcome this problem by providing an Expert System that helps farmers to make early diagnosis of rice plant diseases. Data of rice plant pests and diseases in 2016 were taken from Samarinda, East Kalimantan, Indonesia using an in-depth survey, and rice experts from the Department of Food Crops and Horticulture of East Kalimantan Province were recruited for the project. The Expert System for Rice Plant Disease Diagnosis, ESforRPD2, was developed based on the pest and disease experiences of the rice experts, and uses a Waterfall Paradigm and Unified Modelling Language. This Expert System can detect 48 symptoms and 8 types of diseases of rice plants from 16 data tests with an accuracy of 87.5%. ESforRPD2 is available in Indonesian at: http://esforrpd2.blog.unmul.ac.id Keywords Expert System, Rice Plant Disease, Waterfall, Unified Modelling Language   This article is included in the ICTROPS 2018 collection. 1 1 1 2 1 2  Referee Status:   Invited Referees      version 2 published 21 Feb 2019 version 1 published 06 Dec 2018 1 report , University of Georgia, USAYi Fang1  06 Dec 2018,  :1902 (First published: 7 )https://doi.org/10.12688/f1000research.16657.1  21 Feb 2019,  :1902 (Latest published: 7 )https://doi.org/10.12688/f1000research.16657.2 v1 Page 1 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 https://f1000research.com/articles/7-1902/v1 https://orcid.org/0000-0003-2983-137X https://orcid.org/0000-0003-1358-1329 http://esforrpd2.blog.unmul.ac.id https://f1000research.com/collections/ICTROPS2018 https://f1000research.com/collections/ICTROPS2018 https://f1000research.com/articles/7-1902/v2 https://f1000research.com/articles/7-1902/v1 https://orcid.org/0000-0003-2583-328X https://doi.org/10.12688/f1000research.16657.1 https://doi.org/10.12688/f1000research.16657.2 http://crossmark.crossref.org/dialog/?doi=10.12688/f1000research.16657.1&domain=pdf&date_stamp=2018-12-06    Fahrul Agus ( )Corresponding author: fahrulagus@unmul.ac.id   : Conceptualization, Funding Acquisition, Supervision, Writing – Original Draft Preparation, Writing – Review & Editing; Author roles: Agus F : Data Curation, Formal Analysis, Writing – Original Draft Preparation;  : Supervision;  : Writing – Review &Ihsan M Marisa Khairina D Candra KP Editing  No competing interests were disclosed.Competing interests:  The author(s) declared that no grants were involved in supporting this work.Grant information:  © 2018 Agus F  . This is an open access article distributed under the terms of the  , whichCopyright: et al Creative Commons Attribution Licence permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.  Agus F, Ihsan M, Marisa Khairina D and Candra KP. How to cite this article: ESforRPD2: Expert System for Rice Plant Disease Diagnosis    2018,  :1902 ( )[version 1; referees: 1 approved with reservations] F1000Research 7 https://doi.org/10.12688/f1000research.16657.1  06 Dec 2018,  :1902 ( ) First published: 7 https://doi.org/10.12688/f1000research.16657.1 Page 2 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 http://creativecommons.org/licenses/by/4.0/ https://doi.org/10.12688/f1000research.16657.1 https://doi.org/10.12688/f1000research.16657.1 Introduction Correct diagnosis of symptoms in rice plant diseases, caused by bacteria, nematodes, fungi, phythoplasmal and viruses1–4, is very critical in supporting the productivity of rice plants. However, many regions in Indonesia have a huge problem because of a limited number of rice plant pathologists. The large plan- tation area of rice plants is also a problem due to logistical issues when visiting these sites, leading to difficulty obtaining disease evidence. Along with other rapid technological developments, a technol- ogy known as Expert System (ES)5–8 has been developed to solve health9–12, education13, and business14, including agriculture15,16, problems. ES is usually designed for a specific condition, i.e. variables of climate in cases of agriculture. This article proposes a new software based on ES for the diagnosis of disease in rice plants in the Samarinda region, Indonesia. Waterfall Paradigm applied in designing this ES. The prototype, Expert System for Rice Plant Disease Diagnosis (ESforRPD2) is available at: http:// esforrpd2.blog.unmul.ac.id. Methods Data collection and ES development The ES of rice plant disease diagnosis was designed to help farmers and agricultural officials to diagnose rice plant diseases occurring in the Samarinda region, East Kalimantan province, Indonesia. Rice plant experts were recruited from the Seed Tech- nology Development Division at the Department of Food Crops and Horticulture of East Kalimantan Province and from the Department of Agro-eco-technology of Agricultural Faculty of Mulawarman University (one expert from each). The experts were the primary source for information on rice plant symptoms and diseases. The two rice plant experts have experience in diagnosing rice plant disease in the region of East Kalimantan Province for 20 years. Symptoms, diseases and their relationships (and their ranked importance) were derived from the experts by questionnaire (Supplementary File 1). This information was then used to construct the knowledge base for building the ES software. The ES software was developed using the Waterfall paradigm as recommended by Sommerville17 using five stages, i.e. (i) planning and requirement, (ii) analysis and software design, (iii) imple- mentation and unit testing, (iv) integration and (v) system test- ing and operation and maintenance. ES architecture consists of three parts, namely the user interface, the inference engine and the knowledge base as proposed by Lucas and van der Gaag7. The user interface is used as a consulting interface in order to obtain knowledge and advice from the ES, which would be like consulting an expert. In this ES, the inference engine works as a consultation system in processing input data to build a diagnosis based on the knowledge base developed. Implementation The implementation of the ESforRPD2 application is based on Unified Modelling Language (Figure 1) as proposed by Sommerville17, which consists of use case diagrams, activity diagrams, and class diagrams. We constructed two types of “Use case diagram”, namely “Use case for user” consisting of four cases (Article, Consult- ing, Choose Symptoms and Consulting Result); and “Use case for expert” consisting of three cases (Symptoms, Diseases, and Relation). The use case describes the functions of the ES inter- acting with user and expert. The activity diagram illustrates the flow of various activities being designed in the ES, i.e. how Figure 1. a. Use case diagram of user. b. Use case diagram of expert. Page 3 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 http://esforrpd2.blog.unmul.ac.id/ http://esforrpd2.blog.unmul.ac.id/ Figure 1c. Activity diagram of ESforRPD2 system application. the flow starts, the decision that might occur, and the flow end. The activity diagram also describes parallel processes that might occur in some executions. In this ES, we build four data stores (Expert, Symptoms, Relation, and Diagnosis) in the class diagram. The ESforRPD2 application uses four datasets, namely disease- and symptoms-data, knowledge base, and symptoms- disease-weight relationships table (Dataset 2). The construction of decision trees and forward-chaining tracing for diagnosing of rice plant diseases in the ES is shown in Figure 2. ESforRPD2 is the first version of ES (only in Indonesian) to make it user-friendly for Indonesian users. Users use a consul- tation page to choose the symptoms of the rice plant. The ES performs the calculation process to obtain the trust level using the Dempster-Shafer method18. The user page (Figure 3a) is the main web page for users without logging in. In the user page, there is also a home menu that displays articles about ES, rice plant diseases, and the Dempster-Shafer method. The consultation page starts the user consultation about the disease of rice plants (Figure 3b). The ES will provide an output as a display showing the symptoms, diagnosis of disease and the confidence level (Figure 3c). Operation The ESforRPD2 application is developed using CPU with specifications of Intel Core i3, 4GB RAM, and 300GB HDD. The same specification of CPU is needed to operate this application. Page 4 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 Figure 1d. Class diagram of ESforRPD2 system application. Uses case The ESforRPD2 application was tested applying symptom-data inputs by clicking the symptoms selected (Figure 5b). In a single test using the case of four symptom-data inputs selected, namely (i) Spots on leaf midrib, (ii) Little spots are dark brown or slightly purple rounded shape, (iii) Spots on oval-shaped leaves and evenly distributed on the leaf surface, (iv) The size of spots is 2–10 mm long and 1 mm wide, a display of diag- nosis page (Figure 3c) will appear following clicking of the “submit diagnose” button. The diagnosis page shows the confidence level. In this case test, the ES gave the accuracy of disease type detection of 91%. 16 tests in row were conducted using randomly selected symptoms by user in the ES. The results were approved by the two experts. In total, 14 diagnosis (87.5%) of the 16 results showed by the ES were justified by the two experts (Table 1). Discussion The ESforRPD2 application is showing good reliability. By applying 16 tests, the ESforRPD2 showed a level of performance of 87.5% (Table 1) following justification to two rice plants diseases experts. The performance of the ESforRPD2 during validation was the expected high-performance level of plant diseases diagnosis by the expert system. This performance is much higher than the performance of ES for Chili pepper pest diagnosis invented by Agus et al.16. However other Expert System could show excellent performance of 98.38%19, this evidence advice that the performance of ESforRPD2 could be improved in the next study. Currently, ESforRPD2 has only been tested with data from the Samarinda region. In a future study, we will use data from other regions of East Kalimantan, which have the same climate (tropical rainforest) and soil character as the Samarinda region. Page 5 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 Figure 2. Decision tree and forward chaining tracing. G1 G2 G3 G4 G5 G6 G7 G8 G9 G10 G11 G12 G13 G14 G15 G16 G17 G18 G19 G20 G21 G22 G23 G24 G25 G26 G27 G28 G29 G30 G31 G32 G33 G34 G35 G36 G37 G38 G39 G40 G41 G42 G43 G44 G45 G46 G47 G48 P1 P2 P3 P4 P5 P6 P7 P8 Page 6 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 Figure 3a. User main page. Figure 3b. Consultation page. Page 7 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 Figure 3c. Diagnosis results page. Table 1. System testing with expert justification. Test No. Experts Justification (English/Indonesian) Results Diagnosis of ESforRPD2 (English/Indonesian) Results 1 Blast/Blas Blas/Blast Suitable 2 Brown Spot (Bercak Coklat) Brown Spot (Bercak Coklat) Suitable 3 Narrow Brown Spot (Bercak Coklat Sempit) Narrow Brown Spot (Bercak Coklat Sempit) Suitable 4 Sheath Bligh (Hawar Pelepah) Sheath Bligh (Hawar Pelepah) Suitable 5 False Smut (Noda Palsu/Gosong Palsu) False Smut (Noda Palsu/Gosong Palsu) Suitable 6 Grassy Stunt (Kerdil Rumput) Grassy Stunt (Kerdil Rumput) Suitable 7 Bacterial leaf blight (BLB-Kresek Hawar Daun) Bacterial leaf blight (BLB-Kresek Hawar Daun) Suitable 8 Tungro (Tungro) Tungro (Tungro) Suitable 9 Blast (Blas) Blast (Blas) Suitable 10 Brown Spot (Bercak Coklat) Brown Spot (Bercak Coklat) Suitable 11 Narrow Brown Spot (Bercak Coklat Sempit) Blas/Blas Unsuitable 12 Sheath Bligh (Hawar Pelepah) Blast/Blas Unsuitable 13 False Smut (Noda Palsu/Gosong Palsu) False Smut (Noda Palsu/Gosong Palsu) Suitable 14 Grassy Stunt (Kerdil Rumput) Grassy Stunt (Kerdil Rumput) Suitable 15 Bacterial leaf blight (BLB-Kresek Hawar Daun) Bacterial leaf blight (BLB-Kresek Hawar Daun) Suitable 16 Tungro (Tungro) Tungro (Tungro) Suitable Page 8 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 References 1. Jagan Mohan K, Balasubramanian M, Palanivel S: Detection and Recognition of Diseases from Paddy Plant Leaf Images. Int J Comput Appl. 2016; 144(12): 34–41. Reference Source 2. Chapuis E, Besnard G, Andrianasetra S, et al.: First report of the root-knot nematode (Meloidogyne graminicola) in Madagascar rice fields. Australas Plant Dis Notes. 2016; 11(1): 32. Publisher Full Text 3. Qudsia H, Akhter M, Riaz A, et al.: Comparative Efficacy of Different Chemical Treatments for Paddy Blast, Brown Leaf Spot and Bacterial Leaf Blight Diseases in Rice (Oryza Sativa L.). Appl Microbiol Open Access. 2017; 3(3). Publisher Full Text 4. Kulmitra AK, Sahu N, Kumar VBS, et al.: In vitro evaluation of bio-agents against Pyricularia oryzae (Cav.) causing rice blast disease. Agric Sci Dig - A Res J. 2017; 37(3): 98–101. Publisher Full Text 5. Todd BS: An Introduction to Expert Systems. Issue 95, Oxford University Computing Laboratory, Programming Research Group; 1992. Reference Source 6. Turban E, Frenzel LE: Expert Systems and Applied Artificial Intelligence. Prentice Hall Professional Technical Reference; 1992. Reference Source 7. Lucas PJF, van der Gaag LC: Principles of Expert Systems. Amsterdam: Addison-Wesley; 1991. Reference Source 8. Gaines BR: Designing Expert Systems for Usability. Knowl Creat Diffus Util. 1986; 1–40. Reference Source 9. Gudu J, Gichoya D, Nyongesa P, et al.: Development of a medical expert system as an expert knowledge sharing tool on diagnosis and treatment of hypertension in pregnancy. Int J Biosci Biochem Bioinforma. 2012; 2(5): 297–300. Publisher Full Text 10. Abu Naser SS, Ola AZA: An expert system for diagnosing eye diseases using clips. Theor Appl Inf Technol. 2008; 923–930. Reference Source 11. Ayangbekun OJ, Jimoh IA: Expert System for Diagnosis Neurodegenerative Diseases. Int J Comput Inf Technol. 2015; 4(4): 694–698. Reference Source 12. Ayangbekun OJ, Bankole FO: An Expert System for Diagnosis of Blood Disorder. Int J Comput Appl. 2014; 100(7): 975–8887. 13. Divayana DGH: Utilization of cse-ucla model in evaluating of digital library program based on expert system at universitas teknologi indonesia: A model for evaluating of information technology-based education services. J Theor Appl Inf Technol. 2017; 95(15): 3585–3596. Reference Source 14. Arias-Aranda D, Castro JL, Navarro M, et al.: A fuzzy expert system for business management. Expert Syst Appl. 2010; 37(12): 7570–7580. Publisher Full Text 15. Ihsan M, Agus F, Khairina DM: Penerapan Metode Dempster Shafer Untuk Sistem Deteksi Penyakit Tanaman Padi. Prosiding Seminar Ilmu Komputer dan Teknologi Informasi. 2017; 2(1). Reference Source 16. Agus F, Wulandari HE, Astuti IF: Expert System With Certainty Factor For Early Diagnosis Of Red Chili Peppers Diseases. JAIS. 2017; 2(2): 52–66. Reference Source 17. Sommerville I: Software Engineering. 10th Edition. Pearson; 2016. Reference Source 18. Maseleno A, Mahmud Hasan M: Skin infection detection using Dempster-Shafer theory. In: 2012 International Conference on Informatics, Electronics and Vision, ICIEV 2012. 2012. Publisher Full Text 19. Sabzi S, Abbaspour-Gilandeh Y, García-Mateos G: A fast and accurate expert system for weed identification in potato crops using metaheuristic algorithms. Comput Ind. 2018; 98: 80–89. Publisher Full Text 20. Newbery F, Qi A, Fitt BD: Modelling impacts of climate change on arable crop diseases: progress, challenges and applications. Curr Opin Plant Biol. 2016; 32: 101–109. PubMed Abstract | Publisher Full Text 21. Xu C, Wu W, Ge Q: Impact assessment of climate change on rice yields using the ORYZA model in the Sichuan Basin, China. Int J Clim. 2018; 38(7): 2922–2939. Publisher Full Text 22. fahrulagus, Fajar MI: fahrulagus/paper: ESforRPD2 (Version V.10). Zenodo. 2018. http://www.doi.org/10.5281/zenodo.1490641 23. Agus F, Ihsan M, Khairina DM, et al.: Knowledge base for rice plant disease diagnosis [Data set]. Zenodo. 2018. http://www.doi.org/10.5281/zenodo.1490658 24. Agus F, Ihsan M, Khairina DM, et al.: Dataset for rice plant diseases expert interview [Data set]. F1000Research. Zenodo. 2018. http://www.doi.org/10.5281/zenodo.1953383 In addition, we will test data from other regions in Indonesia, which have a different climate. Newbery et al.20 showed that different climate conditions affect symptoms of arable crop disease; therefore, the ESforRPD2 will need continuous evaluation because climate change effects21. Consent Written informed consent was obtained from the two experts for participation in the study. Software availability Software application is available from: http://esforrpd2.blog. unmul.ac.id. Source code: https://github.com/fahrulagus/paper. Archived source code as at time of publication: https://doi. org/10.5281/zenodo.149064122 License: GNU GPL v3.0 Data availability Underlying data Zenodo: Knowledge base for rice plant disease diagnosis, https:// doi.org/10.5281/zenodo.149065823 Extended data Zenodo: Dataset for rice plant diseases expert interview, http:// doi.org/10.5281/zenodo.195338324 Grant information The author(s) declared that no grants were involved in supporting this work. Acknowledgements The authors are grateful to both experts in this research, the Rector of Mulawarman University and Islamic Development Bank Project. Page 9 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 https://pdfs.semanticscholar.org/8bf2/acd54720b36a1d7e9c8347b6fbd7889c13b6.pdf http://dx.doi.org/10.1007/s13314-016-0222-5 http://dx.doi.org/10.4172/2471-9315.1000138 http://dx.doi.org/10.18805/asd.v37i03.8989 https://www.cs.ox.ac.uk/files/3425/PRG95.pdf https://dl.acm.org/citation.cfm?id=573712&preflayout=flat https://www.cs.ru.nl/P.Lucas/proe.pdf https://pages.cpsc.ucalgary.ca/~gaines/reports/KBS/HFES/HFES.pdf http://dx.doi.org/10.7763/IJBBB.2012.V2.120 http://www.jatit.org/volumes/research-papers/Vol4No10/5Vol4No10.pdf https://www.ijcit.com/archives/volume4/issue4/Paper040413.pdf http://www.jatit.org/volumes/Vol95No15/17Vol95No15.pdf http://dx.doi.org/10.1016/j.eswa.2010.04.086 http://e-journals.unmul.ac.id/index.php/SAKTI/article/download/249/pdf https://publikasi.dinus.ac.id/index.php/jais/article/download/1455/1182 https://books.google.co.in/books?id=tW4VngEACAAJ&dq=Software+Engineering,+10th+Edition+PDF&hl=en&sa=X&ved=0ahUKEwij6dX544LfAhVJK48KHUvfD9sQ6AEILjAB http://dx.doi.org/10.1109/ICIEV.2012.6317330 http://dx.doi.org/10.1016/J.COMPIND.2018.03.001 http://www.ncbi.nlm.nih.gov/pubmed/27471781 http://dx.doi.org/10.1016/J.PBI.2016.07.002 http://dx.doi.org/10.1002/joc.5473 http://www.doi.org/10.5281/zenodo.1490641 http://www.doi.org/10.5281/zenodo.1490658 http://www.doi.org/10.5281/zenodo.1953383 http://esforrpd2.blog.unmul.ac.id/ http://esforrpd2.blog.unmul.ac.id/ https://github.com/fahrulagus/paper https://dx.doi.org/10.5281/zenodo.1490641 https://dx.doi.org/10.5281/zenodo.1490641 https://dx.doi.org/10.5281/zenodo.1490658 https://dx.doi.org/10.5281/zenodo.1490658 http://dx.doi.org/10.5281/zenodo.1953383 http://dx.doi.org/10.5281/zenodo.1953383   Open Peer Review Current Referee Status: Version 1 24 January 2019Referee Report https://doi.org/10.5256/f1000research.18205.r43488    Yi Fang Nano Electrochemistry Laboratory, College of Engineering, University of Georgia, Athens, GA, USA The author applied Expert System (ES) for rice plant disease and diagnosis; the background information and introduction is sufficient and well organized. The entire manuscript is also presented well. However, the author used the word “accuracy” which is not quite a scientific term. If “accuracy” is defined as sensitivity of the method, how about the specificity of the method? If a method has low specificity, it may not be able to solve the problem from false positive. For other comments, please see below: Page 1: Change “is the lack of knowledge of farmers on early symptoms…” to “is that farmers lack of knowledge of early symptoms…”.   Page 1: “accuracy of 87.5%” accuracy is not a scientific terminology, do you refer to sensitivity or specificity?   Page 3: change “… education, and business, including agriculture, problems” to “…education, business, and agriculture problems.”   Page 3: please change to “Waterfall Paradigm was applied in designing this ES.”   Page 5: “In this case test, the ES gave the accuracy of disease type detection of 91%”. Do you refer to sensitivity? Please also try to apply this comment to the other “accuracy” you mentioned in the manuscript. Is the rationale for developing the new software tool clearly explained? Partly Is the description of the software tool technically sound? Yes Are sufficient details of the code, methods and analysis (if applicable) provided to allow replication of the software development and its use by others? Yes Is sufficient information provided to allow interpretation of the expected output datasets and any results generated using the tool? Yes Page 10 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 https://doi.org/10.5256/f1000research.18205.r43488 http://orcid.org/0000-0003-2583-328X   Are the conclusions about the tool and its performance adequately supported by the findings presented in the article? Yes  No competing interests were disclosed.Competing Interests: Reviewer Expertise: electrochemistry, plant diseases, biosensors, sensor I have read this submission. I believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above. Author Response 12 Feb 2019 , Mulawarman University, IndonesiaFahrul Agus We agree with your judgment regarding the term of accuracy. We meant the accuracy is the sensitivity, for that reason we change the term accuracy to sensitivity. Regarding the term of specificity, we explain that this system has high specificity for rice plants because all data used in constructing the algorithm were collected specifically for rice plant diseases.   No competing interests were disclosed.Competing Interests: The benefits of publishing with F1000Research: Your article is published within days, with no editorial bias You can publish traditional articles, null/negative results, case reports, data notes and more The peer review process is transparent and collaborative Your article is indexed in PubMed after passing peer review Dedicated customer support at every stage For pre-submission enquiries, contact   research@f1000.com Page 11 of 11 F1000Research 2018, 7:1902 Last updated: 01 MAR 2019 
work_5bftzq2t35edjgifosexdmf6t4	----	Automatic Schaeffer’s Gestures Recognition System Francisco Gomez-Donoso Miguel Cazorla Alberto Garcia-Garcia Jose Garcia-Rodriguez Computer Science Research Institute. University of Alicante P.O. Box 99. E-03080. Alicante. Spain April 10, 2016 Abstract Schaeffer’s sign language consists of a reduced set of gestures designed to help children with autism or cognitive learning disabilities to develop adequate communication skills. Our automatic recognition system for Schaeffer’s gesture language uses the information provided by an RGB- D camera to capture body motion and recognize gestures using Dynamic Time Warping combined with k-Nearest Neighbors methods. The learning process is reinforced by the interaction with the proposed system that accelerates learning itself thus helping both children and educators. To demonstrate the validity of the system, a set of qualitative experiments with children were carried out. As a result, a system which is able to recognize a subset of 11 gestures of Schaeffer’s sign language online was achieved. keywords: Schaeffer’s gestures, 3D gesture recognition, Human-Machine Interaction, RGB-D sensors 1 Introduction Disabled people are a group that requires special attention from governments and people with cognitive disabilities (learning difficulties, cerebral palsy, etc.) are a special group within that one. Caregivers and educators need a way to communicate with them. In the case of children with autism, aided commu- nication systems from low (pictures) to high tech (speech devices), have been used. However, gestural communication or sign languages are still recommended as well to provide autistic children with augmentative or alternative means of communication [29]. Several studies have been presented in fields like psychology, neuroscience or linguistics, demonstrating that the combination of speech with gestures forms an integrated and useful system during language production and comprehension [27]. In particular, several works have studied how that combination can improve children comprehension in many teaching contexts [28]. For that purpose, a 1 Usuario Texto escrito a máquina This is a previous version of the article published in Expert Systems. 2016, 33(5): 480-488. doi:10.1111/exsy.12160 Usuario Nota adhesiva Completed definida por Usuario Usuario Nota adhesiva Accepted definida por Usuario Usuario Nota adhesiva None definida por Usuario http://dx.doi.org/10.1111/exsy.12160 special gesture set was developed by Schaeffer et al. [18]. To the best of our knowledge, there exists no system that is able to recognize these gestures. Gesture languages based on hand poses (static gestures) or movement pat- terns (dynamic gestures) have been used for implementing command and control interfaces [1–4]. Gestures, which involve spontaneous hand and arm movements that complement speech, have been proven to be a very effective tool for multi- modal user interfaces [5–9]. For instance, object manipulation interfaces [10–12] use the hand for navigation, selection and manipulation tasks in virtual envi- ronments. Several applications, such as heavy machinery control or manipulators, han- dling computer-generated avatars or musical interaction [13], use the hand as an efficient control device and with a high number of Degree of Freedom (DoF). In addition, some applications such as surgical immersive virtual reality simula- tions [14] and training systems (VGX, nd), include the manipulation of complex objects in their own definition. Almost all Human-Computer Interaction (HCI) systems based on gestures use hand movements as their main input. Currently, the most effective hand motion capture tools are electromechanical or magnetic detection devices (data gloves) [15, 16]. These devices are placed on the hand to measure the loca- tion and angles of the finger joints. They offer the most comprehensive set of real-time measurements, they are application-independent and allow full func- tionality of the hand in HCI systems. However, they have several disadvantages in terms of use and cost. In addition, they also hinder the movement of the hand and require complex calibration and installation procedures to obtain accurate measurements. Computer vision represents a promising alternative to data gloves because of its potential to provide a more natural interaction without intrusive devices. However, there are still several challenges to overcome for it to become more widely used, namely precision, processing speed, and generality. Among the different parts of the body, the hand is the most effective tool for general purpose interaction due to its communicative and manipulative functionalities. Some trends in interaction tend to adopt both modalities, thus allowing an intuitive and natural interaction. This work focuses on part of HCI, which is the branch of computer science that studies and develops new paths in the communication process between hu- mans and machines. HCI is a multidisciplinary field that includes topics such as artificial intelligence, design, social sciences, and natural language understand- ing. In particular, we present a gesture recognition system that is specially de- signed for Schaeffer’s gestures. The system, apart from gesture recognition, is able to show the user examples of the correct way to make the gestures, com- bined with speech reproduction of the described concept. Once the user imitates the gesture, the system provides the most similar one in the dictionary. This document is structured as follows. In Section 2 we briefly introduce Schaeffer’s gestures, what they are used for, and how they can help cognitively disabled people. Section 3 describes the general system architecture and explains the modules of the system: the input data, preprocessing and classification. Next, in Section 4 we present some experiments that were carried out using the gesture recognition system with different parameters over a set of recorded gestures in order to test its performance and its impact on children learning. 2 Finally, we draw some conclusions and outline future works in Section 5. 2 Schaeffer’s gestures Research on hand gesture falls within the behavioural analysis of interactions. Behavioural studies often involve observational methods, i.e., the adoption and/or adaptation of reliable coding systems shared by the scientific community [30]. There are many research studies that propose different taxonomies of gesture languages [30–34]. However, we are interested in a special group of patients: children with autism or cognitive learning disability. This kind of patients need a special way to learn to talk. In the year 1980, Schaeffer, Musil and Kollinzas published a book entitled “Total Communication: A signed speech program for non-verbal children” [18] in which they lay the foundations for interaction among people who are not able to speak. That work describes a complete sign language so that these people can interact with others more effectively. The speech signed program is an example of a system of signs, as classified by Kiernan [19], in which the therapist introduces the user to the speech signed language. It follows the structure of oral language with some spoken words which are accompanied with signs. The real strength of this system is that its use is based on the child’s overall development framework. The study of common development enables us to understand the communicative disorders that certain diseases cause. We can use this speech signed program without special authorization or training and it can be modified to meet the personal needs of people who might use it. Its learning and use does not obstruct or hinder, or therefore slow the onset of language, quite the opposite, it promotes and motivates language development. Both this Alternative Communication System (ACS) and other alternative sys- tems can be more than augmentative speech enhancers since they unlock this unique way of communication and allow others to be developed. The theoretical basis for ACS appeared in the USA in the year 1980, and a revised edition was published in 1994. The proposed system can recognize a subset of Schaeffer’s gestures (see Table 1). We aim to recognize the complete Schaeffer’s language in the future. The system has been designed as a tool which educators can use with autistic children that do not speak at all. The curiosity and interest of children in computers can accelerate the learning process that begins with the children just repeating the gesture, but after some time it includes the pronunciation of the words. The system can help to promote self-correction. There are three main components in a sign: the final movement, the position of the hand (relative to the body), and the shape of the hand. Firstly, it is necessary to use tactile help to show the children how to develop the gesture. Next, imitation aspects are very important and one of the steps where the system could be helpful. Finally, verbal support include full words, parts of words or even sentences. That part could also be improved with the system. 3 Water Help Sandwich Sleep Shower Clean Mom Dad Want Sick Dirty Table 1: Schaeffer’s gestures. 3 Gestures recognition In this section we explain the full system by giving an overview and describing the main device: the Microsoft Kinect v21. Next, we outline the gesture acqui- sition and characterization process and describe the classification methodology. Finally, we provide description of the model database. 3.1 System Overview Figure 1 shows a diagram of the system. When a person makes a gesture in front of the camera, the motion is captured by the Kinect v2. This information, summarized and packaged, is what the system understands as a gesture. The 1https://dev.windows.com/en-us/kinect 4 gesture object is sent to the Gesture Class Pre Selection (GCPS) module, which quickly executes with a naively selected subset of possible classes for the gesture, discarding others to improve performance. Then, both the subset of possible candidates classes and the gesture itself are sent to the classifier. The classifier compares the unknown gesture with every gesture present in the model using Dynamic Time Warping (DTW) [21], and it uses the Nearest Neighbor (NN) algorithm [22] to select the one with the shortest distance. The class is returned as the tag for the unknown gesture. This result is then sent to the user. The whole process is executed online. There is also an offline stage that handles the training of the model. Editing and condensing processes are applied in order to obtain a fitted and error free model. Figure 1: System architecture 3.2 Acquiring gestures: Microsoft Kinect v2 Microsoft Kinect v2 is an RGB-D camera capable of capturing the color and depth of a scene separately. The color stream is obtained with a common high definition RGB digital camera and for the depth stream it sends light beams that reflect on the surfaces and return to the sensor. By measuring the time difference between the emission and reception it calculates the depth of the element reflecting the beam. This process is known as Time of Flight (ToF). This device is capable of capturing color images with a resolution of 1920 × 1080 pixels, while the depth images are captured with a resolution of 512 × 424. The device is also able to capture audio and its direction of origin, segment elements such as bodies and other objects, and most importantly for the task at hand: by using its API, it is also able to detect the joints of the skeleton of a person, which makes us able to detect gestures. Kinect v2 is able to capture up to 25 joints, although our recognition system only uses 11 joints for the upper part of the body of the person, the others are ignored. 5 3.3 Gesture Characterization Once Kinect has captured these 11 joints of interest, two tasks are carried out before proceeding to the next module of the system: firstly, the points are grouped by joint type in order to facilitate the DTW comparison process in the classifier module, and then it runs a downsampling process in order to speed up the system. Kinect captures information at 30 fps, which means 330 points per second as long as a gesture lasts. Working with this amount of data implies a high computational cost, so the system reduces the information. The downsampling method used is k-means clustering with 20 centroids. In addition, it is necessary to obtain independence of the angle at which the subject is located when performing the gesture and its position in the scene. For this purpose, a change on the reference system of the captured points is performed. Firstly, the system obtains two vectors, one from the neck joint to the right shoulder one, and another from the neck joint to the head one. These represent the X and Y axes of the new reference system. The Z axis is obtained by performing the cross product of these two vectors. Then, the transformation matrix is calculated from the rotation and translation between the new reference system and the camera one. At last, the transformation matrix is multiplied by all the points that compose the gesture. Some sample gestures and their associated point clouds are shown in Figure 2. Figure 2: Clean and Sleep gestures with their associated joint point clouds. 3.4 Gesture Class Pre Selection The classification process compares the unknown gesture with all the gestures that compose the model, making it possible to accelerate the process by com- paring it only with a subset of gestures. This GCPS module implements a series of naive but fast to evaluate rules which examine the gesture features, such as hand position or point cloud cen- troids, and it is able to discard some classes. For example, if the gesture for Want is performed with the hand below the shoulders continuously, all gestures performed above them are automatically discarded. In this way, some classes are taken as impossible and pruned, so when the comparison with the model occurs, it contains a reduced set of examples, which leads to an improved run- time. The rules are evaluated in order and if the gesture does not meet any of the rules, the classifier runs with the full model. See Table 2 for an example of how the rules are checked. The rules implemented by the GCPS are: 6 • Rule 1: If the performer keeps the hand all the time over his head and there is a z-index variation below a threshold of 0.1 meters, the gesture will be classified as “Shower”. • Rule 2: If the performer moves the hand over and between his shoulders all the time, and there is a z-index variation below a threshold of 0.05 meters, the gesture will be classified as “Dirty”. • Rule 3: If the performer has his hands all the time closer than 0.22 meters one from the other, the selected classes will be “Help” or “Sleep”. • Rule 4: If the performer keeps his hand under and between his shoulders, and far from his head 0.2 meters, the selected classes will be “Help”, “Want” or “Sandwich”. • Rule 5: If the performer moves his hand over his neck all the time, the selected classes will be “Sick”, “Shower” or “Water”. • Rule 6: If the centroid of the points that describe the hand motion when performing a gesture is located to the right of the performer, the selected classes will be “Water”, “Dad” or “Clean”. • Rule 7: If the average distance between the hand and the head is below a threshold of 0.28 meters, the selected classes will be “Mom”, “Dad”, “Sick” or “Water”. “Help” gesture Rule 1: Is the performer keeping the hand all the time over his head and there is a z-index variation below a threshold of 0.1 meters? No. Evaluate the next rule. Rule 2: Is the performer moving the hand over and be- tween his shoulders all the time, and is there a z-index variation below a threshold of 0.05 meters? No. Evalu- ate the next rule. Rule 3: Is the performer having his hands all the time closer than 0.22 meters one from the other? Yes. Run the classifier with examples of the “Help” and “Sleep” classes. Table 2: The GCPS module evaluates the rules from first to last (the order is important). If at some point a rule is evaluated to true, the classifier will be executed with the examples bonded to that rule. 3.5 Classifiers In the classification process, the distance provided by DTW is calculated be- tween the unknown gesture and every gesture that composes the model. The 7 distance between two gestures is calculated by adding the partial distances aris- ing from comparing each point collection of each joint. That is why they are packed in such a way in the Gesture module. The NN algorithm is then executed and its class is returned as the label for the unknown gesture. An early abandon technique is also applied as follows: after each comparison between the unknown gesture and a gesture of the model, a check is performed to determine if the distance returned is the minimum distance so far, in which case it is stored. This distance is sent to the following comparison as a threshold value and if at some point of the comparison between two gestures the partial distance obtained is greater than that threshold, the algorithm finishes. In this way, we improve the runtime of this module [24]. 3.6 Model The model is composed of all the gestures that the system has learned. These gestures have been obtained from multiple persons who were recorded as they performed every gesture several times. Then, the gestures were labeled and stored. Within the model there are gestures that were not properly performed, mislabeled, or provide redundant or useless information. To eliminate these problems, two processes are performed. Firstly, the editing algorithm [25] is applied so those mislabeled gestures are discarded, and then the condensing or CNN algorithm [26] is executed. The latter extracts only those examples that actually provide new information to form the final model. This whole training process is performed offline. The model used by the final recognition system is composed of 253 different gestures spread over 11 classes . 4 Experiments Two different sets of experiments were carried out. On the one hand, we tested the system with a number of samples to prove its accuracy recognizing the proposed gestures. On the other hand, a small group of children with autism used the system under the supervision of educators to study the impact of its use on the sign language and speech learning process. A screenshot of the developed application is shown in Figure 3. 4.1 Gesture Recognition Accuracy The experimentation consists of using the system to classify a collection of 264 gestures captured with Kinect v2. In this collection, there exist 24 samples of each gesture type performed by five different persons. The classifier was set up with various parameters in order to determine the configuration that provides the best performance: • Downsampling with 20 centroids, with GCPS activated and using k-Nearest Neighbors (k-NN) (k = 3) for the classifier. • Downsampling with 20 centroids, with GCPS deactivated and using NN for the classifier. 8 Figure 3: Screenshot of the application which shows a person choosing a gesture and performing it. The recognized gesture is shown on the right side. • Downsampling with 20 centroids, with GCPS activated and using NN for the classifier. • Downsampling with 10 centroids, with GCPS activated and using NN for the classifier. Wa He Sa Sl Sh Si Cl Mo Da Wt Di 0 0.2 0.4 0.6 0.8 1 Gesture Type S u cc es s R a te 3NN, GCPS, 20 centroids NN, no GCPS, 20 centroids NN, GCPS, 20 centroids NN, GCPS, 10 centroids Figure 4: Success rate with the different parameter configurations. Figure 4 shows the results for the different parameter configurations. The system setup that provides the best success rate is the one downsampled with 20 centroids, with the GCPS module deactivated and using NN for the classi- fier method. However, this configuration requires too much runtime, making it impractical for real-time uses, which is what this project aims to address. Figure 5 shows that the fastest method for gesture classification tasks is the 10 centroids, with the GCPS module activated and using NN, but its success rate is below the threshold of acceptance. The second fastest system configuration, 9 the one downsampled with 20 centroids with the GCPS module activated and using NN, provides a very high success rate, making it the best option with a reasonable ratio of success rate to elapsed time. Figure 5 shows the average elapsed run time of a five cross validation round for these four configurations (a round is composed of 55 unknown gestures classifications). 3N N G C P S 20 ce nt ro id s N N no G C P S 20 ce nt ro id s N N G C P S 20 ce nt ro id s N N G C P S 10 ce nt ro id s 0 2,000 4,000 6,000 E la p se d T im e (m s) Figure 5: Execution time for a five cross validation round. Table 3: Timing for different configurations. All values are in ms. 5CV round 3NN GCPS 20 centroids NN no GCPS 20 centr. NN GCPS 20 centroids NN GCPS 10 centroids Total time GCPS time Class. time Total time Total time GCPS time Class. time Total time GCPS time Class. time 1 5642 97 5545 1519 720 99 621 212 62 150 2 5396 58 5338 1463 563 59 504 179 29 150 3 6113 66 6047 1465 644 68 576 201 34 167 4 5708 60 5648 1399 632 63 569 190 31 159 5 5158 54 5104 1129 473 51 422 138 24 114 As seen in Table 3, running the system with the GCPS module activated gives almost no impact in the overall runtime yet it improves the classification stage runtime. As exposed in Section 3.4 the GCPS module preselects some classes, decreasing the model size thus producing less comparisons in the clas- sification stage. Although the GCPS module introduces an error, it improves the execution time in every case while providing a high success rate, so its use is justified. We can see how the configurations with 20 centroids improve the execution time as well, while the 10 centroids one provides the best runtime but fails in terms 10 success rate. Regarding the classification algorithm, the k-NN (k = 3) provides a high success rate but with a prohibitive processing time. Instead of that, the best option is to use the nearest neighbor algorithm, which not only provides a high success rate but it is also faster. So, in the light of the experiments, the best setup for the gesture classification task is provided by the NN, GCPS activated and 20 centroids system. The evaluation of the evolution of the gesture learning process should be based on two main groups of features: sign component (shape, position respect body and final movement), and conditions (in presence of the object or in ab- sence of the object). Finally, it is also important to consider if it was necessary full help, partial help or absence of help by the educator. 4.2 Learning Improvement To evaluate the impact of the system in the children learning process there are three main stages: gesture imitation, gesture and speech imitation, and autonomous speech. The study with a large group of children in parallel with and without the use of the system was unable because of the absence of a large number of children in the same learning stage. The system is used mainly with children between 2 and 5 years and only a few children (1 or 2) begin at the same time. However, the qualitative evaluation of the educators indicates that, in all cases, the learning process is accelerated. Once they pass the phase in which a tactile help is necessary, the use of the system stimulates the visual imitation and fosters self-correction thus reducing the average time to achieve autonomous spoken communication. 5 Conclusions This approach provides an innovative, customizable and reliable system for Schaeffer’s gesture language detection and learning using Kinect v2. The main use case of the system is providing a tool to help children with autism to com- municate with gestures and speech. The system has been tested with excellent results with a group of children between two and five years old. It is also oriented to human-machine interaction for everyone, including people with cognitive dis- abilities, who can use this system to communicate with another person or a robot companion. The system can only recognize a subset of 11 different gesture classes, but we aim, as a future work, to recognize the whole of Schaeffer’s sign language and implement a system to detect when a gesture starts and ends in order to create a continuous real-time classification system. The main drawback of the proposed system is the GCPS module, as it is necessary to achieve real- time response, but its use decreases the accuracy of the system. In order to improve the performance of the system and make it more suitable for a real- time application with more gestures in the model, we also aim as a future work to parallelize the NN module. In addition, the whole system could be deployed in a heterogeneous computing platform for mobile robots such as the NVIDIA Jetson TK1, which features a multi-core CPU together with a many-core GPU for parallelization with low power consumption. 11 The proposed system has been tested with children, obtaining a positive qualitative evaluation from the educators. We could not get a quantitative evaluation due to the small population we can reach in a short period of time. We plan as a future work evaluate the system with more children in order to get a representative population. Acknowledgments. This work has been supported by the Spanish Government DPI2013-40534-R Grant, supported with Feder funds. We would like to thank APSA, an asso- ciation that helps people with mental handicaps in Alicante (Spain) for their collaboration in the experiments with autistic children. References [1] F. Quek, ”Unencumbered gestural interaction”. IEEE Multi-Media 3, 1996. [2] M. Turk, ”Handbook of Virtual Environments: Design, Implementation, and Applications”. Hillsdale: K.M. Stanney, 2002. [3] S. Lenman, ”Using marking menus to develop command sets for computer vision based hand gesture interfaces”. NY: ACM Press, 2002. [4] M. Nielsen, ”A procedurefor developing intuitive and ergonomic gesture in- terfaces for HCI”. 5th International Gesture Workshop, 2003. [5] A. Wexelblat, ”An approach to natural gesture in virtual environments”. 1995. [6] F. Quek, ”Multimodal human discourse: gesture and speech”. 2002. [7] R. Bolt, ”Put-that-there: voice and gesture at the graphics interface”. NY: ACM Press, 1980. [8] D.B. Koons, ”Iconic: speech and depictive gestures at the human-machine interface”.NY: ACM Press, 1994. [9] M. Billinghurst,” Put that where? Voice and gesture at the graphics inter- face”. 1998. [10] D. Bowman, ”Principles for the design of performance-oriented interaction techniques”. Hillsdale: Lawrence Erlbaum Associates, 2002. [11] J. Gabbard, ”A taxonomy of usability characteristics in virtual environ- ments”. Department of Computer Science, University of Western Australia, 1997. [12] V. Buchman, ”ingARtips: gesture based direct manipulation in augmented reality”. NY: ACM Press, 2004. [13] D. Sturman, ”Whole hand input”. MIT, 1992. 12 [14] A. Liu, ”A survey of surgical simulation: applications, technology, and education”. 2003. [15] D. Sturman, ”A survey of glove-based input”. 1994. [16] E. Foxlin, ”Motion tracking requirements and technologies”. Hillsdale: Lawrence Erlbaum Associates, 2002. [17] I. Oikonomidis, N. Kyriazis and A.A. Argyros, ”Efficient model-based 3D tracking of hand articulations using Kinect”, in Proceedings of the 22nd British Machine Vision Conference, BMVC 2011, University of Dundee, UK, Aug. 29-Sep. 1, 2011. [18] B. Schaeffer, A. Musil and G.Kollinzas, ”Total Communication: A Signed Speech Program for Nonverbal Children”. Research Press, 1980. [19] C. Kiernan, ”Alternatives to speech: A review of research and manual and other forms of communication with the mentally handicapped and other noncommunication populations”. British Journal of Mental Subnormality, 1977. [20] A. Rebollo et al., ”Diccionario de signos para alumnos con necesidades educativas especiales en el área de comunicación/lenguaje : programa de comunicación total habla signada de B. Schaeffer”. Conserjeŕıa de Educación y Universidades de la región de Murcia, 2011. [21] H. Sakoe and S. Chiba, Dynamic programming algorithm optimization for spoken word recognition, Acoustics, Speech and Signal Processing, IEEE Transactions on, vol 26, 1978. [22] S. Arya, et al, ”An optimal algorithm for approximate nearest neighbor searching in fixed dimensions”. University of Maryland, 1994. [23] S. P. Lloyd, ”Least square quantization un PCM”. Bell Telephone Labora- tories Paper, 1982. [24] J. Li and Y. Wang, ”EA DTW: Early abandon to accelerate exactly warp- ing matching of time series”. College of Computer Science and Technology, Huazhong University of Science and Technology, 2007. [25] D. S. Wilson, ”Asymptotic properties of nearest neighbor rules using edited data”. IEEE Transactions on Systems, Man, and Cybernetics, vol. smc-2, no 3, 1972. [26] P.E. Hart, The condensed nearest neighbor rule. IEEE Transactions on Information Theory, IT-14(3):515516, 1968. [27] S.D. Kelly, S.M. Manning and S. Rodak, Gesture Gives a Hand to Language and Learning: Perspectives from Cognitive Neuroscience, Developmental Psychology and Education,Language and Linguistics Compass 2 (2008). [28] S. Wagner-Cook and S. Goldin-Meadow, The Role of Gesture in Learning: Do Children Use Their Hands to Change Their Minds?, Journal of Cognition and Development, 7(2), 211232, 2006. 13 [29] J.B. Schwartz, C. Nye, Improving Communication for Children with Autism: Does Sign Language Work?, EBP Briefs, 1(2), 1-17, 2006. [30] R. Bakeman, J.M. Gottman, Observing interaction. An introduction to sequential analysis, II edn. Cambridge University Press, New York 1997. [31] J.B. Bavelas, N. Chovil,D.A. Lawrie, A. Wade, Interactive gestures. Dis- course Processes 15, 469489, 1992. [32] P. Ekman, W.V Friesen, The repertoire of nonverbal behavior. Semiotica 1, 4998, 1969. [33] A. Kendon, Gesture and speech: How they interact. In: Wiemann, J.M., Harrison, R.P. (eds.) Nonverbal Interaction, pp. 1345. Sage Publications, Beverly Hills, 1983. [34] D. McNeill, Hand and Mind. The University of Chicago Press, Chicago, 1992. 14 
work_5c6x6vtjqrg2jmbpbikwekbpta	----	An evolutionary voting for k-nearest neighbours An evolutionary voting for k-nearest neighbours Daniel Mateos-García, Jorge García-Gutiérrez, José C. Riquelme-Santos Keywords: Evolutionary computation Nearest-neigbour Weighted voting a b s t r a c t This work presents an evolutionary approach to modify the voting system of the k-nearest neighbours (kNN) rule we called EvoNN. Our approach results in a real-valued vector which provides the optimal relative con-tribution of the k-nearest neighbours. We compare two possible versions of our algorithm. One of them (EvoNN1) introduces a constraint on the resulted real-valued vector where the greater value is assigned to the nearest neighbour. The second version (EvoNN2) does not include any particular constraint on the order of the weights. We compare both versions with classical kNN and 4 other weighted variants of the kNN on 48 datasets of the UCI repository. Results show that EvoNN1 outperforms EvoNN2 and statistically obtains better results than the rest of the compared methods. e m i 2 i A ( m c G s t r c c K n P o C u m t I i K o o f p n l M t t s T c t m l w t n 1. Introduction Weighting in machine learning is a common technique for mpha-sizing some characteristics of data to improve the resulting odels. For example, weighting has been used to outline the mportance of some particular instances (Blachnik & Duch, 011) or features (Zhi, Fan, & Zhao, 2014), or rank a set of techniques n the context of en-sembles (Berikov, 2014). In a broad sense, rtificial Neural Networks (ANNs) and Support Vector Machines SVMs) can be also seen as ex-amples of using weights in learning odels but the k-nearest neigh-bours (kNN) has been the most ommon technique to benefit from weights (Mateos-García, García- utiérrez, & Riquelme-Santos, 2012). kNN and its variants have been widely used in the literature to olve real problems. Rodger (2014) used a hybrid model to predict he demand of natural gas. The system was implemented integrating egression, fuzzy logic, nearest neighbour and neural networks, and onsidering several variables such as the price, operating expenses, ost to drill new wells, etc. If we focus on biological data, Park and im (2015) selected significant genes from microarrays by using a earest-neighbour-based ensemble of classifiers. On the other hand, ark, Park, Jung, and Lee (2015) tackled the problem of designing rec- mmender systems. For this purpose the authors presented Reversed F (RCF), a fast item-based collaborative filtering algorithm which tilizes a k-nearest neighbour graph. i d e t c p A The main goal of a weighting system lies in the optimization (com- only by metaheuristics) of a set of weights in the training step to ob- ain the highest accuracy but trying not to overfit the resulting model. f we focus on kNN weighting methods, many proposals weight- ng features or instances can be found. In Raymer, Punch, Goodman, uhn, and Jain (2000) a weighting method to obtain an optimal set f features was provided. The set of features was selected by means f a kNN-based genetic algorithm using a bit vector to indicate if a eature was in the selection or not. In a later work, the same authors resented a hybrid evolutionary algorithm using a Bayesian discrimi- ant function (Raymer, Doom, Kuhn, & Punch, 2003) and trying to iso- ate characteristics belonging to large datasets of biomedical origin. oreover, Paredes and Vidal (2006) used different similarity func- ions to improve the behaviour of the kNN. In a first approximation, hey considered a weight by feature and instance on training data re- ulting in a non-viable number of parameters in the learning process. hen, the authors proposed three types of reduction: a weight by lass and feature (label dependency), a weight by prototype (proto- ype dependency) and a combination of the previous ones. The opti- ization process was carried out by descendant gradient. In the same ine, Tahir, Bouridane, and Kurugollu (2007) showed an approach that as able to both select and weight features simultaneously by using abu search. Furthermore, Mohemmed and Zhang (2008) presented a earest-centroid-based classifier. This method calculated prototyp- cal instances by considering arithmetic average from the training ata. To classify an instance, the method calculated the distance to very prototype and then selected the nearest one. The optimiza- ion of the best centroids that minimized the classification error was arried out through particle swarm. Fernandez and Isasi (2008) also roposed a weighting system by using a prototype-based classifier. fter a data normalization that was based on the position of the http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.08.017&domain=pdf mailto:mateosg@us.es mailto:jorgarcia@us.es mailto:riquelme@us.es http://www.lsi.us.es p v w o c b t t x d y x w δ e e t m t n s b s a t 2 t s m n 2 w n v s E i u i i m instances with respect to regions, the weights were iteratively calcu- lated. More recently, AlSukker, Khushaba, and Al-Ani (2010) used dif- ferential evolution to find weights for different aspects of data. They described four approaches: feature weighting, neighbour weighting, class weighting and mixed weighting (features and classes), with the latter being the one providing the best results. Weighting has also been applied to the vote system of the kNN. Thus, the distance-weighted k-nearest neighbour rule (WKNN) pro- osed by Dudani (1976) has been known for long. WKNN weights the otes of the k nearest neighbours (wi) according to Eq. (1) where di is the distance of the ith nearest neighbour (being d1 the distance of the nearest) regarding an instance to be classified. A similar ver- sion using a uniform weighting (UWKNN) has also been proposed where a weight is inversely proportional to the position reach among the neighbours (i.e., wu i = 1/i). Recently, both techniques have been explored working together as a new classifier called Dual-Weighted kNN (DWKNN) showing promising results where each weight was calculated according to Eq. (2) (Gou, Xiong, & Kuang, 2011). A later ork of Gou, Du, Zhang, and Xiong (2012) provided another version f DWKNN where the calculation of the weights were different ac- ording to Eq. (3). wwi = ⎧⎨ ⎩ (dk − di) (dk − d1) if di �= d1 1 if di = d1 (1) wdw1 i = wwi ∗ wui (2) wdw2 i = ⎧⎨ ⎩ (dk − di) (dk − d1) ∗ (dk + d1) (dk + di) if di �= d1 1 if di = d1 (3) Although all the previous approaches provided improvements re- garding the classical kNN performance, they have not explored the possible better suitability of evolutionary computation for the op- timization of the neighbours weights. Thus, we propose an evolu- tionary method to improve the kNN rule by altering the vote system knowledge obtained in the training phase. We also explore the use of constraints in learning weights with two different versions of our approach and we study their reliability compared with classical kNN and other 4 weighted variants on UCI datasets (Lichman, 2013). Fi- nally, results are statistically validated to reinforce the final conclu- sions. The rest of the paper is organized as follows. Section 2 presents the elements of the two versions of the evolutionary algorithm de- signed to weight the vote system of the kNN. The results and several statistical tests are specified in Section 3. Finally, Section 4 presents the main findings and future work. 2. Method In this section, two variants of our voting optimization system called Evolutionary Voting of Nearest Neighbours (EvoNN) are de- scribed. 2.1. Purpose and functionality The aim of our work was to find a set of weights to modify the influence of every nearest neighbour when they voted to assign a la- bel to an unlabelled instance. Moreover, our approach also provided a measurement about the influence of the proximity of neighbours by means of the optimized weights. Thus, the evolutionary process provides the optimal weights that have been found to improve the classification accuracy of the domain under study. To formalize our approach we assume that the set of classes (or labels) L is represented by the natural numbers from 1 to b, with b being the number of labels. Thus, let D = {(e, l) | e ∈ R f and l ∈ L = {1, 2, . . ., b}} be the dataset under study with f being the num- er of features and b the number of labels. Let label be a function hat assigns to every element e the real class to which it belongs o. Let us suppose that D is divided in the sets TR and TS, each of them being the training set and the testing set, respectively, such that D = T R ∪ T S and T R ∩ T S = ∅. In the training step the classification er- ror is minimized with participation only by instances from TR. This classification error is calculated as follows. For each x ∈ TR we com- pute its k nearest neighbours according to a distance function d. Let i with i = 1. . .k be the neighbours to x but ranked by distance, i.e.: (x, x1) ≤ d(x, x2). . . ≤ d(x, xk) and ∀y ∈ TR with y �= xi, d(x, xk) < d(x, ). According to the standard kNN rule, the prediction of the label of from the labels of its neighbours can be formalized: predLabel(x) = arg max l∈{1..b} k∑ i=1 δ(l, label(xi)) (4) here (l, label(xi)) = { 1 if label(xi) = l 0 otherwise (5) If instead of a unitary vote, we consider that the ith neighbour contributes with the weight wi ∈ R, the function (4) is redefined: predLabel(x, w) = arg max l∈{1..b} k∑ i=1 ωiδ(l, label(xi)) (6) The function to minimize is the sum of all prediction errors of ev- ry instance x from TR. Thus, if we define the error function as rror(x, w) = { 1 if predLabel(x, w) �= label(x) 0 otherwise (7) hen the function to optimize is: in w∈Rk ∑ x∈T R error(x, w) (8) With regard to the two versions of the evolutionary algorithm, hey were called EvoNN1 and EvoNN2. For EvoNN1, the nearest eighbours (those with lowest distances regarding the unlabelled in- tance) were “ heavier” and therefore, their influence (weight) had to e greater. In EvoNN2 weights had no constraints. Regarding the de- ign of the evolutionary algorithm, it is easy to suppose that EvoNN1 nd EvoNN2 were similar except on the encoding, crossover and mu- ation. .2. Voting optimization This subsection details the search algorithm to calculate the op- imum contribution of k nearest neighbours carried out by two ver- ions of an evolutionary algorithm . It is then necessary to define their ain characteristics i.e., individual encoding, genetic operators, fit- ess function and generational replacement policy. .2.1. Individual encoding In EvoNN1 and EvoNN2, an individual was a real-valued vector ith size k symbolizing the relative contribution of the k-nearest eighbours in the voting system of the kNN rule. Position 1 in the ector was associated with the nearest neighbour in distance and po- ition k with the furthest. Moreover, a constraint was established in voNN1 to assure that the closest neighbours were more important .e., ω1 ≥ ω2 ≥ . . .ωk. Regarding the initial population, both designs integrated individ- als with k random values between 0 and 1 (ordered in EvoNN1). To nclude individuals representing classic kNN, initial population also ncluded several vectors with the first k values set to 1 and the re- aining set to 0 in the initial population e.g., (1.0, 0.0, . . ., 0.0) for 1: Fitness(W, k, TR) : error 2: error = 0 3: for i = 1 to m do 4: Divide TR in s bags: B1...Bs 5: partialError = 0 6: for j = 1 to s do 7: partialError = partialError + Evaluate(W, k, TR − Bj, Bj) 8: end for 9: partialError = partialError/s 10: error = error + partialError 11: end for 12: error = error/m 13: return error 14: Evaluate(W, k, Train, Test) : error 15: error = 0 16: for each instTest in Test do 17: lab = NearestN(W, k, Train, instTest) 18: if lab = label(insTest) then 19: error = error + 1 20: end if 21: end for 22: error = error/size(Test) 23: return error 24: NearestN(W, k, Train, y) : labY 25: sortedInst and kNeighbours are empty sorted sets 26: for each x in Train do 27: insert x in sortedInst sorted by d(x, y) 28: end for 29: kNeighbours = sortedInst.get(k) 30: labY = majorityLabel(kNeighbours, W ) 31: return labY Fig. 1. Fitness function. k 1 s f 2 i l c o B S r m o o o d a g g t i i 2 T m d a o t v w c ( f fi d s ( fi ( v = 1, (1.0, 1.0, . . ., 0.0) for k = 2, etc. Finally, the maximum value of for a weight could be surpassed during the evolutionary process to tand out the importance of a concrete neighbour as explained in the ollowing section. .2.2. Crossover and mutation The goal of the crossover operator was the generation of a new ndividual (offspring) from the genotypic characteristics of a set of se- ected parents (two in our case, parent1 and parent2). There is also a onstraint in the order of the genes in EvoNN1. Thus, the crossover perator in the ith gene for EvoNN1 was defined as in Eq. (9) where LX − α stands for the crossover operator defined in Eshelman and chaffer (1993) and calculated from parent1(i) and parent2(i), γ is a andom value between 0 and 1, max = o f f spring(i − 1) and min = inimum(parent1(i), parent2(i), o f f spring(i − 1)). f f spring(i) = { BLX − α if i = 1 (max − min) ∗ γ + min otherwise (9) In EvoNN2, because the individuals were freely built, the crossover perator was simpler, see Eq. (10): f f spring(i) = BLX − α from parent1(i) and parent2(i) (10) Regarding the mutation operator, the ith gene of the individual in- iv could change according to Eq. (11) in EvoNN1 where δ was initially random value between 0 and 1. Then, to improve the fit through the enerations, the δ value was in the interval [0, 1] during the first 10 enerations. For the next 10, it was in [0, 0.9], etc. On the other hand, he mutation operator in EvoNN2 worked as can be seen in Eq. (12). ndiv′(i) = ⎧⎪⎪⎨ ⎪⎪⎩ indiv(i) + indiv(i) ∗ δ if i = 1 indiv(i) − indiv(i) ∗ δ if i = k (indiv(i − 1) − indiv(i + 1)) ∗ δ +indiv(i + 1) otherwise (11) ndiv′(i) = indiv(i) ± indiv(i) ∗ δ (12) .2.3. Fitness function EvoNN1 and EvoNN2 were defined by the same fitness function. his is because the proposed evolutionary designs were thought to inimize the classification error and therefore, the order of weights id not therefore have any influence on the fitness function. Both pproaches used the TR ⊂ D exclusively to obtain the contributions f the neighbours in the training step. Since we know the labels of he instances from TR, the fitness function was based on a cross- alidation error rate by using kNN and the weighted voting system. Fig. 1 shows the fitness calculation with m × s cross validations, here m stands for the number of iterations of the validation pro- ess (line 3) and s is the number of partitions of the training data TR line 4). Thus, the set TR is randomly divided in the bags B1, B2...Bs or each validation. Then, every bag Bj is evaluated through a classi- cation process by using T R − B j as a training set. This evaluation is riven by the function Evaluate which we will describe later. The clas- ification error on every Bj is accumulated on average by partialError lines 7 and 9), and by error in every validation (line 10). Finally, the tness value is the result of calculating the mean of all the validations line 12). The input parameters of the function Evaluate are the weighted ector W, the k value, and the subsets T R − B j and Bj (line 7). a i a e a w t f t W 2 d t r i Table 1 Comparison of accuracies of EvoNN1 and EvoNN2. In bold, the best. Dataset EvoNN1 EvoNN2 Dataset EvoNN1 EvoNN2 anneal 97.98 97.95 heart h 86.78 86.04 arrhythmia 60.13 59.76 heart stat 78.38 78.13 audiology 67.12 65.97 hepatitis 83.44 81.26 australian 83.89 84.11 hypothyroid 93.06 93.04 autos 75.42 75.12 ionosphere 90.45 90.36 balance 86.27 86.23 iris 99.78 99.77 breast c. 71.27 69.64 kr vs kp 98.5 98.55 breast w. 97.15 97.17 labor 81.69 81.22 bridges v1 70.7 67.86 landsat 93.86 93.77 bridges v2 69.79 66.42 liver disorders 61.49 60.72 car 83.43 86.89 lung cancer 81.4 75.79 cmc 63.21 63.07 lymph 80.87 80.47 colic 79.67 79.55 mfeat factors 97.1 97.13 credit a 84.34 83.75 mfeat fourier 86.19 85.91 credit g 72.17 72.81 mfeat karhunen 98.04 98 cylinder b 80.19 80.19 mfeat morph. 97.24 97.31 dermatology 95.39 95.62 mfeat pixel 96.93 96.92 diabetes 71.73 71.6 mfeat zernike 92.05 92.05 ecoli 93.17 92.81 molecular bio. 38.34 32.47 flags 55.4 54.73 monks 1 47.39 54.79 glass 60.8 62.57 monks 2 47.12 54.35 haberman 72.77 68.49 monks 3 47.78 46.22 hayes 62.35 62.26 mushroom 100 100 heart c 82.52 82.14 nursery 95.91 96.22 i t d p e o e t d r g m W t u m f T c s t ( t c s d u g i p t a t t b Therefore, the result of this function is the error rate on Bj taking T R − B j as reference to calculate the neighbours. For every single instance from the set used to measure the fitness (line 16), the returned label by the function NearestN is the majority one according to the k nearest instances belonging to the set used as training data (line 17). If the returned label does not correspond to the real label of the testing instance, the error is increased by 1 (line 19). Then, the resulting error is normalized with the size of the set used as testing data (line 22). Therefore, the value returned by Evaluate is real number between 0 (all instances are well-classified) and 1 (all nstances are misclassified). The function NearestN calculates the nearest instances to the ex- mple y belonging to the set under evaluation (line 24 and seq.). Ev- ry example of the neighbours bag is then inserted in a sorted set ccording to the distance to y. Thus, the example at the first position ill be the nearest to y and the one at the last position will be the fur- hest (line 27). When the algorithm selects the k nearest neighbours rom the sorted set (line 29), the majority label is returned according o the relative contribution of each neighbour expressed by the vector and according to the Eq. (6) (lines 30 and 31). .2.4. Generational policy Regarding the transition between generations, we chose an elitist esign where the best individual was transferred from one genera- ion to the next without being affected by the mutation operator. The emaining population is built as follows: with N being the number of ndividuals, C − 1 individuals were created by cloning the best indi- vidual from the previous generation, and the next N − C individuals resulted from the crossover operation. The selection of the individu- als to cross was carried out by the tournament method. Furthermore, all individuals except the first one were under mutation with a prob- ability of p. 3. Results Applying optimization techniques to a kNN classifier is a line of re- search that has been widely discussed in the literature with consider- able success. Finding an optimal weighting (for features or instances) increases the degrees of freedom of kNN. This allows the model to achieve a better fit in the training step and consequently improve the accuracy in testing. The novelty of this work lies in that we do not weight features or instances but the influence of each of the k near- est neighbours. In the traditional kNN model, the selected neighbours participate with a unitary vote. Our approach modifies the weight of the votes to modulate the contribution of each neighbour. Thus, our approach introduces a flexibility mechanism able to “ learn” geomet- ric arrangements of the different classes that depend on the domain under study. For example, if we focus on a dataset where the first neighbour is twice as important as the second one, and the second neighbour is in turn twice as important as the third, fourth and fifth one, we know that the traditional kNN is not able to exploit this sin- gularity but our proposal could suggest a weighted vector (4,2,1,1,1) which would be an optimal solution. To assess the quality of our approaches, we use 48 datasets from the repository UCI (Lichman, 2013) with different types of fea- tures and number of classes. All data were preprocessed with the same techniques i.e., binarization of nominal features, replacement of missing values and normalization to avoid the Hughes effect. Regard- ing the evolutionary search configuration and after a trial-and-error process, we used a population of 100 individuals, 100 generations, 10% of elitism and a mutation probability of 0.1. Regarding the pa- rameters α (crossover), g (mutation) and k (number of neighbours) their values were set at 0.5, 20 and 5 respectively for both EvoNN1 and EvoNN2. Although the experiments were carried out on four Intel® Xeon® Processor E7-4820, an increase of the computational cost s usually derived when using evolutionary computation. Thus, he CPU time for the optimization process can be expressed as #Generations×#Individuals×kNN time #threads where the number of available threads epends on the workload. Taking the German Credit Data as an exam- le (“credit g” in Table 1 with 1000 instances and 20 features), the volutionary algorithm took about thirteen and a half minutes to run nce. In a first level, a comparison between EvoNN1 and EvoNN2 was stablished. Table 1 shows the results obtained. Every row represents he mean accuracy after five 10-fold cross-validations (10FCV) with ifferent random seeds on a dataset. We provided the mean accu- acy to decrease the influence of randomness in the evolutionary al- orithms. The results show that EvoNN1 obtained better results for ost of the datasets. After testing the performance of both versions, we applied the ilcoxon signed-rank test to try to assure that the differences be- ween the two versions were statistically significant. The reason for sing a non-parametric test lies in the fact that the results did not eet the necessary conditions to apply parametric tests, especially or the sphericity condition (Demšar, 2006; García & Herrera, 2008). he statistic for Wilcoxon was 290.0 and the p-Value 0.0062, so we ould state that they behaved significantly different from each other. To measure the quality of our best approach, we established a econd level of comparison among IBk (implementation of kNN in he framework WEKA (Hall et al., 2009)), WKNN, UKNN, DWKNNv1 Gou et al., 2011) and DWKNNv2 (Gou et al., 2012), and EvoNN1. All he weighting algorithms were developed from the original IBk but hanging its voting system, and keeping the same k value of 5. Table 2 hows the mean accuracy obtained by the analyzed algorithms. Every ataset was evaluated for each technique as the mean of five 10FCV sing five different seeds. As can be seen, the performance of our al- orithm was the best in 25 out of the 48 datasets, and the second best n 7 out of the remaining 23. Although our approach seemed to outperform the rest of com- etitors, the results were statistically validated again to reinforce his conclusion. Thus, we carried out a non-parametric Friedman test nd a Holm post-hoc procedure to elucidate if the performance of he different algorithms was statistically different. The first step for he Friedman test is the calculation of the mean rankings reached y each technique (a ranking of 1 is the best). Table 3 shows the Table 2 Accuracy of every studied algorithm throughout 48 datasets from UCI. Datasets EvoNN1 IBk WKNN UWKNN DWKNN1 DWKNN2 anneal 97.98 97.1 97.85 98.43 99.14 98.23 arrhythmia 60.13 59.39 57.53 58.45 54.9 57.23 audiology 67.12 60.73 68.4 68.78 66.48 68.44 australian 83.89 84.38 82.16 83.46 80.4 82.11 autos 75.42 59.79 73.22 75.87 76.19 76.13 balance 86.27 88.28 86.81 82.65 82.01 86.81 breast c. 71.27 73.86 69.63 70.95 68.12 69.63 breast w. 97.15 97 96.09 96.36 95.39 96.04 bridges v1 70.7 58.34 65.11 64.93 64.08 65.24 bridges v2 69.79 60.47 62.75 63.37 62.15 62.75 car 83.43 93.13 93.13 88.16 88.16 93.13 cmc 63.21 45.83 45.26 45.37 44.16 44.56 colic 79.67 79.2 73.7 77.32 70.79 73.5 credit a 84.34 83.46 81.41 83.65 80.12 81.47 credit g 72.17 72.59 71.87 72.9 71.33 71.79 cylinder b 80.19 72.29 78.7 77.37 79.16 78.81 dermatology 95.39 96.26 96.32 95.79 95.16 96.32 diabetes 71.73 74.72 72.67 72.91 70.59 72.54 ecoli 93.17 86.49 82.84 83.78 80.24 82.18 flags 55.4 54.31 59.12 60.31 57.93 59.12 glass 60.8 66.13 70.27 68.05 69.49 70.07 haberman 72.77 71.08 70.46 67.55 64.92 69.77 hayes 62.35 27.66 67.52 67.47 74.62 67.52 heart c 82.52 83.31 80.6 80.14 76.66 80.27 heart h 86.78 79.41 78.56 78.56 76.67 78.54 heart stat 78.38 78.36 76.59 79.39 75.7 76.31 hepatitis 83.44 82.67 81.8 83.91 79.95 80.64 hypothyroid 93.06 93.25 92.62 93.13 91.2 92.5 ionosphere 90.45 85.61 87.46 86.33 86.89 87.62 iris 99.78 95.59 95.19 95.66 95.66 95.19 kr vs kp 98.5 96.2 95.95 95.12 93.7 95.95 labor 81.69 80.07 84.01 85.78 85.67 84.13 landsat 93.86 90.29 90.53 90.37 90.13 90.56 liver disorders 61.49 58.36 61.28 60.57 59.34 61.01 lung cancer 81.4 80.95 71.44 78.02 67.87 71.44 lymph 80.87 78.53 82.82 84.97 82.37 82.85 mfeat factors 97.1 96.56 96.69 96.51 96.08 96.68 mfeat fourier 86.19 81.56 81.02 81.26 80.07 80.94 mfeat karhunen 98.04 96.11 96.8 96.75 96.18 96.77 mfeat morph. 97.24 71.08 68.03 68.44 65.9 67.7 mfeat pixel 96.93 95.92 96.71 96.78 96.36 96.7 mfeat zernike 92.05 80.53 79.09 79.39 79.03 79.07 molecular bio. 38.34 32.57 33.56 34.76 35.24 33.56 monks 1 47.39 52.25 35.8 39.51 37.19 35.8 monks 2 47.12 54.13 38.27 40.78 38.03 38.27 monks 3 47.78 38.48 35.28 37.59 38.93 35.45 mushroom 100 100 100 100 100 100 nursery 95.91 98.36 93.63 96.1 93.23 93.63 Table 3 Mean ranking reached by every com- pared technique. Technique Ranking EvoNN1 2.270 UWKNN 3.062 IBk 3.458 WKNN 3.594 DWKNN2 3.781 DWKNN1 4.833 r s w 2 c c c m Table 4 Results for Holm post-hoc procedure. DataSet p z Holm’s adjusted α DWKNNv1 1.94E−11 6.710 0.010 DWKNNv2 7.65E−5 3.955 0.013 WKNN 5.32E−4 3.464 0.017 IBk 0.002 3.109 0.025 UWKNN 0.038 2.073 0.050 F ( t ( s v b a b c ankings in our case. Then, we carried out the Friedman test. The tatistic for Friedman was 48.95, distributed according to a chi-square ith 5 degrees of freedom. The p-Value for Friedman was around .302E−9 so the null hypothesis (no statistical difference among the ompared techniques) could be rejected with an α = 0.05. Because Friedman’ test should be accompanied by a post-hoc pro- edure for multiple pairwise comparison, we used the Holm pro- edure. This procedure took our approach (EvoNN1) as the control ethod and compared it with each other technique by a pairwise riedman test. The results of the procedure can be seen in Table 4 p-Value, Friedman statistic and adjusted α for Holm procedure). In his case, every test rejected the hypothesis of no pairwise difference p-Value below adjusted α), so we could state that our algorithm was ignificantly better than its competitors from a statistical point of iew taking into account that it obtained the best ranking and also ehaved statistically different from the rest. From our experimentation, some facts should be outlined. First, n evolutionary algorithm to adjust the contribution of the neigh- ourhood improved the behaviour of the classical kNN rule. This fact ould not completely be assured for the rest of weighting techniques B D D E F G G G probably due to a better adaptability of metaheuristics compared with strict functions as the suggested by DWKNN or UWKNN. Second, the correct working of the evolutionary approach was highly correlated with the distance of the weighted neighbours to the unlabelled instances in the training phase. In other words, although EvoNN2 had more degrees of freedom in the search space, EvoNN1 reached the best performance. Although this proved true, we think this fact can be related to problems in the evolutionary search in EvoNN2 so more work could be needed to make it reach its poten- tially better weights. 4. Conclusions This work presented an evolutionary method to optimize the kNN algorithm. Unlike the classical approaches that use weights on fea- tures and instances of data, we focused on adjust the contribution of each one of the k-nearest neighbours. The results showed a satisfac- tory performance that was statistically supported. On the other hand, despite the use of multitasking architectures, optimization techniques and more specifically evolutionary heuris- tics usually consume more computational resources than lazy classi- fiers. Therefore it will be necessary to establish flexible and scalable parallelization pathways that are in line with the growing amount of data. In future works, we will focus in alternative metrics to accuracy to use in unbalanced data classification. Thus, precision, recall, false positive rate, false negative rate or F-measure could be considered as the objective function to optimize. In addition, Big Data is a clear target to research. Although there are not yet many studies about using kNN in Big Data, we will try to extrapolate our method to apply in massive datasets. Regression problems are approachable too. The kNN algorithm works effectively when the class to predict is numeric so our method can be used in a natural way. Finally, the strengths of our system were tested on 48 hetero- geneous domains and the obtained results encourage us to imple- ment future experiments on more specific problems. In this con- text, we have experience with remote sensing data (García-Gutierrez, Gonçalves-Seco, & Riquelme-Santos, 2011) and this approach could compliment the cited study in a new work. References AlSukker, A., Khushaba, R., & Al-Ani, A. (2010). Optimizing the k-NN metric weights us- ing differential evolution. In Proceedings of international conference on multimedia computing and information technology (MCIT), 2010 (pp. 89–92). Berikov, V. (2014). Weighted ensemble of algorithms for complex data clustering. Pat- tern Recognition Letters, 38, 99–106. lachnik, M., & Duch, W. (2011). LVQ algorithm with instance weighting for generation of prototype-based rules. Neural Networks, 24(8), 824–830.(Artificial Neural Net- works: Selected Papers from ICANN 2010). emšar, J. (2006). Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research, 7, 1–30. udani, S. A. (1976). The distance-weighted k-nearest-neighbor rule. IEEE Transactions on Systems, Man and Cybernetics, 6(4), 325–327. shelman, L. J., & Schaffer, J. D. (1993). Real-coded genetic algorithms and interval- schemata. In D. L. Whitley (Ed.), Foundation of genetic algorithms 2 (pp. 187–202). San Mateo, CA: Morgan Kaufmann. ernandez, F., & Isasi, P. (2008). Local feature weighting in nearest prototype classifica- tion. IEEE Transactions on Neural Networks, 19(1), 40–53. arcía-Gutierrez, J., Gonçalves-Seco, L., & Riquelme-Santos, J. C. (2011). Auto- matic environmental quality assessment for mixed-land zones using lidar and intelligent techniques. Expert Systems with Applications, 38(6), 6805–6813. http://dx.doi.org/10.1016/j.eswa.2010.12.065. arcía, S., & Herrera, F. (2008). An Extension on “Statistical Comparisons of Classifiers over Multiple Data Sets” for all Pairwise Comparisons. Journal of Machine Learning Research, 9, 2677–2694. ou, J., Du, L., Zhang, Y., & Xiong, T. (2012). A new distance-weighted k-nearest neighbor classifier. Journal of Information & Computational Science, 9(6), 1429–1436. Gou, J., Xiong, T., & Kuang, J. (2011). A novel weighted voting for k-nearest neighbor rule. Journal of Computers, 6(5), 833–840. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., & Witten, I. H. (2009). The WEKA data mining software: an update. SIGKDD Explorations, 11(1). Lichman, M. (2013). UCI machine learning repository. http://archive.ics.uci.edu/ml. Mateos-García, D., García-Gutiérrez, J., & Riquelme-Santos, J. C. (2012). On the evolu- tionary optimization of k-NN by label-dependent feature weighting. Pattern Recog- nition Letters, 33(16), 2232–2238. doi:10.1016/j.patrec.2012.08.011. Mohemmed, A. W., & Zhang, M. (2008). Evaluation of particle swarm optimization based centroid classifier with different distance metrics. In Proceedings of IEEE congress on evolutionary computation’08 (pp. 2929–2932). Paredes, R., & Vidal, E. (2006). Learning weighted metrics to minimize nearest- neighbor classification error. IEEE Transactions on Pattern Analysis and Machine In- telligence, 28(7), 1100–1110. Park, C. H., & Kim, S. B. (2015). Sequential random k-nearest neighbor feature selec- tion for high-dimensional data. Expert Systems with Applications, 42(5), 2336–2342. http://dx.doi.org/10.1016/j.eswa.2014.10.044. Park, Y., Park, S., Jung, W., & Lee, S-g. (2015). Reversed CF: a fast collaborative filtering algorithm using a k-nearest neighbor graph. Expert Systems with Applications, 42(8), 4022–4028. http://dx.doi.org/10.1016/j.eswa.2015.01.001. Raymer, M., Doom, T., Kuhn, L., & Punch, W. (2003). Knowledge discovery in medical and biological datasets using a hybrid bayes classifier/evolutionary algorithm. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, 33(5), 802–813. Raymer, M., Punch, W., Goodman, E., Kuhn, L., & Jain, A. (2000). Dimensionality reduc- tion using genetic algorithms. IEEE Transactions on Evolutionary Computation, 4(2), 164–171. Rodger, J. A. (2014). A fuzzy nearest neighbor neural network statistical model for predicting demand for natural gas and energy cost savings in pub- lic buildings. Expert Systems with Applications, 41(4, Part 2), 1813–1829. http://dx.doi.org/10.1016/j.eswa.2013.08.080. Tahir, M. A., Bouridane, A., & Kurugollu, F. (2007). Simultaneous feature selection and feature weighting using hybrid tabu search/k-nearest neighbor classifier. Pattern Recognition Letters, 28(4), 438–446. Zhi, X., Fan, J., & Zhao, F. (2014). Robust local feature weighting hard c-means clustering algorithm. Neurocomputing, 134, 20–29. http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0007 http://dx.doi.org/10.1016/j.eswa.2010.12.065 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0012 http://archive.ics.uci.edu/ml http://dx.doi.org/10.1016/j.patrec.2012.08.011 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0015 http://dx.doi.org/10.1016/j.eswa.2014.10.044 http://dx.doi.org/10.1016/j.eswa.2015.01.001 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0019 http://dx.doi.org/10.1016/j.eswa.2013.08.080 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00562-X/sbref0022 An evolutionary voting for k-nearest neighbours 1 Introduction 2 Method 2.1 Purpose and functionality 2.2 Voting optimization 2.2.1 Individual encoding 2.2.2 Crossover and mutation 2.2.3 Fitness function 2.2.4 Generational policy 3 Results 4 Conclusions References 
work_5de6wpiw7zhctcpenowyonn4sm	----	Document downloaded from: This paper must be cited as: The final publication is available at Copyright http://dx.doi.org/10.1016/j.eswa.2011.07.098 http://hdl.handle.net/10251/34848 Elsevier Rodríguez Molins, M.; Salido Gregorio, MA.; Barber Sanchís, F. (2012). Intelligent planning for allocating containers in maritime terminals. Expert Systems with Applications. 39(1):978-989. doi:10.1016/j.eswa.2011.07.098. Intelligent Planning for Allocating Containers in Maritime Terminals M. Rodriguez-Molinsa,∗, M. A. Salidoa, F. Barbera aInstituto de Automática e Informática Industrial Universidad Politécnica de Valencia. Valencia, Spain Abstract Maritime container terminals are facilities where cargo containers are transshipped between ships or between ships and land vehicles (tucks or trains). These terminals involve a large number of complex and combinatorial problems. One of them is related to the Container Stacking Problem. A container yard is a type of temporary store where containers await further transport by truck, train or vessel. The main efficiency problem for an individual stack is to ensure easy access to containers at the expected time of transfer. Stacks are ’last-in, first-out’ storage structures where containers are stocked in the order they arrive. But they should be retrieved from the stack in the order (usually different) they should be shipped. This retrieval operation should be efficiently performed, since berthing time of vessels and the terminal operations should be optimized. To do this, cranes can relocate containers in the stacks to minimize the rearrangements required to meet the expected order of demand for containers. In this paper, we present a domain-dependent heuristically guided planner for obtaining the optimized reshuffling plan, given a stacking state and a container demand. The planner can also be used for finding the best allocation of containers in a yard-bay in order to minimize the number of reshuffles as well as to be used for simulation tasks and obtaining conclusions about possible yard configurations. Keywords: Planning, Heuristics, Optimizing, Container Stacking Problem 1. Introduction Maritime container terminals are the most important locations for transshipment and intermodal container trans- fers (Figure 1). [5] shows how this transshipment market is growing fast (container throughput has increased by 58 per cent over 2000-2004) and needs further studies to an- alyze it. In order to ensure reliability, e.g. delivery dates or handling times, to the different shipping companies as well as increasing productivity and container throughput from the quayside and landside and vice versa, there are several issues which need optimization. [18, 17] provide an extensive survey about operations at seaport container terminals and methods for their optimization. Moreover, other problems could be faced as for instance planning the routes for liner shipping services to obtain the maxi- mal profit [2]. Another important issue for the success at any container terminal is to forecast container throughput accurately [1]. With this data they could develop better operational strategies and investment plans. Containers are an ISO standardized metal box and can be stacked on top of each other. Loading and offloading ∗Corresponding author Email addresses: mrodriguez@dsic.upv.es (M. Rodriguez-Molins), msalido@dsic.upv.es (M. A. Salido), fbarber@dsic.upv.es (F. Barber) containers on the stack is performed by cranes following a ’last-in, first-out’ (LIFO) storage. In order to access a con- tainer which is not at the top of its pile, those above it must be relocated. It occurs since other ships have been un- loaded later or containers have been stacked in the wrong order due to lack of accurate information. This reduces the productivity of the cranes. Maximizing the efficiency of this process leads to several requirements: 1. Each incoming container should be allocated a place in the stack which should be free and supported at the time of arrival. 2. Each outgoing container should be easily accessible, and preferably close to its unloading position, at the time of its departure. In addition, there exist a set of hard/soft constraints re- garding the container locations, for example, small differ- ences in height of adjacent yard-bays, dangerous containers must be allocated separately by maintaining a minimum distance and so on. Nowadays, the allocation of positions to containers is usually done manually. Therefore, using appropriate Ar- tificial Intelligent techniques is possible to achieve signif- icant improvements of lead times, storage utilization and throughput. Figure 2 left shows a container yard. A yard consists of several blocks, and each block consists of 20-30 yard- Preprint submitted to Elsevier July 29, 2011 Figure 1: Container Terminal at Valencia bays [10]. Each yard-bay contains several (usually 6) rows. Each row has a maximum allowed tier (usually tier 4 or tier 5 for full containers). Figure 2 right shows a gantry crane that is able to move a container within a stacking area or to another location on the terminal. For safety reasons, it is usually prohibited to move the gantry crane while carrying a container [12], therefore these movements only take place in the same yard-bay. Figure 2: A container yard (left) and gantry cranes (right) (Photos by Stephen Berend) When a container arrives at the terminal port, a trans- fer crane picks it up and stacks it in a yard-bay. During the ship loading operation, a transfer crane picks up the container and transfers it to a truck that delivers it to a quay crane. In container terminals, the loading operation for export containers is pre-planned by load planners. For load plan- ning, a containership agent usually transfers a load profile (an outline of a load plan) to a terminal operating com- pany several days before a ship’s arrival. The load profile specifies only the container group. In order to have an ef- ficient load sequence, storage layout of export containers must have a good configuration. The main focus of this paper is to present a planning system which optimally reallocates outgoing containers for the final storage layout from which a load planner can con- struct an efficient load sequence list. In this way, the ob- jective is therefore to plan the movement of the cranes so as to minimize the number of reshuffles of containers in a complete yard. To this end, the yard is decomposed in yard-bays, so that the problem is distributed into a set of subproblems. Thus, each yard-bay generates a subprob- lem, but containers of different yard-bays must satisfy a set of constraints among them, so that subproblems will be sequentially solved taken into account the set of con- straints with previously solved subproblems. In the literature, generally this problem can be seen in two different ways according to when it should be done the optimization: 1. minimizing the number of relocations during the pickup operation. 2. getting a desirable layout for the bay before the pickup operation is done in order to minimize (or eliminate) the number of relocations during this process. [11] proposes a methodology to estimate the expected number of rehandles to pick up an arbitrary container and the total number of rehandles to pick up all the containers in a bay for a given initial stacking configuration. In a sim- ilar way, [9] compares two methods, branch-and-bound al- gorithm and a heuristic rule based on an estimator, which they minimize the number of relocations during the pickup operation. In [8], they also propose a methodology to convert the current bay layout into the desirable layout by moving the fewest possible number of containers (remarshalling) and in the shortest possible travel distance although it takes a considerable time since they use mathematical program- ming techniques. Cooperative coevolutionary algorithms have been developed in [13] to obtain a plan for remar- shalling in automated container terminals. This paper focuses on this latter issue. But we present a new heuristic with a set of optimization criteria in or- der to achieve efficiency and take into account constraints that should be considered in real-world problems in the provided solutions. 2. Problem description (The Container Stacking Problem) The Container Stacking Problem can be viewed as a modification of the Blocks World planning domain [19], which is a well-known domain in the planning community. This domain consists of a finite number of blocks stacked into towers on a table large enough to hold them all. The Blocks World planning problem is to turn an initial state of the blocks into a goal state, by moving one block at a time from the top of a tower onto another tower (or on a table). The optimal Blocks World planning problem is to do so in a minimal number of moves. Blocks World problem is closed to the Container Stack- ing Problem, but there are some important differences: • The number of towers is limited to 6 because a yard- bay contains usually 6 rows. 2 • The height of a tower is also limited to 4 or 5 tiers depending on the employed cranes. • There exist a set of constraints that involve differ- ent rows such as balanced adjacent rows, dangerous containers located in different rows, etc. • The main difference is in the problem goal specifi- cation. In the Blocks World domain the goal is to get the blocks arranged in a certain layout, specify- ing the final position of each block. In the container stacking problem the goal state is not defined as ac- curately, so many different layouts can be a solution for a problem. The goal is that the most immedi- ate containers to load are in the top of the towers, without indicating which containers must be in each tower. We can model our problem by using the standard en- coding language for classical planning tasks called PDDL (Planning Domain Definition Language) [3] whose purpose is to express the physical properties of the domain un- der consideration and it can be graphically represented by means of tools as [4]. A classical AI planning problem can be defined by a tuple ⟨A, I, G⟩, where A is a set of actions with preconditions and effects, I is the set of propositions in the initial state, and G is a set of propositions that hold true in any goal state. A solution plan to a problem in this form is a sequence of actions chosen from A that when ap- plied transform the initial state I into a state of which G is a subset. Following the PDDL standard, a planning task is de- fined by means of two text files. The domain file, which contains the common features for problems of this domain and the problem file, which describes the particular char- acteristics of each problem. These two files will be de- scribed in the following subsections. 2.1. Domain specification In this file, we will specify the objects which may ap- pear in the domain as well as the relations among them (propositions). Moreover, in order to make changes to the world state, actions must be defined. • Object types: containers and rows, where the rows represent the areas in a yard-bay in which a tower or stack of containers can be built. • Types of propositions: – Predicate for indicating that the container ?x is on ?y, which can be another container or, directly, the floor of a row (stack). on ?x - container ?y - (either row container) – Predicate for indicating that the container ?x is in the tower built on the row ?r. at ?x - container ?r - row – Predicate for stating that ?x, which can be a row or a container, is clear, that is, there are no containers stacked on it. clear ?x - (either row container) – Predicate for indicating that the crane used to move the containers is not holding any con- tainer. crane-empty – Predicate for stating that te crane is holding the container ?x. holding ?x - container – Predicates used to describe the problem goal. The first one specifies the most immediate con- tainers to load, which must be located on the top of the towers to facilitate the ship loading operation. The second one becomes true when this goal is achieved for the given container. goal-container ?x - container and ready ?x - container – Numerical predicates. The first one stores the number of containers stacked on a given row and the second one counts the number of con- tainer movements carried out in the plan. height ?s - row and num-moves • Actions: – The crane picks the container ?x which is in the floor of row ?r. pick (?x - container ?r - row) – The crane puts the container ?x, which is hold- ing, in the floor of row ?r. put (?x - container ?r - row) – The crane unstacks the container ?x, which is in row ?r, from the container ?y. unstack (?x - container ?y - container ?r - row) – The crane stacks the container ?x, which is cur- rently holding, on container ?y in the row ?r. stack (?x - container ?y - container ?r - row) – Finally, we have defined two additional actions that allow to check whether a given (goal) con- tainer is ready, that is, it is in a valid position. When a container is clear: fict-check1 (?x - container) The container is under another (goal) container which is in a valid position. fict-check2 (?x - container ?y - container) As an example of PDDL format, we show in Figure 3 the specification of the stack operator. Preconditions describe the conditions that must hold to apply the action: crane must be holding container ?x, container ?y must be clear and at row ?r, and the number of containers in that 3 row must be less than 4. With this constraint we limit the height of the piles. The effects describe the changes in the world after the execution of the action: container ?x becomes clear and stacked on ?y at row ?r, and the crane is not holding any container. Container ?y becomes not clear and the number of movements and the containers in ?r is increased in one unit. (:action stack :parameters (?x - container ?y - container ?r - row) :precondition (and (holding ?x) (clear ?y) (at ?y ?r) (< (height ?r) 4)) :effect (and (clear ?x) (on ?x ?y) (at ?x ?r) (crane-empty) (not (holding ?x)) (not (ready ?y)) (not (clear ?y)) (increase (num-moves) 1) (increase (height ?r) 1))) Figure 3: Formalization of the stack operator in PDDL. 2.2. Problem specification Once the problem domain has been defined, we can de- fine problem instances. These files describe the particular characteristics of each problem: • Objects: the rows available in the yard-bay (usually 6) and the containers stored in them. • Initial state: the initial layout of the containers in the yard. • The goal specification: the selected containers to be allocated at the top of the stacks or under other se- lected containers. • The metric function: the function to optimize. In our case, we want to minimize the number of reloca- tion movements (reshuffles). Since the Container Stacking Problem can be formal- ized with these two files, we can use a general domain independent planner to solve our problems as Metric FF [7]. The plan, which is returned by the planner, is a totally ordered sequence of actions or movements which must be carried out by the crane to achieve the objective. Fig- ure 4 shows an example of the obtained plan for a given problem. The performance of this general planner will be analyzed in Section 6, which will be compared with the domain-oriented planner presented in next Sections. 3. A Domain-Dependent Heuristically Guided Plan- ner Metric FF planner might obtain plans, but it is very in- efficient. Therefore, we propose a domain-dependent plan- Figure 4: The obtained plan solution to be carried out by the transfer crane. ner in order to provide more efficiency, it means at least re- ducing the number of crane operations required to achieve a desirable layout. The proposed planner is built on the basis of a lo- cal search domain-independent planner called Simplanner [16]. This planner has several interesting properties for the container stacking problem: • It is an anytime planning algorithm. This means that the planner can found a first, probably subopti- mal, solution quite rapidly and that this solution is being improved while time is available. • It is complete, so it will always find a solution if exists. • It is optimal, so that it guarantees finding the opti- mal plan if there is time enough for computation. It follows an enforced hill-climbing [6] approach with some modifications: • It applies a best-first search strategy to escape from plateaux. This search is guided by a combination of two heuristic functions and it allows the planner to escape from a local minima very efficiently. • If a plateau exit node is found within a search limit imposed, the hill-climbing search is resumed from the exit node. Otherwise, a new local search iteration is started from the best open node. The initial approach, based on Simplanner, was firstly used to solve individual subproblems (yard-bays). To im- prove the solutions obtained by Simplanner we have fur- ther developed a domain-dependent heuristic to guide the search in order to accelerate and guide the search toward a optimal or sub-optimal solutions. This heuristic (called h1) was developed to efficiently solve one yard-bay. h1 computes an estimator of the num- ber of container movements that must be carried out to reach a goal state (see Algorithm 1). The essential part of this algorithm is to count the number of containers lo- cated on the selected ones, but also keeps track of the 4 containers that are held by the crane distinguishing be- tween whether they are selected containers or not. When the crane is holding a selected container, the value h has a smaller increase since, although this state is not a solution, this container will be at the top of some row in the next movement. Algorithm 1: Pseudo-code of the domain- dependent heuristic h1 Data: b: state of the yard-bay; Result: h: heuristic value of b; h ← 0; // Container hold by the crane if ∃x−container / Holding(x) ∈ b then if GoalContainer(x) then h ← 0.1; else h ← 0.5; end end // Increasing the ∆h value for r ← 1 to numRows(b) do ∆h ← 0; for x−container / At(x, r) ∧ GoalContainer(x) ∈ b do if @y−container / GoalContainer(y) ∧ On(y, x) ∈ b then ∆h ← max(∆h, NumContainersOn(x)); end end h ← h + ∆h; end 4. Optimization criteria for one-bay yards Despite we are able to obtain good solutions (layouts) from Simplanner enhanced with h1, we also want solu- tions more realistic for instance taking into account safety standards. From this heuristic h1, we have developed some opti- mization criteria each one of them achieving one of the requirements we could face at Container Terminals [15]. These criteria are centered in the next issues: 1. Reducing distance of the goal containers to the cargo side (OC1d). 2. Increasing the range of the move actions set for the cranes allowing to move a container to 5th tier (OC1t). 3. Applying different ways of balancing within the same bay in order to avoid sinks (OC1b). These criteria have been easily incorporated in our planner by defining a heuristic function as a linear combi- nation of two functions: h(s) = α · h1(s) + β · h2(s) (1) being this secondary function a combination of these three criteria described: h2(s) = OC1d + OC1t + OC1b (2) Note that although we want to guarantee balancing with this last optimization criterion, unbalanced states (states with sinks) are allowed during this process of re- marshalling in order to get better solutions according to the number of reshuffles done. 4.1. OC1d: Placing goal containers close to cargo side Given an initial state, several different layouts can be usually achieved making the same number of reshuffles and some of them can be more interesting than the rest accord- ing to other important questions. In this case, since the transfer crane is located at the right side of the yard-bay, we want to obtain a layout where it is minimized the dis- tance of the goal containers to this side of the yard-bay. Achieving this we can spend considerably less time during the truck loading operations. Figure 5: Obtained plan with the initial domain-dependent heuristic. Following the heuristic function presented in Equation 1: • h1(s) is the main heuristic function, which estimates the number of movements required to reach the goal layout (outlined in Algorithm 1). Since this is the main optimization function, α value should be sig- nificantly higher than β. • h2(s) is the secondary function we want to optimize. In this case, it is just OC1d. This means the sum of the distances of the selected containers to the right side of the yard-bay, which can be computed as Al- gorithm 2 shows. Algorithm 2: Pseudo-code to calculate the distance Data: s: state to evaluate Result: d: distance value of s d ← 0; for r ← 1 to numRows(s) do for x−container / At(x, r) ∈ s ∧ GoalContainer(x) do d ← d + (numRows(s) − r); end end The benefits of using this combined heuristic function can be observed in Figure 5 and Figure 6. In the first one we want only to minimize the number of reshuffles, i.e. h(s) = h1(s). In the second one, we also want to minimize 5 the distance of the selected containers to the forklift truck, so we have set h(s) = 9 ∗ h1(s) + h2(s). As a result, none of the selected containers (the red ones) are placed in the most left rows, reducing the required time to load the truck. Figure 6: Obtained plan with the distance optimization function. 4.2. OC1t: Allowing the 5th tier during the remarshalling process In this optimization criterion as well as the next ones, we will include the new given heuristic value with the same factor as the initial one. One of the decisions that must be done in Container Terminals is about which cranes have to be bought depending on how many tiers cranes work. This topic has been considered in [14]. But, another approach is to reach the fifth tier only during the remarshalling pro- cess. Thereby, there would be 4 tiers at the beginning and the end keeping the first requirements. Following this concept, we will use instances of prob- lems < n, 4 > with a domain whose move actions allow 5 tiers at the stacks. This function is showed in Algorithm 3 and it follows the same steps than the original but in- creasing the value of h when the height of one of the stacks is higher than 4. Thereby, we assure that the final layout will always have 4 tiers. 4.3. OC1b: Balancing one yard-bay In this section we present an extension for the heuris- tic h1 (Algorithm 1) to include the balancing of the stacks within one yard-bay as a requirement. It is considered that there is a sink when the height difference between two adjacent stacks in the same yard-bay is greater than a maximum number of containers, in our case two contain- ers. Considering the time when the goal containers are re- moved from the yard, we can distinguish three ways to get balanced one yard-bay presented in the next subsections. The last mode is the consequence of applying the first two ones. 1. Balanced before loading operation In this case, we consider that the layout must be balanced before the goal containers are removed from that yard-bay. This function is showed in Algorithm 4, it compares the height of each row of the yard-bay with the next Algorithm 3: Pseudo-code of the domain- dependent heuristic function to allow 5 tiers Data: s: state to evaluate Result: h: heuristic value of s h ← 0; if ∃x−container / Holding(x) ∈ s then if GoalContainer(x) then h ← 0.1; else h ← 0.5; end end for r ← 1 to numRows(s) do ∆h ← 0; if Height[r, s] > 4 then if x−container / Clear(x, r) ∈ s ∧ GoalContainer(x) then ∆h ← 0.5; else ∆h ← 1; end end for x−container / At(x, r) ∈ s ∧ GoalContainer(x) do if @y−container / GoalContainer(y) ∧ On(y, x) ∈ s then ∆h ← max(∆h, NumContainersOn(x)); end end h ← h + ∆h; end one, and if the difference is higher than 2, the value heuristic h is increased. As it appears in Figure 7, this criterion avoids the sinks in the final layout while all the containers are still in the yard-bay. However, when these containers are removed, it might cause that the new layout is unbalanced as it hap- pens in Figure 7(c). Algorithm 4: Pseudo-code to balance before the goal containers are removed Data: s: state to evaluate; h: Initial heuristic; Result: h: heuristic value of s; for r ← 1 to numRows(s) − 1 do ∆h ← Abs(Height[r, s] − Height[r + 1, s]) − 2; if ∆h > 0 then h ← h + ∆h; end end 2. Balanced after loading operation In contrast to the method seen above, we can consider that the layout must remain balanced after the goal containers are removed from the yard-bay. Figure 8 shows the layouts we get after execute the plan returned by our planner. Algorithm 6 shows this function. It uses the Func- tion HeightsWithoutGoals (Algorithm 5) in order to calculate for the yard-bay b the height for each stack where the first no-goal container is. These values are employed to get the difference of height between two adjacent stacks once the goal containers have been removed from the yard. Heights of each row are stored as soon as the planner gets the final so- lution plan for one yard-bay. After we obtain these values, we increase the heuristic value h according to whether or not there are goal containers on the floor. 6 (a) Initial Layout (b) With goal containers (c) Without goal containers Figure 7: Effects of using function seen in Algorithm 4 Algorithm 5: Function HeightsWithoutGoals to calculate heights of each row without taking into ac- count the goal containers at the top Data: b: state of the yard-bay; Result: MinHeight, heights calculated; for r ← 1 to numRows(b) do MinHeight[r, b] ← Height[r, b]; // Decrease till the first no goal-container while MinHeight[r, b] > 0 ∧ GoalContainer(MinHeight[r, b], r) ∈ b do MinHeight[r, b] ← MinHeight[r, b] − 1; end end Then, we use the values given by HeightsWithout- Goals to calculate the difference between two adja- cent stacks, when this difference is higher than 2 we consider that there is a sink, so h is increased again. However, this process might also cause some unbal- anced layouts (Figure 8(b)). But in this case, non- desirable layouts will appear while the goal contain- ers are in the yard-bay. Once they have been re- moved from it, these layouts will be balanced ones (Figure 8(c)). 3. Balanced before and after loading operation Finally, we present an optimization criterion which obtains a layout where is balanced both before and after the goal containers are removed from this yard- bay. With this function we want to solve the prob- lems seen in the last subsections as we can see it in Figure 9. This function (Algorithm 7) is a mixture of the last two ones. First, we increase h when there are goal containers on the floor. When this is achieved, we in- crease h when the difference between the heights val- ues obtained by the function HeightsWithoutGoals (Algorithm 5) are higher than 2 for two contiguous rows. And finally, if h value is low enough (in our case lower than 1), we increase h again if the differ- Algorithm 6: Pseudo-code to balance after the goal containers are removed Data: s: state to evaluate; h: Initial heuristic; Result: h: heuristic value of s; HeightsWithoutGoals(s); ∆h ← 0; // Not allow containers on the floor for r ← 1 to numRows(s) do if ∃x−container / On(x, r) ∧ GoalContainer(x) then if MinHeight[r, s] > 0 then ∆h ← ∆h + NumContainersOn(x); end end end h ← h + ∆h; for r ← 1 to numRows(s) − 1 do ∆h ← Abs(MinHeight[r, s] − MinHeight[r + 1, s]); if ∆h > 2 then h ← h + ∆h − 2; end end (a) Initial Layout (b) With goal containers (c) Without goal containers Figure 8: Effects of using function seen in Algorithm 6 ence between the actual heights of two contiguous rows is higher than 2. 5. Optimization criteria for one block This initial heuristic (h1) was unable to solve a com- plete yard or block (in our case, one block consists of 20 yard-bays) due to the fact that they only solve individual yard-bays. In this paper, we also have developed two opti- mization criteria that include new constraints that involve several yard-bays. These constraints are: • Balancing contiguous yard-bays: rows of adjacent yard-bays must be balanced, that is, the difference between the number of containers of row j in yard- bay i and row j in yard-bay i − 1 must be lower than a maximum (in our case lower than 3). Figure 10 shows which rows must be get balanced when we consider one yard-bay and Figure 11 left shows an example of non-balanced yard-bays (rows in dotted points). 7 Algorithm 7: Pseudo-code to balance the yard-bay before and after the goal containers are removed Data: s: state to evaluate; h: Initial heuristic; Result: h: heuristic value of s; HeightsWithoutGoals(s); ∆h ← 0; // Not allow containers on the floor for r ← 1 to numRows(s) do if ∃x−container / On(x, r) ∧ GoalContainer(x) then if MinHeight[r, s] > 0 then ∆h ← ∆h + NumContainersOn(x); end end end h ← h + ∆h; if h < 2 then ∆h ← 0; // Balancing with containers which are not objective for r ← 1 to numRows(s) − 1 do ∆h ← Abs(MinHeight[r, s] − MinHeight[r + 1, s]); if ∆h > 2 then h ← h + 0.6 × (∆h − 2); end end if h < 2 then // Balancing with containers which are objective for r ← 1 to numRows(s) − 1 do ∆h ← Abs(Height[r, s] − Height[r + 1, s]); if ∆h > 2 then h ← h + 0.4 × (∆h − 2); end end end end (a) Initial Layout (b) With goal containers (c) Without goal containers Figure 9: Effects of using function seen in Algorithm 7 • Dangerous containers: two dangerous containers must maintain a minimum security distance. Figure 11 right shows an example of two dangerous containers that does not satisfy the security distance constraint. These constraints interrelate the yard-bays so the prob- lem must be solved as a complete problem. However, it is a combinatorial problem and it is not possible to find an optimal or sub-optimal solution in a reasonable time. Following the previous philosophy of solving each subprob- lem independently (each yard-bay separately), we can dis- (i,j)(i-1,j) (i,j+1) (i+1,j) (i,j-1) Bays S ta c k s Figure 10: Balancing scheme tribute the problem into subproblems and solve them se- quentially taken into account related yard-bays. Thus a solution to the first yard-bay is taken into account to solve the second yard-bay. A solution to the second yard-bay is taken into account to solve the third yard-bay. Further- more, if there exist a dangerous container in a first bay, its location is taken into account to solve a dangerous con- tainer located in the third yard-bay (if it exists); and so on. Taken into account this distributed and synchronous model, we present two different optimization criteria to manage these types of constraints. These two criteria are added to the heuristic function seen in Equation 1 as h3 (Equation 3); and Equation 4 shows the exact combination of them. This makes possible to follow a criterion with major priority than the other one. h = α · h1 + β · h2 + γ · h3 (3) h3 = δ1 · OCnB + δ2 · OCnD (4) As a consequence of the solving mode followed, depend- ing on the order the yard-bays are resolved may not be possible to achieve a solution. Moreover, as mentioned in Section 4, although we want to guarantee balancing and/or minimum distance between dangerous containers, during relocation of container process we will allow the presence of non-desirable sates, e.g. with some sinks between two contiguous rows or bays. These intermediate states are allowed because through them we will be able to get bet- ter solutions taking into account as metric function the number of reshuffles done. 5.1. OCnB: Balancing contiguous yard-bays In this section we present an extension for the heuristic h1 (Algorithm 1) to include the balancing of continuous yard-bays as a requirement. It is considered that there is a sink when a difference higher than two containers exists between two adjacent rows in contiguous yard-bays. This criterion is an extension of the balanced heuristic presented in Algorithm 7, which avoids sinks in the same yard-bay (horizontal balance) both before and after the outbound 8 i-1 i i-2 i ( ) ( ) ( ) 45 2223)2( min 222 =<= −+−+−−= Dd iid d Figure 11: (Left) Non-balanced yard-bays. (Right) Proximity of two dangerous containers. containers have been removed from the yard. However, in this case a sink represents a constraint between two subproblems. Thus, we also consider that there is a sink when a difference of two exits between the same row r in two contiguous yard-bays (vertical balance). This process is showed in Algorithm 8. This also uses the Function HeightsWithoutGoals (Algorithm 5) in order to calculate for the yard-bay b the height for each stack where the first no-goal container is. Heights of each row are stored as soon as the planner gets the final solution plan for one yard-bay. First, we apply the criterion seen in Algorithm 7 on the yard-bay b. Through heights’ calculated by Algorithm 5 and the real heights of the actual yard-bay we obtain the differences between the row r and r − 1 to calculate the value of h. When this value is zero (the yard-bay b is horizontally balanced), then we introduce our function to balance it with respect to the last yard-bay bl. To do so, we must also calculate the heights’ through the Algorithm 5 over bl and use the real heights of it in order to obtain the differences between the row r situated in b and bl. When these differences are higher than 2, we increase h propor- tionally. After that process, b will be balanced horizontally with respect to their rows, and vertically with respect to the last yard-bay. Repeating this process for each yard-bay in the block, this will be completely balanced. 5.2. OCnD: Dangerous containers Within a block, there are different types of contain- ers depending on the goods they transport, being some of them dangerous. If they do not satisfy certain restrictions, it may become a hazard situation for the yard since e.g. if one of them explodes and they are not enough far between them, it will set off a chain of explosions. With this added objective, the next optimization cri- terion (Algorithm 9) ensures a minimum distance (Dmin) between every two dangerous containers (Cd) in the yard. Dmin is set as one parameter for the planner and the dis- tance is calculated as the Euclidean distance, considering each container located in a 3-dimensional space (X,Y,Z) Algorithm 8: Pseudo-code to balance two adjacent yard-bays Data: b: state of the actual yard-bay; h: Initial heuristic; bl: last yard-bay; Result: h: heuristic value of b // Getting the balance horizontally HeightsWithoutGoals(b); h ← h + BalBeforeAfter(b); // This heuristic will be executed after a partial solution if h < 1 ∧ NumBay(b) ̸= 1 then ∆h ← 0; HeightsWithoutGoals(bl); // Balancing with containers which are not objective for r ← 1 to numRows(b) do ∆h ← Abs(MinHeight[r, bl] − MinHeight[r, b]); if ∆h > 2 then h ← h + 0.6 × (∆h − 2); end end if h = 0 then // Balancing with containers which are objective for r ← 1 to numRows(b) do ∆h ← Abs(Height[r, bl] − Height[r, b]); if ∆h > 2 then h ← h + 0.4 × (∆h − 2); end end end end where X is the number of yard-bays, Y is the number of rows and Z is the tier. Generally, in container terminals, at most, there is only one dangerous container in two contiguous yard-bays, so that we take into account this assumption in the develop- ment of this function. This function increases h value when a dangerous con- tainer Cd1 exists in a yard-bay b and the distance con- straints between dangerous containers are not hold. Thereby, for each dangerous container Cd2 allocated in the previous Dmin yard-bays is calculated by Euclidean distance to Cd1. If this distance is lower than Dmin, for any dangerous con- tainer Cd2, then h value is increased with the number of containers n on Cd1 because it indicates that removing those n containers is necessary to reallocate the container Cd1. 6. Evaluation In this section, we evaluate the behavior of the heuris- tic with the set of optimization criteria presented in this paper. The experiments were performed on random in- stances. A random instance of a yard-bay is characterized by the tuple < n, s >, where n is the number of containers in a yard-bay and s is the number of selected containers in the yard-bay. Each instance is a random configuration of all containers distributed along six stacks with 4 tiers. They are solved on a personal computer equipped with a Core 2 Quad Q9950 2.84Ghz with 3.25Gb RAM. First, we present a comparison between our basic do- main dependent heuristic h1 against a domain independent one (Metric FF ). Thus, Table 1 presents the average run- ning time (in milliseconds) to achieve a first solution as well as the best solution found (number of reshuffles) in 10 9 Algorithm 9: Pseudo-code to avoid locating two dangerous containers closer to a distance Dmin Data: B: whole block; b: state of the actual yard-bay; h: Initial heuristic; Dmin: Minimum distance; Result: h: heuristic value of b; nBay ← NumBay(b); if nBay > 1 ∧∃Cd1 ∈ b then ∆h ← 0; L1 ← Location(Cd1); foreach bl ∈ Y / NumBay(bl) ∈{max(nBay−Dmin + 1, 1), nBay−1} do if ∃Cd2 ∈ bl then L2 ← Location(Cd2); dist ← EuclideanDistance(L1, L2); if dist < Dmin then ∆h ← ∆h + NumContainersOn(Cd1); if Clear(Cd1) ∈ b then ∆h ← ∆h + (Dmin − dist); end end end end h ← h + ∆h; end Algorithm 10: Sinks within a whole block Data: B: whole block; Result: nSinks: number of Sinks; nSinks ← 0; for b ← 1 to numYards(B) do for r ← 1 to numRows(b) − 1 do ∆h ← Abs(Height[r, b] − Height[r + 1, b]); if ∆h > 2 then nSinks ← nSinks + 1; end end if NumBay(b) > 1 then for r ← 1 to numRows(b) do ∆h ← Abs(Height[r, b] − Height[r, b − 1]); if ∆h > 2 then nSinks ← nSinks + 1; end end end end Algorithm 11: Unfeasible relationships between two dangerous containers within a whole block Data: B: whole block; Result: nDang: number of Sinks; nDang ← 0; for b ← 1 to numYards(B) do nBay ← NumBay(b); if nBay > 1 ∧∃Cd1 ∈ b then L1 ← Location(Cd1); foreach bl ∈ Y / NumBay(bl) ∈ {max(nBay − Dmin + 1, 1), nBay − 1} do if ∃Cd2 ∈ bl then L2 ← Location(Cd2); dist ← EuclideanDistance(L1, L2); if dist < Dmin then nDang ← nDang + 1; end end end end end seconds for our domain-dependent planner and the aver- age running time (in milliseconds) and the quality of the solution for Metric FF. Both planners have been tested in problems < n, 4 > evaluating 100 test cases for each one. Thus, we fixed the number of selected containers to 4 and we increased the number of containers n from 15 to 21. It can be observed that our new domain-dependent heuristic is able to find a solution in a few milliseconds, meanwhile the domain-independent planner (Metric FF ) needs much time for finding a solution and also, this solu- tion needs more moves to get a goal state. Furthermore, due to the fact that our tool is an anytime planner, we evaluate the best solution found in a given time (10 sec- onds). Table 1: Average number of reshuffles and running time of MetricF F and h1 in problems < n, 4 >. Metric FF Heuristic (h1) Instance Running Solution Time first Best Solution time solution in 10 secs < 13, 4 > 22 3.07 2 3.07 < 15, 4 > 3102 4.04 5 3.65 < 17, 4 > 4669 5.35 11 4.35 < 19, 4 > 6504 6.06 22 4.72 < 20, 4 > 22622 7.01 33 5.22 < 21, 4 > 13981 6.82 62 5.08 Now we show the effects of using each one of the criteria described in Section 4 separately. In Table 2, we present the average sum of distances between the selected contain- ers and the right side of the layout in both our domain- independent heuristic and our domain-dependent heuristic with distance optimization for problems < n, 4 >. As men- tioned above, we fixed the number of selected containers to 4 and we increased the number of containers n from 13 to 21. It can be observed that distance optimization func- tion helps finding solution plans that place the selected containers closer to the cargo side of the yard-bay. Table 2: Average distance obtained by considering distance or not in our domain-dependent heuristic < n, 4 > with 4 tiers. Instance Metric FF OC1d Distance Reshuffles Distance Reshuffles < 13, 4 > 11.28 3.07 10.91 3.07 < 15, 4 > 10.60 4.04 9.21 3.65 < 17, 4 > 10.58 5.35 8.87 4.46 < 19, 4 > 12.28 6.06 8.33 4.85 < 20, 4 > 12.71 7.01 7.75 5.55 < 21, 4 > 12.20 6.82 8.22 5.33 Applying the criterion or function showed in Algorithm 3 we obtain the results appeared in Table 3. These results are the comparison between the number of solved problems over 100 problems < n, 4 > using or not that criterion in just one second. Through this table we can conclude that: • The greater number of containers, the fewer prob- lems are solved. This is because as we increase the number of containers there are less positions or gaps where containers could be remarshalled. 10 • Allowing movements to the 5th helps us to solve more problems. It is remarkable with instances < 23, 4 > with H1 only three problems could be solved, how- ever OC1t solves 84 over 100 problems. Table 3: Number of solved problems < n, 4 > with 4 and 5 tiers during the process. Instance 4 tiers h1 5 tiers OC1t < 19, 4 > 100 100 < 20, 4 > 100 100 < 21, 4 > 95 99 < 23, 4 > 3 84 Last criterion for solving problems where we only take into account one yard-bay is showed in Section 4.3. As we mentioned in this section, since the last function (Algo- rithm 7) presents the best results after the whole process of remarshalling, we do the comparison in Table 4 among the solutions given by Metric FF planner, the initial one h1 and OC1b (Both) in 50 test cases. These results are the average of the best solutions found given a time limit of 1 second for the instances of both < 15, 4 > and < 17, 4 >. Sinks are calculated by Algorithm 10. As we men- tioned above, we consider that there is a sink where the difference in tiers between two adjacent rows is higher than 2. Thereby, in this algorithm we are counting sinks pro- duced between two contiguous stacks at the same yard- bay as well as between two rows in one yard-bay and the previous one. This process takes into account the goal containers in final yard-bays. Table 4: Average number of movements, sinks and time for the first solution in problems < 15, 4 > (1) and < 17, 4 > (2) using or not balanced heuristics. Metric FF h1 OC1b (1) (2) (1) (2) (1) (2) Reshuffles 3.72 4.24 3.42 3.72 4, 76 5.04 Sinks 0.62 0.50 0.94 0.66 0 0 Time First 2621 2961 5 9 32 44 From here we realize an evaluation for the criteria pre- sented in Section 5. Table 5 shows the performance of the criteria for solving the whole block of yard-bays. These experiments were performed in blocks of 20 yard-bays and each one of them are instances < 15, 4 >. This evaluation was carried out in a yard with 2 blocks of 20 yard-bays. Thus, the results showed in Table 5 represent the average number of reshuffles, the average number of sinks gener- ated along the block and the average number of unsatisfied dangerous containers. Results given by these optimization criteria are the average of the best solutions found in 10 seconds. The number of unfeasible relationships between dan- gerous containers is calculated by means of Algorithm 11. Basically, we look for those pairs of dangerous containers whose distance between them is shorter than minimum distance (Dmin). In this table, it can be observed that h1 still outper- forms Metric FF in the average number of reshuffles. How- ever, due to the fact that they do not take into account the balancing constraints, Metric FF generated an average of 18.00 sinks in the block of yard-bay and h1 generated and average of 29.50 sinks. And the same thing happens for the average number of unfeasible constraints for dangerous containers, Metric FF gives us 16.00 and h1 obtains 7.50. Taking into account that OCN is a junction of OCnB and OCnD, both OCnB and OCnD solved their problems, that is, OCnB obtained its solutions with no sinks and OCnD obtained its solutions by satisfying all dangerous constraints. Furthermore, OCN was able to solve its prob- lems by satisfying both types of constraints. However we could state that balancing problem is harder than the problem related to dangerous containers because OCnB needs more reshuffles to obtain a solution plan than OCnD. Moreover, we observe with OCnB, OCnD and OCN ensure the established requirements however the average reshuf- fles is increased with respect to h1. Table 5: Average results with blocks of 20 yard-bays each one being a < 15, 4 > problem. Metric FF h1 OCnB OCnD OCN Reshuffles 3.65 3.38 4.85 4.00 5.65 Sinks 18.00 29.50 0 40.33 0 Non-Safe 16.00 7.50 8.00 0 0 Dangerous 7. Conclusions This paper presents domain-dependent heuristics and a set of optimization criteria for solving the Container Stack- ing problem by means of planning techniques from Artifi- cial Intelligence. We have developed a domain-dependent planning tool for finding optimized plans to obtain an ap- propriate configuration of containers in a yard-bay. Thus, given a set of outgoing containers, our planner minimizes the number of necessary reshuffles of containers in order to allocate all selected containers at the top of the stacks. This proposed planner is able to satisfy both balancing constraints and keeping a security distance between dan- gerous containers, as well as reducing the distance of the goal containers to the cargo side or allowing a fifth tier during the remarshalling process. Additional criteria have been defined for management of blocks of yard-bays. However, as the problems involve a larger number of constraints, the solution becomes harder and the number of reshuffles increases. Due to the fact that a solution of a yard-bay influences on the solution of the following yard-bay, the order of solving the yard-bays will vary and determine the minimal number of reshuffles. 11 This proposed planner with a domain-dependent heuris- tic allows us obtaining optimized and efficient solutions. This automatic planner can help to take decisions in the port operations dealing with real problems. Moreover, it can help to simulate operations to obtain conclusions about the operation of the terminal, evaluate alternative configurations, obtain performance measures, etc. Partic- ularly, in [14] the proposed planner has been applied for obtaining an evaluation of alternative 4 or 5 tiers stacks configuration. Acknowledgment This work has been partially supported by the research projects TIN2010-20976-C02-01 (Min. de Ciencia e Inno- vación, Spain), P19/08 (Min. de Fomento, Spain-FEDER) and the VALi+d Program of the Conselleria d’Educació (Generalitat Valenciana), as well as with the collaboration of the maritime container terminal MSC (Mediterranean Shipping Company S.A.). References [1] Chen, S.-H., Chen, J.-N., 2010. Forecasting container through- puts at ports using genetic programming. Expert Systems with Applications 37 (3), 2054 – 2058. URL http://www.sciencedirect.com/science/article/ B6V03-4WNXTWY-M/2/1a5e0fe084ba3ea36303bd280acecc04 [2] Chuang, T.-N., Lin, C.-T., Kung, J.-Y., Lin, M.-D., 2010. Plan- ning the route of container ships: A fuzzy genetic approach. Expert Systems with Applications 37 (4), 2948 – 2956. URL http://www.sciencedirect.com/science/article/ B6V03-4X7YNF2-7/2/19d11092d4e05e1daabb09291c7aa78d [3] Ghallab, M., Howe, A., Knoblock, C., McDermott, D., Ram, A., Veloso, M., Weld, D., Wilkins, D., 1998. PDDL - the planning domain definition language. AIPS-98 Planning Committee. [4] Hatzi, O., Vrakas, D., Bassiliades, N., Anagnostopoulos, D., Vlahavas, I., 2010. A visual programming system for automated problem solving. Expert Systems with Applications 37 (6), 4611 – 4625. URL http://www.sciencedirect.com/science/article/ B6V03-4XX23MG-C/2/3b6b43536b6d90488631a7c780a02df1 [5] Henesey, L., 2006. Overview of Transshipment Operations and Simulation. In: MedTrade conference, Malta, April. pp. 6–7. [6] Hoffman, J., Nebel, B., 2001. The FF planning system: Fast planning generation through heuristic search. Journal of Artifi- cial Intelligence Research 14, 253–302. [7] Hoffmann, J., 2003. The metric-ff planning system: translating "ignoring delete lists" to numeric state variables. J. Artif. Int. Res. 20 (1), 291–341. [8] Kim, K., Bae, J., 1998. Re-marshaling export containers in port container terminals. Computers & Industrial Engineering 35 (3-4), 655 – 658, selected Papers from the 22nd ICC and IE Conference. URL http://www.sciencedirect.com/science/article/ B6V27-3WXX91K-2D/2/a56e762628a5a1107f0d067a9dd3af36 [9] Kim, K., Hong, G., 2006. A heuristic rule for relocating blocks. Computers & Operations Research 33 (4), 940–954. [10] Kim, K., Park, Y., Ryu, K., 2000. Deriving decision rules to locate export containers in container yards. European Journal of Operational Research 124, 89–101. [11] Kim, K. H., 1997. Evaluation of the number of rehandles in container yards. Computers & Industrial Engineering 32 (4), 701 – 711, new Advances in Analysis of Manufacturing Sys- tems. URL http://www.sciencedirect.com/science/article/ B6V27-3SN7946-2/2/4cf015e963d2d47389862ee31e485b58 [12] Lee, Y., Hsu, N.-Y., 2007. An optimization model for the container pre-marshalling problem. Computers & Operations Research 34 (11), 3295 – 3313. URL http://www.sciencedirect.com/science/article/ B6VC5-4JBGKD8-1/2/2a8c21229ed83e829dd4ef7bdc1fab48 [13] Park, K., Park, T., Ryu, K., 2009. Planning for remarshaling in an automated container terminal using cooperative coevolution- ary algorithms. In: Proceedings of the 2009 ACM symposium on Applied Computing. ACM, pp. 1098–1105. [14] Salido, M., Sapena, O., Barber, F., 2009. What is better: 4 tiers or 5 tiers in the container stacking problem? The In- ternational Workshop on Harbour, Maritime and Multimodal Logistics Modelling and Simulation. [15] Salido, M., Sapena, O., Rodriguez, M., Barber, F., 2009. A planning tool for minimizing reshuffles in containers terminals. ICTAI 2009: 21st International Conference on Tools with Arti- ficial Intelligence. [16] Sapena, O., Onaindía, E., 2002. Domain independent on-line planning for strips domains. In proc. IBERAMIA-02, 2527, 825– 834. [17] Stahlbock, R., Voß, S., 2008. Operations research at container terminals: a literature update. OR Spectrum 30 (1), 1–52. [18] Vis, I., De Koster, R., 2003. Transshipment of containers at a container terminal: an overview. European Journal of Opera- tional Research 147, 1–16. [19] Winograd, T., 1971. Procedures as a representation for data in a computer program for understanding natural language. MIT. Cent. Space Res., Cambridge, MA. 12 http://www.sciencedirect.com/science/article/B6V03-4WNXTWY-M/2/1a5e0fe084ba3ea36303bd280acecc04 http://www.sciencedirect.com/science/article/B6V03-4WNXTWY-M/2/1a5e0fe084ba3ea36303bd280acecc04 http://www.sciencedirect.com/science/article/B6V03-4X7YNF2-7/2/19d11092d4e05e1daabb09291c7aa78d http://www.sciencedirect.com/science/article/B6V03-4X7YNF2-7/2/19d11092d4e05e1daabb09291c7aa78d http://www.sciencedirect.com/science/article/B6V03-4XX23MG-C/2/3b6b43536b6d90488631a7c780a02df1 http://www.sciencedirect.com/science/article/B6V03-4XX23MG-C/2/3b6b43536b6d90488631a7c780a02df1 http://www.sciencedirect.com/science/article/B6V27-3WXX91K-2D/2/a56e762628a5a1107f0d067a9dd3af36 http://www.sciencedirect.com/science/article/B6V27-3WXX91K-2D/2/a56e762628a5a1107f0d067a9dd3af36 http://www.sciencedirect.com/science/article/B6V27-3SN7946-2/2/4cf015e963d2d47389862ee31e485b58 http://www.sciencedirect.com/science/article/B6V27-3SN7946-2/2/4cf015e963d2d47389862ee31e485b58 http://www.sciencedirect.com/science/article/B6VC5-4JBGKD8-1/2/2a8c21229ed83e829dd4ef7bdc1fab48 http://www.sciencedirect.com/science/article/B6VC5-4JBGKD8-1/2/2a8c21229ed83e829dd4ef7bdc1fab48 
work_5fgifeevezgirlwhykdwka2s4a	----	Parameterizing neural networks for disease classification Received: 17 September 2018 Revised: 22 March 2019 Accepted: 13 July 2019 DOI: 10.1111/ exsy.12465 S P E C I A L I S S U E P A P E R Parameterizing neural networks for disease classification Guryash Bahra1 Lena Wiese2 1 Institute of Computer Science, University of Göttingen, Göttingen, Germany 2 L3S Research Center/ Knowledge Based Systems Group, Leibniz University Hannover, Hannover, Germany Correspondence Lena Wiese, L3S Research Center/ Knowledge Based Systems Group, Leibniz University Hannover, Appelstraße 4, Hannover 30167, Germany. Email: wiese@l3s.de Abstract Neural networks are one option to implement decision support systems for health care applications. In this paper, we identify optimal settings of neural networks for medical diagnoses: The study involves the application of supervised machine learning using an artificial neural network to distinguish between gout and leukaemia patients. With the objective to improve the base accuracy (calculated from the initial set-up of the neural network model), several enhancements are analysed, such as the use of hyperbolic tangent activation function instead of the sigmoid function, the use of two hidden layers instead of one, and transforming the measurements with linear regression to obtain a smoothened data set. Another setting we study is the impact on the accuracy when using a data set of reduced size but with higher data quality. We also discuss the tradeoff between accuracy and runtime efficiency. K E Y W O R D S artifical neural network, disease classification, MIMIC-III, supervised machine learning 1 I N T R O D U C T I O N Future precision medicine can vastly profit from data analytics and machine learning to obtain a patient-specific personalized diagnosis based on various individual health indicators. Based on a body of prior experience documented in electronic medical records of other patients, machine learning techniques can form the underpinning of medical decision support systems, ideally relying on an integration of both the capabilities of storage and data analytics as for example in Tashkandi, Wiese, and Wiese (2018). An artificial neural network (ANN) is an established method to perform classification (one of the supervised learning techniques) on clinical data. In medical data analysis, machine learning is a common procedure used for classification of patients suffering from different diseases. In this paper, an ANN is used to predict and classify the diagnoses of either gout or leukaemia based on uric acid measurements. However, we address the topic from a more technical perspective to find out which enhanced settings (in terms of data preprocessing and neural network configuration) provide the most benefit. 1.1 Medical background We concentrate on the differentiation of two diseases that both can be identified by measuring the concentration of uric acid. In particular, one of the common factors in leukaemia and gout diseases is the abnormal uric acid signature in blood. Uric acid concentration in a healthy human in developed countries ranges from 3.5 mg/ dl in infants to about 6 mg/ dl in adults (Alvarez-Lario & Macarron-Vicente, 2011; Lasko et al., 2013; Wilcox, 1996). In patients suffering from gout or leukaemia, the uric acid concentration increases more than the normal range and is therefore regularly monitored and treated. In case of gout, a combination of genetic mutations and environmental factors causes uric acid concentration to increase. This further results in formation of uric acid crystals, which precipitates into joints and causes painful arthritis (Lasko et al., 2013). As for leukaemia, turnover of white blood cells increases the uric acid concentration. The two diseases have different pathophysiology, which in turn combined with their treatment procedures, resulting in different signatures of uric acid concentrations. Therefore, in this article, supervised learning is performed on uric acid measurements to classify the two diseases leukaemia and gout. In this way, we can derive an automated recommendation for medical staff regarding the likeliness of one disease or the other. This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes. © 2019 The Authors. Expert Systems published by John Wiley & Sons Ltd Expert Systems. 2019;e12465. wileyonlinelibrary.com/ journal/ exsy 1 of 14 https:/ / doi.org/ 10.1111/ exsy.12465 https://doi.org/10.1111/exsy.12465 https://orcid.org/0000-0003-3515-9209 2 of 14 BAHRA AND WIESE 1.2 Machine learning techniques Machine learning is one of the top growing fields of recent times and is being utilized in various areas for many applications, such as financial trading, health care, recommendations, online search, and many more. Machine learning involves building of models from the given data set, which can be utilized to make future predictions. This process is executed in two phases: (a) From the input data set, calculate unknown dependencies, and (b) from those dependencies, predict outcomes on so far unseen data sets. The two common types of machine learning are supervised and unsupervised learning. Supervised learning involves use of a labelled set of input data to predict the output. In contrast, unsupervised learning involves unlabelled data. Because of this, there is no designated output such that the learning model has to identify patterns in the input data. The work in this article addresses the supervised learning problem of classification, which involves categorizing the data into finite classes. More precisely, uric acid concentrations are classified into the two classes: leukaemia and gout. 1.3 Objectives The main objective of this article is to perform supervised machine learning with neural networks and assess the accuracy of the system to distinguish between the patients with either gout or leukaemia. We verify and extend our previous preliminary analysis in Bahra and Wiese (2018) by considering several advanced settings of the neural network; we also analyse the tradeoffs between these different settings of the neural network model. To achieve this objective, several tasks have to be performed, and these are summarized as follows: 1. Identify files to be used from the Medical Information Mart for Intensive Care (MIMIC) data set provided by Goldberger et al. (2000) and Johnson et al. (2016). 2. Identify data (patients with either gout or leukaemia and their uric acid signatures) required for the study. 3. Perform preprocessing on the identified data (uric acid concentrations) with linear regression. 4. Perform supervised learning with threefolds cross validation on original and linear regression transformed uric acid measurements in different settings. 5. Improve the data quality by reducing the data set to patients with at least three uric acid measurements and observing the impact on the accuracy. 6. Assess the impact on the runtime efficiency of these enhancements. In more detail, our study starts with identifying the tables from the MIMIC-III database. From those tables, patients with gout and leukaemia diseases and their corresponding uric acid measurements are identified. The data are then cleansed and are further used for supervised learning. The neural network for the supervised learning is designed using one hidden layer and sigmoid activation function. Accuracy is then calculated to measure the effectiveness of the model. Furthermore, to improve the initial accuracy of the model, a couple of enhancements are tried. These enhancements include the use of the hyperbolic tangent (tanh) activation function, two hidden layers, and linear regression for transformation of the original uric acid measurements. Lastly, the size of the data set is reduced in order to avoid a bias that results from singular uric acid values in the measurement sequence of a patient. It is achieved by removing the sequences with less than three measurements. 1.4 Outline of the article After discussing the background and objectives of the article in Section 1, Section 2 surveys related work in the area. Section 3 covers the detailed description of the MIMIC database and includes tables identification and data set selection and creation. Section 4 covers the supervised learning performed on our use case, introducing the basics of neural networks along with their implementation. Results and observations of supervised learning are then discussed in Section 5; this section also covers various enhanced techniques that include the use of normalization, hyperbolic tangent function as activation function, two hidden layers instead of one, and linear regression for data preprocessing or transformation. Results with a reduced data set size are presented in Section 6. The runtime behaviour of our settings is discussed in Section 7. Section 8 summarizes our work. 2 R E L A T E D W O R K Several studies in health care are based on machine learning. In this section, we survey some applications of machine learning in health care in the last two decades. Lee, Liao, and Embrechts (2000) analysed heart disease databases with techniques like data visualization and correlation analysis to identify important features in heart disease and to build a multivariate relationship model to visualize the relationship between any two features. They describe two different approaches, discriminant analysis, and neural networks classification to identify high-risk patients. García-Gómez et al. (2004) study classification of soft tissue tumors into either benign or malignant classes based on a pattern-recognition approach. They use MR images as input data and perform machine learning with several approaches—ANN, support vector machine (SVM), and k-nearest neighbour—for the classification task. They conclude that neural networks give relatively good results compared with the other two techniques. BAHRA AND WIESE 3 of 14 Joshi, Pakhomov, Pedersen, and Chute (2006) perform machine learning on electronic medical records to unambiguously expand acronyms in clinical reports. The learning is carried out with three different algorithms, Naïve Bayes, decision tree, and SVM. They observe that accuracy of the model is better with all the features combined as compared with when features are used individually and in pairs, independent of the algorithm used. Huang, McCullagh, Black, and Harper (2007) perform supervised learning on diabetes data to classify patients into diabetic and nondiabetic classes. The study also involves identification of features that give a good accuracy for the model. They used several learning algorithms—Naïve Bayes, IB1 (an instance-based learning algorithm), and C4.5 (a decision tree algorithm)—to achieve the objective. Nguyen, Moore, McCowan, and Courage (2007) use SVM to classify lung cancer patients into multiple classes of T (the tumor stage that describes size and position of the tumor) and N (the node stage that describes presence of spread into the lymph nodes) stages of lung cancer. Juhola (2008) classifies otoneurological data into six classes: vestibular schwannoma, benign positional vertigo, Menière's disease, sudden deafness, traumatic vertigo, and vestibular neuritis. Seven techniques—k-nearest neighbour searching, discriminant analysis, Naïve Bayesian decision rule, k-means clustering, decision trees, multilayer perceptron neural networks, and Kohonen networks—are employed for this task. It is observed that linear discriminant analysis performed better than the other six techniques. Maes, Twisk, and Johnson (2012) apply supervised learning methods, such as linear discriminant analysis, pattern recognition, and factor analysis, to differentiate between myalgic encephalomyelitis, chronic fatigue syndrome, and chronic fatigue. Lasko, Denny, and Levy (2013) use an unsupervised learning method to identify features in gout and leukaemia diseases with the use of the uric acid signatures. The work described in their paper is implemented on data extracted from electronic medical records Roden et al. (2008). In their paper, it is mentioned that the data are noisy, sparse, and irregular; therefore, it is smoothened with the use of Gaussian process regression. On this data, deep learning is performed with the use of sparse autoencoders. Furthermore, the features learned from first and second layers of sparse autoencoders are then utilized for a supervised learning classification task (to classify between gout and leukaemia) using logistic regression. Shouval et al. (2014) survey different supervised learning algorithms, namely, ANN, decision tree, and SVM, on a haematopoietic stem cell transplantation database for the classification task. Kourou, Exarchos, Exarchos, Karamouzis, and Fotiadis (2015) review supervised learning methods, namely, ANN, Bayesian Networks, SVM, and decision trees for prognosis of cancer and its prediction. Their paper even highlights the case studies used for machine learning tools to predict cancer susceptibility, cancer recurrence, and cancer survival. Beaulieu-Jones and Greene (2016) develop a semisupervised learning method and use it to improve survival predictions of patients suffering from ALS. Semisupervised learning results from a combination of supervised and unsupervised learning. The method for semisupervised learning is developed using denoising autoencoders (an unsupervised learning approach) for phenotype stratification, in combination with random forests (as supervised learning) for classification. In their paper, unsupervised learning is performed to reduce the amount of features of the input data set, which is then followed by performing supervised learning on this reduced data set. Weng, Reps, Kai,Garibaldi, and Qureshi (2017) use supervised learning for cardiovascular risk prediction. To achieve this, the following algorithms are used: random forest, logistic regression, gradient boosting machines, and neural networks. It is concluded in their paper that neural networks perform better than the rest. It can be seen from this survey that classification of diseases is a popular use case and ANN is a widely used technique for this classification task. In contrast to the surveyed approaches, in this paper, we apply supervised learning on a data set different from the ones used before. In addition, we study the effect of several enhancements (in particular of linear regression for data smoothening). Moreover, we assess the runtime behaviour of the different settings. 3 D A T A S E T The MIMIC-III is the third iteration of a large clinical database called MIMIC as reported by Goldberger et al. (2000) and Johnson et al. (2016). It comprises the medical data of patients admitted to critical care units, Coronary Care Unit, Cardiac Surgery Recovery Unit, Medical Intensive Care Unit (ICU), Neonatal ICU, Surgical ICU, and Trauma Surgical ICU, at the hospital Beth Israel Deaconess Medical Center in Boston. According to Goldberger et al. (2000), the current version of the database is 1.4 as of September 4, 2016, and consists of health-related records of deidentified 46,520 subjects out of which 38,645 are adults, and 7,875 are neonates. The patients are deidentified according to Health Insurance Portability and Accountability Act (HIPAA) standards that involved removal of 18 fields, such as patients' name, telephone numbers, addresses, and so on, as listed in HIPAA. It also involved shifting of the dates, including date of birth, by a random offset, as directed by the HIPAA. The database not only includes the information of the vital sign measurements, medicines administered, laboratory measurements, fluid balance, imaging reports, and out-of-hospital mortality but also includes patients' demographics, nurses' and physicians' notes, procedure and diagnostic codes, and more. Note that the MIMIC database was prepared by compiling data from two data sources, CareVue and Metavision ICU databases, used at the hospital. 3.1 Data set identification There are 26 comma-separated values files in the MIMIC-III data set. Out of those 26 files, the following four files are considered for our case study: 4 of 14 BAHRA AND WIESE TA B L E 1 S e le ct io n o f h o sp it a la d m is si o n s fo r a n a ly si s A d m is si o n s N u m b e r o f u n iq u e a d m is si o n s N u m b e r o f a d m is si o n s le u k a e m ia N u m b e r o f a d m is si o n s g o u t In it ia la d m is si o n s 2 ,8 3 7 6 1 8 2 ,2 1 9 A d m is si o n s a ft e r e lim in a ti n g p a ti e n ts w it h b o th d ia g n o se s 2 ,7 7 3 5 8 4 2 ,1 8 9 A d m is si o n s a ft e r e lim in a ti n g p a ti e n ts w it h n o u ri c a ci d m e a su re m e n ts 1 ,0 7 6 3 9 8 6 7 8 F in a la d m is si o n s a ft e r e lim in a ti n g a d m is si o n s w it h n o u ri c a ci d m e a su re m e n ts 6 4 0 3 0 6 3 3 4 BAHRA AND WIESE 5 of 14 TA B L E 2 S e le ct io n o f p a ti e n ts fo r a n a ly si s P a ti e n ts N u m b e r o f u n iq u e p a ti e n ts N u m b e r o f p a ti e n ts le u k a e m ia N u m b e r o f p a ti e n ts g o u t In it ia lo b se rv a ti o n s 2 ,2 5 9 4 5 4 1 ,8 0 5 O b se rv a ti o n s a ft e r e lim in a ti n g p a ti e n ts w it h b o th d ia g n o se s 2 ,2 1 5 4 3 2 1 ,7 8 3 O b se rv a ti o n s a ft e r e lim in a ti n g p a ti e n ts w it h n o u ri c a ci d m e a su re m e n ts 7 6 1 2 7 3 4 8 8 F in a lp a ti e n ts a ft e r e lim in a ti n g a d m is si o n s w it h n o u ri c a ci d m e a su re m e n ts 5 6 7 2 5 6 3 1 1 6 of 14 BAHRA AND WIESE TABLE 3 Selection of uric acid measurements for analysis Measurements Number of uric acid measurements Initial observations 19,906 Observations after eliminating patients with different diagnosis 7,076 Final observations corresponding to final admission IDs 5,665 • D_ICD_DIAGNOSES gives the disease codes (ICD-9 codes) for gout and leukaemia diagnoses. • DIAGNOSES_ICD identifies the hospital admissions and patients suffering from gout and leukaemia. • D_LABITEMS gives the ID for uric acid signatures. • LABEVENTS gives the data about uric acid measurements of the patients identified with gout and leukaemia. There are 78 ICD-9 codes for leukaemia and 11 for gout identified from the D_ICD_DIAGNOSES table. Then, from the DIAGNOSES_ICD table, a total of 2,837 hospital admissions are identified with the above diagnoses, out of which 618 hospital admissions are for leukaemia and 2,219 are for gout. These many admissions correspond to 2,259 patients or unique SUBJECT_IDs; in R, the duplicated function with logical negation operator (!) was used to find these unique SUBJECT_IDs. Among these subject IDs, 454 patients suffered from leukaemia, and 1,805 suffered from gout. Furthermore, 22 patients are common for both the diagnoses. After removing IDs of those patients, a total of 2,773 hospital admissions were identified for the patients suffering either leukaemia (584 HADM_IDs) or gout (2,189 HADM_IDs) but not both. These 2,773 admissions correspond to 2,215 patients, out of which 1,783 patients suffered from gout and 432 from leukaemia. Moreover, the hospital admissions are reduced from 2,773 to 1,076 by removing records with no uric acid measurements for those patients. The number of admissions further decreased to 640 as there are no uric acid observations corresponding to those admissions. And finally, these 640 unique admissions correspond to 567 unique patients, out of which 311 suffered from gout and the remaining 256 patients suffered from leukaemia. Tables 1 and 2 summarize the above information. As for the uric acid signatures, 3 IDs were identified from the D_LABITEMS table. Corresponding to those IDs, 19,906 observations representing uric acid measurements are identified from the LABEVENTS table. These 19,906 observations reduced to 7,076, as the removed observations did not correspond to the identified SUBJECT_IDs. Furthermore, 7,076 observations reduced to 5,665, as those observations did not correspond to the identified hospital admission IDs. The Table 3 summarizes the above information. 3.2 Data set creation The data, that is, uric acid concentrations (from Table 3), are then arranged into 567 sequences, grouped according to the patient IDs (from Table 2). These sequences are then broken down to a size of 17 values per row, where the first two values are the label (1 for leukaemia and 0 for gout) and patient ID, and the remaining 15 values are the uric acid concentrations. Note that the sizes of measurement sequences are unequal. In order to make use of all measurements in the data set, patients with more than 15 measurements occur multiple times: each sequence of 15 consecutive measurements is treated as a new sequence. On the other hand, for sequences with less than 15 measurements, value 0 is used for the remaining part of the sequence. This resulted in a total of 813 sequences. The sequences are then shuffled with the use of sample function provided by R Team (2014). 4 M E T H O D 4.1 Neural networks Neural networks are one class of several supervised learning classification techniques (as introduced in Section 1.2) and are based on the concept of perceptrons, which was originally introduced by Rosenblatt (1958) and is now widely discussed in standard textbooks and surveys like Kotsiantis (2007) and Nielsen (2015). In this work, a three-layered neural network is used, where the first layer, L1 , is the input layer, the second layer, L2 , is the hidden layer, and the last layer, L3 , is the output layer. Note that the output of one layer is the input of the next layer, and there are sl number of nodes in layer l. Figure 1 illustrates these settings. 4.1.1 Defining the layer sizes To begin, the number of nodes sl in all the layers are defined. According to our data set (from Section 3.2), the input size is 15, that is, the number of nodes (s1 ) in Layer 1 (L1 ) is 15. This is because, from Section 3.2, out of 17 values per row, 15 values are the uric acid measurements. The number of output nodes is 1, as the label (leukaemia or gout) per row is single valued (from Section 3.2). The number of nodes in the hidden layer s2 is varied in the upcoming tests in order to assess the changes (positive, negative, or no change) in the result. We chose to test s2 = 5, s2 = 10, and s2 = 25. 4.1.2 Forward propagation As the nodes are calculated starting from the Layer L1 up to Layer L3 in the network, this step is called forward propagation. The activation unit, a(l) i , is used to define the output of the ith unit in layer l. Therefore, for the L1 layer, a (1) i = xi consists of one measurement value from an input BAHRA AND WIESE 7 of 14 FIGURE 1 Our neural network settings: weight matrix W and bias b are recalibrated in the learning step; weight decay parameter 𝜆 and the amount of nodes in Layer 2 s2 are varied in our tests sequence; as for the Layers L2 and L3 , the nodes are computational, and therefore, activations are calculated as a function of an input vector a (l) (the activations of the previous layer), a weights matrix W and a bias vector b: 1. For Layer l = 2: a(2) = f ( W(1) ∗ a(1) ) + b(1), note that a(1) is an input data sequence and f is the sigmoid function (see below). 2. For Layer l = 3: a(3) = f(W(2) ∗ a(2)) + b(2), this is the final output of the neural network denoting the classification into either gout or leukaemia. Here, W(l−1) ij is the weight or the parameter of the connection between the jth unit in Layer l − 1 and the ith unit in Layer l, bias unit bl i corresponds to the ith unit in Layer l. 4.1.3 Initialization of weights W and biases b and activation function The weights in matrix W are to be initialized to a value close to 0 and therefore are randomly initialized from the interval [−0.5, 0.5] (as in Nguyen & Widrow, 1990). The biases b are initialized to 0. Function f ∶ R → R is the activation function and is usually defined with sigmoid function as shown by Equation (1); it produces an output ranging over (0, 1). f(z) = 1 1 + exp(−z) . (1) An important identity to note is that the derivative f′(z) of sigmoid function f(z) (in Equation 1) is given by Equation (2) and is used in the learning phase. f′(z) = f(z)(1 − f(z)). (2) In order to adapt the bias vector b and the weight matrix W, we apply the squared-error cost function and minimize its value by backpropagation in the neural network. 4.1.4 Cost function For a single training example (x, y), where x is one input sequnence of 15 uric acid measurements and y is the class label for either gout or leukaemia, the squared-error cost function is given in Equation (3). J(W, b; x, y) = 1 2 ||hW,b(x) − y||2, (3) where, hW,b(x) = a (3) 1 is the final activation of the neural network; the size of hW,b(x) is 1 × 1, as there is one output node and one training example. More generally, for a given training set { (x(1), y(1)), .., (x(r), y(r)), .., (x(m), y(m)) } of m training examples, the cost function is given as Equation (4), J(W, b) = [ 1 m m∑ r=1 J ( W, b; x(r), y(r) )] + 𝜆 2 nl−1∑ l=1 sl∑ i=1 sl+1∑ j=1 ( W(l) ij )2 , (4) where the first term is the sum of squares error term averaged over all input sequnences and the second term is the regularization term (or the weight decay term) that decreases the value of the weights and avoids overfitting. Parameter 𝜆 in Equation (4) is the weight decay parameter. To minimize the cost function J(W, b) as a function of weights matrix W and bias vector b, every parameter W(l) ij and b(l) i is initialized to a random value near to 0. Then, an optimization algorithm is applied for minimization. We chose the limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) method introduced by Liu and Nocedal (1989) for the minimization of the cost function because it usually faster than a basic gradient descent algorithm. The L-BFGS algorithm is employed with the use of function minFunc (see Schmidt, 2005). 8 of 14 BAHRA AND WIESE 4.1.5 Define 𝜆, and update parameters W and b The weight decay parameter 𝜆 (in Equation 4) needs to be defined. In the upcoming tests, the value of weight decay parameter 𝜆 is varied in order to assess the changes (positive, negative, or no change) in the result. We chose to compare three different settings for 𝜆, namely, 𝜆 = 0.0001, 𝜆 = 0.00001, and 𝜆 = 0.000001. The weights matrix W and the bias vector b for Layers L1 and L2 , which minimize the cost function J(W, b), are recalibrated after every iteration of the L-BFGS algorithm. In other words, W and b are equal to final values of the partial derivatives of the overall cost function, 𝜕 𝜕W(l) J(W, b) and 𝜕 𝜕b(l) J(W, b), respectively. 4.2 Accuracy calculation To calculate the accuracy of the machine learning model on the training and the testing sets, forward propagation (as described in Section 4.1.2) is performed such that for Layers l = 2, 3: a(l) = f(W(l−1) ∗ a(l−1)) + b(l−1). Note that the activation matrix of Layer l = 1, a(1) = data (either training or testing), f is the sigmoid activation function, and W and b are calculated using the L-BFGS algorithm. Then, the activation vector of the output layer, a(3), is used to assign labels to the data (either training set or testing set). If the value of each element in a(3) is greater or equal to 0.5, then Label 1 (corresponding to leukaemia) is assigned to the element, else 0 (corresponding to gout) is assigned. These assigned labels then form the prediction vector of size 542 × 1 in case of training set and of 271 × 1 in case of testing set. Each element in prediction vector is then compared with the corresponding actual label of the data. If the labels are the same, 1 is assigned to a comparison vector; if labels are not the same, then 0 is assigned to the comparison vector. Furthermore, the average is calculated for the comparison vector, which is of size 542 × 1 in case of training set and of 271 × 1 in case of testing set. The average multiplied with 100 gives the accuracy of the model in percent. 4.3 K-fold cross validation To create training and testing sets used for the learning, the k-fold cross-validation method is employed. In k-fold cross validation, the data are randomly divided into k subsets of equal size, and a single subset is referred to as fold. Of the k folds, k − 1 folds are combined to form the training set and the remaining fold is used as the testing set, and the accuracy is calculated for the training and the testing sets, which describes the stability of the model. This is then repeated for k iterations, and for every iteration, the testing set comprises a fold, used exactly once. Moreover, to use the model for new predictions and to estimate the overall accuracy of the model, consider the classifier for which the highest accuracy is achieved for the testing set. We applied threefold cross validation as well as 10-fold cross validation. As described in the previous paragraph, first, the data are divided into equal subsets. Therefore, 813 sequences (from Section 3.2) are divided into three equal subsets as well as 10 subsets, respectively. Then, supervised learning is performed on these subsets for three iterations—each iteration either executing threefold or 10-fold cross validation. For each iteration, the testing set is formed with a single subset used exactly once and the remaining subsets are used as the training set. The accuracy (calculated as in Section 4.2) is reported as the average of all iterations. 5 R E S U L T S O N C O M P L E T E D A T A S E T This section describes the results of supervised learning. The results represent the neural network model's ability to distinguish between patients suffering from either gout or leukaemia based on abnormal uric acid measurements. The accuracies are determined for different cases, resulting from the change in the values of weight decay parameter 𝜆 and the number of hidden layer nodes s2 (as in Section 4.1.1). The accuracy is computed three times (iteration I1-I3) per case; this is because weights in W are randomly initialized (see Section 4.1.3) and therefore give slightly different values for each iteration. For final accuracy of the case, these accuracies are averaged out. Accuracy is measured in percent. 5.1 Results of cross validation on original data set The results in Figure 2 are of threefold cross validation performed on the original data set with one hidden layer. The following observations can be seen from Figure 2. All average test accuracies are very similar. The highest average test accuracy is 83% in case of five neurons in the hidden layer—that is, with the lowest number s2 of nodes in the hidden layer—and the lowest setting for the weight decay parameter 𝜆. However, the lower weight decay parameter incurs a much more increased runtime. The results of 10-fold cross validation in Figure 3 present a similar picture with the highest test accuracy of roughly 83% in case of 10 neurons in the hidden layer and the lowest setting for the weight decay parameter 𝜆. We tested different settings with two hidden layers. In the first case, the first hidden layer was fixed to Size 5, and only the size of the second hidden layer is increased; in the second case, both the first and second hidden layer sizes are increased. In the first case in Figure 4, adding the second layer reduced the accuracy for the largest weight decay parameter (𝜆 = 0.01). For the other two weight decay parameters, the accuracy BAHRA AND WIESE 9 of 14 FIGURE 2 Accuracies (in percent) and runtime (in seconds) of supervised learning with threefold cross validation on original data set using neural network with one hidden layer FIGURE 3 Accuracies (in percent) and runtime (in seconds) of supervised learning with 10-fold cross validation on original data set using neural network with one hidden layer remained in the same range as with one hidden layer. The second case in Figure 5, with both layer sizes increased, also shows a decrease in accuracy for the case of the largest weight decay parameter and low sizes of the first layer. When trained with 10-fold cross validation, due to the larger training set size, the decrease in accuracy did not occur (full results are not shown here due to space restrictions). In other words, the results of 10-fold cross validation with two hidden layers are comparable with Figure 3. 5.2 Preprocessing of the data set using linear regression The time-series data present in MIMIC-III are incomplete, inconsistent, sparse, and noisy. In order to regularize the data, we applied linear regression to condition the data so as to handle irregularities in the data. We expected to smoothen the data by linear regression to obtain a general trend of each of the uric acid signatures. Linear regression tries to model the relationship between input and output variables by fitting a linear equation to the input (or observed) data. To perform linear regression to smoothen the data, the lm1 function provided by R, is used. The lm function is called for a single patient ID at a time. The result of linear regression for two different patient IDs (or sequences) can be seen in Figure 6. 1 Syntax: lm(x ∼ y). In this work, x is uric acid concentrations, and y corresponds to time measurements. 10 of 14 BAHRA AND WIESE FIGURE 4 Accuracies (in percent) and runtime (in seconds) of supervised learning with threefold cross validation on original data set using neural network with two hidden layers and the first layer size fixed to 5 FIGURE 5 Accuracies (in percent) and runtime (in seconds) of supervised learning with threefold cross validation on original data set using neural network with two hidden layers and varying both layer sizes FIGURE 6 Illustration of Linear regression transformation for two different sequences BAHRA AND WIESE 11 of 14 FIGURE 7 Accuracies (in percent) and runtime (in seconds) of supervised learning with threefold cross validation on data set with linear regression and using neural network with one hidden layer FIGURE 8 Accuracies (in percent) and runtime (in seconds) of supervised learning with threefold cross validation on data set with linear regression and using neural network with two hidden layers The results presented in the Figure 7 are of threefold cross validation performed on on the data set transformed with linear regression before supervised learning with one hidden layer. Interestingly, for the high weight decay parameter (𝜆 = 0.01), the linear regression of the data set increased the accuracy to 83.3%). In the other cases, the linear regression did not show any effect on the accuracy. The same observations were made with 10-fold cross validation. With two hidden layers, the observations with linear regression resemble the observation without linear regression: For the largest weight decay parameter, the accuracy for low layer sizes is decreased when adding a second layer (shown in Figure 8). When fixing the first layer size to 5, the accuracy decreased even more in this case. In all other cases, that is, for different layer sizes as well as for all 10-fold cross validation cases, the linear regression with two hidden layers did not show any effect on the accuracy (results not shown here due to space restrictions). 6 R E D U C E D D A T A S E T S In this section, sequences with less than three nonzero data points are removed in order to remove biases in the result due to too short uric acid measurement sequences. In other words, threefold cross validation is carried out on the data of the patients, which have more than two nonzero data points (uric acid measurements) per sequence. Therefore, the number of sequences reduced to 462 (out of 813 from Section 3.2). Subsequently, the size of a single fold for cross validation was reduced accordingly, too. As in Section 5, the accuracies are determined for different settings; the accuracy is computed three times per setting, and for the final accuracy, these accuracies are averaged out. 12 of 14 BAHRA AND WIESE 6.1 Cross validation on reduced original data set The results presented here are of cross validation performed on the reduced original data set. The following observations can be seen from Figure 9 for one hidden layer with threefold cross validation. The accuracy increased when using the higher quality data set to up to 88.4% in the case of medium weight decay parameter (𝜆 = 0.001). A similar improvement was observed with 10-fold cross validation (not shown here due to space restrictions). In the case of two hidden layers, these improvements also manifest for the case that both layer sizes are increased. Our main observation is that accuracies are overall better for the reduced data set (as compared with the original one). 6.2 Cross validation on reduced transformed data set We next tested the case of one hidden layer for the data set that is both reduced to retain only sequences of length at least three as well as transformed by linear regression. The results are shown in Figure 10. The accuracy increased to up to 89.1% for the case of 𝜆 = 0.01 with hidden layer size 20 as well as 25. This is the best setting that could be obtained; in all other cases (two hidden layers or 10-fold cross validation), this improvement could not be achieved. The overall observation is hence that the best result can be achieved with the reduced and transformed data set lowest setting for the weight decay parameter 𝜆. Transformation with linear regression in this case pays off with improved accuracies (as compared with the nontransformed case). In particular, again we can observe that accuracies are overall better for the reduced data set (as compared with the original one). FIGURE 9 Accuracies (in percent) and runtime (in seconds) of supervised learning with threefold cross validation on reduced data set using neural network with one hidden layer FIGURE 10 Accuracies (in percent) and runtime (in seconds) of supervised learning with threefold cross validation on reduced data set with linear regression and using neural network with one hidden layer BAHRA AND WIESE 13 of 14 7 R U N T I M E C O M P A R I S O N We implemented the steps to carry out the supervised learning presented in the previous sections in Octave (Eaton, 2019). Data preprocessing was done in R. We measured the runtime of the neural network model learning phases for all settings for which we reported the accuracies in the previous sections. Executions are run on a Ubuntu 16.04.2 LTS system with 8-GB RAM and 1 TB of hard disk. The right-hand sides of all the above figures show the runtime in seconds. In case of the original data set, we can observe that the higher amount s2 of nodes in the hidden layer(s) has a noticeable impact on the runtime. In those cases where the higher amount of hidden layer nodes only gives a marginal improvement of accuracy (as in Section 5.2), the lower amount of hidden layer nodes might be preferred in terms of runtime efficiency. Notably, the lowest value for the weight decay parameter (𝜆 = 0.001) has a very strong impact on the runtime. With this 𝜆 value the backpropagation has difficulties finding the optimization quickly. Overall, this 𝜆 value does not pay off neither in terms of accuracy nor in terms of efficiency. In case of the reduced data set, we can observe that not only the accuracies are better but also the overall runtime decreased. Hence, having a smaller set of data but with a higher overall data quality can lead to better classification results. 8 D I S C U S S I O N A N D C O N C L U S I O N In our experiments, a neural network is designed to classify gout and leukaemia patients based on their uric acid measurements. Tests showed that using more layers only improved the accuracy insignificantly. Yet it can be observed for our use case that the lower value for the weight decay parameter leads to a runtime increase due to a more involved optimization steps and low values for this parameters should be avoided. In our settings, using a high weight decay parameter and 20 hidden layers turned out to be the best setting for both accuracy and efficiency. Moreover, learning on the reduced data set (with better data quality) performed better than on the complete data set both in terms of accuracy and efficiency. Hence, regarding the tradeoff of having higher data quality with a reduced data set size versus having overall lower data quality with a larger data set size, we observed that a reduced data set size provided the most benefit. It can also be observed that the using linear regression transformed data did improve the accuracy of the system best when used in combination with the reduced data set. Overall, we can conclude that for our use case, this additional preprocessing step only provides a benefit on the reduced data set in terms of accuracy but not in terms of performance. To sum up, we conclude that several enhancements and settings of neural networks might not lead to optimal accuracy results. Hence, it should be carefully assessed which settings provide optimal results (both in terms of accuracy and efficiency) for the use case at hand. In future work, our study can be extended by more features in addition to the uric acid signatures in order to improve the accuracy results. An extension to cover other diseases than gout and leukaemia can also be a worthwhile topic of future work. Moreover, an in-depth comparison and combination with other related approaches (in particular, the feature learning approach in Lasko et al. (2013)) can be performed in order to assess the overall reliability of disease classification as well as quantify their runtime impact. F U N D I N G I N F O R M A T I O N None reported. C O N F L I C T O F I N T E R E S T The authors declare no potential conflict of interests. O R C I D Lena Wiese https:/ / orcid.org/ 0000- 0003- 3515- 9209 R E F E R E N C E S Alvarez-Lario, B., & Macarron-Vicente, J. (2011). Is there anything good in uric acid? QJM: An International Journal of Medicine, 104, 1015–1024. Bahra, G., & Wiese, L. (2018). Classifying leukemia and gout patients with neural networks. In International conference on database and expert systems applications workshops, pp. 150–160. Beaulieu-Jones, B. K., & Greene, C. S. (2016). Semi-supervised learning of the electronic health record for phenotype stratification. Journal of Biomedical Informatics, 64, 168–178. Eaton, J. W. (2019). GNU Octave version 5.1.0 manual: A high-level interactive language for numerical computations. https:/ / octave.org/ doc/ interpreter/ García-Gómez, J. M., Vidal, C., Martí-Bonmatí, D. L., Galant, J., Sans, N., Robles, M., & Casacuberta, F. (2004). Benign / malignant classifier of soft tissue tumors using mr imaging. Magnetic Resonance Materials in Physics, Biology and Medicine, 16(4), 194–201. Goldberger, A. L., Amaral, LuisA. N., Glass, L., Hausdorff, J. M., Ivanov, P. C., Mark, R. G., ... Stanley, H. E. (2000). Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals, Vol. 101: Circulation Electronic Pages. https://orcid.org/0000-0003-3515-9209 https://orcid.org/0000-0003-3515-9209 https://octave.org/doc/interpreter/ 14 of 14 BAHRA AND WIESE Huang, Y., McCullagh, P., Black, N., & Harper, R. (2007). Feature selection and classification model construction on type 2 diabetic patients' data. Artificial Intelligence in Medicine, 41(3), 251–262. Johnson, A. E., Pollard, T. J., Shen, L., wei H. Lehman, L., Feng, M., Ghassemi, M., ... Mark, R. G. (2016). MIMIC-III, a freely accessible critical care database: Scientific Data. Joshi, M., Pakhomov, S., Pedersen, T., & Chute, C. G. (2006). A comparative study of supervised learning as applied to acronym expansion in clinical reports. AMIA Annual Symposium Proceedings, 399–403. Juhola, M. (2008). On machine learning classification of otoneurological data. In ehealth beyond the horizon - get it there, IOS Press, pp. 211–216. Kotsiantis, S. B. (2007). Supervised machine learning: A review of classification techniques. In Maglogiannis, I. G., Karpouzis, K., & Wallace, M. (Eds.), Emerging artificial intelligence applications in computer engineering: Real word ai systems with applications in eHealth, HCI, information retrieval and pervasive technologies (pp. 3–24). IOS Press. Kourou, K., Exarchos, T. P., Exarchos, K. P., Karamouzis, M. V., & Fotiadis, D. I. (2015). Machine learning applications in cancer prognosis and prediction. Computational and Structural Biotechnology Journal, 13, 8–17. Lasko, T. A., Denny, J. C., & Levy, M. A. (2013). Computational phenotype discovery using unsupervised feature learning over noisy, sparse, and irregular clinical data: PLOS ONE 8(8). Lee, I.-N., Liao, S.-C., & Embrechts, M. (2000). Data mining techniques applied to medical information. Medical Informatics and the Internet in Medicine, 25(2), 81–102. Liu, D. C., & Nocedal, J. (1989). On the limited memory BFGS method for large scale optimization. Mathematical Programming, 45(1-3), 503–528. Maes, M., Twisk, FrankN. M., & Johnson, C. (2012). Myalgic encephalomyelitis (ME), chronic fatigue syndrome (CFS), and chronic fatigue (CF) are distinguished accurately: Results of supervised learning techniques applied on clinical and inflammatory data. Psychiatry Research, 200(2), 754–760. Nguyen, A., Moore, D., McCowan, I., & Courage, M. J. (2007). Multi-class classification of cancer stages from free-text histology reports using support vector machines. In 29th Annual International Conference of The IEEE Engineering in Medicine and Biology Society, pp. 5140–5143. Nguyen, D., & Widrow, B. (1990). Improving the learning speed of 2-layer neural networks by choosing initial values of the adaptive weights. In Ijcnn International Joint Conference on Neural Networks (Vol.3). Nielsen, M. A. (2015). Neural networks and deep learning: Determination Press. http:/ / neuralnetworksanddeeplearning.com/ Roden, D. M., Pulley, J. M., Basford, M. A., Bernard, G. R., Clayton, E. W., Balser, J. R., & Masys, D. R. (2008). Development of a large-scale de-identified DNA biobank to enable personalized medicine, Vol. 84: Clinical Pharmacology and Therapeutics. Rosenblatt, F. (1958). The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 65(6), 386. Schmidt, M. (2005). minFunc: unconstrained differentiable multivariate optimization in MATLAB. https:/ / www.cs.ubc.ca/ ~schmidtm/ Software/ minFunc.html Shouval, R, Bondi, O, Mishan, H, Shimoni, A, Unger, R, & Nagler, A (2014). Application of machine learning algorithms for clinical predictive modeling: A data-mining approach in SCT. Bone Marrow Transplantation, 49, 332–337. Tashkandi, A., Wiese, I., & Wiese, L. (2018). Efficient in-database patient similarity analysis for personalized medical decision support systems. Big Data Research. Team, R. C. (2014). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. https:/ / www.r- project. org/ Weng, S. F., Reps, J., Kai, J., Garibaldi, J. M., & Qureshi, N. (2017). Can machine-learning improve cardiovascular risk prediction using routine clinical data? PLOS ONE, 12(4), 1–14. Wiese, L. (2015). Advanced data management for SQL, noSQL, cloud and distributed databases. DeGruyter. Wilcox, W. D. (1996). Abnormal serum uric acid levels in children, Vol. 128: The Journal of Pediatrics. A U T H O R B I O G R A P H I E S Guryash Bahra received her B.Tech degree from Guru Tegh Bahadur Institute of Technology, India, and M.Sc. degree from University of Göttingen, Germany, in 2011 and 2018 respectively. Her research interests include topics like machine learning, data analysis, data mining and database systems. Lena Wiese is a member of the L3S Research Center Hannover. She also leads the research group ‘‘Knowledge Engineering’’ (at the Institute of Computer Science, University of Goettingen). She holds a PhD and a Master degree from TU Dortmund. After her PhD she worked as a postdoctoral researcher at the National Institute of Informatics in Tokyo and as a visiting lecturer at the University of Hildesheim and the University of Salzburg. Dr. Wiese is author of the text book Wiese (2015) on Advanced Data Management. Her research interests lie in the area of efficient and secure data management and analysis. She is an active member of the German Informatics Society (GI) and regularly acts as a reviewer for conferences and journals. How to cite this article: Bahra G, Wiese L. Parameterizing neural networks for disease classification. Expert Systems. 2019;e12465. https:/ / doi.org/ 10.1111/ exsy.12465 http://neuralnetworksanddeeplearning.com/ https://www.cs.ubc.ca/~schmidtm/Software/minFunc.html https://www.r-project.org/ https://www.r-project.org/ https://doi.org/10.1111/exsy.12465 Parameterizing neural networks for disease classification Abstract INTRODUCTION Medical background Machine learning techniques Objectives Outline of the article RELATED WORK DATA SET Data set identification Data set creation METHOD Neural networks Defining the layer sizes Forward propagation Initialization of weights Wand biases band activation function Cost function Define , and update parameters Wand b Accuracy calculation K-fold cross validation RESULTS ON COMPLETE DATA SET Results of cross validation on original data set Preprocessing of the data set using linear regression REDUCED DATA SETS Cross validation on reduced original data set Cross validation on reduced transformed data set RUNTIME COMPARISON DISCUSSION AND CONCLUSION Funding information Conflict of interest REFERENCES 
work_5fiiil2k3jagvoabh3rwcw3i6e	----	Expert systems and robotics Volume 93, Number 3, May-June 1988 Journal of Research of the National Bureau of Standards Accuracy in Trace Analysis Expert Systems and Robotics T. L. Isenhour Department of Chemistry Kansas State University Manhattan, KS 66506 and J. C. Marshall Department of Chemistry Saint Olaf College Northfield, MN 55057 In this paper, we will discuss the interface be- tween expert systems and laboratory robotics. We will use examples from our recent research to illus- trate how we are building an effective interface and indicate where we think this research will lead. What are expert systems? As an operational defi- nition we will adopt the following: Expert Systems are a sub-specialty in artificial intelligence (AI). The term is generally under- stood to mean a "knowledge-based" or "knowledge-driven" system designed to repre- sent and apply factual knowledge in specific, very limited areas of expertise. In the early sixties, AI researchers attempted to simulate the complicated process of thinking by finding general methods for solving broad classes of problems. This proved too difficult and such at- tempts failed. In the early seventies the problem was reformulated to include careful attention to data structures but the emphasis was still on gen- eral knowledge. Progress was still limited. In the late seventies the problem was further refined to focus almost completely on the knowledge repre- sentation. The goal was to make intelligent pro- grams by providing them with high quality, domain-specific knowledge about some limited problem area. This strategy is much like that used by a human expert and gives rise to the term "ex- pert system." What domains are appropriate for expert system work? First and foremost, for the present state of expert systems technology the problem domain must be of limited scope. A majority of the people within the application field must agree that real ex- perts do exist. The problem must be knowledge, not data, intensive. A problem is knowledge inten- sive if there is substantial variability in people's ability to solve it. The problem must not require information from visual input. Multiple answers from the same input data can be handled but with limited success. Perhaps the best test of all for a potential candidate for expert system work is the so-called "telephone test." If you have a problem and you are confident that if you called some known expert in the field, he or she could solve the problem for you in 30 minutes or less over the phone, then the problem is likely to be amenable to an expert system solution. How do expert systems compare with human ex- perts? The popular press has tended to be wildly optimistic about the present state of expert systems development. While many useful expert systems are available, they apply to very limited problem domains. In such domains expert systems can quickly provide answers that are consistent and ob- jective. Expert systems can capture human exper- tise and make it permanent, widely available and easily portable. However, current expert systems lack the creativity and adaptability expected of a human expert. How do expert systems work? Regardless of the details of the implementation, an expert system is a program driven by an inference engine towards a specific goal. It is, in the limit, a remarkably simple 209 Volume 93, Number 3, May-June 1988 Journal of Research of the National Bureau of Standards Accuracy in Trace Analysis process involving a cleverly ordered series of "if tests." A potential difficulty is, when the problem gets large and consequently the number of rule structures in the data base increases, an expert sys- tem can become difficult to modify, hard to debug and slow to execute. Expert Systems for Data Management Chemists, particularly analytical chemists, have historically been very concerned with the storage and retrieval of information. Spectral libraries are a common example. There is presently much interest in the potential of so-called "smart data bases" [11. The fundamental idea is that a data base is repre- sented as a collection of executable statements rather than facts. The smart data base concept is a subtle strategy that can be illustrated with a trivial example in- volving the periodic table. The entries of the data base all conform to the PROLOG predicate ATOMIC and become executable statements; as such they are no longer passive facts. The PRO- LOG definitions and data base entries for a small part of this periodic table are shown in figure 1. In this example, apart from the definitions, there are no program statements other than the data base. A compelling advantage is that all of the features of the Al language used (in this case PROLOG) are available to form queries and the inference engine interrogates the file automatically. This is illus- trated in figure 2. Methods Development Using Expert Sys- tems and Robotics A central theme in our research for the past sev- eral years has been the idea of the Analytical Direc- tor. Laboratory robots can carry out simple repetitive tasks, following an invariant set of rules. However, when a robot becomes a mechanical ex- tension of a control program that has logic capabil- ity the whole becomes greater than the sum of the parts. The Analytical Director project is an expert system driven robot that combines knowledge about analytical chemistry with laboratory robotics. The system is presently capable, in a lim- ited way, of designing procedures for analysis, test- ing and modifying such procedures, and finally archiving the modified procedure for future refer- ence. The flexible library facilities of the Analytical Director are possible because of the "smart data" capabilities inherent in the logic based program- ming languages. The current implementation of the Analytical Di- rector is a Zymark robot running under control of the ARTS [2] software system, an expert system driven robotics language. The control computer is a simple PC., To demonstrate an application of the Analytical Director, the development of a complexometric ti- tration procedure [3] is shown as a flow chart in figure 3. A successful complexometric titration requires that the conditions be adjusted so as to insure a conditional stability constant of about IX 105. Choices to be made include the pH, the titrant, the masking agent or agents used and the method of endpoint detection. A vast literature exists on corn- plexometric titrations. Some of this information is part of the knowledge base used by the ARTS sys- tem. The system not only starts with a knowledge base, but can continually update that knowledge base using results of experiments. The user is given the opportunity to specify some or all of the parameters that he/she wishes. The system will not override user input even though it may be wrong. The system will fill in missing user input from its knowledge base. The success or failure of a deter- mination is stored by the system for future use. Experimental results for the triplicate determina- tion of Ni+' by complexometric titration are shown in table I and compared with results obtained with manual titrations. Table 1. Comparison of expert system and human counterpart. Results for the titration of a Nil+ solution using 0. 1004M EDTA without an indicator. Absorbance data were collected at 480 nm Expert system Human Trial 1 0.1006 0.0981 Trial 2 0.1004 0.0981 Trial 3 0.1007 0.0983 Average 0.1005, 0.0981, Standard deviation 0.0001, 0.0001, %Standard deviation 0.15 0.12 210 Volume 93, Number 3, May-June 1988 Journal of Research of the National Bureau of Standards Accuracy in Trace Analysis Building Expert Systems from Chemical Data One of the most difficult problems with expert system work is creating an efficient knowledge base. When the knowledge base gets large, it be- comes imperative to create the most efficient possi- ble production rules. The knowledge base used by an expert system can be most efficiently structured as a set of rules that describe the minimal decision tree spanning the data. The root node of this tree is the attribute of the data that minimizes the num- ber of branches from the root. Each branch from the root node contains a different value of the root attribute and creates second level nodes. These sec- ond level nodes may be branched further using at- tributes different from the previous attributes used to split the data. The class attributes and values will occupy terminal nodes in the decision tree. If more than one attribute is used to describe the data, the decision tree will not be unique. As the number of attributes required to describe the data increases, the number of possible decision trees increases combinatorially. For this task we have implemented the ID3 (iter- ative dichotomizer 3) algorithm [4-6], originally developed for organizing and optimizing chess end-game strategies. The ID3 algorithm is based on information the- ory and uses the entropy of classification. The en- tropy of classification is a measure of the entropy resulting from classifying an object in a particular class. The algorithm will first determine the at- tribute to use for the root node of the tree so that the number of branches from the root node are minimum. Each branch from the root node repre- sents a unique value of the root attribute. The al- gorithm is then applied recursively to all the second level nodes, and all subsequent nodes spawned by each of the second level nodes. We have implemented the ID3 algorithm (7) in PROLOG so that it accepts classification data and determines an efficient set of rules spanning the data. The program will then produce a file of rules that can be used directly by an expert system as an efficiently ordered knowledge base. A simple example of how this works uses the infrared data in table 2. These data are applied to identifying substituted benzenes from their infrared absorption spectra. There is enough information in the first two bands to answer the question. There is no informa- tion in the last two bands relevant to this question. Table 2. Infrared data for some substituted benzenes Compound Degree IR ranges in cm-' name of 650- 700- 750- 800- 850- substitution 699 749 799 849 899 toluene MONO Sa s w w w m-xylene META S w S w w o-xylene ORTHO w S S w w p-xylene PARA w w S w w 'S=strong; W~weak. From these data, the algorithm outputs the infor- mation tree shown in figure 4. Not shown is the set of syntactically correct PROLOG production rules generated by the program that span the tree in fig- ure 4. Conclusion The purpose of this research is the combination of logic programming with laboratory robotics. The goal of this research is the creation of the Ana- lytical Director, an intelligent laboratory robotics system that will be able to develop, test and modify laboratory procedures without human supervision. Acknowledgment This research was supported by the National Sci- ence Foundation under grant number CHE- 8415295. /-PROLOG data base example-/ domains name,symbl = symbol number = integer weight = real predicates atomic(name~symblnumber,weight) clauses atomic("Hydrogen" ,"H",1,1.008). atomic("Helium","He",2,4.003). atomic("Lithium","Li",3,6.941). atomic("Beryllium". "Be",4,9.012) atomic( "Boron, "3" 6,5,10. 81) Figure 1. PROLOG data base example. 211 Volume 93, Number 3, May-June 1988 Journal of Research of the National Bureau of Standards Accuracy in Trace Analysis Goal: atoaimlaeNse.Svabol.flumber.Welght) . Number>1,.oight<:0 Navo-Heliua. Symbol-He. Nuber-2. WeIlght=4.003 Nane.tLithim..Sito~. Nujsber3. Weight=6.941 Namcarervllilu, Symbolue. Nunber.4. Weight=0.012 3 solutions Goal: atomic(Naae, 3"b",Number.Welghtl Nan-Sboron. Number2-, Weight-oOsl I soluti.n Goal' atomictNaneSyz~bol.5,Wolqht) Na.e'..oron, Symbo.1'. WeightiO. 31 I SolUtlon Figure 2. PROLOG data base interrogation examples. [4] Quinlan, J. R., Learning Efficient Classification Procedures and Their Applications to Chess End Games, in Machine Learning-An Artificial Intelligence Approach, Michalski, R. S., Carbonell, J. G., Mitchell, T. M., (Eds.), Tioga Pub- lishing Company, Palo Alto, CA (1983). [5] Thompson, B., Thompson, W., Byte 11, 149 (1986). [6] Derde, M., Buydens, L., Guns, C., and Massart, D., Anal. Chem. 59, 1868 (1987). [7] Schlieper, W. A., Isenhour, T. L., and Marshall, J. C., Using Analytical Data to Build Expert Systems, submitted, J. Chem. Inf. Comput. Sci. USER INTERFACE AND SYSTEM CONTROLLER (PROLOG) EXPERIMENTAL CONDITIONS (PROLOG) CONTROL INSTRIUMENTS (C) ANALYZE STANDARDS AND UNKNOWNS (C) Figure 3. Complexometric titrations under trol. expert system con- 650-6 99 S 700-749 m S W 70 0-749 m W S O T MONO META W I I ORtTHO PARA Figure 4. Decision tree generated from data in table 2. References [I] Schur, S., Al Expert 3, 26 (1988), (2] Schlieper, W. A., Isenhour, T. L., and Marshall, J. C., Anal. Chem., in press. [3] Schlieper, W. A., Isenhour, T. L., and Marshall, J. C., Com- plexometric Analysis Using an Artificial Intelligence Driven Robotic System, J. Chem. Inf. Comput. Sci., in press. 212 I 
work_5i2amsy5abfjnectpyizxhpmfy	----	06 April 2021 AperTO - Archivio Istituzionale Open Access dell'Università di Torino Original Citation: Leveraging social semantic components in executable environments for learning Published version: DOI:10.1111/exsy.12044 Terms of use: Open Access (Article begins on next page) Anyone can freely access the full text of works made available as "Open Access". Works made available under a Creative Commons license can be used according to the terms and conditions of said license. Use of all other works requires consent of the right holder (author or publisher) if not exempted from copyright protection by the applicable law. Availability: This is the author's manuscript This version is available http://hdl.handle.net/2318/143214 since 2016-06-29T12:17:04Z Leveraging social semantic components in executable environments for learning Rossana Damiano Dipartimento di Informatica and CIRMA Università di Torino - Italy rossana@di.unito.it Cristina Gena Dipartimento di Informatica and CIRMA Università di Torino - Italy vincenzo@di.unito.it Vincenzo Lombardo Dipartimento di Informatica and CIRMA Università di Torino - Italy VRMMP - Torino - Italy cgena@di.unito.it Abstract Learning can benefit from the modern web structure through the convergence of top–down encyclopedic institutional knowledge and bottom–up user–generated annotations. A promising approach to such convergence consists in leveraging the social functionali- ties in 3.0 executable environments through the recommendation of tags with the mediation of lexical and semantic resources. This paper addresses such issues through the design and evaluation of a tag recommendation system in a Web 3.0 web portal, “150 Digit”. Designed for schools, “150 Digit” encourages students and teachers to interact with a set of four exhibitions on the histori- cal and social aspects of the Italian unification process in a virtual environment. The web site displays the exhibits and their related documents promoting the users’ active participation through tag- ging, voting, and commenting the exhibits. Tags become a way for students to create and explore new relations among the site con- tents, orthogonal to the institutional viewpoint. In this paper, we illustrate the recommendation strategy incorporated in “150 Digit”, which relies on a semantic middleware to mediate between the input expressed by the users through tags and the top-down institutional classification provided by the curators of the exhibitions. Following on, we describe the evaluation process conducted in a real experi- mental setting, and discuss the evaluation results and their implica- tions for learning environments. Keywords: tag recommendation, Web 3.0, evaluation, learning environments 1 Introduction According to W.L. Hosch’s definition , Web 1.0 can be described – using an analogy to file system permissions – as “read-only”, Web 2.0 as “read-write” and Web 3.0 as “read-write-execute”.1 Fol- lowing the Web 3.0 principle of executability, the “150 Digit” web portal (http://www.150digit.it) has been designed with the goal of creating a virtual environment where schools interact with the ex- hibitions that celebrate the 150th anniversary of the Unification of Italy. In 150 Digit, a 3D reconstruction allows users to visit the exhibitions; and encourages them to be an active part of the site community by tagging, voting, and commenting the exhibits, and by uploading new contents. The site contents include both the ex- hibits, with their related documents (such as the curators notes), and user–generated contents, such as multimedia presentations created by teachers and students on the Unification of Italy. As a result of the user centered methodology on which the site was designed, tagging emerged as a primary issue starting from the de- 1http://www.britannica.com/blogs/2007/07/web-30-the-dreamer-of-the- vine/ sign phase, in the focus groups organized with the teachers involved in the project. Since the pioneering work by [Bateman et al. 2007] tags have been pointed out as an important resource in learning: “The information provided by tags provides insight on learner’s comprehension and activity” (p. 1). Being primarily targeted at educational users, tags play a two-fold role in 150 Digit: on one side the tagging activity, as stressed by the focus groups partici- pants, is part of the educational processes and promotes the linguis- tic reflection of the students over the site contents; on the other, tags complement the institutional stance on the themes covered by the exhibitions, mirroring the users own understanding and letting new opinions emerge. Also, tagging fosters new correlations of the site contents and can be exploited by the users to navigate in an alter- native way to following the paths offered by the site’s information architecture. Finally, user preferences and tags are used to generate recommendations of contents and promote the exploration of the site in a “bottom-up” perspective. In 150 Digit, the support provided to the users (and educational users in particular) to improve tagging is a recommender module, which was designed and developed in order to meet the project’s specific needs. Suggesting tags to users aims at overcoming the well known trend according to which a site folksonomy stops or slows its growth after some time because the users start to use the same tags, and do not introduce new tags anymore [Trant 2009]. The novelty of the approach developed for 150 Digit is that the tag recommender relies on the semantic description of the exhibitions provided by the curators. Basically, the generation of new tags is obtained through lexical resources, which allow the recommender to expand the meaning of the existing tags, while non relevant tags are filtered out by consulting the semantic description given by the curators. The rationale for this approach is to leverage the exhibi- tions’ institutional perspective to focalize the generation of new tags and support the teachers work more effectively. At the same time, this strategy aims to avoid the generation of a semantic gap between the institutional categorization of the contents and the folksonomy, keeping them aligned as the latter is expanded. Notice that this can be seen as a variation of the well known “vocabulary problem” (first identified by [Furnas et al. 1987]), i.e. the lack of convergence of terms in user–generated vocabularies. In this paper we describe and evaluate the tag recommendation sys- tem of 150 Digit as a method to improve the contribution of the so- cial semantic components in executable environments designed for learning. The semantic layer of the site relies on a light ontology, WordNet Domains [Bentivogli et al. 2004], a taxonomy of domains originally developed to add semantic information to the meanings of the terms in WordNet [Miller 1995]. In 150 Digit domains are used to categorize the exhibits and the user-contributed contents, providing a background description against which new tags are sought by the recommender. The recommendation of tags exploits the meaning relations encoded in the Italian version of WordNet, MultiWordNet [Pianta et al. 2002], to expand on the existing tags and propose new ones. From March to June 2012, students and teachers visited the large exhibition “Fare gli italiani” (“Making Italians”), where they at- tended a post hoc laboratory in which they were asked to interact with the “150 Digit” portal. These laboratories were the basis for a thorough evaluation of the system. In this paper, we analyze the results of the evaluation, assessing the impact of our approach on the accuracy of the inserted tags, their quantity and typology. Given the data recorded in log files (a sort of indirect observation), we also analyze the users’ behavior in general and compare the folksonomy generated through the system use with a baseline. The paper is structured as follows: after surveying the related work (Section 2), in Section 3 an overview of the “150 Digit” web portal is given, in terms of its design, goals and functionalities, including a prelimi- nary evaluation conducted on the prototype system. Section 4 gives an evaluation of the portal through an experiment with real users. Discussion and conclusions end the paper. 2 Related work Since the advent of Web 2.0, tagging has attracted much interest in scholars and has been studied under many perspectives. In partic- ular we acknowledge three main areas in the corpus of tag-related research. Tags have been studied with the goal of understanding the behavior and interests of users, letting different tagging styles emerge; more recently, they have been studied as a resource, in an attempt to extract ontologies from user-generated folksonomies; fi- nally, promising attempts have been made to exploit tags in order to provide personalized recommendations and services to users, rang- ing from tag recommendation to recommendation. In e-learning, tags can be viewed as a means to gain insight on students’ learn- ing [Bateman et al. 2007]; from them, information can be gathered to build user profiles, aimed at making the learning environment adaptive [Ferreira-Satler et al. 2011]. Several attempts have been made to interpret sets of user tags. Users can tag with different purposes: to categorize or describe a resource for future retrieval, or to give an opinion [O’ Donovan 2009]. Concerning the action of tagging in the artwork domain, the results of the experiments in the Steve.museum project [Trant 2006] show that when users add tags on artworks, professional users and non–expert users insert complementary information: non-expert users insert information on the subject of the artwork (such as, in case of a painting, the people and place depicted, the ideas it suggests, the emotions, etc.), while experts provide only “external” information regarding the authors, the historical period, materials and so on. Moreover it emerged that users are generally keen on leaving a trace of what they think and feel. Other works envisage the complementarity of top-down classi- fication and user-driven classification. [Szomszor et al. 2007] suggest that the best solution for resources accessibility would be to integrate the users’ subjective perspective with traditional classification systems. This could exploit the benefits of both approaches limiting their respective problems. This idea also inspired the work of “150 Digit”, which integrates the curators’ knowledge encoded in the semantic component and the users’ perspective expressed by tags. Regarding the user participation to both content creation and tag insertion, Nonnecke et al. [Nonnecke et al. 2004] and Preece et al. [Preece et al. 2004] identify the roles of “lurkers” and “posters”, where lurkers are members of online communities who read, but do not post, and posters are the few members who post content. These results have been recently confirmed by Gena et al. [Gena et al. Accepted for publication.]: the results show that the most partici- pating users contribute in the form of small contributions (clicking on a tag for insertion, clicking on like/dislike) and just a few of them generate bottom-up contents. Analyzing the users’ tagging activity, they reported that 84% of user tags were the ones proposed by the system and just clicked on by the users, while the remaining 16% were inserted by users as free text. This demonstrates that, when available, users tend to select proposed tags instead of inserting new ones, thus providing support to the use of tag recommenders. These findings are also stated in [Ames and Naaman 2007]. The authors reported that in the same domain (photos), users tag more in sys- tems that recommends tags (ZoneTag) than in a system that does not offer tag recommendations (Flickr). Concerning the recommendation of tags, most approaches rely on statistical techniques (PageRank, evolved into FolkRank [Hotho et al. 2006]) to learn correlations among tags from their co- occurrence in a folksonomy, and use this information to suggest suitable tags for a resource. In our approach, the tag recommen- dation mechanism relies on the meaning relations encoded in an external resource, i.e., WordNet [Miller 1995] and WordNet Do- mains [Bentivogli et al. 2004], following the approach proposed by [Xu et al. 2006]. Similarly, [Cantador et al. 2011] proposed a method that uses the YAGO ontology (containing information from Wordnet and Wikipedia) for filtering and classifying tags into a set of purpose-oriented categories (content-based, context-based, sub- jective, and organizational). The results show that content- and context- based tags are considered superior to subjective and or- ganizational tags in helping a tag recommender component. They found that the transformation of tags into ontology concepts con- sents inferring semantic relations among concepts for recommen- dation purposes. In content-based recommenders, the use of WordNet to improve recommendations is not new. [Degemmis et al. 2007] transform the classic keyword based profiles into semantic user profiles uti- lizing Wordnet and experienced that semantic user profiles produce more accurate recommendations. [Laniado et al. 2007] propose in- tegrating WordNet in the navigation interface of a folksonomy. In particular using WordNet to build a hierarchy (top-down classifica- tion) of related tags (the relatedness is calculated according to well known similarity metrics in Wordnet) can help users navigate and find related resources in del.icio.us. This approach is quite similar to our tag recommendation strategy, in which related tags are sug- gested on the basis of the hierarchical relation encoded in Wordnet. Finally [Djuana et al. 2011] have found that a backbone ontology, such as the 43 categories incorporated in WordNet, may improve tag recommendation. They automatically learn the ontology from user tags, and use this ontology to improve recommendations by re-ranking the proposed tags on the basis of a collaborative filtering algorithm. They have found that the re-ranking procedure improves precision and recall. We currently do not consider WordNet cate- gories in our recommendation process, though an automatic learn- ing component could be added in order to perform this task. A strong focus on semantics characterizes the content-based rec- ommendation systems [Pazzani and Billsus 2007], as is the case of 150 Digit. An essential component for content recommenders is a system to describe the items that may be recommended, and this description very often relies on an ontology. For instance, in e- learning, Protus 2.0 tutoring system [Vesin et al. 2012] is a content- based recommender that uses an ontology for knowledge represen- tation and inference engines for reasoning. [Tang et al. 2012] start- ing from a mining approach combined with fuzzy logic techniques generate a Personal Web Usage Ontology (written in OWL), which enables personalized web resources recommendation. On the side of content-based recommender in the artwork domain we mention the CHIP artwork recommender2. Similarly to 150 Digit, where most content items are constituted by artworks, one of the main goals of CHIP is to demonstrate how Semantic Web and recom- mendation technologies can be deployed together to improve the access to digital museum collections [Wang et al. 2009]. Results from a user test demonstrate that users prefer content-based rec- ommendations that leverage artwork features, and conclude that domain-specific terms are generally more useful for content-based techniques than generic ones. This finding is in line with our de- cision to use domain knowledge, encoded in the semantic catego- rization of the contents provided by the curators, to improve the recommendation of tags. 3 System Overview The goal of 150 Digit is to provide an open environment where stu- dents can visit the exhibitions online and access a wide repository of multimedia items related with the subject of the Unification of Italy. The site contains both institutional contents, taken from the exhibi- tions, and user–generated contents. Contents can be commented on and tagged by users, thus generating new connections over them. Tags are exploited to group contents on the fly, through a dedicated tag-based search tool; tags and preferences are exploited to recom- mend contents to the users. By doing so, the site integrates the top-down perspective reflected in the institutional categorisation of content with the bottom-up perspective induced by the users’ activ- ity. 3.1 Functionalities and Design The project encompasses three user profiles: the editor, who is in charge of editing and publishing the institutional contents provided by the exhibition curators, and validating the contents uploaded by the students; the profile of the classes, student and teacher, who can visit the exhibitions, add tags and comments to the exhibits, vote them, and upload new items; and the registered user, who does not belong to a class but can visit the exhibitions, vote and tag the ex- hibits, and create her/his own playlist in a private area. Dedicated tools like the “virtual classroom” (a separate space to comment site contents shared by a group of students under the guidance of one or more teachers) are aimed at improving the quality of the inter- action with the site for educational users (for full description of the system, see [Damiano et al. 2011]). Given these profiles, the portal has three main functions: content management, content editing and navigation. • A content management system allows the site editors to cre- ate the site main sections (in 150 Digit, they consist of exhi- bitions) and categories within these sections, to add contents to the categories and describe them through tags and semantic labels. These labels constitute the semantic layer of the site (as described in Section 3.2). • A simple content editor lets the educational users edit and publish contents in the existing exhibitions and categories, de- scribing them through tags. During the tagging process, users can ask the system to recommend tags. Suggesting tags to users is a way to contrast the trend according to which a site folksonomy slows its growth because the users stop introduc- ing new tags [Trant 2009]. In 150 Digit, given the educa- tional goals of the site, tag recommendation also serves the purpose of supporting the teachers’ work on the linguistic de- scription of the exhibits and documents contained in the site. Web 2.0 functionalities, such as tagging and commenting, are also available as part of the site navigation. 2http://www.chip-project.org/ • Site navigation, open to non–registered users, is the same for the three profiles. In addition to the standard navigation en- forced by the site’s information architecture, users can nav- igate the contents by following content recommendations or by using the tag-based search tool. In a didactic perspective, exploring the site through the recommendations provided by the site or through the tag-based search can be seen as a way to support the theachers’ work in relating the items of the ex- hibitions into alternative, coherent narratives. The interaction design of 150 Digit relies on the ‘visit’ metaphor to structure the information. The user can visit the four exhibitions with a standard hypertext-based format, by following the connec- tions over the items induced by tags through the tag-based search, or in a 3D modality (see Figure 1). The portal features a plugin, tested on major browsers, to navigate the exhibitions in a 3D environment, with the aim of making the access to the exhibits more compelling. This approach is borrowed from entertainment (videogames in par- ticular) in order to offer students with an immersive, non textual access modality they are familiar with. • The standard, hypertext-based navigation follows a classical top-down approach from general categories to detailed infor- mation. The information architecture encompasses three lay- ers, namely exhibitions, categories and items; at item level, the user can move across items by following recommenda- tions. • The 3D navigation contains a set of navigation paths that mir- ror those experienced by the visitors in the real exhibitions. The use of the same structure in both the standard and the 3D visit is aimed at providing guidance to users in the 3D space. The 3D visit relies on the paradigm of constrained spatial nav- igation [Burigat and Chittaro 2007], i.e., it is constrained to some fixed positions, in sequential order, where the visitor is “transported” through a stepwise flight simulation (briefly de- scribed in [Damiano et al. 2012]). • The tag-based navigation provides a bottom-up approach to site contents. In this modality, users can take advantage of the search functionalities (by keyword, artwork’s title, author, tag, etc.) and sort the items by number of views, users’ pref- erences, and so on. Users can switch from one modality to another (for instance from 3D to hypertext, from hypertext to tag-based navigation) anytime during navigation, and remain in the same (virtual) location (e.g. the same category or item) after the switch. This approach is intended to stress the parallelism between the various navigation modalities, taking user’s need of orientation into account, and giv- ing them the possibility to easily switch among multi-modal infor- mation and different viewpoints of navigation. 150 Digit was developed by a multi–disciplinary team, involving AI, computer graphics, interaction design and media experts, and with the participation of the target users in all the phases of the project, from design to prototyping, according to a user-centered, iterative design methodology. The resulting portal integrates differ- ent components (social, didactic, informative) in a seamless inter- face that overcomes the challenges posed by the software integra- tion issues and the content production process. The web 3.0 portal interface design was inspired by usability heuristics and guidelines, as well as by information architecture principles. Moreover the web pages were created in respect of the Italian accessibility law (Stanca Act). A usability expert supervised the interface design together with the web designers, and reported heuristics and guidelines that guided the design decision process. Figure 1: The visit modalities in 150 Digit. Left, standard hypertext; center, 3D visit; right, tag–based visit. Different types evaluations were carried out by the project team at different stages of development. In the system design stage, and in particular during the requirement elicitation, a focus group of 5 users, 4 males and 1 females, aged 40-62, was selected. The par- ticipants were shown to a set of 15 scenario based static interfaces and the main systems functionalities, labeling and layout with the designers, for 3 hours. In general this group of teachers highlighted the need for textual content to be associated to the exhibits, and for dedicated tools for content creation. They appreciated the pro- posed interfaces/functionalities, and considered them as valid tools for classroom work and students’ involvement. The main findings emerged from the focus group affected the project with changes in both labeling (e.g., “favourites” instead of “playlist”) and func- tionalities. In particular, some of the existing functionalities were modified (for example, teachers suggested to show tag recommen- dations only on request), and new ones were added: mainly, the possibility of creating a virtual class where students can discuss the exhibits and insert comments that are visible only within this class. A preliminary evaluation was conducted on a static prototype, which consisted of interface screenshots. This evaluation aimed at verifying the navigation issues (such as breadcrumbs, home button, etc.) in the graphical interface and the users’ reception of the so- cial (tagging) and semantic (tag recommendations) functionalities. 5 users were tested. These were teachers, 3 males and 2 females, aged 25-55. The test on the static interface consisted of showing the users screenshots to and discussing the solutions with respect to the aforementioned functionalities, while tag recommendation module test consisted of the accomplishment of a set of tasks, such as tag- ging or voting an item. The issues that emerged from this evaluation concerned the understanding of the social aspects of the site, such as the role of tags in the fruition of the contents. So, in the re- design, tooltips to explain the meaning of social functions and the possibility of increasing the size of the pictures which illustrate the contents were added. 3.2 Semantic Framework and Tag Recommendation The need to support prototyping, development and production within a tight time schedule has determined the choice to rely on ‘light’ semantic tools to leverage the portal recommendation func- tions. The system semantics rely on WordNet Domains [Bentivogli et al. 2004], a hierarchy of domain labels (169 labels) integrated in MultiWordNet. While most ontologies require expert knowledge to understand their structure (consider for example, foundational on- tologies like SUMO [Niles and Pease 2003] or DOLCE [Gangemi et al. 2002]), WordNet Domains lends itself to the use by non expert users, providing an off-the-shelf, portable middleware on the top of which semantic tools can be built. Semantic Categorization of Contents. In 150 Digit, the recom- mendations provided by the system rely on a semantic categoriza- tion of the contents, with the aim of integrating the social compo- nent with the institutional perspective conveyed by the curators in the conceptual organization of the exhibitions. For each exhibition in 150 Digit, each category was associated by the curators to the do- mains, which, according to them, better describe the category cov- erage in semantic terms. For example, the “Timeline” category was associated with the “Time Period” domain, the “Mass media” with multiple domains, “Linguistics”, “Photography”, “Telecommuni- cation”, “Cinema”, “Radio”, “Telephony” and “Tv”. The underly- ing assumption is that semantic tools (and taxonomies in particular) can provide an effective “external grounding” to the relations over tags, as exemplified by the work of [Markines et al. 2009], that em- ploys taxonomies (such as WordNet) to measure the reliability of the emergent semantic relations among tags in folksonomies, thus providing a sound foundation to the Social Semantic approach. Tag Expansion and Disambiguation. Differently from standard approaches, which exploit statistical techniques to recommend tags (as in the case of PageRank, re–cast into FolkRank [Hotho et al. 2006]), the tag recommendation mechanism in 150 Digit consists of a constrained expansion of the meaning of existing tags, based on the semantic relations over the lexical items incorporated in Word- Net [Miller 1995]. In WordNet, words are gathered into sets of synonyms (i.e. words with same meaning), called synsets; synsets are linked according to meaning relations, such as hyperonymy (more general meaning) and hyponymy (more specific meaning). MultiWordNet includes the Italian language and is aligned to WordNet 1.6. The basic ex- pansion relies on the synonymy relations among lexical items en- coded in synsets: 1. For each user tag, get the corresponding lexical entry from the lemmatizer; 2. Given the lemma, get the synsets from MultiWordNet in which it appears; 3. For each synset found, get all the lemmas contained; 4. Merge the obtained synsets by deleting the repeated entries. Further expansion relies on querying MultiWordNet for related synsets based on hyperonymy and hyponymy relations at step 3. The simple expansion mechanism described above however does not guarantee that the recommended tags are actually related to the user tags, due to the polysemy of natural language. In other words, a tag may correspond to more than one lexical entry. To Figure 2: Screenshots of the tag recommendation interface. The slider allows the user to regulate the quantity of recommended tags (from left, “Pochi–Few” tags, to right, “Molti–Many” tags). Recommended tags (here, given the input tag “folla”, i.e., “crowd”) are arranged in a tag cloud. The terms referring to “crowd”, include “mass”, “army”, “bunch”, “swarm”, etc. . overcome this difficulty, two disambiguation strategies are incor- porated in the tag recommender. The disambiguation relies both on ‘syntactic’ knowledge provided by the context of other tags and on the ‘semantic’ knowledge contained in the semantic layer. The ‘syntactic’ disambiguation relies on the context of the other tags as- sociated with the item: for each proposed tag, if it co-occurs in the same synset with one of the context tags, it is included in the rec- ommended tags; otherwise it is discarded. For example, consider the situation in which the tags associated with an exhibit are “emi- grants” (emigrant), “pescatore” (fisherman) and “giovane” (youth). Following this strategy, a tag which is a synonym of one of the three tags (for example, “ragazzo”, i.e., young man) will be recom- mended, while a tag which is not a synonym of any of the three tags (such as “garzone”, i.e. shop boy) will be not recommended. The ‘semantic’ disambiguation relies on the domains attached to the categories, inspired by [Magnini et al. 2002]. Each exhibit in- herits the domain labels associated with the category it belongs to (each exhibit belongs to only one category, parallel with the actual arrangement of the exhibition) and with the exhibition itself. These domains provide the semantic context against which the proposed tags are filtered to eliminate the non relevant ones. For example, consider the Italian word “quadro”. This word has two different meaning, “painting” and “control panel”, the first one associated with the “Art” domain in MultiWordNet, the second one associated with the “Electronics” domain. If the disambiguation occurs in the category “Painters and patriots” (associated, among others, with the “Art” domain), only the first meaning of the word “quadro” is con- sidered, while the second one (with its synonyms and other related terms) is discarded because its domains don’t match the category domains. Interactive Tag Recommendation. In order to let the user control the combination of the expansion and disambiguation techniques described above, the recommendation of tags is accomplished in an interactive fashion (see Fig. 2). If the user enters one or more tags in the system, an auto–completion function shows the possible words given the letters inserted so far; then, the user can then ask the sys- tem to propose new, related tags. If no tags have been inserted by the user, the recommendation takes the tags that are already asso- ciated with the current item as input (if any, otherwise, the recom- mendation cannot be made). The amount of recommended tags is regulated by a slider: the user can move the slider from the “Few tags” position (the starting position) to the “Many tags” position, through intermediate positions. Each position corresponds to a dif- ferent combination of tag expansion and filtering. Figure 2 shows how the cloud of recommend tags grows as the user moves the slider from left (“less tags”) to right (“more tags”), with two intermediate positions between the initial recommendation to the highest expan- sion of the user inserted tag. Although the interface allows the user to control the tag expan- sion mechanism, the presence of hyponyms and hyperonyms may still disorientate them, since their introduction in the set of recom- mended tags may not be obvious, especially the first time the sys- tem is used. In order to overcome this problem, a tag cloud presents the recommended tags, so as to alert the user of the possible pres- ence of unexpected tags. As the user moves the slider, the tag cloud grows or shrinks, and the user can accept one or more of the rec- ommended tags by clicking on them. In the suggested tags cloud, the font size of each tag is given by a combination of two factors: tag frequency in the folksonomy and in the language use. The use of word frequency in language use (taken from a frequency lexi- con, “Corpus e Lessico di Frequenza dell’Italiano Scritto”, CoLFIS [Laudanna et al. 1995]) has the function of making more unusual terms less visible in the cloud. Recommender Architecture. The architecture of the tag recom- mendation system includes the following components: • Lemmatizer: performs the morphological analysis of the user tag, returning its non flexed form, needed to access the lex- ical knowledge. For example: “persone” (people), the plu- ral form of “persona” (person) is converted into the singular form. Since most tags are nouns, we chose to consider only the plural to singular conversion. The latter is achieved by using a data base of forms, implemented in mySql. • Expansion Module: written in PHP, implements the expansion of the user tags along the semantic relations incorporated in MultiWordNet, as described above. Again, MultiWordNet is stored in a mySql data base and is accessed by a set of PHP APIs. • Disambiguation Module: implements the context–base and the semantic–based disambiguation strategies. This module interacts with the site CMS to get the set of tags that have al- ready been added to the item and the domain labels that are associated with it. • Tag Cloud Generator: this module determines the size of the tags in the generated cloud based on the frequency of tags in the folksonomy and in the lexicon. Item recommendation. The content recommendation function relies on two complementary approaches: a collaborative filtering approach [Schafer et al. 2007] and a semantic approach. So, the user is presented with two sets of Figure 3: The semantic architecture of the web 3.0 portal. recommendations, one of which is based on the preferences given by the other users, and the other is based on the tags added to the items. The semantic–based recommendation selects the items to recom- mend based on the shared tags with the current item. Items are ranked according to the number of tags they have in common with the given item. Items with the same ranking are re-ranked according to the category to which they belong: items from the same category (and the same exhibitions) as the current item are preferred. The social recommendation is based on the preferences expressed by the community of the users, and is inspired by the technique of collaborative filtering [Xu et al. 2006; Sarwar et al. 2001]: 1. Given the current content, select its highest vote; 2. Select all the users who have given the same vote to that item; 3. For each of these users, select the items to which the user has given the same (or higher) vote ; If the set is empty, set the vote to vote – 1; 4. Rank the selected items by their highest votes; 5. Select the first n contents; The user is presented with the two sets of recommendations (tag– based and preference–based); the difference between the two is communicated by different labels, “150 Digit recommends” and “Other schools recommend ” respectively. In case the same item appears in the two sets, the duplication is eliminated. 3.3 Preliminary Evaluation Given the logs of the first six months of publication of the web site, a preliminary evaluation of the users’ acceptance of the site functionalities was conducted, and of the recommender system in particular. Only front-end users were considered for social func- tionalities (content generation and tagging), while for the semantic analysis of tags back-office users were also taken into considera- tion, as they benefit from this kind of recommendation because they are requested to tag exhibits as part of the publication process During the first six months, 347 users logged onto the site, 149 (42.93%) active teachers, 199 (57.35%) regular visitors. Of the regular visitors, 61 users (31%) were associated to classes. It is im- portant to note that teachers and classes were explicitly contacted by the committee in order to promote the portal. These teach- ers/classes were randomly selected over a set of teachers/classes who regularly participate in trials organized by the Ministry of Ed- ucation. This evaluation, conducted for prototype refinement and experiment design, was split into two parts: the first concerning the generic Web 2.0, i.e. social, functionalities and the second relating to the 3.0, i.e. executable, functionalities. Web 2.0 functionalities. The users’ behavior in relation to social functionalities can be ana- lyzed in terms of: • their participation in content creation, • their tagging activity, • the quantity and the typology of inserted tags. With regards to the uploading of user-generated content, 11 virtual classes of the 51 registered classes (21.57%) inserted new contents (a total of 29 new contents, while the institutional contents are 271). In detail, 3 classes with same teacher inserted more than half of con- tents (51.72%), one class inserted 13.79% of contents and another one inserted 10,34% of contents. Thus 5 classes out of 11 (45.45%) inserted 76% of contents. In total 404 tags (duplicates included), either freely inserted by users or selected among the tags suggested by the system, have been collected since the beginning of the experimentation. More specifically 297 tags (73.51%) were proposed by the system and just clicked on, while the remaining 107 ( 26.41%) were inserted by users in free text. 28 users out of 347 (8.07%) inserted tags. Of these 17 were teachers working with their virtual classes (61%) and 11 were regular visitors (39%). In particular a teacher using the site both as a regular visitor and with her 5 classes inserted al- most half the tags (201 tags, 49.75%). Note that this teacher, and her classes, were the same that inserted more than half the con- tents. Another teacher both as visitor and with her class inserted 16.83% of tags (namely 68 tags), while another class created an above average number of tags (31 tags, 7.67%). The remaining classes (31.71%) inserted an average of 5.9 tags per class, while regular visitors (32.14%) inserted an average of 5 tags per user. The low user participation in both content creation and tag insertion con- firms the results of [Nonnecke et al. 2004] and [Preece et al. 2004] and replicates the dichotomy between “lurkers” and “posters” men- tioned above. Not all the tagged contents in exhibition received the same number of tags, see Table 1. The frequency of tags with respect to exhibi- tions/sections needs to be balanced with the number of contents present in each exhibition/section. In general, the number of tags is proportional to the number of content items in the exhibitions, with two exceptions: “La bella Italia” and the “Extra contents” sections. “La bella Italia” received the least user attention in term of tags, despite its relevant number of contents; the “Extra contents” section, i.e. the section containing schoolgenerated contents that are not strictly related to the main exhibitions, was particularly successful. This success is not surprising, as the contents generated by classes receive more attention by the classes themselves. Moreover, the insertion of a new content implies the insertion of tags. It is interesting to compare the number of tags received by each ex- hibition with the number of visits of the real and virtual exhibitions. In the real world, the exhibition “Fare gli italiani” had the highest number of visitors, followed by “La bella Italia”, “Stazione futuro”, and “Il futuro nelle mani”. The number of visits received by the ex- hibitions on the web site “Fare gli italiani” is still the most visited exhibition followed by “La bella Italia”, “Il futuro nelle mani”, and “Stazione futuro” (the last two being almost on a par). While the Figure 4: Content recommendation in 150 Digit. On the left, the tag–based recommendations (“The systems recommends you”) ; on the right, the preference–based recommendations (“Schools recommend you”). Exhibition/Section Number of tags Number of contents “Fare gli italiani” 179 tags (44.31%) 135 (45%) “Extra contents” 70 tags (17.33%) 28 (9.33%) “Il futuro nelle mani” 37 tags (9.16%) 48 (16%) “Stazione futuro” 33 tags (7.92%) 30 (10%) “La bella Italia” 16 tags (3.96%) 38 (12.6%) “The places (of current exhibitions)” 16 tags (3.96%) 9 (%3) Table 1: The most tagged exhibitions/sections trend regarding the former exhibitions has also been confirmed by the 150 Digit taggers activity, the latter ones, in particular “La bella Italia”, reveal a much lower number of tags with respect to their virtual visits. An explanation could be that the sections/exhibitions receiving more tags are those whose contents are more pertinent to the topics covered by the study programme. Regarding the tagged contents, 109 items have been tagged, with an average number of 3.71 tags per item. However the distribution of tag per item is not homogeneous. 10 items (9.17%) received a number of tags more than twice the average, as detailed in Table 2. The other items (90.83%) received a number of tags ranging from 1 to 7. More specifically 4 items (3.67%) received 7 tags, 5 items (4.59%) re- ceived 6 tags, 10 items (9.17%) received 5 tags, 7 items (6.42%) re- ceived 4 tags, 17 items (15.60%) received 3 tags, 32 items (29.36%) received 2 tags, 24 items (32.02%) received 1 tag. Notice that most of the items received a low number of tags. The most used tags (“history”, “unification”, “Italy”, “risorgimento”, “tradition”, etc) reflect the historical context of the web site contents (the celebra- tion for the 150th anniversary of the unification of Italy), while the others reflect the artwork content (“woman”, “women”, “food”), namely subject related tags. The remaining 254 tags have been used with these frequency values: 198 tags (49%) have been used once, 47 tags (11.63%) have been used twice, 9 tags have been used 3 times. To sum up we can conclude that a few tags are used more than once, while most of the tags are used once, or twice at most. However these considerations must also take into account the lim- ited sample of users involved in the trial. From these data, we concluded that the users’ behavior with the 150 Digit system is coherent with the social functionalities largely reported in the literature. Different groups (less active and more ac- tive users) emerge for the quantity of uploaded contents and added tags, with the distribution of tags featuring the most common tags, in line with the themes of the exhibitions. Web 3.0 Functionalities The data set collected in the 6-month testing of the prototype sys- tem was employed to conduct a preliminary evaluation of the ade- quacy of our approach to the recommendation of tags. Our work- ing hypothesis is that, if the approach is correct, the semantics of the folksonomy should, to some degree, match the institutional cat- egorization of contents, thanks to the use of the semantic layer in the recommendation of tags. In order to obtain a semantic descrip- tion of the folksonomy, comparable with the description of the cate- gories in the semantic layer, we adopted the very same resource em- ployed to encode the semantic layer itself, i.e., WordNet Domains (see Section 3.2). To do so, tags were associated with the domains of the corresponding terms in MultiWordNet, thus obtaining a pic- ture (though a coarse one) of the semantics of the folksonomy. The obtained representation can be straightforwardly compared to the description of the categories encoded in the semantic layer, which were entered with the relevant domains by the curators. As a preliminary step of the analysis, duplicated tags were elim- inated from the folksonomy; then each tag was associated to the corresponding lemma in MultiWordNet and all the synsets in which the lemma appears were collected. Following this, the relevant do- mains for each tag were retrieved through the synset ids (in Multi- WordNet, domains are associated with synsets). With this mapping, the distribution of domain labels in the folkson- omy were investigated and compared with the institutional labels, both at site and category level. In order to compare the domains associated with the folksonomy with the institutional domains as- sociated with the website categories the folksonomy domains were ranked according to their frequency. Table3.3 gives the ranking of domains according to their frequency of association to the user tags at site level. As expected, the top domain is Factotum (44.97%), the domain assigned to lemmas having a generic meaning. Beside other generic domains, such as “Quality” (5.35%) or “Person” (5.27%), most of the domains in this rank, such as “Geography” (5.46%), “Military” (4.24%), “Politics” (4.04%) and Art (3.66%) are highly relevant to the themes dealt with by the exhibitions, which narrate Italy’s complex evolution after its unification in terms of political, social and cultural aspects. So, if “Buildings” (5.64%) can be re- lated to the monuments and buildings that appear as pictures or location in many exhibits, “Administration” (4.21%) refers to the institutions and administrative regions mentioned all along the nar- ration of Italys historical evolution. In order to gather more insight, we extended the semantic compari- son of tags and institutional domains to the highest level of detail of the categories (each exhibition includes several categories). How- ever, since most data in the tagset concerned the exhibition “Fare gli Italiani”, we limited the categorylevel comparison to this exhi- Table 2: Most tagged items Item Exhibition/Section Number of tags “Adunata” “Fare gli italiani” 34 tags (8.42%) “Calendario 2011” “Fare gli italiani” 16 tags (3.96%)* “Blog 150 anni insieme” “Fare gli italiani” 12 tags (2.97%)* “Cibo per le feste” “Extra contents” 12 tags (2.97%)* “L’Italia, Selargius, la Sardegna nei 150 anni” “extra contents” 12 tags (2.97%)* “Carabiniere in alta uniforme e Corazziere” “Il futuro nelle mani” 10 tags (2.48%) “Il pane, simbolo dell’unita” “Stazione futuro” 9 tags (2.23%) “Foggia e l’Unita’ d’Italia” “Extra contents” 8 tags (1.98%)* “La classe....di una volta” “Fare gli italiani” 8 tags (1.98%)* “Rete-Mondo Moda” “The places” 8 tags (1.98%)* Rank Domain Hit number 1 Factotum 1092 (44.97%) 2 Buildings 137 (5.64%) 3 Geography 133 (5.46%) 4 Quality 130 (5.35%) 5 Person 128 (5.27%) 6 Military 103 (4.24%) 7 Administration 102 (4.21%) 8 Politics 98 (4.04%) 9 Art 89 (3.66%) 10 Sociology 84 (3.45%) Table 3: Ranking of domains according to the matching tags. bition. “Fare gli Italiani” contains 20 categories, with an average of about 3.68 labels for each category. By comparing the domains associated with each category to those associated with tags added to contents of the same category, we found a significant overlap. The comparison showed that the average overlap between folkson- omy domains and institutional domains is 61.02%, that is, 61.02% of the domains extracted from the folksonomy of a certain category matched the institutional domains in the corresponding category. More interestingly, for each category the topmost ranked domain in the folksonomy (the domain that was associated with most tags in the category) always matches one of the institutional domains. For example, the topmost ranked domain in the folksonomy for the category “The Migrations” is “Geography”: this domain, together with “Sociology”, was associated by the curator with that category. The evaluation methodology described has two obvious limitations. Firstly, the limited coverage of the folksonomy by MultiWordNet: of the 1957 tags in the folksonomy, only 865 can be found in Multi- WordNet (44.2%). Secondly, we did not perform any kind of dis- ambiguation of tags, so the representation of the folksonomy in terms of domains may reflect the ambiguity of tags. Notwithstand- ing these limitations, we considered the overlap of the institutional and the folksonomy domains satisfactory, so the tag recommenda- tion strategy was maintained, extending it to work with no text input by the user: in the current version of the website, the input to the tag recommender can be given explicitly, by inserting one or more tags, or implicitly, by requesting the systems to suggest tags based on the tags associated with the current item. 4 Experimental Evaluation During a few months, students and teachers visited the large and permanent exhibition “Fare gli italiani” (“Making Italians”), and then attended one of the three post hoc laboratories where they were asked to interact with the 150 Digit portal. These laborato- ries constituted the basis for a thorough evaluation of the system Laboratory Control Experimental On the trail of migrants 11 10 The suitcase of the historian 2 3 Making Italians E-book: 2 3 Table 4: Laboratory sessions attended, with numbers of the control and experimental groups, respectively. with the goal of assessing the acceptance of the tag recommender and the effectiveness of the Web 3.0 (a Social Semantic Web ap- proach, borrowing the words of [Markines et al. 2009]) over the Web 2.0 approach in tag recommendation. During the lab sessions, the students were requested to analyze the items related to the ex- hibits and to create and upload new materials, also adding tags in the meantime. In the experimental group, the students interacted with the system, receiving tag and content recommendations, while for the control group, the recommendation module was disabled in order to gather the control data. Notice that the users in the control group could not use the tag recommender, but were able to see the other users tags (in the style of Web 2.0), when adding new tags to the contents. In the following, we analyze the results, evaluating the impact of our approach on • the tagging activity of users, i.e. the quantity and the typology of the inserted tags; • the semantics they convey. The latter point implies an analysis of the tag sets generated by the two versions of the system: one is completely user-generated, while the other one mixes user-generated tags and semantic recommenda- tions. 4.1 Tag Analysis: Methodology and Results Hypothesis. We hypothesized that the tag set generated by the ex- perimental group would be quantitatively and qualitatively different from the tag set generated by the control group. In particular, the folksonomy of the experimental group, having benefited from the tag recommender, should be larger and more heterogeneous. Design. The first group of students (control group) interacted with a modified version of the system, without the semantic recommender (independent variable). In this phase subjects received only social recommendation, namely the recommended tags that are the most used tags. The second group (experimental group) of students in- teracted with the regular version of the system, with the semantic recommender enabled. Participants. 15 classes (9 secondary school classes, and 6 high school classes) for the control phase with 33 registered users3 on the web site vs. a total of 298 users participating to the laboratories. 16 classes (10 secondary school classes, and 6 high school classes) for the experimental phase with 42 registered users vs. a total of 255 users participating to the laboratories. Apparatus and Materials. The laboratory rooms were equipped with a set of computers (Windows-based PCs) and one Interactive whiteboard (mainly for the lab conductor). Users browsed the web site using MS Explorer 8. The performances were traced by means of an ad hoc logging system. Users were given written instructions. Procedure. The classes used the system during one of the follow- ing laboratories: • On the trail of migrants: Through multimedia workstations students faced the journey of Italian emigrants of the early twentieth century: the baggage, the passenger list, landing at Ellis Island and the interrogation. Every student group took the role of a (emigrant) character. Then, with a leap through time, they relived the journey of new migrants from Guinea, Afghanistan, Kurdistan and other countries different from Italy. At the end of the laboratory students had to tag the characters on 150 Digit; • The historian’s suitcase: Before visiting the exhibition, by means of 150 Digit, students were introduced to the narra- tive choices made by the historians that curate the exhibition. They then chose an exhibition category and looked for its his- torical sources. At the end of the lab they uploaded (on 150 Digit) and tagged a document containing their work; • Making Italians E-book: In groups students were assigned specific roles to make observations, collect data and produce images along the way. After the visit, in 150 Digit, each group, created a humorous and original page of text and ani- mated images using Prezi editor, on the themes of the exhibi- tion or on how it was received. The e-book was then uploaded into the system and tagged. Table 4 summarizes which laboratories were attended during the control and the experimental phase. Each class was free to attend their preferred laboratory. In the first month of experimentation each class interacted with the modified version of the system, with- out the semantic recommender (control group). During the sec- ond month of experimentation each class interacted with the regu- lar version of the system, with the semantic recommender enabled (experimental group). In both conditions, students were required to analyze the items related to the exhibits and to create and up- load new materials, adding tags while doing both these operations. Users in the experimental group were asked to insert at least one of the suggested tags. It is important to note that users interacted with the system in groups. Every class was split into 2-4 groups. So the number of users registered into the system was less than the total number of real users attending the laboratories. Control group results. The 33 users of the control group inserted a total of 133 tags, with an average of 4.03 tags per registered user, and 8.06 tags per class. 10 users (24.24%) inserted almost half the tags (49.62%), in this way confirming the presence of active users in a community that are more active than others in content contributions [Nonnecke et al. 2004]. The same consideration ap- plies to classes: 4 classes (26.67%) inserted more than half the tags (55.64%). Table 5 shows the most used tags, namely all the tags 3Users interact with the system in group. Every class was split into 2-4 groups. Thus the number of registered users on the systems was less than the total number of users attending the laboratories. Tags Frequency Percentage Cosa Nostra 6 4.51 % mafia 6 4.51 % assault 5 3.76 % migration 5 3.76 % war 5 3.76 % emigration 4 3.01 % Clash 3 2.26 % collective act 3 2.26 % education 3 2.26 % Gold Rush 3 2.26 % immigration 3 2.26 % maid 3 2.26 % civil war 2 1.50 % conflict 2 1.50 % fight 2 1.50 % massacre 2 1.50 % murder 2 1.50 % Naples 2 1.50 % presentation 2 1.50 % Public Schools 2 1.50 % school 2 1.50 % Topic 2 1.50 % trench 2 1.50 % Table 5: Most used tags - Control Group Categories Frequency Percentage Migrations 46 34.59% Mafias 21 15.79% The First World War 18 13.53% The school 15 11.28% The power of the unity 10 7.52% The Second World War 9 6.77% The massmedia 5 3.76% The Church 3 2.26% The campaigns 3 2.26% The consumption 1 0.75% Italy cities 1 0.75% Painters and Patriots (2011) 1 0.75% Table 6: Most tagged categories - Control Group used more than once. Notice that i) the two most used tags are syn- onyms (i.e. “Cosa Nostra” and “mafia”); ii) the other top-ranked tags, i.e. migration-emigration-immigration, and assault-war-clash, share the same meaning; iii) these mentioned tags reflect the con- tent of the most tagged categories, namely “Migrations”, “Mafia”, and “The First World War”; iv) finally, a great number of tags (62, which is 46,62% of the total number of tags) were used only once, which shows great sparsity in the use of tags. The control group classes uploaded 20 new contents in total (5 new contents per class). Only 4 classes uploaded new contents since the “On the trail of mi- grants” laboratory does not require users to upload material. Table 6 shows the most tagged categories, which clearly reflect the themes of the most attended laboratories. Experimental group results. The 42 registered users of the experimental group inserted in total 214 tags, with an average of 5.09 tags per user, and 13.37 tags per class. 8 users (19.05%) inserted almost half the tags (49.07%), thus confirming also in this case the presence of active users in a community [Nonnecke et al. 2004]. The same consideration applies to classes: 5 classes (31.25%) inserted more than half the tags (54.87%). Table 7 gives the most used tags, namely all the tags used more than twice. Tags Frequency Percentage farmer 8 3.74% worker 6 2.80% merchant 5 2.34% illiterate 4 1.87% Migrant 4 1.87% Cameo 3 1.40% maid 3 1.40% miner 3 1.40% mother 3 1.40% poor 3 1.40% shopkeeper 3 1.40% unemployed 3 1.40% Veteran 3 1.40% well-off 3 1.40% woman 3 1.40% young 3 1.40% Table 7: Most used tags - Experimental Group The tags used once are 108 (50.47%), while tags used twice are 10.75%. Thus most tags (more than 60%) are re-used very little or not at all. Concerning the most used tags, i.e. farmer, worker, merchant, illiterate, migrant, etc., we should notice these are mainly the tags reflecting the subjects presented in the “Migration” category. The experimental group classes uploaded in total 21 new contents (3.5 new contents per class). Notice that just 6 classes uploaded new contents, since the “On the trail of migrants” laboratory does not require to upload material. Table 8 shows the most tagged cate- gories, which reflect the themes of the most frequented laboratories. For what concerns the slider that regulates the number of recom- mended tags, its use was analyzed in terms of the positions of the cursor selected by users when they accepted the recommendations. The hypothesis was that an even distribution of the cursor posi- tions would confirm the users’ understanding and acceptance of the slider: • Tags from position 1 were selected 84 times (39.25%): in this position, the expansion strategy considers the tighter seman- tic relations encoded in MultiWordNet, i.e., synonymy, and the candidate tags are disambiguated against the domains at- tached to the current category; • Tags from position 2 were selected 54 times (25.23%): in this position, the expansion is extended to the hyponymy and hy- peronymy relations, but the disambiguation is also extended to take the context of the existing tags into account; • Tags from position 3 were selected 40 times (18.69%): in this position, the domain–based disambiguation is removed; • Tags from position 4 were selected 36 times (16.82%): in this position, both disambiguation methods are removed; As shown by the usage of the slider tool, users seem to accept and understand its usage correctly, since all the cursor’s positions on the slider were employed, with an obvious prevalence of the initial position. 4.2 Tag Recommender Evaluation Using the same approach described in Section 3.3, the semantic rep- resentation of the two tag sets (one generated by the control group and the other by the experimental group), was compared and given Categories Frequency Percentage Migrations 150 70.09% The Futurist 17 7.94% The New Officine 11 5.14% The mafia 6 2.80% Campaigns 6 2.80% Unification of Italy 4 1.87% The First World War 4 1.87% It began with their 4 1.87% Consumption 3 1.40% Gallery of shops 3 1.40% Painters and Patriots (2011) 2 0.93% Officine Grandi Repairs 2 0.93% The Second World War 2 0.93% Table 8: Most tagged categories- Experimental Group in terms of the domains associated with the tags they contain. For each tag set, we computed the overlap of its domains with the insti- tutional domains, category by category (since the association with domains is at category level, i.e., each category was associated by the curators with a set of domains). For each tag, we collected the domains to which it is associated (us- ing the synset-domain mapping encoded in MultiWordNet ). For each category of the exhibition, we obtained a set of domains, (the folksonomy domains) ranked according to the number of tags to which each domain is associated4. Each set of ranked domains con- stitutes a rough semantic representation of the tag set in terms of the taxonomy of domains encoded in MultiWordNet. Tables 10 and 9 report the top ten domain labels for each tag set. The control group tag set refers to 44 different domains; the experimental group tag set of the refers to 55 different domains. Each domain is accompanied by the number of times it is associated with a tag in the folkson- omy (“Hit number” in the tables). For each set, some tags could not be employed for this evaluation because they are not present in MultiWordNet (for example, because they are proper nouns, neolo- gisms, etc.) or because they dont have a domain label associated in MultiWordNet. The two sets of ranked domains were compared. First of all, in order to assess if the two sets of domains (obtained in the two ex- perimental sessions) were significantly different from the statistical point of view, the χ2 test was calculated. The statistic shows that the difference in the two distribution are significant (χ2(67)=108.54, p<0.001). A more thorough comparison shows that the two sets have 29 do- mains in common; also, it was observed that only 4 of the 29 shared domains are in the ten topmost ranked domains of both sets (“Soci- ology”, “Psychological features”, “Military”, “Person”). By extending the comparison to the categories, the χ2 test shows that only the different frequencies of distribution of domains be- longing to “Migrations” and “The First World War” are significant, obtaining respectively χ2(40)=91.39, p<0.001, and χ2(13)=52.60, p<0.001. Secondly, the overlap of each set of folksonomy domains with the set of institutional domains was measured. Similarly to what emerged from the dataset collected during the system tuning phase (described in Section 3.3), the top ranked domains in the folkson- omy overlap with the institutional domain labels. Moreover, by observing the overlap at category level, we found that, in both tag 4Again, notice that the same tag can be associated to different domains due to its intrinsic polysemy (i.e., a tag belonging to more than one synset) or because a synset is associated to more domains. Domain Hit number person 113 (21.86%) commerce 21 (4.06%) military 21 (4.06%) agriculture 17 (3.29%) sociology 17 (3.29%) psychological features 15 (2.90%) quality 15 (2.90%) biology 13 (2.51%) medicine 12 (2.32%) animals 10 (1.94%) Table 9: Top ranked domain labels in experimental group. sets, the topmost ranked domains match at least one of the institu- tional domains assigned to that category, notwithstanding the low number of domains associated with the categories by the curators (an average 3.68 domains per category). For example, in the “Mi- grations” category, the domain in common between the folksonomy and the institutional domains is “Sociology” (17 hits in the control group and 10 in the experimental group) where the institutional do- mains of the category are “Geography” and “Sociology”); for the “The Second World War” category, the most frequent domain (9 hits in the control group and 6 in the control group) is “Military”, which matches the institutional domains associated with that cat- egory (“Diplomacy”, “History”, “Military”). The only exceptions are given by some categories for which less than 5 tags were col- lected. To summarize, although the two distributions of domains are statistically different, they are not significantly different in terms of their overlap with the institutional domain labels. However, it is necessary to specify that a full comparison is impossible. Due to the uneven distribution of tags in the category and to the partial coverage by MultiWordNet , the comparison was possible for only 6 categories out of 35; in all other cases, the tags of the users could either not be found in MultiWord-Net or there were no tags at all in that category (for one of the two tag sets). The analysis of the semantics of the folksonomy conducted through the domains confirms the prototype analysis findings (Section 3.3), i.e., that the folksonomy domains significantly overlap with the in- stitutional domains. In summary, given the domain based analysis, we can conclude that the assumption on which the recommendation strategy relies, that the institutional domains match the semantics of the site categories as perceived by the users, is confirmed by the data. As for the effectiveness of the recommendation strategy with re- spect to the baseline (no recommendation at all), no significant dif- ferences in the overlap of folksonomy domains with institutional domains were found between the control and the experimental group; i.e., the overlap does not grow using the recommender. In this way the tag recommendation can seem ineffective. However, the limited size of the tag sets collected during the experimentation makes the comparison difficult, and although a statistical difference does exist, it cannot be traced back to the overlap with the institu- tional domains. On the other hand, this result can be considered a success in an educational context, because tag insertion in focused trials driven by teachers leads students to select the tags in a very accurate and focused way (with and without the recommenders), and this may have contributed to reducing the difference between the two situations. 5 Discussion This section summarizes the results of the evaluation and analyzes the tag sets of the experimental and control groups in further de- Domain Hit number sociology 52 (13.07%) military 37 (9.3%) person 34 (8.54%) history 23 (5.78%) school 17 (4.27%) law 15 (3.77%) pedagogy 14 (3.52%) psychological features 10 (2.51%) time period 8 (2.01%) linguistics 7 (1.76%) Table 10: Top ranked domain labels in control group. tail, with the goal of explaining the difference between the two sets emerged from the domain-based analysis. As can be expected, users in the experimental group inserted more tags than users in the control group (5.09% tags vs. 4.03% tags per user), showing that the use of the recommender helps users find more tags. Also, users in the experimental group tended to insert more new tags compared to the control group: there are 85 (64%) distinct tags in the control group tag set, and 147 (69%) distinct tags in the experimental group tag set. This finding is positive for the evaluation of the tag recommender, because it shows that its use can increase the variety of tags. However, this is not clear-cut: since users were requested to insert tags at the end of each laboratory session and to reason about which tag to insert (e.g., they were explicitly asked to tag the discussed artworks and the uploaded material), the number and the variety of inserted tags may have been biased by these given instructions, blurring the differences between the two groups. In addition to this, the laboratory context and the presence of the teachers may have led students to be more accurate in tagging. Concerning the emergent semantics of the generated folksnomy, the number of domains referred to by the tags in the experimental tag set is higher than the number of domains referred to by the tags in the control tag set (55 domains vs. 44 domains). Even if these results do not show a clear benefit in the use of the tag recommender (again, also due to the structure of the experimental context), the number and distribution of tags in the experimental group seem to have been positively affected by the use of the tag recommender. To better understand the meaning of the tags inserted by the users, further semantics may be required. In order to gain more insight, two more investigations were conducted: first, the tags were an- alyzed according to their lexical types, trying to register the dif- ference in the two tag sets; second, the distribution of tags in the WordNet Domains in the two tag sets were compared, trying to re- late it to the use of the tag recommender. To study the distribution of tags based on their lexical types, the collected tags were clustered into WordNet categories. A semi- automatic procedure5 performed a WordNet dictionary lookup to obtain the top-level categories that could be deduced from the cor- respondence with the lexicographer file organization6. Only tags contained in the WordNet dictionary were mapped to WordNet cat- egories. The obtained classification for control and experimental group is detailed in Table 11. 5Disambiguation was performed manually. Each times a tag belongs to more than one category the right meaning was manually checked by con- fronting the tag and tagged item. 6http://wordnet.princeton.edu/man/lexnames.5WN.html Lexnames Experimental group Control group adj.all 31 2 adj.pert 1 0 compound 7 6 invented 12 0 noun.act 19 45 noun.animal 4 0 noun.artifact 14 7 noun.attribute 2 1 noun.cognition 2 10 noun.communication 3 5 noun.event 3 8 noun.feeling 7 0 noun.food 1 1 noun.group 6 18 noun.location 7 10 noun.object 0 1 noun.person 77 15 noun.plant 0 1 noun.relation 1 0 noun.state 12 0 noun.substance 2 1 noun.time 0 1 NULL 1 1 verb.stative 2 0 Table 11: Comparison between Wordnet lexnames The χ2 test was calculated in order to assess if the frequency of distribution of tags obtained in the two experimental session were significant. The statistics show that the difference in the two dis- tributions is significant (χ2(23)=125.47, p<0.001). What seems to emerge from this classification and from the comparison of the results of the two experimental conditions is that the presence of the semantic recommender led users to tag adjectives frequently (14.5%) and nouns denoting people (noun.person: 36%), such as farmer, merchant, worker, etc very frequently. The use of these words may be partly explained by the topics covered in “On the trail of migrants” lab, which focuses on people. However, this was the most frequented laboratory, as well as the “Migrations” cate- gory, in both groups (not only in the experimental one), and the most used tags in the control group do not reflect this trend. In- deed , students in the control group often used nouns denoting acts or actions (noun.act: 33.8%), and nouns denoting groups of people or objects (noun.groups: 13.51%) more than nouns denoting peo- ple (noun.person: 11.28%). From this observation we can conclude that the semantic recommendations of tags promote a more pre- cise description of the subject represented in the artwork. This is demonstrated by the wide use of tags belonging to the Wordnet per- son category. This finding also confirms the results obtained in the Steve.museum project [Trant 2006], according to which users pre- fer tags related to the subject of the artwork and to the description of what is represented in it. As illustrated in the previous section, the comparison of the two tag sets conducted by analyzing their implicit semantics through Multi- WordNet Domains confirmed the statistical difference between the two sets, implying that the use of the tag recommender actually affects the tagging behavior of users, although it does not prove that the overlap of the folksonomy is increased with the use of the recommender. In particular, we tried to relate the distribution of domains in the tag set of the experimental group with the function- ing of the tag recommender. Remember that, by using a slider, the users can affect the behavior of the recommender, reducing or in- creasing the number of recommended tags (see Section 3.2). In the starting position, the recommendation strategy is more restric- tive, as it applies different disambiguation techniques, which are progressively removed as the slider is moved to the following posi- tions. In parallel, the search for recommended tags is extended from tighter semantic relations (synonymy) to looser ones (hyponymy and hyperonymy). As a result, in the the last slider position, no dis- ambiguation is performed and all available semantic relations are considered. By analyzing the distributions of domain labels that emerged from the two sets of tags (control and experimental group), we note that the difference resides mainly in two facts: • In the experimental group, when the recommender was in use, there were 26 more domains (such as “Exchange”, “Adminis- tration”, etc.) than in the control group (there are 16 shared domains in the two distributions).These domains tend to be more specific than the shared ones; • In the experimental group, the most abstract domains in the hi- erarchy have more tags related to them than the control group. For example, “Person” moves from 34 related tags in the con- trol group tag set to 113 related tags in the tag set of the ex- perimental group. These two complementary findings show that the growth observed in the experimental tag set is related both with the diversity of do- mains it contained (with more specific domains added) and with the distribution of the tags in the domains, with more of the tags associ- ated with the generic domains. The users generated more tags with the help of the recommender, and this increase affected both the generic and the specific domains, which increased in number due to the focalization brought by the use of the recommender. These findings are consistent with the use of the tag recommen- dation slider. The increased number of specific domains can be attributed to the 39.25% of times the users accepted the recom- mended tag with the position “A” in the slider: in this position, tags are constrained to the institutional domains, which are generally quite specific, such as “Telecommunications” or “Transport”. The higher number of tags related with generic and abstract domains (such as “Psychological features”, or “Person”) are consistent with the fact that the remaining positions of the slider expand the existing tags along the hyperonymy relations encoded in MultiWordNet, in which 16.82% cases where no disambiguation was conducted (with the slider in D position, where the highest number of recommended tags is generated by the system). Last, but not least, a difference between the two tag sets lies in the domain/tag relationship. For example, if we consider the most tagged category, “Le migrazioni” (Migrations), we observe that in the tag set generated by the experimental group, tags and categories tend to increase proportionally. Notwithstanding this growth, the set of domains emerging from this tag set remain consistent with actual semantics of the category: for example, consider the domain “metrology”, brought in by the tag related with the distances cov- ered by the migrants, or “religion”, which is related to the descrip- tion of the ethnic groups involved in the migrations. Given the analysis conducted, we can conclude that a mixed ap- proach, which relies on leveraging the curatorial knowledge (en- coded in semantic format) to support social functions such as tag- ging, is suitable to the development of active learning environments. The use of the tag recommender contributes to keeping the tagging activity of the educational users aligned with the categorization of contents provided by the curators, at the same time sustaining the growth of the obtained folksonomy and its focalization in terms of variety. By promoting the linguistic reflection on the use of tags, the recommender helps the teacher focalize the students’ work with the beneficial input of the curatorial knowledge. 6 Conclusions In this paper we described “150 Digit”, a web portal on the 150th anniversary of the Unification of Italy, which has been designed and implemented with a Social Semantic Web approach in mind. The 150 Digit portal displays the contents of the exhibitions organized in 2011 for the anniversary, and encourages the active participation of the educational users through social functionalities such as tag- ging, voting, and commenting, and through the creation and upload of new contents. A 3D visit contributes to making the environment more compelling for students. In particular, we described the semantic-based tag recommenda- tions incorporated in “150 Digit”. The users’ tagging activity intro- duces new connections over the site contents, expanding and com- plementing the categorization of the exhibitions provided by the cu- rators. The tag recommender exploits the semantic description of the curatorial categorization of contents to sustain the growth of the folksonomy, keeping it aligned with the institutional perspective, with benefits for the focalization of the activities in the schoolwork. After describing the design of “150 Digit”, mainly based on an it- erative, user–centered methodology, we illustrated the evaluation conducted of the tag recommendation function, discussing its re- sults and possible benefits for learning environments. Given the users’ activity logs collected for 150 Digit, which in- clude tagging, commenting, uploading and saving content, current methodologies (for example, see [Ferreira-Satler et al. 2011]) con- sent the use of this information to generate more adaptive recom- mendations. As future work, we envisage the improvement of the tag recommender by extracting user profiles from the users’ behav- ior. Acknowledgments This work was partially supported by the project “150digit. L’Italia delle scuole” funded by Comitato Italia 150. The authors would like to thank: Comitato 150 and Esperienza Italia 150 (especially (Clau- dia Cugnasco and Marina Bertiglia), CSL – Università di Firenze; Indire – MIUR, Firenze; CIRMA – Università di Torino. This work was carried out with Virtual Reality & Multi Media Park S.p.A. (Fabrizio Nunnari, Marco Squeo, Shanti May, Davide Di Giannan- tonio). We also thank the users who took part in the focus group and in the tests and the anonymous reviewers for their help and advise. 7 Authors Rossana Damiano, PhD, is an Assistant Professor at the Depart- ment of Computer Science of the University of Torino, Italy. Her interdisciplinary research activity is centred on intelligent applica- tions for cultural heritage, ranging from interactive multimedia sys- tems to applications of virtual agents. In particular, she is interested in the role of semantic models in new media production and appli- cations. Vincenzo Lombardo, PhD, is an Associate Professor of Informat- ics at the University of Torino, Italy. He works on models and ap- plications of knowledge-based system to multimedia communica- tion. He carries on a production activity in multimedia production, hosted by events at the international level. Cristina Gena is an Assistant Professor at the Department of Com- puter Science of the University of Torino, working in the area of intelligent user interfaces. She completed her Ph.D. in Communi- cation Science (University of Torino) in 2003, with a thesis on the evaluation of user-adaptive systems. Her current research activities address user modeling, adaptive web systems and their evaluation, context-aware systems, semantic web, web 2.0, usability and in- teraction design. Her contribution is based on experiences gained within these fields. References AMES, M., AND NAAMAN, M. 2007. Why we tag: motivations for annotation in mobile and online media. In CHI, ACM, M. B. Rosson and D. J. Gilmore, Eds., 971–980. BATEMAN, S., BROOKS, C., MCCALLA, G., AND BRUSILOVSKY, P. 2007. Applying collaborative tagging to e-learning. In Proceedings of the Workshop on Tagging and Metadata for Social Information Organization (WWW’07). BENTIVOGLI, L., FORNER, P., MAGNINI, B., AND PIANTA, E. 2004. Revising the wordnet domains hierarchy: semantics, cov- erage and balancing. In Proceedings of the Workshop on Mul- tilingual Linguistic Ressources, Association for Computational Linguistics, 101–108. BURIGAT, S., AND CHITTARO, L. 2007. Navigation in 3d virtual environments: Effects of user experience and location-pointing navigation aids. International Journal of Human-Computer Studies 65, 11, 945–958. CANTADOR, I., KONSTAS, I., AND JOSE, J. M. 2011. Categoris- ing social tags to improve folksonomy-based recommendations. J. Web Sem. 9, 1, 1–15. DAMIANO, R., GENA, C., LOMBARDO, V., NUNNARI, F., SUP- PINI, A., AND CREVOLA, A. 2011. 150 digit. integrating 3d visit and social functions into a web 3.0 learning-oriented ap- proach. In Broadband and Wireless Computing, Communication and Applications (BWCCA), 2011 International Conference on, IEEE, 136–143. DAMIANO, R., LOMBARDO, V., GENA, C., AND NUNNARI, F. 2012. Guidance for web 3d in cultural heritage dissemination. In Proceedings of the 17th International Conference on 3D Web Technology, ACM, 186–186. DEGEMMIS, M., LOPS, P., AND SEMERARO, G. 2007. A content-collaborative recommender that exploits wordnet-based user profiles for neighborhood formation. User Model. User- Adapt. Interact. 17, 3, 217–255. DJUANA, E., XU, Y., LI, Y., AND JOSANG, A. 2011. Ontol- ogy learning from user tagging for tag recommendation making. In Proceedings of the 2011 IEEE/WIC/ACM International Con- ferences on Web Intelligence and Intelligent Agent Technology - Volume 03, IEEE Computer Society, Washington, DC, USA, WI-IAT ’11, 310–313. FERREIRA-SATLER, M., ROMERO, F., MENENDEZ- DOMINGUEZ, V., ZAPATA, A., AND PRIETO, M. 2011. Fuzzy ontologies-based user profiles applied to enhance e- learning activities. Soft Computing-A Fusion of Foundations, Methodologies and Applications, 1–13. FURNAS, G., LANDAUER, T., GOMEZ, L., AND DUMAIS, S. 1987. The vocabulary problem in human-system communica- tion. Communications of the ACM 30, 11, 964–971. GANGEMI, A., GUARINO, N., MASOLO, C., OLTRAMARI, A., AND SCHNEIDER, L. 2002. Sweetening ontologies with dolce. Knowledge engineering and knowledge management: Ontolo- gies and the semantic Web, 223–233. GENA, C., CENA, F., VERNERO, F., AND GRILLO, P. Accepted for publication. The evaluation of a social adaptive web site for cultural events. User Model. User-Adapt. Interact.. HOTHO, A., JÄSCHKE, R., SCHMITZ, C., AND STUMME, G. 2006. Folkrank : A ranking algorithm for folksonomies. In LWA, University of Hildesheim, Institute of Computer Science, K.-D. Althoff and M. Schaaf, Eds., vol. 1/2006 of Hildesheimer Informatik-Berichte, 111–114. LANIADO, D., EYNARD, D., AND COLOMBETTI, M. 2007. A semantic tool to support navigation in a folksonomy. In Proceed- ings of the eighteenth conference on Hypertext and hypermedia, ACM, New York, NY, USA, HT ’07, 153–154. LAUDANNA, A., THORNTON, A., BROWN, G., BURANI, C., AND MARCONI, L. 1995. Un corpus dell’italiano scritto contempo- raneo dalla parte del ricevente. III giornate internazionali di analisi statistica dei dati testuali 1, 103–109. MAGNINI, B., STRAPPARAVA, C., PEZZULO, G., AND GLIOZZO, A. 2002. The role of domain information in word sense disam- biguation. Natural Language Engineering 8, 04, 359–373. MARKINES, B., CATTUTO, C., MENCZER, F., BENZ, D., HOTHO, A., AND STUMME, G. 2009. Evaluating similarity measures for emergent semantics of social tagging. In Proceed- ings of the 18th international conference on World wide web, April, Citeseer, 20–24. MILLER, G. 1995. Wordnet: a lexical database for english. Com- munications of the ACM 38, 11, 39–41. NILES, I., AND PEASE, A. 2003. Mapping WordNet to the SUMO ontology. In Proceedings of the IEEE International Knowledge Engineering conference, 23–26. NONNECKE, B., PREECE, J., AND ANDREWS, D. 2004. What lurkers and posters think of each other. Hawaii International Conference on System Sciences 7, 70195a. O’ DONOVAN, J. 2009. Capturing trust in social web applica- tions. In Computing with Social Trust, J. Golbeck, Ed., Human- Computer Interaction Series. Springer London, 213–257. PAZZANI, M. J., AND BILLSUS, D. 2007. Content-based recom- mendation systems. In The Adaptive Web, 325–341. PIANTA, E., BENTIVOGLI, L., AND GIRARDI, C. 2002. Develop- ing an aligned multilingual database. In Proc. 1st Intl Conference on Global WordNet. PREECE, J., NONNECKE, B., AND ANDREWS, D. 2004. The top five reasons for lurking: improving community experiences for everyone. Computers in Human Behavior 20, 2 (March), 201– 223. SARWAR, B., KARYPIS, G., KONSTAN, J., AND REIDL, J. 2001. Item-based collaborative filtering recommendation algorithms. In Proceedings of the 10th international conference on World Wide Web, ACM, 285–295. SCHAFER, J., FRANKOWSKI, D., HERLOCKER, J., AND SEN, S. 2007. Collaborative filtering recommender systems. The adap- tive web, 291–324. SZOMSZOR, M., CATTUTO, C., ALANI, H., O’HARA, K., BAL- DASSARRI, A., LORETO, V., AND SERVEDIO, V. D. 2007. Folksonomies, the semantic web, and movie recommendation. In Proceedings of the Fourth European Semantic Web Confer- ence, 71–85. TANG, J., HUI, S., ZHOU, B., FONG, A. C. M., AND HONG, G. 2012. Generation of personalized ontology based on consumer emotion and behavior analysis. IEEE Transactions on Affective Computing 3, 2, 152–164. TRANT, J. 2006. Exploring the potential for social tagging and folksonomy in art museums: Proof of concept. New Review of Hypermedia and Multimedia 12, 1, 83–105. TRANT, J. 2009. Studying social tagging and folksonomy: A re- view and framework. Journal of Digital Information 10, 1. VESIN, B., IVANOVI?, M., KLA?NJA-MILI?EVI?, A., AND BUDIMAC, Z. 2012. Protus 2.0: Ontology-based semantic rec- ommendation in programming tutoring system. Expert Systems with Applications 39, 15, 12229 – 12246. WANG, Y., STASH, N., AROYO, L., HOLLINK, L., AND SCHREIBER, G. 2009. Semantic relations for content-based recommendations. In K-CAP, 209–210. XU, Z., FU, Y., MAO, J., AND SU, D. 2006. Towards the seman- tic web: Collaborative tag suggestions. In Collaborative web tagging workshop at WWW2006, Edinburgh, Scotland, Citeseer. 
work_5kks6ebnnbfdhpdqj4inpbgo6e	----	Anomaly detection in monitoring sensor data for preventive maintenance HAL Id: lirmm-00670917 https://hal-lirmm.ccsd.cnrs.fr/lirmm-00670917 Submitted on 16 Feb 2012 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Anomaly Detection in Monitoring Sensor Data for Preventive Maintenance Julien Rabatel, Sandra Bringay, Pascal Poncelet To cite this version: Julien Rabatel, Sandra Bringay, Pascal Poncelet. Anomaly Detection in Monitoring Sensor Data for Preventive Maintenance. Expert Systems with Applications, Elsevier, 2011, 38 (6), pp.7003-7015. �10.1016/j.eswa.2010.12.014�. �lirmm-00670917� https://hal-lirmm.ccsd.cnrs.fr/lirmm-00670917 https://hal.archives-ouvertes.fr Expert Systems with Applications 38 (2011) 7003–7015 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Anomaly detection in monitoring sensor data for preventive maintenance Julien Rabatel a,b,⇑, Sandra Bringay a,c, Pascal Poncelet a a LIRMM, Université Montpellier 2, CNRS, 161 rue Ada, 34392 Montpellier Cedex 5, France b Fatronik France Tecnalia Cap Omega, Rond-point Benjamin Franklin - CS 39521, 34960 Montpellier, France c Dpt MIAp, Université Montpellier 3, Route de Mende, 34199 Montpellier Cedex 5, France a r t i c l e i n f o Keywords: Anomaly detection Behavior characterization Sequential patterns Preventive maintenance 0957-4174/$ - see front matter � 2010 Elsevier Ltd. A doi:10.1016/j.eswa.2010.12.014 ⇑ Corresponding author at: LIRMM, Université Mon 34392 Montpellier Cedex 5, France. E-mail addresses: rabatel@lirmm.fr (J. Rabatel), b poncelet@lirmm.fr (P. Poncelet). 1 http://www.smartmotorist.com. a b s t r a c t Today, many industrial companies must face problems raised by maintenance. In particular, the anomaly detection problem is probably one of the most challenging. In this paper we focus on the railway main- tenance task and propose to automatically detect anomalies in order to predict in advance potential fail- ures. We first address the problem of characterizing normal behavior. In order to extract interesting patterns, we have developed a method to take into account the contextual criteria associated to railway data (itinerary, weather conditions, etc.). We then measure the compliance of new data, according to extracted knowledge, and provide information about the seriousness and the exact localization of a detected anomaly. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction Today, many industrial companies must face problems raised by maintenance. The most common solution is called curative maintenance, i.e., equipment is replaced or repaired after the appearance of obvious failures, once the damage is occurred. This solution poses many problems. Curative maintenance is too be- lated and is particularly costly on several aspects. On the financial side first, for many companies, a few hours of downtime can result in millions of dollars in losses. It is generally much less expensive to make predictive maintenance to prevent a serious breakdown. In addition, the corrective maintenance is also a problem for secu- rity aspects. In many sensitive areas, equipment failures can cause death. For example, it is estimated that approximately 5% of motor vehicle accidents are caused by equipment malfunction or a lack of maintenance.1 Another aspect is related to the environment and energy saving. Indeed, equipment that is worn or subject to mal- functions often consumes more energy than equipment that oper- ates optimally. In addition, a systematic maintenance planning is not a satisfactory solution as too expensive compared to real needs. To reduce the problems of equipment maintenance, to propose ways to make maintenance both faster and more effective by antic- ipating serious breakdowns represents a particularly critical issue. Such a preventive maintenance consists in detecting anomalous behavior in order to prevent further damages and avoid more ll rights reserved. tpellier 2, CNRS, 161 rue Ada, ringay@lirmm.fr (S. Bringay), costly maintenance operations. To this end, it is necessary to mon- itor the working equipment. Usually, monitoring data is available through embedded sensors and provides us with important infor- mation such as temperatures, humidity rates, etc. Nevertheless, data collected by sensors are difficult to exploit for several reasons. First, a very large amount of data usually available at a rapid rate must be managed to provide a relevant description of the observed behaviors. Furthermore, they contain many errors: sensor data are very noisy and sensors themselves can become defective. Finally, when considering data transmission, very often lots of information are missing. In this paper, we focus on the field of train maintenance. Trains monitoring is also ensured by sensors positioned on the main components (wheels, motors, etc.) to provide much information (temperature, acceleration, velocity). In this context, we are subject to the difficulties we have described above: voluminous and noisy data, information transmission problems, etc. Moreover, it is important to take into account the different types of data available. We therefore wish to propose a method to exploit this informa- tion in order to assist the development of an effective predictive maintenance. The needs in the context of train maintenance are twofold. First, it is important to provide a better understanding of monitored sys- tems. Indeed, as they are often complex and contain many compo- nents, the experts have little knowledge about their actual behavior. This lack of knowledge makes the problem of mainte- nance very difficult. From another point of view it could be inter- esting to get an overview of the normal behavior (e.g., in case of monitoring) and then it is necessary to propose a way for charac- terizing such normal behaviors from a huge amount of historical data. Another challenging point that we must consider is that http://dx.doi.org/10.1016/j.eswa.2010.12.014 mailto:rabatel@lirmm.fr mailto:bringay@lirmm.fr mailto:poncelet@lirmm.fr http://www.smartmotorist.com http://dx.doi.org/10.1016/j.eswa.2010.12.014 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 7004 J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 normal behavior strongly depends on the context. For example, a very low ambient temperature will probably affect a train behav- ior. Similarly, each itinerary with its own characteristics (slopes, turns, etc.) influences a journey. Consequently it is essential, in or- der to efficiently characterize the behavior of trains as well as to detect anomalies, to consider the surrounding context. In this pa- per, we will show how these elements can be directly addressed by data mining techniques and how they can be used to design a system for detecting anomalies in train behavior and help experts so that a detected problem can be dealt as promptly as possible, and the right decisions can be made. Our approach follows the framework presented in Fig. 1. We ad- dress the two issues mentioned above: (i) the knowledge discovery process about normal train behavior and (ii) the anomaly detection in new data. The characterization of normal behavior is divided into three steps. First, we consider the data describing the normal behavior (i.e., containing no anomalies) that were recorded in the past. These so-called historical data are segmented into different classes, which are defined by the context in which the data were recorded. This organization brings together all trips that were conducted un- der similar conditions. The criteria used to create the classes are, for example, the outside temperature, the itinerary, etc. Then, we extract the most frequent behaviors to characterize each of these classes. We thus obtain knowledge classes that describe very pre- cisely normal train behavior and also provide us with essential information about the impact of contextual criteria. For example, we can answer questions such as ‘‘What are the behaviors that are specific to a high outside temperature?’’. After having characterized the train behavior in the first step of our framework, we wish to detect anomalies in newly recorded data. To this end, we use the previously obtained knowledge. We have developed a method to compare new monitoring data with a class of knowledge. Thus, we can detect critical events and notify experts that a maintenance operation may be necessary. This paper is organized as follows. Section 2 describes the data representation in the context of train maintenance. Section 3 shows the characterization of normal behaviors by discovering sequential patterns. Then we present the anomaly detection for predictive maintenance approach in Section 4. Experiments con- ducted with real and simulated data are described in Section 5 and related work is presented in section 6. Finally, we conclude in Section 7. Fig. 1. General 2. Data preprocessing In this section, we address the problem of preprocessing railway data. From raw data collected by sensors, we design a suitable rep- resentation for data mining tasks. The train monitoring system exploited in this study is such that a large set of sensors is distributed on the main components of each monitored train. The key element of this system is the bogie, because failures and anomalous behaviors are most of the time associated with it. Each of the 8 bogies of a train has 32 sensors col- lecting information such as temperatures, accelerations and veloc- ity. Every five minutes during a journey, all sensors collect a value stored in a central database. A complete description of these data is available in Carrascal, Díez, and Azpeitia (2009). 2.1. Sensor data for train maintenance The data resulting from sensors for train maintenance is com- plex for the two following reasons: (i) very often errors and noisy values pervade the experimental data; (ii) multi-source informa- tion must be handled at the same time. For instance, in train main- tenance following data must be considered. 2.1.1. Sensors Each sensor describes one property of the global behavior of a train which can correspond to different information (e.g., temper- ature, velocity, acceleration). 2.1.2. Measurements They stand for numerical values recorded by the sensors and could be very noisy for different reasons such as failures, data transfer, etc. Note that the numerical values collected by the sen- sors are then discretized to obtain a set of data more suited to the step of data mining described in Section 3. 2.1.3. Readings They are defined as the set of values measured by all the sensors at a given date. The information carried out by a reading could be considered as the state of the global behavior observed at the given moment. Due to the data transfer, some errors may occur and then readings can become incomplete or even missing. framework. Table 1 Extract from sensor data. TIME A B C . . . 2008/03/27 06:36:39 0 16 16 . . . 2008/03/27 06:41:39 82.5 16 16 . . . 2008/03/27 06:46:38 135.6 19 21 . . . 2008/03/27 06:51:38 105 22 25 . . . J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 7005 We consider that handled data are such as those described in Table 1, where a reading for a given date (first column) is de- scribed by sensor measurements (cells of other columns). The behavior of a sensor is described by the numerical values it measures. The discretization of these sensor values is a preprocess- ing that, for example, has been used in Chong, Krishnaswamy, Loke, and Gaben (2008), Mihail Halatchev (2005) to group values expressing the same information into classes of values (e.g., a value could be low, medium or high). For example, temperature values 25�C and 26�C are different but very close, so they can be regarded as carrying the same information. Moreover, for velocity as well as acceleration sensors, we also consider a zero value because it is a particular value that can not be confused with non-zero values. For example, a zero velocity indicates that the train is stationary, while a non-zero velocity, no matter how low it is, concerns a train in motion. 2.2. Granularity in railway data Data collected from a train constitutes a list of readings describ- ing its behavior over time. As such a representation is not appropri- ate to extract useful knowledge, we decompose the list of readings at different levels of granularity and then we consider the three fol- lowing concepts journeys, episodes and episode fragments which are defined as follows. 2.2.1. Journey The definition of a journey is associated to the railway context. For a train, a journey stands for the list of readings collected during the time interval between the departure and the arrival. Usually, a journey is several hours long and has some interruptions when the train stops in railway stations. We consider the decomposition into journeys as the coarsest granularity of railway data. Let minDuration be a minimum duration threshold, maxStop be a maximum stop duration, and J be a list of readings (rm, . . ., ri , . . .rn), where ri is the reading collected at time i. J is a journey if: 1. (n � m) > minDuration, 2. 9=ðru; . . . ; rv; . . . rwÞ# Jj ðw � uÞ > maxStop; and 8v 2 ½u; w�; velocityðvÞ¼ 0: � 2.2.2. Episode The main issue for characterizing train behavior is to compare elements which are similar. However, as trains can have different routes the notion of journey is not sufficient (for instance, between two different journeys, we could have different number of stops as well as a different delay between two railway stations). That is the reason why we segment the journeys into episodes to get a finer level of granularity. Episodes are obtained by relying on the stops of a train (easily recognizable considering the train velocity). An episode is defined as a list of readings (rm, . . .ri, . . . , rn) such as: � velocity(m) = 0 and velocity(n) = 0,2 � if m < i < n, velocity(i) – 0. 2 Here, the velocity of the train at time t is denoted as velocity(t). Fig. 2 describes a segmentation of a journey into episodes by considering the velocity changes. This level of granularity is con- sidered as the most relevant because it provides us with a set of homogeneous data. However, we can segment episodes in order to obtain a more detailed representation and a finer granularity level. 2.2.3. Episode fragment The level of granularity corresponding to the fragments is based on the fact that the behavior of a train during an episode can easily be divided into three chronological steps. First, the train is station- ary (i.e., velocity 0) then an acceleration begins. We call this step the starting step. More formally, let E = (rm, . . . , rn) be an episode. The starting fragment Estarting = (rm, . . . , rk) of this episode is a list of readings such as: 8i; j 2 ½m; k�; i < j () velocityðiÞ < velocityðjÞ: At the end of an episode, the train begins a deceleration ending with a stop. This is the ending step. More formally, let E = (rm, . . . , rn) be an episode. The ending fragment Eending = (rk, . . . , rn) of this epi- sode is a list of readings such as: 8i; j 2 ½k; n�; i < j () velocityðiÞ > velocityðjÞ: The traveling fragment is defined as the sublist of a given epi- sode between the starting fragment and the ending fragment. During this fragment, there are accelerations or decelerations, but no stop. More formally, let E be an episode, Estarting its starting fragment, and Eending its ending fragment. Then, the traveling fragment of E, de- noted as Etraveling, is a list of readings defined as: Etraveling ¼ E � Estarting � Eending: Fig. 2 shows the segmentation of an episode into three frag- ments: the starting fragment, the traveling fragment and the end- ing fragment. From now we thus consider that all the sensor data are stored in a database, containing all information about the different granular- ity levels. For example, all the sensor readings composing the frag- ment shown in Fig. 2 are indexed and we know that a particular fragment f is an ending fragment included in an episode e, belong- ing to the journey J. J is associated with the itinerary I and the index of e in I is 2 (i.e., the second portion of this route). 2.3. Data classes Previously, we have presented how to process the data in the railway context in order to obtain a relevant representation. How- ever, we have not considered the contextual aspects of such data. Indeed, in the railway field of application as well as other industrial applications, the behavior strongly depends on different exterior parameters. The nature of such context is inevitably dependant on the field of application. For example, the behavior of a train during a trip de- pends on contextual criteria such as the weather conditions or its geographical position. We have previously exploited different lev- els of granularity in data, we now use different existing contexts to consider the variability of train behavior according to them. Example 1. Table 2 presents a set of fragments, identified by the id column. Each fragment is associated with contextual criteria (the humidity rate, the exterior temperature, the global route, and the index of the episode in this route). For example, the fragment f2 was performed with a low humidity rate and a low exterior temperature. In addition, f2 is part of the second episode (E2) of the itinerary denoted by I1. Table 3 Some contextual classes. Class Humidity Exterior temperature [⁄,⁄] ⁄ ⁄ [low,⁄] low ⁄ [high,⁄] high ⁄ [low, high] low high [⁄, high] ⁄ high . . . . . . . . . Fig. 2. Segmentation of a journey into episodes. Table 2 Fragments and contextual information. ID Humidity Exterior temperature Itinerary Index f1 low high I1 E1 f2 low low I1 E2 f3 high high I2 E1 f4 low low I1 E1 f5 high low I1 E2 7006 J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 2.3.1. Data classes Now we present how contextual criteria are handled, by divid- ing the data into classes according to the different contextual crite- ria. To this end, we will at first develop a formal representation of the different contexts that can be met. Based on the general principle of contextualization, here we de- fine how classes are constructed. We consider now a formal description of a context. Each fragment is described in a set of n contextual dimensions (e.g., Duration or Exterior Temperature), denoted by DC. Let c be a class, defined in DC. A class c is denoted by ½cD1 ; . . . ; cDi ; . . . ; cDk�, where cDi is the value of c for the dimension Di, Di 2 DC. We use a wildcard value, denoted by ⁄, which can sub- stitute any value on each dimension in DC. In other words, "X 2 DC, "x 2 Dim(X), {x} �⁄. Thus, an episode e belongs to a class c if the restriction of e on DC 3 is included in c: 8X 2 DC; eX # cX: We have identified a lattice structure, called the contextual lat- tice, to represent the set of contextual classes. We now explain how this structure is defined. 2.3.2. Contextual lattice The set of contextual classes can be represented in a multidi- mensional space containing all the combinations of different crite- ria as well as their possible values. Example 2. Table 3 shows some of the contextual classes corre- sponding to two dimensions: Humidity and Exterior Temperature (respectively denoted as Hum and ExtT below). A class c is denoted by [cExtT, cHum]. For example, the class denoted by [low,⁄] is equivalent to the context where the temper- ature is low (i.e., cExtT = low), for any humidity rate (i.e., cHum = ⁄). Using the dataset of Table 2, we can see that the set of fragments belonging to the class [low,⁄] is {f1, f2, f4}. Similarly, the set of fragments belonging to the class [low, high] is {f1}. 3 The restriction of e in DC is the description of e, limited to the dimensions of DC. Contextual classes and their relationships can be represented as a lattice. Nevertheless, we first have to define a generalization/ specialization order on the set of environmental classes. Definition 1. Let c, c0 be two classes. c > c0 ()8X 2 DC; c0X � cX . If c > c0, then c is said to be more general than c0 and c0 is said to be more specific than c. Moreover, if there is no class c0 such that c0 – [;, . . . ,;] and c > c0, then c is said to be a specialized class. In order to construct contextual classes, we provide a sum oper- ator (denoted by +) and a product operator (denoted by �). The sum of two classes gives us the most specific class generalizing them. The sum operator is defined as follows. Definition 2. Let c, c0 be two classes. z ¼ c þ c0 ()8X 2 DC; zX ¼ cX if cX ¼ c0X; � elsewhere: � Example 3. Given the set of classes in Table 3, we note that [low, low] + [low, high] = [low,⁄], and [low, high] + [high, low] = [⁄,⁄]. The product of two classes gives the most general class special- izing them. The product operator is defined as follows. Definition 3. Let c, c0 be two classes. The class u is defined as follows: 8X 2 DC; uX ¼ cX \ c00X . Then, z ¼ c � c0 () z ¼ u if 9= X 2 DCjuX ¼;; ½;; . . . ;;� elsewhere: � Example 4. Given the set of classes in Table 3, we note that [low, high]�[low,⁄] = [low, high], and [⁄, high]�[low, low] = [;,;]. We can now define a lattice, by using the generalization/specializa- tion order between classes and the operators defined above. The or- dered set hCS, > i is a lattice denoted as CL, in which Meet ( V ) and Join ( W ) elements are given by: 1. "C � CL, V C = +c2Cc 2. "C � CL, W C = �c2Cc Fig. 3. A contextual lattice. J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 7007 Fig. 3 illustrates an extract of the lattice of contextual classes of the dataset provided in Table 2. We can now very precisely index historical data in appropriate classes. 3. Normal behavior characterization In this section, we focus on the data mining step in the knowl- edge discovery process and more precisely on the extraction of pat- terns characterizing normal behavior. 3.1. How to extract normal behavior? The objective of the behavior characterization is, from a data- base of sensor measurements, to provide a list of patterns depict- ing normal behavior. We want to answer the following question: which patterns often appear in the data? Such a problem, also known as pattern mining, has been extensively addressed by the data min- ing community in the last decade. Amongst all the data mining methods, we cite the sequential patterns mining problem. Sequential patterns were introduced in Agrawal and Srikant (1995) and can be considered as an extension of the concept of association rule Agrawal, Imieliński, and Swami (1993) by handling timestamps associated to items. The sequential patterns goal is to extract sets of items commonly associated over time. In the ‘‘basket market ’’ concern, a sequential pattern can be for example: ‘‘40 % of the customers buy a television, then buy later on a DVD player ’’. In the following we adapt the sequential pattern mining problem to our sensor data. In the following definitions and examples, we will consider a set of sensors, denoted by X, and the set of possible values collected by a sensor A 2 X, denoted by dom(A). Definition 4. An item Av is a pair {A, v}, where A 2 X and v 2 dom(A). It stands for the value v collected by the sensor A at a given time. The item Av is called an A-item. Moreover, we denote by value(Av) the value v, and by sensor(Av) the sensor A. Definition 5. Let x = {A1, A2, . . . , Ai, . . . , An} a set of sensors such that x # X. An itemset I is a reading or a part of a reading at a given time, i.e., a non-ordered set of items, denoted as: I ¼ A1v 1 A 2 v 2 . . . Aiv i . . . A n v n � � : Note that if Aiv i 2 I; 9= A i v0 i j Aiv0 i 2 I. Indeed, a sensor cannot read two different values at the same time. Definition 6. A sequence s is a list of itemsets denoted as: s ¼hI1I2 . . . Ii . . . Ini; where the itemset Ii stands for the ith itemset of s. An empty sequence is denoted as ;. Example 5. Data described in Table 1, are translated into the fol- lowing sequence: hðA0B16C16ÞðA82:5 B16C16ÞðA135:6 B19C21ÞðA105B22C25Þi: When data are discretized, the sequence becomes: hðA0BlowClowÞðAavg BlowClowÞðAhighBavg CavgÞðAhighBavg ChighÞi: Definition 7. Given two sequences s = hI1, I2, . . . , Imi and s0 ¼ hI01I 0 2 . . . I0ni, if there exist integers 1 6 i1 < i2 <� � �< im 6 n such that I1 # I 0 i1 ; I2 # I 0 i2 ; . . . ; Im # I 0 im , then the sequence s is a subsequence of the sequence s0, denoted as s v s0, and s0 supports s. Table 4 Sample of sequences. ID Sequence sa h(AhighBhigh)i sb h(Ahigh)(AavgBavg)(Blow)(Alow) i sc h(Ahigh)(Blow)i sd h(Ahigh)(Bavg)i 7008 J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 Example 6. From Table 4, we can note that sc v sb and sd v sb. Definition 8. If a sequence s is not a subsequence of any other sequences, then we say that s is maximal. Example 7. From the set of sequences contained in Table 4, we can note that sa and sb are maximal sequences. Definition 9. Let s be a sequence. The length of s, denoted as jsj, is the number of itemsets in s. The size of s, denoted as ksk, is the number of items in s. Example 8. The length of sa is jsaj = 1, and its size is ksak = 2. Definition 10. The support of a sequence is defined as the fraction of total sequences in a sequence database DB that support this sequence. A sequence is said to be frequent if its support is greater than or equal to a threshold minimum support (minSupp) specified by the user. The sequential pattern mining problem is, for a given threshold minSupp and a sequence database DB, to find all frequent sequences in DB. 3.1.1. Contiguous subsequences mining To extract interesting patterns in a database of behavioral sequences, it is important to note that a frequent sequence is inter- esting only if the gap between each itemset is limited. By extract- ing frequent sequences in a database of behavioral sequences, we want to highlight the frequent interactions and correlations between sensors, but also between readings. However, the interest of those patterns is strongly associated with the temporal gap between each pair of itemsets in a sequence. To take this into consideration, we modify the concept of subsequence in order to consider only the consecutive itemsets in a sequence but first of all we define the notion of concatenation. Definition 11. The concatenation of sequences is denoted as s + s0, and the result is the sequence obtained by appending s0 to the end of s, so that we have js + s0j = jsj + js0j and ks + s0k = ksk + ks0k. Example 9. The sequence corresponding to the concatenation of sa and sc is: sa þ sc ¼hðAhighBhighÞðAhighÞðBlowÞi: Definition 12. A sequence s is a contiguous subsequence of the sequence s0, denoted as s v cs0, if there exist three sequences s1, s2, and s3 such that s0 = s1 + s2 + s3, jsj = js2j, and s v s0. Example 10. The sequence sd is a contiguous subsequence of sb, i.e., sd v csb. Indeed, there exist three sequences s1 = ;, s2 = h(Ahigh)- (AavgBavg)i and s3 = h(Alow)i, such that sd = s1 + s2 + s3, jsdj = js2j, and sb v s2. sb ¼ ; z}|{s1 þhðAhighÞðAavg BavgÞi zfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{s2 þhðBlowÞðAlowÞi zfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflffl{s3 ; sd ¼hðAhighÞðBavgÞi: 3.1.2. Aggregated sequences Handled behavioral data are highly redundant: the behavior described by the sensor measurements are generally stable for a reading to another. This characteristic is particularly visible for data that change slowly over time (e.g., a wheel temperature). Therefore, we are dealing with very large sequences with consecutive itemsets containing redundant information. Such sequences bring two problems: (i) the mining of this kind of se- quences is particularly difficult (the repetition of identical itemsets considerably increases the search space) and (ii) it yields no addi- tional information. We therefore propose the concept of aggre- gated sequence. Definition 13. Let s = hI1, I2, . . . , Ini be a sequence. The correspond- ing aggregated pattern, denoted by s⁄, is the maximal subse- quence of s respecting the following condition: s� ¼ hI�1 I � 2 . . . I � i . . . I � mi; such that 8I � i 2 s �; I�i – I � iþ1: Note that s⁄v s. Example 11. Let s = h(AlowBavg)(Alow)(Alow)(Bhigh)i. The aggregated pattern of s, denoted by s*, is: s� ¼ hðAlowBavgÞðAlowÞðBhighÞi: 3.2. Contextualized characterization We have seen in Section 2 that normal train behavior is related to contextual parameters. We have described the necessary con- cepts and methodologies to extract knowledge in a set of historical data. However, in Section 2 we have underlined the fact that nor- mal behavior are very dependent on the context. In consequence, to properly characterize such behavior, we take into account the data classes described previously. 3.2.1. Knowledge classes A class describes a context according to different criteria. To ex- tract knowledge about the influence of these criteria on a train behavior, each class is associated with (i) all behavioral data col- lected in the associated context (see Section 2), (ii) the patterns characterizing the normal behavior in this context. In Section 2.3.1, we have seen that the context of a class is pro- vided through the description of this class. We present below how are defined the corresponding set of data and the set of character- izing patterns. Definition 14. A sequence s is a c-general sequence if s is frequent in all child classes of c. If c is a specialized class, then the set of general sequences in c is the set of frequent sequences in c. A class c, defined in DC, is associated to the set of behavioral sequences contained in c as well as to the set of c-general se- quences. Consequently, a class c is now denoted by a triplet c ¼hD;E;Si, where: � D¼ descðcÞ, is the description of c in DC. � E ¼ dataðcÞ, is the set of fragments contained in c. � S ¼ seqðcÞ, is the set of c-general sequences. By using the previous definitions, the set of general sequences for each class in the class lattice is constructed in the following manner: 1. Let c be a specialized class. seq(c) is obtained by extracting all frequent sequences in data(c) (see Section 3.1). 2. Let c be a non-specialized class. seq(c) is defined as follows: Fig. 4. Sequences spreading in the contextual lattice. J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 7009 seqðcÞ¼ \ ci2�c seqðciÞ: Thus, by extracting frequent sequences in specialized classes only, we construct all sets of general sequences. Fig. 4 is depicting the construction of general sequences in a contextual lattice, by following the previously described method. This method of construction of all general sequences for each class has the advantage of limiting the application of the algorithm for discovering sequential patterns. Indeed, the extraction is per- formed only on specialized classes, and the remaining sets of gen- eral sequences are deducted from them. We have presented in this section the approach for extracting knowledge from historical behavioral data collected by sensors. In particular, we have seen how to exploit contextual data classes, and thereby obtained results are not associated with all historical data, but with contextual criteria. Therefore, we can precisely de- scribe the impact of these criteria on the normal behavior of a train and answer questions such as: � What are the behaviors that appear only when the outside tem- perature is high? � What behaviors are specific to a given itinerary? � What behaviors are specific to the ending fragments? � What are the most general behaviors, which do not depend on the context? This accuracy greatly improves the lack of knowledge about the normal train behavior. The obtained knowledge can be used be- sides to detect anomalies in the data collected, as we will see in the next section. 4. Anomaly detection In this section, we present how anomaly detection is performed. We consider that we are provided with both one database contain- ing normal behavior on which knowledge have been extracted (see Section 3) and data corresponding to one new journey. During a journey, each component (and each sensor) has a behavior which is related to the behavior of other components. These interactions, both temporal and inter-components, are de- scribed by a set of sequential patterns that we have extracted ear- lier (see Section 3). Therefore, we want to make an evaluation of the behavior of a component based on obtained knowledge. The main idea is relying on the following remarks: � we can consider that a sensor behaves normally if we find enough patterns in our knowledge base that validate its current state, � we can say that its behavior is anomalous if we find enough pat- terns that contradict it, � if there are not enough patterns to validate or contradict the current sensor state, then the behavior is considered uncertain because we do not have sufficient knowledge to assess its cur- rent behavior. Thus, for each reading and each sensor, we compute two scores: a concordance score and a discordance score. Depending on the va- lue of these scores, we can then indicate whether the behavior is normal, anomalous, or uncertain. Below, we describe how the var- ious needed scores are calculated. 4.1. Preliminary definitions In order to use the extracted sequential patterns for measuring the compliance of a sequence describing a new journey, first of all we introduce the concept of covering sequences. Definition 15. Let I be an itemset and s = hI1, . . . , Ii, . . . , Ini a sequence. I is covering s if "Ii 2 s, I # Ii. Example 12. The itemset (Alow) is covering the sequence h(Alow Blow)(Alow)i. Definition 16. Let p ¼hI�1 . . . I � i . . . I � l i be an aggregated pattern and s = hI1, . . ., Ij,. . . , Imi a sequence. p is covering s, denoted by p � s, if there exists a set of sequences {s1, s2, . . . , sm} such that: (i) s = s1 + s2 + � � � + sm, (ii) 8ij1 6 i 6 m; I�i is covering Ii. Moreover, I � i is called the corre- sponding itemset of Ii in p. Example 13. Let s be a sequence, and p an aggregated pattern, such that: s ¼hðAlowÞðAlowBlowÞ zfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflffl{s1 ðAavg Bav gÞðAavg Bav gÞ zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{s2 ðBhighÞ zfflfflffl}|fflfflffl{s3 i p ¼hðAlowÞ|fflffl{zfflffl} I1 ðAavg BavgÞ|fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl} I2 ðBhighÞ|fflfflffl{zfflfflffl} I3 i: We can note that s can be broken down into 3 sequences s1, s2 and s3, such that I1 v s1, I2 v s2, and I3 v s3. Thus, p � s. On the other hand, p is not covering the sequence s0 ¼ hðAlowÞðAlowBlowÞðAav gÞðAavg Bav gÞðAavg BavgÞðBhighÞi: Using the notion of covering sequence, we can now describe two types of patterns: (1) concordant patterns, validating the behavior of a sensor at a given time, and (2) discordant patterns contradicting this behavior. Definition 17. Let A 2 X, s = hI1, I2, . . . , Ini a sequence, and p ¼ hI�1I � 2 . . . I � mi an aggregated pattern. p is a concordant pattern for A in the ith itemset of s, i.e., a (A, i)-concordant pattern in s, if: (i) there exist integers h, j such that 1 6 h 6 i 6 j 6 n, and p �hIh, . . ., Ii , . . .Iji. (ii) let I⁄ be the corresponding itemset of Ii in p, there exists an item Av 2 I⁄. Table 5 Concordant patterns. ID Aggregated pattern Support p1 h(Alow)(AavgBavg)i 25% p2 h(AlowBavg)i 70% p3 h(AlowBavg)(Aavg)i 30% p4 h(AlowBlow)(AlowBavg)i 55% Table 6 Discordant patterns. ID Aggregated pattern Support p5 h(AlowBlow)(AavgBavg)i 45% p6 h(AhighBavg)i 20% 7010 J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 Example 14. Let A, B and C be sensors, p1 = h(AavgBlow)(Bavg)i, p2 = h(Aavg)i and p3 = h(Alow)(Aavg)i three aggregated patterns, and s a sequence such that: s ¼hðAavg BlowClowÞðAhighBavg CavgÞðAhighBhighChighÞi: The aggregated patterns p1 and p2 are (A, 1)-concordant patterns. On the other hand, p3 is not a (A, 1)-concordant pattern. A discordant pattern for the state of a sensor at a given time de- scribes an unexpected behavior. Definition 18. Let A 2 X, s = hI1, . . ., Ii , . . .Ini a sequence such that Ii contains an A-item, denoted by is, and p ¼hI�1I � 2 . . . I � j . . . I � mi an aggregated pattern such that I�j contains an A-item, denoted by ip. p0 is the sequence p where ip has been replaced by is in Ij. p is a discordant pattern for A in the ith itemset of s, i.e., a (A, i)- discordant pattern in s if: (i) p is not a (A, i)-concordant pattern, (ii) p0 is a (A, i)-concordant pattern. The items is and ip are called discordant items. More precisely, is is the discordant item of s, and ip is the discordant item of p. Example 15. Let A, B and C be sensors, p1 = h(AlowBlow)(Bavg)i and p2 = h(AhighBavg)i two aggregated patterns, and s a sequence such that: s ¼hðAavg BlowClowÞðAhighBavg CavgÞðAhighBhighChighÞi: The aggregated pattern p1 is a (A, 1)-discordant pattern. On the other hand, p2 is not a (A, 1)-discordant pattern. 4.2. Conformity score In the following examples, we will consider the sequence s, such that: s ¼hðAlowBlowÞðAlowBavgÞðAlowBav gÞðAavg Bav gÞi: Table 5 (respectively Table 6), contains the set of (A, 3)-concor- dant patterns (respectively (A, 3)-discordant patterns) in s as well as their support. 4.2.1. Concordance score Computing a concordance score for a given sensor A in the ith itemset of a given sequence s consists in distinguishing, in a set of patterns such as that described in Table 5, the (A, i)-concordant pattern in s. Note that two concordant patterns do not always have the same weight. In Table 5, we can not give the same weight to all the patterns in the overall concordance score of sensor A in the 3th itemset of s. Indeed, the size of the pattern p2 is too low to consider that it will affect the final score. In addition, we believe that the support of a pattern also influences its weight. Therefore, we define the weight of a sequence as follows: Definition 19. Let p be an aggregated pattern and s a sequence such that p is an (A, i)-concordant pattern in s. The weight of such a concordant pattern is defined as follows: weightðpÞ¼ kpk supportðpÞ: Example 16. The weight of p1 is: weightðp1Þ¼ kp1k supportðp1Þ¼ 3 0:25 ¼ 0:75: We can now define the concordance score of a sensor A in the ith itemset of a sequence s. Definition 20. Let Pc the set of all (A, i)-concordant patterns. The concordance score of a sensor A in the ith itemset of a sequence s is defined as follows: scoreconcðA; iÞ¼ X p2Pc weightðpÞ: Example 17. The concordance score of A in the 3rd itemset of s is: scoreconcðA; iÞ¼ X p2Pc weightðpÞ¼ weightðp1Þþ . . . þ weightðp4Þ ¼ 0:75 þ 1:4 þ 0:9 þ 2:2 ¼ 5; 25: 4.2.2. Discordance score In this part, we focus on discordant patterns, which tend to inval- idate the behavior of a sensor A in the ith itemset of a sequence. Discordance degree To evaluate the weight of a discordant pat- tern, it is important to consider the gap between this value and the ‘‘expected value’’. Let A 2 X a sensor, and D¼ domðAÞ. D is such that D¼ðv 1; . . . v i; . . . ; v nÞ, where vi is a discrete value. Definition 21. Let p be a (A, i)-discordant pattern in a sequence s, ip the discordant item of p, and is the discordant item of s. We define the discordance degree of p as follows. Let us consider v k 2D and v l 2D, such that vk = value(ip) and vl = value(is). The discordance degree of p, denoted by discDegree(p), is: discDegreeðpÞ¼ jl � kj n : Example 18. In the previous sequence s, dom(A) = dom(B) = (i1, i2, i3), such that i1 = low, i2 = avg, and i3 = high. By considering the sequence s and the discordant pattern p5, we can note that the discordant item of s is Alow and the discordant item of p5 is Aavg. Thus, the discordance degree of p5 is: discDegreeðp5Þ¼ j2 � 1j jdomðAÞj ¼ 1=3: We can now define the weight of a discordant pattern. This weight must take into account three features: the discordance de- gree, the size of a pattern, and its support. Concerning the size of a discordant pattern, we believe that the important part of a discor- dant pattern is the part covering the tested sequence. Therefore, we consider the size of the sequence s without the discordant item. Definition 22. Let p be an aggregated pattern and s a sequence such that p is an (A, i)-discordant pattern. The weight of such a discordant pattern is defined as follows: weightðpÞ¼ ðkpk� 1Þ supportðpÞ discDegreeðpÞ: J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 7011 Example 19. The weight of p5 is: weightðp5Þ¼ ðkp5k� 1Þ supportðp5Þ discDegreeðp5Þ ¼ ð4 � 1Þ 0:45 1 3 ¼ 0:45: Definition 23. Let Pd the set of all (A, i)-discordant patterns. The dis- cordance score of a sensor A in the ith itemset of a sequence s is: scorediscðA; iÞ¼ X p2Pd weightðpÞ: Fig. 5. Influence of the smoothing window on the anomaly score. Example 20. The discordance score of A in the 3rd itemset of s is calculated as follows: scorediscðA; 3Þ¼ X p2Pd weightðpÞ¼ weightðp5Þþ weightðp6Þ ¼ 0:45 þ 0:13 ¼ 0:58: 4.2.3. Conformity score For each sensor and each itemset in a sequence, we defined how to calculate a concordance score and a discordance score. We can now address the problem of defining a global conformity score of a sensor A in the ith itemset of a sequence, denoted as score (A, i). This score, defined between �1 and 1, must meet the following requirements: � if score(c, t) is close to 1, the state of A in i is considered as normal, � if score(c, t) is close to �1, the state of A in i is considered as abnormal, � if score(c, t) is close to 0, the state of A in i is considered as uncertain. Definition 24. Let A 2 X be a sensor and s a sequence. The confor- mity score of A in the ith itemset of s, denoted by score(A, i) is defined as follows: scoreðA; iÞ¼ scoreconcðA; iÞ� scorediscðA; iÞ maxðscoreconcðA; iÞ; scorediscðA; iÞÞ : Example 21. By considering the previous examples, we can now calculate the conformity score of A in the 3rd itemset of s as follows: scoreðA; 3Þ¼ scoreconcðA; 3Þ� scorediscðA; 3Þ maxðscoreconcðA; 3Þ; scorediscðA; 3ÞÞ ¼ 5; 25 � 0:58 5; 25 ¼ 0:89: 4.2.4. Smoothing window One important issue when addressing the problem of detecting anomalies concerns false alarms, especially when data are noisy. Thus, it is possible that a sensor has a bad score on a reading, which do not correspond to a real problem. In this case, the score of the sensor then quickly goes back to normal values. To avoid this, it is preferable to take into account the surrounding score values of the same sensor. We have thus developed the possibity to set a parameter called smoothing window and noted w. The score of a sensor for a given date is the average score on the smoothing window. Fig. 5 shows the evolution of the score of a sensor during a frag- ment, with and without smoothing (with w = 3). Without smooth- ing, we can see a decrease of the value that could be interpreted as an anomalous behavior. However, the immediate increasing of the score indicates that it is not a real anomaly. The smoothed score is an efficient method for restricting this phenomenon. We have described throughout this section how to use a list of sequential patterns, illustrating the normal behavior of the moni- tored system, to detect abnormal behavior among new collected data. This approach is particularly suited to the problem of moni- toring complex industrial systems because it does not only declare a data sequence as normal or abnormal, but provides also very pre- cise information about the anomaly detected: the localization of the fault (which sensor, which component, etc.), the precise mo- ment when this anomaly occurred, and a score quantifying the seriousness of the detected problem. Moreover, the described approach can also obtain information about the nature of the detected anomaly. We can in particular dis- tinguish the case where the detected anomaly corresponds to a sensor failure (which means that the behavior of the component on which the sensor is located is normal, but the measurements are corrupted) from the more serious case where the behavior of the component is actually involved in the anomaly. To differentiate those cases, we exploit the fact that several sensors are generally installed on a same component (e.g., 4 temperature sensors are in- stalled on each of the wheels of a train). We can therefore infer that if only one sensor from a group of sensors has a low score, then the corresponding anomaly relates to that sensor only. We can there- fore analyze the data according to several levels of granularity: the sensor level (by considering only the score of each sensor), the level of a component (e.g. by interpreting the average score of sensors installed on it), the level a car (by calculating the average score of components in the car), etc. 5. Experimental results In order to evaluate our proposal, several experiments were conducted on a real dataset. They correspond to the railway data collected on 12 trains where each train has 249 sensors, over a one-year period. A reading is collected every five minutes. 232 temperature sensors and 16 acceleration sensors are distributed on the different components (e.g., wheels, motors, etc.) and a sen- sor measures the overall speed of the train. The experimental protocol follows the organization of the proposals: 1. Characterization of normal behavior (see Section 3). We have studied the impact of contextualization on the characterization of normal behavior, and the benefit of the search for specific patterns in various contexts. Table 7 Number of general sequences and aggregated patterns, according to the contextual class. Class General sequences Aggregated patterns [T = ⁄; FT = ⁄; I = ⁄; E = ⁄] 5 5 [T = low; FT = ⁄; I = ⁄; E = ⁄] 20 17 [T = low; FT = middle; I = ⁄; E = ⁄] 123 89 [T = low; FT = middle; I = 1; E = ⁄] 270 189 [T = low; FT = middle; I = 1; E = 2] 542 455 Fig. 6. Simulated anomaly: blocked values. Fig. 7. Simulated anomaly: shifted values. 7012 J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 2. Anomaly detection (see Section 4). We have evaluated the con- formity score by applying it on normal and abnormal data. 5.1. Normal behavior characterization The discovery of frequent sequences has been performed with the PSP algorithm described in Masseglia, Cathala, and Poncelet (1998) with the minimum support set to 0.3. To build a contextual lattice such as that presented in Section 3.2, we used four criteria: � the fragment type (denoted as FT): beginning, middle, or ending fragments. � the exterior temperature (denoted as T): low or high. � The itinerary (denoted as I): I1, I2, I3. � The episode in an itinerary (denoted as E): 1st, 2nd, 3rd, or 4th episode. Table 7 shows, for some classes extracted from the obtained hierarchy, the respective number of general sequences, as well as the number of aggregated patterns among them. For example, the class of middle fragments which have been achieved with a low exterior temperature for any itinerary and any episode, de- noted as [T = low; FT = middle; I = ⁄; E = ⁄], contains 123 general se- quences. 89 (72%) among them are aggregated patterns. We can note that the more a class is general, the less it contains general sequences. This shows the necessity of contextualizing the behavior characterization. Indeed, very few behaviors are common to all contexts: a normal behavior is actually dependent on the context in which it operates. Moreover, aggregated patterns have reduced the number of extracted sequential patterns, by removing all very long and redundant sequences. 5.2. Anomaly detection We described in Section 4 the approach used to detect anoma- lies in new data. Thus, we can use it to classify new fragments of journeys into two categories: normal data and abnormal data. To evaluate our approach in an appropriate manner, it is necessary to conduct experiments on both types of data. However, if it is easy to find data which do not contain faults, it is often more difficult to obtain a large data set containing anomalies. For this reason, we have simulated a set of anomalous data, on which we have con- ducted our experiments. To this end, we have used the real normal data set, and we have corrupted the data in order to reproduce classic behavioral anom- alies on the values collected by sensors. Anomalies generated are threefold: � Blocked values (see Fig. 6): the value collected by a sensor is not refreshing. This anomaly is usually related to a faulty sen- sor, or a problem of transmitting information between the sen- sor and the database, but very rarely describes a real problem on a component. � Shifted values (see Fig. 7): a value that is shifted in comparison with the normal behavior of a sensor (or group of sensors): a constant value is added (or substracted) to the real collected value. This type of anomaly may, for example, describe an abnormal overheating of a component. � Random values (see Fig. 8): in this case, collected values are randomized. They describe an aberrant behavior without any link with the expected behavior. We simulated these anomalies in order to build a data set con- taining about 600 fragments of episodes (i.e., 200 episodes) com- posed as follows: � 300 are fragments of real data validated as normal (i.e., without anomalies), � 300 fragments containing anomalies were generated from nor- mal data. The three types of anomalies described above are dis- tributed equally between these fragments. The approach was tested on the basis of cross-validation, by segmenting the total set of data into 10 equal parts. Thus, each part contains 600 fragments, of which 300 are normal data. Note that the learning step (i.e., characterization of normal behavior) is per- formed on normal data only. To quantify the number of anomalous fragments in our data set, we consider a fragment is abnormal if its anomaly score falls below �0.5. Thus, we have obtained the confusion matrices presented in Table 8, containing the average results obtained by cross validation Fig. 8. Simulated anomaly: random values. Table 8 Global confusion matrix. Predict. normal Predict. anomalous Recall Precision Real normal 272 28 90.67% 91.58% Real anomalous 25 275 91.67% 90.76% Global results 297 303 91.17% 91.17% Table 9 Confusion matrix for random values anomalies. Predict. anomalous Predict. normal Real anomalous 98 2 Table 10 Confusion matrix for blocked values anomalies. Predict. anomalous Predict. normal Real anomalous 85 15 Table 11 Confusion matrix for shifted values anomalies. Predict. anomalous Predict. normal Real anomalous 92 8 J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 7013 on the entire test set. We can note that the system is well balanced and has similar qualities to both recognize the normal behaviors, but also to identify anomalies. In order to evaluate the results, we have calculated two mea- sures of quality widely used: the precision and the recall. More- over, recall and precision is in all cases above 90%, so the approach limit both the number of false alarms and the number of undetected anomalies. Tables 9–11 contain the results for each type of anomaly. We note that the results are very good to recognize a ‘‘random values’’ or a ‘‘shifted values’’ anomaly. This is due to the fact that in this case the simulated behavior is very different from a normal behavior. In contrast, ‘‘blocked values’’ anomalies are different: the anomalous value may not be sufficiently distant from ‘‘expected’’ values for detecting the problem. 5.2.1. Additional information about detected anomalies We have seen how to detect anomalies in new collected data. Moreover, it is possible to obtain additional information about de- tected anomalies. In particular, we use the lattice of classes to bet- ter understand the anomalies. When an anomaly is detected in a specialized class it may be useful, for users taking maintenance decisions, to better under- stand anomalies sources in order to assess its seriousness. In par- ticular, the user can recalculate the conformity score in other classes of the contextual lattice, and particularly in the super- classes of its specialized class. Several cases are then possible. � The score is low in the more specific classes of the lattice, but is improved in the more general classes. In this case, the tested journey is anomalous, but the anomaly is not serious. It does not meet the most specific behaviors expected in its class, but nonetheless meets the most general behaviors. � The score is low in all classes of the lattice, including the general classes. Here, the observed journey is problematic. The behavior of the train does not correspond to what is expected. � The score is low in a branch of the lattice, but still good in another branch. For example, consider a journey that contains an anomaly in its specialized class [high, low] (i.e., high exterior temperature and low humidity rate). So it is particularly helpful for decision makers to know the scores obtained by this journey in the superclasses [high,⁄] or [⁄low]. If the anomaly is related to one of these classes, i.e., if the score is low in [high,⁄] and correct in [⁄, low] for example, then the user gets additional information, which will allow him to investigate more effec- tively the cause of the anomaly. This information, combined with the fact that user may at any time consult the list of concordant or discordant patterns when an anomaly is detected, can provide a better understanding of the trains behavior. Therefore, the approach not only detects prob- lems but also allows decision makers to investigate the causes and potential seriousness of an anomaly, and help them to make the right decisions. 6. Related work The needs to develop our approach are twofold: (i) extract knowledge from sensor data, and (ii) detect anomalies in such data. Recently, the problem of mining sensor data has been addressed by the data mining community. Different approaches focusing either on the data representation by performing for example sensor clus- tering Ci, Guizani, and Sharif (2007), Rodrigues and Gama (2006), or knowledge extraction by mining association rules Boukerche and Samarah (2007), Boukerche and Samarah (2008), Chong et al. (2008), Ma et al. (2004), Mihail Halatchev (2005), Yairi, Kato, and Hori (2001), or sequential patterns Cook et al. (2003), Guralnik and Haigh (2002), Tseng and Lu (2009), Wu, Peng, and Chen (2001) were proposed. The second problem we address in this paper concerns the detection of anomalous behavior from data collected by the sen- sors. This topic, usually known as the anomaly detection problem, has received considerable attention in recent years. It can be viewed as a classification problem in which a system behavior can be clas- sified as normal or anomalous. Defined in Grubbs (1969), an anom- aly or outlier is an observation that appears to deviate markedly from other members of the sample in which it occurs. More recently, different definitions have been proposed in literature (Barnett & Lewis, 1994; Douglas M., 1980; Ramaswamy, Rastogi, & Shim, 2000). More generally, we will consider an anomaly as 7014 J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 behaving differently from the norm. A lot of approaches have been developed in order to meet the requirements of very different appli- cation domains, such as network intrusion detection (Jayakumar et al., 2008; Siaterlis & Maglaris, 2004; Wang, Guan, & Zhang, 2008; Wu and, 2006), industrial systems monitoring (Keogh, Lin, Lee, & Van Herle, 2006; Yairi et al., 2001), medical condition moni- toring (Lin, Keogh, Fu, & Van Herle, 2005; Wong, Moore, Cooper, & Wagner, 2003), fraud detection (Brause, Langsdorf, & Hepp, 1999, Phua, Alahakoon, & Lee, 2004), etc. Of course, the techniques employed are strongly related to the nature of input data. In particular, numerous anomaly detection methods have been developed considering transaction data sets. However, a lot of real-world domains have to handle temporal data that can be stored as sequential data sets. In particular, the behav- ioral data collected by sensors in the complex systems monitoring problem belongs to this category. The temporal nature of sequen- tial data requires new techniques to detect anomalies within. Sev- eral methods for detecting anomalies in sequential data have been developed for different application domains (Lee, Stolfo, & Chan, 1997; Michael & Ghosh, 2000; Sun, Chawla, & Arunasalam, 2006), but they cannot meet all our requirements. We have previously seen that an anomaly behave differently from the norm. Conse- quently, in order to properly detect anomalies in a data set, we need to define what is the norm, i.e., to characterize normal behav- ior. Some application domains can benefit from a priori knowledge of experts to construct this characterization. However, in other do- mains such as train monitoring for example, a priori knowledge is difficult to obtain. In consequence, we have to develop a system for exploiting the extracted results to characterize normal behavior. These patterns are describing the expected norm. We also have to manage several difficulties inherent to the problem that we address. For example, developing a hierarchy of contexts is necessary to (i) take into account variations in the train behavior according to various criteria, and (ii) provide more expla- nations to maintenance experts when an anomaly is detected. In addition, we handle sequences particularly large and redundant (see Section 2). Therefore, we propose the concept of aggregated sequences to accomplish our goals. Moreover, we must attach importance to the interpretability of results in the normal behavior characterization process as well as in the anomaly detection. Indeed, the problem is very sensitive (er- rors can lead to heavy financial costs), the experts need to under- stand the results to make and justify the right decisions. For all these reasons, we must develop a new appropriate method. 7. Conclusion In this paper, we have addressed a problem involved in many areas: the maintenance of complex systems. There are several solu- tions to perform maintenance of such systems. Firstly, corrective maintenance, consisting in making the necessary maintenance operations after the occurrence of a failure, is not suited to this contexte as too costly and too dangerous. A planned and system- atic maintenance is too expensive, although it can usually avoid serious failures. Thus, we addressed the problem of developing an approach to allow preventive maintenance, which is a good compromise between the two previous solutions. Preventive maintenance consists in detecting abnormal behavior that may be harbingers of major failures, to perform only the necessary pre- ventive maintenance operations. However, the complexity of such systems (e.g., trains, indus- trial machinery, etc.) makes their behavior difficult to understand and interpret. In these circumstances it is particularly difficult to detect abnormal behavior that often herald significant and costly failures. The problem that we address is particularly challenging. Indeed, errors in diagnosis may cause many inconveniences. We have therefore developed an approach to use data collected by sensors in order to analyze behaviors and allow the detection of anomalies. Our contribution is divided into three parts. First, we have pro- posed an adequate representation of data in order to extract knowledge based on sensor data describing the past normal behav- ior of systems. Secondly, we studied the possibility to extract sequential patterns in historical data both to improve the under- standing of systems for experts, but also to provide a knowledge database used to develop a method for detecting anomalies. To characterize normal behavior adequately, we also took into ac- count the context. Indeed, the context in which a system is running often has an influence on its behavior, and on the definition of an anomaly. Finally, we proposed an approach to compare new pat- terns with all sequential patterns describing normal behavior. We provide experts with an anomaly score for each sensor and each component of the systems studied. The approach allows to lo- cate precisely the anomalies and to quantify the extent to which the anomaly seems problematic. One main advantage of the presented approach is that all ob- tained results are easily interpretable for decision makers. This is particularly important because users must be able to make deci- sions with certainty. Although the presented work meets our expectations in terms of results, it opens interesting perspectives. In particular, the devel- opment of a graphical user interface will allow users to access re- sults quickly and efficiently. Another important aspect may be to adapt the approach to a real-time context. This will detect the anomalies on the running trains. Furthermore, an interesting ques- tion may be addressed: how to manage the evolution of normal behavior over time? This will conduct the knowledge database to be incrementally updated in order to ensure handled knowledge to be valid over time. References Agrawal, R., Imieliński, T., & Swami, A. (1993). Mining association rules between sets of items in large databases. SIGMOD Record, 22(2). Agrawal, R., & Srikant, R. (1995). Mining sequential patterns. In P. S. Yu & A. S. P. Chen (Eds.), Eleventh International Conference on Data Engineering. IEEE Computer Society Press. Barnett, V., & Lewis, T. (1994). Outliers in statistical data. Wiley series in probability & statistics. Wiley. April. Boukerche, A., & Samarah, S. (2007). A new in-network data reduction mechanism to gather data for mining wireless sensor networks. In MSWiM ’07: Proceedings of the 10th ACM symposium on modeling, analysis, and simulation of wireless and mobile systems. New York, NY, USA: ACM. Boukerche, A., & Samarah, S. (2008). A novel algorithm for mining association rules in wireless ad hoc sensor networks. IEEE Transactions on Parallel Distributed Systems, 19(7), 865–877. Brause, R., Langsdorf, T., & Hepp, M. (1999). Neural data mining for credit card fraud detection. In ICTAI ’99: Proceedings of the 11th IEEE international conference on tools with artificial intelligence (pp. 103). Washington, DC, USA: IEEE Computer Society. Carrascal, A., Díez, A., & Azpeitia, A. (2009). Unsupervised methods for anomalies detection through intelligent monitoring systems. In Proceedings of the fourth international conference on hybrid artificial intelligence systems (pp. 144). Springer. Chong, S. K., Krishnaswamy, S., Loke, S. W., & Gaben, M. M. (2008). Using association rules for energy conservation in wireless sensor networks. In SAC ’08: Proceedings of the 2008 ACM symposium on Applied computing. ACM. Ci, S., Guizani, M., & Sharif, H. (2007). Adaptive clustering in wireless sensor networks by mining sensor energy data. Computer Communications, 30(14–15), 2968–2975. Cook, D. J., Youngblood, M., Heierman, E. O., III, Gopalratnam, K., Rao, S., Litvin, A., et al. (2003). Mavhome: An agent-based smart home. In PERCOM ’03: Proceedings of the first IEEE international conference on pervasive computing and communications. Springer. Grubbs, F. E. (1969). Procedures for detecting outlying observations in samples. Technometrics, 11, 1–21. Gruenwald, L., & Halatchev, M. (2005). Estimating missing values in related sensor data streams. In J. R. Haritsa & T. M. Vijayaraman (Eds.), Proceedings of the 11th international conference on management of data (COMAD ’05). Computer Society of India. J. Rabatel et al. / Expert Systems with Applications 38 (2011) 7003–7015 7015 Guralnik, V., & Haigh, K. Z. (2002). Learning models of human behaviour with sequential patterns. In Proceedings of the AAAI-02 workshop ‘‘Automation as Caregiver’’. Hawkins Douglas, M. (1980). Identification of outliers. Chapman and Hall. Keogh, E., Lin, J., Lee, S.-H., & Van Herle, H. (2006). Finding the most unusual time series subsequence: algorithms and applications. Knowledge Information Systems, 11(1), 1–27. Lee, W., Stolfo, S. J., & Chan, P. K. (1997). Learning patterns from unix process execution traces for intrusion detection. In In AAAI workshop on AI approaches to fraud detection and risk management (pp. 50–56). AAAI Press. Lin, J., Keogh, E., Fu, A., & Van Herle, H. (2005). Approximations to magic: Finding unusual medical time series. In CBMS ’05: Proceedings of the 18th IEEE symposium on computer-based medical systems (pp. 329–334). Washington, DC, USA: IEEE Computer Society. Masseglia, F., Cathala, F., & Poncelet, P. (1998). The psp approach for mining sequential patterns. In J. M. Zytkow & M. Quafafou (Eds.), PKDD. Lecture Notes in Computer Science (Vol. 1510). Springer. Ma, X., Yang, D., Tang, S., Luo, Q., Li, S., & Zhang, D. (2004). Online mining in sensor networks. In H. Jin, G. R. Gao, Z. Xu, & H. Chen (Eds.), NPC. Lecture Notes in Computer Science (Vol. 3222). Springer. Michael, C. C., & Ghosh, A. (2000). Two state-based approaches to program-based anomaly detection. In ACSAC ’00: Proceedings of the 16th annual computer security applications conference (pp. 21). Washington, DC, USA: IEEE Computer Society. Phua, C., Alahakoon, D., & Lee, V. (2004). Minority report in fraud detection: Classification of skewed data. SIGKDD Explorations Newsletter, 6(1), 50–59. Ramaswamy, S., Rastogi, R., & Shim, K. Efficient algorithms for mining outliers from large data sets. In SIGMOD ’00: Proceedings of the 2000 ACM SIGMOD international conference on Management of data (pp. 427–438). New York, USA: ACM. Rodrigues, P. P., & Gama, J. (2006). Online prediction of streaming sensor data. In: Jesus, S., Gama, A. -R. J., & Roure, J., (Eds.), Proceedings of the 3rd international workshop on knowledge discovery from data streams (IWKDDS 2006), in conjuntion with the 23rd international conference on machine learning. Shaikh, R. A., Jameel, H., d’Auriol, B. J., Lee, S., Song, Y.-J., & Lee, H. (2008). Trusting anomaly and intrusion claims for cooperative distributed intrusion detection schemes of wireless sensor networks. In ICYCS ’08: Proceedings of the 2008 the 9th international conference for young computer scientists (pp. 2038–2043). Washington, DC, USA: IEEE Computer Society. Siaterlis, C., & Maglaris, B. (2004). Towards multisensor data fusion for dos detection. In SAC ’04: Proceedings of the 2004 ACM symposium on applied computing (pp. 439–446). New York, NY, USA: ACM. Sun, P., Chawla, S., & Arunasalam, B. (2006). Mining for outliers in sequential databases. In D. Lambert, J. Ghosh, J. Srivastava, & D. B. Skillicorn (Eds.), SDM. SIAM. Tseng, V. S., & Lu, E. H.-C. (2009). Energy-efficient real-time object tracking in multi- level sensor networks by mining and predicting movement patterns. Journal of Systems and Software, 82(4), 697–706. Wang, W., Guan, X., & Zhang, X. (2008). Processing of massive audit data streams for real-time anomaly intrusion detection. Computer Communications, 31(1), 58–72. Wong, W.-K., Moore, A., Cooper, G., & Wagner, M. (2003). Bayesian network anomaly pattern detection for disease outbreaks. In Proceedings of the twentieth international conference on machine learning (pp. 808–815). AAAI Press. Wu, P.-H., Peng, W.-C., & Chen, M.-S. (2001). Mining sequential alarm patterns in a telecommunication database. In DBTel ’01: Proceedings of the VLDB 2001 international workshop on databases in telecommunications II. Springer-Verlag. Wu, N., & Zhang, J. (2006). Factor-analysis based anomaly detection and clustering. Decision Support Systems, 42(1), 375–389. Yairi, T., Kato, Y., Hori, K. (2001). Fault detection by mining association rules from house-keeping data. In Proceedings of the sixth international symposium on artificial intelligence, robotics and automation in space. Anomaly detection in monitoring sensor data for preventive maintenance Introduction Data preprocessing Sensor data for train maintenance Sensors Measurements Readings Granularity in railway data Journey Episode Episode fragment Data classes Data classes Contextual lattice Normal behavior characterization How to extract normal behavior? Contiguous subsequences mining Aggregated sequences Contextualized characterization Knowledge classes Anomaly detection Preliminary definitions Conformity score Concordance score Discordance score Conformity score Smoothing window Experimental results Normal behavior characterization Anomaly detection Additional information about detected anomalies Related work Conclusion References 
work_5mo3nifidfctpcvtbeobiho6ki	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 221004241 Params is empty 221004241 exception Params is empty 2021/04/06-03:05:50 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221004241 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:50 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_5nuvugk7enet7cebmlc4yjulxq	----	PII: 0020-0255(85)90023-4 INFORMATION SCIENCES 36,5-16 (1985) 5 From Fuzzy Logic to Expert Systems* BRIAN R. GAINES and MILDRED L. G. SHAW Department of Computer Science, York University, 4700 Keele Street, Downsview, Ontario, Canada MU 1 P3 ABSTRACT The history and achievements of fuzzy set theory (FST) are considered in relation to developments in expert systems (ESs). An overview is given of early ES developments and the role of linguistic rules and metarules. It is suggested that FST and ESs are not just mathemati- cal and technological advances but also represent major paradigm shifts in system theory. The main shift is away from the normative application of technology to change the world to be theoretically tractable, and towards increasing model realism. The limitations of classical system theory when applied to natural systems were the impetus behind the development of FST. INTRODUCTION Lotfi Zadeh first discussed the need for a “mathematics of fuzzy or cloudy quantities” in a paper entitled “From circuit theory to system theory” published in 1962. This led to his publishing his seminal paper, “Fuzzy sets,” proposing such a mathematics in 1965. After two decades it is reasonable to look back and review the impact of this proposal. Has it led to a shift in our modes of thinking and problem-solving? Has it led to new perspectives on systems? Has it led to new tools? Is it being used in applications? Has the original purpose been achieved? If one judges a concept by the activity it generates and the literature it generates, then that of fuzzy set theory (FST) must be rated very highly. A comprehensive bibliography for the first decade shows an increase from 2 papers published in 1965 to over 227 in 1975, with a cumulative total in 1975 of *This paper is dedicated to Lotfi Zadeh, who has combined scientific work of the highest standard with tenacity, kindness, and patience. OElsevier Science Publishing Co., Inc. 1985 52 Vanderbilt Ave., New York, NY 10017 0020-0255/85/$03.30 6 BRIAN R. GAINES AND MILDRED L. G. SILAW some 620 items [lo] and in 1979 of some 1400 items [19]. The number of papers a year and cumulative total fit well to exponential growth at 60% a year for the first decade. However, it is now almost impossible to track the growth of a literature which as grown from the output of a small group of specialists to that of an international community involving almost every nation and discipline. The rapid growth in the output of papers on FST during the decade from 1965 is the characteristic initial exponential form of the learning curve for a new discipline [4]. The overall form of the learning curve is logistic, with the initial exponential growth eventually becoming an exponential approach to an asymp- totic output rate. Fitting empirical data to such a curve is notoriously sensitive to minor fluctuations [2]. The best fit to the 1965-1979 data gives an asymptote at 300 papers a year, with 10% of this being reached in 1970,50% in 1975, and 90% in 1980. This suggests that the first half of the learning curve was the decade 1965-1975, and that we are now completing the second half in the decade 1975-1985. The 300 papers a year figure corresponds to about four times the output in the main journal on FST and is probably a realistic figure for refereed journal papers. The growth rate of the dissemination of knowledge about, interest in, and work on FST and its applications has been spectacular. There has been a shift in the modes of thinking and problem-solving for a significant community of theoretical and applied scientists and technologists. In this paper we attempt to put this shift in perspective by examining the role of fuzzy reasoning in an applications area whose growth has itself been as spectacular, that of expert system (ESs) and their applications [23,15,25]. This is an interesting area, not only for its high intrinsic value, but also because it enables us to contrast differing aspects of the role of FST in modem informa- tion science. Expert systems development leads to requirements for reasoning with imprecise data, where FST provides an alternative paradigm to those of classical logic and probability theory. The most well-recognized breakthroughs in ESs, such as MYCIN [28], were not based on FST, but on heuristic methods that turn out to resemble FST closely. Other early breakthroughs such as linguistic process controllers [20] were based directly on FST. The next section gives an overview of early ES development commencing with control systems based on FST. EXPERT SYSTEMS: AN OVERVIEW The computer simulation of people in the roles of experts on some topic has become an important application of interactive computer systems. It has gener- ated a new industry based on creating expert systems to make the practical working knowledge of a human expert in a specific subject area such as medicine or geology widely available to those without direct access to the FROM FUZZY LOGIC TO EXPERT SYSTEMS 7 original expert [25]. Programs now exist that have made practical achievements in medical diagnosis, interpretation of mass spectroscopy results, analysis of geological survey data, and other problems where one would normally go to a human expert for advice. EARLY EXPERT SYSTEMS One of the first ES developments was the fuzzy logic control system devel- oped in 1974 by Mamdani and Assilian. The system accepted human knowledge of control strategies expressed verbally and encoded it directly as computer programs which acted on the environment [20]. This work was undertaken as part of a study of machine learning in process control, and the system controlled was a small steam engine. The verbal rules were of the form shown in Figure 1. The history of this discovery that human control expertise could be expressed linguistically and encoded directly on a computer is itself interesting. The original objective of the work was to investigate learning controllers in a realistic situation, and the steam engine used was real, not a computer simulation. It was found that learning a control strategy from a zero base was problematic, since it did not happen, at least not before the steam engine ran out of water. Games and Andreae f9] had reported a similar problem with simulated control situa- tions and introduced the notion of priming a learning machine with an initial suboptimal control strategy. Gaines [7] reported positive results with an adap- tive threshold logic learning controller using a crude priming technique to simulate the effect of verbal instructions on a human trainee. Mamdani and Assilian decided to prime their system in a similar way, but using Zadeb’s [36] suggestion of directly translating linguistic statements into decision rules using FST. What was surprising at the time and made the 1974 results a recognized breakthrough was that the control rules derived from the verbal statements were extremely effective. They compared favorably with those derived by tuning a standard PID ~ropo~on~-steak-de~vative) controller for optimum perfor- mance. Mamdani and Ass&n also found that the learning machine then proceeded only to learn less effective strategies. Hence interest switched to the process whereby human expression of verbal rules that appear vague can lead to highly effective control strategies. In the past ten years Mamdani and Assilian’s If the pressure error is positive and big and the change in pressure error is not negative medium or big Then make the heat change negative and big, Fig. 1. Rule from fuzzy logic controller BRIAN R. GAINES AND MILDRED L. G. SHAW When the burning zone temperature is drastically low: (a) reduce kiln speed; (b) reduce fuel. When the burning zone temperature is slightly low: (a) increase I.D. fan speed; (b) increase fuel rate. Fig. 2. Rules from a lime kiln operator’s manual. results have been replicated in many different countries for many different control processes, including a number of significant industrial processes such as pig iron smelting where effective automatic control had been thought to be impossible [21]. The concept of an ES was not prevalent at the time of the initial fuzzy control studies, and their significance as examples of early ES development was noted only later. It has also been noted that the transmission of decision and control knowledge through linguistic rules is a technique that has been used in the past quite independently of computer systems. Figure 2 is an excerpt from a training manual for lime kiln operators [24]. It seems that the encoding of expertise into a set of fairly vague but highly applicable rules is a standard way of making it available to others. In parallel with the controller development, other rule-based ESs were being developed for completely different domains. The system widely recognized as an early breakthrough, MYCIN, is a medical diagnosis ES which aids a clinician to act as a consultant on infectious diseases [28]. It uses rules of the form shown in Figure 3. These rules are obtained from specialists in microbial infections, and their application to particular data is fairly simple data processing. The rules are validated through their application to many cases and revised when they fail to give the correct diagnosis. MYCIN is designed to interact with a clinician in order to make a diagnosis and suggest therapy for a particular patient with suspected microbial infections. It first gathers data about the patient and then uses them to make inferences about the infections and their treatment. Note that the MYCIN rule of Figure 3 involves an assertion that is evidential rather than true. Shortliffe found it necessary to encode rules of inference that RULE 50 If (1) the infection is primary-bacteremia, and (2) the site of the culture is one of the sterile sites, and (3) the suspected portal of entry of the organism is the gastrointestinal tract, Then there is suggestive evidence (0.7) that the identity of the organism is bacteroides. Fig. 3. A MYCIN rule. FROM FUZZY LOGIC TO EXPERT SYSTEMS 9 were imprecise and could not be encoded simply in terms of truth and falsity. He ascribed a verbal label, “suggestive evidence,” and a numerical truth value, “0.7,” to a rule and developed a calcuhts for combining such truth values in chains of logical inference. Thus, linguistic reasoning and multivalued logics were key components of early ES developments, although the MYCIN developers were initially unaware of FST and the linguistic controller developers were initially unaware of ES concepts. Since the early success with MYCIN a very wide range of ESs have been developed. Gevarter [15] has summarized some well-known expert systems and their applications, but the numbers and domains have since increased so rapidly that it is now impossible to make any accurate count. Most universities have some activity in this field, and many industrial ESs are regarded as highly proprietary. The problem of extracting knowledge from experts and encoding it for computers is now termed knowledge engineering and is regarded as the major bottleneck in ES development ‘[17]. Direct elicitation of a complete, noncontradictory set of linguistic rules from experts for a large domain has proved very difficult and time-consuming. However, Boose [3] has reported success in rapid prototyping of ESs for a large number of applications using knowledge engineering tools based on an FST model of personal construct psychology [27, 111. EXTENSION OF RULES TO METAKNO WLEDGE Mamdani, Procyk, and Baaklini [22] found that learning could be introduced effectively in the steam engine controller through meturules that expressed the way in which the basic rules should be changed as a result of performance feedback. The learning level of their controller operated on rules of the form shown in Figure 4. This was sufficient for the fuzzy controller to acquire a control strategy similar to that induced through verbal rules from a human expert. Metarules were also introduced independently by Davis to aid the debugging of MYCIN. It was difficult to set up the MYCIN rules initially and also difficult to trace errors in the deductions. To overcome these problems TEIRESIAS [6] was added as an auxiliary ES with expertise about MYCIN to explain MYCIN’s decisions and help the clinician amend the rules when they led to incorrect If time is small and error is negative big Then desired change is big. If time is big and error is positive zero qm desired change is zero. Fig. 4. Metarules from fuzw leaming controller. 10 BRIAN R. GAINES AND MILDRED L. G. SHAW METARULEOO3 : If (1) there are rules which do not mention the current goal in their premise (2) there are rules which mention the current goal in their premise Then it is definite that the former should be done before the latter. Fig. 5. A TEIRESIAS metande. conclusions. TEIRESIAS uses a rule-based approach to reasoning like MYcIN’s, but the rules are now rules about the forms of rules and the use of rules. A typical one of these metarules is shown in Figure 5. Whereas MYCIN’S rules are specific to microbial infections, those of TEIRESIAS are more general and can be used in other domains. Davis [5], for example, shows TIXRELSIAS being used as an investment decision system for clients of a stockbroker. One of the important features of MYCIN/TEIRESIAS that has become an essential characteristic of ESs is their capability to provide explanations of the deductions given. “Why?” questions are accepted as responses when data are requested, and are interpreted as a request for the rule to be shown that requires the data requested. A “why?’ question may also be asked when conclusions are drawn, and is then interpreted as a request for the complete chain of logic used in arriving at that conclusion to be shown. The facility to answer such questions make ESs accountable for their behavior and conclusions. This is itself a major new feature of systems programmed for computers. Another important feature of ESs is that they are not static representations of knowledge but can continue to acquire knowledge as they are used. Essen- tially, the use of me&rules allows ESs to be programmed interactively by their users. From one perspective the metarules of the fuzzy learning controller and TEIRESIAS can be seen as an important development in automatic programming. From another they can be seen as a way in which a machine acquires knowledge through interaction with its environment or with a person. These are analogous to the fundamental ways in which people acquire knowledge [12]. For such applications computational logics capable of dealing with the uncertainties of imprecise data and fallible hypotheses are essential. SHIFTS IN SYSTEMS PARADIGMS The previous section has shown how the early applications of FST to control and decision systems paralleled the development of early expert systems in the use of linguistic rules, fuzzy reasoning, and metarules. This role of FST, significant as it is in itself, is only an indication of the deeper paradigm shifts from which FST and ESs both stem. The classical approach in decision and FROM FUZZY LOGIC TO EXPERT SYSTEMS 11 control system design is: (1) thoroughly instrument the system to be controlled or about which decisions are to be made; (2) use the instrumentation to gather data about the system behavior under a wide variety of circumstances; (3) from these data build a model of the system that accounts for this behavior; (4) from this model derive algorithms for decision or control that are optimal in terms of prescribed performance parameters. This positivistic paradigm underlies the methodologies of the physical sciences and technologies based on them. It has the merit that it has been extremely successful in engineering much of the technological infrastructure of our current civilization. However, this paradigm is successful only to the extent that the systems under consideration are amenable to instrumentation and modeling. Its greatest successes have been where this amenability can be achieved normatively, that is, in cases where the system to be controlled is itself a human artifact. For example, linear system theory has not become a major tool in systems engineer- ing because most natural systems are linear-they are not. The implication is in the opposite direction: that linear systems are mathematically tractable and that we design artificial systems to be linear so that we may model them readily. The application of “a linear model with quadratic performance criterion” to natural systems is often attempted, but in general it does not work. We have done so not because the tool was appropriate, but because it was the only one we had. However, the use of a hammer to insert screws, although partially effective, tends to distort, destroy, and generally defeat the purpose of using a screw. Similarly, the use of an inappropriate system theory to model a system may give useful, but limited, results when we have no other, but it distorts reality, destroys information, and generally defeats the purpose of modeling that system. Much of our current technology succeeds to the extent that it is normative. In agriculture we reduce the complexity of a natural ecology to a comprehensible simplicity by the use of pesticides, herbicides, and chemical fertilizers [13]. We reduce the system to one which is amenable to our modeling techniques. That simpler is not necessarily better and that reengineering nature to impose uniformity destroys variety which is itself valuable have only been realized in recent years. The four shifts in perspective that we see in FST and ESs are: PROBLEM 1. The models available are inadequate to capture the system. Old approach: Procrustean design. Change the world to fit the model-normative technology. 12 BRIAN R. GAINES AND MILDRED L. G. SHAW Paradigm shift: Model realism. Use system methodologies and information technology that enable the natural world to be modeled without distortion and destruction. PROBLEM 2. Optimal control is oversensitive to system uncertainties. Old approach: Suboptimality. Use a suboptimal controller that is robust. Paradigm shift: Model uncertainty. Model the uncertainty as part of the system. PROBLEM 3. Data are unavailable or inadequate for modeling. Old approach: Managing. Do not automate-leave to human decision/con- trol. Paradigm shift: Expert systems. Model the person as a decision-maker or controller. PROBLEM 4. Neither a human nor an automatic system alone is adequate. Old approach: Ad hoc system design, Use a mixture of automatic and human decision/control. Paradigm shift: Accountable integration. Integrate automatic and human activity-make the automation accountable (“why?” in ESs). The last three perspectives all stem from the first. The importance of this first perspective to Zadeh is apparent in his 1962 paper, where he discusses the fundamental inadequacy of conventional mathematics for coping with the analysis of biological systems, noting also that the need for a new mathematics was becoming increasingly apparent even in the realm of inanimate systems. The second perspective is that which led to EST. Optimal control theory was regarded as the peak achievement of system theory in the 1950s and 1960s. However, it proved limited in application because it demanded precision in system modeling that was impossible in practice. It was too sensitive to the nuances of system structure expressed through overprecise system definition. The third perspective is that which led to the success of linguistic fuzzy controllers and later ESs. Hayes-Roth [16] has noted the many problems that have been felt to require human management are now amenable to ESs. Modeling the way the expert performs the task rather than modeling the task itself is the primary characteristic of an ES. The fourth perspective is an important one for both ESs and FST. They are knowledge-based systems because they make provision for explaining the deci- sions reached in terms of the data and inferences used. It is interesting to note that logics of uncertainty that aggregate evidence, such as probabilistic logics, do not provide a simple mechanism for explanation. Explicable logics have to be truth-functional and nonaggregative; fuzzy logic satisfies these requirements (uniquely among those logics satisfying the weak axioms of a standard tmcer- FROM FUZZY LOGIC TO EXPERT SYSTEMS 13 tainty logic [8]). It is also interesting to note that the capability to give explanations is seen by some philosophers of science as a key difference between models that make accurate predictions and scientific theories that in addition provide causal explanations [26]. The “why?” question has important implica- tions for both the practical and theoretical significance of ESs and the log&s on which they are based. ZADEH’S ROLE IN THE PARADIGM SHIFTS The development of FST and ESs is not just a technological improvement but represents a major paradigm shift in system theory. Zadeh was ideally placed to understand the strengths and weaknesses of system theory, and the significance of developments in artificial intelligence. The sixty-five papers he published in the fifteen years preceding “Fuzzy sets” [35] represent the peak achievements of mathematical system theory on a broad front. One of his earliest papers is on thinking machines as a new field in electrical engineering [29]. His technical papers commence with filtering and prediction techniques at the frontiers of linear system theory, for example on linear time-varying systems. Many of the papers contain the term “general” in their title, and many of them are applied to circuit and control systems. Zadeh’s aspirations were clearly defined from the start: to develop general mathematical theories that could be applied in engineering design. In the later papers of that era Zadeh grapples with the problem of formulat- ing engineering concepts whose intuitive significance is of paramount important but which are difficult to encompass theoretically. He discusses the ‘“identifica- tion problem” [30], “what is optimal” [31], “the definition of adaptivity” [33], and the “concept of state” [34]. These papers represent the most advanced system thinking of that era and were seminal in content. They were based on classical mathematical system theory and widely recognized as breaking new ground. It was a shock for many when Zadeh turned away from this work after 1965 and began to develop and promote first the concepts of FST and second its instantiation through fuzzy linguistic reasoning. FST was from the outset an attempt to create a new mathematical system theory that corresponds to paradigm shift 1 and fits the realities of the world without distorting them. It was created by a person who had extended the boundaries of current system theory, attempted to encompass in generality the key concepts of applied systems engineering, and recognized the failure of that theory in this task. FST was also highly controversial from the start and aroused strong emotional responses. Kalman [18] comments on an FST paper by Zadeh: His proposals could be severely, ferociously, even brutally criticized from a technical point of view. . No doubt Professor Zadeh’s enthusiasm for fuzzy sets has been reinforced by the prevailing political climate in the U.S.: one of unprecedented permissiveness. 14 BRIAN R. GAINES AND MILDRED L. G. SHAW Arbib [l], in a book review, is more gentle in his criticism but casts doubts on the appropriateness of FST (“A horrifying thought-what if Newton had rejected the concept of mass, and sought to base his theory on a ‘degree of heaviness’ between 0 and 1 for each object in the universe?“) and chides those writing on FST for their insularity in not linking it with existing material on multivalued logics, linguistic semantics, and so on. Criticism and controversy are the hallmarks of significant paradigm changes and key components of the process of scientific development and progress. So is the accompanying literature explosion, much of which is inadequately founded due to the inexperience of not only authors but also referees in the material of the new paradigm. Those days are now past. The key journals on FST are well refereed, the mathematical foundations are solid, and the many links with related topics and literature have been well documented. Against the back- ground of ESs, human reasoning processes are now treated with greater respect and linguistic reasoning is being widely investigated. The notion of “degree of heaviness” of an object does precede that of its mass, and the development of naive physics on the basis of such informal concepts is now regarded as a major challenge for artificial intelligence and cognitive science research [14]. CONCLUSIONS After two decades of fuzzy set theory, with the development of mathematical foundations and a wealth of applications, it is easy to look back and wonder what all the controversy was about. We have suggested that FST and ESs, and the application of one to the other, are not just mathematical and technological advances but also represent major paradigm shifts in system theory. The controversy has been about quite fundamental changes in system philosophy and technology, shifts from a positivistic, normative approach to a more realistic and naturalistic approach. These shifts are apparent throughout science and technology and its application to our world and society. Lotfi Zadeh recognized the changes necessary in system theory at an early stage and pioneered in developing new approaches. The technical success of his work and the wide-ranging theoretical and practical advances stemming from it are tributes in themselves. However, technical tributes alone are not enough to indicate the courage needed to drop well-established lines of research and swim against the current of scientific thought. His perseverance in the face of criticism, his patient explanations, his kindness in supporting others, and his willingness to present the material to unsympathetic audiences, make our tributes as much personal as technical. Fuzzy set, theory cannot be either right or wrong. It is applicable mathematics tested by its uses. However, the rationale behind it, the systemic principles FROM FUZZY LOGIC TO EXPERT SYSTEMS 15 involved, can be right or wrong, They are right for our time, for the objectives of dealing adequately with a complex universe and extending the capabilities of the person with computer enhancements. The redevelopment of system theory is not yet complete, and the seminal notions of stability, adaptivity, modeling, and so on still need adequate expression. However, we now have the Foundations on which to build a system theory that combines realism with power and provides applicable mathematics for our ~ow~~~~b~~ saiety. REFERENCES 1. M. Arbib, Book review of applications of fuzzy sets to systems analysis, J. Sot. Industrial and Appl. Math. 19:753 (1977). 2. W. Ascher, Forecasting: An Appmisal far Policy-Makers and Planners, iohns Hopkins W.P., Baltimore, 1978. 3. J. H. Boose, Personal construct theory anil the transfer of human expertise, in Proceedings AAAI-84, Amer. Assoc. for Act&&l Intelligence, C&f., 1984, pp. 27-33. 4. D. Crane, ~~o~~b~e Co&es: -niffusioon of knowledge in Scient@c Commrrrrities. Univ. of Chicqo Press, Chicqo, f972. 5. R. Davis, TEIRESIAS: Experiments in communicating with a ~owledg~~~ed system, in Designing for Ham-Computer Commwrication (M. E. sime and M. J. Coombs. Eds.), Academic, London, 1983, pp. 87-137. 6. R. Davis and D. B. Lenat, KnowledgeBused Systems in Artificial Intelligence, McGraw-Hill, New York, 1982. 7. B. R. Gaines, The learning of perceptual-motor skills by men and machines and its relationship to training, Znstructianal Sci. 1(3):263-312 (Oct. 1972). 8. B. R. Gaines, Precise pas-fuzzy Future, Internat. J. Man-Much& Stud, 19(1):117-134 (July 1983). 9. B. R. Gaines and J. H. Andreae, A lear+g machine in he context of the general control prablem, in Fr~ee~n~ cif t&-e 3rd Congress of tke ~~ter~~jo~~ Federation for Auromatic Con&o& Butterworth, London, 1966. 10. B. R. Gaines and L. 3. K&out, The fuzzy decade: A bibfiograpity of fuzzy systems and closely related topics, hernat. J. Man-Muchine Stud. 9(lf:l-68 (Jan. 1977). 11. B. R. Gaines and M. L, G. Shaw, Logical famdations of expert systems, in Proceedings of IEEE Internationaf Conference on Systems, Man and Cybernetics, Halifax, Nova Scotia, Oct. 1984, pp. 238-247. 12. B. R. Gaines and M. L. G. Shaw, Xhe Art of Computer Conversation: A New Medium for Communicution, Prentice-Hall, E~~glcwcud Cliffs, NJ., 1984. 13. B. R. Gaines and M. L. G. Shaw, Expert systems: The substitution of howledge for materials in world food production, Pmsible tiorkh 1(2):5-8 (July 1984). 14. D. Gentner and A. L. Stevens, (&Is.), Mental Mode&, Lawrence Erlbaum, Hillsdale, N.J., 1983. 15. W. B. Gevarter, Expert systems: Limited but powerful, IEEE Spectrum l&39-45 (1983). 16. F. Hayes-Rot& The industria&ation of knowkdge exxgkerkg, in Artificial ~~te~~~g~ce Applications for Bt&ness (W. Reitman, Ed.), Able%, Norwood, W.I., 19&Q, pp. 159-177. 16 BRIAN R. GAINES AND MILDRED L. G. SHAW 17. 18. 19. 20. 21. 22. 23. F. Hayes-Roth, D. A. Waterman, and D. B. Lenat (Eds.), Building Expert Systems, Addison-Wesley, Reading, Mass., 1983. P. J. Kalman, Discussion of “A new approach to system analysis” by L. A. Zadeh, in Man nnd Computer (M. Marois, Ed.), North-Holland, New York, 1974, p. 93. A. Kandel and R. R. Yager, A 1979 bibliography of fuzzy sets, their applications and related topics, in Advances in Fuzzy Set Theory and Applications (M. M. Gupta, R. K. Ragade, and R. R. Yager, Eds.), North-Holland, Amsterdam, 1979, pp. 621-744. E. H. Mamdani and S. Assilian, An experiment in linguistic synthesis with a fuzzy logic controller, in Fuzzy Reasoning and its Applications (E. H. Mamdani and B. R. Gaines, Eds.), Academic, London, 1981, pp. 31-l-323. E. H. Mamdani, J. J. Ostergaard, and E. Lembessis, Use of fuzzy logic for implementing rule-based control of industrial processes, in Fuzzy Sets and Decision Analysis (H. J. Zimmermann, L. A. Zadeh, and B. R. Gaines, Eds.), TIMS Series Studies in the Manage- ment Sciences, 20, North-Holland, Amsterdam, 1984. E. H. Mamdani, T. Procyk, and N. Baaklini, Application of fuzzy logic to controller design based on linguistic protocol, in Discrete Systems and Fuzzy Reasoning (E. H. Mamdani and B. R. Gaines, Eds.), Queen Mary College, London, 1976, pp. 125-149. D. Michie (Ed.), Expert Systems in the Micro Electronic Age, Edinburgh, U.P., Edinburgh, 1970. 24. K. E. Peray and J. J. Waddell, The Rotary Cement Kiln, Chemical Publishing, New York, 1972. 25. W. Reitman (Ed.), Artificial Intelligence Applications for Business, Ablex, Nonvood, N.J., 1984. 26. 27. 28. 29. 30. 31. 32. 33. 34. 35. 36. W. Salmon, Scientific Explanation and the Causal Structure of the World, Princeton U.P., Princeton, N.J., 1984. M. L. G. Shaw, Interactive knowledge elicitation, in Proceedings of CZPS Session 84, Canad. Inform. Processing Sot., Calgary, May 1984, pp. 202-208. E. H. Shortliffe, Computer-Based Medical Consultations: MYCIN, Elsevier, New York, 1976. L. A. Zadeh, Thinking machines-a new field in electrical engineering, Columbia Engrg. Quart. 3:12-13 (1950). L. A. Zadeh, On the identification problem, IRE Trans. Circuit Theory CT-3:277-281 (1956). L. A. Zadeh, What is optimal?, IRE Trans. Circuit Theory CT-4:3 (1957). L. A. Zadeh, From circuit theory to system theory, Proc. Inst. Radio Engrs. 50:856-865 (1962). L. A. Zadeh, On the definition of adaptivity, Proc. IEEE 51469-470 (1963). L. A. Zadeh, The concept of state in system theory, in Proceedings of the Second Systems Symposium, Wiley, New York, 1964. L. A. Zadeh, Fuzzy sets, Inform. and Control 8:338-353 (1965). L. A. Zadeh, Outline of a new approach to the analysis of complex systems and decision processes, IEEE Trans. Systems Man Cybernet. SMC-2:28-44 (1973). Received 18 March 1985; revised 31 March 1985 
work_5pmq7rfw4bdkncq5sdteafpdiu	----	Hierarchical camera auto-calibration for traffic surveillance systems Hierarchical camera auto-calibration for traffic surveillance systems S. Álvarez !, D.F. Llorca, M.A. Sotelo Computer Engineering Department, Polytechnic School, University of Alcalá, Madrid 28871, Spain a r t i c l e i n f o Keywords: Auto-calibration Pan-tilt-zoom cameras Vanishing points Intelligent transportation systems Urban traffic infrastructures a b s t r a c t In this paper, a hierarchical monocular camera auto-calibration method is presented for applications in the framework of intelligent transportation systems (ITS). It is based on vanishing point extraction from common static elements present on the scene, and moving objects as pedestrians and vehicles. This pro- cess is very useful to recover metrics from images or applying information of 3D models to estimate 2D pose of targets, making a posterior object detection and tracking more robust to noise and occlusions. Moreover, the algorithm is independent of the position of the camera, and it is able to work with variable pan-tilt-zoom (PTZ) cameras in fully self-adaptive mode. The objective is to obtain the camera parame- ters without any restriction in terms of constraints or the need of prior knowledge, to deal with most traf- fic scenarios and possible configurations. The results achieved up to date in real traffic conditions are presented and discussed. ! 2013 Elsevier Ltd. All rights reserved. 1. Introduction Camera calibration is a fundamental stage in computer vision. The process is the determination of the relationship between a ref- erence plane and the camera coordinate system (extrinsics), and between the camera and the image coordinate system (intrinsics). These parameters are very useful to recover metrics from images or applying prior information of 3D models to estimate 2D pose of targets, giving an idea of the size of the objects and making their detection and tracking more robust to noise and occlusions. In a previous paper Álvarez et al. (2012), the authors presented a target detection system for transport infrastructures based on manual camera calibration through vanishing points. After that, the approach was improved, as described in Álvarez et al. (2013), with a preliminary automatic calibration method based on camera zooming and zebra-crossings. The current paper extends these works with a hierarchical camera auto-calibration system, which deals with most traffic scenarios and configurations with no restrictions. The work begins with the paper presented in Álvarez et al. (2011). The standard method to calibrate a camera is based on a set of correspondences between 3D points and their projections on im- age plane as presented by Hartley and Zisserman (2000) and Tsai (1986). However, this method requires either prior information of the scene or calibrated templates, limiting the feasibility of surveillance algorithms in most possible scenarios. In addition, cal- ibrated templates are not always available, they are not applicable for already-recorded videos and if the camera is placed very high their small projection can derive in poor accurate results. Finally, in case of having PTZ cameras, using a template each time the cam- era angles or zoom changes is not feasible. One novel method which solves the problem of the template is the orthogonal calibra- tion proposed by Kim (2009). The system extracts the world coor- dinates from aerial pictures (on-line satellite images) or GPS devices to make the correspondences with the image captured. However this system is dependent on prior information from an external source and it does not work indoor. Therefore auto-cali- bration seems to be the more suitable way to recover camera parameters for surveillance applications. One of the distinguished features of the perspective projection is that the image of an object that stretches off to infinity can have finite extent. For example, parallel world lines are imaged as con- verging lines, which image intersection point is called vanishing point. Caprile and Torre (1990), developed a new method for cam- era calibration using simple properties of vanishing points. In their work, the intrinsic parameters of the camera were recovered from a single image of a cube. In a second step, the extrinsic parameters of a pair of cameras were estimated from an image stereo pair of a suitable planar pattern. The technique was improved by Cipolla et al. (1999), who computed both intrinsic and extrinsic parameters from three vanishing points and two reference points from two views of an architectural scene. However these assumptions were incomplete, because as demonstrated by Hartley, Zisserman and Liebowitz in different publications, and summarized in Hartley and Zisserman (2000), it is possible to obtain all the parameters needed to calibrate a camera from three orthogonal vanishing points. From the mentioned works, a lot of research has been done to calibrate cameras in architectural environments (Rother, 2002; Tardif, 2009, etc.. . .). All these methods are based on scenarios 0957-4174/$ - see front matter ! 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.08.050 ! Corresponding author. Tel.: +34 91 885 6702. E-mail address: sergio.alvarez@aut.uah.es (S. Álvarez). Expert Systems with Applications 41 (2014) 1532–1542 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2013.08.050&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.08.050 mailto:sergio.alvarez@aut.uah.es http://dx.doi.org/10.1016/j.eswa.2013.08.050 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa where the large number of orthogonal lines provide an easy way to obtain the three orthogonal vanishing points. Nevertheless, in absence of so strong structures, as usual in the case of traffic scenes, the vanishing point-based calibration is not applicable. In this context, a different possibility is to make use of object motion. The complete camera calibration work using this idea was introduced by Lv et al. (2006). The method uses a tracking algorithm to obtain multiple observations of a person moving around the scene. Three orthogonal vanishing points are then com- puted by extracting head and feet positions in their leg-crossing phases. The approach requires accurate localization of these posi- tions, which is a challenge in traffic surveillance videos. Further- more, the localization step uses FFT based synchronization of a person’s walk cycle, which requires constant velocity motion along a straight line. Finally it does not handle noise models in the data and assumes constant human height and planar human motion, so the approach is really limited. Based on this knowledge, in Junejo (2009) it is proposed a quite similar calibration approach for pedestrians walking on uneven terrains. There are no restrictions as with Lv’s work, but the intrinsic parameters are estimated by obtaining the infinite homography from all the extracted points in multiple cameras. To manage these inconveniences, the solution lies in computing the three vanishing points by studying three orthogonal compo- nents with parallel lines in the moving objects or their motion pat- terns. Zhang et al. (2013) presented a self-calibration method using the orientation of pedestrians and vehicles. The method seems to extract a vertical vanishing point from the main axis direction of the pedestrian trunk, perpendicular to the ground plane. Addition- ally, two horizontal vanishing points are extracted by analysing the histogram of oriented gradients of moving cars. The idea is inter- esting and it was initially implemented for this work. However, the straight vehicles used by Zhang differ from the modern ones, usually with more irregular and rounded shapes. Finally, the pe- destrian detection step is not described and results are not de- picted in the paper. Hodlmoser et al. (2010) present a different approach. They use zebra-crossings with known metrics to obtain the ground plane information, and pedestrians to obtain the verti- cal lines. The problem is the maximum distance that the camera can be from the ground and the necessity of knowing real distances from the scene. In this paper, a self-calibration procedure based on vanishing points is presented. It is done through a hierarchical process which covers most of traffic scenarios and possible configurations. The objective is to obtain both intrinsic and extrinsic camera parame- ters without restrictions in terms of constraints (restrictions men- tioned in previous paragraphs, vehicles driven in only one road direction (Hue et al., 2008), deprecated camera roll (Schoepflin and Dailey, 2003), etc.) or the need of prior information, except for the camera height. After the present introduction, the remainder of the document is organized as follows. Section 2 describes the developed camera auto-calibration method, based on vanishing points, and the hier- archical system proposed. In Section 3, an application of this tech- nique in the context of traffic surveillance is depicted with the developed segmentation and tracking algorithms. The results ob- tained are presented and discussed in Section 4 and finally Section 5 contains the conclusions and future work. 2. Camera autocalibration The camera model used and the equations to obtain the calibra- tion parameters from orthogonal vanishing points are described in the previous paper, Álvarez et al. (2013). In summary, the conclu- sion is that it is possible to calibrate a camera if the principal point and two orthogonal vanishing points are known; or by computing the principal point as the orthocentre of the triangle formed by three orthogonal vanishing points as vertices. The current work is focused on the way to extract these points from common ele- ments of traffic scenarios. Fig. 1 summarizes the proposed camera calibration process. 2.1. Hierarchical auto-calibration This section presents the proposed method to extract the van- ishing points from the image through a hierarchical process. Depending on which elements appear in the scene and the chance of using camera zoom, 5 levels have been established to determine the hierarchy of each developed method and the priority of the solution adopted. Before presenting the hierarchical tree of Fig. 2, and to make its comprehension easier, the different options devel- oped to obtain the vanishing points and optical center are de- scribed in the following paragraphs: ! Zoom: when zooming, if several features of the image are matched between frames they converge in a common point which corresponds to the optical center. ! Crosswalk (cross): the alternate white and gray stripes painted on the road surface provide a perfect environment to obtain two perpendicular sets of parallel lines. It means that two van- ishing points of the ground plane can be extracted. ! Pedestrians (ped): humans are roughly vertical while they stand or walk. This characteristic makes them very useful to extract perpendicular lines to the ground. ! Vehicle motion (vmot/vperp): if one vanishing point from the ground plane is needed, it can be obtained from vehicles mov- ing along the main motion direction (vmot). In case of a perpen- dicular intersection (in 3D coordinates), vehicles along the two main directions will provide perpendicular sets of parallel lines corresponding to the two ground plane vanishing points (vperp). ! Structured scene (struct): in case of scenes with a considerable number of architectural elements, the orthogonal vanishing point extraction can be done by brute force gradient analysis. ! Optical center assumption (OC): when it is not possible to obtain one of the three vanishing points, the optical cener can be assumed as the center of the image, although a small error is committed. The different possible cases are also summarized in Fig. 2 and Table 1. 2.2. Principal point through camera zoom When zooming, if several features of the image are matched be- tween frames, the lines which join the previous and new feature positions converge in a common point which corresponds with the optical center. This effect is demonstrated in Álvarez et al. (2013) and represented in Fig. 3: an image was taken before and Fig. 1. Camera auto-calibration process. S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 1533 after zooming and the matched features converge to the same point, the optical center. 2.3. Zebra crossing vanishing point extraction The alternate stripes painted on the road surface provide a perfect environment to obtain two perpendicular sets of parallel lines. It means that two vanishing points from the ground plane can be computed. The crosswalk detection method is also ex- plained in Álvarez et al. (2013)), and illustrated in the Fig. 4. Firstly, the background model image is binarized, and the lines are extracted by gradient analysis and grouped by angle. After that, a RANSAC-based filter is applied to get the final candidates. The red line is the one which best fits the candidate. Bipolarity and transition analysis is then done in order to obtain a confi- dence factor. Finally, the vanishing points are computed. 2.4. Pedestrian vanishing point extraction Humans are roughly vertical while they stand or walk. This property makes them very useful to get perpendicular lines to the ground, to compute the vertical vanishing point. One option is to extract the vertical component of each pedestrian to form the necessary set of parallel lines, as done by Hodlmoser et al. (2010). However, the cameras in common traffic scenarios are usu- ally located quite higher than the situations proposed by the authors in the paper, and small pedestrians can derive into errone- ous lines extractions. Traffic scenes provide a lot of structured ele- ments with vertical components (walls, lampposts, traffic lights, etc.), that can be used to increase the performance and quality of the system. The developed algorithm is based on this idea, and it is divided into the following steps: pedestrian detection with ver- tical component extraction, scene analysis and vanishing point computation. The aim of this method is to detect pedestrians with no false positives, to avoid lines that are not perpendicular to the ground. Therefore, it is not crucial to detect all the pedestrians in the image but it is important to be sure that the detected objects are humans. In order to obtain useful candidates for vertical lines extraction two kinds of parameters for every moving object are obtained: the motion direction and the main axis direction. The difference of these directions is quite significant for moving pedestrians while it is very small for vehicles. It is evident that this classification is not very accurate, but in practice it is good enough to get valid pedestrians useful to extract vertical lines. To compute the main axis direction of the blob h, three different approaches have been used: moment analysis, principal compo- nent analysis and RANSAC estimation. The direction estimated by moment analysis is defined as: Fig. 2. Hierarchical calibration tree used. Note: Perp. Intersec means perpendicular intersection. Table 1 Cases of the hierarchical tree. CASE ZOOM CROSS PED VMOT PERP STRUCT OC MANUAL 1 x x 2 x x x 3 x x 4 x x 5 x x 6 x x 7 x x 8 x x 9 x x x 10 x x 11 x 12 x Fig. 3. Principal point computation through camera zoom. (a) Image before zooming and extracted features. (b) Image after zooming and extracted features. (c) Feature matching. The common point corresponds to the optical center. 1534 S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 hmoment " tan#1 2l11 l20 # l02 ! " $1% where lpq is the central moment of order (p,q). Principal Component Analysis (PCA) is equivalent to major axis regressions, so the largest axis can be considered as the vertical component. And finally RANSAC algorithm takes the centroid of each candidate row to estimate the line that corresponds to the main axis of the pedestrian. When these three methods obtain sim- ilar results and the blob aspect ratio is valid, the candidate is con- sidered a pedestrian. At the same time, a gradient line extraction of the image is done in order to extract all the possible structured ele- ments. The angle of the vertical components of the pedestrians is compared to the lines extracted and, in case of matching, the lines will be saved to compute afterwards the vanishing point. Due to the perspective of the camera, a perpendicular line to the ground in the image has different angles depending on the position. More- over, because of the negative pitch the vertical vanishing point has to be positive. Therefore the image is divided into five quadrants. Fig. 5 depicts an example of the developed method. Fig. 5(a) represents the lines extracted from the scene, with different colors depending on the belonging quadrant. Fig. 5(b) shows the detected pedestrian inside a green box with the estimated vertical compo- nent in red, and the matched vertical lines in magenta. Finally, Fig. 5(c) depicts the estimation of the vertical vanishing point with all the accumulated vertical lines. Red lines are the outliers and green lines the inliers for a RANSAC-based method to obtain the intersection point. 2.5. Vehicle motion vanishing point extraction One of the properties of the traffic scenarios is that many vehi- cles drive in the same or inverse direction of the 3D world. There- fore the main axis of these vehicles are parallel to each other, and also parallel to the ground plane. This supplies important informa- tion to extract horizontal vanishing points. As explained in the hierarchical calibration tree (Fig. 2), there are cases that need only one ground plane vanishing point while others need two. In case of computing the optical center (either by zooming analysis or assuming it as the image center) and detecting pedestrians, only one vanishing point from the ground plane is needed, in any direction. On other hand, in case of needing two ground plane vanishing points and if a perpendicular intersec- tion (in 3D coordinates) is present in the scene, vehicles moving along the two main directions will provide perpendicular sets of parallel lines corresponding to the two ground plane vanishing points. In both cases the followed process is similar, done either for one direction or two respectively. Firstly, the main motion directions are extracted. For this pur- pose, a feature optical flow analysis of the foreground blobs is done and their motion direction is saved into an histogram. Once it is Fig. 4. Crosswalk detection example. (a) Binarized background model. (b) Line extraction. (c) Grouped candidates with testing lines in red. (d) Parallel lines to compute the vanishing points. (For interpretation of the references to colour in this figure caption, the reader is referred to the web version of this article.) Fig. 5. Vertical vanishing point extraction example. (a) Extracted scene lines divided by 5 quadrants. (b) Detected pedestrians with red vertical component and vertical matches in magenta. (c) RANSAC vanishing point estimation with red outliers and green inliers. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 1535 constructed after a determined number of frames, an EM algorithm is used to fit the histograms into gaussians in order to get the prin- cipal components of the movement. Fig. 6 shows an example of a perpendicular intersection, where the features of the foreground objects are tracked by optical flow and the motion direction histo- gram with the gaussian components in red is computed. The verti- cal axis corresponds to the frequency of the angle, and the horizontal axis corresponds to the angle value between 0" and 180". These values are not perpendicular in image coordinates due to the perspective projection. After getting the main directions of the scene, the motion of each foreground blob is analysed. In case of detecting motion in the computed directions, the gradients of the blob are extracted in order to look for parallel lines with the mentioned angles. Once obtaining a representative number of parallel segments, a RANSAC- based method to obtain the intersection point is used. Fig. 7 shows an example of two ground plane vanishing point extraction using the method explained in this section. 2.6. Structured scenarios vanishing point extraction In the case of having a considerable number of architectural ele- ments in the scene, a last option for an autonomous calibration is available (although less common and effective). It consist on extracting the vanishing points by brute force gradient analysis, assuming that the three sets of parallel lines with most number of lines are orthogonal. To group the lines, J-Linkage algorithm (Toldo and Fusiello, 2008) is used. This method is based on the work of Tardif (2009), although he does not look for orthogonal vanishing points. Fig. 8 shows the orthogonal lines extracted in a structured scenario to compute the three orthogonal vanishing points. 3. Traffic target detection and tracking After calibrating the camera, an approximate size of pedestrians and vehicles in the image can be obtained using a standard size for them in world coordinates. This step will give the system a notion of how big are the searched elements. In this section, a multilevel framework to detect and track pedestrians and vehicles is pre- sented. Fig. 9 illustrates the flowchart of the proposed framework, which consists of 3 levels: 1) Image segmentation level, to create and handle a background model and to obtain the foreground ob- jects without image noise, camera vibrations or illumination ef- fects; 2) features level, which extracts and follows features of the foreground objects; 3) objects level, which is in charge of managing occlusions and create and track object clusters. The first and sec- ond levels are similar than the ones described by the authors in Álvarez et al. (2012), and the objects level is improved in the current work as explained next. 3.1. Objects level Usually, feature grouping works associating features directly into objects using proximity and motion history. However, the dis- tance between two features that belong to the same object can be much larger than two features that belong to two nearby objects, which can confuse the system. To efficiently deal with the problem, Fig. 6. Example of main motion directions extracted in a perpendicular intersection. (a) Perpendicular intersection. (b) Foreground optical flow analysis. (c) Histogram of directions and fitted gaussians in red. (For interpretation of the references to colour in this figure caption, the reader is referred to the web version of this article.) Fig. 7. Example of ground plane vanishing point extraction in a perpendicular intersection. Fig. 8. Extracted lines in a structured scenario to obtain three orthogonal vanishing points automatically. 1536 S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 a multilevel grouping algorithm is presented. First, an occlusion reasoning step is done in order to split foreground blobs from dif- ferent objects. After that, the individual features are associated to a blob and grouped into clusters depending on their motion and 3D sizes. Finally, the objects are tracked. 3.1.1. Partial occlusion reasoning The first step when considering this problem is to observe the shapes of the objects involved in an occlusion. A common charac- teristic is that the shapes generated by an occlusion are not uni- form: non-occluded objects are generally convex, whereas the shape of partially occluded objects become concave. An example of non-occluded and occluded objects is given by comparing their convex hull in Fig. 10. It canbeseenthatnon-occludedobjectscanreachagoodfitbytheir convex hull, which does not hold for occluded objects. Accordingly, if there is an approximate idea of the searched objects sizes (through the camera calibration), an occlusion can be figured out by studying the blob shapes and their convex hulls. In particular one simple shape descriptorhasbeenwidelyusedinthistask:theshapecompactness. Itis an intrinsic characteristic of the object shapes defined by: C " P2 A $2% where A is the shape area and P is the shape perimeter or boundary length. This way to measure shape compactness is taken from the isoperimetric inequality (Montero and Bribiesca, 2009). The next step to evaluate if a blob is the result of an occlusion is to compare the shape compactness of the object (Co) and the one of its convex hull (Ch). Obviously Co is always greater than Ch, because the area of an object is smaller than the one of its convex hull, whereas the boundary length of an object is greater. Therefore, for non-occluded objects Ch is close to Co, and for occluded ones Ch is smaller than Co. The ratio between both descriptors is used to discriminate both sit- uations. It is called compactness ratio and it is defined by: CR " Ch Co $3% Another parameter used to detect occlusions is the convexity, determined by the ratio between areas as: RA " Ao Ah $4% where Ao and Ah represent the area of the object and the area of the object’s convex hull respectively. Since the denominator is always greater than the numerator, RA is always less than one. For a non- occluded object its shape is convex and RA is close to 1, whereas for occluded objects RA is far less than 1. The third estimator to consider a blob as an occlusion is its size. After calibrating the camera, the approximate sizes of the pedestri- ans and vehicles located in the ground plane are known. In case of occlusions these sizes will be considerably increased. If the three parameters described above indicate an occlusion, the occlusion reasoning method is run as described in the flowchart of Fig. 11. Fig. 9. Flowchart of the proposed framework to detect and track pedestrians and vehicles. Fig. 10. Object occlusions and convex hull. The convex hull is represented in white and the foreground blob in gray color. Fig. 11. Flowchart of the occlusion reasoning method. S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 1537 An useful way to understand the shape of an object contour is to compute its convex hull and convexity defects. Fig. 12 illustrates these concepts using the image of a vehicle occlusion. The gray area corresponds to the foreground blob and the coloured areas represent the different defects of the convex hull. Finally, the red marks correspond to the farthest points from the convex hull with- in each defect, also called defect points. The distance between the farthest defect point and the convex hull is taken, and this point is selected as the first cutting point. The next objective is to find an optimum second cutting point to create a cutting line which separates the blob into two different ob- jects. To extract the second point, the occluded object is sequen- tially cut by segments that join the cutting point with the rest of defect points. For every line, the area and compactness ratios for each new blob are computed. The chosen cutting line is the one that brings the maximum ratio given by the Eq. (5). Ratio " X2 i"1 RAi & CRi 2 $5% Fig. 13 depicts some examples of occlusion reasoning using the method explained before. This procedure does not require prior knowledge but the known measures from camera calibration. By using this method, most partial occlusions can be effectively handled. 3.1.2. Feature clustering and model fitting To group all the features from the same object, a 2-stage 3D clustering algorithm is used. First the individual features are as- signed to a blob (after the occlusion reasoning) and grouped into clusters depending on their motion. Finally these clusters are grouped into objects depending on the 3D sizes and motion. Therefore, if a blob corresponds to a single object, all its features will have a similar motion and will be grouped together. Other- wise, they will be clustered into multiple objects associated to dif- ferent motion characteristics. As an unsupervised stage, it is necessary to identify the number of clusters and the correspon- dence of the samples automatically. Mean Shift (Comaniciu and Meer, 2002) is used as a non-parametric method which does not require prior knowledge of the number of clusters, and does not constrain their shape. The main idea behind mean shift is to treat the points in the d-dimensional feature space as an empirical probability density function where dense regions in the feature space correspond to the local maxima or modes of the underlying distribution. For each data point in the feature space, one performs a gradient ascent procedure on the local estimated density until convergence. The stationary points of this procedure represent the modes of the dis- tribution. Furthermore, the data points associated with the same stationary point are considered members of the same cluster. The quality of the output is controlled by a kernel bandwidth, and it is not critical due to objects moving with different angles or veloc- ities generate features with a strong different component. Fig. 14 depicts an example of the feature clustering step. As mentioned before, an approximate size of vehicles is known thanks to the information provided by the camera calibration. Therefore, a vehicle which has been split into several blobs due to errors in the foreground or a misclassified occlusion can be merged. If the clusters fits into the 3D size of an standard target in the corresponding 2D coordinates and have similar motion, the clusters are merged into a final object. Fig. 15 shows an exam- ple of blob merging after splitting the blob due to an occlusion with a tree. 3.1.3. Cluster tracking After detecting consecutively a cluster several times, a tracking stage combined with a multi-frame validation process takes place. This final step is used to reinforce the coherence of the detected objects over time, obtaining a more stable position, avoiding occlu- sions in case the previous methods fail, and minimizing the effect of both false-positive and false-negative detections. The multi- frame validation and tracking algorithm relies on the Kalman filter theory in 2-D space, with a state vector based on the ellipse parameters: centroid, axis and angle. For the data association, Hun- garian assignment is used. Fig. 16 depicts a performance example of the detection system after the steps described. Fig. 12. Blob, convex hull and convexity defects in an occlusion example. Fig. 13. Examples of occlusion management by the proposed algorithm. 1538 S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 4. Experimental results 4.1. Camera auto-calibration For the auto-calibration method, 30 sequences from different scenarios and conditions have been used, testing the 12 cases of the hierarchical tree. As a result, a comparative table (Table 2) has been constructed with the average errors of the main intrinsic and extrinsic parameters extracted (focal distance, pitch and roll), compared to the cases 5 and 12 which are considered the ground- truth. Yaw is not used because its variation does not modify the ground plane and does not have impact into the 3D projection. As can be seen, case 1 is the best solution due to the strong par- allel component of their orthogonal elements and the zooming chance. Near it, cases 6 and 7 have similar results. It was expected because they are the relative cases to the first one, but without zoom. On the other hand, the worst options are cases 3, 4, 10 and 11, based on perpendicular intersections (not always available or strictly perpendicular), and structured scenes (not always with strictly orthogonal components). The obtained results are really satisfactory: the low error ob- tained proves the strength of the system, and the multiple options of the hierarchical tree provide high versatility to cover most of the possible traffic scenarios. Furthermore, the system is able to adapt the calibration parameters in case of PTZ camera displacements without manual supervision. Even if there is no chance to auto-cal- ibrate the camera (due to absence of orthogonal components), the manual input of lines remains as a valid option which allows the user to control the system in a short time. 4.2. Target detection system Firstly, the performance of the proposed object occlusion rea- soning framework has been quantitatively evaluated. The results are summarized in Table 3, separated by occlusion class and depending on the level of the algorithm where they were detected and managed. Detected columns stand for the number of occlusions detected by each level, and handled is the number of occlusions correctly managed. The total column contains all the detected Fig. 14. Examples of feature clustering represented by coloured features. Fig. 15. Examples of cluster merging. Fig. 16. Example of the target detection system. Table 2 Auto-calibration errors comparative table. Case Focal (%) Pitch (") Roll (") 12/5 0.00 0.00 0.00 1 2.29 1.68 0.30 2 4.69 2.83 0.34 3 5.14 2.55 0.65 4 6.68 3.05 0.67 6 3.52 1.46 0.51 7 3.88 2.05 0.51 8 4.05 2.25 0.69 9 4.40 2.57 0.26 10 7.47 3.11 0.64 11 7.18 3.16 0.65 S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 1539 and non-detected occlusions and it is used to evaluate the rate values as Handled/Total. Occlusion level always takes part in the process and only if it can not detect or handle the occlusion, the algorithm passes through the next level. The global occlusion management ratio (91%) is very reason- able. It is important to emphasize that this analysis is single frame. Therefore an error due to an occlusion in a particular frame is not important in the whole path of an object. Moreover, after the track- ing stage this value is increased to 95% because several occlusions are managed by the multi-frame validation. The advantage of the approach is the use of a multi-level framework that allows to solve an occlusion from different and complementary points of view. To analyse the performance of the global target detection sys- tem, the algorithm has been tested on over 2 h of traffic videos with more than 2000 objects between vehicles and pedestrians. The sequences include different camera views, illumination effects, shadows, etc., in order to evaluate the method in a wide range of situations. Some examples of the testing scenarios used are shown in Fig. 17 and described in Table 4. The global results of the application are depicted in Table 5 in terms of object detection rate, recall and precision. The Detection Rate (DR) is the percentage of correctly detected objects, the Recall (R) measures the system’s ability to identify positive samples, and the Precision (P) is the fraction of retrieved instances that are rele- vant. where TP stands for the number of true positives (objects cor- rectly detected at least the 80% of their path), FP stands for the number of false positives (unexpected detections or object splits) and FN is the number of false negatives (missing detections). From a total amount of 2269 objects, the system has obtained a detection rate of 93.3%, which is considered a good result and valid for the proposed application. To analyse the importance of the cal- ibration stage, this value is compared with the one obtained with- out calibrating the camera as represented in Fig. 18. Firstly, the system works correctly with the parameters obtained by an initial camera auto-calibration. Between frames 44 and 54, the camera changes its angles and zooms out. Then, a comparison between Table 3 Quantitative evaluation of the occlusion reasoning framework. Occlusion level Clustering level Together Detected Handled Detected Handled Detected Handled Total Rate Ped& Ped 226 213 11 11 237 224 251 0.89 Car& Car 124 115 19 18 143 133 142 0.93 Car& Ped 53 51 29 26 82 77 85 0.90 Result: 403 379 59 55 462 434 478 0.91 Fig. 17. Samples of testing scenarios. Table 4 Description of testing videos. Video number of frames Resolution Conditions Source video1 16402 640 ' 480 Cloudy Own sequence video2 5244 640 ' 480 Dusk (dark) Own sequence video3 3332 640 ' 480 Dusk (bright) Own sequence video4 18296 640 ' 480 Sunny Lunds Univ. video5 15921 640 ' 480 Cloudy Own sequence video6 3585 640 ' 480 Sunny Own sequence video7 630 768 ' 576 Fog/rain Karlsruhe Univ. video8 4290 352 ' 288 Cloudy Candela Table 5 Results of the target detection system. N: number of samples. TP: number of true positives. FP: number of false positives. FN: number of false negatives. DR: detection rate. R: recall. P: precision. Scenario N TP FP FN DR R P Sunny (shadows) 901 832 39 32 0.923 0.963 0.955 Cloudy 885 841 23 43 0.950 0.951 0.973 Dusk 312 291 17 15 0.933 0.951 0.945 Rain/snow 171 152 17 13 0.889 0.921 0.899 Total 2269 2116 96 103 0.933 0.954 0.957 1540 S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 the system with and without auto-recalibration is done (through OC computation and vanishing points extracted from a crosswalk). The blue line represents the detection rate of the auto-recalibrated system (near 93%) and the red line represents the DR in case of the calibration parameters remain constant (near 50–60%) with graphical examples. It demonstrate the need to work in fully self-adaptive mode. 5. Conclusions and future work In this paper, a novel hierarchical camera self-calibration proce- dure based on vanishing points has been presented. Depending on which elements appear in the scene and the chance of using cam- era zoom, 5 levels have been established to determine the hierar- chy of each developed method and the priority of the solution adopted. It is an important step for many possible applications, be- cause it provides very useful information to compute an approxi- mate size of the searched objects. In this context, a monocular system has been developed to detect and track vehicles and pedes- trians for applications in the framework of Intelligent Transport Systems. Trough the auto-calibration step, the algorithm requires no object model or prior knowledge (only an approximate size of the searched objects in world coordinates), it can work indoor and outdoor, in different conditions and scenarios. Moreover, it is completely autonomous (‘‘plug & play’’), independent of the posi- tion of the camera and able to manage PTZ changes in fully self- adaptive mode. From the results and conclusions of the present work, several future lines for each treated topic are devised. With respect to the camera auto-calibration, an interesting improvement is related to the recalibration process in case of PTZ displacements. The idea is to develop a segment tracking, to use the same set of orthogonal lines to find the new position of the previously used vanishing points. Besides that, due to the high diversity of camera views, operating conditions and observation objectives in traffic surveil- lance, there is an important lack of a common framework and most authors use their proprietary sequences. This condition has gener- ated a large diverse body of work, where it is difficult to perform direct comparison between the proposed algorithms. It would be very important to generate a public traffic database, with a wide range of scenarios and conditions, to be able to make these comparatives. Acknowledgment This work has been supported by the Spanish Ministry of Sci- ence and Innovation by means of Research Grant ONDA-FP TRA2011-27712-C02-02. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.eswa.2013.08. 050. References Álvarez, S., Sotelo, M. A., Llorca, D. F., & Quintero, R. (2011). Monocular vision-based target detection on dynamic transport infraestructures. In Lecture notes in computer science (pp. 576–583). Álvarez, S., Llorca, D. F., Sotelo, M. A., & Lorente, A. G. (2012). Monocular target detection on transport infrastructures with dynamic and variable environments. In IEEE intelligent transportation systems conference. Álvarez, S., Llorca, D. F., & Sotelo, M. A. (2013). Camera auto-calibration using zooming and zebra-crossing for traffic monitoring applications. In IEEE intelligent transportation systems conference. Caprile, B., & Torre, V. (1990). Using vanishing points for camera calibration. International Journal of Computer Vision, 4, 127–140. Cipolla, R., Drummond, T., & Robertson, D. (1999). Camera calibration from vanishing points in images of architectural scenes. Comaniciu, D., & Meer, P. (2002). Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24, 603–619. Hartley, R., & Zisserman, A. (2000). Multiple view geometry in computer vision. Cambridge University Press. Hodlmoser, M., Micusik, B., & Kampel, M. (2010). Practical camera auto- calibration based on object appearance and motion for traffic scene visual surveillance. In IEEE conference on computer vision and pattern recognition (pp. 1–8). Hue, T., Lu, S., & Zhang, J. (2008). Self-calibration of traffic surveillance camera using motion tracking. In IEEE conference on intelligent transportation systems. Junejo, I. N. (2009). Using pedestrians walking on uneven terrains for camera calibration. Machine Vision and Applications, 22, 137–144. Kim, Z. (2009). Camera calibration from orthogonally projected coordinates with noisy-ransac, In IEEE workshop on application of computer vision. Lv, F., Zhao, T., & Nevatia, R. (2006). Camera calibration from video of a walking human. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(9), 1513–1518. Montero, R., & Bribiesca, E., (2009). State of the art of compactness and circularity measures. In International mathematical forum. Rother, C. (2002). A new approach to vanishing point detection in architectural environments. Image and Vision Computing, 20, 647–655. Fig. 18. Comparative example of the system with and without recalibration after a camera change. S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 1541 http://dx.doi.org/10.1016/j.eswa.2013.08.050 http://dx.doi.org/10.1016/j.eswa.2013.08.050 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0005 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0005 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0010 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0010 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0010 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0015 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0015 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0020 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0020 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0025 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0025 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0025 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0030 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0030 Schoepflin, T., & Dailey, D. (2003). Dynamic camera calibration of roadside traffic management cameras for vehicle speed estimation. IEEE Transactions on Intelligent Transportation Systems, 4(2), 90–98. Tardif, J. P. (2009). Non-iterative approach for fast and accurate vanishing point detection. In IEEE conference on computer vision. Toldo, R., & Fusiello, A. (2008). Robust multiple structures estimation with j-linkage. In European conference on computer vision (pp. 537–547). Tsai, R. (1986). An efficient and accurate camera calibration technique for 3d machine vision. In IEEE conference on computer vision and pattern recognition. Zhang, Z., Tan, T., Huang, K., & Wang, Y. (2013). Practical camera calibration from moving objects for traffic scene surveillance. IEEE Transactions on Circuits and Systems for Video Technology, 23, 518–533. 1542 S. Álvarez et al. / Expert Systems with Applications 41 (2014) 1532–1542 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0035 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0035 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0035 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0040 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0040 http://refhub.elsevier.com/S0957-4174(13)00671-4/h0040 Hierarchical camera auto-calibration for traffic surveillance systems 1 Introduction 2 Camera autocalibration 2.1 Hierarchical auto-calibration 2.2 Principal point through camera zoom 2.3 Zebra crossing vanishing point extraction 2.4 Pedestrian vanishing point extraction 2.5 Vehicle motion vanishing point extraction 2.6 Structured scenarios vanishing point extraction 3 Traffic target detection and tracking 3.1 Objects level 3.1.1 Partial occlusion reasoning 3.1.2 Feature clustering and model fitting 3.1.3 Cluster tracking 4 Experimental results 4.1 Camera auto-calibration 4.2 Target detection system 5 Conclusions and future work Acknowledgment Appendix A Supplementary data References 
work_5pqm2k3m6rdqpf2t7c7shx2ziq	----	PII: 0898-1221(90)90153-B Computers Math. Applic. Vol. 19, No. 11, pp. 105-119, 1990 0097-4943/90 $3.00+0.00 Printed in Great Britain. All rights reserved Copyright © 1990 Pergamon Press plc C O M P U T A T I O N A L M O D E L S O F U N C E R T A I N T Y R E A S O N I N G I N E X P E R T S Y S T E M S J. F. BALDWIN Department of Engineering Mathematics, University of Bristol, Bristol BS8 1TR, England A b s t r a c t - - T h e use o f support pairs associated with the facts and rules o f a knowledge base o f an expert system to capture various aspects o f inductive reasoning is discussed. The concept o f semantic unification is introduced with reference to fuzzy sets theory. In this respect a probabilistic interpretation for this semantic unification is described using a population voting model. Examples are discussed including default reasoning using support logic. I N T R O D U C T I O N In this paper the use o f support pairs associated with Prolog type clauses to capture various aspects o f inductive reasoning will be discussed. This represents a support logic programming system and it has been implemented in the form o f the language Fril [1]. Fril can operate as a pure Prolog system if no uncertainties are involved. False facts, equivalent statements and a true logic negation can be used. Fill is being used on a variety o f applications involving reasoning with uncertainty. These include AI applied to scene analysis, medical and fault diagnosis, expert systems in command and control, analogical reasoning, probabilistic grammars, program evaluation etc. Support logic programming [2-5] is an evidential reasoning system which, rather than proving theorems, collects evidence to support an hypothesis and also to support the negation o f the hypothesis. These supports do not have to add up to unity. If no evidence is available then a support pair (0 1) is returned corresponding to zero support for and zero support against, i.e. total uncertainty. A logic o f support has been studied under various names by different people. K o o p m a n called it a logic o f intuitive probability and Carnap a theory o f confirmation [6]. These and other theories are based on a single number representing the supports and some are comparitive only. Zadeh's fuzzy sets theory forms a basis for a possibility theory [7]. Bellman made contributions to the analytical formalism o f the theory o f fuzzy sets [8]. Support logic programming uses this theory. Fuzzy sets can be used as the referents o f concepts and semantic unification is used to match two fuzzy terms which are syntactically different but which semantically have something in common. An interpretation o f fuzzy sets which shows how semantic unification can be given a probabilistic interpretation, necessary for support logic programming, is given below. The calculus o f Fril is consistent with probability theory and this supposes that all propositions are true or false, although it may not be possible to acquire evidence which will allow a probability o f 1 or 0 to be obtained. Expert systems and other knowledge engineering systems such as vision understanding and speech recognition programs must be able to cope with knowledge bases with incomplete information. Incomplete information can be o f two types: one type concerns lack o f data and the other, lack o f concept definition. For example, in a vision system only part o f the object may be in view because o f occlusion. This gives rise to uncertainty concerning what the object may be since the part that can be seen could belong to several different objects. Possible extensions o f the part o f the object will give rise to different interpretations and each extension will have a probability associated with it. This probability is assessed from other evidence picked up from other parts o f the picture and may not be able to be assessed with complete accuracy. It may be possible to assess its value as being contained within a certain interval. On the other hand, the whole object may be in complete view but it is still difficult to say exactly what it is. F o r example, if the object in total view was a bushy tree like object, no complete support 105 106 J.F. BALDWIN could be given for it being a tree or for it being a bush, simply because o f a lack o f precise definition for bush and tree. Most concepts which we use in our daily lives are o f this nature. F o r example we cannot prescribe sufficient and/or necessary conditions for classifying what we mean by a "humane society", " a good business venture", " a comfortable seat", "a well-structured program", "a stable system", " a reliable system" etc. Even accepting that these definitions are context dependent, we will still have difficulty in giving exact definitions within a given context. Furthermore, the relevant context may again be a border line decision. In law, cases are often looked at by considering similar cases from the past where the judgements are known. The present case may not precisely fit any of the historical ones but only have similarities with each. The judgement on the present case will then be influenced in some form by the combination o f those judgements of similar past cases. What form this so called combination should take is not at all obvious. Heuristics used by experts are probabilistic in nature. Truth is not guaranteed when certain conditions hold. Difficulties o f entailment when true propositions are replaced by highly probable propositions are well-known. Contraposition is valid for deductive entailments but does not hold for high probabilities. Deductive entailment is transitive but strong inductive support is not. More importantly the following valid argument o f deductive logic does not carry over into the inductive case. I f A entails B then A A N D C entails B. In fact if Pr(B [A) is high, this provides no constraint on the value o f Pr(B I A, C) since Pr(C 1,4, B). Pr(B I A) Pr(B I A, C) = P r ( C I A ) This, o f course, is obvious from sample space considerations. All relevant criteria must be considered when giving supports to predicates as suggested by Hempel's maximum specificity conditions [9]. Causal connections are important in expert systems. Any sensible theory o f causation is probabilistic. Frequent conjunctions often occur, constant conjunctions rarely [10], The modelling o f causality is also discussed in Ref. [11]. Probabilistic reasoning can be viewed as constraint reasoning in which the various probabilistic statements given provide evidence to constrain the probability o f another statement to be contained within a certain interval. If we know that: Pr(P --~ Q) = 2/3, Pr(P) = 4/5, what can we conclude a b o u t Pr(Q)? A point value probability cannot be determined since the above two probabilistic statements gives insufficient information for this. The statements constrain the interval which contains Pr(Q). Three possible cases must be analysed, since the case corresponding to N O T ( P - - . Q) A N D (NOT P ) is not possible because o f the inconsistency of the two statements. Case 1 Case 2 C a s e 3 P--*Q NOT(P---~ Q ) P.--*Q P P N O T P Q NOT P W o r l d I: Q W o r l d 2: N O T Q Q: {1,1} Q: {0,0} Q: {0,1} XI X 2 X 3 where xt is Pr(case i) and {a, b} means that N E C ( Q ) = a, POS(Q) = b, where a = 0 if N E C ( Q ) is false and 1 if it is true, and b = 0 if POS(Q) is false and 1 if it is true. N E C and POS are modal logic operators. Then since Xl + X2+X3---- l, X I + X 2 ~--- 4/5, Computational models 107 since P r ( P ) = Pr(P A N D P - - - . Q ) + Pr(P A N D N O T (P---.Q)) x l + x 3 = 2 / 3 , since Pr(P---*Q)= Pr(P--*Q A N D P ) + Pr(P---~Q A N D N O T P ) so that x1=7/15, x2=1/3, x3=1/5. Hence Pr(NEC Q) = 7/15 and Pr(POS Q) = 7/15 + 1/5 = 2/3. From Pr(NEC Q) ~< Pr(Q) ~< Pr(POS Q) it follows that Pr(Q) lies in the interval [7/15, 2/3]. This interval can be determined using linear programming formulations and this is discussed in Ref. [5]. S U P P O R T P A I R S The theory o f uncertainty which forms the basis o f support logic programming is based on the association o f support pairs with Horn clauses as used in Prolog. Any proposition P is assumed to be true or false. A two valued logic is assumed. There is no mention o f truth values lying between 0 and 1. Furthermore any valid formula o f first order logic, F say, will be such that there is support o f 1 for and 0 against. Evidence, E, is used to assign a necessary support, Sn(P J E) for, and a necessary support, Sn(NOT P I E) against any proposition P being true. Possible supports Sp(P I E) and Sp(NOT P I E ) are defined as Sp(P I E ) = 1 - Sn(P I E); Sp(NOT P I E ) = 1 - Sn(NOT P I E ) . These can be further interpreted in terms o f the modal logic necessity and possibility operators, namely Sn(P I E) = Pr(NEC P I E ) ; Sp(P I E) = Pr(POS P I E ) , where modal operators are to be understood in the context o f possible world semantics. The belief that the truth value 1 can be assigned to P using evidence E, Pr(P I E), lies in an interval determined by the necessary and possible supports for P : Pr(P [E) lies in [Sn(P I E), Sp(P [E)]. It is necessarily true that S n ( P A N D N O T P ) = 0 and Sp(P A N D N O T P ) -- 0, Sn(P O R N O T P ) = I and Sp(P O R N O T P ) = I. S U P P O R T L O G I C P R O G R A M M I N G A support logic program consists o f a sequence o f support clauses. A support clause is a clause with an associated support structure. A clause is a list o f one or more atoms. An atom is an atomic formula which is a list whose first element is apredicate symbol or a relation and the remaining elements are terms. A term is a number, constant, variable or list. The elements o f a list are terms. A support structure can be a single support pair or a list o f two support pairs. A support pair is a list o f two elements; the first element being called the necessary support and the second element the possible support. Variables, constants, numbers and lists have their usual meaning. Support clauses can be further divided into simple support clauses and compound support clauses. CAMWA 19/I I - - H 108 J . F . BALDWIN An example o f a simple support clause is: ((coml2 is a large senate committee)):(0.6 0.8), which could mean that the degree o f belief that c o m l 2 is a large senate committee is some number in the interval [0.6 0.8]. The doubt expressed by using this interval arises because o f the imprecise definition o f large. This support pair could be determined by asking a large representative sample o f university members to vote whether they accepted that senate committees o f various sizes were large. The vote could be "yes", " n o " or "abstain" for each size presented. The proportion who vote "yes" for a given size would represent the necessary support for it being a large committee. This number plus the number o f abstentions would give the possible support. The number o f " n o ' s " would give the necessary support against and this with the abstentions would give the possible support against. O f course, the doubt could also arise because o f the uncertainty in the actual numbers on a committee. The final support pair used must take account of both these cases of uncertainty and this could be done using more rules to determine the support pair. An example o f a compound statement with a single support pair is: ((committee C contains a professor) (C is a large senate committee)):(0.9 1), which says that at least 90% of large senate committees contain a professor. In other words the conditional probability Pr(committee C contains a professor lC is a large committee) lies in the interval [0.9 1]. In Prolog terms this corresponds to a fact. The simple clause ((p)): (0 0) says that p is false. An example o f a compound statement with two support pairs is: ((performance X good) (engineers_report X ok) (efficiency X near_optimal)): ((0.9 1) (0 0.2)), which says that if the body o f the rule, in this case the conjunction o f the two atoms (engineers_report X ok) and (efficiency X near_optimal), is true then the probability that the performance o f X is good lies in the interval [0.9 1], while if the body is false this probability lies in the interval [0 0.2]. A special case o f this rule, namely ((p)(q)):((l 1)(0 0)) says that p is equivalent to q. THE C A L C U L U S OF S U P P O R T LOGIC The calculus used in support logic programming is fully described in Ref. [5]. We will not repeat this here but discuss a simple example to illustrate the main points o f the calculus. Since propositions are assumed to be either true or false, assuming one accepts the scoring argument o f De Finetti and Lindley [12] then if the support pairs correspond to single numbers, a probability calculus must be used. The calculus for the support pairs is then easily determined since probabilities are contained within intervals determined by the support pairs. A unique probability is not determined and a simple constrained optimisation problem gives the required support pair for any compound statement in terms o f the support pairs o f its parts. Consider the following example in which we know that P r ( a l q ) = 0 . 5 , P r ( a I N O T q ) = 0 . 4 , Pr(a Is) = 0.8, Pr(a LNOT s) = 0.4, Pr(q) = 0.7, Pr(s) = 0.175, Computational models 109 then we can determine Pr(a) in two ways, namely Pr(a) = Pr(a Iq)" Pr(q) + Pr(a tNOT q). Pr(NOT q) = 0.47, Pr(a) = Pr(a Is). Pr(s) + Pr(a I N O T s). Pr(NOT s) = 0.47. If the answers in each case had been different, we would have concluded that the knowledge base was inconsistent. The problem expressed in Fril is: ((a) (q)):((0.5 0.5)(0.4 0.4)), ((a) (s)):((0.8 0.8)(0.4 0.4)), ((q)) :(0.7 0.7), ((s)):(0.175 0.175), and the query yields the solution qs((a)), (0.47 0.47). Fril uses the two p r o o f paths to provide the answer to the query and gives the answer (0.47 0.47) in each case. These intervals are intersected to give (0.47 0.47) as the final solution. We will now consider a modified problem in which the point probabilities are not precisely known. Pr(a Jq) is in [0.45, 0.55], Pr(a Is) is in [0.75, 0.85], Pr(q) is in [0.65, 0.75] Pr(s) is in [0.1, 0.2] Pr(a I N O T q) is in [0.35, 0.45], Pr(a INOT s) is in [0.35, 0.45], yields the solution and the query ((a)) (q)):((0.45 0.55)(0.35 0.45)), ((a) (s)):((0.75 0.85)(0.35 0.45)), ((q)):(0.65 0.75), ((s)):(0.1 0.2), qs((a)), (0.38 0.545). The basic rule used to combine support pairs from different p r o o f paths is the intersection rule. An alternative method o f combining p r o o f paths is available in Fril and corresponds to using a Dempster type rule [13]. This should only be used when the p r o o f paths correspond to independent viewpoints. In this case conflicts can occur and the Dempster rule is one way o f resolving the conflicts. If the user has some other way he wishes to combine solutions from different viewpoints he can express this as a rule in Fril. Nothing has been said a b o u t finding the support pairs o f a conjunction or disjunction when given support pairs for each atom. The rules used for Fill are consistent with probability theory. We can use the theorem o f total probability as before to obtain P(a) but we must use interval arithmetic. The two methods give [0.38, 0.57] and [0.355, 0.545], respectively for Pr(a). Any point in the final interval containing the point probability Pr(a) must lie in both these intervals using a consistency argument. Therefore we must intersect the intervals to obtain the final interval. This defines the rule o f how solutions are combined from different p r o o f paths in Fill. For this case the final answer for Pr(a) is that it is contained in the interval [0.38, 0.545]. The Fril program for this case is 110 J . F . BALDWIN D E F A U L T R E A S O N I N G Consider the Fill program: ((live_.another_five_years X) (english X) (age X 30) (not suffers_from_lung_cancer X)):(0.9 l) ((live_another_five_years X) (english X) (age X 30) (suffers_from_lung_cancer X)):(0 0.1). I f we do n o t k n o w a n y t h i n g a b o u t the health o f a 30-year old English person, these two rules will use ((suffers.from_lung_cancer person)):(0 1) and conclude (0 l) as the support pair for (live an- other_five_years person). Intuitively we m a y feel t h a t an answer something like (0.85 1) should have been given, since we k n o w that most 30-year old Englishmen do live a n o t h e r five years. We could therefore add the additional rule to our program: ((live_another_five_years X) (english X) (age X 30)):(x y ) , where x a n d y are chosen appropriately. W h a t is appropriate? Strictly (x y ) = (0 1) to be consistent with the other two rules. But this would not satisfy the reason we are introducing this rule. We will choose (x y ) = 0.85 1). I f n o t h i n g is k n o w n a b o u t the health o f the person then Fril will use each o f the rules, obtaining (0 1) f r o m the first two and (0.85 1) from the last giving the answer (0.85 1). I f the second rule is applicable then rules 2 a n d 3 will give inconsistent answers. Fill recognizes this a n d because the b o d y o f rule 3 is contained in the b o d y o f rule 2, ignores rule 3 a n d uses the first two rules only. This is a consequence o f the m a x i m u m specificity requirement. N o n - m o n o t o n i c logic is used to avoid problems like this but these logics have inconsistencies [14]. By using Fril there is no reason to introduce these various additional logics a n d default reasoning. In the case o f the s t a n d a r d problem that "all birds can fly", " a penguin is a bird . . . . a penguin c a n n o t fly" it is, o f course, false to say t h a t all birds can fly. M o s t birds can fly so t h a t if all t h a t is k n o w n is that X is a bird there is a high probability that it can fly. This will n o t be the case if X is a penguin and is treated as the above problem. Details o f this and similar problems can be f o u n d in Refs [3, 5]. A recursive definition o f a concept " t a l l " , for example, can be written in Fill. This uses the fact t h a t if y o u remove a little height from a tall m a n there is still a high support, b u t n o t a certain support, for him still being tall. Details can be f o u n d in Refs [1, 4]. V O T I N G M O D E L I N T E R P R E T A T I O N OF A F U Z Z Y SET A fuzzy subset f with respect to the set F is defined by means o f a membership function M f: F--~ [0, 1 ]. In other words an element, e, o f the set F belongs to the fuzzy subset f with a degree o f membership Mf(e). H o w can we interpret this membership level in more specific terms which will give some justification to its actual value a n d also its existence a n d use? One possible interpretation is in terms o f the voting behaviour o f a p o p u l a t i o n P, say, o f persons, all o f w h o m have their own understanding o f the meaning o f f We all use the term " t a l l " in relation to a person's height w i t h o u t having a precise understanding o f w h a t it means. W h e n we use the word in o r d i n a r y conversation, we assume others will be able to interpret it in more or loss the same way as ourselves. It is certainly true t h a t there is a set o f heights which everyone would accept as satisfying the concept o f "tall height" a n d there is a set o f heights which every one would accept as n o t satisfying this concept. The difficulty arises for the set o f heights in between these two sets. I f this intermediate set is null then we have an exact definition for " t a l l " b u t otherwise we do not. Each member o f this intermediate set can have a degree o f membership in the set o f heights representing " t a l l " , but how do we choose the actual degree? C o m p u t a t i o n a l m o d e l s 111 Consider a set F and l e t f b e a fuzzy subset o f this set. Let each person belonging to a population P vote on whether to accept or reject the membership o f a given element e o f set F as belonging t o f . Each person must accept or reject, agree or n o t agree to the elements membership. Abstentions a n d partial agreements are n o t allowed. M f ( e ) is equated to the p r o p o r t i o n o f persons who vote for accepting e as a m e m b e r o f f . Similarly (1 - Mf(e)) is the p r o p o r t i o n o f persons who reject e as belonging to f so t h a t this is the membership level for e n o t belonging to f . Each individual will have a threshold level such that if his d o u b t in an element e belonging to f increases above this level he will reject its membership, otherwise he will accept it. It is similar to a member o f a j u r y having to say guilty or innocent for the person on trial. The evidence presented at the trial m a y n o t be conclusive but the person must still m a k e a final judgement. I f people are alowed to abstain we return to an analogue with support pairs. Example Let F be the set o f positive integers {50 55 60 65 70 75} and f be defined by the following membership function Mf(55) = 0.2, Mf(60) = 0.5, Mf(65) = 0.8, Mf(70) = 1, Mf(50) = Mf(75) = 0. The fuzzy set f can therefore be represented as f = 5510.2 + 6010.5 + 6510.8 + 701 I. I f we take P as m a d e up o f 10 people then the voting pattern to give consistency with this definition could be Person 1 2 3 4 5 6 7 8 9 10 70 70 70 70 70 70 70 70 70 70 65 65 65 65 65 65 65 65 60 60 60 60 60 55 55 The integers given are those integers which the person accepted as satisfying f . Therefore person 1 accepts {70, 65, 60, 55} as satisfying f while person 7 only accepts {70, 65}. This interpretation assumes that persons who vote yes for 55 also vote yes for 60 and for 65. Similarly it assumes persons who vote yes for 60 vote yes for 65. The assumption that this interpretation uses is t h a t a person who votes for an element h in the set F with membership value M f ( h ) as belonging to f will also vote for a n y other element o f F satisfying f if it has a higher membership value t h a n Mf(h). We will call this the constant threshold model since it corresponds to each person having a threshold level for acceptance o f an element o f F i n f w h i c h does not vary with the element o f F chosen. A n alternative interpretation could be: Person 1 2 3 4 5 6 7 8 9 10 70 70 70 70 70 70 70 70 70 70 65 65 65 65 65 65 65 65 60 60 60 60 60 55 55 Other possible interpretations can be given but the one which intuitively seems more reasonable is the c o n s t a n t threshold model. I N T E R S E C T I O N A N D U N I O N O F F U Z Z Y SETS Consider two fuzzy sets f l , f 2 , with membership functions M f l , Mf2, b o t h defined as fuzzy subsets o f the set F. 112 J . F . BALDWIN Consider an element h o f F. The proportion o f persons o f population P who vote for h satisfying both the concepts defined by f l and f 2 is contained in the interval Iconj = [ m a x { M f l ( h ) + M f 2 ( h ) - 1,0}, min{Mfl(h), Mf2(h)}]. If we use the constant threshold model then one assumes that the threshold levels o f the persons P stay constant for judging different concepts. This means that if a person votes yes for one concept when having a certain degree of doubt, that person will also vote yes for another concept if faced with the same degree o f doubt. In this case we assume that those people who voted yes for the concept with the lower membership value will also have voted yes for the other. Thus for this assumption the membership value for element h for the intersection o f f l and f 2 will be min{Mfl(h), Mf2(h)}. Similarly for the union o f f l and f 2 the membership value for element h will lie in the Idisj = [max{Mfl(h), Mf2(h)}, min{Mfl(h) + Mf2(h), 1}]. With the same assumption as above, the minimum number of persons would vote yes for the union, so that the membership level for element h belonging to the union o f f l and f 2 is max{Mfl(h), Mf2(h)}. This is the assumption we make in Fril for fuzzy sets and is the usual definition for fuzzy conjunction and disjunction. More generally we can define a mapping T T: [0, 1].[0, 1]--,[0, 1] The mapping S: [0, 1]*[0, 1]--~[0, 1]. satisfies the same axioms as for T except that (l) is replaced by (l') where (1') S ( a O) = a. Examples o f instances o f S conorms corresponding to the T norms (1)-(3) above are respectively (1) S(a b ) = max{a b}, (2) S(a b ) = a + b - a . b , (3) S(a b) = m i n { a + b 1}. Assumptions can be made about the voting model to obtain each o f these answers. For example, if it is assumed that no preference can be made for any possible voting pattern for P in relation to f l or f 2 , then all possible distributions must be allowed. For each pai~ o f distributions the proportion o f those persons voting for both and the proportion o f those persons voting for at least one can be determined. This gives the values o f the conjunction and disjunction, respectively for this pair o f distributions. This is repeated for all possible pairs o f distributions and the values for the conjunction and disjunction determined in each case. If it is assumed that any pair o f which satisfies the axioms (1) T(a, 1) = a, (2) T(a, b) = T(b, a), (3) T ( a , b ) > t T ( e , d ) if a > i c and b > i d , (4) T(a, T(b, c)) = T(T(a, b), c). T is called a T-norm and generalizes the A N D corresponding to conjunction Examples o f instances o f T are (1) T(a, b) = rain{a, b}, (2) T(a, b) = a .b, (3) T(a, b) = max{a + b - 1, 0}. A dual norm, called the T-conorm, S, exists which generalizes disjunction. For any T-norm T there exists a dual norm S such that S(a, b) = 1 - T((1 - a), (1 - b)). C o m p u t a t i o n a l m o d e l s 113 distributions is as likely as a n y other, then the expected values for the conjunction a n d disjunction will be equal to M f l (h). Mf2(h) a n d M fl (h) + M f2(h) - M f l (h). Mf2(h), respectively. S E M A N T I C U N I F I C A T I O N Let f l , f 2 be two fuzzy subsets o f the set F and suppose that each o f these can be associated with some object X. Then we can ask the question, what is the probability o f " X i s f l " given that we k n o w that " X is f 2 " . Consider the more specific case in which we k n o w that J o h n is between 5 ft 10 in. and 6 ft. Then the probability t h a t J o h n is over 5 ft 10 in. is 1. The probability that J o h n is below 5 ft 9 in. is 0. The probability that J o h n is between 5 ft 9 in. a n d 5 ft 11 in. lies between 0 a n d 1. This is so since J o h n ' s actual height could belong to the interval [5 ft 10 in., 5 ft 11 in.] which would give a probability o f 1 o f J o h n being between 5 ft 9 in. a n d 5 ft 11 in., but it could also belong to [5 ft 11 in., 6 ft] which would give zero probability. The first can occur with a probability x~ and the second case with a probability x2. I f all t h a t we k n o w a b o u t x~ and x2 is that both are non-negative a n d they sum to one, then the probability o f J o h n being between 5 ft 9 in. a n d 5 ft 11 in. lies anywhere in the interval [0, 1]. I f we can estimate x~ and x~ then Pr(John is between 5 ft 9 in. and 5 ft 11 in.) = x~. I f an equally likely distribution is assumed over the interval [5ft 1 0 i n . , 6 f t ] then X l = 1 / 2 . This example illustrates the non-fuzzy version o f the situation posed above, with respect to the fuzzy subsets f l and f 2 . We can ask a similar question for the fuzzy case. W h a t is the probability that J o h n is tall given that we k n o w that J o h n is a little above average height? We should be able to arrive at an answer using a similar approach to that used for the non-fuzzy case but taking into account t h a t n o t every height has membership level o f 1 or 0 in the sets " t a l l " and " a little above average height". S E M A N T I C We will now return to the example above actual definitions for f l , f 2 a n d F. The Pr(X is f l [ X is f 2 ) can be interpreted X is f 2 ) . N o w N o w Therefore U N I F I C A T I O N A N D P O P U L A T I O N V O T I N G M O D E L o f determining Pr(X is f l I X is f 2 ) when we are given in this voting model as Pr(P accepts X is f l I P is told P r ( P accepts X is f l I P is told X is f 2 ) = S U M {Pr(P accepts X is f l I P accepts X is h, P is told X is f 2 ) . h Pr(P accepts X is h I P is told X is f2)} = S U M {Pr(P accepts X i s f l IP accepts X is h). h P r ( P accepts X is h I P is told X is f 2 ) } . Pr(P accepts X i s f l [P accepts X is h ) = Pr(h is accepted a s f l ) = M f l ( h ) . P r ( P accepts X is f l I P is told X is f 2 ) = S U M M f l ( h ) . P r ( P accepts X is h IP is told X is f 2 ) . h Pr(P accepts X is h I P is told X is f 2 ) = S U M Pr(person i chooses X is h IX is f 2 ) . i 1 1 4 J , F . BALDWIN I f person i is told t h a t X is f 2 , then this person has an interpretation for the label f 2 in the f o r m o f a set o f acceptable values. I f h is n o t one o f these values P r ( p e r s o n i chooses X is h l X is f 2 ) --- 0. I f the set o f values consists only o f h then P r ( p e r s o n i chooses X is h IX is f 2 ) = 1. I f the set contains m o r e than o n e value including h then probabilities m u s t be assigned to each value. All that is k n o w n a b o u t these probabilities is that they sum to 1. E x a m p l e Consider then P interprets f 2 as f l = 5510.2 + 6010.5 + 6510.8 + 7011, f 2 -- 5511 + 6010.2, F = {50, 55, 60, 65, 70, 75}, P f r s o u 1 2 3 4 5 6 7 8 9 10 55 55 55 55 55 55 55 55 55 55 60 60 so that P r ( P accepts X is 551P is told X is f 2 ) = 4/5 + x . 1/5, P r ( P accepts X is 601P is told X is f 2 ) = (1 - x ) . 1/5 following optimization p r o b l e m m a x / m i n z = 0 . 2 x I q- subject to xl + x 2 ~ < 1, x2 ~< 0.2, X 3 = 0 , X 4 ~--- 0 , X 1 " ~ ' X 2 " ~ X 3 " 3 t - X 4 ~ - 1, 0.5x2 + 0.8x3 + lx4, subject to for the c o n s t a n t threshold m o d e l used for interpreting f 2 . I f all possible distributions corresponding to interpretations o f f 2 are considered then the following optimization m o d e l results m a x / m i n z ffi 0.2xl + 0.5x2 + 0.8x3 + 1 X 4 , x l ~ < l , x2 ~< 0.2, X 3 "~- 0 , x 4 = 0 , xl + x2 + x3 + x4 = 1. where 0 ~< x ~ 1. This follows since persons 3 - 1 0 accept X is 55 as this is the only value they can choose. Persons 1 a n d 2 have a choice a n d x represents their p r o b a b i l i t y o f choosing 55. T h e r e f o r e P r ( P accepts X is f l I P is told X is f 2 ) = 0.2(0.8 + 0.2x) + 0 . 5 , 0 . 2 ( 1 - x); 0 ~ x ~< 1, so t h a t P r ( P accepts X is f l I P is told X is f 2 ) lies in the interval [0.2, 0.26]. W e thus conclude that P r ( X is f l IX is f 2 ) is in [0.2, 0.26]. This is equivalent to solving the Computational models 115 In both cases: min z gives lower b o u n d a n d m a x z gives upper b o u n d for Pr(X is f l IX is f 2 ) . In this example the support pair is the same whatever interpretation is used for f 2 . The next example will yield different results for different interpretations. A M O R E C O M P L E X E X A M P L E F = {el, e2, e3, e4, eSe6}, f l = e l l 0 . 1 + e 2 1 0 . 3 + e 3 1 0 . 5 + e 4 1 0 . 7 e 5 1 1 , f 2 = e l 10.2 + e2J 1 + e310.7 + e410.1, Pr(X is f l IX is f 2 ) lies in [z rain, z max], where z rain and z max are determined by solving one o f the following optimization models. (l) Using c o n s t a n t threshold model: m i n / m a x z = 0.1x~ + 0.3x2 + 0.5x3 + 0.7x4 + xs, subject to Therefore Thus x4 ~< 0.I, xl + x4 ~ 0.2, xj + x3 + x4 ~< 0.7, xl + x2 + x3 + x4 = 1. z min = 0 . 1 , 0 . 2 + 0.3*0.8 = 0.26, z max = 0 . 7 , 0 . 1 + 0 . 5 , 0 . 6 + 0 . 3 , 0 . 3 = 0.46. Pr(X is f l IX is f 2 ) lies in [0.26, 0.46]. (2) Allowing for all possible interpretations o f f 2 m i n / m a x z = 0.1x~ + 0.3x2 + 0.5x3 + 0.7x4 + xs, subject to X4 ~<0.1 xl ~< 0.2 x3 ~ 0.7 X2~<l X I " ~ - X 2 " ~ - X 3 " ~ - X 4 = 1. Therefore Thus z min = 0.1 *0.2 + 0.3*0.8 = 0.26, z m a x = 0 . 7 , 0 . 1 + 0 . 5 , 0 . 7 + 0 . 3 , 0 . 2 = 0.48. Pr(X is f l IX is f 2 ) lies in [0.26, 0.48] R E S T R I C T I O N M O D E L The generalization o f the linear p r o g r a m m i n g solutions for general fuzzy subsets f l a n d f 2 , defined on F, when X is f 2 is given a n d Pr(X is f l I X is f 2 ) is to be determined, is as 116 J.F. BALDWIN follows: Let F = { e , } ; i = l . . . n , f l = S U M {e, lMfl(e~)}, e i f 2 = S U M {e,[ M f 2 ( e i ) } . el (1) Using the c o n s t a n t threshold m o d e l for f 2 : max/min z = S U M Mfl(e~)'Xei , e i subject to S U M Xk <. Mf2(ei); all el, S U M x ~ , = 1 . {k: Mf2(k) ~< Mf'2(ei)} e i (2) Using all possible interpretations for f 2 : m a x / m i n z = S U M M f l (e~)- xei, e i subject to xe, ~< Mf2(ei); all ei, S U M xei = 1. e i I f z min, z max c o r r e s p o n d to the min z and m a x z, subject to the given constraints, respectively then P r ( X is f l I!X is if2) lies in [z min, z max]. This c o r r e s p o n d s to interpreting m e m b e r s h i p functions o f fuzzy sets as possibility restrictions. F o r s o m e element e~ in F, Mfl(e~) gives the u p p e r b o u n d to the possibility that ei belongs t o f l . M f l restricts the possibility that h belongs to f l to M f l (e~). I f we k n o w the p r o b a b i l i t y distribution over the elements o f F consistent with the statement that X is f 2 , i.e. we k n o w P r ( X is e~lX is f 2 ) for all e i in F, then P r ( X is f l l X is f 2 ) = S U M Mfl(ei)" P r ( X is e,[ X is f 2 ) . e i This uses a weighted sum o f the conditional probabilities where the weights are the m e m b e r s h i p values o f the fuzzy set f l . F o r the non-fuzzy case the characteristic values o f the characteristic function for f l w o u l d be used. In fact we d o n o t k n o w w h a t the values for {Pr(X is e i l X is f 2 ) } are. All we k n o w is that the value o f P r ( X is ei l X is f 2 ) will be constrained b y the fact that X is f 2 . W e will use one o f the following a s s u m p t i o n s for the f o r m that this constraint can take: (1) Using the c o n s t a n t threshold m o d e l for f 2 : P r ( X is one o f Fs I X is f 2 ) ~< max{Mf2(ei):e~ in F s } for any subset Fs o f F. (2) Using all possible interpretations for f 2 : P r ( X is e~lX is f 2 ) ~< Mf2(e~); all e~ in F. I f we p u t P r ( X is ei l X is f 2 ) = x,,, then this gives the constraints given a b o v e for each o f the models. There is n o w n o unique solution for P r ( X is f l IX is ./2) so that we can find u p p e r and lower b o u n d s for this by maximizing and minimizing S U M { M f l ( e ~ ) . x , , } subject to the constraints o n xe, given above. Computational models I 17 F u r t h e r m o r e , as given below, this can be generalized once again to handle continuous membership functions. S P E C I A L C A S E We will consider the special cases o f determining Pr(X is f i X is f ) . This often causes difficulty since the value o f this is n o t necessarily 1 as m a n y seem to expect. We will consider an example a n d then justify this answer by using the p o p u l a t i o n model for interpreting the results. Let F = {el, e2, e3, e4, es, e6, eT, es, eg, el0} a n d f = el 10.1 + e210.2 + e310.3 + e410.4 + e510.5 + e610.6 + e710.7 + e810.8 + e910.9 + e~011.0, then (1) Using the c o n s t a n t threshold model: m a x / m i n z = 0.1xl + 0.2x2 + 0.3x3 + 0.4x4 + 0.5x5 + 0.6x6 + 0.7x7 + 0.8x s + 0.9x 9 + xi0 , subject to so that X I ~<0.1, Xl + X2 ~< 0.2, Xl + X2 + X3 Xl +X2 +X3 + Xl + X2 + Xa + xl + x2 + x3 + X l + X 2 + X 3 + Xi Xi X1 0.3, X4 ~< 0.4, X4 + X5 ~< 0.5, X4 + Xs + X6 <~ 0.6, X4 + Xs + X6 + XT <~ 0.7 + X2 + X3 + X4 + X5 + X6 + X7 + Xs ~ 0.8, + X2 + X3 + X4 + X5 + X6 + XT + Xs + Xg ~ 0.9, + X2 + X3 + X4 + X5 + X6 + Xy + Xs + X9 + Xlo ~ 1, z min = 0.1,(0.1 + 0 . 2 + 0 . 3 + 0 . 4 + 0 . 5 + 0 . 6 + 0 . 7 + 0 . 8 + 0 . 9 + 1 ) = 0 . 1 , 5 . 5 = 0.55, z m a x = 1 . 1 = 1. Therefore using the c o n s t a n t threshold model Pr(X is f i X is f ) = [0.55, 1]. We might note t h a t i f we had chosen f to be f = et l0 + e210.1 + e310.21 + e410.3 + e~10.4 + e610.5 + e710.6 + esl0.7 + e910.8 + et0[0.9, then Pr(X is f i X is f ) would lie in [0.45, 1]. (2) Using all possible interpretations for f : m a x / m i n z = 0.1xl + 0.2x2 + 0.3x3 + 0.4x4 + 0.5x5 + 0.6x6 + 0.7x7 + 0.8x s + 0.9x9 + xl0, 118 subject to so t h a t J. F. BALDWIN xl ~<0.1, x2 ~< 0.2, x 3 ~ 0.3, x4 ~< 0.4, x~ ~<0.5, x6 ~< 0.6, x7 ~< 0.7, xs ~<0.8, x9 ~ 0.9, X~o ~< 1, z min = 0 . 1 , 0 . 1 + 0 . 2 , 0 . 2 + 0 . 3 , 0 . 3 + 0 . 4 , 0 . 4 = 0.3, z max = 1, Pr(X is f i X is f ) lies in [0.3, 1]. We might note t h a t if we h a d chosen f to be f = el 10 + e210.1 + e310.21 -{- e410.3 -t- e5t0.4 + e610.5 -q- e710.6 + esl0.7 + e910.8 -{- ej010.9, then Pr(X is f i X is f ) would lie in [0.4, 1]. It is quite easy to justify this result using the population voting model. I f the population P is told t h a t X is f then P will interpret this. There will be some elements o f the set F which some members o f P, but n o t all, will accept as possible values for X. W h e n asked if X is f t h e s e members o f P will choose these values with a certain probability as possible values for X, b u t n o t all members o f P will choose these values. Consider the situation t h a t the p o p u l a t i o n is told t h a t J o h n is tall. Some members o f the p o p u l a t i o n k n o w t h a t other members accept heights corresponding to tall which they do not. Therefore these members k n o w t h a t there is a probability t h a t J o h n ' s height will be a value which is unacceptable to them as representing tall. Therefore the probability t h a t J o h n is tall when the p o p u l a t i o n is told t h a t J o h n is tall c a n n o t be 1 since some members o f the p o p u l a t i o n will n o t vote yes for certain. C O N T I N U O U S C A S E The above is easily generalized to the c o n t i n u o u s case. C O N C L U S I O N S A theory of reasoning with uncertainties applicable to expert systems and other AI applications has been described. It is being applied to many applications and evidence to date indicates that it is relatively easy to apply. Fril is a powerful AI systems language and ideal for writing expert system shells. R E F E R E N C E S I. J. F. Baldwin, T. P. Martin and B. W. Pilsworth, Fril Manual. Equipu AIR Ltd, Bristol (1987). 2. J. F. Baldwin, Support logic programming. Proc. NATO Advanced Study Institute on Fuzzy Sets Theory and Applications. Louvain-La-N©uve, Belgium 0955). Computational models I 19 3. J. F. Baldwin, Support logic programming. In Fuzzy Sets Theory and Applications (Eds A. Jones and H. J. Zimmermann), pp. 133-170. Reidel, Dordrecht, Holland (1986). 4. J. F. Baldwin, Evidential support logic programming. Fuzzy Sets Systems 10, 1-26 (1987). 5. J. F. Baldwin, A theory of support pairs (in press). 6. I. Hacking, Logic o f Statistical Inference. Cambridge Univ. Press (1965). 7. L. A. Zadeh, Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets Systems 1, 3-28 (1978). 8. R. Bellman and M. Giertz, On the analytic formalism o f the theory of fuzzy sets. Inform. Sci. 5, 149-156 (1973). 9. C. G. Hempel, Maximum specificity and lawlikeness in probabilistic explanation. Phil. Sci. 35, 116-133 (1968). 10. P. Suppcs, Probabilistic Metaphysics. Blackwell, Oxford (1984). 11. W. C. Salmon, Scientific Explanation and the Causal Structure o f the World. Princeton Univ. Press, N.J. (1984). 12. D. V. Lindley, Scoring rules and the inevitability of probability. Int. Star. Rev. 50, 1-26 (1982). 13. G. Sharer, A Mathematical Theory o f Evidence. Princeton Univ. Press, N.J. (1976). 14. V. O. Homolka, The role of nonmonotonic reasoning in an intelligent maintenance system, ITRC73. Information Technology Research Centre, Bristol Univ. (1985). 
work_5q5kihz46jaevgqzflwcshzajm	----	This article was originally published in a journal published by Elsevier, and the attached copy is provided by Elsevier for the author’s benefit and for the benefit of the author’s institution, for non-commercial research and educational use including without limitation use in instruction at your institution, sending it to specific colleagues that you know, and providing a copy to your institution’s administrator. All other uses, reproduction and distribution, including without limitation commercial reprints, selling or licensing copies or access, or posting on open internet sites, your personal or institution’s website or repository, are prohibited. For exceptions, permission may be sought for such use through Elsevier’s permissions site at: http://www.elsevier.com/locate/permissionusematerial http://www.elsevier.com/locate/permissionusematerial A ut ho r's pe rs on al co py Educational data mining: A survey from 1995 to 2005 C. Romero *, S. Ventura Department of Computer Sciences, University of Cordoba, Cordoba, Spain Abstract Currently there is an increasing interest in data mining and educational systems, making educational data mining as a new growing research community. This paper surveys the application of data mining to traditional educational systems, particular web-based courses, well-known learning content management systems, and adaptive and intelligent web-based educational systems. Each of these systems has different data source and objectives for knowledge discovering. After preprocessing the available data in each case, data mining tech- niques can be applied: statistics and visualization; clustering, classification and outlier detection; association rule mining and pattern min- ing; and text mining. The success of the plentiful work needs much more specialized work in order for educational data mining to become a mature area. � 2006 Elsevier Ltd. All rights reserved. Keywords: Data mining; Educational systems; Web mining; Web-based educational systems 1. Introduction During the past decades, the most important innova- tions in educational systems are related to the introduction of new technologies (Ha, Bae, & Park, 2000) as web-based education. This is a form of computer-aided instruction virtually independent of a specific location and any specific hardware platform (Brusilovsky & Peylo, 2003). It has con- siderably gained in importance and thousands of web courses have been deployed in the past few years. But many of the current web-based courses are based on static learn- ing materials, which do not take into account the diversity of students. Adaptive and intelligent web-based educa- tional systems have been seen as a solution to individually richer learning environments. These systems try to offer learners personalized education by building a model of the individual’s goals, preferences, and knowledge. Data mining or knowledge discovery in databases (KDD) is the automatic extraction of implicit and interesting pat- terns from large data collections (Klosgen & Zytkow, 2002). KDD can be used not only to learn the model for the learning process (Hamalainen, Suhonen, Sutinen, & Toivonen, 2004) or student modeling (Tang & McCalla, 2002) but also to evaluate and to improve e-learning sys- tems (Zaı̈ane & Luo, 2001) by discovering useful learning information from learning portfolios (Hwang, Chang, & Chen, 2004). In conventional teaching environments, educators are able to obtain feedback on student learning experiences in face-to-face interactions with students, enabling a con- tinual evaluation of their teaching programs (Sheard, Ced- dia, Hurst, & Tuovinen, 2003). Decision making of classroom processes involves observing a student’s behav- ior, analyzing historical data, and estimating the effective- ness of pedagogical strategies. However, when students work in electronic environments, this informal monitoring is not possible; educators must look for other ways to attain this information. Organizations, which run distance education sites, collect large volumes of data, automatically generated by web servers and collected in server access logs. Web-based learning environments are able to record most learning behaviors of the students, and are hence able to provide a huge amount of learning profile. Recently, there is a growing interest in the automatic analysis of lear- ner interaction data with web-based learning environments 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.04.005 * Corresponding author. Tel.: +34 957 212172; fax: +34 957 218630. E-mail address: cromero@uco.es (C. Romero). www.elsevier.com/locate/eswa Expert Systems with Applications 33 (2007) 135–146 Expert Systems with Applications A ut ho r's pe rs on al co py (Muehlenbrock, 2005). In order to provide a more effective learning environment, data mining techniques can be applied (Ingram, 1999). Data mining is a step in the overall process of KDD that consists of preprocessing, data min- ing and postprocessing. Data mining has already been suc- cessfully applied in e-commerce (Srivastava, Cooley, Deshpande, & Tan, 2000), and it has begun to be used in e-learning with promising results. Although the discovery methods used in both areas (e-commerce and e-learning) are similar (Hanna, 2004), there are some important differ- ences between them: • Domain. The e-commerce purpose is to guide clients in purchasing while the e-learning purpose is to guide stu- dents in learning (Romero, Ventura, & Bra, 2004). • Data. In e-commerce the used data are normally simple web server access log, but in e-learning there is more information about a student’s interaction (Pahl & Don- nellan, 2003). The user model is also different in both systems. • Objective. The objective of data mining in e-commerce is increasing profit, that is tangible and can be measured in terms of amounts of money, number of customers and customer loyalty. And the objective of data mining in e-learning is to improving the learning. This goal is more subjective and more subtle to measure. • Techniques. Educational systems have special character- istics that require a different treatment of the mining problem. As a consequence, some specific data mining techniques are needed to address in particular the pro- cess of learning (Li & Zaı̈ane, 2004; Pahl & Donnellan, 2003). Some traditional techniques can be adapted, some cannot. The application of knowledge extraction techniques to educational systems in order to improve learning can be viewed as a formative evaluation technique. Formative evaluation (Arruabarrena, Pérez, López-Cuadrado, & Vadillo, 2002) is the evaluation of an educational program while it is still in development, and with the purpose of con- tinually improving the program. Examining how students use the system is one way to evaluate the instructional design in a formative manner and it may help the educator to improve the instructional materials (Ingram, 1999). Data mining techniques can discover useful information that can be used in formative evaluation to assist educators establish a pedagogical basis for decisions when designing or modifying an environment or teaching approach. The application of data mining in educational systems is an iter- ative cycle of hypothesis formation, testing, and refinement (see Fig. 1). Mined knowledge should enter the loop of the system and guide, facilitate and enhance learning as a whole. Not only turning data into knowledge, but also fil- tering mined knowledge for decision making. As we can see in Fig. 1, educators and academics respon- sible are in charge of designing, planning, building and maintaining the educational systems. Students use and interact with them. Starting from all the available informa- tion about courses, students, usage and interaction, different data mining techniques can be applied in order to discover useful knowledge that helps to improve the e-learning pro- cess. The discovered knowledge can be used not only by pro- viders (educators) but also by own users (students). So, the application of data mining in educational systems can be oriented to different actors with each particular point of view (Zorrilla, Menasalvas, Marin, Mora, & Segovia, 2005): • Oriented towards students (Heraud, France, & Mille, 2004; Farzan, 2004; Lu, 2004; Tang & McCalla, 2005; Zaı̈ane, 2002). The objective is to recommend to learners activities, resources and learning tasks that would favour and improve their learning, suggest good learn- ing experiences for the students, suggest path pruning and shortening or simply links to follow, based on the tasks already done by the learner and their successes, and on tasks made by other similar learners, etc. Educational Systems (traditional classrooms, e-learning systems, adaptive and intelligent web-based educational systems) Educators Students Data Mining (clustering, classification, outlier, association, pattern matching, text mining) Academics Responsible Students usage and interaction data, course information, academic data, etc. To show recommendations To show discovered knowledge To design, plan, build and maintenance To use, interact, participe and communicate Fig. 1. The cycle of applying data mining in educational systems. 136 C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 A ut ho r's pe rs on al co py • Oriented towards educators (Ha et al., 2000; Hamalainen et al., 2004; Merceron & Yacef, 2004; Minaei-Bidgoli & Punch, 2003; Mor & Minguillon, 2004; Muehlenbrock, 2005; Pahl & Donnellan, 2003; Romero et al., 2004; Silva & Vieira, 2002; Talavera & Gaudioso, 2004; Tang et al., 2000; Ueno, 2004b; Zaı̈ane & Luo, 2001). The objective is to get more objective feedback for instruc- tion, evaluate the structure of the course content and its effectiveness on the learning process, classify learners into groups based on their needs in guidance and mon- itoring, find learning learner’s regular as well as irregular patterns, find the most frequently made mistakes, find activities that are more effective, discover information to improve the adaptation and customization of the courses, restructure sites to better personalize course- ware, organize the contents efficiently to the progress of the learner and adaptively constructing instructional plans, etc. • Oriented towards academics responsible and administra- tors (Becker, Ghedini, & Terra, 2000; Grob, Bensberg, & Kaderali, 2004; Luan, 2002; Ma, Liu, Wong, Yu, & Lee, 2000; Peled & Rashty, 1999; Sanjeev & Zytkow, 1995; Urbancic, Skrjanc, & Flach, 2002). The objective is to have parameters about how to improve site effi- ciency and adapt it to the behavior of their users (opti- mal server size, network traffic distribution, etc.), have measures about how to better organize institutional resources (human and material) and their educational offer, enhance educational programs offer and determine effectiveness of the new computer mediated distance learning approach. There are many general data mining tools that provide mining algorithms, filtering and visualization techniques. Some examples of commercial and academic tool are DBMiner, Clementine, Intelligent Miner, Weka, etc. (Klos- gen & Zytkow, 2002). However these tools are not specifi- cally designed and maintained for pedagogical purposes and it is cumbersome for an educator who does not have an extensive knowledge in data mining to use these tools (Zaı̈ane, Xin, & Han, 1998). In order to solve this problem, some specific educational data mining, statistical and visu- alization tools have been developed to help educators in analyzing the different aspects of the learning process (see Table 1). We have divided this paper into the following sections. We first review some different types of educational systems and how data mining can be applied in each of them. We then describe the data mining techniques that have been applied in educational systems grouping them by task. Finally, we summarize the main conclusions and we draw some future research. 2. Educational systems: data and objectives Data mining can be applied to data coming from two types of educational systems: traditional classroom and distance education. It is necessary to deal separately with the application of data mining techniques in each type due to the fact that they have different data sources and objectives. 2.1. Traditional classrooms Traditional classroom environments are the most widely used educational systems. It is based on face-to-face con- tact between educators and students organized through lec- turers. There are a lot of different subtypes: private and public education, elementary and primary education, adult education, higher, tertiary and academic education, special education, etc. They have been criticized because they encourage passive learning, ignore individual differences and needs of the learners, and do not pay attention to problem solving, critical thinking, or other higher order thinking skills (Johnson, Arago, Shaik, & Palma-Rivas, 2000). In conventional classrooms, educators attempt to enhance instructions by monitoring student’s learning pro- cesses and analyzing their performances by paper records and observation. They can also use information about stu- dent attendance, course information, curriculum goals, and individualized plan data. And educational institution has many diverse and varied sources of information (Ma et al., 2000): traditional databases (with a student’s infor- mation, educator’s information, class and schedule infor- mation, etc.), online information (online web pages and course content pages), multimedia databases, etc. Data mining can help each actor of the learning process. Institutions would like to know which students will enroll in a particular course and which students will need assis- tance in order to graduate. An administrator may wish to find out information such as the admission requirements and to predict the class enrollment size for timetabling. Students may wish to know how best to select courses Table 1 Some specific educational data mining, statistics and visualization tools Tool name Authors Mining task Mining tool Za€ıane and Luo (2001) Association and patterns MultiStar Silva and Vieira (2002) Association and classification Data Analysis Center Shen et al. (2002) Association and classification EPRules Romero et al. (2003) Association KAON Tane et al. (2004) Text mining and clustering TADA-ED Merceron and Yacef (2005) Classification and association O3R Becker et al. (2005) Sequential patterns Synergo/ColAT Avouris et al. (2005) Statistics and visualization GISMO/CourseVis Mazza and Milani (2005) Visualization Listen tool Mostow et al. (2005) Visualization TAFPA Damez et al. (2005) Classification iPDF_Analyzer Bari and Benzater (2005) Text mining C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 137 A ut ho r's pe rs on al co py based on prediction of how well they will perform in the courses selected. Instructors may wish to know what learn- ing experiences are most contributive to overall learning outcomes, why is one class outperforming the other, similar groups of students, etc. There are some works about the application of data mining in traditional education. One of the first articles about the use of data mining in educa- tion to understand the student enrollment was written by Sanjeev and Zytkow (1995). They apply knowledge discov- ery in the form of statements ‘‘Pattern P holds for data in Range R’’ to university databases. The results were pre- sented to a senior university administrator in order to make strategic decisions about the institutional policies. Another work on the use of KDD to identity and understand whether curriculum revisions can affect students in a Bra- zilian university was done by Becker et al. (2000). They ver- ify the qualitative impact of revisions and evaluate it using a number of techniques, such as summarization, associa- tion, classification. In a related work, the objective is to select the weak students to attend remedial classes (Ma et al., 2000). They use a scoring function that is based on association rules. First, they identify the potential weak students and then select the course that each weak student is recommended to take. Finally, an application in higher education for doing a comprehensive analysis of student characteristics is done by Luan (2002). He proposes to use different unsupervised (Kohonen nets) and supervised (C5.0, genetic algorithms, etc.) data mining algorithms to do clustering and prediction in order to enable educational institutions to better allocate resources and staff, proac- tively manage student outcomes, and improve the effective- ness of alumni development. 2.2. Distance education Distance education or distance learning consists of techniques and methods providing access to educational programs for students who are separated by time and space from lecturers. e-Learning systems lack a closer student–educator relationship (one to one). There are dif- ferent subtypes of distance education: paper-based corre- spondence education, videotape education, computer- aided education (multimedia education, internet education or web-based education), etc. Currently, the most used is web-based education allowing students to conveniently learn via the Internet. Web-based education is a form of distance education delivered over the Internet (Johnson et al., 2000). Today, there are a lot of terms used to refer to web-based education such as e-learning, e-training, online instruction, web-based learning, web-based training, web-based instruction, etc. And there are different types of web-based systems: synchronous and asynchronous, col- laborative and non-collaborative, closed corpus and open corpus, etc. These web-based education systems can nor- mally record the student’s accesses in web logs that provide a raw trace of the learners’ navigation on the site. There are several types of logs (Srivastava et al., 2000): • Server log file. This constitutes the most widely used data source for performing data mining, containing just the bare details of timing, path, and input-response. There are a variety of formats, such as common log for- mat (CLF), extended log format (ELF), etc. (Koutri, Avouris, & Daskalaki, 2004). Normally, there is a single log file for all students. • Client log file. This consists of a set of log files, one per student, and contains information about the interaction of the user with the system. Can be implemented by a remote agent (such Javascripts, Java Applets), modify- ing the source code of an existing browser, or using cookies. • Proxy log file. This consists of a set of log files of caching between client browsers and web servers. This informa- tion complements server log file information. It should be noted that the concept of logging may include restrictions by law. Therefore, whenever a log sys- tem authenticates users it should not relate to a person’s true identity but primarily they as individual persons (Rahkila & Karjalainen, 1999). Log files also have several inherent limitations, tracking for files not users, simple clicks and not learning activities, not capturing contextual information, recognizing specific computers not specific people, having incomplete and incorrect information prob- lem, and some technical aspects of web browser (as the cache) may prevent to record logs. To address these prob- lems, authors have proposed several solutions. Yu, Own, and Lin (2001) propose to use another way to record a lear- ner’s portfolio that includes the learning path, preferred learning course, grade of course, and learning time, etc. Li and Zaı̈ane (2004) use more information channels to model user navigational behavior: web access logs, the structure of a visited web site, and the content of visited web pages. Avouris, Komis, Fiotakis, Margaritis, and Voyiatzaki (2005) expand automatically generated log files by introduc- ing contextual information as additional events and by associating comments and static files. Monk (2005) com- bines data on the activity with content and user profiles in a composite information model. Ingram (1999) combines data with other inquiry methods, such as informal chatting with students, in class shows of hands, surveys, or written feedback about the web site. Iksal and Choquet (2005) pro- pose to use a specific usage tracking meta-language to describe the track semantics recorded by web-based educa- tional systems. Markham et al. (2003) propose to use soft- ware agents to extract data from the e-learning environment and to organize them in intelligent ways. Next, we distinguish between three different types of web- based education systems: particular web-based courses, well-known learning content management systems, and adaptive and intelligent web-based educational systems. 2.2.1. Particular web-based courses Particular web-based courses are specific courseware that use standard HTML (HyperText Markup Language). 138 C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 A ut ho r's pe rs on al co py There are a lot of courses, tutorials, etc. of this type on the Internet, and as another web site, they have the same kinds of data sources (Srivastava et al., 2000): • Content: The real data in the web pages, i.e. the data the web page were designed to convey to the users. This usu- ally consists of texts, graphics, videos, sounds, etc. • Structure: Data which describe the organization of the content. Intra-page structure information includes the arrangement of various HTML or XML tags within a given page. This can be represented as a tree structure, where the HTML tag becomes the root of the tree. The principal kind of inter-page structure information is hyper-links connecting one page to another. • Usage: Data that describe the pattern of usage of web pages. There are two main types of students’ informa- tion (Silva & Vieira, 2002): information about the stu- dent’s actions and communications, and information about the student’s activities in the course. • User profile: Data that provide demographic informa- tion about users of the web site. This includes registra- tion data and customer profile information. Data mining can be used to know how students use the course, how a pedagogical strategy impacts different types of students, in which order the students study subtopics, what are the pages/topics that students skip, how much time the students spend with a single page, a chapter or the full course, etc. 2.2.2. Well-known learning content management systems Well-known learning content management systems (LCMS) are platforms that offer a great variety of channels and workspaces to facilitate information sharing and com- munication between participants in a course, let educators distribute information to students, produce content mate- rial, prepare assignments and tests, engage in discussions, manage distance classes and enable collaborative learning with forums, chats, file storage areas, news services, etc. Some examples of commercial LCMS are Blackboard, Vir- tual-U, WebCT, TopClass, etc. and some example of free LCMS are Moodle, Ilias, Claroline, aTutor, etc. (Paulsen, 2003). These systems accumulate large log data of the stu- dents’ activities and usually have built-in student monitor- ing features (Mazza & Milani, 2005). They can record whatever student activities it involves, such as reading, writ- ing, taking tests, performing various tasks in real or virtual environments, even communicating with peers (Mostow, 2004). They normally also provide a database that stores all the systems information: personal information of the users (profile), academic results, user’s interaction data, etc. Although some platforms offer reporting tools, when there is a great number of students, it becomes hard for a tutor to extract useful information. Data mining can be applied to explore, visualize and analyze data in order to identify useful patterns (Talavera & Gaudioso, 2004) and to evaluate web activity to get more objective feedback for your instruction and knowing more about how the stu- dents learning on the LCMS (Zaı̈ane & Luo, 2001). 2.2.3. Adaptive and intelligent web-based educational systems Adaptive and intelligent web-based educational systems (AIWBES) provide an alternative to the traditional just- put-it-on-the-web approach in the development of web- based educational courseware (Brusilovsky & Peylo, 2003). AIWBES attempt to be more adaptive by building a model of the goals, preferences and knowledge of each individual student and using this model throughout the interaction with the student in order to adapt to the needs of that student. AIWES are the result of a joint evolution of intelligent tutoring systems (ITS) and adaptive hyperme- dia systems (AHS). Some examples of ITS are SQL-Tutor, German Tutor, ActiveMath, VC-Prolog-Tutor, and some examples of AHS are AHA!, InterBook, KBS-Hyperbook, WebCOBALT (Brusilovsky & Peylo, 2003). The data from AIWEBS are semantically richer and can lead to more diag- nostic analysis than data from traditional web-based educa- tion system (Merceron & Yacef, 2004). The available data come from the domain model (which may be structured into an ontology), pedagogical dataset (set of problems, their answer and complexity information), interaction log files (data related with user interaction) and student model (list of the satisfactions and violations constraints). AIWBES use a standard student model (used internally by the tutor- ing system), but, for the purpose of data mining, it is neces- sary to have a new model of student interaction with augmented information with contextual data. These stu- dent’s interaction can be analyzed at a number of different layers of granularity: course, sessions, problems, attempts and constraints (Nilakant & Mitrovic, 2005). Data mining can be used in order to know the causes of problems in the system, for example, incorrect feedback statements (Nilakant & Mitrovic, 2005), to adapt the level to the progress of the learner (Romero et al., 2004), to sug- gest personalized learning experiences and activities for the students (Tang & McCalla, 2005). 3. Data preprocessing Data preprocessing allows to transform the original data into a suitable shape to be used by a particular mining algorithm. So, before applying the data mining algorithm, a number of general data preprocessing tasks have to be addressed (Koutri et al., 2004; Zorrilla et al., 2005): • Data cleaning. It is one of the major preprocessing tasks, to remove irrelevant items and log entries that are not needed for the mining process such us graphics, scripts. • User identification. Process of associating page refer- ences to the connected user. • Session identification. It takes all of the page references for a given user and course in a log and breaks them up into user sessions. In our particular case, we have C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 139 A ut ho r's pe rs on al co py initially considered a new session when a change in a user course happens or when the time interval between two successive inter-transaction clicks ups 30 min (Zorrilla et al., 2005). • Path completion. It fills in page references that are miss- ing due to browser and proxy server caching. • Transaction identification. It breaks down sessions into smaller units, referred to as transactions or episodes. • Data transformation and enrichment. It consists of calcu- lating new attributes from the existing ones, conversing of numerical attributes into nominal attributes, provid- ing meaning to references contained in the log, etc. • Data integration. It is the integration and synchroniza- tion of data from heterogeneous sources. • Data reduction. It is for reducing data dimensionality. Additionally, data preprocessing of web-based educa- tional systems has some specific issues: • Most of the systems use user authentication (password protection) in which logs have entries identified by users since the users have to log-in, and sessions are already identified since users may also have to log-out (Rahkila & Karjalainen, 1999). • Most of the systems record the students’ interactions not only in log files but also directly in databases. If this is not the case, during the preparation process, data for each individual student (profile, logs, etc.) can be aggre- gated to a database (Talavera & Gaudioso, 2004). Dat- abases are more powerful than typical log text files and provide an analysis easier, more flexible and less bug prone. • Data transformation is more oriented to a better inter- pretation of data. Numerical values are discretized or transformed into ranges that provide a much more com- prehensible view of the data. New attributes result from other current attributes in a specific attribute derivation. The derivation performs some kind of aggregation, for example, each attempt is grouped into a new number of attempt attribute (Nilakant & Mitrovic, 2005). • In the division of individual visit sessions into transac- tions, can be identified subsessions or missions with coherent information needs in which the identified sequence is based on the real content of pages (Li & Zaı̈ane, 2004). Besides different meanings of interaction at different levels of abstraction can be distinguished (Pahl & Donnellan, 2003): learning and training interac- tion, human–computer interface and multimedia and service interaction. • The data filtration uses specific educational concepts as number of attempts, number of repeated reading, level of knowledge, etc. Normally, data is filtered by defining some condition on one or more attributes and removing the instances that violate it (Nilakant & Mitrovic, 2005). The educators have to actively participate in the previ- ous preprocessing task, for example, indicating specific data filtration and attribute derivation or transformation, etc. So, it is needed to enhance preprocessing facilities that prepare the e-learning data in a meaningful and useful way. 4. Data mining techniques in educational systems Data mining is a multidisciplinary area in which several computing paradigms converge: decision tree construction, rule induction, artificial neural networks, instance-based learning, Bayesian learning, logic programming, statistical algorithms, etc. (Klosgen & Zytkow, 2002). Next, we are going to describe some specific application of data mining techniques grouped by tasks, in web-based educational sys- tems (see Table 2). 4.1. Statistics and visualization Student’s usage statistics are often the starting point of evaluations of an e-learning system, although they are usu- ally not considered as data mining techniques (Tsantis & Castellani, 2001). Formal statistical inference is assumption driven in the sense that a hypothesis is formed and then tested against the data. Data mining, in contrast, is discov- ery driven in the sense that the hypothesis is automatically extracted from the data. Usage statistics may be extracted using standard tools designed to analyze web server logs as AccessWatch, Ana- log, Gwstat, WebStat, etc. (Zaı̈ane et al., 1998). But there are some specific statistical tools in educational data as Synergo/ColAT (Avouris et al., 2005). Some example of usage statistics are simple measures such as the total num- ber of visits and number of visits per page (Pahl & Donne- llan, 2003). Some other general statistics show the connected learner distribution over time, the most frequent acceded courses, how learners establish many learning ses- sions over time (Zorrilla et al., 2005). Besides, some specific statistical in AIWBES can show the average number of constraint violations, the average problem complexity, the total time spent in attempts (Nilakant & Mitrovic, 2005). More complex statistical tests of procedures such as regression analysis, correlation analysis, multivariate statistical methods. (Zarzo, 2003) need to use a more pow- erful statistical tools as SPSS, SAS, S, R, Statistica, etc. (Klosgen & Zytkow, 2002). If data are stored in a relational database, then SQL queries (Heiner, Beck, & Mostow, 2004; Merceron & Yacef, 2003) can provide functionality for a number of simple statistical operations such as stan- dard deviation, mode, sample size, etc. (Nilakant & Mitro- vic, 2005). But the information obtained from usage statistics is not always easy to interpret to the educators and then other techniques have to be used. Information visualization techniques can be used to graphically render complex, multidimensional student tracking data collected by web-based educational systems (Mazza & Milani, 2005). These techniques facilitate to analyze large amounts of information by representing the data in some visual display. Normally large quantities of 140 C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 A ut ho r's pe rs on al co py raw instance data are represented or plotted as spreadsheet charts, scatterplot, 3D representations, etc. The informa- tion visualized in statistical graphs can be about assignment complement, admitted question, exam score, etc. (Shen, Yang, & Han, 2002). Visualization techniques have been used to visualize social aspects in computer-supported collaborative learning, community relationships in peer- to-peer systems, and conversations in online groups. Instructors can manipulate the graphical representations generated, which allow them to gain an understanding of their learners and become aware of what is happening in distance classes. There are some specific visualization tools in educational data as GISMO/CourseVis (Mazza & Milani, 2005) and Listen tool (Mostow et al., 2005). 4.2. Web mining Web mining (Srivastava et al., 2000) is the application of data mining techniques to extract knowledge from web data. There are three main web mining categories from the used data viewpoint: web content mining is the process of extracting useful information from the contents of web documents; web structure mining is the process of discover- ing structure information from the web; and web usage mining (WUM) that is the discovering of meaningful pat- terns from data generated by client–server transactions on one or more web localities. But there are two types of web mining categories from the used system viewpoint (Li & Zaı̈ane, 2004): offline web mining, that is used to dis- cover patterns and other useful information to help educa- tors to validate the learning models and restructure the web site; and online or integrated web mining in which the pat- terns automatically discovered are fed into an intelligent software system or agent that could assist learners in their online learning endeavours. The mined patterns are used on-the-fly by the system to improve the application or its functions. There are different web mining techniques applied to educational systems, but almost all of them can be grouped in one of the three next ones: clustering, classification and outlier detection; association rule mining and sequential pattern mining; and text mining. 4.2.1. Clustering, classification and outlier detection Clustering is a process of grouping physical or abstract objects into classes of similar objects. Clustering and clas- sification (Klosgen & Zytkow, 2002) are both classification methods. Clustering is an unsupervised classification and classification is a supervised classification. Classification and prediction are also related techniques. Classification predicts class labels, whereas prediction predicts continu- ous-valued functions. On the other hand, an outlier is an observation (or measurement) that is unusually large or small relative to the other values in a dataset. Outliers typ- ically are attributable to one of the following causes: the measurement is observed, recorded, or entered into the computer incorrectly; the measurements come from a dif- ferent population; the measurement is correct, but repre- sents a rare event. Table 2 Works about applying data mining techniques in educational systems Authors Mining task Educational system Sanjeev and Zytkow (1995) Sequence pattern Traditional education Za€ıane et al. (1998) Statistic and sequence pattern LCM systems Beck and Woolf (2000) Prediction AIWBE system Becker et al. (2000) Association and classification Traditional education Chen et al. (2000) Classification Web-based course Ha et al. (2000) Association Web-based course Ma et al. (2000) Association Traditional education Tang et al. (2000) Text mining AIWBE system Yu et al. (2001) Association Web-based course Za€ıane and Luo (2001) Sequence pattern LCM system Luan (2002) Clustering and prediction Traditional education Pahl and Donnellan (2003) Sequence pattern and statistics LCM system Shen et al. (2002) Visualization LCM system Wang (2002) Association and sequence pattern Web-based course Merceron and Yacef (2003) Statistic AIWBE system Minaei-Bidgoli and Punch (2003) Classification Web-based course Shen et al. (2003) Sequence pattern and clustering Web-based course Zarzo (2003) Statistic Web-based course Arroyo et al. (2004) Prediction AIWBE system Baker et al. (2004) Classification AIWBE system Chen et al. (2004) Text mining Web-based course Freyberger et al. (2004) Association AIWBE system Hamalainen et al. (2004) Classification AIWBE system Heiner et al. (2004) Statistic AIWBE system Lu (2004) Association AIWBE system Merceron and Yacef (2004) Association AIWBE system Minaei-Bidgoli et al. (2004) Association Web-based course Mor and Minguillon (2004) Clustering LCM system Romero et al. (2004) Association AIWBE system Talavera and Gaudioso (2004) Clustering LCM system Ueno (2004b) Outlier detection Web-based course Ueno (2004a) Text mining Web-based course Wang et al. (2004) Sequence pattern and clustering LCM system Li and Za€ıane (2004) Association LCM system Avouris et al. (2005) Statistic Web-based course Castro et al. (2005) Outlier detection LCM system Dringus and Ellis (2005) Text mining LCM system Feng et al. (2005) Prediction AIWBE system Hammouda and Kamel (2005) Text mining Web-based course Markellou et al. (2005) Association Web-based course Mazza and Milani (2005) Visualization LCM system Mostow et al. (2005) Visualization AIWBE system Muehlenbrock (2005) Outlier detection AIWBE system Nilakant and Mitrovic (2005) Statistic AIWBE system Tang and McCalla (2005) Clustering AIWBE system Zorrilla et al. (2005) Statistic LCM system Damez et al. (2005) Classification AIWBE system Bari and Benzater (2005) Text mining LCM system C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 141 A ut ho r's pe rs on al co py All these methods have been applied to web-based edu- cational systems. Clustering can group together a set of pages with similar contents, users with similar navigation behavior or navigation sessions. Classification allows char- acterizing the properties of a group of user profiles, similar pages or learning sessions. And outlier detection can detect students with learning problems. Next, we describe some works about the application of these techniques in different types of web-based educational systems: • Particular web-based courses. Chen, Liu, Ou, and Liu (2000) apply decision tree (C5.0 algorithm) and data cube technology from web log portfolios for managing classroom processes. The induction analysis discovers potential student groups that have similar characteristics and reaction to a particular pedagogical strategy. Min- aei-Bidgoli and Punch (2003) classify students based on features extracted from the logged data in order to predict their final grades. They use genetic algorithms to optimize a combination of multiple classifiers by weighing feature vectors. Ueno (2004b) proposes a method of online outlier detection of learners’ irregular learning processes by using the learners’ response time data for the e-learning contents. The outlier detection method uses a Bayesian predictive distribution and it assists a two way instruction by using mining results for the learners’ learning processes. • Well-known learning content management systems. Tala- vera and Gaudioso (2004) propose mining student data using clustering to discover patterns reflecting user behaviors. They propose models for collaboration man- agement to characterize similar behavior groups in unstructured collaboration spaces. Mor and Minguillon (2004) extend the sequencing capabilities of the SCORM standard to include the concept of recommended itiner- ary, by combining educators expertise with learned experience acquired by system usage analysis. They use clustering algorithms for grouping students. Castro, Vel- lido, Nebot, and Minguillon (2005) detect atypical behavior on the grouping structure of the users of a vir- tual campus. They propose to use a generative topo- graphic mapping model and a clustering model to characterize groups of online students. The model neu- tralizes the negative impact of outliers on the data clus- tering process. • Adaptive and intelligent web-based educational systems. Tang et al. (2000) use data clustering for web learning to promote group-based collaborative learning and to provide incremental learner diagnosis. They find clusters of students with similar learning characteristics based on the sequence and the contents of the pages they visited. Currently, they are working in smart recommendation for evolving e-learning systems (Tang & McCalla, 2005) using clustering and collaborative filtering. This is a paper recommender system that can personalize and adapt the course content based on the system’s observation of the learners and the accumulated ratings given by the learners. Hamalainen et al. (2004) introduce a hybrid model, which combines both data mining and machine learning techniques in constructing a Bayesian network to describe the student’s learning process. The goal is to classify students to give them differentiated guiding according to their skills and other characteris- tics. Beck and Woolf (2000) construct a learning agent for high-level student modeling with machine learning in intelligent tutoring systems. The agent learns to pre- dict the probability the student’s next response will be correct, and how long it will take the student to generate that response. They use linear regression to predict observable variables. Arroyo, Murray, Woolf, and Beal (2004) infer unobservable learning variables from stu- dents ITS log files. They star from a correlation analysis between variables and construct a Bayesian network that infers students’ attitudes (negative and positive) and predictions of the system. They use a maximum likelihood method to learn conditional probabilities from students’ data. Baker, Corbett, and Koedinger (2004) use machine-learned latent response model to detect student misuse of intelligent tutoring systems. They build a classifier to identify if a student is gaming the system in a way that leads to poor learning and in need of an intervention. Feng, Heffernan, and Koe- dinger (2005) look for sources of error in predicting a student’s knowledge. They perform a stepwise regres- sion to predict what metrics help to explain poor predic- tion of state exam scores. Muehlenbrock (2005) detects regularities and deviations in the learner’s or educator’s actions among others, in order to provide educators and learners with additional information to manage their learning and teaching. Damez, Marsala, Dang, and Bouchon-Meunier (2005) use a fuzzy decision tree for user modeling and discriminating a novice from an experimented user automatically. They use an agent to learn the cognitive characteristics of an user’s interac- tions and classify users as experimented or not. 4.2.2. Association rule mining and sequential pattern mining Association rule mining is one of the most well studied mining methods. Such rules associate one or more attri- butes of a dataset with another attribute, producing an if–then statement concerning attribute values. Mining asso- ciation rules between sets of items in large databases was first stated by Agrawal, Imielinski, and Swami (1993) and it opened a brand new family of algorithms. The original problem was the market basket analysis that tried to find all the interesting relations between the bought products. Sequential pattern mining (Agrawal & Srikant, 1995) attempts to find inter-session patterns such as the presence of a set of items followed by another item in a time-ordered set of sessions or episodes. These methods have been applied to web-based educa- tion systems. Associations could reveal which contents stu- dents tend to access together, or which combination of tools 142 C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 A ut ho r's pe rs on al co py they use. Sequential patterns can reveal which content has motivated the access to other contents, or how tools and contents are entwined in the learning process. Next, we describe some works about the application of these tech- niques in different types of web-based educational systems: • Particular web-based courses. Ha et al. (2000) perform web page traversal path analysis for customized educa- tion, and web page associations for virtual knowledge structures, which can be formed by learners themselves as they navigate web pages. Yu et al. (2001) find incor- rect student behavior. They modify traditional web logs, and apply fuzzy association rules to find out the rela- tionships between each pattern of a learner’s behavior; including the time spent online, number of read and published articles, number of asked questions, etc. Wang (2002) develops a portfolio analysis tool based on data mining techniques. He uses associative material clusters and sequences among them. This knowledge allows educators to study the dynamic browsing struc- ture and to identify interesting or unexpected learning patterns. To do this, he discovers two types of relations: association relations and sequence relations between documents. Shen, Han, Yang, Yang, and Huang (2003) use data mining and case-based reasoning for distance learning. They use clustering to classify students based on their learning actions and they find sequential associ- ation rules of different knowledge points. Minaei- Bidgoli, Tan, and Punch (2004) propose mining interesting contrast rules for web-based education systems. Con- trast rules help one to identify attributes characterizing patterns of performance disparity between various groups of students. Markellou, Mousourouli, Spiros, and Tsakalidis (2005) propose an ontology-based frame- work and discover association rules, using the a priori algorithm. The role of the ontology is to determine which learning materials are more suitable to be recom- mended to the user. • Well-known learning content management systems. Zaı̈ane and Luo (2001) propose the discovery of useful patterns based on restrictions, to help educators evalu- ate students’ activities in web courses. Li and Zaı̈ane (2004) also use recommender agents for e-learning sys- tems which use association rule mining to discover asso- ciations between user actions and URLs. The agent recommends online learning activities or shortcuts in a course web site based on a learner’s access history. Pahl and Donnellan (2003) analyze a student’s individual ses- sions. They first define the learning period (of time) of each student and then split web server log files into indi- vidual sessions, calculate session statistics, and search for session patterns and time series. Wang, Weng, Su, and Tseng (2004) propose a four phase learning portfo- lio mining approach, which use sequential pattern min- ing, clustering and decision tree creation sequentially, to extract learning features to create a decision tree to pre- dict which group a new learner belongs to. • Adaptive and intelligent web-based educational systems. Lu (2004) uses association fuzzy rules in a personalized e-learning material recommender system. He uses fuzzy matching rules to discover associations between stu- dent’s requirements and a list of learning materials. Romero et al. (2004) propose to use grammar-based genetic programming with multi-objective optimization techniques for providing a feedback to courseware authors. They discover interesting relationships from student’s usage information. Merceron and Yacef (2004) use association rule and symbolic data analysis, as well as traditional SQL queries to mining student data captured from a web-based tutoring tool. Their goal is to find mistakes that often occur together. Frey- berger, Heffernan, and Ruiz (2004) use association rules to guide a search for best fitting transfer model of stu- dent learning in intelligent tutoring systems. The associ- ation rules determine what operation to perform on the transfer model that predict a student’s success. 4.2.3. Text mining Text mining methods can be viewed as an extension of data mining to text data and it is very related to web con- tent mining. It is an interdisciplinary area involving machine learning and data mining, statistics, information retrieval and natural language processing (Grobelnik, Mladenic, & Jermol, 2002). Text mining can work with unstructured or semi-structured datasets such as full-text documents, HTML files, emails, etc. Next, we describe some works on the application of these techniques in differ- ent types of web-based educational systems: • Particular web-based courses. Ueno (2004a) uses data mining and text mining technologies for collaborative learning in an ILMS. She uses text mining for discussion board an expanded correspondence analysis. Learners select the relevant category which represent his/her com- ment and the system provides evaluations for a learner’s comments between peers. Chen, Li, Wang, and Jia (2004) propose to automatically construct an e-textbook via web content mining. They use a ranking strategy to evaluate the web page suitability and they extract con- cept features and build concept hierarchies. Tane, Sch- mitz, and Stumme (2004) propose an ontology-based tool to make the most of the resources available on the web. They use text mining and text clustering tech- niques in order to group documents according to their topics and similarities. Hammouda and Kamel (2005) propose to perform data mining on documents, which serves as a basis for knowledge extraction in e-learning environments. In the process of text mining, a grouping (clustering) approach is also employed to identify groups of documents. • Well-known learning content management systems. Drin- gus and Ellis (2005) propose to use text mining as a strategy for assessing asynchronous discussion forums. C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 143 A ut ho r's pe rs on al co py Text mining techniques improve the educator’s ability to evaluate the progress of a thread discussion. Bari and Benzater (2005) retrieve data from pdf interactive multi- media productions for helping the evaluation of multi- media presentations, for statistics purpose and for extracting relevant data. They identify the main blocks of multimedia presentations and retrieve their internal properties. • Adaptive and intelligent web-based educational systems. Tang et al. (2000) propose to construct a personalized web tutor tree by mining both context and structure of the courseware. They use a key-word-driven text min- ing algorithm to select articles for distance learning students. 5. Conclusions and future research Educational data mining is an upcoming field related to several well-established areas of research including e-learn- ing, adaptive hypermedia, intelligent tutoring systems, web mining, data mining, etc. The application of data mining in educational systems has specific requirements not present in other domains, mainly the need to take into account pedagogical aspects of the learner and the system. Although the educational data mining is a very recent research area there is an important number of contribu- tions published in journals, international congress, specific workshops and some ongoing books (Romero & Ventura, 2006) that show it is one new promising area. Some of the most promising work line is the use of e-learning recommendation agents (Lu, 2004; Zaı̈ane, 2002). These recommender agents sees what a student is doing and recommends actions (activities, shortcuts, contents, etc.) they think would be beneficial to the student. Recom- mender agents can also be integrated in evolving e-learning systems in which materials are automatically found on the web and integrated into the system (Tang & McCalla, 2005). In this way, they help educators to detect which parts of existing materials from heterogeneous sources as the Internet are the best to use for composing new courses. Besides recommenders can also be integrated with domain knowledge and ontologies, combining web mining and semantic web in semantic web mining (Markellou et al., 2005). Semantic web mining is a successful integration of ontological knowledge at every stage of the knowledge discovery process (Becker, Vanzin, & Ruiz, 2005). Educational data mining is a young research area and it is necessary more specialized and oriented work educa- tional domain in order to obtain a similar application suc- cess level to other areas, such as medical data mining, mining e-commerce data, etc. We believe that some future researches lines are: • Mining tools more easy to use by educators or not expert users in data mining. Data mining tools are normally designed more for power and flexibility than for simplic- ity. Most of the current data mining tools are too com- plex to use for educators and their features go well beyond the scope of what a educator may want to do. So, these tools must have a more intuitive and easy to use interface, with parameter-free data mining algo- rithms to simplify the configuration and execution, and with good visualization facilities to make their results meaningful to educators and e-learning designers. • Standardization of methods and data. Current tools for mining data from a specific course may be useful only to its developers. There are no general tools or re-using tools or techniques that can be applied to any educational system. So, a standardization of data, and the preprocess- ing, discovering and postprocessing tasks is needed. • Integration with the e-learning system. The data mining tool has to be integrated into the e-learning environment as another author tool. All data mining tasks (prepro- cessing, data mining and postprocessing) have to be car- ried out into a single application. Feedback and results obtained with data mining can be directly applied to the e-learning environment. • Specific data mining techniques. More effective mining tools that integrate educational domain knowledge into data mining techniques. Education-specific mining tech- niques can help much better to improve the instructional design and pedagogical decisions. Traditional mining algorithms need to be tuned to take into account the educational context. Acknowledgement The authors gratefully acknowledge the financial sup- port provided by the Spanish Department of Research of the Ministry of Science and Technology under TIN2005- 08386-C05-02 Project. References Agrawal, R., Imielinski, T., & Swami, A. (1993). Mining association rules between sets of items in large databases. In Proceedings of the 1993 ACM SIGMOD international conference on management of data, Washington, DC (pp. 207–216). Agrawal, R., & Srikant, R. (1995). Mining sequential patterns. In Eleventh international conference on data engineering (pp. 3–14). Taipei, Taiwan: IEEE Computer Society Press. Arroyo, I., Murray, T., Woolf, B., & Beal, C. (2004). Inferring unobservable learning variables from students’ help seeking behavior. In Intelligent tutoring systems (pp. 782–784). Arruabarrena, R., Pérez, T. A., López-Cuadrado, J., & Vadillo, J. G. J. (2002). On evaluating adaptive systems for education. In Adaptive hypermedia (pp. 363–367). Avouris, N., Komis, V., Fiotakis, G., Margaritis, M., & Voyiatzaki, E. (2005). Why logging of fingertip actions is not enough for analysis of learning activities. In Workshop on usage analysis in learning systems at the 12th international conference on artificial intelligence in education. Baker, R., Corbett, A., & Koedinger, K. (2004). Detecting student misuse of intelligent tutoring systems. In Intelligent tutoring systems (pp. 531– 540). 144 C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 A ut ho r's pe rs on al co py Bari, M., & Benzater, B. (2005). Retrieving data from pdf interactive multimedia productions. In International conference on human system learning: Who is in control? (pp. 321–330). Beck, J., & Woolf, B. (2000). High-level student modeling with machine learning. In Intelligent tutoring systems (pp. 584–593). Becker, K., Ghedini, C., & Terra, E. (2000). Using kdd to analyze the impact of curriculum revisions in a Brazilian university. In Eleventh international conference on data engineering. Proceedings of the SPIE 14th annual international conference on aerospace/defense, sensing, simulation and controls, Orlando (pp. 412–419). Becker, K., Vanzin, M., & Ruiz, D. D. A. (2005). Ontology-based filtering mechanisms for web usage patterns retrieval. In 6th International conference on e-commerce and web technologies (pp. 267–277). Brusilovsky, P., & Peylo, C. (2003). Adaptive and intelligent web-based educational systems. International Journal of Artificial Intelligence in Education, 13, 156–169. Castro, F., Vellido, A., Nebot, A., & Minguillon, J. (2005). Detecting atypical student behaviour on an e-learning system. In I Simposio Nacional de Tecnologas de la Informacin y las Comunicaciones en la Educacin, Granada (pp. 153–160). Chen, G., Liu, C., Ou, K., & Liu, B. (2000). Discovering decision knowledge from web log portfolio for managing classroom processes by applying decision tree and data cube technology. Journal of Educational Computing Research, 23(3), 305–332. Chen, J., Li, Q., Wang, L., & Jia, W. (2004). Automatically generating an e-textbook on the web. In International conference on advances in web- based learning (pp. 35–42). Damez, M., Marsala, C., Dang, T., & Bouchon-Meunier, B. (2005). Fuzzy decision tree for user modeling from human–computer interactions. In International conference on human system learning: Who is in control? (pp. 287–302). Dringus, L., & Ellis, T. (2005). Using data mining as a strategy for assessing asynchronous discussion forums. Computer & Education Journal, 45, 141–160. Farzan, R. (2004). Adaptive socio-recommender system for open-corpus e-learning. In Doctoral consortium of the third international conference on adaptive hypermedia and adaptive web-based systems. Feng, M., Heffernan, N., & Koedinger, K. (2005). Looking for sources of error in predicting student’s knowledge. In Proceedings of AAAI’05 workshop on educational data mining. Freyberger, J., Heffernan, N., & Ruiz, C. (2004). Using association rules to guide a search for best fitting transfer models of student learning. In Workshop on analyzing student–tutor interactions logs to improve educational outcomes at ITS conference. Grob, H., Bensberg, F., & Kaderali, F. (2004). Controlling open source intermediaries – a web log mining approach. In Proceedings of the 26th international conference on information technology interfaces (pp. 233–242). Grobelnik, M., Mladenic, D., & Jermol, M. (2002). Exploiting text mining in publishing and education. In Proceedings of the ICML-2002 workshop on data mining lessons learned (pp. 34–39). Ha, S., Bae, S., & Park, S. (2000). Web mining for distance education. In IEEE international conference on management of innovation and technology (pp. 715–719). Hamalainen, W., Suhonen, J., Sutinen, E., & Toivonen, H. (2004). Data mining in personalizing distance education courses. In World confer- ence on open learning and distance education, Hong Kong. Hammouda, K., & Kamel, M. (2005). Ch. Data mining in e-learning. Hanna, M. (2004). Data mining in the e-learning domain. Computers & Education Journal, 42(3), 267–287. Heiner, C., Beck, J., & Mostow, J. (2004). Lessons on using its data to answer educational research questions. In Proceedings of the ITS2004 workshop on analyzing student–tutor interaction logs to improve educational outcomes (pp. 1–9). Heraud, J., France, L., & Mille, A. (2004). Pixed: an its that guides students with the help of learners’ interaction log. In International conference on intelligent tutoring systems (workshop analyzing student– tutor interaction logs to improve educational outcomes), Maceio (pp. 57–64). Hwang, W., Chang, C., & Chen, G. (2004). The relationship of learning traits, motivation and performance-learning response dynamics. Computers & Education Journal, 42(3), 267–287. Iksal, S., & Choquet, C. (2005). Usage analysis driven by models in a pedagogical context. Ingram, A. (1999). Using web server logs in evaluating instructional web sites. Journal of Educational Technology Systems, 28(2), 137–157. Johnson, S., Arago, S., Shaik, N., & Palma-Rivas, N. (2000). Comparative analysis of learner satisfaction and learning outcomes in online and face-to-face learning environments. Journal of Interactive Learning Research, 11(1), 29–49. Klosgen, W., & Zytkow, J. (2002). Handbook of data mining and knowledge discovery. New York: Oxford University Press. Koutri, M., Avouris, N., & Daskalaki, S. (2004). Ch. A survey on web usage mining techniques for web-based adaptive hypermedia systems. Li, J., & Zaı̈ane, O. (2004). Combining usage, content, and structure data to improve web site recommendation. In International conference on e- commerce and web technologies (pp. 305–315). Lu, J. (2004). Personalized e-learning material recommender system. In International conference on information technology for application (pp. 374–379). Luan, J. (2002). Data mining, knowledge management in higher educa- tion, potential applications. In Workshop associate of institutional research international conference, Toronto (pp. 1–18). Ma, Y., Liu, B., Wong, C., Yu, P., & Lee, S. (2000). Targeting the right students using data mining. In KDD ’00: Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 457–464). Markellou, P., Mousourouli, I., Spiros, S., & Tsakalidis, A. (2005). Using semantic web mining technologies for personalized e-learning experi- ences. In Proceedings of the web-based education (pp. 461–826). Markham, S., Ceddia, J., Sheard, J., Burvill, C., Weir, J., Field, B., et al. (2003). Applying agent technology to evaluation tasks in e-learning environments. In Proceedings of the exploring educational technologies conference. Mazza, R., & Milani, C. (2005). Exploring usage analysis in learning systems: Gaining insights from visualisations. In Workshop on usage analysis in learning systems at 12th international conference on artificial intelligence in education. Merceron, A., & Yacef, K. (2003). A web-based tutoring tool with mining facilities to improve learning and teaching. In Proceedings of 11th international conference on artificial intelligence in education (pp. 201–208). Merceron, A., & Yacef, K. (2004). Mining student data captured from a web-based tutoring tool: Initial exploration and results. Journal of Interactive Learning Research, 15(4), 319–346. Merceron, A., & Yacef, K. (2005). Tada-ed for educational data mining. Interactive Multimedia Electronic Journal of Computer-Enhanced Learning, 7(1), 267–287. Minaei-Bidgoli, B., & Punch, W. (2003). Using genetic algorithms for data mining optimization in an educational web-based system. In GECCO (pp. 2252–2263). Minaei-Bidgoli, B., Tan, P., & Punch, W. (2004). Mining interesting contrast rules for a web-based educational system. In International conference on machine learning applications. Monk, D. (2005). Using data mining for e-learning decision making. Electronic Journal of e-Learning, 3(1), 41–54. Mor, E., & Minguillon, J. (2004). E-learning personalization based on itineraries and long-term navigational behavior. In Proceedings of the 13th international world wide web conference (pp. 264–265). Mostow, J. (2004). Some useful design tactics for mining its data. In Proceedings of the ITS2004 workshop on analyzing student–tutor interaction logs to improve educational outcomes. Mostow, J., Beck, J., Cen, H., Cuneo, A., Gouvea, E., & Heiner, C. (2005). An educational data mining tool to browse tutor–student interactions: Time will tell! In Proceedings of the workshop on educational data mining (pp. 15–22). Muehlenbrock, M. (2005). Automatic action analysis in an interactive learning environment. C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 145 A ut ho r's pe rs on al co py Nilakant, K., & Mitrovic, A. (2005). Application of data mining in constraint-based intelligent tutoring systems. In Proceedings of the artificial intelligence in education, AIED (pp. 896–898). Pahl, C., & Donnellan, C. (2003). Data mining technology for the evaluation of web-based teaching and learning systems. In Proceedings of the congress e-learning, Montreal, Canada. Paulsen, M. (2003). Online education and learning management systems. Bekkestua: NKI Forlaget. Peled, A., & Rashty, D. (1999). Logging for success: Advancing the use of www logs to improve computer mediated distance learning. Journal of Educational Computing Research, 21(4), 413–431. Rahkila, M., & Karjalainen, M. (1999). Evaluation of learning in computer based education using log systems. In ASEE/IEEE frontiers in education conference, San Juan, Puerto Rico (pp. 16–21). Romero, C., & Ventura, S. (2006). Data mining in e-learning. Southamp- ton, UK: Wit Press. Romero, C., Ventura, S., Bra, P., & Castro, C. (2003). Discovering prediction rules in aha! courses. In User modeling (pp. 25–34). Romero, C., Ventura, S., & Bra, P. D. (2004). Knowledge discovery with genetic programming for providing feedback to courseware author. User Modeling and User-Adapted Interaction: The Journal of Person- alization Research, 14(5), 425–464. Sanjeev, P., & Zytkow, J. M. (1995). Discovering enrollment knowledge in university databases. In KDD (pp. 246–251). Sheard, J., Ceddia, J., Hurst, J., & Tuovinen, J. (2003). Inferring student learning behaviour from website interactions: A usage analysis. Journal of Education and Information Technologies, 8(3), 245–266. Shen, R., Han, P., Yang, F., Yang, Q., & Huang, J. (2003). Data mining and case-based reasoning for distance learning. Journal of Distance Education Technologies, 1(3), 46–58. Shen, R., Yang, F., & Han, P. (2002). Data analysis center based on e- learning platform. In Proceedings of the 5th international workshop on the internet challenge: Technology and applications (pp. 19–28). Silva, D., & Vieira, M. (2002). Using data warehouse and data mining resources for ongoing assessment in distance learning. In IEEE international conference on advanced learning technologies, Kazan, Russia (pp. 40–45). Srivastava, J., Cooley, R., Deshpande, M., & Tan, P. (2000). Web usage mining: Discovery and applications of usage patterns from web data. SIGKDD Explorations, 1(2), 12–23. Talavera, L., & Gaudioso, E. (2004). Mining student data to characterize similar behavior groups in unstructured collaboration spaces. In Workshop on artificial intelligence in CSCL. 16th European conference on artificial intelligence (pp. 17–23). Tane, J., Schmitz, C., & Stumme, G. (2004). Semantic resource manage- ment for the web: An e-learning application. In Proceedings of the WWW conference, New York, USA (pp. 1–10). Tang, C., Yin, H., Li, T., Lau, R., Li, Q., & Kilis, D. (2000). Personalized courseware construction based on web data mining. In Proceedings of the first international conference on web information systems engineer- ing, Washington, DC, USA (pp. 204–211). Tang, T., & McCalla, G. (2002). Student modeling for a web-based learning environment: A data mining approach. In Eighteenth national conference on artificial intelligence, Menlo Park, CA, USA (pp. 967– 968). Tang, T., & McCalla, G. (2005). Smart recommendation for an evolving e-learning system. International Journal on E-Learning, 4(1), 105– 129. Tsantis, L., & Castellani, J. (2001). Enhancing learning environments through solution-based knowledge discovery tools. Journal of Special Education Technology, 16(4). Ueno, M. (2004a). Data mining and text mining technologies for collaborative learning in an ILMS ‘‘samurai’’. In ICALT. Ueno, M. (2004b). Online outlier detection system for learning time data in e-learning and its evaluation. In International conference on computers and advanced technology in education (pp. 248–253). Urbancic, T., Skrjanc, M., & Flach, P. (2002). Web-based analysis of data mining and decision support education. AI Communications, 15, 199–204. Wang, F. (2002). On using data-mining technology for browsing log file analysis in asynchronous learning environment. In Conference on educational multimedia, hypermedia and telecommunications (pp. 2005– 2006). Wang, W., Weng, J., Su, J., & Tseng, S. (2004). Learning portfolio analysis and mining in SCORM compliant environment. In ASEE/ IEEE frontiers in education conference (pp. 17–24). Yu, P., Own, C., & Lin, L. (2001). On learning behavior analysis of web based interactive environment. In Proceedings of ICCEE, Oslo/Bergen, Norway. Zaı̈ane, O. (2002). Building a recommender agent for e-learning systems. In ICCE (pp. 55–59). Zaı̈ane, O., & Luo, J. (2001). Web usage mining for a better web-based learning environment. In Proceedings of conference on advanced technology for education, Banff, Alberta (pp. 60–64). Zaı̈ane, O., Xin, M., & Han, J. (1998). Discovering web access patterns and trends by applying OLAP and data mining technology on web logs. In Advances in digital libraries (pp. 19–29). Zarzo, M. (2003). E-learning in the new era of data mining. In International conference on network universities and e-learning, Valen- cia, Spain. Zorrilla, M. E., Menasalvas, E., Marin, D., Mora, E., & Segovia, J. (2005). Web usage mining project for improving web-based learning sites. In Web mining workshop, Cataluna. 146 C. Romero, S. Ventura / Expert Systems with Applications 33 (2007) 135–146 
work_5ra3eporv5glfexqgreko6dnzy	----	Proceedings of the 10th ASAT Conference, 13-15 May 2003 Paper 1E-1 1043 Military Technical College Kobry El-Kobbah Cairo, Egypt 10th International Conference On Aerospace Sciences& Aviation Technology OPTIMAL LOAD SHEDDING FOR MINIMUM UNDER FREQUENCY USING EXPERT SYSTEMS EBTISAM MOSTAFA SATED* ABSTRACT The under-frequency problem due to disturbances in large interconnected power systems, may lead - if not corrected -to an extensive serious damage. Disturbances in large power systems sometimes cause this interconnected systems to become separated into isolated subsystems, and this may lead to excess load, however, overloads the generating units causing the frequency to drop to levels that may cause permanent damage to steam turbines. Load shedding are often applied throughout the system to avoid the turbine damage by restoring the frequency. It helps in restoring frequency in appropriate amount taking into consideration that there is no need in shedding excessive amounts of load. So minimum load shedding for under-frequency is needed which is the main aim of this paper with the help of an expert system. This paper treats the under-frequency problem by using an adaptive traditional method for setting of the under-frequency relays. And an expert system is then suggested and developed as a perfect and fast method to accelerate the optimal shedding process for minimum frequency limit. These methods are implemented and applied on a simple power system. The obtained results are presented and discussed with a comparison between the traditional and suggested expert system methods. KEYWORDS: Under frequency , Load shedding , Power systems, Expert system , Minimum frequency limit. *Associate Professor of Electrical Power Engineering, Faculty of Engineering ( Shoubra ) , Zagazig University Proceedings of the 10th ASAT Conference, 13-15 May 2003 Paper 1E-1 1044 1.INTRODUCTION Electrical power systems are always subjected to experiencing emergency conditions, which may result in a mismatching between generation and load. This may lead to isolated subsystems with either over or under frequency conditions. These subsystems are often suffer insufficient generating capacity to carry the load, and this condition almost lead to a decay of system frequency and may collapse within a short time, if measures are not taken, to restore the load-generation balance [1]. The problem of under-frequency and load shedding has been studied for many years [2-4]. It has been noted through these studies, that the power plants themselves are subject to failure at low frequencies since plant auxiliaries are unable to maintain normal output with the critical frequency being about 10 to 15% below normal frequency [5]. Several treatments of under-frequency load shedding have been presented in the past [6], [7].To maintain a load to generation balance during the system disturbance, practically , automatic under-frequency relays were used to drop the load. In the past, system engineers began investigating the need for turbine protection and initiated an off- frequency study to review all aspects of system abnormal frequency protection [7]. In recent years, expert systems have been used at the inclusion of the operators experience in some areas. And it is certainly possible to bring the solution closer to the operator's routine. The most recent development in expert systems for power system operation shows a number of concepts, which used for solving different power system problems [8-11]. One of these problems is the under frequency problem. In this paper, a combination between a traditional analytical program and a rule- based program is made for the purpose of minimum under-frequency and load shedding problem. A comparison between results of these two programs is also made. 2-THEORETICAL FORMULATIONS 2.1 System Modeling and Frequency Analysis As shown in fig.1, a linear system frequency response model is considered to represent the frequency performance of an isolated subsystem [5], [8]. From the block diagram shown in Fig.1, the frequency response in per unit (p.u.) can be computed as in the following equations: D Aw, , where , Aw indicate the per unit frequency deviation - [ I( + F,TRS) 1'w R(1+ T R S) Proceedings of the 10th ASAT Conference, 13-15 May 2003 Paper 1E-1 1045 r K (1+ F H T,S) 3Aw (3) = Pd,„ Pm = Pau' R(1+ TR S) Aw = (P„ – A,)[ 2HS (4) From eq. (1), and (3), eq. (4) can be rewritten as; = K + F HT,S) la,w)–DAw)[—] R(1+ TR S) 2HS Rearrange this equation in S-domain; (1 + T,S)P„,, = 2 S2 + 2Mw„S + w„ Where, DR+K „ and M = [ 2HR + (DR + K „FH )T, L v DR + K „, = 2HRT, 2(DR + K „,) For a sudden disturbance, Pdist will be in the form of step function in s-domain, Pchange=Pdist /s, and the per unit frequency deviation in time-domain will be; Aw(t)= ( R )(1+ sin((w„ M + c))13„,„„g„ (5) Dr + K „, Where, B - 111- 2TR Mw" +2 TR2 w" and 1 - M M c =tan -I ( w " TR ) tan-I 1 – Mw„ TR ( – M Through these last equations, the value of frequency or speed can be computed for any Pdist. 2.2 Load Shedding Design Generally, load-shedding design depends mainly on the frequency responses of the subsystem model according to many factors. These factors are, the frequency set points, number of frequency steps, and the amount of load shed per step. The minimum allowable operating frequency must be selected to a certain value. Proceedings of the 10th ASAT Conference, 13-15 May 2003 Paper 1E-1 1046 Sufficient load should be shed to limit the frequency decline to above this selected minimum operating frequency value [7]. After setting the initial load shed frequency, load shedding is made in this paper in two protective plans. The first one, consists of four equal load-shedding steps for different values of power disturbance. And the second plan, consists of six load shedding steps of unequal size with the same value of total minimum load shed and for different values of load disturbance. 2.3 Expert System Expert systems are computer programs that constructed to do the kinds of activities that human experts can do. Its programs are usually set up to operate in a manner that will be perceived as intelligent, as if there a human expert on the other side of the video terminal. The appearance of the expert system intelligence is enhanced to the extent that its program can accept free form input in simple English sentences and can state its conclusions in the same way. A program that can do this is said to have a natural language interface. Expert system programs are usually written in languages like LISP and PROLOG, [9], [10]. In this paper Prolog is used as a language for designing the suggested expert system. Fig.2 presents the fundamental components of the expert system which are, [10], [11]; The database: which is composed of different data concerning the power system (the information of the system components). The knowledge base: it contains the knowledge required to perform the expert system's tasks well, where the quality of the expert system relies on its knowledge base). Inference engine: it is needed to perform the exploitation of the knowledge base in order to find solutions. 3. APPLICATIONS AND RESULTS Frequency Response and Load Shedding Results To compute the optimum settings of under-frequency relays, an under frequency turbine protection schemes are designed for a system with assumed sudden disturbance lead to a sudden load change of 0.6, 0.8, and 1.0 per unit. The minimum allowable operating frequency is selected to be 57 Hz. And the system parameters in per unit (p.u.) are [5]; Krr= 0.85, D=1.0, H=3.5, FH=0.3, TR=8.0, R=0.06. All system parameters are estimated based on common knowledge of typical system designs 3.1 Scheme 1: Linear Frequency Response Model Applications and Results Table 1, illustrates the frequency response of sudden load change of 0.6, 0.8, and 1.0 p.0 . By using the linear system response model, fig.1,where the frequency Proceedings of the 10th ASAT Conference, 13-15 May 2003 Paper 1E-1 1047 response is obtained from Eq. (5), two protective plans are designed and compared. Plan1, for load shedding in 4-equal steps, and plan2 for load shedding in 6-steps. Since in large turbine-generator sets are not rated for continuous operation below about 59.5Hz [5], let the first step of load shedding frequency is set at 59.5 Hz. Plan 1: As shown in Table 2, where for each load disturbance the corresponding minimum amount of load to be shed for four equal load shedding steps are tabulated, taking into consideration that the load shedding plan uses a time delay of 0.1 second for the first load shedding step. Table 3, illustrates the frequency response after load shedding in four equal steps (plant) for sudden disturbance equal 0.6, 0.8, and 1.0 p.u. Fig. 3, (a, b, and c), shows the frequency response for 0.6, 0.8, and 1.0 p.u. load disturbance, respectively, before and after load shedding. Plan 2: As shown in Table4, it consists of six load shedding steps of unequal size, for 0.6, 0.8, and 1.0 p.u. load disturbance. The load shedding steps are taken at closer steps than plan1. i.e. for load disturbance=0.6p.u.at frequency =59.5 Hz, load shedding is 0.046 p.u.(time delay=0.lsecond), then at frequency=59.3Hz, load shedding is 0.046p.u., then by step frequency equal 0.2 Hz, the frequency will reach 58.5Hz, at the sixth step and the load to be shed is found 0.036p.u. Note that, the total minimum load shedding for 0.6, 0.8, and 1.0p.u.load disturbance is the same as in plan1,0.24,0.48,and0.64, respectively. Both load shedding in plan1 and plan2 use a time delay of 0.1 sec. for the first step. Note that the larger load shedding is taken first, followed by the smaller one. It is found that through this study, the results of the two plans are close, which indicates that the total amount of load is more important than the exact time of shedding. 3.2 Scheme 2: Expert System Applications and Results Since it is important to detect the need for shedding very quickly and to complete all load shedding before the frequency can decay to a hazardous level, expert system is suggested and provided to help in this part of the problem. The simulation program has been developed to integrate the expert system module. As shown in Fig.4, a load shedding with frequency response program is used to simulate the power subsystem behavior. Hence the expert system module exchanges information with this program. The interface module takes care of the presentation of the solutions proposed by the expert system with the sequence of applied rules. This expert system is applied on the two protective plans, i.e., it is used to identify very fast, 1- the total amount of minimum load to be shed for any load disturbance occurred as a first objective. 2- after the first objective is fulfilled, the second objective will be the identification of the amount of load to be shed at any step, 4- steps (plant), or 6-steps (plan2). 3- identification of step frequency value corresponding to the amount of load shedding at this step. Plan 1 Rule3 Rule4 Rule5 Rule6 IF THEN IF THEN IF THEN IF THEN Proceedings of the 10th ASAT Conference, 13-15 May 2003 Paper 1E-1 1048 3,2.1. Production rules The following rules are proposed to determine the minimum amount of load shedding for any unknown load disturbance and also the amount of load to be shed at any step or frequency level for the two plans ( see flow chart in Fig. 5). Rule1 IF the frequency of the subsystem is declined under 57.0Hz THEN this subsystem is subjected to under frequency problem Rule2 IF the load disturbance is matched with that in database THEN total minimum load shedding will identified the study is for plant load will be shed in four steps the number of steps is four load shedding is equal for all steps load shedding for plant is in stepl set the initial frequency and frequency interval the load shedding for plant is in last step frequency must be 58Hz Pian2 Ruie7 IF THEN RuIe8 IF THEN shedding Ruler9 IF THEN Rule i0 IF THEN the study is for plan2 load will be shed in six unequal steps the number of steps is six the larger steps are taken first followed by smaller steps of load load shedding for plan2 is in step1 set the initial frequency and frequency interval the load shedding for plan2 is in step six frequency must Pe 58.5Hz. Results Sequence of Rules Rule1 the subsystem is subjected to under frequency problem. Rule2 consider the subsystem is subjected to new load disturbances 0.70p.u, and 0.90p.u. If these values are not matched with those in database, they must be approximated Plan1 Rule3 for plant, number of steps =4 Rule4 total minimum load shed and load shed for each step; for 0.70p.0 load disturbance, total load shed=0.36p.u. load shed for each step=0.09p.u. for 0.90p.0 load disturbance, total load shed=0.56p.u,load shed for each step=0.14p.u. Rule5 initial frequency is set at 59.5Hz,and frequency interval is 0.5Hz Proceedings of the 10th ASA T Conference, 13-15 May 2003 Paper 1E-1 1049 Rule6 : load to be shed at different frequency levels with 0.5Hz interval step frequencv(Hz.), load to be shed(p.u)for load disturbance 0.70p.0 0.90p.0 1 59.5 0.09 0.14 2 59.0 0.09 0.14 3 58.5 0.09 0.14 4 58.0 0.09 0.14 Plan2 Rule7 : for plan2, number of steps=6 Rule8 : total minimum load shed is 0.36p.0 for 0.70p.0 load disturbance and 0.56p.0 for 0.90p.0 load disturbance. Rule9 : initial frequency is set at 59.5Hz,and frequency interval is 0.2Hz Rule10 : load to be shed at different frequency level with 0.2Hz interval step frequencv(HZ), load to be shed(p.u)for load disturbance 0.700 0.900 1 59.5 0.069 0.106 2 59.3 0.069 0.106 3 59.1 0.057 0.089 4 58.9 0.057 0.089 5 58.7 0.054 0.085 6 58.5 0.054 0.085 Hint: If the new load disturbance value is mismatched with those in data base, this value is approximated to the nearest value in data base. 3.2.2 Comparisons and Discussions For the first scheme, as a comparison between the two plans, it is obvious that, they both have a close view of the time frame when the load shedding is being initiated time delay = 0.1second). And, since the amount of the total load shed is the same for the two plans, the comparison is based on the running time taken for the two programs. It is found that, for each step of the two plans, the execution time was 0.39seconds. Therefore, for plant, the total time taken for running is 1.56 seconds, while for plan2, the running time was found 2.34seconds. For the second scheme; beside the right and accurate action for using this expert system, the time taken for searching through this expert system, to make the decision, is found too small compared with that time for the first scheme (0.59 seconds for plant, and 0.81seconds for plan2). Therefore expert system is considered simple, fast, and accurate in this area of study. Both of the two scheme's programs would be executing in a computer with main processor is Intel Pentium W 733MHz., Cash memory is 256k, and 64MB Ram. Proceedings of the 10m ASAT Conference, 1345 May 2003 Paper 1E4 1050 4.CONCLUSUIONS Two schemes were proposed and suggested for determining the minimum amount of load to be shed for under frequency problem caused by a sudden load disturbance. The first scheme is based on the frequency detection by using a linear system frequency response model. This scheme is applied for two plans. The first plan deals with four equal load shedding steps, while the second plan consists of six load shedding steps of unequal size, and taken at steps that are closer together than the first plan but with the same total shed. As the size of the disturbance may increases, there will be greater need for rapid action. From this point of view, it is important to determine very quickly how much load shedding is required for a given disturbance. Expert system is suggested and introduced as a second scheme to solve this problem. It helps the operator to take his decision on a higher level without the need of applying the analytic programs and search for optimal total load shedding and load to be shed at every frequency level. A comparison between the two schemes is made and good results are obtained. REFERENCES [1] IEEE Committee Report, "Survey of Under frequency Relay Tripping of Load Under Emergency Conditions", IEEE Trans., v PAS-87, March, pp.1362-1366, (1968). [2] Malizewski, R. M., Dunlop, R. D. and Wilson, G. L.,"Frequency Actuated Load Shedding and Restoration. Part I. Philosophy", IEEE Trans. V PAS-90, pp.1452-1459, (1971). [3] Horowitz, S. H., Politis, A., and Gabrielle, A. F., "Frequency Actuated Load Shedding and Restoration. Part 11. Implementation", IEEE Trans., v PAS-90, pp.1460-1468, (1971). [4] IEEE Working Group on Methods of System Preservation During Under- Frequency Conditions, "A Status Report on Methods Used for System Preservation During Under Frequency Conditions", Power System Relaying Commetti Report, IEEE Trans. V PAS-94, n2, pp.360-366 (1975). [5] P. M. Anderson and M. Mirheydar, "An Adaptive Method for Setting Under Frequency Load Shedding Relays", IEEE Trans. On Power Systems, vol.7,no.2, 1992, pp.647-653. [6] R. S. Hahn, S. Disgupta, E. Baytch and R. D. Willoughby, "Maximum Frequency Decay Rate for Reactor Coolant Pump Motors", IEEE Trans., v NS26, pp.863-870, (1977). [7] D. W. Smaha, C. R. Rowland and J. W. Pope, "Coordination of Load Conservation with Turbine-Generator Underfrequency Protection", IEEE Trans., v PAS-99, pp.1137-1150, (1980). [8] P. M. Anderson and M. Mirheydar, "A low-Order System Frequency Response Model", IEEE Trans. On Power System, v 5, n.3, pp.720-729, (1990). [9] D. Reichelt and H. Glavitsch, "A knowledge-Base System for Security Enhansment Combining Al-Techniques and Analytic Algorithms, 10th Power Systems Computation Conference PSCC, Graz, Austria, (1990). Proceedings of the 10th ASAT Conference, 13-15 May 2003 Paper 1E-1 1051 [10]D.Reichelt and H. Glavitsch, "Features of Hybrid Expert System for Security Enhancement", IEEE Trans. On Power Systems, vol.7, No.2, pp.907-913, (1992). [11]Tang, Guoqing, Xu, Hongyuan, and Chen Heng, "An Expert System for VoltageNAR Control of Interconnected Power Systems", ESAP'93, pp.292- 295, (1993). [12]Josiane Razanamampandry, Pierre-Andre'Chamorel, and Alain J. Germond, "Expert System for Voltage and VAR Control", ESAP'93, pp.296-302,(1993). Proceedings of the le ASAT Conference, 13-15 May 2003 Paper 1E-1 1052 is the high-pressure power fraction of the reheat turbine. Km: is the gain ID : is the damping factor : is the inertia constant of the subsystem. constant. TR: is the reheat time Fig.1 A linear system frequency response model for an isolated subsystem with P th,t. as input. Power system Network database system Expert Knowledge base (Search prorsiprodfct system) Inference engine (analytical Programs) t language interface 1, t user result display Fig.2 The expert system configuration is 5, a 10 A- I.% a 2 4.0 0.484 8 10 Proceedings of the 10th ASA T Conference, 13-15 May 2003 Paper 1E-1 1053 (a) Pd,„„r,3.6 p.u. (b) p.u. (c) Pd„„w=1.0 p.u. Fig. 3 Frequency response, before and after load shedding for load disturbance, 0.6, 0.8, and 1.0 p.u. Expert system module Load shedding with frequency I response programming I decisions Interface module Fig. 4 Simulation program scheme. Data base, Pte= 0.6, 0.8, 1.0.p.u. For plansl&2, - fu & (f) for each step. - Total load to be shed. - Load to be shed r each step. Check: If the subsystem frequency f<5714z? Yes The system is subjected to under Frequency problem. Get minimum load shedding and all details for plank & plan2 (from knowledge base) for Plant : get -Total load to be shed in 6-steps - Load shedding at any frequency step with:- * Initial frequency 59.5Hz * Frequency interval 0.2Hz * Final frequency 58.5Hz Proceedings of the 10th ASA T Conference, 13-15 May 2003 Paper 1E-1 1054 Check: If the disturbance match with those in data base? Pa„,,,,„. Fla = 0.6, 0.8, 1.0.p.u.? Expert system (knowledge base) Approximation is made No To the nearst case with 4 Those in data base. If using Prolog program Use "round" statement For approximation Plank : get -Total load to be shed in 4-steps. -Load shedding at any frequency step with:- * Initial frequency 59.5Hz * Frequency interval 0.5Hz * Final frequency 58Hz. L $ [ Results Fig.5 Flow chart of the suggested expert system. fr uen f Hz Pchan e--0.8 .u. Pchan 0 60.00 57.16 55.31 54.20 53.62 53.43 53.49 53.71 54.02 54.38 54.74 55.09 55.41 55.69 55.99 56.15 56.32 u. • 56.45 Proceedings of the 10°' ASAT Conference, 13-15 May 2003 Paper 1E-1 1055 Tablet Frequency response of sudden load change (Pc h",) of 0.6, 0.8,1.0 p.u. Time sec 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 Pchan e=0.6 60.00 57.87 56.48 55.65 55.22 55.07 55.12 55.28 55.52 55.78 56.06 56.26 56.56 56.77 56.96 57.11 57.24 Table2 Minimum load shedding and frequency settings of plant Step step load shedding (p.u.) frequency (Hz) Pchange=0.6pu Pchange=0.8pu Pchange=1.0pu 59.5 0.06 0.12 0.16 2 59.0 0.06 0.12 0.16 3 58.5 0.06 0.12 0.16 4 58.0 0.06 0.12 0.16 total min. load shedding 0.24p.u, I 0.48p.u. I 0.64p.u. Proceedings of the 10m ASAT Conference, 13-15 May 2003 Paper 1E-1 1056 Table3 Frequency response after load shedding of plant time (sec.) frequency f (Hz) Pchange=0.6 p.u. Pchange=0.8 p.u. Pchange=1.0 p.u. 0.0 60.00 60.00 60.00 0.5 58.72 58.86 58.92 1.0 1.5 57.89 57.39 58.12 58.29 57.68 57.79 2.0 57.13 57.45 57,53 2.5 57.04 57.37 57.44 3.0 57.07 57.39 57.49 3.5 57.17 57.48 57.57 4.0 57.31 57.61 57.81 4.5 57.47 57.75 57.97 5.0 57.63 57.89 58.11 5.5 57.79 58.04 58.26 6.0 57.93 58.16 58.39 6.5 58.06 58.28 58.53 7.0 58.17 58.38 58.67 7.5 58.26 58.46 58.81 8.0 58.34 58.53 58.94 Table 4 Minimum load shedding and frequency settings of plan2. step frequency (Hz) step load shedding (D.u.1 Pchamo=0 6pu I 0.046 Pceance=0.8Pu 0.091 Pchanae=1 .0pu 0.122 1 59.5 2 59.3 0.046 0.091 0.122 3 59.1 0.038 0.077 0.102 4 58.9 0.038 0.077 0.102 5 58.7 0.036 0.072 0.096 6 58.5 0.036 0.072 0.096 total min. load shed. 0.24p.u. 0.48p.u. 0.64p.u. Page 1 Page 2 Page 3 Page 4 Page 5 Page 6 Page 7 Page 8 Page 9 Page 10 Page 11 Page 12 Page 13 Page 14 
work_5tcfhjy3nvfqpg7ip4bhhly63i	----	Microsoft Word - 03_2010.docx CENTRE FOR EMEA BANKING, FINANCE & ECONOMICS Technical Efficiency in Saudi Arabian Banks Albert Assaf, Carlos P. Barros and Roman Matousek No 03/10 Working Paper Series 2 Technical Efficiency in Saudi Arabian Banks Albert Assaf 1 Carlos Pestana Barros 2 and Roman Matousek 3 1 Instituto de Economia e Gestão, Technical University of Lisbon, Rua Miguel Lupi, 20, 1249-078 Lisbon, , Portugal.cbarros@iseg.utl.pt 2 Centre for Tourism and Services Research, Victoria University, Australia. albert.assaf @ vu.edu.au. 3 Corresponding author: Centre for EMEA Banking, Finance and Economics, London Metropolitan Business School, London Metropolitan University, 84 Moorgate, London, EC2M 6SQ. E-mail: r.matousek@londonmet.ac.uk. Abstract This study analyses the technical efficiency of Arabian banks using a DEA-Data Envelopment Analysis frontier model. In the first stage, a bootstrapped variable returns to scale (VRS) DEA model is used to identify the efficient scores. In the second stage, a bootstrapped truncated regression is adopted to identify the covariates that explain technical efficiency. Policy implications are derived. Keywords: Arabian banks, DEA-data Envelopment Analysis, Double Bootstrap, Truncated Regression 3 1. Introduction The Saudi banking sector has undergone substantial changes over the last decade. Banks have expanded their operations and have taken advantage of scale and scope economies as well as product diversification. The driving force behind these changes has been the recent gradual liberalization of financial sector, globalization of financial markets, changes in technology, product innovation and the growth of business activities by Islamic countries in the West (El-Gamal, 2006). The Saudi banking system is quite unique compared to the traditional banking system. It is under strict regulation imposed by SAMA (Saudi Monetary Agency) and has several distinguished factors. Firstly, Saudi banks are unique in the fact that they provide a combination of conventional banking and Islamic banking. Secondly, the overwhelming dependence on oil causes difficulties for domestic banks to diversify credit risk which consequently requires from banks to hold high levels of capital. Thirdly, the banks are funded by low cost demand deposits. 1 Empirical research on banks efficiency in Arabic peninsula is still very limited as opposed to extensive research in Europe and USA. There are only several studies that tackle this issue in Islamic countries, see for example Avkiran (2009), Mostafa (2009), Emrouznejad and Anouze (2009), Hisham et al (2008), Essayyad and Madani (2003). The main motivation of the present research is to analyse the technical efficiency of Saudi Arabian banks, a richer leading oil country that aims to develop supported in the foreign earnings obtained in oil. The banks are the institutions that channel the oil funds to the companies and families and therefore are strong determinants in the allocation of capital, financial stability and the competitiveness and development of manufacturing and services (Beck et al., 2003). This paper also aims to analyse the differences in technical efficiency among purely domestic banks and foreign banks that operate on the basis of join-venture with financial institution in Saudi Arabia. Based on the results we aim to provide and discuss implications for Saudi banking. The present paper contributes to the literature on banking performance analysing the technical efficiency of Saudi Arabia with a simultaneous two-stage DEA bootstrap model. In the first stage, Data Envelopment Analysis (DEA) is used to estimate the relative efficiency scores ranking banks according to their efficiency. In the second stage, the Simar and Wilson (2007) procedure is applied to bootstrap the DEA scores with a truncated regression. This approach allows us to obtain more reliable evidence compared to previous studies analysing the efficiency. The Simar and Wilson (2007) 1 IMF claims that about 40% of total assets is funded by demand deposits. 4 procedure ensures the efficient estimation of the second-stage estimators, which is not a property of alternative methods. The remainder of the paper is organised as follows: in Section 2, we describe developments in the Arabian banking sector. In Section 3, we present the literature survey whilst the data is presented in Section 4. Section 5 discusses methodology. Empirical results are presented in Section 6. Finally, Section 7 concludes. 2. Contextual Setting The Saudi banking system is small compared with most OECD banking systems. The monetization of the banking system measured in terms of private credits to GDP was just 37per cent in 2008. The banking system has displayed a high degree of stability and strong resilience to external shocks till 2007. The stability of the sector has been enhanced by its strict regulatory rules imposed by the Saudi Arabia Monetary Fund (SAMA). The distinguished characteristics of the Saudi banking sector as opposed to traditional west banking systems is the blend of Islamic banking and “Islamic Windows” of conventional banking. Islamic banking also known as Islamic Shariah based banking system is different from conventional banking. The concept of Islamic banking is based on its profit-and-loss sharing paradigm (PLS). This concept is underpinned by five code of belief in Islamic finance, i.e., avoidance of Riba (interest), Gharar (uncertainty), Mysur (gambling), Haram (prohibited) and sale of the items not owned or possessed. The main features of Islamic banking have been outlined, for example, by Chong and Liu (2009), Taylor (2004). The Islamic Banking contributes to a significant profitability of Saudi Banking sector. Islamic financial products dominate the Saudi market. Islamic banks control some 62% of total assets. It is estimated that about 40% of deposit are non interest bearing because of riba. The gradual deregulation process of financial services allowed foreign financial institutions to provide financial services in Saudi Arabia. As a reaction to this process, domestic Saudi banks have also introduced a large scale of new products and services. The proportion of Islamic banking has increased recently significantly. Table 1 shows the share of Sharia-Compliant assets of banking sector. 5 <Insert Table 1 around here> During the period 1996-2005, Saudi bank assets grew by 213 per cent, deposits by 224 per cent and capital and reserves by 248 per cent. Table 2 indicates the performance of the Saudi Banking sector from 2000 to 2007, measured in terms of ROA and ROE. The profits of the banking system showed strong growth - return on equity averaged well over 20 per cent and return on assets more than 2 per cent. <Insert Table 2 around here> Saudi banks have a dominant position in retail banking with about 1,300 branches. It is expected that foreign banks increase competition for business in project finance, initial public offerings and Islamic financial instruments. But the retail side is also facing competition. There are currently thirteen commercial banks in Saudi Arabia. Table 3 lists the largest commercial banks and describes their main activities. Three banks are fully Saudi owned, seven have minority foreign ownership and one foreign bank has a branch presence — Gulf Investment Bank (Bahrain). Five banks also have a joint venture agreements with major international banks under which the latter provide management and technical support. Banks operate on the universal banking model and provide a broad range of products and services including retail and corporate banking, investment management and advice, and both domestic and international securities brokerage services. It is important to stress that most domestic banks have some government participation. In 2005, Saudi Arabia formally joined WTO (World Trade Organization) and consequently foreign investors have been allowed to own up to 60% (previously 40%) of any bank. But this change has not yet been fully reflected in the changes of the ownership structure. SAMA has granted ten new licenses to foreign banks since 2004, but only five banks so far opened for business. These new foreign banks are restricted to a single branch. The Gulf International Bank (Bahrain) arrived first in 2000, followed by the Emirates Bank International (United Arab Emirates), the National Bank of Kuwait, and the National Bank of Bahrain in 2002. 2 <Insert Table 3 around here> 2 The list includes Deutsche Bank, BNP Paribas, J.P. Morgan, National Bank of Kuwait, Emirates Bank, , State Bank of India and National Bank of Pakistan. 6 3. Literature Survey Empirical research analysing banks efficiency has been widely addressed in the literature. There has been extensive use of non-parametric and parametric methods such as DEA (Data Envelopment Analysis) and stochastic frontier to measure technical efficiency (Alam, 2001; William, Peypoch and Barros, 2009 Pasiouras et al., 2008; Hahn, 2007; Ataullah and Le, 2008). The present study analyses the technical efficiency of Saudi Arabian banks with a two-stage DEA model, where in the first stage the technical efficiency is bootstrapped (Charnes et al., 1978) and in the second stage the Simar and Wilson (2007) procedure is used to bootstrap the DEA scores with a truncated regression. The research on banks efficiency in Islamic countries is still rather limited. Hisham et al. (2008) analysed production efficiency of Islamic Banks and Conventional Bank Islamic Windows in Malaysia by using the variable returns to scale DEA model. The study showed that Islamic banks are more efficient at controlling costs than profits. The driving force of cost efficiency was resource management and economies of scale. In a follow up study, Avkiran (2009) applied network data envelopment analysis (NDEA) for banks operating in the United Arab Emirates. It was argued that NDEA is superior to the standard DEA technique since it provides adequate detail information for management to identify the specific determinants of inefficiency. Avrikan (2009) discussed that this method can be used in countries where an industry exhibits high levels of inefficiency and managerial strategies of how to eliminate the remaining inefficiencies have to be implemented. A further strand of applied research on banking in Arabic peninsula is focused on competition by using the traditional structural and non-structural models (Turk-Arris , 2008; Al- Muharani et al., 2006). Essayyad and Madani (2003) investigated concentration, efficiency, and profitability of commercial banks operating in Saudi Arabia. They found that the banking system was highly concentrated and lacked sound competitive environment. However their results focused on the 7 1989- 2001 period which was characterized by major structural changes resulting from the membership of Saudi Arabia in WTO. To our best knowledge there are no studies that apply an advanced methodological approach - two stage DEA bootstrap model - for Saudi banking. 4. Data Our analysis includes nine banks that currently operate in the Saudi Arabia. Data are collected from annual reports over the period 1999 – 2007 (81 observations). The DEA two stage procedures are adopted. In the first stage a Bootstrapped DEA model is used to estimate the technical efficiency. To model the bank production process, we follow the intermediation approach (see Sealey and Lindley, 1977) and assume that banks purchase liabilities that are transformed into earning assets. Banks are assumed to produce four outputs that cover both on and off-balance sheet activities: (i) total customer loans, (ii) Securities and (iii) interbank loans. Three inputs are used to produce bank output: (iv) total employees; (v) fixed assets and (vi) total deposits. The descriptive statistics are shown in Table 4. <Insert Table 4 around here > 5. Methodology 5.1 Efficiency Measurement DEA is used in the first stage for estimating the technical efficiency of Arabic banks. The motivation and early versions of the DEA models have appeared in several previous studies in the literature, so they will not be reiterated here. For a detailed review refer to Coelli et al. (1998). The model used in this study follows an output oriented assumption and can be derived for the i-th bank by solving the following linear programming: 8 ˆ , ˆ ˆmax ; ; ; i n n n i i i i i i i y Y x X δ λ δ δ δ λ λ λ λ =1 =1 =1   = > 0 ≤ ≥ = 1 ≥ 0    ∑ ∑ ∑ , i =1….n banks (1) where Y is vector of bank outputs , X is s vector of bank inputs, λ is a 1I × vector of constants. The value of ˆ i δ obtained is the technical efficiency score for the i-th firm. A measure of ˆ i δ =1 indicates that the bank l is technically efficient, and inefficient if ˆ i δ >1. This linear programming problem must be solved n times, once for each bank in the sample. Note that the DEA model can also be estimated using either the constant return to scale (CRS) 3 or variable return to scale (VRS) assumptions and the shape of the frontier will differ depending on the scale assumptions that underline the model. In this paper we rely on the VRS assumption, as the CRS is only correct as long as it is appropriate to assume that banks are operating at an optimal level of scale. Technological advances and regulatory changes might vary across banks in various size groups, so allowing for VRS would permit modelling of the entire range of technology. changes might vary across 5.2 The Bootstrap Approach A new debate has recently been raised in the literature regarding the statistical limitations of DEA scores. Simar and Wilson (1998, 1999, and 2007) emphasise that efficiency scores generated by DEA are strongly dependent on each other in the statistical sense, and thus using them in a second stage regression might violate the basic model assumption required by regression. A main reason for this problem is the well-known fact that the DEA efficiency score is a relative efficiency index, not an absolute efficiency index. One cannot also obtain statistical properties of DEA, as the efficiency scores are calculated rather than estimated. Recently, Simar and Wilson (2007) proposed a procedure, based on a double bootstrap, which enables consistent inference in the second stage regression, while simultaneously constructing confidence internal and producing standard errors for the DEA efficiency scores. The bootstrap is a computer-based method which is based on the idea of resampling from an original data in order to assign statistical properties for the quantities of interest. Efron (1979) was the first to introduce method, and since then it has become a popular and powerful statistical tool. For more technical details on the method refer to Efron (1979) and Efron and Tibshirani (1993). 3 A production function is said to exhibit constant return to scale (CRS) if a proportionate increase in inputs results in the same proportionate increase in outputs. The variable return to scale (VRS), on the other hand, doe not assume full proportionality between the inputs and outputs. 9 In the case of DEA, Simar and Wilson (1998) were the first to introduce the method to obtain statistical properties of the efficiency scores, and in a follow-up paper they extended their approach to account for the impact of environmental variables 4 on efficiency (Simar and Wilson, 2007). Before illustrating their procedure we first present the following model: ˆ i i i zδ β ε= + (8) where i z is a vector of management related variables which is expected to affect the efficiency of firms under consideration and β refers to a vector of parameters with some statistical noise i ε . A popular procedure in the literature is to use the Ordinarily Least Square (OLS) regression to estimate this relationship. However, as described in Simar and Wilson (2007), this might lead to estimation problems mainly related to the correlation and dependency problems of the efficiency scores which violate the regression assumption that i ε are independent of i z . The importance of the Simar and Wilson (2007) procedure is that it produces with bias corrected estimates of ˆ i δ valid estimates for the parameters in the regression model. We describe their bootstrap algorithm in the following steps: i. Calculate the DEA output-orientated efficiency score ˆ i δ for each bank, using the linear programming problem in (1): ˆ , ˆ ˆmax ; ; ; i n n n i i i i i i i y Y x X δ λ δ δ δ λ λ λ λ =1 =1 =1   = > 0 ≤ ≥ = 1 ≥ 0    ∑ ∑ ∑ for 1, ...i n= ii. Use the maximum likelihood method to estimate the truncated regression of ˆ i θ on i z , to provide and estimate β̂ of β and an estimate ˆεσ of εσ . iii. For each bank 1, ...i n= , repeat the next four steps (1-4) B times to yield a set of bootstrap estimates{ }*,ˆ , b 1,...Bi bδ = . 1. Draw i ε from the 2ˆ(0, )N εσ distribution with left truncation at ˆ(1 ) i zβ− 2. Compute * ˆ i i i zδ β ε= + 3. Construct a pseudo data set * * ( , ) i i x y , where * i i x x= and * *ˆ / i i i i y y d d= 4 These are variables that are neither inputs nor outputs but are used to mainly explain the variation in the efficiency scores. 10 4. Compute a new DEA estimate * i d on the set of pseudo data * * ( , ) i i x y , i.e. Y and X are replace by { }* * , 1,...iY y i n= = and { } * * , 1,... i X x i n= = in () iv. For each bank, compute the bias corrected estimate ˆ̂ ˆ ˆ i i i biasd d= - , where ˆ i bias is the bootstrap estimator of bias obtained as: * , ˆ ˆˆ B i i b i b bias B δ δ =1 1 = −∑ . v. Use the Maximum likelihood method is again used to estimate the truncated regression of ˆ̂ i δ on i z , providing estimates ( ˆ ˆˆ ˆ,β σ ) of ( , εβ σ ). vi. Repeat the next three steps (1-3) B2 times to obtain a set of bootstrap estimates * *ˆ ˆˆ ˆ, , ,..... b b b Bβ σ 2    = 1      1. For , , , , , ,i n= 1 , i ε is drawn from ( )ˆ̂,N σ0 with left truncation at ˆ̂ izβ 1 −    2. For , , , , , ,i n= 1 , compute ** ˆ̂ i i i zδ β ε= + 3. The Maximum likelihood method is again used to estimate the truncated regression of ** i δ on i z , providing estimates ( * *ˆ ˆˆ ˆ,β σ ) vii. Use the bootstrap results to construct confidence intervals. 6. Empirical Results and Discussions 6.1 Efficiency measurement 11 The VRS technical efficiency estimates the different Arabic banks, obtained from 2000 5 bootstrap iterations are reported in Table 5. In order to compare our method with the traditional DEA model, we report in Table 5 the original (non-bootstrapped) DEA estimates. Due to the upward-bias in the original estimates and due to the bootstrap correction in the confidence intervals, the original estimates lie for every observation outside but close to the lower bound for the confidence interval. However, the bias corrected estimates lie for every observation inside the confidence interval. As indicated before, the applied methodology is based on the bootstrapped DEA given its statistical advantage over the traditional DEA method. The results clearly indicate that the average efficiency score of Saudi banks has increased since 1999 to reach an average efficiency level of 90.21% in 2007, when several banks operated at a high efficiency level such as Fransi, SAAB, Jazira, and Hollandi. The lowest performing banks include Riyad and Samba. Their average technical efficiency was 86.71% and 88.84% respectively in 2007. On average Saudi banks were nearly 9.79% away from their frontier - maximum efficiency. An investigation of each individual year indicates that the average efficiency score has gradually decreased in the period 1999-2003- before starting to increase consistently until 2007. The period of low efficiency might be related to the economic slowdown in the late 1990s and the beginning of new millennium. Low oil prices with relatively high interest rates forced banks to scale down their business activities and this might have impacted negatively on the efficiency results. In terms of the increase in efficiency, it is possible to discuss the long term impact of the restructuring and consolidation process which started in the second half of 1990s and lasted up to 2003. Banks increased their capital base that strengthened their position. This enabled them to increase the deposit- raising potential and their provision for doubtful and non-performing loans. Another important dimension of the restructuring process may be seen in the improvement of information technology, and development of new financial products and services. The gradual openness for new foreign banks not only from the Gulf Cooperation Council countries but also from the West is also a possible influential factor. <Insert Table 5 around here> 6.2 Determinants of Technical Efficiency 5 Simar and Wilson (1998) recommended the use of 2000 bootstrap iterations to obtain reliable bootstrap estimates. 12 In order to account for the sources of efficiency changes we have also regressed the efficiency scores on variables that are expected to have the impact on the technical efficiency. The bootstrap procedure used in this stage was explained in Section 5. The model can be expressed as follows: ˆ ln it it it it it it it TA Liq NPM PR Dummy tδ β β β β β β β ε 0 1 2 3 4 5 6 = + + + + + + + (9) where ˆ it δ is the technical efficiency scores; ln it TA is log of total assets , it Liq is the liquidity, it NPM is the net profit margin, it PR is the payout ratio, Dummy is the ownership dummy, which is one if one for foreign banks and zero elsewhere t is a time trend to capture any missing dynamics, and it ε is random error representing statistical noise. The results of the model are displayed in Table 6. Our estimation shows that technical efficiency increases with assets, signifying that large size contributes to higher technical efficiency. Another variable, the net profit margin is significant and positive; however, the value of the coefficient is rather marginal. It is expected that more efficient bank would have their net profit margins lower compared to less efficient banks. The coefficient of the liquidity ratio is also positive and significant. Moreover, we find that the PR-payout ratio has a negative impact on the technical efficiency. In other words, less profit is retained by banks the lower is their technical efficiency. This finding has an important implication for managers. It shows that the shareholders should sacrifice their dividends and allow banks to re-invest their profit. It seems that a low level of retained profit imposes constraints on banks activities. The Saudi banks need to invest extensively in technologies, product development, human capital and branch networking. Therefore, the policy of a high PR-payout ratio does not seem to be appropriate. A dummy variable that distinguishes purely domestic banks and Saudi banks with foreign capital is negative. This result is apparently in contradiction with the role of foreign banks, see for example, De Haas and van Lelyveld (2006) and Lensink et al. (2008). The presence of foreign capital in joint venture institutions should enhance managerial skills and technical support. In the case of 13 Saudi banks, the results show that domestic banks have superior information and managerial skills within the Saudi Arabia market. Thus, it indirectly supports the hypothesis that managerial skills cannot be easily imported to the Saudi banking system mainly due to lack of expertise in specific area of Islamic banking. 6. Conclusion This study has analysed the technical efficiency of Saudi Arabian banks from 1999 to 2007 by applying a two stage DEA bootstrap model. The limitations of the popular DEA approach, extensively used to estimate the efficiency of the international banks, were corrected with the use of an advanced bootstrap approach. The overall goal was to improve the accuracy and consistency of DEA results and to provide bank policy makers with more reliable evidences on the possible reasons of efficiency variations between the different Arabic banks. The general conclusion is that Saudi Arabian banks are currently operating in a high efficient environment. Notwithstanding, our results show that there is room for Saudi banks to increase their efficiency. The most efficient bank in the period (Ryad) should be used as benchmark for the least efficient banks. We have also found that Saudi banks with foreign capital have to improve their technical efficiency. This result is in contradiction with the general notion that foreign capital brings managerial skills. The present study can provide a starting point for further investigation and validation into the efficiency of the Arabic bank sector. This strand of research can provide important information for policy makers as for the openness of Saudi banking to new banks. Therefore, more investigation with alternative models can cross validate the present research. 14 7. References Al-Muhrani, S., Matthews, K., & Khabari,Y. (2006). Market Structure and competitive conditions in the Arab GCC banking system. Journal of Banking and Finance, 30, 3487-3501. Alam, I. M. S. (2001). A Non-Parametric Approach for Assessing Productivity Dynamics of Large Banks. Journal of Money, Credit and Banking, 33, 121-139. Ataullah, A.,& Le, H. (2008). Economic reforms and bank efficiency in developing countries: the case of the Indian banking industry. Applied Financial Economics, 16, 653- 663. 15 Avkiran, N.K.(2009). Opening the blac box of efficiency analysis: An Illustration with UAE banks. Omega, 37, 930-941. Barros, C.P., Ferreira, C., & Williams, J. (2007). Analysing the Determinants of Performance of the Best and Worst European banks: A Mixed Logit Approach. Journal of Banking and Finance, 31, 2189-2203. Beck, T., Demirgüç-Kunt, A., & Levine, R. (2003). Law, endowments, and finance. Journal of Financial Economics, 70, 137-181. Coelli, T., Prasada Rao, D.S., & Battese, G. E. (1998). An Introduction to Efficiency and Productivity Analysis. Boston: Kluwer, Academic Publishers). De Haas, R., & van Lelyveld, I. (2006). Foreign banks and credit stability in Central and Eastern Europe: A panel data analysis. Journal of Banking & Finance, 30, 1927-1952. El-Gamal, M. A. (2006). Islamic Finance, Law, Economics, and Practice. Cambridge: Cambridge University Press. Essayyad, M., & Madani, H. (2003). Investigating bank structure of an open petroleum economy: the case of Saudi Arabia. Journal Managerial Finance, 29, 73 - 92. Emrouznejad, A., Anouze, A.L. (2009), A note on the modeling the efficiency of top Arab banks, Expert Systems with Applications, 36, 5741-5744. Hahn, F.R. (2007). Domestic mergers in the Austrian banking sector: a performance analysis. Applied Financial Economics, 17, 185-196. Hisham, K.M., Sauman, M.S., & Rohani, M. (2008). Assessing production efficiency of Islamic banks and conventional bank Islamic windows in Malaysia. International Journal of Business and Management Research, 1, 31-48. Lensink, R., Meesters, A., & Naaborg, I. (2008). Bank efficiency and foreign ownership: Do good institutions matter?. Journal of Banking and Finance, 32, 834-844. Mostafa, M., M. (2009). Modeling the efficiency of top Arab banks: A DEA–neural network approach, Expert Systems with Applications, 36, 309-320. Pasiouras, F., Liadaki, A., & Zopounidis, C. (2008). Bank efficiency and share performance: Evidence from Greece. Applied Financial Economics, 18, 1121-1130. 16 Sealey, C., & Lindley, J.T. (1977). Inputs, outputs and a theory of production and cost at depository financial institution. Journal of Finance, 32, 1251-1266. Simar, L., & Wilson, P. W. (1998). Sensitivity Analysis of Efficiency Scores: How to Bootstrap in Nonparametric Frontier Models. Management Science, 44, 49-61. Simar, L.,& Wilson, P.W.( 1999). Estimating and bootstrapping Malmquist indices. European Journal of Operational Research, 115, 459-471. Simar, L., &Wilson, P.W. (2007). Estimation and inference in two stage, semi-parametric models of productive efficiency. Journal of Econometrics, 136, 31-64. Taylor, J.M. (2004). Understanding and Supporting Islamic Finance: Product differentiation and international standards. Forum on Islamic Finance, Harvard University. Turk-Ariss, R. (2009). Competitive behaviour in Middle East and North Africa banking systems. The Quarterly Review of Economics and Finance, forthcoming Williams, J., Peypoch, N., & Barros, C.P. (2009). The Luenberger indicator and productivity growth: A note on the European savings banks sector. Applied Economics (forthcoming). Table 1. Sharia-Compliant assets of banking sector (percentage of total banking sector assets) 17 1999 2000 2001 2002 2003 On balance sheet 11.2 13.0 15.0 17.0 21.0 Off balance sheet 1.7 2.1 2.3 2.0 2.3 Source: International Monetary Fund Table 2. Performance of Saudi Arabia Banking Sector Performance (%) Table 3. Saudi Arabia Banks 2000 2001 2002 2003 2004 2005 2006 2007 ROA 2.15 2.24 2.11 2.23 2.66 3.75 4.26 2.92 ROE 20.12 20.58 19.68 21.19 24.10 30.08 31.30 22.85 Net Profit Margin 50.43 51.44 50.12 50.07 58.17 66.77 69.41 61.79 18 Bank name Description of Bank’s Activity Al Jazira Bank Bank that provides Shari'ah-compliant financial services, including personal banking, business banking, investment banking & e-banking. Al Rajhi Banking & Investment corporation Bank formed by the takeover of Arab Bank Limited in 1980; offers commercial and Islamic banking products besides services in investment, mutual funds and assets management, local and international equity trading, foreign exchange and treasury Arab National Bank (ANB) Bank formed by the takeover of Arab Bank Limited in 1980; offers commercial and Islamic banking products besides services in investment, mutual funds and assets management, local and international equity trading, foreign exchange and treasury. Banque Saudi Fransi Saudi Arabian joint stock company affiliated with Calyon of France; it is a full service commercial bank that provides comprehensive financial services and products in Saudi Arabia and other markets; offers several Islamic banking products Riyad Bank Retail & corporate bank with a network of nearly 200 branches and 618 automated teller machines across the Kingdom. SABB (formerly Saudi British Bank) Commercial bank with a network of 73 branches across Saudi Arabia (13 exclusive ladies' sections/branches); associate company of the HSBC Group; head office is in Riyadh; activities: personal banking, investment banking, Islamic banking, treasury etc. Samba Financial Group Internet access service of Samba, one of the largest banks in the Middle East with 62 branches, 253 ATMs, three global investment centres. Saudi Hollandi Bank First bank in Saudi Arabia, originally set up to serve pilgrims from the Dutch East Indies (now Indonesia); today it is a Saudi joint stock company, 40% owned by ABN Amro; offers personal banking and corporate banking services and mutual funds. Saudi Investment Bank (SAIB) Commercial bank whose activities include personal banking, Islamic banking, corporate banking, investment banking, treasury & electronic banking; shareholders include J.P. Morgan Chase, Mizuho Corporate Bank & Saudi investors Table 4. Descriptive Statistics of the Data 19 Variable Mean St.Dev Median Min Max Total Employees 6584 5830 3642 672 19493 Fixed Assets 33960654963 1305400000 512640000 38930000 80553000000 Total Deposits 39614000000 24048000000 34980000000 2769111000 115810000000 Customer Loans 23082000000 19182000000 19162000000 226884000 80553000000 Securities 2365487062 1555812392 2000000000 10000000 7096219000 Interbank Loans 5532705704 4379740749 4427328000 157165000 17798000000 20 Year Bank Original Bootstrapped Bias Lower Bound Upper Bound Year Bank Original Bootstrapped Bias Lower Bound Upper Bound 1999 Riyad 1.0000 0.9364 0.0031 0.8517 0.9975 2000 Riyad 1.0000 0.9432 0.0025 0.8594 0.9966 1999 Jazira 1.0000 0.8368 0.0581 0.6749 0.9973 2000 Jazira 0.8685 0.8283 0.0019 0.7498 0.8660 1999 SAIB 1.0000 0.8298 0.0631 0.6599 0.9974 2000 SAIB 0.9971 0.9414 0.0037 0.8377 0.9944 1999 Hollandi 0.9413 0.8963 0.0012 0.8335 0.9384 2000 Hollandi 0.8624 0.8246 0.0005 0.7846 0.8602 1999 Fransi 1.0000 0.8367 0.0602 0.6609 0.9971 2000 Fransi 0.6931 0.6687 0.0001 0.6524 0.6909 1999 SAAB 1.0000 0.8371 0.0577 0.6621 0.9967 2000 SAAB 0.9837 0.9370 0.0011 0.8756 0.9808 1999 ANB 0.7156 0.6968 0.0001 0.6707 0.7140 2000 RAJHI 0.6267 0.6110 0.0001 0.5887 0.6253 1999 SAMBA 0.9232 0.8954 0.0002 0.8638 0.9203 2000 SAMBA 0.8835 0.8599 0.0001 0.8335 0.8810 1999 RAJHI 1.0000 0.9647 0.0006 0.9176 0.9979 2000 ANB 0.9725 0.9384 0.0008 0.8732 0.9699 Average 0.9533 0.8589 Average 0.8764 0.8392 Year Bank Original Bootstrapped Bias Lower Bound Upper Bound Year Bank Original Bootstrapped Bias Lower Bound Upper Bound 2001 Riyad 0.9861 0.9465 0.0014 0.8735 0.9837 2002 Riyad 0.9707 0.9418 0.0004 0.9019 0.9682 2001 Jazira 0.8401 0.8125 0.0005 0.7638 0.8384 2002 Jazira 0.7962 0.7775 0.0002 0.7460 0.7945 2001 SAIB 1.0000 0.9410 0.0020 0.8708 0.9970 2002 SAIB 0.9134 0.8593 0.0021 0.7865 0.9111 2001 Hollandi 0.9946 0.9513 0.0007 0.9070 0.9919 2002 Hollandi 0.7644 0.7398 0.0001 0.7183 0.7628 2001 Fransi 0.7058 0.6800 0.0002 0.6480 0.7041 2002 Fransi 0.7306 0.7080 0.0001 0.6896 0.7283 2001 SAAB 0.7218 0.6975 0.0001 0.6749 0.7199 2002 SAAB 0.7746 0.7506 0.0001 0.7296 0.7728 2001 ANB 0.5707 0.5540 0.0001 0.5327 0.5695 2002 RAJHI 0.5235 0.5108 6.5147 0.4956 0.5221 2001 SAMBA 0.8644 0.8460 0.0001 0.8232 0.8624 2002 SAMBA 0.8743 0.8565 0.0001 0.8346 0.8723 2001 RAJHI 0.9503 0.9082 0.0012 0.8372 0.9472 2002 ANB 1.0000 0.9114 0.0049 0.8354 0.9970 Average 0.8482 0.8152 Average 0.8164 0.7840 Year Bank Original Bootstrapped Bias Lower Bound Upper Bound Year Bank Original Bootstrapped Bias Lower Bound Upper Bound 2003 Riyad 0.9707 0.9386 0.0004 0.8934 0.9686 2004 Riyad 0.9729 0.9407 0.0003 0.9110 0.9701 2003 Jazira 0.7899 0.7730 8.4424 0.7563 0.7878 2004 Jazira 0.8217 0.7941 0.0002 0.7698 0.8198 2003 SAIB 1.0000 0.8862 0.0115 0.7868 0.9976 2004 SAIB 0.9479 0.8944 0.0019 0.8256 0.9457 2003 Hollandi 0.8902 0.8561 0.0003 0.8236 0.8875 2004 Hollandi 0.8920 0.8619 0.0002 0.8350 0.8890 2003 Fransi 0.7754 0.7511 0.0001 0.7330 0.7735 2004 Fransi 0.8485 0.8182 0.0002 0.7930 0.8460 2003 SAAB 0.8710 0.8465 0.0001 0.8256 0.8688 2004 SAAB 0.8635 0.8347 0.0001 0.8129 0.8606 2003 ANB 0.4723 0.4612 4.6606 0.4472 0.4707 2004 RAJHI 0.4198 0.4082 6.9305 0.3913 0.4188 2003 SAMBA 0.8536 0.8341 0.0001 0.8094 0.8517 2004 SAMBA 0.8936 0.8478 0.0016 0.7750 0.8905 Table 5 Bootstrapped Efficiency Results 21 2003 RAJHI 0.9530 0.9198 0.0004 0.8807 0.9503 2004 ANB 0.9641 0.9182 0.0011 0.8663 0.9613 Average 0.8418 0.8074 Average 0.8471 0.8131 Year Bank Original Bootstrapped Bias Lower Bound Upper Bound Year Bank Original Bootstrapped Bias Lower Bound Upper Bound 2005 Riyad 1.0000 0.9183 0.0035 0.8594 0.9972 2006 Riyad 0.9711 0.9269 0.0012 0.8651 0.9690 2005 Jazira 0.8207 0.7930 0.0002 0.7685 0.8188 2006 Jazira 1.0000 0.8844 0.0103 0.8067 0.9975 2005 SAIB 1.0000 0.9141 0.0060 0.8051 0.9973 2006 SAIB 1.0000 0.8821 0.0132 0.7895 0.9970 2005 Hollandi 1.0000 0.9494 0.0011 0.8914 0.9975 2006 Hollandi 1.0000 0.9332 0.0046 0.8149 0.9971 2005 Fransi 0.9829 0.9519 0.0004 0.9160 0.9810 2006 Fransi 1.0000 0.9270 0.0033 0.8498 0.9974 2005 SAAB 0.9845 0.9576 0.0003 0.9263 0.9816 2006 SAAB 1.0000 0.9481 0.0016 0.8841 0.9971 2005 ANB 0.6235 0.6003 0.0004 0.5613 0.6219 2006 RAJHI 0.9159 0.8823 0.0009 0.8211 0.9132 2005 SAMBA 0.9212 0.8861 0.0006 0.8414 0.9197 2006 SAMBA 0.9965 0.9464 0.0016 0.8750 0.9939 2005 RAJHI 0.9865 0.9397 0.0009 0.8926 0.9839 2006 ANB 0.9878 0.9338 0.0024 0.8485 0.9849 Average 0.9244 0.8789 Average 0.9857 0.9182 Year Bank Original Bootstrapped Bias Lower Bound Upper Bound 2007 Riyad 1.0000 0.8671 0.0215 0.7438 0.9974 2007 Jazira 0.9421 0.9011 0.0006 0.8629 0.9381 2007 SAIB 0.9999 0.9418 0.0037 0.8313 0.9976 2007 Hollandi 1.0000 0.9070 0.0077 0.7864 0.9973 2007 Fransi 0.9758 0.9008 0.0071 0.8196 0.9972 2007 SAAB 0.9985 0.9240 0.0035 0.8498 0.9972 2007 ANB 0.8992 0.8958 0.0084 0.8087 0.9978 2007 SAMBA 0.9321 0.8884 0.0098 0.7997 0.9268 2007 RAJHI 0.9452 0.8926 0.0088 0.8022 0.9376 Average 0.9659 0.9021 22 Table 6. Bootstrapped Truncated Regression Variable Coefficient t-statistic Constant -0.2612** 2.790 ln TA 0.0479** 3.139 0.0026* 2.166 Liq 0.3566** 3.822 PR - 0.0008** -2.337 Dummy -0.0303* 2.185 t 0.0020** 3.333 Variance 0.1400** 7.690 ** significant at the 5% confidence level; * significant at the 10% confidence level; total number of iterations=2000 NPM 
work_5vqe5iapgfam3egdt4zhjfgrdi	----	Keybook: Unbias object recognition using keywords Expert Systems with Applications 42 (2015) 3991–3999 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Keybook: Unbias object recognition using keywords http://dx.doi.org/10.1016/j.eswa.2015.01.019 0957-4174/� 2015 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. Tel.: +60 3 7967 6433. E-mail addresses: wailam88@siswa.um.edu.my (W.L. Hoo), ch_lim@siswa.um. edu.my (C.H. Lim), cs.chan@um.edu.my (C.S. Chan). Wai Lam Hoo, Chern Hong Lim, Chee Seng Chan ⇑ Centre of Image and Signal Processing, Faculty of Computer Science & Information Technology, University of Malaya, 50603 Kuala Lumpur, Malaysia a r t i c l e i n f o a b s t r a c t Article history: Available online 17 January 2015 Keywords: Dataset bias Object recognition Codebook generation Bag-of-Words model The presence of bias in existing object recognition datasets is now a well-known problem in the computer vision community. In this paper, we proposed an improved codebook representation in the Bag-of-Words (BoW) approach by generating Keybook. In specific, our Keybook is composed from the keywords that significantly represent the object classes. It is extracted utilizing the concept of mutual information. The intuition is to perform feature selection by maximize the mutual information of the features between the object classes; while minimize the mutual information of the features between the domains. With this, the Keybook will not bias to any of the domain and consists of valuable keywords among the object classes. The proposed method is tested on four public datasets to evaluate the classification performance in seen and unseen datasets. Experiment results have showed the effectiveness of our proposed methods in undo the dataset bias problem. � 2015 Elsevier Ltd. All rights reserved. 1. Introduction Fritz, & Darrell, 2010) (in this case, one dataset is treated as one Object recognition has been an active research since decades due to the demands in many real-world applications, for example object tracking (Choi & Christensen, 2012; Serratosa, Alquézar, & Amézquita, 2012), visual surveillance (Guo, Xia, & Xiaofei, 2014; Lim, Tang, & Chan, 2014; Szpak & Tapamo, 2011), human-robot interactions (Stasse, Foissotte, & Kheddar, 2008; Rudinac, Kootstra, Kragic, & Jonker, 2012; Bielecki, Buratowski, & Śmigielski, 2013), etc. Though successful attempts, Torralba and Efros (2011) raised an important question – ‘‘how well does a typ- ical object detector trained on one dataset generalize when tested on a representative set of other datasets, compared with its performances on the native test set?’’, or in other words, unbiased. Unfortunately, extensive experiments in Torralba and Efros (2011) found out that there is existence of various types of strong build-in bias (e.g selection bias, capture bias, and negative set bias) in popular object recognition datasets (e.g Caltech101, ImageNet, SUN09, LabelMe). These biases are embedded to these datasets uninten- tionally, during the data collection stage. Such situation is very critical as it causes one object representation and recognition algo- rithm will only work well in the specific dataset that one choose to use; and resulting in poor recognition rate when another dataset is selected. We visualize this problem in Fig. 1(a). This lead to some solutions (Duan, Xu, Tsang, & Luo, 2012; Kulis, Saenko, & Darrell, 2011; Li, Shi, Liu, Hauptmann, & Xiong, 2012; Saenko, Kulis, domain) where the main objective in these works is to achieve good recognition rate across all the datasets as depicted in Fig. 1(b). Khosla, Zhou, Malisiewicz, Efros, and Torralba (2012) proposed to learn a visual world SVM model to undo the bias factor. This approach learns the bias information for each dataset from a set of BoW representation in different datasets. However, there is room for improvement as the current approach is not feasible for the rapid dataset augmentation in the learning process. This is because their method needs to learn the bias weight for each of the unseen dataset to achieve better performance. That is to say, the approach will need to re-learn bias weight whenever there is a new set of images. To cope with this, our idea is to undo the bias during the codebook generation stage, and we name this as Key- book generation. From our investigation, biases exist because of the strong preference on certain information or features that cause the inclination towards a specific dataset. For an example, Fig. 2 shows a set of images of car class from two different datasets. Dataset 1 (Fig. 2(a)) has strong preference on cars with rectangle windows (representing in red boxes); while sports car with slanted front (representing in red boxes red boxes) is expected on dataset 2 (Fig. 2(b)) in order for an object to be recognized as a car. Literally, these features are not significant representation for car and such features are biased towards the corresponding training dataset. Thus, it deteriorates the recognition performance when another dataset is used in the testing stage. A better solution is to discover features for an object class that are generalized across all datasets. For example, the wheel (representing in green box) can be a choice for a car class as illustrated in Fig. 2. http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.01.019&domain=pdf http://dx.doi.org/10.1016/j.eswa.2015.01.019 mailto:wailam88@siswa.um.edu.my mailto:ch_lim@siswa.um.edu.my mailto:ch_lim@siswa.um.edu.my mailto:cs.chan@um.edu.my http://dx.doi.org/10.1016/j.eswa.2015.01.019 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. (a) The recognition accuracy using a classifier that train on a specific domain of car images and then test on all other domains. Generally, it is known that a high accuracy will be obtained when the training and testing images are originated from the same domain. However, the accuracy will drop drastically when other domains are used as the testing sets because of the existence of dataset bias problem as reported in Torralba and Efros (2011). (b) The main objective of this work is to generate a trained model that are able achieve consistent accuracy when tested across all the domains. Fig. 2. Good features vs bad features. Red boxes indicate features that are bias to a particular dataset, while green boxes indicate features that share across datasets. One can notice that windows in (a) and car front in (b) are feature that biased to respective datasets. Car tyres, instead, are the common features that exist in both datasets. This image is best view in color. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) 3992 W.L. Hoo et al. / Expert Systems with Applications 42 (2015) 3991–3999 To achieve this, we extend the work in Khosla et al. (2012) by proposing Keybook generation with the aim to discover the key- words from codebook that learned from the different datasets. In general, keywords are the visual features that significantly rep- resent the respective object classes. We utilize the mutual infor- mation (MI) to compute the semantic correlation of the visual features between the object to the classes and the datasets. According to Zitnick and Parikh (2013), MI is a good choice to dis- cover the subset of image specific information that is semantically meaningful. In this context, the MI is a measurement between the visual features (codeword in our case) to obtain the relative degree that they are required in object classes and datasets for semantic understanding. Then, by maximizing MI between the object classes and minimize MI between the datasets, the bias towards the datasets could be undo and keywords candidate are preserved to form the Keybook. Taking example in Fig. 2 with the use of the proposed Keybook idea, the green box features that are generalized over the object class will be retained, while the red box features that are bias towards a particular dataset will be discarded. The contribution of our work can be highlighted as follow. We propose Keybook generation to discover keywords from the con- ventional codebook model. The collected Keybook that is general- ized over the datasets will greatly enhance the object recognition performance. For this purpose, the MI is utilized to undo the bias between the objects to the dataset. This approach is in contrast with the state-of-the-art solution (Khosla et al., 2012) where our proposed are performed in the initial training stage, which allows the framework to learn the model in an unbias manner as early as in the codebook generation stage. Fig. 3. Overall framework for our proposed Keybook generation. Keybook is generated via preserving the significant codewords of each object class, and remove codewords that show strong preference towards a dataset. This is make possible using the mutual information. W.L. Hoo et al. / Expert Systems with Applications 42 (2015) 3991–3999 3993 The rest of the paper is organized as follows. Section 2 covers the related works in domain adaptation. Our proposed framework is detailed in Section 3. Sections 4 and 5 discuss the experiments settings and results using seen and unseen dataset. Finally, we con- clude our work in Section 6. 2. Related work Detecting and recognizing objects is one of the most important research areas in computer vision community till date. However, Khosla et al. (2012) showed that conventional object recognition methods performed poorly when cross testing datasets were employed. Since then, several solutions such as transfer learning, semi-supervised learning, and unsupervised learning had been applied in this context. To begin with, in transfer learning, the notable work was Prasath Elango, Tommasi, and Caputo (2012) whom studied learn- ing of visual perceptual tasks such as recognizing a place or an object, taking the advantage of what the peer system had learned beforehand. However, the work is unable to handle when the sys- tem is exposed to a new environment or to deal with similar object but in a different domain, as they did not account the bias in differ- ent domains. Instead, they only transfer the discriminative model on this problem to the second system. In semi-supervised approaches, one of the pioneer works is Daumé and Marcu (2006). The work modeled the domain adapta- tion problem as such there are huge labels from the source domain and scarce labels from the target domain. Then, the work formu- lated it in terms of a simple mixture model and applies in the context of conditional classification models and conditional lin- ear-chain sequence labeling models. As such, inference could be efficiently solved using the conditional expectation maximization algorithm. In a more advanced work, Blitzer, Crammer, Kulesza, Pereira, and Wortman (2007) proposed to explicitly model the inherent trade-off between training on a large but inaccurate source dataset, and a small but accurate target training set. Their approach achieve much lower target error compare to the standard empirical risk minimization method. Unfortunately in these aforesaid approaches, minimum amount of labels in target domain is still required. This is not a feasible approach in domain adaptation as the target dataset and label will not be available in the training process. The above issue leads to unsupervised domain adaptation solu- tions as such approaches do not require labeled target data. Bergamo and Torresani (2010) proposed transductive SVMs while Bruzzone and Marconcini (2010) iteratively relabeling the target domain. Though successful, these works suffered from extensive computational cost as their approach depend very much on tuning parameters during the SVM training stage. A resolution for this was proposed by Gong, Shi, Sha, and Grauman (2012) which inspired from the idea of geodesic flows (Gopalan, Li, & Chellappa, 2011) to derive intermediate subspaces that interpolate between the source and target domain. Also, Gong et al. (2012) contributed in eliminating the need of tuning many parameters needed in previ- ous works and achieve improvement in terms of computational complexity. In what constitute closer to our work is the bias learning in dif- ferent domains, for instance Khosla et al. (2012). Their idea was inspired by Torralba and Efros (2011) whom found out that there is existence of bias in popular object classification datasets. Khosla et al. (2012) exploited dataset bias during training stage, and learnt two sets of weights: (1) bias vectors associated with each individual dataset, and (2) visual world weights that are com- mon to all datasets, which are learned by undoing the associated bias from each dataset. Their approach outperforms the SVM clas- sification that does not account for the presence of the bias. How- ever, the requirement of prior knowledge on the unseen domain in the their work is impractical to the real world application. Despite the aforementioned methods achieve good accuracy, none of them considering undo the bias and collecting the signif- icant features at the early stage (e.g during codebook generation stage). Note that, one cannot always train on augmented dataset to achieve better recognition rate. Instead, the idea of Keybook is useful to achieve better classify the unseen datasets, as well as providing a more feasibility solution for the rapid dataset aug- mentation. From our study, there is a surge of interest recently in understanding the semantic meaning of the image depends on the presence of objects, their features, and their relations to Fig. 4. Keybook generation pipeline. First, objects from different d are obtained. Then, we build V d for each d. BoW is build based on all V d generated. Based on the BoW, we calculate MIc and MId based on class label c and dataset label d. Finally, by maximizing MIc and minimizing MId , Keybook, V key is generated. 3994 W.L. Hoo et al. / Expert Systems with Applications 42 (2015) 3991–3999 the scene properties (Zitnick & Parikh, 2013). The quantitative measurement to study the semantic importance of various scene properties and features is using the mutual information (MI). Inspired from this, in our work, similar concept was employed to understand the importance of the presence of features towards class and dataset. For instance, if MI between a feature and a class is large, it indicates that the feature provides important informa- tion about an object class. On the other hand, if MI between a fea- ture to a specific dataset is high, this show that the particular feature may be bias to that dataset. From this, one can undo the bias by maximize MI between features and object class labels; while minimize MI between features and dataset labels. With this, the significant differences between our proposed solution and those conventional solutions are firstly, our proposed solu- tion do not require label from the target domain during the train- ing step. This is in contrast to the semi-supervised methods. Secondly, our proposed method undo the dataset bias during the codebook generation stage. This is in contrast to the solution in Khosla et al. (2012). 3. Proposed framework In the section, we first detail how a visual codebook is gener- ated as illustrated in Fig. 3. Then, we elaborate our proposed method, Keybook generation that uses the MI computation. Finally, we provide the explanation on the classification method that based on Khosla et al. (2012). 3.1. Codebook generation BoW model which is initially being employed in Natural Lan- guage Processing (NLP) has now established its contribution in the field of computer vision (Fei-Fei & Perona, 2005; Niebles, Wang, & Fei-Fei, 2008; Yang, Jin, Sukthankar, & Jurie, 2008), partic- ularly in the area of object recognition. In general, BoW is a sparse vector of occurrence count of codewords, that is a sparse histogram over the vocabulary. In NLP, the vocabulary is a database of words such as the popular WordNet (Miller, 1995) that covered almost every words available and is online. However in computer vision, yet we had own the vocabulary of all the objects in the world, the best representation of an object is still in the discovering pro- gress. There is still a milestone before the codebook generation which is a vital step in the BoW approach to achieve that. To begin with the codebook generation, we represent an image I ¼fxi; yc; ydg, where x are the extracted image features using fea- ture extraction algorithms (for instance SIFT, DSIFT, HOG, etc.), and each x is associated with a class label yc and domain label yd . A codebook V ¼fw1; . . . ; wkg is obtained by finding the cluster cen- ters w from xi , extracted from training images Itrain. Conventionally, we build V d from the training set of domain d. Then, each I is represented by the BoW model based on the obtained V via a quantization process: BoWðwÞ¼ 1 n Xn i¼1 1; if w ¼ arg min w2V ðDðw; xiÞÞ 0; otherwise ( ð1Þ Table 1 Experiments on seen datasets (i.e the train set of the dataset used for testing is W.L. Hoo et al. / Expert Systems with Applications 42 (2015) 3991–3999 3995 where n is the number of features in I, and Dðw; raÞ is the distance between w and feature xi . available during training), where all four datasets are used in training and only one of the datasets is used for testing at one time. PAS, LAB, CAL, SUN refer to the four datasets, Pascal VOC2007, LabelMe, Caltech101 and SUN09, respectively. We compare both results (i.e without V_{key} and with V_{key}), where the better results are in bold. Overall, our proposed (i.e with V_{key}) outperforms in 19 out of 25 cases. The underlined results indicate that these implementations are similar to conventional classification implementation (i.e train and test on same dataset). Train Test W PAS W LAB W Cal W SUN W vw Without V key All PAS 16.89 27.19 33.40 28.19 25.70 All LAB 51.01 67.82 65.68 69.13 68.24 All CAL 5.75 37.86 96.68 25.92 59.60 All SUN 16.90 25.46 27.68 50.97 40.31 Average 22.63 39.68 55.86 43.55 48.46 With V key All PAS 17.36 27.46 32.67 27.96 25.81 All LAB 52.53 67.96 65.16 68.72 67.64 All CAL 8.37 41.96 96.99 26.81 62.59 All SUN 18.70 26.58 28.53 51.43 41.76 Average 24.24 40.99 55.84 43.73 49.45 3.2. Keybook generation Unfortunately, conventional codebook generation induce domain bias problem, as illustrated in Fig. 2. This motivates us to propose Keybook that consists of keywords that are significant rep- resentation of the object classes. The principle of the Keybook is to discover the underlay unique features for an object class that are shared across all domains. On the other hand, we reject those fea- tures that cause bias in each specific domain. In order to achieve that, we obtain Keybook V key from a set of V d from different d by utilizing the computation of MI. In general, MI is useful in finding the correlation between two random variables A and B: MIðA; BÞ¼ X a2A X b2B pða; bÞ log pða; bÞ pðaÞpðbÞ � � : ð2Þ In our paper, A would be the BoW that is generated from a set of V d , and B would the labels, either class label c or dataset label d. In the proposed framework, V key contains image features that share among c across d, and discard those image features that bias to d. To build V key, we chose keyword w� 2 V key discovered from all V d (we denote as V all), by select the w that is maximizing the mutual information among the object classes, MIc , and minimizing the mutual information of domains, MId : w� ¼ arg max w2V all ðMIc � MIdÞ; ð3Þ where MIc ¼ MIðBoWðwÞ; ycÞ and MId ¼ MIðBoWðwÞ; ydÞ. Then, we build a new BoW representations based on the V key, and learn a classifier from these BoW for the object recognition task. Fig. 4 shows a step-by-step process on how we obtain the Keybook. Fig. 5. Sample images of five different classes in Ca 3.3. Classification Since our proposed framework is almost similar to the conven- tional codebook generation method, we adopt the visual world SVM model as proposed in Khosla et al. (2012) for the classification stage. The visual world weight wvw and bias vector Md is taken into account and learn as below for different d: min wvw;Md;�;/ 1 2 kwvwk 2 þ k 2 Xn d¼1 kMdk 2 þ C1 Xn d¼1 Xsd j¼1 �dj þ C2 Xn d¼1 Xsd j¼1 /dj ð4Þ subject to wd ¼ wvw þMd ð5Þ yij wvw � x i j P 1 �� i j; i ¼ 1 . . . n; j ¼ 1 . . . sd ð6Þ yij wi � x i j P 1 �� i j; i ¼ 1 . . . n; j ¼ 1 . . . sd ð7Þ � P 0; /ij P 0; i ¼ 1 . . . n; j ¼ 1 . . . sd ð8Þ ltech101, LabelMe, SUN09 and Pascal VOC07. 3996 W.L. Hoo et al. / Expert Systems with Applications 42 (2015) 3991–3999 Eqs. (5)–(8) define the constraints for the visual world SVM model, while C1 and C2 are the hyperparameters that employed to bal- ance the learning terms in the objective function. This is in order to control the relative importance of the constraints. k is used to define the weight between independent weights of different data- sets, and a common weight for all datasets. � represents losses across datasets when using the visual world weights wvw, which need to be minimized because wvw is expected to generalize across all datasets. / represents the losses when test images are wrongly classified by the biased weights. Further details can be found in Khosla et al. (2012). We summarize the proposed framework in Algorithm 1. Algorithm 1. Unbiased framework by learning Keybook (V key) and visual world SVM model Require: A set of training images with fxi; yc; ydg. Ensure: All parameters are set: total number of codewords for each dataset codebook V d, total number of codewords for Keybook V key. 1. Learn V d for each dataset d using Eq. (1). 2. Find w� for V key using mutual information (MI) as described in Eqs. (2) and (3). 3. Learn visual world SVM model using V key by optimizing Eq. (4) based on constraints in Eqs. (5)–(8). 4. Experiments We tested our proposed framework on four different datasets that share the common object classes, but have huge variation in terms of orientation, viewpoint and environment, namely: Cal- tech101 (CAL) (Fei-Fei, Fergus, & Perona, 2004), LabelMe (LAB) (Russell, Torralba, Murphy, & Freeman, 2008), SUN09 (SUN) (Choi, Lim, Torralba, & Willsky, 2010), and Pascal VOC07 (VOC) (Everingham, Van Gool, Williams, Winn & Zisserman). The five object classes are ‘bird’, ‘car’, ‘chair’, ‘dog’, and ‘person’ and some of the example images are shown in Fig. 5. Fig. 6. The graphs show improvement in percent average precision (%AP) of classification et al. (2012)) in 4 dataset. The labels on the x-axis ‘P’, ‘L’, ‘C’, and ‘S’ represent the datasets the mean AP (mAP) increment over all datasets. Y-axis is the %AP increment over the base 25 cases, with an overall improvement of 5% mAP. 4.1. Implementation details We employed the experimental settings as to Khosla et al. (2012), except that we sample SIFT features on all training set using patch size = 8 and step size = 8 for simplicity and efficiency. We build V d and V key that consist of 256 codewords and keywords respectively to ensure a fair comparison. Then, we use LLC coding (Wang et al., 2010) with 3 level spatial pyramid to pool SIFT fea- tures based on V key. The evaluations are tested on both seen and unseen datasets. 4.2. Testing on seen dataset Table 1 shows the experimental results when we trained the model using training data from all four datasets, and test on one dataset at a time. Since the training data consist of images from the same dataset for the testing data, we consider this as seen data- set classification task. Table 1 (top: Without V_{key}) shows the baseline approach based on visual world SVM model (Khosla et al., 2012) without V key implementation, while Table 1 (bottom: With V_{key}) shows our proposed method that combines both benefits from V key and the visual world SVM model. In general, the visual world SVM model that use V key shows around 1% of mean average precision (mAP) improvement com- pared to the model that use conventional codebook. This shows that V key enhance the model for cross-domain classification task. This is intuitively possible as Keybook attempts to find a more gen- eral representation of the objects among the same classes. There- fore it discards visual codewords that are biased to a particular dataset. Takes note that, the usage of more simpler feature space in feature representation will cause some variations to the results published in Khosla et al. (2012). However, this is a fair comparison as both methods used the same features during testing. Besides that, the underlined results indicate results using the conventional SVM model, where wd is used to classify d, instead of wvw . Compar- ing both results using the conventional SVM model, our proposed on unseen datasets using the proposed method over the baseline approach (Khosla Pascal VOC2007, LabelMe, Caltech101 and SUN09 respectively, while ‘M’ represents line Khosla et al. (2012). Overall, our algorithm outperforms the baseline in 20 out of Fig. 7. Qualitative analysis on experiments that classify unseen datasets. Blue region indicates correct classification, and red means false classification. For (a) and (b), first row: Pascal VOC07, second row: SUN09, third row: LabelMe, fourth row: Caltech101. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) W.L. Hoo et al. / Expert Systems with Applications 42 (2015) 3991–3999 3997 method achieved 0.35 mAP improvement on all four datasets com- pared to those methods without V key. 4.3. Testing on unseen dataset We tested our proposed method on unseen dataset, which indi- cates that for each testing iteration, we only select three datasets instead of four for training, and perform testing on the dataset that was leave out. For example, if we train on Caltech101, LabelMe and SUN09 datasets, then we will test on PASCAL VOC07 dataset. In here, the proposed method employs wvw during the classification stage. Fig. 6 shows most of the average precision (AP) for unseen dataset is increased, except for Caltech101 dataset. This is because from Khosla et al. (2012), Caltech101 has the best AP improvement, while the Keybook generation obtained a more generalized model across the dataset. This generalized model might had slightly affected the performance in Caltech101. This also implies the bot- tleneck of the unbias algorithm, which further improve classifica- tion result would be hard, by comparable results could be maintained. In spite of that, we have significant improvement over 5% mAP in the overall performance. 5. Discussion Fig. 7 shows testing images that are correctly classified and miss-classified in our proposed method for unseen dataset. From our visual inspection, one can notice that the correct classified images are those cars that the car wheels are visible. For the miss-classified images, the car wheels are either occluded or miss- ing. This reflects our claim in the proposed method, where the car wheels could be the keywords found in the V key, and we gain improve classification result in the car class when the car wheels 3998 W.L. Hoo et al. / Expert Systems with Applications 42 (2015) 3991–3999 are visible. In contrast, our model will work poorly if the keywords (e.g. wheels) are occluded or missing. Besides, with the help of V key, the proposed method could correct images that wrongly classified in the system that without V key. Note that this observation is some- how not applicable for Caltech101 samples. This is maybe due to the characteristics of Caltech101 samples posses strong capture bias problem, as indicated in Torralba and Efros (2011). 6. Conclusion Domain adaptation has risen as an important issue in object recognition research and the recent trend towards the solution has been focused on the study of the domain bias. Failing to handle the domain bias problem will deteriorate the overall object recog- nition system performance, especially when the trained model is used in different unseen domains. This is due to the trained model is tailored to learn for a specific domain. In this paper, we proposed a framework that undo the domain bias thru finding the most sig- nificant features of an object class which are generalized across all the domains. Specifically, we employed mutual information (MI) computation to analyse the correlation between the visual feature of the object classes and the datasets; and generate the Keybook by keyword selection. The proposed approach was tested on four pub- licly available datasets and achieved better performance in either seen or unseen dataset evaluation. Although our proposed method achieves better results in the overall performance, it is still bounded with several limitations. First, the recognition is very much depends on the significant features of the object in an image, and so an occlusion on the significant feature will deteriorate the overall performance. In con- junction, deciding the number of codewords for the Keybook remains a key issue that decides the performance of the final classification. The proposed framework will be very useful in application such as visual surveillance, ambient assisted living, and robotic vision that requires object recognition. Such systems that are employed to assist or protect human in their daily live are required to be able to perform consistently in different environments. With our pro- posed solution, since the significant features of the object class are retained, so it is able to cope with environments change. There are a few potential future works to enhance the current approach. First, the discriminative characteristic (i.e hard assign- ment) in the current work can be relaxed with the integration the soft assignment codebook representation (Hoo, Kim, Pei, & Chan, 2014). Secondly, alternatives MI that might be more useful in Keybook during the generation of the dataset codebook could be investigated. Other possible research directions are performing the unbias attempt in the feature level (i.e when feature extraction from images occur), and dataset collection level (i.e how researcher could collect bias-free dataset). Finally, the work could be extended to space–time domain such as human motion analysis (Lim, Vats, & Chan, in press); or to fine-grained classification problem (Yao, Khosla, & Fei-Fei, 2011). Acknowledgments This research is supported by the High Impact Research MoE Grant UM.C/625/1/HIR/MoE/FCSIT/08, H-22001-00-B0008 from the Ministry of Education Malaysia. References Bergamo, A., & Torresani, L. (2010). Exploiting weakly-labeled web images to improve object classification: A domain adaptation approach. In Advances in neural information processing systems (NIPS) (pp. 181–189). Bielecki, A., Buratowski, T., & Śmigielski, P. (2013). Recognition of two-dimensional representation of urban environment for autonomous flying agents. Expert Systems with Applications, 40, 3623–3633. Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Wortman, J. (2007). Learning bounds for domain adaptation. In Advances in neural information processing systems (NIPS) (Vol. 20, pp. 129–136). Bruzzone, L., & Marconcini, M. (2010). Domain adaptation problems: A dasvm classification technique and a circular validation strategy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32, 770–787. Choi, C., & Christensen, H. I. (2012). 3d textureless object detection and tracking: An edge-based approach. In IEEE/RSJ international conference on intelligent robots and systems (IROS) (pp. 3877–3884). Choi, M. J., Lim, J. J., Torralba, A., & Willsky, A. S. (2010). Exploiting hierarchical context on a large database of object categories. In IEEE conference on computer vision and pattern recognition (CVPR) (pp. 129–136). Daumé, H., III, & Marcu, D. (2006). Domain adaptation for statistical classifiers. Journal of Artificial Intelligence Research, 26, 101–126. Duan, L., Xu, D., Tsang, I. W.-H., & Luo, J. (2012). Visual event recognition in videos by learning from web data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34, 1667–1680. Everingham, M., Van Gool, L., Williams, C. K. I., Winn, J., & Zisserman, A. The PASCAL visual object classes challenge 2007 (VOC2007) results. <http://www.pascal- network.org/challenges/VOC/voc2007/workshop/index.html>. Fei-Fei, L., & Perona, P. (2005). A bayesian hierarchical model for learning natural scene categories. In IEEE conference on computer vision and pattern recognition (CVPR) (Vol. 2, pp. 524–531). Fei-Fei, L., Fergus, R., & Perona, P. (2004). Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE computer vision and pattern recognition workshop (CVPRW) (pp. 178–178). Gong, B., Shi, Y., Sha, F., & Grauman, K. (2012). Geodesic flow kernel for unsupervised domain adaptation. In IEEE conference on computer vision and pattern recognition (CVPR) (pp. 2066–2073). Gopalan, R., Li, R., & Chellappa, R. (2011). Domain adaptation for object recognition: An unsupervised approach. In IEEE international conference on computer vision (ICCV) (pp. 999–1006). Guo, W., Xia, X., & Xiaofei, W. (2014). A remote sensing ship recognition method based on dynamic probability generative model. Expert Systems with Applications, 41, 6446–6458. Hoo, W. L., Kim, T., Pei, Y., & Chan, C. S. (2014). Enhanced random forest with image/ patch-level learning for image understanding. In 22nd international conference on pattern recognition, ICPR (pp. 3434–3439). Khosla, A., Zhou, T., Malisiewicz, T., Efros, A., & Torralba, A. (2012). Undoing the damage of dataset bias. In European conference on computer vision (ECCV) (pp. 158–171). Kulis, B., Saenko, K., & Darrell, T. (2011). What you saw is not what you get: Domain adaptation using asymmetric kernel transforms. In IEEE conference on computer vision and pattern recognition (CVPR) (pp. 1785–1792). Lim, C. H., Vats, E., Chan, C. S. (in press). Fuzzy human motion analysis: A review. Pattern Recognition. http://dx.doi.org/10.1016/j.patcog.2014.11.016. Lim, M. K., Tang, S., & Chan, C. S. (2014). iSurveillance: Intelligent framework for multiple events detection in surveillance videos. Expert Systems with Applications, 41, 4704–4715. Li, H., Shi, Y., Liu, Y., Hauptmann, A. G., & Xiong, Z. (2012). Cross-domain video concept detection: A joint discriminative and generative active learning approach. Expert Systems with Applications, 39, 12220–12228. Miller, G. A. (1995). Wordnet: A lexical database for english. Communications of the ACM, 38, 39–41. Niebles, J. C., Wang, H., & Fei-Fei, L. (2008). Unsupervised learning of human action categories using spatial–temporal words. International Journal of Computer Vision, 79, 299–318. Prasath Elango, S., Tommasi, T., & Caputo, B. (2012). Transfer learning of visual concepts across robots: A discriminative approach. Idiap-RR Idiap-RR-06-2012 Idiap. Rudinac, M., Kootstra, G., Kragic, D., & Jonker, P. P. (2012). Learning and recognition of objects inspired by early cognition. In IEEE/RSJ international conference on intelligent robots and systems (IROS) (pp. 4177–4184). Russell, B. C., Torralba, A., Murphy, K. P., & Freeman, W. T. (2008). Labelme: A database and web-based tool for image annotation. International Journal of Computer Vision, 77, 157–173. Saenko, K., Kulis, B., Fritz, M., Darrell, T. (2010). Adapting visual category models to new domains. In European conference on computer vision (ECCV) (pp. 213–226). Serratosa, F., Alquézar, R., & Amézquita, N. (2012). A probabilistic integrated object recognition and tracking framework. Expert Systems with Applications, 39, 7302–7318. Stasse, O., Foissotte, T., & Kheddar, A. (2008). Treasure hunting for humanoids robot. In IEEE RAS/RSJ international conference on humanoids robots, workshop on cognitive humanoid vision. Szpak, Z. L., & Tapamo, J. R. (2011). Maritime surveillance: Tracking ships inside a dynamic background using a fast level-set. Expert Systems with Applications, 38, 6669–6680. Torralba, A., & Efros, A. A. (2011). Unbiased look at dataset bias. In IEEE conference on computer vision and pattern recognition (CVPR) (pp. 1521–1528). http://refhub.elsevier.com/S0957-4174(15)00034-2/h0010 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0010 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0010 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0020 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0020 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0020 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0035 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0035 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0040 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0040 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0040 http://www.pascal-network.org/challenges/VOC/voc2007/workshop/index.html http://www.pascal-network.org/challenges/VOC/voc2007/workshop/index.html http://refhub.elsevier.com/S0957-4174(15)00034-2/h0070 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0070 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0070 http://dx.doi.org/10.1016/j.patcog.2014.11.016 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0095 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0095 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0095 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0100 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0100 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0100 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0105 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0105 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0110 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0110 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0110 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0125 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0125 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0125 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0135 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0135 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0135 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0145 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0145 http://refhub.elsevier.com/S0957-4174(15)00034-2/h0145 W.L. Hoo et al. / Expert Systems with Applications 42 (2015) 3991–3999 3999 Wang, J., Yang, J., Yu, K., Lv, F., Huang, T., & Gong, Y. (2010). Locality-constrained linear coding for image classification. In IEEE conference on computer vision and pattern recognition (CVPR) (pp. 3360–3367). Yang, L., Jin, R., Sukthankar, R., Jurie, F. (2008). Unifying discriminative visual codebook generation with classifier training for object category recognition. In IEEE conference on computer vision and pattern recognition (CVPR) (pp. 1–8). Yao, B., Khosla, A., & Fei-Fei, L. (2011). Combining randomization and discrimination for fine-grained image categorization. In IEEE conference on computer vision and pattern recognition (CVPR) (pp. 1577–1584). Zitnick, C. L., & Parikh, D. (2013). Bringing semantics into focus using visual abstraction. In IEEE conference on computer vision and pattern recognition (CVPR) (pp. 3009–3016). Keybook: Unbias object recognition using keywords 1 Introduction 2 Related work 3 Proposed framework 3.1 Codebook generation 3.2 Keybook generation 3.3 Classification 4 Experiments 4.1 Implementation details 4.2 Testing on seen dataset 4.3 Testing on unseen dataset 5 Discussion 6 Conclusion Acknowledgments References 
work_5rzbrjfjkvdyhidl7nbgqjrpi4	----	Earthquake Prediction Using Expert Systems: A Systematic Mapping Study sustainability Review Earthquake Prediction Using Expert Systems: A Systematic Mapping Study Rabia Tehseen, Muhammad Shoaib Farooq * and Adnan Abid Department of Computer Science, University of Management and Technology, Lahore 54770, Pakistan; f2017288003@umt.edu.pk (R.T.); adnan.abid@umt.edu.pk (A.A.) * Correspondence: Shoaib.farooq@umt.edu.pk Received: 24 December 2019; Accepted: 15 March 2020; Published: 19 March 2020 ���������� ������� Abstract: Earthquake is one of the most hazardous natural calamity. Many algorithms have been proposed for earthquake prediction using expert systems (ES). We aim to identify and compare methods, models, frameworks, and tools used to forecast earthquakes using different parameters. We have conducted a systematic mapping study based upon 70 systematically selected high quality peer reviewed research articles involving ES for earthquake prediction, published between January 2010 and January 2020.To the best of our knowledge, there is no recent study that provides a comprehensive survey of this research area. The analysis shows that most of the proposed models have attempted long term predictions about time, intensity, and location of future earthquakes. The article discusses different variants of rule-based, fuzzy, and machine learning based expert systems for earthquake prediction. Moreover, the discussion covers regional and global seismic data sets used, tools employed, to predict earth quake for different geographical regions. Bibliometric and meta-information based analysis has been performed by classifying the articles according to research type, empirical type, approach, target area, and system specific parameters. Lastly, it also presents a taxonomy of earthquake prediction approaches, and research evolution during the last decade. Keywords: Expert systems; Systematic Mapping Study (SMS), earthquake prediction; seismic data; Early-warning systems 1. Introduction Earthquakes have been one of the most hazardous but least predictable natural disaster [1,2]. The occurrence of catastrophic earthquakes results in casualties, massive damage to the infrastructure, the vanquishing of societies in a flash and a sudden downfall in the country’s economy [3,4]. There are many geographical factors that may cause an earthquake, including ground motion, heavy rainfall, rock bed material, regional tectonics and altitude [5]. There is a tremendous pressure on geologists and seismologists for the prediction of the time, place and strength of earthquakes [6]. Many researchers have claimed to predict earthquakes by observing multiple precursors such as recording the behavior of animals, observing an increase in temperature, emission of radon gas, and observing the change in seismicity patterns of the region, etc. References [7,8]. However, it is very hard to generalize and standardize these prediction algorithms as the precursors do not necessarily appear before every earthquake [9]. Earthquake prediction is a highly complicated task and many investigators have used different approaches for making forecasts. Among these different approaches the methods and algorithms that are based on a variety of expert systems (ES)have exhibited promising results in this area. The literature survey reveals that different approaches of expert systems including fuzzy, rule-based, neuro-fuzzyand machine/deep learning methods have been used to forecast future earthquake from historic and instrumental data. In practice, ES have also been efficiently used for risk analysis and assessment in Sustainability 2020, 12, 2420; doi:10.3390/su12062420 www.mdpi.com/journal/sustainability http://www.mdpi.com/journal/sustainability http://www.mdpi.com https://orcid.org/0000-0002-4095-8868 http://www.mdpi.com/2071-1050/12/6/2420?type=check_update&version=1 http://dx.doi.org/10.3390/su12062420 http://www.mdpi.com/journal/sustainability Sustainability 2020, 12, 2420 2 of 32 multiple areas such as, information technology (IT) [10], engineering [11], economics, healthcare [12], and civil engineering [13–15]. The motivation behind applying expert system technique for earthquake prediction lies in its noticeable effectiveness and reliability [16] of such approaches in other disciplines. This research presents a systematic mapping study to facilitate the researchers and practitioners in understanding the fundamentals of earthquake prediction systems; evolution of research in this area, and prominent research directions in this area of research. To these ends, it presents an analysis of eighty four articles while presenting a classification scheme showing multiple aspects covered in the literature addressing earthquake prediction. The study presents a summary of various aspects of the expert systems given in multiple articles for earthquake prediction. It also determines the most frequently used variants of ES for the prediction of earthquake while highlighting the accuracy in prediction results claimed in multiple articles. It is pertinent to mention that to the best of our knowledge, no mapping study has been found about using ES for earthquake prediction. Furthermore, this mapping study not only covers the bibliometric aspects of the selected research articles, but also presents a taxonomy of approaches, variants of ES, their strengths and comparative analysis of their effectiveness in prediction accuracy, and the widely used seismic data sets. Lastly, the open research areas and future directions in this area have also been presented to the researchers working in this area. There are certain limitations of this survey. These limitations relate to the selection of primary studies [17]. In order to ensure that as many relevant publications as possible have been included, we have identified search terms in several iterations. Terms related to ES and earthquake prediction were used in the search string. However, the list might not have been complete, and additional or alternative terms might have altered the final list of papers found [18]. The search was performed by using the Elsevier Scopus Digital Library. According to the statistics of the publications retrieved, we believe that most of the research on earthquake can be found in this electronic library. However, certain papers may have been overlooked due to the subscription limitations. Another threat is related to the handling of duplications, which might have slightly changed our results. Kappa measure has been used for making decisions about possible duplications. The data has been extracted from the primary studies and classified to generate the final results. The decision about which data to collect and how to classify the papers therefore depended on the judgement of the authors conducting the systematic mapping study [19]. The Kappa coefficient has resulted in 0.95 which indicated an agreement among the authors about data inclusion. Data extraction from prose could also result in a misclassification, but this problem was addressed by developing a classification scheme on the basis of widely accepted guidelines [17] and terminology proposed for use in [20]. It would, therefore, only have a minor influence on the classification scheme developed in this mapping study. Validity limitation refers to the missing studies, incorrect data extraction [21] and determining the incorrect relationship among multiple facets. To overcome this threat, we have clearly described the activities involved in publication selection and data extraction in multiple sections. The traceability between the data extracted and the conclusion drawn has been presented through bubble plots and frequency plots. The quality assessment problem is concerned with the quality of study selection [5,22,23]. The systematic mapping results were considered in regard to the seismic domain, and the validity of the conclusions drawn concerns the earthquake prediction context only. To focus on state of the art methods, we have applied time restriction in searching for published studies and have included the papers published during January 2010 till January 2020. The search string and the classification scheme presented in this paper may serve as a starting point for researchers working on the problem of earthquake prediction, and they can search for and categorize additional papers accordingly. The rest of the article has been structured in the following manner. Section 2 presents the background of this study, while the research methodology of this study has been presented in Section 3. The analysis of all the selected and reviewed articles has been presented in Section 4. The analysis about the findings of the literature review has been discussed in Section 5. Lastly, the article has been concluded in Section 6. Sustainability 2020, 12, 2420 3 of 32 2. Background In this section, we have presented the general outline of the research work about using expert system for earthquake prediction. Multiple approaches, including rule-based, fuzzy, neuro-fuzzy and machine / deep learning methods have been used for earthquake prediction. 2.1. Fuzzy Expert System (FES) The concepts of fuzzy set and fuzzy logic were introduced in 1965 by [20]. Fuzzy expert system accepts input as crisp variables and converts it into fuzzy variables. Fuzzy inference engine applies the rules suggested by an expert to formulate the knowledge base. Fuzzy variables combined with linguistic variables to generate membership functions. Techniques based on Fuzzy logic have the benefits over multiple procedures due to their ability to combine with linguistic variables. These fuzzy variables would be converted back into the crisp variable to generate output through the process called defuzzification. Fuzzy logic is more suitable in the situations where a greater number of uncertainties have been involved, such as, earthquake prediction and in the scenarios where an approximate but quick solution is required. Fuzzy logic is not a logic that is fuzzy itself, but a logic that can be used to demonstrate fuzziness [21]. The vagueness of fuzzy logic has been highlighted in [10] by examining the events that cannot be recorded statistically such as crack in the undergroundfault, etc. A new attenuation relationship has been proposed in [8] using three fuzzy input sets including epicentral distance, earthquake magnitude and intensity using earthquake data set of Taiwan and United states of America (USA). A normalized fuzzy ground motion model has been demonstrated using a rational design tool through a combination of natural language with seismic data statistics to quantify response frequency. The earthquake pattern in the Zagros range has been examined in [9] using fuzzy rule-based ES model for some earthquakes. The proposed model has been evaluated using the Molchan statistical procedure by comparing complicated reasoning procedure of the forecasting model with knowledge simulation provided by human experts using the datasets of Iran. A rock burst forecasting model has been presented in [13] by studying the seismic features of coal mining in China. In this study, Gaussian shaped membership function has been combined with the exponential distribution function using reliability theory. The comprehensive forecasting result was obtained by integrating the maximum membership degree principle (MMDP) and the variable fuzzy pattern recognition (VFPR) method. The performance of the proposed model has been evaluated using seismic data collected over the period of four months. The proposed model has been able to forecast the rock burst incident in the coal mine of China. Multiple algorithms have been combined for development of the hybrid prediction model [24,25]. Ionospheric disturbance has been examined in [26] and a fuzzy logic-based gradient descent method has been proposed to forecast the ionospheric change parameters. The gradient descent estimated values were used to tune the membership function. The satisfactory performance has been observed during evaluated of the proposed model using data collected from two geomagnetic storms on the low latitude. Reference [1] has claimedearthquake prediction on the bases of the classification of seismicsignals. 2.2. Rule Based Expert System (RBES) In RBES domain knowledge is represented by a set of rules and the current situation is presented with the set of facts stored in the database. An inference engine is responsible to match the rule with the fact. The fired rule may change the set of facts and add new facts. Many researchers have used rule based expert system for earthquake prediction. A belief rule based expert system has been presented in [27] to predict the earthquake under uncertainty. Specific animal behavior in response to environmental and chemical changes has been examined for earthquake prediction. Reference [20] developed rules from historical earthquake data using predicate logic. These rules have been mathematically validated on real time data. Prediction is performed through RBES that takes current earthquake attributes for prediction of future earthquake. Sustainability 2020, 12, 2420 4 of 32 2.3. Neuro Fuzzy Expert System (NFES) Fuzzy logic is combined with neural networks to develop expert systems. Fuzzy logic provided a high level reasoning procedure by including domain information from the domain expert and neural network has been used to develop low level computational structures. The Neuro fuzzy expert system has been used in many articles to analyze multiple aspects of data for earthquake predictions. Reference [28] combined grid partition, subtractive clustering and fuzzy C-means (FCM) for the development of models using NFES structure. Reference [5] applied NFES to compute land sliding susceptibility using statistical index (WI). Reference [22] collected geographical information to pass through six different membership functions for measuring land sliding susceptibility using NFES. Many researchers have analysed combination of artificial neural network and fuzzy inference system [29–32]. Earthquake attribute such as magnitude, depth, longitude and latitude has been studied in [33] to provide input to NFES for computation of the future earthquake. 2.4. Machine Learning (Ml) Machine Learning has been widely used for making earthquake predictions due to their ability to improve over time. With the huge amount of earthquake instrumental data, machine learning approaches are capable enough to improve efficiency and accuracy in earthquake prediction. Multiple machine learning methods including, Artificial Neural Network (ANN), Support Vector machine (SVM), K-nearest neighbour (KNN), Native Bayes (NB) and random forest algorithms have been exercised for earthquake prediction. Reference [34] applied Artificial Neural Networks, Support Vector Machines and Random Forests to perform temporal investigations on earthquake catalogue of Cyprus region and calculated sixty seismic indictors for making short term earthquake prediction. Reference [35] applied different machine learning algorithms namely support vector machine (SVM), K-nearest neighbor (KNN), random forest (RF), and Naïve Bayes (NB) algorithms in R programming language for earthquake prediction using seismic dataset of India. Reference [36] studied the thermal anomalies that happened before the earthquake occurred in Imphal, India, in 2016 and investigated multiple seismic facts through satellite data using machine learning algorithms for an earthquake. Reference [37] collected records of aftershocks of the Kermanshah (Iran) Earthquake and applied different machine learning (ML) algorithms, including Naive Bayes, k-nearest neighbors, a support vector machine, and random forests to predict future earthquakes by observing aftershock patterns. Reference [38] exercised neural networks for earthquake signal detection. Reference [39] listed the detailed description of the monitoring techniques used for earthquake prediction. References [40,41] presented a comprehensive review of machine learning methods used for earthquake prediction. Reference [42] made seismic hazards forecasts by using two different machine learning based methods for both spatial and space-time prediction of strong earthquakes. Reference [43] determined the significance of shallow land slide triggers in making earthquake forecasts using machine learning methods. Reference [44] improved the conventional waveform correlation method and presented a new method for detection of seismic signals for monitoring the false alarms using machine learning. Reference [45] identified, classified and reviewed the prominent machine/deep learning models used in energy systems. Reference [46] discussed multiple artificial intelligent models utilized for hydrologic model prediction in past decade. Reference [47] highlighted the opportunities and challenges presented by big data for informed decision-making. Reference [48] developed a food forecast model using multiple optimization methods. 3. Research Methodology The objective of systematic mapping study is to present an overview of the research area and quantify the results presented by the selected studies. We intend to determine the research trends by mapping the frequency of publications over time. For this purpose, we have adopted the methodologies Sustainability 2020, 12, 2420 5 of 32 of [45,46] and performed a number of activities as shown in the Figure 1. Our main goal is to provide deep inside of expert system based solutions proposed for earthquake prediction in literature. Sustainability 2020, 11, x FOR PEER REVIEW 5 of 33 Figure 1. Research Methodology of Systematic Mapping Study (SMS). 3. Research Methodology The objective of systematic mapping study is to present an overview of the research area and quantify the results presented by the selected studies. We intend to determine the research trends by mapping the frequency of publications over time. For this purpose, we have adopted the methodologies of [45,46] and performed a number of activities as shown in the Figure 1. Our main goal is to provide deep inside of expert system based solutions proposed for earthquake prediction in literature. 3.1. Defining Research Questions We have addressed three research questions in this mapping study. These research questions have served as a guide for the classification of research articles. A set of research questions has been described in Table 1. Table 1. Set of research questions (RQs) and their motivation. Research Question (RQ) RQ Statement Motivation RQ 1 What are the bibliometric key facts of expert systems (ES) based earthquake prediction publications? RQ 1.1 How many studies have been contributed from January 2010 to January 2020? The intentions of this research question is to find out the number of publications that have been contributed in the selected time period and the main venues where the studies have been published. RQ 1.2 What are the venues where these studies have been published? RQ 2 Which research type facets do the identified publications address? RQ 2.1 What is the type of research conducted in the publication? The main intention is to categorize the selected publications through the schema established by [17,49]. Therefore, we use the research type facets given by Zhang et al. [49]. Based on these type facets, we wanted to find out multiple research contexts, including the type of the research, empirical type of the research, approaches used in the research and areas targeted by the researchers for data extraction. RQ 2.2 What is the empirical type of the research conducted in the publication? Figure 1. Research Methodology of Systematic Mapping Study (SMS). 3.1. Defining Research Questions We have addressed three research questions in this mapping study. These research questions have served as a guide for the classification of research articles. A set of research questions has been described in Table 1. Table 1. Set of research questions (RQs) and their motivation. Research Question (RQ) RQ Statement Motivation RQ 1 What are the bibliometric key facts of expert systems (ES) based earthquake prediction publications? RQ 1.1 How many studies have been contributed from January 2010 to January 2020? The intentions of this research question is to find out the number of publications that have been contributed in the selected time period and the main venues where the studies have been published.RQ 1.2 What are the venues where these studies have been published? RQ 2 Which research type facets do the identified publications address? RQ 2.1 What is the type of research conducted in the publication? The main intention is to categorize the selected publications through the schema established by [17,49]. Therefore, we use the research type facets given by Zhang et al. [49]. Based on these type facets, we wanted to find out multiple research contexts, including the type of the research, empirical type of the research, approaches used in the research and areas targeted by the researchers for data extraction. RQ 2.2 What is the empirical type of the research conducted in the publication? RQ 2.3 What approach has been used by the researcher? RQ 2.4 Which area has been targeted by the research for data collection? RQ 3 What is the type and other key aspects of proposed Expert System (ES) in the classified publications? RQ 3.1 What type of expert system has been proposed in the selected studies? The main aim is to determine the types of proposed ES used for earthquake prediction in the articles published during January 2010 till January 2020. This question is helpful in highlighting the other parameters of the proposed ES like input domain, number of input attributes passed, type of the input attributes, prediction logic, the tools and techniques used in the articles have been categorized. RQ 3.2 Which input domain does the proposed ES address? RQ 3.3 How many input attributes are passed to the proposed ES? RQ 3.4 What is the type of the input attributes passed to the proposed ES? RQ 3.5 Which type of prediction logic has been used by the proposed ES? RQ 3.6 Which tool or technique has been used to develop the proposed ES? 3.2. Search and Selection Strategy After defining research questions, next activity was to select the sources from where the articles would be retrieved. For this purpose, we have adopted the searching strategy given in [47,48] and Sustainability 2020, 12, 2420 6 of 32 have developed a comprehensive search string based on the key terms given in Table 2. Articles have been collected from Elsevier Scopus digital library (www.scopus.com).The terms stated in Table 2 have been used to develop the search string for searching articles from the given sources. Table 2. Distinct key terms used in developing the search strings. AND Terms OR Terms Earthquake Rule based, Fuzzy, Frame based Machine Learning, Deep learning, Expert system Seismic, Tremor Indicator Precursor, Feature Prediction Predict* (* means wildcard) (“Rule based” OR “*Fuzzy” OR “Frame based” OR “neural” OR “machine learning” OR “deep learning” OR “Expert system”) AND (“Earthquake” OR “Seismic” OR “Tremor”) AND (“Predict*”) AND (“Indicat*” OR “Precursor” OR “features”) 3.2.1. Identification of Search String The primary keywords were selected as key identifiers of work in the field of earthquake prediction using expert system. A set of distinct keywords is listed in Table 2. This mapping study has been conducted to examine the literature about earthquake prediction using expert systems. We have used multiple key terms like fuzzy, frame based and rule based methods for extracting such articles that describe the use of expert systems for earthquake prediction, but does not necessarily have an expert system explicitly written in their titles. In the same way, earthquake has been presented as a seismic event or a tremor in some studies, so we have also included these keywords. The change in the behaviour of precursors has been studied in many articles for earthquake prediction. Therefore, we have also included a few keywords describing the indicators of future earthquakes. Search string has been developed after defining key terms. The following search string has been used to search articles from Elsevier Scopus digital library. We have included the articles published from January 2010 till January 2020. 3.2.2. Screening and Selection Criteria Screening of the articles has been performed after retrieval. The main purpose is to select most relevant articles. Every paper was retrieved and evaluated by considering title, abstract, keywords, introduction and conclusion. The inclusion and exclusion criteria given in Table 3 have been used for article selection. We have included only those articles that satisfy the inclusion criteria given in Table 3. We have excluded those papers which have been retrieved from multiple sources or representing different stages of the same project. For inclusion of the articles that have identical abstracts we have calculated Kappa coefficient. The papers that are not written in the English language have also not been included. The thesis has also been excluded because they normally cover multiple aspects of the problem. Papers with unclear methodology and not satisfying our quality criteria have also been excluded. Table 3. Inclusion and Exclusion Criteria. Inclusion Criteria (IC) Criteria Description IC1 Articles in which an expert system has been developed for earthquake prediction IC2 Articles in which earthquake precursors have been analyzed IC3 Articles presenting unique and new ideas IC4 Literature published as book chapter and technical reports for earthquake prediction IC5 Articles with identical abstracts (on the basis of Kappa coefficient) Exclusion Criteria (EC) EC1 Duplicates and identical titles EC2 Papers not in English language EC3 Thesis (cover several different aspects) EC4 Papers with unclear methodology EC5 Papers not satisfying quality criteria www.scopus.com Sustainability 2020, 12, 2420 7 of 32 We have taken a good care of article quality selection in the article selection process. To ensure high quality in article selection, criteria given in [19] have been adopted and elaborated in Table 4. Table 4. Quality Criteria. Quality Ranking Sr. Criteria Type Weight a. Study Presents contribution Yes 1 No 0 Partially 0.5 b. Study presents solution Yes 1 No 0 Partially 0.5 c. The study presents empirically validated results Yes 1 No 0 Partially 0.5 Search Process resulted in the retrieval of 2137 articles that passed through multiple phases. In the first phase, our search string has retrieved 2137 research articles. After passing through multiple screening phases of selection criteria seventy articles have been considered original, non-duplicate, with clear methodology and satisfying our quality criteria. The coefficient has been calculated to determine the relevance among the articles. In every screening phase, two authors of this mapping study have been asked to make judgement about the relevance of the article by selecting any choice from “accept”, “reject” or “differ” options. In case of difference in the opinion of both judges, comprehensive discussion has been carried out to a decision point has been reached in the form of acceptance or rejection. 3.3. Data Extraction and Synthesis It is based on providing a set of answers to the research questions. Table 5 represents the data that we intend to extract by asking research questions prescribed in Table 1. Table 5. Data extracted through each research question. Research Questions (RQs) Data Extracted RQ 1 RQ 1.1 Number of publications contributed in the given time period has been determined. RQ 1.2 A main venue where the study has been published has been noted. RQ 2 RQ 2.1 Research type (solution, evaluation, experience) has been determined. RQ 2.2 Empirical type (Experiment, survey, case study) has been determined. RQ 2.3 The approach used (model, method, guideline, framework, tool) has been noted. RQ 2.4 Seismic zone (global, regional) focused by the study has been determined. RQ 3 RQ 3.1 Type of the proposed expert system (Fuzzy expert system, rule based expert system, Neuro fuzzy expert system) has been noted. RQ 3.2 Identification of the input domain i.e., quake or precursive RQ 3.3 Number of input attributes, i.e., single or multiple that have been passed to the proposed ES for earthquake prediction. RQ 3.4 Type of the input attributes (numeric or discrete) has been determined. RQ 3.5 Prediction logic (inductive or deductive) used by the proposed expert system has been noted. RQ 3.6 Tools and techniques used to develop the proposed expert system given in the studies have been categorized. RQ1 extracts bibliometric facts, including a number of publications contributed in the period of ten years from January 2010 till January 2020 and the main venue where the publication has been submitted e.g., journal, conference, book chapter, etc. RQ1 highlights the trend of the researchers regarding article submission in the last ten years. RQ2 deals with various research type facets of the articles, including research type, empirical type, approach and the targeted area. Type of the research determined that the article has presented a novel solution or illustrated an extension of already existing technique. Evaluation research examined the Sustainability 2020, 12, 2420 8 of 32 techniques used in the articles have passed through an evaluation process before its implementation. Experience papers presented the personal experiences of the author explaining how something has been done in practice. Empirical type illustrated that an experiment, survey, or case study has been performed in the selected article. We have also collected the information regarding focused seismic zone through RQ2. Detailed type facets of the proposed ES for prediction of earthquake have been listed in the classification Table 6. We have provided a set of distinct keywords in Figure 2 to explain the contents of Table 6. The relationship between number of publications and the type facets determined through RQ2 has been shown in Figure 3. Sustainability 2020, 11, x FOR PEER REVIEW 9 of 33 RQ 3 RQ 3.1 Type of the proposed expert system (Fuzzy expert system, rule based expert system, Neuro fuzzy expert system) has been noted. RQ 3.2 Identification of the input domain i.e quake or precursive RQ 3.3 Number of input attributes, i.e. single or multiple that have been passed to the proposed ES for earthquake prediction. RQ 3.4 Type of the input attributes (numeric or discrete) has been determined. RQ 3.5 Prediction logic (inductive or deductive) used by the proposed expert system has been noted. RQ 3.6 Tools and techniques used to develop the proposed expert system given in the studies have been categorized. Classification of Articles We have classified the articles to determine multiple parameters summarized in Table 6 presenting the bibliographic values, research type facets, type of the expert system used and other key aspects of the ES proposed in these articles for earthquake prediction. Quality criteria defined in Table 4 have also been applied to categorize the articles accordingly. Classification Table 6 had been developed on the basis of research questions given in Table 1 and the quality criteria prescribed in Table 4. 4. Analysis The results obtained from the research questions (as given in Table 1) have been presented in Table 6. All articles have been selected to illustrate their relevance and contribution in the earthquake prediction process by providing an answer to every research question. 4.1. Basic Analysis After deep investigationeighty four articles have been selected and main approaches for earthquake prediction including machine learning, neuro-fuzzy, fuzzy and rule-based approaches have been reviewed. These articles were thoroughly analyzed to answer the research questions (RQs) given in Table 1. We have summarized the results obtained from RQs in Table 6. A set of distinct keywords to explain the contents of Table 6 has been given in Figure 2. Figure 2. Set of distinct keywords used to populate classification Table 6. Figure 2. Set of distinct keywords used to populate classification Table 6. Sustainability 2020, 12, x FOR PEER REVIEW 5 of 33 Last type facet of this mapping study is the target area that has been classified as regional and global. 75% of the included researches have been focused to a certain specific regional area where as 24% of the researchers have used global data to forecast earthquake. Figure 3 shows the number of publications (n) according to the type facets given in RQ2. Figure 3. Relationship between number of publications (n) and the type facets. 4.4. ES type and System Specific Key Aspects of Proposed ES Three types of the proposed ES have been identified by RQ3 and presented in Figure 4. Fuzzy expert system (FES) has been proposed for making earthquake predictions by 29% of the studies. Neuro-Fuzzy expert system for making an earthquake prediction has been proposed by 11% of the studies. The rule based expert system has been proposed by 7% of the included studies. 36% of the articles have used machine learning for earthquake prediction and 12% of the studies have used other computational approaches like index development or classifier calculation for earthquake prediction. In this mapping study RQ3 also deals with exploring other key aspects of the proposed ESs like input domain, input attributes and their types, data type, prediction logic and the prediction technique covering the use of computational supporting software or algorithm for performing earthquake prediction task. Input domain can be the earthquake parameters like magnitude, b-value, etc. or precursors calculated from real time earthquake data. In this mapping study, 51% researches have directly consumed earthquake parameters (listed as quake variables in Table 6) whereas 38% have studied precursors that serve as indicators for an upcoming devastating earthquake event. Further, we have explored the type of above stated input attributes to analyze that whether predictions have been made by consuming single input attribute or combination of multiple attributes has been used for earthquake predictions. In this mapping study, 27%, researchers have made predictions on the basis of a single attribute where as 68% studies have used a combination of multiple input attributes for earthquake predictions. Through data type we have worked to find that input variables passed to the proposed ES are quantitative type (numeric form like magnitude, slope, etc.) or qualitative type (discrete form like water elevation, emission of radon gas etc.). It has been observed that 57% of researchers have used numeric inputs for earthquake forecast whereas 43% of the studies have performed the earthquake prediction task using a discrete type of input. This mapping study has also captured the data regarding prediction logic and prediction methodology. Through prediction logic, we have determined that the article have used the inductive or deductive Figure 3. Relationship between number of publications (n) and the type facets. Sustainability 2020, 12, 2420 9 of 32 Table 6. Classification Table. Ref. Bibliometric Facts Type Facets System Specific Information Quality Ranking Publication Channel Publication Year Research Type Empirical Type Approach Target Area Proposed ES type Input Domain Input Attribute Input Attribute type Data Type Prediction Logic Tools and Techniques (a) (b) (c) Score [1] Journal 2014 Eva CS Mod RL FES PR ML DV Dis IN SW 1 0.5 0.5 2.0 [2] Journal 2017 Eva Ext Met RL FES QE SL DV Num IN AM 1 0.5 1 2.5 [3] Journal 2017 Eva Ext Mod RL Other QE ML DV Num DD SW 1 0.5 1 2.5 [5] Journal 2016 Eva CS Met RL NFES PR SL DV Dis IN AM 1 0.5 0.5 2.0 [6] Journal 2018 Eva Ext Met RL FES QE ML DV Num IN SW 1 0.5 1 2.5 [7] Journal 2015 Eva Ext Mod RL FES QE ML PE Num IN AM 1 0.5 1 2.5 [8] Confe 2012 Sol Ext Met RL FES PR ML DV Dis IN Oth 0.5 1 1 2.5 [9] Journal 2014 Eva Ext Mod RL FES PR ML DV Dis IN SW 1 0.5 1 2.5 [10] Journal 2017 Eva CS Met RL Other QE ML DV Num IN SW 1 0.5 1 2.5 [11] Journal 2017 Eva CS Mod RL FES PR SL DV Dis IN SW 1 0.5 0.5 2.0 [12] Journal 2017 Eva CS Mod RL FES QE ML DV Num IN SW 1 0.5 0.5 2.0 [13] Journal 2018 Sol Ext Mod RL FES QE SL DV Num IN SW 0.5 1 1 2.5 [14] Book 2017 Eva CS Met RL FES PR SL DV Dis IN SW 1 0.5 0.5 2.0 [15] Journal 2017 Exp Sur Met RL FES QE ML PE Dis IN Oth 0.5 0.5 0 1.0 [16] Journal 2013 Exp Sur Mod GL RBES PR SL PE Dis DD SW 0.5 0.5 0 1.0 [17] Journal 2013 Eva Ext Mod RL RBES PR ML PE Dis IN Oth 1 0.5 1 2.5 [18] Journal 2016 Eva Sur Met RL FES QE ML PE Num DD AM 1 0.5 0 1.5 [19] Confe 2015 Eva Ext Mod GL RBES QE ML DV Dis IN SW 1 0.5 1 2.5 [20] Journal 2014 Eva Ext Mod GL RBES QE ML PE Dis IN SW 1 0.5 1 2.5 [21] Journal 2018 Eva Ext Met GL FES QE ML DV Dis DD AM 1 0.5 1 2.5 [22] Journal 2018 Eva Ext Met RL NFES QE ML DV Dis IN SW 1 0.5 1 2.5 [23] Journal 2015 Exp CS Met RL FES QE SL PE Dis IN Oth 0.5 0.5 0.5 1.5 [24] Confe 2016 Exp CS Met RL FES QE ML DV Dis IN Oth 0.5 0.5 0.5 1.5 [25] Confe 2010 Eva Ext Met RL Other PR SL PE Num DD AM 1 0.5 1 2.5 [26] Journal 2017 Sol CS Mod RL FES PR SL PE Dis IN SW 0.5 1 0.5 2.0 [27] Journal 2018 Eva Ext Mod GL RBES PR SL PE Dis IN AM 1 0.5 1 2.5 [28] Journal 2012 Sol Sur Mod GL FES PR ML PE Num IN AM 0.5 1 0 1.5 [30] Journal 2015 Exp Sur Gle GL NFES PR ML PE Num IN SW 0.5 0.5 0 1.0 [31] Journal 2015 Eva Ext Mod RL FES QE ML PE Num IN AM 1 0.5 1 2.5 [32] Journal 2018 Eva CS Mod RL NFES PR SL DV Dis IN SW 1 0.5 0.5 2.0 [33] Journal 2014 Exp Sur Gle GL NFES QE ML DV Num DD SW 0.5 0.5 0 1.0 Sustainability 2020, 12, 2420 10 of 32 Table 6. Cont. Ref. Bibliometric Facts Type Facets System Specific Information Quality Ranking Publication Channel Publication Year Research Type Empirical Type Approach Target Area Proposed ES type Input Domain Input Attribute Input Attribute type Data Type Prediction Logic Tools and Techniques (a) (b) (c) Score [34] Journal 2020 Eva Ext FW RL Ml QE ML PE Num DD AM 1 0.5 1 2.5 [35] Confe 2020 Eva Ext Mod RL Ml QE ML DV Num DD AM 1 0.5 1 2.5 [36] Journal 2020 Exp CS Met RL Ml PR SL DV Dis DD SW 0.5 0.5 0 1 [37] Journal 2019 Exp Ext Met RL Ml QE ML PE Num IN AM 0.5 0.5 1 2 [38] Confe 2019 Exp Ext Met GL Ml PR SL PE Dis DD SW 0.5 0.5 1 2 [39] Confe 2019 Sol Ext Mod GL Ml QE ML PE Num DD AM 0.5 1 1 2.5 [40] Confe 2019 Exp Sur Gle GL Ml PR ML DV Dis IN AM 0.5 0.5 0 1 [41] Confe 2019 Exp Sur Gle GL Ml PR ML PE Dis DD AM 0.5 0.5 0 1 [42] Journal 2019 Eva Ext Met RL Ml QE ML DV Num DD AM 1 0.5 1 2.5 [43] Confe 2019 Eva Ext Mod GL Ml PR ML DV Dis DD AM 1 0.5 1 2.5 [44] Journal 2018 Sol Ext Met GL Ml QE ML DV Num IN AM 0.5 1 1 2.5 [49] Journal 2019 Eva Ext Met RL NFES PR ML DV Num DD AM 1 0.5 1 2.5 [50] Journal 2013 Exp Ext Met RL FES PR ML PE Num IN Oth 0.5 0.5 1 2.0 [51] Confe 2010 Sol CS Mod RL FES PR SL DV Dis DD AM 0.5 1 0.5 2.0 [52] Journal 2018 Eva Ext Mod RL FES PR ML PE Dis IN Oth 1 0.5 1 2.5 [53] Journal 2011 Sol Sur Mod RL Other QE SL DV Num DD AM 0.5 1 0 1.5 [54] Journal 2019 Exp CS Gle RL Other PR SL PE Num DD AM 0.5 0.5 0.5 1.5 [55] Confe 2010 Sol Ext Met RL FES PR SL DV Dis IN Oth 0.5 1 1 2.5 [56] Confe 2018 Exp Ext FW GL NN QE ML DV Num DD AM 0.5 0.5 1 2 [57] Confe 2018 Exp Sur Met GL Ml PR ML PE Num IN AM 0.5 0.5 0 1 [58] Journal 2018 Exp Ext Met RL NN QE ML PE Num DD AM 0.5 0.5 1 2 [59] Journal 2018 Sol CS Met RL Ml QE ML PE Num DD SW 0.5 1 0.5 2 [60] Journal 2018 Eva Ext Met GL Ml QE ML DV Num IN SW 1 0.5 1 2.5 [61] Confe 2018 Sol Ext Met RL Ml QE ML PE Num DD SW 0.5 1 1 2.5 [62] Journal 2017 Eva CS Mod RL Ml QE ML PE Num DD SW 1 0.5 0.5 2 [63] Confe 2017 Exp Ext Mod GL Ml PR SL PE Dis IN AW 0.5 0.5 1 2 [64] Journal 2017 Sol CS Mod RL NN PR ML DV Num IN SW 0.5 1 0.5 2 [65] Journal 2017 Eva CS Met RL Ml QE ML DV Num DD AW 1 0.5 0.5 2 [66] Journal 2017 Sol CS Met RL Ml PR SL PE Dis IN SW 0.5 1 0.5 2 [67] Journal 2017 Eva CS Met RL NN PR ML DV Num DD AW 1 0.5 0.5 2 Sustainability 2020, 12, 2420 11 of 32 Table 6. Cont. Ref. Bibliometric Facts Type Facets System Specific Information Quality Ranking Publication Channel Publication Year Research Type Empirical Type Approach Target Area Proposed ES type Input Domain Input Attribute Input Attribute type Data Type Prediction Logic Tools and Techniques (a) (b) (c) Score [68] Confe 2017 Eva CS Mod RL Ml QE ML PE Num IN SW 1 0.5 0.5 2 [69] Confe 2015 Eva CS Met RL Ml QE ML DV Num IN SW 1 0.5 0.5 2 [70] Journal 2015 Eva Ext Mod GL Ml QE ML PE Num IN SW 1 0.5 1 2.5 [71] Journal 2013 Eva CS Mod RL Ml PR ML DV Dis DD SW 1 0.5 0.5 2 [72] Journal 2013 Eva CS Met RL Ml QE ML DV Num DD SW 1 0.5 0.5 2 [73] Journal 2012 Exp CS Met RL Ml PR SL DV Dis DD SW 0.5 0.5 0.5 1.5 [74] Journal 2016 Sol Ext Met GL Ml QE ML DV Num IN SW 0.5 1 1 2.5 Notes: Publication Channel: Confe = Conference; Research Type: Eva: Evaluation, Exp: experience, Sol: Solution; Emperical Type: CS: Case study, Ext: Experiment, Sur: Survey; Approach: Mod: Model, Met: Method, Gle: Guideline, FW: Framework; Target Area: RL: Regional, GL: Global; Proposed ES type: Ml: Machine learning, FES: Fuzzy expert system, NFES: Nuero fuzzy expert system, RBES: Rulebased expert system; Input Attribute: ML: Multiple, SL: Single; Input Attribute Type: PE: Primitive, DV:Derived; Data Type: Num: Numeric, Dis: Discrete; Input domain: QE: Quake, PR: Precursive; Prediction Logic: IN: Inductive, DD:Deductive; Tools and Techniques: SW:Software, Alg:Algorithm. Sustainability 2020, 12, 2420 12 of 32 We have collected information about the type of ES such as FES, RBES and NFES proposed for earthquake prediction in the selected articles through RQ3. The key aspects of the proposed ES including input domain, number of input attributes, type of input attributes, data type, prediction logic, tools and techniques used for prediction of earthquake have been discussed in Figure 4. Input domain describes the input variables taken by the system to predict an earthquake. It can be either ‘Quake variables’ like latitude, longitude, magnitude, primary wave (P-Wave) attributes, etc. or ‘Precursors variables’ like ionosphere readings, earth’s thermal variations via satellites etc. Input attributes can be single or multiple. Some systems predict using single quake or precursor variable. Many articles used multiple input variables for earthquake prediction. Input attributes type can be primitive or derived. Some techniques directly consume an input parameter and others manipulate (derive) before use. The proposed ES uses variables in primitive form or transforms it into some other form before consuming it. The data type can be numerical or discrete. Prediction logic can be inductive or deductive. Inductive logic has no absolute proof from premises to conclusions, e.g., fuzzy sets. Deductive prediction logic determined an absolute proof from premises to conclusions Sustainability 2020, 12, x FOR PEER REVIEW 6 of 33 logic for making earthquake predictions. Inductive logic has been used by 62% of the studies, whereas deductive logic has been presented in 32% of the included studies for giving absolute results. Next we have worked to determine the predictive methodology to describethe use of any software or an algorithm for evaluation of the proposed system. 42% of the studies have used multiple algorithms for evaluation of their predictions, 48% have used the MATLAB software for evaluation of the proposed ES and 10% have adopted other evaluation mechanisms. Figure 4. Relationship between number of studies and system specific parameters of ES. 4.5. Quality Assessment The quality assessment score obtained by each paper has been presented in Table 6. These articles have been collected in three distinct categories and presented in Table 11 according to the scores obtained by every article given Table 6. By observing the quality scores, 74% of the papers included in the mapping study had obtained above average score, 11% of the studies had obtained average score and 15% of the studies remained below average. This quality assessment may facilitate researchers in the selection of articles for their work. Table 11. Scores. Reference Score Average =1.5 Total % [1–14,17,19–22,25–27,31,32,34,35,37–39,42–44,49–52,55,56,58– 72,74] Above average 55 79 [18,23,24,28,53,54,73] Average 7 10 [15,16,30,33,36,40,41,57] Below average 8 11 5. Discussion In this mapping study, main techniques used for earthquake prediction including rule-based, fuzzy, neuro-fuzzy and machine learning have been explored. Expert systems based approaches have been used for seismic risk assessment for landslide susceptibility through seismic hazard analysis [50–52] and soil classification [53]. ES have also been applied in earthquake engineering for seismic hazard analysis and assessment of bridges and buildings under multiple hazards [24,54,55]. The expert systems have been used to analyze multiple aspects of earthquake prediction, but due to the non-existence of the time-dependent global earthquake forecasting model, the regional earthquake likelihood models have been popular [87]. On the other hand, some initial research has been conducted to design earthquake early-warning systems that work globally [32,49]. 28 46 27 39 48 51 34 13 38 58 45 36 33 40 13 10 25 5 ty pe FE S/ N FE S/ RB ES / M L/ O th er In pu t D om ai n (Q ua ck / Pr ec ur si ve ) In pu t a tt rib ut e (s in gl e / m ul tip le ) in pu t a tt rib ut e ty pe (P rim iti ve /D er iv e d) D at a ty pe (n um er ic / de sc re te ) Pr ed ic tio nl og ic (in du ct iv e/ de du ct iv e) To ol s & te ch ni qu e (A lg o/ SW /o th er ) RQ3 Figure 4. Relationship between number of studies and system specific parameters of ES. Classification of Articles We have classified the articles to determine multiple parameters summarized in Table 6 presenting the bibliographic values, research type facets, type of the expert system used and other key aspects of the ES proposed in these articles for earthquake prediction. Quality criteria defined in Table 4 have also been applied to categorize the articles accordingly. Classification Table 6 had been developed on the basis of research questions given in Table 1 and the quality criteria prescribed in Table 4. 4. Analysis The results obtained from the research questions (as given in Table 1) have been presented in Table 6. All articles have been selected to illustrate their relevance and contribution in the earthquake prediction process by providing an answer to every research question. 4.1. Basic Analysis After deep investigationeighty four articles have been selected and main approaches for earthquake prediction including machine learning, neuro-fuzzy, fuzzy and rule-based approaches have been reviewed. Sustainability 2020, 12, 2420 13 of 32 These articles were thoroughly analyzed to answer the research questions (RQs) given in Table 1. We have summarized the results obtained from RQs in Table 6. A set of distinct keywords to explain the contents of Table 6 has been given in Figure 2. Table 6 describes the list of the selected articles and classifies them by extracting bibliometric facts, type facets, system specific information and quality ranking. Quality scores obtained by every article have been summed up to find the total score. According to Table 6, the average scores obtained by the articles are 1.5. The publications having scored greater than 1.5 are above average and the publications having scores below 1.5 are considered below average. Detailed description of classified results after application of quality criteria has been given in Table 6. Table 7 illustrates the trend of researcher regarding article submission in multiple sources. It lists down multiple sources, publication channels, and the frequency of articles in each source. Table 7. Article submission trend of the researchers. Source Channel Reference International Conference on Natural Computation (ICNC) Conference [75] Pure and Applied Geophysic Journal [76] Expert Systems with Applications Journal [1,9] IEEEACCESS Journal [49] International Journal of Computer Applications Journal [77] Proceedings of Indian National Science Academy Journal [78] Earth Science Informatics Journal [65,79] Journal of Indian Society of Remote Sensing Journal [80] Bulletin of Engineering Geology and Environment Journal [23,60,81,82] Natural Hazards Journal [2,4,12,22,28,51,83] Knowledge Based Systems Journal [17,19] Journal of Environmental Radioactivity Journal [54] International Journal of Coal Geology Journal [53] Computer-Aided Civil and Infrastructure Engineering Journal [15] Applied Sciences Journal [10] International Journal of Disaster Risk Reduction Journal [11] Tunnelling and Underground Space Technology Journal [13] PLoS ONE Journal [18,58] Environmental Earth Sciences Journal [5,52,84–86] International Journal of Fuzzy Systems Journal [32] Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications Journal [27] Applied Soft Computing Journal [87] Journal of Intelligent Information Systems Journal [20] Geodesy and Geodynamics Journal [26] Environmental Monitoring Assessment Journal [21] Earth Science Informatics Journal [65,79] International Journal of Computer Information Systems and Industrial Management Applications Journal [30] Biostatistics and Biometrics Journal [6] International Journal of Engineering Research & Technology Journal [24] Journal Geological Society of India Journal [50] Journal of Sustainability Science and Management Journal [16] Journal of Chemical and Pharmaceutical Sciences Journal [33] Acta Geophysica Journal [3] International Conference on Fuzzy Systems and Knowledge Discovery (FSKD) Conference [25,55,88] International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems Journal [7,71] International Conference on Information Management, Innovation Management and Industrial Engineering Conference [51] Analysis & Computation Specialty Conference Conference [8] Soil Dynamics and earthquake engineering Journal [34,64,69] Lecture notes on electrical engineering Conference [35] Advances in intelligent system and computing Journal [36] ISPR- International Journal of geo information Journal [37] Seismological Research Letter Conference [38,40] Geophysical Research Letter Conference [39,63] CEUR workshop procedings Conference [41] Geosciences Journal [42,59] Proceedings of SPIE-the international society of optical engineering Conference [43] Bulletin of seismological society of America Journal [44] Neural processing letter Conference [56] Proceedings-IEEE 4th International conference on big data, computing services and applications Conference [57] Lecture notes on computer science Conference [61] Geomagnatics, Natural hazards and risks Journal [62] International Journal of SWARM intelligence research Journal [66] Neural computing and application Journal [67] Proceedings- 14th international conference on frontiers of information technology Conference [68] Proceedings- 9th international conference on application of information and commucation technology Conference [70] Bollettino deGesfisica Teorica ed applicata Journal [71] Applied soft computing journal Journal [73] Journal of King Saud University Journal [74] Sustainability 2020, 12, 2420 14 of 32 Two dimensional studies (as described in Table 8) have been conducted to highlight seismically active zones of the world. In the first dimension, researchers have evaluated the data of specific regions, whereas in the second dimension, researchers have analyzed overall earthquake data of the world in their articles. Table 8 presented the zones (regional or global) with their geographical dimension and their location on the tectonic plates. Table 8 will facilitate the researchers in determining that which tectonic zone is being mostly explored and analyzed by the practitioners. Table 8. Target area representing the Geographic dimensions. Target Area Geographic dimension Ref. Worldwide Global [7,16,19–21,27,30,33,38–41,43,44, 56,57,60,63,70,74,77,79] Japan North American plate+Pacific plate+Pphilipine sea plate [61,85] California North American plate+Pacific plate [10,31,42,58,59,64,87] China Eurasian plate+Indian plate+Philipine sea plate [2,12,13,23,25,49,51,55,75,88] Taiwan Eurasian plate + Philipine sea plate [8] Pakistan Eurasian plate + Indian plate [3,51,53,54,68,89] Italy, Turkey Eurasian plate+African plate [15,32,70,76] Greece, Azarbaijan Eurasian plate+Arabian plate [86] [69] Morocco African plate [1] Nepal, Israel Indian plate+African plate [18,50] India, Indian plate [24,26,35,36,78] Iran Iranian plate [5,6,9,22,28,37,71,82,83] Saudi Arabia Arabian plate [52] Ethiopia Arabian plate+Somali plate+Nubian plate [86] Caraga Philipine sea plate [14] Vietnam, Malaysia Somali plate [80,84] Chile Nazca plate [10,17,58,71,72] Republic of Croatia Apulian Plate [65] Cyprus African plate+Eurasian Plate +Arabian plate [34] The researchers of the selected articles have evaluated their work using various tools and techniques including software, algorithms and other such as index normalization etc. These tools and techniques have been summarized in Table 9. Table 9. Tools and Techniques applied in the literature for ES development. Tools and Techniques % Reference MATrix Laboratory (MATLAB) 41 [1,3,6,7,9–14,16,19,22,26,30,32,33,76,79,81,84,87] Database Index normalization 4 [23,55] Generalized Langevin equation (GLE) 1.8 [51] Subsidence Coefficient calculator 1.8 [75] Predicate (PRED) in C++ 1.8 [88] Annealing, Sparsespike 1.8 [25] Classification and regression trees(CART) 1.8 [49] Fuzzy C-mean 4 [28,77] Upgraded IF THEN ELSE 4 [27,83] Normalized fuzzy peak ground acceleration (FPGA) 1.8 [8] Sustainability 2020, 12, 2420 15 of 32 Table 9. Cont. Tools and Techniques % Reference Predicate Logic 7 [17,24,50,86] Mean absolute error(MAE), Root mean square error(RMSE) 1.8 [54] Earth resources data analysis system (ERDAS) model maker 1.8 [51] 3Dimensional seismic tomography 1.8 [78] Mean square error(MSE) 4 [31,53] Rapid miner software, frequency, pattern growth algorithm 1.8 [20] Adobe 1.8 [89] Geological carbon storage (GCS) analyzer- Monecarle 1.8 [85] Fuzzy probablistic seismic hazard analyzer (FPSHA) 1.8 [2] FURIA 1.8 [80] AriGIS 1.8 [81] Saga 1.8 [83] Aeronautical reconnaissance coverage Geographic information system (ARC/INFO GIS) 1.8 [84] Geographic information system (GIS), Multi criteria decision analysis (MCDA) 4 [15,82] Multilayer Preceptron -Rule Based (MLP-RB) 1.8 [21] Nearest neighbor Invariant Riemannian metric (AIRM) 1.8 [52] WI (Weighted index) 1.8 [5] Knowledge extraction based on evolutionary learning (KEEL) 1.8 [10] Particle SWARM Optimization (PSO) 1.8 [56] Apache SPARK 1.8 [59] Kernal Fisher Discriminant Algoritthm (KFDA) 1.8 [60] Novel earthquake early warning system (NEEWS) 1.8 [64] Accuracy of results obtained through the proposed expert system for making earthquake predictions using a training set (TS) or independent test set (ITS) has been listed in Table 10. 4.2. Key Facts of Expert System Based Earthquake Prediction Publications Different publication channels, and frequency of publications per source have been presented in Table 7. Number of articles published per year from January 2010 to January 2020 along with the primary source where the article has been submitted is determined. Different publication channels have been identified, including conferences, journals, technical report and book chapters. Around 75 % of the selected papers have been published in peer reviewed journals, 24 % have presented at conferences, workshops, and symposia. 4% technical reports and book chapters have also included in this systematic mapping study. 4.3. ResearchType Facets Addressed by the Identified Publications For deep investigation, we have listed down the type of facets addressed by the selected publications given in Table 6. Type of research presented in the paper, its empirical type, the approach used and focused area has been inquired through RQ2 and listed in Table 6. As identified earlier, the research could be of evaluation type, solution type or experience type. In this mapping study, 23% of the research has presented a solution proposal to illustrate the novelty of the proposed solution for a problem or as an extension of an existing technique. 42% studies belonged to an evaluation type which represents the techniques has passed through the evaluation stages prior to its implementation or solutions have been evaluated after implementation. 22% of the mapping study presented experience Sustainability 2020, 12, 2420 16 of 32 papers which showed the personal experience of the author about earthquake prediction techniques to be used in practice. Table 10. Accuracy claimed in multiple articles. Reference Number of Records (EQ) Accuracy Magnitude Range Data Set [6] 60 78% 5.2–7.7 TS [9] 343 seismograms 99.71% ≥5.0 TS [10] 47 93.54 ≥5.5 TS [13] 12 indices 91% ≥4.5 TS [18] 9531 69.8% ≥2.0 ITS [20] 677245 87.85 3.6–9.1 TS [21] 12690 50.14% ≥3.0 ITS [22] 522 95.8% ≥4.0 TS [23] 1773 85.73% ≥3.5 TS [24] 337 63% ≥3.0 ITS [38] 1000 80.1% < 5.5 TS [43] 227 70% <5.0 TS [50] 77 80.11% ≥5.0 TS [55] 26481 78% 2.5–7.5 TS [63] 10567 40% 0.1–5.9 ITS [76] 100 99.99% 5.5–7.7 TS [86] 476 87.2% ≥5.0 TS [80] 248 84% ≥5.5 TS [83] 78 88% ≥5.0 TS [84] 1059846 86.28% ≥1.5 TS Empirically, the studies included in this systematic mapping study have been categorized into the survey, experiment and case study. 13% of the researchers have conducted surveys to study the already presented prediction models. 39% have conducted experiments on earthquake data retrieved from multiple seismic zones to make earthquake predictions and 32% have presented the case studies in which earthquake prediction task has been worked out using the data of specific areas. In this mapping study, the next type facet is the prediction approach which has been further categorized into model, method, guideline, framework and tool. 40% of the included researches have computational model for making earthquake predictions, 40%, researchers have proposed a method for earthquake prediction, 13% have given the guidelines for earthquake prediction, only 2% of the included studies have presented a framework for the earthquake prediction and analyzed earthquake prediction tools. Last type facet of this mapping study is the target area that has been classified as regional and global. 75% of the included researches have been focused to a certain specific regional area where as 24% of the researchers have used global data to forecast earthquake. Figure 3 shows the number of publications (n) according to the type facets given in RQ2. 4.4. ES type and System Specific Key Aspects of Proposed ES Three types of the proposed ES have been identified by RQ3 and presented in Figure 4. Fuzzy expert system (FES) has been proposed for making earthquake predictions by 29% of the studies. Neuro-Fuzzy expert system for making an earthquake prediction has been proposed by 11% of the studies. The rule based expert system has been proposed by 7% of the included studies. 36% of the articles have used Sustainability 2020, 12, 2420 17 of 32 machine learning for earthquake prediction and 12% of the studies have used other computational approaches like index development or classifier calculation for earthquake prediction. In this mapping study RQ3 also deals with exploring other key aspects of the proposed ESs like input domain, input attributes and their types, data type, prediction logic and the prediction technique covering the use of computational supporting software or algorithm for performing earthquake prediction task. Input domain can be the earthquake parameters like magnitude, b-value, etc. or precursors calculated from real time earthquake data. In this mapping study, 51% researches have directly consumed earthquake parameters (listed as quake variables in Table 6) whereas 38% have studied precursors that serve as indicators for an upcoming devastating earthquake event. Further, we have explored the type of above stated input attributes to analyze that whether predictions have been made by consuming single input attribute or combination of multiple attributes has been used for earthquake predictions. In this mapping study, 27%, researchers have made predictions on the basis of a single attribute where as 68% studies have used a combination of multiple input attributes for earthquake predictions. Through data type we have worked to find that input variables passed to the proposed ES are quantitative type (numeric form like magnitude, slope, etc.) or qualitative type (discrete form like water elevation, emission of radon gas etc.). It has been observed that 57% of researchers have used numeric inputs for earthquake forecast whereas 43% of the studies have performed the earthquake prediction task using a discrete type of input. This mapping study has also captured the data regarding prediction logic and prediction methodology. Through prediction logic, we have determined that the article have used the inductive or deductive logic for making earthquake predictions. Inductive logic has been used by 62% of the studies, whereas deductive logic has been presented in 32% of the included studies for giving absolute results. Next we have worked to determine the predictive methodology to describethe use of any software or an algorithm for evaluation of the proposed system. 42% of the studies have used multiple algorithms for evaluation of their predictions, 48% have used the MATLAB software for evaluation of the proposed ES and 10% have adopted other evaluation mechanisms. 4.5. Quality Assessment The quality assessment score obtained by each paper has been presented in Table 6. These articles have been collected in three distinct categories and presented in Table 11 according to the scores obtained by every article given Table 6. By observing the quality scores, 74% of the papers included in the mapping study had obtained above average score, 11% of the studies had obtained average score and 15% of the studies remained below average. This quality assessment may facilitate researchers in the selection of articles for their work. Table 11. Scores. Reference Score Average =1.5 Total % [1–14,17,19–22,25–27,31,32,34,35,37–39,42–44,49–52,55,56,58–72,74] Above average 55 79 [18,23,24,28,53,54,73] Average 7 10 [15,16,30,33,36,40,41,57] Below average 8 11 5. Discussion In this mapping study, main techniques used for earthquake prediction including rule-based, fuzzy, neuro-fuzzy and machine learning have been explored. Expert systems based approaches have been used for seismic risk assessment for landslide susceptibility through seismic hazard analysis [50–52] and soil classification [53]. ES have also been applied in earthquake engineering for seismic hazard analysis and assessment of bridges and buildings under multiple hazards [24,54,55]. The expert systems have been used to analyze multiple aspects of earthquake prediction, but due to the non-existence of the Sustainability 2020, 12, 2420 18 of 32 time-dependent global earthquake forecasting model, the regional earthquake likelihood models have been popular [87]. On the other hand, some initial research has been conducted to design earthquake early-warning systems that work globally [32,49]. A detailed discussion regarding the history and theory of accelerating seismic release and preparedness for an earthquake has been conducted in [90,91]. Extended earthquake sequences with stable features have been observed over long time periods and explained accelerated seismicity before the occurrence of devastating earthquakes with in time. The aftershock decay models proposed in the articles from 1894 till 2014 about accelerating seismology, has been analyzed in [92] and valuable information about multiple scientific process including earthquake cascaded events have been collected. Multiple anomalies have been investigated in [93] to formulate the criteria for identification of the genuine precursors from a preliminary list of identifying precursors, methods or case studies. A collection of one hundred articles related to accelerating seismology has been studied in [94] to classify the precursor into two types including critical processes such as cascading triggering of earthquake events or normal processes such as pressure on main fault. Many other studies investigated earthquake precursors like [95] examined multiple parameters collected from anomalies present in geophysical fields such as ionospheric disorder for short term earthquake prediction. Genetic algorithm (GA) has been used to optimize the hybrid artificial neural network model for the prediction of peak particle velocity in [96]. An artificial neural network has been applied to predict shear wave velocity in [97]. A seismic data-driven tool has been proposed for seismic fracture identification using large post-stack seismic dataset in [98]. The dynamic response of geogrid machine foundation bed has been studied in [99]. Rockfall hazard assessment using artificial neural network has been performed in [100]. The properties of the lower ionosphere have been examined in [101] using random matrix theory for the prediction of earthquakes. Both spatial and spatio temporal earthquake predictions have been made in [102] using machine learning methodologies. Reference [103] presented a natural time analysis of seismic-electric signal emerged before two earthquake events for the prediction of next expected earthquake events.The density of foreshocks and aftershocks has been analysed in [72] for prediction of future earthquakes that may occur in the selected seismically active regions. Reference [56] examined the impact of deep learning algorithms for classification of earthquake precursors for extraction of seismic patterns and unique features from big data. Reference [57] distinguished between seismic signals and non seismic signals using logistic regression method on the data collected from National Seismological Network of Colombia. Reference [58] applied Support Vector regression and Hybrid neural Network for earthquake prediction in Hindukush, Chile and Southern California regions with prediction accuracy rate of 82.7%, 84.9%, 90.6% respectively. Reference [59] analyzed earthquake magnitude prediction on the basis of regression algorithms and cloud based big data infrastructure. Reference [60] used grid-search method to construct support vector machine (SVM) based model for earthquake prediction. Reference [61] launched a web based platform for automatic calculation of seismic hazard fields to predict earthquakes. Reference [62] highlighted the vital role of temporal strong ground motion parameters in earthquake engineering and risk assessment using machine learning methods. Reference [63] used gradient boosted trees algorithm of machine learning to perform fivefold cross validation on training data set from earthquake catalogs to make earthquake predictions. References [64] developed an earthquake early warning system using image recognition techniques. Reference [65] studies multiple machine learning methods including random forest, artificial neural network, recurrent neural network, Naïve Bayesian and regression for earthquake prediction. Reference [66] examined the efficiency of bio-inspired algorithms for supervised classification of on real datasets to handle emergencies. Reference [67] discussed the emerging Regional Earthquake likelihood models in making earthquake predictions with improved accuracy. Reference [68] used tree-based ensemble methodologies for earthquake prediction within time period of 15 days and calculated seismic features of Hindukush region applying machine learning methods for macro earthquake prediction. Reference [69] proposed a multi-step prediction method for short-term prediction of strong Sustainability 2020, 12, 2420 19 of 32 earthquake with relatively high accuracy. Reference [70] proposed a monitoring system for preparing machine learning data-sets for earthquake prediction based on seismic-acoustic signals from stations in Azerbaijan. Reference [71] estimated the magnitudes of earthquake events recorded on daily bases using artificial neural network (ANN) to prove that training set of global data is more effective in earthquake prediction than making earthquake prediction using local data. References [72,73] studied the importance of machine learning methods in earthquake prediction and highlighted the impact of accurate prediction on country’s economy. Reference [74] presented a scheme for large earthquake prediction based on radial basis function (RBF) neural network (NN) models. This section summarizes and discusses the results of the systematic mapping study. Taxonomy of earthquake prediction is given in Figure 5. Sustainability 2020, 12, x FOR PEER REVIEW 8 of 33 prediction than making earthquake prediction using local data. [72,73] studied the importance of machine learning methods in earthquake prediction and highlighted the impact of accurate prediction on country’s economy. [74] presented a scheme for large earthquake prediction based on radial basis function (RBF) neural network (NN) models. This section summarizes and discusses the results of the systematic mapping study. Taxonomy of earthquake prediction is given in Figure 5. Figure 5. Taxonomy of earthquake prediction approaches. There are two distinct types of earthquake prediction approaches presented in Figure 5, i.e. deterministic forecasts and probabilistic forecasts. Deterministic forecasts are made on the base of earthquake characteristics like rupture length, modified Mercalli and return period where as probabilistic forecasts deals with precursors, implications of earthquake physics and elastic rebound. Table 12 clusters the articles according to the taxonomy given in Figure 5. Table 12. Approaches used to predict particular seismic phenomena. Deterministic Domain s Seismic Phenomena Approach Reference C ha ra ct er is ti c ea rt hq ua ke Rupture length Modified Mercalli Return Period Classification, Clustering, Machine Learning (ML), Neural network (NN) [2,6,8,10,16,17-– 20,28,33– 35,37,39,42,44,58– 62,65,69,70,72,74,80,86] Probabilistic P re cu rs or Animal behavior Predicate Logic, [27,66] Seismic velocity [9,32,57] Figure 5. Taxonomy of earthquake prediction approaches. There are two distinct types of earthquake prediction approaches presented in Figure 5, i.e., deterministic forecasts and probabilistic forecasts. Deterministic forecasts are made on the base of earthquake characteristics like rupture length, modified Mercalli and return period where as probabilistic forecasts deals with precursors, implications of earthquake physics and elastic rebound. Table 12 clusters the articles according to the taxonomy given in Figure 5. From Table 12 it is clear that multiple approaches have been used for eathquake prediction including Rule based, Fuzzy and Neuro-fuzzy and machine learning. For deterministic estimations, earthquakes have been grouped according to their magnitude range by applying classification, clustering and machine learning techiques. For probablistic forcasts, multiple artificial intelligence methods have been exercised for making earthquake predictions. Comparitive studies have been conducted [13], systematic methods of predicate logic have also been applied to analyze the precursor that may have vital importance for earthquake prediction [9,27,32,57,65]. To estimate the mutual relationship of precursors, regression line has been calculated [26]. Machine learnining approaches have shown greater improvement in prediction accuracy presented in Table 10. Multi criteria decision making (MCDM) approaches including Technique for order of preference by similarity of ideal solutions (TOPSIS) and Sustainability 2020, 12, 2420 20 of 32 Aggrigated indices randomization (AIRM) have also been applied for earthquake prediction. Multiple precursors may occur before catastrophic earthquakes which have been analyzed through MCDM methods by assigning weightage to every precurssor for the prediction of expected earthquake events. Pattern recognition techniques have also been exercised to analyse the seismic patterns generated by the energy released in previous earthquakes and radon emission recorded before an earthquake. Table 12. Approaches used to predict particular seismic phenomena. Deterministic Domains Seismic Phenomena Approach Reference Characteristic earthquake Rupture length Modified Mercalli Return Period Classification, Clustering, Machine Learning (ML), Neural network (NN) [2,6,8,10,16–20,28,33–35,37,39,42,44, 58–62,65,69,70,72,74,80,86] Probabilistic Precursor Animal behavior Predicate Logic, Aggregated Indices Randomization Method(AIRM), Regression, Comparison, Clustering, ML, NN [27,66] Seismic velocity [9,32,57] Seismic resistivity [84] Topography uplift [5,10,14,15,22,23,50,52,77] Radon emission [13,104] Seismic electric signal [21] Electromagnetic signals [30,63,64,71] Ground water elevation [105,106] Land sliding [41,82,83,86] Earthquake physics Earthquake light Ionosphere disorder ML, NN Technique for Order of Preference by Similarity to Ideal Solution [36,85] [26] Elsticrebound Seismic Gap Seismic Pattern Pattern recognition Clustering, ML [37,38,40,51] [49,88] 5.1. Comparitive Analysis of Methods Multiple techniques have been used for earthquake prediction in the literature. We have formulated the foundation of this mapping study on expert systems.However, there are many other approaches as well for the prediction of earthquake presented in the literature. We have compared expert system developed for earthquake prediction with other artificial intelligence techniques used for the same in Table 13. Table 13. Comparison of three Artificial intelligence techniques. Method Comparison Neural networks and Expert systems Expert system is about capturing and encoding (often manually) rules that experts use so as to develop a program that can mimic their behavior in a very specific domain. It often involved chaining these rules together. With ANN the rules are encoded automatically by presenting examples, good and bad, to the network. The network adjusts weightings over many iterative cycles, honing its output to the correct value. Feed Forward Neural Networks can predict long term and short term earthquakes but it cannot get feedback of output from multiple layers and Back Propagation Neural Network mostly trapped in different local conditions during the training phase of earthquake data sets. However, probability of getting desired output raises when it is tested with ideally designed inputs. Machine learning and Expert systems Machine learning (ML) focuses on modeling of data statistically and expert is involved at the time of decision. Supervised learning algorithms are used to copy the ending decisive behavior of the Expert systems are based upon set of rules prescribed by human expert and learn by directly injecting the domain level knowledge of human expert. The knowledge obtained from the expert is completely converted into membership functions and used in decision making. Explanation facility is also available as an expert describes all the steps till decision, the basis and exception handling procedures. A rigid system is developed that follows exact rules as described by the expert. Rigidness of the expert system makes it most suitable from all other techniques for predicting future earthquakes. Table 14 presented the comparision of multiple methods including fuzzy, neuro fuzzy, rule based and machine learning used for earthquake prediction in the literature.It can be observed that both deterministic and probablistic methods are being excercised for earthquake prediction. Many researchers have performed numerical experiments to achieve success in predicting earthquakes. Some have developed tools while others have explored multiple dimensions of application area. Table 14 clearly presents that most recent studies have worked on exploring machine learning based models for earthquake prediction. Sustainability 2020, 12, 2420 21 of 32 Table 14. Comparision of Methods. Method Ref. Prediction Approach Algorithm Defined Application Area Dataset Deterministic Probabilistic Analytical Work Global Approximation Numerical Experiment Exploration with actual forecasts Success Achieved Characteristic Earth quake Precursors Zone Studied Fuzzy Expert System [2] 4 4 4 4 China [6] 4 4 4 4 Iran [8] 4 4 Taiwan [9] 4 4 4 Iran [10] 4 4 4 4 4 California [13] 4 4 4 China [14] 4 4 4 Caraga [15] 4 4 4 Turkey [18] 4 4 Nepal [21] 4 4 4 4 [23] 4 4 4 4 China [24] 4 4 4 India [26] 4 4 4 India [50] 4 4 4 4 Nepal [51] 4 4 4 China [52] 4 4 Saudi Arabia [80] 4 4 4 4 4 Malaysia [82] 4 4 Iran [83] 4 4 4 4 Iran [84] 4 4 4 4 Malaysia [86] 4 4 Ethiopia Neuro Fuzzy Expert System (NFES) [5] 4 4 4 Iran [17] 4 4 Chile [19] 4 4 4 [20] 4 4 4 4 Sustainability 2020, 12, 2420 22 of 32 Table 14. Cont. Method Ref. Prediction Approach Algorithm Defined Application Area Dataset Deterministic Probabilistic Analytical Work Global Approximation Numerical Experiment Exploration with actual forecasts Success Achieved Characteristic Earth quake Precursors Zone Studied Neuro Fuzzy Expert System (NFES) [22] 4 4 4 Iran [27] 4 4 4 [28] 4 4 Iran [30] 4 4 4 4 [32] 4 4 Turkey [33] 4 4 4 4 [49] 4 4 4 4 China [77] 4 4 4 [81] 4 4 4 Greece Machine Learning (ML) [34] 4 4 4 Cyprus [35] 4 4 4 India [36] 4 4 India [37] 4 4 4 4 Iran [38] 4 4 4 4 [39] 4 4 4 4 4 [40] 4 4 4 [41] 4 4 4 California [42] 4 4 4 4 [44] 4 4 4 4 [56] 4 4 4 4 4 4 [57] 4 4 4 California [58] 4 4 4 4 California [59] 4 4 4 4 Sustainability 2020, 12, 2420 23 of 32 Table 14. Cont. Method Ref. Prediction Approach Algorithm Defined Application Area Dataset Deterministic Probabilistic Analytical Work Global Approximation Numerical Experiment Exploration with actual forecasts Success Achieved Characteristic Earth quake Precursors Zone Studied Machine Learning (ML) [60] 4 4 4 4 Japan [61] 4 4 4 4 [62] 4 4 4 [63] 4 4 4 4 4 [64] 4 4 4 California [65] 4 4 Croatia [66] 4 4 4 [68] 4 4 Pakistan [69] 4 4 Greece [70] 4 4 4 4 Turkey [71] 4 4 Iran [72] 4 4 Chile [74] 4 4 4 4 Sustainability 2020, 12, 2420 24 of 32 5.2. Principal Findings The goal of this systematic mapping study is to examine the current status of ES used to predict earthquakes by selecting the appropriate number of recently published papers. We have passed all the studies from our selection and quality criteria to classify the studies according to the scores obtained by every research work on the basis of rules listed in Table 6. The categories of the studies according to their quality scores has been listed in Table 6. In this mapping study, 74% articles have above average scores, 15% articles scored average quality results and 11% researches are below average.Table 11 will facilitate the researchers in selecting quality studies. In Figure 6, we have combined the answers from sub-parts of RQ2 for presenting an overview of earthquake prediction activity. This mapping has allowed us to obtain more information on how the results from each RQ relate to each others. Figure 6 presents the research type facet related to earthquake prediction, distributed over approaches and its empirical type. It allows us to conclude that only one model was proposed by the authors that mainly report their experience in the earthquake prediction process. However, guidelines have been provided by six authors. The majority of the solution proposals have developed models for earthquake prediction. However, just one researcher has developed a tool for earthquake prediction evaluation. Further information regarding the relationship between type facets is shown in Figure 6. Sustainability 2020, 12, x FOR PEER REVIEW 12 of 33 above average scores, 15% articles scored average quality results and 11% researches are below average.Table 11 will facilitate the researchers in selecting quality studies. In Figure 6, we have combined the answers from sub-parts of RQ2 for presenting an overview of earthquake prediction activity. This mapping has allowed us to obtain more information on how the results from each RQ relate to each others. Figure 6 presents the research type facet related to earthquake prediction, distributed over approaches and its empirical type. It allows us to conclude that only one model was proposed by the authors that mainly report their experience in the earthquake prediction process. However, guidelines have been provided by six authors. The majority of the solution proposals have developed models for earthquake prediction. However, just one researcher has developed a tool for earthquake prediction evaluation. Further information regarding the relationship between type facets is shown in Figure 6. Figure 6. Relationship between type facets given in Research Question 2 (RQ2). RQ3 deals with identification of the type of proposed ES. Figure 7 shows the relationship between the proposed ES with other system specific details. Machine / deep learning has emerged as most focused and recent trend in the earthquake prediction process. However, It can be clearly seen from Figure 7 that FES has been proposed by a majority of the researchers for a long time period. FES is representing a balance in the selection of all input parameters except the use of deductive logic in earthquake prediction in Figure 7. It might be due to the nature of FES basically reflecting uncertainty [15] that's why most of the researchers have used inductive logic to propose FES for the earthquake prediction. However, some of the researchers have also worked to proposed NFES and RBES but their frequency is very low. Figure 6. Relationship between type facets given in Research Question 2 (RQ2). RQ3 deals with identification of the type of proposed ES. Figure 7 shows the relationship between the proposed ES with other system specific details. Machine / deep learning has emerged as most focused and recent trend in the earthquake prediction process. However, It can be clearly seen from Figure 7 that FES has been proposed by a majority of the researchers for a long time period. FES is representing a balance in the selection of all input parameters except the use of deductive logic in earthquake prediction in Figure 7. It might be due to the nature of FES basically reflecting uncertainty [15] that’s why most of the researchers have used inductive logic to propose FES for the earthquake prediction. However, some of the researchers have also worked to proposed NFES and RBES but their frequency is very low. Sustainability 2020, 12, 2420 25 of 32 Sustainability 2020, 12, x FOR PEER REVIEW 13 of 33 Figure 7. Relationship between ES approaches and system specific parameters. The findings of our systematic mapping study havethe following implications for practitioners and researchers. This study will allow them to discover the existing use of expert system in the literature concerning earthquake prediction. In order to improve the reliability in earthquake prediction, researchers and practitioners may consider the following advices: 1. Globally, we need a program of identification and characterization of potentially hazardous faults in multiple seismic zones. From those studies, site-specific expected seismic shaking maps can be developed that would facilitate in developing expert system for earthquake prediction process. 2. By comparing different forecasts that are computed from common data, contrasts in performance can be tied to specific features of the computational prediction method. Enforcing the need to create a testable prediction, hypothesis that may reveal shortcomings or incomplete features of the prediction method is needed. 3. Activities focusing on comparative testing of computational prediction methods based on seismicity and fault information that provide probabilistic predictions of moderate magnitude earthquakes on a geographic grid are needed. This approach can be optimized to achieve useful statistics in a short time and can also advance the research field by providing insights into the computational predictability of earthquakes. However, visible hypotheses such as the M8/MSc predictions of global earthquakes, the “reverse detection of precursors” method, or the Retrograde Intravenous Pressure Infusion “RIPI” method, each of which analyze temporal and spatial variations in seismicity, or other methods based on observable quantities such as the electromagnetic field, ground temperature, gaseous emissions, geodetic deformation, or changes in seismic wave speed. Many of the most visible and influential earthquake predictions are posed as “alarms” or “times of increased probability” (TIPs) within some specified region rather than as probabilities on a grid of points. 4. Evaluation of emerging situations such as earthquake swarms, the likelihood of damaging aftershocks or triggered earthquakes following major quakes, or the likelihood of re-rupture of a fault following a major earthquake should be examined. Likewise, a broader suite of statistical tests, spanning the range from straightforward to sophisticated, would allow some prediction methods to be easily disproven in a way that’s clear to researchers, the media and the public, while providing the rigorous analysis required for comparative testing. These should include statistical tests applicable to alarm-based computational prediction methods. 5.3. Evolution of Tools and Techniques Figure 7. Relationship between ES approaches and system specific parameters. The findings of our systematic mapping study havethe following implications for practitioners and researchers. This study will allow them to discover the existing use of expert system in the literature concerning earthquake prediction. In order to improve the reliability in earthquake prediction, researchers and practitioners may consider the following advices: 1. Globally, we need a program of identification and characterization of potentially hazardous faults in multiple seismic zones. From those studies, site-specific expected seismic shaking maps can be developed that would facilitate in developing expert system for earthquake prediction process. 2. By comparing different forecasts that are computed from common data, contrasts in performance can be tied to specific features of the computational prediction method. Enforcing the need to create a testable prediction, hypothesis that may reveal shortcomings or incomplete features of the prediction method is needed. 3. Activities focusing on comparative testing of computational prediction methods based on seismicity and fault information that provide probabilistic predictions of moderate magnitude earthquakes on a geographic grid are needed. This approach can be optimized to achieve useful statistics in a short time and can also advance the research field by providing insights into the computational predictability of earthquakes. However, visible hypotheses such as the M8/MSc predictions of global earthquakes, the “reverse detection of precursors” method, or the Retrograde Intravenous Pressure Infusion “RIPI” method, each of which analyze temporal and spatial variations in seismicity, or other methods based on observable quantities such as the electromagnetic field, ground temperature, gaseous emissions, geodetic deformation, or changes in seismic wave speed. Many of the most visible and influential earthquake predictions are posed as “alarms” or “times of increased probability” (TIPs) within some specified region rather than as probabilities on a grid of points. 4. Evaluation of emerging situations such as earthquake swarms, the likelihood of damaging aftershocks or triggered earthquakes following major quakes, or the likelihood of re-rupture of a fault following a major earthquake should be examined. Likewise, a broader suite of statistical tests, spanning the range from straightforward to sophisticated, would allow some prediction methods to be easily disproven in a way that’s clear to researchers, the media and the public, while providing the rigorous analysis required for comparative testing. These should include statistical tests applicable to alarm-based computational prediction methods. Sustainability 2020, 12, 2420 26 of 32 5.3. Evolution of Tools and Techniques To reveal the development trend of researchers, various tools and techniques used to develop ES for earthquake prediction given in the literature have been listed in Table 10. It is clear from Table 10 that 41% of the researchers have used the MATLAB software for development of an ES. MATLAB is in the market since 1984. MATLAB is used for multi domain purposes like signal processing, image processing and automation, etc. It has Simulink, state flow, embedded coder and Simulink coder. Simulink facilitates the model based development where as code is automatically generated through embedded coder. Moreover, traceability of the code is much easier than legacy coding. Simulink helps in the development of a system in block diagrams by providing many elements like transfer function, summing junction, fuction generators and oscilloscopes which makes the work easier for the researchers. Through its debugging option, it has gained the trust of the researchers from last thirty five years. It is also clear from Table 6 that for earthquake prediction rule-based expert systems have been focused till 2013 then, till 2017 fuzzy and neur- fuzzy expert systems have been explored. But, now a prominent change in the research trend has been observed from 2018 onwards that earthquake predictions are being carried out using machine learning and deep learning approaches. Changing trend of earthquake prediction approaches has been shown in the timeline presented in the Figure 8. Sustainability 2020, 12, x FOR PEER REVIEW 14 of 33 To reveal the development trend of researchers, various tools and techniques used to develop ES for earthquake prediction given in the literature have been listed in Table 10. It is clear from Table 10 that 41% of the researchers have used the MATLAB software for development of an ES. MATLAB is in the market since 1984. MATLAB is used for multi domain purposes like signal processing, image processing and automation, etc. It has Simulink, state flow, embedded coder and Simulink coder. Simulink facilitates the model based development where as code is automatically generated through embedded coder. Moreover, traceability of the code is much easier than legacy coding. Simulink helps in the development of a system in block diagrams by providing many elements like transfer function, summing junction, fuction generators and oscilloscopes which makes the work easier for the researchers. Through its debugging option, it has gained the trust of the researchers from last thirty five years.It is also clear from Table 6 that for earthquake prediction rule-based expert systems have been focused till 2013 then, till 2017 fuzzy and neur- fuzzy expert systems have been explored. But, now a prominent change in the research trend has been observed from 2018 onwards that earthquake predictions are being carried out using machine learning and deep learning approaches. Changing trend of earthquake prediction approaches has been shown in the timeline presented in the Figure 8. 2013 2017 2020 Figure 8. Timeline of earthquake prediction approaches. 6. Conclusion and Future Directions This paper has presented a systematic mapping study to summarize the existing research on multiple approaches involving machine learning, neuro-fuzzy, fuzzy and rule based approaches that have been used for earthquake prediction. Out of 2137 studies, 70 articles published between January 2010 till January 2020 were carefully selected and classified on the basis of research type, empirical type, approach, target area, and other system specific parameters. Publication source and trend have also been identified. Most of the articles considered in this study have been selected from peer reviewed journals and (Computing Research and Education) CORE ranked conferences. Majority of the papers included in this mapping study involve empirical validation for the proposed solutions to predict earthquakes. The use of various types of ES presented in this study may help researchers to identify approaches that can be adopted in order to improve the quality of earthquake prediction in their work. For future research more attention should be given to theapplication of machine learning and deep learning methods for earthquake prediction. Due to ever increasing volume of data, there is a need to employ machine learning and deep learning for the prediction of earthquakes. Moreover , there is a need to conduct more of evaluation research for the validation of the already presented prediction models based on fuzzy logic. The analysis of the presented research works show that most of the approaches focused on analyzing the precursors generating direct warning of an earthquake. However, recent trends show that there is a need to extend this work by involving other factors which may include volcanic eruption, nuclear explosion, hurricanes, tsunamis etc. The results obtained showed that an increasing amount of attention has been paid to the use of an ES for earthquake prediction since 2011. It has been noticed that till 2013 the reserch has been more focused on rule based methods for earthquake prediction. Then the trend has changed towards using fuzzy and neuro fuzzy methods till 2018. Machine learning and deep learning has emerged as the most focused approach for earthquake prediction in the recent time. The classification developed in this mapping study has presented increased trend of applying FES and ML in making long term earthquake predictions. However, In the future, more advanced RBES FES / NFES Machine/ Deep Learning Figure 8. Timeline of earthquake prediction approaches. 6. Conclusion and Future Directions This paper has presented a systematic mapping study to summarize the existing research on multiple approaches involving machine learning, neuro-fuzzy, fuzzy and rule based approaches that have been used for earthquake prediction. Out of 2137 studies, 70 articles published between January 2010 till January 2020 were carefully selected and classified on the basis of research type, empirical type, approach, target area, and other system specific parameters. Publication source and trend have also been identified. Most of the articles considered in this study have been selected from peer reviewed journals and (Computing Research and Education) CORE ranked conferences. Majority of the papers included in this mapping study involve empirical validation for the proposed solutions to predict earthquakes. The use of various types of ES presented in this study may help researchers to identify approaches that can be adopted in order to improve the quality of earthquake prediction in their work. For future research more attention should be given to theapplication of machine learning and deep learning methods for earthquake prediction. Due to ever increasing volume of data, there is a need to employ machine learning and deep learning for the prediction of earthquakes. Moreover, there is a need to conduct more of evaluation research for the validation of the already presented prediction models based on fuzzy logic. The analysis of the presented research works show that most of the approaches focused on analyzing the precursors generating direct warning of an earthquake. However, recent trends show that there is a need to extend this work by involving other factors which may include volcanic eruption, nuclear explosion, hurricanes, tsunamis etc. The results obtained showed that an increasing amount of attention has been paid to the use of an ES for earthquake prediction since 2011. It has been noticed that till 2013 the reserch has been more focused on rule based methods for earthquake prediction. Then the trend has changed towards using fuzzy and neuro fuzzy methods till 2018. Machine learning and deep learning has emerged as the most focused approach for earthquake prediction in the recent time. Sustainability 2020, 12, 2420 27 of 32 The classification developed in this mapping study has presented increased trend of applying FES and ML in making long term earthquake predictions. However, In the future, more advanced deep learning based model should be designed to make pinpoint predictions. Moreover, there is a continuing need to develop a suite of basic tools and reference models to rapidly establish an unbiased framework to evaluate prediction methods, which enforces strict adherence to the scientific method, motivates investigators to accurately and unambiguously express prediction hypotheses, and provides guidance and tools for formal testing of those hypotheses. These features would lead to progress in evaluating seismicity-based models. Author Contributions: “conceptualization, R.T. and M.S.F.; methodology, R.T.; software, R.T.; validation, R.T., M.S.F and A.A.; formal analysis, R.T.; investigation, M.S.F; resources, A.A.; data curation, Rabia; writing—original draft preparation, R.T.; writing—review and editing, R.T., M.S.F. and A.A.; visualization, A.A.; supervision, M.S.F, A.A.; project administration, M.S.F.; funding acquisition, R.T.” All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding Acknowledgments: The authors appreciate the anonymous reviewers for their valuable feedback on the initial version of this paper. Conflicts of Interest: The authors declare no conflict of interest. References 1. Laasri, H.A.; Akhouayri, E.-S.; Agliz, D.; Zonta, D.; Atmani, A. A fuzzy expert system for automatic seismic signal classification. Expert Syst. Appl. 2015, 42, 1013–1027. [CrossRef] 2. Andrić, J.M.; Lu, D.-G. Fuzzy probabilistic seismic hazard analysis with applications to Kunming city, China. Nat. Hazards 2017, 89, 1031–1057. [CrossRef] 3. Asim, K.M.; Awais, M.; Martínez–Álvarez, F.; Iqbal, T. Seismic activity prediction using computational intelligence techniques in northern Pakistan. Acta Geophys. 2017, 65, 919–930. [CrossRef] 4. Asim, K.M.; Martínez-Álvarez, F.; Basit, A.; Iqbal, T. Earthquake magnitude prediction in Hindukush region using machine learning techniques. Nat. Hazards 2016, 85, 471–486. [CrossRef] 5. Aghdam, I.N.; Varzandeh, M.H.M.; Pradhan, B. Landslide susceptibility mapping using an ensemble statistical index (Wi) and adaptive neuro-fuzzy inference system (ANFIS) model at Alborz Mountains (Iran). Environ. Earth Sci. 2016, 75, 553. [CrossRef] 6. Ahmadi, M.; Nasrollahnejad, A.; Faraji, A. Prediction of Peak ground acceleration for earthquakes by using intelligent methods. In Proceedings of the 2017 5th Iranian Joint Congress on Fuzzy and Intelligent Systems (CFIS), Qazvin, Iran, 7–9 March 2017; pp. 7–12. 7. Bahrami, B.; Shafiee, M. Fuzzy Descriptor Models for Earthquake Time Prediction Using Seismic Time Series. Int. J. Uncertain. Fuzziness Knowl. -Based Syst. 2015, 23, 505–519. [CrossRef] 8. Ahumada, A.; Altunkaynak, A.; Ayoub, A. Fuzzy logic-based attenuation relationships of strong motion earthquake records. Expert Syst. Appl. 2015, 42, 1287–1297. [CrossRef] 9. Aboonasr, S.F.G.; Zamani, A.; Razavipour, F.; Boostani, R. Earthquake hazard assessment in the Zagros Orogenic Belt of Iran using a fuzzy rule-based model. Acta Geophys. 2017, 65, 589–605. [CrossRef] 10. Fernández-Gómez, M.J.; Asencio-Cortés, G.; Troncoso, A.; Martínez-Álvarez, F. Large Earthquake Magnitude Prediction in Chile with Imbalanced Classifiers and Ensemble Learning. Appl. Sci. 2017, 7, 625. [CrossRef] 11. Andrić, J.M.; Lu, D.G. Fuzzy–Based Method for the Prediction of Seismic Resilience of Bridges; Elsevier: Amsterdam, The Netherlands, 2016. 12. Azadeh, A.; Mohammadfam, I.; Khoshnoud, M.; Nikafrouz, M. Design and implementation of a fuzzy expert system for performance assessment of an integrated health, safety, environment (HSE) and ergonomics system: The case of a gas refinery. Inf. Sci. 2008, 178, 4280–4300. [CrossRef] 13. Cai, W.; Dou, L.; Zhang, M.; Cao, W.; Shi, J.-Q.; Feng, L. A fuzzy comprehensive evaluation methodology for rock burst forecasting using microseismic monitoring. Tunn. Undergr. Space Technol. 2018, 80, 232–245. [CrossRef] 14. Cerna, M.A.D.; Maravillas, E.A. Mamdani Fuzzy Decision Model for GIS-Based Landslide Hazard Mapping. In Transactions on Engineering Technologies; Springer: Singapore, 2017; pp. 59–73. http://dx.doi.org/10.1016/j.eswa.2014.08.023 http://dx.doi.org/10.1007/s11069-017-3007-z http://dx.doi.org/10.1007/s11600-017-0082-1 http://dx.doi.org/10.1007/s11069-016-2579-3 http://dx.doi.org/10.1007/s12665-015-5233-6 http://dx.doi.org/10.1142/S0218488515500221 http://dx.doi.org/10.1016/j.eswa.2014.09.035 http://dx.doi.org/10.1007/s11600-017-0055-4 http://dx.doi.org/10.3390/app7060625 http://dx.doi.org/10.1016/j.ins.2008.06.026 http://dx.doi.org/10.1016/j.tust.2018.06.029 Sustainability 2020, 12, 2420 28 of 32 15. D’Urso, M.G.; Masi, D.; Zuccaro, G.; Gregorio, D. Multicriteria Fuzzy Analysis for a GIS-Based Management of Earthquake Scenarios. Comput. Civ. Infrastruct. Eng. 2017, 33, 165–179. [CrossRef] 16. Dutta, P.K.; Mishra, O.P.; Naskar, M.K. A review of operational earthquake forecasting methodologies using linguistic fuzzy rule-based models from imprecise data with weighted regression approach. J. Sustain. Sci. Manag. 2013, 8, 220–235. 17. Martínez-Álvarez, F.; Reyes, J.; Morales-Esteban, A.; Rubio-Escudero, C. Determining the best set of seismicity indicators to predict earthquakes. Two case studies: Chile and the Iberian Peninsula. Knowl. -Based Syst. 2013, 50, 198–210. [CrossRef] 18. Last, M.; Rabinowitz, N.; Leonard, G. Predicting the Maximum Earthquake Magnitude from Seismic Data in Israel and Its Neighboring Countries. PLoS ONE 2016, 11, e0146101. [CrossRef] 19. Ikram, A.; Qamar, U. Developing an expert system based on association rules and predicate logic for earthquake prediction. Knowl. -Based Syst. 2015, 75, 87–103. [CrossRef] 20. Ikram, A.; Qamar, U. A rule-based expert system for earthquake prediction. J. Intell. Inf. Syst. 2014, 43, 205–230. [CrossRef] 21. Ghorbani, S.; Barari, M.; Hoseini, M. Presenting a new method to improve the detection of micro-seismic events. Environ. Monit. Assess. 2018, 190, 464. [CrossRef] 22. Ghorbanzadeh, O.; Rostamzadeh, H.; Blaschke, T.; Gholaminia, K.; Aryal, J. A new GIS-based data mining technique using an adaptive neuro-fuzzy inference system (ANFIS) and k-fold cross-validation approach for land subsidence susceptibility mapping. Nat. Hazards 2018, 94, 497–517. [CrossRef] 23. Meng, Q.; Miao, F.; Zhen, J.; Wang, X.; Wang, A.; Peng, Y.; Fan, Q. GIS-based landslide susceptibility mapping with logistic regression, analytical hierarchy process, and combined fuzzy and support vector machine methods: A case study from Wolong Giant Panda Natural Reserve, China. Bull. Int. Assoc. Eng. Geol. 2015, 75, 923–944. [CrossRef] 24. Tahernia, N. Fuzzy-Logic Tree Approach for Seismic Hazard Analysis. Int. J. Eng. Technol. 2014, 6, 182–185. [CrossRef] 25. Wang, S.; Liu, H.; Wang, S.; Tong, S.; Wang, R. Pseudo-acoustic inversion method and its application. In Proceedings of the 2010 Seventh International Conference on Fuzzy Systems and Knowledge Discovery, Yantai, China, 10–12 August 2010; Volume 2, pp. 598–601. 26. Ratnam, D.V.; Vindhya, G.; Dabbakuti, J.R.K.K. Ionospheric forecasting model using fuzzy logic-based gradient descent method. Geodesy Geodyn. 2017, 8, 305–310. [CrossRef] 27. Hossain, M.S.; Al Hasan, A.; Guha, S.; Andersson, K. A Belief Rule Based Expert System to Predict Earthquake under Uncertainty. J. Wirel. Mob. Netw. Ubiquitous Comput. Dependable Appl. 2018, 9, 26–41. 28. Mirrashid, M. Earthquake magnitude prediction by adaptive neuro-fuzzy inference system (ANFIS) based on fuzzy C-means algorithm. Nat. Hazards 2014, 74, 1577–1593. [CrossRef] 29. Abayon, R.C.; Apilado, J.R.; Pacis, D.B.; Chua, M.G.; Aguilar, A.V.; Calim, J.; Padilla, S.M.A.; Puno, J.C.S.; Apsay, M.R.B.; Bustamante, R. A Weather Prediction and Earthquake Monitoring System. In Proceedings of the 2018 IEEE Conference on Systems, Process and Control (ICSPC), Malacca, Malaysia, 14–15 December 2018; pp. 203–208. 30. Shibli, M. A novel approach to predict earthquakes using adaptive neural fuzzy inference system and conservation of energy-angular momentum. Int. J. Comp. Inf. Syst. Ind. Manag. Appl. 2011, 2150, 371–390. 31. Torres, V.M.; Castillo, O. A Type-2 Fuzzy Neural Network Ensemble to Predict Chaotic Time Series. In Studies in Computational Intelligence; Springer: Berlin/Heidelberg, Germany, 2015; Volume 601, pp. 185–195. 32. Tosunoğlu, N.G.; Apaydin, A. A New Spatial Algorithm Based on Adaptive Fuzzy Neural Network for Prediction of Crustal Motion Velocities in Earthquake Research. Int. J. Fuzzy Syst. 2018, 20, 1656–1670. [CrossRef] 33. Kamath, R.S.; Kamat, R.K. Earthquake Magnitude Prediction for Andaman-Nicobar Islands: Adaptive Neuro Fuzzy Modeling with Fuzzy Subtractive Clustering Approach. J. Chem. Pharm. Sci. 2017, 10, 1228–1233. 34. Asim, K.M.; Moustafa, S.S.; Niaz, I.A.; Elawadi, E.A.; Iqbal, T.; Martínez-Álvarez, F. Seismicity analysis and machine learning models for short-term low magnitude seismic activity predictions in Cyprus. Soil Dyn. Earthq. Eng. 2020, 130, 105932. [CrossRef] 35. Vasti, M.; Dev, A. Classification and Analysis of Real-World Earthquake Data Using Various Machine Learning Algorithms. In Lecture Notes in Electrical Engineering; Springer: Singapore, 2019; pp. 1–14. http://dx.doi.org/10.1111/mice.12335 http://dx.doi.org/10.1016/j.knosys.2013.06.011 http://dx.doi.org/10.1371/journal.pone.0146101 http://dx.doi.org/10.1016/j.knosys.2014.11.024 http://dx.doi.org/10.1007/s10844-014-0316-5 http://dx.doi.org/10.1007/s10661-018-6837-6 http://dx.doi.org/10.1007/s11069-018-3449-y http://dx.doi.org/10.1007/s10064-015-0786-x http://dx.doi.org/10.7763/IJET.2014.V6.692 http://dx.doi.org/10.1016/j.geog.2017.05.003 http://dx.doi.org/10.1007/s11069-014-1264-7 http://dx.doi.org/10.1007/s40815-018-0483-6 http://dx.doi.org/10.1016/j.soildyn.2019.105932 Sustainability 2020, 12, 2420 29 of 32 36. Mukhopadhyay, U.K.; Sharma, R.N.K.; Anwar, S.; Dutta, A.D. Correlating Thermal Anomaly with Earthquake Occurrences Using Remote Sensing; Springer: Berlin/Heidelberg, Germany, 2019; pp. 863–875. 37. Karimzadeh, S.; Matsuoka, M.; Kuang, J.; Ge, L. Spatial Prediction of Aftershocks Triggered by a Major Earthquake: A Binary Machine Learning Perspective. ISPRS Int. J. Geo-Inf. 2019, 8, 462. [CrossRef] 38. Zhou, Z.; Lin, Y.; Zhang, Z.; Wu, Y.; Johnson, P. Earthquake Detection in 1D Time—Series Data with Feature Selection and Dictionary Learning. Seism. Res. Lett. 2019, 90, 563–572. [CrossRef] 39. Corbi, F.; Sandri, L.; Bedford, J.; Funiciello, F.; Brizzi, S.; Rosenau, M.; Lallemand, S. Machine Learning Can Predict the Timing and Size of Analog Earthquakes. Geophys. Res. Lett. 2019, 46, 1303–1311. [CrossRef] 40. Kong, Q.; Trugman, D.T.; Ross, Z.E.; Bianco, M.; Meade, B.J.; Gerstoft, P. Machine Learning in Seismology: Turning Data into Insights. Seism. Res. Lett. 2018, 90, 3–14. [CrossRef] 41. Galkina, A.; Grafeeva, N. Machine Learning Methods for Earthquake Prediction: A Survey. In Proceedings of the Fourth Conference on Software Engineering and Information Management (SEIM-2019), Saint Petersburg, Russia, 13 April 2019; full papers. p. 25. 42. Gitis, V.G.; Derendyaev, A. Machine Learning Methods for Seismic Hazards Forecast. Geosciences 2019, 9, 308. [CrossRef] 43. Al-Najjar, H.A.H.; Kalantar, B.; Pradhan, B.; Saeidi, V. Conditioning factor determination for mapping and prediction of landslide susceptibility using machine learning algorithms. In Earth Resources and Environmental Remote Sensing/GIS Applications X; International Society for Optics and Photonics: California, CA, USA, 2019; Volume 11156, p. 111560K. 44. Ganter, T.; Sundermier, A.; Ballard, S. Alternate Null Hypothesis Correlation: A New Approach to Automatic Seismic Event Detection. Bull. Seism. Soc. Am. 2018, 108, 3528–3547. [CrossRef] 45. Mosavi, A.; Salimi, M.; Ardabili, S.F.; Rabczuk, T.; Shamshirband, S.; Várkonyi-Kóczy, A.R. State of the Art of Machine Learning Models in Energy Systems, a Systematic Review. Energies 2019, 12, 1301. [CrossRef] 46. Mosavi, A.; Ozturk, P.; Chau, K.-W. Flood Prediction Using Machine Learning Models: Literature Review. Water 2018, 10, 1536. [CrossRef] 47. Dineva, A.; Mosavi, A.; Ardabili, S.F.; Vajda, I.; Shamshirband, S.; Rabczuk, T.; Chau, K.-W. Review of Soft Computing Models in Design and Control of Rotating Electrical Machines. Energies 2019, 12, 1049. [CrossRef] 48. Nosratabadi, S.; Mosavi, A.; Shamshirband, S.; Zavadskas, E.K.; Rakotonirainy, A.; Chau, K.-W. Sustainable Business Models: A Review. Sustainability 2019, 11, 1663. [CrossRef] 49. Zhang, L.; Si, L.; Yang, H.; Hu, Y.; Qiu, J. Precursory Pattern Based Feature Extraction Techniques for Earthquake Prediction. IEEE Access 2019, 7, 30991–31001. [CrossRef] 50. Kayastha, P.; Bijukchhen, S.; Dhital, M.R.; De Smedt, F. GIS based landslide susceptibility mapping using a fuzzy logic approach: A case study from Ghurmi-Dhad Khola area, Eastern Nepal. J. Geol. Soc. India 2013, 82, 249–261. [CrossRef] 51. Lu, J.; Hu, S.; Niu, Z.; You, R. The Application of Fuzzy Comprehensive Evaluation Model in Landslide Prediction. In Proceedings of the 2010 3rd International Conference on Information Management, Innovation Management and Industrial Engineering, Kunming, China, 26–28 November 2010; Volume 4, pp. 612–615. 52. Mallick, J.; Singh, R.K.; Alawadh, M.A.; Islam, S.; Khan, R.A.; Qureshi, M.N. GIS-based landslide susceptibility evaluation using fuzzy-AHP multi-criteria decision-making techniques in the Abha Watershed, Saudi Arabia. Environ. Earth Sci. 2018, 77, 276. [CrossRef] 53. Mohsin, S.; Azam, F. Computational seismic algorithmic comparison for earthquake prediction. Int. J. Geol. 2011, 5, 53–59. 54. Sengar, S.S.; Kumar, A.; Ghosh, S.K.; Wason, H.R.; Raju, P.L.N.; Murthy, Y.V.N.K. Earthquake-induced built-up damage identification using fuzzy approach. Geomat. Nat. Hazards Risk 2013, 4, 320–338. [CrossRef] 55. Sun, D.; Sun, B. Rapid prediction of earthquake damage to buildings based on fuzzy analysis. In Proceedings of the 2010 Seventh International Conference on Fuzzy Systems and Knowledge Discovery, Yantai, China, 10–12 August 2010; Volume 3, pp. 1332–1335. 56. Huang, L.; Xiang, L.-Y. Method for Meteorological Early Warning of Precipitation-Induced Landslides Based on Deep Neural Network. Neural Process. Lett. 2018, 48, 1243–1260. [CrossRef] 57. Li, W.; Narvekar, N.; Nakshatra, N.; Raut, N.; Sirkeci, B.; Gao, J. Seismic Data Classification Using Machine Learning. In Proceedings of the 2018 IEEE Fourth International Conference on Big Data Computing Service and Applications (BigDataService), Bamberg, Germany, 26–29 March 2018; pp. 56–63. http://dx.doi.org/10.3390/ijgi8100462 http://dx.doi.org/10.1785/0220180315 http://dx.doi.org/10.1029/2018GL081251 http://dx.doi.org/10.1785/0220180259 http://dx.doi.org/10.3390/geosciences9070308 http://dx.doi.org/10.1785/0120180074 http://dx.doi.org/10.3390/en12071301 http://dx.doi.org/10.3390/w10111536 http://dx.doi.org/10.3390/en12061049 http://dx.doi.org/10.3390/su11061663 http://dx.doi.org/10.1109/ACCESS.2019.2902224 http://dx.doi.org/10.1007/s12594-013-0147-y http://dx.doi.org/10.1007/s12665-018-7451-1 http://dx.doi.org/10.1080/19475705.2012.746242 http://dx.doi.org/10.1007/s11063-017-9778-0 Sustainability 2020, 12, 2420 30 of 32 58. Asim, K.M.; Idris, A.; Iqbal, T.; Martínez-Álvarez, F. Earthquake prediction model using support vector regressor and hybrid neural networks. PLoS ONE 2018, 13, e0199004. [CrossRef] 59. Asencio–Cortés, G.; Morales-Esteban, A.; Shang, X.; Martínez–Álvarez, F. Earthquake prediction in California using regression algorithms and cloud-based big data infrastructure. Comput. Geosci. 2018, 115, 198–210. [CrossRef] 60. Hoang, N.-D.; Bui, D.T.T. Predicting earthquake-induced soil liquefaction based on a hybridization of kernel Fisher discriminant analysis and a least squares support vector machine: A multi-dataset study. Bull. Int. Assoc. Eng. Geol. 2016, 77, 191–204. [CrossRef] 61. Gitis, V.G.; Derendyaev, A. Web-Based GIS platform for automatic prediction of earthquakes. In Proceedings of the International Conference on Computational Science and Its Applications, Melbourne, Australia, 2–5 July 2018; Springer: Berlin/Heidelberg, Germany, 2018; pp. 268–283. 62. Thomas, S.; Pillai, G.; Pal, K. Prediction of peak ground acceleration using �-SVR, ν-SVR and Ls-SVR algorithm. Geomatics, Nat. Hazards Risk 2016, 8, 177–193. [CrossRef] 63. Rouet-Leduc, B.; Hulbert, C.; Lubbers, N.; Barros, K.; Humphreys, C.J.; Johnson, P.A. Machine Learning Predicts Laboratory Earthquakes. Geophys. Res. Lett. 2017, 44, 9276–9282. [CrossRef] 64. Rafiei, M.H.; Adeli, H. NEEWS: A novel earthquake early warning model using neural dynamic classification and neural dynamic optimization. Soil Dyn. Earthq. Eng. 2017, 100, 417–427. [CrossRef] 65. Asencio–Cortés, G.; Scitovski, S.; Scitovski, R.; Martínez–Álvarez, F. Temporal analysis of croatianseismogenic zones to improve earthquake magnitude prediction. Earth Sci. Inform. 2017, 10, 303–320. [CrossRef] 66. Rahmani, M.E.; Amine, A.; Hamou, R.M. A Novel Bio Inspired Algorithm Based on Echolocation Mechanism of Bats for Seismic States Prediction. Int. J. Swarm Intell. Res. 2017, 8, 1–18. [CrossRef] 67. Asencio-Cortés, G.; Martínez-Álvarez, F.; Troncoso, A.; Morales-Esteban, A. Medium–large earthquake magnitude prediction in Tokyo with artificial neural networks. Neural Comput. Appl. 2015, 28, 1043–1055. [CrossRef] 68. Asim, K.M.; Idris, A.; Martinez-Alvarez, F.; Iqbal, T. Short Term Earthquake Prediction in Hindukush Region Using Tree Based Ensemble Learning. In Proceedings of the 2016 International Conference on Frontiers of Information Technology (FIT), Islamabad, Pakistan, 19–21 December 2016; pp. 365–370. 69. Yang, D.; Yang, K. Multi-step prediction of strong earthquake ground motions and seismic responses of SDOF systems based on EMD-ELM method. Soil Dyn. Earthq. Eng. 2016, 85, 117–129. [CrossRef] 70. Vahaplar, A.; Tezel, B.T.; Nasiboglu, E.; Nasibov, E. A monitoring system to prepare machine learning data sets for earthquake prediction based on seismic-acoustic signals. In Proceedings of the 2015 9th International Conference on Application of Information and Communication Technologies (AICT), Rostov on Don, Russia, 14–16 October 2015; pp. 44–47. 71. Buscema, P.M.; Massini, G.; Maurelli, G. Artificial Adaptive Systems to predict the magnitude of earthquakes. Bollettino di GeofisicaTeorica ed Applicata 2015, 56, 227–256. 72. Kamogawa, M.; Nanjo, K.; Izutsu, J.; Orihara, Y.; Nagao, T.; Uyeda, S. Nucleation and Cascade Features of Earthquake Mainshock Statistically Explored from Foreshock Seismicity. Entropy 2019, 21, 421. [CrossRef] 73. Reyes, J.; Morales-Esteban, A.; Martínez-Álvarez, F. Neural networks to predict earthquakes in Chile. Appl. Soft Comput. 2013, 13, 1314–1328. [CrossRef] 74. Farrokhzad, F.; Choobbasti, A.; Barari, A. Liquefaction microzonation of Babol city using artificial neural network. J. King Saud Univ.-Sci. 2012, 24, 89–100. [CrossRef] 75. Gu, T.-F.; Wang, J.-D. Application of fuzzy neural networks for predicting seismic subsidence coefficient of loess subgrade. In Proceedings of the 2010 Sixth International Conference on Natural Computation, Yantai, China, 10–12 August 2010; Volume 3, pp. 1556–1559. 76. Korkmaz, K.A.; Demir, F. Ground Motion Data Profile of Western Turkey with Intelligent Hybrid Processing. Pure Appl. Geophys. 2016, 174, 293–303. [CrossRef] 77. Dutta, P.K.; Mishra, O.P.; Naskar, M.K. Decision analysis for earthquake prediction methodologies: Fuzzy inference algorithm for trust validation. Int. J. Comput. Appl. 2012, 45, 13–20. 78. Mishra, O.P. Seismological research in India. Proc. Indian Natl. Sci. Acad. 2012, 76, 361–375. 79. Pandit, A.R.; Biswal, K.C. Prediction of earthquake magnitude using adaptive neuro fuzzy inference system. Earth Sci. Inform. 2019, 12, 513–524. [CrossRef] 80. Pham, B.T.; Bui, D.T.T.; Pham, H.V.; Le, H.Q.; Prakash, I.; Dholakia, M.B. Landslide Hazard Assessment Using Random SubSpace Fuzzy Rules Based Classifier Ensemble and Probability Analysis of Rainfall Data: http://dx.doi.org/10.1371/journal.pone.0199004 http://dx.doi.org/10.1016/j.cageo.2017.10.011 http://dx.doi.org/10.1007/s10064-016-0924-0 http://dx.doi.org/10.1080/19475705.2016.1176604 http://dx.doi.org/10.1002/2017GL074677 http://dx.doi.org/10.1016/j.soildyn.2017.05.013 http://dx.doi.org/10.1007/s12145-017-0295-5 http://dx.doi.org/10.4018/IJSIR.2017070101 http://dx.doi.org/10.1007/s00521-015-2121-7 http://dx.doi.org/10.1016/j.soildyn.2016.03.015 http://dx.doi.org/10.3390/e21040421 http://dx.doi.org/10.1016/j.asoc.2012.10.014 http://dx.doi.org/10.1016/j.jksus.2010.09.003 http://dx.doi.org/10.1007/s00024-016-1379-8 http://dx.doi.org/10.1007/s12145-019-00397-w Sustainability 2020, 12, 2420 31 of 32 A Case Study at Mu Cang Chai District, Yen Bai Province (Viet Nam). J. Indian Soc. Remote. Sens. 2016, 45, 673–683. [CrossRef] 81. Polykretis, C.; Chalkias, C.; Ferentinou, M. Adaptive neuro-fuzzy inference system (ANFIS) modeling for landslide susceptibility assessment in a Mediterranean hilly area. Bull. Int. Assoc. Eng. Geol. 2017, 78, 1173–1187. [CrossRef] 82. Razifard, M.; Shoaei, G.; Zaré, M. Application of fuzzy logic in the preparation of hazard maps of landslides triggered by the twin Ahar-Varzeghan earthquakes (2012). Bull. Int. Assoc. Eng. Geol. 2018, 78, 223–245. [CrossRef] 83. Pourghasemi, H.R.; Pradhan, B.; Gokceoglu, C. Application of fuzzy logic and analytical hierarchy process (AHP) to landslide susceptibility mapping at Haraz watershed, Iran. Nat. Hazards 2012, 63, 965–996. [CrossRef] 84. Pradhan, B. Use of GIS-based fuzzy logic relations and its cross application to produce landslide susceptibility maps in three test areas in Malaysia. Environ. Earth Sci. 2010, 63, 329–349. [CrossRef] 85. Zhang, Y.; Oldenburg, C.M.; Finsterle, S. Percolation-theory and fuzzy rule-based probability estimation of fault leakage at geologic carbon sequestration sites. Environ. Earth Sci. 2009, 59, 1447–1459. [CrossRef] 86. Meten, M.; Bhandary, N.P.; Yatabe, R. Application of GIS-based fuzzy logic and rock engineering system (RES) approaches for landslide susceptibili mapping in Selelkula area of the Lower Jema River Gorge, Central Ethiopia. Environ, Earth Sci. 2015, 74, 3395–3416. [CrossRef] 87. Thomas, S.; Pillai, G.N.; Pal, K.; Jagtap, P. Prediction of ground motion parameters using randomized ANFIS (RANFIS). Appl. Soft Comput. 2016, 40, 624–634. [CrossRef] 88. Wu, A. Design and practice of a digital seismic waveform analyzing tool. In Proceedings of the 2012 9th International Conference on Fuzzy Systems and Knowledge Discovery, Chongqing, China, 29–31 May 2012; pp. 2299–2303. 89. Ismail, N.; Khattak, N. Reconnaissance Report on the Mw 7.5 Hindu Kush Earthquake of 26th October 2015 and the Subsequent Aftershocks; United Arab Emirates University: Al Ain, UAE, 2015. 90. Mignan, A. Retrospective on the Accelerating Seismic Release (ASR) hypothesis: Controversy and new horizons. Tectonophysics 2011, 505, 1–16. [CrossRef] 91. Keilis-Borok, V. Earthquake prediction: State-of-the-art and emerging possibilities. Annu. Rev. Earth Planet. Sci. 2002, 30, 1–33. [CrossRef] 92. Mignan, A. The debate on the prognostic value of earthquake foreshocks: A meta-analysis. Sci. Rep. 2014, 4, 4099. [CrossRef] [PubMed] 93. Wyss, M. Evaluation of proposed earthquake precursors. EOS Trans. Am. Geophys. Union 1991, 72, 411. [CrossRef] 94. Mignan, A. A preliminary text classification of the precursory accelerating seismicity corpus: Inference on some theoretical trends in earthquake predictability research from 1988 to 2018. J. Seism. 2019, 23, 771–785. [CrossRef] 95. Cicerone, R.D.; Ebel, J.E.; Britton, J. A systematic compilation of earthquake precursors. Tectonophysics 2009, 476, 371–396. [CrossRef] 96. Azimi, Y.; Khoshrou, S.H.; Osanloo, M. Prediction of blast induced ground vibration (BIGV) of quarry mining using hybrid genetic algorithm optimized artificial neural network. Measurement 2019, 147, 106874. [CrossRef] 97. Mehrgini, B.; Izadi, H.; Memarian, H. Shear wave velocity prediction using Elman artificial neural network. Carbonates Evaporites 2017, 34, 1281–1291. [CrossRef] 98. Udegbe, E.; Morgan, E.; Srinivasan, S. Big data analytics for seismic fracture identification using amplitude-based statistics. Comput. Geosci. 2019, 23, 1277–1291. [CrossRef] 99. Sharma, S.; Venkateswarlu, H.; Hegde, A. Application of Machine Learning Techniques for Predicting the Dynamic Response of Geogrid Reinforced Foundation Beds. Geotech. Geol. Eng. 2019, 37, 4845–4864. [CrossRef] 100. Fanos, A.M.; Pradhan, B. A Novel Hybrid Machine Learning-Based Model for Rockfall Source Identification in Presence of Other Landslide Types Using LiDAR and GIS. Earth Syst. Environ. 2019, 3, 491–506. [CrossRef] 101. Břizová, L.; Kříž, J.; Studnička, F.; Slegr, J. Methods for the Evaluation of the Stochastic Properties of the Ionosphere for Earthquake Prediction—Random Matrix Theory. Atmosphere 2019, 10, 413. [CrossRef] http://dx.doi.org/10.1007/s12524-016-0620-3 http://dx.doi.org/10.1007/s10064-017-1125-1 http://dx.doi.org/10.1007/s10064-018-1235-4 http://dx.doi.org/10.1007/s11069-012-0217-2 http://dx.doi.org/10.1007/s12665-010-0705-1 http://dx.doi.org/10.1007/s12665-009-0131-4 http://dx.doi.org/10.1007/s12665-015-4377-8 http://dx.doi.org/10.1016/j.asoc.2015.12.013 http://dx.doi.org/10.1016/j.tecto.2011.03.010 http://dx.doi.org/10.1146/annurev.earth.30.100301.083856 http://dx.doi.org/10.1038/srep04099 http://www.ncbi.nlm.nih.gov/pubmed/24526224 http://dx.doi.org/10.1029/90EO10300 http://dx.doi.org/10.1007/s10950-019-09833-2 http://dx.doi.org/10.1016/j.tecto.2009.06.008 http://dx.doi.org/10.1016/j.measurement.2019.106874 http://dx.doi.org/10.1007/s13146-017-0406-x http://dx.doi.org/10.1007/s10596-019-09890-z http://dx.doi.org/10.1007/s10706-019-00945-7 http://dx.doi.org/10.1007/s41748-019-00114-z http://dx.doi.org/10.3390/atmos10070413 Sustainability 2020, 12, 2420 32 of 32 102. Andrén, M.; Stockmann, G.; SkeltoniD, A.; Sturkell, E.; Mörth, C.; Guðrúnardóttir, H.R.; Keller, N.S.; Odling, N.; Dahrén, B.; Broman, C.; et al. Coupling between mineral reactions, chemical changes in groundwater, and earthquakes in Iceland. J. Geophys. Res. Solid Earth 2016, 121, 2315–2337. [CrossRef] 103. Sarlis, N.; Skordas, E.S.; Varotsos, P. Natural Time Analysis: Results Related to Two Earthquakes in Greece during 2019. Proceedings 2019, 24, 20. 104. Tareen, A.D.K.; Asim, K.M.; Kearfott, K.; Rafique, M.; Nadeem, M.S.A.; Iqbal, T.; Rahman, S.U. Automated anomalous behaviour detection in soil radon gas prior to earthquakes using computational intelligence techniques. J. Environ. Radioact. 2019, 203, 48–54. [CrossRef] [PubMed] 105. Orihara, Y.; Kamogawa, M.; Nagao, T. Preseismic Changes of the Level and Temperature of Confined Groundwater related to the 2011 Tohoku Earthquake. Sci. Rep. 2014, 4, 6907. [CrossRef] 106. SkeltoniD, A.; Andrén, M.; Kristmannsdóttir, H.; Stockmann, G.; Mörth, C.-M.; Sveinbjörnsdóttir, A.; Jónsson, S.; Sturkell, E.; Guðrúnardóttir, H.R.; Hjartarson, H.; et al. Changes in groundwater chemistry before two consecutive earthquakes in Iceland. Nat. Geosci. 2014, 7, 752–756. [CrossRef] © 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/). http://dx.doi.org/10.1002/2015JB012614 http://dx.doi.org/10.1016/j.jenvrad.2019.03.003 http://www.ncbi.nlm.nih.gov/pubmed/30861489 http://dx.doi.org/10.1038/srep06907 http://dx.doi.org/10.1038/ngeo2250 http://creativecommons.org/ http://creativecommons.org/licenses/by/4.0/. Introduction Background Fuzzy Expert System (FES) Rule Based Expert System (RBES) Neuro Fuzzy Expert System (NFES) Machine Learning (Ml) Research Methodology Defining Research Questions Search and Selection Strategy Identification of Search String Screening and Selection Criteria Data Extraction and Synthesis Analysis Basic Analysis Key Facts of Expert System Based Earthquake Prediction Publications ResearchType Facets Addressed by the Identified Publications ES type and System Specific Key Aspects of Proposed ES Quality Assessment Discussion Comparitive Analysis of Methods Principal Findings Evolution of Tools and Techniques Conclusion and Future Directions References 
work_5vwgkea44rbdzmzx5o2wmh3qai	----	AUTHOR(S): TITLE: YEAR: Publisher citation: OpenAIR citation: Publisher copyright statement: OpenAIR takedown statement: This publication is made freely available under ________ open access. This is the ______________________ version of an article originally published by ____________________________ in __________________________________________________________________________________________ (ISSN _________; eISSN __________). This publication is distributed under a CC ____________ license. ____________________________________________________ Section 6 of the “Repository policy for OpenAIR @ RGU” (available from http://www.rgu.ac.uk/staff-and-current- students/library/library-policies/repository-policies) provides guidance on the criteria under which RGU will consider withdrawing material from OpenAIR. If you believe that this item is subject to any of these criteria, or for any other reason should not be held on OpenAIR, then please contact openair-help@rgu.ac.uk with the details of the item and the nature of your complaint. Improving e-Learning Recommendation by using Background Knowledge Blessing Mbipom Susan Craw Stewart Massie School of Computing Science & Digital Media Robert Gordon University, Aberdeen, UK b.e.mbipom/s.craw/s.massie@rgu.ac.uk Abstract There is currently a large amount of e-Learning resources available to learners on the Web. How- ever, learners often have difficulty finding and retrieving relevant materials to support their learning goals because they lack the domain knowledge to craft effective queries that convey what they wish to learn. In addition, the unfamiliar vocabulary often used by domain experts makes it difficult to map a learner’s query to a relevant learning material. We address these challenges by introducing an innovative method that automatically builds background knowledge for a learning domain. In cre- ating our method, we exploit a structured collection of teaching materials as a guide for identifying the important domain concepts. We enrich the identified concepts with discovered text from an en- cyclopedia, thereby increasing the richness of our acquired knowledge. We employ the developed background knowledge for influencing the representation and retrieval of learning resources to im- prove e-Learning recommendation. The effectiveness of our method is evaluated using a collection of Machine Learning and Data Mining papers. Our method outperforms the benchmark, demonstrating the advantage of using background knowledge for improving the representation and recommendation of e-Learning materials. 1 Introduction Learning-focused content is increasingly available on the Web, thus providing an excellent source of information for building e-Learning systems (Clarà and Barberà, 2013). However, learners often have difficulty finding the right learning materials because they lack the domain knowledge required to formulate effective queries (Chen et al., 2014). In addition, a mismatch in the vocabulary used by learners when crafting their queries and that used by domain experts to describe learning concepts poses a further challenge for systems recommending resources to learners. Another challenge with e-Learning recommendation is that the learning resources are often un- structured text, and so are not properly indexed for retrieval (Nasraoui and Zhuhadar, 2010). The challenge of dealing with unstructured learning resources creates a difficulty in finding and retrieving relevant learning resources. Hence the need for an effective method of representing learning materials with the aim of improving recommendation. This paper proposes the automated acquisition of background knowledge about a domain that can then be employed for enhancing e-Learning recommendation. In our method, we create a concept- aware representation that contains a good coverage of relevant topics from the domain. First, we exploit a structured collection of teaching materials as a guide for identifying the important concepts. Next, we enrich the identified concepts with discovered text from an encyclopedia source, thereby increasing the richness of our representation. Our developed method is demonstrated in Machine Learning and Data Mining, although the method we present can be applied to learning materials in other domains. Other projects such as DeepQA (Ferrucci et al., 2013) and DBpedia (Lehmann et al., 2015) use a range of knowledge-rich representations to enhance retrieval. Such knowledge-rich sources are usu- ally in the form of important topics that describe a domain. While these projects generally rely on handcrafted knowledge sources, they highlight the advantage in exploiting knowledge-rich represen- tations as a basis for improving recommendation. A good coverage of domain topics is useful for representing learning materials. These domain topics contain rich vocabulary and provide a good knowledge source for mapping learners’ queries 1 to learning materials. Thus allowing us to address the mismatch in the vocabulary used by learners and domain experts. We address this issue by introducing a method that automatically creates custom background knowledge in the form of a rich set of domain topics. Further, we explore building a richer vocabulary to achieve a better coverage of the domain, and this method is employed to improve e-Learning recommendation. We make several contributions in this work. Firstly, the creation of background knowledge for an e-Learning domain. We describe how we take advantage of the knowledge of experts contained in e-Books to build a knowledge-rich representation that is used to enhance recommendation. Secondly, we present a method that harnesses the developed background knowledge to augment the represen- tation of learning resources in order to improve e-Learning recommendation. Finally, we explore a larger concept vocabulary which provides a better coverage of the domain. We refine our method pre- sented in (Mbipom et al., 2016) to generate a richer and focused set of domain concepts. The results from our evaluation show the improvement in e-Learning recommendation when the richer concept vocabulary is used for representing learning resources. The rest of this paper is organised as follows. In Section 2 we present related text representation approaches that underpin this work. Section 3 describes the development of our background knowl- edge using available knowledge sources. Section 4 discusses the representation of learning resources using our methods. Then Section 5 presents the evaluation of the learning resource representation. In Section 6 we present our refined method of generating background knowledge with an evaluation using the richer vocabulary and a larger dataset for recommendation. Finally, Section 7 presents our conclusions. 2 Related Work E-Learning recommendation is challenging because learning resources are often unstructured text, and so are not properly indexed for retrieval. A possible solution to addressing this challenge is the creation of effective representations that capture the content of learning resources. However, building suitable representations for learning resources in e-Learning environments is not easy (Dietze et al., 2012), as the resources do not have a pre-defined set of features by which they can be indexed. We propose the creation of a knowledge-rich representation that captures the domain-specific vocabulary contained in learning resources. Figure 1 illustrates two broad approaches often used to address the challenge of text representation. These are corpus-based methods, such as topic models (Blei and McAuliffe, 2007; Chen and Liu, 2014); and structured representations, such as those that take advantage of ontologies (Boyce and Pahl, 2007; Yarandi et al., 2011). In Figure 1, the lower row of items identifies various knowledge sources that can be employed to build a range of knowledge- light to knowledge-rich text representation approaches. Corpus-based methods usually involve the use of statistical models to identify topics from a cor- pus. The identified topics are often keywords (Beliga et al., 2015; Matsuo and Ishizuka, 2004) or phrases (Coenen et al., 2007; Witten et al., 1999). Coenen et al. showed that using a combination of keywords and phrases was better than using only keywords (Coenen et al., 2007). These topics can be extracted from different text sources such as: learning resources (Rodrigues et al., 2007; Yang et al., 2016), metadata e.g. Tables of contents (Bousbahi and Chorfi, 2015), and Encyclopedia e.g. Wikipedia (Milne and Witten, 2008; Qureshi et al., 2014). A drawback of the corpus-based methods is that, they normally rely on the coverage of the document collection used, so the topics produced may not be representative of the learning domain. Figure 1: Two broad approaches used for text representation Structured representations capture relationships between important domain concepts. This often entails using an existing ontology e.g ACM taxonomy (Nasraoui and Zhuhadar, 2010; Ruiz-Iniesta et al., 2014), or creating a new one (Gherasim et al., 2013; Panagiotis et al., 2016). Although ontolo- gies are designed to have a good coverage of their domains, the output is still dependent on the view 2 of its builders and, because of handcrafting, existing ontologies cannot easily be adapted to new do- mains. e-Learning is dynamic because new resources are becoming available regularly, and so using fixed ontologies limits the potential to incorporate new content. The approach adopted in this paper draws insight from both the corpus-based methods and struc- tured representations highlighted in Figure 1. We leverage on a structured corpus of teaching materials such as Tables of contents of e-Books, in order to identify important topics in an e-Learning domain. These topics are a combination of keywords and phrases as recommended in (Coenen et al., 2007). The identified topics are then enriched with discovered text from Wikipedia in order to enhance our representation. In addition, we refine the methods developed in previous work (Mbipom et al., 2016) so that we can generate a richer set of relevant topics that provide a good coverage of the learning do- main. Consequently, our approach is employed to influence the representation and retrieval of relevant learning resources. 3 Creation of Background Knowledge Background knowledge refers to information about a domain that is useful for general understanding and problem-solving (Zhang et al., 2013). We attempt to capture background knowledge as a set of domain concepts, each representing an important topic in the domain. For example, in a learning domain, such as Machine Learning, you would find topics such as Classification, Clustering and Regression. Each of these topics would be represented by a concept, in the form of a concept label and a pseudo-document which describes the concept. The concepts can then be used to underpin the representation of e-Learning resources. Our knowledge extraction process is shown in Figure 2. The input to this process are domain knowledge sources, and we use a structured collection of teaching materials and an encyclopedia source. Next, ngrams are automatically extracted from our structured collection to generate a set of potential concept labels. Then a domain lexicon is used to validate the extracted ngrams to ensure that the ngrams are also being used in another information source. The encyclopedia provides text descriptions for the identified ngrams. The output from this process is a set of domain concepts, each having a concept label and an associated pseudo-document. We discuss the stages of the background knowledge creation in the following sections. Figure 2: An overview of the background knowledge creation process 3.1 Knowledge Sources Two knowledge sources are used as initial inputs for discovering concept labels. A structured col- lection of teaching materials provides a source for extracting important topics identified by teaching experts in the domain, while a domain lexicon provides a broader but more detailed coverage of the relevant topics in the domain. The lexicon is used to verify that the concept labels identified from the 3 teaching materials are directly relevant. Thereafter, an encyclopedia source, such as Wikipedia pages, is searched and provides the relevant text to form a pseudo-document for each verified concept label. The final output from this process is our set of domain concepts each comprising a concept label and an associated pseudo-document. Our approach is demonstrated with learning resources from Machine Learning and Data Mining. We use e-Books as our collection of teaching materials; a summary of the books used is shown in Table 1. Two Google Scholar queries: “Introduction to data mining textbook” and “Introduction to machine learning textbook” guided the selection process, and 20 e-Books that met all of the following 3 criteria were chosen. Firstly, the book should be about the domain. Secondly, there should be Google Scholar citations for the book. Thirdly, the book should be accessible. We use the Tables-of-Contents (TOCs) of the books as our structured knowledge source. We use Wikipedia to create our domain lexicon because it contains articles for many learning domains (Völkel et al., 2006; Zheng et al., 2010), and the contributions of many people (Yang and Lai, 2010), so this provides the coverage we need in our lexicon. The lexicon is generated from 2 Wikipedia sources. First, the phrases in the contents and overview sections of the chosen domain are extracted to form a topic list. Then, a list with the titles of articles related to the domain is added to the topic list to assemble our lexicon. Overall, our domain lexicon contains a set of 664 Wiki-phrases. Table 1: Summary of e-Books used Book Title & Author Cites Machine learning; Mitchell 264 Introduction to machine learning; Alpaydin 2621 Machine learning a probabilistic perspective; Murphy 1059 Introduction to machine learning; Kodratoff 159 Gaussian processes for machine learning; Rasmussen & Williams 5365 Introduction to machine learning; Smola & Vishwanathan 38 Machine learning, neural and statistical classification; Michie, Spiegelhalter, & Taylor 2899 Introduction to machine learning; Nilsson 155 A First Encounter with Machine Learning; Welling 7 Bayesian reasoning and machine learning; Barber 271 Foundations of machine learning; Mohri, Rostamizadeh, & Talwalkar 197 Data mining-practical machine learning tools and techniques; Witten & Frank 27098 Data mining concepts models and techniques; Gorunescu 244 Web data mining; Liu 1596 An introduction to data mining; Larose 1371 Data mining concepts and techniques; Han & Kamber 22856 Introduction to data mining; Tan, Steinbach, & Kumar 6887 Principles of data mining; Bramer 402 Introduction to data mining for the life sciences; Sullivan 15 Data mining concepts methods and applications; Yin, Kaku, Tang, & Zhu 23 3.2 Generating Potential Domain Concepts In the first stage of the process, the text from the TOCs is pre-processed. We remove punctuations, symbols, and numbers from the TOCs, so that only words are used for generating concept labels. After this, we remove 2 sets of stopwords. First, a standard English stopwords list, which allows us to remove common words and still retain a good set of words for generating our concept labels. Second, an additional set of words which we refer to as TOC-stopwords are removed. It contains: structural words, such as chapter and appendix, which relate to the structure of the TOCs; roman numerals, such as xxiv and xxxv, which are used to indicate the sections in a TOC; and words, such as introduction and conclusion, which describe parts of a learning material and are generic across domains. In addition, words referring directly to the name of the domain used for demonstration are removed, as we wish to generate concepts that describe the domain. We do not use stemming because we found it harmful during pre-processing. When searching an encyclopedia source with the stemmed form of words, relevant results would not be returned. The output from pre-processing is a set of TOC phrases. In the next stage, we apply ngram extraction to the TOC phrases to generate all 1-3 grams from the entire set of TOC phrases. The output from this process are TOC-ngrams containing a set of 2038 unigrams, 5405 bigrams and 6133 trigrams, which 4 are used as the potential domain concept labels. Many irrelevant ngrams are generated from the TOCs because we have simply selected all 1-3 grams. 3.3 Verifying Concept Labels using Domain Lexicon A domain lexicon is used to verify the generated TOC-ngrams to confirm which of the ngrams are relevant for the domain. Our domain lexicon contains a set of 664 Wiki-phrases, each of which is pre-processed by removing non-alphanumeric characters. The distribution of Wiki-phrases is shown in Figure 3. The 84% of the Wiki-phrases that are 1-3 grams are used for verification. The comparison of TOC-ngrams with the domain lexicon identifies the potential domain concept labels that are actu- ally being used to describe aspects of the chosen domain in Wikipedia. During verification, ngrams referring directly to the title of the domain, e.g. machine learning and data mining, are not included in the Wiki-phrases because our aim is to generate concept labels that describe specific topics within the domain. Overall, a set of 17 unigrams, 58 bigrams and 15 trigrams are verified as potential concept labels. Bigrams yield the highest number of ngrams, which indicates that bigrams are particularly useful for describing topics in this domain. Figure 3: Distribution of Wiki-phrases used for verifying concept labels 3.4 Domain Concept Generation Our domain concepts are generated after a second verification step is applied to the ngrams returned from the previous stage. Each ngram is retained as a concept label if all of 3 criteria are met. Firstly, if a Wikipedia page describing the ngram exists. Secondly, if the text describing the ngram is not contained as part of the page describing another ngram. Thirdly, if the ngram is not a synonym of another ngram. For the third criteria, if two ngrams are synonyms, the ngram with the higher frequency is retained as a concept label while its synonym is retained as part of the extracted text. For example, 2 ngrams cluster analysis and clustering are regarded as synonyms in Wikipedia, so the text associated with them is the same. The label clustering is retained as the concept label because it occurs more frequently in the TOCs, and its synonym, cluster analysis is contained as part of the discovered text. The concept labels are used to search Wikipedia pages in order to generate a description for the identified concept label. The search returns discovered text that forms a pseudo-document which in- cludes the concept label. So, the concept label and pseudo-document pair make up a domain concept. Overall, 73 domain concepts are generated. Each pseudo-document is pre-processed using standard techniques of English stopwords removal and Porter stemming (Porter, 1980). The pseudo-document terms form the concept vocabulary that can be used to represent resources. 4 Representing Learning Resources Using Background Knowledge Our background knowledge contains a rich representation of the learning domain and by harnessing this knowledge for representing learning resources, we expect to retrieve documents based on the do- 5 main concepts that they contain. These concepts are designed to be effective for e-Learning, because they are assembled from TOCs of teaching materials (Agrawal et al., 2012). We present two ap- proaches which have been developed by employing our background knowledge in the representation of learning resources. 4.1 The CONCEPTBASED Document Representation approach Representing documents with the concept vocabulary allows retrieval to focus on the concepts con- tained in the documents. Figures 4 & 5 illustrate the CONCEPTBASED method. Firstly, in Figure 4, the concept vocabulary, t1 ... tc, from the pseudo-documents of concepts, C1 ... Cm, is used to create a term-concept matrix and a term-document matrix using TF-IDF weighting (Salton and Buckley, 1988). In Figure 4a, ci j is the TF-IDF of term ti in concept C j , while Figure 4b shows dik which is the TF-IDF of ti in Dk . (a) Term-concept matrix (b) Term-document matrix Figure 4: Term matrices for concepts and documents Next, documents D1 ... Dn are represented with respect to concepts by computing the cosine similarity of the term vectors for concepts and documents. The output is the concept-document matrix shown in Figure 5a, where y jk is the cosine similarity of the vertical shaded term vectors for C j and Dk from Figures 4a and 4b respectively. Finally, the document similarity is generated by computing the cosine similarity of concept-vectors for documents. Figure 5b shows zkm, which is the cosine similarity of the concept-vectors for Dk and Dm from Figure 5a. So, the CONCEPTBASED approach uses the document representation and similarity in Figure 5 to influence retrieval. We expect to retrieve documents that are similar based on the domain concepts that they contain. (a) Concept-document matrix representation (b) Document-document similarity Figure 5: Document representation and similarity using the CONCEPTBASED approach 4.2 The HYBRID Document Representation Approach The HYBRID approach exploits the relative distribution of the vocabulary in the concept and document spaces to augment the representation of learning resources with a bigger, but focused, vocabulary as shown in Figure 6. So the TF-IDF weight of a term changes depending on its relative frequency in both spaces. First, our 73 domain concepts, C1 ... Cm from section 3.4, and the documents we wish to represent, D1 ... Dn, are merged to form a corpus. Next, a term-document matrix with TF-IDF weighting is created using all the terms, t1 ... tT from the vocabulary of the merged corpus as shown in Figure 6a. Entry qik is the TF-IDF weight of term ti in Dk . If ti has a lower relative frequency in the 6 concept space compared to the document space, then the weight qik is boosted. So, distinctive terms from the concept space will get boosted. Although the overlap of terms from both spaces are useful for altering the term weights, it is valuable to keep all the terms from the document space because this gives us a richer vocabulary. The shaded term vectors for D1 ... Dn in Figure 6a form a term- document matrix for documents whose term weights have been influenced by the presence of terms from the concept vocabulary. (a) Hybrid term-document matrix representation (b) Hybrid document similarity Figure 6: Representation and similarity of documents using the HYBRID approach Finally, the document similarity in Figure 6b, is generated by computing the cosine similarity between the augmented term vectors for D1 ... Dn. Entry r jk is the cosine similarity of the term vectors for documents, D j and Dk from Figure 6a. The HYBRID method exploits the vocabulary in the concept and document spaces to influence the retrieval of documents. 5 Evaluating Learning Resource Representation Our methods are evaluated on a collection of topic-labelled learning resources by simulating an e- Learning recommendation task. We use a collection from Microsoft Academic Search (MAS)(Hands, 2012), in which the author-defined keywords associated with each paper identifies the topics they contain. The keywords represent what relevance would mean in an e-Learning domain and we exploit them for judging document relevance. The papers from MAS act as our e-Learning resources, and using a query-by-example scenario, we evaluate the relevance of a retrieved document by considering the overlap of keywords with the query. This evaluation approach allows us to measure the ability of the methods to identify relevant learning resources. We compare the performance of our CONCEPTBASED and HYBRID methods against that of Bag of Words (BOW). The BOW is a standard Information Retrieval method where documents are repre- sented using terms from the document space only with TF-IDF weighting. For each of the 3 methods, the documents are first pre-processed by removing English stopwords and applying Porter stemming. Then, after representation, a similarity-based retrieval is employed using cosine similarity. 5.1 Evaluation Method and Dataset Evaluations using human evaluators are expensive, so we take advantage of the author-defined key- words for judging the relevance of a document. The keywords are used to define an overlap metric. Given a query document Q with a set of keywords KQ, and a retrieved document R with its set of keywords KR, the relevance of R to Q is based on the overlap of KR with KQ. The overlap is computed as: Overlap(KQ,KR) = |KQ ∩KR| min ( |KQ|,|KR| ) (1) We decide if a retrieval is relevant by setting an overlap threshold, and if the overlap between KQ and KR meets the threshold, then KR is considered to be relevant. Figure 7 shows the number of keywords per document and the overlap of document pairs for the first dataset used. Our first dataset which we refer to as dataset 1 contains 217 Machine Learning and Data Mining papers. A distribution of the keywords per document is shown in Figure 7a, where the documents are sorted based on the number of keywords they contain. There are 903 unique keywords, and 1,497 keywords in total. A summary of the overlap scores for all document pairs is shown in Figure 7b. There are 23,436 entries for the 217 document pairs, and 20,251 are zero, meaning that there is no overlap in 86% of the data. So only 14% of the data have an overlap of keywords, indicating that the distribution of keyword overlap is skewed. There are 10% of document 7 pairs with overlap scores ≥ 0.14, and 5% are ≥ 0.25. For experiments with this dataset we use 0.14 and 0.25 as thresholds, thus avoiding extreme values that would allow either very many or few of the documents to be considered as relevant. (a) # of keywords per Document (b) Overlap of document pairs Figure 7: Number of keywords per document and overlap profile of document pairs in dataset 1 Our interest is in the topmost documents retrieved, because we want our top recommendations to be relevant. We use precision@n to determine the proportion of relevant documents retrieved: Precision@n = |retrievedDocuments∩relevantDocuments| n (2) where, n is the number of documents retrieved each time, retrievedDocuments is the set of documents retrieved, and relevantDocuments are those documents that are considered to be relevant i.e. have an overlap that is greater than the threshold. 5.2 Evaluation Results The methods are evaluated using a leave-one-out retrieval. In Figure 8, the number of recommenda- tions (n) is shown on the x-axis and the average precision@n is shown on the y-axis. RANDOM(N) has been included to give an idea of the relationship between the threshold and the precision values. RANDOM results are consistent with the relationship between the threshold and the proportion of data in Figure 7b. Overall, HYBRID(�) performs better than BOW(×) and CONCEPTBASED(•), showing that aug- menting the representation of documents with a bigger, but focused vocabulary, as done in HYBRID, is a better way of harnessing our background knowledge. BOW also performs well because the doc- ument vocabulary is large, but the vocabulary used in CONCEPTBASED may be too limited. The complexity of the representation method in HYBRID overcomes the limitation of CONCEPTBASED. All the graphs fall as the number of recommendations, n increases. This is expected because the ear- lier retrievals are more likely to be relevant. However, the overlap of HYBRID and BOW at higher values of n may be because the documents retrieved by both methods are drawn from the same neigh- bourhoods. (a) Results at Threshold of 0.14 (b) Results at Threshold of 0.25 Figure 8: Precision of the methods at overlap thresholds of 0.14 and 0.25 on dataset 1 The relative performance at a threshold of 0.25 in Figure 8b, is similar to the performance at 0.14. However, at this more challenging threshold, HYBRID and BOW do not perform well on the first retrieval. Generally, the results show that the HYBRID method is able to identify relevant learning 8 resources by highlighting the domain concepts they contain, and this is important in e-Learning. The graphs show that augmenting the representation of learning resources with our background knowledge is beneficial for e-Learning recommendation. 6 Refined Background Knowledge One issue with the previous concept generation method is that the concept vocabulary produced was limited. A suitable representation for e-Learning resources should have a good coverage of relevant domain topics. In this section, we discuss the steps taken to refine our method used for generating domain concepts in order to improve our background knowledge and increase the coverage of our concept vocabulary. 6.1 Enriched Domain Concepts In developing this method, we go through the phases described in sections 3.2 - 3.4. First, in addition to the TOC stopwords, the SMART stopwords (Salton, 1971) are also removed during pre-processing. This allows us to remove words that do not contribute to learning terms, and still retain a good set of words for generating our concepts. Second, words referring to the name of the domain used for demonstration such as: machine, learning, data, and mining are not removed during pre-processing, as we observed that removing these words before ngram generation prevents other relevant ngrams such as instance based learning or reinforcement learning, that contain any of these words from being identified. Third, we increase our ngram extraction to generate 1-5 grams from our TOC-phrases because, a distribution of the Wiki-phrases in Figure 3 showed that 99% of phrases are 1-5grams; this allows us to increase the number of concepts we can generate. We apply ngram extraction to the TOC-phrases to produce the following TOC-ngrams: 2467 Unigrams; 5387 Bigrams; 3625 Trigrams; 1668 Fourgrams; and 576 Fivegrams. The TOC-ngrams are verified as described in Section 3.3 using the Wiki-phrases to produce a set of potential concept labels containing 24 Unigrams; 96 Bigrams; 38 Trigrams; 6 Fourgrams; and no Fivegrams. A second verification step as described in Section 3.4 is applied to the potential concept labels. This entails using the verified ngrams to search Wikipedia pages in order to generate a domain concept. The search returns discovered text that forms a pseudo-document and a concept label. Overall, our refined method has 150 domain concepts that pass the second verification, each having a concept label and pseudo-document pair. The pseudo-document terms are pre-processed using standard techniques of English stopword removal and Porter Stemming. These terms now form the concept vocabulary of our refined background knowledge which we refer to as the CONCEPTBASED+ method. 6.2 Recommendation using the CONCEPTBASED+ approach The CONCEPTBASED+ method employs the richer concept vocabulary of our refined background knowledge for representing documents. We expect the representation created using the CONCEPT- BASED+ method to contain a better coverage of the learning domain because of the richer concepts it contains. Our aim is to address the issue of the limited concepts contained in the CONCEPTBASED method. For recommendation using CONCEPTBASED+, we use the same representation and docu- ment similarity as the CONCEPTBASED method illustrated in Figures 4 & 5, but with a richer concept vocabulary. So documents are represented with respect to concepts by computing the cosine similarity of term vectors for concepts and documents to produce a concept document matrix. Then, the simi- larity between documents can be generated by computing the similarity between respective concept vectors for documents. By using the CONCEPTBASED+ method for representation, we expect to retrieve documents that are similar based on the concepts they contain, and this is obtained from a document-document sim- ilarity matrix as shown in Figure 5b. A standard approach of representing documents would be to define the document similarity based on the term document matrix illustrated in Figure 4b, but this exploits the concept vocabulary only. In our approach, we put more emphasis on the domain concepts, so we use the concept document matrix illustrated in Figure 5a, to underpin the similarity between documents. The CONCEPTBASED+ method combines the focus with breadth of a richer set of domain concepts when representing documents. 6.3 Evaluating the Refined Representation This section investigates whether the domain concepts generated using a refined approach i.e. CON- CEPTBASED+ are better for representing documents than concepts generated with a standard method 9 i.e.CONCEPTBASED. The same evaluation method and dataset 1 presented in Section 5.1 is adopted here, and a leave-one-out retrieval is applied for evaluating the methods. In Figure 9, the number of recommendations is shown on the x-axis while the average precision@n is shown on the y-axis. An overlap threshold of 0.14 is used because there are 10% of document pairs in this dataset with overlap scores ≥ 0.14. The performance of CONCEPTBASED+(�) is shown by the darker line, and CONCEPTBASED(•) by the gray line. BOW(×) is included as the benchmark and RANDOM(N) gives an idea of the rela- tionship between the threshold used and the precision values. The graphs of all the methods fall as the number of recommendations, n increases. This is expected as earlier retrievals are more likely to be relevant. Overall, CONCEPTBASED+ outperforms CONCEPTBASED, BOW, and RANDOM, by producing better recommendations for all values of n. This performance shows the advantage of using the richer concept vocabulary for representing learning materials. The results confirm that CONCEPT- BASED+ contains concepts that have a better coverage of the learning domain than CONCEPTBASED which has a limited set of concepts. So we adopt CONCEPTBASED+ as a background knowledge representation for learning materials in this domain. Figure 9: Comparing CONCEPTBASED and CONCEPTBASED+ at a threshold of 0.14 on dataset 1 6.4 Evaluation Using a Larger Dataset We compare the performance of our HYBRID and CONCEPTBASED+ methods against that of the standard BOW approach on a larger dataset, in order to confirm our findings from the previous ex- periments. Figure 10 contains the number of keywords per document and the overlap of document pairs for the second dataset used. Our second dataset which we refer to as dataset 2 contains 1000 Machine Learning and Data Mining papers also from Microsoft Academic Research. Figure 10a contains a distribution of the keywords per document, where the documents are sorted based on the number of keywords they contain. There are 3063 unique keywords, and 4551 keywords in total. We take advantage of these author-defined keywords for judging relevance. A summary of the overlap profile of document pairs for dataset 2 is shown in Figure 10b. There are 499,500 entries for the 1000 document pairs, and 480,129 entries are zero, meaning that there is no overlap in 96% of the data. So only 4% of the data have an overlap of keywords, indicating that the distribution of keyword overlap is skewed. There are 3% of document pairs with overlap scores ≥ 0.2. The same evaluation method presented in 5.1 is used here. Then a leave-one-out retrieval method is applied, and precision@n as given in Equation 2 is used to determine the proportion of relevant documents retrieved. With dataset 2, we use a threshold of 0.2 thus preventing values that allow either too many or few documents to be considered as relevant. In Figure 11, the number of recommendations is shown on the x-axis and the average precision@n is on the y-axis. The average precision values are based on the overlap of keywords between document pairs and the threshold value chosen for the experiment. RANDOM(N) gives an idea of the relationship between the threshold and the precision values, and the results are consistent with the overlap profile in Figure 10b. 10 (a) # of keywords per document (b) Overlap of document pairs Figure 10: Number of keywords per document and overlap profile of document pairs in dataset 2 On this bigger dataset, CONCEPTBASED+(�) method outperforms HYBRID(�), BOW(×), and CONCEPTBASED(•), confirming that using a richer and focused vocabulary to represent documents is useful for e-Learning recommendation. The results also show HYBRID performing better than BOW, again confirming that augmenting the representation of learning resources with domain con- cepts is better than using the content only for e-Learning recommendation. Experiments were also run at thresholds of 0.25 and 0.33 and the relative performance at these thresholds is similar to the performance at 0.2, so the graphs are not shown. Our results show that we are able to leverage on the vocabulary from CONCEPTBASED+ which is not only a larger vocabulary, but one focused on domain concepts, thus allowing our method to influence the retrieval and recommendation of relevant learning resources. Figure 11: Precision of the methods at overlap threshold of 0.2 on dataset 2 7 Conclusions The growing availability of e-Learning materials on the Web provides opportunities for learners to easily access new and valuable information. However, finding good materials is difficult because retrieval has to overcome the challenge of ineffective queries often input by learners. e-Learning recommendation offers a possible solution to this difficulty. Though, recommendation in e-Learning environments is challenging because the learning materials are often unstructured text, and so are not properly indexed for retrieval. We address this challenge by creating a method that automatically acquires background knowledge in the form of a rich set of concepts related to the selected learning domain. In building our method, we take advantage of the knowledge of experts contained in the TOCs of e-Books to identify relevant domain topics. By using e-Books we benefit from the prove- nance associated with these teaching materials. The identified topics are enriched with discovered 11 text from Wikipedia, and this extends the coverage and richness of our representation. CONCEPTBASED method takes advantage of similar distributions of concept terms in the concept and document spaces to define a concept-term driven representation. Although the concept vocab- ulary in CONCEPTBASED is limited, HYBRID exploits the relative distribution of the vocabulary in the concept and document spaces to augment the representation of learning resources with a larger vocabulary influenced by domain concepts. CONCEPTBASED+ provides a richer concept vocabulary that allows concept-based distinctiveness to be helpful in the representation and retrieval of docu- ments. This refined method allows us to generate a richer and focused set of domain concepts, which provides a better coverage of the domain. The performance of CONCEPTBASED+ in our evaluation shows the advantage of using the richer concept vocabulary for representing learning materials. Our results confirm an improvement in e-Learning recommendation when a rich concept vocabulary is used for representing learning resources. References Agrawal, R., Chakraborty, S., Gollapudi, S., Kannan, A., and Kenthapadi, K. (2012). Quality of textbooks: An empirical study. In ACM Symposium on Computing for Development, pages 16:1–16:1. doi: 10.1145/2160601.2160623. Beliga, S., Meštrović, A., and Martinčić-Ipšić, S. (2015). An overview of graph-based keyword ex- traction methods and approaches. Journal of Information and Organizational Sciences, 39(1):1– 20. Blei, D. M. and McAuliffe, J. D. (2007). Supervised topic models. In Neural Information Processing Systems, pages 121–128. Bousbahi, F. and Chorfi, H. (2015). MOOC-Rec: A case based recommender system for MOOCs. Procedia - Social and Behavioral Sciences, 195:1813 – 1822. doi: 10.1016/j.sbspro.2015.06.395. Boyce, S. and Pahl, C. (2007). Developing domain ontologies for course content. Journal of Educa- tional Technology & Society, 10(3):275–288. Chen, W., Niu, Z., Zhao, X., and Li, Y. (2014). A hybrid recommendation algorithm adapted in e-learning environments. World Wide Web, 17(2):271–284. doi: 10.1007/s11280-012-0187-z. Chen, Z. and Liu, B. (2014). Topic modeling using topics from many domains, lifelong learning and big data. In 31st International Conference on Machine Learning, pages 703–711. Clarà, M. and Barberà, E. (2013). Learning online: Massive Open Online Courses (MOOCs), connectivism, and cultural psychology. Distance Education, 34(1):129–136. doi: 10.1080/01587919.2013.770428. Coenen, F., Leng, P., Sanderson, R., and Wang, Y. J. (2007). Statistical identification of key phrases for text classification. In Machine Learning and Data Mining in Pattern Recognition, pages 838–853. Springer. doi: 10.1007/978-3-540-73499-4_63. Dietze, S., Yu, H. Q., Giordano, D., Kaldoudi, E., Dovrolis, N., and Taibi, D. (2012). Linked educa- tion: Interlinking educational resources and the Web of data. In 27th Annual ACM Symposium on Applied Computing, pages 366–371. 10.1145/2245276.2245347. Ferrucci, D., Levas, A., Bagchi, S., Gondek, D., and Mueller, E. T. (2013). Watson: Beyond Jeopardy! Artificial Intelligence, 199:93–105. doi: https://doi.org/10.1016/j.artint.2012.06.009. Gherasim, T., Harzallah, M., Berio, G., and Kuntz, P. (2013). Methods and tools for automatic construction of ontologies from textual resources: A framework for comparison and its appli- cation. In Advances in Knowledge Discovery and Management, pages 177–201. Springer. doi: 10.1007/978-3-642-35855-5_9. Hands, A. (2012). Microsoft academic search. Technical Services Quarterly, 29(3):251–252. doi: 10.1080/07317131.2012.682026. Lehmann, J., Isele, R., Jakob, M., Jentzsch, A., Kontokostas, D., Mendes, P. N., Hellmann, S., Morsey, M., Van Kleef, P., Auer, S., et al. (2015). DBpedia–a large-scale, multilingual knowledge base extracted from Wikipedia. Semantic Web, 6(2):167–195. doi: 10.3233/SW-140134. Matsuo, Y. and Ishizuka, M. (2004). Keyword extraction from a single document using word co-occurrence statistical information. International Journal on Artificial Intelligence Tools, 13(01):157–169. doi: 10.1142/S0218213004001466. 12 Mbipom, B., Craw, S., and Massie, S. (2016). Harnessing background knowledge for e-learning rec- ommendation. In Research and Development in Intelligent Systems XXXIII: Incorporating Ap- plications and Innovations in Intelligent Systems XXIV, pages 3–17. Springer. doi: 10.1007/978- 3-319-47175-4_1. Milne, D. and Witten, I. H. (2008). Learning to link with Wikipedia. In 17th ACM Conference on Information and Knowledge Management, pages 509–518. doi: 10.1145/1458082.1458150. Nasraoui, O. and Zhuhadar, L. (2010). Improving recall and precision of a personalized semantic search engine for e-learning. In 4th IEEE International Conference on Digital Society, pages 216–221. doi: 10.1109/ICDS.2010.63. Panagiotis, S., Ioannis, P., Christos, G., and Achilles, K. (2016). APLe: Agents for personalized learn- ing in distance learning. In 7th International Conference on Computer Supported Education, pages 37–56. Springer. doi: 10.1007/978-3-319-29585-5_3. Porter, M. F. (1980). An algorithm for suffix stripping. Program, 14(3):130–137. Qureshi, M. A., O’Riordan, C., and Pasi, G. (2014). Exploiting Wikipedia to identify domain-specific key terms/phrases from a short-text collection. In Proceedings of the 5th Italian Information Retrieval Workshop, pages 63–74. Rodrigues, L., Antunes, B., Gomes, P., Santos, A., Barbeira, J., and Carvalho, R. (2007). Using textual CBR for e-learning content categorization and retrieval. In 4th International Conference on Case-Based Reasoning workshop on textual case-based reasoning. Ruiz-Iniesta, A., Jimenez-Diaz, G., and Gomez-Albarran, M. (2014). A semantically enriched context-aware OER recommendation strategy and its application to a computer science OER repository. IEEE Transactions on Education, 57(4):255–260. doi: 10.1109/TE.2014.2309554. Salton, G. (1971). The SMART Retrieval System-Experiments in Automatic Document Processing. Prentice-Hall, Upper Saddle River, NJ, USA. doi: 10.1109/TPC.1972.6591971. Salton, G. and Buckley, C. (1988). Term-weighting approaches in automatic text retrieval. Informa- tion Processing & Management, 24(5):513–523. doi: 10.1016/0306-4573(88)90021-0. Völkel, M., Krötzsch, M., Vrandecic, D., Haller, H., and Studer, R. (2006). Semantic Wikipedia. In Proceedings of the 15th International Conference on World Wide Web, pages 585–594. ACM. doi: 10.1145/1135777.1135863. Witten, I. H., Paynter, G. W., Frank, E., Gutwin, C., and Nevill-Manning, C. G. (1999). KEA: Practical automatic keyphrase extraction. In 4th ACM Conference on Digital libraries, pages 254–255. doi: 10.1145/313238.313437. Yang, H.-L. and Lai, C.-Y. (2010). Motivations of Wikipedia content contributors. Computers in Human Behavior, 26(6):1377 – 1383. doi: 10.1016/j.chb.2010.04.011. Yang, K., Chen, Z., Cai, Y., Huang, D., and Leung, H. F. (2016). Improved automatic keyword extraction given more semantic knowledge. In International Conference on Database Systems for Advanced Applications, pages 112–125. Springer. doi: 10.1007/978-3-319-32055-7_10. Yarandi, M., Tawil, A.-R., and Jahankhani, H. (2011). Adaptive e-learning system using ontology. In In Proceedings of the 22nd International Workshop on Database and Expert Systems Applica- tions, pages 511–516. doi: 10.1109/DEXA.2011.9. Zhang, X., Liu, J., and Cole, M. (2013). Task topic knowledge vs. background domain knowledge: Impact of two types of knowledge on user search performance. In Advances in Information Systems and Technologies, pages 179–191. Springer. doi: 10.1007/978-3-642-36981-0_17. Zheng, Z., Li, F., Huang, M., and Zhu, X. (2010). Learning to link entities with knowledge base. In Human Language Technologies: The Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 483–491. 13 MBIPOM 2018 Improving e-learning recommendation coversheetJournalArticles MBIPOM 2017 Improving e-learning recommendation MBIPOM 2018 Improving e-learning (FINAL) OA: GREEN OA Logo: AUTHORS: MBIPOM, B., CRAW, S. and MASSIE, S. TITLE: Improving e-learning recommendation by using background knowledge. YEAR: 2018 Publisher citation: MBIPOM, B., CRAW, S. and MASSIE, S. 2018. Improving e-learning recommendation by using background knowledge. Expert systems [online], Early View. Available from: https://doi.org/10.1111/exsy.12265 OpenAIR citation: MBIPOM, B., CRAW, S. and MASSIE, S. 2018. Improving e-learning recommendation by using background knowledge. Expert systems, Early View. Held on OpenAIR [online]. Available from: https://openair.rgu.ac.uk Version: AUTHOR ACCEPTED Publisher: WILEY Series: Expert systems ISSN: 0266-4720 eISSN: 1468-0394 Set statement: This is the accepted version of the following article MBIPOM, B., CRAW, S. and MASSIE, S. 2018. Improving e-learning recommendation by using background knowledge. Expert systems, Early View, which has been published in final form at https://doi.org/10.1111/exsy.12265. This article may be used for non-commercial purposes in accordance with the Wiley Self-Archiving Policy. License: BY-NC 4.0 License URL: https://creativecommons.org/licenses/by-nc/4.0 CC Logo: 2018-01-26T16:06:55+0000 OpenAIR at RGU 
work_5xpfvjtqk5bbfefatdohgjss3a	----	Real-time credit card fraud detection using computational intelligence Real-time credit card fraud detection using computational intelligence Jon T.S. Quah *, M. Sriganesh School of Electrical and Electronic Engineering, Nanyang Technological University, Nanyang Avenue, Singapore 639798, Republic of Singapore Abstract Online banking and e-commerce have been experiencing rapid growth over the past few years and show tremendous promise of growth even in the future. This has made it easier for fraudsters to indulge in new and abstruse ways of committing credit card fraud over the Internet. This paper focuses on real-time fraud detection and presents a new and innovative approach in understanding spending patterns to decipher potential fraud cases. It makes use of self-organization map to decipher, filter and analyze customer behavior for detection of fraud. � 2007 Elsevier Ltd. All rights reserved. Keywords: Self-organizing map; Unsupervised learning; Risk scoring; Transaction 1. Introduction The fast and wide reach of the Internet has made it one of the major selling channels for the retail sector. In the last few years, there has been a rapid increase in the number of card issuers, card users and online merchants, giving very little time for technology to catch-up and prevent online fraud completely. Statistics shows that on-line banking has been the fastest growing Internet activity with nearly 44% of the population in the US actively participating in it (Fighting Fraud on the Internet, 1999). As overall e-com- merce volumes continued to grow over the past few years, the figure of losses to Internet merchants was projected to be between $5 and $15 billion in the year 2005. Recent sta- tistics by Garner group place online fraud rate between 0.8% and 0.9%, with auction fraud accounting to nearly half of the total incidents of fraud on the Internet (Online Banking, 2005). Considering the current trends of e-com- merce volumes, the projected loss is $8.2 billion in the year 2006, with $3.0 billion in the US alone (Statistics for Gen- eral & Online Fraud, 2007). 1.1. How and where does fraud begin In order to understand the severity of credit card fraud, let us briefly look into the mechanisms adopted by fraud- sters to commit fraud. Credit card fraud involves illegal use of card or card information without the knowledge of the owner and hence is an act of criminal deception. Fraudsters usually get hold of card information in a variety of ways: Intercepting of mails containing newly issued cards, copying and replicating of card information through skimmers or gathering sensitive information through phish- ing (cloned websites) or from unethical employees of credit card companies. Phishing involves acquiring of sensitive information like card numbers and passwords by masquerading as a trust- worthy person or business in an electronic communication such as e-mail (Schneck, 2007). Fraudsters may also resort to generation of credit card numbers using BIN (Bank Iden- tification Numbers) of banks. A recent scheme of Triangula- tion takes fraud fighters many days to realize and investigate (Bhatla, Prabhu, & Dua, 2003). In this method, the fraudster operates through an authentic-looking website, where he advertises and sells goods at highly discounted prices. The unaware buyer submits his card information and buys goods. The fraudster then places an order with a genuine merchant using the stolen card information. He then uses 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.08.093 * Corresponding author. Tel.: +65 67905871; fax: +65 62701556. E-mail address: itsquah@ntu.edu.sg (J.T.S. Quah). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 35 (2008) 1721–1732 Expert Systems with Applications mailto:itsquah@ntu.edu.sg the stolen card to purchase other goods or route funds into intractable accounts. Its only after several days that the mer- chant and card owners realize about the fraud. This type of fraud causes initial confusion that provides camouflage for the fraudster to carry out their operations. 1.2. Impact of fraud It is interesting to note that credit card fraud affects card owners the least because their liability is limited to the trans- actions made. The existing legislations and cardholder pro- tection policies as well as insurance schemes in most countries protect the interests of the cardholders. However, the most affected are the merchants, who, in most situations, do not have any evidence (eg. Digital signature) to dispute the cardholders’ claim of misused card information. Mer- chants end up bearing all the loses due to chargeback, ship- ping cost of goods, card issuer fees and charges as well as their own administrative costs. Excessive fraudulent cases involving the same merchant can drive away customers, cause card issuer banks to withdraw service and also result in loss of reputation and goodwill (Yu & Singh, 2002). Card issuer banks have to bear the administrative cost of investi- gations into fraud cases as well as infrastructure costs of set- ting up the required software and hardware facilities to combat fraud. They also incur indirect costs through trans- action delays. Studies show that the average time lag between the fraudulent transaction date and chargeback notification can be as high as 72 days, thereby giving fraudsters sufficient time to cause severe damage (Bhatla et al., 2003). 1.3. Fraud detection and prevention The negative impacts of fraud make it very clear and necessary to put in place an effective and economical fraud detection system. Recent technological advancements to combat fraud have contributed number of solutions in this area (Bhatla et al., 2003; All points protection, 2007). Fraud detection techniques involving sophisticated screen- ing of transactions to tracking customer behaviour and spending patterns are now being developed and employed by both merchants as well as card issuer banks (Tan & Thoen, 2000). Some of the recently employed techniques include transaction screening through Address Verification Systems (AVS), Card Verification Method (CVM), Per- sonal Identification Number (PIN) and Biometrics. AVS involves verification of address with zip code of the cus- tomer while CVM and PIN involve checking of numeric code that is keyed in by the customer. Biometrics might involve signature or fingerprint verification. Rule-based methods and maintaining of positive and negative lists of customers and geographical regions are also used in prac- tice. Data mining and credit scoring methods focus on sta- tistical analyses and deciphering of customer behaviour and spending patterns to detect frauds (Huang, Chen, & Wang, 2007). Neural networks are capable of deriving pat- terns out of databases containing historical transactions of customers. These neural networks can be ‘trained’ and are ‘adaptive’ to the emerging new forms of frauds. Deployment of sophisticated techniques and screening of every transaction alone will not reduce losses. It is necessary to employ an effective and economical solution to combat fraud (Turney, 1995). Such a solution should not only detect fraud cases efficiently but also turn out to be cost-effective. The idea is to strike a balance between the cost involved in transaction screening and review and the losses due to fraud- ulent cases. Analyses show that review of only 2.0% of trans- actions can result in reducing fraud losses accounting to 1.0% of total value of transactions. While a review of as high as 30% of transactions can reduce the fraud loses drastically to 0.06%, but that increases review costs exorbitantly (Bhatla et al., 2003). The estimated cost of not using anti-fraud soft- ware was about $60 billion in 2005 (Statistics for General & Online Fraud, 2007). The key to minimize total costs is to categorize transac- tions and review only the potentially fraudulent cases. This should involve deployment of a step-by-step screening, filter- ing and review mechanism. A typical deployment can involve initial authentication of transactions through PIN, expiry date on card, AVS and CVM. A second level of screening can involve comparing with positive and negative lists as well as rules based on customers, geographical regions, IP addresses and policies. Risk and credit scoring with pattern and behaviour analyses can come next, followed by manual review. This classifies and filters out transactions as genuine or fraudulent in every step and as a result only a few transac- tions would require further manual review. Such a solution reduces the overall processing delay as well as total costs involved in manpower and administration. The focus of this paper will now shift to risk scoring and behavioral pattern detection using neural networks. 2. Neural networks in fraud detection – literature review Neural Networks have been extensively put to use in the areas of banking, finance and insurance. They have been successfully applied into credit scoring of customers, bank- ruptcy or business failure prediction, stock price forecast- ing, bond rating, currency prediction and many more areas (Stock Analysis, 2007; Quah & Srinivasan, 1999). In the area of fraud detection and prevention, neural net- works like feed-forward networks with back-propagation have found immense applications (Fraud Brief – AVS & CVM, ClearCommerce Corporation, 2003; The Evolving Threat of Card Skimming, FairIsaac, 2007; ClearCom- merce Fraud Prevention Guide, 2002). Usually such appli- cations of neural networks systems involve knowing about the previous cases of fraud, to make systems learn the var- ious trends. Fraud cases are statistically analyzed to derive out relationships among input data and values for certain key parameters in order to understand the various patterns of fraud. This knowledge of fraud trends is then iteratively taught to feed-forward neural networks, which can success- fully identify similar fraud cases occurring in the future. 1722 J.T.S. Quah, M. Sriganesh / Expert Systems with Applications 35 (2008) 1721–1732 https://isiarticles.com/article/17706 
work_5xv3i4veezb63kw7jpg2ahioku	----	Dominancebased rough set theory over intervalvalued information systems Article DOI: 10.1111/exsy.12022 Dominance-based rough set theory over interval-valued information systems Bingzhen Sun,1,2 Weimin Ma1* and Zengtai Gong3 (1) School of Economics and Management, Tongji University, Shanghai 200092, China E-mail: mawm@tongji.edu.cn (2) School of Traffic and Transportation, Lanzhou Jiaotong University, Lanzhou, Gansu 730070, China (3) College of Mathematics and Information Science, Northwest Normal University, Lanzhou, Gansu 730070, China E-mail: bzh--sun@163.com Abstract: This paper proposes a new generalization of classical real-valued information systems, that is, interval-valued information systems. By defining an interval-valued dominance relation on a condition attribute, a rough set model and attribute reduction are established over interval-valued information systems. Moreover, several interesting properties are investigated by constructive approach. Furthermore, knowledge reductions of consistent and inconsistent interval-valued dominance decision information systems are studied, respectively. Subsequently, some descriptive theorems of knowledge reduction are presented for interval-valued dominance decision information systems. Finally, the validity of the model and conclusions is verified by numerical example. Keywords: rough set, interval-valued dominance relation, knowledge reduction 1. Introduction Rough set theory, introduced by Pawlak (1982), has become well established as an approach to studying information systems characterized by insufficient and incomplete information in a wide variety of applications. Various generalizations of rough approximations and applications have been made over the past years. Recently, rough set theory has become a popular mathematical tool in areas such as pattern recognition, image processing, feature selection (Pawlak & Skowron, 2007), conflict analysis (Pawlak, 1998), decision support (Tsumoto, 1998; Kim & Han, 2001;Tsumoto, 2004), data mining and knowledge discovery processes from large data sets (Lingras & Yao, 1998). The concept of rough set theory is founded on the assumption that every object of the universe of discourse is associated with some information. Objects characterized by the same information are indiscernible in view of their available information. This kind of generalization for indiscernibility relation forms the mathematical basis of rough set theory (Yao, 2003; Beynon et al., 2000; Skowron & Stepaniuk, 1996; Huynh & Nakamori, 2005). Classical rough set adopts two definable sets, namely the lower and upper approximations, to deal with the approximation computation for any subset of universe. Then, potential knowledge will be revealed from information systems by lower and upper approximations, and a decision rule can be induced. One important usage of rough set theory is knowledge reduction for information systems. Knowledge reduction deletes unnecessary attributes from the original attribute set, while the same ability to classify remains (Kryszkiewicz, 1998; Ziarko, 2003; Xu et al., 2010). Many valuable conclusions have been drawn through the method of knowledge reduction in consistent information systems (Zhang et al., 2003a, 2003b). Meanwhile, more attention has been paid to knowledge reduction in inconsistent systems in research on rough set theory. Many types of knowledge reduction also have been proposed in this field (Zhang et al., 2006; Wang & Lin, 2006; Xu & Zhang, 2008; Xu et al., 2010). Information systems or decision information systems transacting with Pawlak rough set theory are self-contained, and the value of the attributes is discrete or certain. At the same time, decision rule and knowledge discovery of information systems are acquired by attribute reduction based on equivalence relation (satisfying reflexive, symmetric and transitive properties). Recently, many authors have generalized equivalence relation to various forms such as similarity relation, dominance relation and covering relation in terms of Pawlak rough set theory. More detailed studies on generalized information systems follow from the perspective of knowledge reduction (Greco et al., 1999; Greco et al., 2001; Greco et al., 2002; Kazimierz, 2004; Zhu, 2007; Leung et al., 2008; Wu et al., 2005; Leung et al., 2006; Huang et al., 2012a, 2012b). Moreover, a large number of concepts of knowledge reduction such as generalized decision reduct, b � reduct (Kryszkiewicz, 2001; Zhang et al., 2006), probability reduct, dynamic reduct (Grzymala, 1997), entropy reduct and approximation entropy reduct (Slezak, 2002), distribution reduct and allocation reduct (Zhang et al., 2003a, 2003b, Zhang et al., 2006; Xu & Zhang, 2008; Xu et al., 2010) also have been developed for decision information systems. However, the value of attribute may be symbolic-valued, set-valued, fuzzy-valued or fuzzy number-valued and continuous-valued for the existing database in practice. Therefore, Pawlak rough set has some limitations in handling these values of attributes in information systems. Moreover, apart from different definitions of the knowledge reduction established for classical real-valued information © 2013 Wiley Publishing Ltd Expert Systems, xxxx 2013, Vol. 00, No. 00 systems, various generalized information systems are proposed, such as set-valued information systems (Guan & Wang, 2006; Zhang et al., 2006; Qian et al., 2008, 2009; Wang & Zhang, 2006), ordered information systems (Xu & Zhang, 2008; Xu et al., 2010), fuzzy information systems (Zhang et al., 2003a, 2003b; Yang et al., 2009), continuous- valued information systems (Ziarko, 2003; Zhang et al., 2003a, 2003b), interval-valued information systems and interval-valued fuzzy information systems (Sun et al., 2008; Gong et al., 2008; Zhang et al., 2009; Yang & Sheng, 2011; Huang, 2012; Dai et al., 2012; Chen & Miao, 2011), interval-valued intuitionistic fuzzy information systems (Zhang, 2010; Huang et al., 2012a, 2012b) and vague objective information systems (Feng et al., 2010). Though there is different background in theory and application for these generalized information systems or decision information systems, these methods go along with classical real-valued information systems. Moreover, the concepts of knowledge reduction for classical real-valued information systems have been successfully generalized in these information systems. The terms with the reduction of aforementioned various generalized information systems also use the same concepts, such as the classical real-valued information systems (Zhang et al., 2003a, 2003b; Zhang et al., 2006). So far, most research has been conducted on interval- valued information systems. Actually, the interval value of an attribute is a closed interval of the universe of real number for existing interval-valued information systems (Huang et al., 2006; Leung et al., 2008; Qian et al., 2008; Yang et al., 2009). There are many useful applications based on these interval-valued information systems, such as pattern recognition and automated knowledge acquisition. However, it would be of great value to understand what restricts the value of an attribute on the closed interval of [0,1] where the uncertainty decision or comprehension evaluation is concerned. Therefore, research on the information systems with the value of an attribute in closed interval [0,1] is a necessity. In this paper, we tentatively discuss the rough set model and knowledge reduction in interval-valued information systems. This paper studies interval-valued information systems by defining an interval-valued dominance relation over attribute set, that is, we establish a rough set model on interval-valued information systems. Like the Pawlak rough set, several properties of this model are also studied. Meanwhile, we also study interval-valued decision infor- mation systems by defining an interval-valued dominance relation over condition attribute and an equivalence relation over decision attribute, respectively. The attribute reduction of interval-valued information systems is investigated in a manner similar to classical real-valued information systems (Zhang et al., 2003a, 2003b; Mi et al., 2004; Wu et al., 2005; Zhang et al., 2006; Wu, 2008; Xu & Zhang, 2008; Xu et al., 2010). Furthermore, knowledge reduction of both consistent and inconsistent interval-valued decision information systems is discussed. In fact, all conclusions presented in this paper are generalization of the corresponding conclusions of classical real-valued decision information systems (Zhang et al., 2003a, 2003b; Mi et al., 2004; Wu et al., 2005; Zhang et al., 2006; Xu & Zhang, 2008; Xu et al., 2010). However, there also exist some differences with the corre- sponding conclusions of classical real-valued information systems because the interval values in closed interval [0,1] have some specialized properties. Generally speaking, all similar research focuses on the definition of a binary relation on condition and decision attributes. The existing research related to our work focuses on the following aspects. One is to define two dominance relations on both condition and decision attributes, as in ordered information systems (Xu & Zhang, 2008; Xu et al., 2010). Another is to define two equivalence relations over both condition and decision attributes, as in real-valued information systems (Zhang et al., 2003a, 2003b; Zhang et al., 2006; Mi et al., 2004). Some others define a binary relation or partial relation on condition and decision attributes, as in set-valued or language-valued information systems (Guan & Wang, 2006; Wang & Zhang, 2006; Wang & Lin, 2006; Qian et al., 2008; Qian et al., 2009). Therefore, some conclusions reached in the existing related work could not hold in the interval-valued (decision) information systems studied in this paper. The rest of this paper is organized as follows. Section 2 briefly introduces some important notions of the classical Pawlak rough set theory and the concept of interval-valued systems. Section 3 establishes the rough set model over interval-valued information systems. Furthermore, attribute reduction is studied in detail for interval-valued information systems. In Section 4, knowledge reduction for both consistent and inconsistent interval-valued dominance decision information systems is discussed, respectively. A numerical example is employed to test our conclusions. We conclude our research and set further future research directions in Section 5 2. Preliminary In this section, we present some concepts and conclusions that will be used in this paper (Pawlak & Skowron, 2007; Gong et al., 2008; Sun et al., 2008; Zhang, 2010). Let U be a non-empty finite set, which is called universe. Let R be an equivalence relation on U. That is, R is a binary relation and satisfies reflexive, symmetric and transitive properties. Then, we call (U,R) an approximation space. Moreover, binary relation R divides the universe U into piecewise disjoint classes such that any two different elements x and y belong to the same class if and only if (x,y) 2 R. Furthermore, U/R = {X1,X2, . . .,Xm} denotes all equiv- alence classes in universe U induced by equivalence relation R. If both of any two different elements x and y of U belong to Xi 2 U/R, then x and y are called indiscernible. The equivalence class of R and ∅ is called the primary set in approximation space (U,R). We describe the definition of information systems (Huynh & Nakamori, 2005; Gong et al., 2008; Sun et al., 2008) as follows. In general, a triple (U,A,{Va}a 2 A) is called an information system, where U is the set of object, A is the set of attribute and Va is the domain of attribute a, understood as a mapping a : U ! Va. It can be easily seen that each subset of attribute set A induces an equivalence relation: ind Að Þ ¼ x; yð Þ 2 U � U a xð Þ ¼ a yð Þ; a 2 Aj g:f © 2013 Wiley Publishing LtdExpert Systems, xxxx 2013, Vol. 00, No. 00 Then, ind(A) is called an indiscernibility relation and denotes ind(A) = ∩ a 2 Aind(a) (where ind(a) means ind({a})). If (x,y) 2 ind(A), we then say that the objects x and y are indiscernible with respect to A. In other words, we cannot distinguish x from y in terms of the attributes in A and vice versa. For any X ⊆ U, we define the two subsets of approximation space (U,R) as follows: R � X ¼ x 2 U x½ �R⊆ X �� �; R�X ¼ x 2 U x½ �R∩ X 6¼ ∅�� ��� We call R � X and R � X the lower and upper approximations of X with respect to (U,R). If R � X ¼ R�X, then X is called the definable set. Otherwise, X is called the rough set. The lower approximation R � X is the union of all elementary sets that are the subset of X, and the upper approximation R � X is the union of all elementary sets that have a non-empty intersection with X. The lower (upper) approximation R � X R � Xð Þ is interpreted as the collection of those elements of U that definitely (possibly) belong to X.(Huynh & Nakamori, 2005; Pawlak & Skowron, 2007; Gong et al., 2008; Sun et al., 2008). In what follows, we give the definition of interval value in [0,1] and its operators. We assume throughout that I is a unit closed interval, that is, I = [0,1]. Let [I] = {[a,b]|a ≤ b, a, b 2 I}. For any a 2 I, define �a ¼ a; a½ �, then �a 2 I½ �: Definition 2.1. Gorzafczary, 1988; Turksen, 1986; Gong et al., 2008; Sun et al., 2008; Zhang, 2010 For any ai 2 I, i 2 J, we define ∨i2Jai ¼ sup aiji 2 Jf g; ∧i2Jai ¼ inf aiji 2 Jf g; ∨i2J ai; bi½ � ¼ ∨i2Jai; ∨i2Jbi½ �; ∧i2J ai; bi½ � ¼ ∧i2Jai; ∧i2Jbi½ � where ∨ means maximum and ∧ means minimum. For any interval-valued [ai, bi] 2 [I], i = 1, 2, we define the following operations. [a1, b1] ¼ [a2, b2] if and only if a1 = a2, b1 = b2, [a1, b1] ≤ [a2, b2] if and only if a1 ≤ a2, b1 ≤ b2, [a1, b1] < [a2, b2] if and only if [a1,b1] ≤ [a2, b2] but [a1, b1] 6¼ [a2, b2]. Remark 2.1. It can be easily known that the following relation does not hold for any [ai, bi] 2 [I], i = 1, 2. a1; b1½ �≰ a2; b2½ � , a1; b1½ �> a2; b2½ �; a1; b1½ �≮ a2; b2½ � , a1; b1½ � ≥ a2; b2½ �: In general, the aforementioned two relations hold when [ai, bi] 2 [I], i = 1, 2 is real-valued. That is, ai = bi for any ai, bi 2 I, i = 1, 2. 3. Dominance-based rough set model over interval-valued information systems In this section, we discuss the basic rough set theory by defining an interval-valued dominance relation for interval- valued information systems. 3.1. Rough set model on interval-valued information systems The interval-valued information system was first defined by Sun and Gong (Sun et al., 2008; Gong et al., 2008). Here, we present the definition of interval-valued information systems as follows. Let U = {x1,x2, . . .,xk}, A= {a1,a2, . . .,am}. F = {fi|i ≤ m} is a family of mapping set between U and A, where fi is an interval- valued mapping from U to A. That is, fi : U ! [I](i ≤ m), and [I] is the domain of attribute ai. Then, we call triple (U,A,F) as an interval-valued information system(Gong et al., 2008; Sun et al., 2008). In general, every attribute set has determined an indiscernibility relation or equivalence relation in real- valued information systems. So, every attribute set defines an equivalence relation over interval-valued information systems as well. However, the equivalence relation could not be satisfied in practice, such as the multi-object comprehensive evaluation decision-making, the supplier selection decision-making with uncertainty and group decision-making in emergency events based on interval- valued information systems. So, there should be given a more suitable binary relation for interval-valued information systems. Therefore, we define an interval- valued dominance relation for interval-valued information systems in this section. It is also a generalization of the dominance relation in real-valued information systems (Zhang et al., 2003a, 2003b; Zhang et al., 2006; Xu & Zhang, 2008; Xu et al., 2010). Definition 3.1. Let (U,A,F) be interval-valued information systems. For any B ⊆ A, we define R≤B ¼ xi; xj � � fk xið Þ ≤ fk xj � � ; 8ak 2 B; xi; xj 2 U �� �� Then, R≤B is called an interval-valued dominance relation over interval-valued information systems. Furthermore, we call triple (U,A,F) as interval-valued information systems based on interval-valued dominance relation. For the sake of clarity, denote U; R≤A; F � � ¼ U; A; Fð Þ, and we call U; R≤A; F � � as interval-valued dominance information systems. Remark 3.1. If the value of attribute a 2 A is real-valued, U; R≤A; F � � will be real-valued information systems. Moreover, the interval- valued dominance relation will be real-valued dominance relation (Zhang et al., 2003a, 2003b; Zhang et al., 2006; Xu & Zhang, 2008; Xu et al., 2010). Denote x½ �≤B ¼ y 2 U fk xð Þ ≤ fk yð Þ; ak 2 Bj g:f It is easy to see that the interval-valued dominance relation satisfies the following properties and that all of © 2013 Wiley Publishing Ltd Expert Systems, xxxx 2013, Vol. 00, No. 00 them are similar to real-valued information systems, ordered information systems, set-valued information systems and other information systems with dominance relation (Zhang et al., 2003a, 2003b; Wang & Zhang, 2006; Xu & Zhang, 2008; Xu et al., 2010). Proposition 3.1. Let R≤A be an interval-valued dominance relation of U; R≤A; F � � . Then, the following conclusions hold. 1. R≤A is reflexive and transitive but not symmetric. Therefore, it is not an equivalence relation. 2. If B1 ⊆ B2 ⊆ A, then R≤A⊆R≤B2⊆R ≤ B1 : 3. If B1 ⊆ B2 ⊆ A, then x½ �≤A⊆ x½ �≤B2⊆ x½ � ≤ B1 : 4. If y 2 x½ �≤A, then y½ �≤A⊆ x½ �≤A and x½ �≤A ¼ ∪ y½ �≤A y 2 x½ �≤A �� �:� 5. y½ �≤A ¼ x½ �≤A if and only if f(x,a) = f(y,a), a 2 A. 6. C≤ ¼ x½ �≤A x 2 Uj g � is a covering of universe U. In what follows, we give the rough set model for interval- valued information systems. Let U; R≤A; F � � be interval-valued dominance information systems. For any X⊆ U, we define the lower and upper approximations of X as follows: R � ≤ A Xð Þ ¼ x 2 U x½ �≤A⊆X ��� o; R�≤A Xð Þ ¼ x 2 U x½ �≤A∩X 6¼ ∅�� �� n If R � ≤ A Xð Þ ¼ R �≤ A Xð Þ, then X is called a definable set with interval-valued dominance relation R≤A . Otherwise, X is called a rough set in U; R≤A; F � � : Like the rough set model over the existing generalized information systems (Zhang et al., 2003a, 2003b; Wang & Zhang, 2006; Xu & Zhang, 2008; Xu et al., 2010), the following properties are clear. Theorem 3.1. Let U; R≤A; F � � be an interval-valued dominance information system. Then, the lower and upper approximations satisfy the following properties: 1. R � ≤ A Uð Þ ¼ U; R �≤ A ∅ð Þ ¼ ∅; 2. R � ≤ A Xð Þ ¼ ≃R �≤ A ≃Xð Þ; R �≤ A Xð Þ ¼ ≃R� ≤ A ≃Xð Þ; 3. R � ≤ AðX∩YÞ¼R � ≤ AðXÞ∩R� ≤ AðYÞ; R � ≤ AðX∪YÞ¼R �≤ AðXÞ∪R �≤ AðYÞ; 4. R � ≤ AðX∪YÞ⊇R � ≤ A Xð Þ∪R� ≤ A Yð Þ; R �≤ AðX∩YÞ⊆R �≤ A Xð Þ∩R �≤ A Yð Þ; 5. R � ≤ A Xð Þ⊆X⊆R �≤ A Xð Þ; 6. R � ≤ A Xð Þ⊆R� ≤ A R� ≤ A Xð Þ � � ; R �≤ A R �≤ A Xð Þ � � ⊆R�≤A Xð Þ; where ≃ X stands for the complement of X. Proof It can be easily proven by the definitions earlier. Example 3.1. Table 1 gives an example of interval-valued information systems. It is easy to obtain the following dominance classes by Definition 3.1: x1½ �≤A ¼ x1; x2; x5f g x2½ �≤A ¼ x2; x5f g x3½ �≤A ¼ x2; x3; x4; x5f g x4½ �≤A ¼ x4f g x5½ �≤A ¼ x5f g Setting X = {x2, x3, x5}. Then, come the following relations: R � ≤ A Xð Þ ¼ x2; x5f g; R �≤ A Xð Þ ¼ x1; x2; x3; x5f g Clearly, R � ≤ A Xð Þ⊆X⊆R �≤ A Xð Þ: Remark 3.1. By Example 3.1, the following relations could not hold. 1. � R ≤ A Xð Þ ¼ ∪ x½ �≤Aj x½ �≤A⊆X; x 2 U � � ¼ x 2 Uj x½ �≤A⊆X � � : 2. �R ≤ A Xð Þ ¼ ∪ x½ �≤Aj x½ �≤A∩X 6¼ ∅; x 2 U � � ¼ x 2 Uj x½ �≤A∩X 6¼ ∅ � � : Similar to the definition by Huynh and Nakamori (Huynh & Nakamori, 2005; Gong et al., 2008; Zhang, 2010), we give the following definitions. Definition 3.2. Let U; R≤A; F � � be interval-valued dominance information systems. For any X, Y ⊆ U, X and Y are called lower rough equal if R � ≤ A Xð Þ ¼ R� ≤ A Yð Þ. Denote as X≂Y. X and Y are called upper rough equal if R �≤ A Xð Þ ¼ R �≤ A Yð Þ. Denote as X ’ Y. X and Y are called rough equal if they are both lower and upper equal. Denote as X � Y. It is easytoknow that the operators≂, ’ and� are equivalence relations. Then, the following conclusions are satisfied. Theorem 3.2. Let U; R≤A; F � � be interval-valued dominance information systems. For any subset X, Y, X0, Y0 ⊆ U, the following properties hold: 1. X ≂ Y , (X∩ Y ) ≂X and (X ∩Y ) ≂Y, 2. X ’ Y , (X∪Y ) ’ X and (X ∪ Y ) ’ Y, 3. If X ≂ X 0 and Y ≂ Y 0, then (X ∩Y ) ≂ (X 0 ∩Y 0 ), 4. If X ’ X 0 and Y ’ Y 0, then (X ∪Y ) ’ (X 0 ∪Y0 ), 5. If X ≂ Y, then X∩ (∼Y ) ≂ ∅, 6. If X ’ Y, then X∪ (∼Y ) ’ U. Proof It is easy to prove by the aforementioned definition. In general, the formula will not be satisfied when ’ is replaced by ≂ in Theorem 3.2 and vice versa. Table 1: An interval-valued information systems U a1 a2 a3 x1 [0.1, 0.5] [0.2, 0.6] [0.1, 0.6] x2 [0.3, 0.8] [0.2, 0.6] [0.2, 0.7] x3 [0.1, 0.5] [0.1, 0.6] [0.2, 0.6] x4 [0.2, 0.9] [0.1, 0.7] [0.3, 0.8] x5 [0.3, 1] [0.3, 0.9] [0.2, 0.8] © 2013 Wiley Publishing LtdExpert Systems, xxxx 2013, Vol. 00, No. 00 Remark 3.2. Let U; R≤A; F � � be interval-valued dominance information systems. For any X,Y⊆ U, the following relations do not hold according to Remark 3.1 and Theorem 3.1. 1. R � ≤ A Xð Þ ¼ ∩ Y⊆UjY≂Xf g; 2. R �≤ A Xð Þ ¼ ∪ Y⊆UjY ’ Xf g: However, these two equations hold for real-valued information systems. This is the difference between the interval-valued dominance and real-valued information systems (Zhang et al., 2003a, 2003b). Definition 3.3. Let U; R≤A; F � � be interval-valued dominance information systems. For any X⊆U, the precision of X in U; R≤A; F � � is defined as follows: aA Xð Þ ¼ R � ≤ A Xð Þ ��� ��� R �≤ A Xð Þ ��� ��� Moreover, rA(X) = 1 � aA(X) is called the roughness of X in U; R≤A; F � � : The ratio gA Xð Þ ¼ R � ≤ A Xð Þ ��� ��� Uj j defines the quality of approximation of X by R≤A in U; R ≤ A; F � � : Clearly, there is 0 ≤ aA(X) ≤ 1 and aA(X) ¼ 1 if and only if R � ≤ A Xð Þ ¼ X ¼ �R ≤ A Xð Þ: 3.2. Dominance-based attribute reduction for interval-valued information systems In this section, we study attribute reduction for interval- valued dominance information systems. In general, objects are described by different attributes. However, it is not necessary to know all attributes for the classification of information systems. That is, some attributes are unnecessary and do not affect the result of classification when removed from the attribute set. Meanwhile, some attributes are indispensable to the result of classification when removed from the attribute set. Furthermore, some attributes are relatively necessary for the classification and may determine the result by associating with other attributes. The attribute reduction presents a minimum attribute subset completely describing the classification as the original attribute set for information systems. First of all, we define the reduction for the interval-valued dominance information systems. Let U; R≤A; F � � be interval-valued dominance information systems. If B ⊆ A satisfies the following relations, 1. R≤B ¼ R≤A; 2. For any b 2 B; R≤B� bf g 6¼ R≤A: then, B is called a reduction of U; R≤A; F � � : Clearly, it is easy to know that there is not only one reduction of interval-valued dominance information systems. In what follows, we give the concepts of discernibility attribute set and discernibility matrix. Definition 3.4. Let U; R≤A; F � � be interval-valued dominance information systems. DP x; yð Þ ¼ ai 2 A fai xð Þ ≰ fai yð Þ; x; y 2 Uj gf is called a discernibility attribute set of an interval-valued dominance information system. Remark 3.3. According to Remark 2.1, we know that the discernibility attribute set will be defined as DP x; yð Þ ¼ ai 2 A fai xð Þ > fai yð Þ; x; y 2 Uj gf when U; R≤A; F � � is the real-valued information system. Remark 3.4. The relation of DP(x,y) ∩DP(y,x) = ∅, x, y 2 U could not hold in U; R≤A; F � � , but it is correct for real-valued information systems. This is the difference between the interval-valued domi- nance and real-valued information systems. The following Example 3.2 shows this difference directly. Example 3.2. Continued from Example 3.1 By Table 1, we have, respectively, the values of elements x1 and x4 with respect to attribute A as follows: fa1 x1ð Þ ¼ 0:1; 0:5½ �; fa1 x4ð Þ ¼ 0:2; 0:9½ �; fa2 x1ð Þ ¼ 0:2; 0:6½ �; fa2 x4ð Þ ¼ 0:1; 0:7½ �; fa3 x1ð Þ ¼ 0:1; 0:6½ �; fa3 x4ð Þ ¼ 0:3; 0:8½ � By Definition 3.4, we obtain the discernibility attribute set of x1 and x4 as follows: DP x1; x4ð Þ ¼ ai 2 Aj fai x1ð Þ≰ fai x4ð Þf g ¼ a2f g; DP x4; x1ð Þ ¼ ai 2 Aj fai x4ð Þ≰ fai x1ð Þf g ¼ a1; a2; a3f g Then, DP(x1,x4)∩DP(x4,x1) = {a2} 6¼ ∅. In what follows, we discuss the attribute reduction for interval-valued dominance information systems. Denote DP ¼ DP x; yð Þ x; y 2 Uj gf Then, DP is called the discernibility matrix of interval- valued dominance information systems U; R≤A; F � � : Denote DP0 ¼ DP x; yð Þ DP x; yð Þ 6¼ ∅j gf By the aforementioned definitions, the following conclusions are clear. Theorem 3.4. Let U; R≤A; F � � be interval-valued dominance information systems. For any B ⊆ A, the following propositions are equivalent. 1. R≤B ¼ R≤A: 2. B ∩B0 6¼ ∅, for any B0 2 DP0: 3. If B ∩B0 = ∅, for any B0 ⊆A, then B0=2DP0: © 2013 Wiley Publishing Ltd Expert Systems, xxxx 2013, Vol. 00, No. 00 Proof Firstly, we prove that 1 is equivalent to 2. 1 ⇒ 2: For any x, y 2 U, there is x½ �≤A ¼ y½ �≤B according to R≤B ¼ R≤A .So, if y =2 x½ �≤B , it can be easily obtained that fai xð Þ≰ fai yð Þ for any ai 2 B, x 2 U by the definition of x½ �≤B. That is, ai 2 DP(x,y). It proves B0 ∩B 6¼ ∅ for any B0 2 DP0; x; y 2 U: 2 ⇒ 1: Clearly, there exists ai 2 B that satisfies ai 2 B0 according to B∩B0 6¼ ∅ for any B0 2 DP0 .This shows fai xð Þ ≰ fai yð Þ; x; y 2 U. So, x½ �≤B∩ y½ �≤B ¼ ∅. Furthermore, there are x½ �≤A⊆ x½ �≤B and y½ �≤A⊆ y½ �≤B by the relation B⊆A. So, there are x½ �≤A ¼ x½ �≤B and y½ �≤A ¼ y½ �≤B for any x, y 2 U. Then, R≤B ¼ R≤A: Therefore, 1 is equivalent to 2. The equivalence of 2 and 3 has a similar proof. By the aforementioned discussion, the following theorem is clear. Theorem 3.5. Let U; R≤A; F � � be interval-valued dominance information systems. For any B ⊆ A, then B is a reduction of U; R≤A; F � � if and only if the following conditions are satisfied: 1. B ∩B0 6¼ ∅, for any B0 2 DP0: 2. For any b 2 B, there exists B0 2 DP0 that satisfies (B � {b}) ∩B0 = ∅. Proof It is easy to prove by the definition of reduction and Theorem 3.4. Example 3.3. Continued from Example 3.1 Table 2 gives the discernibility matrix of interval-valued dominance information systems in Table 1. By the aforementioned definition, we can obtain the discernible attribute set as follows: DP0 ¼ a2f g; a3f g; a1; a2f g; a1; a3f g; Ag Let B ¼ {a2,a3}. By Table 1, it is easy to obtain the interval-valued dominance classes as follows: x1½ �≤B ¼ x1; x2; x5f g x2½ �≤B ¼ x2; x5f g x3½ �≤B ¼ x2; x3; x4; x5f g x4½ �≤B ¼ x4f g x5½ �≤B ¼ x5f g The conclusion (1) of Theorem 3.4 R≤B ¼ R≤A can be easily verified. Meanwhile, it is easy to verify that B ∩B0 6¼ ∅ holds for any B 0 2 DP0: Let B0 = {a1} and B0 = ∅, respectively. It also can be easily seen that the conclusion (3) of Theorem 3.4 holds. Thus, the validity of Theorem 3.4 is tested. Furthermore, Theorem 3.5 also can be easily illuminated as the way to Theorem 3.4. Therefore, B = {a2,a3} is a reduction of interval-valued dominance information systems U; R≤A; F � � : 4. Dominance-based knowledge reduction for interval- valued decision information systems In this section, we establish some conclusions of knowledge reduction for interval-valued decision information systems by the rough set theory based on interval-valued dominance relation. Actually, another similar problem in knowledge reduction is feature selection, which has been a hot problem investigated by many researchers from the artificial intelligence and machine learning community (Kohavi & John, 1997; Praczyk, et al., 1999; Dash & Liu, 1997, 2003; Xiong & Funk, 2006, 2010; Vale et al., 2010). Feature selection is an effective technique in dealing with dimensionality reduction. Many valuable algorithms and results have been established for feature selection problem by these scholars. Moreover, rough set theory also has been used as an effective approach to feature selection problems in recent years (Kuncheva, 1992; Thangavel & Pethalakshmi, 2009; Tsenga & Huang, 2007; Maria & Maite, 2011). In this paper, we discuss the knowledge reduction or feature selection problem for interval-valued decision information systems by using dominance-based rough set theory over interval-valued information systems in detail. Decision information systems are special information systems with condition and decision attributes simultaneously. The decision information systems study the interrelationship between condition and decision attributes, and then acquire the decision rule from information systems. Let U; R≤A; F; D; gD � � be interval-valued dominance decision information systems, where U; R≤A; F � � is an interval-valued dominance information system, and D is a decision attribute set and satisfies A∩D ¼ ∅. F ¼ fai U ! I½ �; ai 2 Aj gf is a family of mapping between U and A; gD : U ! VD is a discrete integer-valued mapping from universe U to decision attribute set D, where VD ¼ {1,2, . . .,k} is the domain of decision attribute. So, an equivalence relation RD is defined by the integer- valued mapping gD. That is, RD¼ {(x,y)|gD(x)¼ gD(y),x,y2 U} is an equivalence relation over universe U. Denote U=RD ¼ x½ �D x 2 Uj g � where [x]D = {y|(x,y) 2 RD, x, y 2 U}. In what follows, we present the concept of consistent and inconsistent interval-valued dominance decision information systems. Table 2: Discernibility matrix DP U x1 x2 x3 x4 x5 x1 ∅ ∅ {a2} {a2} ∅ x2 {a1, a3} ∅ A {a1, a2} ∅ x3 {a3} ∅ ∅ ∅ ∅ x4 A A A ∅ {a3} x5 A A A {a1, a2} ∅ © 2013 Wiley Publishing LtdExpert Systems, xxxx 2013, Vol. 00, No. 00 Definition 4.1. Let U; R≤A; F; D; gD � � be interval-valued dominance decision information systems. If R≤A⊆RD , that is, x½ �≤A⊆ x½ �D . Then U; R≤A; F; D; gD � � is called a consistent interval-valued dominance decision information system. Otherwise, U; R≤A; F; D; gD � � is called an inconsistent interval-valued dominance decision information system. Remark 4.1. The binary relation based on condition and decision attributes is different over consistent and inconsistent interval-valued dominance information systems, that is, a dominance relation on condition attribute and an equivalence relation on decision attribute. However, the binary relation defined on condition and decision attributes is identical for real-valued decision information systems. That is, they define two equivalence relations or two dominance relations on condition and decision attributes (Zhang et al., 2003a, 2003b; Zhang et al., 2006; Xu & Zhang, 2008; Xu et al., 2010). In what follows, we discuss the methods of knowledge reduction for consistent and inconsistent interval-valued dominance decision information systems, respectively. 4.1. Consistent interval-valued dominance decision information systems We first introduce the concept of inclusion degree. Definition 4.2. Zhang et al., 2003a, 2003b Let (X,≤) be a partial order set. For any x, y 2 X, there exists a number D(y/x) that satisfies the following conditions: 1. 0 ≤ D(y/x) ≤ 1. 2. x ≤ y ⇒ D(y/x) ¼ 1. 3. x ≤ y ≤ z ⇒ D(x/z) ≤ D(x/y). Then, D(•) is called the inclusion degree over X. Let U; R≤A; F; D; gD � � be interval-valued dominance decision information systems for any B⊆A. Denote U=A≤B ¼ x½ �≤B x 2 Uj g � U=RD ¼ D1; D2 . . . ; Drf g where Dj ¼ [xj]D, j ¼ 1, 2, . . ., r. Obviously, U=A≤B is a covering, but U/RD is a partition of universe U. Denote D Dj= x½ �≤B � � ¼ Dj∩ x½ � ≤ B �� �� x½ �≤B �� �� ; Dj 2 U=RD where | • | stands for the cardinality of the set. Then, D Dj= x½ �≤B � � is an inclusion degree of P(U) (where P(U) stands for the power set of U) according to Definition 4.2. On the basis of the inclusion degree D Dj= x½ �≤B � � , the following conclusion is clear. Theorem 4.1. Let U; R≤A; F; D; gD � � be consistent interval-valued dominance decision information systems. For any B ⊆ A, the following propositions are equivalent. 1. D x½ �D= x½ �≤B � � ¼ 1, for any x 2 U. 2. R≤B⊆RD: 3. 1Uj j X X2U=RD � R≤B Xð Þ ��� ��� ¼ 1: Proof Because D x½ �D= x½ �≤B � � ¼ 1 , x½ �≤B⊆ x½ �D, this proves that 1 is equivalent to 2. According to Theorem 3.1, we have � R≤B Xð Þ⊆X: Therefore, 1 Uj j X X2U=RD � R≤B Xð Þ ��� ��� ¼ 1 means that � R≤B Xð Þ ¼ X: That is, R≤B⊆RD: So, we prove that 3 is equivalent to 2. Lemma 4.1. Zhang et al., 2003a, 2003b Let P be a partition of universe U, P; ≤ð Þ be a partial order set and D be an inclusion degree over (P(U),⊆), for any two partitions A and B of universe U, where A ¼ E1; ; E2; . . . ; ; Ekf g; B ¼ F1; ; F2; . . . ; ; Flf g: Denote D B=Að Þ ¼ ∧ k i¼1 ∨ l j¼1 D Fj=Ei � � Then, B is an inclusion degree over P; ≤ð Þ. Proof It is easy to obtain that 0 ≤ D B=Að Þ ≤ 1 by 0 ≤ D(Fj/Ei) ≤ 1. Meanwhile, the relation A ≤ B means that there exists Fj 2 B and satisfies the relation Ei ⊆Fj for any Ei 2 A. This is equivalent to ∨lj¼1D Fj=Ei � � ¼ 1: Hence, it is also equivalent to D B=Að Þ ¼ 1: Given C ¼ G1; G2 . . . ; Grf g and A ≤ B ≤ C: Then, for any Ei 2 A, there exists Fj 2 B , and Gp 2 C satisfies the relation Ei ⊆ Fj ⊆ Gp. Because D is the inclusion degree over P Uð Þ; ⊆ð Þ, so there is D(Ei/Gp) ≤ D(Ei/Fj). Therefore, it proves the relation D A=Cð Þ ≤ D A=Bð Þ: Lemma 4.2. Let U; R≤A; F; D; gD � � be consistent interval-valued dominance decision information systems. For any B⊆A, denote CB R≤B � � as the covering of universe U that is generated by the interval- valued dominance relation R≤B. Then, D U=RDð Þ=CB R≤B � �� � ¼ min x2U D x½ �D= x½ �≤B � � and D U=RDð Þ=CB R≤B � �� � ¼ 1 Uj j X X2U=RD � R≤ B Xð Þ ��� ��� are the inclusion degree of universe U. Proof It can be easily proved by Theorem 4.1 and Lemma 4.1. In the following, we establish the definition of reduction and also provide the judgement theorem of knowledge © 2013 Wiley Publishing Ltd Expert Systems, xxxx 2013, Vol. 00, No. 00 reduction based on the inclusion degree for consistent interval-valued dominance decision information systems by Theorem 4.1 and Lemma 4.2. Definition 4.3. Let U; R≤A; F; D; gD � � be consistent interval-valued dominance decision information systems. For any B ⊆ A, if the following equations are satisfied 1. D U=RDð Þ=CB R≤B � �� � ¼ 1: 2. D U=RDð Þ=CB R≤B� bf g � �� � < 1 for any b 2 B. then B is called a reduction of (U,A,F,D,gD). Definition 4.4. Zhang et al., 2006; Xu & Zhang, 2008; Xu et al., 2010; Li et al., 2010 Let a ¼ (a1,a2, . . .,an), b ¼ (b1,b2, . . .,bn) be two n dimension vectors. If ai ¼ bi(i = 1,2, . . ., n), then a is called equal to b, denoted as a ¼ b. If ai ≤ bi(i = 1,2, . . ., n), then a is called less than b, denoted as a ≤ b. Otherwise, if there exists i0(i0 2 {1,2, . . .,n}), and ai0 > bi0 is satisfied, then a is defined as no greater than b, denoted as a ≰ b: Remark 4.2. If ai, bi 2 [I], then a and b are called n-dimension interval- valued vectors. Therefore, similar to Definition 4.4, the definition of the relations a ¼ b, a ≤ b and a ≰ b is clear for an n-dimension interval-valued vector. Remark 4.3. It can be easily seen that a≰b is not equivalent to a > b because both a and b are interval-valued vectors. So is a < b and a 6¼ b. In what follows, we present the description of knowledge reduction for consistent interval-valued dominance decision information systems. Definition 4.5. Let U; R≤A; F; D; gD � � be consistent interval-valued dominance decision information systems. Define Dg x; yð Þ ¼ ak 2 Ajfak xð Þ ≰ fak yð Þf g; gD xð Þ 6¼ gD yð Þ;A; gD xð Þ ¼ gD yð Þ; where fak xð Þ stands for the value of object x 2 U about attribute ak. Then, Dg(x,y) is called the dominated discernibility attribute set of objects x and y with respect to consistent interval-valued dominance decision information systems. Moreover, Dg = (Dg(x,y)|x, y 2 U) is called the dominated discernibility matrix of consistent interval-valued dominance decision information systems. Here, the term for the aforementioned definition comes from Zhang et al. (2006), but the conditions in the definition are defined in a different manner. Theorem 4.2. Let U; R≤A; F; D; gD � � be consistent interval-valued dominance decision information system. For any B ⊆ A, 1. 8x; y 2 U; R≤B⊆RD , B∩Dg x; yð Þ 6¼ ∅; 2. R≤B⊆RD , 8B 0⊆A, if B∩B0 = ∅, then B0=2Dg: Proof 1. If R≤B⊆RD, then there is Dg(x,y) = A when gD(x) = gD(y). So, B∩Dg(x,y) 6¼ ∅. Meanwhile, there is y =2 x½ �D when gD(x) 6¼ gD(y). Then, x½ �≤B⊆ x½ �D is satisfied by R≤B⊆RD . So, y =2 x½ �≤B and ak 2 B satisfy fak xð Þ≰ fak yð Þ. That is, ak 2 Dg(x,y). This proves B ∩Dg(x,y) 6¼ ∅. Conversely, there is gD(x) 6¼ gD(y) when y =2 x½ �D for any x, y 2 U. Then, by B∩Dg(x,y) 6¼ ∅, we know there exists ak 2 B satisfying fak xð Þ≰ fak yð Þ. So, y =2 x½ �≤B: That is, x½ �≤B⊆ x½ �D: This proves R≤B⊆RD: 2. By the conclusion of 1, R≤B⊆RD , for any B0 2 Dg, and there is B ∩B0 6¼ ∅. By Theorem 4.2, the following conclusion is clear. For B ⊆A, denote g Bð Þ ¼ Πx;y2UwB Dg x; yð Þ � � where wB Dg x; yð Þ � � ¼ 1; B∩Dg x; yð Þ 6¼ ∅; 0; B∩Dg x; yð Þ ¼ ∅ Theorem 4.3. Let U; R≤A; F; D; gD � � be consistent interval-valued dominance decision information systems. Dg = {Dg(x,y)|x, y 2 U} is a dominated discernibility matrix. For any B⊆A, there is g Bð Þ ¼ 1 , R≤B⊆RD Proof It is easy to prove by the aforementioned definition. Theorems 4.2 and 4.3 give the conditions of sufficiency and necessity of knowledge reduction for consistent interval-valued dominance decision information systems. Next, we discuss the knowledge reduction for inconsistent interval-valued dominance decision information systems. 4.2. Inconsistent interval-valued dominance decision information systems In general, we also can define the distribution reduction, generalized distribution reduction and maximum distribution reduction with corresponding judgement theorems of knowledge reduction similar to real-valued inconsistent decision, inconsistent ordered and set-valued decision and others (Wang & Zhang, 2006; Xu & Zhang, 2008; Xu et al., 2010; Li et al., 2010; Zhang, 2010). In this section, we discuss another two reductions for inconsistent interval-valued dominance decision information systems. We begin with some basic definitions as follows. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dominance decision information systems. © 2013 Wiley Publishing LtdExpert Systems, xxxx 2013, Vol. 00, No. 00 For any B ⊆ A, denote �B ¼ 1 Uj j Xr j¼1 �R ≤ B Dj � � j �� dB xð Þ ¼ Dj Dj∩ x½ �≤B 6¼ ∅; x 2 U �� �� where Dj 2 U/RD ¼ {[x]D|gD(x) ¼ gD(y), x, y 2 U}. It is easy to know that the following two propositions are clear by dB(x) and �B. Proposition 4.1. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dominance decision information systems. For any B⊆A, 1. dA(x) ⊆dB(x) for any x 2 U. 2. If x½ �≤B⊆ y½ �≤B, then dB(x) ⊆dB(y)x, for any y 2 U. Proposition 4.2. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dom- inance decision information systems. For any B ⊆A, the following conditions hold. 1. 1 ≤ �B ≤ |U/RD|. 2. For any Dj 2 U/RD, �B ¼ 1 , �R ≤ B Dj � � ¼ Dj; �B ¼ U=RDj j , �R ≤ B Dj � � ¼ U; Dj 2 U=RD 3. If �R ≤ A Dj � � ⊆�R≤B Dj � � , then �A ≤ �B. On the basis of dB(x) and �B, we give two reductions for inconsistent interval-valued dominance decision information system as follows. Definition 4.6. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dom- inance decision information systems. For any B⊆ A, 1. If dA(x) ¼ dB(x) for any x 2 U, then B is called a dominated allocation consistent set. Furthermore, if B � {a} for any a 2 B is not a dominated allocation consistent set, then B is called a dominated allocation reduction; 2. If �B ¼ �A for any x 2 U, then B is called an approximation-dominated consistent set. Furthermore, if B � {a} for any a 2 B is not an approximation- dominated consistent set, then B is called an approximation-dominated reduction. Theorem 4.4. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dom- inance decision information systems. For any B ⊆ A, then B is a dominated allocation consistent set if and only if B is an approximation-dominated consistent set. Proof Let B be a dominated allocation consistent set. That is, dB (x) ¼ dA(x) for any x 2 U. Then, x 2 �R≤B Dj � � , x½ �≤B∩Dj 6¼ ∅ , Dj 2 dB xð Þ , Dj 2 dA xð Þ , x½ �≤A∩Dj 6¼ ∅ , x 2 �R ≤ A Dj � � This proves �R ≤ B Dj � � ¼ �R≤A Dj � � : Therefore, �B ¼ �A, and B is the approximation- dominated consistent set. Conversely, if B is an approximation-dominated consistent set, then �B ¼ �A. That is, Xr j¼1 �R ≤ B Dj � ��� �� ¼ Xr j¼1 j�R≤A Dj � � j: Meanwhile, �R ≤ A Dj � � ⊆�R≤B Dj � � holds for any Dj 2 U/RD. Therefore, �R ≤ A Dj � � ¼ �R≤B Dj � � : So, Dj 2 dB xð Þ , x½ �≤B∩Dj 6¼ ∅ , x 2 �R≤B Dj � � , x 2 �R≤A Dj � � , x½ �≤A∩Dj 6¼ ∅ , Dj 2 dA xð Þ This proves dB(x) ⊆dA(x). Furthermore, we know that dA(x)⊆dB(x) by Proposition 4.2. Thus, we prove that dA(x)¼ dB(x). This proves that B is a dominated allocation consistent set. Example 4.1. Table 3 gives an inconsistent interval-valued dominance decision information system. It is easy to obtain the interval-valued dominance classes and equivalence classes by condition attribute A and decision attribute D, respectively, as follows: x1½ �≤A ¼ x1; x2; x5f g x2½ �≤A ¼ x2; x5f g x3½ �≤A ¼ x2; x3; x4; x5f g x4½ �≤A ¼ x4f g x5½ �≤A ¼ x5f g and x1½ �D ¼ x5½ �D ¼ x1; x5f g x2½ �D ¼ x4½ �D ¼ x2; x4f g x3½ �D ¼ x3f g It can be easily seen that ∪x2U x½ �≤A is a covering, and ∪ x 2 U[x]D is a partition of universe U. Let D1 ¼ x1½ �D ¼ x5½ �D D2 ¼ x2½ �D ¼ x4½ �D D3 ¼ x3½ �D Table 3: Inconsistent interval-valued dominance decision information system U a1 a2 a3 D x1 [0.1, 0.5] [0.2, 0.6] [0.1, 0.6] 3 x2 [0.3, 0.8] [0.2, 0.6] [0.2, 0.7] 2 x3 [0.1, 0.5] [0.1, 0.6] [0.2, 0.6] 1 x4 [0.2, 0.9] [0.1,0.7] [0.3, 0.8] 2 x5 [0.3, 1] [0.3,0.9] [0.2, 0.8] 3 © 2013 Wiley Publishing Ltd Expert Systems, xxxx 2013, Vol. 00, No. 00 We calculate the value of dA(x)(x 2 U) as follows: dA x1ð Þ ¼ D1; D2f g; dA x2ð Þ ¼ D1; D2f g; dA x3ð Þ ¼ D1; D2; D3f g; dA x4ð Þ ¼ D2f g; dA x5ð Þ ¼ D1f g Taking B ¼ {a2,a3} ⊆A ¼ {a1,a2,a3}. for any x 2 U, we can easily obtain that x½ �≤A ¼ x½ �≤B: Then, dB(x) ¼ dA(x) holds. So, B is a dominated allocation consistent set. In what follows, we give the two descriptive theorems of knowledge reduction for inconsistent interval-valued dominance decision information systems. Theorem 4.5. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dom- inance decision information systems. For any B⊆A, x, y 2 U, B is a dominated allocation consistent set if and only if x½ �≤B⊈ y½ �≤B when dA xð Þ⊈dA yð Þ: Proof 1. Suppose that x½ �≤B⊈ y½ �≤B does not hold when dA xð Þ⊈dA yð Þ for any x, y 2 U. That is, x½ �≤B⊆ y½ �≤B; then we know dB(x) ⊆ dB(y) by Proposition 4.1. Moreover, B is a dominated allocation consistent set. So, dA(x) ¼ dB(x) and dA(y) ¼ dB(y). Therefore, there is dA(x) ⊆ dA(y). This is a contradiction. Conversely, we only need to prove that dB(x) ¼ dA(x) for any x 2 U. It is easy to know that dA(x)⊆ dB(x) holds. Then, we only prove dB(x) ⊆dA(x). In what follows, we prove this conclusion by two different cases. • If dB(x) ¼ ∅, the conclusion holds. • If dB(x) 6¼ ∅, x½ �≤B ⊈ y½ �≤B when dA xð Þ ⊈ dA yð Þ: That is, if x½ �≤B⊆ y½ �≤B , then dA(x) ⊆dA(y) holds for any x, y 2 U. Because dB(x) 6¼ ∅ andDj∩ x½ �≤B 6¼ ∅hold for any Dj 2 U/RD, let y 2 Dj∩ x½ �≤B . Then, there are y 2 x½ �≤B and y 2 Dj. So, y½ �≤B⊆ x½ �≤B: This proves that dA(y) ⊆dA(x) holds. It is easy to know that y 2 y½ �≤A, that is, y 2 Dj∩ y½ �≤A: So, Dj∩ x½ �≤A⊆Dj∩ y½ �≤A: This proves that dB(x) ⊆dA(y). The aforementioned two cases prove that dB(x)⊆dA(x) hold. So, the conclusion is proved. The following is a corollary. Corollary 4.1. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dom- inance decision information systems. For any B ⊆ A, the following two conclusions are equivalent. 1. B is an approximation-dominated consistent set. 2. x½ �≤B⊈ y½ �≤B when dA xð Þ⊈dA yð Þ; x; y 2 U: On the basis of dominated consistent attribute sets, we can obtain the reduction for inconsistent interval-valued dominance decision information systems. In what follows, we give another descriptive theorem of knowledge reduction for inconsistent interval-valued dominance decision information systems. First of all, we present the following definition. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dominance decision information systems. Denote D�d ¼ x; yð Þ dA xð Þ⊆dA yð Þj gf Definition 4.7. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dom- inance decision information systems. Define Dd x; yð Þ ¼ fak 2 A fak xð Þ≰ fak yð Þj g; x; yð Þ 2 D�d; A; x; yð Þ =2 D�d ( Then, Dd(x,y) is called the dominated allocation discernibility attribute set of x and y. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dominance decision information systems. Denote Dd ¼ Dd x; yð Þ x; y 2 Uj Þð Then, Dd is called the dominated allocation discernibility matrix for inconsistent interval-valued dominance decision information systems. On the basis of the aforementioned definitions, we have the following descriptive theorem. Theorem 4.6. Let U; R≤A; F; D; gD � � be inconsistent interval-valued dom- inance decision information systems. For any B ⊆ A, B is a dominated allocation consistent set if and only if B ∩ Dd(x,y) 6¼ ∅ for any x; yð Þ 2 D�d: Proof 1. For any x, y 2 U, dA xð Þ⊈dA yð Þ holds when (x,y) 2 U. Furthermore, we have x½ �≤B⊆ y½ �≤B because B is a dominated allocation consistent set. Then, there are three cases for x½ �≤B and y½ �≤B: (a) x½ �≤B⊈ y½ �≤B; (b) x½ �≤B∩ y½ �≤B ¼ ∅; (c) x½ �≤B∩ y½ �≤B⊆ x½ �≤B and x½ �≤B∩ y½ �≤B⊆ y½ �≤B: In what follows, we prove B ∩Dd(x,y) 6¼ ∅ for the aforementioned three cases. (a) If x½ �≤B⊈ y½ �≤B, there exists z 2 y½ �≤B that satisfies z =2 x½ �≤B. Then, there exists ak 2 B that satisfies fak xð Þ≰fak zð Þ: So, fak yð Þ≤fak zð Þ according to z 2 y½ �≤B: Therefore, fak yð Þ≤fak xð Þ: This proves ak 2 Dd(x,y). That is, B ∩Dd(x,y) 6¼ ∅. © 2013 Wiley Publishing LtdExpert Systems, xxxx 2013, Vol. 00, No. 00 (b) If x½ �≤B∩ y½ �≤B ¼ ∅ , there exists ak 2 B that satisfies fak xð Þ≰fak yð Þ. That is, B ∩Dd(x,y) 6¼ ∅. Otherwise, there is fak xð Þ ≤ fak yð Þ for any ak 2 B. So, it proves that y 2 x½ �≤B: This is contradicted by x½ �≤B∩ x½ �≤B ¼ ∅: (c) If x½ �≤B∩ y½ �≤B⊆ x½ �≤B and x½ �≤B∩ y½ �≤B⊆ y½ �≤B , then it can be proved in the same way as (a). Considering the aforementioned discussion, B ∩Dd (x, y) 6¼ ∅ when B is a dominated allocation consistent set for x; yð Þ 2 D�d: Conversely, there is B ∩Dd(x,y) 6¼ ∅ for x; yð Þ 2 D�d: Then, there exists ak 2 B satisfying ak 2 Dd(x,y). So, we have fak xð Þ≰fak yð Þ. That is, y=2 x½ �≤B: Furthermore, there is x½ �≤B⊈ y½ �≤B because y 2 y½ �≤B . Meanwhile, there is dA xð Þ⊈dA yð Þ for any x; yð Þ 2 D�d: Hence, we have x½ �≤B⊈ y½ �≤B when dA xð Þ⊈dA yð Þ: Thus, B is a dominated allocation consistent set. Example 4.2. Continued from Example 4.1 Table 4 gives the discernibility matrix for inconsistent interval-valued dominance decision information systems. From Table 4, it easy to know that B = {a2,a3} is a dominated allocation reduction of an inconsistent interval-valued dominance decision information system U; R≤A; F; D; gD � � : Moreover, the following results can be easily calculated according to Example 4.1: x1½ �≤B ¼ x1; x2; x5f g; x2½ �≤B ¼ x2; x5f g; x3½ �≤B ¼ x2; x3; x4; x5f g; x4½ �≤B ¼ x4f g; x5½ �≤B ¼ x5f g and dA x1ð Þ ¼ D1; D2f g; dA x2ð Þ ¼ D1; D2f g; dA x3ð Þ ¼ D1; D2; D3f g; dA x4ð Þ ¼ D2f g; dA x5ð Þ ¼ D1f g By these results, we can verify the condition: If dA xð Þ⊈dA yð Þ, then x½ �≤B⊈ y½ �≤B, for any x, y 2 U. This is the conclusion of Theorem 4.5. Furthermore, it can be easily known that Dd(x,y) ¼ {{a2}, {a3}, {a1,a2}, {a1,a3}, A} by Definition 4.7. Then we have B ∩Dd(x,y) 6¼ ∅, for any x, y 2 U. This is the conclusion of Theorem 4.6. 5. Conclusions and remarks Knowledge reduction over information systems and the generalization of real-valued information systems are two main research directions both in theory and application of rough set theory. This paper studies a type of extended real-valued information systems, that is, interval-valued information systems. Moreover, we study the basic rough set theory over interval-valued information systems. We systematically discuss attribute reduction for interval-valued dominance information systems based on the proposed rough set model. To discuss knowledge reduction for interval-valued dominance decision information systems, some concepts concerning the reduction of both consistent and inconsistent interval-valued dominance decision information systems have been established by the definitions of an interval-valued dominance relation and an equivalence relation in U; R≤A; F; D; gD � � .Some descriptive theorems for knowledge reduction are given in detail for interval-valued dominance decision information systems. It is also an extension of real-valued decision information systems. A numerical example is applied to test the methods and conclusions of the interval-valued dominance decision information systems as well. There are at least two aspects in the study of rough set theory: constructive and axiomatic approaches (Wu and Zhang 2004). The same is true for interval-valued information systems. In this paper, we define the rough lower (upper) approximation operator and discuss the basic properties by the constructive method. So further work should consider the axiomatic approaches to rough lower (upper) approximation operator and decision-making in the context of an interval-valued environment. Acknowledgements The authors thank the editor-in-chief Dr Jon Hall and the anonymous reviewers for valuable comments to improve the quality of this paper. The work was partly supported by the National Science Foundation of China (71161016 and 71071113), the Foundation for the Author of National Excellent Doctoral Dissertation of PR China (200782), the Shuguang Plan of Shanghai Education Development Foundation and Shanghai Education Committee (08SG21), the Shanghai Pujiang Program, and Shanghai Philosophical and Social Science Program (2010BZH003) and the Fundamental Research Funds for the Central Universities. The third author is supported by the National Science Foundation of China (71061013). References BEYNON, M., B. CURRY and P. MORGAN (2000) Classification and rule induction using rough sets theory, Expert Systems, 17, 136–148. CHEN, Y. and D.Q. MIAO (2011) Rule extraction based on granulation order in interval-valued fuzzy information system, Expert Systems with Applications, 38, 12249¨C12261. DAI, J.H., W.T. WANG, Q. XU and H.W. TIAN (2012) Uncertainty measurement for interval-valued decision systems based on extended conditional entropy, Knowledge-Based Systems, 27, 443–450. DASG, M. and H. LIU (2003) Consistency-based search in feature selection, Artificial Intelligence, 151, 155–176. DASH, M. and H. LIU, (1997) Feature selection for classification, Intelligent Data Analysis, 1, 131. Table 4: Dominated allocation discernibility matrix Dd U x1 x2 x3 x4 x5 x1 ∅ ∅ {a2} A A x2 {a1, a3} ∅ A A A x3 {a3} ∅ ∅ A A x4 A A A ∅ A x5 A A A {a1, a2} ∅ © 2013 Wiley Publishing Ltd Expert Systems, xxxx 2013, Vol. 00, No. 00 FENG, L., G.Y. WANG and X.X. LI (2010) Knowledge acquisition in vague objective information systems based on rough sets, Expert Systems, 17, 129–142. GONG, Z.T., B.Z. SUN and D.G. CHEN (2008) Rough set theory for the interval-valued fuzzy information systems, Information Scineces, 178, 1968–1985. GORZAFCZARY, B. (1988) Interval-valued fuzzy controller based on verbal modal of object, Fuzzy Sets and Systems, 28, 45–53. GRECO, S., B. MATARAZZO and R. SLOWINSKI (1999) Rough approximation of a preference relation by dominance relations, European Journal of Operational Research, 117, 63–83. GRECO, S., B. MATARAZZO and R. SLOWINSKI (2001) Rough sets theory for multicriteria decision analysis, European Journal of Operational Research, 129, 1–47. GRECO, S., B. MATARAZZO and R. SLOWINSKI (2002) Rough approximation by dominance relation, International Journal of Intelligent Systems, 17, 153–171. GRZYMALA, B.J.W. (1997) A new version of the rule induction system LERS, Fundamenta Informaticae, 31, 27–39. GUAN, Y.Y. and H.K. WANG (2006) Set-valued information systems, Information Sciences, 176, 2507–2525. HUANG, B. (2012) Graded dominance interval-based fuzzy objective information systems, Knowledge-Based Systems, 24, 1004–1012. HUANG, J.J., C.S. ONG and G.H. TZENG (2006) Interval multidimensional scaling for group decision using rough set concept, Expert Systems with Applications, 31, 525–530. HUANG, B., D.K. WEI, H.X. LI and Y.L. ZHANG (2012a) Using a rough set model to extract rules in dominance-based interval- valued intuitionistic fuzzy information systems, Information Sciences, DOI.org/10.1016/j.ins.2012.09.010. HUANG, B., H.X. LI and D.K. WEI (2012b) Dominance-based rough set model in intuitionistic fuzzy information systems, Knowledge- Based Systems, 28, 115–123. HUYNH, V.N. and Y. NAKAMORI (2005) A roughness measure for fuzzy sets, Information Scineces, 173, 255–275. KAZIMIERZ, Z. (2004) Rough approximation of a preference relation by a multi-attribute dominance for deterministic, stochastic and fuzzy decision problems, European Journal of Operational Research, 159, 196–206. KIM, K.J. and I. HAN (2001) The extraction of trading rules from stock market data using rough sets, Expert Systems, 18, 194–202. KOHAVI, R. and JOHN, G.H. (1997) Wrapper for feature subset selection, Artificial Intelligence, 97, 273–324. KRYSZKIEWICZ, M. (1998) Rough set approach to incomplete information systems, The Information of the Science, 112, 39–49. KRYSZKIEWICZ, M. (2001) Comparative study of alternative type of knowledge reduction in inconsistent systems, International Journal of Intelligent Systems, 16, 105–120. KUNCHEVA, L.L. (1992) Fuzzy rough sets: application to feature selection, Fuzzy Sets and Systems, 51, 147–153. LEUNG, Y., W.Z. WU and W.X. ZHANG (2006) Knowledge acquisition in incomplete information systems: a rough set approach, European Journal of Operational Research, 168, 164–180. LEUNG, Y., M.F. MANFRED, W.Z. WU and J.S. MI (2008) A rough set approach for the discovery of classification rules in interval- valued information systems, International Journal of Approximate Reasoning, 47, 233–246 LINGRAS, P.J. and YAO, Y.Y. (1998) Data mining using extensions of the rough set model, Journal of the American Society for Information Science, 49, 415–422. LI, Y., J. ZHAO, N.X. SUN and SAMKAR, K.P. (2010) Generalized distribution reduction reduction in inconsistent decision systems based on dominance relations, J. Xu et al., (Eds.): RSKT, 6401, 151–158. MARIA, S. and L.S. MAITE (2011) Rough set based approaches to feature selection for case-based reasoning classifiers, Pattern Recognition Letters, 32, 280–292. MI, J.S., W.Z. WU and W.X. ZHANG (2004) Approach to knowledge reduction based on variable precision rough set model, Information Sciences, 159, 255–272. PAWLAK, Z. (1982) Rough sets. International Journal Information Science, 11, 341–356. PAWLAK, Z. (1998) An inquiry into anatomy of conflicts, European Journal of Operational Research, 109, 184–190. PAWLAK, Z. and SKOWRON, A. (2007) Rough sets: some extensions, Information Sciences, 177, 28–40. PRACZYK, K., H. KIENDL and T. SLAWINSKI (1999) Finding relevant process characteristics with a method for data-based complexity reduction, Lecture Notes in Computer Science 1625, 548–555. QIAN, Y.H., J.Y. LIANG and C.Y. DANG (2008) Interval ordered information systems, Computers and Mathematics with Applications, 56, 1994–2009. QIAN, Y.H., C.Y. DANG, J.Y. LIANG, and D.W. TANG (2009) Set- valued ordered information systems, Information Sciences, 179, 2809¨C2832. SKOWRON, A. and J. STEPANIUK (1996) Tolerance approximation space, Fundanation Information, 27, 245–253. SLEZAK, D. (2002) Approximation entropy reducts, Fundamenta Information, 53, 365–390. SUN, B.Z., Z.T. GONG and D.G. CHEN (2008) Fuzzy rough set theory for the interval-valued fuzzy information systems, Information Scineces, 178, 2794–2815. THANGAVEL, K. and A. PETHALAKSHMI (2009) Dimensionality reduction based on rough set theory: a review, Applied Soft Computing, 9,1–12. TSENGA, T.L. and C.C. HUANG (2007) Rough set-based approach to feature selection in customer relationship management, Omega, The International Journal of Management Science, 35,365–383. TSUMOTO, S. (1998) Automated extraction of medical expert system rules from clinical databases based on rough set theory, Information Sciences, 112, 67–84. TSUMOTO, S. (2004) Mining diagnostic rules from clinical databases using rough sets and medical diagnostic model, The Information of the Science, 162, 65–80. TURKSEN, B. (1986) Interval-valued fuzzy sets based on normal forms, Fuzzy Sets and Systems, 20, 191–210. VALE, K.O., A.F. NETO and A.M.P. CANUTO (2010) Using a reinforcement-based feature selection method in classifier ensemble, Hybrid Intelligent Systems (HIS), 10th International Conference on, 23–25, 213–218. WANG, J.Y. and R.B. LIN (2006) A new knowledge reduction in inconsistent decision information system, IEEE International Conference on Granular Computing, 377–380. WANG, H and W.X. ZHANG (2006) Knowledge reduction in set- valued decision information systems, Lecture Notes in Control and Information Sciences, 344, 927–932. WU, W.Z. (2008) Attribute reduction based on evidence theory in incomplete decision systems, Information Sciences, 178, 1355–1371. WU, W.Z., W.X. ZHANG (2004) Constructive and axiomatic approaches of fuzzy approximation operators, The Information of the Science, 159, 233–254. WU, W.Z., Y. LEUNG and J.S. SHENG (2005) On characterizations of (I, J)-fuzzy rough approximation operators, Fuzzy Sets and Systems, 154(1), 76–102. WU, W.Z., W.X. ZHANG and H.Z. LI (2003) Knowledge acquisition in incomplete fuzzy information systems via the rough set approach, Expert Systems, 20, 280–286. XIONG, N. and P. FUNK (2006) Construction of fuzzy knowledge bases incorporating feature selection, Soft Computing, 9(10), 796–804 XIONG, N. and P. FUNK (2010) Combined feature selection and similarity modeling in case-based reasoning using hierarchical memetic algorithm, In: Proceedings of the IEEE World Congress on Computational Intelligence, 1537–1542. XU, W.H. and W.X. ZHANG (2008) Methods for knowledge reduction in inconsistent ordered information systems, Journal of Applied Mathematics and Computing, 26, 313–323. XU, W.H., X.Y. ZHANG, J.M. ZHANG and W.X. ZHANG (2010) Attribute reduction in ordered information systems based on evidence theory, Knowledge and Information Systems, 25, 169–184. YANG, H. and C. SHENG (2011) New roughness measures of the interval-valued fuzzy sets, Expert Systems with Applications, 38, 2849–2856. YANG, X.B., D.J. YU, J.Y. YANG and L.H. WEI (2009) Dominance- based rough set approach to incomplete interval-valued information system, Data & Knowledge Engineering, 68, 1331–1347. YAO, Y.Y. (2003) Probabilistic approaches to rough sets, Expert Systems, 20, 2871–297. © 2013 Wiley Publishing LtdExpert Systems, xxxx 2013, Vol. 00, No. 00 http://DOI.org ZHANG, Z.M. (2010) An interval-valued rough intuitionistic fuzzy set model, International Journal of General Systems, 39, 135–164. ZHANG, W.X., Y. LEUNG and W.Z. WU (2003a) Information Systems and Knowledge Discovery, Beijing: Science Press, 13–56. ZHANG, M., L.D. XU, W.X. ZHANG and H.Z LI (2003b) A rough set approach to knowledge reduction based on inclusion degree and evidence reasoning theory, Expert Systems, 20, 298–304. ZHANG, W.X., Y.Y. YAO and Y. LEUNG (2006) Rough Sets and Concept Lattices. Xi’an: Xi’an Jiaotong University Press, 325–354. ZHANG, H.Y., W.X. ZHANG and W.Z. WU (2009) On characterization of generalized interval-valued fuzzy rough sets on two universes of discourse, International Journal of Approximate Reasoning, 51, 56–70. ZHU, W. (2007) Generalized rough sets based on relations, Information Sciences, 177, 4997–5011. ZIARKO, W. (2003) Acquisition of hierarchy structured probabilistic decision tables and rules from data, Expert Systems, 20, 305–310. The authors Bingzhen Sun Bingzhen Sun is currently a PhD candidate of the School of Economics & Management, Tongji University in Shanghai, China. He received the BS degree from the Department of Mathematics and Applied Mathematics at the Northwest Normal University in Lanzhou, China, in 2003. He received the MS degree in applied mathematics from the College of Mathematics and Information Science, Northwest Normal University in Lanzhou, China, in 2006. His research interests include rough sets theory and applications, fuzzy sets and system, decision-making with uncertainty and operations research. Weimin Ma Weimin Ma, a professor of the School of Economics & Management, serves as an authorized PhD supervisor in management science and engineering at Tongji University in Shanghai, China. He received the BS degree from the Department of Mechanical Manufacturing Engineering at the Northwest Polytechnic University in Xi’an, China, in 1993. He received the MS and PhD degrees in management science and engineering from the School of Economics & Management, Xi’an Jiaotong University, Xi’an, China, in 1999 and 2003, respectively. His research interests include on-line computation, fuzzy sets and system, information system and technology, decision-making with uncertainty, algorithm design and operations research. Zengtai Gong Zengtai Gong received his PhD degree in science from the Harbin Institute of Technology in China. He is a professor and doctor supervisor at the College of Mathematics and Statistics, Northwest Normal University in China. His research directions include rough set theory, fuzzy analysis, real analysis and their applications in environment problems and management of water resources. © 2013 Wiley Publishing Ltd Expert Systems, xxxx 2013, Vol. 00, No. 00 
work_5ygoblmckzhlho4hus7dwkp4iy	----	Microsoft Word - Pinter _2010_ --- Calibrating ANNs by GO.docx Calibrating Artificial Neural Networks by Global Optimization János D. Pintér1,2 1: Faculty of Engineering, Özyeğin University, Kusbakisi Caddesi, �o.2, 34662 Istanbul, Turkey; www.ozyegin.edu.tr; janos.pinter@ozyegin.edu.tr. 2: Pintér Consulting Services, Inc., Canada; www.pinterconsulting.com; janos.d.pinter@gmail.com. Technical Report, Özyeğin University, Istanbul Submitted for publication: July 2010 ABSTRACT An artificial neural network (ANN) is a computational model − implemented as a computer program − that is aimed at emulating the key features and operations of biological neural networks. ANNs are extensively used to model unknown or unspecified functional relationships between the input and output of a “black box” system. In order to apply such a generic procedure to actual decision problems, a key requirement is ANN training to minimize the discrepancy between modeled and measured system output. In this work, we consider ANN training as a (potentially) multi-modal optimization problem. To address this issue, we introduce a global optimization (GO) framework and corresponding GO software. The practical viability of the GO based approach is illustrated by finding close numerical approximations of (one-dimensional, but non-trivial) functions. KEYWORDS Artificial Neural Networks, ANN model calibration by global optimization, Lipschitz Global Optimizer (LGO) solver suite, ANN implementation in Mathematica, MathOptimizer Professional (LGO for Mathematica), Illustrative numerical examples. 1. Introduction A neuron is a single nerve cell consisting of a nucleus (cell body, the cell’s life support center), incoming dendrites with endings called synapses that receive stimuli (electrical input signals) from other nerve cells, and an axon (nerve fiber) with terminals that deliver output signals to other neurons, muscles or glands. The electrical signal traveling down on the axon is called neural impulse: this signal is communicated to the synapses of some other neurons. This way, a single neuron acts like a signal processing unit; the entire neural network has a complex, massive multi-processing architecture. Neural systems are essential for an organism, in order to adapt properly and quickly to its environment. An artificial neural network (ANN) is a computational model − implemented as a computer program − that is aimed at emulating the essential features and operations of neural networks. ANNs are used extensively to model complex relationships between input and output data sequences, or to find hidden patterns in data sets when the analytical (explicit model function based) treatment of the problem is tedious or currently not possible. ANNs replace these unknown functional relationships by adaptively constructed approximating functions (approximators). Such approximators are typically designed on the basis of training examples of a given task, such as recognising hand- written words. For in-depth expositions related to the ANN paradigm consult e.g. Hertz et al., (1991), Cichocki and Unbehauen (1993), Smith (1993), Golden (1996), Ripley (1996), Mehrotra et al. (1997), Bar-Yam (1999), Steeb (2005), Cruse (2006). The mathematical model of an ANN is based on the key features of a neuron as outlined above. We shall consider first a single artificial neuron (AN), and assume that a finite sequence of input signals s=(s1, s2, …, sT) is received by its synapses: the signals st t=1,…T may be real numbers or vectors. The AN’s information processing capability will be modeled by introducing an aggregator (transfer) function a that depends on a certain parameter vector x=(x1, x2, …, xn). Subsequently, x will be chosen according to a chosen optimality criterion. The optimization is typically carried out in such a way that the sequence of model output signals m=(m1, m2,…, mT) generated by the function a=a(x, s) is as close as possible to the corresponding set of observations o=(o1, o2,…, oT). This generic description can be summarized symbolically as follows. s → AN → m s=(s1, s2, …, sT), mt=a(x, st), t=1,…T. (1) minimize d(m, o) m=(m1, m2,…, mT), o=(o1, o2,…, oT), d is a discrepancy measure. A single AN modeled as above is called a perceptron. Flexible generalizations of a single perceptron based model are multiple perceptrons receiving the same input signals: such a structure is called an artificial neuron layer (ANL). To extend this structure further, multilayered networks can also be defined: in these several ANLs are connected sequentially. It is also possible to model several functional relationships simultaneously by the corresponding output sequences of a multi-response ANN. Figure 1 illustrates a multilayered feed-forward ANN that includes an input layer with multiple inputs, an intermediate (hidden) signal processing layer, and an output layer that emits multiple responses. The hidden layer’s AN units receive the input information, followed by certain decomposition, extraction, and transformation steps to generate the output information. Hence, the ANN implementation versions used in practice all share the features of adaptive, distributed, local, and parallel processing. Figure 1 Schematic representation of a multilayered ANN Source: http://en.wikipedia.org/wiki/Artificial_neural_network Following the generic description (1), the optimal parameterization strategy is summarized as follows. Given the functional form of a and the sequence s of input values, we want to find a parameterization x such that the difference between the output set m=a(x, s) and the observation set o is minimal. The discrepancy d(m, o) is typically expressed by a suitable norm function: d(m, o)=||m-o||. Once the parameter vector x becomes known − i.e. it has been estimated as a result of using the training I/O sequences s and o − the trained ANN can be used to perform various tasks related to the assessment of further input data such as {st} for t=T+1, T+2, and so on. Some important application examples of this paradigm are signal processing, data classification, and time series forecasting; further application areas will be mentioned later on. 2. Function Approximation by A��s: Theoretical Background and a Corresponding Optimization Model Formulation The key theoretical feature of ANNs is their ability to approximate (unknown) continuous functions by learning from observed data, under general analytical conditions. Specifically, the approximation theorem of Cybenko (1989) states that a feed-forward ANN with a single hidden layer that contains a finite number of neurons equipped with a suitable transfer function can serve as a universal approximator of any given continuous real-valued function f: Rk→R defined on [0,1]k. The approximation is uniform, in terms of the supremum norm of functions defined on [0,1]k. We will discuss this result below in some detail, and then introduce a corresponding optimization model. For a somewhat more complete picture of the theoretical foundations of ANNs, let us remark that Funahashi (1989) and Hornik et al. (1989) proved similar results to that of Cybenko. Hornik et al. (1989) proved that feed-forward neural networks with a single hidden layer of suitable transfer function units can approximate any Borel measurable multivariate vector function f:Rk1→Rk2 uniformly, to arbitrary prescribed accuracy on an arbitrary compact set C⊂Rk1. For more detailed technical discussions of these and closely related approximation results, consult also the foundations laid out by Kolmogorov (1957) which, however, do not lead directly to the results cited here since the approximating formula depends on the function f. More recent discussions of related function approximation results are presented by Park and Sandberg (1991, 1993), Lang (2005), Sutskever and Hinton (2008), with further topical references. Let us consider a single input signal st from R k k≥1, and the corresponding scalar output of ot = f(st) for all t=1,..,T: here f is an unknown continuous model function that we wish to approximate by a suitably defined ANN. We remark that in our general model the input signals can be delivered to a single (synapse of an) AN, or to several synapses of several ANs at the same time: the corresponding information will be then aggregated by the ANN. Cybenko (1989) establishes the theoretical possibility of generating a parameterized approximating function a(x, st), using a fixed type of transfer function σ such that the approximation shown below is uniformly valid. f(st)~a(x, st):=∑ j=1,…,J αj σ(∑ i=1,…,I yji sti + θj). (2) To relate this general result directly to an underlying ANN model, in (2) j=1,…,J denotes the index of the individual signal processing units ANj in the hidden layer. For each unit ANj, σ is a given function form to be specified below, αj, yj, and θj are model parameters to be determined depending on the unknown function f. Specifically, σ is a continuous real valued transfer function (in principle, σ could also be a function of j); ∑ i=1,…,I yji sti is the scalar product of the synaptic weight vector yj of ANj and the signal st=(st1,…, stI) component arriving from synapse i of ANj, for i =1,…,I; θj is called the threshold level (offset, bias) of ANj; and αj is the weight of ANj in the function approximation (2) for j=1,…,J. Following Cybenko (1989) and Hornik et al. (1989), the transfer function σ is assumed to be monotonically non-decreasing, and it satisfies the relations lim σ(z)=0 as z→-∞, and lim σ(z)=1 as z→∞. (3) Transfer functions with these properties are called sigmoidal or squashing functions. Let us remark here that other types of transfer functions can also be used in generating function approximations. Suitable sets of radial basis functions (RBF) are another well-known example (Park and Sandberg, 1991, 1993; Röpke et al., 2005). RBFs have the general form φ(z)=φ(||z−−−−z0||): here ||.|| is a given norm function, z0 is the (scalar or vector) centre of the monotonically non-increasing function φ that satisfies the conditions φ(0)=1, and lim φ(z)=0 as ||z−−−−z0||→∞. We will continue the present discussion using a specific sigmoidal transfer function form. An often used function form is σ(z)=1/(1+e-z) z∈R. (4) For proper theoretical and better numerical tractability, we assume and define finite box (variable range) constraints to bound the values of the parameters αj, yji and θj: αjmin≤αj≤αjmax for j=1,…,J. yjimin≤yji≤yjimax for i =1,…,I and j=1,…,J. θjmin≤θj≤θjmax for j=1,…,J. (5) The entire set of model parameters will be denoted by x=({αj}, {yji}, {θj}) where the appropriate components are included for all i=1,…,I and j=1,…,J. The optimization problem induced for a given approximator function form is to minimize the error of the approximation (2) for the entire training sequence (st, ot) t=1,…,T, using a selected norm function. In this study, we shall use the least squares error criterion to measure the quality of the approximation: in other words, the objective function (6) shown below will be minimized under the box constraints (5). e(x)=∑ t=1,…,T (f(st) − a(x, st)) 2 (6) The notation e(x) refers to the fact that (6) is an error function. Let us point out here that other (more general linear or nonlinear) constraints regarding feasible parameter combinations may also be present if dictated by the problem context. Such situations can also be handled by our model development approach, by simply adding these constraints to the optimization model formulation. An important case in point is to attempt the exclusion of identical (symmetric) solutions to in which the component functions of the function a are in principle exchangeable. To break this symmetry, one can assume e.g. that the components of {αj} are ordered so that αj≥αj+1 for all j=1,…,J-1. (7) One could also add further lexicographic constraints if dictated by the problem at hand. Another possible − but not always applicable − option could be to normalize the weights {αj}so that we have ∑ j=1,…,J αj=1, αj≥0 for all j=1,…,J. (8) 3. Postulating and Calibrating A�� Model Instances Cybenko (1989) emphasizes that his cited result only guarantees a uniformly valid approximation of a continuous function f, as opposed to its exact representation. We should also keep in mind that the general approximation result expressed by (2) does not specify the actual number of processing units: hence, the key model parameter J needs to be estimated for each problem instance separately. Cybenko (1989) conjectures that J could become “astronomical” in many function approximation problems ‒ especially so in higher dimensional approximations. These points should be kept in mind, to avoid making unjustifiably optimistic claims. Once we select (test or postulate) the ANN structure to use in an actual case study, the remaining basic requirement is its application-specific parameterization with respect to a given training data set. The objective function in (6) makes this a nonlinear model fitting problem that could have a multitude of local or global optima. This multimodality issue is discussed in detail e.g. by Pintér (1996, 2003) with further references; in the context of ANNs, consult e.g. Auer et al. (1996), Bianchini and Gori (1996), Abraham (2004). For a more complete picture, let us remark that detailed expository discussions related to global optimization models, algorithms and applications are presented e.g. in the topical handbooks edited by Horst and Pardalos (1995), Pardalos and Romeijn (2002). Let us also observe here that following the selection of the function form σ, the number of real-valued parameters (decision variables to optimize) is J(I+2), each with corresponding box constraints; furthermore, the number of optionally added linear constraints − considering (7) and (8) − is J. Since in the general ANN model setting the values of I and J cannot be theoretically specified a priori, in practical applications one can try several computationally affordable combinations. In general, ANNs with more components could serve as better approximators, within reason and in line with the size of the training data set that should at least satisfy T≥J(I+2). At the same time, in order to assure the success of a suitable nonlinear optimization procedure (and a corresponding software), one should restrict I and J to manageable values. This choice will then influence the accuracy of the approximation obtained. Summarizing the preceding discussion, first one has to postulate an ANN structure, and the family of parameterized transfer functions to use. Next, one needs to provide a set of training (input-output) data. The training is then aimed at the selection of an optimally parameterized function from the chosen family of function models, to minimize the error function. The iterative training of ANNs ‒ based on local optimization techniques ‒ is often referred to in the topical literature as back-propagation. Due to the possible multimodality of the error function (6), the simplest “standard” tools of unconstrained local nonlinear optimization (direct search, gradient based methods, Newton-type methods) often fail to locate the best parameter combination. This issue is well-known for ANN researchers: consult e.g. Bianchini and Gori (1996), Sexton et al. (1998), Abraham (2004), Hamm et al. (2007). To overcome this difficulty, the various optimization strategies used in practice to train ANNs include experimental design, sampling by low-discrepancy sequences, theoretically exact global or local scope search approaches, as well as popular heuristics such as evolutionary optimization methods, particle swarm optimization, simulated annealing, tabu search and tunneling functions. ANN model parameterization frameworks and numerical studies are presented and discussed e.g. by Watkin (1993), Prechelt (1994), Bianchini and Gori (1996), Sexton et al. (1998), Jordanov and Brown (1999), Toh (1999), Ye and Lin (2003), Abraham (2004), Hamm et al. (2007). Next we will discuss a theoretically exact global optimization based approach and a corresponding software implementation that will be used subsequently to solve several illustrative ANN calibration problems. Similarly to most other researchers, we will study “only” the problem of estimating the parameters of given ANNs with a given (postulated) architecture. The problem of finding the most appropriate model form in full generality seems elusive, and it is certainly difficult − if not impossible − to address in actual numerical studies. 4. Global Optimization: A General Modeling Framework To match the modeling approach and notations introduced earlier, we shall use the following symbols and assumptions: x∈Rn n-dimensional real-valued vector of decision variables e:Rn→R continuous (scalar-valued) objective function D⊂Rn non-empty set of feasible solutions, a proper subset of Rn: The feasible set D is closed and bounded, defined by l∈Rn, u∈Rn component-wise finite lower and upper bounds on x, and (optionally) by g:Rn→Rm an m-vector of continuous (more general) constraint functions. We shall then consider the continuous global optimization (GO) model min e(x) subject to x∈D:={x: l≤x≤u gj(x)≤0 j=1,...,m}. (9) Without going into details which are not necessary for the present discussion, let us remark that the terse model formulation (9) covers many special cases. Specifically, it also subsumes the ANN model calibration model statement summarized by formulas (4) to (6), with the optionally added constraints (7) and (8). Additionally, the GO problem statement (9) also guarantees that the set of global solutions is non-empty, and its structure supports the application of (theoretically) globally convergent deterministic and stochastic optimization algorithms. For technical details, consult e.g. Pintér (1996). Needless to say, the numerical handling of GO model instances of (9) can be a tough challenge: even small-dimensional model instances can be hard to handle, and they essentially require an appropriate global scope search strategy. These general notes apply also to many ANN model instances as it will be illustrated shortly (in Section 6). Consequently, the calibration of ANNs ‒ as a rule ‒ requires proper GO software: one such software product will be briefly described next. 5. The LGO Software Package for Constrained Global-Local Optimization Let us reiterate the key “black box” characteristic of the ANN modeling framework: namely, that its error function is evaluated by a numerical procedure that − in a model development and testing exercise − may be subject to changes and modifications. Similar situations that require (also) global scope search often arise in the real world of optimization: their solution requires easy-to-use, robust and efficient solver technology. Within the broad GO software category, direct (derivative-free) global search algorithm implementations have the advantage of immediate applicability to “black box” problems. Next, we shall briefly review the key features of the Lipschitz Global Optimizer (LGO) solver suite that has been designed to directly address also “black box” problems. Developed since the late 1980’s, LGO is currently available for a range of compiler platforms (C, C++, C#, Fortran 77/90/95), with seamless links to optimization modeling languages (AIMMS, AMPL, GAMS, MPL), MS Excel spreadsheets, and to the leading high-level technical computing systems Maple, Mathematica, and MATLAB. For theoretical, algorithmic and implementation details not discussed here, we refer e.g. to Pintér (1996, 1997, 2002, 2005, 2007, 2009, 2010a, 2010b), Pintér and Kampas (2005), Pintér et al. (2006). The overall design of LGO is based on the combination of continuous nonlinear optimization strategies, with corresponding theoretical global and local convergence properties. Hence, LGO can be used for both constrained global and local optimization. Let us point out here that LGO’s overall derivative-free design is different from many other nonlinear optimization software packages that require explicit analytical model information to support model parsing and function type specific operations. Of course, we do not claim that one design is “always superior” to the other. A model function decomposition-parsing-bounding based approach supports the precise solution of a range of GO models, assuming sufficient runtime and computational resources. All models, however, have to be defined in terms of a given set of possible component functions. In contrast to this approach, the design of LGO allows the handling of genuine “black box” problems that will remain outside of the scope of the analytical global optimization approaches referred to above. For optimization model examples with embedded computational procedures, consult e.g. Pintér (1996, 2009), Pintér et al. (2006), Kampas and Pintér (2006, 2009). In numerical practice (that is in hardware resource-limited and time-limited runs), LGO’s global search options generate a global solution estimate(s) that is (are) refined by the seamlessly following local search mode(s). This way, the expected result of using LGO is a global and local search based high-quality feasible solution that meets at least the local optimality criteria. (To guarantee theoretical local optimality, standard local smoothness conditions need to apply.) At the same time, one should keep in mind that no global ‒ or, in fact, any other ‒ optimization software will “always” work satisfactorily, with default settings and under resource limitations related to model size, time, model function evaluation, or to other preset usage limits. Extensive numerical tests and an increasing range of widely different practical applications demonstrate that LGO and its platform-specific implementations can find high-quality numerical solutions not only when using academic GO test problems, but also in often far more complicated, sizeable GO models. Examples of real-world optimization challenges handled by various LGO implementations are environmental systems modeling and management (Pintér, 1996), laser design (Isenor et al., 2003), intensity modulated cancer therapy planning (Tervo et al., 2003), the operation of integrated oil and gas production systems (Mason et al., 2007), vehicle component design (Goossens et al., 2007), various packing problems with industrial relevance (Castillo et al., 2008), fuel processing technology development (Pantoleontos et al., 2009), and currency trading strategies (Çağlayan and Pintér, 2010). 6. �umerical Examples and Discussion The detailed testing and comparative assessment of optimization software products is a significant research area. One needs to follow or to establish objective evaluation criteria, and to present (in principle) fully reproducible results. The software tests themselves are based on a given set of test problems and a selection of criteria such as the reliability and efficiency of the solvers used, in terms of the quality of solutions obtained and the corresponding computational effort. For examples of benchmarking studies, in the ANN context we refer to Prechelt (1994), Hamm et al. (2007); in the GO context, see e.g. Ali et al. (2005), Khompatraporn et al. (2005), Pintér (2002, 2007); with further topical references therein. Here we will illustrate the performance of global optimization software (specifically, of an LGO implementation) in the ANN calibration context, by solving merely one- dimensional, but non-trivial function approximation problems. The key features of the computational environment used and the numerical tests are summarized below: the complete set of detailed numerical results is available from the author upon request. • Hardware: PC, Intel™Core™2 Duo CPU P8400 @ 2.26GHz, 2.86 GB RAM. • Operating System: Windows XP Pro 2002 with SP 3 (32-bit). • Global optimization software: MathOptimizer Professional (MOP). Let us remark that MathOptimizer Professional is the software product name introduced for the LGO solver implementation that is linked to Mathematica (Wolfram, 2009). For details, consult e.g. Pintér and Kampas (2005), Kampas and Pintér (2006). • ANN computational models used: 1-J-1 type networks that have a single input node, a single output node, and J processing units. Recall that this setup directly follows the function approximation structure expressed by (2). The ANN models considered are implemented in native Mathematica code. • ANN initialization by assigned starting parameter values is not necessary, since MOP (i.e., LGO) uses a genuine global scope optimization approach. Let us note at the same time that all LGO implementations also support the optional use of set initial values in their constrained local optimization mode, without applying first a global search mode. • A single run is done in all cases; no repetitions are needed, since MOP produces identical runs under the same initializations and solver options. Repeated runs with possibly changing numerical results can be easily generated by setting several options of MOP. The latter are based on selecting one of the several global search operational modes, using different starting points in the local search mode, using different random seeds, and assigning different computational resources to MOP runs. • In all test problems the optimized function approximations are based on a set of training points, in line with the modeling framework of Section 2. • Simulated noise structure: none. In other words, we assume that all training data are exact. This assumption could be easily relaxed, by introducing a probabilistic noise structure as done e.g. in Pintér (2003). • Key performance measure: standard error (the root of the mean square error, RMSE) calculated for the numerical solution found. • Some additional Mathematica features used: documentation and model visualization capabilities. • Some additional MOP features used here (or elsewhere): automatic generation of optimization model code in C or Fortran; and the automatic generation of input and result text files. These files are used and generated by the LGO solver core of MOP. In the numerical examples presented here, our objective is to “learn” the graph of certain one-dimensional functions defined on the interval [0,1] based on a preset number of uniformly spaced training points. As earlier, we shall use the notation T for the number of training data, and J for the number of processing units in the hidden layer. Example 1 Our first (easiest) test problem presented here has been studied also by Toh (1999), as well as by Ye and Lin (2003). f(x)=sin(2πx) cos(4πx) 0≤x≤1. (10) Applying T=100, J=7, and MOP’s local search mode (LS), we obtain a corresponding function approximation characterized by RMSE ~ 0.00294674. With default MOP settings, the number of error function evaluations is 83,473; the corresponding runtime is 4.20 seconds, implying an estimated 20,000 function evaluations per second. (Let us remark that the speed estimate can vary a bit, based on the actual status of the computer used.) For comparison, the function (10) is displayed below (see the left hand side of Figure 2), directly followed by list plot (right hand side) produced by the output of the trained ANN at the uniformly distributed T=100 sample points. Figure 2 Function (10), and the list plot produced by the optimized ANN. Both the numerical result (the RMSE obtained) and Figure 2 indicate a good fit. For a more complete picture, we also solved the same problem using LGO’s most extensive global search mode (stochastic multi-start, MS) followed by corresponding local searches from the best points found. This operational mode is referred to as MSLS, and it has been used in all subsequent numerical examples. In the MSLS mode, we assume (postulate) the variable ranges [lb, ub] = [-20, 20] for all ANN model parameters, and set a limit for the global search steps (error function evaluations) as gss=106. Based on our GO related numerical experience, this does not seem to be an exorbitant setting since we have to optimize 21 parameters globally. (For a simple comparison, one can think of aiming at just 10% precision for each variable setting in 21 dimensions, using a uniform grid based search: this naïve strategy would require 1021 function evaluations.) As a result of using the MSLS operational mode, we receive RMSE ~ 0.00166644 which represents about 43.5% reduction of the average error found using only LS. The corresponding runtime is 59.77 seconds. It is also worth pointing out that the optimized ANN parameterizations found by LS and MSLS are quite different. Several components of these solutions lie on the boundary of the preset interval range [-20, 20] indicating a possible need for extending the search region. These findings, together with our other runs for this and other test problems indicate the potentially massive multi-modality of the objective (4) in similar or more complicated function approximation problems as well as the possible need to search on larger parameter regions. In this particular test example, we could have perhaps reduced the global search effort, or just use the local search based estimate which seems “good enough”. Such experimental “streamlining” may not always work, however, and we are fully convinced that global scope search is typically needed in calibrating ANNs. Hence, we will continue to demonstrate MOP’s standard global optimization capabilities, by solving all further test problems in MSLS mode. The examples presented below are chosen to be increasingly more difficult, although all are “only” approximation problems related to one-dimensional functions. Obviously, one could “fabricate” functions that become very difficult to match by ANNs. 0.2 0.4 0.6 0.8 1.0 -1.0 -0.5 0.5 1.0 20 40 60 80 100 -1.0 -0.5 0.5 1.0 Example 2 f(x)=sin(2πx) sin(3πx) sin(5πx) 0≤x≤1. (11) To handle this problem, we use T=100, J=10, and MOP’s MSLS search mode. Notice that the number of processing units has been increased heuristically, reflecting our belief that this function could be more difficult to reproduce than (10). This leads to a 30- variable GO model: we keep the allocated global search effort the same (gss=106) as in Example 1, however. We conjecture again that the induced GO model is highly multi-extremal. The global search based numerical solution has an RMSE ~ 0.00493399; the corresponding runtime is 76.16 seconds (1,053,014 function evaluations, i.e. ~ 13,800 evaluations per second). For comparison, the function (11) is displayed below, directly followed by the list plot produced by the trained ANN, see Figure 3. Again, the reproduction of the target function seems to be quite satisfactory, both numerically and visually. Figure 3 Function (11), and the list plot produced by the optimized ANN. Example 3 f(x)=sin(5πx) sin(7πx) sin(11πx) log(1+x) 0≤x≤1. (12) In (12), log denotes the natural logarithm function. As above, first we use T=100, J=10, and MOP’s MSLS search mode with 106 global search steps; furthermore, we extend the search region to [-50, 50] for each parameter. The global search based numerical solution has an RMSE ~ 0.0353158 which is noticeably not as good as the previous RMSEs; the corresponding runtime is 82.50 seconds. For comparison, the function (12) is displayed below, directly followed by the list plot produced by the trained ANN. In the corresponding Figure 4, one can notice a more apparent discrepancy in the “low-amplitude” section of the function, estimated by simple inspection as the interval [0, 0.4], while the approximation is quite satisfactory in the remaining “higher-amplitude” interval [0.4, 1]. 0.2 0.4 0.6 0.8 1.0 -0.6 -0.4 -0.2 0.2 0.4 0.6 20 40 60 80 100 -0.6 -0.4 -0.2 0.2 0.4 0.6 Figure 4 Function (12), and the list plot produced by a suboptimal ANN. For comparison, we also ran MOP in its (only) LS mode; the resulting standard error is RMSE ~ 0.0617162; the runtime is 28.19 seconds. Figure 5, corresponding to this case, visibly becomes a much more crude approximation of (12) than the list plot in Figure 4. Again, this finding illustrates the typical need for a global scope model fitting approach. Figure 5 Function (12), and the list plot produced by the locally optimized ANN. Example 4 In several subsequent runs, we increased the required precision of the approximation to function (12) by the ANN, by increasing one or several of the following: the number of training data T, the number of processing units J, and the solver runtime allocated to MOP. Obviously, all the above measures could lead to improved results. Without going into unnecessary further details, Figure 6 (as an example) illustrates that e.g. the combination T=200, J=20, gss=3,000,000, and the additionally set runtime limit of 900 seconds allows the approximation of the entire function rather faithfully. This finding is also indicated by the corresponding RMSE ~ 0.0111696. Let us remark that this MOP run was automatically interrupted at the set time limit: arguably, without this limit even better approximations could be found. 0.2 0.4 0.6 0.8 1.0 -0.2 -0.1 0.1 0.2 0.3 0.4 20 40 60 80 100 -0.2 -0.1 0.1 0.2 0.3 0.4 0.2 0.4 0.6 0.8 1.0 -0.2 -0.1 0.1 0.2 0.3 0.4 20 40 60 80 100 -0.2 -0.1 0.1 0.2 0.3 0.4 Figure 6 Function (12), and the list plot produced by the globally optimized ANN. To summarize our illustrative numerical experiments, we can say that the GO based ANN calibration approach can be used to handle non-trivial function approximation problems within reasonable computational effort (on today’s personal computers). Let us recall that ‒ in light of the theoretical results cited earlier ‒ exact function representations are in general too much to ask for. Our approach is directly applicable to solve at least certain types of multi-input and multi-response approximation problems. If the input sequences need to be aggregated then this will lead to increase GO model dimensionality. However, if the input data sequences can be processed by independently operated hidden layers for each function output, then our approach can be applied iteratively (repetitively) to such multi-response problems. Hence, in case of fully decoupled ANN models, the corresponding sub- problems can be solved independently, thereby avoiding the (expected) exponential increase of computational complexity. Notwithstanding the above, one should be careful to avoid overly optimistic blanket statements when handling general multiple-input multiple-output problems by ANNs. 6. Conclusions Due to its “universal approximator” capability, the ANN computational structure flexibly supports the modeling of many important research problems and real world applications that can be described by continuous functions defined over compact sets. ANN application areas includes data classification and pattern recognition (Ripley, 1996), damage detection and earthquake simulation (Pei et al., 2006), function approximation (Toh, 1999; Ye and Lin, 2003), material science (Bhadeshia, 1999), experimental design of engineering systems (Röpke et al., 2005), nonlinear optimization (Malek et al., 2010), polypeptide structure prediction (Dorn and de Souza, 2010), prediction of trading signals of stock market indices (Tilakaratne et al., 2008), regression analysis (De Veux et al., 1998), signal and image processing (Watkin, 1993; Masters, 1994), time series analysis and forecasting (Franses and van Dijk, 2000; Kajitani et al., 2005). At the same time, the ANN computational model itself has an obvious meta-heuristic flavor in the sense that it needs to be instantiated and calibrated for each new model type. Even if such problem-specific ANN design can benefit from the expertise of domain specialists, manual design can become truly difficult when problem complexity increases. 0.2 0.4 0.6 0.8 1.0 -0.2 -0.1 0.1 0.2 0.3 0.4 50 100 150 200 -0.2 -0.1 0.1 0.2 0.3 0.4 Furthermore, once the ANN structure is selected, the computational effort may still need adjustments to handle the application satisfactorily. These practical aspects require flexible and transparent ANN implementations. Our current research code set up in an interpreted computing environment (Mathematica) is an illustrative example of such an implementation. After defining the function to be approximated, one needs to change only a few input data to produce a new sampling plan, to define the ANN instance (within a postulated ANN family), and to activate the optimization engine. The corresponding results can be then directly observed and visualized (after a few seconds or minutes in the examples presented here, using an average capability PC). A focal point of the present study has been to demonstrate the essential requirement of high-quality ANN model calibration. To meet this challenge, we apply a global optimization (GO) based model fitting strategy. Our approach is based on theoretically sound GO methodology, and the corresponding software implementations are capable of handling non-trivial function approximation problems in reasonable time. The presented illustrative examples directly indicate the applicability of the GO based approach e.g. to pattern or symbol recognition, image or signal processing, and time series analysis. In accordance with the theoretically exponential complexity of GO, the numerical demand of model fitting exercises can be expected to increase considerably with the size of the induced optimization model. This specifically includes ANN fitting problems in higher dimensions, and the consideration of general (possibly complicated) constraints regarding feasible parameter settings. We plan to investigate these issues in forthcoming studies, in relation to practically motivated problems that can be modeled by ANNs. Acknowledgements The author wishes to acknowledge the valuable contributions Dr. Frank Kampas to the development of the MathOptimizer Professional software product. Thanks are due also to Dr. Maria Battarra for her comments on the manuscript. References Abraham, A. (2004) Meta learning evolutionary artificial neural networks. �eurocomputing 56, 1-38. Ali, M.M., Khompatraporn, C. and Zabinsky, Z.B. (2005) A numerical evaluation of several stochastic algorithms on selected continuous global optimization test problems. Journal of Global Optimization 31, 635–672. Auer, P. Herbster, M. Warmuth, M. (1996) Exponentially many local minima for single neurons. In: Touretzky, D. et al., (Eds.), Advances in �eural Information Processing Systems, Vol. 8, pp. 316–322. MIT Press, Cambridge, MA. Bar-Yam, Y. (1999). Dynamics of Complex Systems. Westview Press, Boulder, CO. Bhadeshia, H.K.D.H. (1999) Neural networks in materials science, ISIJ International, 39 (10), 966-979. Bianchini, M. and Gori, M. (1996). Optimal learning in artificial neural networks: A review of theoretical results. �eurocomputing 13, 313-346. Çağlayan, M.O. and Pintér, J.D. (2010) Development and calibration of currency market strategies by global optimization. Research Report, Özyeğin University, Istanbul. (Submitted for publication) Castillo, I., Kampas, F.J., and Pintér, J.D. (2008) Solving circle packing problems by global optimization: numerical results and industrial applications. European Journal of Operational Research 191, 786– 802. Cichocki, A. and Unbehauen, R. (1993) �eural �etworks for Optimization and Signal Processing. John Wiley & Sons, Chichester, UK. Cruse, H. (2006) �eural �etworks as Cybernetic Systems (2nd and Revised Edition). Brains, Minds and Media, Bielefeld, Germany. Cybenko, G. (1989) Approximation by superpositions of a sigmoidal function. Mathematics of Control, Signals, and Systems 2, 303-314. De Veux, R.D., Schumi, J., Schweinsberg, J., and Ungar, L.H. (1998) Prediction intervals for neural networks via nonlinear regression. Technometrics 40, 273–282. Dorn, M., and de Souza, O. N. (2010) A3N: An artificial neural network N-gram-based method to approximate 3-D polypeptides structure prediction. Expert Systems with Applications (forthcoming), doi:10.1016/j.eswa.2010.04.096. Franses, P.H., van Dijk, D. (2000) �onlinear Time Series Models in Empirical Finance. Cambridge University Press, Cambridge, UK. Funahashi, K. (1989) On the approximate realization of continuous mappings by neural networks. �eural �etworks 2, 183-192. Golden, R.M. (1996) Mathematical Methods for �eural �etwork Analysis and Design, MIT Press, Cambridge, MA. Goossens, P., McPhee, J., Schmitke, C., Pintér, J.D., and Stahl, H. (2007) Driving innovation: how mathematical modeling and optimization increase efficiency and productivity in vehicle design. Technical Memorandum, Maplesoft, Waterloo, ON. Hamm, L., Brorsen, B.W. and Hagan, M. T. (2007) Comparison of stochastic global optimization methods to estimate neural network weights. �eural Processing Letters 26, 145–158. Hertz, J., Krogh, A., and Palmer, R.G. (1991) Introduction to the Theory of �eural Computation. Addison- Wesley, Reading, MA. Hornik, K., Stinchcombe, M., and White, H. (1989) Multilayer feedforward networks are universal approximators. �eural �etworks 2, 359–366. Horst, R. and Pardalos, P.M., Eds. (1995) Handbook of Global Optimization, Volume 1. Kluwer Academic Publishers, Dordrecht. Isenor, G., Pintér, J.D. and Cada, M. (2003) A global optimization approach to laser design. Optimization and Engineering 4, 177-196. Jordanov, I. and Brown, R. (1999) Neural network learning using low-discrepancy sequence. In Foo, N., Ed., AI’99, Lecture �otes in Artificial Intelligence 1747, pp. 255-267. Springer-Verlag, Berlin Heidelberg. Kajitani, Y., Hipel, K.W., and McLeod, A.I. (2005) Forecasting nonlinear time series with feed-forward neural networks: a case study of Canadian lynx data. Journal of Forecasting 24, 105–117. Kampas, F.J. and Pintér, J.D. (2006) Configuration analysis and design by using optimization tools in Mathematica. The Mathematica Journal 10 (1), 128-154. Kampas, F.J. and Pintér, J.D. (2009) MathOptimizer: A nonlinear optimization package for Mathematica users. Research Report, Özyeğin University, Istanbul. (Submitted for publication) Khompatraporn, Ch., Pintér, J.D. and Zabinsky, Z.B. (2005) Comparative assessment of algorithms and software for global optimization. Journal of Global Optimization, 31, 613–633. Kolmogorov, A. N. (1957). On the representation of continuous functions of many variables by superposition of continuous functions of one variable and addition. Doklady Akademii �auk SSR, 114, 953-956. Lang, B. (2005) Monotonic multi-layer perceptron networks as universal approximators. In: Duch, W. et al., Eds., Artificial �eural �etworks: Formal Models and Their Applications - ICA�� 2005, pp. 31-37. Lecture �otes in Computer Science 3697, Springer-Verlag Berlin Heidelberg. Malek, A., Hosseinipour-Mahania, N., and Ezazipoura, S. (2010) Efficient recurrent neural network model for the solution of general nonlinear optimization problems. Optimization Methods & Software 25 (4), 489–506. Mason, T.L., Emelle, C., van Berkel, J., Bagirov A.M., Kampas, F. and Pintér, J.D. (2007) Integrated production system optimization using the Lipschitz Global Optimizer and the Discrete Gradient Method. Journal of Industrial and Management Optimization 3 (2), 257-277. Masters, T. (1994) Signal and Image Processing with �eural �etworks, John Wiley & Sons, New York. Mehrotra, K., Mohan, C.K., and Ranka, S. (1997) Elements of Artificial �eural �ets. MIT Press, Cambridge, MA. Pantoleontos, G., Basinas, P., Skodras, G., Grammelis, P., Pintér, J.D., Topis, S., and Sakellaropoulos, G.P. (2009) A global optimization study on the devolatilisation kinetics of coal, biomass and waste fuels. Fuel Processing Technology 90, 762–769. Pardalos, P.M. and Romeijn, H.E., Eds. (2002) Handbook of Global Optimization, Volume 2. Kluwer Academic Publishers, Dordrecht. Park, J. and Sandberg, I.W. (1991) Universal approximation using radial-basis-function networks. �eural Computation 3 (2), 246-257. Park, J. and Sandberg, I.W. (1993) Approximation and radial-basis-function networks. �eural Computation 5 (2), 305-316. Pei, J-S., Mai, E., and Piyawat, K. (2006) Multilayer feedforward neural network initialization methodology for modeling nonlinear restoring forces and beyond. 4WCSCM-306, 4th World Conference on Structural Control and Monitoring, San Diego, CA, July 2006. Pintér, J.D. (1996) Global Optimization in Action. Kluwer Academic Publishers, Dordrecht. Pintér, J.D. (1997) LGO ‒ A program system for continuous and Lipschitz optimization. In: Bomze, I.M., Csendes, T., Horst, R. and Pardalos, P.M., Eds. Developments in Global Optimization, pp. 183-197. Kluwer Academic Publishers, Dordrecht. Pintér, J.D. (2002) Global optimization: software, test problems, and applications. In: Pardalos, P.M. and Romeijn, H.E., Eds. Handbook of Global Optimization, Volume 2, pp. 515-569. Kluwer Academic Publishers, Dordrecht. Pintér, J.D. (2003) Globally optimized calibration of nonlinear models: techniques, software, and applications. Optimization Methods and Software 18 (3) 335-355. Pintér, J.D. (2005) Nonlinear optimization in modeling environments: software implementations for compilers, spreadsheets, modeling languages, and integrated computing systems. In: Jeyakumar, V. and Rubinov, A.M., Eds., Continuous Optimization: Current Trends and Modern Applications, pp. 147-173. Springer Science + Business Media, New York. Pintér, J.D. (2007) Nonlinear optimization with GAMS/LGO. Journal of Global Optimization 38, 79-101. Pintér, J.D. (2009) Software development for global optimization. In: Pardalos, P.M. and T. F. Coleman, eds. Global Optimization: Methods and Applications, pp. 183-204. Fields Institute Communications Volume 55. Published by the American Mathematical Society, Providence, RI. Pintér, J.D. (2010a) LGO ─ A Model Development and Solver System for �onlinear (Global and Local) Optimization. User’s Guide. (Current version) Distributed by Pintér Consulting Services, Inc., Canada; www.pinterconsulting.com. Pintér, J.D. (2010b) Spreadsheet-Based Modeling and �onlinear Optimization with Excel-LGO. User’s Guide. With technical contributions by B.C. Şal and F.J. Kampas. (Current version) Distributed by Pintér Consulting Services, Inc., Canada. www.pinterconsulting.com. Pintér, J.D. and Kampas, F.J. (2005) Nonlinear optimization in Mathematica with MathOptimizer Professional. Mathematica in Education and Research 10 (2), 1-18. Pintér, J.D., Linder, D., and Chin, P. (2006) Global Optimization Toolbox for Maple: an introduction with illustrative applications. Optimization Methods and Software 21 (4), 565-582. Prechelt, L. (1994) PROBEN1 − A set of benchmarks and benchmarking rules for neural network training algorithms. Technical Report 21/94, Fakultat fur Informatik, Universitat Karlsruhe, Germany, September 1994. Ripley, B.D. (1996) Pattern Recognition and �eural �etworks. Cambridge University Press, Cambridge, MA. Röpke, K. et al., Eds. (2005) DoE – Design of Experiments. Produced with the technical collaboration of IAV GmbH. Verlag Moderne Industrie. © sv corporate media GmbH, Munich, Germany. Sexton, R.S., Alidaee, B., Dorsey, R.E. and Johnson, J.D. (1998) Global optimization for artificial neural networks: A tabu search application. European Journal of Operational Research 106 (1998) 570-584. Smith, M. (1993) �eural �etworks for Statistical Modeling, Van Nostrand Reinhold, New York. Steeb, W-H. (2005) The �onlinear Workbook. (3rd Edition) World Scientific, Singapore. Sutskever, I. and Hinton, G.E. (2008) Deep, narrow sigmoid belief networks are universal approximators. �eural Computation 20, 2629–2636. Tervo, J., Kolmonen, P., Lyyra-Laitinen, T., Pintér, J.D., and Lahtinen, T. (2003) An optimization-based approach to the multiple static delivery technique in radiation therapy. Annals of Operations Research 119, 205-227. Tilakaratne, Ch.D., Mammadov, M.A. and Morris, S.A. (2008) Predicting trading signals of stock market indices using neural networks. In: Wobcke, W. and Zhang, M., Eds., AI 2008: Advances in Artificial Intelligence, pp. 522-531. Lecture �otes in Computer Science 5360, Springer-Verlag Berlin Heidelberg, 2008. Toh, K-A. (1999) Fast deterministic global optimization for FNN training. In: Proceedings of IEEE International Conference on Systems, Man, and Cybernetics, Tokyo, Japan, pp. V413 - V418. Watkin, T. L. H. (1993) Optimal learning with a neural network. Europhysics Letters 21 (8), 871-876. Wikipedia (2010) Artificial neural network. http://en.wikipedia.org/wiki/Artificial_neural_network. Wolfram Research (2009) Mathematica (Version 7.0.1). Distributed by WRI, Champaign, IL. Ye, H. and Lin, Z. (2003) Global optimization of neural network weights using subenergy tunneling function and ripple search. Proc. IEEE ISCAS, Bangkok, Thailand, May 2003, V725-728. 
work_62sk5oo6zvc2faonatnepqu43u	----	A Yugoslav Journal of Operations Research 16 (2006), Number 1, 107-124 CONCEPT OF EXPERT SYSTEM FOR MODAL SPLIT IN TRANSPORTATION PLANNING Maja M. POPOVIĆ, Jadranka J. JOVIĆ Faculty of Transport and Traffic Engineering University of Belgrade, Belgrade Serbia and Montenegro Received: October 2004 / Accepted: March 2005 Abstract: The objective of this paper is to develop a concept of expert system based on the survey of experts' opinions and their experience concerning relations in modal split, on the basis of parameters of transport system demand and transport supply, defined through PT travel time and city size, i.e. mean trip length. This expert system could be of use both to experts and less experienced planners who could apply the knowledge contained in this expert system for further improvement, on operational as well as on strategic level. Keywords: Modal split, transport supply, transport demand, expert system. INTRODUCTION The expert system, created in this paper, has been envisaged as a tool to be used on operational and strategic level in some segments of transport planning. The basic idea for creating such an expert system was to define relations in modal split on the basis of parameters of transport system demand and supply in one city. To enable the use of this expert system on both strategic (planning) and operational level, it is necessary to determine parameters which could be forecasted on short-term and long-term basis. For planning purposes, such an expert system should determine margins of share of specific transport modes. These limits should serve as benchmarks in forecast of the relations in modal split during planning period, on the basis of which, specific elements of future transport system are to be defined. In a similar manner this expert system can be used for operational purposes, providing that survey of parameter values is performed in the existing state or that a short-term forecast for these parameters is made. M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 108 As it is known [1] the basic parts of an expert system are the knowledge base and a program which will make application of the knowledge base possible. In this case the knowledge base has been created on the basis of an experts' opinions survey and on the basis of past transport studies in the subject field. The process of a knowledge base creation can be performed in the manner presented in Figure 1. Figure 1: Block diagram of the process of knowledge base creation - EXSYMS The first phase of knowledge base creation is represented by an analysis of the existing transport studies. Characteristics of trip and modal split analyzed for Former Yugoslavia cities can be presented in the scope of this phase. The second phase comprises selection of relevant parameters of transport system demand and supply. As a supply parameter, i.e. parameter of transport system quality, in the first phase, it is recommended to select public transport travel time (tt), and the mean trip length (Lmean), as a parameter for transport system users' demand. The third phase represents definition of scenarios for modal split modelling. These scenarios are being formulated by a combination of four values of mean trip length M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 109 (Lmean and three values of the time spent outside vehicle (t1). The time spent outside vehicle (t1) is a sum of the transfer time (ttr), approach time (tacc), waiting time (tw) and terminating time (tter). In this phase travel time (tt) is replaced by the time spent outside vehicle (t1). In this way twelve scenarios were formulated. Based on these 12 scenarios the relations in the modal split were studied for three modes: PC-passenger car, PT- public transport and OM-other modes (walking, bicycle, motorcycle, etc.) In the fourth phase a survey of experts' opinions on modal split relations was carried out. In the scope of this phase a study methodology was prepared and survey of experts was performed. The fifth phase includes the first part of the analysis of the obtained results, i.e., elaboration of the survey of experts' opinions applying prepared mathematical tests. This results in interval evaluation of shares of defined modes, sample dispersion and mean value of shares of defined modes. In addition, a processing of the obtained data was performed. Correction of the results obtained during survey of experts' opinions was carried out in the sixth phase. These corrections were made in accordance with perceived discrepancies and inconsistencies from the fifth phase. After correcting these results, in the seventh phase, the EXSYMS Essential Base (Expert System for Modal Split) was created. The Essential Base was provided when, applying corrected survey results, regressive equations of dependence of specific transport mode share upon the time spent outside vehicle (t1), as well as upon the mean trip length (Lmean) were formulated. In the next phase, the Essential Base was enlarged by correlating PT travel time (tt) with the time spent outside vehicle (t1). In this way, by applying regression analysis, dependence equations of specific transport mode share upon PT travel time (tt) were provided. In the last phase of creating the EXSYMS Base formed were range dependence diagrams between share of PC, PT and OM, on the one hand and PT travel time, on the other for various mean trip lengths (Lmean). After that, PT travel time (tt) was correlated to parameters of PT operation (line network density in PT (σ) and, thus, the EXSYMS Base was created. The second part of the expert system, the programme for knowledge application (inference engine), should be selected from a group of those programs which are modelling a safe reasoning (in the function of the type of the subject problem, in each phase, it selects a single rule from the set of valid rules), because the nature of the subject problem is such that the queries raised in the expert system require a range solution. Deduction and regression are to be applied as modes of reasoning. In some specific cases this programme should enable providing responses to the following queries: Assuming that transport supply parameter A has value "a" and transport demand parameter B has value "b", what is proportional share of certain transport modes (PC,:PT, OM); and Assuming that percentages of transport mode share are PC = n%, PT=m% and OM=r%, what are the values of transport supply A parameter and transport demand B parameter? This represents a description of the general concept of expert system for modal split. For purposes of this paper, the phase representing creation of knowledge base of the expert system was further elaborated. M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 110 2. CREATION OF THE KNOWLEDGE BASE FOR THE EXSYMS - EXPERT SYSTEM FOR MODAL SPLIT The knowledge base for the concept of this expert system was created on the basis of a study of previous transport surveys in Former Yugoslavia cities, as well as on the basis of survey of experts' opinions with respect to relations in movement modal split in cities. 2.1. Selection of entry parameters for the expert system and definition of scenarios Implementation of this expert system implies formulation of certain scenarios on the basis of the selected (studied) parameters of transport system demand and supply, which would enable definition of relations in movement modal split. This means that for this expert system a modal split method has to be selected in the first place. After that, it is necessary, among a large set of influential parameters, to select parameters most adequately represent the current situation and prepare scenarios, which should represent a combination of parameters of transport system supply and demand by transport system users. For modal split modelling a normative method was selected due to the fact that past experience in Former Yugoslavia cities indicates that this mode was the most frequent mode of modal split modelling in Former Yugoslavia region. Another reason for such a choice was that this mode, by its nature, is dependent upon planners' (experts') assessment. On the other hand, expert system creation implies formalization of the models to be implemented into expert system. Consequently, by selection of normative modelling this process is considerably simplified. Selected parameters for creation of the knowledge base of expert system and for defining scenarios of survey of experts' opinions on relations in modal split are: PT travel time (tt) and city size expressed through mean trip length (Lmean). The idea behind formulation of such scenarios was, from a number of selected values during ride in PT and from the mean trip length, to make combinations (value couples) for which relations in modal split are to be studied. Such scenarios are defined for four values of mean trip length Lmean (km) and three values of travel time tt (min). Based on the review of the transport surveys carried out in Former Yugoslavia cities so far, as well as on the basis of recommendations of the experts in the field of Transportation planning and traffic control, PT, etc., city size parameter, i.e., mean trip length was defined as follows (Table 1): Table 1: Dependence between mean trip length and city size City size (number of citizens) Mean trip length Lmean (km) Remarks 50,000 to 100,000 2 A city with undeveloped PT (or without PT) with commuter transit having function of PT (Šabac, Užice, Valjevo, Čačak, etc.) 100,000 to 200,000 3 A city with mainly one PT subsystem and commuter transit (Subotica, Podgorica, Kragujevac, etc.) 200,000 to 500,000 4 A city with developed PT (one or two subsystems of PT) and commuter transit (Novi Sad, Niš) over 500,000 8 A big city having more PT subsystems (bus, tramway, trolleybus, etc.), as well as developed commuter transit (Belgrade) M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 111 Travel time is a transport system quality parameter. In the process of travelling a series of difficulties are piling up for a user. The travelling benefits are realized only upon satisfying trip purpose at the end of a trip. Measurement of unattractiveness or attractiveness toward a specific form of transport in the scope of one transport system is made possible by comparing all inconveniences and benefits from the realized purpose. The time spent in transportation for the users of transport system is time lost, i.e. "cost", not only in terms of transport fare, but also in reference to "generalized costs" arising from a series of, for a user, undesirable circumstances in the course of operation of transport. PT travel time (tt), as one of the characteristics of a journey, is a complex indicator. It includes the following: • Time of access to a system (tacc) (time needed from the "door" to station/stop), • Waiting time of (tw), • Riding time (time spent in a vehicle) (tr), • Transfer time (ttr), and • Terminating time (tter) (time needed from the last station/stop to the "door"). PT travel time comprises some of the essential transport supply parameters, such as: • PT network density (as a supply indicator) – denser network ⇒ bigger accessibility ⇒ shorter access time ⇒ shorter riding time, • Number of vehicles on one line is included into time intervals, upon which waiting time is depending. Therefore, both number of vehicles and time interval are presented through waiting time – more vehicles ⇒ shorter time intervals ⇒ shorter waiting time. For this reason, in this expert system concept, transport supply was defined in terms of the PT travel time tt. PT travel time tt was specified as dependence upon the riding time tr and upon the time spent outside vehicle t1, i.e., the sum of other components of time travel (access time (tacc) + waiting time (tw) + transfer time (ttr)+ terminating time (tter). The time spent outside vehicle t1 is provided in three options: • t1= 10 minutes - "pink option" • t1= 20 minutes - "grey option" • t1= 30 minutes - "black option" Time spent in a vehicle (tr) is given for each city size (mean trip length). Value of the time spent in a vehicle is calculated for PT operating speed v=15 km/h. This speed value is assumed on the basis of average PT operating speed [2]. o Time spent in a vehicle for a city with mean trip length Lmean= 7 km is tt= 28 min o Time spent in a vehicle for a city with mean trip length Lmean = 4 km is tt=16 min o Time spent in a vehicle for a city with mean trip length Lmean = 3 km is tt =12 min o Time spent in a vehicle for a city with mean trip length Lmean = 2 km is tt =8 min By combining these values, twelve scenarios for determining relations in modal split for three modes of transport – PC, PT and OM - were defined. Since the concept of M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 112 this expert system is based on the survey of experts' opinions, the author of this paper believed that scenarios should be defined in a way which would allow provision of the most precise responses possible from the experts and which would make experts' work as easier as possible. Consequently, PT travel time tt, in the survey of experts was modified by the time spent outside vehicle t1 (access time + waiting time + transfer time + terminating time). After processing the results of this survey, travel time tt parameter was added to the expert system concept. These twelve scenarios, three for each four categories of city size, i.e. mean trip length, (for t1=10, 20, and 30 minutes) are presented in Table 2. Table 2: Scenarios for modal split modelling Scenarios Mean trip length (km) / Time outside vehicle (min) Lmean / t1 S1 8 / 10 S2 8 / 20 S3 8 / 30 S4 4 / 10 S5 4 / 20 S6 4 / 30 S7 3 / 10 S8 3 / 20 S9 3 / 30 S10 2 / 10 S11 2 / 20 S12 2 / 30 Above mentioned scenarios served as a basis for the survey of experts' opinions on the subject of relations in modal split. 2.2. Survey of experts' opinions on relations in modal split The objective of this survey is to provide an expert assessment of the impact of share of transport mode (modal split) to the function of transport supply quality expressed as per time spent outside vehicle (t1), as a "negative" component of the travel time. This survey was carried out for four different values of the mean trip length, i.e. city size, (as a transport demand parameter) and for three values of time spent outside vehicle t1 (as a parameter of quality of PT supply), i.e. for 12 scenarios presented in Table 2. In defining methodology of this study it was necessary to observe the following principles: • objective selection of the expert group • provision of independent judgment of experts, and • unique formulation of questions and answers provided for each and all expert in the course of the survey This methodology specifies: • size of sample, • interview form, M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 113 • survey method, and • mode of data processing. Selection of experts was carried out according to the following criteria: • theoretical knowledge in the field of modal split in transportation planning • practical knowledge in the subject field • experience in planning and organization of transport system. The number of experts who participated in the survey was partially reduced due to such criteria, but objectively, conditions for forming a bigger expert group were not available. For this interview selected were 15 experts from three cities Belgrade, Niš and Subotica who, in the course of their work, had made significant contribution in the field of modal split. Having in mind the delicacy of the issue under consideration in this paper, as well as the fact that it was needed on more occasions to consult the experts and to provide independence of experts' judgement, a modified method of mutual expert evaluation was selected. The experts were requested, by employing their expertise and experience in the field of modal split modelling, to define proportional ratio in movement modal split, for three movement modes – PC, PT and OM – for the defined scenarios. The results of this survey of experts were systemized by scenarios and they were elaborated by statistical tests for "variable mean value". For the needs of an easier data processing an EXCEL program was designed. A "variable mean value" test was applied to each scenario, i.e. to each defined transport mode in the scope of the twelve scenarios, with the following parameters: • sample of volume n =15, • risk α = 0.15 (selected risk assessment is in accord with tolerable variations in planning procedures, i.e. reliability interval of 85%). Output results are: variable mean value, reliability interval of mean value and sample dispersion. 2.3. Survey results The results obtained after data processing are presented in Table 3. Data from this Table 3 represent relations in modal split for three movement modes: PC, PT and OM. They are expressed as per mean value and reliability intervals. It should be mentioned that, in the course of this survey, no weight coefficients were determined, i.e. their scoring was not performed in the first phase of the analysis of results. At a later stage, correction of responses in relation to inconsistency with sample's mean value in responses of some expert groups was made. After processing and analysing these results it was concluded that, by defining interval of confidence 15% around mean values obtained in the first phase of survey processing, reliability interval of 85% was obtained (what is acceptable for purposes of transportation planning), and, to some extent, previously mentioned inconsistencies were corrected. However, due to large sample dispersion, this interval forms functionality intervals which would be mutually overlapped, what could make application of the survey results difficult. Therefore, interval of 15% was reduced to interval of 5%. Correction of the results provided by this interval of ±5% will be made after regression analysis of the results. M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 114 Table 3: Corrected values: proportional share of PC, PT and OM depending on time spent outside vehicle (t1) and mean trip length (Lmean), expressed by mean value Mean trip length Lz 7 km 4 km 3 km 2 km 10 min PC- 24.3.% PT -60.4% OM -15.3% PC-24.2 PT -49.9 OM -25.9 PC- 24.1 PT -43.2 OM -31.7 PC- 24.0 PT -29.7 OM -46.3 20 min PC- 27.9 PT -54.4 OM -17.9 PC- 27.7 PT -44.4 OM -27.9 PC- 27.5 PT -37.7 OM -34.8 PC- 24.3 PT -28.8 OM -48.9 Time Out- Side vehicle(t1) 30 min PC- 31.1 PT -50.0 OM -18.8 PC- 30.2 PT -39.5 OM -30.3 PC-28.7 PT -31.1 OM -40.2 PC- 27.4 PT -21.1 OM -51.5 On the basis of data from Table 3, for each city size, regression dependence equations between share of specific transport modes and the time spent outside vehicle (t1) were calculated. These regression lines are presented in the following diagrams. 50 40 30 20 10 y= 29,8+ 0,17 x y= 21,6+ 0,23 x y= 20,86+ 0,32 x Time spent outside vehicle (t )1 10 20 30 40 50 60 70 80 y= 20,8+ 0,34 x Lmean= 2km Lmean= 3km Lmean= 4km Lmean= 7km Pa ss en ge r ca r sh ar e in to ta l v ol um e of tr ip (% ) Figure 2: PC share in total volume of trip depending on time spent outside vehicle (t1) M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 115 Time spent outside vehicle (t )1 10 20 30 40 50 60 70 80 Lmean= 2km Lmean= 3km Lmean= 4km Lmean= 7km Figure 3: PT share in total volume of trip depending on time spent outside vehicle (t1) O th er m od es s ha re in to ta l v ol um e of tr ip ( % ) 50 40 30 20 10 y= 13,76+ 0,18 x y= 24,4+ 0,18 x y= 27,33+ 0,43 x Time spent outside vehicle (t )1 10 20 30 40 50 60 70 80 y= 45,3+ 0,21 x Lmean= 2km Lmean= 3km Lmean= 4km Lmean= 7km 60 Figure 4: OM share in total volume of trip depending on time spent outside vehicle (t1) On the basis of data from Table 4, regression dependence equations of share between specific movement modes and city size, i.e. mean trip length, were calculated. These regression lines are presented in the following diagrams. M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 116 Mean trip length Lmean Pa ss en ge r c ar s ha re in to ta l v ol um e of tr ip (% ) 1 2 3 4 5 6 7 8 50 40 30 20 10 t = 30 min1 t = 20 min1 t = 10 min1 y= 23,9+ 0,06 x y= 23,8+ 0,57 x y= 26,6+ 0,7 x Figure 5: Share of PC in total volume of trip depending on mean trip length Lmean Mean trip length Lmean Sh ar e of p ub lic tr an sp or t i n to ta l v ol um e of tr ip ( % ) 1 2 3 4 5 6 7 8 70 60 50 40 30 20 10 t = 10 min1 t = 20 min1 t = 30 min1 y= 13,6+ 5,48 x y= 20,05+ 5,11 x y= 22,5+ 5,75 x Figure 6: Share of PT in total volume of trip depending on mean trip length Lmean M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 117 Sh ar e of o th er m od es in to ta l v ol um e of tr ip ( % ) 50 40 y= 52,9-5,74 x y= 56,4-5,83 x y= 59,7-6,17 x Mean trip length Lmean 1 2 3 4 5 6 7 8 30 20 10 t = 30 min1t = 20 min1 t = 10 min1 Figure 7: Share of OM in total volume of trip depending on mean trip length Lmean By applying these diagrams it is possible to determine relations in modal split in the following manner. For the selected value of the time spent outside vehicle in PT t1, and city size presented by mean trip length (Lmean), PC share in total movement volume is read from the diagram in Figure 5. After that, from diagrams presented in Figure 6 and Figure 7, percentage of share of PT and OM for identical values of the time spent outside vehicle in PT t1, and mean trip length (Lmean) is read. The amount of proportional values should be close to 100%. A procedure defined in such a way represents only a sub-phase in the process of creating knowledge base for an expert system. 2.4. Correlation between survey results and PT travel time The next step in creating this expert system base is represented by correlating obtained results based on the time spent outside vehicle and results based on travel time. This was performed as follows: for each city size, i.e. mean trip length (2, 3, 4 and 7 km) calculation of PT travel time for all three values of time spent outside vehicle was made, as well as for the value of PT riding time with operating speed of 15 km/h. • tt 2,10 = 18 min. – travel time (min) for a city with Lmean= 2km, tter= 10 min and tr = 15 km/h • tt 2,20 = 28 min. – travel time (min) for a city with Lmean = 2km, tter = 20 min and tr =15 km/h • tt2,30= 38 min. - travel time (min) for a city with Lmean = 2km, tter = 30 min and tr=15 km/h M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 118 • tt 3,10= 22 min. - travel time (min) for a city with Lmean = 3km, tter = 10 min and tr=15 km/h • tt 3,20= 32 min. - travel time (min) for a city with Lmean = 3km, tter = 20 min and tr=15 km/h • tt3,30= 42 min. - travel time (min) for a city with Lmean = 3km, tter = 30 min and tr=15 km/h • tt 4,10= 26 min. - travel time (min) for a city with Lmean = 4km, tter = 10 min and tr=15 km/h • tt 4.20= 36 min. - travel time (min) for a city with Lmean = 4km, tter = 20 min and tr=15 km/h • tt4,30= 46 min. - travel time (min) for a city with Lmean = 4km, tter = 30 min and tr=15 km/h • tt 7,10= 38 min. - travel time (min) for a city with Lmean = 7km, tter = 10 min and tr=15 km/h • tt 7,20= 48 min. - travel time (min) for a city with Lmean = 7km, tter = 20 min and tr=15 km/h • tt 7,30= 58 min. - travel time (min) for a city with Lmean = 7km, tter = 30 min and tr=15 km/h For these PT travel times, identical relations in modal split apply as in the case of the time spent outside vehicle in PT (t1) 10, 20 and 30 min, for each city size. For example, for a city with mean trip length of 4km, travel time of 26 minutes provides identical relations in modal split as time spent outside vehicle of 10 minutes, i.e. travel time of 36 and 46 minutes provides identical relations in modal split as the time spent outside vehicle of 20 and 30 minutes, respectively. Public transport travel time (t ) in interval 5%t Pa s s en ge r c ar s h a re in to ta l v ol um e of m o v em en t (% ) Lmean Lmean Lmean Lmean Figure 8: PC share in total volume of movement depending on PT travel time (tt) in interval ± 5% Required corrections are to be done in terms of forming intervals around the values read from diagrams or calculated from regression equations. It has been already stated that these corrections can be done by applying values in the interval (-5%, +5%) M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 119 around mean value, because, in such way corrections of variations occurring during analysis of expert group results (performed expert group by expert group) as well as during processing of the results itself. In the following diagrams presented are interval solutions for determination of certain transport mode shares, depending on travel time. Pu bl ic tr an sp or t s ha re in to ta l v ol um e of m ov em en t (% ) Public transport travel time (t ) in interval 5%t Lmean Lmean Lmean Lmean Figure 9: PT share in total volume of movement depending on PT travel time (t in t interval ± 5% ) O t h er m od es s ha re in to ta l v ol um e o f m ov em en t (% ) public transport travel time (t ) in interval 5%t Lmean Lmean Lmean Lmean Figure 10: OM share in total volume of movement depending on PT travel time (tt) in interval ± 5% M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 120 The step forward in outlining the EXSYMS was correlation between travel time and specific transport system parameters. Each travel time components depends on a series of variables, such as: travel speed, time interval, network density, number of vehicles on a line, etc. Because it is difficult to work simultaneously with greater number of parameters, particularly in transport demand forecast, we chose one parameter – network density σ (km/km2), which represents length of line network per square kilometre of area surface. Research conducted by Zilbertal [2] shows how travel time can be expressed in terms of network density. According to Zilbertal [2], travel time (tt) is: 120 1 60 3 4 n m t t r e l L t Nv tv P σ σ ⎛ ⎞ ⎜ ⎟⎛ ⎞ = + + +⎜ ⎟⎜ ⎟ ⎝ ⎠ ⎜ ⎟⎜ ⎟ ⎝ ⎠ ean The expression N/P represents density of transport units per kilometre of a city area. Values of this parameter for certain city categories are given in Table 4. Table 4: Density of transport units per city area kilometre City size (in thous.of citizens) Mean trip length Lmean (km) N/P (in trip /km2) 50-100 2.0-2.2 1.5-1.8 100-250 2.2-2.5 2.0-2.2 250-500 2.5-3.0 2.4-3.2 500-1000 3.0-4.0 3.8-4.5 Source [2] By analysing values from Table 4, as well as by analysing values of city size and mean trip length (Lmean), which are subject of this paper, values of indicators N/P from Table 4, were approximated and modified for mean trip lengths (Lmean) of 2, 3, 4 and 7 km. Table 5: Density of transport units per city area kilometre (N/P) City size (in thous.of citizens) Mean trip length Lmean (km) N/P (in trip/km2) 50-100 2 1.85 100-200 3 2.6 200-500 4 3.8 >500 7 5.0 Assuming that the following applies to all city categories: • interstation distance 400m, • w d 4 km/h, • km/h, travel time depending on network density calculated applying lowing formula [2] alking spee operating speed 15 can be the fol : M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 121 10σ 5 N σ = + 3.9 3t meant L+ + (min.) (2) defi P For ned values from Table 5, travel time in correlation with network density, is presented in Figure 11. Pu bl ic tr an sp or t t ra ve l t im e t (m in ) 1 Network density (km/km)σ Lmean=7 km Lmean=4 km Lmean=3 km Lmean=2 km Figure 11: Dependence of PT travel time upon network density ted to network density, what increases potential application of the proposed concept. 3. APPLICATION OF EXSYMS Two potential groups of users of hereby created "tool" could be recognized. In the scope of strategic operation of transport planners, such an expert system provides response to various scenarios of urban system development and its transport subsystem. Operational application implies an insight into effects of certain, actual, management and organizational activities in the scope of the transport system. The EXSYMS represents a "tool" to be used in determining a framework in which modal split modelling1 is feasible. Defined relations in modal split can be used as benchmarks in determining share of specific transport modes (their definition being done on the basis of two parameters). They should serve as margins within which planners In this manner PT travel time, which is an input value for application of this expert system, is correla 1 Utilisation implies existence of alternative “PT” in the transport system of the region under consideration, either developed PT or a commuter transit which has the role of PT in smaller towns. M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 122 should perform the final defining of these relations. Definition of final relations in movement modal split should be based on a comprehensive and detailed transport study which should illustrate particularities of a given region, in the sense of topographic characteristics, socio-economic parameters, movement characteristics, transport system, etc., as well as on setting up the trends of future modifications of such parameters. The following algorithm presents a method of applying this expert system on planning level. Mean travel time Public transport travel time Network density Ekspert system EXSYMS Other factors: socio-economical, public transport functioning parameters ... Share of personal, public other means of transport in total volume of trip (interval +/- 5%) Corected modal split Number of residents in the city (PV, PT, other) Figure 12: Method of applying EXSYMS on planning (strategic) level In order to apply the EXSYMS on an operational level it is necessary to be acquainted with the current situation in transport system, what implies existence of a solid information base formed on research results. Having this view in mind, it is possible to apply such an expert system in "gaming" domain, i.e. testing a number of possible actions to be undertaken, and not for defining one specific action. This means that this expert system, for actual relations in modal split, can offer values only for those parameters which it contains, and which are defined on the basis of a number of assumptions, that may, more or less, vary for an actual case. M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 123 Mean travel timePlaned modal split Ekspert system EXSYMS Network density σ Headway PT capacity Tarife PT travel time split Exsisting modal Needs, planing Costs / Benefit points, development building-in into the expert stem. This means that it is necessary to, first of all, in more detail and on more mples, estimate parameter values affecting movement modal split as well as their dependencies. After that, it is possible to define such dependencies in terms of mathematical models which should be calibrated in accordance with specific characteristics of each domain, and built in into the expert system's knowledge base. Figure 13: Method of applying EXSYMS on operational level 4. CONCLUSION Considering the fact that expert systems involve use of computers, i.e. application of specific programs and "shells" to build-in an expert system, the procedure presented in this paper should be formalized and modified for sy sa M.M. Popović, J.J. Jović / Concept of Expert System for Modal Split 124 REFERENCES [1] Adeli H., and Balasubramagnyam, K.V., Expert Systems for Structural Design - A New Generation, Prentice Hall, Englewood Clifts, New Jersey, 1988. [2] Samoilov, D.S., Gorodskoj transport, Moskva Stroiizdat, Moskva, SSSR, 1983. [3] Benjamin, "Utilization of attitudinal measurement techniques to analyse demand for transportation: Methods, applications and new directions", Behavioral Research for Transport Policy, Utrecht, The Netherlands, 1986. [4] "Belgrade Transport System study - BETRAS", Institute of Transport and Traffic Engineering Faculty, JUGINUS, , 1985. [5] Dimitrijević, B., and Popović, M., “A new fuzzy TOPSIS method”, Proceedings of Symposium of Operational Research - SYMOPIS 97, Bečići, 1997. [6] Jović, J. et al., "Transport model of Belgrade", TTEF Institute, Belgrade, 2003. [7] Jović, J., “Application of neural networks to modal split modelling”, YUJOR, 7 (1997) 221- 235. [8] Jović, J., "The influence of the users’ behaviour on modal split planning", PhD Thesis, Faculty of Traffic and Transport Engineering, University of Belgrade, Belgrade, 1992. [9] Popović, M., "Concept of expert system for modal split modelling", M.Sc. Thesis, Faculty of Traffic and Transport Engineering, University of Belgrade, Belgrade, 2001. 
work_64dgu5uohbam5dywgp6ermitom	----	Building and Testing the SHYSTER-MYCIN Hybrid Legal Expert System TR-CS-03-01 Building and Testing the SHYSTER-MYCIN Hybrid Legal Expert System Thomas A. O’Callaghan, James Popple and Eric McCreath May 2003 Joint Computer Science Technical Report Series Department of Computer Science Faculty of Engineering and Information Technology Computer Sciences Laboratory Research School of Information Sciences and Engineering This technical report series is published jointly by the Department of Computer Science, Faculty of Engineering and Information Technology, and the Computer Sciences Laboratory, Research School of Information Sciences and Engineering, The Australian National University. Please direct correspondence regarding this series to: Technical Reports Department of Computer Science Faculty of Engineering and Information Technology The Australian National University Canberra ACT 0200 Australia or send email to: Technical.Reports@cs.anu.edu.au A list of technical reports, including some abstracts and copies of some full reports may be found at: http://cs.anu.edu.au/techreports/ Recent reports in this series: TR-CS-02-06 Stephen M Blackburn and Kathryn S McKinley. Fast garbage collection without a long wait. November 2002. TR-CS-02-05 Peter Christen and Tim Churches. Febrl - freely extensible biomedical record linkage. October 2002. TR-CS-02-04 John N. Zigman and Ramesh Sankaranarayana. dJVM - a distributed JVM on a cluster. September 2002. TR-CS-02-03 Adam Czezowski and Peter Christen. How fast is -fast? Performance analysis of KDD applications using hardware performance counters on UltraSPARC-III. September 2002. TR-CS-02-02 Bill Clarke, Adam Czezowski, and Peter Strazdins. Implementation aspects of a SPARC V9 complete machine simulator. February 2002. TR-CS-02-01 Peter Christen and Adam Czezowski. Performance analysis of KDD applications using hardware event counters. February 2002. http://cs.anu.edu.au/techreports/ mailto:Technical.Reports@cs.anu.edu.au Building and Testing the SHYSTER-MYCIN Hybrid Legal Expert System Thomas A. O’Callaghan <tom@tom-ocallaghan.com> James Popple <james@popple.net> Eric McCreath <eric.mccreath@anu.edu.au> Department of Computer Science Faculty of Engineering and Information Technology The Australian National University Technical Report TR-CS-03-01 May 2003 Abstract SHYSTER-MYCIN is a hybrid legal expert system created by combining rule-based and case-based reasoning. The MYCIN part uses a system of rules to reason with provisions of an Act of a parliament; the SHYSTER part uses analogy to reason with cases that explain “open-textured” concepts encountered in legislation. The construction of the expert system is focused upon: creating and evaluating a model of legal reasoning, and improving the reporting made by the MYCIN part. The model of legal reasoning is supported by jurisprudential discussion. The model holds that rules (in the strict sense of the word) cannot be extracted from cases. Cases should therefore be argued by analogy. The only rules that exist in law are those in legislation. The method of evaluating the model of legal reasoning is comparative. Reports by the system are compared with reports by a test group of legally trained people. Both the system and the test group were provided with the same material on which to base their reports. This ensured that the evaluation was of the model of reasoning, rather than the depth of knowledge. The reporting made by MYCIN was improved for use in SHYSTER-MYCIN, so that the system states how it comes to its conclusions. This reporting was then restricted to only the “interesting” conclusions. Further information on the research described in this technical report is available in [21] and [22], and at <http://cs.anu.edu.au/software/shyster/tom/>. 1 mailto:tom@tom-ocallaghan.com mailto:james@popple.net mailto:eric.mccreath@anu.edu.au http://cs.anu.edu.au/software/shyster/tom/ 1 Introduction SHYSTER-MYCIN was created by the first-named author, and is the legal expert system discussed in this paper. SHYSTER-MYCIN combines rule-based reasoning with case-based reasoning. The system is designed as a legal expert system to be consulted by legally trained people. This hybrid system enables the case-based reasoner (SHYSTER) to determine open- textured concepts when required by the rule-based reasoner (MYCIN). The system operates on a reduced version of the Australian Copyright Act 1968, including cases that define the term “authorization”1 (see section 2). The Act is reasoned by a system of rules, whereas cases are reasoned by analogy. This approach is supported by jurisprudential discussions on legal reasoning (section 3). The system was created in three progressive versions (section 4). The focus of the cre- ation of the system was the reporting of reasons for conclusions. The second and third versions were tested against three criteria: validity, conciseness and correctness (section 5). The system performed well against those criteria, indicating that the approach taken is ap- propriate: that is, it is appropriate to use rules to reason with statutes and analogy to reason with cases. 2 Domain In Australia, copyright law is governed by the Copyright Act. All rights, remedies and defences originate from the Act.2 In Australian copyright law, cases provide examples of applications of the Act and interpretations of terms used in the Act. SHYSTER-MYCIN operates upon a restricted area of the Act. Specifically the sections that are represented in SHYSTER-MYCIN are a set regarding ownership of copyrights, the acts that may be exclusively performed and the infringement of those rights. The cases that SHYSTER-MYCIN knows about are those on the meaning of the term “authorization”. The term “authorization” is undefined in the Copyright Act. Consequently, a number of cases have been before the courts seeking answers as to what conduct amounts to au- thorization. The main contexts in which the issue has arisen are: “home taping of recorded materials, photocopying in educational institutions and performing works in public” [19, page 198]. In 2002, the Act was amended to include guidelines on defining “authorization”. How- ever, this does not amount to a definition; the term remains undefined in the Act. Subsec- tions 36(1A) and 101(1A) provide guidance when determining if an act can be said to be an authorization of an infringement of an exclusive right. These new provisions essentially codify the principles drawn from the leading Australian case.3 1The meaning of the term “authorization” in the Copyright Act is one of the four areas of law on which SHYSTER was tested (see [25]). 2See section 8 of the Act: “[s]ubject to section 8A, copyright does not subsist otherwise than by virtue of this Act.” (Section 8A deals with the prerogative rights of the Crown in the nature of copyright.) 3University of New South Wales v Moorhouse (1975) 133 CLR 1, as explained in the Explanatory Memorandum to the Copyright Amendment (Digital Agenda) Bill 2000. The facts to be considered in determining whether an infringing act was authorized should include: (a) the extent (if any) of the person’s power to prevent the doing of the act concerned; (b) the nature of any relationship existing between the person and the person who did the act concerned; (c) whether the person took any other reasonable steps to prevent or avoid the doing of the act, including whether the person complied with any relevant industry codes of practice. 2 The case-base is restricted to those cases up to 1983. The reason for this is that the focus of the construction of SHYSTER-MYCIN was on the creation of the hybrid system and the improving of the MYCIN component. The case law specification that was used with the SHYSTER part consisted only of decisions up to 1983 for reasons explained in [25, page 181]. 3 Method of Reasoning Susskind has criticised all developers of legal expert systems as failing to consider jurispru- dence4 in their construction [33, page 194]. At best some attempted justification of their method after the construction of the system. Susskind suggests that jurisprudence should be the starting point for a legal expert system rather than merely a point of discussion after construction. Whilst this may be a valid point, the works that Susskind suggests (Hart, Dworkin, Fin- nis or Raz) as starting points are not, at least immediately, useful. Jurisprudence almost solely deals with general questions such as “what is law?”and “what is good law?”. The ju- risprudents have rarely studied the question of “how do we argue with law?” or “how does a lawyer reason?” The answers or discussion on these questions must surely provide a more sturdy ground to begin the construction of a legal expert system than a theory of “what law is”. Chandler states that “[m]ore must be known about the mental operations a lawyer per- forms when engaging in case law research before the computer can be programmed to aid him to the full extent of its capacity” [10]. Bing is in agreement: “[a] computer program for legal reasoning cannot be created without first characterising the task to be performed and the means by which the reasoning agent performs it” [5]. Hart [14] views the legal system as heavily rule-based. He claims that rules can be ex- tracted from all cases, and that these are “as determinate as any statutory rule” [page 135]. However, just prior to this statement, he concedes that there is “no authoritative or uniquely correct formulation of any rule to be extracted from cases” [page 134]. Hart attempts to rec- oncile these statements by claiming that whilst “no authoritative or uniquely correct” rule exists, there is “very general agreement” [page 134]; yet it still seems a leap of faith to char- acterise that “rule” as being as determinative as rules extracted from Acts of a parliament. If the “general agreement” is such that the rule is so well defined and understood by all, then why is there litigation? In Hart’s world, in every dispute both parties must already know which one will lose. Eisenberg [13] explains in detail the process of extracting rules from cases. In his discus- sion on reasoning by analogy, he rejects Levi’s [17] characterisation (as described below) of legal reasoning as only involving reasoning by example/analogy: It may be that example [analogy] has a part to play in the intuitive leap of dis- covery. Courts, however, cannot leave matters at that. Courts must justify their results by objective reasons that meet certain criteria, and must reject intuitive 4“The subject matter of jurisprudence, whether the discipline be classified as an art or science, is the nature of law and its working” [18, page 1]. Chinhengo notes the etymology of the word “jurisprudence” from its Latin routes—juris “of the law” and prudens “skilled”—and indicates this vague definition of the term has meant that over the years the term has taken on a number of different meanings [11, page 2]. He continues, stating that “jurisprudence may be said to involve the study of a wide range of social phenomena, with the specific aim of understanding the nature, place and role of law within society”. The two chief divisions of jurisprudential enquiry were defined by Austin as analytical and normative, addressing general questions of “what is law?” and “what is good law?” respectively [4]. 3 conclusions that they cannot justify in this way. In a normative context, justifi- catory reasoning can proceed only from standards, and “reasoning by example,” as such, is virtually impossible. Reason cannot be used to justify a normative conclusion on the basis of an example without first drawing a maxim or rule from the example (or, what is the same thing, without first concluding that the example “stands for” a maxim or rule). [page 86] Eisenberg believes that a rule must be created somewhere in the process of reasoning by analogy. However, in his description above, analogy can occur as the first stage in the process. The “concluding that the example ‘stands for’ a maxim or rule” is simply the process of stating how the two cases are alike. The rule is created with a specific association and application in mind. It is therefore not a general, but instead a very specific “rule”. Aarnio puts forward a view of legal reasoning in which induction is allowable, but it only provides prediction, not certainty. Aarnio contrasts law with nature, declaring that in nature there are “regularities, instances of invariability” which permit generalisation, whereas law is “volative, a result of human will” [1, page 79] and consequently generalisations can never hold. Aarnio explains [at page 79] the inability to generalise in law by an example of the local- isation of any “rule”: “If a person, then, has checked cases a, b, c and d and has stated that legal principle Ni is expressed in all of them, this does not yet entitle him to claim that the principle is general in nature.” Whilst Aarnio rejects the idea of declaring a rule “general in nature”, he believes that prediction is possible. When there are several cases all of which express the same rule, then the “possibility to draw up a plausible prediction increases” [page 256]. He does, however, deny the possibility of making a prediction or rule from a single case. Although more cases give a better chance of defining a rule, Aarnio warns of Dray’s paradox: “a law providing a historical explanation may be sufficiently comprehensive only when it contains such a large number of restrictive conditions that in the end it only concerns the individual case that should be explained” [page 73]. As the certainty of a prediction in- creases with the addition of more cases from which the prediction is inducted, the generality of the prediction decreases. Aarnio cautiously concedes that inducting a rule may be possible from cases. However, this concession is made with such a limitation for it to almost be a denial of the appropriate- ness of inducting a rule. The concept of a rule fits very neatly into computer science, as “if . . . then . . . ” statements have been a part of computing since its inception. However, the ease of handling such representations should not be the motivation for modelling the real-world system in that way. In the second edition of his work Allen separates legal reasoning into two categories which are neatly separated by the line between case-law and statute-law. Allen states [2, page 248]: “Whereas precedent is inductive, enactment clearly imposes the necessity of de- duction upon the Courts. It is general and comprehensive in form, precedent particular and limited. A decision, whatever implications may be read into it by subsequent comparison and interpretation, exists primarily for the settling of a particular dispute: a statute purports to lay down a universal rule.” By the time of his seventh edition [3], Allen states that, whilst the method of arguing with cases is usually termed induction, what really happens is argument by analogy. Allen regards analogy as the best and most common form of argument—“a close analogy is more convincing than a far-fetched illustration” [3, page 286]. “Every ratio is an interpretation of authorities in the light of the facts of the instant case . . . The ratio is thus in a constant 4 state of flux . . . it is not susceptible of any precise and comprehensive definition” [page 60]. When interpreting the ratio of a case in light of the instant case, analogy is necessarily involved. Thus there is a blurring in the four-step process typically taught to law students of “Issue-Rule-Application-Conclusion” (see for example [9, page 58–60]), that is, the rule and its application should be considered as one question—how the rule is to be applied to the facts of the instant case dictate how the rule will be formed.5 Levi [17] explicitly states that the process of “Issue-Rule-Application-Conclusion” is not just blurred, but in fact is reversed (at least in the middle). Levi believes that the use of analogy is the method of arguing with cases in law—“the finding of similarity or difference is the key step in the legal process” [page 2]. By arguing with cases using the method of analogy, “the rules arise out of a process which, while comparing fact situations, creates rules and then applies them” [page 4]. Levi admits that such a description of the process of legal reasoning will not sit well with lawyers and judges, as it “runs contrary to the pretense of the system” [page 9]. However, he sees it as much more dangerous to continue in the belief of a system of rules being es- tablished from cases: “[t]he rule will be useless. It will operate on a level where it has no meaning . . . The statement of the rule is roughly analogous to the appeal to the meaning of a statute or of a constitution, but it has less of a function to perform. It is window dressing. Yet it can be very misleading” [page 9]. Leith [16] begins his discussion of the “AI Man’s View of Law” with the following obser- vation: “it is almost as though when God made computer scientists, he made them all think of law in the same way—as a system of rules.” Leith views the law as being more than simply a system of rules. He states [at page 511] that: “it seems to me to be all very well to draw up a collection of rules from legislation; but, as lawyers all know intimately, a piece of legislation is but one thing in the legal world.” Leith does not explicitly state that rules are an inappropriate way of reasoning with cases, but it is an obvious conclusion to make based on this statement. Leith therefore presents the view that more than rule-based reasoning is required to rea- son in “the legal world”. Leith states that rule-based reasoning is appropriate for statutes, but that it is not appropriate for the rest of the law (for example, cases). Schauer states that whilst we speak of rules in the common law, they are “so malleable so as not to even be rules” [29, page 177]. Schauer appears to be of the view that rules in the proper sense cannot be elicited from previous cases: “[precedent] cannot serve to provide the rule-like constraint” [page 184]. Schauer states that there is an ambiguity in the word “rule”. This ambiguity causes some jurisprudents to believe that the “rule of law” means that the law consists of rules. Schauer explains [at page 167] the use of the word “rule”: “[i]n the sense that we have rulers who rule their subjects, ‘rule’ bears its closest affinity with ‘reign’ or ‘control’, and has only the remotest relationship with a form of decision-making characterized either by generality or by the entrenchment of generalizations.” At page 177, Schauer explains that whilst “rules” may claim to be applied, their applica- tion is not by way of interpretation. Rather “rules” are used as guidelines: [a]lthough lawyers and judges can describe any number of common-law rules, and although both opinions and textbooks can state them in ‘black letter’ fashion, the rules have no single authoritative formulation, and accordingly the process of applying them does not involve an interpretation of the text of the rule . . . it 5This method of reasoning appears to be heading the way of the rule skeptics. The rule skeptics see citing of legal rules in judgments as a ex post facto justification of the decision in a case rather than the sources upon which to reach the decision. 5 appears that common-law ‘rules’ are indeed descriptive rather than prescriptive, functioning merely as temporary guides. Schauer [at page 178] agrees with the rule skeptics, in that “[t]he common law appears . . . to be decision according to justification rather than decision according to rule.” Schauer identifies that there is a problem with claiming to find “rules” in cases. The problem that Schauer identifies [at page 183] is that at the outset of constructing a rule, the predicate (the facts of the case) must be stated. These facts cannot be easily stated: What distinguishes reasoning from precedent from reasoning from rule, how- ever, is the necessity in precedential reasoning of constructing the generaliza- tion/factual predicate that already exists in the case of a rule. As we have seen, the factual predicate of a rule, a generalization necessarily encompassing a mul- tiplicity of events, is part of the rule’s canonical form. But where there is only a previous decision and no rule-formulation, the source of the factual predicate is obscure, and consequently the manner in which the previous decision constrains becomes problematic. For Schauer, the concept of precedent is to ensure the same result on the same facts. He notes [at 183]: “[n]o two events are exactly alike, but the idea of precedential constraint presupposes that a prior decision will control a subsequent set of facts that are like the first.” Here the word “like” describes a definite association. That is, one case is “like” another. Cases are therefore compared by analogy. To create a rule from a case would be to alter the use of “like” to describe indefinite, possible association. Schauer does not agree with the creation of rules from cases, and con- sequently would not condone this extension of the meaning of the word “like”. The comments that Schauer has made as to the problems with constructing “rules” are restricted to the common law. That is, Schauer does not agree with the proposition that cases can be argued with by use of rule-based reasoning. The method of reasoning that Schauer advocates is that of analogy. Some of those who support the idea of extracting rules from cases admit their short- comings. These “rules” are: local not general, temporal rather than permanent, and subjec- tively considered correct rather than universally. Given these limitations, it hardly seems appropriate to refer to whatever is extracted from a case as a “rule”. It is a guide that is extracted or a principle. The use of the word rule (in legal circles) is not designed to convey the same strictness as a rule say in mathematics— something that cannot be broken. The use of analogy to compare cases seems to fit well with the doctrine of precedent.6 Such a method of reasoning has support from some of the above-mentioned jurisprudents. 6To select an appropriate line of authority, the courts adhere to the “doctrine of precedent”. At a practical level, the doctrine provides a method of deciding how binding or applicable a previously decided case is, based upon which court decided that case. At a policy level, the doctrine provides certainty, equality, efficiency, and the appearance of justice (in the sense that the law is seen to operate consistently). Cook et al. have summarised the general rules of precedent [12]: • each court is bound by decisions of courts higher in its hierarchy; • a decision of a court in a different hierarchy may be of considerable weight but will not be binding; • only the ratio decidendi (the judge’s decision on the material facts) of a case is binding; • any relevant decisions, although not binding, may be considered and followed; and • precedents are not necessarily abrogated by lapse of time. 6 The idea of analogy is to show how one thing is like another. The doctrine of precedent purports to treat like cases alike. 4 The Design of SHYSTER-MYCIN SHYSTER-MYCIN combines MYCIN and SHYSTER, two previous expert systems. MYCIN7 is a medical expert system, which was adapted for use in SHYSTER-MYCIN. SHYSTER8 is a legal expert system, and was used without alteration in SHYSTER-MYCIN. MYCIN’s “certainty factor” is not used in SHYSTER-MYCIN.9 The reason for this is the difficulty in scientifically establishing how certain a fact is in a legal domain. In medicine, the “certainty factor” can be established by calculating the error in measurement, or statisti- cally measured certainties of test results. In the law, the vast majority of conclusions cannot be established by scientific methods and therefore a “certainty” cannot be attached to them. For example the speed of a motor vehicle can be observed and the level of certainty of that observation stated. However, the level of certainty that (given that speed) the driver was driving at an excessive speed in all the circumstances10 cannot be determined by way of a formula.11 In SHYSTER-MYCIN, the MYCIN part is used to reason with the provisions of an Act of Parliament only. The MYCIN part is not used to reason with so-called rules from decided cases. Similarly, SHYSTER does not reason with the provisions of an Act—its domain is decided cases. This clear delineation between the use of rule-based and case-based reasoning has not always been made in hybrid legal expert systems. Some previous hybrid systems would represent so-called “clear” cases using rules, storing the remaining cases in a case-based system (see for example CABARET [26, 27, 28, 30, 31, 32] and GREBE [6, 7]). The SHYSTER part of SHYSTER-MYCIN has been left untouched, and is only called upon for questions relating to its knowledge on the definition of “authorization”. In this 7Created through “collaboration between the medical and AI communities at Stanford” [15], beginning in 1972. The version of MYCIN that was used for SHYSTER-MYCIN was created by Peter Norvig [20] (and is available at http://www.norvig.com/paip.html). Consequently, comments about the system are com- ments about the Norvig version of MYCIN, not necessarily the original system itself. For an account of the original system, see [8]. 8Popple created SHYSTER as part of his PhD research at the Australian National University [24]. SHYS- TER “represents the state of the art in statistical legal reasoning” [23]. See also <http://cs.anu.edu.au/ software/shyster/>. 9Or, more precisely, all the certainty factors are set at 1. 10For example, perhaps the driver was travelling faster than the posted speed limit, but was doing so because she was transporting to hospital a person who had just suffered a heart attack. Objectively she is “speeding” as she is travelling at a speed greater than the posted limit. Supposing that there is provision to be excused from a fine for speeding if the speeding was “necessary or not excessive in the circumstances”, the case described in this footnote would fit the “necessary or not excessive in the circumstances” test, as she was speeding in the hope of saving a person’s life. The objective fact can be described, noting the margin of error in making the measurement of the car’s speed. This can be expressed using a “certainty factor”. The subjective fact is a unique, subjective weighting of a multitude of secondary facts. As the weighting of the secondary facts is also subjective, a “certainty factor” cannot be attributed to the primary subjective fact. 11A statistician may claim that, with a sufficient number of cases, a certainty factor could be established. The first problem is that the majority of cases do not make it onto a public record, having been settled outside the court system. However, if we accept a large enough number of cases could be observed, taking the example in the previous footnote, the number of “valid excuse” cases could be compared with the total number of speeding cases, resulting in a “certainty factor”. This would give the probability that the accused may have a valid excuse. Importantly it says nothing about the actual validity of the excuse. Consequently a “certainty factor”, so determined, has a very limited use—it can only provide the user with an estimation of how valuable further facts may be. 7 http://cs.anu.edu.au/software/shyster/ http://cs.anu.edu.au/software/shyster/ http://www.norvig.com/paip.html way the SHYSTER part is like an expert that the MYCIN part calls upon when it cannot answer a question. 4.1 The intended use of SHYSTER-MYCIN SHYSTER-MYCIN is created as a tool for use by lawyers or para-legals. It is designed as a system that would fall into the “internal knowledge systems” quadrant of Susskind’s “Legal Grid” [33, page 9]. That is, the system is designed to speed up internal processes of handling a matter. With an improved method of fact elicitation and classification, the system could be moved into the “online legal services” quadrant. This improvement would require the system to “get the facts right” when questioning a lay person. 4.2 The versions of SHYSTER-MYCIN SHYSTER-MYCIN was produced in three different versions. The first version was very basic, and was used to make a preliminary assessment of the appropriateness of coupling a rule-based and a case-based system to reason with sections and cases, respectively. The second version of the system had a greatly increased rule base, and alterations were made to the reporting of results so that reasons for a conclusion, rather just the conclusion, would be reported. The third version of the system produced more concise reports, by limiting the conclusions that would be reported. 4.2.1 Version 1 The first version of SHYSTER-MYCIN (“SM-v1”) was created to provide a preliminary as- sessment of the approach that would be undertaken with the later versions of SHYSTER- MYCIN. That approach was to provide the MYCIN part with a rule base drawn from pro- visions of the Copyright Act, with SHYSTER available to be called upon when an “open textured” term12 was encountered. When an “open textured” term was encountered, the user was notified that it was a term that might be best answered by consulting the SHYSTER part. The user had a choice at this stage to answer the question based on their own knowledge, or to consult SHYSTER. If SHYSTER was consulted, the user answered SHYSTER’s questions, and, at the end of the consultation, was given the likely result. The user then gave this answer to the MYCIN part. This allowed the user to “over-rule” the SHYSTER part if they so wished. SM-v1 has a rule base of 16 rules; it draws conclusions based on the values of nine para- meters. The rules are used to represent subsections 13(2), 36(1), and 101(1) of the Copyright Act. These are the three provisions of the Act in which the term “authorization” is used. As explained in section 2 above, this term, whilst used in the Act, remains undefined. The parameters for SM-v1 are the facts that are asked of the user or that are determined by applying rules to the facts obtained from the user. The parameters that SM-v1 uses are: 1. the name of the material; 2. the type of the material; 3. whether the accused was the owner of the material; 4. whether the accused had a licence to use the material; 12Specifically, for this system, the term “authorization”. 8 5. whether the accused had authorized someone else to use the material; 6. whether the use of the material occurred in Australia; 7. whether the accused had infringed the owner’s rights under subsection 13(2); 8. whether the accused had infringed the owner’s rights under subsection 36(1); and 9. whether the accused had infringed the owner’s rights under subsection 101(1). The last three parameters are the “goal” parameters. These are the facts that SM-v1 attempts to establish by applying the rules it knows to the facts asked of the user in relation to parameters 1–6. Interactions with SM-v1 indicated that the approach proposed would be valid, and worth continuing. The MYCIN part of SM-v1 was able to work logically through the provisions in the Act that were provided to it as rules. However, SM-v1 only “knew” of three subsec- tions of the Act. This meant that the answers to questions asked by SM-v1 relied upon the user being familiar with the remainder of the Act. For example, the user had to make as- sessments as to “ownership” of copyrights—something covered by other provisions of the Act.13 Also, that the copyright material was used by the accused, and that that use was an exercise of one of the exclusive rights14 for that material, were facts assumed by the system. In advancing from SM-v1, more provisions of the Act had to be added to the system, and better reporting had to be implemented. 4.2.2 Version 2 SM-v2 uses the same approach as SM-v1 in that the MYCIN part reasons with provisions of the Copyright Act and the SHYSTER part reasons with decided cases. Version 2 differed from version 1 in three areas: the size of the rule base, the debugging of the MYCIN part, and the reporting of conclusions. The rule base that the MYCIN part was working from was greatly expanded in version 2 as compared with version 1. The provisions of the Act that the MYCIN part knew were increased, making the system more realistic. This meant that most of the terms or concepts encountered in subsections 13(2), 36(1) and 101(1) were determined in surrounding sections. For example, to determine whether the accused was the owner of the copyright, rules repre- senting subsection 35(2) were added to explain that a person who authored a work owned the copyright. In SM-v2 the provisions that the MYCIN part knows about are: sub-s 13(2) the right to authorize acts; s 31 the acts the owner has an exclusive right to (for works); s 35 determining the owner of a copyright (for works); s 36 how copyright is infringed (for works); ss 85–88 the acts the owner has an exclusive right to (for subject matter other than works); ss 97–100 determining the owner of a copyright (for subject matter other than works); and s 101 how copyright is infringed (for subject matter other than works). The rules that represent these provisions of the Act total 273. The number of parameters used by these rules increased from 9 to 56. 13Sections 35 and 97–100 of the Act. 14Sections 31 and 85–88 of the Act. 9 In the process of expanding the rule base and altering the reporting (discussed below), it became apparent that a method of viewing how the MYCIN part stepped through its rules would be useful. To achieve this, the MYCIN reasoner was altered so that, when each question was asked, information about the rule currently under consideration was recorded in a file. This record was called a “stream of consciousness”. It detailed why the MYCIN part was asking each question, and provided information as to how the system arrived at its conclusions. This record was useful in debugging the rule base. The file provided a step-by-step record, which assisted in checking that the rules were entered in the way that they were meant to be entered, in order to accurately represent the provisions of the Act. This record of the stream of consciousness was also the first attempt to improve the reporting. However, it was immediately obvious that the record was far too long. All rules that came under considerations were recorded. This meant many non-firing15 rules were included in the file. Thus the interesting parts were hidden by the sheer volume of the record. Prior to altering the MYCIN part, the reporting was very limited. When a conclusion was reached, only the conclusion would be reported at the end of the consultation. Importantly, no reasons supporting the conclusion were reported. The reporting of conclusions was improved in SM-v2 to make the MYCIN part fit with the generally accepted definition of an expert system. That is, an expert system should report on reasons for reaching conclusions, rather than simply return the conclusions. The improvement to the reporting was made such that, when the reasoner was conclud- ing a rule, the “report-why” function would be called upon. This function writes to a file the facts that were known, the rule that was applied to these facts, and the conclusion that was consequently made. The report is made using LATEX tags, so that the report from the MYCIN part can be combined with the output from the SHYSTER part (which also produces LATEX output), to produce a more cohesive report. SM-v2 reported upon every conclusion drawn; this approach was slightly—but signifi- cantly—altered in SM-v3. 4.2.3 Version 3 SM-v3 operated on the same rule base that SM-v2 did, however, the reporting of conclusions was altered. The reporting done by the MYCIN part was restricted to only reporting on conclusions that were made by relying on more than one fact. 5 Testing and Results The reporting made by SHYSTER-MYCIN was assessed against three criteria: validity, con- ciseness and “correctness”. The testing of SHYSTER-MYCIN was made by way of compar- ison with reports from three legal professionals. 5.1 The testing methodology To determine the appropriateness of the approach taken in SHYSTER-MYCIN, the reporting that it makes was assessed. To do this the system’s report was compared with reports made by legal professionals. The test group consisted of three legal professionals: a law graduate, a practising solicitor with five years’ experience and one with 30 years’ experience. 15When a rule “fires” it is activated. That is, the premise are all found to be true, and the consequence is to reach the prescribed conclusion, or perform the required action. 10 None of the test group was expert in copyright law. The reason for selecting such a group of people was to make comparisons with SHYSTER-MYCIN on as level a playing field as possible. The test group was provided with the same material that SHYSTER-MYCIN was provided with. This method of testing differs from previous evaluations of expert systems. Previously, expert systems were made to compete with human experts which were allowed to draw on years of experience and knowledge not represented in the expert system. Such an evaluation will result in comments about the inadequacy of the volume of knowledge in expert system. However, the expansion of an expert system’s knowledge is simply a matter of time and resource. The method of testing SHYSTER-MYCIN provides evaluation of the model of legal rea- soning. Using this method of testing, if a model of legal reasoning is evaluated favourably, then appropriate investment can be made in expanding its knowledge. 5.2 The testing process The test group was given a series of short questions to answer. The group was instructed to answer these questions using a short version of the Copyright Act and case summaries. The version of the Act was the same set of sections that SHYSTER-MYCIN knew (except for a few preliminary sections: 1–10). The cases summarised were those in SHYSTER’s case law specification for the meaning of “authorization”. The case summaries provided the test group with: the name of the case, the facts of the case and some commentary on the case. Using only these materials, the test group answered the questions. SHYSTER-MYCIN was used to answer the same set of questions that the test group answered. 5.3 The validity A report by SHYSTER-MYCIN was valid if it referenced the same sections as a majority of the test group. SM-v2 took a lazy or verbose approach, reporting on every section on which it made a conclusion — this was invariably every section that it knew. SM-v3, on the other hand, only reported a handful of conclusions, yet also managed to make valid reports. By limiting the reporting of SM-v3 to conclusions made on more than one fact, SM-v3 would reference the sections that the test group would reference and usually only 2–3 extra sections. SM-v2 would make approximately 11 excess references, as compared with the test group. 5.4 The conciseness The conciseness of each report was measured by counting the number of conclusions stated. The more concise a report was, the better it was considered to be. A concise report is consid- ered to be the ideal because the system should be able to provide reasons for its conclusions, but should not regurgitate a copy of its rule-base as “explanation”. Explanation should involve a condensing or summarising of information. The user should only be told the “in- teresting” parts—however that may be defined. However, a report should not become more concise at the expense of its validity. On average, SM-v3 reported only 24% of the conclusions that SM-v2 reported. Just on this comparison, version 3 seems to have an advantage over version 2. According to SM-v3, most of the conclusions (approximately 3/4 of them) are uninteresting. The criterion that 11 SM-v3 uses to eliminate the uninteresting conclusions is to not report the conclusions made by applying a rule to a solitary fact (see section 4.2.3). This criterion is based upon the idea that conclusions based on a single fact are simply one-to-one mappings between one fact and another and do not give the user any real information. At best the information that the user provided is regurgitated to them with a slightly different wording. Conversely, a conclusion is interesting if it is arrived at by combining several facts. When comparing SM-v3 with the test group, it can be seen (in Figure 1) that SM-v3 does perform much better than SM-v2, yet there is still room for improvement. On average the test group reported 12% of the conclusions that SM-v2 did, or about half as many as SM-v3 did. In answering Question 2, SM-v3 and the test group were fairly equal in the number of conclusions reported. SM-v3 reported on 6 conclusions, with the test group averaging just under 5. The greatest difference was observed in answers to Question 5: SM-v3 reported on 5 conclusions, each member of the test group reported on only one. It is suggested that by the time of Question 5, the test group decided to rely heavily upon references to their earlier answers and, as a consequence, only had a single conclusion each to report. ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� � � � � � � � � � � � � � � � � � � � � � � � � � � ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ����� ����� ����� ��� ��� ��� � �� � Q1 Q2 Q3(i) Q3(ii) Q4(i) Q4(ii) Q5 N um be r of C on cl us io ns 10 15 20 25 30 5 0 SM−v2 SM−v3 Test group Figure 1: The number of conclusions reported for each question 5.5 The “correctness” The “correctness” of the report is assessed by seeing if the “right” result is returned as an answer to the question raised by the factual scenario. The “right” result is not a scientifically observable fact—in the legal domain, several answers may be logically correct. That is, the answer is a logical conclusion that results from the application of Acts of Parliament and the Common Law to the facts of the present case. Therefore, to assess the “correctness” of the results reported by SHYSTER-MYCIN, there must be some method of determining the “right” result for each question. To this end, the results given by the test group were taken to be representative of the “right” result. As there were three members of the test group, the “right” result would be the result returned by a majority of the group. In this way the test group was akin to a panel of three judges hearing a case. It is therefore assumed that the results of the test group are indicative of a good answer and, when taking the majority view, a “right” answer. A “correct” system will provide answers that are the same as the majority of the legal experts in the test group. SM-v2 and SM-v3 operate on the same rule base and use the same methods for accessing and concluding rules. Their difference is only in the reports they make. Therefore, because 12 the rules that they use and their use of them is the same, they will always return the same result for the same set of answers to their questions. For this reason, when assessing the correctness of the system, the system generally (versions 2 and 3) will be referred to. SHYSTER-MYCIN agreed with the experts for each answer to the questions. When opin- ion was divided amongst the experts, the system agreed with the majority view. The test group agreed on the result for each question except for two questions. These splits in the decisions of the test group arose because of disagreements over the classification of facts. When the facts as classified by the minority were used to answer SHYSTER-MYCIN’s ques- tions, the answer given by the minority was returned. These results indicate that the system provides “correct” answers. Further, these results also indicate that a crucial part of reaching a decision is the initial step of classifying the facts of the case. 6 Conclusions The results of testing SHYSTER-MYCIN show that the approach taken in constructing the system is appropriate. That is, it is appropriate to use rule-based reasoning when dealing with statutes, and it is appropriate to use case-based reasoning when dealing with cases. The results suggest that if the entire Copyright Act and associated cases were represented within SHYSTER-MYCIN (that is, with the whole Act represented in the MYCIN part, and all of the open-textured concepts resolvable by reference to an enhanced case-base in the SHYSTER part), the system would be expert in the entire body of copyright law. 6.1 Future research Given sufficient time and resources, the entire Copyright Act and relevant cases could be represented in SHYSTER-MYCIN. Before translating the entire Act into rules for the MYCIN part, some sort of rule man- agement system should be created. This is because, in creating SHYSTER-MYCIN, it was observed that a great number of rules are required to represent only a small number of pro- visions in the Act. Without a rule management system, the rule-base is “fragile”: that is, changing it would most likely result in error. The system could be further improved in the conciseness of the reporting by the MYCIN part. By defining a “positive conclusion”, the reporting could be further restricted to only report conclusions made on more than one fact which was concluded positively. A “positive conclusion” would have the effect of assigning a direction to a fact. A positive answer would be one that brings the system a step closer to a goal.16 Although left untouched in the construction of SHYSTER-MYCIN, the reasoning per- formed by SHYSTER could be altered. Methods of analogy other than nearest neighbour could be employed by SHYSTER to select cases.17 The facts collected by the MYCIN and SHYSTER parts should be shared between the two expert systems. At present they each store their own sets of facts, and do not pass information between each other. If the facts were stored commonly, then the potential for a user to be asked the same question twice is eliminated. This has two benefits: the user does not become annoyed or frustrated by double-questioning, and conflicting answers are not provided.18 16The answer “false” could be “positive” under this system, depending on how the question is phrased. 17As suggested by Popple [25, page 251]. 18Otherwise the SHYSTER part and the MYCIN part could be working on the basis of inconsistent facts. 13 SHYSTER-MYCIN could be tested by comparing the results given by the system when used by lay-people with the answers given by experts. This would test whether SHYSTER- MYCIN can accurately gather facts by questioning a person. Fact elicitation/classification is one of the areas that Susskind identifies as requiring greater research [33]. A system capable of getting the facts right by questioning a lay person would fall within the “online legal service” quadrant of Susskind’s “Legal Grid” [33, page 9]. Such a system could be of great use both commercially and for society. 7 Acknowledgments Thanks to Mr T. E. O’Callaghan, General Counsel, Citibank, N. A. for editorial assistance. References [1] AARNIO, Aulis 1977, On Legal Reasoning, Turun Yliopisto. [2] ALLEN, Carleton Kemp 1930, Law in the Making (second edition), Oxford University Press. [3] ALLEN, Carleton Kemp 1964, Law in the Making (seventh edition), Oxford University Press. [4] AUSTIN, J 1832, The Province of Jurisprudence Determined, Weidenfeld & Nicolson. [5] BING, Jon 1990, “Legal decisions and computerized systems”, in From Data Protection to Knowledge Machines: The Study of Law and Informatics, ed. P Seipel, Kluwer Law and Taxation Publishers. [6] BRANTING, L Karl 1989, “Representing and reusing explanations of legal precedents”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, pp. 103–10. [7] BRANTING, L Karl 1991, “Reasoning with portions of precedents”, in Proceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, pp. 145–54. [8] BUCHANAN, Bruce G and SHORTLIFFE, Edward H (eds) 1985, Rule-Based Expert Sys- tems: The MYCIN Experiments of the Stanford Heuristic Programming Project, Addison- Wesley Publishing Company. [9] CALLEROS, Charles R. 1994, Legal Method and Writing (second edition), Little, Brown & Company. [10] CHANDLER, James P 1974, “Computers and case law”, Rutgers Journal of Computers (Technology) and the Law, vol. 3, pp. 202–18. [11] CHINHENGO, Austin 2000, Essential Jurisprudence (second edition), Cavendish Publish- ing Limited. [12] COOK, Catriona, CREYKE, Robin, GEDDES, Robert and HOLLOWAY, Ian 2001, Laying Down the Law (fifth edition), Butterworths. 14 [13] EISENBERG, Melvin Aron 1988, The Nature of the Common Law, Harvard University Press. [14] HART, H. L. A. 1994, The Concept of Law (second edition), Oxford University Press. [15] JACKSON, Peter 1986, Introduction to Expert Systems, Addison-Wesley Publishing Com- pany. [16] LEITH, Philip 1986, “Legal expert systems: Misunderstanding the legal process”, Com- puters and Law, no. 49, pp. 26–31. [17] LEVI, Edward H. 1961, An Introduction to Legal Reasoning (seventh edition), The Univer- sity of Chicago Press. [18] MCCOUBREY, Hilaire and WHITE, Nigel D. 1999, Textbook on Jurisprudence (third edi- tion), Blackstone Press Limited. [19] MCKEOUGH, Jill, BOWERY, Kathy and GRIFFITH, Philip 2002, Intellectual Property (Com- mentary and Materials) (third edition), Lawbook Co. [20] NORVIG, Peter 1992, Paradigms of Artificial Intelligence Programming: Case Studies in Com- mon Lisp, Morgan Kaufmann. [21] O’CALLAGHAN, Thomas Alexander 2003, A Hybrid Legal Expert System, Honours the- sis, Department of Computer Science, Faculty of Engineering and Information Tech- nology, The Australian National University, Canberra, February. <http://cs.anu.edu. au/software/shyster/tom/thesis.pdf>. [22] O’CALLAGHAN, Thomas A., POPPLE, James and MCCREATH, Eric 2003, “SHYSTER- MYCIN: A hybrid legal expert system”, in Proceedings of the Ninth International Con- ference on Artificial Intelligence and Law (ICAIL-03), Edinburgh, Scotland, 24–28 June, ACM, pp. 103–4. ISBN 1 58113 747 8. <http://cs.anu.edu.au/software/shyster/tom/ icail-03.pdf>. [23] PANNU, Anandeep 1995, “Using genetic algorithms to inductively reason with cases in the legal domain”, in Proceedings of the Fifth International Conference on Artificial Intelli- gence and Law (ICAIL-95), College Park, Maryland, 21–24 May, pp. 175–184. [24] POPPLE, James 1993, SHYSTER: A Pragmatic Legal Expert System, PhD thesis, The Aus- tralian National University, Canberra, April. ISBN 0 7315 1827 6. <http://cs.anu.edu. au/˜James.Popple/publications/theses/phd.pdf>. [25] POPPLE, James 1996, A Pragmatic Legal Expert System, Applied Legal Philosophy Se- ries, May, Dartmouth, Aldershot. ISBN 1 85521 739 2. <http://cs.anu.edu.au/˜James. Popple/publications/books/shyster.pdf>. [26] RISSLAND, Edwina L 1990, “Artificial intelligence and law: Stepping stones to a model of legal reasoning”, The Yale Law Journal, vol. 99, no. 8, June, pp. 1957–81. [27] RISSLAND, Edwina L and SKALAK, David B 1989, “Combining case-based and rule- based reasoning: A heuristic approach”, in Proceedings of the Eleventh International Joint Conference on Artificial Intelligence (IJCAI-89), Detroit, Michigan, pp. 524–530. 15 http://cs.anu.edu.au/software/shyster/tom/thesis.pdf http://cs.anu.edu.au/software/shyster/tom/thesis.pdf http://cs.anu.edu.au/software/shyster/tom/icail-03.pdf http://cs.anu.edu.au/software/shyster/tom/icail-03.pdf http://cs.anu.edu.au/~James.Popple/publications/theses/phd.pdf http://cs.anu.edu.au/~James.Popple/publications/theses/phd.pdf http://cs.anu.edu.au/~James.Popple/publications/books/shyster.pdf http://cs.anu.edu.au/~James.Popple/publications/books/shyster.pdf [28] RISSLAND, Edwina L and SKALAK, David B 1989, “Interpreting statutory predi- cates”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, pp. 46–53. [29] SCHAUER, Frederick F 1991, Playing by the rules: A philosophical examination of rule-based decision making in law and in life, Oxford University Press. [30] SKALAK, David B 1989, “Taking advantage of models for legal classification”, in Pro- ceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, pp. 234–41. [31] SKALAK, David B and RISSLAND, Edwina L 1991, “Argument moves in a rule-guided domain”, in Proceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, pp. 1–11. [32] SKALAK, David B and RISSLAND, Edwina L 1992, “Arguments and cases: An inevitable intertwining”, Artificial Intelligence and Law, vol. 1, no. 1, pp. 3–44. [33] SUSSKIND, Richard E 2001, Transforming the law: Essays on technology, justice, and the legal marketplace, Oxford University Press. 16 [Cover] [Verso cover] Abstract 1 Introduction 2 Domain 3 Method of Reasoning 4 The Design of SHYSTER-MYCIN 4.1 The intended use of SHYSTER-MYCIN 4.2 The versions of SHYSTER-MYCIN 4.2.1 Version 1 4.2.2 Version 2 4.2.3 Version 3 5 Testing and Results 5.1 The testing methodology 5.2 The testing process 5.3 The validity 5.4 The conciseness 5.5 The "correctness" 6 Conclusions 6.1 Future research 7 Acknowledgments References 
work_64wagicu2ncslhlep6cmcmqhly	----	llllllllllllllllllllllllllll~llllllllllll 1402782330 *’ 1; 2 I-- ’ : .. ‘\t\. ,. : ” ’ I , ‘., ;. MARKETING PLANNING AND EXPERT SYSTEMS : AN EPISTEMOLOGY O F PRACTICE BY DR. M A L C O L M H.B. M C D O N A L D P R O F E S S O R O F MARKETING PLANNING AND DIRECTOR O F THE CRANFIELD MARKETING PLANNING RESEARCH CENTRE, CRANFIELD S C H O O L O F MANAGEMENT, CRANFIELD, BEDFORD M K 4 3 O A L TEL: 0234 751122 MARKETING PLANNING AND EXPERT SYSTEMS : AN EPISTEMOLOGY OF PRACTICE Abstract After nearly a quarter of a century, Artificial Intelligence, in spite of all its promise, has made virtually no progress in the domain of marketing, and whilst most interested parties view them as a potentially powerful way of beating the competition, there are few products and no on-line systems available. This paper explores why progress has been so slow in the domain of marketing and describes the experience and progress of a group of major British multinational companies who have joined forces to produce an Expert Marketing Planning System, EXMAR, with the author of this paper as principal expert. A number of conclusions are drawn, but one of the main ones is that the development of EXMNZ shows that it is possible to use Expert System methodologies to build support systems in complex areas of marketing management, especially if the domain is well defined, has a large number of factors to be considered, and relevant expert knowledge is available. Also Expert Systems are shown as being useful in helping both academics and practitioners to structure, validate and use marketing knowledge and to better understand the interrelationships between the elements of marketing. In particular, it forces managers to think deeply and in a structured way about the issues that need to be considered in developing a strategic marketing plan. MARKETING PLANNING AND EXPERT SYSTEMS : AN EPISTEMOLOGY OF PRACTICE Just imagine what would happen to a major industrial company’s profitability if, instead of expert marketing knowledge being hoarded in the heads of an elite but small number of very experienced and successful marketing managers, &l of the company’s worldwide marketing decisions were being made using this expertise. Imagine what would happen to a bank’s profitability if &! the decisions were being made by its very best bankers. Imagine what would happen to a Unit Trust company if & the investment decisions were being made by their very best experts. After nearly a quarter of a century of Expert Systems, dreams such as this now seem possible. But there is still a long way to go, and many formidable technical and methodological obstacles still remain to be overcome. A surprising fact about Expert Systems is that although they have inspired a number of new programming languages and powerful new computer architectures, they have made virtually no progress in the domain of marketing, and whilst most interested parties view them as a potentially powerful way of beating the competition, there are few products and no 1 2 on-line systems available . Because Artificial Intelligence has become the latest buzzword, many software houses are hyping up their old software in advertisements, but most of these can be discounted as irrelevant in the real world of Expert Systems3. The principal reasons for this lack of progress centre around the technical problems associated with getting computers to mimic experts and the costs involved. - 1 - There are no shortcuts to building good expert systems. It takes a considerable amount of skill, patience and several years of effort to develop an expert system in a n e w area and get it into the field 4. The purpose of this paper is: 1. To explore why progress has been so slow in the domain of marketing and to evaluate the impact that Expert Systems are likely to have on marketing management. Consequently, technical issues are discussed only briefly. For a full technical explanation of Artificial Intelligence and Expert Systems, readers should refer to the Marketing Science Institute paper on Expert Systems in Marketing4. 2. To discuss the experience and progress of a group of major companies w h o have joined forces to produce an Expert Marketing Planning System, EXMAR, with the author as the principal expert. W H A T A R E E X P E R T S Y S T E M S ? Expert Systems is a branch of what is known as Artificial Intelligence, which is a loosely grouped activity in which a number of researchers of varying backgrounds have done s o m e research since the mid 1950s. But Artificial Intelligence is still not tightly defined. According to Horwitt 5 “Artificial Intelligence is one of the most misunderstood concepts of our time, and little wonder. The fact that very few real- world AI applications exist only serves to feed our wildest sci-fi fantasies. O n e of AI’s major effects, however, has been the spawning of four critical areas of business computer applications research : Natural Languages; Robotics; Visualisation Systems; and Expert Systems.” Conventional computing deals with simple and unambiguous facts with existing packages being little more than moronic number crunchers. Most software is written in the form of an algorithm, which is a list of c o m m a n d s for the computer to carry out in the order prescribed. It uses data held in a separate file, which is stored in a particular way. Thus, software is data plus algorithm and is useful for boring, repetitive, numerical tasks. The largest selling software has been spreadsheets and word processing packages. Database management was developed from this. However, managers handle more than words and numbers. They are concerned about knowledge, which is information interpreted for a particular application. The British Computer Society definition of an ‘Expert System is: “T h e embodiment within a computer of a knowledge based component, from a n expert skill, in such a form that the system can offer intelligent advice or take a n intelligent decision about a processing function. A desirable additional characteristic, which many would consider fundamental, is the capability of the system, o n demand, to justify its o w n line of reasoning in a m a n n e r directly attributable to the enquirer. T h e style adopted to attain these characteristics is rule-based programming.” Put more simply, Expert Systems capture not only the knowledge of a h u m a n expert, but also the rules he uses to reach his conclusions. This knowledge is then m a d e available to others by means of a computer program. The two main components of &r Expert System are: 8 the knowledge Base 8 the Inference Engine - 3 - The rules used by an expert and his knowledge and experience about a certain domain are interrogated and the captured knowledge becomes the Knowledge Base, which is the heart of the system. The Inference Engine accesses the Knowledge Base, makes the necessary connections, draws conclusions, and generates the answers. The general reasoning strategies are separated from the Knowledge Base so as to allow the system to use knowledge in a variety of ways, requesting additional information if required to solve a particular problem and explaining the reasoning behind its questions and recommendations by reporting the rules and facts used. Since the Knowledge Base and Inference Engine are separate, an Inference Engine can be bought to be used in association with other data bases. This is called a shell. A n Expert System will usually have the following characteristics: 8 8 8 8 8 8 8 it will relate to one area of expertise or knowledge rather than to a set of data it will be restricted to a particular topic it will have collected the rules (heuristics) and knowledge of an expert it will have an Inference Engine it will be capable of extension it will be able to cope with uncertainty it will give advice - 4 - 8 it will explain its reasoning. To summarise, the differences between traditional packages and Expert Systems are 6. as follows . Traditional Packanez handles data uses algorithms goes through repetitive processes based on large data bases ExDert Svstems handles knowledge uses heuristics goes through inferential processes based on knowledge bases W H Y H A S P R O G R E S S B E E N S O S L O W I N T H E D O M A I N O F M A R K E T I N G ? During the 196Os, attention w a s focussed on specific problem-solving applications in scientific fields. M a n y successful Expert Systems have been built, including MYCIN for diagnosing infectious diseases’, and P R O S P E C T O R , a system for evaluating geographical locations for possible mineral deposits 8. Management problems, however, do not lend themselves to quite the s a m e precise logic as scientific problems. People do not solve most of life’s problems by mathematical means, but rather by experience, knowledge and intuition. Marketing problems are dealt with in the s a m e way, as most of them are logical rather than mathematical, and problem-solving knowledge, whilst available, is incomplete. Decision-Support Systems and the like use hard facts and static formulae which, given the correct data, provide correct answers. They belong more naturally to the logical, black-or-white, right-or-wrong world of computers. But managers in the world of marketing deal with uncertainties and often with vague concepts. Decisions invariably are built on a set of “rules”, or heuristics, that reflect the expert’s o w n knowledge and experience about the problem in question. These “rules” are hard to nail down and quantify, because the expert’s experience enables him to think in terms of shades of grey, “more or less”, and “approximately”. Such fuzzy reasoning is commonly used by h u m a n beings to find a path through situations that are too complex and amorphous for the h u m a n mind to handle in a totally conscious, rational, scientific way. Most people would acknowledge that in virtually any walk of life, the true expert has built up his expertise largely from experience and an intuitive grasp of problem- solving in the real world, something which is often referred to ‘as the “University of Life”. Indeed, m a n y of the world’s leading business people acknowledge that they o w e their success not to formal business education and text books, but to their o w n experience, flair and intuitive good judgement. Donald Schon9 describes this phenomenon as follows: “Competent practitioners usually know more than they can say. They exhibit a kind of knowing-in-practice, most of which is tacit”. H e cites an investment banker, w h o makes his decisions based on 70 to 80 per cent instinct, and only 20 to 30 per cent calculable rules. This “gut feel” was a major asset to the bank in question. His point is that artistry is not reducible to discernible routines. H e describes scientific rigour as “describable, testable, replicable techniques derived from scientific research, based on knowledge that is testable, consensual, cumulative and convergent”, but then goes on to argue that m u c h of what passes for scientific - 6 - management is irrelevant because business problems do not c o m e well formed. Certainly, most marketing problems are messy and indeterminate and successful practitioners m a k e judgements using criteria and rules which are difficult to define. M a n y academics would decry this as a lack of rigour, and in so doing exclude as non-rigorous m u c h of what successful practitioners actually do. The following quotation from Schon neatly sums up the problems facing not only teachers and researchers of marketing, but, more importantly, the initiators of expert marketing systems : “In the varied topography of professional practice, there is a high, hard ground which overlooks a swamp. O n the high ground, manageable problems lend themselves to solution through the use of research-based theory a n d technique. In the swampy lowlands, problems are messy a n d confused a n d incapable of technical solution. T h e irony of the situation is that the problems of the high ground tend to b e relatively unimportant to society at large, however great their technical interest may be, while in . the swamp lie the problems of greatest h u m a n concern.” The problem to be addressed by Expert Systems in the marketing domain, then, centres around h o w to take account of the intuitive artistry displayed by experts in situations of complexity and uncertainty in a way that is describable and susceptible to a kind of rigour that falls outside the boundaries of technical rationality. The question, then, is h o w an e&temology of practice can be captured and represented in an Expert System. For an Expert System to mimic an Expert, it needs to be able to deal with the uncertainties, complexities, and vague concepts that h u m a n beings deal with routinely, even though such “rules” are neither simple nor straightforward. For example, a simple rule for a marketing manager might be: “If the market is growing, increase promotional expenditure”. This would appear to be easy for a h u m a n being to understand, but in reality words like “market”, “growing”, “increase” and “promotional expenditure” are open to m a n y different interpretations, as indeed is the whole lexicon of marketing. O n e way of dealing with this problem is the development of fuzzy sets. A “growing market”, for example, is a fuzzy set in the sense that its meaning can vary from situation to situation. Fuzzy numbers approximate the response figures from marketing experts and these numbers are then loaded into, for example, sales projections and promotion analyses. The foundation of any Expert System is the Knowledge Base, which can be extracted from one or more experts in a particular field. The expertise is usually stored in the form of rules of thumb (heuristics), which are, typically “If then” statements. For example, if A is true, then B is true; or if X is true, do Y. Given an initial set of circumstances, the system can m a p out a set of contingencies and further -- contingencies. A heuristic differs from an algorithm in that it does not give a correct answer, nor does it guarantee results. It merely suggests a general direction that is more or less likely to prove more useful than another direction. A n example of a heuristic in chess might be: “If a player stays in control of the centre of the board, he is more likely to win”. In marketing, a heuristic might be: “if the market is growing and if you have appropriate business strengths, then an appropriate marketing objective would be to grow market share”. A system of interlinking heuristics in the form of a decision tree is one way of representing knowledge. These are sometimes “backwards inferencing” and sometimes “forward inferencing”. Backwards inferencing starts with an objective and tries different combinations of rules and/or actions until it is reached. Forward inferencing reasons from initial information until it reaches useful conclusions. This can give rise to what is termed “combinatorial explosion”, which can be avoided by pruning and the use of heuristics which are correct most of the time. This gives probable solutions to less rigorously defined problems that are too complex to be dealt with algorithmically. To date, however, no one has seriously tackled the world of marketing with Expert Systems other than the MSI M X X D ~ system developed to advise on advertising design. After considering a variety of consumer and environmental factors, advertisers use a combination of empirical research, communication theory, and rules of thumb, to select communication objectives and select appropriate creative approaches. The authors themselves list a number of weaknesses in A D C A D , but conclude: “As one advertising executive put it: “it helps us to think a little deeper about the issues w e have to consider in developing ads that are both strategically a n d executionally sound”. Another interesting and relevant conclusion was that most managers, when asked, said they would like to m a k e use of existing theoretical and empirical knowledge of marketing when making decisions. However, few actually did use such knowledge. Expert Systems can bridge this gap by structuring, validating and disseminating marketing knowledge, whilst at a theoretical level, they challenge their creators to understand and critically evaluate the elements of marketing knowledge and their interrelationships. A C A S E HISTORY O F T H E D E V E L O P M E N T O F A N E X P E R T S Y S T E M IN M A R K E T I N G PLANNING During the 198Os, Japanese activity in the field of Expert Systems prompted the E E C to give birth to the ESPRIT programme in an attempt to integrate European efforts. This in turn led to the DTI sponsored ALVI programmes. A n outcrop of these is a n e w DTI-sponsored club by the n a m e of EXMAR, which set out in 1987 to produce an Expert System in the domain of marketing planning, inviting the author of this paper to be the principal expert. The ten founder m e m b e r companies include s o m e of Britain’s biggest and most successful multinational corporations spanning capital goods, industrial goods, consumer goods, and service industries. After almost two years of work and an expenditure of over f) million, all there is to show is a demonstrator model on a Xerox 1186 workstation which exemplifies the scope of the Expert System using a case study specially written for the club by the author of this paper. The purpose of this part of this paper is the explain h o w E X M A R has developed, what obstacles were encountered along the way, h o w these were overcome, and what problems still remain to be solved before a commercially usable P C based system can be m a d e available. The first point to be m a d e is that Expert Systems do play a vital role in the accumulation, synthesis and understanding of the constructs of marketing and their interrelationships. M a n y of the theories, illuminative sights, empirical research findings, models, and experience, are scattered around in books, libraries and inside the heads of both practitioners and academics. They remain, therefore, largely unavailable to most marketing managers, and indeed to most marketing academics. The synthesis of such knowledge in a particular domain into Expert Systems not only benefits those whose task it is to develop the system, by forcing them to turn their knowledge and expertise into actionable marketing propositions, but also those responsible for marketing decisions by making it available where it is likely to have the greatest impact. P R O B L E M S SUITABLE F O R EXPERT S Y S T E M S In deciding whether marketing planning was a sensible domain for the application of Expert Systems methodology, the M S 1 checklist4 proves useful. Four criteria are provided: Are the key relationships in the domain logical rather than arithmetical ? Since the decision area is knowledge-intensive, the answer here is “yes”. Is the problem domain semi structured rather than structured or unstructured? Well-structured problems can use more conventional procedures, but since the marketing planning process is only semi-structured, the answer is “yes”. Is knowledge in the domain incomplete ? Since marketing planning and all its contextual problems remains one of the most under-researched areas of marketing, and since little has been published about the interrelationships of all the techniques of marketing in systems design, the answer is “yes”. This is in fact the key to the whole project and why it was chosen in the first place by the club members. n Will problem solving in the domain require a direct interface between the manager and the computer system ? The intention is to have operational marketing managers using the system for the production of marketing plans, so the answer is “yes”. Marketing Planning remains one of the last bastions of ignorance in the field of marketing. The benefits of marketing planning are‘ well documented and agreed,l’ yet so complicated is the process of marketing planning, and so confusing are the interrelationships between the tools and techniques of marketing planning 11 , that very few British companies enjoy these benefits, as has been shown by a seminal paper by Greenley 12 that reviewed all the major U K empirical research in this area. Indeed, there were as many dysfunctional results from the attempts of companies to initiate marketing planning procedures as there were benefits. The whole thrust of the project, then, was to tackle this problem by means of an Expert Marketing Planning System codenamed EntAR. Marketing planning can be defined as a logical sequence and a series of activities leading to the setting of marketing objectives and the formulation of plans for achieving them. The model taken to represent the marketing planning process was the author’s nine 10 stage breakdown , as given in Figure 1 later in this paper. - 12 - ANALYSIS P H A S E The initial requirements analysis produced a number of interesting problems for the project, which were to s o w the seeds of expensive and time-consuming delay. These problems can be summarised as follows:- (i) it b e c a m e clear that not m a n y of the m e m b e r companies were particularly au fait with the methodology of marketing planning. This led to the problem of setting clear objectives for the project. (ii) the diversity of c o m p a n y industry types, ranging from capital goods to service industries, meant that no subsequent system could possibly be suitable for all circumstances. (iii) problems and subsequent proposed objectives ranged from .“To support a formal planning framework to improve discipline during the planning process” and “To support further understanding of the effects of currency fluctuations” to “To promote discipline in pricing control” For these reasons, it w a s decided to focus on the process of marketing planning itself rather than on any situation-specific system. M E T H O D O L O G Y A firm of software consultants w a s appointed project manager and a knowledge based systems house w a s appointed principal contractor. Considerable confusion surrounded the proposed delivery system with the result that specifications, such as model, functional requirements, system structure, information requirements, enhancements, consequences, knowledge base specification, validation procedures, and so on were never produced. The systems house began a series of twelve half day interviews with the author of this paper in order to develop the Knowledge Base. Unfortunately, although taped and transcribed, they were largely unfocussed due to the inexperience of the interviewers and little progress was m a d e towards formal modelling of the marketing planning process, in spite of very specific guidance given by the author to the interviewers. The problem, centred around lack of proper project control by the project managers, confused expectations by members of the club based on marketing planning naivety, the inexperience of the knowledge engineers, and the passive role of the domain expert, which was necessary in view of the nature of the project. Several attempts on the author’s part to guide the system were brushed aside as politically inexpedient. The result was that the paper outlining the tasks to be performed by the computer system targeted the whole marketing planning process rather than any subset, and because of this breadth, the process to be computerised was not documented in any detail, nor backed up by any substantive models and interrelationships. N E W K N O W L E D G E ENGINEERS APPOINTED At this point, the problems began to assume crisis proportions, and the project manager appointed n e w knowledge engineers to take over the feasibility study and the delivery system. - 14 - The n e w contractor set about finding s o m e c o m m o n requirements a m o n g end users in order to outline the domain model, with a boundary definition showing which parts of the model would be tackled by the computer system. They set about establishing the following areas: \ 8 scope 8 constraints 8 organisational impact 8 maintainability 8 extensibility 8 technology 8 time scales 8 risk and cost versus quantifiable benefits For the first time the E X M A R project was beginning to focus on building a system for appropriate problems that were valued, bounded and routine. The following emerged as the final overview of the objectives of E X M A R as agreed by all members of the club. W H A T W I L L E X M A R D O ? E X M A R is intended to be a Marketing Planner’s Assistant. It will guide a user through the marketing planning process, offering advice at key stages, controlling data input and presenting data in various ways so as to assist in the setting of objectives and strategies. - IS - The full Marketing Planning Process has nine stages, with various feedback loops, as shown in Figure 1. Fbure 1 -THE MARKETINd PLANNING P R O C E S S 1. Corporate Objectives 2. Marketing Audit 3. S W O T Analysis 4. Assumptions A . Marketing Objectives & Strategies LtEstimate Expected Results 7. Identify Alternative Plans & Mixes 8. Programmes 9. Measurement & Review The current vision of E X M A R concentrates on stages 2, 3 and 5 because club members have consistently identified stage 5 (objectives and strategy setting), together with the preceding data collection and analysis, as the main problem areas. Corporate Objectives and Mission Statement are taken as the given inputs (from outside the user’s influence) needed to start the process. All relevant data is then collected in a Marketing Audit phase. This data is then abstracted and analysed in the S W O T phase and relevant assumptions recorded. Various methods are then available in the final phase to assist the user to set realistic and consistent objectives, together with coherent strategies to meet them. - 16 - It is anticipated that E X M A R sessions will be highly interactive and iterative, encouraging scenario planning. They should also permit analysis at different levels of detail, from a corporate overview of key business sectors to in- depth studies of individual market segments. Further details of h o w members believed E X M A R would actually do this, together with implementation and development constraints, are included as Appendix 1. From this it will be seen that members wanted IBM pc compatibility for hardware, with software amenable to change by programmers not involved in its development, and which would be amenable to extension and add-ons. T H E N E X T S T E P A number of refinements and corrections to the methodologies and interrelationships was n o w necessary before the project could proceed. These were detailed by the author in a separate document, the relevant part of which is reproduced in Appendix 2 to this paper. D E V E L O P M E N T O F A D E M O N S T R A T I O N M O D E L S o m e further interviews with the knowledge engineers quickly moved the project towards the production of s o m e deliverables. It was possible, for example, to define those parts of the marketing planning process which seemed the most likely candidates for automated support. The agreed primary objective of E X M A R was to provide automated assistance for the marketing planning process, since it had been agreed a m o n g members that in general marketing decisions are taken without sufficient analysis and understanding of the relevant issues. The reason was seen as being a lack of knowledge and understanding of h o w and why the multifarious factors of marketing interact and serve to form the parameters of any business activity. In this real life situation w e see emerging the perfect role for an Expert System in marketing planning. D E M O N S T R A T I O N M O D E L All that remained n o w was to produce a model to demonstrate h o w such an Expert System would work. For this, the author wrote a special case study based on a multinational company in the bearings industry. The case study contained all the necessary features to demonstrate the scope, methodologies and outputs of the proposed Expert System. A detailed report was a necessary prerequisite for producing a live demonstration model. The report actually produced outlines the scope and functional breakdown, the data model, and the technique interrelationships. This report is included as Appendix 3. It is recommended that this should be carefully studied, as it describes the basic model and outlines the technique interrelationships. Not included in this paper are other parts of the report relating to technique descriptions, model testing and technical details relating to the demonstration itself. From this it will be seen that the 9 step model shown in Figure 1 was m a d e more amenable to computerisation, as shown in Figure 2. A n example of the detail included in one of these stages is shown in Figure 3. The basic Data Model used and s o m e of the techniques relating to it are shown in Figure 4. - 18 - F U R T H E R R E F I N E M E N T O F S C O P E Figure 2 Various related areas are outside EXMAR'S scope, on the grounds that, though important, they are peripheral to the central concerns of EXhUR, and should not be studied in detail in the interests of timely focus. These areas are summarised in the boxes on the diagram below outside the “scoping” dotted line. Brief notes on these follow. I I I I I I I I I I I I I I I I I I I I I I I I I I I I I i I I I I I I S E L E C T / D E F I N E B U S i N E S S UNIT I D E F I N E UNIT M I S S t O N I I I I I I I 1 I I C O N O U C T AUDIT 1 S U M M A R I S E I I I I S E T O B J E C T I V E S -[SET] I I I 1 POSiTlONING I I m--1-1-------- ----------m- 1 ---1--------------------- : . 1 I I P R O D U C E S T R A T E G I C ; M A R K E T I N G P L A N F O R A I I I I B U S I N E S S UNIT I S E G M E N T A T I O N Conduct Audit Figure 3 C O N D U C T AUDIT .- F A C T O R S .- .B CHECKLIST * / A S S E S S CRITICAL - S T R E N G T H S S U C C E S S F A C T O R S T A B L E I A N D W E A K N E S S E S I A S S E S S O P P O R T U N I T I E S AND- T H R E A T S c L .w Of .- .D C H E C K L I S T * O T LIST A S S E S S M A R K E T -~ A T T R A C T I V E N E S S L ‘+=j A~~tk: czg The objective is to assess the state and prospects of the products and markets already identified. Information needed at this point m a y have been collected in advance of the planning process, or it m a y be collected now. - 2 0 - Techniaue Interrelationshins DATA USED BY TECHNIQUES Figure 4 The diagram below shows the data used as input by some of the techniques modelled. MARKET MAET ;;;; mj _-_. SECh IA-*- ,.~~~~:~~~ \ \ (actual orpote$aI) 0 Oefmmon compe titer) l Planning G l name OR . nnaA ‘,.l‘illrA ’ i l differentiatior/ (l-10) . . / PRODUCT FOR l market share’ -. I ‘). .costs --__ I -. 0 score on CSFi,- ;-o~a.U&re 0 newlexisti ng . , --S_ 0 price/average priq ’ I ’ I ’ I ’ I ’ , ’ , l Relation to 0 sales volume ‘. *. I ‘\ I , -. , -. / .‘- [-+q is sold in to .’ PRODUCT I I 1 l name: i I l costs’ : D newlexistind (self or compe tr tar) J Production of a Demonstration Model At a packed meeting of the members in December 1988, the demonstrator model was unveiled. Its purpose was: 8 to demonstrate how such a system would meet the club’s primary objectives; 8 to provide evidence of the feasibility of building such an Expert System in technical terms; 8 to provide a basis for feedback about the systems’s utility. It was developed on a Xerox 1186 workstation running the Interlisp environment to minimise the time required to build the demonstrator and because of Interlisp’s power and maturity. The demonstrator provided: 8 guidance and support for the marketing planning process at various stages and help in managing the interactions; 8 variable forms of information presentation and manipulation, such as data forms, diagrams and text. Relationships and constraints between information are managed by the system, for example by calculation and iconic cross- references; 8 a free interface which allows the user to take the initiative in determining precisely what he wants to do next, and what he wishes to have displayed to assist his actions. This is done by the provision of a number of means of navigating around the window-based system. - 22 - The demonstrator model was spectacularly successful with club members and clearly illustrated the large amount of iteration that would need to occur in generating a plan. It also gave s o m e indication of the processes of information gathering and debate that would typically have to occur in the real world whilst using the system. Conclusion Although the actual demonstration model using the case study is not included in this paper, for reasons both of confidentiality and brevity, the E X M A R project has clearly reached a stage of development that demonstrates the value of Expert Systems in marketing. A number of conclusions can be drawn: (9 The development of E X M A R shows that it is possible to use Expert Systems methodologies to build support systems in complex areas of marketing management, especially if the domain is well defined, has a large number of factors to be considered and relevant expert knowledge is available. (ii) The more complex and amorphous the expertise to be captured, the longer it takes both the expert and the knowledge engineer to reach an acceptable approximation. It is clear that to develop an Expert System that is of s o m e practical use requires both time and resources of massive proportions. This is supported by the M S 1 research paper4, which concludes: “There are no shortcuts to building a good Expert System. It takes a considerable amount of skill, patience, and years of effort to develop an Expert System in a n e w area and get it into the field”. - 23 - (iii) Expert Systems provide a consistency to h u m a n decision making which is valuable, since people tend to forget or ignore knowledge. (iv> E X M A R has generated considerable interest and support a m o n g the major multinational companies that form the club, because it forces them to think deeply and in a structured way about the issues that need to be considered in developing a strategic marketing plan. (VI Expert Systems are useful in helping both academics and practitioners to structure, validate, and use marketing knowledge and to better understand the interrelationships between the elements of marketing. (4 Tight project control is vital. This view is supported by Mumford13. In particular, the following issues need to be considered: 6) Subject matter - h o w well it is defined ? - is it likely to change during the project’s life ? - can adequate inputs be provided by both experts and.knowledge engineers ? (ii) T h e User - do they understand the likely time of the project ? - do they know exactly what they want ? - are they willing to work constructively to solve problems ? (iii) Time - are the project deadlines realistic and achievable ? - 24 - (iv) Resources - is the budget sufficient ? - is sufficient skilled human resource available ? - will facilities requirements be catered for ? (v> Project Management - is the project management strong enough and sufficiently disciplined? (vii) The potential advantages of Expert Systems are: n consistent advice a secure knowledge bases n making better use of experts n enhanced decision making n improved analysis (viii) The stages in building an Expert System are: (8 problem identification and definition (ii) the acquisition of relevant knowledge . (iii) the representation of relevant knowledge (iv) the selection of a reasoning approach (VI system selection (vi) prototype development (vii) system refinement and validation. (ix> Since we live in an imperfect world, with imperfect problems and imperfect tools, it is unreasonable to expect a perfect Expert System until there are perfect experts and perfect technology. On the other hand, if an Expert System gives better advice than you would have had without it, it is probably worthwhile. In conclusion, it is unlikely that Expert Systems will ever be able to give the s a m e value as real h u m a n experts, although clearly they can offer reasonable advice. Nor will they guarantee that you m a k e the right decisions. But they can help you gain a proper perspective of the alternatives. In a sense, Expert Systems will always be a bit like Distance Learning programmes, which can replace a bad teacher, but never a good one. R E F E R E N C E S 1. 2. 3. 4. 5. Horwitt E. “Exploring Expert Systems” Business Computer Systems, March, 1985 6. Bryant N. “Managing Expert Systems” John Wiley and Sons, 1988 7. Buchan B. and Shortliffe E. “Rule-based Expert Programs : the MYCIN Experiments of the Stanford Heuristic Programming Project, Reading, MA., Addison- Wesley, 1984 8. 9. 10. 11. 12. 13. Moutinho L. and Paton R. “Expert Systems : a N e w Tool in Marketing” Quarterly Review of Marketing, Summer, 1988 Foster E. “Artificial Intelligence” Personal Computing, April, 1985 Cebrzyask “Artifical Intelligence : the goal is to store an expert’s real knowledge on a disk” Marketing New, Vol. 21, No. 5, 1987 Rangaswamy A., Burke R., Wind J. and Eliashberg J. “Expert Systems for Marketing” Marketing Science Institute Working Paper Report, Nos. 87- 107, 1988 Duda R., Gaschnig J. and Hart P. “Model Design in the Prospector Consultant System for Mineral Exploration” in “Expert Systems in the Micro- electronic Age” ed. Michie D., Edinburgh: Edinburgh University Press, 197’: Schon D. “The Crisis of Professional Knowledge and the Pursuit of an Epistemology of Practice”, Research Paper for the Harvard Business School 75th Anniversary Colloquim on Teaching by the Case Method, April, 1984 McDonald M. “The Theory and Practice of Marketing Planning for Industrial Goods in International Markets”, Cranfield Institute of Technology, PhD., 1984 Leppard J. ‘“Marketing Planning and Corporate Culture - a Conceptual Framework which examines Attitudes in the Context of Marketing Planning” Cranfield Institute of Technology, M.Phil., 1987 Greenley G. “A n Exposition into Empirical Research into Marketing Planning”, Journal of Marketing Management, 3.l., July, 1987 Mumford E. “Designing Computer Based Systems”, University of Wales Review Business and Economics, 3, 1988 APPENDIX 1 H O W WILL E X M A R ACHIEVE ITS OBJECTIVES ? Firstly the system will prompt the user for information to define the business unit to be analysed. This will include a basic definition, mission statement and top level objectives. The user will then be asked to specify the market segments and products to be analysed (the system uses a simple Pareto 80/20 rule to help the user to focus on the most important business areas). The result is a comprehensive list of existing and potential product-in-market combinations which can be organised in the form of an Ansoff Matrix (Figure 2). Figure 2 ,4NSOFF MATRIX Increasing Technological N e w n e s s D P R O D U C T S Existing Potential M Existing M A R K E T P R O D U C T A P R E S E N T A T I O N D E V E L O P M E N T R K E T M A R K E T DIVERSIFI- S Potential E X T E N S I O N CATION The system will ask the user to order the products and markets by h o w n e w they are to the business unit’s existing area of operation. - .28 - At this (audit) stage the Ansoff Matrix is used to drive the data collection process. It can be used in later stages to a s s e s s strategic direction b y reference to the classification of product/market combinations in each box (eg. Market Extension versus Product Development). O n c e a complete list of Markets and Products has been established the s y s t e m will prompt for k e y information required for the S W O T analysis and later stages. For each market s e g m e n t the user m u s t supply two sets of factors: one to m e a s u r e the attractiveness of the market to the business unit; the other to m e a s u r e h o w a product m a y be evaluated b y that market. T h e s e are k n o w n a s Market Attractiveness Factors and Critical S u c c e s s Factors ( M A F s and CSFs). E a c h product m u s t then be evaluated for each relevant market s e g m e n t b y inputting scores against the C S F s previously specified, to a s s e s s business strength (relative to the competition). This has to be done for both the current position and the forecast position. Forecasts are also required of performance level. T h e S W O T stage also requires information on opportunities and threats, in terms of their impact and likelihood. T h e s e can be s u m m a r i s e d in an Impact/Urgency matrix and referenced at later stages, where the user is reminded of threats of high impact and likelihood w h e n setting strategies. A large a m o u n t of analysis is necessary to support the S W O T stage and E X M A R will be able to a c c e s s other packages for supplementary analyses. A s s u m p t i o n s are input a s part of the forecast and once the S W O T is complete the user can proceed to set objectives and strategies. - 9a - Objective setting is driven by the concept of G a p Analysis which portrays the target level of performance and a ‘status quo’ forecast figure. The forecast is obtained from summarising the performance level of all Product/Markets in the top left hand corner of the Ansoff Matrix. The user attempts to close the gap by: a) selecting existing product/markets and improving operational performance b) selecting n e w product/markets for inclusion in the portfolio. The key aid for the user in this process is the Directional Policy Matrix (Fig. 3) which essentially summarises a large amount of the S W O T analysis. 3 Figure DIRECTIONAL POLICY MATRIX Business Strengths (CSF Scores) H I G H M E D L O W Market Attractiveness (MAF Scores) H I G H M E D Size of circle shows performance level This is a very versatile tool which visually displays a large number of the measurable criteria relevant to the selection decision, including performance (size of circle) and potential for improvement (position on the matrix). The user will be able to m o v e circles to more favourable positions on the matrix to represent changes in objectives and such movement is recorded by the system. Later, the user will be prompted for strategies to achieve the movement. The result of this stage will be a set of revised product-in-market objectives with associated strategies. At any point the evolving strategy can be evaluated and the system will produce reports and various displays to assist this evaluation. For example, portfolio balance can be evaluated using the Directional Policy Matrix itself, plus other tools including the Ansoff and I Boston Matrices. This implies considerable feedback from strategy to objectives setting and the system will document reasons for changes. This phase is potentially very rich in expertise, and research into ways of capturing this is continuing within the club. W H Y D O W E W A N T E X M A R ? The System will provide:- I 1) an automated implementation of a rigorous marketing planning process. Historically the process has been difficult to implement rigorously. 2) a comprehensive statement of the data requirements of the marketing planning process, with particular emphasis on the quantification of previously nebulous concepts such as business strengths. - 31 - 3) powerful visual displays of key information. These aid understanding and communication. They also free the user to concentrate on other expertise-rich concepts such as coherence and consistency of strategy. 4) an opportunity to build a hierarchical structure of plans, from business overview to detailed product analysis. This will depend on the quality of implementation. The benefits of the above, in terms of the quality of plans (and of the debate during their construction) are similar to those claimed by formalised marketing planning. N o existing software approaches the functionality of that envisaged by EXMAR. JMPLEMENTATIONCONSIDERATIONS Oreanisation Organisationally, E X M A R simply requires the existence of a marketing manager to use the system. Naturally such a user will need access to the required data, s o m e of which m a y not be available immediately. O n e of the spin-off benefits of E X M A R m a y be to act as a catalyst to prompt change both organisationally and in the data collected. D E V E L O P M E N T C O N S T R A I N T S The club has consistently specified I B M P C compatibility for hardware and this has not changed. Software is a more flexible issue but the following are requirements which should be borne in mind. a) Club m e m b e r s are very likely to want to develop and customise their copy of the system. This implies:- 1) Software which is amenable to change by programmers not involved in its development. 2) A preference for software which has a wide user base, particularly a m o n g club members. 3) S o m e level of system documentation. b) Expertise is likely to be gained in using E X M A R over time, which will generate a need to ‘build-in’ further levels of expertise. Thus the software should be amenable to extension of the expert system aspects, implying a rule-based or list processing capability. c) The software should have adequate data communications for access to and from other packages. This requirement also indicates a likely future requirement for a multi-tasking environment. - 33 - d) S o m e club members have been conditioned to expect Goldworks to be the chosen software and m a y already have committed themselves to this package. APPENDIX 2 Refinements a n d corrections to the methodoloeies a n d interrelationshiDs outlined in a letter to the club workine bv Professor Malcolm McDonald 1. T h e Directional Policv Matrix 6) It’s O K to have nine boxes, which is in any case the w a y it w a s originally conceived. I personally keep it to four because it is conceptually easier and fits more comfortably into what “students” have b e c o m e used to via Ansoff, Boston, Porter, et al. Nonetheless, the nine box matrix does provide more options and greater flexibility. C a n I suggest that, rather than confusing users at the construction stage with a nine box matrix, w e only put the lines in after the calculations have been completed. W e must, then, ensure that the dividing points are 33.3 (or 3.3) along each axis. (ii) It is imperative that you do not try to use profit as a measure of circle diameter. Take it from me, every time this measure is .used, it distorts the truth. For example, there m a y be a product or market that accounts for, say, 50 per cent of sales value, but 10 per cent 0f‘“profits”. This would appear as a small circle, so masking its true use of resources. In any case, profit is an accounting notion which depends on an arbitrary allocation of overheads. There is also the tricky question of whether it is products or customers that determine profitability. It is usually the latter, which is rarely catered for in accounting systems. In any case, profit is almost certain to be strongly reflected in the Marketing Attractiveness criteria. (iii) W e should m a k e it clear that there are m a n y different levels of analysis. This could involve any of the following:- . . - Regions (of the world) - Countries - Areas (of countries) - Companies - Strategic Business Units - Divisions - Product Groups often synonymous with markets - Products - Segments - Customers - Distributors/Agents/Wholesalers Etc. Each one of these can be further sub analysed, if necessary. (iv> W e should ensure that users are m a d e aware of the pitfalls, which are as follows: (a> Users must beware of becoming emotionally involved in their o w n interpretation of “attractiveness”, which often leads to them “fiddling” the system to ensure their business comes out in the yoper quadrants. It is clearly illogical (or at least unusual), if everything is seen as highly attractive. In such a case, either all are equally attractive, or the scoring is wrong. The scale 0 - 10 is meant to represent relative attractiveness according to their o w n criteria, so that something that is near to 0 is nothing like as. attractive to the company as something that comes out at, say 9.5. But is does not necessarilv m e a n it is unattractive. To m a k e this effective, perhaps w e should put in a suggestion that users might think in terms of “potential” if they feel (after being given due “warning” of the pitfalls) that the word “attractiveness” might cause problems. O n e other warning. Users m be prepared to score 0, where appropriate. W e might even put in a proposal that, if appropriate, the scale might have negative values to go along with a negative scoring system. This might be appropriate where there are very wide extremes. lb) O n e final point on this. Recently, I had a case of a company which was experiencing decline in u its segments. In this case, “attractiveness” hardly seemed like the appropriate description. So, instead w e used “potential”, since s o m e divisions had greater potential for growing sales and profits than others. For example, the “shipping” market was in decline, and the company had a high market share. In the “food” market, on the other hand, (also in decline), the company had a m u c h smaller market share, so the potential for taking market share (and improving profit), was greater, hence it appeared in the upper quadrant. Not surprisingly, “Food” also appeared on the right of the horizontal axis. The point is that had w e not used this device, everything would have appeared in the bottom part of the matrix. Whilst this is obviously a possibility, it would not have been particularly helpful in this somewhat sad case. Whether the total picture in this case is acceptable or not is irrelevant. The truth is that this company has diversified into other unrelated business areas that are m u c h more attractive. H a d these newer S B U s been included in the analysis, then clearly even “Food” would have appeared as low in attractiveness. (VI A propos the two situations (t-3 to t.0 and t.0 to t+3) for the vertical axis, I must stress m y strong reservation that his will almost certainly confuse most users, and might even irritate them having to do it twice. Nonetheless, it’s perfectly logical, consistent and feasible. What w e must stress, however, is that the first part of the exercise must be for t-3 to t.0 and u reflect what has happened historicallv. The m part of the exercise, t.0 to t+3, is a forecast of attractiveness and must reflect their view of what will happen over the next three years. - 2. Ouantifvinn Opportunities a n d Threats First of all, let’s consider the following generalised list of macro and micro factors which might be relevant: demographic economic technological macro political legal social/cultural customers competitors distribution channels suppliers potential competitors . A m u c h more detailed checklist will be provided with the actual system. These might be considered to be either Qpoortunitia or threats. I suggest w e provide a matrix (similar to the ISSUES MATRIX) for each, ie. two matrices, one an O P P O R T U N I T Y MATRIX, the other a THREATMATRIX. W e could m a k e it work as follows:- (a) List Threats (no more than ten) (b) Probability of occurrence (within t.0 to t+3) (.05 to .95) (c) Impact on the organisation (score 1 to 10) The matrix would look as follows: IMPACT 10. 6 3 I . 95 - 1 - ---.,---- - ----------_ (3 . 6 1 I I PROBABILITY I op I O C C U R R E N C E - 3 , ’ I I . 05 1 It can be seen here that Threat 1, say something specific to do with C A P ( C o m m o n Agricultural Policy), will have a big impact on the organisation (score 8), and that there is a high probability that it will happen (probability .8). All Threats can be plotted using this methodology, which would need s o m e guidelines similar to those provided in m y ISSUES PRIORITY MATRIX. The whole process can then be repeated for Qrmortunities. 3. Strategies arising out of Directional Poiicv Matrix a n a h + If w e get users to predict the scores of Critical Success Factors, then clearly they will need to convert these into strategies. W e must be careful not to lead them too m u c h “by the nose”, and I suggest w e don’t need to go beyond the overall guidelines suggested by the Shell Directional Policy Matrix. Clearly, these must be converted into 4 x Ps jargon by the user, but it would be a gigantic task to attempt to list all possible combinations of marketing mix strategies. Also, the gverall objectives for each product/market should be consistent with the guidelines suggested by the Porter Matrix a n d by Life Cyc!e Analysis. L. g i. e i’l S”$ ;*a. 1. y. li i - 1.n - 4. Market Life Cvcle ‘. W e must be careful here. There is no general recognition iri’*’ ” I’ I”- theory of anything called a “Market Life Cycle”. I have sent under separate cover a detailed explanation of what I mean. But yes, of course the guidelines can be included to help users select appropriate strategies. . A P P E N D I X 3 ScoDe. Functional Breakdown. Data Model a n d Techniaue Interrelations of Exmar Scone a n d Functional Breakdown This section defines the functions performed during the relevant stages of the Marketing Planning process, and in s o m e cases breaks down the functions into simpler functions. The functions are related to techniques and methods used in carrying out the function, and to deliverables that form part of the Marketing Plan, as defined by McDonald. The top level breakdown is used to refine further the scope beyond the definition contained in Rl Initial Findings Report. N O T A T I O N O F F U N C T I O N A L B R E A K D O W N D I A G R A M S The diagram beiow summarises the notation used in the Functional Breakdown diagrams. Function boxes represent tasks to be performed as a step towards production of a marketing plan. Technique/method boxes have icons that illustrate the style of representation used by the techniques or method. d F ”N C T lON j .~~~l.-:“i”““““I:-- I /,,,I . I “m a y mvolve feedback co” “Deliverable” T O D Level Breakdown and ScoDinp SCOPING DEFINED BY INITIAL FINDINGS REPORT The Initial Findings Report defined the scope of EXMAR Phase 2 as being the marketing audit, S W O T analysis and objectives and strategies stages of McDonald’s 9 stage breakdown of the Marketing Planning process. This is taken as the starting point for this section. The diagram below summaries this. Il. Coy orate objectives 1 --- + ,,-----------\ \ , \ ‘\, 12. Mafketing audit ( ‘,, The Marketing Planning Process ‘1, 13. S W Y T analysis ( ‘,, \ \ \ \ \ \ \ , 4. Assumptions I I \ \ \ \ \ \ \ \ \ , \ \ I . ‘\ 5. Marketing objectives and strategies ‘N \ \ \ L.,,-e-. ----a------------ A input to the model. 6. Estimate expected results \ ) 7. Identify alternative plans and mixes 1 18. Programmes 1 9. Measurement and review Objectives are set subject to certain assumptions: other than this, little formalism has yet emerged with regard to assumptions. The setting of corporate objectives is outside Phase 2 scope. So any information from the corporative objectives required is regarded as an T O P L E V E L F U N C T I O N A L B R E A K D O W N It is useful to produce a slightly differing top level breakdown than that contained in McDonald’s g-stage diagram. This is given below. Explanatory notes follow. P R O D U C E S T R A T E G I C M A R K E T I N G P L A N F O R A B U S I N E S S UNIT S E L E C T / D E F i N E B U S I N E S S UNIT D E F I N E UNIT M I S S I O N C O N D U C T AUDIT 1-j S U M M A R I S E 1 -1 S E T S T R A T E G Y 1 - 44 - Produce Strategic Marketing Plan for a Business Unit This describes the task being modelled. .It is strategic because it involves ignoring some details to aid clearer thinking about the important parts of a business. It is for a business unit because this process can be carried out at any level of an organisation, or for a subset of the business that crosses organisational boundaries. Select/Define Business Unit Identify which area of the business the marketing plan is for. Define Unit Mission Define what the business unit is in existence to achieve. Focus Identify which of the unit’s products and markets are of interest. Conduct Audit Assess the products and markets identified in Focus stage. Summarise Summarise the products in the business unit in a form suitable as a starting point for the setting of objectives. Set Objectives Set objectives for the business unit based on the information collected, analysed and summarised. Set Strategy Define strategy by which the objectives are to be met. F U R T H V R R E F I N E M E N T O F scope Various related areas are outside E X M A R ’S scope, on the grounds that, though important, they are peripheral to the central concerns of EXMAR, and should not be studied in detail in the interests of timely focus. These areas are summarised in the boxes on the diagram below outside the “scoping” dotted line. Brief notes on these follow. r ------------------------- 1 I I I I I I I I I I I 1 I P R O D U C E S T R A T E G I C I M A R K E T I N G P L A N F O R A f B U S I N E S S UNIT I I I I I S E L E C T / D E F I N E B U S I N E S S UNIT I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I ----I D E F I N E UNIT M I S S I O N [piiF=] C O N D U C T AUOIT S U M M A R I S E S E T O B J E C T I V E S 1 S E T S T R A T E G Y 1 ---------we------ i I I I I I I I I I I I I I I I I i I I I I I I 1 W - W - 1 C O R P O R A T E O B J E C T I V E S E n l N G ORGANIS- ATION D I A G N O S I S 1 1 M A R K E T R E S E A R C H M A R K E T S E G M E N T A T I O N I P R O O U C T POSITIONING L I Organisation Diagnosis Such issues as diagnosis of the health of an organisation, Blake- Mouton Matrix etc. Corporate Objective Setting The means by which corporate objectives are arrived at is not within EXMAR'S scope. Where corporate objectives (or business unit objectives derived from them) are required by later parts of the marketing process, they are regarded as an input to the model. Market Research, market segmentation, product positioning Such techniques as research into the needs or wants of customers, positioning products within markets by finding criteria with which to map the market, and related market segmentation techniques are not covered. The results of, market segmentation are important to the functions modelled, so this is essentially an input to the model, though some assistance may be offered. TECHNIQUESCONSIDERED Porter Matrix Critical Success Factors table Directional Policy Matrix . Ansoff Matrix Boston Matrix Product Life Cycle Gap Analysis Objectives Typology Threat Assessment Market Attractiveness Table Cost Experience Curve Porter 5-Force Model Downside Risk Assessment T E C H N I Q U E S LEFT O U T Opportunity Matrix Product Positioning M a p Customer Preference M a p Market Segmentation M a p Diffusion of Innovation Blake/Mouton Matrix Organisation Diagnosis McDonald Productivity Matrix Size/Diversity Graph (part of organisation diagnosis) Market Segmentation Studies - detail to investigate Financial S u m m a r y - both part of Marketing Audit Response Elasticities Second Level Functional Breakdown PEFINE UNIT MISSION BUSiNESS UNIT DEFfNfTfON This involves definition of what the unit is for, including any financial targets. This will be a corporate mission statement if the whole organisation is being considered. Otherwise it will identify the specific role of the unit within the organisation. SELECT/DEFINE BUSINESS UNIT r * FiNANClAL DEFINE UNIT . - SUMMARY MISSION WSiNESS DEFINI- * TION / UNIT MISSION This involves defining which business unit the plan is for. Where a plan is being produced for an organisational unit, this simply involves identifying the unit. But it may be .more complex: one may wish to carry out the plan just for a subset of an organisational unit’s business of particular interest, or for an area of the business that crosses organisational boundaries. For example, a plan for tinned foods within a foods company may cross department boundaries of design, production, finance, etc. - 49 - I The output is a definition of the business unit, including a title that can be used to head all documents associated with the plan. It m a y be possible to produce a checklist to assist in this function. Financial Summary Any financial targets set for the unit, particularly for revenue or profit. This also involves specification of the planning period to the end of which the targets relate (typically 3 years). Business Definition/Unit Mission Statement A statement in words to cover aspects of the mission not covered by the Financial Summary. Brief statements should be made which cover the followings points: 9 Role or Contribution of the Unit e8. - profit generator, - service department, - opportunity seeker ii) Definition of the Business - the needs satisfied or the benefits provided. Should not be too specific (eg. “we sell milking machinery”) or too general (eg. “we’re in the engineering business”). iii) Distinctive Competence - this should be a brief statement that applies only to the specific unit. A statement that could equally apply to any competitor is unsatisfactory. . iv) Indications for Future Direction - a brief statement of the principal things that serious consideration would be given to (eg. move into a new segment). Focus C O S T EXPERIENCE C U R V E J Focus The object is to identify which market segments and products are to be considered in production of the marketing plan. This involves ignoring some detail for the sake of aiding understanding about the critical issues involved. For example, an audit of tinned foods may decide to focus on baked beans and pet foods, and ignore the small market for anchovies. Identify 20% Critical to Business The basic rule of thumb is that the 20% of the organisation’s markets and products most critical to its success are those that should be included in a strategic marketing plan. This is a guideline only: the planner may wish to conduct a more or less exhaustive plan. The Porter Matrix may assist by showing the relative strength of the products in their markets in terms of differentiation and cost leadership, as an indication of the possible future importance of the products. Segment the Market The relevant markets should be identified and, where appropriate, segmented. This is in general a creative and important step. Limited guidance’ only is incorporated in this model. The Porter matrix may be of assistance in market segmentation, as clusters of products in similar positions might reasonably be placed in a segment. The Standard Industrial Classification (SIC) used as a basis of statistics collection by the Government can form a useful starting point for market definition, as a checklist from which to select, though it is not always appropriate. Predict Next 3 Years Prediction of the future prospects of the products, all other things being equal, is important as an input into the audit of the current position. It is also an important validation step, as it may affect which 20% of the products and markets are d,eemed to be critical. For example, if the demand for anchovies is expected to rise steeply in the next three years, it may be decided to include them in the tinned foods audit after all. Consideration of where the product is in its life cycle may assist in prediction. The Ansoff matrix may already at this point suggest new markets and products that should be defined and considered. The cost experience curve may suggest what is likely to happen to the costs of the products, which may have implications for its future prospects. Conduct Audit ' C O N D U C T -I AUDIT .s F A C T O R S .S * A S S E S S C R ITICAL - S T R E N G T H S . S U C C E S S A N D F A C T O R S TABLE W E A K N E S S E S I .- O T .s * .v CHECKLIST e. O P P O R T U N I T I E S T H R E A T S s_ .- .- NESS F A C T O R S .B r- A S S E S S M A R K E T A T T R A C T l V E N E S S The objective is to assess the state and prospects of the products and markets already identified. Information needed at this point m a y have been collected in advance of the planning process, or it m a y be collected now. Assess Strengths and Weaknesses The strengths and weaknesses of the company’s products in its markets can be summarised in a Critical Success Factors table. It is very important to get this right, and to validate it against information on the competitors in the market and their strength in the markets. If the information is not available to sufficient accuracy, it should be obtained. After all, one is identif\ .‘irg factors critical to the success of the business. A checklist is available of possible factors to consider. Assess Opportunities and Threats The Porter 5-force model of pressures on you can assist in identification of threats. The Threat Assessment matrix gives guidance on whether to include ihe threats in the summary list. A checklist of possible opportunities and threats is available. Assess Market Attractiveness The Market Attractiveness table summarises the attractiveness of a market to the company. It thus complements the Critical Success Factors (CSF) table: CSF summarises the company’s prospects of success in the market if it chooses to compete, whereas this table summarises the desirability of competing. One important aspect of the market’s attractiveness is the expected future of the market: in this way, the market attractiveness table may be more forward looking than the CSF table. - 56 - _- S U M M A R I S E ._____________ i;-ll \ \ \ \ \ \ \ “.‘\f \ -* \ \ x DIRECTIONAL \ \ \ x POLICY MATRIX \ \ I * m ANALYSIS - S U M M A R Y I I The objective is to summarise the products in the business unit in a form suitable as a starting point for setting of objectives. The essential component of this is the Directional Policy Matrix, with the current picture of the portfolio, and current projections. The projections can then be modified during the setting of objectives. I The axes of the D P M have already been determined during the Audit, being the C S F factors and weightings, and the market attractiveness factors and weightings. Guidelines for the reduction of the number of products to be displayed to a sensible number m a y be used; and the axes m a y be changed and/or relabelled in order more effectively to differentiate between products, if initially they are excessively clustered. Groups of products, including portfolios, m a y meaningfully be plotted on the D P M , as well as single products: McDonald gives an example of Cranfield School of Management’s courses. - c7 - The Boston matrix m a y be used if it is appropriate in this case, on the grounds of its greater simplicity. Similar remarks apply to those above about the D P M . A financial gap m a y be ascertained at this point between a unit financial objective and the current projections. This gap is notated on the diagram as a thermometer, as in essence it simply records a gap between two values, though the traditional graphic representation has the advantage of recording the value of a third dimension of the current position. Similarly, a “strategic gap” m a y be identified between other objectives of the unit included in the mission statement, and their anticipated fulfilment on the basis of current predictions of the unit’s products and consequent work. This m a y be to do with maintaining the synergy of the organisation. Such a strategic gap would be recorded in an Analysis Summary. S E T O B J E C T I V E S --____---______ 1 The purpose of this function is to produce a list of objectives. These should be quantified, but beyond this the possible types of objectives have not been identified. Gap Analysis m a y be used to drive this process, by attempting to close the gap starting with productivity improvements, then considering n e w markets ‘and n e w products in the order suggested by the Ansoff Matrix, and finally by considering changing the business’s assets (changing the nature of the business) in order to meet the objective. At any point, changing the objective m a y also be an option. This is the reason for the feedback line to “Define Unit Mission”. The D P M suggests “directional policy guidelines” for each product/product group plotted on it. These are taken into account in setting objectives for the product or product group. The Porter matrix m a y provide further help in this. Boston m a y be used, if, again, its implications are acceptable in this case. m/ z ;TRATEGY - * The strategic steps needed to meet the objectives are identified and recorded. The model has not yet been extended to cover this in any more detail. Data Model NOTATION The data model diagrams presented later in the section are in a format known as Entity Relationships Diagrams. The diagram below is used to explain the notation, IS sold Into PRODUCT Boxes represent “entities” and lines represent “relationships”. An entity is anything you wish to hold information about, such as Products and Markets. The information can be represented by blobs by the box, with text describing the information. Each item is called an attribute, such as a market’s size. A star in place of a blob indicates an attribute that can be used to identify the particular entity concerned. A relationship represents some connection between the entities. For example, products are related to markets in that a product may be sold into a given market. An arrow leading from entity A to entity B indicates that a given instance of entity A may be related to more than one of entity B. Text by the line may be used to indicate the nature of the relationship. So a product may be sold into more than one market, and a market m a y have more than one product sold into it. The case, where there is an arrow at each end, is called a many- to-many relationship. Data Model (1) l Oeflnttlon p iLfj? l Planning / Z?Zt~on to U N ’T (self or organrrarion cornpet, tOr~sf’u~u’e I M A R K E T E G M E N T 1 : i:oZ.vtk Into (L C) is sold in to P R O J N C T l name . costs 0 newlexistlng (self of competrtor) This gives a simplified data model, as a step towards the full model, described in the next section. Products are in a many-to-many relationship with markets they are in, as are business units. Products, . markets a n d business units m a y all b e nested within others. Composite products are products consisting of several other products, which are sold individually as well. A n example might be a variety pack of cat food. A portfolio is a set of products that, by contrast, is not sold as a set, but which is in s o m e way related. The total range of cat foods offered in an example. If a product is neither of these, it is called a basic product. The critical success factors and weightings that apply to a given market are c o m m o n to all competitors in the market, so they are an attribute of the market itself in the model. This diagram is inadequate when you consider information such as market share. Market share is not an attribute of products: a given product m a y be sold into two markets, in each of which it has a different market share. S o a n e w entity is needed ‘between” Product and Market. Similarly, the attractiveness of a market to a given firm is specific to that firm, so a n e w entity is needed between Market and Business Unit. Data Model (il) *Market aRramvenesS / factors ( M A W by 1 l name INVOLVE- W A R K E T O R l need fulfilled (defn ) l Score on MAFs. iZ+fI$;;~n 47 ‘-- ‘M A C e ’ y + M E N T I N and overall score B U S ! N E S S / M A R K E T l Sac M A R K E T - UNIT S E G N IENT 0 growtn info LC) (actual of p o tential) (self or l Oeflnmon I I l C.S.f ‘I l C.S.f ‘I (actual of p o tentw com p e titor) l Plannmg l differentlatlon (l-10) ‘R O D U C T F O R l market share . costs 0 score on CSFs. t overall score 0 new/exisung l price/average prrce 0 sales volume is sold P R O D U C T * name 0 costs 0 newlexistlng period 0 Relation co orgamsation structure I OBJECTIVE (self or compecrcor) A n important area in which this model needs extension is in modelling of features that change over time. This is only loosely described at present, for example by the attribute “Growth info” for markets. O n e possibility is to have a different entity for each year (or other period) under consideration - so you might have six Market entities, one for each year from three years ago to the end of the planning period in three years’ time, with differing information as to market size and critical success factors. A n intermediate possibility would be to have s o m e information that is static over time, and other information in a separate dynamic entity. This needs investigation. Techniaue InterrelationshibS PATAUSEDBY TECmIQUU The diagram below shows the data used as input by some of the techniques modelled. ~~~~~~~~~~~~~~ 0 C.S.F ‘s \ ~~~:~~ \ *. \ (actual or pote$al) competitor) l Phlnm g ! l Aclarion tc l differentiatlqh (l- 10) . . : I l market share’ *. I PRODUCT FOR . costs ---_ l score on CSF;,-~‘oweralJ&re l new/existing ., T 0 pnce/average prq 0 sales volume -. *. I ‘\ I ’ , -. I is sold in to .’ PRODUCT . ‘\ ** +q / ’ : ii I I I I I I I’ : , I , I I I I I I I I i I I I I , , I I I I l name: l name: : : I I 0 costs’ : 0 costs’ : 0 new/existind 0 new/existind (self or CompetHor) Techniaue interrelationshins The diagrams below show various connections identified between techniques. They assume that by using a technique, any data required by it is entered into the model by s o m e means, so that data is available for another technique. T E C H N I Q U E INTERRELATIONSHIPS (1) horuontal a m vemcal a m poky guldelmes current growh EXPERIENCE + : provides x as input to - Technique interrelationships (II) dwemonal pohcy gwdehner c\ I xl B O S T O N qap I n I d~fferwm~twn. to atd JO forecastmq C O S T EXPERIENCE C U R V E - 69 - 
work_655yqr7luzbtva5htquwp75r6e	----	easypublic.com -&nbspThis website is for sale! -&nbspeasy public Resources and Information. easypublic.com Buy this domain The domain easypublic.com may be for sale by its owner! This webpage was generated by the domain owner using Sedo Domain Parking. Disclaimer: Sedo maintains no relationship with third party advertisers. Reference to any specific service or trade mark is not controlled by Sedo nor does it constitute or imply its association, endorsement or recommendation. Privacy Policy 
work_65xfr2qt4zboda2elkpprp4zra	----	AUTHOR(S): TITLE: YEAR: Publisher citation: OpenAIR citation: Publisher copyright statement: OpenAIR takedown statement: This publication is made freely available under ________ open access. This is the ______________________ version of an article originally published by ____________________________ in __________________________________________________________________________________________ (ISSN _________; eISSN __________). This publication is distributed under a CC ____________ license. ____________________________________________________ Section 6 of the “Repository policy for OpenAIR @ RGU” (available from http://www.rgu.ac.uk/staff-and-current- students/library/library-policies/repository-policies) provides guidance on the criteria under which RGU will consider withdrawing material from OpenAIR. If you believe that this item is subject to any of these criteria, or for any other reason should not be held on OpenAIR, then please contact openair-help@rgu.ac.uk with the details of the item and the nature of your complaint. Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis Anil Bandhakavi, Nirmalie Wiratunga, Stewart Massie and Deepak P. Abstract Theoretical frameworks in psychology map the relationships between emotions and sentiments. In this paper we study the role of such mapping for com- putational emotion detection from text (e.g. social media) with a aim to understand the usefulness of an emotion-rich corpus of documents (e.g. tweets) to learn polar- ity lexicons for sentiment analysis. We propose two different methods that lever- age a corpus of emotion-labelled tweets to learn word-polarity lexicons. The pro- posed methods model the emotion corpus using a generative unigram mixture model (UMM), combined with the emotion-sentiment mapping proposed in Psychology for automated generation of word-polarity lexicons that capture emotion-rich vo- cabulary. We comparatively evaluate the quality of the proposed mixture model in learning emotion-aware sentiment lexicons with those generated using supervised latent dirichlet allocation (sLDA) and word-document frequency (WDF) statistics. Sentiment analysis experiments on benchmark Twitter data sets confirm the quality of our proposed lexicons. Further a comparative analysis with sLDA, WDF based emotion-aware lexicons and standard sentiment lexicons that are agnostic to emo- tion knowledge suggest that the proposed lexicons lead to a significantly better per- formance in both sentiment classification and sentiment intensity prediction tasks. 1 Introduction Sentiment analysis concerns the computational study of natural language text (e.g. words, sentences and documents) in order to identify and effectively quantify its polarity (i.e positive or negative) [28]. Sentiment lexicons are the most popular re- sources used for sentiment analysis, since they capture the polarity of a large col- Anil Bandhakavi, Nirmalie Wiratunga, Stewart Massie School of Computing, Robert Gordon University, Aberdeen, UK, e-mail: a.s.bandhakavi, n.wiratunga, s.massie @rgu.ac.uk Deepak P. Queen’s University, Belfast, UK, e-mail: deepaksp@acm.org A.Bandhakavi, N.Wiratunga, S.Massie, and P.Deepak lection of words. These lexicons are either hand-crafted (e.g. opinion lexicon [15], General Inquirer [36] and MPQA subjectivity lexicon [38]) or generated (e.g. Senti- WordNet [9] and SenticNet [7]) using linguistic resources such as WordNet [10] and ConceptNet [20]. However, on social media (e.g. Twitter), text contains special sym- bols resulting in non-standard spellings, punctuations and capitalization; sequence of repeating characters and emoticons for which the aforementioned lexicons have limited or no coverage. As a result domain-specific sentiment lexicons were developed to capture the in- formal and creative expressions used on social media to convey sentiment [24, 11]. The extraction of such lexicons is possible with limited effort, due to the abun- dance of weakly-labelled sentiment data on social media, obtained using emoti- cons [13, 14]. However, sentiment on social media is not limited to conveying posi- tivity and negativity. Socio-linguistics suggest that on social media, people express a wide range of emotions such as anger, fear, joy, sadness etc [6]. Following the trends in lexicon based sentiment analysis, research in the textual emotion detec- tion also developed lexicons that can not only capture the emotional orientation of words [25, 31], but also quantify their emotional intensity [35, 33]. Though research in psychology defines sentiment and emotion differently [26], it also provides a relationship between them [4]. Further research in emotion clas- sification [37, 12] demonstrated the usefulness of sentiment features extracted using a lexicon for document representation. Similarly emoticons used as fea- tures to represent documents improved sentiment classification [14, 24]. However, the exploration of emotion knowledge for sentiment analysis is limited to emoti- cons [14, 16, 17], leaving a host of creative expressions such as emotional hash- tags (e.g. #loveisbliss), elongated words (e.g. haaaappyy!!!) and their concatenated variants unexplored. An emotion-corpus crawled on Twitter using seed words for different emotions as in [23, 37] can potentially serve as a knowledge resource for sentiment analysis. Adopting such corpora for sentiment analysis, e.g. sentiment lexicon extraction is particularly interesting, given the challenges involved in devel- oping effective models which can cope with the lexical variations on social media. Therefore, in this work we explore the role of a Twitter emotion corpus for ex- tracting a sentiment lexicon, which can be used to analyse the sentiment of tweets. We do a qualitative comparison between standard sentiment lexicons that are agnos- tic to the emotion-knowledge, emotion-aware sentiment lexicons generated using techniques such as supervised latent dirichlet allocation (slDA) and the proposed sentiment lexicons. Our contributions in this paper are as follows: 1. We propose two different methods to generate sentiment lexicons from a cor- pus of emotion-labelled tweets by combining our prior work on domain-specific emotion lexicon generation [1, 2], with the emotion-sentiment mapping pre- sented in Psychology (see figure 1) [4]; and 2. We comparatively evaluate the quality of the proposed sentiment lexicons, emotion- aware sentiment lexicons learnt using sLDA and the standard sentiment lexicons found in literature through different sentiment analysis tasks: sentiment intensity prediction and sentiment classification on benchmark Twitter data sets. Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis In the rest of the paper we review related literature in Section 2. In Sections 3 and 4 we formulate the methods to extract sentiment lexicons from an emotion corpus of tweets. Section 5 presents the baseline methods to extract sentiment lexicons from an emotion corpus of tweets. In Section 6 we describe our experimental set up and analyse the results. Section 7 presents our conclusions. 2 Related Work In this section we review the literature concerning sentiment lexicons, followed by a review of different emotion theories and their relationship with sentiments proposed in Psychology. 2.1 Lexicons for Sentiment Analysis Broadly sentiment lexicons are of two types: hand-crafted and automatic. Hand- crafted lexicons such as opinion lexicon [15], General Inquirer [36] and MPQA subjectivity lexicon [38] have human assigned sentiment scores. On the other hand automatic lexicons are of two types: corpus-based and resource-based. Lexicons such as SentiWordNet [9] and SenticNet [7] are resource-based, since they are ex- tracted using linguistic resources such as WordNet [10] and ConceptNet [20]. A common limitation of resource-based and hand-crafted lexicons is that, they have static vocabulary, making them limitedly effective to mine sentiment on social media, which is inherently dynamic. Corpus-based lexicons such as in [24, 11], gauge the corpus level variations in sentiment using statistical models, and are found to be very effective on social media. Further with the abundance of weakly- labelled sentiment data on social media, these lexicons can be updated with very low costs. Similarly research in emotion analysis lead to the development of resource- based [25, 31] and corpus-based emotion lexicons [35, 33]. Prior research in sentiment analysis developed models that exploit emotion knowledge, such as emoticons to gain performance improvements [14, 16, 17]. However, other forms of emotion knowledge such as an emotion corpus and the lexicons learnt from it, could potentially have richer sentiment-relevant informa- tion, compared to that of emoticons. Therefore it is interesting to study role of such emotion knowledge for sentiment analysis, in particular for sentiment lexicon gen- eration and validate its usefulness. Our work focusses on this aspect, by exploiting an emotion-labelled corpus of tweets to learn sentiment lexicons. We achieve this by combining our prior work on generative mixture models for lexicon extraction and the emotion-sentiment mapping provided in psychology. 2.2 Emotion Theories Research in psychology proposed many emotion theories, wherein each theory or- ganizes a set of emotions into some structural form (e.g. taxonomy). In the following sections we detail the most popular emotion theories studied in psychology. A.Bandhakavi, N.Wiratunga, S.Massie, and P.Deepak 2.2.1 Ekman Emotion Theory Paul Ekman, an American psychologist focused on identifying the most basic set of emotions that can be expressed distinctly in the form of a facial expression. The emotions identified as basic by Ekman are anger, fear, joy, sadness, surprise and disgust [8]. 2.2.2 Plutchik’s Emotion Theory Unlike the Ekman emotion model Plutchik’s emotion model defines eight basic emotions such as anger, anticipation, disgust, joy, fear, sadness and surprise [30]. These basic emotions are arranged as bipolar pairs namely: joy-sadness, trust- disgust, fear-anger, surprise-anticipation. 2.2.3 Parrot’s Emotion Theory Parrot organised emotions in a three level hierarchical structure [29]. The levels represent primary, secondary and tertiary emotions respectively. Parrot identified emotions such as love, joy, surprise, anger, sadness and fear, as the primary emo- tions. Though Ekman and Plutchik emotion models are popular, research in Twitter emotion detection [37], [32] focussed on emotions that largely overlap with that of Parrot, given their popular expressiveness on social media. We use the Parrot emotion-labelled twitter corpus [37] in this study for generating sentiment lexicons. 2.2.4 Emotion-Sentiment Relationship in Psychology One of the popular approaches for emotion modelling in Psychology is the dimen- sional approach, wherein each emotion is considered as a point in the continuous multidimensional space where each aspect or characteristic of an emotion is rep- resented as a dimension. Affect variability is captured by two dimensions namely valence and arousal [18]. Valence (pleasure - displeasure) depicts the degree of pos- itivity or negativity of an emotion. Arousal (activation- deactivation) depicts the excitement or the strength of an emotion. The dimensional approach depicting par- rot’s primary emotions in the valence arousal 2D space is shown in Figure 1 [3]. 3 Emotion-Aware Models for Sentiment Analysis In this section we formulate two different methods which utilize a corpus of emotion-labelled documents for sentiment analysis of text. The first method learns an emotion lexicon and further transforms it into a sentiment lexicon using the emotion-sentiment mapping (refer section 2.2) proposed in Psychology. The sec- ond method on the other hand learns the sentiment labels for the documents in the emotion corpus using the emotion-sentiment mapping, followed by a sentiment lex- icon extraction. The two proposed methods are illustrated visually in figures 2a and 2b. Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis Fig. 1: Parrot’s emotions in the valence-arousal plane of the dimensional model 3.1 Emotion Corpus-EmoSentiLex A simple way to utilize a corpus of emotion-labelled documents, XE for sentiment analysis is to first learn an emotion lexicon, and further transform it into a sentiment lexicon. An emotion lexicon U MM EmoLex in our case is a |V|×(k + 1) matrix, where U MM EmoLex(i, j) is the emotional valence of the ith word in vocabulary V to the jth emotion in E (set of k emotions) and U MM EmoLex(i,k + 1) corresponds to its neutral valence (refer section 4). Observe that k emotions are considered and the k +1 is the neutral class. Further using the emotion-sentiment mapping proposed in Psychology we transform the emotion lexicon U MM EmoLex into a sentiment lexicon U MM EmoSentiLex, which is a |V|×1 matrix as follows: U MM EmoSentiLex(i) = Log ( ∑m∈E+ U MM EmoLex(i,m) ∑n∈E− U MM Emolex(i,n) ) (1) where E+ ⊂ E and E− ⊂ E are the set of positive and negative emotions according to the emotion-sentiment mapping. Further m and n are iterators over the set of positive and negative emotions. Note that the log scoring assigns a positive value for words having stronger associations with emotions such as Joy, Surprise and Love and negative values for words having stronger associations with emotions such as Anger, Sadness and Fear. Therefore we expect that sentiment knowledge for words is implicitly captured in an emotion lexicon, which can be easily extracted using this simple transformation. Using the above method, any automatically generated emotion lexicon can be converted into a sentiment lexicon. For example automatic emotion lexicons learnt from a corpus of emotion labelled tweets using methods such as latent dirichlet al- location (LDA) can be used to learn emotion-aware sentiment lexicons. We refer A.Bandhakavi, N.Wiratunga, S.Massie, and P.Deepak (a) Emotion Corpus-EmoSentiLex (b) Emotion Corpus-SentiLex Fig. 2: Emotion-Aware Models for Sentiment Analysis the readers to section 5 for the lexicon generation process using LDA and word- frequency statistics. Though the above method induces a sentiment lexicon it does not model the document-sentiment relationships to learn the lexicon, which is im- portant to quantify word-sentiment associations. Therefore we introduce an alter- nate method which overcomes this limitation while utilizing an emotion corpus for sentiment lexicon generation. 3.2 Emotion Corpus-SentiLex An alternate way to utilize the emotion corpus, XE for sentiment analysis is to trans- form it into a sentiment corpus, XS by learning the sentiment label for each document d ∈ XE . This is done by using the emotion-sentiment mapping as follows: Sentiment(d) = { positive if emotion(d)∈ E+ negative if emotion(d)∈ E− (2) Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis The sentiment lexicon U MM SentiLex learnt from the corpus XS is a |V|×3 ma- trix, where U MM SentiLex(i,1), U MM SentiLex(i,2) and U MM SentiLex(i,3) are the positive, negative and neutral valences corresponding to the ith word in vocab- ulary V . Observe that unlike the method which learns U MM EmoSentiLex, by ag- gregating word-level emotion scores into sentiment scores, this method learns the sentiment-class knowledge corresponding to the documents, before learning a word- sentiment lexicon. We expect this additional layer of supervision, to benefit perfor- mance, following the findings of earlier research in supervised and unsupervised sentiment analysis. In the following section we briefly explain our proposed method to generate U MM SentiLex and U MM EmoLex. Further details about our proposed method can be found in [1, 2] 4 Mixture Model for Lexicon Generation In this section we describe our proposed unigram mixture model (UMM) applied to the task of emotion lexicon (U MM EmoLex) generation. Sentiment lexicon (U MM SentiLex) generation is a special case of emotion lexicon generation, where the k emotion classes are reduced to positive and negative classes. Therefore we continue the presentation for the general case, i.e. U MM EmoLex generation. We model real-world emotion data to be a mixture of emotion bearing words and emotion-neutral (background) words. For example consider the tweet going to Paris this Saturday #elated #joyous, which explicitly connotes emotion joy. How- ever, the word Saturday is evidently not indicative of joy. Further Paris could be associated with emotions such as love. Therefore we propose a generative model which assumes a mixture of two unigram language models to account for such word mixtures in documents. More formally our generative model is as follows to de- scribe the generation of documents connoting emotion et : P(Det ,Z|θet ) = |Det | ∏ i=1 ∏ w∈di [(1−Zw)λet P(w|θet ) +(Zw)(1−λet )P(w|N)] c(w,di) (3) where θet is the emotion language model and N is the background language model. λet is the mixture parameter and Zw is a binary hidden variable which in- dicates the language model that generated the word w. The estimation of parameters θet and Z can be done using expectation maximiza- tion (EM), which iteratively maximizes the complete data (Det , Z) by alternating between E-step and M-step. The E and M steps in our case are as follows: E-step: P(Zw = 0|Det ,θ (n) et ) = λet P(w|θ (n) et ) λet P(w|θ (n) et )+(1−λet )P(w|N) (4) M-step: A.Bandhakavi, N.Wiratunga, S.Massie, and P.Deepak P(w|θ (n+1) θet ) = ∑ |Det | i=1 P(Zw = 0|Det ,θ (n) et )c(w,di) ∑w∈V ∑ |Det | i=1 P(Zw = 0|Det ,θ (n) et )c(w,di) (5) where n indicates the EM iteration number. The EM iterations are terminated when an optimal estimate for the emotion language model θet is obtained. EM is used to estimate the parameters of the k mixture models corresponding to the emotions in E. The emotion lexicon U MM EmoLex is learnt by using the k emotion language models and the background model N as follows: U MM EmoLex(wi,θe j ) = P(wi|θ (n) e j ) ∑ k t=1[P(wi|θ (n) et )]+ P(wi|N) (6) U MM EmoLex(wi,N) = P(wi|N) ∑ k t=1[P(wi|θ (n) et )]+ P(wi|N) (7) where k is the number of emotions in the corpus, and U MM EmoLex is a |V|×(k + 1) matrix. 5 Baseline Domain-specific Lexicon Generation Methods In this section we formally present the other lexicon generation methods proposed in the literature using latent dirichlet allocation and word-document frequency statis- tics. These lexicon generation methods can be used to induce emotion and sentiment lexicons from documents labelled for emotions and sentiments respectively. 5.1 Supervised Latent Dirichlet Allocation based Emotion Lexicon Latent Dirichlet Allocation (LDA) [5] is a popular topic detection algorithm which models documents to exhibit characteristics of multiple topics. In sentiment analysis LDA is applied to capture the relationships between words and sentiment (positiv- ity, negativity) in addition to the topics [22, 19]. Similarly in emotion detection, LDA has been applied in a semi-supervised manner using a minimal set of domain- independent seed emotion words to learn emotion-relevant topics [39]. However su- pervised LDA (sLDA) [21] offers a more accurate means to learn emotion-relevant topics from labelled/weakly-labelled emotion corpora, because the usage of a min- imal set of seed emotion words, does not guarantee the same level of coverage for all domains, thereby affecting the accuracy of the topics generated. Accordingly sLDA can be used to learn topic (emotion) distributions and map these into a word- emotion lexicon. More formally, let θe1 ,θe2 ,...,θen be the topic distributions learnt for emotions e1,e2,...,en, then the emotion lexicon is induced as follows: sLDA EmoLex(w j,en) = P(w j|θen ) ∑ |E| i=1 P(w j|ei) (8) Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis where θen is the topic distribution for emotion en obtained from sLDA, where w j is the jth word in the vocabulary V . We learn the emotion-aware sentiment lexicons (sLDA EmoSentiLex and sLDA SentiLex) using sLDA EmoLex following the same process as illustrated in sections 3.1 and 3.2. 5.2 Word-Document Frequency based Emotion Lexicon Crowd-sourced emotion annotations provided by readers of the documents (e.g news stories) are used to learn word-emotion lexicon. These emotion annotations are in the form of numerical ratings, which can be normalized to define a probabil- ity distribution of emotions on each document. [33] proposed a lexicon generation method by combining the document-frequency distributions of words and the emo- tion distributions over documents. Since this method involves modelling of frequen- cies of words in emotional documents and emotion ratings, we refer to the method as word-document-frequency (WDF) lexicon. The generation method for the lexicon can be formally described as follows: W DF −EmoLex(w j,en) = ∑ |X| i=1 P(w j|di)rin ∑ |E| n=1 ∑ |X| i=1 P(w j|di)rin (9) where w j is the jth word in vocabulary V and rin is the normalized emotion rating of the nth emotion in E on the ith document in the corpus X . Observe that unlike the sLDA and UMM lexicons, the WDF lexicon requires the emotion labels for the documents in the form of numerical ratings. As in the case of sLDA based emotion lexicon we also learn emotion-aware sentiment lexicons (W DF EmoSentiLex and W DF SentiLex) using W DF EmoLex following the process illustrated in sections 3.1 and 3.2. 6 Evaluation Our evaluation is a comparative study, of the performance of the standard senti- ment lexicons, and the proposed emotion corpus based sentiment lexicons through a variety of sentiment analysis tasks on benchmark Twitter data sets. Significance is reported using a paired one-tailed t-test using 95% confidence (i.e. with p-value ≤ 0.05). Observe that in all our experimental results, the best performing methods are highlighted in bold. 6.1 Evaluation Tasks Our evaluation includes the following sentiment analysis tasks. 1. Sentiment intensity prediction: Given a collection of words/phrases extracted from sentiment bearing tweets, the objective is to predict a sentiment intensity score for each word/phrase and arrange them in decreasing order of intensity. A.Bandhakavi, N.Wiratunga, S.Massie, and P.Deepak The predictions are validated against a ranking given by humans. Formally, given a phrase P, the sentiment intensity score for the phrase is calculated as follows: SentimentIntensity(P) = ∑ w∈P Log ( Lex(w,+) Lex(w,−) ) ×count(w,P) (10) where w is a word in the phrase P, count(w,P) is the number of times w appears in P. Lex(w,+),Lex(w,−) are the positive and negative valences for the word w in a lexicon. Some lexicons offer the sentiment intensity scores (e.g. SenticNet, S140 lexicon), in which case we use them directly. The aforementioned computation applies to the UMM based lexicons like Sentilex and S140-UMM lexicon. 2. Sentiment classification: Given a collection of documents (tweets), the objective is to classify them into positive and negative classes. The predictions are vali- dated against human judgements. Formally, given a document d, the sentiment class is predicted using a lexicon as follows: d[+] = ∑ w∈d Lex(w,+)×count(w,d) (11) where d[+] is the positive intensity of d. Similarly d[−] indicates the negative intensity of d. Finally the sentiment class of d is determined as follows: Sentiment(d) = { positive if d[+] > d[−] negative if d[−] > d[+] (12) 6.2 Datasets We use four benchmark data sets in our evaluation. Note that the emotion corpus is used in two different ways to learn sentiment lexicons (refer sections 3 and 4). Further the S140 training data is used to learn a sentiment lexicon using the pro- posed method (refer section 4). The remaining data sets are used for evaluation. We expect our evaluation to test the transferability of each of the lexicons, given that the training and test data are not always from the same corpus, albeit from similar genre. 6.2.1 Emotion Dataset A collection of 0.28 million emotional tweets crawled from Twitter streaming API1 using emotion hashtags provided in [37]. The emotion labels in the data set correspond to Parrot’s [29] primary emotions and were obtained through distant- supervision2. Parrot’s emotion theory identifies an equal number of positive and negative emotions. Therefore we expect the sentiment lexicons learnt on this corpus to be able to mine both positive and negative sentiment in the test corpora. 1 https://dev.twitter.com/streaming/public 2 http://www.gabormelli.com/RKB/Distant-Supervision-Learning-Algorithm Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis 6.2.2 S140 Dataset A collection of 1.6 million (0.8 million positive and 0.8 million negative) sentiment bearing tweets harnessed by Go et.al [13] using the Twitter API. Further the data set also contains a collection of 359 (182 positive and 177 negative) manually annotated tweets. We generate a sentiment lexicon using the proposed method in section 4 on the 1.6 million tweets and compare it with the S140 lexicon [24]. 6.2.3 SemEval-2013 Dataset A collection of 3430 (2587 positive and 843 negative) tweets hand-labelled for sen- timent using Amazon Mechanical Turk [27]. Note that unlike the S140 test data, there is high skewness in the class distributions. Therefore it would be a greater challenge to transfer the lexicons learnt on the emotion corpus and also those learnt on the S140 training corpus to sentiment classification. 6.2.4 SemEval-2015 Dataset A collection of 1315 words/phrases hand-labelled for sentiment intensity scores [34]. A higher score indicates greater positivity. Further the words/phrases are arranged in decreasing order of positivity. We used this data set to validate the performance of different lexicons in ranking words/phrases for sentiment. 6.3 Baselines and Metrics The following different models are used in our comparative study: 1. Resource-based sentiment lexicons SentiWordNet and SenticNet; 2. Corpus-based sentiment lexicons S140 lexicon [24] and NRCHashtag lexicon [24]; 3. Corpus-based sentiment lexicon (S140-UMM lexicon) learnt using the proposed method on S140 corpus (refer section 4); 4. Corpus-based sentiment lexicons (W DF EmoSentiLex, W DF SentiLex, sLDA EmoSentiLex and sLDA SentiLex) learnt on the emotion corpus (refer section 6.2.1) using WDF, sLDA methods (refer sections 5);and 5. Corpus-based sentiment lexicons (U MM EmoSentiLex and U MM SentiLex) learnt on the emotion corpus (refer section 6.2.1) using the proposed method (refer sec- tions 3 and 4) Performance evaluation is done using using Spearman’s rank correlation coefficient and F-score for sentiment ranking and sentiment classification respectively. F-score is chosen for the classification task since it measures the performance of an algo- rithm in terms of both precision and recall. A.Bandhakavi, N.Wiratunga, S.Massie, and P.Deepak 6.4 Results and Analysis In this section we first analyse the quality of the different lexicon generation meth- ods visually using word clouds, thereafter we analyse the sentiment ranking results and the sentiment classification results obtained using the different lexicons. 6.5 Emotion word clouds for Lexicons In this section we analyse the word-emotion associations learnt by the different lexi- cons, WDF, sLDA and UMM lexicon from the twitter emotion corpus. This analysis is particularly interesting, as we expect it to reveal interesting trends that could effect the performance on the sentiment analysis tasks, whose results are discussed later in this section. Figures 3 and 4show the most expressive words for emotions fear and joy (other emotions not shown due to space constraints) identified by WDF, sLDA and UMM lexicons. It is evident from the figures that all these lexicons capture the domain-specific vocabulary that is expressed informally. This is very important in order to be able to effective emotion detection in dynamic domains such as Twitter. The word clouds presented in the figures are the top 100 words for each emotion, after removing the common words in English language. We observed that WDF lexicon is biased towards the majority class (joy here) in the corpus in learning the word-emotion associations. This done by observing the class level performance metrics like accuracy and F-score. For example it identified words connoting joy such as succeed! and Ha! as top anger words, similarly for other emotions. This is due to the fact that WDF lexicon is designed for emotion rated documents and it is less effective in capturing word-emotion associations on a corpus that have discrete emotion labels. On the other hand sLDA lexicon, because of the assumption of its underlying generative model that documents are a mixture of multiple topics (emo- tions) learnt better word-emotion associations compared to WDF lexicon. However sLDA lexicon was not able to discriminate effectively between words that strongly convey a particular emotion and those that are weakly associated with an emotion. For example words such as scared, worried and nervous are not well distinguished from other words for emotion fear and similarly for other emotions. As a result it was observed in the word clouds for the sLDA lexicon that top words for each emotion have similar size. This is not desirable since the word-emotion association scores form an important knowledge resource for sentiment analysis. It was observed that the proposed UMM lexicon discriminate between strong and weak words for each emotion effectively. This is very promising, since this knowledge will be very useful for the sentiment intensity prediction task. Further UMM is also observed to capture words that are emotion-relevant but are rare. For example words such as :) and fun! for the emotion joy. We expect this word-level analysis to help infer useful insights about the performance gains of the proposed lexicon over the baselines in different sentiment analysis tasks. In the following section we analyse the performance of the different lexicons in sentiment analysis tasks. Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis Fig. 3: Top fear words for WDF (top), sLDA (middle) and UMM (bottom) lexicons 6.5.1 Sentiment Ranking Table 1 summarizes the sentiment ranking results obtained for different lexicons. In general resource-based lexicons SentiWordNet and SenticNet are outperformed by all the corpus-based lexicons. This is expected, because the vocabulary coverage of these lexicons relevant to social media is limited compared to other lexicons. Fur- thermore, the results also suggest that the sentiment intensity knowledge captured by the corpus-based lexicons is superior to that of resource-based lexicons. NRCHashtag lexicon performed significantly better than the remaining base- lines and the proposed U MM EmoSentiLex. The significant performance differ- ences between NRCHashtag lexicon and S140 lexicon and NRCHashtag lexicon and S140-UMM lexicon clearly suggests the superiority of the NRCHashtag cor- pus over the S140 corpus in learning transferable lexicons for sentiment intensity prediction. It would be interesting to compare the performance of these lexicons in the sentiment classification tasks. In the case of emotion-aware sentiment lexicons learnt using WDF, sLDA we observed that W DF SentiLex and sLDA SentiLex out- performed their counterparts W DF EmoSentiLex and sLDA EmoSentiLex. This is expected since the former ones models document-sentiment relationships. Further A.Bandhakavi, N.Wiratunga, S.Massie, and P.Deepak Fig. 4: Top joy words for WDF (top), sLDA (middle) and UMM (bottom) lexicons sLDA based lexicons performed better over the WDF ones given that it was bet- ter able to model word-emotion associations from the documents as illustrated in section 6.5. It is extremely promising to see that the proposed lexicons outperform most of the baselines significantly. Amongst the proposed lexicons, U MM SentiLex per- formed significantly better than U MM EmoSentiLex. This is not surprising, since U MM SentiLex has the ability to incorporate the sentiment-class knowledge of the documents in the learning stage. This exactly follows the findings of earlier research in supervised and unsupervised sentiment analysis. 6.5.2 Sentiment Classification Sentiment classification results for the S140 data set are shown in Table 2. Here unlike in the sentiment intensity prediction task, SentiWordNet demonstrated com- parable performance with that of corpus-based lexicons. However, SenticNet does perform the worst amongst all the lexicons. This suggests that SentiWordNet is bet- ter transferable onto social media compared to SenticNet. Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis Table 1: Sentiment Ranking Results Method Spearman’s Rank Correlation Coefficient Emotion-agnostic Sentiment Lexicons (Baselines) SentiWordNet 0.479 SenticNet 0.425 S140 lexicon 0.506 NRCHashtag lexicon 0.624 S140-UMM-lexicon 0.517 Emotion-aware Sentiment Lexicons (Baselines) WDF-EmoSentiLex 0.489 WDF-Sentilex 0.497 sLDA-EmoSentiLex 0.493 sLDA-SentiLex 0.514 Emotion-aware Sentiment Lexicons (Proposed Methods) UMM-EmoSentiLex 0.572 UMM-SentiLex 0.682 The S140 corpus based lexicons significantly outperform NRCHashtag lexicon, given their advantage to train on a corpus, that is similar to the test set. In the case of lexicons based on sLDA and WDF we observed similar trends as seen in the sen- timent intensity prediction task. Overall sLDA SentiLex performed better than the other sLDA and WDF lexicons given that it has the ability to model word-emotion associations and document-sentiment relationships more effectively. However, the proposed lexicon U MM SentiLex recorded the best performance on this data set. once again the superiority of U MM SentiLex over U MM EmoSentiLex is evidenced, given its ability to incorporate sentiment-class knowledge of the doc- uments in the learning stage. The performance improvements of emotion corpus based sentiment lexicons over a majority of baseline lexicons, clearly suggests that emotion knowledge when exploited effectively is very useful for sentiment analysis. Table 3 summarizes the results for different lexicon on the SemEval-2013 data set. Unlike the previous, this data set has a very skewed class distribution. The im- pact of this is clearly reflected in the results. Majority of the lexicons recorded strong performances in classifying positive class documents. Once again SentiWordNet demonstrated that it is better transferable onto social media compared to SenticNet. Similar to the previous data set, S140 corpus based lexicons performed better than NRCHashtag corpus based lexicon. Overall comparison across the evaluation tasks suggests that S140 corpus based lexicons record better performance in sen- timent classification, whereas NRCHashtag lexicon records better performance in sentiment quantification. This offers interesting directions for future work on com- posing different corpora for learning sentiment lexicons. In the case of sLDA and WDF based lexicons we observed that they perform poorly compared to baseline lexicons that are emotion-agnostic. This could be due to the sensitive nature of these methods to class distributions while learning lexi- A.Bandhakavi, N.Wiratunga, S.Massie, and P.Deepak Table 2: Sentiment Classification Results on S140 test data set Method Positive F-score Negative F-score Overall F-score Emotion-agnostic Sentiment Lexicons (Baselines) SentiWordNet 69.42 67.60 68.51 SenticNet 59.88 59.84 59.86 S140-lexicon 71.55 69.42 70.48 NRCHashtag-lexicon 66.66 64.75 65.70 S140-UMM-lexicon 75.14 69.36 72.25 Emotion-aware Sentiment Lexicons (Baselines) WDF-EmoSentiLex 58.57 48.24 53.40 WDF-SentiLex 59.63 50.12 54.87 sLDA-EmoSentiLex 61.68 57.23 59.45 sLDA-SentiLex 62.93 58.11 60.52 Emotion-aware Sentiment Lexicons (Proposed Methods ) UMM-EmoSentiLex 67.51 71.14 69.32 UMM-SentiLex 72.93 74.11 73.52 cons. In general we observed that these two lexicon generation methods were not able to leverage the emotion corpus effectively to learn sentiment polarity lexicons suggesting the usefulness of the proposed UMM method which has consistently recorded best performances in all the sentiment analysis tasks. We highlight our findings for the UMM lexicon on the SemEval-2013 data set below. The proposed lexicon U MM EmoSentiLex performed significantly below most of the lexicons on this data set. We believe the inability to learn the document- sentiment relationships, coupled with the skewed class distribution characteris- tics of the data set resulted in such performance degradation. However, our pro- posed lexicon U MM SentiLex significantly outperformed all the remaining lexi- cons. The consistent performance of U MM SentiLex in all the evaluation tasks, strongly evidences the correlation between emotions and sentiments. We believe that the emotion-sentiment mapping in psychology effectively clusters the emotion corpus into sentiment classes, thereafter the ability of the UMM model to effectively capture the word-sentiment relationships resulted in the performance improvements for U MM SentiLex. 7 Conclusions In this paper we study the mapping proposed in psychology between emotions and sentiments, from a computational modelling perspective in order to establish the role of an emotion corpus for sentiment analysis. By combining a generative uni- gram mixture model (UMM) with the emotion-sentiment mapping, we propose two different methods to extract lexicons for Twitter sentiment analysis from an emo- tion labelled Twitter corpus. Further we also evaluate how the proposed UMM lexi- Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis Table 3: Sentiment Classification Results on SemEval-2013 data set Method Positive F-score Negative F-score Overall F-score Emotion-agnostic Sentiment Lexicons (Baselines) SentiWordNet 80.14 50.38 65.26 SenticNet 54.95 55.94 55.45 S140-lexicon 80.13 57.87 69.00 NRCHashtag-lexicon 80.25 53.98 67.11 S140-UMM-lexicon 78.87 55.85 67.36 Emotion-aware Sentiment Lexicons (Baselines) WDF-EmoSentiLex 59.83 47.39 53.61 WDF-SentiLex 62.18 59.88 61.03 sLDA-EmoSentiLex 67.84 58.99 63.41 sLDA-SentiLex 73.45 60.02 66.73 Emotion-aware Sentiment Lexicons (Proposed Methods) UMM-EmoSentiLex 64.51 48.37 56.44 UMM-SentiLex 83.06 60.98 72.02 con generation method fares in comparison with other automatic lexicon extraction methods proposed using supervised latent dirichlet allocation (sLDA) and word- document frequency statistics (WDF) in learning emotion-aware sentiment polarity lexicons. We comparatively evaluate the quality of the proposed emotion-aware sen- timent lexicons, those generated using sLDA, WDF and standard sentiment lexicons that are agnostic to emotion knowledge through a variety of sentiment analysis tasks on benchmark Twitter data sets. Our experiments confirm that the proposed senti- ment lexicons, yield significant improvements over standard lexicons in sentiment classification and sentiment intensity prediction tasks. It is extremely promising to see the potential of an emotion corpus as a useful knowledge resource for sentiment analysis, especially on social media where emotions and sentiments are widely ex- pressed. Further the cost-effectiveness of the emotion-sentiment mapping to cluster the emotion corpus into positive, negative classes (0.28 million tweets in a second) makes it practically possible to adopt large emotion corpora, in order to extract sen- timent lexicons with improved coverage. References 1. Bandhakavi, A., Wiratunga, N., Deepak, P., Massie, S.: Generating a word-emotion lexicon from #emotional tweets. In: Proc of the 3rd Joint Conference on Lexical and Computational Semantics (*SEM 2014) (2014) 2. Bandhakavi, A., Wiratunga, N., Massie, S., Deepak, P.: Lexicon generation for emotion detec- tion from text. IEEE Intelligent Systems, January/February (2017) 3. Binali, H., Potdar, V.: Emotion detection state-of -the-art. In: Proc of the CUBE International Information Technology Conference, pp. 501–507 (2012) 4. Binali H. Potdar, V., Wu, C.: Computational approaches for emotion detection in text. In: 4th IEEE International Conference on Digital Ecosystems and Technologies DEST (2010) A.Bandhakavi, N.Wiratunga, S.Massie, and P.Deepak 5. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent dirichlet allocation. the Journal of machine Learning research 3, 993–1022 (2003) 6. Boyd, D., Golder, S., Lotan, G.: Tweet, tweet, retweet: Conversational aspects of retweeting on twitter. In: Proc of the 43rd Hawaii International Conference on System Sciences. (2010) 7. Cambria, E., Olsher, D., Rajagopal, D.: Senticnet 3: A common and common-sense knowledge base for cognition-driven sentiment analysis. In: 28th AAAI conf on Artificial Intelligence, pp. 1515-1521 (2014) 8. Ekman, P.: An argument for basic emotions. Cognition and Emotion, 6(3), pp. 169-200 (1992) 9. Esuli, A., Baccianella, S., Sebastiani, F.: Sentiwordnet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining. In: Proc of LREC (2010) 10. Fellbaum, Christiane: Wordnet and wordnets. In: Encyclopedia of Language and Linguistics pp. 665–670 (2005) 11. Feng, S., K.Song, D.Wang, G.Yu: A word-emotion mutual reinformcement ranking model for building sentiment lexicon from massive collection of microblogs. World Wide Web, 18(4):949-967 (2015) 12. Ghazi, D., Inkpen, D., Szpakowicz, S.: Hierarchical approach to emotion recognition and clas- sification in texts. In: Proc of the 23rd Canadian conference on Advances in Artificial Intelli- gence (2010) 13. Go, A., Bhayani, R., Huang, L.: Twitter sentiment classification using distant supervision. Processing, pp 1-6 (2009) 14. Hogenboom, A., Bal, D., Frasincar, F., Bal, M.: Exploiting emoticons in polarity classification of text. Journal of Web Engineering (2013) 15. Hu, M., Liu., B.: Mining and summarizing customer reviews. In: Proc of the ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2004) 16. Hu, X., Tang, J., Gao, H., Liu, H.: Unsupervised sentiment analysis with emotional signals. In: Proc of the International World Wide Web Conference (WWW) (2013) 17. Jiang, F., Liu, Y.Q., Luan, H.B., Sun, J.S., Zhu, X., Zhang, M., Ma, S.P.: Microblog sentiment analysis with emoticon space model. Journal of Computer Science and Technology, vol 30(5), pp 1120-1129 (2015) 18. Jin, X., Wang, Z.: An emotion space model for recognition of emotions in spoken chinese. In: Proc of the First international conference on Affective Computing and Intelligent Interaction (2005) 19. Lin, C., He, Y.: Joint sentiment/topic model for sentiment analysis. In: Proceedings of the 18th ACM conference on Information and knowledge management, pp. 375–384. ACM (2009) 20. Liu, H., Singh., P.: Conceptnet- a practical commonsense reasoning tool-kit. BT Technology Journal, 22(4), pp. 211-226 (2004) 21. Mcauliffe, J.D., Blei, D.M.: Supervised topic models. In: Advances in Neural Information Processing Systems, pp. 121-128 (2007) 22. Mei, Q., Ling, X., Wondra, M., Su, H., Zhai, C.: Topic sentiment mixture: modeling facets and opinions in weblogs. In: Proceedings of the 16th international conference on World Wide Web, pp. 171–180. ACM (2007) 23. Mohammad, S.M.: #emotional tweets. In: Proc of The First Joint Conference on Lexical and Computational Semantics, pp. 246-255 (2012) 24. Mohammad, S.M., Kiritchenko, S., Zhu, X.: Nrc-canada: Building the state-of-the-art in sen- timent analysis of tweets. In: 7th International Workshop on Semantic Evaluation (SemEval 2013), pp 321-327 (2013) 25. Mohammad, S.M., Turney, P.: Crowdsourcing a word-emotion association lexicon. Computa- tional Intelligence, 29(3), pp. 436-465 (2013) 26. Munezero, M., Montero, C.S., Sutinen, E., Pajunen, J.: Are they different? affect, feeling, emotion,sentiment, and opinion detection in text. IEEE Transactions on Affective Computing, Vol 5 No 2 (2014) 27. Nakov, P., Rosenthal, S., Kozareva, Z., Stoyanov, V., Ritter, A., Wilson, T.: Semeval-2013 task2: Sentiment analysis in twitter. In: Proc of the 7th International Workshop on Semantic Evaluation (SemEval-2013) (2013) Emotion-Aware Polarity Lexicons for Twitter Sentiment Analysis 28. Pang, B., Lee, L.: Opinion mining and sentiment analysis. Foundations and Trends in Infor- mation Retrieval 2(1), 1–135 (2008) 29. Parrott, W.: Emotions in social psychology. Psychology Press, Philadelphia (2001) 30. Plutchik., R.: A general psychoevolutionary theory of emotion. In R. Plutchik & H. Kellerman (Eds.), Emotion: Theory, research, and experience: Vol. 1., (pp. 3–33) (1980) 31. Poria, S., Gelbukh, A., Cambria, E., Hussain, A., Huang, G.B.: Emosenticspace: A novel framework for affective common-sense reasoning. Knowledge-Based Systems 69, pp. 108- 123 (2014) 32. Qadir, A., Riloff, E.: Bootstrapped learning of emotion hashtags #hashtags4you. In: the 4th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis (WASSA 2013) (2013) 33. Rao, Y., Lei, J., Wenyin, L., Li, Q., Chen, M.: Building emotional dictionary for sentiment analysis of online news. World Wide Web, Vol 17, pp. 723-742 (2014) 34. Rosenthal, S., Nakov, P., Kiritchenko, S., Mohammad, S.M., Ritter, A., Stoyanov, V.: Semeval- 2015: Sentiment analysis in twitter. In: Proc of the 9th International Workshop on Semantic Evaluation (SemEval-2015) (2015) 35. Song, K., Feng, S., Gao, W., Wang, D., Chen, L., Zhang, C.: Build emotion lexicon from microblogs by combining effects of seed words and emoticons in a hetereogeneous graph. In: Proc of the 26th ACM Conference on Hypertext & Social Media, pp. 283-292 (2015) 36. Stone, P.J., Dexter, D.C., Marshall, S.S., Daniel, O.M.: The general inquirer: A computer approach to content analysis. The MIT Press (1966) 37. Wang, W.: Harnessing twitter ”big data” for automatic emotion identification. In: Proc of the ASE/IEEE International Conference on Social Computing and International Conference on Privacy, Security, Risk and Trust (2012) 38. Wilson, T., Wiebe, J., Hoffmann., P.: Recognizing contextual polarity in phrase-level sentiment analysis. In: Proc. of HLT-EMNLP-2005 (2005) 39. Yang, M., Peng, B., Chen, Z., Dingju Zhu, a.K.C.: A topic model for building fine-grained domain-specific emotion lexicon. In: Proc of the 52nd Annual Meeting of the Assoc for Computational Linguistics, pp. 421-426 (2014) coversheetJournalArticles BCS_AI_2016_ExpertSystems_ABandhakavi_Revised.pdf OA: GREEN OA Logo: AUTHORS: BANDHAKAVI, A., WIRATUNGA, N., MASSIE, S. and DEEPAK, P. TITLE: Emotion-aware polarity lexicons for Twitter sentiment analysis. YEAR: 2018 Publisher citation: BANDHAKAVI, A., WIRATUNGA, N., MASSIE, S. and DEEPAK, P. 2018. Emotion-aware polarity lexicons for Twitter sentiment analysis. Expert systems [online], Early View. Available from: https://doi.org/10.1111/exsy.12332 OpenAIR citation: BANDHAKAVI, A., WIRATUNGA, N., MASSIE, S. and DEEPAK, P. 2018. Emotion-aware polarity lexicons for Twitter sentiment analysis. Expert systems, Early View. Held on OpenAIR [online]. Available from: https://openair.rgu.ac.uk Version: AUTHOR ACCEPTED Publisher: WILEY Series: Expert systems ISSN: 0266-4720 eISSN: 1468-0394 Set statement: This is the peer reviewed version of the following article: BANDHAKAVI, A., WIRATUNGA, N., MASSIE, S. and DEEPAK, P. 2018. Emotion-aware polarity lexicons for Twitter sentiment analysis. Expert systems [online], Early View, which has been published in final form at https://doi.org/10.1111/exsy.12332. This article may be used for non-commercial purposes in accordance with Wiley Terms and Conditions for Use of Self-Archived Versions. License: BY-NC 4.0 License URL: https://creativecommons.org/licenses/by-nc/4.0 CC Logo: 2018-10-25T11:08:40+0100 OpenAIR at RGU 
work_65i36k6w6rbvlbd47zygz7n6ae	----	Dynamic User Profiles for Web Personalization Ahmad Hawalaha,b, Maria Faslia,∗ aSchool of Computer Science and Electronic Engineering, University of Essex Wivenhoe Park, Colchester CO4 3SQ, UK bCollege of Computer Science and Engineering, University of Taibah, Medina, Saudi Arabia Abstract Web personalization systems are used to enhance the user experience by providing tailor-made services based on the user’s interests and preferences which are typically stored in user profiles. For such systems to remain effective, the profiles need to be able to adapt and reflect the users’ changing behaviour. In this paper, we introduce a set of methods designed to capture and track user interests and maintain dynamic user profiles within a personalization system. User interests are represented as ontological concepts which are constructed by mapping web pages visited by a user to a reference ontology and are subsequently used to learn short-term and long-term interests. A multi-agent system facilitates and coordinates the capture, storage, management and adaptation of user interests. We propose a search system that utilizes our dynamic user profile to provide a personalized search experience. We present a series of experiments that show how our system can effectively model a dynamic user profile and is capable of learning and adapting to different user browsing behaviours. Keywords: dynamic user profile, modelling user behaviour, web personalization, ontology, multi-agent systems 1. Introduction The dramatic growth of information on the WWW has inadvertently led to information overload and hence finding a specific piece of information has become difficult and time consuming (Challam et al., 2007). Web personalization systems have emerged in recent years in order to deal with this problem aiming to provide a personalized experience to users based on their individual preferences, interests and needs. Such systems have been developed for different domains of application (Pignotti et al., 2004; Challam et al., 2007; Sieg et al., 2007; Pan et al., 2007). In e-commerce, such systems have been used to recommend new items and products to users based on their previous purchasing history (Gorgoglione et al., 2006; Huang et al., 2004), while in e-learning, they are used to provide personalized e-learning services (Sun and Xie, 2009; Zhuhadar and Nasraoui, 2008). Personalization systems require and maintain information about users, and this may include demographic data, interests, preferences, and previous history. One of the main challenges in such systems is that user interests, preferences and needs are not fixed, but change over time. If user profiles contain just static information, this eventually leads to constraining the personalization process and recommending irrelevant services and items over time. To overcome this problem, methods are required for learning and understanding different user behaviours, and then adapting the profiles accordingly. ∗Corresponding author Email addresses: ahawalah@taibahu.edu.sa (Ahmad Hawalah), mfasli@essex.ac.uk (Maria Fasli) Preprint submitted to Expert Systems with Applications 3rd December 2014 In this paper, we utilize ontological profiles to capture user interests and provide recommendations based on these. The contribution of our work is threefold. Firstly, we introduce two algorithms in order to improve the mapping process between web pages visited by the user that contain implicit information about the user interests and a reference ontology to explicitly represent these interests. Secondly, we introduce novel techniques to construct ontological short-term and long-term profiles that are tailored to the users, and adapt them based on their ongoing behaviour. Thirdly, the methods introduced attempt to recognize and handle potential interest drift and interest shift in the user interests. To demonstrate our work, we introduce a personalization system that consists of three phases. The first phase is the information retrieval phase which involves preparing a reference ontology, collecting user navigation behaviour, and mapping visited web pages to the reference ontology. Indeed, this phase is very important as capturing inaccurate user interests would directly affect the subsequent phases and eventually the personalization performance. In this phase, we utilize two novel algorithms based on our work in (Hawalah and Fasli, 2011a) to improve the mapping process. The second phase is the profile adaptation and learning phase which utilizes previous work in (Hawalah and Fasli, 2011b). This phase plays a major role in our model as it is responsible for learning, adapting and modelling ontological-user profiles. This also includes methods to adapt the ontological profiles to any shift or drift that might occur in the users’ behaviour. This phase makes use of a multi-agent system that coordinates the various processes and ensures that the user profile remains up-to-date. In the last phase, a re-ranking search system is introduced that utilizes the dynamic user profile to provide a personalized search experience. The re-ranking search system takes advantage of the user interests to provide more personalized search results. To evaluate this work, we have conducted experiments with users over a period of time to assess the ability and effectiveness of our methods in tracking and adapting to changes in the user behaviour. The rest of the paper is structured as follows. First we discuss related work. Section 3 presents the main architecture for modelling dynamic user profiles which consists of three phases with each phase being dis- cussed in more detail in a subsequent section. We introduce the evaluation in section 4. Section 5 describes the evaluation of the mapping and profile construction methods, while section 6, details the evaluation of the dynamic user profiling methods in the context of their deployment within a personalized search system. The paper ends with the conclusions and pointers to future work. 2. Related Work As information on the WWW continues to proliferate, users find it increasingly difficult and time- consuming to sift through this information. To aid the user in his/her quest for the right information (be it on items, products, movies or articles), recommender and web personalization systems have emerged. Re- commender systems use two broad categories of techniques: content-based and collaborative-filtering (CF) techniques. The first technique views users as individuals. Such systems track the user interests and prefer- ences and create an explicit profile that characterizes the user and any ensuing recommendations are guided by the profile. The second technique does not make use of complex user profiles, instead information in the form of ratings (for instance on a scale from 1-5) for items, products, etc. is collated from each user and then similarities between the users and items are calculated and exploited to make new recommendations. Our work in this paper, is related to the first family of techniques used for personalization systems. Personalization systems may rely on different knowledge bases to learn and model user profiles includ- ing taxonomies (Eirinaki et al., 2006; Mooney et al., 1998), flat databases (Wu et al., 2001) and ontologies (Middleton et al., 2004; Weng and Chang, 2008; Felden and Linden, 2007; Liu et al., 2008). Different techniques have been proposed to discover and recommend new items to users based on the modelled pro- files, such as using content-based models (Liu et al., 2008; Mooney and Roy, 2000; Middleton et al., 2004), 2 spreading activation techniques (Blanco-Fernandez et al., 2011; Liang et al., 2008; Gao et al., 2008; Weng and Chang, 2008) and classification techniques (Xu et al., 2008). Ontologies encapsulate knowledge about a domain of application (Razmerita and Lytras, 2008) and as such they provide a highly expressive medium for describing user interests and preferences and rich inter- relations among them. Unlike simple methods of representing information such as weighted keywords and semantic network profiles, an ontology provides a more powerful, deeper and broader concept hierarchy representation for user profiles (Gauch et al., 2007). Using ontologies to model profiles has already been proposed in various applications in the field of information systems in general and personalization in par- ticular. Trajkova and Gauch (2004) and Zhang et al. (2007) for example, introduced a general mechanism for modelling user profiles by implicitly tracking user visited web pages. These visited web pages are then mapped to different concepts in the Open Directory Project (ODP) domain ontology1. In (Weng and Chang, 2008), user browsing and search activities are tracked and processed in order to build user profiles. The user profiles are learnt based on an ontology that consists of a hierarchical representation of different topics. A spreading activation model is then applied using this ontology to provide adequate recommendations to users. Middleton et al. (2004) described two hybrid recommender systems that employ ontological user profiles to recommend appropriate research papers to academic staff and students. The novelty in this work is in the ability of these profiles to infer more preferences based on the collected user data, while it also offers users a visualization of their profiles so that they can explicitly modify them. Sieg et al. (2007) presented an ontological user profile for tracking user behaviour in a web search environment. A spreading activation mechanism was proposed to learn and maintain user interests. User profiles are then used to re-rank the search results based on the users’ current interests. In (Challam et al., 2007), a web search system based on ontological user profiles is proposed. The authors suggest that using contextual information, i.e. the user’s current task, can provide a more effective personalized experience. However, these studies do not focus on the process of learning user behaviour over time; instead they focus on providing accurate search personalization at a particular time. Anand et al. (2007) proposed using an ontology to represent user interests. In this approach, the content descriptor of the items in an ontology is used with user actual ratings to provide a recommendation. Some studies have proposed sophisticated approaches for modelling user profiles in a more dynamic way. Grcar et al. (2005) presented a dynamic user profile that tracks implicitly user navigation and consists of short and long-term models. However, this research assumed that all the visited web pages are of interest to the user no matter how long s/he spends reading through them. As all pages are treated the same, the potential strength of interest in various topics is ignored. Another issue with this study is that the short and long-term folders have been limited to predefined sizes (i.e. 5 and 300 respectively). The short-term folder stores the most recent visited web pages, while the long-term folder stores the last viewed web pages. This technique limits the user profile to a small number of interests, and at the same time can suffer from instability due to highly changeable interests. Along the same lines, Li et al. (2007) proposed a short-term model that uses a page-history buffer (PHB) which emulates the functionality of the cache and database disk buffer. Again the size of the buffer is fixed, which leads to the same problems as in (Grcar et al., 2005). Another limitation relates to the representation of the user interests in the profiles as the interest-topic and associated weight. An interest’s weight is represented as the total number of visits for this interest. The model does not make use of a time discount factor, hence, the weight of the interest would remain the same over time. For example, if a user visited pages associated with the topic IPhone 20 times over a period of time, then this number remains the same even if the user has shown no interest in IPhone in the more recent past. 1http://www.dmoz.org/ 3 Cantador et al. (2008) proposed an adaptation strategy for a dynamic user profile. This work deploys a reference ontology that is used to map user preferences and current context to provide personalization. User preferences can be seen as users’ long-term interests, while current user context represents the users’ interests at the current runtime. All user interests are represented as weighted concepts. However, user preferences are stored in a stack history which is limited to a predefined size and as such it would only be able to store a small proportion of user history but not all user preferences. Furthermore, the mechanism for detecting current user context does not consider the user’s previous sessions. For instance, if a user shows interest in a concept Car in many sessions, and shows interest in a concept Football in the last session, both of these concepts would be treated similarly without taking into account that a user might be more interested in Car as s/he has shown an interest in this concept over many sessions. Although the above studies provide different ways to model user profiles, most of them do not deal with the user’s changing interests over time, while others only attempted to model low-level dynamic user profiles. The distinction made in these works between short and long-term interests, and interests at current runtime is often blurred and most such approaches treat users as having fixed size interests. 3. Capturing and Modelling Dynamic User Profiles In this section, and following on from the identification of a number of drawbacks in existing works in personalization, we propose a dynamic user profiling approach. In developing our research, we have taken into account a number of factors and desirable properties for personalization systems: • Users are reluctant to provide information about their interests and preferences explicitly through completing questionnaires, etc. (Montaner, 2001). Hence, a flexible personalization system should try and capture interests through tracking user behaviours implicitly without user intervention, but as accurately as possible. • In order to be able to provide advanced services to users, their interests and preferences should be captured and stored in an appropriate format that would enable further processing to be performed and useful information to be extracted. The use of ontologies has been shown to provide a signific- ant improvement in the performance of personalization systems (Challam et al., 2007; Trajkova and Gauch, 2004; Middleton et al., 2004). To this end, we have decided to deploy ontologies for repres- enting user interests to support more effective, and semantically-driven exploration of the profile and hence richer recommendations. • User interests very rarely remain static, they constantly change and evolve over time. Most existing personalization systems do not deal with this challenge in capturing user behaviour or they do not do so in a way that looks at the user behaviour in a holistic way. Our aim is to develop profile methods that should be able to track the change in the user interests including any interest-shift or drift and adapt accordingly. • As discussed in the previous section, user interests can be distinguished into short-term, which are the user current interests and can be highly changeable, and long-term, which tend to be more stable over time. Such a distinction of user interests is useful to capture and it provides an extra dimension in trying to understand the individual user and his/her needs. However, this distinction is very often blurred in existing approaches or not captured at all. We aim to capture this aspect of user interests in our system in a clear manner. 4 User Browsing Activities Data processing A multi-agent system (Learning and adaptation) Reference Ontology User ontological profile FeedBack Personalized Services Data source Personalized system Mapping process Create an instance IR phase Learning and adaptation phase Personalization system phase Figure 1: The phases and architecture of the personalization system. • Typically work in personalization is developed to address specific domain problems. We aim to develop models and methods that will have wider applicability and can be integrated and deployed to provide a range of personalization services. Building upon these, we propose a novel personalization system that consists of three phases as illus- trated in Figure 1: • The Information Retrieval phase. • The Learning and Adaptation phase. • The Personalization phase. 3.1. Information Retrieval (IR) Phase The aim of this phase is to collect user browsing behaviour to discover their interests. Personalization is aimed at enhancing the process of retrieving information as a user receives tailored contents to his/her interests, but if the system collects the wrong interests, then it would provide inaccurate and ineffective services. The tracking and discovery of user interests is done in three stages. 3.1.1. Stage One: Tracking User Browsing Behaviour Since collecting data explicitly adds more of a burden on users (Montaner, 2001), we aim at collecting such behaviour unobtrusively. The data that we need to observe in this system is the visited websites, contents of each web page, timestamp which denotes the date/time of the visit and finally the duration of 5 the visit. The Browsing Activities component in Figure 1 is used to collect and then store this information in a log file. This component records all the visited web pages W = {W1,W2, ...,Wn} during the user browsing sessions and fetches all the textual contents txi for each Wi ∈ W . For each visited web page Wi, the timestamp q is recorded as well as the duration k of the user’s visit to the web page. This raw information is stored in a Raw log file which is then used by the Data-processing component. First, all the noise in txi such as HTML tags is removed2. This component then tokenizes all the txi for each Wi to discover terms ti. Once all the txi are cleaned and tokenized, it is important to reduce the dimensionality of terms by removing all unnecessary ones which have low discriminating values as well as reducing their ambiguity. Hence, we first remove all the common terms such as ‘and’, ‘or’ and ‘the’ (stop words) using a stop list. Subsequently, we apply the Porter stemming algorithm (Porter, 1997) on all the ti in order to return each term to its stem (e.g. computer to compute). The outcome of this phase is a P-log file that can be processed further. 3.1.2. Stage Two: Reference Ontology Preparation Ontologies comprise rich knowledge representation structures and their use has been shown to provide a significant improvement in the performance of personalization systems (Challam et al., 2007; Trajkova and Gauch, 2004; Middleton et al., 2004). In the proposed system, an ontology plays a significant role in modelling dynamic user profiles. A reference (or domain) ontology describes a particular domain of application and is usually modelled in a hierarchical way in which super-classes are linked to sub-classes. Each concept in an ontology is typically associated with a document that explains and represents it. Unlike flat representations, a reference ontology provides a richer representation of information in that semantic and structural relationships are defined explicitly. We use a reference ontology for two purposes: firstly, to identify user interests based on the visited web pages, and secondly, to represent user interests as ontological concepts. Let O be a reference ontology that represents a domain and C a set of concepts that belongs to O. R = {r1, ...rn} is a set of relations that links two concepts. A reference ontology might have different types of relations such as r1: ‘is-a’, r2: ‘has-a’, etc. As we propose a content-based personalization system, each ci ∈ C in O is associated with a document di that represents the ci. We first convert all the terms in all the documents associated with concepts in the reference ontology to vectors using the TF ∗ IDF classifier. That is, each term tj ∈ di where di → ci (associated to ci) in O is given a weight value w that measures the importance of the term tj within a document di. Using the TF ∗ IDF , this weight can be computed by dividing the number of occurrences of term tj by the total number of terms | t | in a document di. Then the inverse document frequency (IDF) which represents the general importance of the term is calculated by dividing the total number of documents D in O (| D |) by the total number of documents di that contain the term tj in O | tj . di ∈ D |. Finally, each term in the ontology is given a weight from 0 to 1. wtj = (tfji ∗ idfj) idfj = log( | D | | tj . di ∈ D | ) (1) Although the vector space model is one of the simplest classification methods, it has one drawback: it distinguishes between terms or vectors that have the same root. Words such as ‘play’, ‘plays’ and ‘played’ 2Although works have shown that HTML structure can be used as part of the indexation process, in this phase of our work, we do not make use of such information. 6 are processed as different words. This makes the classifier less effective as the dimensionality of the terms increases. In order to avoid this, and similarly to the first stage, we remove stop words and apply the Porter stemming algorithm (Porter, 1997) to remove term suffixes and return each term in the ontology to its stem. Once all the terms in a reference ontology are converted to vectors, the ontology will be ready to be used to discover user interests from user visited web pages. 3.1.3. Stage Three: Mapping Documents to a Reference Ontology Once the term weights are calculated for each term in the ontology, any vector similarity method can be used to map visited web pages to appropriate concepts (or classes) in the reference ontology. In this paper, the well-known cosine similarity algorithm (Baeza-Yates and Ribeiro-Neto, 2011) is applied to classify web pages to the right concepts. Cosine similarity is used to ascertain the similarity between two vectors and measures the cosine of the angle between them – in our case, the vector representing the web page visited against the vector representing a specific concept in the reference ontology. simCosine(d1,d2) = ∑n i=1 wi1,wi2√∑n i=1 wi1 2 ∗ √∑n i=1 wi2 2 (2) Several studies have tried to improve the accuracy of the mapping process. One such approach is by using an ontology with a limited number of levels. Liu et al. (2002) for instance, map user interests to a reference ontology that is limited to just two levels. Studies such as (Speretta and Gauch, 2005; Chen et al., 2002; Trajkova and Gauch, 2004) use a three level ontology to map user interests. When it comes to the retrieval precision, using a limited number of levels has been reported to improve the overall accuracy as it reduces the number of concepts in a reference ontology which in turn minimizes the error of mapping contents to wrong concepts. However, levels two or three of an ontology can be too general and broad to represent actual user interests, and hence this risks no specific interests being recognized. For instance, in the Open Directory Project (ODP)3 ontology, level two Computers/Programming or three Computers/ Programming/Languages are too general to represent interests such as Java or C# programming languages. Another method to improve the mapping performance is by adding a pre-defined percentage of each sub- class’s weight to its super-class. The idea is that if a user is interested in a particular class, then s/he is also interested in its super-class. For example, if a user is interested in football, then s/he also has some interest in its super-class which is likely to be Sport. Middleton et al. (2004) and Kim et al. (2007) implemented this idea by adding an extra 50% for each interest to its super-class. The process is then repeated until the root class. Although this method showed an improvement over the original cosine similarity, the accumulation behaviour that it induces puts more emphasis on the top level classes which are too general to represent actual user interests, while middle and low level classes receive less attention. Daoud et al. (2008) concur that representing interests with level two of an ontology is too general, while a leaf node representation is too detailed. They suggested that the most relevant concept is the one that has the greater number of dependencies. To this end, they proposed a sub-concept aggregation scheme where the main goal is to represent all user interests with level three in an ontology. The weight of the level three in this system is calculated by adding the weights of all its sub-concepts. However, representing user interests by level three in an ontology is restrictive as ontologies can have different structure and complexity. Some ontologies may have few levels, while others may extend to several; the ODP has seven levels. To address the limitations of the aforementioned methods, we need a method that improves the mapping process and at the same time maintains both general and specific interests. To this end, we introduce two 3http://www.dmoz.org/ 7 new methods called Gradual Extra Weight (GEW) and Contextual Concept Clustering (3C) (Hawalah and Fasli, 2011a). The Gradual Extra Weight Algorithm (GEW). The GEW algorithm is based on the idea that if a user is interested in a particular class, then s/he also has some interest in its super-class. However, we make no assumption about the number of levels an ontology has. Moreover, we do not assign a specific percentage of a sub-class to be added to its super-class such as in (Middleton et al., 2004) and (Kim et al., 2007). We propose an auto-tuning mechanism in which the percentage value of each sub-concept that is added to its super-concept is tailored to different levels of the ontology. In this mechanism, we assume that the concepts deeper in the ontology express more specific interests than a general one which is expressed with a concept higher up in the ontology (e.g. Java as opposed to Programming Languages). Therefore, the concepts in the lower levels would add more weight to their super-concepts than those in higher levels. Equation 3 shows how the extra percentage (EP ) for each level is calculated in order to transfer weight: EP = ci.Current level∗α O.Max levels (3) Where ci.Current level is the level of a sub-concept, α is a parameter that controls how much weight is transferred from one concept to another (i.e. α = 0 means no weight is transferred, while α = 1 means the maximum weight is transferred) and O.Max levels is the total number of levels in the ontology O. Using the total number of levels in an ontology in equation 3 allows us to gradually reduce the per- centage of weight that is transferred as we move up towards the root and keep a balance between general and specific interests. The default α value is 0.5, as in this case the maximum weight transferred from the concepts is 50% from the leaf concept to its super-concept and this percentage is reduced as we move up towards the root. However, α is not fixed but it can be changed based on the ontology and the number of levels at hand. When changing the α value, extra care is required to avoid selecting too high or too low an α value, as the former might cause inflation at top levels, while the latter would make the GEW ineffective as the weight which is transferred from a concept to its super-concept would be too small and insignificant. Finally, the GEW algorithm is applied to each visited web page after the cosine similarity is computed between this web page and all concepts in an ontology. However, one problem that might arise at this stage is that the ontology might have a huge number of concepts and hence, it would be too expensive to apply the GEW to all concepts. To reduce the computational load, we only apply the GEW on just the top γ results that have the highest similarity weights. Determining the best value for γ allows us to remove the concepts that do not add any value when applying the GEW algorithm4. The GEW algorithm is explained in more detail in Figure 2. To demonstrate how the GEW algorithm works, consider a reference ontology as in Figure 3 with 6 levels. For each visited web page that is mapped to a reference ontology using the cosine similarity measure, all the concepts in this ontology would be assigned a similarity weight ∈ [0,1]. The GEW algorithm starts from the highest similar concept which in this example is the concept with ID 6 that has a weight of 0.3. The extra weight based on the level of this concept which is level 6 is computed as in the dashed-box on the left. The EP which is 0.5 is then multiplied by the original weight of this concept which is 0.3. The total extra weight (0.15) will be added to the upper-concept. As the similarity of this upper-concept and the visited web page is 0.12, this weight will be modified by adding the extra 0.15 to it, so the new weight is 0.27. This process is then repeated until the root node is reached. 4In the evaluation section, we describe how the value for γ can be identified experimentally. 8 Figure 2: The GEW algorithm. 9 Figure 3: An example of the GEW algorithm. The Contextual Concept Clustering Algorithm (3C). Although the GEW method may improve the pro- cess of mapping web pages to concepts, correct mapping cannot be guaranteed as not all the visited web pages usually have good representative contents or web pages may contain a lot of noise. We aim to further improve the mapping process by taking advantage of having a log file that stores the user’s browsing behaviour. When users browse, they tend to visit several web pages that represent one interest during the same session. All current works in the field of personalization that map each visited web page to a reference ontology, do so in isolation and without any consideration about other web pages that have been visited during the same session. However, a set of visited web pages in one session might hold important hidden information that can be exploited to improve the mapping process. Assume that a user has visited 3 web pages about the XML topic, and some of these pages contain contents that are not very representative. A traditional mapping approach would map each web page to a concept from an ontology separately. As some of these web pages have poor representative contents, such an approach might assign wrong concepts to them. But if we grouped similar web pages together and applied the mapping process on this group, the risk of mapping these web pages to the wrong concept could potentially be reduced. This is because the web pages with good representative contents can “hold up” the pages with poor contents and eventually help the mapping process to map all the pages in this group (and even the ones with noise) to the correct concept in the reference ontology. Another reason why grouping web pages may be effective is that such a grouping may in essence better encapsulate the context of the user behaviour. Consider a user that visits a web page that provides a comparison between the C# and Java languages and hence contains two concepts. An assignment to either of these would be correct from the mapping process point of view. However, if this user is mainly interested in the C# language and is just curious to find out why it might be better than Java, s/he would not be interested in receiving recommendations about Java. Although classifying such a web page to either 10 C# or Java would be conceptually correct, it would not be appropriate for this specific user who is mainly interested in C#. Therefore, taking into account the user session (group of pages) would help to classify them to the most appropriate concept; in our example the C# concept. To address these issues, we introduce the Contextual Concept Clustering (3C) algorithm which aims to group related web pages into clusters which are then assigned to a particular concept in a reference ontology. The 3C algorithm works as follows (Figure 4): 1) For each browsing session, the GEW algorithm is applied in order to find the top γ similarities between each visited web page and concepts in the reference ontology. 2) After applying the GEW, the highest β similar concepts from the ontology to each web page are used as possible candidates to represent it. We need to consider all the top β results because in some cases the concept with the highest similarity value does not give a correct representation of a web page. This could be due to poor or irrelevant information in a web page, or it could be simply due to a high level of noise. 3) The “context” is then exploited by selecting the common concept that is associated with different web pages. This common concept eventually is selected to represent a web page. If there is no common concept, the concept with the highest similarity weight is selected. The β value can be identified by conducting an experiment to find the most appropriate value (see evaluation section). To illustrate the 3C algorithm, consider a user who has visited six web pages as in Figure 5. First the GEW algorithm is applied to all these web pages and then the top β similar concepts are selected as possible candidates to represent this concept. In this example, each web page is linked to 5 concepts (concepts are represented as circles and unique id numbers from the ontology). The 3C mechanism clusters these top similar concepts into groups where each group contains a set of web pages that have the same common candidate concept mapped to them. Hence there are three groups. Group 1 consists of four web pages that share concept 25 as common. The weight of this group can then be computed as the total of all the semantic similarity weights between the common concept 25, and all other web pages that are members of group 1 namely web pages 1, 2, 3 and 4. However, web pages can also be mapped at the same time to other candidate concepts that might be clustered in different groups. For example, web pages 1 and 2 can be assigned to either group 1 or 2, while web page 3 can be assigned to either group 1 or 3. If there are more than one groups, the page is assigned to the group with the highest total weight. So, in this example web pages 1, 2, 3 and 4 would be represented by concept 25, while web pages 5 and 6 would be represented by concept 21. 3.2. Learning and Adaptation Phase Typically, user interests change over time as they may lose/acquire interest in an object. If a user profile is not able to detect and adapt to such changes, this will inevitably affect the personalization performance. According to (Cantador et al., 2008; Picault and Ribiere, 2008), a number of heuristics can be used when modelling a user profile to recognize patterns in browsing behaviour: • A concept may occur in a short period with a very low occurrence. This should not be considered to be a concept of interest to a user. For example, a user might open a web page that does not meet his needs so he closes it immediately. • A concept may occur in a short period and its occurrence is very high. This concept should be considered as a short-term interest, and it should be removed once the user shows no more interest in it. For example, a user may be planning a day out in London which would involve looking for train tickets, museum and theater tickets. The London interest will exist for a while, but once the trip is taken, the user would have no further interest in London. 11 Figure 4: The 3C algorithm. 12 Figure 5: An example of the 3C algorithm. • A concept may occur over a long period and its occurrence is high. This concept should be considered as a short-term interest for a start, and once it is confirmed with time, it should become a long-term interest. For instance, a new Computer Science student will become interested in topics related to his subject. These topics should be treated as long-term interests. • A concept may occur in a short or long period, and its occurrence is very high. Then, this concept disappears, but after a while the user shows interest in this concept again. Such behaviour is difficult to recognize as a user might shift or drift his interests to new or urgent ones, and once he is done with these new interests, he returns to the previous ones. An example would be when a user is interested in reading topics related to his work, but if he decided to take a holiday and travel abroad his interests would probably shift to new topics related to the visiting country. However, once back from the holiday, the user would likely return to his earlier work-related topics. We make use of these heuristics in our learning and adaptation mechanism and to this end we introduce a multi-agent system which deals with the complex processes of recognizing, adding, updating and deleting user interests. We make use of a multi-agent system architecture for two main reasons. Firstly, the use of agents with specific responsibilities within the learning and adaptation process enables us to delineate func- tionalities and interactions between them clearly; the flow and processing of information are clearly defined. This also facilitates the use and processing of the same piece of information from multiple perspectives as different agents in the system may have access to the same information, but they may be viewing it from a different perspective and dealing with it in a different way. The second main reason is that of flexibility and extensibility. The various agents within the system can be upgraded and modified on their own without affecting the rest of the agents. Hence, we can easily make changes to the behaviour of the agents and their underlying algorithms in the future separately. This also enables us to test the system in a more systematic way and reconfigure algorithms in isolation and in conjunction with each other. The multi-agent system handles the users’ browsing behaviour in general, and the change in their in- terests in particular and it consists of three layers to provide dynamic learning and adaptation: the session based, short-term and long-term layer (Figure 6). Each layer consists of one or more agents that are re- sponsible for a (set of) task(s). 13 Figure 6: A multi-agent system for modelling dynamic user profiles. 3.2.1. Session-Based Layer In order to learn user profiles, we need to observe, process and then learn from user browsing beha- viours for each session, hence this process is activated after every session. The learning and the adaptation processes consist of adding, forgetting and deleting concepts from a user profile, computing the interest weights for the visited concepts and preparing a list of candidate concepts to be analyzed by the short-term and long-term layers. This layer consists of a number of agents each responsible for one or more tasks. Session-based Agent: This is a key agent in the multi-agent system as it is responsible for four main tasks: (i) collecting data from the P-log file; (ii) communicating with other agents to calculate the latest interest weight for all collected concepts from the P-log; (iii) storing all the processed concepts and their interest weights in a session-based profile (SBP); (iv) communicating with the Short-term and Long-term Agents to enable them to discover short-term and long-term interests respectively. Once a session finishes, the data in the P-log file are processed and stored in a SBP. One critical differ- ence between the P-log and the SBP is that the P-log consists of URLs, timestamps, durations and concepts that represent these URLs, while in the SBP there are no URLs but just concepts and their associated at- tributes. These attributes include: (i) The Status which can be: positive status such as Browsed-concept or Confirmed-concept, or negative status such as Forgotten-concept or Deleted-concept. (ii) The Relev- ance size which refers to how much a concept is relevant to a user interest. The relevance size can be measured based on user feedback about each concept. If the user feedback is positive, then the relevance size increases, but if it is negative, then it decreases. (iii) The third attribute is the Frecency which repres- ents the interest weight that indicates how much a user is interested in a concept. (iv) Finally, the Frequency attribute represents how many URLs from the P-log file have been mapped to a particular concept. The Session-based Agent starts by extracting from the P-log file the concepts that represent a web page, the Status and Frequency attributes. Next it communicates with the other agents in order to set the Relevance size and compute the Frecency weight for each concept. We borrow the term frecency, which is a combination of frequency and recency, from the field of web browsing to represent the weight of an interest (MozillaWiki, 2010). This process is performed daily and hence all new interests as well as existing ones 14 are processed daily to adapt them to any change in user behaviour. Insert Agent: This agent is responsible for processing all concepts with positive status that are received from the Session-based Agent. The positive status may be associated with different events where each event has different value to represent its importance. The different events in the system (and depending on the domain of application these may vary), need to be assigned numerical values in order to illustrate their difference and their relative importance. We identify two events: Browsed-concept and Confirmed-concept. The Browsed-concept event is the default one that is assigned to any concept that is browsed by a user. If such a concept appears in two subsequent sessions, it is assigned the Confirmed-concept event. This event has a higher weight than the Browsed-concept one, as concepts appearing in more than one session would likely be of more interest to users than those that appear in just one. Other events could be added based on the domain of application, such as Purchased-item, Printed-concept or Bookmarked-concept. Whenever a web page is browsed that has a classified concept, the interest weight (i.e. frecency) for that concept is accumulated using equation 4. The frecency value is computed based on two factors: the event and the duration that are associated with the web page. For instance, if the concept has been browsed, the event weight is assigned to be 100, while if it is a Confirmed-concept, then we assign a weight of 150. We could have assigned different values for the weights at this stage (e.g. 1 and 1.5 respectively); the important thing is to be able to differentiate between these two events and to reflect the difference in their importance. The second factor, the duration, is the time spent on a web page and as it is in seconds, it is normalized. We consider the duration when computing the interest weight for a concept since time spent on web pages has been shown by studies including (Konstan et al., 1997; Claypool et al., 2001) to be a good indicator of user interest. Fre = n∑ ci∈C ( ci.k 100 ) ∗E.weight; E.weight = { 100 if Browsed-concept 150 if Confirmed-concept } (4) Where ci.k is the time a user spends reading ci, E.weight is a weight that is added based on the event of the concept. The interest value of a concept ci is computed by summing up all the frecency values for each web page mapped to this concept. After assigning a frecency weight to each new concept, the Insert Agent adds this concept and its properties to the SBP. Forget Agent: This agent handles the behaviour that occurs when a user loses interest in a concept. An effective personalization system should be able to adapt to this change in user behaviour by forgetting such an interest from the user profile. However, such a concept should not be deleted immediately, but instead it should be forgotten gradually until the concept is confirmed as no longer being of interest to the user. Of course, not all interests are forgotten with the same pace. To account for this, the pace of the forgetting process depends on three factors. The first is the Relevance size that is associated with each concept and is an indicator of the user’s strength of interest. The Relevance size is increased/decreased depending on whether a user shows positive feedback towards a concept (i.e. by either searching for or browsing web pages related to this concept) or negative. The second factor is the recency of a concept as old concepts are forgotten faster than new ones. The final factor is related to the introduction of new interests to a user profile. If a user has started to lose his/her interest in a concept, and at the same time started to acquire new interests, then this behaviour might indicate that there is a drift in the user interests. In related work, and in order to encode the assumption that the user’s preferences gradually decay as time passes, Sugiyama et al. (2004) introduced a forgetting factor for interests in the user’s profile: e− log2 hl (d−dtinit) (5) 15 Where dtinit is the day when term t occurs initially and d is the number of days after the occurrence on day dtinit . The parameter hl is used to indicated the half-life span and is set to 7 under the assumption that user interests reduce to 1/2 in a week (Sugiyama et al., 2004). In order to apply a forgetting process to our profiles, we borrow the idea of the time-based forgetting factor from (Sugiyama et al., 2004). In contrast to using a predefined half-life span, our forgetting factor is more dynamic in order to account for the three aforementioned factors and hence the new frecency of a concept (New Fre) is calculated as follows: ci.New Fre = ci.Old Fre ∗ e − log2 (ci.Relevance size∗2) (Gm−Gl)+(#new interests : dt) (6) Where ci.Relevance size is the relevance size of a concept ci, Gm is today’s day, Gl is the day of the last occurrence and # new interests is the number of new interests that are introduced in the Gm. Unlike (Sugiyama et al., 2004) who divided log2 by a predefined half time, in this equation we divide log2 by (c.Relevance size ∗ 2) and use the recency and the number of new concepts, so the larger the value of these elements, the faster the forgetting process. This makes our forgetting factor more dynamic and we take into account individual user behaviour as users, for instance, may acquire new interests at different rates. In Equation 6, we rely on the concept’s Relevance size, recency and the number of new concepts introduced to the user profile to handle the user’s interest change and interest drift. As explained in (Godoy and Amandi, 2009), interest drift occurs when an interest experiences a gradual change. The proposed mechanism handles such behaviour by initializing the process once a user shows no more interest in a concept. The pace of forgetting further depends on whether a user has got interested in new concepts or not. The more new concepts in the SBP, the faster the forgetting process. This is particularly important in coping with interest drift as the introduction of new concepts at the time of losing interest in older concepts indicates implicitly that there is a drift in the user interests. Hence, older concepts should be forgotten and replaced by new ones. Finally, it is important to note that the forgetting process for a concept might be paused when a user shows interest again. The Relevance size would also increase to reflect this change. But if a user continues not to show interest in the concept, this will be gradually forgotten until the end of the Relevance size (the size of zero). Then this concept is sent to the Delete Agent. Delete Agent: This agent manages the gradual deletion of a concept from a user profile. The difference in the operation of the Forget and Delete Agents is that a concept with a Forgotten status can still be an interesting concept to a user and appear in his/her profile, while a concept with a Deleted status means that this concept is no longer interesting and it should not appear in the profile. When a concept is passed on to the Delete Agent, this is removed much faster based on the time of the last appearance of the concept and until the weight reaches a predefined threshold and then it is removed altogether. We use the next equation to compute the new frecency of a Deleted concept: ci.New Fre = ( ci.Old Fre Gm −Gl ) (7) Where Gm is the current date and Gl is the date of the last occurrence of ci. Hence, if there has been no interest shown for long periods of time, the frecency will be reduced accordingly. Once all concepts are processed, the Session-based Agent stores the final results into the SBP. Sub- sequently these are used to discover short-term and long-term user interests, therefore, the last task of the Session-based Agent is to communicate with the Short-term and Long-term Agents. 16 In this work, we distinguish between long-term and short-term interests as they are different in nature. Discovering the short-term interests helps a personalization system to shed light on users’ current interests which are likely to change frequently. To illustrate, a user might be interested in a holiday in France. After returning from the holiday, the user would most likely lose interest in France and therefore this should con- stitute a short-term interest. A typical personalization system would provide relative personalized contents to this user that match his/her interest in France. The long-term interests on the other hand, can play an important role in providing more effective personalization services to users that match their long-term pref- erences. If the same user has a long-term interest in science fiction movies, this can be combined with the short-term one by suggesting new science fiction movies and cinemas while in France. However, the process of discovering short and long-term interests presents challenges. The main one arises from the question of how long (short) is long (short). Users may exhibit very different browsing behaviours: some may change their interests frequently, while others may rarely change. Therefore, identi- fying length/duration is subjective to each user. We address this issue in our work by incorporating long and short-term layers that consist of adaptable mechanisms to discover short and long-term interests specific to each user. 3.2.2. Short-term Layer Our goal in this section is to develop a new mechanism to learn the user’s short-term interests and to be able to adapt to the various browsing behaviours listed in the beginning of section 3.2. This layer includes two components: the short-term profile (STP) that is used for storing all the concepts identified as short- term interests, and the short-term agent (STA) that is responsible for tasks such as discovering, maintaining and storing short-term interests in the STP. The STA discovers short-term interests by examining the concepts stored in the SBP. As each concept in the SBP is associated with a frecency that represents its importance, the STA uses these values to determine which of these concepts are short-term interests. One way to discover short-term interests is to pre-define a threshold so all concepts with weights above it are selected as short-term interests. This method has been adopted by studies such as (Li et al., 2007) and (Grcar et al., 2005). However, users are different and using the same threshold for every user would very likely result in modelling inaccurate interests. We overcome this problem by developing a self-tuning mechanism to calculate a threshold which is ad- aptable to different users as well as to changes in each user’s browsing behaviour. To identify this threshold, we first observe the frecency average of a user’s browsing behaviour for each session s. Normally, when browsing, the user would navigate to both interesting and uninteresting pages. The uninteresting topics would be assigned low frecency weights, while interesting ones would be assigned high weights (i.e. based on the time spent on each page). The threshold is computed as the ratio of the total frecency values for each visited concept in each session to the total number of concepts that have been visited in all sessions: Threshold = ∑n all s∈ui Frecency values | Total number of concepts | (8) Hence, all the concepts in the SBP that have weights above this threshold are stored in the STP. As this threshold adapts to various browsing behaviours, each user would have a different threshold. Relying solely on a threshold is not adequate, as we also need to control the size of the STP. This is because users might be interested in a large number of topics on one day, but on another they may have less interesting topics or even no interests at all. Unlike some studies such as (Li et al., 2007) and (Grcar et al., 2005) which rely on fixed size profiles to store interests, in this layer we propose a mechanism to adjust the STP size by either expanding or limiting it to reflect the actual user interests. Taking into account the user’s 17 browsing behaviour, we compare the average of the current session to the average of the last session and if it is larger/smaller, then the size of the STP is increased/decreased accordingly, otherwise it remains the same. We use the average of each session as it provides an indication of a user’s behaviour. To prevent the STP size from being increased indefinitely or diminished we set a maximum/minimum STP size which can be identified based on the domain of application at hand. The frecency average for a session is computed by: Fre average = ∑n c∈si Frecency values | Total number of concepts | (9) Where ∑n c∈si Frecency values is the total frecency values for all concepts c that have been visited during session si. Finally, the STA also needs to be able to deal with the problem of interest shift. An interest shift can be defined as abrupt change in user interests (Godoy and Amandi, 2009). Unlike interest drift which is a gradual interest change and which has been dealt with by the Forget Agent in section 3.2.1, the interest shift is more difficult to recognize as it happens suddenly (Gorgoglione et al., 2006). Users can sometimes become suddenly interested in some topics and then once these topics lose their importance, they simply abandon them. For instance, when the Football World Cup is about to start, sports fans would become interested in this and might not exhibit interest in other sports. But when the event finishes, they would likely shift their interest to other sports, or new events. As it is not easy to recognize such shifts at the time of their occurrence, personalization systems can suffer from lower performance. The STA handles interest shift by adding any new concept with a frecency weight above the computed threshold to the STP. However, if the STP is already full, then the concept would not be added. To deal with this problem, the STA replaces the concept that has the lowest frecency weight in the STP with this new concept, even if this concept has a lower frecency weight than the replaced one. We attempt to replace new concepts with existing concepts in the STP when it is full in order to discover a user’s interest shift. If this new concept appears again in the subsequent session, the Insert Agent in the session-based layer would process it and it would be treated similarly to the other concepts, while if the concept does not get confirmed, then it will be forgotten. In this case, our system would recognize and handle the interest shift when it happens. We call this the Replacement-task as older concepts are replaced by new ones. 3.2.3. Long-term Layer The Long-term layer is responsible for learning, recognizing and storing long-term interests that are confirmed with time and consists of the long-term profile (LTP), which stores all interests that are recognized as long-term ones, and the long-term agent (LTA), whose task is to recognize, maintain and store long-term interests in the LTP. Long-term interests are different in nature from short-term ones as they do not change as frequently, but instead they appear more regularly over a longer period in a user profile. For instance, users who work as programmers might have a long-term interest in programming languages which might appear regularly in their profiles. Unlike the short-term interests which rely mostly on recency when computing their interest weights, the long-term interests should rely more on the frequency of occurrence. The LTA computes the frequency weights for each concept in the SBP using the following equation: ci.FW = (Nocc ∗Nday)∗ ( Gm −Gf P.logage ) (10) 18 Where ci.FW is the frequency weight of a concept ci, Nocc is the number of times the concept ci occurs in the P-log file, Nday is the number of days that the concept ci occurs in, Gm is the day when the process is launched (i.e. today’s day), Gf is the day of the first appearance of the concept ci and finally, the P.logage is the age of the P-log. By using the age of the P-log we can detect which concepts have been appearing for longer. In this equation, the more frequent occurrences of a concept, the higher the weight assigned to it. Moreover, the interests that occur over a longer period receive higher weights than those that occur in a short period as they seem to be more stable. Unlike the STA which computes the short-term interests’ frecency after each session, the LTA computes the frequency weights for long-term interests periodically such as once or twice a month based on the domain of application. As long-term interests do not change as frequently as short-term ones, there is no need to compute them as frequently. Once the frequency weights for all concepts in the SBP are computed, the LTA needs to identify the in- terests that should be treated as long-term interests. We introduce a new mechanism to compute a threshold so that all concepts above it are stored in the LTP. Similar to the short-term layer, we propose a dynamic threshold that is based on each user’s browsing behaviour. We first compute the frequency for each concept using equation 10 and then we compute the Standard Deviation of all the concepts in the SBP: σ = √√√√ 1 N N∑ i=1 (ci.FW −AV )2 (11) Where σ is the Standard Deviation, N is the total number of concepts, ci.FW is the frequency weight of a concept ci, and AV is the average of all the frequency values for all concepts. The threshold is then: Threshold = σ + (∑N i=1 ci.FW | N | ) (12) Where ∑N i=1 ci.FW is the total of all the frequency weights for all concepts. Finally, all the concepts with frequency weights higher than the computed threshold would be stored in the LTP. Again, instead of predetermining a threshold and fixing it for all users at the same level, we utilize the information on the frequency weights of the user’s interests and their standard deviation and this enables us to tailor the threshold to the specific user’s pattern of behaviour. Long-term interests are more stable than short-term ones, so the process of discovering them is much easier. It is not appropriate to use the same technique to discover the short-term interests as they are unstable and change frequently, and they are also subject to interest-shift or drift. When finding the short-term interests, we should not use the standard deviation with the frecency average as a threshold as we do not want to limit the threshold to select just the outstanding concepts, but indeed we need to select all the concepts with frecency weights larger than the frecency average. This is because, when a user changes his/her interest, the new interest would likely have a low frecency weight in the beginning, so using just the frecency average would allow us to capture such interests. However, in order to deal with the problem of selecting too many short-term interests, in the short-term layer, we limit the size of the STP by setting a maximum number of interests. So for this particular reason, we do not have to set a specific size for the long-term layer as using the long-term threshold would likely limit the long-term interests to only those most frequent and outstanding ones. 19 Figure 7: The re-ranking process for the personalized web search system. 3.3. Personalization System Phase We integrate the work described so far with an experimental web search personalization system in the third and final phase. As part of this, we also introduce an experimental re-ranking approach. We aim to demonstrate that our approach on dynamic user profiles can effectively capture user interests as well as the change in user behaviours including interest shift and drift and can be used as part of a personalization sys- tem to provide more effective and accurate services to users. However, the methods that we have developed are generic enough and can be integrated and utilized by diverse personalization systems. In the proposed system, which is illustrated in Figure 7, when a user submits a query, two processes take place simultaneously: 1. The query is mapped to the ontological user profile to identify the most similar concept that represents it. For this reason, the cosine similarity is used to compute the similarity between a query and all the documents for all concepts in the user profile. We also pre-identify a threshold, so the highest similar concept which is above this threshold will be selected to represent this query. 2. The query is passed to any search engine (e.g. Google or Yahoo) in order to retrieve initial search results. The contents of each retrieved result are then extracted and stored in a separate document using various text analysis techniques such as the ones that have been used in section 3.1.1. We subsequently compute the similarity between each retrieved search document and the contents of the interesting concept that represents the initial user query. Finally, we re-rank all the retrieved search results in descending order and present them the user. The ranking approach is described in Figure 8. 4. Evaluation In general, the evaluation of recommender, web personalization or systems that deploy user profiles is recognized to be difficult and expensive (Yang et al., 2005). As such systems may have different purposes and they tend to be very complex, making direct comparisons is difficult. The evaluation strategies com- monly used can be divided into three main categories (Shani and Gunawardana, 2011): offline, user-centered studies and online evaluations. In offline evaluations, existing real datasets or artificial datasets are employed to assess the performance of the system. The use of such datasets minimizes the cost of the evaluation and can facilitate reuse. Although there are datasets such as the MovieLens5 one that can be used for collaborative filtering systems, rich user datasets that can be used for personalization systems are almost non-existent. 5http://MovieLens.umn.edu 20 Figure 8: The re-ranking algorithm. 21 In user-centered evaluations, real users are recruited in order to collate realistic behaviours to test the performance of the system (Shani and Gunawardana, 2011). Such evaluations involve three steps: (i) re- cruiting users; (ii) creating a set of tasks to be completed by the users; and (iii) collecting and analyzing user interactions and behaviours. Such evaluations are expensive to design and carry out and inevitably tend to be limited in scale as it is very difficult and costly to recruit large numbers of users. In particular, the tasks need to be carefully drawn so that they are able to provide appropriate and relevant results. To be able to draw meaningful comparisons among systems that are being tested, users need to be able to repeat the same or very similar tasks, hence these usually take the form of simulated work tasks. Borlund (2003) made a number of recommendations when creating tasks for such evaluations. User evaluations are able to provide a better indication of the effectiveness, accuracy and efficiency of personalization systems than using artificial datasets as systems are tested in realistic settings. They can also provide answers to a wide range of questions directly by the users and therefore afford a more well-rounded evaluation. In online evaluations, a system is essentially deployed and evaluated in a real environment. This is considered to be the best way to evaluate any system as in a real environment a system can be thoroughly tested against real users’ behaviours (Montaner, 2001; Shani and Gunawardana, 2011). However, online evaluations present a number of challenges. A system needs to be sufficiently well developed to make it available to real users. This is expensive in terms of time and effort. The experimental algorithms deployed may have a low performance, and hence, exposing them as part of an online evaluation may affect the user experience and potentially if the evaluation is carried out by a company, the users’ trust in a company, product etc. may be negatively affected. To evaluate our work, we have undertaken a user-centered evaluation that enabled us to collect real user data which we subsequently analyzed to examine the performance of the various aspects of our work. We have split the evaluation into two parts (i) evaluation of the mapping and user profile modelling methods; (ii) evaluation of the methods within a search personalization system. 5. Evaluation I: Mapping & User Profile Modelling Methods In setting up our user-centered study, we chose the domain of computers as the application domain. Our methods require the use of a reference ontology. Though we do not make any assumptions on the ontology used, for the purposes of this study, we first created a reference ontology using information extracted from the computer category from the Open Directory Project (ODP) Ontology. We used the ODP as it is con- sidered to be the largest manually constructed directory consisting of websites categorized in related topics. The computer directory contains more than 110,000 websites categorized in more than 7,000 categories. In order to train the classifier, all the websites under each category were fetched. Furthermore, all the contents of all websites were extracted and combined in one document. Hence, each category ci ∈ O is represented by one document di that includes textual contents tx from web pages W associated with the ci. All the non- semantically related classes (e.g. alphabetical order) were removed from this ontology. This resulted in a total of 4,116 categories and about 100,000 training websites whose contents were extracted and combined in 4,116 documents in our reference ontology. For each document, we apply dimensionality reduction tech- niques as explained in section 3.1.2 to minimize the number of terms. After applying these techniques, a vector space model is used to convert the terms in these documents to vectors. For this purpose, the TF-IDF classifier is used which gives each term tj ∈ di a weight from 0 to 1 using equation 1. Next, five scenarios were created to address different user browsing behaviours with respect to the heuristics that were discussed in section 3.2. The aim of these scenarios is to test the learning and adaptation ensuing from our methods when different behaviours occur and in particular interest shift or drift. For each scenario, a set of tasks were created to simulate an environment where users would interact with the system 22 as they perform their own tasks. We selected 35 random topics from the computer ontology (e.g. XML, C#, Computer security) as the basis for the experimental tasks. A total of 90 unique tasks and sub-tasks were created based on these 35 topics. The tasks take the form of finding a specific piece of information, or writing a short paragraph. Finally, 30 participants with a computer science background were invited to participate in our experiment during a 20 day period. The participants were divided so that each scenario was undertaken by six participants. These scenarios are described in more detail in appendix A. To track user browsing behaviour, a Firefox browser was used with a modified add-on component called Meetimer6. Meetimer was modified to implicitly track user browsing behaviour including all the visited web pages, timestamps and durations. An SQLite database was used to store all the user sessions. After 20 days, 30 log files from 30 different users were collected. These 30 users together surfed 9,360 web pages, and an average of 15.6 web pages a day for each user. We then processed the collected data to create processed log files (P-log files) by fetching all the visited web pages, and extracting all their contents. Moreover, we removed all the HTML tags and other noise data such as advertisements. All the stop words were removed and then the Porter stemming algorithm was applied (Porter, 1997) to reduce the terms’ dimensionality. In the next stage, we analyzed the collected data by first examining the performance of the GEW and 3C algorithms, and then the performance of the learning and adaptation processes. 5.1. Evaluation of the GEW and 3C Algorithms To evaluate the proposed mapping algorithms GEW and 3C, we conducted three experiments to analyze different aspects that impact on their overall performance. Tuning of the GEW Algorithm. In this experiment, we examine the best values for the parameters α and γ. The former controls how much weight is transferred from one concept to its upper-concept. The latter refers to the number of highest similar concepts to a web page from the reference ontology that GEW would be applied on. Determining the best γ values allows us to remove the concepts that add no value when applying the GEW algorithm. We have experimented with α = {0.1, ...,1} and γ = {5,10,20,50,70,100,n} where n refers to considering all concepts when applying the GEW. We examine all of these settings to measure the accuracy of mapping the user visited web pages to a reference ontology using equation 13. The accuracy is computed as the ratio of the total number of positively mapped web pages to correct concepts from a reference ontology to the total number of web pages visited by all users. Accuracy = | Positive mapped web pages | | All web pages | (13) Figure 9 shows the accuracy percentages for applying the GEW on all the visited web pages by all users in all scenarios using different α and γ values. It can be clearly seen that the accuracy of applying the GEW algorithm to all the similar concepts is relatively low, and the accuracy increases whenever γ decreases. In particular, when applying the GEW to the top 5, 10 and 20 concepts, the accuracy of the mapping process shows considerable improvement. Surprisingly, the best accuracy result is achieved when γ = 10 but not when γ = 5. A possible explanation for this is that the top 10 similar results to a web page could hold the most important concepts that are likely to be related to this page, while considering just the top 5 results may miss some very important concepts. When it comes to α, interestingly the accuracy results improve when α is close to 0.5, and clearly tends to decrease when α tends to the maximum or minimum. This is because when α = 1, most of the inaccurately mapped concepts are too general to represent user interests, while 6https://addons.mozilla.org/en-US/firefox/addon/meetimer/ 23 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 A cc u ra cy α value Top γ =5 Top γ= 10 Top γ= 20 Top γ= 50 Top γ= 70 Top γ= 100 All Figure 9: The accuracy results of applying the GEW algorithm on all web pages visited by all users with different α and γ values. β value 5 10 20 30 40 Accuracy 0.7805 0.7448 0.731 0.7108 0.669 Table 1: Accuracy results of the 3C algorithm with different β values. when α = 0.1 most of web pages were mapped to leaf nodes from the reference ontology which happen to be too specific to represent these interests. For α = 0.5, a balance between general and specific interests is maintained. Based on these results, we assign α = 0.5 and γ = 10 for all subsequent experiments. Tuning the 3C Algorithm. To apply the 3C algorithm, we need to set β which acts as a threshold and determines the number of concepts that are used as candidates. To identify the best β value, we apply the 3C algorithm to the top 40, 30, 20, 10 and 5 concepts. For each value, the 3C algorithm attempts to cluster web pages with common candidate concepts together in order to identify the most accurate concept that correctly represents them. Table 1 shows the accuracy results using the measure in equation 13. Surprisingly, it can be seen from Table 1 that all thresholds have achieved very close accuracy results. The top 5 threshold achieved the highest accuracy result of 0.7805, while the top 10 and top 20 performed slightly worse with 0.7448 and 0.731 respectively. This result may be interpreted as suggesting that the most appropriate concept that should be mapped to a web page is likely to be among the top 5 most similar concepts to a web page. Finally, these results support our hypothesis that mapping web pages as groups can enhance the mapping process as the accuracy of applying just the GEW is 0.6779 while by using both the GEW and 3C algorithms, the accuracy increased to 0.7805. Based on these results, we assign β = 5 in all subsequent experiments. Comparing Against Other Mapping Techniques. So far, we have examined different values to find the best possible configuration that can enhance the GEW and 3C performance. However, it is essential to examine how these algorithms perform against other techniques in the literature. Hence, in this experiment, we compare our mapping algorithms (GEW and 3C) to three different mapping techniques in the literature. The first is the original cosine similarity (OCS) which computes the similarity between each collected URL and documents in the reference ontology. Each URL is then represented by the highest similarity 24 Method Accuracy GEW & 3C 0.7805 OCS 0.4489 Adding 50% 0.6158 SAS 0.4223 Table 2: Comparison of four different mapping approaches. concept from the ODP. The second technique, which was suggested by (Middleton et al., 2004) and (Kim et al., 2007) is adding 50% of each sub-concept’s weight to its super-concept, and repeats this process until reaching the root. The third technique is the Sub-class Aggregation Scheme (SAS) that was proposed in (Daoud et al., 2008) where user interests are represented using level three of the ontology. The weight for each level-three concept is computed by adding the weights of its sub-concepts. In this experiment, each visited web page for each user is mapped by applying all four techniques. Table 2 shows the overall average accuracy for each mapping technique for all the users. According to Table 2, the OCS and SAS performed poorly as the former achieved an accuracy of 0.4489 while the latter a slightly lower result at just 0.4223. It is somewhat surprising that the OCS performed slightly better than the SAS. This poor performance of the SAS could be attributed to that all the visited web pages are mapped only to concepts in level three of the reference ontology. This causes mapping these web pages to concepts which are too general to represent them. On the other hand, adding 50% of each sub-class to its super-class shows considerable improvement in the mapping with accuracy 0.6158. This could be due to the main underlying assumption that if a user is interested in any concept, then s/he is likely have some interest in its super-concept. Although this technique improved the overall mapping accuracy, the percentage which is added to the super-classes is very high (50%). As a result, 44% of all the incorrectly mapped web pages were mapped to level 1 and 2 super-concepts that are too general and broad to represent user interests. The overall accuracy of the GEW and 3C algorithms shows a marked improvement over all other techniques, achieving an accuracy result of 0.7805. This improvement shows that GEW and 3C can overcome some of the drawbacks of other techniques as well as keep a balance between the general and the more specific interests. 5.2. Evaluation of the Dynamic User Profiling Methods Here we aim to evaluate the performance of the proposed dynamic user profile and the effectiveness of the learning and adaptation processes. Hence, we first evaluate different features of our system, and then compare its overall performance against other modelling approaches in the literature. For all of these eval- uations, we measure the precision of learning and adapting the user profiles for each day of the experiment duration (20 days) using equation 14: Precision = | Number of correctlearnedand adapted interests | | Total number of actual interests | (14) The precision in this equation is computed as the ratio of correctly modelled interests in user profiles to the actual number of interests in the scenarios. Modelling Dynamic User Profiles. First, we examine the effectiveness of the Replacement-task feature which was proposed in section 3.2.2, and aimed at dealing with the interest-shift behaviour that might occur 25 Figure 10: Average precision for each user over the 20 day period. when a user browses the Internet. We compare how our system performs when the Replacement-task is enabled and when it is disabled for each day of the experiment and for each user profile. As Figure 10 illustrates when the feature is enabled the average precision for all user profiles is 0.7649, while when it is disabled it drops slightly to 0.7228. In more detail, we can see that in some days such as days 5 and 11, a sudden drop in the precision is observed when the Replacement-task is enabled as well as when it is disabled. These drops were somehow expected as in these days, the user profiles experienced a shift of interests. Interestingly, when the Replacement-task is enabled, the precision results during the interest- shifts are considerably better than when it is disabled. Moreover, after each interest-shift such as after days 5 and 11, the system with the feature enabled managed to effectively learn and adapt to the changes in user behaviours quickly as the precision consistently improved. These results show that using the Replacement- task feature can detect and handle the interest-shift quite well. Next, we evaluate the performance of different threshold techniques that are used to discover short-term and long-term interests. First, we evaluate three such techniques for discovering short-term interests: • Fixed number of interests (Fixed): A fixed number of topics in the SBP is selected and considered as short-term interests. This threshold technique has been applied by studies such as (Grcar et al., 2005) and (Li et al., 2007). In this experiment, we select just the top 5 topics from the SBP as it was suggested in (Grcar et al., 2005). • Largest gap (LGap): In this technique, we first find the largest gap among the frecency values that are associated with the topics in the SBP, and then consider all the topics with frecency values larger than the largest gap as short-term interests. • Our approach: We employ the system described with all the aforementioned features and use the average frecency and the Replacement-task feature to select the correct short-term interests for a user. Figure 11 shows the overall comparison between the above three threshold techniques over 20 days. The LGap technique achieved the lowest precision performance at an average of 0.649. A possible explanation for this is that when a new interesting concept is browsed by a user, its interest weight would not be high. 26 Figure 11: Comparison of three techniques for discovering short-term interests. As a result when the largest gap is computed, only the existing interests with high interest weights are considered as short-term interests, while the new browsed interests are not (as their interest weights are likely to be less than the largest gap value). In addition, when using LGap, sharp drops in the precision can be observed after days 5 and 11. These drops occur because of the interest-shift and drift in users’ interests in the scenarios that have been tested which cannot be captured by LGap. When it comes to the Fixed number of interests threshold, surprisingly the performance is slightly better than LGap with an average precision of 0.683. What is interesting, is that this technique is not affected by the interest-shift and drift that occur in the tested scenarios. In days 10, 12 and 13 this technique managed to achieve the best precision in comparison to the others which shows that it can handle all the interest-shifts and drifts quite well. Since the five most recent concepts are considered as short-term interests on each day, regardless of their interest weights, it manages to capture the interest-shift and drift; when these occur, the new interests are simply added to the user short-term profiles. Although this technique has no problem with the interest-shift and drift behaviours, it has two limitations that affect its overall precision. The first limitation is that the size of the short-term profile is limited to just five interests per session. Hence, if a user has more than five interests at the same time, this technique would eliminate and ignore some of the interesting concepts. The second limitation is that the interest weights for concepts are not considered when selecting short-term interests. As a result, any concept even uninteresting ones would be considered as interesting to a user by this technique. The last threshold technique is our approach of using the average frecency with the Replacement-task feature. As it can be seen from Figure 11, our technique achieves an average precision of 0.764 which is a significant improvement on the other techniques. This result demonstrates that our approach can provide an effective mechanism to learn and adapt to the changes in user browsing behaviours. When it comes to discovering long-term interests, we evaluate and compare the performance of the following four techniques: • Fixed number of interests (Fixed): We apply the threshold technique in (Grcar et al., 2005) where the short-term profile stores the most recent visited pages, while the long-term one stores the last viewed ones. • Largest gap (LGap): In this technique, we first find the largest gap among the frequency values that 27 Threshold Technique Fixed LGap Our technique JFA Average Precision 0.411 0.86 0.91 0.84 Table 3: Comparison of four techniques for discovering long-term interests. are associated with the topics in the SBP, and then consider all the topics with frequency values larger than the largest gap as long-term interests. • Our technique: This is in essence our proposed approach of modelling a dynamic user profile by using the frequency average and the standard deviation to discover long-term interests. • Just the frequency average (JFA): In this technique, we deploy our proposed approach of modelling a dynamic user profile to learn user interests by using just the average frequency as a threshold. Unlike short-term interests that are discovered and learnt daily, long-term interests are discovered and learnt at the end of the experiment period (i.e. day 20). In Table 3, we show just the average precision of modelling long-term interests for all the user profiles for the four different threshold techniques. The results show that the Fixed technique recorded the lowest performance. This is because in (Grcar et al., 2005), all the last visited concepts are considered as long-term interests, which means that uninteresting and short- term interesting concepts are also treated as long-term interests. On the other hand, LGap surprisingly achieved a relatively good average precision of 0.86 which is better than both the Fixed and JFA. This result can be attributed to that most of the long-term interests in user profiles have higher frequency weights which are much higher than the short-term interests. However, one limitation of this technique which may not be obvious in our scenarios might appear when the list of concepts has more than one largest gap or the largest gap is between the uninteresting and short-term concepts, but not between short-term and long-term interests. In this case, the largest gap would likely capture inaccurate long-term interests from such a list. The results show that using both the frequency average and standard deviation instead of just the frequency average can provide better average precision, 0.91 and 0.84 respectively. This is because when using both frequency average and standard deviation just those long-term interests with the highest frequency weights are selected to be long-term interests, but not the short-term interests with high interest weights as in the case of using just the frequency average as a threshold. Comparison Against Other State-of-the-Art Works. In this experiment, we compare the performance of our modelling system against two other approaches, namely: (Grcar et al., 2005) and (Li et al., 2007) (both were discussed in section 2). We test all the five scenarios which are described in more detail in appendix A. Figure 12 shows the comparison results between our work and these two approaches for all users over 20 days. It can be clearly seen that the approach of (Grcar et al., 2005) has achieved the lowest result with an average precision of 0.41 for all days. This is due to a number of limitations. The main limitation is that as all visited web pages are treated as interesting, even uninteresting ones, inevitably inaccurate interests would be captured and modelled. Another problem is that the user short-term profile is limited to a small number of interests, which means that when this gets full, many interesting concepts would be left out. This is further accentuated by the fact that abandoned interests are not removed from the profile. The approach of (Li et al., 2007), performs slightly better with average precision over all days of 0.534. This result is due to different limitations. One problem is associated with the representation of a user profile where the interests are represented as concepts while the interest weights are the total number of visits. Even if the user had stopped visiting such interests as the number of visits would not change, the user interests 28 Figure 12: Comparison of three user profile modelling approaches. would not be removed from the profiles. Another issue is that the size of the buffer in the short-term profiles is fixed, which leads to the same problem of capturing just a limited number of interests as in (Grcar et al., 2005). As a consequence, this approach could not model correct user profiles when a user is interested in a relatively big number of interests such as in day 7 in scenario 3. Our approach on the other hand, managed to achieve a good average precision of about 0.75 for all days. This result is due to that user profiles are not limited to a fixed number of interests, but the size adapts in the case of short-term profiles and is unlimited in long-term profiles. This performance can also be attributed to the ability to recognize the shifts and drifts in user interests as well as forgetting uninteresting concepts more effectively. Overall, although all the designed scenarios are different and exhibit behaviours that are highly changeable, our dynamic user profile managed to maintain good precision overall for all users during all days. This shows that our proposed methods are able to effectively capture user interests and adapt to different browsing behaviours. 6. Evaluation II: Dynamic User Profiling for Personalized Search We wish to evaluate the performance of the learning and adaptation processes in the context of a per- sonalization system. To this end, we implemented the proposed search personalization system in section 3.3 that utilizes the dynamic user profiles to provide re-ranked search results based on user interests. We used the same reference ontology that we created for the first evaluation and used Google7 as the search engine to retrieve the search results. As before, the Firefox browser with the Meetimer component was used to collect user behaviour. We invited 30 users to participate in an experiment over six days. All users were postgraduate computer science students, so they are expert computer users. They were asked to search and browse for interesting Programming Languages during the first three days. During the following three days, users are asked to search and browse for any interesting topic in the field of computer science. All users were asked to search and browse for at least three different interests during each of the three days. We do this so that we can examine the shift of interests and how our models adapt to this. For each day, users were asked to write down all the visited interests in order to compare them to the modelled dynamic 7http://www.google.com 29 # of queries submitted by all users 723 # of interests visited by all users 216 # of visited web pages by all users 1,290 Table 4: Collected data from all users. user profiles. After six days, we collected all the 30 log files for all the 30 participants. Table 4 summarizes the collected information. In the next step, we retrieved all the visited web pages for all users and extracted all their contents. All the web pages that belong to one interest were aggregated in one document. Then, in order to examine the personalization performance of our proposed search system, we asked each user to select three queries related to programming languages which were visited during the first three days, and to select another three queries related to the interests that were visited during the last three days of the experiment. For each of these queries, we retrieved the top 30 results from the Google search engine. We then randomized all of these results and asked users to select just the results that were relevant to each query. Users were free to click and navigate any result for more detail. Finally, we recorded the users’ responses with regards to all of their queries. 6.1. Evaluating the Accuracy of Modelling User Profiles We are interested in examining the accuracy of modelling user profiles and in particular, the quality of identifying and mapping visited web pages to concepts from the reference ontology. As users were asked to write down all the topics that they visited for each day, we compare these recorded topics to the concepts that were discovered in the dynamic users profiles using equation 13. Figure 13 shows the average mapping accuracy results for all users in each day of the experiment for two approaches. The first approach is using our system with the GEW and 3C algorithms, while the second is using our system with the simple cosine similarity algorithm. According to Figure 13, the average accuracy for all users when our system uses the GEW and 3C algorithms is better than with the simple cosine similarity algorithm (0.791 and 0.602 respectively). In more detail, the accuracy of our model with the GEW and 3C on day 1 was only just 0.52. This could be attributed to that the dynamic user profiles have just started to learn and adapt to the users’ browsing behaviour which means that they only have a small amount of information. However, the trend from day 2 to day 3 has improved significantly with accuracy being at 0.877 and 0.91 respectively. This is because after three days of learning, the profiles adapted effectively to the new interests which in turn improved the accuracy results. On day 4, unsurprisingly a sudden decrease occurred in the overall accuracy from 0.91 to 0.692. This drop was expected as users started to shift their browsing behaviour from navigating through programming languages to navigating through their own interests. Once users started to navigate more interests and browse more web pages on days 5 and 6, the overall accuracy experienced a considerable improvement from 0.692 on day 4 to 0.89 on day 6. This demonstrates that the sudden change of user interests would lower the performance of the dynamic user profile. Nevertheless, our system shows it can recover quickly as it effectively adapted to the abrupt changes. Finally, it is clear from this figure that using the GEW and 3C outperforms the cosine similarity on all days. 6.2. Evaluating the Re-ranking Approach In this experiment, our goal is to evaluate the performance of the developed search personalization system that utilizes dynamic user profiles to re-rank search results. An effective re-ranking approach should 30 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 d1 d2 d3 d4 d5 d6 A v e rg a e A cc u ra cy Day Average accuracy for user profiles Our approach Simple Cosine similarity Figure 13: Average accuracy for each user profile over the six day period. place the most relevant results at the top of the retrieved results. For this experiment, we evaluate the quality of our proposed system by comparing its performance against three different systems. The first system is the Google original ranking (GOR) that retrieves results for each query without any further processing. The second system is using our proposed search system (PSS) that utilizes dynamic user profiles. The third system computes the simple cosine similarity (SCS) between a user query and all the retrieved results from the Google search engine and then re-ranks these results based on their weights. The last system uses the rich concept descriptions in the reference ontology (RCD). We use a range of metrics to measure the quality of these re-ranking methods. The Average Ranking Error (ARE) is calculated as follows: ARE(uq) = ∑n p∈R p.position | Total number of p | (15) Where uq is a query that is sent by a user u, R is a set of all the retrieved results for the query uq, p.position is a position of a retrieved page p in the ranking list, and | Totalnumberof p | is the total number of results that are selected by users as related to their queries. This metric which has been used by many studies such as (Li et al., 2007; Dou et al., 2007) has one important advantage as it computes the error of the ranking order of any system in comparison to the actual ranking provided by users taking into account the order of all retrieved results. Therefore, in this metric, a smaller error indicates better performance. Figure 14 shows the ARE for all users for each system. During the first three days, the PSS achieved the best overall performance in terms of ARE over all other three methods, while the RCD achieved the second best. The SCS and GOR, on the other hand, recorded lower results during the first three days. In more detail, the performance of our method (PSS) improved steadily from 13.53 on day 1 to 12.92 on day 3. This improvement is due to the fact that during the second and third days, our system managed to learn and collect sufficient information about the users’ interests. The RCD, on the other hand, achieved lower performance than the PSS, but at the same time better performance than the SCS. This is because in this approach, the concepts in the reference ontology are associated with documents that represent these concepts. These documents contributed to the good performance as they hold rich information about the associated concepts which are used to re-rank the retrieved results. Finally, 31 11 11.5 12 12.5 13 13.5 14 14.5 15 15.5 16 d1 d2 d3 d4 d5 d6 A v e ra g e R a n k in g E rr o r Day Average Ranking Error for All Users GOR SCS RCD PSS Figure 14: ARE across all four systems for all users over the six day period. the GOR system has shown the worst performance and was the only approach that the performance kept getting worse during the second and third days. This is because during days 2 and 3, users started to navigate more interests, as well as visiting more web pages regarding the interests that they had visited earlier on day 1. As a result, new concepts were discovered, and more information about each concept was recorded. The emergence of new concepts resulted in the GOR mechanism scoring the lowest ranking average of about 14.8 and 14.9 on days 2 and 3 respectively, while all the other methods managed to improve their overall performance. On day 4, all four methods experienced a sudden drop in their performance. This trend was expected as on that day users started to shift their interests from navigating topics related to programming languages to topics related to their own interests in the field of computers. Similar to the results recorded on day 1, the PSS recorded better performance than the other methods and this performance kept improving during days 5 and 6 as a result of collecting more information about user interests. Next, we look at the performance of all of the tested methods using the precision at N results (P @N) in order to examine the proportion of the relevant results ranked at different top N results. Unlike the previous metric, the order of the result is not important; what is important is to examine the users’ satisfaction over the retrieved top N results for their submitted queries. We also use the Mean Average Precision (MAP) metric to report the mean of the ranking precision for all the queries submitted by all users. Figure 15 shows the precision performance of four methods at different top N results namely: P@5, P@10, P@15, P@20 and P@25, while Table 5 illustrates the MAP for these four methods. The results in Figure 15 show that the PSS system achieved the best precision performance at all P @N results. RCD and SCS scored the second and third best performances while GOR achieved the lowest precision. At P@5, the PSS method achieved the best precision result of 0.881 which represents about 12%, 13% and 18% improvement over the RCD, SCS and GOR respectively. This indicates the effectiveness of the system that provides a re-ranking search result based on the proposed method of modelling dynamic user profiles. We also note that the RCD achieved better performance than the other methods due to that all the retrieved results were re-ranked based on the information that is associated with the concepts in the reference ontology. Finally, both the SCS and GOR achieved the lowest precision performance at all P @N results. When it comes to the MAP metric, the results reflect those of Figure 15. That is, the PSS has the highest MAP of 0.807, while the RCD and SCS scored the second and third best MAP (i.e. 0.748 and 32 0.4 0.5 0.6 0.7 0.8 0.9 1 P@5 P@10 P@15 P@20 P@25 P re ci si o n P@N for four ranking methods PSS GOR SCS RCD Figure 15: P @N results for the four different ranking methods. Method MAP GOR 0.677 PSS 0.807 SCS 0.73 RCD 0.748 Table 5: MAP results for the four ranking methods. 0.73 respectively), while GOR achieved the lowest MAP of 0.677. Overall, PSS effectively outperformed the three other methods, and was able to provide more accurate and reliable results when the user profiles had sufficient amounts of information. This demonstrates that the proposed dynamic user profile adapted effectively to the users’ interests, and managed to provide an improved personalized search service. 7. Conclusions In this paper, we presented a dynamic user profile modelling approach for web personalization. The starting point of our work was the identification of a number of issues in existing approaches, namely, the limitations of such systems in dealing with the user’s changing interests over time, the lack of a clear delineation of the short and long-term aspects of user interests and weaknesses in modelling dynamic aspects of the user’s behaviour. In our work, we aimed to address these issues. Firstly, the developed methods are able to deal with the constant change and transitions in a user’s interests including sudden change. Secondly, we have developed methods that are able to capture to a satisfactory extent the differences in a user’s short and long-term interests and the transition from one to another. These methods are also able to adapt based on the patterns of behaviour of individual users and hence we do not apply a one-size-fits-all approach. Finally, the methods developed have wider applicability and can be used in a number of domains and applications in which ongoing user behaviour can be captured and information about user preferences and interests can be extracted and modelled. 33 In more detail, our theoretical contribution is three-fold. Firstly, we introduce two mapping algorithms to improve the mapping process between web pages visited by the user that contain implicit information about his/her interests, and a reference ontology to explicitly represent these interests. Secondly, we intro- duce novel techniques to construct ontological short and long-term profiles that are tailored to the users, and adapt them based on their ongoing behaviour. Thirdly, within the methods for the construction and adapt- ation of user profiles, we introduce techniques to recognize and handle potential interest drift and interest shift in the user interests. Unlike other approaches in the literature such as (Challam et al., 2007; Zhang et al., 2007; Trajkova and Gauch, 2004), the methods developed are able to handle the dynamic nature of user interests and behaviours that change frequently. Our methods can handle the addition of new interests to a user profile, but also other more convoluted processes such as updating, gradually forgetting and delet- ing user interests. We also distinguish between long-term and short-term interests. Unlike approaches that suggest pre-defined thresholds to discover and store such interests (Grcar et al., 2005; Li et al., 2007), we presented mechanisms that are able to adapt to each user according to his/her behaviour. Another contri- bution is that the proposed methods are flexible and can be imported and used with diverse personalization systems. In this paper, we have illustrated this contribution by developing a re-ranking algorithm that uses our system for modelling dynamic user profiles to provide personalized search results to users. In order to evaluate our work, we introduced a two-part evaluation. The first part aimed at evaluating the mapping and user profile modelling methods, while the second part aimed at evaluating the methods within a re-ranking personalization system. Both of these user-centered evaluations used the same refer- ence ontology which was extracted from the ODP. The first evaluation showed that the methods developed are capable of mapping web pages to the reference ontology more accurately than other methods in the literature. In addition, we showed that by examining different user behaviours, the learning and adaptation strategies are capable of dynamically capturing user interests and forgetting the ones that are no longer of interest. We also demonstrated that our methods are able to deal with the interest-shift and interest-drift behaviours by minimizing the negative impact of these behaviours when modelling user profiles. In the second set of evaluations, we examined our work within a personalized search system. We found that our system achieved higher performance than the other search approaches. Our evaluation results demonstrated that the proposed methods can effectively capture user interests, adapt to the changes occurring in user behaviours and can enhance the performance of a personalization system. Although in this work we have addressed some of the limitations of existing personalization methods and we have showed that our system is effective at learning and adapting dynamic user profiles, our proposed approach has a number of limitations. It is essential for our system to collect sufficient information about each user in order to learn user profiles, and without such information, the system would not be able to provide effective personalization services. This problem of having just a small amount of information is known as the cold start problem and is a typical problem in content-based personalization systems. In addition, in order for our adaptation methods to work effectively, the user needs to be using the system in an ongoing basis. If the user stops using the system for a period of time and returns while interests have changed significantly, the system will still be producing recommendations based on older information. As it is often the case with content-based systems, our system could suffer from the problem of over- specialisation. In essence, as the user profile adapts, the profile information becomes ever more specialised and the system may be unable to make novel suggestions to the user. Another issue is that the effectiveness of the techniques that we have proposed to infer and exploit semantic information from user profiles depend heavily on the richness and the quality of the reference ontology. We assumed the existence of such an ontology, but it is essential to note that the quality of the recommendations will depend on the quality and richness of the underlying reference ontology deployed. Another weakness of our work is that some of the 34 models and mechanisms depend on parameters and settings that need to be pre-optimized. For example, in order to compute the semantic relatedness between ontological concepts, we need to pre-identify the weights of each relation type in an ontology. Similarly, the processes of learning, adapting and exploiting dynamic user profiles have a number of parameters that need to be identified. For our evaluation, such parameters were defined by running experiments using training datasets where the values and settings that provide the optimum results were selected. However, such a mechanism is too slow and requires conducting a large number of experiments. Hence when our methods are applied in different domains, these settings need to identified and carefully selected. Ontological user profiles comprise a rich representation of user interests which can help us to understand the user needs in a more effective way and hence deliver better services. There are a number of avenues for future work and extensions that we would like to explore. Currently, although we are using ontologies to model the user profiles, we are not making use of the axioms that may be encoded into an ontology. Such axioms may allow more useful information to be extracted from the user profiles and more complex inferences to be made. We would like to explore how such axioms can play a role in further shaping the user profiles and also as part of the recommendation process. With the huge growth of social networks, we believe that some of the developed methods can be ex- tended and applied in order to provide social recommendations to users. In particular, we could extend the use of our methods to combine information included in social network profiles of individual users, but also make use of information in their connections’ profiles in order to understand user preferences and interests better and also identify similarities with connections that can help enrich recommendations further. The latter could also help alleviate the problem of over-specialisation that our profiles may suffer from. In ad- dition, such social network profiles contain multi-modal information which includes tags, videos, pictures, etc. and another direction that we could extend our work would be toward developing methods that could make use of different types of information. In order to deal with the cold start issue, additional information can be used as suggested in (Middleton et al., 2004) which could be an external ontology that includes more information about users such as their job positions, publications, or the projects that they are working on. Such information could for instance be collated from sites like LinkedIn8 (even if not available in an ontology format). Another potential solution would be to develop a hybrid recommendation technique which combines our ontological user profile with a collaborative filtering approach such as in (Basiri et al., 2010). It would be interesting to use such hybrid system to also take advantage of the features of both techniques in order to extend and enhance our system to provide more accurate and diverse services to users. Acknowledgements Dr Ahmad Hawalah was sponsored for his PhD studies by the University of Taibah, Medina, Kingdom of Saudi Arabia. Anand, S.S., Kearney, P., Shapcott, M., 2007. Generating semantically enriched user profiles for web personalization. ACM Transactions on Internet Technology 7. Baeza-Yates, R., Ribeiro-Neto, B., 2011. Modern Information Retrieval: The Concepts and Technology behind Search (2nd Edition). 2 ed., Addison-Wesley Professional. Basiri, J., Shakery, A., Moshiri, B., Hayat, M., 2010. Alleviating the cold-start problem of recommender systems using a new hybrid approach, in: Proceedings of the 5th International Symposium on Telecommunications (IST 2010), pp. 962–967. Blanco-Fernandez, Y., Nores, M.L., Gil-Solla, A., Cabrer, M.R., Arias, J.J.P., 2011. Exploring synergies between content-based filtering and spreading activation techniques in knowledge-based recommender systems. Information Sciences 181, 4823–4846. 8https://uk.linkedin.com 35 Borlund, P., 2003. The IIR evaluation model: a framework for evaluation of interactive information retrieval systems. Information Research 8. Cantador, I., Fernandez, M., Vallet, D., Castells, P., Picault, J., Ribire, M., 2008. A multi-purpose ontology-based approach for personalised content filtering and retrieval, in: Advances in Semantic Media Adaptation and Personalization. Springer Verlag. volume 93, pp. 25—51. Volume title: Advances in Semantic Media Adaptation and Personalization. Challam, V., Gauch, S., Chandramouli, A., 2007. Contextual search using ontology-based user profiles, in: Large Scale Semantic Access to Content (Text, Image, Video, and Sound) (RIAO ’07), Pittsburgh, Pennsylvania. pp. 612—617. Chen, C.C., Chen, M.C., Sun, Y., 2002. PVA: A Self-Adaptive personal view agent. Journal of Intelligent Information Systems 18, 173—194. Claypool, M., Le, P., Wased, M., Brown, D., 2001. Implicit interest indicators, in: Proceedings of the 6th International Conference on Intelligent User Interfaces, ACM. pp. 33—40. Daoud, M., Tamine-Lechani, L., Boughanem, M., 2008. Learning user interests for a session-based personalized search, in: Proceedings of the Second International Symposium on Information Interaction in Context, ACM, New York, NY, USA. pp. 57—64. Dou, Z., Song, R., Wen, J., 2007. A large-scale evaluation and analysis of personalized search strategies, in: Proceedings of the 16th International Conference on the World Wide Web, ACM, New York, NY, USA. pp. 581—590. Eirinaki, M., Mavroeidis, D., Tsatsaronis, G., Vazirgiannis, M., 2006. Introducing semantics in web personalization: The role of ontologies, in: Ackermann, M., Berendt, B., Grobelnik, M., Hotho, A., Mladeni, D., Semeraro, G., Spiliopoulou, M., Stumme, G., Svtek, V., Someren, M. (Eds.), Semantics, Web and Mining. Springer. volume 4289, pp. 147–162. Felden, C., Linden, M., 2007. Ontology-Based user profiling, in: Abramowicz, W. (Ed.), Business Information Systems. Springer Berlin / Heidelberg. volume 4439 of Lecture Notes in Computer Science, pp. 314—327. Gao, Q., Yan, J., Liu, M., 2008. A semantic approach to recommendation system based on user ontology and spreading activation model, in: Proceedings of the 2008 IFIP International Conference on Network and Parallel Computing, pp. 488–492. Gauch, S., Speretta, M., Chandramouli, A., Micarelli, A., Brusilovsky, P., Kobsa, A., Nejdl, W., 2007. User profiles for personalized information access the adaptive web, in: The Adaptive Web. Springer Berlin / Heidelberg. volume 4321, pp. 54—89. Godoy, D., Amandi, A., 2009. Interest drifts in user profiling: A Relevance-Based approach and analysis of scenarios. The Computer Journal 52, 771—788. Gorgoglione, M., Palmisano, C., Tuzhilin, A., 2006. Personalization in context: Does context matter when building personalized customer models?, in: Proceedings of the IEEE International Conference on Data Mining, pp. 222—231. Grcar, M., Mladeni, D., Grobelnik, M., 2005. User profiling for interest-focused browsing history, in: Proceedings of the Workshop on End User Aspects of the Semantic Web (in conjunction with the 2nd European Semantic Web Conference), pp. 99–109. Hawalah, A., Fasli, M., 2011a. Improving the mapping process in ontology-based user profiles for web personalization systems, in: Proceedings of the International Conference or Agents and Artificial Intelligence ICAART, pp. 321—328. Hawalah, A., Fasli, M., 2011b. A multi-agent system using ontological user profiles for dynamic user modelling, in: Proceedings of the 2011 IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology, pp. 430—437. Huang, Z., Chung, W., Chen, H., 2004. A graph model for E-Commerce recommender systems. Journal of the American Society for Information Science and Technology 55, 259—274. Kim, J., Kim, J., Kim, C., 2007. Ontology-Based user preference modeling for enhancing interoperability in personalized services, in: Stephanidis, C. (Ed.), Universal Access in Human-Computer Interaction. Applications and Services. Springer, pp. 903—912. Konstan, J.A., Miller, B.N., Maltz, D., Herlocker, J.L., Gordon, L.R., Riedl, J., Volume, H., 1997. Grouplens: Applying collabor- ative filtering to usenet news. Communications of the ACM 40, 77—87. Li, L., Yang, Z., Wang, B., Kitsuregawa, M., 2007. Dynamic adaptation strategies for long-term and short-term user profile to personalize search, in: Proceedings of the Joint 9th Asia-Pacific Web and 8th International Conference on Web-age Information Management Conference on Advances in Data and Web Management, pp. 228—240. Liang, T.P., Yang, Y.F., Chen, D.N., Ku, Y.C., 2008. A semantic-expansion approach to personalized knowledge recommendation. Decision Support Systems 45, 401–412. Liu, F., Yu, C., Meng, W., 2002. Personalized web search by mapping user queries to categories, in: Proceedings of the Eleventh International Conference on Information and Knowledge Management (CIKM), ACM, New York, NY, USA. pp. 558—565. Liu, W., Jin, F., Zhang, X., 2008. Ontology-based user modeling for e-commerce system, in: Proceedings of the Third International Conference on Pervasive Computing and Applications (ICPCA) 2008, IEEE. pp. 260–263. Middleton, S.E., Shadbolt, N.R., De Roure, D.C., 2004. Ontological user profiling in recommender systems. ACM Transactions on Information Systems (TOIS) 22, 54—88. Montaner, M., 2001. A taxonomy of personalized agents on the internet. Technical Report, TR-2001-05-1, Departament d’Electronica, Informatica i Automatica. Universitat de . Mooney, R.J., Bennett, P.N., Roy, L., 1998. Book recommending using text categorization with extracted information, in: Papers from 1998 Workshop on Recommender systems, AAAI Press. pp. 49–54. 36 Mooney, R.J., Roy, L., 2000. Content-based book recommending using learning for text categorization, in: Proceedings of the Fifth ACM Conference on Digital Libraries, ACM, New York, NY, USA. pp. 195–204. MozillaWiki, 2010. User:Mconnor/Past/PlacesFrecency - MozillaWiki. See https://wiki.mozilla.org/User:Mconnor/PlacesFrecency. Pan, X., Wang, Z., Gu, X., 2007. Context-Based adaptive personalized web search for improving information retrieval effectiveness, in: Proceedings of the International Conference on Wireless Communications, Networking and Mobile Computing, IEEE. pp. 5427—5430. Picault, J., Ribiere, M., 2008. An empirical user profile adaptation mechanism that reacts to shifts of interests, in: Proceedings of the European Conference in Artificial Intelligence ECAI. Pignotti, E., Edwards, P., Grimnes, G.A., 2004. Context aware personalised service delivery, in: Proceedings of the European Conference in Artificial Intelligence ECAI 2004, Valencia. pp. 1077–1078. Porter, M.F., 1997. An algorithm for suffix stripping, in: Sparck Jones, K., Willett, P. (Eds.), Readings in Information Retrieval. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, pp. 313—316. Razmerita, L., Lytras, M.D., 2008. Ontology-Based user modelling personalization: Analyzing the requirements of a semantic learning portal, in: Lytras, M.D., Carroll, J.M., Damiani, E., Tennyson, R.D. (Eds.), Emerging Technologies and Information Systems for the Knowledge Society. Springer. volume 5288, pp. 354—363. Shani, G., Gunawardana, A., 2011. Evaluating recommendation systems, in: Ricci, F., Rokach, L., Shapira, B., Kantor, P.B. (Eds.), Recommender Systems Handbook. Springer, pp. 257–297. Sieg, A., Mobasher, B., Burke, R., 2007. Ontological user profiles for representing context in web search, in: 2007 IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology Workshops, IEEE. pp. 91—94. Speretta, M., Gauch, S., 2005. Personalized search based on user search histories, in: Proceedings of the 2005 IEEE/WIC/ACM International Conference on Web Intelligence, IEEE. pp. 622—628. Sugiyama, K., Hatano, K., Yoshikawa, M., 2004. Adaptive web search based on user profile constructed without any effort from users, in: Proceedings of the 13th International Conference on the World Wide Web WWW ’04, New York, NY, USA. p. 675. Sun, J., Xie, Y., 2009. A recommender system based on web data mining for personalized E-Learning, in: Proceedings of the International Conference on Information Engineering and Computer Science, 2009. ICIECS, IEEE. pp. 1—4. Trajkova, J., Gauch, S., 2004. Improving Ontology-Based user profiles, in: Proceedings of Large Scale Semantic Access to Content (Text, Image, Video, and Sound) RIAO, pp. 380—389. Weng, S.S., Chang, H.L., 2008. Using ontology network analysis for research document recommendation. Expert Systems with Applications 34, 1857–1869. Wu, K.L., Aggarwal, C.C., Yu, P.S., 2001. Personalization with dynamic profiler, in: Advanced Issues of E-Commerce and Web-Based Information Systems, Third International Workshop on WECWIS 2001, IEEE. pp. 12–20. Xu, K., Zhu, M., Zhang, D., Gu, T., 2008. Context-aware content filtering & presentation for pervasive & mobile information systems, in: Proceedings of the 1st International Conference on Ambient Media and Systems, pp. 1–20. Yang, Y., Padmanabhan, B., Rajagopalan, B., Deshmukh, A., 2005. Evaluation of online personalization systems: A survey of evaluation schemes and a knowledge-based approach. Journal of Electronic Commerce Research 6, 112–122. Zhang, H., Song, Y., Song, H., 2007. Construction of Ontology-Based user model for web personalization, in: Proceedings of the User Modeling Conference, pp. 67—76. Zhuhadar, L., Nasraoui, O., 2008. Personalized cluster-based semantically enriched web search for e-learning, in: Proceedings of the 2nd International Workshop on Ontologies and Information Systems for the Semantic Web, ACM, New York, NY, USA. pp. 105—112. AppendixA. The Evaluation Scenarios In this appendix, we present the five scenarios that were used as part of the first evaluation to address different user browsing behaviours. Scenario (1): ‘Normal behaviour’: In the first scenario, we look at how our model would learn and adapt to normal behaviour that consists of uninteresting topics and short-term and long-term interests. Figure A.16 represents this scenario. This scenario consists of a total of 12 topics and 126 tasks (i.e. some of the tasks are repeated in different days). Users were asked to complete as many tasks as possible during 20 days. All the tasks in this scenario have been designed to simulate three browsing behaviours. The first behaviour is browsing uninteresting topics (e.g. topics 2, 4, 6, 8, 10 and 12 in table 1). The tasks in these topics occur in a short period and their occurrence is very low. In the second behaviour, we simulate a browsing behaviour when a user has long-term interests in some topics. In this scenario, topics 1 and 3 simulate this behaviour 37 Figure A.16: Scenario (1): Normal Behaviour. Figure A.17: Scenario (2): Browsing-stopped. as both these topics occur over a long period and their occurrence is very high. In the final behaviour, we simulate a user behaviour when he has short-term interests in some topics. Examples of this behaviour are topics 5, 7, 9 and 11 in Figure A.16 which appear for a short period and their occurrence is high. Scenario (2): ‘Browsing-stopped behaviour’: The second scenario again examines how our system would learn and adapt to normal behaviour. However, the difference in this scenario is that in the second third of the experiment’s duration (i.e. from day 8 to day 12) users will not be asked to complete any task. The main aim of such behaviour is to see how our system reacts when users suddenly stop browsing the Internet for a short while and then return to their normal browsing behaviour (i.e. from day 13). Figure A.17 shows the ‘browsing-stopped’ behaviour scenario. Scenario (3): ‘Gradual interest-drift behaviour’: Scenario (3) examines interest drift. For this purpose, we simulate a browsing behaviour where user interests are gradually changed from a set of topics to new topics. In order for the interest drift to be gradual, we designed the tasks in such a way so that short-term topics (i.e. topics 3, 5, 7, 9 and 12) start as interesting topics, and then they become less interesting after a while and just before a user starts to drift his interests to new topics. Figure A.18 shows the gradual interest drift scenario. Scenario(4): ‘Sudden interest-shift behaviour’: Scenario (4) has been designed to test our system 38 Figure A.18: Scenario (3): Gradual interest-drift. Figure A.19: Scenario (4): Sudden interest-shift. when user interests suddenly shift to new ones. In this scenario, the interest-shift occurs three times: from day 1 to day 6, from day 7 to day 11 and from day 12 to day 20 (see Figure A.19). In order to simulate the interest-shift behaviour, we designed all the tasks to be stopped suddenly, and new tasks about new topics take place abruptly. Our system would have to adapt to such behaviour in terms of recognizing new interests and forgetting uninteresting concepts rapidly. Scenario (5): ‘Interrupted browsing behaviour’: In scenario (5), we simulate a situation when a user’s regular behaviour is interrupted. In addition, a user might have regular interests on some topics, but in some situations he might lose this interest temporarily as he might get interested in new topics. But after a while, the user would get back to his regular behaviour and get interested again in the previous topics. For example, a user might have a long-term interest in reading topics related to his work, but if he decided to take a holiday in France for example, his interests would probably shift to topics related to France. However, once he gets back from his holiday, he would most likely return to his previous behaviour by browsing again for the earlier topics that related to his work. Figure A.20 shows this scenario. 39 Figure A.20: Scenario (5): Interrupted browsing behaviour. AppendixB. Example of adaptation of ontological user profile In this appendix, we present a brief example of how a user’s ontological profile adapts based on the interaction of the user with the system and through browsing topics. This example has been taken from a real user profile that was created during our experiment as described in section 5. The profile shown in Figures B.21-B.23 represents only a fraction of the actual user profile and is provided for illustrative purposes. The numbers indicated alongside the concepts are the interest weights. Figure B.21 illustrates a fragment of the user’s profile after this has been mapped to the reference ontology ODP on Day 1 of the experiment. The user has the following short-term interests: HCI, Dell, CSS, and Data mining. The average accuracy of the user profile on that day was 0.66 and this was due to the limited information present in the user’s profile. In Figure B.22, we can see the adapted fragment of the user profile on Day 5 of the experiment. Our approach captured the users interests as being in: HCI, Dell, CSS and XML. By Day 5 certain concepts that were ascertained as being of no interest to the user such as Data Mining and Database were forgotten as they just appeared only once during Day 1. The concept Directories was also forgotten as the user showed more interest on the HCI concept. The accuracy during day 5 was in fact close to 1 for this specific user. Finally, Figure B.23 shows the adapted user profile by Day 15 of the experiment. By Day 15, the user profile has experienced a number of changes: • As the user developed new short-term interests (AI and AJAX) this caused the accuracy to drop. The logfiles reveal that there was much noise in the collected user interests which caused the Insert Agent to add wrong concepts to the user profile (i.e. Javascript, Agent and Fuzzy). • Long-term interests were recognized as XML and HCI as they recorded high frequency weights (.793 and .778 respectively) compared to the short-term interests. • The Forget Agent processed and forgot three concepts that used to be short-term interests (i.e. Data Mining, Dell and CSS). • Two concepts were deleted from the user profile as the user did not show any interest in them. These concepts are Database and Directories and their frecency weights were less than the calcu- lated threshold for concepts to remain in the user profile (even as forgotten concepts). 40 Figure B.21: Day 1: Ontological user profile after mapping user interests to reference ontology. Figure B.22: Day 5: Adapted ontological user profile after interaction with the system. 41 Figure B.23: Day 15: Adapted ontological user profile after further interaction with the system. 42 
work_663xupkdkzhanoy2e2fbo7p2jm	----	Self-Adaptive Attribute Weighting for Naive Bayes Classification Jia Wua,b, Shirui Panb, Xingquan Zhuc, Zhihua Caia, Peng Zhangb, Chengqi Zhangb aSchool of Computer Science, China University of Geosciences, Wuhan 430074, China. bQuantum Computation & Intelligent Systems (QCIS) Centre, Faculty of Engineering & Information Technology, University of Technology Sydney, NSW 2007, Australia. cDepartment of Computer & Electrical Engineering and Computer Science, Florida Atlantic University, Boca Raton, FL 33431, USA. Abstract Naive Bayes (NB) is a popular machine learning tool for classification, due to its simplicity, high computational efficiency, and good classification accuracy, especially for high dimensional data such as texts. In reality, the pronounced advantage of NB is often challenged by the strong conditional independence assumption between attributes, which may deteriorate the classification perfor- mance. Accordingly, numerous efforts have been made to improve NB, by using approaches such as structure extension, attribute selection, attribute weighting, instance weighting, local learning and so on. In this paper, we propose a new Artificial Immune Sys- tem (AIS) based self-adaptive attribute weighting method for Naive Bayes classification. The proposed method, namely AISWNB, uses immunity theory in artificial immune systems to search optimal attribute weight values, where self-adjusted weight values will alleviate the conditional independence assumption and help calculate the conditional probability in an accurate way. One noticeable advantage of AISWNB is that the unique immune system based evolutionary computation process, including initialization, clone, section, and mutation, ensures that AISWNB can adjust itself to the data without explicit specification of functional or distribu- tional forms of the underlying model. As a result, AISWNB can obtain good attribute weight values during the learning process. Experiments and comparisons on 36 machine learning benchmark data sets and six image classification data sets demonstrate that AISWNB significantly outperforms its peers in classification accuracy, class probability estimation, and class ranking performance. Keywords: Naive Bayes, Self-Adaptive, Attribute Weighting, Artificial Immune Systems, Evolutionary Computing 1. Introduction Naive Bayes (NB) (Friedman et al., 1997), a special Bayesian network, is a Bayes’ theorem oriented learning model par- ticularly useful for learning tasks involving high dimensional data (Hernández-González et al., 2013), such as text classifi- cation (Kim et al., 2006; Chen et al., 2009) and web mining (Zhang et al., 2009). In general Bayesian models, the clas- sification is derived by using the dependency (or conditional dependency) between random variables. This process is typi- cally time consuming because examining relationships among all random variables is a combinatorial optimization task. Al- ternatively, NB relaxes the restriction of the dependency struc- tures between attributes by simply assuming that attributes are conditionally independent, given the class label. As a result, examining relationships between attributes is no longer needed and the derivation of an NB model can linearly scale to the training data. In reality, attributes in many learning tasks are correlated to each other, so NB’s conditional independence assumption may impair its classification performance (Webb et al., 2012). Email addresses: jia.wu@student.uts.edu.au (Jia Wu), shirui.pan@student.uts.edu.au (Shirui Pan), xzhu3@fau.edu (Xingquan Zhu), zhcai@cug.edu.cn (Zhihua Cai), peng.zhang@uts.edu.au (Peng Zhang), chengqi.zhang@uts.edu.au (Chengqi Zhang) In order to relax the conditional independence assumption and simultaneously retain NB’s efficiency, many approaches have been proposed by using solutions in five main categories: (1) structure extension (Liu et al., 2011; Jiang et al., 2012a); (2) attribute weighting (Zaidi et al., 2013; Wu et al., 2013a); (3) attribute selection; (4) instance weighting; and (5) instance se- lection. In this paper, we propose to use attribute weighting to mitigate NB’s primary weakness (the conditional indepen- dence assumption of the attributes) by assigning a weight value to each individual attribute. Because weight values enforce at- tributes to play different roles in classification, the correspond- ing Weighted Naive Bayes (WNB) will help relax the condi- tional independence assumption and make NB efficient for data (Wu et al., 2007) with strong attribute correlations. Indeed, the objective of assigning different weight values to attributes share striking similarity as feature selection, where the later intends to discover a subset set of features (with equal importance) to train a classification model. Assume weight values of all attributes are suitably determined, feature selec- tion can be achieved by using a subset of features ranked in a preference order. Therefore, feature weighting can be consid- ered as a generalization of feature selection, and many methods have been using feature selection to help improve NB classi- fication. For example, Langley & Sage (1994) proposed the Selective Bayes Classifier (SBC), by using feature selection to accommodate redundant attributes in the prediction process Preprint submitted to ELsevier September 6, 2014 and to augment Naive Bayes with the ability to exclude at- tributes that introduce dependencies. Meanwhile, in order to discover proper weight values for weighted NB classification, researchers have proposed many useful methods to evaluate the importance of attributes, including gain ratio (Zhang & Sheng, 2004), correlation-based algorithm (Hall, 2000), mutual infor- mation (Jiang et al., 2009, 2012b), and ReliefF attribute rank- ing algorithm (Robnik-Šikonja & Kononenko, 2003). Zhang & Sheng (2004) investigated the gain ratio based weighting scheme and several wrapper based methods for finding attribute weights in order to improve the Area Under Curve (AUC), which is a common metric used to compare algorithms by comparing their performance with respect to different param- eter settings. Hall (2007) proposed a new attribute weight- ing method to improve the AUC value, where the weights as- signed to the attributes are inversely proportional to the mini- mum depth at which the attributes are first tested in an unpruned decision tree. The above methods for weighted Naive Bayes classification have achieved good performance to solve domain specific prob- lems, through the employment of some external criteria, such as gain ratio, to determine suitable weight values for attributes. By doing so, the assessment of the attributes and the deriva- tion of the NB models are separated into two steps, with the attribute weights being determined without taking the NB ob- jective function into consideration. In order to address the prob- lem and seamlessly integrate attribute weighting and NB learn- ing into an integrated process, in this paper we first carry out a systematic experimental analysis for the existing improved al- gorithms for naive Bayes via attribute weighting (WNB), and then propose a new method to automatically calculate opti- mal attribute weight values for WNB, by directly working on WNB’s objective function. To this end, we employ an evolu- tionary computation based method, namely Artificial Immune System (AIS) (Er et al., 2012; Cuevas et al., 2012; Haktanir- lar Ulutas & Kulturel-Konak, 2012), to assign proper weight values for NB classification. In our previous study (Wu et al., 2013b), we have successfully proposed an AIS based method to automatically and self-adaptively select optimal terms and values for probability estimation. By proposing evolution com- putation for attribute weighting, our method in this paper will further advance weighted Naive Bayes to ensure that attribute weighting can automatically adapt to different learning tasks. In order to enable adaptive attribute weighting for NB clas- sification, we will propose to use AIS mechanism to adaptively determine attribute weights, where an automated search strat- egy is used to find optimal attribute weight values for each data set. The unique immune system computation processes, including initialization, clone, mutation, and selection, ensure that our method can adjust itself to the data without any explicit specification of functional or distributional form of the underly- ing model. Experiments and comparisons, on 36 UCI machine learning benchmark data sets (Bache & Lichman, 2013) and six image classification data sets (Li & Wang, 2008), demon- strate that the proposed artificial immune systems based weight- ing scheme for Naive Bayes classification (AISWNB) can suc- cessfully find optimal weight combinations for different learn- ing tasks, and its performance consistently outperforms other state-of-the-art NB algorithms. The corresponding superior- ity is demonstrated through three major performance metrics, including classification accuracy, class probability estimation, and class ranking performance (Zhang & Su, 2004; Jiang et al., 2009, 2012a). AISWNB is a self-learning algorithm by utilizing the im- munological properties, such as memory property and clonal selection. In contrast to the conventional statistical probabilis- tic evaluation in NB, the niche and advantages of AISWNB can be understood from the following four aspects: • AISWNB is a data-driven self-adaptive method because it does not require explicit specification of functional or distributional form of the underlying model. • AISWNB is a nonlinear model and is flexible in modeling complex real-world relationships. • AISWNB inherits the memory property of human immune systems and can recognize the same or similar antigen quickly at different times. • AISWNB can self-adaptively select suitable affinity func- tions to meet different types of learning tasks. The remainder of the paper is organized as follows. Section 2 reviews related work. Preliminary concepts and problem state- ments are addressed in Section 3. Section 4 introduces our new AISWNB framework, followed by the experiments in Section 5. We conclude the paper in Section 6 . 2. Related Work By proposing to use artificial immune system based method to search optimal weight values for weighted naive Bayes clas- sification, our method is related to attribute weighting in ma- chine learning and AIS based evolutionary computation. 2.1. Attribute Weighted Methods In real-world learning tasks, attributes often play different roles for classification. Therefore, assigning different weight values to attributes can potentially help improve the classifica- tion performance. During the whole process, the way of learn- ing the attribute weights plays an essential role. In this subsec- tion, we review existing work on attribute weighting by sepa- rating them into two categories: methods which consider each single attribute’s correlation with the class, and methods which consider multiple attributes’ joint correlations with the class. 2.1.1. Attribute Weighting via Single Attribute Correlation Mutual Information (MI) between two random variables pro- vides a quantity measure to evaluate the mutual dependence of two variables. A high MI value indicates a large reduction in uncertainty and low MI reflects a small reduction, and a zero MI value between two random variables means that the variables are independent. Friedman et al. (1997) provides a complete definition of mutual information between a pair of variables. 2 Mutual information has a long history of being used for mea- suring correlation between attributes and the class variable for classification. For instance, Jiang et al. (2012b) applied mutual information help to improve the accuracy of AODE (Averaged One-Dependence Estimators). Jiang et al. (2009) proposed a Hidden Naive Bayes (HNB) classifier, which uses MI based at- tribute weighting method to weight one-dependence estimators. Han et al. (2001) proposed a new algorithm called AWKNN (Attribute Weighted K-Nearest-Neighbor). In our experiments, we will apply mutual information to calculate the weight value between each attribute and the class attribute for WNB, and will use this approach, MIWNB, as a baseline for comparisons. Information Gain (IG), originally used by Quinlan (1993) in the decision tree leaning algorithm, is a commonly used mea- sure to evaluate the correlation of the attribute to the class. A notable drawback of IG is that the resulting score is biased to attributes with a large number of distinct values. Accordingly, information gain ratio (Quinlan, 1993) was proposed to solve the drawback by dividing each attribute’s IG score by the infor- mation encoded in each attribute itself. Zhang & Sheng (2004) argued that an attribute with a higher gain ratio value deserves a larger weight in WNB. In their studies, they proposed a gain ratio weighted method that calculates the weight of an attribute from a data set. 2.1.2. Attribute Weighting via Multiple Attribute Correlation Correlation-based Feature Selection (CFS) for attribute weighing uses a correlation-based heuristic evaluation function as an attribute quality measure (Hall, 2000) to calculate the weight value of each attribute. It uses a best-first search to tra- verse the feature space. CFS starts with an empty set and gener- ates all possible single feature expansions. The subset with the highest evaluation is selected and expanded in the same man- ner by adding new features. If expanding a subset results in no improvement, the search drops back to the next best unex- panded subset and continues from there. The best subset found is returned after the search terminates. The core of CFS is the heuristic process that evaluates the worth or “merit” of a feature subset. Hall (2007) employed this method to evaluate the importance of attributes according to the heuristic “merit” value. Relief is a feature selection method based on attribute estima- tion (Kira & Rendell, 1992). Relief assigns a grade of relevance to each feature by examining the change of the feature values with respect to instances within the same class (i.e., the near- est hit) and instances between classes (i.e., the nearest miss). If a feature’s values remain relatively stable for instances within the same class, the feature will receive a higher weight value. The original Relief only handles binary classification problems. Its extension, Relief-F, can be applied for multi-class classi- fications (Kononenko, 1994). Besides, Tucker et al. (2010) applied Relief-F attribute weighted approach to deal with top- down product, which is an engineering optimization problem. Attribute Correlation-based Weighting is a method which explicitly considers the correlation of each attribute to all other attributes to calculate the attribute’s weight value (Hall, 2007). Selection AntigenB-cell Differentiation M M Proliferation/ Clone Matuation M e m o ry c e lls P la s m a c e lls Figure 1: A conceptual view of immune response in immune systems: A B-cell contains the antibody (the middle rings on the left) that allows it to recognize the antigen (triangle), which denotes pathogenic materials invading to the sys- tem. The binding between B-cell and antigen can be evaluated by using cer- tain affinity (i.e., degree of binding). In a learning system, this resembles to the assessment of how good a solution (i.e., antibody) recognizes/resolves the training data (i.e., antigen). After the recognition, the system will respond and result in proliferation, differentiation, and maturation process of the B-cell as secondary antibodies. The secondary antibodies with high affinity becomes a memory cell, and others become plasma cells. Memory cells are retrained in the system to allow faster response to the same (or similar) attacks in the future (if the body is re-infected by the same pathogenic materials). A large weight value will be assigned to the attributes with strong dependencies on other attributes. In order to estimate each attribute’s dependence, an unpruned decision tree is con- structed from the training instances with a minimum depth, which indicates the depth for testing the tree. The weight as- signed to each attribute is inversely proportional to the mini- mum depth at which they are first tested in an unpruned deci- sion tree. Attributes that do not appear in the tree receive zero weight values. 2.2. Artificial Immune Systems 2.2.1. Human Immune System The human immune system contains two major parts: (1) hu- moral immunity, which deals with infectious agents in the blood and body tissues, and (2) cell-mediated immunity, which deals with body cells that have been infected. In general, the humoral system is managed by B-cells (with help from T-cells), and the cell-mediated system is managed by T-cells. Each cell (B or T) has a unique type of molecular receptor (location in shape space), which allows for the binding of the antigens (shown as triangles in Fig. 1). A higher affinity between the receptor and antigens indicates a stronger binding. In immunology, immune system contains two types of lym- phocyte cells (B- and T-cells), each of which has a unique type of molecular receptor allowing others to bind to them. When pathogens (i.e., biological agents that may cause diseases or ill- ness) invade the body, antibodies which are produced from B- cells are response for the detection/binding of a foreign protein or antigen (i.e., pathogenic materials). Once the binding be- tween B-cells and antigens are established, B-cells will undergo a series of process including proliferation, differentiation, and maturation, and eventually result in memory cells. The memory cells are retrained in the system to allow faster response to the same (or similar) attacks in the future (if the body is re-infected by the same pathogenic materials). This response process could 3 be explained by clonal selection theory, and the conceptual re- view is shown in Fig. 1. In this paper, humoral immunity is delegated to the natural immune system and the action of T- cells is not explained. The clonal selection followed by the B- cells of human immune system is the fundamental mechanism on which Artificial Immune Systems (AIS) is modeled. 2.2.2. AIS: Artificial Immune Systems Artificial Immune Systems denotes a class of evolutionary computational methods which intend to exploit and simulate the functions and behaviors of mammalian immune systems’ learn- ing and memorization capability to solve a learning task. The theme of an AIS is to resemble a biological immune systems’ ability to distinguish foreign molecules (or elements) which can attack/damage the body, and provide a learning system with ca- pability of distinction between self vs. non-self. This capability eventually leads to the assessment of the fitness scores of can- didates with respect to the underlying system. More specifically, AIS consists of three major components, including representation, recognition, and clone selection when dealing with learning algorithms. The representation, known as shape-shape problem, focuses on how to model antibodies and antigens. When the immune system is attacked by antigen, antibodies try to neutralize the infection by binding to the anti- gen through the recognition process. Binding strength, also re- garded as affinity, is used as a threshold for the immune system to respond to the antigen. The clone selection is corresponding to an affinity maturation process, which means immune individ- uals with high affinity will gradually increase during clone and mutation process. At the same time, some immune individuals will polarize into memory individuals. Similar to the AIS, evolutionary algorithms (EAs), such as Genetic Algorithms (GA) (Park & Ryu, 2010), Evolution Strategies (ES) (Huang et al., 2011) and Differential evolution (DE) (Storn & Price, 1997) are all designed based on the basic idea of biological evolution to control, and optimize artificial systems. Evolutionary computation shares many concepts of AIS like a population, genotype phenotype mapping, and pro- liferation of the most fit. On the other hand, AIS models based on immune networks resemble the structures and interactions of neural network models. The key advantages of AIS over neural networks are the benefits of a population of solutions and the evolutionary selection pressure and mutation. Mean- while, the underlying mechanisms are fundamentally different in many aspects. First and foremost, the immune system is highly distributed, highly adaptive, self-organising, maintains a memory of past encounters and has the ability to continu- ously learn about new encounters. AIS is the system devel- oped around the current understanding of the immune system. Second, AIS is a general framework for a distributed adaptive system and could, in principle, be applied to many domains. Compared to most other evolutionary algorithms, AIS is much more simple and straightforward to be implemented, which is important for practitioners from other fields. In addition, be- cause AIS is self-organizing, it requires much less system pa- rameters than other evolutionary computation methods. Some works have also pointed out the similarities and the differences between AIS and other heuristics (Zheng et al., 2010; Aickelin et al., 2013; Castro & Timmis, 2002). In recent years, there has been considerable interests in ex- ploring and exploiting the potential of AIS for applications in computer science and engineering including pattern recognition (Yuan et al., 2012), clustering (de Mello Honorio et al., 2012), optimization (Woldemariam & Yen, 2010), and Remote Sens- ing (Zhong & Zhang, 2012). However, the advantage of AIS for Bayesian classification has received very little attention. In this paper, we propose a new AIS based attribute weighting method for Naive Bayes classification. The performance of this design is validated through numerous performance metrics, including classification accuracy, class probability estimation, and class ranking performance. It is worth noting that some works exist to improve AIS for domain specific problems, such as an im- proved artificial immune system for seeking the Pareto front of land-use allocation problem in large areas (Huang et al., 2013). However, in this paper, we do not consider the improved AIS for WNB. This is mainly because that we aim at proposing a self-adaptive attribute weighting framework based on the im- mune system for WNB, and our designs can be easily general- ized to any AIS based algorithms. 3. Preliminaries and Problem Definition Given a training set D = {x1, · · · , xN} with N instances, each of which contains n attribute values and a class label, we use xi = {xi,1, · · · xi, j, · · · xi,n, yi} to denote the ith instance xi in the data set D. xi, j denotes the jth attribute value of xi and yi denotes the class label of xi. The class space Y = {c1, · · · , ck, · · · , cL} denotes the set of labels that each instance belongs to and ck denotes the kth label of the class space. For ease of understanding, we use (xi, yi) as a shorthand to represent an instance and its class label, and use xi as a shorthand of xi. We also use a j as a shorthand to represent the jth attribute. For an instance (xi, yi) in the training set D, its class label satisfies yi ∈Y, whereas a test instance xt only contains attribute values and its class label yt needs to be predicted by a weighted naive Bayes classification model, which can be formally defined as c(xt) = arg max ck∈Y P(ck) n∏ j=1 P(xt, j |ck) w j (1) In Eq. (1), p(ck) represents the prior probability of class ck in the whole training set. P(xt, j |ck) denotes the conditional proba- bility distribution of attribute xt, j conditioned by the given class ck. w j denotes the weight value of the jth attribute. In this paper, we focus on the calculation of the conditional probability p(xi, j|ck)w j by finding optimal attribute weight val- ues w j, j = 1, · · · , n. While all existing attribute weighting ap- proaches define the weight without considering the uniqueness of the underlying training data, we intend to resolve the opti- mal w value selection problem as an optimization process. As- sume that the calculation of each conditional probability value p(xi, j|ck)w j has an optimal w j value, there are n w j vectors 4 In it ia l P o p u la ti o n Global Optimal Weight Clone (b) (c) Mutation (d) (f) (e) Re-selection of the weight with the best affinity :weight vector (a) individual wi lation is generated through certain random mechanisms. So ∗ (wtc best individual wc, corresponding to the obtained attribute Figure 2: A conceptual view of self-adaptive weighting strategy for AISWNB: An initial population contains many antibodies (i.e., weight vectors w) that allow themselves to recognize antigens (i.e., training instances Da) with certain affinity (i.e., classification performance on Db). After recognition, the system will respond and select the weight vector wtc (in the tth iteration) with the best affinity (a), and then clone it (b) to replace the weight vectors with low affinity (c). After that, the mutation strategy is adopted to maintain the diversity of the weight vectors (d). The mutation population will further replace the old population (e) to reselect the best weight vector as the memory antibody (a). Through the evolutionary process, the global optimal weigh vector wc will be obtained (f). needed for NB classification. As a result, the WNB classifica- tion can be transferred to an optimization problem as follows. w∗ = arg max w j∈w f (xt, w) s.t. 0 ≤ w j ≤ 1 (2) where w = {w1, · · · , w j, · · · , wn} denotes the attribute weight vector for WNB. And f (xt, w) is calculated by Eq. (1). 4. Self-Adaptive Attribute Weighted Naive Bayes 4.1. AIS Symbol Definitions and Overall Framework 4.1.1. AIS Symbol Definitions In this paper, we propose to use AIS to learn optimal attribute weight values for NB classification. In our solution, antigens in AISWNB are simulated as training instances which are pre- sented to the system during the training process. Antibodies represent attribute weight vector w with different set of values (i.e., candidates). The binding of the antibodies and antigens will resemble the fitness of a specific weight vector with re- spect to the given training data. This can be evaluated by using the affinity score. During the learning process, the antibodies with good affin- ity will experience a form of clonal expansion after being pre- sented with the training data sets (analogous to antigens). When antibodies are cloned they will undergo a mutation process, in which a specific mutation function will be designed (and de- ployed). The evolving optimization process of the AIS system will help discover optimal w vector with the best classification performance. Before introducing algorithm details, we briefly define fol- lowing key notations, which will help understand the learning of the weight values using AIS principle. In Table 1, we also summarize the mapping of the symbols between immune sys- tems and AIS based weighting scheme for Naive Bayes. • Antibodies: W represents the set of antibodies, W = {w1, · · · , wL}, where L represents the size of antibodies. wi = {wi,1, · · ·wi, j, · · ·wi,n} represents a single antibody (i.e., attribute weight vector). So wi, j will represent the jth value of the ith antibody wi. • Antigens: Da represents the set of antigens, Da = {xa1, · · · , x a Na }, where Na represents the size of antigens. xai represents a single antigen. In AISWNB, xai denotes an instance in the data set Da. • Affinity: A measure of closeness between antibodies and antigens. In the current implementation, this value is cal- culated as accuracy (ACC), the area under the ROC curve (AUC), or conditional log likelihood (CLL) on a given data set, when dealing with classification accuracy, rank- ing, and probability estimation learning tasks, respectively. • Memory Cell: wc represents the memory cell for the an- tibody which has the best affinity (i.e., best classification performance on the test data sets Db = {xb1, · · · , x b Nb }. • Clone rate: An integer value used to determine the number of mutated clones for a given antibody (weight vector). Specifically, a selected antibody is allowed to produce up to mutated clones with clone rate value after responding to a given antigen set. • Mutation rate: A parameter between 0 and 1 that indicates the probability of an antibody being mutated. For a given antibody, its mutation rate is equal to 1 minus its affinity. By doing so, the antibody with high affinity will have a low probability of being mutated. 4.1.2. AISWNB Overall Framework A conceptual view of the proposed self-adaptive weighting strategy for AISWNB is shown in Fig. 2. In our settings, we 5 Table 1: Symbol mapping between Immune System and AISWNB. Immune systems AISWNB Antibody Attribute weight vector w Antigens Training instances in Da Shape-space Possible values of the data vectors Affinity The fitness of the weight vector w on the testing datasets Clonal Expansion Reproduction of weight vectors that are well matched with antigens Affinity Maturation Specific mutation of w vector and removal of lowest stimulated weight vectors. Immune Memory Memory set of mutated weight vectors use antibody to simulate weight vector w of the naive Bayes models, so an initial population of random antibodies, which correspond to a set of random weight value vectors W, are se- lected. The antibodies will recognize the antigens (which cor- respond to the training instances Da) with certain affinity (i.e., classification performance on Db). This recognition process re- sembles to the assessment of evaluating how good the weigh value solutions fit the underlying training data. After the recognition process, the system will respond and select the weight vector wtc with a good affinity, which corre- sponds to step (a) in Fig. 2, and then clone some high promising weight vectors to replace some weight vectors with low affin- ity values. After that, the mutation strategy is carried out to maintain the diversity of the weight vectors, as shown on step (d) of Fig. 2. The mutation population will further replace the old population to reselect the best weight vector as the memory antibody. Through the repetitive evolutionary process, the final optimal weigh vector wc will be obtained, as show on step (f) in Fig. 2. The performance of a classifier is often measured by clas- sification accuracy (ACC). So one can use ACC to calculate the affinity as shown in the above process. In reality, many data mining applications also require the calculation of the class distributions and ranking of the classes, in addition to the classification accuracy (Zhang & Su, 2004; Jiang et al., 2009; Wu & Cai, 2014). In recent years, the area under the ROC curve (AUC) has been used by machine learning and data mining community, and researchers believe (Ling et al., 2003) that AUC is a more discriminant evaluation method than error rate for learning algorithms that also produce class probability estimations. For class probability estimation, conditional log likelihood (CLL) has also been used to evaluate the quality of class probabilities from a classifier (Grossman & Domingos, 2004). Meanwhile, some existing research works (Zhang & Sheng, 2004; Hall, 2007) have proposed to use attribute weight- ing to improve the AUC performance of NB. However, attribute weighting for Naive Bayes on accuracy performance or class probability estimation has received very little attention. Ac- cording to the experimental analysis in Section 5.2.1, most ex- isting attribute weighting approaches cannot work well on all the above mentioned three learning tasks. This is mainly be- cause that the attribute subset evaluation used in the traditional attribute weighted Naive Bayes (e.g., SBC and CFS), intends to maximize the classification accuracy, which may lead to mis- matching between the learning process and the other learning goals (Jiang et al., 2012a). In order to address this challenge, the affinity function in AISWNB can be dynamically adjusted to match the learning process and the learning goal. The de- tails related to the affinity function will be addressed in Section 4.2.2. 4.2. AISWNB: AIS based Attribute Weighted Naive Bayes The proposed AISWNB is achieved through the following two major steps: (1) use AIS algorithm to train models from the training instances, with the purpose of obtaining optimal at- tribute weight values; and (2) the test instances are classified by the AISWNB classifiers with the learned attribute weight val- ues. Algorithm 1 reports the details of the proposed AISWNB framework, which is described as follows: 4.2.1. Initialization During the initialization process, we generate a set of L weight candidates: W = {w1, · · · , wL}, where each individual wi = {wi,1, · · ·wi, j, · · ·wi,n} represents an antibody (i.e. a weight value vector with wi, j representing the weight value for the jth attribute). To generate random weight values for all candidates, we set each wi, j value as a uniformly distributed random vari- able within range (0, 1]. In our experiments (detailed in Section 5), we use 80% of instances in a given data set D as the antigens set Da to learn optimal weight values wc, and the remaining in- stances are used as the test set Db. 4.2.2. AISWNB Evaluation The AISWNB evaluation process intends to resemble the recognition and the evolution process of the immune systems to find good antibodies (i.e., weight vectors) as shown in Fig. 2. In a weighted NB learning context, the above process corresponds to finding and selecting good weight vectors, and then apply- ing clone and mutation actions to the selected weight vectors to generate new candidates. Some newly generated good can- didates will further be retained to train weighted naive Bayes networks. In the following, we briefly explain the actions in each indi- vidual step: • Calculation of affinity function: For the learning task concerning about the maximization of the classification accuracy (ACC), the affinity of the ith individual of the tth generation wti can be obtained by applying the current attribute weight vector wti to the WNB model, and then evaluate its affinity function as follows, f [wti] = 1 Nb Nb∑ i=1 δ[c(xbi ), y b i ] (3) In Eq. (3), c(xbi ) is the classification result of the ith in- stance in a test data set Db with Nb instances, by using an AISWNB classifier with attribute weight values wti. y b i is 6 Algorithm 1 AISWNB (Weighted Naive Bayes by AIS) Input: Clone Factor c; Threshold T ; Maximum Generation MaxGen; Antibody Population W; Antigen Population Da; Test affinity set Db; Output: The target class label c(xt ) of test instance xt ; 1: W ← The wi, j value of wi for each individual is initialized using a random number distributed between (0, 1]. 2: while t ≤ MaxGen and f [wt+1c ] − f [w t c] ≤ T do 3: f [wti ] ← Apply antigen population D a, test affinity set Db w.r.t. antibody wti , and calculate the affinity of w t i . 4: wtc ← Apply the sequence of each f [w t i ] to the whole antibody population Wt and find the wtc with the best affinity. 5: (Wr )t ← Select the temporary antibodies set with the lowest affinity with clone factor c. 6: (Wc)t ← Clone wtc with clone factor c and obtain clone anti- body set. 7: Wt ← [Wt − (Wr )t ] ∪ (Wc)t ; 8: for all each wti in W t do 9: vt+1i ← Apply w t c and a normally distributed random variable N(0,1) to wti and obtain the mutation individual. 10: wt+1i ← Apply v t+1 i to w t i and obtain the new individual in t + 1th generation. 11: end for 12: end while 13: c(xt ) ← Apply wc to instance xt to predict the underlying class label . the actual class label of the ith instance. δ[c(xbi ), y b i ] is one if c(xbi ) = y b i and zero otherwise. For learning tasks concerning about other types of learning goals, such as the ranking of the classes or the estimation of class probability distributions, the affinity function can be changed by using the corresponding evaluation crite- ria, AUC or CLL. So the attribute weight values can be learned in line with the underlying learning goals. The corresponding details are addressed in Section 5.1.3. • Antibody Selection: We sort the individuals in the initial antibody population according to the affinity of each indi- vidual, and choose the individual wtc with the best affinity performance in the tth generation as the memory antibody. • Antibody Clone: To ensure that the population size of every generation is fixed, the best individual wtc will be cloned under the clone factor c. After that, we use the clone set to replace the individuals with low affinity ac- cording to the same rate c. • Antibody Mutation: Using the mutation operation to treat the individuals in tth generation Wt. It means that we obtain the middle generation composed with the new variation individuals from the parent generation. For any individual wti from the tth generation, the new variation individual vt+1i can be generated as follows: vt+1i = w t i + F ∗ N(0, 1) ∗ (w t c − w t i) (4) Among them, N(0,1) is a normally distributed random variable within the range [0,1]. F, as the variation factor during the process of evolution, can be adaptively obtained according to the different clones (Zhong & Zhang, 2012). F = 1 − f [wti] (5) where f [wti] denotes the affinity of the ith individual from the tth generation. 4.2.3. AISWNB Updating To determine whether a variation individual vt+1i can replace a target individual vector wti, as a new individual w t+1 i for the t + 1th generation, AISWNB adopts a greedy search strategy. The target individual wti is replaced by v t+1 i , if and only if v t+1 i ’s affinity is better than that of wti. In addition, the system also chooses the individual wt+1c with the best affinity performance in the t + 1th generation as a new memory antibody. An unabridged evolutionary process for the population in- cludes Evaluation and Update, which continuously repeats un- til (1) the algorithm surpasses the pre-set maximum number MaxGen, or (2) the results obtained from two consecutive it- erations are less than the threshold (i.e., T). After obtaining the best individual wc (i.e., attribute weight value vector), we use the weight values to build a WNB classifier to classify test data. 4.3. Time Complexity The time complexity of AISWNB is mainly attributed to the following two processes: (1) evaluation of AISWNB, and (2) updating of the weight values. Prior to the evaluation of AISWNB model, an NB classi- fier needs to be trained from Da with Na instances, which will take O(Na · n), where n is the number of attribute (NB needs to scan the whole training set and build prior probabilities for all classes and conditional probabilities for all n attributes). For the weight population W in each generation, the calculation of affinity function for each weight individual w ∈ W is similar to testing an NB classifier on a test set Db with Nb instances, which will take O(Nb · n · L), where L is the size of the weight populations (i.e. the number of weight vectors). The rest four operations (e.g., selection, clone, mutation, and update) are all based on weight vectors. The corresponding time complexity is O(L · logL). Assume the average number of evolution genera- tions is M, the total time complexity U is given by Eq. (6). U = O(Na · n) + M × [O(Nb · n · L) + O(L · logL)] (6) Because Na+Nb = N, where N is the total number of training data, Eq. (6) can be rewritten as U =O[(N − Nb) · n] + O(Nb · n · L · M) + O(L · logL · M) ≤O(N · n) + O(Nb · n · L · M) + O(L · logL · M) ≤O(N · n · L · M) + O(L · logL · M) ≤O(N · n · L2 · M) (7) Eq. (7) shows that the total time complexity of AISWNB is bounded by four important factors: (1) the total number of train- ing samples N; (2) the number of the attribute n; (3) the size of 7 weight pollution L; and (4) the average number of evolution generations M. In our experiments, we use a threshold T to automatically determine the termination process by following the principe that if the results obtained from two consecutive iterations are less than T , the algorithm will terminate. This process will further reduce the number of iterations and save computational costs. 5. Experiments 5.1. Experimental Settings 5.1.1. Benchmark Data and Parameters We implement the proposed method using WEKA (Witten & Frank, 2005) data mining tool and validate its performance on 36 benchmark data sets from UCI data repository (Bache & Lichman, 2013) and six image classification data sets from Corel Image repository (Li & Wang, 2008). Because Bayesian classifiers are designed for categorical attributes, in our exper- iments, we first replace all missing attribute values using the unsupervised attribute filter ReplaceMissingV alues in WEKA. Then, we apply unsupervised filter Discretize in WEKA to dis- cretize numeric attributes into nominal attributes. The similar data preprocessing could also be found in previous works (Jiang et al., 2009, 2012a). The three parameters L, M and T in Algorithm 1 are set to 50, 50, and 0.001, respectively. All results are obtained via 10 runs of 10-fold cross validation, and our algorithm is carried out on the same training data sets and evaluated on the same testing data. Moreover, all experiments are conducted on a Linux clus- ter node with an Interl(R) Xeon(R) @3.33GHZ CPU and 3GB fixed memory size. 5.1.2. Baseline Methods For comparison purposes, we compare AISWNB with the following baseline methods: • NB: A standard Naive Bayes classifier with conditional at- tribute independence assumption (Friedman et al., 1997); • CFSWNB: An attribute weighted Naive Bayes based on correlation-based feature selection (Hall, 2000); • GRWNB: An attribute weighted Naive Bayes using gain ratio based feature selection (Zhang & Sheng, 2004); • MIWNB: An attribute weighted Naive Bayes using mu- tual information weighted method for feature selec- tion (Jiang et al., 2012b); • ReFWNB: An attribute weighted Naive Bayes using a feature selection method based on attribute estima- tion (Robnik-Šikonja & Kononenko, 2003); • TreeWNB: An attribute weighted Naive Bayes with the weighting method according to the degree to which they depend on the values of other attributes (Hall, 2007); • SBC: A bagged decision-tree based attribute selection fil- ter for Naive Bayes (Langley & Sage, 1994); Table 2: Detailed information of the 36 UCI benchmark data sets Data set Instances Attributes Classes Missing Numeric anneal 898 39 6 Y Y anneal.ORIG 898 39 6 Y Y audiology 226 70 24 Y N autos 205 26 7 Y Y balance-scale 625 5 3 N Y breast-cancer 286 10 2 Y N breast-w 699 10 2 Y N colic 368 23 2 Y Y colic.ORIG 368 28 2 Y Y credit-a 690 16 2 Y Y credit-g 1000 21 2 N Y diabetes 768 9 2 N Y Glass 214 10 7 N Y heart-c 303 14 5 Y Y heart-h 294 14 5 Y Y heart-statlog 270 14 2 N Y hepatitis 155 20 2 Y Y hypothyroid 3772 30 4 Y Y ionosphere 351 35 2 N Y iris 150 5 3 N Y kr-vs-kp 3196 37 2 N N labor 57 17 2 Y Y letter 20000 17 26 N Y lymph 148 19 4 N Y mushroom 8124 23 2 Y N primary-tumor 339 18 21 Y N segment 2310 20 7 N Y sick 3772 30 2 Y Y sonar 208 61 2 N Y soybean 683 36 19 Y N splice 3190 62 3 N N vehicle 846 19 4 N Y vote 435 17 2 Y N vowel 990 14 11 N Y waveform-5000 5000 41 3 N Y zoo 101 18 7 N Y • RMWNB: An attribute weighted Naive Bayes with the at- tribute weights randomly selected from (0, 1]; 5.1.3. Evaluation Criterion In our experiments, the selected algorithms are evaluated us- ing three performance metrics, including classification accu- racy (measured by ACC), class ranking performance (measured by AUC), and class probability estimation (measured by CLL). The ACC of each method is calculated by the percentage of correctly predicted samples in the test set. In some data mining applications, learning a classifier with accurate class ranking or class probability distributions is also desirable (Zhang & Su, 2004). For example, in direct market- ing, the limitation of the resources only allows promotion of the top x% customers during gradual roll-out, or different promo- tion strategies are deployed for customers with different likeli- hood of buying certain products. To accomplish these learning tasks, ranking customers according to their likelihood of buy- ing is more useful than simply classifying customers as: buyer or non-buyer (Jiang et al., 2012a). To evaluate the classifier performance in terms of class ranking and class probability dis- tributions, we use AUC and CLL, where AUC of the classifier is calculated as follows: 8 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 AISWNB N B (a) AISWNB vs. NB 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 AISWNB S B C (b) AISWNB vs. SBC 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 AISWNB C F S W N B (c) AISWNB vs. CFSWNB 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 AISWNB G R W N B (d) AISWNB vs. GRWNB 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 AISWNB M IW N B (e) AISWNB vs. MIWNB 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 AISWNB R e F W N B (f) AISWNB vs. ReFWNB 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 AISWNB T re e W N B (g) AISWNB vs. TreeWNB 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 AISWNB R M W N B (h) AISWNB vs. RMWNB Figure 3: AISWNB vs. competing algorithms: classification accuracy (ACC). Table 3: Detailed experimental results on classification accuracy (ACC) and standard deviation %. Data set AISWNB NB SBC CFSWNB GRWNB MIWNB ReFWNB TreeWNB RMWNB anneal 96.99±2.02 94.32±2.23 • 94.03±2.37 • 93.35±2.19 • 94.78±2.29 • 86.01±3.97 • 94.28±1.99 • 89.73±1.94 • 93.31±2.74 • anneal.ORIG 91.08±2.92 88.16±3.06 • 84.66±3.74 • 87.81±2.55 • 85.21±4.23 • 77.45±4.33 • 76.54±0.84 • 88.11±2.53 • 84.16±3.63 • audiology 75.49±8.00 71.40±6.37 70.91±7.09 69.84±6.46 • 71.58±6.76 70.82±6.99 61.22±6.63 • 71.92±6.67 62.67±9.22 • autos 70.13±11.42 63.97±11.35 • 70.10±9.38 68.41±10.53 65.14±10.92 65.52±11.44 65.58±9.57 67.76±10.51 64.01±11.24 • balance-scale 91.41±1.32 91.44±1.30 91.44±1.30 82.13±3.32 • 90.27±1.89 • 90.27±1.91 • 87.71±3.11 • 90.03±1.99 • 81.01±7.58 • breast-cancer 72.91±7.98 72.94±7.71 73.25±7.60 71.73±7.40 71.07±6.30 70.63±8.76 70.30±1.37 72.39±7.47 72.60±6.61 breast-w 97.24±1.68 97.30±1.75 97.30±1.75 97.21±1.84 97.28±1.78 97.33±1.77 97.25±1.91 97.34±1.81 96.94±2.00 colic 81.46±6.07 78.86±6.05 82.28±5.86 83.45±5.35 81.77±6.12 83.56±5.79 84.07±5.27 83.64±5.47 78.78±6.22 colic.ORIG 73.99±7.22 74.21±7.09 74.57±5.85 74.61±6.62 75.29±5.92 71.78±6.70 66.20±1.37 • 76.00±6.53 73.04±6.39 credit-a 85.13±3.90 84.74±3.83 85.75±4.16 86.14±4.06 85.51±3.96 85.51±3.96 85.67±3.97 86.46±3.85 82.32±5.85 credit-g 75.80±3.59 75.93±3.87 72.43±3.61 • 76.13±3.64 74.06±2.85 70.43±4.55 • 70.00±0.00 • 76.14±3.62 73.23±3.63 diabetes 75.86±4.87 75.68±4.85 75.93±5.07 77.02±4.87 75.03±3.95 75.94±5.49 65.16±0.47 • 76.91±5.07 73.70±5.30 glass 57.74±10.16 57.69±10.07 56.22±10.36 56.70±9.79 57.87±9.28 56.63±9.84 55.56±8.67 57.44±9.37 55.49±9.63 heart-c 82.41±6.66 83.44±6.27 84.20±6.37 83.64±6.37 84.14±6.19 83.50±6.60 83.20±6.83 83.57±6.04 81.48±7.15 heart-h 82.42±5.98 83.64±5.85 82.59±6.40 83.31±6.46 83.15±6.77 81.75±6.58 80.04±5.91 83.34±6.28 81.47±7.66 heart-statlog 83.52±6.19 83.78±5.41 84.19±6.07 84.22±5.99 83.59±6.52 83.41±6.08 81.44±5.97 84.04±5.90 81.85±6.24 hepatitis 84.52±9.61 84.06±9.91 83.60±9.77 84.52±9.22 82.97±9.89 85.15±9.45 82.28±4.54 83.35±8.24 85.22±8.87 hypothyroid 93.42±0.62 92.79±0.73 • 93.52±0.48 93.60±0.51 93.33±0.45 75.98±2.04 • 93.30±0.44 93.58±0.50 93.37±0.53 ionosphere 90.69±4.05 90.86±4.33 90.89±4.72 92.20±3.94 92.02±4.19 92.00±4.08 91.08±4.18 92.00±4.06 90.66±4.74 iris 94.87±6.28 94.33±6.79 96.87±4.29 95.33±5.40 95.93±4.73 95.93±4.73 96.07±4.65 95.53±5.19 91.87±7.38 kr-vs-kp 95.84±1.55 87.79±1.91 • 92.38±1.56 • 91.22±1.45 • 89.79±1.63 • 90.83±1.72 • 90.51±1.60 • 94.21±1.29 • 82.63±4.99 • labor 95.80±8.73 96.70±7.27 84.97±12.91 • 90.20±11.28 92.30±9.98 90.57±11.18 88.63±12.10 88.10±12.66 92.17±11.15 letter 67.75±2.17 65.80±2.04 • 67.32±2.22 66.23±2.07 • 68.03±2.10 68.33±2.15 68.50±1.99 65.99±2.08 • 60.17±3.54 • lymph 84.94±8.42 85.97±8.88 82.72±9.39 83.00±9.35 81.03±9.15 82.85±8.96 78.78±8.66 • 82.20±10.01 83.33±9.51 mushroom 98.32±0.99 93.58±2.03 • 98.12±1.14 98.28±1.04 98.33±0.98 98.21±1.01 97.83±1.25 97.82±1.10 93.54±3.18 • primary-tumor 46.52±6.28 47.20±6.02 45.78±6.84 43.78±5.05 46.77±6.08 45.64±6.93 24.78±1.47 • 45.54±5.39 42.60±6.44 segment 91.54±1.82 89.03±1.66 • 90.46±2.10 • 89.92±1.76 • 85.82±2.06 • 86.23±1.89 • 85.57±2.03 • 90.48±1.56 • 85.56±3.95 • sick 97.27±0.79 96.78±0.91 • 96.81±0.89 97.30±0.88 96.47±0.93 • 96.26±0.95 • 96.64±0.71 • 96.94±0.92 95.58±1.38 • sonar 76.76±10.78 76.35±9.94 73.54±9.45 75.71±9.58 74.99±9.54 76.14±9.55 76.39±9.45 75.36±8.81 75.22±9.93 soybean 92.94±2.54 92.20±3.23 91.00±3.31 • 92.61±2.82 90.61±3.40 • 91.74±3.28 88.58±3.46 • 92.85±2.90 89.90±3.56 • splice 95.71±1.07 95.42±1.14 95.84±1.03 95.39±1.28 94.08±1.28 • 94.50±1.21 • 88.30±2.44 • 96.14±1.03 92.61±2.17 • vehicle 62.81±3.81 61.03±3.48 56.32±4.01 • 60.59±3.57 60.39±3.35 58.91±3.35 • 61.16±3.31 61.75±3.44 60.11±3.42 vote 94.35±3.67 90.21±3.95 • 94.46±2.81 93.49±3.55 93.03±3.57 93.01±3.54 95.24±2.91 94.83±3.01 89.86±4.56 • vowel 67.26±4.86 66.09±4.78 62.75±5.10 • 65.32±4.59 64.82±4.63 64.65±4.66 66.14±4.67 66.65±4.73 57.41±7.79 • waveform-5000 82.60±3.52 79.80±2.97 • 79.71±2.97 77.78±2.92 • 78.71±3.05 • 78.65±3.05 • 80.43±3.38 79.56±3.00 • 78.38±3.52 • zoo 95.95±5.62 94.37±6.79 91.51±7.68 • 95.15±4.98 95.55±5.30 94.39±5.96 94.25±4.93 90.65±7.29 • 92.58±7.25 • :Statistically significant degradation. E = P0 − t0(t0 + 1)/2 t0t1 (8) where t0 and t1 are the number of negative and positive in- stances, repressively. P0 = ∑ ri, with ri denoting the rank of the ith negative instance in the ranked list. It is clear that AUC 9 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1 AISWNB N B (a) AISWNB vs. NB 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1 AISWNB S B C (b) AISWNB vs. SBC 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1 AISWNB C F S W N B (c) AISWNB vs. CFSWNB 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1 AISWNB G R W N B (d) AISWNB vs. GRWNB 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1 AISWNB M IW N B (e) AISWNB vs. MIWNB 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1 AISWNB R e F W N B (f) AISWNB vs. ReFWNB 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1 AISWNB T re e W N B (g) AISWNB vs. TreeWNB 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1 AISWNB R M W N B (h) AISWNB vs. RMWNB Figure 4: AISWNB vs. competing algorithms: area under the ROC curve (AUC). Table 4: The detailed experimental results on area under the ROC curve (AUC) and standard deviation %. Data set AISWNB NB SBC CFSWNB GRWNB MIWNB ReFWNB TreeWNB RMWNB anneal 98.88±1.67 98.76±1.84 98.27±2.54 98.66±1.89 98.63±1.90 98.82±1.73 98.77±1.77 98.58±1.96 98.74±1.81 anneal.ORIG 97.90±3.22 96.79±5.42 95.68±6.71 96.61±6.55 97.59±3.91 97.32±4.28 95.21±7.68 96.58±6.09 96.46±4.59 • audiology 84.32±1.56 83.85±1.44 84.05±1.61 84.06±1.59 83.75±1.45 84.26±1.56 83.98±1.58 83.99±1.59 83.43±1.37 • autos 94.68±3.27 91.96±3.32 • 94.26±3.13 93.78±3.40 92.98±3.68 93.27±3.73 93.50±3.48 94.44±3.70 92.39±3.85 balance-scale 89.76±4.24 85.00±4.03 • 85.00±4.03 • 76.84±4.78 • 76.29±3.84 • 83.64±3.79 • 74.67±3.75 • 82.40± 4.07 • 71.81±6.47 • breast-cancer 67.02±14.05 71.32±13.81 ◦ 69.45±14.71 69.93±13.90 70.10±14.49 70.61±13.08 67.13±12.81 70.98±13.89 ◦ 66.31±15.36 breast-w 99.13±0.94 99.23±0.83 99.23±0.83 99.19±0.87 99.20±0.88 99.16±0.89 99.19±0.90 99.29±0.76 98.89±1.21 colic 87.08±5.05 84.42±5.45 • 86.91±7.08 87.72±5.92 88.35±5.47 88.23±6.13 88.72±5.32 88.06±5.56 84.75±4.97 colic.ORIG 82.93±7.23 81.70±7.23 81.19±5.32 84.58±5.29 84.75±4.77 82.10±4.97 83.15±4.33 85.11±5.78 78.96± 6.56 credit-a 91.74±3.72 91.97±3.14 91.66±3.45 92.34±3.28 92.22±3.64 91.95±3.31 91.83±3.12 92.07±3.33 91.62±3.43 credit-g 79.22±5.20 79.42±4.52 74.26±5.55 • 79.13±5.01 77.90±5.62 78.15±5.84 77.91±6.15 79.64±4.85 78.45± 4.12 diabetes 84.44±4.68 82.74±4.94 • 84.08±5.04 84.10±4.66 84.00±4.80 83.81±4.80 83.79±4.64 83.67±4.81 78.49±4.81 • glass 88.43±3.18 82.63±6.07 • 83.33±7.29 84.25±4.60 • 81.84±5.59 • 84.97±6.12 85.63±4.36 • 82.90±5.23 • 81.11±5.19 • heart-c 84.17±0.66 84.17±0.50 84.05±0.61 84.13±0.54 84.07±0.57 84.05±0.57 84.15±0.56 84.14±0.55 83.79±0.67 heart-h 83.97±0.51 83.92±0.63 84.07±0.68 83.91±0.68 83.97±0.72 83.99±0.72 83.91±0.65 83.82±0.72 83.56±0.81 heart-statlog 91.00±4.80 91.33±5.15 89.61±5.01 91.39±4.17 91.28±4.84 91.28±4.13 91.56±3.94 91.39±4.26 88.28± 6.60 hepatitis 88.11±10.12 89.90±8.24 86.29±11.02 88.07±9.17 87.79±9.55 88.46±9.07 87.35±8.74 86.79±9.78 86.61±10.60 hypothyroid 89.05±9.74 87.53±9.20 • 85.43±8.19 • 86.96±8.70 • 86.30±9.11 • 88.46±9.98 83.68±7.65 • 87.03±8.80 • 85.60±8.07 • ionosphere 95.93±2.75 93.90±3.21 93.14±4.49 94.99±2.82 94.86±2.97 94.54±3.05 94.39±3.31 95.55±2.51 94.44± 2.89 iris 99.20±1.80 98.93±2.16 99.40±1.27 99.20±1.80 99.20±1.80 99.20±1.80 99.20±1.80 99.20±1.80 97.87±4.08 kr-vs-kp 99.29±0.36 95.20±1.28 • 96.94±0.83 • 97.75±0.92 • 98.10±0.77 • 97.95±0.81 • 98.48±0.69 • 98.72±0.67 • 87.00±2.22 • labor 98.75±3.95 98.75±3.95 80.63±33.13 100.00±0.00 100.00±0.00 92.50±23.72 90.00±27.51 96.25±8.44 95.00±12.08 letter 96.18±0.62 95.68±0.73 • 96.13±0.57 95.44±0.71 • 96.10±0.61 96.20±0.59 95.41±0.67 • 95.60±0.71 • 94.47±0.81 • lymph 94.88±4.92 95.01±4.87 94.64±4.87 95.30±5.32 94.67±4.71 94.81±4.92 92.99±4.15 94.10±4.65 94.68±4.74 mushroom 99.98±0.23 99.59±0.17 • 98.88±0.80 • 99.82±0.15 • 99.81±0.16 • 99.91±0.07 • 99.87±0.11 99.92±0.10 99.67±0.28 • primary-tumor 85.07±2.67 85.05±2.96 84.81±2.92 85.44±2.17 85.76±2.21 85.40±2.53 85.20±1.82 85.34±2.54 84.33± 2.64 segment 99.30±0.27 98.37±0.52 • 98.72±0.55 • 98.51±0.55 • 97.97±0.62 • 97.95±0.64 • 97.90±0.63 • 98.59±0.46 • 98.16±0.53 • sick 97.33±1.41 95.92±2.48 94.10±2.89 • 95.88±2.66 95.67±2.72 95.32±3.09 96.04±2.66 96.11±2.66 91.82±2.61 • sonar 87.86±10.08 86.79±9.83 81.11±11.83 • 84.52±10.52 84.61±10.17 85.27±10.16 85.47±10.10 83.05±9.80 86.03±10.79 soybean 99.97±0.07 99.90±0.07 • 99.87±0.11 99.93±0.07 99.91±0.07 99.92±0.06 99.89±0.07 99.92±0.06 99.68±0.19 • splice 99.50±0.21 99.41±0.22 • 99.45±0.26 99.45±0.27 99.27±0.29 • 99.32±0.27 • 99.40±0.27 99.54±0.22 98.34±0.34 • vehicle 84.41±3.74 80.85±3.73 • 78.53±4.19 • 80.57±4.24 • 79.21±4.19 • 78.91±3.95 • 80.64±4.60 • 81.50±3.92 • 80.46±3.73 • vote 98.93±1.10 96.79±1.95 • 98.25±1.64 98.02±1.44 97.82±1.58 97.94±1.47 98.53±1.35 98.56±1.12 95.15±2.45 • vowel 96.57±0.63 96.19±0.72 • 95.33±0.83 • 96.18±0.90 95.99±0.99 95.80±1.08 96.20±0.92 96.31±0.78 94.81±0.95 • waveform-5000 95.79±1.50 95.41±1.36 95.38±1.43 95.44±1.45 94.97±1.46 94.94±1.50 96.28±1.16 95.85±1.28 94.12±1.86 • zoo 98.57±1.66 98.57±1.66 97.86±2.37 98.57±1.66 98.57±1.66 99.05±1.23 98.57±1.66 97.86±2.37 98.57±1.66 ◦,•: Statistically significant upgradation and degradation, respectively. is essentially a measure of the quality of ranking. The above measure can only deal with two-class problem. For multiple classes, Hand & Till (2001) proposes an improved AUC calcu- lating measure: E′ = 2 g(g − 1) ∑ i< j<L E(ci, c j) (9) 10 −3 −2.5 −2 −1.5 −1 −0.5 0 −3 −2.5 −2 −1.5 −1 −0.5 0 AISWNB N B (a) AISWNB vs. NB −3 −2.5 −2 −1.5 −1 −0.5 0 −3 −2.5 −2 −1.5 −1 −0.5 0 AISWNB S B C (b) AISWNB vs. SBC −3 −2.5 −2 −1.5 −1 −0.5 0 −3 −2.5 −2 −1.5 −1 −0.5 0 AISWNB C F S W N B (c) AISWNB vs. CFSWNB −3 −2.5 −2 −1.5 −1 −0.5 0 −3 −2.5 −2 −1.5 −1 −0.5 0 AISWNB G R W N B (d) AISWNB vs. GRWNB −3 −2.5 −2 −1.5 −1 −0.5 0 −3 −2.5 −2 −1.5 −1 −0.5 0 AISWNB M IW N B (e) AISWNB vs. MIWNB −3 −2.5 −2 −1.5 −1 −0.5 0 −3 −2.5 −2 −1.5 −1 −0.5 0 AISWNB R e F W N B (f) AISWNB vs. ReFWNB −3 −2.5 −2 −1.5 −1 −0.5 0 −3 −2.5 −2 −1.5 −1 −0.5 0 AISWNB T re e W N B (g) AISWNB vs. TreeWNB −3 −2.5 −2 −1.5 −1 −0.5 0 −3 −2.5 −2 −1.5 −1 −0.5 0 AISWNB R M W N B (h) AISWNB vs. RMWNB Figure 5: AISWNB vs. competing algorithms: averaged conditional log likelihood (CLL). Table 5: The detailed experimental results on averaged conditional log likelihood (CLL) and standard deviation %. Data set AISWNB NB SBC CFSWNB GRWNB MIWNB ReFWNB TreeWNB RMWNB anneal -11.73±8.27 -15.69± 9.70 • -15.73±7.33 • -14.67±4.88 -17.17±10.48 • -78.00±39.77 • -19.06±3.85 • -21.47±3.93 • -16.39±5.44 • anneal.ORIG -25.95±9.51 -26.55± 8.07 -34.30±7.60 • -29.77±5.65 • -42.43±18.60 • -152.86±51.66 • -48.21±3.17 • -29.11±6.19 -36.35±4.96 • audiology -210.64±61.79 -290.06±91.97 • -111.32±24.90 ◦ -120.43±25.33 ◦ -265.53±85.38 • -272.67±103.55 -134.29±10.89 ◦ -100.85±18.51 ◦ -200.75±47.87 autos -114.18±51.49 -218.20±102.17 • -85.52±29.92 -88.69±32.47 -172.38±85.19 • -242.85±136.90 • -89.03±15.96 -83.84±30.97 -131.85±59.60 balance-scale -38.92±4.81 -50.83± 2.27 • -50.83±2.27 • -62.01±2.28 • -79.05±0.99 • -51.87± 2.24 • -82.57±0.91 • -56.55±2.16 • -68.96±1.97 • breast-cancer -58.21±14.01 -63.30±20.73 -62.27±18.57 -57.88±14.47 -55.79±7.11 -88.72±35.20 • -59.81±1.75 -57.16±13.78 -58.29±10.45 breast-w -10.93±8.17 -26.10±23.29 -26.10±23.29 -16.06±12.78 • -19.42±16.26 -29.87±25.75 • -11.75±8.98 -14.52±12.15 -17.54±14.12 colic -51.43±15.04 -82.31±26.04 • -45.03±15.68 -43.71±14.50 -76.49±33.29 • -114.24±59.00 • -48.55±2.67 -42.76±12.86 -56.87±16.17 colic.ORIG -53.74±15.12 -55.54±15.53 -47.46±7.72 -45.33±7.31 -47.45±11.37 -107.18±33.73 • -58.99±1.54 -45.18±8.06 -51.61±7.10 credit-a -35.52±7.02 -41.34±11.26 • -36.20±8.35 -35.36±8.19 -71.13±24.35 • -135.06±44.97 • -46.36±2.28 • -35.73±8.14 -37.44±8.19 credit-g -51.02±6.80 -52.42± 7.29 -53.45±5.47 -49.80±5.24 -51.48±3.74 -110.29±33.21 • -58.28±0.45 • -49.67±5.52 -51.05±3.23 diabetes -47.40±7.06 -53.18±10.46 • -47.43±6.96 -47.53±6.59 -51.03±2.71 -66.08±15.95 • -58.98±0.99 • -48.71±7.58 -53.29±5.43 • glass -110.74±16.18 -112.89±19.39 -105.31±23.28 -101.69±14.84 -104.39±18.75 -116.55±24.63 -117.45±6.44 -102.39±13.45 ◦ -111.84±14.83 heart-c -37.25±11.66 -45.10±16.99 • -42.37±13.85 -37.98±10.62 -40.89±13.51 -61.93±27.50 • -53.89±3.23 • -37.57±11.28 -45.11±12.67 • heart-h -39.29±10.79 -46.47±17.46 • -41.49±18.32 -41.16±15.97 -58.45±32.43 -110.01±61.17 • -50.58±3.53 • -43.24±17.01 -47.18±18.08 heart-statlog -39.77±11.61 -45.09±16.73 -45.77±12.75 -38.98±9.43 -42.12±12.19 -63.07±24.21 • -54.17±2.20 • -38.55±9.52 -45.27±13.86 hepatitis -42.24±19.93 -55.52±26.52 • -51.14±24.18 -37.18±15.13 -54.35±25.26 -90.12±47.24 • -36.72±5.87 -35.63±12.40 -36.06±16.12 hypothyroid -23.06±4.09 -25.83± 5.10 • -23.45±2.47 -22.77±3.48 -33.92±6.80 • -229.14±15.95 • -24.28±1.92 -22.94±3.71 -23.68±2.90 ionosphere -58.28±27.66 -99.71±39.08 • -28.78±11.61 -32.56±10.80 -69.17±25.88 • -102.56±38.33 • -27.78±8.98 ◦ -25.73±9.04 ◦ -45.20±15.87 iris -16.05±21.20 -17.09±18.44 -12.21±11.73 -14.31±13.53 -13.95±11.59 -15.09±19.29 -15.47±9.07 -15.03±10.38 -31.23±11.86 • kr-vs-kp -22.44±3.44 -29.17± 2.61 • -29.42±1.56 • -33.64±1.67 • -24.86±2.72 -23.07± 5.52 -51.42±0.90 • -33.97±1.40 • -43.70±2.89 • labor -32.62±58.47 -16.33±18.27 -53.33±55.16 -25.05±22.81 -26.12±26.51 -50.12±86.71 -40.81±12.46 -34.04±21.59 -25.18±25.41 letter -131.06±16.50 -141.18±13.19 • -126.77±10.75 -129.18±8.46 -135.78±11.98 -159.59±15.95 • -147.16±5.25 • -124.01±9.14 -140.77±10.48 • lymph -41.50±29.51 -43.67±21.80 -44.23±23.18 -40.17±16.98 -48.92±24.70 -77.21±46.06 • -61.21±7.59 -40.45±14.42 -43.34±14.44 mushroom -6.47±4.47 -21.35±10.02 • -7.94±3.93 -7.54±4.35 -14.55±9.96 • -21.49±15.63 • -7.47±3.38 -4.61±1.98 -9.45±5.63 primary-tumor -192.48±29.10 -192.48±29.10 -191.85±26.09 -189.01±17.16 -191.34±28.77 -205.01±35.39 -231.81±8.00 • -183.30±18.38 -196.23±20.31 segment -28.01±7.28 -53.79±16.55 • -29.68±7.00 -33.61±7.60 • -73.20±22.92 • -95.99±31.85 • -44.91±9.49 • -33.68±7.74 • -40.80±7.24 • sick -8.94±2.02 -12.13± 3.07 • -9.37±1.74 -9.05±2.35 -20.08±5.08 • -159.49±38.79 • -9.04±1.51 -9.47±2.34 -14.36±1.00 • sonar -87.88±49.02 -99.84±57.80 -73.80±31.79 -54.72±23.58 -112.51±61.06 -170.49±99.29 • -53.32±5.03 -53.80±17.28 -70.07±36.18 soybean -24.61±7.16 -38.71±14.36 • -39.72±15.49 • -28.24±6.95 -44.63±18.14 • -62.61±26.64 • -39.72±6.83 • -25.84±5.37 -35.62±8.96 • splice -14.32±2.97 -14.63± 2.70 -13.59±2.86 -15.11±1.79 -42.06±9.42 • -59.55±13.56 • -46.68±2.66 • -12.85±1.84 -26.96±1.78 • vehicle -103.14±20.56 -200.65±32.03 • -108.54±15.01 -113.99±17.20 -179.37±32.99 • -288.16±53.62 • -92.71±7.80 -128.15±18.98 • -144.67±21.50 • vote -19.57±11.92 -62.16±29.81 • -17.21±10.75 -25.04±14.94 -59.49±37.62 • -74.89±47.32 • -15.25±8.06 -18.01±11.83 -51.92±18.60 • vowel -86.72±9.97 -88.30± 9.00 -96.55±8.84 • -97.03±7.38 • -91.60±13.72 -112.48±22.63 • -141.45±3.53 • -92.19±6.77 • -106.41±6.30 • waveform-5000 -43.30±8.36 -74.37±17.55 • -66.35±15.03 • -48.84±7.92 -113.49±28.22 • -173.81±43.93 • -59.68±2.02 • -51.10±10.63 -54.77±12.34 • zoo -9.96±10.34 -11.82±10.79 -21.17±18.37 • -17.44±10.34 -9.65±9.97 -12.35±13.26 -19.57±10.44 • -34.69±14.08 • -18.46±13.37 ◦,•: Statistically significant upgradation and degradation, respectively. where g is the number of classes and E(ci, c j) is the AUC of each pair of classes ci and c j. In our experiments, CLL value of a classifier h on data set D with N instances is evaluated by using Eq. (10). CLL(h|D) = ∑N i=1 ∑n j=1 logPh(yi|xt, j) (10) 11 Table 6: Two-tailed t-test on classification accuracy (ACC). NB SBC CFSWNB GRWNB MIWNB ReFWNB TreeWNB RMWNB SBC 6/25/5 CFSWNB 8/25/3 4/28/4 GRWNB 5/25/6 2/30/4 2/27/7 MIWNB 5/22/9 2/26/8 2/26/8 2/28/6 ReFWNB 5/19/12 3/23/10 4/19/13 2/24/10 5/23/8 TreeWNB 7/26/3 6/27/3 4/30/2 8/24/4 10/25/1 11/23/2 RMWNB 1/23/12 1/25/10 0/22/14 0/28/8 3/25/8 6/24/6 1/20/15 AISWNB 11/25/0 10/26/0 8/28/0 9/27/0 11/25/0 14/22/0 8/28/0 15/21/0 Table 7: Two-tailed t-test on area under the ROC curve (AUC). NB SBC CFSWNB GRWNB MIWNB ReFWNB TreeWNB RMWNB SBC 4/25/7 CFSWNB 9/24/3 10/23/3 GRWNB 8/24/4 9/23/4 2/29/5 MIWNB 8/25/3 5/27/4 4/24/8 6/26/4 ReFWNB 5/27/4 8/24/4 3/30/3 7/25/4 6/26/4 TreeWNB 13/19/4 9/24/3 7/28/1 9/26/1 9/23/4 8/27/1 RMWNB 0/19/17 3/19/14 0/16/20 2/18/16 2/16/18 2/20/14 0/15/21 AISWNB 15/20/1 10/26/0 8/28/0 8/28/0 6/30/0 7/29/0 7/28/1 17/19/0 where h is a learning model. In Friedman et al. (1997), max- imizing Eq. (10) amounts to best approximate the conditional probability of Y given each text instance xt, and is equivalent to minimizing the conditional cross-entropy. 5.2. UCI Benchmark Learning Tasks We first test the performance of the proposed method on 36 UCI benchmark data sets (Bache & Lichman, 2013), which include a wide range of domains with data characteristics de- scribed in Table 2. These 36 UCI data sets have always been considered as benchmark data sets for data mining and machine learning related classification tasks (Jiang et al., 2009), in order to validate the effectiveness of new algorithm. In General, an algorithm is effective if it can significantly outperform its peers on at least 8 UCI data sets (Wu et al., 2013a,b; Jiang & Zhang, 2005; Zhang & Sheng, 2004). In order to demonstrate the performance of weighted NB, compared to the generic NB classifiers, our experiments will first report the performance of NB and WNB classifiers with re- spect to classification accuracy measured by ACC, ranking per- formance measured by AUC, and probability estimation mea- sured by CLL. More specifically, we compare an NB classi- fier (Friedman et al., 1997) with SBC (Langley & Sage, 1994), CFSWNB (Hall, 2000), GRWNB (Zhang & Sheng, 2004), MIWNB (Jiang et al., 2012b), RefWNB (Robnik-Šikonja & Kononenko, 2003) and TreeWNB Hall (2007). The purpose of the second experiment is to compare the pro- posed attribute weighted Naive Bayes, namely AISWNB, with each attribute weighted approach in literature. Finally, we com- pare related algorithms via two-tailed t-test with a 95% confi- dence level. Based on the statistical theory, the difference is statistically significant only if the probability of significant dif- Table 8: Two-tailed t-test on averaged conditional log likelihood (CLL). NB SBC CFSWNB GRWNB MIWNB ReFWNB TreeWNB RMWNB SBC 16/16/4 CFSWNB 23/8/5 12/18/6 GRWNB 11/15/10 6/9/21 3/7/26 MIWNB 1/8/27 2/2/32 2/3/31 1/5/30 ReFWNB 12/13/11 3/13/20 3/9/24 14/9/13 24/8/4 TreeWNB 25/4/7 12/17/7 11/16/9 22/10/4 31/2/3 21/11/4 RMWNB 14/13/9 2/18/16 0/7/29 17/13/6 28/5/3 17/9/10 3/5/28 AISWNB 21/15/0 8/27/1 6/29/1 17/19/0 29/7/0 18/16/2 7/26/3 16/20/0 ference is at least 95 percent, i.e., the p-value for a t-test be- tween two algorithms is less than 0.05. 5.2.1. Attribute Weighted NB vs. Standard NB Tables 3, 4, and 5 report the detailed results (the average ACC, AUC, and CLL values and the respective standard devia- tion values) of AISWNB and other baseline algorithms, respec- tively. In these three tables, the symbols ◦ and • represent sta- tistically significant upgradation and degradation over the pro- posed AISWNB with the p-value less than 0.05. In addition, Tables 6, 7, and 8 illustrate the compared results of two-tailed t-test, in which each entry w/t/l means that the algorithm in the corresponding row wins in w data sets, ties in t data sets, and loses in l data sets on the 36 UCI data sets, compared to the algorithm in the corresponding column. Overall, the results can be summarized as follows: 1. TreeWNB shows the best performance. Compared to NB, TreeWNB wins 7 data sets, ties 26 data sets, loses 3 data sets on ACC, and shows great superiority to NB on AUC (13 wins and 4 losses) and CLL (25 wins and 7 losses). 2. CFSWNB outperforms NB on ACC and AUC. Compared to NB, CFSWNB wins 8 data sets, ties 25 data sets and loses 3 data sets on ACC, with the better performance (9 wins and 3 losses) on AUC. But CFSWNB significantly outperforms NB on CLL (23 wins and 5 losses). 3. GRWNB slightly outperforms NB. The results show that GRWNB has a slightly better performance on AUC (8 wins and 4 losses), and almost ties NB both on ACC (5 wins and 6 losses) and CLL (11 wins and 10 losses). 4. SBC’s performance is comparable with NB. More specifi- cally, SBC almost ties NB (6 wins and 5 losses) on ACC, and is sightly worse than NB (4 wins and 7 losses) on AUC. On the contrary, SBC shows better performance on CLL (16 wins and 4 losses) compared with NB. 5. MIWNB is inferior to NB. Because, although MIWNB performs a little better than NB on AUC (8 wins and 3 losses), it shows bad performance on ACC (5 wins and 9 losses), and is significantly inferior to NB on CLL (1 wins and 27 losses). 6. ReFWNB is an ineffective attribute weighting method for improving accuracy performance. In terms of ACC, it is 12 inferior to NB (5 wins and 12 losses), SBC (3 wins and 10 losses) and all other attribute weighted method: CF- SWNB (4 wins and 13 losses), GRWNB (2 wins and 10 losses), MIWNB (5 wins and 8 losses) and TreeWNB (2 wins and 11 losses). Besides, ReFWNB almost ties NB both on ACC (5 wins and 4 losses) and CLL(12 wins and 11 losses). Overall, our experiments show that most existing attribute weighting approaches cannot achieve good performance with respect to the measures studied in our experiments, including AUC, CLL, and ACC. What is more, the classification accuracy (ACC) is a fundamental evaluation criteria in many real-world applications (e.g., image retrieval). In this case, it is impor- tant to design a novel attribute weighting approach for WNB that could improve the classification performance for different learning tasks. 5.2.2. Accuracy Comparisons Between AISWNB and Baselines In Figs. 3, 4, and 5, we report the performance of the pro- posed AISWNB, which uses artificial immune system princi- ples to calculate attribute weight values for NB classification in term of classification accuracy (ACC), and area under the ROC curve (AUC), averaged conditional log likelihood (CLL), respectively. In these figures, data points below the y = x di- agonal line are data sets on which AISWNB achieves better results than the rival algorithm. The detailed results are further presented in Tables 3, 4, and 5. Our experimental results indi- cate that AISWNB has very significant gain compared to other attribute weighting methods on the above three evaluation cri- teria. In summary, our experimental results in Tables 6, 7, and 8 show: 1. AISWNB not only significantly outperforms NB (11 wins and 0 loss) in ACC, but is also in AUC (15 wins and 1 loss) and CLL (21 wins and 0 loss). 2. AISWNB significantly outperforms SBC (10 wins and 0 loss) in ACC and AUC, and also outperforms SBC in CLL with 8 wins and 1 loss. 3. AISWNB outperforms CFSWNB (8 wins and 0 losses), GRWNB (9 wins and 0 loss), MIWNB (11 wins and 0 losses), ReFWNB (14 wins and 0 losses) and TreeWNB (8 wins and 0 losses) in ACC. The similar superiority for AISWNB could also been found in other two evaluation criteria (AUC and CLL). 5.3. Image Retrieval Learning Tasks In image classification, an image is classified according to its visual content. An important application of image classification is image retrieval: searching through an image data set to obtain (or retrieve) those images with particular visual content. For example, finding pictures containing a car. In this part of experiment, we report AISWNB’s performance for automatic image retrieval tasks. The original data are color images from Corel benchmark data set (Li & Wang, 2008). For Figure 6: Example images used in the experiment are from the COREL im- age categorization database. Each row shows images in one category, and the selected images in our data set are from six categories (“people”, “elephant”, “car”, “fox”, “tiger”, and “bus”). each image, four sets of visual features are extracted, includ- ing color histogram (Wu et al., 2014), color histogram layout (Liu & Yang, 2013), color moments (Chen et al., 2013), and co-occurrence texture (Luo et al., 2013). We choose the color histogram approach in the HSV (Hong et al., 2013) color space as color features. The HSV color space is divided into 32 sub- spaces (32 colors with 8 ranges of H and 4 ranges of S ). For each image, the value in each dimension of the color histogram represents the density of the corresponding color in the entire image. So the whole image is represented as 32-dimensional color histogram features. For the color histogram layout, each image is divided into 4 sub-images (one horizontal split and one vertical split), in which 4×2 color histogram for each subimage is computed. In this case, we can obtain another 32-dimensional features. In addition, the color moment feature has 9 dimensions, in which one (mean, or standard deviation, or skewness) for each of H, S , and V in HSV color space. At last, for texture feature, im- ages are converted to 16 gray-scale images, then co-occurrence in 4 directions is computed (horizontal, vertical, and two diag- onal directions). The corresponding 16 texture feature values are: one for each direction, second angular moment, contrast, inverse difference moment, and entropy. The similar image fea- ture exploration approach could also been found in Ortega et al. (1998). Some sample images from the benchmark data sets are shown in Fig. 6. In our experiment, we use six different cat- egories (“people”, “elephant”, “car”, “fox”, “tiger”, and “bus”) to form six binary learning problems (one for each category). To obtain negative classes for each of the six positive classes, we randomly selected 100 images from the remaining classes. Therefore, we form six binary image data sets, each of which is evenly balanced and contains exactly 200 images. Tables 9, 10, and 11 report the detailed experimental results in terms of ACC, AUC, and CLL, respectively. Overall, the pro- posed AISWNB can achieve the best performance in all cases, which demonstrate that AISWNB can outperform other base- line methods for visual feature based image classification. 13 Table 9: Experimental results on the image classification data sets with respect to the clarification accuracy (ACC) %. Data set AISWNB NB SBC CFSWNB GRWNB MIWNB ReFWNB TreeWNB bike 80.03 71.46 74.66 74.28 74.03 73.94 73.90 72.66 car 75.05 66.47 66.74 67.36 66.14 68.03 59.25 67.25 elephant 82.48 73.50 77.15 75.95 73.00 76.40 75.05 74.25 fox 70.34 56.25 53.10 55.90 53.15 56.75 54.35 55.95 people 86.67 75.67 72.89 73.85 73.3 74.14 73.40 74.96 tiger 89.87 78.60 78.85 78.15 77.3 80.10 81.05 80.85 Table 10: Experimental results on the image classification data sets with respect to the area under the ROC curve (AUC) %. Data set AISWNB NB SBC CFSWNB GRWNB MIWNB ReFWNB TreeWNB bike 89.76 79.80 81.51 81.53 80.96 81.63 80.96 80.42 car 82.54 72.33 72.21 73.35 72.55 73.10 72.14 73.20 elephant 90.62 84.96 86.48 86.62 86.56 85.56 84.41 85.04 fox 72.07 57.14 54.62 58.53 56.46 59.07 56.34 57.74 people 88.54 83.48 81.13 84.28 83.25 83.56 83.84 84.28 tiger 91.23 86.18 85.23 85.43 85.94 85.95 86.85 86.41 Table 11: Experimental results on the image classification data sets with respect to the averaged conditional log likelihood (CLL) %. Data set AISWNB NB SBC CFSWNB GRWNB MIWNB ReFWNB TreeWNB bike -43.78 -159.58 -75.70 -63.64 -129.31 -387.03 -61.66 -72.12 car -32.54 -116.60 -74.20 -63.51 -78.12 -237.54 -67.75 -66.45 elephant -44.12 -60.95 -48.83 -48.22 -137.56 -742.52 -64.49 -51.20 fox -28.40 -94.12 -77.35 -69.83 -130.99 -601.20 -69.20 -69.88 people -37.90 -124.87 -57.82 -60.13 -200.71 -428.08 -57.44 -58.12 tiger -30.88 -66.62 -55.99 -49.50 -133.67 -603.65 -62.52 -48.48 5.4. Convergence and Learning Curves In order to investigate the convergence of the AISWNB al- gorithm, we report the relationship between the number of it- erations and one of the evaluation criteria (i.e., classification accuracy) on 12 UCI data sets (half of them are from data sets with a relatively large number of instances, the remaining half represent data sets with a relatively large number of attributes). The results are shown in Figs. 7 and 8, where each point in the curves corresponds to the mean accuracy from 10-fold cross validation under the underlying iteration with the current optimal attribute weight values. The results in Figs. 7 and 8 show that AISWNB has a quick convergence to achieve a higher classification accuracy than other algorithms. To further study the convergence of the al- gorithm, let’s take the “kr-vs-kp” data set, which is a high- dimensional data set (37 attributes) with 3196 instances, as an example. Existing research by Kohavi (1996) has shown that “kr-vs-kp” data set has strong attribute dependencies. Our re- sults show that AISWNB achieves 96.9% classification accu- racy, which is significantly higher than NB’s accuracy 88.1% on the same data set. The accuracy of the final convergence is better than CFSWNB (91.3%), GRWNB (90%), MIWNB (90.8%), ReFWNB (90.4%), and TreeWNB (94.4%). Similar levels of improvement can also be observed from other data sets. In some situations, the proposed AISWNB can achieve a better accuracy than a number of attribute weighted baseline methods within only one iteration. This demonstrates the good convergence performance of AISWNB. Meanwhile, in order to determine whether the improvement of AISWNB is attributed to random attribute weights or simple random weight values can also have a high classification performance, we use random weight values and denote the results by RMWNB in the last col- umn of Tables 3 (ACC), 4 (AUC) and 5 (CLL). The results on 36 UCI benchmark data sets clearly show that random attribute weight selection is ineffective to improve the performance of WNB. According to the t-test results in Tables 6, 7, and 8, some detailed explanations can be discussed as follows: 1. RMWNB is severely inferior to AISWNB (0 win and 15 losses) in ACC, (0 win and 17 losses) in AUC, and (0 win 14 0 50 100 150 200 0.714 0.716 0.718 0.72 0.722 0.724 0.726 0.728 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy letter (NumInstances:20000, NumAttributes:17) AISWNB CFSWNB=0.701 GRWNB=0.713 MIWNB=0.715 ReFWNB=0.718 TreeWNB=0.710 RMWNB=0.644 NB=0.712 (a) 0 50 100 150 200 0.89 0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy kr−vs−kp (NumInstances:3196, NumAttributes:37) AISWNB CFSWNB=0.913 GRWNB=0.900 MIWNB=0.908 ReFWNB=0.904 TreeWNB=0.944 RMWNB=0.809 NB=0.881 (b) 0 50 100 150 200 0.972 0.973 0.974 0.975 0.976 0.977 0.978 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy sick (NumInstances:3772, NumAttributes:30) AISWNB CFSWNB=0.974 GRWNB=0.965 MIWNB=0.963 ReFWNB=0.967 TreeWNB=0.970 RMWNB=0.960 NB=0.969 (c) 0 50 100 150 200 0.85 0.86 0.87 0.88 0.89 0.9 0.91 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy waveform−5000 (NumInstances:5000, NumAttributes:41) AISWNB CFSWNB=0.805 GRWNB=0.803 MIWNB=0.802 ReFWNB=0.823 TreeWNB=0.833 RMWNB=0.822 NB=0.835 (d) 0 50 100 150 200 0.95 0.955 0.96 0.965 0.97 0.975 0.98 0.985 0.99 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy mushroom (NumInstances:8124, NumAttributes:23) AISWNB CFSWNB=0.983 GRWNB=0.983 MIWNB=0.982 ReFWNB=0.979 TreeWNB=0.980 RMWNB=0.937 NB=0.938 (e) 0 50 100 150 200 0.9355 0.936 0.9365 0.937 0.9375 0.938 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy hypothyroid (NumInstances:3772, NumAttributes:30) AISWNB CFSWNB=0.937 GRWNB=0.935 MIWNB=0.761 ReFWNB=0.933 TreeWNB=0.937 RMWNB=0.935 NB=0.930 (f) Figure 7: Convergence learning curves of AISWNB for 6 data sets with large number of instances. (a) letter: 20000 instances. (b) kr-vs-kp: 3196 instances. (c) sick: 3772 instances. (d) waveform-5000: 5000 instances. (e) mushroom: 8124 instances. (f) hypothyroid: 3772 instances. 0 50 100 150 200 0.96 0.965 0.97 0.975 0.98 0.985 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy anneal (NumInstances:898, NumAttributes:39) AISWNB CFSWNB=0.938 GRWNB=0.952 MIWNB=0.869 ReFWNB=0.949 TreeWNB=0.902 RMWNB=0.934 NB=0.950 (a) 0 50 100 150 200 0.96 0.962 0.964 0.966 0.968 0.97 0.972 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy splice (NumInstances:3190, NumAttributes:62) AISWNB CFSWNB=0.957 GRWNB=0.943 MIWNB=0.947 ReFWNB=0.890 TreeWNB=0.963 RMWNB=0.932 NB=0.959 (b) 0 50 100 150 200 0.78 0.8 0.82 0.84 0.86 0.88 0.9 0.92 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy audiology (NumInstances:226, NumAttributes:70) AISWNB CFSWNB=0.754 GRWNB=0.790 MIWNB=0.849 ReFWNB=0.667 TreeWNB=0.820 RMWNB=0.711 NB=0.779 (c) 0 50 100 150 200 0.905 0.91 0.915 0.92 0.925 0.93 0.935 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy anneal.ORIG (NumInstances:898, NumAttributes:39) AISWNB CFSWNB=0.886 GRWNB=0.871 MIWNB=0.782 ReFWNB=0.766 TreeWNB=0.889 RMWNB=0.847 NB=0.898 (d) 0 50 100 150 200 0.934 0.936 0.938 0.94 0.942 0.944 0.946 0.948 0.95 0.952 0.954 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy soybean (NumInstances:683, NumAttributes:36) AISWNB CFSWNB=0.935 GRWNB=0.913 MIWNB=0.923 ReFWNB=0.902 TreeWNB=0.939 RMWNB=0.919 NB=0.933 (e) 0 50 100 150 200 0.925 0.93 0.935 0.94 0.945 0.95 Generation/Iteration M e a n C la ss ifi ca tio n A cc u ra cy ionosphere (NumInstances:351, NumAttributes:35) AISWNB CFSWNB=0.930 GRWNB=0.927 MIWNB=0.928 ReFWNB=0.927 TreeWNB=0.937 RMWNB=0.921 NB=0.923 (f) Figure 8: Convergence learning curves of AISWNB for 6 data sets with large number of attributes. (a) anneal: 39 attributes. (b) splice: 62 attributes. (c) audiology: 70 attributes. (d) anneal.ORIG: 39 attributes. (e) soybean: 36 attributes. (f) ionosphere: 35 attributes. 15 and 16 losses) in CLL. 2. RMWNB is inferior to SBC, CFSWNB, and TreeWNB in all above three evaluation criteria. Specifically, for ACC it is around (0 wins and 13 losses), AUC around (0 wins and 20 losses), and CLL around (2 wins and 28 losses). 3. Although RMWNB shows a better performance than GR- WNB (17 win and 6 losses) and MIWNB (28 win and 3 losses) in terms of CLL values, its performs is inferior to GRWNB and MIWNB on the other two evaluation criteria ACC and AUC. 4. Compared with ReFWNB, RMWNB seems competitive with ReFWNB. This is due to the fact that RMWNB ties ReFWNB (6 wins and 6 losses) in ACC, and shows bet- ter CLL performance with (17 wins and 10 losses). On the contrary, RMWNB is worse than ReFWNB in AUC (2 wins and 14 losses). It is worth noting that in Section 5.2.1 we have investigated accuracy performance for a va- riety of attribute weighting methods for WNB in the liter- ature. ReFWNB has shown to be an inappropriate method to improve the ACC of WNB. 6. Conclusions In this paper, we proposed to improve Naive Bayes classi- fication by relaxing the attribute conditional independence as- sumption. Because attributes may play different roles in real- work applications, many existing works have proposed to use attribute weighting, which is an expansion of feature selection, to improve NB classification. In the paper, we argued that exist- ing attribute weighing methods separate the attribute weighting and the Naive Bayes learning into two separated steps, without taking the Naive Bayes objective function into consideration. As a result, the selected attribute weights may not directly cor- respond to the final classification performance. Alternatively, our research intends to design an integrated process with attribute weighting being seamlessly integrated into the Naive Bayes learning objective. To this end, we proposed to improve the Naive Bayes classification by self- adaptively assigning proper weight values to the attributes. Our method, namely AISWNB, uses immunity theory in artificial immune systems, including initialization, clone, section, and mutation, to self-adaptively search optimal weight values for weighted Naive Bayes classification. In each generation (i.e., iteration), AISWNB will choose the appropriate affinity function to meet the different learning tasks. After that, each individual (i.e., weight vector) will be sorted by the affinity in order to select the best individual as the memory antibody. By cloning the best individual to replace the indi- viduals with low affinity, the performance of the current pop- ulation will be enhanced. Meanwhile, in order to maintain the diversity, the mutation operation is adopted. The iterative pro- cess will continue until the algorithm converges, through which, the selected attribute weight values can adapt to the underlying training data to ensure good performance gain. Experiments and comparisons on 36 UCI benchmark data sets and six image classification data sets demonstrated that AISWNB outperforms state-of-the-art attribute weighted Naive Bayes approaches in terms of three performance metrics, in- cluding classification accuracy (measured by ACC), class rank- ing performance (measured by AUC), and class probability esti- mation (measured by CLL). The convergence study also shows that the proposed AISWNB has good convergence speed. Overall, this paper provided an effective approach to self- adaptively calculate the attribute weight for Naive Bayes clas- sifiers. Our principle of using evolutionary machine learning can be applied/extended to other weight optimization prob- lems, such as weighted one-dependence estimators (Wu & Cai, 2014) in Bayesian networks. In addition, the affinity function in AISWNB may be changed when facing different evaluation criteria. In this case, further research can consider integrating affinity functions to achieve a trade-off between different eval- uation criteria for optimal performance gain. Acknowledgments The work was supported by the Key Project of the Nat- ural Science Foundation of Hubei Province, China (Grant No. 2013CFA004), and the National Scholarship for Build- ing High Level Universities, China Scholarship Council (No. 201206410056), and National Natural Science Foundation of China (Grant No. 61403351 and 61370025). It is also par- tially supported by the Australian Research Council Discovery Projects under Grant No. DP140100545 and DP140102206. This research was also partially done when the first author vis- ited Sa-Shixuan International Research Centre for Big Data Management and Analytics hosted in Renmin University of China. This Center is partially funded by a Chinese National “111” Project “Attracting International Talents in Data Engi- neering and Knowledge Engineering Research”. References Aickelin, U., Dasgupta, D., & Gu, F. (2013). Artificial immune systems (intros 2). CoRR, abs/1308.5138. Bache, K., & Lichman, M. (2013). UCI machine learning repository. URL: http://archive.ics.uci.edu/ml. Castro, L. N. D., & Timmis, J. (2002). Artificial immune systems: A novel paradigm to pattern recognition. In Artificial Neural Networks in Pattern Recognition (pp. 67–84). Springer Verlag, University of Paisley, UK. Chen, J., Huang, H., Tian, S., & Qu, Y. (2009). Feature selection for text classification with naive bayes. Expert Systems with Applications, 36, 5432– 5435. Chen, Y., Sampathkumar, H., Luo, B., & wen Chen, X. (2013). ilike: Bridging the semantic gap in vertical image search by integrating text and visual fea- tures. IEEE Transactions on Knowledge and Data Engineering, 25, 2257– 2270. Cuevas, E., Osuna-Enciso, V., Wario, F., Zaldı́var, D., & Pérez-Cisneros, M. (2012). Automatic multiple circle detection based on artificial immune sys- tems. Expert Systems with Applications, 39, 713–722. Er, O., Yumusak, N., & Temurtas, F. (2012). Diagnosis of chest diseases using artificial immune system. Expert Systems with Applications, 39, 1862–1868. Friedman, N., Geiger, D., & Goldszmidt, M. (1997). Bayesian network classi- fiers. Machine Learning, 29, 131–163. 16 http://archive.ics.uci.edu/ml Grossman, D., & Domingos, P. (2004). Learning bayesian network classifiers by maximizing conditional likelihood. In Proceedings of the Twenty-first In- ternational Conference on Machine Learning ICML’04 (pp. 361–368). New York, NY, USA. Haktanirlar Ulutas, B., & Kulturel-Konak, S. (2012). An artificial immune system based algorithm to solve unequal area facility layout problem. Expert Systems with Applications, 39, 5384–5395. Hall, M. (2007). A decision tree-based attribute weighting filter for naive bayes. Knowledge-Based Systems, 20, 120–126. Hall, M. A. (2000). Correlation-based feature selection for discrete and numeric class machine learning. In Proceedings of the Seventeenth International Conference on Machine Learning ICML ’00 (pp. 359–366). San Francisco, CA, USA. Han, E.-H., Karypis, G., & Kumar, V. (2001). Text categorization using weight adjusted k-nearest neighbor classification. In Proceedings of the 5th Pacific- Asia Conference on Knowledge Discovery and Data Mining PAKDD’01 (pp. 53–65). London, UK, UK. Hand, D. J., & Till, R. J. (2001). A simple generalisation of the area under the roc curve for multiple class classification problems. Machine Learning, 45, 171–186. Hernández-González, J., Inza, I. n., & Lozano, J. A. (2013). Learning bayesian network classifiers from label proportions. Pattern Recogn., 46, 3425–3440. Hong, Z., Mei, X., Prokhorov, D., & Tao, D. (2013). Tracking via robust multi- task multi-view joint sparse representation. In International Conference on Computer Vision ICCV’13 (pp. 649–656). Sydney, Australia. Huang, K., Liu, X., Li, X., Liang, J., & He, S. (2013). An improved artificial immune system for seeking the pareto front of land-use allocation problem in large areas. International Journal of Geographical Information Science, 27, 922–946. Huang, Y.-P., Chang, Y.-T., Hsieh, S.-L., & Sandnes, F. E. (2011). An adaptive knowledge evolution strategy for finding near-optimal solutions of specific problems. Expert Systems with Applications, 38, 3806–3818. Jiang, L., Cai, Z., Zhang, H., & Wang, D. (2012a). Not so greedy: Randomly selected naive bayes. Expert Systems with Applications, 39, 11022–11028. Jiang, L., & Zhang, H. (2005). Learning instance greedily cloning naive bayes for ranking. In Proceedings of the Fifth IEEE International Conference on Data Mining ICDM’05 (pp. 202–209). Washington, DC, USA. Jiang, L., Zhang, H., & Cai, Z. (2009). A novel bayes model: Hidden naive bayes. IEEE Transactions on Knowledge and Data Engineering, 21, 1361– 1371. Jiang, L., Zhang, H., Cai, Z., & Wang, D. (2012b). Weighted average of one- dependence estimators?. Journal of Experimental and Theoretical Artificial Intelligence, 24, 219–230. Kim, S.-B., Han, K.-S., Rim, H.-C., & Myaeng, S.-H. (2006). Some effective techniques for naive bayes text classification. IEEE Transactions on Knowl- edge and Data Engineering, 18, 1457–1466. Kira, K., & Rendell, L. A. (1992). A practical approach to feature selection. In Proceedings of the ninth international workshop on Machine learning (pp. 249–256). San Francisco, CA, USA. Kohavi, R. (1996). Scaling up the accuracy of naive-bayes classifiers:a decision-tree hybrid. In Proceedings of Second International Conference on Knowledge Discovery and Data Mining KDD’96 (pp. 202–207). Protland, Oregon, USA. Kononenko, I. (1994). Estimating attributes: analysis and extensions of re- lief. In Proceedings of the 7th European Conference on Machine Learning ECML’94 (pp. 171–182). Secaucus, NJ, USA. Langley, P., & Sage, S. (1994). Induction of selective bayesian classifiers. In Proceedings of the Tenth International Conference on Uncertainty in Artifi- cial Intelligence UAI’94 (pp. 339–406). Morgan Kaufmann. Li, J., & Wang, J. Z. (2008). Real-time computerized annotation of pictures. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30, 985– 1002. Ling, C. X., Huang, J., & Zhang, H. (2003). Auc: A statistically consistent and more discriminating measure than accuracy. In Proceedings of the 18th International Joint Conference on Artificial Intelligence IJCAI’03 (pp. 519– 524). San Francisco, CA, USA. Liu, G.-H., & Yang, J.-Y. (2013). Content-based image retrieval using color difference histogram. Pattern Recogn., 46, 188–198. Liu, W.-Y., Yue, K., & Li, W.-H. (2011). Constructing the bayesian network structure from dependencies implied in multiple relational schemas. Expert Systems with Applications, 38, 7123–7134. Luo, Y., Tao, D., Xu, C., Xu, C., Liu, H., & Wen, Y. (2013). Multiview vector- valued manifold regularization for multilabel image classification. Neural Networks and Learning Systems, IEEE Transactions on, 24, 709–722. de Mello Honorio, L., Leite da Silva, A., & Barbosa, D. (2012). A cluster and gradient-based artificial immune system applied in optimization scenarios. IEEE Transactions on Evolutionary Computation, 16, 301–318. Ortega, M., Rui, Y., Chakrabarti, K., Porkaew, K., Mehrotra, S., & Huang, T. (1998). Supporting ranked boolean similarity queries in mars. IEEE Transactions on Knowledge and Data Engineering, 10, 905–925. Park, T., & Ryu, K. R. (2010). A dual-population genetic algorithm for adap- tive diversity control. IEEE Transactions on Evolutionary Computation, 14, 865–884. Quinlan, J. R. (1993). C4.5: Programs for Machine Learning. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc. Robnik-Šikonja, M., & Kononenko, I. (2003). Theoretical and empirical anal- ysis of relieff and rrelieff. Machine Learning, 53, 23–69. Storn, R., & Price, K. (1997). Differential evolution &ndash; a simple and efficient heuristic for global optimization over continuous spaces. Journal of Global Optimization, 11, 341–359. Tucker, C., Kim, H., Barker, D., & Zhang, Y. (2010). A relieff attribute weight- ing and x-means clustering methodology for top-down product family opti- mization. Engineering Optimizaiton, 42, 593–616. Webb, G. I., Boughton, J. R., Zheng, F., Ting, K. M., & Salem, H. (2012). Learning by extrapolation from marginal to full-multivariate probability dis- tributions: Decreasingly naive bayesian classification. Machine Learning, 86, 233–272. Witten, I. H., & Frank, E. (2005). Data Mining: Practical Machine Learning Tools and Techniques. The Morgan Kaufmann Series in Data Management Systems (2nd ed.). San Francisco, CA: Morgan Kaufmann Publishers. URL: http://www.cs.waikato.ac.nz/ml/weka/. Woldemariam, K. M., & Yen, G. G. (2010). Vaccine-enhanced artificial im- mune system for multimodal function optimization. IEEE Transactions on Systems, Man, and Cybernetics–Part B: Cybernetics, 40, 218–228. Wu, J., & Cai, Z. (2014). Learning attribute weighted aode for roc area rank- ing. International Journal of Information and Communication Technology, 6, 23–38. Wu, J., Cai, Z., Zeng, S., & Zhu, X. (2013a). Artificial immune system for attribute weighted naive bayes classification. In Proceedings of the Inter- national Joint Conference on Neural Networks IJCNN’13 (pp. 798–805). Dallas, TX, USA. Wu, J., Cai, Z., & Zhu, X. (2013b). Self-adaptive probability estimation for naive bayes classification. In Proceedings of the International Joint Confer- ence on Neural Networks IJCNN’13 (pp. 2303–2310). Dallas, TX, USA. Wu, J., Hong, Z., Pan, S., Zhu, X., Zhang, C., & Cai, Z. (2014). Multi-graph learning with positive and unlabeled bags. In Proceedings of SIAM Inter- national Conference on Data Mining SDM’14 (pp. 217–225). Philadelphia, Pennsylvania, USA. Wu, X., Kumar, V., Ross Quinlan, J., Ghosh, J., Yang, Q., Motoda, H., McLach- lan, G. J., Ng, A., Liu, B., Yu, P. S., Zhou, Z.-H., Steinbach, M., Hand, D. J., & Steinberg, D. (2007). Top 10 algorithms in data mining. Knowledge and Information Systems, 14, 1–37. Yuan, J., Zhang, L., Zhao, C., Li, Z., & Zhang, Y. (2012). An improved self- organization antibody network for pattern recognition and its performance study. Pattern Recognition, 321, 96–103. Zaidi, N. A., Cerquides, J., Carman, M. J., & Webb, G. I. (2013). Alleviat- ing naive bayes attribute independence assumption by attribute weighting. Journal of Machine Learning Research, 14, 1947–1988. Zhang, C., Xue, G.-R., Yu, Y., & Zha, H. (2009). Web-scale classification with naive bayes. In Proceedings of the 18th International Conference on World Wide Web WWW ’09 (pp. 1083–1084). Zhang, H., & Sheng, S. (2004). Learning weighted naive bayes with accurate ranking. In Proceedings of the Fourth IEEE International Conference on Data Mining ICDM’04 (pp. 567–570). Washington, DC, USA. Zhang, H., & Su, J. (2004). Naive bayesian classifiers for ranking. In Proceed- ings of the 15th European Conference on Machine Learning ECML’04 (pp. 501–512). Pisa, Italy. Zheng, J., Chen, Y., & Zhang, W. (2010). A survey of artificial immune appli- cations. Artificial Intelligence Review, 34, 19–34. Zhong, Y., & Zhang, L. (2012). An adaptive artificial immune network for supervised classification of multi-/hyperspectral remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 50, 894–909. 17 http://www.cs.waikato.ac.nz/ml/weka/ Introduction Related Work Attribute Weighted Methods Attribute Weighting via Single Attribute Correlation Attribute Weighting via Multiple Attribute Correlation Artificial Immune Systems Human Immune System AIS: Artificial Immune Systems Preliminaries and Problem Definition Self-Adaptive Attribute Weighted Naive Bayes AIS Symbol Definitions and Overall Framework AIS Symbol Definitions AISWNB Overall Framework AISWNB: AIS based Attribute Weighted Naive Bayes Initialization AISWNB Evaluation AISWNB Updating Time Complexity Experiments Experimental Settings Benchmark Data and Parameters Baseline Methods Evaluation Criterion UCI Benchmark Learning Tasks Attribute Weighted NB vs. Standard NB Accuracy Comparisons Between AISWNB and Baselines Image Retrieval Learning Tasks Convergence and Learning Curves Conclusions 
work_66jumwpewnfhfbuvxt6k33exle	----	PII: 0957-4174(96)00018-8 Pergamon Expert Systems With Applications, Vol. 10, No. 3/4, pp. 393--401, 1996 Copyright © 1996 Elsevier Science Lid Printed in Great Britain. All rights reserved 0957-4174/96 $15.00+0.00 S0957-4174(96)00018-8 Interactive Induction of Expert Knowledge BINGCHIANG JENG, TING-PENG LIANG AND MINYANG HONG Department of Information Management, National Sun Yat-Sen University, Kaohsiung, Taiwan 80424, Republic of China Abstract--The process o f extracting, structuring and organizing elicited knowledge (called knowledge acquisition) is a bottleneck in developing knowledge-based systems. A manual approach that elicits domain knowledge by interviewing human experts typically has problems, because the experts are often unable to articulate their reasoning rules. An automatic approach that induces knowledge from a set o f training cases also suffers from the unavailability o f sufficient training cases. We present an integrated approach that combines the strengths o f both methods to compensate f o r their weaknesses. In this approach, human experts are responsible f o r solving problems, whereas an inductive learning algorithm is responsible f o r reasoning and consistency checking. Copyright © 1996 Elsevier Science Ltd 1. I N T R O D U C T I O N K N O W L E D G E ACQUISITION is a process o f extracting, structuring and organizing elicited knowledge from domain experts or other knowledge sources and convert- ing the knowledge into rules or other forms o f representation accessible with a computer program [16,25]. As the power o f an expert system comes from the knowledge that it possesses [13], the acquisition o f knowledge is a major stage in the development of an experts system. The transfer of expertise from various sources to a knowledge base is difficult and challenging. A manual process that encodes domain knowledge by interviewing human experts suffers from several difficulties [31,40]. For example, human experts are typically not also expert in articulating their rules of reasoning because they are not explicitly aware o f the structure o f their knowledge. A paradox o f expertise [20] claims that the better one is an expert, the worst one is to tell the details. Knowledge acquired during interview may be unreliable, as verbal reports and mental behavior are not necessarily corre- lated [2]. Incongruities may occur among what an expert says that he does, what he actually does and what he should have done [17]. Communication between the knowledge engineer and the domain expert is another difficult problem. The knowledge engineer may know little about the problem domain, and may not understand clearly the jargon used by the domain expert. On the other hand, the domain expert may not understand the process o f knowledge acquisition and what the knowledge engineer needs. Even the cultural differences between the two parties o f knowledge acquisition may undermine the development o f a knowledge-based system. When historical examples are available, an automatic process can be used to induce rules from fragments o f knowledge [35,36]. The unique power o f this approach is the "automatic" learning capability o f its inductive algorithm. Many successful applications are reported, such as disease diagnoses [42,23,29], business applica- tions [5,8,39,22] and others [5,10,26]. A major constraint o f automatic learning is that it requires many training cases from which to induce knowledge. The quality o f the training result depends on the quality o f the data. Another drawback is that induction may generate knowledge that uses paths of reasoning different from those used by the expert. This may make human experts reluctant to accept the induced knowledge base regardless o f its performance. Given that each approach alone has limitations and their respective strengths and weaknesses seem com- plementary, we seek to integrate the two approaches to gather the advantages and to avoid the deficiencies. The idea is that the weaknesses o f one approach be compensated by the strengths o f the other. For example, the task o f articulating a reasoning process for a human expert can be left to an inductive method to solve, or the restriction o f inductive learning on training examples can be given to an experienced human expert to provide help. This approach is useful as human experts are good at solving problems whereas inductive learning algorithms are good at extracting rules o f reasoning. Previous research that aimed at a similar goal exists. For example, Parsaye [31] proposed an approach that combines interactive knowledge acquisition with rule induction. Based on the theory o f repertory grids [3,21], the system can interactively interview an expert by 393 394 Bingchiang Jeng et aL asking questions to capture the expert's knowledge. The captured knowledge is then generalized using rule induction. Buntine and Stirring [7] presented another approach that allows a human expert to interact with a set o f automatically induced rules so that each other's knowledge is cross-checked. Recently Evans and Fisher [12] reported a similar approach that is successfully applied to a problem domain where the experts have only weak causal knowledge. A limitation o f Buntine and Stirling's approach is that it still depends on a set o f pre-existing training examples. In some cases, these examples are hard to come by. Furthermore, the way an expert interacts with the preliminary rules is informal. The approach presented in this paper extends their ideas to allow interactive knowledge acquisition from (nearly) scratch. A modified algorithm for inductive learning is designed so that an expert communicates with the system in an interrogative style during the inductive process. In addition to inducing a decision tree from existing training cases, the system also identifies near- miss cases falling on the classification borders for clarification and verification by human experts. These cases, after classification, become new training cases for accurate learning by the modified inductive algorithm. This way, the expert's knowledge is accumulated incre- mentally and the learning process can be repeated until the induced knowledge is satisfied by the human expert. The remainder o f the paper is organized as follows. The process of automatic knowledge acquisition and its related techniques is reviewed in Section 2. An inter- active process for inductive learning that shows how to elicit knowledge from a domain expert is described in Section 3. Experimental results from the evaluation of the performance o f the proposed approach is presented in Section 4. Section 5 summarizes the research. 2. A U T O M A T I N G K N O W L E D G E A C Q U I S I T I O N There are many different methods for knowledge acquisition, which can be classified as: manual, semi- automatic and automatic [11,25]. Eliciting knowledge by manual methods is highly labor-intensive. Methods belonging to this category include structured/unstruc- tured interviews, analysis o f protocol, observations and so forth [40]. Semi-automatic methods are primarily designed to support either the expert or the knowledge engineer to perform the necessary task more effectively. Tools to elicit domain knowledge include ETS, KRITON, AQUI- NAS, MORE, M O L E etc. [4,24]. Others include TEIRESIAS [17] that assists the knowledge engineer to update a knowledge base, and KADS [14] that provides a collection o f tools to support a knowledge engineer to extract, to structure, to analyze and to document expert knowledge. Although semi-automatic methods expedite the work o f knowledge engineers and/or domain experts, they still share problems o f manual methods: the acquired knowl- edge is difficult to validate; correlation between verbal reports and mental behavior may be weak. In addition, Michie [28] stated that knowledge o f certain types could not be elicited using manual methods, simply because the domain was so large and complicated that the expert would be unable to explain how it operates. Automatic methods that use machine-learning tech- niques to extract knowledge require less or no participation by either knowledge engineers or domain experts. Therefore, they do not have the difficulties associated with human experts. Currently, the most popular automated method is rule induction. Induction is a process o f general inference from particular instances. Rule induction (or inductive learn- ing) refers to a concept learning process by which a set o f rules is created from training cases to explain or to predict a problem-solving behavior. When the induced structure o f knowledge is represented in the form o f a decision tree, it is also called (decision) tree induction [27,35]. A decision tree is considered as a set o f rules in a compact form; it can be transformed into rules easily. Early work on rule induction is traced to 1966 when Hunt, Martin and Stone developed a method for induction. The method was later implemented and expanded b y Paterson and Niblett [32] to create ACLS (A Concept-Learning System) and by Quinlan [34,35] to develop the popular ID3. In the following, we shall use the ID3 algorithm to illustrate how rule induction works. Input to ID3 is a collection o f training cases. Each is described by a set o f attributes associated with a class name (or outcome). An attribute can be either categorical (e.g. color) or numerical (e.g. age). A numerical attribute can adopt discrete or continuous values. The basic procedure o f ID3 applies a divide-and-conquer approach to partition recursively the data set, based on a test on selected attribute values, into mutually exclusive subsets. The procedure continues until each subset contains cases o f the same class (to avoid over-fitting the training data, termination conditions may be defined) or no attribute is available for further decomposition. If S= {(ai~ . . . . . ain; ci)laij~A j, ci~C)} is a set o f training cases, where Aj denotes the domain o f attribute j and C for the class, the algorithm is shown in Fig. 1. An example shown in Fig. 2 is the decision tree created by ID3 from financial data to analyze the risk of bankruptcy. A new case is classified by examining which leaf it reaches when traveling down through the tree. The traversed path is determined b y a sequence of tests to the new case (i.e. a branch is selected depending on the outcome o f a test that the new case generates). When it reaches a leaf in this way, it is considered similar to other existing cases contained in the leaf as they have all passed the same tests. Hence, the outcome o f the new case is the same as its neighbors and is assigned with the I n t e r a c t i v e I n d u c t i o n o f E x p e r t K n o w l e d g e 3 9 5 Procedure I n d u c t i o n (S: TrainSet, T: Tree) Begin If S contains cases of the same class Then label T with the class name and exit Else Begin For each numerical attribute A i, Do Find a value v/to decompose the training set into two subsets, Calculate the entropy of the decomposition, Choose the decomposition whose entropy value is the smallest; For each categorical attribute, Decompose S by its classes and Calculate its entropy; Partition S into two or more mutually exclusive subsets: S i, i = 1,..., k, based on the attribute A s whose entropy value is the smallest after decomposition. Create a node T i for each subset S i, and link it to the parent node as its child. For i = 1 to k, call I n d u c t i o n (S i, T i) End End, FIGURE 1. T h e ID3 algorithm. class n a m e labeled at the leaf. F o r e x a m p l e the prediction o f b a n k r u p t c y for case (0, 0, 0, 0.02, 0.05, 0.6, 0.04) is YES, as it follows the path at the far left d o w n to a leaf labeled " Y E S " . 3. I N T E R A C T I V E I N D U C T I O N T h e fundamental basis for methods o f decision tree induction (or rule induction) is that t w o cases with similar features are classified into the same class. The success o f such an a p p r o a c h relies on sufficient training cases to c o v e r every aspect o f the p r o b l e m d o m a i n w i t h o u t inconsistency. Otherwise, the i n d u c e d k n o w - ledge m a y be unreliable. In practical applications, however, this a s s u m p t i o n is c o m m o n l y violated. T h e sources o f training cases are, in general, f r o m a domain expert o r f r o m historical data in a maintained database. Difficulties arise typically because the domain expert can provide only a few selected examples, whereas historical data, if they exist, m a y be obsolete o r contain errors and missing values. A further difficulty with inductive learning is in its explanation capability. Explaining the reasoning process b y which a conclusion is r e a c h e d is a key requirement for an expert system. Rules i n d u c e d f r o m the training set may, however, differ f r o m those used b y the expert. This makes its explanation s o m e t i m e s less acceptable to h u m a n beings because the u n d e r l y i n g process o f reason- ing used b y the expert s y s t e m m a y be incomprehensible to users. It is thus useful to have interactions between h u m a n experts and inductive learning algorithms. This allows h u m a n experts to provide useful knowledge, such as hand-crafted tutorial examples, rules o f thumb, general hints and problem-solving strategies, to help an inductive algorithm. The inductive algorithm presented b e l o w is an example that supports interaction with experts during its learning process. After a k n o w l e d g e structure is induced f r o m its training cases, the a l g o r i t h m identifies cases that cannot be correctly classified. T h e s e cases are b r o u g h t to the expert in the f o r m o f questions to be solved. O n c e they are solved, they b e c o m e n e w training cases to the inductive algorithm for further learning. In this way, expert k n o w l e d g e is incrementally elicited and incorpor- ated into the induced k n o w l e d g e structure. A typical scenario o f the interactive inductive process is as follows. Initially, the k n o w l e d g e e n g i n e e r collects k n o w l e d g e f r o m all possible sources. Different forms o f k n o w l e d g e are stored differently. Rules (mostly f r o m domain experts) are stored in the rule set and tutorial cases (mostly f r o m the historical database) are stored in the training set. Then, the algorithm tests any inconsistencies L e a f 1 < ' ~ Leaf4 Leaf2 Leaf3 L e a f 5 L e a f 6 B a n k r u p t c y : Y e s , N o C I : C o n s i s t e n c y O p i n i o n C 2 : Subject to O p i n i o n C 3 : G o i n g C o n c e r n QI : Net Income/Total Assets Q 2 : Current Assets/Total Assets ( ~ : Current R a t i o : Cash/Total Assets FIGURE 2. A n e x a m p l e of a decision tree for bankruptcy prediction. 396 Bingchiang Jeng et al. Y Xc a l a a a x < . x / • • • " b ~ - ~ Y < ~ / ~Y>Ye a a • b Sb Sc Xc a a b b Sc Y c Sb b X (a) A decision tree (b) Partition of feature space FIGURE 3. S p a c e partition by a decision tree. in the knowledge structure represented by these two sets. As both sets are incomplete at the beginning, contra- dictory cases (i.e. cases classified differently by those two knowledge sources) are identified. These cases are then presented to the domain expert for further review. Prompting experts to review contradictory cases has two purposes. First, certain knowledge ignored by the expert may be use to provoke him to describe the inference rules in more detail. Then the contradictory cases, after correction, become new examples to the inductive algorithm for further learning. Revision o f the two sets triggers the cross-validation process again and the above procedure is repeated. A sketch of this approach is shown in Fig. 5. An important feature of the above algorithm is its ability to detect inconsistencies during the process o f interactive induction. This is important when two different sources o f knowledge are merged because Sa X inconsistencies may exist. In the following we shall present an algorithm for detecting inconsistencies. `% Sb G i v e n b o rd er • • m e m o . • a e o Y Con-ect border 3.1. S t r a t e g y f o r B o r d e r V a l i d a t i o n A decision tree created from training cases can be represented as a partitioned feature space. The induction o f decision trees (or rules) in some sense is a process to maximize the internal similarity within the partitioned subspaces. Each leaf o f the tree corresponds to a subspace, and each case in the problem domain corre- sponds to a point in it. For example, a decision tree shown in Fig. 3 partitions a two-dimensional feature space with attribute (X, Y) into three subspaces. Sub- spaces S, and St represent class A and subspace Sb represents class B. A new case is evaluated based on the subspace into which it falls (e.g. a case with X<Xc is classified into class A). Similarly, rules acquired from human experts can also be illustrated as partitions o f the feature space. The problem o f verifying whether two knowledge sources (i.e. rules and the training set) have inconsistencies is thus equivalent to verifying whether their partitions are identical, which can be tested by simply determining if there are cases that fall into different subspaces in these two partitions. This is explained in detail in the following. A border is defined as a boundary that separates two subspaces. A border is redundant if both subspaces adjacent to the border represent the same class (e.g. the one between subspace S, and Sc in Fig. 3). Incon- sistencies occur when more than one non-redundant border can be found to partition a space into subspaces. = Sa : S c a Y : S b (a) ~ X a n d Y are shifted X Sa x u Sc Y S b O N point o O F F p o i n t x Sa , % x • e e • • w • • I • , • , • i m o Sb X X (b) ]Border shifts are d ~ with O N . O F F poinm FIGURE 4. A n example of border shift a n d t h e strategy to detect IL Y Interactive Induction of Expert Knowledge 397 TABLE 1 A Table of Path Conditions for the Bankrupt Decision Tree Path F 2 F 3 F4 F 5 Class 1 - - o 0 / . 0 2 9 5 - - o o / . 0 4 3 5 Yes 2 = 0 - o 0 / - . 0 0 9 3 - . 0 4 3 5 / o 0 N o 3 = 1 - o 0 / - . 0 0 9 3 - . 0 4 3 5 / o 0 Yes 4 - - . 0 0 9 3 / . 0 2 9 5 - . 0 4 3 5 / o 0 Yes 5 - . 0 2 9 5 / o 0 - o 0 / . 7 4 8 7 - N o 6 - . 0 2 9 5 / o 0 . 7 4 8 7 / o 0 - Yes [, Local Refinement L RGURE 5. The process for interactive Induction. k Such a situation is called a border shift, since at least one o f them is incorrect. Inconsistency exists i f and only if a border shift exists. Figure 4(a) shows two examples o f border shifts. A strategy for detecting border shifts, that requires only two points (ON and OFF), can be found in [18]. The ON point lies exactly on a given border to be verified, while the OFF point lies slightly o f f in the open side o f the border. A border shift exists i f the actual location o f the border fails outside the OFF point or inside the O N point. In other words, no border shift exists i f it passes exactly between the two points. For example, i f border Y lies in a different location (the broken line) in the left diagram o f Fig. 4(b), the OFF point would have been classified differently. Similarly, the O N point would b e classified differently in the right diagram o f Fig. 4(b), i f the solid-line border is shifted to the broken-line one. With carefully chosen testing points, the strategy can detect a border shift effectively, i f it exists. 3.2. Algorithm for Test-Case Generation The following is an algorithm for finding non-redundant borders and verifying their correctiveness in a partitioned feature space. We use the decision tree shown in Fig. 2 as an example to illustrate it. The first step o f the algorithm is to create a table o f path conditions according to the decision tree. Each path from the root to a leaf o f the decision tree corresponds to a row in the table. Attributes to partition the training cases correspond to columns. Each field in a row records an attribute condition defined by the path. The last column denotes the outcomes. Table 1 shows the path conditions o f the decision tree in Fig. 2. Each cell in the table gives the upper and lower bounds o f an attribute. The purpose o f the path condition table is to determine the adjacency relationship between two sub- spaces. In the following, we show how to identify a non-redundant border from the table. As non-redundant borders require their adjacent subspaces to represent different classes, adjacency relationships are determined only between paths that represent distinct outcomes. In our example, paths 1, 3, 4 and 6 belong to one group and paths 2 and 5 belong to another. A non-redundant border is identified when a path from one group is adjacent to a path from the other. The adjacency relationship o f two paths is calculated b y assessing whether the intersection o f their path conditions recorded in the table is non-empty. This step amounts to assessing whether corresponding attribute conditions o f the two paths overlap. The results o f the intersection define the border. For example, we can determine whether paths 1 and 2 have an intersection by examining whether their attributes overlap. The data in Table 1 indicate that their intersections are F 2 = 0 , F3 ~< - 0 . 0 0 9 3 , F 4 = " d o n ' t care" and F5 =0.0435. As the intersection is non-empty, paths 1 and 2 are considered adjacent to each other. Other non-redundant borders are calculated similarly. Given a non-redundant border, one can select an ON point and an OFF point to verify its correctness. For instance, cases to test the previously generated border might be F 2 = 0 , F 3 = - 0 . 0 0 9 3 , F 4 = " d o n ' t care", F 5 = 0 . 0 4 3 5 with c l a s s = Y E S (ON point) and F 2 = 0 , F 3 = - 0 . 0 0 9 3 , F 4 = " d o n ' t care", F 5 = 0 . 0 4 3 6 with class = N O (OFF point). In our strategy, whether the O N point is located exactly on the border is relative unimportant, as long as the OFF point is chosen to be close to it. The test cases are then compared with the outcomes predicted by inferences from the rule set. I f knowledge o f the human expert is incompletely included in the rule set or the training data are imperfect, contradictions would occur. These test cases are then presented to the human expert for review. Otherwise the process proceeds to identify the next non-redundant border and to generate other test cases. At the beginning o f the interactive induction process, a relatively large number o f contradictory cases m a y be identified. After certain revisions, contradictory cases reduce. 3.3. The Process of Interactive Induction To summarize, the process o f interactive induction as shown in Fig. 5 consists o f six phases: rule description, 398 Bingchiang Jeng et al. case induction, contradiction identification, revision and augmentation, local refinement and result validation. Iterations occur in the first four phases until a termination criterion is satisfied. Then it passes through the phases o f local refinement (optional) and result validation. This process differs f r o m the traditional inductive learning that only performs case induction and result validation. No interaction between the expert and the induced knowledge is allowed. The phase o f rule description is a major step where an expert can provide judgments and transform into rules. Any manual techniques for eliciting knowledge described before can be applied here. The set o f rules is not required to be complete as it will be refined in later iterations. It is not very difficult to create such a rule set as an expert can generally provide rules o f thumb easily. The more complete the rule set which describes the expertise, the fewer iterations it needs to perform later. I f the rule set is not available, the step o f identifying contradictory cases will check with a domain expert directly. Although ID3 is the example that we used in this research, the induction phase can use any inductive learning algorithm. The only concern is how the algorithm handles new training cases. It would be inefficient if the algorithm flushes its old memory and redoes the learning process from the beginning every time a modification is involved. U t g o f f has described an incrementally inductive algorithm ID5R [41] which processes new training cases by updating the existing decision tree, instead o f creating a new one. Thus the incremental learning cost o f the algorithm is relatively low. The training cases for induction can be historical data (if they exist), or hand-crafted tutorial examples given by the expert. The phase o f contradiction identification detects inconsistency between rules and training cases. It uses the validation strategy described in Section 3.1. Contra- dictory cases are passed over to the revision and augmentation phase in which human experts re-inspect the underlying process o f reasoning and revise the rule or the case descriptions. Cases after correction are included in the training set that initiates another iteration o f the induction process. In addition to simply resolving conflicts, defining a sufficient feature set for an application domain is a challenging problem for machine learning. The ID3 algorithm, for example, would perform poorly, i f the attributes o f a set o f training cases are not sufficient to describe a domain. This problem however, is mitigated in our approach as human experts m a y add features to solve contradictions that cannot be solved by rule revision. This way, the features in the feature set can grow. The local refinement phase is optional. It is primarily designed to fine-tuning the overall predictive accuracy o f inductive learning. For a problem domain to be modeled, the predictive accuracy cannot be improved infinitely. A limit is often reached after several iterations. This is primarily due to the orthogonal partition o f feature space in ID3, which m a y not match the real situation exactly. In this case, predictive accuracy cannot be improved by further training. Local refinement that performs learning independently on selected subspaces can then c o m e into play. Finally, the induced knowledge needs to b e validated. A set o f criteria for evaluating the quality o f the inductive result and determining when to terminate the inductive process are critical. Other tools that facilitate the validation and maintenance o f the knowledge base can also be useful in this stage. 4. P E R F O R M A N C E E V A L U A T I O N To evaluate this process, a prototype system implement- ing the method has been developed. The kernel o f the system consists o f three modules: one for inductive learning based on ID5R (an incremental version o f ID3) [41], one for rule inference based on CLIPS and another to select test cases. The user interface integrates these modules. To evaluate whether the interactive induction approach is practically useful, we conducted experiments on four hypothetical and one real problem. The hypothet- ical problems are to learn a concept that defines a geometric pattern, such as a circle, a polygon and so on. The real world problem is to learn the rules that an expert uses at a blackjack game to decide whether to call for another card. 4.1. Learning Geometric Patterns In this experiment, the system needs to learn a geometric pattern. First, a circle was used for experiment. A point falling inside a circle was assigned c l a s s 1 and others were assigned class O, as shown in Fig. 6(a). Initially, nine training cases were selected randomly from the diagram. The rule provided by the expert was Rule 1 : I f X ~ < 3 a n d X ~ > - 3 and Y~<3 and Y ~ > - 3 Then Class = 1 Else Class=0; The concept learned after the first iteration o f the induction is given in Fig. 6(b), from which the contra- dictory cases identified are shown in Fig. 6(c). The final concept induced after three iterations o f the learning process is given in Fig. 6(d). One can see that the result approaches the actual circle pattern. In addition to the nine given training cases, the system generates a total o f 20 questions for the human expert to answer in the learning process. Figure 7 shows how other geometric patterns can be learned in experiments. They all achieve satisfactory results. The learned patterns are near the actual patterns. Interactive Induction of Expert Knowledge 399 y X X (a) Initial Training Cases ~ . ~ l a s s 0 (b) First Inductive Result Y ~ . . . . , , o . . - _x :::Cia s i : : : . : . ; . ; . ; . : ( c ) Identified Conflicting (d) Final Inductive Result Cases FIGURE 6. A learning process for the circle experiment. The only limitation is that the nature o f the orthogonal partition with ID3 does not allow a real curvilinear draw. 4.2. Learning Blackjack Games Another experiment shows how a system can learn the rules that a human expert uses in playing blackjack. The game is played between a dealer and players. Each player receives two reversed cards, while the dealer receives one reversed and one obverse cards. An ace can count as either 1 point or 11 points. During the game, a player m a y request for more cards (to "hit") or to stay with the current hand (to "stand"). A player wins if the total points o f cards in hand is below 21 points but greater than the dealer's. A human expert decides whether to hit or to stand based on his current count in hand, with or without an ace, and the dealer's obverse card. These three attributes determine a decision. The expert's rules can be drawn in a decision diagram as shown in Fig. 8(a). Initially five randomly chosen cases were given for training. The rules learned after three iterations are shown in Fig. 8(b), which are pretty close to the actual rules used b y the expert. 5. RELATED WORK The subject discussed in this article can also be seen as a problem o f concept learning in the machine learning field. In particular, the way a domain expert interacts with a learning system, can be formulated as a model o f learning with membership queries [1], in which an instructor or oracle exists to answer any queries placed by a learner. In this model, the learner has control o v e r what part o f information it receives from a problem domain next. Although our work starts initially from a different perspective, its result happens to be in accordance with most recent work in machine learning. Cohn et al. [9] presented an approach to learn a concept by generating queries from a region o f uncertainty, an area in the domain where misclassification is still possible. An interesting coincidence in their neural network's imple- mentation is that it uses two networks S and G to identify an uncertain case, when outcomes o f the case are inconsistent between them. They demonstrated in several domains that the approach gives better learning results for a fixed number o f training cases than simply learning from examples alone. This is encouraging since the rationale o f identifying regions o f uncertainty is similar to our identification o f contradictory cases. Another learning model similar to the above is the on- line (or incremental) learning model in which the learner answers a sequence o f yes/no questions with immediate feedback provided after each question. A variant o f the on-line learning model, called self-directed learning, has been recently proposed b y Goldman and Sloan [15] which allows the learner to select the presentation order for the instance. Thus it can be seen as a variation o f learning with membership queries in which the learner is only "charged" for queries whose outcomes are unpre- Y . : . : . . . : - : . --i.',';:i:i:i;:i:::i.;.,~'~" ' • .~ ~ (a) ~ Class 0 x ~,:'J ..v.k-.~ t ; [ . " x . ' . ~.::;: !!!i (b) FIGURE 7. L e a r n i n g other geometric patterns. Class 0 (c) 400 Bingchiang Jeng et al. P l a y e r ' s Current Count Player's Current Count 21 2 0 19 18 17 I 6 15 14 13 12 11 10 D e a l e r ' s o b v e r s e c a r d 3 4 5 6 7 g 9 1 0 A - - S t a n d . . . . - - - H i t . . . . 21 2 0 19 15 17 16 15 14 13 12 11 10 2 3 4 5 6 7 8 9 1 0 A - - - S t a n d _ - - With Ace Without Ace ( a ) E x p e r t s ' r u l e s to p l a y b l a c k j a c k Dealer's obverse card 3 4 5 6 7 8 9 1 0 A 3 4 5 6 7 g 9 1 0 A 21 21 2o - - S t a n d - - - ~ 2o _ _ _ Stand 19 19 18 18 17 17 16 16 15 15 14 - H i t - - - H i t - - 14 13 13 12 12 11 11 1o lo Hit W i t h A c e W i t h o u t A c e ( b ) R u l e s l e a r n e d a c c o r d i n g t o t h e i n d u c t i v e a l g o r i t h m FIGURE 8. Learning the blackjack playing rules. • S t a n d . . _ H i t _ dictable. Theoretical results given by Goldman and Sloan show that the performance of self-directed learning is the best among all other commonly studied on-line and query learning models. Our process of learning by detecting contradictory cases near the classification border is also in accordance with the main results discovered from an autonomous exploratory learning system (i.e. without an external teacher). Winston [43] first drew attention to the critical role of near-miss training examples. Recently the Protos system, developed by Porter et al. [33], also used near- miss cases to learn an examplar difference before matching a new case to an existing examplar. 6. D I S C U S S I O N A N D C O N C L U S I O N S In this article, we have presented an algorithm for interactive induction and evaluated the performance of the implemented system in learning goemetric patterns and blackjack game strategies. The algorithm is simple but efficient. The contribution of this work is two-fold. First, it suggests an effective way o f integrating manual knowl- edge acquisition and inductive learning to achieve a more accurate knowledge base. Second, it provides a new way of creating near-miss training examples and learning from these examples. This allows critical knowledge to be learned without a large number o f training cases. Further research following this includes replacing ID3 with other learning methods, testing in other domains and exploring the handling o f certainty factors in interactive induction. The current approach can also be extended to acquire knowledge from multiple experts. The main difficulty for knowledge acquisition from multiple experts is how to combine their expertise and find out any inconsistencies systematically. Due to different subjective opinions, two competitive domain experts may disagree with each other occasionally. Our approach should be helpful in identifying conflicting rules or cases during the knowledge acquisition process. R E F E R E N C E S F I. Angluin, D. (1988). Queries and concept learning. Machine Learning, 2, 319-342. 2. Bainbridge, L. (1986). Asking questions and accessing knowledge. In Future computing systems. New York: Elsevier. 3. Boose, J. (1984). Personal construct theory and the transfer of human expertise. Proc. of the Nat'l Conf on Artificial Intelligence, Austin, TX. 4. Boose, J. & Gaines, B. R. (1989). Knowledge acquisition for knowledge-based system: Notes on the state-of-the-art. Machine Learning, 4, 131-143. 5. Braun, H. & Chandler, J. S. (1987). Predicting stock market behavior through rule induction: an application of the learning- from-example approach. Decision Sciences, 18, 415-429. 6. Buchanan, B. G. et al. (1983). Constructing an expert system. In E Hayes-Roth, D. Waterman & D. Lenat (Eds), Building expert systems. Reading, Addison-Wesley. 7. Buntine, W. & Stirling, D. (1990). Interactive induction. In J. E. Hayes-Michie, D. Michie & E. Tyugu (Eds), Machine intelligence, Vol. 12, pp. 121-138, Oxford: Oxford University Press. 8. Carter, C. & Catlett, J. (1987). Assessing credit card applications using machine learning. IEEE Expert, 2, 71-79. 9. Cohn, D., Atlas, L. & Ladner, R. (1994). Improving generalization with active learning. Machine Learning, 15, 201-221. 10. Chung, H. M. & Silver, M. S. (1992). Rule-based expert systems and linear models: A n empirical comparison of learning-by- example methods. Decision Science, 23, 687-707. 11. Eriksson, H. (1992). A survey of knowledge acquisition techniques and tools and their relationship to software engineering. Journal of System and Software, 19, 97-107. 12. Evans, B. & Fisher, D. (1994). Overcoming process delays with decision tree induction. IEEE Expert, 60--66. 13. Feigenbaum, E. A. (1977). The art of artificial intelligence: Themes and case studies of knowledge engineering. International Joint Conference on Artificial Intelligence, Vol. 5, pp. 1014-1029. 14. Gaines, B. R. (1988). Knowledge acquisition: Development and advances.. In M. D. Oliff, (Ed.) Expert system and intelligent programming. New York: Elsevier. 15. Goldman, S. A. & Solan, R. H. (1994). The power o f self-directed learning. Machine Learning, 14, 271-294. 16. Hart, A. (1992). Knowledge acquisition for expert systems, 2nd edn. New York: McGraw-Hill. 17. Hayes-Roth, E , Waterman, D. A. & Lenat, D. (1983). Building expert systems. Reading MA: Addison-Wesley. 18. Jeng, B. & Weyuker, E. (1994). A simplified approach to domain testing. ACM Transactions on Software Engineering and Method- ology, 3, 254-270. 19. Jeng, B., Liang, T. P, & Jeng, Y. M. FILM: A fuzzy inductive learning method for automatic knowledge acquisition. Forthcoming in Decision Support Systems. 20. Johnson, P. E. (1983). W h a t kind of expert should a system be? Journal of Medicine and Philosophy, g, 77-97. 21. Kelly, G. A. (1955). The psychology of personal constructs., New York: Norton. 22. Liang, 1". P. (1992). A composite approach to inducing knowledge for expert systems design. Management Science, 3g, 1-17. Interactive Induction o f Expert Knowledge 401 23. Marchand, A., VanLente, E & Galen, R. (1983). The assessment of laboratory tests in the diagnosis of acute appendicitis. American Journal of Clinical Pathology, 80:3, 369-374. 24. Marcus, S. (1988). Automating knowledge acquisition for expert systems. Boston, MA: Kluwer. 25. McGraw, K. L. & Harbison-Briggs, K. (1989). Knowledge acquisition: Principles and guidelines. Englewood Cliffs, NJ: Prentice-Hall. 26. Messier, W. E & Hansen, J. V. (1988). Inducing rules for expert system development: an example using default and bankruptcy data. Management Science, 34, 1403-1415. 27. Michalski, R. S. & Chilausky, R. L. (1980). Knowledge acquisition by encoding expert rules versus computer induction from examples: A case study involving soybean pathology. Int. J. Man-Machine Study, 12, 63-87. 28. Michie, D. (Ed.) (1984). Introductory readings in expert systems. New York: Breach. 29. Michalski, R., Mozetic, I., Hong, J. & Lavrac, N. (1986). The multi-purpose incremental learning system AQ25 and its testing application to three medical domains. In Proc. of the 5th Annual Nat'l Conference on Artificial Intelligence, pp. 2041-2045, Phil- adelphia, PA. 30. Mitchell, T. M. 0982). Generalization as search. Artificial Intelligence, lg, 203-226. 31. Parsaye, K. (1988). Acquiring and verifying knowledge automat- ically. A1 Expert, pp. 48-63. 32. Paterson, A. & Niblett, T. 0982). ACLS user manual. Scotland: Intelligence Terminal Ltd. 33. Porter, B.W., Bareiss, R. & Holm, R.C. (1990). Concept learning and heuristic classification in weak-theory domains. Artificial Intelligence, 45, 229-263. 34. Quinlan, J. R. (2979). Discovering rules from large collections of examples: a case study. In D. Michie (Ed.), Expert systems in the micro electronic age. Scotland: Edinburgh University Press. 35. Quinlan, J. R. (1986). Induction of decision trees. Machine Learning, 1, 81-106. 36. Rendell, L. A. (1986). A general framework for induction and a study of selective induction. Machine Learning, 1, 177-220. 37. Ruff, R. A. & Dietterich, T. G. (1989). What good are experiments. Proc. of the Sixth International Workshop on Machine Learning, pp. 109-222, Ithaca, New York. 38. Scott, P. D. & Markovitch, S. (1993). t~xperience selection and problem choice in an exploratory learning system. Machine Learning, 12, 49-67. 39. Shaw, M. J. & Gentry, J. A. (1988). Using an expert system with inductive learning to evaluate business loans. Financial Manage- ment, 17, 45-56. 40. Turban, E. (1992). Expert systems and applied artificial intelli- gence. New York: Macmillan. 41. Utgoff, P. E. (1989). Incremental induction of decision trees. Machine Learning, 4, 161-186. 42. Wardle, A. & Wardle, L. (1978). Computer aided diagnosis--a review of research. Meth. Inform. Med., 17, 15-28. 43. Winston, P. H. (1975). Learning structural descriptions from examples. In P. H. Winston (Ed.), The psychology o f computer vision. New York: McGraw-Hill. 
work_67pjqcjuqvgt5m6kqlcnfgkxae	----	Editorial: Knowledge Discovery and Business Intelligence Paulo Cortez1 Manuel Filipe Santos1 1 Centro Algoritmi, Dep. Information Systems, University of Minho, 4800-058 Guimarães, Portugal Email: pcortez@dsi.uminho.pt, mfs@dsi.uminho.pt 1 Introduction Due to advances in Information Technology, nowadays it is easy to collect, store, process and share data. In effect, the amount of data stored by organizations or indi- viduals is estimated to be growing exponentially [Lyman, 2003]. While Expert Sys- tems (ES) were originally devised to solely mimic human experts [Buchanan, 1986], there is a pressure to extract as much useful information as possible from past data. Hence, the current trend to include data-driven models, possibly integrated with expert-driven models, into the decision-making process. Within this context, there are two relevant terms: Knowledge Discovery (KD) and Business Intelligence (BI). KD is a branch of the Artificial Intelligence (AI) field that aims to extract use- ful and understandable high-level knowledge from complex and/or large volumes of data [Fayyad et al., 1996]. BI is an umbrella term that represents several computer architectures, tools, technologies and methods (e.g., Data Warehousing, On-line ana- lytical processing and KD) to access past data and support decision-making in public and corporate enterprises, from operational to strategic level [Turban et al., 2010]. KD and BI are faced with new challenges. For instance, recent communication technologies, such as WiMAX, lead to interesting network optimization problems, which can be solved using a KD approach. Also, when adopting fuzzy clustering methods, it is not clear how to correctly identify the ideal number of clusters. More- over, a large effort has been put on static analysis of social networks. Yet, these social networks exist in time and thus may evolve under a dynamic environment. Furthermore, while there are several public dictionaries with synonymy information, the building of thesaurus often requires manual work and is usually incomplete. This special issue, entitled “Knowledge Discovery and Business Intelligence” (KDBI), focuses on new KD approaches that aim to solve new challenges (such as previously described), leading to a potential valuable impact in several BI domains. This special issue consists of extended versions of papers from the 2nd KDBI work- shop of the 15th Portuguese Conference on Artificial Intelligence (EPIA 2011) that was held in Lisbon, Portugal. A total of 27 papers were submitted to the 2nd KDBI workshop, from which the best 10 papers were invited for this special issue. Each extended paper was reviewed by a minimum of three reviewers (related with both 1 the 2nd KDBI workshop and Expert Systems journal), and passed through two rounds of reviews. Finally, the best four papers were accepted, corresponding to an acceptance rate of 15%, when considering the initial 2nd KDBI submitted papers, and 40%, when considering the extended invited papers. In the next section, we provide a brief introduction to these papers. 2 Contents of the special issue For this special issue, four outstanding papers were selected. All these papers pro- pose novel intelligent data analysis methods that are tested over real-world data and with an analysis of their valuable impacts in the business domain. The paper of [Stolojescu et al., 2013] proposes an original KD methodology for the telecommunications domain. In particular, a Long Range Dependence (LRD) perspective is used to estimate the quality of base stations positioned within a wire- less network that uses the new WiMAX communication technology. The proposed approach can be used to optimize the functioning of the WiMAX network (e.g., turning the traffic more uniform and allowing to accommodate an increased num- ber of users). Moreover, it could also increase network security by detecting traffic anomalies. The work of [Nascimento et al., 2013] presents further research of a novel fuzzy additive spectral clustering method, termed FADDIS. Such method has the advan- tage of providing guidance for choosing the number of clusters, while it compares favorably with state of the art methods. In particular, FADDIS is potentially useful when modeling pairwise similarities by additive properties (e.g., finding thematic clusters of the activities conducted in an organization). The paper of [Oliveira and Gama, 2013] introduces a new approach for analyzing dynamic temporal changes in social networks, at both node-level and community level. In particular, three-order tensors are used to project the trajectories of social entities into a 2D space, which helps in understanding latent properties of the dy- namic social network. While the approach was tested in friendship networks among university freshmen, there are also potential benefits in business applications, such as churn prediction and viral marketing. The work of [Oliveira and Gomes, 2013] discusses a novel method for the enrich- ment of thesaurus by including information acquired automatically from dictionar- ies. While the method (which includes a clustering approach) was applied to enrich a Portuguese thesaurus with information from three dictionaries, it can be adapted to other languages. Such thesaurus is useful within several domains, including not only text mining and information retrieval (e.g., stemming) but also BI. For in- stance, it could help in the mapping of attributes with similar meaning but with different names, thus facilitating the extraction of data from multiple sources into a data warehouse. All these papers represent the best contributions to the Knowledge Discovery and Business Intelligence workshop of EPIA 2011. Following the original purpose of the KDBI workshop, we hope that this special issue contributes for enrichment of the KD and BI fields, including in particular an increased interaction between KD and BI. 2 Acknowledgments Many individuals contributed to this special issue. We would like to thank Dr. Jon G. Hall, Editor-in-Chief of Expert Systems, for his interest and support. We also wish to thank the other KDBI 2011 track (of EPIA) co-organizers, Nuno Mar- ques, Lúıs Cavique and João Gama, without their help this issue would not have been possible. Finally, we thank the authors, who contributed with their papers, and the reviewers (from the KDBI 2011 program committee and also others), who contributed with their expert and valuable opinions. References [Buchanan, 1986] Buchanan, B. G. (1986). Expert systems: working systems and the research literature. Expert systems, 3(1):32–50. [Fayyad et al., 1996] Fayyad, U., Piatetsky-Shapiro, G., and Smyth, P. (1996). Ad- vances in Knowledge Discovery and Data Mining. MIT Press. [Lyman, 2003] Lyman, P. (2003). How much information? http://www. sims. berke- ley. edu/research/projects/how-much-info-2003/. [Nascimento et al., 2013] Nascimento, S., Felizardo, R., and Mirkin, B. (2013). Laplacian normalization for deriving thematic fuzzy clusters with an additive spectral approach. Expert systems, pages –. [Oliveira and Gomes, 2013] Oliveira, H. and Gomes, P. (2013). Towards the auto- matic enrichment of a thesaurus with information in dictionaries. Expert systems, pages –. [Oliveira and Gama, 2013] Oliveira, M. and Gama, J. (2013). Visualization of evolv- ing social networks using actor-level and community-level trajectories. Expert systems, pages –. [Stolojescu et al., 2013] Stolojescu, C., Lenca, P., Moga, S., and Isar, A. (2013). Wimax traffic analysis and base stations classification in terms of lrd. Expert systems, pages –. [Turban et al., 2010] Turban, E., Sharda, R., Aronson, J., and King, D. (2010). Business Intelligence, A Managerial Approach. Prentice-Hall, 2nd edition. The authors Paulo Cortez Paulo Cortez is Associate Professor at the Department of Information Systems, University of Minho, Portugal. He is also researcher at the Intelligent Data Systems group of Centro Algoritmi, with interests in the fields of: Business Intelligence (Decision Support, Data Mining and Forecasting); and Artificial Intelligence (Neural 3 Networks, Evolutionary Algorithms and Applications). Currently, he is associate editor of the Neural Processing Letters journal and participated in 9 R&D projects (principal investigator in 2). He is co-author of more than eighty indexed (ISI, Scopus) publications in international journals (e.g., Expert Systems, Information Sciences) and conferences. Web-page: http://www3.dsi.uminho.pt/pcortez Manuel Filipe Santos Manuel Filipe Santos is Associate Professor at the Department of Information Sys- tems, University of Minho, Portugal, and leader of the Intelligent Data Systems group of Centro Algoritmi, with interests in the fields of: Business Intelligence, Data Mining and Learning Classifier Systems. He participated in several R&D projects (principal investigator in 3). He is also co-author of several publications in international journals (e.g., Artificial Intelligence in Medicine) and conferences (e.g., GECCO). 4 
work_6bfvqtsj4bb4blfcoovfzbfgoi	----	A novel Artificial Immune System for fault behavior detection Expert Systems with Applications 38 (2011) 6957–6966 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A novel Artificial Immune System for fault behavior detection C.A. Laurentys a, R.M. Palhares b,⇑, W.M. Caminhas b a Department of Electrical Engineering, Federal University of Minas Gerais, Av. Antônio Carlos 6627, 31270-010 Belo Horizonte, MG, Brazil b Department of Electronics Engineering, Federal University of Minas Gerais, Av. Antônio Carlos 6627, 31270-010 Belo Horizonte, MG, Brazil a r t i c l e i n f o a b s t r a c t Keywords: Artificial intelligence Artificial Immune System Fault detection Immune system Model-based fault diagnosis 0957-4174 � 2010 Elsevier Ltd. doi:10.1016/j.eswa.2010.12.019 ⇑ Corresponding author. Tel.: +55 (31) 3409 3457; E-mail address: palhares@cpdee.ufmg.br (R.M. Pal Open access under the El This paper presents an error detection methodology to enable fault detection inspired on recent immune theory. The fault detection problem is a challenging problem due to processes increasing complexity and agility necessary to avoid malfunction or accidents. The key challenge is determining the difference between normal and potential harmful activities. A promising solution is emerging in the form of Artifi- cial Immune System (AIS). In this article, Natural Killer (NK) immune cells mechanisms inspired an AIS. The AIS proposed uses recent biological mechanism such as: NK activation and education machinery. DAMADICS benchmark was applied to compare the proposed AIS performance to others fault detection algorithms. The results show that the novel approach developed provides better detection rate and false alarms tradeoff when compared to other methods in literature. � 2010 Elsevier Ltd. Open access under the Elsevier OA license. 1. Introduction Brazil has been facing an increasing rate of incidence related to manufacturing facilities, reaching more than 234 thousands in 2007. According to MPS studies between 1999 and 2007, the an- nual accidents related to manufacturing lays around 35% of total accidents (MPS, 2009). In 2007, for instance, manufacturing facilities accidents were the major cause of accidents as depicted by Fig. 1. Industrial statistics shows that 70% of the accidents are caused by human mistakes and operators need decision making tools to create a more trustable industrial environment (Venkatasubrama- nian et al., 2003). In this context, fault detection is one of the vital components for the Abnormal Event Management (AEM) Luo, Zhang, Zhao, & Zhang, 2004; Venkatasubramanian et al., 2003. AEM becomes challenging due to the size and complexity of proce- dures and the broad activities scope, encompassing a variety of factors such as: degradation of the process, inadequate, incomplete and not reliable measurements. The AEM has been turning into a challenging problem due to the size and complexity of procedures and the broad activities scope. Today’s challenge is the AEM fault detection automation, and the usage of computer systems to improve it, applying artificial intelligence methods such as fuzzy (Mendonca, Sousa, & Sa da Costa, 2009; Saravanan, Cholairajan, & Ramachandran, 2009) and neural networks (Tang & Yao, 2008; Wu & Liu, 2009). Natural Computing (De Castro, 2005; Tang & Yao, 2008), more specifically, Artificial Immune System (AIS) Timmis et al., 2008, has been sup- porting this context (Gan, Zhao, & Chow, 2009; Ji & Dasgupta, fax: +55 (31) 3409 4850. hares). sevier OA license. 2009). The AIS have emerged from attempts to model and apply immunological principles or inspirations into the development of new computational tools. Artificial Immune Systems (AIS) are de- fined as a new computational paradigm based on metaphors of the biological immune systems. Immune-based techniques have been applied in a wide area of applications such as intrusion detection (Powers & He, 2008), opti- mization algorithms (Paquete & Stutzle, 2009), robust control (Gui- maraes, Palhares, Campelo, & Igarashi, 2007) and others (DeCastro and Von Zuben, 2009). The powerful information processing capabil- ity, pattern recognition, learning, memory and immune distributive nature provide rich metaphors for its artificial (computational) counterpart. The novel technique is based on a synergy of Immunological inspiration, its formalization using Natural Computing , and the Fault detection context as shown by Fig. 2. In this article the Fault detection block shown by Fig. 2 defined the performance indicators to allow the comparison with literature algorithms; immunology provided the inspiration to develop the AIS fault detection scheme named NKC Algorithm and Natural Computing act a driver to translate the immune inspirations into algorithms. Firstly, it is important to stress that the key article contribution is an AIS inspired on recent literature of Natural Killer (NK) cell mechanisms (Lanier, 2008; Larrea, Alvarez, Soto, Vazquez, & Gonz- alez, 2008; Luci & Tomasello, 2008) to enable AEM fault detection. Secondly, the novel AIS is a population-based algorithm. The immune system has a set of NK cells that dynamically interacts to reach an overall outcome and it is not based on a single NK cell as proposed by other literature models (Elmeziane, Berrada, & Kas- sou, 2008). http://dx.doi.org/10.1016/j.eswa.2010.12.019 mailto:palhares@cpdee.ufmg.br http://dx.doi.org/10.1016/j.eswa.2010.12.019 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa http://www.elsevier.com/open-access/userlicense/1.0/ http://www.elsevier.com/open-access/userlicense/1.0/ 0 100 200 300 400 500 600 700 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 # Total #Manufacturing 35.91% 15.29% 7.07% 6.62% 5.93% 5.31% 12.90% 2.72% 8.25% 0.00% 5.00% 10.00% 15.00% 20.00% 25.00% 30.00% 35.00% 40.00% Ma nu fac tur ing Se rvi ce s Co mm erc e Bu ild ing He alt hy Tr an sp or tat ion Ign or ed Pr im ary Se cto r Ot he rs (a) (b) T ho us an ds Fig. 1. (a) Accidents chart showing the number of total accidents and only the ones related to the manufacturing environment during 1999 and 2007 (last year published by MPS). (b) MPS accidents reasons pareto chart regarding 2007: Manufacturing leads the key root cause of Brazilian accidents. Fig. 2. Immune inspired fault detection: A dynamic feedback between fault detection, immunology and natural computing are the major components adopted in this article. 6958 C.A. Laurentys et al. / Expert Systems with Applications 38 (2011) 6957–6966 Finally, this novel proposition is applied to a DAMADICS bench- mark (Barty, 2006) resulting with a better detection rate and false alarms tradeoff when compared to other methods in literature. The article is structured in the following sections: � Natural Killer Immune Cell: describes Natural Killer Cells features that inspired the AIS, � Materials and Methods: presents analogies between the pro- posed AIS and NK cells, � Calculation: describes the DAMADICS benchmark used for AIS tests, � Results and Discussions: discusses algorithm results and compare its performance, � Conclusions: points out the major benefits and AIS improvements. 2. Natural killer immune cell This section describes Natural Killer cells recent immune litera- ture and details key features that inspired the proposed AIS. 2.1. NK cell immune overview NK cells are an essential component of immunity against tu- mour and viral infected cells (Raulet, Vance, & McMahon, 2001). NK cells are involved in early defenses (Luci & Tomasello, 2008). In the context of engineering, i.e., in error detection, this specific characteristic is useful once it is desired to detect a fault as soon as it began, especially in the case of incipient faults. The NK cells triggering is regulated by activating and inhibitory surface cell receptors (Lanier, 2008; Vivier, Tomasello, Baratin, Walzer, & Ugolini, 2008). It means that NK cell machinery is com- posed by activating and inhibitory molecules that are able to con- nect them into target cells surface molecules. NK cells detect changes in the target cell surface – malignant transformation or infection – resulting in loss or gain of molecules that are detected by the NK cell surface receptors. Based on the features described, the NK cell immune model that defines the NK lifecycle is described in the following steps: � Initiation of receptors expression: is described by generation of NK cells inhibitory and activation receptors through a stochastic process to distribute the NK Cell receptors – this procedure will generate immature NK cells, � Education process: aims to maximize the discriminatory proper- ties of the immature NK Cells receptors maximizing its discrim- inatory properties – this procedure will generate mature NK cells, � Recognition phase: mature NK cells interacts with others cells applying the balance between negative and positive signals; they may be triggered (generating activated NK cells) or not (generating inactivated NK cells). If it is triggered, it releases molecular substances called cytokine to signal it to other cells. These key steps are further detailed in the next sections. 2.2. NK Immune cell receptors initiation The NK Cell receptors initiation is a process that will produce the NK Cells receptors. According to literature the receptors are generated by a stochastic mechanism. A model found in the litera- ture (Lanier, 2008) describes it as an adjustment mechanism based on ‘‘At least one theory’’ and ‘‘Zero sum balance’’. The ‘‘At least one’’ hypothesis is imposed to describe NK cells developmental process. According to this theory, to achieve self- tolerance (Raulet & Vance, 2006) (does not activate improperly) NK cells must have at least one inhibitory receptor. A process must equip the NK cell with at least one inhibitory receptor; otherwise it will be activated improperly. C.A. Laurentys et al. / Expert Systems with Applications 38 (2011) 6957–6966 6959 The ‘‘Zero sum balance’’ fundamentals are illustrated by Fig. 3. The shaded section indicates a stimulatory signal. According to (Luci & Tomasello, 2008), the NK cells development states are: � Mature – has already achieved ‘‘Zero sum balance’’ (e.g. see Fig. 2 cell ‘‘A’’) through inhibitory and activating receptors balance, � Immature or hyporesponsive (e.g. see Fig. 2 cell ‘‘B’’) – NK cell will not activate due to absence of inhibitory or activating receptors, � Possible mature (e.g. see Fig. 2 cell ‘‘C’’) – might achieve the ‘‘Zero sum balance’’ through a further education process. 2.3. NK immune cell education process Although a stochastic mechanism appears to underlie the initial expression of the receptors, the final functional NK cell repertoire is shaped by an education process (Raulet & Vance, 2006). The goal of the education process is to maximize the discrimina- tory properties of the NK cells. Their ultimate effects are to ensure a repertoire of NK cell is both self-tolerant and useful. The self-toler- ant property avoids NK cell inappropriate activation, specifically to activate in a normal situation, destroying cells of the own organism. There are evidences that the education mechanism maximizes the NK cells discrimination properties but this process and its machinery is still poorly understood (Raulet et al., 2001). 2.4. NK immune cell recognition phase NK cells functions are regulated by a dynamic balance between negative and positive signals initiated after engagement of cell sur- face inhibitory and activating receptors (Raulet et al., 2001). A dy- namic balance is illustrated by Fig. 4. The states of a mature NK cell are: � Triggered by Absence of Inhibitory Signals: NK Cell receptors interacted with target cells receptors and it was triggered by absence of target cell inhibitory receptors, � Triggered by Activating Signals: NK Cell receptors interacted with target cell stimulatory receptors and the NK cell was triggered by the presence of stimulatory receptors, � Not Triggered Region: NK cell receptors interacted with target cell inhibitory receptors and the NK cell will not be triggered. Fig. 3. Model for stimulatory and inhibitory initiation. Cell ‘‘A’’ has achieved balance in the amplitudes of positive and negative signals (influenced by the number of receptors, ligand affinities) and is therefore maximally sensitive to changes in host cell. Cell ‘‘B’’ lacks inhibitory receptors; these cells are rendered hyporesponsive. Cell ‘‘C’’ has both stimulatory and inhibitory receptors but has not reach Zero sum balance. It is important to stress the role of innate cytokines – a molec- ular communication substance released by innate cells – is a mech- anism that might alter a predefined dynamic balance. In general, NK cells are able to release different types of cytokine depending on the molecular interaction with other cells. The dynamic balance of a particular NK cell might be altered by innate cytokines re- leased by other cells as illustrated by Fig. 4 (d) facilitating the trig- ger or turning it more difficult. 3. Materials and methods The Section 2 described NK cells immune mechanisms and its machinery that inspired an Artificial Immune System (AIS) for fault detection. The AIS proposed in this article was inspired on the NK cells described on item II. The sources of inspiration were: (a) Initiation of the NK Cells receptors: It is generated stochastically using the ‘‘At least one’’ hypothesis and the ‘‘Zero sum balance’’, (b) Education process: NK immune response is guided by a dynamic balance of two signal receptors – activating and inhibitory. NK receptors are adjusted by an education mech- anism to maximize its performance, (c) Recognition phase: The generated NK cells will monitor the other cells in other to verify if they are activated based on their inhibitory and activating receptors. The NK cells immune context is interpreted as an analogy to propose a fault detection AIS. In this scenario, the normal and anomalous available patterns will behave as inhibitory and activat- ing signal of the fault detection methodology. Therefore, the nor- mal patterns are understood as inhibitory receptors and the anomalous ones as activating receptors. A stochastic mechanism will initiate NK cells population recep- tors based on the available patterns (normal and abnormal). Fur- ther, an education process will be necessary to optimize the NK cell population. Fig. 5 summarizes key analogies of the NK cells im- mune system mechanisms inspirations. The Methodology section will detail how each of these steps were artificially developed generating the AIS proposed. 3.1. Initiation of the NK cells receptor This phase applies a metaphor for how stimulatory and inhibi- tory interactions of the NK cell receptors are integrated during the receptor acquisition phase of NK cell development generating immature NK cells. Fig. 6 describes the key steps of the Artificial NK Receptors Ini- tiation algorithm which is immune inspired in the initiation of the NK Cells receptors. It is important to mention the biological analogy with the algo- rithm using the ‘‘At least one’’ hypothesis and the ‘‘Zero sum mod- el’’ for initiating the NK Cells receptors expression (Raulet et al., 2001). The artificial At least one algorithm represented by block (1) at Fig. 6 forces a NK Cell receptor repertoire with at least one inhibitory receptor in order to create responsive NK cells (Matured or possible Matured). The artificial Zero sum algorithm represented by block (2) at Fig. 6 defines the NK cell activation region. It is firstly based on the number of receptors as a first indicative of Zero sum balance. In this case the type of the receptor alternate between {F} and {N} using a random choice, producing a balanced set of inhibitory and activating receptors. Secondly, it considers the affinity be- tween the receptors of inhibitory and activating to define NK cell response. The Zero sum model is represented by a line equation that fits the average Euclidian distance points between a specific inhibitory receptor and all activating receptors. Fig. 4. NK cell recognition phase. Left panels show the NK cell balances. Activating target cells often down regulate inhibitory – polygons (b) and up regulate activating receptors – circles (c). They also induce innate cytokines that bias NK-cell recognition and signaling in favor of triggering (d). Innate cytokines (stars and sliding weight) alter the equilibrium signaling machinery. Right panels shows interaction between the two cells. 6960 C.A. Laurentys et al. / Expert Systems with Applications 38 (2011) 6957–6966 These cells might need a further educational process to balance the strengths of each inhibitory and activate signaling and there- fore are considered immature NK cells population. In order to illustrate the stochastic mechanism of NK cell forma- tion a two class problem, a two dimensional space, is depicted at Fig. 7. It should be stressed that the number of NK cells activating and inhibitory receptors will be equal and the NK cells generated will have at least one inhibitory receptor and lines to define the Zero sum balance of it. The algorithm takes as input: � Fault Set ‘‘f’’ – [x1 . . . xn]: These are the n process variables val- ues affected when the fault occurs. These components will gen- erate the NK cells activating receptors. � Normal Set ‘‘n’’ - [x1 . . . xn]: These are the same n process vari- ables values in normal conditions. These components will gen- erate the NK cells inhibitory receptors. The key algorithm parameters are: � Cells_Pop –‘‘m’’: It is a constant real number defined by the user. It defines the number of Immature Artificial NK cells that the population will have. � ReceptorsbyCell – ‘‘r’’: It is a constant real number defined by the user. It represents the number of receptors (inhibitory and acti- vating) that the NK cells will posse. � Size_Receptors – ‘‘s’’: It is a constant real number defined by the user. It represents the length of the receptor generated. Fig. 5. Comparison of the NK cell immune model and the AIS key steps. C.A. Laurentys et al. / Expert Systems with Applications 38 (2011) 6957–6966 6961 3.2. Education mechanism The education mechanism goal is to maximize the NK popula- tion discriminatory properties such as NK cell self-tolerance and used by stochastic mechanisms. The education mechanism gener- ates mature NK Cells through the immature population. The Artificial Education Mechanism algorithm proposed consid- ers the false positive rate as a self-tolerance measurement and individual/population performance as usefulness measurement. The algorithm is presented at Fig. 8 along its key parameters. The key of the Artificial Education Mechanism algorithm is repre- sented by block (1) in Fig. 8, where the discriminatory properties (self-tolerance, performance and maturation) are assessed. Fig. 6. Artificial NK Receptors Initiation algorithm – Implements the stochastic initiation of the NK cells receptors. Firstly, a threshold mature is checked to verify if the minimum performance is reached otherwise the cell will be discarded. Secondly, the Artificial Education Mechanism algorithm considers the NK cell self-tolerance, meaning in this context that cells will not be triggered improperly, generating false alarms. Therefore, a mechanism of avoiding a NK cell detecting own inhibitory recep- tors as activating receptors is implemented. In this case, the algo- rithm will erase the artificial Zero sum model (line equations) that triggers in inappropriate situation. If all zero sum models are not self tolerant, the cell is deleted. Finally, the performance is assessed as being the NK cell useful- ness. NK cell correct triggers for normal and fault situation, consid- ering own cell receptors and a disjoint validation set of known receptors, are applied. Fig. 7. Example of receptors generated through Artificial NK Receptors Initiation algorithm. The activating receptors marked with ‘‘+’’ refer to anomalous conditions and the inhibitory receptors (‘‘�’’) to normal conditions. The circles around the receptors indicate the stochastic receptors selected by Artificial NK Receptors Initiation algorithm (activating and inhibitory receptors). The lines marked with ‘‘�’’ indicate the zero sum lines generated. Fig. 8. Artificial Education Mechanism algorithm – Implement the educational process to maximize performance of the NK cell population. Fig. 9. (a) Illustrative example of a mature NK cell generated. The activating receptors ‘‘+’’ refers to possible anomalous situation and the inhibitory receptors marked with ‘‘�’’. The circles in the lines (zero sum models) indicated the Zero sum self-tolerant marked with circles. (b) A NK cell is only added to the mature population if it increases overall population performance. In this example, the population has four mature NK cells. 6962 C.A. Laurentys et al. / Expert Systems with Applications 38 (2011) 6957–6966 The NK cell will be added to the mature population if it im- proves the overall population performance, turning into a mature NK cell (see Fig. 9 for details). This constraint aims to create a pop- ulation that does not have redundant knowledge, and always increasing its capacity to improve performance with a restricted number of NK cells (parameter Max_cell). The key algorithm parameters are: � Max_Cells – ‘‘max’’: its constant real number defined by the user. It defines the maximum number of Immature Artificial NK cells that the mature population will have. � Threshold_mature – ‘‘tr’’: its constant real number defined by the user. It represents minimum NK cell performance to join the mature NK cell population. � Attempts_to_join – ‘‘aj’’: its constant real number defined by the user. It represents the number of attempts to maturate NK cells. It is a stop criterion of the education mechanism. 3.3. Recognition phase The mature NK cells population will interact with unknown cell receptors that it desires to classify in safe or dangerous situa- tion using the interaction machinery described by Fig. 3. The cells will interact with the stimulatory and inhibitory signal from the target cell generating: � Innate cytokine – based on the Euclidian distance of the target cell receptors and its Zero sum balances: positive (above Zero sum balance) and negative (below it) signals. � Status (triggered or not triggered) – triggered when positive innate cytokines are greater than negative innate cytokines, otherwise set as not triggered. The fault detection AIS considers fault situation using the majority vote of the mature NK cells status indicates triggered, otherwise the fault detection AIS indicates normal condition as shown by Fig. 10. 4. Calculation In this paper the DAMADICS benchmark is used to validate the proposed methodology and to compare its performance to other fault detection methods. The benchmark consists of an actuator controlling the water in- put flow in a boiler. The boiler presented is part of a process of evaporation comprised of five stages of a sugar factory in Poland. The benchmark was developed by a research group in Europe called DAMADICS: Development and Application Methods for Diagnosis of Actuator in Industrial Control Systems. The process actuator is composed of three elements: a control valve V, the pneumatic servo motor S and the positioner P. The valve V controls the flow of water that passes through a pipe feed- ing the boiler. The pneumatic servo motor S set the place of valve V plug’s in order to act on the rate of flow of water. The positioner is the device used to correct a mispositioning of the shaft of the en- gine, caused by internal or external sources, such as: friction, vari- ations in pressure supplied to the servo motor and so on. Fig. 11 illustrates the schematic diagram of the actuator. Fig. 10. (a) The proposed AIS is a population-based algorithm that allows interaction between the NK cells output. Fig. 11. DAMADICS industrial process schematic diagram. Table 1 Input variables applied to the model. Input Range Unit Description GV [0 1] % Output signal from controller PI – Pa The value of the input pressure P2 – Pa The value of the output pressure Tl – C Fluid temperature Table 2 Output variables applied to the model. Output Range Unit Description F – t/h Medium flow X [0 100] % Disturbed value of rod displacement C.A. Laurentys et al. / Expert Systems with Applications 38 (2011) 6957–6966 6963 4.1. Normal operation process modeling The process described was modeled using input and output variables defined by Tables 1 and 2. In order to provide a dynamic system model to simulate normal conditions, a neural network was applied. The neural network Fig. 12. Feed-forward multilayer perceptron ap structure was generated by a threshold validation error and it is depicted by Fig. 12. For training phase 5,000 points in normal operation were used: 50% for training and last 50% for test. The sampling time applied was 1 s. The sample time was picked in order to avoid aliasing. The stop training criterion was the medium square error below 0.1%. The neural network parameters generated by training are de- tailed in Appendix A. plied to model the normal plant dynamics. Table 3 Artificial fault available at the DAMADICS. Fault Description Control valve faults fl Valve clogging f2 Valve plug or valve seat sedimentation f3 Valve plug or valve seat erosion f4 Increased cf valve or hush ng friction f5 External leakage (leaky bushing, covers, terminals) f6 Internal leakage (valve tightness) f7 Medium evaporation or critical flow Pneumatic servo-motor faults f8 Twisted servo-motor’s piston rod f9 Servo-motor’s housing or terminals tightness f10 Servo-motor’s diaphragm perforation f11 Servo-motor’s spring fault Positioner faults f12 Electro-pneumatic transducer fault fl3 Rod displacement sensor fault f14 Pressure sensor fault f15 Positioner feedback fault General faults/external faults f16 Positioner-supply pressure drop f17 Unexpected pressure change across the valve f18 Fully or partly opened bypass valves f19 Flow rate sensor fault Fig. 13. k-fold cross-validation using k = 10. Table 4 List of parameters and corresponding levels considered for the experiments. Id Algorithm and parameter name Levels � + A Artificial NK Receptors Initiation: Size_Receptors 10 30 B Artificial NK Receptors Initiation: ReceptorsbyCell 3 5 C Artificial Education Mechanism: Threshold_mature 40 60 D Artificial Education Mechanism: Attempts_to_join 3 5 E Artificial Education Mechanism: Max_Cells 3 5 6964 C.A. Laurentys et al. / Expert Systems with Applications 38 (2011) 6957–6966 It is important to stress that the generated model was achieved by neural network training using normal data. It means that the model is able to describe the plant behavior without faults. Therefore, the expected behavior of the residue (difference be- tween process outputs and model outputs) is to be near to zero in absence of faults. Also, none filtering approach was applied to the residue. Fig. 14. Main effect plot for the exp 4.2. DAMADICS artificial faults The DAMADICS simulation tools allowed several artificial faults simulation, to validate the AIS proposed. Table 3 presents the faults available to be applied in this context. In order to compare the proposed AIS performance to others found in the literature, the following faults were considered: f1, f7, f15. The restricted faults set are used once other authors only provided their result for this set. For each fault a 5000 points database is available to train and assess the AIS overall performance. The output residues (the differ- ence between the neural network model and the real measured outputs) from X and F in normal and abnormal situations were ap- plied to generate the inhibitory and activating receptors. The resi- due in normal and abnormal situation was picked to represent NK cell receptors once faults manifest as errors and will enable further fault detection. Therefore, the selected model has a key role on the proposed algorithm. 5. Results and discussions In order to compare the proposed AIS performance to other methods in the Literature applied to the DAMADICS, a cross-valida- tion method was applied. The following DAMADICS faults were as- sessed: f1, f7, f15, f17. A train and test disjoint set was generated for each fault. The database used a stratified k-fold cross-validation where k-1 folds were training folds and one for test. A stratified 10-fold cross-val- idation (Demsar, 2006) was applied to estimate the average error as depicted by Fig. 13. eriments defined in this article. Fig. 15. Interaction plot for the experiments defined in this article. Table 5 Performance indicators for DAMADICS fault detection. Id AIS proposed Paim Previdia dr (%) fa (%) dr (%) fa (%) dr (%) fa (%) f1 98.3 ± 1.1 0 ± 0 98.22 0 83.0 0.02 f7 98.5 ± 1.2 0 ± 0 98.89 0 83.0 0.02 f15 98.7 ± 2.8 0 ± 0 98.89 0 92.0 0.02 f17 95.9 ± 5.6 0 ± 0 98.78 0 99.0 0 C.A. Laurentys et al. / Expert Systems with Applications 38 (2011) 6957–6966 6965 The normal and fault data were uniformly distributed in order to generate balanced folds. Each of the 10-folds has a total of 1000 points (500 belonging to normal dataset and 500 to fault dataset). Since only one 10-fold cross-validation run may not be enough to obtain a reliable error estimate, the overall performance of the AIS was computed as the average validation error for 100 runs of the stratified 10-fold cross-validation. An important decision that needs to be discussed is about the algorithm parameters definition. Design of Experiments (DOE) is needed for experiment with real-life systems, and with either deterministic or random simulation models (Kleijnen, 2008). DOE was applied to verify the proposed AIS parameter sensitiv- ity analysis and define what set of parameters produce best performance. The experimental design considered all AIS algorithm parame- ters that were described in the Methodology section. Table 4 lists the algorithm parameters and their associated values (or levels) applied on the experiments stressing the best ones for the DAMA- DICS benchmark. The simulations average results for each set of parameters are shown in Appendix B. In order to understand the effect of each parameter in the detection rate the main effect plot was generated for the simulations and is shown by Fig. 14. Regarding the algorithm sensibility analysis, the parameter ‘‘Size_Receptors’’ has shown to be the most sensitive one (differ- ence of 5.2%) for the levels experimented followed by the Recep- torsbyCell reaching 0.7% difference as shown by Fig. 14. Also to verify how each parameter affects the others the inter- action plot was generated and it is shown by Fig. 15 The AIS proposed performance using best parameters levels (indicated in bold at Table 4) are shown by Table 5 along with its standard deviations. Table 5 also presents the results found in lit- erature using the following performance indicators: � dr – Average detection rate – as the % of correct classifications, � fa – Average false alarm – as the % of false positive classifications. Analyzing faults ‘‘f1’’, ‘‘f7’’ and ‘‘f15’’ the proposed algorithm achieved higher detection rate than the Previdia (Previdia & Parisin, 2006) algorithm. In general aspects, the proposed AIS detection rate is very high when compared to Previdia algo- rithm with statistically the same false alarms rate, except the fault ‘‘f17’’. Compared the assessed result for fault ‘‘f17’’, it was verified that the algorithms have statistically the same detection rate (confi- dence level of 95%) with the proposed AIS presenting significantly more false alarm rate (almost reaching 1% of false alarm) than Pre- vidia algorithm. This difference is due to the residue analysis does not clearly provide separation between normal and abnormal situation, gener- ating for the AIS proposed an increased false alarm rate. This fact reinforces the coupling of the modeling procedures to the AIS proposed algorithm performance. 6. Conclusions This paper has proposed and validated a novel Artificial Im- mune System (AIS) based approach for fault detection inspired on the recent NK cell immune theory. The approach validation used a stratified 10-fold cross-validation and DAMADICS fault detection database. The DOE was applied to check sensibility analysis and define the best parameter performance. Comparing the proposed AIS perfor- mance to the Previdia algorithm as shown in Table 4, it is clear that the proposed AIS is able to provide a better detection rate with the same or in some cases lower false alarm rate for several DAMADICS database faults. The results show that the proposed AIS automates the AEM fault detection step providing a better tradeoff among detection rate and false alarm rate through the new approach proposed. Acknowledgment This work has been supported by the Brazilian agencies CNPq and FAPEMIG. 6966 C.A. Laurentys et al. / Expert Systems with Applications 38 (2011) 6957–6966 Appendix A. Neural network parameters IWð1; 1Þ¼ �44:7848 15:2394 1:3094 �4:4656 8:8382 �619705 216:2751 65:2473 �40:1468 32:6456 24:0152 42:3434 32:4858 �30:7896 41:79 62:4357 2 6664 3 7775 LWð2; 1Þ¼ �0:65707 0:0073388 �0:074369 0:088101 0:085571 �0:02992 0:2211 �0:079982 � � bð1Þ¼ 16:695 �107:0308 �31:901 �264309 2 6664 3 7775 bð2Þ¼ 0:52836 0:59037 � � Appendix B. Individual experiments results Id A B C D E dr (%) fa (%) 1 + + + + � 98.24 ± 2.38 0 ± 0 2 + � + + + 93.73 ± 4.23 0.01 ± 0.01 3 � � � + � 91.92 ± 5.79 0.24 ± 0.48 4 � � + � + 93 ± 4.65 0 ± 0 5 + + + � + 98.8 ± 1.39 0 ± 0 6 + � + � � 92.14 ± 5.29 0.09 ± 0.18 7 + + � � 98.1 ± 1.32 0 ± 0 8 � � � + + 93.41 ± 4.64 0 ± 0 9 � + + + + 98.07 ± 2.26 0 ± 0 10 � � + + � 94.09 ± 3.95 0 ± 0 11 + + � + � 98.72 ± 1.09 0 ± 0 12 � � + + + 91.77 ± 5.57 0.6 ± 1.21 13 � + � � � 96.88 ± 2.37 0 ± 0 14 � + � + + 97.09 ± 4.04 0 ± 0 15 + + � � + 98.29 ± 1.32 0 ± 0 16 � + � � 96.58 ± 3.51 0 ± 0 17 + � + � + 93.35 ± 4.58 0 ± 0 18 + + + + + 99.33 ± 1.32 0 ± 0 19 + + � + + 98.3 ± 1.44 0 ± 0 20 + � � + � 93.63 ± 4.54 0.03 ± 0.06 21 � + + � + 97.96 ± 1.41 0 ± 0 22 � + � � + 98.02 ± 1.38 0 ± 0 23 + � � � � 92.51 ± 5.07 0.09 ± 0.18 24 + � � � + 91.96 ± 5.49 0.09 ± 0.15 25 + � � � 98.15 ± 1.27 0 ± 0 26 � + + � � 96.95 ± 2.82 0 ± 0 27 � � � � + 92.74 ± 4.89 0 ± 0 28 + � + � 92.08 ± 5.26 0.09 ± 0.18 29 � + + + � 97.83 ± 1.49 0 ± 0 30 � � + � � 92.81 ± 4.94 0.03 ± 0.05 31 + � � + + 92.98 ± 4.73 0 ± 0 32 � � � � � 92.54 ± 5.18 0.09 ± 0.18 References Barty, M. Z. (2006). In Specification of actuators intended to use for benchmark definition. Warszawa, Poland: Warsaw University of Technology, Institute of Automatic Control and Robotics. De Castro, L. N. (2005). Fundamentals of natural computing: Basic concepts, algorithms, and applications. Chapman and Hall/CRC Computer and Information Sciences, Taylor & Francis, Inc. ISBN: 1584886439. DeCastro, P. A. D., & Von Zuben, F. J. (2009). BAIS: A Bayesian Artificial Immune System for the effective handling of building blocks. Information Sciences, 179, 1426–1440. Demsar, J. (2006). Statistical Comparisons of Classifiers over Multiple Data Sets. Journal of Machine Learning Research, 7, 1–30. Elmeziane, R., Berrada, I., Kassou, I. (2008). A new artificial immune system for the detection of abnormal behaviour, studies in computational intelligence. Springer Berlin/Heidelberg, 149, 113–122. doi:10.1007/978-3-540-70560-410. Gan, Z., Zhao, M., & Chow, T. (2009). Induction machine fault detection using clone selection programming. Expert Systems with Applications, 36(4), 8000–8012. ISSN 0957-4174, doi:10.1016/j.eswa.2008.10.058. Guimaraes, F. G. R., Palhares, M., Campelo, F., & Igarashi, H. (2007). Design of mixed control systems using algorithms inspired by the immune system. Information Sciences, 177, 4368–4386. Ji, Z., & Dasgupta, D. (2009). V-detector: An efficient negative selection algorithm with ‘probably adequate’ detector coverage. Information Sciences, 179, 1390–1406. Kleijnen, J. P. C. (2008). Design of experiments: Overview. In S. J. Mason, R. R. Hill, L. Mönch, J. O. Rose, & J. W. Fowler (Eds.), Proceedings of the 2008 Winter Simulation Conference (pp. 479–488). The Netherlands: Tilburg. Lanier, L. L. (2008). Up on the tightrope: natural killer cell activation and inhibition. Nature Immunology, 5, 495–506. Larrea, C. L., Alvarez, B. S., Soto, A. L., Vazquez, A. L., & Gonzalez, S. (2008). The NKG2D receptor: sensing stressed cells. Cell Press - Elsevier Ltd. Luci, C., & Tomasello, E. (2008). Natural killer cells: Detectors of stress. The International Journal of Biochemistry & Cell Biology, 40, 2335–2340. Luo, M., Zhang, D. H., Zhao, Y. Z., & Zhang, J. B. (2004). A structured methodology for the development of anticipative event management systems with self-recovery capability, IECON 2004. 30th Annual Conference of IEEE (vol. 3(2–6), pp. 2614– 2619). Mendonca, L. F., Sousa, J.M.C., & Sa da Costa, J.M.G. (2009). An architecture for fault detection and isolation based on fuzzy methods. Expert systems with applications (vol. 36(2), part 1, pp. 1092–1104). ISSN 0957-4174, doi: 10.1016/ j.eswa.2007.11.009. MPS, Brazilian Official Work on Statistics (2009). <http://creme.dataprev.gov.br/ AEAT/GLiq/Liq01/Liq01.htm>. Paquete, L. T., & Stutzle, L. (2009). Design and analysis of stochastic local search for the multiobjective traveling salesman problem. Computers & Operations Research, 36, 2619–2631. Powers, S. T., & He, J. (2008). A hybrid artificial immune system and Self Organising Map for network intrusion detection. Information Sciences, 178, 3024–3042. Previdia, F., & Parisin, T. (2006). Model-free actuator fault detection using a spectral estimation approach: the case of the DAMADICS benchmark problem. Control Engineering Practice, 14, 635–644. Raulet, D. H., & Vance, R. E. (2006). Self-tolerance of natural killer cells. Nature Reviews, 6, 520–531. Raulet, D. H., Vance, R. E., & McMahon, C. W. (2001). Regulation of the Natural Killer Cell Receptor Repertoire. Annual Review of Immunology, 19, 291–330. Saravanan, N., Cholairajan, S., & Ramachandran, K.I. (2009). Vibration-based fault diagnosis of spur bevel gear box using fuzzy technique. Expert systems with applications (vol. 36(2), part 2, pp. 3119–3135). ISSN 0957-4174, doi:10.1016/ j.eswa.2008.01.010. Tang, K., & Yao, X. (2008). Special issue on ‘Nature inspired problem-solving’. Information Sciences, 178, 2983–2984. Timmis, J., Hart, E., Hone, A., Neal, M., Robins, A., Stepney, S., et al. (2008). Immuno- engineering. In 2nd IFIP International Conference on Biologically Inspired Collaborative Computing (pp. 3–17). Milan, Italy: 20th IFIP World Computer Congress. Venkatasubramanian, V. et al. (2003). A review of process fault detection and Diagnosis. Computers and Chemical Engineering, 27, 293–311. Vivier, E., Tomasello, E., Baratin, M., Walzer, T., & Ugolini, S. (2008). Functions of natural killer cells. Nature Immunology, 9, 503–510. Wu, J., & Liu, C. (2009). An expert system for fault diagnosis in internal combustion engines using wavelet packet transform and neural network. Expert systems with applications (vol. 36(3), part 1, pp. 4278–4286). ISSN 0957-4174, doi:10.1016/j.eswa.2008.03.008. http://www.creme.dataprev.gov.br/AEAT/GLiq/Liq01/Liq01.htm http://www.creme.dataprev.gov.br/AEAT/GLiq/Liq01/Liq01.htm A novel Artificial Immune System for fault behavior detection Introduction Natural killer immune cell NK cell immune overview NK Immune cell receptors initiation NK immune cell education process NK immune cell recognition phase Materials and methods Initiation of the NK cells receptor Education mechanism Recognition phase Calculation Normal operation process modeling DAMADICS artificial faults Results and discussions Conclusions Acknowledgment Neural network parameters Individual experiments results References 
work_6cljm6hxu5gz3nywbedzajaeaq	----	People | The University of Edinburgh Skip to main content Schools & departments MyEd Search: Search School of GeoSciences Menu School of GeoSciences home Undergraduate degrees Postgraduate degrees New and Returning Students - Who to Contact Current Research News & events Facilities For Staff & Students About Us Our People Equality Diversity & Inclusion Alumni and donors FAQs for UG and PGT students (secured) Home GeoSciences Our People Contact us People People working in and with the School of GeoSciences Key Contacts Academic Research Technical Administrative Computing Honorary Visitors Everyone One Page profiles/feed/ limit=20;pages=4; {{{script}}} {{{css}}} {{#if alpha}}{{{alpha}}}{{else}}{{/if}} {{#if singleton}} {{#each people}} {{title}} {{firstname}} {{surname}} {{jobtitle}} {{roomnum}} {{building_info.building_name}} {{building_info.address1}} {{building_info.address2}} {{building_info.address3}} Contact details {{#if phone}} Work: {{phone}} {{/if}} Email: {{email}} {{#if website}} Web: Personal website {{/if}} {{#if pureurl}} Web: Research Explorer publication list {{/if}} {{#if biography}} Biography {{{biography}}} {{/if}} {{#if research}} Research {{{research}}} {{/if}} {{/each}} {{else}} {{#each group}} {{{anchor}}} {{#each people}} {{"title"}} {{"firstname"}} {{"surname"}} {{#if phone}}{{phone}}{{else}}{{/if}} {{#if job}}{{job}}{{else}}{{job_title}}{{/if}} {{/each}} {{/each}} {{/if}} {{#with pager}} {{#each before}} {{this}}{{/each}} {{current}} (current){{#each after}} {{this}}{{/each}} Single Page {{/with}} profiles/students/ xalphabet=10 {{#each group}} {{title}} {{#each people}} {{#if @first}} {{fullname}} {{else}} {{fullname}} {{/if}} {{#if @last}} {{/if}} {{/each}} {{/each}} This article was published on 27 Nov, 2020 Banner Image courtesy of Magnus Hagdorn Contact us The University of Edinburgh Terms & conditions Privacy & cookies Complaints procedure Modern slavery Website accessibility Freedom of information publication scheme Data Protection MyEd login The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336, VAT Registration Number GB 592 9507 00, and is acknowledged by the UK authorities as a “Recognised body” which has been granted degree awarding powers. Unless explicitly stated otherwise, all material is copyright © The University of Edinburgh 2021. CMS login 
work_6dybqb64nrdshp7ppfsruknp5y	----	PII: 0743-1066(89)90034-4 J. LOGIC PROGRAMMING 1989:163-178 163 METAINTERPRETERS FOR EXPERT SYSTEM CONSTRUCTION LEON STERLING AND RANDALL D. BEER D We discuss the use of metainterpreters for building expert systems in PROLOG. Three issues are covered. The first is a technique for mixing a metainterpreter into an object program imbuing it with the functionality specified by the metainterpreter. Mixing a metainterpreter into a PROLOG program consists of two steps: partially evaluating the metainterpreter with respect to the object program and pushing down the metaarguments of the metainterpreter into the object program. The second issue is a classification of metainterpreters into structural, contextual, and behavioral enhance- ments. Examples are given of useful enhancements for building expert systems. Finally, we discuss the combination of several metainterpreters and how a programming environment for building expert systems could be built based on the ideas in this paper. a 1. INTRODUCTION Expert systems have been commonly decomposed into a knowledge base and inference engine. This decomposition is not entirely appropriate for expert systems written in PROLOG. Much of what an inference engine does is provided by PROLOG itself; knowledge bases are executable. However, simple knowledge bases written in PROLOG are generally insufficient as expert systems. Three limitations of PROLOG for building expert systems are commonly cited. There is no explanation capability, no mechanism for reasoning with uncertainty, and an inflexible control regime. This paper discusses how metainterpreters can be used to overcome these limitations. Clark and McCabe [7] suggested adding features such as explanations or uncer- tainty reasoning by modifying the PROLOG program constituting the knowledge Address correspondence to Professor L. Sterling, Department of Computer Science, Case Institute of Technology, Case Western Reserve University, Cleveland, Ohio 44106. Received 10 July 1987; accepted 11 October 1987. THE JOURNAL OF LOGIC PROGRAMMING QElsevier Science Publishing Co., Inc., 1989 655 Avenue of the Americas, New York, NY 10010 0743-1066/89/$3.50 164 LEONSTERLINGANDRANDALLD.BEER base. Extra arguments are added to each predicate to maintain the necessary information-the uncertainty, for example, or a representation of the proof tree. This approach retains the efficiency and structure of the knowledge base, but sacrifices modularity and clarity. The entire knowledge base must be altered to incorporate a new feature. Furthermore, a purely declarative program is altered to include procedural concerns such as an uncertainty calculus. Blurring of tasks in this way is generally confusing. An alternative approach for adding extra features is based on metainterpreters [24]. The knowledge base of the expert system is kept intact, and a metainterpreter is written to specify how to associate the feature with the knowledge base. Meta interpreters can also be used to express alternate control. Solving a query with a metainterpreter rather than PROLOG is known in the folklore to cost an order of magnitude. Some confirming statistics are presented in [28] and [25]. One claim of this paper is that partial evaluation can be realistically used to solve this problem. The partial evaluation of metainterpreters has been discussed by Gallagher [9] and Takeuchi and Furukawa [28]. It is a program transformation technique which can be used to specialize a metainterpreter to a given object program in such a way that the extra level of interpretation is removed while the functionality is retained. Controlling partial evaluation is difficult in general. Here, by restricting our task to removing the level of intepretation introduced by the metainterpreter, we avoid many of these difficulties. An important pragmatic issue in the use of metainterpreters for building expert systems is their combination. For example, a user may wish to have both an explanation capability and an uncertainty calculated for her program where both capabilities are expressed as separate metainterpreters. Rather than writing an entirely new metainterpreter, however, she should have some way to simply inte- grate the functionality of the existing ones. In considering the ways in which various metainterpreters can be combined, we have found a classification of their enhance- ments to an object program into structural, contextual, and behavioral to be indispensible. We advocate a particular methodology for building expert systems in PROLOG. The domain knowledge is cleanly separated from its use. The functionality of the system is incrementally developed by selecting from a library of existing metainter- preters, tailoring and combining them as required. Finally, the knowledge base is imbued with the functionality of the combined metainterpreter via partial evalua- tion, resulting in an efficient object level program. This style of development, where the programmer assembles a system from modular pieces, is reminiscent of the use of inheritance in the object-oriented programming paradigm, especially the multiple inheritance of FLAVORS [6,19,30]. In FLAVORS, the functionality of some base flavor may be enhanced by selecting from a set of potential mix-ins. Consequently, we draw on the concepts and terminology of object-oriented programming in describing our work. In this paper we concentrate on metainterpreters for PROLOG. The discussion, however, could be applied easily to metainterpreters for other logic programming languages such as FCP [17]. Most comments referring to PROLOG would equally apply to other languages. This research continues the outline given in [24]. It is an application of general work in the logic programming community on metainterpreters [8,11,18,21] and METAINTERPRETERS FOR SYSTEM CONSTRUCTION 165 partial evaluation [3,13,14,28,20] to expert system construction. Implicit in our research, but not addressed in this paper, are many interesting metalinguistic issues such as those raised in [4] and [29]. An overview of the paper is as follows. The next section introduces the notion of metainterpreters for PROLOG. Sections 3 and 4 discuss program transformation techniques which remove the extra level of interpretation necessitated by a metain- terpreter. We focus particularly on the use of partial evaluation and demonstrate how the control of a general partial evaluator may be simplified in the context of removing a level of interpretation. Section 5 presents a classification of the possible ways in which a metainterpreter can enhance an object program. Section 6 considers the problem of combining the functionality of several metainterpreters into a single one and presents a program transformation technique which accomplishes the combination for two of the three classes of enhancements considered in Section 5. Finally, we conclude with a discussion of directions for future research. 2. METAINTERPRETERS AND FLAVORS A metainterpreter for a language, sometimes called a metacircular interpreter [l], is an interpreter for a language written in that language. It is a special case of a metalevel interpreter where the object language and the metalanguage have been amalgamated. For a discussion of amalgamation in logic programming see [4]. Many different metainterpreters can be written for a given language. An impor- tant characteristic of a metainterpreter is the level of computational detail it makes accessible. This characteristic is called granularity. The granularity is course if there is little access, and jine if it is detailed. Throughout the paper the predicate solve is used to denote a metainterpreter. The arity of solve changes, however, depending on how the metainterpreter has been enhanced. Conventions of Edinburgh PROLOG as laid out in [27] are used. The simplest metainterpreter calls PROLOG directly. It is defined as solve(Goal) :- Goal. It corresponds to amalgamating object languages and metalanguages by making them identical. The granularity of this metainterpreter is very coarse, allowing little scope for enhancements other than alternative top level shells, as described in [27]. A far more useful metainterpreter is at the clause reduction level. The three clause metainterpreter for pure PROLOG given below is well known. It makes explicit the choice of clause being used to reduce a goal, and the choice of literal to reduce in the resolvent. Unification and backtracking are handled implicitly, relying upon the behavior of PROLOG: solve( true). solve((A,B)) :- solve(A),solve(B). solve(A) :- clause(A,B), solve(B). Metainterpreters can be written at finer levels of granularity than the clause reduction level. For example, a metainterpreter can model backtracking by keeping explicit stacks, or can perform unification. In our experience, the metainterpreter at the clause reduction level has the granularity most suited for building expert systems. 166 LEONSTERLINGANDRANDALLD.BEER Our underlying motivation for discussing metainterpreters is how they can be used to build complex programs for applications such as expert systems and program development environments. Many uses of inheritance in the object-ori- ented programming paradigm are essentially concerned with this issue [23]. Inheri- tance allows a new object to inherit features of an existing one, supporting reusable software components which minimize redundancy and encourage modularity. Multi- ple inheritance is an especially powerful form of inheritance in which a new object inherits features from several other objects. It has been most popularized by the FLAVORS system, an object-oriented extension to LISP found on many LISP machines [6,19,30]. Flavors allows the functionality of some base flavor to be modified or enhanced by combining it with mix-in flavors selected from a library of potential augmentations. We claim that this style of development is also desirable for expert systems. Where possible, a programmer should be able to pick and choose from a library of stereotypic expert system features and integrate them into his knowledge base. Consequently, we draw heavily on the concepts and terminology of FLAVORS, speaking of “flavored” metainterpreters and of “mixing” a number of metainter- preters into a knowledge base. Rather than intending this as simply a verbal analogy, however, we believe that our use of metainterpreters shares many of the fundamental concerns of inheritance in object-oriented programming. Such basic questions as how one decomposes a given domain into a number of modular components, how one expresses these components so that a given subset of them may be integrated, and how one automates this integration are very much at the heart of our work. It should be noted that our use of the concepts of object-oriented programming emphasizes different issues than previous work on the integration of logic program- ming and object-oriented programming. Such work has tended to emphasize the modeling of message passing and state [15,10,22,31], while we have focused on the uses of inheritance to modularly develop complex software systems. A flavor in the context of logic programming is a metainterpreter. Associated with each flavor is the relation it computes, called its metagoal, and the program used to compute it, called its theory. Flavors are represented accordingly as a ternary relation, Jlauor(Nume, Metu Goal, Theory). For reasons that are probably obvious the above metainterpreter is the vanilla flavor. Its metagoal is solue(Goul), and its theory is the three clauses. A flavor is enhanced to have greater functionality or extended performance. For example, the prooftree flavor shown in Figure 1 builds a proof tree of the proof of an object goal. Throughout the remainder of this paper, we will make reference in our examples to an expert system for evaluating credit requests to a bank [2]. Two clauses from this expert system are shown in Figure 2. flavor(prooftree, solve(Goal,Proof), ._ [solve(true,nil), (sclve((A,B),(ProofA =d ProofB)) :- FIGURE I. hhnnrd flavor solve(A,ProofA), solve(B,ProcfB)), tree. (solve(A,(A if Pro&B)) :- clause(A,B), solve(B,ProofB))]). for hddina a nroof METAINTERPRETERS FOR SYSTEM CONSTRUCTION 167 credit(Client,Answer) :- ok_profile(Client), collateral_rating(Cent,CRating), financial_rating(CIient,FRating), bank_yield(CIient,Yield), FIGURE 2. Credit evaluation expert evaluate(profile(CRating,FRating,Yield),Answer). financial_rating(Client,FRating) :- system fragment. financial_factors(Factors), score(Factors,Client,O,Score), calibrate(Score,FRating). 3. THE PARTIAL EVALUATION OF METAINTERPRETERS 3. I. Introduction The use of partial evaluation to remove the extra level of interpretation necessitated by a metainterpreter was first discussed by Gallagher [9] and developed by Take&i and Furukawa [28]. Conceptually, the partial evaluation of a PROLOG program is simple. A more general program is specialized for a specific use based upon partial information concerning that use (such as partial instantiation of some of the arguments). A metainterpreter may be specialized to a given object program by combining the metainterpreter with the object program and partially evaluating the combined program with respect to a given object goal. The program transformations which are the basis of partial evaluation use existing techniques. Komorowski [16] describes partial evaluation in PROLOG as consisting of three methods: pruning, forward data structure propagation, and backward data structure propagation. Pruning, the simplest, is essentially only choosing the clauses of a program actually used in a computation, Forward data structure propagation consists in propagating any partial argument instantiations of a goal to its subgoals. Backward data structure propagation involves replacing any calls to a deterministic goal with the body of its associated clause. Our partial evaluator is similar to that of Takeuchi and Furukawa [28] in its use of the techniques of pruning, forward data structure propagation, and selective unfolding. The last is essentially a generalization of backward data structure propagation in which the requirement of determinism is dropped and a clause containing a goal to be unfolded is duplicated for each clause unifying with that goal. The result of partially evaluating the prooftree flavor and the credit evaluation system with respect to the goal credit(Client,Answer) is shown in Figure 3. 3.2. Controlling Partial Evaluation A general partial evaluator can be difficult to control, particularly with respect to the decision whether or not to unfold a given goal. In general, our partial evaluator takes advantage of the constraints imposed by the context of its intended use to simplify its control. The primary goal is the removal of the extra level of interpreta- tion necessitated by a metainterpreter. Rather than giving the user the ability to 168 LEON STERLING AND RANDALL D. BEER solve(credit(Client,Answer), credit(Client,Answer) if (Proof1 and Proof2 and Proof3 and Proof4 and Proofs)) :- solve(ok_profile(Client),Proofl), solve(collateral_ratiing(Client, CRating),Proof2), solve(financial_rating(Client,FRating),Proof3), solve(bti_yield(Client,Yield),Proof4), solve(evaluate@rofile(CRating,FRating,Yield),Answer),ProofS). solve(financial_rating(Client,FRating), financial_rating(Client,FRating) if (Proof1 and Proof2 and ProoD)) :- solve(financial_factors(Factors),Proofl), solve(score(Factors,Client,O,Score),Proof2), solve(calibrate(Score,FRating),ProoD). FIGURE 3. Partially evaluating the credit system and the firooftree flavor. control the partial evaluator with explicit declarations, we prefer to give a set of general rules for removing the extra level of -interpretation from the combination of object program and metainterpreter. The top level relation for the partial evaluation is peual(ObjGoa1, Program, NewProgram). This takes Program, which is the concatenation of the object program and the metainterpreter, and partially evaluates it with respect to ObjGoal. A new program is produced which is the specialized metainterpreter with the extra level of interpretation removed, but the metaargument structure in place. A compat- ibility issue immediately arises between the representation of object clauses and the format expected by a metainterpreter. We represent object clauses as clause/:! relations to be consistent with their representation in the database. The first, and most important, control decision is whether or not to unfold a given goal. One extreme would be to unfold every goal, essentially converting the program to a set of ground facts. However, this is an exponential process and is usually not desirable for very large programs. Our partial evaluator provides a declaration, shou~d_unfoZd/l, which specifies that a given goal should be unfolded. However, rather than intending that this declaration be provided by the user for each partial evaluation, we have provided a set of them which seems to be sufficient for removing the extra level of interpretation from the metainterpreters which we have so far considered. For example, metainterpreter calls with conjunctive object goals are unfolded. Figure 4 shows the should_unjold/l declarations we have found sufficient to date. A second control decision concerns the handling of infinite loops. A computation which terminates at run time can still cause an infinite loop during partial evalua- tion if an insuthcient number of arguments are instantiated. Our system maintains a stack of pending goals, and detects an infinite loop whenever a goal to be partially evaluated is an instance of a goal on the stack. The detection of an infinite loop terminates partial evaluation and returns the goal itself. Another important control issue is that of the evaluability of goals. The notion of an evaluable goal is useful primarily for simplifying a program resulting from partial METAINTERPRETERS FOR SYSTEM CONSTRUCTION 169 should_unfold(Goal) :- functor(Goal,solve,N), arg(l,Goal,(A,B)). should_unfold(Goal) :- functor(Go&olve,N), arg(l,Goal,true) should_unfold(Goal) :- FIGURE 4. Should-unfold declara- tions. functor(Goal,solve,N), arg(l,Goal,A), system(A). should_unfold(Goal) :- system(Goal). evaluation. For example, if enough arguments are instantiated in a call to append/3, then that call may be performed during partial evaluation rather than at nmtime. However, it is an open research issue to determine whether or not an arbitrary goal is sufficiently instantiated to be completely evaluated. Further, while such evalua- tions can produce more efficient and aesthetic programs, they are not strictly necessary for removing the extra level of metainterpretation. Because the goal evaluability issue is primarily a matter of aesthetics, we provide no general mechanism for declaring a given goal to be evaluable. However, our partial evaluator handles the evaluability of most system goals. evaluating for example X is 0 + 1 and clause( solue( credit( Client, Answer), Tree), Body) without any explicit declarations from the user. A final control issue to be addressed by a general partial evaluator is that of open programs. A program is open if some of the goals are left intentionally undefined at partial evaluation time. This can be useful, for example, if data for an expert system are to be provided at a later time. We handle open programs by assuming that a goal which fails during partial evaluation time will also fail at run time, that is, we assume that a system to be partially evaluated is closed. This assumption is motivated by the fact that the partial evaluator is intended to be used to assemble a final, efficient version of a debugged system. However, we are currently examining ways to allow restricted kinds of open programs without requiring an explicit declaration. To conclude this section, a comparison is given between our partial evaluator and the more general partial evaluator discussed by Takeuchi and Furukawa in [28]. Their approach on each of the four control issues is to provide explicit hooks in the form of user-specified declarations. We look at the nature of the declarations used for each control issue. In [28], the default behavior is that every goal is unfolded unless a declaration, inhibit_unfoIding(Goal), is present. The onus is on the user to know which goals may cause trouble. Our basic assumption is the opposite. No goal is unfolded unless there is a declaration should_unfold(Goal), and declarations sufficient to remove a level of interpretation are provided in the system as given in Figure 4. Infinite loops are detected in [28] by maintaining a stack of pending goals and checking whether a goal being evaluated is an instance of a goal on this stack. Unlike our approach, there is the possibility to nonetheless continue partial evalua- tion. This is specified by the user with a declaration expand-loop (Goal )_ Being able to recursively expand goals is irrelevant for our system, since we are only unfolding one level of interpretation. 170 LEON STERLING AND RANDALL D. BEER Another type of declaration is used in the partial evaluator of 1281 to specify evaluability of goals, namely facts of the form type(GoaZ,e). A type fact is also used to handle open goals. The declaration type( Goal, t) terminates the partial evaluation of Goal. In our system, there is no classification of goals into types according to how they should be handled by a partial evaluator. Allowing declarations provides greater flexibility, but also demands more of the user. Our system demonstrates that the necessary declarations can be specified ahead of time, in the context of removing one level of interpretation. The expert system builder need not be concerned with details of the partial evaluator. 4. PUSHING DOWN METAARGUMENTS A metainterpreter can have two distinct effects upon a given object program. It can affect the structure of the proof for, and indeed the eventual success of, a given object goal, and it can add additional arguments to the proof. We consider each in turn with respect to what needs to be done to mix a flavor into an object program. The prooftree flavor of Figure 1 faithfully reflects PROLOG’s standard proof structure. However, more complex proof structures are possible, such as a planner which dynamically reorders conjunctive goals. Other metainterpreters affect the provability of certain goals. The askable flavor, based on the relation askable which will be introduced in the next section, for example, prompts the user for a solution to a goal that would fail in the standard PROLOG model of computation. The effects of metainterpreters on the structure of the proof tree for a given object are made explicit by the process of partial evaluation. The other effect that a metainterpreter can have on an object program is computing extra arguments while solving a given goal. Such arguments are called metaarguments. Two flavors which compute a single metaargument are prooftree, whose metaargument is the proof tree of the goal being solved, and count, whose metaargument is the number of reductions used in solving a goal. The argument- computing effects of a metainterpreter can be made explicit by “pushing down” metaarguments into the object program. The top-level relation which we use for pushing down the extra arguments in a goal is pu.sh_down_meta_args( MetaGoal, MetaProg, ObjGoal, ObjProg). Input for push_down_meta_args is a metaprogram such as Figure 3 and its associated metagoal, in this case solue(credit( Client, Answer), Tree). The output is an object program and object goal which no longer contain any explicit mention of the me&interpreter but which have been augmented with any additional arguments that the flavor com- puted. For simplicity, our system assumes that the first argument to every metainter- preter is always the object goal. The result of pushing down the metaargument computed by prooftree into the credit evaluation expert system fragment is shown in Figure 5. Partial evaluation followed by the pushing down of metaarguments makes entirely explicit both effects of a given metainterpreter on an object program. The result of this process is simply another object program, augmented by the function- ality of the flavor. METAINTERPRETERS FOR SYSTEM CONSTRUCTION 171 credit(Client,Answer, credit(Client,Answer) if (Proofl and Proof2 and Proof3 and Proof4 and Proofs)) :- ok_profile(CIient,Proofl), colIateral_rating(CIient, CRating, Proof2), tinancial_rating(CIient,FRating,Proof3), bank_geld(CIient,Yield,Proof4), evaluate(profile(CRating,FRating,Yield),Answer,ProofS). FIGURE 5. Expert system fragment with prooftree mixed in. financial_rating(CIient,FRating, fmancial_rating(Client,FRating) if (Proofl and Proof2 and Proof3)) :- financial_factors(Factors,Proofl), score(Factors,CIient,O,Score,Proof2), caIibrate(Score,FRating,Proof3). 5. CLASSIFYING METAINTERPRETER ENHANCEMENTS This section concentrates on how to enhance flavors, building complex flavors from simple ones. The examples are presented to demonstrate the ease of enhancing flavors in logic programming. More examples can be found in [27], with particular emphasis on the application of enhanced flavors for building expert system shells and program debuggers. There are two ways of enhancing a flavor. Firstly arguments can be added to the solve predicate, and secondly its behavior can be changed through the addition, modification, or deletion of clauses or goals within a clause. These two forms of enhancements correspond to the two effects a metainterpreter can have on a given object program: adding extra arguments to the proof of a given goal and affecting the structure of the proof itself. Enhancements due to additional arguments are further of two kinds. The extra arguments can be used as a structure to be computed while solving a goal, or can be used as a context representing information needed during the computation. We consider each of these three forms of enhancement. DeJnition. A flavor solve, is a structural enhancement of a flavor solve, if (1) each argument of solve, corresponds to an argument of solveM, (2) corresponding arguments behave identically on identical inputs, and (3) arguments of solve, not corresponding to arguments of solve, compute a structure as the goal is being solved. To facilitate specifying the correspondence, arguments in enhanced flavors are in the same order as the simpler flavor. Additional arguments appear last. This will be true also for contextual enhancements to be introduced below. The prooftree flavor shown in Figure 1 is a typical structural enhancement. A second example is the count flavor in Figure 6. Its metagoal so/ve(Goal, N) is true if Goal requires N reductions to be solved. It is clear that the behavior of count in 172 LEON STERLING AND RANDALL D. BEER solve(true,O) solve((A,B),N) :- solve(A,Nl), solve(B,N2), N is Nl + N2. FIGURE f ,.- ~~-_ solve(A,N) :- counting gc 0. xrucum.uy ennancea navor Ior ml reductions. clause(A,B), solve(B,Nl), N is Nl + 1. solving a goal is identical to uanillu. The structure computed is the number of goal reductions. In general, structural enhancements can introduce arbitrary PROLOG code to maintain the extra arguments, for example the arithmetic additions in Figure 6. Structural enhancements are normally used to instantiate their extra arguments. The count flavor, for example, would instantiate its second argument to the number of goal reductions used in solving a particular goal. Note that the base fact of counr, soZue(true,O), is ground. This is a typical feature of a structural enhancement. The general metalevel predicate demo(Goa1, Proof) described by Bowen and Kowalski in [4], where Proof is the proof of Goal, is a structural enhancement of the metalevel interpreter demo. In fact, all additional arguments computed by structural enhancements, such as the number of goal reductions, can be regarded as “alterna- tive proofs.” Dejnition. A flavor solve, is a contextual enhancement of a flavor solve, if (1) each argument of solve, corresponds to an argument of solve,, (2) corresponding arguments behave identically on identical inputs, and (3) arguments of solve, not corresponding to arguments of solue, carry a context as the goal is being solved. A typical contextual enhancement of the vanilla flavor is given in Figure 7. The flavor is called brunch, and its metagoal solve( GouZ,Context ) is true if Goal is solved; that is, Context has no declarative relationship with Goal. Its intended use is to convey the search tree further in the program. All the flavor does is insist that the associated search trees are logically consistent from one goal to another. The branch flavor can form the basis of a why not explanation facility [26], for example. The additional arguments for contextual enhancements must be instantiated to an initial context; otherwise an error is likely. The initial context for the branch flavor is the empty list. If the inital context is not instantiated, brunch will construct an incomplete list. The third class of enhancements are called behavioral enhancements. A behav- iorally enhanced flavor extends the computation performed by the flavor being solve(true,Context). solve((A,B),Context) :- solve(A,Context), solve(B,Context). solve(A,Context) :- FIGURE 7. Contextually enhanced flavor for carrying the proof tree branch. clause(A,B), solve (B,[A if B]Context]). METAINTJZRPRETERS FOR SYSTEM CONSTRUCTION 173 solve( true). solve((A,B)) :- solve(A), solve(B). solve(A) :- clause(A,B), solve(B). FIGURE 8. Behaviorally enhanced flavor for querying solve(A) :- system(A), A. the user. solve(A) :- askable( ask(A). enhanced. The metagoal of a behaviorally enhanced flavor remains the same. Typically, extra clauses are included in the theory. Definition. A flavor solve, is a behavioral enhancement of a flavor solve,,, if they produce different proofs for identical object goals. A metainterpreter can poten- tially be behaviorally enhanced by the addition, modification, or deletion of clauses or goals within a clause. Any modification to a metainterpreter which is not a structural or contextual enhancement is a behavioral enhancement. Behav- ioral modification by side effects, for example tracing a computation on the screen, we regard as a behavioral enhancement. Figure 8 shows an example of a behavioral enhancement of vanilla. The metagoal of askable is the same as that of vanilla. Its theory, however, contains two additional clauses which handle system goals and ask the user if a given goal is true if it can be proven in no other way. Behavioral enhancement can also arise from the addition or modification of goals within a clause. In general, a behavioral enhance- ment is any change to a metainterpreter which affects its proof of an object goal. The forms of enhancement are by no means exclusive. Flavors can be enhanced in several ways at once. For example, in the explanation shell for expert systems described in [26], the metainterpreter for the why not component exhibits all three enhancements. The argument constituting the contextual enhancement is a branch of the proof tree, the argument constituting the structural enhancement returns the proof tree, and the behavioral enhancement adds extra clauses so that the metainter- preter always succeeds even if PROLOG fails. 6. COMBINING METAINTERF’RETERS An important underlying principle of our methodology for building expert systems, and complex programs more generally, is that they should be constructed from simple building blocks. The individual units must be easy to combine and use. If metainterpreters are to be useful as modular components of expert systems, it must be easy to mix several metainterpreters and incorporate their joint effect in an object program. This section discusses how to combine the effect of different metainterpreters when solving a given object goal. Two strategies for combination, nesting and separative combination, are presented. For each, a metalevel goal or program is given which specifies what is computed. Further strengths and weaknesses of each are discussed. We suggest there are two major issues relevant for the combination of metainter- preters. The first is whether an efficient program can be derived which incorporates 174 LEON STERLING AND RANDALL D. BEER the effects of the different flavors. More particularly in this paper, we consider whether the mixing technique of Sections 3 and 4 is applicable. The second issue is commutativity, that is, whether the combination strategy is sensitive to the order in which the metainterpreters are given. A minor issue is handling name clashes. In this paper, all metainterpreters are referred to as solve, which would cause problems in most current PROLOG implementations. Each metainterpreter could be renamed, or a good module system used. The best solution is to use an extension of PROLOG such as metaPROLOG [5], where theories are treated as genuine objects and can be referred to and manipulated. Each metainterpreter would be a theory. In this section all metainterpreters are assumed to be structural or contextual enhancements of a common metainterpreter, unless explicitly stated otherwise. Each enhancement has a single extra argument. For example, prooftree, branch, and count are all of this form. Restricting to enhancements with only one extra argument is in fact no loss of generality. Several extra arguments can be grouped into a single structured argument with a technique similar to the pushing down of metaargu- ments discussed in Section 4. Combining behavioral enhancements is not discussed in this paper. Different ways of modifying control using behavioral enhancements are not obviously conso- nant. It is a research issue to find a useful way of specifying behavioral enhance- ments to allow them to be combined commutatively. The first strategy for combining metainterpreters is called nesting. Suppose solve, and solve, are two metainterpreters. The effect of both solve, and solve, in solving a given object level query Goal is captured with the metalevel query solve,( solve,( Goal, Argl), Arg2). It is straightforward to mix two metainterpreters into an object program to achieve the effect of a given nested query. The first metainterpreter is mixed into the object program, and the resultant program is the new object program into which the second metainterpreter is mixed. The process can be iterated to mix in several metainterpreters. A discussion of this approach is given in [25] together with some statistics on the speedup of solving the final object program in contrast with going through levels of interpretation. The nesting strategy is not commutative, however. Consider mixing both prooftree and branch into an object program. If prooftree is mixed in after branch, the argument constituting the proof tree will contain references to the context that has been added, which would not be there if prooftree had been mixed in before branch. The problem is more noticeable with flavors with extra goals such as count. If count and prooftree are mixed, the proof tree may or may not include the arithmetic calculation, depending on the order of mixing in the metainterpreters. The reason for the lack of commutativity of the nesting strategy is the loss of distinction between object level and metalevel entities. Goals in the unenhanced metainterpreter, called basic goals, and goals introduced to manipulate the extra structure are indistinguishable once a metainterpreter has been mixed into an object program. The partial evaluator cannot distinguish easily between object level and metalevel goals. Although this may be overcome with a more sophisticated partial evaluator taking account of appropriate metaknowledge denoting basic goals, it is preferable to find a simpler form of combination which is commutative. METAINTERPRETERS FOR SYSTEM CONSTRUCTION 175 solve(Goal,[Flavor]Flavors],[Result [Results]) :- flavor(Havor,MetaGoa.l,Theory), amalgarnated(Goal,MetaGoal,Result), solve_flavor(Theory,MetaGoal), FIGURE 9. Solving goals separatively. solve(Goal,Flavors,Results). solve(GoaU I,[ I). The second strategy is called separative combination. The extra argument of each metainterpreter is computed separately in the same context. Figure 9 contains a definition for a predicate solve(Goa1, Flavors, Results), where Goal is the object goal, Flavors is a list of flavors to be mixed, and Results is a list of arguments computed by each flavor. The goal is solved sequentially for each flavor. There are two points of interest in the code. Firstly, if the goal being solved is not ground initially, but is instantiated by the first flavor, then later flavors will solve the instantiated goal. This guarantees consistency. Secondly, there is no commitment to the representation of the flavor or the form of the result in the code. The predicate amalgamated/3 is responsible for respecting representational details. More importantly, it communicates information between the object and metalevel goals. An appropriate definition is amalgamated(Goal,solve(Goal,Result) ,Result) . The order of the flavors is unimportant for the separative strategy. They com- mute with respect to the program in Figure 9. In other words, the argument corresponding to each flavor will be the same no matter in which order it appears. The code for solue/3 solves the object goal as many times as there are flavors. It is preferable to solve a given object goal only once. This can be achieved by generating a metainterpreter which combines the effects of the flavors. The top level code for combining metainterpreters is given in Figure 10. The basic relation is combine( Flavors, NewFlavor), where Flavors is a list of flavor names and NewFlavor combine([~avorIFlavors],flavor(comb,solve(Goal,Args),Meta )) :- flavor(Flavor,_,_), basic_Aavor(Flavor,Vanilla), flavor(Vanilla,_,VauillaTheory), combine([FlavorlFlavors],Args,VanillaTheory,Meta ). combine([Flavor(Flavors],[Arg@rgs],InMeta,Meta ) :- flavor(Flavor,solve(Goal,Arg),Theory), combine_theories(Theory,In Meta,OutMeta), combine(Flavors,Args,OutMeta,Meta ). combine([ I,[ ],Meta,Meta ) :- complete(Meta ). FIGURE 10. Combining flavors separatively. 176 LEON STERLING AND RANDALL D. BEER is a new flavor which represents the combination of all of them. The metagoal of the combined flavor is solue(Goal,Args), where Args is the list of arguments which are the enhancements of the flavors being combined. The code for combine theories is technical but straightforward and is omitted here. It is an example, similar to the pushing down of metaarguments, of a problem that has been sufficiently constrained that syntactic transformations suffice to solve it. The last clause in Figure 10 completes the incomplete structures generated by combine_theories. 7. CONCLUSIONS AND DIRECTIONS FOR FUTURE RESEARCH The paper has discussed how different metainterpreters can be combined and mixed into a knowledge base. A classification of metainterpreters has been developed to specify the combination. A basic metainterpreter of a given granularity is identified. Three types of enhancements of the basic metainterpreter can then be described: structural enhancements, contextual enhancements, and behavioral enhancements. A strategy for combining structural and contextual enhancements called separa- tive combination is described. It is best implemented as follows. Given a collection of metainterpreters, combine them to produce a new me&interpreter, and then mix the new metainterpreter into the knowledge base. The mixing in is performed by a restricted form of partial evaluation augmented by an additional step of pushing down metaarguments. Future work is possible for the three areas of classification, combining metainter- preters and mixing them into the knowledge base. The classification can be sharp- ened with respect to behavioral enhancements. Hopefully a clearer understanding of the types of behavioral enhancements will enable them to be freely combined with structural and contextual enhancements. For the partial evaluation, the assumptions about expert systems that are used could be more clearly stated, and it could be shown how they affect the partial evaluator. The example flavors have all been small. We feel, nonetheless, that they are representative of what will be needed to develop large applications. PROLOG programs are concise. A large application should still be developed as concrete proof of the utility of these ideas. Finally, the relationship to LISP flavors needs to be explored further. Both the separative and nesting strategies of mixing flavors treat the flavors as independent. It is both more interesting and more general purpose to allow interactions between flavors. Indeed, the bulk of work on flavors in LISP is in the development of a language for expressing interaction. The examples in logic programming that we have seen have not needed interaction. The appropriate language or form for expressing interactions for logic programming flavors thus remains essentially academic, but is a topic for research currently under investigation. This work was partially supported by a core research grant from the Cleveland Advanced Manufacturing Program and the State of Ohio. Anm Lakhotia helped with running some test examples. METAINTERPRETERS FOR SYSTEM CONSTRUCTION 177 REFERENCES 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. Abelson, H. and Sussman, G. J., The Structure and Interpretation of Computer Programs, MIT Press, 1985. Ben David, A. and Sterling, L. S., A Prototype Expert System for Credit Evaluation, in: L. F. Paul (ed.), Artificial Intelligence in Economics and Management, Elsevier North- Holland, 1986, pp. 121-128. Bloch, C., Source-to-Source Transformations of Logic Programs, M.Sc. Thesis, CS84-22, Weizmann Inst. of Science, 1984. Bowen, K. and Kowalski, R., Amalgamating Language and Meta-language, in: Clark and Tarnlund (eds.), Logic Programming Academic, 1982, pp. 153-172. Bowen, K. and Weinberg, T., A meta-level extension of PROLOG, in: J. Cohen and J. Conery (eds.), Proceedings of the 1985 Symposium on Logic Programming, IEEE Com- puter Society Press, 1985, pp. 48-53. Cannon, H. I., Flavors: A Non-Hierarchical Approach to Object-Centered Programming, unpublished paper, 1982. Clark, K. L. and McCabe, F. G., PROLOG: A Language for Implementing Expert Systems, in: Hayes, Michie, and Pao (eds.), Machine Intefhgence IO Ellis-Hotwood, 1982, pp. 455-470. Dincbas, M., Meta control of logic programs in METALOG, in: Proceedings of FGCS ‘84, Tokyo, Nov. 1984, pp. 361-370. Gallagher, J., Transforming Logic Programs by Specializing Interpreters, in: The Proceed- ings of the Seventh European Conference on AI, Brighton Center, UK, 1986. Gallaire, H., Merging Objects and Logic Programming: Relational Semantics, in: Pro- ceedings AAAI-86, Philadelphia, 1986, pp. 754-758. Gallaire, H. and Lasserre, C., Metalevel Control for Logic Programs, in: Clark and Tamlund (eds.), Logic Programming, Academic, 1982, pp. 173-185. Hammond, P., micro-PROLOG for expert systems, in: micro-prolog: Programming in logic, Prentice-Hall International, 1984. Kahn, K. M., The Compilation of PROLOG Programs without the Use of a PROLOG Compiler, Tech. Report, UPMAIL, Uppsala Univ., Sweden, 1984. Kahn, K. M., Partial Evaluation, Programming Methodology, and Artificial Intelligence, AI Magazine, Spring, 1984, pp. 53-57. Kahn, K., Tribble, E. E., Miller, M. S., and Bobrow, D. G., Objects in Concurrent Logic Programming Languages, in: Proceedings of the Conference on Object-Oriented Program- ming Systems, Languages, and Applications, 1986, pp. 242-257, Komorowski, H. J., A Specification of an Abstract PROLOG Machine and Its Applica- tion to Partial Evaluation, Linkoping Studies in Science and Technology Dissertations, No. 69, Software Systems Research Center, Linkoping Univ., Sweden. Mierowsky, C., Taylor, S., Shapiro, E. Y., Levy, J., and Safra, S., The Design and Implementation of Flat Concurrent PROLOG, Tech. Report CS85-09, Dept. of Com- puter Science, Weizmann Inst. of Science, Rehovot, Israel, 1985. Pereira, L., Logic Control with Logic, in: Proceedings of the First International Logic Programming Conference, Marseille, 1982, pp. 9-18. Moon, D. A., Object-Oriented Programming with Flavors, in: Proceedings of the ACM Conference on Object-Oriented Programming Systems, Languages, and Applications, Port- land, Ore., 1986, pp. l-8. Shapiro, E. Y. and Safra, M., Meta-interpreters for real, in: Proceedings of IFIP-86, 1986. Shapiro, E. Y., Unpublished lecture notes on Concurrent PROLOG, 1985. Shapiro, E. Y. and Takeuchi, A., Object Oriented Programming in Concurrent PROLOG, New Generation Comput. 1~25-48 (1983). LEON STERLING AND RANDALL D. BEER 23. 24. 25. 26. 27. 28. 29. 30. 31. Stefik, M. and Bobrow, D. G., Object-Oriented Programming: Themes and Variations, AI Magazine 6(4):40-62 (Winter 1986). Sterling, L. S., Meta interpreters for Expert Systems, CAISR TR 134-85, Case Western Reserve Univ., 1985. Sterling, L. S. and Beer, R. D., Incremental Flavor-Mixing of Meta-interpreters for Expert System Construction, in: Proceedings of the Third Logic Programming Symposium, Salt Lake City, 1986, pp. 20-27. Sterling, L. S. and Lalee, M., An Explanation Shell for Expert Systems, Comput. Intelligence (1986). Sterling, L. S. and Shapiro, E. Y., The Art of Prolog, MIT Press, 1986. Takeuchi, A. and Furukawa, K., Partial Evaluation of PROLOG programs and its Application to Metaprogramming, ICOT Tech. Report, 1985. Weyhrauch, R. W., Prolegomena to a Theory of Mechanized Formal Reasoning, Artifi- cial Intelligence 13:133-170 (1980). Weinreb, D. and Moon, D., Flavors: Message Passing in the Lisp Machine, MIT-AI Memo No. 602, 1980. Zaniolo, C., Object-Oriented Programming in PROLOG, in: Proceedings 1984 Znterna- tional Symposium on Logic Programming, Atlantic City, 1984, pp. 265-270. 
work_6elfbs2jczduxf3rj3hx67c7be	----	Converting a Rule-based Expert System into a Belief Network∗ M. Korver & P.J.F. Lucas Department of Computer Science, Utrecht University P.O. Box 80.089 3508 TB Utrecht, The Netherlands e-mail: lucas@cs.uu.nl Abstract The theory of belief networks offers a relatively new approach for dealing with uncertain information in knowledge-based (expert) systems. In contrast with the heuristic tech- niques for reasoning with uncertainty employed in many rule-based expert systems, the theory of belief networks is mathematically sound, based on techniques from probability theory. It therefore seems attractive to convert existing rule-based expert systems into belief networks. In this article, we discuss the design of a belief network reformulation of the diagnostic rule-based expert system HEPAR. For the purpose of this experiment, we have studied several typical pieces of medical knowledge represented in the HEPAR system. It turned out that, due to the differences in the type of knowledge represented and in the formalism used to represent uncertainty, much of the medical knowledge re- quired for building the belief network concerned could not be extracted from HEPAR. As a consequence, significant additional knowledge acquisition was required. However, the objects and attributes defined in the HEPAR system, as well as the conditions in produc- tion rules mentioning these objects and attributes were useful for guiding the selection of the statistical variables for building the belief network. The mapping of objects and attributes in HEPAR to statistical variables is discussed in detail. Keywords & Phrases: medical expert systems, belief networks, causal graphs, decision support systems. 1 Introduction In heuristic, diagnostic expert systems, knowledge from a given domain is typically represented in the form of production rules, or rules for short. To express uncertainty in the domain, each conclusion of a rule is associated with a measure of confidence in its correctness. The exact meaning of such non-probabilistic measures of uncertainty usually is not clearly defined. An example of a method for handling uncertainty frequently applied in rule-based expert systems is the certainty-factor model developed by E.H. Shortliffe and B.G. Buchanan for the (E)MYCIN system [24, 1]. Certainty factors can be given a probabilistic interpretation, but ∗Published in: Medical Informatics, 18(3): 219–241, 1993 (also: 1994 Yearbook of Medical Informatics). 1 2 OVERVIEW OF THE HEPAR SYSTEM 2 it turns out that in doing so usually an inconsistent specification of a probability distribution results [5, 3]. HEPAR is such a rule-based expert system; it aims at supporting the clinician in the initial assessment of the patient with a disorder of the liver or biliary tract [14]. The HEPAR system has been designed in such a way that it is capable of generating diagnostic explanations, based on clinical knowledge, with a similar amount of detail as employed by the clinician. The system incorporates the certainty-factor model to deal with uncertain medical knowledge. Although its diagnostic performance has been shown to be quite reasonable [11, 13], one of the problems with the HEPAR system is that its conclusions may be difficult to interpret, due to the unclear meaning of certainty factors. In contrast with the heuristic models for reasoning with uncertainty, such as the certainty- factor model, the recent theory of belief networks is based on mathematically sound tech- niques, derived straight from probability theory [20]. A belief network permits the qualitative representation of dependencies and independencies among statistical variables. Since a causal relationship between two variables implies their statistical dependence, a belief network can be used to represent causal medical knowledge. The theory of belief networks then permits the use of this causal knowledge for diagnostic problem-solving. Such a causal model of a medical domain may be more easy to comprehend by medical students and unexperienced clinicians than a similar diagnostic rule-based system. Typically, in a rule-based expert system the heuristic knowledge encoded is too far removed from the underlying (patho)physiological pro- cesses to be understood by the non-expert physician. Furthermore, a belief network designed for diagnostic problem-solving can be applied to predict findings associated with (groups of) disorders. Although, in principle, the production rule formalism allows for the implementa- tion of causal reasoning models, such a rule-based system cannot be used for diagnosis in a straightforward way. The sound mathematical basis and the expressiveness of the formal- ism of belief networks make it attractive to consider converting an existing rule-based expert system into a belief network. In this article, we investigate the potentials of converting a diagnostic rule-based expert system into a belief network. This study is based on an actual experiment in which a repre- sentative part of the HEPAR system was converted into a belief network. Similar work has been carried out in the probabilistic reformulation of the INTERNIST-1/QMR knowledge base [25, 17]. However, the knowledge-representation and inference techniques used in this expert system differ considerably from those used in rule-based expert systems. We start with an introduction to the field of hepatology, and we will subsequently review the most im- portant characteristics of the HEPAR system. The steps taken in converting the rule-based representation formalism applied in the HEPAR system into a belief network representation are discussed in Section 3. We conclude by summarizing the potentials and limitations of our approach to knowledge-base reformulation. 2 Overview of the HEPAR system 2.1 Clinical diagnosis in hepatology Hepatology is a subspeciality of internal medicine which concerns the diagnosis and manage- ment of patients with disorders of the liver and biliary tract. Diagnosis of such disorders on purely clinical grounds is a difficult task, requiring much experience, as has been shown in several studies [15, 22, 4, 32, 33]. Clinicians with little experience in the field of hepatology 2 OVERVIEW OF THE HEPAR SYSTEM 3 have shown to produce a correct specific diagnosis in patients with jaundice in less than 45% of the cases [31]. The main problems in the diagnosis of disorders of the liver and biliary tract are [34]: • to distinguish between disorders of the liver (hepatocellular disorders) and of the biliary tract (biliary obstructive disorders), • to differentiate between acute and chronic disorders, and • to recognize whether a disorder is benign or malignant in nature. These different disease groups require different diagnostic procedures, treatment plans and prognostic assessments. Especially the distinction between biliary obstructive and hepato- cellular disorders is important, since the former typically require surgical treatment, whereas diseases from the latter group are managed conservatively. Both biliary obstructive and hep- atocellular disorders may give rise to an impaired secretion of bile into the bile ducts, called cholestasis. This manifests itself by jaundice through the accumulation in the serum of the biliary substance bilirubin. A large number of diagnostic methods is available to analyse patients with disorders of the liver and biliary tract. These include serological tests for several forms of viral hepatitis and autoimmune liver disease, and ultrasound imaging of the liver and biliary tract, which is of major importance in the diagnosis of biliary obstruction. In those cases where insufficient information is obtained from ultrasound, invasive techniques, such as percutaneous transhep- atic cholangiography (PTC) and endoscopic retrograde cholangiopancreatography (ERCP), can offer help. In addition, liver biopsy can be employed in order to determine the aetiology and severity of chronic liver disease. Of overriding importance in the diagnosis of hepatobiliary disease, however, is a detailed history and thorough physical examination. Together with a small number of routine labora- tory tests, they frequently suffice for the formulation of a differential diagnosis. Supplemen- tary diagnostic tests as mentioned above need then only be performed in selected cases, which minimizes not only expenses, but also diagnostic risk, since PTC, ERCP and liver biopsy are each associated with considerable morbidity and even a small mortality rate. 2.2 HEPAR, a rule-based system for the diagnosis of hepatobiliary disease As in many clinical specialities, in hepatology establishing an early working diagnosis is im- portant. Firstly, because treatment must often be initiated as soon as possible. Secondly, because it determines which diagnostic tests are selected for further assessment of the pa- tient. The application of simple data from the medical history, physical examination and laboratory tests has therefore been taken as the point of departure for the development of the HEPAR expert system. Using merely these data, the system is capable of providing an initial assessment of the patient with a disorder of the liver or biliary tract, determining: • whether the patient is suffering from an acute, subacute or chronic disorder, • whether the disorder is hepatocellular or biliary obstructive in nature, and • whether benign or malignant features are present. 2 OVERVIEW OF THE HEPAR SYSTEM 4 patient age sex complaints signs ... ... type disorder nature disorder diagnosis complab duration pain location character ... labresults PTT PTT-K ... u.s. liver size density ... radiology x-ray abdomen x-ray thorax scintiscan haematology ESR HB ... biochemistry γ-GT ASAT ... serology hepatitis A hepatitis B ... u.s. bileducts intrahepatic extrahepatic other u.s. ascites gallbladder ... complab = complaints, clinical signs, lab abnormalities u.s. = ultrasound Figure 1: The object tree of the HEPAR system. When additional information derived from some more specific, but generally available tests like ultrasound is included, the HEPAR system is also capable of producing a differential diagnosis consisting of a subset of possible diagnoses out of a set of some 80 disorders, ordered by the amount of evidence for each diagnosis. No use is made of data obtained by PTC, ERCP and liver biopsy. Thus closely following the clinical approach to the hepatological patient, the HEPAR system may be a suitable tool to assist the doctor in the initial assessment of the patient. The knowledge base of the HEPAR system consists of two separate components: • a collection of object definitions, structured as a tree, where each object definition describes an entity from the field of hepatology by listing a collection of its features, and • a rule base, containing the problem-solving knowledge concerning diagnosis in hepatol- ogy. The features of an object are usually called attributes. For example, in the HEPAR object tree, which is depicted in Figure 1, the attributes ‘age’, ‘sex’, ‘complaints’, ‘signs’ and ‘diagnosis’ of the object named ‘patient’ describe identically named features of a patient. The object tree specifies for each attribute: 2 OVERVIEW OF THE HEPAR SYSTEM 5 if lessthan(biochemistry,AP,60) and greaterthan(biochemistry,gamma-GT,100) and between(biochemistry,ASAT,[50,200]) and between(biochemistry,ALAT,[50,200]) and lessequal(biochemistry,total_bili,17) then conclude(patient,type-disorder,biliary_obstructive) with CF = 0.25 conclude(patient,type-disorder,hepatocellular) with CF = 0.60 fi Figure 2: Example of a production rule from the HEPAR system. • a value type (singlevalued, multivalued, yes/no), • a domain (integer, real, boolean, textvalue), and • a trace-class. A singlevalued attribute can only take one value with absolute certainty at a time. The domain of singlevalued attributes may be a real or integer interval, boolean or textvalue (a set of strings). Contrary to singlevalued attributes, multivalued attributes can take more than one (text)value at the same time. For instance, a patient can be of only one age (singlevalued attribute), but can suffer from several complaints at the same time (multivalued attribute). A yes/no attribute is a special singlevalued attribute which can take only one of the two values ‘yes’ and ‘no’. The trace-class specifies for each attribute whether it should be asked from the user, de- noted by the keyword askfirst (or initialdata if the attribute should be asked at the beginning of the consultation), should be inferred from the rule base, in which case it is declared as a goal, or neither of the two, in which case it is only considered by the inference engine when required for determining the values of some other attribute. All goal attributes in the system are contained in the ‘patient’ object. The main goal attributes are: • the nature of the disorder (benign or malignant), • the type of the disorder (hepatocellular or biliary obstructive), and • the final diagnosis. Furthermore, portal hypertension is a goal attribute, since it is a complication whose presence is very important to assess in view of the far-reaching consequences for therapy and prognosis. The goal attribute ‘type of the disorder’ is multivalued. In this way, it is expressed that hepatocellular and biliary obstructive features are not mutually exclusive alternatives, but can be present in the patient simultaneously. Similarly, several interrelated diseases may be present within the same patient, so the attribute ‘diagnosis’ is also multivalued. The rule base of HEPAR consists of a collection of 533 production rules having 1032 conclusions. An example of a production rule selected from HEPAR is shown in Figure 2. 3 DESIGN OF THE BELIEF NETWORK 6 This production rule consists of five conditions and two conclusions which are drawn only if all conditions are satisfied. This depends on given data and derived facts concerning a specific case. With each conclusion, a certainty factor is associated. A certainty factor expresses on a scale from −1 to +1 the confidence in the correctness of the conclusion, provided that all conditions of the rule are satisfied with absolute certainty. Under the given conditions, the example rule concludes that a hepatocellular disorder is more likely (CF = 0.60) than a biliary obstructive disorder (CF = 0.25). Most of the production rules concern data obtained from medical history, physical exami- nation and routine blood-chemistry. A brief analysis according to the conclusions of the rules will give an impression of the complexity of the rule base. The duration of complaints, clinical signs and lab abnormalities (attribute ‘duration’) is deduced in only 3 rules (one for each of the values acute, subacute and chronic). A total of 58 rules conclude about the nature of the disorder, with 44 conclusions having malignant and 54 having benign as a value. Conclusions about the type of the disorder are drawn in 257 rules (249 conclude hepatocellular, 244 biliary obstructive); of these, 225 are similar to the one shown in Figure 2. About the diagnosis of the patient 339 conclusions are drawn in 224 rules. During a consultation, the production rules are applied by the inference engine to derive values for the goal attributes. The algorithm used is known as top-down inference or backward chaining. The process of data collection is guided by the course of the inference, except as far as the initialdata attributes are concerned (see above). Part of the reasoning in the system concerns the manipulation of the certainty factors to propagate the uncertainty in problem-solving knowledge to the final conclusions concerning the goal attributes. The representation of knowledge by means of production rules with object-attribute-value triples, as well as top-down inference and the certainty-factor model are described in more detail in [1, 12]. 3 Design of the belief network 3.1 Basic notions of belief networks The formalism of belief networks offers an intuitively appealing approach for expressing in- exact causal relationships between domain concepts [7, 20]. A belief network consists of two components [3]: • A qualitative representation of the variables and relationships between the variables dis- cerned in the domain, expressed by means of a directed acyclic graph G = (V (G), A(G)), where V (G) = {V1, V2, . . . , Vn} is a set of vertices, taken as the variables, and A(G) a set of arcs (Vi, Vj ), where Vi, Vj ∈ V (G), taken as the relationships between the variables. • A quantitative representation of the ‘strengths’ of the relationships between the vari- ables, expressed by means of assessment functions. Note that the expressive power of the belief network formalism for representing causal knowl- edge is restricted by the acyclic nature of the directed graph; cyclic causal influences cannot be represented in the formalism. More about this restriction will be said in Section 3.3.4. Each of the variables Vi ∈ V (G) may take one value from the variable’s outcome space. The presence of an arc (Vi, Vj ) in A(G) represents a direct influential or causal relationship from Vi to Vj ; its absence denotes that the variables Vi and Vj are assumed not to influence 3 DESIGN OF THE BELIEF NETWORK 7 each other directly. Because of the causal meaning of the arcs, the qualitative representation of the domain will be called a causal graph in this article. For each vertex Vi an assessment function, denoted by γVi , is defined by the conditional probability distributions P (Vi = vi|cπ(Vi)), where vi denotes an arbitrary value of Vi and cπ(Vi) a conjunction of arbitrary values for each of the variables in the set π(Vi) of parents of Vi. In case the set of parents of the vertex Vi is empty, the probability function to be specified is just the prior probability distribution P (Vi = vi), for each value vi. The assessment functions together define a probability distribution on the domain. After a belief network has been specified, it can be used for reasoning about uncertain information in the domain concerned. For example, one may enter into a medical belief network information concerning certain symptoms and signs observed in a patient, after which an updated probability distribution will be computed. The process of computing the updated probability distribution after entering specific evidence into the belief network is called evidence propagation. Currently, there are two frequently applied algorithms for evidence propagation in use in belief networks. The first, and oldest, scheme originates from J.H. Kim and J. Pearl [7]. This scheme is only applicable in belief networks in which the causal graph has the form of a singly connected graph, i.e. a directed acyclic graph in which at most one path exists between any two vertices. The second scheme for evidence propagation originates from the work of S.L. Lauritzen and D.J. Spiegelhalter [8]. This algorithm is applicable in arbitrary belief networks. However, several techniques are known for extending the method of Kim and Pearl to arbitrary belief networks [20]. For an in-depth account of belief networks, the interested reader is referred to [18, 20]. For an overview of the principles of belief networks, the reader is referred to [12]. There are now several expert system shells available for building belief networks. For the experiments described in this article, we used the IDEAL system [27], to which some small programs for the graphical display and printing of belief networks were added. 3.2 Design considerations In developing the belief network, the HEPAR system was taken as the point of departure. However, as we shall see, using the HEPAR system as a source for statistical variables would result in a belief network consisting of over 300 vertices. We have therefore confined the scope of the experiment by focussing on three prototypical aspects of the field: • cholestasis, which is an important pathophysiological concept used in the early assess- ment of the patient; • cirrhosis, which is an important concept embodying a large number of chronic disorders of the liver and biliary tract; • Wilson’s disease, which is a specific chronic disorder affecting the liver. In this recessively inherited derangement of copper metabolism, cirrhosis may develop as a consequence of progressive copper accumulation in the liver; copper deposits in other organs may cause extrahepatic disease and peripheral cornea pigmentations, called Kayser-Fleischer rings. In this paper, we confine ourselves to discussing the last two aspects of the field. The development of the collection of causal graphs was undertaken in several steps: 3 DESIGN OF THE BELIEF NETWORK 8 1. The variables, with associated outcome spaces, playing a central role in the domain were discerned, yielding an initial set of vertices. 2. The possible set of states for each variable was determined. 3. The causal relationships between the various variables were established, resulting in a set of arcs and several additional intermediate vertices. After this initial design of the causal graphs, which was based on medical considerations only and not on considerations with respect to the final implementation, cycles introduced in the graph were cut, and some vertices were removed or added to simplify probability assessment. Finally, the causal graph concerning Wilson’s disease was transformed into a belief network by filling in the assessment functions. In the following sections, the above-mentioned steps will be discussed one by one. 3.3 Design of the causal graphs 3.3.1 Choice of statistical variables The object tree of HEPAR served as the point of departure for the choice of the statistical variables in the design of the belief network. The concept of attribute closely resembles the concept of variable, and the notion of the domain of an attribute is similar to the notion of the outcome space of a statistical variable. More precisely, a statistical variable may be viewed as a singlevalued attribute, because the values in the outcome space of a variable are mutually exclusive. A detailed analysis of the HEPAR system revealed that its object tree contains 102 dif- ferent attributes grouped by 11 objects. Subdivided by value type, there are 30 yes/no, 49 singlevalued and 23 multivalued attributes. Whereas translating yes/no and singlevalued attributes into statistical variables is straightforward, converting a multivalued attribute re- quires splitting up its domain into mutually exclusive subsets, thus obtaining a collection of variables with associated outcome spaces. As an example of this translation process, consider the multivalued attribute ‘blood count’ which is contained in the object ‘haematology’. Among others, the domain of this attribute contains the values ‘normal’, ‘lymphocytosis’, ‘leukocytosis’ and ‘leukocytopenia’. In the translation process, related attribute values were combined as the possible values of a new statistical variable. To the outcome space of each of these variables we added the special value ‘normal’, thus obtaining a non-binary variable. In this case, we obtained the statistical variable ‘leukocyte count’ with outcome space {leukocytosis, normal, leukocytopenia}. At- tribute values having no direct relationship to any other value were translated into binary statistical variables. For example, the value ‘lymphocytosis’ was translated into the variable ‘lymphocytosis’ with outcome space {yes, no}. The number of values in the domain of the multivalued attributes in the HEPAR object tree varies from 2 (e.g. for the attribute ‘type disorder’) up to 87 for the attribute ‘diagnosis’ (of which 9 not yet concluded in production rules). Most multivalued attributes have domains in which neither of the values excludes any one of the others. The conversion of such a mul- tivalued attribute with a domain consisting of n values will thus result in n binary variables. As a consequence, splitting up the 23 multivalued attributes in HEPAR will yield almost 250 variables. Hence, using only the HEPAR system as a source for variables to be discerned will already render a belief network consisting of over 300 vertices. 3 DESIGN OF THE BELIEF NETWORK 9 The theory underlying belief networks does not provide a means for grouping of variables; consequently, no analogy of the concept of object exists. By converting the object tree into a set of statistical variables, the structural coherence between closely related concepts provided in HEPAR is lost; variables with other than causal relations are dispersed over the network. 3.3.2 State assessment of variables In the previous section, we have shown how the translation of the attributes of HEPAR into statistical variables was carried out. Now, if V is any statistical variable, we may want to express probabilistic statements of the form P (V = v|cπ(V )) where V is an arbitrary variable and v a value of V , or P (v1 ≤ V < v2|cπ(V )) for V a numerical (integer or real) variable. Expressions such as V = v and v1 ≤ V < v2 will be called states of a variable in this article. One of the problems in distinguishing states for the statistical variables discerned was that the theory of belief networks is not capable of dealing with continuous probability distribu- tions. As a practical solution to representing a continuously distributed variable in a belief network, the variable is simply made discrete by subdividing its outcome space into a finite number of intervals, and taking these intervals as states. In case of the HEPAR system, such a conversion was required for singlevalued attributes with a real or integer domain. The assessment of variable states was guided by the conditions and conclusions in the production rules of HEPAR. For non-numerical attributes, the states distinguished for the associated variables corresponded in most cases to the values enumerated in the associated outcome space. Most of these values were referred to in the conditions and conclusions of the rules. Numerical attributes were transformed into discrete statistical variables with a finite number of interval states, where the assessment of the cut-off points of the intervals was based on the conditions in which the corresponding attributes appeared. If necessary, these intervals were further refined by using cut-off points mentioned in the hepatological literature [16]. Consider for example Figure 2. This rule provides us with one cut-off point for ‘AP’ (alkaline phosphatase), ‘gamma-GT’ and ‘total bili’ (total bilirubin); two cut-off points can be derived for ‘ASAT’ and ‘ALAT’ from this rule. In HEPAR, 20 of the 49 singlevalued attributes have a numerical domain. For 19 of these singlevalued attributes, from one up to nine cut-off points were distinguished. One attribute is only used in an arithmetical expression, so no cut-off points could be determined for this attribute. ‘Age’ was the attribute for which nine different cut-off points were discerned. Since many disorders are age-specific, it seems inevitable to distinguish so many states for this feature. In [2] for instance, its range is subdivided in seven categories. The results of this translation process are summarized in Table 1. So far, the conversion of the HEPAR system into a belief network did not cause insurmountable problems. Only a small number of attributes could not be converted into discrete statistical variables without using information from other sources than the HEPAR system. 3 DESIGN OF THE BELIEF NETWORK 10 Attributes in HEPAR Derived variables No. of values Type No. No. 2 3 4 5 6 7 8 9 10 unknown yes/no 30 30 30 singlevalued 49 textvalue 29 29 20 5 4 numerical interval 20 20 8 3 2 1 2 2 1 1∗ multivalued 23 > 246 242 4 −∗∗ Table 1: Summary of the results of the conversion of HEPAR attributes into discrete statistical variables. ∗) No cut-off points could be derived from the HEPAR rule base. ∗∗) In the current implementation, no domain is specified for the multivalued attribute ‘drug’, nor is the attribute applied in the rules. 3.3.3 Detecting causal relations between variables In rule-based systems, attributes are placed into relation with each other by means of pro- duction rules. In a diagnostic expert system like HEPAR, most of these relations concern associations between disease manifestations on the one hand and diagnostic conclusions on the other hand, which are largely heuristic and not causal in nature. Although the notion of causality employed in belief networks is interpreted somewhat broadly as any relationship in which one variable influences another, the differences between heuristic and causal knowledge imply that rules cannot be translated into causal graphs in a straightforward way. The analysis of the rule base of the HEPAR system was thus guided by the question of how the associations had to be interpreted in terms of causality. It turned out that in broad outline three different possibilities can be distinguished: 1. The condition (e.g. age) is viewed as one of the ‘causes’ of the conclusion (e.g. Wilson’s disease). 2. The condition (e.g. Kayser-Fleischer rings) is viewed as one of the effects of the conclu- sion (e.g. Wilson’s disease). 3. Both a condition (e.g. Kayser-Fleischer rings) and a conclusion (e.g. chronic duration of complaints) are viewed as the effects of a common cause (e.g. Wilson’s disease). In the last case, we see that a variable taking part in a causal relation is actually missing in the rule concerned. As we shall see in the following, in designing a belief network, taking a set of production rules as a point of departure will yield an incomplete belief network. So, it will be necessary to add several intermediate vertices in order to obtain a satisfactory description of the causal disease mechanisms involved. We have used additional knowledge from hepatological textbooks, in particular [34], in order to extend the heuristic associations in the rules to causal relations. We shall illustrate this process with the help of the HEPAR rule shown in Figure 3. In this rule, findings from history and physical examination are associated with a conclusion stating that the patient’s disorder is chronic. An analysis of the relations between the variables corresponding to the mentioned object-attribute-value triples in terms of causes and effects is summarized in the following paragraph. Variables occurring in the rule have been indicated in italic. 3 DESIGN OF THE BELIEF NETWORK 11 if greaterthan(patient,time_course,26) or same(patient,complaint,haematemesis) or same(patient,signs,xanthomata) or same(patient,signs,testicular_atrophy) or same(patient,signs,spider_angiomas) or same(patient,signs,palmar_erythema) or same(patient,signs,Kayser_Fleischer_rings) or same(patient,signs,gynaecomastia) or same(patient,signs,oesophageal_varices) or same(patient,signs,caput_medusae) or same(patient,signs,butterfly_erythema) then conclude(complab,duration,chronic) with CF = 1 fi Figure 3: Heuristic knowledge in the HEPAR system. The time course of the disorder is a direct result of its acute, subacute or chronic duration. Two sorts of chronic disorder of the liver are of importance in this respect: cirrhosis and chronic active hepatitis. The two major complications of cirrhosis are hepatocellular failure and portal hypertension. Portal hypertension may lead to the development of portasystemic collaterals such as oesophageal varices and caput medusae (dilated skin veins surrounding the navel); rupture of an oesophageal varix causes haematemesis (vomiting of blood). The pathogenesis of testicular atrophy, gynaecomastia, spider angiomas and palmar erythema in subjects with cirrhosis is not entirely clear, but seems related to abnormal sex hormone metabolism, among others a raised oestrogen concentration as a consequence of decreased liver function. (Oestrogens are metabolized largely by the liver.) Kayser-Fleischer rings are almost pathognomonic for Wilson’s disease, but can also be found in primary biliary cirrhosis. In both disorders, the liver is usually cirrhotic. Primary biliary cirrhosis causes obstruction of the small intrahepatic biliary ducts, and since in chronic cholestasis very high plasma lipid levels can occur, xanthomata (deposits of lipids in the skin) may also appear in this disease. Chronic active hepatitis, in itself a type of chronic liver disease, is also a likely cause of cirrhosis. One of the disorders resulting in chronic active hepatitis is autoimmune chronic hepatitis, which is associated with butterfly erythema. Part of the knowledge lacking in the production rule analysed above is represented in other HEPAR rules, namely the rules concerning portal hypertension, Wilson’s disease, pri- mary biliary cirrhosis and autoimmune chronic hepatitis. When represented as a causal graph Figure 4 is obtained. As can be seen, all conditions in the example rule from Figure 3, except the first, correspond to variables directly or indirectly influenced by a variable which also influences the variable corresponding to the rule’s conclusion. The vertex ‘cirrhosis’ has been added to enable connecting the conclusion to the conditions 2, 4-6 and 8-10; the vertices la- belled ‘portal hypertension’, ‘portasystemic collaterals’, ‘liver function’ and ‘oestrogens’ have been added here to obtain a more natural causal, pathophysiological description in corre- spondence with that provided in the textbooks. The disease causally related to condition 11 (butterfly erythema) is ‘autoimmune chronic hepatitis’, which has been added to the graph 3 DESIGN OF THE BELIEF NETWORK 12 PALMAR ERYTHEMA YES NO SPIDER ANGIOMAS YES NO TESTICULAR ATROPHY YES NO CAPUT MEDUSAE YES NO HAEMATEMESIS YES NO XANTHOMATA YES NO KAYSER-FLEISCHER RINGS YES NO TIME COURSE < 12 WEEKS 12 − 26 WEEKS > 26 WEEKS BUTTERFLY ERYTHEMA YES NO GYNAECOMASTIA YES NO OESOPHAGEAL VARICES YES NO PLASMA LIPIDS OESTROGENS CHOLESTASIS DURATION ACUTE SUBACUTE CHRONIC LIVER FUNCTION PORTASYSTEMIC COLLATERALS PORTAL HYPERTENSION YES NO CIRRHOSIS YES NO WILSON’S DISEASE YES NO PRIMARY BILIARY CIRRHOSIS YES NO CHRONIC ACTIVE HEPATITIS AUTOIMMUNE CHRONIC HEPATITIS YES NO Figure 4: Causal graph corresponding to the Hepar rule in Figure 3; additional causal knowl- edge is derived from textbooks. Added variables are dashed. 3 DESIGN OF THE BELIEF NETWORK 13 cause or risk factor specific disease disease-specific symptom or sign disease group group-specific symptom or sign ... ... ... ... ... ... Figure 5: Global graph structure for a diagnostic belief network. as a common cause of ‘cirrhosis’ (via ‘chronic active hepatitis’) and ‘butterfly erythema’. In a similar way, condition 7 (Kayser-Fleischer rings) caused addition of the vertices ‘primary biliary cirrhosis’ and ‘Wilson’s disease’, while condition 3 (xanthomata) has been explained by means of ‘cholestasis’ and ‘plasma lipids’. This example brings into the picture a third type of reasoning in addition to causal and heuristic reasoning: taxonomic reasoning. Taxonomic reasoning is based on hierarchical classification of concepts in groups and subgroups according to common features. The higher a concept is placed into a hierarchy, the more general it is. In HEPAR, for instance, this type of reasoning is applied in the rules concerning specific diseases. One of the questions that arose in representing taxonomic knowledge in a causal graph is whether the arcs between the vertices should be drawn from vertices representing specific concepts to vertices representing more general concepts, or the other way around. It is easy to see that, as far as taxonomies of disorders are concerned, we should follow the first approach, since disorders are categorized according to common features resulting from the disorders. The resulting method is depicted in Figure 5. Part of the graph in Figure 4 may thus also be interpreted as a hierarchy, with the vertex ‘duration’ representing the most general concept, ‘cirrhosis’ and ‘chronic active hepatitis’ as intermediate concepts, and ‘primary biliary cirrhosis’, ‘Wilson’s disease’ and ‘autoimmune chronic hepatitis’ as the most specific concepts, being specific disorders. 3.3.4 Cycles in graphs and pathophysiological knowledge Representing causal knowledge as a graph as described in the previous subsection does not necessarily render a causal graph. As we have discussed in Section 3.1, a causal graph is by definition acyclic; in representing pathophysiological knowledge, however, cycles may be introduced. So, the restricted expressive power of belief networks in representing cyclic in- fuences may hinder a natural representation of medical knowledge. In developing our causal graphs, we obtained directed cycles representing positive feedback influences, as well as cycles representing negative feedback influences. Although we do not provide a theoretical solution to dealing with directed cycles in the article, we show that by means of domain-specific ar- 3 DESIGN OF THE BELIEF NETWORK 14 guments, the cycles may be cut. Here, we only discuss the handling of negative feedback influences. The process is illustrated by a causal graph for cirrhosis which we developed in converting the HEPAR system. The graph in Figure 6 represents the causal relations between cirrhosis, portal hyper- tension and ascites (fluid in the peritoneal cavity). As depicted, the distortion of the liver architecture in cirrhosis may lead to portal hypertension and, as a consequence of a partial block to hepatic venous outflow, to an increased exudation of hepatic lymph possibly exceed- ing the reabsorption capacity. In that case, the additional protein-rich lymph exudes from the liver surface into the peritoneal cavity. Another complication of cirrhosis is an impaired liver cell function, which may result in a reduced albumin synthesis. Both the albumin loss to the peritoneal cavity and the impaired synthesis of albumin affect the amount of serum albumin; the plasma volume determines its actual concentration. When portal hypertension is transmitted back to the splanchnic area, transudation of fluid into the peritoneal cavity will result, which is also favoured by a reduced colloid osmotic pressure consequent to reduced serum albumin. Because of this loss of fluid into the peritoneal compartment and of pooling of blood in the splanchnic circulation, the ‘effective’ blood volume is reduced. Renal perfu- sion therefore falls and the kidneys retain sodium (and water), so that the plasma volume increases. This renal regulatory mechanism is represented as the cycle numbered ‘1’ in Fig- ure 6. This negative feedback process, which is also effective under normal conditions, takes care of a stable plasma volume. Splanchnic pooling results through this cycle in a resetting of the plasma volume to a higher level. Because of this and since the renal regulatory mech- anism is in fact beyond the scope of the domain, it seems justified to draw an arc between ‘splanchnic pooling’ and ‘plasma volume’, and to remove the remainder of the cycle from the graph. The second cycle depicted in the graph represents a pathological negative feedback process. The effect of splanchnic fluid transudation on the plasma volume, however, is undone by the above-mentioned renal regulatory mechanism. Therefore, cycle 2 may safely be cut by removing the arc from ‘splanchnic fluid transudation’ to ‘plasma volume’. 3.3.5 Final adjustment of the causal graph In addition to the problem of handling cyclic causal influences in a causal graph, we have to solve the problem of keeping the number of vertices as small as possible. The main reasons for this are: • Since many of the corresponding variables are hardly accessible for observation, prob- abilities with regard to them will not be available in literature nor obtainable from experts. • Inference speed will decrease considerably as the number of variables in the belief net- work increases. We recognize two methods for reducing the number of vertices. The first one involves cut- ting off dead-end paths without any vertices representing diagnostic hypotheses or disease manifestations. Because paths not involving such vertices are not of interest in diagnostic problem-solving, they may be removed. In the second approach, we omit some of the in- termediate vertices; the remaining ones are connected in an obvious way, as is illustrated in Figure 7 which has been obtained from Figure 6. The gain expected from a reduction of the number of additional vertices should be weighted against the loss of information involved. Leaving out intermediate steps in causal mechanisms 3 DESIGN OF THE BELIEF NETWORK 15 SODIUM RETENTION ASCITES YES NO OEDEMA YES NO RENAL PERFUSION SPLANCHNIC FLUID TRANSUDATION EFFECTIVE BLOOD VOLUME SPLANCHNIC POOLING SERUM ALBUMIN < 30 g/l 30 − 40 g/l 40 − 55 g/l COLLOID OSMOTIC PRESSURE LIVER SYNTHESIS CAPACITY PLASMA VOLUME PRESSURE SPLANCHNIC AREA ALBUMIN LOSS LIVER CELL MASS HEPATIC LYMPH EXUDATION PORTAL HYPERTENSION YES NO LIVER ARCHITECTURE CIRRHOSIS YES NO 1 2 Figure 6: Cycle 1 represents a negative feedback process for keeping the plasma volume up to a constant level. The effect of splanchnic pooling on this process basically is an adjustment of the plasma volume to a higher steady state. It therefore seems justified to eliminate the dashed vertices and redirect the arc departing from ‘splanchnic pooling’ to ‘plasma volume’. Since the effect of splanchnic fluid transudation on the plasma volume is undone by the above-mentioned regulatory mechanism, cycle 2 can harmlessly be interrupted by removing the crossed arc. 3 DESIGN OF THE BELIEF NETWORK 16 OEDEMA YES NO ASCITES YES NO SERUM ALBUMIN < 30 g/l 30 − 40 g/l 40 − 55 g/l PORTAL HYPERTENSION YES NO CIRRHOSIS YES NO Figure 7: A strongly simplified version of the graph in Figure 6. Vertices not discerned in HEPAR have been removed. A pair of vertices has been connected by an arc iff in the original graph a path existed between them. renders the network harder to interpret and harder to adjust. We have therefore been very reluctant in applying this method. 3.4 Probability assessment 3.4.1 Sources of probabilistic information After having carried out the steps discussed above, several causal graphs were obtained. In turning these into belief networks, we had to find a fill-in for the assessment functions by (conditional) probabilities. On first thought, we hoped it might be possible to derive this probabilistic information from the certainty factors specified in the production rules in the HEPAR system. However, recall from the preceding sections that none of the production rules could directly be translated into a causal graph with the conclusion as a vertex having incoming arcs originating from vertices representing the conditions. So, the certainty factors specified in the production rules could not be applied as the basis of probability assessment. We have therefore consulted standard medical textbooks for collecting such probabilistic information [34, 21]. One of the difficulties encountered in obtaining assessment functions from medical litera- ture is that usually in describing diseases a non-numerical notion of uncertainty is employed, featured by terms such as ‘rare’, ‘common’, ‘almost all’, etcetera. In case probabilistic in- formation is specified, it is often incomplete. Also, available conditional probabilities may concern variables not causally related to each other, usually rendering these numbers useless for specifying the assessment functions in a belief network. Such information can, however, be applied for the evaluation of a belief network. Furthermore, probabilistic information from the literature is not always based on the population of the country concerned. Because factors such as hospital referral patterns and geographical variation in disease presentation play a role, probabilistic information available from studies carried out elsewhere may not be useful. Non-conditional probabilities seem to be more susceptible to population characteristics than conditional probabilities are. 3 DESIGN OF THE BELIEF NETWORK 17 WILSON’S DISEASE (= D) YES (= d1) NO (= d2) SERUM CAERULOPLASMIN (= SC) < 200 mg/l (= sc1) 200 − 300 mg/l (= sc2) ≥ 300 mg/l (= sc3) WILSONIAN SYMPTOMS (= S) YES (= s1) NO (= s2) HEPATIC COPPER (= HC) 20 − 50 µg/g (= hc1) 50 − 250 µg/g (= hc2) ≥ 250 µg/g (= hc3 WILSON’S DISEASE GENOTYPE (= G) HOMOZYGOUS (= g1) HETEROZYGOUS (= g2) NORMAL (= g3) AGE (= A) 0 − 6 (= a1) 6 − 10 10 − 16 16 − 25 . . . 25 − 40 ≥ 40 (= a6) Figure 8: Causal graph representing some qualitative knowledge with regard to Wilson’s disease. One way to simplify obtaining probabilistic information is to reduce the number of parents of a given vertex in a causal graph. This may be done by introducing ‘summarizing’ interme- diate vertices, a technique known as ‘divorcing multiple parents’ [19]. The main advantage of this technique is that conditional probabilities P (V = v | v1 ∧ · · · ∧ vn) where n is small, say less than 4, are much easier to estimate than for n large. For our application, the number of parents of vertices was always limited. 3.4.2 Estimating probabilities We shall illustrate the process of probability assessment using information from medical liter- ature, by discussing the belief network we obtained for Wilson’s disease in some detail. This disorder is attended with low levels of serum caeruloplasmin and progressive copper accumu- lation in the liver. To be more precise, it is the genotype with regard to this disease which influences serum caeruloplasmin and hepatic copper levels. The capacity of the hepatocytes to store copper is eventually exceeded and release into blood and uptake in extrahepatic sites occurs, causing extrahepatic disease and Kayser-Fleischer rings. In Figure 8, the rela- tionships among the above-mentioned concepts have been represented by means of a causal graph. Clinical manifestations, here summarized in the vertex ‘Wilsonian symptoms’, can be due to liver involvement as well as to involvement of other organs. From the observation that Wilson’s disease is a recessively inherited disorder all homozy- gous subjects will eventually have symptoms of, the values for the assessment function for the vertex ‘Wilson’s disease’ follow immediately. Only subjects with a pair of ‘Wilson’s alleles’ have the disease. Hence, the values of the assessment function γD are as depicted in the right half of Table 2. According to Mendelian laws, we have P (g1) = P (g1) · P (g1) + 1 2 P (g1) · P (g2) + 1 4 P (g2) · P (g2) in non-consanguineous societies. Rewriting this equation results in P (g2) = −P (g1) + √ P (g1) · (4 − 3P (g1)). 3 DESIGN OF THE BELIEF NETWORK 18 γG(gi) γD(dj | gi) Wilson’s disease Wilson’s disease genotype yes no Homozygous 0.00003 Homozygous 1 0 Heterozygous 0.01092 Heterozygous 0 1 Normal 0.98905 Normal 0 1 Table 2: Assessment functions γG(gi) and γD(dj | gi). From the literature we have that the world-wide prevalence of Wilson’s disease is about one in 30 000, we can now assess γG. Its values are given in the left part of Table 2. Here, we have to keep in mind, however, that we are developing a belief network for use in hepatological patients, in whom the prevalence of Wilson’s disease is much higher. With regard to ‘age’, a similar problem occurs: age distribution in the population of a specialized liver unit is very likely to differ from that in normal population. For serum caeruloplasmin, we found that 94.9% and 20.4% of homo- and heterozygotes for the Wilson’s disease gene, respectively, have levels below 200 mg/l [30]. A table containing a summary of analytical data (means and standard deviations) in patients with Wilson’s disease, heterozygous carriers, and control subjects [29], allowed us to estimate values for the assessment function γSC for caeruloplasmin levels, using statistical techniques assuming normal distribution. Hepatic copper (µg/g dry weight) Group No. Range Mean ± SD Wilson’s disease: Asymptomatic 36 152 – 1828 983.5 ± 368 Symptomatic 33 94 – 1360 588.3 ± 304 Heterozygous carriers 14 39 – 213 117.0 ± 51 Normal subjects 16 20 – 45 31.5 ± 6.8 Table 3: Summary of analytical data in patients with Wilson’s disease, heterozygous carriers, and control subjects. Source: [29]. The same techniques were applied to the information with regard to hepatic copper levels, reproduced in Table 3. Contrary to serum caeruloplasmin, the hepatic copper concentration is not only influenced by the Wilson’s disease genotype, but also by age. This applies only to homozygous patients, however. In normal subjects and heterozygous carriers, hepatic copper concentration does not vary with age, so that we have P (hci | gj ∧ ak) = P (hci | gj ) if j 6= 1. These values are given in the lower half of Table 4, which has been derived from Table 3 as described above. With regard to homozygotes, conditioning on the values of S (Wilsonian symptoms) gives us P (hci | g1 ∧ ak) = ∑ s P (hci | g1 ∧ ak ∧ s) · P (s | g1 ∧ ak). We next assumed that the influence of age on hepatic copper can be neglected once we know whether a Wilson’s disease patient is symptomatic or not. The last equation can then be 3 DESIGN OF THE BELIEF NETWORK 19 Hepatic copper (µg/g dry weight) Group < 50 50 − 250 ≥ 250 Wilson’s disease: Asymptomatic 0.0056 0.0176 0.9768 0.0020 – 0.0139 0.0079 – 0.0356 0.9505 – 0.9901 Symptomatic 0.0384 0.0951 0.8665 0.0166 – 0.0793 0.0542 – 0.1473 0.7734 – 0.9292 Heterozygous carriers 0.0951 0.9004 0.0045 0.0294 – 0.2327 0.7666 – 0.9494 0.0007 – 0.0212 Normal subjects 0.9967 0.0033 0 0.9857 – 0.9994 0.0006 – 0.0143 Table 4: Conditional probabilities of the values of hepatic copper concentration given Wilson’s disease genotype and presence or absence of Wilsonian symptoms. The probabilities are derived from Table 3 by statistical methods assuming normal distribution. The probability intervals on the lower lines are meant as an indication of the accuracy of the corresponding probabilities on the upper lines. simplified to: P (hci | g1 ∧ ak) = ∑ s P (hci | g1 ∧ s) · P (s | g1 ∧ ak) Note that eventually all untreated subjects with a pair of Wilson’s disease genes will become symptomatic Wilsonian patients. Then, Figure 9 provides us with the required information for assessing the conditional probabilities P (s | g1 ∧ ak). The conditional probabilities P (hc | g1 ∧ s) can be assessed from Table 4, so that we can compute the remaining values of γHC . The resulting belief network for Wilson’s disease is shown in Figure 10. The bar charts in this figure show the prior probability distributions for the variables. We have added vertices for the inheritance of Wilson’s disease (Wilson’s disease genotype of mother, father and sibling). Wilsonians symptoms have been futher specified as Kayser-Fleischer rings, cirrhosis, psychiatric, neurological and renal disease; the corresponding assessment functions have been obtained from [34]. 3.4.3 Preliminary evaluation of the belief network Evaluation of the belief network for Wilson’s disease has been carried out by using data from patients with Wilson’s disease, and by comparing the probabilities concerning symptoms and signs, when certain evidence was entered into the graph, with statistical information concerning Wilson’s disease in the literature. For purpose of the evaluation, we have tried to simulate the situation of an internal medicine clinic. In that case, the prevalence of Wilson’s disease is much higher than in the general population. We assumed that Wilson’s disease will be encountered in about one in 200 patients. As an example, using data of a patient that has been used in the evaluation of the HEPAR system, resulted in a probability of Wilson’s disease of 0.774; the HEPAR system was also capable of concluding the patient to suffer from Wilson’s disease with a certainty factor equal to 0.69. Similar results were obtain for five patients reported in the medical literature [6]. 4 DISCUSSION 20 0 5 10 15 20 25 30 35 40 45 years of age 0 20 40 60 80 100 % F (x) 0.2% 9.0% 40.7% 72.6% 94.1% 99.9% Figure 9: Cumulative percentage of 121 patients, F (x), with onset of overt symptoms of Wilson’s disease before age x. Source: [29]. Since eventually all untreated homozygotes for the Wilson’s disease gene will become symp- tomatic Wilsonian patients, the dashed lines and corresponding percentages give us the proba- bilities of symptomatic Wilson’s disease in homozygous subjects from the distinct age groups. The belief network was next evaluation in a test in which, after entering certain combi- nations of evidence mentioned in the literature, computed probabilities were compared with qualitative and quantitative statements concerning Wilson’s disease that have appeared in literature. Because most of these statements were qualitative in nature, it was not possible to obtain precise evaluation results in this study. We had the impression that the probabilities computed using the belief network, fitted the descriptions in the literature quite closely. 4 Discussion Diagnostic problem-solving in medicine can be considered from many different perspectives. The same is true for computer-based decision-support systems, such as medical expert sys- tems. On the one hand, assistance in medical diagnosis can be offered by expert systems employing heuristic reasoning models, based on the experience of clinicians in the manage- ment of their patients. In such systems, decisions as to which diagnostic procedures ought be untertaken for a given patient, are implicitly embodied in the knowledge encoded and is applied by means of some diagnostic strategy [14]. On the other hand, decision theory, being a combination of probability theory and utility theory, makes it possible to explicitly reason about such decisions, allowing the analysis of cost/benefit considerations of any action undertaken by the clinician, an application not readily accessible by the heuristic models. However, the development of such decision-theoretical systems for areas of the size that can be handled by current expert system technology, is a major undertaking. It therefore seems attractive to consider starting the development of such decision-theoretical models by taking existing heuristic expert systems as a point of departure. Belief networks may be taken as a starting point for the development of decision-theoretical models. In the present article, we have described an experiment on converting a rule-based expert system into a belief network, looking at the notion of belief network primarily as a means for representing causal medical knowledge. Rule-based systems for medical diagnosis, however, 4 DISCUSSION 21 KAYSER-FLEISCHER RINGS YES 0.004 NO 0.996 RENAL DISEASE YES 3.L-5 NO 1.000 PSYCHIATRIC DISEASE YES 3.L-4 NO 1.000 CIRRHOSIS YES 0.011 NO 0.989 WILSON’S DISEASE YES 3.L-5 NO 1.000 URINARY COPPER < 0.5 µmol/24h 0.945 0.5 − 1.6 µmol/24h 0.050 ≥ 1.6 µmol/24h 0.005 TOTAL SERUM COPPER DECREASED 0.026 NORMAL 0.943 INCREASED 0.031 MOTHER: WILSON’S DISEASE YES 3.L-5 NO 1.000 SIBLING: WILSON’S DISEASE YES 3.L-5 NO 1.000 FATHER: WILSON’S DISEASE YES 3.L-5 NO 1.000 NEUROLOGICAL DISEASE YES 9.L-4 NO 0.999 CAERULOPLASMIN SERUM COPPER < 9.5 µmol/l 0.005 9.5 − 14.3 µmol/l 0.420 14.3 − 19.0 µmol/l 0.574 SIBLING: WILSON’S DISEASE GENOTYPE HOMOZYGOUS 3.L-5 HETEROZYGOUS 0.011 NORMAL 0.989 TISSUE COPPER NORMAL 0.995 MOD. INCREASED 0.002 TOXIC 0.003 FREE SERUM COPPER 0.8 − 1.6 µmol/l 0.994 1.6 − 8.0 µmol/l 0.006 SERUM CAERULOPLASMIN < 200 mg/l 0.003 200 − 300 mg/l 0.419 ≥ 300 mg/l 0.578 HEPATIC COPPER 20 − 50 µg/g 0.985 50 − 250 µg/g 0.015 ≥ 250 µg/g 8.L-5 WILSON’S DISEASE GENOTYPE HOMOZYGOUS 3.L-5 HETEROZYGOUS 0.011 NORMAL 0.989 AGE 0 − 6 0.075 6 − 10 0.050 10 − 16 0.080 16 − 25 0.145 25 − 40 0.240 ≥ 40 0.410 FATHER: WILSON’S DISEASE GENOTYPE HOMOZYGOUS 3.L-5 HETEROZYGOUS 0.011 NORMAL 0.989 MOTHER: WILSON’S DISEASE GENOTYPE HOMOZYGOUS 3.L-5 HETEROZYGOUS 0.011 NORMAL 0.989 Figure 10: Belief network for Wilson’s disease. REFERENCES 22 mainly represent heuristic knowledge; they lack knowledge about causal directions in relation- ships between variables. As a consequence, it is generally not possible to design a mapping that permits the automatic conversion of a rule-based expert system into a causal graph. Moreover, our study shows that for estimating a probability distribution for a belief network, the measures of uncertainty associated with the rules in the corresponding rule-based expert system are of little value. Due to these differences in the type of knowledge represented and in the formalism used to represent uncertainties, much knowledge required in a belief network cannot be derived from a rule-based system. Hence, converting a rule-based system into a belief network requires (re)consulting medical literature and renewed knowledge elicitation from experts. However, our experiment shows that the knowledge represented in a rule-based system may be well suited as a foundation for building a belief network upon. Major variables as well as their outcome spaces can rather easily be derived from it. Moreover, although the relationships expressed in the rules do not represent causal direction, they still can serve as an aid in searching for causal relations. The main bottleneck in the conversion of a rule-based system into a belief network is the phase in which assessment functions are to be estimated. In our case, it turned out to be mandatory to consult other sources than the original rule base, in particular the medical literature. Unfortunately, the medical literature usually contains little suitable probabilistic information and uncertainty is often described in qualitative terms. When probabilistic in- formation is provided, the probability distribution is often incomplete or conditioned on the ‘wrong’ variables. Furthermore, probability assessment is essentially population specific; the patients visiting a specialized medical centre form a population different from the normal pop- ulation, which is likely to affect especially unconditional probabilities. Variables such as age and disease prevalence often are associated with unconditional probabilities, whereas these very probabilities are usually mentioned in literature for normal population. In for instance [28, 9, 10, 23], similar problems have been recognized in transferring diagnostic systems to other populations. The problem of probability assessment can partly be solved by keeping the causal graph as simple as possible. A great deal of thought should therefore be paid to the construction of the causal graph, in which the availability of probabilistic knowledge should be taken into account. Using subjective probabilities may also be a solution. In a study concerning the assessment of simple conditional probabilities (viz. of certain findings given certain diseases), Spiegelhalter et al. showed that reliable probability assessments can be obtained from experts [26]. Moreover, they propose a practical procedure for improving such subjective probabilities by combining them with observed data. However, since we have tried to construct a belief network that provides a clear representation of the causal, pathophysiological knowledge involved, subjective probability assessment becomes much more complicated. In order to obtain a reliable belief network in such cases, the availability of probabilities computed from a large series of real-life patients therefore seems inevitable. References [1] B.G. Buchanan and E.H. Shortliffe (Eds.), Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project (Addison-Wesley Publishing Company, Reading, Massachusetts, 1984). REFERENCES 23 [2] F. Burbank, A computer diagnostic system for the diagnosis of prolonged undifferenti- ating liver disease, American Journal of Medicine 46 (1969) 401–415. [3] L.C. van der Gaag, Probability-based models for plausible reasoning, Ph.D. Thesis, Uni- versity of Amsterdam, Amsterdam, 1990. [4] A. Haubek, J.H. Pedersen, F. Burcharth, J. Gammelgaard, S. Hancke and L. Willumsen, Dynamic sonography in the evaluation of jaundice, American Journal of Radiology 136 (1981) 1071–1074. [5] D.E. Heckerman, Probabilistic interpretations for MYCIN’s certainty factors, in: L.N. Kanal and J.F. Lemmer (Eds.), Uncertainty in Artificial Intelligence 3 (North- Holland, Amsterdam, 1986) 167–196. [6] T.U. Hoogenraad, C.J.C. van den Hamer, J. van Hattum, Effective treatment of Wilson’s disease with oral zinc sulphate: two case reports, British Medical Journal 289 (1984) 273–276. [7] J.H. Kim and J. Pearl, A computational model for causal and diagnostic reasoning in inference systems, in: Proceedings of the 8th International Joint Conference on Artificial Intelligence (Karlsruhe, West Germany, 1983) 190–193. [8] S.L. Lauritzen and D.J. Spiegelhalter, Local computations with probabilities on graphical structures and their application to expert systems, Journal of the Royal Statistical Society (Series B) 50 (1988) 157–224. [9] G. Lindberg, Studies on diagnostic decision making in jaundice, M.D. Thesis, Karolinska Institute, Stockholm, 1982. [10] G. Lindberg, C. Thomsen, A. Malchow-Møller, P. Matzen and J. Hilden, Differential diagnosis of jaundice: Applicability of the Copenhagen Pocket Chart proved in Stockholm patients, Liver 7 (1987) 43–49. [11] P.J.F. Lucas, R.W. Segaar and A.R. Janssens, HEPAR, an expert system for the diag- nosis of disorders of the liver and biliary tract, Liver 9 (1989) 266–275. [12] P.J.F. Lucas and L.C. van der Gaag, Principles of Expert Systems (Addison-Wesley Publishing Company, Wokingham, England, 1991). [13] P.J.F. Lucas and A.R. Janssens, Second validation of the HEPAR system, an expert system for the diagnosis of disorders of the liver and biliary tract, Liver 11 (1991) 340– 364. [14] P.J.F. Lucas and A.R. Janssens, Development and validation of HEPAR, an expert system for the diagnosis of disorders of the liver and biliary tract, Journal of Medical Informatics 16 (1991) 259–270. [15] W.B. Martin, P.C. Apostolakos and H. Roazen, Clinical versus actuarial prediction in the differential diagnosis of jaundice, American Journal of Medical Sciences 240 (1960) 571–578. REFERENCES 24 [16] P. Matzen, A. Malchow-Møller, J. Hilden, C. Thomsen, L.B. Svendsen, J. Gammelgaard, E. Juhl and the Copenhagen Computer Icterus Group, Differential diagnosis of jaundice: a pocket diagnostic chart, Liver 4 (1984) 360–371. [17] B. Middleton, M.A. Shwe, D.E. Heckerman, M. Henrion, E.J. Horvitz, H.P. Lehmann and G.F. Cooper, Probabilistic diagnosis using a reformulation of the INTERNIST-1/QMR knowledge base, II. Evaluation of diagnostic performance, Methods of Information in Medicine 30 (1991) 256–267. [18] R.E. Neapolitan, Probabilistic Reasoning in Expert Systems — Theory and Algorithms (John Wiley & Sons, New York, 1990). [19] K.G. Olesen, U. Kjaerulff, F. Jensen, F.V. Jensen, B. Falck, S. Andreassen and S.K. An- dersen, A MUNIN network for the median nerve — A case study on loops, Applied Artificial Intelligence 3 (1989) 301–319. [20] J. Pearl, Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference (Morgan Kaufmann Publishers, San Mateo, California, 1988). [21] R.G. Petersdorf, R.D. Adams, E. Braunwald, K.J. Isselbacher, J.B. Martin and J.D. Wil- son (Eds.), Harrison’s Principles of Internal Medicine, tenth edition (McGraw-Hill In- ternational Book Company, Auckland, 1983). [22] S. Schenker, J. Balint, L. Schiff, Differential diagnosis of jaundice: Report of a prospective study of 61 proved cases, American Journal of Digestive Diseases 7 (1962) 449–463. [23] R.W. Segaar, J.H.P. Wilson, J.D.F. Habbema, A. Malchow-Møller, J. Hilden and P.J. van der Maas, Transferring a diagnostic decision aid for jaundice, Netherlands Jour- nal of Medicine 33 (1988) 5–15. [24] E.H. Shortliffe, Computer-Based Medical Consultations: MYCIN (Elsevier, New York, 1976). [25] M.A. Shwe, B. Middleton, D.E. Heckerman, M. Henrion, E.J. Horvitz, H.P. Lehmann and G.F. Cooper, Probabilistic diagnosis using a reformulation of the INTERNIST-1/QMR knowledge base, I. The probabilistic model and inference algorithms, Methods of Information in Medicine 30 (1991) 241–255. [26] D.J. Spiegelhalter, R.C.G. Franklin and K. Bull, Assessment, criticism and improve- ment of imprecise subjective probabilities for a medical expert system, in: M. Henrion, R.D. Shachter, L.N. Kanal and J.F. Lemmer (Eds.), Uncertainty in Artificial Intelli- gence 5 (North-Holland, Amsterdam, 1990) 285–294. [27] S. Srinivas and J. Breese, IDEAL: Influence diagram evaluation and analysis in Lisp; documentation and users guide, Technical Memorandum No. 23, Rockwell International Science Center, Palo Alto Laboratory, Palo Alto, 1990. [28] R.B. Stern, R.P. Knill-Jones and R. Williams, Use of computer program for diagnosing jaundice in district hospitals and specialized liver unit, British Medical Journal 2 (1975) 659–662. REFERENCES 25 [29] I. Sternlieb and I.H. Scheinberg, Prevention of Wilson’s disease in asymptomatic patients, New England Journal of Medicine 278 (1968) 352–359. [30] I. Sternlieb and I.H. Scheinberg, Wilson’s disease’, in: [34] 949–961. [31] A. Theodossi, An assessment of the value of diagnostic techniques in hepatobiliary dis- ease, M.D. Thesis, University of London, London, 1986. [32] A. Theodossi, R.P. Knill-Jones, A. Skene, G. Lindberg, B. Bjerregaard, J. Holst- Christensen and R. Williams, Inter-observer variation of symptoms and signs in jaundice, Liver 1 (1981) 21–32. [33] A. Theodossi, D. Spiegelhalter, B. Portmann, A.L.W.F. Eddleston and R. Williams, The value of clinical, biochemical, ultrasound and liver biopsy data in assessing patients with liver disease, Liver 3 (1983) 315–326. [34] R. Wright, G.H. Millward-Sadler, K.G.M.M. Alberti and S. Karran (Eds.), Liver and Biliary Disease, second edition (W.B. Saunders Company, London, 1985). 
work_6fsgua7v2bf3bhu2cimvomab7m	----	Document downloaded from: This paper must be cited as: The final publication is available at Copyright http://dx.doi.org/10.1016/j.eswa.2012.01.111 http://hdl.handle.net/10251/50835 Elsevier Reynoso Meza, G.; Sanchís Saez, J.; Blasco Ferragud, FX.; Herrero Durá, JM. (2012). Multiobjective evolutionary algorithms for multivariable PI controller design. Expert Systems with Applications. 39(9):7895-7907. doi:10.1016/j.eswa.2012.01.111,. Multiobjective Evolutionary Algorithms for Multivariable PI Controller Design Gilberto Reynoso-Meza∗, Javier Sanchis, Xavier Blasco, Juan Herrero Grupo de Control Predictivo y Optimización Heuŕıstica (CPOH) Instituto Universitario de Automática e Informática Industrial Universitat Politècnica de València Camino de Vera 14, 46022 - Valencia, Spain http://ctl-predictivo.upv.es Abstract A multiobjective optimisation engineering design (MOED) methodology for PI controller tuning in multivariable processes is presented. The MOED pro- cedure is a natural approach for facing multiobjective problems where several requirements and specifications need to be fulfilled. An algorithm based on the differential evolution technique and spherical pruning is used for this purpose. To evaluate the methodology, a multivariable control benchmark is used. The obtained results validate the MOED procedure as a practical and useful technique for parametric controller tuning in multivariable processes. Keywords: multiobjective optimisation, controller tuning, pid tuning, multiobjective evolutionary optimisation, decision making. 1. Introduction PI and PID controllers currently represent a reliable digital control solu- tion because of their simplicity and efficacy [3]. They are often used in indus- trial applications and there is ongoing research on new techniques for robust tuning in single-input single-output (SISO) systems, as well as multiple-input multiple-output (MIMO) systems. MIMO systems are very common in in- ∗Corresponding author. Tel.:+34963877007;fax:+34963879579 Email address: gilreyme@posgrado.upv.es (Gilberto Reynoso-Meza) URL: http://ctl-predictivo.upv.es/ (Gilberto Reynoso-Meza) Preprint submitted to Expert Systems with Applications April 7, 2011 dustrial processes, and their complexity relies on the dynamic interaction between inputs and outputs. New PI-PID controller tuning techniques mainly search for a trade-off so- lution among several control and operational requirements. Some approaches state the design problem as an analytical/numerical optimisation procedure [14, 51, 15, 2, 36], or as an evolutionary optimisation statement [24, 22, 23]. In both cases, a variety of specifications with several requirements and spec- ifications must be faced. Such problems involving multiple objectives are known as multiobjective problems (MOP). In an MOP, the designer (control engineer) has to deal with a list of re- quirements and searches for a solution with a desired trade-off (preferences) among objectives. A traditional approach to handle preferences in an MOP is to translate it into a single-objective problem using weighting factors. More elaborate methods have been developed [30], such as goal programming, lex- icographic methods, physical programming [55], and recently, global physical programming [33, 45]. Multiobjective optimisation (MOO) can handle these issues in a simpler manner because of its simultaneous optimisation approach. In MOO, all of the objectives and constraints are significant from the designer’s point of view. Consequently, each is optimised to obtain a set of optimal non- dominated solutions. In this set of solutions, no solution is better than the others in every objective - but each solution offers different balances between design objectives. As a result, the decision maker (DM) can obtain a better insight into the trade-off for different solutions and can analyse the tendencies. This approach produces more information for selecting the most preferable solution that meets the DM’s preferences. The difficulty involved in the PI-PID tuning process based on optimisa- tion increases when: • MIMO systems are considered instead of SISO systems. • The number of engineering requirements (objectives) increases. • MOO is required instead of single objective optimisation. • Constrained problems are treated instead of unconstrained problems. It is, therefore, worthwhile searching for new algorithms and strategies to tackle constrained MOO for PI-PID tuning in multivariable processes. 2 Therefore, this paper proposes an MOP statement for constrained MIMO PI tuning that demonstrates its viability in an easy and intuitive way. This is fulfilled by defining the MOO statement with well-known performance indexes and a graphical visualisation of the Pareto front. This is a very im- portant issue since the DM requires a useful and interpretable approximation for the decision making stage. The remainder of this paper is organised as follows: in Section 2 a review of MOO is presented; in Section 3 a multiobjective optimisation engineer- ing design (MOED) methodology for multivariable PI controller tuning is explained. In Section 4 the MOED methodology is evaluated in a multivari- able benchmark process. Finally, some concluding remarks are given. 2. Multiobjective optimisation review An MOP, without loss of generality,1 can be stated as follows: min θ∈ℜn J(θ) = [J1(θ), . . . , Jm(θ)] ∈ ℜ m (1) where θ ∈ ℜn is defined as the decision vector and J(θ) as the objective vector (see Figure 1). A unique solution does not generally exist for an MOP because no solution is better than the others for all the objectives. Let ΘP be defined as the Pareto set, or set of solutions of the MOP, and JP be defined as the Pareto front or the projection of ΘP in the objective space. Each point in the Pareto front is said to be a non-dominated solution (see Figure 2). Definition 1. (Dominance relation): given a solution θ1 with objective vec- tor J(θ1) dominates a second solution θ2 with objective vector J(θ2) if and only if: {∀i ∈ [1, 2, . . . m], Ji(θ 1) ≤ Ji(θ 2)} ∧ {∃q ∈ [1, 2, . . . m] : Jq(θ 1) < Jq(θ 2)} which is denoted as θ1 ≺ θ2. 1A maximisation problem can be converted to a minimisation problem taking into account that max Ji(θ) = min(−Ji(θ)) is applied. 3 q 1 J2 q2 q3 J1 q 2 q 3 q 4 q1 E s p a c i o d e v a r i a b l e s d e d e c i s i ó n D E s p a c i o d e f u n c i o n e s o b j e t i v o J( )q 1 J( )q 2 J( )q 3 J( )q 4J(·) Á r e a d e d o m i n a n c i a Figure 1: Pareto set (left) and Pareto front (right). Objective vector J(θ4) is dominated by J(θ2) . Two useful vectors can be defined: the ideal solution Jmin and the nadir solution Jmax: Jideal = Jmin = [ min J(θ)∈J∗ P J1(θ), . . . , min J(θ)∈J∗ P Jm(θ) ] Jnadir = Jmax = [ max J(θ)∈J∗ P J1(θ), . . . , max J(θ)∈J∗ P Jm(θ) ] (2) MOO techniques search for a discrete approximation Θ∗ P of the Pareto set ΘP capable of generating a good quality description J ∗ P of the Pareto front JP (see Figure 3). In this way, the DM has a set of solutions for a given problem and more flexibility for choosing a particular or desired solution. There are several widely used algorithms for calculating this Pareto front approximation (normal boundary intersection method [7, 35], normal constraint method [1, 44, 32, 31], and the successive Pareto front optimisation [43]). Recently, multiobjective evolutionary algorithms (MOEAs) have been used due to their flexibility in dealing with non-convex and highly constrained functions [6, 5]. For this reason, MOEAs are considered in this work. A general framework is required to successfully incorporate the MOO approach into any engineering process. A multiobjective optimisation engi- neering design (MOED) methodology is shown in Figure 4. It consists in 4 Figure 2: Dominance concept. A given objective vector A dominates the objective vectors with a better or equal cost value in all objectives (with, at least, one of them being better). Two important points are defined: the ideal solution and the nadir solution (see Equation 2). 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.5 1 1.5 2 J1 J 2 True Pareto front (continous) J P Solutions describing a Pareto front approximation J P * Dominated Solutions (+) Figure 3: Pareto front concept (example of two objectives). Points are a possible Pareto front approximation obtained by a particular optimisation algorithm. 5 three main steps: MOP definition: at this stage the following are defined: the design concept (how to tackle the problem at hand); the engineering requirements (what it is important to optimise); and the constraints (which solutions are not practical/allowed). The design concept implies the existence of a parametric model that defines the parameter values (the decision space) that leads to a particular design alternative and its performance [34]. MOO process: at this stage, the MOO statement, as well as the MOEA, are defined. It is important to select an MOEA that assures reasonable diversity, spread, and convergence to the Pareto front and is an efficient constraint handling mechanism. Decision making stage: finally, with the calculated approximation J∗ P , the DM can analyse the trade-off along the Pareto front. The DM will select the best vector solution according to his/her needs. A reli- able tool or methodology is required for this final step, since it is not a trivial task to perform an analysis on m-dimensional Pareto fronts. Figure 4: Multiobjective optimisation engineering methodology. The MOED methodology for PI tuning the multivariable process is then defined. 6 3. Multiobjective optimisation engineering design applied to mul- tivariable PI controller tuning MIMO systems are common in industry. Their complexity is due to their coupling effects between inputs and outputs. Consider a NxN multivariable process modeled by the following transfer matrix: Gp(s) =    Gp11(s) . . . Gp1N(s) ... ... ... GpN1(s) . . . GpNN(s)    (3) The selected controller design concept must fulfill a set of requirements, in accordance with the given process. Common choices for controlling MIMO system are: decoupled PI-PID controllers [22]; centralised PI-PID controllers [23]; state space feedback techniques [13, 40]; or predictive control [12, 47, 28]. The selection of one technique over another depends on the desired balance between complexity and trade-off between design specifications. 3.1. MOP definition In accordance with Figure 4, the first step is to define the design technique (leading to the decision space definition), the operational constraints, and the objective space (optimisation objectives). Using MOEAs in the MOO process gives a greater flexibility to use any type of parametric controller and define any type of performance objective. In this work, a set of decoupled PI controllers is proposed to tackle the control problem in a MIMO system. PI controllers are a simple but successful solution, and they can be improved with complementary techniques (see [3]). Equation (4) shows the structure of the chosen PI controller: Gc(s) = kc ( 1 + 1 Tis ) E(s) (4) where kc is the proportional gain, Ti the integral time (secs), and E(s) the error signal. The decoupled PI controller Gc(s) design has N SISO controllers: Gc(s) =    Gc1(s) . . . 0 ... ... ... 0 . . . GcN(s)    (5) 7 Therefore, the decision space is defined as: θ = [kc1, Ti1, . . . , kcN, TiN] ∈ ℜ 2N (6) The non-convex optimisation developed by [2] will be used as guideline for the SISO PI controllers. This optimisation procedure is analytical and model oriented and does not require any time domain function computations (simulations). It defines a given value of the maximum sensitivity func- tion as a design constraint Ms = max ∣ ∣ ∣ 1 1+Gc(ω)Gp(ω) ∣ ∣ ∣ and/or the maximum complementary sensitivity function Mp = max ∣ ∣ ∣ Gc(ω) 1+Gc(ω)Gp(ω) ∣ ∣ ∣ . A numerical non-convex optimisation is then used, by increasing as much as possible the integral gain ki = kc/Ti subject to the values of Ms and Mp, to obtain a desired trade-off between load rejection and robustness. The previous tuning procedure can be adapted for MOEAs and defining as engineering control objectives ki, Ms and Mp. Such objectives give the DM some insight regarding the trade-off for robustness, load rejection, and set point response as in [2]. To apply this tuning procedure in a multivariable process, an index to measure the overall MIMO system stability is required. Here, the closed loop log modulus (Lcm) will be used as a robustness indicator. This index leads to the well-known largest log modulus (BLT) tuning criteria for diagonal PID controllers in MIMO processes [29]. The criteria is defined as: Lcm = 20 log ∣ ∣ ∣ ∣ W(s) 1 + W(s) ∣ ∣ ∣ ∣ (7) where W(s) = −1 + det (I + Gp(s)Gc(s)). Therefore, the MOP at hand is to find a trade-off solution θ, that is: J(θ) = [−ki1, Ms1, Mp1, . . . , −kiN, MsN, MpN, Lcm] ∈ ℜ 3N+1 (8) The objective vector as defined by Equation (8) does not guarantee to give the DM a useful Pareto front with a good degree of flexibility to select a reliable and practical solution. It is well-known that certain practical limits to Ms, Mp and Lcm values are needed to guarantee a minimum of stability margin. Therefore, the MOP statement must consider the following practical limits: 8 kc1 + ν1 · kc1/Ti1 ≤ Ku1 ... kcN + νN · kcN/TiN ≤ KuN 1.2 ≤ Ms1,...,N ≤ 2.0 1 ≤ Mp1,...,N ≤ 1.5 0 ≤ Lcm ≤ 2N (9) Where ν is the maximum value between the time delay process and 1. Constraint kc +ν ·kc/Ti ≤ Ku is used to bound the maximum allowed control action effort to the ultimate gain Ku. Constraints 1.2 ≤ Ms and 1 ≤ Mp are used to avoid controllers with a sluggish performance, while constraints Ms ≤ 2.0 and Mp ≤ 1.5 guarantee a minimum of stability margin [2]. The empirical rule of keep Lcm ≤ 2N [29] is accepted. 3.2. The MOO process As constraints are considered in the MOP, a constraint handling mecha- nism is used. According to the practical and empirical limits defined for J∗P by Equation (9), any unfeasible solution is punished. In [8], a penalty func- tion without penalty parameter is proposed. Such penalty function enforces the following criteria: 1. Any feasible solution is preferred to any infeasible solution. 2. Between two feasible solutions, the solution with the better objective function value is preferred. 3. Between two infeasible solutions, the solution with the smaller con- straint violation is preferred. Following these ideas, the MOO vector objective takes the form: min θ∈ℜ2N J(θ) =        J(θ) ∈ ℜ3N+1 if 7 ∑ k=1 φk(θ) = 0 offset + ( 7 ∑ k=1 φk(θ) ) · R ∈ ℜ3N+1 otherwise (10) 9 where: offset = max (Jmax) · R φ1(θ) = max{0, kc1 + ν1kc1 Ti1 − Ku1, . . . , kcN + νNkcN TiN − KuN} φ2(θ) = max{0, 1.2 − Ms1 . . . , 1.2 − MsN} φ3(θ) = max{0, 1.0 − Mp1 . . . , 1.0 − MpN} (11) φ4(θ) = max{0, Ms1 − 2.0, . . . , MsN − 2.0} φ5(θ) = max{0, Mp1 − 1.5, . . . , MpN − 1.5} φ6(θ) = max{0, Lcm − 2N} φ7(θ) = max{0, −Lcm} (12) and R is a vector of 1s with dimension 1x7. An extensive list of MOEAs are available for MOO. Some examples are NSGA-II [10], MOGA [11], ǫv-MOGA [19], paǫ-MyDE [17], sp-MODE [37], among others. They have been used for SISO PID tuning [20, 41, 42, 38, 50], as well for MIMO tuning [53]. In this case, the differential evolution (DE) algorithm [49, 48] was selected as the evolutionary technique. Multiobjective optimisation algorithms based on DE have shown a remarkable performance in a variety of multiobjective optimisation problems [21]. The DE algorithm has two main operators: mutation and crossover. Mutation: At generation k for each target (parent) vector θi|k, a mutant vector vi|k is generated according to Equation (13): v i|k = θ r1|k + F(θ r2|k − θ r3|k) (13) Where r1 6= r2 6= r3 6= i and F is known as the scaling factor. Crossover: For each target vector θi|k and its mutant vector v i|k, a trial (child) vector ui|k = [u i 1|k, u i 2|k, . . . , u i n|k] is created as follows: uij|k = { vij|k if rand(0, 1) ≤ Cr θij|k otherwise (14) where j ∈ 1, 2, 3 . . . n and Cr is the crossover probability rate. 10 For single objective optimisation, a child is selected over its parent (for the next generation) if it has a better cost value. In MOO, a child is selected over his parent if child ≺ parent. The DE by itself is only capable of evolving its population towards the Pareto front, and cannot improve diversity or spread along the Pareto front. As a result, it is necessary to use a mechanism to improve diversity. The use of an external archive is very common and widely accepted in MOEAs. It consists in using an evolutive population plus an external archive where quality solutions are stored and usually used in the evolutionary strat- egy itself. Several techniques have been used to maintain diversity and spread solutions in this archive. In [27], a relaxed form of Pareto dominance, known as ǫ−dominance, is proposed. The main idea is to use an archiving strategy, where instead of using the classical relation for dominance (Definition 1), the following ideas are used: • A solution dominates the solutions that are less fit for all the objectives. • A solution dominates the solutions inside a distance that is less than a parameter ǫ (ǫ-dominance concept). Algorithms based on this approach include the ǫ−MOEA [9], ǫ−MyDE [46], ǫv−MOGA [18, 19], paǫ−MyDE [17], and paǫ−ODEMO [16]. Pruning techniques are commonly based on using some kind of measurement to select individuals in less crowded areas. The crowding distance from [10] or a based measurement are used in algorithms such NSGAII [10] or GDE3 [26]. Nevertheless, it is worthwhile to look for new techniques to deal with any geometrical characteristic of the Pareto front, such as concavity, convexity, disconnected segments, or mixed characteristics. In this work, a pruning technique is employed, based on spherical relations in the objective space [39]. The technique shows a good flexibility in dealing with diverse geometries in m-dimensional Pareto fronts and achieves a well-spread set of solutions. The basic idea of the spherical relations is to analyse the proposed solu- tions in the current Pareto front approximation by using normalised spherical coordinates from a reference solution (see Figure 5a). The spherical pruning works over a non-dominated set of solutions. This can be interpreted as if the designer stands at the ideal solution (or any desired solution) with a given direction in the objective space. The DM will then be searching for the closest non-dominated solution (Figure 5b). 11 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 J 1 J 2 J 3 (a) Spherical relations (b) Set of solutions  → (c) Dominance on  → (d) Spherical pruning on  Figure 5: Spherical relations (up) and spherical pruning (down) for MOO problems (‖ · ‖2 case). Extreme solutions are always preserved. 12 The spherical relations are compatible with any MOO algorithm or evolu- tionary technique and will be merged together with the DE algorithm. Such an approach, presented as sp-MODE in previous works, has been shown to be effective in addressing control engineering problems [37, 42, 40]. 3.3. Decision making stage An m-dimensional Pareto front J∗P is difficult to analyse without an ef- fective visualization tool. If J∗P is not clearly displayed to the DM, it will be complicated to perform a practical analysis in the Pareto front approxima- tion, or select a solution with a desired trade-off. The graphical visualisation is not a trivial task when the number of objectives is more than three. The level diagram (LD) tool [4] is used because it is flexible in performing a useful analysis on the obtained Pareto front J∗P . LDs are based on the classification of the J∗P approximation obtained. Each objective Jq(θ) is normalised with respect to its minimum and maximum values. That is: Ĵq(θ) = Jq(θ) − J min q Jmaxq − J min q , q ∈ [1, . . . , m]. (15) On each normalised objective vector Ĵ(θ) a p-norm ‖x‖p is applied to evaluate the distance to an ideal solution Jideal ≈ Jmin 2. The LD tool displays a two-dimensional graph for every objective and every decision vari- able. The ordered pairs ( Jq(θ), ‖Ĵ(θ)‖p ) in each objective sub-graph and ( θl, ‖Ĵ(θ)‖p ) , l ∈ {1, 2, . . . , n} in each decision variable sub-graph are plot- ted. Therefore, a given solution will have the same y-value in every graphic. This correspondence will help to evaluate general tendencies along the Pareto front and compare solutions in accordance with the selected norm. A deeper explanation of the LD tool capabilities can be found in [4]. A MIMO benchmark will next be considered to validate the MOED for the multivariable PI controller tuning defined in this work. 4. Procedure validation To show the applicability of the MOED proposal for multivariable PI tuning, the well-known distillation column model defined by Wood and Berry will be used [52]: 2Since Jideal is not always available 13 Gp(s) =   Gp11(s) Gp12(s) Gp21(s) Gp22(s)   =   12.8e−s 16.7s+1 −18.9e−3s 21s+1 6.6e−7s 10.9s+1 −19.4e−3s 14.4s+1   (16) As mentioned earlier, any kind of parametric controller can be tuned with the MOED methodology, but for comparison purposes two PI controllers will be used: Gc(s) =     kc1 ( 1 + 1 Ti1s ) 0 0 kc2 ( 1 + 1 Ti2s )     (17) 4.1. Engineering design process Given equations (16) and (17), the MOP at hand is to find a trade-off solution θ = [kc1, Ti1 kc2, Ti2] for the design objectives: J(θ) = [−kc1/Ti1, Ms1, Mp1, −kc2/Ti2, Ms2, Mp2, Lcm] (18) subject to: kc1 + kc1/Ti1 ≤ Ku1 ≈ 2.0 |kc2 + 3kc2/Ti2| ≤ |Ku2| ≈ | − 0.42| 1.2 ≤ Ms1,2 ≤ 2.0 (19) 1 ≤ Mp1,2 ≤ 1.5 0 ≤ Lcm ≤ 4 4.2. Multiobjective optimisation process The MOO objective vector shown in Table 1 is in accordance with Equa- tion (10). The optimisation process is performed with three different MOEAs: • A DE algorithm without archiving strategy (NA); namely, a child will be selected over his parent if child ≺ parent. Parameter values F = 0.5, Cr = 0.8 are used (which are standard parameters in accordance with [48]) and an initial population of 50 random decision vectors. 14 • A DE algorithm with spherical pruning (sp-MODE). Parameter values F = 0.5, Cr = 0.8, a population of 50 solutions, and a spherical grid resolution of 5 are used. • The gamultiobj algorithm provided by MatLab c© is used to calculate a Pareto front for reference. This algorithm uses a controlled elitist genetic algorithm (a variant of NSGA-II [10]). Diversity is maintained by controlling the elite members of the population as the algorithm progresses by using a crowding distance index. Default parameters are used and the BLT solution [29] is used in its initial population. The maximum allowable function evaluations (FEs) for each method is bound to 6000, and 25 independent runs will be evaluated to analyse their performance. Each execution from the sp-MODE and the NA strategy will be compared with the Pareto front J∗P |GA built with the executions of the gamultiobj algorithm. To evaluate the performance of each MOEA, the Iǫ binary indicator [54, 25] is used. The indicator indicates the factor Iǫ(A, B) by which an approx- imation set A is worse than another set B with respect to all the objectives. Using a comparison method (see Table 2) CIǫ,E(A, B) = E(Iǫ(A, B), Iǫ(B, A)) = {false, true} the Eps binary indicator is a compatible and complete oper- ator 3 and this is useful to determine if two Pareto fronts are incomparable, equal, or if one is better than the other [54]. The optimisation experiments were carried on an a standard PC, with a Pentium(R) processor running at 3.40 GHz and 2 GB RAM. The results after 25 independent trials with each proposal are shown in Table 3 (performance indicators) and Table 4 (non-dominated solutions attained). As evidenced by the given results, the sp-MODE algorithm represents a viable approach for generating the Pareto front. The sp-MODE algorithm outperforms the gamultiobj algorithm, since Iǫ(sp − MODE, GA) < 1. Be- sides, the sp-MODE algorithm has a better improvement over J∗P |GA than the NA-strategy (Iǫ(sp − MODE, GA) < Iǫ(NA, GA)). 3Given a binary relation on approximation sets (·), the comparison method is com- patible if CI,E(A, B) → A · B ∨ B · A. However, the comparison method is complete if A · B ∨ A · B → CI,E(A, B). 15 min θ∈ℜ4 J(θ) =        J(θ) ∈ ℜ7 if 7 ∑ k=1 φk(θ) = 0 offset + 7 ∑ k=1 φk(θ) · R ∈ ℜ 7 otherwise J(θ) = [J1(θ), J2(θ), J3(θ), J4(θ), J5(θ), J6(θ), J7(θ)] θ = [kc1, Ti1, kc2, Ti2] J1(θ) = −ki1 = −kc1/Ti1 Kc1 ∈ [0.001, Ku1] J2(θ) = Ms1 = max ∣ ∣ ∣ 1 1+Gc1(ω)Gp11(ω) ∣ ∣ ∣ Kc2 ∈ [Ku2, −0.001] J3(θ) = Mp1 = max ∣ ∣ ∣ Gc1(ω) 1+Gc1(ω)Gp11(ω) ∣ ∣ ∣ Ti1,i2 ∈ [0.001, 40] J4(θ) = −ki2 = −kc2/Ti2 J5(θ) = Ms2 = max ∣ ∣ ∣ 1 1+Gc2(ω)Gp22(ω) ∣ ∣ ∣ J6(θ) = Mp2 = max ∣ ∣ ∣ Gc2(ω) 1+Gc2(ω)Gp22(ω) ∣ ∣ ∣ J7(θ) = Lcm = 20 log ∣ ∣ ∣ W (S) 1+W (S) ∣ ∣ ∣ , W(s) = −1 det (I + Gp(s)Gc(s)) offset = max(Jnadir) ∗ R φ1(θ) = max (0, kc1 + kc1/Ti1 − 2.1, |kc2 + 3kc2/Ti2| − 0.42) φ2(θ) = max (0, 1.2 − Ms1, 1.2 − Ms2) φ3(θ) = max (0, 1.0 − Mp1, 1.0 − Mp2) φ4(θ) = max (0, Ms1 − 2.0, Ms2 − 2.0) φ5(θ) = max (0, Mp1 − 1.5, Mp2 − 1.5) φ6(θ) = max{0, Lcm − 4.0} φ7(θ) = max{0, −Lcm} Table 1: MOO statement for the multivariable PID controller approach. 16 Iǫ(A, B) < 1 → Every J(θ b) ∈ B is strictly dominated by at least one J(θa) ∈ A. Iǫ(A, B) ≤ 1 ∧ Iǫ(B, A) > 1 → Every J(θ b) ∈ B is weakly dominated by at least one J(θa) ∈ A and A 6= B. Iǫ(A, B) ≤ 1 → Every J(θ b) ∈ B is weakly dominated by at least one J(θa) ∈ A. Iǫ(A, B) = 1 ∧ Iǫ(B, A) = 1 → A = B. Iǫ(A, B) > 1 ∧ Iǫ(B, A) > 1 → Neither A weakly dominates B nor B weakly dominates A. Table 2: Comparison methods using the Eps indicator. Eps Indicator Iǫ(NA, GA) Iǫ(sp − MODE, GA) Worst 2.34E-001 1.01E-001 Pareto Best 7.74E-002 4.69E-002 Set Median 1.07E-001 7.55E-003 Reference Mean 1.19E-001 7.53E-002 (676 Sol) Std 3.32E-002 1.44E-002 Table 3: Performance achieved by MOEAs. NA sp-MODE Worst 0 109 Best 48 243 Median 43 153 Mean 4.03E+001 1.56E+002 Std 9.15E+000 3.09E+001 Table 4: Number of solutions achieved by MOEAs. 17 kc1 Ti1 kc2 Ti2 J1(θ) J2(θ) J3(θ) J4(θ) J5(θ) J6(θ) J7(θ) BLT [29] 0.3750 8.2900 -0.0750 23.6000 -0.0452 1.2953 1.1081 -0.0032 1.2513 0.9998 3.8599 WIB [22] 0.8485 326.3462 -0.0132 1.9130 -0.0026 1.6663 1.0178 -0.0069 2.0569 1.7259 0.6024 min ‖Ĵ(θ)‖2 0.4245 15.6135 -0.0397 7.0977 -0.0272 1.3090 1.0014 -0.0056 1.3090 1.1047 1.5054 min ‖Ĵ(θ)‖1 0.3351 34.1079 -0.0476 8.9239 -0.0098 1.2263 1.0000 -0.0053 1.2496 1.0427 0.7488 min ‖Ĵ(θ)‖∞ 0.7415 11.2697 -0.0431 5.4571 -0.0657 1.6220 1.0809 -0.0079 1.5097 1.2914 2.1913 J7(θ) = 2.9922 0.7687 6.9516 -0.0408 5.1598 -0.1106 1.6989 1.2144 -0.0079 1.5316 1.3170 2.9922 J7(θ) = 3.4956 0.8458 12.4453 -0.0858 17.6735 -0.0680 1.7434 1.1414 -0.0049 1.3092 1.0000 3.4956 J7(θ) = 3.995 0.92489 8.7357 -0.0783 5.8147 -0.1059 1.8880 1.2790 -0.0135 1.6731 1.4436 3.9950 Table 5: Controllers selected for further evaluation. 4.3. Decision making stage To validate the MOED method approach as a competitive and practical solution for controller tuning, the Pareto front with 153 solutions (median value in Table 4) is selected and used for controller evaluation.4 In Figure 6 the Pareto set and Pareto front using the LD tool are shown respectively. The controller selection procedure lies on the DM preferences and desired specifications. To illustrate the tradeoff achieved by different solutions, six controllers Gc(s) were selected from the Pareto front for further evaluation (see Table 5). The controllers with the lowest ‖Ĵ(θ)‖1, ‖Ĵ(θ)‖2 and ‖Ĵ(θ)‖∞ norm are selected. An overall trade-off between objectives is expected using these controllers. The remaining controllers are selected according to DM preferences. Let’s assume, for example, that the DM is interested in controllers over the A-line in objective J7(θ) (see Figure 6b) and decides to perform a further analysis on three controller from such a geometric locus. The controller resulting from the BLT tuning [29] (oriented to MIMO- stability using the Ziegler-Nichols procedure), as well as the controller pro- posed in [22] (WIB) that minimises the integral of the absolute error for a specific test, are finally included. These controllers will be tested in the multivariable model in three dif- ferent experiments: 1. Set point change in controlled Variable 1; consequently, the perfor- mance to reject the disturbance in controlled Variable 2 is evaluated. 2. Set point change in controlled Variable 2; consequently, the perfor- mance to reject disturbance in controlled Variable 1 is evaluated. 4The best solution attained could be used for this analysis, but this will not be entirely realistic, since it is not always possible to run an optimisation algorithm several times. 18 0 0.2 0.4 0.6 0.8 1 0.6 0.7 0.8 0.9 1 1.1 θ 1 : Proportional Gain 1 | |J (θ )| | ∞ 0 5 10 15 20 25 30 35 40 0.6 0.7 0.8 0.9 1 1.1 θ 2 : Integral Time 1 (secs) −0.12 −0.1 −0.08 −0.06 −0.04 −0.02 0 0.6 0.7 0.8 0.9 1 1.1 θ 3 : Proportional Gain 2 | |J (θ )| | ∞ 0 5 10 15 20 25 30 35 40 0.6 0.7 0.8 0.9 1 1.1 θ 4 : Integral Time 2 (secs) (a) Θ∗P −0.15 −0.1 −0.05 0 0.6 0.8 1 J 1 : Integral Gain (−−−) | |J (θ )| | ∞ −0.015 −0.01 −0.005 0 0.6 0.8 1 J 4 : Integral Gain (−−−) 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 0.6 0.8 1 J 2 : Sensitivity Function | |J (θ )| | ∞ 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 0.6 0.8 1 J 5 : Sensitivity Function 1 1.1 1.2 1.3 1.4 1.5 0.6 0.8 1 J 3 : Complementary Sensitivity Function | |J (θ )| | ∞ 1 1.1 1.2 1.3 1.4 1.5 0.6 0.8 1 J 6 : Complementary Sensitivity Function 0 0.5 1 1.5 2 2.5 3 3.5 4 0.6 0.8 1 J 7 : Biggest Log Modulus | |J (θ )| | ∞ A (b) J∗P Figure 6: Pareto set and Pareto front used for analysis. ‖Ĵ(θ)‖∞ norm is used. A-line indicates controllers that match a hypothetical preference of the DM. 19 3. Simultaneous set point change in both controlled variables. In all cases, the integral of the absolute error (IAE), the integral of the absolute derivative of the control action (IADU), the settling time (ST) at ±2%, the rise time (RT) from 10% to 90%, the maximum deviation (MD), and the overshoot (OS) will be evaluated. In Table 6 and in Figures 7,8 and 9 the obtained results for each controller are shown. Some expected behaviours are noted: • For controllers in the A-line (see J7 at Figure 6b) the greater the Lcm, the greater the control action and the worse are the trade-offs. That is evident since such controllers are incapable of performing well in all the experiments. Notice how these controllers become more oscillating as J7(θ) increases (Figures 7b, 8b, and 9b). • Controller WIB obtains the best value in IAE for Experiment 3; this was expected since this controller was tuned to minimise IAE for the same experiment. Notice how this outstanding performance has a lower trade-off when single set-point change in controlled variable 1 is applied (Figures 7a and 8a). • Controllers with min ‖Ĵ(θ)‖2, min ‖Ĵ(θ)‖1 and min ‖Ĵ(θ)‖∞ have a balanced trade-off between objectives, achieving good overall perfor- mance (Figures 7b, 8b, and 9b). It is important to remark that there are no bad controllers, just controllers with different trade-offs between objectives. As we can see, performances dif- fer. This analysis could assist in scheduling strategies where more than one controller is used. As a final remark, it can be noticed that operational aspects such saturation, initial states, and operational ranges are not con- sidered. MOEA flexibility allows the use of time function computations to incorporate operational aspects and re-define the MOO statement with more meaningful objectives. 5. Conclusions In this work, an MOED methodology for multivariable PI controller tun- ing has been presented. The obtained results validate the methodology as a practical approach. Thanks to the visualisation capabilities of the LD tool, 20 it is easier to perform the controller selection procedure. As the simulations reveal, the MOO approach is validated as a useful tool for control purposes. With this approach, most of the optimisation procedure uses classical control techniques supported with well-known performance objectives. The continuous use of these objectives by the control engineer community ensures practical bounds and quick interpretations for selecting suitable controllers. The Pareto front enables us to have a better insight into the objective trade- off and how it changes between solutions. The MOP definition for the Wood and Berry distillation column will allow further comparisons of MOEA performance. This MOP provides a useful multiobjective constrained problem for controller tuning in the multivariable process, and will help focus these algorithms for a specific class of engineering design problem. Acknowledgment This work was partially supported by the FPI-2010/19 grant from the Universitat Politècnica de València and project DPI2008-02133/DPI from the Spanish Ministry of Science and Innovation. [1] A. Messac, A. Ismail-Yahaya, C. M., 2003. The normalized normal con- straint method for generating the pareto frontier. Structural and multi- disciplinary optimisation (25), 86 – 98. [2] Astrom, K., Panagopoulos, H., Hagglund, T., 1998. Design of pi con- trollers based on non-convex optimisation. Automatica 34 (5), 585 – 601. [3] Åström, K. J., Hägglund, T., 2005. Advanced PID Control. ISA - The Instrumentation, Systems, and Automation Society, Research Triangle Park, NC 27709. [4] Blasco, X., Herrero, J., Sanchis, J., Mart́ınez, M., 2008. A new graphical visualization of n-dimensional pareto front for decision-making in mul- tiobjective optimisation. Information Sciences 178 (20), 3908 – 3924. [5] Coello, C. A. C., Lamont, G. B., 2004. Applications of Multi-Objective evolutionary algorithms, adavances in natural computation vol. 1 Edi- tion. World Scientific Publishing. 21 IAE IADU ST RT MD OS BLT 4.54E+002 8.71E-001 2.29E+001 3.68E+000 – 10.38% Unit Step WIB 2.48E+003 2.02E+000 +2.00E+002 1.39E+000 – 7.24% Reference min ‖Ĵ(θ)‖2 5.78E+002 8.99E-001 5.81E+001 4.02E+000 – 0.24% Y1 min ‖Ĵ(θ)‖1 1.57E+003 6.86E-001 1.40E+002 5.53E+001 – 0.00% min ‖Ĵ(θ)‖∞ 3.34E+002 1.76E+000 3.27E+001 1.50E+000 – 13.78% Y1 J7(θ) = 2.9922 3.32E+002 1.97E+000 2.18E+001 1.35E+000 – 24.43% J7(θ) = 3.4956 3.18E+002 2.28E+000 2.90E+001 1.27E+000 – 29.11% J7(θ) = 3.995 3.16E+002 2.74E+000 1.99E+001 1.11E+000 – 30.04% BLT 1.65E+003 1.02E-001 1.38E+002 — 6.70E-001 — Unit Step WIB 1.01E+003 7.02E-002 6.78E+001 — 8.07E-001 — Reference min ‖Ĵ(θ)‖2 9.54E+002 5.34E-002 5.12E+001 — 6.48E-001 — Y1 min ‖Ĵ(θ)‖1 9.92E+002 5.43E-002 7.00E+001 — 5.36E-001 — min ‖Ĵ(θ)‖∞ 8.20E+002 6.97E-002 5.93E+001 — 8.47E-001 — Y2 J7(θ) = 2.9922 8.59E+002 7.35E-002 6.35E+001 — 9.27E-001 — J7(θ) = 3.4956 1.10E+003 1.63E-001 8.40E+001 — 8.94E-001 — J7(θ) = 3.995 6.44E+002 1.72E-001 5.27E+001 — 9.79E-001 — BLT 3.38E+002 1.92E-001 4.58E+001 — 1.82E-001 — Unit Step WIB 4.22E+003 1.58E-001 +2.00E+002 — 1.49E-001 — Reference min ‖Ĵ(θ)‖2 5.63E+002 1.53E-001 7.27E+001 — 1.36E-001 — Y2 min ‖Ĵ(θ)‖1 1.56E+003 1.53E-001 1.57E+002 — 1.89E-001 — min ‖Ĵ(θ)‖∞ 2.32E+002 1.54E-001 4.03E+001 — 9.63E-002 — J7(θ) = 2.9922 1.40E+002 1.57E-001 2.86E+001 — 8.68E-002 — Y1 J7(θ) = 3.4956 2.25E+002 2.62E-001 3.85E+001 — 1.39E-001 — J7(θ) = 3.995 1.47E+002 2.42E-001 2.61E+001 — 1.44E-001 — BLT 3.26E+003 1.63E-001 1.73E+002 8.58E+001 — 0.00% Unit Step WIB 1.80E+003 1.24E-001 6.85E+001 2.05E+001 — 6.19% Reference min ‖Ĵ(θ)‖2 1.85E+003 1.04E-001 6.78E+001 3.59E+001 — 0.00% Y2 min ‖Ĵ(θ)‖1 1.94E+003 1.04E-001 9.48E+001 3.72E+001 — 0.00% min ‖Ĵ(θ)‖∞ 1.38E+003 1.10E-001 3.84E+001 2.34E+001 — 1.33% J7(θ) = 2.9922 1.43E+003 1.14E-001 5.91E+001 2.35E+001 — 2.11% Y2 J7(θ) = 3.4956 2.13E+003 1.84E-001 1.11E+002 5.19E+001 — 0.00% J7(θ) = 3.995 9.00E+002 1.80E-001 4.75E+001 6.62E+000 — 2.57% BLT 5.75E+002 1.01E+000 3.23E+001 2.98E+000 – 26.57% Unit Step WIB 2.35E+002 1.95E+000 8.33E+000 1.39E+000 – 8.04% Reference min ‖Ĵ(θ)‖2 5.41E+002 9.57E-001 5.03E+001 3.28E+000 – 13.84% Y1,Y2 min ‖Ĵ(θ)‖1 6.35E+002 7.30E-001 8.34E+001 4.68E+000 – 7.36% min ‖Ĵ(θ)‖∞ 3.74E+002 1.78E+000 2.55E+001 1.50E+000 – 19.39% Y1 J7(θ) = 2.9922 3.67E+002 1.99E+000 1.30E+001 1.35E+000 – 29.46% J7(θ) = 3.4956 3.94E+002 2.33E+000 2.50E+001 1.27E+000 – 26.38% J7(θ) = 3.995 3.86E+002 2.74E+000 2.53E+001 1.11E+000 – 35.27% BLT 1.81E+003 2.37E-001 2.01E+001 4.77E+000 – 29.61% Unit Step WIB 7.97E+002 6.91E-002 1.73E+001 2.23E+000 – 6.91% Reference min ‖Ĵ(θ)‖2 1.10E+003 1.19E-001 1.83E+001 4.59E+000 – 11.94% Y1,Y2 min ‖Ĵ(θ)‖1 1.07E+003 1.23E-001 1.80E+001 5.28E+000 – 12.28% min ‖Ĵ(θ)‖∞ 1.01E+003 1.48E-001 1.72E+001 3.61E+000 – 37.50% Y2 J7(θ) = 2.9922 1.04E+003 1.51E-001 1.67E+001 3.49E+000 – 43.74% J7(θ) = 3.4956 1.42E+003 3.31E-001 1.61E+001 3.89E+000 – 55.55% J7(θ) = 3.995 1.15E+003 3.32E-001 2.97E+001 3.68E+000 – 79.98% Table 6: Controller performance in experimental setup. In bold appears the best value and in italics the worst value in each experiment. 22 0 50 100 150 200 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Y 1 0 50 100 150 200 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time U 1 0 50 100 150 200 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 Y 2 0 50 100 150 200 0 0.01 0.02 0.03 0.04 0.05 0.06 Time U 2 BLT WIB min 2−norm min 1−norm (a) 0 50 100 150 200 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Y 1 0 50 100 150 200 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time U 1 0 50 100 150 200 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 Y 2 0 50 100 150 200 0 0.02 0.04 0.06 0.08 0.1 0.12 Time U 2 min inf−norm J7=2.9922 J7=3.4956 J7=3.995 0 5 10 15 20 0.9 1 1.1 1.2 1.3 (b) Figure 7: Performance in Experiment 1. 23 0 100 200 300 400 500 0 0.05 0.1 0.15 0.2 Y 1 0 100 200 300 400 500 −0.15 −0.1 −0.05 0 Time U 1 0 100 200 300 400 500 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Y 2 0 100 200 300 400 500 −0.12 −0.1 −0.08 −0.06 −0.04 −0.02 0 Time U 2 BLT WIB min 2−norm min 1−norm (a) 0 100 200 300 400 500 −0.05 0 0.05 0.1 0.15 Y 1 0 100 200 300 400 500 −0.2 −0.15 −0.1 −0.05 0 Time U 1 0 100 200 300 400 500 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Y 2 0 100 200 300 400 500 −0.14 −0.12 −0.1 −0.08 −0.06 −0.04 Time U 2 min inf−norm J7=2.9922 J7=3.4956 J7=3.995 0 10 20 30 40 0 0.05 0.1 (b) Figure 8: Performance in Experiment 2. 24 0 50 100 150 200 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Y 1 0 50 100 150 200 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time U 1 0 50 100 150 200 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Y 2 0 50 100 150 200 −0.1 −0.08 −0.06 −0.04 −0.02 0 0.02 Time U 2 BLT WIB min 2−norm min 1−norm 0 5 10 15 20 0 0.5 1 1.5 (a) 0 50 100 150 200 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Y 1 0 50 100 150 200 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time U 1 0 50 100 150 200 0 0.5 1 1.5 2 Y 2 0 50 100 150 200 −0.15 −0.1 −0.05 0 0.05 Time U 2 min inf−norm J7=2.9922 J7=3.4956 J7=3.995 0 5 10 15 20 0.9 1 1.1 1.2 1.3 1.4 (b) Figure 9: Performance in Experiment 3. 25 [6] Coello, C. A. C., Veldhuizen, D. V., G.B. Lamont, e., 2002. Evolutionary algorithms for solving multi-objective problems. Kluwer Academic Press. [7] Das, I., Dennis, J., 1998. Normal-boundary intersection: a new method for generating the Pareto surface in non-linear multicriteria optimisation problems. SIAM J. Optim. 8, 631 – 657. [8] Deb, K., 2000. An efficient constraint handling method for genetic al- gorithms. Computer Methods in Applied Mechanics and Engineering 186 (2-4), 311 – 338. [9] Deb, K., Mohan, M., Mishra, S., 2003. Towards a quick computation of well spread Pareto-optimal solutions. In: Fonseca, C., Fleming, P., Zitzler, E., Deb, K., editors., L. T. (Eds.), Evolutionary Multi-criterion optimisation: Second International Conference, EMO. pp. 222 – 236. [10] Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multiobjective genetic algorithm: nsga − ii. IEEE Transactions on Evolutionary Computation Vol. 6 (2), 124 – 141. [11] Fonseca, C., Fleming, P., 1993. Genetic algorithms for multiobjective optimisation: formulation, discussion an generalization. Proceedings of the Fifth International Conference on Genetic Algorithms., 416 – 423. [12] Garćıa-Nieto, S., Mart́ınez, M., Blasco, X., Sanchis, J., 2008. Nonlinear predictive control based on local model networks for air management in diesel engines. Control Engineering Practice 16 (12), 1399 – 1413. [13] Garcia-Alvarado, M., Ruiz-Lopez, I., 2010. A design method for robust and quadratic optimal mimo linear controllers. Chemical Engineering Science 65 (11), 3431 – 3438. [14] Ge, M., Chiu, M.-S., Wang, Q.-G., 2002. Robust pid controller design via lmi approach. Journal of Process Control (12), 3 – 13. [15] Goncalves, E. N., Palhares, R. M., Takahashi, R. H., 2008. A novel approach for h2/h∞ robust pid synthesis for uncertain systems. Journal of process control (18), 19 – 26. [16] Gong, W., Cai, Z., Zhu, L., 2009. An efficient multiobjective differential evolution algorithm for engineering design. Structural and Multidisci- plinary Optimisation 38, 137 – 157, 10.1007/s00158-008-0269-9. 26 [17] Hernández-Dı́az, A. G., Snatana-Quintero, L. V., Coello, C. A. C., Molina, J., 2007. Pareto-adaptive ǫ-dominance. Evolutionary Compu- tation (4), 493 – 517. [18] Herrero, J., Blasco, X., Mart́ınez, M., Ramos, C., 2005. Nonlinear robust identification using multiobjective evolutionary algorithms. In: Mira, J., Álvarez, J. (Eds.), Artificial Intelligence and Knowledge Engineering Ap- plications: A Bioinspired Approach. Vol. LNCS 3562. Springer-Verlag, pp. 231 – 241. [19] Herrero, J., Mart́ınez, M., Sanchis, J., Blasco, X., 2007. Well-distributed Pareto front by using the epsilon-moga evolutionary algorithm. In: et al., F. S. (Ed.), Computational and Ambient Intelligence. Vol. LNCS 4507. Springer-Verlag, pp. 292 – 299. [20] Herreros, A., Baeyens, E., Perán, J. R., 2002. Design of pid-type con- trollers using multiobjective genetic algorithms. ISA Transactions 41 (4), 457 – 472. [21] Huang, V., Qin, A., Deb, K., Zitzler, E., Suganthan, P., Liang, J., Preuss, M., Huband, S., 2007. Problem definitions for perfor- mance assessment on multi-objective optimisation algorithms. Tech. rep., Nanyang Technological University, Singapore. [22] Iruthayarajan, M. W., Baskar, S., 2009. Evolutionary algorithms based design of multivariable pid controller. Expert Systems with Applications 36 (5), 9159 – 9167. [23] Iruthayarajan, M. W., Bskar, S., 2010. Covariance matrix adaptation evolution strategy based design of centralized pid controller. Expert Sys- tems with Applications 37 (8), 5775 – 5781. [24] Kim, T.-H., Maruta, I., Sugie, T., 2008. Robust pid controller tun- ing based on the constrained particle swarm optimisation. Automatica 44 (4), 1104 – 1110. [25] Knowles, J., Thiele, L., Zitzler., E., Feb. 2006. A tutorial on the per- formance assessment of stochastic multiobjective optimizers. Tech. Rep. TIK report No. 214, Computer Engineering and Networks Laboratory. ETH Zurich. 27 [26] Kukkonen, S., Lampinen., J., 2005. Gde3: The third evolution step on generalized differential evolution. Vol. 1. IEEE Congress on Evolutionary Computation (CEC 2005), pp. 443 – 450. [27] Laumanns, M., Thiele, L., Deb, K., Zitzler, E., 2002. Combining conver- gence and diversity in evolutionary multiobjective optimisation. Evolu- tionary Computation (3), 263 – 282. [28] Lauŕı, D., Salcedo, J., Garćıa-Nieto, S., Mart́ınez, M., 2010. Model pre- dictive control relevant identification: multiple input multiple output against multiple input single output. IET Control Theory & Applica- tions 4 (9), 1756–1766. [29] Luyben, W. L., 1986. Simple method for tuning siso controllers in mul- tivariable systems. Industrial and Engineering Chemistry Process De- sign (25), 654 – 660. [30] Marler, R., Arora, J., 2004. Survey of multi-objective optimisation meth- ods for engineering. Structural and Multidisciplinary Optimisation (26), 369 – 395. [31] Mart́ınez, M., Herrero, J., Sanchis, J., Blasco, X., Garćıa-Nieto, S., 2009. Applied pareto multi-objective optimisation by stochastic solvers. Engineering Applications of Artificial Intelligence 22, 455 – 465. [32] Mart́ınez, M., Nieto, S. G., Sanchis, J., Blasco, X., 2009. Genetic al- gorithms optimisation for normalized normal constraint method under Pareto construction. Advances in Engineering Software 40, 260 – 267. [33] Mart́ınez, M., Sanchis, J., Blasco, X., 2006. Multiobjective controller de- sign handling human preferences. Engineering Applications of Artificial Intelligence 19, 927 – 938. [34] Mattson, C. A., Messac, A., 2005. Pareto frontier based concept selection under uncertainty, with visualization. Optimisation and Engineering 6, 85 – 115, 10.1023/B:OPTE.0000048538.35456.45. [35] Miettinen, K. M., 1998. Nonlinear multiobjective optimisation. Kluwer Academic Publishers. 28 [36] Panagopoulos, H., Astrom, K., Hagglund, T., Jan. 2002. Design of pid controllers based on constrained optimisation. Control Theory and Ap- plications, IEE Proceedings - 149 (1), 32 – 40. [37] Reynoso-Meza, G., 2009. Design, coding and implementation of a mul- tiobjective optimisation algorithm based on differential evolution with spherical pruning: applications for system identification and controller tuning. Master’s thesis, Universidad Politécnica de Valencia. URL http://personales.alumno.upv.es/gilreyme [38] Reynoso-Meza, G., Blasco, X., Sanchis, J., 2009. Diseño multiobjetivo de controladores pid para el benchmark de control 2008-2009. Revista Iberoamericana de Automática e Informática Industrial 6 (4), 93 – 103. [39] Reynoso-Meza, G., Blasco, X., Sanchis, J., Herrero, J., (Submitted). Spherical relations to achieve diversity in multiobjective evolutionary algorithms. Evolutionary Computation. [40] Reynoso-Meza, G., Garćıa-Nieto, S., Sanchis, J., Blasco, X., (Submit- ted). Controller tuning using multiobjective optimisation algorithms: a global tuning framework. Control Engineering Practice. [41] Reynoso-Meza, G., Sanchis, J., Blasco, X., 30 November - 2 Decem- ber 2009. Multiobjective design of a digital controller for the throttle control benchmark. In: Proceedings of the IFAC Control Workshop on Engine and Powertrain Control, Simulation and Modeling. IFP Rueil- Malmaison. [42] Reynoso-Meza, G., Sanchis, J., Blasco, X., Mart́ınez, M., 2010. Mul- tiobjective design of continuous controllers using differential evolution and spherical pruning. In: Chio, C. D., Cagnoni, S., Cotta, C., Eber, M., Ekárt, A., I.Esparcia-Alcaráz, A., Goh, C.-K., J.Merelo, J., Neri, F., Preuss, M., Togelius, J., N.Yannakakis, G. (Eds.), Applications of Evolutionary Computation, Part I. Vol. LNCS 6024. Springer-Verlag, pp. 532 – 541. [43] Ruzika1, S., Wiecek, M., 2009. Successive approach to compute the bounded Pareto front of practical multiobjective optimisation problems. SIAM J. Optim. 20, 915 – 934. 29 [44] Sanchis, J., Mart́ınez, M., Blasco, X., Salcedo, J. V., 2008. A new per- spective on multiobjective optimisation by enhanced normalized normal constraint method. Structural and Multidisciplinary Optimisation (36), 537 – 546. [45] Sanchis, J., Mart́ınez, M. A., Blasco, X., Reynoso-Meza, G., 2010. Modelling preferences in multiobjective engineering design. En- gineering Applications of Artificial Intelligence, Article in press. doi:10.1016/j.engappai.2010.07.005. [46] Santana-Quintero, L., Coello, C. C., 2003. An algorithm based on difer- ential evolution for multiobjective problems. In: Dagli, C., Buczak, A., Enke, D., Embrechts, M., editors, O. E. (Eds.), Smart systems engineer- ing design: neural networks, evolutionary programming and artificial Life. Vol. 15. pp. 221 – 220. [47] Song, Z., Kusiak, A., Feb. 2009. Optimisation of temporal processes: A model predictive control approach. Evolutionary Computation, IEEE Transactions on 13 (1), 169 –179. [48] Storn, R., 2008. Sci: Differential evolution research: Trends and open questions. Vol. LNCS 143. Springer, Heidelberg, pp. 1 – 31. [49] Storn, R., Price, K., 1997. Differential evolution: A simple and effi- cient heuristic for global optimisation over continuous spaces. Journal of Global Optimisation 11, 341 – 359. [50] Tavakoli, S., Griffin, I., Fleming, P. J., September 2007. Multi-objective optimisation approach to the pi tuning problem. In: Proceedings of the IEEE Congress on Evolutionary Computation (CEC2007). pp. 3165 – 3171. [51] Toscano, R., 2005. A simple robust pi/pid controller design via numerical optimisation approach. Journal of Process Control (15), 81 – 88. [52] Wood, R. K., Berry, M. W., 1973. Terminal composition control of a binary distillation column. Chemical Engineering Science 28 (9), 1707 – 1717. 30 [53] Xue, Y., Li, D., Gao, F., 2010. Multi-objective optimisation and selec- tion for the pi control of alstom gasifier problem. Control Engineering Practice 18 (1), 67 – 76. [54] Zitzler, E., Thiele, L., Laumanns, M., Fonseca, C., da Fonseca, V., 2003. Performance assessment of multiobjective optimizers: an analysis and review. Evolutionary Computation, IEEE Transactions on 7 (2), 117 – 132. [55] A. Messac, S.M. Gupta, and B. Akbulut. Linear physical programming: A new approach to multiple objective optimization. Trans. on Opera- tional Research, 8:39–49, 1996. 31 
work_6g4jvpdji5ao5czenrtajbs3mm	----	Index evaluations and business strategies on communities of practice Index evaluations and business strategies on communities of practice Mei-Tai Chu *, Rajiv Khosla Business Systems and Knowledge Modeling Laboratory, School of Business La Trobe University, Melbourne, Victoria – 3086, Australia Abstract Business strategies and index evaluations on communities of practice (CoPs) could be the prevailing way for group learning and inno- vation building within firms. As firms grow in size, scope, and complexity, CoPs members who regularly engage in sharing and learning based on common interests, could improve organizational performance. Due to multi-criteria consideration and uncertain information handling, the purpose of this research is to use the fuzzy multi-criteria decision making (MCDM) method to analyze various index pri- orities and strategy preferences of CoPs, by undertaking empirical studies of Industrial Technology Research Institute (ITRI) in Taiwan. Fourteen units in survey case were given fifty-seven questionnaires about their priorities towards sixteen different pairs of criteria. Addi- tionally, they were also asked to estimate their four highest achievable business strategies. These evaluation criteria include satisfying multi-dimensions to capable operators. Under each of the four first-tier dimensions, four second-tier criteria are used to assess and echo their first-tier dimensions. The findings of this paper can promote performance value of implementing knowledge management systems and modelling of competitive strategies for CoPs. � 2007 Elsevier Ltd. All rights reserved. Keywords: Communities of practice (CoPs); Business Strategies; Fuzzy multi-criteria decision making (MCDM); Industrial Technology Research Institute (ITRI) 1. Introduction CoPs simultaneously emphasizes storage and distribu- tion of explicit and tacit knowledge, enhances member interaction and knowledge sharing, enables organization learning, and induces innovation to maximize the value of Knowledge Management (KM). Global enterprises, such as IBM, 3M, Xerox, Cisco, and Dell, meet transfor- mation needs by operating CoPs in emerging economies, have taken CoPs as a new central role in the value chain (Chu, Shyu, Tzeng, & Khosla, 2007a, 2007b). As knowl- edge complexity increases, and cross-field cooperation grows, the knowledge lifecycle shortens. Therefore, CoPs tends to focus on important issues which face dynamic and evolving nature of organizations. ITRI is a non-profit organization dedicated to the research and development of industrial technologies, in order to promote Taiwan’s high-tech industry. ITRI’s vision for 2008 is to maintain the same scale but triple the revenue growth of 2002. They face an enormous chal- lenge reaching such a high target, especially if they rely on traditional work methods and cultural legacy (Buckley, 1985). ITRI with many knowledge workers wants to pro- vide industrial total solution and continue to innovate. CoPs could be the best practice to break the organizational fence and establish a sharing culture. From 2000, ITRI began implementing KM. In 2003, they focused on cus- tomer driven values and gradually generated KM benefits. Each ITRI unit’s resources are limited, and preferences and problems toward CoPs are different. Every unit will have varying index priorities and goals. Therefore, the maxi- mum benefit and impact of CoPs can be achieved with careful planning (ITRI, 2003, 2007). The purpose of this research is to use fuzzy MCDM method to identify the index priority and to measure the four-business strategies of CoPs. This evaluation fea- tures two distinct stages. First, we brainstorm with KM 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.11.053 * Corresponding author. Tel.: +61 3 94793064; fax: +61 3 94795971. E-mail address: debbiechu0421@hotmail.com (M.-T. Chu). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 36 (2009) 1549–1558 Expert Systems with Applications mailto:debbiechu0421@hotmail.com consultants in ITRI to identify dimensions and criteria for the questionnaire. This research proposed four-dimen- sion architecture with objective and trade-off criteria as follows, with four criteria designed for each dimension, respectively. � Locus of Leadership: related to enforcement or volun- teer, whole or partial adoption. � Incentive Mechanism: related to award or punishment. � Member Interaction: related to sharing or security. � Complementary Asset: related to infrastructure and resource. Secondly, the questionnaire was distributed to a broad sampling of CoPs experts, to seek their views and calculate their index values. The aim is to provide a valuable refer- ence when choosing suitable CoPs business strategies. Although many scholars assert that CoPs create organiza- tion value, there has been relatively little systematic and quantitative study on the linkage between community out- comes and the underlying functional structure, the majority of papers focus on individual and subjective viewpoints, this research attempts to determine these insufficiencies, and aims at objective and trade-off driven criteria for future analysis. When evaluating business strategies of CoPs, many dif- ferent aspects could be taken into consideration. There are numerous evaluation indexes. Moreover, their struc- tures are hierarchical (Kerzner, 1989). Many scholars and experts have adopted the analytic hierarchy process (AHP) (Saaty, 1977, 1980) method to evaluate the prob- lems of relative level factors of hierarchy and to provide a more complete depiction of the structural and func- tional aspects of whole system. For instance, Hwang and Yoon (1981) discussed multi-attribute decision meth- ods and application, In addition, Cheng and Mon (1994) evaluated weapon system by AHP, Tsaur, Tzeng, and Wang (1997) analyzed tourist risk using fuzzy perspec- tives, and Tang and Tzeng (1999) researched e-business promotion strategy for information service industry. Other examples of AHP and fuzzy MCDM usage include: Yu and Lee (2002) evaluated the policy issuing for national 3G telecommunications license using the fuzzy multi-evaluation strategic method, Shyu (2003) used fuzzy MCDM to identify foundry factors and preferred location of IC design. In the questionnaire, dimensions were measured by pair- wise comparison, participants found it easier to decide dimension A is more important than B dimension instead of dimension A versus B is 5 to 1. Recently some scholars used fuzzy AHP (Buckley, 1985) to handle linguistic scale problems, which is more convenient to help participants to express opinion and concept precisely, just as Cheng and Mon (1994) applied to the choosing of weapon systems. This research uses the fuzzy AHP method proposed by Buckley (1985). The questionnaires and consultant inter- views were designed to reveal the perspectives of ITRI CoPs experts, in fourteen units. The questionnaires and interviews were also designed to weight their comparative importance of business strategies and indexes. Sixty-two surveys were returned out of the seventy-five distributed. For the purpose of the research, only five are deemed inva- lid. AHP, EXCEL, and SPSS were used to analyze the questionnaires, which were collected at different stages of the hierarchy and for different types of work. For CoPs dimension priority, the results show that par- ticipants chose Member Interaction, Incentive Mechanism, Complementary Asset, and Locus of Leadership in that order. For criteria importance, Emphasize Cross-Field Sharing scored highest, next Achievements Appraisal Basis, then Bottom-Up Teaming, and Independent IT Platform ranked lowest. For business strategies, Increasing Core Competency scored highest, next Enhancing Working Effi- ciency, Inducing Innovation Learning, and finally Promoting Responsiveness scored the lowest. Section 2 briefly introduces the literature review and comparison analysis used: the implications, benchmarks, and values as well as a description of expected benefits. The benefit assessment follows the firm’s missions and objectives, making a holistic understanding of managerial structure, a critical success index to evaluate benefits. Sec- tion 3 illustrates the characteristics of fuzzy MCDM with models. Section 4 designs a research model to meet analyt- ical needs. Section 5 conducts a questionnaire using above model to perform and verify the results from this study. Finally, Section 6 provides the conclusions and suggestions of this research. 2. Literature review and comparison analysis Most KM projects stress explicit knowledge. However, very little tacit knowledge of the individual or community is inquired upon. In the last two years KM projects empha- size people interface, especially cross organization bound- aries. As the tacit knowledge is transmitted in cross- regions, CoPs has become the critical concept in design and operation of KM systems. 2.1. CoPs implications There is different positioning and goals in various CoPs. Although the names differ, the different terms describe sim- ilar content and concept. In 2000, Verna thought knowl- edge should include and utilize KC to create organizational knowledge. In 1998, Wenger first proposed CoPs in the Harvard Business Review. He believes CoPs is an informal group sharing knowledge, points out CoPs is composed by three critical elements shown in Table 1. 2.2. CoPs benchmarking There is other team formation in an organization besides CoPs, including formal divisions, project teams, 1550 M.-T. Chu, R. Khosla / Expert Systems with Applications 36 (2009) 1549–1558 https://isiarticles.com/article/12638 
work_6h2k5norr5fhrjwguy5aqkmgti	----	Design facial appearance for roles in video games Expert Systems with Applications 36 (2009) 4929–4934 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Design facial appearance for roles in video games Shang Hwa Hsu *, Ching-Han Kao, Muh-Cherng Wu Department of Industrial Engineering and Management, National Chiao Tung University, 1001 Ta Hsueh Road, HsinChu 30010, Taiwan, ROC a r t i c l e i n f o Keywords: Video game Role Fuzzy AHP BP neural network 0957-4174/$ - see front matter � 2008 Elsevier Ltd. A doi:10.1016/j.eswa.2008.05.049 * Corresponding author. Tel.: +886 3 5726731; fax: E-mail address: shhsu@cc.nctu.edu.tw (S.H. Hsu). a b s t r a c t Roles in video games often serve as avatars of players. Different game players may have their particular preferences on a role’s facial appearance. It would be desirable to allow players to customize the design of roles. This paper presents two methods for recommending a roles’ facial appearance for a particular game player and illustrates the two methods by using heroic roles as an example. The two recommendation methods are designated as the text-input and the picture-input approaches. The text-input approach requests the game player to carry out pairwise comparisons for determining the relative weights of 16 personality traits of heroes. The recommendation mechanism for the text-input approach is based on the fuzzy AHP (analytic hierarchy process). Whereas the picture-input approach requests the game player to view a sample set of pictures and rate his/her preferences on each picture. The recommendation mech- anism for the picture-input approach is based on the BP (back-propagation) neural network. Experiments indicated that the text-input approach is more effective in terms of recommending an appropriate facial appearance, yet at the expense of needing more user time. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction In playing a video game (called a game hereafter), a player ex- presses his or her intentions by manipulating the actions of a role. A game role is a character, which serves as the player’s avatar or competitor. Roles to a game are as important as actors to a film. Casting appropriate actors can lead to the success of a film. In the same way, designing appropriate roles is very important to the success of a video game. Previous researchers have published numerous studies on the role design of video games. Most of them attempted to create a life-like role by emulating human behaviors—such as dialogue (Brusk & Eladhari, 2006; Gustafson, Boye, Fredriksson, Johanneson, & Königsmann, 2005; Jan & Traum, 2005), intelligence (Frasca, 2001; Lair & Duchi, 2000; Vala, Paiva, & Prada, 2004), motor-skill actions (Blumberg & Galyean, 1997), and emotional expressions (Rizzo, Neumann, Enciso, Fidaleo, & Noh, 2001; Wallraven, Breidt, Cunningham, & Bülthoff, 2005). Of these studies, those which analyze emotional expressions attempt to automatically create a facial expression to represent a character’s mood (e.g. happy, angry, sad, and disgusted). Such fa- cial expression studies have an implicit objective—mood manifesta- tion (Bartlett, Hager, Ekman, & Sejnowski, 1999; Ekman, 1993; Zhang & Ji, 2005). That is, a computer-generated facial expression should model a particular mood (e.g. sad) that is easily recogniz- able by humans. ll rights reserved. +886 3 5722392. In a film, proper recognition of an actors’ mood by interpreting their body languages is never enough. To produce a popular film, the attractive appearance of actors is often much more important. Likewise, the appearance of a game role may influence player involvement in the game. An attractive appearance may induce players to have affectionate feelings for the avatar, and in turn, make game play more fun (Hsu, Lee, & Wu, 2005). Even though the facial appearance of a role is very important to game design, this topic has rarely been investigated. This study proposes two methods for automatically recom- mending attractive facial appearances of heroic roles to game play- ers—in a customization manner. The observations in this study indicate that different game players have different preferences for the appearance of a hero. According to further interviews, this difference is due to the fact that players have different preferences for a hero’s personality traits. This implies that personality traits may be manifested by facial appearances. Based on this implication, this study proposes a research frame- work to design a hero’s facial appearance as favored by a particular game player (Fig. 1). This research framework involves two phases: creation and application. The creation phase develops a database that relates facial appearance to personality traits, which involves two steps. Firstly, multimedia software is used to create samples of her- oes’ facial appearances. Secondly, game players evaluate the personal- ity traits for each facial appearance. In this phase, a vector comprising numeric data can model both a facial appearance and its personality traits. The created database is called a face-trait database. The application phase implements the customization design of the heroes’ facial appearance. That is, for any given game player, mailto:shhsu@cc.nctu.edu.tw http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Text-input recommendation mechanism (Fuzzy AHP) Picture-input recommendation mechanism (Neural network) Establish face-trait database Creation phase Application phase Fig. 1. Research framework. 4930 S.H. Hsu et al. / Expert Systems with Applications 36 (2009) 4929–4934 a highly favored heroes’ facial appearance can be quickly recom- mended from the face-trait database. Two recommendation mech- anisms can be used to retrieve these highly favored facial appearance profiles. One mechanism is a text-input approach, which requires the game player to determine each personality trait’s relative weight. This mechanism uses the fuzzy AHP technique (Laarhoven & Ped- rycz, 1983; Saaty, 1980; Shamsuzzaman, 2000; Zadeh, 1975). The other mechanism is a picture-input approach, which requires the game player to evaluate their degree of preferences for a sample set of pictures (facial appearances). This mechanism is imple- mented using the back-propagation (BP) neural network technique. The remainder of this paper is organized as follows. Section 2 describes how the face-trait database is created. Section 3 first introduces the fuzzy AHP technique, and then presents the text- input recommendation mechanism. Section 4 describes the picture-input recommendation mechanism. Section 5 presents the experiments and results for justifying the effectiveness of the two recommendation mechanisms. The last section contains concluding remarks. Fig. 2. The facial appearance creation so 2. Face-trait database A prototype face-trait database has been developed. In develop- ing the database, heroes’ facial appearances were created by com- mercially available multimedia software (Live Studio Head Tool V2.6). We use 16 personality traits of heroes to characterize each fa- cial appearance created. A hero’s facial appearance could then be encoded by a vector consisting of 16 elements, which provides a key for the recommendation facial appearances. 2.1. Facial appearance creation The multimedia software used to create a facial appearance is based on a feature-based approach. That is, the configuration of a face is composed of several features. A feature may denote a part of a face such as eyes, lips, and nose or denote a face’s characteristic such as skin color and texture. For each feature, there are many op- tions for selection. Various combinations of feature options result in different facial appearances. For features associated with geometric shapes, their feature op- tions are created by the parametric-geometry approach (Myung & Han, 2001; Verroust, Schonek, & Roller, 1992). That is, changing some of its geometric parameters can vary the shape of an object. Considering an object with rectangular shape as an example, we could take the length-to-width ratio as one parameter and vary shape of the object by changing the parameter value. Likewise, a facial feature such as eyes can be modeled by a parametric-geom- etry shape with more than one parameter; for example, the curva- ture of upper eye lid and that of lower eye lid. By varying these two parameters, we could have many different shapes for modeling eyes. Example features associated with geometric shapes include lips, noses, eyes, eyebrows, and facial outline. These shapes are com- monly controlled by a set of parameters, and each parameter value ftware Live Studio Head Tool v.2.6. S.H. Hsu et al. / Expert Systems with Applications 36 (2009) 4929–4934 4931 is a real number in a predefined range such as [0, 1]. The input of such a real-number value is through a scale-bar input device (rightmost part of Fig. 2). Due to the completeness property of real numbers, there are theoretically an infinite number of options available for features associated with geometric types. By contrast, some other features provide only a finite number of feature options; for example—type of hair and type of moustache. The input of such a nominal-type feature option is through the clicking of a menu box (middle part of Fig. 2). The multimedia software we used provides 12 features. In this research, only five shape-oriented features are chosen as variables in designing a facial appearance for the prototype database. These five features are facial outline, eye brows, eyes, lip, and nose, which are selected because most prior research has concluded their importance on facial expression and recognition (Adolphs, 2002; Brunelli & Poggio, 1993; Sadrô, Jarudi, & Sinha, 2003). The prototype database which is essentially expandable, at its present state, includes 243 different facial appearances. That is, each of the five facial features has only three options for selection, which leads to 35 = 243 different facial appearances. Three in- stances of the facial appearances are shown in Fig. 3. 2.2. Personality traits evaluation We use 16 personality traits to characterize each facial appear- ance. These personality traits are a hero’s characters reported in a prior research by Hsu, Kao, and Wu (2007). According to their re- search, by the technique of factor analysis (principal components factoring) (McDonald, 1985), these 16 personality traits can be cat- egorized into three groups—bravery, visionary, and moral as shown in Fig. 4. In the characterization of a facial appearance, a five-point scale evaluates each personality trait. The higher the value, the higher degree does the facial appearance reveal—in terms of the personal- ity trait. In the characterization process, we asked 112 subjects (61males, 51 females at the age of 17–25) to evaluate the 16 per- Fig. 3. Example of facial appearances. He Bravery Visio C ourageous U nbeatable C harism atic M ighty C ool A m bitious P ersevering V isionary D aring Fig. 4. Three levels fuzzy A sonality traits for each of 243 facial appearances. All these subjects are all game player, who are either senior high school or college students. Results of the characterization process yield 243 trait vectors, each of which represents a particular facial appearance. The value of each element, ranging from 1 to 5, is the mean score reported by all the subjects. Let X = [x1, . . ., x16] denote a trait vector so obtained, which is further transformed into a normalized vector, denoted by Y ¼ ½y1; . . . ; y16� where yi ¼ xiP16 k¼1 xk and P16 i¼1 yi ¼ 1. 3. Text-input recommendation mechanism To create a hero face favored by a particular game player, it is important to know how the player weights each personality trait. The procedure for determining the weights of personality traits is by applying the fuzzy AHP technique (Saaty, 1990). The technique, widely applied in various areas (Durán & Aguilo, 2007; Kim & Yoon, 1992; Mamaghani, 2002; Muralidar & Santhanam, 1990; Wu, Lo, & Hsu, 2007) is briefly stated below. Firstly, the 16 personality traits are hierarchically clustered as shown in Fig. 4. This hierarchy indicates that these 16 personality traits, based on a prior research (Hsu et al., 2007), can be catego- rized into three groups—bravery, visionary, and moral, which are called group-level traits. Each personality traits within a group is called a member-level trait. Secondly, the relative weights for group-level traits and that for member-level traits in each group have to be determined. We there- fore have four weight-determination problems, one for group-level and three for member-level traits. To each weight-determination problem, we asked the game player to carry out a pairwise com- parison experiment, and use a fuzzy AHP algorithm to process the experiment data for determining the relative weights. Thirdly, we use the relative weights to compute a recommenda- tion-priority value to retrieve a hero face from the face-trait data- base for the game player. 3.1. Pairwise comparison experiment The pairwise comparison experiment is explained by using the weight-determination of group-level traits as an example. Of the three group-level traits, any two (a pair) is chosen for comparison. Assume the pair (bravery, visionary) is chosen. Then, the player is asked to answer the following question: ‘‘Consider a person to be recognized as a hero. Of the two groups of personality traits, which one is more important? Try to compare the degree of importance between them.” Now suppose the player answers bravery is more important than visionary. Then, five linguistic variables listed in Table 1 are provided for the player to express his/her opinion. These linguistic variables range from ‘‘absolutely important” to ‘‘equally impor- tant” on a five level scales, where between any two consecutive ro nary Moral L eadership C apable T rustw orthy C alm G racious Just U nselfish HP hierarchy model. Table 1 Linguistic variables used in the fuzzy AHP Fuzzy number Linguistic variables ~1 ¼ð1; 1; 1Þ Equally important ~3 ¼ð2; 3; 4Þ Weakly important ~5 ¼ð4; 5; 6Þ Essentially important ~7 ¼ð6; 7; 8Þ Very strongly important ~9 ¼ð8; 9; 10Þ Absolutely important ~2 ¼ð1; 2; 3Þ; ~4 ¼ð3; 4; 5Þ; ~6 ¼ð5; 6; 7Þ; ~8 ¼ð7; 8; 9Þ Intermediate values between two adjacent judgments 4932 S.H. Hsu et al. / Expert Systems with Applications 36 (2009) 4929–4934 scales an intermediate scale is additionally defined so that nine scales are finally created. As shown in Table 1, each linguistic var- iable is represented by a triangular fuzzy number—for example, ~7 ¼ð6; 7; 8Þ denotes ‘‘very strongly important”. The adoption of lin- guistic variables is to resolve the vagueness occurred in human judgment (Zadeh, 1975). Based on the pairwise comparison experiment, a n � n matrix à = [ãij] could be obtained, where n denotes the number of traits to be compared, ~A ¼ 1 ~a12 � � � ~a1n ~a21 1 � � � ~a2n .. . .. . . . . .. . ~an1 ~an2 � � � 1 2 66664 3 77775 Notice that ãij = 1/ãji, which ensures that the comparison for each pair (i, j) is consistent; and ãii = 1, which denotes that a self-compari- son is always ‘‘equally important” and is not needed. 3.2. Fuzzy AHP algorithm Two procedures are used to obtain the relative weights for group-level and member-level traits. The first one Compute_Rela- tive_Weight is intended to compute the relative weights, and the second one Validity_Check_for_Relative_Weights is developed for checking the validity of the obtained relative weights. These two procedures are respectively described below, where the definitions of arithmetic operators (i.e., �, �, and Defuzzy) are introduced in the Appendix. 3.2.1. Procedure Compute_Relative_Weight Step 1: Calculate ~W i , the fuzzy weight for each row i in à (Buckley, 1985) ~Zi ¼ð~ai 1 � ~ai 2 ����� ~ainÞ 1 n; 8i ¼ 1; 2; . . . ; n ~W i ¼ ~Zi �ð~Z1 � ~Z2 ��� �� ~ZnÞ �1 ; 8i ¼ 1; :::; n Step 2: Defuzzication of ~W i and à (Teng & Tzeng, 1993) Ŵ i ¼ Defuzzyð ~W iÞ aij ¼ Defuzzyð~aijÞ; whereA ¼ ½aij� and ~A ¼ ½~aij� Step 3: Compute relative weights W_{i} W i ¼ Ŵ iPn i¼1 Ŵ i Table 2 Values of RI n 1 2 3 4 5 6 7 8 9 10 11 RI 0.00 0.00 0.58 0.90 1.12 1.24 1.32 1.41 1.45 1.49 1.51 3.2.2. Procedure Validity_Check_for_Relative_Weights Step 1: Compute W�i W�1 W�2 .. . W�n 2 66664 3 77775 ¼ A W 1 W 2 .. . W n 2 66664 3 77775 Step 2: Compute kmax, the maximum eigenvalue kmax ¼ 1 n W�1 W 1 � � þ W�2 W 2 � � þ�� �þ W�n W n � �� � Step 3: Compute CI, the consistency index CI ¼ kmax � n n � 1 Step 4: Compute CR CR = CI/RI the values of RI are shown in Table 2 (Saaty, 1980). Step 5: Consistency check If CR 6 0.1 (pairwise comparison data à is reasonably consistent) Output the resulting relative weights Wi (1 6 i 6 n) Else (CR > 0.1 pairwise comparison data à are inconsistent) Repeat the pairwise comparison experiment. Endif 3.3. Recommending mechanism To recommend a hero face favored by a particular game player, we first compute a recommendation-priority value for each facial appearance in the face-trait database and recommend the one with the highest recommendation-priority value. Define S = {(Yi, Fi)|1 6 i 6 n} as the face-trait database developed for facial design, where Fi denotes ith facial appearance, and Yi = [yij], 1 6 j 6 16 denotes the personality-trait vector of Fi Notice that Fi is a 2D picture while Yi is a vector with 16 numeric elements, where yij denotes the degree of jth personality traits that picture Fi manifest to a ‘‘common people”. Let W = [wj], 1 6 j 6 16 represent the preferences of a particular game player on each of the 16 per- sonality traits, where wj denotes the relative weight of jth person- ality trait, perceived by the game player—an ‘‘individual people” rather than a ‘‘common people”. The procedure for recommending facial appearances most favorable to a particular game player proceeds as follows. Step 1: Compute the recommendation-priority value pi ¼ X16 j¼1 wj � yij; 1 6 i 6 n Step 2: Output the recommended one i� ¼ arg maxðpiÞ 4. Picture-input recommendation mechanism The BP neural network technique has been widely used as a pre- dictor (Dutta & Shekhar, 1988; Liau & Chen, 2005; O’Leary, 1998; Salchenberger, Cinar, & Lash, 1992; Tam & Kiang, 1992). Given a sample set of input/output data obtained from a real-world sys- tem, we can use the technique to establish a BP network that serves as an input/output mapping mechanism—a three-layer architecture as shown in Fig. 5. The established BP network can be used to predict the output for a new input data set. Details of the BP neural network technique can be referred to Wasserman (1989) and Hertz, Krogh, and Palmer (1991). The picture-input recommendation mechanism is to establish a BP network for a particular game player. For such a BP network, its Input Hidden Output fi1 fi2 fi3 fi4 fi15 Pi Fig. 5. Architecture of the neural network. S.H. Hsu et al. / Expert Systems with Applications 36 (2009) 4929–4934 4933 input is a facial appearance denoted by Fi = [fik], 1 6 i 6 243, 1 6 k 6 15, where fik (a binary number) denotes the existence of the kth feature option in the ith facial appearance. As stated, a facial appearance has five features, each of which has three options for selection. This leads to 15 (5 � 3) types of feature options. A facial appearance can then be modeled by a vector with 15 binary ele- ments, in which 1 denotes a feature option exists and 0 denotes its nonexistence. The output of such a BP network is denoted by Pi, which denotes the game player’s preference on the ith facial appearance. The use of the BP neural network technique involves two phases: training and predicting phases. In the training phase, we at- tempt to establish a BP network for the game player. Out of the 243 facial appearances, we randomly sample 81 ones and use them to build up a BP network. The algorithms of the training process can be referred to Grossberg (1974) and Rumelhart, Hinton, and Wil- liams (1986). In the predicting phase, we attempt to use the estab- lished BP network to predict the player’s preferences on the other facial appearances not considered in training phase. That is, the BP network will predict the player’s preference on the remaining 162 facial appearances. 5. Experiments An experiment is carried out to compare the performance of the text-input and the picture-input approaches. Twenty game players (10 males, 10 females at the age of 17–25) are invited as experi- ment subjects, and 243 facial appearances are created in the face-trait database. The experiment proceeds as follows. Firstly, we carried out the text-input approach in which each subject i is requested to perform a pairwise comparison for deter- mining his/her relative weights on the 16 personality traits. The system will output a picture (say, X�i ) with the highest recommen- dation-priority value. Secondly, we carried out the picture-input approach in which each subject i is requested to view 81 pictures and give his/her preference on each picture. The input/output data of the 81 pic- tures are used to establish a BP network. Then, all the 243 pictures are fed into the BP network, and the one (say, Y�i ) with the highest preference value will be recommended. Table 3 Comparing effectiveness Subjects 1 2 3 4 5 6 7 8 9 10 Text-input xi 0 0 0 0 0 1 0 1 0 0 Picture-input yi 0 0 3 0 1 2 0 2 0 1 Finally, each subject is requested to view the remaining 162 pic- tures and give his/her preference on each one. Out of the 243 pic- tures, the one with the highest preference (say, Z�i ) is selected. Define rðP�i Þ as the preference of picture P evaluated by subject i. The effectiveness of the text-input approach is measured by the distribution of the metric: xi ¼ rðZ�i Þ� rðX � i Þ, and that for the pic- ture-input approach is measured by the distribution of yi ¼ rðZ�i Þ� rðY � i Þ . The distributions of xi and yi are shown in Table 3. The table indicates that the text-input approach is superior to the picture-in- put approach. In the text-input approach, 85% game players (17 out of the 20 subjects) will get their most favor picture, while only 35% can get so in the picture-input approach. However, the text-input approach needs more user time. The average time for performing a pairwise comparison is about 26 min. and that for evaluating 81 pictures is about 2.6 min. The experiment results lead to the following implication. For a video game equipped with relatively few numbers of pictures, we would suggest an exhaustive display of pictures to a game player. In contrast, for a video game equipped with a great amount of pictures, we would suggest the use of the text-input approach. 6. Conclusions Playing video games through manipulating a role is quite com- mon. Role design therefore has been a significant research issue. Most prior research attempted to create a life-like role by develop software for emulating human’s capabilities, such as dialogues, intelligence, motor-skill, and emotional expressions. This research is unique in providing a customized facial appearance for each game player. A video game with such a customization function would become more popular. Two methods for providing such a customization function have been developed. One is called the text-input approach whose mechanism is based on the fuzzy AHP technique. The other is called the picture-input approach whose mechanism is based on the BP neural network technique. The text-input approach requires a game player to perform a pairwise comparison on 16 personality traits. The picture-input approach requires a game player to view and evaluate a sample set of pictures. Experiment results indicated that the text-input approach is superior to the picture-input approach, in terms of recommending an appropriate picture to a game player; however, at the price of needing more user input time. Appendix. Arithmetical operators for fuzzy numbers The fuzzy arithmetic operators used in this research are defined below (Laarhoven & Pedrycz, 1983; Zadeh, 1975), by referring two fuzzy numbers ã1 = (l1, m1, r1) and ã2 = (l1, m1, r1). (1) Addition operator: �ã1 � ã2 = (l1 + l2, m1 + m2, r1 + r2) (2) Multiplication operator: �ã1 � ã2 = (l1 � l2, m1 � m2, r1 � r2) (3) Defuzzication operator (Teng & Tzeng, 1993) Defuzzy(ã1) = |(r1 � l1) + (m1 � l1)|/3 + l1 11 12 13 14 15 16 17 18 19 20 0 0 0 0 0 0 1 0 0 0 0 1 1 2 0 2 1 1 3 1 4934 S.H. Hsu et al. / Expert Systems with Applications 36 (2009) 4929–4934 References Adolphs, R. (2002). Recognizing emotions from facial expressions: Psychological and neurological mechanisms. Behavioral and Cognitive Neuroscience Reviews 1, 1, 21–61. Bartlett, M. S., Hager, J. C., Ekman, P., & Sejnowski, T. J. (1999). Measuring facial expressions by computer image analysis. Psychophysiology, 36, 253–263. Blumberg, B., & Galyean, T. (1997). Multi-level control for animated autonomous agents: Do the right thing. . . oh, not that. . .. In: R. Trappl & P. Petta (Eds.), Creating personalities for synthetic actors. Berlin: Springer-Verlag. (pp. 74–82). Brunelli, R., & Poggio, T. (1993). Face recognition features versus templates. IEEE Transactions on PAMI, 15(10), 1042–1052. Brusk, J., & Eladhari, M. (2006). Playing the character. RPG seminar, Tampere, March (pp. 30–31). Buckley, J. J. (1985). Ranking alternatives using fuzzy numbers. Fuzzy Sets and Systems, 15(1), 21–31. Durán, O., & Aguilo, J. (2007). Computer-aided machine-tool selection based on a Fuzzy-AHP approach. Expert Systems with Applications. doi:10.1016/ j.eswa.2007.01.046. Dutta, S., & Shekhar, S. (1988). Bond rating: A non-conservative application of neural networks. In Proceedings of the IEEE international conference on neural networks (pp. 443–450). Ekman, P. (1993). Facial expressions and emotion. American Psychologist, 48, 384–392. Frasca, G. (2001). Rethinking agency and immersion: Videogames as a means of consciousness-raisin (essay). SIGGRAPH 2001 N-Space art gallery. Grossberg, S. (1974). Classical and instrumental learning by neural networks. Progress in theoretical biology (Vol. 3, pp. 51–141). New York, NY: Academic Press. Gustafson, J., Boye, J., Fredriksson, M., Johanneson, L., & Königsmann, J. (2005). Providing computer game characters with conversational abilities. In Proceedings of intelligent virtual agent (IVA05), Kos, Greece. Hertz, J., Krogh, A., & Palmer, R. G. (1991). Introduction to the theory of neural computation. Redwood City, CA:: Addison-Wesley. Hsu, S. H., Kao, C. H., & Wu, M. C. (2007). Factors influencing player preferences for heroic roles in role-playing games. CyberPsychology & Behavior, 10(2), 293–295. Hsu, S. H., Lee, F. L., & Wu, M. C. (2005). Designing action games for appealing to buyers. CyberPsychology & Behavior, 8(6), 585–591. Jan, D., Traum, D. R. (2005). Dialog simulation for background characters. In IVA (pp. 65–74). Kim, C. S., & Yoon, Y. (1992). Selection of a good expert system shell for instructional purposes in business. Information & Management, 23(5), 249–262. Laarhoven, P. J. M., & Pedrycz, W. (1983). A fuzzy extension of Saaty’s priority theory. Fuzzy Sets and Systems, 11(3), 229–241. Lair, J. E., & Duchi, J. C. (2000). Creating human-like synthetic characters with multiple skill levels: A case study using the Soar Quakebot. AAAI 2000 fall symposium on simulating human agent, technical report FS-00-03 (pp. 75–79). AAAI Press. Liau, L. C. K., & Chen, B. S. C. (2005). Process optimization of gold stud bump manufacturing using artificial neural networks. Expert Systems with Applications, 29, 264–271. Mamaghani, F. (2002). Evaluation and selection of an antivirus and content filtering software. Information Management & Computer Security, 10(1), 28–32. McDonald, R. (1985). Factor analysis and related techniques. Hillsdale, NJ: Lawrence Erlbaum. Muralidar, K., & Santhanam, R. (1990). Using the analytic hierarchy process for information system project selection. Information & Management, 18(1), 87–95. Myung, S., & Han, S. (2001). Knowledge-base parametric design of mechanical products based on configuration design method. Expert Systems with Applications, 21, 99–107. O’Leary, D. E. (1998). Using neural networks to predict corporate failure. International Journal of Intelligent Systems in Accounting, Finance and Management, 7, 187–197. Rizzo, A. A., Neumann, U., Enciso, R., Fidaleo, D., & Noh, J. Y. (2001). Performance- driven facial animation: Basic research on human judgments of emotional state in facial avatars. CyberPsychology & Behavior, 4(4), 471–487. Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning internal representation by error propagation. Parallel and Distributed Processing, 1, 318–362. Saaty, T. L. (1980). The analytic hierarchy process: Planning, priority setting, resource allocation. New York: McGraw-Hill (p. 20). Saaty, T. L. (1990). How to make a decision: The analytic hierarchy process. European Journal of Operational Research, 48(1), 9–26. Sadrô, J., Jarudi, I., & Sinha, P. (2003). The role of eyebrows in face recognition. Perception, 32, 285–293. Salchenberger, L. M., Cinar, E. M., & Lash, N. A. (1992). Neural networks: A new tool for predicting thrift failures. Decision Sciences, 23(4), 899–916. Shamsuzzaman, M. (2000). Selection of a FMS based on fuzzy set theory and AHP methods. ISE-Thesis, Asian Institute of Technology, Bangkok. Tam, K. Y., & Kiang, M. Y. (1992). Managerial applications of neural networks: The case of bank failure predictions. Management Science, 38(7), 926–947. Teng, J. Y., & Tzeng, G. H. (1993). Transportation investment project selection with fuzzy multi-objective. Transportation Planning and Technology, 17, 91–112. Vala, M., Paiva, A., & Prada, R. (2004). From motion control to emotion influence: Controlling autonomous synthetic characters in a computer game. In AAMAS’04, New York, USA (pp. 1300–1301). Verroust, A., Schonek, F., & Roller, D. (1992). Rule-oriented method for parameterized computer-aided design. Computer Aided Design, 24(10), 531– 540. Wallraven, C., Breidt, M., Cunningham, D. W., & Bülthoff, H. H. (2005). Psychophysical evaluation of animated facial expressions. In Proceedings of the 2nd symposium on Applied perception in graphics and visualization (pp. 26–28). Wasserman, R. (1989). Neural computing: Theory and practice. New York, NY: Van Nostrand Reinhold. Wu, M. C., Lo, Y. F., & Hsu, S. H. (2007). A fuzzy CBR technique for generating product ideas. Expert Systems with Application, 34, 530–540. Zadeh, L. A. (1975). The concept of a linguistic variable and its application to approximate reasoning. Information Sciences, Part 1, 8, 199–249; Part 2, 8, pp. 301–357; Part 3, pp. 43–80. Zhang, Y., & Ji, Q. (2005). Active and dynamic information fusion for facial expression understanding from image sequences. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(5), 699–714. http://dx.doi.org/10.1016/j.eswa.2007.01.046 http://dx.doi.org/10.1016/j.eswa.2007.01.046 Design Facial Appearance facial appearance for Roles roles in Video Gamesvideo games Introduction Face-trait database Facial appearance creation Personality traits evaluation Text-input recommendation mechanism Pairwise comparison experiment Fuzzy AHP algorithm Procedure Compute_Relative_Weight Procedure Validity_Check_for_Relative_Weights Recommending mechanism Picture-input recommendation mechanism Experiments Conclusions Arithmetical operators for fuzzy numbers References 
work_6hpkgpx7orhhpc2hu5ryelt3j4	----	Estimating translation probabilities for social tag suggestion Expert Systems with Applications 42 (2015) 1950–1959 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Estimating translation probabilities for social tag suggestion http://dx.doi.org/10.1016/j.eswa.2014.10.002 0957-4174/� 2014 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). ⇑ Corresponding author at: Room 4-506, FIT building, Tsinghua University, Beijing 100084, China. Tel.: +86 138 1035 7951. E-mail addresses: cxx.thu@gmail.com (X. Chen), liuzy@tsinghua.edu.cn (Z. Liu), sms@tsinghua.edu.cn (M. Sun). 1 The original record was obtained from the book review website Douban (www.douban.com) in Chinese. Here we translate it into English for comprehension. Xinxiong Chen ⇑, Zhiyuan Liu, Maosong Sun Department of Computer Science and Technology, State Key Lab on Intelligent Technology and Systems, National Lab for Information Science and Technology, Tsinghua University, Beijing 100084, China a r t i c l e i n f o Article history: Available online 12 October 2014 Keywords: Natural language processing Tag suggestion Translation model Word alignment model Pointwise mutual information a b s t r a c t The task of social tag suggestion is to recommend tags automatically for a user when he or she wants to annotate an online resource. In this study, we focus on how to make use of the text description of a resource to suggest tags. It is intuitive to select significant words from the text description of a source as the suggested tags. However, since users can arbitrarily annotate any tags to a resource, tag suggestion suffers from the vocabulary gap issue — the appropriate tags of a resource may be statistically insignifi- cant or even do not appear in the corresponding description. In order to solve the vocabulary gap issue, in this paper we present a new perspective on social tag suggestion. By considering both a description and tags as summaries of a given resource composed in two languages, tag suggestion can be regarded as a translation from description to tags. We propose two methods to estimate the translation probabilities between words in descriptions and tags. Based on the translation probabilities between words and tags estimated for a large collection of description-tags pairs, we can suggest tags according to the words in a resource description. Experiments on real-world datasets indicate that our methods outperform other methods in precision, recall and F-measure. Moreover, our methods are relatively simple and efficient, which makes them practical for Web applications. � 2014 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). 1. Introduction In Web 2.0, Web users often use tags to collect and share online resources such as Web pages, photos, videos, movies and books. As an example, we consider a social tagging system for books. Table 1 presents a book entry annotated with several tags by multiple users.1 On the top of Table 1 we list the title and a short introduction of the book ‘‘The Count of Monte Cristo’’. The bottom of Table 1 shows the annotated tags, each of which is followed by a number in brack- ets, which is the total number of users who used the tag to annotate the book. As the tags of online resources are annotated collabora- tively by multiple users, we also refer to these tags as social tags. For a resource, we refer to the additional information, such as the title and the introduction of a book, as description, and the user- annotated social tags as annotation. The task of social tag suggestion is to automatically recommend tags for a user when he or she wants to annotate a resource. Social tag suggestion, as a crucial component for social tagging systems, can help users annotate resources. Moreover, social tag suggestion is usually considered as an equivalent problem to modeling social tagging behaviors, which is playing an increasingly important role in social computing and information retrieval. Most online resources have descriptions, usually containing abundant information about resources (Liu, Chen, & Sun, 2011). For example, on a book review website, each book entry contains a title, the author(s) and an introduction of the book. Thus, a num- ber of researchers (Liu et al., 2011; Katakis, Tsoumakas, & Vlahavas, 2008; Mishne, 2006; Xu, Fu, Mao, & Su, 2006) propose to automat- ically suggest tags based on resource descriptions, which is collectively known as the content-based approach (Xu et al., 2006). In this study, we focus on how to make use of the text description of a resource to suggest tags. Note that besides descrip- tions, online resources may also have multimedia data (e.g., images, videos and audio files) and a survey of multimedia tagging can be found in Wang, Ni, Hua, and Chua (2012). One may think to suggest tags by selecting important words from descriptions. This approach is far from sufficient because descriptions and annotations use diverse vocabularies, which is typically referred to as the vocabulary gap problem (Liu et al., 2011). The vocabulary gap is usually reflected in two primary issues: http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.10.002&domain=pdf http://creativecommons.org/licenses/by-nc-nd/3.0/ http://dx.doi.org/10.1016/j.eswa.2014.10.002 http://creativecommons.org/licenses/by-nc-nd/3.0/ mailto:cxx.thu@gmail.com mailto:liuzy@tsinghua.edu.cn mailto:sms@tsinghua.edu.cn http://www.douban.com http://dx.doi.org/10.1016/j.eswa.2014.10.002 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Table 1 An example of social tagging. The number in the bracket after each tag is the total count of users that annotated the book with the tag. Description Title: The Count of Monte Cristo Intro: The Count of Monte Cristo is one of the most popular fiction by Alexandre Dumas. The writing of the work was completed in 1844. . . . Annotation Dumas (2748), Count of Monte Cristo (2716), foreign literature (1813), novel (1345), France (1096), classic (1062), revenge (913), famous book (759), . . . X. Chen et al. / Expert Systems with Applications 42 (2015) 1950–1959 1951 1. A portion of tags in the annotation do appear in the correspond- ing description, but they may not be statistically significant. 2. A portion of tags may even not appear in the description. Taking the book entry in Table 1 as example, the tag ‘‘classic’’ had been annotated by 1062 users but it did not appear in the description; another appropriate tag ‘‘famous book’’ also did not appear in the description. Many approaches have been proposed to reduce the vocabulary gap and find the semantic correspondence between descriptions and annotations. Several researchers regard social tag suggestion as a classification problem by considering each tag as a category label (Fujimura, Fujimura, & Okuda, 2008; Heymann, Ramage, & Garcia-Molina, 2008; Katakis et al., 2008; Lee & Chun, 2007; Mishne, 2006; Ohkura, Kiyota, & Nakagawa, 2006). Various classi- fiers such as Naive Bayes, kNN and SVM have been explored and words are used as features. Some researchers propose to use the topic information between words and tags to suggest tags (Iwata, Yamada, & Ueda, 2009; Si, Liu, & Sun, 2010). In this paper, we propose a new perspective on social tag sug- gestion to solve the vocabulary gap problem. By regarding both the description and the annotation as parallel summaries of a resource, we want to build a translation model to estimate the translation probabilities between the words in descriptions and tags in annotations. The translation probabilities are able to cap- ture the semantic relation between words and tags. After obtaining the translation probabilities, the tagging behavior associated with a resource can then be regarded as a word translation process: 1. A user reads the resource description and understands its sub- stance according to the important words in the description. 2. Triggered by the important words in the description, the user translate these words into corresponding tags and annotate the resource with these tags. In Fig. 1, we provide a simple example to demonstrate the basic idea of using word translation for tag suggestion. In this figure, some words in the first sentence of book description are translated to tags in the annotation. The translation is denoted with various arrows from words or phrases in the description to tags in the annotation. For example, the phrase Count of Monte Cristo in the description is translated to two tags, including Dumas and Count of Monte Cristo, and the word fictions is translated to novel. In this paper, we propose two methods to estimate transla- tion probabilities between words in descriptions and tags in Fig. 1. An example of the word translation method for suggesting tags from a given description. annotations. One method is the word alignment model (WAM) in statistical machine translation (SMT) and the other method is mutual information (MI) (Lin, 1998). It is straightforward to use WAM since it is the basic model in SMT to estimate the translation probabilities. For training, WAM requires a collection of parallel documents, where each document pair should have the compara- ble length. In this paper we propose a sampling method to prepare length-balanced description–annotation pairs for WAM. Moreover, we propose the second method, MI, to estimate the translation probabilities. Mutual information is a popular measure that can utilize co-occurrence information to measure semantic similarities between two words (Lin, 1998). Our model can solve the vocabulary gap problem because the translation probabilities estimated by WAM and MI are able to capture the semantic relation between words and tags. Thus we can suggest tags that are not statistically significant or even not appear in the descriptions based on the translation probabilities. We hypothesize that our approach is better than the methods mentioned above and conduct experiments to investigate the per- formance of our model in the task of tag suggestion. Experiments on real-world datasets indicate that our method outperforms other methods in precision, recall and F-measure. Moreover, our method is relatively simple and efficient, as proven with the computational complexity, which makes it practical for Web applications. The remainder of this paper is organized as follows: In Section 2 we briefly introduce some of the most commonly used methods for tag suggestion. Sections 3 and 4 introduce the details of our approach. Section 5 presents the experimental evaluation of our approach compared to other existing techniques. Finally Section 6 concludes the paper. 2. Related work Many researchers have built social tag suggestion systems based on collaborative filtering (CF) (Herlocker, Konstan, Borchers, & Riedl, 1999; Herlocker, Konstan, Terveen, & Riedl, 2004). CF is a widely used technique in recommender systems (Resnick & Varian, 1997). The collaboration-based methods typically base their suggestions on the tagging history of the given resource and user, without considering resource description. Matrix Factor- ization (Rendle, Balby Marinho, Nanopoulos, & Schmidt-Thieme, 2009) and FolkRank (Jaschke, Marinho, Hotho, Schmidt-Thieme, & Stumme, 2008) are representative CF methods for social tag sug- gestion. Most of these methods suffer from the cold-start problem (Lam, Vu, Le, & Duong, 2008), i.e., they are not able to provide effec- tive suggestions for resources that no one has annotated yet. The content-based approach for social tag suggestion ameliorates the cold-start problem of the collaboration-based approach by suggesting tags according to resource descriptions. Therefore, the content-based approach plays an important role in social tag suggestion, especially for new resources and new tagging systems without tagging history. Several researchers regard social tag suggestion as a classifica- tion problem by considering each tag as a category label (Fujimura et al., 2008; Heymann et al., 2008; Katakis et al., 2008; Lee & Chun, 2007; Mishne, 2006; Ohkura et al., 2006). Various 1952 X. Chen et al. / Expert Systems with Applications 42 (2015) 1950–1959 classifiers such as Naive Bayes, kNN, SVM and neural networks have been explored to solve the social tag suggestion problem. There are two issues emerging from the classification-based methods: (1) the annotations provided by users are noisy, and the classification-based methods cannot handle the issue well (Liu et al., 2011); (2) the training cost and classification cost of many classification-based methods are usually proportional to the number of classification labels (Si et al., 2010). Thus, these methods may be inefficient for a real-world social tagging system, where hundreds of thousands of unique tags should be considered as classification labels. Inspired by the emerging popularity of latent topic models such as Latent Dirichlet Allocation (LDA) (Blei, Ng, & Jordan, 2003), var- ious methods have been proposed to model tags using generative latent topic models. One intuitive approach is assuming that both tags and words are generated from the same set of latent topics. By representing both tags and descriptions as the distributions of latent topics, this approach suggests tags according to their likelihood given the description (Krestel, Fankhauser, & Nejdl, 2009; Si & Sun, 2009). Bundschus et al. (2009) proposed a joint latent topic model of users, words and tags. Iwata et al. (2009) pro- posed an LDA-based topic model, Content Relevance Model (CRM), which aims to identify the content-related tags for suggestion. Empirical experiments revealed that CRM outperformed both classification methods and Corr-LDA (Blei & Jordan, 2003), a gener- ative topic model for contents and annotations. Most latent topic models have to pre-specify the number of topics before training. We can either use cross-validation to deter- mine the optimal number of topics or employ the infinite models for automatically adjusting the number of topics during training. Both of the solutions are usually computationally complicated. More importantly, topic-based methods suggest tags by measuring the topical relevance of tags and resource descriptions. The latent topics are at the concept level (Liu, Huang, Zheng, & Sun, 2010), which are usually too coarse-grained to precisely suggest fine- grained tags such as named entities, e.g., the tags ‘‘Dumas’’ and ‘‘Count of Monte Cristo’’ in Table 1. To remedy the problem, Si et al. (2010) proposed a generative model, Tag Allocation Model (TAM), which considers the words in descriptions as the possible topics to generate tags. Our model is also a content-based approach so we compare some of the content-based approach mentioned above (kNN, Naive Bayes, CRM and TAM) to investigate the performance of our approach. 3. Learning translation probabilities Given a resource description, the ranking score of a tag can be calculated from two probabilities: (1) the translation probabilities between words and tags, (2) the probabilities of a word given the description. In Section 3, we will present how to use two different methods, word alignment model and mutual information, to esti- mate the translation probabilities between words in description and tags in annotations. We will introduce how to calculate the probabilities of a word given the description and perform tag sug- gestion in Section 4. First we give formal definitions to description and tags. In this paper, the description of a resource is the textual information of the resource, including the title and the short introduction. The description can be treated as a bag of words. Here we do not use stemming techniques or lemmatization techniques to preprocess the description. We just removed the stop-words in the descrip- tion. Tags of a resource are labels with count information to describe the resource. Before introducing the methods in details, we introduce the notation. A resource is denoted as r 2 R, where R is the set of resources. Each resource in the training set contains a description and an annotation containing a set of tags. The description dr of resource r can be regarded as a bag of words wr ¼fðwi; cwiÞg Nr i¼1 , where cwi is the count of word wi, and Nr is the number of unique words in r. The annotated tags ar of the resource r is represented as tr ¼fðti; ctiÞg Mr i¼1 , where cti is the count of the tag ti, and Mr is the number of unique tags for r. 3.1. Word alignment model (WAM)-based approach WAM, as a traditional machine translation method, requires a parallel training dataset consisting of a number of aligned sentence pairs. We assume the description and the annotation of a resource are written in two distinct languages. Thus, we prepare our parallel training dataset by pairing descriptions and annotations. Accord- ingly, the WAM-based approach contains two steps. First, given a collection of annotated resources, we prepare description– annotation pairs for using the word alignment model. Second, given a collection of description–annotation pairs, we adopt IBM Model-1, a widely used word alignment model, to learn the trans- lation probabilities between words in descriptions and tags in annotations. We will introduce the two steps separately. 3.1.1. Preparing description–annotation pairs for WAM In a typical tag suggestion system, the length of a resource description is usually limited to hundreds of words. In addition, it is common for some popular resources to be annotated by multi- ple users with thousands of tags. For example, the tag Dumas is annotated by 2,748 users for the book in Table 1. In another extreme, a resource may be annotated with only several tags. We have to address the length-unbalance between a resource descrip- tion and its corresponding annotation for two reasons. (1) When the number of annotated tags is large, it is impossible to list all annotated tags on the annotation side of a description–annotation pair. The performance of word alignment models will also suffer from the unbalanced length of aligned pairs in the parallel training data set (Och & Ney, 2003). (2) Moreover, the annotated tags may have different importance for the resource. It would be unfair to treat these tags without distinction. In this study, we propose a sampling method to prepare length- balanced description–annotation pairs for word alignment. The basic idea is to sample a bag of tags from the annotation according to tag weights and make the generated bag of tags have compara- ble length with the words in description. For example, the length of the description in Table 1 is 54 and the number of unique tags is 21. If we list 54 words in one side and 21 tags in another side, we will get a sentence pair with unbalanced length. Thus we propose to sample a bag of tags with comparable length with the words, for example, 54 tags. We consider two parameters when sampling tags. First, we have to select a tag weighting type for sampling. In this paper, we investigate two straightforward weighting types, including tag frequency (TFt) within the annotation, which considers the local importance, and tag-frequency inverse-document-frequency (TF-IDFt), which also considers global specification (IDFw) besides TFt. Given a resource r, TFt and TF-IRFt of the tag t are defined as TFt ¼ ctP t ct ; TF-IDFt ¼ ctP t ct � log jRj jfr 2 R : ct > 0gj � � ð1Þ where jr 2 R : ct > 0j indicates the number of resources that have been annotated with the tag t. Another parameter is the length ratio between the description and the sampled annotation. We denote the ratio as d ¼ jwrjjtrj , where jwrj is the number of words in the description and jtrj is the num- ber of tags in the annotation. Still take the book in Table 1 for 3. X. Chen et al. / Expert Systems with Applications 42 (2015) 1950–1959 1953 example, if the length of the description is 54 and the length ratio is 10=5, then we will sample 27 tags from the annotations. 3.1.2. Learning translation probabilities for WAM After preparing aligned description–annotation pairs for WAM, next we will choose an appropriate WAM model to obtain the translation probabilities between words in description and tags in annotations. Note that the annotated tags (tr ) form a bag of labels with no position information, thus, we select IBM Model-1 (Brown, Pietra, Pietra, & Mercer, 1993) for training, which does not take word position information into account on both sides for each aligned pair. Suppose the source language is the description, and the target language is the annotation. We use word alignment models to learn the translation probabilities between words in descriptions and labels in annotations. In IBM Model-1, the relationship of the source language w ¼ wJ1 and the target language t ¼ t I 1 is con- nected via a hidden variable describing an alignment mapping from source position j to target position aj: Pr wJ1jt I 1 � � ¼ X aJ 1 Pr wJ1; a J 1jt I 1 � � : ð2Þ The alignment aJ1 also contains empty-word alignments aj ¼ 0 which align source words to the empty word. IBM Model-1 can be trained using an Expectation–Maximization (EM) algorithm in an unsupervised fashion, and obtains the translation probabilities of two vocabularies, i.e., PrðwjtÞ, where t is a tag and w is a word. IBM Model-1 only produces one-to-many alignments from source language to target language. The learned model is thus asymmetric, i.e., the model learned from description–annotation pairs is different from the model learned from annotation– description pairs. So we establish learned translation models in two directions: one regards description as the source language and annotations as the target language, and the other is in the reverse direction of the pairs. We denote the first model as Prd2a and the latter as Pra2d . We further define PrðtjwÞ as the harmonic mean of the two models: PrðtjwÞ/ k Prd2aðtjwÞ þ 1 � k Pra2dðtjwÞ � ��1 ; ð3Þ where k is the harmonic factor for combining the two models. When k ¼ 1 or k ¼ 0, it simply uses model Prd2a or Pra2d correspondingly. Finally we get the translation probabilities PrðtjwÞ using the WAM model, which can be regarded as the sematic relatedness between words and tags. 3.2. Mutual information (MI)-based approach From Section 3.1.1, we can see that WAM has to use a sampling technique to prepare description–annotation pairs. Unlike WAM, MI only needs the co-occurrence information between words in description and tags in annotations, which does not require sampling. We obtain translation probabilities using MI as follows. First, for each pair of a word w in the descriptions and a tag t in the annotations, we compute their mutual information score. Informally, mutual information divides the probability of observ- ing w and t together in the same resource by the probabilities of observing w and t independently. The mutual information between a word w and a tag t is calculated as follows: Iðw; tÞ¼ X Xw¼0;1 X Xt¼0;1 pðXw; XtÞ � log PrðXw; XtÞ PrðXwÞPrðXtÞ ð4Þ where Xw and Xt are binary variables indicating whether w or t is present or absent, respectively. The estimation of the probabilities PrðXwÞ; PrðXtÞ and PrðXw; XtÞ can be found in Karimzadehgan and Zhai (2010). For a word w, we set a tag co-occurrence threshold c and will remove the mutual information between word w and tag t if cðXw ¼ 1; Xt ¼ 1Þ6 c, where cðXw ¼ 1; Xt ¼ 1Þ is the number of resources that contain both w and t. We set the co-occurrence threshold c for two reasons. (1) We are usually less confident with the translation probabilities estimated using infrequent word-tag pairs, which are usually noisy and unimportant. With the thresh- old, we can filter out a lot of noisy information. (2) Moreover, we can largely reduce the computation cost of estimating mutual information scores. Of course, when c ¼ 0, we will not remove any word-tag pairs. Thus, we normalize the mutual information score to obtain a translation probability PrðtjwÞ between a word u and a tag t: PrðtjwÞ¼ Iðw; tÞP t0 Iðw; t 0Þ ð5Þ Intuitively, the probability is higher if the word u and the tag t are more likely to co-occur. For example, we get the probabilities Prfrevengejrevengeg¼ 0:127 and Prfmartialartsjrevengeg¼ 0:088 from MI model, standing for the word ‘‘revenge’’ is more likely to co-occur with the tag ‘‘revenge’’ than the ‘‘martial arts’’. 3.3. Emphasize self-translation probability As a word may not always appear as a tag in annotation, the approaches described in Section 3.1 and Section 3.2 may under- estimate the self-translation probabilities, i.e., it is possible that Prðt – wjwÞ > Prðt ¼ wjwÞ. Here we propose to emphasize the self-translation probability for two reasons. (1) Under estimation of self-translation probabilities may lead to a situation where a proper tag t that also appears frequently in the description may receive a less recommendation score (i.e., Prðtjw ¼ tÞ) compared to other tags t0 (i.e., Prðt0jw ¼ tÞ). (2) In some real tag suggest systems, the tags that appear in the resource description are more likely to be selected by users for annotation. Hence, we introduce a parameter a to emphasize self-translation probabilities. This idea can be applied to adjust the translation probabilities from any translation models. PrðtjwÞ¼ a þð1 � aÞPrðt ¼ wjwÞ t ¼ w ð1 � aÞPrðtjwÞ t – w � ð6Þ When a ¼ 1:0, the method will suggest tags simply according to their importance scores in the description, whereas when a ¼ 0, it will not emphasize the tags that appear in the description and suggest tags only according to translation probabilities. For exam- ple, if a ¼ 0:5 and Prfrevengejrevengeg¼ 0:127, then we will get a emphasized probability Prfrevengejrevengeg¼ 0:5635. 4. Suggesting tags with translation probabilities After estimating translation probabilities between words and tags PrðtjwÞ in Section 3, we will show how to suggest tags in this section. Given a resource description dr , our model for tag suggestion is a 3-step process: 1. Measure the importance score PrðwjdrÞ of each word w in description dr . 2. Compute the ranking score of tag t by Prðtjdr ¼ wrÞ¼ X w2wr PrðtjwÞPrðwjdrÞ ð7Þ According to PrðtjdrÞ, suggest top-ranked tags to users. Table 2 Statistical information of two datasets. D; W; L; �Nd and �Na are the number of resources, the vocabulary of descriptions, the vocabulary of tags, the average number of words in each description and the average number of tags in each resource, respectively. Data D W T �Nd �Na BOOK 70,000 174,748 46,150 211.6 3.5 BIBTEX 158,924 91,277 50,847 5.8 2.7 1954 X. Chen et al. / Expert Systems with Applications 42 (2015) 1950–1959 Since we get PrðtjwÞ in Section 3, we focus on how to measure PrðwjdrÞ in this section. Here we investigate three methods to compute the importance score of a word in a resource description: TF-IDFw, TextRank and their product. TF-IDFw and TextRank are the two most widely adopted methods to weight words in a document. Similar to TF-IDFt mentioned in Section 3.1.1, TF-IDFw (Salton & Buckley, 1988) considers both the local importance (TFw) and global specification (IDFw). TextRank (Mihalcea & Tarau, 2004) is proposed to compute term importance using graph-based algo- rithms such as PageRank (Page, Brin, Motwani, & Winograd, 1998), which considers the semantic relations between terms as a term graph of the given resource description. We also use the product of TF-IDFw and TextRank to weight terms, which poten- tially takes both global information and term relations into account. Finally we get the ranking of tags according to Eq. (7). Because WAM and MI can estimate the translation probabilities between words and tags, which can be regarded as the semantic relatedness between words and tags, we hypothesize that our methods are bet- ter than the baseline methods mentioned in Section 2 (kNN, Naive Bayes, CRM and TAM). In next section we run experiments to val- idate this hypothesis. 5. Experiments 5.1. Datasets and evaluation metrics Datasets In our experiments, we select two real-world datasets of social tagging systems that have diverse properties to evaluate our methods. In Table 2 we show the detailed statistical informa- tion of the two datasets. The first dataset, denoted as BOOK, was obtained from a popular Chinese book review website, www.douban.com, which contains the descriptions of books and the tags collaboratively annotated by users. The second dataset, denoted as BIBTEX, was obtained from an English online bibliography website, www.bibsonomy.org.2 The dataset contains the descriptions for academic papers and the tags annotated by users. As shown in Table 2, the average length of descriptions in the BIBTEX dataset is much shorter than the BOOK dataset. Moreover, the BIBTEX dataset does not provide how many times each label is used in a resource annotation. 3 In more detail, the training phase of WAM contains preparing parallel training dataset with OðD �NaÞ and learning translation probabilities using word alignment models with OðID �Nd �NaÞ, where I is the number of iterations for learning trigger probabilities, and �Na is the average number of tags for each description after 5.1.1. Evaluation metrics We use precision that measures exactness, recall that measures completeness and F-measure, which is the harmonic mean of pre- cision and recall, to evaluate the performance of tag suggestion methods. For a resource, we denote the original tags (gold stan- dard) as T a, the suggested tags as T s, and the correctly suggested tags as T s \ T a. Precision, recall and F-measure are defined as follows: p ¼ jT s \ T aj jT sj ; r ¼ jT s \ T aj jT aj ; F ¼ 2pr p þ r : ð8Þ The final evaluation scores are computed by micro-averaging (i.e., averaging on resources of test set). We performed 5-fold cross- validation for each method for both datasets. In the experiments, the number of suggested tags Mr ranges from 1 to 10. 2 The dataset can be obtained from http://www.kde.cs.uni-kassel.de/bibsonomy/ dumps. 5.2. Comparing results 5.2.1. Baseline methods We select four algorithms as the baselines for comparison: Naive Bayes (NB) (Manning, Raghavan, & Schtze, 2008), k nearest neighborhood (kNN) (Manning et al., 2008), Content Relevance (CRM) Model (Iwata et al., 2009) and Tag Allocation Model (TAM) (Si et al., 2010). Our methods are denoted as WAM and MI. NB and kNN are two representative classification methods. NB is a simple generative model, which models the probability of each tag t given a description d as PrðtjdÞ/ PrðtÞ Y w2d PrðwjtÞ: ð9Þ PrðtÞ is estimated by the frequency of the documents annotated with the tag t. PrðwjtÞ is estimated by the frequency of the word w in the resource descriptions annotated with the tag t. kNN is a widely used classification method for tag suggestion, which anno- tates tags to a resource according to the annotated tags of similar resources measured using vector space models (Manning et al., 2008). CRM and TAM are selected to represent topic-based methods for tag suggestion. CRM is an LDA-based generative model. The number of latent topics K is the key parameter for CRM. In the experiments, we evaluated the performance of CRM with different K values, and here, we only present the best one obtained by set- ting K ¼ 1; 024. TAM is also a generative model that considers the words in descriptions as the topics to further generate tags for the resource. We set parameters for TAM as in Si et al. (2010). We compare the complexity of these methods. We denote the number of training iterations in CRM, TAM and WAM as I, and the number of topics in CRM as K. For the training phase, the com- plexity of NB is OðD �Nd �NaÞ; kNN is Oð1Þ; TAM is OðID �Nd �NaÞ; CRM is OðIKD �Nd �NaÞ; WAM is OðID �Nd �NaÞ,3 and MI is OðD �Nd �NaÞ. When sug- gesting for a given resource description with length Nd, the complex- ity of NB is OðNdTÞ; kNN is OðD �Nd �NaÞ; CRM is OðIKNdTÞ; TAM is OðINdTÞ; WAM is OðNdTÞ, and MI is OðNdTÞ. From the analysis, we can see that WAM and MI are relatively simple methods for both training and suggestion. This is especially valuable because WAM and MI also present good effectiveness for tag suggestion compared with other methods as we will shown later. 5.2.2. Parameter settings For WAM, we use GIZA++ (Och & Ney, 2003)4 with the IBM Model-1 to determine translation probabilities using description– annotation pairs for WAM. The experimental results of WAM are obtained by setting parameters as follows: label weighting type as TF-IDFt, length ratio d ¼ 1, harmonic factor k ¼ 0:5 and the type of sampling. 4 GIZA++ is freely available from code.google.com/p/giza-pp. The toolkit is widely used for word alignment in SMT. In this paper, we use the default setting of parameters for training. http://www.douban.com http://www.bibsonomy.org http://www.kde.cs.uni-kassel.de/bibsonomy/dumps http://www.kde.cs.uni-kassel.de/bibsonomy/dumps http://code.google.com/p/giza-pp 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Precision R ec al l MI WAM TAM CRM kNN NB (a) BOOK 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 Precision R ec al l MI WAM TAM CRM kNN NB (b) BIBTEX Fig. 2. Performance comparison between NB, kNN, CRM, TAM, WTM and MI for the two datasets. Table 3 Comparison of NB, kNN, CRM, TAM, WTM and MI results for the BOOK dataset when suggesting M ¼ 3 tags. We have used t-test to ensure the differences between other results and the best result (in bold) in each column of the table are statistically significant at p < 0:05. Method Precision Recall F-measure NB 0.271 0.302 0.247 ± 0.004 X. Chen et al. / Expert Systems with Applications 42 (2015) 1950–1959 1955 word importance scores as TF-IDFw. For MI, we set the tag co-occurrence threshold c ¼ 0 and the type of word importance scores as TF-IDFw. The values are used as default values by maximiz- ing the F-measure on a development set of 1000 instances from the website where BOOK dataset are obtained (not included int the BOOK dataset). The influence of parameters to WAM and MI can be found in Section 5.3. kNN 0.280 0.314 0.258 ± 0.002 CRM 0.292 0.323 0.266 ± 0.004 TAM 0.310 0.344 0.283 ± 0.001 WAM 0.368 0.452 0.355 ± 0.002 MI 0.422 0.493 0.397 ± 0.002 5.2.3. Experiment results and analysis In Fig. 2 we present the precision–recall curves of NB, kNN, CRM, TAM, WAM and MI for the two datasets. Each point of a pre- cision–recall curve represents different numbers of tags from M ¼ 1 (bottom right, with higher precision and lower recall) to M ¼ 10 (upper left, with higher recall but lower precision). The closer the curve to the upper right is, the better the overall perfor- mance of the method is. From Fig. 2, we observe the following: 1. The method based on MI consistently performs the best for both datasets. The method based on WAM achieves the second best performance for both datasets. These results indicate that our method is robust and effective for social tag suggestion. 2. The advantage of our method based on WAM is more significant on the BOOK dataset than the baseline method. The reason is that WAM has a good advantage in count information of tags compared with other baseline methods. 3. Although WAM has a good count information advantage for the BOOK dataset, MI is still better than WAM. The reason is that MI tries to translate a word to more tags, whereas the translation probabilities of a word in WAM always focus on only one or two tags. This leads to the result that the tags suggested by MI have a better coverage than WAM. We will present the translation probability table of words in the next subsection. 4. The average length of resource description is short in BIBTEX, which makes it difficult to determine the importance score of words, but even for the BIBTEX dataset with no count informa- tion of tags, our method still outperforms other methods. To further demonstrate the performance of our word transla- tion method and other baseline methods, in Table 3 we show the precision, recall and F-measure of NB, kNN, CRM, TAM,WAM and MI applied to the BOOK dataset when suggesting M ¼ 3 labels.5 Here we also show the variance of F-measure. In fact, MI achieves the best performance when M ¼ 2, where the F-measure of MI is 5 We elected to show this number because it is near the average number of labels for the BOOK dataset. 0.399, indicating an outperforming of both CRM (F ¼ 0:263) and TAM (F ¼ 0:277) by more than 10%. 5.2.4. An example In Table 4, we show the top 10 tags suggested by NB, CRM, TAM, WTM and MI applied to the book in Table 1. The number in brack- ets after the name of each method is the count of correctly suggested labels. The correctly suggested tags are marked in bold face. We elected not to show the kNN results because the tags suggested by kNN are totally unrelated to the book due to the insufficient finding of nearest neighbors. From Table 4, we observe that NB, CRM and TAM, as generative models, tend to suggest coarse-grained tags, such as ‘‘novel’’, ‘‘literature’’, ‘‘classic’’ and ‘‘France’’, and fail in suggesting fine- grained tags such as ‘‘Alexandre Dumas’’, ‘‘Count of Monte Cristo’’, ‘‘revenge’’ and ‘‘suspense’’. On the contrary, WAM and MI succeed in suggesting both the coarse-grained and the fine-grained tags related to the book. To find out how our model can suggest these fine-grained tags, we list four important words (using TF-IDFw as weighting metric) of the description and their corresponding tags with the highest translation probabilities in Tables 5 and 6. The values in brackets are the probability of tag t given word w; PrðtjwÞ. For each word, we eliminated the tags with a probability less than 0.05. We can see that the translation probabilities can map the words in descrip- tions to their semantically corresponding tags in annotations. Take the word ‘‘Count of Monte Cristo’’ in Table 5 for example, besides the tag identical to itself, it has a high probability to the tag ‘‘Alexander Dumas’’, which indicates the tag ‘‘Alexandre Dumas’’ is highly related to the word ‘‘Count of Monte Cristo’’. In fact, the word ‘‘Count of Monte Cristo’’ appears in 19 books (12 of them are the different versions of ‘‘Count of Monte Cristo’’ and other are the novels written by ‘‘Alexander Dumas’’) and 16 of them 0.22 0.24 0.26 0.28 0.3 0.32 0.34 0.36 0.38 0 1 2 3 4 5 6 7 8 9 F- m ea su re Number of Suggested Tags λ = 0.0 λ = 0.2 λ = 0.4 λ = 0.5 λ = 0.6 λ = 0.8 λ = 1.0 Fig. 3. F-measure of WAM versus the number of suggested tags for the BOOK dataset when harmonic factor k ranges from 0.0 to 1.0. Table 4 Top 10 tags suggested by NB, CRM, TAM, WAM and MI for the book in Table 1. NB (+6): novel, foreign literature, literature, history, Japan, classic, France, philosophy, America, biography CRM (+5): novel, foreign literature, literature, biography, philosophy, culture, France, British, comic, history TAM (+5): novel, sociology, finance, foreign literature, France, literature, biography, France literature, comic, China WAM (+7): novel, Alexandre Dumas, history, Count of Monte Cristo, foreign literature, biography, suspense, comic, America, France MI (+7): Alexandre Dumas, novel, Count of Monte Cristo, foreign literature, France, revenge, French literature, Liang Yusheng, martial arts, Comedie Humaine Table 5 Four important words (in bold face) in the book description in Table 1 and their corresponding tags with the highest translation probabilities in WAM. Count of Monte Cristo: Count of Monte Cristo (0.728), Alexandre Dumas (0.270), . . . Alexandre Dumas: Alexandre Dumas (0.966), . . . Revenge: foreign literature (0.168), classic (0.130), martial arts (0.123), Alexandre Dumas (0.122), . . . France: France (0.99), . . . Table 6 Four important words (in bold face) in the book description in Table 1 and their corresponding tags with the highest translation probabilities in MI. Count of Monte Cristo: Count of Monte Cristo (0.274), Alexandre Dumas (0.244), revenge (0.093), French literature (0.069), France (0.057), . . . Alexandre Dumas: Alexandre Dumas (0.352), France (0.105), French literature (0.069), Count of Monte Cristo (0.067), foreign literature (0.053), revenge (0.052), . . . Revenge: Liang Yusheng (0.154), revenge (0.127), martial arts (0.088), . . . France: France (0.309), France literature (0.069), . . . 6 GIZA++ restricts the values of length ratio within 19 ; 9 � � by setting the parameter axfertility = 10. From Fig. 4, we can see when d ¼ 10, the performance becomes uch worse as GIZA++ will cut off the sentences out of range. 1956 X. Chen et al. / Expert Systems with Applications 42 (2015) 1950–1959 are labeled with the tag ‘‘Alexander Dumas’’. This proves that our model can capture the semantic relation between words and tags. Note that ‘‘Count of Monte Cristo’’ and ‘‘Alexander Dumas’’ correspond to the title and the author of the book in Table 1, they may be easily derived from other metadata of the book (although it is not the case of our dataset). So it is more interesting to see that our model can suggest the fine-grained tags like ‘‘revenge’’. From Table 6, we can see that each of the words ‘‘Count of Monte Cristo’’, ‘‘Alexander Dumas’’ and ‘‘revenge’’ has a probability to the tag ‘‘revenge’’. The tag ‘‘revenge’’ is suggested jointly by combining the scores from these important words in the description. The abil- ity to suggest a tag jointly enables our model to suggest the tags that are not statistically significant or even not appear in the descriptions. Thus our model can solve the vocabulary gap problem. 5.3. Parameter influences 5.3.1. Parameter influences for WAM We explore the parameter influences on WAM for tag sugges- tion. The parameters include harmonic factor, length ratio, tag weighting types, and types of word translation power. When inves- tigating one parameter, we set the other parameters to the values inducing the best performance as mentioned in Section 5.2. Finally, we also investigated the influence of training data size for classifi- cation performance. In the experiments we found that WAM reveals similar trends for both the BOOK dataset and the BIBTEX dataset. Thus, we only show the experimental results for the BOOK dataset for analysis. 5.3.1.1. Harmonic factor. In Fig. 3 we investigate the influence of harmonic factor via the curves of F-measure of WAM versus the number of suggested tags on the BOOK dataset when harmonic factor k ranges from 0.0 to 1.0. As shown in Section 3.1.2, the harmonic factor k controls the proportion between model Prd2a and Pra2d . From Fig. 3, we observe that neither single model Prd2a (k ¼ 1:0) nor Pra2d (k ¼ 0:0) achieves the best performance. When the two models are combined by the harmonic mean, the performance is consistently better, especially when k ranges from 0.2 to 0.6. This is reasonable because IBM Model-1 constrains that only the term in the source language that can be aligned to multiple terms in target language, which makes the translation probability learned by a single model asymmetric. 5.3.1.2. Length ratio. Fig. 4 shows the influence of length ratios for WAM on the BOOK dataset. From the figure, we observe that the performance for tag suggestion is robust as the length ratio varies, except when the ratio breaks the default restriction of GIZA++ (i.e., d ¼ 10).6 5.3.1.3. Tag weighting types. The influence of two weighting types, TFt and TF-IDFt, on tag suggestion when M ¼ 3 on the BOOK dataset is shown in Table 7. TF-IDFt tends to select the tags more specific to m m 0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34 0.36 0.38 0 1 2 3 4 5 6 7 8 9 F- m ea su re Number of Suggested Tags δ = 10/1 δ = 10/3 δ = 10/5 δ = 1/1 δ = 1/2 δ = 1/5 Fig. 4. F-measure of WAM versus the number of suggested tags for the BOOK dataset when length ratio d ranges from 10=1 to 1=5. Table 7 Evaluation results for different tag weighting types for WAM when M ¼ 3 on the BOOK dataset. Weighting Precision Recall F-measure TFt 0.356 0.437 0.342 ± 0.002 TF-IDFt 0.368 0.452 0.355 ± 0.002 Table 8 Evaluation results for different methods for computing word importance scores to WAM when M ¼ 3 for the BOOK dataset. Weighting Precision Recall F-measure TF-IDFw 0.368 0.452 0.355 ± 0.002 TextRank 0.345 0.424 0.332 ± 0.002 Product 0.368 0.451 0.354 ± 0.002 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.1 0.2 0.3 0.4 0.5 0.6 0.7 R ec al l Precision 8,000 16,000 24,000 32,000 40,000 48,000 56,000 Fig. 5. Precision–recall curves of WAM when the training data size increases from 8000 to 56,000 on the BOOK dataset. 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Precision R ec al l γ=0 γ=1 γ=2 γ=5 γ=10 Fig. 6. Precision–recall curves of MI when the tag co-occurrence threshold c increases from 0 to 10 on the BOOK dataset. X. Chen et al. / Expert Systems with Applications 42 (2015) 1950–1959 1957 the resource, whereas TFt tends to select the most popular tags, because the latter does not consider global information (the IDFt part). Table 7 verifies the analysis, where TF-IDFt is slightly better than TFt. 5.3.1.4. Methods for computing word importance scores. In Table 8, we show the performance of WAM applied to the BOOK data- set with different methods for computing word importance scores. From the table, we can see that there is no significant difference between TF-IDFw and the product of TF-IDFw and TextRank, and TextRank performs the worst. This performance indicates that TextRank is less competitive when measuring word importance scores, as it does not take global information into consideration. 5.3.1.5. Training data size. We also investigated the influence of training data size for WAM. As shown in Fig. 5, we increased the training data size from 8000 to 56,000 step by 8000, and performed evaluation on 4000 resources. The figure shows that: (1) when the training data size is small (e.g., 8000), WAM can still achieve good performance, and (2) when the training data size increases, the performance will be improved, but the improvement speed will be slowed down as the training data size increases. This indicates that WAM does not require a huge size dataset to achieve good performance. 5.3.2. Parameter influences for MI The parameters of the MI-based method include tag co- occurrence threshold and types of word importance scores. When investigating one parameter, we set the other parameters to be the values inducing the best performance as mentioned in Section 5.2. Similar to WAM, we only present the experimental results for the BOOK dataset for analysis. 5.3.2.1. Tag co-occurrence threshold. Fig. 6 shows the influence of the tag co-occurrence threshold for MI on the BOOK dataset. We set the tag co-occurrence threshold c to different values. The figure shows that the MI-based method achieves the best performance with the BOOK dataset when c ¼ 0. This indicates that the set of tags that have low co-occurrences with a word not only contains noisy tags, but also contains proper tags that need to be suggested by the word. 5.3.2.2. Methods for computing word importance scores. In Table 9, we show the performance of MI applied to the BOOK dataset with different methods for computing word importance scores. From the table, we can see that for MI, TF-IDFw performs the best and TextRank performs the worst. It is similar to WAM, and these Table 9 Evaluation results for different methods for computing word importance scores with MI when M ¼ 3 for the BOOK dataset. Weighting Precision Recall F-measure TF-IDFw 0.422 0.493 0.397 ± 0.002 TextRank 0.393 0.461 0.370 ± 0.002 Product 0.407 0.475 0.382 ± 0.002 1 2 3 4 5 6 7 8 9 10 0.24 0.26 0.28 0.3 0.32 0.34 0.36 0.38 0.4 0.42 Number of Suggested Tags F− m ea su re α=0.0 α=0.1 α=0.3 α=0.5 α=0.7 α=0.9 Fig. 7. F-measure of MI versus the number of suggested tags for the BOOK dataset when the self-translation parameter a ranges from 0.0 to 0.9. Table 10 The evaluation results for emphasizing the self-Translation probability on MI with different methods for computing word importance scores when M ¼ 3 with the BOOK dataset. Weighting Precision Recall F-measure TF-IDFw 0.385 0.472 0.371 ± 0.001 TextRank 0.344 0.423 0.332 ± 0.002 Product 0.374 0.457 0.360 ± 0.001 1958 X. Chen et al. / Expert Systems with Applications 42 (2015) 1950–1959 results indicate that TextRank is less competitive for measuring word importance scores, as it does not take global information into consideration. By analyzing the influences of parameters on WAM and MI, we find that the word translation model is robust to parameter variations. 5.4. Performance of emphasizing self-translation probability In Fig. 7 we investigate the influence of self-translation param- eter via the curves of F-measure of MI versus the number of sug- gested tags on the two datasets when the self-translation parameter a ranges from 0.0 to 0.9. As shown in Section 3.3, the parameter a controls the self-translation probabilities. From Fig. 7, we observe that MI achieve the best performance when alpha ¼ 0:2 for the BOOK dataset, whereas alpha ¼ 0:4 on BIBTEX dataset. These results indicate that for both dataset, for one word, we need to emphasize the translation probability to itself. This is reasonable because without self-translation, a important tag may not be suggested as it does not suggest itself using the translation probabilities. We also see that for different dataset, the self-translation parameter is not a constant and varies from the need of emphasizing the words in the current document. Finally, we tested the performance of emphasizing self- translation probability for WAM with different word translation methods for the BOOK dataset. As shown in Table 10, emphasizing the self-translation probability improves the performance of WAM (in Table 8) as applied to the BOOK dataset when using TF-IDFw and the product as the methods for computing the word trigger powers, but decays when using TextRank. This result verifies that TF-IDFw is the best method to measure word importance scores for WAM, which indicates that emphasizing the tags appearing in the descriptions may enhance the classification performance of the word translation method. However, the performance when emphasizing the self- translation probability on the BIBTEX dataset decays much compared with WAM. The F-measure of emphasizing the self-translation probability is only F ¼ 0:229 compared with WAM F ¼ 0:267. The main reason for the decay is that the average length of descriptions in the BIBTEX dataset is too short to provide sufficient information to precisely emphasize tags, and usually emphasize wrong tags and drop correct tags. The experimental results for emphasizing the self-translation probability suggest that, we have to analyze the characteristics of the tag suggestion systems to decide whether to emphasize the tags that appear in the corresponding descriptions. It is also worth investigating the problem when combining with collaboration- based methods for social tag suggestion. 6. Conclusions In this paper, we present a new perspective on social tag suggestion and propose two methods to estimate translation prob- abilities between words in descriptions and tags. One method is the word alignment model in statistical machine translation and the other method is mutual information. Based on the transla- tion probabilities between words and tags, we propose the word translation method. The experiments revealed that our method is effective and efficient for social tag suggestion compared with other baseline methods. There are several open issues for further investigation: 1. Our model focuses on suggesting social tags according to the resource descriptions. We will take advantages of more social information such as user information to improve the perfor- mance of social tag suggestion. 2. Other metadata of resources (author, title, images and videos) can be also taken into account to improve the performance of social tag suggestion. 3. Our model is a supervised model which requires a large collec- tion of annotated resources. We will explore to use large-scale unlabeled text corpora to estimate the translation probabilities between words, which could be used to enhance the estimation of translation probabilities between words in descriptions and tags in annotations. 4. In this paper we suggest each tag in isolation, without consider- ing the correlations between tags. We will investigate the hier- atical structure and the semantic relatedness between tags to regularize the granularity of tags. Acknowledgments This work is supported by the National Natural Science Founda- tion of China (NSFC) under Grant Nos. 61170196 and 61202140. The authors would like to thank Peng Li for his insightful suggestions. References Blei, D., & Jordan, M. (2003). Modeling annotated data. In Proceedings of SIGIR (pp. 127–134). Blei, D., Ng, A., & Jordan, M. (2003). Latent dirichlet allocation. JMLR, 3, 993–1022. http://refhub.elsevier.com/S0957-4174(14)00625-3/h0010 X. Chen et al. / Expert Systems with Applications 42 (2015) 1950–1959 1959 Brown, P., Pietra, V., Pietra, S., & Mercer, R. (1993). The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2), 263–311. Bundschus, M., Yu, S., Tresp, V., Rettinger, A., Dejori, M., & Kriegel, H. (2009). Hierarchical bayesian models for collaborative tagging systems. In Proceedings of ICDM (pp. 728–733). Fujimura, S., Fujimura, K., & Okuda, H. (2008). Blogosonomy: Autotagging any text using bloggers’ knowledge. In Proceedings of WI (pp. 205–212). Herlocker, J., Konstan, J., Borchers, A., & Riedl, J. (1999). An algorithmic framework for performing collaborative filtering. In Proceedings of SIGIR (pp. 230–237). Herlocker, J., Konstan, J., Terveen, L., & Riedl, J. (2004). Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems, 22(1), 5–53. Heymann, P., Ramage, D., & Garcia-Molina, H. (2008). Social tag prediction. In Proceedings of SIGIR (pp. 531–538). Iwata, T., Yamada, T., & Ueda, N. (2009). Modeling social annotation data with content relevance using a topic model. In Proceedings of NIPS (pp. 835–843). Jaschke, R., Marinho, L., Hotho, A., Schmidt-Thieme, L., & Stumme, G. (2008). Tag recommendations in social bookmarking systems. AI Communications, 21(4), 231–247. Karimzadehgan, M., & Zhai, C. (2010). Estimation of statistical translation models based on mutual information for ad hoc information retrieval. In Proceedings of SIGIR (pp. 323–330). Katakis, I., Tsoumakas, G., & Vlahavas, I. (2008). Multilabel text classification for automated tag suggestion. ECML PKDD discovery challenge 2008, p. 75. Krestel, R., Fankhauser, P., & Nejdl, W. (2009). Latent dirichlet allocation for tag recommendation. In Proceedings of ACM RecSys (pp. 61–68). Lam, X. N., Vu, T., Le, T. D., & Duong, A. D. (2008). Addressing cold-start problem in recommendation systems. In Proceedings of the 2nd international conference on ubiquitous information management and communication (pp. 208–211). ACM. Lee, S., & Chun, A. (2007). Automatic tag recommendation for the web 2.0 blogosphere using collaborative tagging and hybrid ann semantic structures. In Proceedings of WSEAS (pages 88–93). Lin, D. (1998). An information-theoretic definition of similarity. In Proceedings of ICML (Vol. 98, pp. 296–304). Liu, Z., Chen, X., & Sun, M. (2011). A simple word trigger method for social tag suggestion. In Proceedings of the conference on empirical methods in natural language processing (pp. 1577–1588). Association for Computational Linguistics. Liu, Z., Huang, W., Zheng, Y., & Sun, M. (2010). Automatic keyphrase extraction via topic decomposition. In Proceedings of the 2010 conference on empirical methods in natural language processing (pp. 366–376). Association for Computational Linguistics. Manning, C., Raghavan, P., & Schtze, H. (2008). Introduction to information retrieval. New York, NY, USA: Cambridge University Press. Mihalcea, R., & Tarau, P. (2004). TextRank: Bringing order into texts. In Proceedings of EMNLP (pp. 404–411). Mishne, G. (2006). Autotag: A collaborative approach to automated tag assignment for weblog posts. In Proceedings of WWW (pp. 953–954). Och, F., & Ney, H. (2003). A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1), 19–51. Ohkura, T., Kiyota, Y., & Nakagawa, H. (2006). Browsing system for weblog articles based on automated folksonomy. In Proceedings of WWW. Page, L., Brin, S., Motwani, R., & Winograd, T. (1998). The pagerank citation ranking: Bringing order to the web. Technical report, Stanford Digital Library Technologies Project. Rendle, S., Balby Marinho, L., Nanopoulos, A., & Schmidt-Thieme, L. (2009). Learning optimal ranking with tensor factorization for tag recommendation. In Proceedings of KDD (pp. 727–736). Resnick, P., & Varian, H. (1997). Recommender systems. Communications of the ACM, 40(3), 56–58. Salton, G., & Buckley, C. (1988). Term-weighting approaches in automatic text retrieval. Information processing and management, 24(5), 513–523. Si, X., Liu, Z., & Sun, M. (2010). Modeling social annotations via latent reason identification. IEEE Intelligent Systems. Si, X., & Sun, M. (2009). Tag-LDA for scalable real-time tag recommendation. Journal of Computational Information Systems, 6(1), 23–31. Wang, M., Ni, B., Hua, X.-S., & Chua, T.-S. (2012). Assistive tagging: A survey of multimedia tagging with human–computer joint exploration. ACM Computing Surveys (CSUR), 44(4), 25. Xu, Z., Fu, Y., Mao, J., & Su, D., 2006. Towards the semantic web: Collaborative tag suggestions. In Collaborative web tagging workshop at WWW2006. http://refhub.elsevier.com/S0957-4174(14)00625-3/h0015 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0015 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0015 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0035 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0035 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0035 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0050 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0050 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0050 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0070 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0070 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0070 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0070 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0085 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0085 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0085 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0090 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0090 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0090 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0090 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0095 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0095 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0110 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0110 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0130 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0130 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0135 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0135 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0140 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0140 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0145 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0145 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0150 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0150 http://refhub.elsevier.com/S0957-4174(14)00625-3/h0150 Estimating translation probabilities for social tag suggestion 1 Introduction 2 Related work 3 Learning translation probabilities 3.1 Word alignment model (WAM)-based approach 3.1.1 Preparing description–annotation pairs for WAM 3.1.2 Learning translation probabilities for WAM 3.2 Mutual information (MI)-based approach 3.3 Emphasize self-translation probability 4 Suggesting tags with translation probabilities 5 Experiments 5.1 Datasets and evaluation metrics 5.1.1 Evaluation metrics 5.2 Comparing results 5.2.1 Baseline methods 5.2.2 Parameter settings 5.2.3 Experiment results and analysis 5.2.4 An example 5.3 Parameter influences 5.3.1 Parameter influences for WAM 5.3.1.1 Harmonic factor 5.3.1.2 Length ratio 5.3.1.3 Tag weighting types 5.3.1.4 Methods for computing word importance scores 5.3.1.5 Training data size 5.3.2 Parameter influences for MI 5.3.2.1 Tag co-occurrence threshold 5.3.2.2 Methods for computing word importance scores 5.4 Performance of emphasizing self-translation probability 6 Conclusions Acknowledgments References 
work_6lnqupedgjfqpc2rrrox7k5aaa	----	King’s Research Portal DOI: 10.1016/j.future.2017.05.037 Document Version Peer reviewed version Link to publication record in King's Research Portal Citation for published version (APA): Cioara, T., Anghel, I., Salomie, I., Barakat, L., Miles, S., Reidlinger, D., ... Pop, F. (2017). Expert system for nutrition care process of older adults. DOI: 10.1016/j.future.2017.05.037 Citing this paper Please note that where the full-text provided on King's Research Portal is the Author Accepted Manuscript or Post-Print version this may differ from the final Published version. If citing, it is advised that you check and use the publisher's definitive version for pagination, volume/issue, and date of publication details. And where the final published version is provided on the Research Portal, if citing you are again advised to check the publisher's website for any subsequent corrections. General rights Copyright and moral rights for the publications made accessible in the Research Portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognize and abide by the legal requirements associated with these rights. •Users may download and print one copy of any publication from the Research Portal for the purpose of private study or research. •You may not further distribute the material or use it for any profit-making activity or commercial gain •You may freely distribute the URL identifying the publication in the Research Portal Take down policy If you believe that this document breaches copyright please contact librarypure@kcl.ac.uk providing details, and we will remove access to the work immediately and investigate your claim. Download date: 05. Nov. 2018 https://doi.org/10.1016/j.future.2017.05.037 https://kclpure.kcl.ac.uk/portal/en/publications/expert-system-for-nutrition-care-process-of-older-adults(c98d37e3-a92f-4217-931b-98ce90fb5018).html Accepted Manuscript Expert system for nutrition care process of older adults Tudor Cioara, Ionut Anghel, Ioan Salomie, Lina Barakat, Simon Miles, Dianne Reidlinger, Adel Taweel, Ciprian Dobre, Florin Pop PII: S0167-739X(17)31105-6 DOI: http://dx.doi.org/10.1016/j.future.2017.05.037 Reference: FUTURE 3485 To appear in: Future Generation Computer Systems Received date : 31 May 2016 Revised date : 25 November 2016 Accepted date : 28 May 2017 Please cite this article as: T. Cioara, I. Anghel, I. Salomie, L. Barakat, S. Miles, D. Reidlinger, A. Taweel, C. Dobre, F. Pop, Expert system for nutrition care process of older adults, Future Generation Computer Systems (2017), http://dx.doi.org/10.1016/j.future.2017.05.037 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. http://dx.doi.org/10.1016/j.future.2017.05.037 Expert System for Nutrition Care Process of Older Adults Tudor Cioara a, Ionut Anghel a, Ioan Salomie a, Lina Barakat b, Simon Miles b, Dianne Reidlinger c and Adel Taweel d Ciprian Dobre e, Florin Pop e* aTechnical University of Cluj-Napoca, Memorandumului 28, 400 114 Cluj-Napoca, Romania E-mail: {tudor.cioara, ionut.anghel, ioan.salomie}@cs.utcluj.ro bKing's College London, Strand, London, WC2R 2LS, United Kingdom E-mail: {lina.barakat, simon.miles}@kcl.ac.uk cBond University, Queensland Area, Australia E-mail: dreidlin@bond.edu.au dBirzeit University, Birzeit, Palestine E-mail: ataweel@birzeit.edu eUniversity Politehnica of Bucharest, Romania E-mail: {ciprian.dobre, florin.pop}@cs.pub.ro Abstract. This paper presents an expert system for a nutrition care process tailored for the specific needs of elders. Dietary knowledge is defined by nutritionists and encoded as Nutrition Care Process Ontology, and then used as underlining base and standardized model for the nutrition care planning. An inference engine is developed on top of the ontology, providing semantic reasoning infrastructure and mechanisms for evaluating the rules defined for assessing short and long term elders’ self -feeding behaviors, to identify unhealthy dietary patterns and detect the early instauration of malnutrition. Our expert system provides personalized intervention plans covering nutrition education, diet prescription and food ordering adapted to the older adult’s specific nutritional needs, health conditions and food preferences. In-lab evaluation results are presented proving the usefulness and quality of the expert system as well as the computational efficiency, coupling and cohesion of the defined ontology. Keywords: Expert system, Nutrition care, Inference engine, Malnutrition, Ontology 1. Introduction Over the past decade, healthcare systems have under-gone a paradigm shift from being a solely treatment-based focus towards a more personalized, person-centred, and prevention- oriented approach. Such change is driven by the increasing health cost burden to non-sustainable limits, of treatment-based systems, due to the overall ageing of the population, sedentary life-styles and poor nutrition habits, which has led to the increased proliferation of chronic illnesses (e.g. diabetes) (Antos et al., 2013). Studies have shown that in Europe more than 15% of the older population is affected by poor nutrition including malnutrition caused by age-related risk factors such as sensory changes (taste, smell, eye sight), poor dental health, lack of transportation, physical difficulties, forgetfulness and other issues (Sieber, 2010). Malnutrition is defined as a state of nutrition in which a deficiency, excess or imbalance of energy, protein, and other nutrients causes measurable adverse effects on body form (body shape, size and composition), function, and clinical outcome (Elia, 2001). According to the British Association for Parenteral and Enteral Nutrition (Elia&Russell, 2008), malnutrition affects over 3 million people in the UK alone, and of these, about 1.3 million are over the age of 65. If unman-aged, malnutrition may significantly impact on the older person’s health (such as exacerbation of chronic conditions, delayed recovery from illness, etc.), thus causing significant increases in related healthcare costs. In fact, the cost associated with malnutrition in Europe is estimated to amount to a * Corresponding author staggering 170 billion Euro each year (Ljungqvist and Man, 2009). The rapid identification of malnutrition and early prevention through the provision of nutritional assistance to the elderly would thus help to avoid such high public health costs, and enhance both the mental and physical conditions of older adults including their quality of life. It is generally agreed that the best strategy for malnutrition prevention is to lead a healthy lifestyle which can be enacted through a personalized nutrition care process. In Europe, it has been estimated that 77% of the disease burden can be ac-counted for disorders related to unhealthy lifestyle and furthermore, 70% of stroke and colon cancer, 80% of coronary heart disease, and 90% of type II diabetes could be prevented and managed through nutrition care (Brown, 2013). Lifestyle behavioural factors (poor nutrition habits, physical inactivity, tobacco and alcohol use) are classified as modifiable indirect risk factors which can be influenced by individuals and if not managed, could lead to metabolic and physiological changes including high blood pressure, high blood glucose, overweight, obesity and high cholesterol, which all represent direct factors for the development of chronic diseases (Willett et al., 2006). At the same time targeting obesity and overweight, promoting healthy eating, physical activity, smoking/alcohol cessation have been shown to reduce the incidence of “type 2” diabetes (Knowler, et al., 2002). In this context, advances in the ICT (Information and Communication Technology) sector have made feasible the development of solutions for nutrition care through prevention and self-management. Most contemporary nutrition management solutions aim at offering nutritional information and advice for popular commercial products. Their healthy lifestyle plans are targeting weight loss and do not consider the specific nutritional and physiological problems of the older adults, or the detection and prevention of malnutrition. This paper contributes towards achieving these goals by proposing an expert system for nutrition care process tailored for the specific needs of older adults. Led by nutritionists, we first investigate benchmarks and nutritional guidelines to evaluate diets based on published recommendations suitable for elders, as well as identify suitable nutrition problems and interventions for the elders. The older adults nutrition related information (provided by nutritionists) is utilized to construct a semantic dietary knowledge encoded as ontology, named the Nutrition Care Process Ontology. It is composed of four sub- ontologies: Nutrition Monitoring Ontology, Nutrition Assessment Ontology, Nutrition Problem Identification Ontology, and Nutrition Intervention Ontology. The Nutrition Monitoring Ontology defines and semantically represents information regarding the older adult relevant information for assessing their nutrition and self-feeding behaviour. The Nutrition Assessment Ontology covers information facilitating the assessment of older adult’s food intake and converts these to associated nutrient values. The Nutrition Problem Identification Ontology captures potential nutrition related problems and associated symptoms. Finally, the Nutrition Intervention Ontology models suitable intervention actions for identified nutrition problems and unhealthy behaviour. A nutrition inference engine is developed on top of the Nutrition Care Process Ontology, to provide a semantic reasoning infrastructure for evaluating the rules defined for assessing short term and long term older adult’s self-feeding behaviours, for identifying un-healthy dietary patterns and proactively detecting the early instauration of malnutrition and for helping nutritionists to define personalized intervention plans. The rest of the paper is organized as follows. Section 2 discusses related work. The proposed Nutrition Care Process Ontology and the rules for assessing unhealthy behaviours are detailed in Section 3, while Section 4 presents a corresponding use case validation. Finally, Section 5 concludes the paper. 2. Related work Most of the state of the art of diet management models and services aim to provide nutrients information for popular products and to define customized weight loss and healthy lifestyle plans (see CaloriesCount web applications). Although these models are intended to be used by all kinds of people regardless of age, the specific problems of older adults regarding nutrition and self-feeding behaviour are not specifically considered. Sensory changes, side effects of medication, physical difficulty or forgetfulness can cause nutrition problems for older adults which cannot be solved by using simple weight loss plans. Main challenges addressed by existing research efforts focus on monitoring food intake and nutritional habits, definition of appropriate knowledge to assess unhealthy behaviours and the development of expert systems which may take nutrition intervention decisions based on the monitored nutrition data and knowledge base. Defining and representing nutrition related knowledge diet is fundamental for allowing ICT systems to reason about it, and to provide personalized diet intervention and feedback. As the knowledge base become more structured rule based reasoning can be employed to assess the nutrition related behaviour of a person (Vassányi et al., 2014). While the use of ontologies has proven to be effective in establishing standard models, taxonomies, vocabularies and domain terminology (Valencia- García et al., 2008; Rivero et al., 2013) few approaches use the ontologies for evaluating nutrition related behaviour, and the provision of intervention plans is mostly limited to the management of some chronic conditions such as diabetes (Quinn et al., 2015; Lee et al., 2008). In (Tumnark et al., 2013) ontology-based personalized dietary recommendation for weightlifting to assist athletes in meeting their nutritional requirements is developed and used to provide personalized daily menus. The provided ontology is limited to weightlifting nutritional knowledge while the inference engine is not suitable for complex reasoning processes considering various age related factors. In (Quinn et al., 2015) the authors present a conceptual architecture for web-based personalized patient education experience having as central element the patient ontological model which captures knowledge related to medical conditions, physical activities and educational background. Ontology based daily menu assistance system for suggesting daily menus based on reference values of daily calories of a person is the subject of (Fudholi et al., 2009). However, the developed fuzzy ontology is limited to some food related criteria such as price, rate, vote and taste but the use of daily calories benchmark values makes it suitable for losing weight based on low calories intervention plans. In (Lee et al., 2008) an ontology model for diabetic food recommendation is proposed containing Taiwanese food ontology and a set of personal food ontologies. An intelligent agent based on a fuzzy inference engine is developed and used to create a meal plan according to a person’s lifestyle and health needs for diabetes as a chronic condition. The results show great potential in supporting the dietician efforts but the main disadvantage is that the ontology focuses on Taiwanese food only and lacks the reliability of fuzzy reasoning. In (Snae and Brückner, 2008) a counselling system for menu planning in a restaurant is developed. The system is based on a food ontology which contains specifications of ingredients, substances, nutrition facts and recommended daily intakes, an inference system based on the defined ontology, and a web interface for dieticians. The system’s disadvantage is its static nature in not being able to adapt the provided menus and recipes for specific nutritional profiles, for diabetics for example, and the lack of an automated assessment of dietary plans. The PIPS (Personalized Information Platform for Health and Life Services) food ontology (Dominguez et al., 2006) is a food taxonomy that uses the Eurocode food coding used by software agents to generate personalized advice for people with type II diabetes. Our approach for nutritional assessment uses PIPS ontology for assessing the behaviour of older adults focusing on factors relevant to nutrition. Nutrition expert systems have proven to be effective for offering advice and menu planning out of nutritional knowledge for preventing malnutrition (Quinn et al., 2015; Lee et al., 2008; Tumnark et al., 2013; Snae and Brückner, 2008; Vassányi et al., 2014). In (Espín et al., 2015) the authors describe a nutritional recommender system, for helping older adults to follow dietary plans that are based on nutritionists’ guidelines. The proposed system uses a reasoning process based on SWRL (Semantic Web Rule Language) rules upon nutritional and user profile ontologies to generate recommendations through semantic similarity measures. Similarly in (Quinn et al., 2015) semantic rules are used to infer associations between ontology concepts in order to create educational content for health education. (Al- Dhuhli et al., 2013) propose an expert system for evaluating nutrition conditions and for suggesting a specific type of food or the required time to exercise each day. The knowledge used as input to the system is based on if-then statements and decision tables and is represented using the e2go freeware rule-based shell. Although the system is evaluated using experts and different user groups, it does not involve complex reasoning processes, being limited to generating simple outputs such as body mass index values, nutrient values or weight information. In (Vassányi et al., 2014), a system which analyses dietary logs to assess the diet of an individual and to construct a personalized menu is developed. The system is based on predefined rules for foods and dishes employing genetic algorithms to calculate the fitness of candidate solutions using personalized target values of various nutrients. The solution was developed for mobile devices, collects nutrition information only through user interaction which is not very accurate and does not represent the knowledge in computer interpretable manner thus no advanced reasoning and inference is possible. In (van der Merwe et al., 2014) the authors propose an expert system based on a rule- based inference engine and multi-objective linear programming models for dietary recommendations. The system generates an eating-plan that conforms to the nutritionist specifications, considering end-users’ preferences and food item costs. A knowledge base was developed consisting of predefined foods grouped according to the proportions of carbohydrate, protein and fat. The inference engine uses if-then production rules while linear programming is applied to select foods from the knowledge base and construct the overall personalized diet. The system does not involve monitoring the patient following an intervention plan nor does allow defining complex inference rules due to the if-then approach. In (Arens-Volland et al., 2009) an ICT based system for the detection and prevention of food allergies as well as for diet management is proposed. The authors propose an expert system for mobile devices for analysis of nutrition-based allergies based on end-user food diaries filled using a barcode scanner, electronic patient records for identifying the specific allergies and food product databases. The system helps the end-users to avoid stressful and time consuming blood and skin tests but cannot take into consideration the food that is not labelled in a specific way, particularly cooked food. The opportunities offered by an ICT based learning environment to deliver nutrition health learning and education for children are investigated in (Raiha, 2013). The end-user eating habits and nutrition knowledge are recorded by means of questionnaires. The study revealed that the use of the ICT-based learning environment developed the pupils’ critical reading skills for nutrition and health information that can be found on the Internet. In (Atienza et al., 2008) 27 healthy adults aged over 50 were involved in an 8-week nutrition intervention study using mobile devices. They used PDAs to monitor their vegetable and whole-grain intake levels twice per day and to provide daily individualized age-appropriate feedback, goal- setting, and support. The results have shown that the participants reported a significantly greater increases in vegetable servings (1.5-2.5 servings/day), as well as a trend toward greater intake of dietary fibre from grains (3.7-4.5 servings/day). A prototype expert system for human nutritional diagnosis based on service oriented architecture was proposed in (Quesada and Jenkins, 2013). The system performs a nutrition diagnosis using Body Mass Index (BMI) evaluation of the end-users and sends notifications to the end-users regarding the meal time and suggest healthy food places nearby. Our proposed solution takes a different approach and is the first (to the best of our knowledge) that attempts semantic modelling of nutrition related knowledge covering all phases of the older adult’s nutrition care process. In addition, the proposed inference engine allows the automatic assessment of early instauration of nutrition related problems, and of older adults’ short and long term unhealthy behaviours, which may lead to malnutrition. Unlike other approaches our expert system will not provide only menu planning but it will go one step further to provide personalized intervention plans that includes nutrition education, diet recommendations and food ordering adapted to the older adult’s specific nutrition needs, health and preferences. 3. Nutrition care process ontology There are a number of standardized models of the nutrition care planning process (see the Academy of Nutrition and Dietetics’ Nutrition Care Process and Model). The British Dietetic Association has a similar version (see the Model and Process for Nutrition and Dietetic Care), and the same applies to countries such as US, Canada and Australia. The international dietetics community is working to develop a standardized language for documenting the nutrition care process. Within Europe, take up by the profession has been slower, but there is a vision for the adoption of a standardized care process to be used by all dietitians, and in the training of dietitians, by 2020 (see FAD Professional Practice Committee report). Nutrition care planning is a cyclical and systematic process, with each stage dependent on those preceding it. The four stages are: (i) monitoring, (ii) assessment, (iii) nutrition problem identification and finally (iv) intervention implementation. The first stage (monitoring) feeds new data into the second stage (assessment), i.e. it triggers a re- assessment to determine if the previous plan is still relevant or if changes need to be made to reflect the progression of previously identified, or the development of new, nutritional problems. Factors influencing the implementation of the intervention (e.g. social and financial changes) are also identified during the monitoring stage. In practice, screening in the community or by non-nutrition specialists triggers a referral to a nutritionist or dietitian for more in depth assessment and a ‘nutrition care plan’ is then developed. Hence screening is almost a ‘pre-step’ of nutrition care process, which we have decided to incorporate in the monitoring phase, but does not need to be executed each time the nutrition care process runs. Our aim is to provide an appropriate computational representation for the clinically informed diet knowledge and to construct a knowledge base on which reasoning and query processes can be executed for dietary assessment of older adults. This will enact the development of a nutrition expert system which will automate the nutrition care process by providing nutrition assistance, advices, explanations, nutrition evaluation and intervention and rational decision support without emotional overhead. Also interactions of nutritionists with the older adults are minimized making the nutrition care process more efficient and less invasive. A common contemporary approach to represent knowledge of a specific domain (in this case the older adult’s nutrition care process domain) is in the form of ontologies (Rodríguez et al., 2014). The main idea is to establish standard models, taxonomies, vocabularies and domain terminology. These can be further used to infer new knowledge and relationships in the modelled domain. Here, ontologies are used to capture nutrition related knowledge resulting in the Nutrition Care Process Ontology, composed from four sub- ontologies: Nutrition Monitoring Ontology, Nutrition Assessment Ontology, Nutrition Problem Identification Ontology and Nutrition Intervention Ontology. These sub- ontologies are detailed next. 3.1. Nutrition Monitoring Ontology Nutrition Monitoring Ontology (see Fig. 1) semantically annotate data describing the older adult self-feeding behaviour such as personal information (such as age and gender), lifestyle information (including living arrangements, financial context, and skills and equipment for food preparation), social information (e.g. supports already in place including family and wider social network), and physical activity information. The rest of the assessment is based on the “ABCD approach” (Knox et al., 2013) (which includes collection of anthropometry, biochemistry, clinical and dietary information) as outlined below. Anthropometry is the measurement of different aspects of the human body. For older adults, the most relevant data to collect will be their measured weight and height, although proxy or estimated measures can be used particularly for height in case where it may be more difficult to measure due to specific health conditions. These include ulna length and demi span as measures that can be used where measured height is not possible. This information needs to be interpreted to inform the assessment; commonly body mass index (BMI, weight/ height2) and any percentage change in weight (previous – current weight/ previous weight x 100) in the preceding 6 months are calculated. Biochemistry information is collected to identify those individuals requiring intensive follow up (individualised care by a nutrition professional) and flag up where dietary modification might be targeted. It is unlikely that older people living in the community will be able to provide blood chemistry results for this assessment factor. However, those being seen by a nutritionist or dietitian with a referral from a doctor may require the system to record this information. Typically in the older age group blood chemistry results may include blood lipids (cholesterol, triglycerides), blood glucose, or full blood counts (which include white cell count, red blood cell count, and haemoglobin levels), and if diabetes is present, glycosylated haemoglobin (HbA1C). These tests must be ordered by a doctor but are useful adjuncts to the nutritional assessment; an evaluation of whether dietary modification might help to address values outside of the normal range can be then made (dietary modification may be instead of or in conjunction with medication). They are also very useful outcome indicators to evaluate nutrition and dietary care. The absence of biochemistry data does not preclude further assessment. Clinical information collected includes details of: (i) health problems including normal ageing processes likely to affect food intake e.g. poor dentition, problems with digestion and information on mobility/function (activities of daily life if available); (ii) medical diagnoses and/or information on treatment from other health professionals such as physiotherapists, speech and language therapists (who may prescribe food and drink texture modifications), home care nurses etc.; (iii) use of prescribed and over the counter medications and supplements (if available from the general practitioner). Dietary and food preference information collected includes likes and dislikes of individual, unsuitable foods (due to texture modification needs, allergies, intolerances, personal or religious and cultural exclusions) and details of texture modification prescriptions (e.g. soft foods, pureed food, thickened drinks). This information ensures tailoring to the individual in the planning and implementation of the intervention. 3.2. Nutrition Assessment Ontology The objective of the Nutrition Assessment Ontology is to semantically represent food and nutritional information and to enact the assessment of older adult food intake and associated nutritional values. To achieve this objective the nutritionists are interested in nutrients intake as well as at food consumption data, thus the ontology stores data about different types of foods Figure 1: Nutrition Monitoring Ontology and their associated nutritional values. The Food concept is at the root of this ontology, and all the other concepts inherit its properties, allowing the description of an aliment in terms of its nutrients. There are two types of foods modelled in the ontology: basic food and combined food. The basic food taxonomy classifies food items according to the PIPS (Personalized Information Platform for Health and Life Services) food ontology (Dominguez et al., 2006), which has been imported and used in our assessment ontology. The PIPS ontology constructs the food taxonomy using the Eurocode 2 food coding which provides a mono-hierarchical classification of foods according to groups and subgroups that are useful in dietary studies. Specifically, the PIPS ontology defines 14 major food groups: 1) Milk, milk products and dishes, 2) Eggs, 3) Meat, 4) Poultry, 5) Fish, molluscs, reptiles and crustaceans, 6) Oils and fats, 7) Grains, 8) Pulses, seeds, nuts and kernels, 9) Vegetables, 10) Fruits, 11) Sugar, 12) Beverages (except milk), 13) Miscellaneous, soups, sauces and 14) Foods for special nutritional use. The combined foods are prepared foods which are based on a recipe of basic foods (see Fig. 2). The recipe concept represents a collection of basic foods in different proportions or quantities and each recipe is associated with a combined food. There are different types of combined food: flavoured food, regional cuisine food and dish, and each of them has different subtypes. For example, flavoured food has the following subtypes: sour food, bitter food, spicy food, salted food and sweet food. Regional cuisine food can be further classified into: Italian cuisine food, Spanish cuisine food, etc. Finally, a dish could be a starter dish, a main course dish, or a dessert dish. Another important part of this ontology, for assessing the nutritional values of the older adult food intake, is represented by food quantities. There are different types of food quantities such as slice, cup, bowl, dish, pound, piece, box, bag, carton, jar, etc. We have defined the FoodQuantity concept and associated relation for modelling their conversion in grams. Nutritional intake will be calculated from the food the older consumes during the day and will include the nutrient values such as: energy (kcal or kJ), fat (g), carbohydrates (g), fatty acids (g), protein (g), potassium (mg), calcium (mg), sodium (mg), vitamin D (ug), alcohol (g), and water (g). These values are important because malnutrition is prevented by maintaining them within ideal limits. The nutrient values for each type of food are extracted from the McCance and Widdowson's food composition tables, which provide the description of around 3400 food items. Based on these nutritional values, the ontology individuals that represent specific food items are created and classified in the food ontology. For each food item, the main properties extracted are: the food code (Eurocode format), name and description, number of kilocalories, the amount of water, protein, fat, and other nutrients per 100 g. Three tables from McCance and Widdowson are important for creating the ontology instances: inorganics, proximates and vitamins. The inorganics table contains information on inorganic nutrients per 100 g such as the quantity of sodium (mg), potassium (mg), calcium (mg), magnesium (mg), phosphorus (mg), iron (mg), copper (mg), zinc (mg), chloride (mg), manganese (mg), selenium (ug), and iodine (ug). The proximates table contains information about the macro nutrients and their sub-components per 100 g such as: water (g), total nitrogen (g), protein (g), fat (g), carbohydrate (g), energy (kcal), energy (kJ), starch (g), oligosaccharide (g), total sugars (g), glucose (g), galactose (g), fructose (g), sucrose (g), maltose (g), lactose (g), alcohol (g), cholesterol (mg), etc.. Finally, the vitamins table contains information per 100 g on the content of vitamins including retinol (ug), carotene (ug), retinol equivalent (ug), vitamin D, vitamin E, vitamin K1, thiamine, riboflavin, niacin, and vitamin C. 3.3. Nutrition Problem Identification Ontology The objective of this ontology is to define, classify and semantically represent potential nutrition related problems and the associated symptoms. Based on the values obtained from nutrition monitoring and nutrition assessment, this ontology provides the nutritionist with the ability to detect and evaluate nutrition related problems early in older adults, thus enabling proactive definition of nutrition intervention schemes that will allow the prevention of a problem before its actual instauration. We have defined this ontology by adapting the Nutrition Diagnosis step in the Academy of Nutrition and Dietetic Nutrition Care Process and Model, which specifies a problem as an alteration in the person’s nutritional status. Whilst the model proposes 60 approved standardized language terms, we have selected smaller number of problems and defined their associated symptoms for older adults using SWRL rules, which Figure 2: The Nutrition Assessment sub-Ontology are then evaluated by reasoning on the defined ontologies. The rules to proactively detect the early instauration of specific nutrition related problems in older adults as provided by nutritionists are defined in Table 1, while the Nutrition Problem Identification Ontology is depicted in Fig. 3. For example, to detect the nutrition problem related to inadequate fluid intake, 3rd SWRL rule from Table 1 is used. We have chosen this rule due to the fact that dehydration has a high occurrence rate in older adults and the assessment of this problem is quite complex due to the fact that daily water needs are strongly dependent on various factors like ambient temperature, fluid losses and dietary composition (see Hydration in the Aging). In our approach each older adult consumes different amounts of food and drinks during the day, which are monitored, with the information being stored in the nutrition monitoring and assessment ontologies. Using the McCance and Widdowson's tables, the nutrient values associated with the older adult’s intake are computed (i.e. amount of protein, carbohydrates, liquid, and so on are consumed). In the case of inadequate fluid intake, the reasoning works as follows: given the monitored intake of an older adult during a whole day, if the amount of water consumed is less than or equal to 1.5 litters (equivalent of six glasses), the older adult’s fluid intake for that day is deemed suboptimal. Besides the nutrition problems included in Table 1, we have defined rules for assessing short term and long term unhealthy behaviours which may lead to the development and instauration of malnutrition. As previously noted, the impact of malnutrition on the elderly population includes the exacerbation of chronic and acute diseases, acceleration in the development of degenerative diseases, delayed recovery from illness and, in very extreme cases, even death so identifying behaviours which allow the early identification and prevention of malnutrition is paramount (see Time to recognise malnutrition in EU). Addressing malnutrition is not exclusively addressing insufficient nutrition intake and starvation but also addressing obesity due to the fact that older adults may eat an abundance of calories that do not deliver all the nutrients their body needs. Table 1: SWRL rules for detecting the early instauration of nutrition related problems Nutrition Problem Rule definition SWRL Rule Implementation Dietary Plan for Intervention Suboptimal Energy Intake Weight loss > 3kg in the past month AnthropometryInformation (?a) ˄ hasWeightChange(?a, ?w) ˄ lessThanOrEqual(?w, -3.0) → hasNutritionProblem(?a, SuboptimalEnergyIntake) High Calories Diet, Provide Additional Snacks Excessive Energy Intake Weight gain > 3kg in the past month AnthropometryInformation (?a) ˄ hasWeightChange(?a, ?w) ˄ greaterThanOrEqual(?w, 3.0) → hasNutritionProblem(?a, ExcessiveEnergyIntake) Low Calories Diet Suboptimal fluid intake Less than 6 glasses of fluid (~1.5 L) per day IntakeNutrientValues (?i) ˄ hasWaterIntakeValue (?i, ?w) ˄ lessThanOrEqual(?w, 1.5) → hasFluidIntakeProblem(?i, SuboptimalFluidIntake) Increase Fluid Intake Plan Excessive Salt Intake Over 6 g per day for elders with heart problems IntakeNutrientValues (?i) ˄ hasSaltIntakeValue(?i, ?v) ˄ greaterThan(?v, 6) → hasNutritionProblem(?a, ExcesiveSaltIntake) Low Salt Diet Underweight BMI < 18.5 kg/m2 AnthropometryInformation (?a) ˄ hasBMI(?a, ?bmi) ˄ lessThan(?bmi, 18.5) → hasWeightProblem(?a, Underweight) High Calories Diet, Provide Additional Snacks Overweight BMI between 25.0 and 29.9 kg/m2 AnthropometryInformation (?a) ˄ hasBMI(?a, ?bmi) ˄ greaterThanOrEqual(?bmi, 25.0) ˄ lessThan (?bmi, 30.0) → hasWeightProblem(?a, Overweight) Low Calories Diet Obesity BMI > 30.0 kg/m2 AnthropometryInformation (?a) ˄ hasBMI(?a, ?bmi) ˄ greaterThanOrEqual(?bmi, 25.0) → hasWeightProblem(?a, Obesity) Figure 3: Nutrition Problem Identification Ontology To assess the older adult’s short term behaviours which may lead to malnutrition, the older adult’s self-feeding process is evaluated against the benchmarks and dietary recommendations for older adults (regarding daily nutrients intake) as defined by nutritionists. As before, these recommendations are represented using SWRL (see Table 2) and evaluated by means of a reasoning engine on the defined ontologies. If the conjunctions specified in the SWRL rule’s antecedent are false for a specific older adult, the rule’s consequent classifies the corresponding unhealthy behaviour. If unhealthy behaviours are observed over several days, nutrition interventions are triggered in the form of notifying the older adult or the associated carer and also recommending and delivering food to the older adult to rebalance the nutrients intake (see Nutrition Intervention Ontology). In addition to short term behaviours to assess the older adult’s long term unhealthy behaviours which may lead to malnutrition, the older adult’s level of adherence to the Table 2: SWRL rules for detecting older adult’s short term nutrition related un-healthy behaviors Evaluation Criteria Nutrition Assessment Rule SWRL Rule Implementation Energy Men: ~ 9.8 MJ/day Women: ~ 8.0 MJ/day AnthropometryInformation (?z) ˄ IntakeNutrientValues (?x) ˄ PersonalData (?p) ˄ PhysicalActivity (?ph) ˄ hasGender (?p, "male"^^string) ˄ hasHeight (?z, ?h) ˄ hasPAF (?ph, ?f) ˄ hasWeight (?z, ?w) ˄ add (?s1, ?r1, ?r2) ˄ multiply (?r1, ?h, 2.57) ˄ multiply(?r2, ?w, 0.0478) ˄ subtract(?s2, ?s1, 1.07) ˄ multiply(?rez, ?s2, ?f) → hasEstimatedEnergyValue(?x, ?rez) Total fat ~ 20-35% energy IntakeNutrientValues (?x) ˄ hasEstimatedEnergyValue(?x, ?e) ˄ divide(?r, ?m, 100.0) ˄ divide(?rez, ?r, 9.0) ˄ multiply(?m, ?e, 35.0) → hasEstimatedFatValueUpperLimit(?x, ?rez) Saturated fatty acids < 11% energy IntakeNutrientValues (?x) ˄ hasEstimatedEnergyValue(?x, ?e) ˄ divide(?r, ?m, 100.0) ˄ divide(?rez, ?r, 9.0) ˄ multiply(?m, ?e, 11.0) → hasEstimatedSatFatAcidValueUpperLimit(?x, ?rez) Trans fatty acids < 1% energy IntakeNutrientValues (?x) ˄ hasEstimatedEnergyValue(?x, ?e) ˄ divide(?r, ?m, 100.0) ˄ divide(?rez, ?r, 9.0) ˄ multiply(?m, ?e, 1.0) → hasEstimatedTransFatAcidValueUpperLimit(?x, ?rez) Protein > 0.75g/kg body weight / day AnthropometryInformation (?z) ˄ IntakeNutrientValues (?x) ˄ hasWeight(?z, ?w) ˄ multiply(?r, ?w, 0.75) → hasEstimatedProteinValueLowerLimit(?x, ?r) Potassium > 3.5g/d FoodIntake(?x) → hasEstimatedPotassiumValueLowerLimit (?x, 3.5) Calcium > 700 mg/ day FoodIntake(?x) → hasEstimatedCalciumValueLowerLimit (?x, 700.0) Vitamin D ~10 micrograms/day FoodIntake(?x) → hasEstimatedVitaminDValue (?x, 10.0) Salt < 6 g / day FoodIntake(?x) → hasEstimatedSaltValueUpperLimit (?x, 6.0) Alcohol Men: < 28 units / week Women: < 21 units/ week IntakeNutrientValues (?x) ˄ PersonalData(?p) ˄ hasGender(?p, "male"^^string) → hasEstimatedAlcoholValueUpperLimit(?x, 21.0) Table 3: 14-item Mediterranean diet adherence score questionnaire Questions Criteria for 1 point Do you use olive oil as main culinary fat? Yes How much olive oil do you consume in a given day (i.e. oil used for frying, salads, out-of-house meals, etc.)? ≥ 4 tablespoons How many vegetable servings do you consume per day (1 serving: 200g; consider side dishes as half a serving) ≥ 2 How many fruit units (portions) (including natural fruit juices) do you consume per day? ≥ 3 How many servings of red meat, hamburger, or meat products (ham, sausage, etc.) do you consume per day (1 serving: 100 – 150g). < 1 How many servings of butter, margarine or cream do you consumer per day (1 serving: 12 g) < 1 How many sweet or carbonated beverages do you drink per day? < 1 How much wine do you drink per week? ≥ 7 glasses How many servings of legumes do you consume per week? (1 serving: 150g) ≥ 3 How many servings of fish or shellfish do you consume per week? (1 serving 100-150g fish or 4-5 units (pieces) or 200g of shellfish) ≥ 3 How many times per week do you consume commercial sweets or pastries (not homemade), such as cakes, cookies, biscuits, or custard? < 3 How many servings of nuts (including peanuts) do you consume per week? (1 serving 30g) ≥ 3 Do you preferentially consume chicken, turkey, or rabbit meat instead of veal, pork, hamburger or sausage? Yes How many times per week do you consume vegetables, pasta, rice, or other dishes seasoned with sofrito (sauce made with tomato and onion, leek, or garlic and simmered with olive oil)? ≥ 2 Mediterranean diet assessed using a previously published (Martinez-Gonzalez et al., 2012) Mediterranean Diet Adherence Score Questionnaire (see Table 3). The effectiveness of the Mediterranean diet in reducing the prevalence of cardiovascular and chronic diseases has been largely evidenced in the state of the art literature (Féart et al., 2010). Also it has been shown that there is a 13 percent lower risk of cognitive impairment in Mediterranean diet adherents, and that following a Mediterranean diet may help prevent diabetes in people who are at risk of heart disease (Preidt, 2014). The Mediterranean diet is based on few principles such as (see Table 3): eating primarily plant-based foods such as fruits and vegetables, whole grains, legumes and nuts; replacing butter with healthy fats, such as olive oil; using less salt to flavor foods; limiting red meat to no more than a few times a month and eating other kind of low fat meat such as chicken or turkey; eating fish and poultry at least twice a week; and finally an optional principle is drinking red wine in moderation due to the anti-oxidants it contains. To assess the adherence of a person to the Mediterranean diet, the nutritionist uses the Mediterranean Diet Adherence Score questionnaire presented in Table 3. It is well known that the use of questionnaires is not always effective for older adults due to the effect of age on memory recall. The problem can be addressed by defining SPARQL rules (see SPARQL recommendation) for each corresponding question and using them on the defined nutrition monitoring and assessment ontologies to calculate the older adult adherence to the Mediterranean diet score. If the calculated score is less than 3 points there is poor adherence, if the calculated score is between 4 and 7 points there is medium adherence, and if the score is higher than 8 points there is good adherence. For example, Fig. 4 presents the SPARQL query rule equivalent to question 3 of Table 3, i.e. “How many vegetable servings do you consume per day (1 serving: 200g; consider side dishes as half a serving)”. The query starts with the declaration of prefixes to the ontologies defined above: PIPS food, nutrition assessment and nutrition monitoring. The query retrieves the food intake and the quantity of vegetables consumed during a day for an older adult by searching the food taxonomy defined in the PIPS ontology. For each type of food intake that is classified as of type Vegetables in this ontology, the foodQuantitiesInFoodIntake data Property is considered. The sum of these data values is the overall quantity of vegetables consumed during the day. 3.4. Nutrition Intervention Ontology The Nutrition Intervention Ontology (see Fig. 5) models knowledge regarding the type of actions that may be taken in case a nutrition problem or unhealthy behaviour is identified for an older adult. The nutrition intervention process consists of two steps consisting of different actions modelled in our ontology: "PREFIX foods: <http://www.semanticweb.org/foodsontology/food/#>" + "PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>" + "PREFIX food: <http://www.semanticweb.org/dorin/ontologies/food#>" + "PREFIX nutritionassessment: <http://www.semanticweb.org/ ontologies/nutritionassessment#>" + "PREFIX nutritionmonitoring: <http://www.semanticweb.org/ ontologies/nutritionmonitoring#>" + "SELECT ?foodIntake (SUM(xsd:double(?coefficient) * xsd:double(?grams)) AS ?Grams) " + " WHERE {" + " ?foodIntake nutritionmonitoring:hasUsername \"" + username + "\" . " + " ?foodIntake nutritionmonitoring:hasDate \"" + dateString + "\"^^xsd:dateTime . " + " ?foodQuantityInFoodIntake nutritionassessment:foodQuantityHasFoodIntake ?foodIntake . " + " ?foodQuantityInFoodIntake nutritionassessment:foodIntakeHasFoodQuantity ?foodQuantity . " + " ?foodQuantityInFoodIntake nutritionassessment:foodQuantityInFoodIntakeHasCoefficient ?coefficient . " + " ?foodQuantity nutritionassessment:hasGramsValue ?grams . " + " ?foodQuantity nutritionassessment:hasBasicFood ?basicFood . " + " ?basicFood rdf:type food:Vegetables . " + " } GROUP BY ?foodIntake "; Figure 4: SPARQL rule equivalent to question 3 of Table 3 “How many vegetable servings do you consume per day” Figure 5: Nutrition Intervention Ontology planning the intervention and implementing it. In the first step, the recommendation of a new diet is performed by a nutritionist aiming to change an identified unhealthy behaviour by setting goals and expected outcomes in form of a recommendation that needs to be followed by the older adult. The recommendation will be defined in close correlation with the nutrition problems identified and nutrition-related monitored information. Ideally the plan should be developed in conjunction with or by the older person (and carers or family) and should reflect clinical guidelines. For goals or intervention planning, the following factors are considered: changing nutrients or food provided to affect intake, influencing nutrition related knowledge or behaviour, change to environment and access to supportive care and services. Table 4: Generic nutrition intervention plan for elders Energy Protein BREAKFAST: E.g. fruit juice, cereal and semi-skimmed milk, bread and spread, preserves; 400 kcal 5.0g MAIN MEALS AND DESSERTS: Two main courses and desserts offered each day should average out to meet energy and protein values; 1050 kcal 29.0 g HIGHER ENERGY SNACKS INCLUDING A SUPPER SNACK: Two to three daily; total nutrients provided should average out to meet the energy and protein values. 400 kcal 3.0g MILK FOR DRINKS 400ml semi-skimmed/skimmed milk. Minimum of Fluid intake 6-8 glasses of water, juice and milky drinks. A small portion (15g or half a matchbox) (OR additional 150ml milk) 190 kcal 60 kcal 14g 4.0g For practical purposes the total to be provided is rounded to 2100 kcal and 55g protein. These are minimum values and not targets. Our semantic based approach enables the dynamic selection, Table 4 presents a generic nutrition intervention plan for an older adult incorporating minimum nutritional standards with no chronic condition (adapted from (Cartz, M. et al., 2012) and NACC Nutritional Standards for Adults). In the second step, the intervention plan is implemented by means of defined intervention alternatives presented in Table 5. Table 5: Nutrition intervention implementation options Intervention Automated goals Example of actions Food ordering Implement the nutrition prescription plan:  regular meals and/or regular snacks  modify distribution, type or amount of food within meals or snacks;  encourage / provide specific foods, beverages or food groups. Provide additional snack foods that are high in energy and protein (and in line with older person’s preferences) e.g. cheese with biscuits, yoghurt, custard, scones or pancakes. Feeding assistance Assistance with eating:  special equipment;  feeding position recommendation;  meal set up. Involve carers and family. Incorporate some of the strategies recommended by the Dementia Mealtime Assistance Tool or similar. Feeding environment Improvement to eating environment:  lighting, temperature distractions, table set up adjustment, etc. Nutrition education Brief Nutrition Education:  build or reinforce basic nutrition related knowledge Video clips with tips, recipes, skills education; highlighting foods on the menu that meet goals set. Regarding food ordering, different foods may be suggested to comply with the nutritionist’s recommendation and nutrition- SELECT ?foodIntake " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?energy) / 100) AS ?Energy) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?fat) / 100) AS ?Fat) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?carb) / 100) AS ?Carb) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?satFatAcid) / 100) AS ?SatFatAcid) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?transFatAcid) / 100) AS ?TransFatAcid) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?potassium) / 100) AS ?Potassium) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?calcium) / 100) AS ?Calcium) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?sodium) / 100) AS ?Sodium) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?vitaminD) / 100) AS ?VitaminD) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?alcohol) / 100) AS ?Alcohol) " + "(SUM(xsd:double(?coefficient) * xsd:double(?grams) * xsd:double(?water) / 100) AS ?Water) " + " WHERE {" + "?foodIntake rdf:type nutritionmonitoring:FoodIntake . ?foodIntake nutritionmonitoring:foodIntakeHasId " + foodIntakeId + ". ?foodQuantityInFoodIntake nutritionassessment:foodQuantityHasFoodIntake ?foodIntake" + ". ?foodQuantityInFoodIntake nutritionassessment:foodIntakeHasFoodQuantity ?foodQuantity " + ". ?foodQuantityInFoodIntake nutritionassessment:foodQuantityInFoodIntakeHasCoefficient ?coefficient" + ". ?foodQuantity nutritionassessment:hasGramsValue ?grams . ?foodQuantity nutritionassessment:hasBasicFood ?basicFood" + ". ?basicFood nutritionassessment:hasEnergyValue ?energy . ?basicFood nutritionassessment:hasFatValue ?fat" + ". ?basicFood nutritionassessment:hasCarbValue ?carb . ?basicFood nutritionassessment:hasSatFatAcidValue ?satFatAcid" + ". ?basicFood nutritionassessment:hasTransFatAcidValue ?transFatAcid . ?basicFood nutritionassessment:hasProteinValue ?protein" + ". ?basicFood nutritionassessment:hasPotassiumValue ?potassium . ?basicFood nutritionassessment:hasCalciumValue ?calcium" + ". ?basicFood nutritionassessment:hasSodiumValue ?sodium . ?basicFood nutritionassessment:hasVitaminDValue ?vitaminD" + ". ?basicFood nutritionassessment:hasAlcoholValue ?alcohol . ?basicFood nutritionassessment:hasWaterValue ?water" + " } " + " GROUP BY ?foodIntake "; Figure 6: Evaluating the nutritional information of older adult’s daily food intake related monitored information. If it is inferred that using food ordering can no longer balance the daily intake (i.e. there is severe violation of the dietary plan), alarms are generated for the older adult, associated carer and nutritionist. Our semantic based approach enables the dynamic selection based on the prescribed diet, of suitable food service providers, potentially enabling automated shopping. In addition, it allows the assistance of older adults and their informal carers during daily self-feeding activities to enable the detection and prevention of malnutrition. The nutrition education actions aim to educate the older adult to eat healthier food by building or reinforcing basic nutrition or diet related knowledge in strict alignment with the assessed nutrition related problems. In this step the nutritionist will also use the system for process evaluation purposes, i.e. to check that the plan has been carried out (for example, that a meal with additional snack items has been delivered as agreed), and that any existing problems are identified and rectified. This is done by means of continuously evaluating the food intake by calculating the nutritional intake on daily basis using SPARQL rules such the one presented in Fig. 6 and by food ordering to balance the intake. In this phase, it is important that the elder can provide feedback on the acceptability of the intervention, as well as communicate any practical issues with its implementation. 4. Use case validation The validation process aims to ensure that the defined nutrition expert system satisfies the requirements of its users, which are in our case the nutritionists. In the state of the art literature, there is a general agreement on two main classes of expert system validation (Preece, 1990): laboratory validation and field validation. Field validation for evaluating our approach in depth will require large scale deployment and data gathering over a long period which may be costly in terms of time and money and may primarily discover and address problems with front-end interface aspects of the system rather than the underlying technical contributions as described in the above paper. Moreover, prior to that, a preliminary evaluation is required to ensure that the idea is feasible, as determined by expert nutritionists. This is why our evaluation process is based on laboratory validation, which measures the usefulness and quality of the system, aiming to expose problems regarding the defined nutrition knowledge base and inference system. The validation process will consist of comparing the nutrition expert’s assessment with that of the expert system for a test case (an older adult), with the system being accepted if it is as competent as the nutritionist (Grogono, et al., 1993). This process requires the definition and generation of older adult patients’ situations describing unhealthy self-feeding behaviours, the execution of the expert system to assess the nutrition related problems, and the comparison of the results provided with the ones given by the nutritionist. In addition, the cohesion and computational efficiency of the knowledge base are assessed. 4.1. Nutrition Expert System in-lab Prototype Fig. 7 presents the architecture of the nutrition expert system developed and used for the in-lab validation process. The core part of the system is the nutrition related knowledge represented by means of the Nutrition Care Process Ontology and associated nutrition diagnosis SWRL and SPARQL rules. The Nutrition Monitoring module generates data describing older adult unhealthy self-feeding behaviours in accordance with the foreseen validation scenarios. Examples of the data generated are the older adult’s food and beverage intake, weight, physical activity levels, etc. The Nutrition Assessment and Problem Identification modules implement the nutrition expert system’s inference engine, which evaluates the defined nutrition related rules on elder’s food intake. The inference engine is based on semantic reasoning tools and techniques. The Nutrition Intervention module selects the appropriate intervention option according to the assessed nutrition problem (i.e. it associates dietary goals with specific conditions and behaviours identified). We have designed a semantic inference engine consisting of the following tools (see Fig. 8): (i) Protégé (Fudholi et al., 2009) for implementing the concepts and properties of the Nutrition Care Process Ontology, (ii) Jena framework which provides the API to process RDF data, (iii) D2RQ for mapping individuals from the ontology with entries from an SQL database and (iv) Pellet reasoned for evaluating the defined nutrition rules on the ontology. Due to the large number of ontology concepts describing food items that need to be instantiated and classified in the defined taxonomy, this process cannot be done manually. On the other hand, there is an evident need for ensuring a semantic validation Figure 7: Architecture of the nutrition expert system’s in-lab prototype and automated reasoning on the data provided which cannot be achieved using the relational database representation of data. Our solution is to store data regarding food items and their nutritional information in a relational database and to map the ontology concepts onto the database tables to obtain concept instances. Different alternatives to connect and map ontology to a relational database are reported in the literature (Konstantinou et al., 2008). When choosing the solution we have considered the following criteria: how the individuals will be persisted in the relational database, whether the ontology will be written in a specialized tool, whether a reasoner will be used to infer new knowledge from the already existing data, and whether data will be queried by using a specialized language. We chose to use D2RQ due to the fact that classes, object properties and data properties can be managed in the form of database tables, there is a clear separation between the ontology data (individuals) and ontology structure, and D2RQ can be easily integrated with Jena which is a powerful tool for reasoning on ontologies. A simplified architecture of how the ontology is mapped to the database using D2RQ and how this mapping can be used in order to get information is presented in Fig. 8 Figure 8: Food Ontology mapping onto the Food Composition DB and Inference Engine Data about individuals is persisted in a database. A mapping file is used to define the relation between the ontology specific elements (i.e. ontology classes, data properties, object properties, and subclass relations) and the database tables. Reasoning and query rules are expressed using the concepts provided by the food ontology and the mapping defined data from the database. The steps followed by the in-lab prototype to make decisions are: firstly the nutrition monitoring module generates data describing the older adult’s nutrition related situation by creating the corresponding individual in the Nutrition Care Process ontology; secondly the data regarding the nutrient composition of the older adult’s consumed food items is queried by the D2RQ platform and passed to Jena; thirdly the defined nutrition related rules are loaded by Jena and evaluated by Pellet on the ontology; and finally the inference result is queried to extract and process the nutrition assessments and decisions. 4.2. Ontology usefulness evaluation The usefulness of the Nutrition Care Process ontology and inference engine is evaluated by using the developed in-lab system prototype for nutrition knowledge management to assess the self-feeding behaviour of one individual (called Miguel). Two nutrition experts who provided the initial nutrition related knowledge had interacted with the in-lab prototype and conducted independent assessment processes on the patient to evaluate the outcome. Miguel is a 78 year old retired man who lives alone in his home in a small town of the North of Spain. He was diagnosed with type II diabetes 7 years ago and was later also diagnosed with cardiovascular disease after he experienced a Myocardial Infarction two years ago. He wakes up every day at 8.00AM and prepares himself a simple breakfast, coffee with some toast. After that, Miguel takes a walk towards a newspaper kiosk located in the town’s main square, which is 15 minutes’ walk from his home. Miguel buys the newspaper and then sits in one of the main square benches. He reads the news for around one hour and gets distracted with the bustle of the place for some more time. On his way back home, Miguel stops at a supermarket, close to his place, and buys some food including precooked meals. Miguel is able to prepare simple meals, but not complex ones, so he prefers to consume precooked food which makes his life easier. Also every weekend Miguel’s friends come to visit, often bringing some sweets or treats. Being diagnosed with type II diabetes and cardiovascular disease, the diet that Miguel is actually following is not suitable for his condition as he is consuming too many calories, too much salt and saturated fat, and does not have balance among the main food groups. He gained 4 kg in weight over the last 12 months, so his local doctor has referred him to a nutritionist. Table 6 presents Miguel’s profile after the screening conducted by the nutritionist. Table 6: Miguel’s nutrition related screening Monitored Data Measured Values Anthropometry Height: 178 cm (measured using stadiometer) Current weight: 88 kg BMI: 27.8 kg/m2 Weight history: 82 kg (2 year ago), 84 kg (1 years ago) Physical activity factor (PAF): 1.2 Biochemistry HbA1C: 7.5% Random blood glucose: 9.8 mmol/L Blood pressure: 140/90 mm Hg Clinical Medical history: Myocardial Infarction, Arthritis Medications: Metformin 1 g bd, Simvastatin 20 mg od Dietary Texture: Normal Requires diabetic meals Average Meal: 550 kcal; 18g protein Preferences: Spanish Cuisine Current Mediterranean diet assessment score: 5 The data resulting from Miguel’s screening and nutrition monitoring as well as his profile are inserted as individuals in the Nutrition Care Process ontology. Using this information, the inference engine of the prototype is able to calculate the following parameters regarding Miguel’s condition: – The estimated basal metabolic rate (BMR). The value of BMR is computed using the Henry equation (Henry, 2005) where 0.0478, 2.57 and 1.07 are constants for male elders aged over 60 years: 𝐵𝑀𝑅 = (0.0478 ×𝑤𝑒𝑖𝑔ℎ𝑡) + (2.57×ℎ𝑒𝑖𝑔ℎ𝑡) − 1.07 = 7.711 𝑀𝐽/𝑑𝑎𝑦 (1842 𝑘𝑐𝑎𝑙/𝑑𝑎𝑦) (1) – The estimated daily energy requirement. The value of the estimated daily energy requirement is the estimated BMR multiplied by a physical activity factor (PAF): 𝐵𝑀𝑅 ×𝑃𝐴𝐹 = 9.25 𝑀𝐽/𝑑𝑎𝑦 (2211 𝑘𝑐𝑎𝑙/𝑑𝑎𝑦) (2) – The body mass index: 𝐵𝑀𝐼 = 𝑤𝑒𝑖𝑔ℎ𝑡(𝑘𝑔)/(ℎ𝑒𝑖𝑔ℎ𝑡(𝑚)×ℎ𝑒𝑖𝑔ℎ𝑡(𝑚) ) = 27.8 𝑘𝑔/𝑚2 (3) Next we determine our prototype’s ability to evaluate the nutrition problem identification rules defined in SWRL in Section 3.3 (see Table 2) considering Miguel’s screening, profile and the above assessment, to infer the Recommended Daily Intake values for different nutrients personalized for Miguel’s condition. Table 7 presents the Recommended Daily Intake values for Miguel’s daily menu calculated by our prototype and benchmark values provided by nutritionists. Table 7: Miguel inferred recommended daily nutrient intake Nutrient Miguel’s Inferred Daily Values Nutritionist Benchmark for Men Energy 9.25 MJ/day ~ 9.8 MJ/day Total fat ~ 49 - 86 g/day (~19.6- 34.4% energy) ~ 20-35% energy Saturated fatty acids ~ 27 g/day (~10.8% energy) < 11% energy Trans fatty acids ~ 2.5 g/day (~1% energy) < 1% energy Protein ~ 66 g/day (~11.4% energy) > 0.75g/ kg body weight / day (>11% energy) Potassium ~ 3.5 g/day > 3.5 g/day Calcium ~ 700 mg/day > 700 mg/ day Vitamin D ~ 10 micrograms /day ~10 micrograms/day Salt ~ 2.4 g/day < 6 g / day Alcohol ~ 32 g/day < 28 units / week Water ~ 2.5 L/day ~2.5 L/day One can notice that all calculated values fall between the limits traced by nutritionists. At the same time the prototype inference engine makes a nutrition diagnosis assessment to identify the symptoms of nutrition related problems defined in Table 1 by evaluating the associated SWRL rules on Miguel’s specific situation. For example due to the fact that Miguel’s BMI is between 25 and 30 (i.e. 27.8) our prototype classifies him as overweight. For nutrition intervention the system first tries to associate the proper intervention plans from the ones already defined by nutritionists and formalized in our ontology for Miguel’s condition and profile. Being previously assessed as overweight our prototype successfully selects a low calorie based diet (see Table 1) as appropriate intervention plan suggesting a daily menu with ~836 kJ lower than his estimated daily energy requirements (~ 9.25 MJ). Also because he suffers from diabetes and cardiovascular diseases, our prototype infers that in addition to the low calories diet, Miguel should also follow specific dietary plans defined for those diseases. For cardiovascular diseases the system recommends a personalized low salt diet (see Table 1) with a moderate restriction estimated based on Miguel’s daily energy requirements. It is likely that potassium intakes, at a higher intake of fruit and vegetables than recommended to the general population, will have an effect on blood pressure. The association of blood pressure with alcohol is particularly strong at intakes above 2 drinks per day so it is carefully evaluated. For diabetes, the system’s dietary plan is based on a reduction in calorie intake in those who are overweight, coupled with an improvement in physical activity to improve the action of insulin. However for diabetes patients requiring insulin as Miguel’s case, the system recommends eating carbohydrate containing foods at regular intervals and/or matched to their insulin dose. More slowly absorbed carbohydrate foods (i.e. those with a lower glycaemic index) should be emphasized for improved glycaemic control. Once the intervention plan is developed for Miguel’s individual conditions, the system prototype will continuously evaluate the nutritional adequacy of Miguel’s daily food intake using the SWRL rules. Table 8 presents different simulated situations of Miguel’s daily intake, for which the prototype was used to evaluate the intake adequacy and the degree of violation of Miguel’s dietary goals. The same situations were presented to the nutrition experts who conduct their own independent assessment, and in the end the results are compared. As it can be seen our prototype manages to successfully determine the daily menus that are inadequate for Miguel assessing the adherence to the goals defined by nutritionists. Table 8: Miguel nutrition assessment for daily food intake Miguel’s Daily Intake Nutrient Analysis Our prototype results Nutritionist results Day 1: Breakfast – Mug of coffee with 60ml whole milk and 1 teaspoon sugar; 2 slices white bread spread thickly with butter Lunch – 250ml orange juice; 1 x 400g packs of macaroni cheese; 2 tablespoons canned sweet corn, 2 slices tomato, 200g rice pudding Dinner – 1 x 400g pack chicken curry with rice. Half-pint glass of mango juice. Snack – Banana (1 medium) Energy: 2169.8 kcal Exceeds recommendation: Salt: 7.4g % en from fat: 37.29% % en from SFA: 18.9% Below recommendation: Fluid: 1645.32 ml Potassium: 2.9g Energy: 2144 kcal Exceeds recommendation: Salt: 7.5g % en from fat: 37% % en from SFA: 19% Below recommendation: Fluid: 1620 ml Potassium: 2.8g Day 2: Breakfast – Mug of coffee with 60ml whole milk and 1 teaspoon sugar; 2 slices white bread spread thickly with butter Lunch – 1 x 440g pack of Beef curry with rice. 200g rice pudding, 330ml can of lemonade Dinner -Turkey pie– 1 medium slice; 2 x 65g Ice lollies with real fruit juice Snack: 1 x 40g bags peanuts, raisins and chocolate chips. 1 pint lager, 1 x 37g bag potato crisps Energy: 2145.0 kcal Exceeds recommendation: Salt: 8.3 g % en from fat: 39.12% % en from SFA: 18.47% Below recommendation: Fluid: 1218.67 ml Potassium: 3.4 g Energy: 2098 kcal Exceeds recommendation: Salt: 6.3 g % en from SFA: 16% Below recommendation: Potassium: 2.8 g Day 3: Breakfast – Mug of coffee with 60ml whole milk and 1 teaspoon sugar; 2 slices white bread spread thickly with butter Lunch – 1 x 400g pack chicken curry with rice, 2 dessertspoons plain yoghurt, granary bread 1 slice spread thickly with butter, half-pint glass mango juice. Dinner – 1 x 300g packet of sausages and mash, 1 medium slice apple pie (1 crust) with ready-made custard (1 pot). Snack – 1 medium nectarine, 1 medium banana Energy: 2255.1 kcal Exceeds recommendation: Salt: 8.3 g % en from fat: 39.37% % en from SFA: 18.07 % Below recommendation: Fluid: 1495.31 ml Potassium: 2.9 g Energy: 2239 kcal Exceeds recommendation: Salt: 8.4 g % en from fat: 39% % en from SFA: 17% Below recommendation: Potassium: 3.0 g Miguel’s Daily Intake Nutrient Analysis Our prototype results Nutritionist results Day 4: Breakfast – Mug of coffee with 60ml whole milk and 1 teaspoon sugar; 2 slices white bread spread thickly with butter Lunch: 400g pack vegetable lasagne, garlic bread – half of 150g stick, fruit salad 1 cup with 2 scoops vanilla ice cream Dinner: 1 x 450g pack cottage pie, 1 half-pint glass mango juice Snack: 2 glasses tomato juice, 1 portion camembert cheese with 5 water biscuits Energy: 2401.8 kcal Exceeding recommendation: Salt: 15 g % en from fat: 41.57% % en from SFA: 17.41% Below recommendation: Fluid: 1694.40 ml Energy: 2379 kcal Exceeding recommendation: Salt: 14.3 g % en from fat: 41% % en from SFA: 17% Below recommendation: - Legend: % en from fat = proportion of energy derived from total fat; % en from SFA = proportion of energy derived from saturated fatty acids; kcal = kilocalories 4.3. Ontology quality evaluation To assess the quality of the ontology, the computational efficiency, coupling and cohesion metrics were used (Burton- Jones et al., 2005; Hlomani and Stacey, 2014). We have simulated an increasing number of older adults (up to 210) by generating and inserting the corresponding individuals in the ontology, and evaluated the nutrition knowledge and knowledge management using metrics reported in literature. The ontology computational efficiency refers to the time overhead for running SPARQL queries and SWRL reasoning rules and its variation with the number of older adults stored in the ontology. First we have evaluated the time needed to run and assess the Mediterranean Diet adherence questions, with the average time values obtained being reported in Fig. 9. Even though the relation between time overhead and number of elders is exponential, reasonable time results are obtained for 210 elders (~ 27 seconds). Second we have determined the Pellet’s incremental reasoning performance on the Nutrition Care Process ontology by evaluating the variation of the incremental classification time with the number of older adults stored in the ontology. The incremental classification is used to update the ontology classification results when the class hierarchy changes (new concepts are defined). Figure 9: Computational time overhead results The coupling metric evaluates the number of classes from external ontologies imported and used in the Nutrition Care Process ontology. We have reused only the PIPS food ontology for classifying the food items so the coupling value is given by the number of concepts used from this ontology (i.e. 176 out of 277 named classes). The cohesion metric measures the modularity and the degree of relatedness of concepts from our ontology in terms of number of root concepts, number of leaf concepts, and depth of inheritance tree, etc. For computing cohesion we have used an ontology evaluation tab for Protégé (see Ontology evaluation in Protégé) and the obtained results are summarized in Table 9. Regarding ontology consistency, besides the evaluation done by nutrition experts regarding ontology concepts definition, taxonomy and relation among them, we have used Pellet (Incremental) reasoner to check the ontology’s hierarchies, domains, ranges, conflicting disjoint assertions, etc. No Inconsistent Ontology Exception was generated. 4.4. Discussion and Limitations In this section we discuss our findings and potential limitations form the perspective of our system’s end users: older adults, Table 9: Nutrition care process ontology metrics results Evaluation Metric Monitoring Ontology Assessment Ontology Problem Identification Ontology Intervention Ontology PIPS Food Ontology No. Named Classes 43 34 7 7 176 Average No. Parents 0.8 0.82 0.85 0.94 0.92 Average No. Siblings 3.4 4.7 6 3.2 12.5 Max Depth 2 3 1 5 1 Total No. Nodes 44 35 8 18 177 Total No. Roots 9 6 1 1 13 Total No. Internal Nodes 10 6 1 5 13 Total No. Children 34 28 6 16 163 Total No. External Nodes 33 28 6 12 163 No. Properties 53 48 4 6 0 No. Object Properties 8 10 0 1 0 No. Data Type Properties 45 38 2 5 0 Properties with Domain 100% 93.75% 100% 100% 0% Properties with Range 39.60% 62.50% 100% 100% 0% nutritionists and food providers. Effective older adults’ nutrition monitoring is crucial for our system to capture nutrition problems, tracking trends and later for intervention decision-making. The nutritional assessment of food intake done by nutritionists during elderly regular visits using questionnaires implies great memory effort from elderly and may lead to imprecise results due to elderly memory impairment. We had approach this problem allowing the elders to provide information regarding their food intake directly right after they had eaten the meals using semantic concepts from the food ontology familiar to them and the interfaces provided by our system. This requires some minor technical knowledge from elders making the accessibility and usability features our system of high priority. Before commercializing such product great focus should be put on making the user experience positive with people in general and for those with age related impairments in particular and understanding the type of barriers they have when interacting and understanding the system. At the same time the ontology based inference system proposed is open for integrating sensor based intake monitoring (when available) for providing a definitive answer to the question of “what is the actual food intake?”. We innovatively use ontologies, databases and interference systems for the analysis of non‐medical, nutrition data to identify patterns and relations between lifestyle behavior, self- feeding process and adherence to a prescribed diet. A challenge in this case is the way in which the large amount of data collected is stored considering system scalability, performance and usability. Even though the databases provide great scalability and are used widely, there are some limitations: they store only data that is explicitly known, the performance of the system is reduced significantly if the number of tables is very large, and the extraction of semantic meaning from data is slow. Our data representation approach which uses both an ontology and a relational database presents as advantages the fact that reasoning capabilities may be used, in order to infer new knowledge from information which is already represented in the ontology, while new information can be inserted easily using Object/Relational Mapping tools that are not specific to ontologies. The computational efficiency evaluations show reasonable processing time results even for complex inference processes such that the long term unhealthy behaviours and Mediterranean diet assessments. As from nutritionists point of view their main benefit for using the system is the automatic identification of older adults’ behavioural patterns that are usually associated with unhealthy eating and the deviations from prescribed diets especially in the case of pre-existing health problems such as diabetes and cardiovascular dieses. Our inference system is able to correctly assess the total nutrient values from daily food intake and more over it is successful in selecting the appropriate diet and food suggestions to compensate potential self-feeding problems. The nutrition assessment rules defined with the help of nutritionists and implemented in SWRL are generic enough to allow the nutritionists to define more specialized self-feeding problems and to associate and monitor the adherence to a wide variety of diets (see Table 10) which are usually prescribed to older adults. From food providers perspective our system provides the opportunity of new revenue streams from delivery of personalized meals packages, potentially enabling automated shopping. Following a healthy diet plan that meets an older adult’s nutritional needs, while accommodating their medical conditions and food preferences, is essential for addressing the malnutrition problem. Whilst those living on their own may wish (and are encouraged) to remain independent, their ability to cook and prepare food may be affected by various aging- related diseases, making it difficult for them to maintain a well- balanced diet. Utilizing external food providers offers an opportunity for such older adults to obtain the quality, nutritious meals they are no longer capable of cooking at home. Moreover, meal delivery services can also bring other benefits for older adults including cost-efficiency and social interaction. Table 10: Main Classes of diets and their variations Clases Sub-clases BASAL DIET Normal diet, Without salt added, Soft diet with/without salt added, Puree diet with/without salt added DIABETIC DIET Normal diabetic diet, Diabetic diet without salt added, Soft diabetic diet with/without salt added, Puree diabetic diet with/without salt added. WEIGHT CONTROL DIET Weight control diet with/without salt added, Soft weight control diet with/without salt added, Puree weight control diet with/without salt added DIET FOR ANTI- COUAGULANT THEREAPY Normal sintrom diet, Sintrom diabetic diet with/without salt added, Sintrom weight control diet with/without salt added, Soft sintrom diet with/without salt added, Puree sintrom diet with /without salt added SPECIAL DIET Vegetarian diet, Gastric protection diet, Astringent diet, Intolerance diets, Allergy diets, Uric Acid diet, Purine restricted diet, Dialysis patient diet, Renal patient’s diet, Low protein diet. Such reliance on external food providers, however, also imposes challenges for our system. Although a range of food services supporting older adults in achieving a healthy diet may be available, potential clients (older adults, their carers, or software applications acting on their behalf) might not be aware of these services, thus demanding a means to increase their visibility. Additionally, such food services need to be sufficiently described in order to enable an appropriate selection (e.g. to ensure the meals selected are both appealing and comply with specific dietary criteria). 5. Conclusion In this paper we have presented an expert system for older adults’ nutrition care process. We have defined a semantic dietary knowledge base capturing various aspects that are relevant to older adults and their nutrition and a suitable inference system based on reasoning and query techniques (against such knowledge base) to assess the elderly’s short term and long term self-feeding behaviours, to identify potential nutrition associated problems and to help nutritionists in defining a corresponding personalized intervention plan. The semantic dietary knowledge is encoded as the Nutrition Care Process Ontology, consisting of four sub-ontologies: Nutrition Monitoring, Nutrition Assessment, Nutrition Problem Identification, and Nutrition Intervention. The evaluation is carried out in-lab and assesses the usefulness and quality of the system aiming to expose problems regarding the defined nutrition knowledge base and inference system. We have compared the nutritionist assessment with the expert system for a test case older adult diagnosed with cardiovascular disease and diabetes, with our system proving to be as competent as the nutritionist. The computational efficiency tests show that even though the relation between the ontology processing time overhead and number of elders is exponential, reasonable time results are obtained for 210 elders. Also it is shown that the defined ontology features strong coupling, cohesion and consistency. Acknowledgements This work has been carried out in the context of Ambient Assisted Living project DIET4Elders (www.diet4elders.eu). Also it was partially supported by a grant of the Romanian National Authority for Scientific Research, CCCDI – UEFISCDI. References ANTOS, J., BAICKER, K., CHERNEW, M., et al. (2013) Person-Centered Health Care Reform: A Framework for Improving Care and Slowing Health Care Cost Growth. Engelberg Center for Health Care Reform Report, http://www.brookings.edu/~/media/Research/Files/Reports /2013/04/person-centered-health-care- reform/person_centered_health_care_reform.PDF?la=en AL-DHUHLI, B.A., et al. (2013) Developing a Nutrition and Diet Expert System Prototype. 21 International Business Information Management Association Conference, http://www.ibima.org/AT2013/papers/balq.html. ARENS-VOLLAND, A., RÖSCH, N., FEIDERT, F., HARPES, P., HERBST, R., MÖSGES, R. (2009) ICT- supported Allergy and Diet management, Proceedings of the Med-e-Tel Conference. ATIENZA, A.A., et al. (2008) Using hand-held computer technologies to improve dietary intake. American Journal of Preventive Medicine, Volume 34, Issue 6, Pages 514– 518, doi:10.1016/j.amepre.2008.01.034. British Dietetic Association Nutrition and Dietetic Care pro- cess, Accessed 2015, http://www.rgu.ac.uk/979ED290- 7EB7-11E3-8B2A0050568D00BF BROWN, R.C.H. (2013) Moral responsibility for (un)healthy behaviour. J Med Ethics, doi: 10.1136/medethics-2012- 100774. BURTON-JONES, A. et al. (2005) A semiotic metrics suite for as-sessing the quality of ontologies”. Data & Knowl. Eng. 55(1), http://dx.doi.org/10.1016/j.datak.2004.11.010. caloriecount, Accessed 2015, http://www.caloriecount.com/ CaloriesCount, Accessed 2015, http://www.caloriescount.com CARTZ, M. et al. (2012) Improving Outcomes through Food and Beverage Services. The Food Counts Group in consultation with The British Dietetic Association Report, https://www.bda.uk.com/publications/professional/Nutritio nHydrationDigest.pdf D2RQ framework, Accessed 2015, http://d2rq.org/ DOMINGUEZ, D., GRASSO F., MILLER, T., SERAFIN R. (2006) PIPS: An Integrated Environment for Health Care Delivery and Healthy Lifestyle Support. 4th Working notes on Agents Applied in Health Care, ECAI 2006, http://www.findanexpert.unimelb.edu.au/display/publicati on122874. EFAD Professional Practice Committee (PPC), The implementation of a Nutrition Care Process (NCP) and Standardized Language (SL) among dietitians in Europe, (2014). Accessed 2015, http://www.drf.nu/wp- content/uploads/2014/08/EFAD-Prof-Practice-Committee- 2014.pdf ELIA, M. (2001), The Malnutrition Advisory Group consensus guidelines for the detection and management of malnutrition in the community. Nutrition Bulletin, 26: 81– 83. doi: 10.1046/j.1467-3010.2001.00111.x ELIA, M. AND RUSSELL, C.A. (2008) Combating malnutrition; Recommendations for Action. A report from the Advisory Group on Malnutrition. Report from The Advisory Group On Malnutrition, Led By BAPEN, http://www.bapen.org.uk/pdfs/reports/advisory_group_rep ort.pdf. ESPÍN, V., HURTADO, M. V., AND NOGUERA, M. (2015) Nutrition for Elder Care: a nutritional semantic recommender system for the elderly. Expert Systems, doi: 10.1111/exsy.12143. FÉART, C., SAMIERI, C., BARBERGER-GATEAU, P. (2010) Mediterranean diet and cognitive function in older adults. Current Opininion Clin Nutr Metab Care, 13(1):14- 8, http://www.ncbi.nlm.nih.gov/pubmed/19834324. FUDHOLI, D.H., MANEERAT, N., VARAKULSIRIPUNTH, R., KATO, Y. (2009) Application of Protégé, SWRL and SQWRL in Fuzzy Ontology-Based Menu Recommendation. Int. Symposium on Intelligent Signal Proc. and Communication Systems, http://dx.doi.org/10.1109/ISPACS.2009.5383759. GROGONO, P.D., PREECE, A. D., et al. (1993) Review of Expert Systems Evaluation Techniques. AAAI Technical Report https://www.aaai.org/Papers/Workshops/1993/WS- 93-05/WS93-05-016.pdf. HENRY, C.J.K. (2005) Basal metabolic rate studies in humans: measurement and development of new equations. Public Health Nutrition, 8(7A), 1133–1152, http://www.ncbi.nlm.nih.gov/pubmed/16277825. HLOMANI, H. AND STACEY, D. (2014) Approaches, methods, metrics, measures, and subjectivity in ontology evaluation: A survey. IOS Press, http://www.semantic- web-journal.net/system/files/swj657.pdf. Hydration in the Aging, Accessed 2015, http://www.h4hinitiative.com/h4h-academy/hydration- lab/hydration-aging/concerns Jena framework, Accessed 2015, https://jena.apache.org/ KNOWLER, W.C., et al. (2002) Reduction in the incidence of type 2 diabetes with lifestyle intervention or metformin. New England Journal of Medicine, vol 346, no 6: pp 393- 403, doi: 10.1056/NEJMoa012512. KNOX, T.A. et al. (2013) Assessment of Nutritional Status, Body Composition, and Human Immunodeficiency Virus- Associated Morphologic Changes. Clinical Infections Diseases, Vol 36, http://www.ncbi.nlm.nih.gov/pubmed/12652373. KONSTANTINOU, N., SPANOS, D., MITROU, N. (2008) Ontology and database mapping: a survey of current implementations and future directions. Journal Web Eng., 7, 1, 1-24, http://dl.acm.org/citation.cfm?id=2011263. LEE, C.S., WANG, M.H., LI, H-C., CHEN, W-H. (2008) Intelligent Ontological Agent for Diabetic Food Recommendation. IEEE World Congress on Computational Intelligence, http://dx.doi.org/10.1109/FUZZY.2008.4630615. LJUNGQVIST, O. AND MAN, F. (2009) Under nutrition: a major health problem in Europe. Nutr Hosp. 24(3):369-70, http://www.ncbi.nlm.nih.gov/pubmed/19721916. MARTINEZ-GONZALEZ, M.A. et al. (2012) A 14-item Mediterranean diet assessment tool and obesity indexes among high-risk subjects: The PREDIMED Trial. PlosOne, 7 (8): e43134, DOI: 10.1371/journal.pone.0043134. McCance and Widdowson's food composition tables, Accessed 2015, http://tna.europarchive.org/20110116113217/http://www.f ood.gov.uk/science/dietarysurveys/dietsurveys/ NACC Nutritional Standards for Adults (2011) http://www.thenacc.co.uk/assets/downloads/105/NACC% 20Nutrition%20Standards%202010.pdf Nutrition and Dietetics’ Nutrition Care Process and Model, Accessed 2015, http://www.andeal.org/category.cfm?cid=2 Ontology evaluation in Protégé, Accessed 2014, http://protegewiki.stanford.edu/wiki/Ontology_Evaluation Pellet reasoner, Accessed 2015, http://clarkparsia.com/pellet/ PREECE, A. D. (1990) Towards a methodology for evaluating expert systems. Expert Systems, 7: 215–223. doi: 10.1111/j.1468-0394.1990.tb00234.x PREIDT, R. (2014) Take Heart: Mediterranean Diet Combats Type 2 Diabetes, http://www.webmd.com/diabetes/news/20140327/take- heart-mediterranean-diet-combats-diabetes-study-says QUESADA, C., AND JENKINS, M. (2013) A Prototype Mobile Expert System for Nutritional Diagnosis. Proc. of the Twenty-Sixth International Florida Artificial Intelligence Research Society Conference, http://www.aaai.org/ocs/index.php/FLAIRS/FLAIRS13/pa per/view/5871. QUINN, S., BOND, R., AND NUGENT, C. (2015) Ontological modelling and rule-based reasoning for the provision of personalized patient education. Expert Systems, doi: 10.1111/exsy.12134. RAIHA, T. (2013) Nutrition Health Project using ICT-based Learning Environment. Participatory Action Res. in Est. Finland, http://epublications.uef.fi/pub/urn_isbn_978-952- 61-1058-5/urn_isbn_978-952-61-1058-5.pdf. RIVERO, C. R., HERNANDEZ, I., RUIZ, D., CORCHUELO, R. (2013) Benchmarking Data Exchange among Semantic- Web Ontologies, IEEE Transactions on Knowledge & Data Engineering, vol.25/9, pp. 1997-2009, http://dx.doi.org/10.1109/TKDE.2012.175. RODRÍGUEZ, N.D. et al. (2014) A survey on ontologies for human behavior recognition. ACM Computing Survey Volume 46 Issue 4, 46, http://dx.doi.org/10.1145/2523819. SIEBER, C.C. (2010) Malnutrition and appropriate nutritional care. Nutrition Day Conference, http://www.european- nutrition.org/images/uploads/pdf_pdf_87.pdf. SNAE, C. AND BRÜCKNER, M. (2008) FOODS: A Food- Oriented Ontology-Driven System. Int. Conf. on Digital Ecosyst. and Tech., http://dx.doi.org/10.1109/DEST.2008.4635195. SPARQL recomandation (2013) http://www.w3.org/TR/rdf- sparql-query/ Time to recognise malnutrition in EU (2011). EU Food Information Council Report, Accessed 2015, http://www.eufic.org/article/en/artid/Time-to-recognise- malnutrition-Europe/ TUMNARK, P., OLIVEIRA, L., SANTIBUTR, N. (2013) Ontology-Based Personalized Dietary Recommendation for Weightlifting. International Workshop on Computer Science in Sports, DOI: 10.2991/iwcss-13.2013.13. VALENCIA-GARCÍA, R., FERNÁNDEZ-BREIS, J. T., RUIZ-MARTÍNEZ, J. M., GARCÍA-SÁNCHEZ, F. AND MARTÍNEZ-BÉJAR, R. (2008), A knowledge acquisition methodology to ontology construction for information retrieval from medical documents. Expert Systems, 25: 314–334. doi: 10.1111/j.1468-0394.2008.00464.x VAN DER MERWE, A., KRÜGER, H., STEYN, T. (2014) A diet expert system utilizing linear programming models in a rule-based inference engine. Lecture Notes in Management Science, Vol. 6: 74–81, http://dspace.nwu.ac.za/handle/10394/10898?show=full. VASSÁNYI I., KÓSA I., PINTÉR B. , GAÁL B. (2014) Personalized Dietary Counseling System Using Harmony Rules in Tele-Care. European Journal for Biomedical Informatics, 2014; Volume 10, Issue 2. http://www.ejbi.org/en/ejbi/artinfo/191-en-24.html WILLETT, W.C., KOPLAN, J.P., NUGENT, R., et al. (2006) Prevention of Chronic Disease by Means of Diet and Lifestyle Changes. Disease Control Priorities in Developing Countries - 2nd edition. Chapter 44. http://www.ncbi.nlm.nih.gov/books/NBK11795/. Tudor Cioara, PhD is Assistant Professor and member of the Distributed Systems Research Lab of the Technical University of Cluj-Napoca. His current research interest is focused on artificial intelligence, semantic web, ambient assisted living, context awareness, autonomic computing and green IT. He is reviewer at several international conferences and journals and is/was involved in eight national/international research projects. Also, he has published more than 40 peer-reviewed papers. Ionut Anghel received the PhD degree in computer sciences at the Technical University of Cluj-Napoca in 2012. Currently he is a lecturer of computer science and also a member of the Distributed Systems Research Lab with main research areas: artificial intelligence, semantic web, autonomic computing and ambient assisted living. Ioan Salomie received the PhD degree in computer sciences at the Technical University of Cluj-Napoca in 1994 where he is professor of computer science. His research background includes artificial intelligence, semantic web, knowledge engineering, context awareness and ambient assisted living. He is reviewer at several international conferences and journals and published more than 80 peer-reviewed papers. He was also invited professor at the University of Limerick and Loyola College in Maryland and manager of three EU projects. Lina Barakat received her PhD degree at King’s College London. She is a member of AIS research group at the Department of Informatics, with background in semantic web and knowledge engineering. She has published several papers in international conferences and proceedings. Simon Miles is the head of the Agents and Intelligent Systems research group in the Department of Informatics at King’s College London and also a Senior Lecturer. He received his PhD degree at University of Warwick in 2003. He has published a number of research papers. Dianne Reidlinger is the Public Health Nutrition lead for the Nutrition and Dietetics program at Bond University Australia. She was a research dietitian at King’s College London for eight years. Her main research interest is in public health nutrition, particularly population approaches to improving dietary intakes. She has published a number of papers related to clinical dietetics and dietetics education. Adel Taweel received his PhD in Soft Engineering at Keele University in 2002. He is currently a member of the Birzeit University Palestine and was a lecturer of computer science at King’s College London for several years where he was involved in a number of research projects. His research interests include health informatics, service-oriented architectures, distributed software engineering and model-driven development. *Biographies (Photograph) Ciprian Dobre, Associate Professor, PhD, Habil., has scientific and scholarly contributions in the field of large scale distributed systems concerning mobile applications and smart technologies to reduce urban congestion and air pollution (TRANSYS), context-aware applications (CAPIM), opportunistic networks and mobile data offloading (SPRINT, SENSE), monitoring (MonALISA), high-speed networking (VINCI, FDT), Grid application development (EGEE, SEE-GRID), and evaluation using modeling and simulation (MONARC2, VNSim). These contributions led to important results, demonstrating his qualifications and potential to go significantly beyond the state of the art. Ciprian Dobre was awarded a PhD scholarship from California Institute of Technology (Caltech, USA), and another one from Oracle. His results received one IBM Faculty Award, two CENIC Awards, and three Best Paper Awards (in 2013, 2012, and 2010). The results were published in 6 books, 10 chapters in edited books, 34 articles in major international peer-reviewed journal (19 as main author, cumulated impact factor of 13.24), over 100 articles in well-established international conferences and workshops. Florin Pop received his PhD in Computer Science at the University POLITEHNICA of Bucharest in 2008. He received his MSc in Computer Science in 2004 and the Engineering degree in Computer Science in 2003, at the same University. He is Associate Professor within the Computer Science Department and also an active member of Distributed System Laboratory. His research interests are in scheduling and resource management (decentralized techniques, re-scheduling), multi-criteria optimization methods, Grid middleware tools and applications development (satellite image processing an environmental data analysis), prediction methods, self- organizing systems, contextualized services in distributed systems. Expert System for Nutrition Care Process of Older Adults     Research Highlights  expert system for a nutrition care process tailored for the specific needs of elders   An inference engine is developed on top of the ontology, providing semantic  reasoning infrastructure and mechanisms for evaluating the rules defined for  assessing short and long term elders’ self‐feeding behaviors   Define and semantically represent information    Performance evaluation  
work_6mldueq24zfvdlyaui7qr3x7ym	----	Verifying regional climate model results with web-based expert-system Procedia Technology 1 ( 2012 ) 24 – 30 2212-0173 © 2012 Published by Elsevier Ltd. doi: 10.1016/j.protcy.2012.02.007 INSODE 2011 Verifying regional climate model results with web-based expert-system Emrullah Sonuca *, Baha Sena, Burak Senb aComputer Engineering Department/ Faculty of Engineering/ Karabuk University, Karabuk, Turkey bGeneral Directory of State Meteorological Service, Ankara, Turkey Abstract The verification system aims at monitoring the forecast quality over time. Verification helps improving the forecast quality by knowing the strengths and weaknesses of the existing forecasting system and by comparing the quality of different forecasting methodologies. Thus, the web-based verification system has been developed for verification of forecast results that is produced by International Center for Theoretical Physics Regional Climate Model v4 for our country. The forecasters and analysts can analyze the data in real-time with this web-based system. In this study, model values obtained from the system provided by ULAKBIM High Performance and Grid Computing Center. Model and station values were compared with each other for verification of model results with observation values. Therefore model grid values are transferred to station by using bi-linear and nearest neighbor (k-NN) (proximal) interpolation methods. This process in meteorological literature is called grid to point technique. Verification methods for forecasts of continuous variables are used to verify forecast values with observation values. Some verification methods; Mean Error, Mean Absolute Error and Root Mean Square Error, are calculated for validation. Verification results are shown as table and graphics on web-based system which is developed by the power of PHP (PHP: Hypertext Preprocessor). © 2011 Published by Elsevier Ltd. Keywords expert verification system; climate model; grid to point verification; regcm; grads 1. Introduction Numerical weather prediction (NWP) uses atmospheric state variables (temperature, wind, humidity and pressure), some physical equations (motion, thermodynamics, continuity, hydrostatic equation) and mathematical models to make the weather forecast. The first try of the numerical weather prediction resulted in failure by Lewis Fry Richardson in 1920s. Basically, two group of models used for NWP; Global Models (GCM’s) and Limited Area Models. Limited area models are nested within global models (GCM’s) but some of these models don't consider hydrostatic equilibrium equations [1]. In a typical global climate model, the horizontal (spatial) resolution of atmospheric component is 250 km and ocean component is between 125 and 250 km. This resolution is not enough for lots of the details of the regional * Emrullah Sonuc. Tel.: +90 370 433 2021 . E-mail address: esonuc@karabuk.edu.tr . Available online at www.sciencedirect.com Open access under CC BY-NC-ND license. Open access under CC BY-NC-ND license. http://creativecommons.org/licenses/by-nc-nd/3.0/ http://creativecommons.org/licenses/by-nc-nd/3.0/ 25 Emrullah Sonuc / Procedia Technology 1 ( 2012 ) 24 – 30 climate characteristics. Turkish State Meteorological Service (TSMS) is working about the climate changes and trends for Turkey, to predict climate of the future. As the regional climate models, RegCM and PRECIS models are chosen because of the ability to represent Turkey [1]. In this study, observation data were provided by TSMS and forecast data were calculated by regional climate model RegCM V4.0 [2]. This process was performed on the system provided by ULAKBIM High Performance and Grid Computing Center. Forecast data decoded by The Grid Analysis and Display System (GRADS) which is an interactive desktop tool that is used for easy access, manipulation, and visualization of earth science data. Forecast and observation data are recorded in database with the web-based system coding in PHP. Before the verification, results are transferred from grid to point using bi-linear interpolation and the nearest neighbor interpolation. 2. Materials and Evaluation Methods 2.1 Regional Climate Model (RegCM) Each region or area has a characteristic features. For this reason, GCMs are not enough to make prediction on a limited area. Therefore regional climate models are developed. The Regional Climate Model System (RegCM), developed by The Abdus Salam International Centre for Theoretical Physics (ICTP). RegCM is hydrostatic and compressible atmospheric model in a limited area and consists of sigma-pressure levels. The first version of the model, RegCM1, was developed in 1989 and the last version that is RegCM4 has been released in 2010[2]. There are four main steps for RegCM: 1) The model equations, 2) parameterization, 3) projection and grid structure of the model, and 4) running the model[3]. Topography map of RegCM model which is used at this study is shown in Figure 1. This model has been run on the system provided by ULAKBIM High Performance and Grid Computing Center. Figure 1 - Topography map of RegCM4 model. 2.2 Observation and Forecast Data The observation data has taken as a regular file from TSMS. This file includes station number, year, month, day, air temperature, wind speed and more. And in another file includes station information like station name, latitude and longitude of station, address of the station etc. The air temperature at 2m (t2m) was used for verification. T2m values have been read through this file and the data of the stations were recorded in the database with the web-based application which was coded in PHP. Thus observation data of the stations has become available for verification. The forecast data has been created by RegCM. Model has been run on the super computers provided by ULAKBIM High Performance and Grid Computing Center. The 4th version of the RegCM has been installed on system. It has been run to create monthly surface output files between 1989 and 2007. Model used the ERA Interim data sets which are prepared by European Centre for Medium-Range Weather Forecasts (ECMWF) used to create the model initial and boundary data. At the end of the process surface output files are downloaded from ULAKBIM to the computer which has a web-based system for verification analysis. Surface output file is a binary data file and it has a data descriptor file contains a description of the binary data. The general contents of a gridded data descriptor file are as follows [4]: Filename for the binary data Missing or undefined data value 26 Emrullah Sonuc / Procedia Technology 1 ( 2012 ) 24 – 30 Mapping between grid coordinates and world coordinates Description of variables in the binary data set The data of these files has been decoded by web-based application with the user parameters using GRADS. Also the forecast data has been recorded in the database with the application. 2.3 Interpolation Interpolation is the process of using known data values to estimate unknown data values. [5] In this study, bi- linear interpolation and nearest neighbor interpolation method is applied to obtain forecast data from output files before verifying process. Bi-linear interpolation is an extension of linear interpolation for interpolating functions of two variables on a regular grid. The four red dots show the data points and the green dot is the point at which we want to interpolate [6]. Figure 2 – Interpolation to one point. In the figure 2; P: Desired Point, Q: Known Points (four grid values), R: Point on the line with the known points. Suppose that we want to find the value of the unknown function f at the point P = (x, y). It is assumed that we know the value of f at the four points Q11 = (x1, y1), Q12 = (x1, y2), Q21 = (x2, y1), and Q22 = (x2, y2). [6] We first do linear interpolation in the x-direction. This yields where R1 = (x,y1), (1) where R2 = (x,y2). (2) We proceed by interpolating in the y-direction. (3) This gives us the desired estimate of f(x, y). (4) The nearest neighbor interpolation is a simply method to reconstruct a function is to take for each position the value of the nearest sampling point. The nearest neighbor algorithm selects the value of the nearest point and does not consider the values of neighboring points at all, yielding a piecewise-constant interpolation. 27 Emrullah Sonuc / Procedia Technology 1 ( 2012 ) 24 – 30 2.4 Verification At first, standard statistical methods as well as graphical techniques were used. This method is called subjective verification. On the other hand, objective verification is numerical expression of verification result that is generated by comparing forecast and observation values . Of course, the difference values of all points on the map can be found. However, the difference values of a particular plane may not make sense alone. Therefore to see trend of difference values for the same parameter in a fixed point continue in a time is effective determining the accuracy of the forecasts in the calculation of objective verification. Because of this, verification of variation with time in a point is considered instead of a broad area [1]. In this study, the following verification statics were computed for compared values: Mean Error (ME), Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). (5) The ME is the simplest and most familiar of scores and can provide very useful information on the local behaviour of a given weather parameter (e.g. maximum temperature close to the coastline or minimum temperature over snow-covered ground). The ME range is from minus infinity to infinity, and a perfect score is = 0. However, it is possible to reach a perfect score for a dataset with large errors, if there are compensating errors of a reverse sign. The ME is not an accuracy measure as it does not provide information of the magnitude of forecast errors. The MAE range is from zero to infinity and, as with the ME, a perfect score equals = 0. The MAE measures the average magnitude of forecast errors in a given dataset and therefore is a scalar measure of forecast accuracy. It is advisable to always view the ME and the MAE simultaneously. For small or limited data sets the use of MAE is preferred. The mean absolute error is given by (6) The most common accuracy measure is the RMSE [15]. (7) The RMSE is sensitive to interpolation, phase error and the general variability or anomaly [7]. 3. Results 18 stations were selected for verification. These stations have been doing regularly synoptic observation on and basis at the airports. The data for a one year (365 observation data) belong to 1996 and 00 GMT (Greenwich Mean Time). These data compared with model prediction using two different interpolation methods and three separate statistical evaluations were analyzed. Results; verification score for each station with the Turkey average, location of the stations (continental or maritime) are in the table 1. Table 1. Verification scores of the year 1996 No. Station Number & Name L_ME L_MAE L_RMSE N_ME N_MAE N_RMSE M-C 1 17038-Trabzon -9.66 9.67 10.42 -9.33 9.33 10.11 M 2 17060-Istanbul Ataturk -4.24 4.94 5.95 -4.02 4.89 5.92 M 3 17082-Tokat -3.45 4.34 5.18 -3.45 4.34 5.19 C 4 17090-Sivas. -4.66 5.07 5.79 -4.61 5.03 5.76 C 28 Emrullah Sonuc / Procedia Technology 1 ( 2012 ) 24 – 30 5 17096-Erzurum -1.10 4.11 5.30 -1.01 4.07 5.27 C 6 17112-Canakkale -4.11 4.39 5.21 -4.05 4.35 5.17 M 7 17115-Balikesir -4.03 4.64 5.75 -4.04 4.65 5.77 M 8 17124-Eskisehir -3.89 4.42 5.22 -3.85 4.42 5.23 C 9 17128-Ankara Esenboga -2.97 3.84 4.65 -2.89 3.78 4.59 C 10 17170-Van -6.04 6.18 7.09 -5.54 5.73 6.69 C 11 17195-Kayseri -3.14 3.85 4.72 -3.15 3.87 4.74 C 12 17200-Malatya Erhac -5.16 5.60 6.73 -4.66 5.20 6.32 C 13 17202-Elazig -3.70 4.19 5.21 -3.71 4.19 5.22 C 14 17219-Izmir Adnan Menderes -1.66 3.26 4.01 -1.80 3.24 3.97 M 15 17244-Konya -4.43 4.68 5.45 -4.41 4.65 5.41 C 16 17260-Gaziantep -4.73 5.13 6.15 -4.69 5.08 6.09 C 17 17280-Diyarbakir -2.88 3.63 4.65 -2.79 3.58 4.60 C 18 17300-Antalya -6.12 6.15 6.80 -6.44 6.46 7.08 M Turkey_average -4.22 4.89 5.79 -4.14 4.83 5.73 C=Continental, M=Maritime, L_ME=Bi-linear interpolation mean error, L_MAE=Bi-linear interpolation mean absolute error, L_RMSE=Bi-linear interpolation root mean square error, N_ME=nearest neighbor interpolation mean error, N_MAE=nearest neighbor interpolation mean absolute error, N_RMSE=nearest neighbor interpolation root mean square error According to the results of the analysis; there is not a significant difference between bi-linear interpolation and nearest neighbor interpolation. In 18 stations, bi-linear interpolation results are better for 5 stations, nearest neighbor interpolation results are better for 12 stations and in 1 station, both methods have same results. In verification scores about both interpolation methods, differences are over than ±0.1 for 4 stations and are lower than ±0.1 for 14 stations. Given this assessment, it can be said that there is not a significant difference for selected stations between both methods and evaluated data. The bias, which is calculated for each day, is shown as a graphically in figure 3. Figure 3 – BIAS Line Graphic for Istanbul Ataturk Airport Station (Sample for 15 days) ME and MAE results are close to each other. It shows that the model gives over/lower estimate continuously. In the cases of over-estimate and lower estimate (such as in the Erzurum and Izmir examples) ME value is significantly smaller than MAE value. There is not a significant difference for RMSE and MAE scores. According to the technique of nearest neighbor interpolation; for the scores of 18 stations, RMSE values are between 3.97 and 10.11 and, MAE values are between 3.2 and 9.33, ME are between -9.33 and -1.8. Average of Turkey (for 18 stations) were determined as 5.73, 4.83, -4.14 respectively. According to the technique of bi-linear interpolation; for the scores of 18 stations, RMSE are between 4.01 and 10.42, MAE are between 3.26 and 9.67, ME are between -9.66 and -1.66. Average of Turkey ( for 18 stations) were determined as 5.79, 4.89, -4.22 respectively. 29 Emrullah Sonuc / Procedia Technology 1 ( 2012 ) 24 – 30 In two interpolation methods, the climate model gives lower estimate for 00 GMT. The reason of this situation is reflecting the radiation-induced cooling at night more than usual by model. To determine this condition exactly, it should perform the verification using observation values including 06, 12, 18 GMT and daily average temperature values and analyze the results. The aim of this study is to give information of the software and to show sample application. Therefore mentioned evaluations will be done next studies. 4. Conclusion All over around the world, many meteorological services develop a forecast verification system to verify forecast data. The accuracy of the model predictions are increasingly important to changing climate conditions. Therefore verification data sets created to use in forecast verification models. Demirtas, Nance, Bernardet, Lin, Chuang, Loughe, Ma-honey, Gall and Koch (2005) [8] have developed the verification system includes National Centers for Environmental Prediction (NCEP)’s surface and upper-air verification package, which employs a grid-to-point (station based) approach. Mahoney, Henderson, Brown, Hart, Loughe, Fischer and Sigren (2005) [9] have developed the real-time verification system (RTVS) to provide a statistical baseline for weather forecasts and model based guidance products, and to support real-time forecast operations, model-based algorithm development, and case study assessments. Şen and Aydın [10] have developed a real-time verification system. In their study, inverse distance weighting interpolation method is applied to forecast data before verification. In this study, the mentioned web-based application is an alternative system to verify t2m values by meteorologists or relevant persons. The forecast values are created by RegCM which is running on the system provided by ULAKBIM. This application supports both objective and subjective verification. The subjective verification provided with the graphics and charts and the objective verification provided with the tables including forecast parameter values. Furthermore mean error (ME), mean absolute error (MAE) and root mean square error (RMSE) values are calculated with the objective verification. As a result of this study, the mentioned web-based application is developed for verifying surface output files including t2m values. This application could develop to add new parameters such as precipitation etc. And radiation output files can be added to system to verify new parameter values. References 1. Sayısal Hava Tahmini Şube Müdürlüğü Sayısal Hava Tahmini, Retrieved May 19, 2011 from the World Wide Web: http://www.dmi.gov.tr/FILES/genel/sss/sayisalnedir.pdf. 2. ICTP - Regional Model: REGCM4, Retrieved May 20, 2011 from the World Wide Web: http://www.ictp.it/research/esp/models/regcm4.aspx 3. Morris, D. (2010). E-confidence or incompetence: Are teachers ready to teach in the 21st century?. World Journal on Educational Technology, 2(2), 142-155. 4. B. ŞEN, Bölgesel İklim Modelleri Kullanılarak Çukurova Yöresi’nde İklim Değişikliğinin 1. Ve 2. Ürün Mısır Verimine Olası Etkilerinin Belirlenmesi, Ph.D. Thesis, Çukurova University, Adana, 2007. 5. Components of a GrADS Data Descriptor File, Retrieved May 20, 2011 from the World Wide Web: http://www.iges.org/grads/gadoc/descriptorfile.html 6. Interpolation Techniques, Retrieved May 21, 2011 from the World Wide Web: http://iridl.ldeo.columbia.edu/dochelp/StatTutorial/Interpolation/ 7. Bilinear interpolation, Retrieved May 21, 2011 from the World Wide Web: http://en.wikipedia.org/wiki/Bilinear_interpolation 8. The Root Mean Square Error (RMSE) Retrieved May 21, 2011 from the World Wide Web: http://www.ecmwf.int/products/forecasts/guide/The_RMSE.html 30 Emrullah Sonuc / Procedia Technology 1 ( 2012 ) 24 – 30 9. M. Demirtaş, L. Nance, L. Berdardet, Y. Lin, H.Y. Chuang, A. Loughe, J. Mahoney, R. Gall and S. Koch. The Developmental Testbed Center Verification System. WRF/MM5 User’s Workshop. 27-30 June, 2005 10. J.L. Mahoney, J.K. Henderson, B.G. Brown, J.E. Hart, A. Loughe, C.Fischer and B.Sigren, . The real-time verification system (RTVS) and its application to aviation weather forecasting 10th Conference on Aviation, Range, and Aerospace Meteorology, 2002. 11. B. Şen and N. Aydın, A Real-Time Verification System Development Using Regional Climate Model Results. The 1st International Symposium on Computing in Science & Engineering ISCSE, 2010. 
work_6mn6et62cjcqzfxjoxsjpa2xmi	----	A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments Applied Computing and Informatics (2015) 11, 116–129 Saudi Computer Society, King Saud University Applied Computing and Informatics (http://computer.org.sa) www.ksu.edu.sa www.sciencedirect.com ORIGINAL ARTICLE A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments * Corresponding author. E-mail address: folorunsoo@funaab.edu.ng (O. Folorunso). Peer review under responsibility of King Saud University. Production and hosting by Elsevier http://dx.doi.org/10.1016/j.aci.2014.07.001 2210-8327 ª 2014 King Saud University. Production and hosting by Elsevier B.V. All rights reserved. Olusegun Folorunso *, Olusegun Afeez Mustapha Department of Computer Science, Federal University of Agriculture, Abeokuta, Nigeria Received 13 January 2014; revised 20 June 2014; accepted 3 July 2014 Available online 18 July 2014 KEYWORDS Trust-based access; Control; Fuzzy-expert systems; Crowdsourcing environment; Quality assurance; Trust value Abstract Crowdsourcing has been widely accepted across a broad range of application areas. In crowdsourcing environments, the possibility of performing human computation is characterized with risks due to the openness of their web-based platforms where each crowd worker joins and participates in the pro- cess at any time, causing serious effect on the quality of its computation. In this paper, a combination of Trust-Based Access Control (TBAC) strategy and fuzzy-expert systems was used to enhance the quality of human computation in crowdsourcing environment. A TBAC-fuzzy algorithm was developed and imple- mented using MATLAB 7.6.0 to compute trust value (Tvalue), priority value as evaluated by fuzzy inference system (FIS) and finally generate access decision to each crowd-worker. In conclusion, the use of TBAC is feasible in improving quality of human computation in crowdsourcing environments. ª 2014 King Saud University. Production and hosting by Elsevier B.V. All rights reserved. http://crossmark.crossref.org/dialog/?doi=10.1016/j.aci.2014.07.001&domain=pdf mailto:folorunsoo@funaab.edu.ng http://dx.doi.org/10.1016/j.aci.2014.07.001 http://www.sciencedirect.com/science/journal/22108327 http://dx.doi.org/10.1016/j.aci.2014.07.001 A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments 117 1. Introduction Crowdsourcing is sometimes referred to as human computation, a methodology that lets humans’ process tasks which are difficult to implement in software. These tasks may include transcription of documents, reviewing of articles or eval- uating the quality of ranking algorithms [20]. Crowdsourcing has recently shown enormous potential in solving highly decentralized target localization tasks [13]. In a crowdsourcing system, tasks are distributed to a population of anonymous Internet users for completion. Man-Ching et al. [12] and John et al. [14] defined crowdsourcing as the use of large distributed groups of people to complete micro tasks or to generate information. Because traditional search relevance evaluation requiring expert assessment is a lengthy process [2,6,8,14], crowdsourcing has gained traction as an alternative solution for these types of high volume tasks [1,2,8]. A crowdsourcing site has two groups of users: requesters and workers. The crowdsourcing site exhibits a list of available tasks, associating with reward and time period, that are presented by requesters; and during the period, workers com- pete to provide the best submission [12]. Meanwhile, a worker selects a task from the task list and completes the task because the worker wants to earn the asso- ciated reward. At the end of the period, a subset of submissions is selected, and the corresponding workers are granted the reward by the requesters [12]. Crowdsourcing is now being pushed by large IT companies such as Amazon, Google, or Yahoo!. They have recognized the opportunities behind such mass col- laboration systems [14] for both improving their own services as business case. In particular, Amazon focuses on a task-based crowdsourcing platform called Amazon Mechanical Turk (AMT) [20]. Requesters are invited to issue human-in- telligence tasks (HITs) requiring a certain qualification to the AMT. These crowd- sourcers post mostly simple tasks that, however, require human capabilities. In particular, 50% of tasks are processed at a cost of $0.10 and less, most of the tasks are usually also offered in chunks to multiple AMT workers [12,20]. Meanwhile, the major problem with current crowdsourcing environments as described in [5,26] is the lack of manageability as a result of the openness of Web based platforms, where anybody can join and participate which make quality assurance challenging. From literature, Quality Assurance (QA) is seen as a major determinant of e-satisfaction in most business-to-consumer (B2C) and crowdsourcing environ- ments inclusive. QA is further designed into a theoretical framework or a model. These models are Rayport and Jaworski 7Cs Framework [19], DeLone and McLean Electronic Commerce Model [7], The ISO 9126 Quality Model (ISO/ IEC 9126, 2011), SERVQUAL [17], WebQual 4.0 Model [3], Palmer’s Model [16] and Stefani and Xenos Quality Model [22]. These models were drawn from the fields of information systems, marketing, human–computer interaction and design and the researchers ‘‘focus has been on online consumers’’ perspective [4]. 118 O. Folorunso, O.A. Mustapha QA factors of online transactions as aggregated from QA models is usually expressed by performance characteristics such as response time, throughput, pre- dictability, availability, scalability, stability, usability, reliability, efficiency, func- tionality, speed of download (bit rate), ease of finding information, timeliness, confidentiality and transaction integrity. In this case, we looked into confidential- ity and transaction integrity by focusing on Trust-Based Access Control (TBAC). Trust exists in our daily life and thus can be used as a mechanism to make access decisions in networks [27]. Trust models have also been proposed to control anonymity and uncertainty [9,10]. Several novel access control models based on trust have been proposed in recent years. TrustBAC extends the conventional Role Based Access Control (RBAC) with the notion of trust level [25] in which users are assigned to trust levels instead of the roles based on user identities, behavior history, recommendation, etc. The role that is assigned to a user changes along with the change of the trust level of the user. Trust management systems have been introduced in order to create a coherent framework to deal with trust [11]. Since the importance of building trust models has become vital for the development of some nowadays computer systems the way this trust is derived, that is, the metrics also becomes crucial. Metrics become very important for the deployment of trust management systems as the way of quantifying trust [11]. The simplest way to define a trust metric is by using a dis- crete model where an entity can be either ‘‘trusted’’ or ‘‘not trusted’’. This can also be expressed by using numerical values such as 1 for trusted and 0 for not trusted [11]. However, trust management models cannot address security issues at fullest [18]. There is also need of explicit trust model which address trust access control for crowdsourcing environment. Access control mechanism based on trust cal- culations using fuzzy approach is presented in [24] where access feedback is used for access control. This paper proposes the fuzzy expert systems approach for TBAC which is necessary in capturing the trust calculations and crowd worker identity of context based trust relationship in crowdsourcing environments as inputs and access deci- sion as output to the system. The rest of the paper is organized as follows. In Section 2, we discuss the rele- vant work in the areas of trust and crowd-sourcing, Section 3 is on trust and access model, fuzzy-expert systems design, and CrowdTBAC-fuzzy framework, Section 4 presents results and discussion; and Section 5 concludes the work. 2. Related work This section presents discusses some related works that are relevant to our research ideas and purpose stated earlier. Sara et al. [21] presented a model for users’ reputation in Wikipedia. In their research they first parse and mine the entire English Wikipedia history pages in A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments 119 order to extract detailed user edit patterns and statistics. The pattern and statistics were then used to derive three computational models of a user’s reputation. Finally, these models were validated using ground-truth Wikipedia data associated with vandals and administrators and when used as a classifier, the best model pro- duces an area under the Receiver Operating Characteristics (ROC) curve of 0.98. An automated system was developed to estimate the reputation Ri(t) of a Wikipedia user i at time t based on its past behavior. The reputation index Ri(t) was assumed to be positive and scaled between 0 and 1 and, for the moment, and is assumed to be loosely interpretable as the probability that i produces high-quality content. A step was taken toward this long term goal by developing several computational models of Ri(t) and testing them, in the form of classifiers, on the available ‘‘ground-truth’’ associated with Wikipedia-known administrators and vandals. Their approach, which is fairly standard in machine learning appli- cations, requires some explanations. It was reasonable to assume that there exists a true reputation function that is scaled between 0 and 1 and grows monotonically from the user with the lowest reputation to the user with the highest reputation. The first of the three models that was proposed in their paper measures the rep- utation of a user as tokens are inserted by the user. In the work, it was discovered that a users’ reputation decreases when its earlier tokens inserted were deleted by another user. The second reputation model measures the time interval between insertion and deletion of content. While the third model was developed as a varia- tion of the second model. (Other works are summarized in Supplementary Table 1.) 3. TBAC-fuzzy framework design This section explains the Trust-Based Access Control model; trust metrics, fuzzy- expert systems design, TBAC-fuzzy framework and a crowd-sourcing scenario. In this case we first consider a scenario. 3.1. A crowd-sourcing scenario Let consider an Internet encyclopedia as a crowd-sourcing environment that allows everyone that has access to the Internet to make contributions or edit arti- cles online. In such environment, workers are classified based on their compe- tences and experiences. The crowd workers can freely access the encyclopedia network but are allowed to make contributions or edit some certain articles based on its trust value acquired during the interaction with the host network. The trust value of each crowd worker is greatly affected by the number of activities per- formed. The priority level of workers is determined by the rules generated. The crowd-sourcer specifies a priority value to an existing articles which denotes the sensitive level of that articles which means a worker must have such specified 120 O. Folorunso, O.A. Mustapha priority value or more before access can be granted to make contribution to such an article (Supplementary Fig. 1). 3.2. Trust based access control model for crowdsourcing environments Trust based access control model computes trust quantitatively based on their interaction in a crowdsourcing environment. In order to achieve this, a trust model metric is introduced to evaluate the trust of the workers and then use the trust value to enforce the access control policy with fuzzy expert system. Our trust based access control model is described as a 3-tuple set model {Wr, Td, Wt}. This describes the set of workers (Wr), trust degree (Td), and workers’ trust degree (Wt). We further arrived at the following definition: Definition 1. The set of worker (Wr) is a finite set of group workers. Crowd- workers are classified into; highly skilled, skilled, semiskilled and un-skilled worker. Wr ¼fwr1;wr2; . . . ;wrng Definition 2. Trust degree (Td) measures the degree of trust of a worker. This value is derived from a trust metric, which is later classified in a linguistic (mem- bership) form. td 2 Td; tdfhigh;average; low;verylowg Definition 3. Worker’s trust degree: is a Cartesian set of a worker to a trust degree. Wt # Wr�Td; wtðwrÞ¼ td;fwr 2 Wr and td 2 Tdg 3.3. Trust metric An access control policy considers the workers’ trust degree when deciding whether to grant a worker access to perform tasks. The trust value of a worker is determined as the worker visits the host. A trust metric equation from [15] was adapted so as to compute the trust values of every worker from the crowd. The Tvalue is the trust value of a worker wr 2 Wr on a visit to the host, Ei is an events, action or service carried out on the host site by a worker Wr. The value of the weight Wi is assigned to each event Ei by the host [15].P TvalueðwrÞ¼ T0 þ n i¼0WiEiPn i¼0Wi þ fðtÞ ð1Þ A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments 121 T0 is the initial trust value assigned to a crowd-worker. The initial value has a rela- tion to the evaluation of trust value of a worker. Hence, the crowd-sourcer defines T0 value for a new and existing worker. The term f(t) is added to the equation to reflect any time-dependent activity (or inactivity) to suggest gain or loss of reliability. The f(t) value represents error in the model when the worker does not have access to the environment due to a net- work being down or system itself is down. The trust value of a worker is evaluated based on event performed by the worker. Such an event carries a value and it’s also weighted. An event can be nega- tive if the worker tries to act maliciously; such could be a failure to login or a worker performing brute force attack on the host. The positive value event which varies are the normal event that a worker is expected to carry out on a host. An anonymous worker is classified as unskilled worker with initial value of 0.0 since the worker is new to the environment and the competence is yet to be ascer- tained while other workers (semiskilled, skilled, highly skilled) are known in the environment. Anonymous worker can eventually gain recognition to become known in the environment. An existing worker which is known to the environment has a level of skills and can be classified either as skilled worker, semiskilled and highly skilled depending on the worker’s performance history. 3.4. Fuzzy expert system design The fuzzy expert system has 2 input values and 1 output. The fuzzy sets (i.e. lin- guistic value) for the input variables and the output variable are as follows: i. Degree of trust denoted as ‘‘TrustDegree’’ is defined with four fuzzy sets {high, average, low, verylow}. The ‘‘TrustDegree’’ membership functions are shown in Fig. 4 of the next section. The fuzzy set ‘‘VeryLow’’, ‘‘Low’’, ‘‘Average’’ and ‘‘High’’ sets are in trapezoidal form. ii. Type of crowd worker denoted as ‘‘CrowdWorker’’ with four fuzzy sets (highly skilled worker, skilled worker, semiskilled worker, and un-skilled worker). iii. Priority level denoted as ‘‘priority_Level’’ with four fuzzy sets {high, average, low, verylow}. The ‘‘Priority_Level’’ membership functions are also shown in Fig. 5 where the ‘‘VeryLow’’, ‘‘low’’, ‘‘average’’ and ‘‘High’’ sets are in trape- zoidal form. The linguistic input variable TrustDegree and CrowdWorker are defined in the Table 2, while Table 3 shows the linguistic output variable Priority_Level. The crisp range value was selected based on the degree of membership in the interval [0,1], where 0 and 1 confirms no membership and full membership respectively. The membership function is trapezoidal as it is specified by four parameters (a, b, c, d) as follows: Table 2 The fuzzy set for the two inputs with their linguistic value and crisp values. Linguistic variable Fuzzy set (linguistic value) Crisp range Trust degree High l P 0.75 Average 0.5 6 l 6 0.7 Low 0.25 6 l 6 0.6 Very low l 6 0.35 Crowd workers Highly skilled worker (HS) HS P 0.75 Skilled worker (S) 0.5 6 S 6 0.7 Semiskilled worker (SS) 0.25 6 SS 6 0.6 Un-skilled worker (US) US 6 0.35 Table 3 Fuzzy set for the output. Output Fuzzy set Crisp value Priority_level High_priority 0.75 above Average_priority 0.5–0.75 Low_priority 0.25–0.5 Very low_priority 0.0–0.25 122 O. Folorunso, O.A. Mustapha Trapezoidal ðx; a;b;c;dÞ¼ 0; x � a x�a b�a ; a � x � b 1; b � x � c d�x d�c ; c � x � d 0; d � a 8>>>>>>< >>>>>>: 9>>>>>>= >>>>>>; ð2Þ An alternative expression using min and max:� �� � Trapezoidal ðx; a;b;c;dÞ¼ max min x�a b�a ;1; d�x d� c ;0 ð3Þ 3.5. TBAC-fuzzy framework for crowdsourcing environments The proposed framework comprises of the Trust-Based Access Control and fuzzy expert system. The trust based access control (TBAC) has a model that computes the trust value of workers. The trust metric operates at the preprocessing while the access control module operates at postprocessing module. Supplementary Fig. 2 shows the TBAC and Fuzzy expert system components incorporate into a single framework. The TBAC-fuzzy expert system can be explained with the following module of the framework. Preprocessing module: The first module of the framework contains the user module and trust metric module of the TBAC. At this preprocessing module, A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments 123 the user module classifies workers into different classes while the trust metric module computes the trust value of the worker using Eq. (1). The preprocessing module provides a crisp value which serves as the input value to the fuzzy- expert system. These crisp values are forwarded to the fuzzification module and decode to a fuzzy value. Fuzzification module: In the process of fuzzification, membership functions defined on input variables are applied to their actual values so that the degree of trust for each rule premise can be determined. The crisp values are fuzzified into linguistic values. Fuzzy inference engine module: The inference engine consists of the knowledge base which contains the rules for rating workers (generated rules are available in the supplementary material). Defuzzification: In defuzzification, the fuzzy output set is converted to a crisp number. Some commonly used techniques are the centroid and maximum methods. � In the centroid method, the crisp value of the output variable is computed by finding the variable value of the center of gravity of the membership function for the fuzzy value. � In the maximum method, one of the variable values at which the fuzzy set has its maximum truth value is chosen as the crisp value for the output variable. Postprocessing module: In this module, access right is defined on the workers based on the output result of the fuzzy expert system. The output of the system shows the priority level of the worker. The environment filters the worker based on the output and decides access right to workers that falls within a priority level and thereby improves the quality of the computation. A CrowdTBAC Fuzzy algorithm is described in (Algorithm available in supplementary material). 4. Simulation, results and discussion 4.1. Result Our fuzzy expert system trust model is evaluated using a MATLAB fuzzy logic toolbox version 7.0.6. The fuzzy interference system of our TBAC-fuzzy expert system has two inputs and an output, a mamdani-type inference and a centroid method were used for defuzzification. The implementation process was carried out as shown in Figs. 4–8. Firstly, we defined the number of input and the output on fuzzy interference system editor and define the membership functions of the lin- guistic variable. The surface-viewer above reflects the trust value relative to trust and workers that may help us to analyze trust variance. Fig. 8 shows the output surface for Figure 4 Membership function of TrustDegree. Figure 5 The membership function of an output variable ‘‘TaskPriority’’. 124 O. Folorunso, O.A. Mustapha the priority value versus trust value and worker. This priority value output is a fac- tor that controls and enhances the quality assurance of the crowd-sourcing activities. 4.1.1. Simulation TBAC-fuzzy expert system is simulated in our crowd-sourcing environment sce- nario. The simulation parameters are shown in Table 5 with the range of values based on assumptions. In this process, 10 workers of different classes (5 unskilled, 2 semiskilled, 2 skilled and 1 highly skilled worker) are considered to make con- tribution in encyclopedia environment. Within the range of values for event and weight defined below, three distinct values are generated using a java program- ming language. The values generated are further used to compute the workers’ trust values on the assumption that same events were performed in Table 4. Figure 6 The membership function of CrowdWorkers. Figure 7 The rule surface shows how the membership function of the input variable influences the outcome result. A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments 125 The initial values (T0) of the workers are considered in the computation to determine their actual trust value. In this case, we defined T0 to be 0.00, 0.01, 0.025, and 0.05 for a new worker (unskilled), semiskilled, skilled and highly skilled workers respectively. The last three workers are the existing workers and are clas- sified based on their past achievements and experiences. The designed fuzzy rules are simulated using the trust values of crowd-workers with a fuzzy tool box in MATLAB. The simulation graph above revealed that a task with a priority value of 0.37 can only be performed by workers that gain such a value or more. The threshold value of 0.37 is been defined by the crowd-sourcer Figure 8 The surface view which shows the dependency of the output on the two inputs. Table 4 Results of simulation on MATLAB. Workers crisp value Trust crisp value Priority crisp value Dval = 0.37 (user-defined value to grant access) 0.25 0.3547 0.1271 Denied 0.25 0.47668 0.1342 Denied 0.25 0.40471 0.3988 Granted 0.5 0.53168 0.4864 Granted 0.5 0.59501 0.6 Granted 0.25 0.53451 0.2257 Denied 0.25 0.58843 0.3765 Granted 0.5 0.60677 0.3987 Granted 0.5 0.65498 0.6 Granted 1.0 0.73293 0.7508 Granted Table 5 simulation parameters. Variable Data type Values range Probability distribution Events (E) Real 0.0–0.5 Uniform Weight (W) Real 0.0–0.1 Uniform T0 for highly skilled Real 0.05 Uniform T0 for skilled Real 0.025 Uniform T0 for semiskilled Real 0.01 Uniform T0 for un-skilled Real 0.0 Uniform 126 O. Folorunso, O.A. Mustapha for basic trust level. The threshold/user-defined (Dval = 0.37) value represented by the red line as shown in Fig. 9 gives the crowd-source an insight into the worker that is allowed to performed a task. Any worker below the threshold level will be denied access on the task. It was observed furthermore that not all the unskilled workers can be allowed to carry out a task. Individual workers’ priority level is determined by fuzzy rules Figure 9 Simulation chart. A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments 127 generated using the two inputs (classes of worker and trust value) to determine the output (priority value). The graph also clearly shows that not all two semiskilled workers can be given access to a task of 0.45 priority value that could be defined by the crowd-sourcer. 4.2. Discussion The defined value chosen by the environment signifies the sensitive level of the activities to be performed. A lower value say 0.2 which could signifies an activities which required worker with little knowledge or experience and this will accept more workers which means 2 workers will be rejected and denied access to make contributions. Whenever a worker visits the host, trust computations were made based on some events performed on the system. The events on the system have a value assigned and also they are weighted. In the design architecture, trust metric module compute workers trust and transform into a crisp value which is then decoded at the fuzzification module. The surface view in Fig. 7 shows the dependency of the output on the two inputs. Postprocessing module of the fuzzy expert system grants access to a worker. This module makes the decision based on the output of the fuzzy system. This module filters and classifies the needed worker based on a defined priority value to make contribution or suggestion. The trust metric equation of TBAC (Eq. (1)) was tested with the fuzzy expert system by some generated values. The out- come of the result has it that majority of the workers gained access to perform activities in the crowdsourcing environment while others are denied. Meanwhile, this can be justified for the fact that half of the workers that participated have little 128 O. Folorunso, O.A. Mustapha or more experience while the other workers have little or no experience. In this case a task with a priority value 0.37 can be worked on by 70% of the workers as shown in the simulation chart in Fig. 9. This analysis with the simulation chart helps to enhance the quality assurance of the crowdsourcing activities. 5. Conclusions and future work This paper addressed the issues of trust in crowd-sourcing alongside its incentive and how the quality of response from the crowd-workers would be improved. The paper identified answer ranking as a problem in crowd-sourcing leading to cheat- ing by workers. To look for more efficient and reliable way to solve this problem, a fuzzy expert system and trust based access control were employed as a single framework to enforce access control. A trust based access control fuzzy algorithm was used to compute and assigned trust value to workers. The trust value assigned to crowd-worker serves as one of the inputs to the fuzzy system and access is granted to workers whose priority value is above threshold level based on the fuzzy output. In this case, the crowdsourcing could be assured that only workers with little or more experience could perform certain tasks. This however helps to enhance the quality of response from the crowd. Experiment was conducted on Trust-Based Access Control fuzzy algorithm using systematic generated values on MATLAB. The results indicated that high quality response and service could be obtained from CrowdWorkers. In the future, effort on mood positivity with desired model would be conducted to determine the performance of workers in a crowdsourcing environment. Appendix A. Supplementary material Supplementary data associated with this article can be found, in the online ver- sion, at http://dx.doi.org/10.1016/j.aci.2014.07.001. References [1] O. Alonso, Guidelines for designing crowdsourcing-based relevance evaluation, ACM SIGIR, 2009. [2] O. Alonso, D.E. Rose, V. Stewart, Crowdsourcing for relevance evaluation, SIGIR Forum 42 (2) (2008) 9– 15. [3] S.J. Barnes, R. Vidgen, Measuring web site quality improvements, a case study of the forum measuring web site quality improvements: a case study of the forum on strategic management knowledge exchange, Ind. Manage. Data Syst. 103 (5) (2003) 297–309. [4] Z. Balfagih, N. Mohamed, M. Mahmud, Search of a model for evaluating the quality of e-commerce websites, in: Proceedings of ITSIM’08, Malaysia, 2008. [5] D. Brabham, Crowdsourcing as a model for problem solving: an introduction and cases, Convergence: Int. J. Res. New Media Technol. 14 (1) (2008) 75–90. [6] B. Carterette, P.N. Bennett, D.M. Chickering, Here or there: preference judgments for relevance, in: Proceedings of the European Conference on Information Retrieval (ECIR), 2008. [7] W.H. DeLone, E.R. McLean, The DeLone and McLean model of information systems success, a ten-year update, J. Manage. Inf. Syst. 19 (4) (2003) 9–30. http://dx.doi.org/10.1016/j.aci.2014.07.001 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0010 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0010 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0020 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0020 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0020 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0030 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0030 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0040 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0040 A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments 129 [8] G. Dupret, V. Murdock, B. Piwowarski, Web search engine evaluation using click-through data and a user model, in: Proceedings of the Workshop on Query Log Analysis WWW, 2007. [9] Eric Yuan, Jin Tong, Attributed based access control for web services, in: Proceedings of IEEE International Conference on Web Services, 2005, pp. 561–569. [10] V. George, M. Ian, C. Maria, A probabilistic model for trust and reputation, in: van der Hoek, Kaminka, Lespérance, Luck and Sen (Eds.), Proc. of 9th Int. Conf. on Au-tonomous Agents and Multiagent Systems (AAMAS 2010), May 10–14, Toronto, Canada, 2010. [11] Isaac Agudo, Carmen Fernandez-Gago, Javier Lopez, A model for trust metrics analysis, in: 5th International Conference on Trust, Privacy and Security in Digital Business (TrustBus ‘‘08), LNCS 5185, Springer, Turin, Italy, 2008, pp. 28–37. [12] Y. Man-Ching, K. Irwin, L. Kwong-Sak, A survey of crowdsourcing systems, IEEE International Conference on Privacy, Security, Risk, and Trust, and IEEE International Conference on Social Computing, 2011, pp. 766–773. [13] V. Matteo, R. Alex, J. Nicholas, Trust-based fusion of untrustworthy information in crowdsourcing applications, in: Ito, Jonker, Gini, Shehory (Eds.), Proceedings of the 12th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2013), May 6–10, Saint Paul, Minnesota, USA, 2013, pp. 829–836. [14] L. John, E. Andy, H. Vaughn, B. Lukas, Ensuring quality in crowdsourced search relevance evaluation: the effects of training question distribution, in: Proceedings of the SIGIR 2010 Workshop on Crowdsourcing for Search Evaluation (CSE 2010), July 23, 2010, pp. 17–20. [15] A.R. Patnaik, A.G. Srivastava, A.R. Nayak, Building flexible trust models for E-commerce applications, 2006, <http://tecsis.ca/services/2.Trust%20Model.pdf> (accessed 21.11.13). [16] J.W. Palmer, Web site usability, design, and performance factors, Inf. Syst. Res. 13 (2) (2002) 151–167. [17] A. Parasuraman, V.A. Zeithaml, L.L. Berry, SERVQUAL: a multiple-item scale for measuring consumer perceptions of service quality, J. Retail. 64 (1) (1988) 12–40. [18] Parikshit N. Mahalle, Pravin A. Thakre, Neeli Rashmi Prasad, Ramjee Prasad, A fuzzy approach to trust based control in internet of things, IEEE 3rd International Conference on Wireless Communications, Vehicular Technology, Information Theory and Aerospace & Electronics Systems Technology (Wireless ViTAE – 2013), Atlanta City, USA: IEEE, 2013. [19] J.E. Rayport, B.J. Jaworski, E-Commerce. McGraw-Hill Higher Education, NewYork, 2001. [20] B. Satzger, H. Psaier, D. Schall, S. Dustdar, Auction-based crowdsourcing supporting skill management, Inf. Syst. 38 (4) (2012) 547–560, http://dx.doi.org/10.1016/j.is.2012.09.003. [21] J. Sara, G. Yasser, L. Cristina, B. Pierre, User contribution and trust in Wikipedia, in: Proceedings of the 5th International Conference on Collaborative Computing: Networking, Applications and Worksharing, Washington DC, USA, 2009. [22] A. Stefani, M. Xenos, A model for assessing the quality of e-commerce systems, Proc. PC-HCI 2001 (2001) 105–109. [23] M. Shunan, H. Jingsha, Z.W. Xunboshuaiand, Access control mechanism based on trust quantification, IEEE Second International Conference on Social Computing (SocialCom 2010), volume issue, Minneapolis, USA, August 20–22, 2010, pp. 1032–1037. [24] C. Sundip, R. Indrajit, TrustBAC: integrating trust relationships into the RBAC model for access control in open systems. in: Proceedings of 11th ACM Symposium on Access Control Models and Technologies, 2006, pp. 49–58. [25] M. Vukovic, Crowdsourcing for enterprises, in: Proceedings of the 2009 Congress on Services, IEEE Computer Society, 2009, pp. 686–692. [26] N. Xudong, L. Junzhou, A trust aware access control in service oriented grid environment, in: Proceedings of 6th International Conference on Grid and Cooperative Computing, 2007, pp. 417–422. http://tecsis.ca/services/2.Trust%20Model.pdf http://refhub.elsevier.com/S2210-8327(14)00020-9/h0100 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0105 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0105 http://dx.doi.org/10.1016/j.is.2012.09.003 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0135 http://refhub.elsevier.com/S2210-8327(14)00020-9/h0135 A fuzzy expert system to Trust-Based Access Control in crowdsourcing environments 1 Introduction 2 Related work 3 TBAC-fuzzy framework design 3.1 A crowd-sourcing scenario 3.2 Trust based access control model for crowdsourcing environments 3.3 Trust metric 3.4 Fuzzy expert system design 3.5 TBAC-fuzzy framework for crowdsourcing environments 4 Simulation, results and discussion 4.1 Result 4.1.1 Simulation 4.2 Discussion 5 Conclusions and future work Appendix A Supplementary material Appendix A Supplementary material References 
work_6nikb2xfjnhytgasnakllqgncy	----	A Yugoslav Journal of Operations Research 13 (2003), Number 2, 187-197 AN EXPERT SYSTEM FOR RANKING COMPANIES AND INVESTMENTS: WOOD INDUSTRY CASE* Lazaros HLIADIS+, Theodoros KOUTROUMANIDIS‡, Garyfallos ARABATZIS+, Charalambos ARAPATSAKOS‡ +Department of Forestry and Environmental Management and Natural Resources Democritus University of Thrace omega@netfiles.gr, liliadis@fmenr.duth.gr garamp@fmenr.duth.gr ‡Department of Agricultural Development Democritus University of Thrace xarapat@agro.duth.gr Communicated by Byron Papathanassiou Abstract: Wood industry is a very important part of both the Greek Rural and industrial sector. The discovery of the differentiation in the level of growth and in the quality of financial management between the Greek wood companies can provide very important aid in the design of an effective rural development policy. The evaluation and ranking of Greek wood companies based on actual financial data is a very complicated task and it requires expertise knowledge and skills. On the other hand a computer expert system can perform validation and evaluation in an efficient way and can substitute human experts. An expert system was designed and developed towards this direction. It uses multicriteria analysis for each one of the wood companies based on actual financial data and it applies fundamental principles of fuzzy logic in order to calculate the expected intervals of flows for the following years. Keywords: Multicriteria analysis, computer expert systems, wood industries, fuzzy logic. * Presented at 6th Balkan Conference on Operational Research mailto:omega@netfiles.gr mailto:garamp@fmenr.duth.gr mailto:xarapat@agro.duth.gr L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 188 1. INTRODUCTION The systematic and organized processing of wood by the utilization of contemporary means and the application of basic technological rules to the production line has a long history in Greece. The two most important time-periods for the development of Greece's wood industry are the periods 1938-1940 and 1965-1970. Especially from 1965 to 1970 the rapid technological development (in means and in production methods) the development of new wood products and the development of the international commerce, resulted in historic changes in wood industry [13]. On the other hand Greece had not developed any kind of serious policy, regarding the size and the structure of the units, the adequacy of raw materials and the consuming requirements. Consequently, this evolution took place without any programming and without any design. The lack of raw materials, the large cost of production, the small demand of the market, the development of many micro units (most of them were family business) and the intense competition, have resulted in the reorganization of the brunch in a vertical way. The results were the development of bigger units and the redistribution of land-planning [13]. The wood companies with more than 10 employees (130 establishments) participate with 1.2% to the gross value of production, with 1.7% to the employment and with 0.5% to the exportation of the aggregation of manufacture [5]. The eight main wood industries of Greece were chosen to be evaluated, based on financial data. The financial data were provided by ICAP at our request. The aim of this paper is to describe the development of an expert system that takes into account financial data that affect the function of eight Wood Processing Companies (W.P.C) in order to evaluate and rank them. The expert system uses multicriteria analysis and fuzzy logic in order to carry out the evaluation and the ranking of the eight W.P.C. The project was designed to evaluate the W.P.C. for a period of ten years, from 1991 to 2000. The validation and the ranking of the eight W.P.C. is carried out in the following way: ♦ First the financial data are processed and eight annual indexes of pure numbers are created. ♦ Each index is assigned a weight (equal or uneven weights might be used). ♦ The Expert System uses multicreteria analysis in order to output the net annual flow for each one of the eight W.P.C. The net flow is the difference between the outgoing flow and the incoming flow. ♦ The Expert system ranks annually the eight W.P.C. according to the value of their net flow. ♦ Eight Fuzzy Expected intervals are produced by the Expert System. They are the intervals in which the values of the net flows of the eight W.P.C. are expected to be included in the following year 2001. L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 189 2. THE MULTICRITERIA ANALYSIS METHODOLOGY The PROMETHEE II methods are part of the theory of relevance superiority [2]. They use six types of general tests with the corresponding test's functions in order to determine the superiority between two alternative solutions. In this case the aim is the determination of the superiority of one W.P.C. over a W.P.C. . The type of general level test criterion was selected to be used in this project, with the corresponding criterion function, because it has an indifferent region, for the determination of the superiority [3]. This type of general criterion is the most appropriate to be used in this case, due to the fact that it does not apply a strict choice. Only pairs of W.P.C. are tested in the form iX ) jX ( , , ,..., ,= 1 2 8i jv v i in order to determine which one or has the superiority according to the financial indexes. The function iv jv ( )H d is used to express the superiority: Equation 2.1. Level criterion function that uses preference functions. The value of variable d is the difference between the financial indexes of each pair of W.P.C. for the criterion under evaluation. ( , )i jv v (2.1.) ( , ), ( ) ( , ), ≥ =  ≥ superiority of W.P.C. if 0 superiority of W.P.C. if 0 i j i j i j P v v v d H d P v v v d Where ( , ), ( , )i j j iP v v P v v are the functions of preference. Equation 2.2. The level criterion function. It should be mentioned that p and are parameters that usually have a fixed value. q | | ( ) / | | | | ≤  = <  ≤ 0 if 1 2 if 1 if d q ≤H d q P d d p (2.2.) When it is examined which of two the W.P.C. is the superior, the superiority function ( , )i jv v ( )H d q is applied according to the price of (positive or negative) for each criterion. The and d p parameters are partly estimated in this project and they do not have a fixed value. The estimation of p and q is performed in the following way. ♦ First of all the annual performances of the eight W.P.C. is calculated for each criterion. ♦ If there exists a W.P.C. with a very high value of performance that is clearly much higher than the performance of the other seven W.P.C. it is excluded for the criterion under testing. This is done in order to avoid problems that might be caused in the calculation of p and q . L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 190 ♦ Afterwards, all of the differences d are calculated, for each pair of W.P.C. (under examination) for each criterion. If the preference function takes into account | | (the absolute value of ) only the positive values of d are considered. d d ♦ Afterwards the range between the maximum and the minimum values of is calculated using equation 2.3. E d Equation 2.3. Calculation of the range max min= −E d d (2.3.) ♦ Finally are estimated using the following equations 2.4. and 2.5. ,q p Equation 2.4. Calculation of p min λ= + ∗q d E (2.4.) Equation 2.5. Calculation of q min µ= + ∗p d E (2.5.) The coefficients λ and µ are considered to be threshold values that will be used for the estimation of p and q respectively. The parameters λ and µ can be assigned specific values, depending on the type of the problem and on the degree of sensitivity of the superiority control. In this case λ �has been assigned the value of 0.2 and µ the value of 0.4. In this way the q, p were calculated for each criterion and for each year [10]. The multicriteria indicator of preference ( , )Π i jv v which is a weighted mean, of the preference functions Π with weights defined by the researcher, expresses the superiority of the U.R.C against U.R.C. v after all the criteria are tested. The values of are calculated using the following equation 2.6 [4]. ( , )i jv v iv j Π Equation 2.6. Calculation of the multicriteria indicator ( , ) ( , ) = = ∗ Π = ∑ ∑ 1 1 k t t t i j k t t w Pt v v v v w j (2.6.) It should be mentioned that is defined to be the number of criteria and k ( )= 8k ( , )t i jP v v the preference functions for the criterions. The multicriteria preference indicator takes values between 0 and 1. When two W.P.C. are compared to each other each one is assigned two values of flows the outgoing flow and the incoming flow. k ( , )Π i jv v ( , )i jv v The outgoing flow is calculated that by the following equation 2.7 [1]. L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 191 Equation 2.7. Calculation of the outgoing flow ( ) ( , )ϕ + ∈ = Π∑ j i v A v vi jv i jv iv j (2.7.) In both cases A is defined to be the number of the alternative solutions W.P.C. . (Which in this case are seven). The outgoing flow expresses the total superiority of the W.P.C. against all the other W.P.C. for all the criterions. The incoming flow is determined by the following equation 2.8 [1]. jv iv jv Equation 2.8. Calculation of the incoming flow ( ) ( , )ϕ − ∈ = Π∑ j i v A v v (2.8.) The incoming flow expresses the total superiority of all the other W.P.C. against W.P.C. v for the criteria. The net flow for each W.P.C. is estimated by the following formula: . i iv ( ) ( ) ( )ϕ ϕ ϕ+ −= −i iv v The net flow is the number that is used for the comparison between the W.P.C. in order to obtain the final ranking. Each W.P.C. that has a higher net flow is considered to be superior in the final ranking. The superiority of W.P.C. over the W.P.C. viv j can be expressed using the following expression: j jV Pv ( v is superior to ) or , when i jv →iv v ( ) ( )ϕ ϕ>i jv v When ( ) ( )ϕ ϕ=iv v j jφthe superiority relation is written as follows: (This means that the relation between is neutral). Iiv v ,i jv v 3. DESCRIPTION OF THE INFERENCE ENGINE The expert system was designed to be rule-based and it consists of facts, rules and object-frames. It was designed and constructed to have a main rule set and local rule sets within the object frames [7]. The most important part of an Expert System is the Inference Engine, which is the mechanism that leads to the goal. The Inference engine strategy that was applied was backward-chaining with opportunistic forward, which means that it was designed to be a goal driven expert system, to use Forward Chaining only for the phase of Data Gathering in order to make it faster. It starts from the goal and it evaluates only the necessary rules in order to reach the final conclusion [11]. Knowledge about real world objects is stored in the object frames that contain various types of slots. Each slot describes the properties and the characteristics of the associated object [7]. L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 192 4. INPUT DATA The data that were used as input to the expert system come from balance sheets of the W.P.C. for the period 1991 - 2000. According to these balance sheets, the financial indexes were calculated. These indexes express the efficiency and the performance of the management of the W.P.C. These indexes were used (in past research projects) for the evaluation of investments, using multicriteria analysis [6]. The weights of the financial indexes that were used in the analysis are the following: . ( ,...,= =0 125 1 8iw i ) with = =∑ 7 1 1i i w Table 4.1: Financial indexes used for the determination of the initial input data =1x Reserves*360/Sales =2x Receivable*360/Sales =3x Gross Profit/Sales =4x Profit before taxes/Equity capital =5x Sales/Total Assets =6x Current liabilities*360/Cost of Sales 5. RESULTS OF THE ANALYSIS Initially the expert system performed the calculation of the annual net flows of the eight most important Greek W.P.C. from 1991 to 2000. The calculation of the net flows was performed according to the financing indexes that were mentioned in table 1. Afterwards, all of the W.P.C. have been ranked in proportion to their annual net flows and for the entire period of 1991 -2000. These rankings can be seen clearly in the following tables 5.1. and 5.2. Table 5.1: Annual evaluations of the eight W.P.C. according to their net flows from 1991 to 2000. 91 92 93 94 95 96 97 98 99 00 ABX 1.559 -0.16 -0.48 -1.47 -0.80 -2.80 -1.79 -1.46 0.19 -0.166 AKRITAS 1.49 -0.16 3.506 0.818 0.478 2.138 3.166 3.838 2.16 1.162 DRITSA 0.978 1.162 -0.85 -2.51 3.49 1.158 -0.84 0.470 1.80 -0.166 KARAMPELA -1.40 -0.16 -0.17 -0.83 -3.15 -4.16 -3.15 -3.15 -3.82 -0.166 KOYNDOYRI -1.60 -0.16 -5.16 0.146 0.47 2.462 1.798 2.470 1.31 -0.166 MOYRIKIS 0.109 -0.16 -0.13 -1.15 -0.47 -0.81 -1.15 -1.17 -0.65 -0.166 SELMAN 0.324 -0.16 1.16 1.854 0.51 3.182 0.518 -1.16 2.01 -0.166 XYLEMBORIKI -1.48 -0.16 2.146 3.158 -0.51 -1.15 1.474 0.178 -3 -0.166 The average net flows from 1991 to 2000 of all the eight W.P.C. that were used in the project are selected and presented in table 5.2. L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 193 Table 5.2: The average net flows from 1991 to 2000 for the W.P.C. and their ranking W.P.C. Average value 1991-2000 Ranking according to the average value ABEX 7.411/8 = 0.92 3 AKRITAS 18.59/8 = 2.32 1 DRITSA 4.674/8 = 0.584 4 KARAMPELA −20.193/8 = −2.524 8 KOYNTOYRI 1.555/8 = 0.194 5 MOYRIKIS −5.789/8 = −0.723 7 SELMAN 8.232/8 = 1.029 2 XYLEMPORIKI 0.468/8 = 0.0585 6 The ranking of each W.P.C. and the average ranking for each one, for the total period 1991-2000 is shown in table 5.3. In this table it is clearly shown that Akritas has been characterized 4 out of 10 times as the first company. The position of Akritas Company has become very strong after 1997 and it is obvious that it is very strong up to now. Dritsa company was first twice, but the last years after 1996 its position has dropped significantly. There are four W.P.C. that were first in the past years, but recently they are not so strong. Table 5.3: Annual position for each one of the eight W.P.C. in the rankings of the period 1991-2000 and the average position of each U.R.C. in the same rankings. Rankings of the companies of wood W.P.C 91 92 93 94 95 96 97 98 99 00 Average ABEX 1 2 6 7 7 7 7 2 5 2 3 AKRITAS 2 2 1 3 3 3 1 3 1 1 1 DRITSA 3 1 7 8 1 4 5 4 3 2 4 KARAMPELA 6 2 5 5 8 8 8 5 8 2 8 KOYNTOYRI 8 2 8 4 4 2 2 1 4 2 5 MOYRIKIS 5 2 4 6 5 5 6 6 6 2 7 SELMAN 4 2 3 2 2 1 4 8 2 2 2 XYLEMBORIKI 7 2 2 1 6 6 3 7 7 2 6 6. THE CONCEPT AND THE USE OF FUZZY EXPECTED INTERVALS 6.1. General One of the main features of the expert system is the calculation of the Fuzzy Expected Interval (F.E.I) for each one of the Wood companies of Greece. This means that it can produce a narrow characteristic interval of values. The flow of the company is expected to fall into this interval for the following years. L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 194 For example the F.E.I. could be (1.200, 1.480). This would mean that total flow for the company would fall between 1.200 and 1.480 in most of the cases. In this way the F.E.I can be used to forecast the future flow of each W.P.C. of Greece. Thus, a classification of all W.P.C. of the country, according to their expected flow, can be achieved. It is important that the system manages to produce an interval that is as narrow as possible. The central idea is that statistically and practically there is no interest in forecasting the exact number of the future flow, but rather in finding the general tendency and its direction. The main point is to know if the flow will increase from 1.200 to 1.900, or if it will drop to 0.600 and not to estimate the precise number concerning the past flows of the W.P.C. [14] This means that data can be grouped in an imprecise way (using various keywords) and thus Fuzzy Logic can be applied [12]. For example if the past data of net flows are 0.980, 1.010, 1.090 and 9.99 for four years, they can be grouped in the following way: On four occasions the net flow was almost 1.000. In this way the data can be grouped imprecisely. There are four types of sentences that can be used during classification of the data. Keywords Lower Bound Upper Bound 1st type almost %− 20x − 1x 2nd type more or less %− 20x %+ 20x 3rd type over + 1x %+ 20x 4rth type much more than 2x +∞ In a hypothetical situation using this approach, the net flows can be classified imprecisely into groups in the following way. 5 times the flow was almost 0.600 8 times the flow was more or less 0.850 3 times the flow was over 1.100 2 times the flow was much more than 1.500 This is very flexible way of classifying existing data. Fuzzy logic was introduced by Zadeh in 1965. All the theorems that are used in the following section were described by Kandel and Byatt [8]. 6.2. Functions used in the first two steps of the calculation of the F.E.I. After the classification, the first two steps that should be followed according to Kandel [9] are: A. The first step is to input data from the imprecise classification, into the characteristic function and find all C s [9]. ( )C X ' L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 195 The characteristic function is described by the following ( )C X Equation 6.2.1. ( ) ≤  = ≤   0 IF IF 100 100 1 otherwise X X C X X 0 (6.2.1.) where the number 100 is used as the maximum number of flow that was ever calculated according to the data existing so far. (It is the most extreme case according to the designers' judgment). This function is used for the forecast of the total flow. B. The second step, �9� is to find all 'µ s, which are the candidate Fuzzy Expected Intervals. The 'µ s are intervals of the form [LB, UB] and they can be calculated from the following equations 6.2.2. and 6.2.3. Equation 6.2.2. This equation is used to find the upper bound of every interval µi . max( , ) max( , ) min( , ) = − = = = + ∑ ∑ ∑ 1 2 1 1 2 1 2 1 n i j jnj i j i pi pi U B pi pi pi pi (6.2.2.) Where 1pi is the lowest bound of group and i 2pi is the upper bound of group . i Equation 6.2.3. This equation is used to find the lower bound of every interval µi . min( , ) min( , ) max( , ) = − = = = + ∑ ∑ ∑ 1 2 1 1 2 1 2 1 n i j jnj i j i pi pi L B pi pi pi pi (6.2.3.) Where 1pi is the lowest bound of group and i 2pi is the upper bound of group . i 6.3. Fuzzy set theorems applied in the third and fourth steps C. The third task is to find the minimum interval of each line using Theorems 6.3.1., 6.3.2. and 6.3.3. according to Kandel [9]. Theorems 6.3.1. to 6.3.6. are used to compare pairs of intervals of values and to determine which interval is larger and which is smaller. Theorem 6.3.1. is the following: max( , ) > >  =   1 1 if if m n R r s S R S s r (3.1) Where and { ,..., } { ,..., }= =1 1n mS S S R r r ∩ = ∅R S L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 196 Theorem 6.3.2. is the following: max( , ) > =  > if if m n m nR r sS R S s r (6.3.2.) Where { ,..., } { ,..., } , , ,ϕ= = ≠ ∉1 1m n ∉R r r S S S R S S R R S . Theorem 6.3.3. is the following: If { ,..., } { ,..., }= =1 1m nR r r S S S and ⊆R S then min( , ) [ ,..., ]= 1 mS R S r (6.3.3.) D. The final task is to find the maximum interval over the minima using the Thorems 6.3.4., 6.3.5., and 6.3.6. according to Kandel [9]. Theorem 6.3.4. is the following: If { ,..., } { ,..., }= =1 1m nR r r S S S and ∩ = ∅R S (6.3.4.) Then ma if r and x( , ) =S R R >1 nS max( , ) =S R S if >1 mS r Theorem 6.3.5. is the following: If { ,..., } { ,..., }= =1 1m nR r r S S S and , ,ϕ≠ ∉ ∉R S S R R S (6.3.5.) Then max( , ) =S R R if and >mr Sn max( , ) =S R S if >n mS r Theorem 6.3.6. is the following: If { ,..., } { ,..., }= =1 1m nR r r S S S and ⊆R S (6.3.6.) Then max( , ) [ ,..., ]= 1 nS R r S The maximum interval found is the Preliminary Fuzzy Expected Interval. The maximum number of flow (which in this case is 100) should be multiplied to the bounds of the Preliminary Fuzzy Expected in order to produce the real fuzzy expected interval. This interval could indicate the expected situation for the specific W.P.C. It is obvious that the narrower this interval is, the more useful it is. To achieve a narrower interval, for example, [1.500-1.700] for the net flow of the following year, the classification of the groups of frequencies should be successful. 7. DISCUSSION OF THE W.P.C.'S EXPECTED INTERVALS OF VALUES Actually the testing was done for the eight W.P.C. of Greece. The initial knowledge base of the system included financial data for the eight W.P.C. from 1991 to 2000. It is estimated that the values of the net flows of these W.P.C. will fall inside these intervals for the following year 2001. L. Hliadis, T. Koutroumanidis, G. Arabatzis, C. Arapatsakos / An Expert System 197 Table7.1: F.E.I.'s for the eight W.P.C. for 2001. W.P.C. F.E.I. for 2001 ABEX (0.3, 0.32) AKRITAS (2.09, 2.18) DRITSA (0.66, 1.15) KARAMPELA (0, 0.01) KOYNTOYRI (1.4, 1.4) MOYRIKIS (0.01, 0.02) SELMAN (0.53, 0.62) XYLEMPORIKI (1.4, 1.4) According to the F.E.I. that were produced by the computer expert system, Akritas is going to be first for 2001, with a significant difference from Koundouri and Xylemboriki that are classified as second. Selman is going to be in the third position and Abex in the fourth position. The expert system will be used and tested again with future data. This means that the task of the evaluation of the W.P.C. will continue and the system's credibility will also be evaluated. REFERENCES [1] Baourakis, G., Kalogeras, N., Doumpas, M., Stavropoulou, Ch., Bankova, M., and Zopunidis, K., "Multicriteria methodology for the development of the financial performance of the enterprises in the agricultural sector: The case of cooperative and juice enterprises. Resolution of financial decisions with multicriteria", Anikoulas Publications, Thessaloniki, 2001, 317-333. (in Greek). [2] Brans, J.P., L' ing›nierie de la d›cision, l'›laboration d'instrument d'aide a la d›cision, Colloque d'aide a la d›cision", Universit› Laval, Qu›bec, 1982. [3] Brans, J.P., and Vincke, Ph., "A preference ranking organization method: The PROMETHEE method for multiple criteria decision making", Management, 31(6) (1985) 647-656. [4] Brans, J.P, Vinke, Ph., and Mareschal, B., "How to select and how to rank projects: The PROMETHEE method", European Journal of Operational Research, 24 (1986) 228-238. [5] Epilogi Magazine 2001, Swing of 2001 Annual financial Review. [6] Evrard, Y., and Zisswiller, R. "Une analyse des d›cisions d'investissement fend›e sur les modšles de choix multi-attributs", Finance, 3(1) (1982) 51-68. [7] Jackson, M., Understanding Expert Systems Using Crystal, John Willey & Sons, 1992. [8] Kandel, A., and Byatt, W.J., "Fyzzy sets, fuzzy algebra, and fuzzy statistics", Proc. IEEE, 66(12) (1978) 1619. [9] Kandel, A., Fuzzy Expert Systems", CRC Press, 1992. [10] Koutroumanidis, T., Papathanasiou, J., and Manos, V., "Multicriteria analysis of primarysector's effectiveness in the region of Eastern Macedonia and Thrace", Fourteenth Hellenic Operational Research Society. (H.O.R.S), Duth Xanthi, November 2001. (in Greek) [11] Leonardo "User guide", By Bezant Limited, 1992. [12] Partridge, D., and Hussain, K., Knowledge Based Information Systems, McGraw Hill, 1995. [13] Peteinarakis, J., "Integrated survey for the industry of wood in Greece", LE/2, 1992. [14] Zadeh, L.A., "Fuzzy sets", Inf. Control, 8 (1965) 338. 
work_6ovnzpg4crgrvb5mq3es6kfyc4	----	An Expert System to Assess Fire Safety in Dwellings H. A. DONEGAN, I. R. TAYLOR and R. T. MEEHAN 1nstitute of 1nformatics University of Ulster Shore Road, Newtownabbey BT37 OOB, UK ABSTRACT This paper describes the application of an expert system to the evaluation of fire safety in dwellings based on the body of knowledge developed by the Fire Engineering Research Group at the University of Ulster. The background and philosophy of the evaluation procedure together with the associated reasoning with respect to the choice of system and its implementation are outlined in some detail. This demonstration syat.em is intended as a pilot study for a more ambitious programme. A discussion relating to problems with the system and future developments concludes the paper. KEYWORDS: Fire safety evaluation, expert system, Delphi, points scheme. INTRODUCTION In recent years the notion of fire safety evaluation with respect to buildings has become inextricably bound up with the perception of the prioritisation of those entities which taken together comprise the fire safety components of a specific building type. The prioritisations, often but not exclusively the result of expert opinion, are used in the development of points schemes. These are essentially research into practice devices which facilitate the economical allocation of scarce resources in the design and refurbishment of buildings. The philosophy is clearly extendible to any form of shelter, e.g. ships, planes and trains, where life safety and property protection are paramount. This paper will focus on dwelling fire safety with specific reference to the determination of a fire safety quality measure which can be optimised interactively using the expert system shell Xi Plus in a PC environment. The level of optimisation achieved is a function of the user's requirements and resources given the existence of acceptable norms. The chronology of events leading to this work began with the study by Nelson and Shibe [ll who produced a system for the Fire Safety Evaluation FIRE SAFETY SCIENCE-PROCEEDINGS OF THE THIRD INTERNATIONAL SYMPOSIUM, pp. 485-494 485 Copyright © International Association for Fire Safety Science of health care facilities in the USA. Following these developments, the Department of Health and Social Services (DHSS) sponsored the Fire Engineering Department at the University of Edinburgh to produce a fire safety evaluation points scheme for patient areas within hospitals. The work was conducted by Marchant [2] and completed in 1982. Stollard [3] reported on the development of the points scheme at Edinburgh and outlined the procedures which were used. In 1983 a pilot study [4] based on the methodology of Marchant was undertaken to consider the application of the DHSS scheme to dwellings. This led to a programme of work in 1984 at the University of Ulster, funded by the Science and Engineering Research Council. The theoretical consequences [5] and practical considerations [6] which followed provide the immediate environment for the present expert system. The motivation for the expert system derives from the complexity of the total fire safety scheme and the need to create a 'what if' environment for the user. The ideas in [2] and [4] are directed towards the production of a weighted ranking, referred to as a'priority vector, of fire safety components. Typical components might be management, occupants and visitors, contents, fire brigade, detection systems, fire fighting equipment and so on. The weighted rankings for n such components (w 1,w2,w 3 ' " .w n) emerge from the consensus of expert opinion generally obtained through an application of the Delphi process [7]. Notwithstanding certain shortcomings of the latter [8] and assuming that a priority vector can be arrived at, the fundamental and as yet unanswered question is not with regard to its existence, but with regard to its significance in the field of application. More particularly its optimal effectiveness - can it be used to minimize safety maintenance costs or to allocate scare resources in the pursuit of maximum safety? The authors would contend that such issues .i.n the environment of fire safety analysis are the driving force behind research into practice. On the premise that total safety while being a desirable objective is something which can only be approximated within technological development periods, the approximation is presumably the stability of professional expert opinion within each time period. Fire safety history provides a wealth of examples of connected time periods, each new period arising as a result of at least one catastrophe at the conclusion of or within a previous time period. The overall effect has been the production of more and more prescriptive legislation whi'ch inhibits an engineering approach to the solution of many problems. At the present time, within the constraints of existing legislation, consideration is being given to engineering strategies which will facilitate a degree of trade-off between passive and active fire safety measures. This is where the application of an expert system capable of logical repeatability has a distinct advantage in that it provides a basis for uniformity in the decision making process. It is clear from Figure 1 that an expert system which provides the user with a 'what if' environment fits comfortably as part of a generalisation of the simple fire safety framework proposed in [5J. Prior to the introduction of the expert system the procedure in using a points scheme involved taking the scalar product of the priority vector v: (w 1,w 2,w 3 ' " .w n) with each of two corresponding sets of component 486 (DELPHI INFORMATION I ANALYTICAL MODEL r--------------- ----------------, I I I I I I I I I I E I NORM SURVEY : ; I VECTOR VECTOR I E I I R I ~ ~ IT I ~ II I I I S I I Y I NORM CHARACTERISTIC I S : SCORE Q SCORE S i ~ : i M I I I IE I IN I I v I I I I I R I 1 0 I I N I 1 M : I WHAT IF SURVEY VECTOR VC CHANGES I i ~ 1 ~ : T I I I OUTPUT I I I I ANALYSIS I I I I IL J 1'1C.i GENERALISED FRAMEWORK FOR FIRE SAFETY EVALUATiON 487 points score allocations - a norm set v N (Q1,Q2,q3'" .Qo) for a specific c dwelling category and a survey set V (sl,s2,s3'" .so) for an actual dwelling in the category (4). The corresponding products, the norm score Q and characteristic score S were then compared. The advantage of the above generalisation became apparent when it was realised that it was quite feasible for a dwelling to satisfy the overall criterion, viz: S > Q and yet fail in one or more subsets of the components. Given that there are 2 0 - 1 mathematically possible subsets for the n components this can lead to a large number of possibilities for large n . However in this case with n = 11, (6), it was decided to cluster the components from a fire engineering point of view as follows: A: Human measures B: Building specific or passive measures and C: Supportive or active measures. Table 1 exemplifies the categorisation used in this study. TABLE 1. Fire Safety Component Clustering A:Human B:Passive C:Active Occupants and Internal Design Fire Brigade Visitors Survey Volume Detection Systems Contents Means of Escape First Aid Fire Management Hazard Protection Fighting Equipment External Envelope An important feature of the conceptualisation was to distinguish the notions of passive/active safety from passive/active measures. Literature on safety (9J defines active safety as accident prevention and passive safety as accident protection, with the proviso that every critical event is a series of casual events within which it is possible to practice accident prevention (active safety). Once the critical event occurs it is only possible to practice accident protection (passive safety). Given the situation of a critical event (an unwanted fire), fire engineering literature refers to active and passive measures designed to combat the event. It is clear that trade-off between active and passive measures is possible but there is no potential for trade-off between active and passive safety. Figure 2 shows the fundamental relationship between these variables. From this diagram it is possible to see that total fire safety will only arise as a result of complete active safety or total prevention whereas within a technological time period passive safety is the limiting state of prevailing expert opinion. The implicit assumption is that as technological time periods progress there should be a statistically observable decrease in fatalities and in property destruction. The efficiency of any system can only be measured in the future. Meanwhile the present approach is a contribution to the organised thinking which is essential for any level of improved safety. The next sections will describe the choice of application and its implementation. 488 HUMAN MEASURES FIGURE 2 Passive/Active Safety and Passive/Active Measures CHOICE OF SYSTEM Bearing in mind the above considerations the resultant expert system must be capable of: 1 - deciding overall if a dwelling is adequate for fire safety, 2 - allowing some trade off between active and passive measures, 3 - suggesting acceptable requirements on each of the measures A,B and C and of 4 - allowing changes in the dwelling specification to test 'what .if' questions. It was decided to investigate if an' expert system could meet the above requirements with the possibility that such a system could also answer "why" questions, justifying its conclusions. Such a computer system should also be consistent in its answers and allow a variety of people in different locations to access its expertise. By building in help information at various levels, the system can be employed by more users, allowing them to learn from the inbuilt expertise. To aid such use, information about the dwelling under consideration must be input, with the system asking appropriate questions and using default values if no information is available. To test its feasibility the authors have applied the expert system to the study of the previously mentioned points system for evaluating fire safety in dwellings, [4,6,10]. The eleven components of fire safety, as identified by the Delphi technique, are clustered into the A, Band C categories described above. The expertise of the system includes identification of the eleven components, their relative weighting as given by the vector V' and the norm score Q for which a particular kind of dwelling would be deemed to satisfy fire safety requirements. An expert system requires a knowledge base within which such expertise is represented, an inference engine to process the knowledge and an interface to users and developers. Within an expert system shell the latter two facilities are already provided allowing the user to concentrate on the development of the knowledge base. Since the shell is 'domain-free' and so contains no information on fire safety, the knowledge base must contain all 489 the expertise of the system, which as well as rules and facts includes appropriate questions and help for the user. For thig investigation the PC based shell xi Plus was chosen, in which the knowledge is represented in a rule-based form. The shell allows links to, for example, databases, spreadsheets and C programs and this was utilised for some of the data. The shell provides a comprehensive set of tools for developing and testing the knowledge base, within a user friendly environment. While it does not allow for probabilities within the data this was not required in the present project. Relatively little work on the application of expert systems within fire safety has been pub Li s he d , with most of it on the compliance of buildings with fire regulations. The best known example is the commercial program BRIGADE [11], a system with 4000 rules based on the shell LevelS, while there are discussions of ongoing work in [12]. IMPLEMENTATION In developing an expert system the main problems are obtaining the expertise and then deciding how to represent it within the knowledge base. In basing this expert system on the work of [4,6], the acquisition of knowledge using the Delphi technique has been performed and it remains to represent it within the shell. The components of Table 1 are structured as illustrated as in figure 3, with appropriate questions, on-line help and corresponding rules for each component. I FIRE SAFETY I I HUMAN I I PASSIVE I I ACTIVE IMEASURES MEASURES MEASURES I I I I Occupants Management Contents Fire Detection Fire and Brigade Systems Fighting Visitors Equipment I I I Int~rnall I Survey I [Means ofl IHazardl IExternal I Deslgn Volume Egress Protection Envelope FIGURE 3 External Knowledge Structure 490 In the system as described in [6], worksheets were used to assess each of the eleven components with a rating on a scale from 0-5. Each component set was then weighted by the priority vector V' to give an overall score, S in a scale 0-500. Expert opinion indicates that for the house type considered in the investigation an overall score of 375 or 75% would be sufficient for the dwelling to satisfy fire safety standards. This equates with norm scores for each measure of (A, 154) , (B, 154) and (C, 67), each being 75% of the corresponding maximum allocations of (A,205), (B,205) and (C,90). To allow for trade-off between the different measures the selected norm scores for A, Band C must be less than 75% of the overall maximum allocation. This is illustrated in the Results section below. Within the expert system, separate knowledge bases were constructed for the three measures A, Band C, aiding separate development and testing. The information required is obtained by prompting the user with various questions. The answers are tested by the inbuilt rules of the knowledge base and along with other information stored as data the program decides on an overall score for the dwelling. Unlike the manual system, the expert system can allow different levels of trade-off between A, Band C by setting separate norms for these measures. For example in obtaining a rating for "occupants and visitors" the information considered is the number of occupants compared with the number of bedspaces, the number of floors and the ratio of adults to dependents. A screen of the form of figure 4 is used to request information, with help available if required. Within this component there are rules such as IF THEN IF THEN number of occupants occupant score = 1.7 dwelling is two storey survey score = 1.3 bed spaces with a total of 17 rules for the first component. Under A for human measures there are 46 rules, in B on building specific measures there are 80 rules with 34 rules in C. To test a system with a total of 160 rules, as many as possible paths through the system must be tried. A set of.test cases representing standard and extreme situations was developed, to probe the system for potential limitations and weaknesses. RESULTS Applying the program to an actual dwelling, typical output might be A: Human Measures B: Building Specific Measures C: Supportive Measures OVERALL 491 actual score 170 165 50 385 selected norm score 141 135 58 375 pass/fail p p f P Application: Fire Safety Knowledgebase: Human Measures ,-----occupants and visi tors------------------------, How many occupants in dwelling? 3 How many bed spaces in dwelling? 2 single storey What size is dwelling? *two storey greater than two storey How many able adults in dwelling? 1 How many dependants in dwelling? 2 PRESS Fl FOR HELP Tab/Backtab next/previous field Esc cancel I I CTRL+Rtn end I I F3 why II FIGURE 4: Information requested on occupants and visitors, within human measures. To allow trade-off the selected norm scores in this example are (A, 141), (B,135) and (C, 58). The overall required score is of course 375 (75% of 500) . Trade-off is occuring be t vee n C, below standard, and A and B to give an overall pass for the dVlelling. In contrast for another dwelling the results may be actual selected score norm score pass/fail A: Human Measures 145 141 P B: Building Specific Measures 140 135 P C: Supportive Measures 70 58 P OVERALL 355 375 f so that the dwelling fails to reach the required standard even though it passes in each of A, Band C. If trade-off is allowed, by setting the norms for A, Band C beloVl the overall standard required, then inevitably some dwellings will pass each aspect but fail to reach the required overall standard, as in this example. By setting the required norm scores for A, B and C at 154,154 and 67 respectively, ie 75% of each maximum possible score of 205,205 and 90, no trade-off would be allowed between the measures. The philosophy of trade-off is discussed further in [13]. If a dwelling fails to reach the overall standard, the expert system allows a 'what if' type investigation to be performed to ascertain what improvements may be made to bring the house up to specification. with the addition of some financial information the, alternate cost of a range of possible alterations to the building may be explored using the expert system. The program can be easily adjusted to allow trade-off only between the passive and active measures, Band C. 492 CONCLUSIONS This paper shows that a points scheme for the fire safety evaluation of dwellings may be computerised using an expert system. The program illustrates the merits of using an expert system shell to encapsulate the knowledge, especially in comparison to using a conventional programming language. The program can be run at all stages of development with a built-in user friendliness. On any consultation, context sensitive help is available to the user at the level required, the user may ask the system why some information is required and explore 'what if' type situations. When loaded in a portable PC, the system could be employed on-site to assess the status of a dwelling and, if required, to suggest options for improving its level of fire safety to any required level. The program allows trade-off between different aspects of fire safety, but the amount of trade-off may be adjusted. The system could be developed to assess the fire safety of for example public assembly buildings as discussed in [14]. On the basis of this pilot study the authors are convinced that subject to appropriate funding, the expert system incorporating a points scheme could be extended to more complex buildings where the risk factors are likely to be much higher. REFERENCES 1. Nelson, H.E., and Shibe, A.J.,"A system of Fire Safety Evaluation of Health Care Facilities", Report NBSIR 78-1555-1, NBS, 1978. 2. Marchant, E.W., (ed), "Fire Safety Evaluation (points) Scheme for Patient Areas within Hospitals", A report on its origins and development sponsored by DHSS Department of Fire Safety Engineering, University of Edinburgh, 1982. 3. Stollard, P., "The Development' of a Points Scheme to Assess Fire Safety in Hospitals", Fire Safety Journal, 7, 145-153, 1984. 4. Shields, T.J. Silcock, G.W. and Bell Y., "Fi-;::e Safety Evaluation of Dwellings", Fire Safety Journal, 10 , 29-36, 1986. 5. Donegan, H.A., Shields T.J., and Si1cock G.W., "A Mathematical Strategy to Relate Fire Safety Evaluation and Fire Safety Policy Formulation for Buildings" in Fire Safety Science - Proceedings of the 2nd International Symposium, Tokyo, pp. 433-441, 1988. 6. Shields, T.J, Silcock G.W. and Donegan, H.A., "Assessing Fire Risk in Dwellings", University of Ulster, 1989. 7. Linstone, H.A. and Turoff, M., (eds), The Delphi Method, Techniques and Applications, Addison Wesley, 1975. 8. Shields, T.J., Silcock, G.W. and Donegan, H.A., "Methodological Problems Associated with the Use of the Delphi Technique", Fire Technology, 23, 175-186, 1987. 9. Wilson, G.A., "Techniques of Safety and Risk Management", Proc. One Day Seminar, Dept. of Mech. and Ind. Eng., University of Ulster and Plant Safety, 22, 2, 1990. 10. Shields, T.J., Silcock, G.W. and Donegan, H.A., "The Development of a Fire Safety Evaluation Points Scheme for Dwellings, Part I - Some Theoretical Considerations", Fire Safety Journal, 15, 313-324, 1989. 11. Peregrine Expert Systems Ltd, "Brigade", Dublin, 1988. 12. Hamilton G., Harrison A.P., Pascall J.R., Directory of Research and development of expert systems in the construction and building 493 services industries Vol III, BSRIA Ref 7177, 1983. 13. Shields T.J., Silcock G.W. and Donegan H.A., "A Philosophy for Trade- Off", Report for Department of Environment, NI, 1989. 14. Shields, T.J., Silcock G.W. and Donegan, H.A., " Towards the development of a fire safety systems evaluation for public assembly buildings", Construction Management and Economics, ~, 147-158, 1990. 494 
work_6serihyiu5cz7k6sfc7vnxvykm	----	International Journal of Recent Technology and Engineering (IJRTE) International Journal of Recent Technology and Engineering (IJRTE) ISSN: 2277-3878, Volume-8 Issue-4, November 2019 8870 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D9485118419/2019©BEIESP DOI:10.35940/ijrte.D9485.118419  Abstract: The number of accidents that occur in construction sites are increasing day by day. Due to that, the damages and fatality rate due to construction site failures are alarming. During construction, many accidents may occur in the site. It’s the responsibility of site supervisors and Engineers to control the accidents during construction. An attempt has been made to analyse the different causes for accidents in a construction site based upon the surveys conducted at different construction sites located in Rajiv Gandhi salai, Chennai. Based upon the survey analysis, a smart system has been developed to categorize the construction of safety site based upon safety conditions adopted. This research also brings into light many effects due to improper safety management in the construction site. The research paper concludes by providing certain recommendations and strategies to all construction sites for improving their safety performance, which thereby reduces the number of accidents. Keywords : accidents; Constructions; likert rating, safety management; smart system. I. INTRODUCTION Recent years, the nature of accidents caused in a construction industry is found to be more hazardous when compared to many Industries. The severity of accidents which is happening in construction sites is also increasing day by day. Due to the prevailing unsafe condition in Construction sites,it has become a very difficult task to get skilled labours this has resulted in shortage of getting skilled personals. In the recent years, need for safety awareness among construction industries is much realized. This is due to the high cost associated with work related to money spent towards labour injuries, compensation paid to the workers, premium paid for insurance , indirect costs asscoaited with site accidents ,etc., Status of construction safety at workplace is reviewed in this paper. This paper mainly focuses on creating a safe working environment for construction employees. The study was carried out by physically inspecting several construction sites located in Rajiv Gandhi Salai, collecting the data and feedback regarding number of workers, safety measures adopted, types of accidents occurred, severity of accidents,etc., from construction site workers using questionnaires., rectification is not possible. Revised Manuscript Received on November 15, 2019 * Correspondence Author Dr.P.S.Aravind Raj*, Department of Civil Engineering, Aarupadai Veedu Institute of Technology, Paiyanoor, India. Dr.S.P.Sangeetha, Department of Civil Engineering, Aarupadai Veedu Institute of Technology, Paiyanoor, India. Dr.R.Divahar, Department of Civil Engineering, Aarupadai Veedu Institute of Technology, Paiyanoor, India. II OBJECTIVES OF THE STUDY The paper aimed at: 1. Conducting survey in construction sites and to determine the presence of hazards and then establishing procedures to overcome or meet those hazardous situations. 2. To promote safe practices at construction sites and to protect workers from hazardous situations. 3. To create a knowledge-based decision-making model to ensure safety during construction. 4. To categorize the construction sites based on the model created and to propose safety measures to be adopted. III METHODOLOGY The entire project was categorized to three major phases as shown in Fig. 1. Phase I During the first phase, several construction sites were inspected to determine the presence of hazards or hazardous situations. A set questionnaire were listed to measure the health and safety measured adopted in site. Due to time constraints it was decided to choose 8 construction sites for the study. The information assessed through this survey was to determine whether the chosen construction sites are competent to undertake the work safely without any major accidents. Phase II Since many clients are not aware of construction, health and safety, accidents are very common in construction sites. In order to assist such clients, a decision supporting model that is knowledge assisted was created on the basis of the assessing health and safety competence and rating the sites based on safety Phase III Based upon the questionnaires and with reference to the Likert scaling for rating construction site safety is a much more simple system to rate and catergorize health and safety performance with respect to site conditions and standards. The taxonomy of construction sites based on Likert scale, which was considered in the survey is as given in Table 3.1.The rating of sites as per Likert scaling method is a bipolar rating method used for measuring positive or negative responses for a given statement. An Expert System to Rate Construction Site Based on Safety S. P. Sangeetha, R. Divahar, P. S. Aravind Raj , P. M. Aboobacker Minaz, Abdulla Nahavir Namzeer An Expert System to Rate Construction Site Based on Safety 8871 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D9485118419/2019©BEIESP DOI:10.35940/ijrte.D9485.118419 IV RESULTS AND DISCUSSION The site data collected from six residential building sites and two commercial building sites are compared. Site safety survey has been conducted for a period of two months. The construction sites reviewed in this study are located in Rajiv Gandhi salai , Chennai , India. The survey was conducted once in a week for the duration of two months in each site. Direct survey was conducted with labors; site Engineers, safety engineers, Fig. 1 Methodology adopted to analyse safety in construction site Table 1.Scaling indicator for safety and health competence assessment based on Likert S.NO. Taxonomy Description about taxonomy 1. Poor Safety Standard of site is not up to the standards 2. Fair Partial safety standards achieved 3. Acceptable Construction safety is not adequate but are in standard limits 4. Good Site meets construction safety standards and its compliance is substantial 5. Excellent Exceeds standards and given the best rating 2.8% 9.4% 4.19% 10.4% 2.3% 22.12 % 1.04% 0 5 10 15 20 25 Fatal Accident. Fracture. Loss of body parts. Skin infection, Eye irritation. Loss of Hearing and Blindness. Body Injury. Any Other Injury. Fatal Accident. Fracture. Loss of body parts. Skin infection, Eye irritation Loss of Hearing and Blindness. Body Injury. Any Other Injury. Fig. 2 Percentage of Accidents vs. Reason for Accident International Journal of Recent Technology and Engineering (IJRTE) ISSN: 2277-3878, Volume-8 Issue-4, November 2019 8872 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D9485118419/2019©BEIESP DOI:10.35940/ijrte.D9485.118419 supervisors and safety officers. Safety survey visited were randomly but once in each week. In each site safety items were checked during site visit and necessary data were collected. The site supervisors, site engineers, masons and labourers were interviewed from about eight sites and interviewed about 102 candidates including all categories. During the survey, details of labor accidents which happened in last 2 years were collected and accident rate in each site was calculated the data collected includes, number of accidents, Injury types and reasons for accidents. From the survey results the percentage of accidents occurred during the past years in construction site along with reason for accident in each category was calculated. From the above figure, it was found that the fatal accident was 2.8% ; Fracture 9.4%;Loss of body part 4.19%; Skin infection & Eye irritation 10.4%; Loss of hearing and blindness 2.3%; Body injury 22.12%; other injuries were 1.04%. It was estimated that the average number of accidents in construction site is 7.46%. Based on the survey reports a smart system was developed to rate the construction sites based on Likert scale. After login, one format will appears which include the site names and site measurement which can be filled up by the user or the site engineers as shown in Figure 2.Based on the safety precautions adopted in the site the construction site will be rated. Table 2 summarizes the findings from the broad study performed on the practices for safety in construction sites. From the surveys conducted we found that maximum numbers of construction sites located in the study region have implemented prescribed construction safety practices. From the site survey conducted in all construction sites we found that Site 1, Site 4, Site 7 and site 8 practices a safety policy for their workers. They also provide education and training for all workers at construction sites and thereby creating an awareness among them. In addition to this, auditing ledger for safety and site inspection is maintained properly. Safety meetings were also periodically for the laborers on and off the site explaining the safety aspects and its importances. V CONCLUSIONS From the survey studies conducted, maximum numbers of projects have implemented safety practices. From the site survey in each construction site it was found that Site 1, Site 4, Site 7 and site 8 that they adopt safety policies for their workers on site and also provides awareness and training on safety topics for all their laborers at the work place. Also, safety inspection on tools and workers were done with proper auditing procedures regularly. Addition to the above procedures regular meetings with site stakeholders were conducted to have discussion about the safety, its importance and methods of implementation. Based on the safety practices mentioned in the above table, construction sites were rated based on a rating scale of 1 to 5.With the help of the smart system developed all sites can be checked for safety and rated depending upon the safety measures adopted. Suggestions on safety measures to be followed were also distributed to the sites. Fig. 3 Safety analysis in smart system Sl. No. Safety Practices S 1 S 2 S 3 S 4 S 5 S 6 S 7 S 8 1. Fall protection systems         2. Allocation of budget for safety         3. Personal protective equipment         4. Safety policy  X   X    5. Safety promotions  X   X X   6. Education and training  X       7. Safety auditing  X    X   8. Safety meetings         9. Site safety organization  X X X X X   10. Other safety practices Safety Access available Nil Nil Safety Access available Nil Nil Safety Access available Safety Access available An Expert System to Rate Construction Site Based on Safety 8873 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D9485118419/2019©BEIESP DOI:10.35940/ijrte.D9485.118419 REFERENCES 1. Choudhry, R. M, & Fang, D, “Why Operatives Engage in Unsafe Work Behavior, Investigating Factors on Construction Sites”, Safety science, 46(4), 2008, 566-584. 2. Fang, D. P., Huang, X. Y and Hinze, J., “Benchmarking Studies on Construction Safety Management in China”, Journal of Construction Engineering and Management, 130(3), 2004, 424-432. 3. Hassanein, A. A., & Hanna, R. S, “Safety Performance in the Egyptian Construction Industry”, Journal of Construction Engineering and Management, 134(6), 2008, 451-455. 4. S. Kumar and V. K. Bansal, “Construction safety knowledge for practitioners in the construction industry”,Journal of Frontiers in Construction Engineering,” vol. 2, no. 2, 2013, 34–42. 5. Lehtinen, E., B. Wahlström, “Management of Safety through Performance Indicators for Operational Maintenance.” IAEA Specialist Meeting on Methodology for Nuclear Power Plant Performance and Statistical Analysis. Vienna, 1996, 35-39. 6. L. S. Pheng and S. C. Shiua, “The maintenance of construction Safety: riding on ISO 9000 quality management systems”. Journal of Quality in Maintenance Engineering”, vol. 6, no. 1, 2000,28–44. 7. S. R. Meena, P. M. Nemade, S. N. Pawar, and A. S. Baghele,“ Implementation of safety management through review of construction activities in M.S. building projects,” International Journal of Engineering Research and Technology, vol. 2, no. 5, pp.1656–1662, 2013. 8. Sawacha, E., Naoum, S., and Fong, D (1999), Factors Affecting Safety Performance on Construction Sites, International Journal of Project Management, 17(5), 309-315. 9. Sawacha, E., Naoum, S., and Fong, D, “Factors Affecting Safety Performance on Construction Sites”, International Journal of Project Management, 17(5), 1999, 309-315. 10. ZubairMemon, A. Khatri, K. L., &Memon, A. B, “Identifying the Critical Factors Affecting Safety Program Performance for Construction Projects within Pakistan Construction Industry, Mehran University”, Research Journal of Engineering & Technology, 32(2),2013, 269-276. AUTHORS PROFILE Dr.S.P.Sangeetha, Vice Principal (Academics),AVIT has completed her PG and Doctoral degree in Structural Engineering. She has rich experience in teaching. She has received the “Best faculty award “from DKIRF and “Best recognition award “from Rotaract club of Chennai. Dr.R.Divahar, Associate Professor and Head, Department of Civil Engineering. He has published several Journal papers and patents too. He has several years of experience in teaching. Dr.P.S.Aravind Raj is working as Associate Professor, in Departmnet of Civil Engineering. He have done projects in several engineering problems and developed solutions on Composite constructions, environmental aspects, transportation and site retated requirements. 
work_6sosbe2fdngyjcsfrbqjs3qvwy	----	Review of the main developments of the Analytic Hierarchy Process [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Review of the main developments in the Analytic Hierarchy Process Alessio Ishizaka and Ashraf Labib University of Portsmouth, Portsmouth Business School, Richmond Building, Portland Street, Portsmouth PO1 3DE, United Kingdom Alessio.Ishizaka@port.ac.uk Ashraf.Labib@port.ac.uk ABSTRACT. In this paper the authors review the developments of the Analytic Hierarchy Process (AHP) since its inception. The focus of this paper is a neutral review on the methodological developments rather than reporting its applications that have appeared since its introduction. In particular, we discuss problem modelling, pair-wise comparisons, judgement scales, derivation methods, consistency indices, incomplete matrix, synthesis of the weights, sensitivity analysis and group decisions. All have been important areas of research in AHP. Keywords: AHP, Multicriteria Decision Making, Review 1. Introduction The Analytic Hierarchy Process (AHP) is a multi-criteria decision making (MCDM) method. The oldest reference that we have found dates from 1972 (T. Saaty, 1972). Then, a paper in the Journal of Mathematical Psychology (T. Saaty, 1977) precisely described the method. The vast majority of the applications still use AHP as described in this first publication and are unaware of successive developments. This paper provides a sketch of the major directions in methodological developments (as opposed to a discussion of applications) and further research in this important field. AHP has been inspired by several previous discoveries. The use of pair-wise comparisons (called paired comparisons by psychologists), the essence of AHP, instead of direct allocation of weights has been used long time before by psychologists, e.g. (Thurstone, 1927; Yokoyama, 1921). The hierarchic formulation of the criteria, a major feature of AHP, was first proposed by Miller in his 1966 doctoral dissertation (J. Miller, 1966) and applied in (J. Miller, 1969) and (J. Miller, 1970). The 1-9 scale is based on psychological observations (Fechner 1860; Stevens, 1957). The number of items in each level is inspired by (G. A. Miller, 1956), who recommends seven plus or minus two items. Since its introduction, AHP has been widely used, for example in banks (Haghighi, Divandari, & Keimasi, 2010; Seçme, Bayrakdaroglu, & Kahraman, 2009), manufacturing systems (Iç & Yurdakul, 2009; T.-S. Li & Huang, 2009; Yang, Chuang, & Huang, 2009), operators evaluation (Sen & ÇInar, 2010), drugs selection (Vidal, Sahin, Martelli, Berhoune, & Bonan, 2010), site selection (Önüt, Efendigil, & Soner Kara, 2009), software evaluation (Cebeci, 2009; Chang, Wu, & Lin, 2009), evaluation of website performance (Liu & Chen, 2009), strategy selection (M. K. Chen & Wang, 2010; S. Li & Li, 2009; Limam Mansar, Reijers, & Ounnar, 2009; Wu, Lin, & Lin, 2009), supplier selection (Chamodrakas, Batis, & Martakos, 2010; A. W. Labib, 2011; H. S. Wang, Che, & Wu, 2010; T.-Y. Wang & Yang, mailto:Alessio.Ishizaka@port.ac.uk mailto:Ashraf.Labib@port.ac.uk [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 2009), selection of recycling technology (Y.-L. Hsu, Lee, & Kreng, 2010), firms competence evaluation (M. Amiri, Zandieh, Soltani, & Vahdani, 2009), weapon selection (Dagdeviren, Yavuz, & KilInç, 2009), underground mining method selection (Naghadehi, Mikaeil, & Ataei, 2009) and its sustainability evaluation (Su, Yu, & Zhang, 2010), software design (S. H. Hsu, Kao, & Wu, 2009), organisational performance evaluation (Tseng & Lee, 2009), staff recruitment (Celik, Kandakoglu, & Er, 2009; Khosla, Goonesekera, & Chu, 2009), construction method selection (Pan, 2009), warehouse selection (W Ho & Emrouznejad, 2009), technology evaluation (Lai & Tsai, 2009), route planning (Niaraki & Kim, 2009), project selection (M. P. Amiri, 2010), customer requirement rating (Y. Li, Tang, & Luo, 2010; C.-L. Lin, Chen, & Tzeng, 2010), energy selection (Kahraman & Kaya, 2010), university evaluation (Lee, 2010) and many others. Several papers have compiled the AHP success stories (EH Forman & Gass, 2001; Golden, Wasil, & Harker, 1989; William Ho, 2008; Kumar & Vaidya, 2006; Liberatore & Nydick, 2008; Omkarprasad & Sushil, 2006; T. Saaty & Forman, 1992; Shim, 1989; Sipahi & Timor, 2010; Vargas, 1990; Zahedi, 1986). However AHP has also received strong criticisms. Despite the predicted demise by some researchers, there has been a strong response leading to steady increase in its usage. 2. The AHP method AHP is a multi-criteria decision making (MCDM) method helping decision-maker facing a complex problem with multiple conflicting and subjective criteria (e.g. location or investment selection, projects ranking, etc). Several MCDM methods have been developed (e.g. ELECTRE, MacBeth, SMART, PROMETHEE, UTA,… see (Barthélemy, 2003; Valerie Belton & Stewart, 2002)) and all are based on four steps: problem modelling, weights valuation, weights aggregation and sensitivity analysis. In the next sections we will review these four steps used by AHP and its evolutions. 2.1 Problem modelling As with all decision-making processes, the facilitator will sit a long time with the decision- maker(s) to structure the problem. AHP has the advantage of permitting a hierarchical structure of the criteria (figure 1), which provides users with a better focus on specific criteria and sub-criteria when allocating the weights. This step is important, because a different structure may lead to a different final ranking. Several authors (Pöyhönen, Hamalainen, & Salo, 1997; Stillwell, von Winterfeldt, & John, 1987; Weber, Eisenführ, & von Winterfeldt, 1988) have observed that criteria with a large number of sub-criteria tend to receive more weight than when they are less detailed. Brugha (2004) has provided a complete guideline to structure a problem hierarchically. A book (T. Saaty & Forman, 1992) compiling hierarchies in different applications has been written. When setting up the AHP hierarchy with a large number of elements, the decision maker should attempt to arrange these elements in clusters so they do not differ in extreme ways (Ishizaka, 2004a, 2004b; T. Saaty, 1991). [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Figure 1: Example of a hierarchy (Akarte, Surendra, Ravi, & Rangaraj, 2001) 2.2 Pair-wise comparisons Psychologists argue that it is easier and more accurate to express one’s opinion on only two alternatives than simultaneously on all the alternatives. It also allows consistency cross checking between the different pair-wise comparisons (see section 2.5). AHP uses a ratio scale, which, contrary to methods using interval scales (Kainulainen, Leskinen, Korhonen, Haara, & Hujala, 2009), requires no units in the comparison. The judgement is a relative value or a quotient a / b of two quantities a and b having the same units (intensity, meters, utility, etc). The decision maker does not need to provide a numerical judgement; instead a relative verbal appreciation, more familiar in our daily lives, is sufficient. Comparisons are recorded in a positive reciprocal matrix (1). In special cases, such as in currencies exchanges, not reciprocal matrices can be used (Hovanov, Kolari, & Sokolov, 2008). A =              1...... ....../1... ...... 1 1 21 112 n ijji ij n a aa aa aa (1) where aij is the comparison between element i and j If the matrix is perfectly consistent, then the transitivity rule (2) holds for all comparisons: aij = aik · akj (2) For example, if team A beats team B two-zero and team B beats team C three-zero, then it is expected with the transitivity rule (2) that team A beats team C six-zero (3 · 2 = 6). However, this is seldom the case because our world is inconsistent by nature. As a minimal consistency is required to derive meaningful priorities, a test must be done (see section 2.5). Webber et al. (1996) state that the order in which the comparisons are entered in the matrix may affect the successive judgments. [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 2.3 Judgement scales One of AHP’s strengths is the possibility to evaluate quantitative as well as qualitative criteria and alternatives on the same preference scale. These can be numerical, verbal (table 1) or graphical. The use of verbal responses is intuitively appealing, user-friendly and more common in our everyday lives than numbers. It may also allow some ambiguity in non-trivial comparisons. This ambiguity in the English language has also been criticised (Donegan, Dodd, & McMaster, 1992). Due to its pair-wise comparisons AHP needs ratio scales. Barzilai (2005) claims that preferences cannot be represented with ratio scales, because in his opinion an absolute zero does not exists, as with temperature or electrical tension. Saaty (1994) states that ratio scales are the only possible measurement if we want to be able to aggregate measurement, as in a weighted sum. Dodd and Donegan (1995) have criticised the absence of a zero in the preference scale. To derive priorities, the verbal comparisons must be converted into numerical ones. In Saaty’s AHP the verbal statements are converted into integers from one to nine. Theoretically there is no reason to be restricted to these numbers and verbal gradation. Although the verbal gradation has been little investigated, several other numerical scales have been proposed (table 2). Harker and Vargas (1987) have evaluated a quadratic and a root square scale in only one simple example and argued in favour of Saaty’s 1 to 9 scale. However, one example seems not enough to conclude the superiority of the 1-9 linear scale. Lootsma (1989) argued that the geometric scale is preferable to the 1-9 linear scale. Salo and Hämäläinen (1997) point out that the integers from one to nine yield local weights, which are unevenly dispersed, so that there is lack of sensitivity when comparing elements, which are preferentially close to each other. Based on this observation, they propose a balanced scale where the local weights are evenly dispersed over the weight range [0.1, 0.9]. Earlier Ma and Zheng (1991) have calculated a scale where the inverse elements x of the scale 1/x are linear instead of the x in the Saaty scale. Donegan, Dodd and McMaster (1992) have proposed an asymptotic scale avoiding the boundary problem, e.g. if the decision-maker enters aij = 3 and ajk = 4, s/he is forced to an intransitive relation (2) because the upper limit of the scale is 9 and s/he cannot enter aik = 12. Ji and Jiang (2003) propose a mixture of verbal and geometric scale. The possibility to integrate negative values in the scale has been also explored (Millet & Schoner, 2005; T. Saaty & Ozdemir, 2003). Intensity of importance Definition 1 Equal importance 2 Weak 3 Moderate importance 4 Moderate plus 5 Strong importance 6 Strong plus 7 Very strong or demonstrated importance 8 Very, very strong 9 Extreme importance Table 1: The 1 to 9 fundamental scale [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Scale type Definition Parameters Linear (T. Saaty, 1977) c = a · x a > 0 ; x = {1, 2, …, 9} Power (Harker & Vargas, 1987) c = x a a > 1 ; x = {1, 2, …, 9} Geometric (Lootsma, 1989) c = a x-1 a > 1 ; x = {1, 2, …, 9} or x = {1, 1.5, …, 4} or other step Logarithmic (Ishizaka, Balkenborg, & Kaplan, 2010) c = log a(x+(a-1)) a > 1 ; x = {1, 2, …, 9} Root square (Harker & Vargas, 1987) c = a x a > 1 ; x = {1, 2, …, 9} Asymptotical (Dodd & Donegan, 1995) c =          14 )1(3 tanh 1 x x = {1, 2, …, 9} Inverse linear (Ma & Zheng, 1991) c = 9/(10-x) x = {1, 2, …, 9} Balanced (Salo & Hamalainen, 1997) c = w/(1-w) w = {0.5, 0.55, 0.6,…, 0.9} Table 2: Different scales for comparing two alternatives (for the comparison of A and B, c = 1 indicates A = B; c > 1 indicates A > B; when A < B, the reciprocal values 1/c are used) Among all the proposed scales, the linear scale with the integers one to nine and their reciprocals has been used by far the most often in applications. Saaty (1980; 1991) advocates it as the best scale to represent weight ratios. However, the cited examples deal with objective measurable alternatives such as the areas of figures, whereas AHP mainly treats decision processes as subjective issues. We understand the difficulty of verifying the effectiveness of scales through subjective issues. Salo and Hämäläinen (1997) demonstrate the superiority of the balanced scale when comparing two elements. The choice of the ―best‖ scale is a very heated debate. Some scientists argue that the choice depends on the person and the decision problem (Harker & Vargas, 1987; Pöyhönen, et al., 1997). 2.4 Priorities derivation The goal is to find a set of priorities p1,…,pn such that pi/pj match the comparisons aij in a consistent matrix and when slight inconsistencies are introduced, priorities should vary only slightly. Different methods have been developed to derive priorities. Psychologists using pair-wise matrices before Saaty used the mean of the row. This old method is based on three steps (see example 1): 1. Sum the elements of each column j:   n i ij a 1 ji, 2. Divide each value by its column sum:    n i ij ij ij a a a 1 ' ji, 3. Mean of row i: n a p n j ij i    1 ' [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Example 1: Consider the following comparison matrix: The method ―mean of row‖ derives the priorities as follow: 1. Add the elements of the columns: (1.75; 7; 3.5) 2. Normalize the columns: 3. Calculate the mean of the rows: (a= 0.57; b=0.14; c=0.29). The result can be verified simply: 0.57 ≈ 4 · 0.14, which is equivalent to the entered comparison a = 4 · b 0.57 ≈ 2 · 0.29, which is equivalent to a = 2 · c 0.14 ≈ 0.5 · 0.29, which is equivalent to b = 0.5 · c In the case of the introduction of small inconsistency, we can decently think that it induces only a small distortion. Based on this idea, Saaty (1977) uses the perturbation theory to justify the use of the principal eigenvector p as the desired priorities vector (3). He argues that slight variations in a consistent matrix imply slight variations of the eigenvector and the eigenvalue. A · p = λ · p (3) where A is the comparison matrix p is the priorities vector λ is the maximal eigenvalue Only two years later after the publication of the original AHP, Johnson et al. (1979) show a rank reversal problem for scale inversion with the eigenvalue method. The solution of the eigenequation (3) gives the right eigenvector p, which is not necessary the same as the left eigenvector p’, solution of p’ T · A = λ · p’ T  A T · p’ = λ · p’. The solution depends on the formulation of the problem. This right and left inconsistency (or asymmetry) arises only for inconsistent matrices with a dimension higher than three (T. Saaty & Vargas, 1984a). In order to avoid this problem, Crawford and Williams (1985) have adopted another approach in minimizing the multiplicative error (4): aij = j i p p eij (4) where aij is the comparison between object i and j pi is the priority of object i eij is the error a b c a 1 4 2 b 1/4 1 1/2 c 1/2 2 1 a b c a 0.57 0.57 0.57 b 0.14 0.14 0.14 c 0.29 0.29 0.29 [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 The multiplicative error is commonly accepted to be log normal distributed (similarly the additive error would be assumed to be normal distributed). The geometric mean (5) will minimize the sum of these errors (6). pi = n n j ij a 1 (5) min 2 11 )ln()ln(           n j j i ij n i p p a (6) The geometric mean (also sometimes known as Logarithmic Least Squares Method) can be easily calculated by hand (see example 2) and has been supported by a large segment of the AHP community (Aguarón & Moreno-Jiménez, 2000, 2003; Barzilai, 1997; Barzilai & Lootsma, 1997; Budescu, 1984; MT Escobar & Moreno-Jiménez, 2000; Fichtner, 1986; Leskinen & Kangas, 2005; Lootsma, 1993, 1996). Its main advantage is the absence of rank reversals due to the right and left inconsistency: in fact geometric mean of rows and columns provide the same ranking (which is not necessarily the case with the eigenvalue method). Example 2: The priorities derived with the geometric mean (5) from the matrix of the Example 1 are: p1 = 2241 3  , p2 = 5.0 2 1 1 4 1 3  , p3 = 112 2 1 3  Normalizing, we obtain: p  = (0.57; 0.14; 0.29) If mathematical evidences testify clearly for the geometric mean over the eigenvalue method, there is no clear differences between these two methods when simulations are applied (Budescu, Zwick, & Rapoport, 1986; Cho & Wedley, 2004; Golany & Kress, 1993; Herman & Koczkodaj, 1996; Ishizaka & Lusti, 2006; Jones & Mardle, 2004; Mikhailov & Singh, 1999), apart from special cases (Bajwa, Choo, & Wedley, 2008). Perhaps in the light of this lack of practical evidence, Saaty’s group has always supported the eigenvalue method (Harker & Vargas, 1987; T. Saaty, 2003; T. Saaty & Hu, 1998; T. Saaty & Vargas, 1984a, 1984b). Other methods have been proposed, each one based either on the idea of the distance minimisation (like the geometric mean) or on the idea that small perturbation inducing small errors (like the eigenvalue method or the arithmetic mean of rows). Cho and Wedley (2004) have enumerated 18 different methods, which are effectively 15 because three are equivalent to others (C. Lin, 2007). 2.5 Consistency As priorities make sense only if derived from consistent or near consistent matrices, a consistency check must be applied. Saaty (1977) has proposed a consistency index (CI), which is related to the eigenvalue method: [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 CI = 1 m ax   n n , (7) where n = dimension of the matrix λmax = maximal eigenvalue The consistency ratio, the ratio of CI and RI, is given by: CR = CI/RI, (8) where RI is the random index (the average CI of 500 randomly filled matrices) If CR is less than 10%, then the matrix can be considered as having an acceptable consistency. Saaty (1977) calculated the random indices shown in table 3: n 3 4 5 6 7 8 9 10 RI 0.58 0.9 1.12 1.24 1.32 1.41 1.45 1.49 Table 3: Random indices from (T. Saaty, 1977) Other researchers have run simulations with different numbers of matrices (Aguarón & Moreno-Jiménez, 2003; Alonso & Lamata, 2006; Lane & Verdini, 1989; Tummala & Wan, 1994) or incomplete matrices (E. Forman, 1990). Their random indices are different but close to Saaty’s. This consistency index has been criticised because it allows contradictory judgements in matrices (Bana e Costa & Vansnick, 2008; Kwiesielewicz & van Uden, 2004) or rejects reasonable matrices (Karapetrovic & Rosenbloom, 1999). Techniques based on the transitivity rule (2) have been developed in order to discover contradictory judgements and correct them (Ishizaka & Lusti, 2004; Y. Wang, Chin, & Luo, 2009). Several other methods have been proposed to measure consistency. Peláez and Lamata (2003) describe a method based on the determinant of the matrix. Crawford and Williams (1985) prefer to sum the difference between the ratio of the calculated priorities and the given comparisons in the Geometric Consistency Index (GCI): GCI=   )2)(1( loglog2 2    nn a ji p p ij j i (9) Aguarón and Moreno-Jiménez (2003) determines a threshold that provides an interpretation of the inconsistency threshold analogous to the CR=10%: GCI = 0.3147 for n = 3, GCI = 0.3526 for n = 4 and GCI = 0.370 for n > 4. The transitivity rule (2) has been used by Salo and Hamalainen (1997) and later by Ji and Jiang (2003) in another formulation:   2/)1( log 1 1 1       nn a n i n ij p p ij j i (10) [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Alonso and Lamata (2006) have computed a regression of the random indices and propose the formulation: λmax < n + 0.1(1.7699n-4.3513) (11) Stein and Mizzi (2007) use the normalised column of the comparison matrix. For all consistency checking, some questions remain: what is the cut-off rule to declare my matrix inconsistent? Should this rule depend on the size of the matrix? How should I adapt my consistency definition, when I use another judgement scale? 2.6 Incomplete pair-wise matrix The number of pair-wise comparisons requested can be very high: (n 2 -n)/2 for n alternatives/criteria. For example 8 alternatives and 6 criteria request 183 entries. This high number of questions can quickly become overwhelming and comparisons may be entered with a small reflexion time due in order to speed up the process. Therefore it has been proposed to enter fewer comparisons, which are well evaluated, than the whole number of comparisons, which may be approximate evaluation. Another reason for incomplete comparisons matrix is that the decision-maker may not have formed a strong opinion on a particular judgement and rather that forcing him to give an often wild guess or to have the entire process slowed down due to one comparison; one can simply skip this question. In a Monte-Carlo simulation study, where comparisons are deleted from large matrices (rank 10, 15 and 20), it has been discovered that one can randomly delete as much as 50% of the comparisons without significantly reducing the results (Carmone, Kara, & Zanakis, 1997). The minimal number of comparisons required is n-1, one for each row or column of the pariwise comparison matrix. The other comparisons are redundant and only necessary to check consistency and possibly improve accuracy. They can be calculated by the transitivity rule (2). This transitivity rule can be extended: aij = aik · akm · ... ·avj (12) The literature related to incomplete comparisons falls into three categories: calculation of missing comparisons, starting rules and stopping rules. 2.6.1 Calculation of missing comparisons A natural way to fill in the missing matrix element is to take the geometric average of all the indirectly calculated missing comparisons with the extended transitivity rule (12) (P. Harker, 1987). The drawback of this method is that the number of indirect comparisons grows with the number of alternatives n in such a way that the calculation requires a long processing time. In order to overcome this problem, eigenvector can be derived directly without estimating unknown comparisons (P. T. Harker, 1987b). If we consider the matrix A with missing comparisons, the method to calculate the priorities has two steps: i. A new matrix B is created from the incomplete matrix A: bij = aij if aij is a real number > 0 = 0 otherwise bii = the number of unanswered questions in the row i [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 ii. The eigenvalue method is applied on the matrix B = A + I, where I is the identity matrix. Other methods have been proposed to decrease the number of comparisons:  To use clusters and pivots (Ishizaka, 2008; Shen, Hoerl, & McConnell, 1992). Objects are divided into several clusters such that all these clusters have one common object: the pivot. Then, pair-wise comparisons are performed for each cluster and priorities are calculated. Finally, the global priority is derived by using the pivot and priorities of each cluster.  To make comparisons for a node that has an overall high impact on the final priorities and froze node with a very low global weight (Millet & Harker, 1990). 2.6.2 Starting rule As only n-1 comparisons are required, the question is which one should we estimate? Harker (1987a) used random selection. Ishizaka and Lusti (2004) prefers the first upper diagonal of the comparison matrix as all items are compared exactly same number of time: two. They prefer to reject a common row or column approach because they felt that it compromised the psychological independence of the comparisons. Wedley et al. (1993) investigated starting rules for the selection of the first n-1 comparisons in an experiment with 144 business students for a problem with known answer. The students have to estimate the proportions of five distinct colours in a rectangular area. Six different referents were considered for the n-1 initial comparisons: first column, bottom row, upper diagonal, lowest ranked item, median ranked item and highest ranked item. The lowest ranked item for the first n-1 comparisons proved to be statistically more accurate than any of the other starting methods. However instead of imposing a starting rule, it may be preferable to leave the choice to the decision- maker, who can select the comparisons (s)he is the most comfortable to evaluate. 2.6.3 Stopping rule Harker (1987b) uses a gradient procedure to select the next comparison which will have the greatest impact on the priorities. He suggested three different stopping criteria:  Subjective satisfaction of the user with the priorities.  Tolerance percentage change in absolute attribute weights from one question to another.  Ordinal rank will not be reversed whatever further comparison is entered. This stopping criteria is not effective if two alternatives have almost equivalent priorities. Wedley (1993) has simulated several matrices with different degrees of inconsistency and then used then to develop regressions equations that predict consistency ratio at each step beyond n comparisons. This method is satisfying only if the decision-maker does not enter a too inconsistent comparison. Alternatively, the consistency index can be calculated only based upon the entered comparisons (P. T. Harker, 1987b): the priority vector (pi, i = 1,2,...,n) is calculated (see section 2.6.1) which in turn is used to calculate estimates missing values (pi/pj, i = 1,2,...,n; j = 1,2,...,n). Since these estimates are based only on the known comparisons, the consistency index tends to underestimate the true degree of inconsistency that would occur if all redundant comparisons were entered. In order to correct this bias, new random indexes were calculated (E. H. Forman, 1990). This method is used in Expert Choice, the leading supporting software of AHP, to estimate the consistency ratio for incomplete matrices. [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Lin and Swenseth (1993) have proposed to stop the process when one alternative becomes dominant to such a degree that regardless of the effects of the remaining comparisons to evaluate, it cannot be overtaken as the preferred choice. In fact, in a scenario with n criteria, if the difference between the cumulative scores for the two best alternatives, after considering r criteria, is greater than the remaining total eigenscores of the n-r remaining criteria, the best alternative is identified. The drawback of this method is that only the best choice is identified and the decision-maker must use a top-down sequence for establishing priorities. 2.7 Aggregation The last step is to synthesize the local priorities across all criteria in order to determine the global priority. The historical AHP approach (called later distributive mode) adopts an additive aggregation with normalization of the sum of the local priorities to unity:   j ijji lwp (13) where pi: global priority of the alternative i lij: local priority wj: weight of the criterion j The distributive mode is subject to rank reversal, a phenomenon that has been extensively discussed in the literature. In particular, a memorable debate has appeared first in Omega and then in Management Science and in the Journal of the Operational Society. The saga in Omega has four episodes. The first article (Valerie Belton & Gear, 1983) came as a bombshell. It describes an example where the introduction of a copy of an alternative changes the ranking. The rank reversal is due to the modification of the relative values between the local priorities (which is a different and independent cause of the rank reversal due to the right and left inconsistency described in the section 2.4). As the sum of the local priorities to unity changes with the introduction of a new alternative, the local priorities are also modified when normalised and therefore the global priorities may be reversed. The rank reversal phenomenon is therefore independent of the consistency of the matrix and the derivation method of the priorities. As this phenomenon is not unique to AHP but to all additive models (Triantaphyllou, 2001; Y. Wang & Luo, 2009), Belton and Gear, inspired by the weighted sum model, suggested using a normalisation by dividing the score of each alternative only by the score of the best alternative under each criteria. This normalisation will be called later in the literature B-G normalisation or ideal mode. In the second article of the saga, Saaty and Vargas (1984c) provided a counter-example to show that the ideal mode is also subject to rank reversal. In this case, they introduce an alternative, which is a copy on only two of three criteria. On the last criterion, the new alternative has the largest value, which implies different normalisation and priorities. In the third episode, Belton and Gear (1985) responded that if a new alternative is introduced, then the weight criteria should also be modified. This contradicts the AHP philosophy of independence of weights and alternatives. In the fourth episode, Vargas (1985) claims that a method must not be applied only because it gives the results we want (i.e. preserve ranks). He argued that the process used if one has a set of alternatives and adds or deletes some alternatives should be the same if when starting from scratch. The same remark applies to the work of Wang and Elhag (2006), which propose a normalisation by the sum of all priorities with the exception of the new added one. Later, Schoner and Wedley (1989), showed a tangible example (i.e. all criteria are monetary measurable), that a weight change is required to preserve ranks. In a successive paper, Schoner et al. (1993) proposed the linking pin AHP. The local priorities of a specific alternative is normalised to unity. The rank is preserved because the normalisation [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 is always the same. However the final solution depends on which alternative is selected to link across criteria. This phenomenon, appearing also for near identical copy (Dyer, 1990b) or when a copy is removed (Troutt, 1988), has then been debated in Management Science, in the Journal of the Operational Society and in the European Journal of Operational Research where one side has criticized the rank reversal phenomenon (Dyer, 1990a, 1990b; Holder, 1990, 1991; Stam & Duarte Silva, 2003) and the other has legitimized it (Harker & Vargas, 1987, 1990; Pérez, 1995; T. Saaty, 1986; T. Saaty, 1990; T. Saaty, 1991, 1994; T. Saaty, 2006). Millet and Saaty (2000) gave some guidance on which normalisation to use. However, due to the absence of a causal effect demonstration, we believe that the occasional rank reversals are more side-effects of the procedure rather than credible results of the modelling procedure. The multiplicative aggregation (14) has been proposed to prevent the rank reversal phenomenon (Barzilai & Lootsma, 1997; Lootsma, 1993).  j jw iji lp (14) The multiplicative aggregation has non-linearity properties allowing a superior compromise to be select, which is not the case with the additive aggregation (Ishizaka, et al., 2010; Stam & Duarte Silva, 2003). However, Vargas (1997) showed that additive aggregation is the only way to retrieve exact weights of known objects. 2.8 Sensitivity analysis The last step of the decision process is the sensitivity analysis, where the input data are slightly modified in order to observe the impact on the results. As complex decision models may be inherently unstable, it allows the generation of different scenarios, which may results in other rankings and further discussion may be needed to reach a consensus. If the ranking does not change, the results are said to be robust otherwise it is sensitive. In AHP, the sensitivity analysis can be done on three levels: weights, local priorities and comparisons. The sensitivity analysis in Expert Choice allows the variation the weights of the criteria only as input data. Its interactive graphical interface allows a better visualization of the impact of the changes (figure 2). A sensitivity analysis to single and multiple changes of local priorities has also been studied (H. Chen & Kocaoglu, 2008; Huang, 2002; Masuda, 1990) but not yet implemented in a software. Armacost and Hosseini (1994), inspired from the dual questioning approach attribute (DQDA), have described the way to calculate the most determinant criteria. In fact, it is not necessarily the most weighted criterion that is the most critical; the weight must be multiplied by the difference of the local priorities of the alternatives. Triantaphyllou and Sánchez (1997) have defined a sensitivity coefficient of the weights and local priorities. It is calculated with the minimum change of the current weight/local priority such as the ranking of two alternatives is reversed. The sensitivity coefficient of criterion ck is given by: Sensitivity (ck) = 1/Dk, (15) where Dk is the smallest percent amount by which the current value must change such that the ranking is reversed: Dk = min{σk,i,j} for all k,i,j (16) [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 where k: criterion k i, j: alternative i, j and the perturbation σk,i,j is given by: σk,i,j = (pj-pi)/(lij-lii) AND σk,i,j ≤ wj (17) where pi: global priority of the alternative i lij: local priority of alternative i on the weight j The same paper (Triantaphyllou & Sánchez, 1997) contains also the formulas for the most sensitive weight and criteria for the multiplicative AHP. These parameters are critical and careful attention should be given to them. A sensitivity analysis at a micro level has been developed to calculate the interval a single comparison can vary without changing the rank of the alternatives (Aguarón & Moreno- Jiménez, 2000) and to remain in an acceptable inconsistency index (Aguarón & Moreno- Jiménez, 2003) for AHP using the geometric mean. Figure 2: An example of four possible graphical sensitivity analyses in Expert Choice 3. AHP in group decision making As a decision affects often several persons, the standard AHP has been adapted in order to be applied in group decisions. Consulting several experts avoids also bias that may be present when the judgements are considered from a single expert. There are four ways to combine the preferences into a consensus rating (table 4). [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 mathematical aggregation Yes No a g g re g a ti o n o n : judgements geometric mean on judgements consensus vote on judgements priorities weighted arithmetic mean on priorities consensus vote on priorities Table 4: Four ways to combine preferences. The consensus vote is used, when we have a synergistic group and not a collection of individuals. In this case, the hierarchy of the problem must be the same for all decision- makers. On the judgements level, this method requires the group to reach an agreement on the value of each entry in a matrix of pair-wise comparisons. A consistent agreement is usually difficult to obtain with increasing difficulty with the number of comparison matrices and related discussions. In order to bypass this difficulty, the consensus vote can be postponed after the calculation of the priorities of each participant. O’Learly (1993) recommends this version because an early aggregation could result ―in a meaningless average performance measure‖. An aggregation after the calculation of priorities allows to detect decision-makers from different boards and to discuss further any disagreement. If a consensus is difficult to achieve (e.g. with a large number of persons or distant persons), a mathematical aggregation can be adopted. Two synthesizing methods exist and provide the same results in case of perfect consistency of the pair-wise matrices (T. L. Saaty & Vargas, 2005). In the first method, the geometric mean of individual evaluations are used as elements in the pair-wise matrices and then priorities are computed. The geometric mean method (GMM) must be adopted instead of the arithmetical mean in order to preserve the reciprocal property (Aczél & Saaty, 1983). For example, if person A enters a comparison 9 and person B enters 1/9, then by intuition the mathematical consensus should be 9 1 9  =1, which is a geometric mean and not (9 + 1/9)/2 = 4.56, which is an arithmetic mean. Ramanathan and Ganesh (1994) give an example where the Pareto optimality (i.e. if all group members prefer A to B, then the group decision should prefer A) is not satisfied with the GMM. Van den Honert and Lootsma (1997) argue that this violation could be expected because the pair-wise assessments are a compromise of all the group members’ assessments and therefore it is a compromise that does not represent any opinion of the group member. Madu and Kuei (1995), Bryson (1996) and then Saaty and Vargas (2007) introduce a measure of the dispersion of the judgements (or consensus indicator) in order to avoid this problem. If the group is not homogenous, further discussions are required to reach a consensus. In the second method, decision-makers constitute the first level below the goal of the AHP hierarchy. Priorities are computed and then aggregated using the weighted arithmetic mean method (WAMM). Applications can be found in (A. Labib & Shah, 2001; A. Labib, Williams, & O’Connor, 1996). Arbel and Orgler (1990) have introduced a further level above the stakeholders’ level representing the several economics scenarios. This extra level determines the priorities (weights) of the stakeholders. In a compromised method individual’s derived priorities can be aggregated at each node. However according to Forman and Peniwati (1998), this method is ―less meaningful and not commonly used‖. Aggregation methods with linear programming (Mikhailov, 2004) and Bayesian approach (Altuzarra, Moreno-Jiménez, & Salvador, 2007) have been proposed in order to take a decision even when comparisons are missing, for example when a stakeholder [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 does not feel to have the expertise to judge a particular comparison. Uncertainty has also been taken into account in proposing several ranking with an attached probability (M Escobar & Moreno-jiménez, 2007; Van den Honert, 1998). Group decision may be skewed because of collusion or distortion of the judgements in order to advantage its preferred outcome. As individual identities are lost with an aggregation, we prefer to avoid an early aggregation. Condon, Golden, & Wasil (2003) have developed a programme in order to visualise the decision of each participant, which facilitate the detection of outliers. 4. Conclusion and future developments Decisions that need support methods are difficult by definition and therefore complex to model. A trade-off between prefect modelling and usability of the model should be achieved. It is our belief that AHP has reached this compromise and will be useful for many other cases as it has been in the past. In particular, AHP has broken through the academic community to be widely used by practitioners. This widespread use is certainly due to its ease of applicability and the structure of AHP which follows the intuitive way in which managers solve problems. The hierarchical modelling of the problem, the possibility to adopt verbal judgements and the verification of the consistency are its major assets. Expert Choice, the user-friendly supporting software, has certainly largely contributed to the success of the method. It incorporates intuitive graphical user interfaces, automatic calculation of priorities and inconsistencies and several ways to process a sensitivity analysis (Ishizaka & Labib, 2009). Today, several other supporting software packages have been developed: Decision Lens, HIPRE 3+, RightChoiceDSS, Criterium, EasyMind, Questfox, ChoiceResults, AHPProject, 123AHP... not to mention that a template in Excel could also be easily generated. Along with its traditional applications, a new trend, as compiled by the work of Ho (2008), is to use AHP in conjunction with others methods: mathematical programming techniques like linear programming, data envelopment analysis (DEA), fuzzy sets, house of quality, genetic algorithms, neural networks, SWOT-analysis and so on. There is little doubt that AHP will be more and more frequently adopted. AHP still suffers from some theoretical disputes. The rank reversal is surely the most debated problem. This phenomenon is still not fully resolved and maybe it will never be because the aggregation of preferences transposed from scales of different units is not easily interpretable and even questionable according to the French school (Roy, 1996). The assumption of criteria independence (no correlation) may be sometimes a limitation of AHP (and other MCDM methods). The Analytic Network Process (ANP), a generalisation of AHP with feed-backs to adjust weights, may be a solution. However the decision-maker must answer a much larger number of questions, which may be quite complex: e.g. ―Given an alternative and a criterion, which of the two alternatives influences the given criterion more and how much more than another alternative‖ (T. Saaty & Takizawa, 1986). A simplified ANP, while still keeping its proprieties, would be beneficial for a wider adoption of the method. Another direction of the research will probably be on a more soft side. The choice of a hierarchy and a judgement scale is important and difficult. Problem structuring methods could help in the construction of AHP hierarchies, which is its less formalised aspect (Petkov & Mihova-Petkova, 1997; Petkov, Petkova, Andrew, & Nepal, 2007). [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 5. References Aczél, J., & Saaty, T. (1983). Procedures for synthesizing ratio judgements. Journal of Mathematical Psychology, 27, 93-102. Aguarón, J., & Moreno-Jiménez, J. (2000). Local Stability Intervals in the Analytic Hierarchy Process. European Journal of Operational Research, 125, 113-132. Aguarón, J., & Moreno-Jiménez, J. (2003). The Geometric Consistency Index: Approximated Thresholds. European Journal of Operational Research, 147, 137-145. Akarte, M., Surendra, N., Ravi, B., & Rangaraj, N. (2001). Web based casting supplier evaluation using analytical hierarchy process. Journal of the Operational Research Society, 52, 511-522. Alonso, J., & Lamata, T. (2006). Consistency in the Analytic Hierarchy Process: a New Approach. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 14, 445-459. Altuzarra, A., Moreno-Jiménez, J., & Salvador, M. (2007). A Bayesian priorization procedure for AHP-group decision making. European Journal of Operational Research, 182, 367-382. Amiri, M., Zandieh, M., Soltani, R., & Vahdani, B. (2009). A hybrid multi-criteria decision- making model for firms competence evaluation. Expert Systems with Applications, 36, 12314-12322. Amiri, M. P. (2010). Project selection for oil-fields development by using the AHP and fuzzy TOPSIS methods. Expert Systems with Applications, In Press, Uncorrected Proof, doi: 10.1016/j.eswa.2010.1002.1103. Arbel, A., & Orgler, Y. (1990). An application of the AHP to bank strategic planning: The mergers and acquisitions process. European Journal of Operational Research, 48, 27- 37. Armacost, R., & Hosseini, J. (1994). Identification of determinant attributes using the Analytic Hierarchy Process. Journal of the Academy of Marketing Science, 22, 383- 392. Bajwa, G., Choo, E., & Wedley, W. (2008). Effectiveness Analysis of Deriving Priority Vectors from Reciprocal Pairwise Comparison Matrices. Asia-Pacific Journal of Operational Research, 25, 279-299. Bana e Costa, C., & Vansnick, J. (2008). A Critical Analysis of the Eigenvalue Method Used to Derive Priorities in AHP. European Journal of Operational Research, 187, 1422- 1428. Barthélemy, J. (2003). The seven deadly sins of outsourcing. Academy of Management Executive, 17, 87‑100. Barzilai, J. (1997). Deriving Weights from Pairwise Comparisons Matrices. Journal of the Operational Research Society, 48, 1226-1232. Barzilai, J. (2005). Measurement and preference function modelling. International Transactions in Operational Research, 12, 173-183. Barzilai, J., & Lootsma, F. (1997). Power relation and group aggregation in the multiplicative AHP and SMART. Journal of Multi-Criteria Decision Analysis, 6, 155-165. Belton, V., & Gear, A. (1983). On a Shortcoming of Saaty's Method of Analytical Hierarchies. Omega, 11, 228-230. Belton, V., & Gear, A. (1985). The Legitimacy of Rank Reversal—A Comment. Omega, 13, 143-144. Belton, V., & Stewart, T. J. (2002). Multiple Criteria Decision Analysis: An Integrated Approach. Boston: Kluwer Academic Publishers. [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Brugha, C. (2004). Structure of multi-criteria decision-making. Journal of the Operational Research Society, 55, 1156-1168. Bryson, N. (1996). Group decision-making and the analytic hierarchy process: Exploring the consensus-relevant information content. Computers & Operations Research, 23, 27- 35. Budescu, D. (1984). Scaling binary comparison matrices: A comment on Narasimhan’s proposal and other methods. Fuzzy Sets and Systems, 14, 187-192. Budescu, D., Zwick, R., & Rapoport, A. (1986). A comparison of the eigenvalue method and the geometric mean procedure for ratio scaling. Applied psychological measurement, 10, 69–78. Carmone, F., Kara, A., & Zanakis, S. (1997). A Monte Carlo investigation of incomplete pairwise comparison matrices in AHP. European Journal of Operational Research, 102, 538-553. Cebeci, U. (2009). Fuzzy AHP-based decision support system for selecting ERP systems in textile industry by using balanced scorecard. Expert Systems with Applications, 36, 8900-8909. Celik, M., Kandakoglu, A., & Er, D. (2009). Structuring fuzzy integrated multi-stages evaluation model on academic personnel recruitment in MET institutions. Expert Systems with Applications, 36, 6918-6927. Chamodrakas, I., Batis, D., & Martakos, D. (2010). Supplier selection in electronic marketplaces using satisficing and fuzzy AHP. Expert Systems with Applications, 37, 490-498. Chang, C.-W., Wu, C.-R., & Lin, H.-L. (2009). Applying fuzzy hierarchy multiple attributes to construct an expert decision making process. Expert Systems with Applications, 36, 7363-7368. Chen, H., & Kocaoglu, D. (2008). A sensitivity analysis algorithm for hierarchical decision models. European Journal of Operational Research, 185, 266-288. Chen, M. K., & Wang, S.-C. (2010). The critical factors of success for information service industry in developing international market: Using analytic hierarchy process (AHP) approach. Expert Systems with Applications, 37, 694-704. Cho, E., & Wedley, W. (2004). A Common Framework for Deriving Preference Values fro m Pairwise Comparison Matrices. Computers and Operations Research 31, 893-908. Condon, E., Golden, B., & Wasil, E. (2003). Visualizing group decisions in the analytic hierarchy process. Computers & Operations Research, 30, 1435-1445. Crawford G, & C, W. (1985). A Note on the Analysis of Subjective Judgement Matrices. Journal of Mathematical Psychology, 29, 387-405. Dagdeviren, M., Yavuz, S., & KilInç, N. (2009). Weapon selection using the AHP and TOPSIS methods under fuzzy environment. Expert Systems with Applications, 36, 8143-8151. Dodd, F., & Donegan, H. (1995). Comparison of priotization techniques using interhierarchy mappings. Journal of the Operational Research Society, 46, 492-498. Donegan, H., Dodd, F., & McMaster, T. (1992). A new approach to AHP decision-making. The Statician 41, 295-302. Dyer, J. (1990a). A clarification of ―Remarks on the Analytic Hierarchy Process‖. Management Science, 36, 274-275. Dyer, J. (1990b). Remarks on the Analytic Hierarchy Process. Management Science, 36, 249- 258. Escobar, M., & Moreno-Jiménez, J. (2000). Reciprocal distributions in the analytic hierarchy process. European Journal of Operational Research, 123, 154-174. [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Escobar, M., & Moreno-jiménez, J. (2007). Aggregation of Individual Preference Structures in Ahp-Group Decision Making. Group Decision and Negotiation, 16, 287-301. Fechner , G. (1860). Elemente der Psychophysik (Vol. 2): Breitkopf und Härtel. Fichtner, J. (1986). On deriving priority vectors from matrices of pairwise comparisons. Socio-Economic Planning Sciences, 20, 341-345. Forman, E. (1990). Random Indices for Incomplete Pairwise Comparison Matrices. European Journal of Operational Research, 48, 153-155. Forman, E., & Gass, S. (2001). The Analytic Hierarchy Process – An Exposition. Operations Research, 49, 469-486. Forman, E., & Peniwati, K. (1998). Aggregating individual judgments and priorities with the analytic hierarchy process. European Journal of Operational Research, 108, 165-169. Forman, E. H. (1990). Random indices for incomplete pairwise comparison matrices. European Journal of Operational Research, 48, 153-155. Golany, B., & Kress, M. (1993). A multicriteria evaluation of the methods for obtaining weights from ratio-scale matrices. European Journal of Operational Research, 69, 210–220. Golden, B., Wasil, E., & Harker, P. (1989). The Analytic Hierarchy Process: Applications and Studies. Heidelberg: Springer-Verlag. Haghighi, M., Divandari, A., & Keimasi, M. (2010). The impact of 3D e-readiness on e- banking development in Iran: A fuzzy AHP analysis. Expert Systems with Applications, 37, 4084-4093. Harker, P. (1987). Incomplete Pairwise comparisons in the Analytic Hierarchy Process. Mathematical and Computer Modelling. Harker, P., & Vargas, L. (1987). The Theory of Ratio Scale Estimation: Saaty's Analytic Hierarchy Process. Management Science, 33, 1383-1403. Harker, P., & Vargas, L. (1990). Reply to ―Remarks on the Analytic Hierarchy Process‖. Management Science, 36, 269-273. Harker, P. T. (1987a). Alternative modes of questioning in the analytic hierarchy process. Mathematical Modelling, 9, 353-360. Harker, P. T. (1987b). Incomplete pairwise comparisons in the analytic hierarchy process. Mathematical Modelling, 9, 837-848. Herman, M., & Koczkodaj, W. (1996). A Monte Carlo Study of Pairwise Comparison. Information Processing Letters 57, 25-29. Ho, W. (2008). Integrated analytic hierarchy process and its applications - A literature review. European Journal of Operational Research, 186, 211-228. Ho, W., & Emrouznejad, A. (2009). Multi-criteria logistics distribution network design using SAS/OR. Expert Systems with Applications, 36, 7288-7298. Holder, R. (1990). Some Comment on the Analytic Hierarchy Process. Journal of the Operational Research Society, 41, 1073-1076. Holder, R. (1991). Response to Holder's Comments on the Analytic Hierarchy Process: Response to the Response. Journal of the Operational Research Society, 42, 914-918. Hovanov, N., Kolari, J., & Sokolov, M. (2008). Deriving weights from general pairwise comparisons matrices. Mathematical Social Sciences 55, 205-220. Hsu, S. H., Kao, C.-H., & Wu, M.-C. (2009). Design facial appearance for roles in video games. Expert Systems with Applications, 36, 4929-4934. Hsu, Y.-L., Lee, C.-H., & Kreng, V. B. (2010). The application of Fuzzy Delphi Method and Fuzzy AHP in lubricant regenerative technology selection. Expert Systems with Applications, 37, 419-425. Huang, Y.-F. (2002). Enhancement on sensitivity analysis of priority in analytic hierarchy process. International Journal of General Systems, 31, 531 - 542. [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Iç, Y. T., & Yurdakul, M. (2009). Development of a decision support system for machining center selection. Expert Systems with Applications, 36, 3505-3513. Ishizaka, A. (2004a). The Advantages of Clusters in AHP. In 15th Mini-Euro Conference MUDSM. Coimbra. Ishizaka, A. (2004b). Développement d’un Système Tutorial Intelligent pour Dériver des Priorités dans l’AHP. Berlin: http://www.dissertation.de. Ishizaka, A. (2008). A Multicriteria Approach with AHP and Clusters for Supplier Selection. In 15th International Annual EurOMA Conference. Groningen. Ishizaka, A., Balkenborg, D., & Kaplan, T. (2010). Influence of aggregation and measurement scale on ranking a compromise alternative in AHP. Journal of the Operational Research Society, 62, 700-710. Ishizaka, A., & Labib, A. (2009). Analytic Hierarchy Process and Expert Choice: benefits and limitations. OR Insight, 22, 201–220. Ishizaka, A., & Lusti, M. (2004). An Expert Module to Improve the Consistency of AHP Matrices. International Transactions in Operational Research, 11, 97-105. Ishizaka, A., & Lusti, M. (2006). How to derive priorities in AHP: a comparative study. Central European Journal of Operations Research 14, 387-400. Ji, P., & Jiang, R. (2003). Scale transitivity in the AHP. Journal of the Operational Research Society, 54, 896-905. Johnson, C., Beine, W., & Wang, T. (1979). Right-Left Asymmetry in an Eigenvector Ranking Procedure. Journal of Mathematical Psychology, 19, 61-64. Jones, D., & Mardle, S. (2004). A Distance-Metric Methodology for the Derivation of Weights from a Pairwise Comparison Matrix. Journal of the Operational Research Society, 55, 869-875. Kahraman, C., & Kaya, I. (2010). A fuzzy multicriteria methodology for selection among energy alternatives. Expert Systems with Applications, In Press, Corrected Proof, DOI: 10.1016/j.eswa.2010.1002.1095. Kainulainen, T., Leskinen, P., Korhonen, P., Haara, A., & Hujala, T. (2009). A statistical approach to assessing interval scale preferences in discrete choice problems. Journal of the Operational Research Society, 60, 252-258. Karapetrovic, S., & Rosenbloom, E. (1999). A Quality Control Approach to Consistency Paradoxes in AHP. European Journal of Operational Research, 119, 704-718. Khosla, R., Goonesekera, T., & Chu, M.-T. (2009). Separating the wheat from the chaff: An intelligent sales recruitment and benchmarking system. Expert Systems with Applications, 36, 3017-3027. Kumar, S., & Vaidya, O. (2006). Analytic hierarchy process: An overview of applications. European Journal of Operational Research, 169, 1-29. Kwiesielewicz, M., & van Uden, E. (2004). Inconsistent and Contradictory Judgements in Pairwise Comparison Method in AHP. Computers and Operations Research 31, 713- 719. Labib, A., & Shah, J. (2001). Management decisions for a continuous improvement process in industry using the Analytical Hierarchy Process. Journal of Work Study, 50, 189- 193. Labib, A., Williams, G., & O’Connor, R. (1996). Formulation of an appropriate productive maintenance strategy using multiple criteria decision making. Maintenance Journal, 11, 66-75. Labib, A. W. (2011). A supplier selection model: a comparison of fuzzy logic and the analytic hierarchy process. International Journal of Production Research. Lai, W.-H., & Tsai, C.-T. (2009). Fuzzy rule-based analysis of firm's technology transfer in Taiwan's machinery industry. Expert Systems with Applications, 36, 12012-12022. http://www.dissertation.de/ [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Lane, E., & Verdini, W. (1989). A Consistency Test for AHP Decision Makers. Decision Sciences, 20, 575-590. Lee, S.-H. (2010). Using fuzzy AHP to develop intellectual capital evaluation model for assessing their performance contribution in a university. Expert Systems with Applications, In Press, Corrected Proof, DOI: 10.1016/j.eswa.2009.1012.1020. Leskinen, P., & Kangas, J. (2005). Rank reversal in multi-criteria decision analysis with statistical modelling of ratio-scale pairwise comparisons. Journal of the Operational Research Society, 56, 855-861. Li, S., & Li, J. Z. (2009). Hybridising human judgment, AHP, simulation and a fuzzy expert system for strategy formulation under uncertainty. Expert Systems with Applications, 36, 5557-5564. Li, T.-S., & Huang, H.-H. (2009). Applying TRIZ and Fuzzy AHP to develop innovative design for automated manufacturing systems. Expert Systems with Applications, 36, 8302-8312. Li, Y., Tang, J., & Luo, X. (2010). An ECI-based methodology for determining the final importance ratings of customer requirements in MP product improvement. Expert Systems with Applications, In Press, Uncorrected Proof, doi: 10.1016/j.eswa.2010.1002.1100. Liberatore, M., & Nydick, R. (2008). The analytic hierarchy process in medical and health care decision making: A literature review. European Journal of Operational Research, 189, 194-207. Lim, K., & Swenseth, S. (1993). An iterative procedure for reducing problem size in large scale AHP problems. European Journal of Operational Research, 67, 64-74. Limam Mansar, S., Reijers, H., & Ounnar, F. (2009). Development of a decision-making strategy to improve the efficiency of BPR. Expert Systems with Applications, 36, 3248-3262. Lin, C.-L., Chen, C.-W., & Tzeng, G.-H. (2010). Planning the development strategy for the mobile communication package based on consumers' choice preferences. Expert Systems with Applications, In Press, Corrected Proof, DOI: 10.1016/j.eswa.2009.1011.1009. Lin, C. (2007). A Revised Framework for Deriving Preference Values from Pairwise Comparison Matrices. European Journal of Operational Research, 176, 1145-1150. Liu, C.-C., & Chen, S.-Y. (2009). Prioritization of digital capital measures in recruiting website for the national armed forces. Expert Systems with Applications, 36, 9415- 9421. Lootsma, F. (1989). Conflict Resolution via Pairwise Comparison of Concessions. European Journal of Operational Research, 40, 109-116. Lootsma, F. (1993). Scale sensitivity in the multiplicative AHP and SMART. Journal of Multi-Criteria Decision Analysis, 2, 87-110. Lootsma, F. (1996). A model for the relative importance of the criteria in the Multiplicative AHP and SMART. European Journal of Operational Research, 94, 467-476. Ma, D., & Zheng, X. (1991). 9/9-9/1 Scale Method of AHP. In 2nd Int. Symposium on AHP (Vol. 1, pp. 197-202). Pittsburgh Madu, C., & Kuei, C.-H. (1995). Stability analyses of group decision making. Computers & Industrial Engineering, 28, 881-892. Masuda, T. (1990). Hierarchical sensitivity analysis of priority used in analytic hierarchy process. International Journal of Systems Science, 21, 415 - 427. Mikhailov, L. (2004). Group prioritization in the AHP by fuzzy preference programming method. Computers & Operations Research, 31, 293-301. [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Mikhailov, L., & Singh, M. G. (1999). Comparison analysis of methods for deriving priorities in the analytic hierarchy process. In IEEE International Conference on Systems, Man, and Cybernetics (Vol. 1, pp. 1037-1042 ). Tokyo. Miller, G. A. (1956). The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychological Review, 63, 81-97. Miller, J. (1966). The assessment of Worth: a systematic procedure and its experimental validation. MIT. Miller, J. (1969). Assessing alternative transportation systems. In Memorandum RM-5865- DOR: The RAND Corporation. Miller, J. (1970). Professional decision-making: a procedure for evaluation complex alternatives. New York: Praeger Publishers. Millet, I., & Harker, P. (1990). Globally effective questioning in the Analytic Hierarchy Process. European Journal of Operational Research, 48, 88-97. Millet, I., & Saaty, T. (2000). On the Relativity of Relative Measures-Accommodating both Rank Preservation and Rank Reversals in the AHP. European Journal of Operational Research, 121, 205-212. Millet, I., & Schoner, B. (2005). Incorporating negative values into the Analytic Hierarchy Process. Computers and Operations Research 32, 3163-3173. Naghadehi, M. Z., Mikaeil, R., & Ataei, M. (2009). The application of fuzzy analytic hierarchy process (FAHP) approach to selection of optimum underground mining method for Jajarm Bauxite Mine, Iran. Expert Systems with Applications, 36, 8218- 8226. Niaraki, A. S., & Kim, K. (2009). Ontology based personalized route planning system using a multi-criteria decision making approach. Expert Systems with Applications, 36, 2250- 2259. O'Leary, D. (1993). Determining Differences in Expert Judgment: Implications for Knowledge Acquisition and Validation. Decision Sciences, 24, 395-408. Omkarprasad, V., & Sushil, K. (2006). Analytic hierarchy process: an overview of applications. European Journal of Operational Research, 169, 1-29. Önüt, S., Efendigil, T., & Soner Kara, S. (2009). A combined fuzzy MCDM approach for selecting shopping center site: An example from Istanbul, Turkey. Expert Systems with Applications, In Press, Uncorrected Proof, doi:10.1016/j.eswa.2009.1006.1080. Pan, N. (2009). Selecting an appropriate excavation construction method based on qualitative assessments. Expert Systems with Applications, 36, 5481-5490. Peláez, P., & Lamata, M. (2003). A New Measure of Consistency for Positive Reciprocal Matrices. Computers & Mathematics with Applications, 46, 1839-1845. Pérez, J. (1995). Some comments on Saaty’s AHP. Management Science, 41, 1091-1095. Petkov, D., & Mihova-Petkova, O. (1997). The Analytic Hierarchy Process and Systems Thinking. In 13th International MCDM Conference (Vol. 465, pp. 243-252). Cape Town Springer. Petkov, D., Petkova, O., Andrew, T., & Nepal, T. (2007). Mixing Multiple Criteria Decision Making with soft systems thinking techniques for decision support in complex situations. Decision Support Systems, 43, 1615-1629. Pöyhönen, M., Hamalainen, R., & Salo, A. (1997). An Experiment on the Numerical Modelling of Verbal Ratio Statements. Journal of Multi-Criteria Decision Analysis, 6, 1-10. Ramanathan, R., & Ganesh, L. (1994). Group preference aggregation methods employed in AHP: An evaluation and an intrinsic process for deriving members' weightages. European Journal of Operational Research, 79, 249-265. [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Roy, B. (1996). Multicriteria Methodology for Decision Analysis. Dordrecht: Kluwer Academic Publishers. Saaty, T. (1972). An eigenvalue allocation model for prioritization and planning. In Working paper, Energy Management and Policy Center: University of Pennsylvania. Saaty, T. (1977). A scaling method for priorities in hierarchical structures. Journal of Mathematical Psychology, 15, 234-281. Saaty, T. (1980). The Analytic Hierarchy Process. New York: McGraw-Hill. Saaty, T. (1986). Axiomatic Foundation of the Analytic Hierarchy Process. Management Science, 32, 841-855. Saaty, T. (1990). An Exposition of the AHP in Reply to the Paper ―Remarks on the Analytic Hierarchy Process‖. Management Science, 36, 259-268. Saaty, T. (1991). Response to Holder’s Comments on the Analytic Hierarchy Process. Journal of the Operational Research Society, 42, 909-929. Saaty, T. (1994). Highlights and critical points in the theory and application of the Analytic Hierarchy Process. European Journal of Operational Research, 74, 426-447. Saaty, T. (2003). Decision-making with the AHP: Why is the Principal Eigenvector necessary? European Journal of Operational Research, 145, 85-91. Saaty, T. (2006). Rank from Comparisons and from Ratings in the Analytic Hierarchy/Network Processes. European Journal of Operational Research, 168, 557- 570. Saaty, T., & Forman, E. (1992). The Hierarchon: A Dictionary of Hierarchies (Vol. V). Pittsburgh: RWS Publications. Saaty, T., & Hu, G. (1998). Ranking by Eigenvector Versus other Methods in the Analytic Hierarchy Process. Applied Mathematics Letters 11, 121-125. Saaty, T., & Ozdemir, M. (2003). Negative Priorities in the Analytic Hierarchy Process. Mathematical and Computer Modelling, 37, 1063-1075. Saaty, T., & Takizawa, M. (1986). Dependence and Independence: from Linear Hierarchies to Nonlinear Networks. European Journal of Operational Research, 26, 229-237. Saaty, T., & Vargas, L. (1984a). Comparison of Eigenvalue, Logarithmic Least Squares and Least Squares Methods in Estimating Ratios. Mathematical Modeling 5, 309-324. Saaty, T., & Vargas, L. (1984b). Inconsistency and Rank Preservation. Journal of Mathematical Psychology, 28, 205-214. Saaty, T., & Vargas, L. (1984c). The legitimacy of rank reversal. Omega, 12, 513-516. Saaty, T., & Vargas, L. (2007). Dispersion of group judgments. Mathematical and Computer Modelling, 46, 918-925. Saaty, T. L., & Vargas, L. G. (2005). The possibility of group welfare functions. International Journal of Information Technology & Decision Making, 4, 167-176. Salo, A., & Hamalainen, R. (1997). On the Measurement of Preference in the Analytic Hierarchy Process. Journal of Multi-Criteria Decision Analysis, 6, 309-319. Schoner, B., & Wedley, W. (1989). Ambiguous Criteria Weights in AHP: Consequences and Solutions. Decision Sciences, 20, 462-475. Schoner, B., Wedley, W., & Choo, E. (1993). A Unified Approach to AHP with Linking Pins. European Journal of Operational Research, 64, 384-392. Seçme, N. Y., Bayrakdaroglu, A., & Kahraman, C. (2009). Fuzzy performance evaluation in Turkish Banking Sector using Analytic Hierarchy Process and TOPSIS. Expert Systems with Applications, 36, 11699-11709. Sen, C. G., & ÇInar, G. (2010). Evaluation and pre-allocation of operators with multiple skills: A combined fuzzy AHP and max-min approach. Expert Systems with Applications, 37, 2043-2053. [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Shen, Y., Hoerl, A., & McConnell, W. (1992). An incomplete design in the analytic hierarchy process. Mathematical and Computer Modelling, 16, 121-129. Shim, J. (1989). Bibliography research on the analytic hierarchy process (AHP). Socio- Economic Planning Sciences, 23, 161-167. Sipahi, S., & Timor, M. (2010). The analytic hierarchy process and analytic network process: an overview of applications. Management Decision, 48, 775-808. Stam, A., & Duarte Silva, P. (2003). On Multiplicative Priority Rating Methods for AHP. European Journal of Operational Research, 145, 92-108. Stein, W., & Mizzi, P. (2007). The Harmonic Consistency Index for the Analytic Hierarchy Process. European Journal of Operational Research, 177, 488-497. Stevens, S. (1957). On the psychophysical law. Psychological Review, 64, 153-181. Stillwell, W., von Winterfeldt, D., & John, R. (1987). Comparing hierarchical and non- hierarchical weighting methods for eliciting multiattribute value models. Management Science, 33, 442-450. Su, S., Yu, J., & Zhang, J. (2010). Measurements study on sustainability of China's mining cities. Expert Systems with Applications, In Press, DOI: 10.1016/j.eswa.2010.1002.1140. Thurstone, L. (1927). A law of comparative judgments. Psychological Review, 34, 273-286. Triantaphyllou, E. (2001). Two new cases of rank reversals when the AHP and some of its additive variants are used that do not occur with the Multiplicative AHP. Journal of Multi-Criteria Decision Analysis, 10, 11-25. Triantaphyllou, E., & Sánchez, A. (1997). A sensitivity analysis approach for some deterministic multi-criteria decision-making methods. Decision Sciences, 28, 151- 194. Troutt, M. (1988). Rank Reversal and the Dependence of Priorities on the Underlying MAV Function. Omega, 16, 365-367. Tseng, Y.-F., & Lee, T.-Z. (2009). Comparing appropriate decision support of human resource practices on organizational performance with DEA/AHP model. Expert Systems with Applications, 36, 6548-6558. Tummala, V., & Wan, Y. (1994). On the Mean Random Inconsistency Index of the Analytic Hierarchy Process (AHP). Computers & Industrial Engineering, 27, 401-404. Van den Honert, R. (1998). Stochastic group preference modelling in the multiplicative AHP: A model of group consensus. European Journal of Operational Research, 110, 99- 111. Van Den Honert, R., & Lootsma, F. (1997). Group preference aggregation in the multiplicative AHP The model of the group decision process and Pareto optimality. European Journal of Operational Research, 96, 363-370. Vargas, L. (1985). A Rejoinder. Omega, 13, 249. Vargas, L. (1990). An overview of the analytic hierarchy process and its applications. European Journal of Operational Research, 48, 2-8. Vargas, L. (1997). Comments on Barzilai and Lootsma Why the Multiplicative AHP is Invalid: A Practical Counterexample. Journal of Multi-Criteria Decision Analysis, 6, 169-170. Vidal, L.-A., Sahin, E., Martelli, N., Berhoune, M., & Bonan, B. (2010). Applying AHP to select drugs to be produced by anticipation in a chemotherapy compounding unit. Expert Systems with Applications, 37, 1528-1534. Wang, H. S., Che, Z. H., & Wu, C. (2010). Using analytic hierarchy process and particle swarm optimization algorithm for evaluating product plans. Expert Systems with Applications, 37, 1023-1034. [Pre-print version], please cite as: Ishizaka A., Labib A. Review of the main developments in the analytic hierarchy process, Expert Systems with Applications, 38(11), 14336-14345, 2011 Wang, T.-Y., & Yang, Y.-H. (2009). A fuzzy model for supplier selection in quantity discount environments. Expert Systems with Applications, 36, 12179-12187. Wang, Y., Chin, K.-S., & Luo, Y. (2009). Aggregation of direct and indirect judgements in pairwise comparison matrices with a re-examination of the criticisms by Bana e Costa and Vansnick. Information Sciences, 179, 329-337. Wang, Y., & Elhag, T. (2006). An Approach to Avoiding Rank Reversal in AHP. Decision Support Systems, 42, 1474-1480. Wang, Y., & Luo, Y. (2009). On Rank Reversal in Decision Analysis. Mathematical and Computer Modelling, 49, 1221-1229. Webber, S., Apostolou, B., & Hassell, J. (1996). The sensitivity of the analytic hierarchy process to alternative scale and cue presentations. European Journal of Operational Research, 96, 351-362. Weber, M., Eisenführ, F., & von Winterfeldt, D. (1988). The Effects of Spitting Attributes on Weights in Multiattribute Utility Measurement. Management Science, 34, 431-445. Wedley, W. (1993). Consistency prediction for incomplete AHP matrices. Mathematical and Computer Modelling, 17, 151-161. Wedley, W., Schoner, B., & Tang, T. (1993). Starting rules for incomplete comparisons in the analytic hierarchy process. Mathematical and Computer Modelling, 17, 93-100. Wu, C.-R., Lin, C.-T., & Lin, Y.-F. (2009). Selecting the preferable bancassurance alliance strategic by using expert group decision technique. Expert Systems with Applications, 36, 3623-3629. Yang, C.-L., Chuang, S.-P., & Huang, R.-H. (2009). Manufacturing evaluation system based on AHP/ANP approach for wafer fabricating industry. Expert Systems with Applications, 36, 11369-11377. Yokoyama, M. (1921). The Nature of the affective judgment in the method of paired comparison. The American Journal of Psychology 32, 357-369. Zahedi, F. (1986). The analytic hierarchy process: a survey of the method and its applications. Interface 16, 96-108. 
work_6svi36sxgvejlhk3ehxl4ihkhu	----	An innovative blemish detection system for curved LED lenses An innovative blemish detection system for curved LED lenses Yuan-ShyiPeter Chiu, Hong-Dar Lin ⇑ Department of Industrial Engineering and Management, Chaoyang University of Technology, 168 Jifong E. Rd., Wufong District, Taichung 41349, Taiwan a r t i c l e i n f o Keywords: Blemish detection system Curved LED lens Discrete cosine transform Energy features Grey clustering a b s t r a c t The functions of LED lenses include focusing, beauty, and protection to avoid the waste of light and light pollution. Nevertheless, LED lens with a transparent and curved surface is more difficult to detect the visual blemishes than electronic and optical components by current computer vision systems. This research proposes an innovative blemish detection system to detect visual blemishes of the curved LED lenses. A spatial domain image with equal sized blocks is converted to discrete cosine transform (DCT) domain and some representative energy features of each DCT block are extracted. These energy features of each block are integrated by the Hotelling’s T-squared statistic and the suspected blemish blocks can be determined by the multivariate statistical method. Then, the grey clustering technique based on the block grey relational grades is applied to further confirm the block locations of real blem- ishes. Finally, a simple thresholding method is applied to set a threshold for distinguishing between defective areas and uniform regions. Experimental results show that the proposed system achieves a high 95.46% probability of correctly discriminating visual blemishes from normal regions and a low 0.13% probability of erroneously detecting normal regions as blemishes on curved surfaces of LED lenses. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction A lens is an optical device with perfect or approximate axial symmetry which transmits and refracts light, converging or diverg- ing the beam. Lenses are typically made of glass or transparent plastic. Optical lenses are transparent components made from optical-quality materials and curved to converge or diverge trans- mitted rays from an object. These rays then form a real or virtual image of the object. There are many types of optical lenses. Optical lenses are widely used in cell phones, notebooks, automotive lights, digital cameras, scanners, head lamps etc. A light-emitting diode (LED) is a semiconductor device that emits visible light when an electric current passes through the semiconductor chip. Compared with incandescent and fluorescent illuminating devices, LEDs have lower power requirement, higher efficiency, and longer lifetime. Typical applications of LED compo- nents include indicator lights, LCD panel backlighting, fiber optic data transmission, etc. To meet consumer and industry needs, LED products are being made in smaller sizes, which increase dif- ficulties of product inspection. The functions of LED lenses include focusing, beauty, and protection to avoid the waste of light and light pollution. An LED without the assistance of lens focus func- tion cannot project light to the intended location. Therefore, LED lenses are invented to improve the light scattering problems of LEDs and they are widely applied to hand flashlights and traffic lights applications. Fig. 1 shows the common LED lens and LED lens product. Lens inspection requires special physical conditions, particu- larly in terms of lighting. In the real working situation, each in- spected lens is brought into the inspector’s field of vision. The lenses are round and transparent; the blemish to be inspected could be located on the external surface of the lenses or inside. A lens presents a certain thickness and a certain curvature, both of which vary. At times, lenses provide the same perceptive result as a magnifying glass, and the blemishes are all the more difficult to track down and to locate in the area of the lens. The majority of blemishes are not only very small but also they are extremely di- verse and can assume various forms. Fig. 2 presents LED lenses with and without visual blemishes. Currently, the most common detection methods for LED lens blemishes are human visual inspection. Human visual inspection is tedious, time-consuming and highly dependent on the inspec- tors’ experiences, conditions, or moods. Erroneous judgments are easily made because of inspectors’ subjectivity and eye fatigues. Difficulties exist in precisely inspecting tiny flaws by machine vi- sion systems because when product images are being captured, the area of a tiny flaw could expand, shrink or even disappear due to uneven illumination of the environment, transparent and curved surfaces of the product, and so on. Seeing the great need for an automated visual detection scheme for LED lens blemishes, we propose an innovative detection system applying block discrete cosine transform (BDCT) and gray clustering technique to over- come the difficulties of traditional machine vision systems. 0957-4174/$ - see front matter � 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2012.07.041 ⇑ Corresponding author. Tel.: +886 4 2332 3000x4258; fax: +886 4 2374 2327. E-mail address: hdlin@cyut.edu.tw (H.-D. Lin). Expert Systems with Applications 40 (2013) 471–479 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2012.07.041 mailto:hdlin@cyut.edu.tw http://dx.doi.org/10.1016/j.eswa.2012.07.041 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 2. Blemish inspection methods Inspection of surface blemishes has become a critical task for manufacturers who strive to improve product quality and produc- tion efficiency (Lin & Lin, 2009; Lin, Lin, Chung, & Lin, 2008). Blem- ish detection techniques, generally classified into the spatial domain and the frequency domain, compute a set of textural fea- tures in a sliding window and search for significant local deviations among the feature values. Latif-Amet, Ertüzün, and Ercil (2000) presented wavelet theory and co-occurrence matrices for detection of blemishes encountered in textile images and classify each sub- window as defective or non-defective with a Mahalanobis distance. Cho, Chung, and Park (2005) applied the adaptive threshold tech- nique and morphology method to detect defects from images of uniform fabrics for developing a real-time vision system. As to techniques in the frequency domain, Chan and Pang (2000) used the Fourier transform to detect fabric defects. Tsai and Hsiao (2001) proposed a wavelet transform based approach for inspecting local defects embedded in homogeneous textured surfaces. By properly selecting the smooth sub-image or the com- bination of detail sub-images in different decomposition levels for backward wavelet transform, regular, repetitive texture patterns can be removed and only local anomalies are enhanced in the reconstructed image. Also, Lin and Ho (2007) developed a novel ap- proach that applies discrete cosine transform based enhancement for the detection of pinhole defects on passive component chips. As to inspecting blemishes of lenses, Rebsamen, Boucheix, and Fayol (2010) described quality control tasks in the optical industry from a work analysis of optical lens inspection to a training pro- gram. Martínez, Ortega, García, and García (2009) developed a vi- sion sensor planning system for automated inspection of headlamp lenses. This system uses the lens CAD, a vision sensor model and the customer requirements describing by a fuzzy ap- proach, to achieve an optimal set of viewpoints by genetic algo- rithm. Bazin, Cole, Kett, and Nixon (2006) proposed a novel method for the industrial inspection of ophthalmic contact lenses in a time constrained production line environment. Perng, Wang, and Chen (2010) presented a new inspection system that uses ma- chine vision to detect optical blemishes in quasi-contact lenses. The optical region of the lens image is first segmented, then the middle axes of each fringe on the optical region are determined. Three features of the fringe are extracted to create a mapping of the original features to a semantic description of the textures. Fi- nally, the quality of the quasi-contact lens is determined by a con- trol chart procedure. Therefore, most of the existing researches focus on inspections of optical lenses, headlamp lenses, and con- tact lenses. They do not detect blemishes with the properties of tiny blemishes on LED lenses. Consequently, we present a new ap- proach using block discrete cosine transform and grey clustering for blemish detection of the transparent and curved LED lens surfaces. Ahmed, Natarajan, and Rao (1974) first defined DCT as one- dimensional (1-D) and suitable for 1-D digital signal processing. Most of the researchers that apply DCT for image processing fo- cused on image compression and image reconstruction, because DCT has the property of packing the most information into the few- est coefficients (Gonzalez & Woods, 2008). However, due to the lack of an efficient algorithm, DCT had not been applied as widely as its properties imply. Only until recently, many algorithms and VLSI architectures for the fast computation of DCT have been pro- posed (Kok, 1997). To enhance edges of remote sensing image data in the DCT domain, Chen, Latifi, and Kanai (1999) developed new and fast algorithms that involve three steps: high-pass filtering, gray levels adding, and contrast stretching. Fig. 1. (a) An LED lens (side view); (b) an LED lens product (top view). Fig. 2. LED lenses with and without visual blemishes (from top view) (a) LED lens without blemish; (b) LED lens with dirt blemishes; (c) LED lens with scratch blemishes. 472 Y.-S. P. Chiu, H.-D. Lin / Expert Systems with Applications 40 (2013) 471–479 https://isiarticles.com/article/2362 
work_6uetnwsvf5bptoiu6ksl6lo2se	----	Fuzzy VIKOR with an application to water resources planning Expert Systems with Applications 38 (2011) 12983–12990 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Fuzzy VIKOR with an application to water resources planning Serafim Opricovic Faculty of Civil Engineering, Bulevar Kralja Aleksandra 73, 11000 Belgrade, Serbia a r t i c l e i n f o Keywords: Fuzzy VIKOR Defuzzification Multicriteria Compromise solution Water resources 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.04.097 E-mail addresses: seropric@yahoo.com, seropric@g a b s t r a c t The fuzzy VIKOR method has been developed to solve fuzzy multicriteria problem with conflicting and noncommensurable (different units) criteria. This method solves problem in a fuzzy environment where both criteria and weights could be fuzzy sets. The triangular fuzzy numbers are used to handle imprecise numerical quantities. Fuzzy VIKOR is based on the aggregating fuzzy merit that represents distance of an alternative to the ideal solution. The fuzzy operations and procedures for ranking fuzzy numbers are used in developing the fuzzy VIKOR algorithm. VIKOR (VIsekriterijumska optimizacija i KOmpromisno Resen- je) focuses on ranking and selecting from a set of alternatives in the presence of conflicting criteria, and on proposing compromise solution (one or more). It is extended with a trade-offs analysis. A numerical example illustrates an application to water resources planning, utilizing the presented methodology to study the development of a reservoir system for the storage of surface flows of the Mlava River and its tributaries for regional water supply. A comparative analysis of results by fuzzy VIKOR and few different approaches is presented. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction There are situations when the evaluation of alternatives must handle the imprecision of established criteria, and the develop- ment of a fuzzy multicriteria decision model is necessary to deal with either ‘‘qualitative’’ (unquantifiable or linguistic) or incom- plete information (Vanegas & Labib, 2001; Zadeh et al., 1987). Imprecision in multicriteria decision making (MCDM) can be mod- eled using fuzzy set theory to define criteria and the importance of criteria. According to Bellman and Zadeh ‘‘much of the decision- making in the real world takes place in an environment in which the goals, the constraints, and consequences of possible actions are not known precisely’’ (Bellman & Zadeh, 1970). Ribeiro pro- vides an overview of the concepts and theories of decision making in a fuzzy environment (Ribeiro, 1996). Von Altrock explains the elements of fuzzy logic system design, presenting case studies of real-world applications, of which the most visible applications are in the realms of consumer products, intelligent control, and industrial systems (Von Altrock, 1995). Less visible, but of growing importance, are applications relating to decision support systems (Zimmermann, 1991, 1987). Although fuzzy set theory has been and still remains somewhat controversial, its successes are too clear to be denied. However, Ribeiro warns that ‘‘too much fuzzifi- cation does not imply better modeling of reality, it can be counter- productive’’. Fuzzy ranking methods have been developed that can be used to compare fuzzy numbers (Chen & Hwang, 1992), but this is still an interesting research area. ll rights reserved. rf.bg.ac.yu There are two approaches to MCDM in a fuzzy environment, ‘‘conventional’’ and ‘‘fuzzy’’ (Perny & Roubens, 1998). The conven- tional approach is based on a nonfuzzy decision model, whereas the fuzziness dissolution (defuzzification) is performed at an early stage (Chen & Hwang, 1992; Wu, Tzeng, & Chen, 2009). The fuzzy approach is based on processing fuzzy data for decision making, then dissolving the fuzziness at a later stage (Opricovic, 2007). In both cases, defuzzification is necessary since MCDM results must provide a crisp conclusion. Defuzzification is selection of a specific crisp element based on the output fuzzy set, and it also includes converting fuzzy numbers into crisp scores. There are several defuzzification methods, although the operation defuzzification cannot be defined uniquely (Chen & Cheng, 2005; Detyniecki & Yager, 2000; Lee & Li, 1988; Opricovic & Tzeng, 2003; Yager & Filev, 1994). The multicriteria decision making (MCDM) procedure consists of generating alternatives, establishing criteria, evaluation of alternatives, assessment of criteria weights, and application of a ranking method (Vincke, 1992). The alternatives are evaluates according to different criteria depending on the objectives of the problem. The evaluation of alternatives should be performed according to each criterion from the set of established criteria. A comparative analysis of MCDM methods is presented in several publications (Escobar & Moreno-Jimenez, 2002; Opricovic & Tzeng, 2007; Triantaphyllou, 2000). The VIKOR method has been developed as an MCDM method to solve a discrete multicritea problem with noncommensurable and conflicting criteria (Opricovic, 1998). It focuses on ranking and selecting from a set of alternatives, and determines compromise http://dx.doi.org/10.1016/j.eswa.2011.04.097 mailto:seropric@yahoo.com mailto:seropric@grf.bg.ac.yu http://dx.doi.org/10.1016/j.eswa.2011.04.097 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 12984 S. Opricovic / Expert Systems with Applications 38 (2011) 12983–12990 solutions for a problem with conflicting criteria, which can help the decision makers to reach a final decision. The compromise solution is a feasible solution which is the closest to the ideal (Opricovic & Tzeng, 2004). VIKOR is based on old ideas of compromise program- ming (Duckstein & Opricovic, 1980; Yu, 1973). An extension of VI- KOR to determine fuzzy compromise solution for multicriteria is presented in (Opricovic, 2007). The fuzzy VIKOR method is developed as a fuzzy MCDM method to solve a discrete fuzzy multicritea problem with noncommensu- rable and conflicting criteria. It is presented in Section 2. The back- ground for this method, including aggregation, normalization, DM’s preference assessment, and operations on fuzzy numbers are discussed, as a study of rationality that in someway justifies the fuzzy VIKOR method and shows the position of its background in the literature on MCDM. This new method provides a contribu- tion to the practice of MCDM. In Section 3, a numerical example illustrates an application of fuzzy VIKOR to water resources plan- ning, aiming to numerical justification. Comparisons of the results by different methods are presented in Section 4. 2. The fuzzy VIKOR method The fuzzy VIKOR method has been developed to determine the compromise solution of the fuzzy multicriteria problem mco j ð~f ijðAjÞ; j ¼ 1; . . . ; JÞ; i ¼ 1; . . . ; n n o where: J is the number of feasible alternatives; Aj = {x1, x2 , . . .} is the jth alternative obtained (generated) with certain values of system variables x; fij is the value of the ith criterion function for the alter- native Aj; n is the number of criteria; mco denotes the operator of a multicriteria decision making procedure for selecting the best (compromise) alternative in multicriteria sense. Alternatives can be generated and their feasibility can be tested by mathematical models (determining variables x), physical models, and/or by exper- iments on the existing system or other similar systems. Constraints are seen as high-priority objectives, which must be satisfied in the alternatives generating process. In this paper we assume the alter- natives are evaluated by the triangular fuzzy numbers ~f ij ¼ðlij; mij; rijÞ; i ¼ 1; . . . ; n; j ¼ 1; . . . ; J. The set of criteria repre- senting benefits (good effects) is denoted by Ib, and a set Ic for costs. Here jIb [ Icj = n, where j�j denotes a cardinal number. The ranking algorithm VIKOR has the following steps: (i) Determine the ideal ~f �i ¼ðl � i ; m � i ; r � i Þ and the nadir ~f �i ¼ ðl�i ; m�i ; r � i Þ values of all criterion functions, i = 1, 2, . . . , n. ~f �i ¼ MAX j ~f ij; ~f � i ¼ MIN j ~f ij; for i 2 Ib; ~f �i ¼ MIN j ~f ij; ~f � i ¼ MAX j ~f ij; for i 2 Ic: (ii) Compute normalized fuzzy difference ~dij; j ¼ 1; . . . ; J; i ¼ 1; . . . ; n: ~dij ¼ð~f �i � ~f ijÞ=ðr � i � l � i Þ for i 2 I b ; ~dij ¼ð~f ij � ~f �i Þ=ðr � i � l � i Þ for i 2 I c ð1Þ (iii) Compute eSj ¼ðSlj; Smj ; SrjÞ and eRj ¼ðRlj; Rmj ; RrjÞ; j ¼ 1; 2; . . . ; J, by the relations eSj ¼ Xn i¼1 � ~wi � ~dij � � ð2Þ eRj ¼ MAX i ~wi � ~dij � � ð3Þ where eS is a fuzzy weighted sum, eR is a fuzzy operator MAX (see Appendix B), ~wi are the weights of criteria, expressing the DM’s preference as the relative importance of the criteria. (iv) Compute the values eQ j ¼ðQ lj; Q mj ; Q rjÞ; j ¼ 1; 2; . . . ; J, by the relationeQ j ¼ v eSj �eS�� �=ðS�r � S�lÞ�ð1 � vÞ eRj � eR�� �=ðR�r � R�lÞ ð4Þ where: eS� ¼ MIN j eSj , S�r ¼ maxj Srj ; eR� ¼ MIN j eRj; R�r ¼ maxj Rrj , and v is introduced as a weight for the strategy of ‘‘the majority of criteria’’ (or ‘‘the maximum group utility’’), whereas 1 � v is the weight of the individual regret. These strategies could be compromised by v = 0.5, and here v is modified as v = (n + 1)/ 2n (from v + 0.5(n � 1)/n = 1) since the criterion (1 of n) related to R is included in S, too. The best values of S and R are denoted by eS� and eR�, respectively. (v) ‘‘Core’’ ranking Rank the alternatives by sorting the core values Q mj ; j ¼ 1; 2; . . . ; J, in decreasing order. The obtained ordering is denoted by fAgQ m . (vi) Fuzzy ranking The jth ranking position in fAgQ m of an alternative A (j), j = 1, . . . , J, is confirmed if MIN k2Jj eQ ðkÞ ¼ eQ ðjÞ, where Jj = {j, j + 1, . . . , J} and eQ ðkÞ is the fuzzy merit for the alternative A(k) at the kth position in fAgQ m (see Appendix A). Confirmed ordering represents ‘‘exact’’ fuzzy ranking fAgeQ , although the set fAgeQ could not be com- plete ordering (it may be partially ranking). (vii) Defuzzification of eSj; eRj; eQ j; j ¼ 1; 2; . . . ; J, by the relations Crisp eN� � ¼ð2m þ l þ rÞ=4 ð5Þ Here the defuzzification method ‘‘2nd weighted mean’’ is applied to convert a fuzzy number into crisp score (see Appendix A). (viii) Rank the alternatives, sorting by the crisp values S, R and Q in decreasing order. The results are three ranking lists {A}S, {A}R, {A}Q. (ix) Propose as a compromise solution the alternative (A(1)) which is the best ranked by the measure Q (in {A}Q) if the fol- lowing two conditions are satisfied: C1. ‘‘Acceptable Advantage’’: Adv P DQ where: Adv = [Q(A(2)) � Q(A(1))]/[Q(A(J)) � Q(A(1))] is the advantage rate of the alternative A(1) ranked first, A(2) is the alternative with second position in {A}Q, and the thresh- old DQ = 1/(J � 1). C2. ‘‘Acceptable Stability in decision making’’: The alternative A(1) must also be the best ranked by S or/and R. If one of the conditions is not satisfied, then a set of compromise solutions is proposed, which consists of: – Alternatives A(1) and A(2) if only the condition C2 is not sat- isfied, or – Alternatives A(1), A(2), . . . , A(M) if the condition C1 is not sat- isfied; A(M) is determined by the relation Q(A(M)) � Q(A(1)) < DQ for maximum M (the positions of these alterna- tives are ‘‘in closeness’’). (x) Determine crisp trade-offs, trik = (Diwk)/(Dkwi), k – i, k = 1, . . . , n, where trik is the number of units of the ith criterion evaluated the same as one unit of the kth criterion; Di ¼ r�i � l � i for i 2 Ib; Di ¼ r�i � l � i for i 2 I c, and w ¼ Crispð~wÞ obtained by defuzzification used in step (vii). The index i is given by the VIKOR user. The VIKOR method introduces these trade-offs as a result of normalization used in Eq. (1) for operations in (2) and (3). S. Opricovic / Expert Systems with Applications 38 (2011) 12983–12990 12985 (xi) The decision maker may give a new value of trik, k – i, k = 1, . . . , n if he or she does not agree with computed values in step (x). The new values of weights are computed wk = j(Dkwi- trik)/Dij, k – i, k = 1, . . . , n; wi is the previous value from the step (x). Then, VIKOR performs a new ranking from step (iii) using ~wk ¼ðwk; wk; wkÞ; k ¼ 1; . . . ; n. The trade-offs determined in step (x) could help the decision maker to assess new values, although that task is very difficult. (xii) The VIKOR algorithm ends if the new values are not given in step (xi). The ranking algorithm VIKOR uses fuzzy operations presented in Appendix B. This method focuses on ranking and selecting from a set of alternatives in the presence of conflicting criteria, and on propos- ing compromise solution (one or more). It is assumed that compro- mising is acceptable for conflict resolution, the decision maker (DM) is willing to approve solution that is the closest to the ideal, and the alternatives are evaluated (fuzzy or crisp) according to all established criteria. The VIKOR method is an effective tool in multicriteria decision making. The obtained compromise solution could be accepted by the decision makers because it provides a maximum group utility of the ‘‘majority’’ (represented by min S, Eq. (2)), and a minimum individual regret of the ‘‘opponent’’ (represented by min R, Eq. (3)). The VIKOR algorithm can be performed without interac- tive participation of DM, but the DM is in charge of approving the final solution and his/her preference must be included. The com- promise solutions could be the base for negotiation, involving the decision makers’ preference by criteria weights. The VIKOR meth- od may be incorporated within a DSS for MCDM (Liu & Stewart, 2004). The fundamental issues of the VIKOR method are discussed in the previous articles, studying aggregation and normalization (Opricovic & Tzeng, 2004), DM’s preference assessment (Opricovic & Tzeng, 2004; Opricovic, 2009), and operations on fuzzy numbers (Opricovic, 2007), that in someway justifies the VIKOR method and its rationality, and shows the position of its background in the lit- erature on MCDM. Several papers presented application of the VI- KOR method (Chang & Hsu, 2009; Ou Yang, Shieh, & Tzeng, 2009; Sanayei, Farid Mousavi, & Yazdankhah, 2010; Tong, Chen, & Wang, 2007). In Section 3, a numerical example illustrates an application of fuzzy VIKOR method aiming to numerical justification. 3. Fuzzy VIKOR application to water resources planning Previous studies of the Mlava water resources system, in Serbia, have selected potential dam sites for reservoirs to provide water. In addition, comprehensive analysis was required to resolve conflict- ing technical, social and environmental features. Even if the topo- graphic surveys confirm that the required reservoir capacity is available, a hydrological solution may conflict with environmental, social, and cultural features. The VIKOR method was applied to evaluate alternative systems on the Mlava River. The alternatives were generated by varying two system parameters, dam site and dam height. The following six alternatives were selected for multicriteria optimization. A1. The alternative A1 is the reservoir Vukan with normal level of 215 m.a.s.l. and useful storage of 86 106 m3 could pro- vide 4.08 m3/s (average) for planed regional water supply. The dam site is 1.5 km downstream of the monastery Gornjak, and the implementation would require the removal of the monastery. There will be a loss of agricultural land of 120 ha. A section (1.5 km) of the regional road and parts of local roads will be flooded by the reservoir. A2. Reservoir Vukan with normal level of 205 m.a.s.l. and useful storage of 40 106 m3 would have less social and environ- mental impacts on local areas and could provide 2.87 m3/s for water supply. It requires the removal of the monastery Gornjak. The loss of agricultural land is less than alternative A1. A3. Reservoir Vitman I with normal level of 215 m.a.s.l. could provide 2.97 m3/s. The dam site is 3 km upstream of the monastery Gornjak, but there will be a loss of agricultural land (120 ha). A4. Reservoir Gradac with normal level of 275 m.a.s.l. could pro- vide 2.73 m3/s. The dam site is in the gorge Ribarska, upstream of the Gornjak gorge. There will be an impact on agricultural area in the region of Zagubica (a loss of 300 ha). The area of several households in two villages will be flooded and they have to be removed. A5. System of three reservoirs, Vitman II (205) and Gradac (251) on Mlava, and Dubocica (255) on the tributary, could provide 2.5 m3/s. All three dam sites are upstream of the monastery Gornjak. The loss of agricultural is relatively small since nor- mal levels are lower. A6. System similar to the alternative A5, Vitman III (203), Gradac (251) and Dubocica (255), which could provide 2.74 m3/s. The Vitman III dam site is shortly downstream of Vitman II. The designed reservoir systems are evaluated according to the following criteria: f1. Investment costs (in 106 US$) including dam construction, expropriation of the area occupied by the reservoir, con- struction of new buildings for the households which have to move, and building new roads that will substitute flooded sections. f2. Water supply discharge – yield (m3/s) is the average annual value of discharge from the reservoir system available for regional water supply. The required reservoir capacity has been determined by the ’’sequent peak’’ algorithm for required total water demands. Water supply discharge has been determined by simulation of reservoir system with required capacity using historical hydrological series. Beside this discharge each reservoir has to realize downstream a biological minimum flow. f3. Social impact (%) on urban and agricultural area expressing local regret as percentage of the regret in the alternative with maximum social impact. f4. Impact on the monastery Gornjak is graded by the experts. The worst grade has the alternative that required the removal of monastery. The construction of a dam could have impact on ambient beauty of the Gornjak gorge. The multicriteria task is to minimize the criterion functions f1, f3, and f4, and to maximize function f2. The four criterion functions are expressed in different units and they are noncommensurable. The values of criterion functions are obtained by a comprehensive study of this reservoir system on Mlava river system, and the re- sults are presented in Table 1. The criteria weights ~Wi ¼ð1; 1; 1Þ; i ¼ 1; 2; 3; 4 express equal importance (no preference), and v = 0.625 (see step (iv) in Section 2). The results obtained by the fuzzy VIKOR algorithm are pre- sented in Tables 2 and 3. Preliminary ranking (‘‘core’’) of alterna- tives by the values Qm is A3, A6, A5, A2, A4, A1. ‘‘Exact’’ fuzzy ranking by fuzzy VIKOR is not complete ordering in this example, since the position of A3 is confirmed (see step (vi) in Section 2, and Fig. 1). Table 1 Performance matrix. Criteria Alternatives Name extr. A1 A2 A3 A4 A5 A6 ~f 1 Investment costs (10 6 $) l 38.00 20.00 24.58 44.54 33.33 33.86 Min m 40.01 21.06 25.87 46.89 33.33 33.86 r 48.00 24.00 29.75 56.27 43.33 42.32 ~f 2 Water supply (m 3/s) l 3.26 2.57 2.82 2.46 2.25 2.47 Max m 4.08 2.87 2.97 2.73 2.50 2.74 r 4.08 2.87 2.97 2.73 2.62 2.85 ~f 3 Social impact (%) l 43 6 38 60 6 6 Min m 47 6 42 62 6 6 r 48 6 50 68 6 6 ~f 4 Impact on monastery (grade) Min l m r 10 10 1 0 2 3 Table 2 Results by fuzzy VIKOR. A1 A2 A3 A4 A5 A6 eS Sl 1.535 1.103 0.791 1.727 0.807 0.796 Sm 2.184 1.661 1.420 2.353 1.402 1.385 Sr 2.897 1.935 1.767 2.885 1.843 1.795 Crisp S 2.200 1.590 1.349 2.330 1.363 1.340 eR Rl 1.0 1.0 0.516 0.871 0.350 0.300 Rm 1.0 1.0 0.607 0.903 0.863 0.732 Rr 1.0 1.0 0.710 1.0 1.0 0.880 Crisp R 1.0 1.0 0.610 0.919 0.769 0.661 eQ Ql 0.087 �0.041 �0.393 0.075 �0.478 �0.508 Qm 0.448 0.293 0.010 0.446 0.142 0.067 Qr 1.0 0.715 0.509 0.996 0.687 0.609 Crisp Q 0.495 0.315 0.034 0.491 0.124 0.059 12986 S. Opricovic / Expert Systems with Applications 38 (2011) 12983–12990 Crisp values (by defuzzification) of eSj; eRj; eQ j; j ¼ 1; 2; . . . ; J, are presented in Table 2. Ranking by crisp values are {A}S = A6, A3, A5, A2, A1, A4, {A}R = A3, A6, A5, A4, A1, A2, {A}Q = A3, A6, A5, A2, A4, A1. The compromise solution for final decision is the set {A3, A6, A5} (see step (ix) in Section 2). A 3. Vitman I (215) (advantage 5.4%). A 6. Vitman III (203), Gradac (251), Dubocica (255). A 5. Vitman II (205), Gradac (251), Dubocica (255). Table 3 Ranking by fuzzy VIKOR. Ordering 1 2 3 4 5 6 ‘‘Core’’ ranking fAgQ m A3 A6 A5 A2 A4 A1 ‘‘Exact’’ fuzzy ranking fAgeQ A6 A5 A2 A4 A1 Defuzzification Q A3 A6 A5 A2 A4 A1 S A6 A3 A5 A2 A1 A4 R A3 A6 A5 A4 A1 A2 0 0.2 0.4 0.6 0.8 1 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 A1 A2 A3 A4 A5 A6 Fig. 1. Fuzzy merit eQðAjÞ and crisp Qj, j = 1, 2, . . . , J. The trade-offs values determined by VIKOR (the step x) are pre- sented in Table 4, showing how many 106$ are evaluated as one unit of kth criterion. The determined tradeoffs are: 19.82 106$/ (m3/s), 0.58 106$/% and 3.63 106$/mark-unit, for example, 1 m3/s of water supply discharge worth as 19.82 106$ of invest- ment costs, and removing monastery Gornjak worth as 25.41 106$. This values seem too high in economic sense, although assessing trade-offs between economic and qualitative criteria is a very difficult task. The trade-offs determined by VIKOR are the result of normalizing noncommensurable criteria. The new trade-offs are given in Table 4. New weights are determined and ranking by VIKOR has been repeated. The results obtained by the fuzzy VIKOR algorithm with given tradeoffs in Table 4 are presented in Tables 5 and 6. New weights in Table 4 are determined by the procedure presented in step (xi) in Section 2. For example, w5 = j(D5w2tr25)/D2j = ((4.08 � 2.25) � 1 � 15)/(56.27 � 20.0) = 0.757, w2 is the previous value from the step (x). Preliminary ranking (‘‘core’’) of alternatives by the values Qm is A3, A2, A6, A5, A1, A4. ‘‘Exact’’ fuzzy ranking by fuzzy VIKOR is not complete ordering in this example, although the positions of five alternatives are confirmed (see step (vi) in Section 2). The position of A2 is not confirmed since Ql(A2) = �0.171 is greater then Ql(A5) = �0.290 (Fig. 2). Table 4 Trade-offs by VIKOR. ~1 ~2 ~3 ~4 tr1k; k ¼ 1; . . . ; n ð10 6 $=~kÞ 1.0 19.82 0.58 3.63 New given trade-offs tr1k, k = 1, . . . , n 1 15 0.2 2 New weights 1.0 0.757 0.342 0.551 Table 5 Results by fuzzy VIKOR with given trade-offs. A1 A2 A3 A4 A5 A6 eS Sl 0.802 0.602 0.368 1.083 0.632 0.607 Sm 1.300 1.052 0.845 1.579 1.102 1.073 Sr 1.894 1.286 1.088 2.012 1.510 1.447 Crisp S 1.324 0.998 0.786 1.563 1.087 1.050 eR Rl 0.551 0.551 0.176 0.566 0.265 0.272 Rm 0.551 0.551 0.459 0.712 0.653 0.554 Rr 0.772 0.624 0.521 1.000 0.757 0.666 Crisp R 0.607 0.570 0.404 0.748 0.582 0.512 eQ Ql �0.095 �0.171 �0.431 0.019 �0.290 �0.296 Qm 0.215 0.121 0.0 0.394 0.186 0.130 Qr 0.851 0.553 0.431 1.000 0.699 0.633 Crisp Q 0.297 0.156 0.0 0.452 0.195 0.149 Table 6 Ranking by fuzzy VIKOR with given trade-offs. Ordering 1 2 3 4 5 6 ‘‘Core’’ ranking fAgQ m A3 A2 A6 A5 A1 A4 ‘‘Exact’’ fuzzy ranking fAgeQ A3 A6 A5 A1 A4 Defuzzification Q A3 A6 A2 A5 A1 A4 S A3 A2 A6 A5 A1 A4 R A3 A6 A2 A5 A1 A4 0 0.2 0.4 0.6 0.8 1 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 A1 A2 A3 A4 A5 A6 Fig. 2. New fuzzy merit eQðAjÞ; j ¼ 1; 2; . . . ; J for given trade-offs. Table 7 Defuzified performance matrix. A1 A2 A3 A4 A5 A6 f1 41.50 21.53 26.52 48.65 35.83 35.97 f2 3.87 2.79 2.93 2.66 2.47 2.70 f3 46.25 6 43 63 6 6 f4 10 10 1 0 2 3 S. Opricovic / Expert Systems with Applications 38 (2011) 12983–12990 12987 Crisp values (by defuzzification) of eSj; eRj; eQ j; j ¼ 1; 2; . . . ; J, are presented in Table 5. Ranking by crisp values are is {A}Q = A3, A6, A2, A5, A1, A4. The compromise solution for final decision is A3 Vitman I (215) with the advantage rate of 33%. The alternative A3 is a single reservoir system (Vitman I (215)). Second ranked is the alternative A6, a three-reservoir system (Vitman III (203), Gradac (251), Dubocica (255)). The decision makers for the Mlava project may adopt alternative A6, which could be developed in three phases, building one reservoir in each phase. 4. Some comparisons Comparisons of the results by different methods are made in this section. These comparisons could challenge the readers to compare fuzzy VIKOR with particular methods. 4.1. Non-fuzzy model Non-fuzzy MCDM methods are compared many times (Triantaphyllou, 2000; Opricovic & Tzeng, 2007). Original (non- fuzzy) VIKOR uses crisp input data. The input data for VIKOR are core (m) values from Table 1. The result from (Opricovic, 2009) is the ranking list: A3, A6, A5, A2, A1, A4, and the set of compromise solutions consists of A 3. Vitman I (215) (advantage 19%). A 6. Vitman III (203), Gradac (251), Dubocica (255). Preliminary ranking (‘‘core’’) of alternatives by the values Qm is A3, A6, A5, A2, A4, A1 (Table 3), the set of compromise solutions is {A3, A6}. Crisp values (by defuzzification) of eSj; eRj; eQ j; j ¼ 1; 2; . . . ; J, are presented in Table 2. Ranking by defuzzified values ofeQ j is {A}Q = A3, A6, A5, A2, A4, A1, and the set of compromise solu- tions is A 3. Vitman I (215) (advantage 5.4%). A 6. Vitman III (203), Gradac (251), Dubocica (255) . A 5. Vitman II (205), Gradac (251), Dubocica (255). The results are similar because of small spread (support) of fuz- zy numbers r � l in Table 1. 4.2. Predefuzzification Two approaches to fuzzy multicriteria decision making are pre- sented, ‘‘fuzzy’’ and ‘‘conventional’’. The fuzzy VIKOR method is a fuzzy approach. The conventional approach is based on a nonfuzzy decision model, here VIKOR method, whereas the fuzziness disso- lution (defuzzification) is performed at an early stage (predefuzz- ification). The predefuzzification approach is used in (Wu et al., 2009), utilizing the COA (center of area) method to find ‘‘the Best Nonfuzzy Performance value (BNP)’’. The ranking of the alterna- tives then proceeds based on the value of the derived BNP for each of the alternatives. Three MCDM analytical tools of SAW, TOPSIS, and VIKOR were adopted to rank alternatives. The same approach is used in (Wu, Chen, & Chen, 2010), adopting VIKOR to rank alternatives. The fuzzy numbers from Table 1 are defuzified by the procedure used in step (vii), Section 2, and discussed in Appendix A. The defuzzified performance matrix is presented in Table 7. The results in Table 7 are used as input data for non-fuzzy VI- KOR. The non-fuzzy VIKOR is published in Opricovic and Tzeng (2007). The ranking result is A3, A6, A5, A2, A1, A4 and the compro- mise solution is A3 (advantage 21.9%). The non-fuzzy VIKOR results (Opricovic, 2009) with core (m) values from Table 1 as input data are: ranking list: A3, A6, A5, A2, A1, A4, and the set of compromise solutions consists of A3 (advantage 19%) and A6. 12988 S. Opricovic / Expert Systems with Applications 38 (2011) 12983–12990 These two results are very close, since the threshold for advan- tage in this example is 20% (see step (ix) in Section 2). Main reason is small spread (support) of fuzzy numbers r � l in Table 1. 4.3. Ranking fuzzy numbers Ranking fuzzy numbers using a-weighted valuations is consid- ered in (Detyniecki & Yager, 2000). Fuzzy numbers eN 1 ¼ð1; 7; 9Þ; eN 2 ¼ð4; 6; 12Þ, and eN 1 ¼ð1; 7; 9Þ; eN 3 ¼ð2; 4; 10Þ are com- pared. The result is eN 1 < eN 2 for ‘‘pro-support’’ (0 6 q < 2) andeN 1 > eN 2 for q > 2 (importance of high a–levels). Analogous result is for eN 1 and eN 3. Applying fuzzy VIKOR the result is as follows: ‘‘Core’’ rankingeN 1; eN 2; eN 3; ‘‘Exact’’ ranking eN 2 > eN 3; Ranking by crisp values (defuzzification) eN 2ðQ ¼ 0:026Þ; eN 1ðQ ¼ 079Þ; eN 3ðQ ¼ 0:132Þ. ‘‘Exact’’ ranking provides consistent result. Ordering eN 1 and eN 2 is an inconsistent result. Crisp values by the center-of-gravity (k = 1, Appendix A) are N1 = 5.667, N2 = 7.333, N3 = 5.333, and by the CFCS method (Opricovic & Tzeng, 2003) N1 = 6.29, N2 = 6.71, N3 = 4.88. An interesting example is ordering fuzzy numbers eN 1 ¼ð5; 6; 13Þ; eN 2 ¼ð3; 7; 13Þ; eN 3 ¼ð5; 8; 9Þ; eN 4 ¼ð2; 9; 10Þ, (Detyniecki & Yager, 2000). Applying fuzzy VIKOR there is no ordering by ‘‘exact’ fuzzy ranking; and ranking by crisp values (defuzzification)eN 1 ¼ eN 2 ¼ eN 3 ¼ eN 4. Crisp values by the center-of-gravity are N1 = 8, N2 = 7.667, N3 = 7.333, N4 = 7, and by the CFCS method (Opricovic & Tzeng, 2003) N1 = 7.128, N2 = 7.366, N3 = 7.574. N4 = 7.872. These are examples of inconsistent ranking. The explanation of an inconsistent result is that low a-levels are compensated with the high a-levels (Detyniecki & Yager, 2000). 4.4. NFWA and fuzzy VIKOR The NFWA method (new fuzzy-weighted average) is applied and an example is presented in Vanegas and Labib (2001). Fuzzy result by NFWA is eDðA1Þ¼ ð0:15; 0:32; 0:58Þ; eDðA2Þ¼ ð0:37; 0:61; 0:85Þ; eDðA3Þ¼ ð0:15; 0:38; 0:61Þ, and ranking result (crisp) is A2(D = 0.61), A3(0.38), A1(0.35). The same problem is solved by the fuzzy VIKOR algorithm and the result is as follows: eQðA1Þ¼ ð�0:54; 0:28; 0:98Þ; eQðA2Þ¼ ð�0:75; 0:0; 0:75Þ; eQðA3Þ¼ ð�0:64; 0:15; 0:93Þ, ‘‘Exact’’ fuzzy rank- ing A2, A3, A1 (complete ranking); and ranking by crisp values (defuzzification) A2(Q = 0.0), A3(0.15), A1(0.25). Ranking results by these two methods are very close, and ordering is the same. 4.5. Fuzzy AHP and fuzzy VIKOR An example of fuzzy multicriteria problem is presented in (Gu & Zhu, 2006). Comparison results between the proposed algorithm and other algorithms are presented. There is a conclusion ‘‘It is apparent that the proposed improving fuzzy AHP algorithm based on fuzzy eigenvector of fuzzy attribute evaluation space is more efficient than others. It has good objectivity and resolution.’’ Fuzzy result by the improved fuzzy AHP algorithm is fWðA1Þ¼ ð0:3375;0:8195;1Þ; fWðA2Þ¼ð0:3164;0:6740;1Þ; fWðA3Þ¼ð0:3770; 0:8941; 1Þ; fWðA4Þ¼ ð0:3387; 0:7491; 1Þ, and ranking result (crisp) is A3(W = 0.791), A1(0.744), A4(0.709), A2(0.666). The same problem is solved by the fuzzy VIKOR algorithm and the result is as follows. eQðA1Þ¼ ð�0:455; 0:051; 0:669Þ; eQðA2Þ¼ ð�0:396; 0:193; 1:0Þ; eQðA3Þ¼ ð�0:491; 0:0; 0:5Þ; eQðA4Þ¼ ð�0:474; 0:110; 0:784Þ. ‘‘Exact’’ fuzzy ranking is A3, A4, A2; and ranking by crisp values (defuzzification) A3(Q = 0.002), A1(0.079), A4(0.132), A2(0.248). Ranking results by these two methods are very close, and ordering is the same. 4.6. ‘‘Distance’’ method and fuzzy VIKOR The procedure FMCGDSS (fuzzy multi-criteria group decision support system) based on metric distance method is presented and applied in (Chen & Cheng, 2005). The ranking results are the following: A3(fuzzy mean = 3.662, fuzzy spread = 0.881), A2(3.577, 0.97), A1(3.477, 0.952). The same problem is solved by the fuzzy VIKOR algorithm and the result is as follows: eQðA1Þ¼ ð�0:86; 0:088; 0:946Þ; eQðA2Þ¼ ð�0:863; 0:054; 0:954Þ; eQðA3Þ¼ ð�0:871; 0:0; 0:877Þ, ‘‘Exact’’ fuzzy ranking is A3, A1; and ranking by crisp values (defuzzification) A3(Q = 0.002), A2(0.05), A1(0.065). Ordering by these two methods is the same. 5. Conclusions The fuzzy VIKOR method focuses on ranking and selecting from a set of alternatives in a fuzzy environment. Imprecision in multi- criteria decision making is modeled using fuzzy set theory to de- fine criteria and the importance of criteria (weights). The triangular fuzzy numbers are used to handle imprecise numerical quantities. The VIKOR method is based on the aggregating fuzzy merit eQ that represents distance of an alternative to the ideal solu- tion. The fuzzy operations and procedures for ranking fuzzy num- bers are used in developing VIKOR algorithm. A numerical example illustrates an application of the fuzzy VI- KOR method to water resources planning, aiming to numerical jus- tification. It is an intention to illustrate the conceptual and operational validation of the application of this method in real world problem. The fuzzy VIKOR method background and compar- isons of the results by different methods are presented in order to show the position of this new method in the literature on fuzzy MCDM. Researchers are challenged to provide a guide for choosing the method that is both theoretically well founded and practically operational to solve actual problems. Acknowledgments This paper is a result of the project 144035A ‘‘Resolving Com- plexity, Fuzziness, Uncertainty, Conflict’’ which is funded by the Ministry of Science of Serbia. The constructive comments of the editor and the reviewers are gratefully acknowledged. Appendix A A.1. Ranking fuzzy numbers and defuzzification MCDM in a fuzzy environment requires the comparison of fuzzy numbers. The problem of comparing fuzzy numbers has been stud- ied and appears to be an important and difficult problem. A fuzzy number is characterized by its shape, spread, height, and relative location on the x-axis. A good ranking method would be one that takes into account all these factors. Since a fuzzy number repre- sents many possible real numbers that have different membership values, one will face a difficult problem of comparing two different fuzzy numbers. Over 20 ranking methods for fuzzy numbers have been proposed, but none of these existing methods is perfect (Chen & Hwang, 1992; Detyniecki & Yager, 2000; Gu & Zhu, 2006; Lai & Hwang, 1994; Lee & Li, 1988; Yager, 1981). In general, two ap- proaches are used: (1) comparison of fuzzy numbers and (2) con- verting fuzzy number into crisp score (deffuzzification). The fuzzy VIKOR algorithm in step (vi) uses a ranking procedure with consistent results providing complete ranking only if fuzzy numbers have separated membership functions S. Opricovic / Expert Systems with Applications 38 (2011) 12983–12990 12989 (no cross-overlapping). If there is cross-overlapping, this proce- dure does not provide complete ordering. Fuzzy operator MIN is used for ranking and to confirm ranking (domination). An alternative Aj is better ranked than Ak if eQ j ¼ MIN eQ j; eQ k� �, or Q lj 6 Q l k; Q m j < Q m k ; Q r j 6 Q r k. The fuzzy VIKOR algorithm uses a defuzzification procedure in step (vii) to convert fuzzy numbers into real (crisp) numbers. Then, in step (viii), the ranking of fuzzy numbers is performed through the comparison of the corresponding real numbers. The kth weighted mean method has been developed to be used as defuzzification procedure in this paper. It uses membership function to the power of k as a weighted factor. The crisp value CrispðeNÞ for the triangular fuzzy number eN ¼ðl; m; rÞ is determined by the following formula CrispðeNÞ¼ Z r l xlkðxÞdx �Z r l lkðxÞdx Integrating the integrals the following formula is obtained: CrispðeNÞ¼ ðkm þ l þ rÞ=ðk þ 2Þ or C ¼ m þðsr � slÞ=ðk þ 2Þ and lðCÞ¼ kþ1 kþ2 þ sr ðkþ2Þsl ; C 6 m kþ1 kþ2 þ sl ðkþ2Þsr ; C P m ( where C ¼ Crisp eN� �; sl ¼ m � l and sr = r � m are left and right sup- port (spread), respectively. The parameter (power) k has the impact on defuzzification re- sult as follows: k ¼ 1 : C ¼ðm þ l þ rÞ=3 or C ¼ m þðsr � slÞ=3 and lðCÞ P 2=3 Increasing k (k = 2, 3, . . .): C moves toward m (core) and membership l(C) increases; for example, for k = 4: C = (m + (sr � sl)/6); and, limk?1C(k) = m, limk?1l(C,k) = 1. The Centroid (Center-of-gravity) method, which provides a crisp value based on the center-of-gravity of the fuzzy set could be considered as a special case for k = 1. With- in multicriteria decision making, k P 4 could be preferred by a ‘‘risk aversion’’ decision maker (increasing membership l(C)), this is ‘‘pro-core’’ defuzzification. A ‘‘gambler’’ decision maker could have different preference (Yu, 1990). A general suggestion could be to use one of the values {2, 3, 4} for power k, and the value of power k should be the same for defuzzifying all fuzzy numbers within a study. The fuzzy VIKOR algorithm in step (vii) uses the 2nd weighted mean as a practical defuzzification tool for converting a fuzzy num- ber into crisp number. A weighted factor include the membership function l(x) that denotes the degree of truth that the fuzzy value is equal to x within the real interval [l, r]. The greater the value of l(x), the higher the confidence in the value of x. The wider the sup- port of the membership function, the higher the fuzziness (impre- ciseness, uncertainty). Appendix B B.1. Operations on triangular fuzzy numbers To express an imprecise value, as ‘‘about m’’ (‘‘approximately m’’), the triangular fuzzy number (TFN) eN ¼ðl; m; rÞ is used, associ- ated with the membership triangular function defined as follows: leNðxÞ¼ ðx � lÞ=ðm � lÞ; x 6 m ðr � xÞ=ðr � mÞ; x P m 0; x R ½l; r 8>< >: The membership function l(x) denotes the degree of truth that the fuzzy value is equal to x within the real interval [l, r]. The fuzzy number eN has the core m with l(m) = 1 and the support [l, r]. The fuzzy VIKOR method has been developed applying mathe- matical operations on TFNs defined as follows Summation : Pn i¼1 � eN i ¼ Pn i¼1 li; Pn i¼1 mi; Pn i¼1 ri � � Scalar summation : eN � K ¼ðl þ K; m þ K; r þ KÞ Subtraction : eN 1 � eN 2 ¼ðl1 � r2; m1 � m2; r1 � l2Þ Scalar subtraction : eN � K ¼ðl � K; m � K; r � KÞ Scalar multiplication : K eN ¼ðK l; K m; K rÞ; for K P 0 Multiplication : eN 1 � eN 2 ¼ðl1 l2; m1 m2; r1 r2Þ; for l1 P 0ðpositiveeN 1Þ Scalar division : eN=K ¼ðl=K; m=K; r=KÞ; for K > 0 Operator MAX : MAX i eN i ¼ðmax i li; max i mi; max i riÞ Operator MIN : MIN i eN i ¼ðmin i li; min i mi; min i riÞ The result of summation or subtraction on TFNs is TFN. The re- sult of fuzzy multiplication is considered as an approximation of TFN, especially when applying a-cut (Giachetti & Young, 1997a). The result of MAX eN 1; eN 2� � (or MIN) is not a TFN only if eN 1 andeN 2 are overlapping in two cases: 1. l1 < l2, m1 – m2, r1 > r2; and 2. l1 < l2, m1 > m2, r1 < r2; in these cases TFN is used as an approxima- tion. Definitions and characteristics of the above operations are discussed in several articles (Chiu & Wang, 2002; Giachetti & Young, 1997b; Klir & Yuan, 1995). References Bellman, R. E., & Zadeh, L. A. (1970). Decision-making in a fuzzy environment. Management Science, 17, 141–164. Chang, C.-L., & Hsu, C.-H. (2009). Multi-criteria analysis via the VIKOR method for prioritizing land-use restraint strategies in the Tseng–Wen reservoir watershed. Journal of Environmental Management, 90(11), 3226–3230. Chen, L. S., & Cheng, C. H. (2005). Selecting IS personnel use fuzzy GDSS based on metric distance method. European Journal of Operational Research, 160(3), 803–820. Chen, S. J., & Hwang, C. L. (1992). Fuzzy multiple attribute decision making: Methods and applications. Berlin: Springer-Verlag. Chiu, C. H., & Wang, W. J. (2002). A simple computation of MIN and MAX operations for fuzzy numbers. Fuzzy Sets and Systems, 126, 273–276. Detyniecki, M., & Yager, R. (2000). Ranking fuzzy numbers using a-weighted valuations. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 8(5), 573–591. Duckstein, L., & Opricovic, S. (1980). Multiobjective optimization in river basin development. Water Resources Research, 16(1), 14–20. Escobar, M. T., & Moreno-Jimenez, J. M. (2002). A linkage between the analytic hierarchy process and the compromise programming models. Omega, 30, 359–365. Giachetti, R. E., & Young, R. E. (1997a). Analysis of the error in the standard approximation used for multiplication of triangular and trapezoidal fuzzy numbers and the development of a new approximation. Fuzzy Sets and Systems, 91, 1–13. Giachetti, R. E., & Young, R. E. (1997b). A parametric representation of fuzzy numbers and their arithmetic operators. Fuzzy Sets and Systems, 91, 185–202. Gu, X., & Zhu, Q. (2006). Fuzzy multi-attribute decision-making method based on eigenvector of fuzzy attribute evaluation space. Decision Support Systems, 41, 400–410. Klir, G. J., & Yuan, B. (1995). Fuzzy sets and fuzzy logic: Theory and applications. Englewood Cliffs: Prentice-Hall. Lai, Y. J., & Hwang, C. L. (1994). Fuzzy multiple objective decision making: Methods and applications. Berlin: Springer-Verlag. 12990 S. Opricovic / Expert Systems with Applications 38 (2011) 12983–12990 Lee, E. S., & Li, R. L. (1988). Comparison of fuzzy numbers based on the probability measure of fuzzy events. Computers and Mathematics with Applications, 15, 887–896. Liu, D., & Stewart, T. J. (2004). Integrated object-oriented framework for MCDM and DSS modeling. Decision Support Systems, 38, 421–434. Opricovic, S. (1998). Multicriteria optimization of civil engineering systems (in Serbian, Visekriterijumska optimizacija sistema u gradjevinarstvu). Belgrade: Faculty of Civil Engineering. Opricovic, S. (2007). A fuzzy compromise solution for multicriteria problems. International Journal of Uncertainty, Fuzziness and Knowledge-based Systems, 15(3), 363–380. Opricovic, S. (2009). A compromise solution in water resources planning. Water Resources Management, 23(8), 1549–1561. Opricovic, S., & Tzeng, G. H. (2003). Defuzzification within a multicriteria decision model. International Journal of Uncertainty, Fuzziness And Knowledge-Based Systems, 11(5), 635–652. Opricovic, S., & Tzeng, G. H. (2004). The compromise solution by MCDM methods: A comparative analysis of VIKOR and TOPSIS. European Journal of Operational Research, 156(2), 445–455. Opricovic, S., & Tzeng, G. H. (2007). Extended VIKOR method in comparison with outranking methods. European Journal of Operational Research, 178(2), 514–529. Ou Yang, Y. P., Shieh, H. M., & Tzeng, G. H. (2009). A VIKOR technique with applications based on DEMATEL and ANP. Communications in Computer and Information Science, 35, 780–788. Perny, P., & Roubens, M. (1998). Fuzzy preference modeling. In R. Slowinski (Ed.), Fuzzy Sets in Decision Analysis, Operations Research and Statistics (pp. 3–30). Boston: Kluwer Academic Publishers. Ribeiro, R. A. (1996). Fuzzy multiple attribute decision making: A review and new preference elicitation techniques. Fuzzy Sets and Systems, 78, 155–181. Sanayei, A., Farid Mousavi, S., & Yazdankhah, A. (2010). Group decision making process for supplier selection with VIKOR under fuzzy environment. Expert Systems with Applications, 37(1), 24–30. Tong, L., Chen, C. C., & Wang, C. H. (2007). Optimization of multi-response processes using the VIKOR method. International Journal of Advanced Manufacturing Technology, 31(11–12), 1049–1057. Triantaphyllou, E. (2000). Multi-criteria decision making methods: A comparative study. Dordrecht: Kluwer Academic Publishers. Vanegas, L. V., & Labib, A. W. (2001). Application of new fuzzy-weighted average (NFWA) method to engineering design evaluation. International Journal of Production Research, 39(6), 1147–1162. Vincke, P. (1992). Multicriteria decision-aid. New York: John Wiley & Sons. Von Altrock, C. (1995). Fuzzy logic & neuro-fuzzy applications explained. Englewood Cliffs: Prentice Hall. Wu, H. Y., Chen, J. K., & Chen, I. S. (2010). Innovation capital indicator assessment of Taiwanese Universities: A hybrid fuzzy model application. Expert Systems with Applications, 37, 1635–1642. Wu, H. Y., Tzeng, G. H., & Chen, Y. H. (2009). A fuzzy MCDM approach for evaluating banking performance based on balanced scorecard. Expert Systems with Applications, 36, 10135–10147. Yager, R. R. (1981). A procedure for ordering fuzzy subsets of the unit interval. Information Sciences, 24, 143–161. Yager, R. R., & Filev, D. P. (1994). Essentials of fuzzy modeling and control. New York: John Wiley & Sons, Inc. Yu, P. L. (1973). A class of solutions for group decision problems. Management Science, 19(8), 936–946. Yu, P. L. (1990). Forming winning strategies: An integrated theory of habitual domains. Heidelberg: Springer-Verlag. Zadeh, L. A. (1987). The concept of a linguistic variable and its application to approximate reasoning. In R. R. Yager et al. (Eds.), Fuzzy sets and applications: Selected papers by L.A. Zadeh (pp. 219–269). New York: John Wiley & Sons. Zimmermann, H. J. (1987). Fuzzy sets, decision making, and expert systems. Boston: Kluwer Academic Publishers. Zimmermann, H. J. (1991). Fuzzy set theory and its applications. Boston: Kluwer Academic Publishers. Fuzzy VIKOR with an application to water resources planning 1 Introduction 2 The fuzzy VIKOR method 3 Fuzzy VIKOR application to water resources planning 4 Some comparisons 4.1 Non-fuzzy model 4.2 Predefuzzification 4.3 Ranking fuzzy numbers 4.4 NFWA and fuzzy VIKOR 4.5 Fuzzy AHP and fuzzy VIKOR 4.6 “Distance” method and fuzzy VIKOR 5 Conclusions Acknowledgments Appendix A A.1 Ranking fuzzy numbers and defuzzification Appendix B B.1 Operations on triangular fuzzy numbers References 
work_6uttfqrwwzazriitygqfm22fii	----	Implementation of an Intelligent Course Advisory Expert System (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 3, No.5, 2014 6 | P a g e www.ijarai.thesai.org Implementation of an Intelligent Course Advisory Expert System Cased-Based Course Advisory Expert System Olawande Daramola, Onyeka Emebo, Ibukun Afolabi, Charles Ayo Department of computer and information sciences, Covenant University, Nigeria Abstract—Academic advising of students is an expert task that requires a lot of time, and intellectual investments from the human agent saddled with such a responsibility. In addition, good quality academic advising is subject to availability of experienced and committed personnel to undertake the task. However, there are instances when there is paucity of capable human adviser, or where qualified persons are not readily available because of other pressing commitments, which will make system-based decision support desirable and useful. In this work, we present the design and implementation of an intelligent Course Advisory Expert System (CAES) that uses a combination of rule based reasoning (RBR) and case based reasoning (CBR) to recommend courses that a student should register in a specific semester, by making recommendation based on the student’s academic history. The evaluation of CAES yielded satisfactory performance in terms of credibility of its recommendations and usability. Keywords—Academic advising; expert system; case-based reasoning; JESS; rule-based reasoning; evaluation I. INTRODUCTION The quality of academic advising received by a student is crucial to the overall performance of the student. Good advising yields a good outcome while bad advising will be frustrating and have a damaging effect on students’ progress. However, a staff advisor needs to keep up with the academic history of advisees in order to be an effective guide. Academic advising requires a lot of patience, commitment and ingenuity, which does not always exist, because humans have their limitations. In many scenarios, the rules for guiding students may change from time to time due to curriculum reviews, changes in course structure, or the circumstances of specific students. This makes it necessary for the human advisor to be adept in all the nuances of academic advising at all times. In many academic departments, the roles assigned to staff may change periodically, making it compulsory for the staff concerned to learn new rules that pertain to advising a new set of students. In addition, academic advising is time-consuming and mentally exacting, requiring the application of psychological and people management skills. All of these present a complex scenario that requires good decision- making, which places huge responsibility on the human advisor. Therefore, there is a need to alleviate the drudgery associated with academic advising by using expert systems to aid decision-making. The use of an expert system will ensure the automation a significant part of the advisory process in a way that allow humans to do what they can do best, while the system complements human expertise by doing what it can do best, thereby creating a synergy that benefits both staff and students. Hence, the essence of a course advisory expert system is not to replace the human advisor, but to minimize the cognitive load and the time expended by the human advisor on academic advising, and to improve the quality of academic advising. Course advising involves an academic staff giving counsel to a student on the courses to register in a semester in order to satisfy established academic requirements that pertain to the student’s academic programme. Students in a University are generally expected to satisfy some performance criteria in order to progress from one level to another, with a specified number of credit units to be passed among a set of compulsory (core), electives, and optional courses. The role of the human course adviser is to ensure that a student makes good decisions on courses that should be registered relative to the student’s current level and academic history in order to satisfy the graduation requirements. The course advisory task is a domain for the application of expert system because – it is based on the use of domain specific knowledge, uses voluminous data, it is difficult to characterize accurately, the curricular do change from time to time, and decisions have to be made based on the specific rules of the University concerned. A lot of the decisions made by a human advisor during the process of advising a student are based on reasoning drawn from previous episodes and experiences that the advisor had gained over time, and known rules of the University that relates to course registration. This suggests that a model of expert system that uses Case Based Reasoning (CBR) and rule-based reasoning for decision-making would be viable for academic advising. CBR is a pattern-based problem solving paradigm that relies on knowledge gained from previous episodes to resolve new problems whenever sufficient similarity can be established between the current case (problem) and past cases that are stored in the case base (repository) [1]. The attraction for using CBR as the mode of reasoning for academic advising is because many similarities exist in the nature of academic problems and concerns that students’ have in the process of course registration. Hence, the combination of CBR and rule- based reasoning – which enables the consideration of specific university rules for decision making – in order to develop an expert system for student advising. A case based approach will seek to emulate human expertise to a reasonable extent, by (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 3, No.5, 2014 7 | P a g e www.ijarai.thesai.org drawing on the similarity that exist in the experiences gained in previous cases of course advising, and an awareness of relevant university rules. This paper describes the implementation of an intelligent course advisory expert system (CAES). The expert system uses the combination of rule-based reasoning and case-based reasoning to generate credible recommendation to guide students on courses to register. The objective of the system is to reduce the effort, and time used in the process of student advising, and to improve quality. The remaining part of the paper is described as follows. In section 2, we present related work. Section 3 discusses the course registration process and the requirements for a Course Advisory Expert System (CAES). Section 4 gives a description of the architecture of the CAES and the process of applying the CAES. Section 5 gives an outline of algorithms that enable some of the core functionalities of the CAES. Section 6 reports a case study of the application of the CAES in a tertiary institution. Section 7 is a preliminary evaluation of the CAES, while the result and discussion was presented in Section 8. The paper is concluded in section 9 with a brief comments and outlook of work for the future. II. RELATED WORK The desire for technology-supported academic advising has been around for a while, and a number of efforts have been reported in the literature. In [2], the evaluation of a Web- based decision support tool that aids student advising was reported. The evaluation of the tool showed that large percentage of respondents regard it as effective and efficient for academic advising, however, the details of its implementation was not provided in the paper. In [3], the design of an i-Counselling system that combines ontology- based information retrieval and optimization-based search technology to provide relevant answers to queries posed by new and current students was reported. The academic advising module of the system is able to answer questions from current students on specific programmes, study plans and graduation requirements. The system was adjudged effective after an evaluation was conducted. The HE-Advisor [4], is a multidisciplinary Web-based higher education advisory system that offers academic advisory services in order to help students make the best decision in selecting a degree to study. It also incorporates guidance on course registration to assist students to stay on the right path towards concluding their degree, information on graduation requirements and statistics for timetable planners were also provided by the system. The ViCurriAs [5] is a visual tool that facilitates the registering of new curriculum plans and track the progress of students enrolled for a degree programme. Other types of expert systems or hybrid intelligent systems that have been used for academic advising include [6, 7, 8]. In addition, in [9], the concepts of intelligent agents and semantic web were used to develop an academic advisory system. The domain knowledge was modeled by using the OWL ontology language, while the agents reason on stored domain knowledge by using an inference engine. The work in [10] presents the architectural framework of an intelligent advisory system that uses the concepts of object-orientation and knowledgebase rules for academic advising. The objective of the system is to help students to know what to do and how to do it. In [11] the Interactive Virtual Expert System for Advising (InVEStA) was reported. InVEStA was designed to assist undergraduate students and their advisors in providing timely, accurate and conflict-free schedules. The system was implemented using Java and an object-relational database. It comprises a Database Layer, Transaction Layer, Scheduler and the web-based Front-End. The Graduate Course Advisor (GCA) is a rule-based expert system that advises graduate students of computer science [12]. The GCA is a Prolog-based system that was modelled after MYCIN. GCA divides advising into four phases such that each phase may apply the inference engine to its own rule base and invoke other procedures. The CBR Recommender for Academic Advising (AACORN) was presented in [13]. AACORN uses course histories to generate recommendations for course advising. By reusing the knowledge embedded in a student’s academic history as captured in student's transcripts, AACORN is able to make reasonable suggestions with a limited amount of domain knowledge. The edit distance was used to determine the similarity between the course history of a new student and other course histories in the case base. The intelligent Course Advisory Expert System (CAES) presented in this work differs from other course advisory system because it integrates the use of CBR and rule-based reasoning to generate intelligent recommendations for students on courses to register. The merit of the CAES when compared to many of the previous approaches is the relatively cheap cost of knowledge acquisition and representation. A CBR system like the CAES is able to acquire new knowledge as usage of the system increases, while its rules can also be modified with minimal effort. This is unlike when an ontology is used for knowledge representation, which although, quite effective, require an advance investment in quality ontology development before efficient course advising can be obtained. Hence, as a contribution, this work offers a cheaper but cost effective way for implementing expert-based academic course advising. In sequel section, we shall discuss the architecture of system in more detail. III. AN OVERVIEW OF THE COURSE REGISTRATION PROCESS The procedure for course registration by a student entails a series of activities. The procedure includes: 1) authenticate the status of student to determine if student qualify to be registered into a particular level based on previous academic performance; 2) select a course to be added to the list of registered courses by student; 3) add or drop a course after initial registration; validate course prerequisites; and 4) check the rules that guides total numbers of course to register and the combination of courses to register. (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 3, No.5, 2014 8 | P a g e www.ijarai.thesai.org It is expected that for a healthy process both the student and the advisor must have adequate understanding of the procedure in order to avoid violations. Fig. 1 shows the key uses cases that pertain to a course registration scenario. A more detailed analysis of the use cases captured in Fig. 1, revealed a number of specific requirements that a course advisory system must meet. These include: Fig. 1. A Use Case Diagram of the Course Registration Process 1) The System shall be able to authenticate the status of every user as either a valid student user or staff user. 2) The System shall only allow courses to be added or dropped during the date period allocated for course registration. 3) The System shall capture detailed general student information including the department, level, and college. 4) The System shall capture detailed information on students examination results including the failed, passed and dropped courses. 5) The System shall capture all relevant university rules that pertain to registration. 6) The System shall be able to give recommendation to a user once the valid status of the user is determined. 7) The System shall provide explanation for all recommendations suggested to the user. 8) The System shall provide real-time feedback when the user requests a recommendation. These set of requirements provided the basis for the design and implementation of the CAES. IV. ARCHITECTURE OF THE COURSE ADVISORY EXPERT SYSTEM (CAES) The CAES is based on a three-tier architecture that consists of a presentation layer, a middle layer and a data layer (see Figure 2). The presentation layer enables the user to access the application via a browser by using client devices such as desktop, laptop, or mobile phones. The various graphic user interfaces (GUIs) through which the user interacts with the system are contained in this layer. The middle layer consists of the Web application server, which facilitates communication in form of requests and notifications between the clients and the CAES application using the HTTP protocol. Apache Tomcat was used as the web application server for the CAES. The middle layer also contains the rule-based engine (RBR), which was implemented using Java Expert System Shell (JESS) 1 in order to enable reasoning on the rules that pertain to student registration; the case-based reasoning (CBR) engine enables case based reasoning. The RBR and CBR engine are deployed on the web application server. The middle layer also contains the Java servlets and JSP components that provided basis to weave java codes round the RBR and CBR engines of the CAES. The Java Data Base Connectivity (JDBC) protocol that enables interaction with the data layer of the architecture is also a contained in the middle layer. The Data Layer contains the data and knowledge artifacts that the system relies on to deliver its functionality. This layer consist of a knowledge base that contains the facts and rules (Jess fact files and rules) that is used by the RBR engine, and the relational database that contains information on all courses that are available in the University. A. Using the CAES for Advising In order to use the CAES for academic advising, the user will need to do the following: 1) Input a valid identification number at the CAES GUI 2) If successful, the CAES interface will display student details from the course information database. Displayed 1 http://herzberg.ca.sandia.gov/ Fig. 2. The 3-tier architecture of the system (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 3, No.5, 2014 9 | P a g e www.ijarai.thesai.org information will include current cumulative grade point average (CGPA), passed courses with grades obtained, failed courses, dropped courses, and the set of courses to register for the current semester. 3) Click recommend to generate a list of suggested courses to register for the new semester 4) Click on view explanation to see rationale for recommended courses. The Inference engine comprising of the rule engine and the CBR engine are used to generate recommendation of courses to be registered in a current semester. Fig. 3. Schematic representation of CAES recommendation The Inference mechanism checks to see if there are previous cases that are similar to the current case by taking note of the courses failed and dropped by a student and his current level. All of these factors are considered in generating an advice for the student. Case based reasoning is used because the system computes a recommendation by scanning the case base for instances that are similar to the one at hand and adapts the most similar old solution in a new scenario. The report is sent back to the student via the CAES GUI. If no similar case exists then rules contained in the knowledge base are used to construct a recommendation based on deductions that can be made using information available on student’s level, failed courses, failed prerequisites, and maximum total of credits that can be registered. CAES retains in the case base, all cases that have been handled successfully. An investment to be made in order to productively engage the CAES is that an administrator must continually maintain the case base to ensure that course information and the rules in the knowledge base are regularly updated. This is to ensure that the CAES system have the correct basis to make its recommendation during academic advising. The Figure 3 is a schematic representation of the recommendation process of the CAES using a program flowchart. V. THE REASONING MECHANISM OF THE CAES In this section, we give some insight into the reasoning behind some of the recommendations of the CAES. When CAES starts, the student course information is considered as a new case. CAES then computes a similarity score for the new case using the algorithm. Similarity (NC, OC) = common common + different Where NC is the new case, OC is the old case present in the case base. Common refers the matching pair between the new case and an old case. Different refers the mismatch pair between the new case and an old case. The case with the highest similarity score is picked as the candidate for adaptation in order to recommend to a user the courses to register. If a similar case does not exist, then a decision algorithm based on the rule engine is used to generate recommendation. The case adaptation procedure is rule-based, whereby university rules are used to guide selections. Three rules were used 1) a course with a higher credit unit should be selected over a course with a lower credit unit; 2) compulsory courses take precedence over electives and optional courses; and 3) a course that is pre-requisite for another that is failed, should be considered over courses that are not prerequisite for any other course. VI. CASE STUDY AND DISCUSSION A case study of Covenant University a tertiary institution based in Ota, Nigeria was undertaken using students of the Computer Science study program of the University as subjects. For a student intending to register a course at the beginning of a new semester these scenarios exist. 1) The student could have just the current semester course to register. 2) The student could have failed course(s) alongside the current semester courses. 3) The student could have dropped course(s) alongside the current semester courses. 4) The student could have failed and dropped course(s) alongside the current semester courses. In recommending the set of courses to register for the current semester, CAES uses the scenario above that is applicable to that particular student together with the set of rules outlined in the University policy for course registration, putting into consideration the different course status (course perquisites, compulsory or elective courses).The different conditions were modelled as rules and stored in the knowledge base of CAES. The following are the set of algorithms (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 3, No.5, 2014 10 | P a g e www.ijarai.thesai.org showing the rationale for specific scenarios that are captured in the knowledge base of CAES. The REGISTERDROPPEDFAILEDCOURSE algorithm in Table 1 caters for the scenarios i) - iii), while the REGISTERCOURSE algorithm in Table 2 caters for scenario iv). Table 3 shows sample JESS rule that states that a compulsory course have precedence over other types of courses. Also, Figure 4.1 and Figure 4.2 show snapshot of the CAES application. TABLE II. REGISTER COURSES ALGORITHM TABLE I. REGISTER DROPPED AND FAILED COURSES ALGORITHM Fig. 4. The CAES Interface Algorithm REGISTERCOURSE (E, S) Input: A vector E of Elective courses and S a vector of courses to register in the current session of the same semester. Output: A vector R containing the list of courses recommended for registration by the student in that semester. R ← NULL [initialize R] for each course Ci ∈ S while registeredCredit < maxRegistrable AND i < count(S) if prequisite(Ci) is passed Add Ci to R registeredCredit ← registeredCredit + courseCredit(Ci) increment i If registeredCredit < maxRegistrable for each course Kj ∈ S that is compulsory ordered by course credit in descending order while registeredCredit < maxRegistrable AND i < count(S) Add Kj to R registeredCredit ← registeredCredit + courseCredit(Kj) increment j If registeredCredit < maxRegistrable for each course Me ∈ E that is elective while registeredCredit < maxRegistrable AND e < count(E) AddMe to R registeredCredit ← registeredCredit + courseCredit(Me) increment e return the vector R containing the list of recommended course for the semester. Algorithm REGISTERDROPPEDFAILEDCOURSE (V, E, S) Input: A vector V of courses failed and/or dropped in the previous session of the same semester, E a vector of elective courses and S a vector of courses to register in the current session of the same semester. Output: A vector R containing the list of courses recommended for registration by the student in that semester. Initialize R. [Considering Failed and Dropped courses] for all courses vi ∈ V ordered by coursecode in ascending order while registeredCredit < maxRegistrable AND i < count(V) Add vi to R. registeredCredit ← registeredCredit + courseCredit(vi) increment i. [Considering failed prequisite course] If registeredCredit < maxRegistrable for each course Cj ∈ S while registeredCredit < maxRegistrable AND j < count(S) if prequisite(Cj) is failed OR dropped then Add Cj to D else Add Cj to R S ← S- Cj registeredCredit ← registeredCredit + courseCredit(Ci) increment j. [For the remaining courses] If registeredCredit < maxRegistrable for each course Kp ∈ S that is compulsory ordered by course credit in descending order while registeredCredit < maxRegistrable AND p < count(S) Add Kp to R registeredCredit ← registeredCredit + courseCredit(Kp) increment p. If registeredCredit < maxRegistrable for each course Me ∈ E that is elective while registeredCredit < maxRegistrable AND e < count(E) Add Me to R registeredCredit ← registeredCredit + courseCredit(Me) increment e. return the vector R containing the list of recommended course for the semester. (defrule recommend-compulsory-course "If there is a compulsory course, recommend for registration" ;; The course belongs to the type department and is compulsory (course (belongsTo department) (ccode ?code)(ctitle ?title) (cunit ?unit) (cstatus compulsory)) ;; and we haven't recommended this type yet (not (recommendcourse (ccode ?code) (ctitle ?title))) => ;; Recommend the course. (assert (recommendcourse (ccode ?code) (ctitle ?title) (cunit ?unit) (because "compulsory departmental course")))) TABLE III. SAMPLE JESS RULE TO SELECT A COMPULSORY COURSES Fig. 5. CAES Recommendation Page (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 3, No.5, 2014 11 | P a g e www.ijarai.thesai.org VII. EVALUATION Human experts conducted a usability evaluation of the prototype in order to assess the level of user satisfaction with the system. This was then validated through the direct method of evaluating expert systems as used by Salim et al. [14]. A small experiment to test the system’s recommendations against those of human advisors was conducted using the direct method. Course Advisers across each level from the Department of Computer and Information Sciences of Covenant University were asked to participate in the survey. Each received an identical set of questionnaire, and had a running version of CAES installed for them. The course advisers were asked to rank the recommendation of CAES on a likert scale of 0-5 to assess the degree of how true or false are the recommendations of CAES. A brief overview of the direct method of expert system evaluation used by each evaluator is as follows: 1) The evaluator is given a sample copy of the software system - CAES to be evaluated. 2) The evaluator selects a benchmark problem, based on his experience, and runs this problem on CAES. 3) After running the bench-mark problem, the evaluator responds to a set of questions (14) in the questionnaire instrument and estimates a quantitative answer to each question on a 0 to 5 scale with 5 being very true and 0 being very false. 4) Each numerical result is multiplied by a weighting factor as given in the weight column. 5) The weighted values are summed and then divided by the sum of the weights (19) to give a result in the numerical range of 0 to 5. The Figure 6 gives a computation of the evaluation experiment conducted by one of the evaluator. A subset of the summary result in calculating the experimental evaluation of the evaluators is given in the Table 4. Evaluator Computed Satisfaction Level 1 4.00 2 4.16 3 4.21 4 3.52 5 3.57 Mean Satisfaction Level 3.89 VIII. RESULT AND DISCUSION From the statistical analysis of the results obtained from the evaluation of the human experts that participated in the experiment, CAES had a mean satisfaction level score of 3.89 out a maximum of 5.0, which is indicative of a 77.8% level of user satisfaction. The implication of this result is that after the experts have considered important metric dimensions such as correctness of answer, accuracy, quality of reasoning technique, sensitivity, reliability, cost effectiveness, and observed limitations of the system, the system obtained a mean rating of 77.8%. This connotes an appreciably good rating for the CAES system, and an indication of its viability to support the task of academic advising. Fig. 6. Evaluator’s questionnaire TABLE IV. RESULT OF EVALUATION EXPERIMENT (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 3, No.5, 2014 12 | P a g e www.ijarai.thesai.org IX. CONCLUSION The CAES system that was developed is intended for use in a mid-range university. Its experimental version was successfully trialed by the Department of Computer and Information Sciences of Covenant University. The modular structure and web-based design of the CAES makes it suitable to be launched and used in other departments of the University. In our future work, we shall improve on the case revision and case adaptation capability of the CAES, because we observed some complex cases, which the system did not handle adequately. This had to do with students that have changed from one programme to another - many of them more than once - , and have failed and dropped courses that are spread among different departments. We observed that in such scenarios, it was difficult for the current implementation of CAES find good cases to use as basis for adaptation to construct a recommendation. We do not consider this a major drawback of CAES, because even for the human course adviser, cases where a student has failed multiply in different departments are more intricate to handle, yet we seek to improve CAES in these areas. As its contribution, this work offers a demonstration of application of artificial intelligence technology (AI) to support academic advising, which is very crucial to the academic well- being of students. The CAES was not intended to eliminate the role of human (staff) advisors, rather it enables students to concentrate on real issues that pertain to course registration, and affords unrestricted access to expert advice thereby reducing the burden placed on the human advisor. REFERENCES [1] Aamodt, A., E. Plaza, 1994. Case-based reasoning: Foundational issues, methodological variations, and system approaches. AI communications, 7 (1): 39-59. [2] Feghali, T., I. Zbib and S. Hallal, 2011. A Web-based Decision Support Tool for Academic Advising. Educational Technology & Society, 14 (1): 82–94. [3] Chun M., Y. Eva, S. Tsang, S. Lam, C. Dominic, and C. Pang, 2010, Intelligent Counseling System: A 24 x 7 Academic Advisor, EDUCAUSE Quarterly 33(4). [4] Albalooshi, F., and S. Shatnawi, 2010. HE-Advisor: A Multidisciplinary Web-Based Higher Education Advisory System, Global Journal of Computer Science and Technology, 10(7):37-49. [5] Zucker, R., 2009. ViCurriAs: a curriculum visualization tool for faculty, advisors, and students. Journal of Computing Sciences in Colleges, 25(2):138-145. [6] Rao T., S. Coleman, and C. Hollenbeck, 1987. ADVISOR – An Expert System for Student Advisement. Proc. 15th Annual Conference on Computer Science, St Louis, pp. 32-35. [7] Murray W., and L. LeBlanc, 1995. A Decision Support System for Academic Advising. Proc. 1995 ACM Symposium on Applied Computing, pp. 22-26. [8] Harlan R. M, 1994. The Automated Student Advisor: A Large Project for Expert Systems Courses. ACM SIGCSE Bulletin 26(1):31- 35. [9] Dunkel J., and R. Bruns, 2005. Software architecture of advisory systems using agent and semantic Web technologies. Proceedings of the IEEE/WIC/ACM International Conference on Web Intelligence, 19:418 – 421. [10] Gunadhi H., L. Kwang-Hui, Y. Wee-Yong, 1995. PACE: a planning advisor on curriculum and enrolment. Proceedings of the Twenty-Eighth Hawaii International Conference on System Sciences, 3:23 – 31. [11] Pokrajac D., M. Rasamny, 2006. Interactive Virtual Expert System for Advising (InVEStA). 36th Annual Frontiers in Education Conference. pp. 18-23. [12] Valtorta M., B. Smith, and D. Loveland, 1984. The graduate course advisor: a multi-phase rule-based expert system. Proceedings of the IEEE Workshop on Principles of Knowledge-Based Systems. [13] Sandvig, J. and R. Burke, 2005. AACORN: a CBR recommender for academic advising. Technical Report, TR05-015, DePaul University. [14] Salim M., A. Villavicencio, and M. Timmerman, 2002. A Method for Evaluating Expert System Shells for Classroom Instruction. Journal of Industrial Technology, Vol. 19(1). 
work_6vaed5kxqndbnfdfz5mbn7dsqm	----	MISNIS: An Intelligent Platform for Twitter Topic Mining Accepted Manuscript MISNIS: An Intelligent Platform for Twitter Topic Mining Joao P. Carvalho , Hugo Rosa , Gaspar Brogueira , Fernando Batista PII: S0957-4174(17)30531-6 DOI: 10.1016/j.eswa.2017.08.001 Reference: ESWA 11470 To appear in: Expert Systems With Applications Received date: 1 February 2016 Revised date: 31 July 2017 Accepted date: 1 August 2017 Please cite this article as: Joao P. Carvalho , Hugo Rosa , Gaspar Brogueira , Fernando Batista , MISNIS: An Intelligent Platform for Twitter Topic Mining, Expert Systems With Applications (2017), doi: 10.1016/j.eswa.2017.08.001 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. http://dx.doi.org/10.1016/j.eswa.2017.08.001 http://dx.doi.org/10.1016/j.eswa.2017.08.001 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Highlights:  An intelligent platform to efficiently collect and manage large Twitter corpora  Circumvents Twitter restrictions that limit free access to 1% of all flowing tweets  An add-on implementing intelligent methods for Twitter topic mining  Intelligent retrieval of tweets related to a given topic  A case study is presented as a demonstration example ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT MISNIS: An Intelligent Platform for Twitter Topic Mining Joao P. Carvalho INESC-ID/Instituto Superior Técnico, Universidade de Lisboa R. Alves Redol, 9, 1000-029 Lisboa, Portugal joao.carvalho@inesc-id.pt Hugo Rosa INESC-ID, Portugal hugo.rosa@inesc-id.pt Gaspar Brogueira INESC-ID/ISCTE-IUL, Portugal gmrba@iscte.pt Fernando Batista INESC-ID/ISCTE-IUL, Portugal Fernando.Batista@inesc-id.pt ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Abstract Twitter has become a major tool for spreading news, for dissemination of positions and ideas, and for the commenting and analysis of current world events. However, with more than 500 million tweets flowing per day, it is necessary to find efficient ways of collecting, storing, managing, mining and visualizing all this information. This is especially relevant if one considers that Twitter has no ways of indexing tweet contents, and that the only available categorization “mechanism” is the #hashtag, which is totally dependent of a user’s will to use it. This paper presents an intelligent platform and framework, named MISNIS - Intelligent Mining of Public Social Networks’ Influence in Society - that facilitates these issues and allows a non-technical user to easily mine a given topic from a very large tweet’s corpus and obtain relevant contents and indicators such as user influence or sentiment analysis. When compared to other existent similar platforms, MISNIS is an expert system that includes specifically developed intelligent techniques that: (1) Circumvent the Twitter API restrictions that limit access to 1% of all flowing tweets. The platform has been able to collect more than 80% of all flowing portuguese language tweets in Portugal when online; (2) Intelligently retrieve most tweets related to a given topic even when the tweets do not contain the topic #hashtag or user indicated keywords. A 40% increase in the number of retrieved relevant tweets has been reported in real world case studies. The platform is currently focused on portuguese language tweets posted in Portugal. However, most developed technologies are language independent (e.g. intelligent retrieval, sentiment analysis, etc.), and technically MISNIS can be easily expanded to cover other languages and locations. Keywords: Twitter; Intelligent Topic Mining; Fuzzy Fingerprints; Text Analytics; Sentiment Analysis. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 1 Introduction When Twitter was launched in 2006 as a simple public social networking service enabling users to send and read short 140-character messages, hardly anyone could predict that it would become a major tool for spreading news, for dissemination of positions or ideas, and for the commenting and analysis of current world events. This became evident with the so-called “Arab Spring” in 2010, where Twitter was used as an alternative means of communicating to the outside world what was censored by state controlled traditional news broadcasters. During the subsequent years, events such as the “Spanish protests”, the “London riots” or the “Taksim Gezi Park protests”, further increased the notion that important events are often commented in Twitter before they become “public news”. This has led to a change in how the public perceives the importance of social networks, and even news agencies and networks had to adapt and are now using Twitter as a potential (and some times preferential) source of information. As an example, Sankaranarayanan (2009), showed how Twitter can be used to automatically obtain breaking news from the tweets posted by users, and exemplifies that when Michael Jackson passed away, “the first tweet was posted 20 minutes after the 911 call, which was almost an hour before the conventional news media first reported on his condition”. In fact, Twitter is so fast that it can even outpace an earthquake: on August, 23rd 2011, when a 5.9 magnitude earthquake struck close to Richmond, Virginia, U.S.A., the effects were first felt in Washington D.C. from where several tweets were posted stating the event; various people reported having read those tweets in New York City (400Km away) before the earthquake reached them! (Ford, 2011) (Gupta et al., 2014). The negative side of this fast paced online news environment is that it discourages fact- check and verification (Chen et al., 2015), and some concern is justified when considering the rise in phenomena such as “fake news”, that were, for example, exploited in recipe-like fashion to impact the 2016 USA Presidential elections (Mustafaraj and Metaxas, 2017). Vosoughi et al. (2017) aimed to reduce the impact of false information on Twitter by automatically predicting, with 75% accuracy, the veracity of rumors on a collection of nearly 1 million tweets, extracted from real-world events such as the 2013 Boston Marathon bombings, the 2014 Ferguson unrest and the 2014 Ebola epidemic. The previous examples show the importance of automatically analyzing the massive amount of information on Twitter. However, using Twitter as a source of information involves many technical obstacles, of which the first is collecting and dealing with the amount of flowing information. As of mid 2015, more than 500 millions tweets covering thousands of different topics are published daily (almost 6000 tweets per second!). Collecting, storing, managing and visualizing such large amounts of information is a far from trivial problem and demands dedicated and intelligent hardware and software platforms. Even assuming one is able to access all tweeted contents, there is still the problem of filtering which content is relevant for a given topic of interest. This is far from trivial, even if we simply consider doing it on a daily basis: on a given day it is very unlikely that more than a few thousand tweets are relevant to a given discussion topic (even when considering major topics). We are talking about detecting 0.001%-0.01% of the 500 million daily tweets, which is basically trying to find a needle on a hay stack. Twitter’s approach to deal with this problem is to provide a list of top trends (Twitter, 2010) and the #hashtag mechanism: when referring to a certain topic, users are encouraged to indicate it using a hashtag. E.g., “#refugeeswelcome in Europe!” indicates the topic of the tweet is the current refugees crisis in Europe. However, not all tweets related to a given ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT topic are hashtagged. In fact, according to (Mazzia, 2010), only 16% of all tweets are hashtagged, numbers that have been confirmed by our experiments. The explanation for this lies partially with the fact that 140 characters is a scarce amount of text to communicate a thought, something that can be aggravated by the inclusion of an #hashtag, which also uses valuable space. It thus becomes clear that, to correctly analyze a given discussion topic, it is of the utmost importance to retrieve as much of the remaining 84% untagged tweets as possible. Since no other tagging mechanisms exist in Twitter, the process of retrieving tweets that are related to a given topic, our needle on a haystack, must use some kind of text classification process in order to detect if the contents of a given tweet is somehow related with the intended topic. This classification process must simultaneously be able to retrieve only the relevant tweets (i.e., have high values of precision and recall), and to be computationally efficient in order to deal with the huge amount of data. Additional tasks of interest when considering the use of Twitter as an information source, include finding which of the topic related tweets have more relevance (e.g. by finding who are most important “actors” discussing the topic), performing sentiment analysis, extract statistics on the topic origin and respective spatial-temporal evolution, etc. In this article, we present an intelligent platform, MISNIS - Intelligent Mining of Public Social Networks’ Influence in Society, which addresses the issues mentioned above, and can be used as an expert system by social scientists when studying social networks’ impact in society. The platform can be divided into two major blocks (Figure 1): 1) Smart mechanisms for collecting, storing and managing Twitter information; 2) Intelligent mechanisms to retrieve, analyze and represent the information that is relevant for a given topic. Figure 1: MISNIS Framework architecture ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT When compared to other existent similar platforms, MISNIS includes specifically developed intelligent techniques that: (1) Circumvent the Twitter API restrictions that limit access to 1% of all flowing tweets. The platform has been able to collect more than 80% of all flowing portuguese language tweets in Portugal when online (Brogueira et al., 2016); (2) Intelligent retrieval of most tweets related to a given topic even when the tweets do not contain the topic #hashtag or user indicated keywords. A 40% increase in the number of retrieved relevant tweets has been reported in real world case studies Carvalho et al., 2017). Despite being operational, MISNIS is an ongoing work and can be improved in several aspects: (1) The platform is currently focused on portuguese language tweets posted in Portugal. However, most developed technologies are language independent (e.g. intelligent retrieval, sentiment analysis, etc.), and technically MISNIS can be easily expanded to cover other languages and locations; (2) Sentiment analysis methods can be improved; (3) Dependence on Twitter and Google APIs: most changes to the APIs endpoints imply changing and recompiling the platform code. The paper is organized as follows: In section 2 we describe some related work relevant to the developed platform; Section 3 describes the architecture of the framework that was implemented for Twitter data acquisition, storage, management and visualization; Section 4 focuses on the expert system for intelligent Twitter data mining added to the framework; Section 5 presents a small case study used to exemplify the developed platform and framework; Finally, Section 6 presents some conclusions. 2 Related Work 2.1 Large Scale Social Data Acquisition and Storage The analysis of the content and information shared on social networks has been proved useful in various fields, including Politics, Marketing, Tourism, Public Health, and Safety. Twitter is amongst the most widely used social networks, making available about 500 million tweets every day1, on average. Twitter provides free access to part of the information produced by its users through public APIs (Application Programming Interface), and the popularity of Twitter as a source of information has led to the development of numerous applications and to new research methods in various fields. For example, Paul et al. (2011) developed a method for tracking disease risk factors by measuring the behavioral level, tracking diseases by geographic regions to analyze the symptoms and medication applied. The study was based on about 1.5 Million tweets related to health that contained references to various ailments including allergies, obesity and insomnia. Santos et al. (2013, 2014) used a set of approximately 2700 tweets produced in Portugal to predict the incidence and spread of the influenza virus through the Portuguese population. Widener et al. (2014) used information extraction and sentiment analysis (through a data mining framework), to try to understand how geolocated tweets can be used to research the prevalence of healthy and unhealthy food in contiguous regions of the United States. Other studies related to public health were also reported by Culotta (2010) and by Scanfeld et al. (2010). Twitter was also used as a source of information to help identifying or locating the occurrence of earthquakes, taking into account that “when an earthquake occurs people produce many posts on Twitter related to the event, which permits the identification of earthquakes simply by observing the increase in the tweet volume” (Sakaki et al. 2010). Kumar et al. (2013b) proposes an approach to identify a subset of users and their location to justify them to be followed in disaster situations in order to get a quick access to useful information 1 https://about.twitter.com/company (accessed in 14-05-2015). ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT about the event. During a crisis a particular user’s location is an important factor in determining whether he/she is likely to publish relevant information on the state of crisis. For instance, in the eventuality of an earthquake, tweets produced in a place close to the earthquake are likely to be more relevant to assess the situation than tweets produced from a more distant location. Other studies have been produced based on similar topics (Mendoza et al. 2010, Qu et al. 2011, Lachlan et al. 2014). Gerber (2014) tried to predict criminal activity in the largest city of the United States of America using tweets marked in space and time. Since tweets are public, officially made available by Twitter services, the development of linguistic analysis models that enable the automatic identification of topics related to the commission of a crime may be considered quite relevant not only in preventing similar crimes but also to support decision making under trial in court of law. The design of software architectures for capturing Twitter information and extracting relevant knowledge from tweets represents a major challenge, not only due to the massive amounts of streamed data, but also due to possible access limits imposed by Twitter (Oussalah et al. 2013). Most of the work reported in literature restricts the data in some way, in order to respect the limits imposed by Twitter. For example, Perera et al. (2010) describes a software architecture based on the Twitter API based on Python and MySQL that collects tweets sent only to specific users. Twython, a Python wrapper, was used to obtain spatial data (location, name, description, etc.) of the authors of the tweets. The collecting process runs over in 5-minute intervals and the collected tweets sent to a particular Twitter id user, including the President Barack Obama. Anderson et al. (2011) reports a concurrency-based software architecture that allows collecting a large volume of data, with a theoretical maximum of 500 million tweets per day. The developed code is multi-threaded and therefore adapted to run on machines with multiple processors, uses Spring, MVC, Hibernate and JPA frameworks, and the infrastructure components: Tomcat, Lucene (for tweet indexation), MySQL (to store the collected data). Marcus et al. (2011) developed TwitInfo, a platform for collecting and processing tweets in real time for sentiment analysis. The architecture proposed by Oussalah et al. (2013) collects tweets continuously and in real time using the Streaming API, aiming at semantically and spatially analyze the collected data. The collected tweets are restricted to a rectangular area bounded by geographical coordinates (longitude and latitude), and the software implementation is based on Django (framework for developing web applications in Python), Lucene, and MySQL. This architecture allows searching for tweets using the text, username or location. 2.2 Tweet Topic Detection One of the goals of this work is to automatically classify tweets into a set of topics of interest. Tweet Topic Detection involves automatically determining if a given tweet is related to a given, usually #hashtagged, topic. This is basically a classification problem, albeit one with its own specificities: (1) it is a text-based classification problem, with an unknown and vast number of classes, where the documents up for classification are very short (maximum of 140 characters in length); (2) it fits the Big Data paradigm due to the huge amounts of streaming data. We purposefully distinguish between Topic Classification and Topic Detection. The former is broadly known in Natural Language Processing (NLP) as Text Categorization, and is defined as the task of finding the correct topic (or topics) for each document, given a closed set of generic categories (subjects, topics) such as politics, sports, music, religion, etc., and a collection of text documents (Feldman, 2006), in this case, tweets; the tweets will commonly belong to one or more of those categories and it is highly uncommon that a tweet goes unclassified. The latter is more detailed in its approach ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT since it attempts to determine the topic of the document, from a predetermined large set of possible topics, where the topics are so unique amongst themselves that there is a good chance that a tweet without a hashtag will probably not belong to any of the current trends. With this difference in mind, the most similar works on topic detection within Twitter are those related with emerging topics or trends, as for example the works of Mathiodakis (2010), Cataldi (2010) Kasiviswanathan (2011) or Saha (2012). In these articles, the authors use a wide variety of text analysis techniques to determine the most common related words and, as consequence, detect topics. In our work, we assume the existence of trending topics and set the goal of efficient detection of tweets that are related to said topics, despite not being explicitly marked (hashtagged) as so. Topic Classification is also a well-documented and commonly studied task. In Lee (2011), Twitter Trending Topics are classified into 18 broad categories like sports, politics, technology, etc., and a classification accuracy of 65% and 70% is achieved when using text-based and network-based classification modelling, respectively. The experiment was performed on a dataset of randomly selected 768 trending topics (over 18 classes). More recently, Cigarrán et al. (2016) proposed an approach based on Formal Concept Analysis (FCA) to perform Twitter topic detection in unsupervised fashion, finding that it outperforms traditional classification, clustering and probabilistic approaches on a benchmark Replab 2013 dataset. Empowered by previous work and supporting our view of topic detection, the most promising method is Twitter Topic Fuzzy Fingerprinting (Rosa et al., 2014a, 2014b). Fingerprint identification is a well-known and widely documented technique in forensic sciences. In computer sciences a fingerprint is a procedure that maps an arbitrarily large data item (such as a computer file, or author set of texts) to a much compact information block, its fingerprint, that uniquely identifies the original data for all practical purposes, just as human fingerprints uniquely identify people. Fuzzy Fingerprints were originally introduced as a tool for text classification by Homem and Carvalho (2011). They were successfully used to detect authorship of newspaper articles (out of 73 different authors). For text classification purposes, a set of texts associated with a given class is used to build the class fingerprint. Each word in each text represents a distinctive event in the process of building the class fingerprint, and distinct word frequencies are used as a proxy for the class associated with a specific text. The set of the fuzzy fingerprints of all classes is known as the fingerprint library. Given a fingerprint library and a text to be classified, the text fingerprint is obtained using a process similar to the one used to create the fingerprint of each class, and then a similarity function is used to fit the text into the class that has the most similar fingerprint. In order to use Fuzzy Fingerprints for tweet topic detection, several procedural changes were proposed by Rosa et. al (Rosa, 2014a, 2014b). According to the authors, the Twitter Topic Fuzzy Fingerprints performed very well on a set of 2 millions English, Spanish and Portuguese tweets collected over a single day, beating other widely used text classification techniques. The training set consisted of 11000 tweets containing the 22 of the top daily trends (hashtagged topics). 350 unhashtagged test tweets were properly classified with an f-measure score of 0.844 (precision=0.804, recall=0.889). Further work by Rosa (2014), used a training set of 21000 tweets, from “21 impartially chosen topics of interest out of the top trends of the 18th of May, 2013”. The test set was made of “585 tweets that do not contain any of the top trending hashtags” and “each tweet was impartially annotated to belong to one of the 21 chosen top trends”. After extensive parameter optimization using a development set, the fuzzy fingerprint method scored an f-measure of 0.833 proving to be not only more accurate than other well- known classifying techniques (kNN and SVM), but also much faster (177 times faster than kNN and 419 times faster than SVM). ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Topic models are commonly reported in the literature as one of the most successful techniques for topic detection/classification/trending on twitter (Hoffman et al., 2010). Non-probabilistic topic models, namely Latent Semantic Analysis (LSA) (Landauer, 1998) appeared first, but most of the current literature refers to generative probabilistic models (Blei, 2012), based on Latent Dirichlet Allocation (LDA). Our previous attempts applying methods based on LDA to the specific problem of tweet topic detection produce weak results, unless very extensive parameterization and testing was done a priori for each new topic, which obviously prevents their use in the developed platform. 2.3 User Influence The concept of influence is of much interest for several fields, such as sociology, marketing and politics. Empirically speaking, an influential person can be described as someone with the ability to change the opinion of many, in order to reflect his own. While Rogers (1982) supports this statement, claiming that “a minority of users, called influentials, excel in persuading others”, more modern approaches (Domingos, 2001) seem to emphasize the importance of interpersonal relationships amongst ordinary users, reinforcing that people make choices based on the opinions of their peers. The point is that “influence” is an abstract concept, which makes it exceptionally hard to quantify. Several studies have attempted to accomplish this goal. In (Cha, 2010), three measures of influence were taken into account, regarding Twitter: in-degree, re-tweets and mentions, where “in-degree is the number of people who follow a user; re-tweets mean the number of times others forward a user's tweet; and mentions mean the number of times others mention a user's name”. It concluded that while in-degree measure is useful to identify users who get a lot of attention, it “is not related to other important notions of influence such as engaging audience”. Instead “it is more influential to have an active audience who re-tweets or mentions the user”. In (Leavitt, 2009), the authors conclude that within Twitter, “news outlets, regardless of follower count, influence large amounts of followers to republish their content to other users”, while “celebrities with higher follower totals foster more conversation than provide retweetable content”. InfluenceTracker (Razis, 2014) is a framework that rates the impact of a Twitter account taking into consideration an Influence Metric, based on the ratio between the number of followers of a user and the users it follows, and the amount of recent activity of a given account. It also calculates a Tweet Transmission rate where the “most important factor (...) is the followers' probability of re-tweeting”. Cha (2010) also shows “that the number of followers a user has, is not sufficient to guarantee the maximum diffusion of information in Twitter (...) because, these followers should not only be active Twitter users, but also have impact on the network”. Even if one agrees on the measures that best represent influence, aggregating and computing the measures is not a trivial task since user interactions should not be ignored. A sound approach consists in using graphs and to compute user relevance recurring to graph centrality algorithms. In graph theory and network analysis, the concept of centrality refers to the identification of the most important vertices within a graph, in this case, the most important users. We therefore define a graph G(V,E) where V is the set of users and E is the set of directed links between them. Currently the most “famous” centrality algorithm is PageRank (Page, 1998, 1999). It is one of Google's search engine methods, with web pages used as nodes and back-links ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT forming the edges of the graph. The PageRank is considered to be a random walk model, because the weight of a page/node is “the probability that a random walker (which continues to follow arbitrary links to move from page to page) will be at a node at any given time”. A damping factor is used as the “probability of the random walk to jump to an arbitrary page, rather than to follow a link, on the Web” and is “required to reduce the effects on the PageRank computation of loops and dangling links in the Web” (Phuoc, 2009). Other less complex and with guaranteed convergence centrality methods exist, such as for example, Katz (Katz, 1953), which are often preferred over PageRank for that reason. 2.4 Sentiment Analysis Sentiment Analysis is a relevant and well-known task that consists of extracting sentiments and emotions expressed in texts. Being the first step towards the online reputation analysis, it is now gaining particular relevance because of the rise of social media, such as blogs and social networks. The increasing amount of user-generated contents constitute huge volumes of opinionated texts all over the web that are precious sources of information, especially for decision support. Sentiment Analysis can be used to know what people think about a product, a company, an event, or a political candidate. Sentiment analysis can be performed at different complexity levels, where the most basic one consists just on deciding whether a portion of text contains a positive or a negative sentiment. Dealing with the huge amounts of data available on Twitter demand clever strategies. One approach combines sentiment analysis and causal rule discovery (Dehkharghani, 2014). Other, by Kontopoulos (2013) uses ontologies. An interesting and simpler idea, explored by Go (2009), consists of using emoticons, abundantly available on tweets, to automatically label the data and then use such data to train machine learning algorithms. The paper shows that machine learning algorithms trained with such approach achieve above 80% accuracy, when classifying messages as positive or negative. A similar idea was previously explored by Pang (2002) for movie reviews, by using star ratings as polarity signals in their training data. This latter paper analyses the performance of different classifiers on movie reviews, and presents a number of techniques that were used by many authors and served as baseline for posterior studies. As an example, they have adapted a technique, introduced by Das and Chen (2001), for modelling the contextual effect of negation, adding the prefix NOT_ to every word between a “negation word” and the first punctuation mark following the negation word. Common approaches to sentiment analysis also involve the use of sentiment lexicons of positive and negative words or expressions (Stone, 1996; Hu, 2004; Wilson, 2005; Baccianella, 2010). Another research approach involves learning polarity lexicons and can be especially useful for dealing with large corpora. The process starts with a seed set of words and the idea is to increasingly find words or phrases with similar polarity, in semi-supervised fashion (Turney, 2002). The final lexicon contains much more words, possibly learning domain-specific information, and therefore is more prone to be robust. The work reported by Kim (2004) is another example of learning algorithm that uses WordNet synonyms and antonyms to learn polarity. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 2.5 Twitter Data Analysis Platforms To the best of our knowledge, there are no Twitter data analysis platforms with the exact same goal as MISNIS. However, similar ones have come up in recent years, albeit more focused on social media marketing by aiding brands to grow or just helping regular users and public personalities to better foster their web engagement. The earliest such alternative that we found was TwitterMonitor (Mathioudakis and Koudas, 2010), which automatically identified emerging trends on Twitter and provided meaningful analytics that synthetized an accurate description of each topic. Other tools focus on eye-catching reports and graphics to display relevant data and often provide insightful analytics on the retrieved data: (i) Warble (www.warble.co) is a web-based solution that allows users to track keywords and hashtags that matter to him/her, while monitoring brands and brand engagement; (ii) Twitonomy (www.twitonomy.com) provides detailed analytics on anyone’s tweets, allowing users to get insights on followers and friends as well as your interactions with other users; (iii) TweetReach (www.tweetreach.com) gives real-time analytics on a user’s reach, performance and engagement while continuously analysing all posts about the topics he/she cares about, including sentiment analysis; (iv) SocioViz (www.socioviz.net) analyses any topic, term or hashtag and identifies key influencers, opinions and contents; (v) Mozdeh (http://mozdeh.wlv.ac.uk/) is a free Windows application for keyword, issue, time series, sentiment, gender and content analyses of social media texts; (vi) Netlytic (http://netlytic.org) is a community-supported text and social networks analyser that can automatically summarize and discover social networks from online conversations on social media sites; (vii) Discovertext (http://discovertext.com) is a commercial company combining machine learning classifiers with cloud and crowdsource services to retrieve relevant items and sort them into topics and sentiment categories; (viii) Visibrain (http://www.visibrain.com) is a commercial “media monitoring tool for PR and communications professionals, used for reputation management, PR crisis prevention, and detecting influencers and trends”, claiming to be able to capture all social media around a brand. Most of these tools share common features with MISNIS. From keyword tracking, to geo- located data, sentiment analysis and user influence, almost all the above-mentioned platforms implement at least one of these capabilities, often more, and with better user interface. However, except for DiscoverText, none of the platforms have the mechanisms to overcome Twitter API limits. Hence they are only able to capture and analyse 1% of all flowing tweets. The exception, DiscoverText, makes use of the Twitter “firehose”, which allows access to 100% of the tweets in real-time streaming. However it is a paid (and quite expensive) solution. Even more important, independently from being paid or a free solution, and far as we could tell, none of the mentioned platforms is able to detect relevant related tweets unless they contain explicit user defined keywords/hashtags, therefore missing important information for the analysis of a given topic. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 3 Twitter Data Acquisition, Storage, Management and Visualization In order to circumvent Twitter data access restrictions and build a tweets repository from where data could be efficiently managed, intelligently retrieved and visualized, we developed an information system consisting of four main modules (Figure 2): 1) Data collection; 2) Data expansion; 3) Data access; 4) Visualization. Figure 2: Architecture of the Twitter data collection, management and visualization system. The presented system focus mainly on Portuguese Twitter data (Tweets produced in Portugal and in European Portuguese), but can easily be adapted and expanded for most countries and languages. In the Collect module we retrieve geolocated tweets produced in Portugal and discard those that are not recognized as written in European Portuguese. The collection uses Twitter StremingAPI, which implies that at most 1% of the data is collected. All collected tweets are stored in MongoDB2. In the Expand module we identify each individual previously collected user, and explore their timelines to retrieve past tweets (and add them to the MongoDB). In the Access module we implemented a REST API3 as an abstraction to intelligently access the database. The Visualization module implements a dashboard to visualize metrics, indicators and any queried information. Each module is detailed in the following sections. 2 https://www.mongodb.org, last accessed, July 2015 3 http://www.restapitutorial.com/, lasta accessed, July 2015 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 3.1 Geolocated Data Collection The Twitter Streaming API statuses/filter allows the access to tweets being tweeted at the time of request. Several filters can be used to tailor the requested information: the API allows filtering by keywords, hashtags, user ID and geographically delimited regions (Kumar 2013a). The number of parameters and the volume of returned information is limited by Twitter. Currently each request allows for a maximum of 400 keywords, 25 geographical regions or 5000 user IDs, and up to a maximum of 1% of all currently flowing tweets are searched for and returned. The geolocation of a tweet can be obtained using two different processes: i) directly from the tweet when the user opts to make his location known at time of publishing; ii) using the information contained in the user profile location field. The percentage of geolocated tweets is low, under one fifth of all tweets4, and as such, it is easier to comply with Twitter API restrictions when filtering for geolocated tweets:  Taking into consideration that currently there are 500 million tweets per day5, it is theoretically possible to retrieve up to 5 million tweets per day using the Streaming API from a single user account;  Previous works (Brogueira, 2014a, 2014b) have shown that in 2014, 60000 geolocated tweets were produced per day, in Portugal and using (European) Portuguese, i.e., almost 2 orders of magnitude lower than the 5 million Streaming API daily limits. The implemented platform works on a 24 hours per day basis using the Twitter Streaming API to collect geolocated flowing tweets within Portugal. The delimitation of the geographical area of mainland Portugal and the archipelagos of Madeira and Azores uses the Twitter REST API geo/search. Note that this area includes small parts of Spain and North Africa that are filtered a posteriori. The collected tweets are compressed and stored on a hard drive (in as “real time” as possible), and later filtered and stored in a MongoDB database. The filtering operation consists of considering only tweets that are produced in Portugal (field place.country = ”pt”) and written in European Portuguese (field lang = “pt”). The double-check allows for the detection of tweets where the Twitter language detection algorithm6 fails, which is not uncommon between Portuguese, Spanish and Galician. In addition to the tweet filter and storage operation, the author of each tweet is identified using the field user.id, and if he is unknown, added to a “Users” MongoDB collection. For each new user it is created a JSON7 document containing: i) the user.id; ii) the date of first detection; iii) control flags (see section 3.2). 3.2 Database Expansion The tweet corpus expansion is based on the retrieval of the timeline of each user present in the “Users” collection. The user’s timeline is the record of the user’s recent Twitter activity, and can be accessed via the REST API statuses/user_timeline. 4 https://pressroom.usc.edu/twitter-and-privacy-nearly-one-in-five-tweets-divulge- user-location-through-geotagging-or-metadata/, last accessed July 2015. 5 https://about.twitter.com, last accessed October 2015. 6 https://blog.twitter.com/2013/introducing-new-metadata-for-tweets, last accessed July 2015. 7 http://json.org/, last accessed June 2015. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT The API returns the most recent 3200 tweets of a given user. However, Twitter imposes several restrictions that hinder the timeline collection process: up to 180 requests are authorized per 15 minute period and per authenticated Twitter API access account (Kumar, 2013a), and each request returns at most 200 tweets. Therefore, 16 API requests are needed in order to retrieve a 3200 tweets timeline. As such, it is only possible to retrieve the timeline of 11 different users per 15-minute period (around 1080 timelines per day) when using a single Twitter account. Since an average of 232 new users are identified per day by the system (Brogueira, 2015a), and the total of registered users is, by August 2015, above 120K, collecting the timelines of every single new user, and updating existing users’ timelines on a continuous basis using a single Twitter API access account is an increasingly lengthy process, that is not viable and far from straightforward due to the above-mentioned restrictions. Timeline retrieval in the developed system uses 15 different Twitter API access accounts that are synchronized and optimized to prevent repeated invocations and failures due to exceeding the API limits during the 15-minute window. The retrieval process considers three different scenarios: i) Integration of a new user, which implies obtaining the complete timeline (up to 3200 tweets); ii) Existing user, for whom it is necessary to retrieve tweets produced since the date of the last retrieved tweet; iii) Blocked access users, i.e., users that have explicitly blocked the access to their timelines. Each account is dynamically assigned to one of the first two scenarios taking into consideration the amount of timelines to process for each. Blocked users are kept out of the loop and checked sporadically in order to detect eventual status change. Each account is also associated to a JSON document containing the strings used for Twitter API authentication (OAuth), and the flags that identify which scenario the account is currently assigned to and how it should operate. More details can be found in (Brogueira et al., 2015b, 2016). With the joint operation of the geolocated data collection and data expansion modules, MISNIS has been able to collect more than 80% of all flowing portuguese language tweets in Portugal when online, which is a huge amount when compared to the theoretical 1% freely made available by Twitter (Brogueira et al., 2016). 3.3 Data Access and Data sharing The Data Access module consists of a REST API developed in order to facilitate the access to the information stored in the database, allow access by third party applications, and enable the developed Dashboard (see section 3.4). REST stands for Representational State Transfer, a set of constraints and principles used in web interface architectures. A REST API is a set of data and functions that facilitate information exchange between applications and web services designed according to the REST principles. The developed REST API makes a set of endpoints available for interaction with the MongoDB. Table 1 presents some of the endpoints developed for information access. They are divided into three categories: i) Access to tweets information; ii) Access to user information; iii) Access to previously processed statistics (performed on the stored data). Table 2 presents endpoints available for adding information to the database, ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT namely information resulting from the intelligent data processing methods presented in Section 4. Table 1: Database access endpoints of the REST API. Endpoint Return Data /api/{collection}/tweet/{id_tweet} All information concerning a specific tweet /api/{collection}/tweet/page/{id_page} Set of 1K tweets ordered by decreasing publishing date /api/{collection}/tweet/hour/{hour} Tweets collected per hour /api/{collection}/tweet/day/{day} Tweets collected per day /api/{collection}/tweet/weekDay/{weekDay} Tweets collected per week day /api/{collection}/tweet/month/{month} Tweets collected per month /api/{collection}/tweet/year/{year} Tweets collected per year /api/{collection}/query/{query} Set of tweets according to the filter specified in parameter “query” /api/user/{id_user} All tweets produced by user /api/user/{id_user}/firstProfile User profile when first tweet included in the database was published /api/user/{id_user}/lastProfile User profile when his last tweet included in the database was published /api/user/{id_user}/ageGender Fields of user profile used to infer age and gender Table 2: REST API endpoints for saving “intelligent analysis” results into MongoDB. Endpoint Return Data /api/t2f2tweets/{topic} Tweets related to a given topic (processed using Twitter Topic Detection) /api/popusers/{topic} Most relevant users on a given topic of discussion 3.4 Data Visualization The Data Visualization module consists on a web dashboard integrating several data indicators. The dashboard is implemented using the Google Charts API8 and the REST API presented in the previous section. The dashboard includes charts and statistics for geolocated tweets, timeline tweets and user information. 8 https://developers.google.com/chart/?csw=1 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Since some of the statistics and charts involve a large volume of data, and the processing of the respective queries is not feasible in real time, such queries are therefore pre- processed automatically on a daily basis. In such cases, the visualized information refers to data collected up to the previous day. Figure 3 shows an example of such queries, where information concerning geolocated tweets is visualized. In the top center of the screen it is shown the total number of tweets and the average number of collected tweets per day. The top graph shows the evolution in the number of collected tweets per day during the analyzed period (the y-scale is in millions of tweets). The bottom 4 graphs show (left-to-right, top-to-bottom): Histogram of tweets per hour of the day (%); Histogram of tweets per day of the week (%); Number of tweets per day during the last month (millions); Number of tweets per month during the last 16 months (millions). It is also possible to query and visualize in “real time”. For example, Figure 4 shows the dashboard for the location of geolocalized tweets collected on May 31st, 2015 (distributed per 6 hour periods). Figure 3: Dashboard containing indicators about collected geolocated tweets ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Figure 4: Visualization of collected geolocated tweets in May, 31st, 2015 4 Intelligent twitter topic mining add-on The framework described in the previous section allows the collection, management and visualization of an extensive tweet corpus. In this section we address how we add to the framework the capabilities to intelligently retrieve relevant information from the stored corpus. Despite the underneath complexity, the process is designed to be easily accessible to users, regardless of specialization and degree of technical knowledge. The process is represented in Figure 5, succinctly described as follows, and detailed in the next subsections. Figure 5: Intelligent Twitter topic mining (FUWS – Fuzzy Uke Word Similarity algorithm; TD – Topic Detection task; PR - PageRank for User Influence task; SA – Sentiment Analysis task). ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT All interaction with the user occurs in the “User interface” layer. A “Twitter Topic Fuzzy Fingerprints” layer is used to process user inputs, interact with the REST API, process the data returned from MongoDB and present the results to the user. When a user wants to retrieve information relevant to a given topic of interest, it must input the search period (begin date, end date) and any keywords/#hashtags using the available web interface (Figure 6-1). The system will then use a fuzzy word similarity algorithm (FUWS – see section 4.2) to search the MongoDB for similar keywords and #hashtags found in the database during the referenced time period. (Figure 6-2). As a result of this step, all tweets containing such keywords/#hashtags are retrieved and stored in a temporary collection, and an output list of all similar keywords/#hashtags is returned and shown to the user. The user can prune the list and indicate which of the returned keywords and hashtags might or might not be useful within the context of the topic to be analyzed. The pruned keywords/hashtags list is then passed back to the Twitter Topic Fuzzy Fingerprints layer, where it is used to create a topic Fuzzy Fingerprint (see section 4.1) based on the tweets stored in the temporary collection. This fingerprint is then used to find tweets in MongoDB that are related to the topic of interest (Figure 6-3). The set of relevant tweets is written back into a separate collection in MongoDB. Tweets are considered relevant to the topic if they match the fingerprint to a given degree. It should be noted that the method can find relevant tweets even when they do not contain any of the items present in the pruned list of keywords and/or #hashtags. A user only needs to provide a single relevant #hashtag since the method is able to create the topic fingerprint based on contents of the tweets found after applying the FUWS. The relevant tweets are then processed in order to find the top-20 most influential users in propagating the topic in study, (see section 4.3), and sentiment analysis is performed (see section 4.4). The resulting data is also written back into MongoDB in separate collections. A Results webpage (Figure 7), containing relevant information and automatically obtained by querying the above-mentioned collections, is presented to the user as a result of the process. The results presented area can be further detailed, and the collections can be queried using any of the developed REST API commands (in a user friendly way). ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Figure 6: Intelligent topic mining steps when viewed from the user dashboard ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Figure 7: Intelligent topic mining results’ visualization 4.1 Twitter Topic Fuzzy Fingerprints In order to perform tweet topic detection, we used the Twitter Topic Fuzzy Fingerprints method (Rosa et al., 2014a, 2014b, 2014c), which adapts the original Fuzzy Fingerprints method to the characteristics of individual tweets. As described in section 2.2, textual fuzzy fingerprinting works by comparing the similarity of the fuzzy fingerprint of an individual text, with the fingerprint of a possible author (created based on a set of texts written by that author). When using Fuzzy Fingerprints to detect if a tweet is related to a given topic, we start by creating the fingerprint of the topic (instead of an author), plus the fingerprint of a set of trending topics. A topic fingerprint is created based on a set of training tweets known to be related to the topic in question. All tweets containing #hashtags related with the topic are retrieved from the database, and their text is preprocessed to remove unimportant information for this task, such as, for example, words with less than 3 characters, as studied in (Rosa et al., 2014a, 2014b). The next step consists in obtaining the word frequency in the processed tweets. Only the top-k words are considered for the fingerprint. The computation of the top-k words and respective frequency is done using an approximated counting method, FSS – Filtered Space Saving (Homem, 2010) for efficiency reasons. This process is repeated for all the trending topics and the topic to be detected. The next step differs from the original method: we account for the Inverse Class Frequency of each word, icfv (1), which is an adaptation of the well-known Inverse Document Frequency (idf), to reorder the top-k word lists of all topics. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT , (1) In (1), J is the cardinality of all considered topics, and Jv is the number of topics where word v is present. The product of the frequency of each word v with its icfv, is used to distinguish the occurrence of common words and get a new ordered k-sized rank for each #hashtagged topic. The next step consists in fuzzifying each top-k list in order to obtain the fingerprint for each #hashtagged topic. A membership value is assigned to each word of the top-k list based on its order (instead of its frequency or its icf). In MISNIS we used the fuzzifying function represented in (2), a Pareto-based linear function, where 20% of the top k words assume 80% of the membership degree: { ( ) ( ) (2) where μji is the membership value of the ith top word in topic j, and k is a constant (the size of the fingerprint), The fingerprint of topic j is a size-k fuzzy vector, where each position contains an element vji (in this approach vji is a word of topic j), and a membership value μji representing the fuzzified value of the rank of vji (the membership of the rank), obtained by the application of (2). Formally, topic j is represented by its size-k fingerprint Φj (3) {( ) ( ) ( )} (3) The set of all computed topic fingerprints constitutes the fingerprint library. Once the fingerprint library is created, it is possible to look in MongoDB for tweets that, despite not containing an intended #hashtag, discuss the same topic. In the original fingerprint method (Homem, 2011), this would be done by computing the fingerprint of an individual tweet and comparing it to the topic fingerprint. However, knowing that a text fingerprint is essentially based on the fuzzification of the order of the word frequency of the text, and that tweets have a maximum of 140 characters, it is not possible (or useful) to create the fingerprint on an individual tweet, since within a tweet, very few relevant words (if any!), are repeated. As such, a new similarity score, the Tweet to Topic Similarity Score (T2S2) was developed to test for the similarity between a given tweet and a topic fingerprint (4). The T2S2 score does not take into account the size of the text to be classified (i.e., its number of words), hence avoids the problem of fuzzy fingerprint similarity computation for short texts. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT In (4), Φj is the fingerprint of #hashtagged topic j, T is the set of distinct words in the preprocessed tweet text, { } is the set of words of Φj, and μji is the membership degree of word vji in the fingerprint Φj. Essentially, T2S2 sums the membership value of every word v that is common between the tweet and the #hashtag j fingerprint, and normalizes this value by dividing it with the sum of the top-x membership values of Φj, being x the minimum between k and the cardinality of T. T2S2 approaches 1 if most to all features of the tweet belong to the top words of the fingerprint. T2S2 tends to 0 when there are no common words between the tweet and the fingerprint, or when the few common words are in the bottom of the fingerprint. When a given tweet has a T2S2 score with #topic j above a given threshold, then it is considered relevant to the topic. Such tweets are retrieved from the database. In the MISNIS platform we use the following parameters when performing topic detection: fingerprint size k=20, words with less than 3 characters removed during preprocessing (no stopwords are removed and no stemming is performed). Tweets with a T2S2 score above 0.10 are retrieved. Stemming is not performed since even though it gives marginal gains in Precision and Recall, there is a high penalty in execution time. Contrary to other tasks, stopwords do not hinder performance, so they are kept. The choice and optimization of the preprocessing steps are detailed in (Rosa et. al 2014, 2014a, 2014b). A real word case study based on a 2011 London Riots dataset has shown the Twitter Topic Fuzzy Fingerprints was able to retrieve 40% more relevant tweets (with an F- measure estimated to be 0.95<F1<1), than simply using #hashtags and common kewyords (Carvalho et al., 2017). 4.2 Fuzzy Uke Word Similarity (FUWS) Obtaining a successful topic fingerprint depends obviously on the data used to create it. The most usual approach to obtain data to create a topic fingerprint consists in using tweets that are undoubtedly associated with the intended topic, i.e., those that contain an #hashtag created to discuss that topic. However, since there are no rules governing #hashtag creation or #hashtag use in Twitter, anyone can create and use any #hashtag referring to a given topic. It is common that many #hashtags are used to refer the same topic (even if some #hashtags naturally end up being more popular than others). It is therefore of interest to have a method that can help finding similar #hashtags in the database in order to extend the dataset of tweets used to create the fuzzy fingerprint. Another characteristic of Twitter, is the increasingly higher use of mobile devices such as smartphones. The habit of tweeting using virtual keyboards is usually associated with a relevant number of “thumbos”, i.e., typing errors (aka “typos”) due to thumbs hitting the wrong virtual key due to the small (virtual) keys and lack of physical key feedback. Not even the use of automatic word correction prevents a high number of word errors in tweets’ text. As a result, many tweets relevant to a given topic might not be retrieved when looking for specific keywords because the keywords might contain errors. This is even more likely in the case of #hashtags, since word correction usually does not affect ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT them. Therefore, it is also of interest to detect such errors when creating the dataset of tweets used to create the fuzzy fingerprint. Detecting similar #hashtags and/or word errors can be accomplished using string similarity techniques. Current research on string similarity offers a panoply of measures that can be used in this context, such as the ones based on edit distances (Levenshtein 1966) or on the length of the longest common subsequence of the strings. However, most of the existing measures have their own drawbacks. For instance, some do not take into consideration linguistically driven misspellings, others the phonetics of the string or the mistakes resulting from the input device. In fact, edit distances, even if universally used for online word typing correction are not adequate at all for the kind of problem we are approaching, since they do not have a strong discriminative power. In MISNIS we used the Fuzzy Uke Word Similarity (FUWS) (Carvalho and Coheur, 2013) to detect #hashtag/keyword similarity. This word similarity function combines the most interesting characteristics of the two main philosophies in word and string matching (edit distance and common subsequence of strings), and by integrating specific fuzzy based expert knowledge concerning typographical errors (like for example taking into consideration which keys are closer to others and are more likely to appear in “thumbos”), can achieve a good discrimination. The similarity threshold value used to process word and #hashtag similarity in MISNIS is 0.67, as suggested by Carvalho and Coheur (2013). 4.3 User Influence In order to assert user influence for a given Twitter topic (using the tweets obtained using the Fuzzy fingerprints), and considering the definitions of influence discussed in Section 2.3, we proposed and implemented a graph representation of user's influence based on mentions (Rosa et. al, 2015). In the proposed user influence representation, whenever a user is mentioned in a tweet's text using the @user tag, a link is made from the creator of the tweet, to the mentioned user:  When @userA tweets “Do you think we can we get out of this financial crisis, @userB?”, the link @userA->@userB is created. This is also true for re-tweets:  The tweet “RT @userC The crisis is everywhere!” from @userA, creates the link: @userA->@userC. In (Rosa et. al, 2015), an empirical analysis showed that, in the context of Twitter User Influence, PageRank (Page 1998, 1999) outperforms other well-known network centrality algorithms, in particular, Katz (1953). As such, PageRank was chosen for the implementation of determining user relevance within MISNIS. PageRank parameterization consists in deciding the damping factor d (see section 2.3). The true value that Google uses as damping factor is unknown, but it has become common to use d=0.85 in the literature. A lower value of the damping factor implies that the graph's structure is less respected, therefore making the ''walker'' more random and less strict. After several experiments we opted to use d=0.85 within the platform. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 4.4 Sentiment Analysis Tweets are a form of short informal texts that may contain misspellings, slang terms, shortened word forms, elongations, leet speech, hashtags, and many other specific phenomena that pose new challenges to Sentiment Analysis (Kiritchenko et al., 2014). Given the limited amounts of annotated data for certain languages, supervised learning is often not possible, and alternate approaches using either manual or automatic sentiment lexicons are often applied (Kiritchenko et al., 2014). In order to take this into account and to maintain language independence, our approach to Sentiment Analysis follows closely the idea explored by (Go, 2009), that consists of using emoticons, abundantly available on tweets, to automatically label the data and then use such data to train machine learning models. In order to build our sentiment models, we have used the knowledge flow described in http://markahall.blogspot.co.nz/2012/03/sentiment- analysis-with-weka.html (retrieved in September, 2015) that uses weka (Hall et al., 2009) to automatically label the training tweets based on emoticons. This approach has the huge advantage of being easily applied to different languages, for which manual labels are scarce or non-existing, as is the case of Portuguese language. We have then adopted an approach based on logistic regression, which corresponds to the maximum entropy (ME) classification for independent events (Berger, 1996), to create our sentiment models, based on the previously labelled data. The ME models used in this study were trained using the MegaM tool (Daume, 2004), which uses an efficient implementation of conjugate gradient (for binary problems). Finally, the MegaM models were used thought an interface available from the NLTK toolkit9. Our Portuguese language models were trained using around 200k tweets, and achieved an accuracy between 69% and 71% for the “Positive”/ “Neutral” / “Negative” classification problem. Moreover, based on the produced model, we can automatically derive a new automatic lexicon that can be used for alternative future classification approaches. 5 Application example Here we present an execution of the MISNIS framework, applied to a real case collected from the Portuguese database of tweets in MongoDB. On the night from 21st to the 22nd of November of 2014, former Portuguese Prime Minister José Sócrates was arrested under suspicion of fraud and money laundering during his time in office. This piece of news was the headline of all media outlets for several days and, to this day, remains quite an important story to follow. Naturally, Twitter was no exception, which made this an interesting story to analyze. The following data was inputted into the interface webforms:  Start Date: 20th November 2014 00:00:00  End Date: 23th November 2014 23:59:59  Keywords: socrates, freesocas With this input, the framework generated the screen shown in Figure 7. Out of a total of 3,309,468 tweets in that time interval, 380 were identified to the inputted or other similar keywords found by the system (including the hashtags #socrates and #freesocas). These 380 tweets were used to train the Twitter Topic Fuzzy Fingerprint method along with 15 other top trends existing within the full three million 9 http://www.nltk.org/_modules/nltk/classify/megam.html ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT tweets. A total of 29,019 tweets containing the 15+1 trends were used as the training set. As an output, the procedure created a dataset of 10,687 tweets regarding José Socrates' arrest (including additional sentiment information per tweet), produced several indicators regarding topic evolution through time, and listed the top 20 users in discussing and forwarding the topic (Table 3). The first retrieved true positive tweet occurs on November 22nd, at 00:08:51, just a few minutes after the arrest occurred. This is one more example of the relevance of Twitter in current news events. The Precision (true positive rate) of the retrieved results is 0.97, with many of the false positives (i.e., tweets that are not related to José Sócrates’ arrest) retrieved on the 48 hours before the arrest (November 20-21st). It was not possible to calculate the Recall since it would be necessary to manually check 3 Million tweets one by one to count the false negatives (this would obviously defeat the purpose of the expert system), but previous tests using the same parameters on other databases have shown that Fuzzy Fingerprint Recall is usually slightly higher than Precision (Rosa et al. 2014a). It should be noted that, at the time, the database did not contain tweets from Twitter users that never produced geolocated contents. As such, there were no tweets collected from the major media players such as TV stations and newspapers. Although 10,687 tweets out of a universe of 3,309,468 tweets, seems a low number, this can be explained by the absence in the database of the major media players, and also by the fact that the majority of Twitter users in Portugal are young teenagers (Brogueira, 2014a, 2014b) that usually do not show any special interest in political matters. Of special relevance is the fact that from 380 hashtagged tweets, the framework retrieved more than 10K tweets that could otherwise go unnoticed among more than 3 million tweets (with a very high Precision). Amongst the top users, it is rather interesting the high incidence of Portuguese comedians, namely @niltoncomedy (ranked 5th), @raminhoseffect (ranked 11th), and @omalestafeito (ranked 13th), and also a few humor oriented internet personalities, such as @avo_idalina (ranked 4th), who is a fictional grandmother character, @miguelluzp (ranked 15th) and @oconguito (ranked 19th), both famous YouTuber teenagers. A proper sociological analysis on the retrieved results is available in (Rebelo et. al, 2016). ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Table 3: Top-20 most influential users on Twitter during the early days of the Sócrates prison arrest events (excluding major media players). Ranking (Page Rank) User PageRank Weight 1 @afonsotorrao 0.00701 2 @likerabos 0.00418 3 @sergiohatesyou 0.00356 4 @avo_idalina 0.00295 5 @niltoncomedy 0.00291 6 @tiagosamuel69 0.00219 7 @marinhskilla 0.00209 8 @pedromvfranco 0.00206 9 @nandinholuz 0.00167 10 @unkndeath2002 0.00158 11 @raminhoseffect 0.00155 12 @im_a_barbieee 0.00143 13 @omalestafeito 0.00141 14 @issodepende 0.00134 15 @miguelluzp 0.00127 16 @inaaraujo99 0.00126 17 @rubencgomes 0.00125 18 @pedroboucherie 0.00124 19 @oconguito 0.00109 20 @daspthebest 0.00109 6 Conclusions Social networks and social networking are here to stay. This is not a controversial or novel statement: independently from how one feels towards adopting the use of social networks, no one can deny their importance in current modern world society. From event advertising or idea dissemination, to commenting and analysis, social networks have become the de facto means for individual opinion making and, consequently, one of the main shapers of an individual’s perception of society and the world that surrounds her/him. Despite the undeniable importance of social networks, too many questions concerning their effect in society are yet to be properly addressed. What makes events become important in social networks? Why and how they become important? How long does it take for an event to make an impact in social networks and society? Can social networks give more importance to an event than it really deserves, i.e., are social networks becoming a factor by themselves? What is the role of social networks’ major actors (important journalists, bloggers, commentators, politicians, etc.) in the propagation of such events? Are such actors in the origin of the events or mere catalysts to the observations of minor role players? In this article we presented a framework, MISNIS, that can help answering such questions:  The framework enables the identification and traction of important events (topics) and of key actors within those topics, as well as identifies their origin and propagation timeline. Measures and indicators to characterize events are provided by the framework contributing with essential information to understand the social network phenomena and its importance in current world society. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT  MISNIS addresses the issues of collecting, storing, managing, mining and visualizing Twitter data. It applies well-known and novel techniques in the fields of Computational Intelligence, Information Retrieval, Big Data, Topic Detection, User Influence and Sentiment Analysis, to social networks Data Mining, in particular, Twitter.  MISNIS can be used as an expert system by social scientists, sociologists, or any other users to retrieve relevant data to study social networks’ impact in society without requiring any computer technical expertise. The framework is currently operational, even if on a prototype form with limited access. External access to its functionalities can be made available by request to the authors. Future work includes:  Expand and facilitate the access to the platform (direct external access via user account);  Expand the platform to use other public social networks, such as public blogs, webpages, public Facebook profiles, etc., as additional sources of data;  Expand multilinguality and region use. The platform is currently focused on portuguese language tweets posted in Portugal, but most developed technologies are language independent and can be presently used for tweets in most languages (e.g. intelligent tweet retrieval, sentiment analysis, etc.). However, some collection and data expansion, and the geo-location mechanisms are either region or language dependent, and should be adapted. Currently we are working on English language tweets in UK and Ireland and in the USA;  Sentiment analysis methods can and should be improved. Even though we opted for a language independent mechanism, it is possible to improve sentiment analysis by combining it with language dependent lexicon. We also find the polarity sentiment approach very limiting, and would like to include more elaborate approaches;  The platform is very dependent on the use of several Twitter and Google APIs: most changes to the APIs endpoints imply changing and recompiling the platform code. We would like to improve the code architecture to allow for external dynamical API changes: the platform administrator would simply need to update the API access data without the need to recompile the code. Acknowledgment This work was supported by national funds through FCT Fundação para a Ciência e a Tecnologia, under project PTDC/IVC-ESCT/4919/2012 and project PEstOE/EEI/LA0021/2013. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT References Baccianella, S., Esuli, A., and Sebastiani, F. (2010). Sentiwordnet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining. In Proc. of LREC’10, Valletta, Malta. ELRA. Berger, A. L., Pietra, S. D., and Pietra, V. D. (1996). A maximum entropy approach to natural language processing. Computational Linguistics, 22(1):39–71. Blei, David M. (2012). Probabilistic Topic Models. Communications of the ACM, 55(4):77– 84. Blei, David M., Ng, Andrew Y., and Jordan, Michael I. (2003). Latent Dirichlet Allocation. Journal of Machine Learning Research, 3:993–1022. Brogueira, G., Batista, F., and Carvalho, J. P. (2015a). Arquitetura e desenvolvimento de um repositório de tweets em português europeu. In 5as Jornadas de Informtica da Universidade de vora - JIUE 2015. Springer. Brogueira, G., Batista, F., and Carvalho, J. P. (2015b). Sistema inteligente de recolha, armazenamento e visualização de informação proveniente do twitter. In Conferência da Associação Portuguesa de Sistemas de Informação, CAPSI 2015. Brogueira, G., Batista, F., and Carvalho, J. P. (2015c). Using geolocated tweets for characterization of portuguese administrative regions. In 18th AGILE International Conference on Geographic Information Science. Brogueira, G., Batista, F., Carvalho, J. P., and Moniz, H. (2014a). Expanding a database of portuguese tweets. In SLATE’14 3rd Symposium on Languages, Applications and Technologies, volume 4569 of OpenAccess Series in Informatics (OASIcs), pages 275– 282. Schloss Dagstuhl. Brogueira, G., Batista, F., Carvalho, J. P., and Moniz, H. (2014b). Portuguese geolocated tweets: an overview. In ISDOC2014 - Proceedings of the International Conference on Information Systems and Design of Communication, pages 178–179. ACM. Brogueira, G., Batista, F., Carvalho, J.P. (2016). A Smart System for Twitter Corpus Collection, Management and Visualization, International Journal of Technology and Human Interaction (IJTHI), IGI Global, vol. 13, n. 3, December 2016 C. Rebelo, I. Pereira, H. Rosa, F. Batista, J.P. Carvalho, “The news will be tweeted: multiple uses of Twitter around a major political event”, submitted to New Media and Society. Carvalho, J. P. and Coheur, L. (2013). Introducing UWS - A fuzzy based word similarity function with good discrimination capability: Preliminary results. In FUZZ-IEEE, pages 1–8. Carvalho, J. P., Pedro, V., and Batista, F. (2013). Towards intelligent mining of public social networks’ influence in society. In IFSA World Congress and NAFIPS Annual Meeting (IFSA/NAFIPS), pages 478 – 483, Edmonton, Canada. Carvalho, J.P., Rosa, H. and Batista, F. (2017). Detecting relevant tweets in very large tweet collections: the London Riots case study. InFUZZ-IEEE, 2017 IEEE International Conference on Fuzzy Systems, Jul, 2017, Naples, Italy Cataldi, M., Di Caro, L., and Schifanella, C. (2010). Emerging topic detection on twitter based on temporal and social terms evaluation. In Proceedings of the Tenth International Workshop on Multimedia Data Mining, MDMKDD ’10, pages 4:1–4:10, New York, NY, USA. ACM. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Cha, M., Haddai, H., Benevenuto, F., and Gummadi, K. P. (2010). Measuring User Influence in Twitter: The Million Follower Fallacy. International AAAI Conference on Weblogs and Social Media, pages 10–17. Chen Y., Conroy N.J. and Rubin, V. L. (2015). News in an online world: the need for an "automatic crap detector". In Proceedings of the 78th ASIS&T Annual Meeting: Information Science with Impact: Research in and for the Community (ASIST '15). American Society for Information Science, Silver Springs, MD, USA, Article 81 , 4 pages. Cigarrán J., Castellanos A., García-Serrano A. (2016), A step forward for Topic Detection in Twitter: An FCA-based approach, Expert Systems with Applications, Volume 57, 2016, Pages 21-36, ISSN 0957-4174, http://dx.doi.org/10.1016/j.eswa.2016.03.011. Culotta, A. (2010). Towards detecting influenza epidemics by analyzing twitter messages. In Proceedings of the First Workshop on Social Media Analytics, SOMA ’10, pages 115–122, New York, NY, USA. ACM. Das, S. and Chen, M. (2001). Yahoo! for Amazon: Extracting market sentiment from stock message boards. In Proceedings of the Asia Pacific Finance Association Annual Conference (APFA). Daume III, H. (2004). Notes on CG and LM-BFGS optimization of logistic regression. http://hal3.name/megam/. Dehkharghani, R., Mercan, H., Javeed, A., and Saygin, Y. (2014). Sentimental causal rule discovery from twitter. Expert Systems with Applications, 41(10):4950 – 4958. Domingos, P. and Richardson, M. (2001). Mining the network value of customers. In Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’01, pages 57–66, New York, NY, USA. ACM. Feldman, R. and Sanger, J. (2006). Text Mining Handbook: Advanced Approaches in Analyzing Unstructured Data. Cambridge University Press, New York, NY, USA. Ford, R., (2011). Hollywood Reporter [Online]. Available at: http://www.hollywoodreporter.com/news/earthquake-Twitter-users-learnedtremors- -226481 [Accessed 30/7/2017] Gerber, M. S. (2014). Predicting crime using twitter and kernel density estimation. Decision Support Systems, 61(0):115 – 125. Go, A., Bhayani, R., and Huang, L. (2009). Twitter Sentiment Classification using Distant Supervision. Technical report, Stanford University. Gupta, M. Li, R. Chang, K. (2014). Towards a Social Media Analytics Platform: Event Detection and Description for Twitter – a Tutorial. 23rd International WWW Conference. [Online] Available at: http://www2014.kr/asset/slide/Towards%20a%20Social%20Media%2 0Analytics%20Platform.pdf, [Accessed 11/ 1/ 2017] Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., and Witten, I. H. (2009). The weka data mining software: An update. SIGKDD Explor. Newsl., 11(1):10–18. Hoffman, M. D., Blei, D. M., and Bach, F. R. (2010). Online learning for latent dirichlet allocation. In Lafferty, J. D., Williams, C. K. I., Shawe-Taylor, J., Zemel, R. S., and Culotta, A., editors, NIPS, pages 856–864. Curran Associates, Inc. Homem, N. and Carvalho, J. P. (2010). Finding top-k elements in data streams. Inf. Sci., 180(24):4958–4974, Elsevier. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Homem, N. and Carvalho, J. P. (2011). Authorship identification and author fuzzy fingerprints. In 30th Annual Conference of the North American Fuzzy Information Processing Society, NAFIPS2011. Hu, M. and Liu, B. (2004). Mining and summarizing customer reviews. In Proc. of the 10th ACM SIGKDD int. conf. on Knowledge discovery and data mining, KDD ’04, pages 168–177. ACM. Kasiviswanathan, S. P., Melville, P., Banerjee, A., and Sindhwani, V. (2011). Emerging topic detection using dictionary learning. In Proceedings of the 20th ACM International Conference on Information and Knowledge Management, CIKM ’11, pages 745–754, New York, NY, USA. ACM. Katz, L. (1953). A new status index derived from sociometric analysis. Psychometrika, 18(1):39–43. Kim, S.-M. and Hovy, E. (2004). Determining the sentiment of opinions. In Proc. of COLING ’04. ACL. Kiritchenko, Svetlana, Zhu, Xiaodan, and Mohammad, Saif M. (2014). Sentiment analysis of short informal texts. Jounal of Artificial Inteligence Research, 50, 1, 723-762. Kontopoulos, E., Berberidis, C., Dergiades, T., Bassiliades, N., (2013). Ontology-based sentiment analysis of twitter posts. Expert systems with Applications, 40(10):4065– 4074, Elsevier. Kumar, S., Morstatter, F., and Liu, H. (2013a). Twitter Data Analytics. Springer, New York, NY, USA. Kumar, S., Morstatter, F., Zafarani, R., and Liu, H. (2013b). Whom should I follow?: Identifying relevant users during crises. In Proceedings of the 24th ACM Conference on Hypertext and Social Media, HT ’13, pages 139–147, New York, NY, USA. ACM. Lachlan, K. A., Spence, P. R., and Lin, X. (2014). Expressions of risk awareness and concern through twitter: On the utility of using the medium as an indication of audience needs. Computers in Human Behavior, 35(0):554–559. Leavitt, A., Burchard, E., Fisher, D., and Gilbert, S. (2009). The influentials: New approaches for analyzing influence on twitter. Lee, K., Palsetia, D., Narayanan, R., Patwary, M. M. A., Agrawal, A., and Choudhary, A. (2011). Twitter trending topic classification. In Proceedings of the 2011 IEEE 11th International Conference on Data Mining Workshops, ICDMW ’11, pages 251–258, Washington, DC, USA. IEEE Computer Society. Levenshtein, V. (1966). Binary codes capable of correcting deletions, insertions, and reversals” Soviet Physics Doklady, 1966. 10:707–710. Marcus, A., Bernstein, M. S., Badar, O., Karger, D. R., Madden, S., and Miller, R. C. (2011a). Tweets as data: Demonstration of tweeql and twitinfo. In Proceedings of the 2011 ACM SIGMOD International Conference on Management of Data, SIGMOD ’11, pages 1259– 1262, New York, NY, USA. ACM. Marcus, A., Bernstein, M. S., Badar, O., Karger, D. R., Madden, S., and Miller, R. C. (2011b). Twitinfo: Aggregating and visualizing microblogs for event exploration. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI ’11, pages 227– 236, New York, NY, USA. ACM. Mathioudakis, M. and Koudas, N. (2010). Twittermonitor: Trend de- tection over the twitter stream. In Proceedings of the 2010 ACM SIGMOD International Conference on Management of Data, SIGMOD ’10, pages 1155–1158, New York, NY, USA. ACM. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Mazzia, A. and Juett, J. (2010). Suggesting hashtags on twitter. Master’s thesis, University of Michigan. Mendoza, M., Poblete, B., and Castillo, C. (2010). Twitter under crisis: Can we trust what we rt? In Proceedings of the First Workshop on Social Media Analytics, SOMA ’10, pages 71–79, New York, NY, USA. ACM. Mustafaraj E. and Metaxas P.T. (2017). The Fake News Spreading Plague: Was it Preventable?. In Proceedings of the 2017 ACM on Web Science Conference (WebSci '17). ACM, New York, NY, USA, 235-239. DOI: https://doi.org/10.1145/3091478.3091523 Oussalah, M., Bhat, F., Challis, K., and Schnier, T. (2013). A software architecture for twitter collection, search and geolocation services. Knowledge-Based Systems, 37(0):105 – 120. Page, L., Brin, S., Motwani, R., and Winograd, T. (1998). 1 Introduction and Motivation 2 A Ranking for Every Page on the Web. World Wide Web Internet And Web Information Systems, 54(1999-66):1–17. Page, L., Brin, S., Motwani, R., and Winograd, T. (1999). The pagerank citation ranking: Bringing order to the web. Pang, B., Lee, L., and Vaithyanathan, S. (2002). Thumbs up? Sentiment classification using machine learning techniques. In Proceedings of EMNLP, pages 79–86. Paul, M. J. and Dredze, M. (2011). You are what you tweet: Analyzing twitter for public health. In Adamic, L. A., Baeza-Yates, R. A., and Counts, S., editors, ICWSM. The AAAI Press. Perera, R., Anand, S., Subbalakshmi, K., and Chandramouli, R. (2010). Twitter analytics: Architecture, tools and analysis. In MILITARY COMMUNICATIONS CONFERENCE, 2010 - MILCOM 2010, pages 2186–2191. Phuoc, N. Q., Kim, S.-R., Lee, H.-K., and Kim, H. (2009). Pagerank vs. Katz status index, a theoretical approach. In Proceedings of the 2009 Fourth International Conference on Computer Sciences and Convergence Information Technology, ICCIT ’09, pages 1276– 1279, Washington, DC, USA. IEEE Computer Society. Qu, Y., Huang, C., Zhang, P., and Zhang, J. (2011). Microblogging after a major disaster in china: A case study of the 2010 yushu earthquake. In Proceedings of the ACM 2011 Conference on Computer Supported Cooperative Work, CSCW ’11, pages 25–34, New York, NY, USA. ACM. Razis, G. and Anagnostopoulos, I. (2014). Influencetracker: Rating the impact of a twitter account. CoRR. Rogers, E. M. (1962). Diffusion of innovations. Rosa, H. (2014). Topic detection within social networks. Master’s thesis, Instituto Superior Técnico, Universidade de Lisboa. Rosa, H., Batista, F., and Carvalho, J. P. (2014a). Twitter topic fuzzy fingerprints. In WCCI2014, FUZZ-IEEE, 2014 IEEE World Congress on Computational Intelligence, International Conference on Fuzzy Systems, IEEE Xplorer, pages 776–783, Beijing, China. Rosa, H., Carvalho, J. P., and Batista, F. (2014b). Detecting a Tweet’s Topic within a Large Number of Portuguese Twitter Trends. In Pereira, M. J. V., Leal, J. P., and Simes, A., editors, 3rd Symposium on Languages, Applications and Technologies, volume 38 of OpenAccess Series in Informatics (OASIcs), pages 185–199, Dagstuhl, Germany. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Rosa, H., Carvalho, J. P., Astudillo, R., and Batista, F. (2015). Detecting user influence in twitter: Pagerank vs katz, a case study. 7th European Symposium on Computational Intelligence and Mathematics. Saha, A. and Sindhwani, V. (2012). Learning evolving and emerging topics in social media: A dynamic nmf approach with temporal regularization. In Proceedings of the Fifth ACM International Conference on Web Search and Data Mining, WSDM ’12, pages 693–702, New York, NY, USA. ACM. Sakaki, T., Okazaki, M., and Matsuo, Y. (2010). Earthquake shakes twitter users: Real- time event detection by social sensors. In Proceedings of the 19th International Conference on World Wide Web, WWW ’10, pages 851–860, New York, NY, USA. ACM. Sankaranarayanan, J., Samet, H., Teitler, B. E., Lieberman, M. D., and Sperling, J. (2009). Twitterstand: News in tweets. In Proceedings of the 17th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, GIS ’09, pages 42–51, New York, NY, USA. ACM. Santos, C. J. and Matos, S. (2013). Predicting flu incidence from portuguese tweets. In International Work-Conference on Bioinformatics and Biomedical Engineering 2013. Proceedings. Scanfeld, D., Scanfeld, V., and Larson, E. L. (2010a). Dissemination of health in- formation through social networks: Twitter and antibiotics. American Journal of Infection Control, 38(3):182–188. Stone, P., Dunphy, D., Smith, M., and Ogilvie, D. (1966). The General Inquirer: A Computer Approach to Content Analysis. MIT Press. Turney, P. D. (2002). Thumbs Up or Thumbs Down? Semantic Orientation Applied to Unsupervised Classification of Reviews. In Proc. of the 40th Annual Meeting on Association for Computational Linguistics, pages 417–424. ACL. Twitter (2010). To trend or not to trend. https://blog.twitter.com/2010/trend-or-not- trend. Accessed: 2014-03-28. Vosoughi S., Mohsenvand M., and Roy D. (2017). Rumor Gauge: Predicting the Veracity of Rumors on Twitter. ACM Trans. Knowl. Discov. Data 11, 4, Article 50 (July 2017), 36 pages. DOI: https://doi.org/10.1145/3070644 
work_6wqaopq7qjhppbpy3n7wzspbay	----	CCC 2018 Proceedings of the Creative Construction Conference (2018) Edited by: Miroslaw J. Skibniewski & Miklos Hajdu DOI 10.3311/CCC2018-007 Creative Construction Conference 2018, CCC 2018, 30 June - 3 July 2018, Ljubljana, Slovenia Evaluation of the construction project success with use of neural networks Alexey Bulgakowa*, Georgii Tokmakovb, Jens Otto c, Katharina Langoschd aSouthwest State University, a.bulgakow@gmx.de bSouth-Russian State Polytechnic University, tokmakov.tun@gmail.com cTechnical University Dresden, Germany, jens.otto@tu-dresden.de dTechnical University Munich, Germany, katharina.langosch@br2.ar.tum.de Abstract Construction project success is determined in terms of cost, schedule, performance and safety through many events and resultant interactions, plans, facilities and changes in participants and the environment. In the construction industry there are myriad uncertainties that make management exceedingly complex. Factors for success vary from project to project. Human experts can often achieve a satisfactory project outcome, however, shortfalls nearly always occur due to managers failing to take all relevant factors into consideration, in addition to lacking access to all relevant information. Statistical methods represent a basic approach to identifying significant factors from historical data or questionnaire results. However, the dynamic nature of critical factors means that changes in project conditions must be monitored continuously. Artificial intelligence techniques have a wide range of applications, including monitoring and forecasting of long-term projects; their main advantage is the ability to track and predict trends in changing project implementation factors. In this article, the authors describe the structure and algorithm of the neural network for assessing the success of construction projects, taking into account the individual influence of the initial conditions as well as their combined impact. © 2018 The Authors. Published by Diamond Congress Ltd., Budapest University of Technology and Economics Peer-review under responsibility of the scientific committee of the Creative Construction Conference 2018. Keywords: Construction project success, neural network, artificial intelligence techniques ; 1. Introduction Once a construction project has been proposed, the primary contract is typically subdivided into multiple subcontracts. Large numbers of participants are, therefore, involved in project planning and implementation phases (see Fig. 1.). Expectations can only be met by conducting a comprehensive analysis of participants [1-4]. Project Fig. 1. Basic information and resource flows of construction production 46 http://2018.creative-construction-conference.com/proceedings/ mailto:tokmakov.tun@gmail.com mailto:jens.otto@tu-dresden.de success is determined in terms of cost, schedule, performance and safety through many events and resultant interactions, plans, facilities and changes in participants and the environment. Project managers able to identify key determinants to project success can monitor project performance continuously and make proper decisions based on objective performance predictions related to project success targets. Artificial intelligence, a novel technology for extracting knowledge, is already widely applied to various civil engineering problems, including project management. To predict project performance, an appreciation of critical factors are crucial to achieve of the project objectives successfully. Statistical methods represent a basic approach to identifying significant factors from historical data or questionnaire results. However, the dynamic nature of critical factors means that changes in project conditions must be continuously monitored. In the construction industry, neuro fuzzy intelligent systems are used not only to predict the technical performance of the construction project, but also to compare them with the results of statistical methods of reducing variables [5,6]. 2. Success in construction industry The final objective of construction management is to accomplish construction projects ‘successfully’. From an owner's perspective, the definition of a successful project is one that meets or surpasses budget and schedule expectations. Consequently, a less-than-successful project fails to achieve budgetary and/or schedule expectations [7]. Project managers are responsible for project success, and must monitor project performance continuously in order to take proper corrective actions essential to controlling project progress. Project outcomes are affected by different factors at various points in time. During the course of a project, predicting project outcomes at different stages requires an ability to analyze dissimilar factors [8]. A dynamic prediction methodology is thus required for project managers to monitor project performance on a continuous basis. However, each time, numerous time-dependent variables can affect project outcome. In addition, due to the nature of the construction industry, such variables remain uncertain [9]. While human experts can predict project outcomes based on their personally accumulated experience and knowledge, the significance of such judgments is restricted by their subjective cognitions and/or knowledge limitations. 3. Neural network for evaluation of the construction project success For the proposed neural network, “hybrid” refers to the combination of traditional neural and high order neural networks. The high order neural network is constructed of three layers with a high order connection and a linear connection between the 1st and 2nd layers and 2nd and 3rd layers, respectively. This study extends the use of high order connections for all connection alternatives, i.e., all layer connections can switch between linear and high order formats (see Fig. 2). An HNN neuron is dominated by an alternative of the following equation: Linear connection: 𝑦𝑖 = 𝑓(∑𝑤𝑗𝑖𝑥𝑖 + 𝑏𝑗0); (1) High order connection: 𝑦𝑖 = 𝑓(∏ 𝑥𝑖 𝑝𝑗𝑖 ∗ 1𝑏𝑗0); (2) Activation function: 𝑓(𝑥) = 1 1+𝑒−𝑎𝑥 ; (3) where wij – coupling weight coefficients; yi – output neuron signal; xi – input neuron signal. An activation function 𝑓 uses a sigmoid function with a slope coefficient of 𝑎. Therefore, each layer connection features an attached connection type that represents the corresponding operation selection (see Fig. 2). This implies that each layer depending on the operating mode may change the type of connection between neurons. For example, for a neural network consisting of 3 layers, the following options are possible (linear - L, higher order - HO): L-L; L-HO; HO-L and HO-HO. CCC 2018 Proceedings DOI 10.3311/CCC2018-007 47 Fig. 2. Hybrid Neural Network Structure Fig. 3. Optimization genetic algorithm Start Population formation p(t) Each neuron evolution in p(t) Satisfy the conditions No Selection of p(t) with the creation of p 0 (t) Mutation of p(t) ,p 0 (t) with the creation of p м (t) Each neuron evolution in p 0 (t) and p м (t) Selection of p(t+1) from p(t), p 0 (t) and p м (t) Stop Yes t=t+1 . . . . . . Fuzzification procedure Hidden layers Defuzzification procedure . . . . . . . . . . . . Output Input CCC 2018 Proceedings DOI 10.3311/CCC2018-007 48 4. Optimization genetic algorithm Optimization of the created neural network is possible with the help of the genetic algorithm adaptation [10, 11] (see Fig. 3). Genetic algorithm, which imitates elements of the natural evolution process, were first proposed in Holland [12]. To apply the genetic algorithm to problem optimization, one must identify all essential parameters to determine chromosome length. The chromosome (i.e., one individual) in this study represents neural network parameters: interconnection coefficients, connection types, a slope coefficient of activation function, and network topology (total layers and layer neurons). Fuzzy logic parameters include membership function summit points, membership function widths and defuzzification weights. Genetic algorithm is a method of random search with elements of adaptation, which is based on principles similar to the Darwin’s evolution process of biological organisms. In this case, three types of operations are performed: crossing, mutation, selection. The fitness degree (how the population corresponds to the given task) is defined through the fitness function that can also include penalty functions for violation of additional restrictions on variables. There are various forms of crossing [13]. They make a selection of the fittest specimen, which constitute a parental pair and the crisscrossing of the chromosomal chains takes place, i.e. the descendant line code inherits fragments of codes of parental chromosomes. The selection operator allows the creation of a new population from a set of specimen, which are generated and modified descendants of specimen after mutation. The genetic algorithm is used to adjust the membership functions that are defined within the accuracy of a few changeable parameters, such as triangular, trapezoidal, and radial functions. When simultaneously configuring several membership functions, the parameters of each are coded by their own segment of the chromosome, so that during the process of crossing the code, sharing occurs only between chromosome segments of the same type. For configure the rule base, to each element of the chromosome is assigned an element of this rule base and based on the received encoding is selected a type of genetic operator. Thus, the architecture of the systems for evaluation of the construction project success can be represented as follows (see Fig. 4). Fig. 4. Architecture of the systems for evaluation of the construction project success Conclusions This paper described an evolutionary fuzzy hybrid neural network to enhance project cash flow management. Neural networks and high order neural networks are combined in the developed evolutionary fuzzy hybrid neural network to Input Data; Desired results Fuzzification Procedure Fuzzy logic block The hybrid neural network Defuzzification Procedure Predicted Results Optimization Procedure Membership Function Neuron Network Parameters Defuzzification Parameters Data flow Control flows Functional objects Databases CCC 2018 Proceedings DOI 10.3311/CCC2018-007 49 form a hybrid neural network, which acts as the major inference engine and operates with alternating linear and non- linear neural networks layer connections. Fuzzy logic is employed to sandwich the hybrid neural network between a fuzzification and defuzzification layer. The authors developed this evolutionary fuzzy hybrid neural network to assess construction industry project success by fusing hybrid neural network, fuzzy logic and genetic algorithm. References [1] Bulgakov A.G., Tokmakov G.E. The analysis of the control systems for building site, problems, possible solutions. VII Internationally scientific conferences: Realisation of the European scientists; 17th-25th of July, 2011. Sofia. S. 43-44. [2] Bulgakow A.G., Jehle P.,Tokmakov G. SCM-logistic and mechatronics systems for ensuring the smooth construction process // Innovation in Mechanical Engineering – Shaping the Future: 56-th International Scientific Colloquium (12–16 September 2011: Conference Proceedings /Ilmenau University of Technology).Ilmenau, 2011. [3] V. Sanvido, F. Grobler, K. Parfitt, M. Guvenis, M. Coyle, Critical success factors for construction projects, Journal of Construction Engineering and Management 118 (1) (1992) 94–111. [4] A.F. Griffith, G.E. Gibson, M.R. Hamilton, A.L. Tortora, C.T. Wilson, Project success index for capital facility construction projects, Journal of Performance of Constructed Facilities 13 (1) (1999) 39–45. [5] D.K.H. Chua, P.K. Loh, Y.C. Kog, E.J. Jaselskis, Neural networks for construction project success, Expert Systems with Applications 13 (4) (1997) 317–328. [6] M.E. Georgy, L.M. Chang, L. Zhang, Prediction of engineering performance: a neurofuzzy approach, Journal of Construction Engineering and Management 131 (5) (2005) 548–557. [7] CII, Predictive tools: closing the performance gap, Research Summary, RS107–1, The Construction Industry Institute, Austin, Tex, 1996. [8] J.S. Russell, E.J. Jaselskis, P. Samuel, S.P. Lawrence, Continuous assessment of project performance, Journal of Construction Engineering and Management 123 (1) (1997) 64–71. [9] J.S. Russell, E.J. Jaselskis, P. Samuel, S.P. Lawrence, Continuous assessment of project performance, Journal of Construction Engineering and Management 123 (1) (1997) 64–71. [10] Kureychik V.M. Genetic algorithms. State problems, prospects., Izvestiya RAN, Theory and control system., 1999. № 1. С. 144-160. [11] Rutkovskij L., Pilinskij M. Neural networks, genetic algorithms and fuzzy systems., Gorzachaza Liniza TELEKOM, 2004. [12] J.H. Holland, Adaptation in Neural and Artificial Systems, The University of Michigan Press, Ann Arbor, 1995. [13] Mecheryakov V.A. Recurrent algorithm neural identification working process earthmoving machinery. Siberian State Automobile and Highway Academy, Omsk, 2007.- S. 63-66 CCC 2018 Proceedings DOI 10.3311/CCC2018-007 50 
work_6xmmkpbyhjeerly5wooinxk3qa	----	MERIT-Infonomics Research Memorandum series Expert Systems: Aspects of and Limitations to the Codifiability of Knowledge Robin Cowan 2001-005 MERIT – Maastricht Economic Research Institute on Innovation and Technology PO Box 616 6200 MD Maastricht The Netherlands T: +31 43 3883875 F: +31 43 3884905 http://meritbbs.unimaas.nl e-mail: secr-merit@merit.unimaas.nl International Institute of Infonomics PO Box 2606 6401 DC Heerlen The Netherlands T: +31 45 5707690 F: +31 45 5706262 http://www.infonomics.nl e-mail: secr@infonomics.nl Expert Systems: Aspects of and Limitations to the Codifiability of Knowledge Robin Cowan March 2001 Keywords: knowledge, codification, expert systems, knowledge creation, information technology JEL: O31 O32 D83 D89 Abstract This paper discusses recent attempts to codify knowledge through the development of expert systems in several different contexts. This paper argues that in the context of expert systems there is some knowledge that can be codified (turned into an expert system essentially in its entirety), some for which this is partly possible, and some for which it is basically impossible given the state of today's technology. We look specifically at the expertise of three different types of workers: the artisan, the repairer and the strategist, and differences in natures of their expertise, and ask what it is about these different tasks that makes human expertise easy, hard or impossible to capture in codified form. The studies also show though that different types of knowledge lend themselves with different degrees of compliance to the codification process. Acknowledgements This paper has been prepared under the EC TSER Programme’s TIPIK Project (Technology and Infrastructure Policy in the Knowledge-Based Economy - The Impact of the Tendency Towards Codification of Knowledge). I are grateful for the excellent research assistance of Lorri Baier and the comments and suggestions received from colleagues in the TIPIK Project, particularly Patrick Cohendet and Dominique Foray. Robin Cowan MERIT, Maastricht University, PO Box 616, 6200 MD Maastricht, The Netherlands Phone: +31-(0)43-3883878, Email: r.cowan@merit.unimaas.nl Introduction Knowledge codification involves turning knowledge, or parts of it, into messages that can be processed as information. Consequently, codified knowledge is constituted by codes or messages. The codes or messages are expressed in symbols, but for these symbolic representations to be useful, both the representational rules (grammar) and the notation (vocabulary) must be stable, and to some extent, standardized. There must be agreement between or among codifiers and users of codified knowledge, who of course could be the same individual or entity, regarding meaning. Knowledge so codified is easier to distribute, store and recall. These activities are all valuable, and codifying knowledge lowers the costs of all of them. Whether or not all knowledge is in principle codifiable, in practice some is not. This can be seen simply as a matter of cost — the activity of codification has an output, and this output has value, so the activity has benefits. But it also has costs in terms of labour, capital and material. Thus whether a piece of knowledge is codified will depend on the relative costs and benefits of doing so at each level of the process. Of course the evaluation of the relative costs and benefits can change over time and with circumstances. Cowan et al. (2000) propose a taxonomy of knowledge as a way in to the issue of codification and codifiability. All deliberate activity involves knowledge on the part of the actor. That knowledge can be partitioned in a variety of ways, but for our purposes, initially we can see it as codified or not.1 To be codified, knowledge must at the least be articulable. So there are two explanations for the non-codification of a piece of knowledge: either it really is inarticulable and therefore uncodifiable; or relative to the benefits, the cost of codification is too high. In either case, the knowledge will remain tacit. When this is so, any storage, transfer, or use of this knowledge must continue to rest in the hands of the experts. Given that time and resources are finite, and making quite natural assumptions about the distribution of features of knowledge it seems safe to state that some knowledge will remain uncodified, and therefore the domain of human experts. To examine the proposition that some codification can be too costly (or simply impossible) in some cases, this paper discusses recent attempts to codify knowledge through 1 We should point out that often activities which appear to use uncodified knowledge are in fact using knowledge that has been codified but internalized by the agent. An economics professor for example, uses the knowledge stored in principles text books all the time, but rarely refers to the books themselves. This is a case of codified knowledge that has been internalized, referred to as an example of a “displaced codebook” by Cowan et al. (2000). 2 the development of expert systems in several different contexts. Because writing an expert system is an explicit attempt to transfer knowledge from a human to a machine through some sort of message creation process, this seems a good place to look for insight into the process of codification. In a discussion of knowledge acquisition for expert systems, Tzafestas writes that the knowledge which permits experts to do their jobs, and indeed to be experts, is largely tacit. “The knowledge engineer has to explore this tacit knowledge and to present it in a form which can be understood by the experts (for verification) and by the users for beneficial use.” (1993, p. 30). This paper argues that in the context of expert systems there is some knowledge that can be codified (turned into an expert system essentially in its entirety), some for which this is partly possible, and some for which it is basically impossible given the state of today’s technology. These studies we refer to demonstrate attempts to model the kind of knowledge typically employed in industrial settings. We look specifically at the expertise of three different types of workers: the artisan, the repairer and the strategist, and differences in natures of their expertise, and ask what it is about these different tasks that makes human expertise easy, hard or impossible to capture in codified form. But before beginning the analysis, a brief excursus on the nature of the codification process is in order.2 The codification process: messages, models, languages The most widely recognised aspect of the codification process, in which knowledge is transformed into information, is the creation of messages, expressing in a symbolic representation existing, or created, knowledge. Creating messages often refers to the actual process of distributing information, for example, one person telling another or writing down what he knows. That a message can be successfully created, though, implies the existence of some infrastructure. There must be a language in which the message can be “written” and “read”. But a language which is suitable to codify the knowledge in question itself pre- supposes a both model of the phenomenon and a vocabulary with which to name parts of the model and their interactions. It is important to point out that in creating an expert system, say a system to do medical diagnosis, the codes we refer to are not the messages to the user: “What is the infection? … From what site was the culture taken? … Is the organism rod or coccus (etc.)? … This is a streptococcal infection, with certainly factor 70%.” but rather the representation 2 For a more complete discussion of this process see Cowan and Foray (1997). 3 of the knowledge (of a doctor) that underlies this diagnosis. What is codified, for example, is the tree structure of the inquiry he uses to arrive at his diagnosis, for example: “If the organism stains grampos AND has coccus shape AND grows in chains THEN there is suggestive evidence (0.7) that the identity of the organism is streptococcus.”3 In its most general form, then, codification is a process which involves not only the creation of messages, but also the creation of models, since modelling knowledge is a prerequisite to transforming this knowledge into information. In this sense, codification is a central method for producing knowledge. Codification cannot in general be considered as a simple transfer or translation operation, then; there is always, at some stage, an aspect of creation. This aspect of codification can entail fundamental transformations in the way knowledge is organised, so the codified knowledge-base cannot cover exactly the tacit knowledge-base for which, in some sense, it acts as a substitute. This is particularly so in the case of expert systems: “the imitation of expertise, because it is a process of automation of knowledge, is possible only at the expense of the “active” transformation of this knowledge; it is hence in itself a creator of expertise” (Hatchuel and Weil, 1995, p.25). Different types of knowledge demand different types of languages — music, mathematics, expert systems, novels, all have different languages associated with their codification. Some languages are “generic” and can express a variety of types of knowledge (for instance, it is possible to some degree to write mathematical problems in natural language); some, like some computer languages or the jargon of esoteric academic disciplines, are very specialised. These languages must be developed before any messages can be written. Development may be undertaken as an explicit activity or it may emerge from the activities of those working with models and messages. But in either case there is a cost implied, as people expend time and other resources in activities that tend to the creations of a stable, standardised vocabulary. A language must exist, but central to any language are concepts and vocabulary. It is the creation of those two things that the modeller is doing. The existence of a vocabulary, which pre-supposes a model (so that if one does not exist it must be created), is necessary for the ability to create messages.4 References to languages mean that a minimal requirement to be a potential user of the information is that one must understand the language in which the knowledge is 3 These examples are taken from the MYCIN system, one of the first expert systems, designed for medical diagnosis of certain blood infections. (This is rule 037.) See Davis et al. (1977). 4 For a discussion of this and related arguments about the philosophy of science, and the philosophy of reference, see the works of Hilary Putnam, for example, 1987. 4 recorded.5 This is true whether the language is English or French, a computer language or mathematics. Knowledge is easier to codify and codified knowledge is more effectively diffuesd within a community made up of agents who can read the codes. (We should point out that the ability to read the code can be an important form of tacit knowledge.) Diffusion and use of codified knowledge are thus dependent upon the initial investment required to build a community of agents, a clique or a network the members of which can read the codes. This ability to produce and receive signals in a language, even a very common one, requires prior and irreversible investment (Arrow, 1974, p. 39). The relations between the three layers are complex: sometimes pre-existing languages and models are available for the codification process; sometimes codification requires the creation of a new language and some new modelling. In most cases, there is some degree of creation at both the level of languages and models. Expert systems designers often refer to the knowledge (or expertise) acquisition issue, (see for example Tzafestas, 1993, chapter 2) and the way they describe it can be interpreted in the terms set out above. For example, there are two approaches to fault diagnosis expert systems: “model based approach (using mathematical techniques of analytical redundancy) and knowledge based approach (which tries to imitate the reasoning of human fault diagnosers and operators).” (Tzafestas, 1993, p. 39). In both cases a model of some sort must be built: in the first case, a model of the machine; in the second, a model of (one part of) the thought patterns of an expert. Of course, few expert systems are pure cases of one type or the other, and typically building a system involves both kinds of knowledge. Some knowledge is acquired from written documents. Construction of the fault diagnosis system for compressors at the Toyoda Automatic Loom Works used books, manuals and service guides. The service guide illustrated each part of the compressor and gave information about possible failure causes. The existence of these documents provided the basis for a consistent language, and each contained structured descriptions (which can be seen as a model) of different parts of the machine in question. This greatly simplified building a model of the machine. Other knowledge comes from the experts themselves.6 Part of their expertise had been codified before the construction of the expert system began. There existed a decision tree diagram, “designed by the human expert … [and] used to instruct workers … on the expert’s fault 5 The word “recorded” is deliberately ambiguous. Codifiers must understand the language through which the knowledge is systematized and stored. Users must understand the language in which messages are written. So, for example, developers of a medical diagnosis expert system must understand Lisp, but users of it must understand something like a dialect of a natural language. 6 The former is sometimes referred to as “deep knowledge”, the latter as “shallow knowledge”. 5 finding technique. Some of the significant rules in the knowledge base are comprised of the cause and effect relationships as represented in the tree diagram.” (Chen and Ishiko, 1993, p. 197). The diagram is clearly a model of the reasoning of the expert. Four Cases of Expert System Creation In this section we examine four expert systems, drawing attention to the types of knowledge at issue, problems in their creation and what in some cases appear to be inherent limitations in codification.7 Each has a different character regarding the extent to which it can be codified, and the problems inherent in its codification.8 Fixed goals and linear processes: the artisan’s expertise The job of the “artisan” involves a achieving some well-specified final goal (transforming an input into an output) through a sequence of known steps. Planning production is a typical task: “Planning is a process that searches for a sequence of actions (or operations) that will achieve a goal statement.” (Chang and Wee, 1993, p. 291). This expertise can often be represented explicitly as a list of instructions (think of a recipe book or an instruction manual) and is sometimes referred to as technical- or doing- know-how. This expertise allows the artisan to determine the steps between an initial state and a desired final state, passing through intermediate states by well-recognised and determinate steps or actions. It is typically easily and effectively archived, and expresses the “simplest” form of knowledge in the breakdown of industrial job expertise, that is, knowledge that can be transformed into a more or less linear process of discrete steps (or actions/transformations) with a predictable outcome. This is the type of situation in the first case we examine. 7 The cases we discuss in detail are those presented in Hatchuel and Weil. We have used their description of the cases but have re- interpreted them in terms of the theoretical views on codification discussed above and in Cowan and Foray (1997) and Cowan et al. (2000). There are many case studies of expert systems in the literature (see for example, Kaewert and Frost, 1990 Badiru, 1992, or Tzafestas, 1993 as three collections), and we refer to some of the similarities between these and those of Hatchuel and Weil. 8 Hatchuel and Weil refer to three different types of knowledge: doing know-how; understanding know-how; and combining know-how. (For a discussion of different types of know-how see Lundvall and Johnson, 1994). Following Cowan et al. (2000), to the extent possible we avoid discussion of types of knowledge. Instead we focus on types of knowledge activities, attempting to avoid debates about whether we are observing a case of this or that type of knowledge. Further, it is, in the end, codification of the knowledge involved in particular activities that is important, rather than codifying certain types of knowledge, thus by focusing on activities rather than knowledge we can, perhaps, avoid having to map from an economic activity to its type of knowledge and thence to its codifiability, and attempt to move directly from a knowledge activity to discussions of the ability or usefulness of codification in that setting. 6 TOTEM: Automated Production Routings Around 1992 a medium-sized firm that processes gold, silver and platinum experienced a rapid increase in the number and variety of products it manufactured. This implied a significant increase in the number of “production routings” the firm had to create and execute. A production routing is a document specifying the form and quantity of some required raw material, the machines best suited for each operation in its transformation, and a sequence of operations for producing the final product. For this firm more than half of the firms orders were unique, so each order created a new problem, and the firm typically processed several orders at the same time, often using the same machines. Rather than produce a huge production routing library, and in order to keep pace with an accelerated and more complex production schedule, the managers decided to develop an expert system to generate production routings. The system was constructed in a shell called TOTEM, (Traitement Optimisé des Temps et des Matieres), chosen for its ability to represent the planners’ expertise in the form of simple rules. TOTEM began as an empty structure and was gradually fed the expertise needed to produce coherent, accurate routings. Expertise was formally expressed as conditional rules of the “if…then” type: “If thickness decreases by 10% then length will increase by 10%.”, which were elicited from the employees’ descriptions of their decision-making processes. As the system was being created, employees enunciated facts about production routings, and each new fact created a new inference rule, thus propagating the system’s reasoning until no new facts were being discovered. At this point the system could be used to create complete routing documents. Each rule represented an extremely small fragment of expertise, but adding it to the system must be seen as the creation of a new message — codifying one more piece of knowledge. These messages were meant to express all of the relations among the elements of the metal refining process, including the transformation of various characteristics of the product as it passed through its route, the choice of operations and machines, as well as specifications for machine settings, and so on. While formulating the rules it became clear that to avoid excessively time-consuming searches by the system (which evaluates between 500 and 1000 parameters per routing calculation), it was necessary to structure the knowledge in different levels. To do this effectively involved parsing the knowledge into pieces that fit together “at the same level”. Implicitly, this involved creating a model both of the decision-making process and of the knowledge involved in it. Herein lay the modelling process. Level one distinguished knowledge according to product families. Level two broke this knowledge down into 7 autonomous areas of expertise such that each could deal with a different aspect of the problem (e.g., expertise on a particular machine or a specified process). The areas of expertise were then arranged hierarchically and divided so that the system only worked on a small number of potentially useful rules at one time. To ease computation requirements, the system was not constructed to search for the optimal routing, but rather merely one that met the requirements satisfactorily. The process of codification created several thousand “messages”, each one representing a rule used by the firm’s routing experts. These “fragments of expertise” corresponded fairly well to the way planners described their own process of creating a routing. At this level the creation of the system represents a relatively straight-forward transfer of knowledge from humans to machine code. The model of the knowledge was a relatively simple construction of a set of “if … then” statements, and their implications regarding connections between different parts of the system. While modelling the knowledge created a coherent system and explicitly described the links between different parts of the knowledge and the process of decision-making, it created few strikingly new concepts. Thus it was possible to express the model using natural language, though due to the “if … then” structure of the rules there was of course heavy emphasis on Boolean expression. In the production planning process the transformation of thoughts into decisions represented a relatively linear process and thus one that was not overly complex to model. However, as the number of rules increased, the costs of using this simple model (of a list of decision rules) rose, and in order to maintain efficiency and flexibility, the model was adapted to better suit the requirements of the orders and the constraints of the shop floor. The general expertise was organised into a detailed structure of different areas of expertise. This micro organisation of the rule-structure created a macro-level process that closely resembled a manually created production routing, but in fact the “reasoning” of the expert system differed from that of a human route designer. This suggests that there was non-trivial knowledge creation in constructing the model on which the system was based. This system successfully automated a task that had been done manually, and did so on the basis of the knowledge or expertise of the operators. It codified the decision making- process, though at the micro level the model implemented turned out to be different than the model implicitly used by the manual operators. Nonetheless, the system created a macro structure which, when used, closely mimicked the performance of the humans. Other examples of similar success stories exist. Digital Equipment Corp. (now absorbed by Compaq) developed expert systems in many areas of production and management. MATCHER matches the components used in cancelled orders to those 8 required by orders in progress in order to save re-stocking time. BuildArea Selector assigns production of orders to build areas and schedules the orders of production (Kaewert and Frost, 1990). Texas Instruments has created an automated manufacturing centre in which an expert system controls robots, vision systems, machine tools, sensors and so on. Given data on monthly part needs, the system plans, creates material lists, schedules production and monitors the equipment. Human involvement is mostly restricted to entering parts lists. (See Herrod, 1988.) A system developed by InSol Inc. creates bills of materials-requirements for warehouse storage systems. Similar to the Texas Instruments system, the operator enters data on requirements, in this case the nature of the structures to be built, and the system produces the bill of materials using processing that is completely transparent to the user. (See Jindia, 1990.)9 In all of these cases, the decision-making of humans has been successfully represented in an expert system. TOTEM and these other examples have several things in common which jointly make it feasible to create an expert system in which human expertise is (close to) entirely captured. In all of the examples the problem solved by the human and then machine expert was a case of straightforward means-ends problems or decision structures. Given the goal, there are specific, necessary steps that must be followed to realize it. In this case “backward chaining” is possible: one can “backtrack from a goal to the paths that lead to that goal.” (Badiru, 1992, p. 24). But backward chaining is possible only when certain conditions are satisfied. First, it must be that the goal of the activity is known in advance, and does not change as the activity progresses. When a goal changes as a consequence of trying to achieve it, backward chaining is simply not possible. Intimately connected to this condition is that the reasoning employed in the activity be relatively straight-forward means-ends reasoning. This means-ends reasoning is itself decomposable into a sequence of linear steps or stages each one of which lends itself to its own backward chaining. So an identifiable sequence of intermediate steps or goals, leading to the final goal, each one subjected to simple means- ends logic creates a natural situation in which backward chaining can work very effectively. For chaining to be effective, though, one must also be able to enumerate the entire set of problem-solution combinations. Without this completeness, and in addition a relatively low level of uncertainty in these combinations, codification will necessarily be incomplete, and any expert system will demand significant human intervention. The TOTEM case exhibited all of these properties, so backward chaining, and a complete expert system, was possible. (The success of the other examples listed above suggests 9 For many more examples see Badiru (1992), especially chapter 11. 9 that they too satisfied these conditions. See Badiru, 1992.) The problem-solution rules in these cases (and in TOTEM in particular) tended to be of the simple if … then structure so the codes or messages they created were relatively straightforward, and as individual messages, self-contained (and already existed largely intact, complete and coherent in the minds of the human planners). It was also possible to identify all of them. The sequential structure of the building process made this feasible. Any time the shop encountered a situation that was not already included in the codes, the operators were able to describe the rule they used to make the required decision. Simple “if … then” logic, combined with the fact that decisions could be taken to a great extent sequentially, created a tree-structure for the route design process.10 Further, the processes involved are very linear and could be described in a sequence of discrete steps which must take place in a particular order. The process between initial and final state can be broken up into a sequence of largely self-contained sub-processes which simplifies greatly the structure of the expert system. This is true both of the physical processes that are being planned (a production path, inserting surplus parts into an existing production plan) but also of the planning itself. In the creation of the system, TOTEM and the others mentioned all exhibited “exploratory neutrality”. That is, the only goal in creating these systems was to automate well-specified operational sequences that already existed. Building the system was not intended to create new knowledge or information. New knowledge was created in TOTEM, in that a new model of the decision-making process emerged, but this can be seen as a new description of an existing process rather than to find new information. Further, the systems were not being designed to make any choices or evaluate any trade-offs between over-arching goals such as quality, cost, delivery times and so on. Development of the knowledge environment did not raise any issues beyond itself. Thus the task of the codifiers was well-defined, and very circumscribed — a very particular process was to be automated. The relative simplicity of the implicit logical structure and the presence of exploratory neutrality together imply that the modelling process in TOTEM (and presumably in the other examples as well) was in important respects relatively simple. This is not to say that it was easy or fast, rather it is to say that the model itself was not overly complex. Knowledge 10 In the case of TOTEM this is in fact a simplification, though not a gross one. The system is given certain parameters, (number of output units, size, material, tolerances and so on) and on the basis of those explicit parameters generates as much of the routing code as is possible, using the Boolean decision structure described. Then, having completed that stage, it calculates further, implicit parameters (such as yield coefficients), and makes further inferences based on those parameters. Note that even in this more complex procedure the algorithm is very linear. 10 was divided into identifiable pieces which interacted in relatively straight-forward ways. When the TOTEM system ran into the performance limits of the first version of the model, the necessary modifications were significant in terms of performance, but not excessively costly. Thus codification created new knowledge in the form of a new model of the process, but because the costs of doing so were relatively low, codification was relatively complete. It is also worth pointing out that in the TOTEM example the knowledge environment was relatively stable.11 Materials, machines and operations may have varied according to the demands of specific orders but the basic, component operations involved in turning the inputs into outputs were not changing, though of course which of them were used and in which sequence changed from order to order. This meant that essentially once the materials were obtained and the machines were available the expert system could do its work without interference. Categorization and analogy: the repairer’s expertise A very common activity in which expert systems are employed is fault diagnosis and repair. These sorts of systems are used in medical diagnosis, machinery repair, fault-finding in industrial processes and so on. It is almost universally the case, though, that these systems are designed to assist experts rather than to replace them: The maintenance of a complex item of equipment involves a diagnostic procedure incorporating many rules as well as judgement decision by the maintenance personnel. Experience is an important factor in determining the ease with which a mechanic can locate a failure in a component and implement the appropriate correction. Expert systems are now being utilized to assist maintenance personnel in performing complex repairs by presenting menu-driven instruction guides for the diagnostic task. Badiru (1992, pp. 294-5, emphasis added). Diagnosing and repairing any fault requires a combination of activities designed to solve a problem that is often well-specified only in a very general sense, and is in addition often a problem that has not been seen before. This task is typically is more complex than 11 See Cowan and Foray (1997) for a discussion of codification in stable and unstable knowledge environments. The general idea is that every knowledge activity takes place within a broader context or knowledge environment. Sometimes this environment is changing, as would be the case, for example, if there was an episode of rapid, extensive advance in knowledge or technology, or if a new field was opening up and there was not yet consensus about how to formalize it or how to proceed in investigating it. On the other hand, often knowledge activity takes place in a stable environment: there is consensus on how to model the class of phenomena at issue, disputes over modelling, jargon, vocabulary and so on have been settled, and much of the activity is standardized and perhaps even routinized. The economics of codification are very different in the two cases. 11 that of the artisan, involving as it does both investigation and analysis in order first to find the cause of the problem and then to determine a solution. Rather than simple deduction, typically what is involved is an inductive process in which identifying the cause of a problem is done through generalisation. That is, a problem is recognised as belonging to a certain class of situations all members of which have the same cause, and thus demand the same or similar corrective measures. To complicate matters, however, this “cause” may in fact be “multiple” in the sense that several causes produce this class of effects, or worse yet, the effect may belong to the “cause unknown” class of situations. In the first case the operation of repair is relatively straightforward, once the identification is complete. In the second case, repair proceeds through trial and error, though on a circumscribed set of causes. In the final case, the repairer’s expertise is often manifest as “having the right hunch” or remembering what appears to be a similar repair done in the past. Even in the first case, where a single cause is determined, the identification of this problem as belonging to a specific class can create significant difficulties. While deductive algorithms have been developed and are highly advanced, inductive algorithms are much less so. Thus a major challenge in codifying the expertise of the repairer lies precisely in this inductive step. Fault diagnosis systems often attempt to do this by classification. Classes of faults are generated through historical records of previous faults, and the system attempts to put the fault it is diagnosing into one of the classes. The systems DIVA and EXACT had similar performance features: their diagnoses were based on historical generalizations and classification. They are fast (EXACT is claimed to be as good as a human in terms of speed and accuracy) but they cannot solve problems “which have never been encountered in the past and are thus not included in the knowledge database.” (Chen and Ishiko, 1993, p. 206).12 Clearly, repair cannot be described as a linear process from start to finish with a rehearsed and predictable outcome; rather “…it intermingles action and investigation in an ever-changing pattern” (Hatchuel and Weil, 1995, p.36), sometimes unsuccessfully. Therefore, unlike knowledge which is codified, and in a sense “fixed” in a recipe book or a dictionary, “the repairer’s expertise cannot be defined out of context, it requires that the practical and social framework of the reparation [sic] be constructed simultaneously.” (Hatchuel and Weil, 1995, p.37). The difficulty, indeed, lies in the fact that the repairer cannot simply follow a set of rules, he must use his judgement. 12 DIVA is used “to help the operating personnel to interpret the evolution of vibrations and to diagnose the developing faults” in turbine generators, (David et al., 1993, p. 212); EXACT is used for compressor trouble-shooting, (Chen and Ishiko, 1993). The quotation refers explicitly to EXACT. When DIVA encounters a problem it does not recognize, it gives the user the message “I do not recognize your problem! I guess I am not (yet) competent enough …” (David et al. 1993, p. 218). 12 Cornélius The manufacturing sector has undergone rapid and extensive changes in recent decades, and the introduction of “mass customization” has increased the complexity of both manufacturing processes and machinery. A firm in the parts manufacturing sector studied by Hatchuel and Weil responded with an automation programme, thereby increasing the complexity of its machinery base, and increasing the need for maintenance of the new, more complex capital. In this context an expert system devoted specifically to industrial diagnosis and maintenance, Cornélius, was proposed. The facility chosen to test the feasibility of automation a flexible manufacturing cell in the machining workshop. It was thought appropriate for three reasons. First, because the factory’s entire output passed through this cell, it was at risk of becoming a production bottleneck. Thus the ability to diagnose and repair problems quickly and accurately, and to the extent possible automatically, was highly desirable. Second, its problems were representative of those experienced by other NC machines in the plant, so an expert system that could handle the problems of this cell could easily be adapted to other areas of the plant. Third, there were several experienced operators and maintenance workers who could lend expertise to the project. The task of the expert system was to find the source of any failure, to diagnose it and to instruct the operator how to effect repairs. The underlying rationale was that an important source of savings would be to shift at least some of the repair tasks from maintenance personnel to the operators, who were not explicitly trained in maintenance. Cornélius was chosen as the software package because of the relative simplicity of its basic concepts and what appeared to be the close relationship of its reasoning to that of a theoretically trained technician. Given the diagnostic nature of the project, a system that represented knowledge in a set of simple “if A then B” statements was not enough since, due to the pervasiveness of multiple possible causes of defects, the statements would actually be of the form “if A then B or C or D or … ”. Further, since a large proportion of the breakdowns have some unique aspect it would often be unclear which rule would apply. The problem faced by the expert system was much more like unravelling a mystery than simply applying diagnostic rules. To solve this puzzle, a system would somehow need access to a model of the internal functioning of the machine, and in particular to the way in which different components related to each other. Eventually, the system was designed in two parts: a generic application that reproduced a certain mode of reasoning, and a knowledge database which provided the raw 13 data to which the reasoning was applied. These two parts combined to create a model of the machine and the way a repairer would normally address this model. How does this system diagnose a machine failure? In its representation in the expert system, the flexible cell is divided into its main component processes, such as a door opening or closing, conveyor movement, tool changes and so on. Symptoms of failure lead the system to invoke particular subroutines: if it receives a “door not open” signal, it invokes the “suspect door not open” function. Since there are a multitude of reasons why a door may fail, the key is rapid identification of the correct cause. Figuratively, the expert system represents each component system and function by a diagram, similar to an electric circuit diagram, the point of which is to describe the components according to their relations with each other. (These “diagrams” are stored in the knowledge database mentioned above.) During operation of the cell, signals are sent by each of the components describing its states and actions. The system relies on maps of normal and deviating machine behaviour (which are logical translations of the circuit diagrams) to monitor and act. The expert system scans the maps, gathering information, (by inference or by asking the user questions) on the state of the components. The user’s answers allow the system to infer either a normal or deviating state of a component. The diagnostic process continues, either by logical inference or by further questioning of the user, until the faulty component is detected. In case one is not, the system suspects a different function according to the way these functions depend on each other, until no further inferences can be made. One early difficult conceptual issue was to decide whose expertise was to be modelled — the technician who had installed it, the maintenance fitters who maintained it, or the operators who were responsible for restarting it when necessary. Each of these actors could hold relevant pieces of the diagnostic puzzle. The answer emerged during the process of codification: none alone had expertise complete enough to create the expert system. A more “general” expertise had to be constructed and then encoded. Interestingly, language was a potential stumbling block. Different jargons were in use by these different types of agents, and this created a potential difficulty in codifying their knowledge in such a way that others could use it. Language began as non-standardised. The first step of the diagnostic process was to determine the most frequently failing functions. As a consequence, the first modelling task was to identify these functions. This identification was a central part of writing the databases that underpin the system. From here, though, the task of modelling the workings of the system, and the technical inter- relations among the different functions, seemed monumental. Fortunately, it was greatly facilitated by an existing maintenance manual which described components and procedures 14 of available tests. 13 This manual provided a complete description of the flexible cell and its components, the most frequent sources of failure, diagnostic procedures and emergency repairs. From the point of view of creating the expert system, one major contribution of the manual was that in the course of its creation the ambiguities in the technical language had been removed. This solved the language problem, and created the necessary standards. The members of the group agreed on clear and distinct definitions for the cell’s components, and hence much time was saved in the course of formalising the knowledge and in developing the language in which to do so. (The procedure took only 2 months.) The modelling process was trickier. It could not consist of a mere translation of functional plans for the cell. Rather, the important diagnostic elements had to be extracted from the exhaustive list of relations in the functional plans. What caused a great deal of difficulty in designing the reasoning rules was that the typical serviceman’s approach could not be described in diagnostic terms alone. In fact, servicemen relied much more on past experience; their first questions seemed to be whether a particular failure was the same as or related to one seen in the past. Thus it was suggested that this system should not start with a list of questions designed to define a cause, but rather with a list of actions to perform based on the preliminary analysis of the causes of the most frequent failures. While modelling the machine was relatively straightforward, explicitly modelling the casual links among the parts of the machine was more difficult. But a dynamic model of the repairer’s expertise, which would be necessary to make his expertise useful to a non-expert, turned out to be much more difficult yet. This is due in large part to the fact that it involved considerable pattern recognition, and an iterative process of recognition and action. Interestingly, after the first version was tested, the system’s designers re-wrote the knowledge databases to reduce the number of components described, and revised the diagnostic maps to be more economical. The dynamic process of creating the expert system led to the creation of new knowledge, necessitating the creation of new messages and models with which to structure them. Even though in the beginning it was not clear to the project leaders what the final outcome would be, in the end this knowledge turned out to be extremely useful in increasing the overall reliability of the cell, as part of the knowledge created indicated precisely where the most important and common failures were occurring and to what they were linked. 13 The manual had been prepared recently by a group consisting of a workshop supervisor, two artisans, a representative of the maintenance service and a representative of t he innovation group. Notice that this group included several different types of expertise, each of which contributed to understanding how the cell functions. 15 This particular expert system, however, has not been used in the way it was originally intended, which was to make it possible for non-experts to take on a more active role in repair. It has turned out to be extremely expensive (and time-consuming) to create and maintain the needed knowledge databases. Further, even though specifically designed to be user-friendly, it has proved difficult to use for those without significant amounts of expertise ex ante. The knowledge codified in this system was useful to those who were in a sense the sources of this knowledge, as it formalised their knowledge and the processes they followed. This permitted them to work faster and more reliably. However, use of the system was dependent on their own pre-existing expertise, and in particular on their practical experience. We observe here an interesting interplay between environmental stability and codifiability. Contrary to a general tendency wherein stability of environment increases net benefits to codification, in this case the opposite seems to be true. Because the environment was stable the experience accumulated by the operators retained its value and led to ever increasing ability to make good and rapid judgements about where to begin and how to proceed. Again the issue is pattern recognition and generalisation, a venue in which humans in general still seem to have an advantage over computer technologies. Imagine, though, for the sake of contrast, a repairer of either a new machine or a machine that is constantly changing. Here, experience is of little use, and human ability to generalise from past experience does not help. Of more value will be the user or instruction manual, which by definition contains codified knowledge, and is thus more easily embedded in an expert system to aid in the repair. Stability of the environment is central to the ability to codify the expertise of the artisan, and indeed raises dramatically the value of doing so. But here the opposite seems to be the case. While stable knowledge makes codification possible, the stability of the environment makes the value of codified knowledge decrease relative to the unstable environment case.14 In the case of a repairer, the final goal is fixed, and readily identifiable. Nonetheless, backward chaining is not possible. The problem is that the steps to achieve the final goal are unclear before the fact. There are too many possibilities and initially not enough is known to be able to tell the effects of acting on them. The Cornélius project was only a partial success. While the initial goal was to make it possible for non-experts (namely operators) to effect repairs to the machine, in fact this did not happen. Nonetheless, some knowledge was codified, and in the process some created, 14 In the jargon of Cowan et al. (2000) by the time the environment has stabilized, the codebook has been displaced — the repairer has gained enough experience that he knows pretty well what is in the manual, and his expertise lies in knowledge (gained from experience) of things not in the manual. 16 and this codified knowledge was of value to repair specialists. Cornélius is not unique. Descriptions of other fault-diagnosis expert systems indicate the same feature. General Electric’s CATS-1, now called DELTA, was perhaps the earliest, and best-known system for trouble-shooting. This system helps maintenance personnel in railroad repair shops to isolate and repair a variety of diesel-electric locomotive problems (Pratt, 1984, cited in Chen et al. 1993). Ford’s DEX.C3 provides assistance to mechanics at Ford dealerships in diagnosing faults in three-gear automatic transmissions (cf. Bungers, 1986). The PAR (Pulse Acquisition Radar) system of the US Army “assists a PAR mechanic in troubleshooting … It is intended to complement technical manual instructions and mechanic expertise.” (Badiru 1992, p. 298) Bell’s Automated Cable Expertise …was designed to … help managers to investigate the data contained in the complaint system.” (Chen et al. 1993, p. 196; cf. Rauch-Hindin, 1986). AT&T Bell’s interactive Repair Assistant (IRA) “… provides troubleshooting advice to the telephone company’s mobile field technicians.” (Chen et al. 1993, p. 196. cf. Horton et al., 1988). 15 What we observe in the Cornélius case, and what is clear in the descriptions of so many fault-diagnosis systems is that the knowledge or expertise involved in fault diagnosis and repair is not simply raw material that needs to be memorised and recorded. Instead it is a theoretical construction interlinking choices and raw data, and making inferences from one to the other. The key to effective “understanding” is very abstract and the expertise includes generalisation and pattern recognition as central activities. With contemporary technology, both computer hardware technology and technologies of representation of (abstract) knowledge, these activities are both still very difficult. Both DIVA and EXACT simply stop when they encounter a situation they “have not seen before”. Thus modelling the knowledge used is extremely expensive, and possibly still impossible. The inference-making aspect of this expertise implies that it can be highly context-dependent since the context will determine which parts of both data and logic should be employed. Without a coherent model of the artefact or process being “repaired”, messages could not be organised in a way that was easily accessible to non-experts. Experts, though, have internalised this model so the messages were 15 The descriptions in the text here are either quotations or paraphrases of descriptions of these systems. (In every quotation the emphasis has been added.) The idea that appears over and over in descriptions of fault diagnosis systems is the assistance that they provide to already-skilled technicians. There are, it should be noted, a small number of exceptions: Cooker checks large sterilizer cookers for Campbell Soup. Used day and night to “diagnose cooker problems without the need of consulting human experts.” Termi-point Expert Maintenance System (TEMS) (an expert system for supporting wire-bonding machine maintenance) “can diagnose certain machine problems in a few minutes while the human experts solve the same problems in an hour or more.” (Chen and Ishiko, 1993, p. 196. cf. Dallimonti, 1986). 17 coherent to them. Thus, the use of the system cannot be separated from the expertise of the user, and codification will necessarily be incomplete. Balancing conflicting goals: the expertise of the strategist Strategy involves both making plans and finding the most effective means to carry them out. This contrasts with the tasks of the artisan or repairer, both of whom take their final goals or desired end state as given and merely seek effective means with which to achieve them. The strategist, often a corporate leader, strategic planner of the co-ordinator of a team of professionals must formulate both plan and means simultaneously. Here, the expertise has three aspects. One concerns the tasks to be performed, resources to be mobilised, goals to be met and so on, all of which can be relatively easily defined and recorded. The second involves combining, arranging and co-ordinating events, activities and individuals into a coherent unit or network to decide upon and achieve the desired ends through the means chosen. The third, and most difficult, involves making trade-offs among conflicting goals or objectives. The forms and methods of decision-making involved in these activities can be very difficult to formalise. The problem is that backward chaining is simply not possible. As the problem is explored, the goal can shift, and when this happens, it becomes extremely difficult to compare (in a mechanical way) different courses of action. Of course, there are final goals that do not change: maximize profits of the organization; get all trains safely and quickly through a station and so on, but the distance between the immediate means and the final end is quite large, and there is no obvious (linear) path linking the possible actions with the final outcome. There can be many possible actions, and tracing each of them to see their effects on the final, immutable, goal is simply too difficult. In this case, a common strategy is to retreat to intermediate goals. For example, the profits of a conglomerate are intimately related to the profits of each subsidiary, thus if it is too difficult to plan for the conglomerate profits, we should look instead at the profits of all the subsidiaries. But notice that in making this move we have changed from a single final goal to several intermediate goals, and these intermediate goals can, and probably typically do, conflict with each other.16 This is the point at which strategic expertise enters: how to trade these goals off against each other. 16 Chang and Wee (1993, p. 291) put it this way: “Normally an original goal statement is too complex or too abstract to achieve through a single operation. It is often necessary to break down the original goal into smaller, simpler subgoals. In an ideal situation, a subplan for each subgoal can be formulated independently and a compete plan can be derived simply by combining subplans. However, in practice, subplans often interact. That is, achieving some goals can actually prevent the accomplishment of 18 The “Naval” Project An on-going problem for a large multi-national oil corporation was to allocate oil rigs across a variety of prospects. Because both rigs and prospects are very heterogeneous, co- ordinating its activities and those of its subsidiaries so that the most appropriate rig was used at each prospect at the appropriate time could be a significant source of savings. Similarly, to get this co-ordination wrong could be quite costly. Traditionally, this activity involved a great deal of negotiation on the part of corporate planners. To a great extent it was an ad hoc process which evolved continuously as problems in projected assignments were encountered and addressed. In an attempt to make co-ordination more effective and less costly, the firm embarked on the Naval project. This was an attempt to develop an expert system that could generate the most coherent compromise between each of the operations of the multi- national, given geographical, physical and time constraints. The system would in principle use two distinct types of knowledge: one that made it possible to list prospective rigs and drilling prospects, and to rank each rig as a candidate for a particular prospect;17 and one that made it possible to combine different preferences across a large number of prospects into a single coherent plan for rig deployment. The expert system worked in two stages. In the first, the basic elements of rigs (type, weight of derrick, capacity etc.) and prospects (geology, drilling cost, time constraints etc.) were described and simply given to the system as data. It then created a list of advantages and disadvantages of each possible assignment, and used those lists to create a score for each rig- prospect combination. More was needed, however to generate an optimal plan for rig assignment since these scores were based on technical issues, and took no account of the objectives of the different subsidiaries involved. Crucially, these scores made no attempt to balance any conflicts in these objectives. Underlying the lists of advantages and disadvantages were constraints, and these were central in the derivation of the assignment scores. But when looked at from different points of view (e.g. from the points of view of different subsidiaries, or when considering different strategic objectives of the form) different constraints had different importance. To bring this into the analysis, the user of the system had to intervene, giving weights to these constraints. This permitted the system to generate a “weighted if … then” allocation algorithm. others.” They were discussing a “simple” production planning problem in which the conflicts were relatively straightforward to resolve. They were considering none of the complexities involved in the compromises discussed in this section. 17 This issue here is that rigs differ from each other in terms of performance, cost, availability and so on, and prospects differ from each other in terms of technical characteristics, cost of drilling, geological value and so on. 19 Using this weighted system of potential assignments, the expert system was able to build an optimal programme. The “compromise algorithm” represented a significant advance in artificial intelligence, codifying a process formerly done by negotiation within a committee. In this sense it was creating expertise, (ore perhaps unifying it) out of the disparate activities of the committee members in negotiation. Unfortunately the system was limited in several ways. The initial assignment rules had to be simple enough to remain operational, yet general enough to apply to a broad range of situations. As well, although formalised rules contained much, and diverse, content, they were all handled in the same way. Only the weights attributed to them differentiated their importance. But assigning weights in the first place proved to be a daunting task. The limits to codification arose because it was impossible to assign objective values a priori since the weights (and to some extent the parameters themselves) only had meaning relative to one another, and relative to the shifting goals of the organisations. Furthermore, there was no system of weights that could adequately (or accurately) account for the tactics previously used by planners in a rich and dynamic environment. Each specific context required a modification of the weights to correspond with the most suitable tactic (an issue that was brought to light when the recommendation of the expert system was compared with the programmes habitually negotiated for real cases in the past). Nonetheless, Naval proved useful in generating acceptable programmes once the experts had learned enough about the system’s reasoning and the effects of different weights. Still, a full-time operator was necessary to master the several types of expertise involved in running the system. He had to have a sound knowledge of rigs and prospects, an understanding of the reasoning processes of the expert system, complete knowledge of general planning and compromise activities in the industry, and finally he had to understand the over-arching goals of the conglomerate and the inter-relations among the goals of its subsidiaries. His expertise was not only necessary for assigning weights and guiding searches, but also for explaining and justifying the results. For this reason Naval turned out to be “a system for experts rather than an expert system.” (Hatchuel and Weil, 1995, p. 214). What we see here is that this assignment activity was too complex to be codified completely. In essence the system was being asked to make compromises that formerly had been worked out through an iterative negotiation process. While some parts of this were encoded, it was only with the help of significant input from an operator who understood why and how some constraints were more important than others (and that this could change over time) that the system could be usefully employed. Combining and negotiating the goals of the different subsidiaries involved, and creating a plan that enabled these goals to 20 be met to the extent possible involved too much iteration between means and ends. Because the goals as initially stated were incompatible, creating an insoluble problem, they had to be “re-negotiated” as the assignments were developed. If the goal was to automate, and duplicate, the rig assignment that had been carried out by committee previously, the Naval project was a failure. The experts were able to provide the necessary rules to create sensible single rig-prospect combinations (also introducing economic rules and preferences concerning contractors or the origin of rigs). However, in trying to create the compromise programme there was an almost total absence of “expertise” that could be formulated by the identified experts. No one person had handled the task before, and thus it was unclear what exactly was being modelled. Eventually it was seen that this was the negotiation or compromise process itself rather than the skills of any expert. There seemed to be no systematic way of preparing these programmes since the procedure had been handled previously by several committees working continuously, intervening when necessary, trying to reconcile, through negotiations, the most obvious contradictions. In this case a huge innovation in knowledge modelling would have been necessary, involving not only modelling the way in which compromise is struck, but also, and especially, the way goals are adjusted to make compromise possible. The first part, finding a compromise among different objectives, was achieved (and was no mean feat), but in its entirety this innovation is simply beyond reach. The key difficulty was in codifying the weighting procedure which allowed trade-offs to be made among conflicting goals and constraints. The weights having been assigned, the artificial intelligence experts were able to find algorithms to create reasonable assignments, but the key step, assigning weights, had to be done manually. In this enterprise, codification was not an imitation or simulation of expert reasoning, but a pure innovation in the context of expertise within the firm. Unlike the Totem and Cornélius projects in which expert know-how was transformed, in the Naval case a method completely unknown to the experts was incorporated. But while Naval provided the specialists with added support and legitimacy in their co-ordination tasks, the experts could not be done away with. GESPI The GESPI project (Gestion Prévisionnelle des Itinéraires) was an attempt to automate the routing of trains through an urban railway station in France. For us it provides another example of the limits to codifying the strategist’s expertise. 21 In any station, planners create a plan at the start of each day. But in this case there were 640 possible routes to enter the station, 30 tracks through it, and six tracks serving 30 platforms. This is an enormous combinatorial problem: a train has 1400 possible routes through the station. But advances in artificial intelligence technology (and the facility of computers in dealing with combinatorial problems) suggested that an expert system could be an attractive way of solving this repeated problem. Any complete expert system would have to cope with three management levels: crfeating the long term service plan twice a year, defining basic summer and winter services; a middle term plan making basic track assignments for each day of the week; and real time traffic control in which plans are sent to the signal cabin for dispatchers’ use in directing traffic. The third tier of management takes place in real time as it includes on the spot, last minute evaluations of incidents and adjustments to plans. After looking initially at the possibility of automating daily routing, the project team decided to restrict their plans to the middle term assignment of trains and tracks to avoid the technical difficulties of operating in real time. In building the GESPI system the engineers discovered that there were far too many possible combinations of available facilities and constraints in the station to index them, so the planners broke the problem down into rules of routine and rules for managing interventions. There were rules to define trains at risk of conflict, and rules defining trains that were (or were not) assigned regularly to a track. These two types of rules let the planners work out a programme in stages. Initially the system checked the state of facilities and work in progress and ensured that all regular traffic was conflict free. It then attempted to assign platforms and tracks to new or relocated trains (about 250 per day) and ensured a clear route for each, checking that the assignments conformed to technical, commercial and traffic constraints. If there were trains it could not schedule, it asked the planner to route them. The system was built iteratively. When a piece of expertise was recognised, it was formalised and codified. In this is was much like TOTEM, and each piece of expertise was translated into a message. This process constituted a stepwise creation of a model of the information processing and decision-making process. As the model was built, it was tested against actual decisions and revised, much like building a model in any scientific endeavour. This process increased the quantity of the experts’ knowledge that was codified (more “rules” were identified for example) and it improved the principles on which the system was based. Initial results were encouraging as “the experts expressed their reasoning in a form close to the heuristics and rules used in artificial intelligence.” (Hatchuel and Weil, 1995. p.181) A small number of simple rules, codifying the constraints that existed, formed the basis of the 22 system. In the first stage of building the system, technical constraints were addressed, as these were the simplest and most standardised. At a second stage of modelling, commercial constraints were included. These were more complex in part because they involved other departments such as sales and marketing, and thus implicitly involved trade-offs between the goals of the various departments. Further, they existed in a less formalised and more diffuse state, (diffuse both in terms of their natures and in that they originated in different parts of the organization) which made codification more difficult. The language in which the expertise was expressed was again much like a natural language, but the model of the planning process became more and more complex as codification proceeded. As a system, GESPI was able partially to automate the planning task. Users remark, however, that checking GESPI’s plans is more work than creating them from scratch. One can conjecture that this is the case because in creating a plan the logic of the plan is developed, and when a planner knows the general logic of it, it is possible to see whether different parts fit with the general logic. The difficulty in altering a plan of this nature is the risk of creating a cascade of secondary changes that must be made to accommodate the first one. If one understands the logic of the overall plan, certain possible responses can be ruled out immediately because they will cause these cascades later in the routing. If one does not understand the logic, one can only proceed by trial and error. When reading a plan created by GESPI the logic is unknown and has to be deciphered before the internal coherence of the plan can be evaluated. This is a part of the expertise that is extremely difficult to codify. It is in a sense the same issue as the evolving means-ends problem of Naval. The evolution of means and ends is essentially an evolution of the logic of the process. It is in understanding and manipulating this logic that the expertise lies.18 Again, in this case it was not possible to transfer expert knowledge completely. Even though there appeared to be a linear approach to the problem of co-ordinating trains in and out of the station, involving mostly an identifiable goal, fixed constraints, and a limited (if large) number of possible actions, it was the practical experience of the planners that enabled them to judge in advance a potentially dangerous situation and intervene if necessary. It was not possible for the programme to imitate the expertise involved in this kind of decision- making. Paradoxically, a system designed to be more rapid than human planners was in fact less flexible when it came to making last minute modifications because it had to reformulate 18 In the Naval case especially, to the extent that the logic of the outcome is discussed in normal circumstances, it would almost certainly be as ex post rationalisation. After the process has been completed the logic has appeared. Before the process is complete, the logic of the ends is itself unclear and evolving. To mimic this process then, an expert system would have to be able to change its internal logic as the process evolves and goals shift. 23 an entire day’s programme, whereas a planner could make these changes just prior to sending out his plan or at any time throughout the day, as discussed in the previous paragraph. GESPI used the rules and experience provided initially by the planners to combine the main artisanal elements with the skills required to find the best possible compromise, but the programme was unable to imitate their reasoning. Close knowledge of the logic of the station and the consequent logic of the plan were central to being able make (especially last minute, ad hoc) adjustments to an existing plan. This logic proved too difficult to model, and is possibly even inarticulable. Thus again the result of the expert system project was to provide a system that was of limited use and could only be used be existing experts who were able to provide the missing logic.19 In both the Naval and GESPI projects, “expertise” was recognised at the corporate level. One or several highly skilled people were using their skills to perform some important task. Expertise was involved in what appeared, at first glance, to be a task with a relatively stable, repeated objective. Automation should have been possible. The cases differ in that in one, GESPI, the expertise could be located in a single individual (though there were several who, individually, had it); in the other, Naval, the expertise was constituted at the group level — it did not reside in a single individual but in the dynamic of the group. In the GESPI case, several individual experts were identifiable, but a certain aspect of their expertise remained difficult to harness. They possessed knowledge that was highly abstract, and possibly inarticulable; knowledge that is only gained through years of experience in routing trains. Here, planners were developing a plan with some internal logic, and this logic itself evolved as they worked. But by the end of the process this logic had been internalized. Was it articulable? This is difficult to answer. Certainly, though, it would have been extremely difficult, and hence costly to articulate it in such a way that the expert system could have used it the way the planners themselves did. In the Naval case two issues complicated matters. First, an “acting expert” could not readily be identified, hence articulating the expertise was a problem. Indeed, the expertise lay in the interactions of the negotiators. Second, and much more important, was that the key part of the expertise expressed in these interactions was the ability to modify means and ends iteratively, that is, the plan revision resulting from negotiations over the different goals of the actors. The fact that intermediate goals were revised as the planners worked meant that the 19 One use of the system was to check the effects of possible modifications to the station, presumably because this was not a time- critical task, and the key question was whether a proposed modification would create impossible problems for train routing. Using GESPI for this sort of task would not remove a planner from his normal duties. 24 means-ends connections were extremely complicated, and reflexively modified. These “shifting sands” caused immense problems for codification. The stumbling block in both cases lay in codifying some highly abstract vision of the entire operation, and its abstract goals, while the details of the problem solution were involved with very intermediate, conflicting goals. These intermediate goals changed as the problems were being solved, but remained connected to the over-arching goal through the internal logic of the solution. Shifting goals, and the need to develop and understand the abstract internal logic of the solution made codification very incomplete at best. In both cases the activity under consideration involved generating plans for actions which demonstrated compromise between multiple objectives. While at some high, over- arching level goals were constant (make profits; get the trains through the station safely and on time) the intermediate goals were very flexible, and, due to the fact that they conflicted with each other, of necessity fluid. It was this fluidity that caused problems. Before concluding, we should ask briefly about the issue of joint versus individuated expertise: Naval versus GESPI. The two cases are different in that regard. GESPI does represent “organizational expertise” pretty clearly. The simple issue of multiple expertises is not important: “… a combination of independent expertises is very easy to be accomodated [sic] in an expert system. Furthermore, modern expert systems can communicate between them and exchange their knowledge and conclusions.” (Tzafestas and Adrianopoulos, 1993, p. 28). This is what the designers of GESPI tried to do. The problem was not in combining their factual knowledge (presumably different members of the team had different knowledge about rigs and prospects), though, the difficulty lay in the negotiation among different goals. Each member of the team brought with him or her ideas about the goals of his subsidiary and about the over-arching goals of the multi-national. These ideas were combined into a single plan by some inarticulable process. Would designing an expert system have been possible had the process been carried out by one individual who was able to bring to the task all of these conflicting goals? Could he have articulated the process by which his decision was reached? It seems unlikely. The problem lay not in the multiple personnel aspect, it lay in the compromising aspect.20 20 This is not to say that expertise that lies in a group of persons is necessarily the same as the expertise of an individual. But drawing out the implications for codification of the differences is not possible using these cases. 25 Conclusions Codifying knowledge involves three distinct but related aspects: creating models; creating languages; and creating messages. Each of these aspects has its own costs, and at each level, as the process occurs, very typically new knowledge is created. The act of codifying is not merely translating the “expert’s” knowledge out of his head and onto “paper”, but is typically an act of knowledge creation.21 As we have seen in these studies, different knowledge activities involve more and less complex codification processes. Generally speaking, relatively linear, direct processes with a known and fixed goal, like those pursued by the artisan, can readily be codified into a list of instructions or decision rules which can be implemented by a machine. Activities that involve significant amounts of pattern recognition, generalisation and use of analogy are more difficult. Here a deep abstract knowledge of the system being “repaired” is part of the expertise, and this becomes difficult to model and thus to codify. Forming strategies, trading different, conflicting goals off against each other is harder yet. Goals can change as constraints are encountered, so they co-evolve with actions in the process of addressing whatever problem is to be solved. This involves a vision of the system at an even higher level of abstraction, and this aspect of expertise is extremely difficult to codify. Costs of codification obviously increase as we move from the artisan to the strategist, and these rising costs can contribute to the less complete expert systems that it seems possible to produce when artisanal or diagnostic expertise is involved. We should also note that environment matters. When an environment is changing it means that the expertise in question has to be both broader (in dealing with more situations) and deeper (to deal with new situations). A “complex” task in a stable environment is likely to be easier to codify than a “simple” task in a changing environment. Speaking again in terms of costs, in the stable environment, the fixed costs of creating models and languages will in general be lower and, further, there is the possibility to spread them over more applications. This seems to be true for many types of expertise. But when the central part of human expertise is created through experience (as is the case of the repairer) a stable environment increases the value of this uncodified knowledge, and it can be this expertise that is difficult to codify and this decreases the net benefits of codification. The TOTEM project was implemented and considered a success. We can conclude from this that benefits outweighed costs. Interestingly, though, economic factors did not play a large part in the decision-making in the project. We can conjecture that costs were 21 This point is stressed by Hatchuel and Weil. 26 relatively small. Benefits came not in the form of immediate productivity gains, but rather in the creation of expertise in a form that was possible to use and evolve. Further, the project created unexpected knowledge (for example an improved definition of ingots) which gave the firm better control over its production process. While Cornélius was a technical success, in that it was able to codify expert knowledge, it was not a success in diffusing it beyond the initial experts whose knowledge was codified. The project as an expert system for repair has been discontinued, but the codified knowledge has been put to other, valuable, uses. Codification of the causal structures of failures enabled the firm to increase the general reliability and productivity of the cell. The Naval project was discontinued in part because of a major change in the oil market, and in part because the labour costs of the project were not bearable in the firm’s new environment. In the new environment, the firm no longer performed task that Naval was meant to automate. This, of course, makes it difficult to evaluate its (potential) costs and benefits. The diagrams produced in the process of creating the GESPI program were similar to those produced manually. The main result of the system, though, was to confirm the previous opinion that the station was operating at or above capacity. But the similarity between the GESPI routing and manual routings implied that the GESPI system could be used to run experiments on proposed station modifications. One highly peripheral benefit that was noted was that GESPI provided a beachhead to further office automation. As an expert system, though, it seems clear that costs outweighed benefits. These four case studies illustrate that expert systems can codify knowledge embedded deeply in skilled humans, and in doing so often create new, sometimes unexpected and valuable knowledge. The studies also show though that different types of knowledge lend themselves with different degrees of compliance to the codification process. 27 Bibliography Arrow, Kenneth J. (1974). The Limits of Organization, Norton. Badiru, A.B (1992). Expert Systems Applications in Engineering and Manufacturing Englewood Cliffs: Prentice-Hall. Callahan, P.H. (1988). “Expert Systems for AT&T Switched Network Maintenance” AT&T Technical Journal, vol. 67, pp. 93-101, January/February. Chang, K.-H. and W.G. Wee “A Knowledge-Based Mechanical Assembly Planning System” in Tzafestas (1993), pp. 291-306. Chen, J.-G. and K. Ishiko (1993). “EXACT — an Expert System for Automobile Air-Conditioner Compressor Troubleshooting” in Tzafestas (1993), pp.193-207. Cowan, Robin. and Dominique Foray (1997) “The Economics of Codification and the Diffusion of Knowledge”, Industrial and Corporate Change vol. 6(3), pp. 595-622. Cowan, Robin, Paul David and Dominique Foray, “The Explicit Economics of Knowledge Codification and Tacitness” Industrial and Corporate Change, 2000. Dallimonti, R. “Smarter Maintenance with Expert Systems” Plant Engineering, pp. 51-56, June 18, 1987. David, J-M., J-P Krivine, J-P Tiarri and B. Ricard (1993). “Using Prototypical Knowledge in Classification-Based Expert Systems” in Tzafestas (1993), pp.209-221. Dosi, G. and M. Egidi (1991). “Substantive and Procedural Uncertainty,” Journal of Evolutionary Economics, 1(1): pp. 145-168. Hatchuel, A. and Weil. B. (1995) Experts in Organizations, de Gruyter, New York. Herrod, Richard A. (1988). “AI: Promises Start to Pay Off”, Manufacturing Engineering, March 1988, pp. 98-103. Horton, E.M., Hsiao, J. and Sielinski, J.E. (1988). "Interactive Repair Assistant: A Knowledge-Based System for Providing Advice to Field Technicians" IEEE Communications Magazine, vol. 26, pp. 56-57, November. Jindia, A.K. (1990). “Expert Systems Remove Repetitive Tedious Work for Customer Order Entry” Industrial Engineering, vol. 22(11) Nov. pp. 51-53. Kaewert, D.W. and J.M Frost (1990) Developing Expert Systems for Manufacturing: A Case Study Approach New York: McGraw-Hill. Kokkenaki, A.A., K.P. Valavanis and S.G. Tzafestas, (1993). “A Survey of Expert System Tools and Engineering-based Expert Systems”, in Tzafestas (1993), pp. 367-378. Kubicek, H. and Seeger, P. (1992) “The negotiation of data standards: A comparative analysis of EAN and EFT/POS systems” in M. Dierkes and U. Hoffman eds., New Technology at the Outset: Social forces in the shaping of technological innovation, Campus Main. Long, P. O. (1991) “The openness of knowledge: an ideal and its context in 16th-century writings on mining and metallurgy”, Technology and Culture Lundvall, B.A. and Johnson, B. (1994) “The Learning Economy”, Journal of Industry Studies, vol.1(2). 28 Pratt, C.A. (1984). “An Artificially Intelligent Locomotive Mechanic” Simulation pp. 40-41. Putnam, Hilary (1987) The Many Faces of Realism Open Court Press. Rauch-Hindin, W.B. (1986) Artificial Intelligence in Business, Science and Industry - Fundamentals. Prentice- Hall, New Jersey Tzafestas, Spyros (ed.). ( 1993). Expert Systems in Engineering Applications Berlin: Springer-Verlag, 1993. Tzafestas, S. and A. Adrianopoulos “Knowledge Acquisition for Expert System Design.” in Tzafestas 1993, pp. 25-51. MERIT-Infonomics Research Memorandum series - 2001- 2001-001 The Changing Nature of Pharmaceutical R&D - Opportunities for Asia? Jörg C. Mahlich and Thomas Roediger-Schluga 2001-002 The Stringency of Environmental Regulation and the 'Porter Hypothesis' Thomas Roediger-Schluga 2001-003 Tragedy of the Public Knowledge 'Commons'? Global Science, Intellectual Property and the Digital Technology Boomerang Paul A. David 2001-004 Digital Technologies, Research Collaborations and the Extension of Protection for Intellectual Property in Science: Will Building 'Good Fences' Really Make 'Good Neighbors'? Paul A. David 2001-005 Expert Systems: Aspects of and Limitations to the Codifiability of Knowledge Robin Cowan 2001-006 Monopolistic Competition and Search Unemployment: A Pissarides-Dixit- Stiglitz model Thomas Ziesemer Papers can be purchased at a cost of NLG 15,- or US$ 9,- per report at the following address: MERIT – P.O. Box 616 – 6200 MD Maastricht – The Netherlands – Fax : +31-43-3884905 (* Surcharge of NLG 15,- or US$ 9,- for banking costs will be added for order from abroad) Subscription: the yearly rate for MERIT-Infonomics Research Memoranda is NLG 300 or US$ 170, or papers can be downloaded from the internet: http://meritbbs.unimaas.nl http://www.infonomics.nl email: secr-merit@merit.unimaas.nl 
work_6z2l7avhtvdyti2zh7yfclr4dm	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220970083 Params is empty 220970083 exception Params is empty 2021/04/06-03:05:09 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220970083 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:09 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_6zijj3t3yzarjnpezhk6rguhx4	----	Domain-dependant information gathering agent Domain-dependent Information Gathering Agent Aleksander Pivk*, Matjaz Gams Jozef Stefan Institute, Jamova 39, 1000 Ljubljana, Slovenia Abstract A universal agent should be capable of gathering information from arbitrary heterogeneous sites and offer intelligent information services on its own based on information so gathered. We present a domain-dependent agent for information gathering. It can visit an arbitrary domain-related site by observing a user perform the first query. By understanding key concepts of the first query, the agent performs subsequent queries autonomously. When a user asks the agent about a particular item, the agent gathers relevant information from various sites. The major advantage of the agent is a semi- automatic creation of a wrapper around a particular site with few human interventions. We have implemented two versions of such information-gathering agents: ShinA (SHoppINg Assistant) for e-trading tasks, and EMA (EMployment Agent), which performs employment and job related functions over the Internet. Keywords: Intelligent agent; Comparison shopping; Information extraction; Inductive learning; Wrapper 1. Introduction Internet services offer several advantages over the classical ones, and are already changing the way companies, customers, and users perform their activities. But the rapid growth of the Internet has resulted in an explosion of available information and online services. To make things worse, the Internet is poorly structured, and disorganized (Lewis, 1999). There are millions of heterogeneous Internet sites offering various services, which makes the Internet unintelligible for any single user conducting a non- trivial task. Therefore, serviceability of the Internet depends to a large degree on the amount of automated uniform user-friendly services, systems, and tools (Levy, 2000; Pazienza, 1997). The negative side of the Internet success is evident when one wants to gather relevant information. One of the first solutions to cope with information overload were search engines. Typically, a user provides a couple of keywords, and a search engine finds links to relevant sites. This simple approach has several drawbacks. A search engine proposes only simple links to hopefully relevant sites. A user must still follow the links manually, and further browse for the desired information (Balabanovic & Shoham, 1995). Some of the links point to vast databases with specialized interfaces. This forces the user to get acquainted with new interfaces, and slows down the user even more. In addition, this approach favors largest information providers due to name-branding and time optimization. As a result, this largely reduces the basic advantage of the Internet. To further complicate matters, a user is unaware of changes in already visited sites unless he or she revisits the sites frequently. A rather simple solution is to provide alert services when a relevant site changes. E.g., several e- commerce sites already allow shoppers to order price alerts that notify a shopper when the price of a product changes, or falls below a specified amount. But some of the services require lengthy surveys to be filled out before they can be used, while at the same time they may provide little or no personalization. It is also undesirable that such services weaken user privacy (Jakobsson, 1998). Another approach involves voluntary ratings and reviews of sites by users, resulting in personalization and recommendations. Again, this takes a lot of user time and introduces problems with privacy. In addition, recommendation systems effectively reduce the size of the marketplace and introduce bias, as it is difficult to obtain a sufficient number of ratings for every existing vendor, and to control the reliability of the sources (Pivk, 2001; Basu, Hirsh & Cohen, 1998; Schafer, Konstan & Riedl, 1999). In this paper we present solutions of the problem of information gathering. We propose a domain- dependent information-gathering agent that extends the existing approaches by providing automatic uniformity, user autonomy, and privacy. We implemented two systems: a semi-uniform intelligent agent for employment tasks (Gams, 2001) and for e- commerce (Pivk & Gams, 2000). Our principal goal was to improve accessibility and expand the benefits of existing information-gathering approaches. The paper is organized as follows: In Section 2 we describe agents, in Section 3 wrappers, in Section 4 the algorithm, in Section 5 the implementation for employment tasks, and in Section 6 for e-trading. * Corresponding author. Tel.: +386-1-447-3380; fax.: +386-1-425-1038. E-mail address: aleksander.pivk@ijs.si Related work is described in Section 7. We summarize our ideas in Conclusion. 2. Agents In information gathering, one of the most attractive new approaches are software agents. They help automate a variety of activities, most of which are time consuming. Software agents differ from “traditional” software in the sense that they are personalized, social, continuously running and semi- autonomous (Gutmann, Moukas & Maes, 1998). In this way, the Internet is becoming more user-friendly, semi-intelligent and human-like. There are many definitions of what the term “agent” denotes, based on different approaches, expectations and visions (Bradshaw, 1997). Shoham (1997) describes a software agent as a software entity which functions continuously and autonomously in a particular environment often inhabited by other agents and processes. The requirement for continuity and autonomy derives from human desire that an agent be able to perform activities in a flexible and intelligent manner, responsive to changes in the environment, and without constant human supervision. An agent that functions over a long period of time should be able to adapt from its experience (Liebowitz, 1999). There are two types of agents, closely related to our research: Information/Internet agents and shopping agents. Information/Internet agents perform the role of managing, manipulating, or collating information from many distributed sources (Maes, 1994; Maes, Guttman & Moukas, 1999). The motivation for developing information agents is at least twofold. Firstly, there is an increasing need for tools that manage the information explosion. Secondly, there are also vast financial opportunities to be gained. Shopping agents find products (the following terms: product/object/item/entity represent synonyms in this paper) under the best terms among different e- commerce sites (Dastani, Jacobs, Jonker, Catholyn & Truer, 1999; Gutmann, Moukas & Maes, 1998). A shopping agent queries multiple sites on behalf of a shopper to gather pricing and other information on products and services. However, basic comparison-shopping agents (Greenwald & Kephart, 1999) still lack several desired properties. They introduce a marketplace that is biased in favor of those e-commerce sites that collaborate with the shopping agents. In addition to a limited number of e-commerce sites to choose from, those participating sites often do not offer the best prices. Typically, e-commerce sites that are included into agent’s repertoire follow specific protocols or include hand-coded wrappers to transform data into agent-readable forms (Hammer, 1997; Muslea, Minton & Knoblock, 1998). An ideal agent would extract contents of heterogeneous Internet databases online, and present the data to a user in a uniform way (Cowie & Lehnert, 1996). While this task is feasible by humans, current computer systems are not intelligent enough to perform it successfully as the type, amount, and organization of the information provided by databases differ from site to site (Baek, Liebowitz, Prasad & Granger, 1999). For an autonomous agent, the desired properties are uniformity, user autonomy and privacy. Uniformity refers to the ability of an agent to automatically enter various sites and adapt to the type, amount, and organization of the information provided by various companies. Since this property demands human-level intelligence, the task is beyond current computer systems. In reality, we accept limited uniformity, i.e. semi-uniformity. Autonomy refers to the idea that an agent provides the best possible service by remaining as independent as possible from both customers and vendors. Autonomy from vendors implies that the service is to remain unbiased by performing wide searches (as opposed to only searching the databases of a few “preferred” vendors). This can be achieved by progress in making interfaces more uniform, and by improved methods for interpreting potential hits. Autonomy from the customer means that users can be relieved from the tedious task of searching for information and of needing to adjust to different sites. Additionally, our research addresses the privacy of the shopper by concealing the identity and behavior of the user. However, we note that the privacy provided is conditional, and should be selectively revoked if abuse is suspected. These aspects of our proposed agent provide for an unbiased marketplace where the user benefits in many respects, and safely stays hidden behind the agent who performs anonymous searches on user’s behalf. 3. Wrappers There is a classical solution for dealing with heterogeneous information sources (in our case domain-specific databases), where each source has a unique way of providing information to the user. Namely, a program may be able to integrate information from the sources if it has at its disposal a communicator-translator, generally referred to as wrapper (Yang, Seo & Choi, 2000), which produces a translation rule for each site. Formally, a wrapper is a program or a rule that understands information provided by a specific source and translates it into a regular form that can be further reused. A wrapper is an essential component of a mediator system (Fig. 2), which accepts queries from users, translates each one into the appropriate query for the individual source, fetches the relevant pages from that source, extracts the requested information from the retrieved pages, and returns the result to the user. Essentially, wrappers make the Web sources look like databases that can be queried through the mediator’s query language (Kusmerick, Weld & Doorenbas, 1997; Cowie & Lehnert, 1996). Wrappers can be either hand-coded or created automatically by software agents (Muslea, 1999). Hand-coding wrappers is a tedious and time- consuming task. In addition, a comparison agent with these manually written wrappers is not scalable because new stores are not automatically integrated. Automatic construction of wrappers by agents for online stores is a challenging issue, mainly because HTML documents on the Web are not agent-friendly (Liu, Pu & Han, 2000). Another difficulty for automatic wrapper-generation comes from the heterogeneity of scores, in the sense that different stores employ different mechanisms for manipulating customer queries, and different styles for displaying product information. Fortunately, domain-dependent sites such as online stores are mostly semi-structured (Atzeni, Mecca & Merialdo, 1997), and we can extract at least some meaningful information without the help of semantic-based modules. 4. Information-gathering algorithm We present our algorithm for an automatic information-gathering domain-dependent agent. Task: Gather information from heterogeneous Internet databases with as little human intervention as possible. Domain-dependent constraint: Heterogeneous Internet databases have to contain common concepts like job definition in employment tasks, or product description in e-commerce. Before presenting the algorithm, we describe the basic data structures used. 4.1. Data structures A record represents a product/object/item/entity, e.g., a job offer or a product offer. Here is an example of a product offer from Amazon, in this case a book: Special Edition Using Java Server Pages and Servlets, by Mark Wutka, Our Price: $27.99, You Save: $12.00 (30%), Used Price: $15.06, Availability: Usually ships within 24 hours, Paperback - 754 pages 1st edition (January 15, 2000), Que; ISBN: 0789724413; Dimensions: 1.70 x 9.09 x 7.36, Amazon.com Sales Rank: 69,177. A reply usually consists of records of items in HTML. The items are stored as lists of records of text, or records in XML format. A form corresponds to an HTML form. There are two variations of this data structure: empty – corresponding to the HTML source, and filled – with data filled in. An example of a form is presented in Fig. 1. The relevant part of the HTML source is: <form name="shinaForm" method="post" action="shinaServlet" onSubmit="doSubmit()"> <center><table border=0 ><tr> <td align=center><font size=-1> <a href="Help.html">Need help?</a> </font></td><td><center> <input type="text" value="" name="searchStr" size="30" maxlength="80" onChange= "this.value=validateQuery(this.value);"> <td><input type="submit" value="ShinA Search"></td></table></center> <hr size=2 width="380" align="center"> <center><table cols=1 width="63%" ><tr> <td align=center><font size=-1> Number of items to be displayed: </font></td></tr><tr><td align=center> <select name="noItems" size="1" > <option value="5>5&nbsp; <option value="10" selected>10&nbsp; <option value="20"> 20&nbsp; <option value="50">50 </font></select></td></tr></table></center> </form> Fig. 1. ShinA’s main interface. The first paragraph of the HTML code introduces the form name (shinaForm), the method (post), and the action (shinaServlet) that is performed upon query submission, a couple of parameters, and a link to Help. The second paragraph presents an input field for a text query, which gets validated by the validateQuery method. The third paragraph describes the ShinA’s submit button. When pressed, the entered text – query, is sent to the shinaServlet program. The last paragraph lets a user choose the number of items per output page. A query is a string of text submitted to a server through an HTML form. An example of a query searching for "Java Servlets" is: http://ai.ijs.si/shinaServlet?searchStr=Java+Servlets (3/1) The query consists of the address of the server (http://ai.ijs.si), the name of the program (shinaServlet), the parameter name (searchStr), and the entered value. Our system also attaches statistics to each query in the form of the success rate (3 attempts / 1 failure). HTML commands are parts of a form. The agent uses them as patterns. An example of an HTML command as a form field is: <input type="text" value="" name="query" size="30" maxlength="80" onChange= "this.value=validateQuery(this.value);"> A query-reply consists of a query and a corresponding reply. With it the system records questions sent by the user together with reply generated. Fig. 2. An information integration system / mediator accesses databases through wrappers. Extraction rules enable parsing of the output of a particular database in order to get a uniform output by the mediator system. Extraction rules represent the syntactic structure of database output (Ashish & Knoblock, 1997). Our implementations are rather straightforward, e.g. a rule searches for a text description inside HTML documents, separated by specified delimiters. Examples of extraction rules are: text := string of characters number := string of digits 0 .. 9 … price := html.body.td[i].number WHERE html.body.td[i].b.text = “EUR” … jobDescription := { text1 jobDefinition text2 delimiter} productDescription := { ID name price text } The first two lines are examples of basic syntactic structures like text and number. The middle part is an example of an extraction rule for price. It consists of a number, which is followed by a currency type, in this case EUR. The last two lines represent top-level extraction rules, describing major concepts like jobDescription or productDescription. A jobDescription consists of some text before an actual jobDefinition and some text afterwards, where a delimiter separates each jobDescription. Schematically, the system communicates with heterogeneous databases using extraction rules for wrapping (Fig. 2). 4.2. Algorithm The algorithm consists of two major phases: 1) the initial query to an unknown database The agent observes a user enter and query a new Internet database (ObserveUser – names of procedures are in italics). Then the agent records and parses communication between the human and the database, and extracts sufficient content (HTML RuleExtraction, concept matching). If necessary, the agent consults the user to get the desired level of understanding of the communication. This enables nearly error-free performance of the system. 2) subsequent queries (MappAndMatch, Retrieve) The agent performs a new query to the Internet database by simulating what a user did. If the query is successful, the agent updates the statistics for the query and presents information through a uniform interface to the user. If the query is not successful, the system updates statistics, and discovers whether the query has become invalid. If the query is found to be invalid, the system demands new rules to be extracted for a particular site. The user is notified accordingly. When connecting to a new database for the first time, the ObserveUser procedure is applied: procedure ObserveUser (query, commands, reply); {A user starts communicating with a database through the agent interface. The agent remembers user commands and HTML commands without active participation. The user is in principle unaware of agent’s presence.} begin start observing a communication between a user and a database interface; repeat remember query and commands; until session is finished; if session finishes with error then forget the whole session; else RuleExtraction(query, commands, extractRules, reply); end; procedure RuleExtraction(query, commands, extractRules, reply) {The agent stores three basic types of information: - database description - name, URL address, title, country and other available information. - query and extraction rules - to perform another query in future; the query is either stored for the first time or modified accordingly - a list of HTML commands - the agent parses HTML commands, especially those that perform some action with the input/output fields. Other commands are most often skipped.} begin store query into appropriate data structure; find a delimiter in the reply; repeat {parse commands and create extract rules} find the next HTML command in the reply; if this command fits a rule then add the rule to a list of extractRules; else ignore the command; until all commands are parsed; set pointers to establish query-reply structures; {end parsing commands, start parsing the reply in order to extract concepts} repeat find the first concept/pattern/item (e.g., a job, a product ... ); add the concept/pattern into the corresponding list of entities {mapping}; until document is parsed and transformed into a list of entities for the mediator system; end; The procedure for creating extraction rules establishes connections between the parsed HTML document and the domain-dependent concepts. In this way, background knowledge in the form of ontologies is used to capture the meaning of essential parts in the HTML document. For example, in the domain of employment tasks, ontologies can be used to identify professions. If the RuleExtraction procedure fails at any stage, it calls for a supervisor intervention. On demand, it shows all intermediate results, e.g. a list of extracted rules, which the supervisor can modify at will. Therefore, if the system decomposes the process well, no human help is needed. In any case, at the end of the session the system assumes that the parsing and mapping procedures are flawless. This does not mean that no mistake can occur in the future, because a page can change. The "Retrieve" procedure connects to databases of previous queries, where queries are modified to reflect user’s question. The "Retrieve" module connects to the database in two ways: • if exact query was previously successfully performed, the agent repeats it; • otherwise the system tries to create a query, based on a similar query. The best way to do this is to “understand” by ontologies, which queries qualify as promising candidates. For example, if a user asks for a specific type of shoes, all shoe-related shops and all shoe-related previous queries seem to be a good basis for a new query. The new query is constructed by changing previous entities with the specific type of shoes. procedure Retrieve(query, replies); {get info from a query database and write it into replies, consisting of a list of entities with corresponding parameters and their values} begin find all queries with identical input fields to the one typed by a new user; if no queries found then CreateQuery (queries); for all queries do extract database address from a query and connect to the Internet address and obtain lastReply; if an error is obtained then if query is invalid then begin RuleExtraction(query, null, extractRules, null); notify the user; if RuleExtraction not successful then notify the supervisor; end; else notify the supervisor; else begin MappAndMatch (lastReply, replies); update statistics; end; show extracted replies to the user through the mediator system; end; procedure CreateQuery (queries) {create a new list of queries based on previous similar ones} begin apply an ontology to create a list of new queries; for all previous queries that correspond to similar items do begin replace items with new ones; set statistics for new queries; end; end; There are different types of ontologies, some of them hard-coded into the system and others consisting of simple or complex data structures. For example, if a query corresponds to a form with just one input field for a database search, and the user enters just one input field in the mediator form, then this is a perfect match. Another ontology is based on clustering of databases based on the entities they provide (i.e., kinds of products). Other ontologies consist of a list of entities that can be entered into a specific input field. For example, practically all job definitions are defined in a national catalog. Ontologies play a substantial role in construction of modifications from a list of queries and commands. For example, a full "looking-for-jobs" query can be described in a form with a flexible frame consisting of many subfields. The query contains information about job profession and additional specifications of area, income, type of work etc. The routine checks that all common major items, such as profession, are filled in and that modified commands are (hopefully) related to the same job specification. procedure MappAndMatch (lastReply, replies) {based on extracted rules, parse the lastReply and add each extracted entity with corresponding values into replies, i.e. into the mediator system} begin repeat match each extraction rule to the lastReply; when the first rule triggers, delete the matched top of the lastReply; if the extracted transformed text corresponds to a part of the entity then add it into replies; until end-of-lastReply; end; Advanced successful performance of the system is based not only on ontologies but also on memory- based learning. The ontologies are more or less static and sometimes even hand-coded. They represent basic knowledge about a specific problem domain. Some ontologies are dynamically created or instantiated, e.g. those describing shops, which makes it possible to search similar shops. Namely, each shop typically presents its product catalogue, basic orientation, and structure. While ontologies are domain-specific, the memory- based learning is a plain uniform implementation of the basic technique published elsewhere (Kitano, 1993). The algorithm was only slightly modified according to the problem domain characteristics. The algorithm performs memory-based learning by storing queries and statistics. The system learns each time a pattern is successful or not by changing its statistical success rate. In this way, the system adapts to the desires of the users and to changes in the domain-related sites on the Internet. Our algorithm is described in (Gams 2001). Additional learning in ShinA is described in Section 6. The system was implemented in two problem domains: employment and e-trading. While the two implementations vary a lot in terms of programming languages and technical details, the algorithm is practically the same. Both implementations offer information about desired objects, which differ in properties, benefits and costs. Users browse through a large number of potential candidates, and compare several similar possibilities to finally choose specific objects. 5. Employment information gathering EMA (EMployment Agent) is an intelligent employment agent for providing employment information in a way similar to human employment agents (Gams, 2001). An Internet-based intelligent employment agent should perform like a human agent, provide basic information about available jobs and available workers, and try to match these two databases as efficiently as possible (Fig. 3). In employment tasks, users typically want to get a job - any job, a good job, better than a current job. In this way, a typical task of the agent is to find relevant information about a specific job offer (analogy – a specific e-commerce item) from various information sources. EMA is an intelligent agent that provides user- friendly information on demand, or when it notices relevant information on the Internet. EMA can store search patterns, perform search on its own or following direct commands, and it can send online replies or offline emails regarding vacant jobs or available workers. EMA can store and observe interesting Web sites chosen by users, and match available jobs and workers. Fig. 3. The basic task of the EMA Internet-based intelligent employment agent is to provide employment information as human agents do. EMA has been implemented a couple of years ago with several modules added subsequently. Ema also incorporated some language and translation capabilities, such as Slovene text-to-speech interface. It also included ontologies for profession definitions, and the ability to connect to arbitrary employment sites on its own. Natural language processing. A common input to EMA is a partially constrained Slovenian domain-dependent text, often consisting of phrases and form inputs. In bulletin boards, the language is a matter of choice, but titles and information are either in Slovene or in English. Consequently, majority of text is either Slovene or English, with a couple of exceptions. No censorship is performed regarding language or specific details as long as the input is dedicated to the desired employment task and inside minimal decency requirements. The system is capable of translating Slovene text into English. The translation is based on a dictionary consisting of up to four words observed previously in the employment data. In the worst case, new combinations are translated word-by-word and stored for further overview by humans. Stored combinations are sorted by frequency and translated by humans if reasonable. In addition, the translation system looks into the morphology dictionary to capture different forms of the same words. Slovene has a rich morphology so this is essential for good performance. Finally, a spell-checker module corrects spelling errors. The translation is currently not yet at the level performed by systems translating between larger European languages; however, it is sufficiently good to enable basic understanding of translated text, since the employment syntax is quite limited. The Internet site of the EMA system is presented in Fig. 4 (http://www.ess.gov.si). Fig. 4. Home site of the intelligent employment agent. Ontologies. EMA performs some of its advanced functions based on ontologies. In this way, EMA understands the basic employment tasks. One of the most common employment tasks (besides basic queries) is to find a relation between different job descriptions and jobs definitions. This is in fact one of the basic concept of all employment tasks. E.g., from a basic search query consisting of applicant’s properties and desires, one should extract relevant job offers. Conversely, from a job definition one should look for available workers. With machine and statistical text- learning methods (Freitag, 1998; Mladenic, 1999) we (Bezek & Gams, 2001) have designed ontologies, i.e. meta-knowledge about job definitions from interesting words. Semi-uniformity. During the first years of the EMA system, several competing Internet-based employment sites were designed. Since EMA was a national employment site sponsored by Employment Services of Slovenia (ESS), an idea emerged to represent also the information from other employment sites. Instead of handwriting wrappers around each particular employment site, we have designed an algorithm that semi-automatically (Ashish & Knoblock, 1997; Kusmerick, Weld & Doorenbas, 1997) attaches to other employment sites and then automatically gathers information from those particular sites. Although the algorithm published in (Gams, 2001) was not fully implemented, we have designed two slightly modified modules of this kind and implemented them in EMA. Another version was independently designed by a student group (JobProvider) on the basis of public lectures and presentations. All versions were functional for at least some time, and on average for several years, thus showing that the idea is sound. On the basis of these experiences we have decided to generalize the algorithm, and design another system, the ShinA system, for e-commerce tasks (Pivk & Gams, 2000). EMA is a big software system especially for our R&D group of intelligent systems at the Jozef Stefan Institute. It consists of several modules and submodules, written in different programming languages and runs on several platforms and operating systems. The system is a 30 thousand lines program written mainly in C, and also in other languages and with Internet programming tools. Together with text and data it occupies 30 MB of space. The EMA agent was, and still is, among the most successful applications of intelligent agents in Slovenia and in Central Europe. In the first year of its implementation, our country was the third in Europe to offer national employment information through the Internet. At that time, we were the first country in the world to provide over 90% of all nationally available jobs on the Internet. On the other hand, in absolute terms there are employment systems in big countries or employment systems connected to major Internet information providers such as Yahoo or AltaVista that provide orders of magnitude bigger amounts of employment information. EMA was among the most often-visited non- entertainment sites in Slovenia. At its peak, there were over 200.000 visits monthly, which represents one tenth of country’s population. Most visitors were driven by unemployment, or by the desire to find a better job. 6. E-trading information gathering Another implementation of our system is ShinA (SHoppINg Assistant)–a mediator system that automatically collects product description from a number of online stores on user’s behalf. ShinA performs a different task than EMA – that of a shopping assistant. ShinA was reprogrammed from scratch, but still based on experiences and the common algorithms of EMA. ShinA’s crucial ability to semi-automatically gather information from various sites replaces hand-coded wrappers of EMA. (Yang, Lee & Choi, 2000). INPUT by User Fig. 5. Workflow diagram of the ShinA system. 6.1. Workflow of the agent The mediator system enables a user to enter two types of data. The first type is a new e-store URL address and the second type is a request for multiple- store product information (Fig. 5). In the first case, where a user enters a new e-store URL address, the system observes the user’s communication with the e-commerce site. Since e- commerce sites have various input mechanisms, the system must locate the position of a query form that accepts user’s search-for-item requests, and must be able to intelligently parse at least the most relevant parts of it. The user’s communication with the e- commerce site finishes either by entering a search query into an input field of the form or by selecting an ontology (i.e., product catalogue link). This suffices to generate a query template. In this way, ShinA generates an appropriate query template for each online store. Afterwards, the mediator system applies a learning algorithm, described in detail in Section 6.3. In the learning phase, appropriate extraction rules are created that correspond to a most representative pattern (i.e. description of an item). In this phase, ontologies are also recognized and extracted. Evaluation and testing of extracted rules is performed by extracting product information from a randomly retrieved e-store’s product page. In case when the extraction of products’ description is unsuccessful, new pattern and extraction rules have to be discovered. The system must be able to ignore redundant and unnecessary fragments of a page. Furthermore, it has to delineate a product description and recognize the attributes of a product such as the price, the manufacturer, the availability, the special offer, the size, etc. Next, the e-store is classified into one or more corresponding clusters (used for later search of items), using learned ontologies and basic information about the e-store. Finally, basic information, query template, rules and ontologies are stored into a knowledge database. Type of input ObserveUser() User finished RuleExtraction() (wrapper generation) OntologyExtraction() Test extracted rules (wrapper) StoreClustering() SAVE: - E-store basic information - E-store query template - Extracted rules (wrapper) - Ontologies Retrieve() for all e-stores in a cluster (except those already stored) CreateQuery() Merge both list of queries (created & stored) MappAndMatch() SAVE: - Newly generated item-query - Item information OUTPUT (multiple-store item-related information results) query for item search new e-store’s URL address no yes wrong rules (wrapper) rules OK (wrapper) Fig. 6. An overall architecture of ShinA. In the second case, where a user enters a keyword search for an item, the system first checks whether an equal or similar search has ever been requested. If so, each stored query (linked with user’s item search) is modified accordingly and added to the list of potential queries. Otherwise, the system takes the most promising cluster of e-stores and creates queries based on query templates. The created queries are then added to the list of potential queries. Next, each potential query is performed on the corresponding e- store, which responds with a page, hopefully containing information about sought after items with their descriptions. The system tries to extract product information from each retrieved page with appropriate extraction rules. The successful queries and related product information are then stored into Knowledge database. Finally, product information (from various sites and various formats) is presented to the user in a user- friendly uniform way. 6.2. System architecture The common data structures, the algorithm and ontologies have already been described in Section 4. Here we describe ShinA in this framework as a particular implementation of the algorithm. Fig. 6 shows the architecture of our ShinA comparison- shopping agent. The task of the observing module ObserveUser() is to transform requested pages into a form that enables the system to control and supervise user’s communication with e-store sites. The wrapper generator RuleExtraction() is the main learning module that constructs a wrapper for each particular store. The wrapper generator learns two things: it recognizes a store’s query scheme and ontologies, and it learns how to extract a store’s content. The wrapper interpreter Retrieve() is a module that executes generated wrappers to get product information. This module forms several actual queries by combining each store’s query template with the keywords that the user actually typed in, and sends them to the corresponding shopping sites. The search results from the stores are then collected and fed to the uniform output generator module MappAndMatch(). The output generator integrates all the search results and generates a uniform output. Information such as extracted rules, query templates, ontologies, item-related information, e-store-related information, etc., is stored into the Knowledge database. 6.3. Learning Algorithm The main learning task specific to ShinA is to build a rule from one or several resulting pages. Such a rule is used to learn the format of product description from successful searches at a new shopping site. We use an inductive learning mechanism to accomplish this task. Here, the examples correspond to the pages of a search result, and the concept to be learned corresponds to the extraction rule (Yang, Seo & Choi, 2001). Each page contains one or more product descriptions that matched the query. A product description is composed of a sequence of product attributes. For example, a music store displays search results in which the attributes are the CD title, the artist name, the price, etc. The wrapper-learning module has to find the starting and ending position of the list of product attributes within the result page, and recognize the pattern of a product description. To do this, our method is divided into three phases. In the first phase, the HTML source of the page is partitioned into three parts. The first and the last part are redundant and irrelevant fragments of the page (header, advertisement, script) that must be ignored. The useful middle part must be extracted, and consists of a set of attributes that describe the product. If there is more than just one product description on the resulting page, more different patterns of product representations may occur. In the next phase, the algorithm recognizes product attributes by examining HTML tags (delimiters) and categorizes them, accordingly. The product description is thus viewed as a sequence of categories, and the algorithm finds repeating patterns in it. The most frequent pattern becomes the representative product description. 6.4. Implementation ShinA’s Web interface is shown in Fig. 1. Unlike EMA, ShinA is a dedicated system without modules for translation, speech, and large amount of locally stored data. However, the core task of parsing and understanding heterogeneous sites is more difficult in e-commerce than in employment because of higher diversity of e-stores. The ShinA system has been implemented on a Windows 2000 operating system, running Apache HTTP Server, version 1.3. As developing tools, JDK 1.3.1 and Jakarta Tomcat 3.2 were used. The system is written in Java and is based on servlet/JSP. It consists of approximately 10 thousand lines of code. Tomcat is a servlet container that is used in the official Reference Implementation for the Java Server and JavaServer Pages technologies. Both Apache Server and Tomcat are developed in an open and participatory environment and therefore freely available. 7. Related work BargainFinder (Krulwich, 1995) and Jango (Jango) are first-stage comparison shoppers. They specify functions that agents must have in order to be applied to Electronic Commerce. Both systems employ the manual rule extraction method. ShopBot (Doorenbos, Etzioni & Weld, 1997), like our system, suggests an automatic rule extraction technique by analyzing and learning the shopping sites. In order to integrate specific product information, Shopbot first removes irrelevant information such as advertisements by using an inductive learning mechanism and then extracts necessary product information. Shopbot, however, uses several strong biases about the structure of HTML and the display format of product search results to learn. Therefore, Shopbot is unable to learn a shopping site that does not conform to these strong biases. By contrast, the only bias in our method is that the result of a product search should be displayed in a semi-structured way, that is, each product description unit has the same output format, which conforms to almost all shopping sites. We expect our method to function more robustly than Shopbot. The thesis can be verified by empirically testing the success rate of correct wrapper constructions. We plan to measure the performance of our agents in the near future. PersonaLogic is a comparison-shopping system that compares specific products themselves rather than shopping sites. Kasbah, AuctionBot, and Tete-a-Tete (see References) are negotiation mediators with which users can buy and sell products based on negotiation strategies between agents in the virtual marketplace. They do not enter e-commerce sites in an automatic or semi-automatic way like our agents do. To our knowledge, today’s commercial comparison shoppers, including MySimon, PriceWatch, and BottomDollar all employ manual rule extraction methods, and consequently suffer from reduced scalability, and the ability to incorporate new e-stores and adapt to changes in incorporated stores. 8. Conclusion We have proposed a general domain-dependent information-gathering agent. The system has been implemented in several versions of the two basic applications: EMA for employment, and ShinA for e- commerce. The agent successfully constructs correct wrappers for “reasonable” domain-dependent databases without assuming many structural constraints. The agent contains a quick, simple and robust inductive learning algorithm that automatically generates wrappers. The extension of the learning method is memory-based learning of success rates of particular queries. This makes it possible to adapt to user habits and to changes at visited sites. There are some limitations in our current system. Firstly, we have assumed that users give a proper URL and path for the test query. Secondly, each major concept must contain a uniform structure. We think that this is not a severe restriction since most domain-related databases stores that produce semi- structured entity information, contain the same attributes. Thirdly, we only extract the major information from an entity description that may also contain several other attributes. Lastly, the system relies heavily on HTML. If an Internet site provides information exclusively by embedding it in graphics or using Java, the system will be unable to handle the site. Overall, the idea of semi-automatic information- gathering from heterogeneous Internet-based databases is becoming technologically mature. The problems we have been facing during implementations were not of principal matter. Technically, similar systems seem to be close to full- scale implementation. One can imagine a uniform e- shopping agent visiting and gathering information from all e-shopping databases on the Internet. We believe that this technique can also be applied to other information integration systems for heterogeneous information sources. Acknowledgement: Financial support was provided by an international project INCO-Copernicus 960154, Cooperative Research in Information Infrastructure, CRII, by the Ministry of education, science and sports in Slovenia, and by ESS. We would like to thank the CEO of the Employment Service of Slovenia, Mr. J. Glazer. References Ashish, N., & Knoblock, C. (1997). Semi-Automatic Wrapper Generation for Internet Information Sources. In Proceedings of Second IFCIS Conference on Cooperative Information Systems (CoopIS). Atzeni, P., Mecca, G., & Merialdo, P. (1997). Semi- structured and Structured Data in the Web: Going Back ad Forth. ACM SIGMOD Workshop on Management of Semistructured Data, pp.1-9. AuctionBot. http://auction.eecs.umich.edu/ Balabanovic, M., & Shoham, Y. (1995). Learning Information Retrieval Agents: Experiments with Automated Web Browsing. In Proc. of the AAAI Spring Symp. on Information Gathering from Heterogeneous, Distributed Environments. Baek, S., Liebowitz, J., Prasad, S.Y., & Granger, M.J. (1999). Intelligent Agents for Knowledge Management - Toward Intelligent Web-Based Collaboration within Virtual Teams. In Knowledge Management Handbook, (ed. Jay Liebowitz), CRC Press, (11-1 - 11-23). Washington, DC. Basu, C., Hirsh, H., & Cohen, W. (1998). Recommendation as classification: Using social and content-based information in recommendation. In Proceedings of the 1998 Workshop on Recommender Systems, AIII Press, pp.11-15. Bezek, A., & Gams, M. (2001). An agent that understands job description. Informatica (Ljubljana), 25(1), pp.99- 105. BottomDollar. http://www.bottomdollar.com/ Bradshaw, J. M. (1997). Software Agents, AAAI Press, Menlo Park, California. Cowie, J., & Lehnert, W. (1996). Information Extraction. Comm. of the ACM, 39(1), pp.80-101. Dastani, M., Jacobs, N., Jonker, Catholyn, M., & Truer, J. (1999). Modelling User Preferences and Mediating Agents in Electronic Commerce. In Proc. of the Agent- Mediated Electronic Commerce (AMEC) SIG-Meeting of AGENTLINK. Doorenbos, R., Etzioni, O., & Weld, D. (1997). A Scalable Comparison-Shopping Agent for the World Wide Web. In Proceedings of the First International Conference on Autonomous Agents, pp.39-48. Freitag, D. (1998). Information Extraction from HTML: Application of a General Machine Learning Approach. In Proceedings of the 15th National Conference on Artificial Intelligence (AAAI-98). Gams, M. (2001). A Uniform Internet-Communicative Agent, Electronic Commerce Research, 1, Kluwer Academic Publishers, pp.69-84. Greenwald, A.R., & Kephart, J.O. (1999). Shopbots and pricebots. In Proceedings of the 16th International Joint Conference on Artificial Intelligence, pp.506-511. Gutmann, R. H., Moukas, A. G., & Maes, P. (1998). Agents as Mediators in Electronic Commerce, Electronic Markets, 8(1), pp.22-27. Hammer, J., Garcia-Molina, H., Nestorov, S., Yerneni, R., Breunig, M., & Vassalos, V. (1997). Template-based wrappers in the TSIMMIS system. In Proceedings of the ACM SIGMOD International Conference on Management of Data, pp.532-535. Jakobsson, M., & Yung, M. (1998). On assurance structures for WWW commerce. In Financial Cryptography ’98. Jango. http://www.jango.com/ Kasbah. http://kasbah.media.mit.edu/ Kitano, H. (1993). Challenges of Massive Parallelism, In Proceedings of the 13th International Joint Conference on Artificial Intelligence, pp.813-834. Krulwich, B. (1995). Bargain Finder agent prototype. Technical Report, Anderson Consulting. http://bf.cstar.ac.com/bf/ Kusmerick, N., Weld, D.S., & Doorenbas, R. (1997). Wrapper Induction for Information Extraction. In International Joint Conference on Artificial Intelligence, pp.729-735. Levy, A.Y., & Weld, D.S. (2000). Intelligent Internet Systems. Artificial Intelligence, 118, pp.1-14. Liebowitz, J., (ed.). (1999). Knowledge Management Handbook, CRC Press, Washington, DC. Liu, L., Pu, C., & Han, W. (2000). XWRAP: An XML- enabled Wrapper Construction System for Web Information Sources. In Proceedings of the Sixteenth International Conference on Data Engineering, pp. 611-621. Lewis, T. (1999). Microsoft Rising : ... and Other Tales of Silicon Valley, ACM. Maes, P. (1994). Agents that reduce work and information overload. Comm. of the ACM, 37(7), pp.31-40. Maes, P., Guttman, R., & Moukas, A. (1999). Agents that buy and sell. Comm. of the ACM, 42(3), pp.81-91. Mladenic, D. (1999). Text-learning and related intelligent agents. IEEE EXPERT, Special Issue on Applications of Intelligent Information Retrieval. Muslea, I., Minton, S., & Knoblock, C. (1998). Wrapper Induction of Semistructured, Web-based Information http://www.bottomdollar.com/ http://www.jango.com/ Sources. In Proceedings of the Conference on Automatic Learning and Discovery. Muslea, I., Minton, S., & Knoblock, C. (1999). A Hierarchical Approach to Wrapper Induction. In Proceedings of the Third International Conference on Autonomous Agents, pp.190-197. MySimon. http://www.mysimon.com/ Pazienza, M.T. (ed.), (1997). Information Extraction, Springer. PersonaLogic. http://www.personalogic.com/ Pivk, A. (2001). Item-based Collaborative Filtering Algorithms, In Proc. of the Fourth International Conference on Information Society, pp. 65-68. Pivk, A., & Gams, M. (2000). E-Commerce Intelligent Agents, In Proceedings of ICTEC’00, pp.418-429. PriceWatch. http://www.pricewatch.com/ Schafer, J., Konstan, J., & Riedl, J. (1999). Recommender systems in e-commerce. In Proceedings of ACM Conference on E-Commerce, pp.158-166. Shoham, Y. (1997). An Overview of Agent-oriented Programming, In Software Agents (ed. Bradshaw, J. M.), AAAI Press, Menlo Park, California, pp.271-290. Tete-a-Tete. http://ecommerce.media.mit.edu/tete-a-tete/ Yang, J., Seo, H., & Choi, J. (2001). MORPHEUS: A Customized Comparison Shopping Agent. In Proceedings of the 5th International Conference on Autonomous Agents (Agents-2001), pp. 63-64. Yang, J., Lee, E., & Choi, J. (2000). A Shopping Agent that Automatically Constructs Wrappers for Semi- Structured Online Vendors. Lecture Notes in Computer Science, Vol. 1983, pp.368-373. http:/// http://www.personalogic.com/ http://ecommerce.media.mit.edu/tete-a-tete/ Introduction Agents Wrappers Information-gathering algorithm Data structures Algorithm Employment information gathering E-trading information gathering Workflow of the agent System architecture Learning Algorithm Implementation Related work Conclusion 
work_72uszokn2vf3rnkfho2i2uuf4i	----	USING POSSIBILITY THEORY IN EXPERT SYSTEMS by Prakash P. Shenoy School of Business, University of Kansas, Lawrence, KS 66045-7585, USA Tel: (785)-864-7551, Fax: (785)-864-5328, Email: PSHENOY@KU.EDU ABSTRACT This paper has two main objectives. The first objective is to give a characterization of a qualitative description of a possibility function. A qualitative description of a possibility function is called a consistent possibilistic state. The qualitative descrip- tion and its characterization serve as qualitative semantics for possibility functions. These semantics are useful in representing knowledge as possibility functions. The second objective is to describe how Zadeh’s theory of possibility fits in the frame- work of valuation-based systems (VBS). Since VBS serve as a framework for managing uncertainty and imprecision in expert systems, this facilitates the use of possibility theory in expert systems. Key Words: Possibility theory, consistent possibilistic state, valuation-based systems, expert systems 1. INTRODUCTION The main goal of this paper is to describe how Zadeh’s theory of possibility can be used for man- aging uncertainty and imprecision in expert systems. We give a characterization of a qualitative de- scription of a possibility function. This is a useful first step in representing knowledge as possibil- ity functions. Also, we describe how possibility theory fits in the framework of valuation-based systems (VBS). Since VBS serve as a framework for managing uncertainty and imprecision in ex- pert systems, this facilitates the use of possibility theory in expert systems. The theory of possibility was first described by Zadeh [1978, 1979] and further developed by Dubois and Prade [1988]. In possibility theory, the basic representational unit is called a possibility function. The two main operations for manipulating possibility functions are called projection and particularization [Zadeh 1979]. In this paper, we are concerned with two issues related to using possibility theory in expert systems. The first issue is in representing knowledge as possibility functions. To do this effec- tively, we need semantics for possibility functions. What does it mean, for example, to say that Appeared in: Fuzzy Sets and Systems, Vol. 52, Issue 2, 10 Dec. 1992, pp. 129--142. 2 Using Possibility Theory in Expert Systems possibility of a proposition is 0.8? In probability theory, for example, the probability of a proposi- tion can be interpreted as a relative frequency or as a betting rate. With this goal in mind, we first define a qualitative description of a possibility function called a consistent possibilistic state. This is analogous to the definition of a consistent epistemic state [Shenoy 1991a] in Spohn’s theory of epistemic beliefs [Spohn 1988]. Next, we give a characterization of a consistent possibilistic state in terms of what we call the content of a consistent possibilistic state. Again, this characterization is analogous to Spohn’s characterization of a consistent epistemic state in terms of its content [Spohn 1988]. This is only one step toward developing semantics for possibility functions. Semantics for the quantitative aspects of possibility functions have been studied by, for example, Dubois et al. [1989] and Ruspini [1991]. The second issue related to using possibility theory in expert systems is making the correspon- dence between concepts in possibility theory and concepts in expert systems. If we interpret pos- sibility functions as knowledge, what do the projection and particularization operations mean in the context of expert systems? How do we use these operations to make inferences in a knowledge- based system? Also, since expert systems typically deal with many propositions, there is a compu- tational issue of the tractability of using possibility theory to solve large problems. To help answer these questions, we describe how possibility theory fits in the framework of VBS. The framework of VBS was first defined by Shenoy [1989] as a general language for incorpo- rating uncertainty in expert systems. It was further elaborated in [Shenoy 1991c] to include axioms that permit local computation in solving a VBS, and a fusion algorithm for solving a VBS using lo- cal computation. VBS encode knowledge using functions defined on frames of variables. The functions are called valuations. VBS include two operators called combination and marginalization that operate on valuations. Combination corresponds to aggregation of knowledge. Marginalization corresponds to coarsening of knowledge. The process of reasoning in VBS can be described sim- ply as finding the marginal of the joint valuation for each variable in the system. The joint valuation is the valuation obtained by combining all valuations. In systems with many variables, it is compu- tationally intractable to explicitly compute the joint valuation. However, if combination and marginalization satisfy certain axioms, it is possible to compute the marginals of the joint valuation without explicitly computing the joint valuation. The framework of VBS is general enough to represent many domains such as Bayesian prob- ability theory [Shenoy 1991c], Dempster-Shafer’s theory of belief functions [Dempster 1967, Shafer 1976, Shenoy 1991d], Spohn’s theory of epistemic beliefs [Spohn 1988, Shenoy 1991c], discrete optimization [Shenoy 1991b], propositional logic [Shenoy 1990a, 1990b], constraint satisfaction [Shenoy and Shafer 1988], and Bayesian decision theory [Shenoy 1990c, 1990d]. Saffiotti and Umkehrer [1991] describe an efficient implementation of VBS called Pulcinella. The correspondence between possibility theory and VBS is as follows. The particularization operation in possibility theory corresponds to the combination operation in VBS. And the projec- tion operation in possibility theory corresponds to the marginalization operation in VBS. This cor- Introduction 3 respondence was first pointed out explicitly by Dubois and Prade [1991]. They also state that the combination and marginalization operations in possibility theory satisfy the three axioms that en- able the use of local computation in computing marginals of the joint possibility function. In this paper, we reiterate this correspondence in greater detail. An outline of this paper is as follows. Section 2 describes the framework of VBS. Section 3 contains an axiomatic definition and a characterization of a consistent possibilistic state. Section 4 describes the main features of possibility theory in terms of the framework of VBS. Section 5 de- scribes a small example illustrating the use of possibility theory for managing uncertainty and im- precision in expert systems. Section 6 contains some concluding remarks. Finally, section 7 con- tains proofs of all results in the paper. 2. VALUATION-BASED SYSTEMS In this section, we will sketch the basic features of VBS. Also, we describe three axioms that permit the use of local computation, and describe a fusion algorithm for solving a VBS using local computation. 2.1. The Framework This subsection describes the framework of valuation-based systems. In a VBS, we represent knowledge by functions called valuations. Valuations are functions that assign values to the ele- ments of frames for sets of variables. We make inferences in a VBS using two operators called combination and marginalization. We use these operators on valuations. Variables and Configurations. We use the symbol WX for the set of possible values of a variable X, and we call WX the frame for X. We assume that one and only one of the elements of WX is the true value of X. We are concerned with a finite set X of variables, and we assume that all the variables in X have finite frames. Given a nonempty set s of variables, let Ws denote the Cartesian product of WX for X in s; Ws = ×{WX | X∈s}. We call Ws the frame for s. We call the elements of Ws configurations of s. Valuations. Given a subset s of variables, there is a set Vs. We call the elements of Vs val- uations for s. Let V denote the set of all valuations, i.e., V = ∪{Vs | s⊆X}. If σ is a valuation for s, then we say s is the domain of σ. Valuations are primitives in our abstract framework and as such require no definition. But as we shall see shortly, they are objects which can be combined and marginalized. Intuitively, a val- uation for s represents some knowledge about the variables in s. Nonzero valuations. For each s⊆X, there is a nonempty subset Ps of Vs whose elements are called nonzero valuations for s. Let P denote ∪{Ps | s⊆X}, the set of all nonzero valuations. Intuitively, a nonzero valuation represents knowledge that is internally consistent. The notion of nonzero valuations is important as it enables us to define combinability of valuations, and it 4 Using Possibility Theory in Expert Systems allows us to constrain the definitions of combination and marginalization to meaningful operators. An example of a nonzero valuation is a possibility function. Combination. We assume there is a mapping ⊗:V×V → V, called combination, such that (i) if ρ and σ are valuations for r and s, respectively, then ρ⊗σ is a valuation for r∪s; (ii) if either ρ or σ is not a nonzero valuation, then ρ⊗σ is not a nonzero valuation; and (iii) if ρ and σ are both nonzero valuations, then ρ⊗σ may or may not be a nonzero valuation. We call ρ⊗σ the combination of ρ and σ. Intuitively, combination corresponds to aggregation of knowledge. If ρ and σ are valuations for r and s representing knowledge about variables in r and s, respectively, then ρ⊗σ represents the aggregated knowledge about variables in r∪s. For possibility functions, combination is pointwise multiplication [Zadeh 1965]. Marginalization. We assume that for each s⊆X, and for each X∈s, there is a mapping ↓(s−{X}): Vs → Vs–{X}, called marginalization to s–{X} such that (i) If σ is a valuation for s, then σ↓(s–{X}) is a valuation for s–{X}; and (ii) σ↓(s–{X}) is a nonzero valuation if and only if σ is a nonzero valuation. We call σ↓(s–{X}) the marginal of σ for s–{X}. Intuitively, marginalization corresponds to coarsening of knowledge. If σ is a valuation for s representing some knowledge about variables in s, and X∈s, then σ↓(s–{X}) represents the knowl- edge about variables in s–{X} implied by σ if we disregard variable X. In the case of possibility functions, marginalization from s to s–{X} is maximization over the frame for X [Zadeh 1979]. In summary, a valuation-based system consists of a 3-tuple {{σ1, ..., σm}, ⊗, ↓} where {σ1, ..., σm} is a collection of valuations, ⊗ is the combination operator, and ↓ is the marginalization operator. Valuation Networks. A graphical depiction of a valuation-based system is called a valuation network. In a valuation network, variables are represented by circular nodes, and valuations are represented by diamond-shaped nodes. Also, each valuation node is connected by an undirected edge to each variable node in its domain. Figure 1 shows a valuation network for a VBS that consists of valuations σ1 for {W}, σ2 for {W, X}, σ3 for {X, Y}, and σ4 for {Y, Z}. Making Inference in VBS. In a VBS, the combination of all valuations is called the joint valuation. Given a VBS, we make inferences by computing the marginal of the joint valuation for each variable in the system. Valuation-Based Systems 5 Figure 1. A valuation network for a VBS consisting of valuations σ1 for {W}, σ2 for {W, X}, σ3 for {X, Y}, and σ4 for {Y, Z}. W X Y Z σ1 σ2 σ3 σ4 If there are n variables in the system, and each variable has two configurations in its frame, then there are 2n configurations of all variables. Hence, it is not computationally feasible to com- pute the joint valuation when there are a large number of variables. In Section 2.3, we describe an algorithm for computing marginals of the joint valuation without explicitly computing the joint val- uation. So that this algorithm gives us the correct answers, we require combination and marginal- ization to satisfy three axioms. The axioms and the algorithm are described in the next two subsec- tions, respectively. 2.2. Axioms In this section, we state three simple axioms that enable local computation of marginals of the joint valuation. These axioms were first formulated by Shenoy and Shafer [1990]. The axioms are stated slightly differently here. Axiom A1 (Commutativity and associativity of combination): Suppose ρ, σ, and τ are valuations for r, s, and t respectively. Then ρ⊗σ = σ⊗ρ, and ρ⊗(σ⊗τ) = (ρ⊗σ)⊗τ. Axiom A2 (Order of deletion does not matter): Suppose σ is a valuation for s, and suppose X1, X2 ∈ s. Then (σ↓(s–{X1}))↓(s–{X1,X2}) = (σ↓(s–{X2}))↓(s–{X1,X2}). Axiom A3 (Distributivity of marginalization over combination): Suppose ρ and σ are valuations for r and s, respectively. Suppose X∉r, and suppose X∈s. Then (ρ⊗σ)↓((r∪s)–{X}) = ρ⊗(σ↓(s–{X})). One implication of Axiom A1 is that when we have multiple combinations of valuations, we can write it without using parenthesis. For example, (...((σ1⊗σ2)⊗σ3)⊗...⊗σm) can be written simply as ⊗{σi | i = 1, ..., m} or as σ1⊗...⊗σm, i.e., we need not indicate the order in which the combinations are carried out. 6 Using Possibility Theory in Expert Systems If we regard marginalization as a coarsening of a valuation by deleting variables, then axiom A2 says that the order in which the variables are deleted does not matter. One implication of this axiom is that (σ↓(s–{X1}))↓(s–{X1,X2}) can be written simply as σ↓(s–{X1,X2}), i.e., we need not indi- cate the order in which the variables are deleted. Axiom A3 is the crucial axiom that makes local computation possible. Axiom A3 states that the computation of (ρ⊗σ)↓((r∪s)–{X}) can be accomplished without having to compute ρ⊗σ. The combination operation in ρ⊗σ is on the frame for r∪s whereas the combination operation in ρ⊗(σ↓(s–{X})) is on the frame for (r∪s)–{X}. In the next subsection, we describe a fusion algo- rithm that applies this axiom repeatedly resulting in an efficient method for computing marginals. 2.3. A Fusion Algorithm In this subsection, we describe a fusion algorithm for making inferences in a VBS using local computation. Suppose {{σ1, ..., σm}, ⊗, ↓} is a VBS with n variables and m valuations. Suppose that combination and marginalization satisfy the three axioms stated in section 2.2. Suppose we have to compute the marginal of the joint valuation for variable X, (σ1⊗...⊗σm) ↓{X}. The basic idea of the fusion algorithm is to successively delete all variables but X from the VBS. The variables may be deleted in any sequence. Axiom A 2 tells us that all deletion sequences lead to the same answers. But, different deletion sequences may involve different computational costs. We will comment on good deletion sequences at the end of this section. When we delete a variable, we have to do a “fusion” operation on the valuations. Consider a set of k valuations ρ1, ..., ρk. Suppose ρi is a valuation for ri. Let FusX{ρ1, ..., ρk} denote the collection of valuations after fusing the valuations in the set {ρ1, ..., ρk} with respect to variable X. Then FusX{ρ1, ..., ρk} = {ρ ↓(r−{X})}∪{ρi | X∉ri} where ρ = ⊗{ρi | X∈ri}, and r = ∪{ri | X∈ri}. After fusion, the set of valuations is changed as follows. All valuations that bear on X are combined, and the resulting valuation is marginalized such that X is eliminated from its domain. The valuations that do not bear on X remain unchanged. We are ready to state the main theorem that describes the fusion algorithm Theorem 1 [Shenoy 1991c]. Suppose {{σ1, ..., σm}, ⊗, ↓} is a VBS where σi is a valuation for si, and suppose ⊗ and ↓ satisfy axioms A1–A3. Let X denote s1∪...∪sm. Suppose X∈X, and suppose X1X2...Xn–1 is a sequence of variables in X–{X}. Then (σ1⊗...⊗σm) ↓{X} = ⊗{FusXn–1{... FusX2{FusX1{σ1, ..., σm}}}}. To illustrate Theorem 1, consider the VBS shown in Figure 1. Suppose we need to compute the marginal of the joint for Z, (σ1⊗...⊗σ4) ↓{Z}. Consider the deletion sequence WXY. After Valuation-Based Systems 7 fusion with respect to W, we have (σ1⊗σ2) ↓{X} for {X}, σ3 for {X, Y}, and σ4 for {Y, Z}. After fusion with respect to X, we have ((σ1⊗σ2) ↓{X}⊗σ3) ↓{Y} for {Y}, and σ4 for {Y, Z}. Finally, after fusion with respect to Y, we have (((σ1⊗σ2)↓{X}⊗σ3)↓{Y}⊗σ4)↓{Z} for {Z}. Theorem 1 tells us that (((σ1⊗σ2)↓{X}⊗σ3)↓{Y}⊗σ4)↓{Z} = (σ1⊗...⊗σ4)↓{Z}. Thus, instead of doing combinations on the frame for {W, X, Y, Z}, we do combinations on the frame for {W, X}, {X, Y}, and {Y, Z}. The fusion algorithm is shown graphically in Figure 2. Figure 2. The first valuation network shows the initial VBS. The second network is the result after fusion with respect to W. The third network is the result after fusion with respect to X. The fourth network is the result after fusion with respect to Y. W X Y Z σ 1 X Y Z Y Z Z σ 2 σ 4 σ 3 σ 3 σ4 σ 4 (σ1⊗σ2) ↓{X} ((σ1⊗σ2) ↓{X}⊗σ3) ↓{Y} (((σ1⊗σ2)↓{X}⊗σ3)↓{Y}⊗σ4)↓{Z} If we can compute the marginal of the joint valuation for one variable, then we can compute the marginals for all variables. We simply compute them one after the other. It is obvious, however, 8 Using Possibility Theory in Expert Systems that this will involve much duplication of effort. Shenoy and Shafer [1990] describe an efficient algorithm for simultaneous computation of all marginals without duplication of effort. Regardless of the number of variables in a VBS, we can compute marginals of the joint valuation for all vari- ables for roughly three times the computational effort required to compute one marginal [Shenoy and Shafer 1990]. Deletion Sequences. Different deletion sequences may involve different computational ef- forts. For example, consider the VBS in the above example. In this example, deletion sequence WXY involves less computational effort than, for example, XYW, as the former involves combi- nations on the frame for two variables only whereas the latter involves combination on the frame for three variables. Finding an optimal deletion sequence is a secondary optimization problem that has shown to be NP-complete [Arnborg et al. 1987]. But, there are several heuristics for finding good deletion sequences [Kong 1986, Mellouli 1987, Zhang 1988, Kjærulff 1990]. One such heuristic is called one-step-look-ahead [Kong 1986]. This heuristic tells us which variable to delete next. As per this heuristic, the variable that should be deleted next is one that leads to combination over the smallest frame. For example, in the VBS described above, if we as- sume that each variable has a frame consisting of two configurations, then this heuristic would pick W over X and Y for first deletion since deletion of W involves combination on the frame for {W, X} whereas deletion of X involves combination on the frame for {W, X, Y}, and deletion of Y in- volves combination on the frame for {X, Y, Z}. After W is deleted, for second deletion, this heuristic would pick X over Y. Thus, this heuristic would choose deletion sequence WXY. 3. CONSISTENT POSSIBILISTIC STATES In this section, first we define axiomatically a consistent possibilistic state. A consistent pos- sibilistic state serves as a qualitative description of a possibility function. The definition of a con- sistent possibilistic state is analogous to the definition of a consistent epistemic state in Spohn’s theory of epistemic beliefs [Shenoy 1991a, Spohn 1988]. Next, we give a characterization of a consistent possibilistic state. Again, this characterization is analogous to the characterization of a consistent epistemic state given by Spohn [1988]. Before we define a consistent possibilistic state, we need to establish our notation. Consider a variable X with frame WX. Suppose WX is defined such that the propositions re- garding X that are of interest are precisely those of the form ‘The true value of X is in s,’ where s is a subset of WX. Thus the propositions regarding X that are of interest are in a one-to-one corre- spondence with the subsets of WX [Shafer 1976, p. 36]. The correspondence between subsets and propositions is useful since it translates the logical notions of conjunction, disjunction, implication, and negation into the set-theoretic notions of inter- section, union, inclusion, and complementation [Shafer 1976, pp. 36-37]. Thus, if r and s are two subsets of WX, and r' and s' are the corresponding propositions, then r∩s corresponds to the Consistent Possibilistic States 9 conjunction of r' and s', r∪s corresponds to the disjunction of r' and s', r⊆s if and only if r' im- plies s', and r is the set-theoretic complement of s with respect to WX (written as r = ~s) if and only if r' is the negation of s'. Notice also that the proposition that corresponds to ∅ is false, and the proposition that corresponds to WX is true. Henceforth, we will simply refer to a proposition by its corresponding subset. The set of subsets of WX will be denoted by 2 WX. In a possibilistic state for X, propositions are said to be either possibly true (or simply, possi- ble) or possibly false (or simply, not possible). Thus a possibilistic state for X is an assignment of labels ‘possible’ and ‘not possible’ to each proposition in WX. Logical consistency requires that a possibilistic state satisfy certain conditions (axioms). A definition of a consistent possibilistic state is as follows. Definition 1. A possibilistic state for X is said to be consistent if the following four axioms are satisfied: B1. For any proposition s, one and only one of the following conditions holds: (i) s is possible. (ii) s is not possible. B2. WX is possible, and ∅ is not possible. B3. If s is not possible and r⊆s, then r is not possible. B4. If r and s are not possible, then r∪s is not possible. Some simple consequences of Definition 1 are as follows. Proposition 1. The following conditions always hold in any consistent possi- bilistic state: B5. If s is possible and r⊇s, then r is possible. B6. If s is not possible, then ~s is possible. B7. If s is possible, then this does not necessarily imply that ~s is not possible; ~s could also be possible. It is clear from axiom B1 that we can specify a consistent possibilistic state simply by listing all propositions that are not possible. Then axiom B1 tells us that the remaining propositions are pos- sible. Let i denote the set of all propositions (subsets) that are not possible. Theorem 2 gives a characterization of i. Theorem 2. Suppose i denotes the set of all propositions that are not possible in some possibilistic state for X. Then the possibilistic state is consistent if and only if there exists a unique proper subset c of WX such that i = {s∈2 WX | s ⊆ c}. Subset c in Theorem 2 is called the content of the consistent possibilistic state. Note that the content constitutes a complete specification of a possibilistic state. Thus a simple corollary of 10 Using Possibility Theory in Expert Systems Theorem 2 is that in a frame WX consisting of n elements, there are exactly 2 n−1 distinct possible consistent possibilistic states (corresponding to each proper subset of W as the content). The con- tent c represents the largest subset (proposition) that is not possible. Example 1. (Consistent possibilistic states) Consider a frame W = {x, y, z}. Table 1 lists all seven possible consistent possibilistic states. Possibilistic state 1 represents a state of complete ig- norance in which every proposition (except the null proposition) is possible. Possibilistic states 5, 6, and 7 represent a state of complete information where one and only one element of the frame is possibly true. Possibilistic states 2, 3, and 4 are states of partial information where one element of the frame is not possible Table 1. The seven consistent possibilistic states for a variable with frame W = {x,y,z}. State Content c Possible Not Possible i 1 ∅ {x}, {y}, {z}, {x, y}, {y, z}, {x, z}, {x, y, z} ∅ 2 {x} {y}, {z}, {x, y}, {x, z}, {y, z}, {x, y, z} ∅, {x} 3 {y} {x}, {z}, {x, y}, {x, z}, {y, z}, {x, y, z} ∅, {y} 4 {z} {x}, {y}, {x, y}, {x, z}, {y, z}, {x, y, z} ∅, {z} 5 {x, y} {z}, {x, z}, {y, z}, {x, y, z} ∅, {x}, {y}, {x, y} 6 {x, z} {y}, {x, y}, {y, z}, {x, y, z} ∅, {x}, {z}, {x, z} 7 {y, z} {x}, {x, y}, {x, z}, {x, y, z} ∅, {y}, {z}, {y, z} 4. POSSIBILITY THEORY In this section, we describe the main features of Zadeh’s theory of possibility in terms of the framework of VBS described in the previous section, i.e., we will use the terminology of VBS in- stead of the terminology used in [Zadeh 1979]. Thus the operation Zadeh calls particularization we call combination, and the operation Zadeh calls projection we call marginalization. The basic unit of knowledge representation is called a possibility function. Definition 2 . A possibility function π for p is a function π: 2Wp → [0, 1] such that P1. There exists a configuration w∈Wp such that π({w}) = 1; P2. For any r∈(2Wg − {∅}), π(r) = MAX{π({w}) | w∈r}; and P3. π(∅) = 0. Possibility Theory 11 Note that although a possibility function is defined for the set of all subsets of Wp, it is com- pletely specified by its values for each singleton subset of Wp. A possibility function is a complete representation of a consistent possibilistic state. To see what propositions are possible and what propositions are not possible in state π, consider the sub- set c = {w∈Wp | π({w}) < 1}. By condition P1 in Definition 2, c is always a proper subset of Wp. c represents the content of the possibilistic state π, i.e., r is not possible in state π iff r ⊆ c. Thus r is possible in state π iff π(r) = 1, and r is not possible in state π iff π(r) < 1. A possibility function consists of more than a representation of a consistent possibilistic state. It also includes degrees to which proposition are possible and degrees to which propositions are not possible. π(r) can be interpreted as the degree to which proposition r is possible, and 1−π(r) can be interpreted as the degree to which proposition r is not possible, i.e., r is more possible than t if π(r) > π(t) and conversely, r is more impossible than t if π(r) < π(t) < 1. Consider the following possibility function for s: π({w}) = 1 for all w∈Ws. This means that the only proposition that is not possible is ∅ (and all other propositions are possible). We shall call such a possibility function vacuous. It represents a state of complete ignorance. Proposition 2. Suppose π is a possibility function for p. Then P4. For each r ∈ 2Wp, either π(r) = 1 or π(~r) = 1 or both. P5. For each r,t∈ 2Wp, π(r∪t) = MAX{π(r), π(t)} Extension of Subsets. By extension of a subset of a frame to a subset of a larger frame, we mean a cylinder set extension. If r and s are sets of variables, r⊆s, and r is a subset of Wr, then the extension of r to s is r×Ws−r. We will let r↑s denote the extension of r to s. For example, consider three variables R, P, Q with frames WR = {r, ~r}, WP = {p, ~p}, and WQ = {q, ~q}, respectively. Then the extension of {(r,~p), (~r,p)} (which is a subset of W{R,P}) to {R,P,Q} is {(r,~p,q), (r,~p,~q), (~r,p,q), (~r,p,~q)}. Note that the propositions corresponding to r and r↑s are logically equivalent. Marginalization. Suppose π is a possibility function for p. Suppose q ⊆ p. We may be in- terested only in propositions about variables in q. In this case, we would like to marginalize π to q. The following definition of marginalization is motivated by the fact that each proposition q∈2Wq about variables in q can be regarded as a proposition q↑p ∈ 2Wp about variables in p. Definition 3 (Marginalization). Suppose π is a possibility function for p, and suppose q ⊆ p. The marginal of π for q, denoted by π↓q, is a possibility function for q given as follows: π↓q(q) = π(q↑p) = MAX{π({(x,y)}) | x∈q, y∈Wp−q} (1) for all q∈2Wq. 12 Using Possibility Theory in Expert Systems In particular, if q is a singleton subset, i.e., q = {x} for some x∈Wq, then (1) simplifies to: π↓q({x}) = MAX{π({(x,y)}) | y∈Wp−q}. Note that if π is a possibility function for p, and X1, X2 ∈ p, then (π ↓(p–{X1}))↓(p–{X1,X2}) = (π↓(p–{X2}))↓(p–{X1,X2}). In words, if we regard marginalization as reduction of π by deletion of variables, then the order in which the variables are deleted makes no difference in the final answer. Example 2. (Possibility function and marginalization) We would like to determine whether a stranger (about who we know nothing about) is a pacifist or not depending on whether he is a Republican or not and whether he is a Quaker or not. Consider three variables R, Q and P. R has two configurations: r (for Republican), ~r (not Republican); Q has two configurations: q (Quaker) and ~q (not Quaker); and P has two configurations: p (pacifist) and ~p (not pacifist). Our knowledge that most (at least 80 percent) Republicans are not pacifists and that most (at least 90 percent) Quakers are pacifists is represented by the possibility function π for {R,Q,P} shown in Table 2 (the construction of this possibility function will be explained later in this section—see Example 3). Note that the marginal of π for R is the vacuous possibility function for R, i.e., π↓{R}({r}) = π↓{R}({~r}) = 1. Thus in possibilistic state π, we have no knowledge whether the stranger is a re- publican or not. Similarly, notice that the marginals of π for {Q} and {P} are also vacuous. The marginal of π for {R, P} is as follows: π↓{R,P}({r,p}) = 0.2, π↓{R,P}({r,~p}) = π↓{R,P}({~r,p}) = π↓{R,P}({~r,~p}) = 1. Thus pacifist Republicans are not possible (to degree 0.8). Similarly, notice that the marginal of π for {Q, P} is as follows: π↓{Q,P}({q,~p}) = 0.1, π↓{Q,P}({q,p}) = π↓{Q,P}({~q,p}) = π↓{Q,P}({~q,~p}) = 1. Thus non-pacifist Quakers are not possible (to degree 0.9). Finally note that the marginal of π for {R, Q} is as follows: π↓{R,Q}({r,q}) = 0.2, π↓{R,Q}({r,~q}) = π↓{R,Q}({~r,q}) = π↓{R,Q}({~r,~q}) = 1. Thus Republican Quakers are not possible (to degree 0.8). Possibility Theory 13 Table 2. A possibility function π for {R,Q,P}. W{R,P,Q} π r p q 0.2 r p ~q 0.2 r ~p q 0.1 r ~p ~q 1.0 ~r p q 1.0 ~r p ~q 1.0 ~r ~p q 0.1 ~r ~p ~q 1.0 Next, we state a rule for combining possibility function. This rule is called particularization by Zadeh [1979]. Zadeh [1965] has given two rules for combining possibility functions—minimization and multiplication. Axiom B2 (WX is possible) requires that we include normalization in the definition of combination. Since minimization followed by normalization is not associative, multiplication followed by normalization is associative, and associativity of combination is a requirement in VBS (axiom A1), we use multiplcation followed by normalization as our definition of combination. Definition 4. Suppose π1 and π2 are possibility functions for p1 and p2, respectively. Suppose K = MAX{π1({w↓p1}) π2({w↓p2}) | w∈Wp1∪p2}. The combination of π1 and π2, denoted by π1⊗π2, is the function for p1∪p2 given by K–1 π1({w↓p1}) π2({w↓p2}) if K ≠ 0 (π1⊗π2)({w}) =  (2) 0 if K = 0 for all w∈Wp1∪p2. If K = 0, then π1⊗π2 is not a possibility function. In this case, π1⊗π2 is not a nonzero valuation. This means that the knowledge in π1 and π2 are inconsistent. If K ≠ 0, then K is a normalization constant that ensures that π1⊗π2 is a possibility function. Thus, combination consists of pointwise multiplication followed by normalization (if normalization is possible). Example 3. (The rule of combination) Consider two pieces of knowledge as follows: 1. Most Republicans (at least 80 percent) are not pacifists. 2. Most Quakers (at least 90 percent) are pacifists. 14 Using Possibility Theory in Expert Systems If π1 is a possibility function representation of the first piece of knowledge, and π2 is a possibility function representation of the second piece of knowledge, then π1⊗π2 will represent the aggrega- tion of these two pieces of knowledge. In particular, suppose π1 is a possibility function for {R,P} as follows: π1({(r,p)}) = 0.2, π1({(r,~p)}) = π1({(~r,p)}) = π1({(~r,~p)}) = 1 (i.e., pacifist Republicans are not possible to degree 0.8), and suppose π2 is a possibility function for {Q,P} as follows: π2({(q,~p)}) = 0.1, π2({(q,p)}) = π2({(~q,p)}) = π2({(~q,~p)}) = 1 (i.e., non-pacifist Quakers are not possible to degree 0.9). Then the possibility function π1⊗π2 = π, say, shown in Table 3, represents the aggregate knowledge. Table 3. The combination of π1 and π2. W{R,P,Q} π1 π2 π1⊗π2 = π r p q 0.2 1.0 0.2 r p ~q 0.2 1.0 0.2 r ~p q 1.0 0.1 0.1 r ~p ~q 1.0 1.0 1.0 ~r p q 1.0 1.0 1.0 ~r p ~q 1.0 1.0 1.0 ~r ~p q 1.0 0.1 0.1 ~r ~p ~q 1.0 1.0 1.0 Proposition 3. The rule of combination described in (2) has the following prop- erties: C1. (Commutativity) π1⊗π2 = π2⊗π1. C2. (Associativity) (π1⊗π2)⊗π3 = π1⊗(π2⊗π3). C3. If π1 is vacuous, then π1⊗π2 = π2. C4. In general, π1⊗π1 ≠ π1. The possibility function π1⊗π2 represents the aggregation of knowledge contained in possibil- ity functions π1 and π2 if π1 and π2 are independent. Like Bayesian probability theory, Dempster- Shafer’s theory of belief function, and Spohn’s theory of epistemic beliefs, possibility theory requires that the possibility functions π1 and π2 be independent. This is because of property C4—double-counting of knowledge may lead to erroneous results. We have already shown that axioms A1 and A2 are valid for possibility functions. Theorem 3 below states that axiom A3 is also satisfied. This result is stated in [Dubois and Prade 1991]. Possibility Theory 15 Theorem 3. Suppose π1 and π2 are possibility functions for p1 and p2, respectively. Suppose X∉p1, and suppose X∈p2. Then (π1⊗π2) ↓((p1∪p2)–{X}) = π1⊗(π2 ↓(p2−{X})). Since all three axioms required for local computation of marginals are satisfied, the fusion al- gorithm described in section 2.3 can be used for reasoning from knowledge expressed as possibil- ity functions. In the context of possibility theory, [Chatalic et al. 1987] describes a local computa- tional method for reasoning with possibility functions. The next section describes a small example to illustrate the use of the fusion algorithm to find marginals. 5. AN EXAMPLE In this section, we describe an example in complete detail to illustrate the use of possibility theory in managing imprecision in expert systems. Is Dick a Pacifist? Consider the following items of evidence. Most Republicans (at least 80 percent) are not pacifists. Most Quakers (at least 90 percent) are pacifists. Dick is a Republican (and we are more than 99 percent certain of this). Dick is a Quaker (and we are more than 99 per- cent certain of this). Is Dick a pacifist or not? We will model the four items of evidence as possibility functions π1 for {R,P}, π2 for {Q, P}, π3 for {R}, and π4 for {Q}, respectively, as displayed in Table 4. Figure 3 shows the valuation network for this example. If we apply the fusion algorithm using deletion sequence RQ, we get (π1⊗π2⊗π3⊗π4) ↓{P} = (π1⊗π3) ↓{P}⊗(π2⊗π4) ↓{P}. The details of the computations are shown in Table 5. The conclusion is that the proposition Dick is a pacifist is possible (to degree 1), and the proposition Dick is not a pacifist is not possible (to degree 0.5). 16 Using Possibility Theory in Expert Systems Table 4. The possibility functions π1, π2, π3, and π4 in the Is Dick a Pacifist? example. W{R,P} π1 W{Q,P} π2 W{R} π3 W{Q} π4 r p 0.2 q p 1 r 1 q 1 r ~p 1 q ~p 0.1 ~r .01 ~q .01 ~r p 1 ~q p 1 ~r ~p 1 ~q ~p 1 Figure 3. The valuation network for the Is Dick a Pacifist? example. R P Q 3π 4π2π1π Table 5. The computation of (π1⊗π3) ↓{P}⊗(π2⊗π4) ↓{P}. W{R,P} π1⊗π3 W{Q,P} π2⊗π4 r p 0.20 q p 1.00 r ~p 1.00 q ~p 0.10 ~r p 0.01 ~q p 0.01 ~r ~p 0.01 ~q ~p 0.01 WP (π1⊗π3) ↓{P} (π2⊗π4) ↓{P} (π1⊗π3) ↓{P} ⊗(π2⊗π4) ↓{P} p 0.20 1.00 1.00 ~p 1.00 0.10 0.50 Proofs 17 6. CONCLUSIONS We have given an axiomatic definition of a qualitative description of a possibility function called a consistent possibilistic state. We have also given a characterization of a consistent possibilistic state in terms of its content. Our objective here is to help develop semantics for possibility theory. Our contribution in this paper is only one step in this direction. We have also described how possibility theory fits in the framework of valuation-based sys- tems. This was first described by Dubois and Prade [1991]. We have expanded on this topic by providing a brief yet complete description of the VBS framework, and by describing the main features of possibility theory in terms of the VBS framework. We have also described a small ex- ample in complete detail to illustrate the use of possibility theory in managing uncertainty and im- precision in expert systems. 7. PROOFS In this section, we give proofs for all results in the paper. First we state and prove a lemma needed to prove Theorem 1. Lemma 1. Suppose {{σ1, ..., σm}, ⊗, ↓} is a VBS where σi is a valuation for si, and suppose ⊗ and ↓ satisfy axioms A1–A3. Let X denote s1∪...∪sm. Suppose X∈X. Then (σ1⊗...⊗σm}) ↓(X–{X}) = ⊗FusX{σ1, ..., σm}. Proof of Lemma 1. Suppose σi is a valuation for si, i = 1, ..., m. Let s = ∪{si | X∈si}, and let r = ∪{si | X∉si}. Let ρ = ⊗{σi | X∉si}, and σ = ⊗{σi | X∈si}. Note that X∈s, and X∉r. Then (σ1⊗...⊗σm) ↓(X–{X}) = (ρ⊗σ)↓((r∪s)–{X}) = ρ⊗(σ↓(s–{X})) (using axiom A3) = (⊗{σi | X∉si})⊗(σ ↓(s–{X})) = ⊗FusX{σ1, ..., σm}. ■ Proof of Theorem 1. By axiom A2, (σ1⊗...⊗σm) ↓{X} is obtained by sequentially marginaliz- ing all variables but X from the joint valuation. A proof of this theorem is obtained by repeatedly applying the result of Lemma 1. At each step, we delete a variable and fuse the set of all valuations with respect to this variable. Using Lemma 1, after fusion with respect to X1, the combination of all valuations in the resulting VBS is equal to (σ1⊗...⊗σm) ↓(X–{X1}). Again, using Lemma 1, after fusion with respect to X2, the combination of all valuations in the resulting VBS is equal to (σ1⊗...⊗σm) ↓(X–{X1,X2}). And so on. When all variables but X have been deleted, we have the result. ■ 18 Using Possibility Theory in Expert Systems Proof of Proposition 1. To show B5, suppose s is possible, and suppose r⊇s. Assume r is not possible. Then from B3, s is not possible contradicting our earlier supposition. Therefore r must be possible. To show B6, suppose s is not possible. If ~s is also not possible, then from B4, s∪~s = WX is not possible contradicting B2. Therefore ~s must be possible. To show that B7 is true, consider a possibilistic state in which the only proposition that is not possible is ∅ (and all other propositions are possible). Clearly, this possibilistic state satisfies ax- ioms B1 to B4. ■ Proof of Theorem 2. (Sufficiency). Assume that the possibilistic state is consistent. Consider the proposition ∪i. It follows from B4 that ∪i is not possible, and follows from B2 that ∪i is a proper subset of WX. Define c = ∪i. We need to show that i = {s∈2 WX | s ⊆ c}. Suppose r∈i. Then r ⊆ ∪i = c, i.e., r∈{s∈2WX | s ⊆ c}. Now suppose r∈{s∈2WX | s ⊆ c}, i.e., r ⊆ c. Then by axiom B3, r∈i. Thus we have shown that i = {s∈2WX | s ⊆ c}. (Necessity). Suppose i = {s∈2WX | s ⊆ c} for some proper subset c of WX. We will show that this possibilistic state satisfies axioms B1 to B4. First note that B1 is satisfied by definition. B2 is satisfied since c is a proper subset of WX. To show B3, suppose that s is not possible and suppose r ⊆ s. Since s is not possible, by hypothesis, s ⊆ c. Therefore r ⊆ c. Therefore r is not possible. To show B4, suppose that r and s are not possible, i.e., r and s belong to i. By hypoth- esis, r ⊆ c, and s ⊆ c. Hence, r∪s ⊆ c. Therefore r∪s ∈ i, i.e., r∪s is not possible. ■ Proof of Proposition 2. P4 follows from condition P1 in Definition 2, and P5 follows from condition P2 in Definition 2. ■ Proof of Proposition 3. All four properties follow trivially from the definition of combination. ■ Proof of Theorem 3. Note that p1∪p2 = (p1–p2)∪(p1∩p2)∪(p2–p1–{X})∪{X}, p1 = (p1–p2)∪(p1∩p2), and p2 = (p1∩p2)∪(p2–p1–{X})∪{X}. Suppose w∈Wp1−p2, u∈Wp1∩p2, and v∈Wp2−p1−{X}, x∈WX. Then (w,u)∈Wp1, and (u,v,x)∈Wp2. First, note that the normalization factor in the combination on the left-hand side, say K1, is the same as the normalization factor in the combination on the right-hand side, say K2, i.e., K1 = K2, as shown below. K1 = MAX{π1({(w,u)}) π2({(u,v,x)}) | (w,u,v,x)∈Wp1∪p2} = MAX{π1({(w,u)}) MAX{π2({(u,v,x)}) | x∈WX} | (w,u,v)∈W((p1∪p2)−{X})} = MAX{π1({(w,u)}) (π2↓(p2−{X}))({u,v}) | (w,u,v)∈W((p1∪p2)−{X})} = K2 = K, say. Next, Proofs 19 (π1⊗π2) ↓((p1∪p2)–{X})({(w,u,v)}) = = MAX{(π1⊗π2)({(w,u,v,x)}) | x∈WX} = MAX{K–1 π1({(w,u)}) π2({(u,v,x)}) | x∈WX} = K–1 π1({(w,u)}) MAX{π2({(u,v,x)}) | x∈WX} = K–1 π1({(w,u)}) (π2 ↓(p2−{X}))({(u,v)}) = (π1⊗(π2 ↓(p2−{X})))({(w,u,v)}). ■ ACKNOWLEDGEMENTS This work was supported in part by the National Science Foundation under grant IRI-8902444. I am grateful to Didier Dubois, Serafin Moral, and Leen-Kiat Soh for comments. REFERENCES Arnborg, S., D. G. Corneil, and A. Proskurowski (1987), “Complexity of finding embeddings in a k-tree,” SIAM Journal of Algebraic and Discrete Methods, 8, 277–284. Dempster, A. P. (1967), “Upper and lower probabilities induced by a multivalued mapping,” Annals of Mathematical Statistics, 38, 325-339. Dubois, D., J. Lang, and H. Prade (1989), “Automated reasoning using possibilistic logic: Semantics, belief revision and variable certainty weights,” Proceedings of the Fifth Workshop on Uncertainty in Artificial Intelligence, 81–87, Windsor, Ontario. Dubois, D. and H. Prade (1988), Possibility Theory: An Approach to Computerized Processing of Uncertainty, Plenum Press, New York, NY. Dubois, D. and H. Prade (1991), “Inference in possibilistic hypergraphs,” in Bouchon-Meunier, B., R. R. Yager and L. A. Zadeh (eds.), Uncertainty in Knowledge Bases, Lecture Notes in Computer Science No. 521, 250–259, Springer-Verlag, Berlin. Chatalic, P., D. Dubois, and H. Prade (1987), “A system for handling relational dependencies in approximate reasoning,” Proceedings of the Third International Expert Systems Conference, 495–502, London, UK. Kjærulff, U. (1990), “Triangulation of graphs—Algorithms giving small total state space,” Technical report R90-09, Institute for Electronic Systems, University of Aalborg, Denmark. Kong, A. (1986), “Multivariate belief functions and graphical models,” Ph.D. dissertation, Department of Statistics, Harvard University, Cambridge, MA. Mellouli, K. (1987), “On the propagation of beliefs in networks using the Dempster-Shafer theory of evidence,” Ph.D. thesis, School of Business, University of Kansas, Lawrence, KS. Ruspini, E. H. (1991), “On the semantics of fuzzy logic,” International Journal of Approximate Reasoning, 5(1), 45–88. Saffiotti, A. and E. Umkehrer (1991), “Pulcinella: A general tool for propagating uncertainty in valuation networks,” in D’Ambrosio, B., P. Smets, and P. P. Bonissone (eds.), Uncertainty in Artificial Intelligence: Proceedings of the Seventh Conference, 323–331, Morgan Kaufmann, San Mateo, CA. Shafer, G. (1976), A Mathematical Theory of Evidence, Princeton University Press, Princeton, NJ. 20 Using Possibility Theory in Expert Systems Shenoy, P. P. (1989), “A valuation-based language for expert systems,” International Journal for Approximate Reasoning, 3(5), 383–411, 1989. Shenoy, P. P. (1990a), “Valuation-based systems for propositional logic,” in Ras, Z. W., M. Zemankova, and M. L. Emrich (eds.), Methodologies for Intelligent Systems, 5, 305–312, North-Holland, Amsterdam. Shenoy, P. P. (1990b), “Consistency in valuation-based systems,” Working Paper No. 216, School of Business, Lawrence, KS. Shenoy, P. P. (1990c), “Valuation-based systems for Bayesian decision analysis,” Working Paper No. 220, School of Business, University of Kansas, Lawrence, KS. To appear in 1992 in Operations Research, 40(3). Shenoy, P. P. (1990d), “A new method for representing and solving Bayesian decision prob- lems,” Working Paper No. 223, School of Business, University of Kansas, Lawrence, KS. To appear in 1991 in Hand, D. J. (ed.), Artificial Intelligence and Statistics III, Chapman & Hall, London, England. Shenoy, P. P. (1991a), “On Spohn’s rule for revision of beliefs,” International Journal of Approximate Reasoning, 5(2), 149–181. Shenoy, P. P. (1991b), “Valuation-based systems for discrete optimization,” in Bonissone, P. P., M. Henrion, L. N. Kanal, and J. Lemmer (eds.), Uncertainty in Artificial Intelligence, 6, 385–400, North-Holland, Amsterdam. Shenoy, P. P. (1991c), “Valuation-based systems: A framework for managing uncertainty in ex- pert systems,” in Zadeh, L. A. and J. Kacprzyk (eds.), Fuzzy Logic for the Management of Uncertainty, John Wiley & Sons, New York, NY. Shenoy, P. P. (1991d), “Using Dempster-Shafer’s belief-function theory in expert systems,” Working Paper No. 234, School of Business, University of Kansas, Lawrence, KS. To appear in Proceedings of the SPIE Conference on Appplications of Artificial Intelligence X, Orlando, FL, April 1992. Shenoy, P. P. and G. Shafer (1988), “Constraint propagation,” Working Paper No. 208, School of Business, University of Kansas, Lawrence, KS. Shenoy, P. P. and G. Shafer (1990), “Axioms for probability and belief-function propagation,” in Shachter, R. D., T. S. Levitt, J. F. Lemmer and L. N. Kanal (eds.), Uncertainty in Artificial Intelligence, 4, 169–198, North-Holland, Amsterdam. Reprinted in Shafer, G. and J. Pearl, eds. (1990), Readings in Uncertain Reasoning, 575–610, Morgan Kaufmann, San Mateo, CA. Spohn, W. (1988), “Ordinal conditional functions: A dynamic theory of epistemic states,” in Harper, W. L. and B. Skyrms (eds.), Causation in Decision, Belief Change, and Statistics, 2, 105–134, D. Reidel Publishing Company, Dordrecht. Spohn, W. (1990), “A general non-probabilistic theory of inductive reasoning,” in Shachter, R. D., T. S. Levitt, J. F. Lemmer and L. N. Kanal (eds.), Uncertainty in Artificial Intelligence, 4, 149–158, North-Holland, Amsterdam. Zadeh, L. A. (1965), “Fuzzy sets,” Information and Control, 8, 338–353. Zadeh, L. A. (1978), “Fuzzy sets as a basis for a theory of possibility,” Fuzzy Sets and Systems, 1, 3–28. Zadeh, L. A. (1979), “A theory of approximate reasoning,” in Ayes, J. E., D. Mitchie, and L. I. Mikulich (eds.), Machine Intelligence, 9, 149–194, Ellis Horwood, Chichester, U.K. Zhang, L. (1988), “Studies on finding hypertree covers of hypergraphs,” Working Paper No. 198, School of Business, University of Kansas, Lawrence, KS. 21 ABOUT THE AUTHOR Prakash P. Shenoy is the Director of the Doctoral Program and Professor in the School of Business at the University of Kansas at Lawrence where he has been since 1978. Before that, he was with the Mathematics Research Center at the University of Wisconsin at Madison for one year. He received a B.Tech. in Mechanical Engineering from the Indian Institute of Technology at Bombay, India in 1973, and a M.S. and a Ph.D. in Operations Research from Cornell University in 1975 and 1977, respectively. His doctoral dissertation titled “On Coalition Formation and Game theory” was written with Professor William F. Lucas as his dissertation advisor. In the Summer of 1984, he attended a six-week Faculty Development Institute on Information Systems at the University of Minnesota organized by the American Association of Collegiate Schools of Business. During 1985–86, he was an Intra-University Visiting Professor in the Department of Computer Science at the University of Kansas. Subsequently, he served as a faculty intern in the MIS Division of Hallmark Cards, Inc., in Kansas City, Missouri, for six months from June to December, 1986. His research interests are in the areas of Artificial Intelligence and Operations Research. He has published many articles on management of uncertainty in expert systems, decision analysis, opti- mization, and the mathematical theory of games. His articles have appeared in journals such as the International Journal of Approximate Reasoning, IEEE Expert, Operations Research, Management Science, and the International Journal of Game Theory. He has received several research grants from the Database and Expert Systems program of the National Science Foundation, the Research Opportunities in Auditing program of the Peat Marwick Main Foundation, the Higher Education Academic Development Donations program of Apple Computer, Inc., and the General Research Fund program of the University of Kansas. He served as a guest-editor for a special 1990 issue of the International Journal of Approximate Reasoning on the subject of Dempster-Shafer’s Theory of Belief Functions in Artificial Intelligence. He serves on the editorial board of Uncertainty in Knowledge Bases: An International Journal, and as an ad-hoc referee for over 25 journals and conferences in Artificial Intelligence and Operations Research. He also serves on the program committees of several conferences in artificial intelligence. His teaching interests are in the areas of optimization, game theory, decision analysis, informa- tion systems, and uncertain reasoning. He has taught graduate-level courses on linear program- ming, non-linear programming, game theory, management information systems, decision support systems, and management of uncertainty in expert systems. He has served on doctoral dissertation committees of eighteen Ph.D. students in Management Science, Marketing, Accounting, Economics, Computer Science, and Geography, two as chairman. In Spring 1991, he was awarded the Mentor Award by the doctoral students in the School of Business at the University of Kansas. 22 SELECTED WORKING PAPERS Unpublished working papers are available from the respective authors at: School of Business University of Kansas Summerfield Hall Lawrence, KS 66045-2003 USA No. 184. “Propagating Belief Functions with Local Computations,” Prakash P. Shenoy and Glenn Shafer, February 1986. Appeared in IEEE Expert, 1(1986), 43-52. No. 190. “Propagating Belief Functions in Qualitative Markov Trees,” Glenn Shafer, Prakash P. Shenoy, and Khaled Mellouli, June 1987. Appeared in International Journal of Approximate Reasoning, 1(1987), 349-400. No. 196. “On the Propagation of Beliefs in Networks Using the Dempster-Shafer Theory of Evidence,” Khaled Mellouli, April 1988. Available as Ph.D. dissertation from University Microfilms, Inc., Ann Arbor, MI. No. 197. “AUDITOR’S ASSISTANT: A Knowledge Engineering Tool for Audit Decisions,” Glenn Shafer, Prakash P. Shenoy, and Rajendra Srivastava, April 1988. Appeared in Auditing Symposium IX: Proceedings of 1988 Touche Ross/University of Kansas Symposium on Auditing Problems, 61–84, School of Business, University of Kansas, Lawrence, KS. No. 198. “Studies on Finding Hypertree Covers of Hypergraphs,” Lianwen Zhang, April 1988. No. 199. “An Axiomatic Framework for Bayesian and Belief-Function Propagation,” Prakash P. Shenoy and Glenn Shafer, July 1988. Appeared in Proceedings of the Fourth Workshop on Uncertainty in Artificial Intelligence, Minneapolis, MN, 307-314. No. 200. “Probability Propagation,” Glenn Shafer and Prakash P. Shenoy, August 1988. Appeared in Annals of Mathematics and Artificial Intelligence, 2(1990), 327-352. No. 201. “Local Computation in Hypertrees,” Glenn Shafer and Prakash P. Shenoy, August 1988. No. 203. “A Valuation-Based Language for Expert Systems,” Prakash P. Shenoy, August 1988. Appeared in International Journal of Approximate Reasoning, 3(1989), 383-411. No. 206. “An Evidential Reasoning System,” Debra K. Zarley, November 1988. No. 208. “Constraint Propagation,” Prakash P. Shenoy and Glenn Shafer, November 1988. No. 209. “Axioms for Probability and Belief-Function Propagation,” Prakash P. Shenoy and Glenn Shafer, November 1988. Appeared in Uncertainty in Artificial Intelligence, 4(1990), 169-198. Reprinted in Readings in Uncertain Reasoning, Glenn Shafer and Judea Pearl, eds. (1990), 575–610, Morgan Kaufmann, San Mateo, CA. No. 210. “The Unity of Probability,” Glenn Shafer, February 1989. Appeared in von Furstenberg, G. M., ed., Acting Under Uncertainty: Multidisciplinary Conceptions, 95-126, Kluwer, 1990. No. 211. “MacEvidence: A Visual Evidential Language for Knowledge-Based Systems,” Yen-Teh Hsia and Prakash P. Shenoy, March 1989. A condensed version of this paper appeared as: Hsia, Y-T. and P. P. Shenoy, “An evidential language for expert systems,” in Ras, Z. W., ed., Methodologies for Intelligent Systems, 4(1989), 9-16. No. 212. “Audit Risk: A Belief-Function Approach,” Rajendra Srivastava, Prakash P. Shenoy and Glenn Shafer, April 1989. 23 No. 213. “On Spohn’s Rule for Revision of Beliefs,” Prakash P. Shenoy, July 1989. Appeared in International Journal of Approximate Reasoning, 5(1991), 149–181. No. 216. “Consistency in Valuation-Based Systems,” Prakash P. Shenoy, February 1990, revised May 1991. No. 218. “Perspectives on the Theory and Applications of Belief Functions,” Glenn Shafer, April 1990. Appeared in International Journal of Approximate Reasoning, 4(1990), 323-362. No. 220. “Valuation-Based Systems for Bayesian Decision Analysis,” Prakash P. Shenoy, April 1990, revised May 1991. To appear in Operations Research, 40(1992). No. 221. “Valuation-Based Systems for Discrete Optimization,” Prakash P. Shenoy, June 1990. Appeared in Bonissone, P. P., M. Henrion, L. N. Kanal, and J. F. Lemmer, eds., Uncertainty in Artificial Intelligence, 6(1991), 385–400, North-Holland, Amsterdam. No. 223. “A New Method for Representing and Solving Bayesian Decision Problems,” Prakash P. Shenoy, September 1990, revised February 1991. To appear in 1991 in Hand, D. J., ed., Artificial Intelligence and Statistics III, Chapman & Hall, London, England. No. 225. “Why Should Statisticians be Interested in Artificial Intelligence?” Glenn Shafer, November, 1990. Appeared in Proceedings of the Fifth Annual Conference on Making Statistics More Effective in Schools of Business, 1990, 16–58, Lawrence, KS. No. 226. “Valuation-Based Systems: A Framework for Managing Uncertainty in Expert Systems,” Prakash P. Shenoy, March, 1991. Revised July 1991. To appear in Fuzzy Logic for the Management of Uncertainty, Zadeh, L. A. and J. Kacprzyk, eds. (1991), John Wiley and Sons, New York, NY. No. 227. “Valuation Networks, Decision Trees, and Influence Diagrams: A Comparison,” Prakash P. Shenoy, June 1991. Revised October 1991. No. 228. “The Early Development of Mathematical Probability,” Glenn Shafer, August 1991. To appear in Grattan-Guiness, I., ed., Encyclopedia of the History and Philosophy of the Mathematical Sciences, Routledge, London. No. 229. “What is Probability?,” Glenn Shafer, August 1991. To appear in Hoaglin, D. C. and D. S. Moore, eds., Perspectives on Contemporary Statistics, Mathematical Association of America. No. 230. “Can the Various Meanings of Probability be Reconciled?,” Glenn Shafer, August 1991. To appear in Keren, G. and C. Lewis, eds., Methodological and Quantitative Issues in the Analysis of Psychological Data, Lawrence Erlbaum, Hillsdale, NJ. No. 231. “Response to the Discussion of Belief Functions,” Glenn Shafer, August 1991. To appear in International Journal of Approximate Reasoning in 1992. No. 232. “An Axiomatic Study of Computation in Hypertrees,” Glenn Shafer, August 1991. No. 233. “Using Possibility Theory in Expert Systems,” Prakash P. Shenoy, September 1991. To appear in 1992 in Fuzzy Sets and Systems. No. 234. “Using Dempster-Shafer’s Belief-Function Theory in Expert Systems,” Prakash P. Shenoy, September 1991. To appear in 1992 in the Proceedings of the SPIE Conference on Applications of Artificial Intelligence X, Orlando, FL. No. 235. “Potential Influence Diagrams,” Pierre Ndilikilikesha, September 1991. No. 236. “Independence in Valuation-Based Systems,” Prakash P. Shenoy, September 1991. Revised November 1991. 
work_73xkppnc7nbolh3mvrfkoljsym	----	Mimicry embedding for advanced neural network training of 3D biomedical micrographs Mimicry embedding for advanced neural network training of 3D biomedical micrographs Artur Yakimovich a, �, Moona Huttunen a, Jerzy Samolej a, Barbara Clough b, Nagisa Yoshida b, c, d, Serge Mostowy c, d, Eva Frickel b, and Jason Mercer a, � a MRC-Laboratory for Molecular Cell Biology, University College London, Gower St, Kings Cross, London WC1E 6B, United Kingdom b Host-Toxoplasma Interaction Laboratory, The Francis Crick Institute, 1 Midland Rd, London NW1 1ST, United Kingdom c Department of Infection Biology, London School of Hygiene & Tropical Medicine, Keppel Street, London WC1E 7HT, United Kingdom d Section of Microbiology, MRC Centre for Molecular Bacteriology and Infection, Imperial College London, London SW7 2AZ, United Kingdom The use of deep neural networks (DNNs) for analysis of com- plex biomedical images shows great promise but is hampered by a lack of large verified datasets for rapid network evolu- tion. Here we present a novel “mimicry embedding” strategy for rapid application of neural network architecture-based analysis of biomedical imaging datasets. Embedding of a novel biolog- ical dataset, such that it mimics a verified dataset, enables ef- ficient deep learning and seamless architecture switching. We apply this strategy across various microbiological phenotypes; from super-resolved viruses to in vivo parasitic infections. We demonstrate that mimicry embedding enables efficient and ac- curate analysis of three-dimensional microscopy datasets. The results suggest that transfer learning from pre-trained network data may be a powerful general strategy for analysis of hetero- geneous biomedical imaging datasets. Deep learning | capsule networks | transfer learning | super-resolution mi- croscopy | vaccinia virus | Toxoplasma gondii | zebrafish Correspondence: jason.mercer@ucl.ac.uk, artur.yakimovich@ucl.ac.uk Introduction Artificial Neural Networks (ANN) excel at a plethora of pat- tern recognition tasks ranging from natural language pro- cessing (1) and facial recognition (2) to self-driving vehicles (3, 4). In biology, recent advances in machine learning and Deep Learning (5–7) are revolutionizing genome sequenc- ing alignment (8), chemical synthesis (9, 10) and biomedi- cal image analysis (11–13). In the field of computer vision, convolutional neural networks (CNNs) perform object detec- tion and image classification at a level matching or surpassing human analysts (14). Despite this, CNN-based architectures often poorly recognise unseen or transformed (e.g. rotated) data due to the use of max or average pooling (15). While pooling allows CNNs to generalize heterogenous data, po- sitional information is ignored. This leads to prioritization of smaller image features and results in an inability of the network to “see the big picture”. To circumvent this, dy- namically routed capsule-based architectures have been pro- posed (15, 16). These architectures are nested allowing for the retention of image feature positional information, and op- timization of CNN performance on images with a larger field of view. However, these architectures remain data-hungry and of- ten perform poorly on small biomedical datasets of high complexity (17). One major reason for this is the lack of large, balanced well-verified biological datasets (18), akin to MNIST (19) and ImageNet (20) that allow for rapid al- gorithm evolution. To circumvent this, ANN analysis of biomedical images can be aided through transfer learning (21, 22). For this, weights of a network trained on one dataset are transferred onto a fully or partially identical untrained network which is then trained on a biomedical dataset of a similar nature (22). This approach shortens training time and is generally considered to be more efficient than random weights initialization strategies (21, 22). Here, we describe a novel data embedding strategy we term, ‘mimicry embedding’ that allows researchers to circumvent the need for verified biomedical databases to perform ANN analysis. Mimicry embedding involves transforming biomed- ical datasets such that they mimic verified non-biomedical datasets thereby allowing for mimicry weights transfer from the latter. By embedding 3D, fluorescent image-based vac- cinia virus and Toxoplasma gondii host-pathogen interaction datasets to mimic grey-scale handwritten digits, we demon- strate that mimicry weights transfer from MNIST (19) allows one to harness the performance of cutting-edge ANN archi- tectures (CapsNet) for the analysis of biomedical data. Fur- thermore, the high accuracy of the embedded datasets may allow for their use as novel verified biomedical databases. Results More often than not host-pathogen biomedical datasets are not large enough for deep learning. However, we rea- soned that advances in high-content fluorescence imaging (23) which allow for 3-D, multi-position single-pathogen res- olution can serve to increase the size of datasets for ANN analysis (13). To classify single-pathogen data in 3D biomed- ical images we developed ‘ZedMate’, an ImageJ-Fiji (24) plugin that uses the Laplacian of Gaussian spot detection engine of TrackMate (25). We challenged ZedMate with multi-channel, 3D fluorescent images of late timepoint vac- cinia virus (VACV) infected cells (Fig. 1a and Fig. 1). Owing to its large size, well-defined structure and multi- ple layers of resident proteins that distinguish different virus forms, VACV has the features needed for complex fluores- Yakimovich et al. | bioRχiv | October 28, 2019 | 1–12 certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 cence microscopy-based biomedical particle analysis (Fig. 1b). By detecting and linking individual virions within an image across the Z-dimension, ZedMate transforms a series of 2D images into a 3D dataset (Fig. 1b). From the original four-fluorescent channel composite Zed- Mate generates grayscale images that preserve the intensity distribution across the z-dimension of each detected channel (Fig. 1c, upper). From this, fluorescence intensity matrices of each channel per Z-plane are then generated for individ- ual particles (Fig. 1c; lower). Using these matrices and ac- counting for the 3D positional information of the detected particles, ZedMate reconstructions can be plotted (Fig. 1d). Intensity analysis across all channels allows for binning of virions into three categories consistent with their biological readouts (Fig. 1b). Initial reconstructions indicated that ZedMate cannot distin- guish between incoming cell-free virions and newly repli- cated cell-associated virions based solely on c1 and c2 in- tensities (Fig. 1d and S1). To improve the precision of ZedMate-based binning we devised a binary ML/DL strat- egy relying on manual annotation to separate cell-free from cell-associated virions. To maintain the spatial information acquired in ZedMate we attempted to train the capsule ANN (CapsNet) (15) on this annotated dataset. These initial at- tempts failed likely due to the small size and complexity of the dataset, two things CapsNet struggles with (17). To circumvent these issues we decided to harness the state- of-the-art performance of CapsNet on the relatively simple grayscale dataset, MNIST (15). To generate weights match- ing our binary classification problem, the handwritten digits in MNIST were separated into two classes: <5 and ≤5. With no changes to CapsNet other than restricting its output to two classes, this network converged with 99.6% accuracy (Fig. 2a). To allow for transfer learning from this network to our biomedical dataset we designed a vector embedding strat- egy we term “mimicry embedding”. For this, the tensors of each virion’s multi-channel, fluorescence Z-profiles from ZedMate are assembled across the X-axis. This is followed by linear interpolation and padding which serve to centre the virion in a 28x28 pixel image such that the resulting data mimics the grayscale MNIST dataset (Fig. 2b). With this approach we aimed to preserve the weights of early Cap- sNet layers by maintaining the binary MNIST CapsNet ar- chitecture and performing weights transfer. Training on our mimicry-embedded real-world dataset achieved 96.5% accu- racy (96.2% precision, 96.2% recall) at separating cell-free from cell-associated virions (Fig. 2b and 1a-d for classifier training). The CapsNet generator was used to visualize how the trained ANN distinguished between cell-free and cell-associated virions with such accuracy. The reconstructions indicated that cell-free virions were elongated with moderate intensity profiles while cell-associated virions were compact and very bright (Fig. 2c). The reconstructions were in agreement with mimicry embedded virions suggesting that these properties yielded the base for the high classification accuracy (Fig. 2d). Fig. 1. ZedMate facilitates detection and classification of VACV particles in infected cells. (see also Fig. 1). a, Merged four channel fluorescent image of a HeLa cell infected with VACV (see Fig. 1a for channel details). Scale bar; 10 µm. b, Illustration of Laplacian of Gaussian (LoG)-based VACV particle detection in 3D. Dumbbell shape (red) represents a particle sliced in optical Z-sections (semi- transparent grey) providing point signal for LoG detection (yellow) and connected in Z (not to scale). c, Intensity measurement from detected particles represented as a Z-profile intensity matrix. d, 3D plot of detected particles color-coded according to detected channels and virion category (see Fig. 1b for details). Quantification of different particle types is inset. N=30 cells, error bars; + SEM. 2 | bioRχiv Yakimovich et al. | Mimicry embedding certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 Fig. 2. Mimicry embedding allows for separation of cell-free and cell-associated VACV particles through weights transfer from a binary MNIST dataset. a, CapsNet architecture for training on the MNIST hand-written digits dataset repurposed into a binary classification problem (<5 or ≤5) prior to CapsNet weights transfer. b, Mimicry embedding of VACV Z-profiles detected by ZedMate. The intensity matrix of fluorescence signal (see Fig. 1) was embedded to mimic MNIST data using linear interpolation and padding Scale bar; 1 µm. CapsNet architecture - with pre-trained weights from a – for training on mimicry embedded VACV particles. c, Reconstructed particle profiles of the virions separated as cell-free and cell-associated by CapsNet. d, Representative mimicry embedded VACV particles for comparison to c. Yakimovich et al. | Mimicry embedding bioRχiv | 3 certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 To verify this strategy, we performed inference on an un- seen (separate from training and validation sets) experimen- tal dataset. Fig. 3 shows the workflow from an input four- channel image ( Fig. 3a and S1a,b), to detection and binning of virions ( Fig. 3b), followed by mimicry embedding and CapsNet separation of cell-free versus cell-associated virions ( Fig. 3c). The results indicate that our model allows for ac- curate classification of virions into four biologically relevant classes within unseen datasets ( Fig. 3d). We’ve established that mimicry embedding and weights transfer allows us to distinguish between incoming cell-free and newly assembled cell-associated virions at late time- points after infection. Next, we asked if this approach could also be used to classify extracellular versus intracellular viri- ons during virus entry, a single-cell assay that often requires specific antibodies or labelling strategies and labour-intensive manual annotation. Considering these common limitations, we generated a training dataset that would allow for gener- alization of this approach. Early infected cells, virions seen in c1, were stained with common fluorescent DNA (c2) and actin (c3) dyes. To circumvent hand-labelling of the train- ing data, immunolabelling to distinguish between intra- and extracellular virus (c4) was used as a weak labelling (26) strategy (Fig. 4a and S3a). After ZedMate detection and transformation of individual particles, intra- and extracellu- lar virus weak labelling (c4) was removed for mimicry em- bedding. By maintaining our binary MNIST CapsNet archi- tecture and performing weights transfer, we could achieve 82% accuracy (81.3% precision, 81.4% recall) in differentiat- ing between intra- and extracellular virions in the absence of specific-antibody labelling and manual annotation (Fig. 1b-e for classifier training). To estimate accuracy, inference was performed on an unseen dataset in which intra- and extracellular virions were quanti- fied using c1-c4 (measured)- inclusive of extracellular virion weak labelling – or only c1-c3 (predicted) (Fig. 4b). A 86% match between measured and predicted quantification of in- tracellular particles was seen (Fig. 4b; inset). This indicates that weak labelling can effectively substitute for manual an- notation of training datasets when classifying intra- and ex- tracellular virion signals. As an additional test of the ANN, we generated a dataset skewed for extracellular virions by blocking virus entry with IPA-3 (27, 28) (Fig. 4c). Consistent with its performance (Fig. 1b-e), a 93% match between mea- sured and predicted quantifications of intracellular particles was seen (Fig. 4d). Finally, when we visualized the recon- structions of intra- and extra- cellular virion classes, extracel- lular virions appeared brighter and more elongated in the Z- direction than intracellular ones (Fig. 4e). This was in agree- ment with their mimicry embedded counterparts (Fig. 4f), explaining the ANNs ability to accurately predict between and quantify these two virion classes. To assess the general applicability of our mimicry embed- ding approach, we acquired a biomedical imaging dataset of cells infected with an EGFP-expressing version of the para- site Toxoplasma gondii (Tg-EGFP). While Tg-EGFP is readily visualized by conventional mi- Fig. 3. Inference demonstrates that mimicry embedding and trained Cap- sNet allows for efficient classification of VACV particles into four biological classes. a, Merged four channel fluorescent image of a HeLa cell infected with VACV previously unseen by CapsNet (see Fig. 1a for channel details). Scale bar; 10 µm. b, Respective ZedMate particle detection and classification by conventional binning of fluorescence intensities. c, Respective inference of cell-free and cell- associated particles detected by ZedMate, mimicry embedded and predicted by a trained CapsNet (see Fig. 2b,c). d, Combined ZedMate particle detection with mimicry embedded and trained CapsNet results in classification of four types of bi- ologically relevant VACV particles. Inset contains quantification of the particle types in the respective image. 4 | bioRχiv Yakimovich et al. | Mimicry embedding certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 croscopy, detecting and quantifying intracellular viability at the single parasite level is challenging (13). To generate a Tg viability training dataset, cells infected with Tg-EGFP (c1) were fixed and stained with fluorescent markers of DNA (c2), and host cell ubiquitin (c3) which was used a weak label to annotate the subset of “unviable” parasites (13, 29) (Fig. 5a). Individual particle detection and transformation in Zed- Mate was followed by mimicry embedding in the absence of c3 weak labelling. After weights transfer from the binary MNIST CapsNet architecture (illustrated in Fig. 2b), and fine tuning on the Tg-EGFP we achieved 70% accuracy (precision 72.5%, recall 70.7%) in the absence of specific viability la- belling (Fig. 1a-d for classifier training). To assure the ANN could accurately distinguish between vi- able and unviable parasites we generated a data set of cells infected with Tg-EGFP using a specific viability label (c3) as ground truth (Fig.5a). To further assess viability, experi- ments were performed in the absence or presence of INFg, which drives parasite killing. Upon model training and val- idation, test inference on this dataset using c1-c2 resulted in a 84% and 80% match between measured (c3) and predicted (c1-c2) viability in the absence or presence of INFg, respec- tively (Fig.5b). CapsNet generator reconstructions showed that “viable” Tg-EGFP appear larger and brighter than “un- viable” parasites in both c1 and c2 (Fig. 5c). This likely ex- plains the ability of the model to accurately predict Tg-EGFP viability in the absence of specific c3 viability labelling. In an attempt to train a general model for in vivo parasite vi- ability assessment using our in vitro dataset, we performed mimicry embedding on Tg-EGFP (Fig. 5a; c2). This resulted in a >10% drop in prediction accuracy when training on Cap- sNet or using Autokeras (30)), a neural architecture search (data not shown). This suggested that single channel mimicry embedding does not provide enough context for training of complex algorithms. However, we reasoned as our mimicry embedding is based on MNIST, we could use any algorithm that performs well on this dataset. By switching to Drop- Connect (31) architecture, which performs among the best on MNIST, our classifier achieved 65% accuracy (precision 65.9%, recall 64.3%) in differentiating between viable and unviable parasites using a single channel (Fig. 1e-h for clas- sifier training). To test this classifier on an in vivo dataset we infected zebrafish (Danio rerio) larvae with Tg-EGFP and imaged them at 0, 6 and 24 h after infection by fluorescent 3D- stereomicroscopy (Fig. 5d). ZedMate was used to detect and quantify Tg-EGFP numbers over time (Fig. 5e). A dra- matic drop-off in parasite count was seen between infection at 0 h and 6 h, followed by increased numbers of Tg-EGFP by 24h. Next the Tg-EGFP Z-profiles were mimicry embed- ded, normalized and their viability inferred using the in vitro infected cell model previously trained on DropConnect. At high pathogen load (0 h) 48% of Tg-EGFP were scored as viable (Fig. 5f). By 6 h this increased to 95% without any significant change within 24 h. These results are consistent with an initial clearing of unviable parasites, and replication of the remaining viable ones (32). Fig. 4. Mimicry embedding can be used for weak-labelling particle classifica- tion. a, Merged four channel fluorescent image of a HeLa cell infected with VACV previously unseen by CapsNet (see Fig. 1a for channel details). b, ZedMate detec- tion and trained CapsNet predicted extracellular and intracellular particles. Quan- tification of intracellular particles is inset. c, Merged four channel image of HeLa cell infected with VACV and treated with the entry inhibitor, IPA 3, previously un- seen by CapsNet. d, ZedMate detection and trained CapsNet of intracellular and extracellular particles. Quantification of intracellular particles is inset. e, Represen- tative reconstruction profiles of extra- and intra- cellular virions. f, Representative mimicry embedded extra- and intra- cellular VACV particles for comparison to e. N=40 untreated and treated cells each. Scale bars a-d;10 µm. Yakimovich et al. | Mimicry embedding bioRχiv | 5 certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 Fig. 5. Mimicry embedding and weight transfer employed for Toxoplasma gondii (Tg) viability detection in cell culture and in vivo.. a, Merged three channel fluorescent image of a HUVEC cells infected Tg EGFP. Individual channels represent DNA stain (c1), Tg EGFP (c2), Ubiquitin (c3). Scale bar; 25 µm. b, Quantification of weak-labelled (measured) and CapsNet inferred (predicted) viable and unviable parasites. c, Representative reconstructions of the trained CapsNet network for viable and non-viable classes of Tg EGFP Z-profiles. d, Representative images (maximum intensity projections) of zebrafish (D. rerio) larvae infected with Tg EGFP at 0, 6 and 24 h pi. Scale bar; 100 µm. e, ZedMate detected Tg counts at 0,6 and 24h pi f, In vivo inference of Tg EGFP viability over time using DropConnect viability model trained on in vitro Tg data. N=10 images, error bars + SEM. Discussion ANN analysis of biomedical datasets has trailed behind the unprecedented advancement of AI analysis seen in other fields. This is largely due to the lack of open source, ver- ified biomedical datasets comparable to MNIST and Ima- geNet (19, 20). Here we present ZedMate and mimicry em- bedding as a strategy to harness the power of datasets like MNIST and transfer learning to train highly accurate mod- els for analysis of 3-D biomedical data. ZedMate, an open source (ImageJ/Fiji) plugin designed for rapid detection and batch-quantification of 3D images at the single spot level made mimicry embedding possible. When used together with CapsNet (15) mimicry embedding proved to be a promising method for detection of complex biomedical phenotypes in vitro. We show that transforming real-world images such that they resemble landmark datasets assures compatibility with, and seamless switching between, cutting-edge architectures. Embedding data in such a way allows one to maintain full compatibility with weights of the first layers thereby improving transfer. Using in vivo biomedical data, we further demonstrate that mimicry em- bedding can yield a model with higher accuracy than one ob- tained through cutting-edge neural architecture search. Thus, mimicry embedding can serve as a common denominator for assessing performance between architectures. Collectively, our results suggest that ZedMate and mimicry embedding, although employed here for the analysis of host-pathogen in- teraction, can be used for AI analysis of any biomedical 3-D dataset. 6 | bioRχiv Yakimovich et al. | Mimicry embedding certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 Materials and Methods Cell culture, antibodies and reagents. HeLa cells (ATCC) were maintained in in Dulbecco’s modified Eagle’s medium (DMEM, Gibco, Life Technologies, Switzerland) with the addition of 10% fetal bovine serum (FBS, Sigma), and 1% penicillin-streptomycin (Pen-Strep, Sigma), 2 mM Gluta- MAX (Life Technologies). Human Umbilical Vein Endothe- lial cells, HUVECs, (C12203, Promocell), were maintained in M199 medium (Gibco) supplemented with 30 mg/mL en- dothelial cell growth supplement (ECGS, 02–102, Upstate), 10 units/mL heparin (H-3149, Sigma) and 20% FBS (Sigma). Cells were cultivated on plates, pre-coated with 1% (w/v) porcine gelatin (G1890, Sigma). Both HUVECs and HeLa were grown as monolayers at 37.0◦C and 5.0% CO2. HU- VEC were not used beyond passage 6. Hoechst 33342 (Sigma) was used post fixation at 1:10,000 dilution throughout. Cell culture grade dimethyl sulfoxide (DMSO), used to dissolve control experimental compounds was obtained from Sigma. VACV and parapoxvirus strains and virus purification. Vac- cinia virus strain Western Reserve expressing A5 mCherry protein (VACV WR) was used throughout (28, 33, 34). VACV mature virions (MVs) were purified from cytoplas- mic lysates by being pelleted through a 36% sucrose cushion for 90 min at 18,000 × g. The virus pellet was resuspended in 10 mM Tris (pH 9.0) and subsequently banded on a 25 to 40% sucrose gradient at 14,000 × g for 45 min. Following centrifugation, the viral band was collected by aspiration and concentrated by pelleting at 14,000 × g for 45 min. MVs were resuspended in 1 mM Tris (pH 9.0), and the titter was deter- mined for PFU per millilitre as previously described (35). Early VACV infection and extracellular virions stain- ing. HeLa cells were seeded onto CELLview Slide (Greiner Bio-One) at 10,000 cells per well 16h before the experiment. VACV A5-mCherry F13-EGFP was added at MOI 20, to in- crease the chances of synchronous infection. Cells were fixed with 4% EM- grade PFA 4 hours post infection (hpi) for 20 min followed by a PBS wash. Staining and labelling was preceded by blocking (without permeabilization) in block- ing buffer (5% BSA, 1% FBS, in PBS) for 60 min at room temperature (RT). Next, L1 mouse (7D11) antibody (36) (1: 1000) in blocking buffer was added for 60 min at RT, fol- lowed by a PBS wash. Anti-mouse antibody (Alexa 647, Invitrogen. 1:1000), Phalloidin 594 (Sigma, 1:1000) and Hoechst in blocking buffer were added for 60 min at RT, fol- lowed by a PBS wash. 1,1’-Disulfanediyldinaphthalen-2-ol VACV entry inhibitor (IPA-3) was obtained and used as de- scribed (33). DMSO concentration was equal to or below 1%. Late VACV infection and staining. HeLa cells we cultured on the coverslips and infected with VACV WR expressing A5 mCherry protein. At 8 hpi cells were fixed with 4% v/v FA. Next, VACV B5 protein antibody (mouse, 1:1000) in block- ing buffer was added for 60 min at RT, followed by a PBS wash. Anti-mouse antibody (Alexa 647), Hoechst in block- ing buffer were added for 60 min at RT, followed by a PBS wash. Toxoplasma gondii (Tg) cultivation learning infection phenotypes. Toxoplasma (RH type I and Prugniaud type II strains) expressing GFP/luciferase were maintained in vitro by serial passage on human foreskin fibroblasts (HFFs) cul- tures (ATCC). Cultures were grown in DMEM high glucose (Life Technologies) supplemented with 10% FBS (Life Tech- nologies) at 37◦C in 5% CO2. Tg cultured cells infection and staining. The day be- fore the infection, type II parasites were passaged onto new HFFs to obtain parasites with a high viability. Tg were pre- pared from freshly 25G syringe-lysed HFF cultures. Para- sites were subsequently 2 x 27G syringe lysed and excess HFF cell debris removed by centrifugation. Then, the par- asites were added to the experimental cells at an MOI=2. The cell cultures with added Tg were then centrifuged at 500 x g for 5 min to synchronize the infection and the cultures incubated at 37◦C in 5% CO2 for 3h. Samples treated with interferon gamma (IFNγ) were subjected to 100 IU/mL human IFNγ (285-IF, RD Systems) for 18h prior to infection. Upon fixation cells were stained with Hoechst 33342 and mouse mAb anti-ubiquitin FK2 (PW8810, Enzo Lifesciences; RRID: AB10541840) and Alexa Fluor 568- conjugated secondary goat anti-mouse (A-11004, Invitrogen; RRID:AB141371). Tg infection in vivo. Tg EGFP parasites (type 1) were pre- pared from freshly 25G syringe-lysed HFF cultures in 10% FBS. Parasites were subsequently 27G syringe-lysed and ex- cess HFF material removed by centrifugation. After wash- ing with PBS, Toxoplasma tachyzoites were resuspended at 2.0x106 tachyzoites/µl in PBS. Larvae were anesthetized with 20 µg/ml tricaine (Sigma- Aldrich) during the injection procedures and for all live in vivo imaging. All experiments were carried out on TraNac background larvae to minimize obstruction of fluorescence signal by pigmentation. 3dpf larvae were anesthetized and injected with 2.5 nl of parasite suspension into the hindbrain ventricle (HBV) as previously described (37). Infected larvae were transferred into individual wells containing 0.5x E2 me- dia supplemented with methylene blue pre-warmed to 33◦C. Zebrafish husbandry and maintenance. Fish were main- tained at 28.5◦C on a 14hr light, 10hr dark cycle. Embryos obtained by natural spawning were maintained in 05x E2 me- dia supplemented with 0.3 µg/ml methylene blue. Ethics statement. Animal experiments were performed ac- cording to the Animals (Scientific Procedures) Act 1986 and approved by the Home Office (Project licenses: PPL P84A89400 and P4E664E3C). All experiments were con- ducted up to 4 days post fertilisation. Yakimovich et al. | Mimicry embedding bioRχiv | 7 certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 Super-resolution imaging of VACV intracellular viri- ons. Supper-resolution microscopy was performed using a 100x oil immersion objective (NA 1.45) on a VT-iSIM mi- croscope (Visitech; Nikon Eclipse TI), using 405 nm, 488 nm, 561 nm, 647 nm laser frequencies for excitation. High-Content Tg EGFP imaging in cells. Black plastic flat-bottom 96-well plates (Falcon 353219) were imaged on an Opera Phenix High Content Imaging Platform using 63x magnification, 8 Z-slices (0.5 µm/slice) and multiple fields of view per well. Images were as single channel 16-bit tiff files and further processed for ZedMate analysis. 3D Tg EGFP imaging in vivo. Progress of the in vivo infec- tion was monitored by fluorescent stereomicroscopy (Leica M205FA, Leica Microsystems, Nussloch GmbH, Nussloch, Germany) at regular time points. All images were obtained with a 10x objective, at 13x magnification (0.79 µm/px) 20 z planes were captured covering a total distance of 171µm (8.55µm intervals). Data processing and deep neural network training. Our training hardware was based on a single Nvidia 1080 Ti GPU set up in Intel Core i7 8700K sys- tem equipped with 32 Gb of RAM and an SSD. In- stallation consisted of Anaconda Python, Keras-gpu 2.2, Tensorflow-gpu 1.10 and KNIME 3.7.1. Some models were trained on 2019 MacBook Pro equipped with Intel Core i5 CPU using Keras 2.2 CPU. Source code is avail- able under https://github.com/ayakimovich/ ZedMate, example dataset under https://github. com/ayakimovich/virus-mnist. Further materials are available upon request. References 1. Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th international conference on Machine learning, pages 160–167. ACM. ISBN 1605582050. 2. Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815–823. 3. Jinkyu Kim and John Canny. Interpretable learning for self-driving cars by visualizing causal attention. In Proceedings of the IEEE international conference on computer vision, pages 2942–2950. 4. Sebastian Ramos, Stefan Gehrig, Peter Pinggera, Uwe Franke, and Carsten Rother. Detect- ing unexpected obstacles for self-driving cars: Fusing deep learning and geometric mod- eling. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 1025–1032. IEEE. ISBN 1509048049. 5. Yann LeCun and Yoshua Bengio. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. 6. Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. ISSN 0018-9219. 7. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436, 2015. ISSN 1476-4687. 8. Christof Angermueller, Tanel Pärnamaa, Leopold Parts, and Oliver Stegle. Deep learning for computational biology. Molecular systems biology, 12(7):878, 2016. ISSN 1744-4292. 9. Jennifer N Wei, David Duvenaud, and Alán Aspuru-Guzik. Neural networks for the prediction of organic chemistry reactions. ACS central science, 2(10):725–732, 2016. ISSN 2374- 7943. 10. Marwin HS Segler and Mark P Waller. Neuralâsymbolic machine learning for retrosynthesis and reaction prediction. Chemistry–A European Journal, 23(25):5966–5971, 2017. ISSN 0947-6539. 11. Juan C Caicedo, Sam Cooper, Florian Heigwer, Scott Warchal, Peng Qiu, Csaba Molnar, Aliaksei S Vasilevich, Joseph D Barry, Harmanjit Singh Bansal, and Oren Kraus. Data- analysis strategies for image-based cell profiling. Nature methods, 14(9):849, 2017. ISSN 1548-7105. 12. Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, and Siân Culley. Content- aware image restoration: pushing the limits of fluorescence microscopy. Nature methods, 15(12):1090, 2018. ISSN 1548-7105. 13. Daniel Fisch, Artur Yakimovich, Barbara Clough, Joseph Wright, Monique Bunyan, Michael Howell, Jason Mercer, and Eva Frickel. Defining host–pathogen interactions employing an artificial intelligence workflow. eLife, 8:e40560, 2019. ISSN 2050-084X. 14. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105. 15. Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. In Advances in neural information processing systems, pages 3856–3866. 16. Edgar Xi, Selina Bing, and Yang Jin. Capsule network performance on complex data. arXiv preprint arXiv:1712.03480, 2017. 17. Rinat Mukhometzianov and Juan Carrillo. Capsnet comparative performance evaluation for image classification. arXiv preprint arXiv:1805.11195, 2018. 18. Cristian Bartolome Yiguang Zhang and Ashwin Ramaswami. Deepcell: Automating cell nuclei detection with. 19. Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE Signal Processing Magazine, 29(6):141–142, 2012. ISSN 1053- 5888. 20. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 248–255. IEEE. ISBN 1424439922. 21. Jeremy West, Dan Ventura, and Sean Warnick. Spring research presentation: A theoret- ical foundation for inductive transfer. Brigham Young University, College of Physical and Mathematical Sciences, 1, 2007. 22. Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, 22(10):1345–1359, 2010. ISSN 1041-4347. 23. Mojca Mattiazzi Usaj, Erin B Styles, Adrian J Verster, Helena Friesen, Charles Boone, and Brenda J Andrews. High-content screening for quantitative cell biology. Trends in cell biol- ogy, 26(8):598–611, 2016. ISSN 0962-8924. 24. J. Schindelin, I. Arganda-Carreras, E. Frise, V. Kaynig, M. Longair, T. Pietzsch, S. Preibisch, C. Rueden, S. Saalfeld, B. Schmid, J. Y. Tinevez, D. J. White, V. Hartenstein, K. Eliceiri, P. Tomancak, and A. Cardona. Fiji: an open-source platform for biological-image analysis. Nat Methods, 9(7):676–82, 2012. ISSN 1548-7091. doi: 10.1038/nmeth.2019. 25. Jean-Yves Tinevez, Nick Perry, Johannes Schindelin, Genevieve M Hoopes, Gregory D Reynolds, Emmanuel Laplantine, Sebastian Y Bednarek, Spencer L Shorte, and Kevin W Eliceiri. Trackmate: An open and extensible platform for single-particle tracking. Methods, 115:80–90, 2017. ISSN 1046-2023. 26. Alexander J Ratner, Christopher M De Sa, Sen Wu, Daniel Selsam, and Christopher Ré. Data programming: Creating large training sets, quickly. In Advances in neural information processing systems, pages 3567–3575. 27. Sean W Deacon, Alexander Beeser, Jami A Fukui, Ulrike EE Rennefahrt, Cynthia Myers, Jonathan Chernoff, and Jeffrey R Peterson. An isoform-selective, small-molecule inhibitor targets the autoregulatory mechanism of p21-activated kinase. Chemistry biology, 15(4): 322–331, 2008. ISSN 1074-5521. 28. J. Mercer and A. Helenius. Vaccinia virus uses macropinocytosis and apoptotic mimicry to enter host cells. Science, 320(5875):531–5, 2008. ISSN 1095-9203 (Electronic) 0036-8075 (Linking). 29. Barbara Clough, Joseph D Wright, Pedro M Pereira, Elizabeth M Hirst, Ashleigh C Johnston, Ricardo Henriques, and Eva-Maria Frickel. K63-linked ubiquitination targets toxoplasma gondii for endo-lysosomal destruction in ifnγ-stimulated human cells. PLoS pathogens, 12 (11):e1006027, 2016. ISSN 1553-7374. 30. Haifeng Jin, Qingquan Song, and Xia Hu. Auto-keras: Efficient neural architecture search with network morphism. arXiv preprint arXiv:1806.10282, 2018. 31. Li Wan, Matthew Zeiler, Sixin Zhang, Yann LeCun, and Rob Fergus. Regularization of neural networks using dropconnect. icml’13, pp. Technical report, III–1058–III–1066. JMLR. org, 2013. 32. Nagisa Yoshida, Marie-Charlotte Domart, Artur Yakimovich, Maria J. Mazon-Moya, Lucy Collinson, Jason Mercer, Eva-Maria Frickel, and Serge Mostowy. In vivo control of toxo- plasma gondii by zebrafish macrophages. bioRxiv, 2019. 33. J. Mercer, S. Knebel, F. I. Schmidt, J. Crouse, C. Burkard, and A. Helenius. Vaccinia virus strains use distinct forms of macropinocytosis for host-cell entry. Proc Natl Acad Sci U S A, 107(20):9346–51, 2010. ISSN 1091-6490 (Electronic) 0027-8424 (Linking). 34. Samuel Kilcher and Jason Mercer. Dna virus uncoating. Virology, 479:578–590, 2015. ISSN 0042-6822. 35. Jason Mercer and Paula Traktman. Investigation of structural and functional motifs within the vaccinia virus a14 phosphoprotein, an essential component of the virion membrane. Journal of virology, 77(16):8857–8871, 2003. ISSN 0022-538X. 36. Elizabeth J Wolffe, S Vijaya, and Bernard Moss. A myristylated membrane protein encoded by the vaccinia virus l1r open reading frame is the target of potent neutralizing monoclonal antibodies. Virology, 211(1):53–63, 1995. ISSN 0042-6822. 37. Maria J Mazon-Moya, Alexandra R Willis, Vincenzo Torraca, Laurent Boucontet, Avinash R Shenoy, Emma Colucci-Guyon, and Serge Mostowy. Septins restrict inflammation and pro- tect zebrafish larvae from shigella infection. PLoS pathogens, 13(6):e1006467, 2017. ISSN 1553-7374. 8 | bioRχiv Yakimovich et al. | Mimicry embedding certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://github.com/ayakimovich/ZedMate https://github.com/ayakimovich/ZedMate https://github.com/ayakimovich/virus-mnist https://github.com/ayakimovich/virus-mnist https://doi.org/10.1101/820076 Appendix 1. Supplementary Information This is a supplementary section to the preprint manuscript by Yakimovich et al. Source code is available under https: //github.com/ayakimovich/ZedMate. All further information is available upon request. Supplementary Figures. Figure S1. Individual channels used in late VACV infected cells and their biological relevance). a, Maximum intensity projections of individual channel images of HeLa cells infected with VACV at 8 hours post infection. Here, DNA stain (c1), VACV core A5-mCherry (c2), VACV outer envelope protein F13 (c3) and VACV outer envelope protein B5 (c4). Scale bar; 10 µm. b, Illustration of the position of markers in virions [MV (mature virions), IEV (intracellular enveloped virions), CEV (cell-associated extracellular virions)], and these virions - with the corresponding markers - in infected cells. Here c1 marks cellular DNA and cytoplasmic VACV replication sites, c2 marks all virions (MVs, IEVs and CEVs), c3 marks a subset of virions (IEVs and CEVs) and c4 marks only CEVs. Yakimovich et al. | Mimicry embedding bioRχiv | 9 certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://github.com/ayakimovich/ZedMate https://github.com/ayakimovich/ZedMate https://doi.org/10.1101/820076 Figure S2. CapsNet training and validation of cell-free vs. cell-associated virus model. a, Late model loss function change upon training iterations (epochs). b, Late model training and validation (unseen data) accuracy change upon training iterations (epochs). c, Late model receiver operational characteristics (ROC) curve of the trained model obtained using unseen data (validation). Here area under the curve (AUC) was 0.989. d, Late model confusion matrix of the trained model obtained using unseen data (validation). Late model precision was 96.2%, recall was 96.2%, F1-score was 96.2%. 10 | bioRχiv Yakimovich et al. | Mimicry embedding certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 Figure S3. CapsNet training and validation of extracellular vs. intracellular virus model. a, Maximum intensity pro- jections of individual channels from HeLa cells infected with VACV. Here, DNA stain (c1), VACV core A5-EGFP (c2), actin stained with phalloidin (c3) and VACV membrane protein L1 as an extracellular virion label (c4) b, Model loss function change upon training iterations (epochs). c, Model training and validation (unseen data) accuracy change upon training iterations (epochs). d, Model receiver operational characteristics (ROC) curve of the trained model obtained using unseen data (valida- tion). Here area under the curve (AUC) was 0.896. e, Model confusion matrix of the trained model obtained using unseen data (validation). Model precision was 81.3%, recall was 81.4%, F1-score was 81.8%. Yakimovich et al. | Mimicry embedding bioRχiv | 11 certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 Figure S4. CapsNet and DropConnect training and validation of in vitro and in vivo Tg EGFP viability model. a, 2-channel CapsNet model loss function change upon training iterations (epochs). b, 2-channel CapsNet model training and validation (unseen data) accuracy change upon training iterations (epochs). c, 2-channel CapsNet model receiver operational characteristics (ROC) curve of the trained model obtained using unseen data (validation). Here area under the curve (AUC) was 0.764. d, 2-channel CapsNet model confusion matrix of the trained model obtained using unseen data (validation). e, 1- channel DropConnect model loss function change upon training iterations (epochs). f, 1-channel DropConnect model training and validation (unseen data) accuracy change upon training iterations (epochs). g, 1-channel DropConnect model receiver operational characteristics (ROC) curve of the trained model obtained using unseen data (validation). Here area under the curve (AUC) was 0.685. g, 1-channel DropConnect model confusion matrix of the trained model obtained using unseen data (validation). The 2-channel CapsNet model precision was 72.5%, recall was 70.7%, F1-score was 70.1%. The 1-channel DropConnect model precision was 65.9%, recall was 64.3%, F1-score was 64.9%. 12 | bioRχiv Yakimovich et al. | Mimicry embedding certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was notthis version posted October 29, 2019. ; https://doi.org/10.1101/820076doi: bioRxiv preprint https://doi.org/10.1101/820076 Supplementary Information 
work_746woqng7fdohiankzdyij7xsu	----	Title of the Paper: Example Paper for Business Systems Research Journal econstor Make Your Publications Visible. A Service of zbw Leibniz-Informationszentrum Wirtschaft Leibniz Information Centre for Economics Pejanović, Radovan et al. Conference Paper Marketing Decision Making by Applying the Expert System Provided in Cooperation with: IRENET - Society for Advancing Innovation and Research in Economy, Zagreb Suggested Citation: Pejanović, Radovan et al. (2016) : Marketing Decision Making by Applying the Expert System, In: Proceedings of the ENTRENOVA - ENTerprise REsearch InNOVAtion Conference, Rovinj, Croatia, 8-9 September 2016, IRENET - Society for Advancing Innovation and Research in Economy, Zagreb, Vol. 2, pp. 280-286 This Version is available at: http://hdl.handle.net/10419/183727 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc/4.0/ www.econstor.eu 280 ENTRENOVA 8-9, September 2016 Rovinj, Croatia Marketing Decision Making by Applying the Expert System Radovan Pejanović University of Novi Sad, Faculty of Agriculture, Serbia Otilija Sedlak University of Novi Sad, Faculty of Economics, Serbia Zoran Ćirić University of Novi Sad, Faculty of Economics, Serbia Leposava Grubić-Nešić University of Novi Sad, Faculty of Technical Sciences, Serbia Slavica Mitrović University of Novi Sad, Faculty of Technical Sciences, Serbia Jelica Eremić-Đođić Elektroprivreda Srbije, Serbia Abstract The main goal of this paper is to develop an expert system based on fuzzy set theory that will provide more successful and efficient decision making in the area of marketing, in relation with dilemma “to produce or to purchase”. Namely, authors will try to develop a model that will be suitable for making marketing decisions in production systems. Methodology in the paper obtained analysis of the theory of marketing, the development of the specific model for decision making, so as the application of developed model on one case from production system. Methodology which allows to model indeterminacy is fuzzy sets theory which is particularly well designed for dealing with non-probabilistic uncertainties. Authors will develop a model for decision making, based on successful integration of marketing and fuzzy theories. They will implement the model in decision making problem related to the debate “to produce or to purchase” on one real decision problem in production system. The main goal of this paper is to make changes in the work of decision makers in marketing sector. Authors pointed the advantages of the model with quality management, but also some limitations and possibilities for the future researches. Keywords: fuzzy logic, decision support systems, ICT, uncertainty JEL classification: C63 Introduction The marketing management comprises all those activities associated with the transformation and flow of goods and services, including their attendant information flows, from the sources of materials to end users. Management refers to integration of all these activities, both internal and external to the firm. The main goal of this paper was to make changes in the work of decision makers in marketing sector. The authors expanded and changed the decision making system, by developing and installing a new model for decision making. In this paper we had a new model with several inputs and one output. Input elements on the lower level become fuzzy. The outcome of the model is a decision, which is 281 ENTRENOVA 8-9, September 2016 Rovinj, Croatia influenced by a considerable number of factors and the soundness of assumptions. This model could simulate sustainable functioning of the production process. Marketing decision makers include marketing managers, analysts, and also formal or informal authorities featuring in the process of choosing marketing alternatives. They feature both on the supply and demand side, given that a social and managerial process by which individuals and groups obtain what they need and want through creating and exchanging products and values with others (Kosko, 1992). High uncertainty and risk levels result from disruptive innovations and great unexpected shocks (Kotler, 2000). Decision makers often prefer to employ oral presentation rather than numerical value. In such conditions the best solution is to make decisions on the basis of multiple conditions and goals to achieve a relatively desirable level of achievement (Kang al., 2011). These issues have caused the nature of decision making to be full of complexities and ambiguities in the most minor to most major cases. Consequently, most decisions are made in a fuzzy environment. Therefore, considering that the fuzzy logic method is proposed for decision making in uncertain and ambiguous situations, using this method can reduce ambiguities and increase the effectiveness of decisions made (Ertugrul al., 2009). Vorhies and Morgan in their paper recognized that the key mechanism for identifying, building, and enhancing of production process capabilities to deliver sustainable competitive advantage, through marketing decision models (Vorhies al., 2005). The paper organization follows these steps: literature review, about of marketing decision problem and the importance of marketing management sustainability. It is also about hesitant fuzzy sets. The third and fourth part presents the problem and the model. The authors posed the model and performed an in-depth analysis and construction of constraints’ coefficients. Results are presented and differences between results of similar studies are pointed out. The last part contains conclusions with advantages and limitations of the given model, the guidelines and possibilities for the future research. Problem Formulation Making marketing decisions is aimed in two basic directions, i.e. towards solving problems on the one hand and using the benefits of business opportunities on the other. At any rate, marketing decision making enables a transition from the current state into a new, i.e. desired state, which marketing decision makers deem to be more favorable for the company. Decision is made under conditions of risk, when the only knowledge available regarding the state of outcomes is their probability distribution. The outcome of these decisions is influenced by a considerable number of factors and the soundness of assumptions. Traditionally, cost had been viewed as the main consideration in sourcing decision (Maltz al., 2012). Also, one more factor which is important in the process of sourcing decision making is the level of technology development. For example, on the basis of 1990–2002 panel data set on Spanish companies and an exogenous proxy for technological change it have been provided thee causal evidence that technological change increases the likelihood of outsourcing (Bartel al., 2012). Methodology It is due to the importance of additional criteria that a prototype of the expert system was developed (including software solutions). As the economic model in Figure 1 shows, the number of attributes taken into account is far greater compared to the classical decision making method. 282 ENTRENOVA 8-9, September 2016 Rovinj, Croatia The decision maker communicates with the expert system by choosing the domains of values for the attribute of the decision making leaf nodes. Using the built- in logarithm, the expert system proposes a solution with explanation based on these data. Managing and modeling of uncertainty by different forms of information, used by experts to provide their preferences, can be done in various ways, including utility vectors, fuzzy preference relations, linguistic variables, interval values, multiplicative preference relations, hesitant fuzzy sets. A solution to the given problem by applying fuzzy set theory, the economic model (Figure 1) can be supplemented with an oriented graph. Figure 1 Decision tree for problem solving: “Produce or purchase” DECISION ECONOMIC EFFECTS · PRODUCE · PURCHASE PRODUCTION LIMITATIONS EXTERNAL FACTORS EXPERTISE & WORKMANSHIP QUALITY CAPACITY UTILISATION PRODUCTION COSTS PURCHASE PRICE FORESEEN TECHNOLOGICAL CHANGES SUPPLIERS' APPROACH · ENORMOUS · SUBSTANTIAL · NEGLIGIBLE · NONE · FAVOURABLE · NEUTRAL · UNFAVOURABLE · SUBSTANTIAL · FAVOURABLE · UNFAVOURABLE · DEVASTATING · HIGH · AVERAGE · LOW · HIGH · AVERAGE · LOW · HIGH · AVERAGE · LOW · HIGH · MODERATE · LOW QUANTITY OF PRODUCT PRODUCTION SCHEDULE SUPPLIER RELATIONS TRANSPORT COSTS · CONFINING · TOLERABLE · NEGLIGIBLE · AGGRESSIVE · MODERATE · PASSIVE · LARGE · MEDIUM · SMALL · WEEKLY · MONTNLY · QUARTERLY · SEMI-ANNUAL · EXCELLENT · GOOD · POOR · HIGH · AVERAGE · LOW RAW MATERIAL COSTS LABOUR COSTS LONG TERM CHANGE MID-TERM CHANGE · HIGH · AVERAGE · LOW · HIGH · AVERAGE · LOW · DEVELOPMENT · STATUS QUO · ELIMINATION · DEVELOPMENT · STATUS QUO · ELIMINATION Source: Author’s illustration The Fuzzy Approximation Theorem says more than that. Theoretically, all equations can be translated into rule patches. Fuzzy systems approximate systems in physics, communication, physiology, etc. Fuzzy systems can be applied wherever the brain is used (Ciric al., 2015). The Model Let us observe a company ‘Y’ consisting of S sectors, which are producing a different type of products. Characteristics (arguments, input variables) include expertise and developing workmanship quality, quality of product, production schedule, raw 283 ENTRENOVA 8-9, September 2016 Rovinj, Croatia material costs, labor costs, supplier relations, transport costs, long term change, mid- term change, supplier approach. These are not arguments in the classical mathematical sense. The input variables (arguments) are linguistic variables (“high raw material costs”, “medium quantity of product”, “aggressive suppliers’ approach”, “status quo in the forthcoming 3-5 years, but development over the subsequent ten plus years”, etc). The same applies to output variables (functions), representing input functions for complex functions up to the last level, such as production costs, purchase price, foreseen technological change, production limitations etc. Such nature of input and output variables, i.e. value domains that cannot be accurately quantified, make the fuzzy set theory extremely convenient for problem solving in marketing decision making. This is how resolving the above described problem was approached. Input and output variables were defined as different fuzzy sets, both continuous and discontinuous. Defining fuzzy sets intrinsically implies that the appropriate degree of belonging is ascribed to all possible values (Sedlak al., 2013). Results The approximate reasoning algorithm for making the business decision whether to produce or purchase consists of three basic steps. Each basic step is divided into several interconnected steps. Each of these substeps, or steps, contains a logical inference rule defined by experts. As specified numeric values directly entered at the given moment of making the “purchase or produce” decision, the input data are as follows: expertise, quantity of product, production schedule, raw material costs, labor costs, supplier relations, transport costs, and long- and mid-term changes. Fuzzy sets are already prepared at all levels. Direct entry of the above listed first- and second-level data activated the above defined algorithm. Production costs are defuzzified (Figure 2 ) by determining the centre of gravity using the known formula. Thus obtained numerical data is the input value for step 2. 284 ENTRENOVA 8-9, September 2016 Rovinj, Croatia Figure 2 Decision making with defuzzification Source: Author’s illustration Conclusion In this article the accent was on presenting a new method in decision making for business decision “purchase or produce”. The problem has all the characteristics of uncertainty. Serious problems, and the limitations, are determining of input variables and forecasting all possible circumstances, which is only possible based on subjective estimate. An estimate with a degree of accuracy is much easier to obtain for a range of data than for an individual value. Certain characteristics of input variables cannot be measured precisely, and an expert’s opinion is needed. Fuzzy sets can be introduced into the existing decision making models in several ways. As an economic institution, a company bases its existence on the environment, both from the aspect of providing input and from the aspect of achieving and valorizing input. Miscellaneous knowledge and experience, and also decision making in the areas of investment, market operations, financial function, production function or research and development, can be considered more fully and exactly applying fuzzy sets. The fuzzy set theory provides a solution to this. 285 ENTRENOVA 8-9, September 2016 Rovinj, Croatia Individual steps of the algorithm are based on fuzzy logic rules. The algorithm was verified on multiple numeric examples, and development of an appropriate software product is in progress. The essential feature of this algorithm is that there is no possibility of “setting up” a desirable decision by tailoring input data. Miscellaneous knowledge and experience, and also decision making in the areas of investment, market operations, financial function, production function or research and development, can be considered more fully and exactly applying hesitant fuzzy sets and these will be fully investigated in our future work. References 1. Bartel, A. P., Lach, S., Sicherman, N. (2012), “Technological change and the make- or-buy decision”, Journal of Law, Economics, and Organization, pp. 1-28. 2. Ćirić, Z., Stojić, D., Sedlak, O. (2015), “Multicriteria HR allocation based on hesitant fuzzy sets and possibilistic programming”, Acta Polytechnica Hungarica, Óbuda University, Hungarian Academy of Engineering and IEEE Hungary Section, Vol. 12 No. 3, pp. 185-197. 3. Coase, R., (1937), “The Nature of the Firm”, Economica, Vol. 4 No. 16, pp. 386-405. 4. Ertugrul, I, & Karakasoglu, N. (2009), “Performance evaluation of Turkish cement firms with fuzzy analytic hierarchy process and TOPSIS methods”, Expert Systems with Applications, Vol. 36 No. 1, pp. 702–715. 5. Kang, H., Hung, M., Pearn, W. L., Lee, A.H. I., M. (2011), “An Integrated Multi-Criteria Decision Making Model for Evaluating Wind Farm Performance”, Energies, Vol. 4, No. 11, pp. 2002-2026. 6. Kosko, B. (1992), “Neural Networks and Fuzzy Systems”, Englewood Cliffs: Prentice – Hall, New Jersey. 7. Kotler, P. (2000), “Marketing management: The millennium edition”, Upper Saddle River, NJ: Prentice-Hall. 8. Maltz, A., Carter, J. R., & Maltz, E. (2011), “How managers make sourcing decisions about low cost regions: Insights from perceptual mapping”, Industrial Marketing Management, Vol 40, No. 5, pp. 796-804. 9. Sedlak, O., Ćirić, Z., Ćirić, I. (2013), “Strategic Management under the Conditions of Uncertainty and Indefiniteness”, Strategic Management, Vol. 18, No. 1, pp. 62-68. 10. Tan, X.C., Wang, Y.Y., Gu, B.H., Mu, Z.K., Yang, C. (2011), “Improved Methods for Production Manufacturing Processes in Environmentally Benign Manufacturing”, Energies, Vol. 4, No. 9, pp. 1391-1409. 11. Vorhies, D.W., Morgan, N.A. (2005), “Benchmarking Marketing Capabilities for Sustainable Competitive Advantage”, Journal of Marketing, Vol. 69, No. 1, pp. 80-94. About the authors Radovan Pejanović is Professor at the Faculty of Agriculture, Department of Agricultural Economics and Rural Sociology, University of Novi Sad. He obtained all titles from assistant to full professor. He teaches at the undergraduate, graduate and doctoral studies at Agricultural Economics studies and studies of Agro-tourism and Rural Development. He has published a number of papers (textbooks, monographs, scientific papers) and is associate (manager and coordinator) in several national and international projects. He is head of the Center for Rural Development, president of the Scientific Society of Agricultural Economists of Balkan and the president of the Editorial Board of the journal Economics of Agriculture as well as a member of the editorial board of the journal Annals of Scientific Work of the Faculty of Agriculture in Novi Sad and Contemporary Agriculture. He is the editor at the publishing house Academic Book from Novi Sad. Prof. Dr. Pejanović is certified accountant and auditor, assessor of capital, as well as an expert in economic and financial issues. The author can be contacted at pejanovic@uns.ac.rs. http://www.mdpi.com/search?authors=He-Yau%20Kang http://www.mdpi.com/search?authors=Meng-Chan%20Hung http://www.mdpi.com/search?authors=W.%20L.%20Pearn http://www.mdpi.com/search?authors=Amy%20H.%20I.%20Lee http://www.mdpi.com/1996-1073/4/11/2002 http://www.mdpi.com/1996-1073/4/11/2002 http://www.mdpi.com/search?authors=Xian-Chun%20Tan http://www.mdpi.com/search?authors=Yan-Yan%20Wang http://www.mdpi.com/search?authors=Bai-He%20Gu http://www.mdpi.com/search?authors=Ze-Kun%20Mu http://www.mdpi.com/1996-1073/4/9/1391 http://www.mdpi.com/1996-1073/4/9/1391 286 ENTRENOVA 8-9, September 2016 Rovinj, Croatia Otilija Sedlak is Professor at the University of Novi Sad, Faculty of Economics Subotica, working within the Department of Business Informatics and Quantitative Methods. She is involved in the realization of the teaching process on basic studies, master studies as well as on PhD studies in subjects close to her scientific interests. She is an author and co-author of several textbooks, scripts, notes and compendiums. Published or co-published over 30 scientific and professional papers in both national and international journals. She participated in the work on more than 80 scientific meetings that were published, either in whole or partly, in monographs or journals. The author can be contacted at otilijas@ef.uns.ac.rs. Zoran Ćirić is Professor at the University of Novi Sad, Faculty of Economics Subotica, working within the Department of Business Informatics and Quantitative Methods. He is involved in the realization of the teaching process on basic studies, master studies as well as on PhD studies in subjects close to her scientific interests. He is an author and co-author of several textbooks, scripts, notes and compendiums. Published or co-published over 20 scientific and professional papers in both national and international journals. He participated in the work on more than 50 scientific conferences that were published, either in whole or partly, in monographs or journals. The author can be contacted at zoran.ciric@ef.uns.ac.rs. Full Professor Leposava Grubić-Nešić, PhD, Department of Industrial Engineering and Management, Faculty of Technical Sciences, University of Novi Sad, Trg Dositeja Obradovica 6, 21000 Novi Sad, Serbia. Research interests: Human Resources Management, Leadership, Work Motivation, Work Psychology, etc. She is the author of the book Human Resources Management and of over 100 scientific research papers published in national and international journals and conferences Participates in several international and national projects; particularly engaged in cooperation with the industry. The author can be contacted at nesle@uns.ac.rs. Assis. Prof. Slavica Mitrovic, PhD, Department of Industrial Engineering and Management, Faculty of Technical Sciences, University of Novi Sad, Trg Dositeja Obradovica 6, 21000 Novi Sad, Serbia. Research interests: production systems, organization and management, decision making, management principles in organization, entrepreneurship. She is the author of the book Principles of modern management and of over 70 scientific research papers published in national and international journals and conferences. Participates in several international and national projects; particularly engaged in cooperation with the industry. The author can be contacted at mslavica@uns.ac.rs. Jelica Eremić-Đođić, PhD, is financial director of ODS (Distribution system operator "EPS Distribution"), EPS Distribucija, Masarikova 1-3, Beograd, Serbia. Research interests: distribution systems, financial management, business informatics, monitoring. She participates in several national projects; particularly engaged in cooperation with SAP. She is the author of about 20 scientific research papers published in national and international journals and conferences. The author can be contacted at jelica.eremic@ev.rs. mailto:otilijas@ef.uns.ac.rs 
work_75upci22jffbth3moq6xa6xhre	----	122 JIT 3,2 June 1988 of several different hosts looking at nominal prices of databases, processing times and 'other cost-effective- ness considerations'. The second section 'Counting the Costs' covers var- ious aspects of this side of the cost-benefit equation. Again, the editor's introduction is easy to understand and well referenced. The other side of the cost-benefit equation is covered in the third section of the book: 'Assessing the Benefits'. This, incidentally, is the briefest section of the five. Does this mean that there are fewer benefits than costs, or is it simply that the experts in the field are all in agreement? The dilemma in 'Facing a Dilemma' is concerned with budgets. The question is whether libraries, within their restricted budgets, can afford to introduce on-line search services and, if so, ought they to charge their clients? In this section, Cooper and DeWath describe a study that compares the costs of average searches when the search was free and when the user was charged. Also in this topic, Rice provides a useful review of this issue together with a discussion of the alternatives. The final section of this book, 'Predicting Future Trends', is the one that taxed my imagination most. Bysouth's introduction was brief, useful and reason- ably up to date. The reprints themselves, however, were originally printed in 1978, 1979, 1981 (x 3), 1982 and 1983. Barwise's 1979 article is even entitled (pre- dictively at the time), 'The cost of literature search in 1985'! Even allowing for the fact that some of the pre- dictions made in this section are concerned with the 1990s, I cannot see the relevance of publishing predic- tions on a future that is now significantly in the past. The series editor, Blaise Cronin, writes: 'The raison d'être of this series is quite simple: to provide the reader with a clear understanding of the disciplinary base of information science.' This volume goes some way towards that objective. However, like the curate' s egg, it is good in parts, and the reader has to work hard to break through the shell of poor presentation in order to reach the good parts. Anthony Booth Expert Systems in Libraries by Gibb, Forbes (ed.). Proceedings of a conference of the Library Association Information Technology Group and the Library and Information Re- search Group, November 1985. Published by Taylor Graham, 1986. ISBN 0 947568 10 7. This is a slim 97-page paperback containing five of the papers presented at a one-day conference on Expert Systems in Libraries held in Birmingham on Novem- ber 8th 1985. The conference, the aims of which were `to provide an introduction to expert systems in gen- eral and to review the types of books that expert systems in general and to review the types of books that expert systems are being applied to within the library and information science community', was organized jointly by the Library Association Information Tech- nology Group and the Library and Information Re- search Group. The topics covered in this book include an introduc- tory overview of expert systems, software for develop- ing expert systems and their possible application to reference work, classification and cataloguing. All the papers are well written and informative. Unfortu- nately, since the field has now moved on considerably, most of the references referring to the prices and avail- ability of software are hopelessly out of date. For example, to name but two changes, ESP Advisor has now been superseded by Advisor 2 and Expert-Ease has been superseded by Super-Expert and more recently by Expert-Rule. However, despite this and the fact that much more detailed research has now been completed in the areas discussed in the book, 1 feel it is still well worth reading, particularly by those people wishing to venture into this relatively new and exciting field. One final comment what happened to the paper which was presented in a lively fashion at the confer- ence by Professor Alty? Anne Morris Man-Computer Interfaces: An Introduction to Software Design and Implementation by Coates, R B and Vlaeminke, I. Published by Blackwell Scientific Publications, Oxford, 1987. Recent studies in the US seem to indicate that up to 70 per cent of the costs of a software system are incurred after it has been written and deemed complete. The bulk of this cost can be explained by the need for modifications to the software and for staff training before the system can be used. Many of the obstacles faced by those learning a new system can be attributed to the fact that the designers probably felt quite at home with computers and could well have been totally unaware of the difficulties that confront the average end-user. In order to avoid this problem, and to produce a system that is easy and per- haps even enjoyable to use, it is important to consider the 'human factors' perspective quite early on in the design of the system. Successfully adopted, such an approach can lead to a significant increase in producti- vity and to a corresponding reduction in the number of errors made by users. The new book, Man-Computer Interfaces: An Intro- duction to Software Design and Implementation, by Coates and Vlaeminke, is an introduction to the human factors aspects of software system design. Expert Systems in Libraries by Gibb, Forbes (ed.). Proceedings of a conference of the Library Association Information Technology Group and the Library and Information Research Group, November 1985. Published by Taylor Graham, 1986. ISBN 0 947568 10 7. 
work_7bdviwrlcfgxjdvl3uvswa2y6a	----	Azam Beg Search this site Azam Beg Home Books Logic Design Fundamentals Fundamentals of Fault Tolerance Azam Beg Azam Beg I received a PhD degree in Electrical/Computer Engineering from Mississippi State University, MS, USA. I have diverse industrial, research, and academic experience in the fields of computer systems, control systems, integrated circuits, circuit design and testing, and applied machine learning. I worked as a test/product engineer and as a team lead at Motion Control Engineering, Rancho Cordova, CA, USA. Then I joined Intel Corporation, Folsom, CA, USA, as a Flash Memory Test Engineer. Afterwards, I moved to the company’s Microprocessor Division. There I first worked as a Pre-Silicon Architecture Validation Engineer and then as a Microprocessor Design Engineer. In 2005, I joined the College of Information Technology, United Arab Emirates University (UAEU), Al-Ain, UAE. I am currently an Associate Professor in the Computer & Network Engineering department at the UAEU. I am the author and co-author of more than 100 journal and conference papers, and book chapters. I hold four patents (one from UK Intellectual Property Office and three from the US Patent Office). My research interests include design and modeling of low-energy/low-power and reliable digital circuits, applied machine learning techniques and their hardware implementation, and academic assessment tools and methods. My Google Scholar page is here. mail @ azambeg.com Report abuse Page details Page updated Google Sites Report abuse 
work_7c2ec2eccraifjvvadb7744yeu	----	ME-OWA based DEMATEL reliability apportionment method Expert Systems with Applications 38 (2011) 9713–9723 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a ME-OWA based DEMATEL reliability apportionment method Cheng-Shih Liaw a,⇑, Yung-Chia Chang a, Kuei-Hu Chang b, Thing-Yuan Chang c a Department of Industrial Engineering and Management, National Chiao Tung University, Hsinchu 300, Taiwan b Department of Management Sciences, R.O.C. Military Academy, Kaohsiung 830, Taiwan c Department of Information Management, National Chin-Yi University of Technology, Taichung 411, Taiwan a r t i c l e i n f o Keywords: Reliability allocation Maximal entropy ordered weighted averaging (ME-OWA) Feasibility-of-objectives technique Decision making trial and evaluation laboratory (DEMATEL) 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.02.029 ⇑ Corresponding author. Tel.: +886 3 5731815; fax: E-mail address: liaw1158@yahoo.com.tw (C.-S. Lia a b s t r a c t The maximal entropy ordered weighted averaging (ME-OWA)-based decision making trial and evaluation laboratory (DEMATEL) method for reliability allocation has been examined. The assessment results show that most conventional reliability allocation methods have five fundamental problems. The first problem is the measurement scale; while the second problem is that the system allocation factors are not equally weighted to one another, the third problem is that most reliability allocations methods often neglect many important features, such as maintainability and risk issues. The fourth problem is that they do not consider indirect relations between subsystems or components, and the fifth problem is that they do not consider predicted failure rate in the apportionment process. This study evaluated reliability allo- cation using a fighter aircraft’s digital flight control computer (DFLCC). The proposed method offers sev- eral benefits compared with current military and commercial approaches. The computational results clearly demonstrate the advantages of the proposed approach for solving the five fundamental problems. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Reliability allocation is a top-down approach for apportioning accuracy goals in a system, which is essential when different de- sign teams, subcontractors, or manufacturers are involved, as it ef- fects the system safety and usability of the product. The purpose of the reliability allocation is to assign limited resources to the most important subsystems or components and ensure that the product can achieve its designed functions under specific operating condi- tions. The allocation technique significantly influences product life cycle cost and system operational effectiveness. Apportioning system reliability is a multiple-criteria decision- making task, which is usually allocated on the basis of the compo- nent performance and/or cost as criteria. Fuqua (1987) introduced the ‘‘Reliability engineering for electronic design’’, and MIL-HDBK- 338B (1988) defines four approaches to allocating reliability, which includes the equal apportionment technique, the ARINC apportion- ment technique, the feasibility-of-objectives (FOO) technique, and the minimization of effort algorithm. Kuo (1999) in his book ‘‘Reli- ability Assurance: Application for engineering and management’’, introduced four approaches to allocating reliability, which in- cluded the equal apportionment technique, the ARINC apportion- ment technique, the average weighting allocation method, and the pair comparison method. These conventional reliability methods have been widely and successfully applied in a great ll rights reserved. +886 3 5722392. w). many domains (Anderson, 1976; Fuqua, 1987; Kuo, 1999; Smedley, 1992). In addition to these methods, Bracha (1964) introduced an allocated reliability method using four factors: state of the art, sub- system complexity as estimated by the number of parts, environ- mental conditions, and relative operating time, whereas Karmiol (1965) evaluated the complexity, state of the art, operational pro- file, and criticality of the system to mission objectives to apportion subsystem reliability. Boyd (1992) proposed the Boyd method to combine the equal method with the ARINC method, while Falcone, Silvestri, and Bona (2002) used the integrated factors method (IFM) using four factors, criticality (C), complexity (K), functionality (F), and effectiveness (O), to calculate system reliability for an aero- space prototype project. However, the Karmiol method, the FOO technique, the average weighting allocation method, and the IFM method assume an equal interval between category labels; therefore, the operations of mul- tiplication and division are not meaningful on ordinal numbers, and addition and subtraction, while sometimes meaningful, must be done carefully because they assume equal intervals between category labels. Thus these methods all share a common weakness in their measurement scale. Take the IFM as an example: the four system factors used in allocation reliability—criticality (C), com- plexity (K), functionality (F), effectiveness (O)—obtain the subsys- tem reliability IGi = Ki ⁄ Fi ⁄ Oi/Ci (IGi: global index relative to the subsystem). In the use of the IFM method, the four system factors C, K, F, and O are rated from range 1 to 10. These four system factors are multiplied and divided to derive the IGi; i.e., IGi = Ki ⁄ Fi ⁄ Oi/Ci—the multiplied results inIG range from 1 to 1000. Higher http://dx.doi.org/10.1016/j.eswa.2011.02.029 mailto:liaw1158@yahoo.com.tw http://dx.doi.org/10.1016/j.eswa.2011.02.029 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 9714 C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 IGs are assumed to be more of an overall rating than those having a lower IG. From a calculation perspective, the IFM method is simple, easy to understand, straightforward to use, and well documented for ease of reference. However, the method has problems with its measurement scale. The first problem is that the four system fac- tors, K, F, O, and C, are evaluated according to discrete ordinal scales of measure, which represent serious flaws from a technical perspective; in particular, multiplication and/or division is not meaningful and in fact is misleading. The second problem is that the four system factors are not equally weighted, which makes the analysis and interpretation of the results problematic. For example, for two components with IG values of (8 � 2 � 2)/2 = 16 and (7 � 3 � 2)/2 = 21, respectively, the former should have had a higher reliability allocation overall rating than the latter, even though it has a lower IG value. In resolving these two problems, Chang, Chang, and Liaw (2009) provided an innovative reliability allocation using the maximal entropy ordered weighted averaging (ME-OWA) method. This approach uses Yager’s OWA (1988) and the ME-OWA (Chang, Cheng, & Chang, 2008; Fuller & Majlender, 2001) operators, which uses Lagrange multipliers on Yager’s OWA equation to derive a polynomial equation, and determines the optimal weighting vector under the maximal entropy operator. This method is a both simple and effective approach that can efficiently resolve the two short- comings of the conventional allocation methods. However, most current reliability allocation methods do not consider indirect rela- tions between subsystems or components; also, most reliability allocation methods do not consider predicted failure rate in the evaluation process. The Battelle Memorial Institute, through its Geneva Research Centre, first developed the decision making trial and evaluation laboratory (DEMATEL) method (Gabus & Fontela, 1973). It is a potent method that gathers group knowledge for cap- turing the causal relationships between criteria, which is an impor- tant analytical tool to prioritize the alternatives based on the type of relationships and severity of influences. The DEMATEL method to analyze indirect relations has been successfully used in many industrial fields, such as marketing strategies, R&D projects, e-learning evaluations, managers’ competencies, control systems, and airline safety problems (Chiu, Chen, Tzeng, & Shyu, 2006; Hori & Shimizu, 1999; Lin & Wu, 2008; Tzeng, Chiang, & Li, 2007). Chang et al. (2009) introduced the ordered weighted geometric averaging (OWGA) operator and DEMATEL method to evaluate the ordering of risks for failure problems. This is a more general risk priority number (RPN) methodology and provides a more general and reasonable risk assessment ranking. The proposed approach, using ME-OWA, was used throughout the DEMATEL calculation to determine subsystem allocation weighting factors, and the DEMATEL technique also can consider indirect relations. After DEMATEL calculation processes, the proposed method also consid- ers each subsystem’s predicted failure rate. The higher ME- OWA-based DEMATEL values should have a higher reliability allocation overall rating and apportion a higher reliability ratio into subsystems or components. Meanwhile, when using the situation parameter (a), considering the indirect relationship and predicted subsystem failure rate at same time, the proposed method can effi- ciently resolve the five shortcomings of the conventional reliability allocation methods. This study evaluates reliability allocation using a fighter aircraft digital flight control computer (DFLCC). The results from comparison with conventional reliability methods show that the proposed method is an effective methodology, with proven effectiveness yet flexible and yielding accurate results. The remainder of this paper is organized as follows: Section 2 introduces the ME-OWA operations and applications, Section 3 introduces the DEMATEL method, Section 4 introduces conven- tional reliability allocation methods, Section 5 proposes the ME-OWA-based DEMATEL method, and in Section 6, an example is drawn from an aircraft fighter’s DFLCC using the proposed approach for reliability allocation assessment. Section 7 is the conclusion. 2. ME-OWA operators and its operations 2.1. ME-OWA operators Yager (1988) first introduced the concept of OWA operators, which are important aggregation operators within the class of weighted aggregation methods. They have the ability to derive optimal weights of the attributes based on the rating of the weight- ing vectors after an aggregation process (see Definition 1). Definition 1. An OWA operator of dimension n is mapped F: Rn ? R, which has an associated n weighting vector W = [w1, w2, . . . , wn] T of the properties P iwi ¼ 1;8wi 2 ½0; 1�; i ¼ 1; . . . ; n, such that fða1; a2; . . . ; anÞ¼ Xn i¼1 wi bi ð1Þ where bi is the ith largest element in the vector (a1, a2, . . . , an), and b1 P b2 P . . . P bn. Yager (1988) also introduced two important characterizing measurements with respect to the weighting vector W of the OWA operator. One of these two measures is orness of the aggregation, which is defined in Definition 2. Definition 2. Assume F is an OWA aggregation operator with a weighting function W = [w1, w2, . . . , wn]. The degree of orness asso- ciated with this operator is defined as: ornessðWÞ¼ 1 n � 1 Xn i¼1 ðn � iÞwi ð2Þ where orness(W) = a is a situation parameter. It is clear that orness(W) 2 [0, 1] holds for any weighting vector. The second characterizing measurement introduced by Yager (1988) is a measure of dispersion of the aggregation, which is defined in Definition 3. Definition 3. Assume W is a weighting vector with elements w1, . . . , wn; then the measure of dispersion of W is defined as: dispersionðWÞ¼� Xn i¼1 wi ln wi: ð3Þ O’Hagan (1988) combined the principle of maximum entropy and OWA operators to propose a particular OWA weight that has maximum entropy with a given level of orness. This approach is based on the solution of the following mathematical programming problem: Maximize � Xn i¼1 wi ln wi ð4Þ Subject to : 1 n � 1 Xn i¼1 ðn � iÞwi ¼ a; 0 6 a 6 1; ð5Þ Xn i¼1 wi ¼ 1; 0 6 wi 6 1; i ¼ 1; . . . ; n: ð6Þ 2.2. Determination of ME-OWA weights Fuller and Majlender (2001) used the method of Lagrange mul- tipliers on Yager’s OWA equation to derive a polynomial equation, C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 9715 which can determine the optimal weighting vector under the max- imal entropy. By their method, the associated weighting vector is easily obtained by Eqs. (7)–(9): wi ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi wn�j1 w j�1 n n�1 q ð7Þ and wn ¼ ððn � 1Þa � nÞw1 þ 1 ðn � 1Þa þ 1 � nw1 ð8Þ then w1½ðn � 1Þa þ 1 � nw1� n ¼ððn � 1ÞaÞn�1½ððn � 1Þa � nÞw1 þ 1� ð9Þ where w is the weight vector, n is the number of attributes, and a is the situation parameter. 3. DEMATEL methodology The Battelle Memorial Institute, through its Geneva Research Centre (Gabus & Fontela, 1973), first developed the DEMATEL method. It is a potent method that gathers group knowledge for capturing the causal relationship between criteria and can precisely ascertain the cause–effect relationship of criteria when measuring a problem. In recent years, the DEMATEL method has been successfully applied in different industries and in many fields (Chiu et al., 2006; Hori & Shimizu, 1999; Lin & Wu, 2008; Tzeng et al., 2007). The original DEMATEL method was aimed at the fragmented and antagonistic phenomena of world societies and the search for integrated solutions. It is especially practical and useful for visualizing the structure of complicated causal relationships with matrices or digraphs. The matrices or digraphs portray a contextual relation between the elements of the system in which a numeral represents the strength of influ- ence. Hence, the DEMATEL method can convert the relationship between the causes and effects of criteria into an intelligible structural model of the system, which are not only the direct influences taken into account but also the indirect influences among multiple factors. In this section, we briefly describe the DEMATEL method and procedure. 3.1. Outline of the DEMATEL method The essential of the DEMATEL method is reviewed below (Seyed-Hosseini, Safaei, & Asgharpour, 2006). Definition 4. The pair-wise comparison scale may be designated into four levels, where the scores of 0, 1, 2, and 3 represent ‘‘No influence’’, ‘‘Low influence’’, ‘‘High influence’’, and ‘‘Very high influence’’, respectively. Definition 5. The initial direct-relation matrix Z is an n � n matrix that is obtained by pair-wise comparisons in terms of influences and directions between criteria, in which Zij is denoted as the degree to which the criterion Di affects criterion Dj. Accordingly, all principal diagonal elements Zii of matrix Z are set to zero: ð10Þ Definition 6. Let s ¼ max 16j6n Xn j¼1 zij ! : ð11Þ Then, the normalized direct-relation matrix X can be obtained through the following formula: X ¼ Z s : ð12Þ Definition 7. The total relation matrix T can be acquired by using formula (13), in which the I is denoted as the identity matrix: T ¼ limit k!1 ðX þ X2 þ�� �þ XkÞ¼ XðI � XÞ�1: ð13Þ Definition 8. Let tij (i, j = 1, 2, . . . , n) be the elements of the total- relation matrix T; then, the sum of the rows and the sum of the col- umns, denoted as Ri and Cj, respectively, can be obtained through the following two formulas: Di ¼ Xn j¼1 tij ði ¼ 1; 2; . . . ; nÞ; ð14Þ Rj ¼ Xn i¼1 tij ðj ¼ 1; 2; . . . ; nÞ: ð15Þ Definition 9. A causal diagram can be acquired by mapping the ordered pairs of (R + C, R � C), where the horizontal axis (R + C) is named ‘‘Prominence’’ and the vertical axis (R � C) is named ‘‘Relation.’’ In the causal diagram, the horizontal axis ‘‘Prominence’’ shows how important the criterion is, whereas the vertical axis ‘‘Relation’’ may divide the criteria into the cause and effect groups. When the value (R � C) is positive, the criterion belongs to the cause group. If the value (R � C) is negative, the criterion belongs to the effect group. Hence, causal diagrams can visualize the complicated causal relationships between criteria into a visible structural model and provide valuable insight for problem-solving. Furthermore, with the help of a causal diagram, this study will allow proper decisions to be made by recognizing the difference between cause and effect criteria. 3.2. The procedure of the DEMATEL method The DEMATEL method can separate the relevant criteria of a system into the cause and effect groups to facilitate accurate decision-making. A DEMATEL procedure is explained as follows (Seyed-Hosseini et al., 2006): (1) A system designer or decision-maker evaluates the relation- ship between sets of paired alternatives. As a result of this evaluation, a matrix M is obtained as the initial data of the DEMATEL analysis. (2) The elements of the direct relative severity matrix (DRSM) are obtained by Eq. (11). It is the normalized version of matrix M. (3) The elements of the direct and indirect relative severity matrix (DIRSM) are obtained by Eq. (12). The DIRSM consists of all of the relations, including direct and indirect relations between alternatives. (4) Using the values of R + C and R � C, where C is the sum of the columns and R is the sum of the rows of the DIRSM, a level of influence and a level of relationship are defined. The value 9716 C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 R � C indicates the severity of influence for each alternative. Similarly, the value of R + C indicates the degree of relation between each alternative with one another. 4. Conventional reliability allocation methods This section presents a literature review of the applications of reliability allocation. Currently, many reliability allocation tech- niques are available, including the AGREE method (1957), ARINC apportionment technique (1976), Bracha method (1964), Karmiol method (1965), equalization allocation method, pair comparison allocation, the FOO technique, minimization of effort algorithm (1988), Boyd method (1992), average weighting allocation method (Kuo, 1999), Base method, and IFM method (Falcone et al., 2002), among others. The allocation technique significantly influences product life cycle cost and system operational effectiveness. Since the ARINC apportionment technique is an important method in military reliability allocation design, the basic definition and pro- cedure of the ARINC apportionment technique and the ME-OWA apportionment method are reviewed in this section. 4.1. ARINC apportionment technique This method assumes a series of subsystems with constant fail- ure rates. Other fundamental assumptions are: (1) series subsys- tems, (2) constant failure rates, (3) same mission duration time T for each subsystem, and (4) a pre-defined, known allowable system failure rate: k⁄. Suppose a system is composed of N subsystems. Let k�i be the failure rate allocated to subsystem i. The objective is to choose k�i such that (1988): Xn i¼1 k�i P k � : ð16Þ Determine the subsystem failure rates (ki) from past observa- tion or estimation; assign a weighting factor (wi) to each subsys- tem according to the failure rates determined by Eq. (17): wi ¼ kiPn i¼1ki : ð17Þ Allocate subsystem failure rate requirements as follows: k�i ¼ wik � : ð18Þ 4.2. ME-OWA apportionment method Chang et al. (2009) presented an innovative reliability allocation using the ME-OWA method for fighter aircraft RADAR programs. This method can determine the optimal weighting vector under maximal entropy, and the OWA operator has the ability to ascer- tain the optimal reliability allocation rating after an aggregation process. With the optimal weighting vector under maximal entro- py with respect to different a values, sensitivity analysis enables the identification of different a values to evaluate their impact on the reliability allocation rating using Eqs. (7)–(9) with n = 4. Re- sults from this analysis are presented in Table 1. Table 1 The optimal weighting vector under maximal entropy (n = 4). Weight wl w2 w3 w4 a = 0.5 0.250000 0.250000 0.250000 0.250000 a = 0.6 0.416657 0.233398 0.130859 0.073547 a = 0.7 0.493805 0.237305 0.113770 0.054918 a = 0.8 0.596466 0.251953 0.106445 0.045018 a = 0.9 0.764099 0.182129 0.043457 0.010365 a = 1.0 1.000000 0.000000 0.000000 0.000000 Suppose a system is composed of m subsystems; n is the num- ber of system factors. Let Rs be the system’s allocated rating, Ri be the allocated rating to the ith subsystem, and T be the mission duration; then the system failure rate ks is determined by Eq. (19) (1988): ks ¼� lnðRÞ=T: ð19Þ This method uses four subsystem allocation factors, which are computed as a function of a numerical rating, such as system intri- cacy (I), state of the art (S), performance time (P), environment (E), mission time (T), complexity (K), functionality (F), effectiveness (E), and operational profile (O). Each rating is based on a scale from 1 to 10 and is estimated using design engineering and expert judg- ments. In order to compare the different method capabilities, the same influential system reliability factors were selected: I, S, P, and E. Also, the same estimated rating derived from design engi- neering and expert judgment. Based on Table 1 and Eq. (1), calcu- late the aggregated values by OWA weights with respect to different values of (a = 0.5, 0.6, 0.7, 0.8, 0.9, 1); a = 1 is used to represent the situation when the decision-maker is maximally optimistic (a pure optimist), and a = 0.5 is used when the deci- sion-maker faces a moderate assessment. Use Eq. (20) to calculate complexity C0k;8k. C0k ¼ w0k W0 ; 8k: ð20Þ Use Eq. (21) to calculate the allocated subsystem failure rate kk, "k. kk ¼ C 0 kks; 8k: ð21Þ 5. Proposed ME-OWA-based DEMATEL apportionment method 5.1. Advantages of the ME-OWA-based DEMATEL apportionment method With the FOO technique, the average weighting allocation method and IFM method, such as the ISPE and IG values, are ordinal measurement scales; therefore, the operations of multiplication and division are not meaningful and misleading. As a result, some (I, S, P, E) scenarios produce an ISPE value that is lower than other combinations but potentially produce a higher reliability allocation overall rating. For example, the scenario with ISPE value 9 � 5 � 2 � 2 = 180 is lower than the scenario with ISPE value 7 � 5 � 3 � 2 = 210, even thought it should have a higher reliabil- ity allocation overall rating. Therefore, I, S, P, and E are not equally weighted with respect to one another in terms of overall rating. Meanwhile, most current reliability allocation methods do not con- sider indirect relations between subsystems or components; also, most reliability allocation methods do not consider predicted fail- ure rate in the evaluation process. In resolving the five conventional reliability allocation problems mentioned above, the proposed approach is to combine the ME-OWA, DEMATEL, and the ARINC methods. The ME-OWA is used to derive ISPE values and then uses the DEMATEL for capturing the causal relationship between criteria; this method also considers a subsystem’s predicted failure rate at the same time. After calculating each subsystem’s total indirect relationship, then apportion a reasonable reliability rating into subsystems or components. To maintain the R � C value in a positive state, this study pro- poses the R � c value. c represents the average severity of influence for each alternative; the c value can be obtained through the fol- lowing formula—then, the R � c value is derived: ci ¼ CiPn i¼1 Ci : ð22Þ C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 9717 5.2. Procedures of the ME-OWA-based DEMATEL apportionment method The detailed procedure of the proposed ME-OWA-based DEMA- TEL apportionment method is organized into 10 steps and is de- scribed as follows: Step 1: List the structure of systems and subsystems. Step 2: Define the system reliability and mission time. Step 3: Compute the system failure rate from system specifica- tions. Based on Eq. (19), derive the system failure rate ks. Step 4: Determine the scales for I, S, P, and E, respectively. Subsystem allocation factors are computed as a function of numerical ratings of I, S, P, and E. Step 5: Perform DEMATEL procedure. The procedure of the DEMATEL process is as follows (Seyed-Hosseini et al., 2006): (1) A system designer or decision-maker evaluates the relationship between subsystems of paired alterna- tives. The pair-wise comparison scale may be desig- nated into 10 levels, where the scores of 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9 represent the influence levels ‘‘None’’, ‘‘Very minor’’, ‘‘Minor’’, ‘‘Low’’, ‘‘Moderate’’, ‘‘Significant’’, ‘‘Major’’, ‘‘Extreme’’, ‘‘Serious’’, and ‘‘Hazardous’’, respectively. (2) Obtain the elements of DRSM by Eq. (11). (3) Obtain the elements of DIRSM by Eq. (12). (4) Calculate the values of R + C, R � C, and R � c. Step 6: Compute reliability allocation values by assigning differ- ent values to the conditional parameter (a = 0.5, 0.6, 0.7, 0.8, 0.9, 1). Based on Table 1 and Eq. (1), calculate the aggregated val- ues by OWA weights with respect to different values of (a = 0.5, 0.6, 0.7, 0.8, 0.9, 1). Step 7: Compute the allocation rating ri for each subsystem and derive the overall rating wi for the kth subsystem. According to the results of DEMATEL’s calculation, using Eq. (22) to derive ci, then calculate R � c values and use Eq. (23) for allocation weighting factors; calculate the aggregated value by ME-OWA-based weights. The appor- tioning weighing ratio also consider the prediction failure rate in the apportionment process. Use Eq. (23) to calcu- late the allocate weighting factors wi, "i: wi ¼ Xn i¼1 ISPEi � Di � ki= Xn i¼1 Xm i¼1 ISPEi � Di � ki ð23Þ where n: number of ISPE, m: number of subsystems, ISPEi: ME-OWA values for subsystem i, Di: R � c values for subsystem i, and ki: pre- dicted failure rate for subsystem i. Step 8: Compute the complexity C0k for the kth subsystem, "k. Use Eq. (20) to calculate the complexity C0k;8k. Step 9: Compute the allocated subsystem failure rate. Use Eq. (21) to calculate the allocated subsystem failure rate kk, "k. Step 10: Analyze the results and select the optimal reliability allo- cation decision. 6. A case study of the ME-OWA-based DEMATEL apportionment method This paper presents a real-world illustrative example for imple- mentation of the ME-OWA-based DEMATEL apportionment method. A case study of a DFLCC that is installed on a fighter aircraft, drawn from an aircraft company in Taiwan, was used to demonstrate the proposed approach. The DFLCC is a triplex redundant digital flight control computer. The DFLCC receives discrete, analog, and digital input signals from various sensors and computers throughout the fighter aircraft. It processes these input signals through a series of redundant management and control law algorithms, which were developed and supplied by the customer. The DFLCC produces various discrete, analog, and digital output signals that are sent to provide the pilot with the digital flight control system status and to the primary servo actua- tors and leading edge flap, used to control the aircraft’s motions. The DFLCC consists of eight major shop replaceable units (SRUs). The SRUs are the CPU circuit card assembly (CPU card), Input/ Output Processor circuit card assembly (IOP card), ADIO circuit card assembly (ADIO card), 1553/DOUT circuit card assembly (1553/DOUT card), power supply assembly (P/S card), ACS circuit card assembly (ACS card), IBU circuit card assembly (IBU card), and motherboard. The DFLCC block diagram is show in Fig. 1. To accurately assess reliability requirements of subsystems or components, system design engineering needs to clearly define the diagram for each system and subsystem. The diagram for the DFLCC is shown in Fig. 2. 6.1. ARINC apportionment technique analysis To illustrate this method, consider a DFLCC that consists of eight major SRUs. The SRUs are the CPU card, IOP card, ADIO card, 1553/ DOUT card, P/S card, ACS card, IBU card, and motherboard. There are eight subsystems with predicted failure rates of k1 = 22.4 (CPU card), k2 = 23.3 (IOP card), k3 = 15.3 (ADIO card), k4 = 13.1 (1553/DOUT card), k5 = 44.8 (P/S card), k6 = 35.5 (ACS card), k7 = 11.4 (IBU card), and k8 = 16.6 (motherboard) failures per 106 h, respectively. The system has a mission time of 3 h, and a 0.999412 reliability is required. The apportioned system reliability goals are found by Eq. (19) as follows: ks ¼� lnðRÞ=T ¼� lnð0:999412Þ=3 ¼ 196:058 failures per 106 h: Based on the predicted subsystem failure rate for each subsystem, as follows: k1 ¼ 22:4; k2 ¼ 23:3; k3 ¼ 15:3; k4 ¼ 13:1; k5 ¼ 44:8 k6 ¼ 35:5; k7 ¼ 11:4; k8 ¼ 16:6: Using Eq. (17), the weighting factor (wi) for each subsystem is as follows: w1 ¼ 22:4 22:4 þ 23:3 þ 15:3 þ 13:1 þ 44:8 þ 35:5 þ 11:4 þ 16:6 ¼ 0:122923 w2 ¼ 23:3 22:4 þ 23:3 þ 15:3 þ 13:1 þ 44:8 þ 35:5 þ 11:4 þ 16:6 ¼ 0:127650 w3 ¼ 15:3 22:4 þ 23:3 þ 15:3 þ 13:1 þ 44:8 þ 35:5 þ 11:4 þ 16:6 ¼ 0:083677 w4 ¼ 13:1 22:4 þ 23:3 þ 15:3 þ 13:1 þ 44:8 þ 35:5 þ 11:4 þ 16:6 ¼ 0:071762 w5 ¼ 44:8 22:4 þ 23:3 þ 15:3 þ 13:1 þ 44:8 þ 35:5 þ 11:4 þ 16:6 ¼ 0:245880 w6 ¼ 35:5 22:4 þ 23:3 þ 15:3 þ 13:1 þ 44:8 þ 35:5 þ 11:4 þ 16:6 ¼ 0:194512 w7 ¼ 11:4 22:4 þ 23:3 þ 15:3 þ 13:1 þ 44:8 þ 35:5 þ 11:4 þ 16:6 ¼ 0:062637 w8 ¼ 16:6 22:4 þ 23:3 þ 15:3 þ 13:1 þ 44:8 þ 35:5 þ 11:4 þ 16:6 ¼ 0:090958: Channel A Channel B Channel C Start Mother board End ACS card IBU card ADIO card IOP card CPU card 1553/ DOUT card P/S card 1553/ DOUT card ACS card IBU card ADIO card IOP card CPU card P/S card ACS card IBU card ADIO card IOP card CPU card 1553/ DOUT card P/S card Fig. 1. The digital flight control computer (DFLCC) block diagram. Fig. 2. Diagram for the digital flight control computer (DFLCC). 9718 C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 9719 Using Eq. (18), the corresponding allocated subsystem reliability requirements are: k�1 ¼ w1ks ¼ 24:10014 k�2 ¼ w2ks ¼ 25:02689 k�3 ¼ w3ks ¼ 16:40557 k�4 ¼ w4ks ¼ 14:06958 k�5 ¼ w5ks ¼ 48:20675 k�6 ¼ w6ks ¼ 38:13554 k�7 ¼ w7ks ¼ 12:28047 k�8 ¼ w8ks ¼ 17:83306: 6.2. ME-OWA method analysis The proposed approach uses the ME-OWA operator for weight calculations. A sensitivity analysis using different values of a is presented to evaluate their impact on the reliability allocation rat- ing. Based on Table 1, the optimal weighting under maximal entro- py (n = 4), and Eq. (1), the failure rate of the CPU card is calculated as follows. For a = 0.5 (used when the decision-maker faces a moderate assessment), if I, S, P, and E for the subsystem CPU card are 10, 9, 8, and 7, respectively, and the weighting vector contains w1 = 0.25, w2 = 0.25, w3 = 0.25, and w4 = 0.25, then w0k ¼ð10 � 0:25Þþð9 � 0:25Þþð8 � 0:25Þþð7 � 0:25Þ¼ 8:5 C0k ¼ 8:5=45:5 ¼ 0:187: As a result, the failure rate of the CPU card (for a = 0.5) = 196.058 � 0.187 = 36.62622 per 106 h. a = 1 is used to represent the situation when the decision- maker is maximally optimistic (a pure optimist). For a = 1, the OWA(a1, a2, a3) = Max(a1, a2, a3) if I, S, P, and E for the subsystem CPU card are 10, 9, 8, and 7, respectively. The weighting vector con- tains w1 = 1, w2 = 0, w3 = 0, and w4 = 0; consequently, w0k ¼ð10 � 1Þþð9 � 0Þþð8 � 0Þþð7 � 0Þ¼ 10 C0k ¼ 10=59 ¼ 0:169: As a result, the failure rate of the CPU card (for a = 1) = 196.058 � 0.169 = 33.23017 per 106 h. Following the calculation above, the aggregated values of OWA weights by different values of a (a = 0.5, 0.6, 0.7, 0.8, 0.9, 1) are cal- culated for the subsystem CPU card. The resulting failure rates are 33.23017, 33.79919, 34.42269, 34.31116, 34.95256, and 36.62622 per 106 h, respectively. The failure rates for the CPU card, IOP card, ADIO card, 1553/DOUT card, P/S card, ACS card, IBU card, and motherboard are also calculated, and the results are summarized Table 2 The reliability allocation results for DFLCC system (ME-OWA method). Method (1) (2) (3) (4) Intricacy (I) State-of-the-art (S) Performance time (P) Environment (E) CPU 10 9 8 7 IOP 9 9 5 6 ADIO 5 6 3 4 1553/DOUT 7 8 7 6 P/S 4 5 10 9 ACS 3 4 4 3 IBU 2 3 3 3 Motherboard 3 4 9 4 Total 43 48 49 42 in Table 2, columns (5) through (10). The reliability allocation re- sults for the DFLCC system (ME-OWA method) are also shown in Table 2. 6.3. Proposed ME-OWA-based DEMATEL apportionment method analysis The proposed approach is to combine the ME-OWA, DEMATEL, and ARINC methods. The ME-OWA uses the four system reliability factors I, S, P, and E after the estimated rating from design engi- neering and expert judgment and under situation parameters (a = 0.5, 0.6, 0.7, 0.8, 0.9, 1) to derive the ISPE values. The results are summarized in Table 2, columns (5) through (10). Then, the DEMATEL is used for capturing the causal relationships between criteria. This method also considers a subsystem’s predicted failure rate at the same time. Based on the results of system engineering and expert evaluation of the relationship between sets of paired subsystems or components, the initial direct-relation matrix Z is shown in Fig. 3. According to Eq. (12), the elements of the direct relative severity matrix (DRSM) are obtained and shown in Fig. 4. According to Eq. (13), the elements of the direct and indirect relative severity matrix (DIRSM) are obtained, which are shown in Fig. 5. According to Eq. (10)–(15) and using Eq. (22), the outcome of DEMATEL implementation for the DFLCC with respect to direct and indirect relationships is shown in Table 3. After calculating each subsystem’s total indirect relationship, then apportion reasonable reliability ratings into subsystems or components. The calculation results for this model are as follows. The subsystem failure rates (kk, "k) are estimated from past observations or estimations. The SRUs are the CPU card, IOP card, ADIO card, 1553/DOUT card, P/S card, ACS card, IBU card, and motherboard. The predicted failure rates of the eight subsystems are k1 = 22.4 (CPU card), k2 = 23.3 (IOP card), k3 = 15.3 (ADIO card), k4 = 13.1 (1553/DOUT card), k5 = 44.8 (P/S card), k6 = 35.5 (ACS card), k7 = 11.4 (IBU card), and k8 = 16.6 (motherboard) failures per 106 h, respectively. The ISPE values for these eight subsystems are summarized in Table 2, columns (1) through (4). Then, using Eq. (1), we compute the overall rating and complexity factors and then ascertain the allocation failure rate for the CPU board as follows. For a = 0.5 (used when the decision-maker faces a moderate assessment), if I, S, P, and E for the subsystem CPU card are 10, 9, 8, and 7, respectively, the CPU card’s predicted failure rate is k1 = 22.4. Based on Table 1 and Eq. (1) to derive ISPE = 8.5, accord- ing to Eq. (10)–(15), the outcome of the DEMATEL R � c values is 5.25; then ME-OWA (5) (6) (7) (8) (9) (10) a = 0.5 a = 0.6 a = 0.7 a = 0.8 a = 0.9 a = 1.0 36.62622 34.95256 34.31116 34.42269 33.79919 33.23017 31.24001 31.26718 30.90323 31.12875 30.76176 29.90715 19.39035 19.69331 19.55413 19.77479 19.86120 19.93810 30.16277 28.23552 27.62428 27.65346 27.01773 26.58414 30.16277 32.21480 32.23624 32.75853 33.23659 33.23017 15.08138 14.34666 14.06539 14.09318 13.75045 13.29207 11.84966 11.11608 12.70847 10.82105 10.41737 9.96905 21.54484 24.23189 24.65510 25.40554 27.21371 29.90715 196.05800 196.05800 196.05800 196.05800 196.05800 196.05800 Fig. 3. Corresponding initial direct-relation matrix Z. Fig. 4. Corresponding DRSM of the DFLCC. Fig. 5. Corresponding DIRSM of the DFLCC. Table 3 The R + C, R � C and R � c values of the DFLCC by the ME-OWA based DEMATEL method. No. R C c R + C R � C R � c 1 5.401 5.301 0.151 10.702 0.100 5.250 2 4.853 5.346 0.152 10.198 �0.493 4.700 3 4.382 4.768 0.136 9.150 �0.386 4.246 4 4.388 4.586 0.131 8.974 �0.199 4.257 5 5.156 4.914 0.140 10.070 0.242 5.016 6 3.502 3.839 0.109 7.341 �0.338 3.392 7 3.367 3.411 0.097 6.778 �0.044 3.270 8 4.018 2.902 0.083 6.920 1.117 3.936 9720 C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 w0k ¼ 8:5 � 5:25 � 22:4 ¼ 999:6 C0k ¼ 999:6=4899:96 ¼ 0:204: As a result, the failure rate of the CPU card (for a = 0.5) = 196.058 � 0.204 = 39.99635 per 106 h. Following the calculation above, the aggregated values of OWA weights (a = 0.5) for the subsystem CPU card, IOP card, ADIO card, 1553/DOUT card, P/S card, ACS card, IBU card, and motherboard are 39.99635, 31.76884, 11.6974, 15.61947, 62.93992, 16.86418, 4.10171, and 13.07012 failures per 106 h, respectively. The results are summarized in Table 4, column (7). a = 1 is used to represent the situation when the decision- maker is maximally optimistic (a pure optimist). For a = 1, the OWA(a1, a2, a3) = Max(a1, a2, a3). IfI, S, P, and E for the subsystem CPU card are 10, 9, 8, and 7, respectively, the CPU card’s predicted failure rate is k1 = 22.4. Based on Table 1 and Eq. (1), ISPE = 10 is de- rived. According to Eq. (10)–(15) and using Eq. (22), the outcome of the DEMATEL R � c values is 5.25; then w0k ¼ 10 � 5:25 � 22:4 ¼ 1176:01 C0k ¼ 1176:01=6426:23 ¼ 0:183: As a result, the failure rate of the CPU card (for a = 1) = 196.058 � 0.183 = 35.87878 per 106 h. Following the calculation above, the aggregated values of OWA weights (a = 1) for the subsystem CPU card, IOP card, ADIO card, 1553/DOUT card, P/S card, ACS card, IBU card, and motherboard are 35.87878, 30.0706, 11.89226, 13.61114, 68.55898, 14.69581, 3.41185, and 17.93859 failures per 106 h, respectively. The aggre- gated values of OWA weights by different values of (a = 0.5, 0.6, 0.7, 0.8, 0.9, 1) for the subsystem CPU card, IOP card, ADIO card, 1553/DOUT card, P/S card, ACS card, IBU card, and motherboard are also calculated, and the results are summarized in Table 4, columns (7) through (12). The reliability allocation re- sults for the DFLCC system (ME-OWA-based DEMATEL method) are shown in Table 4. 6.4. Method comparison As shown in Table 5, the failure rates using the ARINC appor- tionment technique for the CPU card, IOP card, ADIO card, 1553/ DOUT card, P/S card, ACS card, IBU card, and motherboard are cal- culated, and the results are summarized in Table 5, row (13). The ME-OWA method results are shown in Table 5, rows (7) through (12). The ME-OWA-based DEMATEL method results are shown in Table 5, Rows (1)–(6). In this case study, using the ARINC appor- tionment technique, the appointing failure rates were calculated for the subsystem P/S card (48.20675) > ACS card (38.13554) > IOP card (25.02689) > CPU card (24.10014) > Motherboard (17.83306) > ADIO card (16.40557) > 1553/DOUT card (14.06958) > IBU card (12.28047). Compared with the ME-OWA method (a = 0.9), the appointing failure rate of the subsystem CPU card (33.79919) > P/S card (33.23659) > IOP card (30.76176) > Mother- board (27.21371) > 1553/DOUT card (27.01773) > ADIO card (19.8612) > ACS card (13.75045) > IBU card (10.41737). Compared with the ME-OWA-based DEMATEL method (a = 0.9), the appoint- Table 4 The reliability allocation results for DFLCC system (ME-OWA based DEMATEL method). Method (1) (2) (3) (4) (5) (6) ME-OWA I S P E R � c (DEMATEL) Predicted failure rate (7) (8) (9) (10) (11) (12) a = 0.5 a = 0.6 a = 0.7 a = 0.8 a = 0.9 a = 1.0 CPU 10 9 8 7 5.250 22.400 39.99635 37.74110 37.22365 37.07745 36.36190 35.87878 IOP 9 9 5 6 4.700 23.300 31.76884 31.44028 31.22120 31.22401 30.81863 30.07060 ADIO 5 6 3 4 4.246 15.300 11.69740 11.74707 11.71919 11.76663 11.80378 11.89226 1553/DOUT 7 8 7 6 4.257 13.100 15.61947 14.45767 14.21155 14.12477 13.78338 13.61114 P/S 4 5 10 9 5.016 44.800 62.93992 66.46878 66.82744 67.42426 68.32559 68.55898 ACS 3 4 4 3 3.392 35.500 16.86418 15.86289 15.62541 15.54426 15.14792 14.69581 IBU 2 3 3 3 3.270 11.400 4.10171 3.80468 4.37027 3.69459 3.55246 3.41185 Motherboard 3 4 9 4 3.936 16.600 13.07012 14.53553 14.85930 15.20203 16.26432 17.93859 Total 43 48 49 42 34.067 182.400 196.05800 196.05800 196.05800 196.05800 196.05800 196.05800 Table 5 The allocation results for three methods. Method Conditional parameter (a) CPU IOP ADIO 1553/DOUT P/S ACS IBU Motherboard 1 ME-OWA based DEMATEL 0.5 39.99635 31.76884 11.69740 15.61947 62.93992 16.86418 4.10171 13.07012 2 0.6 37.74110 31.44028 11.74707 14.45767 66.48878 15.86289 3.80468 14.53553 3 0.7 37.22365 31.22120 11.71919 14.21155 66.82744 15.62541 4.37027 14.85930 4 0.8 37.07745 31.22401 11.76663 14.12477 67.42426 15.54426 3.69459 15.20203 5 0.9 36.36190 30.81863 11.80378 13.78338 68.32559 15.14792 3.55246 16.26432 6 1 35.87878 30.07060 11.89226 13.61114 68.55898 14.69581 3.41185 17.93859 7 ME-OWA 0.5 36.62622 31.24001 19.39035 30.16277 30.16277 15.08138 11.84966 21.54484 8 0.6 34.95256 31.26718 19.69331 28.23552 32.21481 14.34666 11.11608 24.23189 9 0.7 34.31116 30.90323 19.55413 27.62428 32.23624 14.06539 12.70847 24.65510 10 0.8 34.42269 31.12875 19.77479 27.65346 32.75853 14.09318 10.82105 25.40554 11 0.9 33.79919 30.76176 19.86120 27.01773 33.23659 13.75045 10.41737 27.21371 12 1 33.23017 29.90715 19.93810 26.58414 33.23017 13.29207 9.96905 29.90715 13 ARINC – 24.10014 25.02689 16.40557 14.06958 48.20675 38.13554 12.28047 17.83306 C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 9721 ing failure rate of the subsystem P/S card (68.32559) > CPU card (36.3619) > IOP card (30.81863) > Motherboard (16.26432) > ACS card (15.14792) > 1553/DOUT card (13.78338) > ADIO card (11.80378) > IBU card (3.55246). The allocation results for the three methods are shown in Table 5. From Table 5, because the ME-OWA-based DEMATEL method using ISPE values as a based, and followed by DEMATEL calcula- tion process, which considers direct and indirect relationships between each of the subsystems at the same time. The results indicate that the P/S card and CPU card have higher failure rates, while the ADIO card and IBU card have lower failure rates. The P/S card R � c value is 5.016 and has a higher predicted failure rate (k5 = 44.8) than the CPU card (R � c values is 5.25 and pre- dicted failure rate k1 = 22.4), whereas the IBU card R � c value is 3.27 and has a lower predicted failure rate (k7 = 11.4) than the ADIO card R � c values (4.246 and predicted failure rate k3 = 15.3). From a technical perspective, the subsystems have higher R � c criteria (higher direct and indirect relationships) and higher predicted failure rates; the proposed ME-OWA-based DEMATEL apportionment method is more reasonable in appoint- ing a more reliable ratio in subsystems. As shown in Fig. 6, the results of the ME-OWA-based DEMATEL apportionment method, which incorporates ME-OWA, DEMATEL, and ARINC methods, this method obtains a correct and discrimi- nating allocation ratio. The result is a more flexible and reasonable allocation rating than the conventional ARINC apportionment tech- nique and the ME-OWA method. The results of the proposed ME- OWA-based DEMATEL method are compared with the ARINC apportionment technique and the ME-OWA method, shown in Fig. 6 below. A comparison of the ARINC apportionment technique, the ME-OWA method, and the proposed ME-OWA-based DEMATEL apportionment method is summarized in Table 6. ‘‘O’’ indicates that the related factor is applicable, whereas ‘‘X’’ indicates that the related factor is not applicable. The proposed method has concluded a number of advantages with its potentialities: (1) Proposes a combined reliability allocation method using the ME-OWA-based DEMATEL method in apportioning system reliability, which combines the ME-OWA, DEMATEL, and the ARINC methods. It can overcome the three conventional shortcomings: measurement scale problem, the not equally weighted problem, and no consideration of indirect relation- ships between subsystems during the appointment pro- cesses. It also uses the ARINC’s concept to consider the predicted failure rate for appointing reliability into subsys- tems or components. (2) Considers situation parameter factors: the ME-OWA is based on estimated ratings from design engineering and expert judgment for appointing reliability. The four system reliabil- ity factors I, S, P, and E under the situation parameter (a = 0.5, 0.6, 0.7, 0.8, 0.9, 1) are used to derived the ISPE values. This assessment result shows that the ME-OWA-based DEM- ATEL method yielded results that not only were correct but also resulted in a discriminating allocation ratio that is flex- ible for real-world applications. (3) Considers indirect relationships between subsystems and components: the combined DEMATEL calculates processes and can consider indirect relationships between subsystems and components. This calculation holds that the higher indi- rect relationship subsystems are appointed a higher alloca- tion ratio, which can efficiently allocate limited resources in subsystems or components. Fig. 6. Comparison of the five methods of reliability allocation. Table 6 Comparison of the three methods. Method Consider factor Measurement scale Order weight Indirect relationship Predicted failure rate Proposed method O O O O ME-OWA O O X X ARINC X X X O Note: ‘‘O’’ represents that the factor is applicable, and ‘‘X’’ represents that the factor is not applicable. 9722 C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 (4) Provides an organized approach and a more flexible struc- ture in reliability allocation: the ME-OWA-based DEMATEL method is applicable to the different design phases. The sys- tem reliability factors are not limited to only I, S, P, and E, and the number of factors is not limited to 4 items (3 or more items are acceptable). Depending upon the selection of applicable variables, such as system intricacy, complexity, state of the art (technology), cost, maintenance, risk, failure rate, design maturity, operating environment, and repair times, the allocating ratio is more suitable for different alter- natives. There is no limitation for implementation of DEMATEL in very large and complex systems, and it can therefore provide an improved structured arrangement for reliability allocation. 7. Conclusion This paper has successfully demonstrated the application of an ME-OWA-based DEMATEL apportionment method for reliability allocation using a fighter aircraft DFLCC. It is an easy, proven, and effective approach, which uses the ME-OWA to derive the ISPE val- ues, followed by the DEMATEL calculation processes to ascertain indirect relations. The higher ISPE values, indirect relationship sub- systems, and higher predicted failure rate translates into a higher reliability allocation overall rating. The main advantages of the proposed approach are: (1) It pro- vides an accurate yet flexible reliability allocation method, which combines the ME-OWA, DEMATEL, and the ARINC methods. The proposed approach can efficiently resolve the measurement scale problem, equally-weighted problem, and considers indirect rela- tionships between subsystems during the appointment processes. (2) The ME-OWA-based DEMATEL method uses the conditional parameter (a) to derive the ISPE values. The assessment results show that the ME-OWA-based DEMATEL method yields results that not only are accurate but also yield discriminating allocation ratios that are flexible for real-world applications. The proposed ME-OWA-based DEMATEL method can better help managers or designers make correct decisions. (3) Using DEMATEL indirect rela- tionships between subsystems and components can be considered. The calculation holds that the higher indirect relationship subsys- tems are appointed a higher allocation ratio, which can result in more efficient allocation of limited resources in subsystems or components. (4) It provides an organized approach and a more flexible structure in reliability allocation. The ME-OWA-based DEMATEL method is applicable to different design phases. The sys- tem reliability factors are not limited to only I, S, P, and E (3 or more items are acceptable). Depending upon the selection of applicable variables, such as system intricacy, complexity, state of the art (technology), cost, maintenance, risk, failure rate, design maturity, operating environment, and repair times, the allocating ratio is more suitable for different alternatives. There is no limitation for implementation of DEMATEL in very large and complex systems, and it can thereby provide an improved structured arrangement for reliability allocation. The ME-OWA-based DEMATEL apportion- ment method can also be used in a wide variety of different indus- tries and fields. The results from the comparison with conventional reliability methods show that the proposed method is both accu- rate and flexible. Also, DEMATEL can be considered in Fuzzy envi- ronments that are suggested for further research. References Advisory Group of Reliability of Electronic Equipment (AGREE) (1957). Reliability of military electronic equipment. Washington, DC: Office of the Assistant Secretary of Defense Research and Engineering. Anderson, R. T. (1976). Reliability design handbook. Chicago: ITT Research Institute. Boyd, J. A. (1992). Allocation of reliability requirements: A new approach. In Proceedings Annual Reliability and Maintainability Symposium. Las Vegas, USA: IEEE (NY). C.-S. Liaw et al. / Expert Systems with Applications 38 (2011) 9713–9723 9723 Bracha, V. J. (1964). The methods of reliability engineering. Machine Design, 70–76. Chang, K. H., Cheng, C. H., & Chang, Y. C. (2008). Reliability assessment of an aircraft propulsion system using IFS and OWA tree. Engineering Optimization, 40(10), 907–921. Chang, Y. C., Chang, K. H., & Liaw, C. S. (2009). Innovative reliability allocation using the maximal entropy ordered weighted averaging method. Computers & Industrial Engineering, 57, 1274–1281. Chiu, Y. J., Chen, H. C., Tzeng, G. H., & Shyu, J. Z. (2006). Marketing strategy based on customer behaviour for the LCD-TV. International Journal and Decision Making, 7(2/3), 143–165. Department of Defense of USA (1988). MIL-HDBK-338B, Electronic Design Reliability Handbook, 6/13–6/16. Falcone, D., Silvestri, A., & Bona, G. D. (2002). Integrated factors method (IFM): A new reliability allocation technique, SEA, IASTED. USA, Cambridge. Fuller, R., & Majlender, P. (2001). An analytic approach for obtaining maximal entropy OWA operator weights. Fuzzy Sets and Systems, 124(1), 53–57. Fuqua, N. B. (1987). Reliability engineering for electronic design. Marcel Dekker Inc. Gabus, A., & Fontela, E. (1973). Perceptions of the world problematique: Communication procedure, communicating with those bearing collective responsibility (DEMATEL Report No. 1), Battelle Geneva Research Centre, Geneva, Switzerland. Hori, S., & Shimizu, Y. (1999). Designing methods of human interface for supervisory control systems. Control Engineering Practice, 7(11), 1413–1419. Karmiol, E. D. (1965). Reliability apportionment, preliminary report EIAM-5, Task II (pp. 10–22), General Electric, Schenectady, NY. Kuo, H. E. (1999). Reliability assurance: Application for engineering and management (2nd ed.). Chinese Society for Quality. 3/16–3/17. Lin, C. J., & Wu, W. W. (2008). A causal analytical method for group decision-making under fuzzy environment. Expert Systems with Applications, 34(1), 205–213. O’Hagan, M. (1988). Aggregating template or rule antecedents in real-time expert systems with fuzzy set logic. In Proceedings of 22nd annual IEEE asilomar conference on signals, systems, computers (pp. 681–689). Pacific Grove, CA, Piscataway, NJ: IEEE. Seyed-Hosseini, S. M., Safaei, N., & Asgharpour, M. J. (2006). Reprioritization of failures in a system failure mode and effects analysis by decision making trial and evaluation laboratory technique. Reliability Engineering and System Safety, 91, 872–881. Smedley, K. (1992). Reliability analysis for LEB ring magnet power system in SSC. IEEE Transactions on Nuclear Science, 39(4), 1170–1174. Tzeng, G. H., Chiang, C. H., & Li, C. W. (2007). Evaluating intertwined effects in e-learning programs: A novel hybrid MCDM model based on factor analysis and DEMATEL. Expert Systems with Applications, 32(4), 1028–1044. Yager, R. R. (1988). On ordered weighted averaging aggregation operators in multi- criteria decision making. IEEE Transactions on Systems, Man and Cybernetics, 18(1), 183–190. ME-OWA based DEMATEL reliability apportionment method Introduction ME-OWA operators and its operations ME-OWA operators Determination of ME-OWA weights DEMATEL methodology Outline of the DEMATEL method The procedure of the DEMATEL method Conventional reliability allocation methods ARINC apportionment technique ME-OWA apportionment method Proposed ME-OWA-based DEMATEL apportionment method Advantages of the ME-OWA-based DEMATEL apportionment method Procedures of the ME-OWA-based DEMATEL apportionment method A case study of the ME-OWA-based DEMATEL apportionment method ARINC apportionment technique analysis ME-OWA method analysis Proposed ME-OWA-based DEMATEL apportionment method analysis Method comparison Conclusion References 
work_7dajei6fxzbuvas6pmzkzxmnxu	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 221024247 Params is empty 221024247 exception Params is empty 2021/04/06-03:06:14 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221024247 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:06:14 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_7deiyma4obegjesw7ewjdzdlmy	----	Extended decision template presentation for combining classifiers Expert Systems with Applications 38 (2011) 8414–8418 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Extended decision template presentation for combining classifiers Mehdi Salkhordeh Haghighi ⇑, Abedin Vahedian 1, Hadi Sadoghi Yazdi 1 Department of Computer Engineering, Ferdowsi University of Mashhad, Mashhad, Iran a r t i c l e i n f o a b s t r a c t Keywords: Decision template Classifiers fusion Classifiers combiner 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.01.036 ⇑ Corresponding author. Tel.: +98 511 9151135801 E-mail addresses: haghighi@ieee.org (M.S. Hagh Vahedian), sadoghi@sttu.ac.ir (H.S. Yazdi). 1 Tel./fax: +98 511 8763306. 2 Multiple Classifier System. In this paper, a new method in classifier fusion is introduced for decision making based on internal struc- ture of base classifiers. Amongst methods used in combining classifiers, there are some methods which work on decision template as a tool for modeling behavior of base classifiers in order to label data. This tool models their behavior only based on their final outputs. Our new method, introduces a special struc- ture for decision template such that internal behavior of a neural network base classifier can be modeled in a proper manner suitable for classifiers fusion. The new method builds decision template for each layer of the neural network including all hidden layers. Therefore, the process of making decision in each base classifier is also available for classifiers fusion. Efficiency of the new method is compared with some known benchmark datasets to show how it can improve efficiency of classifiers fusion. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Classification is one of the most frequently encountered deci- sion making processes in a wide range of applications. A classifica- tion problem occurs when an object needs to be assigned into a predefined group or class based on various properties or features. Many problems in business, science, industry, military, security and medicine can be treated as classification problems. Many classification techniques have been used in various appli- cations over the last decade. Moreover, many optimization tech- niques have also been introduced to enhance efficiency and accuracy of classification. Among these, ensemble techniques or MCSs2 also named classifiers fusion or combining classifiers have special interest. As a general rule, the combinational methods have more power, robustness, resistance, accuracy and generality rather than single classification (Analoui, 2008). The motivation for this procedure is based on the intuitive idea that by combining the out- puts of several individual predictors one might improve performance of a single generic one (Krogh & Vedelsdy, 1995). However, the idea of performance improvement in MCS has been proved to be true only when the combined base classifiers are accurate and diverse enough, which requires an adequate tradeoff between these two conflicting conditions (Navone, Gran- itto, Verdes, & Ceccatto, 2001). Also, Kuncheva (2005) using Con- dorcet Jury theorem (Shapley & Grofman, 1984) has shown that combination of classifiers can usually operate better than a single ll rights reserved. ; fax: +98 511 8763306. ighi), vahedian@um.ac.ir (A. classifier. It means that if classifiers with more diversity are used in the ensemble, then total error can considerably be reduced. Drucker, Schapire, and Simard (1993) attempted to gain a good compromise between these properties including elaborations of bagging (Breiman, 1996), boosting (Freund & Schapire, 1995) and stacking (Wolpert, 1992) techniques. In general, three types of strategies in combining classifiers can be identified (Xu, Krzyzak, & Suen, 1992). In the first type each classifier produces output as a single class label such that these la- bels have to be combined to make final decision (Battiti & Colla, 1994). In the second type outputs of the classifiers are sets of class labels ranked in the order of likelihood (Tumer & Ghosh, 1995) and the third type involves the combination of real valued outputs for each class by the respective classifiers, most often posterior prob- abilities (Jacobs, 1995), sometimes evidences (Rogova, 1994). Nevertheless, there are generally two types of combination named classifier selection and classifier fusion (Woods, Kegelmey- er, & Bowyer, 1997). The presumption in classifier selection is that each classifier is an expert in some local area of the feature space. When a feature vector is submitted to each classifier, the classifier designated for the vicinity of x is given the highest credit in assign- ing the class label to x. Therefore, exactly one classifier is responsi- ble to make the final decision, as in Ng and Abramson (1992), or more than one base classifier, as in Jacobs, Jordan, Nowlan, and Hinton (1991), Alpaydin and Jordan (1996). Classifier fusion as- sumes that all classifiers are trained over the whole input feature space, and are thereby considered as competitive rather than com- plementary (Rastrigin & Erenstein, 1982; Xu et al., 1992). However, between all the methods introduced for combining classifiers, DT based methods have special features suitable for wide range of applications. DT method models behavior of base classifiers for making decision on input data. This modeling http://dx.doi.org/10.1016/j.eswa.2011.01.036 mailto:haghighi@ieee.org mailto:vahedian@um.ac.ir mailto:sadoghi@sttu.ac.ir http://dx.doi.org/10.1016/j.eswa.2011.01.036 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. Construction of DT array for Class I data. Fig. 2. Basic DP structure. M.S. Haghighi et al. / Expert Systems with Applications 38 (2011) 8414–8418 8415 scheme is constructed based on final outputs of base classifiers. Operation of the method is based on a set of DT matrices. DTs make a robust classifier fusion scheme that combines classifier outputs by comparing them with a characteristic template for each class. DT based fusion uses all classifier outputs to calculate the final sup- port for each class, which is in sharp contrast to most other fusion methods which use only the support for that particular class to make their decision. DT based methods have been using in wide range of applica- tions for many years. In our research, DT structure is used for mod- eling behavior of base classifiers in a MCS not only based on their output results but also based on internal processing and partial paths taken for making final decision. Using this method to deeply describe partial decisions made inside the structure of base classi- fiers can make classifier combiner operation more robust and effi- cient. The motivation for these properties lies in the fact that the classifiers combiner has more partially processed data with more options to make decisions. 2. Problem definition As mentioned, the primary goal of a MCS is to improve overall efficiency of base classifiers in making decisions over input data. One of the methods widely used in MCSs is DT method. Basic DT method constructs decision template array for each class of train- ing data only based on final outputs of the base classifiers. There- fore, the combiner uses these arrays to make final decisions for test data. Basically, for the combiner to have more primary tools for making decisions, a new structure for constructing DT is pro- posed. The DT is constructed not only based on final outputs of base classifiers, but also based on internal decision paths each base classifier follows to make final decision. In the next section, more details about the new DT are discussed. 2.1. Preliminaries One of the aspects that deeply affect performance of the MCS is that the behavior of base classifiers is modeled in a proper manner such that in almost all input space, the behavior is completely known. However, since such exact model is hard to find, the base classifiers should be trained such that each of them can model part of the input space more efficiently. Moreover, this type of modeling would be more effective when internal path of decision making in each base classifier is also modeled. Therefore, the primary goal of this research is to focus on inter- nal behavior of base classifiers for training and testing data such that the combiner part of MCS posses more effective tools to make decisions for labeling input data. This information is used to im- prove overall generalization capability of the system. For this to happen, neural network based classifiers are used as base classifi- ers. In addition, different strategies for final combiner are tested to find a suitable test bed for comparison. 2.1.1. Standard DT method Generally, in standard DT method, DT is constructed for each class of training data. Fig. 1 shows structure and process of making a standard DT. In this case, to build a DT for class I training data, DP is constructed for each data in this class, according to Fig. 2. Next, average value of these DPs is called DT for class I. The same proce- dure is done for all other classes of training data. The idea behind this DT is to remember the most typical DP for each class. In the test phase, for a test data, DP is constructed, and then compared to each of the DTs based on some similarity measure. Final decision would be the closest match to label the test data. Fig. 3. Structure of the new system. Fig. 4. Structure of a NN based classifier. 8416 M.S. Haghighi et al. / Expert Systems with Applications 38 (2011) 8414–8418 The process exhibits some weaknesses. For instance, final deci- sion is made based on some distance measure between DP of input test data and each one of DTs. These distance measures compare distances between the DP and each DT according to two represen- tative points. Using a point as a representative for a group of data does not yield the required accuracy and robustness while it can- not either describe properties of the data in a proper way. Fig. 5. DP structure used in the new system. 3 Degree of Support. 2.1.2. Fusion system structure Fig. 3 shows structure of our classifiers fusion system from a general view of operation. In this figure, X is a K featured input data, B is the number of base classifiers, and C is the number of classes. The base classifiers used in the system are neural net- works, all of which with the same internal structure, as shown in Fig. 4. This means that the number of hidden layers and the num- ber of neurons in each layer for all the base classifiers are the same. In order to keep diversity, operation of the base classifiers should be different in input space by using different training data. During test step, input vector X is used as input data and DPs for each layer of base classifiers are formed. Since the number of hid- den layers and the number of neurons in each layer for the base classifiers are the same, for a hidden layer with N neurons, DP would be a B � N matrix. However, by using the special structure designed for DP and DT, it would be possible to use base classifiers with different number of neurons in their corresponding hidden layers. It is clear that DTs are formed during base classifiers train- ing. At the end, the combiner makes final decision based on the DTs and DPs formed. Output of the combiner would be a label for the input data. Different strategies for combining these DTs and DPs can be used. In standard DT method, some distance measures are used to determine similarity between each DT and DP. The combining mechanism used in the MCS is based on Euclid- ean distance measure. To do this, in the test step, after forming DPs for all the layers based on X as input data, the value of DOS3 for class m in layer k is calculated by Eq. (1). In the equation, DTK,m is the decision template for class m in layer k, Nk is the number of neu- rons in layer k; DPk is the decision profile for layer k, C is the total number of classifiers. After computing DOS for each class in each layer, maximum DOS is determined. Finally, voting determines final decision among the maximum values of all layers. DOSm;kðXÞ¼ 1 � PB i¼1 PNk j¼1ðDT k;m i;j � DP k i;jÞ 2 B � Nk ð1Þ As a general rule, diversity is a key element in efficiency of MCS. By using a number of different classifiers in a MCS, increase in accuracy and efficiency of the overall system is expected. It is intuitively ac- Fig. 6. Structure of the new DT. Table 1 Properties of some standard datasets. # of samples # of features # of classes Address OCR 1435 64 10 Haghighi@ieee.org Breast 699 9 2 http://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/ Iris 150 4 3 http://archive.ics.uci.edu/ml/machine-learning-databases/iris/ Glass 214 9 6 http://archive.ics.uci.edu/ml/machine-learning-databases/glass/ Wine 178 13 3 http://archive.ics.uci.edu/ml/machine-learning-databases/wine/ Table 2 Error rate obtained by test data. VT DS DTED DTSD SM MAX PT MIN The new method OCR 3.5 3.75 3.75 3.5 3.5 4.75 4.0 5.75 3.15 Breast 3.33 3.33 3.33 3.33 3.33 3.83 3.33 3.83 3.17 Iris 3.57 3.57 2.86 2.86 3.57 3.57 3.57 3.57 2.86 Glass 25 20.67 20.67 26 25 26 35 21.5 19.33 Wine 6.87 6.87 7.5 6.87 6.87 6.87 6.87 6.87 6.87 M.S. Haghighi et al. / Expert Systems with Applications 38 (2011) 8414–8418 8417 cepted that classifiers to be combined should be diverse. If they were identical, no improvements would result in combining them. Therefore, diversity among the team has been recognized as a key point. Since the main reason for combining classifiers is to improve their performance, there is clearly no advantage to be gained from an ensemble that is composed of a set of identical classifiers or clas- sifiers that show the same patterns of generalizations. Therefore, different initial values and training sets with different training parameters are used for the base classifiers. As a result, different behavior is expected from these classifiers. In the new method introduced here, not only a new structure for DT is proposed, but also a new structure for DP is used to match with the new DT structure in the combiner. The criteria for com- bining is provided in Eq. (2) in which L is the number of layers. DOSmðXÞ¼ maxðDOSm;k; k ¼ 1 . . . LÞ; LabelðXÞ¼ argmaxðDOSiðXÞ; i ¼ 1 . . . CÞ ð2Þ Nevertheless, in our new method, DPs are constructed for each of the hidden and output layers of the base classifiers. Therefore, during training, special structure is needed for DPs and DTs such that in test step, the output combiner can easily use them for making decisions. The special DP structure used is shown in Fig. 5. In this figure, for each classifier, outputs of the neurons of each layer are saved as a vector in corresponding cell of DP array. Since each layer may have different number of neurons, each cell of the DP array is designed such that vectors with different length are stored at the same time. This structure gives the DP more flexibility for using in each MCS with different base classifier structures. Fig. 6 shows details about the structure of the new DT. In this figure, each cell of DT array stores a DP like structure which is con- structed based on the DPs in training step. 3. Implementation and results analysis Performance of our new system is compared with other meth- ods based on some known benchmark datasets listed in Table 1. http://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/ http://archive.ics.uci.edu/ml/machine-learning-databases/iris/ http://archive.ics.uci.edu/ml/machine-learning-databases/glass/ http://archive.ics.uci.edu/ml/machine-learning-databases/wine/ 8418 M.S. Haghighi et al. / Expert Systems with Applications 38 (2011) 8414–8418 It should be noted that in the system, the structure of all the base classifiers should be the same. Therefore, the number of hidden layers and the number of neurons in each hidden layer should be identical. As a result, comparison of DTs for each layer is done sim- ilarly. Eq. (2) indicates labeling process using the DOS obtained by Eq. (1). For each of the data sets listed in Table 1, different groups of data are selected randomly as training and test. In the first step, base classifiers are trained such that minimum error for each one is produced. Next, using training data, DTs are built for each layer of base classifiers and for each class of training data. These DTs are stored in the data structure shown in Fig. 6. After the training step, it is time to test performance of MCS with respect to other fusion methods. Performance of the system is measured based on the error rate obtained by test data. As seen in Table 2, comparison has been carried out with these fusion methods: Voting (VT), Dempster Shafer (DS), Decision template and Euclidean distance (DTED), Decision template and symmetric difference (DTSD), simple mean (SM), maximum (MAX), product (PT), and minimum (MIN). 4. Conclusion In the new method presented, fusion of the decisions made by base classifiers are affected by the internal paths each base classi- fier follows to produce final decision. Where the base classifiers are neural networks, the intermediate steps each base classifier fol- lows are presented by outputs of hidden layers. In this work, out- puts of all hidden layers were preserved in a new structure for DT specially designed for this purpose. Therefore, during test step, these preserved data were used for fusion step such that the com- biner could better make final decision to assign a label to the input data. It is expected that the more information available for fusion, the more the decisions are robust. Nevertheless, it is generally ex- pected that efficiency of the combiner increases compared to the methods using only output of the base classifiers for fusion. References Alpaydin, E., & Jordan, M. I. (1996). Local linear perceptrons for classification. IEEE Transactions on Neural Networks, 7(3), 788–792. Analoui (2008). CCHR: Combination of classifiers using heuristic retraining. In International conference on networked computing and advanced information management (NCM 2008), Korea, Sep. Battiti, R., & Colla, A. M. (1994). Democracy in neural nets: Voting schemes for classification. Neural Networks, 7(4), 691–707. Breiman, L. (1996). Bagging predictors. Machine Learning, 24(2), 123–140. Drucker, H., Schapire, R., & Simard, P. (1993). Improving performance in neural networks using a boosting algorithm. In S. J. Hanson, J. D. Cowen, & C. L. Giles, (Eds.), Advances in neural information processing systems (Vol. 5, pp. 42–49). Freund, Y., & Schapire, R. (1995). A decision theoretic generalization of on-line learning and an application to boosting. In Proceedings of the second European conference on computational learning theory (pp. 23–37). Springer Verlag. Jacobs, R. (1995). Method for combining experts’ probability assessments. Neural Computation, 7(5), 867–888. Jacobs, R. A., Jordan, M. I., Nowlan, S. J., & Hinton, G. E. (1991). Adaptive mixtures of local experts. Neural Computation, 3, 79–87. Krogh, A., & Vedelsdy, J. (1995). Neural network ensembles cross validation, and active learning. In G. Tesauro, D. Touretzky, & T. Leen (Eds.). Advances in neural information processing systems (Vol. 7, pp. 231–238). Cambridge, MA: MIT Press. Kuncheva, L. I. (2005). Combining pattern classifiers, methods and algorithms. New York: Wiley. Navone, H. D., Granitto, P. M., Verdes, P. F., & Ceccatto, H. A. (2001). A learning algorithm for neural network ensembles. Inteligencia Artificial, Revista Iberoamericana de Inteligencia Artificial(12), 70–74. Ng, K.-C., & Abramson, B. (1992). Consensus diagnosis: A simulation study. IEEE Transactions on Systems, Man and Cybernetics, 22, 916–928. Rastrigin, L. A., & Erenstein, R. H. (1982). Method of Collective Recognition, Energoizdat, Moscow, (in Russian). Rogova, G. (1994). Combining the results of several neural network classifiers. Neural Networks, 7(5), 777–781. Shapley, L., & Grofman, B. (1984). Optimizing group judgemental accuracy in the presence of interdependencies. Public Choice, 43, 329–343. Tumer, K., & Ghosh, J. (1995). Order statistics combiners for neural classifiers. In Proceedings of the world congress on neural networks (pp. I:31–34). Washington DC: INNS Press. Wolpert, D. (1992). Stacked generalization. Neural Networks, 5, 241–259. Woods, K., Kegelmeyer, W. P., & Bowyer, K. (1997). Combination of multiple classifiers using local accuracy estimates. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19, 405–410. Xu, L., Krzyzak, A., & Suen, C. Y. (1992). Methods of combining multiple classifiers and their applications to handwriting recognition. IEEE Transactions on Systems, Man and Cybernetics, 22, 418–435. Extended decision template presentation for combining classifiers Introduction Problem definition Preliminaries Standard DT method Fusion system structure Implementation and results analysis Conclusion References 
work_7dx2a2i6wvg5rnypgn2ynvyo2a	----	_Supply chain knowledge management: a literature review 1 Supply chain knowledge management: a literature review Marianna Marra, William Ho, John S. Edwards Aston Business School, UK Abstract This paper aims to contribute to the debate on the role of knowledge management in supply chain management by reviewing the published literature. A total of 58 selected referred journal articles were systematically analyzed. This review identifies various theoretical and methodological characteristics of the way in which knowledge management applications are proposed in the supply chain context. The review shows that little evidence exists of the positive relation between the use of IT solutions and firms’ performance. Some issues remain unexplored such as the problem of knowledge obsolescence in supply chain management. A deeper understanding of the knowledge accumulation process could give new insights. The paper concludes with some future directions for theory construction and empirical research. Keywords Supply chain management, Knowledge management, Literature Review. 2 1. Introduction The aim of this paper is to evaluate the relationship between knowledge management and supply chain management through the analysis of the existing theoretical and empirical works, and to contribute to the debate on the role of knowledge management in supply chain management. Lee (2004) pointed out that efficient knowledge flows and knowledge sharing process among supply chain partners give them the following characteristics: agility, adaptability, and alignment. These characteristics allow them to be the best performers. In this paper the results of a structured review of knowledge management literature in supply chain management are presented. Based on the review of 58 journal articles, three issues are examined, (i) To which areas of supply chain management has knowledge management been commonly applied and how? (ii) How can knowledge management be applied in the supply chain management and improve it? (iii) How is knowledge management used to stimulate knowledge creation and sharing across the various supply chain stakeholders? The review of the literature has shown a growing interest in applying knowledge management in supply chains. This seems due primarily to the fragmented nature of such industry sectors and the fragmented knowledge across complex supply chains. Owing to these circumstances the main knowledge management techniques are summarized, and the knowledge management activities most commonly focused on in the supply chain are highlighted. The paper is organized as follows: section two presents the methodology used. Section three deals with the knowledge management activities identified. Sections four, five and six answer the three research questions, respectively. Finally, the key findings of the research are summarised, in the form of knowledge management trends and the main implications for academics and practitioners. 2. Material and method Journal articles, which appeared in the period from 2000 to 2010, were collected. These articles were sourced from the ScienceDirect and ProQuest academic databases. The search was conducted using the key words “supply chain management AND knowledge management”. This very restrictive search criterion has the advantage of avoiding issues of the precise definition of knowledge management and/or supply chain management. It takes the pragmatic definitional viewpoint that an article is about knowledge management and supply chain management if its authors and journal editors think it is. However, the corresponding limitation of this kind of search is that it assumes keywords are perfect and everyone shares the same understanding of what 3 knowledge management and supply chain management mean. The original search yielded 57 papers; to these was added one other paper cited by one or more of the 57 and evidently also addressing knowledge management and supply chain management, according to the citing authors. This gave a final of 58 articles. Despite the intertwined character of the knowledge management stages and activities, an effort will be made to collocate the papers on an imaginary spectrum or continuum. This may be thought in three ways. First, as running from exploitation to exploration (March, 1999), Gray (2001) used the distinction between these two concepts to categorize knowledge management practices, although exploration and exploitation should rather be viewed as the two end-points of a continuum, not as strict categories (Oshri, Pan, & Newell, 2005). Similarly, Brown and Duguid (2001) called for a re- think of the trade-off between the two. The second way is to think of the continuum as running from information and knowledge sharing to knowledge creation and development. This is similar to the first, but by no means identical with it. Thirdly, on the one hand enhancing IT solutions appear to be the most popular means to improve information and knowledge sharing. On the other hand enhancing social aspects such as communities of practices (CoPs) or trust and commitment seems to be useful to improve effective knowledge sharing and knowledge development, following the two “polar” strategies of Hansen, Nohria, and Tierney (1999). Again, this is by no means a perfect correlation with other two ways of looking at the continuum. In the next section, we use the exploitation/exploration version for the main presentation. 3. Knowledge management stages - From exploitation to exploration 3.1. Exploitation In this section, each of the papers will be presented, starting at the exploitation or knowledge sharing/transfer end of the spectrum/continuum. The resulting classification is presented in Table I. Briscoe, Dainty, and Millet (2001) pointed out the knowledge and skills requirements for construction supply chain partnerships. The most important types of skills used by senior executives in small and medium enterprises (SMEs) in the UK were those associated with reading and understanding technical documentation and legal contracts. The next were those linked with the formation and maintenance of main contractor and client relationships. Finally those related to teamwork within the company were highlighted. While IT and computing, as well as financial management, were low down on the rating scale of skills. 4 Corso and Paolucci (2001) investigated the relation between different approaches to knowledge transfer and pattern of adoption of information and communication technology (ICT) application. They also described the economic implications of these alternative approaches. No relation between ICT investments and a firm’s growth was found. Tah and Carr (2001) carried out a study for developing a sharable knowledge-driven approach to risk management. This defined generic risk and remedial action descriptive terms, which can be stored in catalogues. Wu (2001) addressed the problem of coordination among multi-agent systems. Several multi-agent systems for knowledge management were summarized. The issue of coordination problems in supply chain was presented and how to design multi-agent systems to improve information and knowledge sharing was highlighted. Becker and Zirpoli (2003) carried out a research on the theme of knowledge transfer in outsourcing activities. In particular, the focus was on designing an outsourcing strategy to improve knowledge integration. The decomposition strategy for managing dispersed knowledge in outsourcing activities was described through the analysis of a case study of FIAT auto case. Choi, Budny, and Wank (2004) studied intellectual capital in the form of intellectual property management and its strategic importance in corporate success. They referred to a knowledge supply chain and models of licensing relationships. The importance of knowledge as the basis of the transaction in a licensing relationship was highlighted. Holtbrügge and Berg (2004) carried out a study of the knowledge transfer process in German multinational corporations (MNCs). The evidence showed that the source of knowledge (external and internal) and the characteristics of knowledge flows are affected by different firm-specific and country-specific variables, such as the cultural distance between the subsidiary and the home country of the MNCs. Sivakumar and Roy (2004) proposed the concept of knowledge redundancy as a critical factor for supply chain value creation. The concept of redundancy here does not have a negative meaning. On the contrary it deals with there being sufficient overlap of knowledge to give the opportunity to 5 have good communication and, thus, effective operation activities. A conceptual model for managing knowledge redundancy was proposed. Raisinghani and Meade (2005) investigated the links between supply chain, firm’s agility and knowledge management. Their focus was on the strategic decision making perspective. They provided a decision model that supports in determining the best knowledge management construct for an agile supply chain. Douligeris and Tilipakis (2006) carried out a study on the new opportunities provided by the semantic web. They focused their attention on the introduction of web technologies on supply chain management. The use of the semantic web for improving knowledge management and the benefits of supply chain management sector were highlighted. In particular the use of ontologies in improving knowledge management applications was described. Huang and Lin (2010) addressed the problem of managing knowledge heterogeneity in the context of interoperability among multi-entities in a supply chain. They proposed a solution for sharing knowledge using semantic web, while other studies pointed out the use of Web for sharing only information and data. Their solution was based on a semi-structured knowledge model to represent knowledge not only in an explicit and sharable, but also a meaningful format, an agent-based annotation process to resolve issues associated with the heterogeneity of knowledge documents, and an articulation mechanism to improve the effectiveness of interoperability between two heterogeneous ontologies. Koh and Tan (2006) studied the knowledge translation process and proposed the use of a tool for action plan selection (TAPS) as a decision-making tool to enable translation of knowledge of supply chain uncertainty into business strategy and actions. Bandyopadhyay and Pathak (2007) dealt with outsourcing activity as a means to achieving complex knowledge and competencies. Their analysis demonstrated that cooperation plays an important role in enhancing knowledge sharing between employees when there is a high degree of complementarity of knowledge. Thus the authors suggested that the role of top management in outsourcing activities is not only related to negotiating contracts, but also encouraging cooperation between employees. 6 Chow, Choy, and Lee (2007) developed a survey of knowledge management practices in order to examine the applications and technologies adopted in developing a knowledge management system in build-to-order supply chains. Duanmu and Fai (2007) presented a study of the vertical knowledge transfer process to Chinese suppliers by multinational enterprises (MNEs). Types of knowledge transferred and means by which they are transferred were identified. The focus of their analysis was on the relationship created between MNEs and Chinese suppliers. The motivations of the MNEs entering China were efficiency seeking combined with market seeking motives and cost savings. From Chinese suppliers the reasons to be chosen by MNEs were: great security with respect to survival, increased sales, prestige and potential to learn. Fletcher and Polychronakis (2007) developed a framework to capture and disseminate knowledge in the supply chain. Their empirical work was conducted on small and medium enterprises (SMEs). The framework was developed from the previous works and then empirically improved by further fieldwork within a sample of SMEs. The aim of the framework was to enable partners in the supply chain to harness and potentially disseminate skills and knowledge. Joshi, Sarker, and Sarker (2006) examined the factors that affected the knowledge transfer process within an important social unit, the team. Their theoretical model highlighted the importance of the source's capability, credibility, and extent of communication in determining the extent of the knowledge transferred. Otherwise, capabilities did not play an important role. Cheung and Myers (2008) addressed the main problems of sharing knowledge in global strategic networks. They considered factors that contribute to the sustainability of knowledge sharing in global supply chain. These included management fit, market-related fit, resource fit, shared identity, relational capital and flexibility. Myers and Cheung (2008) conducted a study on how knowledge sharing provides value to buyers and suppliers in a global supply chain. The problem of knowledge sharing in such a context is more complex due to cross-cultural differences. The results showed that knowledge sharing is influenced by market structure, and organizational similarities and dissimilarities between buyers and suppliers more than by their needs. While, despite what other literature has claimed (e.g. Ford et al., 2003) cross-cultural differences rarely matter. 7 Wang, Fergusson, Perry, and Antony (2008) adopted a learning perspective, and emphasized the effective knowledge sharing in a supply chain. Their aim was to explore and formulate a model that supports an enterprise with its management of the supply chain members’ knowledge resource sharing. They highlighted the need for mutual learning to increase the competence of supply chain partners, and presented a model for effective knowledge sharing among all partners. Paton and McLaughlin (2008) highlighted the importance of service exchange for innovation. Service exchanges were presented as a determinant for sustainable growth. They focused their attention on the importance of knowledge transfer in SC exchange. In this case the focus was on the role of knowledge centred technological architecture in supporting knowledge workers. Madsen, Riis, and Waehrens (2008) proposed a method for identifying hidden knowledge in outsourcing activities. This consisted of methods for identifying knowledge ties to non-normal task situations, assessing knowledge of various task situations among business partners and a method for transferring of non-normal task situations. Al-Mutawah, Lee, and Cheung (2009) emphasized the importance of integrating information and knowledge flows within the manufacturing supply chain, and highlighted the importance of handling distributed knowledge. A framework based on multi-agent systems was proposed to address the problem of sharing tacit knowledge in the manufacturing supply chain. Blumenberg, Wagner, and Beimborn (2009) demonstrated that the specific knowledge transfer process has a positive impact on outsourcing performance. The mechanisms mentioned were trainings, strategic level agreements (SLAs), and standards. Fugate, Stank, and Mentzer (2009) examined the importance of knowledge management process to overall organizational performance finding that a shared interpretation of knowledge mediates how it is disseminated. A strong positive relationship was found between knowledge management process and operational and organizational performance. The study was framed in the logistic operations context. Pedroso and Nakano (2009) dealt with the complexity and heterogeneity of information and knowledge flows within the supply chain. These were focused on technical information flow. At the 8 same time the importance to facilitate the punctual delivery of information among supply chain members was recognized. The study was framed in the context of the pharmaceutical industry, where punctual technical information delivery is a determinant for the efficiency of the entire process. In the light of this, particular requirements of the technical flows were highlighted. The integrated management of technical information, order information, material and financial flows was required. In addition the use of IT systems and social networks for dissemination was suggested. Corso, Dogan, Mogre and Perego (2010) carried out a study on the knowledge management applications in the supply chain. The study focused on the food industry. They presented a framework for investigating how an IT-based solution for the supply chain fits the knowledge management need of the firms. Pillai and Min (2010) addressed the problem of knowledge calibration. Knowledge calibration deals with the uncertainty of managing incomplete information and with the managers’ confidence in the accuracy of knowledge available. They proposed a conceptual model for a firm’s capability to calibrate knowledge. Xiwei, Blein, and Kan (2010) pointed out the design issue in knowledge supply networks (KSNs). They focused their attention on the role of each member of the KSNs in providing knowledge. In particular, the role of universities and research centres as a key source of technological innovation was highlighted. In addition a risk evaluation method was illustrated. To finish this “end” of the continuum, we review two articles where the three ways of looking at it do not yield the same categorization. Kovacs and Spens (2010) addressed the problem of knowledge sharing in and between relief supply chains. Following Hansen et al. (1999), they suggested the use of CoPs. Khalfan, Kashyap, Li, and Abbott (2010) analyzed knowledge capture and knowledge sharing, showing that these initiatives improve the integration of construction supply chain and the production performance. Again they highlighted the role of CoPs. 3.2. Exploitation and Exploration In the “middle” of our continuum, we review three papers. Corso, Martini, Paolucci and Pellegrini 9 (2001) conducted a literature review on knowledge management in product innovation. Both exploration and exploitation activities were described. They highlighted two main streams in the literature on that topic: one concerned with the scope of the knowledge creation system (single product innovation process, product innovation portfolio, relationship with external actors), the other one dealing with the knowledge management process. Esper, Ellinger, Stank, Flint, and Moon (2010) suggested the integration of the two strategic processes, the demand-focused and the supply-focused, usually separated. They suggested this integration as the basis of a successful value creation through inter-organizational knowledge management. They went on to elaborate a framework for the integration of the two strategies through the knowledge management processes. Halley et al.’s (2010) paper showed how supply chain management and knowledge management fit together. Despite the recognized attitude of the firm to build external networks of collaboration, the natural network of relationships existing in the supply chain itself is presented to be the best network in which knowledge sharing and creation takes place. It is worth noting that this position differs from a prediction of knowledge management theory. We refer to Granovetter’ s argument (1973), which emphasized more the role of weak ties in business relationships than that of the strong and formalized ties. 3.3. Exploration Turning our focus towards exploration, and hence knowledge creation and learning, Choi and Lee (2002) studied the link between knowledge management strategy and the knowledge creation process. Based on a “skewed arc model” the study proposed that companies should align their knowledge strategies along with knowledge creation modes. Ordonez de Pablos (2002) carried out a study on organizational learning and knowledge management in Spanish manufacturing industry from a strategic perspective. The themes of knowledge exploitation and exploration were pointed out. She suggested that knowledge strategy has to be integrated with strategic decision. Spekman, Spear, and Kamauff (2002) analyzed which factors facilitate supply chain learning and whether supply chain performance is improved if learning is fostered. They demonstrated that there 10 is a link between relational variables such as trust, communication, commitment, and performance. Those characteristics seemed to lead to greater collaboration among supply chain partners. Hall and Andriani (2003) highlighted the learning problem associated with the generation of innovation. They carried out a study on the production of new knowledge presenting knowledge management as a useful tool for identifying organizational and technical challenges. Desouza, Chattarai and Kraft (2003) emphasized how supply chain management and knowledge management fit together. Organizations have to learn about the existing knowledge flows, and to identify the high concentration of knowledge around particular functions of the organization. Johanson and Vahlne (2003) outlined a network model of the internationalization process of the firm. They emphasised the importance of learning from experiential relationships to achieve new business relationships that allow firm to enter new country markets, and the importance of commitment in fostering learning. Freeman, Hutchings, Lazaris, and Zyngier (2010) emphasized the role of technological experience in knowledge development instead of the market specific experience and operation experience pointed out by Johanson and Vahlne (2003). They showed that shared technological knowledge allows rapid transfer and development of new knowledge and the drive to commercialize a product before a competitor. Hult, Ketchen, and Slater (2004) analyzed why some supply chains perform better than others, and analyzed how knowledge development process shapes supply chain outcomes. The findings showed that chains possessing more memory tended to seek more knowledge than chains possessing less memory, and knowledge acquisition activities shaped information distribution activities. Piramuthu (2005) proposed an automated supply chain framework. The main aim was reconfiguring the supply chain as per the dictates of order specifications. The paper dealt with the knowledge discovery methods. In particular the framework dealt with how data collected in dynamics supply chain could be used to improve supply chain performance. Batenburg and Rutten (2003) carried out a study of a regional supply chain network, a Dutch knowledge industry cluster. The case study provided useful insight for managing innovation. The 11 importance of the unique contribution of a specific supplier and of creating trust in inter- organizational ties was recognized. Despite the transaction cost theories dominant in supply chain framework, more importance was given to the role of trust. Hult, Ketchen, Cavusgil, and Calantone (2006) adopted a strategic perspective to identify the ideal profile to reach superior performance. The combination of strategy and eight knowledge elements from the literature on the resource-based view are found to be a determinant for reaching superior performance. The eight knowledge elements are: memory, tacitness, accessibility, quality, use, intensity, responsiveness, and learning capacity. Samaddar and Kadiyala (2006) analyzed the conditions under which effective collaboration and knowledge creation take place. Their findings showed that it is important to maintain an optimal ratio between the leader’s and the follower’s marginal gains for the formation of and the maintenance of collaboration. Hult, Ketchen, and Arrfelt (2007) adopted a strategic management perspective to analyze the importance of managing the knowledge development process in the supply chain through a culture of competitiveness. Their results showed that the interaction between a culture of competitiveness and knowledge development has a positive association with performance, but market turbulence may moderate this relation having a positive impact on knowledge development while a negative one on the culture-performance link. Thus if managers are confident about the market turbulence they should emphasize one of the two aspects. Otherwise they have to focus on both. Maqsood, Walker, and Finegan (2007) presented the building of a learning chain instead of a learning organization. The main motivation was the augmented complexity of markets and organizations. The learning chains could be created through a culture of knowledge sharing across the whole supply chain. They emphasized the adoption of a knowledge advantage framework (K- Adv), a knowledge management initiative, which helps creating a culture of knowledge sharing, through which a knowledge advantage for the whole supply chain could be developed. Trust and commitment are the key concepts on which a combined knowledge strategy has to be based. The role of trust in enhancing knowledge sharing was also assumed in the study by Cheng, Yeh, and Tu (2008) of a relief supply chain. They suggested that trust, shared values and participation are positively related with learning capacity. 12 Chen, Kang, Xing, Lee, and Tong (2008) proposed a model based on analytic network process (ANP) and sensitivity analysis to cope with the problem of new product development (NPD) mix selection. In a second stage strategic NPD mix selection and KM methods were integrated. A balanced scorecard evaluates the NPD mix selected. Then KM methods were used to ensure the successful execution of the NPD strategy and enhancing the knowledge conversion process. Wu (2008) studied the four types of knowledge conversion process of the SECI model (Nonaka and Konno, 1998) in the supply chain to inquire how managers can leverage organizational conditions, technology adoption, supplier relationship management and customer relationship management in knowledge creation in a supply chain. Their results showed that each factor could play an important role in the different phases of the knowledge conversion process (SECI). Yeh (2008) examined the knowledge intensive procurements projects. In particular they described the knowledge conversion-process and value creation in order to quantify the success of each project. Craighead, Tomas, Hult, and Ketchen (2009) adopted an economics perspective to measure the impact of knowledge development capacity on supply chain performance. They measured the effects of innovation-cost strategy in the supply chain. They found that knowledge development capacity and intellectual capital efforts are a good complement for other supply chain strategies. Lancioni and Chandran (2009) studied the problem of managing knowledge in an industrial market. The need to achieve shorter NPD cycles and facilitate the learning process is emphasized. They identified intellectual capital and customer relationship management systems as the most critical areas of knowledge management in order to foster exploitation of knowledge and organizational learning. Lau, Ho, Zaho, and Chung (2009) analyzed a process mining system for supporting knowledge discovery in daily logistic operations. The system was applied to a company case and the results showed that it was capable of extracting high-quality and actionable information. Niemi, Huiskonen, and Karkkainen (2009) pointed out the process of knowledge accumulation. They presented it as an ongoing process where the adoption of organizational processes and 13 inventory techniques takes place gradually. They also took into account the organizational aspect that need to be considered to support inventory management techniques. The model, tested in two case studies, was found useful in assessing the current situation on inventory management practices, identifying the development focus areas and prioritizing the development effort. Lopez and Eldrige (2010) presented a working prototype to promote creation and control of knowledge supply chain. In particular their paper dealt with the dissemination of best practices among supply chain practitioners. A diagnosis module was designed and incorporated in a multi- user collaborative working prototype to examine user specified practices and to report a feedback to the user regarding the impact of these practices. The aim of Niemi, Huiskonen, and Karkainen (2010) was to evaluate the adoption of complex practices of supply chain management. In order to do this they used the knowledge maturity model and strategies of accelerating knowledge creation as theoretical frameworks. Through the analysis of two case studies the main strategies adopted were evaluated to conclude that their selection depended on the cultural and organizational environment. 14 Table I – Exploitation and exploration: a continuum Exploitation Exploitation/Exploration Exploration Briscoe et al. (2001) Corso and Paolucci (2001) Tah and Carr (2001) Wu (2001) Becker and Zirpoli (2003) Choi et al. (2004) Holtbrügge and Berg (2004) Sivakumar and Roy (2004) Raisinghani and Meade (2005) Douligeris and Tilipakis (2006) Joshi et al. (2006) Koh and Tan (2006) Bandyopdhyay and Pathak (2007) Chow et al. (2007) Duanmu and Fai (2007) Fletcher and Polychronakis (2007) Cheung and Myers (2008) Myers and Cheung (2008) Wang et al. (2008) Paton and McLaughlin (2008) Madsen et al. (2008) Al-Mutawah et al. (2009) Blumenberg et al. (2009) Fugate et al. (2009) Pedroso and Nakano (2009) Corso et al. (2010) Huang and Lin (2010) Khalfan et al. (2010) Kovacs and Spens (2010) Pillai and Min (2010) Xiwei et al. (2010) Corso et al. (2001) Esper et al. (2010) Halley et al. (2010) Choi and Lee (2002) Ordonez de Pablos (2002) Spekman et al. (2002) Batenburg and Rutten (2003) Hall and Andriani (2003) Desouza et al. (2003) Johanson and Vahlne (2003) Hult et al. (2004) Piramuthu (2005) Hult et al. (2006) Samaddar and Kadiyala (2006) Hult et al. (2007) Maqsood et al. (2007) Chen et al. (2008) Cheng et al. (2008) Wu (2008) Yeh (2008) Craighead et al. (2009) Lancioni and Chandran (2009) Lau et al. (2009) Niemi et al. (2009) Freeman et al. (2010) Lopez and Eldridge (2010) Niemi et al. (2010) 15 4. Results 4.1 To which areas of supply chain management has knowledge management been commonly applied and how? According to table II, the most common areas appear to be outsourcing, the construction industry, decision support, NPD, and risk management. Outsourcing activities appear to be the most popular area in which knowledge management can be applied to the supply chain. Table II-Specific SCM areas Specific SCM areas References Count Outsourcing Becker and Zirpoli (2003) Bandyopadhyay and Pathak (2007) Madsen et al. (2008) Blumenberg et al. (2009) Niemi et al. (2010) 5 New product development Corso et al. (2001) Corso and Paolucci (2001) Becker and Zirpoli (2003) Chen et al. (2008) 4 Construction Briscoe et al. (2001) Tah and Carr (2001) Khalfan et al. (2010) 3 Decision support Raisinghani and Meade (2005) Koh and Tan (2006) Pedroso and Nakano (2009) 3 Risk Management Tah and Carr (2001) Xiwei et al. (2010) 2 Build-to-Order Chow et al. (2007) 1 Procurement Yeh (2008) 1 Organizational Performance Fugate et al. (2009) 1 4.1.1 Outsourcing Outsourcing activities feature in five journal articles. Three of them (Bandyopadhyay and Pathak, 2007; Blumenberg et al., 2009; Niemi et al., 2010) showed an emphasis on cooperation and trust to 16 enhance knowledge management processes among business partners. Becker and Zirpoli (2003) pointed out the risk of hollowing out the knowledge base and suggested counter-balancing measures, such as the creation of communication structures, in order to reduce the dispersion of knowledge in the long term. Madsen et al. (2008) highlighted the importance of a method for identifying hidden knowledge in outsourcing activities. 4.1.2 New product development NPD is another area in which firms have to cooperate in order to share knowledge, and knowledge management activities have to be developed. The literature review conducted by Corso et al. (2001) on knowledge management in product innovation described NPD as a continuous learning process rather than a sporadic event and showed it as one of the most promising areas where knowledge management could be applied. Corso and Paolucci (2001) included the NPD as a part of their analysis on the relationship between the knowledge transfer processes and ICT applications. Becker and Zirpoli (2003) included the product innovation process as part of their investigation into outsourcing in FIAT auto. Chen et al.’s (2008) highlighted the use of KM methods such as informal meeting, experience workshops and expert interviews to ensure the NPD process. 4.1.3 Construction One specific sector, the construction industry, features in three papers (Briscoe et al., 2001; Tah and Carr, 2001; Khalfan et al., 2010). Briscoe et al. (2001) pointed out both technical and relational skills requirements as important factors for construction supply chain partnerships. Tah and Carr (2001) described a framework for project risk management in the construction supply chain. Khalfan et al. (2010) also looked at the construction industry supply chain, and emphasized the positive role of CoPs in improving the integration of construction supply chain. 4.1.4 Decision support Decision support is also a common area in which knowledge management initiatives are applied since knowledge seeking is considered as a means to fill the gap between what decision makers know and what they need to know. Raisinghani and Meade (2005) addressed the need for a strategic decision-making tool to assist management in determining which knowledge management construct is most beneficial in the development of an agile supply chain. Koh and Tan (2006) suggested the application of a decision-making tool, namely TAPS, to enable knowledge translation. Pedroso and Nakano (2009) found that the efficient delivery of the information flows is positively related with an effective decision-making process. 17 4.1.5 Risk management Two papers addressed risk management. As well as Tah and Carr (2001), Xiwei et al. (2010) pointed out the design issue in KSNs and proposed a risk evaluation method based on linguistic operators. 4.1.6 Others Three other application areas were “one-offs”. Chow et al. (2007) studied the knowledge management system in build-to-order supply chains. Yeh (2008) evaluated the knowledge conversion effects for knowledge-intensive procurement projects. Fugate et al. (2009) focused their attention on the importance of knowledge management process to overall organizational performance. 4.2 How can knowledge management be applied in the supply chain management and improve it? Choi et al. (2004) studied intellectual capital and described five models of licensing relationships to extract value from intellectual capital assets. They, also, highlighted the need for integrating intellectual capital licensing with the supply management function. Hult et al. (2004) showed that knowledge acquisition activities, information distribution activities and shared meaning were associated with faster cycle time. Piramuthu (2005) developed a knowledge-based framework for a dynamic re-configuration of supply chains over time. The paper showed performance improvements of the proposed adaptive supply chain configuration framework over static configurations. Hult et al.’s (2006) findings showed that the closer a supply chain matches an ideal profile of knowledge elements and strategy, the better the supply chain’s performance. Paton and McLaughlin (2008) highlighted the role of knowledge centred technological architectures and solutions in supporting knowledge workers. Esper et al. (2010) suggested knowledge management processes as the basis for the integration of the two strategic processes, the demand-focused and the supply-focused, usually separated. Halley, Nollet, Beaulieu, Roy, and Bigras (2010) suggested the management of the relationships existing in the supply chain for sharing and acquiring knowledge, instead of building external business relations. 4.3 How is knowledge management used to stimulate knowledge creation and sharing across the various supply chain stakeholders? Two main approaches to the knowledge management process emerged in the journal articles analyzed. Those who proposed IT solutions as the main basis of every KM activities, and those 18 who, considered the social aspect of knowledge exchange and, to improve it, proposed the improvement of CoPs. This matches the two KM strategies proposed by Hansen et al. (1999). The second approach, focused on the social architecture of knowledge exchange, highlights the importance of trust, cooperation and communication to foster knowledge sharing and learning among actors (Johanson & Vahlne, 2003; Kovacs and Spens, 2010; Khalfan et al., 2010). As emerged in the papers on outsourcing activities the role of trust, cooperation, communication and relational variables is recognized to be a successful factor in knowledge sharing and knowledge creation process (Spekman et al., 2002; Batenburg & Rutten, 2003; Maqsood et al., 2007; Cheng et al., 2008; Niemi et al., 2010). The first approach is based on the extensive use of IT solutions and innovative techniques, such as in the case of the semantic web (Douligeris & Tilipakis, 2006; Huang & Lin, 2010). These kinds of techniques allow information and knowledge sharing as well as the integration of information flow (Pedroso & Nakano, 2009). What is interesting to note is that the choice of one technique or another is a reductionist way of studying the opportunities provided by knowledge management activities. If we look at these through the lens of the continuum adopted in this paper, between exploitation and exploration activities, they both appear a challenge for the firm. As suggested by March (1999) a balance between exploration and exploitation is the challenge of every firm. Hansen et al. (1999) also recommended that one KM strategy should be used to support the other, typically in an 80-20 split. 5. Discussion In this paper, 58 journal articles, which appeared in the period from 2000 to 2010, dealing with knowledge management applications in the supply chain, were collected. Some observations on these journal articles are made in the following subsections. 5.1. Theoretical concerns and research methodological issue To develop a better understanding the articles were analyzed to determine if a specific theoretical perspective was apparent. Where existing strategic and economic theories were being used they consisted of theory related to strategic management (the resource-based view and the knowledge based views of the firms). From the economic perspective, the transaction cost economic (TCE) paradigm is used, but in some cases the evidence contradicted the TCE predictions, such as in the case of the study on a regional supply network (Batenburg & Rutten, 2003), where the evidence showed that trust played a more important role than economic convenience in inter-organizational ties. From the research methodological point of view there is a very large use of qualitative methods as shown in Table III, while a minority of papers adopt the quantitative ones. 19 Table III Research Methods 20 5.2. Observations and recommendations The literature review has shown that knowledge management is considered as a tool for supply chain integration. Despite the interest in studying IT solutions for improving knowledge sharing, the evidence of a positive relationship between their use and the success of supply chain integration is weak. The work of Corso et al. (2010) provides evidence on the role of knowledge management and IT-solutions in supply chains, through a case study. It is worth noting that Niemi et al. (2009), following Shapiro (2001), suggested the need of organizational solutions more than technical ones (Becker, 2001; Edwards, Shaw, & Collier, 2005). On the other hand some studies find positive relationships between other aspects of knowledge management and supply chain performance. Craighead et al. (2009) found such a relationship between the impact of knowledge development capacity and supply chain performance from an economic perspective. The intertwined link between knowledge management and supply chain management is pointed out primarily by the recurrent use of the phrase “knowledge supply chain” (KSC) (Choi et al., 2004) and “knowledge supply networks” (KSN) (Xiwei et al., 2010). Both the phrases point out the need for attention to the knowledge flows among actors of the supply chain network. In particular, KSN emphasizes the role of stakeholder such as university and research centres in creating innovation, in accordance with the conception of network structure as the organizational form in which the role of each actor is important in fostering the learning process and innovation (Choi et al., 2004; Myers & Cheung, 2008; Wang et al., 2008; Lopez & Eldrige, 2010; Xiwei, 2010). On the other hand KSC appears to be an intangible and specific resource, which brings a supply chain competitive advantage. It emerges that knowledge management plays an important role in implementing supply chain management: for the integration processes, for improving collaboration, knowledge capture and knowledge organization. There is a lack of studies measuring of the impact of knowledge management practices on the supply chain performance. A rare example on the role of knowledge management in supply chain comes from the Italian food industry (Corso et al., 2010). Corso et al. presented a framework for investigating how an IT-based solution for the supply chain fits the knowledge management need of the firms. It is possible to conclude that an attempt to measure, in a positivist sense, the influence of KM on supply chain performance is needed. It is important to note that the studies conducted on IT solutions are not able to make a correlation between them and firm’s growth. Corso and Paolucci (2001) did not find a relation between IT adoption and firm’s growth. IT solutions do seem useful in information integration (Pedroso & 21 Nakano, 2009). Blumenberg et al. (2009) demonstrated the positive impact of knowledge transfer on outsourcing activities. Managerial strategies are expected always to have an impact. A more integrated combination of technology and management practices is needed. Further research on how knowledge management strategies improve supply chain performance is needed. Despite the emerging service economy contributing to changing the notion of innovation moving the attention toward intangible aspects, such as information and knowledge, only one paper were found on the service industry (Paton & McLaughlin, 2008). It can be summarized that the learning processes, stimulated by knowledge management activities. (Hult et al., 2004; 2006; 2007) have typically been studied in accordance with a strategic management perspective. On the other hand, a more operational perspective characterizes the studies on knowledge sharing and transfer. The knowledge accumulation process, within the supply chain, can be considered an interesting topic to understand more deeply. More research in this direction could provide the researchers with new insights in improving supply chain performance. Moreover, the fragmented nature of complex supply chains and the complex nature of knowledge, highlighted in many papers, could lead to problems of knowledge obsolescence. This is another theme where better understanding is needed. Thus the role of knowledge, within the supply chain, in achieving superior performance at the firm level needs a deeper understanding. 22 References Al-Mutawah, K., Lee, V., & Cheung, Y. (2009). A new multi-agent system framework for tacit knowledge management in manufacturing supply chains. Journal of Intelligent Manufacturing, 20, 593-610. Bandyopadhyay, S., & Pathak, P. (2007). Knowledge sharing and cooperation in outsourcing projects-A game theoretical analysis. Decision Support Systems, 43, 349-358. Batenburg, R., & Rutten, R. (2003). Managing innovation in regional supply networks: a Dutch case of knowledge industry clustering. Supply Chain Management: An International Journal, 8, 263-270. Becker, M.C. (2001). Managing dispersed knowledge, Journal of Management Studies, 38, 1037-1051. Becker, M.C., & Zirpoli, F. (2003). Organizing new product development. Knowledge hollowing-out and knowledge integration - the FIAT Autocase. International Journal of Operations & Production Management, 23, 1033-1061. Blumenberg, S., Wagner, H., & Beimborn, D. (2009). Knowledge transfer processes in IT outsourcing relationships and their impact on shared knowledge and outsourcing performance. International Journal of Information Management, 29, 342-352. Briscoe, G., Dainty, A.R.J., & Millett, S. (2001). Construction supply chain partnerships: skills, knowledge and attitudinal requirements. European Journal of Purchasing & Supply Management, 7, 243–255. Brown, J.S., & Duguid, P. (2001). Knowledge and organization: A social-practice perspective. Organization Science, 12, 198-213. Chen, H.H., Kang, H., Xing, X., Lee, H.I.A., & Tong, Y. (2008). Developing new products with knowledge management methods and process development management in a network. Computer in Industry, 59, 242-253. Cheng, J., Yeh, C., & Tu, C. (2008). Trust and knowledge sharing in green supply chains. Supply Chain Management: An International Journal, 13, 283-295. 23 Cheung, M., & Myers, M.B. (2008). Managing knowledge sharing networks in global supply chain. International Journal of Management & Decision Making, 9, 581- 599. Choi, T.Y., Budny, J., & Wank, N. (2004). Intellectual property management: a knowledge supply chain perspective. Business Horizons, 47, 37-44. Choi, B., & Lee, H. (2002). Knowledge management strategy and its link to knowledge creation process. Expert Systems with Applications, 23, 173–187. Chow, H.K.H., Choy, K.L., & Lee, W.B. (2007). Knowledge management approach in build-to- order supply chains. Industrial Management & Data Systems, 107, 882-919. Corso, M., Dogan, S.F., Mogre, R., & Perego, A. (2010). The role of knowledge management in supply chains, evidence from the Italian food industry. International Journal of Networking and Virtual Organisations, 7, 163-183. Corso, M., Martini A., Paolucci E., & Pellegrini L. (2001). Knowledge management in product innovation: an interpretative review. International Journal of Management Reviews, 3, 341– 35. Corso, M., & Paolucci, E. (2001). Fostering innovation and knowledge transfer in product development through Information Technology. International Journal of Technology Management, 22, 126-148. Craighead, C.W., Tomas, G., Hult, M., & Ketchen, D.J. (2009). The effects of innovation cost strategy, knowledge and action in the supply chain on firm performance. Journal of Operations Management, 27, 405-421. Desouza, K.C., Chattarai, A., & Kraft, G. (2003). Supply chain perspective to knowledge management: research propositions. Journal of Knowledge Management, 7, 129-138. Douligeris, C., & Tilipakis, N., (2006), A knowledge management paradigm in the supply chain, EuroMed Journal of Business, 1, 66-83. Duanmu, J.L., & Fai, F.M. (2007). A processual analysis of knowledge transfer: from foreign MNE’s to Chinese suppliers. International Business Review, 16, 449-473. Edwards, J., Shaw, D., & Collier, P.M. (2005). Knowledge management systems: finding a way with technology. Journal of Knowledge Management, 9, 113-125. Esper, T.L., Ellinger, A., Stank, T., Flint, D., & Moon, M. (2010). Demand and supply integration: a conceptual framework of value creation through knowledge management. Academy of Marketing Science, 38, 5-18. Fletcher, L., & Polychronakis, Y.E. (2007). Capturing knowledge management in the supply chain. Euromed Journal of Business, 2, 191-207. 24 Ford, D.P., Connelly, C.E., & Meister, D.B. (2003). Information systems research and Hofstadter’s Culture Consequences; An uneasy and incomplete partnerships. IEEE Transactions on Engineering Management, 50, 8-25. Freeman, S., Hutchings, K., Lazaris, M., & Zyngier, S. (2010). A model of rapid knowledge development: The smaller born-global firm. International Business Review, 19, 70-84. Fugate, B.S., Stank, T.P., & Mentzer, J.T. (2009). Linking improved knowledge management to operational and organizational performance. Journal of Operation Management, 27, 247- 264. Granovetter, M.S. (1973). The strength of weak ties. American Journal of Sociology, 78, 1360- 1380. Gray, P.H. (2001). A problem-solving perspective on knowledge management practices. Decision Support Systems, 31, 87-102. Hall, R., & Andriani, R. (2003). Managing knowledge associated with innovation. Journal of Business Research, 56, 145-152. Halley, A., Nollet, J., Beaulieu, M., Roy, J., & Bigras, Y. (2010). The impact of the supply chain on core competencies and knowledge management: direction for future research. International Journal of Technology Management, 49, 297-313. Hansen, M.T., Nohria, N., & Tierney, T. (1999). What’s your strategy fro managing knowledge?. Harvard Business Review, 77, 106-16. Holtbrügge, D., & Berg, N. (2004). Knowledge transfer in multinational corporations, Evidence from Germany firms. Management International Review, 44, 129-146. Huang, C.C., & Lin, S., (2010), Sharing knowledge in a supply chain using the semantic web. Expert Systems with Applications, 37, 3145–316. Hult, G.T.M., Ketchen, D.J., Cavusgil, T., & Calantone, RJ. (2006). Knowledge as a strategic resource in supply chains. Journal of Operations Management, 24, 458-475. Hult, G.T.M., Ketchen, D.J., & Arrfelt, M. (2007). Strategic supply chain management: Improving performance through a culture of competitiveness and knowledge development. Strategic Management Journal, 28, 1035-1052. Hult, G.T.M., Ketchen, D.J., & Slater, S.F. (2004). Information processing, knowledge development, and strategic supply chain performance. Academy of Management Journal, 47, 241-253. Johanson, J., & Vahlne, J.E. (2003). Business relationship learning and commitment in the internationalization process. Journal of International Entrepreneurship, 1, 83-101. 25 Joshi, K.D., Sarker, S., & Sarker, S. (2006). Knowledge transfer within information systems developments teams: examining the role of knowledge sources attributes. Decision Support Systems, 43, 322-335. Khalfan, M.A., Kashyap, M., Li, X., & Abbott, C. (2010). Knowledge management in construction supply chain integration. International Journal of Networking and Virtual Organisations, 7, 207-221. Koh, S.C.L., & Tan, K.H. (2006). Translating knowledge of supply chain uncertainty into business strategy and actions. Journal of Manufacturing Technology Management, 17, 472- 485. Kovacs, G., & Spens, M.K. (2010). Knowledge sharing in relief supply chains. International Journal of Networking and Virtual Organizations, 7, 222-239. Lancioni, R.A. & Chandran, R. (2009). Managing knowledge in industrial markets: New dimensions and challenges. Industrial Marketing Management, 38, 148-151. Lau, H.C.W., Ho, G.T.S., Zhao, Y., & Chung., N.S.H. (2009). Development of process mining system for supporting knowledge discovery in a supply chain network. International Journal of Production Economics, 122, 176-187. Lee, H.L., (2004), The Triple-A Supply Chain. Harvard Business Review, 82, 102-112. Lopez, G., & Eldrige, G. (2010). A working prototype to promote the creation and control of knowledge supply chains. International journal of Networking and Virtual Organizations, 7, 150-162. Madsen, E., Riis, J., & Waehrens, B. (2008). The knowledge dimension of manufacturing transfer: A method for identifying hidden knowledge. Strategic Outsourcing: an International Journal, 1, 198-209. Maqsood, T., Walker, D., & Finegan, A.T. (2007). Extending the "knowledge advantage": creating learning chains. The Learning Organization, 14, 123-141. March, J.C. (1999). The pursuit of organizational intelligence. Blackwell, Malden, Ma. Myers, M.B., & Cheung, M. (2008). Sharing global supply chain knowledge. MIT Sloan Management Review, 49, 67-73. Niemi, P., Huiskonen, J., & Kärkkäinen, H. (2009). Understanding the knowledge accumulation process-Implications for the adoption of inventory management techniques. International Journal of Production Economics, 118, 160-167. Niemi, P., Huiskonen, J., & Karkkainen, H. (2010). Supply chain development as a knowledge development task. International journal of Networking and Virtual Organizations, 7, 132- 149. http://proquest.umi.com/pqdweb?index=18&did=1230576031&SrchMode=1&sid=5&Fmt=3&VInst=PROD&VType=PQD&RQT=309&VName=PQD&TS=1275298671&clientId=10461 http://proquest.umi.com/pqdweb?index=18&did=1230576031&SrchMode=1&sid=5&Fmt=3&VInst=PROD&VType=PQD&RQT=309&VName=PQD&TS=1275298671&clientId=10461 26 Nonaka, I., & Konno, N. (1998). The concept of ‘ba’: building a foundation for knowledge creation. California Management Review, 40, 40-54. Ordonez de Pablos, P. (2002). Knowledge management and organizational learning: typologies of knowledge strategies in the Spanish manufacturing industry from 1995-1999. Journal of Knowledge Management, 6, 52-63. Oshri, I., Pan, S.L., & Newell, S. (2005). Trade-offs between knowledge exploitation and exploration activities. Knowledge Management Research & Practice, 3, 10-23. Paton, R.A., & McLaughlin, S. (2008). Service innovation: knowledge transfer and the supply chain. European Management Journal, 26, 77-83. Pedroso, M.C., & Nakano, D. (2009). Knowledge and information flows in supply chains: A study on pharmaceutical companies. International Journal of Production Economics, 122, 376–384. Pillai, K.G., & Min, S. (2010). A firm's capability to calibrate supply chain knowledge— Antecedents and consequences. Industrial Marketing Management, 39, 1365-1375. Piramuthu, S. (2005). Knowledge-based framework for automated dynamic supply chain configuration. European Journal of Operational Research, 165, 219-230. Raisinghani, M. S., & Meade, L.L. (2005). Strategic decisions in supply-chain intelligence using knowledge management: an analytic-network-process framework. Supply Chain Management: An International Journal, 10, 151-170. Samaddar, S., & Kadiyala, S.S. (2006). An analysis of inter-organizational resource sharing decisions in collaborative knowledge creation. European Journal of Operational Research, 170, 192-210. Shapiro, J. (2001). Modelling the supply chain. Duxbury, Pacific Grove, CA. Sivakumar, K., & Roy, S. (2004). Knowledge redundancy in supply chains: a framework. Supply Chain Management: an International Journal, 9, 241- 249. Spekman, R.E., Spear, J., & Kamauff, J. (2002). Supply chain competency: learning as a key component. Supply Chain Management: An International Journal, 7, 41-55. Tah, J., & Carr, V. (2001). Towards a framework for project risk knowledge management in the construction supply chain. Advanced Engineering Software, 32, 835-846. Wang, C., Fergusson, C., Perry, D., & Antony, J. (2008). A conceptual case-based model for knowledge sharing among supply chain members. Business Process Management Journal, 14, 147-165. Wu, D.J. (2001). Software agents for knowledge management: coordination in multi-agent supply chains and auctions. Expert Systems with Applications, 20, 51-64. 27 Wu, C. (2008). Knowledge creation in a supply chain. Supply Chain Management: an International review, 13, 241-250. Xiwei, W., Blein, M., Kan, W. (2010). Designing knowledge chain networks in China — A proposal for a risk management system using linguistic decision making. Technological Forecasting & Social Change, 77, 902-915. Yeh, H. (2008). A knowledge value creation model for knowledge intensive procurement projects. Journal of Manufacturing Technology Management, 19, 871-892. 
work_7dybzs5z6raivnizas67aqkqye	----	Modeling spatial and temporal dependencies among global stock markets Expert Systems With Applications 43 (2016) 175–185 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Modeling spatial and temporal dependencies among global stock markets Yingliang Weng∗, Pu Gong School of Management, Huazhong University of Science and Technology, Wuhan 430074, China a r t i c l e i n f o Keywords: Spatial dependence Serial correlation Stock returns Monte Carlo simulation a b s t r a c t An intensive analysis of the dependence structure among stock markets is invaluable to financial experts, policy makers, and academic researchers, providing them with important implications for portfolio man- agement, policy-making, and risk assessment. This paper proposes a novel spatiotemporal model to both examine global stock market linkages and investigate what drives stock returns. The newly introduced model allows us to go beyond conventional correlation analyses confined to studying pairwise relationships and seems to be more suitable for detecting the dependence structure of high-dimensional financial time series. Moreover, a new copula-based approach to define the spatial weight matrix is presented that is based on the construction of a dissimilarity matrix using the Spearman’s contagion index. To the best of our knowledge, this paper is the first to incorporate copulas into the definition of the spatial weight matrix. In addition, the maximum likelihood estimator of our model is derived, together with a Monte Carlo simulation study eval- uating its performance compared to two other methods. Finally, the results demonstrate that our proposed measure of the spatial weight matrix, coupled with our model, performs very well in terms of capturing spatial and temporal dependencies among global stock markets, and that the relative values of conditional volatilities are also important factors in determining stock returns. © 2015 Elsevier Ltd. All rights reserved. 1. Introduction The recent global financial crisis (GFC, hereafter), triggered by the collapse of the US mortgage market in August 2007, has shown that global financial markets, especially global stock markets, tend to be more dependent rather than independent on each other during the crisis. Indeed, with the increasing significance of international trades and investments, the spatial interrelationships among global stock markets are becoming increasingly strong. These phenomena indicate that understanding the nature of the dependence structure among global stock markets holds important implications for finan- cial experts, policy makers, and academic researchers. On a practical level, the analysis of the dependence structure among global stock markets can help investors construct optimal asset allocation port- folios and can be beneficial for governments allocate countries’ eco- nomic resources reasonably. The dependence structure between market assets was initially studied using the Pearson correlation coefficient based on the covari- ance of two variables. Nonetheless, the correlation between different market assets might be non-linear and time-varying, which indicates ∗ Corresponding author. E-mail addresses: wyl_hust@163.com (Y. Weng), gongpu@hust.edu.cn (P. Gong). that the hypothesis of linear correlation is not appropriate. Thus, to- wards a better understanding of these relationships, more recently many researchers proposed a number of approaches, which are based on different techniques like multivariate GARCH models (Jaworski & Pitera, 2014), copula-based clustering procedures (Durante & Jaworski, 2010; Durante, Foscolo, & Sabo, 2013; Durante, Foscolo, Jaworski, & Wang, 2014; Durante, Pappadà, & Torelli, 2014; Durante, Fernández-Sánchez, & Pappadà, 2015), and non-parametric clustering methods (Durante, Pappadà, & Torelli, 2015). Besides, there are many other methods to analyze the similarity of different stock markets such as three-phase clustering method (Aghabozorgi & Teh, 2014), clustering method based on topological features (Pereira & de Mello, 2015), multidimensional scaling and clustering analysis (Esmalifalak, Ajirlou, Behrouz, & Esmalifalak, 2015), etc. From a methodological point of view, the dependence can be es- timated and tested in all of these models. It is noticed that most of the previous studies focused on analyzing the dependence structure between one and other stock markets separately rather than simul- taneously by using time series approaches such as GARCH models. In other words, the majority of these models are confined to studying pairwise relationships and there is still much negligence with respect to the estimation of the spatial dependencies among stock markets. Moreover, it is apparent that there might be several nearby stock mar- kets affecting one stock market due to the geographical proximity or http://dx.doi.org/10.1016/j.eswa.2015.09.002 0957-4174/© 2015 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2015.09.002 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.09.002&domain=pdf mailto:wyl_hust@163.com mailto:gongpu@hust.edu.cn http://dx.doi.org/10.1016/j.eswa.2015.09.002 176 Y. Weng, P. Gong / Expert Systems With Applications 43 (2016) 175–185 economic and financial similarities in the real world, that which in- cludes the virtual world. In this context, spatial econometric model has recently emerged as a useful tool to study the spatial interaction effects among geo- graphical units. The spatial econometrics literature has exhibited a growing interest in the specification and estimation of spatial depen- dence models in the last decade. For instance, Baltagi, Bresson, and Pirotte (2012) discussed various forecasts using a panel data model in which the disturbance term follows an error component model with the spatial autoregressive (SAR) and the spatial moving aver- age (SMA) residuals. More recently, Qu and Lee (2015) presented the specification and estimation of an SAR model with an endogenous spatial weight matrix. Wang and Yu (2015) investigated the quasi- maximum likelihood (QML) estimation of spatial panel data models with time varying spatial weights matrices. In addition, the contemporary literature across many disciplines has stressed the importance of incorporating the idea of specification and estimation of serial and spatial dependence from spatial econo- metrics. As stressed by Elhorst (2014), both spatial and serial autocor- relations are likely to be present in the analysis of space-time dataset, so that dynamic rather than static modeling of the economic relation- ships is preferred. In general, the dynamic spatial panel data models often consist of two core components; one is serial correlation be- tween the observations on each spatial unit over time, and the other is spatial dependence between the observations on the spatial units at each point in time. The literature on time-series econometrics is rich with dealing with serial correlation, and the case is true for spa- tial dependence in spatial econometrics. In the spatial literature, two commonly used spatial error processes are the SMA and the SAR er- ror processes (Fingleton, 2008). It is well known that the SMA error process transmit the shocks locally while the SAR error process trans- mit the shocks globally via the power of the spatial weight matrix (Anselin, Gallo, & Jayet, 2008). In an earlier study, Baltagi, Heun Song, Cheol Jung, and Koh (2007) took into account the testing of the spa- tial and serial autocorrelations in a random effects panel data model, where the serial and spatial correlations in the disturbances are sepa- rated. Later, Elhorst (2008) specified a regression model in which the error terms are both serially and spatially correlated jointly. Recently, Lee and Yu (2015) considered a QML estimation of a linear panel data model with serially and spatially correlated disturbances. For a dy- namic panel data model with a spatial error of the SAR type, Su and Yang (2015) derived its QML estimator under both fixed and random effects specifications. Notably, although there are extensive studies on the specification and estimation of dynamic spatial panel models, the applications of such models in economics and finance are not yet very popular, due to the difficulties in constructing the spatial weight matrices in the context of financial markets. Some important recent contributions include that of Fernandez (2011), who employed firm character vari- ables like size and book-to-market to measure the spatial dependence between firms and derived a spatial capital asset pricing model. Eckel, Löffler, Maurer, and Schmidt (2011) measured the effects of geograph- ical distance on stock market correlation via a regression approach. Fernández-Avilés, Montero, and Orlov (2012) and Hierro, Maza, and Villaverde (2013) used spatial techniques to investigate to what ex- tent different linkages between countries affect the degree of stock market co-movements. Among others, for instance, Moscone, Tosetti, and Canepa (2014) examined the relationship between the real estate prices and bank exposures using a first-order dynamic panel data re- gression model with group-specific effects and SAR errors. Arnold, Stahlberg, and Wied (2013) proposed a simplified third-order spa- tial lag model, which has recently been extended by Schmitt, Schäfer, Wied, and Guhr (2015), to model spatial dependence in stock returns and to forecast risk measures. One common conclusion drawn from the above-mentioned two studies is that the combination of spatial modeling and finance allows for superior risk forecasts. The last two papers are closely connected to our approach. However, neither con- siders the intuitive serial dependence of stock returns in the analytic framework. Therefore, in this paper we attempt to ease such restrictions by explicitly illustrating one-to-one spatial linkages among spatial units using spatial techniques. Particularly, this paper investigates spa- tial and temporal dependencies among 24 global major stock mar- kets over the period from 2000 to 2014 using dynamic spatial panel data models. The current paper contributes to the economics and fi- nance literature that studies the dependence structure among finan- cial markets in several aspects. One of the main contributions of this article is that it pays special attention to the use of dynamic spatial panel data models for filtering out the potential spatial and serial correlations among global stock markets. The current paper draws on the methodological contribu- tion by Elhorst (2008), extends the idea to allow for a linear panel data regression model with SMA error disturbances and with first- order serial correlated remainder disturbances, and takes into ac- count whether variously chosen spatial weight matrices have effects on the root mean square error (RMSE) performance of the model. Our main motivation for studying spatial dependence is that currently a country’s stock market is prone to be affected by its nearby coun- tries not only because of geographical proximity (see Eckel et al., 2011; Zhu, Füss, & Rottke, 2013), but also because of economic and financial similarities (see Abate, 2015; Asgharian, Hess, & Liu, 2013; Fernandez, 2011). The main reason for studying temporal correlation is that investors often make decisions in the current period that are influenced by the behavior of their decisions in the previous period and therefore, the current period’s stock returns should be correlated to the previous period’s stock returns (Tam, 2014). Other contributions include a new copula-based approach for the definition of the spatial weight matrix extended from the work by Durante and Foscolo (2013), and an emphasis on the importance of the relative values of the stock return volatilities constructed by the quartile method in the determination of stock returns. To the best of our knowledge, this paper is the first to incorporate copulas into the construction of the spatial weight matrix. The motivation for choos- ing copulas lies in the fact that copulas offer much more flexibility compared to other correlation approach in terms of modeling the de- pendence structure between financial time series. Particularly, this approach is based on the use of suitable copulas associated with the stock markets and on the calculation of the related conditional Spear- man’s correlation coefficients between two stock markets. In addi- tion, although many factors might affect stock returns, such as stock return volatility, industry, social media, investor behavior, and market characteristics (Aghabozorgi & Teh, 2014; Liu, Wu, Li, & Li, 2015), we believe that it is still necessary to propose a new model to reflect the influence of the investor behavior on stock returns from a reference group perspective. A reference group is a group to which an investor or another group is compared. We support the view that stock returns are affected not only by the absolute values of factors like trading vol- umes, volatilities and market caps, but also by the relative values of factors such as ranking of stocks’ trading volumes, ranking of stocks’ volatilities and ranking of stocks’ market caps. In other words, it is reference group or these rankings which form investors’ trading be- haviors that will result in discrepancy between current and desired stock returns. Hence, differently from the previous literature, we use the relative values of the estimated conditional variance, based on the quartile method, to construct the exogenous explanatory variables. One merit of this approach is that we can examine the risk-return tradeoff in view of the fact that the first quartile of the volatilities im- ply high risk stocks while the fourth quartile of the volatilities imply low risk stocks. The paper is divided into five sections. The following section for- mally presents the model specification and the maximum likelihood (ML) estimation of the models. In Section 3, we perform a series of https://isiarticles.com/article/46016 
work_7erdnztlwvcwvicggeo6tps7ui	----	gp_biometrics.dvi HAL Id: hal-00671952 https://hal.archives-ouvertes.fr/hal-00671952 Submitted on 20 Feb 2012 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Genetic Programming for Multibiometrics Romain Giot, Christophe Rosenberger To cite this version: Romain Giot, Christophe Rosenberger. Genetic Programming for Multibiometrics. Expert Systems with Applications, Elsevier, 2012, 39 (2), pp.1837–1847. �10.1016/j.eswa.2011.08.066�. �hal-00671952� https://hal.archives-ouvertes.fr/hal-00671952 https://hal.archives-ouvertes.fr Genetic Programming for Multibiometrics Romain Giot∗, Christophe Rosenberger GREYC Laboratory ENSICAEN - University of Caen - CNRS 6 Boulevard Maréchal Juin 14000 Caen Cedex - France Abstract Biometric systems suffer from some drawbacks: a biometric system can provide in general good performances except with some individuals as its performance depends highly on the quality of the capture... One solution to solve some of these problems is to use multibiometrics where different biometric systems are combined together (multiple captures of the same biometric modality, multiple feature extraction algorithms, multiple biometric modalities. . . ). In this paper, we are interested in score level fusion functions application (i.e., we use a multi- biometric authentication scheme which accept or deny the claimant for using an application). In the state of the art, the weighted sum of scores (which is a linear classifier) and the use of an SVM (which is a non linear classifier) pro- vided by different biometric systems provid one of the best performances. We present a new method based on the use of genetic programming giving similar or better performances (depending on the complexity of the database). We derive a score fusion function by assembling some classical primitives functions (+, ∗, −, ...). We have validated the proposed method on three significant biometric benchmark datasets from the state of the art. Keywords: Multibiometrics, Genetic Programming, Score fusion, Authentication. ∗Corresponding author Email addresses: romain.giot@ensicaen.fr (Romain Giot), christophe.rosenberger@greyc.ensicaen.fr (Christophe Rosenberger) Preprint submitted to Expert Systems with Applications February 20, 2012 1. Introduction1 1.1. Objective2 Every day, new evolutions are brought in the biometric field of research.3 These evolutions include the proposition of new algorithms with better per-4 formances, new approaches (cancelable biometrics, soft biometrics, ...) and5 even new biometric modalities (like finger knuckle recognition [1], for example).6 There are many different biometric modalites, each classified among three main7 families (even if we can find a more precise topology in the literature) :8 • biological : recognition based on the analysis of biological data linked to an9 individual (e.g., DNA analysis [2], the odor [3], the analysis of the blood10 of different physiological signals, as well as heart beat or EEG [4]);11 • behavioural : based on the analysis of an individual behaviour while he is12 performing a specific task (e.g., keystroke dynamics [5], online handwrit-13 ten signature [6], the way of using the mouse of the computer [7], voice14 recognition [8], gait dynamics (way of walking) [9] or way of driving [10]);15 • morphological based on the recognition of different particular physical pat-16 terns, which are, for most people, permanent and unique (e.g., face recog-17 nition [11], fingerprint recognition [12], hand shape recognition [13], or18 blood vessel [14], ...).19 Nevertheless, there will always be some users for which a biometric modality20 (or method applied to this modality) gives bad results, whereas, they are better21 in average. These low performances can be implied by different facts: the quality22 of the capture, the instant of acquisition and the individual itself but they have23 the same implication (impostors can be accepted or user need to authenticate24 themselves several times on the system before being accepted). Multibiometrics25 allow to solve this problem while obtaining better performances (i.e., better26 security by accepting less impostors and better user acceptance by rejecting less27 genuine users) and by expecting that errors of the different modalities are not28 2 correlated. In this paper, we propose a generic approach for multibiometric29 systems.30 We can find different types of biometric multimodalites [15]. They use:31 1. different sensors of the same biometric modality (i.e., capacitive or resistive32 sensors for fingerprint acquisition);33 2. several different representations for the same capture (i.e., use of points34 of interest or texture for face or fingerprint recognition);35 3. different biometric modalities (i.e., face and fingerprint recognition);36 4. different instances of the same modality (i.e., left and right eye for iris37 recognition);38 5. multiple captures (i.e., 25 images per second in a video used for face recog-39 nition);40 6. an hybrid system composed of the association of the previous ones.41 We are interested in the first four cases in this paper. Our objective is to42 automatically generate fusion functions which combine the scores provided by43 different biometric systems in order to obtain the most efficient multibiometrics44 authentication scheme.45 1.2. Background46 1.2.1. Performance Evaluation47 In order to compare different multibiometrics systems, we need to present48 the how to evaluate them. Several works have already done on the evaluation of49 biometric systems [16, 17]. Evaluation is generally realized within three aspects:50 • performance: it has for objective to measure various statistical criteria51 on the performance of the system (Capacity [18], EER, Failure To En-52 roll (FTE), Failure To Acquire (FTA), computation time, ROC curves,53 etc [17]);54 • acceptability: it gives some information on the individuals’ perception,55 opinions and acceptance regarding the system;56 3 • security: it quantifies how well a biometric system (algorithms and de-57 vices) can resist to several types of logical and physical attacks such as58 Denial of Service (DoS) attack.59 In this paper, we are only interested in performance evaluation (because the60 fusion approach is not modality dependant and perception and security depend61 on the used modalities). The main performance metrics are the following ones:62 • FAR (False Acceptance Rate) which represents the ratio of impostors ac-63 cepted by the system;64 • FRR (False Rejection Rate) which represents the ratio of genuine users65 rejected by the system;66 • EER (Error Equal Rate) which is the error rate when the system is con-67 figured in order to obtain a FAR equal to the FRR;68 • ROC (Receiver Operating Characteristic) curve which plots the FRR de-69 pending on the FAR and gives an overall overview of system performance;70 • AUC (Area Under the Curve) which gives the area under the ROC curve.71 In our case, smaller is better. It is a way to globally compare performance72 of different biometric systems.73 We can also present the HTER (Half Total Error Rate) which is the mean74 between the FAR and FRR for a given threshold (this error rate is interesting75 when we cannot get the EER).76 1.2.2. Biometric Fusion77 There are several studies on multibiometrics. The fusion can be operated on78 different points of the mechanism:79 • template fusion: the templates captured by different biometric systems80 are merged together, then the learning process is realized on these new81 templates [19, 20]. Figure 1(a) presents this type of fusion. The fusion82 4 (a) Template fusion. (b) Classical score fusion. (c) Cascade fusion. (d) Hierarchical fusion. Figure 1: Illustration of different fusion mechanisms. process is related to a feature selection in order to determine the most83 significant patterns to minimize errors.84 • decision fusion: the decision is taken for each of the biometric authen-85 tication system, then the final decision is done by fusing the previous86 ones [21].87 • rank fusion: the decision is done with the help of different ranks of bio-88 metric identification systems. The main method is the majority vote [22].89 • score fusion: the fusion is realized considering the output of the classifiers.90 The Figure 1(b) presents this type of fusion.91 Buyssens et al. [23] showed the interest of biometric fusion for face recogni-92 tion combining the image in visible and infrared color spaces with convolutional93 5 neural networks. In [24], Mantalvao and Freire have combined keystroke dynam-94 ics with voice recognition, it seems it is the first time that multibiometrics has95 been done with keystroke dynamics and another biometric modality. In [25],96 Hocquet et al. demonstrated the interest of fusion in keystroke dynamics in97 order to improve the recognition rates: three different keystroke dynamics func-98 tions are used on the same capture. The sum operator (consisting in summing99 the different scores) seems to be the most powerful approach in the literature.100 These fusion architectures are quite simple but powerful. Results can yet be101 improved (in term of error rate or computation time) by using different archi-102 tectures. A cascade fusion [26] is another interesting approach. A first test is103 done, if the user is correctly verified as the attended client or if it is detected104 as an impostor, the algorithm stops. Otherwise, another biometric authentica-105 tion (with another capture from another modality) is proceeded until obtaining106 a decision of acceptance or rejection, or reaching the end of the cascade. So,107 instead of using one decision threshold, each test (except the last one) needs108 two thresholds: one for rejection and one for acceptance. All scores between109 these thresholds are considered in an indecision zone. This mechanism is pre-110 sented in Figure 1(c). Another advantage of this method is to decrease the111 verification time by not using all the modalities, they are used only if necessary.112 This method has been successfully applied on a multibiometric system using113 face and fingerprint recognition in a mobile environment (where acquisition and114 computation times are important) [26].115 Another kind of architecture has been proposed: it is a hierarchical fusion116 scheme [27] (called multiple layers by their authors). Shen et al. have pre-117 sented this method with two different keystroke dynamics methods. The fusion118 is done at different steps, and involves different mathematical operations on119 scores (sum, weighted sum, product, min, max) and logical operations decision120 (comparison to a threshold, or, and) on differents templates extracted from the121 same capture. An extended version to any multibiometric system is presented122 in Figure 1(d). We think our work can be seen as a generalization of this paper.123 124 6 It is also possible to model the distribution of the genuine and impostor125 matching scores, we talk about Density-based score fusion. In [28], scores are126 modelled with a Gaussian Mixture Model and have been tested on three multi-127 biometric databases involving face, fingerprint, iris and speech modalities.128 129 Concerning non linear algorithms, Support Vector Machine (SVM) can also130 be used in a fusion process. Each score to combine is arranged in a vector131 and a training set is used to learn the SVM model. In [29], the SVM fusion132 to improve face recognition gives slightly better performances than weighted133 sum. Voice and online signature have been fused with SVM in [30]. In this134 experiment, arithmetic mean gives best results with noise free data, while SVM135 gives equivalent results with noisy data.136 1.3. Discussion137 In this paper, we are interested in biometric modality independent transformation-138 based score fusion [28] where the matching scores are first normalized and second139 combined. We have previously seen that in this case, arbitrary functions are140 often used. Our work is based on these various fusion architectures based on141 score fusion in order to produce a score fusion function automatically generated142 with genetic programming [31].143 144 By the way, the definition of a fusion architecture is still an open issue145 in the multibiometrics research field [32], because the range of possible fusion146 configurations is very large. We think that using automatically generated fusion147 functions can bring a new solution to solve this kind of problems.148 2. Material and Methods149 In this section, we present all the required information in order to allow150 other researchers to reproduce our experiment.151 7 2.1. Biometric databases152 As it is well known that results can be highly related to the database, for this153 study, we have used three different multibiometric databases: the first one is the154 BSSR1 [33] distributed by the NIST [34] (referenced as BSSR1 in the paper),155 the second one is a database we have created for this purpose (referenced as156 PRIVATE in the paper) and the third one is a subset of scores computed with157 the BANCA [35] database (referenced as BANCA in the text. In fact, BANCA158 database is composed of templates. We have used the scores available in [36]).159 As all these databases are multi-modal, the scores are presented with tuples:160 the ith tuple of scores is represented as si = (s 1 i , s 2 i , ..., s n i ) for a database having161 n modalities (in our case, n ∈ {4, 5}).162 The three databases are presented in detail in the following subsections while163 Table 1 presents a summary of their description.164 2.1.1. BSSR1 database165 The BSSR1 [33] database consists of an ensemble of scores sets from different166 biometric systems. In this study, we are interested in the subset containing167 the scores of two facial recognition systems and the two scores of a fingerprint168 recognition system applied to two different fingers for 512 users. We have 512169 tuples of intra-scores (comparison of the capture of an individual with its model)170 and 512 ∗ 511 = 261, 632 tuples of inter-scores (comparison of the capture of an171 individual with the model of another individual). Each tuple is composed of 4172 scores: s = (s1 bssr1 , s2 bssr1 , s3 bssr1 , s4 bssr1 ), they respectively represent the score of173 the algorithm A of face recognition, the score of algorithm B of face recognition174 (the same face image is used for the two algorithms), the score of the fingerprint175 recognition with left index, the score of fingerprint recognition with right index.176 This database has been used several times in the literature [28, 37].177 2.1.2. PRIVATE database178 The second database is a chimeric one we have created by combining two179 public biometric template databases: the AR [38] for the facial recognition and180 8 the GREYC keystroke [39] for keystroke dynamics.181 182 The AR database is composed of frontal facial images of 126 individuals183 under different facial expression, illumination conditions or occlusions. This is184 a quite difficult database in reason of these specificities. These images have185 been taken during two different sessions with 13 captures per session. The186 GREYC keystroke contains the captures on several session during a two months187 period involving 133 individuals. Users were asked to type the password ”greyc188 laboratory” 6 times on a laptop and 6 times on an USB keyboard by interleaving189 the typings.190 We have selected the first 100 individual of the AR database and we have191 associated each of these individuals to another one in a subset of the GREYC192 keystroke database having 5 sessions of captures. We then used the 10 first193 captures to create the model of each user and the 16 remaining ones to compute194 the intra and inter scores.195 These scores have been computed by using two different methods for the196 face recognition (the scores s1private and s 2 private and three different ones for the197 keystroke dynamics (s3private, s 4 private and s 5 private scores). The face recognition198 algorithms are based on eigenfaces [11] and SIFT keypoints [40] comparisons199 between images from the model and the capture [41]. Keystroke dynamics scores200 have been computed by using different methods [42] based on SVM, statistical201 information and rhythm measures.202 2.1.3. BANCA database203 The lastest used benchmark is a subset of scores produced by the help of204 the BANCA database [36]. The selected scores correspond to the following205 one labelled: IDIAP voice gmm auto scale 25 100 pca.scores for s1banca, SUR-206 REY face nc man scale 100.scores for s2 banca , SURREY face svm man scale 0.13.scores207 for s3banca and208 UC3M voice gmm auto scale 10 100.scores for s4 banca .209 We have empirically chosen this subset. G1 set is used as the learning set,210 9 Table 1: Summary of the different databases used to validate the proposed method Nb of BSSR1 PRIVATE BANCA users 512 100 208 intra tuple 512 1600 467 inter tuple 261632 158400 624 items/tuples 4 5 4 while G2 set is used as the validation set. Users from G1 are different than users211 from G2.212 2.1.4. Discussion213 The main differences between these three benchmarks are:214 • the biometric modalities used in BSSR1 and BANCA have better perfor-215 mances than the ones in PRIVATE;216 • the quantity of intra-scores is more important in PRIVATE (only one tuple217 of intra-score per user in BSSR1 instead of several in PRIVATE);218 • BSSR1 and BANCA are databases of scores (by the way, we do not know219 the biometric systems having generated them) whereas PRIVATE is a220 database of templates (we had to compute the scores);221 • BSSR1 and BANCA are more adapted to physical access control appli-222 cations (i.e., a building is protected by a multi-modal biometric system),223 while PRIVATE is more adapted to logical access control (i.e., the au-224 thentication to a Web service is protected by a multi-modal biometric225 system).226 In the following subsections, we describe the proposed methodology to auto-227 matically generate a score fusion function with genetic programming. We adopt228 the classical score fusion context described in Figure 1(b). Before using the229 scores provided by different biometric systems, we need to normalize them.230 2.2. Score Normalization231 It is necessary to normalize the various scores before operating the fusion pro-232 cess: indeed, these scores come from different classifiers and their values do not233 10 necessarily evolve within the same interval. We have chosen to use the tanh [43]234 operator to normalize the scores of each modality. Equation (1) presents the235 normalization method, where µmgen and σ m gen respectively represents the average236 and standard deviation of the genuine scores of the modality m. The genuine237 scores are obtained by comparing the model and the capture of the same user:238 they are also called the intra scores. In opposition, the inter scores are obtained239 by comparing the model of a user with the capture of other users. score′ and240 score respectively represents the scores after and before normalisation.241 score′ = 1 2 { tanh ( 1 100 ( score − µmgen σmgen ) + 1 } (1) We have selected this normalization procedure from the state of the art242 because it is known to be stable [44] and does not use impostors patterns which243 can be hard or impossible to obtain in a real application. The aim of this244 paper is not to analyse the performance of biometric systems depending on the245 normalization procedure, but to present a new multibiometrics fusion procedure.246 The scores of each modality have been normalized using this procedure.247 2.3. Fusion Procedure248 In this study, we have chosen to use genetic programming [31] in order to249 generate score fusion functions. Genetic programming belongs to the family of250 evolutionary algorithms and its scheme is quite similar to the one of genetic251 algorithms [45]: a population of computer programs (possibly represented by a252 tree) evolves during several generations; different genetic operators are used to253 create the new population. Programs are evaluated by using a fitness function254 which produces a value that is used for their comparisons and gives a probability255 of selection during the tournaments. In a system where the computer programs256 are represented by trees, their leaves mainly represent the entries of the problem,257 the root gives the solution to the problem and the other nodes are the various258 functions taking into arguments the values of their children nodes.259 The leaves are called terminals and can be of several kinds: (a) pseudo-260 variables containing the real entries of the problem (in our case, the list of261 11 scores of each modality), (b) some constants possibly randomly generated, (c)262 functions without any arguments having any side effect, or (d) some ordinary263 variables.264 The different genetic operators usually used during the evolution are (a)265 the crossover, where randomly choose sub-trees have two different trees are266 exchanged, (b) the mutation, where a sub-tree is destroyed and replaced by267 another one randomly generated, or (c) the copy, where the tree is conserved in268 the next generation. The different steps of a genetic programming engine are269 presented as following:270 1. An initial population is randomly generated. This population is composed271 of computer programs using the available functions and terminals. The272 trees are built using a recursive procedure.273 2. The following steps are repeated until the termination criterion is satis-274 fied (the fitness function has reached the right value, or we reached the275 maximum number of generations).276 (a) Computation of the fitness measure of each program (the program-277 ming is evaluated according to its input data).278 (b) Selection of programs with a probability based on their fitness to279 apply them the genetic operations.280 (c) Creation of the new generation of programs by applying the follow-281 ing genetic operations (depending on their probabilities) to the pre-282 viously selected programs:283 • Reproduction: the individual is copied to the new population.284 • Crossover: A new offspring program is created by recombining285 randomly chosen parts from two select programs. An example is286 provided in Figure 2.287 • Mutation: A new offspring program is created by mutating one288 node of the selected program at a randomly chosen place. An289 example is provided in Figure 3.290 3. the single best program of the whole population is designated as the win-291 ner. This can be the solution or an approximate solution to the problem.292 12 A B C D E F G H I J (a) Program source 1 1 2 3 4 5 6 7 8 (b) Program source 2 A B 2 4 5 6 7 8 (c) Program result 1 1 C 3 D E F G H I J (d) Program result 2 Figure 2: Crossover in genetic programming: node C from tree 1 is exchanged with node 2 from tree 2. Program result 1 is the new individual to add to the new generation. 293 13 A B C D E (a) Program source A 1 C 2 3 D E (b) Program result Figure 3: Mutation in genetic programming: node B is replaced by another sub-tree. Different applications to genetic programming are presented in [46] as well294 as their bibliographic references. The fields of these applications can be listed295 in curve fitting, data modelling, symbolic regression, image and signal process-296 ing, economics, industrial process control, medicine, biology, bioinformatics,297 compression... but, it seems, so far of our knowledge, that it has not been298 yet applied to multibiometrics. We only found one reference on genetic pro-299 gramming in the biometrics field. In this paper [47], authors have used genetic300 programming to learn speaker recognition programs. They have used an island301 model where different islands operate their genetic programming evolution, and,302 after each generation some individuals are able to leave to another island. The303 obtained performance was similar to the state of the art in speaker recognition304 in normal conditions, but, the generated systems performed better in degraded305 conditions.306 More information about the configuration of the genetic programming sys-307 tem is presented in the next section.308 2.4. Parameters of the Genetic Programming309 We want to use a score fusion function that returns a score related to the310 performance of a multibiometric system. This score has to be compared with a311 threshold in order to make the decision of acceptance or rejection of the user.312 14 In this case, none logical operation is required in the generated programs and313 different information can be extracted from the result of the fusion function (we314 can compute the ROC curve, the EER, ...).315 2.4.1. Fitness Function316 The EER (Error Equal Rate) is usually used to compare the performance317 of different biometric systems together. A low EER means that FAR and FRR318 are both low and the system has a good performance if its threshold is config-319 ured accordingly to obtain this value. For this reason, we have chosen to use320 this running point to evaluate the performance of the generated score fusion321 functions.322 To compute the EER, we consider the highest and lowest values in the final323 scores generated by the genetic programming. Then, we set a threshold at the324 lowest score and linearly increment it until obtaining the highest score value in325 1000 steps. For each of these steps, we compute the FAR (comparison between326 the threshold and the inter scores) and FRR (comparison between the threshold327 and the intra scores). The ROC curve can be obtained by plotting all these328 couples of (FAR, FRR), while the EER is the mean of FAR and FRR for the329 couple having the lowest absolute difference. So, the fitness function is fitness =330 (FARi + FRRi)/2, where i is the threshold for which abs(FARi − FRRi) is331 minimal.332 2.4.2. Genetic Programming Parameters333 In this section, we present the various parameters used in the genetic pro-334 gramming algorithm. Table 2 presents the various parameters of the evolution-335 ary algorithm.336 To achieve this experiment, we used the PySTEP [48] library. The generated337 programs contain basic functions (+, −, ∗, /, min, max, avg). The terminals338 are the scores of the biometric systems and random constants between 0 and 1.339 The whole fitness cases are completed with a single tree evaluation, thanks to340 the numpy [49] library. Each fitness case is a tuple of scores (where each score341 15 Table 2: Summary of the configuration of the genetic programming iterations. Numbers used in function set can be scores or constants. Configuration Values Objective Generates a function producing a multibiometrics score. Functions set • +: addition of two numbers, • −: subtraction of two numbers, • ∗: multiplication of two num- bers, • /: division of two numbers, • min: returns the minimum of two numbers, • max: returns the maximum of two numbers, • avg: returns the mean of two numbers Fitness function Computes the EER of the multibiometric system Terminal set BSSR1 • a: scores from s1 bssr1 , • b: scores from s2bssr1, • c: scores from s3bssr1, • d: scores from s4 bssr1 , • 50 constants lin- early distributed between 0 and 1. PRIVATE • a, b, c: keystroke dynamics scores (s3private, s 4 private, s5private), • d, e: face recog- nition scores (s1private, s 2 private), • 50 constants lin- early distributed between 0 and 1. BANCA • a: scores from s1banca, • b: scores from s2 banca , • c: scores from s3banca, • d: scores from s4 banca , • 50 constants lin- early distributed between 0 and 1. Initial popula- tion 500 random trees with a depth between 2 and 8 built with the ramped half and half method. Evolution pa- rameters • Number of individuals: 500, • Maximal number of generations: 50, • Depth limited to: 8, • Probability of crossover: 45%, • Probability of mutation: 50% • Probability of reproduction: 5% (with elitism), • Selection: tournament of size 10 with a selection probability of 80%. Termination cri- terion Best individual has a fitness inferior at 0.001 (by the way, this value would never be met . . . ) or maximal number of generations reached. Learning set First half of the intra-scores tuples and first half of the inter-scores tuples. Validating set Second half of the intra-scores tuples and second half of the inter-scores tuples. 16 comes from a different biometric modality) and its result value is the score342 returned by the generated multimodal system. The global fitness value of a tree343 is the EER value computed with the previously generated scores (computation344 of the ROC curve, then reading of the EER value from it).345 PySTEP is a strongly typed genetic programming engine, but, in our case,346 we do not use any particular constraints: the root node can only have a function347 as child (no terminal in order to avoid an unimodal system, and any function of348 the set), while the other function nodes can have any of the functions as children349 as well as any of the terminals.350 The maximal depth of the generated trees is set to 8. In order to avoid351 to stay in a local minimal solution, the mutation probability is set to 50%.352 500 individuals evolve during 50 generations. We have set this few quantities,353 because during our investigations, using a population of 5000 individuals on354 100 generations did not give so much better results (gain not interesting in355 comparison to the computation time). Each database has been splitted in two356 sets of equal size: the first half is the learning set and the second half is the357 validation set.358 The mutation rate is set to 50%, the cross-over rate to 45% and the repro-359 duction rate to 5%. For mutation and cross-over the individuals are selected360 with a tournament of size 10 with a probability of 80% to select the best individ-361 ual. The same individual can be selected several times. For the reproduction,362 the individuals are selected with an elitism scheme: the 5% best individuals are363 copied from generation n − 1 to generation n. During a crossover, only the first364 offspring (of the two generated ones) is kept.365 3. Results366 In this section, we present the results of the generated fusion programs on367 the three benchmark data sets.368 The results are compared to other functions from the state of the art: (a)369 the min rule which returns the minimum score value, (b) the mul rule which370 17 returns the product of all the scores, (c) the sum rule which returns the sum371 of the scores, (c) the weight rule which returns a weighted sum, and (d) an372 SVM implementation. The weighs of the weighted sum have been configured by373 using genetic algorithm on the training sets [50, 51] (in order to give the best374 results as possible). The fitness function is the value of the EER and the genetic375 algorithm engine must lower this value. Table 3 presents the configuration of376 the genetic algorithm.377 Table 3: Configuration of the genetic algorithm to set the weights of the weighted sum Parameter Value Population 5000 Generations 500 Chromosome signification weights of the fu- sion functions Chromosome values interval [−10; 10] Fitness EER on the gen- erated function Selection normalized ge- metric selection (probability of 0.9) Elitism True For the SVM, we have computed the best parameters (i.e., search the C378 and γ parameter giving the lowest error rate) using the learning database on379 a 5-fold cross validation scheme. We have used the easy.py script provided380 with libSVM [52] for this purpose. We have then tested the performance on the381 validation set. We only obtain on functional point (and not a curve) when using382 an SVM. That’s why we have used the HTER instead of the EER.383 Table 4 presents the performances, for the three databases, of each biometric384 systems, fusion mechanisms from the sate of the art, and our contribution.385 Concerning the state of the art performances, can see that the simple fusion386 functions sum and mul tend to give better performances compared to the best387 biometric method of each database, but they are outperform by the weight rule.388 The min operator gives quite bad results (it does not improve the best biometric389 18 system). The SV M method gives good results but is outperform by the weight390 method.391 Table 5 presents the gain of performance against the weight operator (which392 gives the best results in Table 4) in term of EER and AUC.393 This gain is computed as following: gain = 100 (EERweight − EERgpfunc) EERweight (2) where EERweight and EERgpfunc are respectively the EER values of the weighted394 fusion and the generated score fusion function (the same procedure is used for395 the AUC). Better values than the weighted sum are represented in bold. The396 EER gives a local performance for one running point (system configured in or-397 der to obtain an FAR equal to the FRR), while the AUC gives a gives a global398 performance of the whole system. These two information are really interesting399 to use when comparing biometric systems. Figure 4 presents the ROC curves400 of the generated programs against the weighted sum. Performance of the initial401 biometric systems are not represented, because we have already seen that they402 are worst than the weighted sum (same remark for the other fusion functions).403 Logarithmic scales are used, because error rates are quite small.404 We can see from Table 5 and Figure 4 that most of the time, the automati-405 cally generated functions with genetic programming give slightly better results406 than the weighted sum. These improvements can be local and global and vary407 between 16% and 59% for the EER and 0.05% and 76% for the area under408 the curve. When there is no improvement, the results are equal or (in one409 case) slightly inferior. Even if there is some difference between training (not410 represented in this paper) and validating sets, we cannot observe overfitting411 problem. The BSSR1 dataset presents the largest difference of performance412 between training and validation sets, but, the results are still better than the413 ones from the state of the art (and the same problem can be observe with the414 weighted sum). By the way, the fitness criterion has never been met, we did415 not achieve to obtain fusion functions doing no error. So, the evolution always416 19 Table 4: Performance (HTER in %) of the initial methods (s1 ∗ , s2 ∗ , s3 ∗ , s4 ∗ , s5 ∗ ), the state of the art fusion functions (sum, min, mul, weight) and our proposal on the three databases. Bold values represent better performance than the initial biometric systems, and * represents fusion results better than state of the art. (a) BSSR1 Method HTER BSSR1 Biometric systems s 1 bssr1 04.30% s 2 bssr1 06.19% s 3 bssr1 08.41% s 4 bssr1 04.54% Fusion functions sum 00.70% min 05.04% mul 00.70% weight 00.38% SV M 0.77% (FAR=1.16%, FRR=0.39%) Proposal gpI 0.40% (b) PRIVATE Method HTER PRIVATE Biometric systems s 1 private 8.92% s 2 private 11.53% s 3 private 15.69% s 4 private 06.21% s 5 private 31.43% Fusion functions sum 02.70% min 13.72% mul 02.67% weight 02.26% SV M 05.47% (FAR=10.87, FRR= 0.07%) Proposal gpA 01.57%* (c) BANCA Method HTER BANCA Biometric systems s 1 banca 04.38% s 2 banca 11.54% s 3 banca 08.97% s 4 banca 07.32% Fusion functions sum 01.28% min 04.38% mul 01.28% weight 00.91% SV M 01.01% (FAR= 1.71 %, FRR=0.32%) Proposal gpΦ 00.75%* 20 Table 5: Performance gain betwain our proposal and the weighted sum (which gives the best results in the methods of the state of the art). Database EER AUC BSSR1 -5.26% 0.05% PRIVATE 34.85% 23.85% BANCA 17.58% 76.74% ended when reaching the 50th generation.417 Figure 5 represents the fitness evolution during all the generations of one418 genetic programming run on the BSSR1 database. A logarithmic scale has been419 used to give more importance to the low values and track easier the fitness420 evolution of the best individual of each generation. We can observe the same421 kind of results with the other databases. The fitness convergence appears several422 generations before the end of the computation. The worst program of each423 generation is always very bad which implies that the standard deviation of the424 fitness is also always quite huge. This can be explained by the high quantity of425 mutation probability and the low quantity of good programs kept for the next426 generation. When running the experiment several times, we obtain the same427 convergence value. We can say that we reach the maximum performance of the428 system.429 4. Discussion430 The score fusion functions generated by the proposed approach give a slightly431 better performance than the fusion functions used in the state of the art in multi-432 biometrics. We can argue that genetic programming is adapted to automatically433 define score fusion functions returning a score. The tradeoff of this performance434 gain is the need of training patterns which are not necessary for sum, mul or435 min (but this requirement is already present for the weighted sum or the use436 of an SVM). By the way, this is not really a problem, because we already need437 training patterns to configure the threshold of decision (if we do not want to do438 it empirically) or if we need to normalize the scores before doing the fusion.439 Another problem inherent to genetic programming is the complexity of the440 21 generated programs. It is probable that some subtrees could be pruned or sim-441 plified without loosing performance. Another trail would be to add regulariza-442 tion parameter to the fitness function (for example, the number of nodes or the443 depth of the tree). Generated programs would be more readable by an human444 and quicker to interpret. Figure 6 presents a simple generated tree (depend-445 ing on the database, they can be more or less complex). Even if the program446 is quite short (comparing to the other generated functions), it includes useless447 code (e.g., the subtree avg(a, a − 1/12) could be simplified by a − 1/24). Some448 generated trees include preprocessing steps by not using all the modalities in449 the terminal set.450 Genetic programming generated score fusion functions give performance451 slightly equal or better than genetic algorithm configured weighted sum. Even452 if computation time is more important than for genetic algorithm, we can think453 that the gain is not really important between the two methods, but, to obtain454 these results, genetic programming needed a population ten times smaller and455 ten times less of generations.456 5. Conclusion457 We propose in this paper a new approach for multibiometrics based on the458 automatic generation of score fusion functions. We have seen interesting ap-459 proaches in the state of the art and decided to improve them by automatically460 generated score fusion programs by the help of genetic programming.461 Our contribution concerns the designing of multibiometric systems while462 using a generic approach based on genetic programming (and is inspired from the463 state of the art architectures). The proposed method returns a multibiometrics464 score to be compared with a defined threshold. The proposed multibiometric465 system has been heavily tested on three different multibiometric databases. We466 obtained great improvements compared to classical fusion functions used in the467 state of the art. We hope to have opened a new path in the fusion of biometric468 systems thanks to genetic programming.469 22 Results could surely be improved by using different parameters in the genetic470 programming engine (i.e., more individuals and generations, different range of471 constants, different functions, . . . ). It could be interesting to test other perfor-472 mance metrics could be improved by adding quality measures of the capture,473 and if genetic programming could produce template fusion programs.474 6. Acknowledgment475 The authors would like to thank: the author of pySTEP [48], the library476 used during the experiment, for his helpfull help when encoutering problems477 with it, the authors of the various biometric databases used in this experiment,478 as well as the French Basse-Normandie region for its financial support.479 References480 [1] A. Kumar, Y. Zhou, Human Identification Using KnuckleCodes, in: IEEE481 International Conference on Biometrics: Theory, Applications and Systems482 (BTAS 2009), 2009.483 [2] M. Hashiyada, Developement of biometric dna ink for authentication secu-484 rity, Tohoku J. Exp. Med. 204 (2004) 109–117.485 [3] Z. Korotkaya, Biometric person authentication: Odor, Tech. rep., De-486 partment of Information Technology, Laboratory of Applied Mathematics,487 Lappeenranta University of Technology (2003).488 [4] A. Riera, A. Soria-Frisch, M. Caparrini, C. Grau, G. Ruffini, Unobtru-489 sive biometric system based on electroencephalogram analysis, EURASIP490 Journal on Advances in Signal Processing 2008 (2008) 8.491 [5] R. Gaines, W. Lisowski, S. Press, N. Shapiro, Authentication by keystroke492 timing: some preliminary results, Tech. rep., Rand Corporation (1980).493 [6] J. Fierrez, J. Ortega-Garcia, On-line signature, Springer US, 2008, pp.494 189–209.495 23 [7] A. Weiss, A. Ramapanicker, S. Pranav, S. Noble, L. Immohr, Mouse move-496 ments biometric identification: A feasibility study, in: Proceedings of Stu-497 dent/Faculty Research Day, CSIS, Pace University,, 2007.498 [8] D. Petrovska-Delacretaz, A. El Hannani, G. Chollet, Text-independent499 speaker verification: State of the art and challenges, Lecture Notes In Com-500 puter Science 4391 (2007) 135.501 [9] C. Nandini, C. Kumar, Comprehensive framework to gait recognition, In-502 ternational Journal of Biometrics 1 (1) (2008) 129–137.503 [10] K. Benli, R. Duzagac, M. Eskil, Driver recognition using gaussian mixture504 models and decision fusion techniques, in: ISICA 2008, 2008.505 [11] M. Turk, A. Pentland, Face recognition using eigenfaces, in: Proc. IEEE506 Conf. on Computer Vision and Pattern Recognition, Vol. 591, 1991.507 [12] D. Maltoni, A. Jain, S. Prabhakar, Handbook of fingerprint recognition,508 Springer, 2009.509 [13] A. Kumar, D. Zhang, Personal recognition using hand shape and texture,510 IEEE Transactions on Image Processing 15 (8) (2006) 2454.511 [14] Z. Xu, X. Guo, X. Hu, X. Cheng, The blood vessel recognition of ocular512 fundus, in: Proceedings of the 4th International Conference on Machine513 Learning and Cybernetics (ICMLC’05), 2005, pp. 4493–4498.514 [15] A. Ross, K. Nandakumar, A. Jain, Handbook of multibiometrics, Springer,515 2006.516 [16] M. Theofanos, B. Stanton, C. A. Wolfson, Usability & Biometrics: En-517 suring Successful Biometric Systems, National Institute of Standards and518 Technology (NIST), 2008.519 [17] ISO, Biometric performance testing and reporting, Tech. rep., ISO/IEC520 1975-1:2006(E) (2006).521 24 [18] J. Bhatnagar, A. Kumar, On estimating performance indices for biometric522 identification, Pattern Recognition 42 (2009) 1803 – 1815.523 [19] R. Raghavendra, B. Dorizzi, A. Rao, G. Hemantha Kumar, Pso versus524 adaboost for feature selection in multimodal biometrics, in: IEEE 3rd In-525 ternational Conference on Biometrics: Theory, Applications and Systems,526 BTAS 2009, 2009.527 [20] A. Rattani, M. Tistarelli, Robust multi-modal and multi-unit feature level528 fusion of face and iris biometrics, in: International Conference on biometrics529 (ICB2009), 2009.530 [21] A. Ross, A. Jain, Multimodal biometrics: An overview, in: Proceedings531 of 12th European Signal Processing Conference, Citeseer, 2004, pp. 1221–532 1224.533 [22] Y. Zuev, S. Ivanov, The voting as a way to increase the decision reliability,534 Journal of the Franklin Institute 336 (2) (1999) 361–378.535 [23] P. Buyssens, M. Revenu, O. Lepetit, Fusion of ir and visible light modali-536 ties for face recognition, in: IEEE International Conference on Biometrics:537 Theory, Applications and Systems (BTAS 2009), 2009.538 [24] J. Montalvao Filho, E. Freire, Multimodal biometric fusion—joint typist539 (keystroke) and speaker verification, in: Telecommunications Symposium,540 2006 International, 2006, pp. 609–614.541 [25] S. Hocquet, Authentification biométrique adaptative application à la dy-542 namique de frappe et à la signature manuscrite, Ph.D. thesis, Université543 de Tours (2007).544 [26] L. Allano, La biométrie multimodale : stratégies de fusion de scores et545 mesures de dépendance appliquées aux bases de personnes virtuelles, Ph.D.546 thesis, Institut National des Télécommunications (2009).547 25 [27] P. S. Teh, A. B. J. Teoh, C. Tee, T. S. Ong, A multiple layer fusion approach548 on keystroke dynamics, Pattern Analysis & Applications (2009) 14.549 [28] K. Nandakumar, Y. Chen, S. Dass, A. Jain, Likelihood ratio-based bio-550 metric score fusion, IEEE Transactions on Pattern Analysis and Machine551 Intelligence 30 (2) (2008) 342.552 [29] J. Czyz, M. Sadeghi, J. Kittler, L. Vandendorpe, Decision fusion for face553 authentication 7.554 [30] S. Garcia-Salicetti, M. Mellakh, L. Allano, B. Dorizzi, Multimodal bio-555 metric score fusion: the mean rule vs. support vector classifiers, in: Proc.556 EUSIPCO, 2005.557 [31] J. Koza, J. Rice, Genetic programming, Springer, 1992.558 [32] A. Ross, N. Poh, Handbook of Remote Biometrics, Springer, Ch. Multibio-559 metric Systems: Overview, Case Studies, and Open Issues.560 [33] NIST, Nist biometric score set (2006).561 URL http://www.itl.nist.gov/iad/894.03/biometricscores/562 [34] N. I. of Standards, Technology, Nist biometric score set (2006).563 URL http://www.itl.nist.gov/iad/894.03/biometricscores/564 [35] E. Bailly-Bailliere, S. Bengio, F. Bimbot, M. Hamouz, J. Kittler,565 J. Mariéthoz, J. Matas, K. Messer, V. Popovici, F. Porée, et al., The566 BANCA database and evaluation protocol, Lecture Notes in Computer567 Science (2003) 625–638.568 [36] N. Poh, Banca score database.569 URL http://info.ee.surrey.ac.uk/Personal/Norman.Poh/web/570 banca_multi/main.php?bodyfile=entry_page.html571 [37] N. Sedgwick, C. Limited, Preliminary Report on Development and Evalua-572 tion of Multi-Biometric Fusion using the NIST BSSR1 517-Subject Dataset,573 Cambridge Algorithmica Linited.574 26 [38] A. Martinez, R. Benavente, The ar face database, Tech. rep., CVC Techni-575 cal report (1998).576 [39] R. Giot, M. El-Abed, R. Christophe, Greyc keystroke: a benchmark for577 keystroke dynamics biometric systems, in: IEEE International Conference578 on Biometrics: Theory, Applications and Systems (BTAS 2009), 2009.579 [40] D. Lowe, Distinctive image features from scale-invariant keypoints, Inter-580 national journal of computer vision 60 (2) (2004) 91–110.581 [41] C. Rosenberger, L. Brun, Similarity-based matching for face authentication,582 in: Proceedings of the International Conference on Pattern Recognition583 (ICPR’2008), Tampa, Florida, USA, 2008.584 [42] R. Giot, M. El-Abed, C. Rosenberger, Keystroke dynamics with low con-585 straints svm based passphrase enrollment, in: IEEE Third International586 Conference on Biometrics : Theory, Applicationsand Systems (BTAS),587 2009.588 [43] F. Hampel, E. Ronchetti, P. Rousseeuw, W. Stahel, Robust statistics: the589 approach based on influence functions, John Wiley & Sons New York, 1986.590 [44] A. Jain, K. Nandakumar, A. Ross, Score normalization in multimodal591 biometric systems, Pattern Recognition 38 (12) (2005) 2270 – 2285.592 URL http://www.sciencedirect.com/science/article/593 B6V14-4G0DDW4-1/2/d922960ee7ed8928744113dd9494d37a594 [45] M. Mitchell, An introduction to genetic algorithms, The MIT press, 1998.595 [46] R. Poli, W. Langdon, N. McPhee, A field guide to genetic programming,596 Lulu Enterprises Uk Ltd, 2008, freely available at http://www.gp-filed-597 guide.org.uk.598 [47] P. Day, A. K. Nandi, Robust text-independent speaker verification using ge-599 netic programming, IEEE TRANSACTIONS ON AUDIO, SPEECH, AND600 LANGUAGE PROCESSING 15 (2007) 285–295.601 27 [48] M. Khoury, Python strongly typed genetic programming.602 URL http://pystep.sourceforge.net603 [49] T. Oliphant, Guide to NumPy, Spanish Fork, UT, Trelgol Publishing.604 [50] R. Giot, M. El-Abed, C. Rosenberger, Fast learning for multibiometrics sys-605 tems using genetic algorithms, in: The International Conference on High606 Performance Computing & Simulation (HPCS 2010), IEEE Computer So-607 ciety, Caen, France, 2010, p. 8.608 [51] R. Giot, B. Hemery, C. Rosenberger, Low cost and usable multimodal bio-609 metric system based on keystroke dynamicsand 2d face recognition, in:610 IAPR International Conference on Pattern Recognition (ICPR), IAPR, Is-611 tanbul, Turkey, 2010.612 [52] C. Chang, C. Lin, LIBSVM: a library for support vector machines (2001).613 28 (a) Validation with BSSR1 (b) Validation with PRIVATE (c) Validation with BANCA Figure 4: ROC curves of the fusion systems from the state of the art and with genetic programming. The EER of each fusion function is presented in the legend. Note the use of a logarithmic scale. 29 0 10 20 30 40 50 Generation (#) 10 0 10 1 10 2 F it n e ss s c o re M in /A v g /M a x Min Max Mean Std Figure 5: Fitness evolution of one run of the genetic programming evolution. The max, min, mean and std values of the fitness are represented. We want to minimize the fitness value, so lower is better. Figure 6: Sample of a ”simple” generated program. We can observe the complexity of the generated fusion function. 30 
work_7f4opqdmpfd2thi74qiqo37oua	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 221004486 Params is empty 221004486 exception Params is empty 2021/04/06-03:05:51 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221004486 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:51 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_7fesqq4wwncbrg5ce7t5r4apii	----	02 1303.hwp 한국해양공학회지 제26권 제4호, pp 8-14, 2012년 8월 (ISSN 1225-0767) http://dx.doi.org/10.5574/KSOE.2012.26.4.008 휴리스틱 탐색 기법을 이용한 네스 문가 시스템 신동목* *울산 학교 조선해양공학부 Nesting Expert System using Heuristic Search Dongmok Sheen* *School of Naval Architecture and Ocean Engineering, Univerty of Ulsan, Ulsan, Korea KEY WORDS: Nesting 네스 , Expert system 문가 시스템, Pixel 픽셀, Steel cutting 강재 단, Heuristic search 휴리스틱 탐색 ABSTRACT: Two dimensional nesting is a common problem in industries such as the shipbuilding, automotive, clothing, shoe-making, and furniture industries, in which various parts are cut off from stock or packed in a flat space while minimizing waste or unoccupied space. Nesting is known as an NP-complete problem, which has a solution time proportional to the superpolynomial of the input size. It becomes practically impossible to find an optimal solution using algorithmic methods as the number of shapes to nest increases. Therefore, heuristic methods are commonly used to solve nesting problems. This paper presents an expert system that uses a heuristic search method based on an evaluation function for nesting problems, in which parts and stock are represented by pixels. The system is developed in CLIPS, an expert system shell, and is applied to four different kinds of example problems to verify its applicability in practical problems. 교신 자 신동목: 울산 역시 남구 무거동 산 29, 052-259-2152, dmsheen@ulsan.ac.kr 1. 서 론 네스 (Nesting)은 다양한 모양의 여러 형상들을 평 소재에서 잘라내거나 평면 상에 배치할 때 낭비되는 소재 는 공간이 최소화 되도록 형상들을 배치하는 문제라 정의할 수 있다. 조선 소의 강재 단, 자동차 기계부품 공장 등에서 블랭킹 공정, 의 복 공장, 구두공장, 가구공장 등에서의 소재 단 공정에서 형상들 을 배치하는 문제는 표 인 네스 문제로서 가능한 한 공간 으로 빼곡히 배치하면 버려지는 부분을 작게 함으로써 재료 비를 약할 수 있다. 한 조선소 야 장에서의 블록 배치 문 제도 네스 문제로 볼 수 있다. 이러한 네스 문제는 각 형상 별로 자세 배치될 치 등을 찾아야 하는데 NP-complete (Non determimistric polynomial complete) 문제로서(Bennell and Oliveira, 2008) 알고리즘 인 방법으로 최 해를 찾을 수 없 어 통상 으로 휴리스틱 방법을 이용한다. 상업용 CAD 시스템들 에서 사용하고 있는 방법 한 휴리스틱 방법에 기반을 두고 있는 데(한국찬과 나석주, 1994), 부재들 간의 순서를 정한 후 배치할 부재들을 볼록 다각형으로 단순화하여 기존에 배치된 부재에 NFP(Nofit polygon) 알고리즘(Burke et al., 2007) 등을 용하여 집 배치하는 단계 방법을 사용하고 있다. 본 연구에서는 지식기반 추론 시스템인 문가 시스템을 이 용하여 배치할 형상이나 배치될 소재 경계 형상 등에 한 제 약 없이 네스 문제를 해결하고자 하며, 제안된 방법을 이용하여 각각 다른 특성을 갖는 네스 문제에 용하여 그 용 가능 성을 시하고자 한다. 2. 관련 연구 네스 기법은 형상표 방법에 따라 픽셀로 형상을 표 하는 방법과 형상을 근사화한 다각형으로 표 하는 방법으로 나 수 있다. 픽셀 형태를 이용할 경우 표 이 간단하고 형상간 첩여 부를 쉽게 단할 수 있는 반면, 정 한 형상 표 을 해서는 많 은 양의 메모리와 계산시간이 소요된다. 반면 다각형으로 형상을 표 하는 경우 다각형의 경계를 정의하는 과 선들의 정보만을 처리하게 되므로 상 으로 은 메모리를 이용하나, 형상간 첩 여부를 단하는 알고리즘이 복잡해진다. 형상을 표 하기 하여 픽셀형태를 이용하는 경우 형상을 표 하기 한 해상도를 정한 후 형상의 내부 경계선이 걸 쳐지는 픽셀은 ‘1’, 그 외의 공간은 ‘0’을 할당하는 Fig. 1에 시 Fig. 1 Pixel representation 8 휴리스틱 탐색 기법을 이용한 네스 문가 시스템 9 Fig. 2 A variation of pixel representation and overlapping 한 방법이 일반 이나(Weng and Kuo, 2011), 내부와 경계를 다 르게 코딩하여 첩 시 경계 픽셀 정보를 얻거나 우에서 좌로 빈 공간의 픽셀값을 1씩 증가시켜 할당함으로써 첩 시 빈 공 간까지의 상 인 치를 얻는 방법도 있다(Bennell and Oliveira, 2008; Babu and Babu, 2001). Fig. 2와 같이 코딩하여 두 형상을 배치할 경우 두 픽셀의 합이 4 이상인 픽셀은 두 형상이 겹치는 첩 부 를 의미하며 합이 2인 픽셀은 부 , 1 는 0은 첩되지 않는 부 를 의미하게 된다. 픽셀을 이용하여 형상을 표 하는 경우 형상을 배치할 때 BLF(Bottom-left filling) 방법 이 흔히 사용된다(Babu and Babu, 2001; Weng and Kuo, 2011). BLF 방법에서는 평 소재 상에서 빈 곳 좌측 하단 끝 픽셀 을 형상의 기 픽셀이 놓일 치로 선정하여 첩여부를 검사 하고 첩이 될 때는 한 픽셀씩 로 이동하면서 빈 곳을 찾고 최 상단까지 배치할 빈 곳이 없을 때에는 우측의 다음 픽셀 열의 가 장 아래 빈 픽셀부터 검사해 나가는 방법이다. 다각형을 이용하는 방법은 각 형상을 둘러싼 다각형으로 근 사화하며 NFP를 이용하는 방법(Burke et al., 2007)이 표 이 다(Fig. 3 참조). 한 도형(A)의 모서리를 따라 다른 한 도형(B)이 회 없이 움직일 때 도형 B의 한 기 이 움직이는 궤 (NAB) 을 나타내는 NFP를 정의하고 나면 두 다각형의 첩 여부는 B 의 기 이 NFP의 내부에 속하는 지를 단하면 된다. NFP를 구하는 방법은 A의 모서리는 반시계방향으로 정의하고, B의 모 서리는 시계 방향으로 정리한 후 두 도형의 모서리의 합 Fig. 3 Nofit polygon, NAB Fig. 4 Generation of a nofit polygon 집합에 하여 반시계 방향으로 정렬하면 NFP가 구해진다(Fig. 4). 오목한 형상의 경우나 형상에 구멍이 있는 경우 NFP 알고 리즘은 복잡해진다. NFP는 두 형상간의 첩여부를 단하는 효 과 인 도구이나 여러 부재들 간의 배치 순서를 제시하지는 못 하며 따라서 흔히 큰 부재부터 순차 으로 NFP를 용한다. 한편 조선 련 연구(이철수와 박 렬, 1996; 한창 과 박제웅, 1999)에서는 같은 폭을 갖는 부재들을 그룹화 한 후 각 그룹별 로 부재들과 폭이 같은 긴 평 소재에 일렬로 배치하는 문제로 근하 다. 이 경우 일반 인 2차원 네스 문제보다 단순화 할 수 있다. 즉, 폭 방향으로는 하나의 부재만 배치되며 한 부재는 0 o ,180 o 회 형상, 그리고 각각에 하여 뒤집은 형상의 4 가지 자세만 고려하면 된다. 기존 연구(이철수와 박 렬, 1996; 한창 과 박제웅, 1999)에서는 큰 부재부터 순차 으로 연속 배치하는 방 법을 사용하 으며 따라서 단계별로 추가하는 형상에 하여 4 가지 자세 최 의 배치를 찾았다. 이밖에 평 소재 형상이 직사각형이 아닌 비정형 형상인 경 우에 형상들을 배치하는 문제도 주요한 연구주제들로 다루어지고 있다(Tay et al., 2002; Lee et al., 2008; Burke et al., 2007). 3. 네스팅 전문가시스템 3.1 시스템 구성 문가 시스템은 C언어나 Fortran 언어 등을 이용하여 개발 되는 통 로그램과 구별되는 로그램이다. Fig. 5는 표 인 문가 시스템인 규칙기반 시스템(Rule-based system)의 구성을 보여 다. 각 실행 단계에서 확인된 사실들(Facts)로부터 IF<conditions> THEN<action> 형태의 규칙을 순차 으로 용하여 새로운 사 실을 생성해가면서 궁극 으로 원하는 해를 얻는 추론 방법이 다. 실행과정을 살펴보면 역데이터베이스(Global data base, GDB) 에 장된 사실로부터 해석기(Interpreter)는 규칙기반(Rule base) 에 있는 규칙들 용 가능한 규칙들을 찾아내고, 충돌해결기 구(Conflict resolution mechanism)에서 제어기 (Control stra- tegy)을 용하여 하나의 규칙을 선택하며, 선택된 규칙은 추론 엔진(Inference engine)에 의해 실행된다. 표 인 제어기 으 로는 가장 최근에 추가된 사실과 부합하는 규칙을 선택하는 방법 (Recency ordering), 조건부가 복잡한 규칙을 선택하는 방법 (Specificity ordering)등이 있다. 실행 결과에 따라 GDB가 갱신 되고 체 과정이 반복된다. 실행 과정에서 동일한 사실에 기반 Fig. 5 A Schematic diagram of expert system 10 신동목 하여 한 번 실행된 규칙은 다시 실행되지 않으며 해를 얻거나 더 이상 용 가능한 규칙이 없을 때 시스템은 종료된다. 해를 찾는 과정은 탐색 트리(Tree)로 표 할 수 있는데 각 노 드는 단계별 상태를, 각 노드에서 나온 가지들은 각 상태에서 용 가능한 규칙들을 나타낸다. 하나의 규칙이 용되면 새로운 사실이 추가된 새로운 상태가 되며 새로운 가지치기를 할 수 있게 된다. 문제에 한 사 정보가 없을 때 문제를 해결하는 일반 인 탐색 트리는 크게 깊이 방향을 먼 탐색하는 깊이우 선(Depth-first) 탐색 트리와 같은 깊이에서 가능한 경우의 수를 먼 모두 탐색하는 폭우선(Breadth-first) 탐색 트리로 나 수 있다. 문제의 해를 찾는 탐색 방향에 한 정보가 있을 때는 각 노드에서 다음 탐색할 노드들에 하여 해에 근 한 정도와 련된 평가함수(Evaluation function)를 계산하여 그 값이 큰 노 드부터 탐색하는 휴리스틱(Heuristic) 탐색 방법을 사용한다. 본 연구에서는 미 항공우주국(NASA)에서 개발한 문가 시 스템 쉘(Shell)인 CLIPS(http://clipsrules.sourceforge.net/)를 이 용하여 네스 을 한 문가 시스템을 휴리스틱 탐색 방법을 이용하여 구축하며 이를 블록 맞추기 퍼즐, 일반 네스 , 조선 분 야 네스 , 이형 경계 형상 소재에의 배치 문제에 용하여 그 용 가능성을 시하고자 한다. 3.2 데이터 구조 본 논문에서는 네스 문제를 휴리스틱 기법을 이용한 깊이 우선 트리 탐색 문제로 정의한다. Fig. 6은 휴리스틱 탐색 트리를 보여 다. 각 노드에서 다음에 탐색할 아래 계층 노드들은 평가 함수 값에 따라 재정렬되어 평가함수 값이 가장 큰 노드부터 탐색을 한다. 따라서 Fig. 6에서 제일 좌측 가지들은 각 노드에 서 평가함수 값이 가장 큰 노드들이다. 본 연구에서 평 소재 부재들은 픽셀 형태로 표 한다. 흑백 픽셀 데이터는 기본 으로 이진 상(Binary image) 이므로 형상 간 첩을 검증하기 쉬우며 상업용 CAD 시스템과의 연동도 쉽게 할 수 있다는 장 이 있다. Fig. 6에서 시작 노드는 하나의 부재도 배치되지 않 은 평 소재의 상태를 나타내며, 하나의 계층이 내려갈 때마다 하나의 부재가 배치된다. 모든 부재가 다 배치된 계층에 이르면 Fig. 6 A heuristic depth-first search tree 평 소재에는 모든 부재가 다 배치된 상태로서 문제의 해를 나 타낸다. 하나의 노드는 (b111,b211,…)와 같이 정의되는데 여기서 brst는 r번째 부재의 s번째 치, t번째 자세를 의미한다. 하나의 부재는 회 각에 따라 무한한 가짓수의 자세를 가질 수 있으나 본 연구에서는 0 o ,90 o ,180 o ,270 o 회 형상, 그리고 각각에 한 미 러(Mirror) 이미지 등 최 총 8 가지 자매 형상을 고려한다. 회 이나 미러링(Mirroring) 결과 동일한 형상은 생성하지 않 는다. 따라서 각 노드에는 식(1)과 같이 nr개의 가지가 생성된다. 여기서, n(r,i,Br-1)은 재 (r-1)개의 부재가 배치된 평 소재의 상태 Br-1에서 r번째 부재의 i번째 자세에서 배치 가능한 치에 한 경우의 수를 의미한다.       (1) 다음은 CLIPS로 나타낸 노드에 한 데이터 템 리트 (Template)를 나타낸다. Node-id와 Parent 슬롯(Slot)은 자신의 ID와 부모 노드의 ID를, Tile-list와 Tile-orient-list 슬롯은 재 노드까지 배치되어 있는 부재들과 각각의 자세를, Tile-position- row-list와 Tile-position-col-list 슬롯은 각 부재들의 기 치를 행 번호와 열번호 리스트로 나타낸 것이다. 기 은 부재 들을 둘러싼 최소 사각형(Enclosing rectangle, ER)의 좌 상단 픽 셀이다. (Deftemplate MAIN::node (Slot node-id (type NUMBER)) (Slot parent (type NUMBER)) (Multislot tile-list (type FACT-ADDRESS)) (Multislot tile-orient-list (type FACT-ADDRESS)) (Multislot tile-position-row-list (type NUMBER)) (Multislot tile-position-col-list (type NUMBER))) 역 데이터베이스(GDB)에는 노드들에 한 정보 이외에 부 재들의 배치 순서를 장하는 Lay-tile-list, 다음 번에 배치할 확 장할 순서에 따라 노드들을 리스트로 표 한 Open-list가 있다. 3.3 규칙 지식 베이스는 Initialize, Open-node, Goal 등의 규칙과 련 함수들로 구성되어 있다. Initialize 규칙에서는 각 부재를 먼 크기 순서로 정렬한 후 이를 Lay-tile-list에 장한다. 한 각각 의 부재에 하여 8 가지 자세를 생성하며 용 분야에 따라 자 세의 가짓수를 조정한다. 를 들어 의복제조 분야와 같아 소재 의 앞 뒷면을 구별해야 하는 경우 거울 이미지는 고려할 필요가 없고, 조선 블록 생산 시 소요되는 보강재류와 같이 긴 소재에 폭이 같은 부재들을 배치하는 경우 90 o ,270 o 의 회 형상 그에 한 미러(Mirror) 형상은 고려할 필요가 없다. Open-node 규칙은 해에 근 한 노드를 찾아 다음 순서에 따라 탐색트리를 확장한다. Rule: open-node n = first(open-list). where, open-list={n, REST}) ; Get the 휴리스틱 탐색 기법을 이용한 네스 문가 시스템 11 first node from the set open-list and call it n {N} = expand(n) ; Expand the node n and save the result set as {N} {Ns} = usort(N) ; Sort N in descending order of evaluation function value and save it as set {Ns} open-list = {Ns, REST} ; Update the set open-list with Ns front 다음은 CLIPS로 작성된 Open-node 규칙의 조건부를 일부 보 여 다. CLIPS에서는 패턴 매칭 방법에 의하여 규칙의 조건부 와 매칭된 사실들을 찾는다. (defrule MAIN::open-node ?op <- (open-list (nodes ?nid $?rest-n)) ?cn <- (node (node-id ?nid) (parent ?p) (tile-list $?tlist) (tile-orient-list $?torlist) (tile-position-row $?prlist) (tile-position-col $?pclist)) (lay-tile-list $?tlist ?t $?rest-t) 여기서, 가지를 새로 생성할 노드는 Open-list의 Nodes 슬롯에 장된 리스트의 맨 앞에 있는 노드(?nid)이며 이 노드에 한 정보는 GDB에 있는 노드들 패턴이 일치하는 (Node (Node- id ?nid) …)를 찾아 련정보를 가져오게 된다. 재 노드의 Tile-list 슬롯 값인 $?tlist는 재까지 배치된 부재들을 나타내며 이는 Lay-tile-list와 매칭되어 리스트의 요소 ?t가 다음에 배치할 부재를 나타내게 된다. 이 규칙은 탐색 트리에서 한 노드로부터 가지를 확장하는 규칙으로 재 노드에서 가지를 생성할 때 즉, 재까지 배치가 끝난 부재들에 다음 부재를 추가할 때 부재의 8개 자세에 하여 배치 가능한 nr개의 모든 치(식 (1) 참조) 를 좌 상단부터 차례로 찾아서 자식(Children) 가지로 확장한다. 새로 생성된 노드들은 평가함수 u(Br-1,bir**)값이 큰 순서 로 정 렬하여 Open-list에 장함으로써 평가함수 값이 큰 노드부터 탐색이 이루어질 수 있도록 한다. 여기서, Br-1은 재까지 배치된 부재들을 포함한 평 소재의 상태를 나타낸다. 식 (2)는 본 연 구에서 용한 평가함수를 나타낸다. 식에서 count는 픽셀 수를 세는 함수이며, u(Br-1,bir**)는 r 번 째 부재를 bir**에 배치할 때 재 배치가 완료된 (r-1)개의 부재들을 둘러싼 ER들과 부재 r의 ER을 1:1로 비교했을 때 겹쳐지는 픽셀 수들을 합한 것이다.   ∈ ∩ (2) 평가함수는 탐색트리에서 노드의 확장 순서를 결정하는 기 이 되며 함수 값이 클수록 조 하게 형상이 배치됨을 나타내므로 원 하는 해를 도출할 가능성이 높다. 식 (2)의 평가함수는 각 ER의 좌상단과 우하단의 두개의 꼭지 좌표를 이용하여 쉽게 계산할 수 있다. 각 ER의 두 개의 꼭지 의 상 좌표는 8 개의 자세별 로 사 에 정해져 있으므로 좌표는 기 좌표만큼 더하 기만 하면 된다. 다음은 Count 함수를 나타낸다. 여기서 (xi,yi) 는 Fig. 7에 표시한 바와 같이 형상 i의 좌상단 픽셀 치를 나 타내며, (xip,yip)는 우하단 픽셀 치를 나타낸다. Fig. 7은 두 ER 사이 겹쳐지는 부분을 짙은 회색으로 나타내고 있으며, 여 기서 두 ER간 겹쳐지는 픽셀 수, 즉 Count 값은 49이다. Function: count if (xj ≦ xip) ∧ (xi ≦ xjp) ∧ (yj ≦ yip) ∧ (yi ≦ yjp) width = min(xip,xjp) - max(xi,xj) + 1 height = min(yip,yjp) - max(yi,yj) + 1 count = width * height else count = 0 Fig. 7 Graphical representation of the evaluation function Goal 규칙은 해가 구해졌음을 단하는 규칙으로 체 부재의 치가 다 정해지면 실행된다. 다음은 CLIPS로 나타낸 goal 규 칙의 조건부의 일부를 보여 다. 여기서, Lay-tile-list의 값 $?tlist는 배치해야 할 부재 체 리스트이며 Node의 Tile-list 슬 롯 값과 일치하면 모든 부재가 배치되었음을 나타내므로 Goal 규칙은 노드가 해에 이르 을 때 실행됨을 알 수 있다. (defrule MAIN::goal ?open <- (open-list (nodes $?list)) (lay-tile-list $?tlist) (node (node-id ?nid) (tile-list $?tlist) (tile-orient-list $?torlist) (tile-position-row $?prlist) (tile-position-col $?pclist)) 4. 적용 예 4.1 블록 퍼즐 블록 퍼즐은 Fig. 8(a)와 같은 블록들과 빈 보드를 이용하여 (b)와 같이 채우는 게임이며, 블록들의 회 은 허용되지 않는다. 12 신동목 (a) Blocks and board (b) a solution Fig. 8 An example of block puzzle (a) Patterns (b) A solution Fig. 9 Patterns and a solution generated by the system (a) Patterns (b) A solution Fig. 10 Patterns and a solution 따라서, 하나의 블록을 배치할 때 경우를 나타내는 식(1)은 식 (3)과 같이 된다.   (3) 일반 인 네스 문제와 달리 빈 공간을 최소화 하는 최 화 문제가 아니라 사각평 에 형상들이 빈 곳 없이 채워져야 하는 정답이 있는 문제로서 NFP등을 이용한 일반 인 네스 알고 리즘으로는 풀기 어렵다. 본 연구에서 Fig. 8(a)와 같은 형상들은 Fig. 9(a)와 같이 이진 (Binary) 패턴으로 ER안에 표시한다. 각 형상을 구별하기 하 여 1~8의 숫자로 패턴을 표 하 으며, Fig. 8(b)와 동일한 배치 를 나타낸다. 4.2 일반 네스팅 일반 네스 문제에서는 회 칭을 통하여 한 형상당 8 개의 자매 형상이 만들어 진다. 퍼즐 문제와 달리 평 을 완 히 채우는 것인 일반 으로 불가능하며 따라서 배치 결과 여백 을 최소로 하는 해를 찾는다. 본 시스템에서는 주어진 평 의 크기에 맞추어 해를 찾는다. Fig. 10(a)는 여러 형상들을 ER 안 에 픽셀 형태로 표 한 것이다. 부재가 픽셀에 가득 차 있거나 걸쳐져 있는 곳은 ‘1’, 빈 곳은 ‘0’으로 표 된다. 시스템이 실행 되면 먼 각각의 형상에 하여 회 미러링에 의하여 자 매형상들 간에 복 없이 최 7 개의 추가 자매 형상들이 생성 된다. 네스 은 각각의 자매 형상 그룹에서 하나씩 골라 이루어진 다. 탐색 트리에서 한 노드에 하나의 형상을 추가할 때마다 배치 가능한 모든 경우의 수를 생성하며 이들은 식(2)의 평가 함수 값이 큰 순서로 정렬하여 Open-list의 맨 앞에 장함으로써 그 순서 로 탐색이 이루어지도록 하도록 한다. Fig. 10(b)는 35행 x 20열 크기의 평 에 네스 한 결과를 보여주며 해 에 실제 형상을 첩하여 나타내었다. 휴리스틱 탐색 기법을 이용한 네스 문가 시스템 13 Fig. 11 Patterns Fig. 12 The patterns after initialization Fig. 13 A Solution 일반 인 NFP 알고리즘을 이용한 네스 에서는 구멍이 있는 형상이나 오목한 형상의 경우 보통 구멍이나 오목한 부분을 채워 단순한 볼록 형상을 만든 후 배치를 한다. 본 시스템에서는 그 러한 제약이 없으며 Fig. 10(b)에서 형상 (5)와 같은 형상의 구멍 안에 다른 형상 (6)이 제약 없이 배치가 됨을 확인할 수 있다. 4.3 조선 강재류 네스팅 조선산업에서 블록 보강재류 네스 문제는 먼 폭이 같은 부재들끼리 분류한 후 각 그룹에 하여 동일 폭을 갖는 긴 재에 부재들을 배치를 한다. 따라서, 0 o ,180 o 회 형상, 그리고 각각에 한 미러 이미지의 4개 형상이 하나의 형상 가족이 된다. N 개의 형상 가족에 하여 배치 치를 배제하고 정렬 순서만을 고려해 보면 4 N N!가지의 경우의 수가 있다. 를 들어 8 개의 부재를 배치할 때 약 26억가지 이상의 경우의 수가 발생한다. 따라서 기존의 연구들에서는 큰 부재부터 한 쪽 끝부터 순차 으 로 연속 배치하는 방식을 사용하 으나 본 연구에서는 각 부재를 배치할 때 이 부재에 이어서 배치하는 것이 아니라 제약 없이 평 을 활용하여 배치를 한다. 두 형상 간 첩 여부는 양 단의 경계선 형상만 고려하면 되 므로 Fig. 11의 형상은 Initialize 규칙에 의해서 Fig. 12와 같이 코딩된다. 형상 내에서 세로 폭이 평 의 세로 폭과 같은 부분은 해당 부분의 가로 폭만을 기록하는데, 를 들어 (1)번 형상의 경우 Fig. 11에서 체 열이 ‘1’로 코딩된 이웃한 열들은 모두 합하여 13 개의 열이므로 Fig. 12에서 하나의 표 열에 13으로 코딩 되어 있음을 알 수 있다. 각 형상들이 Fig. 12와 같이 코딩된 후에는 형상들이 배치될 평 소재도 그 만큼 가로 폭을 인 상태에서 배치하며 해를 찾은 후에는 평 소재 각 형상들은 원래 가로 폭으로 복원 한다. 형상을 배치할 때 기본 으로 그 형상이 첩 없이 놓일 수 있는 모든 치를 고려하게 되므로 이와 같은 코딩을 통하 여 각 형상들이 배치될 치에 한 경우의 수를 일 수 있다. Fig. 13은 배치된 결과를 형상들과 첩시켜서 보여주고 있다. 4.4 이형 경계형상을 갖는 소재에 네스팅 각 형상을 배치할 소재가 직사각형 형상이 아닌 경우 기존의 Fig. 14 A solution for nesting on a irregular shaped stock 알고리즘의 경우 어려움이 있어 주요 연구주제가 되어 왔다 (Tay et al., 2002; Lee et al., 2008; Burke et al., 2007). 본 시스 템에서는 사각부재에 잘려나간 부분을 사 에 배치된 형상으로 처리함으로써 간단히 처리할 수 있다. Fig. 14는 이형 경계형상 을 갖는 소재에 형상들을 배치한 실행 결과를 보여 다. 5. 결 론 본 연구에서 개발한 네스 문가 시스템은 평가함수를 활 용한 휴리스틱 탐색을 통하여 해를 비교 짧은 시간( 제들의 경우 수천에서 수만개 이내의 노드 탐색)에 제시할 수 있었다. 퍼즐문제, 일반 네스 문제, 그리고 조선 강재류 네스 제 에 용하여 그 용 가능성을 보 으며, 각 문제별로 특성을 고려하여 문제를 효과 으로 해결할 수 있음을 설명하 다. 한 형상들을 배치할 평 소재의 형상이 사각 넬이 아닐 경우에도 가상의 사각 넬을 경계로 실제 경계 형상부분까지 이미 다른 형상이 배치되어 있는 상태로 처리하면 되므로 평 소재의 형 14 신동목 상도 문제가 되지 않음을 제를 통하여 보 다. 강재 네스 문제 등에서 고려해야 할 형상 간 공구 경로 확보 문제는 각 형 상에 공구경로 폭에 해당하는 Offset을 추가하는 방법으로 해결될 수 있다. 문가 시스템은 일반 로그램과 달리 시스템에 한 체 수정 없이 새로운 규칙을 지식베이스에 추가하는 방법으로 시 스템 성능을 쉽게 개선할 수 있으며 따라서 본 시스템도 경험 지 식 등에 의한 새로운 규칙을 추가하므로써 지속 성능 개선이 가능할 것으로 기 된다. 한, 본 연구에서 사용한 픽셀을 이용 한 형상표 방법을 이용할 경우 상업용 CAD 시스템과 쉽게 연 계할 수 있으며, 오목한 형상이나 구멍이 있는 형상들을 배치하는 데 도형의 근사형상을 이용한 방법과 달리 볼록 형상으로 근사화 해야 하는 등의 제약이 없다. 본 시스템에서는 정해진 원소재 크기에 맞추어 배치를 하게 되므로 원하는 수율에 맞추어 평 크기를 정하면 해를 구할 수 있다. 다만 휴리스틱에 의한 탐색을 하므로 해를 찾는 시간( 는 탐색 노드 개수)을 제한하면 해를 찾지 못하는 경우가 발생할 수 있으며, 그 경우 탐색 제한 시간을 늘리거나 평 소재 크기를 키우면 해를 비교 쉽게 찾을 수 있다. 일반 인 네스 문제는 NP-complete 문제로서 형상 개수가 늘어날수록, 표 의 정 도를 높이기 하여 픽셀의 갯수가 늘 어날수록 최 해를 구하는 것은 거의 불가능하다. 본 연구에서 개발한 시스템도 평가함수를 활용한 휴리스틱 탐색 기법으로 탐색 시간을 으나 형상의 개수가 늘어나거나 형상 표 정 도를 높이기 하여 픽셀 수를 늘릴 경우 탐색 시간이 길어 진다. 따라서 이러한 경우에는 픽셀 수를 여 해상도를 낮추거 나 형상들을 일정 단 로 묶어 그룹별로 네스 한 후 그룹 간 네스 하는 방법 등을 강구할 수 있을 것이다. 참 고 문 헌 이철수, 박 렬 (1996). “선박용 랫바의 자동네스 가스/ 라즈마에 의한 NC 단”, 산업공학, 제9권, 제3호, pp 283-297. 한국찬, 나석주 (1994). “ 이 단공정에서 단부재의 최 배치를 한 네스 알고리즘”, 한용 학회지, 제12권, 제2호, pp 11-19. 한창 , 박제웅 (1999). “ 라메트릭 매크로를 이용한 부재생성 네스 에 한 연구“, 한국해양공학회지, 제13권, 제2호, pp 176-185. Babu, A.R. and Babu, N.R. (2001). “A Generic Approach for Nesting of 2-D Parts in 2-D Sheets using Genetic and Heuristic Algorithms”, Computer-Aided Design, Vol 33, pp 879-891. Bennell, J.A. and Oliveira, J.F. (2008). “The Geometry of Nesting Problems: A Tutorial”, European Journal of Ope- rational Research, Vol 184, pp 397–415. Burke, E.K., Hellier, R.S.R., Kendall, G. and Whitwell, G. (2007). “Complete and Robust No-fit Polygon Generation for the Irregular Stock Cutting Problem”, European Journal of Operational Research, Vol 179, pp 27-49. Lee, W.-C., Ma, H. and Cheng, B.-W. (2008). “A Heuristic for Nesting Problems of Irregular Shapes”, Computer-Aided Design, Vol 40, pp 625-633. Tay, F.E., H. Chong, T.Y. and Lee, F.C. (2002). “Pattern Nesting on Irregular-shaped Stock using Genetic Algo- rithms”, Engineering Applications of Artificial Intelli- gence, Vol 15, pp 551-558. Weng, W.C. and Kuo, H.C. (2011). “Irregular Stock Cutting System Based on AutoCAD”, Advances in Engineering Software, Vol 42, pp 634-643. 2012년 5월 17일 원고 수 2012년 7월 10일 심사 완료 2012년 7월 12일 게재 확정 
work_7fzle5euafeo7mwqyrzwm23zau	----	International Journal of scientific research and management (IJSRM) ||Volume||3||Issue||10||Pages|| 3584-3591||2015|| Website: www.ijsrm.in ISSN (e): 2321-3418 Dr. Taruna Yadav, IJSRM volume 3 issue 10 Oct 2015 [www.ijsrm.in] Page 3584 Knowledge Sharing: Collaboration between Universities and Industry Organizations Dr. Taruna Yadav1 Vaibhav Shrivastava2 1- Assistant Professor, School for Management studies, Babasaheb Bhimrao Ambedkar University, Lucknow, (A Central University), 2- Research scholar, School for Management Studies, Babasaheb Bhimrao Ambedkar University, Lucknow (A Central University), Email:vaibhavsri26193@gmail.com Abstract Human being while traveling through Rural, Agricultural and Industrial Societies has now reached to the Knowledge Society. With the advancement of Information and Communication Technology, there are now drastic changes almost in every sphere of life. Due to increased competition as a result of globalization and continuous technological improvements, it is imperative for universities to collaborate with industry in order to enhance the distribution of knowledge, growth research and development, patent inventions and shape the nation’s organizational capacity. For this reason, it has become gradually clear that there is a need for close by a university - industry partnership as a means of national economic prosperity. It is obvious from the study of literature that university-industry partnership and their succeeding knowledge transfers are topics of economic, political, sociological, managerial and academic interest. Indeed, technological knowledge is look as a great source of long-term economic growth factor .This paper presents our literature analysis concerning this research topic and explores one particular means of inter-organizational knowledge transfer. The paper begins with the term knowledge. It then discusses knowledge sharing. For Collaboration between universities & Industrial organizations, collaboration policy, needs for collaboration, challenges and benefits of university – industry partnership. Keywords: Knowledge transfer, University/industry collaboration, Academic growth INTRODUTION Knowledge sharing To share knowledge means to study, understand, feast and duplication the information, the ideas, the views and the resources with each other, connected with, on a detailed ground. If we both exchange 1 Rupee, we both have 1 rupee each. But if we exchange 1 Good Thought, we both have 2 Good Thoughts. So if you share knowledge with me, we both will gain something and the knowledge of both of us will be increased so that means Knowledge sharing is an action through which knowledge is exchanged among people. Supposing that transfer activities between DOI: 10.18535/ijsrm/v3i10.1 Dr. Taruna Yadav, IJSRM volume 3 issue 10 Oct 2015 [www.ijsrm.in] Page 3585 universities and private sectors will add to corporate competitiveness and growth of the economy, researchers intensively examined university-industry associations. This research field has been highly indicated the traditional open sciences channels and other knowledge communications such as informal associations and combined research & development projects. Collaboration for Knowledge Sharing between Universities and Industrial Organization The association between universities and industries may play a vital role in the area of knowledge sharing. As we know, the research work is a constant process, on various levels, at many places, at the same time. In any subject or any field, daily knowledge is calculating new measurements from the angles of the globe. Every year, many students take admission in institutions of higher education. Some of them start research work in their particular subject. Generally, all the research work has done at this place. It is very important is to expose these works. The concept of collaboration for research work is not new. Many countries in the world are involved in this kind of partnerships on global levels also. However, for India, we have not made more efforts in this field. It is very much essential to apply knowledge on practical platform. Globalization strains that our society requirements to move quicker, work smarter and take more risks than at any time in our history. Every academy has its research students and experts and every industrial firm has its experts, employees who practically work on projects having employed knowledge of many years. By association, they all can share the information and can lead the work in a specific way, can add new measurements in knowledge and can create new standards. Any new invention for example, to make fuel from water, the researcher discovers it and the manufacturing firm places it in exercise. The information created in the educational domain takes various tracks before finally attainment a reasonable receiver, from patent and licenses to Research publication or consulting. LITRETURE REVIEW The study of knowledge sharing has its roots within the technology transfer and improvement literature. The research in this area has focused on clarifications for different nation’s successes or failures in nurturing economic growth through technological development. While some philosophers argue that high investment rates in physical and human capital drive national innovation and growth rates (Young, 1993; Kim & Lau, 1994; Krugman, 1994), ‘assimilation theorists’ instead argue that entrepreneurship, effective learning, and innovation are separate, but equally important variables affecting development (OECD, 1971; Freeman, 1982; Kim & Nelson, 2000). Central to both approaches, nonetheless, is an understanding of the importance of the sharing of ideas. In this literature, successful knowledge sharing results in firm understanding and getting into DOI: 10.18535/ijsrm/v3i10.1 Dr. Taruna Yadav, IJSRM volume 3 issue 10 Oct 2015 [www.ijsrm.in] Page 3586 practice product designs, manufacturing processes, and organizational designs that are new to them (Nelson, 1993). As demonstrated by the title of Richard Nelson’s recent volume on technology transfer, Technology, Knowledge, & Innovation (Kim & Nelson, 2000), knowledge sharing is seen as occurring through a dynamic learning process where organizations continually interact with customers and suppliers to innovate or creatively imitate. Consider the case of technology transfer as articulated by Lall (2000, p. 15) Knowledge is a well-ordered combination of information, adapted within a set of rules, procedures and operations learned through experience and practice (Keskin, 2005). The literature classified knowledge into two main types: tacit and explicit. Explicit knowledge is knowledge that can be seen, shared and easily communicated to others. Most explicit knowledge is in the form of raw data, such as documents that contain the work experiences of staff, descriptions of events, interpretations of data, beliefs, guesses, hunches, ideas, opinions, judgment and proposed actions (Choo, 2000). Tacit knowledge is more difficult to share because it is embedded in a person’s memory. These conclusions have led development experts to recommend that activities focused on facilitating knowledge sharing rather than on transmitting Northern knowledge to the South are likely to prove more successful (Ellerman, Denning & Hanna, 2001; Knowledge for Development, 1998; Social Development Group, 2002; Prusak, 1999). In other words, while the communication of knowledge is important, it is the processes through which knowledge is shared that determine whether organizational learning occurs and, therefore, whether a knowledge- sharing process was a success. OBJECTIVES OF THE STUDY  To understand the conception of knowledge sharing between University & Industry.  To explore the need and scope of University & Industry collaboration.  To discuss the collaboration policy for knowledge sharing.  To analyze the benefits and challenges of University – Industry collaboration and provide some alternative solutions. RESEARCH METHEDOLOGY Research and experimental development work is undertaken systematically to increase the stock of knowledge. The first objective of this paper fulfil by the analysis of introduction of the paper. This research paper is carried out with the help of secondary data. The major tools for the collection of the information has been available collected primarily from journals, articles, online database of Indian Economy, websites or newspaper etc. DISCUSSION AND ANALYSIS DOI: 10.18535/ijsrm/v3i10.1 Dr. Taruna Yadav, IJSRM volume 3 issue 10 Oct 2015 [www.ijsrm.in] Page 3587 Knowledge cannot be produced in an unexpected manner – it needs to be managed well. For that Universities & companies who have combined hands for a specific subject or specific drive should first of all, decide a common frame of work. There are numerous needs for make a corporation of industrial organization and educational institute or universities to make a sustainable growth. NEED OF UNIVERSITY-INDUSTRY COLLABRATION The need for sharing information between university and industry has become gradually obvious in recent time. Historically, research institutions were supposed as a Cause of new ideas and industry offered a usual way to 0get the most out of the use of these ideas. However, the past period has seen a substantial change in the roles of both parties. Many corporations are developing open invention approaches to research & development, combining internal and external resources, and pointing to get the most out of economic value from their knowledgeable property, even when it is not directly linked to their central business. It has become clear that universities need to play a livelier role in their association with industry in order to maximize the use of the research results. It is very much essential to understand needs for partnership and decide collaboration policy to better understand and increase the partnerships.  Knowledge which is generated in research works at university level lies in theses unused in libraries. To pull out the Knowledge from thesis.  To get fresh & pure knowledge directly from universities for industries.  To reduce the time for research at the industrial level.  In universities research work goes on and on haphazardly without any specific direction. To give a specific direction  By getting a perfect direction, research work will be done more speedily. Knowledge will grow faster which is very much essential for knowledge society. TABLE 1 . DIFFERENT WAY OF KNOWLEDGE SHARING BETWEEN UNIVERSITY AND INDUSTRY Publications Scientific publications Co-publications Consulting of publications Participation in conference Participation in fairs professional networks & boards Participation in conferences Exchange in professional organizations Participation in boards of knowledge institutions Participation in governmental organizations DOI: 10.18535/ijsrm/v3i10.1 Dr. Taruna Yadav, IJSRM volume 3 issue 10 Oct 2015 [www.ijsrm.in] Page 3588 Collaboration policy for University and Industry Partnership Universities have research students and subject experts while enterprises have their experts, Experienced workers, and mechanism. By collaboration, the knowledge can be shared among both of the parties. Each organization and university should develop a policy specific to its own needs and objectives. This policy should define the: Objectives First of all, it is very much important to clear the objectives for which both the industries & universities have merged hands. Context Research should be developed in a framework of formation and presentation of knowledge. Research should be designed to reflect that knowledge is developed collectively and that the ‘distribution’ of knowledge leads to its continuous application in different frameworks, by different participants with differing backgrounds and experiences. Participants Research scholars and business experts at university level and the experts and employees Mobility of people Graduates Mobility from public knowledge institutes to industry Mobility from industry to public knowledge institutes Trainees Double appointments Temporarily exchange of personnel Cooperation in R&D Joint R&D projects Presentation of research Supervision of a trainee or Ph.D. student Financing of Ph.D. research Sponsoring of research Sharing of facilities Shared laboratories Common use of machines Common location or building (Science parks) Purchase of prototypes Cooperation in education Contract education or training Retraining of employees Working students Influencing curriculum of university programs Providing scholarships Sponsoring of education Contract research and advisement Contract-based research Contract-based consultancy DOI: 10.18535/ijsrm/v3i10.1 Dr. Taruna Yadav, IJSRM volume 3 issue 10 Oct 2015 [www.ijsrm.in] Page 3589 who have practical working knowledge and experience at the other end should be contributing in the project. Processes The research effort should be established to support and improve knowledge intensive processes. The whole work should be a knowledge life cycle by involving the research students and subject experts as a knowledge generator, the experts of firms who execute the knowledge and the works who actually apply the knowledge on real-world. Here, knowledge will flow in both the directions. And at the time of actual execution on practical ground, if any difficulty derives, the feedback will be given to the researcher and by removing the deficiencies, it will be improved and will be made more perfect and more useful for enterprises. Privacy and other rights To keep the maters and materials and the progress of work top confidential between collaborating parties only is a must. All rights should be reserved for enterprises if it is supported financially by firms. BENEFITS OF UNIVERSITY AND INDUSTRIAL COLLOBORATION  Sharing of valuable knowledge.  U-I partnership able to avoid re-inventing the wheel, reducing redundant work.  May reduce the cost for inventions.  Formation of knowledge with the support of experts and experienced persons.  By giving a right direction to the enthusiastic intelligent students, making them experts of future.  Which kind of change manufacturing firms needed? Which kind of difficulties they are facing and to resolve it, which kind of research works they are supposing from the university will be cleared well in advance. Maximum production with the lowest cost is the main aim of all enterprises. If, they were raw materials, or machinery and technology or management deals.  By collaborations, the firms will inform university and university will frame the research work as per the needs to fulfill the aim.  Any kind of the problem ascends, will be solved at the primary level, which will save the time, money and manpower. DOI: 10.18535/ijsrm/v3i10.1 Dr. Taruna Yadav, IJSRM volume 3 issue 10 Oct 2015 [www.ijsrm.in] Page 3590 Source –http://blogs.lse.ac.uk/impactofsocialsciences/files/2014/08/healy1.jpg CHALLENGES IN UNIVERSITY– INDUSTRY INTERACTIONS There are some major challenges in university– industry interactions or collaboration which are-  Generally lack of information about what is on proposal from universities  Quality of material re. innovation provided by universities once contact made  Generally low level of engagement with universities.  Key problems remain in terms of basic mismatches in terms of relevance, time horizons and expectations CONCLUSION There is much more to be done. Universities will have to get better at identifying their areas of competitive strength in research. The government will have to do more to support an industry- university collaboration. Industries will have to learn how to exploit the innovative ideas that are being developed in the university sector. With such collaborations, government, industry and universities, all three parties can sharpen their entrepreneurial skills to effectuate transformation of the nation’s science and technology landscape. One key strategy is that of responsiveness in the higher education sector. Greater responsiveness imply that universities should take the problems and challenges presented by the societal context in which they operate seriously (HSRC 2003-1). Knowledge sharing appears to work best when it is seen not so much as a relay race, but as a team sport. It is ‘a game during which the ball moves continuously between the players and in which all players have to collaborate and share resources to win’ (Entrepreneurial Higher Education Institution 2002-10–11). DOI: 10.18535/ijsrm/v3i10.1 Dr. Taruna Yadav, IJSRM volume 3 issue 10 Oct 2015 [www.ijsrm.in] Page 3591 There are several areas in which research efforts can be focused on the future. For example, similar understanding of blockades, trials and success factors is required. Further research is needed to understand the mechanisms by which universities transfer research & development knowledge in order to increase industry competitiveness and efficiency as well as complete economic and social growth. Research on how association with universities influences the decision making procedures in industrial firms may be a relevant topic, as would be research that explores how intellectual property rights are negotiated between universities and industry partners. To frame a network for academic libraries of a country for knowledge sharing is also the area of future work. It is hoped that this work has contributed to our understanding of knowledge sharing for saving the time, money and man power and contribution of university libraries as providers of scientific R&D knowledge to move the knowledge life cycle faster. In future, to make India a knowledge super power, this kind of collaborations should be increased. References:- 1. Barry, Brooke-Norris. [On line] Information management: human, ethical and epistemological challenges, South African Journal of Information Management, Vol.3 (1), June 2001 available at http:/ /www.sajim.co.za 2. C. Walt, P.A. Brakel and J.A. Kok. [On line] Knowledge sharing via enterprise intranets – asking the right questions, South African Journal of Information Management, Vol.6 (2), June 2004 available at http:/ /www.sajim.co.za 3. Agrawal, A. (2001) ‘University-to-industry knowledge transfer: literature review and unanswered questions’. International Journal of Management reviews, 3 (4), 285-302. 4. Amesse, F. and Cohendet, P. (2001) ‘Technology transfer revisited from the perspective of the knowledge-based economy’. Research Policy, 30, 1459-1478. 5. Autant-Bernard, C. (2001) ‘Science and knowledge flows: evidence from the French case’. Research policy, 30, 1069-1078. 6. Barnes, T., Pashby, I. and Gibbons, A. (2002) “Effective University-Industry Interaction : A Multi-Case Evaluation of Collaborative RandD Projects”. European Management Journal, 20 (3), 272-285. 7. Bozeman, B. (2000) ‘Technology transfer and public policy: a review of research and theory’. Research Policy, 29, 627-655. 8. Cassier, M. (2002) ‘L’appropriation des connaissances dans les partenariats de recherche entre laboratoires publics et enterprises: quelques tendances récentes’. Rapport de Recherche du Programme du CNRS “Les enjeux économiques de l’innovation”, convention 00.77.001 
work_7gj5kurvuzau7m3p6lv7viz5m4	----	1 A Bi-objective Feature Selection Algorithm for Large Omics Datasets Luís Cavique 1,2 [0000-0002-5590-1493] , Armando B. Mendes 3,4 [0000-0003-3049-5852] , Hugo F.M.C. Martiniano 1,5 [0000-0003-2490-8913] and Luís Correia 1 [0000-0003-2439-1168] 1 MAS-BioISI, FCUL, Lisboa, Portugal, 2 Universidade Aberta, Lisboa, Portugal, 3 Universidade Açores, Ponta Delgada, Portugal, 4 Algoritmi, Universidade do Minho, Portugal, 5 Instituto Dr. Ricardo Jorge, Lisboa, Portugal, luis.cavique@uab.pt, armando.b.mendes@uac.pt, hfmartiniano@fc.ul.pt, luis.correia@ciencias.ulisboa.pt Abstract. Feature selection is one of the most important concepts in data mining when dimensionality reduction is needed. The performance measures of feature selection encompass predictive accuracy and result comprehensibility. Consistency based methods are a significant category of feature selection research that substantially improves the comprehensibility of the result using the parsimony principle. In this work, the bi-objective version of the algorithm Logical Analysis of Inconsistent Data is applied to large volumes of data. In order to deal with hundreds of thousands of attributes, heuristic decomposition uses parallel processing to solve a set covering problem and a cross-validation technique. The bi-objective solutions contain the number of reduced features and the accuracy. The algorithm is applied to omics datasets with genome-like characteristics of patients with rare diseases. Keywords: feature selection, logical analysis of data, heuristic decomposition, bi-objective optimization 1. Introduction In feature selection two main objectives sustain the performance measures, the predictive accuracy and the result comprehensibility [Liu, Yu 2005]. Some models do not consider the predictive accuracy, whereas others tend to destroy the underlying semantics of the features after reduction. It would be highly desirable to find a method that could not only reduce the number of features, but also preserve the data semantics. In this context, Rough Set theory emerges as an important tool to discover data dependencies and reduce dimensionality. Rough Set theory was initially proposed as a tool to reason about vagueness and uncertainty in information systems by Pawlak [Pawlak 1982] and later it was also proposed for attribute selection [Pawlak 1991]. Rough Sets determine a lower and an upper approximation for each class, rather than correct or exclude data inconsistencies. In parallel, Peter Hammer's group [Crama et al. 1988], [Boros et al. 2000], with works in discrete optimization, developed LAD, Logical Analysis of Data. The key features of LAD are the discovery of the minimum number of attributes necessary to explain all mailto:luis.cavique@uab.pt mailto:armando.b.mendes@uac.pt mailto:hfmartiniano@fc.ul.pt 2 observations and the detection of hidden patterns in a dataset with two classes. An extension of the Boolean approach uses nominal non-binary attributes. LAD and Rough Set approaches are a subset of filter models that aim to reduce the number of attributes of datasets using two phases: problem transformation and optimization. Their specificity is to keep the semantics of the data by only removing the redundant data. By combining the two approaches, we proposed Logical Analysis of Inconsistent Data, LAID [Cavique et al. 2013], which blends the best characteristics of both methods. The LAID method should deal with integer attributes associated with costs as in LAD and be tolerant to inconsistency as in Rough Sets. The integration of both approaches is so close that LAID can be seen as a Rough Set extension. One key area for the application of the LAID method is high-dimensional omics datasets, such as those generated by high-throughput sequencing technologies. In the past few years, technological developments in this area have resulted in an astonishing growth of the amount of data produced. As most disease related omics datasets are subject to restricted access, artificial datasets with similar characteristics have been created. The goal of this work is to solve a feature selection problem with a dataset involving two thousand observations and one million attributes. To deal with millions of attributes a heuristic decomposition is used in a parallel computer environment. We will consider two performance measures for comprehensibility of the result and accuracy. In the comprehensibility of the solution, we aim to achieve the minimum number of attributes with the maximum actionable knowledge that better explains the dataset to the final user. The accuracy is given by a cross-validation technique. This document extends on works with LAID characterization [Cavique et al. 2011], [Cavique et al. 2013]. The application of LAID to larger datasets with the development of a heuristic decomposition approach with a sub-problem algorithm was presented in [Cavique et al. 2017]. Following the previous work, the novelty of this document is the presentation of the master problem algorithm with a bi-objective approach for omics datasets. The bi-objective reflects both the predictive accuracy concern and the result comprehensibility. This document is organized as follows. In Section 2, we present the related concepts of feature selection, omics datasets and heuristic decomposition. In Section 3, we present the sub-problem algorithm that finds the feature selection with the minimum number of attributes and computes the accuracy. In Section 4, the heuristic decomposition algorithm and the bi-objective master problem algorithm are presented. In Section 5, the computational results are shown. Finally, in the last section we draw some conclusions about this new approach. 3 2. Related work This work combines the areas of attribute selection and problem decomposition to deal with omics datasets. Consequently, in this section we introduce the related topics: feature selection, omics data and heuristic decomposition. As they will be combined in the proposed sub-problem algorithm, two feature selection methods, Rough Sets and LAD are detailed. 2.1. Feature Selection The motivation to reduce the dimensionality of the feature space is closely related to the decreased time required to double information in the world every year. Surveys on feature selection methods can be found in [Liu, Yu 2005], [Chandrashekar, Sahin 2014]. In feature selection, given thousands to millions of features, the goal is to select the most relevant ones. The performance measures have two main goals, the predictive accuracy and the result comprehensibility. In this work we emphasize the comprehensibility of the results, following Occam’s razor principle that aims to obtain the simplest model. We consider a dataset D={O, X∪C} where O={O1,O2,…,On} is a non-empty set of observations (instances or cases), X={X1,X2,…,Xm} is a non-empty set of features (attributes or columns) and C is the class attribute. In this brief review we use the feature selection taxonomies reported in [Liu, Yu 2005]. There are three basic models in feature selection: Filter, Wrapper and Hybrid model. In the Filter model the most popular independent criteria are consistency measures, distance measures, information measures and dependency measures. In this sub-section the distance measure and the consistency measure are detailed because they are used in this document. Distance measure criteria are applied in observations within the same class and in observations of diverse classes. In the same class, to obtain the disagreement value, the logic operator XOR is applied, where it returns True if (Ox, Xa) ≠ (Oy, Xa) and False otherwise. With diverse classes, the logic operator XOR can also be applied, where if (Ox, Xa) ≠ (Oy, Xa), feature Xa should be chosen because it differentiates the classes. Comparisons of observations within the same class measure the incoherence or noise of the feature. On the other hand, comparisons of observations between different classes measure how strong a feature is in the discrimination or separation of the classes. RELIEF algorithm [Kira, Rendell 1992] evaluates a feature subset based on the subtraction between the distance of observations in different classes and the distance of observations within the same class. Using the consistency measure criteria, an inconsistency occurs when two or more observations have the same values in all attributes but belong to different decision classes. This measure is used in algorithms that attempt to find the minimum number of features with the minimum number of inconsistencies. The well-known algorithm FOCUS [Almuallim, Dietterich 1991] uses the concept of consistency, where irrelevant 4 features are removed. The LAD algorithm [Crama et al. 1988], [Boros et al. 2000] is developed in two steps. In the first step, a disjoint matrix is obtained using the XOR operator to compare observations from different classes. In the second step, a set covering algorithm is applied to ensure data consistency. In the feature selection process using addition or removal of attributes, the information measure is given by the information gain from an added or removed feature. The dependency measure of a feature is related to how important the correlation is between the feature and the class. The Filter model is divided into two sequential steps. The feature selection step is executed before the learning phase of the prediction model, and there is no interaction between the selection and the prediction model. The Wrapper model [John et al. 1994] is also divided into two steps, but with strong interaction between the feature selection phase and the learning phase, where the results of the prediction are used as a criterion of feature choice. Filter models are more intuitive and show better performance since they build the solution in a constructive process without iterations. On the other hand, they present the disadvantages of ignoring the predictive process. Consequently, most of the relevant features might not be adequate in the prediction model and the selection criterion is hard to estimate. In the Wrapper models, the selection criterion is easy to estimate since the features are chosen by the prediction model, and therefore they are classified as model-aware as they incorporate the knowledge of the predictor. Contrary to the Filter models, Wrapper models are computationally expensive and less intuitive. The main disadvantage of this method is the increasing overfitting risk when the number of observations is insufficient. In other words, the Wrapper models do not identify statistical dependency, so the features might not be the main explanatory variables and therefore the model can lose its theoretical basis. Hybrid methods have been proposed to reduce features in classification by combining the advantages of the two previous methods. The main disadvantage of the method is the significant low scalability when the feature number increases. To sum up, in Filter methods, the selection criterion is hard to estimate, whereas Wrapper methods tend to destroy the underlying feature semantics. In this work, we opt for Filter methods for two main reasons: (i) given that large volumes of data should be processed, Filter methods are more computationally efficient, (ii) given that the result comprehensibility is an important criterion, Filter methods are preferable since they preserve the semantics of the data. 5 2.2. Rough Sets and Logical Analysis of Data (LAD) In this sub-section two Filter methods are presented, Rough Sets and Logical Analysis of Data (LAD). Both methods try to preserve the semantics of the data and the algorithms are composed of two-phases, the problem transformation and the problem optimization. Rough Sets algorithm transforms data into a discernibility matrix and obtains the result by solving a SAT optimization problem. LAD firstly converts data into a disjoint matrix and then applies the Set Covering problem to obtain the result. Rough Sets Rough Sets theory was initially proposed as a tool to reason about vagueness and uncertainty in information systems by Pawlak [1982] and later it was also proposed for attribute selection by Pawlak [1991]. The applications of the Rough Sets method are wide leading to meaningful results in many fields, such as conflict analysis, finance, industry, multimedia, medicine, and most recently bioinformatics [Polkowski 2002] [Peters, Skowron 2010]. A detailed example of attribute selection with Rough Sets can be found in Cavique et al. [2013]. In Rough Sets, table values can have any integer value rather than binary values. Inconsistencies occur when two cases have the same values for all attributes but belong to different decision classes. A practical example is two sick people that have the same symptoms but different diseases. With real data this is possible, because the table might have a missing attribute that could discriminate between them. Rough Sets do not correct or exclude the inconsistencies, but for each class they determine a lower and an upper approximation. Given D={O,X∪C}, the subset of objects YO and the subset of attributes BX, Pawlak’s Rough Sets theory defines two approximation spaces: the lower and upper rough approximation. The lower approximation BL(Y) is the least composed set that is contained in Y, and the upper approximation B U (Y) is the greatest composed set that contains Y. As a consequence the approximation space leads to BL(Y)YB U (Y). Also, the lower and upper approximations of a subset YO can be seen as operators in the universe of objects dividing it into three disjoint regions, the positive region POS(Y), the negative region NEG(Y) and the boundary region BR(Y): POS(Y) = BL(Y), NEG(Y) = O  B U (Y) and BR(Y) = B U (Y)BL(Y). When the lower and upper approximations are equal, BL(Y)=B U (Y), there are no inconsistencies and the rough set is called crispy rough set. Another way to identify the roughness of the set is using measures. The accuracy approximation measure is given by: 𝛼(𝑌) = | 𝐵𝐿(𝑌)| | 𝐵𝑈(𝑌)| where |Y| denotes the cardinality of Y0 and 0(Y)1. If (Y)=1, X is crisp; otherwise, it is a Rough Set. The goal of Rough Sets is to discover decision rules from dataset D. The minimum number of attributes required to explain all the observations. Discovering the minimum number of attributes is an NP-hard problem. One of the following techniques is normally used: Reduction by Heuristics or Discernibility Matrix. 6 In Reduction by Heuristics, the search for a core is given by the following procedure: for each iteration, one attribute is removed, and the augmentation of inconsistency is verified. As already referred, inconsistency occurs when two or more observations have the same values for all attributes but belong to different decision classes. If the inconsistency does not increase, the attribute should be removed. When no further attributes can be removed, the remaining ones are considered indispensable and thus the core is found. Using a discernibility matrix of D, denoted by M, an (nn) matrix is defined as follows, where M(i,j)= denotes that this case does not need to be considered. 𝑀(𝑖, 𝑗) = { {𝑥 ∈ 𝑋: 𝑥(𝑜𝑖 ) ≠ 𝑥(𝑜𝑗 )} 𝑖𝑓 𝑐(𝑜𝑖 ) ≠ 𝑐(𝑜𝑗 ) ∅ 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 The discernibility matrix keeps the distinct attributes for each pair of observations belonging to different classes. Discernibility function F(B) is a Boolean function, written in the disjunctive normal form (DNF), that is a normalization of a logical formula which is a conjunction of disjunction clauses. F(B) determines the minimum subset of attributes that allows the differentiation of classes: F(B)= { M(i,j): i, j=1,2, …, n; M(i,j) }. The F(B) decision problem is equivalent to the Satisfiability problem (SAT), which was the first known example of an NP-complete problem. Logic Analysis of Data (LAD) The LAD method developed by Peter Hammer´s group [Crama et al. 1988], [Boros et al. 2000] refers to the discovery of the minimum number of attributes that are necessary to explain all observations and the detection of hidden patterns in a dataset with two classes. The method works on binary data. Let D be the dataset of all observations, then each observation is described using several attributes, and each observation belongs to a class. An extension of the Boolean approach is needed when nominal non-binary attributes are used. The binarization (or discretization) of these attributes is performed by associating a string of Boolean variables to each attribute. Dataset D is given as a D + set for ‘positive’ observations and as a set D − set for ‘negative’ observations, where D = D + ∪D − and the sets are disjoint D + ∩D − =Ø. Observations are classified as positive or negative based on a hidden function, and the goal of the LAD method is to approximate this hidden function with a union of intervals. In order to systematize the process, a disjoint matrix [a(i,j)] will be defined as: 𝑎(𝑖, 𝑗) = { 1 ∀𝑖: 𝑥𝑗 (𝑜𝑎) ≠ 𝑥𝑗 (𝑜𝑏 ), 𝐶(𝑜𝑎) ≠ 𝐶(𝑜𝑏 ), (𝑜𝑎, 𝑜𝑏 ) ∈ 𝑂x𝑂 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 Given the disjoint matrix the Set Covering optimization problem is used to find the selected attributes. 7 2.3. Omics Datasets Omics is a neologism associated with biology suffixes such as genomics, transcriptomics, proteomics, or metabolomics. Omics datasets are typically characterized by high dimensionality and small samples, that is, a large number of features and few observations. One prime example is data from DNA sequencing, especially those generated by the third generation, also called next-generation or high- throughput sequencing machines. A typical genome differs from the reference human genome from 4.1 to 5.0 million variants. For large cohorts, the number of genetic variants is even higher. The 1000 Genomes Project sequenced the genomes of 2504 individuals from 26 populations producing a total of 88 million genetic variants [The 1000 Genomes Project Consortium, 2015]. Presently, the amount of sequencing data is doubling every seven months [Stephens et al. 2015] and, with the advent of new, high-throughput sequencing machines, the growth is expected to accelerate even more. When combined with the high dimensionality of the datasets, this growth makes the development of scalable and accurate feature selection methods even more essential. Since biomedical genome sequence datasets are subject to restricted access, we generated artificial datasets with similar characteristics using SeqSIMLA [Yao, Chung 2016; Chung et al. 2014] (http://seqsimla.sourceforge.net/). SeqSIMLA generates simulated sequencing and phenotype data for case-control cohorts with allele frequencies and linkage disequilibrium patterns based on real datasets. The user can specify around 20 parameters, like disease prevalence or odds ratio for the marginal effects or interactions. 2.4. Heuristic Decomposition Approach To solve a problem with hundreds of thousands of attributes, a Decomposition Heuristic is introduced. In Linear Programming optimization some approaches can be mentioned such as Column Generation, Dantzig-Wolfe decomposition and Benders decomposition. A detailed introduction can be found in [Boyd et al. 2008]. Decomposition is the act of breaking a large problem into sub-problems. Decomposition of a problem is possible if the structure of the problem is maintained. If problem A is separable into sub-problems A 1 , A 2 , …, A max_k it can also run in a parallel computer environment. Instead of minimizing the evaluation function f(A) the decomposition method minimizes the sub-problems f(A 1 ), f(A 2 ), ..., f(A max_k ) and finally minimizes the master problem using function g(), which aggregates solutions of the sub-problems. Algorithm 1 presents the parallel version and Algorithm 2 the sequential version. 8 Algorithm 1: General Heuristic Decomposition (parallel version): Input: sub-problems A 1 , A 2 ,…, A 1..max_k Output: solution of the master-problem algorithm 1. solve the sub-problems: sub-problem 1: y1=minimize f(A 1 ) sub-problem 2: y2=minimize f(A 2 ) ... sub-problem k: ymax_k=minimize f(A max_k ) 2. solve the master-problem: master = minimize g(y1, y2,…, ymax_k) In the sequential version of the heuristic decomposition, Algorithm 2, updates the master-problem at each iteration by adding additional information from the solved sub- problem. Algorithm 2: General Heuristic Decomposition (sequential version): Input: sub-problems A 1 , A 2 ,…, A 1..max_k Output: solution of the master-problem algorithm 1. for k=1 to max_k 1.1. solve sub-problem: yk = minimize f(A k ) 1.2. solve master-problem: master = minimize g(master, yk) 2. end for The decomposition principle is applied in exact methods and in metaheuristic methods [Joncour et al. 2010]. Given the large number of attributes, a heuristic decomposition approach, coined by [Smet et al. 2016] is used in this work. In Section 3 and Section 4, we detail the sub-problem and the master problem algorithms for the Logical Analysis of Inconsistent Data to deal with large datasets. 3. Sub-problem Algorithm As already stated, in this work we propose a feature selection technique to deal with inconsistent data in large datasets. The inconsistent data is processed using the LAID algorithm. The sub-problem algorithm returns two criteria, the information about the quality of the filter (number of features) and the prediction quality (accuracy). Thus, the output of this filter method is the pair (number of features, accuracy). To solve the sub-problem, in Algorithm 3, we describe the LAID Algorithm for the bi- criteria feature selection. Firstly, to remove any inconsistency, we add a dummy binary variable named “je ne sais quoi”, ‘jnsq’, since two observations with the same attributes that belong to different classes work against the consistency measure. In a second step, the Disjoint Matrix Generation [Ai,j] and Cost Vector [cj] are created, where both are based on the definition of the distance measure. In the third step, a Heuristic for the Set Covering problem is applied, returning the reduced set of features. In the fourth step, a cross-validation method is applied to retrieve the accuracy of the solution. Finally, the 9 algorithm returns the pair (number of features, accuracy). The algorithm can be specified as follows: Algorithm 3: Logical Analysis of Inconsistent Data Input: dataset D = {O, X∪C} with binary variables Output: (number of features, accuracy) 1. check data inconsistencies and add dummy variable ‘jnsq’ as a discriminant feature 2. disjoint matrix generation [Ai,j] and cost vector [cj] 3. number of features = Minimum Set Covering Problem 4. accuracy = Cross-validation To exemplify the different steps of the algorithm a running example with six features and five observations is used. 3.1. Data inconsistencies Given the inconsistency of two observations Oa and Ob with the same values for all the attributes X(Oa)=X(Ob) and belonging to different classes, a dummy binary variable “je ne sais quoi” [Cavique et al. 2013] is added, assigning a value of 1 whenever the class=1, or: jnsqa = { 1 if (𝑋(Oa) = X(Ob)) (class (Oa) ≠ class (Ob))  (class(Oa) = 1) 0 otherwise Instead of using the complex approximations of the Rough Sets keeping the data inconsistencies, LAID does not exclude the inconsistency by adding a dummy variable that allows the subsequent application of the other steps of the LAD method. Given the dataset with six features {X1, …, X6} and five observations {O1, …, O5}, the first step of LAID is verifying the existence of inconsistencies in the dataset. For each inconsistency the dummy variable ‘jnsq’ is assigned ‘1’ in class 1. In Table 1 the dummy variable ‘jnsq’ is added in the last feature column to avoid data inconsistencies between observation O2 and O3, which present the same values for all the features. Table 1. Variable ‘je ne sais quoi’ is added observation\feature X1 X2 X3 X4 X5 X6 X7 (jnsq) Class O1 1 1 0 1 1 1 0 0 O2 1 0 1 1 1 0 0 0 O3 1 0 1 1 1 0 1 1 O4 1 0 1 0 0 1 0 1 O5 0 0 1 0 0 0 0 1 10 3.2. Disjoint Matrix Generation [Ai,j] and Cost Vector [cj] In the second step the Cost Vector [cj] and Disjoint Constraints Matrix Generation [Ai,j] are created. Both, the Disjoint Matrix Generation [Ai,j] and Cost Vector [cj] use the concept of distance measure for independent criteria in feature selection. The Cost Vector [cj] is obtained by adding the pair comparisons within the same class. This calculation is used to measure the general incoherence, or noise of feature j. On the other hand, the Disjoint Matrix [Ai,j] is obtained by using the pair comparisons belonging to different classes. The disjoint matrix [Ai,j] for each attribute Xj and pair of observations (Oa, Ob), is defined as: Ai,j = { 1 ∀i: Xj(Oa) ≠ Xj(Ob), class(Oa) ≠ class(Ob), (Oa,Ob)∈O×O 0 otherwise The dimension of index i in matrix [Ai,j] depends on the constraint structure of dataset D. Each constraint results from the comparison of two different arbitrary observations Oa and Ob that belong to distinct classes. If one attribute j is different in the observations Oa and Ob the value of Ai,j becomes value 1, denoting that at least one column (attribute) j must be maintained in order to differentiate the rows (or constraints) i. The cost vector [cj] and the disjoint matrix [Ai,j] will be used as input in the Minimum Set Covering problem, where all constraints (or lines i) must be covered, at least once by some attributes. Table 2 presents pairs of observations of the same class. For each observation the features of the same class should be equal to avoid disagreement. The sum of the disagreement of each feature is called cost and stored in the Cost Vector [cj]. As stated before, for observations Oa, Ob and attribute Xj, the value of disagreement is given by the operator XOR that returns True if (Oa, Xj) ≠ (Ob, Xj) and False otherwise. Table 2. Cost vector [cj] shows the disagreement within the same class Pairs same class X1 X2 X3 X4 X5 X6 X7 (jnsq) Class O1,O2 1 1 1 1 O3,O4 1 1 1 1 0 O3,O5 1 1 1 1 0 O4,O5 1 1 0 cost [cj] 2 1 1 2 2 3 2 To obtain the disjoint matrix [Ai,j], for each pair of observations, the features of different classes should also be different in order to be able to discriminate them. The number of lines generated in [Ai,j] is N.M, where N and M are the number of observations of class 0 and class 1, respectively. The value kj corresponds to the number of lines covered for each feature j. 11 Table 3 presents the pairs of observations that belong to different classes, where the logic operator XOR can also be applied, since, if (Oa, Xj) ≠ (Ob, Xj), the feature Xj should be chosen because it distinguishes the two classes. Table 3. Matrix [Ai,j] shows the disjoint between different classes Pairs different classes X1 X2 X3 X4 X5 X6 X7 (jnsq) 1 - O1,O3 1 1 1 1 2 - O1,O4 1 1 1 1 3 - O1,O5 1 1 1 1 1 1 4 - O2,O3 1 5 - O2,O4 1 1 1 6 - O2,O5 1 1 1 kj 2 3 3 4 4 3 2 In this running example the unitary cost of each column is f(cj, kj) = cj/kj. In Table 4 the unitary cost f(cj, kj) is shown. Table 4. Unitary cost of each column f(cj, kj) Features X1 X2 X3 X4 X5 X6 X7 (jnsq) cj 2 1 1 2 2 3 2 kj 2 3 3 4 4 3 2 cj/kj 1 1/3 1/3 1/2 1/2 1 1 In this example two classes were used. However, the method can be extended to multiple classes. In this extension, observations of different classes should be compared in the generation of the disjoint matrix [Ai,j]. 3.3. Minimum Set Covering Problem In the third step of LAID a heuristic is applied to solve the Set Covering Problem. Given matrix [Ai,j] and variables yj that denote which columns/features are selected, the optimization problem that finds the minimum number of columns/features covering all the rows is the Set Covering problem and can be defined in binary linear programming as: Minimize f= ∑ 𝑐𝑗 .yj Subject to ∑ 𝐴𝑖,𝑗 .y𝑗 ≥1 and y j ∈{0,1} j=1,…,m In this sub-section, a greedy heuristic approach is used to solve the Minimum Set Covering problem, which reuses the algorithm proposed in [Chvatal 1979]. Algorithm 4 is composed of two phases, a constructive phase that adds columns to cover all the lines, and in a second phase removes the redundant columns. 12 Algorithm 4: Set Covering Problem Input: sub-matrix [Ai,j] and vector [cj] Output: a subset of attributes with minimum cost -- Constructive Phase 1. Initialize R = [Ai,j], S=Ø 2. While R ≠ Ø do 2.1. Choose the best line i*∈R such that |Ai*,j)|=min |Ai,j|, ∀j 2.2. Choose the best column j* that covers line i*, considering f(cj, kj) 2.3. Update R and S, R=R\Ai,j*, ∀i, S=S∪{j*} 4. End while -- Remove redundant columns 5. Sort cover S by descending order of costs 6. For each Si do if (S\Si is still a cover) then S=S\Si 7. Return S In the Constructive Phase the algorithm chooses line i* with the fewest elements (2.1). Next it selects column j* that covers line i*, with the lowest cost considering f(cj, kj) (2.2). In the first iteration from features {X1, X2, X3, X4, X5, X6, X7} solution {X2} is obtained with cost=3. This solution is not admissible since there are 3 uncovered lines. In the second iteration X4 is added to the solution, with a global cost of 7, where one covered line is missing. In the third iteration the admissible solution {X2, X4, X7} is found, with cost 9. The constructive phase ends and the redundancies removal phase occurs, where a new admissible solution {X4, X7} is found, with cost 6. Given the subset {X4, X7}, in Table 5 the reduced dataset is presented. The number of features is reduced from 7 to 2 and the number of observations is also reduced from 5 to 3. Nearly all the combinations with 2 features are presented, except for X4=0 and X7=1 because no observations occur. Table 5. Reduced dataset observation\feature X4 X7 (jnsq) Class O1, O2 1 0 0 O3 1 1 1 O4, O5 0 0 1 no observations 0 1 ? 3.4. Leave-one-out cross-validation The major method for supervised learning model validation is cross-validation. Cross- validation technique involves the partition of the sample dataset into n subsamples, using (n-1) for repeated training and the remaining for testing. In this work, we adopt the special case of Leave-One-Out cross-validation, which consists in removing one observation from the original sample, training the algorithm with the remaining observations and then, testing the observation using the resulting model. The input file 13 for the Leave-One-Out cross-validation is the reduced dataset and not the original one, which results in a reduced computational effort. To compare observation i with the testing observation x, we use the Hamming distance, H(i,x). The Hamming distance between the known observation O1=(1,1,0,0) and testing observation x=(1,0,1,0) is H(1,x)=2, since there are two differences, namely in the second and third elements. In the predicting process, the original LAD algorithm that compares observations belonging to two disjoint classes uses the so-called discriminant function, which is similar to the k-nearest neighbor classification algorithm. In the k-nearest neighbor prediction algorithm, k observations are chosen which present the lowest Hamming distance in comparison with the testing observation. As the class of the observation is known, we can count the number of observations of class 0 and class 1. The prediction of the x class is given by the most frequent class, which in statistics is the mode. Therefore, the prediction class of x is the mode of the classes of the k observations: predicted_class(x) = mode (class(O1), class(O2), …, class(Ok)) The performance of a classification model is based on the total counts of correct and incorrect predictions. These counts are tabulated in a table known as Confusion Matrix. The Confusion Matrix is a specific square table, actual class versus predicted class, which allows the visualization of the performance of an algorithm. Based on the Confusion Matrix, many performance measures can be extracted. In this work the accuracy or hit-rate measure is used, expressed by: accuracy = number of correct predictions total number of predictions The sub-problem algorithm returns the pair (number of features, accuracy). The number of features is given by Algorithm 4, the Set Covering Problem, and the accuracy by the Leave-one-out cross-validation. This sub-problem algorithm is used by the Heuristic Decomposition presented in the next section. 4. Heuristic Decomposition and Master-problem Algorithm In many Linear Programming optimization problems, the constraints and variables may be decomposed in blocks, where each constraint set only involves a variable set, with the following structure: Minimize x1 + x2 + x3 + x4 + x5 + x6 Subject to x1 + x2 ≥ 1 x3 + x4 ≥ 1 x5 + x6 ≥ 1 14 As our problem is not so regular, we must resort to a Heuristic Decomposition for the Set Covering problem. The matrix [Ai,j] is divided into max_k sub-problems resulting in max_k sub-matrices [Ai,j] k . The structure of the problem is maintained since each sub- problem returns a feasible (or admissible) solution. Comparing this approach with meta-heuristics the sub-matrices correspond to different neighborhoods in Variable Neighborhood Search or to new regions in the diversification process of Tabu Search, where feasible solutions can be found. Each sub-problem returns the pair (number of features, accuracy) for the bi-criteria feature selection, where the goal is to minimize the number of features and to maximize the accuracy. 4.1. Heuristic Decomposition The original problem with one million features is very hard to solve, given its dimensionality. In the decomposition each sub-problem should return an admissible solution, that is, all lines should be covered by the chosen features/columns. Given the dimension of each sub-problem the value of max_k can be found. As each sub-problem returns a feasible solution, the problem is perfectly separable or independent. A parallel version of the Heuristic Decomposition for the bi-criteria feature selection is implemented. Algorithm 5 reuses the LAID procedure, described in Algorithm 4, and we consider the following notation: nf – number of features ar – accuracy Algorithm 5: Heuristic Decomposition for the Bi-criteria Feature Selection Input: sub-problems A[i,j] 1..max_k Output: Pareto front with pairs (nf, ar) 1. solve the sub-problems: sub-problem 1: (nf 1 , ar 1 ) = LAID (A[i,j] 1 ) sub-problem 2: (nf 2 , ar 2 ) = LAID (A[i,j] 2 ) ... sub-problem max_k: (nf max_k , ar max_k ) = LAID (A[i,j] max_k ) 2. solve the master-problem: find Pareto optimal set ((nf 1 , ar 1 ), (nf 2 , ar 2 ), …, (nf max_k , ar max_k )) In our Heuristic Decomposition for feature selection, using parallel processing, the matrix A[i,j] is divided into max_k sub-matrices. Feasible sub-solutions are returned by the LAID procedure with the metrics, number of features and accuracy. Each pair (number of features, accuracy) corresponds to a point in a bi-objective optimization problem. To find the Pareto optimal set, the master problem bi-criteria decision making is detailed in sub-section 4.2. 15 4.2. Bi-objective Master-problem Algorithm Multi-objective optimization (MOO), or multi-criteria optimization, deals with the optimization of two or more conflicting objectives, subject to a set of constraints. Given the vector x=(x1,…, xm) of decision variables, M is the number of objectives and J the number of constrains, multi-objective optimization can be stated as follows [Collette, Siarry 2011]. Minimize / Maximize fm(x), m = 2,…, M objectives Subject to cj(x) {≤, = ,} 0 j = 1, 2,…, J constrains and xi  0 In multi-objective optimization more than one optimal solution can be obtained, whereas classic optimization has only one objective. Variable S represents the set of feasible solutions associated with equality and inequality constraints. F(x) =(f1(x), f2(x), ..., fm(x)) is the vector of objectives to be optimized. In multi-objective optimization the dominance concept is central. In the maximization problem, the objective vector u=(u1, …, um) dominates v=(v1, …, vm), denoted u > v, if and only if, ui ≥ vi: ∀i, and at least one component v is smaller, ui > vi: ∃i. A solution is non-dominated, or Pareto solution, if and only if, there is no solution that dominates it. In other words, a solution x∗  S is Pareto optimal if for every x ∈ S, F(x) does not dominate F(x∗). The Pareto optimal set, P * , includes all the Pareto solutions, i.e., the set of all solutions whose associated vectors are non-dominated. When plotted in space, non-dominated vectors are collectively known as the Pareto front, PF * . The procedure to generate Pareto front is to compute as many points as possible and then build a surface that includes those points. The surface created by the Pareto front can be linear, convex or concave. The Pareto front is defined as PF∗ = {F(x), x  P∗}. The ideal vector contains the best solutions considering the m objectives separately, at the same point. A point y∗ = (y∗1, y∗2, ..., y∗m) is an ideal vector if it optimizes each objective function fi in F(x). Given a large number of feasible solutions in MOO problems, it is important to differentiate two stages: the optimization of the objective functions and the decision- making process. Three strategies are commonly used: (i) a priori, making decisions before optimizing; (ii) a posteriori, optimizing before making decisions and (iii) compromising between optimizing and decision making. 16 In this work, given the criteria number of features and accuracy, the MOO is reduced to a bi-objective optimization problem. Our approach is formulated as a Bi-objective Feature Selection, such that the aim is to:  minimize the number of features (f1) and  maximize the accuracy of the reduced set of features (f2). Figure 1, adapted from [Talbi 2009], shows a bi-objective optimization. The black circles of the Pareto solution dominate the solutions represented by triangles. The efficient front is given by the curve that includes all the Pareto solutions. In this case we want to minimize f1 and maximize f2. The ideal solution is obtained by the combination of the solution that minimizes f1 with a second solution that maximizes f2. Figure 1. Pareto set to minimize f1 (number of features) and maximize f2 (accuracy) After the generation of the bi-objective sub-solutions generated by the sub-problem, the master problem chooses the non-dominant solutions and includes them in the Pareto optimal set. The Pareto set contains only the solutions with the least number of features and the best accuracy, which substantially reduces the number of bi-criteria solutions that will be analyzed by the decision maker, as in this case, the omics dataset analyst. Regarding the optimization and decision-making stages, we select the a posteriori strategy, which does not require previous information from the decision maker. 4.3. Parallel processing for heuristic decomposition and MOO The rapid increase of powerful processors and fast networks has enabled the emergence of computer clusters, large-scale network of machines (grids) and platforms for high- performance computing (HPC). In terms of designing parallel heuristics, three major parallel levels are identified: the algorithmic level, the interaction level and the solution level [Talbi 2009]. 17  In the algorithm level, different instances of the heuristic can run independently in different processors, and the search is equivalent to the sequential execution. The sub-problem is independent, and the sub-solutions are not shared in the processors. The master problem collects the sub-solutions returned by the sub- problems.  In the interaction level solutions, the sub-problems are independent, although the master problem reuses the sub-solutions by sharing them with the different processors. The main objective of this level is to speed up the algorithm by reducing the search time. This approach can be applied in population-based heuristics.  Finally, at the solution level, the problem is dependent, i.e., not separable, and so processors return partial (or non-feasible) sub-solutions. The master algorithm should ensure the feasibility of the final solution. Regarding the parallel levels, we opt for the algorithm level because only feasible sub- solutions are returned from the sub-problems. In this work, parallel processing perfectly matches large volumes of data and decomposition techniques applied to feature selection objectives and MOO. To deal with large volumes of omics datasets, the decomposition is a strategy that combines with parallel processing. Large omics datasets are parceled out for the heuristic decomposition approach. Then the heuristic sub-problems can run in parallel resulting in a sequence schematized as follows: large omics datasets  heuristic decomposition  parallel processing On the other hand, the feature selection with two objectives influences the MOO approach. MOO collects multiple solutions, which perfectly combine with parallel processing that also deals with many solutions simultaneously. Given the pair (number of feature, accuracy) MOO can be applied. Then MOO can run in parallel processing resulting in the following sequence: pair (number of feature, accuracy)  MOO  parallel processing Parallel processing integrates the two concepts, computing multiple solutions for the decomposed heuristic and for MOO. The result is unified in the Pareto optimal set computed by the master algorithm. 18 5. Computational results To implement the computational results of this algorithm, some choices such as the computational environment, the datasets and the performance measures must be made. The computer programs were written in C language and the GCC/Dev-C++ compiler was used. The computational results were obtained from an Intel Core Duo CPU 3.0 GHz processor with 4.0 GB of main memory running under the Windows 10 operating system. The INCD (National Infrastructure for Distributed Computing, Portugal) computer cluster was used to run the LAID algorithm in parallel. As already stated, since omics datasets with rare diseases information are subject to restricted access, we generated artificial datasets with similar characteristics using SeqSIMLA. The dataset under study is comprised of 1700 observations in the affected group, with class 1, associated with rare disease patients and 300 observations in the control group, with class 0, referring to healthy people. The disease prevalence of 0.015 was chosen. Reference haplotypes and recombination rates from chromosome 1 of the European population of the 1000 Genomes Project were used for data generation [Chung et al. 2014]. Initially 30 variants were chosen as disease sites, of which 19 had allele frequencies larger than 0 (ranging from 0.0001 to 0.447). We use a fixed value of odds ratio of 2.0 for the marginal association with the phenotype for all variants and all non-variant features were removed. In the original dataset the number of attributes is close to one million. The number of patient observations, Oa, is 1700 and the number of observations of controls, Ob, is 300. The generation of matrix [Ai,j], for each observation Oa is compared with all the observations Ob, resulting in a number of lines Oa×Ob, as exemplified in Table 3. The number of lines in the matrix [Ai,j] is 510,000 (1700 x 300) and the final dimension of the matrix is 510,000 x 1,000,000. Two performance measures are used to evaluate results, the computational time and quality of the solutions. The quality solution is given by two criteria, the minimum number of attributes that better explains the dataset and the accuracy of the reduced dataset. 5.1. Computational runtime To study the performance of large matrices the time complexity should be considered to estimate the total runtime. For a two-class problem the number of observations is given by the tuple (O0, O1); (100, 100), (500, 500), (700, 700) and (1700, 300). The number of columns is tested from 100 to 5000 variables. 19 Figure 2. Computational time varying the number of columns and observations In the Decomposition Heuristic for the Set Covering problem, for each sub-matrix [Ai,j] k , the computational time is shown in Figure 2, where a linear behavior can be found between runtime and the number of columns. The total runtime for matrix [Ai,j] with dimension 510,000 x 1,000,000, and k=200, that is 200 sub-matrices [Ai,j] k of 510,000 x 5,000 each, can be estimated in 570,800 seconds, or approximately seven days. Since it is possible to decompose the problem, it can run in a computer parallel environment. In the implementation we distributed the task using 10 machines, resulting in a runtime of 58,000 seconds, or approximately 16 hours. The speedup SN is defined as the time T1 taken to complete a program with one processor divided by the time TN taken to complete the same program with N processors. 𝑆𝑁 = 𝑇1 𝑇𝑁 The efficiency EN using N processors is defined as the speedup SN divided by the number of processors N. 𝐸𝑁 = 𝑆𝑁 𝑁 In our work we opt for the algorithm level, where instances of the heuristic can run independently using different processors, and speedup is linear, with an efficiency EN of 100%, meaning all the processors are being fully used all the time. 5.2. Quality of the solutions The challenge of this work is to solve a feature selection problem with a dataset involving one million attributes. In the computational experiments performed with the sub-dataset with less than 1000 columns, we verified that the admissibility of the solutions is found by merging 3 sub-matrices [Ai,j] k . For sub-datasets with 5000 20 columns, feasible solutions are found for each sub-matrix. Running 200 times, hundreds of possible good solutions are found and then used in the master problem phase. Keeping this in mind, we add a new criterion, the accuracy. As previously stated, the solution quality is given by two objectives, the number of attributes of the feasible reduced set and the accuracy in the cross-validation. The number of attributes is given by the Set Covering heuristic where there are no parameters to point out. To obtain the accuracy in the Leave-One-Out cross-validation, we use k=3 in the k-nearest neighbor prediction algorithm. The Pareto optimal set is the result of the master-problem algorithm. Figure 3 represents the Pareto front in a chart with axis, number of features and accuracy. The Pareto front contains the set of efficient points. However, in this case only point (19, 0.84) belongs to the efficient set, all the other points are dominated by this one. After analyzing the results three solutions are found with the performance of (19, 0.84). The feature selection was accomplished given the reduction from 5000 features to 19 features, which is quite expressive. A scale for classifying the accuracy is: excellent (0.91-1.00), good (0.81-0.90), fair (0.71-0.80), poor (0.61-0.70) and fail (0.51-0.60). So, the accuracy found of 0.84 is considered not excellent but good. Figure 3. Pareto front of the objectives, accuracy and number of features Definitive quality assessment of the sub-dataset must be evaluated by human specialists on rare diseases. A huge reduction is attained, from one million features to three reduced sub-datasets with 19 features each. 21 5.3. Final remarks of the computational experiment The final remark of this study regarding the large volumes of data can be summarized considering three aspects: the data, the algorithm and the environment. Regarding the data, large datasets with many dimensions are becoming the new normal, where the in-memory data access should be replaced by the on-disk data access. The Hierarchical Data Format (HDF5), originally developed at the National Center for Supercomputing Applications, is designed to store and organize large amounts of data that can be accessed on-disk in multiple ways. The new algorithms need low time complexity and problem decomposition is essential to parallelize the problem. The intense use of computational expensive algorithms, like meta-heuristics, should be reduced, giving way to simple heuristics. Furthermore, with meta-heuristics the setting of multiple parameters is nontrivial, while with heuristics the number of parameters is substantially reduced. Finally, the desired environment is in the cloud, running the programs in parallel in High-Performance Computing. The European research roadmap High-Performance Computing 2030 [European Commission 2018] should open new perspectives. 6. Conclusions Given the large datasets available in bioinformatics, the aim of this work is to present a feature selection method able to deal with thousands of observations and millions of attributes. In feature selection, two performance measures are considered: the predictive accuracy and result comprehensibility using the parsimony principle. To preserve the data semantics, the minimum number of attributes that better explain the dataset is found, achieving the maximal actionable knowledge. On the one hand, accuracy is essential to measure the quality of the solution. On the other hand, the consistency measure is a core issue in the result comprehensibility. In this work, we opt for Filter methods for two main reasons, the large volumes of data that should be computed and the data semantics preservation. This document extends the previous work of the LAID algorithm to large datasets, using a bi-objective approach that includes the two performance measures. The sub-problem algorithm returns the performance metric (number of features, accuracy). We present LAID using four steps. Initially, the consistency is assured by the inclusion of the dummy variable “je ne sais quoi”. In the set covering problem, the disjoint matrix [Ai,j] and cost vector [cj] reuse the distance measure concept for the independent criteria in feature selection. Therefore, the algorithm combines consistency measure and the distance measure approaches. To compute the accuracy the Leave-One- Out cross-validation is used. In the Decomposition Heuristic algorithm, the disjoint matrix is divided into sub- matrices [Ai,j] max_k . Each sub-problem algorithm returns a pair of measures, which 22 allows the bi-objective optimization. The algorithm finds multiple feasible solutions that are gathered in the master phase. In the master problem a Pareto optimal set is found selecting the subset of non-dominant solutions and reducing the number of feasible sub- solutions, which facilitate the work of the end user. The parallel processing integrates two concepts, computing multiple solutions for the decomposed heuristic and for MOO. The parallel techniques work perfectly to decompose the large volumes of omics datasets. And the parallel approach also combines with the two objectives of feature selection and the diverse sub-solutions of the bi-objective optimization. The computational runtime for the feature selection with dimension 2,000 observations and 1,000,000 features, running in parallel, seems suitable for user requirements. In the experiment the Pareto optimal set includes only one point, where 19 features are chosen, and the accuracy achieves the percentage of 84%. In future work, following the stabilization of the requirements of the end users, we intend to compare our performance measures with other algorithms. Acknowledgements The authors would like to thank the FCT support UID/Multi/04046/2013. This work used the EGI, European Grid Infrastructure, with the support of the IBERGRID, Iberian Grid Infrastructure, and INCD (Portugal). REFERENCES 1. Almuallim, H., Dietterich, T.G. (1991). Learning with many irrelevant features. In: Proceedings of the 9th National Conference on Artificial Intelligence, MIT Press, pp. 547-552. 2. Boros, E., Hammer, P.L., Ibaraki, T., Kogan, A., Mayoraz, E., Muchnik, I. (2000). An Implementation of Logical Analysis of Data. IEEE Transactions on Knowledge and Data Engineering, vol. 12(2), pp. 292-306. 3. Boyd, S., Xiao, L., Mutapcic, A., Mattingley, J. (2008). Notes on decomposition methods. Notes for EE364B, Stanford University, pp. 1-36. 4. Cavique, L., Mendes, A.B., Funk, M. (2011). Logical analysis of inconsistent data (LAID) for a paremiologic study. In: Processing 15th Portuguese Conference on Artificial Inteligence, EPIA. 5. Cavique, L., Mendes, A.B., Funk, M., Santos, J.M.A. (2013). A feature selection approach in the study of azorean proverbs. In: Exploring Innovative and Successful Applications of Soft Computing, Advances in Computational Intelligence and Robotics (ACIR) Book Series, IGI Global, pp. 38‐58. 6. Cavique L., Mendes A.B., Martiniano H.F.M.C. (2017). A Feature Selection Algorithm Based on Heuristic Decomposition, In: Oliveira E., Gama J., Vale Z., Lopes Cardoso H. (eds) Progress in Artificial Intelligence, EPIA 2017, Porto, Portugal, Lecture Notes in Computer Science, Springer, vol. 10423, pp.525-536. 7. Chandrashekar, G., Sahin, F. (2014). A survey on feature selection methods. Computers and Electrical Engineering, vol. 40(1), pp. 16-28. 8. Chvatal, V. (1979). A greedy heuristic for the set-covering problem. Mathematics of Operations Research, vol. 4, pp. 233-235. 23 9. Chung R.H., Tsai W.Y., Hsieh C.H., Hung K.Y., Hsiung C.A., Hauser E.R. (2014). SeqSIMLA2: Simulating Correlated Quantitative Traits Accounting for Shared Environmental Effects in User-Specified Pedigree Structure. Genetic Epidemiology. vol. 39(1) pp.20-24. 10. Collette Y., P. Siarry (2011). Multiobjective Optimization, Principles and Case Studies, Decision Engineering Series, Springer. 11. Crama, Y., Hammer, P. L., Ibaraki, T. (1988). Cause-effect relationships and partially defined Boolean functions. Annals of Operations Research, vol. 16, pp. 299-326. 12. European Commission (2018). The European declaration on High-Performance Computing, url: https://ec.europa.eu/digital-single-market/en/news/european- declaration-high-performance-computing 13. John, G.H., Kohavi, R., Pfleger. K. (1994). Irrelevant Features and the Subset Selection Problem. In: Proceedings of the 11th International Conference on Machine Learning, ICML 94, pp. 121-129. 14. Joncour, C., Michel. S., Sadykov, R., Sverdlov, D., Vanderbeck, F. (2010). Column generation based primal heuristics. Electronic Notes in Discrete Mathematics, Elsevier, vol. 36, pp. 695–702. 15. Kira, K., Rendell, L.A. (1992). The feature selection problem: traditional methods and a new algorithm. In: Proceedings of 9th National Conference on Artificial Intelligence, pp. 129-134. 16. Liu, H., Yu, L. (2005). Toward integrating feature selection algorithms for classification and clustering. IEEE Transactions on Knowledge and Data Engineering, vol.17, vol. 4, pp. 491-502. 17. Pawlak, Z. (1982). Rough Sets. International Journal of Computer and Information Science, vol. 1, pp. 341-356. 18. Pawlak, Z.: Rough Sets (1991). Theoretical Aspects of Reasoning about Data. Kluwer Academic Publishers, Boston. 19. Peters, J.F. & Skowron A. (2010). Transactions on Rough Sets XI. Lecture Notes in Computer Science / Transactions on Rough Sets, Springer. 20. Polkowski, L. (2002). Rough Sets, Mathematical Foundations. Advances in Soft Computing, Physica-Verlag Heidelberg, Germany. 21. Smet P., Ernst, A., Vanden Berghe, G. (2016). Heuristic decomposition approaches for an integrated task scheduling and personnel rostering problem. Computers and Operations Research, vol. 76, pp. 60-72. 22. Stephens Z.D., S.Y. Lee, F. Faghri, R.H. Campbell, C. Zhai, M.J. Efron, R. Iyer, M.C. Schatz, S. Sinha, G.E. Robinson (2015). Big Data, Astronomical or Genomical?, PLoS Biology vol. 13(7): e1002195, doi:10.1371/ journal.pbio.1002195. 23. Talbi E.G. (2009). Metaheuristics, From Design to Implementation, John Wiley & Sons, Inc. 24. The 1000 Genomes Project Consortium (2015). A global reference for human genetic variation, Nature, vol. 526, pp.68-74. 25. Yao PJ, Chung RH (2016). SeqSIMLA2_exact, simulate multiple disease sites in large pedigrees with given disease status for diseases with low prevalence, Bioinformatics, vol. 32(4), pp. 557-562. 
work_7gmsezddm5cbvhtbtse4f4a3si	----	swbc024.tmp Journal of Agricultural and Applied Economics, 28,1 (July 1996): 180-192 @ 1996 Southern Agricultural Economics Association A Decision Support Aid for Beef Cattle Investment Using Expert Systems Lawrence L. Falconer, Charles R. Long, and James M. McGrann ABSTRACT The beef cattle investment decision provides an excellent opportunity to increase the economic effi- ciency of beef cattle production. The investment questions that face beef cattle producers are of interest to beef cattle producers, educators, and financial institutions involved in lending to beef cattle producing firms. This study develops a decision support aid utilizing expert system technology to assist beef cattle producers in making well-founded investment decisions with respect to the firm’s beef cattle herd. Key Words: beef cattle investment, decision support, expert systems. The development of sound tools that use economic theory to assist beef producers in making breeding cow investment decisions is an important effort. A major study focusing on competitive issues in the beef sector for the 1990s, conducted through the Hubert H. Humphrey Institute of Public Affairs, pointed to this need (Johnson et al.). One conclu- sion reached in that study was that beef producers must lower their costs of beef production to prevent further loss of market share to competing meats. The report concluded: Two factors are likely to be important in lowering cost of production in the future. The first is the need to use the most efficient production technol- ogy available. The second is the need to consoli- date production into even larger units so that all economies of size are realized. The fact that con- solidation into larger units has been taking place at such a rapid rate in recent years suggests that Lawrence Falconer is an assistant professor with the Texas Agricultural Extension Service, Corpus Christi; Charles Long and James McGrann are professors with, respectively, the Texas Agricultural Experiment Station, Overton, and the Texas Agricultural Extension Service, College Station. there are real economies of size at the production level (p. ii). If the concentration of beef production into larger units in order to lower costs of production becomes more important to the ranching firm’s survival in the future, then sound investment decision making by ranchers regarding beef breeding cattle will be key to their economic survival. Initial results of an analysis of data gathered in the National Cattlemen’s Association’s national da- tabase of integrated resource management stan- dardized performance analysis (SPA) for cow-calf producers (McGrann et al. 1992) clearly show that large operations over the 1990 to 1991 period have had lower costs of production vis-?i-vis medium and small sized producers. Pre-tax cost of produc- tion was found to range from an average of $78.36 per cwt for firms with less than 200 head of cows, to $66.81 per cwt for firms with 200–500 head, to $60.34 per cwt for firms with over 500 head. (For discussion of SPA guidelines, refer to McGrann.) Investment analysis related to breeding cows differs from the analysis of machinery or land in- vestment in the uncertainty associated with the bio- logical aspects of beef cow reproduction and mor- Falcone~ Long, and McGrann: Decision Support Aid for Beef Cattle Investment 181 tality. The investment in cows relative to all other investments made by the beef cattle firm is signifi- cant, increasing the importance of the investment decision in beef cattle. SPA data reflect an average total investment of $2,097 per breeding cow on a cost basis. Thus, assuming a value of $700 to $1,000 per COW,investment in breeding cows would range from one-third to one-half of total beef cow herd investment. The beef cattle investment decision provides an excellent opportunity to increase the economic ef- ficiency of beef cattle production. Investment ques- tions that face beef cattle producers—such as whether to raise or purchase replacement cows, whether to expand the herd size through raising or purchasing cows, how to cull cows with regard to expected reproductive performance, and what breed type of cattle to invest in—are of concern to beef cow-calf producers, educators, and financial institutions involved in lending to beef cow-calf producing firms. The primary objective of this study is to develop a decision support aid (DSA) utilizing expert sys- tem technology to assist beef cattle producers in making well-founded investment decisions with re- spect to the firm’s beef cattle herd. The DSA will be used to analyze how the proposed expansion/ contraction in beef cow herd influences a represen- tative firm’s financial condition and performance. Previous Research The asset replacement problem has received atten- tion in the agricultural economics literature for over 30 years. In his 1965 seminal study, Burt developed a model for economic analysis of asset life under conditions in which there was a chance of failure or 10SS, with replacement falling into planned and random categories, In a subsequent study, Perrin in- troduced a general replacement/investment prin- ciple applicable to both appreciating and depreciat- ing assets, and considered theoretical implications of changing discount rates and market prices (sal- vage values). The specific problem of beef cow replacement decisions was initially addressed in the literature when Rogers carried out an empirical investigation into the beef cow replacement decision under con- ditions of certainty. Bentley, Waters, and Shumway expanded the literature with an empirical study of the problem of determining the optimal replace- ment age for beef cows given stochastic elements relating to the probability of producing a market- able calf and the probability that a cow might die in a particular period. Kay and Rister developed two models to examine the effects of income tax policy on the decision to buy or raise replacement beef cows with a fixed herd size. Innes and Carman ana- lyzed the effects of the Tax Reform Act of 1986 on the decision to either raise or purchase beef cow re- placements. The impact of market price uncertainty on the beef cow replacement decision and herd size began to be addressed in the 1980s. Yager, Greer, and Burt developed optimal policies for marketing cull cows through the use of discrete stochastic program- ming, resulting in changes in cow salvage prices, Bentley and Shumway examined the planning of cattle herd size over cycles in beef cattle inventories and prices. Trapp developed investment principles that resulted in separation of the investment and disinvestment decisions that allow for firm expan- sion or contraction through unequai rates of culling and additions. More recent work has focused on the effects of herd management strategies on the beef cow re- placement decision. The replacement decision was examined by Tronstad and Gum with a stochastic dynamic programming model that took into ac- count herd productivity under multiple calving sea- sons and market price uncertainty. A model was de- veloped by Frasier and Pfeiffer that incorporated the effects of different feeding regimes on expected herd productivity. The review of previous research suggests that many factors enter into the beef cow investment de- cision process, most of which contribute to the level of economic efficiency of the beef cow operation. As noted by Long, Cartwright, and Fitzhugh, “The net efficiency of different systems of beef produc- tion is a function of the genetic and environmental inputs and their interaction” (p. 409). The traditional measures of environmental in- puts would include such factors as availability of grazing or raised feed resources. A broader defini- tion of environmental inputs should include the financial position and performance of the firm, which measures the availability and application of 182 Journal of Agricultural and Applied Economics, July 1996 financial resources. However, the previous works do not explicitly examine the relationship of pro- posed investment decisions to the financial perfor- mance and condition of the beef cow-calf produc- ing firm. Further, previous research has not focused on the effect of proposed investment in beef cows on the herd age composition and the resulting im- pact on the overall financial structure of the firm. In the following section, we present the development of a model that will take these and other factors into account when analyzing the impacts of beef cow investment decisions on the firm’s financial condi- tion and performance. Beef Cow Investment Analysis System In this study, we develop a computerized simulation model of the beef cow firm, identified as the Beef Cow Investment Analysis System (BCIAS). This system is primarily a decision support aid (DSA). Decision support implies the use of computers: (a) to assist managers in their decision-making pro- cesses in semistructured tasks; (b) to support, rather than replace, managerial judgment; and (c) to im- prove the effectiveness of decision making rather than its efficiency (Keen and Morton). The DSA’S impact is on decisions in which there is sufficient structure for computer and analytic aids to be of value, but where the manager’s judgment is essen- tial. Under the manager’s control, the DSA should be a supportive tool, i.e., the DSA is not designed to automate the decision process, predefine objec- tives, or impose solutions (Keen and Morton). The BCIAS simulation is a mathematical model of an actual ranching system. A system here can be defined as a collection of entities (such as cows, calves, grazing resources, or financial resources) that act and interact together to accomplish an end—which, in the case of the BCIAS, is improve- ment in the financial position and performance of the firm. The BCIAS employs discrete-event simu- lation, which models a system as it evolves over time with state variables that change only at count- able intervals (Law and Kelton). The BCIAS contains several features designed to address shortcomings of previous beef cow in- vestment models. These features include: (a) intro- duction of stochastic elements, such as calving per- centage, calf death loss, and weaning weights, into the model to account for uncertainty relating to re- production and production parameters; (b) an ex- pert system analysis of the financial position and performance of the firm under baseline and alterna- tive beef cow investment scenarios; (c) decision support for the user regarding output pricing, pro- duction, and reproduction parameters; and (d) in- corporation of standardized production, reproduc- tion, grazing, and raised feed parameter definitions. The BCIAS is comprised of five modules: the beef cow herd inventory (BCHI) module, the beef cow herd resource inventory (BCHRI) module, the beef cow herd cost/price expectations (BCHCPE) module, the beef cow herd investment analysis (BCHIA) module, and the Agricultural Financial Analysis Expert Systems (AFAES). Figure 1 illus- trates how these modules relate to each other. The data processed in the BCHI, BCHRI, and BCHCPE modules are used in the BCHIA module to simulate the projected performance of both the current cow herd and the proposed alternative cow herd over the user-specified planning horizon. This analysis includes examination of the firm’s financial position and performance for both the current herd and the proposed investment, in addition to an eco- nomic analysis of the proposed investment. The BCHI Module The beef cow herd inventory (BCHI) module is de- signed to provide a vehicle for entering reproduc- tive, productive, and financial data that are related to the firm’s current beef cow herd. The measures for production and reproduction efficiency parame- ters used in the BCHI module are based on SPA recommendations (refer to McGrann). The SPA guidelines designate three primary measures for re- productive performance and two primary measures for productive performance within a herd. The pri- mary reproductive performance measures are (a) calving percentage, (b) calf death loss, and (c) calf crop; the primary measures for productive perfor- mance are (a) actual weaning weights, and (b) pounds weaned per exposed female. The production and reproduction measures were used to develop default production parameter estimates that are contained in the BCHI module. These default parameters are intended as examples for users who have the capability to generate such values from their own records, in which case these parameters may be overridden. However, in the Falconer Long, and McGrann: Decision Support Aid for Beef Cattle Investment Beef Cow Herd Inventory Module (BCHI) E L 183 Beef Cow Herd Cost/Price Expectations Module (BCHCPE) Investment Analysis I Module (BCHIA) * Figure 1. Beef Cow Investment Analysis System (BCIAS) modules case where production records are missing, the de- fault parameters can provide a basis for beginning the analysis. The data for estimating default production and reproduction parameters were obtained from a study that included 5,903 calving records for 988 cows at the Texas A&M Research Center at McGregor, Texas. (See Long et al, for details of early management techniques applied to these cattle.) Cows were managed so as to conceive whenever they were physiologically able; cows fail- ing to conceive within 18 months of their last partu- rition date and found to be open were culled. When necessary, cows were also culled for physical un- soundness (McElhenney et al,), Following the work of Greer, Whitman, and Woodward, the Bernoulli distribution was esti- mated by breed type for the calving percentage re- production parameter. The Bernoulli distribution for a random variable x is shown in equation ( 1): (1) f(x; f)) = 6P(1 – 0)’-’, forx=O, l. For this study, the definition for success will be that the cow actually calved within any given pe- riod of time. With the additional assumption of in- dependence between stages of the trials, the Ber- noulli distribution can be extended to become the binomial distribution (Freund and Walpole). Here, we assume that at any specific age, the probability of any cow calving does not depend on any other cow calving that is the same age. The binomial dis- tribution for a random variable x is shown in equa- tion (2): (2) f(x; n, 6) = () n W(1 – 6)’-’, x forx=O, l n.,. ..> For simulation purposes, the number of cows that are available to be exposed would be specified as the number of trials, n, with x being the total num- ber of cows calving. The binomial distribution is used to estimate the calf crop reproduction parameter from the data by breed and age group. The probability of success is defined as a calf born. The simulated number of successes (here, live calves born) is adjusted for death loss and then divided by the simulated num- ber of cows exposed in the corresponding breeding season to arrive at the simulated calf crop or wean- ing percentage. Calf death loss and calf crop reproduction per- formance measures are based on the number of calves born option in the SPA guidelines. Breed type has been shown to be a significant source of 184 Journal of Agricultural and Applied Economics, July 1996 Table 1. Calf Death Loss Percentage and Weaning Weights by Breed Types Percent Mean Weaning Breed Death Loss Weight (lbs.) Angus Brahman Hereford Holstein Jersey Angus X Brahman Angus X Hereford Angus X Holstein Angus X Jersey Brahman X Hereford Brahman X Holstein Brahman X Jersey Hereford X Holstein Hereford X Jersey Holstein X Jersey 4.34 10.58 5.18 4.76 8.21 5.96 7.19 5.81 6.05 4.59 9.34 5.05 3.84 6.29 9.04 426 435 396 503 409 456 435 481 440 457 515 458 490 438 467 Source: Falconer variation for calf survival from birth to weaning (McElhenney et al.). The estimated default calf death loss parameters by breed type are presented in table 1. To arrive at the SPA reproduction effi- ciency measure of calf crop for projection pur- poses, the number of calves born is calculated from starting exposed cow numbers and default calving percentages, to arrive at total calves born. The total number of calves born is then adjusted for calf death loss to arrive at total calves weaned. The SPA primary reproduction efficiency measure for calf crop is calculated by dividing total calves weaned by the total number of exposed cows, The SPA productive performance measure for actual weaning weights is estimated for each cow breed type included in the study conducted by Long et al, The distributions used in the simulation process are based on normal curves, with estimated parameters shown in table 1, The average weaning weight is then simulated as a normal distribution that uses the mean and standard deviation of the samples by breed type, with a user-specified adjust- ment available to be used to correct for age of dam effects. The weights are then multiplied by the calves weaned in the respective breed type and age group, and then summed to arrive at the total pounds of calves weaned. The simulated SPA per- formance measure of pounds weaned per exposed cow is calculated by dividing the total pounds weaned by the number of exposed cows. While cull cow sales is not a primary perfor- mance indicator designated by the SPA guidelines, it is addressed within the guidelines. To estimate weight of cull cows for sale, we drew from the work of Nelsen, Long, and Cartwright. The estimation procedure used here is shown by equation (3) (Brody): (3) Y, = A – Be-”, where Y,is the weight of the animal in kilograms at time t,A is the asymptotic weight, B is a constant of integration, k is the measure of the rate at which the curve is approaching the asymptote, and t is the time in months. This model for weight at any given time in months for breed type and condition score is used in the BCHIA module to calculate sale weight of cull cows by breed type and age group. (See Nelsen, Long, and Cartwright for a more de- tailed discussion of this method.) The BCHRI Module To evaluate the proposed beef cow investment, the resources available for input into the beef cow-calf enterprise must be accurately inventoried. This is accomplished through the beef cow herd resource inventory (BCHRI) module. To examine the perfor- mance of grazing and raised feed resources in the beef cow-calf enterprise, the BCIAS model incor- porates the following two primary measures as de- fined by the SPA guidelines: (a) grazing acres per exposed female, and (b) pounds weaned per acre utilized by the cow-calf enterprise. The simulation model calculates the SPA grazing and raised feed performance grazing acres per exposed female measurement by dividing the number of grazing acres by the simulated number of exposed females. To facilitate the establishment of baseline mea- sures of the firm’s financial condition and perfor- mance and to provide a general format for analysis of projected results from investment in beef cattle, the BCHRI module uses the FINYEAR software package (McGrann, Parker, and Neibergs). FIN- YEAR is a computer program that was created to assist in the development of a set of financial state- ments for a single operating year. The program’s formulation has been closely coordinated with the FalconeC Long, and McGrann: Decision Support Aid for Beef Cattle Investment 185 published Recommendations of the Farm Financial Standards Task Force (Farm Financial Standards Task Force). The task force has developed financial analysis standards and terminology for agricultural businesses. The BCHCPE Module The beef cow herd costJprice expectations (BCHCPE) module is designed to aid the user in the development of cost expectations for grazing and forage production, discount rates used in eco- nomic investment analysis, and output prices for the beef cow enterprise. The output price expecta- tions specifications for the beef cow enterprise used in this study are grouped into three major catego- ries: (a) expectations based on past own informa- tion, (b) subjective expectations of the model user, and (c) expectations based on solutions from a structural agricultural sector model, A naive expec- tations model is used to represent the first category. For subjective price expectations, projected mean cattle prices are elicited from the producer, and then random deviates of historical prices are generated to create the price probability distribution (Rich- ardson et al,). Expected cattle prices for the third category are taken from the AG-GEM model (Pen- son and Taylor), a structural econometric model that specifically links the agricultural sector with the general economy. The BCHIA Module To carry out an economic analysis of the proposed investment for the user-specified planning horizon, the beef cow herd investment analysis (BCHIA) module applies a net present value investment model (as suggested by Barry, Hopkin, and Baker), modified to include a time variant discount rate [equation (4)]: “) “v=-’N”+%++,,,I+[UH where NPV is the net present value, INV is the ini- tial investment, P. is the net cash flows attributed to the investment that can be withdrawn in period n, V~ is the terminal investment value, N is the length of the planning horizon, and i,, is the discount rate. For a single proposed investment, the decision rule for the economic analysis would be to accept the investment on an economic basis if the NPV is greater than zero. In the BCHIA module, the net cash flows and terminal values include both sto- chastic and deterministic elements. The BCHIA module uses the net cash flow from operation of the cattle enterprise as the income flow per period stream in the net present value model. The simulation results also show the range that the simulation generates for net cash flow from opera- tions, allowing the user to measure the range of out- comes as a factor of risk. Assuming independence of cash flows, the standard deviation of the net pres- ent value can also be calculated as shown in equa- tion (5) and used as a measure of risk (Bussey): where V is the variance of the net present value, var, is the variance of the cash flow in time period t,and i, is the discount rate by period, BCHIA Module Output as Input for AFAES Program The output from the BCHIA module is used as in- put into the Agricultural Financial Analysis Expert Systems (AFAES) program. To analyze the firm’s projected performance, the AFAES projected op- erating year performance expert system component uses four years of historic balance sheets, three years of income statements, and three years of statements of cash flow data, along with projected balance sheet, projected income statement, and projected cash flow data. Through this AFAES pro- gram component, a summary report is provided of actual historic and projected measures that are ex- amined, as well as a graphic overall diagnostic analysis of the firm’s financial position and perfor- mance. (For a comprehensive discussion of the AFAES program, see McGrann et al. 1990.) Base Simulation Data To validate the BCIAS, we selected an investment problem facing the Texas A&M University (TAMU) Farm, located in the Brazes River Valley of central Texas. The simulations that were carried out examine two possible courses of action: (a) 186 Journal of Agricultural and Applied Economics, July 1996 Table 2. Probability of Angus X Brahman and Brahman X Hereford Cows Calving Whhin 365 Days of Previous Parturition Age in Years Breed 3 4 5 6 7 8 9 10 11 12 Angus x Brahman 0.086 0.897 0.811 0.833 0.919 0.750 0.704 0.833 0.800 0.500 Brahman X Hereford 0.172 0.797 0.830 0.875 0.878 0.714 0.800 0.682 0.733 0.364 Source: Falconer maintaining the cow herd at present levels, or (b) liquidating the cow-calf enterprise and leasing out the land on which the enterprise was operating. Us- ing the three methods of developing expectations (previously discussed), the BCIAS was run to eval- uate differences in economic analysis results under these options. Results from the simulation using the AG-GEM structural model expectations for the continuation and disinvestment strategies were then processed through the AFAES program. The data used as a basis for the simulations were taken from actual SPA reports for the TAMU Farm. The farm’s cattle are comprised of two main breed types—Beefmaster and Brangus. Thus, two sets of calving probabilities were generated. Be- cause of the breed composition of Beefmaster cattle, which includes Brahman, Hereford, and Shorthorn, the Brahman X Hereford data set from the McGregor Experiment Station was selected to represent the Beefmaster cows, The Brangus were likewise represented by the Angus X Brahman data set from the McGregor Experiment Station because of the Brahman and Angus composition of the Brangus breed. The McGregor data were deemed appropriate to represent the calving performance of the TAMU Farm cows due to the proximity of the TAMU Farm to the McGregor, Texas, area and the overall management practices applied to the ani- mals. These practices included adequate levels of nutrition and the use of artificial insemination, which can substantially improve reproductive per- formance. To generate the required Bernoulli probabilities used in the binomial simulation for the model, a transition matrix moving from parturition probabil- ity data to a probability of parturition by age group under the desired culling strategy was developed for both breed types. The transition probabilities for Angus X Brahman and Brahman X Hereford cows are presented in table 2. The TAMU Farm currently devotes 748 total acres to the beef cow enterprise. The grazing re- sources for the farm include 643 acres of improved perennial pasture, in addition to five acres of annual pastures or forage crops. The farm’s raised feed re- sources include 100 acres of improved perennial grasses devoted to hay production. During fiscal year 1991, the TAMU Farm incurred $102,372 in total direct cash expenses relating to the beef cow- calf enterprise, $2,016 in direct noncash expenses relating to the beef cow-calf enterprise, $3,229 in total indirect cash expenses, and $1,245 in total in- direct noncash expenses. This cost structure will serve as the base for cost structure estimates for all simulations in this study, Table 3 shows the base cattle prices used in the simulation runs, The base for the naive expecta- tions simulation for calf prices was the weighted average actual price received for all calves at the TAMU Farm for 1991, which was $83.85 per cwt. The AG-GEM price expectations simulation for calf prices was generated by using the annual per- centage change in AG-GEM forecasted calf prices applied to the $83.85 per cwt base. Calf prices for the subjective expectations simulation were taken from unpublished survey data (Falconer and Neib- ergs). Highest and lowest expected prices, along with expected prices, were gathered from over 60 cow-calf producers in the spring of 1991 for the 199 1–95 period. The mean of the survey data was used as the parameter input into the subjective price distribution. As with calf prices, the base for the naive expec- tations simulation for cow prices was the weighted average actual price received for all cows at the TAMU Farm for 1991, which was $62.92 per cwt. Falcone~ Long, and McGrann: Decision Support Aid for Beef Ca~tleInvestment 187 Table 3. Base Cattle Prices Used in Simulation Runs Year Simulations 1 2 3 4 5 ---- Calf Prices ($/cwt) ----------- --- Naive Expectations 83.85 83.85 83.85 83.85 83.85 AG-GEM Model 87.20 86.33 86.59 87.02 87.46 Subjective Expectations 93.21 90.21 88.21 87.21 87.21 ---- - Cull Cow Prices ($/cwt) - - - - - - - - - - - Naive Expectations 62.92 62.92 62.92 62.92 62.92 AG-GEM Model 67.29 65.62 64.82 64.06 63.31 Subjective Expectations 55.21 52.21 50.21 49.21 49.21 Table 4. Projected Cow Herd Composition and Production by Year cows Weaned Calf Cull cull Exposed Birth Production Cows Sales Year (head) (head) (lbs.) (head) (lbs.) 1 158 117 51,363 28 32,558 2 158 114 50,052 32 37,204 3 159 111 48,736 36 41,745 4 159 117 51,374 20 23,191 5 159 110 48,298 23 22,008 The AG-GEM price expectations simulation for cow prices was generated by using the annual per- centage change in AG-GEM forecasted cow prices applied to the $62.92 per cwt base. The cow prices for the subjective expectations simulation were generated by adjusting the survey data taken from Falconer and Neibergs for a constant basis. The ba- sis between cow and calf prices was assumed to be $38 per cwt (table 3). The initial composition of the TAMU Farm herd by age and breed type was used as the basis for sim- ulation runs. The culling strategy imposed on the herd was to cull any cow over three years of age that did not calve, and to cull all cows after reaching 11 years of age. The heifer must calve to enter the herd at all, All replacement animals were assumed to be retained from the Beefmaster herd. Table 4 presents the projected composition and production of the herd by year. The projected herd becomes younger under the specified culling and replacement strategy (figure 2), In the initial cattle inventory, 39% of the total cows are less than five years of age, while in the fifth year of the simulation, 69% are less than five years of age. The number of cows under five years of age peaks at 72% in the fourth simulated year. The herd becomes less productive over time (table 5), dropping from 325 pounds weaned per cow ex- posed in the first year of the simulation to 304 pounds in the final year of the five-year planning horizon. This productivity decline is due to lower reproductive performance of the younger cows, pri- marily cows that are calved as two-year-olds, and attempts to rebreed and calve at three years of age. The net farm/ranch income from operation of the beef cattle enterprise is literally the bottom line for the income statement in the BCHIA module. The net income is calculated in this model by sub- tracting accrued direct and indirect expenses from the accrued gross revenue, and then adjusting for total interest expense. The projected accrual- adjusted income statement for the beef cow enter- prise utilizing AG-GEM cost and price expecta- tions is shown in table 6. Investment Analysis Results Simulation results for net farm income from opera- tion of the cattle enterprise were generated for each of the three alternative cattle price expectations methods used in the BCHIA module for the final investment analysis. The AG-GEM forecasted in- terest rate for non-real estate loans was used for the investment analysis in each scenario, The interest rate on non-real estate loans was selected for use as the discount rate because the TAMU Farm does have a small amount of debt which could be paid 188 Journal of Agricultural and Applied Economics, July 1996 Table 5. Standardized Performance Measures for Simulated Herd Meas. Year Unit 1 2 3 4 5 Calving percent % 74.1 72.2 69.8 73.6 69.2 Percent calf crop % 70.9 69.6 67.3 70.4 66.7 Average weaning weight Ibs. 459 455 455 459 456 Pounds weaned per exposed female lbs. 325 317 307 323 304 Total acres per exposed female acres 4.7 4.7 4.7 4,7 4.7 Pounds weaned per acre utilized by cow-calf enterprise lbs. 68.7 66.9 65.2 68,7 64.6 Financial cost (noncalf revenue adjusted) per cwt $ I .94 2.08 2,22 2.47 2.81 120- 100 80 60 40 20 0 ,----- r , .– – --T---- I 1 2 3 4 5 Year Figure 2. Age composition of projected herd down (as an alternative to keeping capital tied up in the cattle operation). The results for net farm income from operation of the cattle enterprise were not encouraging for any of the simulations, and in particular for the AG- GEM expectations simulations. The sharp increase in predicted feed costs early in the planning hori- zon causes a large increase in the total cost struc- ture over the AC-GEM expectations planning sce- nario. This cost increase, coupled with steady to declining cattle prices over the AC-GEM expecta- tions planning scenario, leads to a negative net cash flow from operation of the beef cattle enterprise at the TAMU Farm. The net present value of the net cash flow from operation of the cattle enterprise combined with the market value of the ending cattle inventory is – $238,253 for the AC-GEM ex- pectations planning scenario, The results for net farm income from operation of the cattle enterprise were also negative over the entire period for the naive expectations simula- tions. However, with costs held level, net farm income from cattle operations is higher in com- parison to the AC-GEM expectations planning sce- nario. The net present value of the net cash flow from operation of the cattle enterprise combined with the market value of the ending cattle inventory is –$269,430 for the naive expectations planning scenario. The standard deviation of the net present value of the net cash flow from operations for the naive expectations option is $4,731, which repre- sents the smallest of the standard deviations calcu- lated for all expectations options. The subjective expectations simulations like- wise yielded negative results over the entire period for net farm income from operation of the cattle en- terprise. Due to the expectation of declining calf and cow prices over much of the planning horizon, the subjective model produced the lowest returns of all three expectations scenarios. Under the subjec- tive expectations planning scenario, the net present value of the net cash flow from operation of the cattle enterprise combined with the market value of the ending cattle inventory is – $275,275. The stan- dard deviation of the net present value of the net cash flow from operations for the subjective price expectations option is $8,664—the largest of all the standard deviations for any expectations option. A result of increased variability when em- ploying the subjective expectations option relative to the naive and AC-GEM expectations options is not surprising. Since the subjective expectations option specifically introduces uncertainty in the Falconec Long, and McGrann: Decision Support Aid for Beef Cattle Investment 189 Table 6. Beef Cow Herd Investment Analysis (BCHIA) Module’s Projected Income Statement for AFAES Evaluation of Continued Cattle Operation Strategy Year 1 2 3 4 5 Gross Revemre Total Direct Cash Expenses Total Direct Noncash Expenses Total Direct Operating Expenses Gross Margin Total Indirect Cash Expenses Total Indirect Noncash Expenses Total Indirect Operating Expenses Income After Indirect Expenses Total Interest Cash Expenses Total Interest Noncash Expenses Total Interest Expenses Total Pre-Tax Farm/Ranch Expenses Net FarrM?anch Income from Operation 61,557 115,120 2,151 117,272 (55,715) 3,229 1,245 4,474 (60, 189) o 0 0 121,746 (60, 189) 47,058 121,957 2,229 124,185 (77,127) 3,229 1,245 4,474 (81,601) o 0 0 128,659 (81,601) .($) -- 59,880 128,248 2,311 130,559 (70,679) 3,229 1,245 4,474 (75,153) o 0 0 135,033 (75,153) ---- -. 62,258 135,149 2,401 137,551 (75,293) 3,229 1,245 4,474 (79,767) o 0 0 142,025 (79,767) 56,010 142,453 2,495 144,948 (88,938) 3,229 1,245 4,474 (93,412) o 0 0 149,422 (93,412) Note: The AC-GEM expectations option is used as a baseline for this strategy. output price mechanism by specifying a price dis- tribution, it follows that the subjective expectations option will lead to larger variances in returns rela- tive to the single-valued estimates employed in the AG-GEM and naive expectations scenarios. AFAES Comparison of Two Projected TAMU Farm Strategies Given the negative results generated in the previous investment analysis, two projected strategies for the TAMU Farm were selected for AFAES evaluation and comparison. The first alternative is one of con- tinued operation using the AG-GEM expectations option as a baseline (table 6). The second option is to sell all the cattle and lease the land currently occupied by use of the cattle (table 7). The land utilized by the cattle operation at the TAMU Farm was deemed to have 300 acres suit- able for growing cotton, and was assumed to have a cash lease rate of $40 per acre. The balance of the land utilized by the farm’s cattle operation (448 acres) was deemed to be suitable only as grazing land and was assumed to have a cash lease rate of $16 per acre. The projected income statement for the sell at the end of 1992 strategy is shown in table 7, For purposes of the expert system analysis, operations other than the cattle enterprise at the TAMU Farm were treated using a naive expec- tations approach. Under the continued-operation strategy, the expert system analysis on the projected financial performance of the farm fell from accept- able to unfavorable after the 1993 operating year. The AFAES diagnosis cited the extremely poor profitability of the operation (rating profitability at –29.6 based on a range of – 30 to +30) and deteri- oration in firm growth as reasons for the change to an unfavorable performance rating. In contrast, the AFAES diagnosis gave the TAMU Farm a favorable rating of 15 (from a range of – 15 to + 15) for its liquidity position, but noted that the liquidity position of the firm was not show- ing improvement over time. Because of a lack of debt, the AFAES diagnosis also gave the farm a fa- vorable rating for debt repayment capacity of 12,5 (from a range of –25 to +25), In addition, the di- agnosis produced a favorable rating of 6.7 (from a range of – 20 to +20) for the farm’s solvency posi- tion, but warned that the solvency position of the firm has been declining. Conversely, the AFAES evaluation of the sell at the end of 1992 strategy gave the TAMU Farm acceptable ratings for the entire planning horizon. For example, the diagnosis assigned a favorable rating of 15 (from a range of – 15 to + 15) for the 190 Journal of Agricultural and Applied Economics, July 1996 Table 7. Beef Cow Herd Investment Analysis (BCHIA) Module’s Projected Income Statement for AFAES Evaluation of Sell at End of 1992 Strategy Year 1 2 3 4 5 Gross Revenue Total Direct Cash Expenses Total Direct Noncash Expenses Total Direct Operating Expenses Gross Margin Total Indirect Cash Expenses Total Indirect Noncash Expenses Total Indirect Operating Expenses Income After Indirect Expenses Total Interest Cash Expenses Total Interest Noncash Expenses Total Interest Expenses Total Pre-Tax Farm/Ranch Expenses Net Farm/Ranch Income from Operation 39,687 102,372 2,016 104,388 (64,701) 3,229 1,245 4,474 (69,175) o 0 0 108,862 (69, 175) 19,168 0 0 0 19,168 3,229 1,245 4,474 14,694 0 0 0 4,474 14,694 ($) - - 19,168 0 0 0 19,168 3,229 1,245 4,474 14,694 0 0 0 4,474 14,694 19,168 0 0 0 19,168 3,229 1,245 4,474 14,694 0 0 0 4,474 14,694 19,168 0 0 0 19,168 3,229 1,245 4,474 14,694 0 0 0 4,474 14,694 farm’s liquidity position, and cited strong improve- ment over time, Solvency, repayment capacity, and growth were also given favorable ratings, although AFAES noted slow increases in earned equity as a reason not to give the highest possible ratings for firm growth. The TAMU Farm’s projected profit- abilityy received an extremely low rating of – 14.47 (from a range of – 30 to +30) under the sell at the end of 1992 strategy. While the farm is profitable under this strategy, the rate of return on farm assets was cited as being low, and not showing any im- provement over time. Model Validation The SPA primary performance measures for the TAMU Farm simulation were compared with the latest SPA summary data (McGrann et al. 1992) for validation purposes (table 5). Although the simu- lated TAMU Farm calving and calf crop percent- ages are below the weighted average of the SPA summary, they are within the SPA summary’s re- ported range, as are the simulated average weaning weights, pounds weaned per exposed female, and acres per exposed female. The simulated TAMU Farm pounds weaned per acre utilized by the cow- calf enterprise are above the weighted average of the SPA summary, and are also within the summa- ry’s reported range.. . However, the TAMU Farm simulation shows the current operation’s projected cost of production to be extremely high relative to the SPA summary financial data. The observed range in the SPA sum- mary for weaned calf cost per cwt reflected a low of$31 per cwt and a high of $141 per cwt. In none of the projected years did the TAMU Farm simula- tion have a weaned calf cost less than the highest observed cost. Care should be taken in interpreting these particular projections, since they are depen- dent upon the changes in costs based on AG-GEM forecasts. Summary and Concluding Remarks In this study, we have developed a decision support aid that examines the impact of proposed invest- ment decisions in beef cattle on the firm’s financial performance and position. The model depends heavily on the use of electronic spreadsheets for generation of projections that are used as inputs in a computerized expert system. Data requirements are specific, i.e., the initial inventories of animals and related resources, as well as financial statement information, are required. However, these informa- tion requirements are not overly burdensome, as shown in the example where baseline data were taken from SPA reports and cattle inventories by age group. Falcone6 Long, and McGrann: Decision Support Aid for Beef Cattle Investment 191 The model is currently being used at the county agricultural extension level. While the main goals of this study have been achieved, further work is being conducted with the model to exploit the capa- bilities of the latest generation of computer soft- ware. These enhancements center on making the data transfer between modules more transparent to the user, which will reduce the training time and effort required to generate the analysis. References Barry, P. J., J. A. Hopkin, and C. B. Baker. Financial Management in Agriculture. Danville IL: The Inter- state Printers and Publishers, Inc., 1983. Bentley, E., and C, R. Shumway. “Adaptive Planning Over the Cattle Price Cycle.” S. J. Ag~ Econ. 13(1981):139-48. Bentley, E., J. R. Waters, and C. R. Shumway. “Determin- ing Optimal Replacement Age of Beef Cows in the Presence of Stochastic Elements.” S. J. Agr Econ. 8(1976):13-17. Brody, S. Bioenergetics and Growth. New York: Rein- hold Publishing Co., 1945. Burt, 0. R. “Optimal Replacement Under Risk.” J Farm ~COIZ. 47( 1965):324-45. Bussey, L. E. The Economics of Industrial Projects. Englewood Cliffs NJ: Prentice-Hall, Inc., 1978. Falconer, L. L. “Beef Cow Enterprise Investment/Disin- vestment Analysis with Expert Systems.” Unpub- lished Ph.D. dissertation, Texas A&M University, December 1992. Falconer, L. L., and J. S. Neibergs. “Replacement Cow Market Development Study.” Unpublished paper, Dept. of Agr. Econ., Texas A&M University, 1991. Farm Financial Standards Task Force. Recommendations of the Farm Financial Standards Task Force. Ameri- can Banker’s Association, Washington DC, May 1991. Frasier, W. M., and G. H. Pfeiffer. “Optimal Replace- ment and Management Policies for Beef Cows.” Ame~ J Ag~ Econ. 76(November 1994):847-58. Freund, J. E., and R. E. Walpole. Mathematical Statis- tics, 4th ed. Englewood Cliffs NJ: Prentice-Hall, Inc., 1987. Greer, R. C., R.W. Whitman, and R. R. Woodward. “Esti- mation of Probability of Beef Cows Being Culled and Calculation of Expected Herd Life.” J Animal Sci. 51(1980): 10-19. Innes, R., and H. Carman. “Tax Reform and Beef Cow Replacement Strategy!’ West J. Agz ECCM. 13(1988):254–66. Johnson, D. G., J. M. Conner, T. Josling, A. Schmitz, and G. E. Schuh. Competitive Issues in the Beef Sector: Can Beef Compete in the 1990s? Humphrey Institute Rep. No. 1, Hubert H. Humphrey Institute of Public Affairs, University of Minnesota, 1989. Kay, R. D., and E. Rister. “Income Tax Effects on Beef Cow Replacement Strategy.” S. 1 Agz Econ. 9(1977):169-72. Keen, P. G. W., and M. S. Morton. Decision Support Sys- tems: An Organizational Perspective. Reading MA: Addison-Wesley Publishing Co., 1978. Law, A. M., and W. D. Kelton. Simulation Modeling and Analysis. New York: McGraw-Hill, Inc., 1982. Long, C. R., T. C. Cartwright, and H. A. Fitzhugh, Jr. “Systems Analysis of Sources of Genetic Variation in Efficiency of Beef Production: Cow Size and Herd Management.” J. Animal Sci. 40(1975):409-20. Long, C. R., T. S. Stewart, T. C. Cartwright, and T. G. Jenkins. “Characterization of Cattle of a Five-Breed Diallel: I. Measures of Size, Condition, and Growth in Bulk” 1 Animal Sci. 49(1 979):41 8–31. McElhenney, W. H., C. R. Long, J. F. Baker, and T. C. Cartwright. “Production Characters of First- Generation Cows of a Five-Breed Diallel: Reproduc- tion of Young Cows and Preweaning Performance of Inter Se Calves.” J. Animal Sci. 61(1985):55-65. McGrann, J. M. “Guidelines for Production and Finan- cial Performance Analysis for the Cow-Calf Pro- ducer.” Prepared for the National Cattlemen’s Associ- ation, through Dept. of Agr. Econ., Texas A&M University, 199 I. McGrann, J. M., K. L. Karkosh, L. L. Falconer, and J. S. Neibergs. “AFAES-Agricultural Financial Analysis Expert Systems.” Texas Agr. Exp. Sta., Dept. of Agr. Econ., Texas A&M University, 1990. McGrann, J. M., J. Parker, L. L. Falconer, G. Clary, J. S. Neibergs, and P. Gutierrez. “Standardized Perfor- mance Analysis (SPA) Shows Low Profit, Effects of Economy of Size and Opportunities for Change in the Cow-Calf Sector.” Texas Agr. Ext. Ser., Dept. of Agr. Econ., Texas A&M University, 1992. McGrann, J. M., J. Parker, and J. S. Neibergs. “FlN- YEAR—Farm/Ranch Financial Statement Prepara- tion Package.” Computer program, Texas Agr. Exp. Sta., Dept. of Agr. Econ., Texas A&M University, 1991. Nelsen, T. C., C. R. Long, and T. C. Cartwright. “Post- inflection Growth in Straightbred and Crossbred Cattle: I. Heterosis for Weight, Height, and Maturing Rate.” J. Animal Sci. 55( 1982):280-92. Penson, J. B., and C. R. Taylor. “Modeling the Interface Between Agriculture and the General Economy.” AFPC Policy Work. Pap. No. 90-13, Agricultural and Food Policy Center, Texas Agr. Exp. Sta./Agr. Ext. Ser., Dept. of Agr. Econ., Texas A&M University, 1990, 192 Journal of Agricultural and Applied Economics, July 1996 Perrin, R, K. “Asset Replacement Principles.” Arne~ J. Ag~ Econ. 54(1972):60-67. Richardson, J. W., P. Zimmel, D. Gerloff, D. Anderson, G. Helms, A. Lovell, A. Gray, G. Allen, M. Cochran, S. Wong, D. Justice, and M. Rister. “CIRMAN: ,Crop Insurance Risk-Management Analyzer.” Agricultural and Food Policy Center, Texas Agr. Exp. Sta./Agr, Ext. Ser., Dept. of Agr. Econ., Texas A&M Univer- sity, 1992. Rogers, L. “Economics of Replacement Rates in Com- mercial Herds.’’ Animal,Scici. 34(1972):921–25. Trapp, J.N, “Investment and Disinvestment Principles with Nonconstant Prices and Varying Firm Size Ap- plied to Beef-Breeding Herds.’’ Ame~ J. Ag~ Econ. 68( 1986):69 1-703. Tronstad R., and R. Gum. “Cow Culling Decisions Adapted for Management with CART.” Ame~ J. Ag~ Econ.76(May 1994):237-49. Yager, W. A., R. C. Greer, and O. R. Burt. “Optimal Poli- cies for Marketing Cull Beef Cows.” Ame~ J. Agx Econ. 62(1980):456-67. 
work_7if7t4wnyrcnpcsakzndv4r5s4	----	Automated pattern recognition: self-generating expert systems for the future Journal of Rtsearch of the National Buteau of Standards Volume 90, Number 6, November-December 1985 Automated Pattern Recognition: Self- Generating Expert Systems for the Future Thomas L. Isenhour Utah State University, Logan, UT 84322 Accepted: July 1, 1985 Chemometrics and pattern recognition had their start in chemistry in the late 1960's. The most recent review of the area by Michael DeLaney listed 438 journal articles and books. The three most important areas of future development will be Expert Systems, Relational Data Bases, and Robotics. It should now be possible to combine existing robotics and artificial intelligence software to create a system which will generate its own expert systems using relational data bases. The data will be in the chemical domain and the system I describe we are calling the Analytical Director. The Analytical Director will be an artificial intelligence/robotic expert system for the analytical laboratory. The Analytical Director will develop, test, implement and interpret chemical analysis procedures. It will learn from its own experience, the experience of others and communicate what it has learned to others. The Analytical Director will be a self-generating Expert System. I believe that such systems will, in the future, provide all the advantages of pattern recognition, expert systems and relational data bases in experimental settings. Problems will continue to be defined by human beings, but more and more, the laboratory will design, execute and evaluate its own experiments. Key words: artificial intelligence; chemical analysis; expert systems; pattern recognition; relational data bases; mbotics. Chemometrics and pattern recognition had their start in chemistry in the late 1960's. Two areas of application ap- peared almost simultaneously, those being learning ma- chines and project dendral, both applied to spectroscopic interpretation. The former originated in my research group at the Ugiversity of Washington and have been carried on by us, and my two students, Peter Jurs and Bruce Kowalski, and the latter was developed at the Stanford Artificial Insti- tute by a consortium from Chemistry and Computer Sci- ence. Over the past 15 years a variety of applications have occurred which include the following list and probably oth- ers: statistics, modeling and parameter estimation, resolu- tion, calibration, signal processing, image analysis, factor analysis, pattern recognition, optimization, artificial intelli- gence, graph theory and structure handling, and library searching. The most recent review of the area by Michael DeLaney listed 438 journal articles and books. This was an Analytical Chemistry article which covered the last two years of activ- ity. Clearly, pattern recognition and its applications now have an established place in chemistry. The topic of this presentation is, however, not what has happened up to the point. Rather it is what will happen in the forseeable future. I believe that the three most important areas of future development will be expert systems, rela- tional data bases, and robotics. We will talk a little about each one of these and then go on to deal in detail with what may soon become one of the most, if not the most, sophis- ticated tool that the experimental chemist has ever acquired. An expert system is, simply stated, a piece of software that behaves like an expert. The origin of expert systems were instruction manuals that told you what to do based upon what you encountered in a step by step fashion. 521 About the Author: Thomas L. Isenhour is with Utah State University's Department of Chemistry and Biochem- istry. Almost everyone is familiar with manuals on "How to Fix Your Chevy Station Wagon," etc. These are non- mysterious, and sometimes non-useful, written recipies de- signed to lead you by the hand through a repair, construc- tion, gourmet meal preparation, ecc., that an expert could do but you could not, at least on your own. Of course, the quality of such systems depends on the knowledge of the expert, and on the successful transfer of that knowledge from that expert to the author and from the author to you. The modem expert system is usually a computer program that attempts the same thing, with the same limits of suc- cess. That is, the quality is dependent on the knowledge of the expert, the successful transfer of that knowledge from that expert to the author and from the author to the user. It is a little more, however, because the machine can use your input to do computations, etc., and while all of this could be done with a programmed manual, it certainly is faster by computer and potentially more accurate. There is less chance of you failing to follow the directions for correct use than in a written expert system. Relational data bases are collections of data that interre- late along traditional and non-traditional lines. In a sense, the relational data base is the social scientist's dream. Given a set of descriptors for each entry, the relational data base allows very easy cross correlations such as, how many chil- dren in Miami who had braces before the age of 12 also have maternal grandparents alive in Manhattan. Again, a rela- tional data base was always possible with pencil and paper, or even better with three-by-five cards, but it can be greatly facilitated in a computer. However, the use is still defined by the selection of the descriptors and the quality of the queries. A robot, according to one handy dictionary, is "a machine devised to function in place of a human agent." This is quite a broad definition and could refer to an automatic ticket dispenser at a parking garage, an autopilot operating from an inertial guidance system in a jet airplane, or a computer interfaced with an autoanalyser at a clinical laboratory. It is my contention that it is now possible to combine existing robotics and artificial intelligence software to create a system that will generate its own expert systems using relational data bases. The data will be in the chemical do- main and the system I describe we are calling the Analytical Director. Simply stated, the Analytical Director will be an artificial intelligence/robotic expert system for the analytical laboratory. The Analytical Director will develop, test, im- plement, and interpret chemical analysis procedures. It will learn from its own experience, and that of others, and it will communicate what it has learned to others. The subject of this research is neither automation nor the robot's roll in automation, but rather EXPERT laboratory management working through a combination of robotics and artificial intelligence. We propose to combine robotics and artificial intelligence into an expert system for the analytical chemistry laboratory. We propose to demonstrate that an Analytical Director can develop, test, implement, and inter- pret chemical analysis procedures. It is a misconception that the best use of robots will be exhaustive testing of possible solutions to problems. While computationally exhaustive methods are often quite successful, they are rarely useful to analytical chemistry. Artificial intelligence is required for a real breakthrough in automated laboratory methodology. Consider the following analytical problem which might arise from a relatively simple problem. Given: 10 possible components to a mixture 10 reagents 10 possible temperatures 10 pH values If each reaction combination were chemically independ- ent, that is, if the results of any combination could be learned by a linear addition of the separate tests, then 10,000 procedures could be carried out to determine the entire sys- tem. This might be feasible if, for example, each test could be completed in one minute. (This would require just about one week of continuous work assuming the robot suffered no maintenance problems or other delays.) However, chemical reactions are not usually independ- ent. For example, if one of the components were Fe(ll) and two of the reagents were CSN- and citrate ion, there would clearly be complex equilibria interactions. If we redo the calculation considering from I to 10 possible components, from I to 10 possible reagents and any combination of 10 temperatures and 10 pH values, it requires 1.63 x 1015 tests. Again carried out at one minute intervals, assuming the robot could work through the entire set of procedures with- out interruption, 3. lOx 109 years would be needed. Experi- mental design methods might achieve a few orders of mag- nitude improvement but nothing like the 10 orders of magnitude necessary to make this approach feasible. Furthermore, in real analytical situations the number of variables and dimensions is often much greater. It is clear that analytical chemistry cannot be done by exhaustive trial and error. Therefore, if robotics is to have any real effect upon the field an intelligent robot must be created that can choose meaningful experiments and profit from its experi- ence, as well as the experience of others. We propose to construct such a system and test it initially on a very limited analytical problem to prove chat artificial intelligence can be used to seek efficient paths to complex analytical problems without resorting to exhaustive trial and error. To do so our first model system will be very simple, involving 3 ions, 5 reagents, 2 temperatures and 3 pH's. This system can be exhaustively tested with 4320 proce- dures. Assuming I minute tests, 3 days would be required. Exhaustive testing of this model system will produce a set of observations that will facilitate the development and test- 522 ing of generalized data structures and optimization routines to be used by the Analytical Director for more complex problems. For complex problems, the Analytical Director must develop artificial intelligence methods that circumvent the testing approach. We have selected a developmental domain that is a closed system of simple analytical chemical problems. The domain will be wet and photometric analysis of simple cations. The set of manipulative skills required is purposely limited to the abilities of the Zymark system, and the chemical reactions and spectrophotometric measurements possible limited to those that can be performed with the available equipment. This way, we plan to be able to test thoroughly the creative capabilities of the artificial intelligence programs to be de- veloped for the Analytical Director. We have further se- lected the domain of water analysis for an advanced test of the Analytical Director. Only by using an unbounded prob- lem will we be able to demonstrate the true capability of the Analytical Director. Given a set of standards, reagents, and manipulative skills, the Analytical Director will develop its own set of tests for each individual cation. These data will be stored in a relational data base keyed on ions, reagents, conditions and results of spectroscopic measurements. It will be as- sumed initially that no chemistry is known for these ele- ments or reagents. After developing possible individual tests, these tests will be cross compared to identify likely interfering reactions. Tests will be characterized by their quality as defined by time, expense and reproducibility. As the best compromise of these components is a value judg- ment, an adjustable value coefficient will be developed. Then possible mixture methods will be systematically tested and compared for success. As this proceeds the relational data base will continue to expand. Finally, new unknowns will be introduced into the system to test the ability of the Analytical Director to adapt to new circumstances. At this point literature information will be added to the data base. The Analytical Director will thereby "learn" from the expe- rience of others. Further, the Analytical Director will report developed procedures for possible use by others. In summary, the Analytical Director will be a self- generating expert system. I believe that such systems will, in the future, provide all the advantages of pattern recogni- tion, expert systems, and relational databases in experimen- tal settings. Problems will continue to be defined by human beings, but more and more the laboratory will design, exe- cute, and evaluate its own experiments. 523 
work_7jfzxcfh3bbclafu6up5tw3bmi	----	A parallel graph edit distance algorithm HAL Id: hal-01629290 https://hal-espci.archives-ouvertes.fr/hal-01629290 Submitted on 6 Nov 2017 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. A parallel graph edit distance algorithm Zeina Abu-Aisheh, Romain Raveaux, Jean-Yves Ramel, Patrick Martineau To cite this version: Zeina Abu-Aisheh, Romain Raveaux, Jean-Yves Ramel, Patrick Martineau. A parallel graph edit distance algorithm. Expert Systems with Applications, Elsevier, 2018, 94, pp.41 - 57. �10.1016/j.eswa.2017.10.043�. �hal-01629290� https://hal-espci.archives-ouvertes.fr/hal-01629290 https://hal.archives-ouvertes.fr A parallel graph edit distance algorithm Zeina Abu-Aisheha, Romain Raveauxa, Jean-Yves Ramela, Patrick Martineaua aLaboratoire d’Informatique (LI), Université François Rabelais, 37200, Tours, France Email addresses: firstname.lastname@univ-tours.fr Abstract Graph edit distance (GED) has emerged as a powerful and flexible graph matching paradigm that can be used to address different tasks in pattern recognition, machine learning, and data mining. GED is an error-tolerant graph matching problem which consists in minimizing the cost of the se- quence that transforms a graph into another by means of edit operations. Edit operations are deletion, insertion and substitution of vertices and edges. Each vertex/edge operation has its associated cost defined in the vertex/edge cost function. Unfortunately, Unfortunately, the GED problem is NP-hard. The question of elaborating fast and precise algorithms is of first interest. In this paper, a parallel algorithm for exact GED computation is proposed. Our proposal is based on a branch-and-bound algorithm coupled with a load balancing strategy. Parallel threads run a branch-and-bound algorithm to IFully documented templates are available in the elsarticle package on CTAN. Email addresses: support@elsevier.com (Romain Raveaux), support@elsevier.com (Jean-Yves Ramel), support@elsevier.com (Patrick Martineau) URL: www.elsevier.com (Zeina Abu-Aisheh) Preprint submitted to Parallel Graph Edit Distance November 6, 2017 http://www.ctan.org/tex-archive/macros/latex/contrib/elsarticle explore the solution space and to discard misleading partial solutions. In the mean time, the load balancing scheme ensures that no thread remains idle. Experiments on 4 publicly available datasets empirically demonstrated that under time constraints our proposal can drastically improve a sequential ap- proach and a naive parallel approach. Our proposal was compared to 6 other methods and provided more precise solutions while requiring a low memory usage. 1. Introduction Attributed graphs are powerful data structures for the representation of structured entities. In a graph-based representation, vertices and their at- tributes describe objects (or part of objects) while edges represent interre- lationships between the objects. Due to the inherent genericity of graph- based representations, and thanks to the improvement of computer capaci- ties, structural representations have become more and more popular in the field of Pattern Recognition. Graph edit distance (GED) is a graph matching paradigm whose concept was first reported in (Sanfeliu and Fu, 1983). The basic idea is to find the best sequence of edit operation to transform a graph G1 into another graph G2. The allowed operations are insertion, deletion and/or substitution of vertices and their corresponding edges. GED can be used as a dissimilarity measure for arbitrarily structured and arbitrarily attributed graphs. In contrast to other approaches, it does not suffer from any restrictions and can be applied 2 to any type of graph (including hypergraphs (H. Bunke, 1983)). The main drawback of GED is its computational complexity which is exponential in the number of vertices of the involved graphs. Many fast heuristic GED methods have been proposed in the literature (W. Christmas and Petrou., 1995; Zeng et al., 2009; Fankhauser et al., 2012; Fischer et al., 2013; Serratosa, 2015; Ferrer et al., 2015; Bougleux et al., 2016). However, these heuristic algorithms can only find unbounded suboptimal values. On the other hand, only few exact approaches have been proposed (Tsai et al., 1979; Justice D, 2006; Riesen et al., 2007; Abu-Aisheh et al., 2015). Parallel computing has been fruitfully employed to handle time-consuming operations. Research results in the area of parallel algorithms for solving ma- chine learning and computer vision problems have been reported in (Kumar et al., 1990). These researches demonstrated that parallelism can be ex- ploited efficiently in various machine intelligence and vision problems such as deep learning (Deng and Yu, 2014) or fast fourier transform (Van Loan, 1992). In this paper, we take benefit of parallel computing to solve the exact GED problem. The main contribution of this paper is an exact parallel algorithm based on a load balancing strategy for solving the GED problem. This paper lies in the idea that a parallel execution can help to converge faster to the optimal solution. Our method is very generic and can be applied to directed or undirected fully attributed graphs (i.e., with attributes on both vertices and 3 edges). By limiting the run-time, our exact method provides (sub)optimal solutions and becomes an efficient upper bound approximation of GED. A complete comparative study is provided where 6 exact and approximate GED algorithms were compared on 4 graph datasets. By considering both the quality of the proposed solutions and the speed of the algorithm, we show that our proposal is a good choice when a fast decision is required as in a classification context or when the time is less a matter but a precised solution is required as in image registration. This paper is organized as follows: Section 2 presents the important def- initions necessary for introducing our GED algorithm. Then, Section 3 re- views the existing approximate and exact approaches for computing GED. Section 4 describes the proposed parallel scheme based on a load balancing paradigm. Section 5 presents the experiments and analyses the obtained results. Section 6 provides some concluding remarks. 2. Problem Statements In this section, we first introduce the GED problem which is formally defined as an optimization problem. Secondly, to cope with the inherent complexity of the GED problem, the use of parallel computing is argued. However, the parallel execution of a combinatorial optimization problem is not trivial and consequently the load balancing question is being raised. Finally, the load balancing problem is formally defined and presented to establish the basement of an efficient parallel algorithm. 4 2.1. Graph Edit Distance Problem Attributed graph is defined as a tuple of 4 sets (V , E, µ, ζ) such that: Definition 1. Attributed Graph G = (V,E,µ,ζ) V is a set of vertices E is a set of edges such as E ⊆ V ×V µ : V → LV such that µ is a vertex labeling function which associates a label lV to a vertex vi with vi ∈ V , ∀i ∈ [1,|V |] ζ : E → LE such that ζ is an edge labeling function which associates a label lE to an edge ei with ei ∈ E, ∀i ∈ [1,|E|] Definition 1 allows to handle arbitrarily structured graphs with uncon- strained labeling functions. Labels for both vertices and edges can be given by a set of integers L = {1, 2, 3, · · ·}, a vector space L = Rn and/or a finite set of symbolic labels L = {x,y,z, · · ·}. GED is an error-tolerant graph matching paradigm, it defines the dis- similarity of two graphs by the minimum amount of distortion needed to transform one graph into another (H. Bunke, 1983). GED requires that each vertex/edge of graph G1 is mapped to a distinct vertex/edge of graph G2 or to a dummy vertex/edge. This dummy elements can absorb structural modifications between the involved two graphs. More formally, GED can be defined as follows: Definition 2. Graph Edit Distance Problem GED(G1,G2) = min γ∈Γ(G1,G2) ∑ o∈γ c(o) Where Γ(G1,G2) denotes the set of edit paths transforming G1 = (V1,E1,µ1,ζ1) into G2 = (V2,E2,µ2,ζ2), and c denotes the cost function measuring the 5 strength c(o) of edit operation o. A sequence of edit operations γ that transforms a graph G1 into a graph G2 is commonly referred to as edit path between G1 and G2. In order to represent the degree of modification imposed on a graph by an edit path, a cost function is introduced measuring the strength of the distortions caused by each edit operation. Consequently, the edit distance between graphs is defined by the minimum cost edit path between two graphs. Note that the edit operations on edges can be inferred by edit operations on their adjacent vertices, i.e., whether an edge is substituted, deleted, or inserted, depends on the edit operations performed on its adjacent vertices. An elementary edit operation o is one of vertex substitution (v1 → v2), edge substitution (e1 → e2), vertex deletion (v1 → �), edge deletion (e1 → �), vertex insertion (� → v2) and edge insertion (� → e2) with v1 ∈ V1, v2 ∈ V2, e1 ∈ E1 and e2 ∈ E2. � is a dummy vertex or edge which is used to model insertion or deletion. c(.) is a cost function on elementary edit operations o. The cost function c(.) is of first interest and can change the problem being solved. In (Bunke, 1997; Brun, 2012) a particular cost function for GED was introduced, and it was shown that under this cost function, GED computation is equivalent to the maximum common subgraph problem. Neuhaus and Bunke (Neuhaus and Bunke., 2007) showed that if each elementary operation satisfies the criteria of a distance (separability, symmetry and triangular inequality) then GED is metric. Recently, methods to learn the matching edit cost between graphs 6 have been published (Cortés and Serratosa, 2015). The discussion around the cost functions is beyond the topic of this paper that essentially focuses on the GED computation. 2.2. From GED Problem to Load Balancing Problem GED is a discrete optimization problem that faces the combinatorial ex- plosion curse. The complexity of GED was proven to be NP-hard where the computational complexity of matching is exponential in the number of ver- tices of the involved graphs (Zeng et al., 2009). At run time, the evolution of the size and shape of the search space is irregular and unpredictable. The search space is represented as an ordered tree. For the sake of clarity in the rest of the paper, the term vertex refers to an element of a graph while the term tree node or node represents an element of the search tree. The initial and leaf tree nodes correspond to the initial state and the final acceptable state in the search tree, respectively. Each edge represents a possible way of state change. A combinatorial optimization problem is essentially the problem of finding a minimum-cost path from an initial node to a leaf node in the search tree. More concretely in GED, each tree node is a sequence of edit operations. Leaf nodes are complete edit operations sequences (edit path) while intermediate nodes are partial solutions repre- senting partial edit path. An example of search tree corresponding to GED computation is shown in Figure 1. 7 Figure 1: An incomplete search tree example for solving the GED problem. The first floor represents possible matchings of vertex A with each vertex of the second graph (in blue). A tree node is a partial solution which is to say a partial edit path. The parallelism of combinatorial optimization problems is not trivial. In a parallel combinatorial search application, each thread searches for optimal solutions within a portion of the solution space. The shape and size of the search space change as the search proceeds. Consequently, tree nodes are generated and destroyed at run-time. Portions that encompass the most promising solutions with high prob- ability are expanded in priority and explored exhaustively, while portions that have unfruitful solutions are discarded at run-time. To ensure that par- allel threads are always busy, tree nodes have to be dispatched at run-time. Hence, the local workload of a thread is difficult to predict. The parallel execution of combinatorial optimization problems relies on load balancing strategies to divide the global workload of all threads iter- atively at run-time. From the viewpoint of a workload distribution strat- egy, parallel optimizations fall in the asynchronous communication category where no thread waits another thread to finish in order to start a new task 8 (Bertsekas and Tsitsiklis, 1997). A thread initiates a balancing operation when it becomes lightly loaded or overloaded. The objective of the data distribution strategies is to ensure a fast convergence to the optimal solution such that all the tree nodes are evaluated as fast as possible. In this paper, we propose a parallel GED approach equipped with a load balancing strat- egy. This approach ensures that all threads have the same amount of load and all threads explore the most promising tree nodes first. 2.3. Load Balancing Problem A parallel program is composed of multiple threads, each thread is a processing unit which processes one task. Multiple threads can exist within the same process and share resources such as memory (i.e., the values of its variables at any given moment). In our GED case, the task is to evaluate a tree node (a partial edit path) and to generate its children (the next possible states). A thread performs a task on a set of works. A work is a tree node characterized by its workload. A work cannot be shared between threads, it is the smallest unit of concurrency the parallel program can exploit. Creating a parallel program involves first decomposing the overall com- putation into works and then assigning the works to threads. The decom- position together with the assignment steps is often called partitioning. The assignment on its own is referred to static load balancing. Load balancing algorithms can be broadly categorized into two families: Static and dynamic. 9 2.3.1. Static load balancing Static load balancing algorithms distribute works to threads once and for all, in most cases relying on a priori knowledge about the works and the system on which they run. These algorithms rely on the estimated execu- tion times of works and inter-thread communication requirements. It is not satisfactory for parallel programs that dynamic and/or unpredictable. The problem of static load balancing is known to be NP-hard when the number of threads is greater or equal to 2 (Drozdowski, 2009). 2.3.2. Dynamic load balancing Dynamic load balancing algorithms bind works to threads at run-time. Generally, a dynamic load balancing algorithm consists of three components: A load measurement rule, an initiation rule and a load balancing operation. A very detailed definition of load balancing models can be found in (Xu and Lau, 1997). Load Measurement. Dynamic load balancing algorithms rely on the workload information of threads. The workload information is typically quantified by a load index, a non-negative variable which is equal to zero if the thread is idle or takes an increasing positive value when the load increases (Xu and Lau, 1997). Since the measure of workload would occur frequently, its calculation must be efficient. Initiation Rule. This rule dictates when to initiate a load balancing opera- tion. The execution of a balancing operation incurs non-negligible overhead; 10 its invocation must weight its overhead cost against its expected performance benefit. An initiation policy is thus needed to determine whether a balancing operation will be profitable. Load Balancing Operation. This operation is defined by three rules: Loca- tion, distribution and selection rules. The location rule determines the part- ners of the balancing operation, i.e., the threads to involve in the balancing operation. The distribution rule determines how to redistribute workload among the selected threads. The selection rule selects the most suitable data for transfer among threads. 3. Related Work In this section, an overview of the GED methods presented in the lit- erature is given. Since our goal is to speed up the calculations of GED, parallelism is highly required. Therefore, we also cover the parallel methods, dedicated to solving branch-and-bound (BnB) problems, aiming at getting inspired by some of these works for parallelizing the GED calculations. 3.1. State-of-the-art of Graph Edit Distance The methods of the literature can be divided into two categories depend- ing on whether they can ensure the optimal matching to be found or not. 3.1.1. Exact Graph Edit Distance Approaches A widely used method for edit distance computation is based on the A∗ algorithm (Riesen et al., 2007). This algorithm is considered as a foundation 11 work for solving GED. A∗ is a best-first algorithm where the enumeration of all possible solutions is achieved by means of an ordered tree that is con- structed dynamically at run time by iteratively creating successor nodes. At each time, the node or so called partial edit path p that has the least g(p) + h(p) is chosen where g(p) represents the cost of the partial edit path accumulated so far whereas h(p) denotes the estimated cost from p to a leaf node representing a complete edit path. The sum g(p) + h(p) is referred to as a lower bound lb(p). Given that the estimation of the future costs h(p) is lower than, or equal to, the real costs, an optimal path from the root node to a leaf node is guaranteed to be found (Riesen, 2009). Leaf nodes correspond to feasible solutions and so complete edit paths. In the worst case, the space complexity can be expressed as O(|Γ|) (Cormen et al., 2009) where |Γ| is the cardinality of the set of all possible edit paths. Since |Γ| is exponential in the number of vertices involved in the graphs, the memory usage is still an issue. To overcome the A∗ problem, a recent depth-first BnB GED algorithm, referred to as DF, has been proposed in (Abu-Aisheh et al., 2015). This algorithm speeds up the computations of GED thanks to its upper and lower bounds pruning strategy and its preprocessing step. Moreover, this algorithm does not exhaust memory as the number of pending edit paths that are stored at any time t is relatively small thanks to the space complexity which is equal to |V1|.|V2| in the worst case. In both A∗ and DF, h(p) can be estimated by mapping the unprocessed 12 vertices and edges of graph G1 to the unmapped to those of graph G2 such that the resulting cost is minimal. The unprocessed edges of both graphs are handled separately from the unprocessed vertices. This mapping is done in a faster way than the exact computation and should return a good approxi- mation of the true future cost. Note that the smaller the difference between h(p) and the real future cost, the fewer nodes will be expanded by A∗ and DF. Almohamad and Duffuaa in (Almohamad and Duffuaa, 1993) proposed the first linear programming formulation of the weighted graph matching problem. It consists in determining the permutation matrix minimizing the L1 norm of the difference between adjacency matrix of the input graph and the permuted adjacency matrix of the target one. More recently, Justice and Hero (Justice D, 2006) also proposed a binary linear programming formula- tion of the graph edit distance problem. GM is treated as finding a subgraph of a larger graph known as the edit grid. The edit grid only needs to have as many vertices as the sum of the total number of vertices in the graphs being compared. One drawback of this method is that it does not take into account attributes on edges which limits the range of application. Table 1 synthesizes the aforementioned methods in terms of the size of the graphs they could match, the execution time and the complexity. One can see the complexity of the exact GED in terms of the number of vertices that the methods could match. Based on these facts, researchers shed light on the approximate GED side. 13 Reference Size of Graphs Execution Time Complexity Parallel? (Riesen et al., 2007) 10 10 milliseconds Exponential No (Abu-Aisheh et al., 2015) 15 100000 milliseconds Exponential No (Justice D, 2006) 40 150000 milliseconds Exponential No Table 1: Characteristics of exact graph edit distance methods 3.1.2. Approximate Graph Edit Distance Approaches Variants of approximate GED algorithms are proposed to make GED computation substantially faster. A modification of A∗, called Beam-Search (BS ), has been proposed in (M. Neuhaus and Bunke., 2006). The purpose of BS, is to prune the search tree while searching for an optimal edit path. Instead of exploring all edit paths in the search tree, a parameter s is set to an integer x which is in charge of keeping the x most promising partial edit paths in the set of promising candidates. In (Riesen, 2009), the problem of graph matching is reduced to finding the minimum assignment cost where in the worst case, the maximum number of operations needed by the algorithm is O(n3). This algorithm is referred to as BP. Since BP considers local structures rather than global ones, the op- timal GED is overestimated. Recently, researchers have observed that BP ’s overestimation is very often due to a few incorrectly assigned vertices. That is, only few vertex substitutions from the next step are responsible for ad- ditional (unnecessary) edge operations in the step after and thus resulting in the overestimation of the optimal edit distance. In (Riesen and Bunke, 2014), BP is used as an initial step. Then, pairwise swapping of vertices 14 (local search) is done aiming at improving the accuracy of the distance ob- tained so far. In (Riesen et al., 2014), a search procedure based on a genetic algorithm is proposed to improve the accuracy of BP . These improvements increase run times. However, they improve the accuracy of the BP solution. In (Fischer et al., 2015), the authors propose a novel modification of the Hausdorff distance that takes into account not only substitution, but also deletion and insertion cost. H(V1,V2) is defined as follows: H(V1,V2) =∑ g1 ming2 c̄1(u,v) + ∑ g2 ming1 c̄2(u,v) which can be interpreted as the sum of distances to the most similar vertex in the other graph. This approach allows multiple vertex assignments, consequently, the time complexity is re- duced to quadratic (i.e., O(n2)) with respect to the number of vertices of the involved graphs. In (Riesen, 2015; Bougleux et al., 2016), the GED was shown to be equivalent to a Quadratic Assignment Problem (QAP). In (Bougleux et al., 2016), the QAP formulation of the GED problem is solved by two well-known graph matching methods called Integer Projected Fixed Point method (Leordeanu et al., 2009) and Graduated Non Convexity and Concavity Procedure (Liu and Qiao, 2014). These two approximate methods have been adapted and optimized to solve sub-optimally the GED problem. In (Leordeanu et al., 2009), this heuristic improves an initial solution by try- ing to solve a linear assignment problem and the relaxed QAP where binary constraints are relaxed to the continuous domain. Iteratively, the quadratic formulation is linearly approximated by its 1st-order expansion around the current solution. The resulting assignment helps at guiding the minimization 15 Reference Size of Graphs Execution Time Complexity Parallel? (Riesen, 2009) 100 400 milliseconds Cubic No (Riesen et al., 2007) 70 28 seconds |V1|x No (Fischer et al., 2015) 100 500 milliseconds Quadratic No (Bougleux et al., 2016) 70 27 seconds Cubic No Table 2: Characteristics of approximate GED methods of the relaxed QAP. In (Liu and Qiao, 2014), a path following algorithm aims at approximating the solution of a QAP by considering a convex-concave re- laxation through the modified quadratic function. 3.1.3. Synthesis Table 2 summarizes the aforementioned approximate GED methods. Ap- proximate GED methods often have a polynomial computational time in the size of the input graphs and thus are much faster than the exact ones. Nev- ertheless, these methods do not guarantee to find the optimal matching. On the exact GED side, only few approaches have been proposed to postpone the graph size restriction (Tsai et al., 1979; Justice D, 2006; Riesen et al., 2007; Abu-Aisheh et al., 2015). For all these reasons, we believe that proposing a fast and exact GED algorithm is of great interest. 3.2. State-of-the-art of parallel branch-and-bound algorithms Parallel BnB algorithms have been wildly studied in the past. In this section, we report approaches that have been proposed in the literature to solve BnB in a fully parallel manner. The state-of-the-art in this section is divided into two big families, depending on whether or not the exploration 16 of the search tree is done in a regular way. 3.2.1. Regular Exploration of the Search Space Chakroun et al in (Chakroun and Melab, 2013) put forward a template that transforms the unpredictable and irregular workload associated to the explored BnB tree into regular data-parallel GPU kernels 1. A pool of pend- ing nodes is offloaded to the GPU where each node is evaluated in parallel. Moreover, the branching and pruning steps are performed on the CPU side. In fact, besides equivalent operations, the pruning operator on top of GPU reduces the time of transferring the resulting pool from the GPU to the CPU since the non promising generated sub-problems are kept in the GPU mem- ory and deleted there. The authors of (Boukedjar et al., 2012) presented a CPU-GPU model. When a kernel is launched, the works are assigned to idle threads. Each thread performs computation on only one node of the BnB list. Moreover, each thread has its own register and private local memory. Threads can communicate by means of a global memory. Both (Chakroun and Melab, 2013) and (Boukedjar et al., 2012) solved the irregularity of BnB. However, the explorations in both approaches take longer time. That is be- cause in the first kernel each thread generates only one child at each time while the elimination of branches occurs in the second kernel. An OPEN-MP approaches have been put forward in (Dorta et al., 2003). T threads are established by the master program. Moreover, the master 1Kernels are functions executed by many GPU threads in parallel. 17 program generates tree nodes and put them in a queue. Then T tree nodes are removed from the queue and assigned to each thread. The best solution must be modified carefully where only one thread can change it at any time. The same thing is done when a thread tries to insert a new tree nodes in the global shared queue. In this approach, each thread only takes one node, explores it and at the end of its exploration it sends its result to the master that forwards the message to other slaves if the upper bound is updated. Thus, this model did not tackle the irregularity of BnB. 3.2.2. Irregular Exploration of the Search Space A master-slave parallel formulation of depth-first search was proposed in(Rao and Kumar, 1987). Each thread takes a disjoint part of the search space. Once a thread finishes its assigned part, it steals unexplored nodes of the search space of another thread. A dynamic depth eager scheduling method was proposed in (Neary and Cappello, 2005). In the beginning, a depth parameter is set to 2, which means that all the tasks whose level in the search space is 2 are processed by threads with no further subdivision. Then, each thread works on its associated problems. When a thread runs out of work, it requests work from some thread that it knows. This balances the computational load as long as the number of tasks per thread is high. The communication in (Rao and Kumar, 1987) and (Neary and Cappello, 2005) is asynchronous, and thus threads communicate if they succeed in updating the upper bound. An eager scheduling approach is used to make the tasks 18 balanced depending on the difficulty of each tree node. A master-slave hybrid Depth-First/Best-First was proposed in (Chung et al., 2012). The master thread keeps generating the tree nodes at a pre- determined level (i.e., level i) and saves them to a work pool. Then, each of the worker threads takes a node with the minimum lower bound from the pool and explores it in a depth-first way. Generating and exploring nodes are repeated until finding the solution of the problem. This model has a less communication between threads, however, it is irregular as the works given to threads do not have the same difficulty. In (Allen and Yasuda, 1997), a BnB algorithm for solving inexact graph matching was proposed. This algorithm aims at determining a minimum- distance between two unattributed graphs. At each iteration, each thread takes a node from its queue to be solved by expanding it in a depth-first search way until its branch is fully explored, updating the local best permutation and the corresponding degree of mismatch and eliminating test. Afterwards, a global permutation with its corresponding degree of mismatch is updated and given to all threads when all of them finish solving their chosen nodes or problems, then, a next node is chosen by each thread. A thread becomes inactive when it has no node left in its queue. Load balancing is performed if the number of inactive threads with empty queues is above a threshold T. The best permutation and the best degree of match are only updated at the end of each iteration, such a fact will not prune the search space as fast as possible. 19 3.2.3. Synthesis Based on the aforementioned parallel BnB algorithms and to the best of our knowledge, none of these algorithms addressed the GED problem. We believe that proposing a parallel branch-and-bound algorithm dedicated to solving the GED problem is of great interest since the computational time will be improved. The search tree of GED is irregular (i.e., the number of tree nodes varies depending on the ability of the lower and upper bounds in pruning the search tree) and thus the regular parallel approaches (e.g., (Chakroun and Melab, 2013; Boukedjar et al., 2012; Dorta et al., 2003)) are not suitable for such a problem. The approaches in (Rao and Kumar, 1987), (Chung et al., 2012) and (Neary and Cappello, 2005) are interesting since the communication is asyn- chronous 2 and thus there is no need to stop a thread if it did not finish its tasks, unless another thread ran out of tasks. In (Chung et al., 2012), however, load balancing is not integrated. Thus, when there are no more problems to be generated by the master thread, some threads might become idle for a certain amount of time while waiting the other threads to finish their associated tasks. For GED, load balancing is important to keep the amount of work balanced between all threads. On this basis, we propose a parallel GED method equipped with a load balancing strategy. This paper is considered as an extension of the most 2Asynchronous communication indicates that no thread waits another thread to finish in order to start its new task (Bertsekas and Tsitsiklis, 1997). 20 recent BnB algorithm (DF ). When thinking of a parallel and/or a distributed approach of DF, the edit paths can be considered as atomic tasks to be solved. Edit paths can be dispatched and can be given to threads in order to divide the GED problem into smaller problems. It is hard to estimate the time needed by threads to explore a sub-tree (i.e., to become idle). Likewise, the number of CPUs and/or machines have to be adapted to the amount and type of data that have to be analyzed. Some experiments in Section 5.5 illustrate this point and are followed by a discussion. 4. Proposal: Parallel graph edit distance using a load balancing strategy In this section, an overview of our proposal is given. The main objectives of the approach lie in, first, making sure that all the threads have a work to do. Second, balancing the workload of the threads at run-time. Third, exploring the fruitful partial edit paths of the search tree thanks to the cost estimation of lb(p). Generally speaking, our proposal is inspired by some ideas in (Rao and Kumar, 1987; Neary and Cappello, 2005). A best-first procedure is performed before starting to decompose the search tree (composed of edit paths) into sub-trees. The load balancing procedure occurs when any thread finishes all its assigned edit paths. The algorithm terminates when all threads finish the exploration of their assigned editpaths. Our algorithm, denoted by PDFS, consists of three main steps: Initialization-Decomposition-Assignment, Branch- 21 and-Bound and Load Balancing. Figure 2 pictures the whole steps of PDFS. Figure 2: The main Steps of PDFS 4.1. Initialization, Decomposition and Assignment The objective of this step is twofold. First: dividing the problem into sub-problems. Second, making sure, at the beginning of the method, that all threads have an equivalent workload in terms of number of edit paths and their difficulty. Initialization Procedure. As in DF (Abu-Aisheh et al., 2015), this phase con- sists of 3 steps, each of which aims at speeding up the calculations. First, the vertices and edges cost matrices (Cv and Ce) are constructed, respectively. This step aims at getting rid of re-calculating the distances between attributes when matching vertices and edges of G1 and G2. Let G1 = (V1,E1,µ1,ξ1) and G2 = (V2,E2,µ2,ξ2) be two graphs with V1 = (u1, ...,un) and V2 = (v1, ...,vm). A vertices cost matrix Cv, whose dimension is (n + 2) X (m + 2), is constructed as follows: 22 Cv = c1,1 ... c1,κ c1←� ... ∞ ... ... ... ... ... ... ... ... ... ... ... ... cn,1 ... cn,κ ∞ ... cn→� c�→1 ... ∞ ∞ ... ∞ ∞ ... c�←κ ∞ ... ∞ where n is the number of vertices of G1 and κ is the number of vertices of G2 . Each element ci,j in the matrix Cv corresponds to the cost of assigning the ith vertex of the graph G1 to the j th vertex of the graph G2. The left upper corner of the matrix contains all possible node substitutions while the right upper corner represents the cost of all possible vertices insertions and deletions of vertices of G1, respectively. The left bottom corner contains all possible vertices insertions and deletions of vertices of G2, respectively whereas the bottom right corner elements cost is set to infinity which concerns the substitution of �−�. Similarly, Ce contains all the possible substitutions, deletions and insertions of edges of G1 and G2. Ce is constructed in the very same way as Cv. The aforementioned matrices Cv and Ce are used as an input of the following phase. Second, the vertices of G1 are sorted in order to start with the most promising vertices in G1. BP is applied to establish the initial edit path EP (Riesen, 2009). Afterwards, the edit operations of EP are sorted in ascending 23 order of the matching cost where EPsorted = {u → v} ∀u ∈ V1 ∪{�}. At last, from EPsorted, each u ∈ V1 is inserted in sorted -V1. This operation helps in finding the most promising vertices vi ∈ V1 that will be matched first with the unmatched vertices in V2 to speed up the exploration of the search tree while searching for the optimal solution. Third, a first upper bound (UB ) is computed by BP algorithm as it is relatively fast and it provides reasonable results, see (Riesen, 2009) for more details. Decomposition. Before starting the parallelism, a distribution approach is applied aiming at dispatching the workload or sub-problems among threads. For that purpose, N edit paths are first generated using A∗ by the main thread and saved in the heap. Afterwards, the N partial edit paths are sorted as an ordered tree starting from the node whose lb(p) is minimum up to the most expensive one. Note that N is a parameter of PDFS. Assignment. Let Q be the set of partial solutions outputted by A∗. Assigning partial solutions to parallel threads is equivalent to solving the static load balancing problem stated in Section 2.3.1. Due to the complexity of the problem, we chose to avoid an exact computation of load balancing and we adopted an approximated one. Algorithm 1 depicts the strategy we have followed. Once the partial edit paths are sorted in the centralized heap (line 1), the local list OPEN of each thread is initialized as an empty set (lines 2 to 4). Each thread receives one partial solution at a time, starting from 24 the most promising partial edit paths (line 8). The threads keep taking edit paths in that way until there is no more edit path in the centralized heap (lines 6 to 10). Algorithm 1 Dispatch-Tasks Input: A set of partial edit paths Q generated by A∗ and T threads. Output: The local list OPEN of each thread Ti 1: Q← sortAscending(Q) 2: for Tindex ∈ T do 3: OPENTindex ←{φ} 4: end for 5: i=0 . a variable used for thread’s indices 6: for p ∈Q do 7: index = i % |T| 8: OPENTindex .addTask(p) 9: i++ 10: end for 11: Return OPENTindex ∀ index ∈ 1, · · · , |T| Each thread maintains a local heap to keep the assigned edit paths for exploring edit paths locally. Such an iterative way guarantees the diversity of nodes difficulty that are associated to each thread. 4.2. Branch-and-Bound Method In this section we explain the components of BnB that each thread ex- ecutes on its assigned partial edit paths. First, the rules of selecting edit paths, branching and bounding are described. Second, updating the upper bound and pruning the search tree are detailed. Selection Rule. A systematic evaluation of all possible solutions is performed without explicitly evaluating all of them. The solution space is organized 25 as an ordered tree which is explored in a depth-first way. In depth-first search, each edit path is visited just before its children. In other words, when traversing the search tree, one should travel as deep as possible from node i to node j before backtracking. At each step, the most promising child is chosen. Branching Procedure. Initially each thread only has its assigned editpaths in its local heap set (OPEN ) i.e., the set of the edit paths, found so far. The exploration starts with the first most promising vertex u1 in sorted -V1 in order to generate the children of the selected editpath. The children consist of substituting u1 with all the vertices of G2, in addition to the deletion of u1 (i.e., u1 ⇒ �). Then, the children are added to OPEN. Consequently, a minimum edit path (pmin) is chosen to be explored by selecting the minimum cost node (i.e., min(g(p) + h(p))) among the children of pmin and so on. Starting from the second promising vertex u2 in sorted -V1, the edges of both G1 and G2 are handled. Edges of G1 can be either matched with edges of G2 or deleted while edges of G2 can be either inserted in G1 or matched with edges in G1. However, the decision of whether an edge is inserted, sub- stituted, or deleted is done regarding the matching of their adjacent vertices. That is, the neighborhood of edges dominates their matchings. Edges are handled as follows: Let ui and uj ∈ V1 be matched with vk and vz ∈ V2, respectively (i.e., ui → vk and uj → vz). Based on these matchings. One of the following edge operations is selected: 26 ui uj vk vz ui uj vk vz ui uj vk vz 𝜖 𝜖 eij eij ekz Edges Substitution Edges Deletion Edges Insertion G1 G2 G1 G2 G1 G2 Figure 3: Edge mappings based on their adjacent vertices and whether or not an edge between two vertices can be found • If ∃eij ∈ E1 and ∃ekz ∈ E2 then eij → ekz • If ∃eij ∈ E1 and @ekz ∈ E2 then eij → � • If @eij ∈ E1 and ∃ekz ∈ E2 then � → ekz The search for a better edit path continues through backtracking if pmin equals φ. In this case, the next child of pmin is tried out and so on. Pruning Procedure. As in DF, pruning, or bounding, is achieved thanks to h(p), g(p) and an upper bound UB obtained at node leaves. Formally, for a node p in the search tree, the sum g(p) + h(p) is taken into account and compared with UB. That is, if g(p) + h(p) is less than UB then p can be explored. Otherwise, the encountered p will be pruned from OPEN and a backtracking is done looking for the next promising node and so on until finding the best UB that represents the optimal solution of PDFS. Note that OPEN is a local search tree of each thread. This algorithm differs from A∗ as at any time t, in the worst case, OPEN contains exactly |V1|.|V2| elements and hence the memory consumption is not exhausted. 27 Upper Bound Update. The best upper bound is globally shared by all threads (shared UB ). When a thread finds a better upper bound, the shared UB is updated (i.e., a complete path found by a thread whose cost is less than the current UB ). Heuristic. After comparing several heuristics h(p) from the literature, we selected the bipartite graph matching heuristic proposed in (Riesen, 2009). The complexity of such a method is O({|V1|, |V2|}3 + |E1|, |E2|}3). For each tree node p, the unmatched vertices and edges are handled in a complete independent way. Unmatched vertices of G1 and unmatched vertices of G2 are matched at best by solving a linear sum assignment problem. Unmatched edges of both graphs are handled analogously. Obviously, this procedure allows multiple substitutions involving the same vertex or edge and, therefore, it possibly represents an invalid way to edit the remaining part of G1 into the remaining part of G2. However, the estimated cost certainly constitutes a lower bound of the exact cost. 4.3. Load Balancing and Communication Load Measurement. Each thread i provides some information about its work- load or weight index ωi. Obviously, the number of edit paths in OPEN can be a workload index. However, this choice may not be accurate since BnB computations are irregular with different computational requirements. Sev- eral workload indices can be adapted. One could think about h(p). h(p) can be hard to interpret, it can be small either because the p is close to 28 the leaf node or because p is a very promising solution. To eliminate this ambiguity, instead, one can count the number of vertices in G1 that have not been matched yet. This is done on each edit path in the local heap. In our approach, we have selected the latter. Initiation Rule. An initiation rule dictates when to initiate a load balancing operation. Its invocation decision must appear when a thread workload index ωi reaches a zero value that is to say if the thread is idle. Load Balancing Operation. In parallel BnB computations, each process solves one or more subproblems depending on the decomposite procedure. In our problem, two threads are involved in the load balancing operation: Heavy and idle threads. When a thread becomes idle, the heaviest thread will be in charge of giving to the idle thread some edit paths to explore. All the edit paths of the heavy thread are ordered using their lb(p). The heavy thread distributes the best edit paths between it and the idle thread. This proce- dure guarantees the exploration of the best edit paths first since each thread holds some promising edit paths. Threads Communication. All threads share Ce, Ce, sorted -V1 and UB. Since all threads try to find a better UB, a memory coherence protocol is required on the shared memory location of UB. When two threads simultaneously try to update UB, a synchronization process based on mutex is applied in order to make sure that only one thread can access the resource at a given point in time. 29 5. Experiments This section aims at evaluating the proposed contribution through an experimental study that compares 9 methods in terms of precision, execu- tion time and classification rate on reference datasets. We first describe the datasets, the methods that have been studied and the protocol. Then, the results are presented and discussed. 5.1. Datasets To the best of our knowledge, few publicly available graphs databases are dedicated to precise evaluation of graph matching tasks. However, most of these datasets consist of synthetic graphs that are not representative of PR problems concerning graph matching under noise and distortion. We shed light on the IAM graph repository which is a widely used repository dedi- cated to a wide spectrum of tasks in pattern recognition and machine learning (Riesen, 2008). Moreover, it contains graphs of both symbolic and numeric attributes which is not often the case of other datasets. Consequently, the GED algorithms involved in the experiments are applied to three different real world graph datasets taken from the IAM repository (Riesen, 2008) (i.e., GREC, Mutagenicity (MUTA) and Protein datasets). Continuous attributes on vertices and edges of GREC play an important role in the matching pro- cess whereas MUTA is representative of GM problems where graphs have only symbolic attributes. On the other hand, the Protein database contains numeric attributes on each vertex as well as a string sequence that is used to 30 represent the amino acid sequence. For the scalability experiment, the sub- sets of GREC, MUTA and Protein, proposed in the repository GDR4GED (Abu-Aisheh et al., 2015), were chosen. On the other hand, for the classi- fication experiment, the experiments were conducted on the train and test sets of each of them. In addition to these datasets, a chemical dataset, called PAH, taken from GREYCs Chemistry dataset repository 3, was also integrated in the exper- iments. This dataset is quite challenging since it has no attributes on both vertices and edges. Table 3 summarizes the characteristics of all the selected datasets. These datasets have been chosen by carefully reviewing all the publicly available datasets that have been used in the reference works mentioned in section 3 (LETTER, GREC, COIL, Alkane, FINGERPRINT, PAH, MUTA, PROTEIN and AIDS to name the most frequent ones). On the basis of this review, a subset of these datasets has been chosen in order to get a good representativeness of the different graph features which can affect GED computation (size and labelling): Each dataset has specific edit cost functions. Two non-negative meta parameters are associated to GM: (τvertex and τedge) where τvertex denotes a vertex deletion or insertion costs whereas τedge denotes an edge deletion or insertion costs. A third meta parameter α is integrated to control whether 3https://brunl01.users.greyc.fr/CHEMISTRY/index.html 31 Dataset GREC Mutagenicity Protein PAH Size 4337 1100 660 484 Vertex labels x,y coordinates Chemical symbol Type and amino acid sequence None Edge labels Line type Valence Type and length None vertices 11.5 30.3 32.6 20.7 edges 12.2 30.8 62.1 24.4 Max vertices 25 417 126 28 Max edges 30 112 146 34 Table 3: The Characteristics of the GREC, Mutagenicity, Protein and PAH Datasets. the edit operation cost on the vertices or on the edges is more important. Table 4 demonstrates the cost functions of each of the included datasets as well as their meta parameters. Dataset GREC Mutagenicity Protein PAH τvertex 90 11 11 3 τedge 15 1.1 1 3 α 0.5 0.25 0.75 0.5 Vertex substitution function Extended euclidean distance Dirac function Extended string edit distance 0 Edge substitution function Dirac function Dirac function Dirac function 0 Reference of cost functions (Riesen, 2009) (Riesen, 2009) (Riesen, 2009) (Gauzere et al., 2012) Table 4: The cost functions and meta parameters of the datasets. 5.2. Studied Methods We compared PDFS to five other GED algorithms from the literature. From the related work, we chose two exact methods and three approximate methods. On the exact method side, A∗ algorithm applied to GED problem (Riesen et al., 2007) is a foundation work. It is the most well-known exact method and it is often used to evaluate the accuracy of approximate methods. DF is also a depth-first GED that has been recently published and that beats A∗ in terms of running time and precision (Abu-Aisheh et al., 2015). 32 Moreover, a naive parallel PDFS, referred to as naive-PDFS, is implemented and added to the list of exact methods. The basis of naive-PDFS is similar to PDFS. However, naive-PDFS does not include neither the assignment phase (see Section 4.1) nor the load balancing phase (see Section 4.3). Instead of the assignment phase, a random assignment is applied. naive-PDFS does not include a balancing strategy which means that if a thread Ti finished its assigned nodes, it would be idle during the rest of the execution of naive- PDFS. On the approximate method side, we can distinguish three families of methods, tree-based methods, assignment-based methods and set-based methods. For the tree-based methods, the truncated version of A∗ (i.e., BS - x) was chosen where x refers to maximum number of open edit paths. Among the assignment-based methods, we selected BP . In (Riesen, 2009), authors demonstrated that BP is a good compromise between speed and accuracy. Finally, we picked a set-based method. An approach based on the Hausdorff matching, denoted by H, was proposed in (Fischer et al., 2015). All these methods cover a good range of GED solvers and return a vertex to vertex matching, except H, as well as a distance between two graphs G1 and G2 except the lower bound GED which only returns a distance between two graphs. 5.3. Environment PDFS and naive-PDFS were implemented using Java threads. The eval- uation of both algorithms was conducted on a 24-core Intel i5 processor 33 2.10GHz, 16GB memory. In PDFS, the partial edit paths are sorted in the centralized heap OPEN, as mentioned in Section 4.1. Each thread takes one edit path from OPEN and the edit path is then deleted from OPEN. The CPU only needs to move to the next memory location so the spatial local- ity is exploited at best to reduce cache misses. For sequential algorithms, evaluations were conducted on one core. 5.4. Protocol In this section, the experimental protocol is presented and the objectives of the experiment are described. Let S be a graph dataset consisting of k graphs, S = {g1,g2, ...,gk}. Let M = Me ∪Ma be the set of all the GED methods listed in Section 5.2, with Me = {A∗,DF, PDFS} the set of exact methods and Ma = {BP,BS- 1 ,BS-10 ,BS-100 ,H} the set of approximate methods (where x in BS was set to 1, 10 and 100). Given a method m ∈M, we computed all the pairwise comparisons d(gi,gj) m, where d(gi,gj) m is the value returned by method m on the graph pair (gi,gj) within certain time and memory limits. Two types of experiments were carried out scalability experiment and classication experiment. 5.4.1. Scalability Experiment under Time Constraints In the scalability experiment, several metrics were included: The number of best found solutions and the number of optimal solutions. Moreover, a pro- jection of p on a two-dimensional space (R2) is achieved by using speed-score 34 and deviation-score features where speed and deviation are two concurrent criteria to be minimized. First, for each database, the mean deviation and the mean time is derived as follows: dev p k = 1 m×m m∑ i=1 m∑ j=1 dev(gi,gj) p ∀p ∈ P ∀k ∈ #subsets (1) time p k = 1 m×m m∑ i=1 m∑ j=1 time(Gi,Gj) p and (i,j) ∈ J1,mK2 ∀k ∈ #subsets (2) where dev(Gi,Gj) is the deviation of each d(Gi,Gj) and time(Gi,Gj) is the run time of each d(Gi,Gj). To obtain comparable results between databases, mean deviations and times are normalized between 0 and 1 as follows: deviation scorem = 1 #subsets ∑ S∈subsets devmS max devS (3) time scorem = 1 #subsets ∑ S∈subsets timemS max timeS (4) where max devS and max timeS denote respectively the maximal mean deviation and the maximal mean execution time obtained among all the methods M on dataset S. All these metrics have been proposed in (Abu-Aisheh et al., 2015). This 35 experiment was decomposed of 2 tests: Accuracy Test. The aim was to illustrate the error committed by approxi- mated methods over exact methods. In an ideal case, no time constraint (CT ) should be imposed to reach the optimal solution. Due to the large number of considered matchings and the exponential complexity of the tested algo- rithms, we allowed a maximum CT of 300 seconds. This time constraint was large enough to let the methods search deeply into the solution space and to ensure that many nodes will be explored. The key idea was to reach the optimality, whenever it is possible, or at least to get to the Graal (i.e., the optimal solution) as close as possible. This use case is necessary when it is important to accurately compare images represented by graphs even if the execution time is long. Speed Test. The goal was to evaluate the accuracy of exact methods against approximate methods when time matters. That is to say in a context of very limited time. Thus, for each dataset, we select the slowest graph comparison using an approximate method among BP and H as a first time constraint. Unlike BP and H, BS is not included as it is a tree-search algorithm which could output a solution even under a limited CT . Mathematically saying, CT is defined as follows: CT = max m,i,j {timem(gi,gj)} (5) 36 Where m ∈ Ms /BS, (i,j) ∈ J1,kK2 and time is a function returning the running time of method m for a given graph comparison. This way ensures that BP and H could solve any instance. When the time limit is over, the best solution found so far is outputted by BS as well as the exact GED methods. So time and memory limits play a crucial role in our experiments since they impact such methods. In Table 5, we display the time limits used for each dataset. Dataset GREC MUTA Protein PAH CT (milliseconds) 400 500 400 55 Table 5: Time constraints for accuracy evaluation in limited time This case study is representative of a classification stage where many distances have to be quickly computed. 5.4.2. Classification experiment This part of the experiments aimed at showing the performance of the included methods in classifying the graphs of the test set of each of GREC, MUTA and Protein. PAH is not included since we do not have the classes of the test graphs. Two metrics are proposed: Average time (i.e., the time needed to classify each test graph) and classification rate using 1 nearest neighbor (1-NN). The values of CT were the same ones used in the speed test of the aforementioned experiment (speed test). In all the experiments (i.e., scalability and classification), CM was set to 1GB. Among all the aforementioned methods, we expected A∗ to violate 37 CM specially when graphs get larger. In a small CT context, the number of threads in PDFS was set to 3. The reason is that since CT was quite small, we did not want to lose time decomposing the workload among a big number of threads. Moreover, because of the complexity of the calculation of lb, it was removed from each of BS, A∗, DF and PDFS 5.5. Parameters We study the effect of increasing the number of threads T on both accu- racy and speed of naive-PDFS and PDFS. This test was carried out using a 24-core CPU. T is varied from 2 to 128 threads. Moreover, the effect of several values of N, described in Section 4.1, were studied. Five values of N were chosen: -1, 100, 250, 500 and 1000, where N=-1 represents the de- composition of the first floor in the search tree with all possible branches, N= 100 and 250 moderately perform load balancing while N=500 and 1000 is the exhaustive case where threads have much less time dedicated to load balancing since each thread will be assigned sufficient number of works before the parallelism starts. We expected PDFS to perform better when increasing N up to a threshold where the accuracy of the algorithm is degraded. 5.6. Results In this section, the results are demonstrated along with their discussions. We conducted experiments on the involved datasets, however, for the part of parameters selection, we only show the results on GREC-20 (Abu-Aisheh 38 et al., 2015) since this dataset is representative of the other datasets. Time unit is always expressed in milliseconds. 5.6.1. Number of Threads Table 6 displays the effect of the number of threads |T| on the perfor- mance of PDFS. In Table 6, CPU time is the time spent at working by all the threads. One may notice that increasing |T| resulted in increasing the chance to find a better solution, more optimal solutions and a smaller de- viation as we explored more nodes in a parallel manner. Thus, the overall running time decreased (see Figure 4). Since the machine on which we ran this test has a 24-core processor, there was a saturation when increasing |T|. For example, on 128 threads the deviation became bigger (see Figure 4). On a 24-core machine, 32 and 64 threads had got the best results. In addition, increasing the number of threads also increased the load balancing. Method #best found solutions #optimal solutions Idle Time over CPU Time PDFS-2T 67 48 1.7 ∗10−5 PDFS-4T 79 54 9.5 ∗10−5 PDFS-8T 83 66 2.8 ∗10−4 PDFS-16T 92 69 7.7 ∗10−4 PDFS-32T 94 69 0.011 PDFS-64T 95 68 0.043 PDFS-128T 98 66 0.169 Table 6: The effect of the number of threads on the performance of PDFS. Based on the aforementioned results, |T| is set to 64. For naive-PDFS, the same experiment was conducted. At the end, |T| was set to 128. 39 ● ● ● ● ● ●● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Running time score D e v ia ti o n s c o re PDFS−2threads PDFS−4threads PDFS−8threads PDFS−16threads PDFS−32threads PDFS−64threads PDFS−128threads ● ● ● ● ● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Speed score D e v ia ti o n s c o re PDFS−1st Floor PDFS−100editpaths PDFS−250editpaths PDFS−500editpaths PDFS−1000editpaths Figure 4: Time-deviation score: Left (#Threads), right (# Edit Paths). 5.6.2. Number of Edit Paths Table 7 demonstrates the effect of the number of initial edit paths N on the performance of PDFS. One can remark that N equals 100 was the best choice in terms of the number of best found solutions, number of optimal solutions and deviation. Even though N equals 100 remarkably spent much more time on load balancing, it was still 2.3 times more precise than N equals 1000. The latter represented the least precise results (see Figure 4) which was due to the time spent in dispatching the work among threads before the BnB step started. In the rest of the experiments, the number of initial edit paths is set to 100. For naive-PDFS, N equals 100 also demonstrated the best results. 5.6.3. Comparing PDFS with naive-PDFS In this section we compare both PDFS and naive-PDFS. For comparison needs, both algorithms were executed on 128 threads which is slightly in 40 Method #best found solutions #optimal solutions Idle Time over CPU Time PDFS-1st Floor 82 53 0.067 PDFS-100 EP 85 66 0.067 PDFS-250 EP 85 41 0.053 PDFS-500 EP 82 41 0.023 PDFS-1000 EP 83 40 0.018 Table 7: The effect of the number of edit paths on the performance of PDFS favor of naive-PDFS. The results in Table 8 show that PDFS beat naive-PDFS with 28 more optimal solutions. PDFS is equipped with a load balancing scheme which allows the workload variance to be minimized. The workload variance is defined as the deviation between the threads’s workloads and the average workload of all threads at time t. Reducing the variance is important to make sure that all threads have approximately the same amount of work. One can also observe that PDFS was fully parallel where the CPU time was doubled compared to naive-PDFS. In fact, in naive-PDFS, some threads became idle since they finished they assigned works while other threads continued to explore their assigned edit paths. Method #optimal solutions Mean CPU Time (ms) Mean Variance (ms) naive-PDFS 63 4792444 361157.6 PDFS 91 7275476 49880.62 Table 8: The effect of the number of edit paths on the performance of PDFS when executed on the GREC dataset 5.6.4. Comparing Methods under Constraints In this section, we compare the state-of-the-art methods as well as PDFS under small and large time constraints. 41 Large Time Constraint. Regarding the number of best found solutions and the number of optimal solutions, PDFS always outperformed DF on GREC, MUTA, Protein and PAH, see Figures 5, 6 and 7. On MUTA, the deviation of BP was 20%; this fact confirms that the more complex the graphs the less accurate the answer achieved by BP, see Figure 7(b). BP considers only local, rather than global, edge structure during the optimization process (Riesen, 2009) and so when graphs get larger, its solution becomes far from the exact one. Despite the out-performance of PDFS over BP, H and DF, it did not outperform BS in terms of number of best found solutions, see Figure 5(b). The major differences between these algorithms are the search space and the Vertices-Sorting strategies which are adapted in PDFS and not in BS. Since BP did not give a good estimation on MUTA, it was irrelevant when sorting vertices of G1 resulting in the exploration of misleading nodes in the search tree. Since the graphs of MUTA are relatively large, backtracking nodes took time. However, the difference between BS and PDFS in terms of deviation was only 0.1%. On Protein-30, BS-100 was superior to PDFS in terms of number of best found solutions with 50 better solutions. However, this is not the case of a bigger dataset like Protein-40 where BS-100 outputted unfeasible solutions because of the tremendous size of the search tree and thus PDFS outper- formed it. On average, on all databases and among all methods, PDFS got the best deviation, see Figure 7. Exploring the search tree in a parallel way has an advantage when we are 42 also interested in having more optimal solutions, see Figure 6. Results, in Figure 6, demonstrated that the number of optimal solutions found by PDFS was always equal or greater than the number of optimal solutions found by DF and A∗, except on MUTA-20 where A∗ outperformed it. For instance, on GREC, PDFS found 9.6% more optimal solutions when compared to DF and 10% more optimal solutions on PAH. Note that without time constraints all the exact GED algorithms must find all the optimal solutions except A∗ that has memory bottleneck. Small Time Constraint. Concerning the number of best found solutions, even under a small CT , PDFS outperformed DF where the average difference between DF and PDFS was: 10% on GREC, 16% on MUTA, 15% on Protein and 11% on PAH, see Figure 8. A∗ got the highest deviation rates (around 30% on GREC, 73% on MUTA, 86% on Protein and 51.94% on PAH) since it did not have time to output feasible solutions. Despite the fact that PDFS was among the slowest al- gorithms, it obtained the lowest deviation (0% on both GREC and Protein, 5% on MUTA and 6% on PAH), see Figure 9. BS-100 outputted unfeasible solutions on MUTA-50, MUTA-60, MUTA-70, MUTA-MIX and Protein due to the small CT . 5.6.5. Classification Tests Table 9 shows the methods included in the classification experiments. Different versions of DF and A∗ were tested on each dataset taking into ac- 43 GREC5 GREC10 GREC15 GREC20 GRECMIX Number of Vertices # o f b e s t s o lu ti o n s f o u n d ( in % ) 0 5 0 1 0 0 1 5 0 2 0 0 BP H BS−1 BS−10 BS−100 A* DFS PDFS (a) GREC 10 20 30 40 50 60 70 MIX Number of Vertices # o f b e s t s o lu ti o n s f o u n d ( in % ) 0 5 0 1 0 0 1 5 0 2 0 0 BP H BS−1 BS−10 BS−100 A* DFS PDFS (b) MUTA Protein20 Protein30 Protein40 ProteinMIX Number of Vertices # o f b e s t s o lu ti o n s f o u n d ( in % ) 0 5 0 1 0 0 1 5 0 BP H BS−1 BS−10 BS−100 A* DFS PDFS (c) Protein PAH Number of Vertices # o f b e s t s o lu ti o n s f o u n d ( in % ) 0 5 0 1 0 0 1 5 0 BP H BS−1 BS−10 BS−100 A* DF PDFS (d) PAH Figure 5: Number of best found solution under big time constraint 44 GREC5 GREC10 GREC15 GREC20 GRECmix Number of vertices # o f o p tim a l s o lu tio n s fo u n d ( in % ) 0 5 0 1 0 0 1 5 0 2 0 0 BP H BS−1 BS−10 BS−100 A* DFS PDFS (a) GREC 10 20 30 40 50 60 70 MIX Number of vertices # o f o p tim a l s o lu tio n s fo u n d ( in % ) 0 5 0 1 0 0 1 5 0 2 0 0 BP H BS−1 BS−10 BS−100 A* DFS PDFS (b) MUTA Protein20 Protein30 Protein40 Protein−MIX Number of vertices # o f o p tim a l s o lu tio n s fo u n d ( in % ) 0 5 1 0 1 5 2 0 2 5 3 0 BP H BS−1 BS−10 BS−100 A* DFS PDFS (c) Protein PAH Number of vertices # o f o p tim a l s o lu tio n s fo u n d ( in % ) 0 2 0 4 0 6 0 8 0 BP H BS−1 BS−10 BS−100 A* DF PDFS (d) PAH Figure 6: Number of optimal solutions under big time constraint 45 ● ● ● ● ● ● ● ● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Speed score D e v ia ti o n s c o re BP H BS−1 BS−10 BS−100 A* DF PDFS (a) GREC ● ● ● ● ● ● ● ● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Speed score D e v ia ti o n s c o re BP H BS−1 BS−10 BS−100 A* DFS PDFS (b) MUTA ● ● ● ● ● ● ●● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Speed score D e v ia ti o n s c o re BP H BS−1 BS−10 BS−100 A* DFS PDFS (c) Protein ● ● ●● ● ● ● ● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Speed score D e v ia ti o n s c o re BP H BS−1 BS−10 BS−100 A* DF PDFS (d) PAH Figure 7: Time-deviation score under large time constraint 46 5 10 15 20 MIX Number of Vertices # o f b e s t s o lu ti o n s f o u n d ( in % ) 0 5 0 1 0 0 1 5 0 2 0 0 BP H BS−1 BS−10 BS−100 A* DFS PDFS (a) GREC 10 20 30 40 50 60 70 MIX Number of Vertices # o f b e s t s o lu ti o n s f o u n d ( in % ) 0 5 0 1 0 0 1 5 0 BP H BS−1 BS−10 BS−100 A* DFS PDFS (b) MUTA 20 30 40 MIX Number of Vertices # o f b e s t s o lu ti o n s f o u n d ( in % ) 0 5 0 1 0 0 1 5 0 2 0 0 BP H BS−1 BS−10 BS−100 A* DF PDFS (c) Protein PAH Number of Vertices # o f b e s t s o lu ti o n s f o u n d ( in % ) 0 5 0 1 0 0 1 5 0 BP H BS−1 BS−10 BS−100 A* DF PDFS (d) PAH Figure 8: Number of best found solution under small time Constraint 47 ● ●● ● ● ● ● ● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Speed score D e v ia ti o n s c o re BP H BS−1 BS−10 BS−100 A* DFS PDFS (a) GREC ● ● ● ● ● ● ● ● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Speed score D e v ia ti o n s c o re BP H BS−1 BS−10 BS−100 A* DFS PDFS (b) MUTA ● ● ● ● ● ● ●● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Speed score D e v ia ti o n s c o re BP H BS−1 BS−10 BS−100 A* DF PDFS (c) Protein ● ● ● ● ● ● ● ● 0.0 0.2 0.4 0.6 0.8 1.0 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Speed score D e v ia ti o n s c o re BP H BS−1 BS−10 BS−100 A* DF PDFS (d) PAH Figure 9: Time-deviation score under small time constraint. 48 count different combinations of lb and UB. Afterwards, the best combination is selected to be compared with the other methods. Acronym Details DF -UB-LB DF without upper bound and with h(p)=0. DF -UB-LB DF without UB and with h(p)=lb2. DF -UB-LB DF with an initial UB equals to BP, h(p)=0. DF-UB-LB DF with an initial UB equals to BP and lb2 PDFS Parallel GED with the best parameters of DF A∗-LB the A∗ algorithm with lb2 A∗ the A∗ algorithm without lb2 BS-1, BS-10 and BS-100 Beam Search with OPEN size = 1, 10 and 100, re- spectively BP The bipartite GM H The hausdorff algorithm. Table 9: Methods included in the classification experiments Table 10 shows the classification results on GREC and Protein. On GREC, DF with all its variants obtained the same classification rate as BP (i.e., 0.985) even the one without upper and lower bounds (i.e., DF -UB-LB). That shows that DF can also be used to classify graphs even without being obliged to wait for the final, or optimal, solution. DF-UB-LB was the fastest compared to all the variants. This fact shows the importance of UB and LB to make the algorithm faster. Accordingly, and since PDFS is an extension of DF, not all the variants of PDFS are tested. That is, only PDFS-UB-LB has been included in the tests. PDFS-UB-LB was 29% faster than DF-UB- 49 LB. Despite the fact that H was the worst algorithm when evaluating its distances, it was among the algorithms whose classification rate were high. One can see that, on GREC, H beat both BS-10 and BS-100. A∗-LB ob- tained better classification rate than A∗. A∗’ lower bound is time consuming and consequently the number of unfeasible solutions was high. On Protein, one can see a different behavior (see Table 10). DF-UB -LB was the fastest while DF-UB-LB was the slowest. That is because of the time consumed to calculate distances using the cost functions of Protein. Thus, as on GREC, PDFS-UB-LB was included in the tests. Despite the slowness of DF-UB-LB, it was also the best algorithm in terms of classification rate. PDFS-UB-LB was 36% faster than DF-UB-LB. Even though BS took rela- tively enough time to classify graphs (compared to DF ), it was way far from the results obtained by DF. A∗ was not able to find feasible solutions of each pair of graphs. That was not the case of all the variants of DF as they were always able to output feasible solutions before halting. Computing lb(p) and a first upper bound UB was time consuming on such a large database. Since CT of MUTA was set to 500ms, we kept only DF - UB-LB and A∗-LB. Results showed that DF -UB-LB was twice as slow as BP, however, both of them succeeded in finding the best classification rate (i.e., approximately 0.70). PDFS -UB-LB was also able to find the same classification rate and was 40% faster than DF -UB-LB. From all the aforementioned results, one can conclude that even if the deviation of DF and PDFS was better when compared to BP, it did not have 50 GREC Protein R Time (ms) R Time (ms) DF -UB-LB 0.98 171401.54 0.44 128469.57 DF -UB-LB 0.98 163979.45 0.52 124361.61 DF -UB-LB 0.98 140675.00 0.40 147371.86 DF -UB-LB 0.98 140525.48 0.52 145779.68 PDFS-UB-LB 0.98 99850.79 0.52 80038.33 A∗-LB 0.89 358158.76 0.29 1065106.80 A∗ 0.53 222045.94 0.26 194021.88 BS1 0.98 69236.34 0.24 129571.76 BS10 0.94 83928.21 0.26 139294.88 BS100 0.58 83928.20 0.26 141265.41 BP 0.98 62294.60 0.52 59041.84 H 0.96 63563.74 0.43 71990.62 Table 10: Classification on GREC and Protein. The best exact and approximate methods are marked in bold style. Note that the response time is the average time needed to classify each test graph MUTA R Time (ms) DF -UB-LB 0.70089 1139134.29 PDFS-UB-LB 0.70 760861.51 A∗-LB 0.4574 856793.020 BS-1 0.55 1015688.00 BS-10 0.55 1256793.02 BS-100 0.55 1383838.66 BP 0.70 528546.64 H 0.58 525610.25 Table 11: Classification on MUTA. The best exact and approximate methods are marked in bold style 51 an effect on the classification rate. In other words, for such an application, one does not need to have a very accurate algorithm in order to obtain a good classification rate. 6. Conclusion and Perspectives In the present paper, we have considered the problem of GED computa- tion for pattern recognition. GED is a powerful and flexible paradigm that has been used in different applications in PR. The exact algorithm A∗, pre- sented in the literature suffers from high memory consumption and thus is too costly to match large graphs. In this paper, we propose a parallel exact GED algorithm, referred to as PDFS, which is considered as an extension of a recent GED method based on depth-first tree search (Abu-Aisheh et al., 2015). The algorithm in (Abu-Aisheh et al., 2015), referred to as DF, does not exhaust memory as the space complexity in the worst case is quadratic in the number of vertices i.e., O(|V1| × |V2|). In this paper, we speed up the computation of DF by adopting a load balancing strategy. Each thread gets one or more partial edit path and all threads solve their assigned edit paths in a fully parallel manner. A work stealing or balancing process is performed whenever a thread finishes all its assigned threads. Moreover, synchronization is applied in order to ensure upper bound coherence. In the experiments part, we proposed to evaluate both exact and approx- imate GED approaches under large and small time constraints, on 4 publicly available datasets (GREC, MUTA, Protein and PAH). Such constraints are 52 devoted to accuracy and speed tests, respectively. Small time constraints ensured that the approximate methods BP and H were able to find a so- lution. Experiments demonstrated the importance of the load balancing strategy when compared to a naive method that does not include neither static nor dynamic load balancing. Under small and large time constraints, PDFS proved to have the minimum deviation, the maximum number of best found solutions and the maximum number of optimal solutions. However, since our goal was to elaborate methods dealing with rich and complex at- tributed graphs, BS was slightly superior to PDFS in terms of deviation when evaluated on the MUTA dataset under large time constraint. This could be improved by learning the best sorting strategy for a given database. Results also indicated that there is always a trade-off between deviation and running time. In other words, approximate methods are fast, however, they are not as accurate as exact methods. On the other hand, DF and PDFS take longer time but lead to better results (except on MUTA). By limiting the run-time, our exact method provides (sub)optimal solutions and becomes an efficient upper bound approximation of GED with the same classification rate found by the best approximate method. Even though DF and so PDFS were more accurate than PDFS, their classification rate was as good as the best approximate GED (i.e., BP ). On this basis, one could ask: What is the benefit of having a more precise algorithm in an a classification context? A future promising work could be to make PDFS more scalable to have more precise and thus more optimal solutions under large time constraints. 53 This could be achieved by extending PDFS from a single machine algorithm to a multi-machines one. Moreover, learning to sort vertices of G1 based on the structure and characteristics of graphs is another promising perspective towards a faster exact algorithm. References Abu-Aisheh, Z., Raveaux, R., Ramel, J., 2015. A graph database repository and performance evaluation metrics for graph edit distance. In: Graph- Based Representations in Pattern Recognition - GbRPR 2015. pp. 138– 147. Abu-Aisheh, Z., Raveaux, R., Ramel, J.-Y., Martineau, P., 2015. An exact graph edit distance algorithm for solving pattern recognition problems, 271–278. Allen, R., C. L. M. S. T. S. S. L., Yasuda, D., 1997. A parallel algorithm for graph matching and its maspar implementation. IEEE Transactions on Parallel and Distributed Systems 8 (5), 490–501. Almohamad, H. A., Duffuaa, S. O., 1993. A linear programming approach for the weighted graph matching problem. IEEE Trans. Pattern Anal. Mach. Intell. 15 (5), 522–525. Bertsekas, D. P., Tsitsiklis, J. N., 1997. Parallel and Distributed Computa- tion: Numerical Methods. Athena Scientific. 54 Bougleux, S., Brun, L., Carletti, V., Foggia, P., Gauzerec, B., Vento, M., 2016. Graph edit distance as a quadratic assignment problem. Pattern Recognition Letters. Boukedjar, A., Lalami, M., El-Baz, D., 2012. Parallel branch and bound on a cpu-gpu system. In: Parallel, Distributed and Network-Based Processing (PDP). pp. 392–398. Brun, L., 2012. Relationships between graph edit distance and maximal com- mon structural subgraph. Bunke, H., 1997. On a relation between graph edit distance and maximum common subgraph. Pattern Recognition Letter. 18, 689–694. Chakroun, I., Melab, N., 2013. Operator-level gpu-accelerated branch and bound algorithms. In: ICCS. Vol. 18. Chung, C.-S., Flynn, J., Sang, J., 2012. Parallelization of a branch and bound algorithm on multicore systems. journal of Software Engineering and Ap- plications 5, 12–18. Cormen, T. H., et al., 2009. Introduction to Algorithms, 3rd Edition. The MIT Press. Cortés, X., Serratosa, F., 2015. Learning graph-matching edit-costs based on the optimality of the oracle’s node correspondences. Pattern Recognition Letters 56, 22–29. 55 Deng, L., Yu, D., 2014. Deep learning: Methods and applications. Found. Trends Signal Process. 7, 197–387. Dorta, I., Leon, C., Rodriguez, C., 2003. A comparison between mpi and openmp branch-and-bound skeletons. In: Parallel and Distributed Pro- cessing Symposium, 2003. Proceedings. International. pp. 66–73. Drozdowski, M., 2009. Scheduling for Parallel Processing, 1st Edition. Springer Publishing Company, Incorporated. Fankhauser, S., Riesen, K., Bunke, H., Dickinson, P. J., 2012. Suboptimal graph isomorphism using bipartite matching. IJPRAI 26 (6). Ferrer, M., Serratosa, F., Riesen, K., 2015. A first step towards exact graph edit distance using bipartite graph matching. In: Graph-Based Representa- tions in Pattern Recognition - 10th IAPR-TC-15 International Workshop. pp. 77–86. Fischer, A., Suen, C. Y., Frinken, V., Riesen, K., Bunke, H., 2013. A fast matching algorithm for graph-based handwriting recognition. Graph-Based Representations in Pattern Recognition, 194–203. Fischer, A., Suen, C. Y., Frinken, V., Riesen, K., Bunke, H., 2015. Ap- proximation of graph edit distance based on hausdorff matching. Pattern Recognition 48 (2), 331–343. Gauzere, B., Brun, L., Villemin, D., 2012. Two new graphs kernels in chemoinformatics. Pattern Recognition Letters 33 (15), 2038 – 2047. 56 H. Bunke, G. A., 1983. Inexact graph matching for structural pattern recog- nition. Pattern Recognition Letters. 1, 245–253. Justice D, H. A., 2006. A binary linear programming formulation of the graph edit distance. IEEE Trans Pattern Anal Mach Intell. 28, 1200–1214. Kumar, V., Gopalakrishnan, P. S., Kanal, L. N. (Eds.), 1990. Parallel Al- gorithms for Machine Intelligence and Vision. Springer-Verlag New York, Inc., New York, NY, USA. Leordeanu, M., Hebert , M., Sukthankar, R., 2009. An integer projected fixed point method for graph matching and map inference. In: Proceedings Neural Information Processing Systems. pp. 1114–1122. Liu, Z., Qiao, H., 2014. GNCCP - graduated nonconvexityand concavity procedure. IEEE Trans. Pattern Anal. Mach. Intell. 36, 1258–1267. M. Neuhaus, K. R., Bunke., H., 2006. Fast suboptimal algorithms for the computation of graph edit distance. Proceedings of 11th International Workshop on Structural and Syntactic Pattern Recognition. 28, 163–172. Neary, M. O., Cappello, P. R., 2005. Advanced eager scheduling for java- based adaptive parallel computing. Concurrency - Practice and Experience 17, 797–819. Neuhaus, M., Bunke., H., 2007. Bridging the gap between graph edit distance and kernel machines. Machine Perception and Artificial Intelligence. 68, 17–61. 57 Rao, V. N., Kumar, V., 1987. Parallel depth-first search on multiprocessors part i: Implementation. International journal on Parallel Programming 16(6), 479–499. Riesen, K., B. H., 2008. Iam graph database repository for graph based pattern recognition and machine learning. Pattern Recognition Letters. 5342, 287–297. Riesen, K., B. H., 2009. Approximate graph edit distance computation by means of bipartite graph matching. Image and Vision Computing. 28, 950– 959. Riesen, K., 2015. Structural Pattern Recognition with Graph Edit Distance - Approximation Algorithms and Applications. Advances in Computer Vi- sion and Pattern Recognition. Springer. Riesen, K., Bunke, H., 2014. Improving approximate graph edit distance by means of a greedy swap strategy. In: ICISP. pp. 314–321. Riesen, K., Fankhauser, S., Bunke, H., 2007. Speeding up graph edit distance computation with a bipartite heuristic. In: MLG. URL http://dblp.uni-trier.de/db/conf/mlg/mlg2007.html# RiesenFB07 Riesen, K., Fischer, A., Bunke, H., 2014. Improving approximate graph edit distance using genetic algorithms. In: SSSPR14. pp. 63–72. 58 http://dblp.uni-trier.de/db/conf/mlg/mlg2007.html#RiesenFB07 http://dblp.uni-trier.de/db/conf/mlg/mlg2007.html#RiesenFB07 Sanfeliu, A., Fu, K., 1983. A distance measure between attributed relational graphs for pattern recognition. IEEE Transactions on Systems, Man, and Cybernetics 13, 353–362. Serratosa, F., 2015. Speeding up fast bipartite graph matching through a new cost matrix. IJPRAI 29 (2). Tsai, W.-h., Member, S., Fu, K.-s., 1979. Pattern Deformational Model and Bayes Error-Correcting Recognition System. IEEE Transactions on Sys- tems, Man, and Cybernetics 9, 745–756. Van Loan, C., 1992. Computational Frameworks for the Fast Fourier Trans- form. W. Christmas, J. K., Petrou., M., 1995. Structural matching in computer vision using probabilistic relaxation. IEEE Trans. PAMI, 2, 749–764. Xu, C., Lau, F. C., 1997. Load Balancing in Parallel Computers: Theory and Practice. Kluwer Academic Publishers. Zeng, Z., Tung, A. K. H., Wang, J., Feng, J., Zhou, L., 2009. Comparing stars: On approximating graph edit distance. Proc. VLDB Endow. 2, 25– 36. 59 Introduction Problem Statements Graph Edit Distance Problem From GED Problem to Load Balancing Problem Load Balancing Problem Static load balancing Dynamic load balancing Related Work State-of-the-art of Graph Edit Distance Exact Graph Edit Distance Approaches Approximate Graph Edit Distance Approaches Synthesis State-of-the-art of parallel branch-and-bound algorithms Regular Exploration of the Search Space Irregular Exploration of the Search Space Synthesis Proposal: Parallel graph edit distance using a load balancing strategy Initialization, Decomposition and Assignment Branch-and-Bound Method Load Balancing and Communication Experiments Datasets Studied Methods Environment Protocol Scalability Experiment under Time Constraints Classification experiment Parameters Results Number of Threads Number of Edit Paths Comparing PDFS with naive-PDFS Comparing Methods under Constraints Classification Tests Conclusion and Perspectives 
work_7lkw3khzene7xafknvqc7d3w7e	----	Rule-based Expert Systems for Supporting University Students Procedia Computer Science 31 ( 2014 ) 22 – 31 1877-0509 © 2014 Published by Elsevier B.V. Open access under CC BY-NC-ND license. Selection and peer-review under responsibility of the Organizing Committee of ITQM 2014. doi: 10.1016/j.procs.2014.05.241 ScienceDirect Available online at www.sciencedirect.com 2nd International Conference on Information Technology and Quantitative Management, ITQM 2014 Rule-based expert systems for supporting university students Gökhan Engina, Burak Aksoyerb, Melike Avdagicb, Damla Bozanlıb, Umutcan Hanayb, Deniz Madenb, Gurdal Ertekb* a Oracle Turkey, Istanbul, 34398, Turkey bFaculty of Engineering and Natural Sciences, Sabancı University, Istanbul, 34956, Turkey Abstract There are more than 15 million college students in the US alone. Academic advising for courses and scholarships is typically performed by human advisors, bringing an immense managerial workload to faculty members, as well as other staff at universities. This paper reports and discusses the development of two educational expert systems at a private international university. The first expert system is a course advising system which recommends courses to undergraduate students. The second system suggests scholarships to undergraduate students based on their eligibility. While there have been reported systems for course advising, the literature does not seem to contain any references to expert systems for scholarship recommendation and eligibility checking. Therefore the scholarship recommender that we developed is first of its kind. Both systems have been implemented and tested using Oracle Policy Automation (OPA) software. © 2014 The Authors. Published by Elsevier B.V. Selection and peer-review under responsibility of the Organizing Committee of ITQM 2014. Keywords: rule-based systems; expert systems; recommender systems; decision support systems; higher education; university education. 1. Introduction 1.1. Rule-Based Expert Systems Rule-based expert systems have the ability to emulate the decision making ability of human experts. They are designed to solve problems as humans do, by exploiting encoded human knowledge or expertise1. This knowledge can be extracted and acquired directly through interaction with humans, as well as from printed and electronic resources such as books, magazines and websites. The extracted knowledge forms the knowledge * Corresponding author. Tel.: +90-216-483-9568 ; fax: +90-216-483-9550. E-mail address: ertekg@sabanciuniv.edu. © 2014 Published by Elsevier B.V. Open access under CC BY-NC-ND license. Selection and peer-review under responsibility of the Organizing Committee of ITQM 2014. http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2014.05.241&domain=pdf http://creativecommons.org/licenses/by-nc-nd/3.0/ http://creativecommons.org/licenses/by-nc-nd/3.0/ 23 Gökhan Engin et al. / Procedia Computer Science 31 ( 2014 ) 22 – 31 base of the rule-based system. The other major component of rule-based systems draws conclusions from this knowledge, and is referred to as the inference engine. The conclusions and suggestions offered by the rule- based system satisfy users’ needs for expertise within the chosen domain. Rule-based systems are designed to solve problems in a selected domain. Every domain has its own knowledge and reasoning humans, which can be emulated and even replaced through automated rule-based systems. Many domains contain a large amount of knowledge that can be captured fully only through an information system, since humans may not access or immediately retrieve fully the needed information. There are many advantages of rule-based expert systems: They decrease costs since they reduce the need for human experts; they are permanent; they can be used for different knowledge systems, which increases functionality; they increase reliability since they minimize errors that humans are prone to; and if designed by multiple experts, can increase confidence. Finally, they lack human emotions, which are sources of mistakes in human based systems1. The advantages of rule-based expert systems are multifold and they can considerably facilitate human life for the better. In this paper, we present and describe two rule-based recommender systems projects, both in the domain of university education. 1.2. Oracle Policy Automation (OPA) The software used in the reported projects is Oracle Policy Automation (OPA). OPA is capable of reading rules conveniently from spreadsheet and word-processor file formats, where the rules are written in a format akin to natural language. OPA automatically forms a problem solving or decision making application based on the acquired rule base. In OPA, the rules can be written in more than 25 languages, including Turkish, the language selected for the expert system in our second case study. Users codify the information or knowledge via Microsoft Word and/or Microsoft Excel in their native (natural) language and form rules without having to know any programming languages (such as Java or C#) or expert system languages (such as CLIPS). This makes the testing and maintenance of, as well as the changes to the rules easier. If the rules are detailed and need frequent modifications, and/or if there will be auditing and the rules need to be transparent, OPA is especially suitable and convenient. After the program is executed and a suggestion is presented by the program, the program also shows why and how that suggestion is selected. This way, the rules can be checked and the reasons behind the suggestions can be provided to users. Writing and running an OPA program has five stages. First one is to gather the information and data for the rules that will be implemented. Second step is to write the rules using Microsoft Excel and/or Word. Afterwards, the rule base is checked for language mistakes, rule mistakes and other errors that might have occurred. After the rule base is finalized, the users start answering the related questions to obtain a suggestion. In the first project, the suggestion is the list of courses that can be taken by a student, given the courses he has taken so far. In the second project, the suggestion is the eligibility of (a student for) a specific scholarship. After all the questions are answered, the rule-based expert system presents its suggestion. For example, in the second project, the resulting output is in the form “The student/user is eligible for the Sabancı Vakfı Scholarship” or “The student/user is not eligible for the Sabancı Vakfı Scholarship”, corresponding to the suggestions of applying and not applying to the Sabancı Vakfi Scholarship, respectively. The user can see the reason why the program gave the result it gave by clicking the underlined result. If the user is eligible for the scholarship, clicking the underlined hyperlinked text will show all the criteria satisfied by the user. If the user is not eligible for the scholarship, then the underlined hyperlinked text will state which criteria was not satisfied. This part of OPA is very useful for auditing reasons because it clearly states the reasons for the results. In our application this is highly relevant because the school systems, entrance exams and the rules for the eligibility change frequently. 24 Gökhan Engin et al. / Procedia Computer Science 31 ( 2014 ) 22 – 31 1.3. Expert Systems for University Education There are more than 6000 colleges and more than 15 million college students in the US alone2, and academic advising is typically performed by human advisors3. Engaging with students is becoming more complex and student-demand driven4, bringing an immense managerial workload to faculty members, as well as other staff at universities. There is also an information overload on the student side, as the information that needs to be interpreted, clarified, and captured is showing an ever increasing trend4. Expert systems developed for university education can serve a variety of goals, including advising, tutoring, instruction, and evaluation. We will discuss the studies in literature under two groups, namely expert systems for course advising and expert systems for other education-related goals. Within the first group we will also discuss course advising systems that are built using methodologies other than expert systems. Dunstan3, Nambiar and Dutta5, Laghari & Khuwaja6, and Al Ahmar7 report expert systems for advising undergraduate students on course selection. Al-Ghamdi et al.8 describe a system for advising postgraduate students and presents a detailed review of expert systems for course selection. Khanna et al.9 also provide an extensive review, besides listing the main areas where expert systems can be used in education. Onyeka et al.10 propose combining rule-based reasoning with case based reasoning in course advising expert systems. Goodarzi and Rafe11 incorporate fuzzy reasoning into its advisory expert system. Koutrika et al.12 develop a flexible system for generating advisory expert systems. Parameswaran et al.2 present CourseRank, a course recommendation and planning system developed at Stanford University. The authors report that more than 170 universities throughout the world have adopted CourseRank, making the software in the most widely used course advising system in the world. Besides expert systems for course advising, the literature contains a multitude of studies that describe course advising systems built using other methodologies. Studies of Vialardi et al.13, Taha14, Cho and Kang15, and Park et al.16 are examples where course recommendation is based on data mining. Among these three studies, Vialardi et al.13 reports using real world data from a Spanish university; Taha14 uses collaborative filtering; Cho and Kang15 uses hybrid filtering. Sobecki17 uses nature inspired methods, again for course recommendation. Besides course recommendation, expert systems have been developed for other purposes, as well. Buche and Querrec18 describe an expert system for helping human learners in virtual reality environments. Hwang et al.19 introduce an expert system that works as instructor for improving problem solving skills of students. Jaryani et al.20 present an expert system that provides customized learning for each student based on his/her background and realities. The fuzzy expert system in the work of Van Hecke21 identifies the personal profile of the students as one of the three types. Grupe22 presents a web based expert system for selecting academic major, where the student user is directed to one of the most suitable six majors from among 60. Deorah et al.23 also recommend academic majors. Fong and Biuk-Aghai24 describe a system that recommends the university that matches the student’s profile. The system developed by Aslam and Khan25 is one for recommending faculties to students. Finally, Cline and Brewster26 present an expert system for automatically evaluating student concept maps. 2. Case 1: Supporting Course Selection Decisions 2.1. Motivation For this project our aim was to create a rule-based system for Manufacturing Systems Engineering Program students at Sabanci University to prepare their schedules using Oracle Policy Automation (OPA). When preparing the rule base, we investigated rules related to prerequisite courses, student’s GPA (Grade Point Average), which courses open which semester in, which area the student is specializing in, and which courses the student has readily registered to. 25 Gökhan Engin et al. / Procedia Computer Science 31 ( 2014 ) 22 – 31 2.2. Project Progress In order to reach our aim we collected data from the university’s database of courses, namely BannerWeb. First of all we listed all courses in MS Excel and we connected them with their pre-requisites. This way the obtained a lists of the courses and the relations between courses. Furthermore, we represented this data in the yEd graph visualization software. It is possible to assign colors to nodes in yEd. This way, connection between courses and three main course types as area electives, core electives and required courses became clearly distinguishable (Figure 1). Fig. 1. Visualization of co- and pre-requisite relationship between courses We used the logical connectors “and” and “or” to establish logical connection between a course and its pre- requisites. If a user should take all pre-requisites of a course, “and” is the operator to identify relationships. On the other hand, if one of pre-requisites of a course is sufficient to take that course, “or” is the operator to describe the relationship. Also, through the “exists” command, system associates one entity with another entity. We established the course definition as an excel table with four columns, which are CourseID, CourseName, and CourseCredit and CourseCategory. Another Excel document used by OPA is the student database, which receives student number as input. By this single input, OPA matches student number with student name from the student list. Also, OPA saves all historical data of a user for the next session with the user. This way, a user can easily observe her/his historical movement on OPA as a user. For instance, a student can easily follow his/her credit and course level at past semesters. In addition to this, OPA calculates total credits for all sub - groups of courses is required, core and area electives. And then, OPA gives the user total credit by summation of all units. We finally planned and executed various scenarios and test cases for testing the developed export system. In addition to the mentioned rules, several other types of rules have been defined: Global Controls and Determinations: These are rules that globally determine student and credit attributes, such as the following: eligibility check is completed if total credit is known and student name is known the student is sophomore if total credit > 47 and total credit < 81 26 Gökhan Engin et al. / Procedia Computer Science 31 ( 2014 ) 22 – 31 Individual Course Eligibility Checks: These are rules that determine student eligibility for each course such as the following: the student is eligible for ENS204 if the student has taken NS101 and either the student has taken MATH101 or the student has taken MATH102 Course technical rules: These are technical rules that tell which courses are taken, such as the following: the student has taken CS201 Exists(all courses of the student, course id = “CS201”) the student has taken ENS208 Exists(all courses of the student, course id = “ENS208”) the student has taken ENS491 Exists(all courses of the student, course id = “ENS491”) the student has taken ENS492 Exists(all courses of the student, course id = “ENS492”) 2.3. Results The developed rule-based expert system meets the user with a welcome screen. Next screen takes as input the student number and in order to match student name and record action logs of the user. Third screen is designed for taking course list from the user (Figure 2). The user chooses his / her previously taken courses from a list, where the list is read from a source file by OPA. Once the user submits the courses taken, the lists of courses that s/he can take is listed (Figure 3). If a user wants to know the reason why s/he is not eligible for a course, the user should click “Why?” button on the results screen to reach that information, and learn the reason (Figure 4). 3. Case 2: Supporting Scholarship Decisions 3.1. Motivation The goal of the second project was also to assist university students by providing recommendations. After discussions in the early stages of our project, we decided to implement a scholarship advisor that will present, with a student’s criteria, which scholarship s/he is eligible for. There are many institutions that offer scholarships to university students, however every foundation’s criteria and budget is different. Our aim was to match universities students with the scholarships that they are eligible for. This can help students in easily finding what they are looking for and it can decrease a task that would take weeks to minutes. We used OPA again for this project. In our case, the rules would be based on the students’ GPA, their target university, their income, their skills, and other relevant criteria. OPA asks multiple choice questions that are related with the criteria of the scholarships. From the home screen, the student clicks on the scholarship and then answers the related questions, and finally the results screen shows if s/he is eligible for the scholarship or not. 27 Gökhan Engin et al. / Procedia Computer Science 31 ( 2014 ) 22 – 31 Fig. 2. Screen where the user (student, advisor, staff member) enters the courses taken so far Fig. 3. Screen where the courses that the student can take are listed Fig. 4. The screen where the rationale for the suggestion is provided 28 Gökhan Engin et al. / Procedia Computer Science 31 ( 2014 ) 22 – 31 3.2. Project Progress To realize the project, we firstly gathered the data of all the scholarships in Turkey and their criteria. Finding all the rules wasn’t easy since some of the scholarships did not have a website but had just a phone number. Criteria for some of the scholarships were not available on the websites since the application period was over. After scanning through scholarships, we found approximately 200 scholarships, half of which were only reachable by phone and didn’t provide a website. We eliminated those scholarships from further consideration. From among the other 100 scholarships, only 51 had the criteria publicly available on their websites. From that point on, we deducted the rules of eligibility for each scholarship and separated every one of them in an Excel sheet so that while writing the rules for OPA, it would be easier and faster. Furthermore, after gathering all the data we parsed their requirements and pre-requisites. For one institution, there may be an income requirement, while for another, meeting just a GPA threshold is sufficient, and thus we divided all the information into subcategories. To make it more organized, every scholarship’s criteria is listed in an Excel file so that every rule can be seen and plugged into OPA easily. With this system the students’ search and information gathering time can decrease significantly, and they can filter through significant amount of information and work to achieve on their goals. Moreover, we believe students would not be the only people who would appreciate this program. Also, foundations can receive applications from students who are more likely to be selected, and consequently their workload can be reduced, as well. The second stage of the project consisted of writing all the rules in OPA. Figure 5 illustrates rules for a particular scholarship. This rule writing was done for each of the 51 scholarships. Fig. 5. Rules for a particular scholarship 29 Gökhan Engin et al. / Procedia Computer Science 31 ( 2014 ) 22 – 31 Fig. 6. The first screen that the user sees in Case Study 2 3.3. Results The lists of possible scholarships (in this program there are 51 scholarships) appear on the user’s first interface. The user starts the program by clicking on the scholarship she/he wants to test the eligibility for the first. There are some common rules required for most of the scholarships, the question is asked only once and used in other scholarships as well. Examples for these common rules are the school student wants to attend or is attending, students major, and is whether the student is a Turkish citizen or not (most frequent criterion). In Figure 6, the first interface that the user sees is presented. Here in our example, let us assume that the student clicks the link “Anadolu Eğitim ve Sosyal Yardım Vakfı Bursuna başvurmaya hak kazanır mı?” (“Is the student eligible for the Anadolu Eğitim ve Sosyal Yardım Vakfı scholarship?”) and starts answering the questions in order. Questions appear successively as they are answered in a way that preserves eligibility. If the answer violates eligibility, then the expert system automatically stops and prompts that “The student is not eligible for this scholarship”. For the specific scholarship mentioned, the correct sequence of answers would be (yes, yes, yes, yes, no, no, no), and if this sequence is entered the reason page would look like Figure 7. If, for example, the second question is answered no instead of yes, the reason page would look like Figure 8. Fig. 7. Result screen displayed when the student is eligible, with the rationale for the suggestion (eligible, apply) is explained 30 Gökhan Engin et al. / Procedia Computer Science 31 ( 2014 ) 22 – 31 Fig. 8. Result screen displayed when the student is not eligible, with the rationale for the suggestion (not eligible, do not apply) is explained 3.4. Discussions The biggest problem this rule-based system presented was with the questions where variables were used. Rather than the system asking whether someone’s GPA is greater than 3.5, we decided it would be better if the program asked what the GPA was and then decided if it satisfies the rules. The problem arose when for example we wanted to know the students major. It would have been impractical to ask the user to answer questions like “Are you a Business major?” and if the answer is no asking “Are you a Industrial Engineering major?” and so on. So again we decided to use the variables. But since the program is written in Turkish, the user could enter the major using capital letter, small letter, Turkish letters or English letters. This makes a huge number of possibilities. This is where we decided to search for a pull-down menu for these variables. The student using this automation doesn’t have to answer all the questions for the 51 scholarships; by clicking on the few scholarships they want to apply they can answer the related questions. The system also doesn’t require the user to answer the same question more than once even if the question’s answer is required for different scholarships. For example the citizenship criteria is one of the popular ones, so in the first eligibility test the user answers it and the program remembers the answer and does not ask the same question for the other scholarships. Our first thought was for the student to write down all his or her information and then all the available scholarships would have been presented. Unfortunately due to the fact that every scholarship has different rules and the importance of the answers change in different scholarships, that kind of system would not have worked. In the end, we have assembled the information and the available software and completed the case study successfully. 4. Conclusions This paper reported and discussed the development of two educational expert systems: The first expert system is that course advising system which recommends courses to undergraduate students. The second system suggests scholarships to undergraduate students based on their eligibility. While there have been reported systems for course advising, the literature does not seem to contain any references to expert systems for scholarship recommendation and eligible checking. Therefore, the scholarship recommender that we developed is the first such system in the literature and can provide a frame of reference for following studies. Both systems have been implemented and tested using Oracle Policy Automation (OPA) software, a convenient integrated development environment for expert systems development. Future stages of the project can involve the deployment of the developed expert systems on the university intranet and/or the internet. Acknowledgements The authors thank Mete Sevinç for creating the visualization of co- and pre-requisite relationships between the courses. 31 Gökhan Engin et al. / Procedia Computer Science 31 ( 2014 ) 22 – 31 References 1. Giarratano JG. Expert Systems Principles and Programming. USA: PWS Publishing; 2002. 2. Parameswaran A, Venetis P, Garcia-Molina H. Recommendation systems with complex constraints: A course recommendation perspective. ACM Transactions on Information Systems (TOIS), 2011, 29(4), 20. 3. Dunstan N. ET: an Enrolment Tool to Generate Expert Systems for University Courses, In: Expert Systems, Petrica Vizureanu (Ed.), ISBN: 978-953-307-032-2, InTech, DOI: 10.5772/7071. 2010. Available from: http://www.intechopen.com/books/expert-systems/et-an- enrolment-tool-to-generate-expert-systems-for-university-courses or the short link http://bit.ly/19nnbp6. Accessed on November 14, 2013. 4. Oracle Policy Automation for Higher Education. Oracle Data Sheet. Available under short link http://bit.ly/1boStNO. Accessed on November 13, 2013. 5. Nambiar AN, Dutta AK. Expert system for student advising using JESS. In: 2010 International Conference on Educational and Information Technology (ICEIT), IEEE. 2010, 1, p. V1-312. 6. Laghari, MS, Khuwaja GA. Electrical engineering department advising for course planning. In: Global Engineering Education Conference (EDUCON), 2012 IEEE. 2012, p. 1-6. 7. Al Ahmar MA. A Prototype Student Advising Expert System Supported with an Object-Oriented Database. (IJACSA) Int J Adv Comp Sci Appl. 2011. Available under http://www.ijacsa.thesai.org. Accessed on November 13, 2013. 8. Al-Ghamdi A, Al-Ghuribi S, Fadel A, AL-Ruhaili FAAT. An Expert System for Advising Postgraduate Students. (IJCSIT) Int J Comp Sci and Info Tech, 3(3) , 2012, 4529-4532. 9. Khanna S, Kaushik A, Barnela M. Expert systems advances in education. In: Proceedings of the National Conference on Computational Instrumentation NCCI-2010. CSIO, Chandigarh, India. 2010, p. 109-112. 10. Onyeka E, Olawande D, Charles A. CAES: A model of an RBR-CBR course advisory expert system. In: 2010 International Conference on Information Society (i-Society),. IEEE. 2010, p. 37-42. 11. Goodarzi MH, Rafe V. Educational Advisor System Implemented by Web-Based Fuzzy Expert Systems. Journal of Software Engineering and Application (JSEA), 2012. 5(2). 12. Koutrika G, Bercovitz B, Ikeda R, Kaliszan F, Liou H, Garcia-Molina H. Flexible recommendations for course planning. In: IEEE 25th International Conference on Data Engineering, 2009 (ICDE'09). 2009, p. 1467-1470. 13. Vialardi C, Bravo J, Shafti L, Ortigosa A. Recommendation in higher education using data mining techniques. Group on Educational Data Mining, 2009, p. 190-199. Available from: http://eric.ed.gov/?id=ED539088. Accessed on March 28, 2014. 14. Taha K. Automatic academic advisor. In: Collaborative Computing: 8th International Conference on Networking, Applications and Worksharing (CollaborateCom). 2012, p. 262-268. 15. Cho J, Kang EY. Personalized Curriculum Recommender System Based on Hybrid Filtering. In: Advances in Web-Based Learning– ICWL 2010, 2010, p. 62-71. Springer Berlin Heidelberg. 16. Park DH, Kim HK, Choi IY, Kim JK. A literature review and classification of recommender systems research. Expert Syst Appl, 2012, 39(11), 10059-10072. 17. Sobecki J. Comparison of nature inspired algorithms applied in student courses recommendation. In: Computational Collective Intelligence. Technologies and Applications, 2012, p. 278-287. Springer Berlin Heidelberg. 18. Buche C, Querrec R. An expert system manipulating knowledge to help human learners into virtual environment. Expert Syst Appl, 2011, 38(7), 8446-8457. 19. Hwang GJ, Chen CY, Tsai, PS, Tsai CC. An expert system for improving web-based problem-solving ability of students. Expert Syst Appl, 2011, 38(7), 8664-8672. 20. Jaryani F, Sahibudin S, Ibrahim S, Rahman NA, Daruis R. Intelligent reflective e-portfolio framework based on artificial intelligent Expert systems techniques. In: 3rd International Conference on. Computer Research and Development (ICCRD), IEEE, 2011, 2, p. 214- 216. 21. Van Hecke T. Fuzzy Expert System to Characterize Students. PRIMUS, 2011, 21(7), 651-658. 22. Grupe FH. An Internet-based expert system for selecting an academic major: www.MyMajors.com. Internet High Educ, 2002, 5(4), 333-344. 23. Deorah S, Sridharan S, Goel S. SAES-expert system for advising academic major. In: 2nd International Advanced Computing Conference (IACC), IEEE, 2010, p. 331-336. 24. Fong S, Biuk-Aghai RP. An Automated University Admission Recommender System for Secondary School Students. In: The 6th International Conference on Information Technology and Applications. 2009. 25. Aslam MZ, Khan AR. A Proposed Decision Support System/Expert System for Guiding Fresh Students in Selecting a Faculty in Gomal University, Pakistan. Ind Eng Letters, 2011, 1(4), 33-40. Available under http://www.iiste.org. Accessed on November 13, 2013. 26. Cline BE, Brewster CC, Fell RD. A rule-based system for automatically evaluating student concept maps. Expert Syst Appl, 2010, 37(3), 2282-2291. 
work_7ncnetuel5hujl7vae6rqhp3wu	----	Naar de inhoud springen Terms of Use Login Zoeken naar: Zoeken Je moet eerst inloggen, daarna kun je een nieuwe site maken. Technische Universiteit Delft Auteursrecht © 2021 Alle rechten voorbehouden.Thema: Flash door ThemeGrill. Aangedreven door WordPress 
work_7nklpxgverfz5e5m6l7kxzl2cu	----	This is an electronic reprint of the original article. This reprint may differ from the original in pagination and typographic detail. Powered by TCPDF (www.tcpdf.org) This material is protected by copyright and other intellectual property rights, and duplication or sale of all or part of any of the repository collections is not permitted, except that material may be duplicated by you for your research use or educational purposes in electronic or print form. You must obtain permission for any other use. Electronic or print copies may not be offered, whether for sale or otherwise to anyone who is not an authorised user. Jämsä-Jounela, S-L.; Laine, Sampsa; Ruokonen, Eeva Ore Type based Expert Systems in Mineral Processing Plants Published in: Particle & Particle Systems Characterization Published: 01/01/1998 Document Version Peer reviewed version Please cite the original version: Jämsä-Jounela, S-L., Laine, S., & Ruokonen, E. (1998). Ore Type based Expert Systems in Mineral Processing Plants. Particle & Particle Systems Characterization, (15), 200-207. Author’s accepted manuscript, published in Particle & Particle Systems Characterization 15 (1998) 200-207 Ore Type based Expert Systems in Mineral Processing Plants Sirkka-Liisa Jämsä-Jounela*, Sampsa Laine**, Eeva Ruokonen*** Abstract Artificial intelligence (AI) includes excellent tools for the control and supervision of industrial processes. Several thousand industrial applications have been reported worldwide. Recently, the designers of the AI systems have begun to hybridize the intelligent techniques, expert systems, fuzzy logic and neural networks, to enhance the capability of the AI systems. Expert systems have proved to be ideal candidates especially for the control of mineral processes. An expert system based on on-line classification of the ore type has been developed. Self-organizing maps (SOM) are used for pattern recognition of the type of feed. The system has been tested in two concentrators, the Outokumpu Finnmines Oy, Hitura Mine and Outokumpu Chrome Oy, Kemi Mine. The methodology for the development of the ore type based expert system is presented and the preliminary results obtained in the above plants are described. 1 Introduction Ten years ago, expert systems were introduced, and this was followed by an explosion in the use of fuzzy logic and neural networks. Since then, these three technologies have received several thousand industrial applications worldwide: monitoring and analysis of process operations, fault diagnosis, supervisory control, feedback control and scheduling and planning of operations. Recently, the designers of these systems have begun to hybridize these intelligent techniques to enhance the current capability of expert system. Expert systems have proved to be excellent tools especially for the control of mineral and metallurgical processes. Successful industrial applications have been reported in the area of diagnosis of process operations and intelligent control. Ore deposits are almost always heterogeneous in composition and structure. Heterogeneity impairs profitable production because dissimilar parts of an ore deposit require different treatments in beneficiation. The primary reason for concentrator instability is variations in the type of feed. This problem can be approached in two ways. The first approach is to determine explicitly the type of feed and to develop respective strategies. Determination can be based on the location of the source of the ore (geostatistical approach) or on measurements carried out inside the concentrator (on-line determination). The second approach is to treat the problems caused by varying feed types without explicit knowledge of the feed type. An expert system based on the geostatistical approach for concentrator control has been realized at the Kemira Siilinjärvi mine. The concentrator uses homogenization piles into which the ore is hauled from the excavations. The plant geologist determines the type of ore in the excavation, and the composition of the piles can be deduced from the excavation data. A database storing the process behavior of each type is maintained. As the feed type is known, the operator can find, using the system, the best method of treating the type of feed in question [1, 2]. In the second approach, mathematical models are required in the determining the ore type. Cluster analysis [3] has been used for the off-line determination of the ore type [4]. Significant work related to the determination of the feed type has been carried out by Ginsberg and Whiten [5]. They demonstrated that the cluster algorithm is a good mathematical method for classifying data of high dimensionality. The corresponding on-line algorithm has been developed by Ylinen et al. [6]. Laine et al. [7, 8] have studied self-organizing maps for ore-type determination. As successful case projects, expert system based on on-line classification of the ore type is described in this paper. The essential feature of this expert system is the classification of different ore types and their distinct control strategies at the plant. In addition to the classification, the expert system has a database containing information about how to handle the determined ore type. This self-learning database scans historical process data to suggest the best treatment for the ore type under processing. The system has been tested in two concentrators, the Outokumpu Finnmines Oy, Hitura mine and Outokumpu Chrome Oy, Kemi mine. 2 Generic Methodology for the Development of the Ore Type based Expert System The major tasks of the development of the expert system are presented in Figure 1. The first task is the off- line determination of the feed types of the concentrator. The results are used in the development of the on- line classification and also in the feed type based process study that results in the authoring of the knowledge base. Fig. 1. Utilization of off-line classification The different types of information can be used in the off-line determination of the feed type. A spatial model of the mineralogy of the ore body can be created and the type of concentrator feed can be determined using the location of the excavation. Another possibility is to sample the concentrator feed and to study the feed type on the basis of the laboratory analysis. If possible, independent sources of off-line information should be used as the results derived from one source can be verified using another source. The data are classified using a classification algorithm and the results are verified. In this research the self- organizing map (SOM) has been selected for the classification. The development phases of the on-line SOM are presented in Figure 2. The on-line measurements available in the process automation system are studied and the measurements reflecting the feed type are collected. The on-line classification variables are created on the basis of these measurements; the primary objective of the phase is to remove the influence of the operator present in the measurements. An on-line SOM is calculated and verified using these measurements. Fig. 2. Formulation of the on-line SOM The structure of the expert system proposed for the Hitura and Kemi concentrators is presented in Figure 3. The main function of the system is the on-line determination of the feed type performed by the classification block. It is based on the SOM which classifies the on-line information received from the automation system. This block is supported by the update classification block. It contains the functions for the calculation and modification of the SOM. Fig. 3. Structure of the expert system Prior to the classification, the data must be validated to remove invalid data produced by process or measurement disturbances. The second major set of functions in the system is related to the knowledge base block. The block itself stores the data about the proper treatments for various feed types. The block is supported by the update knowledge base block that updates the knowledge if new proper methods of treatment appear. The method is based on a success index indicating the profitability of the relevant treatment. The success index is an on-line measurable parameter. 3 Self-Organizing Map Of the architectures and algorithms suggested for artificial neural networks, the SOM has the special property of effectively creating spatially organized internal representations of various features of input signals and their abstractions. Self-organization is based on competitive training that is able to find clusters from the learning data. 3.1 Training Algorithm In training, the elements of the SOM compete for the input vector; the element that is closest to the input vector is the winner. The selected element (winner) and its neighborhood are moved closer to the presented input vector. The elements of the network gradually learn to represent the training data. Because the neighborhood is taken into account, the properties of adjacent elements become similar. The map becomes ordered in such a way that clusters with similar properties are located near to each other [9]. If the input vector is denoted by x = [x0 x1 … xN−1] T and the location vector of a mapping element by mi = [mi0 mi1 … miN−1] T, it is possible to write the algorithm that describes the self-organizing operation: (I) Initiate the locations of the elements with random values (II) For each vector of the training data compute C1 and C2 (C1) Find the SOM element mc (winner) best matching the data vector x (t) by searching all elements mi |x(t) − mc(t)| = min i {|x(t) − mi(t)|} (1) (C2) and adjust the locations of the elements mi(t + 1) = { mi(t) + α(t){x(t) − mi(t)}for i ∈ Nc (Nc = neighborhood) mi(t)for all other indices (2) In Eqs. (1) and (2) the Euclidean metric can be used as the distance measure. The parameter α(t) in Eq. (2) is a coefficient that determines the movement of the winning element and the neighborhood in the direction of the data vector x(t). The parameter α(t) is equivalent to the learning rate used in the back propagation algorithm for feedforward neural networks. α(t) is recommended to decrease slowly with time. Initially α(t) may be chosen as unity. In the final stages, the values are recommended to be less than 0.01. One method for calculating α(t) is presented in the following equation, in which T, the total number of iterations, and 𝛼0, the initial value of α(t), are provided by the user of the algorithm. 𝛼(𝑡) = 𝛼0𝑇 𝑇+100𝑡 (3) The neighborhood of the winning element c is defined by 𝑁𝑐 = {𝑖|𝑑(𝑖, 𝑐) < 𝑟(𝑡)}, where 𝑑(𝑖, 𝑐) is the distance between map elements i and c, and r(t) is the radius of the neighborhood. To ensure good global ordering, the radius r(t) should initially be more than half the diameter of the network. The radius should slowly decrease with time. In the final stage a small radius gives the map a good local spatial resolution. In addition to arranging the map topologically, the use of the neighborhood equalizes the number of input vectors classified in each cell. The third parameter of the training is the number of iteration steps. To achieve a good statistical accuracy, the number of training steps must be at least 500 times the elements in the SOM. 3.2 Visualization of the SOM The topology of the trained SOM in the training data space RT can be inspected graphically. Each SOM element carries a vector specifying its location in the RT. In one-dimensional SOM the value corresponding to the dimension is collected from the location vector of each element. The values are set into a matrix that can be presented graphically. The elements with high values are colored black and elements with low values are colored white. This is illustrated in Figure 4. In the two-dimensional SOM, the top left corner corresponds to the section of training data closest to the origin of the XY presentation: the values of both variables are close to zero. Similarly, the top right corner of the SOM corresponds to the bottom right corner of the XY presentation: the value of Var1 is at its maximum and the value of Var2 is approximately one. Fig. 4. Graphical illustration of the values of a variable in the elements of the SOM 4 Hitura and Kemi Mines The Hitura mine is a nickel mine located in central Finland. The mine capacity is 530 000 t/a ore and production currently is 38 000 t/a of NiCu concentrates with a grade of 6.5% Ni, 1.7% Cu and 10–12% MgO. The Hitura ore (0.65% Ni, 0.2% Cu) occurs in amphibole-bearing rocks at contact of a pipe-shaped serpentine against mica-gneiss. The main sulfide minerals are pyrrhotite, pentlandite and chalcopyrite. The inner part of the ore body is strongly serpentinized ultramafite and parts of the sulphides have been converted to mackinavite and vallerite. At the same time the grain size of the sulfides decreases. There are many problems in beneficiating the deposit: natural flocculation, grindability variations, slime formation and natural floatability of MgO-rich sheet silicates. This results in low recovery and grade of the concentrate and a high consumption of chemicals and energy. At the Hitura mine the underground ore is truck-hoisted to the three-staged crushing plant at the surface. The fine ore is heap-stored and its fineness is 100% – 0.022 mm. The conventional rod mill-ball mill grinding circuit consumes 20–30 kWh/t-new-200 mesh. The throughput varies from 60 to 75 t/h. Slurry densities in grinding are exceptionally low owing to the tendency for flocculation. Rougher flotation consists of four rougher banks, each including four OK-16 cells. The residence time of roughers is approximately 50 min. The concentrate from the first rougher bank is cleaned once. The concentrates from the other rougher banks are reground and cleaned three times. The cleaner tails are fed into the third rougher bank. Sulfuric acid has been used for pH adjustment (pH5–6), and it also has a dispersing effect on the slurry. The Kemi chromium mine is part of Outokumpu Oy’s stainless steel business segment, which is engaged in the fully integrated production of stainless steel. At present over 106 tons of ore are mined annually at the Kemi open pit mine. The chromite concentrates, upgraded lumpy ore (,36% Cr2O3) and metallurgical grade concentrate (,44% Cr2O3) are used as raw materials in ferrochrome production at Tornio. The output from the Kemi mine is currently 500 000–600 000 tonnes of concentrates per year. Production is based on a chromite deposit with an average content of 26% Cr2O3 and a Cr/Fe ratio of 1.55. Proven and provable ore reserves are about 70 Mt. Roughly 10 Mt of the deposit can be excavated by open-pit methods. The ore deposit is part of an ultrabasic layered formation located in the area around Kemi in Northern Finland. The almost vertical ore horizon can be used for a distance of 4.5 km and has an average thickness of 40 m. The mineable ore consists of 11 orebodies. There are differences between the orebodies as regards mineralogy, mineral chemistry, physical properties and structure. The average content of the major minerals in the Kemi ore are chromite 73,0%, chlorite 10,7%, tal 5,1%, serpentine 3,7%, flogopite 1,03%, tremolite 2,4% and dolomite 2,2%. The ratios of the gangue mineral composition varies in the individual ore bodies. The processing of chromite at the Kemi concentrator is mainly based on gravity concentration. A high magnetic separation is used to recover the chromite from slimes. After crushing, concentration of the ore takes place in two steps: heavy medium separation and concentration. The flowsheet of the Kemi concentrator is presented in Figure 5. Fig. 5. Flowsheet of the Kemi concentrator The ore is separated into upgraded lumpy ore, middling and waste rock in the two stages of heavy media separation. The process consists of feed preparation in a washing screen, heavy media separation in drum separators and medium recovery by means of low intensity magnetic separators. The concentrating plant is divided into four main unit processes: grinding, cone concentrating, slime circuit and dewatering. The grinding circuit consists of a rod mill and a ball mill operating in a closed circuit with vibrating screens. The mill product after processing by the vibrating screens is classified in hydrocyclones prior to the Reichert cones. The cyclones are for dewatering and desliming. Cyclone underflow, which represents most of the material, is concentrated in the Reichert cone concentrators. The overflow is fed to the slime circuit for further dewatering, classification and enrichment. The concentrates from the Reichert cones and the slime circuit are combined to form the metallurgical grade concentrate (~44% Cr2O3). Foundry sand (~47% Cr2O3) is produced from the cone concentrators in the spirals and classified in the cone classifiers. It is produced only when needed and when the ore quality is suitable for making foundry sand. 5 Case 1: Hitura Concentrator At the Hitura mine, changes in the mineralogy of the concentrator feed cause problems in process control. After a change in the feed type a new process control method has to be found. This is done by experiment because the new type is often unknown. These experiments take time and the resulting treatment method may not be optimal. An expert system has been developed for solving the problem. The system receives variables from the automation system of the concentrator process containing information about the type of feed. The SOMs are used to classify the variables to determine the type of feed. The proper treatments are called from a knowledge base and presented to the operator. The structure of the expert system tested at the Hitura concentrator is outlined in Figure 3. The classification program with the interfaces has been implemented with C++ in MS-Windows 3.1. The system is connected to the Proscon 210 automation system by means of the ×10 comm communication program. The system supports the calculation and study of a SOM and a preliminary version of the knowledge base as follows: - on-line measurable classification variables from the concentrator are validated and classified using the Kohonen SOM; - feed type is forwarded to the knowledge base (KB); - KB will give advice on the proper treatment of this feed to the operator. The two modules form the main segment of the system: classification (CLA) and updating of classification (UCLA) together. The CLA provides with the mapping of the n-dimensional ore type data to the two- dimensional ore type SOM, thus performing classification. The on-line SOM is provided by the UCLA module and is updated automatically to new ore types. The two modules form the second segment of the system: knowledge base (KB) and updating of knowledge base (UKB). The KB contains the set points of controllers and written information for various types of feed. This data are provided with the UKB module. The UKB can be updated adaptively or manually. Adaptive operation is based on an on-line measurable success index that indicates how well the ore is treated. The KB is updated if the success index indicates a new good method of treating an ore type. Manual operation is based on knowledge of the process engineer. The knowledge follows from experience, process study or experiments. 5.1 On-line Determination of the Type of Feed 5.1.1 On-line Measurements for Classification The on-line information available at the plant was studied and a search was made for useful on-line measurements. The domain experts of the Hitura concentrator recommended the use of measurements related to grinding, the consumption of sulfuric acid and the channel intensities of the on-line XRF analyzer. The grinding measurements reflect the specific energies required to grind the ore, which in turn reflect the host rock mineralogy; the relevant measurements were the power draw of the mills, the feed rate of ore to the rod mill, the particle size distribution at the hydrocyclone overflow and the pulp density at the hydroyclone feed. The set related to the consumption of sulfuric acid reflects the content of serpentine in the feed; the relevant measurements were the channel intensities of the Courier 30 analyzer relevant to the analysis of the concentrator feed. 5.1.2 On-line Variables for Classification The three formulation methods were used to create five classification variables. Of the original 24 feed types defined by the off-line SOM, only four were used in the task. Representatives of group 1 were selected to represent serpentine feed types with a low nickel content. The representatives of group 19 are serpentine feed types with a high nickel content. Group 22 represents talc-amphibole-serpentine feed types and group 24 talc-amphibole types. These types were chosen to represent the feed types of the Hitura concentrator. The reduction in the number of groups was necessary in order to control the amount of information presented in the plots. The variable NIMO estimates the nickel content of the feed. It indicates the amount of pentlandite in the feed and is a good indicator of the feed type. The nickel content is estimated by two Courier 30 channel intensities: nickel and molybdenum. The variable NISU estimates the nickel content in the sulfide phase of the ore and reflects both the sulfide phase and the host rock mineralogy. The CUFE variable is the ratio between the copper and iron channels of the Courier 30 analyzer analyzing the concentrator feed. The variable GRIN reflects the grindability of the ore. The classification variable ACID represents the consumption of sulfuric acid. The mineralogical justification is the presumed dissolution and/or adsorption of sulfuric acid by serpentine. As talc, amphibole and chlorite are rather insoluble in sulfuric acid, and the consumption can be used to estimate the proportion of serpentine. 5.2 Formulation of the On-line SOM The on-line SOM was trained using the five variables presented above. The product is a mapping of the ore types of Hitura by on-line data. The on-line SOM is used in on-line determination of the feed and in presenting the classification result to the user. The on-line measurement values were collected from the automation system historical database. Daily- average values in the period between 19.3.1994 and 30.10.1995 were used. The data were screened for process shutdown time; days with more than 3 hs of downtime were removed. After this, 526 cases were available for classification. The data were validated with high-low boundaries to remove faulty values due to disturbances in the unit processes or measurement devices. The boundaries were determined from the histograms of the variables. Weighting for variables was not used. The data were scaled into a standard deviation of 1.0 and mean of 0.5. The pre-processed data were classified using the Kohonen SOM. The bubble algorithm was used to train a hexagonal SOM with 8 rows and 12 columns. The locations of the ore types in the on-line SOM were studied. The high values of variable NISU were on the right side and the low values on the left side of the map. Thus the serpentinized ore types are on the right side and the talc-amphibole types on the left. The average and high values of variable NIMO were found at two locations: in the top-left corner and in the middle right area of the map. They represent the variations in the nickel content within the major feed types. The result of the study on the on-line SOM is presented in Figure 6. Fig. 6. On-line SOM presenting the current feed type and the feed type history to the user Table 1 Estimated savings in the Hitura mine 75% limit 50% limit 25% limit Savings (kFIM/a) 933 2088 3962 5.3 Formulating of the Knowledge Base The knowledge base contains the treatments for various types of feed. Methods to formulate the treatment are used. The first method is based on the calculated average values of the controlled and manipulated variables. The calculation was made based on the long term historical values of the variables. The average values are connected to the SOM. The second method is based on the adaptive self-learning database. The optimal processing conditions for each feed type are found based on on-line success indexes. The best treatments are updated to the top-ten list after every successful treatment. 5.4 Estimation of the Economical Results of the Expert System The economic benefits of the expert system are mostly due to the faster and more accurate adaptation for the ore type changes. The calculated estimation is based on the penalties inflicted by excess MgO in the concentrate and on the nickel lost in the concentrator tailings. Each ore type group in the off-line SOM was studied. The system has been in use at the Hitura concentrator since December 1996. During this period, the Hitura personnel have maintained the system. According to the users, the system is capable of approximately indicating the feed type. The operators have accepted the system for daily use. The operators of the concentrator suggest the authoring of the knowledge base as the next step in the development of the system. 6 Case 2: Kemi Concentrator 6.1 Main modules of the Expert System Chromite has to be separated from the constantly varying mineralogical environment. Rapid changes in ore grade, hardness and chromite grain size make it difficult to control the separation. The process control strategies in the different unit processes are based on dissimilar behaviors of the ore types. The modules of the expert system under development at the Kemi concentrator are as follows: - Choice of the best calibration model for the on-line assay analyzers. - Optimization of the grinding circuit. - Optimization of the Reichert cone circuit. - Optimization of the HGMS circuit. The main modules of the expert system at the Kemi concentrator are presented in Figure 7. Fig. 7. Structure of expert system at the Kemi concentrator For the development of the expert system, the classification system developed at the Hitura concentrator was implemented at the Kemi concentrator. The system was used for the study of the present control strategies in the cone classification circuit. After the training of the SOM, the system was implemented to run on-line. At the grinding circuit, the system was trained to give information on the ore hardness and the correct treatment to the operators. 6.2 Testing of Classification The process history data from the grinding process and the Reichert cone circuit were classified using the neural networks. The aim of the cone circuit data classification was to study the control strategies the operators use to control. The choice of the optimum cone settings is critical for the grade and recovery obtained in the concentrates. There are 94 manipulated variables controlling the concentration: 38 variable slot settings, 36 double cone settings, 9 feed density measurements and 10 dilution water additions. At present the cones are controlled manually from the control room based on the grade of the concentrate. The operators select the set point values for the controllers on the basis of experimentation. Determination of the correct set point values takes time and subsequently reduces profits. The control method currently in use does not optimize the recovery/grade and results in reduced economic profit of the plant [10]. The present control strategy was studied in order to develop the instructions for controlling the cones. The initial aim was to study the control methods used by the operators to control the cones by investigating the set point values of the controllers. A secondary aim was to study the correlation between an ore type and the control methods used. The methods used by the operators to control the cones was studied using an expert system developed at the Hitura mine. The off-line training data contained 17 variables from 1232 shifts. The classification variables were shift averages of the cone setting combinations typically (40–100%) used by the operators. The on-line SOM of the system produces a map with a classification result. Each element in the SOM represents a specific type of treatment method and group of cone settings. The on-line values of each variable in the element are shown in the separate variable window in Figure 8. Fig. 8. On-line SOM of the cone circuit and the data for the latest element in the variable window The position of the on-line classification element in the SOM visualizes the classification presentation. In this case the wide cone settings are located in upper corners in the SOM and the narrow cone settings are located in bottom corners. The pattern followed in changing the slot settings was found to be very simple. According to the classification study, the operators keep the cone settings constant (from wide to narrow settings) for roughing and scavenging. The cleaning stage slot settings are changed more frequently in order to achieve the desired grade of concentrate. When these settings are unsuitable, they change the position of the slots for roughing and scavenging on all the individual cones. The study also revealed that a specific ore type was occasionally treated with group of set point values that resulted in reduced economic results. The grinding circuit SOM was trained to give information on the ore hardness and to give suggestions for ore type treatment. The classification variables were 800 shift averages from the feed t/h, mill powerdraw and mill rotation speed. Figure 9 shows the on-line SOM of the grinding circuit. The black circle around the element presents the current type of feed under processing; the window of the top right corner was been used to study the values of a variable in the elements of the SOM. Fig. 9. On-line SOM of the grinding circuit and the suggestion for the ore type treatment The results and experiences during the test period have been promising. In the near future, studies will be focused on developing operating instructions for the operators and adding them to the system. 7 Conclusions Expert systems have proved to be ideal candidates especially for the control of mineral processes [11]. An expert system based on on-line classification of the ore type has been developed and tested in two concentrators. The advantage of the SOM for process control was found to be its capability for visualization and its computational lightness, and also a large amount of data and/or high dimensionality can be used. In addition to the classification, the expert system included a database concerning information about how to handle a determined ore type. This self-learning database scans historical process data to suggest the best treatment for the ore type under processing. The preliminary results of applications have been promising. The economic benefits are based on faster and more correct adaptation for ore changes. References [1] J. Kallioinen, E. Kesti, J. Miettunen, I. Mullen: A sophisticated information system for process control. Proc. Copper 87, Economics in Metallurgy and Process Control, Santiago, Chile 1988. [2] S.-L. Ja¨msa¨: Experiences in the use of expert system in grinding process control in Siilinjärvi concentrator. Proc XIV Int. Mineral Processing Congress, pp. 1759–1770. Stockholm, Sweden 1988. [3] J. James: Classification Algorithms. Collins, London 1985. [4] M. Anderberg: Cluster Analysis for Applications. Academic Press, New York 1973. [5] D. W. Ginsberg, W. J. Whiten: A data based expert system for engineering applications. Proc. IFAC Workshop on Expert Systems in Mining and Metal Processing, Espoo 1991. [6] R. Ylinen, S.-L. Jämsä-Jounela, J. Miettunen: Use of Cluster Analysis in Process Control. 12th World Congress on IFAC, Sydney, 1993. [7] S. Laine, H. Lappainen, S.-L. Jämsä-Jounela: On-line determination of ore type using cluster analysis and neural networks. Miner. Eng. 8 (1995) 637–648. [8] S. Laine: Ore type based expert system for the Hitura concentrator. Lich. Tech. Theses, Helsinki University of Technology, 1995. [9] T. Kohonen: The self-organizing map. Proc. IEEE 78 (1990) 1464–1480. [10] E. Ruokonen, R. Bergström: Bases on designing ore type based expert system for Kemi chromium mine. Proc. IFAC Symposium on Automation in Mining, Mineral and Metal Processing, pp. 349–354. Sun City, 1995. [11] S.-L. Jämsä-Jounela: Modern approaches to control of mineral processing. Acta Polytech Scand. Math. Comput. Sci. Ser. 57 (1990). 
work_7oeuogr3grai7jg3266i56thxa	----	Authenticating the Writings of Julius Caesar Mike Kestemont University of Antwerp, Prinsstraat 13, B-2000 Antwerp, Belgium Corresponding author: mike.kestemont@uantwerp.be; 0032/477.91.86.68 Justin Stover University of Oxford, All Souls College, Oxford OX1 4AL, United Kingdom justin.stover@classics.ox.ac.uk Moshe Koppel Bar-Ilan University, 52900 Ramat-Gan, Israel moishk@gmail.com Folgert Karsdorp Center for Language Studies, Radboud University, P.O. Box 9103, NL-6500 HD, Nijmegen, The Netherlands fbkarsdorp@fastmail.nl Walter Daelemans University of Antwerp, Belgium, Prinsstraat 13, B-2000 Antwerp, Belgium walter.daelemans@uantwerp.be Abstract In this paper, we shed new light on the authenticity of the Corpus Caesarianum, a group of five commen- taries describing the campaigns of Julius Caesar (100-44 BC), the founder of the Roman empire. While Caesar himself has authored at least part of these commentaries, the authorship of the rest of the texts remains a puzzle that has persisted for nineteen centuries. In particular, the role of Caesar’s general Aulus Hirtius, who has claimed a role in shaping the corpus, has remained in contention. Determining the au- thorship of documents is an increasingly important authentication problem in information and computer science, with valuable applications, ranging from the domain of art history to counter-terrorism research. We describe two state-of-the-art authorship verification systems and benchmark them on 6 present-day evaluation corpora, as well as a Latin benchmark dataset. Regarding Caesar’s writings, our analysis allow us to establish that Hirtius’s claims to part of the corpus must be considered legitimate. We thus demonstrate how computational methods constitute a valuable methodological complement to traditional, expert-based approaches to document authentication. Keywords: Authentication, Authorship Verification, Stylometry, Julius Caesar Preprint submitted to Expert Systems with Applications May 26, 2016 1. Introduction Throughout the twentieth century, influential post-structuralist thinkers, such as Foucault or Barthes have fiercely argued against the importance of the notion of ‘authorship’ (Barthes, 1968; Foucault, 1969). Across many fields in the Humanities for instance, this famously led to a temporary devaluation of the importance attached to the relationship between texts and their original producers (Love, 2002). How-5 ever, numerous examples demonstrate that the public interest in authorship currently shows few signs of abating. The highly mediatized discovery of an pseudonymously published novel by the appraised Harry Potter novelist J.K. Rowling is a good example in this respect (Juola, 2015, 2013). In recent years, many other authorship-related research, such as the Shakespeare controversy (Burrows, 2012), has continued to make frequent headlines in the popular media. In academia too, the much debated application of bibliom-10 etry (Cronin, 2001) or well-known cases of plagiarism (Maurer et al., 2006) hardly suggest that the notion of authorship would have suffered a major loss of public interest. Unsurprisingly, automated authorship analysis (Juola, 2006; Koppel et al., 2009; Stamatatos, 2009b) currently receives increasing attention in Computer and Information Sciences too, as a form of document authentication with promising practical applications across various domains, such as plagiarism detection (Stein et al., 2011) or even in forensic15 sciences (Chaski, 2005; Juola, 2015). Most computational authorship studies in computer science are still restricted to present-day docu- ment collections. In this paper, we illustrate the broader applicability of computational authorship verifi- cation by reporting a high-profile case study from Classical Antiquity (Koppel & Seidman, 2013; Stover et al., 2016). The ‘War Commentaries’ by Julius Caesar (Corpus Caesarianum) refers to a group of Latin20 prose commentaries, describing the military campaigns of the world-renowned statesman Julius Caesar (100–44 BC), the founder of the Roman Empire. While Caesar must have authored a significant portion of these commentaries himself, the exact delineation of his contribution to this important corpus remains a controversial matter. Most notably, Aulus Hirtius – one of Caesar’s most trusted generals – is some- times believed to have contributed significantly to the corpus. Thus, the authenticity and authorship of25 the Caesarian corpus is a philological puzzle that has persisted for nineteen centuries. In this paper, we use computational authorship verification to shed new light on the matter. Below, we will first situate our work in the field of stylistic authentication studies, focusing on the style versus content debate, as well as the difference between open set and closed set attribution. We go on to discuss our implementation of two verification systems, a first-order and a second-order approach,30 which represent the state of the art in the field, given the results of the latest relevant competitions on 2 authorship verification. We first benchmark both systems on 6 present-day data sets, before testing them on an evaluation set of Latin documents from Antiquity. Finally, we analyse the Corpus Caesarianum, offering a detailed discussion of the historical implications of our results. 2. Style vs Content35 Traditionally, scholars have long employed a pragmatic distinction between the ‘style’ and ‘content’ of written documents (Stamatatos et al., 2000), the former encapsulating all aspects of an individual au- thor’s language use at the textual level (Hermann et al., 2015). In authorship studies, there is nowadays a general consensus that features related to style are more useful (Juola, 2006; Koppel et al., 2009; Sta- matatos, 2009b), since topical, content-related features vary much more strongly across the documents40 authored by a single individual. Much research nowadays therefore concerns ways to effectively ex- tract stylistic characteristics from documents that are not affected by a text’s specific content or genre (Argamon & Levitan, 2005; Kestemont et al., 2012; Efstathios, 2013; Sapkota et al., 2015; Seroussi et al., 2014; Sapkota et al., 2014). This has not always been the case: historical practitioners in earlier centuries, commonly based attributions on a much looser defined set of linguistic criteria, including, for instance,45 the use of conspicuous, rare words (Love, 2002; Kestemont, 2014). Naturally, an expert reader’s subjec- tive intuitions (Gelehrtenintuition, connoisseurship) would play a much larger role in studies than would nowadays be acceptable. Especially, the focus on striking characteristics would turn out to be problem- atic. Importantly, low-frequency features are typically tied to fairly specific topics, and thus do not scale well to new texts. More importantly, these whimsical items also appeal to imitators and followers: in50 the case of malignant forgeries or benigne epigones, the authentication of documents will fail, if it is restricted to easy-to-copy, low-frequency characteristics (Love, 2002). The pioneering work by Mosteller and Wallace on the pseudonymously published Federalist papers has marked a turning point in this respect (Mosteller & Wallace, 1964). Mosteller and Wallace proposed to rigidly restrict analyses to high-frequency characteristics and only considered an author’s use of function55 words, or the small and closed set of grammatical items in a language which – as opposed to content words as nouns or verbs – do not carry a straightforward semantics when used in isolation (e.g. the article ‘the’ or the preposition ‘of’) (Aronoff & Fudeman, 2005). For authorship studies, function words are extremely attractive: they are frequent and well-distributed variables across documents, and consequently, they are not specifically linked to a single topic or genre. Importantly, psycholinguistic research suggests that60 grammatical morphemes are less consciously controlled in human language processing, since they do not actively attract cognitive attention (Stamatatos, 2009b; Binongo, 2003; Argamon & Levitan, 2005; Peng et al., 2003). This suggests that function words are relatively resistant to stylistic imitation or forgery. 3 With respect to function words, a number of recent developments are relevant. Ever since the Fed- eralist papers, research into English-language documents has dominated authorship studies. In English,65 many functional morphemes are realised as individual words which can be easily identified in running text (Aronoff & Fudeman, 2005). In recent decades, the attention for other, low-resource languages has increased, including languages that display a much higher level of word inflection (e.g. the Finno-Ugric family) (Rybicki & Eder, 2011). Until fairly recently, other types of style markers (e.g. syntactical), rarely outperformed simple, word-level style markers (Holmes, 1994, 1998; Halteren et al., 2005). Later,70 character n-grams were introduced as a powerful alternative to function words (Kjell, 1994; Daelemans, 2013). This representation from Information Retrieval (originally used for automatic language identi- fication) models texts at the sub-word level and segments them into a series of consecutive, partially overlapping groups of n characters; under a third-order trigram model (n = 3), for instance, the word ‘trigrams’ would yield the n-grams {‘tri’, ‘rig’, ‘gra’, ‘ram’, ‘ams’}.75 Multiple studies have demonstrated the excellent performance of character n-grams for modelling authorship, especially when it comes to more highly inflected languages such as Latin (Sidorov et al., 2014; Efstathios, 2013). This modelling strategy has the advantage that it can also capture morphemic information at the subword level, and is thus potentially sensitive to functional morphemes that are not realised as individual words (e.g. word endings) (Kestemont, 2014; Sapkota et al., 2015; Stamatatos,80 2009b). Similarly, certain n-grams also pick up word stems and research increasingly demonstrates that text representations based on function words can be supplemented with information from lower- frequency strata in languages (Burrows, 2007), such as word stems (Koppel et al., 2009). Naturally, such approaches carefully need to avoid overfitting on the content of a specific document collection. Recent research demonstrated that predominantly functional character n-grams (including punctuation (Grieve,85 2007)) are powerful authorship predictors (Sapkota et al., 2015). This helps explain why this family of features proves more robust with respect to cross-genre problems (Efstathios, 2013; Sapkota et al., 2014). Other recent studies have successfully applied Bayesian topic models to automatically separate style from content (Seroussi et al., 2014). This paper will not dwell on feature selection, although we recognise the substantial efforts and ad-90 vances which have been made on the topic of feature engineering in authorship studies. We limit the stylistic properties studied below to two commonly used feature types: word unigrams and character n- grams. These feature types have the advantage that they can be easily extracted from corpora, without requiring the application of preprocessing tools, such as part-of-speech taggers or parsers, which might not be available for all languages. Their relevance has moreover clearly motivated in the existing litera-95 ture (Daelemans, 2013; Kestemont, 2014; Sapkota et al., 2015; Stamatatos, 2009b). While many studies 4 have indeed reported the successful application of other feature types (Stamatatos, 2009b), it is clear from comparative experiments that word unigrams and character n-grams represent state of the art feature types in authorship studies. 3. Methods100 A number of different experimental procedures should be distinguished in present-day authorship studies (Stamatatos, 2009b). A first important distinction is that between authorship attribution and au- thorship verification (also known as open-set attribution). In the simple attribution scenario, the task is to attribute an anonymous text to a known author, through selecting the correct author from a set of candidate authors. In this closed-set scenario, the algorithm can safely assume that the correct target105 author is present in the set of available candidate authors, a scenario resembling a police line-up. It has been shown that the difficulty of this task increases as the number of candidate authors grows, and the length and or number of the available texts decreases (Daelemans & Van den Bosch, 2005). While the attribution setup is not incompletely unrealistic, it has been noted that in many real-world applications, it cannot be guaranteed that a text’s true author is present among the candidates. This is why the verification110 scenario was introduced, in which the task is to decide whether or not an anonymous text was written by a given candidate author (hence, verification). The verification setup is known to be a more generic, yet also more difficult setup. Recent research has explored interested ways of combining both attribution and verification in a single system (Puig et al., 2016), although both setups are usually treated separately. The Caesarian corpus under scrutiny is a textbook example of a problem in authorship verification, since115 we do not have any guarantees as to the identity of the authors involved. For this paper, we will there- fore use generic implementations of two verification methods which represent the state of the art in the field, especially when looking at the results of the latest PAN competitions. Both systems have proven to be successful approaches to authorship verification, and many of the top-performing contestants in competitions have integrated variations of them.120 Authorship verification is a problem which has been studied for a number of years in the annual PAN competition. The design and evaluation of our analyses closely adheres to this competition’s conventions to increase the comparability of our results (Stamatatos et al., 2014). Each dataset in the PAN competition consists of a set of ‘problems’, in which at least one, but possible more ‘known’ documents are available, which were all written by the same target author. Additionally, each problem defined an ‘unknown text’125 for which has to be determined whether or not it has been written by the author of the ‘known’ texts, through assigning a score between 0 (definitely not the same author) and 1 (definitely the same author), 5 with a threshold at .5. Systems are allowed to leave a selection of difficult problems unanswered by as- signing a score of exactly .5. The problems in each dataset fell apart in two non-overlapping sets: one development set of problems, on which systems could be calibrated, and a roughly equal-sized set of test130 problems, on which the calibrated systems were evaluated. The performance of the submitted systems is evaluated on the basis of two metrics: the AUC score (area under the curve, a well-known scalar eval- uation score for binary classifiers) and the more recently proposed c@1 score (Peñas & Rodrigo, 2011). Unlike the AUC score, c@1 extends the traditional accuracy score (i.e. the ratio of correct answers), by rewarding careful systems that choose to leave those problems unanswered which it considers too diffi-135 cult. The final performance of systems is reported as the product of the AUC and c@1 metric. Following the conventions used at the PAN competition, we statistically compare the accuracy of classifiers using approximate randomisation: this non-parametric test is valuable it does not make assumptions about the (potentially highly complex) distributions of the compared system outputs. 3.1. Verification Systems140 The first verification system (termed O1 here) used here was seminally introduced by Kjell et al. (Kešelj et al., 2003) and was subsequently refined (Potha & Stamatatos, 2014; Kestemont et al., 2011; Stamatatos, 2009a). O1 resorts to the direct (or ‘first order’) calculation of a distance metric between a target author’s stylistic profile in a given problem, and the unknown text. Following (Potha & Stamatatos, 2014; Koppel & Seidman, 2013), we define an author’s profile here as the mean centroid of the known145 document vectors for that author (i.e. we average an author’s score for a particular term across all training texts). Originally, O1 was introduced with a specific distance metric, called ‘common n-grams’ (cng). Let A and B be the respective vectors representing an author’s centroid and the unknown document respectively; consisting of n character n-gram values in some fixed order. Let ai and bi represent the value of the i-th feature in both documents respectively:150 cng(A, B) = nX i=1 ✓ 2(ai � bi) ai + bi ◆2 (1) Studies vary in their exact implementation of this method: the earliest papers would calculate this distance function only for character n-grams which were present in both the profile and the unknown document (hence ‘common’ n-grams), but subsequent research showed that it is beneficial to apply the distance function only to the items which are present in the unknown document (Stamatatos, 2007), so that we use this implementation. To verify whether the unknown document was written by the target155 author in the problem, O1 uses thresholding: unknown documents resulting in a distance below this threshold are attributed to the target author, while all others are not. To normalize the resulting distance 6 score to probability scores in the 0-1 range, they are scaled using the set of all non-zero pairwise scores which can obtained between the known documents in a problem set, before their positive complement is taken (Potha & Stamatatos, 2014). While O1 has so far primarily been used with the cng metric, it can160 also be used with the other distance metrics introduced below. The second verification system (termed O2 here) is a generic implementation of the General Imposters (GI) framework (Koppel & Winter, 2014). The general intuition behind the GI, is not to assess whether two documents are simply similar in writing style, given a static feature vocabulary, but rather, it aims to assess whether two documents are significantly more similar to one another than other documents,165 across a variety of stochastically impaired feature spaces (Stamatatos, 2006; Eder, 2012), and compared to random selections of so-called distractor authors (Juola, 2015), also called ‘imposters’. O1 relies on the calculation of a direct, first-order distance measure between two documents to assess whether they are similar enough to be attributed to the same individual. The GI, however, resorts to the calculation of a ‘second-order’ metric (see Alg. 1, SI). Let x be the vector representing an anonymous document which is170 compared to T = {t1, . . . , tn}, a set of documents by the target author. The task is to determine whether the documents in T were or were not written by the same author as x. Additionally, the GI procedure has access to I = {i1, . . . , in}, a set of distractor documents by so-called imposter authors. The GI then starts a bootstrapped procedure: during k iterations, it randomly samples a subset of the available features, as well as a random subset of imposters from I as I0. In each iteration, we determine whether175 x is closer than any of the documents in T than in I0, given the impaired feature space and a distance function. Instead of returning a first-order distance, the GI returns a second-order metric, indicating the proportion of iterations in which x was closer to an item in T than in I0. As a proportion, the second-order score produced by O2 will automatically lie between 0 and 1 (higher scores indicate a higher attribution confidence). A similar thresholding procedure is therefore applied as with O1. O2 too can used with a180 variety of distance metrics, including the cng metric used in O1. Note that O2 is an example of an ‘extrinsic’ verification method (Juola & Stamatatos, 2013): as opposed to the ‘intrinsic’ setup of O1, O2 also uses known documents from other authors in a particular problem set. In this paper, we sample imposter authors from the known documents that are available for other authors in a particular problem set. To ensure the comparability of O1 and O2, we sample author185 profiles (i.e. mean centroids), instead of individual documents from the imposter pool. Previous studies have automatically crawled the web for useful imposter documents, which yields results that might be difficult to reproduce exactly. Additionally, there is the inherent danger that one might obtain imposter documents that were indeed written by the target author, which would compromise the proper working of O2. Naturally, this problem is even more real in the case of the Latin data sets used below, because of the190 7 relatively sparse online availability of Latin documents from Classical Antiquity. 3.2. Vector space models In technical terms, a collection of texts in authorship studies is typically represented using a vector space model (VSM), as is common in text classification research (Sebastiani, 2002; Stamatatos et al., 2000). Both O1 and O2 are applied to such a VSM, yielding a matrix-like representation of a text collec-195 tions, in which each document is assigned an equal-sized vector, which numerically represents a selection of its stylistic and linguistic properties, also called features, such as word unigram frequencies (Salton & Buckley, 1988; Manning et al., 2008). This process of vectorization typically operates under a ‘bag-of- words’ assumption, which models the occurrence of items in a text, but is in many cases insensitive to their relative order or exact position in a document. A number of different VSMs are currently dominant,200 the choice for which clearly reflects the style vs content assumptions outlined above. The simplest vectorization model is the term-frequency model (tf), which records the relative fre- quency of the individual terms (e.g. words or n-grams) in a document in some fixed order. In authorship studies, it is not uncommon to aggressively truncate such VSMs to the most frequent items in the docu- ment collection (sometimes as little as 30 items (Burrows, 2002)). This truncation is a simple yet efficient205 strategy to combat vector sparsity and automatically causes models to focus on functional morphemes, since grammatical items are typically the most frequent ones in corpora (Stamatatos, 2009b). When al- lowing larger vectors, the tf-model has the disadvantage that it quickly comes to suffer from sparsity artefacts. Additionally, tf assigns equal weights to stylistic properties across the frequency spectrum in a language; therefore, it does not provide any form of feature weighing.210 Another commonly used VSM is the tf � idf-model from Information Retrieval (Manning et al., 2008). The tf � idf model extends the plain tf-model by weighing a word with its inverse document frequency (idf) in the collection. Thus, rare words that are present in only a few documents will be attached more importance. In many ways, this model can be contrasted with the assumption that low- frequency items are bad predictors of authorial style (Binongo, 2003). Nevertheless, a few studies suggest215 that it might be useful (Koppel & Winter, 2014). Arguably, this model captures the intuition that if a highly rare feature is present in two documents, this increases the likelihood that the two documents were authored by the same individual. While the method might therefore be sensitive to overfitting on low- frequency properties, this might be an attractive characteristic in certain (e.g. single-domain) authorship problems.220 Thirdly, there is the std-model which weighs the tf-model through scaling term frequencies by their standard deviation across the document in the corpus. The model has initially been suggested by Burrows 8 (Burrows, 2002) as part of a memory-based learning system for authorship attribution and was later theoretically simplified (Argamon, 2008). A similar approach has been proposed in (Kešelj et al., 2003). This model captures the inverse intuition of the tf �idf model, since it will boost the performance of very225 common items in a document collection, which will have a relatively low standard deviation in tf. This is highly uncommon in other applications in Information Sciences (e.g. document retrieval), although the model has been shown to work surprisingly well for authorship attribution in many studies (Stamatatos, 2009b). 3.3. Distance metrics230 Both O1 and O2 crucially depend on distance metrics which can be applied to two vectors, in this case a vector representing an author’s profile and a vector representing an unknown document. In authorship studies, it is a well known fact that the choice for a particular distance metric has a clear effect on the performance of systems (Evert et al., 2015), which is why distance metrics have continued to attract a lot of attention in authorship studies (Kešelj et al., 2003; Hoover, 2004; Stamatatos, 2007; Smith &235 Aldridge, 2011; Luyckx & Daelemans, 2011; Jockers et al., 2008; Evert et al., 2015). Previous studies have amply shown that specific metrics might behave and perform rather differently in different problem setups, stressing the fundamental ad hoc nature of many authorship problems (Juola, 2006; Evert et al., 2015). While many variations have been proposed, only a small set of metrics (or slight variations thereof) seem to have yielded consistent and good performance across studies. The traditional ‘Manhattan’ city240 block distance is a popular choice, which defines the difference between two documents as the sum of the absolute differences between all features. The city block distance predominantly works well for small and dense VSMs, with very limited vocabularies, such as small sets of function word frequencies. Cosine-based metrics are known to scale better to larger, sparse vectors, and they are therefore more common in Information Sciences (Manning et al., 2008). The cosine distance, for instance, is a pseudo-245 distance measure based on the complement (in positive space) of the angular cosine similarity between two document vectors. In this paper, we will also compare these more established metrics to the still fairly novel minmax measure (Koppel & Winter, 2014), originally introduced in geobotanics by M. Ružička (Ružička, 1958). While the metric has re-emerged a number of times in different disciplines (e.g. as the ‘Jaccardized250 Czekanowski index’ (Schubert & Telcs, 2014)), the method is only a recent addition to authorship studies. In mathematical notation, the minmax measure was originally formulated as the following similarity measure (Cha, 2007). Let a and b represent two document vectors, consisting of n features in some fixed order. Let ai and bi represent the value of the i-th feature in both documents respectively (e.g. the relative 9 frequencies of a particular word in both documents, in the case of the simple tf-model):255 minmax(a, b) = nP i=1 min(ai, bi) nP i=1 max(ai, bi) (2) We turn this similarity metric into a true distance measure by taking its complement in positive space (Schubert & Telcs, 2014): 1 � minmax(a, b). So far, minmax has only been studied in the context of larger verification systems (Koppel & Seidman, 2013; Koppel & Winter, 2014; Seidman, 2013; Khonji & Iraqi, 2014), so that its individual contribution has not been clearly studied yet. More importantly, its per- formance has not rigorously been compared yet to other distance measures, under different experimental260 setups or in combination with different VSMs. In this paper, we will therefore elucidate the interplay of this distance metric and the VSMs described. In the context of the tf � idf model, for instance, the minmax metric will naturally boost the importance of features with larger values (i.e. those that are highly document-specific), whereas the opposite will happen in the std-model. We will empirically investigate the effect of this additional feature weighing.265 4. Benchmark results 4.1. PAN data To demonstrate the overall validity of our approach, we first benchmark O1 and O2 on 6 publicly available benchmark corpora which have been used in the 2014 edition of the PAN competition on au- thorship verification (Stamatatos et al., 2014) (pan.webis.de). In this yearly competition, teams can270 participate in a number of challenges involving forensic text analysis, such as plagiarism detection or authorship classification tasks. The organizers release training data that teams can independently develop systems on, before submitting their software. The organizers then run the software on new, unseen test data and rank the submitting teams according to their performance. We focus on the authorship verifica- tion track which has been organised since a number of years. The PAN 2014 verification datasets (see275 SI) only concern present-day writing samples, and vary strongly in both nature, size and difficulty, so that they provide a solid point of reference. The availability of the results reported by competitors on a fixed test set, moreover makes it easy to compare our results to the best performing systems which were entered into the competition. We report our full results in the SI and limit the discussion in the main text to a sample of illustrative examples. First, we calibrate O1 and and O2 on the development problems280 and then apply both systems to the test problems, reporting the AUC · c@1 for the test problems. In the SI, we report results for each combination of a VSM and distance metric, for the following feature 10 pan.webis.de types: word unigrams, character trigrams, and character tetragrams. For each feature type, we used VSMs that represent full vocabularies. To assess whether O1 and O2 produce significantly different results, we have applied an approximate randomisation test to each pair of scores from O1 and O2. Table 1 gives a285 representative list of results in terms of AUC · c@1, namely the test results for using word unigrams in each corpus, for O1 and O2. For each problem set, we also list the performance of the best-performing individual system in that task, as well as the meta-classifier trained on all submitted systems (which often, but not always, yields the strongest overall result) (Stamatatos et al., 2014). 11 Table 1: A representative list of the main verification results on the PAN corpora in terms of AUC · c@1, namely the test results for using word unigrams in each corpus, for O1 and O2. For each problem set, we also list the performance of the best-performing individual system in that task, as well as the meta-classifier trained on all submitted systems (which often, but not always, yields the strongest overall result) (Stamatatos et al., 2014) Combination Dutch essays Dutch reviews English essays English novels Greek articles Spanish articles O1 cng - tf-std 76.89 31.95 23.94 22.26 28.79 59.54 cng - tf-idf 76.81 32.62 24.38 22.04 28.89 60.50 cng - tf 75.85 31.32 23.40 21.83 28.21 61.95 cosine - tf-std 67.51 27.91 23.61 50.22 42.65 57.20 cosine - tf-idf 61.41 27.50 16.58 33.68 44.33 50.71 cosine - tf 48.11 36.79 28.77 27.60 41.13 47.50 minmax - tf-std 71.32 34.66 25.46 45.14 54.98 45.67 minmax - tf-idf 76.95 42.09 24.22 52.69 59.33 45.63 minmax - tf 70.16 40.32 27.12 47.45 67.47 76.98 manhattan - tf-std 61.93 35.06 23.64 21.31 27.37 40.44 manhattan - tf-idf 71.44 37.08 24.33 43.44 33.52 68.51 manhattan - tf 76.35 34.59 23.92 40.61 42.80 67.83 O2 cng - tf-std 83.8 35.91 26.49 34.72 48.26 74.25 cng - tf-idf 81.70 36.55 27.95 35.47 48.83 73.10 cng - tf 80.26 35.99 27.25 35.80 50.90 80.54 cosine - tf-std 87.50 33.58 29.12 37.35 50.20 63.41 cosine - tf-idf 90.96 36.82 18.16 33.80 41.26 64.54 cosine - tf 76.59 36.58 24.95 30.55 48.63 69.52 minmax - tf-std 89.52 38.78 35.13 36.48 57.33 71.61 minmax - tf-idf 93.70 48.42 30.40 40.66 67.32 73.03 minmax - tf 87.44 38.90 30.50 36.93 67.57 83.77 manhattan - tf-std 47.96 33.05 25.31 22.65 27.99 32.94 manhattan - tf-idf 74.22 35.37 27.36 37.61 35.59 50.69 manhattan - tf 84.58 37.38 28.06 37.43 56.68 63.69 2014 Meta-classifier 86.70 42.80 53.10 50.80 72.00 70.90 2014 Best single system 82.30 52.50 51.30 47.60 63.50 69.80 12 A number of high-level trends emerge. The results immediately illustrate the large differences in290 overall difficulty which exist between the various data sets, ranging from the good scores which can be obtained for relative easy corpus of Dutch-language essays, to the more difficult corpus of English essays. Overall, O2 typically yields a higher performance than O1, although O1 produce the single highest scores for the English novels, where the length of documents is considerably longer than elsewhere. In two problem sets, the Dutch essays and Spanish articles, O2 and O1 respectively yield surprisingly strong295 results, even outperforming the meta-classifier and top-performing in the PAN competition. In the Dutch reviews and Greek articles, the performance of O2 can be characterised as very decent, with a performance between between the meta-classifier and that of the best performing individual system. Interestingly, both O1 and O2 perform relatively poorly for the following two data sets: the English essays and English novels (where text length clearly affects performance). With respect to the former corpus, we hypothesise300 that this loss in performance for O2 is due to the fact that we did not crawl the web for suitable imposters (as other studies have done), but limited our distractor pool to the other known documents in the problem set (because of our focus on Latin documents below). In these particular corpora, the algorithm might suffer from sampling documents that are too similar in content to the unknown document to act as a useful comparand. As to the other feature types, the results show that manhattan only yields acceptable results305 for the character trigram features, which is an expected outcome, because character trigrams lead to a much denser corpus representation. For sparser representations, the minmax and cosine distance offer a much better fit. Especially in the case of word unigrams – which produce the strongest results across corpora – the novel minmax metric offers surprisingly strong results in comparison to the established metrics (it is part of every winning combination under O2). Interestingly, the effect of VSMs is much less310 pronounced than distance metrics: the minmax and cosine metric are generally least affected by a change in VSM. 4.2. Latin data We now proceed to benchmarking our system on a corpus of historic Latin authors. For this study we have collected a representative reference corpus, containing works by some of the main Latin prose315 authors from Classical Antiquity, such as Cicero, Seneca or Suetonius. They predominantly include his- toriographical texts (e.g. Livy’s Ab Urbe Condita) which are sufficiently similar to Caesar’s War Com- mentaries. All original texts were cut up in non-overlapping slices of 1000 words; while this constitutes a challengingly limited document size, this procedure allows us to obtain a sufficiently fine-grained anal- ysis of the Caesarian corpus. For modern documents, promising results are increasingly obtained with320 small document sizes (Koppel et al., 2013; Koppel & Winter, 2014), such as the PAN data used above. 13 To create a set of development and test problems, we proceed as follows. We split the available oeuvres at the author-level into two equal-sized sets. For each set we create a balanced set of same-author and different-author problems: for each true document-author pair, we also include a false document-author pair, whereby we randomly assign a different target author to the test document in question. This en-325 sures that there is no overlap between the development and test problems created: therefore we can now parametrize the system on the development set and evaluate it on the test set, in an entirely parallel fashion as with the PAN data. Figure 1: Precision-recall curves for each metric-VSM combination on the Latin benchmark data (test problems), using the O1 ‘first-order’ verification system. The c@1 score is listed in the legend. The cosine and minmax metric consistently yield higher results than cng and manhattan. In Figs. 1 and 2 we graphically show the results for O1 and O2 on the Latin benchmark corpus, again using untruncated vocabularies: for each combination of a VSM a distance metric, we plot a precision-330 14 Figure 2: Precision-recall curves for each metric-VSM combination on the Latin benchmark data (test problems), using the O2 ‘second-order’ verification system. The c@1 score is listed in the legend. Only the manhattan distance now yields inferior results: the bootstrapping greatly reduces the variation between the different metric-VSM combinations. recall curve; the c@1 score is listed in the legend (see SI for detailed results). The following trends clearly emerge: O2 consistently (in most cases significantly) outperforms O1 on the Latin data. O1 shows wildly diverging results, especially across different distance metrics, whereas the effect of VSMs is much less pronounced. In O2, both the cosine distance and minmax distance yield results that are clearly superior to cng and cityblock. Overall, O2 yields much stabler results across most combinations and for most335 combinations the curves can even not be visibly distinguished any longer. Unsurprisingly cityblock is the only metric which yields visibly inferior results for O2. In O2 too, the minmax and cosine distance overall yield the highest c@1, which is invariable in the upper nineties. Our evaluation shows that the recently introduced minmax metric yields a surprisingly good and consistent performance in comparison 15 to more established metrics. While it is not consistently the best performing metric, it produced highly340 stable results for the PAN data (and to a lesser extent for the Latin data). Overall, we hypothesize that the formulation of the minmax metric has a regularizing effect in the context of authorship studies. Due to its specific formulation, the minmax metric will automatically produce distances in the 0-1 range, in contrast to the more extreme distances which can be produced by e.g. Manhattan. Perhaps because of this, the minmax metric interacts well with both std and td � idf, although these VSMs capture inverse345 intuitions. Like cosine, which also naturally scales distances, minmax is relatively insensitive to the dimensionality of the VSM under which the metric is applied. 5. Caesar’s writings After benchmarking our verification systems, we now proceed to apply them to the Caesarian Corpus (Corpus Caesarianum), because it produced more stabler results for the benchmark data set (i.e. on350 average, it produced the highest results across different metric-vector space combinations). The Caesarian Corpus is composed of five commentaries describing Caesar’s military campaigns (Mayer, 2011; Gaertner & Hausburg, 2013): Gallic War Bellum Gallicum, conquest of Gaul, 58–50 BC; Civil War Bellum civile, civil war with Pompey, 49–48 BC;355 Alexandrian War Bellum Alexandrinum, Middle East campaigns, 48–47 BC; African War Bellum Africum, war in North Africa, 47 to 46 BC Spanish War Bellum Hispaniense, rebellion in Spain, 46–45 BC. The first two commentaries are mainly by Caesar himself, the only exception being the final part of the Gallic War (Book 8), which is commonly attributed to Caesar’s general Aulus Hirtius (c90 – 43360 BC). Caesar’s primary authorship of these two works, except for Book 8, is guaranteed by the ancient testimonia of Cicero, Hirtius, Suetonius, and Priscian as well as the unanimous evidence of the manuscript tradition. Caesar’s ancient biographer Suetonius, writing a century and a half after his death, suggests that either Hirtius or another general, named Oppius, authored the remaining works: ‘[Caesar] also left commentarii of his deeds during the Gallic War and the Civil War with Pompey. For the author of the365 Bellum Alexandrinum, Africum, and Hispaniense is uncertain. Some think it is Oppius, others Hirtius, who supplemented the last, incomplete book of the Bellum Gallicum’ (Appendix I). We also have a letter of Hirtius to Cornelius Balbus, a fellow supporter of Caesar, which is transmitted in the manuscripts 16 preceding the Hirtian 8th book of the Gallic War. In this letter, Hirtius lays out his project: ‘I have continued the accounts of out Caesar on his deeds in Gall, since his earlier and later writings did not fit370 together, and I have also finished the most recent and incomplete account, extending it from the deeds in Alexandria down to the end, not admittedly of civil discord, of which we seen no end, but of Caesar’s life’ (Gaertner & Hausburg, 2013). Despite occasional doubts, the most recent analysis has shown that there is no reason at all for doubt- ing the authenticity of the letter (Gaertner & Hausburg, 2013). Hence, a puzzle that has persisted for375 nineteen centuries: what are the relationships of the different war commentaries to one another, to Hir- tius, and to Caesar (Mayer, 2011)? Current scholarship has focused primarily on the authorship of the Alexandrian War. J. Gaertner and B. Hausburg (Gaertner & Hausburg, 2013) concluded that Hirtius knit together disparate sources to complete the text, including a Caesarian core of material in chapters 1–21. He also exercised a role in the formation of the whole corpus, though with much less firm editorial hand.380 Their analysis was based on a painstaking account of all sorts of evidence, including statistical analysis of usage and language. Their account represents the pinnacle of what can possibly by achieved by manual analytical methods, and offers a ripe target for re-analysis with automated computational methods. We are not the first to do so: in 2002 M. Trauth proposed a computer-assisted analysis of the Corpus which failed to reach any definitive conclusions on the authorship of the Bellum Alexandrinum, based on an385 automated tabulation of the most frequent words. (Trauth, 2002). More than a decade of advances in computational philology allow us to go beyond his inconclusive analysis. To shed new light on the authenticity of the Caesarian corpus, we proceed as follows. To obtain docu- ments of a similar size, we have divided all original commentaries in consecutive, non-overlapping slices of 1000 words and treat these slices as individual documents. We label these documents according to the390 assumption that the Gallic and Civil Wars were written by CAESAR, with the exception of 8th book of the former commentary, which we ascribe to HIRTIUS. To label the disputed authors of the Alexandrian, African and Spanish War, we use the provisional labels X, Y and Z respectively. Fig. 3 offers an initial in- spection of the stylistic structure in this corpus, in the spirit of the first-order distance-calculations of O1. We generated a square distance table using the minmax distance metric to every document pair in the Cae-395 sarian collection and we scaled the distances to the 0–1 range. Next, we plotted a heat map of the distance matrix, and ran a conventional cluster analysis on top of the rows and columns. For the generating the hierarchical dendrograms next to the heatmap, we used the default agglomerative clustering routine in the Seaborn library (https://web.stanford.edu/˜mwaskom/software/seaborn/), which is based on the pairwise Euclidean distance between entries and the average linkage method. The labels400 indicate the most plausible authorial provenance of a document (if known), given the annotation labels 17 https://web.stanford.edu/~mwaskom/software/seaborn/ we just described. This rather naive approach demonstrate a clear-cut distinction: a significant portion of the Bellum Alexandrinum (X) clusters with Hirtius’s contribution to the Gallic Wars, under a clade that is clearly separate from Caesar’s accepted writings. Thus, Hirtius’s writings are distinguished from Caesar’s own405 core contributions; Hirtius’s samples are compellingly close in style to X. Samples from the Alexandrian War appear to be stylistically close to Hirtius’s contribution to the Gallic Wars in Book 8 – which itself is surprisingly distinct from the other chapters in it. The more fundamental question now is how close these texts should truly be, in order to proceed to an actual attribution. We therefore turn to a more advanced analysis using O2. As with the problems in the benchmark experiments, each sample in the commentary410 collection was individually paired with the profile of all five Caesarian ‘authors’ available (including X, Y and Z): using the bootstrapped procedure from O2, we calculate a second-order similarity score by assess- ing in which proportion of a series of iterations one of these documents would be attributed to a particular Caesarian author’s profile, instead of a distractor author in the background corpus. This procedure as such yields, per document, 5 second-order scores, reflecting the probability that the sample must be attributed415 to a Caesarian’s authors profile, rather than an imposter. Following the outcome of the benchmark results, we perform this analysis for the five top-scoring metric-VSM combinations. Afterwards, we average the results over these five simulations and we graphically present the results in Fig. 4 (the full results are included in the SI). Note that in this setup we are especially interested in attribution leakage from one potential author to another: the fact that a text is attributed to the profile based on the other samples from420 its own text is an expected result; the attribution to another Caesarian ‘author’, however, is not. Our O2 analyses divide the Caesarian corpus into two branches at the top-level, which might be called ‘Caesarian’ and ‘non-Caesarian’. As we would expect, the Caesarian branch includes both the Civil War and the Gallic War, books 1–7. However, it also includes the first three samples from the Alexandrian War, providing dramatic confirmation of the theory of a Caesarian core in the first 21 chapters of the425 work. The other branch includes Gallic War, book 8, the rest of the Alexandrian War, the African War, and the Spanish War. The first two are closely affiliated with one another, indicating shared authorship. Stylistically there is no good reason for rejecting Hirtius’s authorship of the Alexandrian War, once we remove the Caesarian chapters 1–21. Gaertner and Hausburg (Gaertner & Hausburg, 2013) argue strongly against Hirtius’s authorship of the Alexandrian War, instead assigning him an amorphous role as editor430 of the corpus. It is true that the Alexandrian War shows far great heterogeneity that the Spanish War, for example, but it clearly clusters with the Gallic War, book 8, in a way the other texts do not, and displays no greater stylistic heterogeneity than Caesar’s own commentaries. 18 The African War and the Spanish War are the most internally consistent of the texts, perhaps an indication of separate authorship. They do, however, cluster with one another and with Hirtius, and the435 non-Caesarian texts all show a greater similarity with each other than with the Caesarian texts. While they are not stylistically homogenous enough to allow us to positive single-authorship in a naive sense, they display no greater stylistic heterogeneity than is present in the Caesarian texts. On both branches, we find the stylistic range we ought to expect in the genre of war commentaries, where commanders drawing up the official account of their campaigns would draw upon the dispatches of their legates and subordinates,440 sometimes integrating them into their own style, other times incorporating their texts with few changes. Importantly, Fig. 4 has an additional feature: whereas other X samples could be found scattered across Caesar’s authentic writings in the non-bootstrapped verification, O2 adds a distinct clade for these and a small set of other samples. This is a strong indication that the bootstrapped O2 system is not only able to distinguish authentic Caesarian material from non-authentic writings, but that it can even differentiate445 between a pure Caesarian style from the impure style resulting from collaborative authorship or the use of source texts. Hence, our analyses broadly supports the following conclusions: 1. Caesar himself wrote, in addition to Gallic Wars, books 1–7 and the Civil War, as well as the first 21 chapters of the Alexandrian War. 2. Hirtius wrote Book 8 of the Gallic Wars and the remainder of the Alexandrian War.450 3. At least one other author wrote the African War and the Spanish War. The African War and the Spanish War were probably written by two different authors. 4. Our results do not invalidate Hirtius’s own claim that he himself compiled and edited the corpus of the non-Caesarian commentaries. 5. The significant stylistic heterogeneity we have detected in parts of the Gallic War and the Civil War455 likely represents Caesar’s compositional practice of relying on, and sometimes incorporating, the briefs written for him by his legates. These findings are entirely consistent with a natural interpretation of Hirtius’s own words in his letter to Balbus, that he composed Gallic War, book 8 as a bridge between the preceding 7 books and the Civil War, that he completed the Alexandrian War, and added the two other commentaries to make the whole460 group a continuous narrative of Caesar’s campaigns. Chronologically the corpus thus ends in March, 45 BC with the Battle of Munda in Spain, but since we know that the end of the Spanish War is missing, there is no reason why we cannot assume that it originally continued with a brief epilogue bringing the narrative up to conclude with Caesar’s assassination in 44 BC. 19 6. Acknowledgements465 The authors would like to thank [anonymized] for their valuable feedback on earlier drafts of this arti- cle. Moshe Koppel acknowledges the support of the Intel Collaboration Research Institute for Computa- tional Intelligence. The work of Folgert Karsdorp has been supported by the Computational Humanities Programme of the Royal Netherlands Academy of Arts and Sciences, under the auspices of the Tunes & Tales project.470 7. References Argamon, S. (2008). Interpreting Burrows’s Delta: Geometric and probabilistic foundations. Literary and Linguistic Computing, 23, 131–147. Argamon, S., & Levitan, S. (2005). Measuring the usefulness of function words for authorship attribution. In Proceedings of the Joint Conference of the Association for Computers and the Humanities and the475 Association for Literary and Linguistic Computing (2005). Aronoff, M., & Fudeman, K. (2005). What is Morphology?. Blackwell. Barthes, R. (1968). La mort de l’auteur. Manteia, 5, 12–17. Binongo, J. (2003). Who Wrote the 15th Book of Oz? An application of multivariate analysis to author- ship attribution. Chance, (pp. 9–17).480 Burrows, J. (2002). ‘Delta’: A measure of stylistic difference and a guide to likely authorship. Literary and Linguistic Computing, 17, 267–287. Burrows, J. (2007). All the way through: Testing for authorship in different fre- quency strata. Literary and Linguistic Computing, 22, 27–47. URL: http://llc. oxfordjournals.org/content/22/1/27.abstract. doi:10.1093/llc/fqi067.485 arXiv:http://llc.oxfordjournals.org/content/22/1/27.full.pdf+html. Burrows, J. (2012). A second opinion on Shakespeare and Authorship Studies in the Twenty-First Cen- tury. Shakespeare Quarterly, 63, 355–392. Cha, S.-H. (2007). Comprehensive survey on distance/similarity measures between probability density functions. International Journal of Mathematical Models and Methods in Applied Sciences, 1, 300–490 307. 20 http://llc.oxfordjournals.org/content/22/1/27.abstract http://llc.oxfordjournals.org/content/22/1/27.abstract http://llc.oxfordjournals.org/content/22/1/27.abstract http://dx.doi.org/10.1093/llc/fqi067 http://arxiv.org/abs/http://llc.oxfordjournals.org/content/22/1/27.full.pdf+html Chaski, C. E. (2005). Who’s at the keyboard? authorship attribution in digital evidence investigations. International Journal of Digital Evidence, 4, 1–13. Cronin, B. (2001). Hyperauthorship: A postmodern perversion or evidence of a structural shift in schol- arly communication practices? Journal of the American Society for Information Science and Tech-495 nology, 52, 558–569. URL: http://dx.doi.org/10.1002/asi.1097. doi:10.1002/asi. 1097. Daelemans, W. (2013). Explanation in computational stylometry. In Computational Linguistics and Intelligent Text Processing (pp. 451–462). Springer. Daelemans, W., & Van den Bosch, A. (2005). Memory-Based Language Processing. Studies in Natural500 Language Processing. Oxford University Press. Eder, M. (2012). Computational stylistics and biblical translation: how reliable can a dendrogram be? In T. Piotrowski, & Ł. Grabowski (Eds.), The Translator and the Computer (pp. 155–170). Wrocław: WSF Press. Efstathios, S. (2013). On the robustness of authorship attribution based on character n-gram features.505 Journal of Law and Policy, 21, 421–439. Evert, S., Proisl, T., Schöch, C., Jannidis, F., Pielström, S., & Vitt, T. (2015). Explaining delta, or: How do distance measures for authorship attribution work? URL: http://dx.doi.org/10.5281/ zenodo.18308. doi:10.5281/zenodo.18308. Foucault, M. (1969). Qu’est-ce qu’un auteur? Bulletin de la Société française de philosophie, 3, 73–104.510 Gaertner, J., & Hausburg, B. (2013). Caesar and the Bellum Alexandrinum: An Analysis of Style, Narra- tive Technique, and the Reception of Greek Historiography. Vandenhoeck & Ruprecht. Grieve, J. (2007). Quantitative authorship attribution: An evaluation of techniques. Literary and Lin- guistic Computing, 22, 251–270. URL: http://llc.oxfordjournals.org/content/22/ 3/251.abstract. doi:10.1093/llc/fqm020.515 Halteren, H. V., Baayen, H., Tweedie, F., Haverkort, M., & Neijt, A. (2005). New machine learning methods demonstrate the existence of a human stylome. Journal of Quantitative Linguistics, 12, 65– 77. Hermann, J., Oskam K., V. D., & Schöch, C. (2015). Revisiting style, a key concept in literary studies. Journal of Literary Theory, 9, 25–52.520 21 http://dx.doi.org/10.1002/asi.1097 http://dx.doi.org/10.1002/asi.1097 http://dx.doi.org/10.1002/asi.1097 http://dx.doi.org/10.1002/asi.1097 http://dx.doi.org/10.5281/zenodo.18308 http://dx.doi.org/10.5281/zenodo.18308 http://dx.doi.org/10.5281/zenodo.18308 http://dx.doi.org/10.5281/zenodo.18308 http://llc.oxfordjournals.org/content/22/3/251.abstract http://llc.oxfordjournals.org/content/22/3/251.abstract http://llc.oxfordjournals.org/content/22/3/251.abstract http://dx.doi.org/10.1093/llc/fqm020 Holmes, D. (1994). Authorship attribution. Computers and the Humanities, 28, 87–106. Holmes, D. (1998). The evolution of stylometry in Humanities scholarship. Literary and Linguistic Computing, 13, 111–117. Hoover, D. L. (2004). Testing Burrows’s Delta. Literary and Linguistic Computing, 19, 453–475. doi:10. 1093/llc/19.4.453.525 Jockers, M. L., Witten, D. M., & Criddle, C. S. (2008). Reassessing authorship of the Book of Mormon using delta and nearest shrunken centroid classification. Literary and Linguistic Computing, 23, 465– 491. Juola, P. (2006). Authorship attribution. Foundations and Trends in Information Retrieval, 1, 233–334. Juola, P. (2013). Rowling and Galbraith: an authorial analysis. URL: http://languagelog.ldc.530 upenn.edu/nll/?p=5315. Juola, P. (2015). The Rowling Case: A Proposed Standard Analytic Protocol for Authorship Questions. Digital Scholarship in the Humanities, . doi:10.1093/llc/fqv040. Juola, P., & Stamatatos, E. (2013). Overview of the author identification task at PAN 2013. In Working Notes for CLEF 2013 Conference , Valencia, Spain, September 23-26, 2013..535 Kešelj, V., Peng, F., Cercone, N., & Thomas, C. (2003). N-gram-based author profiles for authorship attribution. In Proceedings of the Conference Pacific Association for Computational Linguistics, PA- CLING’03 (pp. 255–264). Dalhousie University, Halifax, Nova Scotia, Canada. Kestemont, M. (2014). Function words in authorship attribution. from black magic to theory? In Pro- ceedings of the 3rd Workshop on Computational Linguistics for Literature (pp. 59–66). Association540 for Computational Linguistics. Kestemont, M., Luyckx, K., & Daelemans, W. (2011). Intrinsic plagiarism detection using character trigram distance scores - notebook for PAN at CLEF 2011. In CLEF 2011 Labs and Workshop, Note- book Papers, 19-22 September 2011, Amsterdam, The Netherlands. URL: http://ceur-ws.org/ Vol-1177/CLEF2011wn-PAN-KestemontEt2011.pdf.545 Kestemont, M., Luyckx, K., Daelemans, W., & Crombez, T. (2012). Cross-genre authorship verification using unmasking. English Studies, 93, 340–356. 22 http://dx.doi.org/10.1093/llc/19.4.453 http://dx.doi.org/10.1093/llc/19.4.453 http://dx.doi.org/10.1093/llc/19.4.453 http://languagelog.ldc.upenn.edu/nll/?p=5315 http://languagelog.ldc.upenn.edu/nll/?p=5315 http://languagelog.ldc.upenn.edu/nll/?p=5315 http://dx.doi.org/10.1093/llc/fqv040 http://ceur-ws.org/Vol-1177/CLEF2011wn-PAN-KestemontEt2011.pdf http://ceur-ws.org/Vol-1177/CLEF2011wn-PAN-KestemontEt2011.pdf http://ceur-ws.org/Vol-1177/CLEF2011wn-PAN-KestemontEt2011.pdf Khonji, M., & Iraqi, Y. (2014). A Slightly-modified GI-based Author-verifier with Lots of Features (ASGALF). In Working Notes for CLEF 2014 Conference, Sheffield, UK (pp. 977–983). Kjell, B. (1994). Discrimination of authorship using visualization. Information Processing and Manage-550 ment, 30, 141–50. Koppel, M., Schler, J., & Argamon, S. (2009). Computational methods in authorship attribution. Journal of the American Society for Information Science and Technology, 60, 9–26. Koppel, M., Schler, J., & Argamon, S. (2013). Authorship attribution: What’s easy and what’s hard? Journal of Law & Policy, 21, 317–331.555 Koppel, M., & Seidman, S. (2013). Automatically identifying pseudepigraphic texts. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (pp. 1449–1454). Asso- ciation for Computational Linguistics. Koppel, M., & Winter, Y. (2014). Determining if two documents are written by the same author. Journal of the Association for Information Science and Technology, 65, 178–187.560 Love, H. (2002). Attributing authorship. An introduction. Cambridge: Cambridge University Press. Luyckx, K., & Daelemans, W. (2011). The effect of author set size and data size in authorship attribution. Literary and Linguistic Computing, 26, 35–55. Manning, C. D., Raghavan, P., & Schütze, H. (2008). Introduction to Information Retrieval. New York, NY, USA: Cambridge University Press.565 Maurer, H., Kappe, F., & Zaka, B. (2006). Plagiarism - A Survey. Journal of Universal Computer Science, 12, 1050–1084. Mayer, M. (2011). Caesar and the corpus caesarianum. In G. Marasco (Ed.), Political autobiographies and memoirs in antiquity: A Brill companion (pp. 189–232). Brill. Mosteller, F., & Wallace, D. (1964). Inference and disputed authorship: The Federalist. Addison-Wesley.570 Peñas, A., & Rodrigo, A. (2011). A simple measure to assess non-response. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies - Volume 1 HLT ’11 (pp. 1415–1424). Stroudsburg, PA, USA: Association for Computational Linguis- tics. URL: http://dl.acm.org/citation.cfm?id=2002472.2002646. 23 http://dl.acm.org/citation.cfm?id=2002472.2002646 Peng, F., Schuurmans, D., Wang, S., & Keselj, V. (2003). Language independent authorship attribution575 using character level language models. In Proceedings of the Tenth Conference on European Chapter of the Association for Computational Linguistics - Volume 1 EACL ’03 (pp. 267–274). Stroudsburg, PA, USA: Association for Computational Linguistics. URL: http://dx.doi.org/10.3115/ 1067807.1067843. doi:10.3115/1067807.1067843. Potha, N., & Stamatatos, E. (2014). A profile-based method for authorship verification. In580 A. Likas, K. Blekas, & D. Kalles (Eds.), Artificial Intelligence: Methods and Applications (pp. 313–326). Springer International Publishing volume 8445 of Lecture Notes in Computer Sci- ence. URL: http://dx.doi.org/10.1007/978-3-319-07064-3_25. doi:10.1007/ 978-3-319-07064-3_25. Puig, X., Font, M., & Ginebra, J. (2016). A unified approach to authorship attri-585 bution and verification. The American Statistician, advance access, x. URL: http: //dx.doi.org/10.1080/00031305.2016.1148630. doi:10.1080/00031305.2016. 1148630. arXiv:http://dx.doi.org/10.1080/00031305.2016.1148630. Ružička, M. (1958). Anwendung mathematisch-statistischer methoden in der geobotanik (synthetische bearbeitung von aufnahmen). Biológia (Bratislava), 13, 647–661.590 Rybicki, J., & Eder, M. (2011). Deeper Delta across genres and languages: do we really need the most frequent words? Literary and Linguistic Computing, (pp. 315–321). Salton, G., & Buckley, C. (1988). Term-weighting approaches in automatic text retrieval. Information processing & management, 24, 513–523. Sapkota, U., Bethard, S., Montes, M., & Solorio, T. (2015). Not all character n-grams are created equal:595 A study in authorship attribution. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 93– 102). Denver, Colorado: Association for Computational Linguistics. URL: http://www.aclweb. org/anthology/N15-1010. Sapkota, U., Solorio, T., Montes-y-Gómez, M., Bethard, S., & Rosso, P. (2014). Cross-topic authorship600 attribution: Will out-of-topic data help? In COLING 2014, 25th International Conference on Com- putational Linguistics, Proceedings of the Conference: Technical Papers, August 23-29, 2014, Dublin, Ireland (pp. 1228–1237). URL: http://aclweb.org/anthology/C/C14/C14-1116.pdf. 24 http://dx.doi.org/10.3115/1067807.1067843 http://dx.doi.org/10.3115/1067807.1067843 http://dx.doi.org/10.3115/1067807.1067843 http://dx.doi.org/10.3115/1067807.1067843 http://dx.doi.org/10.1007/978-3-319-07064-3_25 http://dx.doi.org/10.1007/978-3-319-07064-3_25 http://dx.doi.org/10.1007/978-3-319-07064-3_25 http://dx.doi.org/10.1007/978-3-319-07064-3_25 http://dx.doi.org/10.1080/00031305.2016.1148630 http://dx.doi.org/10.1080/00031305.2016.1148630 http://dx.doi.org/10.1080/00031305.2016.1148630 http://dx.doi.org/10.1080/00031305.2016.1148630 http://dx.doi.org/10.1080/00031305.2016.1148630 http://dx.doi.org/10.1080/00031305.2016.1148630 http://arxiv.org/abs/http://dx.doi.org/10.1080/00031305.2016.1148630 http://www.aclweb.org/anthology/N15-1010 http://www.aclweb.org/anthology/N15-1010 http://www.aclweb.org/anthology/N15-1010 http://aclweb.org/anthology/C/C14/C14-1116.pdf Schubert, A., & Telcs, A. (2014). A note on the Jaccardized Czekanowski similarity index. Scientomet- rics, 98, 1397–1399. doi:10.1007/s11192-013-1044-2.605 Sebastiani, F. (2002). Machine learning in automated text categorization. ACM Computing Surveys, 34, 1–47. Seidman, S. (2013). Authorship verification using the impostors method. In CLEF 2013 Evaluation Labs and Workshop-Online Working Notes. Seroussi, Y., Zukerman, I., & Bohnert, F. (2014). Authorship attribution with topic models. Comput. Lin-610 guist., 40, 269–310. URL: http://dx.doi.org/10.1162/COLI_a_00173. doi:10.1162/ COLI_a_00173. Sidorov, G., Velasquez, F., Stamatatos, E., Gelbukh, A., & Chanona-Hernndez, L. (2014). Syntactic n-grams as machine learning features for natural language processing. Expert Systems with Applica- tions, 41, 853 – 860. URL: http://www.sciencedirect.com/science/article/pii/615 S0957417413006271. doi:http://dx.doi.org/10.1016/j.eswa.2013.08.015. Methods and Applications of Artificial and Computational Intelligence. Smith, P. W. H., & Aldridge, W. (2011). Improving authorship attribution: Optimizing Burrows’ Delta method. Journal of Quantitative Linguistics, 18, 63–88. Stamatatos, E. (2006). Authorship attribution based on feature set subspacing ensembles. International620 Journal on. Artificial Intelligence tools, 15, 823–838. Stamatatos, E. (2007). Author identification using imbalanced and limited training texts. In Proceedings of the 18th International Conference on Database and Expert Systems Applications DEXA ’07 (pp. 237–241). Washington, DC, USA: IEEE Computer Society. URL: http://dx.doi.org/10. 1109/DEXA.2007.41. doi:10.1109/DEXA.2007.41.625 Stamatatos, E. (2009a). Intrinsic plagiarism detection using character n-gram profiles. In Third PAN Workshop. Uncovering Plagiarism, Authorship and Social Software Misuse (pp. 38–46). Stamatatos, E. (2009b). A survey of modern authorship attribution methods. Journal of the Association for Information Science and Technology, 60, 538–556. Stamatatos, E., Daelemans, W., Verhoeven, B., Stein, B., Potthast, M., Juola, P., Sánchez-Pérez, M. A., &630 Barrón-Cedeño, A. (2014). Overview of the author identification task at PAN 2014. In Working Notes for CLEF 2014 Conference, Sheffield, UK, September 15-18, 2014. (pp. 877–897). 25 http://dx.doi.org/10.1007/s11192-013-1044-2 http://dx.doi.org/10.1162/COLI_a_00173 http://dx.doi.org/10.1162/COLI_a_00173 http://dx.doi.org/10.1162/COLI_a_00173 http://dx.doi.org/10.1162/COLI_a_00173 http://www.sciencedirect.com/science/article/pii/S0957417413006271 http://www.sciencedirect.com/science/article/pii/S0957417413006271 http://www.sciencedirect.com/science/article/pii/S0957417413006271 http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2013.08.015 http://dx.doi.org/10.1109/DEXA.2007.41 http://dx.doi.org/10.1109/DEXA.2007.41 http://dx.doi.org/10.1109/DEXA.2007.41 http://dx.doi.org/10.1109/DEXA.2007.41 Stamatatos, E., Kokkinakis, G., & Fakotakis, N. (2000). Automatic text categorization in terms of genre and author. Computational Linguistics, 26, 471–495. Stein, B., Lipka, N., & Prettenhofer, P. (2011). Intrinsic plagiarism analysis. Language Resources and635 Evaluation, 45, 63–82. Stover, J., Winter, Y., Koppel, M., & Kestemont, M. (2016). Computational authorship verification method attributes a new work to a major 2nd century African author. Journal of the American Society for Information Science and Technology, 67, 239–242. Trauth, M. (2002). Caesar incertus auctor. Ein quantifizierendes Wort zur Kritik von Verfassersfragen640 in Lateinischen Texten. In J. Jährling, U. Meves, & E. Timm (Eds.), Röllwagenbüchlein. Festschrift Walter Röll (pp. 313–334). Niemeyer. 26 Figure 3: Naive heatmap visualisation of the stylistic structure in the Corpus Caesarianum, based on the scaled, pairwise distance matrix on the basis of the first-order minmax distance metric and the tf VSM (full vocabulary). Conventional clustering was ran on top of rows and columns, representing non-overlapping 1000-word samples from the text. A significant portion of the Bellum Alexandrinum (labeled X) clusters with Hirtius’s contribution to the Gallic Wars, under a clade that is separate from Caesar’s accepted writings. 27 Caesar Hirtius Caesar et al. (?) Figure 4: Cluster and heatmap visualisation of the results of the O2 verification procedure on the Caesarian corpus. Cell values represent the average probability of a sample being attributed to one of the five profiles distinguished. Five independent analysis were run with the 5 top-performing metric-VSM combination in the benchmark section. O2 seems not only able to distinguish authentic Caesarian material from non-authentic writings, but arguably also differentiates between a ‘pure’ Caesarian style and the mixed style resulting from e.g. the general’s dependence on pre-existing briefs by legates. 28 Highlights 1. We shed new light on the authenticity of the writings of Julius Caesar. 2. Hirtius, one of Caesar’s generals, must have contributed to Caesar’s writings. 3. We benchmark two authorship verification systems on publicly available data sets. 4. We test on both modern data sets, and Latin texts from Antiquity. 5. We show how computational methods inform traditional authentication studies. Introduction Style vs Content Methods Verification Systems Vector space models Distance metrics Benchmark results PAN data Latin data Caesar's writings Acknowledgements References 
work_7qkqtvrpezcrtoaq3oc7zpsvg4	----	Storage-Optimizing Clustering Algorithms for High-Dimensional Tick Data Krisztian Buzaa, Gábor I. Nagya, Alexandros Nanopoulosb aFaculty of Electrical Engineering and Informatics Budapest University of Technology and Economics, Hungary buza@cs.bme.hu, nagy.gabor.i@gmail.com bUniversity of Eichstätt-Ingolstadt, Germany alexandros.nanopoulos@ku.de Abstract Tick data are used in several applications that need to keep track of values changing over time, like prices on the stock market or meteorological measurements. Due to the possibly very frequent changes, the size of tick data tends to increase rapidly. Therefore, it becomes of paramount im- portance to reduce the storage space of tick data while, at the same time, allowing queries to be executed efficiently. In this paper, we propose an approach to decompose the original tick data matrix by clustering their attributes using a new clustering algorithm called Storage-Optimizing Hierarchical Agglomerative Clustering (SOHAC). We additionally propose a method for speeding up SOHAC based on a new lower bounding technique that allows SOHAC to be applied to high- dimensional tick data. Our experimental evaluation shows that the proposed approach compares favorably to several baselines in terms of compression. Additionally, it can lead to significant speedup in terms of running time. Keywords: Tick data, Clustering, Storage 1. Introduction Representing information about complex objects or phenomena requires a large set of attributes (a.k.a. features). The values of these attributes may change rapidly over time; for instance, prices on the stock market or meteorological measurements. Data with such characteristics, i.e., high complexity/variety, large volume and change velocity, are nowadays being addressed within the paradigm of big data. Beyond the challenges pertaining to the management of such data, analytical tasks pose the crucial requirement of investigating their dynamics, i.e., how their attributes change values over time. What is, therefore, necessary is to keep track of the changes of values, which results in very large collections of rapidly changing data. This particular data type is being referred to as tick data [13], which is the term we adopt in our study, too. Preprint submitted to Elsevier December 23, 2013 One of the most important examples for tick data is the record of stock market transactions. This type of data can be considered as a matrix: the columns of the matrix correspond different properties of a transaction, such as price, volume of the trade, the symbol of an asset, etc. Every time a transaction is executed or a quote is given for a stock, a row is appended to this matrix. The speed of a transaction in liquid financial markets can be measured in the order of fractions of milliseconds therefore, this matrix grows extremely rapidly. It is necessary to record all transactions but this magnitude of data flow requires proprietary technology allowing efficient storage and quick retrieval of data. Solutions are often built over column based database technology such as a KDB database.1 However, as we will describe in more detail, a straightforward application of such technology leads to suboptimal storage of tick data in the sense of occupied storage space on disk or memory. The primary reason for the sub-optimality of the aforementioned storage is the redundancy of conventional techniques: in case of a straightforward solution, one would use orders of magnitudes more storage space (either disk space or main memory) than required, therefore, storage, access, search and analysis of the data becomes computationally more expensive than necessary. One way to alleviate this problem is the usage of regular data compression methods (see e.g. [16] for an excellent overview). This approach is well-suited for cheap storage of large, historical archives of the data. However, it does not support quick access to the data: if the data is compressed, in order to be able to execute an analytic or search query, a large archive (or at least some parts of that) must be decompressed which might be computationally expensive and therefore the procedure could become inefficient. This problem becomes crucial if many queries have to be executed which is usually the case in real-life applications, such as stock market trading. In this paper, we aim at developing a storage structure for tick data that reduces the storage space required by the straightforward approach with the ability to execute search and analytic queries efficiently in the same manor provided by the straightforward approach. In particular, our approach is based on the decomposition of a large tick data matrix into a number of smaller matrices. We achieve this decomposition by clustering the columns of the matrix. Each resulting smaller matrix contains a subset of the original attributes and the corresponding values of all rows of the original matrix for these attributes. Although, conventional clustering algorithms achieve significant improvements, motivated by hierarchical clustering algorithms, we develop a 1http://kx.com/ 2 new clustering algorithm that minimize storage space required for a tick data matrix. We call our approach SOHAC, Storage-Optimizing Hierarchical Agglomerative Clustering. In this paper, we focus on speeding up SOHAC and propose a new lower bounding technique that allows SOHAC to be applied to high dimensional tick data; we refer to the resulting method as QuickSOHAC. According to our experiments, the proposed lower bounding technique combined with sampling leads to substantial speedup without significant loss of quality. We evaluate the proposed approach on three real-world tick data tables provided by an investment bank. Further- more, while the primary motivation of our work is to develop a storage schema that allows efficient storage and quick access to the data, we show that conventional compression algorithms may also substantially benefit from the proposed approach. The remainder of this paper is organized as follows. Section 2 provides an overview of related works. In Section 3 we present the SOHAC algorithm. In Section 4 we focus on efficient implemen- tation of SOHAC and propose a lower bounding technique that allows to speed up the algorithm. We devote Section 5 to the experimental evaluation of our approach. Finally, we conclude in Section 6. 2. Related Work The data describing transactions on financial markets, especially tick data (also known as tick- by-tick data) allows thorough analysis of the markets and their dynamics. Recent works focused on statistical properties [14] and analysis [17], [18] of currency exchange rates and stock market tick data [13]. Dionne focused on risk analysis [8], while the dynamics of stock markets were studied in [4], [15] and [21]. Based on tick data, Akram et al. empirically studied the law of one price on different financial markets [2], while Cartea proposed to model stock price tick-by-tick data via a non-explosive marked point process [7]. Blais and Protter studied various algorithms for the recognition of whether a transaction in a tick-data set was initiated by the buyer or seller [5]. Barany et al. [3] aimed to detect market crashes based on high-frequency data. Yabuuchi and Watada studied fuzzy autocorrelation models for forecasting problems related to tick data [23]. Although, storage of tick data is a core component of the systems performing the above analytic tasks, none of the above works focused on how to develop storage structures for high dimensional tick data. More closely related to our work is that of Ahmad et al. who focused on the summarization of tick data time series [1]. However, instead of storing a summarized approximation of the time series, we aim at storing the entire original tick data. As we will demonstrate it in Section 3.1, the 3 storage of tick data is a non-trivial task: conventional techniques result in redundant and therefore suboptimal solutions. Conventional data compression techniques, see e.g. [16], can efficiently compress the data, and therefore they are well-suited for data archives. For tick data, however, we need storage structures that allow efficient storage and quick access to the data simultaneously while they can be realized in database systems. As we will demonstrate, large tick data tables can usually be decomposed into several smaller ones, the total size of which being substantially less than that of the original table. Simultaneously, this decomposition allows for quick querying of the data as we will discuss in Section 3.5. Clustering is well-known for its ability to summarize data by grouping similar objects together which allows the user to consider the data at a higher level of abstraction [19]. Therefore, our approach for the decomposition of tick data matrices is based on clustering. In the last decades, very large number of clustering algorithms were developed for various tasks (see e.g. [6], [9], [10] and [12]). We refer to [19] and [22] for excellent surveys of clustering algorithms. We will demonstrate that one can achieve substantial improvements if one uses general-purpose clustering algorithms for the decomposition of tick data matrices. However, such conventional clustering algorithms were originally not designed for storage optimization of tick data and therefore they lead to suboptimal decompositions in most cases. In contrast, we propose a clustering algorithm, SOHAC, which directly minimize the required storage space and therefore, as shown in large number of extensive experiments, SOHAC substan- tially outperforms conventional clustering algorithms for the tick data storage problem. An initial, but promising version of our algorithm was presented in [11]. However, due to its computational complexity, the application of SOHAC to high-dimensional tick data still remained a challenge. This is essential, as most of the tick data tables in real-world applications are high dimensional, i.e., contain many columns. Therefore, after reviewing SOHAC, in this article, we focus on speeding up the algorithm. The techniques we propose allow SOHAC to find the decomposition quickly, even in case of high-dimensional tick data tables. 3. Decomposition of Tick Data Matrices based on Clustering In this section, we explain how the decomposition of tick data matrices may lead to more efficient storage. First, we motivate our approach with an illustrative example, then we review our clustering algorithm, SOHAC, that supports efficient storage of tick data. 4 Figure 1: An illustrative example for tick data. Features describing the weather are monitored continuously. Whenever the value of one of the features changes, a new row is inserted into the recordings (see the table in the top). Decomposition of such tables by features (columns) that change their values simultaneously may substantially reduce the required storage space (see the tables in the bottom of the figure). 3.1. An Illustrative Example Suppose that a weather station monitors features of weather conditions. In this example, such features are the temperature, humidity and pressure of the air, the velocity and direction of the wind, the intensity of the radiation of the sun and the overall outlook (such as sunny, cloudy, raining or snowing). These features are monitored continuously over the time. Whenever the value of one of these features changes, a new raw is inserted into the recordings. This new row contains the values of the features as well as a time-stamp indicating when the observations were made. See the matrix in the top of Figure 1. This representation, called tick data, is well-suited for queries: for example, if we are interested for the features of the weather at 10:30 o’clock, we only need to find the raw corresponding the most recent observation before 10:30, i.e., we have to consider the raw at 10:22. This raw describes the ”state of the world”, i.e., it contains the values of all the features that are relevant in the current application. Such queries regarding the ”state of the world” at a given time can be effectively supported by indexing techniques. The only disadvantage of the representation shown in the top of Figure 1 is that the total size of the matrix may become much larger than actually required. In order to illustrate this we stored 5 the same information in two smaller matrices in the bottom of Figure 1. In our approach such decompositions are based on the clustering of columns: in the example, we consider two clusters of columns. One of the clusters contains Humidity and Pressure, while the other cluster contains the other columns, i.e., Temperature, Velocity of the wind, Direction of the wind, Radiation and Outlook. As shown in the example, due to the decomposition, we can save storage space: the total number of cells required to store the data was reduced from 7×7 = 49 to 3×2 + 5×5 = 31 (without counting the cells in the column Time which acts like an index column). This corresponds to a compression ratio of 31/49 ≈ 0, 633. While the decomposition reduces the required storage space, in the worst case, the computa- tional complexity of a query may increase moderately: if we are interested for all the features describing the weather at 10:30, we have to execute two queries instead of one, however, both queries are executed on much smaller datasets (and therefore the overall execution time is ex- pected to grow only moderately compared to the previous case). Whereas if we are only interested for the temperature and radiation we have to execute just one query on a dataset of reduced size (and therefore the overall execution time is expected to be reduced). The example in Figure 1 illustrates the decomposition of a tick data matrix in an intuitive way. Next, we systematically study such decompositions and develop an algorithm that aims at minimizing the storage space required after the decomposition. 3.2. Definitions and Problem Formulation In general, a tick data matrix M is a matrix where columns correspond attributes or features while rows correspond observations of the same features at different moments of time. Rows of the matrix are ordered according to the order of observations, i.e., the values of the i-th row were observed before the values of the j-th row if and only if i < j. While the observations are made, a new row is added whenever the value of an attribute changes. However, as long as none of the attribute-values changes no new row is added to the matrix, therefore two rows of a tick data matrix differ in the value of at least one attribute. There is an additional column that is used to index the rows of a tick data matrix. This additional index column may contain, for example, ascending integer numbers (like the number of the corresponding row) or a time-stamp (see the Time column in the example in Section 3.1). We use the term regular column for all the columns other than the index column. With decomposition of a tick data matrix M we mean the clustering of the regular columns of 6 M into k disjoint clusters Pi, 1 ≤ i ≤ k, i.e., for each regular column cj of M: cj ∈ P1 ∨ cj ∈ P2 ∨ ...∨ cj ∈ Pk and for all i, j with i 6= j Pi ∩Pj = ∅. Note that this clustering refers to the regular columns only, i.e., in this formulation, the index column does not belong to any cluster. Then, for each cluster Pi, a matrix Mi is derived from M by selecting the index column and those columns of M that belong to cluster Pi. Subsequent rows of a derived matrix Mi may contain the same values in all the regular columns. In such cases we only keep the first row. For example, in Figure 1, P1 = {Humidity, Pressure}, P2 = {Temperature, Wind (velocity), Wind (direction), Radiation, Outlook} and the corresponding matrices M1 and M2 are shown in the bottom left and bottom right of the Figure 1. We can easily see that the original matrix can be reconstructed from the decomposition described above, and therefore, instead of the original matrix M, one can use this decomposition to calculate the results of search and analytic queries. In this paper, we target the problem of finding a decomposition so that the required storage space is minimized. In particular, for a given number of clusters k, we aim at finding a decomposition so that the total number of the cells in all the matrices Mi (without counting the cells in the index column) is minimized. Our approach can simply be adapted for the case of more advanced storage models, where we do not assume uniform storage cost for each cells and/or the storage costs of the index cell is also taken into account. We note that k is usually relatively small: for example, for the storage of tick data of financial transactions, the user is most interested for the decomposition into k = 2 or k = 3 clusters (see also Section 3.5). 3.3. Clustering of Columns of Tick Data Matrix In the literature, there are many clustering algorithms that are able to produce non-overlapping clusters in a way that these clusters together cover all the instances. Therefore, one solution for the problem defined in the previous section is to cluster the columns of a tick data matrix. 7 Figure 2: Construction of a binary change indicator matrix from a tick data matrix. The tick data matrix is shown in the top of the figure, while the corresponding indicator matrix is shown in the bottom. The index column is the Time column in this example. In the context of our problem, two regular columns are considered to be similar, if they often change values togehter (i.e., in the same rows). Therefore, we define a binary change indicator matrix I over a tick data matrix M. Except the entries of the index column, all the entries of the binary change indicator matrix I are either 0 or 1 depending on whether or not the value of a cell in the tick data matrix M is equal to the value of the cell in the same column and the previous row of M: I(i, j) =   M(i, j) if the j-th column is the index column in M 0 if i > 1 and M(i, j) = M(i− 1, j) 1 otherwise where M(i, j) and I(i, j) denote the entries in the i-th row and j-th column of the tick data matrix M and binary change indicator matrix I respectively. As an example, Figure 2 shows how the binary change indicator matrix is derived from a tick data matrix. The index column is the Time column in this example. After constructing the binary change indicator matrix I, we can use its regular columns (i.e., all the columns except the index column) as instances in almost any clustering algorithms. Despite the fact that conventional clustering algorithms are not designed to produce optimal clusters in terms of our problem defined in Section 3.2, as we have shown in [11], if we use the binary change indicator matrix representation and a conventional clustering algorithm to find a decomposition of the tick 8 Algorithm 1 SOHAC: Storage-Optimizing Hierarchical Agglomerative Clustering for Tick Data Require: Tick data matrix M, number of clusters k Ensure: Clustering of the columns of M 1: Construct the binary change indicator matrix I from M 2: P = {{c1},{c2}, ...,{cn}} (Initially, each column cj of M is a separate cluster) 3: while |P | > k do 4: s ←∞ (Storage size for the best clustering found so far) 5: for all pairs of clusters (Ci,Cj), with Ci ∈ P , Cj ∈ P do 6: C′i ←Ci ∪Cj (Merge clusters Ci and Cj into the new cluster C ′ i ) 7: P ′ ← P \{Ci,Cj}∪{C′i} 8: s′ = storage size required to store the decomposition corresponding to P ′ (This can be computed based on I.) 9: if s′ < s then 10: P∗ ← P ′ (The best clustering found so far) 11: s ← s′ 12: end if 13: P ← P∗ 14: end for 15: end while 16: return P data matrix we can already achieve notable improvements in terms of storage space compared to the case of storing the original tick data matrix. In the next section, we review SOHAC, our clustering algorithm that directly optimize the storage space and therefore substantially outperforms conventional clustering algorithms for our problem. 3.4. SOHAC: Storage-Optimizing Hierarchical Agglomerative Clustering SOHAC, Storage-Optimizing Hierarchical Agglomerative Clustering is designed for clustering columns of a tick data matrix. The algorithm builds on the hierarchical agglomerative strategy. Therefore, initially, all the objects belong to separate clusters. Then, clusters are iteratively merged together as long as the current number of clusters is more than k, the user-defined number of clusters. Therefore, at the end of this iterative process, k clusters are produced. 9 The key feature of our algorithm is that in each iteration it merges those two clusters that lead to minimal storage size of the decomposed matrix. This storage size can simply be calculated based on the binary change indicator matrix. For each examined clustering of the columns, we decompose the binary change indicator matrix. Then, we consider the rows that contain only zeros in the regular columns. The cells of such rows can be eliminated in the examined decomposition without loss of information. Therefore, in order to determine the number of cells required for the storage of the examined decomposition, we only need to count the cells in the rows that contain only zeros in their regular columns. The pseudocode of our algorithm is shown in Algorithm 1. 3.5. Querying Decomposed Tick-Data Matrix Two common type of queries over a decomposed tick-data matrix are the point queries and the range queries. A point query retrieves the value of one or more features (columns) at a specific time point. A range query retrieves the values of one or more features within a given time range. In this section, we consider the cost to perform point and range queries, and how it is related to the number of clusters k. Let Csp denote the corresponding cost to perform a point query that retrieves the value of a single feature at a specific time point. Also let m denote the number of rows in the original (uncompressed) matrix and rs the compression ratio of the cluster containing the feature of interest. The compression ratio rs is in the range [0, 1], i.e., after decomposition, the number of rows in the cluster is rsm and thus smaller values of rs indicate better compression. It is easy to see that Csp ∝ log(rsm), because the point query can retrieve the requested value by performing binary search in the cluster that contains the single feature of the query.2 Each cluster already contains values sorted in time, thus the cost is logarithmic to rsm, i.e., the number of rows in the cluster of interest. An increase of the number of clusters k will tend to improve the compression, thus will tend to reduce rs. In this case, the cost C s p tends to reduce with increasing k. For a point query that retrieves the values of all features at a specific time point, let Cap the corresponding cost to perform this query and ra the compression ratio of the tick data table, i.e., the number of rows in a cluster is ram on average (i.e., ra is the average compression ratio of all compression ratios rs for each individual column). In this case, C a p ∝ k log(ram), because the query can retrieve the requested values for all features by performing binary search in all clusters 2We do not consider that values within clusters are indexed, because the updating of such indices would incur prohibitive overhead costs that are not feasible for rapidly changing tick data. 10 that contain the values of all features for the given point in time. As before, the query performs binary search in each cluster and there are k clusters in total. Therefore, an increase in the number of clusters k introduces a trade-off for Cap : larger k values cause a reduction of ra and, thus, a reduction of Cap by a logarithmic factor (due to the log(ram) factor above); on the other hand, however, larger k values increase Cap by a linear factor (due to the k factor above). Based on our empirical results in Section 5.1 about the range of resulting ra values w.r.t. to increasing k values, it can be derived that the reduction in Cap by a logarithmic factor due to smaller ra is substantially outweighed by the increase by a linear factor due to larger k. Similar conclusions can be made for range queries. Let Car denote the cost of performing a range query that retrieves the values of all features within a range in time. A range query can be performed by performing binary searching in each cluster, in order to identify the starting point (as done in the case of a point query) and then by performing sequential retrieval of all subsequent values within the query range. If w is the average number of values retrieved during such a sequential searching in each cluster, then Car ∝ k log(ram) + kw. This results in a similar trade-off as before. Moreover, for range queries with larger time ranges, w will tend to be large, which makes the increase in Car to become larger by an additional factor linear to k due to the second term above (i.e., kw). Therefore, in real-world applications, where very large matrices need to be decomposed, the number of clusters k should not get high, because this will incur prohibitive query costs in case of range and point queries that retrieve values for more features. As will be described during the presentation of our experimental results, reasonable k values are usually in the range 2 − 5. 4. Speeding up SOHAC Algorithm 1 shows SOHAC as we presented it in our previous work [11]. Next, we describe techniques that speed up SOHAC and allow its application to high-dimensional tick data. 4.1. Avoiding the recalculation of storage sizes Note that, in each iteration, most of the clusters remain unchanged, except the ones that are merged in that iteration. Furthermore, merging two clusters only changes the storage of the columns belonging to the merged clusters. As most of the clusters remain unchanged in each iteration of SOHAC, at the end of an iteration, we do not need to recalculate the storage size of a cluster pair (Ci,Cj) if both Ci and Cj remained unchanged in that iteration. 11 4.2. Lower Bounding Storage Size In each iteration of SOHAC, those two clusters are merged that lead to minimal storage size. However, as we will show, in order to find out which pair of clusters lead to minimal storage size, it is not necessary to calculate the exact storage size of all pairs of clusters. More importantly, for many of those pairs that are affected by the merging steps at the end of the iteration (see line 13 in Algorithm 1), i.e., pairs for which the storage size changes, we can avoid the calculation of the exact storage size via lower bounding as described below. In particular, it is unnecessary to calculate the exact storage size resulting from merging the clusters Ci and Cj if we know that merging Ci and Cj results in a storage size of s0 at least and there is an other pair of clusters C′i and C ′ j so that merging C ′ i and C ′ j results in storage size of exactly s and s0 > s. This is shown in Algorithm 2. Next, we develop a computationally cheap lower bound of the storage size which allows to speed up the SOHAC algorithm. Suppose we are given two clusters C1 and C2 containing columns c1, . . . , cn and cn+1, . . . cn+m respectively. Assuming uniform storage costs for all the cells of all the columns, the storage size resulting from merging C1 and C2 can not be less than 12 (n + m)-times the storage size of any pair (ci, cj), 1 ≤ i ≤ n, n + 1 ≤ j ≤ m. With storage size of a pair (ci, cj) we mean the storage size associated with merging the two clusters that contain the single columns ci and cj respectively. According to the above observation, a lower bound of the storage size is StorageSize(C1,C2) ≥ max ci∈C1,cj∈C2 1 2 · (n + m) · StorageSize(ci, cj). (1) The calculation of this lower bound could still be computationally expensive when merging large column clusters C1 and C2 because according to (1) we should take into account all the pairs (c1, c2), c1 ∈C1, c2 ∈C2. In order to avoid it, in our current implementation, for each cluster we maintain a variable that always contains the identifier of the most frequently changing column belonging to that cluster and we calculate the lower bound according to those columns: StorageSize(C1,C2) ≥ 1 2 · (n + m) · StorageSize(mfcc(C1), mfcc(C2)), (2) where mfcc(C) denotes the most frequently changing column of cluster C. This lower bound is illustrated in the example shown in Figure 3. We want to calculate the lower bound for merging the column clusters { Wind (direction), Humidity } and { Pressure, Outlook }. For the cluster { Wind (direction), Humidity }, both columns change twice, therefore, we can treat any of them as the most frequently changing column. For this example we will consider Humidity 12 Algorithm 2 QuickSOHAC: SOHAC with Lower Bounding of Storage Sizes Require: Tick data matrix M, number of clusters k Ensure: Clustering of the columns of M 1: Construct the binary indicator matrix I from M 2: P = { {c1},{c2}, ...,{cn} } 3: Initialize all the values s[i][j] of a two-dimensional array as unknown, s[i][j] is the storage size resulting from merging clusters Ci and Cj. 4: while |P | > k do 5: s ←∞ (Storage size for the best clustering found so far) 6: for all pairs of clusters (Ci,Cj), with Ci ∈ P , Cj ∈ P do 7: C′i ←Ci ∪Cj (Merge clusters Ci and Cj into the new cluster C ′ i ) 8: P ′ ← P \{Ci,Cj}∪{C′i} 9: if s[i][j] is unknown then 10: s0 ← lower bound of the storage size resulting from merging clusters Ci and Cj 11: if s0 < s then 12: calculate the exact storage size resulting from merging clusters Ci and Cj and set the value of s[i][j] to the calculated storage size 13: if s[i][j] < s then 14: P∗ ← P ′, s ← s[i][j] , i∗ ← i , j∗ ← j (The best clustering found so far) 15: end if 16: end if 17: else 18: if s[i][j] < s then 19: P∗ ← P ′, s ← s[i][j] , i∗ ← i , j∗ ← j (The best clustering found so far) 20: end if 21: end if 22: P ← P∗ 23: Let the merged cluster be the new i∗-th cluster, while the j∗-th cluster is not used anymore: Cnewi∗ = (Ci∗ ∪Cj∗ ). Set the storage sizes to unknown for all the pairs that include the new cluster: s[i∗][.] ← unknown, s[.][i∗] ← unknown 24: end for 25: end while 26: return P 13 Figure 3: Example illustrating lower bounding. According to the matrix shown in the top, the storage size for the columns Humidity and Pressure is 3 · 2 = 6 (the number of cells that need to be stored when merging the clusters containing these single columns). According to the matrix in the bottom, the actual storage size resulting from merging the clusters { Wind (direction), Humidity } and { Pressure, Outlook } is 6 · 4 = 24 (without the index column). This actual storage size is more than the resulting lower bound which is 1 2 · (2 + 2) · 6 = 12. as the most frequently changing column of the cluster { Wind (direction), Humidity }. Regarding the other cluster, Pressure changes twice, while Outlook changes only once, therefore, Pressure is the most frequently changing column of cluster { Pressure, Outlook }. In the top of the Figure 3, we see the storage size resulting from merging the columns Humidity and Pressure which is 3·2 = 6, i.e., when merging the clusters that contain only the single columns Humidity and Pressure 6 cells of the original tick data matrix need to be stored. For simplicity, the storage cost of the index column (Time) is ignored throughout the example. According to (2), when merging the clusters { Wind (direction), Humidity } and { Pressure, Outlook }, the resulting storage size is at least 1 2 · (2 + 2) · 6 = 12. The actual storage size is 24, as shown in the Figure 3. Note that in the first iteration of QuickSOHAC, i.e. when each cluster consists of a single column, the lower bounds are exactly the same as the storage sizes. We take this into account in our implementation: in the first iteration, we always set the value of s[i][j] in line 12 of Algorithm 2 independently of the outcome of the previous conditional statement. 14 Table 1: The tick data tables used in our experiments, together with the number of their rows and columns (without the index column). Data table Rows Columns TD-1 408 042 30 TD-2 554 854 190 TD-3 50 000 000 33 4.3. Sampling In order to further speed-up the approach, we can run SOHAC on a small sample, such as 5% or 10% of the entire tick data matrix. While this results in substantial speed-up, according to our observations presented in Section 5, the resulting compression ratios are only slightly worse compared to the case of using the entire tick data table. 5. Experiments We evaluated our algorithm on three real-world tick data tables provided by the world’s leading investment bank. The basic properties of these tables, number of columns and rows are shown in Table 1. First, we compare the decomposition produced by our approach and other clustering algorithms in terms of compression ratio. Subsequently, we focus on the runtime of SOHAC and QuickSOHAC, i.e., to which extent the proposed lower bounding technique allows to speed up the algorithm. We also present results for the case when we sampled the data. Finally, we show that our approach can be combined with conventional compression methods: in particular, we show that compression of the tables with gzip3 may also benefit from the decomposition produced by our approach. 5.1. Compression Ratio In the first experiment, we compared the decomposition of a tick data matrix resulting from the clusters produced by our approach, SOHAC, to the decomposition of the same tick data matrix using existing and widely used clustering algorithms. We measured the quality of decompositions in terms of compression ratio (CR), i.e., the ratio of the number of cells in regular columns after 3http://www.gzip.org/ 15 the decomposition and the number of cells in regular columns in the original matrix: CR = number of cells in regular columns after decomposition number of cells in regular columns in the original matrix An example for the calculation of compression ratio can be found in Section 3.1. As our approach, SOHAC, is built on hierarchical agglomerative clustering, in our experiments we focused on comparing the clustering produced by SOHAC to the clustering produced by dif- ferent variants of hierarchical agglomerative clustering algorithms. Additionally, we compared the clustering produced by SOHAC to the clustering produced by k-Means [20]. We tested two variants of k-Means, the first one used Euclidian distance, while the second one used Manhattan distance. Regarding the variants of hierarchical agglomerative clustering algorithms, we tested Single Linkage, Complete Linkage and Average Linkage with the following proximity measures: Euclidean Distance, Cosine Similarity, Chebychev Distance and Manhattan Distance. We used the Weka- implementations of all of the above baseline algorithms [20]. In total, taking all the examined variants of the baselines into account, we compared our approach to 14 clustering algorithms from the literature. To ensure fair comparison, all of the baselines, as well as our approach, used the binary change indicator matrices as input. First, we decomposed the entire TD-2 tick data table into k = 2, 3, 4 and 5 clustering using SOHAC and the baselines. Table 2 shows the resulting compression ratios. As we can see, SOHAC clearly outperforms the baselines for all the examined cases. Subsequently, in order to be able to study whether the improvements are systematic and sta- tistically significant, we split the entire tick data matrix into 10 disjoint sub-matrices, and we repeated all the experiments 10 times. In each of the 10 rounds of the process, we used a differ- ent sub-matrix, and clustered the columns of that sub-matrix. Therefore, we could calculate the average and standard deviation of the compression ratio. In the left of Figure 4, we show the compression ratio and its standard deviation resulting from the decomposition into k = 2 clusters using our approach, SOHAC, and the baselines. In the figures, SingleLink, CompleteLink and AverageLink are denoted by H-SL, H-CL and H-AL respectively. For SingleLink, CompleteLink, AverageLink and k-Means, we only show the best-performing variants. In the right of Figure 4, we show the compression ratios as function of the number of clusters. SOHAC achieves the best compression ratio in all cases. Especially for TD-3, i.e., the largest among the three data sets, the difference is substantial, which indicates the suitability of SOHAC for tick data of very large size. With increasing k values, as expected, all methods converge to 16 Table 2: Compression ratios on the entire TD-2 tick data table for the decompositions using our approach, SOHAC, and the baselines. Approach Distance measure k = 2 k = 3 k = 4 k = 5 SingleLink Euclidean 0.999 0.842 0.694 0.692 Manhattan 0.999 0.842 0.694 0.692 Cosine 0.947 0.938 0.313 0.245 Chebyshev 0.590 0.580 0.574 0.522 CompleteLink Euclidean 0.999 0.693 0.532 0.528 Manhattan 0.999 0.693 0.532 0.528 Cosine 0.947 0.269 0.247 0.167 Chebyshev 0.590 0.580 0.575 0.522 AverageLink Euclidean 0.999 0.842 0.694 0.528 Manhattan 0.999 0.842 0.694 0.528 Cosine 0.947 0.323 0.311 0.189 Chebyshev 0.590 0.580 0.574 0.522 k-Means Euclidean 0.709 0.269 0.232 0.173 Manhattan 0.999 0.989 0.884 0.474 SOHAC 0.276 0.141 0.122 0.101 comparable compression ratios. Nevertheless, based on the discussion in Section 3.5, larger values of k hinder the efficient execution of queries over the compressed tick data. For this reason, in real-world application the focus is only on small values of k, usually equal to 2 or not much higher (i.e., in the range 2−5). It is worth to notice that in these cases SOHAC presents a clear advantage compared to all examined baselines. 5.2. Runtime In order to evaluate the proposed lower bounding technique, we compared the runtimes of two variants of our approach. By SOHAC we denote the variant that is implemented according to Algo- rithm 1, but avoids the unnecessary recalculation of storage sizes (see Section 4.1). QuickSOHAC additionally uses the proposed lower bounding technique (see Algorithm 2). The runtimes relative to the runtime of QuickSOHAC are shown in Figure 5. We can see that, in terms of runtime, SOHAC is comparable to the baselines even without the proposed lower bounding. The proposed lower bounding leads to 4-6 fold speedup, and therefore QuickSOHAC is faster than the baselines in the vast majority of the examined cases. This means that QuickSOHAC is overall the best performing method, because not only it outperforms all baselines in terms of compression ratios, but it also compares favorable in terms of running time, which makes it suitable for large scale 17 Figure 4: The compression ratio and its standard deviation resulting from the decomposition into k = 2 clusters using our approach, SOHAC, and the baselines (in the left). The compression ratio as function of the number of clusters using our approach, SOHAC, and the baselines (in the right). Note that the curves H-SL, H-CL and H-AL overlap for the TD-2 dataset. 18 applications. Figure 5: Relative runtimes compared to the runtime of QuickSOHAC. For the baselines, we only show the runtime of the best performing variant w.r.t. compression ratio. Furthermore, we examined the effect of sampling: in our experiments, we sampled the first 5% and 10% of the tick data matrices, while the effect of sampling in terms of compression ratio is shown in Figure 6. As shown, using a relatively small portion of the entire tick data table, such as 5% or 10%, allows for additional speedup, while maintaining almost as good compression ratios as if we would use the entire tick data table. 5.3. Combination with conventional compression methods In order to measure physical storage sizes, i.e., the amount of disk space occupied by the tables, we stored both the original binary change indicator matrices and the decomposed ones both in 19 Figure 6: Compression ratio of SOHAC for the cases of using the entire tick data table, sampling 10% and 5% of the table. 20 simple text files and gzip-compressed text files. The results are shown in Table 3. As we can see, SOHAC outperforms the baselines in both contexts, i.e., storage as simple text file and gzip- compressed storage. Furthermore, we emphasize that SOHAC with gzip leads to substantially reduced physical storage size compared to the case of compressing the entire table with gzip. This indicates that SOHAC may be able to decompose the table along such patterns which could not be identified by gzip. Table 3: The physical storage sizes (in kBytes) of the binary change indicator matrices and their standard deviation- with and without the usage of gzip. Data table Approach with gzip without gzip TD-1 no decomposition 367 ± 20 4849 ± 32 Single Link 349 ± 21 3344 ± 451 Complete Link 466 ± 32 4743 ± 457 Average Link 372 ± 37 4210 ± 684 k-Means 367 ± 40 3416 ± 427 SOHAC 351 ± 23 3221 ± 78 TD-2 no decomposition 691 ± 5 23892 ± 10 Single Link 583 ± 5 14155 ± 2106 Complete Link 583 ± 5 14155 ± 2106 Average Link 583 ± 5 14155 ± 2106 k-Means 812 ± 61 23330 ± 3564 SOHAC 568 ± 25 9617 ± 1464 TD-3 no demcomposition 4326 ± 75 61436 ± 277 Single Link 4288 ± 82 54795 ± 340 Complete Link 4288 ± 82 54795 ± 340 Average Link 4288 ± 82 54795 ± 340 k-Means 5545 ± 366 61149 ± 6048 SOHAC 4045 ± 97 37610 ± 837 5.4. Discussion As the experiments show, the proposed algorithm, SOHAC, performs favorably to conventional clustering algorithms on the task it was designed for, i.e., decomposition of tick data matrices. The proposed lower bounding allows to speed up the algorithm by a factor of 4-6. Due to the ever-growing datasets, and the rate of growth in particular, this can be considered relevant even in the light of Moore’s law which states that the computational performance doubles every 18 months. Finding the decomposition quickly may also be relevant in real-time applications. With 21 the significant speed up a lower utilization of current computing resources could be achieved. Furthermore, our experiments illustrate that our approach can be successfully used together with conventional compression algorithms, which is relevant in applications where the final compression rate is more important than the execution time of queries. Regarding the limitations of the proposed approach, we point out that SOHAC is clearly not a general-purpose clustering algorithm, but it was designed for a particular task. While SOHAC (or a modified version of it) might be applied to cluster non-tick data, due to its design, we only expect SOHAC to perform well in cases when the data has at least some relevant aspects in common with tick data, while the adaptation of SOHAC to such new data types is not trivial. 6. Conclusion In this paper, we focused on the storage of tick data, a type of data appearing in various appli- cations from finance to meteorological observations. Due to the rapid change of the values of the observed features the size of tick data tables tends to increase substantially. We proposed a storage scheme for tick data that reduces the storage space while, at the same time, it allows to efficiently execute queries. The proposed approach performs a decomposition of the original tick data matrix into a number of smaller matrices. We achieve this decomposition by clustering the columns of original matrix based on a new clustering algorithm called Storage-Optimizing Hierarchical Ag- glomerative Clustering (SOHAC). In this paper, we focused on the efficient implementation of SOHAC. Our observation about the lower bound of storage sizes allowed to speed up SOHAC. This is essential for the application of SOHAC to high-dimensional tick data. Our experimental evaluation demonstrated that the proposed approach compares favorably to several baselines in terms of compression, whereas the lower bounding technique can lead to substantial speedup. The benefits due to the proposed approach are important, due to the several application domains that are based on the ability to efficiently store and analyse tick data. For instance, major financial institutions record historical stock-market values in the form of tick data, which can be even publicly available and help analysts working in financial services in developing risk management, trading, and quantitative analysis.4 Recording of such historical data is performed for many years and, thus, is of massive volumes, since in order “to store decades of market data, the volume of disk 4See: http://www.xignite.com/market-data/nasdaq-historical-stock-tick-data/ 22 space needed expands geometrically”5. Since uncompressed tick data is voluminous, in order to perform useful analysis the proposed compression scheme offers the ability to reduce their storage requirements and to enable their faster processing based on in-memory architectures. Moreover, an appealing option nowadays for storing large volumes of tick data is to follow a cloud-based solution. Storage as a service (STaaS) is increasingly being used by many organizations that manage such large data collections. STaaS allows such organizations to reduce the total cost of ownership by renting storage space from a service provider. However, the costs involved with STaaS usually increase proportionally to the size of the data. Moreover, tick-data are of very large sizes and they grow rapidly. Thus, the compression attained by the proposed scheme can attain significant cost reductions by optimizing the storage of such large and rapidly-changing tick data. The proposed approach is based on the premise that the “pattern of change” in the tick data remains stable, which allows SOHAC to detect clusters over the entire tick-data matrix. In the long term, however, this “pattern of change” may vary and may require to update the clustering scheme of SOHAC. For this reason, in our future work we plan to investigate extensions of SOHAC that will split the original matrix into clusters, each of which can be effectively represented by a single clustering scheme. The detection of such clusters in a dynamic way, i.e., as new data are arriving, is also an interesting point of future work. Instead of storing the exact signal, in many applications, an approximation of the signal may be sufficient. Therefore, we aim to study SOHAC in context of approximate storage of multivariate signals. Furthermore, we want to explore whether SOHAC can be adapted for the clustering of (multivariate) time-series. The key of this application is a technique that converts time-series into a binary vector reflecting at which positions substantial changes happen. While a naive approach using a fixed threshold may not work well enough, approaches based on machine learning, such as segmentation of the time-series by an appropriate Hidden Markov Model could potentially provide such a representation. We hope to explore the possibilities of the usage of QuickSOHAC clustering algorithm in more general datasets, that are binary in nature, for example document-term matrices in text mining and extend the algorithm to bi-cluster documents and frequently appearing terms in documents. We aim at comparing these results to existing bi-clustering algorithms. 5See: http://marketsmedia.com/tick-tick/ 23 Acknowledgment We thank DAAD and MÖB for financially supporting a researcher exchange program with project number: 39859. References [1] S. Ahmad, T. Taskaya-Temizel, K. Ahmad, Summarizing time series: Learning patterns in volatileseries, Intelligent Data Engineering and Automated Learning–IDEAL 2004 (2004) 523– 532. [2] Q. Akram, D. Rime, L. Sarno, Does the law of one price hold in international financial markets? evidence from tick data, Journal of Banking & Finance 33 (2009) 1741–1754. [3] E. Barany, M.B. Varela, I. Florescu, I. Sengupta, Detecting market crashes by analysing long- memory effects using high-frequency data, Quantitative Finance 12 (2012) 623–634. [4] R. Bartiromo, Dynamics of stock prices, Physical Review E 69 (2004) 067108. [5] M. Blais, P. Protter, Signing trades and an evaluation of the leeready algorithm, Annals of Finance 8 (2012) 1–13. [6] K. Buza, A. Buza, P. Kis, A distributed genetic algorithm for graph-based clustering, Man- Machine Interactions 2 (2011) 323–331. [7] Á. Cartea, Derivatives pricing with marked point processes using tick-by-tick data, Quantita- tive Finance 13 (2013) 111–123. [8] G. Dionne, P. Duchesne, M. Pacurar, Intraday value at risk (ivar) using tick-by-tick data with application to the toronto stock exchange, Journal of Empirical Finance 16 (2009) 777–792. [9] S. Guha, R. Rastogi, K. Shim, Rock: A robust clustering algorithm for categorical attributes, Information Systems 25 (2000) 345–366. [10] T. Kanungo, D. Mount, N. Netanyahu, C. Piatko, R. Silverman, A. Wu, An efficient k-means clustering algorithm: Analysis and implementation, Pattern Analysis and Machine Intelligence, IEEE Transactions on 24 (2002) 881–892. 24 [11] G. Nagy, K. Buza, Sohac: Efficient storage of tick data that supports search and analysis, in: P. Perner (Ed.), Advances in Data Mining. Applications and Theoretical Aspects, volume 7377 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2012, pp. 38–51. [12] A. Nanopoulos, H. Gabriel, M. Spiliopoulou, Spectral clustering in social-tagging systems, Web Information Systems Engineering-WISE 2009 (2009) 87–100. [13] K. Oh, K. Kim, Analyzing stock market tick data using piecewise nonlinear model, Expert Systems with Applications 22 (2002) 249–255. [14] T. Ohnishi, T. Mizuno, K. Aihara, M. Takayasu, H. Takayasu, Statistical properties of the moving average price in dollar–yen exchange rates, Physica A: Statistical Mechanics and its Applications 344 (2004) 207–210. [15] T. Qiu, G. Chen, L.X. Zhong, X.R. Wu, Dynamics of bid–ask spread return and volatility of the chinese stock market, Physica A: Statistical Mechanics and its Applications 391 (2012) 2656–2666. [16] D. Salomon, Data compression: the complete reference, Springer-Verlag New York Inc, 2004. [17] N. Sazuka, Analysis of binarized high frequency financial data, The European Physical Journal B-Condensed Matter and Complex Systems 50 (2006) 129–131. [18] M. Takayasu, H. Takayasu, M. Okazaki, Transaction interval analysis of high resolution foreign exchange data, Empirical Science of Financial Fluctuations-The Advent of Econophysics 18 (2002) 25. [19] P. Tan, M. Steinbach, V. Kumar, et al., Introduction to data mining, Pearson Addison Wesley Boston, 2006. [20] I. Witten, E. Frank, Data Mining: Practical machine learning tools and techniques, Morgan Kaufmann, 2011. [21] Z. Wu, On the intraday periodicity duration adjustment of high-frequency data, Journal of Empirical Finance 19 (2012) 282–291. [22] R. Xu, D. Wunsch, et al., Survey of clustering algorithms, Neural Networks, IEEE Transactions on 16 (2005) 645–678. 25 [23] Y. Yabuuchi, J. Watada, Formulation of possibility grade-based fuzzy autocorrelation model and its application to forecasting, International Journal of Intelligent Technologies and Applied Statistics 5 (2012) 321–335. 26 
work_7sil3dhbojfnxb72jnlh2wxb2m	----	Multi-agent framework for third party logistics in E-commerce Multi-agent framework for third party logistics in E-commerce * Wang Ying*, Sang Dayong Tsinghua University-Hongta Group Postdoctoral Station, Yunnan, 653100 China Abstract In an e-commerce environment, the third party logistics (3PL) takes charge of the logistics design, delivery, storage and transportation in a supply chain with its professional and complete value-added services. Beginning with an analysis of the relationships between the 3PL and supply chain members, the authors suggest that only when the 3PL reengineers its logistics business process to accommodate the customer could it maximize the value of the customer. Finally, five intelligent agents, order management agent, logistics process reengineering agent, resource scheduling agent, dynamic union management agent and simulating and evaluating agent are designed to form an e-commerce based 3PL system which, with the collaboration of the five agents, could construct a virtual private logistics teamwork suitable for a certain customer’s need and furthermore, realize the win–win between the customer and the logistics service vendor. q 2005 Published by Elsevier Ltd. Keywords: Third party logistics (3PL); E-commerce; Business process reengineering; Multi-agent 1. Introduction The tendency of the global economic development makes the competition of the supply chains be the main and essential one between two individual enterprises. Only when its whole supply chain keeps high competitive could an enterprise survive for a long time. The rapid development of e-commerce makes it possible to integrate and improve the competitive advantages of the whole supply chain because the openness, globalization, low-cost and high- efficiency of e-commerce extend the internal information network of an enterprise and enable the business to business (B2B) e-business activities among several enterprises. The involved enterprises are connected to be a larger and more interactive supply chain network system in which exist four flows, i.e. flow of goods, flow of information, flow of financing and flow of trading. Integration of four flows is one of the characteristics of e-commerce supply chain even though they are running in different velocities. For the velocity of goods flow is far slower than that of other three ones, the goods flow has been the bottleneck of the supply chain. To increase the flow rate of goods, enterprise are looking for more professional logistics service vendors in succession. 3PL is selected in place of the private logistics systems of enterprises to take charge of part or whole of the supply chain logistics for it is informational, professional, networked and systemized. Therefore, the efficiency of the supply chain logistics is determined by the efficiency of the 3PL’s logistics service. On the other hand, the logistic customer needs to know more, various information of the logistics since there exists venture of route as well as the ownership of the goods and the goods itself are separated. All of these bring several new requirements both on the 3PL service and on its logistics business process. How can a 3PL provide customized services according to different customer requirements and meanwhile its logistics process could be stable enough for long period of development is the main researching topic of this paper. 2. Characteristics of 3PL in e-commerce environment 3PL vendor is defined as a special middleman of logistics in channel who provides other enterprises with whole or part logistics business service, from generic transportation to design, execute and operation whole system of distribution and logistics in certain period by contract form (Duo Zhang, 2000). Expert Systems with Applications 29 (2005) 431–436 www.elsevier.com/locate/eswa 0957-4174/$ - see front matter q 2005 Published by Elsevier Ltd. doi:10.1016/j.eswa.2005.04.039 * Supported by National Natural Science Foundation of China (59990470-4). * Corresponding author. Tel./fax: C86 1082135511. E-mail address: yingwangsang@yahoo.com (W. Ying). http://www.elsevier.com/locate/eswa A 3PL vendor is a professional logistics company getting profit by taking charge of part or whole logistics in the supply chain of a focal enterprise. Nodes including supply chain are joined together by business process (Lambert and Cooper, 2000). As non-supply chain member nodes, 3PL connect their conjoint nodes with logistics agent business services such as customer relation management (CRM), order fulfillment, structural network design, stock manage- ment, transportation management, returns management, etc. The 3PL appeared in advance of supply chain manage- ment and developed firstly within a very slow speed. In spite of its history of several decades, up to now 3PL still occupies a rather low fraction of logistics markets. Even in U.S.A, 3PL only contributes to 6% of related industries (Jian-me, 1999). The 3PL is in the initial stage of development in China and according to the interview of China Storage Association, only 5.9% of commerce enterprises and 18% of production enterprises outsource their logistics to 3PL vendors while the 3PL vendors haven’t taken part in inner production logistics business of the production enterprises (Shao-ji Shen, 2000). At the same time, the popularization of supply chain management provides a good developing environment and a huge required market for 3PL industry. In the fierce global competition an enterprise is faced with a buyer’s market which is ever rapid changing and difficult to be forecasted. The consumers are also becoming more and more dominant in presenting their personalized and customized requirements. Internet-based collaboration used in e-commerce enables the integration of logistics flow, financing flow, information flow, workflow and value-added flow. With the utilization of e-commerce, the 3PL company could frequently reengineer its logistics business process flow and thus improve the customer responding ability and service quality. The focal enterprise outsourcing its logistics to 3PL will decrease its logistics cost and the whole supply chain product stock and therefore, has more ability to accommodate the market’s variation. Developing and improving of e-commerce based 3PL will turn the focal enterprises’ logistics into a socialized, professional one. The professional logistics management of 3PL realizes the fast moving of products among valid supply chains and shortens both the distances from the producers to the consumers and from the supplying markets to the requiring markets. The 3PL companies and all the supply chain node enterprises will win for the improvement of the whole supply chain efficiency and the decreasing of its cost. 3. Integration of 3PL with the supply chain process Being an associating node in a supply chain, 3PL vendor has a consignment-agency relationship with supply chain member enterprises. The building and maintenance of this relationship depends upon the realization of six core values, i.e. shared goals and objectives, trust, mutual dependence, concern for others profitability, open lines of communi- cation, mutual commitment to customer satisfaction (Daniel H. McQuiston, 2001). As the appurtenant of the supply chain, the 3PL must regard the target of the supply chain members, the maximization of the customers’ interest, as one of its main goals and takes its tenet as helping the customers realize their values-added. Therefore, whether the logistics business process or the management of the 3PL vender should be adapted to the needs of business process and management of the supply chain members. From the 3PL vendor’s view, all the business processes of the logistics customers (the focal enterprise) are different and it must build a dynamic logistics business process to make it be able to be integrated with various supply chain business processes. 3.1. Reengineering of logistics business process The most important steps in the inner 3PL logistics business process include management of customer services, product storage and product transportation. Compared with the supply chain business process, 3PL logistics business process has fewer and more stable customers and no supplier involved. The core competition ability of a 3PL vendor is its ability of integrated services to help its customers to optimize their logistics management strategies, build up and operate their logistics systems and even manage their whole distribution systems. Only by providing customized logistics services to various customers as an agency, a 3PL vendor could establish a long-term union relationship with its customers and enhance it continuously. Therefore, logistics agency allowing customization is an available 3PL mode suitable for being integrated with supply chains. In the running of a supply chain, the 3PL vendor is charged with several logistics activities in procurement, transportation and storage of raw materials and machining, packaging and delivering of products. 3PL should take the advantage of its professional logistics ability and reengineer above logistics activities so that to be able to deliver given products to the receiver in a certain period and minimize the logistics cost as well. 3PL will connect the suppliers, the manufacturers and the distributors in supply chains and provide the substance movement and logistics information flow. The logistics value-added services are realized by the changes of the products’ time and value states. Available 3PL logistics business processes suitable for supply chain management include processes of customer relation management, customer service management, cus- tomer order fulfillment, structural logistics network design, stock management, transportation management, returns management, etc. (1) Customer relation management. Its main purpose is to recognize the core customers and customer groups. W. Ying, S. Dayong / Expert Systems with Applications 29 (2005) 431–436432 https://isiarticles.com/article/479 
work_7xjlmdv42zgjrmqnn44vauu7bu	----	13e.xps World Applied Sciences Journal 31 (Applied Research in Science, Engineering and Management): 69-75, 2014 ISSN 1818-4952 © IDOSI Publications, 2014 DOI: 10.5829/idosi.wasj.2014.31.arsem.516 Corresponding Author: Jakub M. Tomczak, Institute of Computer Science, Wroclaw University of Technology, wyb. Wyspianskiego 27, 50-370, Wroclaw, Poland 69 Application of Classification Restricted Boltzmann Machine to Medical Domains Jakub M. Tomczak Institute of Computer Science, Wroclaw University of Technology, wyb. Wyspianskiego 27, 50-370, Wroclaw, Poland Abstract: Recent developments have demonstrated deep models to be very powerful generative models which are able to extract features automatically and obtain high predictive performance. Typically, a building block of a deep architecture is Restricted Boltzmann Machine (RBM). In this work, we focus on a variant of RBM adopted to the classification setting, which is known as Classification Restricted Boltzmann Machine. We claim that this model should be used as a stand-alone non-linear classifier which could be extremely useful in medical domains. Addit ionally, we show how to obtain sparse representation in RBM by adding a regularization term to the learning objective which enforces sparse solution. The considered classifier is then applied to five different medical domains. Key words: Restricted boltzmann machine • classification • sparse • medical domain • diabetes • oncology INTRODUCTION Deep learning paradigm becomes a crucial part of modern machine learning methods because it allows to extract features automatically and obtain high predictive performance [1]. A building block of a deep architecture could be a probabilistic model called Restricted Boltzmann Machine (RBM), used to represent one layer of the deep structure. Restricted Boltzmann Machines are interesting because they are capable of le arning complex featuers and they form a bipartie graph which makes inference easy in them. Moreover, it has been proven that in the sense of Kullback-Leibler divergence RBM is able to arbitrarily well approximate any distribution over binary inputs for properly chosen number of hidden units [11]. Basing on this result it has been proposed to add an additional layer to RBM representing an output variable. In other words, it has been advocated to use RBM as a stand- alone non-linear classifier [8]. Learning flexible classifiers, which are capable of extracting features in an automatic manner and capturing non-linear dependencies, is especially important in medical domain [7]. It has been shown that such classifiers can be very helpful in medical decision support systems, e.g., Boosted SVM for lung cancer patients [21], Graph-based Rule Inducer for diabetes treatment [19], Bagging of decision trees for breast cancer recurrence [17]. Recently, RBM for classification was applied to breast cancer recurrence and in comparison to other methods it obtained very promising results [18]. In this paper, we aim at verifying the applicability of RBM as a predictive model (or diagnostic tool) in five various medical domains. The typical learning algorithm for RBM is \textit {Stochastic Gradient Descent} (SGD) technique which scales well for large data and ensures convergence at rate equal to inverse of the desired accuracy (under mild conditions) [3]. However, there could be two problems with learning RBM using SGD. First, the model could be overcomplete [13], i.e., there are too many hidden units. Second, learnt features could be not enough discriminative, i.e., they are suitable for reconstruction but insufficient for prediction. A possible solution to both of these issues is application of sparse learning. Typical sparse learning is based on adding a regularization term to the learning objective (typically - negative log-likelihood) which penalizes dense solutions. An example of such regularization is L1 norm of model's parameters. It has been shown that this approach allows to learn with overcomplete representations [13] but also leads to worst predictive performance [14]. Another approach is sparse Bayesian learning which aims at introducing latent variable models which enfo rces sparse solutions [14]. In deep learning there are several methods which introduce sparse solutions. Most of them are based on using regularization term which encourages activation of hidden units to be at low rate. One approach applies cross-entropy between desired activation level and World Appl. Sci. J., 31 (Applied Research in Science, Engineering and Management): 69-75, 2014 70 estimated probability of activation [15]. Very similar solution proposes to use L2 norm instead of the corss-entropy [10]. Other one applies Bhattacharyya distance in order to differentiate probability of activation among hidden units for given input [12]. In this paper, we give a different view on how to use probability distance measures and show that for some probabilistic similarity measures (e.g. Kullback-Leibler divergence, Bhattacharyya distance, Mahalanobis distance) one obtains the same regularization term. The paper is organized as follows. In Section II the model of RBM for classification is outlined and its discriminative learning (Section II-B) and sparse learning (Section II-C) are described. In Section III the empirical study is carried out. We compare the RBM for classification with discriminative and sparse learning with well-known classifiers in five medical domains. At the end of the paper (Section IV) conclusions are drawn and directions of future research are indicated. CLASSIFICATION RESTRICTED BOLTZMANN MACHINE Classification Restricted Boltzmann Machine (ClassRBM) [8, 9] is a three-layer undirected graphical model where the first layer consists of visible input variables x∈{0,1}D, the second layer consists of hidden variables (units) h∈{0,1}M and the third layer represents observable output variable y∈{1,2,…,K}. We use the 1-to-K coding scheme which results in representing output as a binary vector of length K denoted by y, such that if the output (or class) is k, then all elements are zero except element yk which takes the value 1. We allow only the inter-layer connections, i.e., there are no connections within layers. A RBM with M hidden units is a parametric model of the joint distribution of visible and hidden variables, that takes the following form: (1) with parameters θ = {b, c, d, W 1, W 2} and where: (2) is an energy function and (3) It can be shown that the following expressions hold true for ClassRBM [8, 9] (Further in the paper, sometimes we omit explicit conditioning on parameters θ) (4) (5) (6) (7) (8) where sigm(⋅) is the logistic sigmoid function, iW  . is the ith row of weights matrix W , , jW  , is jth column of weights matrix W , ijW  is the element of weight matrix W . Prediction: For given parameters θ it is possible to exactly compute the distribution p(y|x,θ) which can be further used to choose the most probable class label. This conditional distribution take s the following form [8, 9]: (9) Pre-computing the terms allows to reduce the time needed for computing the conditional distribution to O(MD+MK) [8, 9]. Learning: We assume given N data D = {xn, yn}, where nth example consists of an observed inputs xn and a target class yn. Typically, learning of a probabilistic model is based on the likelihood function [2]. However, in order to train ClassRBM we may consider two approaches. The first one, called \textit{generative approach}, aims at maximizing the likelihood function for joint distribution p(y|x,θ). The second one, which we refer to as discriminative approach, considers the likelihood function for conditional distribution p(y|x,θ). The problem with generative approach is that it is impossible to calculate exact gradient of the likelihood function for joint distribution (only approximation can be applied, e.g., Contrastive Divergence [6]). On the other hand, the latter approach allows to compute exact gradient [9]. Additionally, we are interested in obtaining high predictive accuracy, thus, it is more advantageous to learn ClassRBM in a discriminative World Appl. Sci. J., 31 (Applied Research in Science, Engineering and Management): 69-75, 2014 71 manner. Therefore, to train ClassRBM we consider minimization of the negative log-likelihood in the following form: (10) A s s tated before, since the distribution p(y|x,θ) can be calculated exactly, the gradient of (10) can be computed exactly too which yields (Function 1a=b i san indicator function which returs 1 if a and b are equal and 0-otherwise): (11) (12) Sparse learning: Sparse representations have been shown to be beneficial in practical applications. From the information-theoretic point of view, sparse representations obtain better generalization performance in comparison to non-sparse ones because the training examples should encoded with as small number of bits as possible. On the other hand, learning sparse features helps to improve discriminative capabilities of features, i.e., sparse features become more class-specific. There are different approaches to introduce sparse representations (see Section 1 for a short review). In this paper, we follow the approach that adds a regularization term to the objective function which enforces sparse solution. Our new objective takes the following form: (13) where λ>0 is a regularization coefficient and Ω(θ) is regularization term. The most popular regularization term is L2 norm of model's parameters which is known as weight decay, i.e., 22( ) || ||Ω θ = θ . However, we will try to sparsify expected activations of hidden units by forcing them to be kept at fixed small level µ. In order to obtain the desired effect the regularization term could be thought as a measure that compares the difference between two probability distributions. The higher is the difference between expected hidden units activations and µ, the stronger is the regularization. Hence, Ω can be, for example, Kullback-Leibler divergence (KLdiv), Bhattacharyya distance (Bdist), Mahalanobis distance (Mdist)(These three mentioned measures can be calculated analytically for normally distributed random variables, which is not true for other measures, e.g., Kolmogorov distance [20]). Let us assume that the hidden unit activity is normally distributed with mean equal p(h j|x,y) and unit variance, i.e., N(p(h j|x,y),1) and the desired activation level is N(µ,1) (For KLdiv, Bdist and Mdist the value of variance turns to be a scaling factor which can be further included in the regularization coefficient. Therefore, for simplicity, we choose unit variance). Then, it turns out that application of KLdiv, Bdist or Mdist to our problem yields the following regularization term: (14) It is interesting that we have obtained exactly the same regularization term as in [10], i.e., squared difference between µ and p(h j|xn, yn). However, our derivation of the regularization term has formal justification while the result in [10] follows from considerations in neuroscience and is given rather ad hoc as a L2 norm of a difference between expected hidden activations and µ. Moreover, our proposition can be further developed by weakening the assumption about unit variance. However, we leave these investigations for further research. The objective function for learning is a sum of the negative log-likelihood and the regularization term. Therefore, we can apply stochastic gradient descent algorithm to the new objective (13). For the negative log-likelihood the gradient is given in (11) and (12) and for the regularization term given in (\ref{eq:regularizationTerm}) we get (for simplicity we denote p(h j|xn, yn) by p jn): (15) Further, we will refer sparse learning of ClassRBM to as sparseClassRBM. EXPERIMENT Setup: During learning ClassRBM and s parseClassRBM for different datasets we kept all parameters fixed. The learning rate was set to 0.001, the number of hidden units was 100. Additionally, we used Nesterov's Accelerated Gradient technique (with parameter's value 0.5) which is special kind of momentum term [16] and a mini-batch of size 100. World Appl. Sci. J., 31 (Applied Research in Science, Engineering and Management): 69-75, 2014 72 ClassRBM and sparseClassRBM were compared with the following classifiers: • CART, • AdaBoost, • LogitBoost, • Tree Bagging, • Random Forest, • SVM, • NeuralNetwork (three hidden layers). Datasets: We evaluate ClassRBM and sparseClassRBM on the following medical datasets: • Heart (joined cleveland and hungarian datasets) [5]: Diagnostic problem of heart disease (less than 50% diameter narrowing in many major vessel or not), • Diabetes [5]: classification problem of the patient as tested positive for diabetes or not, • Indian liver [5]: classification problem of the patient as healthy or with a liver issue, • Sick [5]: diagnostic problem of thyroid disease, • Oncology [17]: predic tion problem of the patient whether there will be a recurrence of breast cancer or not. The datasets are summarized in Table 1. The number of features and the number of examples for each dataset are given. Additionally, we provide the imbalance ratio defined as the number of majority class examples that are divided by the number of minority class examples. RESULTS AND DISCUSSION The results of the experiment are presented in Fig. 1-5 as boxplots. In Table 2 best three methods for each dataset are presented. It can be noticed that ClassRBM together with sparseClassRBM were the most stable methods which; they four times out of five among best three methods (in the 5th case they were on 4th place). Next three best model were: AdaBoost (three times) and Random Forest and LogitBoost (two times). Therefore, we claim that ClassRBM and its sparse version are strong and robust classifiers. However, they are prone to highly imbalanced data (see poor performance on sick dataset, Fig. 4). Table 1:The number of features, the number of examples and the imbalance ratio for the medical datasets used in the experiment Number Number of Imbalance Dataset of inputs examples ration Heart 46 597 1.45 Diabetes 20 768 1.87 Indian liver 48 583 2.49 Sick 57 3772 15.33 Oncology 55 949 1.60 Table 2: Ranking of classifiers according to Kappa for considered five datasets Dataset 1st place 2nd place 3rd place Heart Sparse class RBM Logit boost Class RBM Diabetes Class RBM Sparse class RBM Ada Boost Indian liver Random forest Sparse RBM Ada Boost Sick Cart Random forest Class RBM Oncology Logit boost Ada Boost Class RBM and Sparse class RBM Fig. 1: Boxplot of Kappa values for heart dataset World Appl. Sci. J., 31 (Applied Research in Science, Engineering and Management): 69-75, 2014 73 Fig. 2: Boxplot of Kappa values for diabetes dataset Fig. 3: Boxplot of Kappa values for indian dataset Fig. 4: Boxplot of Kappa values for sick dataset World Appl. Sci. J., 31 (Applied Research in Science, Engineering and Management): 69-75, 2014 74 Fig. 5: Boxplot of Kappa values for onko dataset Comparing ClassRBM to its sparse version basing on the considered datasets is rather inconclusive. It is difficult to tentatively state whether application of sparse regularization term gives better results. However, on heart, diabetes and indian sparsification performes slightly better than ClassRBM, comparably on onko but worst on sic k. Nonetheless, we believe that for more hidden units the results would be more conclusive and the regularization term would prevent from overfitting and maybe even improve predictive performance. We leave this aspect for further research. CONCLUSION In this paper, we have outlined a deep model called Classification Restricted Boltzmann Machine in application to five medical domains. We follow the way of reasoning given in [8] which says that ClassRBM can be used as stand-alone non-linear classifier. Moreover, we claim that this model is very stable and should be used as a state-of-the-art classifier in any domain and especially in medical domain which demands stable solutions. In the experiments, for five different medical problems, we have shown that both discriminative and sparse learning of ClassRBM give very promising results and outperforms well-known strong classifiers like AdaBoost, LogitBoost, TreeBagging and Random Forest. In our study two important issues have arisen which have indicated possible future research. First, ClassRBM fails when data are highly imbalanced and thus there is a need to propose some remedy for that issue. Second, in our proposition of sparse learning we have assumed unit variance. However, such approach is very simplistic and weakening this assumption could give very interesting solutions. ACKNOWLEDGMENT The research conducted by Jakub M. Tomczak has been partially co-financed by the Ministry of Science and Higher Education, Republic of Poland (grant No. B30098I32). REFERENCES 1. Bengio, Y., 2009. Learning deep architectures for AI. Foundations and trends in Machine Learning, 2 (1): 1-127. 2. Bishop, C.M., 2006. Pattern recognition and machine learning. New York: Springer. 3. Bottou, L., 2012. Stochastic Gradient Descent Tricks. In: Neural Networks: Tricks of the Trade. Springer Berlin Heidelberg, pp: 421-436. 4. Cohen, J., 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement. 20 (1): 37-46. 5. Frank, A. and A. Asuncion, 2010. UCI machine learning repository. 6. Hinton, G.E., 2002. Training products of experts by minimizing contrastive divergence. Neural Computation, 14 (8): 1771-1800. 7. Kononenko, I., 2001. Machine learning for medical diagnosis: History, state of the art and perspective. Artificial Intelligence in Medicine, 23 (1): 89-109. 8. Larochelle, H. and Y. Bengio, 2008. Classification using discriminative restricted Boltzmann machines. ICML. 9. Larochelle, H., M. Mandel, R. Pascanu and Y. Bengio, 2012. Learning algorithms for the classification restricted Boltzmann machine. The Journal of Machine Learning Research, 13: 643-669. World Appl. Sci. J., 31 (Applied Research in Science, Engineering and Management): 69-75, 2014 75 10. Lee, H., C. Ekanadham and A. Ng, 2007. Sparse deep belief net model for visual area V2. In: Advances in Neural Information Processing Systems, pp: 873-880. 11. Le Roux, N. and Y. Bengio, 2008. Representational power of restricted Boltzmann machines and deep belief networks. Neural Computation, 20 (6): 1631-1649. 12. Le Roux, N., H. Larochelle and Y. Bengio, 2008. Discriminative Training of RBMs using Bhattacharyya Distance. Learning Workshop, Cliff Lodge, Snowbird, Utah. 13. Lewicki, M.S. and T.J. Sejnowski, 1998. Learning nonlinear overcomplete representations for efficient coding. Advances in Neural Informatio n Processing Systems, pp: 556-562. 14. Mohamed, S., Ghahramani, Z., Heller, K.A. Bayesian and L1, 2012. Approaches for Sparse Unsupervised Learning. In: Proceedings of the 29th International Conference on Machine Learning (ICML-12), pp: 751-758. 15. Nair, V. and G.E. Hinton, 2009. 3D object recognition with deep belief nets. In: Advances in Neural Information Processing Systems, pp: 1339-1347. 16. Sutskever, I., Martens, J. Dahl and G. Hinton, 2013. On the importance of initialization and momentum in deep learning. ICML. 17. Štrumbelj, E., Z. Bosni´c, I. Kononenko, B. Zakotnik and C. Grašic Kuhar, 2010. Explanation and reliability of prediction models: The case of breast cancer recurrence. Knowledge and Information Systems, 24: 305-324. 18. Tomczak, J.M., 2013. Prediction of breast cancer recurrence using Classification Restricted Boltzmann Machine with Dropping. arXiv preprint arXiv:1308.6324. 19. Tomczak, J.M. and A. Gonczarek, 2013. Decision rules extraction from data stream in the presence of changing context for diabetes treatment. Knowledge and Information Systems, 34 (3): 521-546. 20. Zhou, S.K. and R. Chellappa, 2006. From sample similarity to ensemble similarity: Probabilistic distance measures in reproducing kernel hilbert space. IEEE Transactions on Pattern Analysis and Machine Intelligence. 28 (6): 917-929. 21. Zieba, M., J.M. Tomczak, M. Lubicz, J. S´wia?tek, 2014. Boosted SVM for extracting rules from imbalanced data in application to prediction of the post-operative life expectancy in the lung cancer patients. Applied Soft Computing, 14 (A): 99-108. 
work_7yvewrusbbecrkx57q2e6veq4q	----	Automatic expert system based on images for accuracy crop row detection in maize fields Automatic expert system based on images for accuracy crop row detection in maize fields J.M. Guerrero , M. Guijarro , M. Montalvo , J. Romeo , L Emmi , A. Ribeiro , G. Pajares A B S T R A C T This paper proposes an automatic expert system for accuracy crop row detection in maize fields based on images acquired from a vision system. Different applications in maize, particularly those based on site specific treatments, require the identification of the crop rows. The vision system is designed with a defined geometry and installed onboard a mobile agricultural vehicle, i.e. submitted to vibrations, gyros or uncontrolled movements. Crop rows can be estimated by applying geometrical parameters under image perspective projection. Because of the above undesired effects, most often, the estimation results inaccurate as compared to the real crop rows. The proposed expert system exploits the human knowledge which is mapped into two modules based on image processing techniques. The first one is intended for separating green plants (crops and weeds) from the rest (soil, stones and others). The second one is based on the system geometry where the expected crop lines are mapped onto the image and then a correction is applied through the well-tested and robust Theil-Sen estimator in order to adjust them to the real ones. Its performance is favorably compared against the classical Pearson product-moment correlation coefficient. 1. Introduction 1.1. Problem statement Machine vision systems onboard robots are being increasingly used for site specific treatments in agriculture. With such arrange- ments, the robot navigates and acts over a site-specific area of a larger farm (Davies, Casady, & Massey, 1998), where the vision sys- tems can supply abundant information. An important issue related with the application of machine vi- sion methods is that concerning the crop row and weed detection, which has attracted numerous studies in this area (Burgos-Artizzu, Ribeiro, Tellaeche, Pajares, & Fernández-Quintanilla, 2009; Guerrero, Pajares, Montalvo, Romeo, & Guijarro, 2012; López- Granados, 2011; Montalvo et al., 2012; Onyango & Marchant, 2003; Sainz-Costa, Ribeiro, Burgos-Artizzu, Guijarro, & Pajares, 2011; Tellaeche, Burgos-Artizzu, Pajares, & Ribeiro, 2008; Tellaeche, Burgos-Artizzu, Pajares, Ribeiro, & Fernández-Quintanilla, 2008). The goal is to eliminate weeds to favor the growth of crops. The vision system consists of a CCD-based calibrated camera with known intrinsic parameters, i.e. focal length, lens distortion, image center and CCD sensor sizes and pixel resolutions. The cam- era is located in front of the robot, inclined with a tilt angle (pitch) and at a high from the ground. Yaw and roll angles are also known. This allows determining the rotation and translation matrices defining the extrinsic parameters. Thus, areas in the field can be identified onto the image plane. This means that given an element in the field, with its spatial location, we can determine its relative positioning on the image. The vehicle navigates on a real terrain presenting irregularities and roughness. This produces vibrations and also swinging mainly in the pitch and roll angles. The yaw angle is assumed to be correct because otherwise the robot navigates erroneously out of the crop rows. Moreover, the spacing of crop rows in the field is also known. Because of the above, most often, the mapped expected crop rows in the image do not match with the real ones, this inaccurate esti- mation impedes the application of correct site specific treatments. On the other hand the discrimination of crops and weeds in the im- age is a very difficult task because their Red, Green and Blue spec- tral components display similar values. This means that no discrimination is possible between crops and weeds based on the spectral signatures. Thus, the best option is to locate the crop rows in the image with the maximum accuracy as possible. Indeed, if the crop rows are well located, we can accurately identify those pixels along and around the detected line as crops and the remainder, which are moved away, can be considered as weeds. To achieve this goal, we propose an automatic expert system, which exploits the human knowledge, with two main modules based on image processing techniques, as described later. 1.2. Revision of methods Several strategies have been proposed for crop row detection. Fontaine and Crowe (2006) tested the abilities of fourth line detec- tion algorithms to determine the position and the angle of the cam- era with respect to a set of artificial rows with and without simulated weeds. These were stripe analysis, Hough transform, blob analysis and linear regression. The following is a list of crop row detection methods grouped into different categories including the above. 1.2.1. Methods based on the exploration of horizontal strips Sogaard and Olsen (2003) apply RGB color image transformation to gray scale. This is done by first dividing the color image into its red, green and blue channels and then by applying the well-tested methods to extract living plant tissue described in Woebbecke, Meyer, von Bargen, and Mortensen (1995). After this, the gray scale image is divided into horizontal strips where maximum gray values indicate the presence of a candidate row, each maximum deter- mines a row segment and the center of gravity of the segment is marked at this strip position. Crop rows are identified by joining marked points through a similar method to the one utilized in the Hough transform or by applying linear regression. Sainz-Costa et al. (2011) have developed a strategy based on analysis of video sequences for identifying crop rows. Crop rows persist along the directions defined by the perspective projection with respect the 3D scene in the field. Exploiting this fact, they apply gray scale transformation based on the approach proposed by Ribeiro, Ferná- ndez-Quintanilla, Barroso, and García-Alegre (2005) and then the image is binarized applying a thresholding technique. Each image is divided into four horizontal strips. Rectangular patches are drawn over the binary image to identify patches of crops and rows. The gravity centers of these patches are used as the points defining the crop rows and a line is adjusted considering these points. The first frame in the sequence is used as a lookup table that guides the full process for determining positions where the next patches in subsequent frames are to be identified. Hague, Tillet, and Wheeler (2006) transform the original RGB image to gray scale. The transformed image is then divided into eight horizontal bands. The intensity of the pixels across these bands exhibits a periodic variation, due to the parallel crop rows. Since the camera character- istics, pose and the crop row spacing are known a priori, the row spacing in image pixels can be calculated for each of the horizontal bands using a pinhole model of the camera optics. A band-pass filter can then be constructed which will enhance this pattern, and has a given frequency domain response. Sometimes horizontal patterns are difficult to extract because crops and weeds form a unique patch. 1.2.2. Methods based on the Hough transformation According to Slaughter, Giles, and Downey (2008), one of the most commonly used machine vision methods for identifying crop rows is based upon the Hough (1962) transform. It was intended to deal with discontinuous lines, where the crop stand is incomplete with gaps in crop rows due to poor germination or other factors that result in missing crop plants in the row. It has been intended for real- time automatic guidance of agricultural vehicles (Astrand & Baerveldt, 2005; Hague, Marchant, & Tillett, 1997; Leemans & Destain, 2006; Marchant, 1996). It is applied to binary images, which are obtained by applying similar techniques to the ones explained above, i.e. RGB image transformation to gray scale and binarization (Tellaeche, Pajares, Burgos-Artizzu, & Ribeiro, 2011; Tellaeche et al., 2008; Tellaeche, Burgos-Artizzu, Pajares, Ribeiro, et al., 2008). Gee, Bossu, Jones, and Truchetet (2008) apply a double Hough transform under the assumption that crop rows are the only lines of the image converging to the vanishing point, the remainder lines are rejected, additional constraints such as inter-row spacing and perspective geometry concepts help to identify the lines. It is required to deter- mine the threshold required by the Hough transform to determine maximum peaks values (Jones, Gée, & Truchetet, 2009a, 2009b) or predominant peaks (Rovira-Más, Zhang, Reid, & Will, 2005). Depending on the crop densities several lines could be feasible and a posterior merging process is applied to lines with similar parame- ters (Tellaeche et al., 2008; Tellaeche, Burgos-Artizzu, Pajares, Ribeiro, et al., 2008; Tellaeche et al., 2011). Although intended for real-time, as mentioned before, in our images, where crop and weed plants contribute on the Hough parameter estimation, this method becomes computationally expensive (Ji, & Qi, 2011). On the other hand, the randomized Hough transform requires selecting pairs of points to be considered as a line, i.e. pairs of points belonging to a crop row. If we apply this technique in images where edge points have been extracted, the selection of those pairs becomes highly complex because weeds are also involved. 1.2.3. Vanishing point-based Pla, Sanchiz, Marchant, and Brivot (1997) propose an approach that identifies regions (crops/weeds and soil) by applying color im- age segmentation. They use the skeleton of each defined region as a feature to work out the lines that define the crop. The resulting skeletons and their properties, defined as chains of connected con- tour points, allow the identification of crop rows oriented toward the vanishing point. This process is highly dependent of skeletons, which are not always easy to extract, specially taking into account that weed patches are present. Romeo et al. (2012) apply also knowledge concerning the position of the vanishing point and the crop rows arrangement in the field to detect the expected crop rows. The process is based on the identification of maximum accu- mulation of green pixels along lines oriented toward the vanishing point. A supervised fuzzy clustering method is the proposed strat- egy for greenness identification. This makes the method highly dependent on the training phase unlike the one prosed in this ap- proach which is automatic, i.e. unsupervised. 1.2.4. Stereo-based approach Kise, Zhang, and Rovira-Más (2005) or Kise and Zhang (2008) developed a stereo vis ion-based agricultural machinery crop-row tracking navigation system. Stereo-image processing is used to determine 3D locations of the scene points of the objects of interest from the obtained stereo image. Those 3D positions, determined by means of stereo image disparity computation, provide the base information to create an elevation map that uses a 2D array with varying intensity to indicate the height of the crop. This approach re- quires crops with significant heights with respect the ground. Be- cause in maize fields, during the treatment stage, the heights are not relevant, it becomes ineffective in our application. Rovira-Más, Zhang, and Reid (2008) have applied and extended stereovision techniques to other areas inside Precision Agriculture. Only feasible if crops or weeds in the 3D scene display a relevant height. 3.2.5. Methods based on blob analysis This method finds and characterizes regions of contiguous pix- els of the same value in a binarized image (Fontaine & Crowe, 2006). The algorithm searches for white blobs (inter-row spaces) of more than 200 pixels, under the assumption that smaller blobs could represent noise in the crop rows. Once the blobs were iden- tified, the algorithm determined the angle of their principal axes and the location of their centre of gravity. For a perfectly straight white stripe, the centre of gravity of the blob was over the centre line of the white stripe, and the angle was representative of the angle of the inter-row spaces. The algorithm returned the angle and center of gravity of the blob closest to the center of the image. Identification of blobs in areas with weed patches does not distin- guish between blobs caused by weeds and crops. 1.2.6. Methods based on the accumulation of green plants Olsen (1995) proposed a method based on the consideration that along the crop row appear an important accumulation of green parts in the image. The image is gray scale transformed where green parts appear clearer that the rest. A sum-curve of gray levels is obtained for a given rectangular region exploring all col- umns in the rectangle. It is assumed that vertical lines follow this direction in the image. The images are free of perspective projec- tion because they are acquired with the camera in orthogonal po- sition. A sinusoidal curve is fitted by means of least squares to the sum-curve previously obtained. Local maxima of the sinusoid pro- vide row centers locations. 1.2.7. Methods based on frequency analysis Because crop rows are vertical in the 3D scene, they are mapped under perspective projection onto the image displaying some behavior in frequency domain. Vioix et al. (2002) exploit this fea- ture and apply a bi-dimensional Gabor filter, defined as a modula- tion of a Gaussian function by a cosine signal. The frequency parameter required by the Gabor filter is empirically deduced from the 2D-Fast Fourier Transform (Bossu, Gee, Guillemin, & Truchetet, 2006). Bossu, Gee, Jones, and Truchetet (2009) apply wavelets to discriminate crop rows based on the frequency analysis. They exploit the fact that crop rows are well localized in the frequency domain; thus selecting a mother wavelet function with this fre- quency the crop rows can be extracted. Crops, in the images we have studied, do not display clear frequency contents in the Fourier space, therefore the application of filters based on the frequency becomes a difficult task. 1.2.8. Methods based on linear regression Some techniques above apply this approach. Billingsley and Schoenfisch (1997) reported a crop detection system that is rela- tively insensitive to additional visual 'noise' from weeds. They used linear regression in each of three crop row segments considered and a cost function analogous to the moment of the best-fit line to detect lines fitted to outliers (i.e., noise and weeds) as a means of identifying row guidance information. Montalvo et al. (2012) ap- ply a linear regression for crop row detection in images containing high weeds densities. Some templates are used to guide the detec- tion. Linear regression is also applied in Sogaard and Olsen (2003). Linear regression is highly sensitive to isolated weeds patches placed on the inter crop rows and also for weeds patches over- lapped with crops. In this paper we also apply linear regression based on the Theil Sen estimator (Sen, 1968; Theil, 1950) which is free of the above sensitivity and has been proven in statistics with satisfactory results. 2. Design of the automatic expert system 2.1. System architecture The system architecture is inspired on the human expert knowl- edge about the specific application and also considering the require- ments that must be fulfilled. Astrand (2008) and Slaughter et al. (2008) propose a list of requirements for guidance systems that can be also considered for crop row detection, which in essence is a similar problem. Knowledge and requirements are mapped as fol- lows to build the architecture of the proposed automatic expert sys- tem for the accuracy crop row detection based on images. (a) Both crop and weeds display similar color spectral compo- nents and during the treatment their growth stages are sim- ilar, i.e. with similar height in the plants. (b) Crop rows are accumulations of green plants following spe- cific alignments oriented to the vanishing point. Crops are sown, not manually planted, and the inter-line distances in the field are known. (c) Weeds appear on isolated or overlapped patches with respect crops with irregular distributions. (d) Crop rows must be located with the most accuracy as possi- ble, regardless the distribution of weeds patches around crop and also considering that crop plants could miss along crop lines, as a common situation. Original ¡mage Image Segmentation Apply combined vegetation indices & greenness reinforcement Otsu thresholding Binary image (crop lines & weeds patches) Expert system Crop row detection Crop row estimation Trace expected crop lines according to the camera system geometry Apply Theil-Sen estimator Fig. 1. Automatic expert system architecture. (e) Camera s y s t e m g e o m e t r y is k n o w n , i.e. t h e intrinsic a n d extrinsic p a r a m e t e r s . (f) The robot navigates on u n e v e n t e r r a i n s w i t h p e r h a p s a b u n - d a n t irregularities. (g) The s y s t e m m u s t w o r k on r e a l - t i m e . This r e p r e s e n t s a t r a d e - off b e t w e e n t h e speed of t h e robot a n d t h e c o m p u t a t i o n a l cost. Based on this k n o w l e d g e a n d r e q u i r e m e n t s a n d also considering a d v a n t a g e s a n d shortcomings of t h e different crop r o w d e t e c t i o n m e t h o d s , t h e a u t o m a t i c e x p e r t s y s t e m is designed consisting of t w o m a i n m o d u l e s : image segmentation a n d crop rows estimation. Fig. 1 displays schematically t h e s e t w o m o d u l e s w i t h t h e c o r r e - s p o n d i n g processes. This results in a r o b u s t e x p e r t system, m a k i n g t h e c o n t r i b u t i o n of this paper. 2.2. Image segmentation Image s e g m e n t a t i o n is focused on t h e s e p a r a t i o n of g r e e n plants (crops a n d w e e d s ) from t h e rest (soil, s t o n e s a n d o t h e r s ) . According t o point (a) in t h e list of k n o w l e d g e a n d r e q u i r e m e n t s above, t h e b e s t o p t i o n t o identify w e e d s a n d c r o p is t h e application of v e g e t a t i o n indices i n s t e a d of m e t h o d s b a s e d on height discrim- ination. Vegetation indices are well t e s t e d m e t h o d s , Guijarro et al. (2011) p r o p o s e a c o m b i n a t i o n of v e g e t a t i o n indices, w h i c h is t h e o n e c h o s e n in this p a p e r b e c a u s e its p e r f o r m a n c e in m a i z e fields. 2.2.1. Combination of vegetation indices Given a n original i n p u t i m a g e in t h e RGB color space, w e apply t h e following n o r m a l i z a t i o n s c h e m e , w h i c h is usually applied in a g r o n o m i c i m a g e s e g m e n t a t i o n (Gee et al., 2008), Rn C„ , B„ R„ B„ R„ B„ b = Rn B„ (1) w h e r e R, G a n d B are t h e normalized RGB coordinates ranging from 0 t o 1 and are o b t a i n e d as follows: R„ Rn B„ Bn (2) w h e r e Rm3X = Gm a x = B m a x = 255 for our 24-bit color images. Vegetation indices t o b e c o m b i n e d are c o m p u t e d as follows Guijarro et al. (2011), Excess Green ( W o e b b e c k e et al., 1 9 9 5 ; Ribeiro et al., 2005) Color index of vegetation extraction (Kataoka, Kaneko, Okamoto, & Hata, 2 0 0 3 ) Vegetativen ( H a g u e et al., 2 0 0 6 ) (3) ExG = 2g -r -b C/VE = 0 . 4 4 1 r - 0 . 8 1 1 g (4) + 0 . 3 8 5 b + 1 8 . 7 8 7 4 5 VEG = (5) r"b w i t h a = 0.667 as in its reference Excess green minus excess red (6) (Meyer & Camargo-Neto, = _ 2 0 0 8 ; Neto, 2 0 0 4 ) w h e r e excess red is c o m p u t e d as follows (Meyer, Hindman, & Lakshmi 1998): ExR = 1 . 4 r - g . According t o Guijarro et al. (2011) t h e above four indices are c o m b i n e d t o obtain t h e resulting value COM as follows, COM = WEXGEXG + WEXGEXGR + W Q V E C / V E + WVEGVEG (7) w h e r e wExG = 0.25, wExGR = 0.30, wCJV£ = 0.33 a n d wVEG = 0.12 are t h e w e i g h t s for each index, r e p r e s e n t i n g t h e i r relative relevance in t h e combination. The resulting c o m b i n e d image COM, is linearly m a p p e d t o range into t h e interval [0,1]. 2.2.2. Greenness reinforcement Romeo et al. (2012) p r o p o s e a fuzzy clustering strategy w h e r e t h e cluster c o n t a i n i n g pixels belonging t o g r e e n plants has b e e n analyzed. Clusters c o n t a i n pixels w i t h t h e t h r e e spectral c o m p o - n e n t s in t h e RGB m o d e l as features. Obviously, a n d as expected, t h e g r e e n spectral c o m p o n e n t is d o m i n a n t . On a v e r a g e this c o m p o - n e n t in t h e cluster c e n t e r for g r e e n plants r e p r e s e n t s values a b o v e t h e 36% w i t h r e s p e c t t h e o t h e r t w o c o m p o n e n t s . Exploiting this k n o w l e d g e a n d applying t h e trivial reasoning t h a t pixels c o m i n g from plants should have t h e i r g r e e n c o m p o n e n t d o m i n a n t , w e a c c e n t u a t e t h e g r e e n n e s s in COM by multiplying t h e i r values by g in Eq. (1), i.e. a n e w g r e e n n e s s is o b t a i n e d a s : GA = COM*g. The multiplication is carried o u t pixel by pixel a n d GA is linearly m a p p e d to range in [0,1]. Because g r e p r e s e n t s t h e p e r c e n t a g e of t h e g r e e n c o m p o n e n t , t h e result o b t a i n e d r e p r e s e n t s t h e e m p h a s i s in t h e g r e e n n e s s . 2.2.3. Thresholding Given t h e t r a n s f o r m e d i m a g e GA, t h e n e x t s t e p is its binariza- tion for posterior processing. An easy t h r e s h o l d b a s e d on t h e m e a n gray level of t h e i m a g e ( h i s t o g r a m ) has b e e n i m p l e m e n t e d in Gee et al. (2008) w h e r e t h e living plant m a t e r i a l (crop or w e e d ) a p p e a r s as w h i t e spots a n d t h e rest (i.e. soil surface, stones, s h a d o w s ) as black. Also in Guijarro et al. (2011) t h e w e l l - k n o w n Otsu's (1979) m e t h o d , traditionally applied for binarization, has b e e n applied. More c o m p l e x a p p r o a c h e s h a v e b e e n also applied such as t h e o n e u s e d in Bossu et al. (2009), b a s e d on t h e fe-means clustering m e t h o d . W e h a v e c h o s e n t h e Otsu's m e t h o d for its w e l l - k n o w n p e r f o r m a n c e as r e p o r t e d in Meyer a n d Camargo-Neto (2008) a n d also b a s e d on t h e s t u d y of Sezgin a n d Sankur (2004) w h e r e its per- formance has b e e n t e s t e d in images w h e r e t h e n u m b e r of pixels in b o t h p a r t s of t h e i m a g e h i s t o g r a m t h a t Otsu's p r o d u c e s is close t o e a c h o t h e r . Fig. 2(a) displays a n original i m a g e in t h e RGB color space of a m a i z e crop field. The color space t r a n s f o r m a t i o n by a p p l y i n g GA is displayed in Fig. 2(b). Fig. 2(c) displays t h e i m a g e t r a n s f o r m a t i o n from i m a g e in Fig. 2(b) by applying t h e Otsu's m e t h o d . Note t h e l a n d m a r k s in t h e image, w h i c h a r e explained later in Section 3. 2.3. Crop row estimation This m o d u l e is i n t e n d e d t o apply t h e k n o w l e d g e e m b e d d e d in points (b)-(f), Section 2 . 1 , at t h e s a m e t i m e it provides specific solutions for t h e r e q u i r e m e n t s e x p r e s s e d in such points. 2.3.1. Tracing expected crop lines The robot navigates on u n e v e n t e r r a i n s w i t h p e r h a p s a b u n d a n t irregularities, t h e k n o w l e d g e of extrinsic p a r a m e t e r s of t h e vision s y s t e m does not suffice b e c a u s e t h e c a m e r a is c o n t i n u o u s l y in- volved in a p e r m a n e n t swinging. W e p r o p o s e t h e c u s t o m i z a t i o n of t h e Theil-Sen regression e s t i m a t i o n a p p r o a c h , b e c a u s e of its w e l l - t e s t e d p e r f o r m a n c e in statistics. Because t h e c r o p rows a r r a n g e m e n t are k n o w n in t h e field a n d also t h e extrinsic a n d intrinsic c a m e r a s y s t e m p a r a m e t e r s , t h e ex- p e c t e d c r o p r o w locations in t h e i m a g e can b e e s t i m a t e d a n d m a p p e d as k n o w n lines o n t o t h e i m a g e (Fu, González, & Lee, 1987; Hartley & Zisserman, 2006). U n d e r t h e a s s u m p t i o n of ideal s y s t e m g e o m e t r y t h e e x p e c t e d lines should m a t c h a n d overlap t h e imaged real crop r o w s . Nevertheless, d u e t o u n e v e n t e r r a i n s a n d errors in t h e c r o p r o w a l i g n m e n t d u r i n g t h e sowing, this often does not occur. (a) t \ '% (c) Fig. 2. (a) Original image; (b) GA index extracted from the image in (a); (c) binary image after Otsu thresholding. Therefore, under the above consideration, two cases can appear with respect to the expected and the imaged real crop lines: (a) they match; (b) they do not match. In the first case, the detection method needs to verify this matching. In the second case, a line location correction must be applied until the real crop row is lo- cated. Under this approach the system geometry through the intrinsic and extrinsic parameters guides the crop row detection process. Now the question is: how can we verify the expected lines match or not with the real crop ones? Because we have available white pixels representing green plants in the binary image, we can adjust a straight line for specific pixel alignments that are ex- pected to identify crop rows. This will represent the real crop line. So, because we have both, the expected straight line equation and the adjusted one, we are able to verify the correct or incorrect match for both lines. Thus, we focus the effort in methods for esti- mating the parameters defining real crop lines. 2.3.2. Correction of the expected crop lines: Theil-Sen estimator An important problem to be addressed in our approach is that the method selected can cope with specific pixel alignments but also must be robust enough to avoid significant deviations caused by weeds that are not aligned and placed more or less near the main crop row alignments. This is the main issue addressed in this work. Stewart (1999) provides a tutorial oriented toward robust parameter estimation in computer vision. Two frequently tech- niques used are least-median of squares (LMS) (Rousseeuw, 1984) and M-estimators (Hampel, Rousseeuw, Ronchetti, & Stahel, 1986; Huber, 1981), but a huge volume of data implies that param- eter estimation techniques in computer vision are heavily over constrained, even for problems where low-level feature extraction, such as edge detection, are applied. This in turn implies that parameter estimation problems in vision should be solved by least squares or, more generally, maximum likelihood estimation (MLE) techniques. Unfortunately, computer vision data are rarely drawn from a single statistical population as required for effective use of MLE. In our approach we must estimate two parameters, defining the straight line equations associated to the corresponding crop rows, they are the slope a and the intercept /¡. For the linear regression approach, after several studies, we observed that the least squares estimator of a regression, coefficient a is vulnerable to gross errors and the associated confidence interval is, in addition, sensitive to non-normality of the parent distribution. With other measures, for example, the breakdown point (Rousseeuw & Leroy, 1987) a small number of outlying data can cause an estimate to diverge arbitrarily far from the true estimate. From the point of view of our approach, this means that few weed pixels can move the least squares fit far from the true fit, i.e. far of the real crop line. A second measure of robustness is the influence function (Hampel et al., 1986; Huber, 1981), in which the change in an estimate caused by insertion of outlying data, using a function of the distance, also causes false estimations because it should tend to zero with increasing distance to achieve robustness. Alternative estimators for the regression coefficient, a, based on suitable rank tests are proposed by Mood (1950). They apply the estimation of both parameters a and ¡5 simultaneously using the statistical median by trial and error. Adichie (1967) proposes a more restrictive method under the assumption that the set of points to be adjusted is an absolutely continuous and symmetric distribution function with also an absolutely continuous and square integrable density function; Theil (1950) proposes a very simple estimator for a using also the statistical median; Dytham (2011) and Sen (1968) study a simple and robust estimator for a based on Kendall's (1955) tau rank correlation, a simple non- parametric test that can be used instead of normal regression. Hence, the estimator for a and ¡5 is the median of the set of slopes, where a simple slope is computed between every possible pair of pixels i and j with image coordinates (x„ y,) and (x,-, y¡) respec- tively, finally the median slope is then selected as the best estimate for a. Based on the above considerations, we select the Theil-Sen esti- mator as proposed in Massart (1997), because of its statistical effi- ciency and its robustness, even for low image resolutions, resulting in a promising approach in agricultural images containing crop rows. Nevertheless, as we will see later, its effectiveness from the real-time point of view is relatively low. This means that further analysis or new software and hardware implementations should be required for real-time processing. A straight line is represent by its slope a and its intercept ¡5 as follows, Y = aX- (8) Given a distribution of n pixels the goal is to adjust a straight line to such distribution. The Theil-Sen estimator evaluates pairs of pixels i and j and compute the slope over the set of all possible pairs of such pixels, i.e. over the n(n - l)/2 possible combinations. This is carried out as follows, a = med[Sij\Sij = y¡-y¡ x¡ T^Xj, i,j = 1,2, (9) The estimation of the intercept, /¡, is computed as the statistical median of the intercepts obtained with the robust slope a in (9). The E parameter, set to 10~3, is introduced to avoid exactly vertical lines with slope toward oo. Polar coordinates could be used to avoid this problem. Nevertheless, vertical lines do not appear in our real application. This is carried out as follows, • med(y¡ - <xx¡); Vi = 1,2, (10) 3. Results The images used for this study belong to maize crops. They were captured with a BASLER 17FC 1400 color camera during April/May 2011 in a 1.7-ha experimental field of maize on La Pov- eda Research Station, Arganda del Rey, Madrid. All acquisitions were spaced by five/six days, i.e. they were obtained under differ- ent environmental lighting conditions and different growth stages in maize and weed plants. The digital images were captured under perspective projection and stored as 24-bit color images with res- olutions of 1392 x 1038 pixels saved in RGB (Red, Green and Blue) color space in the TIFF format. The images were processed under LabVIEW Real-Time (2012) from National Instruments, release 2011, under a CompactRIO-9082 1.33 GHz dual-core Intel Core i7 processor, including LX150 FPGA with Real-Time Operating Sys- tem. The proposed algorithm is developed in C++ with MS Visual Studio and compiled as a DLL, which is embedded as an additional module in LabVIEW. A set of 240 images was processed. This equipment is intended to fulfill the real-time specifications ex- pressed in point (g) Section 2.1. The camera extrinsic and intrinsic parameters are: pitch an- gle = 20°, roll angle = 0° and yaw angle = 0° with the camera placed at a height of 1.5 m from the ground; the focal length was 8 mm. An illustrative result, displayed in Fig. 3, is the outperformance of the proposed Theil-Sen estimator as compared to linear regres- sion based on the Pearson product-moment correlation coefficient. The green pixels in the right crop line are the annotated pixels ob- tained by considering the expected crop line as given by the appli- cation geometric transformations according to the intrinsic and extrinsic parameters and the margin of tolerance set to 150 pixels for each side around the expected crop line, making a total margin of 300 pixels. As we can easily infer, pixels far away from the cen- tral ones are weeds, i.e. they appear scattered in the inter-rows. Red line1 in the image is estimated by applying the Theil-Sen estimator and blue line is the one estimated by linear regression based on the Pearson product-moment correlation coefficient. As we can easily see, the best adjusting is achieved by the Theil-Sen's method. This is because it is robust enough against pixel dispersions. On the contrary, the regression-based method is sensitive to this kind of dispersion because it is based on the computation of minimum distances and the scattered pixels exert an important attractiveness. The performance of the Theil-Sen estimator against the Pearson product-moment correlation coefficient is studied through a qual- itative analysis, based on the human expert criterion because no ground truth images are available. By visual inspection of real crop row on the images, the expert determines the best adjustment of the estimated crop lines. Because this approach is oriented to work in the future in real- time applications once the quality performance is achieved, we analyze the behavior of both estimators when applied to images with different resolutions, now under a quantitative analysis. Fig. 4(a) and (b) represent an illustrative example representing an image of the 240 ones analyzed. The area of interest is delimited by landmarks with a wide of 2.25 m (covering three crop rows with inter-row spacing of 0.75 m) and 4 m long. This area is the one to be processed during a normal operation of the agricultural robot. The setting of the intrinsic and extrinsic parameters is the one described in Section 2.1. The camera system was placed 2 m from the area of interest, where a horizontal landmark line delimits the bottom part. 1 For interpretation of color in the figures, the reader is referred to the web version of this article. Fig. 3. Example of crop lines correction by Theil-Sen estimator (red line) and Pearson product-moment correlation coefficient (blue line) considering the dispersion of plants (green pixels). 3.1. Qualitative analysis Fig. 4(a) displays in red those pixels representing green plants (crop and weeds). The expected crop lines, according to the system geometry based on extrinsic and intrinsic parameters, are drawn as yellow lines. A simple image inspection based on the human expert criterion allows us to infer that the expected crop lines do not match accurately with the real ones. This observation is particu- larly relevant in the left crop line. Perhaps this situation is due to the fact that this crop row was the outer line applied by the seeder machine and its location differs from the 0.75 m of the ideal inter- line spacing in maize fields. The deviation of the expected central and right crop lines from the real ones is less marked than in the left one. It is obvious that under this situation the expected crop lines need correction. The three real crop lines contain some gaps produced by errors during the sowing or perhaps because the maize seeds have not emerged. Fig. 4(b) displays corrections of the yellow lines by applying both the Theil-Sen estimator (red lines) and regression through the Pearson product-moment correlation coefficient (blue lines). Based on the human expertise, it is easy to see how the three red lines are well adjusted by Theil-Sen. Indeed, they follow the cen- tral part of the furrow. Perhaps in the central crop line a slight deviation can be appreciated, which compared against the one pro- duced by the Pearson product moment becomes irrelevant. The huge deviation produced by this last estimator in the central crop line is due to the presence of isolated patches (marked with the cir- cle), probably weeds, which have been considered during the pro- cedure of estimation. Something similar happens with respect the right crop lines, but the deviation it is less pronounced. This allows us to conclude that Theil-Sen works appropriately under this type of situations which are abundant in maize fields. This is the general behavior observed in the set of 240 images analyzed, which allows us to verify the outperformance of the Theil-Sen estimator as compared to the Pearson product moment. The margin of tolerance used was again 150 for each side. In order to detect the real crop lines over the image we have also applied the Hough transformation (Slaughter et al., 2008). De- spite we apply geometric constraints considering the extrinsic and intrinsic parameters; the method produces abundant peaks on the accumulator cells, making difficult the selection of the correct one that identifies a crop row. A lot of lines, with different slopes, appear for a unique crop row. This requires a careful peak thres- holding selection (Jones et al., 2009a, 2009b; Rovira-Más et al., 2005). 3.2. Quantitative analysis From a point of view of quantitative analysis we compare the average percentage of success based on the human expert criterion Fig. 4. (a) Expected crop lines mapped by considering intrinsic and extrinsic parameters (yellow) and a margin of tolerance of 300; (b) corrected lines using Theil-Sen estimator (red) and Pearson product-moment correlation coefficient (blue). Table 1 Averaged percentages of success and processing times (ms) for WA, Pearson and Theil-Sen for different size reductions and margins of tolerance. Percentage of success Time (ms) Reduction Margin of tolerance WA Pearson Theil-Sen WA Pearson Theil-Sen 0% (original image) 50 100 70 70 78.2 83.1 86.1 93.1 105 105 9.3 9.5 5354 7506 150 70 72.9 94.1 105 9.9 9568 67% 50 70 80.2 85.2 11.3 8 • 10~2 92 100 70 83.1 91.3 11.3 3 . 4 - 1 0 - ' 476 150 70 79.3 84.6 11.3 7 . 5 - 1 0 - ' 1440 84% 50 70 81.1 83.8 3.3 8 . 0 - 1 0 - 4 30 100 70 82.3 83.1 3.3 2.0- 98 i o - 5 150 70 69.3 77.8 3.3 3.9 - IO"5 230 I GA + Otsu + THEIL-SEN Adjustment 16 8 • Marginof tolerance SO pixels —•—Marginof tolerance 100pixels —±— Margin of tolerance 150 pixels Reduction 094 5,375 7,527 9,589 Reduction 67% 0,09152 0,47826 1,44226 Reduction 8494 0,03086 0,09886 0,23066 Fig. 5. Averaged processing times for the full process: greenness extraction (GA), binarization (Otsu) and linear adjustment (Theil-Sen) for three size reductions and three margins of tolerance. for both Pearson and Theil-Sen and also without adjustment. Each image was visually analyzed by an expert to identify the accuracy between the adjust line and the real crop row considered as satis- factory by the expert. We also compute the computational cost measured in processing times for these three approaches. We call without adjustment (WA) the procedure involving ExG greenness extraction and Otsu based binarization, i.e. we only analyze the crop line detection obtained from a direct geometric mapping based on intrinsic and extrinsic parameters. For Pearson and Theil-Sen, times displayed are exclusively the ones obtained for the specific adjustment, i.e. to obtain the total time they must be added to the one displayed without adjustment. Because Theil-Sen is computational expensive but effective, we have reduced the resolution of the original image by down-sam- pling until to achieve values of 67% and 84% to verified the perfor- mances with and without reductions. Table 1 displays averaged percentages and computational times expressed in milliseconds for the set of original images available (i.e. with the 0% of reduction) and also for these images when their sizes are reduced until the 67% and 84% respectively, in horizontal and vertical sizes. For each reduction we display averaged percentages and times for three values (50,100 and 150) in what we call margin of tolerance. This margin represents the horizontal width used to search pixels representing crop and weeds around the expected crop line ob- tained by applying geometrical constraints (extrinsic and intrinsic parameters). Under this consideration all values for WA do not vary because here no search is required. From Table 1 we can see that the best performance, in terms of accuracy, is achieved by the Theil-Sen approach as compared to WA and Pearson, because it obtains the best results. The best abso- lute performance of Theil-Sen, also in terms of accuracy, is ob- tained for the original images with a margin of 150 pixels. This is because these images contain the maximum information and this margin covers the necessary range to capture all information com- ing from crops being unaffected by weeds pixels. As the reduction increases the accuracy decreases, except for a reduction of 67% with margin of 100 pixels. The decreasing can be explained by the fact that the greater the reduction and tolerance, more pixels belonging to weeds are involved in the line estimation. The excep- tion to this general behavior occurs for a reduction of 67% and a margin of 100. We have tested different combinations of margins and reduction without apparent improvements with respect to the ones displayed. Regarding times, it is obvious that the greater the image sizes and margins of tolerance the greater are processing times. This is because a greater number of pixels need to be processed. The main drawback for Theil-Sen is its high computational cost. Assuming the vision system captures an area of 4 m long the robot will need to navigate at speeds below of 4/I(m/s) to gain time for image pro- cessing and actuation. T represents the processing time. So, in the worst case T= 9.568 s and in the best case T= 0.476 s. The corre- sponding speeds are 1.5km/h and 30.25 km/h, where the first one is acceptable for agricultural tasks and the second one is exces- sive. This means that it is unnecessary reductions of 67%, i.e. reduc- tions between 0 and 67 should be appropriate depending on the agricultural treatments. For clarity, Fig. 5 displays averaged processing times, in seconds, over the set of images available for the full process including greenness extraction based on ExG, binarization through the Otsu's method and line adjustment based on the Theil-Sen estimator. As above, these values are obtained for the three values of reduction (0%, 67% and 84%) and with the three margins of tolerance in pixels (50, 100 and 150). As expected, processing times decrease as the reduction and margin of tolerance also decrease. 4. Conclusions We propose a method for accuracy crop row detection in maize fields. The image is segmented to transform the original color im- age into a gray scale. Then a binarization process is carried out based on the Otsu's method. This allows identifying green plants which are white pixels in the binary image. According to the intrin- sic and extrinsic parameters and applying perspective projection, we trace the expected crop lines over the original image. Consider- ing that white pixels in the binary image are crop and weeds in the original one we follow the expected crop lines and explore in the horizontal direction to capture white pixels in the binary image. These pixels are the ones used for estimating a new crop line that can coincide with the expected one or not. If the new line does not match with the expected one, a correction is made, but if it matches the correct location is verified. The estimation is carried out by applying Theil-Sen and also a linear regression based on the Pearson product-moment correlation coefficient. We have verified Theil-Sen outperforms Pearson product-moment based on qualitative and quantitative analysis, in terms of accuracy, and it is acceptable from the point of view of the processing time. Future improvements could be considered when high weed pressure is present in the image and a great number of weed patches invade the inter-row spacing. Acknowledgments The research leading to these results has been funded by the European Union's Seventh Framework Programme [FP7/2007- 2013] under Grant Agreement no. 245986 in the Theme NMP-2009-3.4-1 (automation and robotics for sustainable crop and forestry management). The authors wish also to acknowledge the project the AGL2011-30442-C02-02, supported by the Ministe- rio de Economía y Competitividad of Spain within the Plan Nacion- al of I+D+i. References Adichie, J. N. (1967). Estimates of regression parameters based on Rank tests. Annals of Mathematical Statistics, 38, 894-904. Astrand, B. (2005). Vision based perception or mechatronic weed control. Doctor of Philosophy Thesis, Chalmers and Halmstad Universities, Sweden. Astrand, B., & Baerveldt, A J. (2005). A vision based row-following system for agricultural field machinery. Mechatronics, 15, 251-269. Billingsley, J., & Schoenfisch, M. (1997). The succesful develepmnet of a vision guidance system for agriculture. Computers and Electronics in Agriculture, 16, 147-163. Bossu, J., Gée, Ch., Guillemin, J. P., & Truchetet, F. (2006). Development of methods based on double Hough transform and Gabor filtering to discriminate crop and weeds in agronomic images. In Proc. SPIE 18th annual symposium electronic imaging science and technology (Vol. 6070), paper no. 23, San Jose, USA. Bossu, J., Gée, Ch., Jones, G., & Truchetet, F. (2009). Wavelet transform to discriminate between crop and weed in perspective agronomic images. Computers and Electronics in Agriculture, 6 5 , 1 3 3 - 1 4 3 . Burgos-Artizzu, X. P., Ribeiro, A., Tellaeche, A., Pajares, G., & Fernández-Quintanilla, C. (2009). Improving weed pressure assessment using digital images from an experience-based reasoning approach. Computers and Electronics in Agriculture, 65,176-185. Davies, G., Casady, W., & Massey, R (1998). Precision agriculture: An introduction. Water Quality Focus Guide (WQ450, available on-line http:// extension.missouri.edu/explorepdf/envqual/wq0450.pdf). Dytham, C. (2011). Choosing and using statistics: A biologist's guide (3rd ed.). 9781405198394. John Wiley and Sons (p. 230). Fontaine, V., & Crowe, T. G. (2006). Development of line-detection algorithms for local positioning in densely seeded crops. Canadian Biosystems Engineering, 48(7), 19-29. Fu, K. S., González, R. C, & Lee, C. S. G. (1987). Robotics: Control, sensing, vision and intelligence. New York: McGraw-Hill. Gée, Ch., Bossu, J., Jones, G., & Truchetet, F. (2008). Crop/weed discrimination in perspective agronomic images. Computers and Electronics in Agriculture, 60, 49-59. Guerrero, J. M., Pajares, G., Montalvo, M., Romeo, J., & Guijarro, M. (2012). Support vector machines for crop/weeds identification in maize fields. Expert Systems with Applications, 39(12), 11149-11155. http://dx.doi.org/10.1016/ j.eswa.2012.03.040. Guijarro, M., Pajares, G., Riomoros, I., Herrera, P. J., Burgos-Artizzu, X. P., & Ribeiro, A. (2011). Automatic segmentation of relevant textures in agricultural images. Computers and Electronics in Agriculture, 75, 7 5 - 8 3 . Hague, T., Marchant, J. A., & Tillett, D. (1997). A system for plant scale husbandry. Precision Agriculture, 635-642. Hague, T., Tillet, N., & Wheeler, H. (2006). Automated crop and weed monitoring in widely spaced cereals. Precision Agriculture, 1(1), 9 5 - 1 1 3 . Hampel, F. R., Rousseeuw, P. J., Ronchetti, E., & Stahel, W. A. (1986). Robust statistics: The approach based on influence functions. New York: John Wiley. Hartley, R, & Zisserman, A. (2006). Multiple view geometry in computer vision. Oxford: Cambridge University Press. Hough, P. V. C. (1962). A method and means for recognizing complex patterns. U.S. Patent Office No. 3069654. Huber, P. J. (1981). Robust statistics. New York: John Wiley. Ji, R., & Qi, L. (2011). Crop-row detection algorithm based on random Hough transformation. Mathematical and Computer Modelling, 54,1016-1020. Jones, G., Gée, Ch., & Truchetet, F. (2009a). Modelling agronomic images for weed detection and comparison of crop/weed discrimination algorithm performance. Precision Agriculture, 10, 1-15. Jones, G., Gée, Ch., & Truchetet, F. (2009b). Assessment of an inter-row weed infestation rate on simulated agronomic images. Computers and Electronics in Agriculture, 67, 4 3 - 5 0 . http:// http://extension.missouri.edu/explorepdf/envqual/wq0450.pdf http://dx.doi.org/10.1016/ Kataoka, T., Kaneko, T., Okamoto, H., & Hata, S. (2003). Crop growth estimation system using machine vision. In The 2003 IEEE/ASME intemat. conference on advanced intelligent mechatronics. Kendall, M. G. (1955). Rank correlation methods (2nd ed.). London: Charles Griffin and Company. Kise, M., & Zhang, Q, (2008). Development of a stereovision sensing system for 3D crop row structure mapping and tractor guidance. Biosystems Engineering, 101, 191-198. Kise, M., Zhang, C¿., & Rovira-Más, F. (2005). A stereovision-based crop row detection method for tractor-automated guidance. Biosystems Engineering, 90(4), 357-367. Leemans, V., & Destain, M. F. (2006). Application of the Hough transform for seed row location using machine vision. Biosystems Engineering, 94(3), 325-336. López-Granados, F. (2011). Weed detection for site-specific weed management: mapping and real-time approaches. Weed Research, 5 1 , 1 - 1 1 . Marchant, J. (1996). Tracking of row structure in three crops using image analysis. Computers and Electronics in Agriculture, 15,161-179. Massart, D. L., Vandeginste, B. G. M., Buydens, L. M. C, De Jong, S., Lewi, P. J., & Smeyers-Verbeke, J. (1997). Single median method, In Handbook of chemometrics and qualimetrics: Part. Data handling in science and technology (vol. 20, 355-356), Elsevier. Meyer, G. E., & Camargo-Neto, J. (2008). Verification of color vegetation indices for automated crop imaging applications. Computers and Electronics in Agriculture, 63, 282-293. Meyer, G. E., Hindman, T. W., & Lakshmi, K. (1998). Machine vision detection parameters for plant species identification. Bellingham, WA: SPIE. Montalvo, M., Pajares, G., Guerrero, J. M., Romeo, J., Guijarro, M., Ribeiro, A., et al. (2012). Automatic detection of crop rows in maize fields with high weeds pressure. Expert Systems with Applications, 39(15), 11889-11897. Mood, A. M. (1950). Introduction to the theory of statistics. New York: McGraw-Hill Book Company. Neto, J. C. (2004). A combined statistical-soft computing approach for classification and mapping weed species in minimum tillage systems. Lincoln, NE: University of Nebraska. Olsen, H.J. (1995). Determination of row position in small-grain crops by analysis of video images. Computers and Electronics in Agriculture, 12,147-162. Onyango, C. M., & Marchant, J. A. (2003). Segmentation of row crop plants from weeds using colour and morphology. Computers and Electronics in Agriculture, 39,141-155. Otsu, N. (1979). A threshold selection method from gray-level histogram. IEEE Transactions on Systems Man and Cybernetics, 9, 62-66. Pla, F., Sanchiz, J. M., Marchant, J. A., & Brivot, R. (1997). Building perspective models to guide a row crop navigation vehicle. Image and Vision Computing, 15, 465-473. Ribeiro, A., Fernández-Quintanilla, C, Barroso, J., & García-Alegre, M. C. (2005). Development of an image analysis system for estimation of weed. In Proceedings of the 5th european conference on precision agriculture (5ECPA) (pp. 169-174). Romeo, J., Pajares, G., Montalvo, M., Guerrero, J. M., Guijarro, M., & Ribeiro, A. (2012). Crop row detection in maize fields inspired on the human visual perception. The Scientific World Journal, 2012. http://dx.doi.org/10.1100/2012/484390. Article ID 484390, 10 pages. Rousseeuw, P. J. (1984). Least median of squares regression. Journal of American Statistical Association, 79, 871-880. Rousseeuw, P. J., & Leroy, A M. (1987). Robust regression and outlier detection. New York: John Wiley. Rovira-Más, F., Zhang, Q,, & Reid, J. F. (2008). Stereo vision three-dimensional terrain maps for precision agriculture. Computers and Electronics in Agriculture, 60, 133-143. Rovira-Más, F., Zhang, 0_., Reid, J. F., & Will, J. D. (2005). Hough-transform-based vision algorithm for crop row detection of an automated agricultural vehicle. Journal of Automobile Engineering Part D, 219, 999-1010. Sainz-Costa, N., Ribeiro, A., Burgos-Artizzu, X., Guijarro, M., & Pajares, G. (2011). Mapping wide row crops with video sequences acquired from a tractor moving at treatment speed. Sensors, 11, 7095-7109. Sen, P. K. (1968). Estimates of the regression coefficient based on Kendall's Tau. Journal of American Statistical Association, 63,1379-1389. Sezgin, M. B., & Sankur, B. (2004). Survey over image thresholding techniques and quantitative performance evaluation. Journal of Electronic Imaging, 13(1), 146-165. Slaughter, D. C, Giles, D. K., & Downey, D. (2008). Autonomous robotic weed control systems: A review. Computers and Electronics in Agriculture, 61, 6 3 - 7 8 . Sogaard, H. T., & Olsen, H. J. (2003). Determination of crop rows by image analysis without segmentation. Computers and Electronics in Agriculture, 38, 141-158. Stewart, C. V. (1999). Robust parameter estimation in computer vision. SIAM Review, 41(3), 513-537. Tellaeche, A., Burgos-Artizzu, X. P., Pajares, G., & Ribeiro, A. (2008). A vision-based method for weeds identification through the Bayesian decision theory. Pattern Recognition, 41, 521-530. Tellaeche, A., Burgos-Artizzu, X., Pajares, G., Ribeiro, A., & Fernández-Quintanilla, C. (2008). A new vision-based approach to differential spraying in precision agriculture. Computers and Electronics in Agriculture, 60(2), 144-155. Tellaeche, A., Pajares, G., Burgos-Artizzu, X. P., & Ribeiro, A. (2011). A computer vision approach for weeds identification through support vector machines. Applied Soft Computing, 11, 908-915. Theil, H. (1950). A rank-invariant method of linear and polynomial regression analysis. In I, II, and III, Nederl. Akad. Wetensch. Proc, 5 3 , 3 8 6 - 3 9 2 , 5 2 1 - 5 2 5 and 1397-1412. Vioix, J. B., Douzals, J. P., Truchetet, F., Assemat, L., & Guillemin, J. P. (2002). Spatial and spectral method for weeds detection and localization. EURASIP JASP, 7, 679-685. Woebbecke, D. M., Meyer, G. E., von Bargen, K., & Mortensen, D. A. (1995). Shape features for identifying young weeds using image analysis. Transactions on American Society of Agricultural Engineering, 38(1), 271-281. http://dx.doi.org/10.1100/2012/484390 
work_7zcibx6kbzd6lbndc76iowoxnm	----	Noname manuscript No. (will be inserted by the editor) A probabilistic interpretation of the medical expert system CADIAG-2 David Picado Muiño Received: date / Accepted: date Abstract CADIAG-2 is a well known expert system aimed at providing support for medical diagnose in the field of internal medicine. CADIAG-2 consists of a knowl- edge base in the form of a set of IF-THEN rules that relate distinct medical entities, in this paper interpreted as conditional probabilistic statements, and an inference engine constructed upon methods of fuzzy set theory. The aim underlying this paper is the understanding of the logical structure of the inference in CADIAG-2. To that purpose we provide a (probabilistic) logical for- malisation of the inference of the system and check its adequacy with probabilistic logic. Keywords Knowledge-based systems · rule-based expert systems · fuzzy expert systems · probabilistic inference · CADIAG-2 1 Introduction CADIAG-2 (Computer Assisted DIAGnosis) is a well known rule-based expert system that aims at pro- viding support in diagnostic decision making in the field of internal medicine. Its design and construction was initiated in the early 80’s at the Medical University of Vienna by K.P. Adlassnig –see (Adlassnig et al., 1986), (Adlassnig et al., 1985), (Adlassnig, 1986) or (Leitich et al., 2002) for more on the origins and de- sign of CADIAG-2–. CADIAG-2 consists of two fundamental pieces: the inference engine and the knowledge base. The inference Institut für Diskrete Mathematik und Geometrie Wiedner Hauptstrasse 8 1040 Vienna, Austria Tel.: +43-1-58801-10427 Fax: +43-1-58801-910427 E-mail: david.muino@tuwien.ac.at engine is based on methods of approximate reasoning in fuzzy set theory, in the sense of (Zadeh, 1965) and (Zadeh, 1975). In fact CADIAG-2 is presented in some monographs as an example of a fuzzy expert system, for example in (Klir and Folger, 1988) or (Zimmer- mann, 1991). The knowledge base consists of a set of IF-THEN rules, also known as production rules in the literature, intended to represent relationships between distinct medical entities: symptoms, findings, signs and test results on the one hand and diseases and thera- pies on the other. The number of rules in the knowl- edge base of CADIAG-2 is approximately 40.000. The vast majority of them are binary (i.e., they relate single medical entities) and only such rules are considered in this paper. The rules in CADIAG-2 are defined along with a certain degree of confirmation which intuitively expresses the degree to which the antecedent confirms the consequent. For example, IF suspicion of liver metastases by palpation THEN pancreatic cancer with degree of confirmation 0.3 We can identify such degrees of confirmation with probabilities and the rules themselves with conditional probabilistic statements. In (Adlassnig, 1986) it is stated that degrees of confirmation can be interpreted as fre- quencies and, more generally, as probabilities. An in- terpretation in terms of degrees of belief of the doctor (or doctors) on the truth of the consequent given that the antecedent of the rule holds is also possible. This fact motivates a probabilistic interpretation of the infer- ence in CADIAG-2 and leads to the primary aim of this paper: formalise the inference in CADIAG-2 on proba- bilistic grounds and check its adequacy with probabil- ity logic –see (Halpern, 2003)– or, more generally, with probability theory. We shall not expect big surprises in 2 this respect. The inference mechanism in CADIAG-2 proceeds in a compositional way and thus it is bound to be probabilistically unsound (as will be clarified later). This was soon observed in earlier studies concerning the celebrated expert system MYCIN –see (Buchanan and Shortliffe, 1984) or (Shortliffe, 1976) for a description of MYCIN and (Hajek, 1988), (Hajek, 1989), (Hajek and Valdés, 1994), (Heckerman, 1986), (Valdés, 1992) for probabilistic approaches to it–. How far the inference mechanism of CADIAG-2 is from probabilistic sound- ness remains to be seen though. It is worth mentioning here that, although the inter- est among theoretical AI researchers in rule-based ex- pert systems seems to be lesser today than some years ago, rule-based expert systems are very popular among AI engineers. Many CADIAG-2-like systems are in use and more are being built for future implementation. It is mainly for this reason that we believe that further analysis and understanding of CADIAG-2-like systems is of relevance. This paper is in some way a continuation of (Cia- battoni and Vetterlein, 2009). In (Ciabattoni and Vet- terlein, 2009) the inference in CADIAG-2 is formalised by means of a logical calculus, CadL, and compared to t-norm-based formalisms –see (Hajek, 1998)–. It is shown that CadL does not respond to any t-norm-based (or to any fragment of a t-norm-based) logic. As far as we know (Ciabattoni and Vetterlein, 2009) consti- tutes the first attempt at formalising and understand- ing CADIAG-2 in a logical way. The present paper is the second. This paper is structured as follows. In Section 2 we give some basic definitions and introduce most of the notation used later in the other sections. In Section 3 the inference process in CADIAG-2 is described when restricted to the binary rules of its knowledge base. In Section 4 the formal system CadPL is defined and anal- ysed in the light of probability logic. CadPL is a for- malisation of the inference mechanism of CADIAG-2 based on a probabilistic interpretation of it. 2 PRELIMINARY DEFINITIONS AND NOTATION Throughout we will be working with a finite propo- sitional language L = {p1, ...,pn}, for some n ∈ N. We will denote by SL its closure under classical connec- tives. Within the context of CADIAG-2 the language L will represent the set of medical entities occurring in the system. By S ⊂ L we will denote the set of symptoms, findings, signs and test results (symptoms for short) and by D ⊂ L the set of diseases and therapies (diseases for short). Let us set LLit = {p ,¬p | p ∈ L} ⊂ SL, the set of literals of the language L. Consider ∆ = {φ1, ...,φk} ⊆ SL, for some k ∈ N. We will denote by ∧ ∆ the sentence φ1 ∧ ...∧φk. Definition 1 Let ω : SL −→ [0, 1]. We say that ω is a probability function on L if the following two conditions hold, for all θ,φ ∈ SL: – If |= θ then ω(θ) = 1. – If |= ¬(θ ∧φ) then ω(θ ∨φ) = ω(θ) + ω(φ).1 We define probability on conditional events in the conventional way. For ω a probability function on L and φ,θ ∈ SL, ω(φ|θ) = ω(φ∧θ) ω(θ) . Notice that if ω(θ) = 0 then ω(φ|θ) is not defined. We are interested in knowledge bases (CADIAG-2- like knowledge bases) that consist of finite collections of IF-THEN rules. We will identify such rules with triples of the form 〈θ,φ,η〉, where θ ∈ SL is the antecedent (or, more in keeping with probabilistic terminology, the conditioning event), φ ∈ SL is the consequent (or the uncertain event) and η ∈ [0, 1] is the degree to which θ confirms φ, which we will interpret as the conditional probability of φ given θ. Let K = {〈θ,φ,η〉 | θ,φ ∈ SL, η ∈ [0, 1]} be the set of all such triples in the language L. For the next two definitions let Φ ⊂K. Definition 2 We say that a probability function ω on L (probabilistically) satisfies Φ if, for all 〈θ,φ,η〉 ∈ Φ, we have that ω(θ) > 0 and ω(φ|θ) = η. Definition 3 We say that Φ is satisfiable (or consis- tent) if there exists a probability function ω on L that satisfies Φ.2 Next we define a partial ordering relation relevant in the context of the inference mechanism of CADIAG-2. Definition 4 Let � be the partial ordering relation on [0, 1] defined as follows: for a,b ∈ [0, 1], a � b if and only if 0 < a ≤ b or 0 ≤ a < 1 and b = 0. We define the strict partial ordering ≺ from � in the conventional way. As we will see later, the definition of the ordering � responds to the use of both 0 and 1 as maximal values in 1 Here (and throughout) |= represents classical entailment. 2 We will be using the terms ’satisfiable’ and ’consistent’ indis- tinguishably. 3 CADIAG-2. The value 0 denotes certainty in the non- occurrence of an event or falsity of a statement and the value 1 denotes certainty in its occurrence or its truth. For the next definition let D = [0, 1] × [0, 1] −{(0, 1), (1, 0)}. Definition 5 The function max∗ : D −→ R is defined as follows, for (a,b) ∈ D: max∗(a,b) = { a if b ≺ a b otherwise The definition of max∗ is extended to more than two arguments in its trivial way. 3 THE INFERENCE IN CADIAG-2 In this section we briefly describe the inference mechanism in CADIAG-2 when restricted to the set of binary rules, which from now we will denote by ΦCB. A more detailed description can be found in (Ciabattoni and Vetterlein, 2009). We have three different types of rules in ΦCB accord- ing to the typology defined in (Ciabattoni and Vetter- lein, 2009): – Type confirming to the degree d (c). A rule of this type is formalised in our settings as 〈φ,p,η〉 for p ∈ L, φ ∈ LLit and η ∈ (0, 1]. The one that follows is an example of a rule of this type in ΦCB: IF suspicion of pancreatic tumor by comput- erized tomography THEN pancreatic cancer with degree of confirmation 0.85 – Type mutually exclusive (me). A rule of type me is formalised as 〈q,¬p, 1〉, for p,q ∈ L. We give an example of a rule of this type: IF positive rheumatoid factor THEN not seronegative rheumatoid arthritis with degree of confirmation 1 – Type always occurring (ao). A rule of type ao is formalised as 〈¬q,¬p, 1〉, for p,q ∈ L. What follows is an example of a rule of this type in ΦCB: IF not chorea minor THEN not polymyositis with degree of confirmation 1 The types me and ao are classical in the sense that the degree of confirmation for the rules of these types is 1 and that the antecedent of such rules (or evidence in our settings) needs to be confirmed (it needs to be given value 1 by the system, see below) in order for these rules to be triggered by the inference engine. We will denote by Φc, Φme and Φao the subsets of ΦCB containing the rules of types c, me and ao respec- tively. The inference engine in CADIAG-2 gets started with a set of symptoms and diseases occurring in ΦCB present in the patient. Let Γ ⊂ L be the set of such medical entities. CADIAG-2 starts with an assignment v0 on Γ that gives a value in the interval [0, 1] to each entity in Γ. The values assigned by v0 are, according to most of the literature on CADIAG-2, intended to stand for mem- bership degrees in the context of fuzzy set theory and represent the degrees to which such entities are present in the patient –see (Adlassnig et al., 1986),(Adlassnig et al., 1985) or (Adlassnig, 1986) for more on the in- tended interpretation of v0–. However, other interpre- tations could also be suitable, at least to some extent. In fact, when defining the system CadPL in the next section, the interpretation to which we will commit will be probabilistic. The assignment v0 is defined for the negation of the entities in Γ and logical equivalents according to the following rule: If v0(θ) = η then v0(¬θ) = 1−η, for θ ∈ SL and η ∈ [0, 1].3 After the initial assignment the inference rules in ΦCB come into play. All the rules triggered by the sen- tences in Γ and its negations are used during the in- ference process. At each step in the inference process a rule is applied (that is done, in principle, in no partic- ular order). At the first step in the inference a rule of the form 〈θ,φ,η〉 in ΦCB is triggered, with η ∈ [0, 1] and θ,φ ∈ LLit. In order for that to happen θ or its negation needs to be in Γ and the value v0(θ) has to be strictly posi- tive if 〈θ,φ,η〉 is a rule in Φc and equal to 1 if the rule is in Φme or Φao. The application of the rule 〈θ,φ,η〉 generates a new assignment, v1, on {φ}. The value as- signed to φ by it is calculated as the minimum between η and v0(θ) and the value assigned to ¬φ and logical equivalents (if necessary for the inference) is calculated from v1(φ) as mentioned above for v0. At the nth step in the inference process a new rule of the form 〈θ,φ,η〉 in ΦCB will be triggered, for η ∈ [0, 1] and θ,φ ∈ LLit. At the nth step the application of this new rule will generate a new assignment on {φ} that will give φ the minimum between η and the value of θ considered for triggering the rule at this step in the inference. The value of θ needs to be strictly positive if the rule is of the type c and equal to 1 if it is of one 3 Notice that such a rule is compatible with a probabilistic interpretation of v0. 4 of the types me or ao. If the strictly positive values generated for θ before the nth step are multiple then the inference mechanism in CADIAG-2 will call the rule 〈θ,φ,η〉 again in further steps, if it has not done so previously, until all the values for θ have been used and all the possible values for φ generated. The assignment vn is defined for ¬φ as mentioned above for v0. The inference process goes on until all the rules trig- gered by all the sentences in Γ and its negations have been used and all the possible assignments for the sen- tences involved in the inference have been generated. CADIAG-2 yields as an outcome of the inference the set of medical entities in L occurring in the rules trig- gered by the initial evidence in Γ along with the max- imal value (with respect to the ordering � defined in Section 2) assigned to them during the inference. If a sentence is assigned both value 0 and 1 along the infer- ence process the system generates an error message. It is worth mentioning that the original inference process in CADIAG-2 works in a slightly different way. The update in the value of the distinct sentences in- volved in the inference is done as soon as two different values for the same sentence are produced by the sys- tem. The value chosen in the update for atomic sen- tences in L is the maximal one (with respect to the or- dering �). Notice though that this feature has a highly undesirable result (unless further restrictions on the rules or on the order in which the rules are applied are imposed), which is that the outcome of a run of the inference mechanism can depend on the order in which the rules are applied. Such a drawback is easily avoided by assuming that the update is only done at the end of the process. There are other several undesirable fea- tures about the inference in CADIAG-2, most of them related to the maximal value 0 and negated proposi- tions. Maybe the most evident concerning the maximal value 0 is that a medical entity that at some step along the inference process is assigned value 0 (that is to say, it is considered false with certainty or impossible) trig- gers any rules in which it occurs as the conditioning event if any other value other than 0 is assigned to it along the inference process. For a deeper analysis of such aspects of the inference process in CADIAG-2 see (Ciabattoni and Vetterlein, 2009). We represent sentences together with the assign- ments generated for them at each step in the inference by pairs in SL× [0, 1] along with a subscript indicating the step in the process at which such pairs have been generated. As mentioned above, a step in the inference process is given by the application of a rule in ΦCB and the new assignments that it generates for the sentences involved in the rule. Let p ∈ L and η ∈ [0, 1] be the highest assignment to p in a run of the inference mechanism in CADIAG-2. We will use the subscript max∗ on the pair (p,η) –that is to say, (p,η)max∗ – to denote that η is the maximal value assigned during the inference process for p (with respect to the ordering �). 4 THE FORMAL SYSTEM CADP L As seen in the previous section, the inference mech- anism in CADIAG-2 gets started by assigning to the ini- tial symptoms and diseases present in the patient (Γ) values in the interval [0, 1] (the initial assignment v0 to the entities in Γ and their corresponding negations). As mentioned earlier, such values stand in principle for fuzzy membership within the context of fuzzy set theory and are motivated by the vague nature of some medical entities that occur in CADIAG-2. In this paper though the interpretation that we will give to the assignment v0 will be probabilistic. Let us consider the medical entity ’reduced glucose in serum’. Let us assume that the value assigned by the evaluation system in CADIAG-2 (i.e., v0) to the state- ment ’Patient A has reduced glucose in serum’ out of the evidence given by the corresponding measurement of the amount of glucose in Patient A is η, for some η ∈ [0, 1]. As an example, we could interpret such value as the degree of belief that a medical doctor has in the truth of the statement given the evidence. As such η could be interpreted as a probability. The probabilistic interpretation is certainly favoured by the discretization applied to medical concepts in CADIAG-2 (for example, the concept ’glucose in serum’ generates five distinct medical entities in CADIAG-2: ’highly reduced glucose in serum’, ’reduced glucose in serum’, ’normal glucose in serum’, ’elevated glucose in serum’ and ’highly el- evated glucose in serum’). Notice that such an inter- pretation places us within the subjective probabilistic frame and thus, for the sake of coherence, the knowledge base ΦCB should also be interpreted subjectively. Other interpretations are also possible though. For example, one could regard such values as the ratio given by the number of doctors that agree on the truth of the state- ment out of all the doctors involved in the assessment. In order to accommodate such values into a coherent probabilistic frame along with the statements in ΦCB one could justify them as being subjective probabilities assessed by a group of experts –see (Genest and Zidek, 1986) or (Osherson and Vardi, 2006) for an analysis and justification of such concept–. Let p ∈ L represent a medical entity present in the patient and assume that η ∈ [0, 1] is the initial value 5 assigned to it by v0 at the start of a run of the infer- ence mechanism in CADIAG-2. We can formalise this by means of a probabilistic conditional statement rep- resented by a triple of the form 〈κ,p,η〉, where κ ∈ SL is the evidence that supports the presence of p in the patient. Next we will define the formal system CadPL. Re- call that the ultimate goal when doing so is to de- fine a system which represents the inference process in CADIAG-2 when interpreted from a probabilistic point of view. Although the inference in CADIAG-2 can be closely related to probability theory (given the nature of the rules of inference in ΦCB) it is not based on prob- abilistic methods and so the degree of freedom when choosing the rules of the system CadPL is high. We have chosen the rules by interpreting in the most nat- ural way the steps along the inference process within a probabilistic frame. The main idea behind such in- terpretation consists of the identification of the infer- ence process with the propagation of evidence facili- tated by the rules in ΦCB. For example, from 〈κ,φ,η〉, where κ ∈ SL is evidence supporting the presence of φ in the patient, and 〈φ,θ,ζ〉 in ΦCB we would infer 〈κ,θ, min(η,ζ)〉, where min(η,ζ) is the value (probabil- ity) assigned to θ given the evidence κ. We would have a propagation process of this nature for each single piece of evidence. The evidence would then be brought to- gether in CADIAG-2 by what we call the right conjunc- tion rule: given two outcomes of the inference process, say 〈κ1,p,η〉 and 〈κ2,p,ζ〉, for p ∈ L and κ1,κ2 ∈ SL, CADIAG-2 combines the evidence given by κ1 and κ2 by computing 〈κ1 ∧ κ2,p, max∗(η,ζ)〉. The inference rules of CadPL that we next present formalise this in- terpretation. A theory T in CadPL is a finite subset of K. A the- ory in our settings is not necessarily assumed to be con- sistent. In fact, ΦCB is not consistent as a probabilistic knowledge base, it contains a large number of conflicts (i.e., minimal inconsistent subsets) –see our compan- ion paper (Klinov et al., 2010) for more on the conflicts in ΦCB–. However, the inference engine in CADIAG-2 does not explode in the presence of such conflicts. For what follows let T = Ω ∪Φ1 ∪Φ2 be a theory, with Ω = {〈κ1,φ1,η1〉, ...,〈κm,φm,ηm〉}, for some m ∈ N, and Φ2 a set of triples of the form 〈θ,φ, 1〉, for θ,φ ∈ SL. Let CΩ = {κ1, ...,κm} and Γ = {φ1, ...,φm}. We will assume CΩ to be a consistent set (i.e., there exists a classical valuation that satisfies CΩ). Within the context of CADIAG-2, Γ is intended to represent the set of medical entities present in the pa- tient and Ω the initial assignment (v0) in the inference process given CΩ, the evidence in support of the pres- ence in the patient of the corresponding medical entities in Γ . The set Φ1 represents the rules of type c and Φ2 the rules of types me and ao. The formal system CadPL is defined by the follow- ing inference rules: – Reflexivity rule For φ ∈ SL, κ ∈CΩ and η ∈ [0, 1], 〈κ,φ,η〉 ∈ Ω T ` 〈κ,φ,η〉 – Negation rule For φ,ψ ∈ SL and η ∈ [0, 1], T ` 〈ψ,φ,η〉 T ` 〈ψ,¬φ, 1 −η〉 – Equivalence rule For ψ,φ,θ ∈ SL and η ∈ [0, 1], ψ ≡ φ T ` 〈θ,φ,η〉 T ` 〈θ,ψ,η〉 – Minimum rules For θ,φ ∈ SL, κ ∈CΩ, η ∈ (0, 1] and ζ ∈ [0, 1], First rule: T ` 〈κ,θ,η〉 〈θ,φ,ζ〉 ∈ Φ1 T ` 〈κ,φ, min(η,ζ)〉 Second rule: T ` 〈κ,θ, 1〉 〈θ,φ, 1〉 ∈ Φ2 T ` 〈κ,φ, 1〉 – Right conjunction rule For φ ∈ SL, C1,C2 ⊆CΩ and η,ζ ∈ [0, 1], T ` 〈 ∧ C1,φ,η〉 T ` 〈 ∧ C2,φ,ζ〉 T ` 〈 ∧ {C1 ∪C2},φ, max∗(η,ζ)〉 – Exhaustivity rule For φ ∈ SL, κ ∈CΩ, C ⊆CΩ and η ∈ [0, 1], T ` 〈 ∧ C,φ,η〉 ∀ζ ∈ [0, 1] T 0 〈κ,φ,ζ〉 T ` 〈κ∧ ∧ C,φ,η〉 Notice that the exhaustivity rule does not have any bearing on the decidability of whether 〈κ,φ,ζ〉 is prov- able from T or not for ζ ∈ [0, 1], φ ∈ SL and κ ∈ CΩ. The exhaustivity rule can only be applied after its prov- ability or non-provability has been decided. Given a theory T of CadPL and a triple Θ ∈ K, a proof of Θ from T is defined as a finite sequence of sequents of the form T ` Θ1, ...,T ` Θn 6 with Θn = Θ and where, for i ∈ {1, ...,n}, each Θi in T ` Θi follows from T by the application of one of the rules above, from Θj in a previous sequent (with j < i) or from Θj,Θk in previous sequents (with j,k < i) by one of the rules above. For what follows let Θ be the triple 〈θ,φ,η〉, for some η ∈ [0, 1] and θ,φ ∈ SL. Definition 6 We say that there exists a maximal proof of Θ from T if there exists a proof of Θ from T and there is no proof from T of 〈θ,φ,ζ〉 with η ≺ ζ. Definition 7 We say that Θ follows maximally from T (denoted by T `CadPL Θ) if there exists a maximal proof of Θ from T . For the next proposition let T = Ω ∪Φ1 ∪Φ2, with Ω = {〈κ1,q1,η1〉, ...,〈κm,qm,ηm〉}, Φ1 = Φ c and Φ2 = Φ me ∪ Φao, CΩ = {κ1, ...,κm} ⊂ SL and Γ = {q1, ...,qm} ⊂ L a set of medical entities occurring in ΦCB. Proposition 1 Let p ∈ L and η ∈ [0, 1]. We have that T `CadPL 〈 ∧ CΩ,p,η〉 if and only if (p,η)max∗ is the outcome of a run of the inference engine of CADIAG-2 on T . Proof 4 In order to prove the left implication let us consider a run of the inference engine of CADIAG-2 on T . The inference starts from pairs of the form (q,η)0 and (¬q, 1 − η)0 for some η ∈ [0, 1] for all q ∈ Γ . In CadPL a pair of the form (q,η)0, for q ∈ Γ , corresponds to a sequent of the form T ` 〈κ,q,η〉, for κ ∈ CΩ. The pair (¬q, 1 − η)0 corresponds to the sequent T ` 〈κ,¬q, 1−η〉. The former corresponds to an application of the reflexivity rule. The latter follows from the first one by an application of the negation rule. Let us assume now that we are at the nth step of the inference process and that a rule of the form 〈θ,φ,ζ〉 is triggered, for some ζ ∈ [0, 1] and θ,φ ∈ SL. Let us suppose that we have (θ,µ)n−t, the pair that triggers the rule at the nth step of the process, for µ ∈ (0, 1] and t ≤ n − 1. In CadPL that would cor- respond to a sequent of the form T ` 〈κ,θ,µ〉 de- rived from a previous step in the inference, for κ ∈CΩ. The inference mechanism in CADIAG-2 produces the pairs (φ, min(ζ,µ))n and (¬φ, 1−min(ζ,µ))n which, in the formal system CadPL, corresponds to the sequents 4 For the sake of brevity we will deal with sentences as if they were equivalence classes. If anything applies to a sentence of the form ¬φ, with φ ∈ SL, we also assume that it applies to any logical equivalent of ¬φ without mentioning it. T ` 〈κ,φ, min(ζ,µ)〉 and T ` 〈κ,¬φ, 1 − min(ζ,µ)〉 re- spectively, which follow by an application of one of the minimum rules and, for the latter, an application of the negation rule on the former. At the end of the process CADIAG-2 generates the pair (p,η)max∗ for each sentence p ∈ L involved in the inference, where η is the maximal value (with respect to the ordering �) among those assigned to p along the inference. This maximisation process is achieved in CadPL by means of repeated applications of the right conjunction rule. Instances of the exhaustivity rule (if necessary) complete the inferential counterpart of CADIAG-2 in CadPL. In order to prove the right implication let us suppose that we have a maximal proof of the form T ` Θ1, ...,T ` Θm, where Θm is the triple 〈 ∧ CΩ,p,η〉, for some η ∈ [0, 1] and p ∈ L. The first sequent of the proof needs to respond to an instance of the reflexivity rule, T ` 〈κ,q,η〉, for some q ∈ Γ, κ ∈ CΩ and η ∈ [0, 1]. The corresponding coun- terpart of this sequent in CADIAG-2 is the pair (q,η)0. Let us move now to the nth sequent, with n ≤ m. The nth sequent can be an instance of the reflexivity rule, T ` 〈κ,q,η〉, for some q ∈ Γ, η ∈ [0, 1] and κ ∈CΩ. The counterpart for this sequent in CADIAG-2 is the pair (q,η)0. The nth sequent can follow from a previous one in the proof by an instance of the negation rule. Let us suppose that the nth sequent is T ` 〈θ,¬φ, 1 − η〉 for some η ∈ [0, 1] and θ,φ ∈ SL and that there is a sequent T ` Θi, for some i < n, of the form T ` 〈θ,φ,η〉. The latter corresponds to a pair of the form (φ,η)t in CADIAG-2 and the former to the pair (¬φ, 1 − η)t, where t is the step in the inference process at which such pairs have been generated. The nth sequent can follow from a previous one by an instance of one of the minimum rules. Let us assume that the nth sequent is T ` 〈κ,φ, min(η,ζ)〉, for some φ ∈ SL, κ ∈ CΩ, η ∈ [0, 1] and ζ ∈ (0, 1], that T ` 〈κ,ψ,ζ〉 is a previous sequent in the proof and that 〈ψ,φ,η〉 is a rule in ΦCB. The latter corresponds in CADIAG-2 to the pair (ψ,ζ)t and the former to the pair (φ, min(η,ζ))t+k, where t,t + k indicate the steps at which the pairs have been generated by the inference process. The nth sequent can follow from previous sequents by an application of the right conjunction rule. The counterpart in CADIAG-2 of such an outcome consists of the maximisation process at the end of the inference. Instances of the exhaustivity rule are irrelevant to the inference in CADIAG-2. 7 This completes the proof. It is worth commenting on some features of the in- ference rules of CadPL in connection with probability theory. Soundness with respect to probabilistic seman- tics of the reflexivity, negation, and equivalence rules is clear. Soundness of the second minimum rule is also clear. The first minimum rule is certainly not sound with respect to such semantics. However, the right con- junction rule is not sound and the exhaustivity rule assumes some probabilistic independence among sen- tences that may not actually be independent. Overall, CadPL does not score well within probability theory. This is no surprise. The computation of conditional probabilistic statements in a compositional way, as done by CADIAG-2 primarily by means of the min and max∗ operators, is clearly bound to be probabilistically un- sound. One may wonder though what could be done in order to improve the inference on probabilistic grounds from a knowledge base like ΦCB. The answer seems to be ’not much’. Certainly a ΦCB-like knowledge base (i.e., a knowledge base given by some binary probabilis- tic conditional statements) is not the most convenient for inferential purposes in probability theory for medi- cal applications like CADIAG-2. As is well known, there are other knowledge-base structures better suited for that purpose, Bayesian networks being the most cele- brated among them, see (Castillo et al., 1997) or (Pearl, 1988). In terms of consistency, it is worth noting that CadPL satisfies what we can call weak consistency –called weak soundness in (Hajek, 1988)–, defined as follows: if there is a maximal proof in CadPL of a statement of the form 〈 ∧ ∆,φ, 1〉 (or 〈 ∧ ∆φ, 0〉) from a certain theory T , with φ ∈ SL and ∆ ⊆ SL then, if there is a maximal proof in CadPL of a statement of the form 〈 ∧ ∆∗,φ,η〉, with ∆ ⊂ ∆∗, then η = 1 (or η = 0 respectively). That is to say, if CadPL concludes certainty about the oc- currence of some event or about the truth or falsity of some sentence then adding new evidence does not alter this certainty. Weak consistency is provided in CadPL and so in the inference mechanism of CADIAG-2 by the operator max∗ defined over the ordering �. It is also worth noting that one could guarantee con- sistency (i.e., satisfiability) by relaxing the interpreta- tion of ΦCB and consider η in each triple 〈θ,φ,η〉 ∈ ΦCB a lower-bound probability threshold rather than a point-valued probabilitiy and by restricting the system to a positive fragment of LLit (i.e., only one of p, ¬p can occur in ΦCB). This way consistency is trivially guar- anteed for ΦCB together with any outcomes produced by the system during the inference process. In terms of soundness there does not seem to be much that one can do in order to improve the inference mechanism for knowledge bases like ΦCB, or at least not much that one can do that does not come at the price of generating probabilistic statements with very low prob- abilistic bounds (when working with lower-bound prob- ability thresholds), which would make CADIAG-2 po- tentially useless for practical purposes. There is some room for improvement for some steps in the inference that come by the addition of independence assump- tions among some of the medical entities in ΦCB. Under such independence assumptions a product rule instead of the minimum rule (i.e., composition through the product operator instead of the minimum one) could yield soundness in the context of lower-bound proba- bility thresholds for some inference steps. 5 CONCLUSION CADIAG-2 is a reasonably well-performing medical expert system (Adlassnig et al., 1986), but how it is so is far from clear. The inference engine of CADIAG-2 was built with methods of approximate reasoning in fuzzy set theory but, as such, it was not based on any logical formalism or theory embedded with a clear semantics. This fact motivated the main aim of this paper, which was no other than the understanding of CADIAG-2 in a logical way. The natural interpretation of the inference rules of CADIAG-2 (i.e., probabilistic) placed us upon the at- tempt of interpreting the inference itself probabilisti- cally. We formalised this interpretation by means of the system CadPL, the logical (probabilistic) counterpart of the inference engine of CADIAG-2. The unsoundness of some of the rules of CadPL (and thus of some infer- ence steps in CADIAG-2) and the inconsistency of the calculus (and thus of the inference process in CADIAG- 2) was made clear. Apart from these drawbacks, other- wise expected, some other aspects of CadPL were also stressed and analysed. At the end of the paper some possibilities for an improvement of CADIAG-2 in terms of soundness and consistency were also mentioned. Acknowledgements This research has been supported by the Vienna Science and Technology Fund (WWTF), Grant MA07− 016. I would like to thank A. Ciabattoni, T. Vetterlein and P. Rusnok for useful comments on previous drafts of this paper and for technical support with CADIAG-2. REFERENCES [Adlassnig, 1986]Adlassnig, K. (1986). Fuzzy set theory in medical diagnosis. IEEE Transactions on Systems, Man and Cybernetics, 16(2):260–265. 8 [Adlassnig et al., 1985]Adlassnig, K., Kolarz, G., Effen- berger, W., and Grabner, H. (1985). Cadiag: Ap- proaches to computer-assisted medical diagnosis. Computers in Biology and Medicine, 15:315–335. [Adlassnig et al., 1986]Adlassnig, K., Kolarz, G., Schei- thauer, W., and Grabner, H. (1986). Approach to a hospital-based application of a medical expert system. Informatics for Health and Social Care, 11(3):205–223. [Buchanan and Shortliffe, 1984]Buchanan, B. and Short- liffe, E. (1984). Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Addison-Wesley, Reading, MA. [Castillo et al., 1997]Castillo, E., Gutiérrez, J., and Hadi, A. (1997). Expert Systems and Probabilistic Net- work Models. Springer-Verlag, New York. [Ciabattoni and Vetterlein, 2009]Ciabattoni, A. and Vet- terlein, T. (2009). On the fuzzy (logical) content of cadiag2. Fuzzy Sets and Systems (to appear shortly). [Genest and Zidek, 1986]Genest, C. and Zidek, J. (1986). Combining probability distributions: A critique and an annotated bibliography. Statistical Science, 1(1):114–135. [Hajek, 1988]Hajek, P. (1988). Towards a probabilistic analysis of mycin-like expert systems. In COMP- STAT 88 proceedings. Physica-Verlag Heidelberg. [Hajek, 1989]Hajek, P. (1989). Towards a probabilistic analysis of mycin-like expert systems 2. In Artifi- cial Intelligence and Information-Control Systems of Robots. North-Holland. [Hajek, 1998]Hajek, P. (1998). Metamathematics of Fuzzy Logic. Kluwer, Dordrecht. [Hajek and Valdés, 1994]Hajek, P. and Valdés, J. (1994). An analysis of mycin-like expert systems. Math- ware and Soft Computing, 1:45–68. [Halpern, 2003]Halpern, J. (2003). Reasoning About Un- certainty. MIT Press. [Heckerman, 1986]Heckerman, D. (1986). Probabilistic in- terpretations for mycin’s certainty factors. In Un- certainty in Artificial Intelligence, pages 167–196. North-Holland. [Klinov et al., 2010]Klinov, P., Parsia, B., and Picado- Muino, D. (2010). The consistency of the cadiag- 2 knowledge base: A probabilistic approach. In LPAR. [Klir and Folger, 1988]Klir, G. and Folger, T. (1988). Fuzzy Sets, Uncertainty and Information. Prentice-Hall International. [Leitich et al., 2002]Leitich, H., Adlassnig, K., and Ko- larz, G. (2002). Evaluation of two different mod- els of semiautomatic knowledge acquisition for the medical consultant system cadiag2/rheuma. Arti- ficial Intelligence in Medicine, 25:215–225. [Osherson and Vardi, 2006]Osherson, D. and Vardi, M. (2006). Aggregating disparate estimates of chance. Games and Economic Behavior, 56:148–173. [Pearl, 1988]Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems. Morgan Kaufman, San Mateo, California. [Shortliffe, 1976]Shortliffe, E. (1976). Computer-Based Medical Consultations: MYCIN. Elsevier, New York. [Valdés, 1992]Valdés, J. (1992). An integrated environ- ment for uncertainty processing using graphical models. In IPMU 92 proceedings. [Zadeh, 1965]Zadeh, L. (1965). Fuzzy sets. Information and Control, 8:338–353. [Zadeh, 1975]Zadeh, L. (1975). Fuzzy logic and approxi- mate reasoning. Synthese, 30:407–428. [Zimmermann, 1991]Zimmermann, H. (1991). Fuzzy Set Theory and its Applications. Kluwer Academic Publisher, Boston - Dordrecht - London. 
work_a56n6qovhvgphchpkspltuwrni	----	Microsoft Word - WP_Thorleuchter_2013_2.docx FACULTEIT ECONOMIE EN BEDRIJFSKUNDE   TWEEKERKENSTRAAT 2 B-9000 GENT Tel. : 32 - (0)9 – 264.34.61 Fax. : 32 - (0)9 – 264.35.92 WORKING PAPER   Quantitative Cross Impact Analysis with Latent Semantic Indexing Dirk Thorleuchter1 Dirk Van den Poel2 November 2013 2013/861 1 Corresponding author: Dr. Dirk Thorleuchter: Fraunhofer INT, Appelsgarten 2, 53879 Euskirchen, Germany. Tel.: +49 2251 18305; fax: +49 2251 18 38 305 E-mail address: Dirk.Thorleuchter@int.fraunhofer.de 2 Prof. Dr. Dirk Van den Poel, Professor of Marketing Analytics/Analytical Customer Relationship Management, Faculty of Economics and Business Administration, E-mail: dirk.vandenpoel@ugent.be; more papers are available from: www.crm.UGent.be D/2013/7012/32 Quantitative Cross Impact Analysis with Latent Semantic Indexing Dirk Thorleuchtera,*, Dirk Van den Poelb Abstract Cross impact analysis (CIA) consists of a set of related methodologies that predict the occurrence probability of a specific event and that also predict the conditional probability of a first event given a second event. The conditional probability can be interpreted as the impact of the second event on the first. Most of the CIA methodologies are qualitative that means the occurrence and conditional probabilities are calculated based on estimations of human experts. In recent years, an increased number of quantitative methodologies can be seen that use a large number of data from databases and the internet. Nearly 80% of all data available in the internet are textual information and thus, knowledge structure based approaches on textual information for calculating the conditional probabilities are proposed in literature. In contrast to related methodologies, this work proposes a new quantitative CIA methodology to predict the conditional probability based on the semantic structure of given textual information. Latent semantic indexing is used to identify the hidden semantic patterns standing behind an event and to calculate the impact of the patterns on other semantic textual patterns representing a different event. This enables to calculate the conditional probabilities semantically. A case study shows that this semantic approach can be used to predict the conditional probability of a technology on a different technology. Keywords: Cross Impact Analysis, Latent Semantic Indexing, Text Mining, Conditional Probability. Introduction In literature, cross impact analysis (CIA) is often used to predict the probability that a specific event occur (occurrence probability) as well as the impact of this event on different events (conditional probability) (Blanning & Reinig, 1999; Schuler, Thompson, Vertinsky, & Ziv, 1991). A large number of existing approaches are qualitative. They are based on estimations of human experts (Banuls, Turoff, & Hiltz, 2013; Mitchell, Tydeman, & Curnow, 1977). In recent years, the number of quantitative approaches has increased. This is because the large number of accessible information today makes it possible to use the results of automated data mining approaches instead of using the time- and cost expensive estimations by human experts (Kim, Lee, Seol, & Lee, 2011). Quantitative CIA approaches that are based on textual information are knowledge structure based because they apply multi-label text classification approaches based on well-known text similarity measures to identify the impact of one event on a different event (Thorleuchter, Van den Poel, & Prinzie, 2010). However, this is done by considering aspects of words and not by considering semantic aspects in textual information. _______________________________________________________________________ 2 An example for an event could be the appearance of a new technology in the technology landscape. The appearance of new technologies and the change of existing technologies over time from past to future is a well-known topic for futurists (Bell, 2002). This enables to predict future technological capabilities for decision-makers (Thorleuchter & Van den Poel, 2013d). The technological landscape is characterized by a large number of technologies that are impacted by a large number of other technologies (Yu, Hurley, Kliebenstein, & Orazem, 2012). Technologies impact other technologies in different ways e.g. in an integrative, substitutive, precursive, and successive way (Geschka, 1983). A short example for the substitutive way is given below: The electrical fuel cell technology used in an energy supply application can be substituted by electrical battery or solar cell technology. This is because all three technologies can be used to realize this application. They replace each other based on their advances. Thus, the three technologies impact each other in a substitutive way. Further, these impacts change very often because current results from technological research and development lead to new technological advances and to the appearance of new substitutive technologies as an ongoing process (Kauffman, Lobo, & Macready, 2000). As a result, using CIA for monitoring these complex technological impacts makes it necessary to use quantitative rather than qualitative approaches. Several texts that describe a single event are normally written in several writing styles by different persons. Further, these texts possibly are written in different contexts or in different languages. It is not necessary that two texts describing the same event contain even one common word (Thorleuchter & Van den Poel, 2013b). With semantic approaches the relationship between the two texts can be identified because they share a common meaning (Choi et al., 2012; Tsai, 2012). This is the reason why semantic text classification approaches often outperform knowledge structure based text classification approaches (Thorleuchter & Van den Poel, 2012c). In contrast to existing CIA approaches, we provide a quantitative CIA approach that considers the aspects of meaning in textual information. Latent semantic indexing (LSI) is a well-known representative for semantic approaches (Jiang, Berry, Donato, Ostrouchov, & Grady, 1999). It identifies the hidden meaning of textual information in documents considering occurrences and co-occurrences of terms (D’Haen, Van den Poel, & Thorleuchter, 2013; Luo, Chen, & Xiong, 2011). Both, terms and documents are mapped to a semantic structure that consists of several semantic textual patterns (Christidis, Mentzas, & Apostolou, 2012; Park, Kim, Choi, & Kim, 2012). The impact of terms and documents on the patterns is calculated (Kuhn, Ducasse, & Girba, 2007). A semantic textual pattern that represents e.g. a technology might contain terms and documents that also have an impact on a different semantic textual pattern representing e.g. another technology (Thorleuchter & Van den Poel, 2013c). This indicates a relationship between the technologies and based on this relationship, the cross-impact between technologies can be calculated. To extract semantic patterns from the large number of texts describing events, we use a rank- validation procedure that is taken over from literature (Thorleuchter & Van den Poel, 2013a). This procedure enables to identify a maximal number of semantic patterns where each pattern can be used to represent a specific event. The rank-validation procedure is successfully evaluated by using LSI with singular value decomposition (SVD). Beside LSI, modern semantic approaches exist that outperform LSI in several studies. Examples for these modern approaches are probabilistic latent semantic indexing (Hofmann, 1999), non-negative matrix _______________________________________________________________________ 3 factorization (Lee & Seung, 1999; Lee & Seung, 2001), and latent dirichlet allocation (Blei, Ng, & Jordan, 2003). However, literature has not validated the use of these modern approaches together with the rank-validation procedure until now. Additionally, the modern approaches are of higher computational complexity than LSI (Ramirez, Brena, Magatti, & Stella, 2012). Thus in this paper, LSI is used together with the rank-validation procedure because this combination is already successful evaluated and it is of good computational performance. In a case study, we predict the impact of technologies on different technologies. The used data are descriptions of research projects funded by the German Ministry of Defense (GE MoD) in 2007. These research projects deal with one or several technologies to create an application. Semantic textual patterns in the descriptions are extracted, the technologies standing behind the patterns are identified, and the cross-impacts between the technologies are calculated. This semantic approach is compared to a knowledge structure based approach that uses the same data for calculating the cross-impacts. Overall, we propose a quantitative methodology that combines semantic text classification with CIA. The use of a semantic approach for the CIA calculation is in contrast to related work. The semantic methodology calculates the conditional probabilities of events given different events quantitatively. This enables to depict the complex relationships between events with lower manual effort than qualitative approaches and by considering semantic aspects. Thus, it is helpful for decision makers. Background The proposed approach calculates conditional cross impact probabilities by use of semantic text classification. Below, we describe how conditional cross impact probabilities can be calculated and how quantitative text-based CIA is processed up to now. In 1968, CIA was proposed (Gordon & Haywood, 1968) to calculate the occurrence probabilities of an event and to calculate the conditional probabilities of one event given another. The approach is based on subjective estimations by human experts. The occurrence probability of an event A was simply defined as P(A) and calculated by the number of these human experts who predict the occurrence of A over the number of all human experts. The conditional probability of event B given event A was defined as P(B|A) and calculated by the number of experts who predict both, the occurrence of A and B over the number of all experts who predict the occurrence of A (Dalkey, 1972; Enzer, 1972). This approach was improved many times and nowadays, most of the new improved approaches focus on a more quantitative way to calculate the probabilities. Examples are the use of cumulative sale probabilities over time by (Caselles-Moncho, 1986) and the use of patent data (Choi, Kim, & Park, 2007). These quantitative approaches start with a multi-label data classification step where the data is assigned to different events (classes). Based on this assignment, the calculation of the probabilities is done in a second step. About 80% of all data available today are textual data. Thus, modern approaches use the large number of textual data e.g. available in the internet for CIA. Examples are the use of linguistic expressions in technology descriptions (Jeong & Kim, 1997) and the use of terms from technology taxonomies (Thorleuchter, Van den Poel, & Prinzie, 2010). From text classification point of view, these approaches are knowledge-based and they use instance- based learning algorithms where semantic aspects of the textual data are not considered. This is in contrast to the approach presented here where a new methodology is provided that uses a semantic approach (LSI) for calculating the conditional probabilities from texts. _______________________________________________________________________ 4 Methodology Fig. 1 shows the processing of the methodology in different steps. The methodology (see Fig. 1) starts with a data collection step. Events are defined and a set of documents are used as input. The documents should consist of textual information describing one or several events. As an example, the case study defines an event as a technology and thus, each document contains a description of a research project where one or several technologies occur. In a preprocessing step, specific elements (e.g. scripting code, tags, and images) are removed. The text is split in terms and each term is checked for typographical errors by use of a dictionary. The large number of different terms is reduced by applying term filtering methods e.g. stop word filtering, part-of-speech tagging, and stemming. Further, Zipf’s law (Zeng, Duan, Cao, & Wu, 2012; Zipf, 1949) is applied where many low frequent terms can be discarded. Each document is represented by a term vector based on vector space model. The size of a vector is based on the reduced number of terms (Thorleuchter & Van den Poel, 2012a). Vector components are represented by weighted frequencies as calculated in accordance to Salton et al. (1994). The frequency of the corresponding term in a specific document is multiplied by its inverse document frequency and it is divided by a length normalization factor. _______________________________________________________________________ 5 The term vectors are used to create a term-by-document matrix with rank r. The rank of the matrix is reduced from r to k by LSI. For the selection of on optimal value of k, a rank- validation procedure as introduced in Sect. 2.3 is applied: for each value of k, LSI is applied and the resulting k dimensions are compared to the descriptions of the events. In the case study, this step is done manually by human experts; however, it could be realized automatically by applying a text similarity measure on patterns and events. As a result of this step, the number of semantic textual patterns with a one-to-one correspondence (an exact pairing) to an event is calculated. A small number of k leads to a small number of semantic textual patterns where most of the terms with high impact on one of these patterns stem from different events. Further, a large number of k leads to a large number of semantic textual patterns where a single event is represented by several patterns. A maximal number of one-to-one correspondences can be obtained by varying the value of k and by calculating the number of identified one-to-one correspondences for each k (rank-validation procedure). As a result, k is selected by applying the rank-validation procedure and j<k semantic textual patterns can be identified with one-to- one correspondences in the step event identification. After selecting the value of k, LSI uses singular value decomposition to split the term-by- document matrix in a product of the matrices U, Σ, and Vt A = U Σ Vt Then, three further matrices Uk, Σk and Vk are calculated by discarding the columns of U, Σ, and V from k+1 on. The components of matrix Uk contain values for the impact of each term on each of the k semantic textual patterns. The impact of each document on each of the k semantic textual patterns can be found in matrix Vk (Thorleuchter, Van den Poel, & Prinzie, 2012b). We use matrix Vk to calculate the conditional cross impact probability of an event B given an event A if both events are represented by a specific semantic textual pattern. This is calculated by the number of documents that are assigned to both events A and B divided by the number of documents that are only assigned to event A. While the impact of a document on a semantic textual pattern is a vale in [-1,..,1], a specific threshold q is used to distinguish between documents that are related to a semantic textual pattern and documents that are not. The cross-impact of A on B is calculated by CI(A,B) = P(B|A) = N(A ∩ B) / N(A) The number of documents in matrix Vk where the impact on a specific event A is above q is used to calculate N(A). N(A ∩ B) is the number of documents in matrix Vk where the impact on event A and B is above q in both cases. To evaluate the CI(A,B) score, precision and recall can be used. The CI(A,B) scores that are in [0,..,1] have to be transformed to Boolean variables [false, true] by use of a further threshold. This enables to identify whether event A impacts event B or not. Precision and recall indicators are well-known performance measures in binary classification. For applying precision and recall, the ground truth has to be determined, too. Case Study _______________________________________________________________________ 6 In a case study, we define defense-based technology areas as events. They are taken over from the technology taxonomy of the European Defense Agency (EDA) where 32 technology areas are selected. For the documents, we use research projects funded by the German Ministry of Defence (GE MoD) in 2007. Descriptions of 985 projects have been identified and stored in documents separately. Some of the research projects examine a specific defense-based technology while other combine several technologies to create new approaches. Fig. 2: Number of one to on correspondences (y-axis) based on the value of k (x-axis) After pre-processing, the term-by-document matrix is built. LSI is applied together with the rank-validation procedure from k = 2 to k = 35. For each k, singular value decomposition is processed and the created k semantic textual patterns are assigned to the 32 technology areas by human experts. Some patterns do not fit to a technology area while others are assigned to several technology areas. The number of semantic textual pattern that are assigned to one and only one technology area (one-to-one correspondences) is depicted in Fig. 2. It shows that up to k = 18, the number of identified one-to-one correspondences is smaller than 9 and from k = 20 on, the number is smaller than 10. Selecting k = 19 leads to the identification of 10 one-to- one correspondences and thus, 10 technology areas. For further processing, k = 19 is selected. The identified 10 technology areas lead to the calculation of 90 conditional cross impact probabilities as calculated by two times the binomial coefficient 10 choose 2. The identified 10 technology areas are presented in Table 1. Table 1: Identified technology areas from EDA taxonomy A02 Signature Related Materials A03 Electronic Materials Technology A04 Photonic/Optical Materials & Device Technology A05 Electronic, Electrical & Electromechanical Device Technology A08 Computing Technologies & Mathematical Techniques B02 Propulsion and Powerplants B04 Electronic Warfare and Directed Energy Technologies B05 Signature Control and Signature Reduction 0   2   4   6   8   10   1   3   5   7   9   11   13   15   17   19   21   23   25   27   29   31   33   35   _______________________________________________________________________ 7 B06 Sensor Systems B08 Simulators, Trainers and Synthetic Environments The impact of a document on a technology area in matrix Vk is a value in [-1,..,1]. The threshold q is selected as suggested in literature (Thorleuchter & Van den Poel, 2013a). Based on the component values in matrix Vk, the 90 conditional cross impact probabilities are calculated based on the value of q = 0.4. The results are depicted in Table 2 colored in five different grayscales from bright to dark concerning the five cases: No cross impact: CI(A,B) = 0; Low cross impact: 0 < CI(A,B) ≤ 0.25; Medium cross impact: 0.25 < CI(A,B) ≤ 0.50; High cross impact: 0.50 < CI(A,B) ≤ 0.75; Very high cross impact: CI(A,B) > 0.75. Table 2: Result matrix of the calculated CI(A,B) e.g. CI(A02, A03) = 0.07 A02 A03 A04 A05 A08 B02 B04 B05 B06 B08 A02 - 0.07 0.04 0 0 0 0 0.89 0.04 0 A03 0.05 - 0.13 0.82 0 0.18 0.24 0 0.21 0 A04 0.02 0.12 - 0.19 0 0 0.26 0.16 0.30 0 A05 0 0.35 0.09 - 0 0.27 0.25 0.02 0.24 0.05 A08 0 0 0 0 - 0 0.08 0 0.20 0.29 B02 0 0.16 0 0.53 0 - 0 0.07 0 0 B04 0 0.25 0.31 0.61 0.17 0 - 0.03 0.06 0 B05 0.66 0 0.18 0.05 0 0.08 0.03 - 0.03 0 B06 0.01 0.11 0.17 0.28 0.20 0 0.03 0.01 - 0 B08 0 0 0 0.13 0.58 0 0 0 0 - Evaluation For the evaluation of the results, we use a further study (furthermore named comparative study) from literature (Thorleuchter, Van den Poel, & Prinzie, 2010) that has calculated the conditional cross impact probabilities from the same input data. The comparative study uses a knowledge structure based classification approach based on centroid vectors to calculate the impacts. The conditional cross impact probabilities calculated by the comparative study are assigned to a positive and a negative class concerning a specific threshold r. This value (r = 0.25) is also used to evaluate the results of our proposed approach. Creating a ground truth for calculating precision and recall is mandatory. This can be done manually by human experts. They have to decide whether a technology has an impact on a second technology _______________________________________________________________________ 8 above the specific threshold or not. The decision that an impact is above a specific threshold is a very subjective task for human experts and it might be that two experts make different decisions. In the comparative study, human experts have analyzed the calculated results and they are able to give a heuristic explanation that confirms each single result. They have not identified any misclassification because the project descriptions use a technical language where terms are more strictly defined than terms from the colloquial language. This enables very good classification results - in this case 100 % precision at 100 % recall. Thus, the results from the comparative study can be used as ground truth to evaluate our proposed approach. A selection of 11 from the 90 conditional cross impact probabilities is presented in Table 3. ’Techn. area A’ represents the influencing technology area and ’Techn. area B’ is the influenced technology area. Both technology areas stem from the 10 technologies as depicted in Table 1. The conditional probability of technology area B given technology area A is CI(A,B) as calculated by our proposed approach. The Boolean cross impact score BCI(A,B) is true if CI(A,B) is above threshold r = 0.25 otherwise it is false. CIcom(A,B) and BCIcom(A,B) are the corresponding values as calculated from the comparative study. Each row represents two different technology areas from Table 1 ordered by the CIcom(A,B) score where BCIcom(A,B) is true. Resdiff(A,B) is the difference between the residual from CI(A,B) to the residual from CIcom(A,B). Table 3: Technology area pairs with BCIcom(A,B) is true ordered by CIcom(A,B) in 2007 Techn. area A Techn. area B CI (A,B) BCI (A,B) CIcom (A,B) BCIcom (A,B) Resdiff (A,B) A02 B05 0.89 True 0.92 True -0.01 A03 A05 0.82 True 0.86 True -0.02 B05 A02 0.66 True 0.62 True 0.06 B04 A05 0.61 True 0.61 True 0.02 B02 A05 0.53 True 0.54 True 0.01 B08 A08 0.58 True 0.53 True 0.07 A05 A03 0.35 True 0.32 True 0.05 A08 B08 0.29 True 0.31 True 0.00 A05 B02 0.27 True 0.29 True 0.00 A05 B06 0.24 False 0.27 True -0.01 A05 B04 0.25 False 0.26 True 0.01 Based on the 90 calculated cross-impact probabilities, the comparative study has shown that in 11 cases, BCIcom(A,B) is true and thus, an impact above the threshold can be seen. In 79 cases, _______________________________________________________________________ 9 BCIcom(A,B) is false. This leads to a frequent baseline of about 12%. The calculated cross-impact probabilities from the proposed approach lead to BCI(A,B) = true in 13 cases and to BCI(A,B) = false in 77 cases. In 9 of the 13 positive cases, BCIcom(A,B) is also true while in 4 cases BCI(A,B) is true and BCIcom(A,B) is false. These results are depicted in Table 4. Thus, precision is calculated as 9 / 13 = 69 % and recall is calculated as 9 / 11 = 82 %. This outperforms frequent baseline of 12 % precision at 82 % recall. The differences between the residuals are small. This also shows that the results are similar to those of the ground truth. Table 4: Confusion matrix Predictive Class Yes No Actual class Yes 9 2 No 4 75 To present a detailed example, we discuss the cross-impact among technology area B02 (Propulsion and Powerplants) and A05 (Electronic, Electrical & Electromechanical Device Technology). In 2007, a well-known trend in propulsion and powerplant technology was the creation of a more electric engine. The corresponding research projects that have been processed during that time can be assigned to both technology areas. The comparative study has shown that 54% of all research projects from technology area B02 are also assigned to A05 and that 29% of all research projects from technology area A05 are also assigned to B02. This is because 37 research projects are assigned to B02, 69 are assigned to A05, and 20 are assigned to both. In contrast to this, the proposed approach assigns 34 research projects to B02, 67 are assigned to A05, and 18 are assigned to both. The differences are evaluated manually by human experts. They could not identify concrete hints for a misclassification because the assignment of the corresponding research projects is very subjective. An example is a research project that develops new kinds of lubricants. The research results can be used for improving propulsion and powerplants but also for many further applications. The question whether this project is related to B02 or not is very subjective. Overall, the proposed approach outperforms the baseline that proves its feasibility. Further, only small differences between CI(A,B) and CIcom(A,B) can be seen in the 90 cross-impacts. This also shows the feasibility of the proposed approach. The 2 cases where BCI(A,B) is false and BCIcom(A,B) is true as well as the 4 cases where BCI(A,B) is true and BCIcom(A,B) is false are resulted by the selection of the threshold r. Despite of the small differences, some CI(A,B) and CIcom(A,B) are assigned to different classes because their value is about the value of the threshold. Thus, these 6 cases are not significant. Conclusion We propose a new approach that calculates conditional cross-impact probabilities. In contrast to previous work, this work uses semantic classification by applying LSI together with a rank validation procedure. While knowledge structure based approaches are used for quantitative CIA in literature, the aim of this work is to show that semantic approaches also can be used. _______________________________________________________________________ 10 In a case study the proposed approach is applied to identify conditional cross-impact probabilities between technologies. The evaluation is based on a further study where an extensive evaluation on the same data was already processed. As a result, the evaluation shows that the proposed approach outperforms the frequent baseline. Thus, it can be successful applied for quantitative CIA. Comparing the proposed approach to a knowledge structure based approach fails because the assignment of texts to classes in the case study is a very subjective task. Literature has shown that in contrast to knowledge structure based approaches, semantic approaches have advances by processing colloquial texts (e.g. internet blogs) rather than highly structured texts where each term is pre-defined in a way that synonym and homonym problems normally do not occur. Technological descriptions are rather structured texts than colloquial texts because several technical terms have well-known meanings. Thus, the results of the case study from both approaches are similar. Future work could be compared both kinds of approaches by use of colloquial texts, e.g. by including documents from the internet written in different languages and in different writing styles. We expect different results from both approaches so that a comparison possibly will show advances of the semantic approach. Normally, LSI is a clustering approach. We used it together with a rank validation procedure for classification. In cases where the events are not pre-defined, LSI can be used as clustering approach without the rank validation procedure. This improves performance of the approach on one hand but probably the automatically created events are not comprehensible for the users on the other hand. This might be a further avenue of future research. Future work also should focus on are the implementation of the compared cross impact (CCI) approach with LSI. CCI analysis extends the CIA and up to now, it is only processed quantitatively by knowledge structure based approaches. Literature Banuls, V. A., Turoff, M., & Hiltz, S.R. (2013). Collaborative scenario modeling in emergency management through cross-impact. Technological Forecasting and Social Change, in press. Bell, W. (2002). A community of futurists and the state of the futures field. Futures, 34(3-4), 235-247. Blanning, R. W., & Reinig, B. A. (1999). Cross-impact analysis using group decision support systems: an application to the future of Hong Kong. Futures, 31(1), 39-56. Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet allocation. Journal of Machine Learning Research, 3(4-5), 993-1022. Caselles-Moncho, A. (1986). An empirical comparison of cross-impact models for forecasting sales. International Journal of Forecasting, 2(3), 295-303. Choi, C., Kim, S., & Park, Y. (2007). A patent-based cross impact analysis for quantitative estimation of technological impact: the case of information and communication technology. Technological Forecasting and Social Change, 74(2007), 1296-1314. Choi, O., Kim, K., Wang, D., Yeh, H., & Hong, M. (2012). Personalized Mobile Information Retrieval System. International Journal of Advanced Robotic Systems, doi: 10.5772/50910. Christidis, K., Mentzas, G., & Apostolou, D. (2012). Using latent topics to enhance search and recommendation in Enterprise Social Software. Expert Systems with Applications, 39(10), 9297-9307. Dalkey, N. C. (1972). An elementary cross-impact model, Technological Forecasting and Social Change, 3, 341-351. D’Haen, J., Van den Poel, D., & Thorleuchter, D. (2013). Predicting customer profitability during acquisition: Finding the optimal combination of data source and data mining technique. Expert Systems with Applications, 40(6), 2007-2012. Enzer, S. (1972). Cross-impact techniques in technology assessment. Futures, 4(1), 30-51. _______________________________________________________________________ 11 Geschka, H. (1983). Creativity techniques in product planning and development: A view from West Germany. R&D Management, 13(3), 169-183. Gordon, T., & Haywood, H. (1968). Initial experiments with the cross impact matrix method of forecasting, Futures, 1(2), 100-116. Hofmann, T. (1999). Probabilistic Latent Semantic Indexing. In: Proceedings of the Twenty-Second Annual International SIGIR Conference on Research and Development in Information Retrieval (SIGIR-99). Jeong, G. H., & Kim, S. H. (1997). Aqualitative cross-impact approach to find the key technology. Technological Forecasting and Social Change, 55(3), 203-214. Jiang, J., Berry, M. W., Donato, J. M., Ostrouchov, G., & Grady, N. W. (1999). Mining consumer product data via latent semantic indexing. Intelligent Data Analysis, 3(5), 377-398. Kauffman, S., Lobo, J., & Macready, W. G. (2000). Optimal search on a technology landscape. Journal of Economic Behavior & Organization, 43(2), 141-166. Kim, C., Lee, H., Seol, H., & Lee, C. (2011). Identifying core technologies based on technological cross-impacts: An association rule mining (ARM) and analytic network process (ANP) approach. Expert Systems with Applications, 38(10), 12559-12564. Kuhn, A., Ducasse, S., & Girba, T. (2007). Semantic clustering: Identifying topics in source code. Information and Software Technology, 49(3), 230-243. Lee, D. D., & Seung, H. S. (1999). Learning the parts of objects by non-negative matrix factorization. Nature, 401(6755), 788-791. Lee, D. D., & Seung, H. S. (2001). Algorithms for Non-negative Matrix Factorization. In: Advances in Neural Information Processing Systems 13. Proceedings of the 2000 Conference. MIT Press. pp. 556-562. Luo, Q., Chen, E., & Xiong, H. (2011). A semantic term weighting scheme for text categorization. Expert Systems with Applications, 38(10), 12708-12716. Mitchell, R. B., Tydeman, J., & Curnow, R. (1977). Scenario generation: limitations and developments in cross- impact analysis. Futures, 9(3), 205-215. Park, D, H., Kim, H. K., Choi, I. Y., & Kim, J. K. (2012). A literature review and classification of recommender systems research. Expert Systems with Applications, 39(11), 10059-10072. Ramirez, E. H., Brena, R. F., Magatti, D., & Stella, F. (2012). Topic model validation. Neurocomputing 76(1), 125-133. Salton, G., Allan, J., & Buckley, C. (1994). Automatic structuring and retrieval of large text files. Communications of the ACM, 37(2), 97-108. Schuler, A., Thompson, W. A., Vertinsky, I., & Ziv, Y. (1991). Cross impact analysis of technological innovation and development in the softwood lumber industry in Canada: a structural modeling approach. IEEE Transition of Engineering Management, 38(3), 224-236. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2010). A compared R&D-based and patent-based cross impact analysis for identifying relationships between technologies. Technological Forecasting and Social Change, 77(7), 1037-1050. Thorleuchter, D., & Van den Poel, D. (2012a). Predicting e-commerce company success by mining the text of its publicly-accessible website. Expert Systems with Applications, 39(17), 13026-13034. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2012b). Analyzing existing customers’ websites to improve the customer acquisition process as well as the profitability prediction in B-to-B marketing. Expert Systems with Applications, 39(3), 2597-2605. Thorleuchter, D., & Van den Poel, D. (2012c). Improved multilevel security with latent semantic indexing. Expert Systems with Applications, 38(18), 13462-13471. Thorleuchter, D., & Van den Poel, D. (2013a). Weak signal identification with semantic web mining. Expert Systems with Applications, 40(12), 4978-4985. Thorleuchter, D., & Van den Poel, D. (2013b). Protecting Research and Technology from Espionage. Expert Systems with Applications, 40(9), 3432-3440. Thorleuchter, D., & Van den Poel, D. (2013c). Technology classification with latent semantic indexing. Expert Systems with Applications, 40(5), 1786-1795. Thorleuchter, D., & Van den Poel, D. (2013d). Web Mining based Extraction of Problem Solution Ideas. Expert Systems with Applications, 40(10), 3961-3969. Tsai, H.H. (2012). Global data mining: An empirical study of current trends, future forecasts and technology diffusions. Expert Systems with Applications, 39(9), 8172-8181. Yu, L., Hurley, T., Kliebenstein, J., & Orazem, P. (2012). A test for complementarities among multiple technologies that avoids the curse of dimensionality. Economics Letters, 116(3), 354-357. Zeng, J., Duan, J., Cao, W., & Wu, C. (2012). Topics modeling based on selective Zipf distribution. Expert Systems with Applications, 39(7), 6541-6546. Zipf, G. K. (1949). Human Behaviour and the Principle of Least Effort. Cambridge: Addison-Wesley. _______________________________________________________________________ 12 
work_aajczoflznegllwaw3jeykiofi	----	No Job Name Research Submission An Expert System for Headache Diagnosis: The Computerized Headache Assessment Tool (CHAT) Morris Maizels, MD; William J. Wolfe, PhD Background.—Migraine is a highly prevalent chronic disorder associated with significant morbidity. Chronic daily headache syndromes, while less common, are less likely to be recognized, and impair quality of life to an even greater extent than episodic migraine. A variety of screening and diagnostic tools for migraine have been proposed and studied. Few investigators have developed and evaluated computerized programs to diagnose headache. Objectives.—To develop and determine the accuracy and utility of a computerized headache assessment tool (CHAT). CHAT was designed to identify all of the major primary headache disorders, distinguish daily from episodic types, and recognize medication overuse. Methods.—CHAT was developed using an expert systems approach to headache diagnosis, with initial branch points de- termined by headache frequency and duration. Appropriate clinical criteria are presented relevant to brief and longer-lasting headaches. CHAT was posted on a web site using Microsoft active server pages and a SQL-server database server. A convenience sample of patients who presented to the adult urgent care department with headache, and patients in a family practice waiting room, were solicited to participate. Those who completed the on-line questionnaire were contacted for a diagnostic interview. Results.—One hundred thirty-five patients completed CHAT and 117 completed a diagnostic interview. CHAT correctly identified 35/35 (100%) patients with episodic migraine and 42/49 (85.7%) of patients with transformed migraine. CHAT also correctly identified 11/11 patients with chronic tension-type headache, 2/2 with episodic tension-type headache, and 1/1 with episodic cluster headache. Medication overuse was correctly recognized in 43/52 (82.7%). The most common misdiagnoses by CHAT were seen in patients with transformed migraine or new daily persistent headache. Fifty patients were referred to their primary care physician and 62 to the headache clinic. Of 29 patients referred to the PCP with a confirmed diagnosis of migraine, 25 made a follow-up appointment, the PCP diagnosed migraine in 19, and initiated migraine-specific therapy or prophylaxis in 17. Conclusion.—The described expert system displays high diagnostic accuracy for migraine and other primary headache disorders, including daily headache syndromes and medication overuse. As part of a disease management program, CHAT led to patients receiving appropriate diagnoses and therapy. Limitations of the system include patient willingness to utilize the program, introducing such a process into the culture of medical care, and the difficult distinction of transformed migraine. Key words: headache diagnosis, expert systems, disease management (Headache 2008;48:72-78) Annually, over 10% of the population suffers at least one migraine headache,1 with a significant im- pact on the individual’s quality of life,2 as well as a major economic and societal burden.3 Up to 5% of the adult population may suffer from daily headache syndromes, with medication overuse identified in one- third of these patients.4 The impact of chronic daily headache on the individual and society is even greater than that of episodic migraine. Despite significant advances in acute and pre- ventive therapy, migraine remains underdiagnosed, and “drug rebound headache” an “unrecognized epidemic.”5 A population-based survey found that From the Kaiser Permanente – Family Medicine, Woodland Hills, CA (Dr. Maizels); and California State University Chan- nel Islands – Computer Science, Camarillo, CA (Dr. Wolfe). Address all correspondence to Dr. Morris Maizels, Kaiser Per- manente – Family Medicine, 5601 De Soto Avenue, Woodland Hills, CA 91365-4084. Accepted for publication June 27, 2007. Conflict of Interest: None ISSN 0017-8748 doi: 10.1111/j.1526-4610.2007.00918.x Published by Blackwell Publishing Headache © 2007 the Authors Journal compilation © 2008 American Headache Society 72 only 48% of participants who fulfilled International Headache Society (IHS) criteria for migraine had been diagnosed as having migraine.6 Only 50% of patients who fulfill criteria for migraine and seek medical care are correctly diagnosed.7 The burden of undiagnosed migraine is significant: 24% of patients with undiag- nosed migraine missed at least 1 day of work or school in the previous 3 months, and 45% reported at least a 50% reduction in productivity.6 Several authors have also proposed simplified cri- teria for diagnosis of migraine. These typically include some subset of IHS criteria.8-10 A variety of question- naires to diagnose migraine have been developed, with reported sensitivities of 76–84%, and specificities of 92-99%.11,12 None of these instruments has come to be used commonly in clinical practice. There have been few reports of utilizing computer technology to diagnose migraine and other primary headache disorders. Drummond and Lance utilized a computer algorithm to determine to what extent clusters of symptoms differentiated diagnoses along the “migraine-tension headache spectrum.”13 Andrew et al.14 and Gobel et al.15 incorporated the 1988 IHS classification system16 into a computer program. Go- bel’s group later utilized the program to standardize inclusion of patients into a sumatriptan trial.17 Diagnostic headache diaries18,19 and structured medical records20 have been incorporated into com- puter programs to aid in headache diagnosis. Several authors have described programs but not reported val- idating the instruments with diagnostic interviews.21-23 Most of the above programs are essentially checklists of symptoms which the computer program tries to fit into a specific diagnosis. The author (Morris Maizels) sought to develop a headache diagnostic program which would employ simple branching decisions to mimic the logic of a clin- ician. The Computerized Headache Assessment Tool (CHAT) was developed with the following features in mind (summarized in Table 1): completed by patients on-line; questions are systematically presented de- pending on prior answers (patients are only presented with relevant questions); screens for all common pri- mary headache disorders; distinguishes chronic from episodic subtypes; recognizes medication overuse; and screens for potentially worrisome headaches. It was hoped that the output of the program would be in a format suitable for a physician to review with the pa- tient, and to include in the medical record. This article describes the development of the program, the diag- nostic accuracy for primary headache syndromes, and the outcomes of screening. METHODS Description of the Computerized Headache Assess- ment Tool (CHAT).—CHAT consists of 4 sections: an explanatory and disclaimer page, the interview section, feedback of answers, and diagnostic output. The disclaimer page requires patients to confirm their understanding that the tool is intended for use in conjunction with a health care professional. The interview section begins with 2 screens which all subjects see: the first asks the frequency of moderate-to-severe headaches, and the second the duration of these headaches. Based on these answers, the questioning branches. Patients with headaches last- ing 3 hours or less next answer questions relevant to brief headache syndromes (number of attacks/day, location, quality, autonomic features, triggering). Pa- tients with severe headaches 4 hours or longer see Table 1.—Desired Characteristics of a Computerized Headache Diagnostic System Diagnostic Features • Recognize the major primary headache disorders, including migraine, tension-type headache, cluster headache, and brief headache syndromes • Distinguish chronic from episodic subtypes of migraine, tension-type headache, and cluster headache • Recognize medication overuse • Screen for “worrisome” features User-friendly • Patient self-administered • Rapid completion Human Interviewer Characteristics • Branching logic • Subsequent questions selected based on prior answers (patient only sees relevant questions) • Feedback answers • Recognize and correct inconsistencies Clinically Useful Output • Criteria for generated diagnoses are explained in a format understandable for patients • Hardcopy of output is suitable for physician to review with patient, and to include into a medical record Headache 73 a screen, which asks questions based on IHS crite- ria for migraine and tension-type headache (severity, laterality, throbbing vs dull, exacerbation with phys- ical activity, associated nausea or vomiting, light or noise sensitivity). This screen also determines the fre- quency of milder headaches and medication usage. Pa- tients with headaches lasting longer than 4 hours are also asked questions about their previous headaches: whether they were daily from onset or became daily suddenly or gradually; whether there were any sig- nificant events associated with headache onset or “transformation”; whether there has been a signifi- cant change in headache pattern; and how long the current headache pattern has persisted. For patients whose current headaches did not fulfill migraine cri- teria on the previous screen, they see an additional screen to determine whether previous headaches met migraine criteria. Feedback. A screen summarizes the patients an- swers, and the patient has an opportunity to correct any answers. Diagnostic Output.—The screen shows an “assess- ment” (the word “diagnosis” is not used, to reduce the subject’s reliance on the CHAT assessment with- out clinician review). Assessments were based on the 1988 IHS classification system,16 and the Silberstein– Lipton revised criteria for transformed migraine.24 Po- tential assessments for headaches lasting 4 hours or longer include migraine (episodic or transformed), mi- grainous, episodic or chronic tension-type headache, and new daily persistent headache. The assessment in- cludes “. . . with medication overuse” for subjects who indicated symptomatic medication use 3–4 days/week or more. For brief headache syndromes, diagnoses in- clude cluster (episodic or chronic) and atypical cluster headache, idiopathic stabbing headache, and trigemi- nal neuralgia. Importantly, headaches that do not ful- fill criteria of the above conditions generate an as- sessment of “cannot assess.” All diagnoses are fol- lowed by a brief explanation of how the diagnosis was made. Patients with daily or near-daily headaches see a caution in bold print that frequent or daily headaches require evaluation to exclude worrisome causes. Errors and Inconsistencies.—During initial testing, it was noted that patients might indicate that headache duration was 3 hours or less when in fact it was over 4 hours. To confirm that this answer is correct, pa- tients with episodic headache who indicate duration as 3 hours or less see a screen that says “Do these headaches ever last more than 4 hours?” Further, for patients who are initially routed to the brief headaches screen, if they do not fulfill criteria for any of the brief headache syndromes, they are routed through the questions for patients with headaches lasting 4 hours or longer. Study Setting.—The patient population included members of a suburban health maintenance organi- zation. Patients completed the internet questionnaire at home or work. Patient Selection.—A poster in the triage section of the adult urgent care (AUC) department described the study, and nurses were asked to identify headache patients and hand them a flyer. However, because of poor compliance with this method, we later mailed out a study flyer to patients who had presented to AUC with headache. A flyer describing the study was also posted in the family medicine department at the receptionist’s desk. There were no inclusion or exclusion criteria for participation, other than age over 18. Diagnostic Confirmation.—Study participants who granted permission were contacted by telephone. A headache clinic nurse experienced in headache diagnosis performed diagnostic interviews. The in- terview began with the San Diego migraine ques- tionnaire, a well-validated instrument for migraine diagnosis.12 Patients with frequent headache were further classified by the modified Silberstein–Lipton criteria.24 Patients with brief headache syndromes were classified by IHS criteria. All interviews with diagnostic uncertainty were reviewed by the author (Morris Maizels). Patient Follow-Up.—Patients were offered follow- up appointments to their primary care physician if headaches were diagnosed as episodic migraine or tension-type headache, and to the headache clinic if headaches were diagnosed as chronic migraine or an- other daily headache syndrome. Chart Review.—Medical records of patients who completed their diagnostic interview were reviewed within 6 months to determine if they had kept their 74 January 2008 follow-up appointment, and whether any new treat- ments were initiated as a result. Data Analysis.—Simple descriptive analyses were performed on the data. This study was approved by the Institutional Re- view Board. Patient consent was obtained through the introductory screen online. RESULTS One hundred thirty-five subjects completed the online survey, 103 from AUC and 32 from primary care. Diagnostic interviews were completed on 117 subjects. Physician diagnoses were available for 98 pa- tients from the AUC sample but not for the primary care waiting room sample. The confirmed diagnoses, CHAT assessments, and AUC physician diagnoses are shown in Table 2. CHAT correctly identified 35/35 (100%) patients with episodic migraine and 42/49 (6%) of patients with transformed migraine. Urgent care physicians correctly diagnosed 26/30 (86.7%) of patients with episodic migraine, and recognized 27/40 (66.7%) of patients with transformed migraine as migraine, al- though none as a daily headache disorder. CHAT also correctly identified 12/12 patients with chronic tension-type headache, 2/2 with episodic tension-type headache, and 1/1 with episodic cluster headache. Al- though NDPH was correctly identified in only 3/7 (42.9%) patients, CHAT recognized 6/7 as a daily headache syndrome. CHAT recognized medication overuse in 43/52 (82.7%) patients who met criteria for the diagnosis on diagnostic interview. Three additional assessments of medication overuse by CHAT were not confirmed. Overall, CHAT recognized 85/90 (94.4%) cases of “any migraine,” (ie, migraine, migrainous, or TM) and 63/68 (92.6%) as a daily headache syn- drome (TM, CTTH, or NDPH). The diagnostic ac- curacy for all headache diagnoses, including unclas- sifiable (but not including medication overuse) was 104/117 (88.9%). Patient Follow-Up.—Fifty patients were referred to their primary care physician, and 42 made at least one visit for headache follow-up. Of 29 patients referred with a confirmed diagnosis of migraine, 25 made a follow-up appointment, the PCP diagnosed mi- graine in 19, and initiated migraine-specific therapy or prophylaxis in 17. Sixty-two patients were referred for evaluation to the headache clinic, and 51 made at least one visit. Table 2.—Confirmed Diagnoses, Assessments Generated by CHAT, and Physician Diagnoses From Adult Urgent Care Visits. Percentages (in Parentheses) of Correct Assessment/Diagnosis Confirmed Diagnosis n = 117 CHAT Assessment n = 117 Physician Diagnosis n = 98 Episodic migraine 35 Episodic migraine 35 (100%) “Migraine” 26/30 (86.7%) Migrainous* 6 Migrainous 5 (83.3%) Probable TM*** 1 TH 1 “vascular headache” 1 no specific diagnosis 1 TM** 49 TM 42/49 (85.7%) “Migraine” 27/40 (66.7%) Episodic migraine 2 CTTH 3 atypical cluster 1 “can’t assess” 1 NDPH 7 NDPH 3 (42.9%) TH 2 no specific diagnosis 3 TM 1 CTTH 2 “can’t assess” 1 CTTH 12 CTTH 12/12 (100%) migraine 2 “TTH” 1 no specific diagnosis 9 ETTH 2 ETTH 2/2 (100%) no specific diagnosis 2 Cluster headache Cluster headache • episodic 1 • episodic 1/1 (100%) • no specific diagnosis 1 • atypical 1 • atypical 1/1 (100%) • “rhinitis vs. cluster” 1 Unclassifiable 4 “can’t assess” 3 (75%) no specific diagnosis 4 NDPH 1 Medication overuse 52 Medication overuse 43 (82.7%) (none identified) * “Migrainous” now labeled “probable migraine” by ICHD-II classification. ** “Transformed migraine,” as designated by Silberstein- Lipton criteria, not recognized in ICHD-II classification. *** Probable transformed migraine, ie, patient could not recall whether headaches changed gradually. TM = transformed migraine; NDPH = new daily persistent headache; CTTH = chronic tension-type headache; ETTH = episodic tension-type headache; TH = tension headache. Headache 75 DISCUSSION An expert systems approach to headache diag- nosis, as incorporated in CHAT, achieves diagnostic precision for primary headache diagnoses. Like all of clinical medicine, accurate diagnosis depends on the accuracy of patient responses, and a skilled human in- terviewer is likely to elicit more accurate responses than a fixed computer program. In our patient sample, CHAT recognized 100% of patients correctly as episodic migraine. There were 13 incorrect diagnoses for the entire cohort, of which 11 were in patients classified by interview as TM or NDPH. The diagnosis of TM and NDPH both rely on the patient’s accurate recall of whether headaches began abruptly and/or became daily gradually or abruptly.The accuracy of this recall is difficult to deter- mine. The designation of TM has been replaced in the current ICHD-II by chronic migraine and episodic mi- graine + CTTH. However, the difficulties of this diag- nostic category led to an appendix revision of chronic migraine (appendix 1.5.1 chronic migraine), which al- lows for CM to be diagnosed with �15 headache days/month, of which 8 days are migraine-like.25 From a clinical point-of-view, especially in a primary care set- ting, the distinctions of migraine from non-migraine, and frequent/daily from episodic are important. The nuances of CM and NDPH would not change clinical practice. The ability of the program to recognize certain headaches as unclassifiable, ie, “can’t assess,” is an important feature of this program. This assessment is most often generated when patients entered incorrect data for headache duration (ie, patients with migraine may indicate headache duration less than 3 hours be- cause of relief with medication). The recognition of medication overuse is an im- portant feature of headache evaluation, and one com- monly overlooked in primary care settings. CHAT cor- rectly recognized medication overuse in 43/52 (82.7%) patients, with 3 false positive diagnoses. The false pos- itives occurred in part because some patients confused preventive with acute/abortive medication. Physicians had a high rate of correct diagnosis of migraine, although the study did not determine what percent of the AUC sample identified themselves as having migraine.The label of “transformed” or chronic migraine is not one familiar to primary care physicians. A surrogate marker to indicate whether physicians rec- ognize headache frequency and its significance might be the use of prophylaxis. This measure was beyond the scope of the study. CHAT is a unique system in that it combines simple human-like branching logic to determine the most appropriate diagnostic questions to ask. Re- cently, Sarchielli et al.19 reported on a software pro- gram which could generate ICHD-II diagnoses of all migraine subtypes, tension-type headache, cluster headache, and other trigeminal autonomic cephalgias. However, the program relies on data entered from pa- tients’ headache diaries, and is suggested as useful in tertiary headache centers rather than for primary care diagnosis. Study Limitations and Limitations of CHAT.— Headaches were most difficult to classify, both by human interviewer and by CHAT, in patients with headache of recent onset (often associated with viral- like syndromes). The tested edition of CHAT does not recognize many of the brief headache syndromes, specifically paroxysmal hemicranias, SUNCT, trigeminal neural- gia, and hypnic headache. Further, CHAT cannot generate more than a sin- gle headache diagnosis. This feature would require revision to be consistent with ICHD-II which distin- guishes chronic migraine from episodic migraine + CTTH. The appropriate clinical application of CHAT would have the physician confirm the accuracy of patient replies. This may not occur, leading to the possibility of treatment based on unvalidated patient replies. However, this problem exists for any screening instrument used in medicine. Any automated screening program must take pains to recognize potentially serious causes for headache, or caution both patient and physician users about that possibility. A later version of CHAT added the question, “Has this headache pattern been sta- ble for the past six months?” This question has not been validated as an adequate screen for ominous headaches. Future Development.—The criterion-based portion of this program can be readily modified to the cur- 76 January 2008 rent ICHD-II and future iterations of headache diag- nosis. The branching logic at present appears appro- priate although further study may lead to future modi- fication. The main modification would involve deletion of “transformed migraine,” replacing it with chronic mi- graine, episodic migraine + CTTH, or CTTH alone. “Migrainous” would be replaced with “probable mi- graine,” a change not requiring any change in the logic of the program. The less common primary care headache syndromes (paroxysmal hemicranias, SUNCT, hypnic headache, etc) can readily be included, although their rarity will make validation in a general population difficult. CONCLUSION An internet-based headache assessment tool has demonstrated a high degree of accuracy in recogniz- ing primaryheadache disorders, distinguishing chronic daily from episodic headache, and recognizing medi- cation overuse. No other computer-assisted headache diagnostic program described in the medical literature has a similar scope of diagnosis, or demonstrated ac- curacy. The major challenge of computer-assisted headache diagnosis at present is not in developing better programs, but in facilitating patients to use such programs, and encouraging their use in primary care settings. Further development might integrate online headache assessment with education about headache treatment, identification of headache triggers, and innovative online behavioral modification programs. REFERENCES 1. Stewart WF, Lipton RB, Celentano DD, Reed ML. Prevalence of migraine headache in the United States: Relation to age, income, race, and other so- ciodemographic factors. JAMA. 1992;267:64-69. 2. Dahlof CG, Solomon GD. The burden of migraine to the individual sufferer: A review. Eur J Neurol. 1998;5:525-533. 3. Hu HX, Markson LE, Lipton RB, Stewart WF, Berger ML. Burden of migraine in the United States: Disability and economic costs. Arch Int Med. 1999;159:813-818. 4. Castillo J, Munoz P, Guitera V, Pascual J. Epidemi- ology of chronic daily headache in the general popu- lation. Headache. 1999;39:190-196. 5. Edmeads J. Analgesic-induced headaches: An unrec- ognized epidemic. Headache. 1990;30:614-615. 6. Lipton RB, Diamond S, Reed M, Diamond ML, Stew- art WF. Migraine diagnosis and treatment: Results from the American Migraine Study II. Headache. 2001;41:638-645. 7. Stang PE, Osterhaus JT, Celentano DD. Migraine: Patterns of healthcare use. Neurology. 1994;44(Suppl 4):S47-S55. 8. Gervil M, Ulrich V, Olesen J, Russell MB. Screen- ing for migraine in the general population: Validation of a simple questionnaire. Cephalalgia. 1998;18:342- 348. 9. Solomon S. Criteria for the diagnosis of migraine in clinical practice. Headache. 1991;31:384-387. 10. Lipton RB, Dodick D, Sadovsky R, et al. A self- administered screener for migraine in primary care. The IDMigraine TM validation study. Neurology. 2003;61:375-382. 11. Rasmussen BK, Jensen R, Olesen J. Questionnaire versus clinical interview in the diagnosis of headache. Headache. 1991;31:290-295. 12. Tom T, Brody M, Valabhji A, Turner L, Molgaard C, Rothrock J. Validation of a new instrument for de- termining migraine prevalence: The UCSD migraine questionnaire. Neurology. 1994;44:925-928. 13. Drummond PD, Lance JW. Clinical diagnosis and computer analysis of headache symptoms. J Neurol Neurosurg Psychiatry. 1984;47:128-133. 14. Andrew ME, Penzien DB, Rains JC, Knowlton GE, McNaulty RD. Development of a computer applica- tion for headache diagnosis: The headache diagnostic system. Int J Biomed Comput. 1992;31:17-24. 15. Gobel H, Soyka D. Acceptance, reliability and va- lidity of computerized headache analysis on the ba- sis of the IHS headache classification. Cephalalgia. 1993(Suppl 13):121. 16. Headache Classification Committee of the Interna- tional Headache Society. Classification and diagnos- tic criteria for headache disorders, cranial neural- gias and facial pain. Cephalalgia. 1988;8(Suppl 7): 1-96. 17. Gobel H, Heinze A, Kuhn K, Heuss D, Lindner V. Effect of operationalized computer diagnosis on the therapeutic results of sumatriptan in general practice. Cephalalgia. 1998;18:481-486. Headache 77 18. Nielsen KD, Rasmussen C, Russell MB. The diagnos- tic headache diary – a headache expert system. Stud Health Technol Inform. 2000;78:149-160. 19. Sarchielli P, Pedini M, Coppola F, et al. Applica- tion of the ICHD-II criteria to the diagnosis of pri- mary chronic headaches via a computerized struc- tured record. Headache. 2007;47:38-44. 20. Gallai V, Sarchielli P, Alberti A, et al. Collabora- tive group for the application of the IHS criteria of the Italian society for the study of headache. Appli- cation of the 1988 International Headache Society diagnostic criteria in nine Italian headache centers using a computerized structured record. Headache. 2002;42:1016-1024. 21. Moses AJ, Lieberman M, Kittay I, Learreta JA. Computer-aided diagnoses of chronic head pain: Ex- planation, study data, implications, and challenges. Cranio. 2006;24:60-66. 22. Mainardi F, Maggioni F, Dainese F, Zanchin G. De- velopment of an ICHD-II based computer system for the general practitioner. J Headache Pain. 2005;6:211- 212. 23. De Simone R, Marano E, Bonavita V. Towards the computerisation of ANIRCEF Headache Cen- tres. Presentation of AIDA CEFALEE, a com- puter assisted diagnosis database for the manage- ment of headache patients. Neurol Sci. 2004;25:S218- S222. 24. Silberstein SD, Lipton RB, Solomon S, Mathew N. Classification of daily and near-daily headaches: Proposed revisions to the IHS criteria. Headache. 1994;34:1-7. 25. Olesen J, Bousser M-G, Diener H-C, et al. New appendix criteria open for a broader concept of chronic migraine. Headache Classification Commit- tee. Cephalalgia. 2006;26:742-746. 78 January 2008 
work_aarrsiiyjfelfbx4vhkqxppyhy	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/authorsrights http://www.elsevier.com/authorsrights Author's personal copy Synergistic case-based reasoning in medical domains Cindy Marling a,⇑, Stefania Montani b, Isabelle Bichindaritz c, Peter Funk d a School of Electrical Engineering and Computer Science, Ohio University, Athens, OH 45701, USA b DISIT, Computer Science Institute, Università del Piemonte Orientale, 15121 Alessandria, Italy c Computer Science Department, State University of New York at Oswego, Oswego, NY 13126, USA d School of Innovation, Design and Engineering, Mälardalen University, SE-721 23 Västerås, Sweden a r t i c l e i n f o Keywords: Case-based reasoning Clinical decision support systems Medical informatics a b s t r a c t This paper presents four synergistic systems that exemplify the approaches and benefits of case-based reasoning in medical domains. It then explores how these systems couple Artificial Intelligence (AI) research with medical research and practice, integrate multiple AI and computing methodologies, lever- age small numbers of available cases, reason with time series data, and integrate numeric data with con- textual and subjective information. The following systems are presented: (1) CARE-PARTNER, which supports the long-term follow-up care of stem-cell transplantation patients; (2) the 4 Diabetes Support System, which aids in managing patients with type 1 diabetes on insulin pump therapy; (3) Retrieval of HEmodialysis in NEphrological Disorders, which supports hemodialysis treatment of patients with end stage renal disease; and (4) the Mälardalen Stress System, which aids in the diagnosis and treatment of stress-related disorders. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction Case-based reasoning (CBR) systems have long found fertile ground in health sciences domains (Begum, Ahmed, Funk, Xiong, & Folke, 2011; Bichindaritz & Marling, 2010; Montani, 2008). Eight international Workshops on CBR in the Health Sciences have high- lighted the challenges and showcased the applications of CBR in biomedical fields. At the 2012 workshop, held at the International Conference on Case-Based Reasoning (ICCBR-12) in Lyon, several exemplary systems were featured (Lamontagne & Recio-García, 2012). In the spirit of CBR, which promotes reasoning and learning from concrete examples, four of these systems were selected as cases of medical CBR systems. These systems are: � CARE-PARTNER, which supports the long-term follow-up care of stem-cell transplantation patients. � The 4 Diabetes Support System, which aids in managing patients with type 1 diabetes on insulin pump therapy. � Retrieval of HEmodialysis in NEphrological Disorders, which supports hemodialysis treatment of patients with end stage renal disease. � The Mälardalen Stress System, which aids in the diagnosis and treatment of stress-related disorders. In this paper, we present each of these systems in turn. Then we explore the synergies enabled and exemplified by these systems. We find a tight coupling of Artificial Intelligence (AI) research with medical research and practice. We see integration of multiple AI and computing technologies. We find that complex domains de- mand complex knowledge structures. We identify a need to fully leverage small numbers of available cases. We encounter time ser- ies data and develop new ways to harness it for reasoning. We inte- grate numeric data, including biosensor signal data, with contextual life-event data and subjective patient perceptions. In essence, the synergistic intertwining of CBR and medicine in these systems has led to new insights in both CBR research and develop- ment and medical practice. It is our hope that these experiences will be retrieved, reused, revised and retained (Aamodt & Plaza, 1994) for future CBR research and system development. 2. CARE-PARTNER CARE-PARTNER is a decision support system for the long-term follow-up of oncology patients who have undergone stem cell transplantation (Bichindaritz, Kansu, & Sullivan, 1998). This system was built between 1996 and 2000 at the Fred Hutchinson Cancer Research Center, at the University of Washington, in Seattle. Three physicians, Keith Sullivan, Paul Martin, and Emin Kansu, and a phy- sician assistant, Muriel Siadak, served as the domain experts. While CARE-PARTNER is no longer an active project, the ideas it pio- neered have been carried over into the ongoing Mémoire project 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.05.063 ⇑ Corresponding author. Tel.: +1 740 593 1246. E-mail addresses: marling@ohio.edu (C. Marling), stefania.montani@unipmn.it (S. Montani), ibichind@oswego.edu (I. Bichindaritz), peter.funk@mdh.se (P. Funk). Expert Systems with Applications 41 (2014) 249–259 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy (Bichindaritz, 2006, 2007) and have influenced numerous other CBR systems in health sciences domains. 2.1. CARE-PARTNER system functionality and goals CARE-PARTNER assists clinicians with the long-term follow-up (LTFU) of stem cell transplant patients once they have returned to their home communities. It provides online answers to ques- tions from home care providers, who previously had to telephone nurses, who would then relay their questions to LTFU clinicians be- fore getting back to them with clinical answers. The electronic con- tact management system developed to replace the old phone and paper-based system provides advantages for research and docu- mentation and serves as an example of a medical knowledge man- agement system. CARE-PARTNER’s decision-support is based upon proven and validated practice, helping to implement evidence-based medicine (Sullivan et al., 1997). It provides the following types of decision support: � Interpretation for each laboratory test and procedure result. � List of differential diagnoses, ranked by likelihood; these diag- noses are often not incompatible, since several diagnoses co- occur to cover all the signs and symptoms exhibited by the patient. � List of steps of laboratory tests and/or procedures for diagnostic assessment. � List of steps of planning actions for treatment. � List of pertinent documents hyperlinked to the previous ele- ments, such as guidelines, or textbook excerpts. An important system requirement for CARE-PARTNER is the management of risk. A physician not specialized in a domain may not be able to critique or challenge system advice, and may not notice even severe mistakes. In the domain of stem-cell trans- plantation, transplant complications were quite unfamiliar to home care providers. These complications can be rapidly life- threatening, thus imposing very high standards of safety to protect patients. Therefore, the reliability and safety of the system were of paramount importance. 2.2. CARE-PARTNER system design Fig. 1 shows CARE-PARTNER’s reasoning cycle. This multimodal reasoning cycle combines case-based reasoning, rule-based rea- soning, and information retrieval. CARE-PARTNER’s reasoning steps are generalizations of the steps defined in these respective methodologies. The cooperation of the different knowledge sources is driven by the LTFU domain, in which, as in most medical domains, knowledge takes several forms: 1. Practice guidelines: A practice guideline is composed of sys- tematically developed textual statements, designed for practi- tioners and patients, which will be helpful in making clinical decisions on the prevention, diagnosis, treatment and manage- ment of selected conditions. Guidelines are represented as rules that are embedded in prototypical cases. 2. Practice pathways: A practice pathway covers the same type of knowledge elements as a practice guideline, but it is spe- cialized to the LTFU domain. While practice guidelines are rep- resented via text, practice pathways are expressed in the knowledge representation formalism of the decision support system. Practice pathways were created by a group of LTFU experts exclusively for the CARE-PARTNER system. Pathways correspond to prototypical cases, and they are represented as cases in the system. 3. Practice cases: A practice case is an example of a problem-solv- ing situation as solved by an expert or possibly a group of experts. It is essentially a real patient problem-solving situa- tion, and not a prototypical one as for a practice guideline or a practice pathway. It is represented as a case in the system. 4. Medical textbooks: A collection of documents serves as docu- mentation and explanation during the reasoning process, often in hyperlinked form. Intensive knowledge elicitation efforts were required to build the case base around a knowledge base of the domain. It was deter- mined early on in the project that cases were not available in elec- tronic format at a level of detail required for CBR. For instance, the patient database did not include patient treatments, or most of the signs and symptoms, but only the main events abstracted from the paper charts. The project team had to come up with prototypical cases to bootstrap the system, which took over two years to devel- op at a level of thoroughness and consistency needed to achieve high quality decision support. This system was unique because its proposed recommenda- tions spanned not only diagnosis, but also lab result interpreta- tion, and treatment planning. An extensive ontology was developed including 1109 diseases, 452 functions (also known as signs and symptoms), 1152 labs, 547 procedures, 2684 med- ications, and 460 sites. Most of the terms naming these objects were standardized using the Unified Medical Language System (UMLS) semantic network (National Library of Medicine, 1995). Only terms not corresponding to objects in the UMLS were added to the domain specific ontology. In particular, the planning actions used in the treatment part of a prototypical case did not exist in the UMLS and were all created for the system. The cornerstone of the knowledge acquisition process was the conception of prototypical cases, or clinical pathways. The 91 implemented clinical pathways primarily correspond to clinical diagnostic categories, with some of them corresponding to essen- tial signs and symptoms requiring specific assessment or treat- ment actions. The clinical pathways are knowledge structures represented using the ontology described above. A prototypical case comprises three parts: 1. A list of findings, corresponding to signs and symptoms. 2. A diagnosis assessment plan, which is a plan to follow for con- firming (or informing) the suspected diagnosis. 3. A treatment/solution plan, which is a plan to follow for treating this disease when confirmed, or a solution when the pathway does not correspond to a disease. The diagnosis assessment part and the treatment part of the case can also be viewed as simplified algorithms, since they use if-then-else structures, loop structures, and sequence structures of actions in time. When instantiated with an actual patient’s data, this provides a diagnosis assessment plan or treatment plan tai- lored to the specific patient. In this way, the knowledge structure allows for sophisticated adaptation when reusing a prototypical case. 2.3. CARE-PARTNER evaluation Table 1 shows the results of an evaluation in which two expert clinicians rated the system using the scale Fails to Meet Standards/ Adequate/Meets All Standards (Bichindaritz, 2006). This evaluation covered 163 different clinical situations or cases, corresponding to contacts between the system and a clinician, for three patients. As seen in Table 1, the system was rated 82.2% of the time as Meets all Standards and 12.3% of the time as Adequate, for a total of 94.5% 250 C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 Author's personal copy of results judged to be at least clinically acceptable by the medical experts. Table 1 also shows that the advice provided by the system covers most of the clinicians’ tasks: lab results interpretation, pro- cedure results interpretation, diagnosis, treatment, and pathways retrieval. Another part of the evaluation dealt with measuring the pro- gress of the system when solving new contact cases. This ability of the system to learn was evaluated on the complete charts of three different patients. The performance of the system signifi- cantly improved between patients 1 and 3 to reach 98.6% accept- ability for the 54 contacts in the third patient’s chart (Bichindaritz, 2006). Since the Therac-25 accidents had recently shown that not even 100% reliability was sufficient to ensure pa- tient safety (Leveson & Turner, 1993), the system was further ex- tended to include a safety control module capable of referring cases requiring particular attention to direct clinician supervision. 3. The 4 Diabetes Support System (4DSS) The 4 Diabetes Support System (4DSS) aims to assist patients with type 1 diabetes (T1D) and their professional caregivers (Marling, Shubrook, & Schwartz, 2009, 2011a, 2012). Work on this system began in 2004 and continues to this day. There are two do- main experts: Frank Schwartz, an endocrinologist; and Jay Shu- brook, a diabetologist. They are practicing clinicians, who have treated several hundred T1D patients, as well as faculty members of the Ohio University Heritage College of Osteopathic Medicine. 3.1. The diabetes management domain There are an estimated 346 million people worldwide who have diabetes (World Health Organization, 2012). Approximately 20 million of them have type 1 diabetes, the most severe form, in which the pancreas fails to produce insulin. Because insulin is an essential hormone needed to convert food into energy, T1D pa- tients depend upon external supplies of insulin. The T1D patients involved in the 4DSS project are treated with insulin pump ther- apy. An insulin pump continuously infuses the patient with basal insulin. The patient may instruct the pump to deliver additional insulin boluses to account for meals or blood glucose excursions. While diabetes can not yet be cured, it is actively managed through blood glucose control. Good blood glucose control is known to help delay or prevent long-term diabetic complications, including blindness, amputations, kidney failure, strokes, and heart attacks (Diabetes Control & Complications Trial Research Group, 1993). Effective blood glucose control entails vigilant self-monitor- ing of blood glucose levels. T1D patients prick their fingers from 4 to 6 times a day and use glucometers to measure their blood. They may also wear continuous glucose monitoring (CGM) devices, which produce blood glucose readings every 5 min. Blood glucose monitoring data is relayed to physicians, who must manually interpret it to find blood glucose control problems and recommend appropriate therapeutic adjustments. While volu- minous blood glucose data contributes to data overload for Working Memory Working Memory Working Memory Conflict Set Target Case Solution Physician Interface Interpretation Memorization Update Reuse Conflict resolution Search Problem Patient Electronic Medical Record Knowledge Base Cases, Pathways, Rules Fig. 1. CARE-PARTNER’s Reasoning Cycle. Table 1 CARE-PARTNER system evaluation results. Type of case Number of cases Fails to meet standards (%) Adequate (%) Meets all standards (%) Labs 54 3.7 3.7 92.6 Procedures 67 8.9 3.0 88.1 Diagnosis 68 16.2 13.2 70.6 Treatment 71 9.9 11.3 78.8 Pathways 47 8.5 8.5 83.0 Total 163 5.5 12.3 82.2 C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 251 Author's personal copy physicians, data concerning life events that impact blood glucose levels is not routinely maintained. Physicians may feel, paradoxi- cally, that they have too much data and yet not enough data at the same time. 3.2. 4DSS system overview The 4DSS is a hybrid case-based reasoning (CBR) system that detects problems in blood glucose control and suggests personal- ized therapeutic adjustments to correct them. Fig. 2 shows a graph- ical overview of the 4DSS. The system is data driven. Blood glucose data comes from glucometers and CGM devices. Insulin data comes from the patient’s pump. The patient uploads blood glucose and insulin data to Medtronic’s proprietary CareLink system (Medtron- ic, 2012), where it is extracted and transferred to the 4DSS data- base. The patient manually enters data about life events that impact blood glucose levels, including food, exercise, sleep, work, stress and illness. Originally entered via computer-based browsers, life-event data is now entered via smart phones. The situation assessment module scans patient data. Tradition- ally in CBR systems, situation assessment begins with a static description of the current problem. It is different in this domain, because blood glucose control problems continue over time, and because patients are not necessarily aware of problems when they occur. The 4DSS situation assessment module has three major components: problem detection, glycemic variability classification, and blood glucose prediction. These components were built using rule-based reasoning, machine learning algorithms, and time series prediction techniques, giving the 4DSS its hybrid character. The problem detection component contains 18 rule-based rou- tines that incorporate physician strategies for finding problems in patient data. At a high level, these routines look for problems involving: (1) hyperglycemia, or high blood glucose, which contrib- utes to long-term diabetic complications; (2) hypoglycemia, or low blood glucose, which may result in severe immediate reactions, including weakness, dizziness, seizure or coma; (3) fluctuations be- tween hyper-and hypoglycemia; and (4) lapses in diabetes self- care. The glycemic variability classication component assesses prob- lems involving blood glucose fluctuation. It detects the problem of excessive glycemic variability, which is believed to presage long- term complications caused by oxidative stress. When expert rules proved inadequate for detecting this problem, machine learning algorithms, including multi-layer perceptrons and support vector machines, were introduced. These algorithms classify 24-h blood glucose plots by variability level to match physician gestalt percep- tion of such plots. The blood glucose prediction component, currently under con- struction, incorporates time series prediction techniques to antici- pate problems before they occur. While blood glucose data is not currently available in real time and must be scanned retrospec- tively, we are preparing for its near-term availability. Predicting problems even 30 min in advance would give patients time to take preventative actions, enhancing patient safety. The situation assessment module reports the problems it finds to a physician, who must then select problems of interest. A se- lected problem triggers the case retrieval module of the 4DSS. The case retrieval module obtains the most similar cases from the 4DSS case base. The case base includes 80 cases, each of which contains a spe- cific blood glucose control problem experienced by a T1D patient, a physician’s recommended therapeutic adjustment for that prob- lem, and the clinical outcome for the patient after making the ther- apeutic adjustment. These cases were compiled during clinical research studies in which: (1) T1D patients contributed blood glu- cose, insulin and life event data; (2) physicians manually identified Fig. 2. Overview of the 4 Diabetes Support System. 252 C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 Author's personal copy blood glucose control problems and recommended solutions to pa- tients; (3) patients made the recommended therapeutic adjust- ments (or not); and (4) physicians examined subsequent patient data to determine the efficacy of the solutions. To retrieve the most similar cases from the case base, the case retrieval module employs a traditional two-step process. First, a subset of potentially similar cases is identified, and then the near- est neighbors are selected from that subset. In the first step, cases are partitioned by problem type. In the second step, a standard nearest neighbor metric is used. Domain specific similarity func- tions are combined with empirically determined weights to obtain an overall score for each case. Cases scoring above a similarity threshhold are forwarded to the adaptation module. The adaptation module personalizes a solution from a retrieved case to fit the situation of a current patient. It begins with the most similar case, but if the solution in that case is not adaptable, it con- siders the next most similar case, and so on. A solution is a thera- peutic adjustment composed of one or more actions that a patient can take. During adaptation, individual actions may be deleted or modified. For example, one possible action is to have a bedtime snack. If the current patient is already having an adequate bedtime snack, this action could be removed from the recommendation. In other situations, the advice could be modified so that the patient eats more or less food before bed, eats a different type of food be- fore bed, or has a snack at a different time of day. The adapted therapeutic adjustment is relayed to the physician as decision support. The physician decides whether or not to relay the recommendation to the patient. It has long been a goal to pro- vide low-risk advice directly to patients, in real-time, as well as to their physicians. However, this must remain a future goal until the safety and efficacy of the system is proven through clinical trials and approval is obtained from governmental regulatory agencies. 3.3. 4DSS evaluation Each component of the 4DSS is evaluated upon completion. These evaluations have provided proof of concept, illuminated sys- tem strengths and weaknesses, and guided system development. Note that a definitive clinical trial, assessing system impact on pa- tient outcomes, remains to be conducted. The problem detection component was evaluated after the first and second 4DSS clinical studies. In the first evaluation, a panel of diabetes practitioners rated a sampling of problem detections, and in the second, each patient’s physician rated all of the problem detections for the patient. In the first test, 77.5% of problem detec- tions were rated as correct (Marling, Shubrook, & Schwartz, 2008), while in the second, 97.9% were rated as correct (Schwartz, Ver- nier, Shubrook, & Marling, 2010). The glycemic variability classification component was also eval- uated twice. Here, ten-fold cross validation was used to determine the accuracy, sensitivity and specificity of each potential classifier, where correctness is defined as matching physician classifications. In an early test, a naive Bayes classifier matched physicians 85% of the time (Marling, Shubrook, Vernier, Wiley, & Schwartz, 2011b). The current best classifier, a multi-layer perceptron, has accuracy, sensitivity and specificity of 93.8%, 86.6%, and 96.6%, respectively (Wiley, Bunescu, Marling, Shubrook, & Schwartz, 2011). The case retrieval module was evaluated by a panel of diabetes practitioners after the first and second clinical studies. Leave-one- out testing was used to provide a sampling of case retrievals for evaluation. In the first test, evaluators rated the retrieved cases as similar to test cases 80% of the time and rated the retrieved solu- tions as beneficial for test patients 70% of the time (Marling et al., 2008). In the second test, they rated retrieved cases as similar 79% of the time and retrieved solutions as beneficial 82% of the time (Marling et al., 2011b). The adaptation module was evaluated by showing physicians sample problems, with both original and adapted solutions, and eliciting feedback on a questionnaire. Physicians rated the original solutions as being fine without adjustment 47% of the time, need- ing minor adjustment 40% of the time, and needing major adjust- ment 13% of the time. They judged the adapted solutions to be better than the original solutions 83% of the time. 4. Retrieval of HEmodialysis in NEphrological disorders (RHENE) Retrieval of HEmodialysis in NEphrological Disorders (RHENE) supports physicians working in the domain of end stage renal dis- ease (ESRD). Work on RHENE began in 2004 and continues to this day. The expert for RHENE is Roberto Bellazzi, a nephrologist at the Vigevano Hospital in Italy. 4.1. The end stage renal disease domain ESRD is a severe chronic condition that corresponds to the final stage of kidney failure. Without medical intervention, ESRD leads to death. Hemodialysis is used to treat ESRD patients. During hemodialysis, an electromechanical device called a hemodialyzer clears the patient’s blood of metabolites, re-establishes acid–base equilibrium and removes excess water. A single hemodialysis ses- sion typically lasts for four hours. On average, a hemodialysis pa- tient receives three treatment sessions per week. The hemodialyzer tracks several time series variables during each ses- sion, sampling each at intervals of from 1 to 15 min. These vari- ables are analyzed to assess the efficacy of the hemodialysis treatment session and to ensure that the patient’s treatment ad- heres to his or her therapy plan. Interpreting a hemodialysis session as a case, we must deal with cases with time series features. Interpreting time series features on screen or on paper can be tedious and prone to errors. Physicians may be asked to recognize small or rare irregularities in the series itself, or to identify partial similarities with past patient situations, which do not depend upon the individual values in the series. While extremely important for patient care, such identification crequires a significant amount of expertise. Therefore, having an automated data interpretation and decision support system is desirable. 4.2. RHENE system overview In time dependent domains, the need to describe process dynamics strongly impacts both case representation and case re- trieval, as analyzed in (Montani & Portinale, 2006). Most reported approaches to similarity-based time series retrieval are founded on the common premise of dimensionality reduction, which sim- plifies knowledge representation (see the survey in (Hetland, 2003)). Dimensionality is often reduced by means of a mathemat- ical transform – such as the Discrete Fourier Transform (Agrawal, Faloutsos, & Swami, 1993) – able to preserve or underestimate the distance between two time series. However, mathematical transforms have several limitations, as they can be computation- ally complex, and usually work in a black box fashion with respect to end users. In contrast, RHENE (Montani, Leonardi, Bottrighi, Por- tinale, & Terenziani, 2011) implements a framework for time series retrieval that exploits Temporal Abstractions (TA) (Shahar, 1997) to reduce time series dimensionality, with multi-dimensional in- dex structures to make retrieval efficient. Using TA allows greater interpretability of the output results and understandability of the retrieval process. It maps huge amounts of temporal information to a compact representation, C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 253 Author's personal copy by aggregating adjacent time series points sharing commonalities (e.g., the same qualitative level, the same trend direction) into a single interval, labeled by a symbol. This technique not only sum- marizes the original longitudinal data, but also highlights mean- ingful data characteristics in a clear, symbolic, high level view. The exploitation of TA for case exploration and retrieval, as well as for data interpretation, is, to the best of our knowledge, a novel contribution of RHENE. TA are typically applied for data interpreta- tion, but not for case retrieval. The basic principle of TA methods is to move from a point-based to an interval-based representation of the data (Bellazzi, Larizza, & Riva, 1998), where the input points (events) are the elements of the time series, and the output intervals (episodes) aggregate adjacent events sharing a common behavior, persistent over time. Episodes are identified by symbols. Basic abstractions can be further subdi- vided into state TA and trend TA. State TA are used to extract epi- sodes associated with qualitative levels of the monitored feature, e.g., low, normal, or high values. Trend TA are exploited to detect specific patterns, such as increase, decrease or stationary behavior, from the time series. RHENE supports multi-level abstractions of the original data. Time series values can be abstracted and queried at finer or coarser levels of detail, along two dimensions: a taxonomy of symbols, and a taxonomy of time granularities. For instance, a taxonomy of trend symbols can be introduced, in which the symbol I (increase) is fur- ther specialized into IW (weak increase) and IS (strong increase), according to the slope. As another example, a series of two adjacent intervals of IW, each having a duration of half an hour, can be merged into a single IW interval, with a duration of 1 h. To scale up from two or more values expressed at a finer granularity to a single value expressed at a coarser one, an up function is provided. This function is domain dependent, but obeys some general con- straints, including ‘‘persistence’’ (the result of coarsening two gran- ules with the same symbol x is a larger granule still labeled as x) and ‘‘monotonicity’’ (ordering among symbols, if any, is preserved) (Montani et al., 2011). Once abstracted, two time series can be compared to each other, enabling the retrieval of similar cases. Retrieval depends upon a distance metric that measures the similarity between symbols in the taxonomy. Different distance functions can be employed, as long as the distance of each symbol from itself is zero and other distances are consistent with respect to any symbol ordering. A query language was developed to facilitate case retrieval. Queries are expressed as sequences of symbols at different levels of detail. RHENE supports ground queries, i.e., queries composed of symbols at the lowest abstraction level in both taxonomies, and abstract queries, including those with symbols at different lev- els in the symbol taxonomies. Queries as regular expressions are also supported, making the query process flexible and user friendly. To increase retrieval efficiency, RHENE includes an indexing strategy, which exploits bi-dimensional taxonomic indexes. A for- est of index structures provides a flexible indexing of cases at dif- ferent levels of the symbol and/or time granularity taxonomies. The root node of each index structure is represented by a string of symbols, defined at the highest level in both dimensions. An example is shown in the right box of Fig. 3. Here, the root node I is refined along the leading time dimension from the 4-h to the 2-h granularity, so that the nodes I I, I S and S I stem from it, pro- vided that up(I,S) = I and up(S,I) = I, where S stands for stationarity. From each node of the leading dimension structure, another index can stem, keeping the time granularity abstraction level fixed. The index develops orthogonally with respect to the leading dimension. In summary, RHENE operates in the following manner, as illus- trated by Fig. 3: � Time series features in ESRD cases (i.e., hemodialysis sessions) are pre-processed by a TA server at the ground level of the sym- bol and time granularity taxonomies. � Pre-processed cases are then stored in a relational database. � When the physician needs to evaluate a patient, he or she can query the database to retrieve similar cases at any level of detail. � Indexes enable quick and interactive query answering. � Retrieved cases are then shown to the physician as decision support. The physician remains responsible for making deci- sions regarding the patient’s therapy. 4.3. RHENE evaluation RHENE was experimentally evaluated, using a dataset of 10,388 actual hemodialysis sessions (i.e., ESRD cases) collected at the Vigevano Hospital in Italy. The TA approach was compared to the more classical approach implemented in an earlier version of RHENE (Montani, Portinale, Leonardi, Bellazzi, & Bellazzi, 2006). That approach was based on DFT for dimensionality reduction and on spatial indexing through TV-trees for improving retrieval performance. Quantitative experiments were run to compare the query answering time required by the two approaches, as well as their scalability when dealing with a case base progressively grow- ing in size. Subsets of from 2,000 to 10,388 actual cases were used to evaluate each system’s performance. As shown in Fig. 4, the TA- based method proved to be much more efficient in query answer- ing than the DFT-based method, with query times under 1 s. A qualitative comparison of the two approaches was also made, through examination of individual case studies. Details of this examination are available in (Montani et al., 2011). In brief, it was found that the ability to consider trends was advantageous for TA-based retrieval. DFT sometimes missed relevant cases, be- cause it considers only point-to-point distances between cases, looking for the best overall alignment. Shifts along an axis, such as lower absolute values, could prevent DFT from finding cases with similar trends. Furthermore, the TA-based approach provided the user with more easily understood queries and results than DFT’s more mathematical black box approach. 5. The Mälardalen Stress System (MSS) The Mälardalen Stress System (MSS) provides decision support for the diagnosis and treatment of stress (Ahmed, Begum, & Funk, 2012, 2011). This system was built between 2002 and 2011 at Mälardalen University in Västerås, Sweden. The experts for this system were psychologists Bo von Schéele and Erik Olsson. Dr. von Schéele has over 30 years’ experience in clinical stress diagno- sis and has pioneered new biosensor and biofeedback methods. 5.1. The stress management domain Stress is a factor of daily life that can adversely impact health and wellbeing. While the causes of stress can not typically be elim- inated, patients can be taught to effectively deal with stress, min- imizing its ill effects. A clinical measurement of stress is Finger Temperature (FT). By placing a sensor on a patient’s fingertip, a continous temperature profile can be obtained. In general, lower finger temperatures are indicative of greater stress. However, changes in FT vary from patient to patient, and correctly interpret- ing and analyzing them requires knowledge and experience. More- over, additional factors, such as the patient’s feelings, behaviors, environment and lifestyle, also play a role in stress management. Contextual features, indicating a patient’s perception of stress, are collected as text and also via Visual Analog Scale (VAS) input. 254 C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 Author's personal copy VAS is used to measure subjective characteristics or attitudes on a scale of 0 to 10. Both sensor signals and textual information are used to diagnose a patient’s level of stress. Biofeedback training is used as treatment to control stress. Dur- ing biofeedback training, a patient can alter his or her physiological or psychological state while observing the changes in FT on a graph. As the patient sees how psychophysiological change is rep- resented on the graph, he or she can train the body and/or mind to change the biological response to stress. Relaxation exercises may be recommended to aid the patient in making positive changes. 5.2. MSS system overview The MSS demonstrates how some common CBR techniques need to be modified in order to provide clinicians with effective decision support based on past cases relevant to the current pa- tient. This clinical usage adds a number of specific requirements. For example, in addition to considering case similarity, it is impor- tant to consider case outcomes. A case with a severe outcome may be very relevant, as a clinician may need to take precautions to avoid a similar result. An example is that most patients’ finger tem- peratures decrease during stress and increase during relaxation, a normal reaction from the parasympathetic nervous system. But in some patients, the effect on finger temperature may differ due to rare conditions the clinician must consider during diagnosis. If the current patient is similar to many past patients diagnosed as not at risk for stress-related illnesses and similar to one patient having a condition with severe health consequences if left un- treated, the latter case is of most interest to the clinician, who may take additional measurements to rule this case out. To meet the needs of the medical domain, the case has been structured with the following components: (a) symptom features, including sensor readings and contextual data; (b) diagnosis; (c) action taken, i.e., treatment and biofeedback training; and (d) out- come, e.g., how quickly the patient improves. The case may also contain (e) comments, which can be added by a clinician who wishes to share perceptions and/or important references with col- leagues. Comments may be added by the clinician responsible for the case or by a domain expert. The symptom features for a case are derived from a calibration protocol in which the patient is asked to complete a series of stressful, relaxing and neutral tasks while FT measurements are collected (Ahmed, Begum, Funk, Xiong, & von Schéele, 2011). This calibration phase is necessary, because sensor readings vary widely from patient to patient. Sensor readings that are normal for one person may be alarming for another individual. The patient is also asked to supply contextual data via text and VAS input. A total of 25 sensor and contextual features comprise this section of a case. The diagnosis part of the case holds the stress classification of the patient, which may be: VeryRelaxed, Relaxed, Normal/Stable, Stressed or VeryStressed. A confidence level for the classification, denoted as High, Medium or Low, is also recorded for each diagno- sis. The actions taken comprise the treatment, and the outcome is the result of the treatment, i.e., the degree of improvement after treatment. Fig. 3. RHENE System Overview. Fig. 4. Average time, in seconds, to answer a query as the number of cases increases, for DFT and TA. C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 255 Author's personal copy Biofeedback is a tool used by clinicians in treatment and overall stress reduction. It is included in the section of the case for action taken. The latest diagnosis is used as one of the parameter settings for biofeedback. The patient gets feedback on parameter improve- ment. Another important parameter is recovery time after stress, i.e., the time it takes for a patient to recover after a stress-phase. This parameter can be used to select relaxation exercises/treat- ment for the patient. During the exercise, the biofeedback system monitors the patient and calculates the recovery time after stress. This enables patients to practice biofeedback either with a clinician at hand or alone with a computer system and sensors. The relaxa- tion exercise/treatment is selected to achieve good results, based on how good the effect was on previous similar patients. If the recovery time does not change with the selected exercise, then a different exercise is selected. Exercise selection is based on past cases and the result of the exercise/treatment. Different exercises have different effects on patients. Selecting exercises and following patient progress in this way increases the efficiency of the exercise. The MSS supports stress diagnosis and treatment in three phases: (1) analyze and classify a patient and make a risk assess- ment; (2) determine individual levels and parameters; and (3) adapt and conduct biofeedback training. CBR is used as the core technology for the system, as illustrated in Fig. 5. Knowledge dis- covery contributes part of the domain knowledge used in case re- trieval. Since patients in the case library have already been diagnosed/classified, these cases can be used to identify the fea- tures that are most important for case comparison and retrieval (Funk & Xiong, 2006). Other AI techniques, including fuzzy logic, rule-based reasoning and textual information retrieval, are also incorporated in supporting roles. 5.3. MSS evaluation The MSS system has been evaluated using a number of different methods. Diagnostic performance has been compared with that of clinicians, including clinical experts and junior clinicians with lim- ited practice in diagnosing stress using sensor readings. Using mul- tiple test sets, the system was able to classify over 80% of cases correctly, compared to between 57% and 69% for junior clinicians and 73% for a senior clinician (Ahmed et al., 2012). In comparing system outputs to clinician diagnoses, it was important to consider the consistency of clinician responses when confronted with the same patient more than once. In another test of diagnostic perfor- mance, leave one out cross validation was used. In this evaluation, cases were removed from the case library one at a time to see how well the system could classify each one. Using a fuzzy matching algorithm, the MSS was able to perform close to expert level, as re- ported in (Begum, Ahmed, Funk, Xiong, & von Schéele, 2009). In addition to judging overall accuracy, it was important to consider the sensitivity and specificity of the system. From a clinical per- spective, missing a stressed patient (false negative) is less accept- able than identifying a healthy individual as stressed (false positive). A false positive would likely be identified as such before beginning treatment, based on additional pre-treatment evaluation. 6. Research synergies and trends In this section, we elaborate on the research synergies and trends highlighted by CARE-PARTNER, the 4DSS, RHENE, and the MSS. 6.1. Coupling AI research with medical research and practice All four systems were built through strong collaborations with medical researchers and clinicians. While we have focused above on AI research and development, there has been a tight coupling with medical research and practice, as well. For example, CARE- PARTNER supported evidence-based practice for oncology. At the time of the original study, stem cell transplantation was a very new procedure, and there were no established guidelines for the long-term follow-up care of patients after they left the cancer cen- ter. By compiling cases of patient problems, solutions, and out- comes, CARE-PARTNER aided in defining best clinical practices. The 4DSS, RHENE, and the MSS can all be characterized as promot- ing the practice of personalized medicine. These systems look at how individual patients respond to diabetes therapy, hemodialysis, Fig. 5. Overview of the Mälardalen Stress System. 256 C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 Author's personal copy and stress, respectively, and suggest personalized therapeutic adjustments based on individual patient needs. The 4DSS contributes to medical research in the areas of glyce- mic variability (GV) measurement and blood glucose prediction. Because there was no accepted metric for GV in routine clinical use, but 4DSS experts wanted to detect excessive GV, new metrics were developed for the 4DSS. These metrics were subsequently published and presented to the diabetes technology community (Schwartz et al., 2010), and the 4DSS project is now leading the way in developing a consensus GV metric for routine clinical use. Current research on blood glucose prediction not only aids in intel- ligent decision support, but potentially aids in the development of an ‘‘artificial pancreas,’’ a project of the Juvenile Diabetes Research Foundation. RHENE supports medical practice and research as well. Hemod- ialysis sessions are ordinarily judged only on the basis of macro- scopic observations of time series features. RHENE provides a deeper insight into the clinical situation, highlighting types of anomalies which, if not leading to immediate hemodialysis failure, could produce poor therapeutic results in the long run. (See, for example, the experiments in Montani et al. (2006)). Moreover, the use of RHENE may lead to the identification of systematic map- pings between TA trend behaviors (e.g., an increase in diastolic pressure) and specific pathologies or complications, in an applica- tion domain where this kind of knowledge does not currently exist. Expert clinicians with the MSS project found that their work on the system helped them to refine their own methods and ap- proaches for diagnosing patients. They also found that the system- atic analysis of symptoms required by the system improved the quality and value of the electronic health records (EHRs) that they maintained for their patients. The final outcomes of treatments had not always been recorded in detail prior to system development. As tracking outcomes was essential for case-based reasoning, the EHRs for patients of participating physicians became more com- plete. Furthermore, the close collaboration made it more explicit how clinical experts, while reading symptoms and measurements, would recognize similarities to past patients that would aid in diagnosis and treatment. This led to the discovery of new features that were valuable in the diagnosis process (Funk & Xiong, 2006, 2008). 6.2. Integrating multiple AI and computing methodologies All four systems are multi-modal, or hybrid, systems, which synergistically combine CBR with other AI and computing method- ologies. CARE-PARTNER combines CBR with rule-based reasoning and information retrieval. Its knowledge base encompasses both theoretical knowledge and experiential knowledge, which may be expressed either in a controlled vocabulary or textually. The 4DSS integrates rule-based reasoning for problem detection and case adaptation, multiple machine learning algorithms for glyce- mic variability classification, and support vector regression for blood glucose prediction. RHENE incorporates temporal abstrac- tions for representing, comparing, and retrieving cases. The MSS leverages rule-based reasoning, fuzzy logic, and information retrie- val. Fuzzy rule-based classification is used when the domain knowledge is precise (e.g., if X and Y are high, then we know that Z is high) but the terminology used is not precise (e.g., the same absolute value may be high or low, depending on context). Infor- mation retrieval is used to handle the free text given in the case, e.g., the patient history. Driven by the demands of specific medical domains, these inte- grations push the envelope of how knowledge-based systems can be engineered. For example, hemodialyzers generate data that could not be readily reasoned about without some form of feature reduction or abstraction. So, RHENE introduced a novel combina- tion of TA and CBR that could be extended to other domains having time series data. Similarly, CARE-PARTNER’s early information re- trieval integration has become increasingly relevant as information from electronic health records and large medical corpora become increasingly available. 6.3. Leveraging small numbers of available cases In this era of Big Data, there is much emphasis on extracting useful information from large volumes of available data. For med- ical decision support applications, however, it is still unusual to have all of the relevant data at hand. Although RHENE did have thousands of cases to work with, CARE-PARTNER, the 4DSS, and the MSS effectively leveraged small numbers of available cases. CARE-PARTNER, for example, began with an incomplete patient database containing extracts from paper-based patient charts. Ninety-one prototypical cases were built through extensive knowl- edge engineering efforts, enabling domain knowledge to supple- ment the available data. The 4DSS began with an abundance of blood glucose data, but the contextual life events needed to inter- pret it were not routinely maintained. A series of three clinical re- search studies was conducted, in which 80 problem/solution/ outcome cases were developed from actual patient data. Each case contains over 140 data elements, hierarchically organized, to rep- resent a problem solving experience. The MSS project conducted clinical studies to capture and represent patient stress profiles. However, case coverage was initially incomplete, because some profiles, such as the Very Relaxed profile, were not well repre- sented. Artificial cases were generated to augment the actual pa- tient cases by using generalized sensor features and a fuzzy inference system. This improved the system’s ability to retrieve applicable cases (Ahmed, Begum, Funk, & Xiong, 2009). While there are many ways to leverage large numbers of exemplars, the ability to solve problems with just a few knowledge-rich cases is a strength of the CBR approach. 6.4. Reasoning with time series data All four systems reason about cases that develop over time. This is in contrast to most CBR systems, in which a snapshot of current conditions is sufficient for assessing and reasoning about the prob- lem at hand. CARE-PARTNER, for example, is for long-term follow- up care, and the patient’s past history informs current care. Time series data for the 4DSS comes in the form of blood glu- cose sensor data, collected for up to 90 days at 5 min intervals, jux- taposed with the life-event and insulin data reported over the same period of time. For problem detection, rule-based pattern recognition routines, based on expert strategies, were imple- mented. For glycemic variability classification, several domain dependent and independent features are extracted from the sensor data, for use by classification algorithms. Blood glucose prediction is tackled as a time series forecasting problem, using support vec- tor regression. RHENE’s time series data comes from the hemodialyzer, over a 4-h period, in increments of from 1 to 15 min. RHENE has the tight- est integration with time series data, using TA to represent, com- pare, and retrieve cases. Time series data for the MSS comes from finger temperature sensors, which are worn by patients during 15-min calibration sessions. The system extracts features from the sensor data, including the recovery rate after stress, the slope during periods of stress, and the difference between relaxed and stressed finger temperature readings. The extracted features are used to build cases for the system. C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 257 Author's personal copy 6.5. Integrating numeric data with contextual and subjective information In medical domains, it is often necessary to consider hard nu- meric data in light of contextual information and subjective patient perceptions. For one thing, the world around the patient impacts the patient’s health. For another, therapeutic recommendations that do not suit a patient’s individual lifestyle and preferences may be ignored, no matter how medically advisable. In the 4DSS, numeric blood glucose data is considered in light of the life events that influence blood glucose levels, including when and what the patient eats, when and how intensely the patient exercises, whether the patient is at home or at work, awake or asleep, feels stressed or ill, and so on. The efficacy of a recommended therapeu- tic intervention is evaluated in light of whether or not the patient actually implemented the recommendation. In the MSS, finger temperature sensor signals that are indicative of stress are consid- ered together with the factors in the patient’s life that could cause the stress. CBR allows the flexible use of disparate types of data within cases, suiting it to these medical domains. 7. Summary and conclusion In this paper, we have presented four synergistic systems that exemplify the approaches and benefits of case-based reasoning in medical domains. These systems are CARE-PARTNER, the 4 Diabe- tes Support System, Retrieval of HEmodialysis in NEphrological Disorders, and the Mälardalen Stress System. We have explored how these systems couple AI research with medical research and practice, integrate multiple AI and computing methodologies, leverage small numbers of available cases, reason with time series data, and integrate numeric data with contextual and subjective information. We hope that these cases of medical CBR systems will inform future medical CBR research and development. Acknowledgements Work on CARE-PARTNER was supported by Grant R01HS09407 from the Agency on Health Care Policy and Research (AHCPR), and by a Scholarship Grant from the University of Washington, Tacoma. Work on the 4DSS is supported, in part, with funding from the National Science Foundation under the Smart Health and Wellbe- ing program (award IIS-1117489). Additional support comes from Medtronic, the Ohio University Russ College Biomedical Engineer- ing Fund, the Ohio University Heritage College of Osteopathic Med- icine Research and Scholarly Affairs Committee, and the Ohio University Diabetes Research Initiative. Work on the RHENE system has been supported with funding from the Italian Ministry of Education (within the project MIUR PRIN – 2004–2006 – Intelligent Analysis of Hemodialysis Monitor- ing Data to Improve Care Processes), and by Regione Piemonte (within the project Regione Piemonte-Polo Innovazione ICT – 2010–2012 – MASP: Ricerca e Sviluppo Sperimentale di un Sistema per il Controllo a Distanza di Aree Sensibili e Protette, Finalizzato all’Erogazione di Servizi Innovativi Orientati al Monitoraggio Cog- nitivo ed Attuativo di Politiche Ambientali). Work on the MSS has been supported by funding from: the Swed- ish Knowledge Foundation Project AIM, Artificial Intelligence for medical application: Diagnosing stress based on psychophysiologi- cal sensors, project number 2001/0068 (2002–2004); IPOS, Intelli- gent integrated sensor systems for diagnosis, treatment and healthcare, project number KKS HÖG 2004/0341 (2005–2008); Improved productivity and life quality through reduced stress, Sparbanksstiftelsen Nya, project reference 2081 (2009); and Multi- sensors for stress monitoring, NovaMedTech, FP7 68737 (2008– 2011). All four systems were made possible through the extensive ef- forts of many collaborators, including domain experts, clinical staff, graduate research assistants, and patients. The authors gratefully acknowledge their contributions and support. References Aamodt, A., & Plaza, E. (1994). Case-based reasoning: Foundational issues, methodological variations, and system approaches. AI Communications, 7, 39–59. Agrawal, R., Faloutsos, C., & Swami, A. (1993). Efficient similarity search in sequence databases. In D. Lomet (Ed.), Proceedings of the 4th interational conference of foundations of data organization and algorithms (pp. 69–84). Springer. Ahmed, M., Begum, S., & Funk, P. (2012). A hybrid case-based system in stress diagnosis and treatment. In IEEE–EMBS International Conference on Biomedical and Health Informatics (BHI2012). Ahmed, M., Begum, S., Funk, P., & Xiong, N. (2009). Fuzzy rule-based classification to build initial case library for case-based stress diagnosis. In Proceedings of the 9th international conference on artificial intelligence and applications (AIA 2009) (pp. 225–230). Ahmed, M. U., Begum, S., Funk, P., Xiong, N., & von Schéele, B. (2011). A multi- module case-based biofeedback system for stress treatment. Artificial Intelligence in Medicine, 51, 107–115. Begum, S., Ahmed, M. U., Funk, P., Xiong, N., & Folke, M. (2011). Case-based reasoning systems in the health sciences: A survey of recent trends and developments. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, 41, 421–434. Begum, S., Ahmed, M. U., Funk, P., Xiong, N., & von Schéele, B. (2009). A case-based decision support system for individual stress diagnosis using fuzzy similarity matching. Computational Intelligence, 25, 180–195. Bellazzi, R., Larizza, C., & Riva, A. (1998). Temporal abstractions for interpreting diabetic patients monitoring data. Intelligent Data Analysis, 2, 97–122. Bichindaritz, I. (2006). Mémoire: A framework for semantic interoperability of case bases in biology and medicine. Artificial Intelligence in Medicine, 36, 177–192. Bichindaritz, I. (2007). Prototypical case mining from biomedical literature. Applied Intelligence, 28, 222–237. Bichindaritz, I., Kansu, E., & Sullivan, K. M. (1998). Case-based reasoning in CARE- PARTNER: Gathering evidence for evidence-based medical practice. In B. Smyth & P. Cunningham (Eds.), Advances in case-based reasoning: Fourth European workshop. Proceedings EWCBR-98 (pp. 334–345). Berlin: Springer-Verlag. Bichindaritz, I., & Marling, C. (2010). Case-based reasoning in the health sciences: Foundations and research directions. In I. Bichindaritz, S. Vaidya, A. Jain, & L. C. Jain (Eds.), Computational intelligence in healthcare 4: Advanced methodologies (pp. 127–158). Berlin: Springer. Diabetes Control and Complications Trial Research Group (1993). The effect of intensive treatment of diabetes on the development and progression of long- term complications in insulin-dependent diabetes mellitus. New England Journal of Medicine, 329, 977–986. Funk, P., & Xiong, N. (2006). Case-based reasoning and knowledge discovery in medical applications with time series. Computational Intelligence, 22, 238–253. Hetland, M. (2003). A survey of recent methods for efficient retrieval of similar time sequences. In M. Last, A. Kandel, & H. Bunke (Eds.), Data mining in time series databases. London: World Scientific. Lamontagne, L., & Recio-García, J.A. (Eds.). (2012). Proceedings of the ICCBR 2012 Workshops. Lyon. Leveson, N. G., & Turner, C. S. (1993). An investigation of the Therac-25 accidents. IEEE Computer, 26, 18–41. Marling, C., Shubrook, J., & Schwartz, F. (2008). Case-based decision support for patients with type 1 diabetes on insulin pump therapy. In K.-D. Althoff, R. Bergmann, M. Minor, & A. Hanft (Eds.), Advances in case-based reasoning: 9th European conference. ECCBR (pp. 325–339). Berlin: Springer. Marling, C., Shubrook, J., & Schwartz, F. (2009). Toward case-based reasoning for diabetes management: A preliminary clinical study and decision support system prototype. Computational Intelligence, 25, 165–179. Marling, C., Wiley, M., Bunescu, R., Shubrook, J., & Schwartz, F. (2012). Emerging applications for intelligent diabetes management. AI Magazine, 33, 67–78. Marling, C., Wiley, M., Cooper, T., Bunescu, R., Shubrook, J., & Schwartz, F. (2011a). The 4 diabetes support system: A case study in CBR research and development. In A. Ram & N. Wiratunga (Eds.), Case-based reasoning research and development: 19th international conference on case-based Reasoning. ICCBR 2011 proceedings (pp. 137–150). Berlin: Springer. Marling, C. R., Shubrook, J. H., Vernier, S. J., Wiley, M. T., & Schwartz, F. L. (2011b). Characterizing blood glucose variability using new metrics with continuous glucose monitoring data. Journal of Diabetes Science and Technology, 5, 871–878. Medtronic (2012). Carelink personal software. <http://www.medtronicdiabetes.net/ products/carelinkpersonalsoftware>. [accessed May, 2012]. Montani, S. (2008). Exploring new roles for case-based reasoning in heterogeneous AI systems for medical decision support. Applied Intelligence, 28, 275–285. Montani, S., Leonardi, G., Bottrighi, A., Portinale, L., & Terenziani, P. (2011). Supporting flexible, efficient and user-interpretable retrieval of similar time series. IEEE Transactions on Knowledge and Data Engineering. 25, 677–689. Montani, S., & Portinale, L. (2006). Accounting for the temporal dimension in case- based retrieval: A framework for medical applications. Computational Intelligence, 22, 208–223. 258 C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 Author's personal copy Montani, S., Portinale, L., Leonardi, G., Bellazzi, R., & Bellazzi, R. (2006). Case-based retrieval to support the treatment of end stage renal failure patients. Artificial Intelligence in Medicine, 37, 31–42. National Library of Medicine (1995). The unified medical language system. <http:// umls.nlm.nih.gov>. [accessed April, 2005]. Schwartz, F. L., Vernier, S. J., Shubrook, J. H., & Marling, C. R. (2010). Evaluating the automated blood glucose pattern detection and case-retrieval modules of the 4 diabetes support system. Journal of Diabetes Science and Technology, 4, 1563–1569. Shahar, Y. (1997). A framework for knowledge-based temporal abstractions. Artificial Intelligence, 90, 79–133. Sullivan, K. M., & Siadak, M. F. (1997). Stem cell transplantation. In F. E. Johnson, K. S. Virgo, S. B. Edge, C. A. Pellegrini, G. J. Poston, & S. P. Schantz, et al. (Eds.), Cancer patient follow-up (pp. 490–501). Louis, MO: Mosby Year-Book Publications. Wiley, M., Bunescu, R., Marling, C., Shubrook, J., & Schwartz, F. (2011). Automatic detection of excessive glycemic variability for diabetes management. In Proceedings of the tenth international conference on machine learning and applications (ICMLA-11) (pp. 148–154). New York, NY: IEEE Press. World Health Organization (2012). Diabetes. <http://www.who.int/mediacentre/ factsheets/fs312/en/index.html>. [accessed May, 2012]. Xiong, N., & Funk, P. (2008). Concise case indexing of time series in health care by means of key sequence discovery. Applied Intelligence, 28, 247–260. C. Marling et al. / Expert Systems with Applications 41 (2014) 249–259 259 
work_ae2kbpgag5fgdkr6lp3cvgavjy	----	A Study of the consideration of bio monitoring of trace Elements in the human nail for the Biometric Authentication N. Ambiga1*, Dr. A.Nagarajan2 1Research Scholar, Department of computer Applications, Alagappa University, karaikudi, Tamilnadu, INDIA 2Assistant professor, Department of computer Applications, Alagappa University, karaikudi, Tamilnadu, INDIA I.INTRODUCTION Human bio monitoring is the concept of monitoring the composition and concentration of chemical elements in the human body. It play a major role in identifying the relationships between the nutrition, health status, dietary intake and environmental exposure. The biological fluids such as blood, urine, sweat, exhaled breath condensate, saliva, earwax, body milk and tears, as well as in adipose tissue and, biological tissues such as hair and nail are analysed and measured for the existence of elements[1]. There are many analytical techniques at present available for checking the element composition and element concentrations in nails such as inductively coupled plasma mass spectrometry and atomic absorption spectrometry[2]. II. NAIL AS A BIO MARKER Biological tissues like nails are utilised as bio metrics for evaluating the environmental Reviewed By: Dr. Daniel V. Department: Medical ABSTRACT The studies of trace element composition in the human nail is called as human bio monitoring. The composition of the chemicals and trace elements and concentration of the chemicals and trace elements in the human body fluids and human body tissues are observed .The concentration of trace elements for a healthy human being is stable for a particular period of time in their nail and also the composition of chemical elements in the nail differs to each individual. So the humans nails can be used as the bio marker for monitoring or quantifying the composition of chemicals (major and trace elements) in the human body. So trace element's composition and concentrations in the nail can be considered for Bio-metrics authentication in future. Because the chemical composition of nail and chemical concentration in the human nail differ from one people to another people and so this can be used to identify the individual in future. Index terms–elements, trace elements, human bio monitoring, human body,nail I n n o v a t i v e J o u r n a l o f M e d i c a l H e a l t h S c i e n c e IJMHS 9 (9), 592–597 (2019) ISSN (O) 2277:4939 | (P) 2589:9341 ⋆ Corresponding author. † Email: ambiga2876@gmail.com. DOI: https://doi.org/10.15520/ijmhs.v9i9.2641 pollution level by learning chemical element composition in nail. The people exposed to environmental pollution can be known through the abnormality of chemical element concentrations in nail samples. the explanation for victimisation fingernails as diagnostic material is that trace components accumulated within the nails will replicate the future exposure to environmental pollution. Excluding that nails may be simply collected and hold on at traditional temperature before analysis and tiny amounts of samples (about ten mg) area unit needed for analysis[3]. The chemical compositions of nails might depend on gender, nutrition, occupation, age, unwellness and season[4]. Each fingernails and toenails are used as biomarkers, every has explicit benefits and drawbacks. nail grows quicker than toenails therefore oft sampling is simple. However, the nail is usually additional exposed to exogenous chemical components, And such a contamination therefore results in an huge amount endogenous chemical components. On different hand, such AN exposure also can be converted into a chance for watching the impact of handling deadly chemicals. Toenails is also additional protected against exogenous exposure, and therefore is also an improved choice for measure of the relevant chemical elements[4]. III. BIO MONITORING Bio monitoring is the measurement of the concentrations of chemical substances in human body fluids and tissues. It has been widely applied in industry and public health. The increasing availability of analytical methodologies with a constant decrease in detection limits make bio monitoring both more accessibleandmoresensitive. This lead to an increase in the available knowledge on the extent of human exposure to chemical substances. It may generate number of opportunities for improving human health risk assessment because it triggers new research investigating the links between low-level exposures, adverse health effects, and potentially vulnerable population groups. This is of explicit interest in sight of the enhanced attention for individualised exposure data from the final public once adverse health effects area unit just suspected from environmental exposure to chemicals. It conjointly stresses the importance of scientifically sound and reliable interpretation of bio observance. IV. BIO MONTITORING USING HUMAN NAIL Analysis of the part of chemical element or element in toenails can be used as a tool in bio monitoring the exposure history or assessing the deficiency of a specific element in an exceedingly study population, which might result in a higher under-standing of environmental and health risk. it will be employed in medical diagnosis conjointly. The connected studies unit of measurement Hair and nail clippings have the additional potential of providing a time-window: it's assumed that once the weather unit of measurement immobilized among the nail albuminoid, in order that the quantity measured in an exceedingly sure layer may be a marker for that parts at the instant of its formation within the past. The materials could thus act as a supply of knowledge on the sem ipermanent variations within the health standing, on the impact of nutrition and on activity exposure [5-10], still as in rhetorical sciences. Nail clippings and hair will be simply obtained, even in post-mortem circumstances. The potential non secular or cultural objection against providing hair samples will be overcome by grouping nail clippings. The recognition of nail clippings began to grow towards the top of the 1970. 593 N. Ambiga & Dr. A.Nagarajan. Innovative Journal of Medical and Health Science, Vol 9 Iss 9, 592–597 (2019) V. COMPONENTS OF NAIL ELEMENT The main component of the human nail is keratin l. The growth of the nail normally takes 2 to 3 months to reach the finger tip. So that it can be clipped easily from there. The process of formation of the hair is too slow and the length of formation is .5 to 1mm per week. The indicator of aging can be measured by the finger nail and toenail. The important properties of nail such as thinning in nail, discoloration in nail, grooves in nail, splitting in nail, thickening in the nail, concave and convex shape in the nail are also to be taken in to consideration for various symptoms of disease in the human body. The flatness of a nail can be used as to indicate disease in the body, nutrient deficiencies, drug reaction and poisoning or nail injury. VI. TRACE ELEMENTS IN TOE NAIL The thickness of nail modification or unsnarled and infected with microorganism indicates the unhealthiness sign of sure illness within the organic structure The human nail is leaky than skin and therefore the composition of human nail consists of seven to twelve-tone music of water, in order that it's a solid half in body.The nails square measure littered with pain and conjointly the nails square measure affected with stretched, tight and cosmetics. Nails once growth still keep isolated from various metabolic activities at intervals the physique, that's taken under consideration as associate honest reflection of long-run exposure. The advantages of nails in elements analysis square measure fascinating biomarker as a result of simple assortment, storage convenience, simple handling and dependableness of later analysis results. The nails from various fingers in feet and hands grow in several weeks of it slow between formation and clipping that indicates exposure to elevated concentration contamination integrated over a try of - twelve month quantity . The amount of part in nails square measure subject of interest at intervals the medical science and environmental sciences since recent years. The activity of nail remains the subject of interest as indices for assessing process standing, designation diseases, characteristic general intoxication and environmental exposures. The thought of elements contents at intervals the nails are thought-about as degree indicator of level in various tissues that mirror mineral metabolism at intervals the body. The connected studies square measure Hair and nail clippings have the additional potential of providing a time-window: it's assumed that once the weather square measure immobilized at intervals the nail scleroprotein, therefore the variety measured throughout a certain layer can be a marker for that elements at the moment of its formation at intervals the past. The materials might therefore act as an offer of data on the semi permanent variations at intervals the health standing, on the impact of nutrition and on activity exposure [5-10], to boot as in rhetorical sciences. Nail clippings and hair are merely obtained, even in post-mortem circumstances. The potential religious or cultural objections against providing hair samples are overcome by aggregation nail clippings. The popularity of nail clippings began to grow towards the highest of the 1970 VII. VARIATION OF TRACE ELEMENT Tear and sweat offer data over a brief time- window. These tissues and body fluids will be necessary for selections on procedures to be taken just in case of deficiency disease, contamination, or simply check the organic process standing at the time. Samples like nail and hair are studied as further thanks to appraise organic process aspects throughout a large time-window. Nail and hair square measure fashioned by a brief amount within the body, and receive many nutrients and trace parts that square measure gift within the body in this moment. A Study of the consideration of bio monitoring of trace Elements in the human nail for the Biometric Authentication 594 Innovative Journal of Medical and Health Science, Vol 9 Iss 9, 592–597 (2019) when formation by the body, the nail and hair from traditional folks much stay with constant structure and composition till be sampled Human nails, which are formed by keratin cells grow as new cells eventhough it replace the old ones. Unlike height, human nails grow during the entire lifetime. The growth of nails, as well as their matrix components, is influenced by several physiological, pathological, and environmental factors. It has been reported that the average growth rate of nails is 0.1 mm every day, depending on age and race to our best knowledge, little is known about the growth rate of nails in the Chinese population. Because of the relatively low growth rate, nail clippings have been used as an important biomarker to reflect relatively long- term exposure as a biomarker, nails have the unique advantage of application in population monitoring studies due to non-invasive sample collection and easy storage However, and nails have not been used as blood and urine in research and health diagnosis. Besides polluting metals, nails also contain essential elements including macro elements and microelements, whose determination may provide a way for assessing nutritional status]. Element research in nails has received less attention for essential elements than toxic elements. CONCLUSION The elements on the market within the physique square measure very important for the assorted physiological activities of the physique and it may also be used as a drug within the hindrance of diseases and management of the many diseases. The importance of the part within the physique has taken longer for its recognition. Currently solely it achieved the standing. Nearly half the world’s population is at the chance of not taking the part in a very correct method and in difficiency. The weather within the physique square measure currently being investigated that it's utilised in a very wide selection from therapeutic activity, biometric identification, drug and sequence delivery, to nano technology with promising outcomes within the future. Thus, it may be complete that the role of components gift within the physique and their significance and thought for biometric identification are going to be a vital analysis space within the future. ACKNOWLEDGMENTS This research work has been supported by RUSA phase 2.0, Alagappa university, Karaikudi,India REFERENCES 1. Profiles of Trace Elements in Toenails of Arab-Americans in the Detroit Area, MichiganMELISSA J. SLOTNICK,*,1 JEROME O. NRIAGU,MARY M. JOHNSON,1 AARON M. LINDER,1KATHRYN L. SAVOIE,HIKMET J. JAMIL, AND ADNAN S. HAMMAD2Biological Trace Element Research 113 Vol. 107, 2005 2. A proposed framework for the interpretation of biomonitoring data,Peter J Boogaard*1 and Chris D Money,Published: 5 June 2008,Environmental Health 2008, 7(Suppl 1):S12 doi:10.1186/1476- 069X-7-S1-S12 3. Metrology of Nail Clippings as Trace Element Biomarkers,Proefschrift, ISBN 978- 1-61499-287-5 4. Sains Malaysiana 46(4)(2017): 605– 613,http://dx.doi.org/10.17576/jsm- 2017-4604-13,Neutron Activation Analysis and Assessment of Trace Elements in Fingernail from Residents of Tokyo, Japan (Analisis Pengaktifan Neutron dan Penilaian Unsur Surih dalam Kuku Penduduk di Tokyo, Jepun) B.S. WEE* & M. EBIHARa,Sains Malaysiana 595 N. Ambiga & Dr. A.Nagarajan. Innovative Journal of Medical and Health Science, Vol 9 Iss 9, 592–597 (2019) 46(4)(2017): 605– 613http://dx.doi.org/10.17576/jsm- 2017-4604-13 5. derived from the Canadian Health Measures Survey 2007– 2013,International Journal of Hygiene and Environmental Health ,220 (2017) 189– 200,www.elsevier.com/locate/ijheh 6. versick J & Mc call J.T(1985),trace elements in human body fluids and tissues.CRC critical reviews in clinical laboratory sciences(22)(2),97-184. 7. Michael Dermiencea,, Georges Lognaya, Franc¸ oise Mathieub, Philippe Goyens b,c ,Effects of thirty elements on bone metabolism,Journal of Trace Elements in Medicine and Biology,32 (2015) 86–106 www.elsevier.com/locate/jtemb, doi.org/10.1016/B978-0-12-811353- 0.00006-3 8. Yuen Zhu, Yuzhe Wang, Fanjian Meng, ifen Lib, Shan Wu, Xiaohui Mei, Hua Li, Distribution of metal and metalloid elements in human scalp hair inTaiyuan, China 9. Guixiang Zhangd, Daishe Wua, ecootoxicology and Environmental Safety148 (2018) 538–545 10. M.A.Camina MartínB.de Mateo SillerasM.P.Redondo del Río ,Body Composition in Older Adults,Cons handbook of models for human aging ,2016 11. Gianluigi Mazzoccoli,Body composition: Where and when ,European Journal of Radiology,PII: S0720-048X(15)30140-6 ,DOI: http://dx.doi.org/doi:10.1016/j.ejrad. 2015.10.020 12. Status and interrelationship of toenail elements in Pacific children Shamshad Karatelad,, Neil I. Ward, Irene Suilan Zeng, Janis waterson, Journal of Trace Elements in Medicine and Biology 46 (2018) 10– 16 13. Kan USUDA,, Koichi KONO, Tomotaro DOTE, Misuzu ATANABE,Hiroyasu SHIMIZU, Yoshimi TANIMOTO and Emi YAMADORI ,An Overview of Boron, Lithium, and Strontium in Human Health and profiles of These Elements in Urine of Japanese, [Environmental Health and Preventive Medicine 12, 231–237, November 2007] 14. Tianjun Liu Beijing ,The scientific hypothesis of an “energy system” in the human body University of Chinese Medicine, Beijing 100029, China , Journal of Traditional Chinese Medical Sciences (2018) 5, 29e34 15. Lyudmila V. Bel’skaya, Victor K. Kosenok, Elena A. Sarf physiological Features of the Normal Mineral Composition of HumanSaliva,PII:S0003 9969(17)3020http://dx.doi.org/doi:10 .1016/j.archoralbio.2017.06.024 Reference: AOB 3927 Archives of Oral Biology 16. Sunil Kr. Jha, Kenshi Hayashi ,Body odor classification by selecting optimal peaks of chemical compounds in GC–MS spectra using filtering approaches PII: S1387-380 6(16)30290 http://dx.doi.org/doi:10.1016/j.ijms.2 017.03.003 Reference: MASPEC 15769 , International Journal of Mass Spectrometry 8-3-2017 ,oct 2015 A Study of the consideration of bio monitoring of trace Elements in the human nail for the Biometric Authentication 596 Innovative Journal of Medical and Health Science, Vol 9 Iss 9, 592–597 (2019) 17. Chang, Raymond (2007). Chemistry, Ninth Edition. MCGraw-Hill. pp. 52. 18. Devika Poddalgoda , Kristin Macey, Innocent Jayawardene , Kannan Krishnan, Derivation of biomonitoring equivalent for inorganic tin for interpreting population-level urinary biomonitoring data,Regulatory Toxicology and Pharmacology 81 (2016) 430-436 19. Z. Hu et al. (eds.), Advances in Artificial Systems for Medicine and Education, Detection of Hidden Mineral Imbalance in the Human Body byTesting Chemical Composition of Hair or Nails,Advances in Intelligent Systems and Computing 658, DOI 10.1007/978-3-319-67349-3_20 20. kuljeet kaur,Rajiv Gupta,shubni A.Saraf and shailendra k.saraf, The metal of life,comprehensive reviews in food science and food safety.vol 13,2014 358-376 ,Toxicological importance of human biomonitoring of metallic andmetalloid elements in different biological samples 21. M. Saiki, M. B. A.Vasconcellos, L. J. de Arauz, R. Fulfaro,.Determination of trace elements in human head hair by neutron activation analysis ,Journal of Radioanalytical and Nuclear Chemistry, VoL 236, Nos 12 (1998) 25-28 22. Jytte Molin Christensen 1, Milan Ihnat 2, Markus Stoeppler 3, Yngvar Thomassen 4, Claude Veillon s, and Mark Wolynetz 6,Fresenius Z ,Human body fluids - IUPAC proposed reference materials for trace elements analysis ,Anal Chem (1987) 326:639-642 23. Bernhard Michalke, Bernd Rossbach ,Thomas Go¨en, Anja Scha¨ ,erhenrich ,Saliva as a matrix for human biomonitoring in occupational and environmental medicine,Gerhard Scherer,Received: 7 November 2013 / Accepted: 20 February 2014, Springer-Verlag Berlin Heidelberg 2014 597 N. Ambiga & Dr. A.Nagarajan. Innovative Journal of Medical and Health Science, Vol 9 Iss 9, 592–597 (2019) AUTHOR BIOGRAPHY N. Ambiga, Research Scholar, Department of computer Applications, Alagappa University, karaikudi, Tamilnadu, INDIA Dr. A.Nagarajan, Assistant professor, Department of computer Applications, Alagappa University, karaikudi, Tamilnadu, INDIA 
work_aeju6oagavhibhlxrdwrozirey	----	An integrated decision making approach for ERP system selection An integrated decision making approach for ERP system selection E. Ertugrul Karsak a,*, C. Okan Özogul b a Industrial Engineering Department, Galatasaray University, Ciragan Caddesi No. 36, Ortakoy, Istanbul 80840, Turkey b HAVELSAN, Mustafa Kemal Mahallesi, Ankara 06520, Turkey Abstract Enterprise resource planning (ERP) systems have gained major prominence by enabling companies to streamline their operations, leverage and integrate business data process. In order to implement an ERP project successfully, it is necessary to select an ERP system which can be aligned with the needs of the company. Thus, a robust decision making approach for ERP software selection requires both company needs and characteristics of the ERP system and their interactions to be taken into account. This paper develops a novel decision framework for ERP software selection based on quality function deployment (QFD), fuzzy linear regression and zero–one goal programming. The proposed framework enables both company demands and ERP system characteristics to be considered, and provides the means for incorporating not only the relationships between company demands and ERP system characteristics but also the interac- tions between ERP system characteristics through adopting the QFD principles. The presented methodology appears as a sound invest- ment decision making tool for ERP systems as well as other information systems. The potential use of the proposed decision framework is illustrated through an application. � 2007 Elsevier Ltd. All rights reserved. Keywords: Decision making; Quality function deployment; Fuzzy linear regression; Enterprise resource planning; Software selection 1. Introduction The unprecedented growth of information and commu- nication technologies has influenced all facets of computing applications across organizations. At the same time, the business environment is becoming increasingly complex with functional units requiring more and more inter-func- tional data flow for decision making, timely and efficient procurement of product parts, management of inventory, accounting, human resources, and distribution of goods and services. To deal with these challenges, new software systems known in the industry as enterprise resource plan- ning (ERP) systems have surfaced in the market targeting mainly large complex business organizations. ERP comprises of a commercial software package that promises the seamless integration of all the information flowing through the company – financial, accounting, human resources, supply chain and customer information (Davenport, 1998). ERP systems are configurable informa- tion systems packages that integrate information and infor- mation-based processes within and across functional areas in an organization (Kumar & Van Hillsgersberg, 2000). ERP software market has been and continues to be one of the fastest growing segments of the information technol- ogy (IT) industry. In recent years, globalization and com- petitive business environment compel companies to invest considerable resources in the implementation of ERP sys- tems. Organizations choose and deploy ERP systems for many tangible and intangible benefits and strategic reasons (Kremzar & Wallace, 2001). Although implementing an ERP system may be costly and time-consuming, its benefits are worthwhile. However, there are a number of examples where organizations have not been successful in reaping the potential benefits that motivated them to make large investments in ERP implementations (Davenport, 1998). Motwani, Mirchandani, Madan, and Gunasekaran (2002) emphasized that ERP adoption involves initiating 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.09.016 * Corresponding author. Fax: +90 212 2595557. E-mail address: ekarsak@gsu.edu.tr (E.E. Karsak). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 36 (2009) 660–667 Expert Systems with Applications mailto:ekarsak@gsu.edu.tr appropriate business process changes as well as informa- tion technology changes to significantly enhance perfor- mance, quality, costs, flexibility, and responsiveness. There is a growing consensus among ERP system imple- menters that selecting an inappropriate system is a major reason for ERP implementation failure. Due to the com- plexity of the business environment and the diversity of ERP alternatives, ERP system selection is a tedious and lengthy task. Given the considerable financial investment and potential risks and benefits, the importance of selecting a suitable ERP system cannot be overemphasized since it is a decision on how to shape the organizational business (Teltumbde, 2000). The selection process for determining the most appro- priate ERP software among a set of possible alternatives in the market is a multi-criteria decision making (MCDM) problem. This paper introduces a decision model for ERP system selection based on quality function deployment (QFD), fuzzy linear regression and goal programming. The proposed approach benefits from the fact that QFD focuses on delivering value by taking into account the cus- tomer requirements and then by deploying this information throughout the ERP system selection process. Fuzzy linear regression is considered as an alternative decision aid for ERP system selection problems where imprecise relation- ships among system parameters exist. Weighted zero–one goal programming (ZOGP), which aims to minimize the weighted sum of deviations from the maximum achievable customer satisfaction values obtained using fuzzy linear regression, is employed as a means for determining the most suitable ERP system alternative. The advantages of the proposed decision framework can be noted as its ability to consider both user demands and ERP system character- istics, relationships between them, and interactions among ERP system characteristics, without requiring the unrealis- tic independence assumption frequently encountered in earlier studies addressing ERP system selection. The rest of the paper is organized as follows: Section 2 provides a concise review of previous works on ERP sys- tem selection. In Section 3, a novel decision making frame- work incorporating quality function deployment, fuzzy linear regression and ZOGP is introduced for ERP system selection problems. In Section 4, an illustrative application of the proposed decision making approach is presented. Finally, concluding remarks are given in Section 5. 2. Literature review Over the past two decades, software selection has become an active area of research due to its complex and imprecise nature. Lin, Hsu, and Sheen (2007) provide a comprehensive review of software selection applications. In this paper, we limit our focus to methodologies devel- oped for ERP system selection. The methods which have been applied to ERP or other information system (IS) selection include scoring, mathe- matical programming, and multi-criteria decision analysis. Owing to its simplicity, the scoring method is one of the most popular methods (Ptak, 2000). Analytic hierarchy process (AHP) based approaches constitute a general class of another commonly used IS selection techniques. Tel- tumbde (2000) proposed a methodology based on the nom- inal group technique and the AHP for evaluating ERP systems. In their recent work, Wei, Chien, and Wang (2005) used the AHP to systematically construct the objec- tives of ERP selection to support the business goals and strategies of an enterprise, identify the appropriate attri- butes, and set up a consistent evaluation standard for facil- itating a group decision process. Other methods employing nonlinear programming mod- els and zero–one goal programming models are also pro- posed for the selection of a suitable IS. Santhanam and Kyparisis (1995, 1996) proposed nonlinear zero–one pro- gramming models for IS project selection. Santhanam and Kyparisis (1995) presented a multi-criteria decision model for IS project selection which utilizes nonlinear zero–one goal programming. Santhanam and Kyparisis (1996) developed a nonlinear zero–one programming model which considered technical interdependencies among IS projects. The model is transformed to a linear mixed integer programming model through a linearization procedure. Although both of these models improved upon earlier studies by considering interdependencies inherent in the IS selection process, the solution procedure is likely to get complicated as the number of IS alternatives and inter- actions among them increase. Wei and Wang (2004) suggested a hierarchical attribute structure model to evaluate the ERP alternatives systemat- ically. They proposed a framework employing an integra- tion model that uses the fuzzy average method and fuzzy integral value ranking for ERP system selection combining data obtained from professional studies with that surveyed from interviews with vendors. Data envelopment analysis (DEA) approach has been also applied to the process of selecting an ERP system. Early adopters of DEA for decision making used the meth- odology to screen, respectively limit the number of alterna- tives, for further evaluation by other multiple attribute decision making (MADM) techniques. Fisher, Kiang, Fisher, and Chi (2004) used DEA to analyze and compare the performance of ERP packages. However, their evalua- tions are based on information provided by ERP vendors. Lall and Teyarachakul (2006) provided a case study on how DEA can be applied for ERP performance evaluation based on the real corporate data reflecting the organiza- tion’s needs and requirements. In a recent study, Bernroid- er and Stix (2006) combined the utility ranking method and the DEA to overcome the limitations of DEA in software selection. 3. The proposed decision framework In this section, a decision making approach that inte- grates quality function deployment (QFD), fuzzy linear E.E. Karsak, C.O. Özogul / Expert Systems with Applications 36 (2009) 660–667 661 https://isiarticles.com/article/1168 
work_afdb256hkbd6bpje6zfuhgo7e4	----	PACIFIC RIM PROPERTY RESEARCH JOURNAL Pacific Rim Property Research Journal, Vol 17, No 1, 2011 157 HOW ATTRACTIVE IS A SHOP: A FUZZY-EXPERT SYSTEM MATTHEW LEUNG and K.C. WONG University of Hong Kong ABSTRACT Professional planners of shopping centres know that intangible factors of a shop, including product type, quality, variety, pricing, branding, and image, are all vital in determining the pedestrian flow into the shop. These intangible factors cannot be easily quantified. Yet by experience, professional planners know how to assess them, and hence assign different shops to different strategic locations within a shopping centre, in order to optimize pedestrian flow. The objective of this study is to model experts’ professional judgment on these intangible factors. A model combining a fuzzy expert system and regression is found capable of predicting pedestrian flow at reasonable accuracies. By fine tuning the location of potential tenants, using an integrated simulation model, it is possible to optimise the shopping centre’s rental performance, even at the early stage of the shopping centre design. Keywords: Shopping centre, shop attractiveness, intangible factors, pedestrian flow, fuzzy expert system INTRODUCTION Success of any retail shop or shopping centre relies heavily on the pedestrian flow. Dawson (1983) stated that “sales are closely related to the volume of pedestrian traffic passing the shop.” Northern (1984) believed that the ultimate success of any shopping centre would be directly proportional to the number of shoppers who pass through the centre. Location is one of the significant determinants. Various neoclassical theories about location (described by Brown (1992) as traditional statistical models) have been studied, such as Central Place Theory, the Principle of Minimum Differentiation, Spatial Interaction Theory, etc. Further to these theories, regression models have been adopted to analyze the impact of various tangible factors, such as the location, physical dimensions of the shop, the walking distance between the shop and major traffic and attraction points, etc. Yet only using tangible factors is insufficient to predict the pedestrian flow drawn by a shop or at any particular point within a shopping centre (Northern 1984). He has identified various significant qualitative Pacific Rim Property Research Journal, Vol 17, No 1, 2011 158 factors, namely tenant mix, centre/ shop characteristics, brand name and prestige images, nearby competition, the ‘quality’ of retail space available, etc. Northern has classified shopping centre image as a qualitative factor; however Dennis et al. (2005) has attempted to quantify the impact of image on the performance of shopping centres. He has investigated the relationship between “image” or “attractiveness” and individual shopper behaviour. He asserted image as the complete mix of cues (e.g. sensory) which communicates with customers and influences shopping behaviour. He has consolidated the gravity, spatial interaction and Central Place approaches in his attitude-behaviour theory. The attractiveness of shopping centres has been quantified and triangulated with a branding framework. His study reported that pleasure and enjoyment, deterrence effect of travel and motivation significantly affected a shopping centre as an object of desire. The ‘pleasure’ and ‘enjoyment’ would be affected by atmospheric stimulus, image of the shopping centre environment; image of stores and products, and arousal stimulation affect. Deterrence effect would be influenced by travel distance and time, distance to a competing centre, size and attractiveness of other centres in and near the catchment area. Shopper’s motivation was significantly related to self-esteem and relatedness. Steenkamp and Wedel (1991) stated that retail image was the key parameter reflecting the total value of shopping centres. A unique and favorable image helps create a sustainable competitive advantage and establish a clear marketing position from other competitors. Five identified attributes of shopping centre image are merchandising, accessibility, service, atmospherics and entertainment (Sit et al., 2003). Besides studying shopping centres, Martineau (1958) has studied and shown that the drawing force of a store was the store personality or image – the way in which the store was defined in the shopper’s mind, partly by its functional qualities and partly by an area of psychological attributes. He stated that “whereas the retailer thinks of himself as a merchant concerned with value and quality, there is a wide range of intangibles which also play a critical role in the success or failure of his store”. Current practitioners predict the pedestrian flow mainly with their own experience and judgment. In order to pursue the best performance, as well as its consistency, of a shopping centre, a scientific management system to control the layout design and tenant mix scheme is necessary and desirable. Compared to tangible attributes, intangible attributes are more difficult to measure and study. This paper will investigate the measurement of shop attractiveness, which is one of the most significant intangible attributes, and the relationship between shop attractiveness and pedestrian flow, with the data collected in three shopping centres in Hong Kong. Pacific Rim Property Research Journal, Vol 17, No 1, 2011 159 SHOP ATTRACTIVENESS The attractiveness of a shop is defined as its ability to draw potential customers into, and around the neighborhood of, the shop. Such attractiveness is determined by both tangible (e.g. shop size, location, etc.) and intangible factors (e.g. branding, market position, etc.). Intangible factors are difficult to be quantified, as individuals have different perceptions on various attributes. A fuzzy expert system, which is a codification of the common sense, imitating how people use their imprecise information to make the right decision, is designed to predict pedestrian flow at specific points inside a shopping centre, It is a widely accepted tool for modeling nonlinear functions of arbitrary complexity and dealing with linguistic variables, which transform descriptive words or sentences into quantitative values. Wong and So (1995), Bagnoli and Smith (1997, 1998) and Ng et al. (2002) applied fuzzy logic on solving problems in the real estate and construction industries. Wong and So have constructed a fuzzy reasoning model and applied the model on contract decision making in Hong Kong. Bagnoli and Smith investigated how to apply fuzzy logic on real estate valuation. Ng et al. have demonstrated how to set up the fuzzy membership functions of procurement selection criteria. The system proposed in this paper involves three main components: fuzzification, inferencing and defuzzification. Fuzzification is the procedure to convert raw data (i.e. linguistic variables) into membership functions. The membership function is a generalization of the linguistic input functions in classical sets. It can be represented graphically of the participation magnitude of each input with an interval ranging from zero (false) to one (true). Inferencing refers to the reasoning or logic applied in the system which constitutes the rule base. Rule base specifies conclusions drawn from assertions known or assumed to be true (Jantzen 1999). With the membership functions and truth values of inputs obtained in fuzzification, the rules applied will be invoked to determine the result, which will be mapped onto a membership function and truth value controlling the output variable. In this paper, the centroid method, which determines the centre of the area of the combined membership functions and produces more representative results than other methods (Nurcahyo et al. 2003), is chosen to perform the defuzzification process. FUZZIFICATION The first step in fuzzy logic is to establish membership functions of input terms which are generalization of the indicator functions. In fuzzy logic, it represents the degree of truth as an extension of valuation. In other words, common sense is codified and reflected in the membership function. The membership functions are established with Pacific Rim Property Research Journal, Vol 17, No 1, 2011 160 the aid of a questionnaire, which collects and quantifies the views of interviewees with different classical sets. The framework of the questionnaire (as shown in Appendix A), with reference to Ng et al. (2002), is designed to collect the maximum, minimum and average values of both input and output functions. The questionnaires are distributed to ten current practitioners/ industry experts who could share their experience towards the operation and mechanism of the retail market. Quantitative data of the linguistic variables are collected to develop the membership functions. Common shapes of membership functions are triangular, trapezoids, smooth triangular and smooth trapezoid. During practical application, the number of curves and their placements are far more critical than the shape type. Three to seven curves (i.e. terms) are generally adequate to cover the universe of disclosure, representing all objects that come into consideration, of the input and output values. In order to design a system for market practitioners who do not process advance mathematical knowledge, a simple triangular function is chosen for simple graphical representation. Another advantage is that simple triangular function can be processed with some common softwares; for example, SPSS, Microsoft Excel. With reference to previous studies, eight input variables (IVs) and three output variables (OVs) were suggested for the model. Input variables, including shop youth image, quality of goods sold, pricing competitiveness, branding, variety of goods sold, trading, shop size and shop location, are the settings and characters of the shops. Output variables refer to the percentage of pedestrians, belonging to different age groups, being attracted to the shop. They consist of three age groups: OV1 (age 15- 24); OV2 (age 25-35) and OV3 (age above 35). Each variable is associated with several terms. The definition of input and output variables are shown in Appendix B. The interviewees are required to indicate their views towards the variables which will be expressed in percentage terms according to the specified formulas to elaborate the linguistic variables. Maximum, mean and minimum values of terms are recorded and will be adopted in later inferencing. A descriptive statistical table of linguistic variables together with terms is presented in Appendix C. For example, see IV15 in Appendix C. Experts consider that a shop’s youth image is said to be mature if 65% to 85% of the goods found in the shop are mature – where a mature good is defined as one that shoppers aged above 35 may find interest in. The mean of all 10 experts’ opinions is 76%. This would give an overall triangular distribution of: 65%, 76%, and 85%; against which the membership function of 0.0; 1.0; and 0.0 could be assigned respectively. Given this triangular distribution, any shop with a certain percentage of mature goods found inside the shop (ranging from Pacific Rim Property Research Journal, Vol 17, No 1, 2011 161 65 to 85%) could be assigned, by interpolation, a “degree” of membership (from 0 to 1) to the mature image. RULE BASE Rules of experience constitute the inference engine which is applied to the shops in estimating their drawing power. In this paper, the rules are categorized by trades, namely restaurant, fashion, beauty, jewelry, gift/ furniture, electrical appliance, store/ supermarket and retail services. They are then subcategorized with other aspects, such as image, quality, pricing, branding, goods variety and shop size. 30 rules (as listed in Appendix D) have been formulated and are considered to cover most, if not all, daily situations. The outcome of each rule is the experts’ rules of thumb applied in their practices. For example in Appendix D, Rule 1 – the Youngsters’ Café Rule – could be read as follows: IF the shop’s youth image is teenagers or youngster; AND goods quality is bad or just okay; and pricing competitiveness is very cheap or cheap; AND branding is unknown brand or local unpopular or local popular or international; AND goods variety is limited; AND trade is restaurant/entertainment; AND shop size is tiny or small or normal; AND location is easy to access or convenient or very convenient; then the shop’s attractiveness to teenagers (age 15-24) is very high; AND to middle age (25-45) is medium; AND to mature customers (above 45) is very low. How applicable Rule 1 is to a specific shop would depend on (a) the degree of membership of each individual IVs, as assessed in the previous section on fuzzification ; and (b) the use of the Max-Min Rule. i.e. choosing the maximum of the various degrees of membership wherever the logical syntax in-between the IVs (or groups of IVs) is OR; and minimum when it is AND. PEDESTRIAN FLOW In this paper, three regional shopping centres in Hong Kong are chosen to study, namely Festival Walk, Langham Place and Pacific Place. All of them share many characteristics in common, such as similar size and scale, managed by leasing and property management professionals, attached to railway/ subway, office buildings attached, etc. These help control the external factors of the subjects. The main objective of the fuzzy model is to evaluate the pedestrian flow at a particular point in a shopping centre with the estimated attractiveness of individual shops. Membership functions of intangible variables of individual shops are assessed and Pacific Rim Property Research Journal, Vol 17, No 1, 2011 162 then incorporated into the inference engine. An index representing the drawing power of each individual shop is derived accordingly. It is possible that more than one shop would have an impact on a particular measuring point. Hence, three principles are established to select shops with drawing power that would heavily affect the pedestrian flow at the measuring point. These three principles are used in the following order: 1. Select shops with entrances visible from the measuring point. 2. Choose those with entrances within 10m from the measuring point; 3. Select the best three shops – those with the highest drawing power to customers. A measuring point in Festival Walk is chosen to demonstrate this method. The selected point is at LG1 with a pedestrian flow count of 801 shoppers per a 10-minute interval. The nearest three shops fitting all the above mentioned principles are Page One, Festival China and Cour Carre. Page One is a book shop with a lettable area approximately 12,600 square feet. It is internationally reputable and adopts a premium pricing strategy. Combined with its wide range of book selection, its marketing strategy focuses on middle class customers with relatively strong purchasing power. Experts’ assessment extracted from the questionnaire and its membership functions are summarized in Table 1. Table 1: Assessment of membership functions of Page One Linguistic variable Experts’ assessment Input variable Membership function Image 50% IV14 0.94 Goods quality 45% IV24 0.71 Pricing 110% IV34 0.68 Brand name 60% IV44 0.67 Variety 160% IV54 0.56 Trade 30% IV62 0.94 Area 126% IV75 0.38 Location 60% IV82 0.91 Applying the Max-Min operation, the output strength is 0.38 and the combination of assessments complies with Rule 24 – Large Toy/Book/Music Shop Rule. According to the rule, three output variables are OV15, OV24 and OV34. In order to assess the implication of the three output variables, they are defuzzified with the calculation of centroids of the outputs. The defuzzification results of OV1, OV2 and OV3 are 69.7%, 66.4% and 61.5% respectively. It could be interpreted as 69.7% of shoppers aged 15 to 24, 66.4% of shoppers aged 25 to 35; and 61.5% of shoppers aged above 35 are “pulled” towards Page One respectively. According to the experts’ opinions, the Pacific Rim Property Research Journal, Vol 17, No 1, 2011 163 proportion of the three age groups in shopping centres is approximately 1:2:1. A weighted average index for the shop pedestrian is calculated as 0.66. Festive China is a local traditional Chinese restaurant, with a lettable area approximately 7,000s.f. Compared with existing nearby competitors, its pricing strategy is positioned at the middle-mass market, providing quality food and beverages at affordable price levels. Food choice is plentiful, although limited to “Chiu Chow” cuisine. The assessment of the shop variable experts’ assessment and its representative membership functions are shown in Table 2. Table 2: Assessment of membership functions of Festive China Linguistic variable Experts’ assessment Input variable Membership function Image 50% IV14 0.94 Goods quality 45% IV23 0.71 Pricing 90% IV33 0.94 Brand name 30% IV43 0.39 Variety 50% IV52 0.91 Trade 20% IV61 0.71 Area 70% IV74 0.50 Location 60% IV82 0.91 The assessments conform to Rule 2 – the Young Restaurant Rule. The output strength is 0.39. After defuzzification, the results for OV1, OV2 and OV3 are 14.3%, 66.5% and 42.0% respectively. It could be interpreted as 14.3%, 66.5% and 42.0% of shoppers aged 15 to 24, 25 to 35, and above 35 are “pulled” towards Festive China respectively. The pedestrian index at the point is 0.47. Cour Carre is a local fashion chain store with a lettable area approximately 1,630s.f. It targets at middle-income working class. Compared with other chain fashion stores, its goods quality and pricing level are reasonable, even though the choices are relatively limited. The assessment of the shop variable experts’ assessment and its representative membership functions are given in Table 3. Pacific Rim Property Research Journal, Vol 17, No 1, 2011 164 Table 3: Assessment of membership functions of Cour Carre Linguistic variable Experts’ assessment Input variable Membership function Image 30% IV13 1.0 Goods Quality 40% IV24 0.94 Pricing 85% IV33 0.63 Brand Name 35% IV43 0.77 Variety 50% IV52 0.91 Trade 50% IV63 0.83 Area 16% IV72 0.53 Location 60% IV82 0.91 The assessments comply with Rule 11 – the Reasonably Priced Small Popular Fashion Shop Rule. The output strength is 0.53. The defuzzification results of OV1, OV2 and OV3 are 14.9%, 46.0% and 42.8% respectively. It could be interpreted as 14.9%, 46.0% and 42.8% of shoppers aged 15 to 24, 25 to 35, and above 35 are “pulled” towards Cour Carre respectively. The pedestrian index generated by Cour Carre at the point is 0.37. The combined pedestrian index, through summation of three individual indices at the point, is 1.50, with the assumption that the pedestrian drawing power of individual shops is independent from each other. By applying the same methodology, the combined pedestrian indices at other testing points are estimated. VALIDITY Three shopping centres (Festival Walk, Pacific Place and Langham Place) have been selected for testing the validity. Combined Pedestrian Index at each particular point was evaluated with the similar mechanism. They are tested against the pedestrian flow at each point using regression models. It is expected that there will be a significant relationship between shop attractiveness (represented by the index) and the pedestrian flow. The results are given in Tables 4-6. Pacific Rim Property Research Journal, Vol 17, No 1, 2011 165 Table 4: Festival Walk The pedestrian flow at 54 measuring points of Festival Walk has been counted and tested against the combined pedestrian index. The model is statistically significant at the 99% confidence level (i.e. p<0.01) with adjusted R-squared of 0.62. Combined pedestrian index is identified as significant at the 0.01 level. Table 5: Langham Place The pedestrian flow at 44 measuring points of Langham Place has been counted and tested against the combined pedestrian index. The model is statistically significant at the 99% confidence level (i.e. p<0.01) with adjusted R-squared of 0.68. Combined pedestrian index is identified to be significant at the 0.01 level. Dependent Variable: Pedestrian Flow PEDP Sample Size: 53 Variable Coefficient Std. Error t-Statistic Prob. Combined Pedestrian Index 481.872 52.350 9.205 0.000 C 70.009 49.144 1.425 0.160 R-squared 0.624255 Mean dependent var 494.874 Adjusted R-squared 0.616887 S.D. dependent var 198.466 S.E. of regression 122.8426 Akaike info criterion 12.497 Sum squared resid 769605.3 Schwarz criterion 12.571 Log likelihood -329.1623 84.730 F-statistic Dependent Variable: Pedestrian Flow PEDP Sample Size: 44 Variable Coefficient Std. Error t-Statistic Prob. Combined Pedestrian Index 719.553 74.772 9.623 0.000 C -28.275 70.049 -0.404 0.689 R-squared 0.687985 Mean dependent var 578.705 Adjusted R-squared 0.680556 S.D. dependent var 357.647 S.E. of regression 202.1396 Akaike info criterion 13.500 Sum squared resid 1716137 Schwarz criterion 13.581 Log likelihood -295.004 92.609 F-statistic Pacific Rim Property Research Journal, Vol 17, No 1, 2011 166 Table 6: Pacific Place The pedestrian flow at 34 measuring points of Pacific Place has been counted and tested against the combined pedestrian index. The model is statistically significant at the 99% confidence level (i.e. p<0.01) with adjusted R-squared of 0.64. Combined pedestrian index is identified to be significant at the 0.01 level. Results of all three models have demonstrated a significant relationship between the combined pedestrian index and the pedestrian flow. By comparing the fuzzy indices assessed at different points within the same retail development, the shop attractiveness map / pattern would be estimated. A regression model, compiling all of the above data, is set up to test the predictive power of the indices on the actual pedestrian flow at particular points. The dependent variable is the pedestrian flow at particular points (PEDP), while independent variables are the Combined Pedestrian Index (INDEX) and the Centre’s Pedestrian Flow (PEDC). PEDC represents the factor of total pedestrians being drawn under the influence of the overall setting of the shopping centre, while INDEX represents the drawing power of particular points. Dependent Variable Pedestrian Flow PEDP Sample Size: 34 Variable Coefficient Std. Error t-Statistic Prob. Combined Pedestrian Index 477.332 61.693 7.737 0.000 C -153.784 59.166 -2.599 0.014 R-squared 0.651659 Mean dependent var 271.221 Adjusted R-squared 0.640774 S.D. dependent var 213.865 S.E. of regression 128.1809 Akaike info criterion 12.602 Sum squared resid 525770.6 Schwarz criterion 12.692 Log likelihood -212.2303 59.864 F-statistic Pacific Rim Property Research Journal, Vol 17, No 1, 2011 167 Table 7: Pedestrian flow forecast The model is statistically significant at the 99% confidence level (i.e. p<0.01) with adjusted R-squared of 0.65. Two independent variables are identified to be significant at the 0.01 level. The model shows significant predictive power on the actual pedestrian flow within the retail development. CONCLUSION Fuzzy logic has been widely adopted in pedestrian flow simulation models in various scenarios, such as railway stations and stadiums. The established simulation models are mainly designed for emergency purpose only, with very few applications for commercial uses. This paper is the first to adopt fuzzy logic to assess shop attractiveness within a shopping centre. The results are positive and a map of shop attractiveness would be drawn. The assessment of the performance of different tenant mix schemes would be carried out systematically beforehand and does not need to be relied solely on the experience of the leasing manager. This method is, therefore, potentially beneficial to developers of shopping centres. Although positive results have been obtained, there are several limitations: 1. The combined pedestrian index was assessed with the assumption that the pedestrian drawing power of individual shop is independent from each other. The Dependent Variable PEDP Sample Size: 131 Variable Coefficient Std. Error t-Statistic Prob. Centre's Pedestrian Flow: PEDC 0.548254 0.07314 7.495995 0.000 Combined Pedestrian Index INDEX 573.3924 41.49664 13.8178 0.000 C -401.5748 63.36572 -6.337414 0.000 R-squared 0.654326 Mean dependent var 464.9835 Adjusted R-squared 0.648924 S.D. dependent var 290.2337 S.E. of regression 171.9682 Akaike info criterion 13.1551 Sum squared resid 3785353 Schwarz criterion 13.2210 Log likelihood -858.6611 121.1453 F-statistic Pacific Rim Property Research Journal, Vol 17, No 1, 2011 168 combined drawing power might have been overestimated. Further study of integration of shop attractiveness is suggested. 2. Shop attractiveness is not the sole factor affecting the pedestrian flow of a shopping centre. A comprehensive system should also include other significant tangible factors, such as location, size, shape, layout, number of floors and number of shops, etc. To assess the overall performance of shopping centres, both tangible and intangible attributes should be taken into account. Compared to shop attractiveness, other factors are more tangible and would be included into a regression-expert system. The system will be desirable for real estate investors, developers, shop tenants and designers, which will allow them to modify the alternative design and tenant mix schemes in the early development stage, without going through the pain of the trail and error stage. REFERENCES Areni, C. and D. Kim (1995), “The Influence of In-Store Lighting on Consumers’ Examination of Merchandise in a Wine Store”, International Journal of Research in Marketing, Vol. 11(4): 117-125 Bagnoli, C. and Smith, H.C. (1997) “Fuzzy Logic: The New Paradigm for Decision Making”, Real Estate Issues, Vol. 22(2): 35-41 Bagnoli, C. and Smith, H.C. (1998), “The Theory of Fuzzy Logic and its Application to Real Estate Valuation”, Journal of Real Estate Research, Vol. 16: 169-99. Brown, S. (1992), Retail location: a micro-scale perspective, Aldershot: Avebury. Brown, S. (1992), “Tenant Mix, Tenant Placement and Shopper Behaviour in a Planned Shopping Centre”, The Service Industries Journal, Vol. 12(3): 384-402. Dawson, J.A. (1983), Shopping Centre Development, London: Longman. Dennis, C., A. Newman, and D. Marshland (2005), Objects of Desire-Consumer Behaviour in Shopping Centre Choices, New York: Palgrave MacMillian. Engel, J.F., Blackwell, R.D. and Miniard, P.W. (1986), Consumer Behaviour, 5th edn. Chicago: Dryden Press. G. W. Nurcahyo, S. Shamsuddin, R. Alias and M. Noor Md. Sap (2003), “Selection of Defuzzification Method to Obtain Crisp Value for Representing Uncertain Data in a Modified Sweep Algorithm”, JCS&T, Vol. 3(2) : 22-28. Pacific Rim Property Research Journal, Vol 17, No 1, 2011 169 Herrington, J. and L. Capella (1996), “Effects of Music in Service Environments: A Field Study”, Journal of Services Marketing, Vol. 10(2): 26-41. Jantzen, J. (1999), “Tutorial on Fuzzy Logic”, http://www.cs.baskent.edu.tr/ ~muluer/educational/ai/JJ_fuzzy_tut.pdf. Martineau, P. (1958), “The Personality of the Retail Store”, Harvard Business Review, Vol. 36: 47-55. Michon, R., Chebat, J.C. and Turley, L.W. (2005), “Mall Atmospherics: the Interaction Effects of the Mall Environment on Shopping Behavior”, Journal of Business Research, Vol. 58: 576-583. Ng, S., Luu, D.T., Chen, S.E. and Lam, K.C. (2002), “Fuzzy Membership Functions of Procurement Selection Criteria” Construction Management and Economics, Vol. 20: 285-296. Northen, R.I. (1984), Shopping Centre Development Reading: Centre for Advanced Land Use Studies, College of Estate Management. Sit, J., Merrilees, B. and Birch, D. (2003), “Entertainment-seeking Shopping Centre Patrons: the Missing Segments”, International Journal of Retail & Distribution Management, Vol. 31(2): 80-94. Smith, J. (1985), “Pedestrianisation – Shopping Streets in Scotland”, The Planner, May:12-16. Steenkamp, J. and Wedel, M. (1991), “Segmenting Retail Markets on Store Image Using a Consumer-Based Methodology”, Journal of Retailing, Vol. 67(3): 300-320. Wong, K.C. and So, T.P. (1995), “A Fuzzy Expert System for Contract Decision Making”, Construction Management and Economics, Vol. 13: 95-103. http://www.cs.baskent.edu.tr/%20~muluer/educational/ai/JJ_fuzzy_tut.pdf� http://www.cs.baskent.edu.tr/%20~muluer/educational/ai/JJ_fuzzy_tut.pdf� Pacific Rim Property Research Journal, Vol 17, No 1, 2011 170 Appendix A: Individual shop attractiveness assessment Individual Shop Attractiveness Assessment Q1. What are the best figures to describe the following five catalogues of shop youth image? Please indicate your responses in the following boxes. ** Teenagers % Youngster % Young % Adult % Mature % Q2. What are the best figures to describe the following six catalogues of goods quality? Please indicate your responses in the following boxes. ** Quality is defined as [Material quality + workmanship quality + design concept quality] x 100% Material quality = Workmanship quality = where Design concept quality = Bad % Just Okay % Reasonable % Above average % Good % Prestige % Q3. Please indicate your responses in the following boxes. ** Very cheap % Cheap % Okay % Reasonable % Expensive % Luxury % amount of money paid for extra material for better quality than the norm / total material price for the norm x 100% amount of money paid for extra design concept effort for better quality than the norm / total price for design concept effort for the norm x 100% Pricing competitiveness is defined as [Normal anticipated pricing in the shop / average market price expectation of the same type of goods] x 100% Image is defined as the % of mature goods (shopper's age > 35 may be interested on that goods) found in the shop What are the best figures to describe the following six catalogues of shop pricing competitiveness? amount of money paid for extra workmanship for better quality than the norm / total workmanship price for the norm x 100% Pacific Rim Property Research Journal, Vol 17, No 1, 2011 171 Q4. What are the best figures to describe the following five catalogues of shop branding? Please indicate your responses in the following boxes. ** Rarely heard % Local % Local popular % International % Prestige % Q5. What are the best figures to describe the following four catalogues of goods variety in the shop? Please indicate your responses in the following boxes. ** Limited choices % Average % Plenty % Abundant % Q6. What are the best figures to describe the attractiveness of the following trades? Please indicate your responses in the following boxes. ** % % % % Supermarket/store/retail services (Trade 5) % Q7. What are the best figures to describe the following five catalogues of shop size? Please indicate your responses in the following boxes. ** Size is defined as [Shop size in shopper impression / 10,000sf] Tiny % Small % Normal % Large % Very large % Q7. Please indicate your responses in the following boxes. ** Very inconvenient % Inconvenient % Normal % Convenient % Very inconvenient % Location is defined as [Travel time from the main entrance to the shop / travel time from the main entrance to the most distant shop] Restaurant/entertainment (Trade 1) Book/music/furniture/toy (Trade 2) Fashion/ sports (Trade 3) Beauty/ Jewelry /AV shops (Trade 4) Attractiveness of trade is defined as the [No. of visitor of the shop/ Total no. of visitors of shopping centre] x 100% What are the best figures to describe the following five catalogues of shop location in term of convenience? Branding is defined as [Proportion of shoppers knowing the brand name / Proportion of shoppers knowing the most famous brand name] x 100% Variety is defined as [Normal anticipation of the choices of goods in the shop / Normal anticipation of choices of goods in typical shops] x 100% Pacific Rim Property Research Journal, Vol 17, No 1, 2011 172 Appendix B: Input and Output Variables of the Fuzzy Expert System Input Variable Description Term IV1 – Shop youth image Shoppers’ general impression on the youth image of the shop Represented formula: % of mature goods (i.e. age of shopper > 35 may be interested) found in the shop IV11 – teenagers IV12 – youngster IV13 – young IV14 – adult IV15 – mature IV2 – Goods quality Shoppers’ general impression on the quality of the goods sold - Represented formula1 (Material quality + workmanship quality + design concept quality), where: : Material quality = (amount paid for extra materials for better quality than the norm/ total material price for the norm) x 100% Workmanship quality = (amount paid for extra workmanship for better quality than the norm/ total workmanship price for the norm) x 100% Design concept quality = (amount paid for extra design concept quality for better quality than the norm/ total price for design concept effort for the norm) x 100% IV21 – bad IV22 – Just okay IV23 – reasonable IV24 – above average IV25 – good IV26 – prestige IV3 – Pricing competitiven ess Shoppers’ general impression on pricing competitiveness of goods sold Represented formula: (Normal anticipated pricing in the shop / average market price expectation of the same type goods) x 100% IV31 – very cheap IV32 – cheap IV33 – okay IV34 – reasonable IV35 – expensive IV36 – luxury IV4 – Branding General impression from shoppers on the brand name of the shop Represented formula: (Proportion of shoppers knowing the brand name / proportion of shoppers knowing the most famous brand IV41 – rarely heard IV42 – local unpopular IV43 – local popular 1 With reference to the methodology suggested by Ng et al. (2002) Pacific Rim Property Research Journal, Vol 17, No 1, 2011 173 name) x 100% IV44 – international IV45 – prestige IV5 – Goods Variety General impression from the shoppers on the variety of the goods sold in the shop Represented formula: (Normal anticipation of the choice of goods of the shop/ Normal anticipation of choice of goods in typical shop) x 100% IV51 – limited IV52 – average IV53 – plenty IV54 – abundant IV6 – Trading The drawing power on shoppers by the nature of the shop trade trade 1 – restaurant / entertainment; trade 2 –books, gift, music and similar goods; trade 3 –fashion, sports and similar goods; trade 4 –beauty cosmetic, jewelry, audio & video, & other luxuries trade 5 – supermarket Represented formula: (No. of visitor of the shop/ Total no. of visitors of shopping centre) x 100% IV61 – trade 1 IV62 – trade 2 IV63 – trade 3 IV64 – trade 4 IV65 – trade 5 IV7 – Shop size Shoppers’ impression on shop size Represented formula: (Shop size in shopper impression / 10,000sf) x 100% IV71 – tiny IV72 – small IV73 – normal IV74 – large IV75 – very large IV8 – Location Shopper impression on the convenience of shop location Represented formula: (Travel time from main entrance/ Travel time to visit the most distant shop) x 100% IV81 – very inconvenient IV82 – inconvenient IV83 – normal IV84 – convenient IV85 – very convenient Output Variable Description Term OV1 How many shoppers in the age group between 15 and 24 are attracted to the shop Represented formulas: Number of shoppers (for the age group between 15 and 24) visiting the shop/Total number of shoppers visiting the shopping centre OV11 – very low OV12 – low OV13 – medium OV14 – High OV15 – Very high OV2 How many shoppers in the age group between 25 and 34 are attracted to the shop Represented formulas: Number of shoppers (for the age group between 25 and 34) visiting the shop/ Total number of shoppers visiting the shopping centre OV21 – very low OV22 – low OV23 – medium OV24 – High OV25 – Very high Pacific Rim Property Research Journal, Vol 17, No 1, 2011 174 OV3 How many shoppers in the age group above 35 are attracted to the shop Represented formulas: Number of shoppers (for the age group above 35) visiting the shop/ Total number of shoppers visiting the shopping centre OV31 – very low OV32 – low OV33 – medium OV34 – High OV35 – Very high Pacific Rim Property Research Journal, Vol 17, No 1, 2011 175 Appendix C: Statistical summary of expert opinion on intangible variables Intangible Variable Term Minimum Mean Maximu m 1. Shop youth image IV11 – teenagers 0% 5% 15% IV12 – youngster 5% 12% 25% IV13 – young 15% 30% 40% IV14 – adult 30% 49% 65% IV15 – mature 65% 76% 85% 2. Goods quality IV21 – bad 0% 6% 15% IV22 – Just okay 5% 15% 25% IV23 – reasonable 20% 32% 40% IV24– above average 25% 41% 55% IV25 – good 40% 53% 65% IV26 – prestige 60% 77% 90% 3. Pricing titi IV31 – very cheap 20% 36% 50% IV32 – cheap 50% 69% 80% IV33 – okay 75% 91% 110% IV34 – reasonable 80% 103% 125% IV35 – expensive 125% 140% 160% IV36 – luxury 150% 170% 200% 4. Branding IV41 – unknown brand 0% 3% 10% IV42 – local unpopular 5% 11% 30% IV43 – local popular 25% 38% 55% IV44 – international 45% 55% 70% IV45 – prestige 65% 81% 95% 5. Goods Variety IV51 – limited 15% 35% 50% IV52 – average 30% 52% 80% IV53 – plenty 50% 75% 120% IV54 – abundant 120% 144% 180% 6. Trading IV61 – trade 1 5% 16% 30% IV62 – trade 2 15% 29% 45% IV63 – trade 3 35% 47% 65% IV64 – trade 4 50% 68% 80% IV65 – trade 5 50% 77% 100% 7. Shop size IV71 – tiny 1% 4% 10% IV72 – small 5% 13% 20% IV73 – normal 15% 26% 40% IV74 – large 35% 60% 80% IV75 – very large 70% 86% 150% Pacific Rim Property Research Journal, Vol 17, No 1, 2011 176 8. Location IV81 – very inconvenient 65% 84% 100% IV82 – inconvenient 50% 61% 75% IV83 – normal 28% 36% 50% IV84 – convenient 10% 17% 32% IV85 – very convenient 5% 9% 15% Intangible Variable Term Minimum Mean Maximum 1. Output Variable I OV11 – very low 5% 14% 25% (Age group 1) OV12 – low 10% 24% 40% OV13 – medium 30% 44% 55% OV14 – high 45% 60% 75% OV15 – very high 50% 74% 90% 2. Output Variable II OV21 – very low 10% 18% 25% (Age group 2) OV22 – low 15% 25% 35% OV23 – medium 30% 49% 60% OV24 –high 50% 67% 85% OV25 –very high 60% 73% 80% 3. Output Variable III OV31 – very low 5% 14% 20% (Age group 3) OV32 – low 15% 23% 35% OV33 – medium 30% 43% 55% OV34 –high 50% 61% 75% OV35 –very high 50% 71% 85% Pacific Rim Property Research Journal, Vol 17, No 1, 2011 177 Appendix D: Rule base – 30 rules 
work_ag2w6cadgvej3hd55ptj5do6ku	----	Association rules applied to credit card fraud detection Association rules applied to credit card fraud detection D. Sánchez a,*, M.A. Vila a, L. Cerda a, J.M. Serrano b a Department of Computer Science, A.I., University of Granada, E.T.S.I. Informática y Telecomunicación, C/Periodista Daniel Saucedo Aranda s/n, 18071 Granada, Spain b Department of Informatics, University of Jaen, Spain Abstract Association rules are considered to be the best studied models for data mining. In this article, we propose their use in order to extract knowledge so that normal behavior patterns may be obtained in unlawful transactions from transactional credit card databases in order to detect and prevent fraud. The proposed methodology has been applied on data about credit card fraud in some of the most important retail companies in Chile. � 2008 Elsevier Ltd. All rights reserved. Keywords: Association rules; Data mining; Credit card fraud; Fraud detection; Fraud prevention 1. Introduction Competitiveness in the retail industry is continuing and it is becoming increasingly aggressive as revealed by recent events in the sector and specialist studies such as (Zarufe, 2005). One of the leading businesses in this sphere is hire purchase and one of the main commercial strategies is the emission of department store credit cards to clients, as evident from the publications (Zarufe, 2005) which indi- cate that three companies lead the retail industry in the Latin American Southern Cone (Argentina, Chile, Peru, Colombia). In Chile, they compete for 95% of retail indus- try sales, which in 2003, according to these same publica- tions, exceeded the three thousand, three hundred million dollar mark with market shares of 60.29%, 18.26% and 15.63%, respectively. By the end of 2003, they had issued 3.0, 2.7 and 2.6 million credit cards, and in Chile alone 16 million credit cards had already been issued by the dif- ferent retail distribution chains. This form of payment was 7 times higher than the number of bank-issued credit cards, which were responsible for on average 65% of the sales of their issuing houses, represented almost 20% of Chile’s con- sumption debt in total, and there were over 11,000 estab- lishments who did not issue this type of card but which traded with them. Within the retail industry, the predomi- nant trade component are financial services and the distri- bution of clothing and furnishings through department stores with sales in 2005 reaching US $3,194, sales which are distributed among the four major chain stores with the profile in Table 1. For years, both from the academic and the technological advisory or consultancy perspective, it has been observed and in this case confirmed that in three of these four large companies, the level of technological support used as part of their computerized management systems in their deci- sion-making processes is very different from the level of use of information technologies used in the operational transaction systems, and these computerized management systems may therefore be classified according to the follow- ing levels of computing maturity: Level Zero (non-computerized systems): No incorpora- tion of computer technology in computerized management systems, i.e. information for decision-making is obtained manually. 0957-4174/$ - see front matter � 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2008.02.001 * Corresponding author. Tel.: +34 958 246397; fax: +34 958 243317. E-mail addresses: daniel@decsai.ugr.es (D. Sánchez), vila@decsai.ugr. es (M.A. Vila), lcerda1511@hotmail.com (L. Cerda), jschica@ujaen.es (J.M. Serrano). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 36 (2009) 3630–3640 Expert Systems with Applications mailto:daniel@decsai.ugr.es mailto:vila@decsai.ugr. es mailto:vila@decsai.ugr. es mailto:lcerda1511@hotmail.com mailto:jschica@ujaen.es Level One (semi-computerized systems): Incipient use of technological resources in computerized management sys- tems for decision-making, e.g. the use of spreadsheets to prepare the information. Level Two (departmental computerized systems): Use of departmental information systems as the nucleus of the computerized management system for decision-making, and data is usually gathered from transactional processes associated to a specific function within the value chain. Level Three (integrated computerized systems): Use of the company’s integrated administrative information sys- tem (ERP) as the main element to supply the computerized management system (CMS) for decision-making. These CMS are supplied by data from all the processes support- ing the company’s value chain. Level Four (controlled or synchronized computerized sys- tems): These computerized management systems (CMS) integrate the use of control boards or control panels, enabling decisions to be made as the need arises by having online information for the fulfillment level of the objectives associated with the management indicators. Level Five (predictive computerized systems): Data min- ing models are incorporated into the previous level to extract non-explicit information from the records of trans- actions from daily operation. These enable new behavior patterns to be determined which reinforce, confirm or mod- ify management indicators and allow trends to be recog- nized for decision-making. Level Six (automatic computerized systems): The previ- ous level of computerized management systems is rein- forced and combined with daily operation through the use of expert systems with knowledge bases and inference engines to support decision-making, thereby incorporating intelligence into the operational systems so that given cer- tain management parameters and indicators they activate, restrict or modify business rules. From the critical or strategic decision-making processes for the business in question, two areas were chosen, which are the most representative in this industry: operational risk control, and corporate management and planning. In this article, we present the first work into operational risk control, whereby we worked with the area of the com- pany with available data and with a Level Two computer- ized management system (the others were ruled out on account of them not having any available data for confi- dentiality reasons or having a Level Zero or Level One computerized management system). In our next publica- tion, we will present our work into the area of corporate management, studying the case of one of these leading companies which also has a Level Two computerized man- agement system. The objective of both these works is to transform the computerized management systems of these decision-mak- ing processes from their current computerization levels with their reactive decision-making processes to computer- ization levels with proactive decision-making processes. This first publication therefore presents the result of applying cutting edge information technologies to one of the operational risk control processes and transforming it from Level Two to Level Five or Six. In particular, the work focuses on the process for controlling the risk of fraud through the use of corporate credit cards as a form of payment. In this respect, the selected process supports one of the widest used differentiation and sustained growth strategies in this industry for obtaining client loyalty. While it is true that the mass issue of credit cards by department stores has been successful as a marketing project, it is equally true that this has increased the risk of exposure to illegal activ- ity, as demonstrated by the growing tendency for fraud which is highlighted in specialist publications (e.g. the latest Cybersource report, Sponsored by CyberSource Corpora- tion Conducted by Mindwave Research, 2006). Diversifica- tion of the client portfolio with this mass issue of credit cards and aggressive marketing plans which encourage the diverse use of this payment method, combined with the lack of efficient techniques and intelligent systems to enable effective detection and prevention of their illegal use, without inconveniencing genuine credit card users, involve the challenge of seeking more efficient methods. This effort is reflected in various articles, in particular in specialist publications which offer different approaches for detecting and preventing this illegal behavior. Never- theless, all of these concur with Bhatla’s observations in 2002 (Bhatla, Vikram, & Dua, 2003) that the evaluated sys- tems are prone to guaranteed effectiveness and that none of the reviewed tools and technologies can alone eliminate fraud. Furthermore, since each technique contributes to the ability to detect fraud, he believes that the most success- ful option would be a combination of several of these techniques, since the results of the Cybersource survey (Sponsored by CyberSource Corporation Conducted by Mindwave Research, 2006) seem to indicate that manual control is still the most used method for detecting and pre- venting fraud. A summary of the state of the art in the techniques and methods used in fraud detection and prevention, and a review of various relevant publications over the last three years confirms the effort employed to obtain useful knowl- Table 1 Distribution of sales between the four major chain stores in Chile Indicator Falabella Ripley Paris La Polar Total No. of stores 30 31 19 26 106 Sales surface area m2 177,538 227,909 154,544 81,000 640,991 Surface area/stores 5,917 7,352 8,134 3,115 6,048 Sales (US$M) 1,178 871 782 360 3,194 % market 36.90% 27.30% 24.50% 11.30% 100 Sales/m2 3.55 2.6 2.6 2.3 Cards issued (M) 3.3 2.6 3.0 1.9 10.8 Active cards (M) 2.6 1.4 1.2 1.4 6.6 % card sales 67% 63% 67% 80% Projected investment (US$M) 1130 551 1200 100 D. Sánchez et al. / Expert Systems with Applications 36 (2009) 3630–3640 3631 https://isiarticles.com/article/17710 
work_agxfmr54pzdobbxo2s4aficvle	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220972998 Params is empty 220972998 exception Params is empty 2021/04/06-03:05:13 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220972998 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:13 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_ahibudfxqrgz5odyehpyedlc6u	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy Integrated expert system applied to the analysis of non-technical losses in power utilities Carlos León a, Félix Biscarri a, Iñigo Monedero a, Juan I. Guerrero a,⇑, Jesús Biscarri b, Rocío Millán b a School of Computer Science and Engineering, Electronic Technology Department, Av. Reina Mercedes S/N, 41012 Seville, Spain b Endesa, Non Technical Losses Department, Borbolla Building, Av. Borbolla S/N, 41092 Seville, Spain a r t i c l e i n f o Keywords: Expert system Data mining Text mining Utilities Power a b s t r a c t The detection of non-technical losses (NTLs), in most papers, commonly deals with the utilization of the registered consumption for each customer; besides, some researchers used the economic activity, the active/reactive ratio and the contract power. Currently, utility company databases store enormous amounts of information on both installations and customers: consumption, technical information on the measure equipment, documentation, inspections results, commentaries of inspectors, etc. In this paper, an integrated expert system (IES) for the analysis and classification of all the available useful infor- mation of the customer is presented. Customer classification identifies the presence of an NTL and the problem type. This IES include several modules: text mining module for analysis of inspector commen- taries and extraction of additional information on the customer, data mining module to draw up the rules that determine the customer estimate consumption, and the Rule Based Expert System module to analyze each customer using the results of the text and data mining modules. This IES is used with real data extracted from Endesa company databases. Endesa is the most important power distribution company in Spain, and one of the most significant companies of Europe. This IES is used in the test phase by human experts in the Endesa company. In this phase, the IES is used as a Decision Support System (DSS), as it contains another module which provides a report with additional information about the customer and a summarized result that the inspectors can use to reach a decision. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Today, in the utility distribution companies, cost reduction is one of the main issues of concern. Therefore, identifying and reduc- ing the non-technical losses (NTLs) are the main objectives. NTL is defined as a non-billed energy due to the existence of irregularities or deviations in the customer facilities. These anomalies or frauds cause an imbalance between the company registered consumption and the real consumption and precipitating serious economic losses, due to the lack of billed consumption. Utility companies like Endesa solve the problem by conducting massive inspections on a customer’s set which satisfies generic conditions. The time and economic costs involved in such inspec- tions are very high. Normally, such large scale inspections achieve 12% success rate at best. This paper proposes an integrated expert system (IES) which globally analyzes all the available information on Endesa’s dat- abases, and differentiates between unnecessary and useful infor- mation. The objective of this analysis is classification of customers under one or more different categories, to help Endesa staff identify and categorize NTLs. This system uses real samples extracted from Endesa’s dat- abases. It is currently in the test phase, and used as a Decision Sup- port System (DSS) although it contains a report-generating module that summarizes the result analysis and its classification. The IES is chiefly composed of several modules: text mining module to collect information on the inspectors’ commentaries, data mining module to set up rules to check if the customer con- sumption is normal, and the Rule Based Expert System (RBES), which uses the generated information by other modules. Besides, it uses a set of parameters (from the domain experts) to classify the customers based on the problems they face. These rules are de- rived from the inspectors and the Endesa staff. In most of the earlier published papers, the authors use several features for NTL detection, like consumption data, economic activ- ity, active/reactive rate, and contracted power. The proposed solu- tion uses all the available customer information that the company has like consumption reports, information on the measuring equip- ment, inspectors’ commentaries, documentation, etc. Thus, while the classification process is being performed, time and economic costs are simultaneously reduced. The efficiency accordingly in- creases due to a cut in false NTLs inspections. 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.02.062 ⇑ Corresponding author. Tel.: +34 679231193. E-mail address: juaguealo@us.es (J.I. Guerrero). Expert Systems with Applications 38 (2011) 10274–10285 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy The proposed IES is a part of the development project in collab- oration with Endesa one of Spain’s largest companies, including more than 10 million customers. In this paper, the proposed IES is explained in great detail, under several sections: – The Bibliographical Review. In this section, authors do a complete review of the papers related with the techniques and technolo- gies used in the IES. – The MIDAS Project section explains the project in which this IES is developed. Also, several concepts and procedures related with the management and control of the customers’ consumption are discussed. – The System Architecture section reveals the architecture of the pro- posed IES, and the sections following it describe each of it modules: data mining, text mining and RBES modules. – The Section 9 describe the current results of IES application. In the Section 10, current researches are briefly presented. 2. Bibliographic review Frauds and anomalies represent a serious problem for utility companies. Advancement in the different techniques and method- ologies related with the Artificial Intelligence has enabled the detection, classification, reduction, and prediction of the problems of frauds and anomalies. The proposed IES employs several techniques to detect and clas- sify customer abnormalities related with frauds and other anoma- lies (NTLs). This system combines RBES, data mining and text mining. Several areas of research related with this work are pres- ent, although based on pattern detection: for example, the main fields of research include those with high economic risk: finances and economics, telecommunications and others. 2.1. Finances and economics Sánchez, Vila, Cerda, and Serrano (2009) use the fuzzy associa- tion rules for fraud detection in credit card transactions; Quah and Sriganesh (2008) use the auto-organization maps to decipher, filter and analyze the customers’ behavior in fraud detection; Kirkos, Spathis, and Manolopoulos (2007) use several data mining tech- niques (decision trees, neural networks and Bayesian belief net- works) analyzing the factors related with Fraudulent Financial Statements (FFS) comparing the efficiency among all the methods. Besides, Richardson (1997) compares statistical and neural net- work techniques. Wheeler and Aitken (2000) use case-based rea- soning to detect frauds in the credit approval process. All these papers propose different techniques, related in large part with data mining, demonstrating their efficiency in fraud detection. Also, Hand and Blunt (2001) have made a significant contribution by an in depth treatment of all the prior information analysis pro- cesses for application of data mining techniques. They classify cus- tomers on the economic sector in which the transfer is made. There are others subjects within the scope of this research field which use very similar methods, although based on other features, for in- stance, anti-money laundering (Gao & Xu, 2009), risk evaluation (Yu, Yue, Wang, & Lai, 2010), credit scoring (Huang, Chen, & Wang, 2007), and financial distress (Chen & Du, 2009). 2.2. Telecommunications In this field of research several other methods and tools related with data mining (Daskalaki, Kopanas, Goudara, & Avouris, 2003; Wang, Wang, Zhan, Li, & Wang, 2004), Support Vector Machine (SVM) (Wang et al., 2009), neural networks and fuzzy rules (Estévez, Held, & Pérez, 2006) and Expert Systems (Hilas, 2009) are applied. These techniques identify various types of trouble solving, though they will need to use the domain expert knowledge at a certain level. 2.3. Others There are several other research fields in which these tech- niques are applied for pattern detection: medicine (Cierniakoski, De, & May, 1991; He, Wang, Graco, & Haukins, 1997; Yang & Hwang, 2006), Intrusion Detection or IDS (Depren, Topallar, Anarim, & Ciliz, 2005; Hernández-Pereira, Suárez-Romero, Fontenla-Romero, & Alonso-Betanzos, 2009; Kim, Im, & Park, 2010), transport (Chen, Qu, & van Zuylen, 2010), etc. All these papers use different search methods to detect anomalous of fraudulent patterns, ruling out the others by any of the means, except in the case of Richardson (1997). He establishes a classification system on several groups based on the problems involved in each case. Unlike these papers, the electrical energy customer usually dis- plays more than one type of pattern that indicates the same NTL. Therefore, the IES classifies customer employing several categories. 2.4. NTL detection There are several papers focusing on NTL detection: Galván, Elices, Muñoz, Czernichow, and Sanz-Bobi (1998) pro- posed a general methodology based on using Radial Basis Function Networks (RBFN), with the following steps: (1) variable selection, (2) data filtering, (3) Model fitting, (4) Model analysis, and (5) Model evaluation. The third step, the RBFN input, is taken of the variables monthly periods of each annual consumption pattern and active/reactive consumption. For instance, the methodology is applied in two economic sectors: low-voltage lodging sector and high-voltage farm watering sector. Cabral, Pinto, Onofre, Gontijo, and Filho (2004) proposed an application that used rough sets to classify the categorical attribute values to detect fraud in customer electrical energy use. The contin- uous attributes were converted to discrete. The system achieved a fraud rightness rate around 20%. The authors expressed that the main difficulty to detect electrical energy profiles was the low ‘fraudulent customer’/’normal customer’ ratio, around the 5%, in Brazilian electrical energy distribution companies. They also admit- ted that to add to their problems, many fraudulent customer behav- iors appeared as normal behavior. Cabral, Pinto, Linares, and Pinto (2006) added Knowledge Discovery Databases (KDD) to improve the success of fraud pattern detection. In the previous version, they performed a KDD process by selecting 12 attributes (2 string and 10 numerical). Unlike these papers, the proposed IES does not include a vari- able selection process, as this selection is performed in a previous data mining step, according to the domain expert knowledge. Be- sides, it uses several modules to adapt unstructured and unusual information. It enables the creation of a set of rules to improve and fine tune the NTLs detection results. Therefore, all the reviewed papers reveal several ideas in common: – Using different techniques related with the data mining and the pattern detection. – In the papers directly related with the NTL detection in energy consumption, only a few key indicators are used: the energy consumption, the economic activity, the contracted power and the active/reactive ratio. Much of the interesting information, for instance the reports of the NTL inspectors, is rejected. The present paper proposed a new integrated expert system, which includes all the available information in Endesa’s databases, C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 10275 Author's personal copy using real inspection results to test the system. The domain expert knowledge is obtained from the Endesa staff. To incorporate the knowledge three techniques are used: – The static knowledge is included in the RBES. – The data mining technique involves different statistical tech- niques to create rules related with the estimated consumption that a customer without NTL would have. – The text mining technique comprises knowledge of the custom- ers’ facilities provided by Endesa’s inspectors. The text mining objective is unstructured information, using natural language, and it can be used to generate new information (Yang & Lee, 2004), to extract information (Sung & Chang, 2004), to summa- rize information (Aliguliyev, 2009), etc. On the contrary, Schutzer (1990) suggested applying a business expert system for fraud detection. Liao (2005) submitted refer- ences for fraud management, and Rahman and Lauby (1993) pro- posed the use of expert systems in Power System Planning. These papers advanced the usefulness of an expert system in the utility distribution field, to implement the domain expert knowl- edge. In recently published papers, authors have not produced pa- pers related with electrical NTLs detection containing a RBES. However, there are papers relating to fraud detection in other re- search fields. For example, Cierniakoski et al. (1991) for processing medical insurance claims; Bowen (1994) for police investigators of economic crimes; and Hilas (2009) for fraud detection in private telecommunications networks. 3. The MIDAS Project The aim of the MIDAS Project is NTL detection by utilizing the data mining techniques and other Artificial Intelligence (AI) tech- niques over Endesa’s databases. Initially, the project began with the NTL pattern detection. These advancements are published in various papers (Biscarri et al., 2008; Biscarri et al., 2009), where different techniques of data mining and neural networks are used to detect the consumption pattern by which the NTLs they identify with are proposed. Supervised and non-supervised techniques too are applied. The information flow diagram is shown in Fig. 1 depicting the research methodology and the IES role in the project. The steps of this cycle are: – Sample Selection. A set of customers is selected. Contracted power, economic sector and geographic location features are used in this step. – Sample extraction from corporate databases. The main the diffi- culty in this step is the large number of customers. Two extrac- tion methods may be used: designing of an extraction system for batch processing or designing of an intermediate database. The first method is used at night or during inactive database periods. The second method, used in the IES, works with the off-line information and it is actualized periodically using a batch process. – Application of studies based on data mining and AI using customer consumption information. The output is a list of customers ‘sus- pected’ of NTL. These studies are described in detail by Biscarri et al.. (2008), Biscarri et al.. (2009). – Analysis of customers. In this phase, the remaining available information on the customers is analyzed, to determine if those with a ‘suspected’ anomaly are wrongly classified. Normally, this process demands much time and effort from the inspectors or domain experts, as it is necessitates a review of all the cus- tomer information manually, one at a time. This step will be replaced with the proposed IES that offers an automatic and accurate analysis. – Customer review. The final conclusions, obtained through the application of studies and the customer analysis are verified with ‘in situ’ inspections. – Review of the results. The results obtained by the inspectors are checked. This information improves the future studies and the IES. Fraud identification per number of ‘in-situ’ inspection per- centage as reported (15–20%, or higher depending on customer characteristics). Actually, after analyzing the ‘in-situ’ inspections conducted by the Endesa staff, the authors conclude that the utilization of addi- tional techniques becomes necessary to detect the false-positives, i.e., the customers who present an NTL pattern but who are not fraudulent or anomalous. We had earlier experienced that the number of false-positives was high. This is a serious problem in all the utility companies; and there is no easy solution as customer consumption depends on several factors. Also, the high cost associated with ‘in-situ’ inspections poses a great limitation. The inspections conducted to identify NTL are more expensive than a standard revision or equipment mainte- nance, as more qualified inspectors are needed. The proposed IES attempts to meet this need by adding new verifications to the selected customers utilizing any of the data mining methods, to minimize the false-positive cases. Project development performed by the SPSS Clementine envi- ronment is commonly used in the commercial development of the data mining process. Liao and Wen (2007) and Liao, Hsieh, and Huang (2008) presented an interesting review of some fea- tures of this software, besides proposing utilization examples. 4. The energy distribution management and the expert system Endesa, similar to other power utilities companies, implements several procedures to manage the energy distribution. For exam- ple, this company establishes contract procedures which lay the groundwork to create a new contract for the customer. Likewise, the company inspectors routinely check customer installation. These inspections follow established guidelines, depending on the requests by the company. Normally, the inspectors keep adding on new information to the corporative databases or modify the existing ones on customer installation. This information includes inspector comments on any thing or any event which the inspector has observed. However, in the case of NTL search, these comments summarize inspector observations and the procedural steps taken to rectify it. When the company inspectors identify any abnormality in the customer measurement system, they notify the company. All the information related with identified NTL are reported and stored in Endesa’sFig. 1. The information flow diagram of the MIDAS Project. 10276 C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 Author's personal copy databases. Thus, the automatic processing of this information, involving an adequate expert system, is clearly very useful in deter- mining NTL patterns, and consequently to reduce false NTL ‘in-situ’ inspections. Endesa staff measure customer consumption by installed mea- surement registers. The measurement period could either be once per month (monthly) or one per 2 months (bimonthly). These data are carefully stored in corporate databases for billing purposes. Sometimes, the readings cannot be reported due to a reading error or causes outside the scope of the reader; for example, the worker cannot access the customer register. In such a scenario, the Endesa system automatically calculates an estimate of this measure which the company terms ‘estimated consumption’, and these are usually lower than the real consumption. For the NTL detection process, the estimated consumptions need to be carefully studied, as they could actually hide an NTL. Authors noted several cases with zero real consumption, resulting from an abnormality or a fraud, and a low billed estimated consumption, for several reasons. Therefore, for the expert system, the number of customers estimated readings becomes highly significant data. All measurement actions are carried out at the customer’s instal- lation and the data are stored in Endesa’s databases. As Endesa has more than 10 million customers, such large numbers of customers need large databases and complex hardware architecture. When future customers request the electric energy service, they need to provide the company with personal information and fur- ther details about his economic activity, his electric supply require- ment and his selected contractual power rate. The contractual rate establishes the power range and voltage to be installed on the customer property. Besides, the contractual rate provides the measuring and control equipment features, by which the price of the billed energy can be established, depending on the time band consumption and the power contracted. RBES checks customer consumption relative to the contractual information. For example, there are some customers for whom the consumption is assembled in particular time band discriminations. 5. Expert system architecture The proposed IES is the result of a research of the all informa- tion available in company databases. Each type of information calls for its own technique based on the information type and search objective. The techniques and technologies listed below are used in the following manner: – RBES module. This technology is used to include the domain expert knowledge. This is static knowledge, and normally, excludes the learning processes. This technology is added to the system as the main module. It uses the information gener- ated by the other modules to include the dynamic knowledge or mining learning. – Data mining module. This technique includes several statistical techniques, like basic statistical indicators and regression tech- niques. It is used to determine whether the consumption is cor- rect based on the geographic location, contracted power, billing frequency, time band discrimination, postal code, and economic activity. This module includes a learning process (described in Section 6). – Text mining module. This technique permits the inclusion of the inspectors’ knowledge of certain customers. This information is chiefly used to determine if false-positives exist. It enhances the efficiency of the classification process. This module too, requires a learning process (described in Section 7). The IES includes several modules which interact with the knowledge base. Fig. 2 shows the basic architecture of the system. The Auxiliary Tools perform several tasks, to check on NTL’s ‘suspicious’ customers and to summarize the accumulated results. This module is not used in the classification process, but is very useful for the company to create summary reports. The core of the expert system, text mining and data mining modules are done using SPSS Clementine. The knowledge base and sample storage are done using MySQL. The connection is main- tained through ODBC drivers. The data mining and text mining modules update a certain part of the knowledge base, because of which, they must be applied in anticipation of customers’ studies. These modules require only peri- odic update. The data mining module particularly needs to update its knowledge base, either monthly or bimonthly. The updating time process depends on the number of customers. For example, with all customers in the low voltage sector (this sample is termed ‘Low Voltage Sample’ or ‘LV Sample’) this process could take 3 h (the features of the computer test are specified in Section 9). The text mining module needs to annually update its knowledge base or when there is an important change in the company. Following the earlier example, this process was executed in 2 h. 6. The data mining module The data mining module has to conduct either a monthly or bi- monthly learning process, as described in this section. Then, the information generated by this module is used in the analysis pro- cess, by the IES. The data mining model of this module thus helps to establish relations among customers’ consumption, and it cate- gorizes them using statistical techniques by estimation of the con- sumption curve in customers’ sets, fulfilling several conditions. The process begins with filtering customers who present a spe- cific abnormality in consumption. This module establishes the nor- mal consumption range for various customers’ sets. The filters applied are: – Elimination of customers possessing a high number of esti- mated measurements. – Elimination of customers with three or more consecutive zero consumption readings. – Elimination of customers showing highly dispersed consump- tion. For example, a customer with three or more consumption peaks greater than twice the standard deviation of consumption. Normally, the electric consumption is highly dependent on a few contract features. The system must select the relevant cus- tomer consumption information, available in the corporate dat- abases. The selected parameters are: – Contracted power. First, the customers are divided into two main groups, the low- and high-contracted power types, based on Fig. 2. System architecture. C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 10277 Author's personal copy several practical studies. These studies show that the majority of customers have contracted electric power less than 300 kW. However, customers with higher contracted power reveal different consumption behaviour. Next, each main group is divided into 40 different subgroups using 2 vingtiles, each containing an identical number of cus- tomers. The vingtiles are numbered from 1 to 20 for the 0 to 300 kW range, and 21–40 for customers with contracted power higher than 300 kW. Usually, customers with similar contracted power display a similar range of consumption. Domestic customers, the most numerous of the clusters, contract low power, normally between 3 and 13 kW. Small businesses, like pubs, restaurants and shops, contract a power greater than 13 kW. Normally, the interval ranging between 0 and 3 kW is assigned to stores, warehouses or auxiliary support. Contracted power higher than 300 kW is associated with industrial or distribution activities. – Geographic location and postal code. These data indicate the cus- tomer location enabling the formation of geographic zones including customers with the same climatic and administrative conditions. – Economic sector. This feature is also strongly related to customer consumption. Similar to what Hand and Blunt (2001) proposed, a clear relationship does exist between economic activity and consumption. – Time band discrimination. This parameter enables the formation of customer sets based on the time band consumption. The aim is to determine a pattern of consumption for each time band. – Billing frequency. This determines measurement as a regular recurrence. It allows tuning the statistical studies circumvent- ing the need to make interpolations. However, these interpola- tions are necessary if the cycles of the readings are not the same for the customers. Normally, the time spans are either monthly or bimonthly. The parameters presented allow the establishment of customer sets with similar behavior, by applying statistical indicator. Four sets are established to fix the normal levels of the indicators, which would show possible customers’ statistical abnormalities. These sets are formed by an aggregation of the following parameters: – The geographic location, the contracted power and the billing frequency (Group A). These parameters establish the usual con- sumption, in a particular geographic location. Through this group, the system extracts group consumption patterns of the customers based on these features in the ‘normal’ consumption range. Fig. 3 shows the monthly consumption average graph versus that of the contracted power, for a region in the north of Spain clearly indicating the relationship between customer consumption and contracted power. – The geographic location, contracted power, billing frequency and postal code (Group B). In the IES-conducted analysis this set allows the detection of seasonal behaviours, for example, the customers related the increase in their consumption in summer to tourism. Fig. 4 shows the common contracted power in a particular postal code and the range of consumption within this zone. This postal code corresponded to a particular place in northern Spain. – The geographic location, contracted power, billing frequency, and economic sector (Group C). This group establishes the depen- dence of the customers’ consumption on its economic activity. In the analysis performed by the proposed IES, this information is used to establish the estimated consumption curve for a par- ticular economic sector. The system assumes that the custom- ers’ economic activity data from Endesa’s databases, is true, but it also implements processes to check the veracity. Fig. 5 shows the common contracted power in a particular economic sector and the range of consumption within this sector. This sector corresponded to restaurants and pubs in northern Spain. – The geographic location, contracted power, billing frequency and time band discrimination (Group D). Studying this group is important as it helps to establish the consumption range within a time band, according to the features mentioned above. In the analysis performed by the proposed IES, it is combined with the results from the other sets. Figs. 6(a) and (b) show the differ- ence between customer consumption within or without the dis- Fig. 3. Graph showing the average monthly consumption (kWh) of a geographic location (in northern Spain). 10278 C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 Author's personal copy crimination band. The consumption curve shows that when several discrimination bands are present, a high consumption range will be seen. The statistical indicators used in these studies are the average and the standard deviation of customer consumption calculated for each group mentioned. Also, a time division is made for each group. Temporal parameters depending on the measurement peri- od are used to make the temporal divisions. These divisions are absolute (all the electricity consumption during the study period), yearly, seasonally, and monthly consumption. A consumption regression study is made to identify the consumption trends. Trends can explain the reduction in consumption in the customers analyzed, if the energy demands in a particular geographic location and economic sector decrease. 7. Text mining module The text mining module learning process is discussed in this sec- tion. Following the learning process, the IES uses the information Fig. 4. Graph showing the average monthly consumption in the postal code XXX03. Fig. 5. Graph showing the average consumption of kWh in the economic sector of restaurants and pubs. C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 10279 Author's personal copy generated in the analysis process. The main objective of this text mining module is the extraction of the information from the inspec- tor commentaries and from the company documents on customer facilities. This is non-structured information, couched in natural language. Usually, the power company inspectors check the cus- tomer installations from time to time, and enter their results in the corporate databases. The proposed IES uses this very interesting information to improve on the analysis of the NTL research. A com- mon term dictionary has been compiled to be used on a set of rules that execute the following actions: – Economic sector checking. It is necessary to determine whether the customer informed activity sector concurs with the real information. For instance, if is possible that the customer has specified the economic sector, which has changed or not been correctly informed possibly due to a contract change. Such a scenario would make the analysis of the consumers difficult; therefore, it becomes useful to detect this abnormality and identify the customer’s economic sector. – Characterization of the customers’ consumption. Inspectors’ reports or technical repair reports can justify certain circum- stances that can occur in the customer’s installation: for exam- ple, the estimated measures, the low or even zero consumption or the work inspector’s orders, repetitively, cancelled. We have detected cases ranging from very little to frequent cases with inaccessible or dangerous access; that explains the infrequent reading measurements. Inspectors report these situations. – Other information checking. These include concepts extracted from the commentaries compiled by the technical staff mem- bers, which in turn facilitates collecting additional information Fig. 6. Graph showing the average consumption of kWh on customers (a) without time band discrimination and (b) with three time bands discrimination. 10280 C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 Author's personal copy on the customer’s installation, such as the existence of a capac- itor bank, which explains the low reactive energy consumption or information about when the measurement installation is to occur. The text mining module is applied to every database field that contains information in natural language. This process is per- formed prior to the IES analysis, as it is necessary to update the knowledge base with the information collected by the text mining module. The following steps are necessary to correctly realize this process: 1. Data Preprocessing. Inconsistent, erroneous or missing informa- tion must be corrected. Besides, both data integration and data transformation are performed. 2. Concepts’ Extraction. It attempts to elicit the text field concepts, structured or otherwise. A concept can comprise one or more words. The concept may be a syntagm or a word which repre- sents an entity (action, event, etc.). Natural Language Processing (NLP) methods are used, to extract linguistic (words, phrases, etc.) and non-linguistic (dates, numbers, etc.) concepts. An interesting review of this technique and its use in the informa- tion system management is proposed by Métais (2002). The fol- lowing set of functionalities are included: a. Recognition of punctuation errors. These types of mistakes include the incorrect use of the tilde, the period, the comma, the point and comma, the dividing bar, etc. Frequently, the text fields contain commentaries in very colloquial language, with less attention being paid to the correct placement of these punctuation signs. b. Recognition of spell errors. A grouping fuzzy technology is applied. When concepts of the text are extracted, words with similar spelling (referring to the letters that compose it) or that are closely related are classified together. By applying this algorithm, mistakes of omission of letters, duplication of letters or permutation of letters are corrected. c. Dictionaries. A dictionary of technical words is compiled, as well as a dictionary of synonyms and abbreviations to help the system recognize the concepts of a more sophisticated form. Also, dictionaries of undesirable concepts are estab- lished, to determine the words or concepts that are rejected in the recognition process. In Spain, several dialects exist, which complicate extraction of the concepts. 3. Categorization. In this step, each concept is qualified and classi- fied based on the inspectors’ knowledge. This process classifies the key concepts or words concurrent with the functions previ- ously specified: a. Identification of a possible change of economic sector. This clas- sification allows the detection of 90 different economic sectors. b. Characterization of the customer consumption. Customer con- sumption characterization is mainly focused on detecting the justification of the anomalous measurements or of the consumption anomalies. Another utility is to eliminate minor concepts, categorizing and classifying them within groups that are designated as sets of little interest, which prevent them from being overlapped or confused with others. c. Diverse information of the customer installation. 8. Rule based expert system module The RBES controls and monitors the analysis process. RBES em- ploys a set of rules to determine the customers’ problems. In the proposed IES, this knowledge is represented by rules of type IF- THEN-ELSE. The representation of the objects and results is done with a dynamic table, in which each row represents a contract of a different customer. The result of applying a rule is entered in a column that identifies its origin. More than 500 rules (including the rules that generated the text mining and the data mining mod- ules), help in classifying customers under seven different groups, based on their intentions: – Generation rules. These rules deal with the generation of new information from the existing customers. They are used to update erroneous information and to preprocess the database information. These rules include the rules generated by the text mining process. For example, the cycle consumption calculation (a cycle is the period of time between taking a measurement and the follow up) is performed by applying the rule, as shown in Table 1. This rule calculates the corresponding consumption between a measure (measure) and the previous one (previ- ous_measure) which are registered in the company databases. The information generated is applied to the other groups of rules. – Classification rules for contract mistakes. These rules deal with the detection of information mistakes in the customer’s con- tract information. In Table 1, as shown. This rule, for example, allows the classification of customers with very low contracted power. These customers cannot be analyzed in the same man- ner as the normal or high contracted power customers are stud- ied. They are analyzed using the text mining rules. – Classification rules for facility problem. The objective is to detect information problems related to the measuring equipment. In Table 1, an example of a rule is shown, that selects customers using obsolete measuring equipment. – Classification rules for consumption problem. The application of this rule set is based on the results of the previous rules, along with the information generated in data mining; the customer is classified according to his electrical consumption. Also, it estab- lishes the limits of low consumption, excessive consumption or null consumption. The initial values for these parameters are determined during the knowledge acquisition process. Low con- sumption particularly, usually indicates the possibility of an NTL. Detection of low consumption customers employs the rules generated with data mining as well as the text mining pro- cess. There are several rules to determine a low consumption case, regarding the history of customer consumption or consid- ering customer’s cluster. An example, as shown in Table 1. This rule compares customer consumption (customer_cycle_con- sumption) with the results of statistical studies in which the Table 1 Some RBES rules. Group of rules Rules Generation rules CYCLE CONSUMPTION=IF measure>previous_measure THEN measure- previous_measure ELSE (10maxlmum-number-of-digits- previous_measure-1+measure) Consumption problem classification rules LOW_CONSUMPTiON=IF range_con_month_group_a > customer_cycie_consumption and range_con_month_group_b > customer_cycle_consumption and range_con_month_group_c > customer_cycte_consumption and range_con_month_group_d > customer_cycle_consumption THEN TRUE ELSE FALSE Contract power classification rules POWER_VERY_LOW=IF customer_contracted_power < 1.5 THEN TRUE ELSE FALSE Facility problem classification rules LOW_NUMBER_WHEELS=IF number_wheels<=4 THEN TRUE ELSE FALSE C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 10281 Author's personal copy consumption range is established (range_con_month_group_a, range_con_month_group_b, range_con_month_group_c and ran- ge_con_month_group_d). In this case, the customer’s consump- tion is compared with the low consumption range of each group described in Section 6. – Rules of customer selection for inspection. These sets of rules pro- pose a list of customers to be inspected, identified from the pre- vious rule conclusions. This selection classifies customers under several categories, which indicate the risk of NTL. It is possible to solve the NTL without making an ‘in-situ’ inspection, where the problem can be solved by updating customer information in the company databases. – Verifying rules. Due to the irregularities in the information avail- able in the corporate databases, rules to check the veracity of the systems results become an absolute necessity. These rules determine if the customer has been correctly analyzed, and also if the customer cannot be reported due to too much incoherent information collected. – Explanation rules. This system strengthens the reports using these IF-THEN-ELSE rules, adding interesting information on the customer or his installation. The reported information includes: o Contractual information like billing frequency, time band dis- crimination, contracted power, postal code, economic activ- ity, etc. o Problems related with contract as wrongly applied rate, incorrectly contracted power, incoherent information, etc. o Problems related with measuring equipment: old register, warn register, lack of obligatory power control switch, etc. o Information and problems related with consumption: charac- terization of customer’s consumption, unexpected low con- sumption, etc. o Other consumption operations, as the average consumption on any consumption time band. 9. Experimental results The tests were run on a computer with a double processor AMD Opteron dual core (1, 7 Ghz), 3Gbytes of RAM and 100 Gb of hard disk space. This IES is part of a project named MIDAS, as shown in Fig. 1. The IES have been designed to analyze the samples obtained by the application of other detection studies, based on Artificial Intel- ligence, statistical studies, and neural networks. The IES was applied over a set of contracts ‘suspected’ of fraud or abnormalities that had been obtained from studies described in Biscarri et al.. (2008), Biscarri et al.. (2009). This ‘suspected cus- tomer set’ consists of 134 contracts. These contracts are selected as they reveal an anomalous con- sumption pattern. The IES analyzes this set and configures it in the most restrictive manner. All the selected contracts will be in- spected, and the number of inspections is highly restricted because they are very expensive. In this sense, the IES selected: – Thirty two contracts with a possible NTL. – Sixteen contracts with incoherent information or including some problem which could be solved without inspection. – Eighty six contracts without a clear NTL, as they had been solved previously or those where the anomalous consumption can be explained. These contracts were classified as false NTLs. This set of 32 contracts was inspected ‘in situ’. The results of these inspections were: – Nine contracts have an NTL. – Two contracts have measurements problems. – Fourteen contracts have a non-NTL. – Two contracts are not in force, and not included in Endesa dat- abases as yet. – Five contracts cannot be inspected, because the customers’ businesses had closed down, and it had not been entered in the Endesa databases as yet. The IES thus filtered 86 cases of false NTLs. A review of four fil- tered cases is shown below. These cases have an NTL pattern, although they have several data which explain the anomalous consumption. The first case has an anomalous consumption for several rea- sons. Between 2005 and 2007 there was a period of decrease in consumption. In 2007, the company started a proceeding, which Fig. 7. Consumption graph of a customer with previous proceeding. 10282 C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 Author's personal copy solved the problem. Later, there were some cycles of low consump- tion, which were explained in the inspectors’ commentaries. They specified that the place had closed and the measurement was esti- mated. Fig. 7 shows the consumption of the customer in a different time discrimination band. The second case is a customer with a contract for a fountain and irrigation engine. The consumption of this type of activity is very irregular, and usually involves a high rate of reactive energy. The IES classified the customer as a non-NTL case. The inspectors’ commentaries specified that it was not possible to access and measure the equipment because it was padlocked and damaged. Besides, these types of customers show very irregular consump- tion as they depend on the vagaries of the climatic conditions, and the IES has no information on these. Fig. 8 shows the con- sumption of this customer in a different time discrimination band. The reactive energy is sufficiently high due to the latency period of the engine. The third case is a customer with a hotel. Usually, such clients have five time discrimination bands making the analysis more dif- ficult. Fig. 9 shows this customer’s consumption in a different time discrimination band. The rules generated by the Data and Text Mining modules are very important to correctly classify this case. Using these rules, several reasons were detected: the customer’s consumption is established by the tourist industry (normally, it is seasonal); and the period of zero consumption was explained by the inspectors’ commentaries, as a result of a problem with Fig. 8. Consumption graph of a fountain and irrigation engine. Fig. 9. Consumption graph of a seasonal case. C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 10283 Author's personal copy the measuring equipment and was successfully solved. The IES classified this as a non-NTL case. The fourth case is a customer with zero consumption, as he had several contracts in the same place. The inspectors’ commentaries specified that the supply was used as a supplement and for occa- sional use only. The IES classified this case as a non-NTL case. 10. Conclusions The present paper includes the investigation of the expert sys- tem in the power utility consumption subject, and its combination with other technologies to further enhance the efficiency. This paper particularly significantly contributes to an as-yet lit- tle exploited subject – the automatic analysis of available customer information of the utilities, on NTL classification. The main area of complexity of this research field was the enormous quantity of information required and the great variety of casuistry that is pres- ent in the customer analysis. As evident in the bibliographic review section, customer analysis is based on customer consumption. When the inspector’s knowledge is included, new techniques need to be added to treat it. The main contributions done by the IES in this paper are: – Identification and classification of the casuistry of utility distri- bution on customer analysis. – This system increases the availability of the quantity of useful information on the customer. – Increasing of the efficiency, regarding massive inspections. The utilization of additional available information about the cus- tomer (in addition to customer consumption) helps to greatly increase the success ratio. – Classification of the normal and NTL cases. The normal detection methods are dedicated only to select the anomalous cases: the present IES makes a classification of the normal and NTL cases. The proposed IES is a real framework, actually in the test phase, in Endesa. – The proposed IES can be used as an additional method to detect NTL increasing the efficiency of the studies. In reality, we researched new techniques and technologies to im- prove the efficiency of this IES. From this viewpoint, we are research- ing the use of the following techniques. – Text mining. In the proposed IES, the inspector’s knowledge in the categorization process is used. The customer’s consumption will be added to this process to automatically make the catego- rization process. Thus, each rule created by text mining has sev- eral associated concepts, one associated action and several features related with consumption. – Data mining. An additional improvement in curve consumption calculation will be researched. This new feature will include the time series and Bayesian networks with the use of events or intervention fields like special independent fields that will be used to model the effects of external occurrences. – Data warehousing. The data mining and text mining techniques will be completed by employing data warehousing technology. This technology will improve the efficiency of the mining process. – Real-time analysis. A new module to reduce the analysis time will be added. This module will determine when a customer will have new information available to analyze it. References Aliguliyev, R. M. (2009). A new sentence similarity measure and sentence based extractive technique for automatic text summarization. Expert Systems with Applications, 36, 7764–7772. Biscarri, F., Monedero, I., León, C., Guerrero, J., Biscarri, J., & Millán, R. (2008). A data mining method based on the variability of the customers consumption. In 10th international conference on enterprise information systems, ICEIS2008, June 12–16, Barcelona, Spain. Biscarri, F., Monedero, I., León, C., Guerrero, J. I., Biscarri, J., & Millán, R. (2009). A mining framework to detect non-technical losses in power utilities. In 11th international conference on enterprise information systems, ICEIS2009, May 6–10, Milano, Italy. Bowen, J. E. (1994). An expert system for police investigators of economic crimes. Expert Systems with Applications, 7(2), 235–248. Cabral, J. E., Pinto, P., Onofre, J., Gontijo, E. M., & Filho, J. R. (2004). Fraud detection in electrical energy consumers using rough sets. In 2004 IEEE international conference on systems, man and cybernetics (Vol. 4, pp. 3625–3629). Cabral, J. E., Pinto, J. O., Linares, K. S. C., & Pinto, A. M. A. (2006). Methodology for fraud detection using rough sets. In 2006 IEEE international conference on granular computing. IEEE Press. Chen, W.-S., & Du, Y.-K. (2009). Using neural networks and data mining techniques for the financial distress prediction model. Expert Systems with Applications, 36, 4075–4085. Chen, S., Qu, G., & van Zuylen, H. (2010). A comparison of outlier detection algorithms for ITS data. Expert Systems with Applications., 36(8), 10976–10986. Cierniakoski, J. J., De, R., & May, J. H. (1991). MEDIN: An expert system for processing medical insurance claims. Expert Systems with Applications, 2, 211–218. Daskalaki, S., Kopanas, I., Goudara, M., & Avouris, N. (2003). Data mining for decision support on customer insolvency in telecommunication business. European Journal of Operational Research, 145, 239–255. Depren, O., Topallar, M., Anarim, E., & Ciliz, M. K. (2005). An intelligent intrusion detection system (IDS) for anomaly and misuse detection in computer network. Expert Systems with Applications, 29, 713–722. Estévez, P. A., Held, C. M., & Pérez, C. A. (2006). Subscription fraud prevention in telecommunications using fuzzy rules and neural networks. Expert Systems with Applications, 31, 337–344. Galván, J. R., Elices, A., Muñoz, A., Czernichow, T., & Sanz-Bobi, M. A. (1998). System for detection of abnormalities and fraud in customer consumption. In 12th conference on the electric power supply industry. November 2–6, Pattaya, Thailand. Gao, S., & Xu, D. (2009). Conceptual modelling and development of an intelligent agent-assisted decision support system for anti-money laundering. Expert Systems with Applications, 36, 1493–1504. Hand, D. J., & Blunt, G. (2001). Prospecting for gems in credit card data. IMA Journal of management Mathematics, 12(2), 173–200. He, H., Wang, J., Graco, W., & Haukins, S. (1997). Application of neural networks to detection of medical fraud. Expert Systems with Applications, 13(4), 329–336. Hernández-Pereira, E., Suárez-Romero, J. A., Fontenla-Romero, O., & Alonso- Betanzos, A. (2009). Conversion methods for symbolic features: A comparison applied to an intrusión detection problem. Expert Systems with Applications, 36, 10612–10617. Hilas, C. S. (2009). Designing an expert system for fraud detection in private telecommunications networks. Expert Systems with applications, 36, 11559–11569. Huang, C.-L., Chen, M.-C., & Wang, C.-J. (2007). Credit scoring with a data mining approach based on support vector machines. Expert Systems with Applications, 33, 847–856. Kim, H. R., Im, H. K., & Park, S. C. (2010). DSS for computer security incident response applying CBR and collaborative response. Expert Systems with Applications., 37(1), 852–870. Kirkos, E., Spathis, C., & Manolopoulos, Y. (2007). Data mining techniques for the detection of fraudulent financial statements. Expert Systems with Applications, 32, 995–1003. Liao, S.-H. (2005). Expert system methodologies and applications – A decade review from 1995 to 2004. Expert Systems with Applications, 28, 93–103. Liao, S.-H., Hsieh, C.-L., & Huang, S.-P. (2008). Mining product maps for new product development. Expert Systems with Applications, 34, 50–62. Liao, S.-H., & Wen, C.-H. (2007). Artificial neural networks classification and clustering of methodologies and applications – Literature analysis from 1995 to 2005. Expert Systems with Applications, 32, 1–11. Métais, E. (2002). Enhancing information systems management with natural language processing techniques. Data & Knowledge Engineering, 41, 247–272. Quah, J. T. S., & Sriganesh, M. (2008). Real-time credit card fraud detection using computational intelligence. Expert Systems with Applications, 35, 1721–1732. Rahman, S., & Lauby, M. (1993). Identification of potential areas for the use of expert systems in power system planning. Expert Systems and Applications, 6, 203–212. Richardson, R. (1997). Neural networks compared to statistical techniques. computational intelligence for financial engineering (CIFEr). In Proceedings of the IEEE/IAFE 23–25 March 1997, pp. 89–95. Sánchez, D., Vila, M. A., Cerda, L., & Serrano, J. M. (2009). Association rules applied to credit card fraud detection. Expert Systems with Applications, 36, 3630–3640. Schutzer, D. (1990). Business expert systems: the competitive edge. Expert Systems with Applications, 1, 17–21. Sung, N. H., & Chang, Y. S. (2004). Business information extraction from semi- structured webpages. Expert Systems with Applications, 26, 575–582. Wang, D., Wang, Q.-Y., Zhan, S.-Y., Li, F.-X., & Wang, D.-Z. (2004). A feature extraction method for fraud detection in mobile communication networks. In 10284 C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 Author's personal copy Fifth world congress on intelligent control and automation, WCICA 2004 (Vol. 2, pp. 1853–1856). Wang, S.-J., Mathew, A., Chen, Y., Xi, L.-F., Ma, L., & Lee, J. (2009). Empirical analysis of support vector machine ensemble classifiers. Expert Systems with Applications, 36, 6466–6476. Wheeler, R., & Aitken, S. (2000). Multiple algorithms for fraud detection. Knowledge- Based Systems, 13(2–3), 93–99. Yang, W.-S., & Hwang, S.-Y. (2006). A process-mining framework for the detection of healthcare fraud and abuse. Expert Systems with Applications, 31, 56–58. Yang, H.-C., & Lee, C.-H. (2004). A text mining approach on automatic generation of web directories and hierarchies. Expert Systems with Applications, 27, 645–663. Yu, L., Yue, W., Wang, S., & Lai, K. K. (2010). Support vector machine based multiagent ensemble learning for credit risk evaluation. Expert System with Applications., 37(2), 1351–1360. C. León et al. / Expert Systems with Applications 38 (2011) 10274–10285 10285 
work_ahtiosddlbhgtj7wlpis3mjplm	----	doi:10.1016/j.eswa.2007.05.048 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 3021–3032 Expert Systems with Applications An approach to mining the multi-relational imbalanced database Chien-I Lee a, Cheng-Jung Tsai b,*, Tong-Qin Wu a, Wei-Pang Yang c a Department of Information and Learning Technology, National University of Tainan, Tainan, Taiwan, ROC b Department of Computer Science, National Chiao Tung University, Hsinchu, Taiwan, ROC c Department of Information Management, National Dong Hwa University, Hualien, Taiwan, ROC Abstract The class imbalance problem is an important issue in classification of Data mining. For example, in the applications of fraudulent telephone calls, telecommunications management, and rare diagnoses, users would be more interested in the minority than the majority. Although there are many proposed algorithms to solve the imbalanced problem, they are unsuitable to be directly applied on a multi- relational database. Nevertheless, many data nowadays such as financial transactions and medical anamneses are stored in a multi-rela- tional database rather than a single data sheet. On the other hand, the widely used multi-relational classification approaches, such as TILDE, FOIL and CrossMine, are insensitive to handle the imbalanced databases. In this paper, we propose a multi-relational g-mean decision tree algorithm to solve the imbalanced problem in a multi-relational database. As shown in our experiments, our approach can more accurately mine a multi-relational imbalanced database. � 2007 Elsevier Ltd. All rights reserved. Keywords: Data mining; Classification; Imbalance; Relational database 1. Introduction Most structural data nowadays, like financial transac- tions and medical anamneses, are stored in the multi-rela- tional database. A multi-relational database usually consists of numerous data sheets and each of them would contain a number of tuples. A data sheet is also called a relation and there are usually one target relation and several non-target relations in a multi-relational database. Each tuple in the target relation is composed of several attribute values and a target class, but a tuple in the non-target rela- tions contains only the attribute values. Therefore, a tuple in the target relation is also called target tuple. No matter the target or non-target relation, they all possess a primary key and some foreign keys. The foreign key usually con- 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.05.048 * Corresponding author. Tel.: +886 6 2133111x777; fax: +886 6 3017137. E-mail addresses: leeci@mail.nutn.edu.tw (C.-I. Lee), tsaicj@cis.nctu. edu.tw (C.-J. Tsai), wpyang@cis.nctu.edu.tw, wpyang@mail.ndhu.edu.tw (W.-P. Yang). nects to the primary or foreign key in other relations to make the database operations such as join or select feasible. There are mainly two kinds of connection between the rela- tions, which are (a) the connections between a primary key K and some foreign keys pointing to K; (b) the connections between two foreign keys K1 and K2 which connect to an identical primary key K. Unfortunately, the classification techniques of tradi- tional data mining, such as decision trees (Quinlan, 1993), neural networks (Mitchell, 1997) and support vector machines (Burges, 1998), are only applied for a singular data sheet. In order to extract out the practical and valu- able information from the relational databases, many multi-relational classification approaches had been pro- posed in recent years. The most widely used category of approaches to multi-relational classification is Inductive Logic Programming (Lavrac & Dzeroski, 1994) such as FOIL (Quinlan et al., 1993), Golem (Muggleton & Feng, 1990), Progol (Muggleton, 1995), TILDE (Blockeel, De Raedt, & Ramon, 1998), and CrossMine (Yin, Han, Yang, & Yu, 2004). However, the rules mined by these methods will be guided by the negative/major tuples. That is, mailto:leeci@mail.nutn.edu.tw mailto:tsaicj@cis. nctu.edu.tw mailto:tsaicj@cis. nctu.edu.tw mailto:wpyang@cis.nctu.edu.tw mailto:wpyang@mail.ndhu. edu.tw Fig. 1. A typical decision tree. 3022 C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 although it can generate a classifier with high accuracy, it is unable to provide the rules precisely for the positive/minor tuples. For some practical applications such as fraudulent telephone calls and rare diagnoses, users would be more interested in the minor tuples. Although there are also many proposed solutions for the imbalanced problem (Jap- kowicz & Stephen, 2002; Tsai, Lee, Chen, & Yang, 2007), they are unsuitable to be directly applied in a multi-rela- tional database. Thus, in this paper, we propose a multi- relational g-mean decision tree, called Mr.G-Tree, as the new solutions for an imbalanced multi-relational database. As it is named, Mr.G-Tree is a decision-tree-based algo- rithm. We focus on the decision-tree-based approaches since decision tree has several advantages (Rastogi & Shim, 2000), which are (a) decision tree is more efficient for large training data than neural networks which would spend a lots time on thousands of iterations; (b) a decision tree algorithm does not require the domain knowledge or prior knowledge; (c) decision tree displays the good classification accuracy compared to other techniques. For clear explana- tion, in the paper we will use ‘‘the positive’’ or ‘‘the minor- ity’’ to denote the minor but interesting tuples, and on the contrary ‘‘the negative’’ or ‘‘the majority’’ to represent the major but trivial target classes in a database. The remainder of this paper is organized as follows. Sec- tion 2 is the review of the related works. In Section 3, multi- relational g-mean decision tree will be introduced. The performance evaluation of will be shown in Section 4. Finally, the conclusion and future research directions will be demonstrated in Section 5. 2. Related work In this section, we first introduce the main principle for inducing a decision tree in Section 2.1. We then review some approaches which were proposed to mine a multi- relational database in Section 2.2. The imbalanced problem will be discussed in Section 2.3. 2.1. Decision tree A decision tree (Han & Kamber, 2001; Quinlan, 1993) is a flow-chart-like tree structure, which is constructed by a recursive divide-and-conquer algorithm. In a decision tree, each internal node denotes a test on an attributes, each branch represents an outcome of the test, and each leaf node has an associated target class. The top-most node in a tree is called root and each path from the root to a leaf node represent a rule. A typical decision tree is shown in Fig. 1. To classify an unknown example, beginning with the root node, successive internal nodes are visited until this example reaches a leaf node. The class of this leaf node is then assigned to this example as a prediction. For instance, the decision in Fig. 1 will approve a golden credit card application if the applicant has a salary higher than 85000 and his repayment record is good. 2.2. Multi-relational classifier In the current field of the multi-relational classification, the most common one is Inductive Logic Programming (ILP). The formal definition of the ILP problem is: ‘‘Given background knowledge B, positive tuples TP, and negative tuples TN, find a hypothesis H, which is a set of Horn clauses, such that "tp 2 TP : H [ Bj=tp and "tn 2 TN : H [ Bj5tn’’. The well known ILP systems are FOIL (Quinlan et al., 1993), Golem (Muggleton & Feng, 1990), Progol (Muggle- ton, 1995), TILDE (Blockeel et al., 1998), and CrossMine (Yin et al., 2004). First-order inductive learning (FOIL) takes CN2 algorithm (Clark & Niblett, 1989) as the basis and applies top-down and general-to-specific search to establish numerous rules. Each rule will include as many the positive as possible. On the contrary, Golem adopts bottom-up and specific-to-general search, it uses the tech- niques of RLGG (Relative Least General Generalization) to undertake generalization among a number of specific rules. Regarding with Progol, it is based on the AQ algo- rithm (Michalski, 1969) and integrates the searching meth- ods in FOIL and Golem. However, higher calculating cost will be caused when the database is enlarged as the com- mon disadvantage for traditional ILP approaches. In order to reduce the time cost, TOP-down induction of first order logical decision trees (TILDE) is proposed. TILDE is a binary logical decision tree and based on the C4.5 algorithm. Experiments also show that its accuracy is definitely better than the traditional ILP approaches (Boström, 1995). However, TILDE is inapplicable the large-scale databases. To solve the scalability problem, based on FOIL, CrossMine algorithm (Yin et al., 2004) is proposed. In order to reduce the requirement of memory, CrossMine virtually joins target relation and non-target relation together. It propagates the primary key and the target class in target relation to all non-target relations. Accordingly, there are two additional columns ‘‘IDs’’ and ‘‘class labels’’ in each non-target relation. By this propaga- Table 1 The confusion matrix Predicted negative Predicted positive True negative a b True positive c c C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 3023 tion approach, CrossMine can calculate the foil gain in each relation without joining all the relations into an indi- vidual table and therefore the memory cost can be well reduced. Owing to the well consideration of connections among all relations, its accuracy performs much better than FOIL. 2.3. Class imbalanced problem The proposed techniques aiming at the class imbalanced problem so far could be classified into three categories as follows (Barandela, Sanchez, Garcia, & Rangel, 2003): 2.3.1. Sampling-based Over-sampling and under-sampling are two main tech- niques in this category. Over-sampling could be further classified into random over-sampling and focused over-sam- pling (Aha, Kibler, & Albert, 1991; Han, Wang, & Mao, 2005). Random over-sampling approach over-samples the minority class at random until it matches the size of the majority class. Focused over-sampling approach over-sam- ples the minority class only with data close to the bound- aries between the minority class and the majority class. Similarly, under-sampling could be also classified into ran- dom under-sampling and focused under-sampling (Deh- meshki, Karakoy, & Casique, 2003; Derouin, Brown, Beck, Fausett, & Schneider, 1991; Lewis & and Catlett, 1997). The former approach removes the majority class at random until it contains as many examples as the minor- ity class, and the latter one removes the majority examples lying further away. The main idea of focused under-sam- pling is to remove the noise or outlier data and to reduce the size of majority class by sampling. The combination of over-sampling and under-sampling has also been pro- posed (Cohen, Hilario, Sax, Hugonnet, & Geissbühler, 2006; Zhou & Liu, 2006). However, over-sampling will increase the training set size and therefore enlarge the com- putational burden and the impact of noise data; under- sampling has been proven to be ineffective since it results in excluding some useful information (Barandela et al., 2003; Japkowicz & Stephen, 2002). 2.3.2. Cost-based Cost-Modifying approach (Pazzani et al., 1994; Zadrozny & Elkan, 2001) reduces the relative misclassifi- cation cost of the majority class (or increasing that of the minority class) to make it correspond to the size of the minority class. However, it is hard for a user to assign a proper cost when he/she is unfamiliar with the domain knowledge (Japkowicz & Stephen, 2002). 2.3.3. Imbalance-insensitive This approach is more attractive and has been proven to be more effective than the above two approaches (Jap- kowicz & Stephen, 2002). The main idea of this technique is to develop an approach that is insensitive to the imbal- ance problem. Proposed techniques including example weighting, rule removing, attribute correlation analysis, etc. (Anto, Susumu, & Akira, 2000; Ezawa, Singh, & Nor- ton, 1996; Fawcett & Provost, 1996; Kubat, Holte, & Mat- win, 1998; Lawrence, Burns, Back, Tsoi, & Giles, 1998). Among the proposed imbalance-insensitive approaches, some of them are limited to specific dataset and some take a lot of training time due to the natural property of neural network. Comparatively, SHRINK (Kubat et al., 1998) is applicable to most applications with numeric attribute data and eliminates the disadvantage described above. SHRINK was developed by the principle of BRUTE (Rid- dle, Segal, & Etzioni, 1994), it searches for the most accu- racy for not only the minority, but also the majority simultaneously by the use of g-mean to reach the best par- tition. However, its shortcoming is to only take care of numeric attribute and search for the best interval in a single one to each attribute. Once if the positive are distributed in two extremities of attribute values, SHRINK will have a poor accuracy. 3. Multi-relational g-mean decision tree algorithm Although there are many multi-relational database min- ing classifiers as described in Section 2.1, they can not solve the imbalanced problem. Furthermore, the proposed algo- rithms to solve the imbalanced dataset are unsuitable to be directly applied on a multi-relational database either. Thus, in this paper, we propose the Multi-relational g-mean deci- sion tree, called Mr.G-Tree, to solve this problem. Without loss of generality, only the case that there are two target classes in a database will be discussed in our paper. If more than two target classes are involved, the target class with the smallest ratio or the interesting one can be identified as the minority, and the rest will be regarded as the majority. 3.1. G-mean based classification model The confusion matrix in Table 1 was a widely used matrix to illustrate different classification measurements. In this matrix, a means the number of the negative that is accurately classified, b is the number that the negative is classified into the positive, c denotes the number that the positive is classified into the negative, and d signifies the number of the positive that is accurately classified. The traditional accuracy is defined as (a + d)/ (a + b + c + d). However, to handle the imbalanced prob- lem, g-mean in Eq. (1) was introduced to represent the 3024 C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 average proportion for the accuracy of the minority and the majority (Rastogi & Shim, 2000). g ¼fða � dÞ=½ða þ bÞ�ðc þ dÞ�g1=2: ð1Þ As stated in Section 2.3, the shortcomings of SHRINK are (a) it establishes only a best interval for each attribute; (b) it is inapplicable to the categorical attribute. These two properties would reduce the predicted accuracy when there are categorical attributes in the training data and when the minority distribute over several intervals of an attribute (Tsai et al., 2007). In order to build a classifier which can handle the imbalanced data more accurately, based on g-mean measurement, we establish a best interval for each numeric attribute and a best subset for each categorical attribute in every internal node o. Each best interval/best subset of the attribute i denotes a test for a tuple. To give higher importance to a test with smaller error, each test is associated with a weight as in Eq. (2) according to its g-mean value gi wi ¼ log½gi=ð1 � giÞ�: ð2Þ Then all of them are combined as a splitting function as shown in Function 3 for an internal node of Mr.G-Tree. SFðoÞ¼ X i hi � wi; ð3Þ In Eq. (3), hi = 1 if the value of attribute i falls in this inter- val and hi = �1 otherwise. However, to ensure that all weight is larger than 0, the test with gi < 0.5 is discarded. The best interval of a numeric attribute is defined as [1/2(aip + minai), 1/2(aiq + maxai)], where minai and maxai, respectively denotes the minimal and maximal attribute value of the interval with the maximal g-mean, aip is the attribute value of tuple p with the order prior to the tuple with attribute value minai, and aip is the attribute value of tuple q with the order posterior to the tuple with attribute value maxai. The reason is that when predicting the unseen data, the output hi of an example with attribute value lies in the [aip, minai] and [aiq, maxai] is unclear. To deal with cat- egorical attribute, we adopt Set Theory to establish the best subset. First, for each categorical attribute, we classify all tuples by their target classes and establish a corresponding power set. Each power set would contain 2k-2 subsets, where k is the number of values in this categorical attri- bute. For each subset, the tuples belonging to it are regard Procedure Leaf_Node(n) Begin For each leaf node If NP ≤(NN ×TP) / TN then The target class of this node is negati Else The target class of this node is positiv End if End Fig. 2. The pseudo code of the hand as the positive, and are regard as the negative otherwise. Finally, the g-mean of each subset is computed and the sub- set with the maximal g-mean is selected as the best subset. However, when a categorical attribute contains many val- ues, the computational cost for a categorical attribute by our approach would be nontrivial. To solve this problem, when the number of values in a categorical attribute is lar- ger than 10, we use a simpler approach. That is, for each value, the tuples belonging to it are regard the positive, and are regard as the negative otherwise. As a result, there are only k subsets when k is larger than 10. In addition, when a leaf node is not pure, most decision tree algorithms will assign this leaf node a negative class if there are more negative instances in this node. However, such a method may reduce the accuracy of the positive that the users are interested in. Thus, in order to mine the posi- tive more accurately, when assigning the target class to a leaf node, Mr.G-Tree adds in the consideration of the imbalanced problem as shown in Definition 2. Fig. 2 repre- sents the corresponding pseudo code. Definition 1. Given a training dataset T that includes both the minority and majority. For any leaf node Ni on Mr. G-Tree, assume that NP represents the number of the minority in Ni, NN represents the number of the majority in Ni, TP represents the number of all the majority in T, and TN represents the number of all the majority in T; if NP 6 (NN · TP)/TN, then the target class of leaf node Ni will be negative, and vice versa. Example 1. To make readers understand our approach mentioned in this section more clearly, here we take the sin- gle-relational database in Table 2 as an example. This data- base contains 20 tuples in which 2 tuples are the positive and 18 ones are the negative and the column BMI denotes ‘‘body mass index’’. The Mr.G-Tree trained from Table 2 is illustrated in Fig. 3. In the Node 1 of Fig. 3, the best inter- val of Attribute ‘‘age’’ is [48, 66] and its g-mean is [(13/ 18) · (2/2)]1/2 = 0.85, consequently the weight of this best interval is log [0.85/(1 � 0.85)] = 0.75. Similarly, the best interval, g-mean, and weight of Attribute ‘‘BMI’’ is [13, 18], 0.85, and 0.75, respectively. As for the subsets and the corresponding g-mean of Attribute ‘‘sport’’ and ‘‘smoking’’ in Node 1, we detail them in Table 3 in which S(i) denote a subset that contains the attribute-value i. Obviously, the best subset Attribute ‘‘sport’’ is [occasional, ve; e; ling of leaf nodes in Mr.G-Tree. Table 2 A training database Patient Age BMI Sport Smoking Diagnosis 1 41 12 Daily Yes Health 2 64 14 Occasional Yes Cancer 3 57 17 Occasional No Health 4 27 30 Daily Yes Health 5 31 25 Daily Yes Health 6 35 28 Daily Yes Health 7 37 15 Daily Yes Health 8 49 17 Never Yes Cancer 9 52 15 Never No Health 10 60 15 Occasional No Health 11 47 28 Daily Yes Health 12 23 26 Daily Yes Health 13 33 32 Never Yes Health 14 28 33 Occasional Yes Health 15 45 19 Never Yes Health 16 68 26 Occasional Yes Health 17 72 19 Never Yes Health 18 62 21 Daily Yes Health 19 56 16 Occasional No Health 20 24 20 Occasional Yes Health Fig. 3. The classification model constructed by using Table 2. C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 3025 never] and the g-mean and weight of this subset is [(8/ 18) · (2/2)]1/2 = 0.67 and log [0.67/ (1 � 0.67)] = 0.31, respectively. Note that since the g-mean of the best subset S(yes) in Attribute ‘‘smoking’’ is less than 0.5, it is not used in Node 1. Finally, we can get the splitting function SF(o) in each node o in Fig. 3 is Table 3 The power set and corresponding g-mean of Attribute (a) ‘‘sport’’ and (b) ‘‘smoking’’ in Node 1 of Fig. 3 (a) Subset S(d, o) S(o, n) S(d, n) S(d) S(o) S(n) G-mean ffiffiffiffiffiffiffiffiffiffiffi 4 18 � 1 2 q ffiffiffiffiffiffiffiffiffiffiffi 8 18 � 2 2 q ffiffiffiffiffiffiffiffiffiffiffi 6 18 � 1 2 q ffiffiffiffiffiffiffiffiffiffiffi 9 18 � 0 2 q ffiffiffiffiffiffiffiffiffiffiffi 12 18 � 1 2 q ffiffiffiffiffiffiffiffiffiffiffi 14 18 � 1 2 q (b) Subset S(yes) S(no) G-mean ffiffiffiffiffiffiffiffiffiffiffi 4 18 � 2 2 q ffiffiffiffiffiffiffiffiffiffiffi 10 18 � 0 2 q SFð1Þ: 0:75 � hage½48; 66�þ 0:75 � hBMI ½13; 18� þ 0:31 � hsport½occasional; never�; SFð2Þ: 0:66 � hsmoking½yes�: 3.2. Handling an imbalanced multi-relational database In order to mine the multi-relational database and avoid the shortcomings of waste memory by joining all data- sheets to a single one, Mr.G-Tree extended the concepts of propagation introduced in CrossMine to virtually link the target relation and non-target relation together. Yet, CrossMine will generate redundant target classes in each non-target relation, which will result in the incorrect g-mean calculated. For example, assuming there are two tuples t1 and t2 in target relation R1 and both of them link to the tuple a1 in non-target relation R2. If applying Cross- Mine, tuple a1 will possess two target classes simulta- neously, affecting the original distribution of target classes in non-target relation R2. If g-mean is calculated by such propagation outcomes, the unreal values for all attributes in R2 will be obtained owing to two target classes of a1 are considered. Take the attribute ‘‘smoking’’ in Table 4 as the example, there are only 9 tuples in non- target relation ‘‘Physical’’, but if taking the propagation method proposed in CrossMine, the number of tuples will be increased to 18 and then the original class distribution in non-target relation ‘‘Physical’’ will be changed. In order to solve this problem, Mr.G-Tree introduces the g-mean Tuple ID propagation algorithm, also known as GTIP algorithm. GTIP maintains the original data dis- tribution in each non-target relation by restoring the num- ber of target classes of each tuple to a single one as described in Definition 2. Definition 2. Given a non-target relation R0 and assume that each tuple ti in R 0 contain two additional columns ‘‘IDs’’ and ‘‘class labels’’ after the propagation. If the class labels of a tuple are all the positive, this tuple will be regarded as a positive tuple; if the class labels of a tuple are all the negative, it will be regarded as a negative tuple; if tuple ti contains both positive and negative class labels, this tuple would contains CP/(CP + CN) positive class and CN/(CP + CN) negative class, where CP and CN, respec- tively represents the numbers of the positive and that of the negative in the class labels column. By Definition 2, GTIP can prevent the original distribu- tion of target class in non-target relation from the distor- tion and the inaccurate g-mean. Fig. 4 is the pseudo code of GTIP algorithm. Here we again take Table 4c as an example. The result of GTIP is shown in Table 5. In addition, semantic links is always a major issue in the research field of Multi-relational database. The research shows that the relations with longer semantic links will become less important in the Multi-relational database (Liu, Yin, & Han, 2005). Thus, in order to design a classi- fier suitable for the multi-relational database, Mr.G-Tree Table 4 A relational database in which (a) the connection between two relations ‘‘Patient’’ and ‘‘Physical’’; (b) the details of relation ‘‘Patient’’; (c) the details of relation ‘‘Physical’’ (a) Patient Physical Patient-id Physical-id Condition-id ! Condition-id Age BMP Sport Smoking Class label (b) Patient (c) Physical Patient-id Condition-id Age Sport Class Physical-id Condition-id BMP Smoking 1 12 41 Daily Health 21 24 22 Yes 2 14 64 Occasional Cancer 22 14 16 Yes 3 17 57 Occasional Health 23 17 16 Yes 4 26 27 Daily Health 24 14 18 Yes 5 26 31 Daily Health 25 15 24 No 6 28 35 Daily Health 26 15 21 No 7 15 37 Daily Health 27 28 20 No 8 17 49 Never Cancer 28 28 25 No 9 15 52 Never Health 29 26 14 No 10 15 60 Occasional Health 11 24 47 Daily Health 12 24 23 Daily Health 13 24 33 Never Health 14 33 28 Occasional Health 15 19 45 Never Health 16 26 68 Occasional Health 17 19 72 Never Health 18 14 62 Daily Health 19 16 56 Occasional Health 20 20 24 Occasional Health 3026 C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 takes the semantic link into account and defines a non-tar- get relation as the effective relation or non-effective relation as shown in Definition 3. While constructing each node, Mr.G-Tree will only propagate the IDs and class labels to an effective relation and then calculate the best inter- vals/subsets of all attributes in the effective relation. In other words, the tuples used to build the node in an upper layer will be closer to the target relation, which can solve the problem of weaker semantic links. Definition 3. Assume that the length of semantic links between a non-target relation and the target relation is l, and the depth of Mr.G-Tree is d (the depth of root node is 0). While constructing a node in k-layer of Mr.G-Tree, a Procedure GTIP(R’) Begin For each non-effective relation R’ /* a non-eff For each primary key and foreign-key k If R’ can connect to a relation R th R’.k then Propagate IDs and class label End if For each tuple ti in relation R’ Assign tuple ti with CP / (CP+CN) posi class; Return the new target class to each tupl End Fig. 4. The pseudo code of g-mean non-target relation is defined as an effective relation if l 6 d, and is defined as a non-effective relation otherwise. 3.3. The detailed step of Mr.G-Tree Integrating the methods mentioned in Sections 3.1 and 3.2, we detail the steps of multi-relational g-mean decision tree as follows. Fig. 5 is the pseudo code of Mr.G-Tree. 1. Performing GTIP algorithm to all effective relation. 2. Calculating the best interval or best subset of each attribute in every effective relation and compute the g-mean. ective will be defined in Definition 3 in R’ at contain the column “class labels” via s from R to R’; tive class and CN / (CP+CN) negative e; tuple ID propagation algorithm. Procedure Traintree(D) Begin Initial Mr. G-Tree’s current depth d = 0; If D is pure or each weight wi = 0 in D then Return; Else If the Mr.G-Tree get deeper then d = d + Initial Left_D = φ , Right_D = φ ; End if For each non-effective relation R’ Call GTIP(R’); For each effective relation R For every attribute i If the attribute is numeric then Sort all examples according the v For each positive example j Calculate the interval; Calculate the g-mean gi of the i Select the best interval whose g If gi < 0.5 then discard this Elseif gi = 1 then wi = 1; Else wi = log[gi / (1-gi)] End if Else Building the corresponding power For each subset Classify all examples to the co Calculate the g-mean gi of this Select the best subset whose g If gi < 0.5 then discard this Elseif gi = 1 then wi = 1; Else wi = log[gi / (1-gi)] End if End if Classification function SF = hi ×wi; For each training example t∈ D If SF(t) ≥ 0 then Left_D = Left_D ∪t; Else Right_D = Right_D ∪t; End if Traintree (Left_D); Traintree (Right_D); Call Leaf_Node(D); End Fig. 5. The pseudo co Table 5 The propagation results of Table 4c by using GTIP Physical Physical-id Condition-id BMP Smoking IDs Class labels 21 24 22 Yes 11, 12, 13 c:0, h:1 22 14 16 Yes 2, 18 c:1/2, h:1/2 23 17 16 Yes 3, 8 c:1/2, h:1/2 24 14 18 Yes 2, 18 c:1/2, h:1/2 25 15 24 No 7, 9, 10 c:0, h:1 26 15 21 No 7, 9, 10 c:0, h:1 27 28 20 No 6 c:0, h:1 28 28 25 No 6 c:0, h:1 29 26 14 No 4, 5, 16 c:0, h:1 C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 3027 3. Discarding the attribute if its best g-mean is less than 0.5. 4. Calculating the weight of every attribute as wi= log [gi/(1 � gi)]. 5. Let hidenote the output function of attribute i. If an attribute value lies in the best interval then hi = 1; hi = �1 otherwise. 6. Combining the output function and weight of all the attributes to generate the splitting function in each node o as SF(oÞ¼ P ihi � wi. 7. A tuple is allotted to the left child node if the SF is greater than 0; otherwise, to the right one. 8. When other relations are able to link to an effective relation through primary key and foreign key, the data in the column of IDs and class labels will be 1; alue of this attribute; nterval [1/2(aip + minai), 1/2(aiq + maxai); i is the maximal; attribute; set; rresponding subset by their class; subset; i is the maximal; attribute; de of Mr.G-Tree. Fig. 6. The Mr.G-Tree constructed by using the relational database in Table 4. Table 6 The details of two real multi-relational databases Id Database Financial (1) Mutagenesis (2) # Non-target relation 7 3 # Attribute in target relation 4 4 # Positive tuples in target relation 76 64 # Total tuples in target relation 400 2 # Tuples in all relations 75,982 15,218 3028 C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 propagated to this relation. After that, setting the relation as the effective one. 9. Repeating the steps from 1 to 8 until all the nodes are pure or can not be further spilt. 10. Assigning the target class to each leaf node according to Definition 2. Table 7 The parameters of synthetic multi-relational database generator Name Description Parameter jRj Number of relation x T Number of tuples in each relation y Imbalance The percentage of positive tuples in target relation z% Tmin Minimal number of tuples in each relation 50 Amin Minimal number of attributes in each relation 2 Amax Maximal number of attributes in each relation 5 Vmin Minimal value in each attribute 2 Vmax Maximal value in each relation 1000 Fmin Minimal number of foreign keys in each relation 2 Fmax Maximal number of foreign keys in each relation 5 3.4. An example of Mr.G-Tree Here, we clearly introduce the process of Mr.G-Tree by using Table 4 again. In the beginning, Mr.G-Tree calcu- lates the best interval/subset and the corresponding g-mean of Attribute ‘‘age’’ and ‘‘sport’’ in the effective relation ‘‘Patient’’. As described in Example 1, the best interval of the numeric attribute age sport is [48, 66] and the corre- sponding g-mean is 0.85; as for the categorical attribute ‘‘sport’’, its best subset is [occasional, never] and 0.67 for g-mean. The weight for those two attributes is respectively wage = 0.75 and wsport = 0.31. Hence, the splitting function in the root node of Mr.G-Tree is SFð1Þ¼ 0:75�hage½48;66�þ0:31�hsport½occasional; never�: Since the non-target relation ‘‘Physical’’ can link to effec- tive relation ‘‘Patient’’ through foreign key ‘‘condition- id’’, Mr.G-Tree will propagate the data in the column ‘‘IDs’’ and ‘‘class labels’’ to the relation ‘‘Physical’’ in the next loop. The propagation results have been shown in Table 5. In the meanwhile, the relation ‘‘Physical’’ will be set as effective one. Next, Mr.G-Tree calculates the g-mean of Attribute ‘‘age’’, ‘‘sport’’, ‘‘BMP’’, and ‘‘smoking’’ from all effective relation. The result comes that the g-mean of Attribute ‘‘age’’ and ‘‘sport’’ are both less than 0.5, so the two attributes are discarded. However, the best interval of the numeric attribute ‘‘BMP’’ is [15, 19] and the corre- sponding g-mean is 0.84; as for the categorical attribute ‘‘smoking’’, its best subset is [yes] and 0.77 for g-mean. The weight for those two attributes is wsmoking = 0.52 and wBMP = 0.72, respectively. Hence, the splitting function in Node 2 of Mr.G-Tree is SFð2Þ¼ 0:52 � hsmoking½yes�þ 0:72 � hBMP½15; 19�: Finally, taking use of the rules described in Definition 2 to assign the target class to all leaf nodes, we can get the Mr.G-Tree as illustrated in Fig. 6. 4. Experimental analyses In this section, the software and hardware environment and the experimental databases for our experiments will be shown in Section 4.1. The measurement for evaluation of an imbalanced database will be introduced in Section 4.2. Finally, the comparisons among TILDE, CrossMine and Mr.G-Tree are illustrated in Sections 4.3 and Section 4.4. 4.1. The experimental environment and databases We implemented our Mr.G-Tree and two state-of-the- art multi-relational classifiers TILDE and CrossMine in Microsoft Visual C++ 6.0 for performance analysis. All experiments in this paper are implemented under the Win- dows 2000 professional on a PC equipped with Intel Pen- tium IV 3.0 GHz CPU and 512 MB DDR memory. Our experimental databases can be divided into 2 parts as follows. 4.1.1. Real multi-relational databases Two real multi-relational databases: Financial database and Mutagenesis database which had been used in Cross- Mine (Yin et al., 2004) are again applied. Financial data- base includes 1 target relation, 7 non-target relations and 75982 tuples; Mutagenesis database contains 1 target rela- 90 100 TILDE C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 3029 tion, 3 non-target relations and 15,218 tuples. The details of the two databases are listed in Table 6. 0 10 20 30 40 50 60 70 80 A cc ur ac y of t he p os it iv e (% ) CrossMine Mr.G-Tree 4.1.2. Synthetic multi-relational databases To have further experimental analysis, we code a data generator to generate proper synthetic multi-relational dat- abases (Quinlan et al., 1993). Each synthetic multi-rela- tional database contains one target relation and several non-target relations. Table 7 is the related parameter of our data generator. Five parameters: Tmin, Amin, Amax, Vmin and Fmin, will be fixed by referencing the settings applied in CrossMine. 2 Databases 1 Fig. 7. The comparison of the accuracy of the positive in two real relational databases. 4.2. The measurement Since the accuracy which is widely used to evaluate a classifier is not a proper metric (Joshi, 2002); in this paper, we use the accuracy of the minority as shown in Eq. (4), where c and d have been described in Table 1, to evaluate the accuracy of a classifier on the minority. accuracyþ ¼ d=ðc þ dÞ; ð4Þ Furthermore, ZR and ZP (Joshi, 2002) in Eqs. (5) and (6) are utilized to evaluate the overall accuracy between Clas- sifier A and B. In Eqs. (5) and (6), RA, RB, PA, and PB, respectively represents the recall of Classifier A, the recall of Classifier B, the precision of A and that of B; nc denotes the total number of the positive, nA0 is the number of the po- sitive predicted by A, nB0 is the number of the positive pre- dicted of B, R = (RA + RB)/2, and P = [(n A 0 � P AÞ + (nB0 � P BÞ]/(nA0 þ nB0Þ. Z R ¼ RA � RBffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2Rð1 � RÞ=nc p ð5Þ Z P ¼ P A � P Bffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pð1 � PÞð1=nA0 þ 1=nB0Þ p : ð6Þ If ZR P 1.96, the overall accuracy of RA is better than RB(denoted by RA � RB). If jZRj < �1.96, the overall accuracy of RB is better than RA(denoted by RA � RB). If jZRj<1.96, there is no significant difference between Clas- sifier A and B and will be denoted by RA � RB. The way to use the ZP to evaluate two classifiers is similar. By using ZR and ZP, we can compare two classifiers in the aspect of overall accuracy as follows. Table 8 The comparison of overall accuracy in two real relational databases Database-id Classifier ZR 1 Mr.G vs. TILDE 2.348 Mr.G vs. CrossMine 2.159 2 Mr.G vs. TILDE 2.07 Mr.G vs. CrossMine 1.054 (a) If (RA � RB and PA � PB) or (RA � RB and PA � PB) or (RA � RB and PA � PB), then the over- all accuracy of Classifier A is better than the Classifier B (A > B). (b) The method of judgment for (B > A) is identical to the one mentioned above. (c) If (RA � RB and PA � PB), that will be A � B. (d) It may happen that one metric is significantly better but the other is significantly worse. If (RA � RB and PA � PB) or (RA � RB and PA � PB), F-mea- sure as shown in Eq. (7) would be utilized as the way for the accuracy judgment. However, since F-measure does not have any probabilistic interpre- tation, we cannot apply any significance test to its value. So we use a heuristic such that the improve- ment in F-measure by at least 1% is required to call a classifier is better. F -measure ¼ð2 � P A � RAÞ=ðP A � RAÞ ð7Þ 4.3. The comparison among TILDE, CrossMine and Mr.G-Tree on real multi-relational databases In this section, we apply two real multi-relational dat- abases to compare the accuracy of the positive and the overall accuracy among TILDE, CrossMine and Mr. G-Tree. Fig. 7 is the comparison of the accuracy of the positive in two real relational databases and Table 8 is the comparison of overall accuracy. In Table 8, the com- parison is marked in bold if the Condition d mentioned ZP F-measure Overall accuracy �0.722 40 vs. 31.69 Mr.G > TILDE �2.149 40 vs. 41.32 Mr.G � CrossMine �1.181 71.43 vs. 57.14 Mr.G > TILDE �0.731 71.43 vs. 69.72 Mr.G � CrossMine 0 10 20 30 40 50 60 70 80 90 100 10 20 50 100 Number of relations A cc ur ac y of t he p os it iv e (% ) TILDE CrossMine Mr.G-Tree Fig. 8. The comparison of the accuracy of the positive on different number of relations. 0 10 20 30 40 50 60 70 80 200 500 1000 2000 5000 Number of tuples A cc ur ac y of t he p os it iv e (% ) TILDE CrossMine Mr.G-Tree Fig. 9. The comparison of the accuracy of the positive on different number of tuples. 3030 C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 in Section 4.2 is met and consequently the F-measure is used to compare the classifiers. All experiments related to the overall accuracy in the following will be presented in the same way. From Fig. 7, it is clear that Mr.G-Tree per- forms much better than TILDE and CrossMine in predict- ing the positive. From Table 8, it can be noted that the overall accuracy of Mr.G-Tree is greater than TILDE and is comparable to CrossMine. 4.4. The comparison among TILDE, CrossMine and Mr.G-Tree on synthetic multi-relational database In Section 4.3, the number of relations in two real dat- abases is less than 10. However, in a real case, there might be a lot of relations in a multi-relational database. In order to evaluate if the accuracy of each classifier will be influ- enced when the number of relations is increased, we fix the variable y and z in Table 7, respectively as 500 and 10% to generate 4 synthetic multi-relational databases. Fig. 8 shows the comparison for the accuracy of the posi- tive; Table 9 is the comparison of overall accuracy. Note that, since TILDE combines all relations into a single one, when the number of relations is large, it will cause the shortage of memory space and then be infeasible. Such a condition is occurred in our experiments when the Table 9 The comparison of overall accuracy on different number of relations Number of relations Classifier ZR 10 Mr.G vs. TILDE 3.141 Mr.G vs. CrossMine 1.714 20 Mr.G vs. TILDE 3.262 Mr.G vs. CrossMine 2.824 50 Mr.G vs. TILDE 3.079 Mr.G vs. CrossMine 2.67 100 Mr.G vs. TILDE – Mr.G vs. CrossMine 3.231 number of relation is over 50 and we mark this condition by ‘‘–’’ in Table 9 since ZR and ZP are both incomputable. All experiments related to the comparison of overall accu- racy between Mr.G-Tree and TILDE in the following will be presented in the same way. From Fig. 8, it is found that the accuracy of the positive in Mr.G-Tree is always higher than TILDE and CrossMine. In Table 9, it is worth to note that as the number of relations is larger than 10, the overall accuracy of Mr.G-Tree is always better than TILDE and CrossMine. Next, we turn to fix the variable x and z, respectively as 20 and 10% to generate 5 databases for the analysis of the accuracy in each classifier when the tuples are getting increased. The results are illustrated in Fig. 9 and Table 10. As from Fig. 9, it is found that the accuracy of the posi- tive in Mr.G-Tree is always higher than TILDE and Cross- Mine no matter how many tuples there are. From Table 10, it is shown that among the overall accuracy of TILDE, CrossMine and Mr.G-Tree, none is always better than the other two. In other words, Mr.G-Tree displays a com- parable overall accuracy to TILDE, CrossMine. Similarly, when the tuples in every relation are over 1000, we are unable to show the accuracy of TILDE in Fig. 9 and Table 10. ZP F-measure Overall accuracy �7.487 23.48 vs. 25.34 Mr.G � TILDE �2.853 23.48 vs.26.11 Mr.G < CrossMine 0.328 24.13 vs. 13.47 Mr.G > TILDE 0.754 24.13 vs. 14.12 Mr.G > CrossMine �0.708 18.6 vs. 16.49 Mr.G > TILDE 1.285 18.6 vs. 8.43 Mr.G > CrossMine – 36.78 vs. – Mr.G > TILDE 1.22 36.78 vs. 20.74 Mr.G > CrossMine Table 10 The comparison of overall accuracy on different number of tuples Number of tuples Classifier ZR ZP F-measure Overall accuracy 200 Mr.G vs. TILDE 1.803 �0.663 20.3 vs. 14.99 Mr.G � TILDE Mr.G vs. CrossMine 1.035 �0.152 20.3 vs.17.74 Mr.G � CrossMine 500 Mr.G vs. TILDE 2.881 �0.167 28.78 vs. 16.82 Mr.G > TILDE Mr.G vs. CrossMine 2.582 0.091 28.78 vs. 21.23 Mr.G > CrossMine 1000 Mr.G vs. TILDE 6.548 �2.105 20.83 vs. 21.53 Mr.G � TILDE Mr.G vs. CrossMine 5.521 �2.861 20.83 vs. 22.66 Mr.G � CrossMine 2000 Mr.G vs. TILDE – – 24.52 vs. – Mr.G > TILDE Mr.G vs. CrossMine 3.523 �4.246 24.52 vs. 26.2 Mr.G � CrossMine 5000 Mr.G vs. TILDE – – 21.19 vs. – Mr.G > TILDE Mr.G vs. CrossMine 6.065 �11.77 21.49 vs. 32.75 Mr.G > CrossMine 0 20 40 60 80 100 1 10 20 40 Imbalance level (%) A cc ur ac y of t he p os it iv e (% ) TILDE CrossMine Mr.G-Tree 5 Fig. 10. The comparison of the accuracy of the positive on different imbalanced levels. C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 3031 Ultimately, for the evaluation of the imbalanced prob- lem, we fix variable x and y in Table 8, respectively as 20 and 500 to produce 5 databases. The corresponding imbal- anced level of 5 databases is 1%, 5%, 10%, 20% and 40%. As illustrated in Fig. 10, the accuracy of the positive in Mr.G-Tree is better than TILDE and CrossMine apart from 40% as its imbalanced level. This result is not sur- Table 11 The comparison of overall accuracy on different imbalanced levels Imbalance (%) Classifier ZR 1 Mr.G vs. TILDE 0.982 Mr.G vs. CrossMine 0.982 5 Mr.G vs. TILDE 2.45 Mr.G vs. CrossMine 2.139 10 Mr.G vs. TILDE 3.601 Mr.G vs. CrossMine 3.331 20 Mr.G vs. TILDE 2.9 Mr.G vs. CrossMine 2.362 40 Mr.G vs. TILDE �7.681 Mr.G vs. CrossMine �7.66 prised and demonstrates the ability of Mr.G-Tree to mine an imbalanced multi-relational database. Table 11 shows that among the overall accuracy of TILDE, Cross- Mine and Mr.G-Tree, none is always superior to the other two. 5. Conclusion and future research directions Currently, the structural data are all stored in the rela- tional database, so the techniques of traditional data min- ing are not workable any more. Although there have been a number of multi-relational data mining classifiers pro- posed, such as TILDE, FOIL, CrossMine and so on, those approaches are unable to well handle the imbalanced prob- lem. Hence, Mr.G-Tree algorithm is proposed in this paper as the solutions. In order to build a classifier which can handle the imbalanced problem more accurately, Mr. G-Tree hierarchically pick up the positive by using g-mean mentioned in Section 3.1. Then, Mr.G-Tree makes the min- ing in a multi-relational database feasible by using GTIP algorithm proposed in Section 3.2 to propagate necessary information to all non-target relations. More importantly, GTIP algorithm keeps the g-mean reasonable after the ZP F-measure Overall accuracy �0.843 10.44 vs. 8.69 Mr.G > TILDE �0.713 10.44 vs.8.13 Mr.G > CrossMine �4.358 23.01 vs. 27.14 Mr.G < TILDE �2.874 23.01 vs. 24.42 Mr.G � CrossMine �6.073 37.66 vs. 42.15 Mr.G < TILDE �4.135 37.66 vs. 39.97 Mr.G � CrossMine �6.288 46.15.vs. 53.35 Mr.G < TILDE �3.835 46.15 vs. 49.34 Mr.G � CrossMine 0.122 62.14 vs. 85.29 Mr.G < TILDE 1.636 62.14 vs. 81.68 Mr.G < CrossMine 3032 C.-I. Lee et al. / Expert Systems with Applications 34 (2008) 3021–3032 propagation. Finally, by taking the semantic link into account, Mr.G-Tree can accurately and efficiently mine the imbalanced multi-relational databases. In the aspect of experimental analysis, since the accuracy which was widely used to evaluate a classifier is not a proper metric for the imbalanced datasets; in this paper, not only the accuracy of the positive is applied to evaluate the predic- tion ability of a classifier on the minority, but also ZR and ZP and F-measure is utilized to evaluate the overall accuracy. The results shown in Section 4 demonstrate that Mr.G-Tree can reach better classification accuracy of the positive than TILDE and CrossMine. Also, it shows com- parable consequence on the overall accuracy. However, Mr.G-Tree currently is only suitable for the database containing two target classes. Although Mr.G-Tree can set the most minor target class as the posi- tive and the rest as the negative under the condition of mul- tiple target classes, this way will change the original data distribution and consequently influences the mining results. Therefore, we will expand Mr.G-Tree in future to make it can accurately manage the databases contain multiple tar- get classes. Another line of future work is incremental Mr.G-Tree that considers the additions, deletions, and updates of tuples. References Aha, D., Kibler, D., & Albert, M. K. (1991). Instance-based learning algorithms. Machine Learning, 6(1), 37–66. Anto, S. N., Susumu, K., & Akira, I. (2000). Fog forecasting using self growing neural network CombNET-II – A solution for imbalanced training sets problem. In Proceedings of the international joint conference on neural networks (pp. 429–434). Italy. Barandela, R., Sanchez, J. S., Garcia, V., & Rangel, E. (2003). Strategies for learning in class imbalance problems. Pattern Recognition, 36(3), 849–851. Blockeel, H., De Raedt, L., & Ramon, J. (1998). Top-down induction of logical decision trees. Artificial Intelligence, 101, 285–297. Boström, H. (1995). Covering vs. divide-and-conquer for top-down induction of logic programs. In Proceedings of the fourteenth interna- tional joint conference on artificial intelligence (pp. 1194–1200). San Mateo. Burges, C. J. C. (1998). A tutorial on support vector machines for pattern recognition. Data Mining and Knowledge Discovery, 2, 121–168. Clark, P., & Niblett, T. (1989). The CN2 induction algorithm. Machine Learning, 3(4), 261–283. Cohen, G., Hilario, M., Sax, H., Hugonnet, S., & Geissbühler, A. (2006). Learning from imbalanced data in surveillance of nosocomial infec- tion. Artificial Intelligence in Medicine, 37(1), 7–18. Dehmeshki, J., Karakoy, M., & Casique, M. V. (2003). A rule-based scheme for filtering examples from majority class in an imbalanced training set. In Proceedings of the 13th international conference on machine learning and data mining in pattern recognition (pp. 215–223). Germany. Derouin, E., Brown, J., Beck, H., Fausett, L., & Schneider, M. (1991). Neural network training on unequally represented classes. Intelligent Engineering Systems Through Artificial Neural Networks, 135–145. Ezawa, K. J., Singh, M., & Norton, S. W. (1996). Learning goal oriented bayesian networks for telecommunications management. In Proceed- ings of the 13th international conference on machine learning (pp. 139– 147). San Francisco. Fawcett, T., & Provost, F. (1996). Combining data mining and machine learning for effective user profiling. In Proceedings of 12th international conference on knowledge discovery and data mining (pp. 8–13). Han, J., & Kamber, M. (2001). Data mining: Concepts and techniques. Morgan Kaufman. Han, H., Wang, W., & Mao, B. H. (2005). Borderline-SMOTE: A new over-sampling method in imbalanced data sets learning. In Proceedings of international conference on intelligent computing (pp. 878–887). Hefei, China. Japkowicz, N., & Stephen, S. (2002). The class imbalance problem: A systematic study. Intelligent Data Analysis Journal, 6(5), 429–450. Joshi, M. V. (2002). On evaluating performance of classifiers for rare classes. In Proceedings of 12th IEEE international conference on data mining. Japan. Kubat, M., Holte, R., & Matwin, S. (1998). Machine learning for the detection of oil spills in satellite radar images. Machine Learning, 30(2/ 3), 195–215. Lavrac, N., & Dzeroski, S. (1994). Inductive logic programming: techniques and applications. New York: Ellis Horwood. Lawrence, S., Burns, I., Back, A. D, Tsoi, A. C., & Giles, C. L. (1998). Neural network classification and prior class probabilities. Neural Networks: Tricks of the Trade 1998, Lecture Notes in Computer Science (pp. 299–314). Lewis, D. D., & Catlett, J. (1997). Heterogeneous uncertainty sampling for supervised learning. In Proceedings of the 11th international conference on machine learning (pp. 148–156). San Francisco. Liu, H., Yin, X., & Han, J. (2005). An efficient multi-relational naı̈ve bayesian classifier based on semantic relationship graphs. In Proceed- ings of ACM-SIGKDD workshop on multi-relational data mining. Chicago. Michalski, R. S. (1969). On the quasi-minimal solution of the general covering problem. In Proceedings of the 5th international symposium on information processing (pp. 125–128). Yugoslavia. Mitchell, T. M. (1997). Machine learning. McGraw Hill. Muggleton, S., & Feng, C. (1990). Efficient induction of logic programs. In Proceedings of the 1st conference on algorithmic learning theory (pp. 368–381). Tokyo. Muggleton, S. (1995). Inverse entailment and progol. In New Generation Computing. Special issue on Inductive Logic Programming, 13, 245–286. Pazzani, M., Merz, C., Murphy, P., Ali, K., Hume, T., & Brunk, C. (1994). Reducing misclassification costs. Proceedings of the 11th international conference on machine learning (pp. 217–225). Quinlan, J. R., & Cameron-Jones, R. M. (1993). FOIL: A midterm report. In Proceedings of the European conference on machine learning (pp. 3– 20). Vienna, Austria. Quinlan, J. R. (1993). C4.5: Programs for machine learning. San Mateo, CA: Morgann Kaufmann. Rastogi, R., & Shim, K. (2000). PUBLIC: A decision tree classifier that integrates building and pruning. Data Mining and Knowledge Discov- ery, 4(4), 315–344. Riddle, P., Segal, R., & Etzioni, O. (1994). Representation design and brute force induction in a boeing manufacturing domain. Applied Artificial Intelligence, 8, 125–147. Tsai, C. J., Lee, C. I., Chen, C. T., & Yang, W. P. (2007). A multivariate decision tree algorithm to mine imbalanced data. WSEAS Transactions on Information Science and applications, 4(1), 50–58. Yin, X., Han, J., Yang, J., & Yu, P. S. (2004). Crossmine: Efficient classification across multiple database relations. In Proceedings of the 20th international conference on data engineering (pp. 399–411). Boston, MA. Zadrozny, B., & Elkan, C. (2001). Learning and making decisions when costs and probabilities are both unknown. In Proceedings of the 7th international conference on knowledge discovery and data mining (pp. 204–213). San Francisco. Zhou, Z. H., & Liu, X. Y. (2006). Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Trans- action on Knowledge and Data Engineering, 18(1), 63–77. An approach to mining the multi-relational imbalanced database Introduction Related work Decision tree Multi-relational classifier Class imbalanced problem Sampling-based Cost-based Imbalance-insensitive Multi-relational g-mean decision tree algorithm G-mean based classification model Handling an imbalanced multi-relational database The detailed step of Mr.G-Tree An example of Mr.G-Tree Experimental analyses The experimental environment and databases Real multi-relational databases Synthetic multi-relational databases The measurement The comparison among TILDE, CrossMine and Mr.G-Tree on real multi-relational databases The comparison among TILDE, CrossMine and Mr.G-Tree on synthetic multi-relational database Conclusion and future research directions References 
work_aogmuki5z5cojix6nrwqgrvsk4	----	ESTJ: An Expert System for Tourism in Jordan Procedia Computer Science 65 ( 2015 ) 821 – 826 Available online at www.sciencedirect.com 1877-0509 © 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of Universal Society for Applied Research doi: 10.1016/j.procs.2015.09.032 ScienceDirect International Conference on Communication, Management and Information Technology (ICCMIT 2015) ESTJ: An Expert System for Tourism in Jordan Nada S. Husseina, Musbah J. Aqelb aDepartment of Computer Science, Faculty of Informatiom Technology, Applied Sceince Private University, Amman, Jordan bDepartment of Computer Science, Faculty of Information Technology, Applied Sceince Private University, Amman, Jordan Abstract An expert system for tourism in Jordan (ESTJ) was developed to recommend a suitable travel schedule that satisfies the tourist's interest. The system is useful for tourists, and tourism agencies to select the best package based on the proper time, budget, and preferences of required tourist places. The system was designed as a rule based expert system and implemented with a prolog language. The system was evaluated and tested with a specialist and the results concluded were up to the level of human expert. © 2015 The Authors. Published by Elsevier B.V. Peer-review under responsibility of Universal Society for Applied Research. Keywords: expert system, artificial intelligent, Tourism. 1.1. Introduction Recently many organizations have shown a great interest in using information technology for smoothing their business and to compete with competitors [1]. One of the major of information technology areas that finds a lot of interest is artificial intelligence. Expert systems are one of artificial intelligence fields that use a specialized knowledge to achieve high performance decision in a particular area [2]. Expert systems found many applications in many fields, like, science, business, and medicine [3]. Tourism is the act of travel for predominantly recreational or leisure purpose and also refers to the provision of services in support of this act. However, it has also importance in terms of economic (tourism revenue), and governments are trying to attract tourists to the country by providing an appropriate atmosphere for tourists and tourism programs that attract people. The expert systems play a vital role in tourism industry. Many expert systems were developed for supporting this area. For example, an expert system for tourist information © 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of Universal Society for Applied Research http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.09.032&domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.09.032&domain=pdf 822 Nada S. Hussein and Musbah J. Aqel / Procedia Computer Science 65 ( 2015 ) 821 – 826 management was developed to recommend a suitable travel schedule that satisfies user input constraints such as time period, budget and preferences. The system can provide tourists with information on the route and the distance between any two towns in the region [4].While Moisuc Diana-Aderina, et. al., designed an expert system for rural tourism in Maramures. The system was developed to evaluate the countryside hotels and ranking them. Moreover, the user can evaluate the benefits of tourist destinations from several points of view and obtain information useful for decision making [5]. However, Yunus Dogan, et. al., developed an expert system to support tourism sector in Turkey, where tourists will be able to select the most suitable holiday places for themselves [6]. However, Sindhu B., et. al., designed an expert system for rating the ecotourism destination [7]. In this research, an expert system will be developed to assist tourists and tourist agencies in Jordan to select the best package according to the given budget, and preferred tourist places. Khakzad developed an expert system that containing more than 20 cities and 150 rules for helping tourists to choose best destination town that has maximum matching with their important request [8]. Tourism industry at Jordan plays significant role in its economy. Since human expertise plays an important part in the activities of many sectors of the tourist industry. An expert system is developed to assist The system is useful for tourists, and tourism agencies to select the best package based on the proper time, budget, and preferences of required tourist places. 1.2. Tourism in Jordan Jordan attracted travellers since ancient times. Today it is more charming and beautiful with its modern state, due to progress and prosperity. It allows visitors to enjoy their taste differently as they wanted. It is the lives of simplicity and unspoiled nature calm and charming area. Jordan is one of the most important tourism countries in the world in general and in the Middle East in particular because of its features and environment which makes it a country with a high tourist attraction [9]. However, the tourism in Jordan can be classified into the following categories: 1- Historical 2- Religion 3- Beaches and water sports 4- Reserves 5- Entertainment 6- Shopping 7- Medical Treatment. 1.3. Expert System Design for Tourism in Jordan The proposed system, i.e., Expert System for Tourism at Jordan (ESTJ), is a rule based expert system which is implemented using Prolog program. The system makes use of backward chaining for the inference engine and search in the knowledge base. The system has a graphical user interface where the user can interact with the system through this interface. The user is presented with a series of questionnaire in the window which the user has the option to answer in yes or no. The set of questions are prepared according to the type of tourist's trip to the required destination. Now, according to the feedback given by the user, the search in the knowledge base for possible pattern matches. If there is a rule in the knowledge base which matches the given facts by the user, the system shows the possible diagnosis in the same window. The expert system was designed as a rule based expert system architecture and it is consisted from three parts as shown in figure (1). These parts are as follows: 1. Knowledge base 2. Inference engine 3. User interface 823 Nada S. Hussein and Musbah J. Aqel / Procedia Computer Science 65 ( 2015 ) 821 – 826 Fig.1: Expert System Architecture 3.1 KNOWLEDGE ACQUISITION The knowledge base can be considered as the heart of the Expert system as all the required facts for building the rules are contained in the knowledge base. Taking this knowledge as the source, rules for the Expert System can be formed. The primary source for knowledge acquisition for the ESTJ System was consultation with tourism's experts, internet, journals and tourism books. 3.2 KNOWLEDGE REPRESENTATION The knowledge base contains facts and rules about tourism in Jordan that acquired from specialists, tourism books, and journals issued from ministry of tourism. The knowledge in the knowledge base is represented as a set of rules. Each rule specifies a relation, recommendation and has IF part (condition) and THEN part (conclusion). Whenever the condition part is true, then the conclusion part will be executed. These rules is then represented as a predicate logic and converted to a program as follows: Rule representation IF The tourist likes to have a short trip (one day) AND likes to visit historical places AND likes to visit Romans civilization places THEN The recommended package starting from Amman by bus Jerash: An ancient city built by Romans Ajloun Castel a nice historical place built by Arabs The inference engine performs the reasoning process while the expert system finds a solution. It connects the rules given in the knowledge base with the facts provided in the knowledge base. The inference process is carried out in an interactive mode with the user providing input queries and responses to questions through the user interfaces that stored as dynamic information in the working memory. It tries to derive new information about the given situation using the rules in the knowledge and the situation specific knowledge in the working memory. 1.4. Expert System results and discussion The system provides a friendly user interface where the system starts its session by inquiring about the tourist programs by prompting the following questions during the consultation session: 1. Trip duration (shot trip-one day) or multiple days. 2. Trip type (e.g. historical, scuba diving, shopping, religion, etc.) 3. Trip budget: minimum price, and maximum price for the total trip expenses User Knowledge Base Inference Engine Facts Expertise 824 Nada S. Hussein and Musbah J. Aqel / Procedia Computer Science 65 ( 2015 ) 821 – 826 4. Trip destination: tourist sites that will be visited. 5. Trip services: in case of multiple days trip (rank of hotel, single or double room, etc.) Assume a tourist approached a tourist agency for a trip from Amman–capital of Jordan to a given destination like Petra which is a historical site. There are two packages for this site. The short trip is one day trip directly to Petra, while the other one is three days trip which includes visiting Petra, Wadi Rum that located at the heart of desert where the tourists enjoy camping overnight, and Aqaba city that has a beach and provides scuba diving entertainment. Types of tourist’s agent will consult the expert system for the best package for this trip with minimum price. The system starts asking the following: Do you plan to have one day trip, or long trip? (Select: 1. One day trip 2. Multiple days Select: 2 Does the place you want to visit have historical places (Y/N)? Y Which ancient civilizations do you like to see at Jordan? (Select: 1. Romans 2. Nabataean 3. Arabian) 2 Do you like camping at desert (Y/N) ? Y Do you like to visit religious places (Y/N)? N Do you like to visit places with beaches and that provide water sports (Y/N) Y Do you like to visit places with shopping mall facilities and provides personal entertainment (Y/N)? Y Do you like to stay in a hotel with one of the following ranks ? (Select: 1. Five star hotel 2. Four star hotel 3. Three star hotel) 1 Do you like to stay in a single or double room with your partner? (Select: 1. Single 2. Double) 1 The expert system accordingly will conclude the best package for the tourist with minimum cost as shown in figure (2). The best choice for you is the following route: Starting from Amman by bus: 1. visiting Petra for one night 2. Camping one night at Wadi Rum 3. Visiting Aqaba city for one day 4. The cost of the trip is ($500) which includes the following: 4.1 Transportation 4.2 Entry fees to the historical place – Petra 4.3 Camping one night at Wadi Rum 4.4 Staying at five stars hotel with a breakfast meal. 825 Nada S. Hussein and Musbah J. Aqel / Procedia Computer Science 65 ( 2015 ) 821 – 826 Fig.2: Consultation Session with Expert System Conclusion An expert System for Tourism in Jordan (ESTJ) was developed using rule based expert system architecture. The system was designed to assist the tourist, and the tourist agencies to select the best package for the tourist according to the tourist's preferences. The system was evaluated by specialist and the results were very satisfying and up to the level of human expertise. References: [1] Razieh, AC., Zahra, AC., "A Review on Expert Systems and Their Usage in Management", 2013, Vol. 7, Issue 8, PP. 1460-1465. [2] Samy, S., Abu Nasser, Abu Zaiter and Ola, A., "An Expert System for Diagnosing Eye Disease Using Clips", Journal of Theoritical and Applied Information Technology, 2008, Vol. 4, Issue 10, PP. 923. [3] Borgohain, R., Sanyal, S., "Rule based expert system for diagnosis of neuromuscular disorders", Cornell university library, 2012. [4] Chauhan, R., "An expert system for tourist information management", International journal of Computer Science and Communication, 2010, Vol. 1, No. 2. [5]Diana-Aderina, M., Simona-Alina, S. and Nela, N., "The Use of Expert Systems in Rural Tourism in Maramures", the Annals of the University of Oradea, 2011, Vol. 1, Issue 2. [6] Dogan, Y., and Kut, A., "An Expert System for Summer Tourism in Turkey by Using Text Mining and K-Means++ Clustering", ICT Innovations, 2010. [7] Babu, S., Subramoonium, S. and KV, K.., " Design and Development of an expert system for rating the ecotourism 826 Nada S. Hussein and Musbah J. Aqel / Procedia Computer Science 65 ( 2015 ) 821 – 826 destinations", Conference on Tourism in India-Challenges Ahead, 2008, PP.15-17. [8] Khazad H., " Tourism expert system with clips using PFC", Artificial intelligence and signal processing (AISP), 2012 16th CSA International Symposuim on , May 2-3, 2012, IEEE conference. [9] Badhad, I., "Geography and Tourist Attractions", Alwaraq Publishing company, Amman, Jordan, 2008. 
work_ashml6tmjjbifhwlu2lmefdagm	----	Data mining for grammatical inference with bioinformatics criteria Data mining for grammatical inference with bioinformatics criteria Vivian F. López ⇑, Ramiro Aguilar, Luis Alonso, María N. Moreno Departament Informática y Automática, University of Salamanca, Plaza de la Merced S/N, 37008 Salamanca, Spain a r t i c l e i n f o Keywords: Grammatical inference Bioinformatic Free Context Grammar DNA Sequential patterns a b s t r a c t In this work a novel data mining process is described that combines hybrid techniques of association analysis and classical sequentiation algorithms of genomics, to generate grammatical structures of a spe- cific language. Subsequently, these structures are converted to Context-Free Grammars. Initially the method applies to context-free languages with the possibility of being applied to other languages: struc- tured programming, the language of the book of life expressed in the genome and proteome and even the natural languages. We used an application of a compilers generator system that allows the development of a practical application within the area of grammarware, where the concepts of the language analysis are applied to other disciplines, like bioinformatic. The tool allows measuring the complexity of the obtained grammar automatically from textual data. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction During recent years many approaches were introduced as data mining methods for pattern recognition in biological database. Bio- informatics employs computational and data processing technolo- gies to develop methods, strategies and programs that permit to handle, order and study the immense quantity of biological data that have been generated and are currently generated. To this aim, the computational linguistics has received consid- erable attention in bioinformatics. The study in Searls et al. (1999) indicated that a relation exists between formal languages theory and DNA. Being the linguistic view of DNA sequences, a rich source of ideas to model strings with correlated symbols. Most of the work (Jiménez-Montaño, 2009; Jiménez-Montaño, Feistel, & Diez-Martínez, 2010) has involved examinations of the occur- rences of ‘‘words’’ in DNA. Searls and Dong (1993) found that such a linguistic approach proves useful not only in theoretical charac- terization of certain structural phenomena in sequences, but also in generalized pattern recognition in this domain, via parsing. The information represented on sequences involves grammatical inference for pattern recognition. In this work a novel data mining process is described that com- bines hybrid techniques of association analysis and classical sequentiation algorithms of genomics to generate grammatical structures of a specific language. Subsequently, these structures are converted to Context-Free Grammars (CFG). Initially the meth- od applies to context-free languages with the possibility of being applied to other languages: structured programming, the language of the book of life expressed in the genome and proteome and even the natural languages. We used an application of the compiler generator so named GAS 1.0 system (López, Sánchez, Alonso, & Moreno, 2009), that repre- sents an Integrated Development Environment (IDE) which allow the development of a practical application within the area for the automatic generation of language-based tools, that starts from the traditional solutions and facilitates the use of formal language theory in other disciplines: Grammar-Based Systems (GBSs) (Mernik, Crepinsek, Kosar, Rebernak, & Zumer, 2004). The tool al- lows measuring the complexity of the obtained grammar automat- ically from textual data. 2. Context-Free Grammar A grammar G is defined like G = (N, T, P, S), where N is the set of nonterminals symbols, T is the set of terminals symbols, P is the set of production rules and S is the initial symbol. The language of a grammar L(G) is the set of all terminal strings w that have deriva- tions from the initial symbol. This is: L(G) = {w is in T⁄jS ) ⁄w}. A CFG has production rules like A ? a where A 2 N and a 2 (N [ T)⁄. The substitution of A by a is carried out independently of the place in which appear A (Louden, 1997). The majority of the programming languages are generated by grammars of this type (enlarged with some contextual elements necessary for the lan- guage semantics). 2.1. Grammars and bioinformatics The association analysis involves techniques that are different in its operations but all of them search relations among the attri- butes of a data set. Some techniques are: 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.08.058 ⇑ Corresponding author. E-mail address: vivian@usal.es (V.F. López). Expert Systems with Applications 39 (2012) 2330–2334 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.08.058 mailto:vivian@usal.es http://dx.doi.org/10.1016/j.eswa.2011.08.058 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa � Association rules � Discovery of sequential patterns (DSP), and � Discovery of associations. The association rules (AR) describe the relations of certain attri- butes with regard to others attributes in a database (DB). These rules identify cause-effect implications between the different attri- butes of the DB. Similar to the AR, the discovery of associations (DA) tries to find implications between different couples attri- bute-value so that the appearance of these determine a present association in a good quantity of the registers of the DB. Discovery of sequential patterns (DSP) is very similar to the AR but search for patterns between transactions so that the presence of a set of items precede another set of items in a DB during a per- iod of time. For example, if the data correspond to registers of arti- cles purchased by clients, a description of what articles buys frequently a client can be obtained, and above all, which is the se- quence of its purchase. Thus, the next time, the profile of the client would be known, and it will be able to predict the sequence of its purchase. This criteria can apply to another data control, for exam- ple, in the bioinformatics context, when the data to treat corre- spond to the chain of nucleotides of the genome and sequences are discovered as the patterns that codify genes conform some pro- tein (Fayyad, Piatetsky-Shapiro, Smyth, & Uthurusamy, 1996). Bioinformatics employs hybrid techniques to handle, order and study the immense quantity of biological data that have been gen- erated and are currently generated. For example, for the human genome (HG), the bioinformatics seeks to find meaning to the lan- guage of the more than 37.000 million peers A, C, T and G that have been compiled and stored in the book of life. They offer us the opportunity to understand the gigantic DB that contain the details of the circumstances of time and place in which the genes are activated, the conformation of the proteins that specify, the form in which they influence some proteins on others and the role that such influences can play in the diseases. Besides, what are the relations of the HG with the genomes of the model organisms, like the fly of the fruit, the mice and the bac- teria? Will it be able to discover sequential patterns that show how are related between itself the fragments of information? and will it be able to conform a grammatical structure that show the interpre- tation of the resultant set? If we are able to infer that structure for this type of language we will contribute to understand the real function of the structure of the DNA and we will understand slightly more than the questions presented. One of the applications of the bioinformatics is the pharmacology, offering reviving solu- tions to the old model for the creation of new medicines. It is worth to note that, one of the more elementary bioinformatics operations consists of the search of resemblances between a fragment of DNA recently arranged and the already available segments of diverse organisms (remember and associate this with the DSP). The finding of approximate alignments permits to predict the type of protein that will specify such sequence. This not only provides trails on pharmacological designs in the initial phases of the development of medicines, but suppresses some that will constitute un resolving Fig. 1. Using of the bioinformatics in the pharmacology. V.F. López et al. / Expert Systems with Applications 39 (2012) 2330–2334 2331 https://isiarticles.com/article/22251 
work_asrprf5bwfgjxlzwixdqpn27v4	----	PII: 0898-1221(89)90223-X Computers Math. Applic. Vol. 18, No. 4, pp. 381-398, 1989 0097-4943/89 $3.00 + 0.00 Printed in Great Britain. All rights reserved Copyright © 1989 Pergamon Press pie ELECTRONIC CIRCUIT DIAGNOSTIC EXPERT SYSTEMS--A SURVEY Y. Lmov AT&T Bell Laboratories, Holmdel, NJ 07733, U.S.A. (Recewedll May 1988) Abstract--The electronic circuit diagnostic problem is roughly formulated and subdivided into six subproblems. Current literature and patents are surveyed with respect to the above six subproblems. Some of the existing expert diagnostic systems as well as expert diagnostic shells are described and their limitations are outlined. A review of the relevant terms from AI is included. The bibliography list contains some 300 references. 1. I N T R O D U C T I O N This paper discusses artificial intelligence methods in circuit pack testing. Testing requires investment o f labor, capital and information about testing techniques. The added cost o f testing in the circuit pack production line is justified by several reasons: first, the defective products are not shipped; second, testing helps to develop product repair methodology; and, finally, testing helps to discover manufacturing problems, e.g. low yield due to a single fault may occur because o f malfunctioning insertion robot, a bad batch o f components, or a badly designed in-circuit test. The interested reader should refer to Towill (1987) and Williams et al. (1982) for overviews o f automatic testing. 1.I. F a u l t c a t e g o r i e s Defects in circuit packs, in general, come from three sources: bad components, damage to components or the circuit board, and mistakes in the manufacturing process. Since "different things go wrong at different times because o f different reasons" (Pynn, 1986), determining an optimal manufacturing and test strategy becomes a challenging task. Roughly speaking, one can define three different fault categories: device faults (e.g. unoperative capacitor) assembly faults (e.g. solder shorts, missing components, misoriented components, etc.) and operational faults, such as timing problems arising from apparently good components that fail to work together. According to this somewhat arbitrary fault categorization, a two-stage test strategy has been adopted by most circuit pack manufacturers: the inspection test stage and the functional test stage. Inspection test is designed to examine the circuit pack for proper construction and it consists of shorts test, in-circuit test, and visual test. The in-circuit test examines the function o f each device separately from others. Functional (system) test is designed to examine the operation o f the pack and thus tests the function o f assemblies o f components. As the complexity o f the circuit pack grows, the role o f testing becomes more important and the testing process becomes more difficult. The area o f design for testability is becoming a popular topic by necessity (Williams and Parker, 1983). With the advent of computer technology, automatic test equipment--the equipment for programmed application o f sequences o f stimuli, measurement, and comparison o f the results to the expected values---became the accepted norm. The issue o f software development for such automatic test equipment becomes both increasingly important and difficult as the complexity o f the circuit packs grows. 1.2. A u t o m a t e d d i a g n o s t i c s The ability o f software to generate stimuli and to draw conclusions about the defective components for a given circuit pack is a subtle issue dependent on the experience o f the test engineer responsible for programming o f the automatic test equipment. If the test is to perform two basic CA~aVA t s/*--e 3 81 382 Y. LIROV tasks, i.e. verification (knowing whether everything is functioning as expected), and diagnosis (identifying what causes the unexpected behavior, so it can be fixed) then it is the diagnosis which becomes an increasingly difficult task as a result o f the growing complexity of the circuit packs. 1.3. Importance o f diagnostics Diagnosis o f an electronic circuit usually refers to the process of determining the faulty component(s) that cause an undesired behavior (output) o f the given circuit for some (correctly) given input. Importance o f diagnosis stems from both practical and theoretical considerations. 1.3.1. Practical considerations. The practical considerations are primarily of an economic nature: there is a monetary value associated with every component in the pack, as well as a monetary value associated with the processes of assembly and soldering of the pack. Thus, when a pack is declared faulty, it is desirable to replace only the faulty components. Moreover, the process o f identification o f the faulty components (the diagnostics) consists o f a series o f tests each of which has an associated cost, expressed in such parameters as test setup time, component destruction, etc. The process o f identification o f the faulty components and their subsequent replacement should not be overly expensive. Thus, Observation 1: Diagnostics o f circuit packs is primarily an economic problem (Barber and Yau, 1986). Currently this problem is solved in the majority o f cases by humans who are called test analysts or diagnosticians. Training o f such human analysts is expensive. Their performance depends on the accumulated experience. Indeed, Observation 2: There are human experts that are extremely successful in the area o f diagnostics. Therefore, it is very desirable to have computer programs which are capable o f performing diagnostics. However, Observation 3: Diagnostics problem (i.e. the construction of a computer algorithm which is capable to perform diagnostics) is NP-hard (Hyafil and Rivest, 1976; Ibarra and Sahni, 1976; Karp, 1972). 1.3.2. Theoretical considerations. Therefore, in the absence o f a better alternative, we resort to the construction o f programs that perform approximate diagnostics. In other words, we are willing to give up some o f the precision o f the final description of the faulty component(s) (solution) in order to gain in the length of time spent to obtain that solution. Moreover, in many cases, in the absence o f rigorous mathematical methods to simulate the complex decision making processes which are performed by the human diagnosticians, we are forced to invent different heuristics in order to obtain in some sense similar diagnostic performance. Therefore, the development of such programs is related to the field o f computer science which is commonly known by the name o f artificial intelligence. This relation is further extended to the fields o f control systems, since, as we shall show later, a diagnostics problem is also a planning or control problem. Moreover, approximate solution approaches often employ probabilistic and fuzzy reasonings, thus relating the research in diagnostics to operations research and information theory. The time complexity considerations in the diagnostic and related algorithms requires applications and further developments o f the results belonging to the field o f computational complexity. And, o f course, the area o f application is related to electrical engineering. Diagnostics belongs to the intersection of all o f the above fields. Hence the theoretical importance o f diagnostics. 1.4. Current approaches Two basic approaches and their combinations are currently utilized in industry for conventional diagnostic software development: the fault dictionary approach and the guided probe approach. The fault dictionary approach requires simulation of the faulty behavior of the system and subsequent storage o f the simulation results (as well as the fault assumptions) in the fault dictionary. A "misbehaved" output o f the pack under test can therefore be looked up in the fault dictionary. Basically, two problems prevent the fault dictionary approach from being the widely accepted testing method. These are the overly extensive memory requirements and the diversity o f the faults. As a result, the fault dictionaries are usually incomplete and ambiguous. The guided probe approach requires only simulation o f the correctly functioning circuit pack. The basic tool in utilizing this approach is the blame-shifting mechanism applied after each Electronic circuit diagnostic expert systems---a survey 383 measurement. The upstream (from the test point view) components in the circuit are blamed only when the measurements do not agree with the expected results. The efficient sequencing o f measurements becomes the maii: problem when implementing the guided probe approach. 1.5. Outline o f the overview In order to appreciate and put into perspective the contributions o f the earlier published results on diagnostics, it is convenient to view them with relation to different aspects o f the diagnostics problem. Such a perspective not only allows us to compare different approaches but also provides us with the opportunity to identify yet unsolved problems. The rest o f the paper is organized as follows: first we state the diagnostics problem in a more formal way so as to allow its subdivision into six subproblems. Then we give a brief overview o f each of the subproblems and the tools that are used in their solution. 2. P R O B L E M S T A T E M E N T A circuit pack (plant) P = P(S, B) is modeled by its structure S and its behavior B. Structure S = S(C, T) is defined by the set C of components and topology T on the set C. Behavior B =B(IO, R) is defined by two mathematical models: (1) I 0 model describing the input--output relationships o f every c e C and (2) R model describing the reliability o f every c e C . A component c ~ C is called faulty if its observed behavior bo is different from be, the expected IO behavior (be e IO). Problem 1 (Diagnostic Problem). Given a multicomponent plant P and a symptom indicating an anomaly on the expected IO behavior, identify the subset Cf c C o f faulty components. However, the identification part o f the above problem is itself a complex process. In particular it involves repetitive generation o f hypotheses and their support or disposal. Proving or rejecting a hypothesis requires in turn to perform sequences o f measurements and subsequent sequences of symbolic computations. Therefore, the identification requirement of Problem 1 can be satisfied by recursively solving the following: Problem 2 (Classification Problem). Given a symptom and the results of previously performed measurements, obtain an estimate o f the faulty subset Cf (generate hypothesis), and Problem 3 (Planning Problem). Given an estimate o f Cf, obtain the minimum expected cost plan (sequence) of measurements, so that when performed, the initial hypothesis will be either proved or rejected. Note 1. The actual reasoning technique employed to prove or reject a hypothesis, although an important issue by itself, serves only a secondary role. Diagnostic expert system exploits diverse sources of information to detect failures in a given unit under test (UUT). The relevant information is concerned with the function of the different components, their interconnection, signal characteristics and their paths, and component reliability measures. Moreover, since usually the initial information is insufficient to diagnose correctly, a number o f additional observations (tests) is needed to complete the diagnosis. The description o f available tests, their diagnostic value and costs is yet another kind o f information necessary for diagnosis. Knowledge representation is the term used for the schemes o f storing the different kinds o f information in order to allow for its efficient processing, i.e., failure detection and cost-efficient test sequence generation. Knowledge has to be adequately represented for quick manipulation, for ease in modification, for specifying distinctions between different concepts, and for viewing of multiple levels o f abstractions. Knowledge representation is inherently related to the ways the knowledge is used. Different problem-solving strategies which use the knowledge base define the inference engine o f the diagnostic expert system. Problem 4 (Reasoning Problem). Find an efficient way o f reasoning about electrical circuits. (Such reasoning involves manipulation o f data structures which represent both structural and behavioral data.) 384 Y. L m o v Note 2. Both Problem 2 and Problem 3 involve requirements to reason along both structural and behavioral lines. Such computational requirements require novel approaches for solution of Problem 5 (Representation Problem). Find an efficient way of representing plant P = P(S, B). Efficiency of representation is determined by its level of completeness, consistency, transparency, and computability (Winston, 1984). When solving Problem 1 by using heuristic approaches and in particular by utilizing knowledge of human experts, one can not overestimate the importance of Problem 6 (Knowledge Engineering Problem). Find efficient ways to extract the heuristics about circuit pack troubleshooting from human experts. And, when interfacing to analysts at the shop floor level, one must also face Problem 7 (User Interface Problem). Find efficient ways to conduct a fruitful dialog with the human analysts performing diagnostics. We have identified six problems which belong to the field of artificially intelligent testing. Now we turn to the approaches and tools which are being used to solve those problems. It is more convenient first to address the problems of representation and reasoning, then the planning and classification problems, and, finally the issues about knowledge engineering and user interfaces. 3. R E P R E S E N T A T I O N AND R E A S O N I N G P R O B L E M S Basically two approaches and their combinations have been used to solve diagnostics problems: the rule-based (shallow modeling) approach and the conceptual (deep) modeling approach. Many diagnostic systems have their knowledge in the form of rules, representing separate and modular chunks of knowledge, e.g. "Observations =~ Hypothesis". Often such knowledge may be contra- dictory, incomplete, or too large to be manipulated efficiently. In such cases one models the domain or system under consideration and reasons with this model to generate diagnostics. Such models have been called causal, conceptual, or deep models. 3.1. Performance notions The choice of the correct approach or their combination depends on such factors as size and complexity of the diagnosed system, its reliability on the one hand, and its criticality and stability on the other hand. System size is usually defined in terms of the number of line replaceable units (LRU). System complexity reflects the size of the effort required to track down a fault and it is a function of the number of feedback loops, number of outputs and inputs to LRUs, etc. Reliability of the diagnosed system is defined as the cumulative reliability of its LRUs. Note that the more reliable the system is the more difficult is the diagnostic task. The criticality of the diagnosed system is defined in terms of the length of time interval length within which the diagnosis must be completed. Stability of the diagnosed system is defined by the number of modifications to the system per unit of time. 3.2. Rule-based appoach The rule-based approach can be further subdivided into deterministic decision trees, single relevant rule, and multiple relevant rule methods. 3.2.1. Deterministic decision tree. The deterministic decision tree method is based on the construction of a hierarchy of questions in such a way that each combination of answers can be interpreted as a leaf on the decision tree. The main advantage of this approach is in its simplicity and in its suitability to the systems of high criticality. However, this approach is limited to the systems of low complexity and high stability. 3.2.2. Single relevant rule. The single relevant rule approach is based on searching the rule base for the single rule which can be applied in a given situation or can confirm a given hypothesis. Since such search procedures can be applied recursively, they have been called forward and backward chaining respectively. The advantages and disadvantages of the single relevant rules are similar to the ones of the deterministic decision trees. However, when applying this approach one must be aware of the requirements to maintain a valid rule base, i.e. no two rules may apply to the same situation (consistency), and each situation has to be covered by a rule (completeness). Electronic circuit diagnostic expert systems---a survey 385 Optimality o f the rule base is obtained by ordering the rules in such a way that the expected time to search for the relevant rule is minimized. An entropy-based approach is usually utilized to obtain the optimal ordering o f the rule-base. Sometimes a cost is associated with the firing action o f the rule or/and with the satisfaction o f the antecedent o f the rule (i.e. query o f the user, test setup time, etc.). In such cases the optimality o f the rule-base becomes a difficult problem. The requirements o f consistency, completeness, and optimality are very difficult to satisfy in advance when dealing with systems o f low stability and high complexity. 3.2.3. Multiple relevant rules. To resolve the above problem, the multiple relevant rules approach has been developed by several authors. This approach requires application o f a heuristic to evaluate the chaining strength of the rule on a particular situation as well as a heuristic to combine the strengths o f several rules when forward or backward chaining is involved. A basic expert system building tool (a shell) has been recently patented (Hardy et al., 1987) by Teknoledge. The tool includes interactive knowledge base debugging, question generation, legal response checking, explanation, and certainty factors. (Its source code in Prolog takes 7 pages and is available in the patent documentation.) However the expert systems based on this approach and having several hundreds o f rules again exhibit great difficulties in maintaining their knowledge base. Also the multiple relevant rules approach does not contribute to the problem o f knowledge acquisition for the systems o f low stability. 3.3. Conceptual modeling approach Conceptual modeling approach takes into consideration the original system description rather than using the knowledge compiled by humans into a decision tree or a set o f rules. The conceptual modeling approach can also be subdivided into normal functioning modeling and fault functioning modeling. The overall diagnostic strategies and their relationships were described in Chandrasek- aran and Milne (1985), who used a hierarchy o f four levels: structural, behavioral, functional and pat- tern matching. Given structural representation (list o f components and their connectivity; Forbus, 1987; Bylander and Chandrasekaran, 1985) using qualitative simulation and consolidation pro- posed methods to generate the behavioral description. To generate a functional description o f the de- vice using the behavioral description, one may use the approach o f de Kleer which is based on teleo- logical (device intentions based) reasoning. The pattern matching diagnostic strategy can be applied after compiling (by a human expert or by a machine) the functional description o f the device. Deep (structural, behavioral and functional) modeling can also be used to declare the innocence o f certain components o f the unit under test (UUT; Scarlet al., 1987), or to explain the observed malfunction (Kuipers, 1987). Using de Kleer's (1983) approach, given a low level electronic description o f the components, one is able to deduce the outputs o f the circuit, and consequently diagnose faults in it. Davis (1983) proposed computing the function and the inverse o f each component and then utilizing them to propagate o f combination o f inputs to compute the outputs or vice versa. Digital troubleshooting becomes possible by comparing the expected values with the computed ones. A formal language for U U T description has been developed by (Chandrasekaran, 1985) who compiles the above description into a set o f production rules which are used for diagnosis. M ilne (1985) uses structural reasoning at both the fault isolation to a functional area and possible faults proposal stages. Using his "responsibility theory" appoach only four basic diagnostic rules are needed: two rules to propose faults using description o f structure and another two using description o f function. Forbus (1981) develops a qualitative reasoning methodology to derive the diagnostic rules given a structural description o f objects and their topology. Bylander and Chandrasekaran (1985) propose a methodology to develop the description o f behavior o f composites o f components based only on the descriptions o f the components. 4. P L A N N I N G P R O B L E M Ooel (1980) has shown that the number o f tests needed to pinpoint a fault is at worst linear in the number o f components. However, as we observed earlier, generating the best test sequence is NP-complete (Ibarra and Sahni, 1976). 386 Y. Lmov A test generation algorithm which uses the design information has been proposed some 20 years ago by Roth et al. (1967) and it is called the d-algorithm. The d-algorithm is based on activating simultaneously all the paths to the observable point in the system. Such an approach is impractical for nowadays circuits because o f its high computational requirements. An algorithm o f Cantone (1985) for deciding which test to perform is based on maximizing the ratio o f information gained and expected cost, computed over the available subcircuits. The algorithm is based on the gamma-miniaverage method (Slagle and Lee, 1971). A combi- nation o f structural information (topology o f components) and compiled knowledge (cost o f test and components' failure rates) is used to further isolate the fault to a single functional area. The D A R T algorithm (Bennett et al., 1981) exploits the heirarchy of the circuit in order to deal with the issue o f exponential growth in the number o f computations. Within each level D A R T deduces the suspects by reasoning backward to find a justification for a symptom and generates additional tests by working forward from a behavioral rule for one o f the suspects to observable outputs. For more on planning see also Allen et al. (1983), de Kleer (1977), Hyafil et al. (1976), Ibarra et al. (1976), Konolige et al. (1980), Loveland (1979), McDermott (1981, 1982), Poage et al. (1964), Rosenchein (1981), Sacerdoti (1977), Shirley (1983), Slagle et al. (1971), Vere (1983) and Warren (1976). 5. C L A S S I F I C A T I O N P R O B L E M The diagnostic mode o f reasoning is a term used by Kim and Pearl (1985) to describe the inference process o f updating the causal model o f beliefs due to modifications in the model o f symptoms. Kim and Pearl (1985) describe a computational model for causal and diagnostic reasoning which is a generalization o f the Bayesian methods previously applied to decision trees (DDI, 1973). 5.1. B a y e s i a n m e t h o d s Diagnostic reasoning requires construction o f means to assign and update a blame measure to parts o f the model, based on the observed behavioral deviations o f the physical system. The classical Bayesian probabilistic approach suffers from the following shortcomings: (1) the need to establish a p r i o r i probability o f success o f a test; (2) the need to establish a priori failure rates of the components; (3) the approach is only applicable when the events are mutally exclusive; and (4) conditional probabilities for every combination o f evidence must be computed since the independence o f evidence cannot be assumed. The Dempster-Shafer (Shafer, 1976) theory o f evidence has been proposed to remedy some o f the above problems. The theory still suffers from shortcomings 1, 2 and 4. Recently Thompson and Wojcik o f Westinghouse patented a " M e t h o d and Apparatus for System Fault Diagnosis and Control" (Thompson and Wojcik, 1987). A domain-independent set of metarules is constructed which is able to build a rule network through which belief is propagated to detect defects. 5.2. M u l t i p l e -fault a s s u m p t i o n The multiple-fault assumption poses an additional difficulty in diagnostics problems. Peng and Reggia (1987) developed a probability based approach to guide the backward chaining in such a way that the more likely situations are considered first. By only using the input/output description o f the components and the connectivity information, de Kleer (1986) developed a procedure to incrementally accept or rule-out possible faults. See also Edgar et al. (1984). Fuzzy logic applications for failure analysis were tried by Kitowski and Ksazek (1985). The applied method involves solution o f fuzzy equations derived from a failure-symptom fuzzy relations matrix. Sensor-based systems for realtime monitoring present particularly stringent requirements for expert diagnostic systems. A rule-based system utilizing multiple sensors for obtaining data about chemical parameters was recently patented (Kemper et al., 1987; Westinghouse Electric). The system is utilized to monitor a steam turbine-generator power plant. Electronic circuit diagnostic expert systems--a survey 387 5.3. L e a r n i n g Learning o f diagnostics systems has been explored only at the compiled knowledge base level reasoning: Pazzani (1987) developed a failure-driven learning mechanism to update the rule-base when a diagnostic failure occurs. 6. KNOWLI~DGE E N G I N E E R I N G PROBLEM Problem 6 (knowledge engineering) has been addressed in Cheeseman (1984), who considered the problem o f expert systems learning from data. In particular he proposed a method for extracting information from data to form the knowledge base for a probabilistic expert system. A rule acquisition method for a diagnostic expert system has been recently filed for European Patent (Alexander et al., 1986) by Tektronix. The method includes a grammar for acceptable patterns which are automatically converted to rules. The rule acquisition process includes parsing o f an input sentence as it is received and guiding the expert through the sentence creation. The system is implemented in Small Talk and Prolog. 7. USER I N T E R F A C E PROBLEM Problem 7 (user interface) is considered in Simmons (1985) and Taie et al. (1987) who addressed the issues o f graphical and animated representation o f functioning o f devices. See also Richer and Clancey (1985) and Brown et ai. (1981). 8. O P E R A T I N G I M P L E M E N T A T I O N S Several diagnostic expert systems have been developed in industry during the last few years. A review o f the capabilities and limitations o f some o f the completed projects is given to exemplify the theoretical discussion o f the applicable methods given above. ADVISR (Cooper et al., 1987) is a rule-based expert system developed by AAI Corporation, running on IBM PC/AT, working in conjunction was ATLAS test program software, and designed to indicate faults on a Navy A N / U S M - U U G (V) automatic tester. A single-fault assumption is made as well as the assumption about fully-reliable connectivities. ADS (Magliero et al., 1987) was developed by Harris Corporation as part o f the Navy Integrated Diagnostic Support System. ADS integrates several knowledge representation paradigms such as pre-computed fault-trees, design topology, logical parameters historical records, and production rules. Test selection is done by maximizing the "expected" entropy to cost ratio. The system is implemented in ADA. FCMDS (Bursch et al., 1987) is developed by Honeywell Inc., in Lisp. FCMDS uses frame representation for its knowledge base which is subdivided into Causal Model KB and Test KB. The Causal Model includes the design toplogy. The Test Base contains the basic procedural knowledge for performing tests. Test selection is done similarly as in ADS. Test interpretation procedures involve proprietary algorithms to manipulate failure likelihood measures attached to LRUs. This allows for multiple-fault assumption. A P U M A I D (McCown et al., 1987) is developed by Allied-Signal Aerospace Company (Bendix). It has a temporal, event-based description o f the diagnostic target system, The system is viewed as a sequence o f functional events composed o f location and context, e.g. time and causality. A failure model is obtained as an event-based description. The rest o f the system is implemented in a rule-based fashion. The rules describe the relationships among events, components and parameters. FIS (Pipitone, 1986) is developed at the NRL's Center for Applied Research in AI. It is based on the fault model which is a qualitative causal model composed o f a block diagram o f the U U T combined with a set o f causal rules for each block. Processing hypothesized test failures through the fault model produces an "ambiguity" set o f possibly faulty components. A rule-based test cost generation is used to evaluate the variable costs o f the tests. The best test is chosen similarly to the ADS. It is written in Lisp and runs on Sun. 388 Y. Lmov AI-Ferret (Maguire, 1987) is developed at Hughes Aircraft Company. It is written in Interlisp-D and Loops and runs on Xerox 1109 AI workstation. It is basically similar to FIS and it has been successfully applied to diagnose the TOW missile system. The resulting rule base contains 3950 rules while the connectivity model contains over 900 functions (modes). Haidex (Firdman, 1987) is another diagnostic expert system developed at Hughes. It is based on the conceptual model of the normally functioning system and the set o f deviations from the normal function. The inference engine generates a troubleshooting strategy by decomposing the U U T into its subsystems. It is developed in Zetalisp and runs on Symbolics 3670. M Y C I N (Shortlife, 1976) is one o f the first expert diagnostic systems designed at Stanford to recognize bacterial infections in blood samples. It is based on purely rule-based approach and it contains about 700 rules. Each rule has a confidence factor associated with it and the inference engine is provided with an a d h o c computation scheme to combine the factors when using backward chaining. Other M Y C I N and medical applications related papers include Adams (1976), Benbasset et al. (1976), Ben-Bassat et al. (1980), Bjerregard (1976), Blois (1980), Bosyj (1975), Cantazarite et al. (1979), Chandrasekaran (1982), Chandrasekaran et al. (1979, 1980), Charniak (1983), Cooper (1984), Elstein et al. (1978), Flehinger (1975), Gomez et al. (1981), Gorry (1973), Kolodner et al. (1987), Kulikowski (1970, 1980), Lipkin et aL (1985), Miller et al. (1982), Mittal (1980), Mittal et al. (1979), Patil et al. (1981), Patrick et al. (1981), Regia (1982), Regia et al. (1985), Rubin (1975), Szolovitz (1978) and Weiss et al. (1978). Some other expert systems ( A F H R L , 1984) include A R B Y / N D S ( D U C K ) for communications networks troubleshooting, ACE for preventive maintenance o f telephone cables, LES (frame-based with about 50rules) for electronic maintenance, STAMP, I D T (Shubin et al., 1982), CRIB (Hartley, 1984) and C R I T T E R (Kelly et al., 1982). 9. SYSTEM B U I L D I N G T O O L S The core o f an expert system has a knowledge base and an inference engine that operates on the knowledge base. User interface is considered at two different levels: when developing a system, and when deploying it. Some o f the important concepts are summarized below. 9.1. K n o w l e d g e representation There are three important aspects o f knowledge representation which must be available in any diagnostic system building tool. These are: object descriptions, actions, and certainties. The most convenient way to represent objects is by frames which are tabular data structures (records). A frame consists o f slots which are filled with either data about objects or relations to other frames. A built-in inheritance mechanism allows to inherit or override the information from other frames. Actions are used to modify the relevant database. Actions are usually implemented by rules or procedures. The degree or correctness o f data and/or knowledge is usually expressed by means o f belief functions, certainty factors, or probabilities. 9.2. I n f e r e n c e approach The inference approach in a typical diagnostic expert system may be a combination o f forward and backward chaining, hypothetical reasoning, and blackboard mechanism. 9.2. I. Chaining. Backward chaining is a method o f evaluating hypothesized conclusions to see whether they are supported by the evidence. It is usually implemented via i f - t h e n rules, starting with rules that have the hypothesized conclusions as their consequents. The set o f rules is then searched for those in which the antecedents include the previous conclusions. This process (backward chaining) is continued until the hypothesis is proved. If the p r o o f is unachievable, then some other hypothesis may be tried. Forward chaining starts with the present data with is matched with the antecedents until a relevant rule is found. The consequent o f that rule determines the new present data and the process o f matching is repeated until a desired conclusion is reached. Electronic circuit diagnostic expert systems--a survey 389 9.2.2. Hypothetical reasoning. Hypothetical reasoning (truth maintenance) refers to the solution in which hypotheses have to be assumed when initializing the search (e.g. forward or backward chaining.) Nonmonotonic reasoning refers to the ability o f retracting the consequences o f the assumptions which have been found untrue during the search process. Different tools handle nonmonotonic reasoning by using viewpoints, contexts or worlds, or nonchronological backtrack- ing. Hypothetical reasoning literature includes Doyle (1979), and McDermott (1980, 1982). The blackboard mechanism (Hayes-Roth, 1983) refers to a common data structure shared by a group of cooperating expert systems. An agenda often is used to control the development o f the solution on the blackboard. This mechanism is especially useful when several competing solutions are developed in parallel. 9.2.3. Pattern matching. Pattern matching is usually required in two mechanized reasoning instances: when matching rule atecedents to the current knowledge situation to fire the best applicable rule, and when matching partially filled frames to the existing knowledge base to determine the current knowledge. 9.3. Expert system shells A diagnostic expert system development tool is required to have the following capabilities: (1) Knowledge representation: structured rules, certainty factors, frames; (2) Inference engine: forward chaining, backward chaining, backward chaining, blackboard-like, truth maintenance, pattern matching; (3) Developer interface: KB editor, menu, check for consistency, graphics representation o f KB, graphics utilities to build end user interface, debug; and (4) Interface to general-purpose programming language (e.g. C). Unfortunately, none o f the available products in the current market have all o f the above features, although all o f the products have many features in addition to the above list. The only products that come close to our requirements are ART, KEE and Knowledge Craft (see Expert System Shells List). However, the cost o f these tools is in the range o f $10,000-$65,000, besides additional expensive training. 9.4. Diagnostic tools Two commercially available special purpose diagnostic expert system tools have been developed recently, namely, IN-ATE, and AI-Test. We give a short outline o f their characteristics. 9.4.1. I N - A T E . IN-ATE (Cantone et al., 1985) provides an excellent menu-driven syntax for specifying fault diagnosis expert rules. It also allows specification o f the signal flow and uses it when searching for the defective component. In additon it has frames-based knowledge representation and a mechanism to perform probabilistic reasoning. The system does have an interface to C, and it does have capabilities for automated testing the knowledge base consistency. It runs however only on a Macintosh computer. The system does not have a separate electronics theory and measurements related data-base. Therefore, each test point must have all the relevant information. Thus, some duplication of information and difficulties in knowledge acquisition are inevitable. 9.4.2. AI-Test. AI-Test (Ben-Bassat et al., 1987) is developed by Intelligent Electronics and runs on IBM-PC/AT. AI-Test is comprised o f two knowledge bases---the Universal Knowledge Base which contains general information relating to electronics theory, kinds o f measurements (DC, frequency, etc), signal types, etc. and the UUT-specific knowledge base which includes the block diagram o f the U U T and tests descriptions. The inference engine interprets the test results (by updating the model o f beliefs) and chooses the next best test. No mechanism is provided to update the initial failure rates o f the components. Also, the choice o f the appropriate kind o f test at a specific test point must be performed by the user. An AI system should be able to suggest the most informative test description automatically. The user interface has been improved greatly since the last release. 10. C O N C L U S I O N S New circuit pack manufacturing technologies (e.g. surface mount) contribute further to the complexity o f the circuit pack and thus make the testing even more difficult. For example a new CAMWA 18/4---F 390 Y. Lmov approach in electronic design is the Built-In-Self-Test (BIST) approach (Agarwal, 1987). This approach has the positive features o f testing the equipment at its normal operation speed, as well as running the test in parallel on all the devices. Thus a malfunctioning device can in principle declare its faulty condition faster and seemingly reduce the difficulty o f the diagnostic task. 10.I. New challenges As happened before, the new technology has posed new testing challenges. In particular, BIST technology requires separate independent testing o f the built-in test equipment. Failing to test the built-in hardware may result in two kinds o f errors: a correctly functioning device declaring itself as faulty and thus being replaced for the wrong reason (such repeated mistakes will later be propagated to mistakes in the process control) and a misbehaving device declaring itself as being healthy, thus postponing the troubleshooting to later stages and therefore increasing the manu- facturing and maintenance costs. 10.2. Drawbacks o f the rule-based approach Since the success o f MYCIN, the rule-based approach has been implemented several times in circuit pack diagnostics. The mechanisms that proved useful in medical applications were automatically applied in the electronics industry. However, in this industry only part o f the knowledge (general electronics laws) is static, while the rest o f the knowledge changes with every new design and with every new batch o f components. Moreover, the usually available statistical data is not present any more. One the other hand, another kind o f data (connectivity and signal paths) is available. These differences must be recognized and utilized (e.g. IN-ATE and AI-Test) in order to develop a successful expert system. 10.3. Classification and planning problems The diagnostic problem is still far from being resolved and it is currently an area o f active research and development. In particular in the area o f the classification problem there is a lack o f computationally efficient schemes to propagate the results o f tests. A prevailing assumption is the "perfect info" assumption underlying the construction of most o f the diagnostic expert systems. The mechanisms for dealing with partial, unreliable, and contradictory results are mainly based on the Dempster-Shaffer theory. The conditional probabilities (correlation data), which are usually available in medical applications are not generally available in electronic circuit pack diagnostics applications. Thus, the applicability o f the "classical" belief propagating schemes is questionable. This issue requires a separate investigation in light of dealing with structural circuit description in addition to rules. The issue of dealing with noisy data also requires a specialized effort. In the area o f the planning problem there is lack o f mathematically sound basis for the best test selection. Most of the existing systems either do not address the issue o f test planning or use an ad hoc method. Computational efficiency is another issue which must be addressed separately. 10.4. Deep knowledge representation and reasoning Although it is now quite obvious that a purely rule-based approach is not practical, there are not many pragmatic ways to implement a deeper kind o f knowledge besides the connectivity o f the circuit pack. The area o f knowledge representation and reasoning currently receives most of the AI community's attention. Yet it is still unclear to what extent qualitative reasoning and functional knowledge can be used for practical applications. Consequently all o f the existing diagnostic expert systems implementations are using a combination o f rule-based and connectivity knowlege with a mechanism for belief functions manipulations and an ad hoc (entropy vs cost based) best test selection heuristic. Thus, efficiency o f such systems is difficult to assess. 10.5. Learning and user interface The least amount o f work is done in the areas o f learning and user interface. It seems that the most desirable approach to "engineer" knowledge is via a mechanized interview o f the design engineers. Experiments have not been performed, nor have tools been constructed to achieve such goals. Electronic circuit diagnostic expert systems--a survey 391 10.6. More open problems Only one system (AI-Test) has implemented a separate global electronics knowledge base containing generally true facts about electronic circuits behavior, possible tests, etc. However, there is no known technique to match the best test to its proposed location and to its time occurrence. This test matching problem solution could provide an important aid in diagnostic expert system construction. And finally, applications o f new technologies (e.g. neural networks) have yet to be examined. Acknowledgements--The author is grateful to On-Ching Yue and Patricia Wirth for many constructive suggestions that improved the readability of this paper. Special thanks are due to Nancy Perine for her help in collecting the bibliography. A P A R T I A L L I S T O F E X P E R T S Y S T E M S H E L L S 1. ART, Inference Corporation, 5300 W. Century Bird, LA, CA 90045, U.S.A. 2. KEE, Intelli Corp, 1975 E1 Camino Real West, Mountain Views, CA 94040, U.S.A. 3. Knowledge Craft, Carnegie Group, 650 Commerce Court, Station Square, Pittsburgh, PA 15219, U.S.A. 4. Picon, Lisp Machine, 5 Tech Drive, Andover, MA 01810, U.S.A. 5. S. 1. M. 1, Teknoledge, 1850 Embarcadero Rd, P.O. Box 10119, Palo Alto, CA 94303, U.S.A. 6. KES, Software Architecture and Engineering, 1600 Wilson Blvd, Suite 500 Arlington, VA 22209, U.S.A. 7. Nexpert Object Newron Data Corp., 444 High St., Palo Alto, CA 94301, U.S.A. , B I B L I O G R A P H Y Adams, J., A probabilistic model of medical reasoning and MYCIN model. Math. Biosci. 32, 177-186 (1976). Addis, T. R., Towards an "expert" diagnostic system. 1CL Tech. J. 1, 79-105 (1980). Addis, T. R. and R. T. Hartley, A fault finding aid using a content addressable file source. ICL Research and Advanced Development Centre, Technical Report 79/3, Stevenage, U.K. (1979). Ahuja, S., Artificial intelligence environment of the analysis and classification of errors in discrete sequential processes, Ph.D.Dissertation, Univ. Maryland, Baltimore (1985). Aikins, J., Prototypes and production rules: a knowledge representation for computer consultations. Ph.D. Dissertation, Rep. No. STAN_CSD_80-814, Stanford University (1980). Alexander, J., M. Freiling, B. Phillips and S. Messick, Rule acquisition for expert systems. European Patent Application 85114858.5 (22 Nov. 1985). Ali, M. and D. A. Scharnhorst, Knowledge-based avionics for interactive fault assessment. WATTEC 86: 13th A. Energy Conf. Knoxville, Ten. (February 1986). Ali, M. and D. A. Scharnhorst, Sensor-based fault diagnosis in a flight expert system. Proc. IEEE 2nd Conf. Artificial Intelligence Applications, Miami, Fla., (December 1985). Andersen, S. K. and M. Woldbye, Knowledge representations for diagnostics and test planning in the domain of electromyography. 7th European Conf. Artificial Intelligence Proc., Vol. I, pp. 357-368 (1986). Andow, P. K., Real time analysis of process plant alarms using a minicomputer. Comput. chem. Engng. 4, 143-155 (1980). Andow, P. K., Computer based alarm systems. Proc. 4th IEEE Conf. Trends in On-line Computer Control Systems, IEE Publication No. 208, pp. 89-93 (1982). Andow, P. K., Propagation of faults in process plants: A state-of-the-art review. Chemical Process Hazards With Special Reference to Plant Design, 7th Int. Syrup., IChemE Symposium Series No. 58, pp. 225-243 (1980). Apfelbaum, L., An expert system for in-circuit fault diagnosis. IEEE 1985 Proc. Int. Test Conf. (1985). Arabian, J. H., User requirements for a dynamic high speed functional tester (DHSFT). Proc. 9th Automatic Test Electron. Seminar~Exhibit. June (1981). Armstrong, D. B., A deductive method for simulating faults in logic circuits. IEEE Trans. Comput. C22, 464-471 (1975). AFHRL, Artificial intelligence in maintenance. AFHL, USAF Systems Command, T. R. AFHRL-TR-84-25 (1984). Barber, J. S., Private communication. Barber, J. and C. W. Yau, Private communication. Barr, A. and E. Feigenbaum, The Handbook o f Artificial Intelligence. Kaufmann Press, Calif. (1981). Bartlett, S. L., C. Cole and R. Jain, Expert system for visual solder joint inspection. Proc. 3rd Conf. A I Applications, pp. 58-63 (1987). Basil, W. and L. Felkel, Disturbance analysis systems. In Human Detection and Diagnosis o f System Failures (Ed. J. Rasmussen and W. B. Rouse). Plenum Press, New York (1981). Benbasset, J. and A. Schiffmann, An approach to teaching the introduction to clinical medicine. Ann. internal Med. 84(4), 477-481 (1976). Ben-Bassat, M. et aL, Pattern-based interactive diagnosis of multiple disorders---the MEDAS system. IEEE Trans. Pattern Analysis Machine lntell. PAMI2, 148-160 (1980). Ben-Bassat, M., D. Ben-Arie, I. Beniaminy, J. Cheifetz and M. Klinger, A case study with AI-TEST: an expert system for electronic troubleshooting. Proc. AUTOTESTCON '87, pp. 363-370 (1987). Bennett, J. S. and C. R. Hollander, DART: an expert system for computer fault diagnosis. Proc. 7th lnt. Joint Conf. Artificial Intelligence (1981) Bjerregaard, B., J. Holst-Christensen, E. Kalaja, J. Lund.Kristensen, F. T. DeBombal and H. D. Horroeks, Computer-aided diagnosis of the acute abdomen: a system from Leeds used on Copenhagen patients. Proc. Information Processing Work Conf. TC4 (1976). 392 Y. Lxgov Blois Marsden, S., Clinical judgement and computers. N. Engl. J. Med. 303(4), 192-197 (1980). Bosyj, M. A., A diagnostic program for the design of procurement systems. MIT project MIT Technical Report 53 (1975). Breuer, M. A., Functional partitioning and simulation o f digital circuits. IEEE Trans. Comput. C19, 1038-1046 (1970). Breuer, M. A. and M. Abramovici, Fault diagnosis based on effect-cause analysis. Proc 17th Design Automation Conf., San Francisco, CA (June 1980). Breuer, M. and A. Friedman, Diagnosis and Reliable Design o f Digital Systems. Computer Science Press. New York (1976). Brown, A. L., Qualitative knowledge, causal reasoning and the localization of failures. MIT AI Lab Technical Report 362 (1976). Brown, D. C. and B. Chandrasekaran, Design consideration for picture production in a natural graphics systems. Comput. Graphics 15(2), 174-207 (1981). Brown, J. S., Remarks on building expert systems. Proc. 5th Int. Joint Conf. Artificial Intelligence, San Francisco, CA, pp. 1003-1005 (1977). Brown, J. S. and R. R. Burton, Multiple representations o f knowledge for tutorial reasoning. In Representation and Understanding (Ed. D. Bobrow and A. Collins). Academic Press, New York (1975). Brown, J. S. and R. R. Burton, Diagnostic models for procedural bugs in basic mathematical skills. Cognitive Sci. 2, 155-198 (1978). Brown, J. S., R. R. Burton and A. G. Bell, SOPHIE; a sophisticated instructional environment for teaching electronic troubleshooting (an example of AI in CAI). Bolt Beranek & Newman, Inc., Report 2790 (1974). Brown, J. S., R. R. Burton and A. G. Bell, SOPHIE: a setup towards a reactive learning environment. Int. J. Man-Machine Stud. 7, 675-696 (1975). Brown, J. S., A. Collins and G. Harris, Artificial intelligence and learning strategies. In Learning Strategies (Ed. A. O'Neil). Academic Press, New York (1978). Brown, J. S., R. Rubinstein and R. R. Burton, Reactive learning environment for computer assisted electronics instruction. Bolt Beranek & Newman, Inc., Report 3314 (1976). Brown, J. S., R. R. Burton and J. de Kleer, Pedagogical, natural language and knowledge engineering techniques in SOPHIE I, II and Ill. In Intelligent Tutoring Systems (Ed. D. Sleeman and J. S. Brown). Academic Press, New York (1982). Burger, L. and E. Zobor, On-line alarm analysis in the hierarchical control system of the WWR-SM Research factor. Nuclear Power Plant Control and Instrumentation, Vol. 2, pp. 95-107. Vienna. Academic Press, New York (1978). Burns, M. and J. Pearl Causal and diagnostic inferences: a comparison of validity. Organizl Behav. Hum. Perform. 211, 13-26 (1981). Bursch, P., J. Meisner and K. Winegar, A PC based expert diagnostic tool. Proc. A U T O T E S T C O N '87, pp. 73-78 (1987). Cain, D. and E. Zebroski, The conceptual design of a power plant safety panel. Nucl. Engng 25, 40-44 (1980). Cantone, R., Model-based probabilistic reasoning for electronic troubleshooting. Proc. 8th Int. Joint Conf. Artifical Intelligence, Karlsruhe, Germany, August (1983). Cantone, R., W. Lander, M. Marrone and W. Gaynor, INATE/2: interpreting high level fault modes. Automated Reasoning Corp., Sunnyvale, Va. (1984). Cantone, R., W. B. Lander, M. P. Marrone, and W. Gaynor, Automated knowledge acquisition in IN-ATE using component information and connectivity. Sigart Newsletter No. 93 (1985). Catanzarite, V. and A. Greenburg, Neurologist--a computer program for diagnosis in neurology. Proc. 3rd Symp. Computer Applications in Medical Care, IEEE, pp. 64-72 (1979). Chandrasekaran, B., On evaluating AI systems for medical diagnosis. Proc. M E D C O M P 82, Catalog No. 82, CH 1797-0, IEEE, New York, pp. 335-338 (1982). Chandrasekaran, B. and T. Bylander, Functional representations and behaviour composition by consolidation: two aspects of reasoning about devices. Sigart Newsletter No. 93 (1985). Chandrasekaran, B. and S. Mittal, Deep versus compiled knowledge approaches to diagnostic problem-solving. Int. J. Man-Machine Stud. 425-436 (1983). Chandrasekaran, B., F. Gomex, S. Mittal and J. W. Smith. An approach to medical diagnosis based on conceptual structures. Proc. 6th Int. Joint Conf. Artificial Intelligence, pp. 134-142 (1979). Chandrasekaran, B., S. Mittal and J. W. Smith, RADEX--towards a computer-based radiology consultant. In Pattern Recognition in Practice (Ed. E. S. Gelsema and L. N. Kanal), pp. 463-474. North-Holland, Amsterdam (1980). Chandrasekaran, B., S. Mittal and J. W. Smith, Reasoning with uncertain knowledge: the M D X approach. In Proc 1st Ann. Joint Conf. American Medical Information Association (AMIA ), (Ed. D. A. B. Lindberg et al.) pp. 335-339. Masson, New York (1982). Chandrasekaran, B., D. D. Sharma and D. W. Miller, The application of knowledge-based systems to reactor operations. Trans. A m Spc. (1982). Charniak E., The Bayesian basis of common sense medical diagnosis. Proc. Third Natn. Conf. Artificial Intelligence, AAAI, pp. 70-73 (1983). Charniak, E. and D. McDermott, Introduction to Artificial Intelligence. Addison-Wesley, Reading, Mass. (1985). Cheeseman, P., Learning of expert systems from data. Proc. A A A I (1984). Chen, C. C. J. J. Schwermann and G. W. Sturm, Private communication. Chiang, H. Y. P. et al., Comparison of parallel and deductive fault simulation. 1EEE Trans. Comput. C23, ! 132-1138 (1974). Christenson, J. et al., Implementation of an automated analysis system in an operating nuclear power plant. Engng Des. 67, 297-304 (1981). Clancey, W. J., Acquiring, representing, and evaluating a competence model of diagnostic strategy. Technical Report HPP-84-2, Stanford University, Stanford, Calif. (1984). Clancey, W. J., Heuristic classification. Technical Report KSL-85-5, Stanford University, Stanford, Calif. (1985). Clancey, W. J., The epistemology of a rule-based expert system--a framework for explanation. Artificial lntell. 20, 215-251 (1983). Clancey, W. J., Tutoring rules for guiding a case method dialogue. Int. J. Man-Machine Stud. 11, 25-49 (1979). Clancey, W. J. and R. Letsinger, NEOMYCIN: reconfiguring a rule-based expert system for application to teaching. Proc. 7th Int. Joint Conf. Artificial Intelligence, pp. 829-836 (1981). Clocksin, W. and A. Morgan, Qualitative control. European Conf. Artificial Intelligence 86, Brighton, England, 21-25 July (1986). Electronic circuit diagnostic expert systems---a survey 393 Cooper, G., NESTOR: a computer-based medical diagnostic aid that integrates causal and probabilistic knowledge. Ph.D Dissertation, Department of Computer Science, Stanford Univ. Calif. (1984). Cooper, L. and T. O. Cooper, A performance evaluation of a prototype expert system. Proc. A U T O T E S T C O N '87, pp. 55-59 (1987). Cross, S. E., An application of qualitative simulation to expert system plan justification. Sigart Newsletter No. 93 (1985). Cross, S. E., Towards an expert system architecture for flight domain applications. Proc. IEEE 1984 National Aerospace and Electronics Conf., Dayton, Ohio. Vol. 2, pp. 784-788 (1984). Cunningham, P., J. Gleeson, S. Hakiel and M. Wheatley, Diagnostic heuristics and perspectives. In Research and Development o f Expert Systems, VoL 3 (Ed. M. A. Bramer), pp. 242-253 (1987). Davis, R., Amplifying expertise with expert systems. In The A I Business (Eds P. H. Winston and K. Prendergast). MIT Press, Cambridge (1984). Davis, R., Applications of meta-level knowledge to the construction, maintenance and use of large knowledge bases. Ph.D Thesis, Stanford University [STAN-CS-76-552, HPP-76-7] (1976). Davis, R., Diagnosing via causal reasoning: paths of interaction and the locality principal. Proc. AAAI-83, Washington, D.C. (1983). Davis, R., Diagnosis based on description of structure and function. Proc. Am. Associal. Artificial Intell. 82, pp. 137-142 (1982). Davis, R., Expert systems: where are we? and where do we go from here? A I Mag. 3(2), 3-22 (1982a). Davis, R., Diagnosis reasoning based on structure and behavior. MIT AI Lab Memo, No. 739, June (1984). Davis, R., Diagnostic reasoning based on structure and behavior of digital hardware. Computer October, 75-82 (1983). Davis, R., Diagnostic reasoning based on structure and behavior. Artificial lntell. 24, 347-410 (1984). Davis, R., Metarules: reasoning about control. Artificial Intell. 15, 179-222 (1980). Davis, R., Qualitative Reasoning About Physical Systems. MIT Press, Cambridge (1985). Davis, R., Reasoning from first principles in hardware troubleshooting. Int. J. Man-Machine Stud. 17, 45-69 (1983). Davis, R. and H. Shrobe, The hardware troubleshooting group. Sigart Newsletter No. 93 (1985). Davis, R. and H. Shrobe, Representing structure and behavior. Artificial Intell. 24, 347-410 (1984). Davis, R. and H. Shrobe, Representing structure and behavior of digital hardware. IEEE Comput. October, 75-82 (1983). Davis, R., H. Shrobe, W. Hanscher, K. Wiechert, M. Shirley and S. Polit, Diagnosis based on description of structure and function. Proc. Am. Assoc. Artificial Intell. 13%142 (1982). Dempster, A., A generalization of Bayesian inference. J. R. Statist. Sci. Serv. B. 30, 205-247 (1968). Dhir, V. K. et al., Pressurized water reactor (PWR) system and simulation and disturbance analysis for anomalous transients and degraded system conditions. Proc. Winter Simulation Conf., IEEE, New York (1979). Doyle, J., A truth maintenance system. Artificial Intell., 12, 17-34 (1979). Doyle, J., Analysis by propagation of constraints in elementary geometry problem solving. MIT AI Lab Working Paper 108 (1976). Doyle, J., Truth maintenance systems for problem solving. MIT AI Lab. Technical Report 19 (1978). Duda, R., J. G. Gaschnig and P. F. Hart, Model design in the PROSPECTOR consultant system for mineral exploration. In Expert Systems in the Microelectronic Age (Ed. D. Michie), pp. 25-49. Edinburgh University Press, Edinburgh, Scotland (1979). Duda, R. O., P. E. Hart, P. Barrett, J. G. Gaschnig, K. Konolige, R. Reboh and J. Slocum, Development of the prospector consultation system for mineral exploration: final report. Technical Report SRI International, Menlo Park, Calif. (1978). Duda, R., et al., Subjective Bayesian methods for rule-based inference systems. Proc. AFIPS, Washington D.C., pp. 1075-1082 (1976). Dworzak, F. et al., Design and implementation of a computerised system for evaluation of plant status with respect to safety technical regulations. Int. Symp. Nuclear Power Plant Control and Instrumentation, Munich, Paper IAEA-SM- 265/95IAEA (1982). DDI Handbook for Decision Analysis. Decision & Design Inc., McLean, Va. (1973). Earman, L. D., et al., Hearsay-II. Comput. Surveys 12, 2 (1980). Edgar, G. and M. Petty, Location of multiple faults by diagnostic expert systems. SPIE Proc. 485, Applications o f Artifical Intelligence (1984). Edwards, J., Covers and packings in a family of sets. Bull. Am. math. Soc. 68, 494-499 (1962). Eichelberger, E. B.,, Hazard Detection in combinational and sequential switching circuits. I B M J Res. Dev. March (1965). Elstein, A. S., L. Shulman and S. Sprafka, Medical Problem Solving: An Analysis o f Clinical Reasoning. Harvard University Press, Cambridge, Mass. (1978). Erman, L. D., F. Hayes-Roth, V. R. Lesser and D. R. Reddy, The Hearsay-II speech-understanding system: integrating knowledge to resolve uncertainty. A C M Comput. Surveys 12, 213-252 (1980). Erman, L. D. and V. R. Lesser, A multi-level organization for problem-solving using many diverse cooperating sources of knowledge. Proc. 4th Int. Joint Conf. Artificial Intelligence. pp. 483-490 (1975). Fikes, R. E., Deductive retrieval mechanisms for state description models. Stanford Research Institute AI Technical Note 106 (1975). Fink, P. K. and J. C. Lusth, Expert systems for diagnostic expertise in the mechanical and electrical domains. IEEE Trans. Systems Man Cybernet. SMC17, 340-349 (1987). Firdman, E., HAIDEX: tomorrow available today. Through the Looking Glass 2(6), 2-8 (1987). Fitzgerald, M. W. et al., 5ESS T M switching system software: diagnostics. A T & T Tech. J. January (1986). Flehinger, B. and R. Engle, HEME: a self-improving computer program for diagnosis-oriented analysis of hematologic diseases. I B M J. Res. Dev. November, 557-564 0975). Florcik D. and H. Wong, A tester independent automated test preparation process for loaded boards. Proc. 1982 Int. Test. Conf., San Francisco, CA (1982). Forbus, K. D., Interpreting observation of physical system. I E E E Trans. Systems Man Cybernet. SMC17, 350-359 (1987). Forbus, K., Qualitative reasoning about physical processes Proc. 7th HCAI, Madrid, Spain (1981). Forbus, K., Qualitative process theory. MIT Technical Report 789 (1984). 394 Y. LIRoV Fox, M. S. and D. J. Moslow, Maximal consistent interpretations of errorful data in hierarchically modeled domains. Proc. 5th Int. Joint Conf. Artificial Intelligence, Cambridge, Mass. (1977). Fox, M. S., S. Lowenfeld and P. Kleinovsky, Techniques for sensor-based diagnosis. Proc. 8th IJCAI (1983). Freund, T. G., Applying knowledge engineering to TPS development. Proc. IEEE Autotestcon, pp. 381-321 (1983). Frey, T. and R. A. Kisner, Survey of methods for improving operator acceptance of computerized aids. N U R E G CR-2586 (1982). Friedman, L., Extended plausible inference. Proc. 7th Int. Joint Conf. Artificial Intelligence (1981). Frohwerk, R. A., Signature analysis: a new digital field service method. Hewlett-Packard J. May, 22-28 (1977). Fusaoka, A., H. Seki and K. Takahashi, A description and reasoning of plant controllers in temporal logic. Proc. 8th Int. Joint Conf. Artificial Intelligence, Karlsruhe, Germany, pp. 405-408 (1983). Garvey, T. D. et al., An inference technique for integrating knowledge from disparate sources. Proc. 7th Int. Joint Conf. Artificial Intelligence (1981). Genesereth, M. R., Diagnosis using hierarchical design models. Proc. AAAI-82, pp. 278-283 (1982). Genesereth, M. R., The use of design descriptions in automated diagnosis. Artificial Intell. 2,4, 411-436 (1984). Genesereth, M. R., The use of hierarchical design models in the automated diagnosis of computer hardware faults. HPP-81-20, Stanford University Heuristic Programming Project (1981). Gimmy, K. L. and E. Nomme, Automatic diagnosis of multiple alarms for reactor control rooms. Savannah River Lab Report No. DP-MS-8L-91 (1981). Goel, P., Test generation cost analysis and projections. 17th Design Automation Conf. Proc. Washington, D.C. (1980). Gomez, F., Knowledge organization and distribution for diagnosis. IEEE Trans. Systems Man Cybernet. 9, 31-40 (1979). Gomez, F. and B. Chandrasekaran, Knowledge organization and distribution for medical diagnosis. IEEE Trans. System Man Cybernet S M C l l , 34 42 (1981). Gordon, J. and E. Shortliffe, A method for managing evidential reasoning in a hierarchical hypothesis space. Artificial lntell. 26, 323-357 (1985). Gorry, G. A., Computer-assisted clinical decision-making. Meth. Inf. Med. 12, 45 51 (1973). Grillmeyer, O. and A. J. Wilkinson, The design and construction of a rule base and inference engine for test system diagnosis. Int. Test Conf. 1985 Proc., IEEE Computer Society, pp. 857-867 (1985). Guarro, S. and D. Okrent, The logic flowgraph: a new approach to process failure modeling and diagnosis for disturbance analysis applications. U.S.A. nucl. Technol. 67, 348-359 (1984). Guarro, S. and D. Okrent, On the implementation of disturbance analysis systems for detection and diagnosis of safety-related transients. Nuclear Power Plant nucl. Engng Des. 65, 3-16 (1981). Guida, G., Inference based on causal knowledge: deep versus shallow reasoning. Proc. Unicorn Seminar Expert Systems and Optimisation in Process Control, London, December (1985). Gunhold, R. and J. Zettel, An expert system for automated testing of electrical components and circuits. A I and Advanced Computer Technology Conf. Proc., p. 9 (1986). Hamscher, W. and R. Davis, Using structural and functional information in diagnostic design. MIT A1 Lab Memo 707 (1983). Hamscher, W. and R. Davis, Diagnosing circuits with state--an inherently underconstrained problem. Proc. 1984 Conf. American Association for Artificial Intelligence (1984). Hardy, S. and R. Joyce, A basic expert system tool. US Patent 4648066 (3 March 1987). Hartley, R. T., CRIB: computer fault-finding through knowledge engineering. Computer March, 76-83 (1984). Harvey, A. M. and J. Bordley III, Differential Diagnosis. Saunders, Philadelphia, Penn. (1972). Hasling, D. W., W. J. Clancey and G. R. Rennels, Strategic explanations for a diagnostic consultation system. Int. J. Man Machine Stud. 20, 3-19 (1984). Hayball, C., Fault diagnosis in the wave soldering o f printed circuit assemblies. Alvey IKBS Research Theme: Expert Systems, Deep Knowledge, Workshop No. 2, 38-43 (1986). Hayes-Roth, B., The blackboard architecture: a general framework for problem solving? Heuristic Programming Project Report No. HPP-83-30, Stanford University (1983). Hayes-Roth, B., A blackboard model of control. Heuristic Programming Project Report No. HPP-83-83, Stanford University (1983). Hayes-Roth, B. and F. Hayes-Roth, A cognitive model of planning. Cognitive Sci. 7, 275-310 (1979). Hayes-Roth, F., D. Waterman and D. Lenat (Eds) Building Expert Systems. Addison-Wesley, New York (1983). Hayes-Roth, F., D. A. Waterman and D. B. Lenat, Principles o f pattern-directed inference systems. In Pattern-Directed Inference Systems (Ed. D. A. Waterman and F. Hayes-Roth) pp. 576-601. Academic Press, New York (1978). Herbert, M. R . , A review of on-line diagnostic aids for nuclear power plant operators. Nucl. Energy 23, 259-264 (1984). Hudlicka, E. and V. Lesser, Modeling and diagnosing problem-solving system behavior. IEEE Trans. Systems Man and Cybernet. SMC17, 407-419 (1987). Hudlicka, E. and V. R. Lesser, Meta-level control through fault detection and diagnosis. Proc. 1984 Conf. American Association for Artificial Intelligence (1984). Hughes Aircraft Corp., Knowledge based systems (KBS) study for advanced mate systems. HAC Report No. FR80-75-869 (1980). Hyafil, L. and R. L. Rivest, Constructing optimal binary decision trees in NP-complete. lnfo. Process. Lett. 51, 15-17 (1976). Ibarra, O. H. and S. Sahni, Polynomially complete fault detection problems. I E E E Trans. Comput. C24, 242-250 (1976). Johnson, P., A. Duran, F. Hassebroek, J. Moiler and M. Prietula, Expertise and error in diagnostic reasoning. Cognitive Sci. 5, 235-283 (1981). Josephson, J., Explanation and induction. Ph.D. Dissertation, Department of Philosophy, Ohio State Univ. (1982). Josephson, J. R., B. Chandreseharan, J. R. Smith and M. C. Taumer, A mechanism for forming composite explanatory hypotheses. IEEE Trans. Systems Man Cybernet. SMC17, 445-455 (1987). Karp, R., Reducibility among combinatorial problems. In Complexity o f Computer Computations (Eds R. Miller and J. Thatcher), pp. 85-103. Plenum Press, New York (1972). Kelly, V. and L. Steinberg, The C R I T T E R system--analyzing digital circuits by propagating behaviors and specifications. Proc. 1982 Conf. American Association for Artificial Intelligence, pp. 284-289 (1982). Electronic circuit diagnostic expert systems--a survey 395 Kemeny, J. C., The report of the President's Commission on the accident at Three Mile Island. Kluwer, Washington D.C. (1979). Kemper, C. T., J. C. Bellows and P. J. Klinosky, Diagnostic apparatus. US Patent 4644479 (17 February 1987). Kidd, A. L., What do users ask?--Some thoughts on diagnostic advice. In Expert Systems. Warwick University (1985). Kignchi, T., H. Yoshida, H. Motoda and S. Kobayashi, The development of a plant operation guidance system utilizing knowledge engineering technique. J. atom. Energy Soc. Jpn 25, 298-305 (1983). Kim, J., CONVINCE: A CONversational INference Consolidation Engine. Ph.D. Dissertation, University of California, Los Angeles (1983). Kim, J. H. and J. Pearl, CONVINCE: A conversational inference consolidation engine. 1EEE Trans. Systems Man Cybernet. SMCIT, 120-132 (1982). King, J. J., Artificial intelligence techniques for device troubleshooting. Technical Report CSL-82-9, Hewlett-Pnckard (1982). Kitowski, J. and E. Ksiazek, Fuzzy logic applications for failure analysis and diagnosis of a primary circuit of the H T R nuclear power plant. Comput. Phys. Commun. 38, 323-327 (1985). de Kleer, J., Local methods for localizing faults in electronic circuits. MIT AI Lab. Mem. 394 (1976). de Kleer, J., The origin and resolution of ambiguities in casual arguments. Proc. 6th Int. Joint Conf. Artificial Intelligence (1979). de Kleer, J., An assumption-based truth maintenance system. Artificial Intell. 38(1) 11-23 (1986). de Kleer, J., A theory of plans for electronic circuits. MIT AI Lab. Working Paper 144 (1977) de Kleer, J., Causal and teleological reasoning in circuit recognition. MIT AI Lab. Technical Report 529 (1979). de Kleer, J., Extending the ATMS. Artificial Intell. 12(3), 12-34 (1979). de Kleer, J., Local methods for localization of faults in electronic circuits. MIT AI Lab. Memo 394 (1976). de Kleer, J. and J. S. Brown, Mental models of physical mechanisms and their acquisition. In Cognitive Skills and their Acquisition. Edbaum, New York (1980) de Kleer, J. and J. S. Brown, A qualitative physics based on confluences. Artificial Intell. 24, 7-84 (1984). de Kleer, J. and B. Williams, Reasoning about multiple faults. Proc. 5th Nam. Conf. Artificial Intelligence, Philadelphia, pp. 132-139 (1986). de Kleer, J. and G. J. Sussman, Propagation of constraints applied to circuit synthesis. Memo No. 485, MIT (1978). Kochan, S., N. Landis and D. Monson, Computer guided probing techniques. Proc. 1987 Int. Test Conf. San Francisco, CA (1987). Kolodner, J. L. and R. M. Kolodner, Using experience in clinical problem solving: introduction and framework. IEEE Trans. Systems Man and Cybernet. SMCI7, 420-431 (1987). Konolige, K., A deductive model of belief. Proc. 8th Int. Joint Conf. Artificial Intelligence, Karlsrube, West Germany, pp. 377-381 (1983). Konolige, K. and N. J. Nilsson, Multi-agent planning systems. Proc. AAAI-I, Stanford, Calif. pp. 138-142 (1980). Kuipers, B., Qualitative simulation as causal explanation. IEEE Trans. Systems Man Cybernet. SMC17, 432-444 (1987). Kuipers, B., Commonsense reasoning about causality: deriving behavior from structure. Artificial Intell. 24, 15-121 (1984). Kulikowski, C. A., Pattern recognition approach to medical diagnosis. IEEE Trans. Systems Sci. Cybernet. SSC6, 83-89 (1970). Kulikowski, C. A., Artificial intelligence methods and systems for medical consultation. IEEE Trans. Pattern Analysis med. lntell. 2, 464-476 (1980). Lai, K.-W., Functional testing of digital systems. Ph.D. Thesis, Carnegie-Mellon University Department of Computer Science, Technical Report CMU-CS-81-148 (1987). Lee, F. P., Process computer alarm and disturbance analysis: review of the state-of-the-art. Comput. chem. Engng 7, 669-694 (1983). Leitch, R., Towards intelligence control systems. Proc. Unicom Seminar on Expert Systems and Optimisation in Process Control, London, December (1985). Lenat, D. B. BEINGS: knowledge as interacting experts. Proc. 4th Int. Joint Conf. Artificial Intelligence, Tbilisi, U.S.S.R. pp. 126-133. (1975). Lin, J. C. et al., A deterministic methodology for application to a diagnosis oriented disturbance analysis system. Nucl. Engng Des. 72, 225-234 (1982). Lind, M., The use of flow models for plant diagnosis. In Human Detection and Diagnosis o f System Failures (Ed. J. Rasmussen and W. B. Rouse), pp. 411-431. Plenum Press, New York (1981). Lindsay, R. K., B. G. Buchanan, E. A. Feigenbaum and J. Lederberg Applications o f Artificial Intelligence for Organic Chemistry: The Dendral Project. McGraw-Hill, New York (1980). Lipkin, M. and J. D. Hardy, Mechanical correlation of data in differential diagnosis of hematological diseases. J. Am. med. Assoc. 166, 113-125 (1985). Long, W., A program for the management of heart failure. Sigart Newsletter No. 93 0985). Loveland, D. W., Selecting optimal test procedures from incomplete test set. Proc. 1st Int. Syrup. Policy Analysis and Information Science, Duke University, Durham, N.C., pp. 228-235 (1979). MacAloney, B. G., Fault spectrum: an analysis of system level test with proposed solutions. Proc. 1981 Int. Test Conf. (1981). Maclean, C. and P. Wilde, Knowledge-based electronic circuit diagnosis. Proc. Expert Systems 83, Technical Conf. BCS SGES, Cambridge, U.K. (1983). Magliero, A., R. Leong and R. Bethel, ADS-the IDSS adaptive diagnostic system. Proc. AUTOTESTCON '87, pp. 61-64 (1987). Magulre, J., AI-FERRET troubleshooting assistant. Advanced Products Laboratory, Support Systems, Hughes Aircraft Co., Long Beach, Calif. (1987). Martin-Soils, G. A. et al., Fault tree synthesis for design and real time applications. Trans. IChemE 60, 14-25 (1982). Masui, S., J. McDermott and A. Sobel, Decision making in time-critical situations. Proc. 8th Int. Joint Conf. Artificial Intelligence, Karlsruhe, West Germany (1983). Matsumoto, K., T. Sakagnchi and T. Wake, Fault diagnosis of a power system based on a description of the structure and function of the relay system. Expert Systems July, 2-3 (1985). 396 Y. LIROV McCown, P. and D. Lewy, APU M A I D - - a diagnostic expert system using heuristic and causal reasoning. Proc. A U T O T E S T C O N '87, pp. 371-375 (1987). McDermott, D., Circuit design as problem solving. In Artificial Intelligence and Pattern Recognition in Computer Aided Design (Ed. Latombe). North-Holland, Amsterdam (1978). McDermott, D., Non mototonic logic II: Non-Monotonic modal theories. Research Report No. 174 Department of Computer Science, Yale Univ. (1980). McDermott, D. V., Planning and acting. Cognitive Sci., 1, 17-54 (1978). McDermott, J., RI: a rule-based configurer of computer systems. Technical Report CMU-CS-80-119, Carnegie-Mellon University, Pittsburgh (1980). Merry, M., APEX 3: an expert system shell for fault diagnosis. GEC J. Res. 1, 39 (1983). Miller, R., H. Pople and J. Myers, INTRNIST- 1, an experimental computer-based diagnostic consultant for general internal medicine. N. Engl. J. Med. 307, 468-476 (1982). Milne, R., Strategies for diagnosis. IEEE Trans. Systems Man Cybernet. SMCI7, 333 (1987). Milne, R., Artificial intelligence applied to condition health monitoring. Chart. Mech. Engng May (1986). Milne, R., Diagnosing faults through responsibility. Proc. 1st Int. Conf. Applied Artificial Intelligence, Southampton, England (1986). Milne, R., Diagnosing faults through responsibility. A C M Annual Conf. Association o f Computing Machinery, Denver, Col., pp. 88-91 (1985). Milne, R., Fault diagnosis and expert systems. 6th Int. Workshop Expert Systems and their Applications, Avignon, France (1986). Milne, R., Fault diagnosis through responsibility. Proc. 9th Int. Joint Conf. Artificial Intelligence, Los Angelos, pp. 424-427 (1985). Milne, R., Reasoning about structure, behavior and function. Sigart Newsletter No. 93, pp. 4-59 (1985). Milne, R., Using AI in the testing of printed circuit boards. National Aerospace and Electronics Conf. Dayton, Ohio (1984). Milne, R. W., The theory of responsibilities. Sigart Newsletter No. 93 (1985). Mingers, J., Expert systems--experiments with rule induction. J. Op. Res. Soc. 37, 1031-1037 (1986). Minsky, M., A framework for representing knowledge. In The Psychology o f Computer Vision (Ed. P. Winston), pp. 211-217. McGraw-Hill, New York (1975). Mitchell, T., L. Steinberg, R. G. Smith, P. Schooley, V. Kelley and H. Jacobs, Representations for reasoning about digital circuits. Proc. 7th Int. Joint Conf. Artificial Intelligence, pp. 343-344 (1981). Mittal, S., Design of a distributed medical diagnosis and database system. Ph.D. Dissertation, Department of Computer and Information Science, Ohio State Univ. (1980). Mittal, S., B. Chandrasekaran and J. Smith., Overview of M D S - - a system for medical diagnosis. Proc. 3rd Symp. Computer Applications in Medical Care, IEEE, pp. 34-46 (1979). Moore, R., Expert systems in process control: applications experience. Ist Int. Conf. Applied Artificial Intelligence., Southampton, England (1986). Morris, R. and R. Nado, Representing actions with an assumption-based truth maintenance system. Proc. AAAI-86, Philadelphia, Penn. (1986). Mostow, D. J., A problem solver for making advice operational. Proc. AAAI-3, pp. 179-283 (1983). Mullis, R., Expert system for VLSI tester diagnostics. Proc. IEEE Test Conf. (1984). Nadig, H. J., Signature analysis---concepts, examples, and guidelines. Hewlett-Packard J. May, 15-21 (1977). Nau, D. and J. Reggia, Relationship between abductive and deductive inference. Proc. 1st Int. Workshop Expert Database Systems, pp. 500-509 (1984). Nawal, H., V. Lesser and E. Milies, Diagnosis using a formal theory of a signal processing system. IEEE Trans. Systems Man Cybernet. SMCI7, 369-379 (1987). Nelson, W. R., REACTOR: an expert system for diagnosis and treatment of nuclear reactor accidents. Proc. 2nd Conf. American Association for Artificial Intelligence, pp. 296-301 (1982). Nii, H. P. and N. Aiello, AGE (attempt to generalize): a knowledge-based program for building knowledge-based programs. Proc. 6th Int. Joint Conf. Artificial Intelligence, Tokyo, Japan (1979). Nilsson, N. J., Principles o f Artificial Intelligence. Tioga, Palo Alto, Calif. (1980). Patil, R., P. Szolvits and W. Schwartz, Causal understanding of patient illness in Medical diagnosis. Proc. 7th Int. Joint Conf. Artificial Intelligence, Vancouver, Canada (1981). Patrick, E. A. et al., Review of pattern recognition in medical diagnosis and consulting relative to a new system model. IEEE Trans. Systems Man. Cybernet. SMC6, 34-45 (1977). Pau, L. F., An adaptive signal classification procedure: application to aircraft engine monitoring. Pattern Recognition 9, 121-130 (1977). Pau, L. F., Applications of pattern recognition to failure analysis and diagnosis. In Human Detection and Diagnosis o f System Failures (Ed. J. Rasmussen and W. B. Rouse) pp. 429-446. Plenum Press, New York (1981). Pau, L. F., Failure Diagnosis and Performance Modelling. Dekker, New York (1981). Pau, L. F., Failure Diagnosis Systems. North-Holland, Amsterdam (1982). Pau, L. F., Failure diagnosis by an expert system and pattern classification. Pattern Recognition Lett. 2, 419-425 (1984). Pau, L. F., Integrated testing and algorithms for visual inspection of integrated circuits. IEEE Trans. Pattern Analysis Medicine Intell. PAMI5, 602-608 (1983). Pazzani, M. J., Failure driven learning of fault diagnosis heuristics. IEEE Trans. Systems Man Cybernet. SMC17, 380-394 (1987). Pearl, J., Fusion, propagation, and structuring in belief networks. Artificial Intell. 29, 241-288 (1986). Pearl, J., Heuristics: Intelligent Search Strategies for Computer Problem Solving. Addison Wesley, New York (1985). Peng, Y., A formalization of parsimonious covering and probabilistic reason in abductive diagnostic inference. Ph.D. Dissertation, Department of Computer Science Univ. Maryland, Baltimore (1986). Peng, Y. and J. Reggia, A probabilistic causal model for diagnostic problem-solving--part 1: integrating symbolic causal inference with numeric probabilistic inference. IEEE Trans. Systems Man Cybernetics. SMC17, 146-162 (1987). Peng, Y. and J. Reggia, A probabilistic causal model for diagnostic problem-solving, part two: diagnostic strategy. IEEE Trans. Systems Man Cybernet. SMCI7, 395-406. (1987). Electronic circuit diagnostic expert systems---a survey 397 Perkins, W. A. and T. J. Laffey, LES: a general expert system and its applications. SPIE Proc. 485, Applications o f Artificial Intelligence, pp. l 11-114 (1984). Pipitone, F., The FIS electronic troubleshooting system. Computer. July, 68-76 (1986). Poage, J. F. and E. J. McCluskey, Derivation of optimum tests for sequential machines. Proc. 5th A. Symp. Switching Circuit Theory Logic Designs, pp. 95-110 (1964). Pople, H., The formation of composite hypotheses in diagnostic problem solving: an exercise in synthetic reasoning. Proc. 5th Int. Joint Conf. Artificial Intelligence, Cambridge, Mass. (1977). Pople, H., On the mechanization of abductive logic. Proc. 3rd Int. Joint Conf. Artificial Intelligence, pp. 147-152 (1973). Putzolu, G. R. and J. P. Roth, A heuristic algorithm for testing of asynchronous circuits. IEEE Trans. Comput. C20, 639-664 (1971). Pynn, C., Strategies for Electronics Test. McGraw-Hill, New York (1987). Quinlan, J. R., Learning efficient classification procedures and their application to chess end game. In Machine Learning: An Artificial Intelligence Approach (Ed. R. Michalski, J. Carbonnell and T. Mitchell). Springer, New York (1984). Ramsey, J., Diagnosis using automatic test equipment and an artificial intelligence expert system. Masters Thesis, Air Force Institute of Technology (1984). Rasmussen, J. and A. Jense, Mental procedures in real-life tasks: a case study of electronic trouble shooting. Ergonomics 17, 293-307 (1974). Reggia, J., Abductive inference. Proc. Expert Systems in Government Symp., pp. 484-489 (1985). Reggia, J., Computer-assisted medical decision making. In Applications o f Computers in Medicine (Ed. M. Schwartz), pp. 198-213, IEEE Press (1982). Reggia, J. and D. Nau, An abductive non-monotonic logic. Proc. Workshop Non-Monotonic Reasoning, AAAI, pp. 385-395 (1984). Reggia, J. and S. Tuhrim, An overview of methods for computer-assisted medical decision making. In Computer-Assisted Medical Decision Making, Vol. I1 (Ed. J. Reggia and S. Tuhrim), pp. 3-45. Springer, New York (1985). Reggia, J., D. Nau, P. Wang and Y. Peng, A formal model of diagnostic inference. Infor. Sci. 37, 227-285 (1985). Reggia, J., D. Nau and P. Wang, Diagnostic expert systems based on a set covering model. Int. J. Man-Machine Stud. November, 437-460 (1983). Reggia, J., B. Perricone, D. Nau and Y. Peng, Answer justification in abductive expert systems for diagnostic problem solving. IEEE Trans. Biomed. Engng BE32, 263-272 (1985). Reiter, R., A theory of diagnosis from first principles. Department of Computer Science, Univ. Toronto, Toronto, Report ON TR-187/86 (1986). Richer, M. H. and W. J. Clancey, Guidon watch: a graphic interface for browsing and viewing a knowledge-based system. IEEE Comput. Graphics Applic. 5(11), 51-64 (1985). Rieger, C. and M. Grinbert, The declarative representation and procedural simulation of causality in physical mechanisms. Proc. 5th Int. Joint Conf. Artificial Intelligence, pp. 250-255 (1977). Robinson, J. A., A machine-oriented logic based on the resolution principle. J. Assoc. Comput. Machinery 12, 23-41 (1965). Rosenschein, S. J., Plan synthesis: a logical perspective. Proc. 7th Int. Joint Conf. Artificial Intelligence, Vancouver, Canada (1981). Roth, J. P., Computer Logic, Testing, and Verification. Computer Science Press, New York (1980). Roth, J. P., Diagnosis of automata failures: a calculus and a method. I B M J. Res. Dev. October, 278-281 (1966). Roth, J. P., W. G. Bouricius and P. R. Schneider, Programmed algorithm to compute tests to detect and distinguish faults in logic circuits. IEEE Trans. electronic Comput. ECI6 (1967). Roylance, G. L., Qualitative analysis of operational amplifier circuits. MIT Dept. of EF.&CS, MS Thesis proposal, March (1976). Rubin, A., The role of hypotheses in medical diagnosis. In Proc. 4th Int. Joint Conf. Artificial Intelligence, pp. 856--862 (1975). Ryan, P. and A. J. Wilkinson, Knowledge acquisition for ATE diagnosis. IEEE 1985 Proc. Int. Test Conf. pp. 848-856 (1985). Sacerdoti, E. D., A Structure for Plans and Behavior. Elsevier, North-Holland, Amsterdam (1977). Samuel, A., Some studies in machine learning using the game of checkers II. Recent progress. I B M J. Res. Dev. 11,601-617 (1967). Scarl, E. A., J. R. Jamieson and C. I. Delaune, Diagnosis and validation through knowledge of structure and function. IEEE Trans. Systems Man Cybernet. SMCI7, 360-368 (1987). Scarl, E. A., J. R. Jamieson and C. I. Delaune, Process monitoring and fault locations at the Kennedy Space Center. Newsletter No. 93 (1985). Schneider, P. R., On the necessity to examine D-chains in diagnostic test generation--an example. I B M J. Res. Dev. November, 114 (1967). Sembugamoorthy, V. and B. Chandrasekaran, Functional representation of devices and compilation of diagnostic problem solving systems. Cognitive Sci. 21, 54-67 (1985). Seshu, S., On an improved diagnosis program. IEEE Trans. electronic Comput. ECI2, 76-79 (1965). Seshu, S. and D. N. Freeman, The diagnosis of asynchronous sequential switching systems, lEE Trans. electronic Comput. ECll, 459-465 (1962). Sharer, G., A Mathematical Theory o f Evidence. Princeton Univ. Press (1976). Shirley, M. H., Digital test generation from hierarchical models and failure symptoms. Masters Thesis, MIT AI Lab (1983). Shirley, M. and R. Davis, Digital test generation from symptom information. IEEE VLSI Workshop (1983). Shortliffe, E., Computer-Based Medical Consultations: MYCIN. Elsevier, New York (1976). Shortliffe, E., M Y C I N - - a rule-based computer program for advising physicians regarding antimicrobial therapy selection. Stanford Univ. AI Lab., Memo AIM-251 (1974). Shortliffe, E. and B. Buchanan, A model of inexact reasoning in medicine. Math. Biosci. 23, 351-379 (1975). Simmons, F., C. Chee, S. Laughton, L. Wood, M. Purvis and D. Throop, Self explanatory device descriptions. Sigart Newsletter No. 93 (1985). Singhal, K. and J. Vlach, Symbolic analysis of analog and digital circuits. IEEE Trans. Circuits Systems. CAS24, 598-609 (1977). Slagle, J. R. and R. C. T. Lee, Application of game tree searching techniques to sequential pattern recognition. Commun. ACM, 142, 103-110 (1971). 398 Y. LIROV Stallman, R. M. and G. J. Sussman, Forward reasoning and dependency-directed backtracking in a system for computer-aided circuit analysis. Artificial Intell. 9, 135-196 (1977). Stallman, R. M. and G. J. Sussman, Problem solving about electrical circuits. In Artificial Intelligence: An M I T Perspective (Ed. P. H. Winston and R. H. Brown), pp. 33-91. MIT Press (1979). Stanfill, C., MACK, a program which deduces the behaviour of machines from their forms. Sigart Newsletter No. 93 (1985). Stefik, M. J., Planning with constraints. Artificial Intell. 16, 111-140 (1981). Stefik, M. J., Planning and meta-planning. Artificial lntell. 16, 141-169 (1981). Steinberg, L. and V. Kelly, The C R I T T E R system: analyzing digital circuits by propagating behaviors and specifications. Proc. 2nd National Conf. on Artif. Intelligence, Washington D.C., pp. 284-289 (1982). Sussman, G. J., A computational model of skill acquisition. MIT AI Lab Technical Report 297 (1973). Sussman, G. J., Electrical design--a problem for artificial intelligence research. MIT AI Memo 425 (1977). Sussman, G. J., SLICES at the boundary between analysis and synthesis. In Artificial Intelligence and Pattern Recognition in Computer Aided Design (Ed. B. Latombe). North-Holland, Amsterdam (1978). Sussman, G. J. and R. M. Stallman, Heuristic technique in computer-aided circuit analysis. IEEE Trans. Circuit Systems CAS22(11), 24 39 (1975). Swabey, M., Use of deep and shallow knowledge in electronics diagnosis. Proc. Unicom Seminar on Expert Systems and Optimization in Process Control, London (1985). Swanson, D. B. Computer simulation of expert problem solving in medical diagnosis. Ph.D. Dissertation, Univ. Minnesota (1978). Swartout, W. R., Explaining and justifying expert consulting programs. Proc. 7th Int. Joint Conf. Artificial Intelligence, pp. 815~22 (1981). Szolovits, P. and S. Pauker, Categorical and probabilistic reasoning in medical diagnosis. Artificial lntell. 2, 115-144 (1978). Szygenda, S. A. and E. W. Thompson, Modeling and digital simulation for design verification diagnosis. IEEE Trans. Comput. C25, 1242-1253 (1976). Taie, M. R. and S. Srihari, Modeling connections for circuit diagnosis. Proc. 3rd Conf. A1 Applications, pp. 81-86 (1987). Taie, M. R., S. N. Srihari, J. Geller and S. C. Shapiro, Device representation using instantiation rules and structureal templates. Proc. Canadian A I Conf-86, Presses de l'Universit6 du Qu6bec, Montr6al, pp. 124-128 (1986). Taie M. R., J. Geller, S. N. Srihari and S. C. Shapiro, Knowledge based modelling of circuit boards. Reliability and Maintenance Syrup. Proc., pp. 422-427 (1987). Tanaka, T., Representation and analysis of electrical circuits in a deductive system. Proc. 8th Int. Joint. Conf. Artificial Intelligence, West Germany, pp. 263-267 (1983). Tanaka, T., Simulation of thinking process on electronic circuits in natural language I, I1, III. Technology Report of Kyushu Univ., Vol. 44, p. 45 (1971). Tanaka, T., Structural analysis of electrical circuits in a deductive system. In Computer Expert Systems (Ed. L. Bolc). Springer, New York (1985). Tendolkar, N. N. and R. L. Swann, Automated diagnostic methodology for the IBM 3081 processor complex. I B M J. Res. and Dev. 26, 78-88 (t982). Thompson, T. F. and R. M. Wojcik, Methods and apparatus for system fault diagnosis and control. U.S. Patent 4649515 (10 March 1987). Towill, D. R., Automatic testing of control systems---past present and future. J. Inst. electronic Radio Engrs 57(2), 67-80 (1987). Tsiang, S. H. and W. Ulrich, Automatic trouble diagnosis o f complex logic circuits. Bell Syst. Tech. J. 41(4), 17-35 (1962). Tulloss, R. E., Fault dictionary compression: recognizing when a fault may be unambiguously represented by a single failure detection. Proc. 1980 Test Conf. (1980). Tulloss, R. E. et al., The diagnostic organization and retrieval-algorithm system. The Engineer. Summer (1982). Tversky, A. and D. Kahneman, Causal schemata in judgements under uncertainty. In Progress in Social Psychology (Ed. M. Fishbein). Erlbaum, Hillsdale, N.J. (1979). Ulrich, E. G., T. Baker and L. R. Williams, Fault test analysis techniques based on simulation. Proc. 9th Design Automation Workshop, pp. 111-115 (1972). Van Melle, W., A domain independent production-rule system for consultation programs. Proc. 6th Int. Joint Conf. Artificial Intelligence, pp. 923-925 (1979). Vere, S., Planning in time: windows and durations for activities and goals. IEEE Trans. Pattern Analysis Machine Intell. PAMI5, 246-267 (1983). Wang, D. T., An algorithm for the detection of test sets for combinational logic networks. IEEE Trans. Comput. C25, 742-746 (1975). Warren, D. H. D., Generating conditional plans and programs. Proc. A I S B Summer Conf., University of Edinburgh, U.K., pp. 344-354 (1976). Washio, T., M. Kitamura, K. Kotajima and K. Sugiyama, Semantic network approach to automated failure diagnosis in nuclear power plant. Proc. 1985 ANS Int. Topical Meet. Computer Applications for Nuclear Power Plant Operation and Control, Pasco, Wash. (1985). Weiss, S., C. A. Kulikowski, S. Amarel and A. Safir, A model-based method for computer-aided medical decision-making. Artificial lntell. 11, 145-172 (1978). Weiss, S. M. and C. A. Kulikowski, EXPERT: a system for developing consultation models. Proc. 6th lnt. Joint Conf. Artificial Intelligence, Tokyo, Japan, pp. 942-947 (1979). White, B. Y. and J. R. Fredericksen, QUEST: qualitative understanding o f electrical systems troubleshooting. Sigart Newsletter No. 93 (1985). Wilkinson, A. J., Benchmarking an expert system for electronic diagnosis. Proc. A U T O T E S T C O N ' 8 7 , pp. 964-971 (1986). Wilkinson, A. J., A method for test system diagnostics based on the principles of artificial intelligence. Int. Test Conf., 1984, Proc. IEEE, pp. 188-195 (1984). Wilkinson, A. J., MIND: an inside look at an expert system for electronic diagnosis. Design Test Mag. August, 69-77 (1985). Williams, T. L., Isolating multiple faults in NDS. Technical Report 010, Smart Systems Technology, Falls Church, Va. (1982). Williams, T. W. and K. P. Parker, Design for testability--A survey. I E E E Trans. Comput. C31, 9-27 (1982). Winston, P., Artificial Intelligence. Addison-Wesley, New York (1984). Yau, C. W., Private communication (1986). Yau, S. S. and S. C. Yang, Multiple fault detection for combinational logic circuits. IEEE Trans. Comput. C24, 233-242 (1975). 
work_auqawdjzq5cqfbtap7kr6ihyte	----	PII: S0957-4174(96)00049-8 P e r g a m o n Expert Systems With Applications, Vol. 11, No. 3, pp. 343-349, 1996 Copyright © 1996 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0957-4174/96 $15.00+0.00 PII: S0957-4174(96)00049-8 Using Expert Systems as a Training Tool in the Agriculture Sector in Egypt A . RAFEA Computer Science Department, Institute of Statistical Studies and Research, Cairo University, 5 Tharwat St, Orman, Giza, Egypt K . SHAALAN Computer Science Department, Institute of Statistical Studies and Research, Cairo University, 5 Tharwat St, Orman, Giza, Egypt Abstract--This paper describes the Egyptian experience in using Expert Systems (ES) as a training tool in the agriculture sector. The work described here is part of an ongoing research to study the use orES in human resources development. In particular, we present the use of such a tool as an instructional device f o r increasing the efficiency of extension workers through improving their general decision- making skills in their jobs. To clarify this process, we conducted an experiment and analyzed its results. Copyright © 1996 Elsevier Science Ltd 1. I N T R O D U C T I O N KNOWLEDGE-BASED SYSTEMS (KBS) are computer pro- grams that incorporate heuristic knowledge and emphasize declarative knowledge over procedural prob- lem solving. KBS can be used as a powerful training tool. In general, the goal of training is to produce a motivated user who has the basic skills needed to apply what has been learned and then to continue to learn on the job (Compeau et al., 1995). Two features make KBS an excellent training tool for personnel whose mission is providing advice. The first feature is that KBS incorpo- rate experienced-based knowledge derived from different sources o f a certain domain, e.g. human experts, and is now structured and provided in a very portable and easily accessible medium. For example, by using an expert system for crop management, the crop consultant is forced to go through the entire reasoning process in a systematic manner ensuring consideration o f all factors affecting the decision. Another feature of the KBS is the explanation facility which is inherently an educational tool. Explanation facilities provide for reasoning which is important as a training tool for new personnel, e.g. new extension personnel that are new to certain crop. Literature that sheds light on using KBS in training programs is beginning to emerge. An empirical study o f the use o f ES in U.S.A. business schools and its implications for industry is present ed in (Teer et al., 1994). Several cases and benefits of using KBS as an instructional device i n MBA programs is surveyed in (Dologite, 1991). An interesting study by Mockler (1990) is the use o f KBS for teaching and performing KBS development. This development is to guide both technical and non-technical managers in finding, defining, selecting, evaluating an area, decision or task for potential KBS development. Moreover, an e~pert system can be used to advise managers on selecting employees for training, as in Ntuen et al. (1995), which is time consuming and belongs to a special class o f multiattribute decision making. The study described in this article addresses the use o f ES as a training tool for increasing the efficiency o f extension workers through improving their general decision-making skills in their jobs. The next section briefly presents the needs for ES's technology in the agriculture sector. Then the focus turns to a description o f the agriculture extension service environment. The expert systems involved in this training are then described briefly. It is followed by a description o f the experiment conducted during the training o f extension workers that is the concern o f this study. Next, we discuss the outcome o f applying the experiment and present its results. In a concluding section, we present some final remarks. 2. THE NEEDS F O R EXPERT SYSTEMS T E C H N O L O G Y IN THE A G R I C U L T U R E S E C T O R Agriculture production has evolved into a complex process requiring the accumulation and integration o f knowledge and information from many diverse sources 343 344 A. Rafea and K. Shaalan including marketing, horticulture, insect management, disease management, weed management, accounting and tax laws. Expert systems are tools for agriculture management since they can provide the site-specific integrated and interpreted advice that farmers and consultants need to more efficiently handle management concerns (Rafea, 1995). This section decribes the need for ES as a tool for decision support and as a tool for training. 2.1. The Need for Expert Systems in Decision Support The development o f an agriculture expert system requires the combined efforts o f specialists from many fields of agriculture, and must be developed with the cooperation o f the farmers and extension officers who will use them (Broner et al., 1992). A recent study o f the needs assessment for expert systems in the agriculture sector in Egypt (ESICM, 1994) suggested that the sector needs to use the ES technology to improve the quality o f the products and increase the efficiency of the agri- cultural system. Expert systems are recognized as an appropriate technology because they address the problem o f transfer- ring knowledge and expertise from highly qualified specialists to less knowledgeable personnel. In agri- cuture, this transfer is always taking place from research to extension, from extension to farmers, and even from farmer to farmer. Expert systems present excellent tools for relieving the increasing pressure on the limited expertise available in developing nations. It must be recognized that knowledge, the very foundation o f expertise, is a scarce resource in developing nations. Expert systems can help expand this vital resource by making available, in specific situations, vital knowledge that increases the effectiveness o f less experienced personnel. Expert Systems can be used by decision makers at different levels: operation level and planning level. On the operation level, the extension worker in the village, district, and/or govemorate can use the system to support him in making his decision in giving the appropriate advice to growers. On the planning level, the decision makers can use the expert system for predictions, such as on the needs for water, fertilizers and pesticides for a particular crop in the region given the area cultivated with such a crop. This generated information is very important for different users: the traders, the exporters, the importers o f these materials. Another type o f application is the estimation o f the yield given a simulation model linked with the expert system. The prediction o f yield can serve the decision makers in deciding the amount to be imported in advance, if any, and hence take necessary actions. 2.2. The Need for Expert Systems in Training Although the goal for developing agricultural expert systems in Egypt has to be used as a decision support tool for the extension workers, practical training o f extensionists on the developed ES has revealed that ES can be used for expediting the training o f extension workers. In the near future, it is not expected to install computers in the 4000 villages in Egypt. Installing computers in the 200 district offices, however, is an attainable goal. Therefore, i f these 200 centers could be used to train the extension workers at the village level using the ESs installed at these centers, there will be a tremendous impact. Traditional ways o f training are not sufficient to cope with the fast growing technologies in the different agriculture specialties for the different crops. Using ESs will reduce training time and enhance its quality. 3. AGRICULTURE EXTENSION SERVICES ENVIRONMENT Agriculture development in Egypt depends on the connection between the three sides o f the extension process (CLAES, 1993): (1) research, (2) extension and (3) farmers. The reporting o f problems, and finding solutions to them are the main concern o f the cooperative extension programs. Through the different stages o f technology develop- ment and information transfer to farmers, the extension sector works with the research component to narrow the gap between research results and this application in the field. Extension engineers help in studying the production situation as they can identify farmers' problems through watching the farmers and working with them to diagnose problems and attempt to find the solutions. The Ministry o f Agriculture and Land Reclamation in Egypt is concerned with all activities in agriculture development, and gives special attention to the coopera- tion o f research and extension in order to facilitate continuous training for all people concerned. This is done to spread appropriate agriculture technology all over th e country. 4. A BRIEF DESCRIPTION OF THE EXPERT SYSTEMS USED IN TRAINING The expert systems being used are mainly for crop management which are developed by the Central Labo- ratory for Agriculture Expert System (CLAES) at the Agriculture Research Center o f Ministry o f Agriculture and Land Reclamation in Egypt. They are the Cucumber Expert System (CUPTEX) and the Citrus Expert System (CITEX). CUPTEX (E1-Dessouki et al., 1993; Rafea et al., 1995) is an expert system for cucumber production management in a plastic tunnel. CITEX (Salah et al., 1993) is an expert system for citrus production in open Expert Systems as a Training Tool in Agriculture 345 fields. Both the two expert systems were modeled using the KADS methodology (Schreiber et al., 1993; Wielinga et al., 1991). A laboratory prototype was implemented using the NEXPERT Object shell (Neuron Data Inc., 1991). Currently, they have been transferred to a knowledge representation language based on object- oriented and logic programming paradigms (ESICM, 1992). These expert systems are intended to be used by the agricultural extension service within the Egyptian Ministry o f Agriculture and by the private sector. The following are components or subsystems of the two expert systems: irrigation, fertilization, verification and treatment. The main objective of the irrigation and fertilization subsystems is to generate schedules, that include the water quantity, irrigation interval, nutrient quantity and application interval. These outputs are based on quantita- tive reasoning rather than heuristic reasoning. The objective o f the verification subsystem is to confirm the user suggestion o f particular disorder according to the symptoms provided by the user. The objective o f the treatment subsystem is to recommend treatment opera- tion according to a case description. 5. EXPERIMENT DESCRIPTION This section describes the experiment conducted during the training of extension workers at the project premises for CUPTEX and CITEX at CLAES. CLAES provided an excellent research site for this study. To date, 749 man/days training were realized. This experience pro- vides a useful vehicle for evaluating t h e effect of using ES as a training tool to increase the decision-making skills o f extension workers. Concerning this study, 11 extension workers who specialized in protected cultiva- tion were involved in the evaluation using CUPTEX and 8 extension workers who specialized in horticulture were involved in the evaluation using CITEX. The objective o f conducting this experiment was two-fold: first, to measure the effect o f using expert systems on the performance o f the extension workers and second to assess the decision-making skills o f the extension workers compared with decisions generated by the expert systems. The methodology followed to achieve this objective is presented in the first subsection, whereas, its application is given in the second subsec- tion. 5.1. The Methodology The proposed methodology is based on tests conducted during one training cycle of competent extension work- ers to measure the effect o f using expert systems on their performance, and to assess their decision-making skills compared to decisions generated by the expert systems. It can be summarized in the following steps: (1) Design forms to document the cases and their results. (2) Prepare cases which are described b y a set o f input data. (3) Distribute these cases to the trainees to give their decisions before using the expert system. (4) Train the extension workers on the expert system. (5) Distribute again the same cases without their previous decisions and ask them to give their decisions again. (6) Evaluate the cases before and after training TABLE 1 Irrigation Results Average score Average score Percentage of before (%) after (%) Enhancement enhancement Water qty 38.1 73.6 35.5 93.18 Interval 41.9 71.2 29.3 69.93 Average 40.0 72.4 32.4 81.00 TABLE 2 Fertilization Subsystem Results Average score Average score Percentage of before (%) after (%) Enhancement enhancement Nitrogen 44.8 64.4 19.6 43,75 Phosphorus 36.3 50.9 14.6 40,22 Potassium 40.1 60.6 20.5 51.12 Magnesium 7.0 62.3 55.3 790 Manure 0.0 91.8 91.8 infinity Average 25.64 66.0 40.36 157.41 346 A. Rafea and K. Shaalan TABLE 3 Verification Subsystem Results Average score Average score Percentage of before (%) after (%) E n h a n c e m e n t enhancement Symptoms on leaves 16.5 Symptoms on stem 28.8 Symptoms on root 40.3 Symptoms on fruits 34.0 Average 29.9 38.9 22.4 135.76 46.5 17.7 61.46 87.5 47.2 117.12 36.0 2.0 5.88 52.23 22.33 80.06 taking the expert system results as a reference (the result o f the expert system for these cases had been verified with domain experts before conducting the experiment). The following formula is used to compute the percentage (%) o f the enhancements: Enhancement Average Score before using the ES × 100 where the E n h a n c e m e n t is the difference between the average score before and after using the ES. 5.2. The Application of the Methodology The methodology was applied as follows: (1) Forms were designed for the different sub- systems o f C U P T E X and CITEX, namely, irrigation, fertilization, verification and treat- ment. In effect, the irrigation and fertilization were grouped in one f o r m whereas the ver- ification and treatment were grouped in another form. (2) Sets o f cases covering the different aspects o f the developed expert systems were prepared in forms. Each set consisted o f approximately 20 cases. (3) Each trainee was given around 10 cases before conducting the training and was asked to give his decisions for these cases. The decision is either irrigation schedule, fertilization sched- ule, s y m p t o m s to be observed i f a disorder is suspected, or a treatment schedule. It should be noted that some o f the trainees have the same cases while some others m a y have different cases. (4) The training was conducted by letting the trainees run the expert system, providing the inputs in the cases and observing the outputs o f the expert system. During training, each trainee was given all the cases, and other cases he/she created were also run on the system. (5) The same cases (cases before training) were given to each trainee after conducting the Iraining. H e / s h e was not told that they were the same cases nor had he/she access to the forms completed before training. (6) The cases were evaluated in the following way: • The irrigation subsystem was evaluated taking into account the water quantity and irrigation interval. • The fertilization subsystem was evaluated taking into account the quantifies o f nitro- gen, phosphorus, potassium, magnesium and manure. • The verification subsystem was evaluated taking into account symptoms on root, stem, leaves and fruits. • The treatment subsystem was evaluated taking into account the treatment materials and their corresponding doses. In all cases the forms before and after training were analyzed taking results produced b y the expert system as a reference. I f the decision given by the trainee matches the expert system result, the trainee is given full marks. I f the decision given by the trainee mismatches the TABLE 4 Treatment Subsystem Result Average score Average score Percentage of before (%) after (%) Enhancement enhancement Material (1) 41.8 72.4 30.6 73.21 Dose (1) 25.8 64.8 39.0 151.16 Material (2) 24.7 36.5 11.8 47.77 Dose (2) 10.5 20.0 9.5 90.48 Average 25.7 48.43 22.73 90.66 Expert Systems as a Training Tool in Agriculture 347 TABLE 5 Irrigation Results Average score Average score Percentage of before (%) after (%) Enhancement enhancement Water qty 31.3 67.5 36.2 115.65 Interval 39.7 60.6 20.9 52.64 Average 35.5 64.05 28.55 84.15 expert system result, the trainee is given zero. In cases where the decision is a quantity or a dose, the trainee is given a score relative to the quantity or dose produced b y the expert system. In other cases, a score was estimated b y a domain expert who was responsible for this evaluation. 6. R E S U L T S O F A P P L Y I N G O U R T R A I N I N G M E T H O D O L O G Y T O T H E A G R I C U L T U R E S E C T O R I N E G Y P T Evaluators have used all the components o f the above- mentioned methodology in their effort to measure the effect o f using expert systems on extension personnel performance. The outcome o f applying all components o f the methodology to C U P T E X and C I T E X is discussed below. The verification results are summarized in Table 3. As can be seen, the average enhancement o f 80.06% was noticed after using the expert system for one-day training. The most remarkable enhancement was related to the symptoms on leaves, whereas the s y m p t o m s on fruits has slightly increased. The treatment subsystem results are summarized in Table 4. As can b e seen f r o m the table, an average enhancement of 90.66% was noticed after using the expert system. A remarkable enhancement was noticed in determining the dose for material (1). It is also worth noting that the trainees were not aware o f the second material application nor its dose (the lowest score before and after the training). 6.2. CITEX Results and Discussion 6.1. CUPTEX Results and Discussion The irrigation results are summarized in Table 1. As can be seen, an average increase o f 81% has been achieved in the decision taken b y the trainees after using the expert system. The enhancement in deciding the water quality is m u c h better than the enhancement in deciding the irrigation interval. The fertilization results are summarized in Table 2. It is worth noting that manure was not included b y any o f the trainees although it is well known that manure should be added before cultivation. They might have assumed that this is a well-known fact. So, they did not put it. The second important remark is that they did not know much about magnesium; the percentage o f enhancement regarding m a g n e s i u m was 790%. Eliminating these two odd cases, an average enhancement o f 45.12% can b e observed in three fertilizers, namely, nitrogen, phospho- rus and potassium. The irrigation results are summarized in Table 5. As can be seen, an average increase o f 84.15% has been achieved in the decision taken b y the trainees after using the expert system. The enhancement in deciding the water quantity is much better than the enhancement in deciding the irrigation interval. The fertilization results are summarized in Table 6. An average enhancement o f 35.89% can be observed after using the expert system. It is worth noting that the least enhancement was in adding manure, only 11.65%. This m a y b e due to the fact that manure application is done once a year during preparation. The verification results are summarized in Table 7. As can be seen, the average enhancement o f 1464.08% was noticed after using the expert system for one-day training. However, the scores o f zeros, before using the expert system, in recognizing the symptoms are very odd. But as can be seen, this recognition reached 90% in the case o f symptoms on roots after using the expert TABLE 6 Fertilization Subsystem Results Average score Average score Percentage of before (%) after (%) Enhancement enhancement Nitrogen 43.1 63.6 20.5 47.56 Phosphorus 46.4 66.9 20.5 44.18 Potassium 38.1 53.4 15.3 40.16 Manure 78.1 87.2 9.1 11.65 Average 51.43 67.78 16.35 35.89 348 A. Rafea and K. Shaalan TABLE 7 Verification Subsystem Results Average score Average score Percentage of before (%) after (%) Enhancement enhancement Symptoms on leaves 14.2 Symptoms on stem 0.0 Symptoms on root 0.0 Symptoms on fruits 0.0 Average 3.55 65.5 51.3 361.27 33.3 33.3 infinity 90.0 90.0 infinity 33.3 33.3 infinity 55.53 51.96 1464.08 system. So, even i f the first row o f Table 7 which relates to symptoms on leaves is considered, the enhancement is 361.27% which is still very high. The treatment subsystem results are summarized in Table 8. As can be seen f r o m the table, an average enhancement o f 732.61% was noticed after using the expert system. A remarkable enhancement was noticed in determining the dose. The improvement in determining the material is also very high. 7. C O N C L U S I O N S The results o f this study suggest that the expert system can be an effective training tool in agriculture extension programs. The experiment showed that there is enhance- ment in the performance o f the extension workers after the usage o f the expert systems which developed after a very short time. Although the expert systems developed are thoroughly verified with domain experts, there is always r o o m for further improvement. Regardless o f the quantitative measures which m a y be arguable regarding how these values are calculated, the trend o f the trainees was to accept the advice given b y the system, which is why their measured performances were increasingly taking the decision o f the expert system as a reference. Table 9 presents a summary o f the results for both CUPTEX and CITEX. As can be seen from Table 9 , the average score percentages for trainees on the four subsystems o f C U P T E X and C I T E X were approximately in the same range (30.31 and 27.13). But i f we look at each individual subsystem, we can find a tangible difference between the verification and treatment sub- systems in the favor o f the C U P T E X trainees, while there is a tangible difference in the fertilization subsystem in the favor o f C I T E X trainees. This remark is related to the percentage o f enhancement, we can see that the best enhancement for CUPTEX was in the fertilization subsystem, whereas the best enhancements for C I T E X were in the verification and treatment subsystems. The average enhancement for C I T E X was approximately six times the average enhancement for CUPTEX. This is because the performance o f the C I T E X trainees in the TABLE 8 Treatment Subsystem Result Average score Average score Percentage of before (%) after (%) E n h a n c e m e n t enhancement Material 11.0 63.9 52.9 480.91 Dose 5.1 55.3 50.2 984.31 Average 8.05 59.6 51.55 732.61 TABLE 9 Results Summary CUPTEX CITEX CUPTEX CITEX (average (average (% of (% of score %) score %) enhancement) enhancement) Before After B e f o r e After Irrigation 40.0 72.4 35.5 64.05 81.0 84.15 Fertilization 25.64 66.0 5 1 . 4 3 67.78 157.41 35.89 Verifiction 29.9 52.23 3.55 5 5 . 5 3 80.06 1464.08 Treatment 25.7 48.43 8.05 59.60 90.66 734.61 Average 30.31 59.77 27.13 61.74 102.28 579.18 Expert Systems as a Training Tool in Agriculture 349 v e r i f i c a t i o n a n d t r e a t m e n t s u b s y s t e m s w a s v e r y l o w (3.55 a n d 8.05), a n d c o n s e q u e n t l y t h e i r p e r f o r m a n c e h a s i n c r e a s e d d r a m a t i c a l l y a f t e r u s i n g t h e s y s t e m ( p e r c e n t - a g e s o f e n h a n c e m e n t a r e 1 4 6 4 . 0 8 a n d 7 3 2 . 6 1 ) . Acknowledgements--The authors would like to acknowledge the Expert System for Improved Crop Management Project funded by UNDP, MOA, and executed by FAO for their support while conducting the experiment described in this paper. R E F E R E N C E S Broner, I., Parente, A. & Tompson, K. (1992). Knowledge-based systems for crop management. Computer in agricultural extension programs. Proceedings of the 4th International Conference. Orlando, FL: American Society of Agriculture Engineers. CLAES (1993). Research and agriculture extension system in the Ministry of Agriculture & Land reclamation. Technical Report No. TR/CLAES/CRG[ll]/2. Central Laboratory for Agricultural Expert Systems, NARP Collaborative Research. Compeau, D., Olfman, L., Sein, M. & Webster, J. (Guest Eds) (1995). End-user training and learning: Introduction. Commun. ACM, 38(7), 24-26. Dologite, D. (1991). Developing knowledge-based systems: A survey of its impact on graduate MBAs. Expert Systems With Applications, 2(2), 175-186. E1-Dessouki, A., Edress, S. & E1-Azhary, S. (1993). CUPTEX: An integrated expert system for crop management of cucumber. Proceedings of the Second National Expert Systems and Develop- ment Workshop (ESADW-93). Ministry of Agriculture and Land Reclamation, Cairo, Egypt. ESICM (1992). Specifications of object-oriented language on top of Prolog: KROL. Technical Report No. TR-88-024-25 Expert Sys- tems for Improved Crop Management, UNDP/FAO, EGY/88/024. ESICM (1994). A study of the needs assessment for expert systems in the agriculture sector in Egypt. Technical Report No. TR-88-024-33 Expert Systems for Improved Crop Management, UNDP/FAO, EGY/88/024. Mockler, R. (1990). Using knowledge-based systems for teaching and performing knowledge-based systems development. Expert Systems With Applications, 1(2), 103-116. Neuron Data Inc. (1991). NEXPERT object reference manual, PC Version 2.0. California. Ntuen, C. & Chestnut, J. (1995). An expert system for selecting manufacturing workers for training. Expert Systems With Applica- tions, 9(3), 309-332. Rafea, A. (1995). Expert system as a tool for information technology in agriculture. The International Informatics Access '95, Singapore. Rafea, A., El-Azhari, S., Ibrahim, I., Soliman, E. & Mahmoud, M. (1995). Experience with the development and deployment of expert systems in agriculture. The Proceedings of IAAI-95, Montreal, Canada. Salah, A., Hassan, H., Tawfik, K., Ibrahim, I. & Farahat, H. (1993). CITEX: An expert system for citrus crop management. Proceedings of the Second National Expert Systems and Development Workshop (ESADW-93), Ministry of Agriculture and Land Reclamation, Cairo, Egypt. Schreiber, G., Wielinga, B. & Breuker, J. (Eds) (1993). KADS: A principled approach to knowledge-based system development. San Diego, CA: Academic Press. Teer, E & Teer, H. (1994). The trend in expert systems coverage in business schools and implications for people in industry, Inter- national Journal of Applied Expert Systems, 2, 191-195. Wielinga, B. & Breuker, J. (1991) KADS: A model approach to knowledge engineering. Esprit Project 5248 KADS II. 
work_avjpj6uqijhr5ncpqosqjrbd3a	----	Chapter-6 (Sania Bhatti-MA Hashmani-Naheed Shaikh).pmd The Specification of an Expert System for Building Bylaws Compliance SANIA BHATTI*, MANZOOR AHMED HASHMANI**, AND NAHEED SHAIKH*** RECEIVED ON 25.05.201 ACCEPTED ON 01.12.2011 ABSTRACT An Expert System is a computer program that simulates the human intelligence and behaviour in specific and limited domains. It is used to solve problems with tricks, shortcuts and heuristics i.e. rules of thumb. Checking a Plan (Map) to verify its compliance with building bylaws is a complex task mainly due to various rules and the exceptions to those rules. Humans are prone to make errors in such situations. Due to the problems faced by Building Control Department, HDA ( Hyderabad Development Authority) there is a strong need to develop a computerized system. In this research we have developed a prototype named as ESBBC (Expert System for Building Bylaws Compliance) for HDA that can help in their building plan checking system. The proposed solution is merging three frameworks, i.e. Java an OOP (Object Oriented Programming) language, Prolog- a rule based language and MS Access- for database. The solution is fulfilling the three main requirements of the HDA, i.e. (1) Determination of whether a particular plan is in compliance with predefined building bylaws or not. (2) Offering search facility. (3) Maintaining records of plans which are entered for compliance checking. We have checked plans of 20 properties according to HDA building regulations using ESBBC and presented their results. The results show that ESBBC has capability to identify errors made by humans. Key Words: Expert System, Plan, Bylaws, Java, Prolog. * Assistant Professor, Department of Software Engineering, Mehran University of Engineering & Technology, Jamshoro. * * Professor, Iqra University Main Campus, Karachi. * * * Assistant Professor, Department of Architecture, Mehran University of Engineering & Technology, Jamshoro. 1. INTRODUCTION based on different parameters like locality, plot size etc. After development these plans are submitted to Building Control authorities for approval. The Building Control authorities in Pakistan are still using the manual system for plan checking. There are many shortcomings of this system such as, many persons are involved, many Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 4 3 When an architect develops design of abuilding called "Plan" then he has to followcertain rules, which are pre defined by Building Control authorities. Preparation of certain rules and regulations for controlling development of areas is known as "Building bylaws". These rules are defined Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 4 4 The Specification of an Expert System for Building Bylaws Compliance corrective steps are involved in case of any mistake, more prone to errors, more time consumption in record maintenance and searching previous records. To overcome these problems, there is a need to develop computerized system which can solve the problems happening due to human errors. In this work we have developed ESBBC in which at the back end rule based inference engine is working which is based on knowledge base of rules. Modelling of rules as data rather than as code is one the challenges addressed in this work. Another challenge is the provision of hybrid approach which combines the strengths of expert system and object oriented programming language. The user interface developed in object oriented programming language allows control of rule based engine and data base. ESBBC is based on building regulations for zonal plan HDA scheme 1, 2 and 3. The dimensions of maps are entered as parameters and are checked by the rule based inference engine. In case of violation of any rule, appropriate messages are given by the program. The interface of ESBBC is flexible to take inputs. It is also maintaining the records of plans being entered and is able to search the particular plan. This section presents an overview of the existing system, its short comings and why a computerized system is needed. Section 2 discusses rule based system and its implementation in different applications. Section 3 states the proposed solution, its prerequisites and advantages. In Section 4, the output of the proposed solution is discussed along with comparison with existing system to show its reliability. In Section 5 conclusion is presented. 1.1 Analysis of Existing System HDA [1] was created with the enactment of Sindh Act XIII in the year 1976 with its preamble for development, improvement and beautification of urban areas of Hyderabad division. However, the working area of HDA was limited to only 200 square miles around Hyderabad city only, including a portion of Dadu District out of Hyderabad Division. Following steps are taken by HDA for authorization of residential plans: (i) Building plans in prescribed file containing Original Tracing, 4 copies (Ammonia) along with allied documents is submitted at reception. (ii) Issuance of advance scrutiny challan by bill clerk to the department is done. (iii) Then challan is deposited in a bank. (iv) After payment of scrutiny fees to the building clerk, file is submitted to registrar of concerned ward/Area. (v) File is sent to draughtsman for report. (vi) Sub Inspector building visits site for verification and does scrutiny of building plan. (vii) Then Building plan is passed to assistant director for approval. (viii) Further permission letter of building plan is prepared and send to the sub inspector and assistant director for signature. (ix) Finally building plan is released. 1.2 Results of Unnecessary Delay in Approval The permission to occupy the building is granted only after the verification of the fact that it is constructed as per detail of the approved plan. Procedures laid out by the authority should be easy, simple and should not lead to unnecessary delay in granting the permission for the construction of building. The unnecessary delay in approval process may lead to the following undesirable consequences: (i) Increase in housing shortage that will result in increase of house rents. Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 4 5 The Specification of an Expert System for Building Bylaws Compliance (ii) Formation of more slums that will result in more financial burden on the local authority for their removal or improvement. (iii) The enthusiasm of the owner who wants to construct the building quickly may slow down because of changes of circumstances with the passage of time. (iv) The local authority may also suffer loss in revenue from various taxes such as water tax, sanitation tax, betterment charges etc. (v) The open space remains idle for long time resulting in under hardship and loss of revenue to the landowner. (vi) Considerable depression in construction industry as a whole. 1.3 Significance of Developing a Computerized System Due to the problems faced by building control authorities there is firm need to develop a computerized system to check building bylaws. The following benefits can be obtained from a computerized system: (i) It will take less time to check a building plan as compared to human experts. (ii) Instead of involving many experts in checking process only one person will be sufficient. (iii) Proper checking can be done and in case of any mistake proper guidance will be given by the system. (iv) It will be simple to maintain record of all plans. (v) Searching of a particular record will become straightforward. 2. RULE BASED SYSTEM AND ITS IMPLEMENTATIONS A rule Based system is an expert system based on a set of rules that a human expert would follow in diagnosing a problem. It is also known as "production system" or "expert system" [2]. Rules represent possible actions which are taken when specified conditions hold on items in the working memory. In a rule based system, the separation of knowledge from control further simplifies development of expert systems by enabling an iterative development process where the engineer acquires, implements and tests individual rules. Rule-based systems are one of the most successful AI paradigms. They are used for synthesis type systems and diagnostic or classification type systems [3]. In this section we discuss some of the implementations of rule based system. A client server architecture consisting of Java based client and Prolog [4-5] based server is discussed in [6]. This is an interesting development of internet Prolog applications which are using Java as front end. It presents application architecture along with image search application and an expert system application for diagnosing car faults. The function of image search application is to search images for remote client depending on the input observations. This application consists of a static server component and a mobile client component. The function of expert system application is to accept a set of observations as input and provide the diagnosis as the output. Using the architecture presented in this paper, various Prolog applications could be ported to the internet. An interesting example to automatically develop web- based expert systems from XML parser which consist of knowledge base is discussed in [7]. The domain specific XML parser changes the knowledge base into Prolog code. A case study of university course rules is presented which relies on XML data definition file of common rules related with courses. Web pages dynamically execute Prolog code to provide responses to the user queries. The Prolog predicates for detecting patterns and antipatterns are Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 4 6 The Specification of an Expert System for Building Bylaws Compliance defined in [8]. The detection methods are proposed for both structural and behavioural features of patterns and antipatterns. Roventa, et. al. [9] a medical expert system is designed which gives help to a medical expert in making appropriate diagnosis of a patient. A knowledge base of twenty seven kidney diseases from nine different categories is developed based on symptoms of kidney diseases. Expert system diagnoses a particular disease using the knowledge of symptoms and laboratory tests. The system is incomplete in the sense that limited number of kidney diseases is considered. A similar expert system application for identification of dolphin's species in Malaysian fisheries using PROLOG is discussed in [10]. A rule based expert system for electrical distribution network steady state is discussed in [11]. The developed model is providing the facility to take real time decisions during system operation. The use of Prolog for building robust commercial applications is discussed in [12]. It has been successfully implemented in environmental system to predict weather, analyze water supplies and heavyweight applications. Prolog has been applied by Youbet.com in the advancement of its ebusiness structure [3]. Youbet.com provides detailed information about four different vendors to horse racing customers. It uses a Java based application server which is integrated with Prolog rule base engine. Another applicability of Prolog using RDF library for semantic web applications is discussed in [13]. 3. EXPERT SYSTEM FOR CHECKING BUILDING BYLAWS COMPLIANCE The proposed solution is developed using the combination of an Object Oriented Language Java [14] and a Rule Based Language Prolog [15-18]. Following are the major components of the system as shown in Fig. 1. (i) User Interface Subsystem (JSwing) (ii) Plan Rule Based Engine (Prolog) (iii) Plan Database Management System The ESBBC supports the bylaws collected from HDA for residential buildings of following sizes: (i) Plot sizes up to 120 Square Yards (ii) Plot sizes up to 150 Square Yards (iii) Plot sizes up to 200 Square Yards (iv) Plot sizes up to 240 Square Yards FIG. 1. SYSTEM DIAGRAM OF ESBBC Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 4 7 The Specification of an Expert System for Building Bylaws Compliance (v) Plot sizes up to 400 Square Yards (vi) Plot sizes up to 600 and above Square Yards The ESBBC is provides following options as shown in Fig. 2. (i) Checking of plan (ii) Saving of record (iii) Searching of property. 3.1 Prerequisites ESBBC requires Windows System with JDK1.3 or higher, SICSTUS Prolog and Jasper package to operate. If Java is used as parent application then there are a couple of things that need to be taken care of. The first is to specify the correct class path so that Java can find the jasper classes (SICStus, SPTerm, and so on). This is done by specifying the pathname of the file 'jasper.jar'. The second is to specify where Java should find the Jasper native library ('libspnative.so' or 'spnative.dll'), which the SICStus class loads into the JVM by invoking the method System.loadLibrary ("spnative"). Under Windows, Java must be able to find 'spnative.dll' through the PATH environment variable. 3.2 Java Interfacing with Prolog The low-level interface uses the JNI (Java Native Interface) to call Prolog from Java. The JNI is a native interface standard developed by Sun Microsystems. It permits Java programmers to integrate native code (currently C and C++, Prolog) into their Java applications by loading proper library. Calling Prolog from Java is done by using the Java package se.sics.jasper. This package contains a set of Java classes, which can be used to create and manipulate terms, ask queries and request one or more solutions. Jasper is a bi- directional interface between programs written in Java and programs written in Prolog. The Java-side of the interface consists of a Java package (jasper) containing classes representing the SICStus emulator [19]. The following example shows how one bylaw is checked from 'Bylws.pl' and response is given. FIG. 2. USE CASE MODEL OF ESBBC Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 4 8 The Specification of an Expert System for Building Bylaws Compliance import se.sics.jasper.*; import java.io.*; public class SimpleA { Static { System.loadLibrary("spnative"); } public static void main(String argv[]) throws Exception { SICStus sp; SPPredicate pred; SPTerm way; SPQuery query; int i; sp = new SICStus(argv,null); sp.load("e:/jdk1.4/bin/Bylaws.pl"); pred = new SPPredicate(sp, "covered_at_ground", 2, ""); way = new SPTerm(sp).putVariable(); query = sp.Query(pred, new SPTerm(way) ); System.out.println(way.toString()); } } 3.3 Implementation The class ExpertSystem works as the Control manager. The objects of following classes are used in ExpertSystem class. Build Complaint Plan WelcomePanel SearchForm AboutDialog Fig. 3 shows the plan entry form of the ESBBC in which first details of the plan are entered which are checked against defined bylaws. The form also provides the option to modify previous plan, delete non required plan and search the required plan. 3.4 Advantages (i) Although both database programs and expert systems feature the retrieval of stored information, a database program cannot draw conclusions from facts in its domain while an expert system containing both declarative and procedural knowledge can emulate the behaviour of human reasoning. Therefore the use of Prolog for developing bylaws is beneficial. (ii) Proposed solution is secure, robust and portable due to Java. (iii) Logic based languages are able to represent the real world more accurately. (iv) Prolog is able to derive new rules from the existing rules contained within the knowledge base. FIG. 3. PLAN INTERFACE Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 4 9 The Specification of an Expert System for Building Bylaws Compliance (v) Jasper is a bi-directional interface between Java and SICStus Prolog providing complete functionality to deal with Prolog predicates from Java. 4. RESULTS AND DISCUSSION In this section bylaws of few submission drawings are analysed via ESBBC. To confirm the reliability of ESBBC, submission drawings are also checked via manual system and finally both results are compared. 4.1 Analysis of Submission Drawings Table 1 presents the summary of results after checking the submission drawings using ESBBC. It is evident from Table 1 and Fig. 4 that only 15% of the submission drawings are violating the building regulations which are defined by HDA. 4.2 Discussion After the analysis of results which are presented in Section 4.1, it is identified that out of 20 properties 3 properties are violating the bylaws defined by HDA. It is really necessary to evaluate the results in order to identify which rule is violated. After analysis of results, it identified that following rules are violated by different properties: It is revealed from Table 2 that out of 3 properties, 1 property is violating the bylaw related to the compulsory open space which must be left at the rear and 2 properties are violating the bylaw related to the compulsory open space which must be left at the first floor. These rules are necessary to follow because in case of disobedience of these bylaws, shadowing problems are created. TABLE 1. NUMBER OF BYLAWS VIOLATED BY DIFFERENT PROPERTIES No. Property Area Covered Address No. of Total No. No. (SYD) Violated Bylaws 1 . H. No. 62 1 2 0 62/D Unit-10 Latifabad 0 4 2 . H. No. 63 1 2 0 63/C Unit-7 Latifabad 0 4 3 . H. No. 52 1 2 0 52/B Unit-3 Latifabad 0 4 4 . H. No. 44 1 2 0 44/A Unit-2 Latifabad 0 4 5 . H. No. 36 1 5 0 36/A Unit-1 Latifabad 1 4 6 . H. No. 76 1 5 0 76/B Unit-6 Latifabad 0 4 7 . B. No. 88 1 5 0 88/A Unit-5 Latifabad 0 4 8 . B. No. 121 1 5 0 121/D Unit-8 Latifabad 0 4 9 . B. No. 57 2 0 0 57/B Unit-2 Latifabad 0 6 1 0 . B. No. 28 2 0 0 28/B Unit-1 Latifabad 0 6 1 1 . B. No. 123 2 0 0 123/D Unit-7 Latifabad 0 6 1 2 . B. No. 129 2 0 0 129/D Unit-9 Latifabad 0 6 1 3 . B. No. 96 2 4 0 96/A Unit-1 Latifabad 1 6 1 4 . B. No. 112 2 4 0 112/D Unit-11 Latifabad 0 6 1 5 . B. No. 133 2 4 0 133/C Unit-4 Latifabad 0 6 1 6 . B. No. 149 2 4 0 149/A Unit-2 Latifabad 0 6 1 7 . B. No. 192 4 0 0 192/C Unit-2 Latifabad 0 6 1 8 . B. No. 152 4 0 0 153/D Unit-4 Latifabad 1 6 1 9 . B. No. 75 4 0 0 75/A Unit-1 Latifabad 0 6 2 0 . B. No. 93 4 0 0 93/B Unit-3 Latifabad 0 6 Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 5 0 The Specification of an Expert System for Building Bylaws Compliance 4.3 Comparison To demonstrate that ESBBC is giving reliable results, a comparison of manual system and ESBBC is performed. Manual system is chosen for comparison because of the unavailability of any other computerized system for HDA. The plans which are being checked by ESBBC in Section 4.1 are checked through manual system. Table 3 shows the comparison of manual checking and ESBBC. It is evident from the table that out of 20 plans ESBBC is able to identify that 15% of properties are violating rules and manual system is identifying that 10% of the properties are violating rules. In order to verify which results are correct, manual verification is performed again. As the result of that; it has been proved that the results given by ESBBC are correct which confirms the reliability of ESBBC. 5. CONCLUSIONS In this paper, ESBBC is proposed after analysis of current building control department system. This is first time that a solution based on expert system has been presented for HDA. The proposed solution presents clear idea about the working of building control department and makes it easier to understand for non-professional persons. It shows how HDA functionality can be best implemented using expert system. The proposed solution is able to verify building compliance of plans by implementing rules in Prolog and checking them from Java. It is maintaining records of plans by implementing database and providing their access from Java. Further, it is providing the facility to search a particular plan by entering owner name. Moreover, we have also analysed submission drawing of twenty properties and presented results. The results show that 15% of the submission drawings are violating the rules defined by HDA because of error prone nature of humans. Currently the solution is based on the implementation of bylaws for zonal plan HDA scheme 1, 2 and 3. The bylaws of other wards can also be implemented in the future extension. TABLE 2. BYLAWS VIOLATED BY DIFFERENT PROPERTIES Property No. Area Covered (SYD) Address No. of Bylaws Violated Bylaw Violated H. No. 36 1 5 0 36/A Unit-1 Latifabad 1 5 feet compulsory space shall be left open at the rear B. No. 96 2 4 0 96/A Unit-1 Latifabad 1 Compulsory open space on first floor 50% B. No. 152 4 0 0 153/D Unit-4 Latifabad 1 Compulsory open space on first floor 50% FIG. 4. NUMBER OF BYLAWS VIOLATED BY DIFFERENT PROPERTIES TABLE 3. COMPARISON OF ESBBC AND MANUAL SYSTEM Property No. Results of ESBBC Results of Manual System No. of Bylaws Violated H. No. 36 1 0 B. No. 96 1 1 B. No. 152 1 0 Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 5 1 The Specification of an Expert System for Building Bylaws Compliance ACKNOWLEDGEMENTS The authors would like to thank Mehran University of Engineering & Technology, Jamshoro, Pakistan, for providing required resources to complete this project. Authors also appreciate the guidance and support of Engr. Mohsin Ali Memon, Lecturer, Department of Software Engineering, Mehran University of Engineering & Technology, Jamshoro. REFERENCES [1] Information Technology Cell, WASA, HDA, "Hyderabad Development Authority", [Accessed December, 2010] Available from: http://hda-wasa.com/ [2] Engelmore, R.S., and Feigenbaum, E., "Expert System and Artificial Intelligence", [Accessed November, 2010] Available from: http://www.wtec.org/loyola/kb/ c1_s1.htm. [3] Tung, D., "Practical Application of Prolog to eBusiness: A Racing Case Study", PC AI Magazine, April 2001, [Accessed December, 2010] Available from: http:// www.amzi.com/articles/youbet.htm [4] Brakto, I., "Prolog Programming for Artificial Intelligence", Addison Wesley, 2000. [5] Krzysztof, R.A., "From Logic Programming to Prolog", Prentice-Hall, 1997. [6] El-Beltagy, S.R., Rafea, M., and Rafea, A., "Practical Development of Internet Prolog Applications using Java Front End", 2nd International Workshop on Logic Programming Tools for Internet Applications in Conjunction with ICLP'97, July, 1997. [7] Dunstan, N., "Generating Domain-Specific Web-Based Expert Systems", Expert Systems with Applications, Volume 35, pp. 686-690, 2008. [8] Stoianov, A., and Sora, I., "Detecting Patterns and Antipatterns in Software Using Prolog Rules", Proceedings of International Joint Conference on Computational Cybernetics and Technical Informatics, pp. 253-258, 2010. [9] Roventa, E., and Rosu, G., "The Diagnosis of Some Kidney Diseases in a Small Prolog Expert System", IEEE Proceedings of Third International Workshop on Soft Computing Applications, pp. 219-224, 2009. [10] Mohamad, R., Daut, N., Barukang, L., and Jaaman, S.A., "Development of Expert System for Identifying Dolphin's Species in Malaysian Fisheries Using PROLOG", International Symposium on Information Technology (ITSim), Kuala Lumpur, Malaysia, 26-29 August, 2008. [11] Borlea, I., Vuc, G., Jigoria-Oprea, D., Kilyeni, A., Barbulescu, C., and Slavici, T., "A Rule-Based Expert System for Steady State Diagnosis of Electrical Distribution Networks", IEEE Proceedings of 15th Mediterranean Electrotechnical Conference, pp. 142-147, 2010. [12] Roth, A.I., "The Practical Application of Prolog", December 10, 2002, [Accessed December, 2010] Available from: http://drdobbs.com/high-performance- computing/184405220. [13] Wielemaker, J., Hildebrand, M., and Ossenbruggen, J.V., "Prolog as the Fundament for Applications on the Semantic Web", Second International workshop of Logic Programming to the Web, Semantic Web and Semantic Web Services, ALPSWS 2007. [14] Oracle, "Java Native Interface", [Accessed November, 2010] Available from: http://java.sun.com/j2se/1.4.2/ docs/guide/jni/ [15] Carlsson, M., and Mildner, P., "Sicstus Prolog-the First 25 Years", Theory and Practice of Logic Programming, 2010. Mehran University Research Journal of Engineering & Technology, Volume 31, No. 2, April, 2012 [ISSN 0254-7821] 2 5 2 The Specification of an Expert System for Building Bylaws Compliance [16] Agren, M., "Tracing and Explaining the Execution of CLP (FD) Programs in SICStus Prolog", Master's Thesis, Uppsala University, 2002. [17] Gupta, G., Pontelli, E., Ali, K., Carlsson, M., and Hermenegildo, M., "Parallel Execution of Prolog Programs", ACM Transaction on Programming Languages and Systems, Volume 23, No. 4, pp. 1-13, 2001. [18] Intelligent Systems Laboratory, Swedish Institute of Computer Science, "Sicstus Prolog", [Accessed November, 2010] Available from: http://www.sics.se/isl/ sicstuswww/site/index.html [19] Intelligent Systems Laboratory, Swedish Institute of Computer Science, "Mixing Java and Prolog", October, 1998 [Accessed December, 2010] Available from: http:/ /www.sics.se/sicstus/docs/3.7.1/html/sicstus_12.html 
work_ayay4uyj5zdbxk2sqsflpgwbqi	----	Microsoft Word - CBM Platform (ESWA) 2 An intelligent condition-based maintenance platform for rotating machinery Van Tung Tran a, b , Bo-Suk Yang a,* a Department of Mechanical and Automotive Engineering, Pukyong National University, San 100, Yongdang-dong, Nam-gu, Busan 608-739, South Korea b Faculty of Mechanical Engineering, Hochiminh City University of Technology, 268 Ly Thuong Kiet St., Dist. 10, Ho Chi Minh City, Vietnam ABSTRACT Maintenance is of necessity for sustaining machinery availability and reliability in order to ensure productivity, product quality, on-time delivery, and safe working environment. The costly maintenance strategies such as corrective maintenance and scheduled maintenance have been progressively replaced by superior maintenance strategies in which condition-based maintenance (CBM) is one of the delegates. This strategy commonly consists of sequent modules such as data acquisition, signal processing, feature extraction and feature selection, condition monitoring, etc. However, approaches in literature which have been developed for each module and implemented for different applications are standalone instead of a comprehensive system. Furthermore, these approaches have been demonstrated in a laboratory environment without any industrial validations. For these reasons, an intelligent algorithm based CBM platform is proposed in this paper to be applied for rotating machinery easily and effectively. Subsequently, two case-studies are presented in order to evaluate the effectiveness of this platform in industrial applications. Keywords: Condition-based maintenance, Diagnostics, Prognostics, Signal processing, Feature extraction, Feature selection 1. Introduction A failure in industrial equipment results in not only the loss of productivity but also timely services to customer, and they may even lead to safety and environmental problems. This emphasizes the necessity of maintenance in manufacturing operations. Maintenance can sustain the reliability and availability of product equipments, improve the product quality, increase productivity, and undertake the safety requirements. However, the general opinion is that maintenance is a necessary evil or nothing can be done to improve maintenance costs. Indeed, the cost of maintenance contributes a large part of the total operating and production cost in specific capital-intensive industries. For example, maintenance cost as a percentage of totals 3 value-added could be up to 20-50% for mining, 15-25% for primary metal and 3-15% for processing and manufacturing industries (Campbell, 1995). Additionally, with the augment of mechanization and automation, many modern plants have installed flexible automatic computer- controlled and unmanned equipments, the maintenance costs have been increased substantially. Consequently, an efficient and reasonable maintenance strategy is necessary for implementation so that the minimum maintenance costs could be attained. The maintenance strategies could be broadly classified into two categories, namely corrective maintenance (CM) and preventive maintenance (PM) (Tsang, 1995; Lebold et al., 2003), and summarized in Table 1. CM, also known as breakdown maintenance, is frequently performed by unplanned activities and implemented after the occurrence of an obvious functional failure, malfunction, or breakdown of the equipment. Its actions can store the functional capabilities of failed components by either repairing the defect or replacing the new ones. PM involves scheduled maintenance and condition-based maintenance (CBM). Scheduled maintenance is periodically carried out by lubricating, calibrating, refurbishing, inspecting, and checking equipment for each fixed period of time to lessen the deterioration leading to faults. This helps prevent the functional failures by replacing critical components at regular intervals which are shorter than their expected useful lives. However, random stoppages of equipment in CM or frequent replacements of the expensive components before the end of their lives in scheduled maintenance result in high cost of maintenance. CBM is a method to reduce the uncertainty of maintenance activities that frequently encounter in other strategies. In CBM, equipment operating conditions are continuously monitored to identify the need for maintenance in real time. The actual preventive actions are only taken when an incipient failure is supposed to have been detected. Therefore, CBM can significantly reduce the maintenance costs by reducing the number of unnecessary scheduled PM operations if properly established and effectively implemented. For that reason, CBM has been received great attention of researchers and practical maintainers. Table 1 Maintenance approaches. Generally, a CBM system comprises a number of functional modules: sensing and data acquisition, signal processing, feature extraction and feature selection, condition monitoring and health assessment, diagnostics, prognostics, decision reasoning, and human system interface. Therefore, in order to implement CBM system, it is required to integrate a variety of hardware and software components. In recent times, these components have been continuously improved to be applied and practiced in different applications by several researches. In case of hardware component, thanks to the fast forwarding development of sensor industry, not only the price and the size of sensors have been reduced but also more tasks have been performed in benchmarking with the conventional ones. A new sensor generation which involves micro- 4 electromechanical systems and smart sensors has been employed for CBM to improve the sensor reliability and accuracy. Owing to the advantages of mutual information from multiple sensors, sensor fusion techniques are commonly engaged to lead the superiority. In case of software component, numerous standalone approaches have been fruitfully developed to implement the CBM system for rotating machinery. Most of these approaches focus on fault diagnostics and prognostics and fall into three main categories: statistical-based, model-based, and data driven-based. The applications of these approaches for rotating machinery could be found in the studies of Jardine (Jardine et al., 2006) and Heng (Heng et al., 2009). Even though standalone approaches in progress have been developed in literature, there are several fundamental opportunities to be considered: • Most of existing approaches are able to apply for specific equipment. A scalable methodology or systematic toolbox for generic machinery does not exist. • Several developed algorithms have been demonstrated in a laboratory environment without any industrial validations. • Currently, methods are generally focused on solving the failure prediction issue. Tools for system performance assessment and degradation prediction have not been well addressed. Furthermore, estimating the remaining lifetime of machinery is still a challenge of prognostics. • There is no available platform that can be applied for various rotating machinery. In recent years, several methods in our studies have been developed and applied for various rotating machinery in the range of signal processing to prognostics. For instance, wavelet transform and statistical method, which were used to extract salient features from raw noise and vibration signals, were combined with self-organizing feature map, learning vector quantization and support vector machine (SVM) for detecting and classifying faults of reciprocating refrigerator compressors (Yang et al., 2005). An expert system, namely VIBEX, was proposed to aid plant operators in diagnosing the cause of abnormal vibration for rotating machinery (Yang et al., 2005). The decision tree in this system was used as an acquisition of structured knowledge to obtain the diagnosing rules from decision table which is built by the cause- symptom matrix. The integration of case-based reasoning with other methods such as Petri nets, adaptive resonance theory, the learning strategy of Kohonen neural network, etc was introduced to the fault diagnosis of induction motors (Yang et al., 2004). Other our works in diagnostic area could be found in references (Niu et al., 2007; Widodo et al., 2007; Niu et al., 2008; Widodo et al., 2008; Tran et al., 2008; Son et al., 2009). Alternatively, in prognostic area, we have proposed numerous models for forecasting future states (Tran et al., 2008; Tran et al., 2009; Niu et al., 2009; Pham et al., 2010; Caesarendra et al., 2010), assessing the degradation (Niu & Yang, 2010; Pham et al., 2010; Caesarendra et al., 2010). Although most of our researches were validated in industrial equipment, they still were 5 standalone approaches. In order to compensate for the remaining shortcomings mentioned above, a systematic platform based on intelligent algorithms, namely intelligent CBM (I-CBM), is proposed in this study. This platform consists of sequent modules: data acquisition, signal processing, feature extraction and feature selection, condition monitoring and health assessment, fault diagnostics, and prognostics. Two case-studies are presented to perform the effectiveness of this platform in industrial application. 2. Intelligent condition-based maintenance architecture The architecture of I-CBM platform is shown in Fig. 1. This platform contains modules with the aim of converting the rotating machinery signals into the useful information for the maintainers to take remedial actions, inspect the conditions, and conduct a repair on the defect before the catastrophic failure occurs. In each module, many applicable algorithms could be appropriately selected to obtain the best result. As depicted in Fig. 1, data can be obtained from the sensors installed on the machinery for condition monitoring or manually input working conditions. Subsequently, these data are transformed into features by selecting appropriate algorithms for signal processing, feature extraction, and feature selection. In the feature space, a proper algorithm is employed for each task of classifying the type of faults, performing the degradation, and forecasting the remaining lifetime of machinery. In order to easily and conveniently implement the proposed I-CBM, a toolbox has been developed which the main user interface is shown in Fig. 2 and the algorithms used for each module of this toolbox is summarized in Table 2. A brief introduction of these modules is given as follows: Fig. 1. The architecture of I-CBM platform. Fig. 2. The main user interface of I-CBM platform. Table 2 Algorithms in I-CBM. Data acquisition: A major challenge confronted with CBM is that how the functional symptoms are monitored in terms of measurable machinery states. Data are a requirement for this challenge. Data acquisition is a process of collecting and storing useful data from target system to monitor the condition, diagnoses the faults, and prognosticate the future states and remaining lifetime. According to (Jardine et al., 200), data used for CBM could be categorized into two main types: event data and condition monitoring data. The former includes the information on what has happened (e.g., installation, breakdown, overhaul, etc.) and what has been done (e.g., minor repair, preventive maintenance, oil change, etc.) to the machinery. The latter is the measurements related to the health condition/states of the machinery. Condition monitoring data is very versatile which could be vibration, acoustic, oil analysis, temperature, 6 pressure, moisture, humidity, weather or environment data, etc. In this platform, vibration and current data are commonly used due to the easy-to-measure signals and analysis. Signal processing: is a process of removing distortions and restoring the original shape of signals, removing sensor data which is not relevant, transforming the signal to make relevant features more explicit. Many methods could be applied for this process in I-CBM platform, for example wavelet transform, fast Fourier transforms (FFT) which is demonstrated in Fig. 3. Fig. 3. FFT for vibration signals. Feature representation: Data obtained from signal processing process is rarely usable in its raw form due to the huge dimensionality. The huge dimensionality causes not only difficulties of data storage but also data transfer. Therefore, representing data as features is the demand for reduce the huge dimensionality. Feature representation or feature calculation module is a sub- module of feature-based techniques, as shown in Fig. 4, and plays a crucial role in attaining the performance of I-CBM platform. Here, the represented features include time domain features (e.g. root mean square, variance, shape factor, skewness, kurtosis, crest factor, etc.), frequency domain features (e.g. content at the feature frequency, the amplitude of FFT spectrum, etc.). Fig. 5 presents an example of feature representation module for vibration signals. Fig. 4. The user interface of module of feature-based techniques. Fig. 5. An example of feature representation module. Feature extraction and/or feature selection: Total features obtained in the previous process can cause cures of dimensionality and peaking phenomenon that greatly degrade the classification accuracy. Feature extraction can be viewed as a pre-pruning process to choose a small subset of total features that is necessary and sufficient to describe the overall operations of machine systems. The importance of feature extraction is not only to reduce the search space, but also to speed up the process of classification and also to improve the quality of classification. The extracted feature vectors will serve as one of the essential inputs to fault diagnosis and prognosis algorithms. In this platform, common algorithms used for feature extraction are principal component analysis (PCA), independent component analysis (ICA), kernel PCA, kernel ICA, linear discriminant analysis, etc., as an example shown in Fig. 6. Even though dimensionality is reduced by the feature extraction process, each feature set contains many redundant or irrelevant features as well as salient features in feature space. Consequently, feature selection process is of necessity to find an optimal subset of features that maximizes information content or predictive accuracy. Fig. 7 is the result obtained from feature selection process. 7 Fig. 6. An example of feature extraction in I-CBM platform. Fig. 7. The result of feature selection in I-CBM platform. Diagnostics: This module is used for analyzing the pattern embedded in the features to determine the root causes of previous observed faults or degradation. In order to attain this purpose by using the I-CBM platform, several classification methods are applied. Furthermore, the maintainers can compare the performance of each method to certainly affirm the condition of machine. Fig. 8 describes the diagnosing results of induction motor faults. Health assessment: The health status and the degradation of machinery are performed in this module by using condition parameters. It also provides the unacceptable level or the failure threshold for the operations so that the appropriate actions will be taken to avoid the consequences of failure before the failure occurs. Prognostics: Prognostics is the ability to predict the remaining lifetime, future health states, or reliability of machinery based on current health assessment and historical trends. Thus, there are two main functions of prognostics: failure prediction and remaining lifetime estimation. Failure prediction allows the pending failures to be identified before they come to a serious situation. Remaining lifetime is the time left before a particular fault will occur or any part needs to be replaced. The techniques related to prognostics can be classified as experience- based, model-based, and data driven-based. The prognostics module of I-CBM platform addresses both these functions in which most of methods belong to data driven-based techniques. Fig. 9 shows a forecasted result of kurtosis feature of bearing by using ARMA/ GARCH model (Pham et al., 2010). Fig. 8. Performance of each method for fault diagnostics of induction motor. Fig. 9. Forecasted result of machine state. 3. Industrial case-studies Two industrial case-studies are presented to demonstrate the applications of proposed platform for rotating machinery. In the case of diagnostics, induction motor is considered due to its indispensable roles in several industrial applications. The faults of induction motor may not only cause the interruption of production operation but also increase costs, decrease product quality and effect safety of operators. Most common faults of induction motors are bearing failures, stator winding failures, broken rotor bar or cracked rotor end-rings and air-gap irregularities (Acosta et al., 2006). To diagnose these faults in this case-study, a combination of wavelet transform and SVM, namely W-SVM, are employed. In the case of prognostics, a low 8 methane compressor which is an important equipment in petrochemical plant is used as an object for investigation. Self-organizing map (SOM) and failure threshold determination are used for assessing the degradation of machine and setting an alarm, respectively. Once the degradation value is higher than failure threshold, the Cox’s proportional hazard model (PHM) (Cox, 1972) in association with SVM is triggered to forecast the remaining lifetime of machine. 3.1. Rotating machinery fault diagnostics 3.1.1. Data acquisition and feature calculation Data acquisition was conducted on induction motor of 160 kW, 440 V, 2 poles as shown in Fig. 10. Six accelerometers were installed along vertical, horizontal and axial directions to pickup vibration signal at drive-end and non drive-end. The maximum frequency of the used signals and the number of sampled data were 60 Hz and 16384, respectively. The condition of induction motor is briefly summarized in Table 3. Each condition was labeled as class from 1 to 7. There are totally 126 features calculated from 6 signals, 21 features and 98 samples calculated from 7 conditions, 14 measurements. Fig. 10. Data acquisition of induction motor. Table 3 Condition of induction motor. 3.1.2. Feature extraction Structure of three first original features, those are mean, RMS, and shape factor are plotted in Fig. 11. This figure shows the performance of original features which are containing overlap in some conditions. To make original features well clustered, applying component analysis is suggested in this study. Component analysis via ICA, PCA, and their kernel are used to extract and reduce the feature dimensionality based on eigenvalue of covariance matrix as described in Fig. 12. After performing component analysis, the features have been changed into independent and principal components, respectively. The first three independent and principal components from PCA, ICA, and their kernel are plotted in Fig. 13. It can be observed that the clusters for seven conditions are separated well. It indicates that component analysis can perform feature extraction and all at once do clustering each condition of induction motors. Fig. 11. Original features. Fig. 12. Feature reduction using component analysis. Fig. 13. The first three principal and independent components. According to the eigenvalue of covariance matrix, the features are changed into component analysis and reduce only 5 component analysis needed for classification process. The other 9 features are discarded due to small of eigenvalue of covariance matrix. The selected component analysis is then used as input vectors for W-SVM classifier to diagnose the faults of induction motor. 3.1.3. Result and discussion The SVM based multi-class classification is applied to perform the classification process using one-against-all methods. Vapnik (Vapnik, 1982) describe a method which used the projected conjugate gradient algorithm to solve the quadratic programming (QP) problem in SVM. In this study, 1 and 10 -7 are assigned to the parameter C (bound of the Lagrange multiplier) and λ (condition parameter for QP method), respectively. Furthermore, wavelet kernel function using Daubechies series is performed. The parameter δ in wavelet kernel refers to number of vanishing moment and is set 4. In the training process, the data set is also trained using RBF kernel function as comparison. The parameter γ for bandwidth RBF kernel is user defined equal to 0.5. The complex separation boundaries of W-SVM are presented in Fig. 14. In these figures, the circle refers to the support vector that states the correct recognition in W-SVM. Each condition of induction motor is well recognized using Daubechies wavelet kernel. In the classification process using W-SVM, each condition of induction motors can be clustered well. The good separation among conditions shows the performance of W-SVM doing recognition of component analysis from vibration signal features. The performance of classification process is summarized in Table 4. All data set come from component analysis are accurately classified using Daubechies wavelet kernel and SVM and reached accuracy 100% in training and testing, respectively. SVM using RBF kernel function with kernel width γ = 0.5 is also performed in classification for comparison with Daubechies wavelet kernel. The results show that the performance of W-SVM is similar to SVM using RBF kernel function, those are 100% in accuracy of training and testing, respectively. In the case of number support vectors, SVM with RBF kernel function needs lower than W-SVM except kernel PCA. Fig. 14. Separation boundaries of W-SVM. Table 4 Results of classification. 3.2. Rotating machinery prognostics 3.2.1. Data acquisition Methane compressor shown in Fig. 15 is important equipment used in petrochemical industry where normal production flow is required to maintain. Due to the importance of this machine, condition monitoring and prognostics are of necessity to sustain its operation. This 10 compressor is driven by a 440 kW motor, 6600 V, 2 poles and operating at a speed of 3565 rpm. Other information of the system is summarized in Table 5. The condition monitoring system of this compressor consists of two types: off-line and on- line. In the off-line system, the acceleration sensors are installed along axial, vertical, and horizontal directions at the locations of drive-end motor, non drive-end motor, male rotor compressor and suction part of compressor. In the on-line system, acceleration sensors are located at the same places as in the off-line system but only in the horizontal direction. Vibration signal was recorded from August 2005 to November 2005. The average recording duration was 6 hours during the data acquisition process. Fig. 15. Low methane compressor. Table 5 Information of the system. 3.2.2. Machine health assessment and failure threshold determination After acquiring vibration signal, RMS and envelope features used in this study are extracted. RMS is a common feature, even the only indicator used in ISO 10816 and ISO 7919 for machinery condition monitoring and alarming. Envelope is useful to detect glitches (narrow pulse signals). Both of the two features are often used for condition monitoring of rotating machinery as plotted in Figs. 16 and 17. In these figures, the machine was obviously in normal operating condition during the first 300 points of the time sequence. At the 291st point, the machine condition significantly reduced in comparison with other points. After the 300th point, the machine suddenly changed the condition and was broken down at the 308th point. Thus, there was a process of degrading from normal operating state to failure state in this compressor. Fig. 16. The entire peak acceleration data of low methane compressor. Fig. 17. The entire envelope acceleration data of low methane compressor. To perform this degradation, a process of normalization is conducted to transform values of features into a common scale and group them as input set for feature-fusion. SOM neural network is employed to combine the input set into a single out indicator which is minimum quantization error (MQE) (Qiu et al., 2003), as shown in Fig. 18. Comparing with the plots of RMS and envelope features, MQE indicator maintains a more steady state than envelop curve, whilst enhances a degradation trend than RMS curve, which is especially appropriate for initial fault detection and health degradation prediction. Additionally, MQE indicator in Fig. 18 can be obviously recognized the sudden increase from the 300th point. This appropriates to the sudden change of RMS feature mentioned above as well as indicates the abnormal value at the 291st point. Therefore, MQE is adequate to assess the machine degradation. 11 Fig. 18. Machine health indicator. The next step is to determine the failure threshold so that the prognostic module is triggered. Further reading the method for determining this threshold could be found in reference (Ginart et al., 2006). Failure threshold is also employed to attain the censored data which is used for generating the PHM in the next process. This data consists of a series of “0” and “1” values indicating the normal condition and failure condition, respectively. 3.2.3. Remaining lifetime estimation To estimate the remaining lifetime, The Cox’s PHM is built based on the censored data obtained from previous step. The parameters of this PHM are estimated as β1 = −0.8042 for RMS feature and β1 = 0.1062 for envelope feature. Hazard rate and survival function estimation of this PHM are depicted in Figs. 19 and 20, respectively. In Fig. 19, the hazard rate gradually increases with respect to time. From the 300th point, the hazard rate significantly changes because of the rapid growth of RMS values. Thus, the more the hazard rate increases, the less the reliability is. Fig. 19. Hazard rate estimation. Fig. 20. Survival function estimation. After attaining the survival function, the process of training and forecasting by using SVM in association with time-series forecasting techniques is carried out. The multi-step ahead direct prediction method [20] of time-series forecasting techniques is applied for this study. The values of survival function up to 291st point are used to train SVM model in which the Gaussian kernel 2 2 ( , ) exp( | | /(2 ) )K x y x y σ= − − is employed. The other predefined parameters ε and C are set to 0.001 and 500, respectively. Moreover, 5-fold cross validation is also applied to choose the best SVM model. The predicted results of 17 points from the 292nd point where the machine commences degrading the state are depicted in Fig. 21. Even though the multi-step ahead is employed, the predicted results is closely resemble with the actual values. From the predicted results, the remaining lifetime could be estimated by using the predefined survival probability for fault. For example, if this value is chosen as 0.2, the point where predicted result reaches for 0.2 is 306th. This means the predicted remaining lifetime of machine after degrading state occurred is 84 hours ((306 − 292) × 6 = 84) while the actual remaining lifetime is 84 hours (the 306th point). As a result, the prognostic module of I-CBM platform can estimate accurately the remaining lifetime of the methane compressor. 12 Fig. 21. Forecasting results. 4. Conclusions This study proposes the I-CBM platform based on standalone data-driven approaches as a comprehensive system for rotating machinery. This platform contains necessary modules to adequate the request for CBM system involved hardware and software components. A toolbox has also been developed as a software component to easily and conveniently implement this platform in application. Two industrial cases are used to validate the effectiveness of the I-CBM platform. The accurate results obtained from these cases indicate that the I-CBM platform can fulfill the shortcomings of previous approaches i.e. systematic tool for various rotation machinery, performance degradation assessment, remaining lifetime estimation, and industrial validation. Acknowledgments The authors gratefully acknowledge the support of Brain Korea 21 Project and The Vietnam National Foundation for Science and Technology Development (NAFOSTED) for this study. We would like to send our sincere thanks to all the people who have contributed to and worked on this aspect, especially Dr. Gang Niu, and Dr. Achmad Widodo. References Acosta, G.G., Verucchi, C.J., Gelso, E.R. (2006). A current monitoring system for diagnosis electrical failures in induction motors. Mechanical Systems and Signal Processing, 20(4), 953-965. Caesarendra, W., Niu, G., Yang, B.S. (2010). Machine condition prognosis based on sequential Monte Carlo method. Expert Systems with Applications, 37(3) 2412-2420. Caesarendra, W., Widodo, A., Yang, B.S. (2010). Application of relevance vector machine and logistic regression for machine degradation assessment. Mechanical Systems and Signal Processing, 24(4), 1161-1171. Campbell, J.D. (1995). Uptime: Strategies for excellence in maintenance management. Productivity Press. Cox, D.R. (1972). Regression models and life-table. Journal of the Royal Statistical Society, Series B, 34(2), 187-220. Ginart, A., Barlas, I., Goldin, J., Dorrity, J.L. (2006). Automated feature selection for embeddable prognostic and health monitoring (PHM) architectures. In IEEE systems readiness technology conference (pp. 195-201). New York: IEEE. Heng, A., Zhang, S., Tan, A.C.C., Mathew, J. (2009). Rotating machinery prognostics: state of the art, challenges and opportunities. Mechanical Systems and Signal Processing, 23 (3), 724-739. Jardine, A.K.S., Lin, D., Banjevic, D. (2006). A review on machinery diagnostics and prognostics implementing condition-based maintenance. Mechanical Systems and Signal Processing, 20(7), 1483-1510. Lebold, M., Reichard, K., Boylan, D. (2003). Utilizing DCOM in an open system architecture framework for machinery monitoring and diagnostics. Proceedings of IEEE Aerospace Conference (Vol. 3, pp. 1227-1236). Niu, G., Han, T., Yang, B.S., Tan, A.C.C. (2007). Multi-agent decision fusion for motor fault diagnosis. Mechanical Systems and Signal Processing, 21(3), 1285-1299. Niu, G., Widodo, A., Son, J.D., Yang, B.S., Hwang, D.H., Kang, D.S. (2008). Decision-level fusion based on wavelet decomposition for induction motor fault diagnosis using transient current signal. Expert Systems with Applications, 35(3), 918-928. 13 Niu, G., Yang, B.S. (2009). Dempster-Shafer regression for multi-step-ahead time-series prediction towards data-driven machinery prognosis. Mechanical Systems and Signal Processing, 23(3), 740- 751. Niu, G., Yang, B.S. (2010). Intelligent condition monitoring and prognostics system based on data-fusion strategy. Expert Systems with Applications, 37(12), 8831-8840 Pham, H.T., Tran, V.T., Yang, B.S. (2010). A hybrid of nonlinear autoregressive model with exogenous input and autoregressive moving average model for long-term machine state forecasting. Expert Systems with Applications, 37(4), 3310-3317 Pham, H.T., Yang, B.S. (2010). Estimation and forecasting of machine health condition using ARMA/ GARCH model. Mechanical Systems and Signal Processing, 24(2), 546-558. Qiu, H., Lee, J., Lin, J., Yu, G. (2003). Robust performance degradation assessment methods for enhanced rolling element bearing prognostics. Advanced Engineering Informatics, 17(3-4), 127-140. Son, J.D., Niu, G., Yang, B.S., Hwang, D.H., Kang, D.S. (2009). Development of smart sensors system for machine fault diagnosis. Expert Systems with Applications, 36(9), 11981-11991. Tran, V.T., Yang, B.S., Oh, M.S., Tan, A.C.C. (2008). Machine condition prognosis based on regression trees and one-step-ahead prediction. Mechanical Systems and Signal Processing, 22(5), 1179-1193. Tran, V.T., Yang, B.S., Oh, M.S., Tan, A.C.C. (2009). Fault diagnosis of induction motor based on decision trees and adaptive neuro-fuzzy inference. Expert Systems with Applications, 36(2), 1840- 1849. Tran, V.T., Yang, B.S., Tan, A.C.C. (2009). Multi-step ahead direct prediction for the machine condition prognosis using regression trees and neuro-fuzzy systems. Expert Systems with Applications, 36(5), 9378-9387. Tsang, A.H.C. (1995). Condition-based maintenance: tools and decision making. Journal of Quality in Maintenance Engineering, 1 (3) 3-17. Vapnik, V.N. (1982). Estimation Dependences Based on Empirical Data. Springer Verlag, Berlin. Widodo, A., Yang, B.S. (2007). Application of nonlinear feature extraction and support vector machines for fault diagnosis of induction motors. Expert Systems with Applications, 33(1), 241-250. Widodo, A., Yang, B.S. (2008). Wavelet support vector machine for induction machine fault diagnosis based on transient current signal. Expert Systems with Applications, 35(1-2), 307-316. Widodo, A., Yang, B.S., Han, T. (2007). Combination of independent component analysis and support vector machines for intelligent faults diagnosis of induction motors. Expert Systems with Applications, 32(2), 299-312. Yang, B.S., Han, T., Kim, Y.S. (2004). Integration of ART-Kohonen neural network and case-based reasoning for intelligent fault diagnosis. Expert Systems with Applications, 26(3), 387-395. Yang, B.S., Hwang, W. W., Kim, D. J., Tan, A.C.C. (2005). Condition classification of small reciprocating compressor for refrigerators using artificial neural networks and support vector machines. Mechanical Systems and Signal Processing, 19(2), 371-390. Yang, B.S., Jeong, S.K., Oh, Y.M., Tan, A.C.C. (2004). Case-based reasoning system with Petri nets for induction motor fault diagnosis. Expert Systems with Applications, 27(2), 301-311. Yang, B.S., Lim, D.S., Tan, A.C.C. (2005).VIBEX: an expert system for vibration fault diagnosis of rotating machinery using decision tree and decision table. Expert Systems with Applications, 28(4), 735-742. 14 Rotating machinery Signal processing Removing distortions Restoring the original shape Removing sensor data Transforming the signal Wavelet transform Fast Fourier transform Empirical mode decomposition Feature representation Feature in time domain Feature in frequency domain Feature in time- frequency domain Feature extraction/ feature selection Reducing the dimension of feature space PCA, ICA, Kernel PCA, Kernel ICA, ... Prognostics Prediction remaining lifetime/ machine reliability Feature forecasting ANFIS CART Dempster- Shafer SVM RVM GARCH ... Diagnostics Detecting faults Classifying fault types Bayesian classifier K-nearest neighbors SOM ART-Kohonen SVM RVM ANFIS Data fusion ... Selecting the salient features Condition entropy Backward/Forward feature selection, ... Health assessment Monitoring the rotating machinery conditions Degradation assessment Failure threshold setting SOM ARMA/GARCH Logistic regression PHM Data acquisition Condition monitoring parameters Testing data Training data Creating diagnostic model Validating model Classification Performing classified results Good model? N Y Condition monitoring Trigger? Failure threshold setting Y Time-series recontruction Prediction model Prognostics Fig. 1. The architecture of I-CBM platform. Fig. 2. The main user interface of I-CBM platform. 15 Fig. 3. FFT for vibration signals. Fig. 4. The user interface of module of feature-based techniques. 16 Fig. 5. An example of feature representation module. Fig. 6. An example of feature extraction in I-CBM platform. 17 Fig. 7. The result of feature selection in I-CBM platform. Fig. 8. Performance of each method for fault diagnostics of induction motor. 18 Fig. 9. Forecasted result of machine state. Fig. 10. Data acquisition of induction motor. -0.4 -0.2 0 0.2 0.4 0.6 0.4 0.5 0.6 0.7 -0.2 -0.1 0 0.1 0.2 Bent shaft Eccentricity Magnetic center moved (DE) Magnetic center moved (NDE) Normal Unbalance rotor Weak-end shield Fig. 11. Original features. 19 Fig. 12. Feature reduction using component analysis. -1.5 -1 -0.5 0 0.5 1 -2 -1 0 1 -3 -2 -1 0 1 2 3 PC 1 PC 2 P C 3 Bent shaft Eccentricity Magnetic center moved (DE) Magnetic center moved (NDE) Normal Unbalance rotor Weak-end shield (a) Principle components 0 10 20 30 40 50 60 70 80 90 100 0 0.5 1 1.5 2 2.5 3 3.5 Compone nt a nalysis E ig e n v a lu e Reduced 9988 55 20 -2 -1 0 1 2 -1 0 1 2 3 -2 -1.5 -1 -0.5 0 0.5 1 IC 1 IC 2 IC 3 Bent shaft Eccentricity Magnetic center moved (DE) Magnetic center moved (NDE) Normal Unbalance rotor Weak-end shield (b) Independent components -0.2 -0.1 0 0.1 0.2 -0.2 -0.1 0 0.1 0.2 -0.4 -0.2 0 0.2 0.4 PC 1 PC 2 P C 3 Bent shaft Eccentricity Magnetic center moved (DE) Magnetic center moved (NDE) Normal Unbalance rotor Weak-end shield (c) Kernel principle components 21 -2 -1 0 1 2 -2 -1 0 1 2 -1.5 -1 -0.5 0 0.5 1 1.5 IC 1 IC 2 IC 3 Bent shaft Eccentricity Magnetic center moved (DE) Magnetic center moved (NDE) Normal Unbalance rotor Weak-end shield (d) Kernel independent components Fig. 13. The first three principal and independent components. Support vector -1.5 -1 -0.5 0 0.5 -2.5 -2 -1.5 -1 -0.5 0 0.5 Class 7 Class 3 Class 6 Class 2S u p p o rt v e c to r Class 5 Class 4 Support vector -1.5 -1 -0.5 0 0.5 -2.5 -2 -1.5 -1 -0.5 0 0.5 Class 7 Class 3 Class 6 Class 2S u p p o rt v e c to r Class 5 Class 4 (a) Daubechies kernel with PC data 22 Support vector -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 -1.5 -1 -0.5 0 0.5 1 Class 4 Class 3 Class 1 Class 6 Class 7 Class 2 Class 5 S u p p o rt v e c to r Support vector -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 -1.5 -1 -0.5 0 0.5 1 Class 4 Class 3 Class 1 Class 6 Class 7 Class 2 Class 5 S u p p o rt v e c to r (b) Daubechies kernel with IC data -0.3 -0.2 -0.1 0 0.1 0.2 -0.1 -0.05 0 0.05 0.1 0.15 S u p p o rt v e c to r Support vector Class 5 Class 4 Class 3 -0.256 -0.254 -0.252 -0.25 -0.248 -0.246 -0.244 -0.242 -0.11 -0.105 -0.1 -0.095 -0.09 -0.085 -0.08 -0.075 -0.07 -0.065 -0.06 Class 6 Class 1 Class 7 Class 2 -0.3 -0.2 -0.1 0 0.1 0.2 -0.1 -0.05 0 0.05 0.1 0.15 S u p p o rt v e c to r Support vector Class 5 Class 4 Class 3 -0.256 -0.254 -0.252 -0.25 -0.248 -0.246 -0.244 -0.242 -0.11 -0.105 -0.1 -0.095 -0.09 -0.085 -0.08 -0.075 -0.07 -0.065 -0.06 Class 6 Class 1 Class 7 Class 2 (c) Daubechies kernel with kernel PC data 23 -1 -0.5 0 0.5 1 -1.5 -1 -0.5 0 0.5 1 1.5 Support vector Class 4Class 4 Class 1Class 1 Class 6Class 6 Class 5Class 5 S u p p o rt v e c to r Class 7Class 7 Class 2Class 2 Class 3Class 3 -1 -0.5 0 0.5 1 -1.5 -1 -0.5 0 0.5 1 1.5 Support vector Class 4Class 4 Class 1Class 1 Class 6Class 6 Class 5Class 5 S u p p o rt v e c to r Class 7Class 7 Class 2Class 2 Class 3Class 3 (d) Daubechies kernel with kernel IC data Fig. 14. Separation boundaries of W-SVM. Fig. 15. Low methane compressor. 24 200 400 600 800 1000 1200 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Time A c c e le ra ti o n (g ) Fig. 16. The entire peak acceleration data of low methane compressor. 0 200 400 600 800 1000 1200 0 0.5 1 1.5 2 2.5 3 Time A c c e le ra ti o n (g ) Fig. 17. The entire envelope acceleration data of low methane compressor. 25 Fig. 18. Machine health indicator. 150 200 250 300 0 2 4 6 8 10 12 14 16 18 20 Time H a z a rd r a te Fig. 19. Hazard rate estimation. 26 150 200 250 300 0 0.2 0.4 0.6 0.8 1 Time S u rv ia l p ro b a b il it y Fig. 20. Survival function estimation. 150 200 250 300 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Time S u rv iv a l p ro b a b il it y X= 292 Y= 0.9417 Actual Predicted Fig. 21. Forecasting results. 27 Table 1 Maintenance approaches. Maintenance strategy Frequency of maintenance Criteria of initiating maintenance Condition assessment Results Corrective Unscheduled Upon failure/ work stoppage to fix immediate problems Unusual ·Unpredictable asset availability and reliability Preventive Pre-scheduled Prescribed based on failure history or test data Unusual/ manual data collection ·Maintenance performed more often than may be necessary Condition- based Just-in-time Prescribed based on statistical patterns in operating parameters Continuous/real- time sensor monitoring and data collection ·Maintenance performed when necessary ·Highest asset availability and reliability of mission- critical assets ·Continuous data collection 28 Table 2 Algorithms in I-CBM. Module name Methods Signal processing Wavelet transform Fast Fourier transform Empirical mode decomposition Feature representation Time domain analysis Frequency domain analysis Time-frequency domain analysis Feature extraction Principal component analysis (PCA) Independent component analysis (ICA) Kernel PCA Kernel ICA Fisher discriminant analysis (FDA) Linear discriminant analysis (LDA) Generalized discriminant analysis (GDA) Clustering techniques Feature selection Individual feature evaluation based on space distribution Condition entropy Backward feature selection Forward feature selection Branch and bound feature selection Plus l-take away r feature selection Floating forward feature selection Distance evaluation technique Taguchi method-based feature selection Genetic algorithm Diagnostics Linear classifier Quadratic classifier Bayesian classifier k-nearest neighbors Self-organizing feature map neural network Learning vector quantization neural network Radial basic function neural network ART-Kohonen neural network Support vector machine (SVM) Decision tree Random forest Adaptive neuro-fuzzy integrated system (ANFIS) Data fusion system Prognostics Adaptive network-based fuzzy inference system (ANFIS) Classification and regression trees (CART) Dampster-Shafer regression Autoregressive moving average (ARMA) Generalized autoregressive conditional heteroscedasticity (GARCH) Relevance support vector machine (RVM) Support vector machine (SVM) Logistic regression Proportional hazard model 29 Table 3 Condition of induction motor. Class No. Condition Description Others 1 Bent rotor Maximum shaft deflection 1.45mm 2 Eccentricity Static eccentricity (30%) Air-gap: 0.25 mm 3 MCDE Magnetic center moved (DE) 6 mm 4 MCNDE Magnetic center moved (NDE) 6 mm 5 Normal No faults - 6 Unbalance Unbalance mass on the rotor 10 gr 7 Weak-end shield Stiffness of the end-cover - Table 4 Results of classification. Kernel Accuracy (training/test), % Number of SVs IC PC Kernel IC Kernel PC IC PC Kernel IC Kernel PC Wavelet Daubechies 100/100 100/100 100/100 100/100 35 39 39 17 RBF-Gaussian (γ = 0.5) 100/100 100/100 100/100 100/100 22 22 25 33 Table 5 Information of the system. Electric motor Compressor Voltage 6600 V Type Wet screw Power 440 kW Lobe Male rotor (4 lobes) Pole 2 Pole Female rotor (6 lobes) Bearing NDE:#6216, DE:#6216 Bearing Thrust: 7321 BDB RPM 3565 rpm Radial: Sleeve type 
work_ayd5ovacqfaazalhyttvgeezae	----	Blogger-Centric Contextual Advertising Teng-Kai Fan Chia-Hui Chang Department of Computer Science and Information Engineering, National Central University, Chung-Li, Taiwan 320, ROC tengkaifan@gmail.com, chia@csie.ncu.edu.tw ABSTRACT This paper addresses the concept of Blogger-Centric Contextual Advertising, which refers to the assignment of personal ads to any blog page, chosen in according to bloggers' interests. As blogs become a platform for expressing personal opinions, they naturally contain various kinds of statements, including facts, comments and statements about personal interests, of both a positive and negative nature. To extend the concept behind the Long Tail theory in contextual advertising, we argue that web bloggers, as the constant visitors of their own blog sites, could be potential consumers who will respond to ads on their own blogs. Hence, in this paper, we propose using text mining techniques to discover bloggers’ immediate personal interests in order to improve online contextual advertising. The proposed BCCA (Blogger-Centric Contextual Advertising) framework aims to combine contextual advertising matching with text mining in order to select ads that are related to personal interests as revealed in a blog and rank them according to their relevance. We validate our approach experimentally using a set of data that includes both real ads and actual blog pages. The results indicate that our proposed method could effectively identify those ads that are positively-correlated with a blogger’s personal interests. Categories and Subject Descriptors I.2.6 [Artificial Intelligence]: Learning. H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval. General Terms Measurements, Algorithms, Experiments. Keywords Web Advertising, Sentiment Detection, Marketing, Blog, Intention Recognition, Machine Learning. 1. INTRODUCTION Contextual advertising is based on studies that show that 80% of Internet users are interested in receiving personalized content on sites they visit [4]. Since the topic of a page somehow reflects the interest of visitors, ads delivered to visitors should depend upon page content rather than upon stereotypes created according to their geographical locations or upon other demographic features, such as gender or age [7]. As shown in previous studies, strong relevance increases the number of click-throughs [1] [3] [10] [13]. Some studies [6] [14] have also demonstrated that focusing on relevant topics written with positive sentiment produces high click-through rates. Although a page-relevant topic is a way to capture visitors’ interest, there is no other way to determine their personal interests. However, since bloggers are constant visitors to their own blogs, and their interests or intensions are well expressed in the weblogs, an ad agency could use those expressed intentions to place interest-oriented ads. For example, Figure 1 shows a weblog with five ads related to traveling placed on the right. Since the content of this page describes reasons for cancelling a trip, these traveling ads are unlikely to be clicked. Instead, what the blogger needs is medical information or information on doctors. The point here is that an ad agency system could assign relevant ads according to bloggers’ own interests, especially their immediate interest or intentions, for targeting ads, thus treating bloggers as the main visitors of their own blogs. To this end, even if an ad is related to the content of a linking page, an ad agency should preferentially consider the immediate interests expressed on the page (i.e., intentions) for placing ads. In this paper, we proposed an ad matching mechanism, which we refer to as Blogger-Centric Contextual Advertising (BCCA), and which is based on latent interest detection to associate ads with blog pages. Instead of the traditional placement of relevant ads, BCCA emphasizes that the ad agency’s system should provide relevant ads that are related to different levels of personal preferences in order to increase clicks. To evaluate our proposed method, we used a real-word collection comprised of ads and blog pages from Google AdSense and Google’s Blog-Search Engine, respectively. Our results show that the proposed approach, based on text mining can effectively recognize the latent interests (e.g., intentions) in a blog page, or the personal interests of the blogger. In addition, we further investigated the effects of ad page matching using an ad Click-Through Rate (CTR) experiment, and our results suggest that our proposed method can effectively match relevant ads to a given blog page. Relevant AdsRelevant Ads Figure 1: Example of a blog page with ads Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. CIKM’09, November 2–6, 2009, Hong Kong, China. Copyright 2009 ACM 978-1-60558-512-3/09/11...$10.00. 2. BCCA FRAMEWORK Traditional contextual advertising processes a given page a user visits to find related topics for matching ads, while blogger- centric contextual advertising would assign ads to a given blog page in accordance with the blogger’s interests. Before demonstrating the proposed framework, we will explain how bloggers’ personal interests could be obtained. Generally speaking, an individual blog often contains profile information, tags and posts, which we could classify as indicating different levels of interest. Profile-level: Blog service providers (e.g., BlogSpot.com, Technorati.com) usually ask bloggers’ to enter interests to build their profiles (e.g., music, reading) when they register as a member of the service. In addition to generic interests collected at registration time, the tags on posts or the archive of past posts can be used to construct bloggers’ profiles, showing their specific interests. Since these kinds of interest continue for a period of time, we view them as long-term interests. In this paper, we assumed that each blog-site has an interest profile containing either generic or specific long-term interests. Triggering-level: Blog posts are media in which bloggers express their opinions and interests, as well as their intensions. For example, the sentence, "The Nokia N95 is a good cell phone for several reasons," express the sentiments of the author toward the object in question, a Nokia N95. For blogger-centric contextual advertising, we argue that recognizing the intentions of authors could be even more effective. For instance, the sentence, "We're going to the doctor right now," in Figure 1 indicates that the author has an immediate intention to see a doctor. As another example, the sentence, "I am looking for a new laptop," implies that the author probably will purchase a laptop. As such, targets are immediate interests; ads centered around them might increase click-through. In BCCA framework, the advertising system analyzes the content of the page to recognize intention and detect sentiment for triggering-level interests. If no such targets are found, the system uses targets from the blogger’s profile and searches the ad database to find the best matching ads. These four modules: intention recognition, sentiment detection, term expansion and target-ad matching are the main components in our BCCA framework. The first two components analyze triggering-level interests, while the last two components enhance the target-ad matching procedure. Note that we place a priority on ad assignment based on different levels of interests. That is, triggering-level interests (i.e., immediate interests) have higher priority than profile-level interests (long-term interests). Thus, the sentiment detection module is invoked only when no positive intention is detected. If neither module detects intention or positive sentences, the system should then use targets from the blogger’s profile. Next, the system proceeds to term expansion and the target-ad matching agency. Due to the short form of ads and targets, we designed a term expansion component to enhance the likelihood of intersection between targets and available ads. Finally, a retrieval function based on a query likelihood language model is deployed for the target-ad matching strategy to rank the ads. 2.1 Intention Recognition Given a triggering page, our aim in this section is to explain the process of recognizing whether there exist any intention-bearing sentences. By modeling this problem one of classification, our job here is the preparation of training sentences which must be labeled as intentional or non-intentional. Labeling each sentence as intentional or non-intentional is a time-consuming and costly task. In this study, we propose a novel technique to semi-automatically label training data for this task. Since many auction web sites (e.g., ebay.com, yahoo.com) provide special forums for buyers to post their needs, it naturally becomes a source of intention-filled sentences. In this study, we collected a large set of posts containing buyers' requirements from ebay.com 1 . However, since many buyers describe their requirements (e.g., product name, product type) in a concise sentence, the labeling system filters simple sentences that do not contain intentions by the following criteria. Part-of-Speech (POS) tag: Since an intentional sentence usually contains a verb that is not a form of "to be", we keep candidate sentences that contain terms tagged as verbs, and remove sentences that contain only noun phrases. The length of the sentence: short polite words are usually used in the forums. Hence, we simply disregard sentences whose lengths are less than three words. All the candidate sentences that conform to the above rules are regarded as intentional data. As for non-intentional training data, we manually construct some queries that will collect entry pages from Wikipedia2. For feature selection, we considered a subset of word unigrams chosen via the chi-square feature selection. In this study, we selected the top-K features with a high chi-square value as input features. To apply these machine learning algorithms on our dataset, we used the standard bag-of-features framework. 2.2 Sentiment Detection Our aim in this section is to apply a contextual sentiment technique for detecting the sentiment of a blog page. To build an efficient learning model, we divided the task of detecting the sentiment of a sentence into two steps, each of which belongs to a binary classification problem. The first is an identification step that aims to identify whether the sentence is subjective or objective. The second is a classification step that classifies the subjective sentences as positive or negative. For sentiment detection, a good Web resource is epinions.com, where pros and cons of products or other topics are discussed by reviewers. Such information is thus used by Kim and Hovy to prepare their training data [8]. Although this method is fully automatic, such an idea based on pro and con phrases does not perform well (F- measure: 0.65). As indicated in [5], opinionated content is most often carried by parts of speech used as modifiers (i.e., adverbs and adjectives) rather than parts of speech used as heads (i.e., verbs, nouns). Thus, we combined the concept proposed in [5] and modified Kim & Hovy’s method as follows. For each term tagged as adjective or adverb (e.g., JJ, RB) in pro and con sets, our labeling system checks each sentence to find sentences that 1 http://pages.ebay.co.uk/wantitnow/ 2 http://en.wikipedia.org contain those adjectives or adverbs. Then the system annotates each sentence with appropriate sentiment labels. As for feature selection, we similarly adopted the chi-square statistic method to select the top-K features for both the sentiment identification and sentiment classification steps. For the identification step, all the subjective and objective sentences are represented as feature- presence vectors, where the presence or absence of each feature is recorded. For the classification step, all the positive and negative sentences are similarly represented as feature-presence vectors to train the classifiers. 2.3 Term Expansion In general, a blog page can be about any theme, while the advertisements are concise in nature. Hence, the intersections of terms between ads and pages are very low. In this paper, we followed three term-expansion methods proposed in [6] to expand the specific terms in a blog page. 2.4 Page-Ad Matching We can regard the blogger-centric contextual advertising issue as a traditional information retrieval problem: that is, given a user’s query q, the IR (Information Retrieval) system returns relevant documents d according to the query content. Hence, we intuitively model a triggering page p and relevant ads a as a user’s query q and corresponding documents d, respectively. In this paper, the query likelihood language model is adopted as our ad retrieval model. The language modeling approach to IR models the following idea: A document d is a good match to a query q if the document model is likely to generate query q, which will in turn happen if the document contains the query words often [11]. Hence, we constructed from each ad a in the collection a language model Ma. Our goal is to rank ads by P(a|q), where the probability of an ad a is interpreted as the likelihood that it is relevant to the query q. 3. EXPERIMENTAL RESULTS In this section, we focus on our three experiments: intention recognition, sentiment detection and page-ad matching. We begin by describing the dataset, and we then proceed to a discussion of the experimental results. 3.1 Datasets To evaluate intention recognition performance, we retrieved data from ebay.com and wikipedia.com and then used these as our training dataset for building our learning classifier model. For sufficient detail, we first collected 36,151 buyer's posts and then randomly examined 10,000 sentences. For non-intentional training data, we similarly examined 10,000 sentences from the Wikipedia dataset. Additionally, to evaluate sentiment detection performance, we collected data from epinions.com and used this as our training dataset for building learning classifier models. We gathered many different types of reviews from epinion.com. For sufficient detail, we examined about 30,000 reviews (about 900,000 sentences), with an average number of 30 sentences per review document. For page-ad matching, we collected a total of 138,907 ads. Our triggering page collection was comprised of 200 blog pages on various topics. It included a range of opinions and was comprised of various subjective articles (100 positive and 50 negative articles) and 50 general articles including intentional sentences. 3.2 Evaluation of Intention Recognition For the dataset of triggering pages (i.e., 50 intentional articles), owing to a lack of training data, we chose SVM learning algorithms with different feature sets to train the models on the ebay.com and Wikipedia datasets. We then applied these models to our triggering page dataset. The results are shown in Table 1. SVM with features selected by chi-square gave the better results, yielding an average 94.2% F-measure for the non-intentional class and an average 80.5% F-measure for the intentional class. 3.3 Evaluation of Sentiment Detection The goals of this section are similar to [6] [8]: the target is to investigate how well the trained model performed on a different data source (200 triggering pages). We compared the modified method to the one (regarded as the baseline system) used in [6]. We chose SVM learning algorithms with different feature sets to train the models on the epinion.com dataset. We then applied these models to our triggering page dataset. We conducted the sentiment identification and classification step and the results are shown in Table 2. For the sentiment identification task, the SVM with features selected by chi-square gave the better results, yielding an average 87.1% F-measure in the objective class and an average 69.9% F-measure in the subjective class. For the sentiment classification task, the best F-measure for the positive class (74.8%) and the best F-measure for the negative class (59.1%) are generated by the SVM with features selected by chi- square and the SVM with unigrams respectively. Table 1: Intension Recognition by various feature sets Feature Set (# of features) Class Pre. (%) Recall (%) F-measure (%) Non-Intentional 94.5 93.5 94.0 Unigram (45030) Intentional 77.8 80.9 79.4 Non-Intentional 95.5 92.9 94.2 Chi-Square (10000) Intentional 77.1 84.2 80.5 Table 2: Sentiment Detection for triggering pages Feature Set Task Class Prec. (%) Recall (%) F-measure (%) Obj. 81.1 88.9 84.8 Identification Sub. 75.8 62.8 67.7 Neg. 80.6 46.6 59.1 Uni- grams Classification Pos. 70.0 65.0 67.4 Obj. 80.5 95.0 87.1 Identification Sub. 86.6 58.8 69.9 Neg. 63.8 52.6 57.6 Chi- Square Classification Pos. 96.0 61.3 74.8 Obj. 78.5 64.5 70.8 Identification Sub. 50.1 67.7 58.2 Neg. 56.1 53.1 54.6 Baseline Classification Pos. 56.2 71.1 68.2 Table 3: Click-Through Rate (CTR) for triggering pages Dataset Method CTR BCCA 52% PCA 53% Positive dataset Google AdSense 47% BCCA 58% PCA 32% Intention dataset Google AdSense 42% BCCA 57% PCA 42% All triggering pages Google AdSense 40% 3.4 Evaluation of Page-Ad Matching The goal of this section is to investigate to what extent ad placements are actually related to personal interests evident on the triggering pages. To evaluate our page-ad matching framework, we compared the top-five ranked ads provided by three different ranking methods: our proposed approach with a personalization mechanism (i.e., BCCA), our proposed approach without a personalization mechanism (i.e., PCA (Plain Contextual Advertising) excluding intention recognition and sentiment detection), and Google AdSense. Here we used the SVM with features selected by chi-square method as intention and sentiment models according to experimental results. Since it is difficult to invite the original authors of blog actually to participate in an Ad Click-Through-Rate (CTR) experiment, 14 volunteers participated in our CTR experiment. All the advertisements in each pool were manually clicked by volunteers. To compare with Google AdSense, we only measured the CTR for this experiment. The measure is the fraction of retrieved ads that are clicked. Table 3 shows the results for three page-ad matching methods. In the case of the positive sentiment dataset, there are no significant differences among the three page-ad matching approaches. For the intention dataset, our results show that the proposed BCCA approach can yield a better performance (58%) than other approaches (32% for PCA approach and 42% for Google Adsense). Regarding all triggering pages, our results show that the proposed BCCA approach can yield a better performance (57%) than other approaches (42% for PCA approach and 40% for Google AdSense). According to Table 3, these results lead us to the conclusion that our BCCA framework can place ads that are related to the personal interest content of triggering pages. 4. RELATED WORK Several studies pertaining to advertising research have stressed the importance of relevant associations for consumers [13] and how irrelevant ads can turn off users and relevant ads are more likely to be clicked [3]. As for contextual advertising, Ribeiro- Neto et al. [12] proposed a number of strategies for matching pages to ads based on extracted keywords. To identify the important parts of the ad, the authors explored the use of different ad sections as a basis for the ad vector. In follow-up research [9], the authors proposed a method to learn the impact of individual features using genetic programming to generate a matching function. However, due to the vagaries of phrase extraction, approaches based on “bid phrases” leads to many irrelevant ads. To overcome this problem, Broder et al. [2] proposed a system for contextual ad matching based on a combination of semantic and syntactic features. The results show that the semantic-syntactic approach outperformed the syntactic approach. Prior studies usually focused on the interests of end-users to assign advertisements. There have been very few studies that tried to deliver personal ads geared towards publishers' interests. Besides, even if several prior studies have proven that relevance has a definite impact on contextual advertising and on the proposed effective ranking function that matches ads with pages, they neglect to consider personal intentions in the assignment of relevant ads. In this study, we further investigate the importance of personal interests mining, including intention and sentiment analysis for improved contextual advertising. 5. CONCLUSION In this study, we proposed and evaluated a novel framework for associating ads with blog pages based on intention analysis. Prior work to date has only examined the extent of content relevance between pages and ads. In this paper, we investigated the intentions and the sentiments of blog pages and utilized this information to demonstrate blogger centric contextual advertising. For intention recognition, we adopted the information from ebay.com and Wikipedia as training data to recognize the latent immediate interest. As for sentiment detection, we proposed a new method to label the sentences from epinions.com for training data preparation. For page-ad matching, we evaluated our framework using 200 personal blog pages and over 130,000 ads sampled from Google AdSense. We compared BCCA with Google AdSense and found that our proposed method with personalized advertising mechanism can achieve superior performance (57% accuracy). 6. REFERENCES [1] A. Anagnostopoulous, A. Z. Broder, E. Gabrilovich, V. Josifovski and L. Riedei. Just-in-Time Contextual Advertising. In CIKM'07, page. 331-340, 2007. [2] A. Broder, M. Fontoura, V. Josifovski and L. Riedel. A semantic approach to contextual advertising. In SIGIR'07, page 559-566, 2007. [3] P. Chatterjee, D. L. Hoffman and T. P. Novak. Modeling the Clickstream: Implications for Web-Based Advertising Efforts. Marketing Science 22, 520-541, 2003. [4] ChoiceSteam, ChoiceStream personalization survey: consumer trends and perceptions, http://www.choicestream.com/pdf/ChoiceStream_PersonalizationSurv eyResults2005.pdf, 2005. [5] A. Esuli and F. Sebastiani. SENTIWORDNET: A Publicly Available Lexical Resource for Opinion Mining. In CLRE'06, page. 417-422, 2006. [6] T.-K. Fan and C.-H. Chang. Sentiment-Oriented Contextual Advertising. ECIR, LNCS 5478, page 202-215, 2009. [7] P. Kazienko and M. Adamski. AdROSA-Adaptive personalization of web advertising. Information Sciences 177, 2269-2295, 2007. [8] S.-M. Kim and E. Hovy. Automatic Identification of Pro and Con Reasons in Online Reviews. In ACL'06, page 483-490, 2006. [9] A. Lacerda, M. Cristo, M. A. Goncalves, W. g. Fan, N. Ziviani and B. Ribeiro-Neto. Learning to advertise. In SIGIR06, page 549-556, 2006. [10] OneUpWeb. How keyword length affects conversion rates, http://www.oneupweb.com/landing/keywordstudy_landing.htm. [11] M. Ponte and W. B. Croft. A language modeling approach to information retrieval. In SIGIR'98, page 275-281, 1998. [12] B. Ribeiro-Neto, M. Cristo, P. B. Golgher and E. S. d. Moura. Impedance coupling in content-targeted advertising. In SIGIR'05 page 496-503, 2005. [13] C. Wang, P. Zhang, R. Choi and M. D'Eredita. Understanding consumers attitude toward advertising. In Eighth Americas Conference on Information Systems, page 1143-1148, 2002. [14] Y. Zhang, A. C. Surendran, J. C. Platt and M. Narasimhan. Learning from multi-topic web documents for contextual advertisement. In SIGKDD'08, page 1051-1059, 2008. << /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Dot Gain 20%) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Warning /CompatibilityLevel 1.3 /CompressObjects /Tags /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJDFFile false /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /LeaveColorUnchanged /DoThumbnails true /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams true /MaxSubsetPct 100 /Optimize true /OPM 1 /ParseDSCComments false /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo true /PreserveFlatness true /PreserveHalftoneInfo true /PreserveOPIComments true /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Apply /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true /Academy /AgencyFB-Bold /AgencyFB-Reg /Alba /AlbaMatter /AlbaSuper /Algerian /Arial-Black /Arial-BoldItalicMT /Arial-BoldMT /Arial-ItalicMT /ArialMT /ArialNarrow /ArialNarrow-Bold /ArialNarrow-BoldItalic /ArialNarrow-Italic /ArialRoundedMTBold /ArialUnicodeMS /BabyKruffy /BaskOldFace /Bauhaus93 /BellMT /BellMTBold /BellMTItalic /BerlinSansFB-Bold /BerlinSansFBDemi-Bold /BerlinSansFB-Reg /BernardMT-Condensed /BlackadderITC-Regular /BodoniMT /BodoniMTBlack /BodoniMTBlack-Italic /BodoniMT-Bold /BodoniMT-BoldItalic /BodoniMTCondensed /BodoniMTCondensed-Bold /BodoniMTCondensed-BoldItalic /BodoniMTCondensed-Italic /BodoniMT-Italic /BodoniMTPosterCompressed /BookAntiqua /BookAntiqua-Bold /BookAntiqua-BoldItalic /BookAntiqua-Italic /BookmanOldStyle /BookmanOldStyle-Bold /BookmanOldStyle-BoldItalic /BookmanOldStyle-Italic /BookshelfSymbolSeven /BradleyHandITC /BritannicBold /Broadway /BrushScriptMT /CalifornianFB-Bold /CalifornianFB-Italic /CalifornianFB-Reg /CalisMTBol /CalistoMT /CalistoMT-BoldItalic /CalistoMT-Italic /Castellar /Centaur /Century /CenturyGothic /CenturyGothic-Bold /CenturyGothic-BoldItalic /CenturyGothic-Italic /CenturySchoolbook /CenturySchoolbook-Bold /CenturySchoolbook-BoldItalic /CenturySchoolbook-Italic /Chick /Chiller-Regular /ColonnaMT /ComicSansMS /ComicSansMS-Bold /CooperBlack /CopperplateGothic-Bold /CopperplateGothic-Light /CourierNewPS-BoldItalicMT /CourierNewPS-BoldMT /CourierNewPS-ItalicMT /CourierNewPSMT /Croobie /CurlzMT /EdwardianScriptITC /Elephant-Italic /Elephant-Regular /EngraversMT /ErasITC-Bold /ErasITC-Demi /ErasITC-Light /ErasITC-Medium /EstrangeloEdessa /Fat /FelixTitlingMT /FootlightMTLight /ForteMT /FranklinGothic-Book /FranklinGothic-BookItalic /FranklinGothic-Demi /FranklinGothic-DemiCond /FranklinGothic-DemiItalic /FranklinGothic-Heavy /FranklinGothic-HeavyItalic /FranklinGothic-Medium /FranklinGothic-MediumCond /FranklinGothic-MediumItalic /FreestyleScript-Regular /FrenchScriptMT /Freshbot /Frosty /Garamond /Garamond-Bold /Garamond-Italic /Gautami /Georgia /Georgia-Bold /Georgia-BoldItalic /Georgia-Italic /Gigi-Regular /GillSansMT /GillSansMT-Bold /GillSansMT-BoldItalic /GillSansMT-Condensed /GillSansMT-ExtraCondensedBold /GillSansMT-Italic /GillSans-UltraBold /GillSans-UltraBoldCondensed /GlooGun /GloucesterMT-ExtraCondensed /GoudyOldStyleT-Bold /GoudyOldStyleT-Italic /GoudyOldStyleT-Regular /GoudyStout /Haettenschweiler /HarlowSolid /Harrington /HighTowerText-Italic /HighTowerText-Reg /Impact /ImprintMT-Shadow /InformalRoman-Regular /Jenkinsv20 /Jenkinsv20Thik /Jokerman-Regular /Jokewood /JuiceITC-Regular /Karat /Kartika /KristenITC-Regular /KunstlerScript /Latha /LatinWide /LetterGothicMT /LetterGothicMT-Bold /LetterGothicMT-BoldOblique /LetterGothicMT-Oblique /LucidaBright /LucidaBright-Demi /LucidaBright-DemiItalic /LucidaBright-Italic /LucidaCalligraphy-Italic /LucidaConsole /LucidaFax /LucidaFax-Demi /LucidaFax-DemiItalic /LucidaFax-Italic /LucidaHandwriting-Italic /LucidaSans /LucidaSans-Demi /LucidaSans-DemiItalic /LucidaSans-Italic /LucidaSans-Typewriter /LucidaSans-TypewriterBold /LucidaSans-TypewriterBoldOblique /LucidaSans-TypewriterOblique /LucidaSansUnicode /Magneto-Bold /MaiandraGD-Regular /Mangal-Regular /MaturaMTScriptCapitals /MicrosoftSansSerif /Mistral /Modern-Regular /MonotypeCorsiva /MSOutlook /MSReferenceSansSerif /MSReferenceSpecialty /MVBoli /NiagaraEngraved-Reg /NiagaraSolid-Reg /OCRAExtended /OldEnglishTextMT /Onyx /PalaceScriptMT /PalatinoLinotype-Bold /PalatinoLinotype-BoldItalic /PalatinoLinotype-Italic /PalatinoLinotype-Roman /Papyrus-Regular /Parchment-Regular /Perpetua /Perpetua-Bold /Perpetua-BoldItalic /Perpetua-Italic /PerpetuaTitlingMT-Bold /PerpetuaTitlingMT-Light /Playbill /Poornut /PoorRichard-Regular /Porkys /PorkysHeavy /Pristina-Regular /PussycatSassy /PussycatSnickers /Raavi /RageItalic /Ravie /Rockwell /Rockwell-Bold /Rockwell-BoldItalic /Rockwell-Condensed /Rockwell-CondensedBold /Rockwell-ExtraBold /Rockwell-Italic /ScriptMTBold /ShowcardGothic-Reg /Shruti /SnapITC-Regular /Square721BT-Roman /Stencil /Sylfaen /SymbolMT /Tahoma /Tahoma-Bold /TempusSansITC /TimesNewRomanMT-ExtraBold /TimesNewRomanPS-BoldItalicMT /TimesNewRomanPS-BoldMT /TimesNewRomanPS-ItalicMT /TimesNewRomanPSMT /Trebuchet-BoldItalic /TrebuchetMS /TrebuchetMS-Bold /TrebuchetMS-Italic /Tunga-Regular /TwCenMT-Bold /TwCenMT-BoldItalic /TwCenMT-Condensed /TwCenMT-CondensedBold /TwCenMT-CondensedExtraBold /TwCenMT-Italic /TwCenMT-Regular /Verdana /Verdana-Bold /Verdana-BoldItalic /Verdana-Italic /VinerHandITC /Vivaldii /VladimirScript /Vrinda /Webdings /WeltronUrban /Wingdings2 /Wingdings3 /Wingdings-Regular /ZWAdobeF ] /NeverEmbed [ true ] /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 300 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 300 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.50000 /EncodeColorImages true /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /ColorImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 300 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 300 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /GrayImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 1200 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 600 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 >> /AllowPSXObjects false /CheckCompliance [ /PDFX1a:2001 ] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /Description << /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000410064006f006200650020005000440046002065876863900275284e8e9ad88d2891cf76845370524d53705237300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef69069752865bc9ad854c18cea76845370524d5370523786557406300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /DAN <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> /DEU <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> /ESP <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> /FRA <FEFF005500740069006c006900730065007a00200063006500730020006f007000740069006f006e00730020006100660069006e00200064006500200063007200e900650072002000640065007300200064006f00630075006d0065006e00740073002000410064006f00620065002000500044004600200070006f0075007200200075006e00650020007100750061006c0069007400e90020006400270069006d007000720065007300730069006f006e00200070007200e9007000720065007300730065002e0020004c0065007300200064006f00630075006d0065006e00740073002000500044004600200063007200e900e90073002000700065007500760065006e0074002000ea0074007200650020006f007500760065007200740073002000640061006e00730020004100630072006f006200610074002c002000610069006e00730069002000710075002700410064006f00620065002000520065006100640065007200200035002e0030002000650074002000760065007200730069006f006e007300200075006c007400e90072006900650075007200650073002e> /ITA <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> /JPN <FEFF9ad854c18cea306a30d730ea30d730ec30b951fa529b7528002000410064006f0062006500200050004400460020658766f8306e4f5c6210306b4f7f75283057307e305930023053306e8a2d5b9a30674f5c62103055308c305f0020005000440046002030d530a130a430eb306f3001004100630072006f0062006100740020304a30883073002000410064006f00620065002000520065006100640065007200200035002e003000204ee5964d3067958b304f30533068304c3067304d307e305930023053306e8a2d5b9a306b306f30d530a930f330c8306e57cb30818fbc307f304c5fc59808306730593002> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020ace0d488c9c80020c2dcd5d80020c778c1c4c5d00020ac00c7a50020c801d569d55c002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken die zijn geoptimaliseerd voor prepress-afdrukken van hoge kwaliteit. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <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> /PTB <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> /SUO <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> /SVE <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> /ENU (Use these settings to create Adobe PDF documents best suited for high-quality prepress printing. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.) >> /Namespace [ (Adobe) (Common) (1.0) ] /OtherNamespaces [ << /AsReaderSpreads false /CropImagesToFrames true /ErrorControl /WarnAndContinue /FlattenerIgnoreSpreadOverrides false /IncludeGuidesGrids false /IncludeNonPrinting false /IncludeSlug false /Namespace [ (Adobe) (InDesign) (4.0) ] /OmitPlacedBitmaps false /OmitPlacedEPS false /OmitPlacedPDF false /SimulateOverprint /Legacy >> << /AddBleedMarks false /AddColorBars false /AddCropMarks false /AddPageInfo false /AddRegMarks false /ConvertColors /ConvertToCMYK /DestinationProfileName () /DestinationProfileSelector /DocumentCMYK /Downsample16BitImages true /FlattenerPreset << /PresetSelector /MediumResolution >> /FormElements false /GenerateStructure false /IncludeBookmarks false /IncludeHyperlinks false /IncludeInteractive false /IncludeLayers false /IncludeProfiles false /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe) (CreativeSuite) (2.0) ] /PDFXOutputIntentProfileSelector /DocumentCMYK /PreserveEditing true /UntaggedCMYKHandling /LeaveUntagged /UntaggedRGBHandling /UseDocumentProfile /UseDocumentBleed false >> ] >> setdistillerparams << /HWResolution [600 600] /PageSize [612.000 792.000] >> setpagedevice 
work_azq5fa4xynajvgzrxxcnbzkvuq	----	Design and Evaluation of an Expert System in a Crushing Plant minerals Article Design and Evaluation of an Expert System in a Crushing Plant Claudio A. Leiva 1,* , Katheryn V. Arcos 1, Diego A. Poblete 1 , Eduardo A. Serey 1, Cynthia M. Torres 2 and Yousef Ghorbani 3 1 Department of Chemical Engineering, Universidad Católica del Norte, 1270709 Antofagasta, Chile; kati.arcos@gmail.com (K.V.A.); diego.poblete@ucn.cl (D.A.P.); edoserey90@hotmail.com (E.A.S.) 2 Department of Metallurgical and Mining Engineering, Universidad Católica del Norte, 1270709 Antofagasta, Chile; cynthia.torres@ucn.cl 3 Department of Civil, Environmental & Natural Resources Engineering, Luleå University of Technology, SE-971 87 Luleå, Sweden; yousef.ghorbani@ltu.se * Correspondence: cleiva01@ucn.cl; Tel.: +56-552655903 Received: 1 September 2018; Accepted: 15 October 2018; Published: 19 October 2018 ���������� ������� Abstract: This document presents a proposal for designing an expert system in the Gabriela Mistral Division’s crushing plant belonging to Codelco (Chile) with the objective of maximizing stacked tonnage, allowing the improvement of operational variables that directly interact with the crushing process. In addition, this study considers the impact that occurs in both the process and operational continuity regarding the standardization of the system. In the first stage, a survey and analysis of historic operation data was carried out, which allowed the definition of benchmarking indicators. Subsequently, both modalities of operation were compared, monitoring processed tonnage and detentions related to operational failures. As a result, significant differences were observed in the performance of the critical line operating with expert control, with a 55% reduction in the detentions referred to operational failures. Added to this is the benefit of low cost and improved quality as the control provides an analysis of the variables in reduced time intervals, which is superior to human control. Keywords: improvement; expert system; knowledge-based system; milling; crushing 1. Introduction The mining industry has been forced to improve the efficiency of its operations to reduce both supplies and energy costs due to the abrupt fall in copper prices since 2015. The complexities of the copper production process include variability in the characteristics of the feed material and operational detentions, among others. This dynamic interaction of processes makes it necessary to incorporate advanced technologies in the production process, such as the expert system—a technology that seeks to reach an optimum point of operation to increase the efficiency of the process [1]. There are four main benefits of developing an expert system, as mentioned in Durkin [2]: improved productivity, low cost, improved quality, and improved image. The Gabriela Mistral Division (Chile) produces 125 thousand tons of fine copper, with an average grade of 0.46% of total copper. This process is carried out by comminution operations (primary and secondary crushing), acidification, and stacking in dynamic stacks. In general terms, the amount of energy consumed in the crushing processes is closely related to the degree of reduction in size achieved by the particles; for the Gabriela Mistral Division, the energy cost corresponds to 30% of the total cost of operation. The variables that affect both the energy expenditure and the efficiency of the process were analyzed by Wei and Craig [3]. The ability to crush increasingly more material Minerals 2018, 8, 469; doi:10.3390/min8100469 www.mdpi.com/journal/minerals http://www.mdpi.com/journal/minerals http://www.mdpi.com https://orcid.org/0000-0003-1568-1462 https://orcid.org/0000-0002-7469-0943 https://orcid.org/0000-0002-5228-3888 http://www.mdpi.com/2075-163X/8/10/469?type=check_update&version=1 http://dx.doi.org/10.3390/min8100469 http://www.mdpi.com/journal/minerals Minerals 2018, 8, 469 2 of 15 faster, more efficiently, and with an appropriate granulometry is one of the main challenges facing this process in the mining industry—a challenge that has also led to the search for best practices in design and maintenance. As part of its continuous improvement plan, Gabriela Mistral has designed a study analyzing the development and implementation of an expert system in the crushing plant that allows the critical variables of the process to be controlled on rules-based decisions that operate to maintain critical plant variables in preset stability ranges [4]. The system mimics the behavior of a human operator to provide continuity, efficiency, and safety [5]. The application of this system has been analyzed in several studies (e.g., [6–11]), showing its usefulness in various operations. It should be noted that one of the major advantages of operating the expert system is that it has the ability to process large amounts of information consistently with an optimal performance by operating day and night, thus minimizing operational problems that may be common with manual operation. This has been verified in various works (e.g., References [12–15]), where significant improvements in work efficiency were achieved. Taking into account the above, the present work aims to increase the stacking performance of the plant by 7% and then contribute to operational continuity and a controlled process by standardizing the operation in different shifts, thus increasing the useful life of the equipment. To date, the Gabriela Mistral division crushing plant has been operating in manual mode with the assistance of room operators who have a thorough knowledge of the behavior of equipment that interacts directly with the plant. The experience of the operators is essential for designing optimal scheduling of the expert system to gather the best operational practices of different work shifts. It is worth mentioning that the system can improve over time, either with new operational maneuvers or with additional schedules, similar to what has been seen in three different studies (References [16–19]). The contribution of this work is that it applied methodological techniques for the development of the knowledge base in an industrial application of a multivariate and multiparametric process. In addition, yields in manual mode and expert mode were compared. The main purpose of this study was to validate the operation of the technology implementation. For this, the operating data using the expert system was compared with those obtained using the common method of operation with a worker. The tests consisted of analyzing the impact in the plant when operated in both modes, evaluating the operational variables, and presenting detentions. The focus of this study was on the knowledge base of the expert system, which expanded through interviews, questionnaires, and review of historical plant data among others to find optimum points of operation that could be used later to program the knowledge base in the K-UCN system (K. Arcos y C. Leiva, Antofagasta, Chile). The comparison of both modes of operation indicated significant differences with respect to the performance of the critical line. The operation with expert control yielded an additional 159 tons/h, which presented an improvement of 2.17% and a stacking opportunity of 1,044,630 tons/year. The number of detentions decreased by 55%, allowing operational continuity and stability of the critical variables of the system. 2. Materials and Methodology First, considering the study of the Gabriela Mistral Division production process, a survey was conducted of the conditions in which the plant was located in order to identify critical equipment in each area (secondary and tertiary crushing). In addition, a theoretical and bibliographic study of the Expert System 2013—which covers only the crushing plant—was carried out, evaluating both the operational parameters and the control logic used in order to analyze the behavior presented by the plant when it is operated automatically. At the same time, a list was made of the sensors required to operate with the expert system and their condition. With this, it was possible to design and operate the expert system K-UCN (Expert System Universidad Católica del Norte), which is the new system applied, as explained in this document, that seeks to improve the result obtained previously in the plant by a similar method. The system used previously (2013) consisted of a much smaller information base of knowledge than the K-UCN system, which uses a more extensive questionnaire to cover each Minerals 2018, 8, 469 3 of 15 situation presented in the plant operation. The Expert System 2013 failed to function properly in the plant, so it was replaced by the currently used manual operation system. We hypothesized that the performance of the K-UCN system would be greater than the system’s performances in manual mode. Alternatively, the performance of K-UCN system might be greater than the Expert System 2013. 2.1. Specifications of the Crushing Plant The mineral derived from both the open-cut and ground-based exploitation must be prepared in a crushing plant; then, if necessary, acid is added to achieve a controlled grain size to ensure a good permeability coefficient of the solution. Crushing is a unitary operation of size reduction by the application of compression forces whose main purpose is to carry out the necessary size reductions, until a product of a suitable granulometry is obtained that allows the development of leaching in stacks or deposits in an efficient way. The objective of crushing the ore is to produce particles of the appropriate size and shape for direct use, release valuable materials from the bargain so that they can be concentrated, and increase the surface area available for chemical reaction. The crushing process is performed in two large stages, which require specific equipment to achieve proper granulometry. First there is a primary crushing or coarse crushing to crush big rocks. Second, there is a secundary grinding stage to achieve the right size. For each of these stages, there is appropriate equipment with the following main characteristics: the opening of feed, capacity to different closures, sizes of products, power, etc. They are usually tabulated by the manufacturers of crushers based on density, grain size, and mineral hardness. 2.2. Instrumentation In order to start and monitor the expert control system, field instrument reliability is required; the variables that interact with the system are shown in Table 1. Table 1. Critical Variables for Expert System, Crushing Plant. Secondary Crusher Tertiary Crusher Strap Current [%] Current of tertiary crushers [%] Tonnage feed [ton/h] Inclined plane level [%] Secondary crushers current [%] Tertiary feed hopper level [%] Inclined plane [%] Current (Motors) [%] Tertiary feed hopper average level [%] These variables are defined because of the importance they have on reducing operational failures. In addition, the associated instrumentation presents reliability for the feedback required to take control. The ranges of these variables and the criterion of the rules used by the expert system can be seen in Tables 2 and 3. There are a large number of system-related variables that are of great importance in the process. However, because of the poor reliability of associated instrumentation and the lack of measurement equipment, they do not interact in the expert control but are rather monitored by the operator. 2.3. Working Methodology The methodology for the knowledge base develops the following activities: • selection of the expert and expert domain; • relationship with the expert and extraction of knowledge of the same, where special psychological techniques such as protocol analysis or interview techniques are of special importance. The professionals surveyed are shown in Table A1; Minerals 2018, 8, 469 4 of 15 • acquisition of knowledge by the engineer and selection of the appropriate representation technique; • selection of the most suitable tool for development. • construction of incremental prototypes. To design the philosophy of the K-UCN system, we started by taking interviews with operators and control room analysts in different shifts to form a knowledge base. This allowed us to define the operational variables, critical system variables, operating and prevailing rules, the dynamic time required by the system, and the analysis of the Expert System 2013. We performed tests by area and line as each of them presents different behaviors and therefore has different control logic. The behavior of the variables described below was evaluated: Secondary crush: (a) percentage of current in belts; (b) tonnage straps; (c) percentage level of inclined plane; (d) current percentage of secondary crushers; and (e) percentage of tertiary tolvines level (average). Tertiary crusher: (a) percentage level of inclined plane; (b) current percentage of tertiary crushers; (c) percentage level of tertiary feed hopper; (d) virtual tonnage; and (e) percentage level of fine silos (average). In order to validate the results obtained from the tests mentioned above, a statistical analysis was carried out, which consisted of sampling the performance of each type of operation and comparing the average obtained during the total working time; this allowed us to evaluate the behavior of the manual operation in order to implement improvements in the control logic. In addition, a hypothesis test was performed to compare the operation of the first semester in manual mode 2016 versus the Expert System K-UCN. This was to analyze the effect when the plant was operated with the expert system, mainly in relation to the performance of the critical line and the number of operational detentions. 2.4. Design of the Expert System K-UCN A knowledge base was formed by collecting information about different facts and experiences to analyze strategies used in different shifts. These were obtained through interviews, questionnaires, and direct conversations with analysts. The questionnaires were based on forms with predetermined questions, allowing the capture of abundant and basic information about the process and the problem under study. This allowed the formulation of initial hypotheses and guided the strategies for applying techniques of data collection on the most critical variables of the process. With respect to direct interviews, the experts responded extensively on the subject with the freedom to deviate from the interviewer’s objectives. This allowed us to learn how they faced and analyzed different situations occurring at the plant and the strategy they used to maintain operational stability, thereby creating a closer link with the experts. For the collection and analysis of data, we designed sheets linked to the PI System—a platform that allowed us to obtain historical data of the plant and monitor the behavior of the operational variables that interact directly with the process, thus validating the knowledge base. The distributed control system (DCS) used in the crushing plant is FOXBORO’s I/A Series System (Process & Machinery Control, Georgia, USA), which uses three software programs for the implementation of the expert system: Minerals 2018, 8, 469 5 of 15 Software A: Shell: IACC (Integrated Automation Control Configurator), a software used in the creation and modification of rules that interact in the system with a programming language structured text; Software B: FOXDRAW, where the screens are edited before the interaction with the operator; and Software C: HMI, a man–machine interface, FOXVIEW, where the operator interacts with the expert control. The main features of the K-UCN expert system are as follows: • It is implemented in Line 1, 2, and 3 of the secondary and tertiary crushing area. • The fine-tuning of the rules is independent for each line as they do not present the same behavior due to their location and the segregation that occurs in the stockpile. • Its main control element is the speed regulation of the secondary and tertiary feeders. • Its operating criteria are based on best operational practices. For each of the variables shown in Table 1, there is a range of values that determine the number of percentage points in which the speed of feeders is adjusted. The system is always controlled by the variable that is in a more critical value in order to maintain stability in the system. The main structures of expert control systems are shown in Figures 1 and 2 and the ranges are shown in Tables 2 and 3. Minerals 2018, 8, x FOR PEER REVIEW 5 of 15 Software C: HMI, a man–machine interface, FOXVIEW, where the operator interacts with the expert control. The main features of the K-UCN expert system are as follows: • It is implemented in Line 1, 2, and 3 of the secondary and tertiary crushing area. • The fine-tuning of the rules is independent for each line as they do not present the same behavior due to their location and the segregation that occurs in the stockpile. • Its main control element is the speed regulation of the secondary and tertiary feeders. • Its operating criteria are based on best operational practices. For each of the variables shown in Table 1, there is a range of values that determine the number of percentage points in which the speed of feeders is adjusted. The system is always controlled by the variable that is in a more critical value in order to maintain stability in the system. The main structures of expert control systems are shown in Figures 1 and 2 and the ranges are shown in Tables 2 and 3. Figure 1. Expert System K-UCN, Secondary Crusher. Figure 1. Expert System K-UCN, Secondary Crusher. Minerals 2018, 8, 469 6 of 15 Table 2. Established Operational Ranges: Secondary Crusher Line 1, Expert System K-UCN. Line 1 Rule 1 Rule 2 Rule 3 Rule 4 Rule 5 Belt Current Tonnage Crusher Current Inclined Plane Tertiary Feed Hopper Range (%) Point (%) Range (ton/h) Point (%) Range (%) Point (%) Range (%) Point (%) Range (%) Point (%) (0–70) 5.0 ((−2000)–(−300)) −5.0 (0–60) 5.0 (0–20) 5.0 (0–20) 5.0 (70.01–80) 4.0 ((−299.99)–(−68)) −2.0 (60.01–75) 4.0 (20.01–25) 5.0 (20.01–30) 4.0 (80.01–85) 3.0 ((−67.99)–(−30)) −0.5 (75.01–85) 2.0 (25.01–30) 0.0 (30.01–40) 2.0 (85.01–90) 2.0 ((−29.99)–30) 0.0 (85.01–90) 1.0 (30.01–40) −1.0 (40.01–50) 1.0 (90.01–94) 1.0 (30.01–68) 0.5 (90.01–95) 0.0 (40.01–50) −2.0 (50.01–55) 0.0 (94.01–96) 0.0 (68.01–150) 1.0 (95.01–98) −2.0 (50.01–100) −4.0 (55.01–60) −5.0 (96.01–98) −1.0 (150.01–300) 2.0 (98.01–102) −4.0 (60.01–65) −7.0 (98.01–100) −2.0 (300.01–500) 3.0 (102.01–110) −5.0 (65.01–100) −10.0 (100.01–110) −4.0 (500.01–800) 4.0 (110.01–120) −6.0 (110.01–120) −5.0 (800.01–2000) 5.0 Minerals 2018, 8, 469 7 of 15 Minerals 2018, 8, x FOR PEER REVIEW 7 of 15 Figure 2. Expert system K-UCN—tertiary crusher. Table 3. Established Operational Ranges: Tertiary Crusher Line 1, Expert System K-UCN. Line 1 Rule 1 Rule 2 Rule 3 Rule 4 Crusher Current Inclined Plane Feed Hopper Engine Current Range (%) Point (%) Range (%) Point (%) Range (%) Point (%) Range (%) Point (%) (0–60) 6.0 (0–20) 6.0 (0–10) −5.0 (0–75) 6.0 (60.01–70) 5.0 (20.01–25) 3.0 (10.01–15) −4.0 (75.01–85) 5.0 (70.01–87) 2.0 (25.01–30) 0.0 (15.01–20) −3.0 (85.01–90) 3.0 (87.01–95) 0.0 (30.01–40) −1.0 (20.01–25) −2.0 (90.01–93) 2.0 (95.01–98) −1.0 (40.01–50) −2.0 (25.01–30) −1.0 (93.01–95) 1.0 (98.01–100) −2.0 (50.01–100) −6.0 (30.01–35) 0.0 (95.01–99) 0.0 (100.01–105) −3.0 (35.01–40) 2.0 (99.01–101) −1.0 (105,01–115) −4.0 (40.01–45) 3.0 (101.01–103) −2.0 (115.01–200) −5.0 (45.01–55) 4.0 (103.01–110) −3.5 (55.01–60) 5.0 (110.01–120) −5.0 (60.01–100) 6.0 Each rule makes a decision according to the variables controlled in each section depending on the interval they are in. Each variable that interacts in the expert system presents an associated dynamic time, i.e., the time it takes to make an impact when making a change, as shown in Tables 4 and 5. This time was obtained from experimental tests in the plant by sampling each independent variable to obtain the median of the data. Table 4. Dynamic times K-UCN, Secondary Crusher. Dynamic Time Line 1 Rules Time (s) Current 5 Tonnage 5 Engine Current 50 Inclined plane level 50 Feed hopper average level 90 Figure 2. Expert system K-UCN—tertiary crusher. Table 3. Established Operational Ranges: Tertiary Crusher Line 1, Expert System K-UCN. Line 1 Rule 1 Rule 2 Rule 3 Rule 4 Crusher Current Inclined Plane Feed Hopper Engine Current Range (%) Point (%) Range (%) Point (%) Range (%) Point (%) Range (%) Point (%) (0–60) 6.0 (0–20) 6.0 (0–10) −5.0 (0–75) 6.0 (60.01–70) 5.0 (20.01–25) 3.0 (10.01–15) −4.0 (75.01–85) 5.0 (70.01–87) 2.0 (25.01–30) 0.0 (15.01–20) −3.0 (85.01–90) 3.0 (87.01–95) 0.0 (30.01–40) −1.0 (20.01–25) −2.0 (90.01–93) 2.0 (95.01–98) −1.0 (40.01–50) −2.0 (25.01–30) −1.0 (93.01–95) 1.0 (98.01–100) −2.0 (50.01–100) −6.0 (30.01–35) 0.0 (95.01–99) 0.0 (100.01–105) −3.0 (35.01–40) 2.0 (99.01–101) −1.0 (105,01–115) −4.0 (40.01–45) 3.0 (101.01–103) −2.0 (115.01–200) −5.0 (45.01–55) 4.0 (103.01–110) −3.5 (55.01–60) 5.0 (110.01–120) −5.0 (60.01–100) 6.0 Each rule makes a decision according to the variables controlled in each section depending on the interval they are in. Each variable that interacts in the expert system presents an associated dynamic time, i.e., the time it takes to make an impact when making a change, as shown in Tables 4 and 5. This time was obtained from experimental tests in the plant by sampling each independent variable to obtain the median of the data. Table 4. Dynamic times K-UCN, Secondary Crusher. Dynamic Time Line 1 Rules Time (s) Current 5 Tonnage 5 Engine Current 50 Inclined plane level 50 Feed hopper average level 90 Minerals 2018, 8, 469 8 of 15 Table 5. Dynamic times, tertiary crushing. Rules Time (s) Current 40 Inclined plane level 30 Feed hopper level 30 Engine current 65 To avoid a lack of control in the operational variables, restrictions were programmed as shown in Table 6. Table 6. Secondary and Tertiary Crushing Restrictions. Crushers Current Time (s) Action Current greater than 100% 15 Decrease 9% Current greater than 110% 8 Decrease 10% Inclined plane Level greater than 50% 1 Decrease 15% In addition, the expert system considers edge conditions that limit the speed of the secondary feeder, which is found in the three secondary crushing lines. • If one of the tertiary feeders is stopped and the average tolvin exceeds 40% of the level for a time of 3 s, the speed of the secondary feeders must decrease by 40% of the set value. • If both tertiary feeders are stopped and the tolvin average is higher than 45% of their level, the speed of the tertiary feeder or feeders should decrease by 40%. Before any detention occurs during the use of the expert system, the control will begin to act 50 s after having given a condition to the feeder, allowing the line to stabilize; then, it will begin to maximize the variables. In case of an interlock, the feeder will start only after the alarm has been acknowledged by the analyst, and the same procedure will be performed. The expert system considers edge conditions that limit the velocity of the tertiary feeder to any event occurring in the secondary crushing. This is applied for the six tertiary lines. • The line is operative with a single secondary feeder with the contribution greater than 95%. If it stops and the level of the feed hopper is less than 50%, the speed of the tertiary feeders decreases by 40% of the set value. • The line is operative with the two secondary feeders. If only one of the feeders stops and the level of the feed hopper is less than 50%, the tertiary feeders speed decreases by 20% of the set value. • The line is operative with the two secondary feeders. Both stop when the level of the feed hopper is less than 50%, and the tertiary feeders speed decreases by 40% of the set value. Unlike the secondary crushing, the expert system in the tertiary crushing plant operates with a stability band. This means that if the variables are within the ranges described in Table 7, it is activated above the decision rules presented by the expert control in order to maintain the stability of the variables, imitating the action performed by the analyst of the control room when it cannot continue to maximize performance. If one of these variables is outside of the ranges, they start to control again according to the programmed rules. Minerals 2018, 8, 469 9 of 15 Table 7. Stability Band Ranges. Variable Condition Tertiary crusher current 75–96% Inclined plane level <40% Tertiary feed hopper level 25–50% Engine current <101% 2.5. Design of the Tests in the Crushing Plant The tests were designed to mainly evaluate the behavior and operational impact caused in the crushing plant when operated with the expert control system so that it can be implemented in the Gabriela Mistral Division. Currently, the crushing plant operates in manual mode, and to review the behavior of the expert control system, three lines of the secondary crushing plant and six lines of the tertiary crushing plant were made available, with each line operating with two crushing lines. The evaluation process considered the following stages: • Reconstruction of logic: This first stage corresponded to the analysis of existing logic, created for the Expert System 2013, with the aim of analyzing the operation of the system for that year. During this period, data was collected from the first half of 2016 to create the knowledge base. • Parameter adjustments: During this stage, based on the knowledge base already created, the new operating strategy was designed for the K-UCN expert system, which recorded data without evaluation for the preliminary tests as it was treated of a net stage of operational adjustments. • Feedback: During this period, the actual evaluation was carried out by taking a careful record of the conditions and operational variables. By means of feedback with the different shifts, fine adjustments were made to the parameters. For this stage, a minimum of three months of data was considered necessary so that the evaluation could be done under different operational conditions and type of mineral fed. 3. Results and Discussions 3.1. Critical Line Performance (KPI) A comparison was made of the stacked performance—for manual operation versus expert system operation. Table 8 shows the hypothesis test with two independent samples performed in Minitab. Here, we observed that the mean of the operation with expert control was greater than that of the manual operation at a significance level of 0.05, representing a difference of 159 ton/h in favor of the operation with the expert system. Table 8. Hypothesis test: manual operation vs. operation expert system. Individual Samples Statistics Expert Manual Sample size 5531 12,001 Average 7353.9 7194.8 Confidence interval of 90% (7344, 7364) (7187.3, 7202.2) Standard deviation 444.55 493.53 An increase of 2.17% in the amount of material processed could be seen, which followed the trend of other works where the use of an expert system generated a benefit in the processes. In Chen et al. [12], an increase of 10% was reported in the capacity of feeding; in Bouffard [6], a benefit of between 1 and 16% increase in processing capacity and 1% increase in recovery between various flotation and milling Minerals 2018, 8, 469 10 of 15 plants was achieved. Figure 3 shows the average performance for both the operation with an expert system of 7354 ton/h and a manual mode with 7195 ton/h. The behavior presented in the data showed a similar coefficient of variation, with a difference of 1% (6% versus 7%). The clear difference was the frequency of the yields obtained, with the expert system K-UCN showing a higher frequency in yields closer to 7500 ton/h (target crushed plant). Minerals 2018, 8, x FOR PEER REVIEW 10 of 15 Figure 3. Data Rate: Expert System vs. Manual Operation. In turn, the behavior of four shifts in both modalities was analyzed, as can be seen in the Table 9 and Figures 4 and 5. Here, the mean of expert system was greater than that of the manual system at the significance level of 0.05. The uncertainty quantified was associated with the estimation of the difference in means from the sample data, getting confidence intervals with 90% certainty that the true difference was between 149.33 and 194.67. Table 9. Turn-based performance in both modes. Mode Turn 1 (T1) Turn 2 (T2) Turn 3 (T3) Turn 4 (T4) Expert 7.263 ton/h 7.376 ton/h 7.402 ton/h 7.409 ton/h Manual 7.340 ton/h 7.215 ton/h 7.067 ton/h 7.034 ton/h Figure 4. Anova hypothesis test, yields by turns in both modalities. Figure 3. Data Rate: Expert System vs. Manual Operation. In turn, the behavior of four shifts in both modalities was analyzed, as can be seen in the Table 9 and Figures 4 and 5. Here, the mean of expert system was greater than that of the manual system at the significance level of 0.05. The uncertainty quantified was associated with the estimation of the difference in means from the sample data, getting confidence intervals with 90% certainty that the true difference was between 149.33 and 194.67. Table 9. Turn-based performance in both modes. Mode Turn 1 (T1) Turn 2 (T2) Turn 3 (T3) Turn 4 (T4) Expert 7.263 ton/h 7.376 ton/h 7.402 ton/h 7.409 ton/h Manual 7.340 ton/h 7.215 ton/h 7.067 ton/h 7.034 ton/h Minerals 2018, 8, x FOR PEER REVIEW 10 of 15 Figure 3. Data Rate: Expert System vs. Manual Operation. In turn, the behavior of four shifts in both modalities was analyzed, as can be seen in the Table 9 and Figures 4 and 5. Here, the mean of expert system was greater than that of the manual system at the significance level of 0.05. The uncertainty quantified was associated with the estimation of the difference in means from the sample data, getting confidence intervals with 90% certainty that the true difference was between 149.33 and 194.67. Table 9. Turn-based performance in both modes. Mode Turn 1 (T1) Turn 2 (T2) Turn 3 (T3) Turn 4 (T4) Expert 7.263 ton/h 7.376 ton/h 7.402 ton/h 7.409 ton/h Manual 7.340 ton/h 7.215 ton/h 7.067 ton/h 7.034 ton/h Figure 4. Anova hypothesis test, yields by turns in both modalities. Figure 4. Anova hypothesis test, yields by turns in both modalities. Minerals 2018, 8, 469 11 of 15 Minerals 2018, 8, x FOR PEER REVIEW 11 of 15 Figure 5. Frequency of data, both modalities by turns. It was observed that the best shifts that operate the plant in manual mode had an inverse behavior to the operation of the expert mode, indicating that a better control of the system—and of the variables that are directly associated in the system—is required. This presents an opportunity for improvement in the standardization of the plant operation plant. 3.2. Comparison between Expert Systems 2013 and 2016 K-UCN In Table 10, it can be observed that the average performance in operation with K-UCN expert control was higher by 3% compared to the Expert System 2013, with a difference of approximately 200 ton/h. Taking into account the uncertainty associated with the estimation of the difference between the means from the sampled data, we had 90% certainty that the true difference was between 182 and 215. Table 10. Hypothesis test, expert system K-UCN vs. Expert System 2013. Individual Samples Statistics Expert 2016 Expert 2013 Sample Size 5531 3114 Average 7353.9 7155.2 90% confidence interval (7344, 7364) (7142.2, 7168.2) Standard deviation 444.55 441.30 The K-UCN expert system presented a more controlled behavior—the trend was closer to the required target of 7500 ton/h as opposed to the Expert System 2013 in which the behavior of the performance was uniform in a wider range (Figure 6). Figure 5. Frequency of data, both modalities by turns. It was observed that the best shifts that operate the plant in manual mode had an inverse behavior to the operation of the expert mode, indicating that a better control of the system—and of the variables that are directly associated in the system—is required. This presents an opportunity for improvement in the standardization of the plant operation plant. 3.2. Comparison between Expert Systems 2013 and 2016 K-UCN In Table 10, it can be observed that the average performance in operation with K-UCN expert control was higher by 3% compared to the Expert System 2013, with a difference of approximately 200 ton/h. Taking into account the uncertainty associated with the estimation of the difference between the means from the sampled data, we had 90% certainty that the true difference was between 182 and 215. Table 10. Hypothesis test, expert system K-UCN vs. Expert System 2013. Individual Samples Statistics Expert 2016 Expert 2013 Sample Size 5531 3114 Average 7353.9 7155.2 90% confidence interval (7344, 7364) (7142.2, 7168.2) Standard deviation 444.55 441.30 The K-UCN expert system presented a more controlled behavior—the trend was closer to the required target of 7500 ton/h as opposed to the Expert System 2013 in which the behavior of the performance was uniform in a wider range (Figure 6). Minerals 2018, 8, 469 12 of 15Minerals 2018, 8, x FOR PEER REVIEW 12 of 15 Figure 6. Frequency of data between expert systems. 3.3. Plant Detentions During the testing period, the number of detentions associated with the following variables that directly impacted the process and the expert control were monitored (Figure 7): • overcurrent of the crushers; • high level thin silos; • low-level feed hopper; • inclined plane; and • extreme misalignment. Figure 7. Daily detentions in both modalities, by lines. While evaluating the detentions of the complete plant, it was observed that in manual operation, 55% more detentions were present than when operating the plant with the expert system (22 versus 49). This shows that the implementation and use of the expert system presents a significant improvement opportunity. This was also seen in Fan et al. [7], where the application of a control system to support the existing one increased the stability of the process without altering production. The greatest percentage of the detentions was due to the extreme misalignment that occurred in the feeders and the low level of the feed hopper because this would increase the odds that the feeders became misaligned by the blows of the load in them. 4. Concluding Remark and Future Work In this paper, it has been demonstrated, at an industrial level, that it is feasible to implement the expert control system in the crushing plant as it presents advantages in productivity and improved quality from a cost–benefit point of view, in addition to maintaining a controlled and standardized process. Figure 6. Frequency of data between expert systems. 3.3. Plant Detentions During the testing period, the number of detentions associated with the following variables that directly impacted the process and the expert control were monitored (Figure 7): • overcurrent of the crushers; • high level thin silos; • low-level feed hopper; • inclined plane; and • extreme misalignment. Minerals 2018, 8, x FOR PEER REVIEW 12 of 15 Figure 6. Frequency of data between expert systems. 3.3. Plant Detentions During the testing period, the number of detentions associated with the following variables that directly impacted the process and the expert control were monitored (Figure 7): • overcurrent of the crushers; • high level thin silos; • low-level feed hopper; • inclined plane; and • extreme misalignment. Figure 7. Daily detentions in both modalities, by lines. While evaluating the detentions of the complete plant, it was observed that in manual operation, 55% more detentions were present than when operating the plant with the expert system (22 versus 49). This shows that the implementation and use of the expert system presents a significant improvement opportunity. This was also seen in Fan et al. [7], where the application of a control system to support the existing one increased the stability of the process without altering production. The greatest percentage of the detentions was due to the extreme misalignment that occurred in the feeders and the low level of the feed hopper because this would increase the odds that the feeders became misaligned by the blows of the load in them. 4. Concluding Remark and Future Work In this paper, it has been demonstrated, at an industrial level, that it is feasible to implement the expert control system in the crushing plant as it presents advantages in productivity and improved quality from a cost–benefit point of view, in addition to maintaining a controlled and standardized process. Figure 7. Daily detentions in both modalities, by lines. While evaluating the detentions of the complete plant, it was observed that in manual operation, 55% more detentions were present than when operating the plant with the expert system (22 versus 49). This shows that the implementation and use of the expert system presents a significant improvement opportunity. This was also seen in Fan et al. [7], where the application of a control system to support the existing one increased the stability of the process without altering production. The greatest percentage of the detentions was due to the extreme misalignment that occurred in the feeders and the low level of the feed hopper because this would increase the odds that the feeders became misaligned by the blows of the load in them. 4. Concluding Remark and Future Work In this paper, it has been demonstrated, at an industrial level, that it is feasible to implement the expert control system in the crushing plant as it presents advantages in productivity and Minerals 2018, 8, 469 13 of 15 improved quality from a cost–benefit point of view, in addition to maintaining a controlled and standardized process. The comparison of both modes of operation indicates significant differences with respect to the performance of the critical line. The operation with expert control yielded an additional 159 tons/h, which was an improvement of 2.17% and a stacking opportunity of 1,044,630 ton/y. The number of detentions decreased by 55%, allowing operational continuity and stability of the critical variables of the system. As for the behavior of the shifts in the operation of the plant, it was observed that the shifts that resulted in the best practices while operating in manual mode showed an inverse behavior to operating in expert control. This is probably due to strong paradigms and the reluctance of using a new system. Currently, this problem is being dealt with by training for shifts that show major weaknesses. The development of the incremental prototype significantly improved plant performance and unscheduled detention. The authors of this paper consider the followings topics for future research: • the implementation of the acidification process in the K-UCN system; • analysis of the possibility of using the inference engine of the expert system with fuzzy logic. Author Contributions: K.V.A. and C.A.L. conceived, designed, and performed the experiments; all the authors analyzed the data; K.V.A. also contributed the software that belongs to Gabriela Mistral Company; C.A.L., C.M.T., Y.G., and E.A.S. wrote the paper. Funding: This research received no external funding. Acknowledgments: The authors wish to acknowledge the material support provided by the Gabriela Mistral Division, Codelco and financial support provided by Universidad Católica del Norte. Conflicts of Interest: The authors declare no conflict of interest. Appendix A The expert system has been designed using interviews and questionnaires considering three professional positions (operator, process engineer, and senior engineer) as show in Table A1. Table A1. Number of professionals surveyed. Questions Operator Process Engineer Senior Engineer 1–10 8 2 1 11–20 8 2 1 21–30 8 2 1 31–42 8 2 1 The purpose of developing the questionnaire was to obtain information on how to operate the crushing plant. In order to update the knowledge base on the expert control system. The following questions were considered: 1. At what level of the stockpile should you begin to feed the plant? 2. What criterion is used to decide which feeder to operate—only one or both? 3. What is the optimum operating current of the secondary belts CV-002, 003, and 004? 4. How does the current of the belt relate to the performance of each line in secondary crushing? 5. What is the factor of the CV-002, 003, and 004 strap gauges used for? Is it necessary? 6. What is the optimal level of operating inclination of the secondary crushers? 7. What should be done when the inclined plane starts to rise? What is the percentage of filling where action should be taken? 8. What is the optimal rate of operation of the secondary crushers? Minerals 2018, 8, 469 14 of 15 9. What should be done when the rate level from the secondary crushers increases or decreases? Indicate ranges 10. What is the optimal operating current of the secondary crushers? 11. Define high operating current in secondary crushers (ranges)—what should be done when currents are high? 12. Define low operating current in secondary crushers (ranges)—what should be done when currents are low? 13. Why do the crushers not operate with the currents to the limit (100%)? 14. What is the optimal level of the tertiary hopper for a secondary and tertiary feeder? 15. What should be done when the tertiary hoppers are full and empty? At what high hopper level should action be taken and what action should be taken? 16. What is the optimum current of the tertiary crushers? 17. Define high operating current for tertiary crushers (ranges)—what should be done when currents are high? 18. Define low operating current for tertiary crushers (ranges)—what should be done when currents are low? 19. What is the optimum operating pressure of the crushers? Is it the same for secondary and tertiary? 20. What should be done when the crushing pressure increases and when it decreases (ranges) for both secondary and tertiary? 21. What is the optimum vibration of the crushers? Is it the same for secondary and tertiary? 22. What should be done when the vibration of the crushers increases and decreases (ranges) for both secondary and tertiary? 23. What should be done when there is high temperature in the crusher? Is the temperature range operated by the crushers the same for secondary and tertiary? 24. What should be done when there is low temperature in the crusher? Is the temperature range operated by the crushers the same for secondary and tertiary? 25. Why is ring hopping caused in the crushers? Is it the same for secondary and tertiary crushing? What action is taken when this occurs? 26. What are the setting ranges in the secondary and tertiary crushers? 27. What steps should be followed to perform the measurement of the setting? 28. Evaluation to modify setting—in which case should the crusher be opened or closed? 29. What is the optimum percentage that the CV-010 and CV-011 belt should operate? What is the minimum and maximum percentage that both belts must operate? 30. According to the virtual tonnage of CV-012, should any action be taken? 31. Optimal level of fine silos—what action is taken when these are low and when they are high (ranges)? 32. What should be done when the belt current is optimal but the secondary crusher current is low? 33. What should be done when the belt current is low but the secondary crusher current is high? 34. What is the strategy to work with 5/6? 35. What should be done when the crushers are worn out? Is the same action taken for both the secondary and tertiary crushers? 36. What should be done when there is very hard or over-sized ore? 37. What should be done when there is thin material present? 38. What should be done in the presence of wet material? 39. What parameters or variables should be evaluated to increase tonnage? 40. What should be done when a secondary line is stopped? 41. What should be done when a tertiary line is stopped? 42. What should be done when the silo B or drum 02 is in maintenance? Minerals 2018, 8, 469 15 of 15 References 1. Bergh, L.G.; Yianatos, J.B.; Leiva, C.A. Fuzzy supervisory control of flotation columns. Miner. Eng. 1998, 11, 739–748. [CrossRef] 2. Durkin, J.; Durkin, J. Expert Systems: Design and Development; Prentice Hall PTR: New York, NY, USA, 1998. 3. Wei, D.; Craig, I.K. Grinding mill circuits—A survey of control and economic concerns. Int. J. Miner. Process. 2009, 90, 56–66. [CrossRef] 4. Lucas, P.; Van Der Gaag, L. Principles of Expert Systems; Addision & Wesley: Boston, MA, USA, 1991. 5. Parsaye, K.; Chignell, M. Intelligent Database Tools & Applications: Hyperinformation Access, Data Quality, Visualization, Automatic Discovery; Wiley: New York, NY, USA, 1993. 6. Bouffard, S.C. Benefits of process control systems in mineral processing grinding circuits. Miner. Eng. 2015, 79, 139–142. [CrossRef] 7. Fan, G.Q.; Rees, N.W. An intelligent expert system (KBOSS) for power plant coal mill supervision and control. Control. Eng. Pract. 1997, 5, 101–108. [CrossRef] 8. Iqbal, A.; He, N.; Li, L.; Dar, N.U. A fuzzy expert system for optimizing parameters and predicting performance measures in hard-milling process. Expert Syst. Appl. 2007, 32, 1020–1027. [CrossRef] 9. Torres-treviño, L.M.; Escamilla-salazar, I.G.; González-Ortíz, B.; Praga-Alejo, R. An expert system for setting parameters in machining processes. Expert Syst. Appl. 2013, 40, 6877–6884. [CrossRef] 10. Gao, Y.; Liu, Y.; Wang, C.; Li, X.; Ou, G. Design and evaluation of a high performance distributed expert system (HPDES) for aerospace ground verification system. Procedia Comput. Sci. 2012, 9, 1380–1389. [CrossRef] 11. Taylor, N.K. An Expert System to Assist in Design. Ph.D. Thesis, University of Nottingham, Nottingham, UK, 1990. 12. Chen, X.; Li, Q.; Fei, S. Supervisory expert control for ball mill grinding circuits. Expert Syst. Appl. 2008, 34, 1877–1885. [CrossRef] 13. Zhou, P.; Lu, S.; Yuan, M.; Chai, T. Survey on higher-level advanced control for grinding circuits operation. Powder Technol. 2016, 288, 324–338. [CrossRef] 14. Festa, A.; Cornejo, F.; Orrante, F.; Alanís, R.; Gutiérrez, B. Expert System Implementation at Peñoles Group Concentrators; Technical Report; SGS Minerals Services: Geneva, Switzerland, 2009. 15. Morgan, G.; Cheng, R.; Altintas, Y.; Ridgway, K. An expert troubleshooting system for the milling process. Int. J. Mach. Tools Manuf. 2007, 47, 1417–1425. [CrossRef] 16. Chen, X.; Li, S.; Zhai, J.; Li, Q. Expert system based adaptive dynamic matrix control for ball mill grinding circuit. Expert Syst. Appl. 2009, 36, 716–723. [CrossRef] 17. Iqbal, A.; He, N.; Dar, N.U.; Li, L. Comparison of fuzzy expert system based strategies of offline and online estimation of flank wear in hard milling process. Expert Syst. Appl. 2007, 33, 61–66. [CrossRef] 18. Rubio, L.; De, M.; Longstaff, A.P.; Fletcher, S.; De La Sen, M.; Longstaff, A.P.; Fletcher, S.; De, M.; Longstaff, A.P.; Fletcher, S. Expert Systems with Applications Model-based expert system to automatically adapt milling forces in Pareto optimal multi-objective working points. Expert Syst. Appl. 2013, 40, 2312–2322. [CrossRef] 19. Leiva, C.A.; Flores, V.M.; Salgado, F.; Poblete, D.; Acuña, C. Applying softcomputing for copper recovery in leaching process. Sci. Program. 2017, 2017, 6. [CrossRef] © 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/). http://dx.doi.org/10.1016/S0892-6875(98)00059-4 http://dx.doi.org/10.1016/j.minpro.2008.10.009 http://dx.doi.org/10.1016/j.mineng.2015.06.006 http://dx.doi.org/10.1016/S0967-0661(96)00213-4 http://dx.doi.org/10.1016/j.eswa.2006.02.003 http://dx.doi.org/10.1016/j.eswa.2013.06.051 http://dx.doi.org/10.1016/j.procs.2012.04.152 http://dx.doi.org/10.1016/j.eswa.2007.02.013 http://dx.doi.org/10.1016/j.powtec.2015.11.010 http://dx.doi.org/10.1016/j.ijmachtools.2006.09.019 http://dx.doi.org/10.1016/j.eswa.2007.10.008 http://dx.doi.org/10.1016/j.eswa.2006.04.003 http://dx.doi.org/10.1016/j.eswa.2012.10.034 http://dx.doi.org/10.1155/2017/6459582 http://creativecommons.org/ http://creativecommons.org/licenses/by/4.0/. Introduction Materials and Methodology Specifications of the Crushing Plant Instrumentation Working Methodology Design of the Expert System K-UCN Design of the Tests in the Crushing Plant Results and Discussions Critical Line Performance (KPI) Comparison between Expert Systems 2013 and 2016 K-UCN Plant Detentions Concluding Remark and Future Work References 
work_b3cby3ftcncclfcju2jjbvnoga	----	A Study Regarding the Use of Expert Systems in Economics Field Procedia Economics and Finance 6 ( 2013 ) 385 – 391 2212-5671 © 2013 The Authors. Published by Elsevier B.V. Selection and peer-review under responsibility of Faculty of Economic Sciences, Lucian Blaga University of Sibiu. doi: 10.1016/S2212-5671(13)00152-4 ScienceDirect Available online at www.sciencedirect.com © 2013 The Authors. Published by Elsevier B.V. Selection and peer-review under responsibility of Faculty of Economic Sciences, Lucian Blaga University of Sibiu. Open access under CC BY-NC-ND license. Open access under CC BY-NC-ND license. http://creativecommons.org/licenses/by-nc-nd/3.0/ http://creativecommons.org/licenses/by-nc-nd/3.0/ 386 Claudiu-Leonardo Stoia / Procedia Economics and Finance 6 ( 2013 ) 385 – 391 387 Claudiu-Leonardo Stoia / Procedia Economics and Finance 6 ( 2013 ) 385 – 391 • • • • • 388 Claudiu-Leonardo Stoia / Procedia Economics and Finance 6 ( 2013 ) 385 – 391 389 Claudiu-Leonardo Stoia / Procedia Economics and Finance 6 ( 2013 ) 385 – 391 390 Claudiu-Leonardo Stoia / Procedia Economics and Finance 6 ( 2013 ) 385 – 391 • • • • 391 Claudiu-Leonardo Stoia / Procedia Economics and Finance 6 ( 2013 ) 385 – 391 
work_b3drjz6z2rhdjcak5ekjjx2qwi	----	Calibration of microsimulation traffic model using neural network approach Expert Systems with Applications 40 (2013) 5965–5974 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Calibration of microsimulation traffic model using neural network approach 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.05.003 ⇑ Corresponding author. Tel.: +385 915220819. E-mail addresses: iirena@gfos.hr (I. Ištoka Otković), tomaz.tollazzi@um.si (T. Tollazzi), matjaz.Sraml@um.si (M. Šraml). Irena Ištoka Otković a,⇑, Tomaž Tollazzi b, Matjaž Šraml b a J.J. Strossmayer University of Osijek, Faculty of Civil Engineering, Drinska 16a, 31000 Osijek, Croatia b University of Maribor, Faculty of Civil Engineering, Smetanova 17, 2000 Maribor, Slovenia a r t i c l e i n f o Keywords: Calibration Microsimulation traffic model VISSIM Neural network Roundabout Validation a b s t r a c t This paper presents the results of research on the applicability of neural networks in the process of com- puter calibration of a microsimulation traffic model. VISSIM microsimulation model is used for calibra- tion done at the example of roundabouts in an urban area. The calibration method is based on the prediction of a neural network for one traffic indicator, i.e. for the traveling time between measuring points. Besides the traveling time, the calibration process further/also involves a comparison between the modeled and measured queue parameters at the entrance to the intersection. The process of valida- tion includes an analysis of traveling time and queue parameters on new sets of data gathered both at the modeled and at a new roundabout. A comparison of the traffic indicators measured in the field and those simulated with the calibrated and uncalibrated microsimulation traffic model provides an insight into the performance of the calibration procedure. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction The functioning of a traffic system is under the influence of var- ious aspects of human behavior (Onieva, Milanés, Villagrá, Pérez, & Godoy, 2012). Studies show that the behavior of traffic participants is, among other things, territorially and culturally conditioned (Ol- stam & Tapani, 2004). Microsimulation models include variable behavior of drivers at a level of each particular entity, and the real- ity of modeling results depends on the initial choice of the model (Fang & Elefteriadou, 2005) and success of the calibration process (Transportation Research Bord, 2000). Calibration of a traffic model is actually an adjustment of a model to a specific local traffic network and its users. Calibration is defined as the process of comparing and minimizing the differ- ences between the modeling results and the real data obtained by counting and measuring in the local network (Transportation Research Bord, 2000). Validation of the model is an estimate of the success of the model calibration made through a comparison of indicators obtained from the calibrated model and the ones measured in traffic. Analysis of bigger spacial and time scopes of a traffic network and prediction of a future state both need a larger number of models, including traffic demand models (Ettema, Arentze, & Timmermans, 2011), dynamic traffic distribution (Flötteröd, Chen & Nagel, 2012), daily mobility (Flötteröd, Bierlaire, & Nagel, 2011), flow prediction under typical and atypical traffic conditions (Castro-Neto, Jeong, Jeong, & Han, 2009) etc. It is realistic to expect that modeling of the selection of a travel mode and users’ routes is connected with the modeling of traffic (Park & Kim, 2001; Rickert & Nagel, 2001), economic and environmental indicators, both for real time modeling and for the prediction of the future traffic demand and its consequences. The degree of reliability of results of model- ing a future state of the system is a subject to a debate (Dorothy, Ambadipudi, & Kill, 2006; Stevanović & Martin, 2008), but scien- tific efforts are directed towards an adoption of reliability criteria in selecting tools in planning and optimization processes. Prediction of selection of users’ routs on the microsimulation le- vel (Flötteröd et al., 2011) and calibrations of traffic distribution model are not included. In this paper, the selection of users’ routs, which is the traffic distribution in an examined segment of the traffic network, is a user defined input parameter of the microsim- ulation model. The aim of this paper is to obtain a calibrated micro- simulation model which is able to give good correlation with measured functional characteristics of a particular segment of the network in real traffic conditions. Genetic algorithm (GA) is the most commonly used calibration algorithm for input parameters of the simulation model. Positive utilization experiences of GA are reported in the following exam- ples: in the calibration of FRESIM model (Cheu, Jin, Ng, Ng, & Srinivasan, 1998), PARAMICS model, VISSIM and CORSIM model (Park, Won, & Yun, 2006) and in the calibration of a VISSIM model (Kim, 2006). Analysis of acceptable time gaps and determination of http://crossmark.dyndns.org/dialog/?doi=10.1016/j.eswa.2013.05.003&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.05.003 mailto:iirena@gfos.hr mailto:tomaz.tollazzi@um.si mailto:matjaz.Sraml@um.si http://dx.doi.org/10.1016/j.eswa.2013.05.003 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. The first (a) and the second (b) examined roundabout. 5966 I. Ištoka Otković et al. / Expert Systems with Applications 40 (2013) 5965–5974 a critical time interval by means of Greenshields’ model (Cicu, Illotta, Bared, & Isebrands, 2011) represent a description of the cal- ibration method of the VISSIM microsimulation model at round- abouts in New York. For the purpose of calibrating a microsimulation traffic model, a new calibration method is analyzed. This paper shows results ob- tained by applying neural networks in the microsimulation traffic model calibration process. The first traffic indicator, chosen in the calibration process, is the traveling time between measuring points. Measuring points were chosen in the real environment and connected to existing ob- jects – a light pole at the beginning of the section and an electric supply pole at the end, because it was easier to enter them into the model layout that way. The traveling time is easily measurable in the field, and it is under the influence of the variable driver’s reaction time. In this paper, the mean traveling time is used in or- der to test the response of neural networks for the given indicator. Furthermore, another easily measurable traffic indicator – queue parameters – is introduced in the calibration process. Validation is done for the traveling time and queue parameters for new sets of measured data. The findings of the comparison of modeled and measured data of the microsimulation traffic model, calibrated to one roundabout, are also shown, and the validation is done on another roundabout. Prediction of traveling time, obtained by application of neural network (Park et al., 1999; Dia, 2001) can be analyzed in a broader context of traffic modeling, but a special emphasis is, in this case, used in microsimulation modeling. The microsimulation model, selected to investigate the applica- bility of neural networks in the calibration process, is the VISSIM.1 Analyses are performed on the basis of results obtained by one-hour traffic simulations, corresponding to the duration of the data gather- ing in the field for every set of data measured in real traffic conditions. 2. Data gathering As the basis for testing the neural network applicability in the process of calibration, the one-lane urban roundabout Vink- ovačka–Drinska in Osijek (Croatia) – roundabout I (Fig. 1a) – was used. The calibration process was done in two steps. The first step is based on the traveling time prediction obtained from the neural network, and the other step introduces queue parameters into the analysis. Maximum queue length and the number of vehicle stop- ping at the entrance into the roundabout, both easily collected un- 1 VISSIM is a stochastic, discrete, micro-simulated (models each entity separately) model designed for traffic analyses. It started to develop in Germany at the University of Karlshruhe in the early 70-ies of the last century. der real traffic conditions, are selected. An estimate of the success of the calibration process is made iteratively also in two steps. The first validation is done at the same roundabout, but for new sets of data, and the second validation is done at a new one-lane round- about Kirova–Opatijska in Osijek – roundabout II (Fig. 1b). Data was gathered, during 2010, at both roundabouts by means of one-hour video camera recordings (Table 1). Both points for measuring the traveling time at both roundabouts are included in the video. Detailed data on the traffic load of the roundabouts for all sets of measured data are available in the dissertation (Ištoka Otković, 2011).2 The results of measured traffic indicators for all sets of mea- sured data are shown in the Table 1 with the calibration data set shaded. The other sets of measured data were used for validation of the calibrated model. The traffic load counted on March 3rd 2010, between 3 and 4 p.m., shows that nearly 1600 vehicles, overall, entered the roundabout. An hour later, one hundred vehicles less en- tered the roundabout. Measuring, done on July 14th 2010, shows that 250 less vehicles entered the roundabout, which reflected on the traffic indicators of the roundabout. Traffic loads of pedestrian flows are not large and this should be taken into account when the results of modeling are interpreted. 3. Calibration of microsimulation model using neural network approach Microsimulation models have a considerable number of model input parameters. Examination of all combinations of input param- eters of the model in their range by means of application of micro- simulation model, in order to find a combination of input parameters that approximates the real traffic situation the best, would be an extremely time-consuming task. A computer can examine a large number of combinations of val- ues of model input parameters in real time, in case that it can ob- tain software simulation output values (e.g. traveling time) of the examined microsimulation traffic model (in this case the VISSIM). The role of a neural network is to predict values of the VISSIM sim- ulations in the process of software examination of combinations of input parameters. In Fig. 2 a simplified scheme of program calibration is shown. The program calibration begins with the creation of a VISSIM simulation database for neural network training (Fig. 2). The task of the neural network is to predict the time of travel between mea- suring points for particular values of input parameters, which would be obtained by the microsimulation model. The program calibration (MATLAB) calls the prediction function, provided by 2 Dissertation is available on-line, in Digital library of the University of Maribor (DKUM). Table 1 Values of measured traffic indicators at examined roundabouts. Roundabout I Roundabout II March 3rd 3–4 p.m. March 3rd 4–5 p.m. July 14th 2–3 p.m. July 14th 8– 9 a.m. Mean value of traveling time (s) 21.8 19.9 18.1 13.3 Maximum queue (m) 26 21 15.5 23 Number of stopping at the entrance 89 61 54 56 Fig. 2. Scheme of computer program calibration. 3 The ratio between the number of input parameters and size of the network training set can be one of the causes of the overtraining. If the number of input parameters is large, and the number of examples in the learning set is small, the network will rather remember, than learn and generalize on the training data set. It should be noted that the ratio between the number of input parameters and the size of network learning set is not the only cause of bad network generalization effect. I. Ištoka Otković et al. / Expert Systems with Applications 40 (2013) 5965–5974 5967 the neural network within the calibration program (subroutine), for each combination of values of input parameters within the gi- ven ranges of values and by a chosen/defined step (Table 2). The preset criterion of comparison of modeled (by neural network pre- diction) and measured values of traveling time according to expression (1) generates a set of potentially optimal combinations of input parameters values into an output file. T MOD � T MEAS T MEAS � � � � � � � � 6 5% ð1Þ where: TMOD is the mean value of modeled traveling time between measurement points, and TMEAS is the mean value of measured trav- eling time between measurement points. The final selection of optimal input parameters from a set of potentially optimal solutions, generated by the calibration pro- gram, is performed based on the simulation results of the original microsimulation traffic model (VISSIM). Validation process of the calibration method is also done (Fig. 2) by application of the real microsimulation model (VISSIM). 3.1. Input parameters of the model The number and the extent of input parameters used in the analysis are directly correlated to the number of possible combina- tions of input parameters necessary to be analyzed in the calibra- tion process. In the process of optimizing the number of model input parameters and their ranges, various statistical methods, de- scribed in detail in literature (Kim, 2006; Park & Qi, 2005), are used. In this paper, neural networks were used for optimisation of input parameters and ranges. Backpropagation type of neural networks can provide a coefficient of influence of every input parameter on the output result, which gives a possibility to select only input parameters influential on the modeled traveling time. Neural net- works have the ability of extrapolation of results, which makes the question of selection of initial ranges of input parameters a less sensitive issue in the process of calibration. The selection of input parameters of the model and their ranges in two steps is described in detail in the dissertation (Ištoka Otković, 2011), and the final re- sult is shown in Table 2. 3.2. Application of neural networks to traveling time prediction The NeuroShell2 software package is used for neural network learning as well as prediction results analysis. The following basic neural networks were tested: Ward Nets, Standard Nets, Jump Connection Nets, Jordan–Elman Nets and General Regression Net (Fig. 3). In total, 70 neural networks, which differed in basic architec- ture, number of hidden layers, number of neurons in hidden layers and activation functions, are analyzed. Neural networks are compared by two basic criteria – training and the ability to generalize. Training is a success of prediction on a training set of data (the set network has learned on), and gen- eralization is the success of prediction on an unknown set of data. A sufficiently large training database partially prevents the net- work overtraining3 effect and creates a precondition for achieving better generalization (Lawrence, Giles, & Tsoi, 1997). For the specific problem, it is possible to create a database for neural network learn- ing of desired size and, therefore, a database of 1379 examples of variations of input parameters and output results of VISSIM simula- tions is generated. From the database, within the NeuroShell2 pro- gram, 20% of data (276 data) is separated in order to be used as a network validation test set. The program offers a possibility of mem- orizing the best result (the best prediction) on a test set of data, as well as memorizing the best result on set of data used for the net- work to learn (training data set). The following indicators were chosen to compare our particular trained networks: correlation coefficient between results of neural network prediction and results of VISSIM simulations, mean abso- lute error and maximum absolute error. The names of all researched networks were coded in order to make them easy-to-survey. The numbers indicate the number of neurons in hidden layers as well as the number of hidden layers (e.g. 13-24-13 is a neural network with three hidden layers having 13 neurons in the first and the third hidden layer and 24 neurons in the second hidden layer). The names of the neural network models included the activation functions of the hidden layers in situations where they differ from the (proposed) default activation functions. The best prediction on the validation (test) data set is the one that is the most often memorized, but there is also the possibility to memorize the best prediction on the training data set which is in this case marked with the ‘‘TR’’ mark. For the General Regression Network, adaptive (AD) and iterative (IT) types of network were analyzed. The comparison of those neural networks that achieved the best prediction results, according to the three chosen criteria, is shown in Table 3. According to the obtained neural network learning results, the best correlation was obtained for the iterative type of regression neural network, as evident from Table 3. Considering the fact that traffic flow modeling is a stochastic problem and that a basic VIS- SIM model uses a random number generator, the achieved correla- tion of 88.3% is satisfactory. In graphs, one set of input data on the abscissa represents one combination of input parameter values (Table 2). One set of values of model input parameters gives one modeled mean traveling time Table 2 Selection of input parameters of the VISSIM calibration model. P Input parameters Value range Step Default value P1 Simulation resolution 1–10 1 5 P2 Number of observed proceeding vehicles 1–4 1 2 P3 Max look ahead distance (m) 100–300 1 250 P4 Min look ahead distance (m) 0–20 1 0 P5 Average standstill distance (m) 1–3 0,1 2 P6 Additive part of desired safety distance (m) 1–5 0,1 2 P7 Multiplicative part of desired safety distance (m) 1–6 0,1 3 P8 Desired speed (km/h) 25–50 10 40* * Expected (desired) travel speed at access lanes of the observed urban intersection. The selected range is designed based on measured velocity on access roads of the observed intersection. ADAPTIVE TYPE ITERATIVE TYPE WARD NETS STANDARD NETS JUMP CONNECTION NETS JORDAN-ELMAN NETS GENERAL REGRESSION NET BEST PREDICTION RESULTS FOR TRAINING DATASET BEST PREDICTION RESULTS FOR TEST DATASET Fig. 3. Examined types of neural networks. Table 3 Performance indicators of trained neural networks – networks with a good response. Type of neural network Correlation coefficient Mean absolute error Maximum absolute error Ward Nets-3 hidden layers 13-18-13 0.8270 0.332 3.000 Ward Nets-3 hidden layers 13-18-20 0.8286 0.331 3.132 Ward Nets-3 hidden layers 13-24-13 0.8309 0.330 3.201 Standard Nets-3 hidden layers13-13-13 0.8355 0.333 3.000 Standard Nets-2 hidden layers 19-19-tanh-logist 0.8208 0.292 3.200 Standard Net-3 hidden layers 26-18-15-tanh-logist 0.7955 0.262 4.421 Standard Nets-3 hidden layers 30-20-18-tanh-logist 0.8199 0.280 3.000 Standard Nets-3 hidden layers 18-15-15-TR 0.8480 0.306 3.000 Standard Nets-3 hidden layers 20-15-13-tanh-logist-TR 0.7881 0.288 4.100 Standard Nets-2 hidden layers 19-24-tanh 0.8402 0.334 3.027 Standard Nets-1 hidden layer 38 0.8378 0.328 3.000 General Regression Net-1 hidden layer 2200tanh-AD 0.8554 0.311 3.000 General Regression Net-1 hidden layer 2380logist-AD 0.8491 0.322 3.000 General Regression Net-1hidden layer 2480tanh-AD 0.8554 0.311 3.000 General Regression Net-1hidden layer 2550-AD 0.8501 0.317 3.000 General Regression Net-1hidden layer 3500tanh-AD 0.8581 0.308 3.000 General Regression Net-1hidden layer 3500tanh-IT 0.8811 0.285 3.328 General Regression Net-1hidden layer 5500logist-AD 0.8469 0.323 3.001 General Regression Net-1hidden layer 5500logist-IT 0.8830 0.278 3.432 Jump Connection Nets-2 hidden layers 19-19 0.8374 0.326 3.000 Jordan–Elman Nets-2 hidden layers 38-38-tanh-logist 0.8293 0.334 3.000 5968 I. Ištoka Otković et al. / Expert Systems with Applications 40 (2013) 5965–5974 between measuring points depicted on the ordinate (Figs. 4–7). In Fig. 4, different combinations of input parameter values (1379 combinations) is depicted on the abscissa, and the ordinate shows the corresponding modeled mean traveling time for every set of in- put data. Traveling time obtained from the General Regression Net- work (Table 3) is compared to results of the VISSIM simulations for 1379 different combinations of values of input data used for neural network learning – training data set (Fig. 4). Besides the software testing of the generalization ability, a two step independent estimate of neural networks is also designed. A set of 25 combinations of new input parameters was used for a comparison of results of generalization of all trained networks. Traveling time prediction provided by the neural network was compared with the traveling time obtained from the VISSIM model. A part of results obtained from Ward Net and Standard Net neural network in the first step of an independent validation is shown in the Fig. 5 and those from the General Regression Network is pre- sented in the Fig. 6. From graphs in Figs. 5 and 6, it is noticeable that the General Regression Networks usually give better generalization results (Fig. 6), than Ward and Standard Net neural network types (Fig. 5). The final validation is done on the basis of data received from 64 new combinations of input parameters, for networks which have shown the best results in the trial validation (Fig. 7). Comparison and assessment are made according to the criterion of minimal mean absolute error. Prediction error is an absolute value of Fig. 4. VISSIM simulations vs. neural network GR 5500logist IT – training data set. Fig. 5. Traveling time prediction for a part of examined neural network. I. Ištoka Otković et al. / Expert Systems with Applications 40 (2013) 5965–5974 5969 difference between neural network prediction results and the expected value obtained by VISSIM simulation. The General Regression Network of the iterative type with a one hidden layer, 5500 neurons in that layer and logistic function of activation, with the approximate configuration depicted in the Fig. 8, had the best correlation with the mean traveling time ob- tained by the VISSIM simulation model. The same network has achieved the best generalization result in both steps of the inde- pendent estimate of neural networks, as it is described in detail and presented in the dissertation (Ištoka Otković, 2011), and there- fore it was chosen for traveling time predictions (replacement for the VISSIM) in the program calibration process (Fig. 2). Schematic overview of the configuration of the chosen neural network is shown in the Fig. 8. 3.3. Traffic microsimulation model calibration program The mean traveling time measured in the field, for the calibra- tion data set, is 21.8 s (Table 1). The mean traveling time obtained by VISSIM simulation, with default values of input parameters, is Fig. 6. Test prediction for adaptive and iterative type of General Regression Networks. 18 19 20 21 22 23 24 25 0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64 Sets of input dana VISSIM GR 5500logist IT SN 13-13-13 TR GR 3500tanh AD Fig. 7. Prediction results of the best neural networks on the validation set of data. 5970 I. Ištoka Otković et al. / Expert Systems with Applications 40 (2013) 5965–5974 20.3 s. The aim of the calibration procedure is to find such combi- nation of values of input parameters, which will approximate the mean measured traveling time. The calibration computer program for the analyzed problem is written in MATLAB programming language. The algorithm, a de- tailed description of the program and the calibration program itself are available in Dissertation (Ištoka Otković, 2011). The basic steps of the program are shown in Fig. 2. For every combination of values of input parameters (Table 2) the program calls the prediction function given by the neural network as a subroutine. A practical problem, which occurred at this research stage, was the connection of neural network prediction function and the MATLAB program. Fig. 8. Scheme of artificial neural network for traveling time prediction. I. Ištoka Otković et al. / Expert Systems with Applications 40 (2013) 5965–5974 5971 Since this was a large output file, manual programming of the pre- diction function in MATLAB format was a time-consuming job. Therefore a special MATLB program, which used the output predic- tion function of the trained neural network in MATLAB format (Ištoka Otković, 2011), had been designed. According to the criterion set in the expression (1), the program generated a set of potentially optimal combinations of input parameter values, and the final selection of the optimal combina- tion was done by application of the original microsimulation traffic model (VISSIM), analysis of traveling time and queue parameters for the existing and new sets of measured data. 3.4. Analysis of program calibration results Analysis of the output file of the calibration program of poten- tially optimal input parameters combinations makes it obvious that there is a certain number of combinations with the maximum values of input parameters P5, P6 and P7 (Table 2). Maximum val- ues of those input parameters influence the modeled capacity of the intersection. Results of the simulation in the VISSIM show that the VISSIM does not manage to generate the overall traffic load in the designated time, which indicates that the capacity of the mod- eled section is smaller than the real one. Generally, the neural net- works gave the correct setting, but realistically, combinations of input parameters P5, P6 and P7 with a maximum value or a value Fig. 9. Output data-file of calibration program co close to the maximum, in the combination, significantly reduce the capacity in comparison to the real one and in this case they are not applicable. The output file analysis procedure is described in detail in the dissertation, but when the combination of input parameter with maximum values (P5, P6, P7) is left out, matching of the results of the output file and VISSIM simulations (80% of the results has less than 5% difference), is seen in Fig. 9. The final selection of values of model input parameters, which will make the best approximation of local traffic conditions (cali- brated model), was made through the second step of calibration, by the comparison of modeling results to measured values of trav- eling time and queue parameters at the examined roundabout. 3.5. Queue parameters During the second step of calibration a set of potentially opti- mal combinations of input parameters is analyzed in the context of VISSIM simulation values obtained for queue parameters – max- imum queue length and number of stopping at the entrance into the roundabout. When the results of the modeling in the VISSIM for 80 combinations of values of input parameters are compared with results of traffic indicators measured in the field according to expression (2), 13 combinations of input parameters, which sat- isfy the criterion set for three examined traffic indicators, are ob- tained. In the real traffic conditions, 21.8 s mean traveling time, maximum queue length of 26 m and the number of stopping at the entrance of 89 are measured (Table 1). T MOD � T MEAS T MEAS � � � � � � � � and Q maxMOD � Q maxMEAS Q maxMEAS � � � � � � � � and STOPMOD � STOPMEAS STOPMEAS � � � � � � � � 6 5% ð2Þ where: TMOD is the mean value of modeled traveling time between measurement points, and TMEAS is the mean value of measured trav- eling time between measurement points, and QmaxMOD mean value of modeled maximum queue at the entrance, and QmaxMEAS is the mean value of measured maximum queue at the entrance, and STOPMOD is the mean value of the modeled number of stopping at the entrance, and STOPMEAS is the mean value of the measured num- ber of stopping at the entrance. In Table 4 those combination values of input parameters that were the closest to the measured values in the field are shaded. mpared with the VISSIM simulation results. Table 4 Calibration by the other traffic indicator (1st set of measured data). Input parameters Neural network VISSIM P1 P2 P3 P4 P5 P6 P7 P8 T T Qmax STOP 6 4 164 8 2.4 2 3.3 40 21.267 20.8 25 86 6 4 172 8 2.4 1.9 3.4 40 21.303 20.8 25 90 6 4 172 8 2.4 1.9 3.5 40 21.295 20.9 25 89 6 4 172 8 2.5 1.9 3.3 40 21.295 20.8 25 86 6 4 172 12 2.4 1.9 3.4 40 21.271 20.8 25 90 6 4 172 12 2.4 1.9 3.5 40 21.264 20.9 25 89 6 4 172 12 2.5 1.9 3.3 40 21.283 20.8 25 86 6 4 173 8 2.4 1.9 3.4 40 21.298 21.1 25 85 6 4 173 12 2.4 1.9 3.4 40 21.298 21.1 25 85 6 4 175 8 2.5 1.9 3.5 40 21.262 21.0 25 85 6 4 176 8 2.4 1.8 3.5 40 21.275 21.4 27 88 6 4 176 8 2.5 1.8 3.4 40 21.273 20.9 27 90 6 4 176 12 2.5 1.8 3.4 40 21.273 20.9 27 90 5972 I. Ištoka Otković et al. / Expert Systems with Applications 40 (2013) 5965–5974 Two bolded combinations of input parameters, which present the best result, entered the validation procedure. The value of simu- lated time of 21.4 s was the best result, and, for the second combi- nation, one out of two practically the same combinations was selected. VISSIM, unlike neural networks, provides the same values of traffic indicators (P4 = 8 and P4 = 12), thus one of two combina- tions that differ only by the fourth input parameter P4, was selected. 4. Validation of the calibrated model The validation procedure is comprised of two steps and it in- volves a comparison of simulation results obtained by an uncali- brated and a calibrated VISSIM microsimulation model with traffic indicators measured in the field. The first validation was done for two new sets of data measured at the same roundabout where the calibration is done (roundabout I). The second valida- tion, done on the set of data measured at the second roundabout (roundabout II), gives an insight into the dilemma whether the cal- ibration process is tied to the one particular roundabout or it may be considered in a broader context. 4.1. First validation The first validation was performed at the same roundabout at which calibration procedure was carried out. The first validation data set of traffic parameters was measured on Wednesday, March 3rd 2010, between 4 and 5 p.m., and the second one was gathered on Wednesday, July 14th 2010, between 2 and 3 p.m. Table 5 Comparison of traffic indicators – first validation. Input parameters of VISSIM model P1 P2 P3 P4 P5 P6 Measured values First validation data set Second validation data set Values obtained by VISSIM model simulations Default values of input parameters – first validation measured data set 5 2 250 0 2 2 Default values of input parameters – second validation measured data set 5 2 250 0 2 2 Calibrated model – first validation measured data set 6 4 176 8 2.4 1.8 6 4 173 8 2.4 1.9 Calibrated model – second validation measured data set 6 4 176 8 2.4 1.8 6 4 173 8 2.4 1.9 In Dissertation (Ištoka Otković, 2011), data on traffic volume and traffic distribution in the roundabout for all performed count- ing and the measured traffic parameters are presented and a single measured passing time between the measurement points for each vehicle in the reference period for all counting was given. It is evident from the Table 5 that default values of input param- eters have provided traffic parameters which did not meet one of the three indicators, i.e. maximum queue differs for 28% from the measured values. The first combination of calibrated model input parameters satisfied all three criteria with an error smaller than 5% for the first measured data set. Both parameter combinations, obtained by calibration procedure (shaded), satisfied the limit set by expression (2) approaching the values measured in the field for the second data set. 4.2. Second validation The second validation was performed at the new location, at the second observed roundabout (roundabout II). Data were collected and traffic parameters measured, on Wednesday morning, July 14th 2010, between 8 and 9 a.m. A graphic presentation of the main roundabout project and measured data of traffic volume and traffic distribution, given in Dissertation (Ištoka Otković, 2011), served as the background for the second validation. VISSIM simulation values for input parameters default values and the two best parameter combinations were compared with measured values of traffic parameters and presented in the Table 6. Default values of input parameters provided simulation values of traveling time within specified limits, but the queue parameter Traffic indicators P7 P8 T Qmax STOP 19.9 21 61 18.1 15.5 54 3 40 20.3 15 60 3 40 17.6 23 50 3.5 40 19.8 22 58 3.4 40 20.2 15 63 3.5 40 17.6 15 52 3.4 40 17.8 15 53 Table 6 Comparison of traffic indicators – second validation, second roundabout. Input parameters of VISSIM model Traffic indicators P1 P2 P3 P4 P5 P6 P7 P8 T Qmax STOP Measured values 13.3 23 56 Values obtained by VISSIM model simulations Default values of input parameters 5 2 250 0 2 2 3 40 13.1 27 50 Calibrated model – third validation measured data set 6 4 176 8 2.4 1.8 3.5 40 13.1 22 58 6 4 173 8 2.4 1.9 3.4 40 13.1 22 54 I. Ištoka Otković et al. / Expert Systems with Applications 40 (2013) 5965–5974 5973 results were not within the expected range. The simulation results showed 26% greater maximum queue value and 10.7% lower value for the number of stopping at the roundabout entrance. Combina- tions of input parameters, obtained by calibration procedure, were within limits set by expression (2) and are shaded in the table. By analyzing and comparing the results of values of traffic parameters shown in Tables 4–6, the combination of input parameters, that satisfied all the observed indicators within the pre-set statistical limits, according to expression (2), can be ob- served. The first combination, obtained by calibration procedure, achieved the optimal combination of values of input parameters for single-line roundabouts in local traffic conditions shown in Table 7. 5. Discussion In the dissertation (Ištoka Otković, 2011) several calibration methods were analyzed using computer software calibrations based on the following: – Prediction of a neural network for a traffic indicator – traveling time between measuring points; – Prediction of a neural network for three traffic indicators – time of travel between measuring points, maximum length of a queue and number of stopping at the entrance into a roundabout; – Three independent neural networks which give three indepen- dent predictions for the examined traffic indicators. This paper describes in details the method, which provided the best result. This method is based on the prediction of one neural network for the traveling time between measuring points. Results presented in this paper clearly show that a neural net- work is not an alternative to a microsimulation model. The neural network achieved accurate results of traveling time prediction, but in this research no neural network gave good results for queue parameter prediction. This paper does not implement neural net- works in the prediction of other traffic indicators obtained by means of microsimulation model. A database of 1379 combinations of values of the model input parameters and the simulation traveling time results made in the VISSIM is used for neural network learning. Such a large database is used in order to investigate the real neural networks response and in order to prevent the network overtraining or lack of vari- Table 7 Optimal values of VISSIM model input parameters. Optimal values of VISSIM model input parameters P1 P2 P3 P4 P5 P6 P7 P8 6 4 176 8 2.4 1.8 3.5 40 ability in the network database to be the causes of possible bad results. It is presumed that such a large database is not necessary for achieving optimal results of neural network learning. The calibration computer program calls the neural network pre- diction as a subroutine and each combination of values of input parameters that meet the set statistical criteria – expression (1), is entered in the output file. Upon the analysis of the calibration program output file in VIS- SIM, 80 potentially optimal combinations of input parameters, which satisfy the criterion (1) set for the traveling time, are ob- tained. In the second step, 80 combinations were analyzed by applying the VISSIM, and in the context of the second traffic indi- cator, in order to choose combinations of values of input parame- ters which provide good simulation results for both traveling time and queue parameters. 13 combinations, which satisfy the condition set in accordance with expression (2), were obtained. Two combinations, which were the best in approximating values measured in the field on the calibration data set, were selected out of those thirteen combinations. The validation procedure was done in two steps, and traveling time and queue parameters, obtained by calibrated and uncali- brated traffic model, were compared to values measured in the field. In the first validation, a new data set, measured at the same roundabout for which the model calibration was done, was com- pared (Table 5). From Table 5 it is obvious that the calibrated mod- el gives simulation results for traveling time between measuring points and queue parameters closer to values measured in the field, than the uncalibrated model with default values of input parameters of the model. The second validation is done at the other roundabout (Table 6). Results, obtained by the comparison of modeling results and traffic indicators measured in the field, show that better results are reached with the model calibrated on the second roundabout of the same type in the local traffic network than with the uncali- brated model. The results in Table 6 show that the calibrated microsimulation model is applicable to the one-lane roundabouts in a local condition and that it is not necessarily related to the loca- tion of the intersection where the calibration is made. Such a con- clusion should not be taken too seriously, given that the research was based on two one-lane roundabouts. Possibility of generalized application of the obtained values of input parameters to other types of roundabouts or intersections in a local traffic network in general should be additionally explored. However, such an opinion is not unfounded, because most of calibrated input parameters are related to the behavior (psychology) of driver, which is to some ex- tent territorially and culturally conditioned. The combination of model input parameters, which approxi- mates local traffic conditions the best, is selected by the analysis of achieved results, presented in Tables 4–6. Optimal combination of values of model input parameters (Table 7) represents the cali- brated microsimulation traffic model for one-lane roundabouts in the examined local network. 5974 I. Ištoka Otković et al. / Expert Systems with Applications 40 (2013) 5965–5974 6. Conclusion Within this paper, a new calibration method encompassing the application of a neural network for the prediction of the results of simulations of microsimulation traffic model within the computer calibration program is analyzed. Traveling time and queue parameters are the traffic indicators which are analyzed in the process of calibration and validation of the model, because they are easily measurable in real traffic condi- tions. The microsimulation model selected for the study of the applicability of neural networks in the calibration procedure is the VISSIM, and two urban roundabouts were used as the basis of the experiment. Results have shown that a neural network is applicable in the process of calibration of the examined microsimulation model. Basic steps of the calibration method by means of neural net- work application include: – A minimum of two sets of measured data (the first for the model calibration and the second for the validation), which are: vehicle and pedestrian traffic load, traffic distribution, mean traveling time between two chosen measuring points easy to enter into the model layout, queue parameters at one of the entrances into the roundabout; – Creating the VISSIM simulation database (variations of values of input parameters) for neural network learning; – Selection of neural network (General Regression Net, Iterative type, with a logistic activation function gave the best response in this study) and obtaining the function of prediction for the program calibration; – Program calibration (program designed in MATLAB); – Checking the results of the program calibration output file in the VISSIM model for the mean traveling time and queue parameters as well as checking the best result in the validation procedure; – Validation procedure involves comparison of simulated and measured data for the second set of data measured in the field; – Combination of input parameters, that gives the simulation results which best approximate the measured data, is opti- mal = calibrated VISSIM. – The described basic steps of calibration are applicable, not only in the VISSIM, but in other microsimulation models. The question about the optimal size of database for microsimu- lation modeling of neural network learning and the question about a possibility of finding a configuration of neural network which would provide a better response than the chosen General Regres- sion Network both remain unanswered, since these answers re- quire further research. Regarding the optimisation issues, the question is whether the obtained optimum is local or absolute minimum and how good should possible local minimum approximate the absolute one. In practical terms, the comparison of the measured field results with simulation results obtained with calibrated and default values of input parameters, gives a basic insight into the performance of the calibration procedure. References Castro-Neto, M., Jeong, Y. S., Jeong, M. K., & Han, L. D. (2009). Online-SVR for short- term traffic flow prediction under typical and atypical traffic conditions. Expert Systems with Applications, 36, 6164–6173. Cheu, R. L., Jin, X., Ng, K. C., Ng, Y. L., & Srinivasan, D. (1998). Calibration of FRESIM for a Singapore expressway using genetic algorithm. Journal of Transportation Engineering, 124(6), 526–535. Cicu, F., Illotta, P. F., Bared, J., & Isebrands, H. (2011). VISSIM calibration of roundabouts traffic performance, In Transportation research board annual meeting 2011. Washington DC. Dia, H. (2001). An object-oriented neural network approach to short-term traffic forecasting. European Journal of Operational Research, 131, 253–261. Dorothy, P., Ambadipudi, R., & Kill, R. (2006). Development and validation of large- scale microscopic models. In Transportation research board 85th annual meeting 2006, Paper#06-2392. Washington DC. Ettema, D., Arentze, T., & Timmermans, H. (2011). Social influences on household location, mobility and activity choice in integrated microsimulation models. Transportation Research Part A, 45, 283–295. Fang, F. C., & Elefteriadou, L. (2005). Some guidelines for selecting microsimulation models for interchange traffic operational analysis. Journal of Transportation Engineering, 131(7), 535–543. Flötteröd, G., Chen, Y., & Nagel, K. (2012). Behavioral calibration and analysis of a large-scale travel microsimulation. Networks and Spatial Economics, 12(4), 481–502. Flötteröd, G., Bierlaire, M., & Nagel, K. (2011). Bayesian demand calibration for dynamic traffic simulations. Transportation Science, 45(4), 541–561. Ištoka Otković, I. (2011). Using neural networks in the process of calibrating the microsimulation models in the analysis and design of roundabouts in urban areas, Ph.D. thesis, University of Maribor, Faculty of Civil Engineering; <http:// dkum.uni-mb.si/Iskanje.php?type=napredno&stl0=Avtor&niz0=Irena+I%C5% A1toka+Otkovi%C4%87>. Kim, S.J. (2006). Simultaneous calibration of a microscopic traffic simulation model and OD matrix, Ph. dissertation, Texas A&M University, USA. Lawrence, S., Giles, C. L., & Tsoi, A. C. (1997). Lessons in neural network training: Overfitting may be harder than expected. In Proceedings of the fourteenth national conference on artificial intelligence, AAAI-97 (pp. 540–545). Menlo Park, California: AAAI Press. Olstam, J. J., & Tapani, A. (2004). Comparison of queue models, Swedish national road and transport research institute, Project VTI meddelande 960 A, Linköping, Sweden. Onieva, E., Milanés, V., Villagrá, J., Pérez, J., & Godoy, J. (2012). Genetic optimization of a vehicle fuzzy decision system for intersections. Expert Systems with Applications, 39, 13148–13157. Park, B., & Qi, H. (2005). Development and evaluation of simulation model calibration procedure. 84th annual meeting 2005, reprint CD-ROM. Washington, DC: Transportation Research Board. Park, B., Won, J., & Yun, I. (2006). Application of microscopic simulation model calibration and validation procedure: A case study of coordinated actuated signal system, Transportation Research Record, TRB, National Research Council, Washington, DC. Park, D., Rilett, L. R., & Han, G. (1999). Spectral basis neural networks for real-time travel time forecasting. Journal of Transportation Engineering, 125(6), 515–523. Park, K., & Kim, W. (2001). A systolic parallel simulation system for dynamic traffic assignment: SPSS-DTA. Expert Systems with Applications, 21, 217–227. Rickert, M., & Nagel, K. (2001). Dynamic traffic assignment on parallel computers in TRANSIMS. Future Generation Computer Systems, 17, 637–648. Stevanović, Z., & Martin, P. T. (2008). Assessment of the suitability of microsimulation as a tool for the evaluation of macroscopically optimized traffic signal timings. Journal of Transportation Engineering, 134, 59–67. Transportation Research Bord (2000). Highway Capacity Manual, Transportation Research Board, National Research Council, Washington, DC, USA. http://refhub.elsevier.com/S0957-4174(13)00289-3/h0005 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0005 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0005 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0010 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0010 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0010 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0015 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0015 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0020 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0020 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0020 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0025 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0025 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0025 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0065 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0065 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0065 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0030 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0030 http://dkum.uni-mb.si/Iskanje.php?type=napredno&amp;stl0=Avtor&amp;niz0=Irena+I%C5%A1toka+Otkovi%C4%87 http://dkum.uni-mb.si/Iskanje.php?type=napredno&amp;stl0=Avtor&amp;niz0=Irena+I%C5%A1toka+Otkovi%C4%87 http://dkum.uni-mb.si/Iskanje.php?type=napredno&amp;stl0=Avtor&amp;niz0=Irena+I%C5%A1toka+Otkovi%C4%87 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0035 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0035 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0035 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0035 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0040 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0040 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0040 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0045 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0045 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0045 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0070 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0070 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0050 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0050 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0055 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0055 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0060 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0060 http://refhub.elsevier.com/S0957-4174(13)00289-3/h0060 Calibration of microsimulation traffic model using neural network approach 1 Introduction 2 Data gathering 3 Calibration of microsimulation model using neural network approach 3.1 Input parameters of the model 3.2 Application of neural networks to traveling time prediction 3.3 Traffic microsimulation model calibration program 3.4 Analysis of program calibration results 3.5 Queue parameters 4 Validation of the calibrated model 4.1 First validation 4.2 Second validation 5 Discussion 6 Conclusion References 
work_b3qhs5g3ajbtpohq2awwdjge74	----	[PDF] Expanding the TRANSFAC database towards an expert system of regulatory molecular mechanisms | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1093/nar/27.1.318 Corpus ID: 18774013Expanding the TRANSFAC database towards an expert system of regulatory molecular mechanisms @article{Heinemeyer1999ExpandingTT, title={Expanding the TRANSFAC database towards an expert system of regulatory molecular mechanisms}, author={T. Heinemeyer and X. Chen and H. Karas and A. Kel and O. Kel-Margoulis and I. Liebich and T. Meinhardt and I. Reuter and F. Schacherer and E. Wingender}, journal={Nucleic acids research}, year={1999}, volume={27 1}, pages={ 318-22 } } T. Heinemeyer, X. Chen, +7 authors E. Wingender Published 1999 Biology, Medicine, Computer Science Nucleic acids research TRANSFAC is a database on transcription factors, their genomic binding sites and DNA-binding profiles. In addition to being updated and extended by new features, it has been complemented now by a series of additional database modules. Among them, modules which provide data about signal transduction pathways (TRANSPATH) or about cell types/organs/developmental stages (CYTOMER) are available as well as an updated version of the previously described COMPEL database. The databases are available on… Expand View on PubMed academic.oup.com Save to Library Create Alert Cite Launch Research Feed Share This Paper 313 CitationsHighly Influential Citations 25 Background Citations 55 Methods Citations 79 View All Figures, Tables, and Topics from this paper table 1 figure 1 table 2 figure 2 table 3 View All 5 Figures & Tables Expert system Database Transduction (machine learning) TRANSCRIPTION FACTOR Organ Transcription (software) WWW 313 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency The TRANSFAC system on gene expression regulation E. Wingender, X. Chen, +11 authors S. Urbach Medicine, Computer Science Nucleic Acids Res. 2001 694 PDF Save Alert Research Feed Modeling regulatory pathways with the use of the TRANSFAC system E. Wingender Biology 2002 2 Save Alert Research Feed TRANSFAC®: transcriptional regulation, from patterns to profiles V. Matys, Ellen Fricke, +18 authors E. Wingender Medicine, Computer Science Nucleic Acids Res. 2003 2,094 PDF Save Alert Research Feed The Eukaryotic Promoter Database (EPD): recent developments Rouaïda Périer, Thomas Junier, C. Bonnard, P. Bucher Computer Science, Medicine Nucleic Acids Res. 1999 40 PDF Save Alert Research Feed TRANSCompel®: a database on composite regulatory elements in eukaryotic genes O. Kel-Margoulis, A. Kel, I. Reuter, I. Deineko, E. Wingender Computer Science, Medicine Nucleic Acids Res. 2002 137 PDF Save Alert Research Feed Discovery and modeling of transcriptional regulatory regions. J. Fickett, W. Wasserman Biology, Medicine Current opinion in biotechnology 2000 169 Save Alert Research Feed Pathway Databases J. Wixon Biology, Medicine Comparative and functional genomics 2001 35 View 2 excerpts, cites background Save Alert Research Feed TFBScluster: a resource for the characterization of transcriptional regulatory networks I. Donaldson, M. Chapman, B. Göttgens Biology, Medicine Bioinform. 2005 30 PDF Save Alert Research Feed The TRANSFAC project as an example of framework technology that supports the analysis of genomic regulation E. Wingender Biology, Medicine Briefings Bioinform. 2008 381 PDF View 1 excerpt, cites background Save Alert Research Feed TFBScluster web server for the identification of mammalian composite regulatory elements I. Donaldson, B. Göttgens Biology, Computer Science Nucleic Acids Res. 2006 10 Highly Influenced PDF View 4 excerpts, cites methods Save Alert Research Feed ... 1 2 3 4 5 ... References SHOWING 1-10 OF 33 REFERENCES SORT BYRelevance Most Influenced Papers Recency Databases on transcriptional regulation: TRANSFAC, TRRD and COMPEL T. Heinemeyer, E. Wingender, +9 authors N. Kolchanov Computer Science, Medicine Nucleic Acids Res. 1998 1,532 PDF Save Alert Research Feed TRANSFAC, TRRD and COMPEL: towards a federated database system on transcriptional regulation E. Wingender, A. Kel, +6 authors N. Kolchanov Computer Science, Medicine Nucleic Acids Res. 1997 133 PDF Save Alert Research Feed The Eukaryotic Promoter Database (EPD): recent developments Rouaïda Périer, Thomas Junier, C. Bonnard, P. Bucher Computer Science, Medicine Nucleic Acids Res. 1999 40 PDF Save Alert Research Feed TRANSFAC: a database on transcription factors and their DNA binding sites E. Wingender, P. Dietze, H. Karas, R. Knüppel Biology, Computer Science Nucleic Acids Res. 1996 1,006 PDF Save Alert Research Feed A compilation of composite regulatory elements affecting gene transcription in vertebrates. O. Kel, A. G. Romaschenko, A. Kel, E. Wingender, N. Kolchanov Biology, Medicine Nucleic acids research 1995 102 Save Alert Research Feed A novel method to develop highly specific models for regulatory units detects a new LTR in GenBank which contains a functional promoter. K. Frech, J. Danescu-Mayer, T. Werner Biology, Medicine Journal of molecular biology 1997 113 Save Alert Research Feed MatInd and MatInspector: new fast and versatile tools for detection of consensus matches in nucleotide sequence data. K. Quandt, K. Frech, H. Karas, E. Wingender, T. Werner Biology, Medicine Nucleic acids research 1995 2,801 Save Alert Research Feed Transcription regulatory region analysis using signal detection and fuzzy clustering L. Pickert, I. Reuter, F. Klawonn, E. Wingender Computer Science, Medicine Bioinform. 1998 63 PDF Save Alert Research Feed Towards a genetic basis for kidney development J. Bard, J. McConnell, J. Davies Biology, Medicine Mechanisms of Development 1994 80 Save Alert Research Feed SRS: information retrieval system for molecular biology data banks. T. Etzold, A. Ulyanov, P. Argos Computer Science, Medicine Methods in enzymology 1996 495 Save Alert Research Feed ... 1 2 3 4 ... Related Papers Abstract Figures, Tables, and Topics 313 Citations 33 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_b4flychdojbpfic7qrlqfilb5e	----	1 The Evaluation and Impact of NEPER Wheat Expert System Ahmed Rafea Computer Science Dept., AUC Email: rafea@aucegypt.edu Mostafa Mahmoud Central Lab. for Agriculture Expert System, ARC Email: mostafa@ esic.claes.sci.eg Abstract: This paper presents the laboratory and field evalation results of NEPER Wheat expert system. The laboratory evaluation showed that NEPER performance is comparable with human experts. Field evaluation has revealed that NEPER has good economic and environmental impacts. The field testing results have also shown that NEPER is usable, applicable and needed. Copyright ® 2001 IFAC Keywords: Expert Systems, Diagnosis, Knowledge-based systems, Hierarchical structures, Classification, Intelligence. 1. INTRODUCTION Bread, known as aish, or life, is a vital component of the Egyptian diet. In 1993, the country produced 4.5 million tons of wheat on 2.2 million feddans. Given the crucial role wheat plays in Egypt, CLAES cooperated with the Intelligent Systems laboratory (ISL) at Michigan State University in developing the Egyptian Regional Wheat Management System, funded by NARP a United States Agency for International Development (USAID) project in the period from 1992 to 1995. This project integrates an ES with a crop simulation model and aims at addressing all aspects of irrigated wheat management in Egypt. This integrated system is named NEPER (Kamel et al, 1995). In order to achieve this goal, NEPER is designed to perform the following functions: • Select the appropriate variety for a specific field • Advise the farmer on field preparation • Design schedules for irrigation and fertilization • Control pests and Weeds • Manage harvests • Diagnose malnutrition • Diagnose disorders • Suggest Treatments In 1997, another project between CLAES and ISL was funded by the ATUT which is also a USAID project. One of the objectives of this project was conducting field-testing to measure the ES performance. The objective of conducting the field- testing was to evaluate the economical and environmental impacts and to measure the ES performance from three aspects: usability, applicability and need. The results of this testing were also used to enhance the user interface and extend the knowledge base of Neper .In this project, there is a component for the evaluation of a new enhanced version of NEPER that considers the whole agricultural operations. This new version has been developed according to the results and recommendation of the field testing that is presented in this paper. In this paper, the technical background of NEPER is presented in section 2. The laboratory evaluation is summarized in section3. The experiments description is presented in sections 4. The economical and environmental impact of the ES are summarized in sections 5 & 6, respectively. The ES performance, section 7, has been measured using three aspects namely usability, applicability, and need of ES. ES enhancements as a result of those experiments are presented in section 8. 2. BACKGROUND In developing Neper, the Generic Task Approach to ES development proposed by Chandrasekran (Chandrasekran, 1986) has been used. The idea behind the Generic Task approach, is that the way a problem is to be solved, depends largely on its type e.g., diagnosis, design, planning, etc. Consequently, problems of the same type could share some sort of a generic problem solver. So, according to the Generic Task methodology approaching a diagnosis problem will be inherently the same regardless of the domain in which such a problem is being addressed. The classical example of a problem solver that could be applied to a diagnosis problem is Hierarchical Classification (Gomez & Chandrasekran, 1981; Chandrasekran, 1983) and it is this problem solver that has been used in implementing the Wheat disorders ES, which is a component of NEPER. 2 This system component has been implemented using a Generic Task Tool developed at Michigan State University (MSU). In this tool, the knowledge base is created as a hierarchy of nodes. In each node, the knowledge is represented in a table, where each entry in this table represents either a database variable or a variable pointing to another table. Each database variable is associated with a question. A user will be presented with that question only if the database variable has never been assigned a value. The combination of possible inputs for each question denotes different rules and matching patterns. If a combination of inputs results in a match value greater than a given threshold, the node is said to be established. By asking the user a series of questions, the system is able to pursue or rule out paths in the classification in which the leaves represent disorders. Basically, if a path from a root to a leaf exists, then the disorder at the leaf is presented as the diagnosis. 3. Laboratory Evaluation Laboratory evaluation is conducted before dissemianting the ESs in the field. The Laboratory evaluation methodology consists of three main procedures namely Verification, Validation, and Evaluation. Verification is defined as the demonstration of consistency, completeness, and correctness of software (Adrion et al, 1982). O’Keefe et al. (1987, 1989, and 1990) have defined verification as "Building the system right", that is making sure that the implemented system is functionally matching the proposed design, and free of semantic and syntactic errors. Validation is the process whereby the system is tested to show that its performance matches the original requirements of the proposed system. It is defined as the determination of the correctness of the final program or software produced from a development project with respect to the user needs and requirements (Adrion et al, 19982). As noted by O’Keefe et al. (1987, 1989, and 1990) "Validation means building the right system". Evaluation is the process whereby we ensure the usability, quality, and utility of the ES (O’Keefe et al. 1987, 1989, and 1990). A complete testing cycle is performed in iterations through which, the ES is updated and refined. Verification process evolves through two main stages during the development of the ES: the development stage and the examination stage. In the development stage, the developer practices different functions of the implemented systems, looking for potential errors that may exist. This is accomplished using two broad techniques: non case-based and case-based. Non case-based techniques include tracing, spying and other traditional debugging techniques. Case- based verification techniques are applied by preparing "Typical Cases". These cases should be selected to serve requirements satisfaction as spelled out in the requirement specification. In the examination stage, the ES is tested to make sure that it is running properly, by testing all the functions of the system trying to examine the performance of the system in different situations. The output of this stage is the verification report that is a document of differences between system design and implementation. This report is used to update , the design document and implementation. The validation step is done through conducting meeting with the doamin experts who provided the knowledge to check that the right system has been developed. This is done by going throgh the generated test cases during the meeting with the domain experts. Their comments on the content and user interface are considered. Necessary updating of the design and implementation is done. The evaluation step is to assess the quality, usability, and utility of the ES from the point of view of human experts other than the domain expert, who participate in the system development. Typical cases are created and distributed to three domain experts in the specialty of a specific sub system. If one sub system includes more than one specialty, cases are distributed to all experts in different specialties. For example in the remediation subsystem, we have three specialties: plant pathology, entomology, and nutrition. Therefore 9 experts have participated in the validation of this subsystem. For each specialty, an evaluator is selected to blindly assess the responses of the three human experts and the ES. After the evaluation, the domain expert participated in the development, the evaluator, and the domain experts participate in an evaluation meeting together with the knowledge engineer to discuss the evaluation results till they reach to a consensus. Figure 1 NEPER Diagnosis evaluation result Applying this methodology on NEPER, verification and validation were done sucessfully. In this paragraph we will presnt the evaluation results of the diagnosis and treatment subsystems. Figure 1 shows the evalaution scores of NEPER diagnosis �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� Expert System Expert � Expert � Expert � Diseases Diagnosis Insect Identification Malnutrition Diagnosis 3 subsystem. NEPER diagnosis subsystem over performs human expert in the insect and malnutrition specilties, and its score in the disease diagnosis results (86%) is equivalent to those of the best human expert. The evaluation scores of NEPER treatment subsystem is shown in figure 2. NEPER treatment over performs human expert in disease treatment, and its score in the insecta and malnutrition treatment are 0.95 and 0.85 respectivly of the best expert-group. After this experiment, NEPER has been trained to reach the scores of the best experts.. Figure 2 NEPER Treatment evaluation result 4. Field Test Experiment Description Many experiments were conducted in the last few years for NEPER ESs. The objectives of those experiments were to validate the system in the field, and to measure the impact of using the system. The experiments were conducted in different locations by selecting two fields at same area and location: one is to be cultivated using NEPER Wheat recommendations without any interference from the agriculture engineer or any specialist, and the other one is to be cultivated as usual, this is a control field. In order to get the best results from the experiment, the following issues and activities were considered and followed: ♦ Formal training on the usage of NEPER was conducted for the staff who are going to use the system. ♦ A computer engineers from CLAES were responsible for supporting the site staff on the usage of the system and handling trouble-shooting problems of hardware and software. ♦ A number of the wheat researchers from Field Crop Research Institute (FCRI) were assigned to supervise different fields, i.e., a researcher for each site. ♦ Periodical fields visits were conducted by researchers from CLAES and FCRI Three experiments were conducted in three different seasons for NEPER ES. Two of those three experiments had the same number of fields, both of them consisted of a total number of 32 fields carefully selected for conducting the experiment. The third one consisted of a total number of 44. These fields were equally divided so that 16 fields in the first two expermints and 22 fields in the third one were assigned to utilize NEPER and managed by the ES, and the other fields were to be managed in the usual practice and acts as control. The selected fields were located in four different geographical areas, namely: Noubaria, Gemiza, Sharkia, and Decerns. In Noubaria two sites were selected to cover the different types of soil at that area. One of those sites located in Bostan and the other one located in Banger El-Sokar. The first experiment covered only the diagnosis and treatment part of the NEPER Wheat ES including Weed Identification. The second one included the strategic part and tactic part. Strategic part includes six subsystems called: Variety Selection, Pre- cultivation Pest Control, Tillage, Planting, Irrigation & fertilization, and Harvest. The third one also included the strategic part and tactic part. Strategic part includes six subsystems called: Variety Selection, Planting, Land Preparation, Irrigation, fertilization, and Harvest. Tactical part includes two subsystems called diagnosis and weed identification, each of them includes the treatment function. 5. Economical Impact In the first experiment (CLAES, 1996), the averages of treatment costs, yields, and straw per feddan was calculated for both NEPER and the control fields. By taking the averages of treatment cost, yield, and straw per Feddan, it was found that the average net income per feddan for ES fields is 2049.85 LE and for control fields is 1600.05 LE, consequently, the net production increase in Egyptian Pound was 449.8. This represents 26.78% increase in the production In the second and third experiments (CLAES, 1999, CLAES, 2001), the complate system was tested. Tables (1) and (2) summarize the result of those experiments in the new reclaimed area and the Delta area. The following remarks were observed: • In both the new reclaimed and Delta area, there was an increase in the production and net profit consistently in the two consecutive seasons. • The percentage of increase in the net profit in the newly reclaimed is greater than the percentage of increase in the net profit in the Delta area. • The production in the newly reclaimed area is less than the Delta area because the lack of expertise �� �� �� �� �� �� �� �� �� �� �� �� � �� �� �� �� �� �� �� �� �� ��� Expert System Expert � Expert � Expert � Dis eas esTreatment Insect Treatm ent Malnutrition Treatm ent 4 in the reclaimed area. Hence expertise transfer in this area has led to a relatively high impact. 6. Environmental Impact The conservation of natural resources has two aspects. The first is pertinent to the management of these resources on the macro level, such as controlling the expansion of urban development in order not to loose agricultural land. The second is concerned with the management of these resources on the micro level such as adding chemical fertilizers to the soil. In this paper, the focus will be on the status of the water and land resources because they are the two main resources related to our work on crop management ESs. Water is the scarcest resource in Egypt, since its supply is nearly fixed and water demand for different sectors is continuously increasing. The water supply can be classified into three categories: surface water, ground water, and (isle) water reuse after treatment either from agriculture drainage or domestic usage. The decision makers concerned with water resource management in Egypt are challenged by how to balance the limited water supply with an increasing water demand for the future, since water is the major constraint for land expansion to satisfy food self- sufficiency. Another challenge is how to reduce the water pollution resulting from using chemical fertilizers and pesticides. After water, land is the major limiting factor for sustainable agricultural development (Rafea, 1996). There are two problems facing decision-makers to conserve water resources namely: the efficient utilization of water resources, and the pollution resulting from the usage of chemical fertilizers and pesticides. Regarding soil conservation, there are two main problems namely: the urban expansion, and the soil degradation resulting from excessive use of fertilizers and other bad agricultural practices. Therefore, the main contribution of ESs for soil and water conservation is to transfer the agricultural practices according to certain strategy or a combination of strategies namely: environmental sustainability, economical sustainability and/or social sustainability. In the ESs that have been built so far, we are concerned with economic sustainability taking into consideration the environmental sustainability in the second place. In other words, we are trying to acquire the recommendations that optimize the output relative to the agricultural inputs. As a consequence, environmental conservation is achieved, because no extra input is provided such as water, fertilizers and pesticides without a return in the yield. The results of experiments conducted for the ES agree with the goals of environmental conservation. The fields managed by the ES have used fewer resources in terms of fertilizers and pesticides than the control fields and hence conserve environment. The cost is an indicator of the increase or decrease of using chemicals in general. Hence, we have used the cost as a factor in determining the quantity of used fertilizers and pesticides. The average cost of pesticides used by NEPER Wheat fields in the first experiment was more than the control fields by 15.7 Egyptian pound/Fadden, but the production increased by 449.8 Egyptian pound/Fadden. Notice that the increase of cost in this experiment is negligible. In the second experiment the average cost of fertilizers and pesticides used by NEPER Wheat fields was less than the control fields by 5.57 Egyptian pound/Fadden and the production increased by 247.4 Egyptian pound/Fadden. In the third experiment the average cost of fertilizers and pesticides used by NEPER Wheat fields was less than the control fields by 1.2 Egyptian pound/Fadden and the production increased by 287.12 Egyptian pound/Fadden. In fact, this indicates that changes to the NEPER Wheat system have made it more compliant to the goals of resource management and environmental conservation. In the second experiment of NEPER wheat ES, the average water quantity used to produce one Ardab of wheat in the ES fields was 112.74 M3 water, while in the control fields the farmers used 152.58 M3 water on average to produce the same quantity of wheat. This represents 35% decrease in the use of irrigation water. Table 2: The result of the experiments in the Delta area Season 97/98 Season 98/99 Item ES Control Differance % ES Control Differance % Average Production 2130 1830 300 16 2117 1759 358 20 Average cost 631 597 34 5 445 422 23 5 Average Net profite 1499 1233 265 22 1672 1337 335 25 Table 1: The result of the experiments in the new reclaimed area Season 97/98 Season 98/99 Item ES Control Differance % ES Control Differance % Average Production 1701 1506 195 13 1647 1431 216 15 Average cost 747 861 -114 -13 468 603 -135 -22 Average Net profite 954 645 309 48 1179 828 351 42 5 7. Expert System performance The expert system performance has been measured using three aspects namely usability, applicability, and need of ES. 7.1 Expert Usability In order to measure the usability of the ES, the developers in CLAES have re-run the system on the cases reported in the forms of the fields managed by NEPER and compared the conclusions with the results represented in the field books by the researchers and extension agriculture engineers in different locations. In the first experiment (CLAES, 1996) was examining the comparison results, it was found that in 86% of the cases, the trained researchers have used the system correctly while this percentage has decreased to 38% for untrained researchers. This indicates the importance of training on the usage of the ES. It is worth noting that there is no great difference between the researchers and extension officers in using the system as the differences was only 4%, although the system was in English. This proves the importance of ES. It raised the performance of extension officers to the level of researchers, in the underlying domain of the NEPER. In the second experiment there was discrepancy between the ES recommendation and the agriculture practices documented in the field books. When this discrepancy was discussed with the ES users, we found that this discrepancy was due to their rejection of the ES recommendation and not due to bad use of the system. Therefor, we concluded that the usability of the system in the second experiment was high. 7.2 Expert system applicability The applicability can be measured by comparing the ES recommendation and to what extent the ES users have applied them. This discrepancy must not be due to bad usability of the system. In the first experiment, the discrepancy between the ES results and the applied practices by the users were due to bad use of the system. In the second experiment (CLAES, 1999), it was difficult to quantify the comparison result as it was found that sometimes the recommendations are applied partially. Hence qualitative measures were found more appropriate, especially in the strategic part. The applicability of the modules: Pre- Cultivation Pest Control, Planting, and Weed Control was found low because in Noubaria fields’ users did not accept the ES recommendations of the pre-cultivation pest control and the planting modules. In the weed control module the actual practice is different from the ES advice. The applicability of the modules: Tillage and Fertilization are moderate, as the ES fields’ users did not accept the ES recommendation in about 50% of the cases. The applicability of the modules Diagnosis and Treatment were above moderate as the advice of the ES fields supervisors matches the advice generated by the ES in the range of 80 to 87.5% in diagnosis and 50% of the cases in treatment. The applicability of the modules: Variety Selection, Irrigation, and Harvest are high. In the Variety Selection, the ES recommendations are compatible with the actual varieties cultivated in the ES fields. In the Irrigation, the Delta area (there is only 10% difference). In the Harvest, the ES recommendations are compatible with the actual practice in the ES fields. 7.3 Need of Expert System In order to measure the need of NEPER, a comparison has been done between the advice given by the researchers and extension workers supervising the control field in the experiment locations and the advice that would be generated if NEPER were used. In the first experiment (CLAES, 1996), examining the comparison results it was found that the ES performance is better in 76% of the cases, and hence there is a great need for having the ES. In the second experiment (CLAES, 1999), it was found that there is a high need for the ES modules: Tillage, Irrigation, Fertilization, Diagnosis, and Treatment. In the Tillage module, it was found that the performance of the ES is better as all control fields supervisors did not apply laser and plowers, appropriately. In the Irrigation module, it was found that the ES recommends less water than what was recorded in the control fields books. In the fertilization module, it was found that the ES is better as ES recommends the adequate quantities of phosphorus and potassium fertilizers whereas some control fields did not apply these types of fertilizers at all. In diagnosis and treatment modules, the performance of the ES is better as the advice of the control fields supervisors match the advice generated by the ES in only 37.5% of the diagnosis cases, and 20% of the cases in treatment. The experiment showed that there is a need for such module especially if the treatment part is modified to be more applicable. 8. Expert System Enhancements According to the results obtained from field-testing, the following enhancements were done: • Arabic language support was introduced. • The irrigation module was revised to be accepted by users. • User interface become more flexible. • Basic information about the field and the enviromnent have been included in the reasoning 6 (i.e. drainage system, previous crops, water source, length and width of the field, etc.) • The variety selection module has been enhanced to produce the most suitable variety for each field and produce justification for this selection. • Basin recommendation has been revised completely. • The harvest module has been enhanced to generate real advice about the suitable date of start harvest The following enhancements were also suggested and the ES are going to include them: • Most of the users were unable to understand what was meant by some operations so, more explanation like video clips should be provided. • Some of the terms are difficult to understand, e.g., “spindly stem” and “leaf chlorosis”. Consequently, pictures are necessary for symptoms at different plant growth stage. • Currently, the ES is capable of diagnosing sever nutrition deficiency. However, it is not equally capable of detecting early stages of nutrition deficiency. This should be rectified. A very good example of this is Nitrogen deficiency. • Drought and Water Logging should be covered by the system specially that their symptoms coincide with the symptoms of Nitrogen def. and Potassium def. 9. CONCLUSION The work done in this project has revealed and emphasized the effectiveness and importance of ES as a decision support tool for extension services. It was very clear that there is a difference in the advice quality and consistency given by the ES and the extension agriculture engineers. In the mean time, field experiments showed that Usage of ES has an economic and environmental impact. Currently there are efforts to disseminate NEPER, nation wide, and to avail it on the Internet. The field testing was found to be very useful as many aspects of the usability, applicability, and need were not possible to be identified without this field test. NEPER was found to be user friendly, and can be used by both researchers and extension workers. The recommendations generated by NEPER were applicable in most of the cases. The cases that were not accepted by the researchers and extension workers conducting the experiment, were discussed and the right recommendations were included in the succesor version. Most of the NEPER modules are found to be needed. The modules which were found not needed , were examined. The result was that this was not needed by the researchers and extension workers conducting the experiment but they are badly needed by the growers and extension workers in remote locations. REFERENCES Adrion, W., Branstad, M., Cherniovsky, J.'Validation (1982) "Verification and Testing of Computer Software" ACM Computing Surveys, Vol. 14, No. 2,1982 Chandrasekran, B. (1986). Generic Tasks in Knowledge-Based Reasoning: High-Level Building Blocks for expert system design. IEEE Expert, 1(3), 23-30. Chandrasekran, B. (1983). Towards a Taxonomy of Problem Solving Types. AI Magazine, 4(1), 9-17. CLAES (1996) "Validating NEPER Wheat Expert System and CERES Wheat Simulation Model", Technical report, No: TR/CLAES/ATUT(1)/3/96.12, 1996. CLAES (1999) "Validating NEPER Wheat Expert System - Field testing for season 97/98", Technical report, No: TR/CLAES/ATUT(w4)/5/99.2, 1999. CLAES (2001) "Validating NEPER Wheat Expert System - Field testing for season 98/99", Technical report, No: TR/CLAES/ATUT(w4)/10/2001.3, 2001. Nazareth, D (1989) Issues in the verification of knowledge In Rule- based System; International Journal of Mas-Machine Studies, Vol.30, 1989, PP.255-271. O' Keefe R.M (1990) "Consultant Report" Report No-CR-88-024-08 the Expert Systems for Improved crop management project.Project No EGY/88/024, August 1990 O' Keefe, R.M., O. Balci, and E. P. Smith, (1987) “ Validating Expert System Performance “ IEEE Expert, Vol. 2, No. 4, Winter 1987, PP 81-90. O' Leary, D., O'Keefe, R. (1989) "Verifying and Validating Expert Systems", Tutorial: MP4, IJCAI,1989. Rafea, A. (1996) "Natural Resources Conservation and Crop Management Expert Systems", Workshop on Decision Support Systems for Sustainable Development, UNU/IIST, Macau. 26 February - 8 March, 1996. Gomez,F., & Chandrasekran, B. (1981). Knowledge Organization and Distribution for Medical Diagnosis. IEEE Transactions on Systems, Man, and Cybernetics, SMC-11(1), 34-42. Kamel, A., Schroeder, K., Sticklen, J., Rafea,A., Salah,A., Schulthess,U., Ward, R. and Ritchie, J. (1994). Integrated Wheat Crop Management System Based on Generic Task Knowledge Based Systems and CERES Numerical Simulation. AI Applications 9(1):17- 27 
work_b4jj37b2uje75bm5llggmalcua	----	Management Studies, Nov.-Dec. 2019, Vol. 7, No. 6, 543-551 doi: 10.17265/2328-2185/2019.06.004 Civic and Social Applications to Support Public Policies on Health Eduardo Amadeu Dutra Moresi, Michel Carmo Lopes, Júlio Cezar Alves dos Santos, Marcos Augusto Alves Tito de Morais, Mário de Oliveira Braga Filho, Jair Alves Barbosa Catholic University of Brasília, DF, Brazil The discussion with emphasis on social control is expressed in new guidelines for its effectiveness through normative instruments and the legal creation of institutional spaces that guarantee the participation of organized civil society in the direct supervision of the executive in the three spheres of government. Social participation for the strengthening of public health in Brazil has undergone complex processes of change, which have resulted in an increasingly qualified, deliberative, independent, and representative social control system. However, policy makers need to identify new technologies and carefully analyze their potential before they begin to exert their breaking power in the economy and society. Mobility and unlimited connectivity deserve special attention from the public power for the ability to achieve the digital inclusion of the citizen through smartphones. This requires a transformation of the governance structure into a platform that is based on the tripod transparency, participation, and collaboration, which relies on the Internet, social networks and connectivity. Keywords: social participation, public health policies, mobile devices, civic social applications Introduction  Popular participation in health management is provided by the Federal Constitution of 1998, Article 198, which deals with the guidelines of the Unified Health System (SUS): decentralization, integrality, and community participation. These guidelines guide the organization and operation of the system, in order to make it more adequate to meet the needs of the Brazilian population (Rolim, Cruz, & Sampaio, 2013). Its consolidation and regulation were recommended by the Organic Laws of Health No. 8080/90 (Brazil, 1990a) and No. 8142/90 (Brazil, 1990b), which established the guidelines and norms for the SUS, as well as Eduardo Amadeu Dutra Moresi, Ph.D., professor, School of Education, Technology and Communication, Catholic University of Brasília, DF, Brazil. Michel Carmo Lopes, MBA, assistant professor, School of Education, Technology and Communication, Catholic University of Brasília, DF, Brazil. Júlio Cezar Alves dos Santos, MBA, assistant professor, School of Education, Technology and Communication, Catholic University of Brasília, DF, Brazil. Marcos Augusto Alves Tito de Morais, MBA, assistant professor, School of Education, Technology and Communication, Catholic University of Brasília, DF, Brazil. Mário de Oliveira Braga Filho, M.Sc., professor, School of Education, Technology and Communication, Catholic University of Brasília, DF, Brazil. Jair Alves Barbosa, M.Sc., professor, School of Education, Technology and Communication, Catholic University of Brasília, DF, Brazil. Correspondence concerning this article should be addressed to Eduardo Amadeu Dutra Moresi, Universidade Católica de Brasília, QS 07-Lote 01-EPCT-Bloco N-71966-700, Brasília, DF, Brazil. DAVID PUBLISHING D CIVIC AND SOCIAL APPLICATIONS TO SUPPORT PUBLIC POLICIES ON HEALTH 544 regulated their organization and functioning, criteria of transfers to states and municipalities, besides disciplining social control in accordance with the representations of the state and municipal health criteria. The SUS brought the expansion of health care to the population, allowing a new look at actions, services, and care practices, which were guided by the principles and guidelines: universality of access to health services; completeness of assistance; equity; political and administrative decentralization; community participation; regionalization and hierarchization. Popular participation and social control in health, among the principles of the Unified Health System (SUS), stand out as having great social and political relevance, since they constitute a guarantee that the population will participate in the process of formulating and controlling the public health policies. In Brazil, social control refers to community participation in the decision-making process on public policies and control over state action (Arantes, Mesquita, Machado, & Ogata, 2007). On the other hand, advances in Information and Communication Technology (ICT) should be part of a positive agenda for inclusion, increased transparency of government decisions, greater popular participation in policymaking, and thus a real increase in citizenship and social control. The advancement of cloud computing technologies, mobile devices, and social networks has shown great potential in contributing to this greater integration between government and society. All of them make great use of the possibilities of the Internet as a new arena for social participation, political activism and business. The issue of digital activism is recent and still lacks the tools to effectively promote effective action, integrating people with government decisions and the latter with the commitment to solve their problems. However, the ordinary citizen still cannot use this data without any previous treatment that facilitates its interpretation. Civic technologies seek to fill this gap by using the Internet and mobile devices to facilitate access and interpretation of data opened by ordinary citizens. In this sense, the objective of this work is to present the use of mobile devices to stimulate social participation to strengthen public health policies. Three examples of civic social applications that have been developed in iOS are presented, employing the Challenge Based Learning (CBL) methodology (Nichols, Cator, & Torres, 2016). Citizen Participation and Social Control The discussion with emphasis on social control in the new Constitution is expressed in new guidelines for its effectiveness through normative instruments and the legal creation of institutional spaces that guarantee the participation of organized civil society in the direct supervision of the executive in the three spheres of government. Thus, all citizens’ rights to health are guaranteed by the Federal Constitution, which reiterates that it is the duty of the State to guarantee the right to health. In Brazil, since 1988, public health policies have been guided by the principles of universality and equity in access to actions and services and by the guidelines of decentralization of management, integral care, and community participation, in the organization of a SUS in the national territory (Rocha, 2008). Essential for participation is the reduction of the distance between society and the manager of public policy, understanding this as something constant in everyone’s life (Pêgo, as cited in Coelho, 2012, p. 145). That is, the public policies are not exclusive initiatives of the state apparatus, but fruit of interlocution and agreement between social actors with diversity of interests and needs. Therefore, communication as a dialogic process is fundamental to citizen participation and social control. Mobile devices will soon become a tool for inclusion and social participation. Therefore, it is important to ensure adequate and sufficient CIVIC AND SOCIAL APPLICATIONS TO SUPPORT PUBLIC POLICIES ON HEALTH 545 access to information produced by SUS and the right of each to express them, to be heard and considered (Coelho, 2012). In short, the Brazilian government still faces difficulties in consolidating the practice of social participation, even though it has legally established spaces for this purpose, since the 1988 Constitution. Despite the indisputable quantitative advances observed in recent years, the consolidation of social participation in Brazil still faces numerous challenges, especially in terms of improving the quality and effectiveness of social participation spaces. In fact, many steps must be taken until, in fact, the deliberations of civil society are sent to the appropriate bins of the Federal Public Administration and are materialized in measures and public policies appropriate to the population. Mobile Devices The advance of technology continues to drive economic growth and, in some cases, trigger disruptive change. Economically disruptive technologies—such as semiconductor microchips, the Internet, or steam power in the Industrial Revolution—transform the way people live and work, allowing new business models and providing an opening for new actors to change the established order. Business leaders and policy makers need to identify potentially disruptive technologies and carefully analyze their potential before they begin to exert their breaking power in the economy and society. However, only a few technologies have the potential to transform reality by changing the way people live and work, redirecting the perception of value and leading to entirely new products and services. The McKinsey Global Institute (Manyika, Chui, Bughin, Dobbs, Bisson, & Marrs, 2013) conducted a study that outlines the 12 technologies that have a combined potential economic impact of tens of trillions of dollars per year from 2025. Trends are:  mobile internet: increasingly cheaper mobile computing devices with higher capacities and better Internet connectivity;  internet of things: networks of sensors and low-cost actuators for data collection, monitoring, decision-making and process optimization;  cloud technology: use of hardware and software resources delivered through a network or the Internet, often as a service. The power of new technologies is everywhere. Social media was virtually unknown a decade ago. Today, almost one billion people have Facebook accounts, establishing a new order in how to socialize and interact with friends, family, and colleagues. Technologies such as mobile Internet are helping to accelerate economic development by enabling millions of people in remote areas of developing regions to be included in the 21st century global economy. Gartner (Dreyfuss, 2014) has identified four converging forces that will impact organizations and their relationship to their external environment. The forces are: cloud computing, which leads organizations to rethink their IT infrastructure investments due to decreasing costs to meet their needs; social, which allows a much broader and deeper involvement of citizens in their relationship with government agencies; information, which assumes a prominent role in any organization, with data in various formats and from different sources, offering more opportunities for more accurate analysis and more informed decisions; mobile, with the explosion of devices (smartphones and tablets) as the main means of access to information and social interaction. CIVIC AND SOCIAL APPLICATIONS TO SUPPORT PUBLIC POLICIES ON HEALTH 546 Mobility and the unlimited connectivity of people with the surrounding world through mobile devices is already a reality and the growth of this trend is one of the main bets of companies focused on the future. With the advent of smartphones and the development of a wide variety of applications, people have come to enjoy an explosion of connections. Never before have individuals been so connected to other people, environments, businesses, and objects. Today, interaction is frequent and organizations already understand the potential of this change and invest in the development of new sources of value. It is possible to perceive the presence of new technologies in the creation of new business models and services online, in the growing generation of information in real time, in the effective identification of users when accessing systems and equipment, in the global management of operations, in the refinement of intelligent operations, the innovative offering of cloud computing, the expansive use of social networks, and the care for protection and privacy while exchanging all of these interconnected data. In a few years, Internet-enabled mobile devices will move from a luxury to few to a way of life for more than a billion people who have smartphones and tablets (Manyika et al., 2013). The ubiquitous connectivity and explosive proliferation of applications are allowing users to adapt their daily routines to new ways of knowing, perceiving, and even interacting with the real world. Mobile Internet technology is evolving rapidly, with intuitive interfaces and new formats, including handheld devices. Mobile Internet also has applications for businesses and the public sector, enabling more efficient delivery of many services and creating opportunities to increase workforce productivity. Access to government data enables the creation of applications that promote social control and citizen participation, as well as creating conditions for the improvement of public policies and services. There is a transformation of the governance structure into a platform that is based on the tripod transparency, participation, and collaboration, and relies on Internet technologies, especially on social networks and mobile device connectivity (O’Reilly, 2011). The emergence of civic applications enables the creation of a layer that adds value to the open government data, allowing the application of a business model approach, further enhancing the possibilities of interaction with society (Albano, 2014; Germano, 2013). Bimber (2000) highlights the possibilities of the strong relationship between technology and citizenship, emphasizing that especially the younger ones feel comfortable with the political debate in the digital format. The author stresses that discussion should not be limited to naming specific technologies as a source of civic engagement, but rather to highlight the flow of information as a transformer, whether flowing through the Internet, the television system, radio, mobile applications, or other emerging technology. Another interesting aspect is that more and more technologies tend to integrate, and thus their boundaries of distinction will become more difficult, keeping the debate about the flow of information itself and its potential. As the development of these applications is still in an initial state, there is still a lot of room for improvement in Brazil’s open data infrastructure. Brito, Costa, Garcia, and Meira (2015) mention some aspects that should be observed: the number of public bases which is still low; the need to keep the data up to date; removal of the need to register or obtain a license; standardization of data; the creation of return channels that allow interaction with the citizen. However, only the publication of data by governments does not guarantee citizen participation. New technologies can bring about significant changes in societies, particularly when the citizen becomes the protagonist through his active participation in the process of transformation of society. The possibility of CIVIC AND SOCIAL APPLICATIONS TO SUPPORT PUBLIC POLICIES ON HEALTH 547 questioning the quality of public services or evaluating public policies, guarantees the citizen’s role. Technology represents new ways of doing things, and once mastered, it creates a lasting change that organizations and people do not “unlearn”. Methodology The methodology followed for application development comprised the following steps:  bibliographic research: Studies and bibliographical surveys were carried out to the knowledge of the state of the art of the proposed theme, being carried out consultations to the bases of the CAPES Portal and to the Brazilian Digital Library of Theses and Dissertations, besides the reference documents on the Unique System of Cheers;  definition of the application scope: use of the Challenge Based Learning (CBL) method to define the scope of the application. At this stage, the key actors were identified to achieve the proposed objectives, in addition to face-to-face interactions for needs assessments and requirements for application development;  application development using Scrum (Priklandnicki, Willi, & Milani, 2014);  testing and publishing: The developed application was tested for 30 days. After the suggestions adjustments were received, the application was published to the Apple Store. The Health Map Application The Health Map is an application developed for the iOS platform, whose main objective is to provide useful information to the citizen about the health institutions registered in the CNES (National Register of Health Establishments), which is maintained by the Ministry of Health Datasus of the information query; the application allows the citizen to evaluate the health establishment and report on incorrect data found in the CNES register. In the application of the CBL method the great idea was identified: to develop an application with information useful to the citizen about Brazilian health facilities, indication of routes of access from the current location of the user to the desired organization, and dynamic searches depending on the service area doctor. The above issues were submitted to two public health experts, a Datasus technician, two designers, and two developers for the iOS platform. The synthesis of the answers allowed identifying the challenge of developing the Health Map with the following functionalities: to visualize the establishments closest to the user; to search for the name of the establishment, medical specialties, or category of health institutions; to interact with information from health facilities, such as making calls, sending e-mails, conducting evaluations, adding favorite establishments for easy access, among others; to navigate assistant to the chosen facility; to access to the list of professionals, specialties, and services of the establishment; report of problems regarding the information presented by the institution, and the user can suggest corrections; evaluation of the health establishment. Figure 1 shows some screens of the application. The first screen allows you to identify health facilities close to the user’s location. When choosing an establishment, you can have access to your complete information, as shown on the second screen. The following screen shows the evaluation categories for each health establishment. In order to perform the evaluation, the user must be registered in the application. This was a suggestion from public health experts seeking to avoid non-relevant information. The evaluation comprises the following attributes: general evaluation; waiting time; availability of physicians; infrastructure; medicines CIVIC AND SOCIAL APPLICATIONS TO SUPPORT PUBLIC POLICIES ON HEALTH 548 and products; equipment and beds; accessibility. The last screen displays the specialty search functionality. The result shows the establishments closest to the user’s current location. It should be noted that the application has other screens that are useful to the user. One is the navigation map that allows you to trace a route between the user’s current location and the health facility of interest. (a) (b) (c) (d) Figure 1. App Screens Health Map: (a) home screen, (b) health facility information, (c) evaluation, (d) research on specialties. The Health Map is available for free download in the App Store in Version 1.3.4, being compatible with iPhone, iPad, iPod, and Apple Watch, in Portuguese and English. The application is being used by more than 5,000 people in Brazil and more than 300 countries like: United States, Japan, Portugal, and Dominican Republic. Information is available on more than 270,000 registered health institutions in CNES, covering the entire country. The conceptual approach to application development highlights the importance of mobility and connectivity in generating ideas for solving everyday problems. Yet only ideas are not enough to reach the solution. They must be implemented in products or services that meet the needs of society. The Health Map application presents the following relevant points:  first App published with information from all health facilities registered at CNES;  covers all health facilities in Brazil;  establishes a channel of communication with citizens, allowing them to participate in the evaluation of public health policies;  allows the reporting of problems on the information registered and published in CNES by Datasus;  is available in English and Portuguese, which facilitates use by foreigners;  was conceived as a service that allows the citizen to interact with the public power. The VacinApp Application VacinApp was developed for the iOS platform, whose main objective is to offer an application that allows managing multiple vaccination cards. In the application of the CBL method the great idea was identified: to develop an application that registers and controls the vaccines of the whole family. To identify the challenge, the following key question was raised: How to manage vaccination cards digitally, alert the immunization campaigns in your region and the location of vaccination posts? CIVIC AND SOCIAL APPLICATIONS TO SUPPORT PUBLIC POLICIES ON HEALTH 549 From the answer to these questions the following challenge was defined: to create an application that can manage multiple profiles of vaccines cards, using geolocation for the identification of health posts. The development team counted on the guidance of public health experts who indicated bibliographies of support (Brazil, 2014; 2016). VacinApp is an application for the Brazilian population with the purpose of facilitating the visualization of the user’s vaccination card, as well as identifying the nearest Vaccination Centers where a vaccine can be taken. People who need to better organize their vaccines and stay informed about vaccination campaigns in their region will realize the usefulness of the application. Table 1 presents the main features of the application and their respective descriptions. Table 1 VacinApp Features Functionality Description View the vaccination card The vaccination card is organized and modularized by age, profile, and user needs. Manage family cards It manages the user profiles by cart, through a simple menu that allows the familiar control. Modify the vaccination card You can add new vaccines to your chart, delete existing records, and change expiration dates. Export vaccination card You can export the carton to the iCloud cloud, in pdf, by e-mail (jpeg). View map of health posts A map containing all vaccination posts in Brazil can be chosen and navigated. The presentation always starts with the user’s current location. Publish campaign reminders There is a welcome screen containing reminders of vaccination campaigns and locations they are currently experiencing. VacinApp’s main screen presents the user with a map and the current user location and the nearest health centers, as shown in Figure 2. The lower navigation bar has the following options: map, vaccination card, recent feedbacks, and settings. As soon as the user enters, a login screen is shown. The logon process can be done by registering or by using a social media account with Facebook and Twitter. The app enables the evaluation of health centers and vaccination points. The two last screens consist of the latest evaluations, and the review notes for each evaluation. The facial expression of the icon reflects if the evaluation is positive or negative. This feature is interesting, as it makes the explicit opinion of the community regarding the vaccination centers. Mainly informal registries, health secretariats can use this information to rank and evaluate the services provided by the center. Before this publication, public health experts, auditors from the Tribunal de Contas da União (TCU) and users have reviewed the application. The suggestions were blended into the new version of the app. VaccinApp is available as a free download on the App Store, on the 1.1 version, compatible to iPhone, iPad, and iPod Touch. The conceptual approach to application development underscores the importance of mobile devices in transforming ideas into innovative products or services. The problem to be solved by VacinApp is aimed at the loss of vaccination cards, the difficulty of finding vaccination posts, and the dissemination of immunization campaigns. At the time of application development, many adults were consulted about the vaccine card, most of whom replied that they did not have the document. Others said they were unaware of the vaccines they should take and the timing of immunization. Therefore, the relevant points of the application are: management of multiple profiles, including dependents; identification of the nearest vaccination posts; possibility of exporting each vaccine card; information on community opinion on vaccination posts; information about immunization vaccines. CIVIC AND SOCIAL APPLICATIONS TO SUPPORT PUBLIC POLICIES ON HEALTH 550 (a) (b) (c) (d) Figure 2. App VacinApp screens: (a) home screen, (b) login to the application, (c) evaluation of the Vaccination Point, (d) shared opinions about care at the Vaccination Posts. Conclusion The objective of this article was to present the use of mobile devices to stimulate social participation for the strengthening of public health policies. Initially, aspects were discussed in the legislation creating the Unified Health System emphasizing the importance of popular participation and social control in health, which stand out for creating conditions for the population to participate in the process of formulation and control of public health policies. Institutional channels of co-management with the state have also been created, which have brought about significant changes in political power relations and the distribution of responsibilities between the state and society, and between the three spheres of government—federal, state, and municipal. However, citizen participation and social control presume the reduction of the distance between society and the manager of public policy. Public policies are not exclusive initiatives of the state apparatus, but fruit of interlocution and agreement among social actors with diversity of interests and needs. The shortening of this distance is still a great challenge. The digital inclusion of citizens through mobile devices begins to gain notoriety. The article explores the potential of mobile devices as a means of social inclusion and of stimulating citizen participation. The mobility and unlimited connectivity of people, through smartphones, has transformed the reality of society and consolidated the growth of this trend, being one of the main bets for the future. Smartphones and the development of a wide variety of applications have resulted in an explosion of connections, in which individuals are now connected to other people, environments, businesses, and objects. Exploring this trend of mobility and connectivity with the development of civic applications should be the most viable alternative to strengthen social control and participation. Therefore, access to open government data by civic applications that transform them into intelligible information for ordinary citizens opens up a new perspective by creating conditions for improving public policies and services. It is expected that there will be a transformation of the governance structure into a technological platform that is based on the tripod transparency, participation, and collaboration, mainly in the social networks and the connectivity of the mobile devices. CIVIC AND SOCIAL APPLICATIONS TO SUPPORT PUBLIC POLICIES ON HEALTH 551 References Albano, C. S. (2014). Dados governamentais abertos: Proposta de um modelo de produção e utilização de informações sob a ótica conceitual da cadeia de valor. São Paulo: Universidade de São Paulo. Arantes, C. I. S., Mesquita, C. C., Machado, M. L. T., & Ogata, M. N. (2007). O controle social no Sistema Ú nico de Saúde: Concepções e ações de enfermeiras da atenção básica. Texto & Contexto Enfermagem, 16(3), 470-478. Bimber, B. (2000). The study of information technology and civic engagement. Political Communication, 17(4), 329-333. Brazil. (1990a) Lei nr 8.142, de 18/12/1990: Dispõe sobre as condições para a promoção, proteção e recuperação da saúde, a organização e o funcionamento dos serviços correspondentes e dá outras providências. Brasília: Congresso Nacional. Brazil. (1990b). Lei nr 8.080, de 19/09/1990: Dispõe sobre a participação da comunidade na gestão do Sistema Ú nico de Saúde (SUS) e sobre as transferências intergovernamentais de recursos financeiros na área da saúde e dá outras providências. Brasília: Congresso Nacional. Brazil. Ministério da Saúde. (2014). Para entender o Controle Social na Saúde. Ministério da Saúde, Conselho Nacional de Saúde. Brasília: Ministério da Saúde. Brazil. Ministério da Saúde. (2016). Calendário Nacional de Vacina. Retrieved from http://portalsaude.saude.gov.br/index.php/o-ministerio/principal/leia-mais-o-ministerio/197-secretaria-svs/13600-calendario- nacional-de-vacinacao (access on 24-05-2016) Brito, K. S., Costa, M. A. S., Garcia, V. C., & Meira, S. R. L. (2015). Is Brazilian open government data actually open data?: An analysis of the current scenario. International Journal of E-Planning Research, 4(2), 57-73. Coelho, J. S. (2012). Construindo a Participação Social no SUS: Um constante repensar em busca de equidade e transformação. Saúde Soc. São Paulo, 21(supl. 1), 138-151. Dreyfuss, C. (2014). Nexus of forces, 2014: Unleashing the power of digitalization. Retrieved from http://www.afsug.com/library/documents/saphila2014_presentations/Day1/BALLROOM_A/Nexus%20of%20Forces,%2020 14%20Unleashing%20the%20Power%20of%20Digitalization%20-%20Cassio%20Dreyfuss.pdf (access on 08-07-2015) Germano, E. C. (2013). Modelos de negócios adotados para o uso de dados governamentais abertos: Estudo exploratório de prestadores de serviços na cadeia de valor dos dados governamentais abertos. São Paulo: Universidade de São Paulo. Manyika, J., Chui, M., Bughin, J., Dobbs, R., Bisson, P., & Marrs, A. (2013). Disruptive technologies: Advances that will transform life, business, and the global economy. NY: McKinsey Global Institute. Nichols, M., Cator, K., & Torres, M. (2016). Challenge based learner user guide. Redwood City, CA: Digital Promise. O’Reilly, T. (2011). Government as a platform. Innovations: Technology, Governance, Globalization, 6(1), 13-40. Priklandnicki, R., Willi, R., & Milani, F. (2014). Métodos ágeis para desenvolvimento de software. Porto Alegre: Bookman. Rocha, E. (2008). A Constituição Cidadã e a institucionalização dos espaços de participação social: Avanços e desafios. In Associação nacional dos auditores fiscais da receita federal do Brasil—ANFIP. 20 anos da Constituição Cidadã: avaliação e desafios da Seguridade Social. Brasília: ANFIP. Rolim, L. B., Cruz, R. S. B. L. C., & Sampaio, K. J. A. J. (2013). Participação popular e o controle social como diretriz do SUS: Uma revisão narrativa. Saúde em Debate, 37(96), 139-147. http://www.afsug.com/library/documents/saphila2014_presentations/Day1/ 
work_b4yoshddkndarh4za7xngltfru	----	A semantic fuzzy expert system for a fuzzy balanced scorecard A semantic fuzzy expert system for a fuzzy balanced scorecard Fernando Bobillo a,*, Miguel Delgado a, Juan Gómez-Romero a, Enrique López b a Department of Computer Science and Artificial Intelligence, E.T.S.I. Informática, University of Granada, C. Periodista Daniel Saucedo Aranda, 18071 Granada, Spain b Economy and Business Management Department, University of León, Campus de Vegazana s/n, 24071 León, Spain Abstract Balanced scorecard is a widely recognized tool to support decision making in business management. Unfortunately, current balanced scorecard-based systems present two drawbacks: they do not allow to define explicitly the semantics of the underlying knowledge and they are not able to deal with imprecision and vagueness. To overcome these limitations, in this paper we propose a semantic fuzzy expert system which implements a generic framework for the balanced scorecard. In our approach, knowledge about balanced scorecard vari- ables is represented using an OWL ontology, therefore allowing reuse and sharing of the model among different companies. The ontology acts as the basis for the fuzzy expert system, which uses highly interpretable fuzzy IF–THEN rules to infer new knowledge. Results are valuable pieces of information to help managers to improve the achievement of the strategic objectives of the company. A main contri- bution of this work it that the system is general and can be customized to adapt to different scenarios. � 2007 Elsevier Ltd. All rights reserved. Keywords: Knowledge-based systems; Expert systems; Ontologies semantic web; Applied economy; Balanced scorecard; Fuzzy logic 1. Introduction Knowledge management plays a key role in the search for success in the current business world. Increasing spe- cialization and complexity of companies has given raise to the necessity of an integral management of own and for- eign resources, which involves and generates huge amounts of valuable data. Empresarial intelligence must be conse- quently more a cornerstone of the corporative strategy than simply an amalgam of disperse tools and procedures, if decision processes are expected to be faced with guaran- tees in order to achieve a joint and balanced global performance. Balanced scorecard (BSC) (Kaplan & Norton, 1992) is a decision support tool at the strategic management level which improves the satisfaction of the strategic objectives. Since it was proposed in the early 1990s, it has demon- strated its suitability to assist decision making in management. Nevertheless current balanced scorecard-based systems suffer from two problems. Firstly, variables which are to be measured have associated vagueness, being much more natural to refer to their values using a linguistic label instead a numerical value as frequently is done. Secondly, data do not have an explicit representation of their seman- tics; ad hoc solutions are usually implemented for each problem, making developers duplicate efforts and users cope with their specific details. Some solutions have been proposed to the first problem. Since fuzzy set theory and fuzzy logic (Zadeh, 1965) have proved to be successful in handling imprecise and vague knowledge, they have been combined with the BSC leads to fuzzy balanced scorecard (see Section 5 for details). However, such approaches also leave room for improve- ments in several aspects such as interpretability, modularity and accuracy. On the other hand, to the very well of our knowledge, there has not been any effort in the other direc- tion. Thus we have represented balanced scorecard data using an ontology, which allows to add semantics to them 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.09.020 * Corresponding author. Tel.: +34 958243194; fax: +34 958243317. E-mail addresses: fbobillo@decsai.ugr.es (F. Bobillo), mdelgado@ ugr.es (M. Delgado), jgomez@decsai.ugr.es (J. Gómez-Romero), elopez@ unileon.es (E. López). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 36 (2009) 423–433 Expert Systems with Applications mailto:fbobillo@decsai.ugr.es mailto:mdelgado@ mailto:jgomez@decsai.ugr.es mailto:elopez@ making easier knowledge base maintenance as well as reuse of components among different organizations. In this paper we present a new approach to a fuzzy BSC which improves the state of art by extending the number of variables and perspectives. We also present a fuzzy expert system for this fuzzy BSC. Its knowledge base relies on an ontology and its inference system derives new knowl- edge from fuzzy rules. The system is general and reusable, so every company can personalize it by providing their own meaning for the linguistic labels defined over the variables (e.g. what they consider a ‘‘high’’ value of some variable) and their own rules. The results of the expert system are highly interpretable pieces of information ready to be incorporated to managers’ decision making processes. The remainder of this paper is structured as follows. Section 2 provides some preliminaries on the fundamental theoretical aspects underlying this paper: balanced score- card, fuzzy logic and ontologies. In Section 3 we present our fuzzy balanced scorecard, describing the variables which take part in it. Implementation details are set out in Section 4. The description of our intelligent system starts by sketching the ontology and then we show how the rule- based engine computes the value of the output variables of the system. Section 5 evaluates our proposal with regard to the related work. Finally, some conclusions and ideas for future research are drawn in Section 6. 2. Background This section provides some basic background about the topics covered in the paper: Section 2.1 quickly overviews the original balanced scorecard, Section 2.2 refreshes the basic ideas in fuzzy sets theory and fuzzy logic, and Section 2.3 recalls the notion of ontology. 2.1. The balanced scorecard In 1992 Robert S. Kaplan and David P. Norton pro- posed the balanced scorecard (BSC) Kaplan and Norton (1992), a widely recognized tool to support decision mak- ing at the strategic management level which improves the satisfaction of the strategic objectives. The name reflects the objective of maintaining a balance ‘‘between short and long-term objectives, between financial and non-finan- cial measures, between lagging and leading indicators, and between internal and external performance perspectives’’ (Kaplan & Norton, 1996). The key innovation of the BSC is, as opposite to tradi- tional approaches which only consider the financial data, to supplement this information with additional non-mone- tary measures. In words of the authors, ‘‘financial measures are inadequate, however, for guiding and evaluating the journey that information age companies must make to create future value through investment in customers, suppliers, employees, processes, technology, and innovation’’. In particular, these authors consider four perspectives: Financial perspective obviously, measuring the financial performance of the company, customer perspective, mea- suring the satisfaction of the customers preferences, inter- nal business process perspective, measuring internal business results against measures from financial and cus- tomer perspectives, and innovation and learning perspective, measuring the ability of the company to adapt to changes. A more detailed description of the perspectives is out of the scope of this paper. On the other hand, Section 3 depicts the perspectives that we consider. Since the apparition of the BSC it has become an impor- tant field of theory and research. Many companies have successfully applied this tool and several variations to the original proposal have been investigated (for instance, see Section 5). 2.2. Fuzzy sets and fuzzy logic This section briefly reviews fuzzy sets theory and fuzzy logic; for more details a good reference is (Klir & Yuan, 1995). Fuzzy set theory and fuzzy logic, proposed by (Zadeh, 1965), are acknowledged as an appropriate formal- ism for capturing imprecise and vague knowledge. While in classical set theory elements either belong to a set or not, in fuzzy set theory elements can belong to a certain degree. More formally, let X be a set of elements. A fuzzy subset A of X, is defined by a membership function lAðxÞ which assigns any x 2 X to a value between 0 and 1. As in the classical case, 0 means no-membership and 1 full member- ship, but now a value between 0 and 1 represents the extent to which x can be considered as an element of X. All crisp set operations are extended to fuzzy sets. The complement, intersection and union set operations are per- formed by a negation function, a t-norm function (typi- cally, the minimum) and t-conorm function (typically, the maximum) respectively. Several membership functions can be used in the defini- tion of a fuzzy set. Some of the most used are the triangular and the trapezoidal function. A triangular function tria;b;cðxÞ (see Fig. 1a) is defined over the set of non-negative reals Rþ [f0g with a 6 b 6 c being real numbers. A trap- ezoidal function trza;b;c;dðxÞ is defined over the set of non- negative reals Rþ [f0g as in Fig. 1b, with a 6 b 6 c 6 d being real numbers. Note that a triangular function tria;b;c can be represented using a trapezoidal function trza;b;b;cðxÞ. One of the most important features of fuzzy logic is its ability to perform approximate reasoning (Zadeh, 1973), which involves inference rules with premises, consequences or both of them containing fuzzy propositions. Fuzzy rule- based systems have some advantages over other formal- isms: they provide a natural representation for human knowledge as well as a very interpretable model (since the semantics of the rules can easily be understood even for not experts users), are simpler, cheaper and more robust than their crisp versions and, last but not least, have shown to behave very well in practical applications. A fuzzy IF–THEN system consists of a rule base (a set of IF–THEN rules) and a reasoning algorithm performing 424 F. Bobillo et al. / Expert Systems with Applications 36 (2009) 423–433 https://isiarticles.com/article/351 
work_b5xqj5grzbfg3j54vqhpg3rlga	----	PII: 0306-4573(95)80003-C Pergamon Information Processing & Management, Vol. 31, No. 1, pp. 15-27, 1995 Copyright 0 1994 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0306-4573195 $9.50 + .oo 0306-4573(94)EOOO4-L TERMINOLOGICAL KNOWLEDGE STRUCTURE FOR INTERMEDIARY EXPERT SYSTEMS RAYA FIDEL Graduate School of Library and Information Science, University of Washington, Seattle, WA 98195, U.S.A. and EFTHIMIS N . EFTHIMIADIS Graduate School of Library and Information Science, University of California, Los Angeles, CA 90024, U.S.A. (Received 4 November 1993; accepted in final form 23 January 1994) Abstract -An intermediary expert system (IES) helps both end users and professional searchers to conduct their online database searching. To provide advice about term selec- tion and query expansion, an IES should include a terminological knowledge structure. Terminological attributes as well as other properties could provide the starting point for building a knowledge base, and knowledge acquisition could rely on knowledge-base techniques coupled with statistical techniques. The searching behavior of expert online searchers would provide one source of knowledge. The knowledge structure would include three constructs for each term: frequency data, a hedge, and a position in a classification scheme. Switching vocabularies or languages could provide a meta-schema and facilitate the interoperability of databases in similar subject domains. To develop such knowledge structure, future research should focus on terminological attributes, word and phrase disambiguation, automated text processing, and the role of thesauri and clas- sification schemes in indexing and retrieval. In particular, such research should develop techniques that combine knowledge-base and statistical methods and that consider user preferences. 1. INTRODUCTION An intermediary expert system (IES) helps users, professional searchers, and end users to conduct their searches of online bibliographic databases. Currently, most online bib- liographic databases provide for searching the titles, abstracts, and sources of the biblio- graphic items to be retrieved in addition to descriptors and identifiers which have been assigned by human indexers, if they are available. This article examines the knowledge base of an IES that provides advice about the selection of search terms, or search keys. It presents a proposal for an integrated approach that would include various methods and techniques that are available today. Recognizing that these methods and techniques could be integrated in a variety of combinations, the article presents one option that focuses on terminological attributes that is based on knowledge acquired from professional search- ers. This option creates a scenario that illustrates how various approaches can be used simultaneously and the effect such a combination would have on research. It examines what knowledge and information could be included in the knowledge base and how they could be organized. The article then shows what research would be required to support the devel- opment of this option. Most commercially available search systems require the use of Boolean operators, Thus, before searching a request, a user breaks it down into concepts, the representation of which would be linked with Boolean AND operators. For actual searching, each con- cept is represented by one or more search keys. A search key is a string of characters to Correspondence should be addressed to Raya Fidel, Graduate School of Library and Information Science, FM-30, University of Washington, Seattle, WA 98195. 15 16 R. FIDEL and E.N. EFTHIMIADIS be searched in the database. A search key, which represents a concept of a request, may consist of one or more words. The selection of search keys is at times a straightforward process; however, at other times it requires knowledge and expertise. Consider the request “attitude of students toward themselves during examination period.” The request can be broken down into three concepts: “attitudes toward them- selves,” “ students,” and “examinations.” A straightforward approach to searching would be to search on the keys as they appear in the request in all available fields and then to intersect the resulting sets (using the AND operator). A professional searcher, however, would likely see much more complexity in the request and would probably try a variety of other search keys that would result in better retrieval. A searcher would probably decide to express the concept “attitudes toward themselves” in a phrase such as “self-image” or “self-esteem.” Also, the searcher is likely to prefer to search the key “examination” only in the descriptor field because it is a common term; as a textword, it appears frequently in the text, often referring to concepts other than educational tests. An IES of the kind considered here would advise users of the most promising search keys to be used. It is well established by now that relying only on the words in a request is not suffi- cient for satisfactory retrieval (Svenonius, 1986). Indeed, research into query expansion- the process of supplementing the original query with additional terms-has been motivated by this observation (Efthimiadis, 1991). In addition, databases that use an indexing lan- guage require users to make another decision: whether to enter the search key as a text- word key, which would retrieve all bibliographic database records that include the key in any field of the record, or as a descriptor, which would retrieve only the records whose descriptor field includes the key. An IES of the type considered here should be able to help users in this decision as well. The interaction that takes place in information retrieval between users and the data- base searched can be described with the use of a simple two-stage model (Efthimiadis, 1991; Efthimiadis & Robertson, 1989). The model includes the end user, the intermediary mech- anism, and the database. The intermediary mechanism may be a human being or some soft- ware, such as a front-end system or an expert system. Here we consider an intermediary mechanism that is a machine, as described in Fig. 1. For simplicity in the discussion, the IES is treated as part of the retrieval system. However, an IES could reside anywhere; it could reside between the retrieval system and the user-supporting initial query formu- lation; it could be a front end or client at the user end, a front end at the database end, or an integral part of the retrieval mechanism. Depending on the characteristics of the particular request searched, the user’s level of expertise, and the database searched, an IES could provide help in three modes: 1. The system decides about the search key with no consultation with the user. 2. The system decides about the search key after interrogating the user. 3. The system presents options from which the user is asked to make a selection. The decision about which mode of advice to provide is situational. In general, we can identify two main sources which can provide knowledge to be utilized or incorporated in an IES (Efthimiadis, 1990). The first source of knowledge is the search intermediaries. Here, the approach that has been taken so far is to try to en- capsulate their skills in a system, such as in PLEXUS (Vickery et al., 1987), IR-NLI ----_----_--------_ I I I i I I Database I I I L__________________I Fig. 1. The role of an IES in the two stage model of interaction in information retrieval. Terminological knowledge structure 17 (Brajnik et al., 1986, 1988), and EP-X (Shute & Smith, 1992). The second source is the knowledge structures found in databases or embodied in search aids or indexing languages, such as thesauri or classification schemes like EP-X (Smith et al., 1989a, 1989b), MENUSE (Pollitt, 1988), and UMLS (Humphreys & Lindberg, 1989). It is promising, however, to integrate knowledge and information from both sources. For instance, terminological knowledge (i.e., knowledge about terms and their properties) acquired from professional intermediaries can point to terminological issues that are rel- evant to retrieval. At times, however, professional searchers may not be in a position to provide the best solution to a terminological problem because there is not enough infor- mation for them to make the most useful decisions. Given the existing techniques in in- formation retrieval, help can come from additional sources. Associative retrieval techniques create one of these sources. While not yet widely available, various techniques have been developed and tested over the last two decades. These techniques are not incompatible. It is possible to devise methods based on more than one associative retrieval approach, and such mixed methods may be appropriate for certain retrieval situations. Furthermore, it is also possible to combine associative and Boolean techniques to enhance both through knowledge-based retrieval techniques. Current prototype IES vary in the help they offer in terms of query negotiation aids, the selection of search keys, and query expansion. Because it is difficult to automate assistance offered at the query formulation stage, there have been various attempts to deal with it at the interface level (e.g., Pollitt, 1988; Thompson & Croft, 1989; Vickery, 1988; Vickery et al., 1986). Some of the systems use thesauri and classification schemes to assist query formulation, for navigation and retrieval (e.g., Frei & Jauslin, 1983; Monarch & Carbonell, 1987; Pollitt, 1987, 1988; Shoval, 1981, 1985; Smith et al., 1989a, 1989b; Vickery, 1988). A few experimental expert systems already incorporate in their knowledge bases knowledge that is pertinent to the selection of search keys. MedIndEx at the National Library of Medicine, for instance, incorporates the Medline indexing policy (Humphrey, 1989; Humphrey & Miller, 1987), a system at the American Petroleum Institute employs knowledge acquired from professional API indexers (Brenner et al., 1984; Martinez et al., 1987), and a system at BIOSIS incorporates knowledge on biological concepts in a seman- tic vocabulary (Vleduts-Stokolov, 1987). Although their knowledge bases could be incor- porated into IESs, these expert systems were designed initially to assist indexing rather than searching. To date, terminological knowledge as employed by searchers has not been explored. To focus the discussion on search-key selection, it is assumed that a request is already broken into its concepts, the databases are selected, and the user is prepared to enter search keys. Further, to provide advice in the selection of search keys and the field(s) to be searched, an IES must possess expert knowledge about the particular database that is being searched. Therefore, it is assumed here that such a system would provide advice for search- ing a defined set of databases covering a certain subject domain. 2. KNOWLEDGE USED BY EXPERIENCED SEARCHERS The first type of knowledge to be incorporated into an IES is knowledge about the selection of search keys acquired from experienced online searchers (Fidel, 1991b). This type of knowledge was collected through observations of searchers performing their reg- ular, job-related searches and through interviews with them. The study team analyzed search protocols, verbal protocols of thought processes while searching, and the transcripts of interviews, with 47 searchers performing a total of 281 searches in a variety of subject areas and library types. The analysis of the search and verbal protocols uncovered the intuitive rules that searchers used and resulted in a decision tree for the selection of search keys which is called the selection routine. The selection routine embodied in the decision tree describes the conditions that search- ers considered and the options that each condition generated. For example, the condition “a search key is mapped to a descriptor through an exact match” generated the following options: enter the descriptor, but if recall needs to be improved: add textword synonyms 18 R. FIDEL and E.N. EFTHIMIADIS to descriptors, or use generic descriptors in an inclusive mode (“explode,” or “cascade”), or add the next broader descriptor in the hierarchy. Data were gathered on the frequency with which the different options were selected. Thus, of the 228 cases in which a descrip- tor was an exact match and searchers wanted to increase recall, 72% of the time they entered textwords as synonyms, 25% they did an inclusive search, and 3% of the time they selected a broader descriptor. Also, data were gathered on reasons associated with spe- cial conditions for each option. Thus, searchers selected textwords as synonyms because the user insisted on using the terms, because they needed to perform a multidatabase search, or because they did not trust the descriptors and/or the indexing of the database. They performed an inclusive search when the query formulation included a relatively large num- ber of concepts, and they entered a broader descriptor when they thought the user would be interested in the broader descriptor as well. Data of the type just described could be incorporated into the knowledge base and the inference engine of an IES for a specific set of databases. The frequencies with which options were selected could be used to handle uncertainties. The frequencies, together with other factors, could also determine the mode of advice to be given. For example, if a search key were matched to a descriptor through an exact match, the system might automatically enter the descriptor without consulting the user. This would be reasonable, given the search- ing behavior of the professional searchers: 100% of the time when there was an exact match they entered the descriptor; only if recall needed to be improved they selected additional keys. The next step, then, is for the system to inquire if recall is satisfactory. If it needs to be improved, other questions can be asked and advice given. Selection routines by themselves are not sufficient for the IES to advise about the selec- tion of search keys. Clearly, the system must include the database’s thesaurus to be able to map search keys to descriptors. But it should include additional knowledge about the database as well. For example, it is important to include information about the indexing of the database- how often it is used by professional searchers, and the indexing policy that created it. This requirement is based on the finding that often searchers selected text- words when they did not trust the database’s vocabulary and indexing, and they frequently referred to the indexing policy when giving reasons for their search decisions. For instance, if exhaustive indexing is mandated in an educational database, searchers may enter a fre- quently occurring descriptor such as “students” as a major descriptor, limiting the retrieval to documents in which “students” is a central topic. In addition to the knowledge about the database, its controlled vocabulary, and index- ing, the IES must incorporate terminological knowledge and information about individ- ual search keys that is not available in thesauri and indexing manuals. For example, suppose a user with the request about students wishes to increase recall. The selection routine, as described earlier, shows that the most common option is to add the textword key “student” to the descriptor. A professional searcher, however, is not likely to select this option in an educational database because the textword occurs too frequently. How would an IES know to eliminate this option? To do so, the system would have to incorporate additional knowl- edge and information (for instance, information on the relative frequency of search keys in the database). Further, some databases do not have sources for terminological knowl- edge such as thesauri and indexing manuals because they are not indexed with controlled vocabulary. The system’s knowledge base, however, should also be able to provide advice for the selection of search keys in such textword databases. 3. TERMINOLOGICAL ATTRIBUTES Searchers considered a variety of attributes when they selected search keys (Fidel, 1991a). Some examples are the number of databases to be searched; the number of com- ponents in a request; and specific attributes of a term. The latter includes whether the term was added just to increase recall (ORing it) or used as a limiting factor (ANDing it); whether it occurred in the records of relevant citations and in which fields; whether it was mapped to a descriptor and through what kind of a match, exact or partial; and whether the user insisted it appears in the retrieved text. A close examination of these attributes Terminological knowledge structure 19 shows that most can be incorporated into an IES and used without additional knowledge. Terminological attributes are the one exception. Very few existing searching tools include terminological information perceived by searchers to be pertinent to the selection of search keys. Each search key has a variety of terminological attributes which are determined by the terms it includes. Here we focus only on those attributes that require special considerations for searching. The study of professional online searchers revealed that searchers consid- ered terminological attributes most often when they thought that a search key might not be a “good” textword; that is, if entered as a textword, it would produce unsatisfactory retrieval. Therefore, an IES would incorporate information about those attributes that determine the suitability of a search key to be entered as a textword. The terminological attributes identified by the searchers as being relevant to the choice of search keys conform with those discussed in the literature; see Fugmann (1982b), Svenonius (1983), and Knapp (1988). Thus, according to the study’s searchers, search keys with the following attributes are not suitable for textword searching: l The key has many synonyms. l The key is ambiguous; for example, archival “record” is different in meaning than database “record” or music “record.” l The key is vague; for example, “health promotion” has no agreed upon boundaries to its definition. l The key occurs too frequently in the database’s text.* It is well accepted by now that controlled vocabularies and indexing are necessary to overcome the retrieval problems caused by these terminological attributes. Clearly, an index language controls for synonyms, ambiguity, vagueness, and context dependence, and it is also useful as an alternative to highly frequent textwords. For instance, a study of the ERIC database found that the search key “student” retrieved 115,061 citations when entered as a textword, but only 1,242 as a descriptor (Markey et al., 1980). However, while incorpo- rating a database’s thesaurus into an IES knowledge base is necessary, it is not sufficient because two functions of an IES would require additional knowledge. First, to advise users about the selection of search keys an IES would recognize problems, or “diagnose” each key to determine whether it is good for textword searching. Second, to advise users search- ing textword databases, an IES would require terminological knowledge that is not included in current thesauri. What then would be included in its knowledge base for an IES to diagnose and pro- vide advice about the selection of search keys in all databases? The following are some exploratory ideas based on knowledge acquired from expert searchers. 4. TERMINOLOGICAL KNOWLEDGE STRUCTURE The IES’s knowledge base would include a variety of machine-readable sources of knowledge that are already available. It would include databases’ thesauri that are per- tinent to the subject domain, the relevant terminological databanks, machine-readable dictionaries, and Roget-like thesauri. In addition, the text stored in each database would be accessible to statistical manipulations. Given these sources, a knowledge structure would be constructed to include knowl- edge generated from these sources and from external, intellectual sources, such as subject experts, indexers, and users. The knowledge structure may be an expanded thesaurus or a semantic net; parts of it may reside permanently in the knowledge base, and others may be generated ad hoc, triggered by specific situations. In addition, each content-bearing term that is relevant to the subject matter which appears in the databases’ text would be included in the knowledge structure. *Frequency data by themselves are not considered to be terminological attributes. IPM 31:1-C 20 R. FIDEL and E.N. EFTHIMIADIS To diagnose problematic search keys and to advise in the selection of alternative or additional keys, the knowledge structure would include three constructs for each term, whether textword or descriptor: frequency data; a hedge; and a position in a classification scheme. 4.1 Frequency data The number of times a search key occurs in the database can help determine whether the key occurs too frequently. Frequency data is commonly supplied by current search systems in the form of number of postings. These data by themselves, however, are not sufficient to determine whether a key is too frequent to be used for textword searching. For instance, databases vary in size so it is unlikely that a universal number could be found that would designate the threshold of too-high frequency. One method to determine whether the frequency of a search key presents difficulties in searching is to compare the frequency with which the key occurs as a textword with that with which it occurs as a descriptor. An empirically based rule could then be established to determine which difference is significant. For instance, such a rule may state that if a textword occurs more than a certain number of times and if the difference is more than one order of magnitude, the search key is not suitable for textword searching. Note that even a highly frequent term might be used successfully (for instance, when it is entered in conjunction with other terms or when it is used as a limiting factor). Because a terminological knowledge base is limited to a particular subject domain, it is plausible to assume that a search key that is too frequent in one database is likely also to be frequent in the other databases covering the same subject. While this assump- tion still requires empirical validation, it supports the expansion of the method described earlier to databases which can be searched only with textwords because of lack of a con- trolled vocabulary. For such textword databases, a list of the most frequently occurring keys would be compiled. These keys would then be checked for frequency in databases with controlled vocabularies; the difference between textword and descriptor occurrences would be measured to determine the suitability of these keys for textword searching. Using this method, then, all the search keys that are not suitable for textword searching be- cause of high frequency would be designated as such. Finally, additional information drawn from the statistical approaches to information retrieval could be used to complement the approach discussed here. 4.2 Hedges Frequency data, including frequency of co-occurrence, is also instrumental in the creation of hedges. A hedge is commonly understood to be an “OR string of terms that cover a topic for which no single term has sufficient extension” (Sievert & Boyce, 1983, p. 491). Hedges are essentially clusters of semantically related terms. Thus, they have been used by searchers as a list of synonyms to expand a concept of a request. As clusters, hedges are the equivalent of the term clustering approaches of automatic indexing (strings, stars, cliques, and clumps). However, the latter are based on statistical associations, usually term co-occurrence, rather than on semantic content. The need for hedges was realized in the early 197Os, when online searching started to become common. Since then, hedges have been developed by individual libraries and searchers, as well as by database producers and search system vendors. Figures 2 and 3 show examples of predefined hedges. Other terminological tools, however, can be viewed as hedges. Piternick (1984), for example, shows that as of 1984, a variety of enhanced thesauri, synonym listings, and early forms of switching languages were available to online searchers to enhance their search- ing vocabulary. Thus, an enriched lead-in (entry) vocabulary can provide a list of textword synonyms for a descriptor, and a list of equivalent descriptors in a number of databases can serve the same function. For instance, the now defunct BRS’s TERM database incor- porated the social science descriptors of five major thesauri (Knapp, 1992). In addition to these intellectually developed tools, more ambitious, computer-assisted projects, such as UMLS, the Unified Medical Language System (Humphreys & Lindberg, 1989), are Terminological knowledge structure 21 Topic: tests and measurements 1 (test or tests or testing or subtests or pretesting).de. 2 (posttesting or inven or invent or inventory or inventories).de. 3 (surv or survey or surveys or scale or scales or seal or score).de. 4 (scores or measurement or measures or screening or exam).de. 5 (examination or examinations or questionnaire or questionnaires).de. 6 (rating or validity or psychometrics or sociograms).de. 7 (assessment or sociometry or piagetian-tasks).de. 8 semantic-differential or piagetian-tasks 9 1 or 2 or 3 or 4 or 5 or 6 or 7 or 8 Fig. 2. “Hedges” (Source: Psychological Abstracts: database PSYC on BRS). being developed to establish machine-generated tools to enhance the searching vocabu- lary in specific subject domains. Similarly, global thesauri, such as Roget’s thesaurus or Wordnet,* are also being used in an attempt to test their applicability/suitability as vocab- ularies in searching. An IES would integrate all such tools that are pertinent to its subject domain into its knowledge base. If a synonym list, or an enriched lead-in vocabulary, does not exist for a certain subject matter, it might be possible to develop it for the IES. While some auto- mated methods have been applied, most are still experimental. Experience indicates, how- ever, that it is realistic to develop such tools without automated techniques, or to verify the results intellectually of automated methods (e.g., Anderson & Rowley, 1992). Hedges would be dynamic in nature. Although technical considerations would deter- mine which parts of a hedge would be stored permanently in the knowledge structure and which parts would be developed ad hoc, hedges would evolve during the lifetime of an IES. Changes and developments in the vocabulary would be reflected in the text stored in the databases and in the revisions of thesauri and terminological databanks. Users and requests would also contribute to the refinement of hedges. Through machine-learning procedures, terms would be added to hedges or deleted from them. An example of such mechanism is provided by the experimental system TEGEN, which is designed to construct thesauri *Wordnet is available from Princeton University. Topic: Air pollution Suggestions for terms to OR: aerobic; AIR; air born; aerosol; asthma; aerial; branch; dust; exhaust; EMPHYSEMA; fume; inhal; lung; nose; pleura; pulmon; respirat; smoke; SMOG; throat; trachea; centilat; vapor. Fig. 3. A hedge from the Hedge Book of a medical library prepared for searching Medline. In this partial list, MeSH headings are printed in caps. 22 R. FIDEL and E.N. EFTHIMIADIS automatically. The system first acquires terms and relationships from users’ requests and then verifies relationships with users’ help (Guentzer et al., 1989). In the terminological knowledge structure, each node, or search key, would have a hedge that would include other keys, or hedge terms, which are associated with the node. For simplicity’s sake we assume here only direct relationship between each member and the node. Such a hedge would have to support various searching decisions in addition to providing lists of synonyms for query expansion. As a result, it would be a special kind of hedge, an expanded hedge: It would not only cluster terms, but it would include infor- mation about the relationship between each term in the hedge and the search key (the node), such as frequency of co-occurrence and semantic relatedness, and it may also include terms that are not synonyms. A hedge of a search key would include a list of all the terms, both descriptors and textwords, that co-occur with the key in the text stored in the databases. A designation of the relative level of frequency with which it co-occurs would be given for each term in the hedge, whether derived by co-occurrence analysis or otherwise. In addition, based on terminological tools such as machine-readable dictionaries, terminological databanks, and Roget-like thesauri, a semantic scale might be created that expresses the semantic related- ness between each term and the key would be designated; see Liddy et al., 1991; Krovetz, 1991 and Nutter et al., 1990. With these relationships explicitly expressed, an IES would provide advice concern- ing most of the terminological properties. Although this advice would be generated by the inference engine supported by the knowledge base, it is useful to discuss a few examples to show how this knowledge could be used. Because much research is still required to find effective methods to deal with terminological attributes, we give these examples only to illustrate how the knowledge could be structured, not how it should be structured. 1. A search key has many synonyms. A diagnosis would be facilitated by checking the number of hedge terms that relate highly to the key on the semantic scale. If the number is relatively large, a ranked list of terms would be presented to the user. First on the list would be the descriptors (if available), followed by highly sim- ilar hedge terms arranged in descending order of co-occurrence frequency with the key. This would encourage the user to enter descriptors, which is the preferred action according to the selection routine for the case of a key having many syn- onyms. If the key has already been used as a textword and the user is looking for additional synonyms to increase recall, those textword synonyms that do not co- occur with the key, or those that co-occur the least frequently, would be suggested as most promising. The rule could be further refined based on frequency data for each hedge term. In addition, probability estimates could be used to determine which term to include and how to weight individual terms in the hedge. 2. The search key occurs too frequently. Such a search key is usually not suitable for textword searching and the IES would suggest alternative terms. It would first retrieve from the hedge descriptors that closely relate on the semantic scale to the key. Next, hedge terms that are semantically related would be listed in descending order of co-occurrence frequency, an order that could be refined by term-occurrence frequencies. The user would first view the descriptors and then the synonyms that co-occur frequently with the key but are themselves suitable for textword searching. 3. The search key is ambiguous. Various mechanisms have been suggested to disam- biguate terms. Krovetz and Croft (1989) explain the early methods for word sense disambiguation using machine-readable dictionaries. Disambiguation tools used in the semantic vocabulary and the concept headings of the BIOSIS system are a com- bination of contextual restrictions, multiword entries, and an associated weighting technique, which is used when the corresponding meaning of the words cannot be derived from context (Vleduts-Stokolov, 1987). Ahlswede et al. (1988) and Nutter et al. (1990) describe the use of a machine-readable dictionary, the Webster’s Sev- enth New Collegiate Dictionary, as a source for identifying semantic relationships between index terms and for linking phrases to index terms. For disambiguation Terminological knowledge structure 23 they make use of the lexical-semantic relationships, the selection restrictions, and the verb categories available in Webster’s to create defining formulae. Wherever this information is not available, they rely on Sager’s linguistic parser. Veronis et al. (1990) describe a method which combines different machine-readable dictio- naries and connectionist models. The resulting neural network is used for word sense disambiguation. The knowledge structure proposed here facilitates another approach. In addition to descriptors, the user would be presented with clusters, created according to semantic relat- edness, of hedge terms that frequently co-occur with the ambiguous key. The ORed clus- ter that represents a desired point of view can then be ANDed with the ambiguous key to improve precision. For example, the ambiguous key “record” may have two clusters, one that includes terms typical of the database literature and another with terms pertinent to archives. A similar procedure can be used for search keys whose meaning depends on the context in which they appear. With machine-learning techniques, hedges would increase their usefulness. For exam- ple, it might be possible to address the problem of nonlexical expressions. Fugmann (1982a, p. 141) defines a lexical search key as “one which consists of a linear sequence of alpha- numerical symbols, which by general agreement is used to represent a certain meaning.” He also explains why nonlexical search keys are not suitable for textword searching. For example, the phrase “attitudes of students towards themselves” is nonlexical. It could, however, be expressed by two lexical keys: “students” and “self-image” (or “self-esteem,” etc.). Obviously, it is almost impossible to predict all the nonlexical expressions users would bring to an IES or that may occur in the text. With machine-learning techniques, however, it might be possible to analyze nonlexical expressions when they are presented to the IES. If a lexical presentation is identified, it could be stored for future requests which include this nonlexical phrase. For example, a proximity combination of the terms “attitudes” and “themselves” may reveal a high co-occurrence with “self-image.” Once it is established and verified that “self-image” is representative enough of “attitudes toward self,” the latter can be stored as a lead-in term. If a new type of relationship is found to be useful for retrieval, such as the frequency with which terms co-occur in requests, the new relationship and the associated terms could be added to the hedge. Future advances in theoretical and computational linguistics, in ter- minological research, and in automated text processing and retrieval would enhance these expanded hedges and enrich their capabilities and effectiveness. 4.3 Classification scheme A classification scheme would provide an overall semantic structure. It would estab- lish hierarchical links between keys, or nodes, based on the keys’ meaning and would des- ignate the type of such link: member/class, part/whole, etc. As an overall structure, the classification scheme would also facilitate the combination of nodes into concepts. For example, the nodes “baby, ” “dog,” and “puppy” would probably establish a variety of rela- tionships: both hierarchical and nonhierarchical; semantic; or pure statistical relationships. It would be the function of the classification scheme to point to the semantic equivalence between the combination of the first two keys and the third one. In a classification scheme, a concept may include more than one key, and a key may belong to more than one concept. Other relationships, which are neither hierarchical nor combinatorial but are relevant for retrieval, would already be expressed in the hedges. The classification scheme would support browsing in hierarchy and would help users to navigate the particular terminological knowledge structure it represents. In addition to these overall functions, a classification scheme could help disambiguate search keys by point- ing to their position in the classification structure or to resolve context-dependence issues. 5. DISCUSSION The development of search interfaces for online bibliographic databases that are geared to end-user searching, and the growing body of research about expert systems, has made 24 R. FIDEL and E.N. EFTHIMIADIS the user the focus of research in information storage and retrieval. This trend motivated the proposal that terminological knowledge acquired from expert searchers could guide the construction of a terminological knowledge structure for IESs. Future investigations are necessary for the development of terminological knowledge structures for IESs. Three research areas that would contribute to this development are terminological research, auto- mated text processing, and thesauri and indexing practice. 5.1 Terminological research The terminological attributes discussed in this article are those that were defined by searchers as problematic for textword searching. Clearly, there are additional attributes of terms that are possibly important for retrieval. For instance, even well-defined terms vary in their level of specificity, in their stability (whether they are well established or just a fad), or in how likely they are to appear in a text about the concept they represent. Termino- logical research should identify such attributes and examine their effect on retrieval. As part of this task, methods for recognizing attributes should be developed. Such methods may make use of statistical or other automated text analyses. 5.2 Automated text processing Research in automated text processing aims at developing methods to represent the content of documents for a number of purposes, including information retrieval. The pur- pose in information retrieval is to develop methods that would be effective in represent- ing any text for all users and requests. Such a global approach has been assumed to be necessary. After over four decades of research in automated text processing, there is a growing notion that such an all-encompassing goal is unattainable, and a more promising approach is to develop automated text analysis methods for limited subject domains and groups of users (e.g., Sparck Jones, 1991). Indeed, research in this direction has already begun. For example, Damerau (1993) developed computer-generated domain-oriented vocabularies with the aid of subject headings that were assigned intellectually to a text, and Ingwersen and Wormell (1988) suggested that different types of information needs may require dif- ferent retrieval techniques (e.g., a search for a known item might be best searched with Boolean logic which is based on exact match rather than a partial match technique). The proposed terminological knowledge structure illustrates that methods developed for automated text processing are relevant to the creation of a knowledge base that sup- ports searching. It reinforces the somewhat neglected connection between indexing and searching. The idea that searching methods be considered when procedures for automated text processing are developed is not new. AID, CITE (Doszkocs, 1978, 1983), ZOOM on ESA/IRS (Martin, 1982), and OKAPI (Walker & de Vere, 1990) are among the attempts which use statistical text analysis techniques to suggest additional terms for query expan- sion. While some of these methods are limited to a subject domain, or even to a user group, they are all global on the terminological level; they do not consider terminological attributes. The importance of such attributes in determining search keys has been demon- strated empirically (Fidel, 1991a). Moreover, the description of the terminological knowl- edge structure shows that data generated from statistical or other automated analyses of text can increase retrieval effectiveness. An example of data about term co-occurrence can illustrate this point. Co-occurrence data are used in automated text processing to indicate some similar- ity among terms. Thus, terms that co-occur with a search key can be used to expand a query. Recent research has demonstrated that the use of co-occurrence data is not effec- tive in improving retrieval performance when used in searching to expand a query auto- matically, because terms that highly co-occur tend to also occur frequently in the database (Peat & Willett, 1991). However, research in interactive, semiautomatic query expansion demonstrated that the use of co-occurrence data to rank terms for query expansion does in fact result in improved retrieval performance (Efthimiadis, 1992). It is possible that co-occurrence data could be used more effectively when termino- logical attributes of search keys are taken into consideration. If a search key has many synonyms, it is best to expand the query with terms that co-occur least frequently, but if Terminological knowledge structure 25 a search key itself occurs too frequently, it is best to substitute it with terms that co-occur most frequently but do not occur frequently in the database. Research in automated text processing that focuses on disambiguation, identification of semantic relatedness, and con- text dependency could provide a significant contribution to the creation of IESs. A close collaboration with terminological research would encourage automated text processing to go beyond its attempts to represent the text stored in a database and to concentrate on extracting terminological attributes from that text to help users’ searching. 5.3 Thesauri and indexes With IESs in place, the role of thesauri and indexing would change substantially. Both the process of indexing and the construction of controlled vocabularies would be limited to intellectual processes. The specific functions of thesauri and indexing would depend on the capabilities of the IES. For example, if an IES refers users to descriptors when a request includes terms that occur too frequently, indexers could be required to consult a list of terms that occur too frequently whenever they index a document to guarantee that they do not forget to assign the descriptors when relevant. Similarly, if the IES includes a seman- tic scale for each hedge term, a thesaurus may be developed that does not include the associative relationships (RT). Indexing would not be eliminated because some of its current functions cannot be automated. It is useful to examine these functions more closely. First, indexing enhances retrieval effectiveness when it resolves terminological difficulties caused by vague, ambig- uous, common, or synonymous terms and nonlexical expressions. The proposed IES would also resolve such difficulties. However, one of the options it would suggest in such situa- tions would be the use of descriptors. In fact, using descriptors has proven to be an effec- tive way to deal with terminological difficulties. Descriptors, of course, are assigned in indexing. But with IESs in place, indexing could be limited to assigning descriptors only for concepts whose terms represent terminological difficulties. Second, human indexers assign weights to concepts, albeit in a subjective manner. When a term is assigned in indexing, whether it is a descriptor or a natural language term, it reflects the indexer’s perception that the text is about the subject represented by that index term or that a user who is interested in this subject might want to retrieve the text. In addi- tion, intellectual indexing uses some form of weighting when terms are assigned as major or minor terms. In automatic indexing a weight is assigned to each term in the text based on the statistical properties of the term. Term weighting is an active area of research in information retrieval. Third, indexing provides explicit representation of information that is implicitly embedded in a text. That is, human indexers can infer perceived “aboutness” or potential relevance without textual clues. Given the present state of research in artificial intelligence (Sparck Jones, 1991), it seems that this important function would have to be performed intellectually. Thus, indexing would be limited to making explicit concepts that are implicit and to assigning descriptors for problematic concepts. Vocabulary control would be exercised only for problematic search keys, and data- base thesauri would be limited to those keys which require the use of descriptors; all other search keys would be indexed and searched with natural language. In fact, with highly developed IESs in place, a database thesaurus would be derived from the terminological knowledge structure, tailored to the specific requirements of the database and its users. In summary, a well-developed IES that advises users on the selection of search keys would not only help searchers, it would change the nature of indexing. It would transfer indexing, and possibly vocabulary control, from a primarily a priori process to a process that is determined by specific information needs and other situational factors. While intel- lectual indexing would be performed to resolve textual and terminological difficulties, indexing of other concepts and terms would be situation dependent and would be per- formed according to the requirements of each information request. Acknowledgemen&-Elaine Svenonius played an active role in the creation of this article; we greatly appreciate her help and insights. We also thank Susanne Humphrey, Philip Smith, and Dagobert Soergel for their useful comments on an earlier version. 26 R. FIDEL and E.N. EFTHIMIADIS REFERENCES Ahlswede, T., Anderson, J., Evens, M., Li, SM., Neises, J., Pin-Ngern, S., & Markowitz, J. (1988). Automatic construction of a phrasal thesaurus for an information retrieval system from a machine readable dictionary. RIAO 88 Program. Conference with presentation of prototypes and operational demonstrations: user- oriented content-based text and image handling (vol. 1, pp. 597-608). March 21-24, 1988, Cambridge, MA. Paris: C.I.D. Anderson, J.D., & Rowley, EA. (1992). Building end-user thesauri from full text. In B.H. Kwasnik & R. Fidel (Eds.), Advances in classification research: proceedings of the 2nd ASZS SZGCR Classification Research Workshop (pp. l-10). Medford, NJ: Learned Information. Brajnik, G., Guida, G., & Tasso, C. (1986). An expert interface for effective man-machine interaction. In L. Bolt & M. Jarke (Eds.), Cooperative interfaces to information systems (pp. 259-308). Berlin: Springer-Verlag. Brajnik, G., Guida, G., & Tasso, C. (1988). IR-NLI II: Applying man-machine interaction and Artificial Intel- ligence concepts to Information Retrieval. In Yves Chiaramella (Ed.), ACM-SZGZR, Ilth International Con- ference on Research and Development in Information Retrieval (pp. 387-399). Grenoble, France: ACM Press. Brenner, E.H., Lucey, J.H., Martinez, C.L., & Melekaa, A. (1984). American Petroleum Institute’s machine- aided indexing and searching project. Science and Technology Libraries, 5(l), 49-62. Damerau, F. (1993). Generating and evaluating domain-oriented multi-word terms from texts. Information Pro- cessing & Management, 29(4), 433-447. Doszkocs, T.E. (1978). AID-an associative interactive dictionary for on-line searching. Online Review, 2, 163-173. Doszkocs, T.E. (1983). CITE NLM: natural-language searching in an online catalog. Information Technology and Libraries, 2(4), 364-380. Efthimiadis, E.N. (1990). Online searching aids: a review of front-ends, gateways and other interfaces. Journal of Documentation, 46(3), 218-262. Efthimiadis, E.N. (1991). Approaches to search formulation and query expansion in information systems: DRS, DBMS, ES. (Report No. RDD/G/lOZ). Boston Spa: British Library. Efthimiadis, E.N. (1992). Interactive query expansion and relevance feedback for document retrieval systems. Doctoral dissertation, City University, London. Efthimiadis, E.N., & Robertson, S.E. (1989). Feedback and interaction in information retrieval. In C. Oppen- heim (Ed.), Perspectives in information management (pp. 257-272). London: Butterworths. Fidel, R. (1991a). Searcher’s selection of search keys: I. The selection routine. Journal of the American Society for Information Science, 42(7), 490-500. Fidel, R. (1991b). Searcher’s selection of search keys: Il. Controlled vocabulary or free-text searching. Journal of the American Society for Information Science, 42(7), 501-514. Frei, H.P., & Jauslin, J.F. (1983). Graphical presentation of information and services: a user-oriented interface. Information Technology: Research and Development, 2, 23-42. Fugmann, R. (1982a). The complexity of natural and indexing languages. International Classification, 9(3), 140-144. Fugmann, R. (1982b). Natural versus indexing languages in chemical documentation. Angewandte Chemie Znter- national Edition, 21(S), 608-616. Guentzer, U., Juettner, G., Seegmuueller, G., & Sarre, F. (1989). Automatic thesaurus construction by machine learning from retrieval sessions. Information Processing & Management, 25(3), 265-273. Humphrey, S.M. (1989). MedIndEx system: medical indexing expert system. Information Processing and Mun- agement, 25(l), 73-88. Humphrey, S.M., & Miller, N.E. (1987). Knowledge-based indexing of the medical literature: the Indexing Aid Project. Journal of the American Society for Information Science, 38(3), 184-l%. Humphreys, B.L., & Lindberg, D.A.B. (1989). Building the Unified Medical Language System. In Proceedings of the Thirteenth Annual Symposium on Computer Applications in Medical Cure (pp. 475-480). Washington, DC: IEEE Computing Society Press. Ingwersen, P., & Wormell, I. (1988). Modern indexing and retrieval techniques matching different types of infor- mation needs. In S. Koskiala & R. Launo (Eds.), Information, knowledge, evaluation: proceedings of the forty-fourth FZD Congress (pp. 79-90). Amsterdam: North-Holland. Knapp, S.D. (1988). Free-text searching of online databases. The Reference Librarian, 5(6), 143-153. Knapp, S.D. (1992). The contemporary thesaurus of social science terms and synonyms: a guide for natural lun- guage computer searching. Phoenix, AZ: Oryx Press. Krovetz, R. (1991). Viewing the dictionary as a classification system. In S.M. Humphrey & B.H. Kwasnik (Eds.), Advances in classificution research: Proceedings of the 1st ASZS SZGCR Classification Research Workshop (pp. 87-93). Medford, NJ: Learned Information. Krovetz, R., & Croft, W.B. (1989). Word sensing disambiguation using machine-readable dictionaries. ln N.J. Belkin & C.J. van Rijsbergen (Eds.), SZGZR ‘89: Proceedings of the Twelfth Annual International ACM- SZGZR Conference on Research and Development in Information Retrieval (pp. 127-136). New York: ACM Press. Liddy, E.D., Hert, C.A., & Doty, P. (1991). Roget’s International Thesaurus: conceptual issues and potential applications. In SM. Humphrey & B.H. Kwasnik (Eds.), Advances in classtfication research: Proceedings of the 1st ASZS SZGCR Classification Research Workshop (pp. 95-100). Medford, NJ: Learned Information. Makey, K., Atherton, P., & Newton, C. (1980). An analysis of controlled vocabulary and free text search state- ments in online searches. Online Review, 4(3), 225-236. Martin, W.A. (1982). Helping the less experienced user. In 6th International Online Meeting (pp. 67-76). Oxford: Learned Information (Europe) Ltd. Martinez, C., Lucey, J., & Linder, E. (1987). An expert system for machine-aided indexing. Journal of Chemi- cal Information and Computer Sciences, 27(4), 158-162. Monarch, I., & Carbonell, J. (1987). CoalSORT: a knowledge-based interface. IEEE Expert, 2, 39-53. Nutter, J.T., Fox, E.A., &Evens, M.W. (1990). Building a lexicon from machine-readable dictionaries for improved information retrieval. Literary and Linguistic Computing, 5(2), 129-137. Terminological knowledge structure 27 Peat, H.J., & Willett, P. (1991). The limitations of term co-occurrence data for query expansion in document retrieval systems. Journal of the American Society for Information Science, 42, 378-383. Piternick, A.B. (1984). Searching vocabularies: a developing category of online search tools. Online Review, 8(5), 441-449. Pollitt, AS. (1987). CANSEARCH: an expert systems approach to document retrieval. Information Process- ing and Management, 23(2), 119-138. Pollitt, AS. (1988). A common query interface using MenUSE-A menu-based user interface search engine. In Proceedings of the 12th International Online Meeting (vol. 2, pp. 445457). Oxford: Learned Information. Shoval, P. (1981). Expert/consultation system for a retrieval data-base with semantic network of concepts. SIGJR Forum, 16, 145-149. Shoval, P. (1985). Principles, procedures and rules in an expert system for information retrieval. Information Processing and Management, 21(6), 475-487. Shoval, P. (1986). Comparison of decision support strategies in Expert Consultation Systems. International Jour- nal of Man-Machine Studies, 24, 125-139. Shute, S.J., & Smith, P.J. (1992). Knowledge-based search tactics. Information Processing & Management, 29(l), 29-45. Sievert, M., & Boyce, B.R. (1983). Hedge trimming and the resurrection of the controlled vocabulary in online searching. Online Review, 7(6), 489-494. Smith, P.J., Shute, S.J., & Galdes, D. (1989a). In search of knowledge based search tactics. In N.J. Belkin & C.J. van Rijsbergen (Eds.), SIGIR ‘89: Proceedings of the Twelfth Annual International ACM-SIGIR Con- ference on Research and Development in Information Retrieval (pp. 3-10). New York: ACM Press. Smith, P.J., Shute, S.J., Galdes, D., & Chignel, M.H. (1989b). Knowledge-based search tactics for an intelligent intermediary system. ACM Transactions on Information Systems, 17(3), 246-210. Sparck Jones, K. (1991). Notes and references on early automatic classification work. SIGIR Forum, 25(l), 10-17. Svenonius, E. (1983). Use of classification in online retrieval. Library Resources and Technical Services, 27(l), 76-80. Svenonius, E. (1986). Unanswered questions in the design of controlled vocabularies. Journal of the American Society for Information Science, 37(5), 33 l-340. Thompson, R.H., & Croft, W.B. (1989). Support for browsing in an intelligent text retrieval system. International Journal of Man-Machine Studies, 30, 639-668. Veronis, J., Ide, N.M., & Harie, S. (1990). Automatic building of large neural networks for natural language disambiguation. Tenth International Workshop. Expert Systems and Their Applications. Specialized Con- ference: Natural Language Processing and Its Applications (pp. 105-l 17). May 28-June 1, 1990, Avignon, France. Nanterre, France: EC2. Vickery, A. (1988). The experience of building expert search systems. In David Raitt (Ed.), Proceedings of the 12th International Online Meeting (vol. 1, pp. 301-313). Oxford: Learned Information. Vickery, A., Brooks, H.M., Robinson, B., & Vickery, B.C. (1986). Expert system for referral (final report). London: University of London, Central Information Service. Vickery, A., Brooks, H.M., Robinson, B., & Vickery, B.C. (1987). A reference and referral system using expert system techniques. Journal of Documentation, 4, 198-203. Vleduts-Stokolov, N. (1987). Concept recognition in an automatic text-processing system for life sciences. Journal of the American Society for Information Science, 38(4), 269-287. Walker, S., & de Vere, R. (1990). Improving subject retrieval in online catalogues: 2. Relevance feedback and query expansion. British Library Research Paper 72. London: British Library. 
work_b6lhy6c635gqbgvyl5wplgrxna	----	A profile of information systems research published in expert systems with applications from 1995 to 2008 Expert Systems with Applications 38 (2011) 3999–4005 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A profile of information systems research published in expert systems with applications from 1995 to 2008 Wen-Lung Shiau ⇑ Ming Chuan University, Department of Information Management, Taiwan, ROC a r t i c l e i n f o Keywords: IS research Literature analysis Expert system Research profile Citation analysis 0957-4174/$ - see front matter � 2010 Elsevier Ltd. A doi:10.1016/j.eswa.2010.09.061 ⇑ Address: 12F.-3, No. 3, Lane 24, Wunhua St., Pingjh Taiwan, ROC. Tel.: +886 3 4948766; fax: +886 3 4663 E-mail address: mac@mail.mcu.edu.tw a b s t r a c t Expert systems with applications (ESWA) has been regarded as one of the highly qualified journals in the information system. This paper profiles research published in ESWA from 1995 to 2008. Based on the multidimensional analysis, we identified the most productive author and universities, research paper numbers per geographic region, and the most employed issues and methodologies used by the most highly published authors. Our results indicate that (1) ESWA is clearly an internationalized journal, (2) the most employed methodologies are fuzzy ESs and knowledge-based systems, and (3) the leading highly published authors always have diverse methodologies and applications. Furthermore, the implica- tions for researchers, journal editors, universities, and research institution are presented. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction Expert system is a kind of information system, a subfield of arti- ficial intelligence, and a software that is used to reproduce the per- formance of one or more human experts. The simulation of performance of the expert is to help human workers solve real world problems by expertise, a specific domain of knowledge. There are diverse problems which need to be solved in the real world. Thus, the use of expert system becomes prolific in many fields (Liao, 2005). ESWA is regarded as one of the distinguished international journals which focus on expert and intelligent sys- tems applied to real world problems globally. As a journal of expert and intelligent systems, it addresses many problems, methods, and performance. Therefore, it is useful for IS researchers to examine the major research issues in expert and intelligent systems and their trends in order to continuously progress the research. This paper profiles the types of research that have been pub- lished in ESWA from 1995 to 2008. The overview of research pub- lished in the journal provides a broader understanding of IS for authors, reviewers, and journal editors, universities, and research institutions (Avison, Dwivedi, Fitzgerald & Powell, 2008). It is worthwhile to classify, categorize, and profile this paper into various dimensions, such as research issue (Alavi & Carlson, 1992; Claver, Gonzalez, & Llopis, 2000; Culnan & Swanson, 1986; Farhoomand & Drury, 1999; Vessey, Ramesh, & Glass, 2002), research methodology (Chen & Hirschheim, 2004; Grover, Lee, & ll rights reserved. en City, Taoyuan County 324, 107. Durand, 1993; Ives, Hamilton, & Davis, 1980; Palvia, 2004; Palvia, Mao, Salam & Soliman, 2003), productive authors and university (Athey & Plotnicki, 2000; Im, Kim, & Kim, 1998; Nath & Jackson, 1991), because the results could provide some frequently called for guidelines and suggestions on how to publish in highly rated journals. Thus, this paper analyzes publication trends in ESWA to realize those objectives as follows: 1. To identify the most productive authors. 2. To identify the universities associated with the most research publications. 3. To identify the research impact of the most productive authors and geographic location based on the AIS classification. 4. To determine the issues and methodologies most investigated and analyze their trends. 5. To discuss the issues and methodologies of the most productive authors. To achieve these objectives, a systematic and comprehensive review is needed. This review is constructed as follows. The next section describes the research methodology used in this paper. The third section presents the results of data analysis. The final sec- tion is the conclusion and implications of this research. 2. Research method Expert system with applications published their first issue in 1990 and had full online information from this journal and was opened to general users in 1995. Thus this paper surveys the devel- opment of ESWA from 1995 to 2008. To focus on research articles, http://dx.doi.org/10.1016/j.eswa.2010.09.061 mailto:mac@mail.mcu.edu.tw http://dx.doi.org/10.1016/j.eswa.2010.09.061 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Table 1 Top 11 productive authors. Name Affiliation AIS region Count Ingoo Han Korea Advanced Institute of Science and Technology 3 26 Tzung-Pei Hong National University of Kaohsiung 3 13 Elif Derya Ubeyli TOBB Ekonomi ve Teknoloji Üniversitesi 2 13 Dirk Van den Poel Ghent University 2 13 Salih Gune Selcuk University 2 12 So Young Sohn Yonsei University 3 11 Shyi-Ming Chen National Taiwan University of Science and Technology 3 11 Kemal Polat Selcuk University 2 11 Yueh-Min Huang National Cheng Kung University 3 10 Sang Chan Park Korea Advanced Institute of Science and Technology 3 10 Inan Guler Gazi University 2 10 4000 W.-L. Shiau / Expert Systems with Applications 38 (2011) 3999–4005 we follow Gallivan and Benbunan-Fich’s suggestions to exclude editorials, guest editorials and book reviews from the analysis (Gallivan & Benbunan-Fich, 2007). Research articles between 1995 and 2008 totaled 1836 articles. For a detailed analysis, each article was carefully examined to capture the relevant data. Relevant data include title, author, sub- ject terms, abstract, affiliation, publication, year, and citations. Various items were counted and recorded for each article includ- ing the productivity of authors, geographic regions, author’s back- ground, research issues, and the impact of the research by the most productive authors. We used self citation within ESWA, ISI citation counts, and Google’s scholar citation counts to assess the number of contributions per author. Institutional productivity was assessed by using each publication counted as one for all authors regardless of the number of co-authors from the same university (Avison et al., 2008; Yogesh & Jasna, 2008). The geo- graphic location variables were counted and categories were based on the AIS guidelines. The profile of methodologies em- ployed in this research is from a review of expert system method- ologies and applications (Liao, 2005). Liao (2005) collects 166 articles from 78 journals from the year 1995–2004, surveys and classifies ES methodologies using 11 categories: rule-based sys- tems, knowledge-based systems, neural network, fuzzy ESs, ob- ject-oriented methodology, case-based reasoning (CBR), system architecture development, intelligent agent (IA) systems, model- ing, ontology, and database methodology. We collected relevant subjects of 11 categories. Then those subjects were used to clas- sify the title, keywords, and abstract of the articles. The issues and methodologies of the most productive authors were also counted. It is essential to emphasize that the findings of this study including the most productive authors and universities should be regarded as ‘indicative and not an authoritative statement’(Palvia, Pinjani, & Sibley, 2007). This is because such profiling analysis was applied only within ESWA. Some important researchers and uni- versities might not have published in ESWA. They may have exper- tise and skills to publish in other important outlets. Therefore, similar to previous profiling studies (Claver et al., 2000; Palvia et al., 2007), it is best to interpret these results cautiously. Table 2 Top 10 productive universities. Rank School Counts 1 Korea Advanced Institute of Science and Technology 90 2 National Chiao-Tung University 59 3 National Cheng-Kung University 55 4 Yonsei University 32 5 National Yunlin University of Science and Technology 31 6 Hong Kong Polytechnic University 29 7 Yuan-Ze University 28 8 National Taiwan University of Science and Technology 27 9 National Central University 24 10 Firat University 23 10 National Kaohsiung First University of Science and Technology 23 10 Nanyang Technological University 23 3. Results 3.1. Productive authors For assessing research productivity, an analysis is made of the authors who have published the most during the period 1995– 2008 in the expert system with applications. All publications naming the researcher are counted equally (Palvia, Pinjani, et al., 2007; Yogesh & Jasna, 2008). For example, an article with two co-authors would provide one count for each. This approach results in the combined count of all authors being greater than the total number of articles. The findings of this study only include authors who have published 10 or more articles during the period 1995– 2008 for reporting purposes. Table 1 lists the 11 most productive authors, sorted by the number of publications, along with the asso- ciation for information systems (AIS) region and current affiliation. The top 11 authors with 10 or more publications are: Ingoo Han, Tzung-Pei Hong, Elif Derya Ubeyli, Dirk Van den Poel, Salih Gune, So Young Sohn, Shyi-Ming Chen, Kemal Polat, Yueh-Min Huang, Sang Chan Park, and Inan Guler. The geographical regions sug- gested by AIS are region 1 (America), region 2 (Europe, the Middle East and African), and region 3 (the Asia and Pacific). It is quite obvious that the most productive authors came from region 2 and region 3. This is because ESWA has attracted non-American authors and has become an internationalized IS journal. 3.2. Leading research university To assess the contribution of institutions and or universities, an analysis is made of institutions and or universities which have published the most during the period 1995–2008. As stated previ- ously, the results should be regarded as indicative and not as an authority’s statement of university research. It was misleading to have several researchers from the same university co-authoring an article (Palvia, Pinjani, et al., 2007). Therefore a university was counted only once when it had two or more authors on a single publication. Table 2 lists the top 10 universities having 23 or more articles published in ESWA during the period 1995–2008. A total of 12 are listed as the top 10 universities because there are three uni- versities having the same counts. The top place in the lists is Korea Advanced Institute of Science and Technology, with 90 contribu- tions. The next 11 universities are National Chiao-Tung University, National Cheng-Kung University, Yonsei University, National Yun- lin University of Science and Technology, Hong Kong Polytechnic University, Yuan-Ze University, National Taiwan University of Sci- ence and Technology, National Central University, Firat University, National Kaohsiung First University of Science and Technology, and Nanyang Technological University. It may show that Asian univer- sities and researchers are more likely to prefer this international- ized journal focusing on expert and knowledge domain. 3.3. Research numbers by AIS region The association of information system (AIS) suggests that the geographical regions are region 1 (America), region 2 (Europe, W.-L. Shiau / Expert Systems with Applications 38 (2011) 3999–4005 4001 the Middle East and African), and region 3 (the Asia and Pacific). An analysis is made of the regions. Fig. 1 lists the total number of con- tributions by AIS regions during the period 1995–2008. Before the year 2000, the total number of articles of region 1 (America) reached its highest in 1998. After the year 2001 region 2 (Europe, the Middle East and African) and region 3 (the Asia and Pacific) grew consistently. After the year 2006 all regions grew quickly be- cause the journal received more and more articles and also has more capacity to publish more articles. 3.4. Research methodologies The profile of methodologies employed in this research follows Liao’s (2005) 11 categories: rule-based systems, knowledge-based systems, neural network, fuzzy ESs, object-oriented methodology, case-based reasoning (CBR), system architecture development, intelligent agent (IA) systems, modeling, ontology, and database methodology (Liao, 2005). We collected the keywords from the Liao’s relevant 11 categories and listed them in Appendix A. We used the keywords in three stages to attribute a total of 1836 articles to 11 categories. The first stage was to compare the collected keywords to each title of research article. There are 757 identified articles and 1079 articles are unidentified. The second stage was to analyze the 1079 articles by comparing the collected keywords to each keyword of each research article. There are 392 identified leaving 687 articles unidentified. The third stage was to analyze the 687 articles by comparing the collected keywords to each abstract of each research article. There are 354 identified articles and 333 articles are unidentified. After three stage analysis Fig. 1. Research numb Fig. 2. The number o the totals are 1503 articles identified and 333 articles unidentified. The identified rate is 81.86%. Fig. 2 lists the number of 11 catego- ries published during the period 1995–2008. The most popular methodology is fuzzy ESs, followed by knowledge-based systems and neural network. From 1990, fuzzy became more and more an important issue worldwide because of the uncertainty in problem solving. It is not surprising that the fuzzy systems were of most concern to the expert and intelligence research community. 3.4.1. Trends in research methodologies To see the trends in research methodologies, we display 11 cat- egories year by year during the period 1995–2008. Fig. 3 lists the trends in research methodologies. The results show that fuzzy ESs started high in the mid 1990s, went down in 2001 and reached the top in 2008. The neural network started to grow quickly from 2003 and reached second in 2008. The knowledge-based systems are constantly in the high rate of publications and rise up to rank third in 2008. The rank fourth is rule-based systems. Other meth- odologies always show a low rate in the trend analysis. 3.5. Leading research profiles The profile of leading research provides important information for researchers to realize what leading authors researched. Junior researchers learn much from publications of leading authors and find some niches to continue their careers. Two analyses were made of issues and methodologies researched by leading authors. Table 3 lists the researched issues by top 11 authors. Table 4 pre- sents the preferred research methodologies by the top 11 authors. ers by AIS region. f 11 categories. Fig. 3. Trends in research methodologies. Table 3 Researched issue by top 11 authors. Name Affiliation Issue researched Ingoo Han Korea Advanced Institute of Science and Technology Action mechanism, virtual community recommender, information noise CRM, bankruptcy prediction, stock price index credit risk, mobile, recommendation, IDS behavior of physicians, activity-based costing prediction of interest rate, maintaining system bond rating, exchange-rate forecasting impact of measurement scale Tzung-Pei Hong National University of Kaohsiung Efficient sanitization, FP tree, bit-based feature selection, machine learning, data mining, parallelized indexing, classification, knowledge-integration, medical Elif Derya Ubeyli TOBB Ekonomi ve Teknoloji Üniversitesi Medical Dirk Van den Poel Ghent University CRM, marketing Salih Gune Selcuk University Medical, medical DSS So Young Sohn Yonsei University Mobile service, CRM, finance, dynamic scoring government fund, bipartite scorecard, selecting air force pilot trainee, commercialization, credit operations Shyi-Ming Chen National Taiwan University of Science and Technology Students’ evaluation, fuzzy system, adapting learning systems, fuzzy risk, temperature prediction, document retrieval, text categorization estimating null values Kemal Polat Selcuk University Medical, medical DSS Yueh-Min Huang National Cheng Kung University Net work, E-learning, adaptive learning, multiprocessor system, proportionate flow shop mining interesting patterns, intelligent, human- expert forum system, job-scheduling Sang Chan Park Korea Advanced Institute of Science and Technology Minimizing information gap, visualization of patent analysis, personalization CRM, electronic commerce, process control system evaluation, integrated yield management in semiconductor manufacturing, database marketing Inan Guler Gazi University Image restoration, medical 4002 W.-L. Shiau / Expert Systems with Applications 38 (2011) 3999–4005 Table 4 Preferred research methodology by top 11 authors. Name Affiliation Employed research methodology Ingoo Han Korea Advanced Institute of Scienceand Technology Rule-based Knowledge-based Neural network Fuzzy ESs Object-oriented Cased-based System architecture Intelligent agent system Modeling Ontology Tzung-Pei Hong National University of Kaohsiung Rule-based Knowledge-based Neural network Fuzzy ESs Object-oriented Cased-based Intelligent agent system Ontology Elif Derya Ubeyli TOBB Ekonomi ve Teknoloji Üniversitesi Rule-based Knowledge-based Neural network Fuzzy ESs Cased-based System architecture Intelligent agent system Database Modeling Dirk Van den Poel Ghent University Knowledge-based Neural network Fuzzy ESs Ontology Salih Gune Selcuk University Rule-based Knowledge-based Neural network Fuzzy ESs Cased-based System architecture Database Modeling So Young Sohn Yonsei University Knowledge-based Neural network Fuzzy ESs Object-oriented Shyi-Ming Chen National Taiwan University of Science and Technology Rule-based Fuzzy ESs System architecture Intelligent agent system Kemal Polat Selcuk University Rule-based Knowledge-based Neural network Fuzzy ES Cased-based System architecture Database Modeling Ontology Yueh-Min Huang National Cheng Kung University Rule-based Knowledge-based Neural network Fuzzy ESs Object-oriented Cased-based System architecture Modeling Intelligent agent system Ontology Table 4 (continued) Name Affiliation Employed research methodology Sang Chan Park Korea Advanced Institute of Scienceand Technology Rule-based Knowledge-based Neural network Fuzzy ESs Object-oriented Cased-based Intelligent agent system Modeling Ontology Inan Guler Gazi University Rule-based Knowledge-based Neural network Fuzzy ESs Cased-based System architecture Database Modeling Intelligent agent system W.-L. Shiau / Expert Systems with Applications 38 (2011) 3999–4005 4003 Most leading researchers have focused their attention on the industry or different methods. For example, professor Han focuses on financial field, professors Ubeyli and Gune focus on medical field, professor Poel focuses on marketing, and professors Hong and Chen use diverse methods to solve specific problems in a do- main. Even though top leading researchers seem to be dealing with different issues, they still focus on a few domains. When they pay more attention to a specific research stream, they are able to solve more core problems in a domain. 4. Conclusions and implications It is important for any field or any journal to do self-introspec- tion in order to realize what happened inside of a journal or a field. This paper profiles the researches that have been published in ESWA from 1995 to 2008. Some interesting results emerge by ana- lyzing the data during the period. First, ESWA is really an internationalized journal from the view- point of geography. There are no significant differences in the num- ber of publications between the three regions before the year 2001. After the year 2001 Asia and Pacific (region 3) published more arti- cles, followed by region 2 (the Europe, Middle East and African) and region 1 (America). Second, the results of the leading research universities show that Asian universities and researchers contrib- uted more to this journal, especially after the year 2001. Third, the more popular methodologies are fuzzy ESs, knowledge-based systems, and neural network. Finally, there are many researchers who develop mathematical models or specialize in analyzing prob- lems that are seen in the real world. Most productive authors deal with diverse issues. Similarly most productive authors employed diverse methodologies to solve problems because they confront different difficulties. There are several implications for researchers, journal editors, research institutions or universities. First, the analysis of this study shows that leading authors deal with different issues with diverse methods. But they still focus their interests and preferred method- ologies on certain research streams. Second, leading authors also develop sophisticated mathematical models or specialize in solv- ing problems that emerge in the real world. Upcoming researchers may apply the knowledge learned from these models to develop even better models. Third, the 11 categories of methodology in this paper are not complete. Other social science methodologies includ- ing expert system methodologies need to be taken into consider- ation, such as psychology and human behavior. Thus, prospects Appendix A (continued) Categories First keyword Second keyword Knowledge_base Medical Financ Planning 4004 W.-L. Shiau / Expert Systems with Applications 38 (2011) 3999–4005 from other social sciences methodologies are an important work in the near future. Finally, our recommendation to journal editors and research programs is to continue diverse research and to publish special issues in urgent problems, such as global recession and financial crisis, in order to meet the worldwide requirements of timely fashion. Appendix A Arranged keywords adapted from Liao (2005) Categories First keyword Second keyword Case_Based Case-base Case Base Manufac Process Knowledge management Ultrasonic Medical Appl Fault Diagn Knowledge Model e-Learning Database Geograph Traditional Chinese medicin Medical Expert Fuzzy_ESs Power load forecasting Scheduling Chemical Diagn Power Diagn Control sys Uncertain Reason Knowledge Integ Fault Diagn Power sys Fault detect Demand Wastewater treatment Data Select On-line Medical Diagn Job match Performance Index Secur Recognition Expert sys Intelligent_agent_systems Intelligent agent Tutor Sys Supply Chain System anal System desig Knowledge represent Adaptive Sys Air pollution Building Design Knowledge Internet Waste Production Decision DSS Knowledge KM Power Design Financial Tumor Business Agricult Steel composit Strateg Isokinetics Chemical Process Plant Control Concurrent Case Chip Power Trans Urban Robot Modeling Process Control Medical Software Estima Assembly Planning Project allocation Neural_networks Neural network Fault Diagn Optimal Power Decision Making Inference Diagnostic sys Machine learning Power load forecasting Process control Knowledge Learn Robot Parameter Waste Treat Biomedical Crude oil distillation Neural network Objective_oriented Object-orient Industr Diagn Knowledge Represent Knowledge Learn Knowledge Engineer Manufacturing information network Syntactic Appendix A (continued) Categories First keyword Second keyword Ontology Medical deci Knowledge Reuse Prevent Landscape assessment Knowledge Acquisition Knowledge Model Rule_based Transit Plann Product Psych Teach Advis Syst Devel Knowledge Veri Knowledge Vali Knowledge Base Scheduli Resource util Probab Diagn Sensor Tutor Knowledge represent Rule base Rule-base System_architecture System architect Material Evalu Material Select Computer aid Implement Corpora Decis Military Ferryboat W.-L. Shiau / Expert Systems with Applications 38 (2011) 3999–4005 4005 References Athey, M., & Plotnicki, J. (1992). An evaluation of research productivity in academic. Communications of the AIS, 3. Avison, D. E., Dwivedi, Y. K., Fitzgerald, G., & Powell, P. (2008). The beginnings of a new Era: Time to reflect on 17 years of the ISJ. Information Systems Journal, 15, 5–21. Chen, W., & Hirschheim, R. (2004). A paradigmatic and methodological examination of information systems research from 1991 to 2001. Information Systems Journal, 14, 197–235. Claver, E., Gonzalez, R., & Llopis, J. (2000). An analysis of research in information systems (1981–1997). Information & Management, 37, 181–195. Culnan, M. J., & Swanson, E. B. (1986). Research in management information systems, 1980–1984: Points of work and reference. MIS Quarterly, 10, 289–302. Farhoomand, A., & Drury, D. H. (1999). A historiographical examination of information systems. Communications of the AIS, 1, 1–27. Gallivan, M. J., & Benbunan-Fich, R. (2007). Analysing IS research productivity: An inclusive approach to global IS scholarship. European Journal of Information Systems, 16, 36–53. Grover, V., Lee, C. C., & Durand, D. (1993). Analyzing methodological rigor of MIS survey research from 1980–1989. Information & Management, 24, 305–317. Im, K. S., Kim, K. Y., & Kim, J. S. (1998). An assessment of individual and institutional research productivity in MIS. Decision Line, 8–11. Ives, B., Hamilton, S., & Davis, G. B. (1980). A framework for research in computer- based management information systems. Management Science, 26, 910–934. Liao, S.-H. (2005). Expert system methodologies and applications – A decade review from 1995 to 2004. Expert Systems with Applications, 28(1), 93–103. Nath, R., & Jackson, W. M. (1991). Productivity of management information systems researchers: Does Lotka’s Law apply? Information Processing & Management, 27, 203–209. Palvia, P. (2004). Research methodologies in MIS: An update. Communications of the Association for Information Systems, 14. Palvia, P., Mao, E., Salam, A. F., & Soliman, K. S. (2003). Management information systems research: what’s there in a methodology. Communications of AIS, 11. Palvia, P., Pinjani, P., & Sibley, E. H. (2007). A profile of information systems research. Information & Management, 44(1), 1–11. Vessey, I., Ramesh, V., & Glass, R. L. (2002). Research in information systems: An empirical study of diversity in the discipline and its journals. Journal of Management Information Systems, 19, 129–174. Yogesh, K. D., & Jasna, K. (2008). Profile of IS research. European Journal of Information Systems, 17, 678–693. A profile of information systems research published in expert systems with applications from 1995 to 2008 Introduction Research method Results Productive authors Leading research university Research numbers by AIS region Research methodologies 3.4.1 Leading research profiles Conclusions and implications Appendix A References 
work_b6m7qqhvezfj5mvhumiyb5ihue	----	Off-Line Arabic (Indian) Numbers Recognition Using Expert System (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 397 | P a g e www.ijacsa.thesai.org Off-Line Arabic (Indian) Numbers Recognition Using Expert System Fahad Layth Malallah Faculty of Computer Science, Cihan University Sulaimaniya, Iraq Mostafah Ghanem Saeed Faculty of Computer Science, Cihan University Sulaimaniya, Iraq Maysoon M. Aziz Department of Mathematics, College of Computer Sciences and Mathematics, University of Mosul Mosul, Iraq Olasimbo Ayodeji Arigbabu Department of Computer and Communication Systems Engineering, Faulty of Engineering, Universiti Putra Malaysia, Serdang Malaysia Sharifah Mumtazah Syed Ahmad Department of Computer and Communication Systems Engineering, Faulty of Engineering, Universiti Putra Malaysia, Serdang Malaysia Abstract—This paper proposes an effective approach to automatic recognition of printed Arabic numerals which are extracted from digital images. First, the input image is normalized and pre-processed to an acceptable form. From the preprocessed image, components of the words are segmented into individual objects representing different numbers. Second, the numerical recognition is performed using an expert system based on a set of if-else rules, where each set of rules represents the categorization of each number. Finally, rigorous experiments are carried out on 226 random Arabic numerals selected from 40 images of Iraqi car plate numbers. The proposed method attained an accuracy of 97%. Keywords—Arabic numeral character recognition; Image Processing; Pattern Recognition; Feature Extraction; Object Segmentation; Expert System I. INTRODUCTION Automatic character recognition is becoming very important in many practical applications such as postcode identification and car plate number recognition. A traffic police officer may want to document the license plate numbers of approaching vehicles. Manually performing this task would obviously be very laborious, and incur significant amount of time. Conversely, an automated process that involves the application of a camera to capture the plate numbers and recognize them using a predictive model, would not only be beneficial in terms of computational time, but also ease the amount of human effort required for such task. Furthermore, these systems can be used for surveillance or monitoring of specific events using numbers. However, most research studies in this area are mainly concentrated on Latin character recognition. In this paper, an effective method for Arabic character recognition is presented, which is also applicable to Kurdish and Persian languages [1]. Arabic is the first language in all Arabic countries. In total, the estimated population of these countries is 280 million and some other countries that consider Arabic as a second language have an estimated population of 250 million. Moreover, Arabic language is ranked fifth out of the most commonly used languages in the world. Due to the wide usage of Arabic language, it is highly desirable to develop an effective and automatic Arabic character recognition system. Therefore, in this paper, a technique to recognize Arabic Numerals by using handcrafted features and expert system for decision making is proposed. This method involves extracting geometric features for each object to be further classified using expert system, which is discussed in-depth in the subsequent sections. It is worth noting that in this paper, Indian Numerals are basically denoted as Arabic, Kurdish and Persian numbers. The basic Indian numbers are 10, ranging from 0-to-9as 9,8,7,6,5,4,3,2,1,0. The organization of the rest of the paper is as follows. Section 2 describes the review of previous research works in this domain. Section 3 outlines the methods used for developing the proposed numeral character recognition system. Section 4 describes the experiment implementation. Section 5 provides the analysis of results and discussions on the main findings of the paper. Finally, this research is summarized in Section 6. II. LITERATURE REVIEW Many research studies have been conducted on automatic Arabic numeral recognition. Some studies are focused on hand- written recognition, while others concentrate on printed materials. A recognition system for offline handwritten Arabic numerals that exploited the properties of Hindi (Arabic) numerals as powerful set of features is proposed in [3], This method is mainly based on image processing operations and a decision making stage that uses if-else statements to determine the appropriate character output [3]. Also, a Latin number recognition system for number plate localization & segmentation is presented, here, the authors adopted skeletonization method for feature extraction and recognition of the characters is based on Support Vector Machine (SVM). It is claimed that the method is invariant to translation and illumination variation [4,5]. Olasimbo Ayodeji Arigbabu contributed to this work while he was a graduate student at Universiti Putra Malaysia. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 398 | P a g e www.ijacsa.thesai.org An efficient shift and scale invariant approach for offline machine-printed decimal digit recognition that computes the correlation factor between the reference and test image to perform recognition, is described in [6]. Also, in [7], a technique utilizes a number of statistical methods to perform machine print recognition. In addition, several approaches that are based on Neural Networks and Support Vector Machine (SVM) have been investigated for recognition of on-line and off-line handwritten Arabic and Hindi numerals [8-15]. Likewise, Hidden Markov Models have also been adopted for recognition of off-line handwritten numerals [16]. Furthermore, in [17] a genetic programming is used to perform the recognition of hand-written digits, however, it has lack in terms of recognition rate. In [18] translational motion estimation has been examined for the recognition of offline machine-print Hindi digits. III. METHODOLOGY A. Framework Design The overall steps involved in the proposed method are illustrated in Figure 1. Fig. 1. Flowchart of the proposed offline print-written number recognition system The process starts with an input image that contains object numbers. Initially, the image is converted to black and white. Then, image normalization (crop) is performed to localize the region of interest (ROI), which is mainly composed of the number section. In addition to that, image complement operation is applied to the normalized image to enable further processing on image pixels with value of 1. Before Normalization After Normalization Fig. 2. Car number plate image, before normalization and after normalization operation Further, preprocessing operations including removal of noise and image enhancement are performed. Afterward, image segmentation is applied using region labeling method [19], and finally, the proposed algorithm for number recognition is implemented to obtain the decision of each number’s identity. B. Preprocessing The normalized input image is converted to binary and the complement of the binary image is derived, as shown in Fig 3a. Then, median filter as a 3 x 3 kernel size for noise removal is used, as depicted in Fig 3b. Closing morphological operation is performed on the filtered image using a 7 x 7 structure element and opening operation is used to remove unwanted small pixels from the binary image. Figure 3 depicts the outputs of the four operations adopted in this research for preprocessing before proceeding to number recognition. C. Expert System An expert system is a computer system that emulates the decision-making ability of a human expert [22]. Expert systems are suitable tools for implementing structural pattern recognition techniques and it helps to solve difficult pattern recognition problems. More rules and human experience can be added easily using rule-based systems, especially in closed- system applications with precise inputs and logical outputs [23, 24]. Expert systems have a number of major system components and interface with individuals who interact with the system in various roles as shown in figure 4. In rule-based expert systems, there are two basic techniques; Forward chaining and Backward chaining inference. The domain knowledge is represented by a set of IF-THEN production rules and the data is represented by a set of facts about the current situation. The matching of the rule IF parts to the facts produces inference chains. The inference chain indicates how an expert system applies the rules to reach a conclusion. The inference engine must decide when the rules have to be fired [24]. An inference engine using forward chaining searches the inference rules until it finds one, where the IF clause is known to be true. Forward chaining is used in this paper because of the similarity to the methodology that depends on the data- driven reasoning. The reasoning starts from the known data and proceeds forward with that data. Each time, only the topmost rule is executed, and when fired, the rule adds a new fact to the database. Any rule can be executed only once and the match-fire cycle stops when no further rules can be fired [25, 26]. Image Segmentation Filtering Object by Proposed Algorithm Input Image Output Numbers Converting to Black & White Image Normalizing Image Preprocessing Image (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 399 | P a g e www.ijacsa.thesai.org (a) Black & White with Complement (b) After Median Filter (c) After Closing Image Operation (d) After Area Opening Operation Fig. 3. Preprocessing steps (Black & White, Median filter, closing filter, area opening operation) D. Feature Extraction and Recognition This section discusses the features that are useful for recognition of each numerical object, as well as how the features are obtained or extracted. Prior to that, it is essential to mention that the recognition operation is performed by processing each object once at a time. Therefore, object segmentation operation is considered very crucial in the proposed method. Details of the segmentation operation are elaborated in [27]. For instance, figure 5 shows some examples of segmented Arabic numbers for feature extraction, where each number will be processed separately. Fig. 4. Components of an expert system [24,25] Fig. 5. Some examples of the segmented Arabic object numbers. The feature extraction and recognition of each number are performed sequentially. In other words, the features of a particular number are extracted using the proposed algorithm, and then, the decision of the number’s identity is performed based on the set of rules in the expert system. The process examines every possible match of the facts provided in the inference engine to determine the expected identity of the number. For instance, a random number can be predicted by first checking whether the feature properties align with the facts about number 5 in the inference engine. If an alignment is not found, the system examines the possibility with the facts about number 9. The process is repeated till it reaches number 6, but if a match is found then the expected identity will be presented to the user. Number 5 Recognition: In order to assign an identity ‘5’ to an object, two conditions (facts) should be satisfied. Firstly, the Euler number should be zero. Euler number describes the relation between the number of contiguous parts and the number of holes on a shape. Let S denote the number of contiguous parts and N be the number of holes in a shape. Thus, the Euler number is determined as in (1): (1) For example, the Euler number for Shape (B) is -1, Shape (9) is 0, and shape (3) is 1. Secondly, the number of flips should be greater than or equal to 3 flips. Flips number is (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 400 | P a g e www.ijacsa.thesai.org computed by scanning the object from left to right at the mid- level of the image, as shown in figure 6. The pseudo code for number five is as follows: If [( Euler Number No. =0) && (Flip_number_mid_horizonatally ) >= 3] Then The Object is ‘5’; In figure 6, it can be seen that the arrow indicates three flips by scanning from left to right at the mid-level of the image. Number of flips is simply a count of the alternating transitions of pixel values from “1” to “0” or vice versa. Below is a pseudo code for extracting the number of flips. first_value=(1,mid_y); for x=min: x_max if ( first_value ~= Object_array(x,mid_y)) first_value= Object_array(x,mid_y); flip_num=flip_num+1; end end Fig. 6. Shows scanning Arabic object number five to count number of flips Number 9 Recognition: In case the conditions for number ‘5’ are not satisfied, the object will be examined with the conditions for number ‘9’. The first condition is that the Euler number should be zero. The second is that the aspect ratio (calculated by dividing the minor axes by the major axes as shown in figure 7) should be more than 0.6. The Aspect ratio formula is specified in (2): (2) Fig. 7. Shows scanning Arabic object number nine to count number of flips The third condition is the flips count resulting from object scanning from left to right direction at the lower part of the object, should be less than or equal to 2. As illustrated in figure 7, the arrow indicates the scanning direction to count the number of flips, which in this case is equals to 2. Pseudo code for number 9 recognition is: If [(Euler Number =0) && (Aspect ratio > 0.6 ) && (flips- lower-horizontal <=2) ] Then The Object is ‘9’; Number ‘8’ Recognition: Three conditions are used to examine whether the object is number ‘8’, when the conditions for number ‘5’ and ‘9’ are not satisfied. Firstly, the Euler Number should be equal to one. Secondly, the widths of the object at the upper, middle, and lower segments are checked. Basically, the width at the upper segment should be less than the width at the middle and lower segments. The final condition is that the middle width should be less than the width at lower segment of the object. The pseudo code is as follows: If [( Euler Number No.=0) && ( lower_dist > middle_dist > upper_dist)] Then The Object is ‘8’; Fig. 8. Shows three scanning positions of Arabic object number eight Number ‘7’ Recognition: Three conditions are also considered to determine whether the preprocessed object is number 7, when the conditions of ‘5’, ‘9’, and ‘8’ are not satisfied. The Euler Number of the object should be to one. The width of the object at the upper segment should be greater than the width of the middle and lower segments. The final condition is that, the width at the middle segment should be larger than the lower segment width. The pseudo code is as follows: If [(Euler Number =0) && (lower_dist < middle_dist < upper_dist ) ] Then The Object is ‘7’; upper_dist middle-dist lower-dist (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 401 | P a g e www.ijacsa.thesai.org Fig. 9. Shows three scanning positions of Arabic object number seven Number 3 Recognition: When the conditions for number ‘5’ ‘9’, ‘8’ and ‘7’ are not satisfied, the object will be examined with the following conditions to determine whether the number is ‘3’. Firstly, the Euler Number should be equal to 1. Then, the algorithm calculates the number of flips for two separated parts. Firstly, the part located in the top quarter of the number object, as highlighted with the upper arrow in figure 10. Here, the number of flips must be greater than or equal to 6. This kind of feature is quite discriminating in comparison to the features of other number since 3 is the only object whose number of flips is equal to 6 in the mentioned position. The third condition is achieved by calculating the flips in the lower quarter of the object and the result should be equal to 2. The pseudo code for the number three object is as follows: If [ (EulerNumber= 1) && (flip_top_quarter >=6) && (flip_bottom_quarter=2)] Then The Object is ‘3’; Fig. 10. Shows two scanning positions of Arabic object number three Number 2 Recognition: Now, conditions are described to determine the identity of the input image as 2, when the conditions for number ‘5’ ‘9’, ‘8’, ‘7’ and ‘3’ are not satisfied. However, prior to that, it is imperative to mention that morphological processing based on skeletonization [28] as shown in figure 11, is adopted in order to extract features that are peculiar to number 2, and also enhance the processing of further feature computations. Fig. 11. Shows steps of Arabic object number two The four conditions are as follows: the Euler Number should be equal to one. The ratio fraction of the object should be greater than or equal to 0.25. This factor is determined by dividing the width (W) distance by height (H) distance. In order to compute each of the mentioned distances, three basic points are located on the object as illustrated in figure 11, which are denoted by the following: P1 is positioned at the bottom of the object, P2 is positioned at the upper left, and P3 lays on the upper right. Since, each point has its coordinates x and y as P(x,y), Height (H), width (W) and slop distance are determined according to the Euclidean distance: √ (3) √ (4) √ (5) Then the ratio is computed as: (6) Third condition is that, the angle between the slop (P1-P3) and x-axes must be positive (larger than zero) to indicate that the object is number ‘2’, let P3(x3,y3) and P1(x1,y1) denote two points of the slop line as shown in figure 11, then the angle is computed using (7) and (8): (7) (8) The fourth condition is achieved by counting the number of flips, which should be equal to 2 in order to differentiate the number from number three. The following is the pseudo code for the four conditions: If [ Euler Number = 1 && Width_height_ratio > 0.25 && slop_ orientation > 0 && flip_num_top_horiz <=2] Then The Object is ‘2’; Number 0 Recognition: In case conditions for number ‘5’ ‘9’, ‘8’, ‘7’ ‘3’ and ‘2’ are not satisfied, the object will be examined with the conditions for number ‘0’ which are described as follows: Firstly, the Euler Number should be equal to one. Secondly, upper_dist Middle-dist lower-dist Flip_Top_Quarter Flip_Bottom_Quarter (1) (2) (3) P3 P3 P2 P1 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 402 | P a g e www.ijacsa.thesai.org solidity (S) factor should be greater than 0.9, (S > 0.9). Solidity (S) is a scalar specifying the proportion of the pixels object in the convex hull that are also in the region as shown in figure 12, it is computed in (9) as follows: (9) Fig. 12. Shows the Arabic object number zero Where Area of object (Area_S) is calculated according to the conditional equation in (10): { ∑ ∑ (10) And the convex hull is calculated as follows: ∑ ∑ (11) The third condition is that, the aspect ratio, which is the division of the minor axis by the major axes, should be more than 0.5. This factor is chosen as both of the mentioned axes sum up to 1, thus 0.5 is considered to take the worst case. The pseudo code of three conditions is: If [Euler Number = 1 && Solidity > 0.9 && Aspect_Ratio > 0.5] Then The Object is ‘0’; Number 4 Recognition: In case conditions for number ‘5’ ‘9’, ‘8’, ‘7’, ‘’3, ‘2’ and ‘0’ are not satisfied, the object will be examined with number ‘4’ conditions which are two conditions. Firstly, Euler number should be equal to one. Secondly, the number of flips should be greater than or equal to 4. To retrieve the number of flips, the object is scanned from the top middle point to the bottom of the image, as shown in figure 13. Fig. 13. Shows the Arabic object number four The pseudo code is as follows: If [Euler Number =1 && flip_num_middle_vertically >= 4] Then The Object is ‘4’; Number ‘1’ Recognition By examining the object with the aforementioned conditions, if the output of the object cannot be accurately decided, then the following conditions describing the facts about number ‘1’ will be considered. Similarly, in this case skeletonization is utilized to enable extraction of detailed information about number 1. Afterwards, the three conditions considered are: Firstly, Euler Number must equal to one. Secondly, the ratio fraction of the object shall be less than factor (0.25) (opposite of the number 2 and 6 recognition). This factor is determined by dividing width distance on the height distance. In order to compute each mentioned distance, three basic points have to be located in the object as in the figure 14, as following: P1 is positioned in bottom of the object, P2 is positioned upper left, and P3 lays upper right. As usual each point has its trajectories x and y as P(x,y). Height, width and slop distances are determined according to Euclidean distance and depicted in the figure 14. The determinations have been explained as in equations in (3), (4), (5) and (6). Additionally, third condition, the angle between the slop (P1-P3) and x-axes must be negative (less than zero) to indicate that this object is number ‘1’, let P3(x3,y3) and P1(x1,y1) are two points of the slop line as shown in figure 14. Then, the angle is computed by taking Tan inverse to the theta as explained in equation (7) and (8). Now, the pseudo code for the three conditions is as follows: If [(Euler Number =1 && Width_Height_Ration < 0.25 && Slop_Orientation < 0] Then The Object is ‘1’; Fig. 14. Shows the Arabic object number one Number 6 Recognition: Finally, if conditions for numbers ‘5’ ‘9’, ‘8’, ‘7’, ‘3’, ’2’, ‘0’, ‘4’, and ‘1’ have not been achieved successfully, the object will be examined with facts about number ‘6’ which is described in the following paragraph. Also, in this case skeletionalization is initially used to preprocess the object. Afterwards, the three conditions are: Firstly, Euler Number must equal to one. Secondly, the ratio fraction of the object should be greater than or equal to factor (0.25) (opposite of number ‘1’ recognition). This factor is determined by dividing width distance by the height distance. (1) (2) (3) P1 P3 P2 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 403 | P a g e www.ijacsa.thesai.org In order to compute each mentioned distance, three basic points are located on the object as depicted in the figure 15. P1 is positioned in bottom of the object, P2 is positioned upper left, and P3 lays upper right. Also, since each point has its trajectories x and y as P(x,y) thus the Height distance (H), width distance (W), slop distance, width height ratio are determined based on Euclidean distance using the following equations (12), (13), (14) and(15) : √ (12) √ (13) √ (14) (15) The third condition is that, the angle between the slop (P1- P2) and x-axes should be negative (less than zero) to indicate that this object is number ‘6’, which is computed as follows: (16), (17) The pseudo code for the three conditions is as follows: If [Euler Number =1 && Width Height Ration > 0.25 && slop orientation < 0] Then The Object is ‘1’; Fig. 15. Shows the Arabic object number six IV. EXPERIMENT AND IMPLEMENTATION To test the proposed algorithm and ascertain its ability to generalize to any random input number, thus, evaluating the algorithm with car plate numbers is considered in this experiment. The number of Arabic numerals is 226 characters collected randomly from 40 images of Iraqi car plates. It is important to mention that, the images are captured in real world conditions where several imaging factors such as illumination, shadow, camera view and incidental lighting are not constrained. Normally, in the verification or identification comparison, there are two possible error measures: False Accept Rate (FAR), which results from the forged template that accepted by the computer system falsely during testing. and False Rejection Rate (FRR), which results from the genuine template that the system recognizes as the query template wrongly [29, 30]. Finally, the total accuracy of the system is calculated by subtracting the average error rate from 100% as in (18): (18) In this research, FAR error does not exist, as there are no forge templates in this experiment. Therefore, FAR is mainly equal to zero. However, FRR is largely used for the testing measure to estimate the recognition rate, because the Arabic numbers are considered as genuine templates, if they are wrongly recognized by computer system, then the FRR increases. For example, number ‘5’ is deemed as genuine template, if the computer system recognized it as 5, FRR is going to be zero, otherwise FRR will be increased. Finally, the equations that are used to estimate the accuracy are as (19) and (20): (19) (20) The proposed algorithm is summarized in figure 16 as follows: (1) (2) (3) P1 P3 P2 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 404 | P a g e www.ijacsa.thesai.org Fig. 16. The proposed algorithm for the number recognition Yes No X=2 Yes Yes No No No No Yes Yes Yes No No Yes Yes No No Yes No No X=6 Width Height Ratio(WHR) && Slop Orientation(SO) WHR> 0.25 SO >0 Flip Middle Vertical (FMV) (flip_num_m_v >= 4) FMV>=4 X= 4 S=Solidity, AS=Aspect Ration S > 0.9, AS>0.5 X= 0 Yes Object is not exist WHR< 0.25 SO <0 X=1 WHR> 0.25 SO <0 Flip Lower Horizontal flips (FLH) X= 9 FLH <= 2 && AR >0.6 Aspect Ratio (AR) FMH >=3 X= 5 Start Input X Number X has Circle? Flip Mid Horizontal (FMH) Yes MD<LD UD<LD UD<MD X=8 Upper_Dist(UD), Mid_dist(MD) Lower_Dist (LD) X=7 UD>MD MD>LD UD>LD X= 3 FTQ >=6 FBQ<=2 Flip Top Quarter (FTQ) Flip Bottom Quarter (FBQ) En d Object is not exist (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 405 | P a g e www.ijacsa.thesai.org V. RESULT AND DISCUSSION Two different results are reported as the outcomes of this research. The first result is the recognition error for each distinct Arabic number among the 226 sample set, which is attained by calculating the total output of a specific numeral type divided by the total input (queried) of the same numeral type that have been randomly collected in the dataset. For example, as shown in Table 1, numeral number 8 has been iterated 25 times with no error in the recognition output (FRR=0). For the remaining numeral types, Table 1 shows the details as dataset iteration times with their system output and also shows the successful accuracy. It can be seen in table 1 that the highest False Reject Rate error (FRR) is 11.53%, which is specifically related to the Arabic number 6. TABLE I. SHOWS THE CHARACTER NUMBER TYPES, COMPUTER SYSTEM OUTPUTS, FRR AND SUCCESSFUL ACCURACY IN PERCENTAGE Numerical Type Dataset Numerical Attempt System Recognized Falsely: FRR as in (20) % Accuracy as in (19) % 0 19 0 0 100 1 46 1 2.17 97.83 2 15 1 6.66 93.34 3 24 1 4.16 95.84 4 20 1 5 95 5 18 0 0 100 6 26 3 11.53 88.47 7 18 0 0 100 8 25 0 0 100 9 15 0 0 100 In decreasing order, it can be seen that number 2 has 6.66 % error, number 3 has 4.16 and number 1 has FRR of 2.17%. These results are mainly due to the noise in the car plate images, while numbers zero, five, seven, eight, and nine have no error at all during testing. The second result that is reported in this research is obtained by calculating the recognition rate and FRR error for each car plate number, whether each one might consist of 4 or 5 or 6 numeral numbers. Figure 17 shows a bar chart which describes each attempt among the 40-image in x axes with their successful accuracy as in 100% in y axis of the chart. Fig. 17. Illustrates the 40 car plate numbers in X axis with their corresponding accuracies in Y axis Here, the overall successful accuracy is 97%, which is the average recognition rate for 40 car plate numbers. It is clear in figure 17 that the following images: 7, 17, 18, 19, 20, 22 and 23 have accuracy 83% because one of the numbers is not recognized correctly due to presence of noise in the input image. This rate (83%) is calculated as follows: In this research method, there is no dataset training to be matched against it as matching operation. However, it works by extracting facts by using geometric feature extraction in order to be applied to the set of rules by using if-else statements as an expert system works. VI. SUMMARY An effective algorithm for Arabic offline print written number recognition is proposed in this research. Several preprocessing operations are initially applied to the input image such as conversion to binary image, noise removal, morphological filtering, and segmentation before entering the data to the recognition system. The proposed approach is based on extraction of both local features such as computing number of flips of the only upper part of the object, as well as global geometric features such as computing the overall aspect ratio of the object, width, height and orientation of the object number. The features are further quantified into a set of facts or conditions that are used for classification based on a set of rules as an expert system. The experiment has been conducted on a random 226 numbers collected from 40 Iraqi car plate numbers. The output showed that the recognition error rate in terms of False Rejection Rate (FRR) is 3% or the overall successful accuracy is 97%. However, this algorithm is not robust against object translation and rotation. Finally, improving and investigating the possibility of using the proposed algorithm for handwritten Arabic number recognition rather than print written is considered as a future work. REFERENCES [1] A. Sabri Mahmoud and M. Sameh Awaida “Recognition Of Off-Line Handwritten Arabic (Indian) Numerals Using Multi-Scale Features And Support Vector Machines Vs. Hidden Markov Models,” The Arabian Journal for Science and Engineering, Vol. 34, No. 2B, 2009. [2] M. A. Abuzaraida, A. M. Zeki, “Feature Extraction Techniques of Online Handwriting Arabic Text Recognition,” IEEE conference, 5th International Conference on Information and Communication Technology for the Muslim World, 2013. [3] R. I. Zaghloul, Bader, D. M.K. Enas and F. AlRawashdeh “Recognition Of Hindi (Arabic) Handwritten Numerals. American Journal of Engineering and Applied Sciences,” Vol.5, No. 2,pp. 132- 135, 2012. [4] S. Sonavane, A. Khade, V. B. Gaikwad, ” Novel Approach for Localization of Indian Car Number Plate Recognition System using Support Vector Machine,” International Journal of Advanced Research in Computer Science and Software Engineering, Vol. 3, No.8, 2013. [5] S. Sivanandan, A, Dhanait, Y. Dhepale and Y. Saiyyad “ Automatic Vehicle Identification Using License Plate Recognition for Indian Vehicles,” Emerging Trends in Computer Science and Information Technology -(ETCSIT) Proceedings published in International Journal of Computer Applications® (IJCA), 2012. [6] A. T. Alqudah, H. R. Al-Zoubi and M. A. Al-Khassaweneh, ”Shift and Scale Invariant Recognition of Printed Numerals,” Abhath Al-Yarmouk: Basic Sci. & Eng. , Vol. 21, No.1, pp. 41-49, 2012. 0 20 40 60 80 100 120 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 4, 2016 406 | P a g e www.ijacsa.thesai.org [7] T. E. Milson and K. R. Rao, “A Statistical Model for Machine Print Recognition,” IEEE Transactions on Systems, Man, and Cybernetics, SMC, Vol.6, No.10 (1976), 671-678. [8] F. N. Said, R. A. Yacoub and C. Y. Suen, “Recognition of English and Arabic Numerals Using a Dynamic Number of Hidden Neurons,” Proceedings of the Fifth International Conference on Document Analysis and Recognition, (1999) 237-240. [9] G. Rajput, R. Horakeri, and S. Chandrakant “Printed and Handwritten Kannada Numeral Recognition Using Crack Codes and Fourier Descriptors Plate,” International Journal of Computer Applications, Vol.1, No.1, pp.53-58, (2010). [10] F. Al-Omari and O. Al-Jarrah, “Handwritten Indian Numerals Recognition System Using Probabilistic Neural Networks,” Advanced Engineering Informatics, Vol.18, No.1,pp. 9-16, (2004). [11] E. Kussul and T. Baidyk, “Improved Method of Handwritten digit Recognition Tested on MNIST Database,” Image and Vision Computing, Vol. 22, No.12, pp.971-981, (2004). [12] C. L. Liu, and H. Sako, “Class-Specific Feature Polynomial Classifier for Pattern Classification and its Application to Handwritten Numeral Recognition,”. The Journal of the Pattern Recognition Society, Vol.39, No.4, pp.669-681, (2006). [13] A. Goltsev, and D. Rachkovskij, “Combination of the Assembly Neural Network with a Perceptron for Recognition of Handwritten Digits Arranged in Numeral Strings,” The Journal of the Pattern Recognition Society, Vol.38, No.3, pp.315-322, (2005). [14] M. Kherallah, L. Haddad, A. Alimi and A. Mitiche “ On-line Handwritten Digit Recognition Based on Trajectory and Velocity Modeling,” Pattern Recognition Letters, Vol. 29, No. 5, pp.580-594, (2008). [15] F. Lauer, C. Suen, and G. Bloch “A Trainable Feature Extractor for Handwritten Digit Recognition,” The Journal of the Pattern Recognition Society, Vol.40, No.6, pp.1816-1824, 2007. [16] H. Soltanzadeh and M. Rahmati, “Recognition of Persian handwritten digits using image profiles of multiple orientations,” Pattern Recognition Letters, Vol.25, No.14, pp.1569-1576, (2004). [17] A. D. Parkins and A. K. Nandi, “Genetic Programming Techniques for Hand Written Digit Recognition,” The Journal of Signal Processing, vol. 84, No.12,pp.2345-2365, (2004). [18] H. Al-Zoubi, and M. Al-Khassaweneh, “Offline Machine-Print Hindi Digit Recognition Using Translational Motion Estimation,” Proceedings of IEEE International Conference on Computational Intelligence for Modeling, Control, and Automation, (CIMCA ’08), Vienna, Austria, 10- 12 December ,2789-2792, (2008). [19] M. Haralick, Robert, and L. G. Shapiro “Computer and Robot Vision,” Vo. I, Addison-Wesley, pp. 28-48, 1992. [20] C. R. Gonzales and R. E. Woods: Digital Image Processing. 2 nd edition, Englewood Cliffs, NJ: Prentice-Hull, 2002. [21] R. Van den Boomgard and R. van Balen ”Methods for Fast Morphological Image Transforms Using Bitmapped Images,” Computer Vision, Graphics, and Image Processing: Graphical Models and Image Processing, Vol. 54, No. 3, pp. 254-258, May 1992. [22] Jackson, Peter: Introduction To Expert Systems (3 rd ) Addison Wesley, p. 2,ISBN 978-0-201-87686-4, (1998). [23] M. Ahmed, R. K. Ward ”An expert system for general symbol recognition,” Pattern Recognition, Vol.33, No.12, pp.1975-1988, 2000. [24] M. Negnevitsky: Artificial intelligence: a guide to intelligent systems. Pearson Education, 2005. [25] M. Sasikumar, S. Ramani, S. M. Raman, K. S. R. Anjaneyulu and R. Chandrasekar, “A Practical Introduction to Rule Based Expert Systems,” New Delhi: Narosa Publishing House, 2007. [26] T. Sharma, N. Tiwari, and D. Kelkar, “Study of Difference Between Forward and Backward Reasoning,” Ujjain: IJETAE, 2012. [27] H. M. Robert, and L. G. Shapiro “Computer and Robot Vision,” Vol. I, Addison-Wesley, pp.28-48, 1992. [28] T. K. Yung and A. Rosenfeld “Topological Algorithms for Digital Image Processing,” Elsevier Science ,1996. [29] F. L. Malallah, S. M. S. Ahmad, W. A. W. Adnan, O. A. Arigbabu,V. Iranmanesh and S. Yussof “Online Handwritten Signature Recognition by Length Normalization using Up-Sampling and Down-Sampling,” International Journal of Cyber-Security and Digital Forensics (IJCSDF), Vol. 4, No.1, pp. 302-313, Dec. 2014. [30] F. L. Malallah, S. M. S. Ahmad, W.A.W. Adnan, S. Yussof “Non- Invertible Online Signature Biometric Template Protection via Shuffling and Trigonometry Transformation,” International Journal of Computer Applications, Vol .98, No.4.pp. 0975 – 8887. July, 2014. http://en.wikipedia.org/wiki/International_Standard_Book_Number http://en.wikipedia.org/wiki/Special:BookSources/978-0-201-87686-4 
work_bcj2seb6rzeh3jjwunry2puj34	----	() Appl. Math. Inf. Sci. 9, No. 4, 1709-1718 (2015) 1709 Applied Mathematics & Information Sciences An International Journal http://dx.doi.org/10.12785/amis/090407 Clustering of Telecommunications User Profiles for Fraud Detection and Security Enhancement in Large Corporate Networks: A case Study Constantinos S. Hilas1,∗, Paris A. Mastorocostas1 and Ioannis T. Rekanos2 1 Dept. of Informatics Engineering, Technological Educational Institute of Central Macedonia, GR-62124, Serres, Greece 2 Dept of Electrical and Electronics Engineering, School of Engineering, Aristotle University of Thessaloniki, GR-54124, Thessaloniki, Greece Received: 8 Oct. 2014, Revised: 8 Jan. 2015, Accepted: 9 Jan. 2015 Published online: 1 Jul. 2015 Abstract: A user’s transactions with modern networks and services produce a vast amount of user related data. The byproduct of every phone call a person makes or every web page one visits is translated into a log record with usage data. By studying these log records, the user’s behavior is revealed and one may come up with clues about user preferences, identify security issues, or discover fraudulent use of the network or the service one provides. Thus, the modeling of network users’ behavior may serve as an invaluable tool for the IT manager. In this paper, many of these issues are discussed and emphasis is given on the construction of appropriate user behavior representation in telecommunications. As an example, the application of two clustering techniques is presented, with the task to identify appropriate user behavior representations (profiles) inside a large organization’s telecommunications network, in order to spot fraudulent usage. Through this study a researcher and/or the organization’s network manager may gain more insight into the problems of user profiling and fraud detection. Keywords: User profiling, clustering applications, data mining, fraud detection, telecommunications 1 Introduction A user profile is a collection of personal data associated to a specific user. Ideally a user’s profile would be an explicit digital representation of that person. However, one may have several different profiles depending on the system with which he interacts. For example, from a personal computer’s perspective a user profile is a collection of settings that make the computer look and work the way you want it to. For an e-commerce system, e.g. an electronic book store, a user’s profile may include identity information, like login credentials; billing information, like credit card data; and more importantly, a user’s preferences as regards literature genres he prefers, recent purchases, etc. Some of this knowledge was explicitly entered to the system by the user, while other was implied by the system after analyzing past transactions. In this sense a profile emerges on the basis of monitoring ones usage patterns, so it is only relevant to a user’s attitude against a specific service. Of course, these data may be correlated with data from other databases and yield a more specific inference about one’s preferences. User profiles may be found in operating systems, e-commerce applications, social networking sites, recommender systems, intelligent tutoring systems, etc. The main idea behind user profiling is that the past behavior of a user can be accumulated in order to construct a profile, or a “user dictionary”, or a “user signature” of what might be the expected values of a user’s behavior. In its simplest form this profile is a vector that contains single numerical summaries of some aspect of behavior or some kind of multivariate behavioral pattern. A profile may also contain categorical, censored, or other non-numeric data. The profile follows the logic of the recordable data, with the constraints inherent in computer technology. This issue, the inherently reductive character of a profile, is important because profiles may impact privacy and identity in the strong sense (concerning our sense of self). ∗ Corresponding author e-mail: chilas@teicm.gr c© 2015 NSP Natural Sciences Publishing Cor. http://dx.doi.org/10.12785/amis/090407 1710 C. Hilas et. al. : Clustering of Telecommunications User Profiles for Fraud... Since profiles will often affect our lives (providing or prohibiting access, enabling selection, inclusion and exclusion) it is of utmost importance to clarify in what ways and on what basis they affect our lives, without conflating profile and profiled person [1]. This remark should always be kept in mind when constructing user profiles whether this is done in order to help or protect the user from other dangers. After a profile is constructed, it can be used appropriately. An operating system stores user profiles that contain one’s settings for desktop backgrounds, screen savers, pointer preferences, sound settings, and other features. A recommender system uses past user preferences in order to predict items that the user has not yet considered, e.g. a book, a song, or even a friend [2]. Traditionally, in computer security user profiles are constructed based on any basic usage characteristic such as resources consumed, login location, typing rate and counts of particular commands. Future behavior of the user can then be compared with his profile in order to examine the consistency with it (normal behavior) or any deviation from his profile, which may imply a breach in security or some fraudulent activity. In particular, fraud detection is important to the telecommunications industry because companies and suppliers of telecommunications services lose a significant proportion of their revenue as a result. Moreover, the modeling and characterization of users’ behavior in telecommunications can be used to improve network security, improve services, provide personalized applications, and optimize the operation of electronic equipment and/or communication protocols. Several categories of telecommunications fraud have been reported in the literature. The most prominant are the technical fraud, the contractual fraud, the procedural fraud, and the hacking fraud [3]. The first three usually burden the economics of the service provider, while hacking fraud also harms the subscriber. Hacking fraud is usually met in the form of the superimposed fraud where the fraudster (hacker) uses a service concurrently with the subscriber and burdens his account. The present paper focuses on superimposed fraud identification. The Communications Fraud Control Association (CFCA) recently announced the results of a global survey carried out in 2013. Experts estimate 2013 fraud losses at $46.3 billion (USD), up 15% from 2011. As a percent of global telecom revenues, fraud losses are approximately 2.09%, a 0.21% increase from 2011. The main reason for the relative increase in fraud is due to more fraudulent activity targeting the wireless industry [4]. It should also be noted that the relative decrease in fraud, that was apparent in previous surveys, was not an actual decline in absolute values but it had been attributed to the fact that the growth in global telecom revenues had outpaced the growth in fraud losses in the past (e.g. the CFCA 2011 survey). Exchange of ideas in fraud detection is limited by the fact that it makes no sense to describe the methods in detail, as it gives fraudsters the information they require to evade detection. Moreover, companies and organizations that have been defrauded refrain from revealing the situation due to reputation concerns. Adding to this, fraud detection problems involve huge data sets, which are constantly evolving. Data sets can be as large as tenths of thousands of calls per weekday for a large organization with 3 or 4 thousand employees, to hundreds of millions of calls for national carriers. One should also consider the size of the related metadata. Another difficulty with fraud detection is the fact that, nowadays, the term telecommunications is wider than ever. It includes both wired and wireless systems, mobile and cellular systems, legacy systems (PSTN, ISDN), terrestrial and satellite networks, and a plethora of Internet–based communication applications. The diversity of network types and applications, along with the deregulation of the market and the relocation of services to the cloud, makes fraud detection a complex task. Research in telecommunications fraud detection is mainly motivated by fraudulent activities in mobile technologies [3, 5]. Recent research also focuses on VoIP technologies [6]. Fraud detection methods can be based on statistical or machine learning techniques and may be supervised or unsupervised [7, 8, 9]. Fawcett and Phua [10] have ensembled the bibliography on the use of data mining and machine learning methods for automatic fraud detection up to 2005. This paper proceeds as follows: In the next Section the user modeling procedure and the proposed user profiles are presented. A brief presentation of the clustering techniques and the clustering quality statistics that are used is given in the third Section. The user data and the outcome of the analysis are described in the fourth Section. In the last Section conclusions are discussed. 2 Behavior Modeling and Profiling in Telecommunications The data that can be used to construct the basic profile vector for a telecommunications user are contained in the Call Detail Record (CDR) of any Private Branch Exchange (PBX) or any VoIP switch. The format of the CDR varies among providers or programs. Some programs allow CDRs to be configured by the user. In most cases a CDR contains at least data such as: the caller ID, the chargeable duration of the call, the called party ID, the date and the time of the call, etc [11]. In mobile telephone systems, such as GSM, the data records that contain details of every mobile phone attempt are the Toll Tickets. Location data may also be useful. In computer security, user profiles may be constructed based on any basic usage characteristic such as resources consumed, login location, typing rate and counts of particular commands. c© 2015 NSP Natural Sciences Publishing Cor. Appl. Math. Inf. Sci. 9, No. 4, 1709-1718 (2015) / www.naturalspublishing.com/Journals.asp 1711 Classifier Y -A xi s Suspicious Cases Fraud Normal Events ������ ������ ������ Call Database Billing Services Fig. 1: Both the billing and the security systems draw usage related data from a common database. CDR data are primarily collected in order to charge the user for the service he gets, but may additionally be used to manage the security of the system. The study of CDRs may reveal cases of unauthorized use of the system which may burden both the provider and the subscriber (Fig.1). The system administrator may investigate unusual call patterns, multiple calls to the same number, calls outside normal working hours, calls with long duration, phone calls to suspicious destinations such as international or premium-rate services (e.g. 090, 1-900, 900). Also, in the case of private PBXs, where access to outgoing destinations is done by means of personal authorization codes (PAC), one may check for frequent or simultaneous use of a PAC, attempts to import nonexistent system codes (efforts to find codes), or using PACs that are not associated with a user [12]. In order to develop models that can be applied to distinguish legitimate from fraudulent usage, one needs examples of both cases. Finding data of normal usage is easy. In fact, most of the data generated during the operation of a system are of this type. The data relating to fraud are relatively rare and any effort to characterize them in detail is difficult, time consuming, and requires specialized knowledge. Moreover, the processing and storage of user data is subject to restrictions by the legislation that protects privacy [13, 14]. Several methods appear in the literature for collecting and classifying data usage. An approach that involves the user in the process, and thus may overcome privacy concerns, is block crediting, proposed by Fawcett and Provost [15]. According to this method both the provider and the subscriber are involved in characterizing the data. If the subscriber has any objection on the amount of his bill, he may contact the provider and collaborate in order to characterize the calls in his detailed bill and then pay for the corresponding price. Typically, the account is divided into two periods, one with the calls of the legitimate user and one that also contains calls from the usurper. However, this kind of separation does not provide high accuracy in the characterization of cases. On the other hand, a detailed per phone call characterization of the account is by nature expensive and requires too much human intervention. In general, fraud detection focuses on the analysis of users’ activity and the related approaches are divided into two main subcategories. The absolute one that searches for limits between legal and fraudulent behavior, and the differential approach that tries to detect extreme changes in a user’s behavior. All cases of telecom fraud can actually be viewed as fraud scenarios which are related to the way the access to the network was acquired. One of the most interesting aspects of the problem is the evaluation of different user representations (profiles) and their effect towards the proper discrimination between legitimate and fraudulent activity. In the present analysis, for each user, three different profile types are constructed and tested. The first profile (Profile1) is build up from the accumulated weekly behavior of the user. The profile consists of seven fields which are the mean and the standard deviation of the number of calls per week (calls), the mean and the standard deviation of the duration (dur) of calls per week, the maximum number of calls, the maximum duration of one call and the maximum cost of one call (Fig.2). All maxima are computed within a week’s period. The second profile (Profile2) is a detailed daily behavior of a user which is constructed by separating the number of calls per day and their corresponding duration per day according to the called destination, i.e., national (nat), international (int), and mobile (mob) calls, and the time of the day, i.e., working hours (w), afternoon hours (a), and night (n) (Fig.3). Last, the third profile (Profile3) is an accumulated per day behavior (Fig.4). It consists of the number of calls and their corresponding duration separated only according to the called destination, that is, national, international and mobile calls. The last two profiles were also accumulated per week to give Profile2w and Profile3w. So, overall five different user profile representations are evaluated. c© 2015 NSP Natural Sciences Publishing Cor. www.naturalspublishing.com/Journals.asp 1712 C. Hilas et. al. : Clustering of Telecommunications User Profiles for Fraud... mean(calls) mean(dur)std(calls) max(cost)max(dur)max(calls)std(dur) Fig. 2: Profile1 of telephone calls int_calls_w int_dur_w int_dur_nint_calls_nint_dur_aint_calls_a nat_calls_w nat_dur_w nat_dur_nnat_calls_nnat_dur_anat_calls_a mob_calls_ w mob_dur_w mob_dur_n mob_calls_ n mob_dur_a mob_calls_ a Fig. 3: Profile2 of telephone calls nat_calls int_durint_callsmob_durmob_callsnat_dur Fig. 4: Profile3 of telephone calls In the past, feed-forward neural networks have applied to classify cases of user behavior [16]. No matter how well a neural network classifier may have performed, there is no clue about the features it actually used in order to achieve its performance. So, it is difficult to identify important characteristics that led to a successful classification. In order to further investigate the problem of appropriate user modeling towards fraud detection, we also apply clustering techniques on the data. The aim is to test whether cases from the same class tend to form clusters and under which condition. An important characteristic of most clustering methods is that they are actually unsupervised learning approaches. Thus, one does not have to provide the corresponding algorithms with class examples. On the contrary, the algorithm is left to decide on case similarities by means of a pre-selected distance (similarity) measure [17]. It is hoped that clustering will unveil important information on the nature of user data and on the key features that may be used to distinguish between legitimate and fraudulent usage. 3 Clustering Techniques and Clustering Quality Measures Clustering is one of the most important sets of unsupervised learning tools among the machine learning techniques and algorithms. It deals with finding a structure or an intrinsic grouping in a collection of unlabeled data. During the clustering procedure some objects are labeled “similar” to each other and “dissimilar” to objects belonging to other clusters. So, clustering aims on organizing objects into groups whose members are similar in some way. The researcher needs to decide on the similarity criterion that will be used during the process as well as on the clustering quality criterion that will be applied to decide what constitutes a good clustering. Clustering has been applied to a wide range of topics, such as pattern recognition, compression, and classification, as well as in diverse disciplines like biology, marketing, psychology and business. For a detailed introduction of clustering techniques the reader is referred to Kaufmann and Rousseeuw [17]. Two of the most common clustering techniques are used in the present work, namely the partitioning and the hierarchical clustering. As a main representative of the partitioning techniques we will apply the k-means algorithm. The hierarchical clustering technique that will be used is the agglomerative clustering. Their main difference is that the first needs the user’s input on the expected number of clusters while the latter needs no user intervention. 3.1 K-means clustering The main idea behind the k-means algorithm is that in order to obtain k clusters, the algorithm selects k objects from the data sets. The remaining objects are then assigned to the nearest representative object, as the algorithm attempts to minimize the average squared distance between objects. Other distance measures can be used as well. A graphical representation of the clustering is provided by displaying the silhouettes introduced by Kaufmann and Rousseeuw [17]. Silhouettes are constructed in the following way. Given the number of clusters in the problem, the value s(i) is defined for each object i and these numbers are presented in a plot. The value s(i) is defined by: s(i) = b(i)− a(i) max{a(i),b(i)} , (1) where a(i) is the average dissimilarity of i to all other objects in A, where A is the cluster in which the object i is first assigned, with d(i,C) being the average dissimilarity of i to all objects of any other cluster C different from A. Object i is assigned to the cluster B that satisfies b(i) = d(i,B). 3.2 Agglomerative Clustering During hierarchical agglomerative clustering the user does not specify the expected number of clusters k. Instead, the algorithm constructs a tree-like hierarchy, a dendrogram, which implicitly contains all values of k. The root of the tree structure defines a cluster that contains all data, while its leafs represent n clusters, each one containing one of the n objects. The agglomerative clustering algorithm starts with each object representing a cluster, called a singleton, and proceeds by fusing the c© 2015 NSP Natural Sciences Publishing Cor. Appl. Math. Inf. Sci. 9, No. 4, 1709-1718 (2015) / www.naturalspublishing.com/Journals.asp 1713 closest ones until a single cluster is obtained. Therefore, a measure of dissimilarity between two clusters must be defined [17]. Two different distance measures, namely the Euclidean distance and the correlation between objects, are used for both clustering algorithms. 3.3 Clustering Quality Measures In order to judge on the quality of the clustering structure, we apply three quality statistics. These are the silhouette coefficient (SC), the agglomerative coefficient (AC) and the cophenetic coefficient (CC). Visual inspection of the dendrograms and the consistency of their structure may also reveal certain characteristics. The silhouette coefficient (SC) is given by: SC = max k s̄(k), (2) where s(k) is derived from (1) and the maximum is taken over all k for which the silhouettes can be constructed. The SC is a useful measure of the amount of clustering structure that has been discovered by the classification algorithm. Another measure, that allows us to decide whether the clustering reveals a clear separation between groups, is the comparison of the within–cluster dissimilarity with the between–cluster dissimilarity. This measure in a tree structure is equivalent to the tendency that the tree branches get taller. If for each object i one measures its distance, l(i), from the tree root, normalize the value in [0-1] and compute the average value, then he has a measure of the average tree height. This average, whose larger values imply better clustering, is named agglomerative coefficient (AC), and it is given by: AC = 1 n n ∑ i=1 l(i), (3) The cophenetic coefficient, CC, is a measure of how faithfully a dendrogram preserves the pair-wise distances between the original un-modeled data points. Given the matrix Z with the distances between clusters and the object dissimilarity matrix Y , then CC can be used as a measure of correlation between the object distances and their distance from the tree root. The value of CC is given by: CC = ∑ i, j i< j ( Yi j −Ŷ )( Zi j − Ẑ ) √ ∑ i, j i< j ( Yi j −Ŷ )2 ∑ i, j i< j ( Zi j − Ẑ )2 , (4) where Yi j is the distance of object i from j. CC values closer to 1 reveal better clustering structures [18]. One last criterion that may help choose in favor of one clustering structure over another is the consistency of the structure. If a branch in a dendrogram is at the same height with its neighbors, then this reveals similarity between the corresponding objects at that level of the structure. This behavior can be considered consistent. Whenever a branch height significantly differs from that of its neighbors, this behavior reveals high values of dissimilarity between the corresponding objects and can be named inconsistent. Inconsistent links in a dendrogram may reveal cluster separation points. For a more detailed analysis on cluster selection criteria the reader is referred to [19]. 4 An Application of User Profile Clustering Our experiments are based on real data extracted from a database that holds the CDR for a period of eight years from an organization’s PBX. According to the organization’s charging policy, only calls to national, international and mobile destinations are charged. Calls to local destinations are not charged so they are not included in the examples. In order to properly charge users, for the calls they place, a system of Personal Identification Numbers (PIN) is used. Each user owns a unique PIN which “unlocks” the organization’s telephone sets in order to place costly outgoing calls. If anyone (e.g., a fraudster) finds a PIN he can use it to place his own calls from any telephone set within the organization. Several user accounts, which have been defrauded, have been identified. The detailed daily accounts were examined by a field expert and each phone call was marked as either normal or defrauded. If during a day no fraudulent activity was present, then the whole day was marked as normal. If at least one call from the fraudster was present, then the whole day was marked as fraud. This separation of the classes is named detailed characterization. Adding to this, each day was also marked, according to the first time that fraudulent activity appeared. Thus, each user’s account is split into two sets, one pre- and one post-fraud. This separation of the classes is named coarse characterization. In this work we use the examples of 6 users (it has been stressed earlier that it is difficult to isolate fraud cases), 5 profile representations, and 2 different ways to characterize the user accounts as normal of fraudulent. The above give 60 different data sets. The user daily profiles will be used as an input for the two algorithms. We will ask the k-means algorithm to partition the input space in two distinct groups. If the legitimate and the fraudulent behavior cases are sufficiently different from each other, then the k-means algorithm will provide us with two distinct clusters of data. If this is not the case, then the division will not be good. Then the same input data will be fed into the agglomerative clustering algorithm and we will also check whether there is an output with distinct cases separation. c© 2015 NSP Natural Sciences Publishing Cor. www.naturalspublishing.com/Journals.asp 1714 C. Hilas et. al. : Clustering of Telecommunications User Profiles for Fraud... Table 1: Clustering outcome for a defrauded user account using the five profiles and two similarity measures, Euclidean distance (Eucl) and Correlation (Correl) Profile Distance Correct SC AC CC Measure Clustering (%) Profile1 Eucl 77.9310 0.7734 0.9632 0.9143 Profile1 Correl 84.8276 0.8729 0.9881 0.9299 Profile2 Eucl 78.0374 0.5303 0.9738 0.9514 Profile2 Correl 67.2897 0.2624 0.9952 0.9698 Profile3 Eucl 67.7570 0.8104 0.9824 0.9160 Profile3 Correl 76.1682 0.7668 0.9984 0.9725 Profile2w Eucl 78.0374 0.5264 0.9738 0.9514 Profile2w Correl 77.1028 0.3429 0.9952 0.9698 Profile3w Eucl 71.2329 0.7533 0.9748 0.9338 Profile3w Correl 77.3973 0.7296 0.9956 0.9489 Table 2: Clustering statistics for six defrauded user accounts and two similarity measures – Profile1 has been used in all cases Profile Distance Correct SC AC CC Measure Clustering (%) User1 euclidean 77.9310 0.7734 0.9632 0.9143 correlation 84.8276 0.8729 0.9881 0.9299 User2 euclidean 65.2582 0.9466 0.9770 0.9118 correlation 78.8732 0.7152 0.9975 0.9232 User3 euclidean 73.3154 0.9989 0.9944 0.9910 correlation 67.1159 0.5847 0.9945 0.6782 User4 euclidean 79.4979 0.8570 0.9706 0.8857 correlation 92.4686 0.6865 0.9949 0.8302 User5 euclidean 62.1367 0.5784 0.9544 0.6278 correlation 79.4149 0.8750 0.9706 0.8785 User6 euclidean 76.9031 0.7632 0.9612 0.8949 correlation 86.4668 0.7445 0.9903 0.8302 In Table 1 the outcome of the clustering procedure for a defrauded user’s account is listed for the five profiles and two similarity measures. These are the results for User1. In general, the best results, i.e. the higher percentage of correct clustering, were achieved with the use of Profile1 combined with correlation as the similarity measure. In Table 2 the values of the statistics of Section 3.3 are presented for the cases of four defrauded user accounts. For each user the similarity measure and the correct clustering percentage are also given. Bold numbers are used to stress best statistics values. In general, correlation gave better results. The case of User3 (Table 2) is an exception that will be dealt with shortly. Due to space limitations only the findings for Profile1 will be depicted in the figures that follow. First, the clustering ability of each profile is tested by means of the k-means algorithm using the Euclidean distance as a similarity measure. Comparison of the clustering outcome with the original class separation reveals that the clustering is correct by 77.9%. However, visual observation of the silhouette diagrams (Fig.5(a)) shows the existence of negative silhouette values at the bottom of the figure which imply wrong clustering. The SC for the clustering in this figure is 0.7734. If the procedure is repeated but with the correlation of the input vectors as the similarity measure, then the outcome is depicted in (Fig.5(b)). Now, the percentage of correct clustering is 84.8% and the SC value is 0.8729. In general, the k-means clustering showed that better separation between classes is achieved when user behavior is modeled with Profile1, where the detailed characterization of the cases is used and correlation is employed as the similarity measure. Examples of the agglomerative clustering of the same user’s behavior and the same profile (Profile1) are shown in Fig.6. The former plot (Fig.6(a)) is a case where the Euclidean distance is used as measure of the similarity between objects, while in the latter (Fig.6(b)) the objects’ correlation is used instead. Fig.6 is reprinted here from [8] with permission from Elsevier. In all cases the analysis reveals that whenever the correlation is used as the distance measure then one gets the highest percentage of correct clustering. Visual inspection of the dendrograms is also important. The difference of correct clustering between the dendrograms in Fig.6(a) and Fig.6(b) is only 7%. However, each one reveals two completely different views of the data. In the first one, one distinct cluster is formed that contains all the outliers regardless of their class membership. In the second, clusters are formed which include only pure fraud cases. In this case there are three clusters that contain the fraud cases [8]. This implies a type of mixed behavior of the fraudster. Actually, a dendrogram structure implicitly contains all possible cluster separations and is up to the analyst to decide the correct number of clusters and the point of separation. It is interesting to study the silhouette diagrams in conjunction with their agglomerative clustering counterparts. Especially in the case where the Euclidean distance is used, one may observe that the silhouette diagram shows some kind of “bad” clustering that may mislead the researcher, but its corresponding dendrogram conveys a behavior that needs to be further explored. In fact, it was an expert’s intervention, who examined all the cases in the separate cluster and revealed that these belong to outliers, i.e. rare cases of extreme network usage by the legal user or the fraudster. Fig. 7, which is reprinted here from [8] with permission from Elsevier, shows the dendrogram for User3 of Table 2 when the correlation is used as the similarity measure. From Table 2 one may observe that for the case of User3 the clustering statistics for correlation are worst than the ones for the Euclidean distance similarity measure. However, a distinct cluster with pure legal behavior is formed ((Fig. 7)). On the contrary, the silhouette diagram for the same user with the Euclidean distance, gives two clusters where the second one has only one member. Then the first one includes all legitimate usage cases and gives the high (73.31) correct clustering percentage. c© 2015 NSP Natural Sciences Publishing Cor. Appl. Math. Inf. Sci. 9, No. 4, 1709-1718 (2015) / www.naturalspublishing.com/Journals.asp 1715 −0.5 0 0.5 1 1 2 Silhouette Value C lu st e r User1 − Profile1 (a) 0 0.2 0.4 0.6 0.8 1 1 2 Silhouette Value C lu st e r User1 − Profile1 (b) Fig. 5: Silhouette diagram for the clustering of a user’s profile into two clusters using (a) the Euclidean distance as a similarity measure – the negative silhouette values at the bottom of the figure imply wrong clustering of the user behaviour, and (b) correlation as a similarity measure – here two well separated clusters are formed 5 Conclusions and Discussion As large companies, organizations or institutions seek to grow their customer base and revenue, they must increasingly combat sophisticated security threats while navigating the growing challenge of compliance risk related to adhering with national or international laws. Business needs drive an unprecedented demand for complex IT networks and applications that call for the 0 200 400 600 800 1000 1200 User1 − Profile1 Two clusters cutting point Outliers (a) 0 0.05 0.1 0.15 0.2 0.25 User1 − Profile1 four clusters two clusters Different clustering limits 70% of fraud cases (b) Fig. 6: Hierarchical agglomerative clustering of a user’s profile using as similarity measure (a) the Euclidean distance between objects – revealing outliers, and (b) the correlation between objects – revealing distinct case groups establishment of authentication and fraud detection solutions throughout the enterprise. To effectively improve security and compliance controls, total cost of ownership and the end-customer experience, companies are adopting identity-based security and fraud detection platforms that span enterprise-wide. In particular, for the telecommunications industry, fraud detection is important because companies and suppliers of telecommunications services lose a significant proportion of their revenue as a result. As regards large organizations and companies, problems with internal fraudulent activities, i.e. misuse of the corporate c© 2015 NSP Natural Sciences Publishing Cor. www.naturalspublishing.com/Journals.asp 1716 C. Hilas et. al. : Clustering of Telecommunications User Profiles for Fraud... 0 0.05 0.1 0.15 0.2 0.25 0.3 three clusters almost all legal behaviour Fig. 7: The dendrogram for User3 of Table 2 when the correlation is used as the similarity measure. A distinct cluster with all legal behavior is formed (reprinted from [8] with permission from Elsevier) network or usurpation of colleagues’ identities burden the company’s budget or degrade the management reputation among employees. Adding to these, the modeling and characterization of users’ behavior in telecommunications can be used to improve network security, improve services, provide personalized applications, and optimize the operation of electronic equipment and/or communication protocols. In this context, we present how some well established unsupervised learning techniques may be applied on telecommunications data in order to provide us with more insight on both a legitimate user’s and a fraudster’s behavior. Prior to this, raw usage data must be transformed into appropriate user profiles. The profile construction and selection is also a challenge. From the analysis it is concluded that, as regards user profile building, accumulated characteristics of a user yield better discrimination results. However, aggregating user’s behavior for larger periods than a week was avoided in order to preserve some level of on-line detection ability. Clustering reveals that misclassification occurs due to mixed types of behavior. That is, there are cases for which the legitimate user acts like a fraudster, e.g. making long or expensive calls, while at the same time a fraudster may act like a normal user, e.g. he is more refrained and tries to mimic legitimate user behavior. So, if one forces the clustering procedure to produce two distinct groups of data (one with the legitimate and one with the fraudulent usage) it will fail to do it correctly. It was evident from the analysis that there was also some kind of “noise” in the data. This is due to the fact that even the field expert, that manually characterized the cases in the first place, had a hard time distinguishing between the two classes. This observation reveals the complexity in a fraud detection problem. This “grey” area in the classification problem is depicted as the suspicious cases in Fig.1; a space between pure legal and fraudulent activity. Taking into account the aforementioned observations, i.e. the mixed types of behavior and the noisy data, it is concluded that clustering alone can not be used as the only decision technique on whether a user account is defrauded or not. Input from other methods is also needed, and a more complex framework should be used. Nonetheless, clustering may provide the researcher with invaluable insight on the nature of data and on the nature of user behavior. The user profiles that were presented, in this work, have the benefit of respecting users’ privacy. That is, except from some coarse user behavior characteristics, all private data (e.g. called number, calling location, etc) are hidden from the analyst. Private data would definitely enhance the accuracy of fraud detection. In fact, the expert who characterized the data sets, in the first place, used rules based on private data and his domain specific knowledge. However, our aim is to test the ability to detect fraud given the minimum possible information about the user. A fundamental aspect of any predictive problem in data analysis is the choice of an appropriate criterion for estimation and performance assessment. In the case of fraud, one needs, in particular, to combine both classification accuracy and timeliness of classification. Moreover, the granularity of the timeline of events plays an important role in the predictive ability of models. This, also, means that standard measures of classification performance, such as the error rate, AUC, KS statistic, information value, etc, may not be sufficient. This is evident in the present work by the fact that it is difficult to judge on the most appropriate clustering outcome just by evaluating similarity statistics. As can be seen in Table 1 the percentage of correct clustering is a vital criterion, while visual observation of the corresponding graphs is also needed. One of the main tasks in future research should be, apart from the selection of appropriate user models (profiles), the effort to design suitable measures and performance curves which combine these aspects. The development of generalized frameworks that would be capable of adapting to different environments is highly desired. The presented methodologies may, also, be applied to study similar problems in mobile or data networks. The first step should be the appropriate user modeling, i.e. the selection of the appropriate attributes for profile building. Towards this step, an expert’s intervention is highly needed. Adding to this, due to privacy concerns and restrictions, the user’s collaboration may also be asked for. The user may give his consent for the analysis of his account which helps overcome some legislation problems regarding privacy. His collaboration may also help the analyst to identify fraudulent activity more precisely. c© 2015 NSP Natural Sciences Publishing Cor. Appl. Math. Inf. Sci. 9, No. 4, 1709-1718 (2015) / www.naturalspublishing.com/Journals.asp 1717 Nonetheless, one should keep in mind that security problems often depend on the specific nature of each problem and there is a risk when directly generalizing the findings even to similar environments. Among other risks, there underlies the danger of information leakage between interconnected IT systems which may burden users’ privacy. Future research includes the application of other clustering methods like the subtractive clustering, or other unsupervised techniques like the SOM, on the problem. Social network analysis is also of great interest and seems like an appealing alternative to statistical approaches or computational intelligence ones. What is intriguing about its application on the problem is that we expect to approach user behavior not by means of some usage statistic but by detecting transactions between the persons behind the statistics. Recent evidence suggests that technical controls only detect one third of fraud cases with zero time exposure and loss. More complex fraud is detected with a range of technical and sociotechnical controls from inside and outside the firm [20]. It is hoped that further research will give important findings on how to distinguish between legitimate and fraudulent network usage. Moreover, the departure from a strict technical approach and the use of social and behavioral controls used in the organizational environment may provide more clues about the problem. Acknowledgement The authors would like to thank the members of the Telecommunications Centre of the Aristotle University of Thessaloniki for their contribution of anonymised user data. The authors wish to acknowledge financial support provided by the Research Committee of the Technological Education Institute of Central Macedonia (Serres) under grant SAT/IC/181212-157/7 References [1] Hildebrandt, Mireille and James Backhouse (eds.), D7.2: Descriptive analysis and inventory of profiling practices. FIDIS Deliverable, available online at: http://www.fidis.net/resources/deliverables/profiling/ (2005). [2] Adomavicius, G., Huang, Z., and Tuzhilin, A., Personalization and Recommender Systems. In Z.-L. C. a. S. Raghavan (Ed.), State-of-the-Art Decision Making Tools in the Information-Intensive Age (pp. 55-100): Tutorials in Operations Research (2008). [3] P. Gosset and M. Hyland, Classification, detection and prosecution of fraud in mobile networks, Proceedings of ACTS Mobile Summit, Sorrento, Italy, June (1999). [4] Communications Fraud Control Association. 2013 Global Fraud Loss Survey, Press Release, Roseland, NJ (CFCA) October 10, 2013, available online: http://www.cfca.org/press.php (2013). [5] Moreau Y., Preneel B., Burge P., Shawe-Taylor J., Stoermann C., Cooke C., Novel Techniques for Fraud Detection in Mobile Telecommunication Networks, ACTS Mobile Summit, Granada Spain (1997). [6] Kapourniotis, T. Dagiuklas, T., Polyzos, G., Alefragkis, P., Scam and fraud detection in VoIP Networks: Analysis and countermeasures using user profiling, 50th FITCE Congress (FITCE), Palermo Italy, doi: 10.1109/FITCE.2011.6133427, (2011). [7] R. J. Bolton and D. J Hand, Statistical fraud detection: a review, Statistical Science, 17.3, 235255, (2002). [8] Constantinos S. Hilas and Paris As. Mastorocostas, An Application of Supervised and Unsupervised Learning Approaches to Telecommunications Fraud Detection, Knowledge-Based Systems, 21, 721 726, doi:10.1016/j.knosys.2008.03.026, (2008). [9] Olszewski, Dominik, A probabilistic approach to fraud detection in telecommunications, Knowledge-Based Systems, 26, 246-258, (2012). [10] Fawcett, T. and Phua, C., Fraud Detection Bibliography, Accessed From: http://liinwww.ira.uka.de/bibliography/Ai/fraud.detection.html (2005). [11] S. F. Hinde, Call Record Analysis, Making Life Easier - Network Design and Management Tools (Digest No: 1996/217), IEE Colloquium on, 8/1 8/4, (1996). [12] DEFINITY Enterprise Communication Server (ECS): Technical Articles and Technical Tips, available online: http://esearch.avaya.com/ [13] Directive 95/46/EC of the European Parliament and of the Council of 24 October 1995 on the protection of individuals with regard to the processing of personal data and on the free movement of such data, Official Journal L 281/31 of 23.11.95, pp 31-50, available online at: http://eur- lex.europa.eu/LexUriServ/LexUriServ.do?uri=CELEX: 31995L0046:EN:NOT (1995). [14] Directive 2002/58/EC. Directive on privacy and electronic communications. Official Journal L 201, 31.7.2002, 3747, available online at: http://eur- lex.europa.eu/LexUriServ/LexUriServ.do? uri=CELEX: 32002L0058:EN:NOT (2002) [15] Fawcett, T. and F. Provost, Adaptive fraud detection, Journal of Data Mining and Knowledge Discovery, 1.3, 291316 (1997). [16] Constantinos S. Hilas, and John N. Sahalos, Testing the fraud detection ability of different user profiles by means of FFNN classifiers, in Collias St. et al ed.. Lecture Notes in Computer Science, 4132, Part II, Berlin-Heidelberg: Springer Verlag, 872-883 (2006). [17] Kaufman L., and Rousseeuw P.J., Finding groups in data: an introduction to cluster analysis, New York: John Wiley & Sons, Inc. (1990). [18] Rohlf, F. J. and David L. Fisher, Test for hierarchical structure in random data sets, Systematic Zool., 17, 407-412 (1968). [19] Witten I. H., and Frank E., Data mining: practical machine learning tools and techniques with Java implementations, London: Academic Press, (2000). [20] Goode, S., and Lacey, Detecting complex account fraud in the enterprise: The role of technical and non-technical controls, Decision Support Systems, 50.4, 702-714 (2011). c© 2015 NSP Natural Sciences Publishing Cor. www.naturalspublishing.com/Journals.asp http://www.fidis.net/resources/deliverables/profiling/ http://www.cfca.org/press.php http://liinwww.ira.uka.de/bibliography/Ai/fraud.detection.html http://esearch.avaya.com/ 1718 C. Hilas et. al. : Clustering of Telecommunications User Profiles for Fraud... Constantinos S. Hilas received the B.S. degree in Physics, in 1989, the M.S. degree in Electronic Physics - Radioelectrology, in 1996, and the Ph.D. degree in Physics, in 2007, from the Aristotle University of Thessaloniki (AUTH), Greece. He also received the M.S. degree in Information Systems from the University of Macedonia (UM), Thessaloniki, Greece, in 2000. From 1993 to 2002, he was Computer Administrator and the Administrator of the Telecommunications Network of AUTH. Since 2002, he has been with the Technological Educational Institute of Central Macedonia (Serres, Greece), where he is currently an Associate Professor in the Dept. of Informatics Engineering. His basic research interests include user characterization, service development, and data mining techniques in telecommunications and networking. Paris A. Mastorocostas received the Diploma and Ph.D. degrees in Electrical & Computer Engineering from Aristotle University of Thessaloniki, Greece. Presently he serves as Professor at the Dept. of Informatics Engineering, Technological Education Institute of Central Macedonia, Serres, Greece. His research interests include fuzzy and neural systems with applications to identification and classification processes, data mining and scientific programming. Ioannis T. Rekanos received the M.Sc. and the Ph.D. degree in electrical and computer engineering, in 1993 and in 1998, respectively, both from the Aristotle University of Thessaloniki (AUTH), Greece. From 2000 to 2002, he was a Postdoctoral Senior Researcher in the Radio Laboratory at the Helsinki University of Technology, Finland. From 2002 to 2006, he served as a faculty member of the Department of Informatics and Communications, TEI of Serres, Greece. Since 2006, he has been with the AUTH, where he is currently an Associate Professor in the Department of Electrical and Computer Engineering. His research interests include electromagnetic and acoustic wave propagation, inverse scattering, computational electromagnetics, and digital signal processing. c© 2015 NSP Natural Sciences Publishing Cor. Introduction Behavior Modeling and Profiling in Telecommunications Clustering Techniques and Clustering Quality Measures An Application of User Profile Clustering Conclusions and Discussion 
work_bcpbwz3m2rh6bibnz7jai32lf4	----	Combination of multiple diagnosis systems in Self-Healing networks Expert Systems With Applications 64 (2016) 56–68 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Combination of multiple diagnosis systems in Self-Healing networks David Palacios, Emil J. Khatib, Raquel Barco ∗ Communications Engineering Dept., University of Málaga, 29071, Málaga, Spain a r t i c l e i n f o Article history: Received 8 January 2016 Revised 20 July 2016 Accepted 21 July 2016 Available online 22 July 2016 Keywords: LTE Self-healing Root cause analysis Self-organizing networks (SON) Hybrid ensemble classifier Automatic fault identification a b s t r a c t The Self-Organizing Networks (SON) paradigm proposes a set of functions to automate network manage- ment in mobile communication networks. Within SON, the purpose of Self-Healing is to detect cells with service degradation, diagnose the fault cause that affects them, rapidly compensate the problem with the support of neighboring cells and repair the network by performing some recovery actions. The diagnosis phase can be designed as a classifier. In this context, hybrid ensembles of classifiers en- hance the diagnosis performance of expert systems of different kinds by combining their outputs. In this paper, a novel scheme of hybrid ensemble of classifiers is proposed as a two-step procedure: a modeling stage of the baseline classifiers and an application stage, when the combination of partial diagnoses is ac- tually performed. The use of statistical models of the baseline classifiers allows an immediate ensemble diagnosis without running and querying them individually, thus resulting in a very low computational cost in the execution stage. Results show that the performance of the proposed method compared to its standalone components is significantly better in terms of diagnosis error rate, using both simulated data and cases from a live LTE network. Furthermore, this method relies on concepts which are not linked to a particular mobile communication technology, allowing it to be applied either on well established cellular networks, like UMTS, or on recent and forthcoming technologies, like LTE-A and 5G. © 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/ ). S t t ( 1. Introduction The growing demand for mobile services with ever-increasing bandwidth and the expanding number of users make necessary the deployment of new and more efficient mobile communication networks over the existing ones (GSM, UMTS), such as Long-Term Evolution (LTE). However, the complexity of this heterogeneous scenario, which comprises several Radio Access Technologies (RAT), requires challenging maintenance and complex operational tasks. Mobile operators need to offer new demanding services without increasing either operational expenditures (OPEX) or capital expenditures (CAPEX). In order to deal with that problem, the 3rd Generation Partnership Project (3GPP) has proposed Self- Organizing Networks (SON) ( 3GPP (d) ) as networks that include mechanisms to automate network procedures in order to help mo- bile operators with their management work, providing significant cost reduction. This automation of network management will also be essential in near and future technologies, like LTE-Advanced and 5G ( 3GPP (b) ). ∗ Corresponding author. Fax: +34952132027. E-mail addresses: dpc@ic.uma.es (D. Palacios), emil@uma.es (E.J. Khatib), rbarco@uma.es (R. Barco). http://dx.doi.org/10.1016/j.eswa.2016.07.030 0957-4174/© 2016 The Authors. Published by Elsevier Ltd. This is an open access article u SON comprises three groups of functions: Self-Configuration, elf-Optimization and Self-Healing. The aim of the latter is to au- onomously solve the problems that a cell, with service degrada- ion or outage, could present ( 3GPP (e) ; Barco, Lázaro, and Muñoz 2012) ). This is done by means of four stages: • Fault Detection: Responsible for finding cells with problems, i.e., cells experiencing service outage or just suffering an unaccept- able service degradation. • Diagnosis of the fault cause: In this step, the actions to be per- formed in order to recover the system from the degradation it is suffering are decided. This step can be divided into two sub- stages: Fault Identification, this is, identifying the fault cause based on observable symptoms such as Key Performance Indi- cators (KPI) and alarms; and Action Identification, which corre- sponds to the decision of what tasks to perform to recover the system normal performance. • Fault recovery: In this step, the proposed solutions are carried out. • Fault compensation: Since diagnosing the fault and repairing it normally takes some time, compensation aims to diminish the impact of the fault by changing parameters in neighboring cells. nder the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/ ). http://dx.doi.org/10.1016/j.eswa.2016.07.030 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2016.07.030&domain=pdf http://creativecommons.org/licenses/by-nc-nd/4.0/ mailto:dpc@ic.uma.es mailto:emil@uma.es mailto:rbarco@uma.es http://dx.doi.org/10.1016/j.eswa.2016.07.030 http://creativecommons.org/licenses/by-nc-nd/4.0/ D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 57 f h c m n L p N w p K f r O t s a e a b m w i b n b t e b u C 2 m i t T t m p n t m d e c o a d n ( 2 n t s a t t t Fig. 1. Scheme of an automatic diagnosis system. l c a a l 2 2 o w T i o s ( t c t t i t fi t t n a w o This paper is focused on the diagnosis task, in particular in the ault identification, also called root cause analysis. Once a problem as been detected in a cell, root cause analysis identifies the fault ause given the value of performance indicators, alarms, counters, obile traces, etc. In the context of cellular networks, some diag- osis systems have been recently proposed. Barco, Díez, Wille, and ázaro (2009) and Barco, Lázaro, Wille, Díez, and Patel (2009) pro- osed diagnosis systems based on Bayesian Networks. Szilágyi and ováczki (2012) used a scoring system in order to determine how ell a specific case fits a diagnosis. Nováczki (2013) enhanced the revious system by adding profiling techniques. The method in hatib, Barco, Gómez-Andrades, and Serrano (2015) was based on uzzy logic and genetic algorithms. Gómez-Andrades, Muñoz, Ser- ano, and Barco (2016) proposed a diagnosis system based on Self- rganized Maps (SOM). Each of the previous methods has its pros and its cons. In prac- ice, this makes the selection of the diagnosis technique cumber- ome when the aim is to deploy a automatic diagnosis system in real network. Furthermore, once the technique has been decided, .g., fuzzy logic, operators normally design several standalone di- gnosis models. This is due to the fact that, firstly, different trou- leshooting experts will build different models and secondly, when odels are learnt from historical cases, different training datasets ill result in different models. To cope with the limitations of standard classifying systems n terms of accuracy and dataset-dependent performance, ensem- les of classifiers arose. Within these, homogeneous and heteroge- eous (commonly known as hybrid) ensembles of classifiers may e found, where the former stand for the ensemble of classifiers of he same kind and the latter stand for the combination of differ- nt kinds of systems and datasets. Despite homogeneous ensem- les have been widely studied and as of today still are extensively sed in different fields ( Begum, Chakraborty, & Sarkar, 2015; Liu, hen, Song, & Han, 20 09; Shen & Chou, 20 06; Wiezbicki & Ribeiro, 016 ). In this paper, a method for the generalized combination of ultiple diagnosis systems based on a hybrid ensemble approach s proposed and tested in the context of cellular networks, which o the authors’ knowledge is a research area still to be explored. he proposed work describes a method to gather, combine and use he knowledge held by any kind of expert system in any field that akes use of a classifying or diagnosis system. In this work, the roposed method is applied in the fault cause diagnosis in cellular etworks, where the expertise may be provided either by a human roubleshooting expert or by a database of cases assessed by auto- atic diagnosis systems. The proposed method allows combining iagnosis systems in a wide sense, being able to merge both sev- ral diagnosis models (expertise) and the tools used for their appli- ation (automatic diagnosis techniques) in the form of supervised r unsupervised classifying systems. Up to now, hybrid ensembles of classifiers are mainly based on set of baseline systems which must first assess the cases un- er test and, consequently, provide partial diagnoses which are fi- ally combined into a final decision using a majority vote scheme Ciocarlie, Lindqvist, Nováczki, & Sanneck, 2013; Gandhi & Pandey, 015; Wei et al., 2014 ). This procedure requires a relatively high umber of diagnosis techniques to be run in the test stage and, herefore, a noticeable expenditure of computational and time re- ources. The proposed work, however, presents a method which llows combining the diagnoses that the standalone diagnosis sys- ems would output for a case under test without actually needing hem to be run, thus lightening the computational weight of the est stage. The main contributions of this paper are: • A method to combine any number and kind of different stan- dalone classifiers as well as different sources of expert knowl- edge in order to get an enhanced performance compared to that of the base classifiers. In the context of troubleshooting in cellular networks this comprises the combination of several di- agnosis models and techniques for the automatic diagnosis. • A method to lighten the computational cost of the evaluation stage in hybrid ensembles of classifiers. This work proposes a scheme to model and emulate the behavior of every standalone classifier so these need not to be continuously queried before combining their partial diagnoses. This paper is organized as follows. Section 2 presents the prob- em formulation. Section 3 introduces the proposed method for ombining multiple baseline diagnosis systems. In Section 4 results re analyzed by means of both a network simulator and data from live LTE network. In Section 5 the future lines of work are out- ined. Finally, Section 6 summarizes the main conclusions. . Problem formulation .1. Root cause analysis in mobile communications networks In the same way that a patient is diagnosed by a doctor based n the symptoms he shows, the status of a communications net- ork may be diagnosed based on a set of performance indicators. his diagnosis task, also called root cause analysis or troubleshoot- ng, is often carried out by human experts using their knowledge n the underlying relations that the observed indicators and the tatus of the network have. However, the number of symptoms counters, alarms, KPIs, call traces, etc.) and possible fault causes he expert has to deal with increases as networks grow in size and omplexity, which makes this task to become a very difficult and ime consuming issue. Furthermore, the current manual troubleshooting is a layered ask, guided by a Trouble Ticket (TT) system. In this problem solv- ng system, a group of specialists tries first to diagnose and solve he problem by performing some simple checks. If they can not nd the root of the problem, this is raised to a more specialized eam (and so on), which performs a deeper study on the symp- oms the case exhibits and resorts to field engineers in case they eed to make some on site checks. As a response to this more and more inefficient procedure, utomatic diagnosis systems arose in an attempt of imitating the ay of acting of troubleshooters. Fig. 1 shows the basic scheme f a system for automatic diagnosis. It is composed of an au- 58 D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 t m s p a a n 2 p i m e k i p s o d p t p t w e s s c m a t s s 1 l f t m e e p f p n w c s B V f h a o s i v k l e tomatic diagnosis technique and a diagnosis model. The first is an artificial intelligence system that outputs a diagnosis taking a set of symptoms, e.g., (KPIs) from a test case as its input. The second represents the knowledge a human expert would have on the underlying relations between the symptoms and the fault causes and may take different forms depending on the diagnosis technique it is destined to work with. For example, a diagnosis model may consist of the parameters (e.g., prior probabilities and probability density functions) required by a given diagnosis technique (e.g., bayesian classifier) or a set of rules for other techniques (e.g., Case Base Reasoning, CBR). As it can be seen in this figure, the diagnosis model may be built from a set of training cases by means of a machine learning algorithm or by trou- bleshooting experts by gathering their knowledge. The proposed method aims to combine the knowledge acquired by any number and kind of diagnosis models and automatic diagnosis tech- niques in an attempt to reduce the errors in fault detection and diagnosis. 2.2. Automated diagnosis from the classification theory A diagnosis system is a method that given a set of indicators or symptoms (called case hereafter) intends to infer the cause that provoked them. In this sense, a diagnosis system acts as a classi- fying system in which the attributes from the cases to be classi- fied correspond to the symptoms from the case to be diagnosed, and the classes to be assigned correspond to the causes to be in- ferred. This is an issue long time investigated in data mining the- ory ( Wu et al., 2008 ), and many types of classifiers have been de- veloped over the years in an attempt to get the maximum infor- mation the cases under diagnosis could provide. However, no algo- rithm has proven to be clearly better than the rest for all kinds of input data by now. One reason for the increasing effort s in the re- lated research is that the performance of a classifier normally de- pends on the nature and distribution of the data it has to work with. For this reason, the present paper focuses not only on com- bining different diagnosis models but on offering the possibility to combine multiple classifiers in the form of automatic diagnosis techniques. Let us assume we have a set of M fault causes to diagnose and R diagnosis systems (either diagnosis model or technique) to com- bine, and that each of these systems can have a subset of these causes as their output, namely, W r for the system r . In this sce- nario, the set of causes a diagnosis system can identify may be dif- ferent from one system to another. This can be seen in (1) , where each row stands for a W r and the element w r m stands for the m th fault cause, diagnosed by the r th system. According to this, each row may be different from another. ⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣ w 1 1 . . . w 1 m · · · w 1 M . . . . . . . . . . . . . . . w r 1 . . . w r m · · · w r M . . . . . . . . . . . . . . . w R 1 . . . w R m · · · w R M ⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦ (1) In a diagnosis system, a case, x , is characterized by its symp- toms, x n , where x = { x 1 , x 2 , . . . , x N } , having a total of N pos- sible symptoms. However, each diagnosis system may consider only a subset of these symptoms, namely, N r for the diagnosis system r . In the context of diagnosis systems for mobile communica- tion networks a case corresponds to an observation or measure- ment from the network; a symptom may be an event counter, a Key Performance Indicator (KPI), a call trace or an alarm and he causes are seen as the network states, among which the nor- al and several fault states may be distinguished. In this paper, ome results from theory of classifiers is used, extended and ap- lied in this context in an attempt of combining the knowledge cquired by these R diagnosis systems, developing a more reli- ble and accurate root cause analysis system for communication etworks. .3. State-of-the-art in ensemble-based classification algorithms This section aims to provide a brief survey on the most recently roposed ensemble-based systems, most of which have been used n classifying tasks in areas not related to mobile communications. Ensembles of classifiers may nowadays be classified into ho- ogeneous and heterogeneous or hybrid. The first stand for those nsembles which put together instances of classifiers of the same ind, e.g., several k-Nearest Neighbor (kNN) classifiers. Conversely, n heterogeneous ensembles a set of classifiers of different kind are ut together, e.g., a kNN and a NN (Neural Network). This is the cope of the present work, as the latter also allow the combination f different sources of expert knowledge within a single enhanced iagnosis system. One of the earliest works on ensemble methods proposed to artition the feature space (i.e., the vector space in which the fea- ures of the cases to be diagnosed are defined) and to assign each art to a different classifier which is supposed to be the best for his subset of cases ( Dasarathy & Sheela, 1979 ). This idea has been idely explored and has given birth to the so-called mixture of xperts algorithm ( Jacobs, Jordan, Nowlan, & Hinton, 1991; Yuk- el, Wilson, & Gader, 2012 ), being the paradigm for the classifier- election type of ensemble methods. Under this approach, only one lassifier is working at the same time and its selection is deter- ined by the partition the case under test belongs. Conversely, in classifier-fusion methods all classifiers are usu- lly trained over the entire feature space. The classifier combina- ion process involves merging the individual classifiers to obtain a ystem that outperforms the standalone classifiers. This is the ba- is for the widely used bagging and boosting predictors ( Breiman, 996; Freund & Schapire, 1997 ), being AdaBoost an example of the atter and one of the most known and used algorithms for classi- ying nowadays. Classifier fusion methods can also be divided into hose which work with classification labels only and those which ake use of a continuous valued output for each classifier for ev- ry class. In this case, the outputs can be seen as the support an xpert gives to a class in terms of the class-conditional posterior robabilities ( Kuncheva, 2002 ). Some examples of ensemble methods as enhanced systems for ault disclosure can be found in the literature with many different urposes. In Liu et al. (2009) , an homogeneous ensemble of neural etworks with cross-validation for fault diagnosis of analog circuits ith tolerance is proposed. In Shen and Chou (2006) , several kNN lassifiers are put together on a majority-vote ensemble to clas- ify the patterns that several proteins may exhibit when folded. In egum et al. (2015) , an homogeneous ensemble of SVM (Support ector Machine) is proposed to identify different types of cancer rom a genetic analysis. Wiezbicki and Ribeiro (2016) proposes an omogeneous ensemble of neural networks, combined by means of weighted majority vote in a sensor network for the classification f gases. Regarding the most recent works on hybrid ensembles of clas- ifiers, in Wei et al. (2014) n ensembles are made up by combin- ng 3 n baseline classifiers. Each ensemble comprises three super- ised methods: a decision tree, a support vector machine and a NN algorithm. In each ensemble, the diagnoses from these base- ine classifiers are fused applying a weighted majority vote, where ach vote is weighted by the performance each individual classifier D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 59 Fig. 2. Proposed method for combining diagnosis systems. Stage 1: Construction of the behavior models. s a j v k t a m w s o m b s f c p c c b t c m p p d f g p c A a e t e n e t s ( c q W c v d 3 q s p b a a T a t d u 3 m b ( t n c i c T f I r d w c A u f i t a c n hows during a prior training stage. Then, the n resulting diagnoses re combined into a final diagnosis applying a non-weighted ma- ority vote. In this case, all the baseline classifiers must be super- ised diagnosis systems, as their performance must be previously nown in order to weigh their votes in the first stage. Unlike this, he proposed method allows the user to combine any kind of di- gnosis system, either supervised or unsupervised ones. And even ore important, regarding the operation stage, in Wei et al. (2014) , henever a new case is to be diagnosed it must pass through two teps, one of them made up of 3 n systems which must first each utput a diagnosis, resulting in a high computational cost. The ethod in the proposed work, however, needs the test cases to e assessed only by one step, which, furthermore, only consist of ome algebraic calculations. Once the training stage has been per- ormed, new cases will be diagnosed at a minimum computational ost. As for Gandhi and Pandey (2015) , a two-step method is again roposed. The first step consists of a learning stage for the base lassifiers and the second step consists of a majority vote-based ombining stage. Again and similar to Wei et al. (2014) , every aseline classifier is required to first diagnose every new case in he application step, which results in a high computational cost ompared to that from the application (test) step in the proposed ethod. In the context of cellular networks, Ciocarlie et al. (2013) pro- oses a hybrid ensemble of classifiers to detect anomalies in the erformance indicators of a cell. This work is focused on the fault etection. Unlike this, the proposed work does not just find a per- ormance degradation, but identifies the fault cause behind it. Re- arding its implementation, this method relies on the use of a ool of models. New models are added to this pool whenever a hange in the configuration parameters of the network takes place. number of N CM × ( N uni v ariate × N U KPI + N multi v ariate × N G KPI ) models, nd thus, instances of automatic techniques must be assessed for very single new case under test. In this expression, N CM stands for he number of sets of network configuration parameters consid- red; N U KPI and N univariate stand for the number of univariate tech- iques considered and the number of KPIs acting as their input in ach model; and N multivariate and N G KPI stand for the number of mul- ivariate techniques used and the number of groups of KPIs con- idered in each model. Like in Ciocarlie et al. (2013) , Wei et al. 2014) and Gandhi and Pandey (2015) , before an ensemble decision an be made, a high number of baseline classifiers must be first ueried. And again similarly to Ciocarlie et al. (2013) , according to ei et al. (2014) and Gandhi and Pandey (2015) , all the partial de- isions meet at a combining stage based in a weighted majority ote. t To the authors’ knowledge, no ensemble method for fault cause iagnosis in cellular networks has been proposed as of today. . Method for combining multiple automatic diagnosis systems In this section, a method for combining the knowledge ac- uired by any number and kind of standalone automatic diagnosis ystems by means of a classifier-fusion scheme is proposed. The roposed method consists of two stages: the construction of the ehavior models of the automatic diagnosis systems, Section 3.1 , nd the combination of these models in order to make a more ccurate diagnosis on the cases from a testing set, Section 3.2 . his can be seen in Figs. 2 and 5 . Before this method can be pplied, two sets of N -dimensional cases must be distinguished: he modeling set and the testing set, where each of these N imensions stands for a working KPI. The modeling set will be sed in the first stage and the testing set in the second. .1. Construction of the behavior models The baseline diagnosis systems are to be combined by means of ixing their models of behavior, which need to be extracted first. Once the diagnosis model from each diagnosis system has been uilt (either from training cases via a machine learning method Khatib, Barco, Gómez-Andrades, Muñoz, & Serrano, 2015 ), or from he experts’ knowledge ( Gómez-Andrades et al., 2016 ) each diag- osis system can start the classification (Fig. 1 ). In this stage, every ase from the modeling set is diagnosed by the R systems. That s, each system assigns to each case one of the M possible fault auses; in particular, one of the causes that system can discern. his can be seen in Fig. 2 , where the case x acts as the input or the R systems and, in turn, they assign it R diagnosis labels. f the system r diagnoses the case x with the cause m , this case eceives the label w r∗m . In this way, each diagnosis system makes a ifferent partition of the modeling set into | W r ∗| disjoint subsets, hose maximum is | W r |, that is, the number of causes that system onsiders (Fig. 3 ), where | A | is the number of elements in the set . This leads to finally identify M ∗ different causes, being M ∗ the nion of W r ∗ over r , with M ∗ ≤ M . According to this, a new matrix rom (1) may be written, substituting every row (i.e., every W r ) by ts corresponding W r ∗. Each row would represent one of the parti- ions of the modeling set and each column would represent how cause “is seen” by each diagnosis system regarding the KPIs the ases belonging to that w r m exhibit. It should be noticed that each of these M ∗ subsets contains a umber of | N r |-dimensional cases. At this point, the behavior of he diagnosis system r is modeled through the estimation of the 60 D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 Fig. 3. Modeling set divided into different subsets by means of two different par- titions: on the left, the partition the first diagnosis system makes, having W 1 = { w 1 1 , w 1 2 , w 1 3 } with | W 1 | = | W 1 ∗| = 3 ; on the right, the partition the diagnosis sys- tem R makes, having W R = { w R 1 , w R 2 , w R 3 } and W R ∗ = { w R ∗ 1 , w R ∗ 2 } . In this last case, the diagnosis system R only diagnosed the causes 1 and 2 although being able of also identifying the fault cause 3. Table 1 Families of PDFs considered for the estimation of p(x n | w r∗m ) . Distribution PDF Parameters Beta �(a + b) �(a )�(b) x a −1 (1 − x ) b−1 a, b Normal 1 σ √ 2 π exp ( − 1 2 ( x −μ σ )2 ) μ, σ Log-normal 1 xσ √ 2 π exp ( − 1 2 ( ln (x −μ) σ )2 ) μ, σ Exponential λ exp (−λx ) λ Gen. extreme value 1 σ t (x ) ξ +1 exp (−t (x )) , μ, σ , ξ t (x ) = {( 1 + ( x −μ σ ) ξ )− 1 ξ ξ � = 0 exp (−(x − μ) /σ ) ξ = 0 T-location �( ν+1 2 ) �( ν 2 ) √ π νσ ( 1 + 1 ν ( x −μ σ )2 )− ν+1 2 ν, μ, σ Nakagami 2 m m �(m )�m x 2 m −1 exp ( − m � x 2 ) m, � Gamma 1 �(k ) θ k x k −1 exp ( − x θ ) k, θ Logistic exp ( x −μs ) s ( 1+ exp ( − x −μs ) ) 2 μ, s Log-logistic (β/α)(x/α) β−1 ( 1+(x/αβ ) ) 2 α, β Weibull k λ ( x λ )k −1 exp ( − x λ )k λ, k Rayleigh x σ 2 exp ( − 1 2 ( x σ )2 ) σ Rice x σ 2 exp ( − x 2 + ν2 2 σ 2 ) I 0 ( xν σ 2 ) ν, σ K & t p b t β w t w h c r 3 s a K b l o t j a P s ( r P a ( p f fi i a b p f f f P w g D I ( t i statistical distributions of the N r KPIs for the cases belonging to W r ∗. That is, the behavior of each diagnosis system is modeled by means of N r × M ∗ PDFs. The estimated statistical distribution of the n th KPI for the subset of cases diagnosed as m by the diag- nosis system r is p(x n | w r∗m ) . The choice of the PDF that estimates each one of these distributions is done according to the maximum likelihood (ML) criterion. To do so, some families of PDFs are con- sidered in the fitting procedure (Table 1 ). In a first step, the distri- bution of the KPI x n from the cases labeled as w r∗ m is fitted attend- ing to the ML criterion with each one of the considered families of PDFs. This results in a set of candidates for estimating its dis- tribution. These PDFs are then sorted by their likelihood and the one with the maximum value is chosen to be the estimation for the KPI. The reason for considering these families of PDFs is to get the better estimation of the distribution of the KPI x n given its belong- ing to w r∗m . Fig. 4 a shows a normalized histogram of the KPI “95th percentile RSRP” from the cases labeled as w r∗m . In this figure, two families of PDFs have been used in an attempt of fitting the under- lying histogram, the normal and the generalized extreme value. As it can be seen, the latter fits it better, resulting in a higher value in a likelihood-ratio test. While some KPIs are counters and they do not have an up- per limit, there are others that are inherently bounded, as they are defined as a ratio. Normally, the beta PDF is used to fit these PIs, usually limited between zero and one ( Barco, Lazaro, Diez, Wille, 2008 ). KPIs like the retainability or the accessibility of- en reach these extreme values making the resulting fitted beta resent asymptotes in these values. To avoid this issue the used eta function β′ is slightly different from that from Table 1 , β. In his case, ′ (x ) = (1 − P 0 − P 1 ) β(x ) + P 0 /h β δ(x ) + P 1 /h β δ(x − 1) , (2) here β( x ) stands for the distribution fitted to a set with no ex- reme values; P 0 and P 1 stand for the relative frequency of cases ith value 0 and 1 respectively; δ stands for the Dirac’s delta and β stands for the step (the resolution) when computing β ′ . This an be seen in Fig. 4 b, where a normalized histogram for the KPI etainability is shown. .2. Combination of behavior models This stage uses the cases from the testing set. In the previous tage, the estimated functions have been seen as conditional prob- bility density functions, that is, functions that express how the PIs are distributed over the cases diagnosed with a given cause y a given system. However, this set of functions may be seen as ikelihood functions by just changing the approach. From this point f view, the function depends on w r∗m given that an observation of he random variable x n (that is, the n th KPI) has taken place. Now, assuming the KPIs are independent among each other, a oint probability function of w r∗m , that is, p( x | w r∗m ) , may be written s p( x | w r∗m ) = ∏ n ∈ N r p(x n | w r∗m ) . (3) Given (3) , and assuming that the prior probability of each cause, (w r∗m ) is given by | w r∗m | | W r∗| , the a posteriori probability for a diagnosis ystem r to diagnose a case with the cause m given its KPIs are x i.e., P (w r m | x ) ) can be calculated by just applying the Bayes’ theo- em. That is, (w r m | x ) = ⎧ ⎨ ⎩ p( x | w r∗m ) P (w r∗m ) ∑ w r∗ i ∈ W r∗ p( x | w r∗i ) P (w r∗i ) i f P (w r∗m ) > 0 0 i f P (w r∗m ) = 0 (4) At this point, some diagnosis system may have not diagnosed given cause as seen in Fig. 3 . In such case, P (w r∗m ) and thus 4) would result equal to zero. In any case, M × R a posteriori robabilities may be distinguished. Fig. 5 shows this when a case y rom the testing subset is to be diagnosed. As it can be seen in this gure, the KPIs from the case y act as input values in the behav- or models of the R diagnosis systems, i.e., the probability functions p( y | w r∗m ) for w r∗m ∈ W r∗ and r = 1 , . . . , R . Then, the a posteriori prob- bilities P (w r m | y ) are computed using these together with P (w r∗m ) y means of the Bayes’ theorem. Now, these M × R a posteriori probabilities together with the rior probabilities can be combined over R using some algebraic unctions, producing M probabilities of the kind P (w m | y ) Rule t per unction used, where m again stands for the cause and t is an index or the rule used in the combination, that is, (w m | y ) Rule t = f Rule t ( P (w 1 m | y ) , . . . , P (w R m | y ) ; P (w m ) ) . (5) here P ( w m ) is defined as the average of P (w r∗ m ) over r . Some rules for the combination of a posteriori probabilities iven by several classifying systems are proposed in Kittler, Hatef, uin, and Matas (1998) and studied further in Kuncheva (2002) . n the first, those rules are derived from a maximum a posteriori MAP) estimation in a multiple random variable scenario in an at- empt of lightening the effort s of computing several joint probabil- ty density functions. These rules are summarized in Table 2 . D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 61 Fig. 4. (a) Normalized histogram for the KPI 95th percentile RSRP and two fitted PDFs: a generalized extreme value PDF in blue (round markers) and a normal PDF in red (square markers). (b), Normalized histogram for the KPI Retainability and a β′ PDF estimation. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Fig. 5. Proposed method for combining diagnosis systems. Stage 2: Combining the behavior models. p n d o t c 4 i i c i c b As this point, the fault cause with the maximum a posteriori robability is taken as the final diagnosis per each rule of combi- ation, d t . That is, Rule t = arg max m { P (w m | y ) Rule t } . (6) Note that a situation with M ∗ < M means that there is at least ne fault cause that have not been identified by any system. In his case, it would be impossible for it to be finally diagnosed in onsequence. . Proof of concept In this section, the proposed method is assessed by combin- ng two different diagnosis models. In the first test, each model s provided by a different expert; in the second test, each model omes from using different machine learning algorithms for build- ng the diagnosis models, provided the same set of training ases. The proposed method has been evaluated and compared to the aseline systems by means of the following figures of merit: 62 D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 Table 2 Algebraic rules for the combination of a posteriori probabilities. Rule P(w m | y ) Product rule P (w m ) −(R −1) R ∏ r=1 P (w r m | y ) Sum rule (1 − R ) P(w m ) + R ∑ r=1 P(w r m | y ) Max rule (1 − R ) P(w m ) + R R max r=1 { P(w r m | y ) } Min rule P (w m ) −(R −1) R min r=1 { P (w r m | y ) } Median rule R med r=1 { P(w r m | y ) } o t t G T t b p f t D e • Diagnosis Error Rate (DER): it is the ratio of problematic cases diagnosed as a fault cause different to the real one (misclassi- fied cases), N MPC , to the total number of problematic cases, N PC . • False Positive Rate (FPR): it is the number of normal cases di- agnosed as problematic cases, ( N FP ), to the total number of nor- mal cases, ( N NC ). • False Negative Rate (FNR): it is the number of problematic cases diagnosed as normal cases, N FN , to the total number of prob- lematic cases, N PC . This is the most critical metric, as it gives an idea on how often the diagnosis system interprets there is no problem when actually some cells are suffering from mal- functioning. Given these definitions, an Overall Error Rate (OER) may be de- fined as OER = P N · F P R + P PR · (F NR + DER ) (7) where P N stands for the relative frequency of the normal cases and P PC stands for the relative frequency of the faulty cases. This metric is useful to assess every method at a single glance. Since these fig- ures of merit require the true cause to be known, the used testing set will include the real diagnosis. 4.1. Combination of diagnosis models devised by multiple experts 4.1.1. Scenario In this test, cases are provided by an LTE RAN simulator ( Muñoz et al., 2011 ). This simulator considers an LTE network composed Table 3 Simulation parameters for cells normal functioning. Parameter Configuration Cellular layout Hexagonal grid, 57 cells, cell radius Transmission direction Downlink Carrier frequency 2.0 GHz System bandwidth 1.4 MHz, 6 PRB (Physical Resource B Frequency reuse 1 Propagation model Okumura-Hata with wrap-around, L σs f = 8 dB and correlation distance Channel model Multipath fading, ETU model Mobility model Random direction, 3 kph Service model Full Buffer, Poisson traffic arrival Base station model Tri-sectorized antenna, SISO, P T X max = Azimuth beamwidth = 70 °, Elevati Scheduler Time domain: Round-Robin, Frequen Power control Equal transmit power per PRB Link Adaptation Fast, CQI (Channel Quality Indicator Handover Triggering event = A3, HOM (Hando Measurement type = RSRP Radio Link Failure SINR < −6.9 dB for 500 ms, Mehlfü Traffic distribution Evenly distributed in space Time resolution 100 TTI (Transmission Time Interval Epoch & KPI time 100 s f 57 macro-cells evenly distributed in space and grouped into 19 hree-sector-sites. To perform this test, similar network configura- ion parameters to those used in Gómez-Andrades et al. (2015) and ómez-Andrades et al. (2016) have been used. They can be seen in able 3 . With this simulator, 1196 cases have been obtained. In this case, raining cases are not needed since the diagnosis models have een defined by experts. It is assumed that a detection system is laced before the input of the diagnosis system, so that only the aulty cases are put under test, putting aside the cases belonging o a normal cause of functioning. Therefore, in this test only the ER is taken into account. In this scenario, six typical RAN fault causes have been consid- red ( M = 6 ): • Excessive downtilt: This situation takes place when the coverage area for a cell is too small, making the signal level in the edge of the cell to be too weak and causing a high number of han- dover failures. The quality of the signal in the surroundings of the cell is also decreased. • Coverage hole: A cell has a coverage hole in some point inside its area when the power received by the user at this point from any cell is not enough to hold the service. This excessive atten- uation can be caused by either obstacles or a bad RF planning and it mainly produces a high number of call drops. • Inter-system interference: This fault cause may occur due to other cellular networks, like WCDMA. It is not always an easy issue to solve, since the fault usually comes from an outer sys- tem. This fault normally causes both the SINR and the average throughput decrease. • Too late handover: A too late handover takes place if a radio link failure occurs while the UE (User Equipment) is moving from one cell to another and the corresponding handover between these cells has not taken place yet. In that case, the UE will request the second cell a connection re-establishment using the physical cell ID of the first cell and its Common Radio Network Temporary Identifier (C-RNTI) in that first cell, which will alert the second cell a too late handover has occurred. • Excessive uptilt: A cell suffers from excessive uptilt when its cov- erage area is larger than necessary, normally because of a bad configuration of the radiation parameters of the antennas. This situation can result in the overlapping of coverage areas from 0.5 km lock) og-normal slow fading, = 50 m 43 dBm, Downtilt = 9 ° on beamwidth = 10 ° cy domain: Best Channel ) based, perfect estimation ver Margin) = 3 dB, hrer, Wrulich, Colom Ikuno, Bosanska, and Rupp (2009) ) (100 ms) D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 63 Table 4 Parameters used for modeling fault causes in Section 4.1 and a priori probabilities for each cause. Fault cause Configuration P ( ω m ) Excessive downtilt Downtilt = [16, 15, 14] ° 0 .18 Coverage hole �hole = [49, 50, 52, 53] dBm 0 .09 Inter-system interf. P T X max = 33 dBm 0 .1 Downtilt = 15 ° Azimuth beamwidth = [30, 60] ° Elevation beamwidth = 10 ° Too late HO HOM = [6, 7, 8] dBm 0 .23 Excessive uptilt Downtilt = [0, 1] ° 0 .21 Lack of coverage P T X max = [7, 8, 9, 10] dBm 0 .19 a c P s a u Table 5 Diagnosis models for the diagnosis sys- tems used in test 1: used thresholds. KPI Thresholds Retainability [0.973, 0.996] HOSR [0.899, 0.989] RSRP [dBm] [ −76 . 9 , −72 . 4] RSRQ [dB] [ −18 . 8 , −18 . 2] SINR [dB] [13, 14.5] Throughput [kbps] [96.2, 111.67] Distance [km] [0.838, 0.88] 4 d v i h W t k s f W o & p “ a f s l l c t e r i d s 6 4 M a T t possibly non-adjacent cells, producing a high number of han- dovers and call drops in this cell and its neighbors • Lack of coverage: A user suffers from weak coverage when the Signal-to-Interference-Plus-Noise Ratio (SINR) measured in the cell is below the minimum level needed to maintain a planned performance requirement because the received power is low. The simulation parameters used to model these degradations re shown in Table 4 , as well as the a priori probability of these auses to take place, given by the experts. In this case, P (w 1 ∗m ) = (w 2 ∗m ) � = 0 ∀ m, so P (w m ) = P (w 1 ∗m ) = P (w 2 ∗m ) . As it can be seen, everal values have been used for modeling a single fault cause, ccording to lighter and more severe degradation. In this test, seven observable features or KPIs ( N = 7 ) have been sed to discern among this set of causes: • Retainability , given as a percentage. This performance indicator quantifies the ability of the cell to hold the service once ac- cepted by the admission control. It gives an idea on how often a user experiences a call drop. • Handover success rate (HOSR) , given as a percentage. This KPI measures the ability of the network to provide mobility to a user without losing its connection. It can be calculated as the ratio between the number of successful handovers and the total number of HO. • 95th percentile RSRP , given in dBm. The Reference Signal Re- ceived Power (RSRP) is defined as the linear average over the power contributions (in [W]) of the resource elements that carry cell-specific reference signals within the considered mea- surement frequency bandwidth. • 5th percentile RSRQ , given in dB. The Reference Signal Received Quality (RSRQ) is a signal quality indicator and is defined as the ratio RSRQ = N PRB · RSRP RSSI , (8) where N PRB is the number of resource blocks of the E-UTRA car- rier RSSI measurement bandwidth and RSSI stands for the to- tal received power within the measurement bandwidth. This is, considering the power from the serving cell, the power of the co-channel serving and non-serving cells, the adjacent chan- nel interference and any possible source of noise. In this paper, RSRQ is expressed in dB. • 95th percentile SINR , given in dB. The Signal-to-Interference- plus-Noise Ratio (SINR) is defined as the ratio between the power of the desired data signal and the sum of the powers of all inter-cell interferences and the noise. It is expressed in dB. • 95th percentile distance , given in km. This KPI measures the dis- tance between users and their serving cell, expressed in km. It can be estimated attending to the transmission delay between them and gives an idea of the cell coverage area. • Average throughput , given in kbps. In LTE systems, the user throughput depends on the SINR experienced by the user through the following equation, 3GPP (c) , T k = (1 − BLER (SINR k )) · D k T T I , (9) where BLER is the Block Error Rate obtained from the users’ SINR, D k is the data block payload in bits of user k and TTI is the transmission time interval. In order to show the impact a proper modeling may have in the diagnosis performance of the proposed method the propor- tion of cases used for the modeling to the testing set has been varied from 25% to 75%. To obtain more reliable results when the number of cases are scarce either in the testing or in the modeling set, 50 repetitions have been made per modeling-to- testing ratio, randomizing the cases assigned to each set. Then, the resulting diagnosis error rates have been averaged over the 50 repetitions. .1.2. The standalone classifiers In this test, for a given technique of automatic diagnosis, two iagnosis models are combined, R = 2 , where each of them is pro- ided by a different expert. This test represents the usual case n cellular networks where each troubleshooting expert defines is own set of rules and KPI thresholds to identify problems. hen deploying the diagnosis system in a network, according to he proposed method, instead of choosing one single model, the nowledge from both experts is fused by combining two diagno- is models. Furthermore, both diagnosis models comprise the six ault causes and the seven different KPIs described above. That is, 1 = W 2 with | W 1 | = M and N 1 = N 2 . The artificial intelligence technique used for these tests is based n a Fuzzy Logic Controller (FLC) ( Khatib, Barco, Gómez-Andrades, Serrano, 2015 ). This system contains rules, which are com- osed of the antecedent (the “if ...” part) and the consequent (the then ...” part), being the last the cause the fuzzy logic controller ssigns to a case if the antecedent is fulfilled attending to the uzzyfied observable features of the case. On the one hand, Table 5 hows the thresholds used in both diagnosis models. The lower imit stands for the value below which a KPI is considered to be ow; the upper limits stands for the value above which a KPI is onsidered to be high. On the other hand, Table 6 shows the if ..., hen ... rules that make up each diagnosis model, given by each xpert. From left to right, each column below “KPI” in Table 6 cor- esponds to the KPIs shown in Table 5 . H stands for a high value n that KPI and L for a low value. Regarding the numbering of the iagnoses, 1 means excessive downtilt; 2: coverage hole; 3: inter- ystem interference; 4: too late handover; 5: excessive uptilt and : lack of coverage. .1.3. Results Table 7 shows the diagnosis error rates computed when the ax rule is used for combining ( Table 2 ). In Table 7 , the aver- ge diagnosis error rate and the rate of improvement are shown. his last rate represents the amount of repetitions (among the 50 hat have been performed) in which the diagnosis error rate from 64 D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 Table 6 Diagnosis models for the diagnosis systems used in test 1: used rules. Diagnosis model 1 Diagnosis model 2 KPI Diag. KPI Diag. L L H L – H L 1 – – H L – H L 1 H H – L L H L 1 L – H H – L H 2 L – – H H – H 2 L – H H H – H 2 L – – H L L H 3 L – – H L L H 3 L L H – L L H 3 L – H – L L H 3 – – H H H H – 4 L – H H L L – 3 – H H – H H – 4 L H – H L L – 3 H – H – H H – 4 – H H H L L H 3 – – H – H H H 4 L L – – – H H 4 H H – – H – H 4 L L – L – – H 4 H H – L – – H 4 L L H – H L – 4 H H – – – H H 4 L L H – H – H 4 H H H – – – H 4 L L L L L L – 4 – – H L H – H 4 – – L H – L H 5 – – H L – L H 4 H H – H – L H 5 L L H L – – H 4 H H L – L L H 5 – – L – L L H 5 – – – L L H L 6 – – L – L H L 6 L – – L L H L 6 – H – L L H L 6 H H – – L H L 6 L L L L L – L 6 Table 7 Results of test 1: Combining two versions of the same classifying algo- rithm. Modeling-to-testing ratio 25% 50% 75% Diagnosis syst. 1, average DER 13.81% 13.7% 13.65% Diagnosis syst. 2, average DER 16.34% 16.13% 16.3% Ens. Method: Max rule average DER 8.29% 5.92% 5.34% Rate of improvement 60% 98% 100% Table 8 Main parameters of the real LTE network used in test two. Parameter Configuration Network Layout Urban area Number of cells 8679 System bandwidth 10 MHz Number of PRBs 50 Frequency reuse factor 1 Max. Transmitted Power 46 dBm Max. Transmitted Power of UE 23 dBm Horizontal HPBW (Half-Power Beam Width) 65 ° HOM 3 dB KPI Time Period Hourly Number of observed cells 45 Number of days under observation 6 days per cell (on average) Size of the dataset 14 ,692 labeled cases w f b 4 p I v m d b a i t b 1 b k c the ensemble method is lower than the best one provided by the baseline diagnosis systems. With a 25% of modeling-to-testing ra- tio only 60% of the iterations shows a better ensemble diagnosis error rate than the ones from its base diagnosis systems, showing, therefore, little improvement in the average diagnosis error rate. This result highlights how the scarcity of cases for modeling im- pacts on the classifying performance of the ensemble. However, if the number of cases used for modeling is doubled 98% of the iter- ations shows a better diagnosis error rate, which results also in a lower average diagnosis error rate. In case the modeling-to-testing ratio is set to 75% every diagnosis error rate provided by the en- semble method is lower than the lowest provided by its compo- nents, reaching a 5.34% on average. This means a DER of approxi- mately 1/3 the lowest DER achieved by the standalone classifiers. Regarding the DER of the standalone diagnosis systems, it can be seen how these are held over the modeling-to-testing ratio. This is because of the randomizing process executed over the la- beled cases to be divided into the modeling and testing subsets. When this random permutation is performed a number of times and some subsets (two, in this case) are chosen blindly from this set, the averages of the amount of cases labeled with a given cause in each of these subsets tend to the ratios of the labels from the original set. This is a consequence of the law of large numbers. For this reason, the resulting averaged DER of these baseline systems is independent on the size of the subsets made from the original set of cases. 4.2. Combination of different diagnosis systems on a live network Once the proposed method has been tested with cases provided by a simulator, a second test with cases from a real live LTE net- ork has been performed. In this test, the diagnosis models built rom two different machine learning algorithms have been com- ined. .2.1. Scenario An LTE network composed of more than 80 0 0 different cells roviding coverage to almost 4 million people has been analyzed. ts vastness makes many different cells to coexist and also a wide ariety of problematic causes to come up. Table 8 summarizes the ain parameters of the network. Among all the available candi- ates, 45 random cells have been chosen to represent the network ehavior. These cells have been monitored for almost 6 days on verage and their KPIs have been stored in an hourly basis. Tak- ng into account that the state of a single cell varies substantially hroughout the day due to the traffic fluctuation, several cases have een stored from each cell at different hours, resulting in a total of 4,692 cases. Once these cases were gathered, they were all labeled y the experts, distinguishing four groups of cases ( M = 4 ): three inds of problematic patterns and the normal cell functioning. The auses of malfunctioning that were found are: • Overload : This fault cause is mainly distinguished by a high number of RRC connections in the cell, which makes the CPU processing load and the number of HO attempts raise conse- D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 65 Table 9 Prior probability of occurrence for the causes considered in test two, P ( w m ). P (Overload) P (Lack of cov.) P (Non-operating) P (Normal) 0 .01 0 .22 0 .47 0 .3 c T f p p p i t S w N 4 l c a d i & r d a e Table 10 Diagnosis models for the diagnosis systems used in test 2: used thresholds. KPI Thresholds Retainability [0.99, 0.997] Accessibility [0.992, 0.998] Number of RRC Connections [5846, 20703] Number of ping-pong HO [18, 83] Number of bad cov. reports [217, 1070] CPU average load [%] [22.5, 34.45] i w o g f t c p t m e n t s w i s s l h v s 1 n 4 p f i i A m T m t t f m b m g a n r n c a n i F d s quently. The accessibility and retainability KPIs also hold values quite below the ones for a cell with normal functioning. • Lack of coverage : This issue can be identified based on the num- ber of bad coverage evaluation reports, which should be notice- ably high. • Non-operating cell : In this case, and only if the cell is report- ing any KPI measurement, most of the reported measurements should be near zero: the retainability, the accessibility, the number of performed HO, the number of RRC connections or the number of coverage reports. The a priori probability of occurrence of each class has been omputed as the average of P (w r∗m ) over r within this selection, able 9 . From this table it should be noted that there are more aulty cases than healthy ones. This is because a previous non- erfect faulty cases detecting stage has been applied, which by- assed some normal cases that now are to be diagnosed as such. At this point, a 20% of the total number of cases (holding the roportion shown in Table 9 between them) were used as a train- ng set for the machine learning algorithms and the rest were used o conform the modeling and testing sets in a ratio that, as in ection 4.1.3 , was varied along the test. In this test six of the most representative KPIs in an LTE net- ork have been chosen to discern between the possible diagnoses, = 6 : • Retainability : described in Section 4.1.1 . • Accessibility : It is used to show the percentage of connections that have got access to that cell over the KPI time period. A low value in this KPIs means that many connections have been blocked during the access procedure. • Number of RRC connections : It is the number of successfully es- tablished RRC connections. Related to the Accessibility KPI, it gives an idea of the amount of users served by the cell. • Number of Ping-Pong Handovers : This KPI counts the number of ping-pong HO that takes place in the cell over the measure- ment time period. A high value in this KPI may mean a bad configuration in the handover policy, as the number of connec- tions that goes back and forth over a cell and its neighbors is high for a single call. • Number of bad coverage reports : It counts the number of times a cell is notified that the UE measured a signal level in which the requirements for the Event A2 takes place, 3GPP (a) . This is, the measured signal level is under a certain threshold. • CPU average load : It is the average CPU load due to the pro- cesses carried out by the cell over the KPI time period. .2.2. The standalone classifiers In this test, the two used standalone classifiers share a simi- ar diagnosis system, a fuzzy-logic controller, which diagnoses the ases attending to if . . . , then . . . rules. The difference resides in the lgorithms they use for learning the rules they apply during the iagnosis process. The first is a genetic algorithm and the second s a data driven algorithm ( Khatib, Barco, Gómez-Andrades, Muñoz, Serrano, 2015; Khatib, Barco, Gómez-Andrades, & Serrano, 2015 ) espectively. In genetic algorithms, three main processes may be istinguished: reproduction, by means of which new individuals re created by either mutation or combination of the previously xisting; evaluation, or the calculation of the probability of each ndividual to survive and reproduce, and selection, a process in hich some individuals are chosen to survive and reproduce based n the results from the evaluation stage. Likewise, data driven al- orithms first take a case from the training set and derives the uzzy rule that covers it. Then, it looks for the cases covered by his rule and scores the rule attending to the number of covered ases. New incoming cases are taken until the training set is com- letely explored. Provided this set of scored rules, the algorithm hen fuses them into a lower number of rules in a attempt of aximizing the number of cases (and therefore, the score) cov- red by the resulting fused rules. In these tests, it is assumed that ot only faulty cases, but also some normal cases are inputs for he diagnosis stage. This can happen when there is no detection ystem before the diagnosis system or in the realistic situation in hich the detection system has a given probability of error. As n Section 4.1.2 , both systems take as possible output all the pre- ented diagnoses making use of the six KPIs shown above. Table 10 hows the thresholds used for these KPIs to consider them high or ow and Table 11 shows the rules each machine learning algorithm as derived from the testing set. As in Table 6 , H stands for a high alue of the KPI and L, for a low value. The KPIs are sorted in the ame way as in Table 10 and the numbering of the diagnoses are : CPU overload; 2: lack of coverage; 3: non-operating cell and 4: ormal functioning. .2.3. Results Once the standalone diagnosis systems have been trained, the erformance metrics DER, FPR, FNR and OER have been computed or both the standalone diagnosis systems and the rules described n Table 2 . In this test, the modeling-to-testing ratio has been var- ed from 10% to 90% in steps of 10, making 10 iterations per step. s in Section 4.1.3 , a random permutation of the cases used for odeling and testing has been done in each of these 10 iterations. he resulting metrics have been then averaged. Table 12 shows the etrics that result of using a proportion of 60% in the modeling- o-testing ratio. This ratio has proved to minimize the values of all he metrics in this test. Unlike in Section 4.1 , this scenario is made rom real cases and contains outliers, that is, atypical cases. As the odeling-to-testing ratio rises, the probability for these outliers to elong to the modeling set also rises, thus inducing the behavior odels to deviate from modeling the trend of the typical cases iven a fault cause. On the other hand, if no outliers are taken into ccount during the model-fitting procedure their fault cause will ot be predictable in the second stage and the error rates will also ise up. As it can be seen in Table 12 , in most cases, the combined diag- osis system outperforms the standalone diagnosis systems. Con- retely, the median rule achieves the lowest overall error rate with 5.39%, approximately 2/3 from that of the best standalone diag- osis system. However, the most relevant improvement takes place n the reduction of the FNR, which has been reduced a 46%. The NR gives an idea of the amount of problematic causes wrongly eemed as normal. It is crucial making this metric as low as pos- ible, since considering a problematic case as normal may result 66 D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 Table 11 Diagnosis models for the diagnosis systems used in test 2: used rules. Diagnosis model 1: Diagnosis model 2: from genetic algorithm from data driven algorithm KPI Diagnosis KPI Diagnosis H H – – L L 1 H H L – L L 1 H – H H L L 1 H H – H L L 1 – H H L – L 2 H – H H L L 1 H – – L – H 2 – L H L – H 2 H – – L H – 2 – H H – H H 2 H – H L – – 2 – – H L H H 2 – L H L – H 2 L – H – H H 2 H H – – H – 2 H H H – H – 2 – H H – H – 2 H L – L L H 2 – – H L H – 2 H – H L – H 2 H – H – H H 2 H – H L H – 2 H – H – H H 2 – L L L L L 3 – L L L L L 3 L L L L H – 4 L L – – H H 4 – L L L H H 4 L L L – H – 4 L L L – H H 4 L L H L L L 4 L – L L H H 4 L L L – – H 4 L – L – H H 4 – – L L H H 4 L L L H – – 4 Table 12 Results of test 2: Combining two different algorithms. DER FPR FNR OER Training: Data driven algorithm 2 .62% 16 .91% 6 .47% 11 .43% Training: Genetic algorithm 1 .87% 16 .61% 2 .68% 8 .16% Ensemble method Product rule 2 .6% 12 .21% 1 .32% 6 .2% Sum rule 1 .78% 11 .55% 1 .25% 5 .59% Max rule 1 .78% 11 .51% 1 .25% 5 .57% Min rule 2 .05% 11 .42% 1 .4% 5 .84% Median rule 1 .78% 10 .67% 1 .34% 5 .39% Majority vote rule 1 .78% 11 .23% 1 .25% 5 .49% in the worst case in unnoticed service outages and degradation in the network performance. Regarding this, the proposed method has proved to successfully reduce the FNR. Other indicators are not as critical. For example, misleading a fault cause with another may be to some extent tolerable (DER); although the actual problem is not that one the operator thinks it is, he is still aware of a problem in the network. Even considering normal cases as faulty may be tolerable as the network performance is not really degraded (FPR). These results can also be seen in the normalized confusion matrices from the diagnosis methods. Fig. 6 a shows the normal- ized confusion matrix for the FLC using genetic algorithm for rule learning; Fig. 6 b shows the confusion matrix given the data driven algorithm was used for learning the rules and Fig. 6 c shows the matrix from applying the median rule with a 60% of modeling- to-testing ratio in the ensemble method. In these matrices, the elements from the fourth column (excluding the main diagonal) account for the false negatives and the elements from the fourth row account for the false negatives. It can be seen how the elements from the main diagonal are reinforced in the ensemble method and how only those diagnoses which are mistaken by both baseline systems are slightly inherited by the latter. Fig. 6 c also shows graphically how the FPR and FNR dropped with respect to those from the standalone systems. 5. Future lines of work • Decision templates . The proposed method does not punish or reward the classifiers according to its performance during the training stage. Going a step further from the idea of the weighted majority vote used in Wei et al. (2014) , a score system based on class and classifier aware decision templates applied over the a posteriori probabilities P (w r m | x ) from Eq. (4) could be used to improve the overall accuracy. • Non-parametric PDFs . The proposal of analytically and parameter-defined PDFs results in a really light way of representing a statistical behavior, as only its parameters must be stored Table 1 to model a diagnosis system. However, these distributions may limit to some extent the statistical representation of the features from the training cases and they may eventually introduce a source of error in the posterior computation of P (w r m ) in case these cases follow a distribu- tion that has not been considered. To solve this, the future research could focus on using non-parametric PDFs, like the kernel-based ones. Probability density functions may be classified into parametric and non-parametric functions. The former have analytic expres- sions and their shape depends on the parameters those func- tions hold. The latter, however, are defined by means of a ker- nel function. If all the cases from the dataset are placed along the axis given by a feature of interest (a certain KPI, for exam- ple) and a kernel function is centered wherever a point is, an empirical non-parametric PDF would result from averaging the sum of these functions over the number of cases. The main ad- vantage of this method is its accuracy when modeling an em- pirical distribution. Its main drawback is that, since it is not defined by any parameters, it should be computed and stored point by point, possibly increasing the storage and comput- ing requirements. This method, however, may be used together with (c). First, a reduced set of synthetic KPIs is computed and then, their PDFs are accurately estimated with this method. • Use of synthetic KPIs via feature extraction. As it is described in Section 3.1 , N r × M ∗ × R PDFs should be estimated in order to model all the feature-class-classifier relations. If any of these factors is relatively high, the computing cost for all these PDFs to be computed could be prohibitive. Due to this, working with a reduced group of synthetic/extracted features is proposed in an attempt of mapping the N original features into ˆ N synthetic features with ˆ N < N. D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 67 Fig. 6. Normalized confusion matrices for the second test. 6 k i l k t s t a o t j s e h u g b t t T i r c In the recent years and mainly motivated by the impulse of data mining many methods for dimensionality reduction have arisen. Within these, it is worth highlighting the Principal Component Analysis method (PCA) ( Jolliffe, 2002 ). In an N - dimensional vector space, the simplest version of PCA (linear PCA) is a technique that finds the mutually-uncorrelated vec- tors onto which the projection of the samples generates the highest variances. The result is a set of orthogonal vectors sorted in descending order of achieved variance. The first of these vectors is that onto which the variance of the projec- tion of the samples is maximum. In this sense, the original KPIs constitute the N -dimensional vector space basis, whereas the ˆ N synthetic KPIs represent the orthogonal vectors with the high- est variance. To be rigorous, up to N synthetic orthogonal KPIs may be computed. However, only a small set of them, the first ˆ N , is enough to account for most of the variance of the data. By applying this technique, based on the eigenvalue decompo- sition of the covariance matrix of the original KPIs, these can be mapped into ˆ N , preserving most of the information contained in the former. . Conclusions A hybrid ensemble of classifiers, devised to merge expert nowledge from different sources has been presented and assessed n the context of fault cause diagnosis in cellular networks, al- owing the expertise from several troubleshooting experts and the nowledge contained in databases of cases previously diagnosed o be combined in order to develop a more accurate diagnosis ystem. Unlike the common approach of hybrid ensembles, based on he majority vote of their baseline components, this work proposes hybrid ensemble of classifiers obtained from the combination f the statistical behavior models of the baseline diagnosis sys- ems. This approach allows obtaining and afterwards combining by ust applying some algebraic rules the partial diagnoses from the tandalone classifiers without actually needing them to assess ev- ry case under test, thus reducing the computational cost of usual ybrid ensembles of classifiers. The method has been tested with two different sources of cases nder test: cases provided by an LTE RAN simulator and cases athered from a real live LTE network. Likewise, two use cases have een assessed: the combination of diagnosis models designed by wo different network troubleshooting experts and the combina- ion of two diagnosis systems using different learning algorithms. he proposed method has proved to outperform the behavior of ts base components in both tests in terms of the diagnosis error ate, proving to be an effective tool in the fault cause diagnosis in urrent and future self-healing networks. 68 D. Palacios et al. / Expert Systems With Applications 64 (2016) 56–68 G J J K K K L M N S W W W Y Acknowledgment This work has been partially funded by Optimi-Ericsson, Junta de Andalucía (Consejería de Ciencia, Innovación y Empresa, Ref. 59288 and Proyecto de Investigación de Excelencia P12-TIC-2905) and ERDF. References 3GPP (a). Evolved Universal Terrestrial Radio Access (E-UTRA) Radio Resource Con- trol (RRC); Protocol Specification, Rel-13, Version 13.2.0, (2015-12). TS 36.331. 3rd Generation Partnership Project. 3GPP (b). Feasibility study for Further Advancements for E-UTRA (LTE-Advanced), Rel-13, Version 13.0.0 (2015–12). TR 36.912. 3rd Generation Partnership Project. 3GPP (c) (May 2004). OFDM-HSDPA System level simulator calibration (R1-040500). 3GPP TSG-RAN WG1 37 . 3rd Generation Partnership Project (3GPP) . 3GPP (d). Self-Organizing Networks (SON); Concepts and requirements, Rel-13, Ver- sion 13.0.0 (2015–12). TS 32.500. 3rd Generation Partnership Project. 3GPP (e). Self-Organizing Networks (SON); Self-Healing concepts and requirements, Rel-13, Version 13.0.0 (2015–12). TS 32.541. 3rd Generation Partnership Project. Barco, R., Díez, L., Wille, V., & Lázaro, P. (2009). Automatic diagnosis of mobile com- munication networks under imprecise parameters. Expert Systems with Applica- tions, 36 (1), 489–500. doi: 10.1016/j.eswa.2007.09.030 . Barco, R., Lazaro, P., Diez, L., & Wille, V. (2008). Continuous versus discrete model in autodiagnosis systems for wireless networks. IEEE Transactions on Mobile Com- puting, 7 (6), 673–681. doi: 10.1109/TMC.2008.23 . Barco, R., Lázaro, P., Wille, V., Díez, L., & Patel, S. (2009). Knowledge acquisition for diagnosis model in wireless networks. Expert Systems with Applications, 36 (3), 4745–4752. doi: 10.1016/j.eswa.2008.06.042 . Barco, R., Lázaro, P., & Muñoz, P. (2012). A unified framework for Self-Healing in wireless networks. IEEE Communications Magazine, 50 (12), 134–142. doi: 10.1109/ MCOM.2012.6384463 . Begum, S., Chakraborty, D., & Sarkar, R. (2015). Cancer classification from gene ex- pression based microarray data using SVM ensemble. In 2015 international con- ference on condition assessment techniques in electrical systems (CATCON) (pp. 13– 16). doi: 10.1109/CATCON.2015.7449500 . Breiman, L. (1996). Bagging predictors. In Machine learning (pp. 123–140) . Ciocarlie, G., Lindqvist, U., Nováczki, S., & Sanneck, H. (2013). Detecting anomalies in cellular networks using an ensemble method. In Proceedings of the 9th interna- tional conference on network and service management (CNSM 2013) (pp. 171–174). doi: 10.1109/CNSM.2013.6727831 . Dasarathy, B., & Sheela, B. V. (1979). A composite classifier system design: con- cepts and methodology. Proceedings of the IEEE, 67 (5), 708–713. doi: 10.1109/ PROC.1979.11321 . Freund, Y., & Schapire, R. E. (1997). A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55 (1), 119–139. doi: 10.1006/jcss.1997.1504 . Gandhi, I., & Pandey, M. (2015). Hybrid ensemble of classifiers using voting. In 2015 international conference on green computing and internet of things (ICGCIoT) (pp. 399–404). doi: 10.1109/ICGCIoT.2015.7380496 . Gómez-Andrades, A., Muñoz Luengo, P., Khatib, E., de la Bandera Cascales, I., Ser- rano, I., & Barco, R. (2015). Methodology for the design and evaluation of Self-Healing LTE networks. IEEE Transactions on Vehicular Technology, PP (99). doi: 10.1109/TVT.2015.2477945 . 1–1 ómez-Andrades, A., Muñoz, P., Serrano, I., & Barco, R. (2016). Automatic root cause analysis for LTE networks based on unsupervised techniques. IEEE Transactions on Vehicular Technology, 65 (4), 2369–2386. doi: 10.1109/TVT.2015.2431742 . acobs, R. A., Jordan, M. I., Nowlan, S. J., & Hinton, G. E. (1991). Adaptive mixtures of local experts. Neural Computing, 3 (1), 79–87. doi: 10.1162/neco.1991.3.1.79 . olliffe, I. (2002). Principal component analysis. Springer Series in Statistics (2nd). Springer-Verlag New York . hatib, E. J., Barco, R., Gómez-Andrades, A., Muñoz, P., & Serrano, I. (2015). Data mining for fuzzy diagnosis systems in LTE networks. Expert Systems with Appli- cations, 42 (21), 7549–7559. doi: 10.1016/j.eswa.2015.05.031 . hatib, E. J., Barco, R., Gómez-Andrades, A., & Serrano, I. (2015). Diagnosis based on genetic fuzzy algorithms for LTE Self-Healing. IEEE Transactions on Vehicular Technology . doi: 10.1109/TVT.2015.2414296 . Kittler, J., Hatef, M., Duin, R. P. W., & Matas, J. (1998). On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20 (3), 226–239. doi: 10. 1109/34.667881 . uncheva, L. (2002). A theoretical study on six classifier fusion strategies. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24 (2), 281–286. doi: 10. 1109/34.982906 . iu, H., Chen, G., Song, G., & Han, T. (2009). Analog circuit fault diagnosis using bagging ensemble method with cross-validation. In International conference on mechatronics and automation, 2009. ICMA 2009 (pp. 4 430–4 434). doi: 10.1109/ ICMA.2009.5246675 . ehlführer, C. , Wrulich, M. , Colom Ikuno, J. , Bosanska, D. , & Rupp, M. (2009). Sim- ulating the long term evolution physical layer. In Proc. of 17th European signal processing conference (EUSIPCO) . Muñoz, P., de la Bandera, I., Ruíz, F., Luna-Ramírez, S., Barco, R., Toril, M., et al. (2011). Computationally-efficient design of a dynamic system-level LTE simu- lator. International Journal of Electronics and Telecommunications, 57 (3), 347–358. doi: 10.1155/2012/802606 . ováczki, S. (2013). An improved anomaly detection and diagnosis framework for mobile network operators. In 2013 9th international conference on the design of reliable communication networks (drcn) (pp. 234–241) . Shen, H.-B., & Chou, K.-C. (2006). Ensemble classifier for protein fold pattern recog- nition. Bioinformatics, 22 (14), 1717–1722. doi: 10.1093/bioinformatics/btl170 . zilágyi, P., & Nováczki, S. (2012). An automatic detection and diagnosis framework for mobile communication systems. IEEE Transactions on Network and Service Management, 9 (2), 184–197. doi: 10.1109/TNSM.2012.031912.110155 . ei, H., Lin, X., Xu, X., Li, L., Zhang, W., & Wang, X. (2014). A novel ensemble clas- sifier based on multiple diverse classification methods. In 2014 11th interna- tional conference on fuzzy systems and knowledge discovery (FSKD) (pp. 301–305). doi: 10.1109/FSKD.2014.6980850 . iezbicki, T., & Ribeiro, E. P. (2016). Sensor drift compensation using weighted neu- ral networks. In 2016 IEEE conference on evolving and adaptive intelligent systems (EAIS) (pp. 92–97). doi: 10.1109/EAIS.2016.7502497 . u, X., Kumar, V., Ross Quinlan, J., Ghosh, J., Yang, Q., Motoda, H., et al. (2008). Top 10 algorithms in data mining. Knowledge and Information Systems, 14 (1), 1–37. doi: 10.1007/s10115- 007- 0114- 2 . uksel, S., Wilson, J., & Gader, P. (2012). Twenty years of mixture of experts. IEEE Transactions on Neural Networks and Learning Systems, 23 (8), 1177–1193. doi: 10. 1109/TNNLS.2012.2200299 . http://dx.doi.org/10.13039/501100002878 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0001 http://dx.doi.org/10.1016/j.eswa.2007.09.030 http://dx.doi.org/10.1109/TMC.2008.23 http://dx.doi.org/10.1016/j.eswa.2008.06.042 http://dx.doi.org/10.1109/MCOM.2012.6384463 http://dx.doi.org/10.1109/CATCON.2015.7449500 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0007 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0007 http://dx.doi.org/10.1109/CNSM.2013.6727831 http://dx.doi.org/10.1109/PROC.1979.11321 http://dx.doi.org/10.1006/jcss.1997.1504 http://dx.doi.org/10.1109/ICGCIoT.2015.7380496 http://dx.doi.org/10.1109/TVT.2015.2477945 http://dx.doi.org/10.1109/TVT.2015.2431742 http://dx.doi.org/10.1162/neco.1991.3.1.79 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0015 http://dx.doi.org/10.1016/j.eswa.2015.05.031 http://dx.doi.org/10.1109/TVT.2015.2414296 http://dx.doi.org/10.1109/34.667881 http://dx.doi.org/10.1109/34.982906 http://dx.doi.org/10.1109/ICMA.2009.5246675 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0021 http://dx.doi.org/10.1155/2012/802606 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0023 http://refhub.elsevier.com/S0957-4174(16)30377-3/sbref0023 http://dx.doi.org/10.1093/bioinformatics/btl170 http://dx.doi.org/10.1109/TNSM.2012.031912.110155 http://dx.doi.org/10.1109/FSKD.2014.6980850 http://dx.doi.org/10.1109/EAIS.2016.7502497 http://dx.doi.org/10.1007/s10115-007-0114-2 http://dx.doi.org/10.1109/TNNLS.2012.2200299 Combination of multiple diagnosis systems in Self-Healing networks 1 Introduction 2 Problem formulation 2.1 Root cause analysis in mobile communications networks 2.2 Automated diagnosis from the classification theory 2.3 State-of-the-art in ensemble-based classification algorithms 3 Method for combining multiple automatic diagnosis systems 3.1 Construction of the behavior models 3.2 Combination of behavior models 4 Proof of concept 4.1 Combination of diagnosis models devised by multiple experts 4.1.1 Scenario 4.1.2 The standalone classifiers 4.1.3 Results 4.2 Combination of different diagnosis systems on a live network 4.2.1 Scenario 4.2.2 The standalone classifiers 4.2.3 Results 5 Future lines of work 6 Conclusions Acknowledgment References 
work_bh3tqo6tq5cd7mlzt4sms7bbb4	----	Seminar Features of the Expert-System-Shell SPIRIT Wilhelm Rödder, Elmar Reucher, Friedhelm Kulmann Overview Knowledge processing in SPIRIT Reliability of answers Graphs and hypergraphs Recall by a stimulus Conclusion and remarks CII 04 1 Knowledge processing in SPIRIT Step 1: Definition of a knowledge domain Specify variables Vl with respective values vl; Literal Vl= vl e.g.: MARITAL=single, STUDENT=true. Propositions formed by junctors ∧ (and), ∨ (or), − (not), Denoted by A,B,C ⇒ Propositional language L. Extension to conditional language L|L by binary conditional operator ⎪. e.g.: MARITAL=single | (STUDENT=true ∧ PARENT=true). B|A [x], A, B ∈ L, x ∈ [0;1]. CII 04 2 Excursus CII 04 14 Step 2: Knowledge acquisition Given set of rules R = {Bi|Ai [xi], i=1,…,I}. Adaption of uniform distribution P0 to R by solving P*=arg min R(Q, P0), s.t. Q ╞ R (1) R(Q, P0) relative entropy from P0 to Q. CII 04 3 Excursus CII 04 15 Step 3: Inference Focus E = {Dj|Cj [yj], j=1,…,J}. Adaption of P* to E by solving P**=arg min R(Q, P*), s.t. Q ╞ E. (2) Query: H|G Answer P**(H|G). CII 04 4 Excursus CII 04 16 Reliability of answers Given P*=arg min R(Q, P0), s.t. Q ╞ R (1) H|G = ? Lower bound u = min Q (H|G) s.t. Q ╞ R and Upper bound u = max Q (H|G) s.t. Q ╞ R. Second order uncertainty of H|G m = -ld u - (-ld u ) [bit]. CII 04 5 Excursus CII 04 17 Graphs and hypergraphs Example creditworthiness NB: No Bad earlier credits (t/f) SU: somebody offers SUrety (t/f) KN: client in KNown to the bank (t/f) ME: financial MEans available (t/f) IN: INcome sufficient (t/f) JO: JOb for more than 3 years (t/f) IA: Inquiry Agency (t/f) LO: LOan the money (t/f) GO: GOod credits (yes/no) U: RetUrn of investment. CII 04 6 Graphs and hypergraphs Markov net Given set of finite valued variables V = {V1,…,VL}. With respect to (V;P) if for any variable Vl, Vm: (Vl, Vm) ∉ E ⇔ (Vl⊥ Vm | V\{l,m};P). ⇒ Minimal independency graph CII 04 7 Graphs and hypergraphs Inference net Given set of finite valued variables V = {V1,…,VL}. Variables Vl and Vm are involved in a rule Bi|Ai Vl and Vm connected by an arrow, if a value vl involved in Ai and vm in Bi. Vl and Vm connected by an edge, if vl and vm appear in the conclusion Bi of the same rule. CII 04 8 Excursus CII 04 18 Graphs and hypergraphs Hypertree Given set of finite valued variables V = {V1,…,VL}. Denote Ei(Bi|Ai) ⊆ V set of variables involved in a rule Bi|Ai ⇒ Ei hyperedges of the hypergraph (V, ). E In general (V, ) not acyclic, E use “fill-in”-methods to construct (acyclic) hypertree For propagation: Hypertree ⇒ junctiontree (each node corresponds to an edge of the hypertree) CII 04 9 Excursus CII 04 19 Graphs and hypergraphs Application credit worthiness NB: No Bad earlier credits true KN: client in KNown to the bank true Amount of credits 10.000 € Credit’s lifespan: 4 years U = 723,06 € JO: JOb for more than 3 years (t/f) false SU: somebody offers SUrety (t/f) false ME: financial MEans available (t/f) true IN: INcome sufficient (t/f) true IA: Inquiry Agency (t/f) true LO: LOan the money (t/f) yes GO: GOod credits (yes/no) yes CII 04 10 Excursus CII 04 20 Recall by a stimulus Vl ∈ {V1,…,VL}, P* epistemic state. P** adaption of P* to a certain focus E = {F [1.]}. Impact measure: R((Vl; P**),(Vl; P*)) [bit]. CII 04 11 Excursus CII 04 21 Excursus CII 04 22 Conclusion and remarks Model no. variable s no. rules no. LEGs H(P 0) H(P*) utility yes/no decision yes/no BB 20 340 17 29.91 18.57 no no TS 76 574 50 76.00 12.83 no yes CR 18 38 13 22.68 6.00 no no BS 86 1051 36 104.79 87.12 no yes OD 6 18 3 8.17 4.08 yes yes CW 10 31 6 11.00 7.38 yes yes blue baby (BB) troubleshooter (TS) car repair (CR) business-to-business (BS) oil drilling problem (OD) credit worthiness support system (CW) CII 04 12 Knowledge processing in SPIRIT Reliability of answers Graphs and hypergraphs Recall by a stimulus Conclusion and remarks Step 1: Definition of a knowledge domain Specify variables Vl with respective values vl; Literal Vl= e.g.: MARITAL=single, STUDENT=true. Propositions formed by junctors ( (and), ( (or), ( (not), Denoted by A,B,C ( Propositional language L. Extension to conditional language L|L by binary conditional operator (. e.g.: MARITAL=single | (STUDENT=true ( PARENT=true). B|A [x], A, B ( L, x ( [0;1]. Step 2: Knowledge acquisition Given set of rules R = {Bi|Ai [xi], i=1,…,I}. Adaption of uniform distribution P0 to R by solving P*=arg min R(Q, P0), s.t. Q ╞ R (1) R(Q, P0) relative entropy from P0 to Q. Step 3: Inference Focus E = {Dj|Cj [yj], j=1,…,J}. Adaption of P* to E by solving P**=arg min R(Q, P*), s.t. Q ╞ E. (2) Query: H|G Answer P**(H|G). Given P*=arg min R(Q, P0), s.t. Q ╞ R (1) H|G = ? Lower bound = min Q (H|G) s.t. Q ╞ R and Upper bound = max Q (H|G) s.t. Q ╞ R. Second order uncertainty of H|G m = -ld - (-ld ) [bit]. Example creditworthiness NB: No Bad earlier credits (t/f) SU: somebody offers SUret KN: client in KNown to the bank (t/f) ME: financial MEans av IN: INcome sufficient (t/f) JO: JOb for more than 3 years IA: Inquiry Agency (t/f) LO: LOan the money (t/f) GO: GOod credits (yes/no) U: RetUrn of investment. Markov net Given set of finite valued variables V = {V1,…,VL}. With respect to (V;P) if for any variable Vl, Vm: (Vl, Vm) ( E ( (Vl( Vm | V\{l,m};P). ( Minimal independency graph Inference net Given set of finite valued variables V = {V1,…,VL}. Variables Vl and Vm are involved in a rule Bi|Ai Vl and Vm connected by an arrow, if a value vl involved in Vl and Vm connected by an edge, if vl and vm appear in the Hypertree Given set of finite valued variables V = {V1,…,VL}. Denote Ei(Bi|Ai) ( V set of variables involved in a rule Bi|Ai ( Ei hyperedges of the hypergraph (V, ). In general (V, ) not acyclic, use “fill-in”-methods to construct (acyclic) hypertree For propagation: Hypertree ( junctiontree (each node corresponds to an edge of the hypertree) Application credit worthiness NB: No Bad earlier credits true KN: client in KNown to the bank true JO: JOb for more than 3 years (t/f) false SU: somebody offers SUrety (t/f) false ME: financial MEans available (t/f) true IN: INcome sufficient (t/f) true IA: Inquiry Agency (t/f) true LO: LOan the money (t/f) yes GO: GOod credits (yes/no) yes Vl \( {V1,…,VL}, P* epistemic state. P** adaption of P* to a certain focus E = {F [1.]}. Impact measure: R((Vl; P**),(Vl; P*)) [bit]. blue baby (BB) troubleshooter (TS) car repair (CR) business-to-business (BS) oil drilling problem (OD) credit worthiness support system ( 
work_bdo3y2u4gfdl3hj4ntxfwmmd4u	----	paper.dvi A review of expert systems for  chromatography Bryant, CH, Adam, AE, Taylor, DR and Rowe, RC http://dx.doi.org/10.1016/0003­2670(94)00209­6 Title A review of expert systems for chromatography Authors Bryant, CH, Adam, AE, Taylor, DR and Rowe, RC Type Article URL This version is available at: http://usir.salford.ac.uk/1775/ Published Date 1994 USIR is a digital collection of the research output of the University of Salford. Where copyright  permits, full text material held in the repository is made freely available online and can be read,  downloaded and copied for non­commercial private study or research purposes. Please check the  manuscript for any further copyright restrictions. For more information, including our policy and submission procedure, please contact the Repository Team at: usir@salford.ac.uk. mailto:usir@salford.ac.uk A Review of Expert Systems for Chromatography C�H�Bryant�� A�Adam �Computation Department� D�R�Taylor �Chemistry Department� University of Manchester Institute of Science and Technology� PO Box ��� Manchester� M� QD� United Kingdom� R�C�Rowe Zeneca Pharmaceuticals� Alderley Park� Maccles�eld� Cheshire� SK �NA� United Kingdom� Abstract Expert systems for chromatography are reviewed� A taxonomy is proposed that allows present �and future� expert systems in this area to be classi�ed and facilitates an under� standing of their inter�relationship� All the systems are described focusing on the reasons for their development� what their purpose was and how they were to be used� The engi� neering methods� knowledge representations� tools and architectures used for the systems are compared and contrasted in a discussion covering all the stages of the development life cycle of expert systems� The review reveals that too often developers of expert systems for chromatographydo not justify their decisions on engineering matters and that the literature suggests that many ideas advocated by knowledge engineers are not being used� � Introduction Section � compares this work with previous reviews section introduces some of the con� cepts and terminology of chromatography and describes how a chromatography method is developed section � gives a de�nition of �expert system and speci�es the scope of this re� view section � explains why expert systems for chromatography are needed and section � proposes a taxonomy for these systems� describes its purpose and how it was developed� � Sections �� �� and � are arranged in a taxonomic fashion corresponding the taxonomy and explain why each system was developed� what its purpose was and how it was to be used� Sections ��� ��� � � and �� discuss the engineering methods� knowledge representations� tools and architectures used in the development of expert systems for chromatography� � Previous Reviews Many papers have been published that review arti�cial intelligence for chemistry but their scope is sowide as to prohibit the inclusion of a reviewof expert systems for chromatography as detailed as that given here� For an example see ���� In addition to arti�cial intelligence for chemistry� the more speci�c topic of expert systems for analytical chemistry has been reviewed in a number of papers� ��� provides an introduction to knowledge representation and explanation facilities for expert systems for analytical chemistry but only gives four references for expert systems for chromatography� � � also discusses topics such as knowl� edge representation but it does not mention any expert systems for chromatography� ��� compares the use of arti�cial intelligence languages and expert system shells for building expert systems for limited domains but only reviews one expert system for chromatography� Three papers ��� ��� ��� have been published that speci�cally review expert systems for chromatography� ��� reviews expert systems for liquid chromatography very brie�y� It lacks the authoritative tone that usually accompanies an academic paper� the style of writing is imprecise� some of the analogies used are weak and only two references are given� ��� does not give any references and only reviews four systems of which only one is an expert system� the other three being simulation programs� ��� mentions more than sixty � computer systems but only about a quarter of these are expert systems for chromatography� This paper reviews all the expert systems for chromatography that ��� reviews and a further eleven expert systems for chromatography that ��� does not mention� ��� does not make the relationship between di�erent expert systems for chromatography clear this paper proposes a taxonomy which facilitates an understanding of their inter�relationship� ��� has a useful explanation of how expert systems in general are structured and introduces some aspects of knowledge engineering� However its analysis of how expert systems for chromatography have been engineered is super�cial and� in contrast to this paper� no conclusions are drawn on this engineering� This paper reviews expert systems for chromatography comprehensively� covering more than twenty��ve systems and gives over eighty references� �This number does not include subsystems of expert systems for chromatography� � � An Brief Introduction to Chromatography The aim of this section is to explain brie�y some concepts and terminology of chromatog� raphy and to describe how a chromatography method is developed� a knowledge of which is assumed later in this paper� Complete de�nitions� including any relevant mathematical equations� can be found in the references given� Chromatography is a separation process in which the sample mixture is distributed between two phases in the chromatographic bed �column or plane�� One phase is stationary whilst the other passes through the chromatographic bed� ��� Thus the former is referred to as the stationary phase and the latter as the mobile phase �or eluent�� Substances to be separated by a chromatographic system must have di�erent relative a�nities for these two phases� Thus� a substance with a relatively higher a�nity for the stationary phase moves with a lower velocity through the chromatographic system than does a substance with lower a�nity� This di�erence in migration velocity ultimately leads to physical separation of the components in a sample� ��� A component of the sample mixture that leaves the stationary phase is said to be eluted in a process known as elution� The eluting power is the power of the eluent� that is the mobile phase� to elute the components remaining on the stationary phase� If the substances to be separated do not have widely di�ering a�nities for the stationary phase then the column can be eluted with the same solvent �mobile phase� all the time� However if the a�nities vary widely then the composition of the eluting solvent can be gradually changed� This technique is known as gradient elution� ���� Stationaryphasesareeither a solid� porous� surface�activematerial in small�particle form or a solid support covered with a thin �lm of liquid ���� The particle size is related to the e�ciency� Column e�ciency is inversely related to the rate at which solute molecules spread out as they travel through the stationary phase ����� The potential of a chromatographic system to separate two compounds is referred to as its selectivity� Most expert systems for chromatography concern chromatographic systems in which the eluted compounds are transported to a detector and recorded as Gaussian �bell�shaped� curves� The signals are known as peaks and are shown on the chromatogram� ��� The chromatogram is a record of the concentration or mass pro�le of the sample components as a function of the movement of the mobile phase ����� The degree of separation of successive solute bands or peaks is referred to as the resolution ����� The peaks give qualitative and quantitative information on the mixture in question ����� Qualitative A peak can be identi�ed by injecting the relevant substance and then com� paring retention times� The retention time is equal to the period between sample in� � troduction and the detector sensing the maximum of the retained peak� The retention time of a component is always constant under identical chromatographic conditions� The column dimensions� type of stationary phase� mobile phase composition and �ow velocity� sample size and temperature provide the chromatographic conditions� Since these conditions vary the capacity factor is better for characterising a compound� al� though its measurement does require that an unretained component is timed through the column� The capacity factor is the ratio of the time spent by the solute in the stationary phase to the time spent in the mobile phase �or eluent�� Quantitative The area of a peak is proportional to the amount of a compound injected� Calibration graphs can be drawn to determine unknown concentrations of identi�ed samples� ��� Chromatographic Modes Gas chromatography is the branch of chromatography in which the mobile phase is a gas� In HPLC �High Performance Liquid Chromatography� the mobile phase is a liquid� There are a number of methods� or modes� for liquid chromatography� Adsorption� Reversed�phase and Ion�pair are three which are relevant to this paper� In adsorption chromatography a relatively polar material is used as the stationary phase and a relatively non�polar solvent as the mobile phase� The di�erent rates at which the various types of molecules in the mixture are adsorbed on the stationary phase provide the separation e�ect� In reversed�phase chromatographythe polarity is reversed� the stationary phase is very non� polar and the mobile phase is relatively polar� Ion�pair chromatography can be used for the separation of ionic compounds� Ionic sample molecules are �masked by a suitable counter ion hence the term pair� ��� ��� Method Development Developing a method for chromatography involves �ve steps�� �� Method Selection This step is also known as the �rst guess� The appropriate chro� matographic method� materials and instrumentation are selected� �� Optimisation The three essential characteristics of any chromatographic separation �retention� selectivity and e�ciency� are optimised� �a� Retention OptimisationA chromatogramis obtained in which the components of interest appear as sharp� symmetrical peaks with a retention time in an opti� mum range� The retention of components needs to be su�ciently high to achieve � separation� but su�ciently low to maintain a reasonable analysis time and good sensitivity� �b� Selectivity Optimisation The best possible selectivity is sought within the constraints of optimum retention times for each peak� �c� Method �or Chromatographic or System or E�ciency� Optimisation After the selectivity has been optimised the e�ciency of the chromatographic system is optimised by selecting the most appropriate column� operating condi� tions and instrumentation� � Method Validation Finally the method is validated to prove that it meets the ana� lytical requirements� � De�nition of an Expert System An expert system is a computer program that represents and reasons with knowledge of some specialist subject with a view to solving problems or giving advice� An expert system should exhibit all of the following features to some degree� ���� � It simulates human reasoning about a problem domain� rather than simulating the domain itself� � It performs reasoning over representations of human knowledge� in addition to doing numerical calculations or data retrieval� � It solves problems by heuristic or approximate methods which unlike algorithmic so� lutions� are not guaranteed to succeed� � It deals with subject matter of realistic complexity that normally requires a consider� able amount of human expertise� � It must exhibit high performance in terms of speed and reliability in order to be a useful tool� � It must be capable of explaining and justifying solutions or recommendations to con� vince the User that its reasoning is in fact correct� ���� gives a broad introduction to expert systems� The literature abounds with books on the various aspects of engineering relevant to this paper� Useful texts include �� � for knowledge acquisition� ���� for knowledge representation and ���� for expert system tools and languages� As stated above an expert system does not simulate the domain itself but human reason� ing about the domain� Hence this paper does not review literature that describes work in � which the primary aim was to model� quantitatively or qualitatively� aspects of chromatog� raphy� Thus ���� is not discussed because it is concerned with model based reasoning and Drylab ���� is not included because it simulates HPLC development� � Why Expert Systems are needed for Chromatography The introduction of workable expert systems for chromatography would be of great bene�t because of the wide range of analytes amenable to chromatography and the complexity of this �eld with regard to the choice of materials and instruments ����� An example of the wide range of amenable analytes is seen in the pharmaceutical indus� try� there is an ever increasingvolumeof diversenovel compounds which have to be screened in order to develop compounds with diagnostic or therapeutic properties� In principle� each compound needs its own method of analysis thus the method development process must be repeated in its entirety for each new compound� Most of these new compounds are analysed using some form of chromatography� mainly HPLC� ���� The choice of materials for HPLC is complex� Many factors a�ect the choice of mode of separation� packing material� mobile phase� and instrumental operating conditions� The development of an optimum method and the interpretation of the results usually requires a lot of expertise and experience to solve the problems that arise for each particular case� It is impossible for any individual analyst to become an expert in all the various areas of chromatography� Hence individual analysts working in areas in which they are not experts have to seek assistance from the experts in those areas� Human� as opposed to computer� experts have several drawbacks however� They are often in short supply or unavailable in a particular laboratory� They are usually busy and have little time to spare to help non�experts� Occasionally their expertise is lost because they leave an organisation or die� Sometimes their expertise is temporarily unavailable because they are ill or on leave� When they are available they may not always be in a helpful mood and may not o�er consistent advice� Expert systems o�er an opportunity to encode chromatographicexpertise in computer systems which can be made widely available and which provide advice which is consistent� User�friendly and easily documented� � The Taxonomy This section discusses the proposed taxonomy of expert systems for chromatography� The purpose of the taxonomy is to�� � � facilitate the readers understanding of how the systems are related� The diagrams that show the taxonomy allow it to be quickly assimilated and they can be used for reference� � allow readers to focus their attention on a class of expert system for chromatography that is of particular interest to them� if they should so wish� � allow future expert systems for chromatography to be classi�ed and hence illustrate how they are related to previous systems� The taxonomy was developed as follows� A number of attributes that were candidates for classifying the expert systems were selected� They included the type of expert system� who developed it and where� brief details of how it was implemented and what the domain was� For each system a list of the values of these attributes was compiled� Attempts were made to group the systems into classes where class membership was determined by the values of one of these attributes� It proved impossible to develop a taxonomy that unambiguously classi�ed all the systems if the type of expert system was used to determine class membership� the classes generated were not mutually exclusive for all the systems� Attempts to classify the systems by considering who developed them or where did not give rise to a taxonomy that would� in any way� facilitate the readers understanding of the inter�relationship between the systems� Classifying the systems in terms of how they were implemented was not possible because their implementations could not be accurately compared and contrasted� the range of aspects of the implementations described in the literature varies for di�erent systems� However it was discovered that all the systems could be unambiguously classi�ed if class membership was based on the domain of a system� An analysis of the resulting classes showed that in some cases two or more of these classes were subclasses of a more general class� An examination of these more general classes showed that some of them were subclasses themselves of a yet more general class� Thus a taxonomy was developed� as shown in Figures �� � and � Sections �� �� and � describe the functionality of the various expert systems for chro� matography� presenting them in a taxonomic fashion corresponding to the taxonomy out� lined above� � Systems Not Speci�c to a Chromatographic Method or Class of Compounds This section discusses all the expert systems for chromatography that were not speci�cally designed for a particular method or a particular class of chemical compounds� � ��� Systems Designed for Several Stages of the Process This section corresponds to the branch of the taxonomy which is entitled �Multiple Stage on Figure � and describes three projects �ESCA� ECAT and ESC�ESLC� which all attempted to implement systems that would give advise to an analyst on several stages of the process of developing a successful HPLC separation� ����� ESCA The expert systems in chemical analysis �ESCA� project� Esprit project ����� started in May ����� o�cially �nished in May ����� and involved �ve partners�� � Philips Scienti�c� Cambridge� U�K� � Catholic University Nijmegen� The Netherlands � Organon International B�V�� Oss� The Netherlands � Philips Research Eindhoven� The Netherlands � Philips Research Hamburg� F�R�G� � Brussels Free University� Belgium� ESCA �wasset up to evaluate the merits of expert system technology for use in industrial chemical analysis� ����� Its aim was to provide expert systems that would illustrate the bene�ts and shortcomings of expert system technology� ESCA resulted in the publication of around �� papers� Some of these give general descriptions of the project� �See ���� ���� ���� ������ The knowledge domain chosen for ESCA was HPLC method development in pharmaceu� tical analysis� �In the pharmaceutical industry an ever increasing volume of diverse novel compounds has to be searched in order to develop compounds with diagnostic or therapeutic properties� In principle� each compound needs its own method of analysis� � � �Most of these � � �are analysed using some form of chromatography� mainly HPLC� � ���� It was believed that an expert system for HPLC method development �would speed up the process mak� ing it more consistent and better documented� ���� while o�ering expertise that was not generally available in every laboratory� Initially the following four domains which covered the entire �eld of method development in HPLC were selected� Each domain represented a discrete step in the process� Thus the intention was to build four experts systems that could be integrated later� �� Selection of Initial Conditions � �� Selection of Selectivity Optimisation Criteria It was decided to build an expert system for the selection of optimisation criteria because �it is di�cult to select the most appropriate criterion in di�erent situations� ���� and the choice greatly a�ects the outcome of the optimisation� The resulting system helped the User to use a conventional package of computer optimisation programs� � Optimisation of Chromatographic Parameters Chromatographic �or Method� optimisation follows selectivity optimisation� The latter results in a method that yields adequate separation in an acceptable amount of time� for the given instrumental conditions� The former optimises these conditions� The aim of the optimisation of the chromatographic parameters is to reduce analysis time and increase the sensitivity of the method� An expert system was developed for this task because the �relations between these parameters are complex� Finding the optimal settings requires the evaluation of a number of equations that are di�cult to see through� even after a long period of study�� ���� �� Validation of the Development Method The expert system developed for this domain was limited to precision testing� A successful integration of the work of all the participants in the ESCA project required that the project was managed e�ectively� there was a large number of participants from �ve organisations that were geographically isolated from one another� Although there were no less than four proposals on how various parts of the work could be integrated� not one of these described how all of the parts could be combined� Furthermore the literature does not provide any evidence that workers on the ESCA project understood that the success of the integration of a project as large as ESCA could not be left to chance and that consequently there was a need for a plan that described how all the parts would be integrated� This may be because the actual goal of the ESCA project was the development of stand�alone systems ����� The four parts of the domain covered by the four proposals on how the various parts of the work could be integrated are outlined below� �� The selection of initial conditions� retention optimisation and selectivity optimisation� �� Selectivity and chromatographic �or method� optimisation� � Method validation� �� Repeatability testing and troubleshooting� Figure � shows the ESCA subsystems implemented for each of the ESCA subdomains and shows the four proposals for integrating the subsystems� Proposal One for Intergration of ESCA Work� Selection of Initial Conditions� Retention Optimisation and Selectivity Op� timisation� Some of the workers on the ESCA project performed a feasibility study ���� �� � for the construction of an integrated system for the selection of initial conditions� retention optimisation� and selectivity optimisation� They proposed to construct the system from the stand�alone expert systems �described in the remaining part of this section� and some additional knowledge which was needed to direct the user through the integrated system� In the proposed integrated system �rst�guess conditions were to be selected by one of the expert systems LABEL� DASH or LIT� the choice depending on the application �eld� After carrying out the �rst experiment� the retention time range of the solutes was to be evaluated� If there were solutes with capacity factors outside the desired range� then one of the corresponding retention optimisation expert systems LABEL�� DASH� or LIT� was to be consulted� The result would be a chromatogram in which all the solutes would have eluted within a reasonable time but two or more peaks may still have overlapped� The selectivity optimisation expert system SLOPES would then have been consulted� DASH �Drug Analysis System in HPLC� ���� ���� was originally developed for chro� matographic method selection and retention optimisation for the purity control of basic compounds� namely CNS �central nervous system��active and cardiovascular drugs� The system determined the initial conditions to obtain a capacity factor between and ��� �In more than ��� of all cases� correct predictions were obtained for the original family of substances� ����� DASH� ����� an extension of DASH� was speci�cally designed for retention optimisation� DASH� was a good�second�guess system whereas DASH was a �rst�guess system� DASH had proved satisfactory when limited to basic drugs� However the integrated ESCA system had to allow for a wider range of compounds� DASH� was proposed because DASH alone would not have performed satisfactorily for this wider range� LABEL ���� ���� ���� was an expert system for selecting the initial chromatographic conditions for the label claim analysis of pharmaceutical formulations on a cyanopropyl column used in di�erent chromatographic modes� LABEL was developed by one of the partners involved with the project before ESCA started� �It was included in the project because it covers the situation that one sample must be analysed for di�erent compounds� This is in contrast to DASH� ����� LABEL contained knowledge for the selection of a suitable detection mode� It only considered UV or electrochemical detection� the latter in the oxidation mode� ���� ���� ���� � LABELassumedtheuseof a single stationaryphase type� namelyanitrileor cyanopropyl column� whichcouldbe used in bothnormal�phaseandreversed�phasechromatography� LA� BEL selected the mobile phase system� either normal�phase� reversed�phase with water or reversed�phase with bu�er� using rules� It then decided if the addition of ion�suppressing agents to the eluting agent was necessary and �nally gave the starting composition of the mobile phase� ���� In � of �� tests performed on LABEL ��� �success was achieved in a manner that probably could not have been done better by a human chromatography expert� ����� LABEL�� like DASH�� was speci�cally designed for retention optimisation� the main task of the expert system was to situate the capacity factor in a suitable range� The optimisation was performed by increasing or decreasing the percentage of organic modi�er in the mobile phase� starting from the �rst�guess composition� ���� LIT was a small expert system that helped to select all the important parameters of a literature method and checked whether the method could be treated by SLOPES� ���� SLOPES� the SeLectivity OPtimisation Expert System� was to comprise three di�erent modules� VARIABLES� DESIRE� and CRISE� VARIABLES selected the relevant optimisa� tion parameters �the variables that need to be considered� and their boundaries �minimum and maximum limits�� DESIRE determined the type of the experimental design �the pat� tern according to which the necessary experiments will be performed such as Simplex or Doehlert� and the number and location of the experiments� Finally� the most suitable opti� misation criterion to describe the quality of separation in a chromatogram was selected by CRISE� CRISE �CRIteria SElection� ���� was an expert system for the selection of objective criteria for systematic optimisation of selectivity� Such criteria must characterise the quality of separation in a chromatogram� The system was developed for several reasons� but principally because numerous criteria have been suggested� each of which yield di�erent results and the choice depends on a large number of factors� �It is genuinely di�cult to select the most suitable criterion in a particular situation� ����� CRISE consisted of four modules� In the �rst the most suitable elemental criterion was selected to quantify the extent of separation between two adjacent peaks in the chro� matogram� Resolution �Rs�� separation factors �S�� separation factors corrected for plate counts �SN� but also expressions such as peak�valley ratios could be used� As Rs� SN and S do not take into account a possible loss in resolution due to non�ideal situations module � investigated whether any corrections were required� Module assisted the User in the se� lection of weighting�factors to make a di�erence between peaks according to their relevance �� or relative importance� In the last module� the elemental criteria were incorporated in the global optimisation criterion that was best suited to quantify the separation quality of the entire chromatogram taking into account the purpose of the chromatogram� � �� The main conclusions of the validation of CRISE were that � i� the expert system provided clear and unambiguous answers for each consultation and ii� the expert system and the human expert provided the same answers for all ten cases ��� consultations� considered during the validation� � ��� A hypermedia version of CRISE� CRISEBOOK � ��� is discussed in Section � ��� Proposal Two for Intergration of ESCA Work� Selectivity and Chromatographic �or Method� Optimisation� Some of the workers on the ESCA project proposed an integrated system for just two of the domains� selectivity optimisation and chromatographic �or method� optimisation� They proposed to construct the system from three programs� CRISE �see above�� Diamond and SOS� � �� Diamond was a package of conventional computer programs which required some ex� pertise to use it� The most di�cult decision that a Diamond User needed to make was the selection of the most appropriate optimisation criterion� In the integrated system this decision was to be made with the help of the expert system CRISE� Another expert system� SOS� was to be used to transform the chromatogram with optimum selectivity produced as a result of Diamond into the optimum overall method by establishing the best column conditions� instrumentation� injected amount� etc� SOS �System Optimisation System� optimised separations in terms of i� su�cient sep� aration ii� su�cient sensitivity and iii� shortest possible time � ��� To run the system� an initial chromatogram was needed� together with the relevant information of how it was recorded� The system used two databases� one describing the available columns� and the other the available detectors� �These were created once for a given laboratory� but could be modi�ed at any time�� The system used this information to�� � recommend the optimum column� detector cell� time constant� �ow�rate and sample size� � predict the required analysis time� the �critical resolution �that is the lowest value observed for the resolution between a relevant pair of peaks� and the pressure drop over the column� � provide an explanation of its reasoning in the form of a bar chart and some additional advice to the user� �� The optimum result was de�ned as�� �� The resolution for all relevant pairs of peaks had to exceed a minimum value speci�ed by the User� �� The signal�to�noise ratio for the smallest relevant peak had to exceed a minimum value speci�ed by the User� � The required analysis time had to be as short as possible� A prototype was implemented in Knowledge Craft � but the �nal system was imple� mented in Pascal� The developers described the conventional explain facilities provided by expert�system development tools such as Knowledge Craft as � generally not very helpful� � ��� Thus when SOS was reimplemented in Pascal � � a new set of help and explain facil� ities was implemented� As a result bar charts could be used to establish which factors were limiting the speed of analysis� which factors caused a particular combination to be invalid and whether or not a major reduction in the analysis time was still feasible� Proposal Three for Intergration of ESCA Work� Method Validation� The ESCA expert systems discussed so far have not dealt with method validation� A prototype for an integrated expert system for intralaboratory precision testing was imple� mented � �� � ��� It integrated two expert systems described below� one for repeatability testing and the other for ruggedness testing� The architecture of the integrated system is described in Section � ��� The purpose of a precision test is to establish the random deviation from the mean in a certain analysis� Precision testing normally consists of repeatability and �interlaboratory� reproducibility tests� In the former� the same sample is analysed under the same conditions by the same analyst a number of times� In the latter� the same sample is tested in di�erent laboratories to examine the precision of the method under slightly changing conditions� Since reproducibility tests involve more than one laboratory they are relatively costly to reduce these costs a ruggedness test can be performed after the repeatability test but before the reproducibility test� Like a repeatability test� a ruggedness test can be performed in just one laboratory but the e�ects of using di�erent laboratories are simulated this detects some of the problems that would usually be discovered during the relatively costly reproducibility test� The system was developed because method validation was becoming increasingly impor� tant as stricter rules were applied by regulatory authorities� Precision testing was a vital �Knowledge Craft is described in Section �� �� step in this validation� The system concentrated on precision tests that could be done in the laboratory where an LC method was developed� For most analysts in a routine labora� tory� the performance of a precision test was not straightforward� The system was intended to given the analyst validating the method as much certainty as possible that the method would not fail in a collaborative interlaboratory test� � �� The expert system for repeatability testing was called REPS �REPeatability testing Sys� tem� � �� � ��� The expert system was used to select suitable test procedures for particular applications and to interpret their results� All the algorithms for the calculation of variance were set up in spreadsheets which could interact with the expert system� This is described in Section � � � The system was reimplemented in Pascal in multiple windows environment ����� This version worked as follows� It started by consulting a system optimisation module to provide the fastest analysis time within the required resolution� The method features to be tested were then de�ned as the sample preparation and the injection procedure� The User had to input a description of the HPLC method together with information on its expected usage� The systemrecommendedanexperimentaldesignbasedon this input� TheUser thencarried out the experiments and collected data which was input to the system� The system then diagnosed problems with the repeatability of the HPLC method� Each problem diagnosed had a list of actions that could be taken to try to solve the problem� The order in which the actions were presented to the User was such that those which were most likely to solve the problem appeared �rst� The expert system for ruggedness testing was called RES �Ruggedness Expert System� � �� � �� ����� RES comprised six modules that represented the major steps in the set�up and interpretation of the ruggedness test� The modules were controlled by a supervisor� The Factor Choice Module This selected the factors to be included in the ruggedness test and the factor levels� �Normally� some �� factors may in�uence method per� formance in liquid chromatography� For every method� however� only a small group of factors is expected to be critical in that small changes can cause large e�ects�� ����� The relevant factors will be di�erent for every chromatographic method� All the relevant factors must be tested but the number of experiments required increases dramatically with the number of factors� Thus the number of factors chosen by the system had to be minimal but su�cient� The module was based on a rule based system ����� The expert system approach yielded acceptable results to the problem� it had proved satisfactory for ten of the eleven cases of pharmaceutical formulations tested� This was in contrast to an al� �� gorithmic approach ��� �the problem of factor choice is di�cult to handle in normal programming languages because no algorithms are available� and they are di�cult to design because of the many parameters involved and their complex interactions�� ���� The expert system could cope with the problem because the relations between the parameters were organised in frames� rules� handlers and demons ����� The Design Selection Module The module selected the experimental design from a set stored in the module� However the User could overrule the decision of the module and even add new designs to the system together with the rules for their selection� After the User had consulted the factor and design selection modules the experimental design could be stored in a �le for reference� The User then performed the experimental work and measured the necessary parameters� The Statistical Results Module Thisperformedstatistical calculationsusingalgorithms� The Chemical Results Module This contained heuristic knowledge� The output from this module consisted of a set of warnings that had to be included in the �nal method description� The module decided whether one of the two repair modules �described below� should be activated� The Reselect Factor Levels Module This was a rule based module which modi�ed the factor levels to a narrower range� The module produced a list of factors and levels that could serve as a new input to the design selection module to repeat the ruggedness testing procedure� The Method Improvement Module This was derived from SOS� which was described above� During the validation of RES the factors chosen by an unbiased expert were compared with those suggested by RES for �� test cases� In all but one� RES performed satisfactorily� Proposal Four for Intergration of ESCA Work� Repeatabil� ity Testing and Troubleshooting� The fourth proposed integration combined REPS� SOS� and three other modules which were built to add �exibility to the integrated system� ���� This allowed the User to consult the system in three di�erent situations� It could be used to access the repeatability of a new method� to check the repeatability of a previously validated method and as a trouble� shooting tool� The latter ��� �turned out to be a valuable feature� ����� Conclusions Arrising from the ESCA Project Despite the publication of around �� papers on the ESCA project� the implementation of over a dozen ESCA subsystems and no less than four proposals on how various parts of �� the ESCA work could be integrated there is no evidence in the literature to suggest that a single integrated system was implemented� Future projects to develop expert systems for chromatography which tackle domains as large as that covered by the ESCA project must be managed more e�ectively� ����� ECAT Wokers at Varian Associates in the USA developed ECAT �Expert Chromatographic Assis� tance Team� ���� �� � ����� ECAT was a collection of expert system programs intended to assist the inexperienced chromatographer in HPLC method development� The recommen� dations of the system were intended to lead the User logically through the entire method development process towards a workable HPLC separation� ECAT comprised the following four modules connected to an inference engine�� Column and Mobile Phase This received inputs about the analyte characteristics and made recommendations for the columnpacking� columngeometry� mobilephase liquids and mobile phase modi�ers� The module used heuristics which were represented by a set of rules� The module forward chained from the initial factual input about the analyte class�es�� queryingthe Userwherenecessary� in order todevelopamoregeneral identity for the class�es�� or to conclude the existence of properties associated with the class�es�� When the chain of inferencing was exhausted the module searched for key facts to determine which of its recommendations it was to output� The module then gave the User the option of changing some of the decisions that the module made when the rules were initially �red� If the User decided to change any decisions then the module redesigned the separation based on the alternative choice� Sample Preparation This helped the User to determine whether or not the sample re� quired pre�treatment in order to enhance some quality of the separation� The module s knowledge base contained sets of rules and facts for determining whether a guard col� umn was needed� and whether the analysis could be aided by solid�phase extraction techniques� A list of recommendations was delivered at the end of the session� Method Optimisation This led the User through a series of experiments in which the separation was optimised in both total time and acceptable resolution� This module was limited to a small set of rules which led the User to use a limited range algorithmic optimisation strategies� Database of Chemical Properties This contained factual information about speci�c chemicals and classes of chemicals� �� ����� ESC�ESLC A chromatography expert system was developed by researchers at the Dalian Institute of Chemical Physics in China ���� ���� ���� ����� This system was named ESC �Expert System for Chromatography� in ���� but was referred to as ESLC �Expert System for Liquid Chromatography� in ����� The strategy for the development of the system was to develop a knowledge and chro� matogram base� an inference engine and a user interface� The chromatogram base could be searched by compound� analyte class or author� �The base included literature references for the chromatograms�� Thus the User could either enter the structure of the sample or the sample name or the name of the class to which the analyte belonged� ���� The chro� matogram played an important role in that it not only provided a reliable knowledge base for the system but it was also used to verify the results recommended by the system ����� The system had �ve modules�� Separation Mode� Mobile Phase and Stationary Phase Selection Selection of the Sample Pretreatment and Detection Method Optimisation of the Operating Conditions Peak Identi�cation Four approaches for identifying peaks were considered� one was on� line and the other three were o��line� Diagnosis of the Hardware System Thiswasdevelopedbecause�almostall chromatographs do not possess self�diagnosis at the heart of the system� the column� ����� It proposed a general strategy for hardware diagnosis involving comparison of the hardware to be used with standards� Any di�erences found resulted in rulesbeing �red and adiagnosis being proposed� ��� Systems Designed for One Stage of the Process Only This section describes those systems shown on Figure � and is organised in a taxonomic fashion corresponding to Figure �� All the systems were designed for one stage of the chromatographic process only� ����� Method Selection Systems Researchers at Virginia Tech in the USA developed ESP �Expert Separation Program� ����� This was an expert system designed to aid the inexperienced analytical chemist in choosing a HPLC method� It did not suggest what conditions were required� �� ESP began by conducting an interview to gather information on the separation prob� lem� The program then attempted to reduce the number of possible candidate solutions by decomposing the separation problem into independent sub�problems� ESP was menu driven and incorporated how and why explanation facilities� The system s architecture is described in Section � ��� ThevalidationofESPshowedthat innineoutof�� test casesESPsuggestedareasonable method� ���� ����� Retention Prediction Systems An expert system called CRIPES �Chromatographic Retention Index Prediction Expert System� was developed by researchers at Loughborough University ���� ����� The system used the molecular structure of an analyte to calculate retention indices from empirically derived quadratic expressions for the structural units� The program could also calculate the resolution of pairs of analytes� During the validation stage the retention of a number of test compounds not previ� ously examined were measured and compared with those calculated by CRIPES� �In many instances there was close agreement between the values� suggesting that generally the pre� diction method was satisfactory � � �The database does not yet include su�cient values for interaction between a wide range of substituents to permit a consistently high accuracy in the calculation of retention indices� although in most instances the predictions are within experimental error� ����� ����� Optimisation Systems Researchers from Bradford and Heriot�Watt Universities developed a system for eluent optimisation in reversed phase HPLC ���� �� � ���� ����� The input to the system was spectral information from a multichannel diode array detector which provided retention information� This type of detector �provides an extra dimension of information from a single chromatogram ���a three dimensional spectrochromatogram is produced � � � in which the axes are time� wavelength and absorbance� ����� The approach taken by the system was to perform a gradient elution experiment to determine the appropriate initial solvent strength� followed by response�surface modelling on the resulting spectral information using an iterative regression method to determine the mobile phase composition for optimum resolution� The iterative regressionapproach relied upon the correct and unambiguous identi�cation of each solute in each of the chromatograms� If this was not possible� perhaps due to co� �� elution of two or more peaks� the system then employed a modi�ed Simplex procedure which makes no assumptions based on spectral data� Peak homogeneity �the extent to which each peak corresponds to just one component of the sample� was assessed by a number of independent modules� the output from which was interpretedbythe expert systemandused tovalidate the response surfacemodel constructed by the optimisation procedure� The system required the column to be preselected� and an operator to prepare the mobile phase� operate the pumps and perform the injections the developers stated their intention to automate these tasks in the future ����� Researchers at Zhejiang University in China developed a system for optimisation of the mobile phase composition in reversed phase HPLC ����� Workers in Hungary produced an expert system for the prediction of initial HPLC con� ditions for selectivity optimisation in pharmaceutical analysis ����� The input data corresponded to the structural formulas of the compounds of interest� Using this data and a database� the system calculated partition coe�cients which were correlated with HPLC retention� ����� Troubleshooting Systems Three troubleshooting systems� HPLC�Doctor ���� CATHIE and PECOD� diagnosed prob� lems with instrumentation whereas another� Upjohn s system� diagnosed problems with methods� Workers at the Upjohn company in the USA produced a knowledge�based expert system for troubleshooting �� of the HPLC assay methods frequently used in their laboratories ����� The aim of the project was to construct an expert system that would advise an entry�level laboratory technician to diagnose problems in a HPLC system� The problem solving approach of the system was as follows�� �� It gathered the symptoms and other relevant information from the User� �� It derived the selectivity� capacity factor and resolution from the data� � It analysed the assay problem� obtaining more data if necessary� �� It diagnosed the most probable causes and then ranked them� The results were shown to the User and justi�cations given� �� In depth troubleshooting followed for the highly ranked causes in order to identify the cause of the problem� � Milne� from Intelligent Applications Limited in Livingston� produced an expert system called CATHIE ���� for the automatic interpretation of gas chromatographic data and for the provision of an expert analysis of the state of the instrument in order to detect possible failures or deteriorations� Workers at Tsingham University in China developed a prototype diagnostic expert sys� tem called PECOD� �The domain�speci�c knowledge on the performance� design and op� eration of packed extraction columns was reviewed� structured and encoded� ����� The knowledge was represented by combining a production system with conventional programs� Systems Speci�c to a Particular Class of Compounds This section describes those systems that are shown on Figure and is organised in a taxonomic fashion corresponding to it� ��� Metabolites Twoexpert systems formetabolitesweredeveloped� SPESandHPLC�METABOLEXPERT� SPES ���� was an expert system for capillary gas chromatography analysis of human ester metabolites� HPLC�METABOLEXPERT����wasdevelopedbyresearchersat theHungarianAcademy of Sciences and the CompuDrug company� The system became commercially available from CompuDrug� The system simultaneously predicted the metabolites of an organic compound and their retention data� The system was developed because HPLC had acquired a major role in the identi�ca� tion of metabolites in medicinal chemistry which had been growing in importance� The identi�cation procedure was complicated by the problem that compounds with unknown or incompletely known structures and retention times needed to be identi�ed in a chro� matogram that contained mainly the peaks of non�metabolites� The User had to enter the parent compound via a graphical interface� The system then output a tree�like picture which represented the metabolic transformations� This was produced using a knowledge base of generalised metabolic transformations taken from a standard text� Toobtain the retentionpredictions the Userhad to input the chromatographicconditions and retention times of the parent compounds� On requesting the HPLC retention prediction for metabolites� the necessary mobile phase composition and predicted retention times were displayed together with a note about the necessary pH and detection wavelength changes� � The prediction was based on the structural di�erences between the parent compound and the metabolites� the system used a database that contained the retention changes caused by a substituent of the molecule that appeared or disappeared in the physiological metabolic route� The power of the expert system to predict retention was investigated by measuring synthetic mixtures of parent compounds and metabolites� The di�erences between the measured and predicted retention times was always less than eight minutes and the average di�erence was ��� minutes� ��� Steroids Two similar expert systems for planning separations of steroids by HPLC were designed by workers in Singapore and Australia �� � ����� The �rst system that was developed tried to select an appropriate separation system for a given number of compounds� The User had to input a list of compounds which s�he suspected to be present in the sample� and a list of compounds which had to be especially well separated� Candidate separations were created by�� �� establishing the characteristics of the sample� that is the compounds of interest and their polarity� �� �nding a stationary phase of similar polarity� � �nding an eluent mixture of opposite polarity� �� �nding a compatible detector� However this strategy would have produced a large number of candidate solutions so the expert system imposed constraints� These constraints were determined from the list of compounds which had to be especially well separated and other additional queries made to User during the consultation� Candidate solutions that did not meet the constraints were ignored� The system �rst searched for a case in the knowledge based in which some or all of the compounds of interest in the sample had been separated� If this failed it searched for a case where similar compounds or class of compounds had been separated� Finally if this failed then the system planned the separation from �rst principles� The implementation described was limited to quite simple separations� However this lim� itation was due to insu�cient information in the knowledge base rather than to de�ciencies in the system structure or the approach to problem solving� �� � �� This system is referred to in this paper as the the decision tree system for steroids� as explained in Section ������� The other system for planning separations of steroids was designed to tackle the problem in the same way but using a di�erent formalism to represent the knowledge� Section ������ explains why this was attempted and discusses the results� This system is referred to in this paper as the APN �Augmented Planning Network� system for steroids because it used ATNs� ATNs are discussed in Section ������� ��� Biological Fluids Workers at Glaxo have investigated the feasibility of building an expert system for method development in HPLC for bioanalysis� that is the analysis of drugs in biological �uids ����� They developed a structured approach to method development in this area and they en� visaged that the data resulting from it would be used to build the knowledge base for an expert system� ��� Proteins Protein puri�cation involves several stages from the extraction from the source to the �nal puri�cation with high�resolution chromatography techniques� Three expert systems for planning the puri�cation of a protein were developed� PROTEIN� PPA and P�� A prototype expert system called PROTEIN was developed at the University of Reading to assist in the preliminary selection of operations for the recovery and puri�cation steps in the manufacture of proteins� The aim was �� � �to design large scale protein puri�cation processes with a very high recovery� a virtually pure product and minimum cost �� �Expert knowledge was obtained partially from the literature but mainly from industrial experts � � �� ����� �The work showed that expert systems can be a helpful tool to assist in solving the knowledge intensive and heuristic based problem �� �� ����� However the system could only amplify the e�ectiveness of an expert� it could not replace the expert� Workers in Sweden developed a knowledge�based system called PPA �Protein Puri�ca� tion Advisor� for planning protein puri�cations ����� PPA was developed because most researchers had been adopting a trial�and�error ap� proach to protein puri�cation ��� �without thought for improvement or optimisation of the method� leading to unpredictable and irreproducible results� and rapid exasperation� ����� PPA was completely User driven in that the Users told the system what they knew about the sample� the source material� and the protein they wanted to isolate� and then �� asked the system for a puri�cation plan� The system then presented a plan� The idea of PPA was not to get a �perfect plan but rather to avoid making the most obvious and time consuming errors� PPA was intended to support Users with a varying degree of experience� it was designed to facilitate learning by graduate students and for senior researchers seeking a second opin� ion� Not only could the User request PPA to formulate a plan but s�he could also describe a complete or partial plan and ask PPA for comments� Further the User could ask for alternative plans and ask why or why not a certain operation had or had not been chosen� PPA was validated by testing it with �� di�erent proteins selected by scientists� The resulting plans were discussed with the respective scientists� �In every case the plans were found to be acceptable with some minor variations� ����� When compared to published procedures for the proteins the plans were found to be very similar� Three of the people who worked on PPA also worked on P�� P� was another system for planning protein puri�cations but P� used a di�erent planning technique to PPA and was designed for a di�erent type of User� P� was a planner ��� �that planned and provided recommendations for the chromato� graphic stage of protein puri�cation� The system was developed to meet the requirements for information and decision support among workers in biochemical research laboratories�� ���� P�providedadviceonwhich technique to use� the running conditions for eachoperation� and measurements to be taken between each puri�cation step� The system was restricted to liquid chromatographic techniques and membrane�bound proteins� P� was primarily intended as a research prototype to explore knowledge representation and algorithms for protein puri�cation planning� Thus only ��� �a limited e�ort was spent on the User interface� ����� The techniques for protein puri�cation a�ect the physical properties of a sample di�er� ently but a protein sample requires carefully controlled conditions� Thus a sample has to be adjusted between the steps of a puri�cation but these adjustments need to be kept to a minimum� One of the aims was that P� would design a plan that involved the minimal number of adjustments� P� sknowledgebasewas structuredaroundaset of partial plans� Thesewere represented by a �technique and decision�point tree � which covered di�erent outcomes of operations in the plan� The knowledge in each decision�point was represented as a rule set� The nodes of the tree corresponded to puri�cation techniques� P� acted as follows�� �� P� began by asking questions about the detergent� the amount of sample and protein stability� A rule base then determined what strategy to use� A strategy was essentially �� an ordering of partial plans� �� For each partial plan puri�cation techniques were selected� using the corresponding decision tree� until the User acknowledged that the target protein was su�ciently pure� For each puri�cation technique that was selected �� �a� A check was made that the sample met the requirements for the technique� Con� straintviolationswerepassedontoamodulewhich recommendedthe adjustments that were necessary before the technique could be employed� �b� The running conditions were then selected using a rule base� �c� The technique was then performed� � Once the target protein was su�ciently pure P� halted� Section � �� contrasts the two planning techniques of PPA and P�� ��� Pesticides Researchers at the University of Missouri produced an knowledge�based system called PIA �Pesticide Identi�cation Assistant� ���� designed to assist an analytical chemist with the identi�cation of pesticide residues in food products using gas�liquid chromatography� ���� The intention was that PIA would be able to identify approximately �� di�erent poten� tial residues found in � � di�erent food items based on retention data� analysis information and heuristic rules based on food types� The analysis information was used to constrain possible identi�cations of a chromatographicpeak� There was a possibility that a chromato� graphic response under investigation was due to a residue from outside the set of compounds that the system recognised� Therefore a de�nite identity could not be established and a va� riety of heuristic rules allowed the selection of the most probable identity from among the possibilities� Help functions were provided throughout the system� Explanations and sug� gestions were only available after the probable identi�cation was made� ���� PIA operated by using Prolog s inference engine directly to perform a number of searches ���� �� �� Prior to attempting residue identi�cation� it performed a search based on the food item in question� This had the e�ect of gathering heuristics relevant to the sample type into working memory� �� It tried to identify the residue�s� by�� �a� searching for a single residue that could account for all the data� �� �b� searching for multiple residues� if no single residue could account for all the data� and �nding a possible residue for each peak on the chromatogram� Rather than always performing a complete search of possible residues� the system �rst performed a search among only those residues previously identi�ed in the sample food type� continuing on to a complete search only if the restricted search failed to generate an identi�cation� � Finally it searched for information to use in explanations of results� and for suggestions for con�rming analyses� However the system was incapable of expanding upon these suggestions or explaining why they were appropriate� �Currently� eleven of the � � food items are represented in the system� as well as �� of the �� possible compounds� Identi�cation of items drawn from this initial sample is ����� Obviously� such performance will degrade as the complete database of food items and compounds is incorporated into the system�� ���� �� Fossil Fuels Four of the expert systems for chromatography were limited to fossil fuels� Two of these tried to automate the interpretation of data from gas chromatography�mass spectrometry� This is a highly skilled� labour intensive task which is used to identify compounds in fossil fuels� Researchers at the Georgia Institute of Technology and the University of Alabama� both in theUSA� suggesteddesigncriteria foragaschromatography�massspectrometry�GC�MS� based expert system for arson analysis� Since petroleum distillates are �the overwhelming choice of the arsonist� ���� a system speci�c to this group of compounds was designed� �Advantage was taken of the high selectivity of mass spectrometry towards aliphatic and aromatic hydrocarbons� ����� �In the USA about one out of �ve reported �res is of suspicious origin� � � �The chemical analysis of residual accelerant is an integral part of most investigations into the origin and cause of a �re� � � �The power of GC�MS in accelerant analysis has been demonstrated con� vincingly on real world samples�� ���� The system was developed because of the limitations of human chromatogram interpretation� Data interpretation had almost always been car� ried out by pattern recognition of chromatograms� This had involved the analyst looking for groups of peaks that were characteristic for fuels commonly encountered� In many cases� chromatogramsof suspect samples had been simply arranged side by side to chromatograms of standards� There are many human and technical limitations of chromatogram interpre� �� tation� For example it was obvious that factors such as retention time shifts or a mismatch of the chromatographic scale could have caused problems� ���� Thus a system was designed which tried to automate the process� The system began by acquiring GC�MS data� It tried to categorise chromatographic pro�les by establishing the presence of key components� Comparison was made to data from standards stored in the memory of the computer� Retention time information as well as spectral criteria was used� The system then printed a table giving extensive information about the presence or absence of the key compounds and denoting the accelerant standard which was the best match to the unknown sample� The system could then be instructed to print out ion chromatograms for the sample and simultaneously display the best match� ���� The researchers were aware that their system had limitations� �Chemical analysis of �re debris remains a challenging task� There is no substitute for human intuition and a machine will never replace a chemist� The program does fail on occasion for samples that produce overwhelming interferences or have other peculiarities� � � �The program �� � is of signi�cant help in routine analysis�� ���� Workers in the USA at the University of Houston and at the company Exxon developed an expert system for the interpretation of the GC�MS spectra of fossil fuel distillates� Interpretation �� � �of GC�MS spectra is still very labour intensive� in particular for non�routine samples containing hundreds of components� such as fossil fuel distillates � � � � Manual identi�cation � � �requires a highly skilled professional and is extremely time consum� ing�� ���� Even when a commercially available data system is used� identi�cation �� � �has to be based on a large number of experienced based� inter�connected decisions� that include a detailed examination of the mass spectra� considering the relative order of elution and a very critical evaluation of the library response� ����� The Exxon system used the commercially available identi�cation capabilities but it ex� tended them by de�ning limits of acceptability for the library identi�cation yielded by those systems� These limits were based on automatic examination of the spectra� on retention time considerations and on the quality of the matching itself� The system was shown ��� �to be highly reliable in correctly identifying components in complex hydrocarbon streams� and doing so at a fraction of the time that would be required by a human expert� ����� Two expert systems for diagnosing the insulation condition of power transformers were developed� TOGA �� � and a system from Xian Jiaotong University in China ����� The latter worked from the measured chromatograph for the transformer oil� The insulation condition of power transformer was diagnosed as either normal� in need of checking and repair� or the system recommended that the diagnostic period needed shortening� �� Systems Speci�c to a Particular Chromatographic Method This section corresponds to the branch of the taxonomy which is entitled �Speci�c to Chro� matographic Method on Figure �� �� Systems for Ion Pair Chromatography Three expert systems were developed for ion pair chromatography� One suggested a suitable set of method conditions whereas the other two were mainly concerned with optimisation� The expert system for suggesting a set of suitable method conditions was developed at the University of New South Wales for ion chromatographic methods using dynamically coated ion�interaction �also known as ion�pair� chromatography ����� Ion�pair chromatography involves the use of apolar stationary phases with eluents con� taining a hydrophobic ion of opposite charge sign to that of the analyte ions� that is a counter ion� Two alternative methods exist� �permanent coating or �dynamic coating � The former requires that the column is �rst equated with the counter ion� which is then absent from the eluent during the analysis step� The latter uses an eluent containing the counter ion for both the column conditioning and analysis steps� The knowledge acquisition process for the expert system revealed that permanent coating represented less than ��� of ion�pair applications and it was therefore decided to concentrate on the dynamic coating procedure� The actual rules for de�ning the method conditions were divided into four stages�� Selection of the method type This part ascertained the class of ion to be analysed and selected a suitable preliminary method� Selection of the column The column was selected by examining the nature of the sample and its matrix� Selection of the eluent The eluent was determined chie�y by the class of ions for anal� ysis� Selection of the detector The detector was chosen on the basis of the properties of the solute ions and the availability of detector types� A report was shown to the User which summarised the chosen method conditions� On viewing this it was possible for the User to modify any answers and the system would produce new conclusions� Throughout the consultation help was available for each query that was made of the User� The system allowed the User to change the conditions for activating some of the rules and the conclusion made by these rules� If any such modi�cations were made then the rule �� base was recon�gured to account for these changes� Thus the User could customise the system to include speci�c preferences� Researchers in Brussels developed one of the two systems concerned mainly with optimi� sation ����� The main purpose of the research was to explore how to introduce optimisation procedures into the expert system rather than to build a system for the domain� Thus only some basic drugs were considered� only one ion�pair reagent was investigated and the detector mode was limited to UV� The system was based on LABEL� Like LABEL it gave advice on the detection and mobile phase but it also included an experimental design module� The systems comprised four modules�� Introductory Module This checked whether ion pair chromatography was necessary and possible� and selected the detector settings� First Guess Module This advised on the mobile�phase conditions� It began by deciding whether a �rst guess was indeed possible� If the situation was too complex then it decided that an experimental design should be applied immediately� Optimisation Module This used an experimental design when an initial guess approach was likely to give unsatisfactory results� that is when more than one substance was present� A simple design was chosen because the aim of the expert system was not to establish the best optimisation method but to investigate the feasibility of intergrating optimisation methods in to an expert system� The User then tested the method and entered the results in to the system� The system evaluated the results and either advised that that the method be used� advised that a good separation could not be obtained using the system selected or adapted the method by calling the fourth module� Adaptation Module The input to this module was the chromatographic results of the method proposed by the �rst guess or optimisation modules� The module either adapted the method or advised the User that a good separation could not be ob� tained using the system selected� The system was validated by following the advice of the system for sixty synthetic mixtures created using a random number generator� �The rate of success was satisfactory� ����� A rulebased systemwasdevelopedat TexasA M University in the USA for the determi� nation of solute types in unknown sample mixtures as a �rst step of selectivity optimisation parameter selection in reversed�phase ion�pair chromatography ���a�� �� This work illustrated that the nature of the solutes in totally unknown aqueous sam� ple mixtures can be determined by the combination of an experimental procedure and a rule�based retention�shift evaluation strategy that is implemented in a computer pro� gram� The approach suggested did not require peak tracking or the use of any other extra�chromatographic information� Other expert systems for selectivity optimisation need extensive a priori chemical information about the sample components in order to make predictions about the expected retention behaviour� The system comprised three parts�� Data Input The User entered the retention times of all the peaks observed� together with the respective pH and pairing�ion information� The retention times and chro� matograms could be entered in any order� Type Evaluation The program tried to label each peak as either SA� SB� WA� WB or N where N!neutral� S!strong� W!Weak� A!acid� and B!base� The retention data of all peaks from all the chromatograms were compared� �At least two chromatograms were required�� Initially all �ve possible solute types were assigned to all peaks in all chromatograms� Then� recursively using the retention�shift rule set� the program eliminated the impossible solute�type designations for all the peaks� List Results The solute�type assignments were then displayed to the User� The User could request the list of rules �and their hierarchical sequence� which were used in the solute�type assignment process� for each individual peak� When tested with a variety of complex samples� the program correctly identi�ed the type of the �rst and last eluting peaks and excluded the impossible solute�type designations for the rest of the peaks� �� System for Thin�Layer Chromatography A system was developed for the identi�cation and separation of samples by thin�layer chro� matography�TLC� ���b�� The system was based on an extensive database of pharmaceutical compounds which could be modi�ed by the User� The system was useful for TLC screening and for selecting a TLC separation system for a given sample� ��� � �� Synopsis of Behaviour and Users Table � compares the predominant behaviour of expert systems for chromatography � and the types of User for which they were intended� It indicates that only three out of the �� systems were targeted at Users who were novices in the particular domain of each of the systems� Thus most expert systems for chromatography were designed for scientists with some experience of chromatography� The table also shows that the systems exhibit four predominant behaviours� just over a third of them �� of ��� plan the whole or part of a chromatographic separation� whist the remainder either predict the values of experimental data� diagnose problems or identify sample components� �� How the Knowledge was Acquired for the Systems The process of acquiring the knowledge needed for an expert system is called knowledge acquisition� The knowledge acquisition process is usually divided into three stages� deciding what knowledge is needed� variously referred to as the de�nition stage or initial analysis getting knowledge� predominantly from human experts� and interpreting it� usually called elicitation and �writing the knowledge in the internal language of the system� encoding it� usually called representation� Knowledge acquisition is a notoriously slow process and has become known as the �bottle�neck in the process of developing expert systems� ����� It has accounted for a signi�cant proportion of the total work involved in developing expert systems for chromatography� For example the knowledge acquisition phase for the PPA project required approximately six person�months of expert time and in terms of calendar time the project required one year� Future developers of expert systems for chromatography should allow for the knowledge acquisition �bottle�neck when planning their projects� ���� Sources of Expertise and Knowledge Some of the systems relied on the literature as the main source of expertise� This was the case for the expert system for planning separations of steroids that was implemented as a decision tree the rules were induced from speci�c examples of validated results of successful separations reported in the literature� Examples of how the expert system should behave were taken and then generalised to a higher�level rule that could guide the inference procedure �� �� ESP used rules drawn from a standard text book ����� �The systems that were designed for several stages of the chromatographic process �ESCA� ECAT and ESC�ESLC� are not included because they do not have a predominant behaviour they comprise several subsys tems which behave di�erently� � However the knowledge and expertise for most of the systems was acquired with the help of one or more experts� For example the sources of expertise for the PPA system were specialists on individual separation techniques and researchers on particular classes of proteins ����� The structure of the knowledge for ECAT was � determined through informal interviewingof experts� andbyaccumulationof ideasandexperience from chromatographers through the literature� ����� One of the problems of working with experts is that they are usually busy people with little spare time to devote to the development of an expert system� The developers of some of the systems found ways of minimising this problem� For example the researchers� based in New South Wales� that developed the expert system for ion chromatographic methods using dynamically coated ion�pair chromatography ���� developed their system from an extensive database of previously published ion chromatography methods� �The database was searched systematically to �nd the most commonly used method conditions for di�erent applications� These conditions were then examined by the expert and rules generated for the expert system ���This method considerably reduced the amount of time required from the expert but still resulted in the development of a competent expert system�� ���� When help is sought from experts for an expert system project it is important that the most suitable experts are chosen� For a company project this may involve deciding whether to use experts from within the company or from outside� Such a decision was made by the developers of the Upjohn troubleshooting system� The sources of the expertise for the Upjohn troubleshooting system were publications and chromatographic experts in the Pharmaceutical Quality Control Division of the Upjohn company ���� since the system was to be an �in�house system� �in�house expertise was adequate� Using more than one expert in a project can lead to problems� Experts may disagree or use di�erent approaches� For some of the systems the number of experts involved was deliberately limited to one� The developers of CRISE stressed the importance of consistency in the knowledge� They achieved this for CRISE by obtaining almost all the knowledge from a single expert� Knowledge from another source was included in the periphery of the system� ���� In the ESCA project four experts separately contributed knowledge for each of the four subdomains� �Although more than one expert could have contributed to a domain� it was decided to avoid any discussion between experts������ Thus previous experience suggests that developers of expert systems for chromatography must carefullyconsider thenumberandchoiceof experts tobeconsulted� anddecidewhether �in�house or external expertise is to be sought� �� ���� Elicitation Techniques Employed For many of the systems the literature does not describe which elicitation techniques were used� However interviewing was used for the majority of those for which the techniques used are described� For a few systems other techniques were employed but they were not described in any detail� For example the knowledge elicitation for the ESCA DASH system was mainly done by interview but other techniques were also used� However they were only described in the vague statement �Sometimes information was exchanged in written form� ����� The literature on knowledge engineering describes in detail a wide range of techniques for knowledge elicitation that can supplement or replace interviewing ���� ���� and gives advice on how and when a given technique should be employed� However the literature on expert systems for chromatography does not mention any of these techniques� Despite the emphasis placed on the techniques by the literature on knowledge elicitation there is no evidence that developers of expert systems for chromatography have used them� considered but rejected using them or were aware of them� Future developers of expert systems for chromatography should consider using the various techniques and describe and justify their choice of technique�s� in any published work� ���� Machine Induction and Neural Networks There is an alternative to the developer of an expert system trying to elicit the expertise from an expert a computer can be programed to develop the expertise� This can involve a computer being trained to develop a set of rules or a neural network� Rule Induction Rule induction involvesa computer system being provided with a number of cases and the system using inductive techniques to create its own rules from the cases� Neural Networks A neural network comprises a layer of input elements connected to one or more hidden layers of elements� the last of which is connected to a layer of output elements� Data from the input layer are transmitted through the hidden layer�s� to the output layer in a manner determined by the strengths of the connections between elements� Training modi�es the connection strengths� and it is the set of their values in the trained network which represents the relationship between the input and the output� There are no reports in the literature to suggest that induction was considered as a method for acquiring the expertise for any of the expert systems for chromatography that �� had been implemented� In the conclusion of the paper on the expert system developed in New South Wales for ion chromatographic methods using dynamically coated ion�pair chromatography ���� the authors stated their intention to investigate the potential of neural networks and the Quinlan rule generating system ���� for knowledge acquisition� They proposed to compile rule bases for the ion exclusion and ion exchange methods of chromatography They stated their belief that since these methods are used more frequently the task of manually per� forming a statistical analysis on a database of literature methods� as was done for the ion pair mechanism� would be extremely di�cult due to the large amount of data� �� The Knowledge Representations Adopted ���� Formalisms for Knowledge Representation The literature on knowledge representation describes a number of formalisms that can be used such as productions rules� frames� semantic nets and the predicate calculus ���� ����� � � includes lucid examples showing how all four formalisms can be used to represent chemical knowledge� Nearly all the expert systems for chromatography used rules� Despite the emphasis placed on the alternatives to purely rule based systems by the literature on knowledge engineering ���� ���� there is no evidence that developers of the rule based systems for chromatography considered using them or were aware of them� Future developers of expert systems for chromatography should consider using the various formalisms and describe and justify their choice of formalism�s� in any published work� The following sections discuss those few systems which did use other formalisms� ������ Frames Frames �� � �� � �as data structures for storing expectations about typical objects and events have become quite widespread in arti�cial intelligence applications� p��� of ����� The �� � �additional structure which can be represented or imposed by frames has considerable value� One demonstration of this is the reconstruction of knowledge bases originally ex� pressed in unstructured production rules into frame systems and the consequent improve� ment in system understanding and ease of maintenance� � � �The use of frames to represent knowledge about structured domains or structured knowledge continues to increase in popu� larity � � �� p�� of ���� Frames ��� �are mostly used in conjunction with other representations� such as production rules� p��� of ����� �� It is therefore surprising that the use of the increasingly common combination of produc� tion rules and frames was only reported for a few of the expert systems for chromatography� The ESCA system for precision testing was one of these� The usual features of frames such as instantiation� inheritance and demons were used � �� � �� ����� ������ Augmented Transition Networks The two systems for planning separations of steroids were designed to tackle the same problem in the same way but using di�erent formalisms to represent the knowledge� The �rst representedadecision treebya series of rules� The secondusedan augmentedtransition network �ATN� � and was developed to show that the ATN formalism was ��� �a more e�cient structure for representing the knowledge base of the � � �system� ����� However no evidence was presented that proved this� Furthermore the comparison of the ATN and rule formalisms ignored the possibility of using some of the main control strategies for rule based systems� such asallowing rules to be �red more than once� and using meta and context sensitive rules� These omissions render the work as irrelevant to the contemporary debate as to which formalism is most e�cient for knowledge representation� The work failed to prove that ATNs should be used for future expert systems for chromatography� ���� Uncertainty Uncertainty exists whenever expert systems have to make non�categorical decisions� Un� certainty can be represented using probability theory� certainty factors� fuzzy logic and the Dempster�Shafer theory of evidence ����� The literature mentions the representation of un� certainty in only a few of the systems� The Upjohn troubleshooting system developed by ���� was implemented in M�� which is rule�based expert system software which deals with uncertainty by the use of MYCIN style certainty factors� The capability of M�� in dealing with uncertainty was used to weigh evidence and thereby establish priority for e�cient trou� bleshooting� PROTEIN used certainty factors as well� PIA handled uncertainty �� � � in an ad�hoc fashion� basically by noting any causes for uncertainty in working memory� assigning a subjective weight� and accumulating the weights associated with each proposed identi�� cation� ����� The Pascal reimplementation of the ESCA REPS also incorporated weighting factors ����� �For a description of ATNs see � ��� �� ���� Representations not Speci�c to Expert Systems Some standard software development representations were used� Data �ow diagrams and state transition diagrams were used for the ESCA REPS � ��� Flow charts were used in the development of ESC�ESLC ���� ����� In academic computer science circles �ow charts are associated with poor implementation practice ����� They �� � �are very limited in their capability for modularisation and structuring � � �� ����� It is therefore surprising that �ow charts were used in the late ����s to design a computer program� especially one which was an expert system� Flow charts should not be used to design future expert systems� �� The Software Tools and Architectures Used This section lists the software tools and architectures used to implement expert systems for chromatography� ���� Software Languages and Expert System Shells and Tools Table � shows which expert systems for chromatography used which languages � and the aspects of the each system for which a language was used� The table shows that those systems in which the coding was not in one language alone� the AI languages were mostly used for reasoning� and never fornumericalprocessing� whereas conventional languageswere mostly used for numerical processing and never for reasoning� This re�ects the strengths of the AI languages� they make it possible to represent knowledge quickly and conveniently p�� of ���� making it easier to code reasoning� The way in which the conventional languages were used also re�ects one of their strengths� they usually provide better facilities than the AI languages for numerical processing� The aim of Table is to show which expert system shells and tools were used in the development of the systems� � �Prolog and LISP are referred to as arti�cial intelligence �AI� languages� Third generation languages� such as BASIC� FORTRAN� C and Pascal� will be referred to as conventional languages� �It does not try to review the shells and tools used or to classify them� they are presented in alphabetical order� No description of the shells and tools is given because of the di�erences between the various versions used for di�erent chromatography systems� �� ���� Hypermedia The developers of CRISEBOOK � �� discussed three problems with �traditional expert systems development software�� �� Updating and maintaining systems based on this type of software is di�cult� �With the conventional present shells it is di�cult for the User to make changes owing to the way rules are connected � � �Adding or changing rules could have repercussions all over the system�� � �� �� To use some of this software requires advanced skills of knowledge engineering� � Systems based on this type of software often have a poor User interface� The developers of CRISEBOOK investigated whether Hypermedia could be used as an alternative to conventional expert systems development software� �Hypermedia is the combination of multimedia and hypertext� In a hypertext document the User does not have to read all information in a sequential way� but is guided by his needs and interests by pieces of information that are linked to each other� Multimedia is a collection of tools producing graphics� sound� and animation� to present data in a more �exible way�� � ��� The work on CRISEBOOK did not show whether Hypermedia would make update and maintenance of expert systems easier� The work was inconclusive as to whether advanced skills would be needed to build expert systems with Hypermedia� although the developers of CRISEBOOK reported that the �scripts that are developed for Hypermedia are �� � �entirely readable even for those unfamiliar with programming�� � �� CRISEBOOK suggested that Hypermedia would enable better interfaces to be produced because data can be presented in a more �exible manner and because much of the programming associated with interfaces is already present in Hypermedia� Problemswith expert system development software can not be considered independently� Changing the development software to ease one problem may a�ect the severity of others or even create new ones� Although CRISEBOOK showed that Hypermedia could help to ease two of the afore mentioned problems� it did not prove that Hypermedia would be a better alternative overall to �traditional expert systems development software� ���� Spreadsheets Twosystemsweredevelopedbycombiningexpert systemsdevelopment softwareandspread� sheets packages� �� In REPSspreadsheetswereprogrammedto produce the necessarystatistics� The spread� sheet package used was Lotus ���� ������ �Lotus Development Corporation� Cambridge� MA� USA� the expert system tool used� Goldworks� was able to communicate with this� Rules� de�ned using Goldworks� set up spreadsheets into which the analyst input data� and the rules then interpreted the data processed in the spreadsheet� � �� In CRIPES VP�Exert communicated with spreadsheets written in VP�Planner �Paper� back Software� ����� ���� Architectures It was vital for the proposed integration of the ESCA repeatability and ruggedness systems that all the modules of the integrated system used the same concepts as the basis for reasoning so that �exible communication was possible� Simple transfer of �les between modules would have been insu�cient for two reasons � ���� �� Most of the facts produced by one module had to be available to all other modules� �� The modules were not to be consulted in a standard sequence� Trying to implement all possible consultation sequences would have become unrealistically complex� It was decided to merge all the existing concepts in one common data structure that formed the basis for all the systems� �The blackboard architecture was chosen � because it allows integration of modules which use di�erent interfacing or problem solving techniques� � ��� �Blackboardsarewell knownarti�cial intelligence techniques for the integrationof expert systems� � ��� They allow several modules to communicate with each other via a common data structure� the blackboard� In a blackboard architecture ��� �the knowledge sources trigger themselves when the state of the blackboard is such that they can contribute to the solution of the problem� � � � In an ideal situation several knowledge sources can be activated at the same time� The blackboard architecture o�ers the possibility for parallel processing�� � �� The ESCA integrated system did not utilise the possibility for parallel processing o�ered by a blackboard� In fact all but one of the expert systems were designed for serial processing� The exception was ESP which was implemented as two concurrent processes� The interface� implemented in C� was a parent process and the �intelligence of the system� the Prolog �The architecture used for the ESCA integrated system varied slightly from a traditional blackboard in that a supervisor module controlled when and which knowledge sources were to be activated� �� code� was the child process� Calls to the UNIX operating system kernel allowed concurrent processes to execute� communicating through UNIX pipes� ���� ���� Expert Planners Some of the systems can be thought of as planners� For example PPA was a planner that planned a puri�cation process from the original source to the �nal product� Thus it was a deductive planner as opposed to a reactive planner� Reactive planners� such as P�� immediately respond �without much deduction� to input without forming a complete plan on how the goal should be reached� A deductive planner was developed because reactive planners are �faster but in many senses �dumber than deductive planners� ����� A deductive planner had the advantage that it was possible to plan in advance and� where possible� optimise the plan� and the ability to deal with unanticipated puri�cation tasks by reasoning from general rules� However the designers of the system were aware of the major drawback of deductive planners� they can not provide dynamic responses to di�erent outcomes of applied operations� P� was designed as a reactive planner because reactive planners do not have this drawback� �Armed with a certain amount of knowledge � � �one can devise a separation plan in advance� The e�ciency of the plan will depend on the actual outcome of the individual steps� When the results are known� one is better equipped to plan subsequent steps� ����� �� Validation and Evaluation �The increasing application of expert systems in routine tasks� and the consequent gradual automation of knowledge intensive work� has led to a need for some kind of assurance of the quality of � � �� p��� of ���� the resulting systems� Hence expert systems need to be validated and evaluated� ���� How the Systems were Validated Validation involves testing by the developers to estimate the faithfulness with which the expert knowledge has been captured� This involves asking questions such as the following� Does the system o�er advice that is appropriate as judged by experts" Is the reasoning correct� consistent and complete" Are the explanations adequate" Do the test problems cover the domain" Validation is an important stage in the development of an expert system� Unfortunately� due to the nature of expert systems� the standard principles and practices of software �� engineering for validation are not directly applicable� The workers on the ESCA project stressed this� �Expert systems can only be expected to be useful in practice if they are reliable� Conventional software products can be tested thoroughly through a number of standard procedures� However� for testing expert systems no standard procedures exist� ����� The validation of only a minority of the expert systems was reported in the literature� For those systems for which a validation was reported one of three approacheswas adopted�� �� Advisory systems for planning were validated by comparing the advice given by the systems with the advice given by experts� For example RES was validated by using �� test cases� each of which was a liquid chromatography method for the separation of a pharmaceutical sample� For each method an unbiased expert s factor choice was listed before the expert system was consulted� The factors chosen by the expert and system were then compared� A slight variation of this approach was used for PPA� Various domain experts were asked to pose problems for the expert system and then decide whether the resulting advice given by the system was acceptable� Twenty di�erent proteins were selected by as many researchers� PPA was then run for each one to produce a plan for its puri�cation� Each of the plans were then discussed with the respective scientist� ���� �� Advisory systems for novice Users were validated by comparing the performance of the User with and without the help of the system� For example the designers of ESP suggested that �Future � validation could take place by asking two similar student groups to solve a set of representative separation problems with and without the aid of ESP� ����� � Predictive expert systems were validated by comparing the the results predicted by the systems with corresponding experimental results� To validate CRIPES the retention of a number of test compounds not previously examined were measured and compared with thosecalculatedbyCRIPES�ThepredictivepowerofHPLC�METABOLEXPERT was investigated for seven parent ��� �molecules and their eight main metabolites by measuring their reversed�phase retention data using various reversed�phase columns and conditions� ����� The measured and predicted retention data were then compared� All three approaches involved using some metric of similarity to decide whether two sets of results were similar enough to allow the system to be judged a success� For example for �The performance of ESP had been measured by a HPLC technician answering ESP�s interview questions for a series of compounds from the literature� ���� � RES the criterion for deciding whether or not a factor choice made by the expert system was acceptable was that the choices of the expert and expert system did not di�er by more than two factors� The number of factors selected was allowed to di�er only by one� the number of factors was important for the �nal acceptance of the system because the number had to be small but su�cient� All the approaches to validation required that a set of tests was selected� The tests were selected in di�erent ways�� Random sets The tests for the validation of LABEL were randomlychosen� �Initially � �� pharmaceutical formulationswere selectedat random fromtheBelgiumDrugrepertory ����� ����� Sets chosen by various experts The set of tests for the validation of PPA was chosen by a number of domain experts� Representative sets The compounds selected from the literature for use in the validation of ESP �see above� were ��� �representative of cases covering the range of possible conclusions that ESP can arrive at� ����� Sets including speci�c tests For example the compounds used in the trials of CRIPES ��� � included some that were selected to test speci�c aspects of retention prediction� ����� The �random sets and �sets chosen by various experts can considered as black box testing and the others can be viewed as clear �or white� box testing� A notorious problem encountered when developing software is that when it is repaired� following the discovery of an error� the repair itself may introduce new� undetected� errors� This is particularly relevant for expert systems incorporating production rules where adding or changing rules can have repercussions all over the system� The developers of the ESCA expert systems were conscious of this when planning the validation of their systems� To ensure that changes to a knowledge base did not produce unexpected side e�ects� it was necessary to de�ne a so�called regression test� The expert selected a number of test cases� representing a broad range of possible cases� that were solved by the expert system after each major revision� If the solutions to the test case did not remain the same it was assumed that changes had been made that adversely a�ected previously evaluated knowledge� ���� �The limited nature of the regression test is of course no guarantee that unexpected results will not appear but the choice of a good regression test set will minimise the chance� ����� Six of these were ignored because they contained compounds that could not be determined by UV or elec trochemical detection� �� � � ���� How the Systems were Evaluated Evaluation is an investigation of the extent to which the system is useful and acceptable to the users� It involves asking questions such as the following� Is the input and output convenient" Is the nature of the interaction appropriate" Is the system fast enough" Does it �t in with the normal work pattern" Does it threaten the User" Is it being used" Will it be used" Have the goals been achieved" Are there hidden costs" Given that these questions are so fundamental it seems remarkable that for all but one of the expert systems for chromatography there were no reports in the literature to suggest that any attempt had been made to evaluate the systems� Unless the results of evaluations are published it is impossible for people who were not involved in the development of a system to know how e�ective a given system was� Future developers of expert systems for chromatography should evaluate their systems and publish the results� Only the evaluation of the ESCA systems was reported in the literature� It involved testing the expert systems in practical situations in order to evaluate their performance in daily practice� Generally� these tests were performed by external evaluators ����� An example of such an evaluation� for two of the ESCA subsystems DASH and DASH�� is de� scribed in ���� The literature on the ESCA evaluations included some interesting comments on the process� For example the evaluators of REPS felt the measured performance was �� � �acceptable for this method� but were pleased to �nd that the system suggesting actions that could further improve the method� This is a typical advantage of expert systems over conventional software packages� � � �The evaluation of this system proved extremely valuable as it resulted in several additions which enhanced the software considerably� ����� The ESCA evaluations were carried out by di�erent persons� ranging from experts in method development to students with little or no experience� During the evaluation phase of all the expert systems it became clear that the attitude towards expert systems is strongly dependent on the expertise level of the evaluator� The accessibility of the specialists ex� pertise was clearly appreciated by inexperienced Users� Experienced Users could appreciate the quality of advice given by the systems� However when the strategy implemented in a system did not agree with the expertise of experienced Users they became dissatis�ed with the system because their own experience� probably better adapted to their speci�c situation� was not considered by the system� ���� �� �� Conclusions Expert systems for chromatography have been reviewed� Most are designed for scientists with some experience of chromatography� The predominant behaviour of each is either to plan the whole or part of a chromatographic separation� predict the values of experimental data� diagnose problems or identify sample components� The domains of the systems vary in their coverage of�� � stages of the chromatographic process� � compounds to be separated by the chromatograph� � chromatographic methods� Thus the systems cover a large number of di�erent subdomains of chromatography� A taxonomy has been proposed that allows expert systems for chromatography to be classi�ed and facilitates an understanding of their inter�relationship� The literature suggests that the most successful expert systems for chromatography are those which tackle speci�c aspects of chromatography rather than those which tackle large parts of this domain� Expert system projects which tackle domains as large as that covered by the ESCA project must be managed e�ectively� particularly if the objective is to produce an integrated system� Previous expert systems for chromatography show that spreadsheets can be embedded within expert systems for chromatography but fail to prove that Hypermedia is a better alternative overall to �traditional expert systems development software or that ATNs o�er a formalism which is more e�cient than rules for knowledge representation� ���� Conclusions on the Engineering There are a number of aspects of the software and knowledge engineering that were ap� proached in a similar way for most of the expert systems for chromatography� From this emerges the following stereotypical portrayal of the engineering for a previous expert system for chromatography� The knowledge acquisition accounts for a signi�cant proportion of the total work� and involves interviewing one or more experts� Rules are used to represent the knowledge and they are encoded using an arti�cial intelligence language or expert system shell or tool� The resulting system is validated but is not evaluated� Despite the literature on knowledge engineering stressing the importance of evaluating systemsandadvocatingawide rangeof techniques forknowledgeelicitationandawide range of formalisms for knowledge representation� developers of expert systems for this �real world �� domain have not responded� nearly all of the developers use interviewing for elicitation and rules for representing knowledge and they do not evaluate their systems� Further research is needed to determine why because developers of expert systems for chromatography often do not justify their decisions on engineering matters in their published work� ���� Summary of Main Conclusions � A taxonomy has been proposed that allows present �and future� expert systems for chromatographytobeclassi�edand facilitatesanunderstandingof their inter�relationship� � The literature suggests that many of the ideas advocated by knowledge engineers are not being used by developers of expert systems for chromatography� � Too often developers of expert systems for chromatography do not justify their deci� sions on engineering matters� �� Recommendations Future developers of expert systems for chromatography should�� � Either plan their projects so as to allow su�cient time for �the bottle�neck in the process of developing expert systems� the knowledge acquisition phase� or consider using other techniques such as neural networks or rule�induction to automatically acquire the expertise� � Carefully consider the number and choice of experts to be consulted� and decide whether �in�house or external expertise is to be sought� � Consider using the various techniques for knowledge elicitation and formalisms for knowledge representation and justify their choice in any published work� At the very least they should consider the use of the increasingly common combination of production rules and frames and avoid out�dated formalisms such as �ow charts� � Evaluate their systems� in addition to validating them� and publish the results� Unless the resultsof evaluationsarepublished it is impossible forpeoplewhowerenot involved in the development of a system to know how e�ective a given system was� The funding wasprovidedbySERC�under the remitof theTotalTechnologyprogramme� and by Zeneca Pharmaceuticals� �� References ��� V�Jakus� Collect� Czech� Chem� Commun�� �� ������ ��� � ��� T�P� Bridge� M�H� Wiliams� and A�F� Fell� Chemistry in Britain� �November ����� ����� � � J�Klaessens� and G� Kateman� Fresenius Z Anal Chem�� �� ������ �� ��� A�P� Wade� S�R� Crouch� and D� Betteridge� TRAC� Trends in Analytical Chemistry� � ������ No���� ��� ��� S�A� Borman� Analytical Chemistry� �� ������ No���� ����� ��� J�L� Glajch� LC�GC � ������ No��� �� ��� T�Hamoir� and D�L� Massart� Advances in Chromatography� ���� � ��� ��� V�R� Meyer� Practical High�performance Liquid Chromatography� John Wiley Sons� ����� ��� J�A� Jonsson� Chromatographic Theory and Basic Techniques� Dekker� New York ����� ���� A� Braithwaite� and F�J� Smith� Chromatographic Methods� Chapman and Hall� London� ����� ���� C�F� Poole� and S�A�Schuette� Contemporary Practice of Chromatography� Elsevier� Amsterdam� ����� ���� P� Jackson� Introduction to Expert Systems� �nd Ed�� Addison�Wesley� ����� �� � A� Hart� Knowledge Acquisition for Expert Systems� Kogan Page� London� ����� ���� G�A� Ringland� and D�A� Duce� Approaches to Knowledge Representation� An In� troduction� Research Studies Press Ltd� Taunton� Somerset� England� ����� ���� P� Harmon� R� Maus� and W� Morrissey� Expert Systems� Tools Applications� John Wiley Sons� ����� ���� I� Dickinson� Laboratory Robotics and Automation� � ���� � ��� ���� L�R�Snyder� J�W�Dolan� andD�C� Lommen� JournalofChromatography���������� ��� ���� J�A� van Leeuwen� L�M�C� Buydens� B�G�M� Vandeginste� and G� Kateman� TRAC� Trends in Analytical Chemistry� �� No��� ��� ���� D� Goulder� T� Bla�ert� A� Blokland� L� Buydens� A� Chhabra� A� Cleland� N� Dunand� H� Hindriks� and G� Kateman� Chromatographia� �� ������ �� � �� ���� M� Mulholland� N� Walker� J�A� van Leeuwen� L� Buydens� F�Maris� H� Hindriks� and P�J� Schoenmakers� Mikrochim Acta �Wien�� II ������ �� � ���� L� Buydens� P� Schoenmakers� F� Maris� and H� Hindriks� Analytica Chimica Acta� ��� ���� � ��� ���� T� Hamoir� M� De Smet� H� Pyrins� P� Conti� N�V� Driessche� D�L� Massart� F� Maris� H� Hindriks� and P�J� Schoenmakers� Journal of Chromatography� ��� ������ �� �� � P� Conti� T� Hamoir� M� De Smet� H� Piryns� N�V� Driessche� F� Maris� H� Hindriks� P�J� Schoenmakers� and D�L� Massart� Chemometrics and Intelligent Laboratory Systems� �� ������ �� ���� F� Maris� R� Hindriks� J� Vink� A� Peeters� N�V� Driessche� L� Massart� Journal of Chromatography� ��� ������ ��� ���� G� Musch� M�M� De Smet� andD�L� Massart� Journalof Chromatography� �������� ��� ���� M� De Smet� A� Peeters� L�Buydens� and D�L� Massart� Journal of Chromatography� ��� ������ ��� ���� M� De Smet� G� Musch� A� Peeters� L� Buydens� and D�L� Massart� Journal of Chromatography� ��� ������ � �� ���� G�Musch� and D�L� Massart� Journal of Chromatography� �� ������ �� ���� A� Peeters� L� Buydens� D�L� Massart� P�J� Schoenmakers� Chromatographia� �� ������ ���� � �� B� Bourguignon� P� Vankeerberghen� and D�L� Massart� Journal of Chromatography� ��� ������ ��� � �� P�J� Schoenmakers� A� Peeters� and R�J� Lynch� Journal of Chromatography� ��� ������ ���� � �� P�J� Schoenmakers� N� Dunand� A� Cleland� G� Musch� and T� Bla�ert� Chro� matographia� �� ������ �� � � P�J� Schoenmakers� and N� Dunand� Journal of Chromatography� ��� ������ ���� � �� J�A� van Leeuwen� L�M�C� Buydens� B�G�M� Vandeginste� G� Kateman� and M� Mulholland� Analytica Chimica Acta� � � ������ ��� � �� L�M�C� Buydens� J�A� van Leeuwen� M� Mulholland� B�G�M� Vandeginste� and G� Kateman� TRAC� Trends in Analytical Chemistry� � ������ No��� ��� � �� M� Mulholland� N�Dunand� and A�Cleland� Journal of Chromatography� ��� ������ �� � �� � �� M� Mulholland� J�A� van Leeuwen� and B� Vandeginste� Analytica Chimica Acta� �� ������ �� � � �� J�A� van Leeuwen� L�M�C� Buydens� B�G�M� Vandeginste� G� Kateman� G�P�J� Schoenmakers� and M� Mulholland� Chemometrics and Intelligent Laboratory Sys� tems� �� ������ �� � �� J�A� van Leeuwen� L�M�C� Buydens� B�G�M� Vandeginste� G� Kateman� P�J� Schoen� makers� M� Mulholland� Chemometrics and Intelligent Laboratory Systems� �� ������ �� ���� J�A� van Leeuwen� L�M�C� Buydens� B�G�M� Vandeginste� G� Kateman� P�J� Schoen� makers� and M� Mulholland� Chemometrics and Intelligent Laboratory Systems� �� ������ ���� ���� J�A� van Leeuwen� B�G�M� Vandeginste� G� Kateman� Analytica Chimica Acta� ��� ������ ���� ���� R� Bach� J� Karnicky� and S� Abbot� in T�H� Pierce� and B�A� Hohne� Arti�cial Intelligence Applications in Chemistry� American Chemical Society� Washington� ����� p���� �� � S�S� Williams� J�F� Karnicky� J� Exco�er� and S�R� Abbot� Journal of Chromatog� raphy� ��� ������ ���� ���� S�S� Williams� TRAC� Trends in Analytical Chemistry � ������ No��� � � ���� P� Lu� and H� Huang� Journal of Chromatography� ��� ������ ���� ���� Z� Yukui� Z� Hanfa� and L� Peichang� Journal of Chromatography� ��� ������ � � ���� N� Chen� X� Liang� Y� Zhang� and P� Lu� Chinese Science Bulletin� � ������ No���� ����� ���� X� Liang� H� Huang� Y� Zhang� and P� Lu� Chromatogram� ������ ���� ���� M�A� Tischler� and E�A� Fox� Computers and Chemistry� �� ������ No��� � �� ���� R�M� Smith� and C�M� Burr� Journal of Chromatography� ��� ������ ��� ���� C�M� Burr� and R�M� Smith� Analytical Proceedings �London�� �� ������ No��� ��� ���� T�P� Bridge� M�H� Wiliams� and A�F� Fell� Journal of Liquid Chromatography� �� ������ �� ��� � � �� � T�P� Bridge� M�H� Wiliams� and A�F� Fell� Journal of Chromatography� ��� ������ ��� ���� T�P� Bridge� M�H� Wiliams� G�G�R� Seaton� and A�F� Fell� Chromatographia� �� ������ ���� �� ���� T�P� Bridge� M�H� Wiliams� and A�F� Fell� Journal of Pharmaceutical Biomedical Analysis� � ������ No�s���� ���� ���� X� Wu� S� Hu� J� Hou� and X� Xu Journal of Chemical Engineering of Chinese Universities� � ������ No��� � �� ���� G� Szepesi� and K� Valko� Journal of Chromatography� ��� ������ ��� ���� K�M� Kiyoshi Tsuji Jenkins� Journal of Chromatography� ��� ������ ���� ���� R� Milne� Journal of Chromatography� ��� ������ ��� ���� X� Wen� and W� Fei� Acta Petrolei Sinica �Petroleum Processing Section� � ������ No��� ��� ���� P�B� Ayscough� S�J� Chinnick� R� Dybowski� and P� Edwards� Chem� Ind� �Lond��� �� ������ ���� ���� K� Valko� G� Szabo� J� Rohricht� K� Jemnitz� and F� Darvas� Journal of Chromatog� raphy� ��� ������ ��� �� � H� Gunasingham� B� Srinivasan� and A�L� Ananda� Analytica Chimica Acta� ��� ������ �� � ���� A�L� Ananda� S�M� Foo� and H� Gunasingham� Journal of Chemical Information and Computational Science� �� ������ ��� ���� C�T� Mant� T�W� Lorne Burke� N�E� Zhou� J�M�R� Parker� and R�S� Hodges� Journal of Chromatography� ��� ������ ��� ���� J�C� Pearce� A� Churchill� and A�C� Terry� Chemometrics and Intelligent Laboratory Systems� Laboratory Information Management� �� ������ �� � ���� J�A� Asenjo� B� Byrne� and L� Herrera� Journal of Biotechnology� �� ������ ���� ���� H� Eriksson� K� Sandahl� G� Forslund� and B� Osterlund� Chemometrics and Intelli� gent Laboratory Systems� � ������ �� � ���� H� Eriksson� K� Sandahl� J� Brewer� and B� Osterlund� Chemometrics and Intelligent Laboratory Systems� � ������ ���� ���� P� Schneider� and S�K� Graham� in H� Berghel� J� Talburt� D� Roach �Eds�� � Pro� ceedings of the ���� Symposium on Applied Computing� ���� ���� G� Holzer� W� Bertsch and Q�W� Zhang� Analytica Chimica Acta� ��� ������ No��� ���� ���� T� Aczel� S�G� Colgrove� and L� Le� ASTM Spec� Tech Publ�� ���� �Novel Tech� Fossil Fuel Mass Spectrometry�� ������ ���� �� �� � C�E� Riese� and J�D� Stuart� in T�H� Pierce� and B�A� Hohne� Arti�cial Intelligence Applications in Chemistry� American Chemical Society� Washington� ����� p��� ���� Z� Yan� Y�Z� Wu� and Y� Zhou� in Proceedings of the rd International Conference on Properties and Applications of Dielectric Materials� � ������ � � ���� M� Mulholland� R� Haddad� and D�B� Hibbert� Journal of Chromatography� ��� ������ �� ���� H� Yuzhu� G� Peeters� G� Musch� and D�L� Massart� Analytica Chimica Acta� �� ������ �� ���a� A�Bartha� and G� Vigh� Journal of Chromatography� ��� ������ � � ���b� H� Moll� Ph�D� thesis� University of Bern� ����� ���� D� Diaper� Knowledge Elicitation� Principles� Techniques and Applications� Ellis Horwood� ����� ���� R� Hindriks� F� Maris� J� Vink� A� Peeters� M� De Smet� D�L� Massart� and L� Buydens� Journal of Chromatography� ��� ������ ���� ���� M� Meyer� J� Booker� Eliciting and Analysing Expert Judgement� A PracticalGuide� Academic Press� London� ����� ���� J�R� Quinlan in D� Michie� �Ed�� Expert Systems in the Micro Electronic Age� Ed� inburgh University Press� Edinburgh� ����� ���� B� McGee� in ���� ��� �� � M� Minsky� in P�H� Winston� The Psychology of Computer Vision� McGraw�Hill� New York� ����� ���� M� Mulholland� N� Walker� F� Maris� H� Hindriks� L� Buydens� I� Bla�ert� and P�J� Schoenmakers� Journal of Chromatography� ��� ������ ���� ���� A� Macro� Software Engineering� Concepts and Management� Prentice Hall� Hert� fordshire� UK� ����� ���� A� van Mayrhauser� Software Engineering� Methods and Management� Academic Press� London� ����� ���� M�F� McTear� and T�J� Anderson� �Ed�s� Understanding Knowledge Engineering� Ellis Horwood� ����� ���� E� Hollnagel� The Reliability of Expert Systems� Ellis Horwood� Chichester� ����� Figure �� A Taxonomy of Expert Systems for Chromatography� Figure �� A Taxonomy of Single Stage Expert Systems for Chromatography� Figure � A Taxonomy of Expert Systems for Chromatography that are speci�c to a class of compounds� Figure �� A mapping between the subdomains of ESCA� the subsystems implemented for each of those subdomains and the proposals for integrating these subsystems� The four polygons correspond to the four proposed integrations and each polygon surrounds those systems that were to be integrated by that proposal� Expert Systems for Chromatography Non Specific Specific to Compound Class Specific to Chromatographic Single Stage Multiple Stage Ion Pair Chromatography ECATESCA New Solute South Wales Method Conditions Type Determ- ination Brussels Method Conditions and Optimisation Method Thin Layer Chromatography ESLC ESC/ Moll’s System Figure �� Method Selection Single Stage Systems Optimisation Trouble- shooting Retention Prediction CATHIEESP Hungarian Selectivity Bradford & Herriot Chinese Mobile PhaseWatt Upjohn HPLC CRIPES Doctor Figure �� Specific to Compound Class Metabolites Steriods Decision APN Proteins Fossil Fuels PPA P8 PROTEIN System for Arson Analysis System SystemTree HPLC METABOL- EXPERT SPES Exxon System System for Power Transformer Analysis TOGA Pesticides PIA Figure �� Selection of Initial Conditions LABEL DASH LIT Retention DASH’ LABEL’ LIT’ SLOPES VARIABLES CRISE DESIRE Diamond SelectivityOptimisation Method (or SOS Repeatability REPS Ruggedness RES Validation graphic) Proposal Two Proposal Three Chromato- Proposal One Proposal Four Figure �� Name of Predominant Behaviour Intended Users Expert Plans Predicts Diag� Ident� Experts Novices Unspe� System noses i�es ci�ed Arson Y Y Analysis Brussels Y Y Method Conditions and Optimisation CATHIE Y Y CRIPES Y Y ESP Y Y Exxon Y Y Hungarian Y Y Selectivity Optimisation HPLC Y Y METABOL EXPERT New South Y Y Wales Method Conditions PECOD Y Y PIA Y Y PPA Y Y Y PROTEIN Y Y P� Y Y Power Y Y Transformer Solute Y Y Type Upjohn Y Y Steroid Y Y Systems �� � � � � �� Table �� Behaviour and Users of Expert Systems for Chromatography� Language Systems that Used for Aspects for which the language was used used Language whole Reasoning Numerical Interface Data system" processing capture Artifical Intelligence Languages Prolog ESP N Y MicroProlog Bradford Herriot N Y Watt System Steroid Decision Y Tree System TurboProlog Exxon System Y PIA Y Solute type Y determination system Common ECAT Y LISP PPA N Y SCHEME ESC�ESLC N Y LISP Conventional Languages BASIC Brussels ion N Y FORTRAN pair system LABEL N Y C Chinese mobile Y phase system ESP N Y RES N Y Lattice C APN Steroid Y system Pascal SOS Y reimplementation REPS Y reimplementation Turbo Pascal Exxon System Y Unspeci�ed Bradford Herriot N Y Y conventional Watt system language ESC�ESLC N Y Table �� Languages used by Expert Systems for Chromatography� Name of Tool Supplier Systems that Notes used Tool Epitool Epitec AB PPA Except interface Expert Systems IBM PROTEIN Environment �ESE� Goldworks Gold Hill REPS First implementation Computers� only Cambridge� RES Factor choice module MA� USA only Knowledge Craft Carnegie Group� SOS Prototype only Pittsburgh� PA� USA Knowledge Software Archi� CRISE Engineering tecture and DASH System �KES� Engineering� LABEL Arlington� Brussels ion VA� USA pair system M�� Cim�ex�Teknow� Upjohn ledge� Palo�Alto system CA� USA Nexpert Object Neuron Data� SOS An attempt was made Palo Alto� to use it for the CA� USA reimplementation Personal Consul� Texas PROTEIN tant Plus�PC Plus Instruments Xi�Plus Expertech� UK S�Wales method conditions system VP�Expert Paperback CRIPES Software Table �� Expert System Tools and Shells used by Expert Systems for Chromatography� 
work_bkf2yq62svgubgfcf4nyrfvrra	----	Fuzzy neural network and fuzzy expert system for load forecasting - Generation, Transmission and Distribution, IEE Proceedings- k and fuzz P.K. Dash A.C. Liew S.Rahman Indexing terms: Neural networks, Load forecasting, Expert system, Power system planning Abstract: A hybrid neural network fuzzy expert system is developed to forecast short-term electric load accurately. The fuzzy membership values of the load and other weather variables are the inputs to the neural network, and the output comprises the membership values of the predicted load. An adaptive fuzzy correction scheme is used to forecast the final load by using a fuzzy rule base and fuzzy inference mechanism. Extensive studies have been performed for all seasons, and a few examples are presented in the paper, including average, peak and hourly load forecasts. 1 introduction The short-term load forecast is very important to an electric utility. The quality of control of a power sys- tem, and economy of operation, are highly sensitive to forecasting error. A sound basis for load predictions is generalisation of past, known cases. The science of sta- tistics provides a range of tools for this purpose. They are based on the idea of fitting a particular class of models to data and then hypothesising that future events will conform to the fitted model. Many approaches have been applied to electric load forecast- ing, including linear regression, exponential smoothing, stochastic process and state space methods. While each of these methods demonstrates success in forecasting, they have serious disadvantages: reliance on large his- torical databases with possible obsolete and irrelevant data, assumptions about static load shapes and param- eters etc. A detailed comparison of various statistical approaches is found in [l]. One of the most promising application areas of the artificial neural network (ANN) is load forecasting [2- 131. The neural network is able to perform nonlinear modelling and adaptation and does not rely on the explicitly expressed relationship between input varia- 0 IEE, 1996 ZEE Proceedings online no. 19960314 Paper received 26th October 1994 P.K. Dash is with the Department of Electrical Engineering, Regional Engineering College, Rourkela, India A.C. Liew is with the Department of Electrical Engineering, National University of Singapore, Singapore S. Rahman is with the Bradley Department of Electrical Engineering, Vir- ginia Polytechnic Institute & State University, USA 106 bles and forecast load. When using neural networks for load forecasting, one needs to consider only the selec- tion of variables as the network input. The relationship between the input variables and predicted load will be formulated by a training process. Several authors have attempted to apply the backpropagation learning algo- rithm to train the ANNs for forecasting time series. The fuzzy expert system approach [14] has also been applied to forecasting where the advantage of an oper- ator’s expert knowledge is used. However, the fuzzy decision system for load forecasting requires detailed analysis of data and the fuzzy rule base has to be devel- oped heuristically for each season. The rules fixed in this way may not always yield the best forecast. On the other hand, hybrid solutions [15, 161 have been pro- posed for short-term forecasting of electric loads, whereby the functionality of the fuzzy expert system and the learning capabilities of the neural network can be merged to yield a forecasting system more powerful than either of its components alone. The present work is aimed at achieving the said objective of a robust load forecast with improved accu- racy using a fuzzy neural network for initial forecast and a fuzzy expert system (FES) producing load cor- rections to yield the final forecast. For the neural net- work to be called a FNN, the signal and/or the weights should be fuzzified. This type of FNN is based on the multilayer perceptron, using the backpropagation algo- rithm. The fuzzified input vector consists of the mem- bership values of the past load and weather parameters and the output vector is defined in terms of fuzzy class- membership values of the forecast load. A simple fuzzy-inferencing mechanism is used to yield the magni- tude of the forecast load during the initial phase. In the final phase a fuzzy expert system is used to produce load corrections. The input vector to the fuzzy expert system (FES) consists of differences in the weather parameters between the present and the forecasted instant. The output of the FES gives the load correction which, when added to the initial forecast, yields the final fore- cast. Thus, by using a hybrid approach the load-fore- casting errors are expected to reduce considerably. However, as the lead time increases from 1 h to 24h or 48 h, the forecasting error increases, necessitating the use of an adaptive fuzzy load error correction scheme. Several examples presented in this paper include 24h- ahead average, peak and hourly load forecasts using the hybrid approach. The effects of both linear and nonlinear adaptive load-correction schemes are shown. IEE Proc -Genu Transm Distrcb , Vol 143, No 1, January 1996 2 form Fuzzy pattern representation in linguistic The approach used in this paper is aimed at improving the prediction by handling uncertain and ambiguous variations of load patterns and weather parameters using a fuzzy linguistic approach. Since it is easier to convert exact information into linguistic form than vice versa, we consider the major linguistic properties small, medium, and large as the attributes of the input fea- ture. Any input feature value can be described in terms of some combination of membership values for these properties. In traditional two-state classifiers, an ele- ment x either belongs or does not belong to a given class A ; thus, the characteristic function is expressed as 1 i f a : E A 0 otherwise P a ( X ) = In real-life problems such as load forecasting, the classes are often i l l defined, overlapping, or fuzzy and a pattern point may belong to more than one class; in such situation, fuzzy set theoretic techniques can be very useful. We use the modified n-function [17, 181, lying in the range [0,1] to assign membership values for the input features corresponding to the linguistic properties small, medium, and large. The membership function of the input feature otherwise where hi > 0 is the radius of the n-function with ci as the central point at which p ( x i ) = 1. This is shown in Fig. 1. Xirnin Csmoh ‘medium ‘Iorgeximm - ’medium -4 I---- ~~%ar+-- ’ / 2 L o d Fig. 1 n-function representation 2. I Let x,,,, and x,,,, denote the upper and lower bounds of the observed range of feature x, in all L pattern points, considering numeric values only. Then, for the three linguistic property sets, the following are used: Choice of parameters for the .n-function L d t U m ( ~ ) = 0.5(& m a z - G mtn) c m e d z u m ( 2 , ) xz mzn + X m e c i z u m ( X z ) ( 3 ) Xsma11(&) = ( l / f d ) { C m e c l z u m ( & ) - & m z n } csmall ( 2 % ) = cmedzzlm(2t) - O . 5 X S m a 1 ~ (zZ) IEE Pvoc -Gener Trunsm Distvib Vol 143, No 1, January 1996 ;\large (xz) = (l/fci) { 5% m a x - C m e c i z u m ( x z ) } C i a r g e ( G ) = C m e d t u m ( 2 z ) + 0 . 5 h a r g e ( X z ) where 0.5 5 f d 5 1.0 is a parameter controlling the extent of overlapping. The n-function representation permits a more compact and meaningful representation of each pattern point in terms of its linguistic proper- ties, and ensures better handling during both the train- ing and testing phases of the proposed fuzzy neural network model. 3 The present work attempts to build a fuzzy neural net- work model based on the multilayer perceptron using the gradient descent based backpropagation algorithm by incorporating concepts from fuzzy sets at various stages. Fig. 2 shows the fuzzy neural network model for obtaining the initial forecast. The fuzzy sets for load, temperature and humidity parameters are shown in Figs. 3-5. The input to the fuzzy neural network comprises the membership values to the overlapping partitions of linguistic properties small, medium, and large corresponding to each input feature such as past load, temperature, humidity etc. This provides scope for incorporating linguistic information in both the training and testing phases of the said model and increases robustness in tackling imprecise or uncertain input specifications. The components of the output layer consist of the membership values to the overlap- ping partitions of linguistic properties small, medium and large corresponding to the forecast load magni- tude. During training, the network backpropagates the errors with respect to the desired membership values at the output nodes. After a number of cycles, the neural network converges to a minimum error solution by using a gradient descent algorithm as shown in the Appendix. Fuzzy neural network for initial forecast h I th layer j th layer k t h layer I th layer Fig. 2 Fuzzy neural network for initia1,forecast 7 1 .o 0.8 0.6 TJ e Q ._ L 2 0.4 E E 0.2 0 0.116 0.223 0.417 0.611 0.800 1.0 load( normalised) Fig.3 ~ small -0- medium -x- large Fuzzy membership grade for load 107 temperature (normalised) Fig. 4 ~ small -0- medium -x- large Fuzzy membership grade for temperature humidity (normalised) Fig. 5 ~ small -0- medium -x- large Fuzzy membership grade for humidity After the learning phase is over, when separate load and weather patterns are presented at the input layer, the output nodes automatically generate the member- ship values corresponding to the linguistic properties small, medium and large. Thus a centroid defuzzifica- tion technique [I91 is used to obtain the initial load forecast from the membership values and the corre- sponding loads obtained from the n-functions. 4 Selection of training patterns The utility data studied here are susceptible to large and sudden changes in weather and load, so selection of appropriate training cases plays a vital role in train- ing the network. Several techniques for the selection of training patterns have been suggested in [l l-131. The present paper discusses a different training scheme for the selection of training patterns for hourly load fore- cast. To predict hourly loads the following load model is chosen: y ( i , t ) = f { y ( i - m ) , y ( i , t - m ~ I), . . . , y ( i , t - m - n l ) , (4) and where y and z are the load and weather variables, respectively; i and t indicate the day and the hour, respectively, m indicates the lead time for the hourly load forecast, (i.e. m = 1 for 1 h-ahead forecast, m = 24 for 24h-ahead forecast, m = 48 for 48h-ahead forecast, m = 168 for 168h-ahead forecast); nl indicates the data z(2, t - m ) , z ( Z , t ) , . . . , z(2 - t - m ) } m = n2 108 length for load; and n2 indicates the data length for temperature. For hourly load forecasting, eqn. 4 is used to select the training patterns. Various lengths of the past his- torical load and temperature values are used and their effects on the load-forecast accuracy are studied. It is found that, with nl > 0 and n2 > 0, there are no marked improvements in the results for the utility data used in this paper. Also the training time increases con- siderably with larger values of nl and n2. Therefore, n1 = 0 and n2 = 0 are chosen in this paper. Using the above scheme, the network is trained for 14 days (i = 1, ..., 14) at time t and load is predicated for the next 14 days (i = 15, ..., 28) at time t. Hence to predict for all 24h of a given day, 24 different neural networks are used, each one trained separately with the same parameters. This is desirable because the training set is small for each neural network consisting of a few patterns (14 patterns only in this case) with irrelevant data for other hours being discarded. Further, depend- ing upon the difference of the load responses on the day of the week, a day-of-the-week indicator is intro- duced along with the input vector to the network. As the weather variable temperature is the most important parameter in short-term load predictions, an all-temperature model is used to obtain the hourly fore- casts. The training data used for 1 h-ahead predictions are: Input pattern: P(i,t) = power at tth instant of ith day, e(i,t) = temperature at (t+l)th instant of ith day, @(i,t+ 1) = temperature at (t+ 1)th instant of ith day wd(i) = day of the week indicator i. Output pattern: P(i,t+l) = power at (t+l)th instant of ith day. The training data for 24h-ahead, 48h-ahead and 168h- ahead predications are Input patterns: P(i,t) = power at tth instant of ith day, O(i,t) = temperature at tth instant of ith day, B(i+m,t) = temperature at tth instant of (i+m)th day. Output pattern: P(i+m,t) = power at tth instant of (i+m)th day where m = 1, 2, 7 for 24h-, 48h- and 168h-ahead pre- dictions, respectively. However, for average daily and peak-load predic- tions, the following training data are used: Input and output patterns: P(i-1) = average load on (i-1)th day, Qmln(i-I) = minimum temperature of (i-1)th day, OmaX(i-l) = maximum temperature of (i-1)th day, O,,(i) = minimum temperature of ith day, Omax (i) = maximum temperature of ith day, P(i) = average load for ith day. The same data can be used for the peak load forecast. Although an all-temperature model will produce an accurate forecast in most seasons, the other weather parameters like humidity and wind speed affect the forecasting accuracy during summer and winter, respec- tively. Thus, if humidity records are available in a par- ticular season, they may be included in a training pattern. IEE Proc -Gener Transm Distrib Vol 143, No I , January I996 5 Fuzzy expert system for final forecast During training, the neural net produces an initial fore- cast with a set of initial load, temperature and humid- ity data expressed in terms of fuzzy membership values. The error between the actual load A ( i ) and predicted load P(i) for a given hour or a given day (ith hour or day) is expressed as In a similar way, temperature and humidity errors are expressed as aqi) = ~ ( i ) - ~ ( i ) AO(2) = O ( 2 ) - O ( 2 - 1) AH(2) = H ( 2 ) - H ( 2 - 1) ( 5 ) (6) However, if maximum and minimum temperatures are used, ( 7 ) A Q m a z (2) = Q m a z ( 2 ) - Q m a z (2 - 1) ~ Q m z n ( 2 ) = Q m z n ( 2 ) - Qmzn(i - 1) The errors in the weather parameters and load-correc- tion values are fuzzified using six fuzzy sets such as SP (small positive), MP (medium positive), LP (large posi- tive), SN (small negative), NM (medium negative) and LN (large negative). Figs. 6 and 7 show the fuzzy sets for temperature and humidity errors. These sets are obtained using the n-function given in eqn. 2. Both positive and negative fuzzy sets are symmetrical about the origin. -0.06 -0.03 0 0.03 0.06 error (normalised) Fig. 6 ~ small -0- medium -x- large Fuzzy membership grade for temperature errors -0.12 -0.06 -0.04 0 0.04 0.06 0.12 error (normalised) Fig. 7 ~ small -0- medium -x- large Fuzzy membership grade for humidity errors The load correction output sets have six members and use a linear fuzzification principle for obtaining membership grades as shown in Figs. 8 and 9. Nonlinear membership grades can also be used for obtaining load corrections for the final forecast. The sets for load corrections are classified as SPC (small positive correction), MPC (medium positive correc- tion), LPC (large positive correction) etc. The corre- sponding negative fuzzy sets are SNC, MNC, and LNC, respectively. 1 . O R t I f -1.5 -1.2 -0.6 0 0.6 1.2 1.5 correction (normalised) Fig. 8 ~ small -0- medium -x- large Membership grades for load correction (linear) 0 0.4 0.8 1.2 1.6 2.0 correction (normalised) Membership grades for load correction (linear) Fig. 9 The membership values of the load correction APc output is given by b m a z where C,,, is the slope of the load error correction and eLc is the maximum load error correction for the corre- sponding linguistic set (for which the membership value becomes unity). The value of C,,, for different output linguistic sets SPC, MPC and LPC is C L a x , C$,, CL,, respectively, and these values are obtained by observing the load prediction errors over a two-week period prior to forecasting. The fuzzy rule base is formed by trial and error to reduce the load correction to a very small value during training. However, a fuzzy basis function approach [19] can be used to select the appropriate rules automati- cally out of a large number of possible combinations. Two sample rules for an all-temperature model for average load forecast will be Rule 1: I F AOmaX(i) is LN and AO,,(i) is SP, THEN AP,(i) is MPC Rule 2: I F AOmaX(i) is MP and AOmin(i) is LP, THEN AP,(i) is LPC In a similar way the two-sample rule for the peak- load forecast are Rule 1: IF AOmaX(i) is SP THEN AP,(i) is SPC Rule 2: I F AO,,,(i) is LN THEN AP,(i) is MNC. IEE Proc-Gener. Transm. Distrib., Vol. 143, No. 1, January 1996 109 However, if a load forecasting model with both temper- ature and humidity parameters is used, the rules are of the form Rule 1: I F AQ(i) is SP and H ( i ) is MP THEN AP,(i) is MPC Rule 2: I F AQ(i) is SN and AH(i) is LN THEN AP,(i) is LNC. The total number of production rules in the fuzzy knowledge base using two variables and six categories of sets is 36. Because of partial matching of the fuzzy rules and the fact that preconditions do overlap, more than one fuzzy control rule can fire at a time. For the fuzzy rules used for load forecasting, the truth values of the pre- conditions are (considering a temperature-humidity model) i = 1, 2, ..., k and k = number of rules fired for a given value of AO(i), and AH(i) belonging to fuzzy classifiers SP, MP, LP etc. and A denotes a conjunction operator (usually a minimum operator). The output of rule (i) is calculated by applying the matching strength of its pre- condition on its conclusion as j = 1, 2, ..., 6 and cALax is the value of C,,, for the out- put set Ay If two rules have the same consequent out- put set, the Lukasiewickz OR rule is used as ai = A [ P { A @ ( i ) ) , Pu(AH(411 (9) A P c ( i ) = a,c2a, (10) a; = min[I,p{AO(i)} + p { A H ( i ) } ] (11) A centroid defuzzification technique is used to yield the load correction as A P c ( i ) = aiAPc/ a, = a:C2u,l ai (12) A P c ( i ) = a:c2,z/ a2 (13) However, by taking the load correction as CAAax, for which the output membership function is unity, the value of AP,(i) is obtained as The final value of forecast load is thus obtained by summing the ANN output, and the output from the fuzzy expert system is The block diagram for the integrated fuzzy neural net- work and fuzzy expert system is shown in Fig. 10. p(t-1)- w - 1 ) - defuzzification e o ) -- I Pf(2) = P(2) + APc(i) (14) fuzzyfication i - ANN defuzzification \ - - M t ) -+- fuzzy rule base and fuzzy inference Fig. 10 Integrated fuzzy neural network-fuzzy expert system 5. I Adaptive load correction For small load correction eqns. 10 and 11 are adequate to produce accurate forecasts (usually for a lead time from 1 - 6h). However, as the lead time increases to 110 24, 48, 72 or 168h, the membership function for the load-error correction is adapted as 1 p{apc(z)~ = c,,, f ac,,, apc(~) (15) the value of AC,, is obtained as a function of load- correction errors from the training cases using different lead times for predictions. Fig. 11 shows the linear adaptive correction ACmax as a function of the normal- ised error. The nonlinear adaptive version of the load- error correction is shown in Fig. 12. For adaptive cor- rections, eqns. 12 and 13 are rewritten as and Some of the results are given below for the various models considered in this paper. 80 I increment / decrement \ I -801 I I I I I I I I 0 0.002 0.004 0.006 0.008 ,010 ,012 014 maximum error (normalised) Fig. 11 60 I Linear adaptive load correction decrement -20 -40 increment I 1 I I I 1 I I 0001 0002 0 003 0 004 0005 maximum error (normalised) Fig. 12 Nonlinear adaptive load correction 6 Forecasting results 6. I Average load forecasting To evaluate the fuzzy ANN and fuzzy expert system approach, load forecasting is performed on the load data collected at the Virginia Polytechnic Institute and State University. The fuzzy neural network (FNN) is compared with the combined FNN and fuzzy expert system model using ordinary backpropagation algo- rithm for obtaining one-day-ahead average load fore- casting during a 14 day period in the month of May. The input layer of the FNN comprises 15 neurons for five input features and the output layer has three neu- rons. The number of neurons in the hidden layer is fixed as 17 for this particular forecast, to obtain the best results. If the day-of-the-week indicator is used, one more neuron is added to the input layer. Back- propagated errors are assigned appropriate weightage for weight updating depending on the membership val- ues at corresponding outputs. The learning rate r\ and IEE Pvoc -Genev Tvansm Distrib Vol 143, No I , Januavy 1996 momentum coefficient a are gradually decreased to prevent oscillations as the neural network converges to a minimum-error solution in a maximal number of sweeps through the training set. Note that the parameters q and a traverse a range of values in the course of computations and one may choose 0.6 < a < 1.0, and 0.0001 < q < 0.1. The values of p and y are chosen for best performance as 0 < p < 0.02 0.2 < y < 0.6 For the present study the initial learning rate q and momentum a are and The percentage error PE is evaluated as 7 = 0.01, a = 0.8 /!? = -0.015, 7 = 0.4 forecast load - actual load PE = x 100 forecast load and percentage absolute error PAE = lPEl The average load profile for the 14 day period in May is shown in Fig. 13. Fig. 14 presents the PAE for 24h period ahead forecast. The maximum PAE for a 24h-ahead forecast during a 14 day period in May is 1.75 using the FNN and fuzzy expert system in com- parison with 3.40 for the FNN only. Weekdays, week- ends and Sundays are all included in this model. 28 3 2 7 - - 2 2 6 - E- 25 24 29 t - - - t 30 t "I 6 W Q 4 4 2 0 1 6 11 16 21 26 3' 2311 I I 1 I I I I 8 I I I . ( 1 1 6 11 days Fig. 13 Average loudprojile forecast for the month of May c I davs Fig. 16 month of Mav Percentage absolute error in peak load profile forecast for the ~ ~ N N bnly -+- FNN with fuzzy corrections 1 6 11 d a y s Fig. 14 month of May ~ FNN only -e- FNN with fuzzy corrections I E E Proc-Gener Transm. Distrib., Vol. 143, No. 1. January 1996 Percentage absolute error in average load profile forecast for the 6.2 Peak-load forecast Figs. 15-18 show the actual peak loads and PAEs for a 31 day period during May and December. The maxi- mum temperature errors are used for fuzzy corrections. The combined model produces a maximum PAE of about 1.65 for a 24h-ahead peak-load forecast during the month of December. During May, the maximum PAE is 7.5 for a 24h-ahead peak-load forecast using the FNN model. However, using fuzzy correction, the maximum PAE is reduced to 5.5. Both the forecasts shown in these Figures include weekdays, and week- ends etc. Special holidays like Christmas etc. are included in obtaining the peak-load forecast for the month of December. These errors can be reduced fur- ther by using adaptive load-correction schemes. 6.3 Hourly load forecasts The data from a utility in Virginia is used to produce 24h- and 48h-ahead hourly forecasts. For this utility, both temperature and humidity records are available during all seasons of the year. However, the forecasting errors for 21 and 22 January over a 24h period are shown in Figs. 19-22. The maximum PAEs for 24h- and 48h-ahead forecasts are 2.8 and 3.5 without fuzzy corrections and 0.69 and 1.72 with fuzzy corrections, respectively. The corresponding load profiles are also shown in these Figures. 111 40c 20 1 6 11 16 21 26 31 days Peak load profile forecast for the month of December Fig. 17 4 3 W 2 2 1 6 11 16 21 26 31 0 1 days Fig. 18 month of December ____ FNN only -+- FNN with fuzzy corrections Percentage absolute error in peak load profile jorecasl f o r the 9500 - 9000 - 8500 - 8000- 7500- 7000 - 6500 - *- 0 - 6000 5500 i 1 1 1 1 1 1 * 1 1 I I 1 6 11 16 21 time,h \ Fig. 19 records for utility data 24h ahead hourly load forecasts with temperature and humidity 6.4 Hourly load forecasts using adaptive correction Figs. 23 and 24 show the 24h ahead forecast errors for a summer day (27 May) using an all-temperature model and the data from the experimental setup at the Virginia Polytechnic. Both linear and nonlinear adap- tive fuzzy corrections are used to provide a more accu- rate forecast. From the Figure it is found that the percentage absolute-load-prediction error over the entire 24h period comes down significantly using both adaptive-fuzzy-correction formulations. However, the difference between the two versions is not very signifi- 112 cant, and thus the PAE with the nonlinear version is shown in the Figure. The nonlinear version is pre- ferred, as it is expected to produce significant accuracy for 1 to 2h ahead load forecast. The hourly-load-fore- cast accuracy is significant in this case as the University load profile (shown in the Figure) does not show signif- icant changes during the 24h period. ‘I I -3 1 6 11 16 21 time,h Fig.20 records for utility data ~ without fuzzy corfection -+- with fuzzy correction 24h ahead hourly load forecasts with temperature and humidity 9500 t 6500 1 6 11 16 21 time,h Fig.21 recordr f o r utility data 48h ahead hourly load forecast with temperature and humidity ’I I time, h Fig.22 records for utility data ~ without fuzzy correction -+- with fuzzy correction 48h ahead hourly load forecasts with temperature and humidity IEE Proc.-Gener. Transm. Distrib., Vol. 143, No. I , Januavy 1996 Figs. 25 and 26 show the 24h-ahead forecast errors for 15 February (winter day) using the nonlinear ver- sion of the fuzzy adaptive correction scheme along with FNN and FNN with fuzzy corrections. Significant accuracy in the hourly forecast is also obtained in this case using adaptive corrections. The above two days are nonspecial days and are chosen to illustrate the accuracy of the adaptive correction scheme. 22.55 22 t 21. 21 3 20. 2- 20 d 9 19. 19 18. 18 17.5 1 6 11 16 21 I ’ ’ I I ’ ’ I ‘ I ’ I I I ‘ ‘ I I ‘ I ‘ I time, h Fig. 23 mer day) Hourly loadjorecast with adaptive corrections f o r 27 May (sum- 0.5l””v”ttJis I 0 1 6 11 16 21 time, h Fig.24 Houvly load forecast with adaptive corrections for 27 M a y (sum- mer day) ~ without fuzzy correction -*- with fuzzy correction --A- with nonlinear adaptive correction 7 Discussion The proposed fuzzy-neural-networklfuzzy-expert-sys- tem approach is found to be very powerful and robust for short-term load predictions. Although the results for two seasons of the year are presented in this paper for validating the effectiveness of this approach, extensive tests have been conducted for other seasons, Sundays, holidays and special days of the year. From the results presented in this paper, it is observed that significant accuracy can be achieved for 24h ahead hourly load forecasts and the PES could be even less than 1%. Adaptive fuzzy load correction schemes enhance the accuracy of the predictions in most cases. However, if the lead time increases to 48h, the percentage error will be around 2%. The accuracy of load predictions using both adaptive corrections will be the highest with loads which do not show large IEE Proc.-Gener. Trunsm. Distrib., Vol. 143, No. I , January 1996 hourly variations. However, significant accuracy can still be achieved with highly stochastic load variations, if a fuzzy basis function approach is used to arrive at adaptive corrections instead of the trial-and-error method presented in this paper. 4017----- 35 21 151 I I I I I I I ’ ‘ I I ’ ’ I ’ ’ ’ 1 6 11 16 time,h Fig. 25 (winter day) Hourly load forecast with adaptive corrections for 15 February 0.5 1 1 0 1 6 11 16 21 tirne,h Fig. 26 (winter day) ~ without fuzzy correction -*-- with fuzzy correction -A- with nonlinear adaptive correction Hourly load forecast with aduptive corrections f o r 15 February The results for average load forecast are very encour- aging and the maximum absolute percentage error is around 1.8. From the results for peak-load forecast presented in this paper, it is observed that except for 14 and 15 May, the PAE is less than 2 for most of the days. The mean absolute error is evaluated for this month and is found to be 2.824 with the FNN model only and 1.150 using fuzzy corrections. The large errors for the above two days are probably due to other factors which have not been used in the training patterns. Although the results for 168 h-ahead load forecast using the above models have not been reported in this paper, the computations reveal that the mean absolute percentage error is around 2 and the maximum PE is 4.28 using the FNN model. With fuzzy corrections, the maximum PE is reduced to 2.16. This is quite comparable with the results for the ANN-based load forecasting technique presented by Karady et al. [12, 131. Although the studies reported here have utilised a few simple examples and models, they are extremely 113 valuable in identifying a promising hybrid forecasting methodology which can be investigated for larger number of inputs, more weather parameters, special load patterns, seasonal load changes, peak and total load forecasts one week ahead etc. 8 Conclusion A new hybrid model integrating an artificial neural net- work and a fuzzy expert system is developed for 24 h ahead average and peak load forecasts and tested with historical load data. The hybrid model uses a fuzzy neural network for obtaining the initial forecast from the fuzzified input data and a fuzzy expert system gen- erating the load correction to produce the final fore- cast. The simulation results of the proposed method using historical data show that the forecasting errors are less than 2% for both 24h ahead average and peak load predictions. The selection of training pattern pre- sented in this paper does not classify the patterns to weekday and weekend day and thus the hybrid model is a promising approach for short-term load forecast for all days of the year. The paper also presents an adaptive fuzzy correction scheme to minimise the fore- cast errors more precisely and 24h ahead hourly fore- casts using this scheme yield significant accuracy. The fuzzy neural network and fuzzy expert system approach has also been applied to one week ahead load forecast, and the results are found to be very promising. 9 Acknowledgment The authors acknowledge funds from the US National Science Foundation (NSF grants INT-9209 103 and INT-9 1 17624) for undertaking this research. 10 References 1 RAHMAN. S.. and BHATNAGAR. R.: ‘An exuert svstem based algorithm for ’ short-term load forecast’, I E g E T;ans.. 1988, PWRS-3, ( 2 ) , pp. 392-399 2 PARK. J.. PARK. Y.. and LEE. K.: ‘Comuosite modellinn for adaptive short-term load forecasting’, IEEE Power Systems En&- neering Summer Meeting, 1990 HUBELE, N.F., and CHENG, C.S.: ‘Identification of seasonal short-term forecasting models using statistical decision functions’. IEEE Trans., 1990, PWRS-5, (l), pp. 40-45 4 PAPALEXOPOULOS, A.D., and HESTERBERG, T.C.: ‘A regression-based approach to short-term system load forecasting’, IEEE Trans., 1990, PWRS-5, (4), pp. 1535-1547 5 CAMPO, R., and RUIZ, P.: ‘Adaptive weather-sensitive short term load forecast’, I E E E Trans., 1987, PWRS-2, (3), pp. 592- 600 YEMURI, S.V., HILL, E.F., and BALASUBRAMANIAN, R.: Load forecasting using stochastic models’, IEEE Power Industry Computer Applications Conference, 1973, pp. 31-37 HAGAN, M.T., and BEHR, S.M.: ‘The time series approach to short term load forecasting’, I E E E Trans., 1990, PWRS-2, (3), pp. 785-791 3 6 7 8 PARK, D., EL-SHARKAWI, M., MARKS, R., ATLAS, A., and DAMBORT, M.: ‘Electric load forecasting using an artificial neural network‘, I E E E Trans., 1991, PWRS-6, ( 2 ) , pp. 442-449 EL-SHARKAWI, M.A., OH, S., MARKS, R.J., DAMDORG, M.J., and BRACE, C.H.: ‘Short term electric load forecasting using an adaptively trained layered perceptron’, Proceedings of the first international forum on Applications of Neural Networks to Power Systems, 1991, pp. 3-6 10 CONNER, J.T., ATLAS, L.E., and MARTIN, D.: ‘Recurrent neural networks and load forecasting’, Proceedings of the first international forum on Applications of Neural Networks to Power Systems, 1991, pp. 22-25 11 LEE, K.Y., CHA, Y.T., and KU, C.C.: ‘A study of neural net- works for short-term load forecasting’, Proceedings of the first international forum on Applications of Neural Networks to Power Systems, 1991, pp. 26-30 12 PENG, T.M., HUBELE, N.F., and KARADY, G.G.: ‘Advance- ment in the application of neural networks for short-term load forecasting’, IEEE Power Engineering Society Summer Meeting, San Diego, 1991 13 PENG, T.M., HUBELE, N.F., and KARADY, G.G.: ‘An adap- tive neural network approach to one-week ahead load forecast- ing’, I E E E Trans., 1993, PWRS-8, (3), pp. 1195-1202 14 LAMBERT TORRES, G., VALIQUETTE, B., and MUKHAU- KAR, D.: ‘Short-term feeder load forecasting using fuzzy logic’, IFAC symposium on Power System and Power Plant Control, Korea, 1989, pp. 22-25 15 KIM, K.-H., PARK, D., and PARK, J.: ‘A hybrid model of arti- ficial neural network, and fuzzy expert system for short-term load forecast’, Proceedings of ESAP-IV, Australia, 1993, pp. 164-168 16 SRINIVASAN, D., LIEW, A.C., and CHANG, C.S.: ‘Fuzzy neural network for generation of daily load profiles’, Proceedings of ISDEM Conference, Singapore, 1993, pp. 470477 17 ZADEH, L A . : ‘Fuzzy sets’, Informat. Control, 1965, 8, pp. 318- 358 18 PAL, S.K., and MAJUMDAR, D.D.: ‘Fuzzy mathematical approach to pattern recognition’ (Halstead Press, New York, 1986) 19 WANG, L.-X., and MENDEL, J.M.: ‘Fuzzy basis functions, Universal approximation, and orthogonal least-square learning’, I E E E Trans., 1992, “-3, (5), pp. 807-814 9 11 Appendix Weight updating using error backpropagation for fuzzy neural network The least-mean-square error in output vectors is mini- mised by using a gradient-descent algorithm by starting with any set of weights and repeatedly updating each weight by an amount where q = learning rate a, p = momentum coefficients n = iteration number E = error cost function Further, the learning rate q is adaptively varied as (19) where y determines the tuning of learning rate q . 114 IEE Pvoc.-Genev. Tvansm. Distvib., Vol. 143, N o . 1, January 1996 
work_bm5nxbzdibgs7nrxpam3y4s67u	----	平凉_平凉今日头条娱乐资讯 平凉_平凉今日头条娱乐资讯 您不具备使用所提供的凭据查看该目录或页的权限。 请尝试以下操作： 如果您认为自己应该能够查看该目录或页面，请与网站管理员联系。 单击刷新 按钮，并使用其他凭据重试。 HTTP 错误 401.1 - 未经授权：访问由于凭据无效被拒绝。 Internet 信息服务 (IIS) 技术信息（为技术支持人员提供） 转到 Microsoft 产品支持服务 并搜索包括“HTTP ”和“401 ”的标题。 打开“IIS 帮助”（可在 IIS 管理器 (inetmgr) 中访问），然后搜索标题为“身份验证”、“访问控制”和“关于自定义错误消息”的主题。 平凉_平凉今日头条娱乐资讯 
work_bn33npet6jgr3ksyqtnodn2xpy	----	NASA Technical Reports Server (NTRS) 
work_bp7byanef5hhpjll25h56u7wp4	----	An Expert System on Site Selection of Sanitary Landfill 1 International Journal of Environment and Pollution, Vol. 28, No. 3-4, 2006, pp. 402-411 An Expert System on Site Selection of Sanitary Landfill Kwokwing Chau Department of Civil & Structural Engineering, The Hong Kong Polytechnic University, Hunghom, Kowloon, Hong Kong (Email: cekwchau@inet.polyu.edu.hk) Abstract It is desirable to encapsulate the heuristic and empirical knowledge including hydrological and bio-geochemical considerations into the selection process of a potential landfill site. This paper delineates a prototype expert system, which has been developed using a commercially available microcomputer-based expert system shell VISUAL RULE STUDIO, for the selection of a landfill site. VISUAL RULE STUDIO acts as an ActiveX Designer under the Microsoft Visual Basic programming environment, with hybrid knowledge representation approach under object-oriented design environment. The blackboard architecture, which is able to group all the knowledge together into an integrated system effectively, is adopted. The evaluation is based on the hazardous waste site ranking system recommended by the U.S. Environmental Protection Agency, adapted to Hong Kong conditions. Increase in efficiency, improvement, consistency of results and automated record keeping are among the advantages of such a system. Keywords: Sanitary landfill; Expert system; Blackboard architecture; Hazardous waste site ranking system; Site selection Biographical notes Dr. Kwokwing Chau is currently an Associate Professor in Department of Civil and Structural Engineering of The Hong Kong Polytechnic University. He received both his BSc(Eng) and MPhil in Civil Engineering from the Department of Civil Engineering, The University of Hong Kong, Hong Kong. He received his PhD in Civil engineering from the Department of Civil Engineering, The University of Queensland, Australia. He is very active in undertaking research works and the scope of his research interest is very broad, covering numerical flow modeling, water quality modeling, hydrological modeling, knowledge-based system development and knowledge management. Introduction This is the Pre-Published Version. A sanitary landfill is one of the most popular refuse disposal means in an urban environment. In order to alleviate the nuisances to public health or safety, extensive studies must be carried out to select the most feasible location. This site selection process involves the study of a proliferation of factors including hydrological as well as bio-geochemical considerations such as composition of contained wastes, possible migration paths of the wastes, planned use of the landfill, population and land use, etc. In the evaluation process, it involves many decisions to be made by the designer based on empirical rules of thumb, heuristics, expertise, judgment, code of practice and previous experience. It is difficult for various teams composed of engineers and scientists to possess the broad expertise and knowledge in performing such site selections in a consistent manner. Specifically, a novice designer may face many difficulties in the design process. It is desirable to encapsulate this heuristic and empirical knowledge into the decision making process. Previously, some algorithmic models were built to deal with the problem. In many instances, it is often difficult for anyone other than the model developer to use the model successfully. Moreover, expert knowledge is not necessarily algorithms or effective procedures of computer sciences. In fact, empirical rules are often not complete, or unique, or in specific frameworks. With the advent of artificial intelligence (AI), an expert system furnishes a solution to this decision making process through the incorporation of symbolic knowledge processing on the basis of the pertinent empirical, heuristic and expert rules. During the past decade, the potential of AI techniques for providing assistance in the solution of engineering problems has been recognized. Expert systems are considered suitable for solving problems that demand considerable expertise, judgment or rules of thumb. Areas of early applications of expert system technology include medical diagnosis, mineral exploration and chemical spectroscopy. Expert systems have widespread applications in different fields and are able to accomplish a level of performance comparable to that of a human expert (Adeli & Kao 1996, Chau 1992, Chau & Albermani 2001, Chau & Anson, 2002, Chau & Chen 2001, Chau & Ng 1996, Chau & Yang 1993, Chau & Zhang 1995, Lin & Albermani 2001, Ranga Rao & Sundaravadivelu 1999, Shwe & Adeli 1991). This paper delineates a prototype expert system for the selection of a landfill site, LANDFILL, which has been developed by using a commercially available microcomputer-based expert system shell VISUAL RULE STUDIO. The blackboard architecture with hybrid knowledge representation techniques including production rule system and object-oriented approach is adopted. The expert system developed is based on the uncontrolled hazardous waste site ranking system recommended by the US Environmental Protection Agency (1984), adapted to Hong Kong situations. Through custom-built 2 interactive graphical user interfaces, the user is directed throughout the rating process. Solution strategies and development techniques of the system are addressed and discussed. Expert systems Expert systems are interactive computer programs that mimic and automate the decision making and reasoning processes of human experts in solving a specific domain problem, through delivering expert advice, answering questions, and justifying their conclusions. A typical expert system is constituted by three core components, namely, knowledge base, inference mechanism and context. The knowledge base is a collection of general facts, rules of thumb and causal models of the behavior specific to the problem domain. The inference mechanism guides the decision making process by using the knowledge base to manipulate the context. The context contains facts that reflect the current state of the problem, constructed dynamically by the inference mechanism from the information provided by the user and the knowledge base. Besides, there are three auxiliary components, which are a knowledge acquisition module, a user interface and an explanation facility. The knowledge acquisition module serves as an interface between the experts and the expert system and provides a means for entering domain specific knowledge into the knowledge base. The user interface is responsible for translating the interactive input as specified by the user to the form used by the expert system. The explanation module provides explanations of the inferences used by the expert system, namely, why a certain fact is requested and how a conclusion was reached. The order of execution of a conventional program is predetermined and any updates may require enormous effort. The programming algorithm is designed to ensure completeness and uniqueness of the solution. The model user often considers the program as a blackbox and does not have thorough understanding of the algorithm. On the other hand, an expert system separates the knowledge base from the control strategy, which allows for incremental addition of knowledge without manipulation of the overall program structure and hence the programmer need not guarantee completeness. By ranking several alternatives with inexact inference methods, several solutions with different confidence factors can be provided for a particular input condition. The user can also question the results through the explanation module. The domain knowledge can be represented in several approaches, namely, production rules, frames and object-oriented programming. A production rules system, comprising a collection of rules, is good at depicting heuristic knowledge. A frame system is suitable for a complex 3 and rich representation of knowledge. Object-oriented programming concept is used when it comprises a number of independent objects that process jobs by exchanging information via messages. A hybrid approach can be adopted to take advantages of both representation methods. Several tools are available for developing an expert system, i.e. traditional programming languages, high-level tools or expert system shells. The programming language designed for AI (LISP or PROLOG) entails tremendous programming effort. High-level tools, such as KEE (Intellicorp 1986), can provide an integrated knowledge engineering environment combining features of the AI languages appropriately and efficiently. An expert system shell provides the skeleton of an expert system incorporating inference engine, user interface and knowledge storage medium, thus requiring the developer only to fill in the domain knowledge. Domain knowledge In general, landfills offer an economic and viable solution to the waste disposal problem. During the landfill operation process, engineering principles are applied to confine the refuse to the smallest practicable size, and to cover it with layers of earth at the end of daily operation or at a higher frequency as appropriate. This operation requires systematically depositing, compacting, and covering the wastes in compliance with specifications such as placing 150 to 300 mm of earth over every 600 mm of compacted fill. Besides, the top earth cover should have a minimum designated depth and be grassed to prevent erosion. A common projected land use of a landfill is a park with recreational facilities that are not affected by gradual ground subsidence. Thorough investigation must be undertaken to select the most appropriate location for a landfill site. In the site selection process, four main factors are considered, namely, potential migration routes of the wastes, waste characteristics, planned features of the landfill and potential targets at risks. For potential migration routes, the study of contaminant movement away from the disposal site through air, surface water and ground water is focussed. The movement of wastes depends not only on physical, geographic and climatic characteristics of the site, but also on the waste characteristics. This feature can be quantified by its toxicity, persistence, corrosiveness, reactivity, ignitability, radioactivity, solubility and volatility. The planned features of the facility will affect the degree of contamination. Potential targets subsume population centers, critical habitats and sensitive ecological systems in its vicinity. All the knowledge can be represented so as to minimize or prevent a contaminant from entering into a potential migration route and arriving at a potential target at risk. 4 The site selection process comprises three basic components, namely, (i) data on physical, geographic, climatic, soil, water quality and socioeconomic conditions; (ii) political considerations, guidelines, standards and regulations established by the pertinent authorities; and (iii) expert judgments and heuristics. In this process, the expert plays a key role since an enormous expert effort is entailed in the synthesis and analysis of these components. The following tasks are also required, namely, designing the general scheme and a uniform ranking procedure, identifying constraining regulations, analyzing data, and selecting a specific landfill site together with size. The expert evolves one’s judgments and expertise, which is the result of an iterative and explorative learning process, into a set of explicit rules. Consequently, these rules can be denoted as empirical knowledge, which comprises expert inferences, heuristics and rules of thumb. The prototype system LANDFILL is an expert system designed to provide surrogate consultation during preliminary/detailed hazardous waste site investigations. It provides a versatile framework for the interpretation, classification and diagnosis of environmental conditions at waste disposal sites. In this study we have considered a site selection procedure for a landfill facility with potential ground water migration routes. This system is intended to be used as a consultation tool to assist waste disposal managers in the preliminary phases of site selection. The objective of such a consultation is to obtain a site rating using the expert rules and the decision logic described in the system’s knowledge base. The various options of LANDFILL are designed to assist a user in such an evaluation. Increase in efficiency, improvement and consistency of selection strategies, and automated record keeping are among the expected benefits of such an expert advisory system. The ranking criteria are based on: relative risk or danger, taking into account the population at risk; the hazardous potential of the substances at a facility; the potential for contamination of drinking water supplies, for direct human contact, and for destruction of sensitive ecosystems; and other appropriate factors. The system can be compiled and encrypted to create a run-only system. This run-only system can be installed on a microcomputer for office use. The user can always overrule any design options and recommendations provided by the system. In other words, it plays the role of a knowledgeable assistant only. In order to facilitate development of the expert system on landfill site selection, an expert system shell containing specific representation methods and inference mechanisms is employed. The knowledge base and explanation facility of the system have been 5 implemented with the aid of a microcomputer shell program called VISUAL RULE STUDIO, which is a hybrid application development tool that integrates object-oriented techniques and the expert system technology (Rule Machines Corporation 1998). VISUAL RULE STUDIO acts as an ActiveX Designer under the Microsoft Visual Basic 6.0 programming environment. Production rules as well as procedural methods are used to represent heuristic and standard engineering design knowledge. By isolating rules as component objects, separate from objects and application logic, it produces objects that can interact with virtually any modern development product and thus rule development becomes a natural part of the component architecture development process. Besides, it can activate outside algorithmic programs, and has mathematical computation capabilities. The blackboard architecture has been successfully used in solving a wide range of tasks, such as speech recognition, signal processing, and planning (Engelmore & Morgan 1988). A blackboard system consists of a number of knowledge sources that communicate through a blackboard and are controlled by an inference mechanism. The blackboard architecture is intended to support the development of systems in domains characterized by interaction between diverse sources of knowledge. The blackboard serves as a global data structure, which facilitates this interaction. Objects are used to encapsulate knowledge structure, procedures, and values. An object’s structure is defined by its class and attribute declarations within a RuleSet. Object behavior is tightly bound to attributes in the form of facets, methods, rules, and demons. Figure 1 shows the structure of VISUAL RULE STUDIO components. Each attribute of a class has a specific attribute type. The attribute types are compound, multicompound, instance reference, numeric, simple, string, interval, and time. Facets provide control over how the inference engines process and use attributes. Methods establish developer-defined procedures associated with each attribute. The set of backward-chaining rules that conclude the same attribute is called the attribute’s rule group. The set of forward-chaining demons that reference the same attribute in their antecedents is called the attribute’s demon group. LANDFILL combines expert system technologies, object-oriented programming, relational database models and hypertext/graphics in Microsoft Windows environment. By defining various types of windows as different classes, such as Check Box, Option Button, List Box, Command Button, Text Box, etc., they can inherit common characteristics and possess their own special properties. Besides the usual components in a typical expert system, namely, knowledge base, inference mechanism, session context, user interface, knowledge acquisition and explanation modules, 6 it also incorporates a database. The database chosen is Microsoft Access, which was selected for several reasons including its popularity as a user-friendly relational database within industry, reasonable cost, and Visual Basic support by means of Visual Basic for Applications. The schematic view of this prototype system is shown in Figure 2. Knowledge plays an important role in an expert system. In order to acquire knowledge, it is better to work with the expert in the context of solving particular actual problems, instead of directly posing questions about rules. The knowledge used has been acquired mostly from written documents such as code of practice, textbooks and design manuals and complemented by experts. The domain knowledge is translated into procedures and methods using object-oriented representation. In this prototype system, knowledge is represented in the IF/THEN/ELSE production rules with confidence factors that can be assigned either automatically, or in response to the user’s request. These rules are a formal way of specifying how an expert reviews a condition, considers various possibilities, and recommends an action. The explicit expertise under the production rule format in the knowledge base are employed to rank the potential landfill locations to identify the more feasible sites for further detailed studies. All landfill site selection parameters are categorized into four main factors, namely, groundwater route characteristics, targets at risk, facility characteristics and waste characteristics. The groundwater route characteristics comprise a number of parameters including the depth of the groundwater at the aquifer, net seasonal precipitation, soil permeability, the physical state of the waste at the time of disposal, aquifer soil type and the depth to bedrock. The following is a typical example of the production rules. Rule to find HazardGroundwaterDepth: 1 of 8 IF GroundwaterRoute.DepthOfGroundwater >= 6 AND GroundwaterRoute.DepthOfGroundwater < 25 THEN HazardScore.HazardGroundwaterDepth:= risky CF 70 The system transforms these production rules into user-friendly interactive menus, for instance as shown in Figure 3. Depending on the responses made by the user, appropriate rules will be executed. The sum of executed assigned hazard values will constitute the groundwater route characteristics hazard score (Sgr). The corresponding rules at the second stage characterize the level of risk to potential targets. They include rules about the potential use of the aquifer of concern, the vicinity of nearest 7 production wells, and nearby population served by the ground water. They include three groups of rules. The sum of the three executed assigned values will then yield the target score (St). In order to determine the waste characteristics hazard score (Swc), the system considers the toxicity, persistence, corrosiveness, reactivity, ignitability, radioactivity, solubility and volatility of each potential waste substance, along with its projected disposed quantity. The substance with the highest sum of the assigned values among all projected waste substances is selected as the representative waste material, with the values taken as the waste characteristics hazard score. The remaining stage includes the expert rules about the planned containment characteristics of the proposed landfill facility. The executed assigned value is defined as the facility characteristics score (Sfc). The system, comprising 200 rules or so, accounts for the facts that describe the domain knowledge on preliminary landfill site selection. The production rule also incorporates the fuzzy description: The user has to choose appropriate answers to all these questions in order to assign a hazard rating score to each parameter. The expert system then matches the selected answers with each rule, executes the appropriate ones, and computes the four partial scores. It ultimately derives the overall hazard rating score of the site, as follows: tfcwcgr SSSSS ****57330 100 = (1) The overall site hazard rating score is normalized between 0 and 100, which is utilized as a measure for the relative ranking of the specific site. Figure 4 shows a screen displaying the overall landfill site hazard rating score of a site at Tseng Kwan O. The scores can be employed as a yardstick for the relative safety of different sites. In a preliminary study, locations with high scores might be eliminated as potential landfills, whereas sites with low scores can remain for further comprehensive investigation. The inference engines control the strategies that determine how, from where, and in what order a knowledge base draws its conclusions. These inference strategies model the reasoning processes an expert uses when solving a problem. The Process Control Knowledge module involves meta-level knowledge which establishes the problem solving strategy and controls the execution of the Design Knowledge modules. It evaluates the Design Status and decides what action should be performed mainly in data-driven forward chaining mechanism. The 8 knowledge representations of the Design Knowledge modules, however, need both forward and backward chaining inference mechanism to arrive at the solution. The system offers a friendly user interface. A combination of mouse and keyboard can be used to navigate the application. The use of a mouse or other pointing device renders the data entry a simple task even for novice computer users. As such, users simply point and click their way through the process to appreciate the dynamic behavior of the system. Whilst input data entries are kept at minimum, they are provided by the user mostly through selection of appropriate values of parameters from the menus and answers to the queries made by the system. The input data provided by the user will be rejected if it is not within the specified range. The system provides in multi-window graphics text display where graphic images are combined with valuable textual information. This kind of intelligent graphics is extremely valuable to novice designers because it enhances their confidence in the design provided by the expert system. The HELP command buttons provide definitions of a variety of parameters involved in order that the user can select the appropriate options. Their primary functions are to aid the user to comprehend the expert’s approach to the problem of landfill site selection, and to gain synthetic experience. In fact, the explanation facility is one of the distinct differences between a conventional computer program and an expert system, which is designed to explain its line of reasoning for acquiring an answer. Conclusions A microcomputer expert system, which assists in making decisions on selection of an appropriate landfill site, was developed and implemented. The prototype system illustrates that an expert system can act as storage for empirical knowledge so that advice on landfill site selection can be furnished in the preliminary design stage. It is shown that the hybrid knowledge representation approach combining production rule system and object-oriented programming technique is possible with the implementation of blackboard system architecture under a Windows platform. The knowledge base is transparent and can easily be updated, which render the expert system an ideal tool for incremental programming. Besides, its explanation facilities are capable of offering valuable information to the user, which can lead to a more efficient planning procedure. The educational spin-off of an expert system in training novice engineers or in transferring knowledge cannot be overemphasized. References 9 Adeli, H. & Kao, W.M. 1996. Object-oriented blackboard models for integrated design of steel structures. Computers & Structures 61(3): 545-561. Chau, K.W. 1992. An expert system for the design of gravity-type vertical seawalls. Engineering Applications of Artificial Intelligence 5(4): 363-367. Chau, K.W. & Albermani, F. 2002. Expert system application on preliminary design of liquid retaining structures. Expert Systems with Applications 22(2): 169-178. Chau, K.W. & Anson, M. 2002. A knowledge-based system for construction site level facilities layout. Lecture Notes in Artificial Intelligence 2358: 393-402. Chau, K.W. & Chen, W. 2001. An example of expert system on numerical modelling system in coastal processes. Advances in Engineering Software 32(9): 695-703. Chau, K.W. & Ng, Vitus 1996. A knowledge-based expert system for design of thrust blocks for water pipelines in Hong Kong. Journal of Water Supply Research and Technology - Aqua 45(2): 96-99. Chau, K.W. & Yang, Wen-wu 1993. Development of an integrated expert system for fluvial hydrodynamics. Advances in Engineering Software 17(3): 165-172. Chau, K.W. & Zhang, X.N. 1995. An expert system for flow routing in a river network. Advances in Engineering Software 22(3): 139-146. Lin, S. & Albermani, F. 2001. Lattice-dome design using a knowledge-based system approach. Computer-aided Civil and Infrastructure Engineering 16(4): 268-286. Engelmore, R. & Morgan, T. 1988. Blackboard systems. Wokingham: Addison_Wesley. Intellicorp 1986. KEE software development user manual, 3rd ed. Intellicorp Ranga Rao, A.V. & Sundaravadivelu, R. 1999. A knowledge based expert system for design of berthing structures. Ocean Engineering 26(7): 653-673. Rule Machines Corporation 1998. Developer’s guide for Visual Rule Studio. Indialantic: Rule Machines Corporation. Shwe, T.T. & Adeli, H., 1991. An expert system for earthquake design code processing. Structural Engineering Review 3: 41-48. US Environmental Protection Agency 1984. Uncontrolled hazardous waste site ranking system, a user manual, HW-10. Washington D.C.: USEPA. 10 Class Name Properties Attributes Facets Methods Rules Demons Name Type Instance Attribute Value Figure 1. Structure of VISUAL RULE STUDIO components. 11 User interface Explanation facility Knowledge acquisition facility Context Inference mechanism Knowledge base User Expert Microsoft Access database Figure 2. Schematic view of the expert system. 12 Figure 3. Screen displaying user-friendly data input of groundwater route characteristics. 13 14 Figure 4. Screen displaying overa landfill site hazard rating score. ll 
work_bslkbhrqkbfwbhqxhkjzv35vpm	----	Korean J. Chem. Eng., 19(4), 545-551 (2002) 545 †To whom correspondence should be addressed. E-mail: swpark@mail.kaist.ac.kr ‡This paper is dedicated to Professor Wha Young Lee on the occasion of his retirement from Seoul National University. Development of Refinery Scheduling System Using Mixed Integer Programming and Expert System Jin-Kwang Bok, Heeman Lee*, Jay Woo Chang** and Sunwon Park***† Samsung SDS Co., Ltd., 707-19 Yeoksam-dong, Gangnam-gu, Seoul 135-918, Korea *Technoleaders Co., Ltd., ***Department of Chemical and Biomolecular Engineering, KAIST, 373-1 Guseong-dong, Yuseong-gu, Daejeon 305-701, Korea **SK Engineering & Construction Co., Ltd., 192-18 Gwanhun-dong, Jongno-gu, Seoul 110-300, Korea (Received 22 January 2002 • accepted 8 March 2002) Abstract−−−−This paper presents a hybrid refinery scheduling system combining mathematical programming model and expert system. Mixed-integer linear programming models for crude oil movement between units are merged into the expert system that is for qualitative issues concerning crude vessel unloading operations. The target problem ranging from the crude unloading to the crude charging to distillation towers is decomposed into several module problems for efficiency. Compared with existing scheduling approaches for oil movement, the proposed hybrid refinery schedul- ing system is very effective in dealing with timing decisions involving vessel unloading operations due to the advan- tages of an expert system. Since the proposed scheduling system can generate solutions so fast, it is expected to play a key role in the real processes. Key words: Refinery Scheduling System, Mixed-Integer Programming, Expert System, Oil Movement INTRODUCTION The main role of a refining operation is to convert crude oil mix into a variety of marketable products through a number of refining processes. The major scheduling objective is to provide crude mix- ing with desired sulfur levels which will yield required distillates [Lee et al., 1996]. It is difficult to achieve this due to scheduling constraints imposed by large shipment sizes and pipeline capacity limits between the charge tank and the refinery tower. Scheduling of crude receiving, blending, and CDU operations to meet distillate and fuel oil production requirements currently drives the refinery management problem [Lee et al., 1992]. While long-term refinery planning problems have been exten- sively studied and the developed algorithms have been commer- cially implemented through the software such as RPMS [Bonner and Moore Corps., 1979] and PIMS [Bechtel Corps., 1993], it is quite recent that mathematical programming approaches were applied to the short-term scheduling of refinery processes that are mainly related to the crude oil unloading and inventory management [Lee et al., 1996; Shah, 1996]. Lee et al. [1996] presented a mixed-in- teger linear programming model for crude-unloading operation, oil movement between tanks, and crude-charging schedule, but the ser- ious shortcoming of the model is the large computational expense. The complication is mainly invoked by a large number of 0-1 var- iables regarding the timing decision of vessel unloading over the discrete time domain, whihch makes industrial implementation of the developed model difficult. Furthermore, it cannot take into ac- count the operational contingency that is ubiquitous in the real in- dustry. Shah [1996] also addressed the crude oil management prob- lem using the two-stage mathematical programming model. How- ever, the model did not support the timing decisions with respect to the vessel operation and the mixing operation that is of practical importance in the refinery processes relying on many different crude types imported overseas. In order to overcome the shortcomings that appeared in the previous literature, this paper proposes a hy- brid scheduling system combining a mathematical programming approach and an expert system that is one of the AI-based ap- proaches. For the required amount of crude processed in CDUs given in the planning stage, the mathematical optimization formulation yields the amount of oil movement between units: storage tank- charge tank and charge tank-crude distillation unit (CDU). The move- ment timings between units are determined by the expert rules in order that the number of changeovers involving significant costs should be minimized. The decision concerning vessel docking and crude unloading, where the so-called fast-unloading heuristics are prevailing to reduce demurrage cost in the field, is also established by the expert system. Even though the expert system cannot neces- sarily provide the optimal solution, it significantly contributes to reducing computation time and coping with the operational contin- gencies. In particular, it is also renowned as potentially capable of handling large-sized real problems. This paper is organized as follows. First, the nature of refinery processes with the information flow for decision making in the as- pect of oil movement is outlined. The decision problem is divided into several event-based modules according to distinct operation features. Next, the proposed methodologies implemented in this scheduling algorithm follow. For each module, a mathematical pro- gramming model or expert rules are described. Finally, the devel- oped scheduling system is demonstrated through simulating an indus- 546 J.-K. Bok et al. trial scaled refinery process. NATURE OF OIL MOVEMENT IN REFINERY PROCESS The refinery planning and scheduling department has the respon- sibility for satisfying requirements for product shipment ensured by management headquarters. That is, operation under the short- term scheduling should output the operation results meeting the target established in the planning level. The major detailed activities for scheduling are as follows: • Crude receiving and mixing • Process unit operations scheduling • Offsite inventory management and product blending and ship- ping The scope of the target problem with respect to movement and inventory control is depicted in Fig. 1. The refinery system consists of a storage tank group, a charge tank group and a CDU group. Most of the refinery companies have limited the number of tanks or CDUs due to the initial investment. Crude vessels route many oil produc- ing countries, such as Middle East, South East, Red Sea, West Africa, Fig. 1. Oil movement scheme in the refinery process. Table 1. Specification of each crude oil (per Barrel) Crude name Origin Crude type Component Properties MVOL1RC QSUL1RC QCST1RC ALG Algerian COND 0.0121 0.1713 15.0200 ATK Attaka COND 0.1676 0.2801 156.1000 AMB Arimbi L/S 0.5574 0.4200 320.0000 ORI Oriente L/S 0.4835 1.8400 1591.0000 ARH Arabian Heavy H/S 0.4783 3.6800 3190.6000 KWT Kuwait H/S 0.4078 4.4600 1075.0000 COND: Condensate crude oil MVOL1RC: Yields of reduced crude L/S: Low sulfur crude oil QSUL1RC: Contained sulfur in reduced crude H/S: High sulfur crude oil QCST1RC: Contained viscosity of reduced crude July, 2002 Development of Refinery Scheduling System Using Mixed Integer Programming and Expert System 547 North America, etc., with the lead time of crude delivery ranging from a few weeks to three months or more. The purchased crude oil consists of a complex mixture of hydrocarbons. The mixture from different sources has different properties as shown in Table 1. Delivery vessels arrive at the docking place linking to storage tanks by pipes for the purpose of crude unloading. The demurrage cost is so much greater than the inventory cost, not less than thou- sands-time in the order of magnitude, that the arriving vessel should start operation promptly. One refiner usually incurs demurrage cost of $2 million/year [King, 1993]. When no dock is available due to the occupation by the other previous vessel, the present vessel should wait for the next turn. The delivered crude is unloaded to storage tanks. Each storage tank is managed according to tank characteris- tics, i.e., the concentration of specific component in the tank. The crude in a storage tank should be mixed in a charge tank in order to provide the requirement for distillation. In contrast with the blend- ing of intermediate products, crude mixing has rarely been regarded as an optimization problem [Lee et al., 1996]. The mixed crude blend with the strict range of sulfur concentration is fed into the distilla- tion tower. Thereafter, for complete refinery products, the crude dis- tillate from CDU should go through the other remaining steps that have been presented by many other researchers [Gifford, 1955; Lee, 1992]. HYBRID SCHEDULING MODEL FOR CRUDE OIL UNLOADING AND INVENTORY MANAGEMENT Seward et al. [1985] suggested that Hierarchical Production Plan- ning (HPP) can be more widely applicable in capacity-oriented com- panies such as primary metals and chemical processing rather than Material Requirement Planning (MRP). In such a context, total pro- duction planning in an oil refinery company is viewed at different hi- erarchical levels: aggregated production planning, operational sched- uling, and process operation. The operational scheduling of our con- cern receives information such as monthly distillation amount for each CDU from the aggregated production planning levels in order to provide guidance for the daily process operation. The oil movement problem can be represented as several module problems: docking of vessel, crude unloading, transferring and mix- ing of crude, and feeding to CDU. Each module problem is treated autonomously with the objective specific to the module and with the algorithm efficient to the module. For instance, the crude un- loading module, whose objective is to unload the crude from the vessel as fast as possible adopts an expert system that decides the most adjustable storage tank for each crude set by taking the in- ventory and the characteristic of the tank into consideration. The decision flow in the view of oil movement is shown in Fig. 2. Oper- ation level for each CDU is computed based on the amount of crude processed in CDUs over the horizon given by the aggregated pro- duction planning. Next, the optimization on mixing operation is per- formed to maximize the inventory efficiency of charging. Note that the concentration of a specific component in each charge tank should be kept within the specified range. In the crude unloading module, the crude that is purchased by routing the several oil-producing coun- Fig. 2. Decision flows in the proposed algorithm. Fig. 3. Event-based time representation. Fig. 4. Independence between each module operation. Korean J. Chem. Eng.(Vol. 19, No. 4) 548 J.-K. Bok et al. tries and transported by vessel without being mixed is unloaded to storage tanks. As shown in Fig. 3 each module operation is event-based, and in general two or three module operations can be carried out sim- ultaneously. In spite of these overlapped operations, each module works independently from the other modules since neither storage nor charge tanks involve simultaneous oil inflow and outflow op- eration. This is illustrated in Fig. 4. Fig. 5 shows the framework of the proposed scheduling system. The proposed scheduling system is implemented by the real time expert system G2 [Gensym, 1995] that provides a built-in rule-based inference engine as well as simulation and monitoring functions. The mathematical optimization in the modules is carried out through GAMS [Brooke et al., 1992]. 1. Docking of the Vessel Generally, the demurrage cost is so high that the vessel arriving at the wharf should search for a docking place as soon as possible. The docking decision is made based on the fact that each docking place is designed for accommodating the specified range of vessel size. As Fig. 6 shows, the arrival time is determined considering whether there is any other vessel on the docking place that is suitable for the current vessel, or not. 2. Crude Feeding to CDU The movement amount is obtained through a mixed integer linear programming optimization and the movement time is determined by the expert system implementing operation rules. The deviation of the amount of crude processed in each CDU from the specified amount in the upper planning level is taken as the objective func- tion to be minimized subject to a set of constraints. (1) Subject to (2) (3) The weight factor for each CDU accounts for the different signif- icance of satisfying the target amount of crude processed in each CDU established in the aggregated planning. Note that the amount of crude processed in each CDU below the target amount is taken into account in accordance with the practical sense as shown in Eqs. (2) and (3). (1) Amount of Distillation Product from Crude Feeding (4) (5) The amount of product p at CDU k is computed by Eq. (4). The yield data of each charge tank j to produce product p at CDU k, apjk is reasonably assumed to be constant if the sulfur composition of each charge tank is kept within the specified range. YDij denotes whether charge tank j is connected to CDU k or not. Only when charge tank j is assigned to CDU k, the transferring amount CRjk has the meaning that Eq. (5) implies. CRmax, jk is bounded by the pump capacity between the charge tank j and CDU k and CRmin, jk is zero. (2) Changeover Constraint (6) For each decision, each CDU k should be connected to one of charge tanks since the cost involved in changeovers from a charge tank to another is high. (3) Distillation Capacity Limitation (7) For CDU k, the amount of distilled should be less than the maxi- mum capacity. (4) Charging Amount Limitation (8) The summation of amounts charged to all CDUs from a charge tank j cannot exceed the initial inventory. The timing of assignment of charge tanks to CDUs is determined by the expert rule that prefers selecting a charge tank with the largest Minimize ωk∆k k ∑ ∆k Qk* − Qpk p ∑≥ k∀ ∆k 0≥ k∀ Qpk = ap jkCRjk j ∑ p∀ k∀, CRm in jk, YDjk CRjk CRm ax jk, YDjk≤ ≤ p∀ k∀, YDjk = 1 j ∑ k∀ Qpk Qmax k,≤ j ∑ k∀ CRjk ICOj≤ k ∑ k∀ Fig. 5. Information flows between the optimization solver and the expert system. Fig. 6. Vessel docking module. July, 2002 Development of Refinery Scheduling System Using Mixed Integer Programming and Expert System 549 moving amount for each CDU so as to reduce the number of chang- eovers. 3. Crude Mixing in the Charge Tank In order to prepare a CDU with the crude blend having the spe- cified sulfur composition, the crudes from the storage tanks should be mixed properly in the charge tank. The inventory level of the charge tank should be maintained high enough to provide the crude blend for the CDU with the minimum number of changeovers and to lower the inventory of storage tank in order to ensure enough space for the crude unloading. A mixed integer linear programming model for maximizing the quantity of available crude from storage tanks is introduced: (9) Subject to (1) Sulfur Composition Range For each crude tank, the crude should be mixed within a speci- fied composition range. (10) The nonlinear inequality, which can be linearized by using stan- dard arithmetic techniques, gives two linear inequalities as follows: (10a) (10b) (2) Inventory Mass Balance (11) (12) For each charge tank, the inventory after mixing operation is obtained by adding the transferred amount from storage tanks to the initial amount. Tmax, ij is bounded by the pump capacity between storage tank i and charge tank j and Tmin, ij is zero. (3) Assignment of Storage Tank to Charge Tank (13) Each charge tank j that is not assigned to CDU k should be con- nected to a storage tank that is not involved in crude unloading. (4) Transferring Limitation (14) The maximum transferable amount cannot exceed the initial inven- tory of each storage tank. (5) Inventory Capacity Limitation (15) Each charge tank has a capacity limitation. 4. Crude Unloading to the Storage Tank The target unloading tank and unloading time for the crude on the vessel is determined by considering the assay data of the crude and the state of storage tanks. The main idea in this algorithm is intimately associated with so called fast-unloading. All the storage tanks are classified into several logical groups according to the qual- ity of the present crude in the tank: high sulfur, low sulfur, etc. The concentration calculation is performed under the assumption that two different kinds of crude are perfectly mixed regardless of the density. The implemented rules represent that each crude set should be unloaded to the storage tank with the minimum inventory in the corresponding tank group. EXAMPLE The proposed scheduling system is applied to a general refinery process: 6 storage tanks, 6 charge tanks, and 3 CDUs. Fig. 7 shows the result of unloading scheduling by the expert system. For each set of crudes on the vessel, the start and the end time of unloading and the best suitable unloading place are determined by expert rules. With the movement amount and movement timing obtained from the optimization models and expert rules, respectively, the sched- ule for oil movement is illustrated in Fig. 8. Fig. 9 shows the sulfur concentration trend that should be maintained within specified range. Maximize Tij j ∑ j ∑ SCLj SSiTij + SCjICj i ∑ Tij + ICj i ∑ ---------------------------------------- SCUj≤ ≤ i∀ SSi − SCLj( )Tij + SC j − SCLj( )ICj 0≥ i ∑ j∀ SSi − SCUj( )Tij + SC j − SCUj( )ICj 0≤ i ∑ j∀ ICj = ICOj + Tij i ∑ j∀ Tm in ij, YCij Tij Tm ax ij, YCij≤ ≤ i∀ j∀, YCij = 1 i i Ic{ }∉ ∑ j∀ − Jk{ } Tij ISOi≤ j ∑ i∀ ICj ICmax j,≤ j∀ Fig. 7. Scheduling result of crude unloading. Fig. 8. Inventory trend in inventory tanks 1, 2 and 3. Korean J. Chem. Eng.(Vol. 19, No. 4) 550 J.-K. Bok et al. CONCLUSIONS A hybrid refinery scheduling system combining a mixed integer programming model and an AI-based expert system was developed. Quantitative decisions such as determining the amount of crude mix- ing or the amount of crude charging were made through the mixed integer linear programming models, and qualitative complications regarding vessel unloading operation and the movement timing be- tween the units were resolved with the expert system. For the monthly amount of distillation products given in the aggregated plan- ning step, the amount of crude that should be processed in each CDU is determined. Then, a decision is made on the mixing in charge tanks in order to provide crude to each CDU. The storage tank with the minimum inventory in the corresponding logical group is se- lected for crude unloading. The simulation study indicates that the proposed scheduling system can be applied to industrial scale refin- ery problems. ACKNOWLEDGMENT This work was partially supported by the Brain Korea 21 pro- ject and SK Engineering & Construction Co., Ltd. NOMENCLATURE (a) Indices b : buoy i : storage tank j : charge tank k : CDU v : vessel (b) Sets Ic : storage tanks that receive crude c Jk : charge tanks that is connected to CDU k (c) Variables CRjk : amount of crude transferred from charge tank j to CDU k ICj : available inventory of charge tank j Qpk : total quantity of product p from CDU k SSi : sulfur composition of storage tank i Tij : crude amount transferred from storage tank i to charge tank j YCij : binary variable denoting if storage tank i is connected to charge tank j YDjk : binary variable denoting if charge tank j is connected to CDU k ∆k : deviation of the amount of the crude in each CDU k from the amount given by the aggregated planning step (d) Parameters apjk : yield of product p from charge tank j in CDU k CRmax, jk : maximum transferable amount from charge tank j to CDU k CRmin, jk : minimum transferable amount from charge tank j to CDU k ICmax, j: inventory limitation of charge tank j ICOj : initial inventory of charge tank j Qmax, k : maximum capacity of CDU k SCLj : lower sulfur limit in charge tank j SCUj : upper sulfur limit in charge tank j Tmax, ij : maximum transferable amount from storage tank i to charge tank j Tmin, ij : minimum transferable amount from storage tank i to charge tank j ωk : weight factor that accounts for penalizing deviation of dis- tillation amount REFERENCES An, D. M., Hou, B. K. and Hwang, K. S., “Automatic Synthesis of Op- erating procedures for Safe Emergency Operation,” HWAHAK KONGHAK, 39, 30 (2001). An, D. M. and Hwang, K. S., “A Study on the Expert System for Pre- dicting Hazardous Conditions of Chemical Process,” HWAHAK KONGHAK, 34, 727 (1996). Anthony, B., David, K. and Alexander, M., “GAMS A USER’S GUIDE,” Boyd & Fraser Publishing Company, Massachusetts, USA (1992). Bechtel Corp., PIMS User’s manual, ver.6.0, Houston, USA (1993). Bonner and Moore Management Science, RPMS A system description, Houston, USA (1979). Chae, H., Yoon, Y. H. and Yoon, E. S., “Safety Analysis Using an Ex- pert System in Chemical Processes,” Korean J. Chem. Eng., 11, 153 (1994). Choi, T. H. and Yoon, E. S., “An Expert System to Aid Blast Furnace Operation Using Artificial Neural Network Techniques,” HWAHAK KONGHAK, 29, 270 (1991). Gensym, G2 Reference Manual, version 4.0; Gensym Corp., USA (1995). Hou, B. K., An, D. M. and Hwang, K. S., “An Expert System for Pre- venting Hazardous Conditions in Boiler Plants,” HWAHAK KONG- HAK, 34, 750 (1996). King, M., “Automation of Oil Movement Systems Prevents Mistakes, Saves Money,” Oil & Gas J., 76 (1993). Kim, C. J., Oh, J. K. and Yoon, E. S., “A Fault Diagnostic Expert Sys- tem Using the Symptom Tree Model,” HWAHAK KONGHAK, Fig. 9. Sulfur concentration trend in charge tanks 1, 2 and 3. July, 2002 Development of Refinery Scheduling System Using Mixed Integer Programming and Expert System 551 28, 417 (1990). Kim, Y. J. and Kang, S. K., “Prediction of Binary Azeotrope Formation in Hydrocarbon Mixtures Using a Knowledge-based Expert Sys- tem,” Korean J. Chem. Eng., 12, 306 (1995). Lee, H., Pinto, J. M., Grossmann, I. E. and Park, S., “MILP Model for Refinery Short Term Scheduling of Crude Oil Unloading with In- ventory Management,” Ind. & Eng. Chem. Res., 35(5), 1630 (1996). Lee, J. K., Kim, M. Y., Song, Y. W. and Yoon, H. S., “UNIK-YUKONG: Refinery Scheduling Expert System,” KYOUNGYOUNG KWA- KHAK, 121 (1992). Lee, K. H., Yang, G., Jung, J. Y. and Lee, I. B., “Development of an Ex- pert System for Functional Group Analysis in Group Contribution Method,” HWAHAK KONGHAK, 31, 647 (1993). Seward, S. M., Taylor, S. G. and Bolander, S. F., “Progress in Integrat- ing and Optimizing Production Plans and Schedules,” Int. J. Prod. Res., 23(3), 609 (1985). Shah, N., “Mathematical Programming Techniques for Crude Oil Sched- uling,” Com. Chem. Eng., 20, S1227 (1996). Song, J. and Park, S., “Intellite3: A Knowledge Based Expert System for Control Structure Synthesis,” Korean J. Chem. Eng., 7, 198 (1990). Yoo, C. K., Kim, D. S., Cho, J. H., Choi, S. W. and Lee, I. B., “Process System Engineering in Wastewater Treatment Process,” Korean J. Chem. Eng., 18, 408 (2001). Yoon, B. S., Oh, J. K. and Yoon, E. S., “Knowledge Base Representa- tion for the Fault Diagnostic Expert System Using the Fault-Conse- quence Digraph,” HWAHAK KONGHAK, 29, 19 (1991). Korean J. Chem. Eng.(Vol. 19, No. 4) Development of Refinery Scheduling System Using Mixed Integer Programming and Expert System Jin-Kwang Bok, Heeman Lee*, Jay Woo Chang** and Sunwon Park***† Samsung SDS Co., Ltd., 707-19 Yeoksam-dong, Gangnam-gu, Seoul 135-918, Korea *Technoleaders Co., ... Abstract�-�This paper presents a hybrid refinery scheduling system combining mathematical program... Key words:�Refinery Scheduling System, Mixed-Integer Programming, Expert System, Oil Movement INTRODUCTION NATURE OF OIL MOVEMENT IN REFINERY PROCESS HYBRID SCHEDULING MODEL FOR CRUDE OIL UNLOADING AND INVENTORY MANAGEMENT EXAMPLE CONCLUSIONS ACKNOWLEDGMENT NOMENCLATURE REFERENCES 
work_bsrqzukk65dflfzlpdpo6jxldm	----	PII: 0957-4174(91)90113-S Expert Systems With Applications, Vol. 2, pp. 167-174, 1991 0957-4174/91 $3.00 + .00 Printed in the USA. © 1991 Pergamon Press plc An Expert System for Determining Government Relocation Allowances H A L B E R G H E L , D A V I D R O A C H University of Arkansas, Fayetteville, AR G A R Y G R E E N , J O E M E E H A N National Center for Toxicological Research, Jefferson, AR Abstract-- We describe an expert system whose rule base is a formafization of the subset of the federal travel regulations on relocation allowances. Its purpose is to help government employees making a permanent change of station to identify the reimbursable expenses associated with that change of station. We discuss the structure of the regulations, the implications of this structure for the rule base and inference engine, and the components of the user interface. 1. I N T R O D U C T I O N IN AN EARLIER article, we described a n i m p l e m e n t a t i o n o f a p r o t o t y p e o f the c u r r e n t system, called R A P (for R e l o c a t i o n Allowance Planner), whose rule base con- sisted o f o n l y a small subset o f the applicable regula- tions a n d whose user interface lacked s o m e o f the i m - p o r t a n t features t h a t h a v e since been a d d e d (Roach, Berghel, M e e h a n , & G r e e n , 1989). T h e p r o t o t y p e in- ference engine has also b e e n significantly e n h a n c e d [ R o a c h & Berghel, in press (a), in press (b)]. In this article, we p r o v i d e a n overview o f the m a t u r e version o f the system, with special e m p h a s i s o n the formaliza- tion o f the regulations a n d structure o f the inferential m e c h a n i s m necessary to carry o u t the required deduc- tions. R A P is i n t e n d e d to b e a p l a n n i n g tool for govern- m e n t e m p l o y e e s w h o are to undergo a p e r m a n e n t change o f station. T h r o u g h a consultation with the sys- t e m the e m p l o y e e c a n d e t e r m i n e w h a t kinds o f expen- ses are identified in the regulations as r e i m b u r s a b l e ex- penses. T h e e m p l o y e e c a n p l a n the m o v e to m a x i m i z e the r e i m b u r s e m e n t a n d m i n i m i z e the costs t h a t would h a v e to b e p a i d f r o m personal funds. Because the reg- ulations are c o m p l e x , this a p p r o a c h is m u c h simpler, This work was supported in part by a grant to the first author from the National Center for Toxicological Research. The authors wish to thank the anonymous referees of Expert Systems With Applications for suggesting several improvements to earlier drafts of this paper. Requests for reprints should be sent to Hal Berghel, Center for Artificial Intelligence and Expert Systems, 131 Biomass Research Center, University of Arkansas, Fayetteville, AR 7270 I. less error prone, less t i m e c o n s u m i n g , a n d less costly t h a n a m a n u a l a p p r o a c h . In addition, questions c o n - cerning the validity a n d a c c u r a c y o f the results can be resolved m o r e easily. T h e explanation facility generates a r e p o r t o f the reasoning process with explicit justifi- cations o f the p r o p o s e d allowances for each type o f expense. 2. R E L O C A T I O N A L L O W A N C E R E G U L A T I O N S W i t h i n carefully defined p a r a m e t e r s , the federal gov- e r n m e n t will c o v e r m a n y o f the expenses that o n e o f its employees encounters w h e n a transfer f r o m one post o f d u t y to a n o t h e r is required. T h e conditions, restric- tions, a n d limits pertaining to expense allowances for a change o f station are codified in a set o f regulations k n o w n as the Federal Travel Regulations (U.S. General Services A d m i n i s t r a t i o n [GSA], 1987). T h e relocation allowance regulations are divided into 10 sections (USGSA, 1984, 1987). T h e first section is c o n c e r n e d with the general eligibility conditions for obtaining a n allowance. T h e o t h e r nine sections are d e v o t e d to the different t y p e s o f c o v e r e d expenses. T h e 10 sections o f the m a n u a l are: 1. applicability a n d general rules, 2. travel to the new residence, 3. miscellaneous, 4. travel to seek new residence quarters, 5. o c c u p a n c y o f t e m p o r a r y quarters, 6. b u y i n g a n d selling residences, 7. t r a n s p o r t a t i o n o f h o u s e h o l d goods, 8. t r a n s p o r t a t i o n o f a m o b i l e h o m e , 167 168 9. n o n t e m p o r a r y storage o f household goods, and 10. transportation o f a privately owned vehicle. The first section is c o n c e r n e d with factors such as em- ployee status, cause o f transfer, authorization for transfer, and time and distance restrictions. T h e con- ditions in section (1) must be satisfied before a n y al- lowance can be granted. T h e remaining nine sections discuss conditions u n d e r which an employee is eligible for specific allowances and procedures for calculating those allowances. Section 2 o f the m a n u a l is concerned with the ex- penses for physically m o v i n g from one post to another. This includes subsistence and transportation expenses. T h e parameters o f interest are the size o f the immediate family, the n u m b e r o f days required to complete the move, and the mileage rate. Miscellaneous expenses and allowances (section 3) cover the costs o f a variety o f items and activities such as forfeitures o f nontrans- ferable medical contracts, registration, and license fees, etc. T h e a m o u n t s allowed depend on whether records are kept, the size o f the i m m e d i a t e family, and the a m o u n t o f the employee's regular weekly pay. Section 4 governs the expenses for a pretransfer trip to find housing at the new post o f duty. T r a n s p o r t a t i o n for b o t h the e m p l o y e e and spouse is provided. T h e per diem for the employee and spouse must be calculated as well as the mileage rate i f a personally owned vehicle is used for the trip. T h e n u m b e r o f trips, the length o f the trip, the distance to the new post o f duty, an d the employee's status must be ascertained. If t e m p o r a r y quarters are to be occupied before a p e r m a n e n t new residence is obtained (section 5) the length o f the stay and the n u m b e r and relationship o f other occupants must be k n o w n in order to d e t e r m i n e an allowance a m o u n t . T h e rates vary depending o n the length o f occupancy. T h e r e are n u m e r o u s conditions that must be satisfied before eligibility can be established for this allowance. Section 6 contains the regulations on allowances for the expenses associated with buying and selling homes. Eligibility conditions include having title to old a n d / or new residences, living in the old residence before transferring, and providing evidence o f transaction ex- penses. The covered expenses include broker's fees, real estate commissions, advertising and appraisal costs, le- gal fees, etc. M a x i m u m s are based on percentages o f actual buying and selling prices. T h e expenses for transporting household goods (section 7) are based o n the m e t h o d o f transportation, weight o f the goods, an d costs for t e m p o r a r y storage o f the goods while in transit (if any). All o f these facts must be obtained. M a x i m u m allowances for this category o f expenses exist a n d must not be exceeded. T h e costs for transporting a mobile h o m e (section 8) are also reimbursable. Mileage rates (consequently, allowance decisions) d e p e n d primarily on the m ean s o f transportation: c o m m e r c i a l carrier, private means, H. Berghel et al. some c o m b i n a t i o n o f the two, or a g o v e r n m e n t bill o f lading. Section 9 governs allowances for storing house- hold goods while at the new residence. Conditions for the allowance are that the transfer is to an isolated station where the goods can n o t be stored. Covered costs include packing, crating and transportation to and from the storage warehouse. T h e final section (section 10) is c o n c e r n e d with the transportation o f a privately o w n ed vehicle. T h e r e are n u m e r o u s eligibility condi- tions that m u st be satisfied in order to obtain this al- lowance. T h e size o f the vehicle, who will drive it, why it is being transported, and the cost relative to the length o f the transfer are all issues that must be addressed. 3. T H E M A N U A L S Y S T E M As a large g o v e r n m e n t research institution, the Na- tional Cen t er for Toxicological Research employs a n u m b e r o f research an d support personnel who fall u n d e r the purview o f these regulations. Because there are a significant n u m b e r o f personnel changes occurring each year, the center attempts to address the eligible employees' desires for assistance in obtaining a relo- cation allowance. T h e o b t a i n m e n t o f all an d only those allowances for which an e m p l o y e e is eligible is greatly hindered by the co m p l ex i t y o f the travel regulations. Appendix B o f the Travel Manual, which is devoted to relocation allowances, is roughly 150 pages o f detailed informa- tion (USGSA, 1987). T h e u n w ary employee, if left to his o w n devices in interpreting the regulations, is usu- ally o v erw h el m ed by the m ag n i t u d e o f the problem o f trying to apply t h e m to his particular situation. Thus, the Cen t er provides s o m e o n e who is familiar with the regulations to assist employees in d et erm i n i n g their re- location allowances. T h e process requires that the assistant interrogate the transferee in order to ascertain whether general eli- gibility conditions are met, which expense categories apply, w h et h er specific restrictions an d limitations are met, which rates are to be used, etc. M a n y questions require y e s / n o responses; others require that values for various entities be s u p p l i e d - - t h e n u m b e r o f family members, the distance to be traveled, the weight o f various objects, an d so forth. This, o f course, suggests that the assistant should possess a co m p reh en si v e and detailed knowledge o f the relocation regulations. However, the n u m b e r o f employees transferring, and t h u s requiring some assistance, is not large enough to justify a position d ev o t ed just to providing that kind o f assistance. As a result, the person w h o is charged with the responsibility o f assisting in allowance deter- m i n at i o n does so relatively infrequently; consequently, that person does n o t s p e c i a l i z e in the relocation reg- ulations. This m e a n s that while the expertise at inter- preting an d applying the regulations is greater t h an that o f the layman, it is still less t h an what could be provided An E S for Government Relocation Allowances by a specialist in the area. Because o f the level o f de- m a n d for assistance with the relocation regulations, it is simply not cost effective to dedicate a person full t i m e to the task. T h e result is that the employee w h o is transferring is not getting the level o f assistance that could be provided. 4. T H E A U T O M A T E D S Y S T E M T h e motivation for the a u t o m a t e d system is that it can provide unerring analyses in a cost effective man n er. T h e user interface is intuitive and simple enough that the l a y m a n can use the system to accurately d e t e r m i n e his allowances. T h e suitability o f the allowance deter- m i n a t i o n process to an expert system style solution is a function o f the hierarchical structure and translat- ability o f the regulations and the interactive character o f the d e t e r m i n a t i o n process. T h e character and or- ganization o f the regulations p e r m i t a fairly straight- forward translation to a standard rule base format. Also, the interactive aspects o f the m a n u a l system are au- t o m a t e d in a very natural way via the expert system's user interface. Figure 1 shows RAP's main menu. On the lower part o f the screen the current client is iden- tified, a n d an indication is given whether a trace, a s u m m a r y , a n d / o r a set o f answers exist for that client. These m a k e up the three categories o f i n f o r m a t i o n that result from a consultation. T h e usual options for re- viewing and printing the trace are present. However, R A P also provides a tabular s u m m a r y o f allowance a m o u n t s and a record o f responses given during a con- sultation. T h e tabular s u m m a r y includes a figure for each category o f expenses along with a total allowance a m o u n t . T h e i n f o r m a t i o n gained as a result o f a con- sultation can be stored and retrieved. This includes the trace, the answers supplied by the user during a con- sultation, and the allowance s u m m a r y information. It is possible to resume a previous consultation using a previously stored set o f answers. T h e system proceeds noninteractively using the retrieved answer set. Should a question arise for which there is no previous answer the system will q u e r y the user in the n o r m a l manner. This feature is useful for resuming incomplete consul- F1 HELP Line 1 Col 1 CLIENT: 7248 169 INSERT $yOu$ are Sa civilian officer or employee transferring from one official station or agency to another for permanent dutyS = Syes$. Syou$ have $appreval for travel from the proper official(s)$ = Syes$. Syou$ will $move within 2 years of your date of transfers = Syes$. Syou$ are $transferring in the interest of the Government and not primarily for your convenience or at your requeszS $yesS, $you$ are Sa new appointees = $no$. Syour spouses will Saccompany you on the moves = SyesS. $the numoer of family members age 12 or olderS = $I$. Sthe numoer of family members under age 125 = $i$. $the number of days required to complete the moves = $105. Syou$ will $drive a privately owned vehicles = SyesS. Sthe total miles drivenS = $i0005, Sthe size of your immediate family$ = $25. FIGURE 2. Editor and sample object value response set. tations. T h e previously answered questions need not be repeated. However, even co m p l et ed consultations can be re- sumed. This is usually d o n e with a co m p l et ed answer set that has been modified. An editor is provided that can be used to m o d i fy a previous answer set. Previous answer(s) can be changed, an d the system can generate a new trace a n d allowance s u m m a r y noninteractively. As a result, whole sessions need not be repeated should o n e desire to change a previous response. Figure 2 shows the ed i t o r screen, in which some o f the objects an d associated values resulting fro m a consultation are displayed. Figure 3 shows the screen when a consul- tation is begun. T h e t o p window identifies the section o f the m a n u a l o n which the cu rren t q u e r y is based. T h e q u e r y itself appears in the next window down. T h e middle p o rt i o n o f the screen consists o f a m e n u for obtaining the user responses to the query. YES and N O responses are entered by selecting the appropriate m e n u o p t i o n (see Figure 4). N u m e r i c responses are entered in the small window at the far left o f the m enu, as shown in Figure 5. T h e type o f response that is ap- propriate to a given q u ery is d e t e r m i n e d before it is displayed. Thus, the highlighting defaults to the correct part o f the m en u . Finally, the m e n u provides options for asking why a particular question is asked (the reason is displayed in the b o t t o m window, as in Figure 6) and for aborting the c u r r e n t consultation a n d returning to the main menu. Current Drive: C: RAP MAIN MENU Monday Current Directory: \RAP January 20, 1990 Data Directory: \RAP\DATA 4:00 pm <FI> First Time Consultation <F2> Resume Previous Consultation <F3> Review Allowance Su~unary <F4> Review Trace <FS> Print Allowance Summary <F6> Print Trace <F7> Select Client <FS> Edit Client Answers <F9> Delete Client Information <FI0> Introduction to RAP <ESC> Exit to DOS Client An ..... T .... ))S .... y 7248 YES NO I~ NO FIGURE 1. The mainmenu. Expense Category F__. L Value Query r J ! ! ! YES NO WHY DO YOU ASK? MAIN MENU I Response to Why FIGURE 3. Query screen layout. 170 H. Berghel et al. Current Questions Relate to General Eligibility Are you a civilian officer of employee transferring from ] oee official station or agency to another for permanent J duty? I I I l l i [] No ..Y oo A K? MENo FIGURE 4. Sample yes/no valued query. Current Questions Relate to Actual Moving Expenses How many family members age 12 or older will relocate with you? I I i ] I ~ YES NO WHY DO YOU ASK? MAIN MENU I FIGURE 5. Sample numeric valued query. Current Questions Relate to Actual Moving Expenses How many members are in your immediate family, including yourself, who will relocate with you? i Y~s IwHY Do Yoo AsKs1 MEmo I To determine value for mileage rate. a your I oK FIGURE 8. Reason for asking the question. As a consultation progresses from o n e category o f allowances to the next, the a m o u n t o f the allowance for the completed category is displayed. At the end o f a consultation the trace, s u m m a r y o f allowances, and user responses are saved, and the user is returned to the main menu. At this point, the results o f the consultation can be viewed, printed, modified, etc. The explanation o f the reasoning process for the completed consultation is in the form o f a 'pretty-printed' trace. Figures 7 and 8 show different portions o f a sample trace. The first is a partial justification o f the claim that the target client has met the eligibility conditions, and the second par- tially justifies the a m o u n t allocated for a particular type o f expense. Finally, the client can view or print a tabular s u m m a r y o f the allowances, as s h o w n in Figure 9. Trace of Reasoning for 7248 you are eligible for a relocation allowance because you meet the applicability requirements in 2-1.2 because you are a civilian officer or employee transferring from one official station or agency to another for permanent duty and you meet the general provisions in 2-1.3 because you have approval for travel from the proper officials and you meet the general provisions conditions in 2-1.3a because you are transferring in the interest of the Government and not primarily for you convenience or at your request FIGURE 7. Portion of explanation for client 7248. Trace of Reasoning for 7248 your subsistence/transportation allowance = 1970.00 because your subsistence/transportation allowance = your subsistence allowance per day + the subsistence allowance for your immediate family per day * the number of days required to complete the move + your transportation allowance for relocating and your subsistence allowance per day = 60.00 because your subsistence allowance per day = your per diem and your per diem = 60.00 and the subsistence allowance for your immediate family per day = 120.00 because FIGURE 8. Portion of explanation for client 7248. Allowance Summary for 7248 Subsistence/Transportation S 1970.00 Miscellaneous Expenses S 400.00 Travel to Seek Residence Quarters $ 80O.CC Temporary Quarters $ 200.00 Residence Transactions $ 1700.00 Mobile Home $ 0.00 Household Goods $ 800.00 Nontemporary Storage of Household Goods $ 1300.00 Personally Owned Vehicle $ 600.00 TOTAL $ 7770.00 FIGURE 9. Tabular summary of allowances for client 7248. 5. I M P L E M E N T A T I O N 5.1. Translating Regulations to Rules The prolix r e g u l a t i o n e s e and frequent use o f exclu- sionary clauses are the main impediments to efficient translation o f the regulations. For example, The expenses of travel, transportation, moving and/or storage of household goods, and applicable allowances as provided in these regulations in connection with the transfer or ap- pointment of employees to posts of duty outside the conter- minous United States, or between posts located in (i) separate countries, (ii) separate areas of the United States located out- side the conterminous United States (e.g., Alaska, Hawaii, the Commonwealth of Puerto Rico), or (iii) any combination of the above, shall not be allowed unless and until the em- ployee selected for such transfer or appointment agrees in writing to remain in the service of the Government for 12 An E S for Government Relocation Allowances 171 months following the effective date of the transfer or ap- pointment (or for 1 school year for Department of Defense overseas dependents school system teachers as determined under Chapter 25 of title 20 of the United States Code (U.S.C.)), unless separated for reasons beyond his/her control and acceptable to the agency concerned. (USGSA, 1984, p. 2-5) A careful a b r i d g m e n t o f the regulations is a necessary c o m p o n e n t o f the translation process. On the o t h er hand, their rigorous organization increases their trans- latability. T h e reason is that the hierarchical relation- ship o f all the parts to the whole is readily discernable. T h e following excerpt from the table o f contents gives some indication o f the level o f analysis (USGSA, 1984, p. iii): Allowable travel and transportation . . . . . . 2-1.5h(2) Destination . . . . . . . . . . . . . . . . . . . . . . . 2-1.5h(2)(a) Allowances . . . . . . . . . . . . . . . . . . . . . . . . 2-1.5h(2)(b) Alternate destination . . . . . . . . . . . . . . . . 2-1.5h(2)(c) Limitations . . . . . . . . . . . . . . . . . . . . . . . . . 2-1.5h(3) Husband and wife both employed . . . . . 2-1.5h(3)(a) Local hires not eligible . . . . . . . . . . . . . . 2-1.5h(3)(b) Married persons in area with spouse . . . 2-1.5h(3)(b)(i) Minors in area with parents . . . . . . . . . . 2-1.5h(3)(b)(ii) This organizational scheme is an i m p o r t a n t plus, since p r o d u c t i o n system rule bases t e n d to exhibit this same hierarchical structure (Davis & Buchanan, 1985; Min- sky, 1985; Nilsson, 1980). In the initial translation o f the regulations, the writ- ten m a n u a l was taken to be the " e x p e r t . " T h a t is, the knowledge engineer worked directly f r o m the written regulations in formulating the rule base. T h e c o m m o n knowledge acquisition p r o b l e m o f making explicit the reasoning processes o f the h u m a n expert was not pres- ent. S o m e clarification o f the m o r e obscure statements in the regulations was provided by people who had m o r e familiarity with the regulations than the knowl- edge engineer. T h e system has c o m p l e t e d one phase o f testing a n d is currently undergoing the second phase• In the first, the person responsible for assisting employees in the m a n u a l system verified the rules. In the current testing phase, the system is in actual use, but the results are double-checked. 5.2. The Rule Base Since we are interested in obtaining an answer to the question " a m I eligible for an allowance and, if so, for how m u c h o f one?," a goal-oriented, backward-chain- ing inference engine is employed. Thus, the rule base is organized as a hierarchy o f conditions on which the top-level goal is contingent. (There are currently 144 rules in the rule base.) T h e structure o f the rule base is complicated b y the fact that procedures for calculating allowances m u st be incorporated. Since simple conditional rule bases en- code declarative knowledge, some special provision m u st be m a d e for encoding procedural knowledge (Davis & Buchanan, 1985). [The problem o f procedural a t t a c h m e n t is a relatively old o n e in knowledge-based systems (Winograd, 1985)]. A frame-based approach provides o n e type o f so- lution to the p ro b l em (Men'itt, 1989; Minsky, 1985). However, we e m p l o y e d a simpler m e t h o d that facili- tates the integration o f procedural information into the explanation (trace) o f the reasoning process. T h e pro- cedural i n f o r m a t i o n in the regulations can be stated as arithmetic expressions. These expressions undergo a bipartite interpretation by the system. O n the one hand, they are interpreted as r u l e s - - i n m u c h the same way as o rd i n ary conditional rules. In fact, they are inter- mingled with conditional rules in the rule base. T h e right-hand side o f an expression is treated as i f it stated conditions for the o b t a i n m e n t o f the left-hand side. As a result, the procedural i n f o r m a t i o n can be incorpo- rated into the trace in a fashion similar to that used for the declarative information. O n the o t h er hand, the expressions are also interpreted as true arithmetic expressions. T h e values they represent are calculated. W e refer to the two rule types as conditional and assignment rules. T h e syntax o f the conditional rules is a variation o n the standard f o r m a t (Amble, 1987; Bratko, 1986; Merritt, 1989). We define both rule types in terms o f a hierarchy o f expressions as follows: Let the vocabulary o f an arithmetic expression be 1. a finite set o f real-valued variables, V = {vl, v2, • . . , V k } , 2. the arithmetic operators '*', '/', ' + ' , ' - ' , a n d 3. the p u n c t u a t i o n symbols '(' an d ')'. We define arithmetic expressions inductively as follows: 1. Real-valued variables a n d real n u m b e r s are arith- metic expressions. 2. If A an d B are arithmetic expressions, t h e n so are (A* a), (A/B), (A + a), an d (A - a). If we let the vocabulary o f a relational expression be 1. a finite set o f real-valued variables, V = {vl, v2, . . . . Vk}, an d 2. the set o f relational operators, O = { =, <, >, <, >-_}, t h en a relational expression is an expression o f the fo rm vi o vj o r vi o #, where 1 < i, j < k, vi, vj ~ V, o E O, an d tt is a real n u m b e r . T h e vocabulary o f a sentence evaluation consists o f 1. a finite set o f sentence variables, S -- {Sl, $ 2 , . . . , Sk}, 2. the identity o p e r a t o r ' = ' , a n d 3. the set o f sentence values, E = {yes, no}. T h e resulting sentence evaluations are expressions o f the f o r m si = e, 1 < i < k, where e ~ E. 172 H. Berghel et at,. Let the vocabulary o f a logical expression be 1. the vocabularies o f relational expressions an d sen- tence evaluations, 2. the set o f logical operators, O = {and, or}, an d 3. the p u n c t u a t i o n symbols '(' and ')'. A logical expression is defined inductively as follows 1. Relational expressions and sentence evaluations are logical expressions. 2. If A and B are logical expressions, then so are (A and B) and (A or B). We can now define assignment and conditional rules. T h e vocabulary o f an assignment rule is c o m p o s e d o f 1. the vocabulary o f arithmetic expressions, 2. the assignment operator ' = =', 3. a finite set o f sentence variables, S = {sl, s2 . . . . . Sk}, and 4. the set o f sentence values, E -- {yes, no}. T h e r e are two forms o f the assignment rule. 1. I f A is a real-valued variable and B is an arithmetic expression, then A = = B is an assignment rule. 2. Any expression o f the f o r m s, = - - e, 1 __< i _< k, where e E E is an assignment rule. By letting the vocabulary o f a conditional rule be 1. the vocabularies o f logical expressions and assign- m e n t rules plus 2. the conditional o p e r a t o r " i f t h e n , " we can define a conditional rule as follows: If A is a logical expression and B is an assignment rule, then if A t h e n B is a conditional rule. T h e following rules are instances o f the rule schemata. T h e first rule (a conditional rule) is a translation o f the excerpt q u o t e d in the previous section. T h e second is a typical assignment rule. 5.3 The Inference Engine T h e i n f e r e n c e en g i n e is w r i t t e n in P ro l o g a n d is m o d e l e d aft er t h e b a c k w a r d - c h a i n i n g engines de- scribed in (Amble, 1987; Brat k o , 1986; Merritt, 1989). An i n - d e p t h d e s c r i p t i o n o f the en g i n e can be f o u n d in [ R o a c h & Berghel, in press (b)]. It is a m e t a - level engine, with respect t o t h e i n t ri n si c P r o l o g en- gine, fo r c a p t u r i n g a t race o f the r e a s o n i n g process. A c c o r d i n g t o t h e v an H a r m e l e n classification o f ex- p e r t sy st em a r c h i t e c t u r e s ( H a r m e l e n , 1989), it is a bilingual m e t a - l e v e l i n f e r e n c e system. A m o d e l - t h e o r e t i c analysis o f t h e m u l t i l ev el r e l a t i o n s h i p s be- t w e e n t h e c o m p o n e n t s o f t h e system is given in [ R o a c h & Berghel, in press (a)]. T h e s t a n d a r d " w h y " a n d " h o w " t races are c o n - s t r u c t e d as t h e c h a i n i n g p r o c e e d s t h r o u g h t h e rule base ( A m b l e , 1987; B r a t k o , 1986; M e r r i t t , 1989). S i n ce a c o m p l e t e l y c o n n e c t e d ru l e base f o r m s an a n d / o r g r a p h (Nilsson, 1980), t h ese t r a c e s are best v i e w e d as r e p r e s e n t a t i o n s o f n o d e s a n d arcs c o m - p o s i n g t r a v e r s a l s o f this g rap h . T h e " w h y " t r a c e is i m p l e m e n t e d as a list o f all t h e satisfied goals o n a p a t h f r o m t h e r o o t n o d e t o t h e c u r r e n t goal. T h e s a t i s f a c t i o n o f t h e s u c c e s s i o n o f goals o n t h e p a t h b e c o m e s t h e " r e a s o n " fo r p u r s u i n g t h e c u r r e n t goal. T h e " h o w " t r a c e is i m p l e m e n t e d as a t ree s t r u c t u r e w h i c h i n c l u d e s all p a t h s l e a d i n g t o satisfied goals. T h e s a t i s f a c t i o n o f t h e r o o t goal is " e x p l a i n e d " in t e r m s o f t h e s a t i s f a c t i o n o f all goals o n w h i ch t h e r o o t goal is c o n t i n g e n t . A conditional rule becomes part o f the solution tree when its c o n s e q u e n t unifies with the cu rren t goal and its an t eced en t condition(s) are satisfied. T h e following Prolog clause illustrates the chaining o f a conditional rule into the solution space. if (('$you$ will $sign a 12 month service agreementS = yes or ($you$ are $a Department of Defense overseas dependents school system teacher as determined under Chapter 25 of title 20 of the United States Codes = yes and $you$ will $sign a service agreement for 1 school yearS = yes)) or (Syou$ are Sseparated for reasons beyond your controls = yes and $you$ will Stransfer with approval of the concerned agencyS = yes)) then $you$meet $the service agreement requirements in 2-1.5a(i) (a-b)$ == yes. Syour subsistence/transportation allowances = = ((($your subsistence allowance per dayS + Sthe subsistence allowance for your immediate family per dayS) * $the number of days required to complete the moveS) + $your transportation allowance for relocating$). An E S for Government Relocation Allowances 173 pursue(Goal=yes, Why, t(t(t(Goal,yes),TruthValue),Reason)) :- [! recorded(cond__rule(if Conditions then Goal=yes),_) !], pursue(Conditions,[Goal=yeslWhy],Reason), truth__value(Reason,TruthValue). T h e goal truth_value extracts the truth value fro m R e a s o n , which is instantiated during the recursive at- t e m p t to satisfy the rule's condition(s). In the case o f assignment rules, the backward- chaining engine matches the c u r r e n t goal with the left o p e r a n d o f the r u l e - - t r e a t i n g it as it would the con- sequent o f a conditional rule. T h e right operands are treated as if they were conditions for the satisfaction o f the left operand. During the attempted "satisfaction" o f these operands, values are associated with each o f them. Attempts are m a d e to match each o f the right operands with o t h e r rules (conditional or assignment). Once the assignment rule has b e c o m e a part o f the p r o o f tree, the right-hand side is treated as an arithmetic expression, that is, it is evaluated. The inference engine clause simply passes the expression on to the appro- priate arithmetic routines. T h e following clause illus- trates these points. the appropriate q u e r y is constructed directly fro m the statement portion o f the query. T h e statement is simply c o n v e r t e d fro m declarative to interrogative form. With o t h er types o f relations, the corresponding q u e r y is in- cluded in the rule base. Since the Prolog co m p i l er used (Arity Corp., 1988) permits Prolog an d C to be integrated, an d since C is better suited for handling the user interface, the pro- cedural language C is used for m o st o f the interface routines. C is used for its speed a n d because sophisti- cated screen interface libraries are readily a v a i l a b l e - - resulting in a substantial reduction in d e v e l o p m e n t t i m e and cost a n d a substantial increase in the quality o f the user interface. F o r a detailed discussion o f the relationship between Prolog an d C in the system see (Roach & Berghel, 1990). pursue (Goal, Why, t(t(t(Goal, Value), true), Reason)) :- recorded(assign_rule(Goal = = Expression),_), pursue(Expression,[GoallWhy],R), evaluate(Expression,Value), Reason = ((t(t(t(Goal,Expression),true),'inrulebase') andR), recordz(wasderived(Goal,Value),_). T h e arithmetic expression instantiated to E x p r e s s i o n is treated as a set o f conditions in the recursive call to p u r s u e . T h e 'how' trace for those 'conditions' is re- t u r n e d in R. It is treated as an arithmetic expression by e v a l u a t e . T h e value that the expression evaluates to is r e t u r n e d in V a l u e . T h e explanation for G o al evaluating to V a l u e is that a rule o f the form G o al = E x p r e s s i o n exists in the rule base and R. T h e con- tent o f R is not explicit in this clause, but it consists o f explanations o f the values associated with each o f the operands o f the arithmetic expression E x p r e s s i o n . 5.4. The User Interface T h e user interface provides access to a n u m b e r o f useful features a n d parcels o f information. T h r e e forms o f explanation are available: a " w h y " trace, a " h o w " trace, a n d a s u m m a r y o f allowance a m o u n t s . In addition, the user can access i n f o r m a t i o n provided in previous con- sultations a n d m o d i f y it. T h e user interface employs a r u d i m e n t a r y level o f natural language analysis in the f o r m a t i o n o f some queries. W h e n the user m u s t decide the truth value o f a relation with statements a n d responses as operands, 6. E X T E N S I B I L I T Y A couple o f extensions to R A P have been proposed. First, i f the n u m b e r o f employees using the system be- comes substantial, a n d there is a need for long-term storage o f the results o f their consultations, then a m o r e sophisticated file m a n a g e m e n t scheme should be em- ployed. T h e cu rren t system uses the operating system's file m a n a g e m e n t facilities, so it does n o t m ai n t ain a single a u t o n o m o u s database. Second, some type o f forms generator would be a natural a d d e n d u m to the system. T h e r e are n u m e r o u s forms that m u st be filled o u t by the e m p l o y e e during the d e t e r m i n a t i o n process. S o m e o f the paperwork required o f the employee could be eliminated b y letting the system generate the com- pleted f o r m s - - u s i n g the i n f o r m a t i o n available at the e n d o f a consultation. 7. C O N C L U S I O N W e have discussed the a u t o m a t i o n o f the process o f d e t e r m i n i n g relocation allowances. W e have seen that an expert system ap p ro ach is a viable one. T h e rep- resentation o f the federal regulations in an expert sys- 174 H. Berghel et al. tern rule base has been described. The procedural in- formation relating to the determination o f actual al- lowance a m o u n t s has been e n c o d e d in a manner that maximizes the scope o f the explanations supplied to the user. We have discussed the system's features for e x a m i n i n g the information available at the end o f a consultation. Finally, facilities have been discussed which make it possible to store, modify, retrieve, and resume previous consultations. R E F E R E N C E S Amble, T. (1987). Logic programming and knowledge engineering. Wokingham, England: Addison-Wesley. Arity Corporation. (1988). The Arity/Prolog language reference manual. Concord, MA. Bratko, I. (1986). Prolog programming for artificial intelligence. Wokingham, England: Addison-Wesley. Davis, R., & Buchanan, R. ( 1985 ). Production rules as a representation for a knowledge-based consultation program. In R. Brachman & H. Levesque (eds.), Readings in knowledge representation (pp. 371-388). San Mateo, CA: Morgan Kaufmann. Merritt, D. (1989). Building expert systems in Prolog. New York: Springer-Verlag. Minsky, M. (1985). A framework for representing knowledge. In R. Brachman and H. Levesque (eds.), Readings in knowledge rep- resentation (pp. 245-262). San Mateo, CA: Morgan Kaufmann. Nilsson, N. (1980). Principles of artificial intelligence. Palo Alto, CA: Tioga. Roach, D., & Berghel, H. [in press (a)]. Expert systems as overlapping logical theories. In H. Berghel, J. Talburt, and D. Roach (eds.), Proceedings of the 1990 ACM/IEEE Symposium on Applied Computing. Washington, DC: 1EEE Computer Society Press. Roach, D., & Berghel, H. (1990). Issues in Interfacing High-Level and Logic Programming Languages (Tech. Rep. No. TR-90-04). Fayetteville: University of Arkansas, Center for Artificial Intel- ligence and Expert Systems. Roach, D., & Berghel, H. [in press (b)]. The physiology of a Prolog expert system inference engine. In H. Berghel, E. Unger, and R. Rankin (eds.), Proceedings of the 1990 A CM Symposium on Small Systems. New York: ACM Press. Roach, D., Berghel, H., Meehan, J., & Green, G. (1989). RAP: An expert system for relocation allowance planning. In S. Dhall, H. Berghel, and K. George (eds.), Proceedings t f the 1989 ACM South Central Regional Conference (pp. 106-111). New York: ACM Press. van Harmelen, F. (1989). A classification of meta-level architectures. In P. Jackson, H. Reichgelt, and F. Harmelen (eds.), Logic-based knowledge representation (pp. 13-35). Cambridge, MA: MIT Press. Winograd, T. (1985). Frame representations and the declarative/pro- cedural controversy. In R. Brachman and H. Levesque (eds.), Readings in knowledge representation (pp. 357-370). San Mateo, CA: Morgan Kaufmann. United States General Services Administration. (1984). Federal Travel Regulations, HHS Appendix B. Washington, DC: Government Printing Otfice. United States General Services Administration, GSA Training Center. (1987). Relocation Allowances, Federal Travel Regulations and Joint Travel Regulations--Vol. H, Lecture Outline. Washington, DC: Government Printing O[fice. 
work_bt7pyfijobbtbm4uwb77i2hgfm	----	NASA Technical Reports Server (NTRS) 
work_bvshrmnbrvdyjmr3keeijjnsce	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220988282 Params is empty 220988282 exception Params is empty 2021/04/06-03:05:31 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220988282 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:31 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_bxbmpzxsijfjpapcjt5kdw6u3q	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220993610 Params is empty 220993610 exception Params is empty 2021/04/06-03:05:37 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220993610 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:37 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_bxbstu7d2zbghpbu2ajkkwvcmy	----	Hybrid Stage Shop Scheduling Accepted Manuscript Hybrid Stage Shop Scheduling Andrea Rossi, Sauro Soldani, Michele Lanzetta PII: S0957-4174(14)00829-X DOI: http://dx.doi.org/10.1016/j.eswa.2014.12.050 Reference: ESWA 9775 To appear in: Expert Systems with Applications Please cite this article as: Rossi, A., Soldani, S., Lanzetta, M., Hybrid Stage Shop Scheduling, Expert Systems with Applications (2015), doi: http://dx.doi.org/10.1016/j.eswa.2014.12.050 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. http://dx.doi.org/10.1016/j.eswa.2014.12.050 http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2014.12.050 Revised manuscript 1 Hybrid Stage Shop Scheduling Andrea Rossi(1), Sauro Soldani(2), Michele Lanzetta(3) Department of Civil and Industrial Engineering, University of Pisa, Italy (1) arossi@ing.unipi.it (2) sauroviola@gmail.com (3) lanzetta@unipi.it Corresponding author Department of Civil and Industrial Engineering Largo Lazzarino 56122 Pisa, Italy Tel. +39-050-2218122 Cel. +39-320-4212172 Abstract The proposed Hybrid Stage Shop Scheduling (HSSS) model, inspired from a real case in the high- fashion industry, aims to fully exploit the potential of parallel resources, splitting and overlapping concurrent operations among teams of multifunctional machines and operators on the same job. The HSSS formally extends mixed shop scheduling (a combination of flowshop and open shop), which is able to model routing flexibility, and hybrid shop scheduling, which provides resource flexibility. To also include operational flexibility through alternative plans obtained by reordering operations linked by undefined or arbitrary (immaterial) precedence constraints, the proposed model integrates process planning and group shop scheduling. A mixed integer linear programming model and a theory based on disjunctive graphs have been proposed to explore the composite relations between nodes involving immaterial relations and deploying their routing rules. A constructive O(resources×jobs2) algorithm to generate a feasible plan/schedule in the most general case has been developed and applied to a case study. Keywords: manufacturing systems, planning/scheduling integration, mixed shop scheduling, nonlinear routing, resource flexibility Revised manuscript 2 1 Introduction We consider a rather general model of mixed shop in which a set of operations for a given set of jobs has to be scheduled on a set of machines, which includes two extensions to the standard scheduling problem as defined by Dauzère-Pérès, Roux, and Lasserre (1998): 1. an operation can be processed by one resource chosen in a given set (resource flexibility); for the sake of generality, we use the standard term resource from the scheduling theory instead of machine; 2. the routing of products in the shop floor is not necessarily linear, i.e. an operation can have more than one predecessor and more than one successor on the routing (nonlinear routing). The mixed shop type considered here is a hybrid (or flexible) shop type combination of flowshop and open shop. A hybrid flowshop is a flowline with parallel resources. In a flowshop, the sequence of operations of each job (routing) is linear and predefined; in an open shop the sequence of operations is immaterial (or undefined). In a mixed shop, the set of constraints between operations is partitioned into two sets: flowshop type set and open shop type set (Masuda, Ishii, & Nishida, 1985). 1.1. Integration of process planning and shop scheduling Mixed shop is the paradigm for the integration of process planning and shop scheduling (Tan & Khoshnevis, 2000). Process planning has been defined by the Society of Manufacturing Engineers as the systematic determination of the methods by which a product is to be manufactured economically and competitively. Traditionally, process planning and shop scheduling are applied separately and sequentially. If a single output of process planning (the plan) is considered as the input to flowshop scheduling, routing constraints from planning may create bottleneck situations on some resources while other can be starving. Also the line balancing may be affected. Consequently, the global system performance can be improved by integrating planning and scheduling. However, the integration of process planning and shop scheduling does not consider operations belonging to the set of open shop type but rather the assignment of optimal process plans among a number of (predefined) alternatives. Stecke and Raman (1995) described a scheme for classifying different types of flexibility conventionally associated with the ability to manufacture a variety of part types by flexible manufacturing systems. In this classification operation flexibility assumes that more alternative plans can be generated by the process planner for a given job. Kis (2003) and Leung (2010) modeled an integrated process planning and shop scheduling system by disjunctive AND/OR graphs. The branches of an OR-subgraph constitute a set of alternative subroutes: exactly one of them must be chosen during scheduling. AND/OR graphs are a generalization of the resource Revised manuscript 3 alternatives of individual operations; however, immaterial constraints among operations cannot be considered effectively. Go, Wahab, Rahman, Ramli, and Hussain (2012) and Bentaha, Battaïa, and Dolgui (2014) approached the design of disassembly lines for end-of-life products with the objective to maximize the line profit. An AND/OR graph was used to model the precedence relationships among tasks and subassemblies and the disassembly alternatives. Doh, Yu, Kim, Lee, and Nam (2013) considered alternative machines for each operation (resource flexibility), in addition to specifying multiple process plans alternative operations and their sequence by a network model with AND/OR nodes. Otto and Otto (2014) described a precedence graph approach that is based on learning from past feasible production sequences and forms a sufficient precedence graph that guarantees feasible assembly line balancing in the automotive industry. The assignment of tasks to stations is due to restrictions, which can be expressed in a precedence graph that includes direct and indirect conjunctive precedence relations. Phanden, Jain, and Verma (2013) developed a simulation-based genetic algorithm (GA) to integrate the process planning and scheduling function that can be quickly implemented in a company with existing process planning and scheduling departments. Bensmaine, Dahane, and Benyoucef (2013) proposed a new heuristics to integrate the process planning and scheduling problem for reconfigurable machine tools, each with multiple configurations, and can perform different operations with different capacities. They considered only direct precedence graph relations. 1.2. Mixed and group shop scheduling In order to reduce the gap with real manufacturing systems, the mixed shop scheduling problem has been regarded as a mixture of flow (or job) and open shop scheduling problems, where operations with immaterial precedence constraints are grouped in the route of the related job. In 1997, the group shop scheduling problem was first introduced in the context of a mathematical competition (Whizzkids, 1997). Regarding the group shop scheduling problem, Blum and Sampels (2004) used a disjunctive graph representation for group shop scheduling and applied an ant colony algorithm to tackle the problem complexity. Liu, Ong, and Ng (2005) proposed a tabu search for group shop scheduling and evaluated the algorithm performance on a set of benchmark problems. Ahmadizar and Shahmaleki (2014) considered the stochastic group shop scheduling problem where both release dates and processing times are random variables, normally, exponentially or uniformly distributed. From the literature above, it can be observed that the mixed shop model includes the models on integration of process planning and scheduling and those on group shop scheduling, by allowing alternative plans produced simply reordering operations connected by immaterial constraints (Figure 1). Revised manuscript 4 According to Stecke and Raman (1995), in addition to operation flexibility, routing flexibility is another aspect of the scheduling flexibility related to the ability of generating alternative paths, which can be followed through the system for a given process plan. As discussed by Rossi and Lanzetta (2013), shared buffers between stages allow routing flexibility, by the permutation of job sequences on resources. Figure 2 shows as an (exclusive) OR node (node 0 towards O31 and O32) that can be reworded as a no-exclusive OR by an immaterial relation, which allows more alternative routings for the scheduler module. Revised manuscript 5 FLEXIBILITY PRECEDENCE GRAPH ALTERNATIVE ROUTINGS conjunctive relations deployment O pe ra ti on f le xi bi li ty (a lt er na ti ve p ro ce ss pl an s) R ou ti ng f le xi bi li ty M ix ed s ho p 1 2 4 5 * 3 1 2 4 5 * 3 2 4 5 3 1 1 1 2 4 5 * 3 1 2 * 3 3 2 4 5 2 3 3 2 4 5 1 1 1 1 2 4 3 5 * 1 2 * 3 3 2 4 5 1 2 3 0 2 3 3 2 4 5 1 1 1 1 2 3 Revised manuscript 6 Figure 1. Example of mixed shop precedence graphs achieved by operation and routing flexibility (according to Stecke & Raman, 1995). No-exclusive OR nodes described by immaterial relations give alternative routing for the scheduler module in order to minimize the completion time. Figure 1 about here Revised manuscript 7 As shown by Masuda, Ishii, and Nishida (1985), the mixed shop problem is NP-hard. Relatively few papers were proposed on the subject. Shakhlevich et al. (2000) discussed the complexity of mixed shop problems under various criteria and clarified the boundary between polynomially solvable and NP-hard problems. Blazewicz and Kobler (2002) reviewed the properties of simple precedence graphs for scheduling problems and exhibited several new polynomial cases for various problems on unrelated parallel machines under arbitrary resource constraints. Ferrell, Sale, Sams, and Yellamarju (2000) approached the problem by heuristics and evaluated the performance in a promising set of dispatching rules. Nasiri and Kianfar (2011) proposed a stage shop scheduling, an open job shop scheduling problem where all the operations of the same job with immaterial precedence constraints are grouped in a stage of the job shop-type. Aloulou and Artigues (2010) and Amin-Naseri and Ashfadi (2012) modeled simple or primitive non-linear precedence constraints, in order to split macro operations into micro operations at stage level. Revised manuscript 8 Table 1. Main contributions from the literature to the mixed shop scheduling problem as an integration of alternative nonlinear routing approaches (group shop, AND/OR) Authors Graph type MILP Theory Disjunctive arc Node Flexible or hybrid Nonlinear routing Group shop Mixed shop AND/OR Masuda, Ishii, and Nishida (1985) • • Dauzère-Pérès, Roux, and Lasserre (1998) • • • Kis (2003 ) • Blum and Sampels (2004) • • Liu, Ong, and Ng (2005) • Leung (2010) • Aloulou and Artigues (2010) • • Nasiri and Kianfar (2011) • • Amin-Naseri and Ashfadi (2012) • • • Doh, Yu, Kim, Lee, and Nam (2013) • • • Bensmaine, Dahane, and Benyoucef (2013) • • Ahmadizar and Shahmaleki (2014) • • Otto and Otto (2014) • • Current work • • • • From this literature analysis chronologically synthesized in Table 1, it can be observed that the most recent works (towards the bottom) converge to the generalization of integrated process planning and group shop scheduling. For this purpose, further research for the definition and use of new precedence relations is required. Revised manuscript 9 1.3. From mixed to hybrid shop scheduling Resource flexibility and nonlinear routing allow a higher level of manufacturing flexibility within each stage of the flowshop: multiple resources can process in parallel subparts of the same product (multi-resource systems). Resource flexibility in the standard flowshop model was reviewed by Ribas, Leisten, and Framinan (2010) and Ruiz and Maroto (2005). In recent years, approaches to hybrid flowshops were proposed by Behnamian and Zandieh (2011), considering the minimization of earliness and quadratic tardiness penalties by a discrete colonial competitive algorithm. Rossi, Pandolfi, and Lanzetta (2013) considered a non-permutation hybrid flow shop scheduling problem, with parallel batching machines, parallel batch families and a machine loading system modeled as sequence-independent uniform set-up times, with the purpose of reducing the number of tardy jobs and the makespan. They proposed two dynamic heuristics based on the critical ratio of allowance set-up and processing time in the scheduling horizon. Li, Meng, Liang, and Zhao (2014) proposed a heuristic-search genetic algorithm, which results in lower complexity and higher efficiency for multi-stage hybrid flow shop scheduling with batch processing machines. Amin-Naseri and Ashfadi (2012) considered some primitive non-linear precedence relations in a flexible job shop environment and proposed a local search problem-specific to speed up e genetic algorithm. Rossi (2014) proposed an ant colony optimization approach based on a disjunctive graph model in order to schedule a manufacturing system with resource flexibility, separable sequence dependent/independent setup and transportation times. Benavides, Ritt, and Miralles (2014) proposed an extension to flowshop with heterogeneous resources among stages, to optimize worker assignment to workstations of health care departments, with the objective of minimizing the makespan, while respecting the diverse capabilities and paces of heterogeneous workers. To our knowledge, only Amin-Naseri and Ashfadi (2012) considered some aspects of nonlinear routing with resource flexibility using a genetic algorithm, without exploring the theoretical implications. 1.4. Problem motivation and approach The research was motivated by a real manufacturing case, a high-fashion company based in Florence, Italy, for the production of high-fashion goods by skilled artisans. An operation (e.g. the manual assembly of a woman bag) can be split in a number of micro- operations (e.g. stitching a handle and sewing a slide fastener), which can be performed on the same job in a fixed or arbitrary routing by different resources of the belonging flowshop stage (e.g. by two operators of the assembly shop). Splitting operations of multi-resource systems affects process planning providing further flexibility and increases the number of scheduling alternatives for the Revised manuscript 10 shop system and, potentially, shorter schedules can be obtained. Alternative plans produced simply reordering operations reduce the process planner workload. Generating alternative plans improves machine utilization and line balancing. Figure 2 shows an example problem derived from the examined case represented as a disjunctive graph. The three shaded areas are examples of macro operations that have been split by detailing the sub operations and resources in order to process them in parallel (overlapping) to reduce the makespan. Precedence graphs allow different routings (sequence permutations) by the inclusion of the immaterial precedence arcs in addition to directed (or conjunctive) arcs. As shown in the example, mixed routing can also span across consecutive stages. The most general case, which represents the focus of current paper, is the shaded graph relation O122-O133-O252 in the example. The cited operations represent respectively: gluing the two halves of the bag, stitching a pocket on just one of the two sides, stitching a buckle to both halves, with realistic times indicated. Gluing is on the previous stage because of a precedence constraint. Sequence O122-O133-O252 is potentially shorter versus sequence O122-O252-O133 for the possible parallelization of O122 and O252. Revised manuscript 11 Figure 2. A disjunctive graph modeling a mixed shop for hybrid flowshop scheduling, with processing times t i j k at nodes O i j k for N (=3) jobs on S (=3) stages, with respectively Q1=2 (hybrid), Q2=1, Q3=1 resources. Colored disjunctive arcs represent the candidate resource for the connected nodes (operations). Bottom: the top digraph partially transformed in acyclic digraph by directing disjunctive arcs on two paths starting from node 0 and ending at node * (green and brown) and one ending in O222 (yellow). Figure 2 about here Stage 2 Stage 3 t253 0 O111 O122 resource assignment: resource 1, stage 1 job routing O222 O211 t111 t122 t211 t222 * O34 O143 O252 t321 t332 t343 t143 O133 t133 O242 t242 O232 t232 t311 immaterial routing resource assignment: resource 2, stage 1 resource assignment: resource 1, stage 2 resource assignment: resource 1, stage 3mixed routing O32 O33 O31 24 Stage 2 Stage 3Stage 1 0 O111 O222 O211 19 * O33 O143 O252 O32 O33 O133 O242 O232 O31 20 11 12 3 7 4 26 41 28 37 O122 36 Stage 1 Revised manuscript 12 We propose the hybrid stage shop scheduling (HSSS), a mixed shop model for hybrid flowshop scheduling with unrelated parallel resources per stage. The proposed underlying mixed model, in addition to routing flexibility that emerges with permutations of the operations linked by immaterial relations, allows: (i) splitting operations of the flowshop type set on the same job and assigning them to alternative resources; (ii) overlapping the operations of consecutive stages linked by immaterial routing, yielding shorter schedules. The proposed hybrid stage shop scheduling model completes the mixed stage shop model by Nasiri and Kianfar (2011) and the hybrid stage shop scheduling model by Amin-Naseri and Ashfadi (2012). Extensions include the addition of operation overlapping and splitting within stages and across consecutive stages with the composition of mixed relations, which will be discussed throughout the paper. The proposed model can be applied to Computer Aided Design (CAD), where designers share the same digital model of the product under development (Leu et al., 2013) , in robotic assembly and in disassembly of end of life goods, where operations on subparts can be carried out concurrently, and in train traffic management by railway line branching (Kozan & Liu, 2012). In section 2 we define the hybrid stage shop scheduling problem and its MILP (section 3). In section 4 we describe the model based on digraphs and in section 5 we propose a set of primitive and mixed rules and their combinations, derived from the precedence graph, in order to deploy and explore various alternative routings. The hybrid stage shop scheduling model is based on a digraph, which is a powerful tool to design scheduling optimization (metaheuristics) algorithms. In section 6 a heuristics is presented to obtain a conjunctive graph, which represents a feasible solution of the hybrid stage shop scheduling problem. Also, the computational complexity is evaluated. Application to a case study is discussed in section 7. 2 Hybrid Stage Shop Scheduling (HSSS) In hybrid stage shop scheduling, N jobs have to be scheduled on S stages with R resources in all, in accordance with its nonlinear routing, represented by a precedence graph, which includes conjunctive (directed) and disjunctive precedence constraints (described below) among operations belonging to a set of il operations Oi j k with i=1,..,N, j=1,.., il , k=1,..,S. Each operation Oi j k has to be processed according to its precedence constraint, denoted by a level Li j, for time ti j k with resource flexibility represented by the assignment to a single resource h among a set of alternative identical resources {Qk-1+1,.., Qk} of the stage k (k=1,..,S). Each job i is subject to a release date, ri, an Revised manuscript 13 initial level ijlj Li,..,1min = =Li 1=1. st(Oi j k) and ct(Oi j k) denote respectively the starting and the completion time of an operation, with iCt being the total completion time for each job. No resource can process more than one operation at a time and an operation must be processed by one, and only one, resource; no precedence constraints between operations of different jobs is allowed. All resources are available since the initial time and resource breakdown is not considered. Setup is negligible or is included in the processing time. The handling system offers the flexibility to move a part with negligible transportation time and includes a buffer for each stage. The buffer size of stages k=1 and k=S+1 is N (i.e. the input and output buffers contain up to N jobs). The number of slots of the buffer at stage k, k=2,…,S, each containing a single job, is defined by the following Lemma 1. Lemma 1. The hybrid flowshop scheduling with N jobs and R resources is Bi=+∝ if and only if the shared buffer size for stage k , k=2,…,S is at least: )( 2−−− kk QQN (1) This proof extends the result achieved by Rossi and Lanzetta (2013). In the worst case, the system becomes congested in a single stage. Let k be this stage, with )( 1−− kk QQ the number of its resources. Hence, the number of jobs waiting in the buffer of stage k or in the previous stages is )( 1−−− kk QQN . Similarly, in the previous stage there are )( 21 −− − kk QQ resources. The jobs can wait after processing in the previous stage k-1 before a resource of the stage k will be free. The minimum buffer size for each of the intermediate stages is given by: )()()( 2211 −−−− −−=−−−− kkkkkk QQNQQQQN (2) � Operations of job i are partitioned in k subsets Gi k to be performed by resources of stage k according to the following equations: NilG i S k ki ,..,1 1 ==∑ = (3) zwSzwGG ziwi ≠=∅= ,,..,1,,∩ (4) Revised manuscript 14 The quantity P = mink=1,..,S )( 1−− kk QQ represents the degree of resource flexibility, i.e. the parallelization capability of the system. We define a hybrid stage shop scheduling model with degree of resource flexibility P as P-HSSS model. The assignment of operations to a given resource of current stage is a source of flexibility, in addition to the flexibility represented by sequencing operations on resources. This double source of flexibility increases the problem complexity along with the performance increase. 3 Hybrid stage shop scheduling mixed integer model A mixed integer linear programming (MILP) model for the hybrid stage shop scheduling is described. Decision variable ijkhx 1, if operation j of job i is assigned to resource h in stage k 0, otherwise The objective function selected is the total completion time (makespan) minimization and can be formulated as: maxCMin (5) subject to constraints I.1 to I.15. ∑ ∑ ∑ = += =− ≥ S k Q Qh l j ijhkijki k k i xtCt 1 1 11 Ni ,...,1=∀ (I.1) ∑ += − = k k Q Qh ijhkx 11 1 Ni ,...,1=∀ , ilj ,...,1=∀ , Sk ,...,1=∀ (I.2) ∑ = = S k ijhkx 1 1 , , kk QQh ,...,11 +=∀ − (I.3) ∑ = ≤ il j ijhkx 1 1 , , (I.4) Ni ,...,1=∀ ilj ,...,1=∀ Ni ,...,1=∀ kk QQh ,...,11 +=∀ − Sk ,...,1=∀ Revised manuscript 15 ∑ = ≤ N i ijhkx 1 1 , , (I.5) ⎪⎩ ⎪ ⎨ ⎧ = =ℜ∈ = + 0 1 0 ijhk ijhk ijk x x if if t , , , (I.6) ijhkijkijkijk xtstct =− , , , (I.7) ⎪⎩ ⎪ ⎨ ⎧ =⇒=ℜ∈ ⇒<>− ' '0 ' '' jjLLse jjLLsectst ijij ijijijkkij ≺ ≺ , , (I.8) 0≥ir Ni ∈∀ (I.9) 0≥ijkst , , (I.10) 0≥ijkt , , (I.11) 0≥ijkct , , (I.12) 0≥iCt (I.13) ℵ∈ijL , (I.14) }{ 1,0∈ijhkx , , , (I.15) The constraint equation (I.1) forces the total completion time of job i, to be less or equal to the sum of all processing times affecting job i, indexed by operations and stages. The constraint equation (I.2) ensures that each operation is assigned to exactly one resource. The constraint equation (I.3) ensures that each operation is specific to exactly one stage. The constraint equation (I.4) shows that any operation can (sum equal to 1) or cannot (sum equal to 0) be assigned to a particular pair of job and resource. ilj ,...,1=∀ kk QQh ,...,11 +=∀ − Sk ,...,1=∀ Ni ,...,1=∀ ilj ,...,1=∀ kk QQh ,...,11 +=∀ − Sk ,...,1=∀ Ni ,...,1=∀ ilj ,...,1=∀ kk QQh ,...,11 +=∀ − Sk ,...,1=∀ Ni ,...,1=∀ ilj ,...,1=∀ Sk ,...,1=∀ Ni ,...,1=∀ ilj ,...,1=∀ Sk ,...,1=∀ Ni ,...,1=∀ ilj ,...,1=∀ Sk ,...,1=∀ Ni ,...,1=∀ ilj ,...,1=∀ Sk ,...,1=∀ Ni ,...,1=∀ Ni ,...,1=∀ ilj ,...,1=∀ Ni ,...,1=∀ ilj ,...,1=∀ kk QQh ,...,11 +=∀ − Sk ,...,1=∀ Revised manuscript 16 The constraint equation (I.5) represents the possibility (sum equal to 1) or not (sum equal to 0) that a job involves the use of a resource for a specific operation. The constraint equation (I.6) requires that the processing time of resource h of stage k performing operation j of job i is equal to 0 if the transaction is not assigned to any resource. The constraint equation (I.7) requires that the difference between the completion time and the starting time of a resource busy for an operation of a job be equal to its processing time. The constraint equation (I.8) ensures that the difference between the starting time of a certain operation and the completion time calculated for the immediately preceding operation is greater than 0 if the value of the variable level of the previous operation is smaller compared to the variable level of the following operation. It is worth dwelling on the condition in which it has the same value of the variable level for two successive operations; in this case the order of execution can be chosen arbitrarily: the result can be greater, less or equal to 0. Finally, the constraint equations ranging from (I.9) to (I.15) are simply conditions of existence and non-negativity of the variables used in the model. 4 Hybrid stage shop scheduling disjunctive graph representation The hybrid stage shop scheduling problem can be represented by a weighted disjunctive graph as in Rossi and Lanzetta (2014), applied to the non permutation flowshop scheduling (Figure 2 top): DG = (ℵ, Wℵ, C, D, Eh k) (6) where ℵ is the set of nodes (operations ijkO ) plus the dummy start and finishing nodes 0 and *; WN is the set of weight on nodes, represented by the processing times ijkt ; C is the set of conjunctive (directed) arcs (a,b) between every pair of nodes a and b on a job routing, representing the precedence constraints between the corresponding operations (conjunctive relations, C); it also includes conjunctive arcs between 0 (*) and every first (last) node on a routing; D is the set of disjunctive arcs [a,b] (equal to [b,a]), between every pair of nodes a and b on a job routing with immaterial precedence constraints between the corresponding operations (disjunctive relations, D); Eh k is the h th set (h∈{Qk-1+1,..., Qk}) of disjunctive arcs [u,v] between every pair of nodes u and v belonging to the same stage k; u,v∈Gi k, k=1,..,S; Eh k also includes disjunctive arcs between 0 and u Revised manuscript 17 (v) and between u (v) and * for all u∈Gi k. The sets Eh k are represented by colored disjunctive arcs in Figure 2 top. It can be noticed that the DG includes one kind of conjunctive arc (arrows in the digraph of Figure 2 top) and two kinds of disjunctive arcs: D for nonlinear routing (dashed lines) and Eh k for resource flexibility (colored lines). In general, a P-HSSS problem, i.e. a problem with degree of resource flexibility P, is represented by S stages of P kinds of arcs E*k. The number of disjunctive arcs in the digraph is O(P⋅S⋅N2), because the arcs of E*k connected with at least N nodes are O(N 2) and the total flexibility is obtained by P⋅N2 arcs of E*k (Kacem, Hammadi, & Borne, 2002). Consequently, in the P-HSSS total flexibility is present, but only within each stage. The set D describes nonlinear routing for open and consequently for the mixed shop considered here. D is not required to model linear problems, such as flowshop and hybrid flowshop scheduling. In the next section the interaction between C and D type relations will be discussed. A finite sequence of conjunctive arcs between two nodes is called path. The length of an arc is equal to the processing time of the node at which it ends. The path length is equal to the sum of the lengths of its arcs. A path which starts from 0 and ends at * is the loading sequence on a resource. Figure 2 bottom shows two paths with lengths, respectively, 24+20 (green path) and 19+11 (brown path), which start from 0 and end at *; they are the loading sequences of resources of stage 1. A cycle is a path that starts and ends at the same node. If no cycle is present in a conjunctive graph achieved by directing some disjunctive arcs, the conjunctive graph is acyclic (Figure 2 bottom). In particular, a digraph is an acyclic conjunctive graph. If an acyclic conjunctive graph includes all the nodes, the related loading sequences on the resources are feasible schedules and the makespan is the length of the critical path, i.e. the longest path between the dummy start and finishing nodes. Finally, for each job, the length of the longest path between the dummy start node 0 and a given node is the total completion time ct(Oi j k) of the corresponding operation. 5 Hybrid stage shop scheduling routing rules Hybrid stage shop scheduling is a more general model for mixed shop scheduling and it is more suitable to describe real manufacturing cases, but it requires the definition of new relations among nodes split at the stage level or across consecutive stages. A digraph as defined in (6) includes C and D type primitive relations. By the presence of D type relations, routing is nonlinear. C and D type relations in (6) between nodes involve precedence constraints, which can be described by the Revised manuscript 18 routing rules (feasible sequence of operations for a given job) introduced in this section. The proposed wildcard node mechanism will also be described. The complete set of basic relations includes two primitive C and D type relations defined in (6) listed in Table 2 and five mixed relations introduced in Table 3. The complete set of basic relations between nodes j and j + 1, j=1,…, 1−il , of the same job i and their levels Li j is: (a). Conjunctive relation (C), Li (j+1) = Li j + 1. The routing is linear as in flowshop scheduling (Table 2). (b). Disjunctive relation (D), Li (j+1)=Li j. The routing is immaterial as in open shop scheduling, where a sequence of operations is a job permutation (Table 2). (c). Mixed relation (M). Mixed relations are present, where a disjunctive relation can anticipate, interpose or postpone a direct relation (Table 3). Levels are the mechanism to represent the precedence constraints between operations. Table 2. Primitive relations (2 nodes) and deployment of their routing rules Case Relations Notation Precedence graph Routing deployment #1. Conjunctive (directed) C (1,2) #2. Disjunctive D [1,2] (or [2,1]) Table 2 shows linear and arbitrary relations among operations of the same job and their routing rules. A C relation is a conjunctive arc between a starting and ending node and is represented in round brackets (1,2); the starting node represents the operation processed before in the feasible sequence 1�2, as imposed by the level Li 2 = Li 1+1. A D relation is a disjunctive arc between two nodes; the nodes represent operations processed in an arbitrary order within the feasible sequence; the two feasible sequences are 1�2 and 2�1 associated with a unique level, where Li 2=Li 1. A D relation has no orientation between nodes and is differentiated from a C relation by using square brackets: [1,2] = [2,1]. Both conjunctive and disjunctive (primitive) relations are denoted by { }DC,=Γ . 2 1 Li 1 Li 1+1 1 2 Li 1 Li 1 2 1 1 2 2 1 Revised manuscript 19 Table 3. Mixed relations (3 nodes) and deployment of their routing rules Case Notation Precedence graph Routing deployment #3. M1 #4. M2 #5. M3 #6. M4 Table 3 shows the mixed relations (Mx, x=1,..,4) among nodes. A M relation, yx ∪ , is a union of primitive relations x,y ∈ Γ = {C, D}. M1 and M2 include one disjunctive relation, [2,3], and two conjunctive relations, yielding two feasible routing rules. 1 2 3 Li 1 Li 1+1 Li 1+1 1 2 3 1 3 2 1 2 3 Li 2+1 Li 3=Li 2 Li 2 2 3 1 3 2 1 1 2 3 Li 2 {Li 1,Li 2} Li 1 1 2 3 1 3 2 3 1 2 1 2 3 Li 1 Li 1 Li 1 1 2 3 1 3 2 2 1 3 2 3 1 3 1 2 3 2 1 Revised manuscript 20 M3 includes two D relations and one C relation, (1,2). The directed arc compels a clockwise path which encompasses a wildcard node. Node 3, belonging to both D relations, [2,3] and [1,3], can be interposed in the sequence compelled by the C relation to obtain feasible routing rules: 1�2�3, 1�3�2 and 3�1�2. M4 includes all three disjunctive arcs (similarly to group shop scheduling); the feasible routing rules are all the six permutations of the three nodes. An instance of the problem can be expressed in compact form by the tuple P-HSSS|ℵα|Cβ|Dχ|Rδ|Nε|Sφ|max Li j γ (7) where quoted symbols α, β, χ, δ, ε, φ and γ are values of the respective parameter (base). An example is the triangular inset of Figure 3 derived from the 2-HSSS|ℵα|C17|D5|R4|N1|S3|max Li j8 case in Figure 2. Figure 3. Examples of precedence relations on 2-HSSS|ℵα|C17|D5|R4|N1|S3|max Li j8: conjunctive relations C ((O111, O122), (O122, O133), etc.) and disjunctive relations D ([O222, O232], [O222, O242], Stage 2 Stage 3 t253 Stage 1 0 O111 O122 resource assignment: resource 1, stage 1job routing: conjunctive relation C O211 t111 t122 t211 t222 * O34 O143 O252 t321 t332 t343 t143 O133 t133 O242 t242 O232 t232 t311 job routing: disjunctive relation D resource assignment: resource 2, stage 1 resource assignment: resource 1, stage 2 resource assignment: resource 1, stage 3 job routing: constrained-disjunctive relation M2 (Nasiri & Kianfar, 2011) O32 O33 O31 job routing: mixed relation M3 within stage 2 (Amin-Naseri & Ashfadi, 2012) job routing: mixed relation M3 across stages 2-3 (proposed) O222 Revised manuscript 21 [O311, O321]) from Table 2; constrained-disjunctive relations M2 and M3 from Table 3 (respectively cases #4. and #5.) in triangular insets. Figure 3 about here 5.1. Properties of basic relations The combination of primitive relations has been shown to yield the mixed relations M1 to M4. Their precedence graph encompasses nonlinear routing; consequently, a digraph (6) can be formed by mixed relations. In this subsection, some properties of primitive and mixed relations are introduced to allow more complex composition and decomposition rules of digraphs (6). Proposition 1. All the introduced relations can be represented by a digraph (6) with maximum level equal to 2. As an example, we consider M3. Its digraph is: DG(M3) = (ℵ, Wℵ, C, D, Eh k) = { } { } { } { } { }( ),]3,2[],3,1[,)2,1(,,,,3,2,1 321 iii ttt where, without loss in generality, the set of the disjunctive arcs of each resource Eh k is not considered. Lemma 2. The union operator ∪ is closed to the set of the primitive relations Γ . For Proposition 1, if Γ∈yx, include an acyclic graph, also yx ∪ includes no cycle on its graph. A cycle breaks the level of the node, which closes the path on itself. Obviously, only a C relation can close a cycle. If x = (a,b) ∈ C and y ∈ Γ, then yx ∪ becomes the mixed relation denoted by, respectively, M1 (if y = (a,c) ∈ C) and M2 or M3 (if y ∈ D) relations. In fact, the level of nodes b and c forces either a disjunctive relation (obtaining the mixed relations M1 and M3) or a conjunctive relation (obtaining the mixed relation M2). Vice versa, if x = (b,a) ∈ C and y ∈ Γ, then becomes, similarly to the previous case, the mixed relation denoted by, respectively, M2 (if y = (a,c) ∈ C) and M1 or M3 (if y ∈ D) relations. yx ∪ Revised manuscript 22 In any case, M1 to M3 relations are disjunctive graphs. Hence, considering the former primitive, i.e. the C relation, the thesis is accomplished. Analogously, if x = [a,b] ∈ C and y ∈ Γ, then becomes the mixed relation denoted by, respectively, M1 (if y = (c,a) ∈ C, M2 (if y = (a,c) ∈ C) and M4 (if y ∈ D). Hence, the thesis is accomplished because M1 and M2 relations are digraphs. � Corollary 1. The level of the wildcard node inherits the levels of the nodes to which it is linked to (through D relations). Corollary 2. The level of all the nodes connected only by D relations is unchanged (i.e. M4). 5.2. Composition and decomposition of basic relations In this subsection, it will be shown how to assemble mixed routing rules to form a nonlinear precedence graph and achieve more complex rules. A list of compositions of mixed relations is shown in Table 4. All possible outcomes of compositions are listed. The same outcome can be obtained by different compositions (not shown). Compositions may involve nodes belonging to the same stage or to two consecutive stages; however, two consecutive stages can be linked by conjunctive arcs only. Decomposition rules can be inferred by following the opposite path. To explain the composition of mixed relations shown in Table 4, case #12. is considered: 33 MM ∪ . Both mixed relations have the same wildcard node 3. From Proposition 1, the two relations can be represented by the digraphs: DG1(M3)= (8) DG2(M3)= { } { } { } { } { }( ),]4,3[],3,2[,)4,2(,,,,4,3,2 432 iii ttt (9) The composite relation is: )3()3( 21 MDGMDG ∪ = { } { } { } { } { }( ),]4,3[],3,2[],3,1[,)4,2(),2,1(,,,,,4,3,2,1 4321 iiii tttt (10) yx ∪ { } { } { } { } { }( ),]3,2[],3,1[,)2,1(,,,,3,2,1 321 iii ttt Revised manuscript 23 which has a node (node 2) with one starting and one ending conjunctive arc. The nodes opposite to node 2 in these two conjunctive arcs have a level reduced by one or increased by one, respectively, if the node is on the conjunctive arc that ends to 2 and if the node is on the conjunctive arc that starts from 2. For Corollary 1 the level of the wildcard node inherits the levels of the nodes to which it is linked to (through D relations): Li 3={Li 1, Li 2= Li 1+1, Li 4= Li 1+2}. As a consequence, the routing rules are obtained by interposing 3 in the sequence compelled by the 2 conjunctive arcs, (1,2) and (2,4): 3�1�2�4, 1�3�2�4, 1�2�3�4, 1�2�4�3. Revised manuscript 24 Table 4. Composition of mixed relations (4 nodes) and deployment of their routing rules Case # Relation Combination of mixed relations Precedence graph Routing deployment #7. 21 MM ∪ 1 2 3 1 3 2 4 4 M1 M2 1 2 3 Li 1 Li 1+1 Li 1+1 4 2 3 L i 2 +1 Li 3=Li 2 Li 2 1 2 3 Li 1 Li 1+2 Li 1+1 4 Li 1+1 1 2 3 Li 1 Li 1+1 4 Li 1 1 3 2 4 1 3 4 2 Revised manuscript 25 #8. 31 MM ∪ M1 M3 1 2 3 1 3 2 4 4 1 2 4 3 #9. 41 MM ∪ 1 1 1 1 1 1 1 M4M1 2 3 4 2 4 3 3 2 4 3 4 2 4 2 3 4 2 1 2 3 Li 1 Li 1+1 Li 1+1 4 2 3 Li 2 {Li 1,Li 2} L i 1 1 2 3 Li 1 Li 1+2 Li 1+1 4 {Li 1+1, Li 1+2} 1 2 3 Li 1 Li 1+1 Li 1+1 4 2 3 L i 1 Li 1 Li 1 1 2 3 Li 1 Li 1+1 Li 1+1 4 Li 1+1 Revised manuscript 26 #10. 12 MM ∪ #11. 23 MM ∪ 1 2 3 1 3 2 4 4 3 1 2 4 M3 M2 4 2 3 L i 2 +1 Li 3=Li 2 Li 2 1 2 3 Li 1 Li 1+1 Li 1+1 1 2 3 Li 1 Li 1+1 4 Li 1 Li 1+1 1 3 2 4 1 3 4 2 4 2 3 Li 2 +1 Li 3=Li 2 Li 2 1 2 3 Li 2 {Li 1,Li 2} Li 1 1 2 3 Li 1 Li 1+2 Li 1+1 4 {Li 1, Li 1+1} Revised manuscript 27 #12. 33 MM ∪ 1 2 3 Li 2 {Li 1,Li 2} Li 1 1 2 3 Li 1 Li 1+1 4 {Li 1, Li 1+1, Li 1+2} Li 1+2 1 2 4 1 2 3 3 4 1 3 2 4 3 1 2 4 1 2 3 Li 1 Li 1+1 4 {Li 1, Li 1+1} { Li 1+1, Li 1+2} 1 2 3 1 3 2 4 4 3 1 2 4 1 3 4 2 Revised manuscript 28 #13. 43 MM ∪ 1 2 3 Li 2 {Li 1,Li 2} Li 1 3 2 Li 2 Li 2 4 Li 2 1 2 3 Li 1 Li 1+1 4 {Li 1, Li 1+1} {Li 1, Li 1+1} 1 2 3 1 2 4 4 3 1 3 2 4 1 3 4 1 4 2 2 3 1 4 3 2 3 1 2 4 3 1 4 2 3 4 1 2 4 1 2 3 4 1 3 2 4 3 1 2 Revised manuscript 29 Composite relations can be recombined into two mixed relations x and y with conjunctive arcs from x to y, which cross the boundaries of the two relations in the x-y direction (by construction), according to the following two alternatives (Table 5): 1. by a C relation, e.g. (3,2) in the precedence graph of row 1 for case #7. in Table 4; 2. by a D relation, e.g. [2,3] in row 2 for all other cases of Table 4 except case #7. (only case #13. is represented). In the first alternative, the combined mixed rule is modeled by two D relations between nodes with the same level. The routing rule of this recombination is unchanged with respect to the original case #7. in Table 4. In the second alternative, the combined mixed rule is modeled by splitting nodes with the same level and by connecting the split nodes by disjunctive arcs. Two mixed relations Mx are created. Each node of the D relation is split into two nodes, the original node and the dummy node, which are connected by a D relation. For example, node 2 is placed facing dummy node 2’ and they are connected by arc [2,2’]. The routing rule of the recombination imposes that only one node, between the original and the dummy node, will be included into the loading sequence of the selected resource. The not selected node will have no connection to the rest of the digraph and will not be considered by the routing rule. In both alternatives, in the combined mixed rule the red arrows cross the boundaries of the two relations from left to right and no arrow is present in the opposite direction in order to process more complex relations which satisfy the properties introduced above. Revised manuscript 30 Table 5. Decomposition rules for the mixed relations of C type (M1 ∪ M2) and of D type (M3 ∪ M3), yielding respectively a D to D relation and two split M3 relations. Case Precedence graph Combined mixed rule #7. 1 2 3 Li 1 Li 1+1 4 D D #12. 1 2 3 Li 1 Li 1+1 { L i 1, Li 1+1 } 2’ 3’ L i 1+1 4 { L i 1 , L i 1 +1 } { L i 1 +1, L i 1 +2 } Splitting operation 2 Splitting operation 3 M3 M3 6 A scheduling heuristic (H•HSSS) This section describes an example heuristic to show the usability of the proposed HSSS model for scheduling purposes. The developed algorithm is a list scheduler heuristic derived from Giffler and Thompson (1960), which was proven to generate a feasible schedule on the digraph, by visiting every node once and 1 2 3 Li 1 Li 1+1 4 Li 1 Li 1+1 1 2 3 Li 1 Li 1+1 4 {Li 1, Li 1+1} { Li 1+1, L Revised manuscript 31 only once. The list scheduler algorithm was originally proposed for classic job shop scheduling. At every step, a node is connected by means of a feasible move to the partially acyclic conjunctive graph, which represents the partial schedule. A feasible move is a disjunctive arc, which can be directed in the partial conjunctive graph without creating a cycle. The following pseudo code produces an acyclic graph, which includes the critical path. Heuristic for hybrid stage shop scheduling (H•HSSS) Input: a weighted digraph DG = (ℵ, WN, C, D, Ehk) O ← {Oi j k ⏐ i=1,..,N, j=1,..,li, k=1,..,S} i. AL1w←{Oi j k ∈O⏐ Oi j k ∩ O = ∅} ii. AL2w←{Oi j k ∈ AL1w | NiLij ,..,1,min = } iii. cti j k =ri, for each Oi j k ∈ AL2w for each k =1 to S do Initialization of Candidate Nodes: build, in two steps, the allowed list ALw for the current iteration w: i. AL1w←{Oi j k ∈ O⏐ Oi j k-1 ∩ O = ∅} ii. AL2w←{Oi j k ∈ AL1w | } for each w =1 to (Σi=1,..,N li) do 1. Initialization of Feasible Moves: mark as a feasible move each disjunctive arc (Oi’ j’ k, Oi j k) of Eh k where Oi j k ∈AL2w and Oi’ j’ k is the last operation of the loading sequence of resource h of the stage k (it creates the possibility for the candidate operation to become the new last operation of that loading sequence); 2. Move Selection: select a feasible move (Oi’ j’ k, Oi j k) of Ehk by directing the related disjunctive arc (Oi’ j’ k =‘dummy 0’,if k =1); 3. Update Routing: if ,*ijij LL = where Oi j* k is in the path which starts from 0 and ends at Oi j k, direct the last disjunctive arc in the path [Oi j* k, Oi j k]∈ D (after this action (Oi j* k, Oi j k) ∈ C) and update 1* += ijij LL ; finally, update the level on all the disjunctive arcs connected to Oi j k (i.e. 1** += ijij LL , if [Oi j k, Oi j** k] ∈ D); NiLij ,..,1,min = Revised manuscript 32 4. Arcs Removal: remove all the remaining disjunctive arcs connected to Oi’ j’ k (i.e. no other operation can be immediately subsequent to Oi’ j’ k in the loading sequence); remove all the remaining disjunctive arcs of Eh’k connected to Oi j k, i.e. h’≠h (i.e. no other loading sequence can include the operation); finally, remove all the remaining disjunctive arcs of D connected to Oi j* k (i.e. the routing must be linear after updating); 5. Computing Length: the length of the arc (Oi’ j’ k, Oi j k) is evaluated as the processing time of node Oi j k; this length is placed on the arc (Oi’ j’ k, Oi j k) and on the arc of the job routing (Oi j* k, Oi j k) which ends at Oi j k; also, the completion time cti j k = max{cti’j* k, cti j* k}+ti j k is placed as a mark of the node Oi j k, if Li j*<Li j; otherwise, if Li j*=Li j and two resources are available, min{ΤA k’, cti j* k}+ti j k, k’≠k, where T is the available time of resource k’; 6. Updating Structures: update O by removing operation Oi j k: O ← O \ {Oi j k}; end for 7. Directing the remaining disjunctive arcs: the arcs are connected to the dummy node *; Output: the acyclic conjunctive graph with the completion times for all the operations Lemma 3. The algorithm generates in: ))(( S l P S l NlO iii ++ (11) a complete selection of arcs of DG i.e. an acyclic conjunctive graph that includes all nodes. This property results from the following considerations: a) the achieved graph includes all the nodes: the main loop is performed |O| times and initially the candidate list includes all the nodes; at each iteration one and only one node is removed from candidate list (step 6); b) the achieved graph is conjunctive: i) for each iteration, the selected feasible move directs an arc of E which ends at the node removed from candidate list (step 2); ii) for each iteration, the node removed from candidate list is inserted in the path which starts from 0 (step 3); iii) Revised manuscript 33 all the disjunctive arcs of E and D which starts from the starting node of the conjunctive arc are removed (step 4); c) the conjunctive graph is acyclic: i) one and only one feasible move ends to a node which is in the candidate list (steps 1, 2 and 6); ii) one and only one routing rule ends to a node which is in the candidate list (steps 3 and 6). In order to evaluate the computational complexity of the algorithm, it can be noted that the main loop is performed n⋅m times and the most time-consuming steps are Arcs Removal and Computing Length step. The selection of a feasible move (Oi’ j k’, Oi j k) entails that the following alternative arcs are removed and the following connected nodes are marked: i) alternative sequencing arcs: all disjunctive arcs connected to the last operation Oi’j k’ in the loading sequence of resource h; they are at most N-1, one for each alternative job in order to approach the sequencing problem (the first member of expression (11complessità)); ii) alternative assigning arcs: all disjunctive arcs of Eh’ k , where h’ belonging to the stage k and h’≠h, connected to the candidate operation Oi j k; they are at most P-1, one for each alternative resource in order to approach the assigning problem (the second member of expression (11complessità)); iii) selecting nodes level: all nodes Oi j k, where Li j*≤ Li j, have marked with the completion time cti j k = max{cti’ j* k, cti j* k }+ti j (the third member of expression (11complessità)). � Because of these computational complexity considerations, the proposed algorithm finds a feasible schedule by means of an implicit visit of a large number of disjunctive arcs. Another consequence is that the three sequencing, assigning and selecting decisional points are considered at the same time in the selection of a feasible move because, at the same time, it is both an alternative sequencing arc and alternative assigning arc with the selected level. 7 Case study The case considered from the fashion industry includes five stages (Soldani, Rossi, & Lanzetta, 2013): 1. cutting – manual, die and machine cutting; Revised manuscript 34 2. preparation – primary and secondary part preparation; 3. varnishing; 4. assembly – manual and machine stitching; 5. finishing – finishing operations and packaging. Jobs include single parts (€€ 25.000 woman bags) or batches. A 5×5 benchmark, with 5 stages and 5 jobs, characterized as 2-HSSS|ℵ34|C23|D6|R10|N5|S5|max Li j6, by a total of 34 operations with 23 conjunctive arcs, 6 disjunctive arcs (6/34=17%) and a maximum number of levels of 6. The simplest non-trivial case of P=2 parallel (or identical) resources per stage has been considered, to make it a hybrid stage shop scheduling benchmark. The release dates for jobs and resources are zero. The case 1-HSSS|ℵ34|C23|D6|R5|N5|S5|max Li j6, with P=1 machine per stage has also been analyzed for comparison and can be considered as a special case of hybrid stage shop scheduling. Consequently, the general H•HSSS pseudo code can be applied to both benchmarks. Revised manuscript 35 111O 122O 132O 143O 154O 164O 175O 211O 221O 231O 285O242O 253O 263O 274O 333O 355O344O311O 322O 411O 422O 475O 485O 565O554O511O 522O 533O 543O 432O 453O 464O Stage 1 Stage 2 Stage 3 Stage 4 Stage 5 54 49 30 16 25 41 58 30 33 3 39 50 58 56 15 11 49 31 20 71 33 26 15 68 40 45 442O 40 77 56 29 60 78 53 Revised manuscript 36 Figure 4. Proposed P-HSSS|ℵ34|C23|D6|R5P|N5|S5|max Li j6 benchmark with 5 stages (columns) and 5 jobs (rows), including the precedence relations in Table 2 and Table 3. Colored arcs for P not shown for clarity. Figure 4 about here The benchmark in Figure 4 has been designed in order to include all the precedence relations introduced above, and recalled in Table 6; inter-stage relations are denoted by the two corresponding consecutive stage numbers, consequently it can be noticed there is no biunivocal relation between stages and primitive relations. Table 6. Map of job routing per precedence relations for the benchmark in Figure 4 STAGE JOB 1 1-2 2 2-3 3 3-4 4 4-5 5 1 Node M1 D M2 Node C Node 2 M4 Node M1 D M2 Node Node 3 Node Node Node Node Node 4 Node M1 M3 Node Node M1 D 5 Node Node C Node Node Table 6 shows 6 occurrences of precedence graphs spanning across two stages. By applying the H•HSSS pseudo code on the two benchmarks considered, the Gantt diagrams of the obtained schedule are shown in Figure 5. Revised manuscript 37 OPERATOR 1 15 S T A G E 1 2-23-1 1-1 5-1 45 99 OPERATOR 3 26 S T A G E 2 3- 4-31-2 4-4 129 155 195 OPERATOR 5 75 S T A G E 3 3-3 4-52-6 5-4 125 306210 OPERATOR 8 5-5 183 364230 2-7 1-6 OPERATOR 9 126 S T A G E 5 3-5 288 333 1-7 4-8 OPERATOR 2 2-1 S T A G E 1 2-3 4-1 20 53 124 176 OPERATOR 4 2 S T A G E 2 4-2 5-2 56 148 181 237 1-3 OPERATOR 6 2-5 S T A G E 3 5-3 95 164 1-4 266 OPERATOR 4 106 S T A G E 4 3-4OPERATOR 7 106 S T A G E 4 3-4 4-6 189 278 1-5 S T A G E 4 OPERATOR 10 S T A G E 5 5-6 239 397 = Makespan318 2-8 4-7 Revised manuscript 38 Figure 5. Gantt for the 2-HSSS|ℵ34|C23|D6|R10|N5|S5|max Li j6 (top) and for the the 1-HSSS|ℵ34|C23|D6|R5|N5|S5|max Li j6 (bottom) benchmark in Figure 4. Larger numbers in colored bars express jobs with their respective operations (smaller numbers) and completion time (superscripts) Figure 5 about here RESOUR CE 1 15 2-1 S T A G E 1 2-2 3-1 2-3 1-1 4-1 5-1 35 65 98 223 300152 RESOUR CE 2 26 S T A G E 2 4-2 4-3 1-2 4-4 5-2 101 182 231 290 386257 1-3 330 RESOUR CE 3 75 2-5 S T A G E 3 3-3 4-5 2-6 5-3 5-4 140 190 247 455345 1-4 415 RESOUR CE 4 106 S T A G E 4 3-4 4-6 5-5 248 273 513314 413 2-7 1-5 1-6 RESOUR CE 5 126 S T A G E 5 3-5 5-6 304 372 546 = Makespan 498453 2-8 1-7 4-84-7 3-2 2-4 Revised manuscript 39 At stage 1 in the top Gantt of Figure 5 for the 2-HSSS benchmark, the routing rule of the mixed relation M4 allows the overlapping among (micro) operations of job 2, renamed as 2-1, 2-2 and 2-3 for readability. Resources 1 and 2 overlap in processing, respectively, operations 2-1 and 2-2, while operation 2-3 is processed on resource 1 as soon as the resource becomes available. The three mixed relations M1 between stage 1 and 2, between stage 2 and 3 and between stage 4 and 5 are merged in three composite relations M1 ∪ M2. All these relations are managed as D relations (also see Table 6). The corresponding routing rule allows overlapping operations 1-1 and 1-2 on the alternative resources of stage 2; analogously, for D relations of stage 3 and 5. The routing rule of the mixed relation M3 of job 4 at stage 2 is scheduled by overlapping operations 4-2 and 4-3 on the two alternative resources; operation 4-4 is processed after operation 4-3 has been completed, because of the conjunctive relation (4-3, 4-4) included in the mixed relation M3. The adopted routing rule is 4-3�4-4; it is only applied to resource 1, because operation 4-2 is processed on resource 2. At stage 1 in the bottom Gantt of Figure 5 for the 1-HSSS benchmark, the routing rule of the mixed relation M4 does not allow the overlapping among (micro) operations 2-1, 2-2 and 2-3, which are processed sequentially by resource 1. Similarly, the three mixed relations M1, including operations 1-2 and 1-3 at stage 2, 4-3, 4-2 and 4-4 at stage 2 and 2-5 and 2-6 at stage 3, which are managed as D relations (also see Table 6), are processed sequentially by resource 1. The routing rule of the mixed relation M3 of job 4 at stage 2 is processed sequentially by resource 1 with routing 4-3, 4-2 and 4-4. It can be noticed that splitting operations does not affect the completion time on each resource, because only permutations are possible. Allowed permutations are according to the selected routing rules of Table 2, Table 3 and Table 4. Inversely, the presence of two or more resources allows further potential makespan reduction by overlapping. As designed, the proposed heuristics provides feasible schedules. The Move Selection at step 2. can be improved by the use of dispatching rules to obtain better solutions. It can also be observed that the proposed theory (model, description and code) is applicable also with only one resource per stage and that it has been profitably applied to the examined case. 8 Conclusion This work has been inspired by a real manufacturing case from the high-fashion industry. A hybrid stage shop scheduling model has been proposed. The problem has been described by mixed integer linear programming. The solution has been formalized by a disjunctive graph representation, a powerful tool to design scheduling optimization (metaheuristics) algorithms. An O(P⋅S⋅N2) Revised manuscript 40 constructive method (heuristics) to generate a feasible plan/schedule in the most general case has been proposed and applied to a case study; it is based on priority levels obtained from the precedence graphs by deploying alternative routings. To find an optimal solution, metaheuristics like ant colony optimization, iterated local search and tabu search that produce constructive solutions may benefit of the graph representation. The hybrid stage shop scheduling model is able to provide shorter scheduling, by overlapping operations and by splitting on the same job, increasing the number of scheduling alternatives. These advantages, obtained at the cost of increased complexity, require resource and organization flexibility: more complex material flow and multifunctional resources (machines and/or operators). The proposed hybrid stage shop scheduling model completes the mixed stage shop model by resource flexibility and by integrating process planning and group shop scheduling (operational and routing flexibility). Greater flexibility and shorter makespan are available versus AND/OR graphs. Other contributions include the addition of operation overlapping and splitting within stages and across consecutive stages with the composition and the decomposition of primitive (2 nodes) and mixed (3 and 4 nodes) relations. The universe of new more complex relations has been explored and discussed both concerning the application implication and the theoretical background. The wildcard node mechanism is the key concept to fully exploit the potential of parallel resources, by allowing alternative plans obtained by permutation of operations linked by undefined or arbitrary routing, modeled as immaterial precedence constraints. Immaterial links, allow the integration of process planning and scheduling, with possible relapse in assembly/disassembly, computer aided design and process planning, manufacturing and service scheduling, and more generally in project management. 9 Future work Mixed relations of the hybrid stage shop scheduling are included in a single stage or may span across two consecutive stages. The extension of immaterial constraints to multiple stages in hybrid shop scheduling (from HSSS to HSS), represents a new research area. The most general case of precedence graphs involving more than two consecutive stages has been examined and solved in a case study. Combining further the new mixed type relation, more complex nonlinear relations (5+ nodes) can be speculated both on the theoretical and algorithmic sides. Among developments is the unifying theory for α nodes and χ immaterial relations (and γ levels). An alternative interpretation of immaterial precedence constraints is considering resources as in a local job shop. In a perspective where the flowshop can be considered as a special case of the job Revised manuscript 41 shop, the hybrid stage shop scheduling model can be considered as a smooth transition between a flowshop and the more general case of job shop. In this regard the mentioned χ parameter may represent a measure of the flow grade of a given job shop. From a practical viewpoint, job shop solutions can be transferred from the literature. Acknowledgements The authors would like to thank the management of the anonymous high-fashion company from a reputed international group for providing the information about their manufacturing process. References Ahmadizar, F., & Shahmaleki, P. (2014). Group shop scheduling with sequence-dependent setup and transportation times. Applied Mathematical Modelling, 38, 5080-5091. Aloulou, M.A., & Artigues, C. (2010). Flexible solutions in disjunctive scheduling: General formulation and study of the flow-shop case. Computers & Operations Research, 37, 890- 898. Amin-Naseri, M.R., & Afshari, A.J. (2012). A hybrid genetic algorithm for integrated process planning and scheduling problem with precedence constraints. International Journal of Advanced Manufacturing Technology, 59, 273–287. Behnamian, J., & Zandieh, M. (2011). A discrete colonial competitive algorithm for hybrid flowshop scheduling to minimize earliness and quadratic tardiness penalties. Expert Systems with Applications, 38, 14490-14498. Benavides, A. J., Ritt, M., & Miralles, C. (2014). Flow shop scheduling with heterogeneous workers. European Journal of Operational Research, 237, 713-720. Bensmaine, A., Dahane, M., & Benyoucef, L. (2013). A non-dominated sorting genetic algorithm based approach for optimal machines selection in reconfigurable manufacturing environment. Computers & Industrial Engineering, 66, 519-524. Blazewicz, J., & Kobler, D. (2002). Review of properties of different precedence graphs for scheduling problems. European Journal of Operational Research, 142, 435-443. Bentaha, M. L., Battaïa, O., Dolgui, A., & Hu, S. J. (2014). Dealing with uncertainty in disassembly line design. CIRP Annals-Manufacturing Technology, 63, 21-24. Blum, C., & Sampels, M. (2004). An Ant Colony Optimization Algorithm for Shop Scheduling Problems. Journal of Mathematical Modelling and Algorithms, 3, 285-308. Revised manuscript 42 Dauzère-Pérès, S., Roux, W., & Lasserre, J.B. (1998) Multi-resource shop scheduling with resource flexibility. European Journal of Operational Research, 107, 289-305. Doh, H. H., Yu, J. M., Kim, J. S., Lee, D. H., & Nam, S. H. (2013). A priority scheduling approach for flexible job shops with multiple process plans. International Journal of Production Research, 51, 3748-3764. Ferrell, W., Sale, J., Sams, J., & Yellamarju, M. (2000). Evaluating simple scheduling rules in a mixed shop environment. Computers & Industrial Engineering, 38, 39–66. Giffler, D., & Thompson, G.L. (1960) Algorithms for solving production scheduling problems. Operation Research, 8, 487–503. Go, T. F., Wahab, D. A., Rahman, M. A., Ramli, R., & Hussain, A. (2012). Genetically optimised disassembly sequence for automotive component reuse. Expert Systems with Applications, 39, 5409-5417. Kacem, I., Hammadi, S., & Borne, P. (2002). Approach by localization and multiobjective evolutionary optimization for flexible job-shop scheduling problems. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, 32, 1-13. Kis, T. (2003). Job-shop scheduling with processing alternatives. European Journal of Operational Research, 151, 307–332. Kozan, E., & Liu, S.Q. (2012) A demand-responsive decision support system for coal transportation. Decision Support Systems, 54, 665-680. Leu, M.C., Bernard, A., ElMaraghy, H.A., Nee, A.Y.C., Ong, S.K., Lanzetta, M., Putz, M., & Zhu, W. (2013). CAD Model Based Virtual Assembly Simulation, Planning and Training. CIRP Annals-Manufacturing Technology, 62, 799-822. Leung, C.W. (2010). Integrated process planning and scheduling by an agent-based ant colony optimization. Computers & Industrial Engineering, 59, 166–180. Li, D., Meng, X., Liang, Q., & Zhao, J. (2015). A heuristic-search genetic algorithm for multi-stage hybrid flow shop scheduling with single processing machines and batch processing machines. Journal of Intelligent Manufacturing. Article in press. Liu, S.Q., Ong, H.L., & Ng, K.M. (2005). A fast tabu search algorithm for the group shop scheduling problem. Advances in Engineering Software, 36, 533-539. Masuda, T., Ishii, H., & Nishida, T. (1985) The mixed shop scheduling problem. Discrete Applied Mathematics, 11, 175-186. Nasiri, M.M., & Kianfar, F. (2011). A GA/TS algorithm for the stage shop scheduling problem. Computers & Industrial Engineering, 61, 161–170. Revised manuscript 43 Otto, C., & Otto, A. (2014). Multiple-source learning precedence graph concept for the automotive industry. European Journal of Operational Research, 234, 253-265. Phanden, R. K., Jain, A., & Verma, R. (2013). An approach for integration of process planning and scheduling. International Journal of Computer Integrated Manufacturing, 26, 284-302. Ribas, I., Leisten, R., & Framinan, J.M. (2010). Review and classification of hybrid flowshop scheduling problems from a production system and a solutions procedure perspective. Computers & Operations Research, 37, 1439–1454. Rossi, A. (2014). Flexible job shop scheduling with sequence-dependent setup and transportation times by ant colony with reinforced pheromone relationships. International Journal of Production Economics, 153, 253-267. Rossi, A., & Lanzetta, M. (2013). Scheduling Flow Lines with Buffers by Ant Colony Digraph. Expert Systems with Applications, 40, 3328-3340. Rossi, A., & Lanzetta, M. (2014). Native Metaheuristics for Non-Permutation Flowshop Scheduling. Journal of Intelligent Manufacturing, 25, 1221–1233. Rossi, A., Pandolfi, A., & Lanzetta, M. (2013). Dynamic set-up rules for hybrid flow shop scheduling with parallel batching machines. International Journal of Production Research, 52, 3842-3857. Ruiz, R., & Maroto, C. (2005) A comprehensive review and evaluation of permutation flowshop heuristics. European Journal of Operational Research, 165, 479–494. Shakhlevich, N.V., Sotskov, Y.N., & Werner, F. (2000). Complexity of mixed shop scheduling problems: A survey. European Journal of Operational Research, 120, 343–351. Soldani, S., Rossi, A., & Lanzetta, M. (2013). Modellazione di un ciclo produttivo nel settore moda, Automazione Integrata, 16, 37-39. Stecke, K. E., & Raman, N. (1995). FMS planning decisions, operating flexibilities, and system performance. Engineering Management, IEEE Transactions on, 42, 82-90. Tan, W., & Khoshnevis, B. (2000). Integration of process planning and scheduling—a review. Journal of Intelligent Manufacturing, 11, 51-63. Whizzkids '97, http://www.win.tue.nl/whizzkids/1997. 
work_bycggifu5rbtvaonvhceqkdy4y	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright An optimum feature extraction method for texture classification Engin Avci *, Abdulkadir Sengur, Davut Hanbay Firat University, Department of Electronic and Computer Science, 23119, Elazig, Turkey a r t i c l e i n f o Keywords: Pattern recognition Texture classification Optimum feature extraction Discrete wavelet transform Entropy Energy Genetic algorithm Neural networks Intelligent systems a b s t r a c t Texture can be defined as a local statistical pattern of texture primitives in observer’s domain of interest. Texture classification aims to assign texture labels to unknown textures, according to training samples and classification rules. In this paper a novel method, which is an intelligent system for texture classifi- cation is introduced. It used a combination of genetic algorithm, discrete wavelet transform and neural network for optimum feature extraction from texture images. An algorithm called the intelligent system, which processes the pattern recognition approximation, is developed. We tested the proposed method with several texture images. The overall success rate is about 95%. � 2008 Published by Elsevier Ltd. 1. Introduction Texture classification aims to assign texture labels to unknown textures, according to training samples and classification rules. Two major issues are critical for texture classification: the texture classification algorithms and texture feature extraction. The main aim of texture classification is to find a best matched category for a given texture among existing textures. During the past decades wavelet analysis has become a power- ful tool for multi-resolution analysis. Important applications can be found in various fields, ranging from remote sensing to biomedical imaging. Intuitively, multi-scale wavelet analysis is an ideal ap- proach to analyze texture because it is well recognized that scale is one of the most important aspects of texture information. Many previous works are proposed to solve texture classification prob- lem (Chellappa, Chatterjee, & Bagdazian, 1985; Chen & Chen, 1999; Jain & Farrokhnia, 1991). A comparison of textural features from Fourier power spectrum, second-order gray level statistics and first-order statistics of gray level differences were shown in Weszka, Dyer, and Rosenfeld (1976). Other textural features including co-occurrence features, Gabor features, Markov random fields (MRF) based features and fractal features were compared in another work (Chen & Chen, 1999). In Muneeswaran, Ganesan, Arumugam, and Ruba Soundar (2005), a new rotational and scale invariant feature set for textural image classification was pre- sented and a combined invariant feature (CIF) set was proposed. It is an integration of the crude wavelets like Gaussian, Mexican Hat and orthogonal wavelets like Daubechies to achieve a high quality rotational and scale invariant feature set. Also it is added with features obtained using the newly proposed weighted smoothening Gaussian filter masks to improve the classification results. Li et al. proposed a dyadic wavelet, wavelet frame, Gabor wavelet, and steerable pyramid for texture classification problem with individual and combined multi-resolution features (Shutao & Shawe-Taylor, 2005). They used support vector machines (SVM) as classifier. An SVM for texture classification, using transla- tion-invariant features generated from the discrete wavelet frame transform is also proposed in Shutao, James, Kwok Hailong, and Yaonan (2003). In Arivazhagan and Ganesan (2003), Arivazhagan and Ganesan used wavelet statistical features, wavelet co-occur- rence features and combination of wavelet statistical features and co-occurrence features of one-level wavelet transformed images with different feature databases. A distance classifier is used for measurement of similarity and consequent labeling. Moj- silovic et al. investigated in their paper whether the properties of decomposition filters play an important role in texture description, and which feature was dominant in the selection of an optimal fil- ter bank (Mojsilovic, Rackov, & Popovic, 2000). They performed classification experiments with 23 Brodatz textures. The experi- mental results indicated that the selection of the decomposition filters has a significant influence on the result of texture character- ization. Finally, the paper ranked 19 orthogonal and biorthogonal filters, and establishes the most relevant criteria for choice of decomposition filters in wavelet-based texture characterization algorithms. Chen and Chen gave a comparative study for filtering methods for texture classification according to their discrimination ability (Chen & Chen, 1999). Their main aim was to evaluate the performance of four filtering methods including Fourier transform, spatial filter, Gabor filter and wavelet transform for classifica- tion. The superiority of the wavelet transform is shown by their 0957-4174/$ - see front matter � 2008 Published by Elsevier Ltd. doi:10.1016/j.eswa.2008.06.076 * Corresponding author. Tel.: +90 4242370000x4257; fax: +90 4242367064. E-mail address: enginavci23@hotmail.com (E. Avci). Expert Systems with Applications 36 (2009) 6036–6043 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a experimental study. Wang et al; proposed a multi-resolution MRF (MRMRF) modeling to describe textures (Lei & Jun, 1999). MRMRF modeling is a method trying to fuse filtering theory and MRF models. In this model, both high pass and low pass components are considered. Huang and Aviyente analyzed the dependence of energy values from different subbands, which may be from the same wavelet basis or different wavelet bases (Huang & Aviyente, 2006). They proposed and information-theoretic measure, mutual information, to select subbands for sparse representation and texture classification. Sengur et al.; used wavelet packet neural networks (WPNN) for texture classification (Sengur, Turkoglu, & Ince, in press).The proposed schema composed of a wavelet packet feature extractor and a multilayer perceptron classifier. Entropy and energy features are integrated wavelet feature extractor. The performed experimental studies show the effectiveness of the WPNN structure. In this paper, a novel method which is an expert system for tex- ture classification is introduced. It used a combination of genetic algorithm, discrete wavelet transform and neural network for opti- mum feature extraction from texture images. An algorithm called the intelligent system, which processes the pattern recognition approximation, is developed. In texture image classification area, the novelties presented of this paper can be summarized as follows: 1. The presented first novelty in this study is, the using an effec- tively optimum feature extraction method that increases per- centage of the correct texture image classification. 2. The presented second novelty in this study is the using of genetic-discrete wavelet-neural network model for selecting of the feature extraction method and finding the optimum wavelet filter and wavelet entropy parameter values for opti- mum feature extraction from texture images. This wavelet filter type and entropy parameters are used for obtaining the wavelet entropy values from wavelet layer as a newness and efficiently method in texture image classification area. The paper is organized as follows. In Section 2, discrete wavelet transform (DWT), artificial neural networks (ANN) are mentioned. Methodology and applications using genetic-discrete wavelet-neu- ral network (GDWNN) is described in Section 3. This GDWNN method enables a large reduction of the texture image data while retaining problem specific information, which facilitates an effi- cient texture classification process. The effectiveness of the pro- posed method for classification of texture images is demonstrated in Section 4. Finally, Section 5 presents discussion and conclusion. 2. Theory 2.1. Discrete wavelet transform Discrete wavelet transform (DWT) is finding inverse use in fields as diverse as telecommunications and biology. Because of their suitability for analyzing non-stationary signals, they have be- come a powerful alternative to Fourier methods in many medical applications, where such signals abound (Daubechies, 1988). The main advantages of wavelets is that they have a varying window size, being wide for slow frequencies and narrow for the fast ones, thus leading to an optimal time-frequency resolution in all the fre- quency ranges. Furthermore, owing to the fact that windows are adapted to the transients of each scale, wavelets lack the require- ment of stationary. The (continuous) wavelet transform of a 1-D signal f(x) is de- fined as ðW wfÞða; bÞ¼ hf ; wða; bÞi¼ Z þ1 �1 fðxÞw�ða;bÞðxÞdx; wða;bÞ ¼ a �1=2w x � b a � � ; ð1Þ where a is the scaling factor, b is the translation parameter related to the location of the window, and w*(x) is the transforming func- tion. The latter is also called the mother wavelet, which is the pro- totype for generating the other window functions. The extension to 2-D is usually performed by using a product of 1-D filters. The transform is computed by applying a filter bank as shown in Fig. 1. L and H to denote the 1-D low pass and high pass filter, respectively. The rows and columns of image are processed separately and down sampled by a factor of 2 in each direction, resulting in one low pass image LL and three detail images HL, LH, and HH. (Fig. 2)a shows the one-level decomposition of Fig. 1 in the spatial domain. The LH channel contains image information of low horizontal frequency and high vertical frequency, the HL channel contains high horizontal frequency and low vertical fre- quency, and the HH channel contains high horizontal and high ver- tical frequencies. Three-level frequency decomposition is shown in Fig. 2b. Note that in multi-scale wavelet decomposition only the LL subband is successively decomposed. 2.2. Genetic algorithms The genetic algorithms use an evolutionary process for solving a problem. The genetic algorithm begins with a set of solutions which are represented by individuals. These sets of solution are called as population. It is taken the solutions from one population and use to form a new population. This iterative process is main- tained during the new population will be better than the old one. Solutions that are then selected to form new solutions are selected according to their fitness values. If fitness value of an individual is better than other’s, this individual is more lucky than other for be reproduced at next population. This iterative process is repeated until some conditions (for example, number of populations or improvement of the best solution, etc.) are satisfied (http:// cs.felk.cvut.cz/~xobitko/ga/ (accessed: 20.8.04)). Outline of the ba- sic genetic algorithm can be given as below: � It is generated random population of n individuals which pro- vide suitable solutions for the problem. � It is evaluated the fitness f(x) of each individual x in the popula- tion (http://cs.felk.cvut.cz/~xobitko/ga/ (accessed: 20.8.04)). � It is created a new population by repeating following steps until the new population is complete. Fig. 1. A one-level wavelet analysis filter bank. E. Avci et al. / Expert Systems with Applications 36 (2009) 6036–6043 6037 � It is selected two parent individuals from a population according to their fitness. The better fitness is interpreted as the bigger chance to be selected for next population. � It is crossed over the parents with a crossover probability to form new individuals. If crossover wasn’t performed, individual would be the exact copy of parents. � It is mutated new individual with a mutation probability at each locus which is position in individual. � It is placed new individuals in the new population. � It is used new generated population for a further run of the algorithm. � It is stop the genetic algorithm, if the end condition is satisfied, and return the best solution in current population. � It is gone to step second. 2.3. The artificial neural networks An artificial neural network (ANN) is a mathematical model consisting of a number of highly interconnected processing ele- ments organized into layers, the geometry and functionality of which have been likened to that of the human brain. The ANN may be regarded as processing learning capabilities. It has natural propensity for storing experimental knowledge. By virtue of its parallel distribution, an ANN is generally robust for tolerant of faults and noise, able to generalize well and capable of solving non-linear problems (Haykin, 1994). 3. Methodology and applications Feature extraction is the key for pattern recognition so that it is the most important component of designing the intelligent system based on pattern recognition since the best classifier will perform poorly if the features are not chosen well (Avci, Turkoglu, & Poyraz, 2005). A feature extractor should reduce the pattern vector (i.e., the original waveform) to a lower dimension, which contains most of the useful information from the original vector. 3.1. Entropy and energy values as features Entropy is a quantity that is widely used in information theory and is based on probability theory (Tzanakou, 2000). Entropy is a common concept in many fields, mainly in mechanic, image pro- cessing and signal processing. The general form of the entropy function is shown as follows: HðXÞ¼� Xn i¼1 pilog2pi ð2Þ where X is a random variable which can be one of the values x1, x2, . . . , xn with probability p1, p2, . . . , pn. Note that if pi = 0, then 0log2 0 is defined to be 0. Thus, H(X) can be interpreted as represent- ing the amount of uncertainty that exists in the value of X. In infor- mation theory, entropy value is considered to be an average amount of information received when the value of X is observed. In this pa- per, we used norm entropy and Sure entropy, respectively. The Norm entropy HN(s) is defined as follows: HNðxÞ¼ Xn i¼0 jxij p for ð1 6 p < 2Þ ð3Þ where x is signal and xi is ith component of the given x variable (Coifman & Wickerhauser, 1992; MATLAB, 2003). The Sure entropy HS(x) is defined as below: jxij6 e ) HSðxÞ¼ Xn i¼0 minððx2i ; e 2Þ ð4Þ There, e is a positive threshold value (Coifman & Wickerhauser, 1992; MATLAB, 2003). Where x is variable and xi is ith component of the given variable. Energy is one of the most commonly used features for texture analysis. In this study we used the averaged l2-norm which is de- fined as follows: El2 ¼ 1 N Xn i¼1 ðxiÞ 2 ð5Þ where N is the number of the components in the x variable. 3.2. Optimum feature extraction and classification by using GDWNN In this study, the genetic algorithm is used for obtaining the optimum wavelet filter and values of the entropy parameter. For this reason, the feature extraction methods are generated using db2, db3, db5, db10, sym2, sym3, sym5, sym7, sym8, bior1.3, bior2.2, bior2.8, bior6.8, coif1, coif2, coif3, coif5 wavelet filters sep- arately. These feature extraction methods can be named as the following: (1) The feature extraction method by using db2 wavelet filter (FEM- 1): For wavelet decomposition of the various texture images, the decomposition and reconstruction structures at level 1 Fig. 2. Wavelet frequency decomposition. 6038 E. Avci et al. / Expert Systems with Applications 36 (2009) 6036–6043 are shown in Fig. 2. DWT is applied to the texture images by using the db2 wavelet decomposition filters. Thus, three detail images and one approximate image are obtained: one-approximation coefficients LL and three detail coeffi- cients LH, HL, HH. A representative example of the wavelet decomposition of the texture image is shown in Fig. 2. One-level DWT decomposition of the Brickwall texture using db2 is shown in Fig. 3. (2) The feature extraction method by using db3 wavelet filter (FEM- 2): In FEM-2, the previous method’s (FEM-1) descriptions are valid. In this method, only db3 wavelet filter is used differ- ently from the FEM-1 for DWT decomposition stage of GDWNN. (3) The feature extraction method by using db5 wavelet filter (FEM- 3): In FEM-3, the descriptions of FEM-1 are valid. In this method, only db5 wavelet filter is used differently from the FEM-1 for DWT decomposition stage of GDWNN. (4) The feature extraction method by using db10 wavelet filter (FEM-4): In FEM-4, the descriptions of FEM-1 are valid. In this method, only db10 wavelet filter is used differently from the FEM-1 for DWT decomposition stage of GDWNN. (5) The feature extraction method by using sym2 wavelet filter (FEM-5): In FEM-5, the previous method’s (FEM-1) descrip- tions are valid. In this method, only sym2 wavelet filter is used differently from the FEM-1 for DWT decomposition stage of GDWNN. (6) The feature extraction method by using sym3 wavelet filter (FEM-6): In FEM-6, the previous method’s (FEM-1) descrip- tions are valid. In this method, only sym3 wavelet filter is used differently from the FEM-1 for DWT decomposition stage of GDWNN. (7) The feature extraction method by using sym5 wavelet filter (FEM-7): In FEM-7, the previous method’s (FEM-1) descrip- tions are valid. In this method, only sym5 wavelet filter is used differently from the FEM-1 for DWT decomposition stage of GDWNN. (8) The feature extraction method by using sym8 wavelet filter (FEM-8): In FEM-8, the previous method’s (FEM-1) descrip- tions are valid. In this method, only sym8 wavelet filter is used differently from the FEM-1 for DWT decomposition stage of GDWNN. (9) The feature extraction method by using bior1.3 wavelet filter (FEM-9): In FEM-9, the previous method’s (FEM-1) descrip- tions are valid. In this method, only bior1.3 wavelet filter is used differently from the FEM-1 for DWT decomposition stage of GDWNN. (10) The feature extraction method by using bior2.2 wavelet filter (FEM-10): In FEM-10, the previous method’s (FEM-1) descriptions are valid. In this method, only bior2.2 wavelet filter is used differently from the FEM-1 for DWT decompo- sition stage of GDWNN. (11) The feature extraction method by using bior2.8 wavelet filter (FEM-11): In FEM-11, the previous method’s (FEM-1) descriptions are valid. In this method, only bior2.8 wavelet filter is used differently from the FEM-1 for DWT decompo- sition stage of GDWNN. (12) The feature extraction method by using bior6.8 wavelet filter (FEM-12): In FEM-12, the previous method’s (FEM-1) descriptions are valid. In this method, only bior6.8 wavelet filter is used differently from the FEM-1 for DWT decompo- sition stage of GDWNN . (13) The feature extraction method by using coif1 wavelet filter (FEM-13): In FEM-13, the previous method’s (FEM-1) descriptions are valid. In this method, only coif1 wavelet fil- ter is used differently from the FEM-1 for DWT decomposi- tion stage of GDWNN. (14) The feature extraction method by using coif2 wavelet filter (FEM-14): In FEM-14, the previous method’s (FEM-1) descriptions are valid. In this method, only coif2 wavelet fil- ter is used differently from the FEM-1 for DWT decomposi- tion stage of GDWNN. (15) The feature extraction method by using coif3 wavelet filter (FEM-15): In FEM-15, the previous method’s (FEM-1) descriptions are valid. In this method, only coif3 wavelet fil- ter is used differently from the FEM-1 for DWT decomposi- tion stage of GDWNN. (16) The feature extraction method by using coif5 wavelet filter (FEM-16): In FEM-16, the previous method’s (FEM-1) descriptions are valid. In this method, only coif5 wavelet fil- ter is used differently from the FEM-1 for DWT decomposi- tion stage of GDWNN. The feature extraction methods and binary equivalent of the feature extraction methods are shown in Table 1. Structural and operational process of the GDWNN: The proposed GDWNN algorithm is developed for determining the most efficient method of 16 texture feature extraction methods for texture dis- crimination which was explained in Section 3.2. Besides choosing the most appropriate wavelet decomposition filter, the proposed algorithm has the ability of selecting the optimum P-parameter va- lue of the norm entropy and optimum e-parameter value of the sure entropy for efficient features which characterize the texture Fig. 3. One-level DWT decomposition of the Brickwall image using db2. Table 1 The feature extraction methods and binary equivalent of the feature extraction methods Feature extraction method Binary equivalent of the feature extraction method Feature extraction method Binary equivalent of the feature extraction method FEM-1 0 0 0 0 FEM-9 1 0 0 0 FEM-2 0 0 0 1 FEM-10 1 0 0 1 FEM-3 0 0 1 0 FEM-11 1 0 1 0 FEM-4 0 0 1 1 FEM-12 1 0 1 1 FEM-5 0 1 0 0 FEM-13 1 1 0 0 FEM-6 0 1 0 1 FEM-14 1 1 0 1 FEM-7 0 1 1 0 FEM-15 1 1 1 0 FEM-8 0 1 1 1 FEM-16 1 1 1 1 E. Avci et al. / Expert Systems with Applications 36 (2009) 6036–6043 6039 images. The operational processes for employing the genetic algo- rithm are listed below: 3.2.1. Step 1. Composing of the initial population The initial population of the genetic algorithm is consisted of 20 individuals. Each individual is a word of 10 bits. These words are consisted of three segments. The first segment of an individual rep- resents the wavelet filter type and four bits are enough for repre- senting this segment because we have 16 wavelet decomposition filters. Fifth, sixth, and seventh bits of the each individual represent the P-parameter value of the norm entropy which is mentioned un- der Section 3.2.1. Moreover, eighth, ninth, and tenth bits of each individual represent e-parameter value of the sure entropy. In norm entropy, P is the power and should be chosen between [1, 2). In sure entropy, e is the threshold and it also should be se- lected between [1, 8). Sensitivity for P-parameter is 1/7 since P- parameter is represented by using three bits for each of individual of the population. Thus, the P-parameter gets one of the values from the set Xp. Xp is consist of the following components: 1, 1.142, 1.285, 1.426, 1.568, 1.71, 1.852 and 1.994. e-Parameter is represented by three bits for each individual of the population. The e-parameter gets one of the following values 1–8. According to this, we can convert these values to binary form in Table 2. For example, two individuals of population may be shown as below: 0 0 0 0 FEM-1 0 0 1 value of the P parameter = 1.142 0 1 1 value of the parameter = 4 a random individual = 0 0 0 0 0 0 1 0 1 1 FEM-2 0 1 0 value of the P parameter = 1.285 0 0 1 value of the parameter = 2 a random individual = 0 0 0 1 0 1 0 0 0 1 0 0 0 1 3.2.2. Step 2. Feature extraction mechanism 1. Each individual of the population represent one of the feature extraction methods which are given in Section 3.2. P-parameter value is used for norm entropy and e-parameter value is used for sure entropy. When a random individual is selected, the appropriate wavelet filter type and the pre-defined parameter values are assigned to the related parameters (P and e). This individual is then sent to the feature extraction mechanism. 2. The feature extraction mechanism tries to obtain the most appropriate feature vector which characterizes the input tex- ture images by adjusting the P and e-parameter values of the related entropy functions. The proposed feature extraction mechanism is tested with several Brodatz texture images, which are shown in Fig. 4. A number of randomly selected tex- ture images of size 96 � 96 are formed. Thus, totally 1000 tex- ture images are processed in the feature extraction mechanism. For each of this 1000 texture image, the norm entropy, the sure entropy and the energy values of the each wavelet decomposition coefficients were obtained by using the feature extraction method. We calculated the norm entropy values of the texture images at the terminal node obtained from wavelet decomposition as defined in Eq. (1). Where, the wavelet entropy H is a real number, x is the termi- nal node signal and (xi) i is the waveform of terminal node. In norm entropy, P is the power and should be such that 1 6 P < 2. We also calculated the sure entropy and energy val- ues as defined in Eqs. (3), (4) and (5). In sure entropy, e is the threshold and must be such that 1 6 e 6 8. All obtained entropy values are normalized by dividing to N=1000. Thus, total normalized 6 entropy values and 3 energy values are found for each of these 10 texture images. 3.2.3. Step 3. Classification and calculation of the fitness function for an individual mechanism This mechanism is realized the intelligent classification by using features obtained from feature extraction mechanism. The training parameters and the structure of the MLP are shown in Ta- ble 3. These parameters are selected for ANN structure after several different experiments. In these experiments, the ANN is employed with different parameters such as the number of hidden layers, the size of the hidden layers, value of the moment constant and learn- ing rate, and type of the activation functions. The operations in this stage are ordered as below: 1. 200 � 9 feature vector which is obtained in feature extraction mechanism is given to input of the artificial neural network (ANN) classification. The decision space at the output of ANN classifier is formed from 10 texture images. 2. The mean square error (MSE) of the ANN is obtained at the final of training of the ANN classifier. 3. We call a constant desired error rate (DER) for the ANN classi- fier. The individuals are selected as winner if MSE is equal to DER or less than DER. If this status is occurred, this means that individual has high fitness value. Therefore, the comparison between MSE and DER is used for our genetic algorithm as fit- ness function of an individual of the population. In this applica- tion, fitness values of the all individuals of population are calculated in the same way. Later, obtained fitness values of all the individuals at the population are ordered from 1 to 20. If fitness value of an individual is less than 11, fitness value of this individual is low. The fitness values, which are higher than 11, are evaluated as high. The individuals which have high fit- ness values at the current population are saved for composing of the next generation. The individuals which have low fitness values at the current population are eliminated. If there are individuals which have same fitness value, one of these individ- ual is selected as random for population in next generation. As a result, optimum 10 individuals of current population are saved for next generation. 3.2.4. Step 4. Crossover operation The 30 % portion of the optimum 10 individuals obtained in Stage-4 are randomly selected and subjected to crossover operator. Namely, 3 individuals are subjected to crossover operator. 2 bits of the each of the random 2 individuals are randomly selected and Table 2 P-parameter value of the norm entropy and e-parameter value of the sure entropy P-parameter value of the norm entropy Binary equivalent of the P-parameter value of the norm entropy e-parameter value of the sure entropy Binary equivalent of the e-parameter value of the sure entropy 1 0 0 0 1 0 0 0 1.142 0 0 1 2 0 0 1 1.285 0 1 0 3 0 1 0 1.426 0 1 1 4 0 1 1 1.568 1 0 0 5 1 0 0 1.71 1 0 1 6 1 0 1 1.852 1 1 0 7 1 1 0 1.994 1 1 1 8 1 1 1 6040 E. Avci et al. / Expert Systems with Applications 36 (2009) 6036–6043 replaced each other for crossover operations. At the final of the crossover operations, 6 new individuals are obtained. Structures of feature extraction and classification mechanism of the GDWNN are shown in Fig. 5. 3.2.5. Step 5. Mutation operation In there, the bit inversion method is used for mutation operator (http://cs.felk.cvut.cz/~xobitko/ga/ (accessed: 20.8.04)). The muta- tion operation is realized by using the 0.3% portion of the total bits numbers of other seven individuals. If a bit is equal to 1, it is chan- ged to 0. If it is equal to 0, it is changed to 1. At the final of the mutation operation, seven new individuals are obtained. There are total 20 individuals in population of next generation at the final of Step 4 and Step 5. The processes in Steps 1–5 are realized respectively for each of the 20 individuals in population for a generation. Aim of using GDWNN in this study is the selecting the most appropriate feature extraction method from among four different feature extraction methods, optimum value of P-norm entropy parameter, and optimum value of e sure entropy parameter. The combination of selected these values are given to feature extrac- tion mechanism. 4. Experimental results Experiments are conducted with 10 Brodatz texture images, each of size 512 � 512 obtained from http://sipi.usc.edu/data- base/database.cgi?volume=textures&image=1#top. All texture images which are used in this study are shown in Fig. 4. We make 20 sub images of size 96 � 96 by randomly chosen from the each original input texture images. Thus, we obtain totally 200 images before feature extraction and training of the GDWNN. 100 texture images are used for test. Each of texture image pass through the process which is explained in Section 3.2. Each texture image is decomposed to 4 subbands according to the 1 level wavelet decomposition using various filters. The experimental results for proposed system are given in Table 4. Here, only the best 10 meth- od results, which the proposed algorithm was found, are given. We also compared the proposed system with several randomly chosen wavelet filters and entropy parameter values. These randomly cho- sen wavelet types, the entropy parameter values and the classifica- tion performance of the texture images were indicated in Table 5. Fig. 4. Texture images. Table 3 MLP architecture and training parameters The number of layers 3 The number of neuron on the layers Input: 9 Hidden: 15 Output: 1 The initial weights and biases Random Activation functions Tangent-sigmoid Training parameters Learning rule Levenberg-Marquart Back-propagation Sum-squared error 0.0000001 E. Avci et al. / Expert Systems with Applications 36 (2009) 6036–6043 6041 As it can be seen from Table 4, the overall classification rates for all texture types are over 90% except one method because the GDWNN could select the most appropriate parameters and the wavelet type. Population An individual in the population is randomly selected The Feature Extraction Method Represented by This Individual Value of the P Norm Entropy Parameter Represented by This Individual Value of the Sure Entropy Parameter Represented by This Individual ε Calculating of The Wavelet Entropies and Wavelet Energies 200 x 9 Feature Vectors Multi Layer Neural Network Classifier No If MSE <= DER Yes The individual is saved for population of the next generation Evaluating of Fitness Function of The GDWNN Fig. 5. Structure of feature extraction and classification mechanism of the GWNN. Table 4 The obtained optimum values by using GDWNN algorithm and classification performance of the GDWNN. Obtained optimum feature extraction method and the related parameters The correct classification rates for texture images (%) Method Wavelet Filter P e Brick wall Grass Gravel Bark Herringbone weave Plastic bubbles Rough wall Straw Woolen cloth Wood grain Total correct classification rates (%) FEM-14 coif2 1.142 8 78 72 99 100 100 100 88 95 100 100 93.2 FEM-3 db5 1.571 1 80 100 100 97 100 100 82 97 100 100 95.6 FEM-15 coif3 1.427 8 75 71 100 85 91 98 78 100 98 100 89.6 FEM-1 db2 1.142 3 78 96 100 100 98 100 71 96 98 100 93.7 FEM-13 coif1 1.713 5 71 100 100 98 82 100 86 86 91 100 91.4 FEM-10 bior2.2 1 5 82 88 100 92 82 100 100 98 100 100 94.2 FEM-15 coif3 1.285 6 78 100 100 82 100 98 96 98 100 100 95.2 FEM-2 db3 1 3 75 87 100 90 100 100 98 93 100 100 94.3 FEM-6 sym3 1.142 8 78 88 92 96 96 96 100 75 100 100 92.1 FEM-7 sym5 1.285 2 75 92 100 97 98 100 86 84 100 100 93.2 Table 5 The randomly chosen wavelet filters and entropy parameter values and their classification performance Randomly selected wavelet filter and parameter values The correct classification rates for texture images (%) Wavelet filter P e Brick wall Grass Gravel Bark Herringbone weave Plastic bubbles Rough wall Straw Woolen cloth Wood grain Total correct classification rates (%) bior2.2 1.1 8 74 80 96 100 93 98 86 90 100 100 82.7 db4 1.5 2 83 93 100 68 93 100 64 88 100 100 88.9 6042 E. Avci et al. / Expert Systems with Applications 36 (2009) 6036–6043 5. Discussions and conclusion This work indicates the use of GDWNN for texture image classi- fication. The feature choice was motivated by a realization that wavelet transform essentially is a representation of textures at a variety of resolutions. In brief, the wavelet decomposition has been demonstrated to be an effective tool for extracting information from the texture images. In this paper, the application of the wavelet entropy and energy in the wavelet layer of GDWNN to the feature extraction from tex- ture images was shown. Wavelet energy proved to be very useful features for characterizing the texture images; furthermore the information obtained from the wavelet entropy is related to the energy. The most important aspect of the intelligent system is the ability of self-organization of the GDWNN without require- ments of programming and the immediate response of a trained net during real-time applications. GDWNN finds optimum wavelet filter type and the optimum parameter values that are used for norm and sure entropy. The disadvantage of the GDWNN for tex- ture classification is that it needs more computation time than a standard MLP structure. Based on theoretical analysis and experimental results, the GDWNN features and classification is shown to be capable of cap- turing image structures which are crucial for texture characteriza- tion. It is shown that GDWNN can classify the given texture images with an overall 93.25% success rate. References Arivazhagan, S., & Ganesan, L. (2003). Texture classification using wavelet transform. Pattern Recognition Letters, 24, 1513–1521. Avci, E., Turkoglu, I., & Poyraz, M. (2005). Intelligent target recognition based on wavelet packet neural network. Experts Systems with Applications, 29(1). Chellappa, R., Chatterjee, S., & Bagdazian, R. (1985). Classification of textures using Gaussian Markov random fields. IEEE Transactions on Acoustics, Speech and Signal Processing, 30, 959–963. Chen, C. C., & Chen, C. C. (1999). Filtering methods for texture discrimination. Pattern Recognition Letters, 20, 783–790. Coifman, R. R., & Wickerhauser, M. V. (1992). Entropy-based algorithms for best basis selection. IEEE Transactions on Information Theory, 38(2), 713–718. Daubechies, I. (1988). Orthogonal bases of compactly supported wavelets. Communications on Pure and Applied Mathematics, 41, 909–996. Haykin, S. (1994). Neural networks. A comprehensive foundation. New York: Macmillan College Publishing Company Inc.. Huang, K., & Aviyente, S. (2006). Information-theoretic wavelet packet subband selection for texture classification. Signal Processing, 86(7), 1410–1420. Jain, A. K., & Farrokhnia, F. (1991). Unsupervised texture segmentation using Gabor filters. Pattern Recognition, 24, 1167–1186. Lei, Wang, & Jun, Liu. (1999). Texture classification using multiresolution Markov random field models. Pattern Recognition Letters, 20(2), 171–182. MATLAB (2003). 5.3 version Wavelet Toolbox, MathWorks Company. Mojsilovic, A., Rackov, D., & Popovic, M. (2000). On the selection of an optimal wavelet basis for texture characterization. IEEE Transactions on Image Processing, 9(12). Muneeswaran, K., Ganesan, L., Arumugam, S., & Ruba Soundar, K. (2005). Texture classification with combined rotation and scale invariant wavelet features. Pattern Recognition, 38(10), 1495–1506. Sengur, A., Turkoglu, I., Ince, M.C., in press. Wavelet packet neural networks for texture classification. Expert Systems with Applications, Uncorrected proof. Available online 13 January , 2006. Shutao, Li., James, T., Kwok Hailong, Zhu, & Yaonan, Wang (2003). Texture classification using the support vector machines. Pattern Recognition, 36(12), 2883–2893. Shutao, Li., & Shawe-Taylor, John (2005). Comparison and fusion of multiresolution features for texture classification. Pattern Recognition Letters, 26(5), 633–638. Tzanakou, E. M. (2000). Supervised and unsupervised pattern recognition feature extraction and computation. CRC Press LLC. Weszka, J. S., Dyer, C. R., & Rosenfeld, A. (1976). A comparative study of texture measures for terrain classification. IEEE Transactions on Systems, Man and Cybernetics, 6, 269–285. E. Avci et al. / Expert Systems with Applications 36 (2009) 6036–6043 6043 EXPERT SYSTEMS WITH APPLICATIONS An International Journal EDITOR-IN-CHIEF Jay Liebowitz, D.Sc. 966 Farm Haven Drive, Rockville, Maryland 20852, USA EDITORIAL ADVISORY BOARD Edward Feigenbaum Stanford University, USA Lance B. Eliot University of Southern California, USA Dianne C. Berry University of Reading, UK Walter Reitman Rensselaer Polytechnic Institute, USA Shri K. Goyal GTE Laboratories, USA Daniel E. O’Leary University of Southern California, USA Richard G. Vedder University of North Texas, USA James L. Rash NASA Goddard Space Flight Center, USA Randall Shumaker Naval Research Laboratory, USA David Bendel Hertz University of Miami, USA A. Desai Narasimhalu Singapore Management University Patricia Lightfoot NASA Goddard Space Flight Center, USA Mark Fox Toronto, Ontario, Canada Jae Kyu Lee Korea Advanced Institute of Science and Technology I. Burhan Turksen University of Toronto, Canada Ching Suen Concordia University, Canada Ernest H. Forman George Washington University, USA Roar Fjellheim Computas A.S., Norway Jan Vanthienen Catholic University, Leuven, Belgium Jon R. Wright AT&T Bell Laboratories, USA Roy Rada University of Maryland, USA John P. Coyne George Washington University, USA John H. Carson George Washington University, USA Daniel Schutzer Citicorp Investment Bank, USA Paul Harmon Harmon Associates, USA Francisco J. Cantu Instituto Tecnologico y de Estudios Superiores de Monterrey, Mexico Brian Gaines University of Calgary, Canada Michel Clerget Cognitech, France Kenneth J. Fordyce IBM, USA James R. Slagle University of Minnesota, USA Eliot Weinman Software Productivity, USA Randall Davis Massachusetts Institute of Technology, USA E. Balagurusamy Mahareer Academy of Technology and Sciences, India Dieter Specht Technische Universitat Cottbus, Germany Vladimir Milacic University of Belgrade, Yugoslavia Hans Bergkvist ATTEXOR S. A., Switzerland Laura C. Davis U.S. Navy Center for Applied Research in Artificial Intelligence, USA Jerald Feinstein MITRE, USA Efraim Turban University of Hawaii, 3435 Kehala Dr., Kiehi HI 96753, USA Bernt Bremdal GeoKnowledge, Norway Milton White Datanamics Inc., USA Editorial Office: Jay Liebowitz, 966 Farm Haven Drive, Rockville, Maryland 20852, USA. Tel./Fax: (+1) 301-770-2978; e-mail: jliebow1@jhu.edu Author enquiries: For enquiries relating to the submission of articles (including electronic submission where available) please visit this journal’s homepage at http://www.elsevier.com/ locate/eswa. You can track accepted articles at http://www.elsevier.com/trackarticle and set up e-mail alerts to inform you of when an article’s status has changed. Also accessible from here is information on copyright, frequently asked questions and more. Contact details for questions arising after acceptance of an article, especially those relating to proofs, are provided after registration of an article for publication. Publication information: Expert Systems with Applications (ISSN 0957-4174). For 2009, volumes 36, 37 are scheduled for publication. Subscription prices are available upon request from the Publisher or from the Regional Sales Office nearest you or from this journal’s website (http://www.elsevier.com/locate/eswa). Further information is available on this journal and other Elsevier products through Elsevier’s website: (http://www.elsevier.com). Subscriptions are accepted on a prepaid basis only and are entered on a calendar year basis. Issues are sent by standard mail (surface within Europe, air delivery outside Europe). Priority rates are available upon request. Claims for missing issues should be made within six months of the date of dispatch. Advertising information: Advertising orders and enquiries can be sent to: James Kenney, Elsevier Ltd., 32 Jamestown Road, London NW1 7BY, UK; phone: (+44) 207 424 4216; fax: (+44) 1865 853 136; e-mail: j.kenney@elsevier.com. Customers in the US and Canada can also contact: Mr Tino DeCarlo, Advertising Department, Elsevier Inc., 360 Park Avenue South, New York, NY 10010-1710, USA; phone: (+1) (212) 633 3815; fax: (+1) (212) 633 3820; e-mail: t.decarlo@elsevier.com. Orders, claims, and journal enquiries: please contact the Regional Sales Office nearest you: St. Louis: Elsevier, Customer Service Department, 11830 Westline Industrial Drive, St. Louis, MO 63146, USA; phone: (877) 8397126 [toll free within the USA]; (+1) (314) 4537076 [outside the USA]; fax: (+1) (314) 5235153; e-mail: JournalCustomerService-usa@elsevier.com Amsterdam: Elsevier, Customer Service Department, PO Box 211, 1000 AE Amsterdam, The Netherlands; phone: (+31) (20) 4853757; fax: (+31) (20) 4853432; e-mail: JournalsCustomer- ServiceEMEA@elsevier.com Tokyo: Elsevier, Customer Service Department, 4F Higashi-Azabu, 1-Chome Bldg, 1-9-15 Higashi-Azabu, Minato-ku, Tokyo 106-0044, Japan; phone: (+81) (3) 5561 5037; fax: (+81) (3) 5561 5047; e-mail: JournalsCustomerServiceJapan@elsevier.com Singapore: Elsevier, Customer Service Department, 3 Killiney Road, #08-01 Winsland House I, Singapore 239519; phone: (+65) 63490222; fax: (+65) 67331510; e-mail: JournalsCustomer- ServiceAPAC@elsevier.com USA mailing notice: Expert Systems with Applications (ISSN 0957-4174) is published 10 issues per annum in January, March, April, May, July, August, September, October, November, December by Elsevier Ltd. (The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, UK). Periodical postage paid at Rahway NJ and additional mailing offices. USA POSTMASTER: Send change of address: Expert Systems with Applications, Elsevier, Customer Service Department, 11830 Westline Industrial Drive, St. Louis, MO 63146, USA. Periodical postage paid at Rahway NJ and additional mailing offices. AIRFREIGHT AND MAILING in USA by Mercury International Limited, 365, Blair Road, Avenel, NJ 07001. 
work_bz6mzwgotffdppuuwaungqojhu	----	PDF hosted at the Radboud Repository of the Radboud University Nijmegen The following full text is a publisher's version. For additional information about this publication click this link. http://hdl.handle.net/2066/125438 Please be advised that this information was generated on 2021-04-06 and may be subject to change. http://hdl.handle.net/2066/125438 Expert systems for fetal assessment in labour (Protocol) Lutomski JE, Meaney S, Greene RA, Ryan AC, Devane D This is a reprint of a Cochrane protocol, prepared and maintained by The Cochrane Collaboration and published in The Cochrane Library 2013, Issue 8 http://www.thecochranelibrary.com Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. http://www.thecochranelibrary.com T A B L E O F C O N T E N T S 1HEADER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1BACKGROUND . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3OBJECTIVES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3METHODS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10CONTRIBUTIONS OF AUTHORS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10DECLARATIONS OF INTEREST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10SOURCES OF SUPPORT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iExpert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. [Intervention Protocol] Expert systems for fetal assessment in labour Jennifer E Lutomski1 , Sarah Meaney1 , Richard A Greene2, Anthony C Ryan3 , Declan Devane4 1National Perinatal Epidemiology Centre, Wilton, Ireland. 2Department of Obstetrics and Gynaecology, National Perinatal Epidemiol- ogy Centre, Wilton, Ireland. 3Neonatal Intensive Care Unit, Cork University Maternity Hospital, Wilton, Ireland. 4 School of Nursing and Midwifery, National University of Ireland Galway, Galway, Ireland Contact address: Jennifer E Lutomski, National Perinatal Epidemiology Centre, 5th Floor, Cork University Maternity Hospital, Wilton, Cork, Ireland. j.lutomski@ucc.ie. Editorial group: Cochrane Pregnancy and Childbirth Group. Publication status and date: New, published in Issue 8, 2013. Citation: Lutomski JE, Meaney S, Greene RA, Ryan AC, Devane D. Expert systems for fetal assessment in labour. Cochrane Database of Systematic Reviews 2013, Issue 8. Art. No.: CD010708. DOI: 10.1002/14651858.CD010708. Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. A B S T R A C T This is the protocol for a review and there is no abstract. The objectives are as follows: To evaluate the effectiveness of continuous or intermittent CTG monitoring during labour with an ES compared with (1) continuous or intermittent CTG monitoring during labour without an ES or (2) intermittent auscultation with a Pinard stethoscope or hand-held Doppler ultrasound device. B A C K G R O U N D Intrapartum fetal hypoxia (shortage of oxygen) is a serious com- plication of labour which increases the risk of perinatal mortal- ity and morbidity, including hypoxic ischaemic encephalopathy (acute or subacute brain injury due to asphyxia), cerebral palsy and developmental delay (Dilenge 2001; Fatemi 2009; McIntyre 2013). Whereas a healthy, term baby has a baseline heart rate be- tween 110 and 160 beats per minute, moderate variations in heart rate amplitude and natural accelerations from its baseline heart rate, this may not be the case for the hypoxic baby (ACOG 2009; Liston 2007; NICE 2007; RANZCOG 2006). In contrast, a hy- poxic baby may display minimal or a complete absence of heart rate variation and accelerations as well as have marked deceler- ations from its baseline heart rate values (ACOG 2009; Liston 2007; NICE 2007; RANZCOG 2006). Since abnormal fetal heart rate patterns may signify poor oxygenation and thus potential fe- tal compromise, maternal care professionals often perform fetal assessment during labour. Cardiotocography (CTG, also known as electronic fetal heart rate monitoring) and intermittent auscultation of the fetal heart rate are the most common forms of fetal assessment during labour (ACOG 2009; Liston 2007; NICE 2007; RANZCOG 2006). In brief, CTG records the fetal heart rate and maternal uterine ac- tivity via two transducers usually placed on the mother abdomen; an ultrasound transducer measures fetal heart rate and a pressure transducer measures uterine contractions. Fetal heart rate and uter- ine contractions are printed on a strip known as a CTG trace or presented directly on a visual display unit to help identify abnor- mal readings. In contrast, intermittent auscultation is a method of listening to the baby’s heart rate using a Pinard (fetal stetho- scope) or a hand-held Doppler ultrasound device. The rationale underpinning both forms of fetal assessment is that babies who are at risk of becoming compromised can be identified early so that appropriate recourse can be made (i.e. more intensive monitoring or expedited birth of the baby) (ACOG 2009; Liston 2007; NICE 2007; RANZCOG 2006). 1Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. mailto:j.lutomski@ucc.ie Description of the condition In the early half of the twentieth century, intermittent ausculta- tion served as the primary form of fetal assessment (Liston 2007). Technological advancements during this time period led to the de- velopment of CTG, and this method of fetal assessment was sub- sequently introduced into maternity care in the late 1960s (Chez 2011). Despite early beliefs that CTG could potentially supersede intermittent auscultation (Liston 2007), since its induction, there has been a growing professional debate on the efficacy of CTG monitoring (Alfirevic 2013; Chauhan 2008; Devane 2012; Hill 2012; Miller 2011a). Although CTG monitoring is typically re- served for high-risk pregnancies, in the many parts of the devel- oped world, more than three-quarters of babies will be assessed us- ing CTG (ACOG 2009; Devane 2007; Devane 2012). Yet, CTG has a relatively low specificity (a high false positive rate) for iden- tifying fetal hypoxia and associated complications (Miller 2011a). For example, the false positive rate for the most abnormal of fetal heart rate patterns, the detection of cerebral palsy, has been re- ported as high as 99.8% (Nelson 1996). This lack of specificity can be attributed, in part, to the variability in interpretation of fetal heart rate traces. Despite guidelines for CTG interpretation, substantial inter- and intra-observer variation in interpretation has been reported among maternity care providers (Blix 2003; Chauhan 2008; Devane 2005; Figueras 2005). Misinterpretation of fetal heart rate traces can lead to poor decisions, which can result in unnecessary intervention or delay or withholding of necessary intervention. Together, this can impact on the risks of operative birth, perinatal asphyxia, hypoxic ischaemic encephalopathy and perinatal death. Given these clinical implications, substantial research has investi- gated CTG monitoring in its current form versus alternative meth- ods. In a recent systematic review, women with signs of labour who received a short, 20-minute CTG tracing on admission to the maternity ward were more likely to have a caesarean birth than women who were monitored using intermittent auscultation (risk ratio (RR) 1.20; 95% confidence interval (CI) 1.00 to 1.44) (Devane 2012). A related review found that in comparison to in- termittent auscultation, continuous CTG during labour signifi- cantly decreased the risk of neonatal seizures (RR 0.50; 95% CI 0.31 to 0.80), though it did not decrease the risk of cerebral palsy (RR 1.75; 95% CI 0.84 to 3.63) (Alfirevic 2013). Further, con- tinuous CTG was associated with significantly higher rates of cae- sarean (RR 1.63; 95% CI 1.29 to 2.07) and instrumental births (RR 1.15; 95% CI 1.01 to 1.33). Thus, while CTG can provide critical information to maternity care providers, it is not without limitations. Description of the intervention First developed in the 1960s, expert systems (ESs) represent a type of applied artificial intelligence designed to assist in complex deci- sion-making (Liao 2005). In a healthcare context, ESs synthesise a computerised knowledge base derived from expert opinion with individual patient data to guide users towards possible diagnosis or treatment decisions (Liao 2005). To process data, an ES may ap- ply rule-based algorithms or neural networks (i.e. a model of pat- tern recognition based on previously collected data) (McCartney 2000; McCartney 2011); notably, however, there are numerous other ES methodologies (Liao 2005). Requirements for ESs vary; systems may be web-based or supported on a stand alone personal computer. ESs are paperless and often represent real-time, which is critical in healthcare environments where changes in health sta- tus can occur rapidly. Advancements in ESs have led to their inte- gration in a wide range of clinical settings, including cardiac care (Chi 2012; Seto 2012), cancer diagnosis (Issac Niwas 2012; Yang 2012) and diabetes diagnosis (Basciftci 2011; Picon 2012). The potential for ESs in maternity care is also well recognised, and as a result, there has been an increasing interest in developing ESs for CTG monitoring (Ayres-de-Campos 2010; Greene 1996). Such an ES would combine information on maternal and fetal characteristics, which may include fetal heart rate, electrocardio- gram (ECG) waveform, uterine contractions and/or gestational age, to provide a comprehensive overview of the labour and/or issue an alert if the baby’s status becomes critical. ES-issued alerts may not only advise the user on potentially suitable intervention but also provide underlying reasons for this recommendation, sim- ulating reasoning capacity that can be interpreted easily by the user (Liao 2005). How the intervention might work Despite controversy surrounding its ubiquitous use, CTG mon- itoring continues to be recommended by a number of profes- sional organisations for the monitoring of at least specific sub- groups of pregnant women (Miller 2011a). Yet, there is general consensus that measures must be taken to improve its interpreta- tion (Chez 2011). For this reason, CTG clinical definitions and guidelines have undergone extensive review (ACOG 2009; Liston 2007; NICE 2007; RANZCOG 2006) and training programmes have been recommended and evaluated (Pehrson 2011). Arguably, however, guidelines and educational interventions alone may be insufficient in reducing inter-observer differences (Devoe 2000). Although complementary measurements to enhance CTG inter- pretation, such as fetal pulse oximetry, lactate level measurements and fetal ECG waveform analysis, have been investigated, to date, these techniques have conferred limited advantages over conven- tional CTG monitoring (East 2007; East 2010; Neilson 2012). In particular, issues in interpretability were reported for ECG wave- form analysis, which may have impacted, in part, its utility as an additional indicator of fetal compromise (Steer 2008). Conse- quently, conventional CTG monitoring has remained the norm across many maternity units. 2Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. In this context, an ES for fetal assessment represents a potential alternative mechanism to counter issues with observer bias and improve interpretation of fetal heart rate tracings. Developing an ES to improve CTG monitoring has been a goal in maternity care for decades. Whereas earlier versions displayed only limited suc- cesses, substantial advances have been made in intelligence soft- ware (Devoe 2000; Greene 1996; Guijarro-Berdiñas 2002; Steer 2008). Moreover, several observational studies have reported sig- nificantly improved levels of agreement between practitioners in- terpreting fetal heart rate patterns and identifying adverse out- comes when assisted by an ES (Ayres-de-Campos 2010; Costa 2010; Costa 2010a). Thus, improving diagnostic interpretation may result in improved identification of truly compromised babies who warrant intervention. Why it is important to do this review Given that a high proportion of women will receive a CTG on their admission to the labour ward and/or will be monitored con- tinuously throughout their labour (ACOG 2009; Devane 2012), it is critical to ensure that these CTGs are interpreted to the high- est level of accuracy to minimise sub-optimal clinical decisions and outcomes. Improved CTG interpretation has clear short- and long-term health benefits for both mothers and babies through possible prevention of serious neonatal complications and unnec- essary operative birth. There are also important economic impli- cations for the healthcare system. Due to long-term care require- ments, neonatal cases of hypoxia and cerebral palsy can lead to costly malpractice suits (Miller 2011b); thus, reducing morbid- ity incidence through enhanced fetal assessment could potentially have profound impact in this regard. Further, reducing unneces- sary caesarean birth also benefits the healthcare system. Not only is caesarean birth typically twice the cost of a vaginal birth (Fawsitt 2013; Henderson 2001), but also, particularly in a nulliparous woman, may initiate a legacy of increased health costs (Grobman 2000) as a result of higher risk of subsequent caesarean birth and medical complications (Garmi 2012; Lydon-Rochelle 2010). O B J E C T I V E S To evaluate the effectiveness of continuous or intermittent CTG monitoring during labour with an ES compared with (1) contin- uous or intermittent CTG monitoring during labour without an ES or (2) intermittent auscultation with a Pinard stethoscope or hand-held Doppler ultrasound device. M E T H O D S Criteria for considering studies for this review Types of studies We will include all randomised, cluster-randomised and quasi-ran- domised trials comparing continuous or intermittent CTG mon- itoring during labour with an ES with continuous or intermit- tent CTG monitoring without an ES. We will also include trials that compare continuous or intermittent CTG monitoring during labour with an ES with intermittent auscultation with a Pinard or hand-held Doppler. Given the rapidity at which the health status of the mother and baby can change during labour, cross-over trials are unsuitable for the intervention under review and therefore will be excluded. Types of participants Pregnant women in labour and their babies. Types of interventions We will make the following comparisons. 1. Continuous CTG monitoring with an ES versus continuous CTG monitoring alone. 2. Continuous CTG monitoring with an ES versus intermittent CTG monitoring alone. 3. Continuous CTG monitoring with an ES versus intermittent auscultation with a Pinard stethoscope or hand-held Doppler ultrasound device. 4. Intermittent CTG monitoring with an ES versus intermittent CTG monitoring alone. 5. Intermittent CTG monitoring with an ES versus continuous CTG monitoring alone. 6. Intermittent CTG monitoring with an ES versus intermittent auscultation with a Pinard stethoscope or hand-held Doppler ultrasound device. For the purposes of this review, we define continuous CTG mon- itoring as monitoring of the fetal heart rate and uterine contrac- tions, which commenced at some point during labour, and upon commencement, was only discontinued for short periods of time. We define intermittent monitoring as monitoring of the fetal heart rate and uterine contractions which were not continuous but rather at select intervals during labour. We define an ES as any applied artificial intelligence tool designed to assist in complex decision-making by integrating a comput- erised knowledge base derived from expert opinion with individu- alised patient data (Liao 2005). Given that the underlying factors built into one ES may differ from others, where sufficient detail is provided by trial authors, we will describe which maternal and baby characteristics are included in the ES. 3Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. Types of outcome measures Primary outcomes Maternal 1. Incidence of caesarean birth. Baby 1. Incidence of perinatal mortality defined as fetal deaths (a baby delivered without signs of life at > 22 weeks’ gestation and/ or with a birthweight > 500 g) and neonatal deaths (death of a liveborn baby > 22 weeks’ gestation and/or with a birthweight > 500 g occurring within 28 days after birth) excluding lethal congential anomalies. 2. Incidence of neonatal seizures. 3. Acidemia as evidenced by a pH less than 7.0 and/or a base deficit greater than 12 mmol/L in umbilical arterial cord blood or neonatal blood sample within the first hour of life, or both. Secondary outcomes Maternal 1. Incidence of operative vaginal birth (ventouse or forceps). 2. Incidence of fetal blood sampling. 3. Incidence of artificial rupture of amniotic membranes. 4. Incidence of oxytocin augmentation of labour. Baby 1. Incidence of hypoxic ischaemic encephalopathy as defined by trial authors. 2. Incidence of admission to neonatal special care and/or neonatal intensive care unit. 3. Apgar score less than seven at five minutes. Search methods for identification of studies Electronic searches We will contact the Trials Search Co-ordinator to search the Cochrane Pregnancy and Childbirth Group’s Trials Register. The Cochrane Pregnancy and Childbirth Group’s Trials Register is maintained by the Trials Search Co-ordinator and contains trials identified from: 1. monthly searches of the Cochrane Central Register of Controlled Trials (CENTRAL); 2. weekly searches of MEDLINE; 3. weekly searches of Embase; 4. handsearches of 30 journals and the proceedings of major conferences; 5. weekly current awareness alerts for a further 44 journals plus monthly BioMed Central email alerts. Details of the search strategies for CENTRAL, MEDLINE and EMBASE, the list of handsearched journals and conference pro- ceedings, and the list of journals reviewed via the current aware- ness service can be found in the ‘Specialized Register’ section within the editorial information about the Cochrane Pregnancy and Childbirth Group. Trials identified through the searching activities described above are each assigned to a review topic (or topics). The Trials Search Co-ordinator searches the register for each review using the topic list rather than keywords. For the purposes of this review, we will also include studies pub- lished as abstracts if we are able to extract sufficient information on the trial. In cases where insufficient data are published, we will first contact the trial authors to access required information. If after contacting the trial authors data remain insufficient, the abstract will be excluded from our review. In addition, we plan to search grey literature through Open Grey and ProQuest Dissertation & Theses Database using the search terms given in Appendix 1 Searching other resources We will review citations of reference lists of included papers iden- tified through the above search strategy and assess their suitability for inclusion in the review. We will not apply any language restrictions. Data collection and analysis The methods that will be carried out in this review were designed in accordance with recommendations described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). Selection of studies Two review authors (Jennifer E Lutomski (JEL) and Declan De- vane (DD)) will independently assess for inclusion all the poten- tial studies we identify as a result of the search strategy. We will resolve any disagreement through discussion or, if required, we will consult a third review author (CA Ryan (CAR), Sarah Meaney (SM) or Richard A Greene (RAG)). Data extraction and management We will design a form to extract data. For eligible studies, two review authors (JEL and SM) will extract the data using the agreed 4Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. http://www.mrw.interscience.wiley.com/cochrane/clabout/articles/PREG/frame.html http://www.mrw.interscience.wiley.com/cochrane/clabout/articles/PREG/frame.html http://www.mrw.interscience.wiley.com/cochrane/clabout/articles/PREG/frame.html http://www.mrw.interscience.wiley.com/cochrane/clabout/articles/PREG/frame.html http://www.mrw.interscience.wiley.com/cochrane/clabout/articles/PREG/frame.html http://www.opengrey.eu/ http://www.opengrey.eu/ form. We will resolve discrepancies through discussion or, if re- quired, we will consult a third review author (DD). We will enter data into Review Manager software (RevMan 2011) and check for accuracy. When information regarding any of the above is unclear, we will attempt to contact authors of the original reports to pro- vide further details. Assessment of risk of bias in included studies Two review authors (JEL and DD) will independently assess risk of bias for each study using the criteria outlined in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). We will resolve any disagreement by discussion or by involving a third review author. (1) Random sequence generation (checking for possible selection bias) We will describe for each included study the method used to gen- erate the allocation sequence in sufficient detail to allow an assess- ment of whether it should produce comparable groups. We will assess the risk of bias for sequence generation as: • low risk (any truly random process, e.g. random number table; computer random number generator); • high risk (any non-random process, e.g. odd or even date of birth; hospital or clinic record number); or • unclear risk. (2) Allocation concealment (checking for possible selection bias) We will describe for each included study the method used to con- ceal allocation to interventions prior to assignment and will assess whether intervention allocation could have been foreseen in ad- vance of, or during recruitment, or changed after assignment. We will assess the risk of bias for allocation concealment as: • low risk (e.g. telephone or central randomisation; consecutively numbered sealed opaque envelopes); • high risk (e.g. open random allocation; unsealed or non- opaque envelopes, alternation; date of birth); or • unclear risk. (3.1) Blinding of participants and personnel (checking for possible performance bias) Given the additional equipment required to support an ES, it is not possible for personnel to be blinded to the intervention in these trials. However, we will describe the methods used, if any, to blind women from knowledge of the intervention. We will consider that studies are at low risk of bias if women were blinded or if we judge that the lack of blinding would not likely impact results. We will assess the risk of bias for blinding of participants as: • low risk; • high risk; or • unclear risk. (3.2) Blinding of outcome assessment (checking for possible detection bias) We will describe for each included study the methods used, if any, to blind outcome assessors from knowledge of which intervention a woman received. We will assess blinding separately for different outcomes or classes of outcomes. We will assess the risk of bias for blinding of outcome assessment as: • low risk; • high risk; or • unclear risk. (4) Incomplete outcome data (checking for possible attrition bias due to the amount, nature and handling of incomplete outcome data) We will describe for each included study, and for each outcome or class of outcomes, the completeness of data including attrition and exclusions from the analysis. We will state whether attrition and exclusions were reported and the numbers included in the analysis at each stage (compared with the total randomised par- ticipants), reasons for attrition or exclusion where reported, and whether missing data were balanced across groups or were related to outcomes. Where sufficient information is reported, or can be supplied by the trial authors, we will re-include missing data in the analyses which we undertake. We will assess the risk of bias for incomplete outcome data as: • low risk (e.g. no missing outcome data; missing outcome data balanced across groups; < 20% missing data); • high risk (e.g. frequency or reasons of missing data imbalanced across groups; > 20% missing data); or • unclear risk. (5) Selective reporting (checking for reporting bias) We will investigate the possibility of selective outcome reporting bias by cross-checking outcomes of interest reported in the meth- ods section to those reported in the results section of the trial pub- lications. We will assess the risk of bias for selective reporting as: • low risk (where it is clear that all of the study’s pre-specified outcomes and all expected outcomes of interest to the review have been reported); • high risk (where not all the study’s pre-specified outcomes have been reported; one or more reported primary outcomes were not pre-specified; outcomes of interest are reported incompletely and so cannot be used; study fails to include results of a key outcome that would have been expected to have been reported); or • unclear risk. 5Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. (6) Other bias (checking for bias due to problems not covered by (1) to (5) above) We will describe for each included study any important concerns we have about other possible sources of bias. For instance, given that certain biases are inherent to cluster-randomised trials due to study design, in these types of trials we will investigate potential biases due to recruitment, baseline imbalance, loss of clusters, in- correct analysis and comparability with individually-randomised trials. We will assess the risk of other forms of bias as: • low risk; • high risk; or • unclear risk. (7) Overall risk of bias We will make explicit judgements about whether studies are at high risk of bias, according to the criteria given in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). With reference to (1) to (6) above, we will assess the likely magni- tude and direction of the bias and whether we consider it is likely to impact on the findings. We will explore the impact of the level of bias through undertaking sensitivity analyses (see Sensitivity analysis) and will assess the overall risk of bias for each included study as: • low risk; • high risk; or • unclear risk. Measures of treatment effect Dichotomous data For dichotomous data, we will present results as summary risk ratios with 95% confidence intervals (95% CI). Continuous data For continuous data, we will use the mean difference with 95% CI if outcomes are measured in the same way between trials. We will use the standardised mean difference with 95% CI to com- bine outcomes from trials that measure the same outcome but use different scales (Higgins 2011). Unit of analysis issues Cluster-randomised trials We will include cluster-randomised trials in the analyses along with individually-randomised trials. We will adjust their sample sizes using the methods described in the Cochrane Handbook for Sys- tematic Reviews of Interventions (Higgins 2011) using an estimate of the intra-cluster correlation co-efficient (ICC) derived from the trial (if possible), from a similar trial or from a study of a similar population. If we use ICCs from other sources, we will report this and conduct sensitivity analyses to investigate the effect of varia- tion in the ICC. If we identify both cluster-randomised trials and individually-randomised trials, we plan to synthesise the relevant information. We will consider it reasonable to combine the re- sults from both if there is little heterogeneity between the study designs and the interaction between the effect of intervention and the choice of randomisation unit is considered to be unlikely. We will also acknowledge heterogeneity in the randomisation unit and perform a sensitivity analysis to investigate the effects of the randomisation unit, i.e. to determine the sensitivity of the effect estimates to inclusion and exclusion of cluster trials. Multi-armed trials We will include multi-armed trials in this review. To overcome po- tential issues due to multiple, correlated comparisons, we will anal- yse multi-armed trials using methods described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). If feasible in the context of the research question, we will first de- termine if we can combine the multiple comparison groups to cre- ate one relevant intervention group and one relevant comparison group. If an appropriate pair-wise comparison cannot be created, we will then derive an average (or weighted average) and variance for all relevant intervention/comparison groups while accounting for correlation between these comparisons. Dealing with missing data For included studies, we will note levels of attrition. We will explore the impact of including studies with high levels of missing data (we judge this a priori to be greater than 20% for primary outcome) in the overall assessment of treatment effect by using sensitivity analysis. For all outcomes, we will carry out analyses, as far as possible, on an intention-to-treat basis, i.e. we will attempt to include all women randomised to each group in the analyses, and all women will be analysed in the group to which they were allocated, regardless of whether or not they received the allocated intervention. The denominator for each outcome in each trial will be the number of women randomised minus any women whose outcomes are known to be missing. Assessment of heterogeneity Where studies are considered similar enough (based on consider- ation of populations and interventions) to allow pooling of data using meta-analysis, we will assess the degree of heterogeneity by visual inspection of forest plots and by examining the Chi² test for heterogeneity. We will assess statistical heterogeneity in each meta-analysis using the T², I² and Chi² statistics. We will regard 6Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. statistical heterogeneity as substantial if an I² is greater than 30% and either the T² is greater than zero, or there is a low P value (less than 0.10) in the Chi² test for heterogeneity. Where we identify substantial clinical, methodological or statisti- cal heterogeneity across included studies, we will not report pooled results from the meta-analysis but will instead use a narrative ap- proach to data synthesis. In this event, we will attempt to explore possible reasons for the heterogeneity by grouping studies that have similar populations and interventions. Assessment of reporting biases If there are 10 or more studies in the meta-analysis, we will in- vestigate reporting biases (such as publication bias) using funnel plots. We will assess funnel plot asymmetry visually. If asymmetry is suggested by a visual assessment, we will perform exploratory analyses to investigate it. Data synthesis We will carry out statistical analysis using the Review Manager software (RevMan 2011). We will use fixed-effect meta-analysis for combining data where it is reasonable to assume that studies are estimating the same underlying treatment effect: i.e. where trials are examining the same intervention, and the trials’ populations and methods are judged sufficiently similar. If there is clinical het- erogeneity sufficient to expect that the underlying treatment ef- fects differ between trials, or if substantial statistical heterogeneity is detected, we will use random-effects meta-analysis to produce an overall summary if an average treatment effect across trials is considered clinically meaningful. The random-effects summary will be treated as the average range of possible treatment effects and we will discuss the clinical implications of treatment effects differing between trials. If the average treatment effect is not clin- ically meaningful, we will not combine trials. If we use random-effects analyses, the results will be presented as the average treatment effect with 95% CIs and the estimates of T² and I². Subgroup analysis and investigation of heterogeneity If we identify substantial heterogeneity, we will investigate this heterogeneity using subgroup analyses and sensitivity analyses. We plan to carry out the following subgroup analysis. 1. Low-risk pregnancies versus high-risk pregnancies as defined by the trial authors. We will limit subgroup analyses to primary maternal and baby outcomes (see Types of outcome measures). We will assess subgroup differences by interaction tests available within RevMan (RevMan 2011). We will report the results of subgroup analyses quoting the Chi2 statistic and P value, and the interaction test I² value. Sensitivity analysis We will perform sensitivity analyses based on trial quality by re- peating our analysis among only those trials judged of ’high qual- ity’. For the purposes of this review, ’high quality’ trials will be defined as trials with low risk of bias due to allocation concealment and low risk of bias due to incomplete outcome data. We will limit sensitivity analyses to primary maternal and baby outcomes (see Types of outcome measures). Sensitivity analyses will also assist in investigating substantial statistical heterogeneity if present (see Assessment of heterogeneity). A C K N O W L E D G E M E N T S We would like to acknowledge Lynn Hampson (Trials Search Co-ordinator), Frances Kellie (Managing Editor) and Jim Neil- son (Contact Editor) of the Cochrane and Pregnancy Childbirth Group for their support throughout the drafting of this protocol. JEL has been awarded a HRB Cochrane Review Training Fellow- ship from the Health Research Board (HRB), Ireland. As part of the pre-publication editorial process, this protocol has been commented on by three peers (an editor and two referees who are external to the editorial team) and the Group’s Statistical Adviser. The National Institute for Health Research (NIHR) is the largest single funder of the Cochrane Pregnancy and Childbirth Group. The views and opinions expressed therein are those of the authors and do not necessarily reflect those of the NIHR, NHS or the Department of Health. 7Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. R E F E R E N C E S Additional references ACOG 2009 ACOG Practice Bulletin No. 106. Intrapartum fetal heart rate monitoring: Nomenclature, interpretation, and general management principles. Obstetrics and Gynecology 2009; 114(1):192–202. Alfirevic 2013 Alfirevic Z, Devane D, Gyte GML. Continuous cardiotocography (CTG) as a form of electronic fetal monitoring (EFM) for fetal assessment during labour. Cochrane Database of Systematic Reviews 2013, Issue 5. [DOI: 10.1002/14651858.CD006066.pub2] Ayres-de-Campos 2010 Ayres-de-Campos D, Ugwumadu A, Banfield P, Lynch P, Amin P, Horwell D, et al.A randomised clinical trial of intrapartum fetal monitoring with computer analysis and alerts versus previously available monitoring. BMC Pregnancy Childbirth 2010;10:71. [DOI: 10.1186/ 1471-2393-10-71] Basciftci 2011 Basciftci F, Hatay OF. Reduced-rule based expert system by the simplification of logic functions for the diagnosis of diabetes. Computers in Biology and Medicine 2011;41(6): 350–6. Blix 2003 Blix E, Sviggum O, Koss KS, Oian P. Inter-observer variation in assessment of 845 labour admission tests: comparison between midwives and obstetricians in the clinical setting and two experts. BJOG: an international journal of obstetrics and gynaecology 2003;110(1):1–5. Chauhan 2008 Chauhan SP, Klauser CK, Woodring TC, Sanderson M, Magann EF, Morrison JC. Intrapartum nonreassuring fetal heart rate tracing and prediction of adverse outcomes: interobserver variability. American Journal of Obstetrics and Gynecology 2008;199(6):623 e1-5. Chez 2011 Chez BF, Baird SM. Electronic fetal heart rate monitoring: where are we now?. Journal of Perinatal and Neonatal Nursing 2011;25(2):180-92; quiz 93-4. Chi 2012 Chi CL, Nick Street W, Robinson JG, Crawford MA. Individualized patient-centered lifestyle recommendations: an expert system for communicating patient specific cardiovascular risk information and prioritizing lifestyle options. Journal of Biomedical Informatics 2012;45(6): 1164–74. Costa 2010 Costa A, Santos C, Ayres-de-Campos D, Costa C, Bernardes J. Access to computerised analysis of intrapartum cardiotocographs improves clinicians’ prediction of newborn umbilical artery blood pH. BJOG: an international journal of obstetrics and gynaecology 2010;117:1288–93. Costa 2010a Costa MA, Ayres-de-Campos D, Machado AP, Santos C, Bernardes J. Comparison of a computer system evaluation of intrapartum cardiotocographic events and consensus of clinicians. Journal of Perinatal Medicine 2010;38:191–5. Devane 2005 Devane D, Lalor J. Midwives’ visual interpretation of intrapartum cardiotocographs: intra- and inter-observer agreement. Journal of Advanced Nursing 2005;52(2): 133–41. Devane 2007 Devane D, Lalor J, Bonnar J. The use of intrapartum electronic fetal heart rate monitoring: a national survey. Irish Medical Journal 2007;100(2):360–2. Devane 2012 Devane D, Lalor JG, Daly S, McGuire W, Smith V. Cardiotocography versus intermittent auscultation of fetal heart on admission to labour ward for assessment of fetal wellbeing. Cochrane Database of Systematic Reviews 2012, Issue 2. [DOI: 10.1002/14651858.CD005122.pub4] Devoe 2000 Devoe L, Golde S, Kilman Y, Morton D, Shea K, Waller J. A comparison of visual analyses of intrapartum fetal heart rate tracings according to the new national institute of child health and human development guidelines with computer analyses by an automated fetal heart rate monitoring system. American Journal of Obstetrics and Gynecology 2000;183(2): 361–6. Dilenge 2001 Dilenge ME, Majnemer A, Shevell MI. Long-term developmental outcome of asphyxiated term neonates. Journal of Child Neurology 2001;16:781–92. East 2007 East CE, Begg L, Colditz PB. Fetal pulse oximetry for fetal assessment in labour. Cochrane Database of Systematic Reviews 2007, Issue 2. Art. No.: CD004075. DOI:. [DOI: 10.1002/14651858.CD004075.pub3] East 2010 East CE, Leader LR, Sheehan P, Henshall NE, Colditz PB. Intrapartum fetal scalp lactate sampling for fetal assessment in the presence of a non-reassuring fetal heart rate trace. Cochrane Database of Systematic Reviews 2010, Issue 3. [DOI: 10.1002/14651858.CD006174.pub2] Fatemi 2009 Fatemi A, Wilson MA, Johnston MV. Hypoxic-ischemic encephalopathy in the term infant. Clinics in Perinatology 2009;36(4):835–58. Fawsitt 2013 Fawsitt CG, Bourke J, Greene RA, Everard CM, Murphy A, Lutomski JE. At what price? A cost-effectiveness analysis comparing trial of labour after previous caesarean versus elective repeat caesarean delivery. PLoS One 2013;8(3): e58577. DOI: 10.1371/journal.pone.0058577. 8Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. Figueras 2005 Figueras F, Albela S, Bonino S, Palacio M, Barrau E, Hernandez S, et al.Visual analysis of antepartum fetal heart rate tracings: inter- and intra-observer agreement and impact of knowledge of neonatal outcome. Journal of Perinatal Medicine 2005;33(3):241–5. Garmi 2012 Garmi G, Salim R. Epidemiology, etiology, diagnosis, and management of placenta accreta. Obstetrics and Gynecology International 2012;2012:873929. [DOI: 10.1155/2012/ 873929] Greene 1996 Greene KR. Intelligent fetal heart rate computer systems in intrapartum surveillance. Current Opinion in Obstetrics and Gynecology 1996;8(2):123–7. Grobman 2000 Grobman WA, Peaceman AM, Socol ML. Cost-effectiveness of elective cesarean delivery after one prior low transverse cesarean. Obstetrics and Gynecology 2000;95(5):745–51. Guijarro-Berdiñas 2002 Guijarro-Berdiñas B, Alonso-Betanzos A, Fontenla- Romero O. Intelligent analysis and pattern recognition in cardiotocographic signals using a tightly coupled hybrid system. Artificial Intelligence 2002;136:1–27. Henderson 2001 Henderson J, McCandlish R, Kumiega L, Petrou S. Systematic review of economic aspects of alternative modes of delivery. BJOG: an international journal of obstetrics and gynaecology 2001;108(2):149–57. Higgins 2011 Higgins JPT, Green S, editors. Cochrane Handbook for Systematic Reviews of Interventions Version 5.1.0 [updated March 2011]. The Cochrane Collaboration, 2011. Available from www.cochrane-handbook.org. Hill 2012 Hill JB, Chauhan SP, Magann EF, Morrison JC, Abuhamad AZ. Intrapartum fetal surveillance: review of three national guidelines. American Journal of Perinatology 2012;29(7): 539–50. Issac Niwas 2012 Issac Niwas S, Palanisamy P, Chibbar R, Zhang WJ. An expert support system for breast cancer diagnosis using color wavelet features. Journal of Medical Systems 2012;36 (5):3091–102. Liao 2005 Liao SH. Expert system methodologies and applications - a decade review from 1995 to 2004. Expert Systems with Applications 2005;28(1):93–103. Liston 2007 Liston R, Sawchuck D, Young D, Society of Obstetrics and Gynaecologists of Canada, British Columbia Perinatal Health Program. Fetal Health Surveillance: Antepartum and intrapartum consensus guideline No. 197. Journal of Obstetrics and Gynaecology Canada: JOGC 2007;29(9 Suppl 4):S3–S56. Lydon-Rochelle 2010 Lydon-Rochelle MT, Cahill AG, Spong CY. Birth after previous cesarean delivery: short-term maternal outcomes. Seminars in Perinatology 2010 Aug;34(4):249–57. McCartney 2000 McCartney PR. Computer analysis of the fetal heart rate. Journal of Obstetric, Gynecologic, and Neonatal Nursing 2000; 29(5):527–36. McCartney 2011 McCartney PR. Computer fetal heart rate pattern analysis. MCN; American Journal of Maternal Child Nursing 2011;36 (6):397. McIntyre 2013 McIntyre S, Taitz D, Keogh J, Goldsmith S, Badawi N, Blair E. A systematic review of risk factors for cerebral palsy in children born at term in developed countries. Developmental Medicine and Child Neurology 2013;55(6): 499–508. Miller 2011a Miller DA. Intrapartum fetal heart rate definitions and interpretation: evolving consensus. Clinical Obstetrics and Gynecology 2011;54(1):16–21. Miller 2011b Miller L. Intrapartum fetal monitoring: liability and documentation. Clinical Obstetrics and Gynecology 2011;54 (1):50–5. Neilson 2012 Neilson JP. Fetal electrocardiogram (ECG) for fetal monitoring during labour. Cochrane Database of Systematic Reviews 2012, Issue 4. [DOI: 10.1002/ 14651858.CD000116.pub3] Nelson 1996 Nelson KB, Dambrosia JM, Ting TY, Grether JK. Uncertain value of electronic fetal monitoring in predicting cerebral palsy. New England Journal of Medicine 1996;334(10): 613–8. NICE 2007 National Institute for Health and Clinical Excellence. CG55 Intrapartum care: Care of healthy women and their babies during childbirth. NICE Clinical Guidelines. UK 2007. Pehrson 2011 Pehrson C, Sorensen JL, Amer-Wahlin I. Evaluation and impact of cardiotocography training programmes: a systematic review. BJOG: an international journal of obstetrics and gynaecology 2011;118(8):926–36. Picon 2012 Picon AP, Ortega NR, Watari R, Sartor C, Sacco IC. Classification of the severity of diabetic neuropathy: a new approach taking uncertainties into account using fuzzy logic. Clinics (Sao Paulo) 2012;67(2):151–6. RANZCOG 2006 Royal Australian and New Zealand College of Obstetricians and Gynaecologists. Clinical Guidelines: Intrapartum 9Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. Fetal Surveillance. 2nd ed. East Melbourne, Australia: RANZCOG. 2nd Ed. 2006. RevMan 2011 The Nordic Cochrane Centre, The Cochrane Collaboration. Review Manager (RevMan). 5.1. Copenhagen: The Nordic Cochrane Centre, The Cochrane Collaboration, 2011. Seto 2012 Seto E, Leonard KJ, Cafazzo JA, Barnsley J, Masino C, Ross HJ. Developing healthcare rule-based expert systems: case study of a heart failure telemonitoring system. International Journal of Medical Informatics 2012;81(8):556–65. Steer 2008 Steer PJ. Has electronic fetal heart rate monitoring made a difference?. Seminars in Fetal and Neonatal Medicine 2008; 13(1):2–7. Yang 2012 Yang F, Tian J, Xiang Y, Zhang Z, Harrington Pde B. Near infrared spectroscopy combined with least squares support vector machines and fuzzy rule-building expert system applied to diagnosis of endometrial carcinoma. Cancer Epidemiology 2012;36(3):317–23. ∗ Indicates the major publication for the study A P P E N D I C E S Appendix 1. Search terms for Open Grey and ProQuest Open Grey medical equipment AND (fetal OR foetal OR birth OR labour OR labor OR childbirth) expert systems AND (fetal OR foetal OR birth OR childbirth OR labor OR labour) cardiotocograph OR cardiotocography fetal monitoring OR foetal monitoring (We will search for each line separately and assess results of each). ProQuest Dissertations and Theses all(fetal OR foetal OR birth OR childbirth OR obstetrics) AND all((electronic NEAR/1 monitoring OR expert NEAR/2 systems OR intelligent NEAR/2 analysis OR computer NEAR/2 analysis OR cardiotocography OR cardiotocograph)) C O N T R I B U T I O N S O F A U T H O R S JEL drafted the initial protocol and is guarantor for the review. DD supervised the review process, critically assessed the methodology and drafted part of the protocol. RAG and CAR provided a clinical perspective; SM provided a consumer perspective. JEL, DD, CAR, SM and RAG revised the protocol for intellectual content and approved its submission for publication. D E C L A R A T I O N S O F I N T E R E S T None known. 10Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. S O U R C E S O F S U P P O R T Internal sources • No sources of support supplied External sources • The Health Research Board, Ireland. Cochrane Review Training Fellowship 11Expert systems for fetal assessment in labour (Protocol) Copyright © 2013 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd. 
work_bzykyau3grhq7dsab54mppmy5e	----	PII: 0743-1066(90)90055-A J. LOGIC PROGRAMMING 1990:X:145-1X5 145 AN EXPERT SYSTEM FOR HARMONIZING CHORALES IN THE STYLE OF J. S. BACH* KEMAL EBCiOS;LU D This paper describes an expert system called CHORAL, for harmonization of four-part chorales in the style of Johann Sebastian Bach. The system contains about 350 rules, written in a form of first-order predicate calculus. The rules represent musical knowledge from multiple viewpoints of the chorale, such as the chord skeleton, the melodic lines of the individual parts, and Schenkerian voice leading within the descant and bass. The program harmonizes chorale melodies using a generate-and-test method with intelligent backtracking. A substantial number of heuristics are used for biasing the search toward musical solutions. The CHORAL knowledge base provides for style-specific modulations, cadence patterns. and complex encounters of simultaneous inessential notes; it imposes difficult constraints for maintaining melodic interest in the inner voices. Encouraging results have been obtained, and output examples are given. BSL, a new and efficient logic-programming language fundamentally different from PROLOG. was designed to implement the CHORAL system. a INTRODUCTION In this paper, we will describe CHORAL, a knowledge-based expert system for harmonization and hierarchical voice-leading analysis of chorales in the style of J. S. Bach. We will first briefly outline a logic-programming language called BSL that was designed to implement the project, and then describe the CHORAL system itself. The research that we are about to report covers vast and highly complex areas in both artificial intelligence and music, so we will strive to use a language as comprehensi- ble as possible. ,4r/dws.v ~ww\po~drnce IO Kcmal Ebcioglu, IBM T. J. Watson Research Center. P.O. Box 21X. Yorktown Heights. NY 10598. Rcceivcd May 1987; accepted May 1988. *This paper is based on the author’s Ph.D. dissertation “An Expert System for Harmonization of Chorales in the Style of J. S. Bach”. Technical Report X6-09. Dept. of Computer Science. S.U.N.Y. at BuRjlo. Mar. 1986, and on about a year of additional work done at IBM Research. This research was supported by NSF grant DCR-83 16665. and the maJor portion of it was done at S.U.N.Y. at BuH'alo. under rhc direction of the author’s advisor John Myhill. 1 E,lw\icr Science Publishing Co.. Inc.. 1990 655 Awnuc of the Americas. New York, NY 10010 0743.1066/YO/$3.50 146 KEMALEBCIO(:ILU BSL: AN EFFICIENT LOGIC-PROGRAMMING LANGUAGE At the outset of our CHORAL project, we considered a number of alternative knowledge representation techniques, and we finally decided to use first-order logic for representing musical knowledge. First-order logic was felt to be well suited to the application, because it allowed us to make precise, concrete assertions about properties of a piece of music, and because it was more formal and tractable than some other AI paradigms, such as unrestricted production systems [27]. We started with over a hundred assertions in first-order predicate calculus, which later formed the seed of the knowledge base. These assertions were not in clausal form, and made free use e.g. of existential and universal quantifiers, as in the assertions one would use to extend English in a formal treatment such as [64]. However, the PROLOG interpreter then available to us on the VAX 11 architecture did not have a natural way of coding quantifiers; moreover, it did not offer the most efficient way for utilizing the native resources of a traditional CPU. On the other hand, our music application was well suited to the native data types and operations of a traditional architecture, and was also known to be extremely computation-intensive (we did have a fair idea of the potential problems of the application, because we had previously written a smaller-scale 16th-century strict-counterpoint program using a similar heuristic search method [17,18]). We were thus led to look for a different logic-programming language for implementing our project. Our requirements were: (1) the language had to have a natural way of coding universal and existential quantifiers directly; (2) the language had to utilize the native resources of a traditional architecture efficiently, in a manner competitive with deterministic Algol-class languages, so that we could use it to produce very high-quality music in a reasonable time; (3) the language had to have a natural way of specifying preferred solutions as well as just correct ones (the musical importance of this will be explained in the sequel); (4) the language had to have a streamlined design in order to increase its chances of being theoretically tractable; moreover, we felt that striving to use a streamlined design was a better way to approach a large project. While we were going back and forth between the logical assertions and ways of “executing” them, a logic-programming language called BSL (Backtracking specifica- tion Language) was designed, which appears to satisfy each of the abovementioned requirements. The design of the BSL language constitutes an unusual approach to the use of logic in computer programming, but is extremely traditional in the sense of the execution paradigm. Unlike languages such as PROLOG, BSL is not a descendant of resolution theorem-proving research [63]: BSL is merely a nondeterministic language with Pascal-style data types, where more than one explicit assignment to a variable is forbidden. BSL has a Lisp-like syntax and is compiled into c via a Lisp program. We have provided BSL with formal semantics, in a style inspired by [lo] and [30]. The semantics of a BSL program F is defined via a ternary relation ‘k, such that 9( F, u, a’) means program F leads to final state u’ when started in initial state u, where a state is a mapping from variable names to elements of a “computer” universe, consisting of integers, arrays, records, and other ancillary objects. Given an initial state, a BSL program may lead to more than one final state, since it is nondeterministic, or it may lead to none at all, in case it never terminates. What makes BSL different from ordinary nondeterministic languages [26,70,7], and relates it to logic, is that there is a simple mapping that translates a BSL program to a AN EXPERT SYSTEM FOR HARMONIZING CHORALES 147 formula of a first-order language, such that if a BSL program terminates in some state u, then the corresponding first-order formula is true in u [where the truth of a formula in a given state (I is evaluated in a fixed “computer” interpretation after replacing any free variables x in the formula by a(x)]. A BSL program is very similar in appearance to the corresponding first-order formula, and for this reason, we call BSL programs formulas. In this paper, we will only give an informal overview of the BSL language. A formal description of BSL and a proof of its soundness can be found in [19]. We will start with an example of a simple BSL program to solve a little puzzle, followed by its first-order translation: Place eight queens on a chess board, so that no queen attacks another (i.e. no two queens are on the same row, column, or diagonal). Assume that the rows and columns are numbered from 0 to 7, and that the array elements p[O], . . . , p[7] represent the column number of the queen in row 0,. . . ,7, respectively: (include stdmac) ;include standard macro definitions (options registers (k j n)) ;allocate kj,n in registers (E ((P (array (8) integer))) (AnO(< n8)(1+ n) (EjO(< j@(L+j) (and(AkO(<kn)(l+ k) (and(! = j (p k)) (! = (- j (P k)) (- n k)) (! = (- j (P k)) (- k n)))) ( := (P n) j))))) First-order translation: (3p 1 type(p) = “(array (8) integer)“) (vn ) 0 2 n < 8) (3j ]O<j<S) [(Vk ] 0 < k < n) [j # p[k] & j - p[k] # n - k & j - p[k] + k - n] & p[nl =_il Because of the similarity between a BSL formula and its logical counterpart, a BSL formula is like a specification for its own self: it describes what it computes. As the reader can readily see, the BSL formula shown above specifies what a solution to the eight-queens problem should satisfy, assuming we read an assignment symbol as equality, and we translate the quantifiers to a conventional notation. This BSL formula compiles into an efficient backtracking program in c that finds and prints instantiations for the array p that would make the (3p)-quantified part of the corresponding first-order formula true in the fixed interpretation. The register declarations shown in the option list are passed to c, and cause the c compiler to place the quantifier indices kj,n in registers if possible, for faster execution. The original BSL compiler was written in Franz Lisp, and generated code acceptable by the Berkeley UNIX’ c compiler, on a VAX 11 computer. We have now ported the BSL ‘UNIX is a trademark of AT&T Bell Laboratories. 148 KEMAL EBCIOGLU compiler to Lisp/VM and IBM 3081-3090 computers: it now generates code acceptable by the c version of the PL.8 compiler [77], and also the AT&T c compiler. We can observe some examples of BSL language features in this eight-queens program: The basic building blocks of BSL are constants, consisting of integers (such as - 2, 0, 3) and record tags (which are identifiers such as ssn, salary), and variables, which are identifiers such as x, p, n, or emp (for convenience, we assume that variables are distinct from record tags). Each variable and constant is a BSL term, and if I, and t, are BSL terms, the binop is one of the binary operators +, -, * , /, sub, and dot, then (binop t, t2) is also a BSL term. Examples of BSL terms are 0, (+ x 2) and (* 2 (dot emp salary)). The constructs (1 + x), (1 - x) may be used as abbreviations for (+ x 1) and (- x l), respectively. A BSL lvalue is either a variable, or a term of the form ( fi . . . (f,_ 1 (f,, x . . . ) . . . ) . . . ) where each of ft.. . . , f,, is either sub or dot, and where x is a variable. Lvalues are terms that can appear as the left-hand operand of an assignment, and are exemplified by x, (dot emp salary), or (sub p n). Lvalues can also be abbreviated as long as their normal notation can be inferred from context; for example, the latter two lvalues can be written as (salary emp) and (p n), in the proper contexts. A BSL atomic formula is either an assignment of the form (:= 1 tt), or a test of the form (relop t, t2), where 1 is an Ivalue, t,, t, are terms, and relop is one of = = (equal), ! = (not equal), < , >=, < = , or > . A BSL atomic formula is a BSL formula. Assuming Fl and F2 are BSL formulas, then so are the following: (and F, F2), (or Fl F2),* (A x init cond incr F,), (E x inir cond incr F,), and (E ((x typ)) F,), where x is a variable; init, incr are terms where init does not contain x; cond is a BSL formula not containing any occurrences of A, E, or := ; and fyp is a type. The BSL types are similar to the type declarations of an Algol-class language, and allow integer, array, and record declarations. Examples of BSL types are integer, (array (3) integer), and (record (ssn integer) (salary integer)). In general, “integer” is a BSL type, and if typ, typl,. . . , typ, are BSL types, k 2 1, and y,, . . . , yk are distinct record tags, and n is a positive integer, then (array (n) fyp) and (record (yi fypJ . . . (y, typ,)) are BSL types. We give here an informal description of the nondeterministic program semantics of BSL: The variables of BSL can range over objects, each of which has a correspond- ing type. Objects of type integer are constants such as -2, 0, 3, and U (called the unassigned constant). An object can also be an array, which is a list of objects of the same type, or a record, which is a list of alternating record tags and objects, not necessarily of the same type. Arrays and records are exemplified by (1 2 U), which is an object of type (array (3) integer), and (ssn 999123456 salary 25000) which is an object of type (record (ssn integer) (salary integer)). The value of a BSL term, in a particular state during execution, is computed by using the usual meanings of the binary operators +, -, * , /, sub, and dot. Here sub is defined to be the subscript operator which takes an array object and an integer i and returns the i th element of the array object (the array elements are numbered starting from 0); and dot is ‘In the eight-queens program above, the construct (and F, F2 6) abbreviates (and F, (and F2 F3)). In general, “and” and “or” associate to the right, and thus (and .) and (or ) can contain more than two subformulas. AN EXPERT SYSTEM FOR HARMONIZING CHORALES 149 defined to be an operator that extracts a subobject of a given record object as determined by a given record tag (it performs a function similar to the dot within the expression “employee.salary” in PL/I). BSL atomic formulas, i.e. assignments and tests, are executed in the conventional manner: the tests are executed by performing the indicated comparison operation after computing the current values of the two terms to be compared; and the assignments are executed by computing the current value of the right-hand-side term, and then destructively changing the value of the left-hand-side term to reflect the current value of the right-hand-side term. If the comparison operation indicated in a test comes out to be true, the effect of the test is a no-op. However, if a test does not come out to be true, or if an assignment is attempted when the current value of the left-hand side is not U, or when the current value of the right-hand side is not an integer, or if an attempt is made to perform an illegal computation (such as using a variable whose value is U in an arithmetic operation or comparison, or dividing by zero), execution does not terminate. The formula (and Ft F’) is executed by first executing F,, then F,_. The formula (or FI F,) is executed by executing one of FI or F2. The formula (A x init cond incr FI) is similar to the c “for” loop; it is executed by saving the old value of x, setting x to init, while cond is true repetitively executing FI and setting x to incr, and restoring the old value of x if and when cond is finally false. (E x init cond incr F,) is executed by saving the old value of x, setting x to init, setting x to incr an arbitrary number of times (possibly zero times), and finally deciding not to set x to incr any more, executing F,, and then restoring the old value of x. Here cond must be true after x is set to init and after each time x is set to incr, or else execution does not terminate. (E ((x typ)) FI) is similar to a “begin-end” block with a local variable; it is executed by saving the old value of x, setting x to an object of type typ all of whose scalar (i.e. integer) subobjects have the value U, executing F,, and then restoring the old value of x. The translation of a BSL program to the first-order assertion that it is true at its termination states is for the most part obvious, as exemplified by the eight-queens program above; however, both the assignment symbol (:=) and the equality test ( = =) of BSL get translated to the equality symbol in the logical counterpart, that is, the program contains procedural information not present in its logical counterpart. First, assume that F’ denotes the first-order translation of a BSL term, formula, or operator F. The translation of BSL terms to first-order logic is straightforward: for example (+ x 2) (dot (sub emp i) salary) translate into +(x,2), dot(sub(emp,i),salary) (which can also be abbreviated as x + 2, emp[i].salary). The first-order translations of the comparison operators -C , > = , < = , > , = = , ! = are the predicate symbols < , 2 , I , > , = , f , respectively. Tests such as (relop t, tz) and assignments such as (:= 1 tl) translate into the first-order atomic formulas t,’ refop’t; and I’ = t;, respectively. (and FI F,) translates into [F,’ & F;], (or FI F2) translates into [F,’ V F;], and (E ((x typ)) FI) translates into (3x ]type(x) =“typ”)[ F,‘]. For a simple subset of BSL, where the only allowable looping constructs are of the form (A x t, (-c x t2) (1 + x) F), (E x t, (< x t2) (1 + x) F), and variants thereof, the translation of these to bounded quantifiers of first-order logic, namely (Vx 1 t,’ 5 x -c t;)[ F’], (3x 1 t; 5 x < t;)[ F’], . . . , works, where t;, t; are the first-order translations of BSL terms t, and t,, respectively, and where x does not occur in either t, or t,. However, for the general case involving arbitrary cond and incr expressions, which we will not elaborate here, the rigorous translation of BSL 150 KEMAL EBCIOiiLU formulas involves associating a different function symbol of the first-order language with every quantified formula of BSL, and is less natural.3 The following translation examples should demonstrate the intuition behind the relationship of a BSL program to its first-order translation: When either (:= x 0) is successfully executed (i.e., x is initially U) or (= = x 0) is successfully executed (i.e., x is initially 0), the assertion x = 0 is true at the termination state. When (or (= = x 0) (= = x 1)) is successfully executed (i.e., x is initially 0 or 1, and the proper subformula of the “or” is chosen for execution), the assertion [x = 0 V x = l] is true at the termination state. When (A i 0 (< i 10) (l+ i) (E (6 integer)) (and (or (:= j 0) (:= j 1)) (:= (sub a i) j)))) is successfully executed (i.e., “a” is initially an array object whose first ten elements are U), (Vi ]O 5 i < 10)(3j ) type(j) = “integer”)[/j = 0 V j = 11 & a[i] = j] is true in the termination state. This assertion says that the first 10 elements of “a” are an arbitrary sequence of O’s and 1’s. To see why this assertion is true at the termination state of the program, observe that during the execution of the program, for each i = 0,. . . ,9, the assertion (3j 1 type(j) = “integer”)[[ j = 0 V j = l] & a[i] = j] was made true by creating (for each i) an integer j equal to 0 or 1, and then making a[;] = j true by assigning j to a[i]. The first-order translation of (and (:= x 0) (:= x (1 + x))) is [x = 0 & x = x + 11, but such a BSL formula can never reach a termination state, no matter what the initial value of x is, because it violates the single-assignment rule enforced by the program semantics of BSL (the single-assign- ment rule is the one that verifies that the left-hand-side is U and the right-hand side is an integer before each explicit assignment). The intuitive purpose of the single- assignment rule is to ensure that the continuation of execution does not destroy the truth of the assertions that were previously made true. Top-level formulas (i.e. complete programs) of the BSL subset we are describing, such as the eight-queens program given above, do not contain free variables, so the truth of the assertions corresponding to such formulas is not affected by the value of any variable in the termination state. Successfully executing such a top-level BSL formula is equivalent to constructively proving that the corresponding first-order sentence is true in a fixed interpretation that involves integers, arrays, records, and operations on such objects (or in all models of a suitably axiomatized “theory of integers, arrays, and records”). A BSL program of the form (E ((x tvp)) F) is implemented on a real, determinis- tic computer via a modified backtracking method, which in principle attempts to simulate all possible executions of the BSL program, and prints out the value of x just before the end of every execution that turns out to be successful. Whenever a choice has to be made between executing Fl and executing F2 in the context (or F, F2), the current state is pushed down to enable restarting by executing F2, and Fl is ?ke [19] for details. In practice, the general case is rarely needed, because BSL programs are often first conceived as first-order assertions rather than, say, “while” loops. AN EXPERT SYSTEM FOR HARMONIZING CHORALES 151 executed. Whenever a choice has to be made between executing F and setting n to incr in the context (E n init cond incr F), the current state is pushed down to enable restarting by setting n to incr, and F is executed. If a test (refop t, tz) is found to be false, or if cond is found to be false in the context (E n init cond incr F), or if the top level (E ((x typ)) . . .) is successfully executed and x is printed, then the state that existed at the most recent choice point is popped from the stack, and execution restarts at that choice point. Attempting to make more than one explicit assignment to a scalar variable or to a scalar subpart of an aggregate variable, and illegal computations (such as attempting to add a number to a variable whose value is U), are considered errors and should never occur during the backtracking execution of a correct BSL program (however, the run-time checks for detecting such errors may be omitted for efficiency reasons). Execution begins with an empty choice-point stack and ends when an attempt is made to pop something from an empty stack. A modification is made to this basic backtracking technique for the case of assignment-free formulas Fl in the context (or Fl F2) or (E n ._. Fl). After a formula Fl in such a context is successfully executed, the most recent choice point on the stack is discarded (which would be the choice point for restarting at F2, or Fi with a different value of n, assuming the modification is uniformly applied). This convention, similar to the “cut” operation of PROLOG, serves to prevent duplicate solutions for x from being printed out (or redundant failures from occurring) when Fl and F2 do not express mutually exclusive conditions, or when Fl is true for more than one n in its quantifier range. Here is an example that demonstrates the motivation behind this modification to backtracking: suppose that many elements of an array “a” are equal to 0 in a particular state during backtracking execution; if in this state an assignment-free subformula (E i 0 (< i N) (1 + i) (= = (a i) 0)) is executed and succeeds after finding that a[i] = 0 for a particular i, and immediately thereafter a failure occurs (or some solution is printed), then there is no point in backtracking to the point in the subformula where i is set to i + 1, and then succeeding again after finding another element of “a” that is equal to 0, because the program will then fail in exactly the same way as before (or will print the same solution that it printed before). So the choice point for backtracking to i = i + 1 is discarded when (E i . . .) succeeds. Since backtracking is a notoriously inefficient search technique, we had to be careful about its implementation in BSL, in the hope of making the language usable for substantial applications. From the compiler-implementation viewpoint, the backtracking semantics described above, which entails saving the entire program state for later restarting at the next alternative, would appear to be an inefficient mechanism, since the program state would seemingly include the current values of all the variables that are active at the point of nondeterministic choice [the active variables are the variables that are declared within the quantifiers enclosing the choice point; nested quantifiers that use the same variable x, such as (E x . . . (A x . . .) . . .), are eliminated by renaming the inner x throughout its scope]. But in case run-time checks of single assignment are omitted, as they are in the present implementation, the single-assignment rule of BSL allows a significant reduction in the amount of information that needs to be saved at a choice point. In the present implementation, where variables are allocated in static storage or in registers (for the BSL subset we are describing), the following observation applies to a typical scalar variable or scalar subpart of a variable, when the state is being pushed down 152 KEMAL EBCIOGLU at a choice point: if the variable already has an integer value, then it does not need to be saved, since it will not have been assigned again and its storage space will have remained intact when backtracking later occurs to this choice point; and if the variable is currently unassigned, then it still does not need to be saved, since its current value will not be used after a backtracking return is made to this choice point. In fact, the only variables that are pushed down at a choice point (called the destructibfe variables) are precisely the variables that are both active at the choice point and declared within the scope of a universal quantifier (A n . . .) enclosing the choice point.4 These variables typically consist of quantifier indices. It is this smallness of state that enables a BSL program to rapidly push down the entire program state at a nondeterministic choice point, and to return to the most recent choice point directly when a failure later occurs, without having to execute state- ments in the backward direction as in older nondeterministic languages [26,7], and without having to restore variables on a “trail” stack to their previous value, as in many PROLOG implementations (e.g. [24,76]). The BSL compiler also uses a technique for eliminating or reducing the need for pushing down choice points: If a subformula Fi in the context (or Fl F2) or (E n init cond incr Fl) is assignment-free, the BSL compiler does not generate code for pushing down the state before the execution of F,, and emits compare-and-branch statements for Fl that directly jump to next alternative when Fl fails (the next alternative is F2, or the setting of n to incr), and that directly jump to the continuation of (or Fl F2) or (E n . . . Fl) when Fl succeeds, using a standard compilation technique for Boolean expressions [l], extended with bounded quanti- fiers. This standard compilation technique is equivalent to, but much more efficient than, the original “cut”-like semantics given above, which involves first pushing down and then discarding a choice point. 5 Moreover, even when there are assign- ments in a subformula Fl in the context (or Fl F2) or (E n init cond incr F,), the compiler delays the pushdown operations that would normally precede the code for Fl. and emits compare-and-branches for as large an initial segment of Fl as possible, as long as assignments are not yet encountered in F,; the code for this initial segment jumps directly to the next alternative when Fl fails (the possibility exploited here is that F, may fail before the state needs to be pushed down).6 It should also be noted that BSL operates on the native data types of a traditional computer, such as integers and arrays of integers, rather than machine words which contain both data and tags, as is the case in a typical PROLOG implementation. Moreover, bounded quantifiers in BSL are native to the language and often compile into simple loops, which enable traditional compiler optimizations such as code motion, strength reduction, or induction-variable elimination. 4This is because n may be reincremented, and the other variables declared within (A n ) may be destroyed through reuse of their storage space during the continuation of execution, due to the static nature of the storage. For example, in the eight-queens program above, j and n (but not p) need to be pushed down within (E j .) for the purpose of later backtracking to the setting of j to j + 1. ‘Gergely and Szots [28] have proposed a logic-programming language called Logic of Cuttable Formulas, whose formulas are compiled into efficient programs that have essentially the same semantics as the assignment-free subset of BSL. ‘For example, in the eight-queens example above, no pushdown operations are executed within (E j (and )) until just before the assignment (:= (p n) j). If (and .) fails before reaching the assignment, it jumps directly to the setting of j to j + 1. A detailed compilation algorithm is given in [19]. ANEXPERTSYSTEM FOR HARMONIZINGCHORALES 153 Some performance comparisons between BSL and PROLOG and Lisp on exhaus- tive search algorithms, which may help to quantify the advantage of using BSL, were given in [20]. BSL'S method of combining extended Boolean tests and backtracking is perhaps a natural way to “execute” a logical specification on a computer; however, the possible logical specifications are limited to those that correspond to valid BSL programs, and it is required that the programmer indicate which equalities in the specification are to be executed as assignments, and which are to be executed as tests.’ The language subset described up to here is called L*, and constitutes the “pure” subset of BSL. The full BSL language also allows user-defined predicates in addition to <, > ,..., user-defined functions in addition to +, sub,. . . , global variable declarations, macro and constant definitions, “if’ and “case” statements, enumera- tion types, real types, and a richer set of primitive operations. The user-defined predicates of BSL are nondeterministic recursive procedures analogous to PROLOG procedures, but whether a parameter of a BSL predicate is used (i.e., is an input) or is assigned to (i.e., is an output) during the predicate call is fixed by the programmer. Facilities include a “(with . . .)” construct that allows convenient abbreviations for certain lvalues that would otherwise have to be written out with long chains of sub and dot operators. A “not” connective is allowed as long as we can move the “not” in front of the atomic formulas with de Morgan-like transformations, and then change (not (= = . . .)) to (! = . . . ), etc., and still get a valid BSL formula. BSL is also extended with heuristics, which are BSL formulas themselves, and which can guide the choices made during the backtracking execution of a BSL program. As a preparation for the next section, which depicts the use of BSL for implementing expert systems, we will describe the heuristics feature of BSL below. Normally, the order of enumeration of the possible successful executions, or termination states, of a BSL formula F during a backtracking simulation is deter- mined in a fairly trivial way via factors such as which subformula occurs first in an (or . . . , . .). This order is fine for applications where all solutions have to be found, but in applications such as music generation, the list of all solutions is of impractical length and is quite boring. It is thus necessary to alter the order of enumeration of termination states so that a better solution will tend to come out first. A more sophisticated order of enumeration of the termination states of a BSL formula F can be obtained by enclosing F in the construct (H F (I, _. . f,) Fk . . . F,), where I I,..., I,, are (not necessarily scalar) lvalues that are assigned during F, and Fk , . . . , F, are side-effect-free BSL formulas, called heuristics. (H F . . .) is simulated as follows: First all executions of F are simulated, and whenever an execution of F terminates successfully, the termination state of the current execution, as represented by the assignments to I,, _. . , I,, is assigned a ‘In contrast to BSL. the unification algorithm [63] and certain nonlogical systems such as “Con- straints” 1741 defer the choice between making equality and checking for it to run time. But the unification algorithm has the elegant consequence of being able to answer different questions about a relation without reprogramming, such as using the same code for finding the parents of a given x, or finding the children of a given y, or checking if a given y is a parent of a given x, or finding pairs (x,y) such that y is a parent of x. BSL is suitable for generate-and-test applications where such versatility, which is useful but usually costly, is not of prime importance, and where the question is fixed (e.g., given the result of laboratory experiments, find the solutions to a molecular-genetics problem, not the other way around, as exemplified by [72]). 154 KEMALEBCIOGLU numerical worth by executing each heuristic Fk,. . . , F, in the current termination state. The heuristics are weighted by decreasing powers of two, and the worth of a termination state is computed to be the sum of the weights of the heuristics that it makes true. Thus, if a heuristic F, is true in the current termination state, k 2 i 2 0, it increases the worth of the current termination state by 2’; otherwise, it does not affect the worth of the current termination state. Then the assignments to fi,. . . , I, in the current termination state are saved in a list along with their worth, and a failure return is forced in order to obtain more termination states of F. If and when all termination states of F are exhausted (as defined by the modified backtracking simulation), the resulting list is sorted according to the worth of each termination state (i.e. assignment to I,, . . , I,,). Ties are resolved with explicit randomness, by shuffling the list randomly before sorting, in order to defeat any extra unwanted “heuristics” that may result from the regularity in the generation of the list. Then (H F . . . ) succeeds first with the highest-valued termination state of F, then, if backtracking occurs, with the next-highest-valued termination state, etc., and finally backtracks to a previous choice point when there are no more assignments left in the list. This feature of BSL forms the basis for the BSL generate-and-test paradigm, which is described next. THE GENERATE-AND-TEST PARADIGM IN BSL Despite its Spartan data types, BSL can be used for designing large and complex expert systems in a structured manner. The formal analog of a knowledge-based system based on the generate-and-test method [72] can be implemented in BSL via a very long formula of the following form: (E ((s (array(N) WI))) (E ((inp ~2)) (and “initialize inp” (AnO(< nN)(l+ n) (H (and (or (and conditions, actions,) . . . (and conditions, actions,)) constraint, ; ; ; generate section ; . . . ; test section constraint,) ; KS W heuristic, . . . ; recommendations section heurisric,))))) ; In the generate-and-test paradigm of BSL, the computation proceeds by “gener- ate-and-test steps”, where each step consists of selecting and assigning an acceptable value to the n th element of the solution array “9 depending on the elements 0,. . . , n - 1 (and also on the input data structure “inp”). The condition-action pairs given here are the formal analogs of production rules [9] as they are used in a generate-and-test application. The conditions are subformulas that typically per- ANEXPERTSYSTEM FORHARMONIZINGCHORALES 155 form certain tests about elements 0 , . . . , n - 1 of the solution array, and the actions are subformulas that typically involve assignments to element n of the solution array. Thus a condition-action pair has the informal meaning “IF conditions are true about the partial solution, THEN a new element as described by the actions can be added to the partial solution”.’ The constraints are subformulas that assert absolute rules about the elements 0,. . . , n of the solution array. They have the procedural effect of rejecting certain assignments to element n (this effect is also called ear& pruning in AI literature [32]). The heuri.stics are subformulas that assert what is desirable about elements 0,. . , , n of the partial solution; they have the procedural effect of having certain assignments to element n tried before others are. The condition-action pairs are called the generate section, the constraints are called the test section, and the heuristics are called the recommend&ions section of the knowledge base. Each step of the program is executed as follows [we are repeating the explanation given above for the (H . . .) construct]: All possible assignments to the n th element of the partial solution are sequentially generated via the production rules. If a candidate assignment does not comply with the constraints, it is thrown away; otherwise its worth is computed by summing the weights of the heuristics that it makes true, and it is saved in a list, along with its worth. When there are no more assignments to be generated for solution element n, the resulting list is sorted according to the worth of each candidate. The program then attempts to continue with the best assignment to element n, and then, if a dead end is later encountered and a backtracking return is made to this point, with the next best assignment, etc., as defined by the sorted list. The reasons we chose the particular powers-of-two weighting scheme described above for the heuristics were its clarity, freedom from unconstrainted numerical weights, and efficient implementation. Heuristic search is an important topic that has attracted considerable research interest, and many alternative approaches have been studied (e.g., [57.X$60,55]). Although BSL's heuristic search paradigm is itself simple, the heuristic criteria that one can specify with ease in BSL can be quite sophisticated, because heuristics have the generality of logic formulas. Heuristics have no effect on the non!eterministic semantics of a BSL formula, or on its first-order translation: the first-order transla- tion of (H F . . .) is taken to be the same as the first-order translation of F; moreover, in the nondeterministic execution semantics of BSL (H F . . .) is executed by executing merely F. Within the production rules, constraints, and heuristics, the existential and universal quantifiers of BSL can provide capabilities equivalent to the pattern-match- ing capabilities of a true production system [27]. For example, assume that we are dealing with a molecular-genetics application similar to the one described in [72]. In this problem, the solution object can be represented as a sequence of elements, each having two attributes, a site and a segment. 9 The problem is to find some or all of the solutions that are consistent with the rules of the domain and the results of ‘Note that this condition-action paradigm captures only the generate-and-test application of produc- tion rules, which is the intended application of EM.. In general. production systems can allow very arbitrary control, such as self-modification [78] or blackboards [31]. The sites are enzyme names labeling points on a circular DNA molecule where that enzyme has made a cut, and the segments are integers indicating the length of the DNA molecule segment from one site (cut point) to the next. Stetik [72] describes the problem in detail. 156 KEMAL EBCIOCLU laboratory experiments given as input. In the context of this application, in order to specify a production rule that says “IF certain conditions are true, THEN the segment whose length is the smallest among a given array of segments can be added to the partial solution”, one could write” (and “certain conditions” (EiO(< imaxsegs)(l+ i) (and(A j 0 ( < j maxsegs) (1 + j) (imp(! = i j) (< (seg_list i) (seg list j)))) (:= (segment (s n)) (seg - list;))))) Assuming appropriate type declarations for seg_list and s, the logical translation of this subformula is [“certain conditions” & (3 (0 5 i < maxsegs) [(Vj 10 < j < maxsegs)[i # j =j seg_list[i] < seg_listb]] & s[n].segment = seg_list[i]]]. Similarly, a constraint asserting “IF certain conditions are true, THEN the site that has just been added to the solution cannot have more than one previous occurrence in the solution” can be written as (imp “certain conditions” (not (E i (l- n) (> i 0) (l- i) (Ej (l- i) (> = j 0) (l- j) (and( = = (site (s i)) (site (s n))) (= = (site (s j)) (site (s i)))))))) whose logical translation is [“certain conditions” =j not[@i In - 1 2 i > 0)(3j ]i - 1 2 j 2 O)[s[i].site = s[n].site & s[j].site = s[i].site]]]. Operations that may normally require more than one recognize-act cycle in an ordinary production system can also be performed in a single generate-and-test step in the present paradigm; e.g., more than one attribute of the next item to be added to the solution, where each attribute involves several nearly independent choices, can be decided in a single step. For example, assuming each solution element has two attributes, site and segment, the generate section of the knowledge base can be constructed as follows: (and (or “condition-action pairs to choose the n th site”) (or “condition-action pairs to choose the n th segment”)) where the n th segment may depend on the n th site. (Note that in a production-sys- tem language, one normal way of achieving the same effect could take two “‘Here. (imp F, Fi) is a macro that expands into (or(not F,) F,) ANEXPERTSYSTEMFORHARMONIZINGCHORALES 157 recognize-act cycles: choosing the site during the first recognize-act cycle and changing the “current task” to be that of choosing the segment, and choosing the segment during the second recognize-act cycle.) In fact, the generate section of the Jill-in knowledge base of the CHORAL system decides the attributes of the three voices bass, tenor, alto, as well as other relevant attributes, in a single step, and has the form (and . . . (A v bass ( < v soprano) (1 + v) (or “condition-action pairs to choose attributes of voice v at the n th step”)) . . . ) Our experience while writing large knowledge bases in BSL has suggested that such knowledge bases can attain a very high degree of complexity, and some discipline is required for managing them (PROLOG and general-purpose expert-system shells are also not immune to this problem). The production rules, constraints and heuristics of a knowledge base need not be enumerated in an unstructured manner as shown here: to enhance legibility, they can be hierarchically grouped according to subject, similarly to chapters and paragraphs of a musical treatise (in a larger project, perhaps different teams can design the different chapters, based on an agreement on the shared data structures and a written outline of the knowledge base). Similarly, distinguishable concepts (e.g. parallel motion of two voices, dou- bling the fifth of a chord), can be implemented through hierarchies of predicate, function, or macro definitions, so constraints and heuristics are short and are as close as possible to an English paraphrasing of them. Other points we learned through experience are that nested AND-OR-AND-OR structures must be avoided (multiplied out, normalized), and that long lists of similar constraints or production rules should be replaced by a compact table that is interpreted by a single production rule or constraint, and constraints or heuristics longer than a screenful of lines should be broken down. But if a proper design methodology is followed, the BSL paradigm indeed allows the benefits of a true production system in an important class of generate-and-test applications. REPRESENTING KNOWLEDGE WITH MULTIPLE VIEWPOINTS The paradigm shown above is suitable only for simple generate-and-test problems, such as Stefik’s GA1 system [72]. It uses a single model of the solution object, as represented by the primitives allowed by the solution array’s type declaration. Representing knowledge about multiple viewpoints, or multiple models of a solution object, is a need that often arises in the design of complex expert systems: the Hearsay-II speech-understanding system [23] was such an example, where there was a need to observe the interpretation of speech simultaneously as mutually consistent streams of syllables, words, and word sequences. In logic, a good way to describe an object from different viewpoints is to use different primitive functions and predi- cates for each view, since without the appropriate primitives, logic formulas for 158 KEMALEBCIOfiLU describing a concept can be unnecessarily long. But since BSL allows efficient manipulation of the native data types of a traditional computer, such as arrays and records, it is preferable to implement multiple viewpoints in BSL with pseudofunc- tions and predicates, through data structures. In BSL, each viewpoint is represented by a different data structure, typically an array of records, that serves as a rich set of primitive pseudofunctions and predicates for that view. For example, assuming that we wish to have a viewpoint that observes the chord skeleton of a musical piece with two primitive functions p(n,v) and a(n,v), represent- ing the pitch and accidental of voice v of chord n, BSL lvalues of the form c[n].p[v] and c[n].a[v], where c is the array of records of the view, can be used as a pseudonotation to abbreviate p(n,v) and a(n,v). BSL'S multiple-view paradigm has the following procedural aspect, which amounts to interleaved execution of generate-and-test: It is convenient to visualize a separate process for each viewpoint, which constructs that particular view of the solution, in close interaction with other processes constructing their respective views. A process typically executes in units of “generate-and-test steps”. The purpose of each step, as before, is to assign acceptable values to the n th element of an array of records, depending on the values of the array elements 0,. . . , n - 1, and external inputs, e.g. elements of external arrays of records, whose values have been assigned by other processes. The processes, implemented as BSL predicate definitions, are arranged in a round-robin scheduling chain. With the exception of the specially designated process called the clock process, each process first attempts to execute zero or more steps until all of its inputs are exhausted, and then schedules (calls) the next process in the chain with parameters that indicate how far each process has progressed in assigning values to its output arrays. The specially designated clock process at- tempts to execute exactly one step when it is scheduled; all other processes adjust their timing to this process. In certain cases a view may be completely dependent on another, i.e., it may not introduce new choices on its own. In the case of such redundant views, it is possible to maintain several views in a single process, and share heuristics and constraints, provided that one master view is chosen to execute the process step and comply with the paradigm. One way to do this is as follows: at the n th step of such a process, the generate section is executed to produce a candidate assignment to the attributes of the n th element of the master view, then the subordinate views are updated according to the chosen master-view attributes, and then a mixture of constraints and heuristics from both the master and subordinate views are used to decide if the candidate assignment to the n th element of the master view is acceptable and desirable. It is evident that the framework described here is in contrast with the more common techniques for constructing expert systems, where some emphasis is placed on sophisticated control structures. We should therefore explain why we have chosen such a streamlined architecture for designing an expert system, rather than a more complex paradigm such as the multiple-demon queues of [71], or the oppor- tunistic scheduling of [23]. We strongly believe that striving to use simpler control structures is a better approach to the design of large systems. Our design approach is in fact a deliberate choice, and is analogous to a recent approach to computer architecture [59,33,62]: It is a preliminary attempt at reducing the semantic gap between the top and bottom levels of the hardware-software complex that imple- AN EXPERT SYSTEM FOR HARMONIZING CHORALES 159 ments an expert system, by designing a streamlined set of system primitives that directly correspond to the target problem” (cf. [53]). The paradigm described here has served to simultaneously represent knowledge about and construct multiple models of the solution object for the chorale program. We suspect that it can also be used for any generate-and-test application where (1) execution efficiency is manda- tory during all stages of the development phase, and (2) the solution can be conveniently represented as one or more Pascal-style data structures. Note that programming such a demanding application in BSL would be much easier than programming it in c or Pascal, since BSL is indeed a high-level declarative language that gives access to the expressive richness of concepts of first-order predicate calculus, despite the fact that there is little tradeoff of efficiency in choosing BSL over conventional low-level languages. However, like some of the other knowledge-en- gineering paradigms, such as diagnosis-oriented skeletal systems [6], BSL has a limited scope of applicability. In particular, the BSL paradigm would be unsuitable for applications that cannot do without list processing: in music, we could get away with mere arrays and records, because music can be represented as a uniform sequence of events. A COMPILATION TECHNIQUE FOR INTELLIGENT BACKTRACKING Ordinary, or chronological, backtracking may sometimes be inefficient when no choices can be found for successfully executing the current generate-and-test step and the immediately preceding step is irrelevant to the failure of the current step. In this case, a substantial amount of computation that will look useless to a human observer will be done until the most recent step that caused the failure is reached. For example, assuming the current step is the n th step and the choice responsible for the failure of the current step was made at step n - k, k > 2, then there will be a seemingly useless sequence of backtrackings to step n - 1 until all of the choices at step n - 1 are exhausted, and then there will be a backtracking to step n - 2, and then there will again be many backtrackings to step n - 1, and so on. The BSL compiler attempts to alleviate the overhead associated with backtracking by a special compilation technique triggered by a compiler option. We have expressly tried to find an intelligent backtracking heuristic with very low overhead, since we have observed that overly ambitious intelligent backtracking schemes can introduce so much overhead that they can actually slow down the typical applica- tions. In our technique, it is assumed that the computation proceeds as a sequence of generate-and-test steps. Otherwise the technique is domain-independent, and will produce the same solutions as ordinary backtracking would. Each scalar variable, or each scalar member of an aggregate variable, has a tag associated with it. At run time, things are arranged so that the tag always contains the stack level to backtrack to in order to get a different choice for the value of the corresponding variable. “An alternative successful approach is to reduce the semantic gap between existing AI software paradigms and hardware, by designing specirrlized hardware for Lisp and PROLOG. The BSL paradigm, on the other hand, is destined for well-understood RISC or supercomputer architectures, As for the parallel execution of BSL. we expect that significant speedup of BSL programs will be achievable in the future, via the emerging “ very long instruction word” (VLIW) architectures and compilation techniques [22,56,21], which are intended for parallel execution of general, ordinary, sequential programs. 160 KEMAL EBCIOC?LU During the execution of a step, a running maximum is maintained of the tags of all variables that occur in the failing tests. When a step cannot be executed and backtracking is necessary, the program returns to this computed most recent responsible step for the failure, which is not necessarily the chronologically preced- ing step. l2 There have been a number of research projects in AI and logic programming that also have addressed the intelligent backtracking problem, using alternative approaches (e.g. [71,13,5,61,11]). The main use of this heuristic is for eliminating the need for Conniver-style [73] programmed return to an earlier-than-normal step, or the PROLOG-style “cuts” to labels specified by the programmer. This sort of inelegant intrusion into the backtracking mechanism would have otherwise been mandatory in the chorale program, since when a step of the chord skeleton view fails, it must at least backtrack to the previous step of the chord skeleton view, which is not necessarily the immediately preceding step (the immediately preceding step can be totally irrelevant to the failure). Because of this property of the scheduling order of the viewpoints, the intelligent backtracking technique causes the chorale program to run much faster than it would without it. However, we have encountered cases in the chorale program where this conservative and domain-independent intelligent back- tracking mechanism is not intelligent enough. In particular, it appears to be desirable to detect not only the responsible step, but also the precise change that is required at that step (as was done in [68]); but we do not presently know of an easy way to compile such an intelligent backtracking algorithm; similarly, we do not know whether the additional overhead would be justified. To remedy the problem, we have added an incomplete-search feature to the compiler that gives a fixed number of chances to the intelligent backtracking technique when there are repeti- tive failures at a given step, and then forces the program to backtrack to a step earlier than the step recommended by the intelligent backtracking technique. The earliness of the step to backtrack to is increased, while the failures continue to occur in the same step as they did before. This feature cannot be used in more mundane applications where all solutions must be found, but it did give satisfactory results in the present application.13 ‘“Here is some more detail: In normal code, when an assignment is made to a scalar variable or a scalar subpart of an aggregate variable, the stack level to backtrack to for undoing this assignment is placed in the tag of the variable (the additional assignment is compiled inline). Tests are executed as usual in normal code. During the execution of a generate-and-test step (a subformula specially designated by the programmer, where we are interested in finding the most recent responsible stack level to backtrack to, in case this subformula fails), whenever a variable is assigned a value, its tag is set to the maximum of the tags of the variables on the right-hand side of the assignment. Whenever a test within the generate-and-test step succeeds, execution proceeds as usual, but when the test fails. a “current estimate” ( a running maximum) for the most recent responsible stack level is updated, if it is less than any of the tags of the variables that were part of the failing test, via additional code compiled inline after the test. If the current generate-and-test step fails, this estimate is used for selecting the stack level to backtrack to; otherwise, the tags of the variables assigned during the current step are changed to point to the next alternative for the current step, and execution continues, typically with the next generate-and-test step. Various optimizations are performed on top of these basic principles. See [19] for further details. “The incomplete search technique was later disabled on the IBM 3081-3090 version of the program, because we felt we could afford more search on the faster hardware. ANEXPERTSYSTEM FORHARMONIZINGCHORALES 161 THE H SOWLEDGE MODELS OF THE CHORAL SYSTEM We are now in a position to discuss the CHORAL system itself. We will be able to give here only a brief overview of the knowledge base of CHORAL, which is in reality very long and complex. The CHORAL system uses the backtrackable process schedul- ing technique described above to implement the following viewpoints of the chorale: The chord-skeleton view, which corresponds to the clock process, observes the chorale as a sequence of rhythmless chords and fermatas, with some unconventional chord symbols underneath them, indicating key and degree within key. The primi- tives of this view allow referencing attributes such as the pitch and accidental of a voice v of any chord n in the sequence of skeletal chords. This is the view where we have placed, e.g., the production rules that enumerate the possible ways of modulat- ing to a new key, the constraints about the preparation and resolution of a seventh in a seventh chord, and the heuristics that prefer “Bachian” cadences. The jil/-in view observes the chorale as four interacting automata that change states in lockstep, generating the actual notes of the chorale in the form of suspensions, passing tones, and similar ornamentations, depending on the underly- ing chord skeleton. This view reads the chord skeleton output. For each voice v at fill-in step n, the primitives allow referencing attributes of voice v at a weak eighth beat and an immediately following strong eighth beat, and the new state that voice v enters at fill-in step n (the states are suspension, descending passing tone, and normal). At each one of its steps, the fill-in view generates the cross product of all possible inessential notes (suspensions, passing tones, neighbor notes, miscellaneous embellishments specific to the chorale style, etc.) in all the voices, and then filters out the unacceptable combinations and selects the desirable combinations, using a rather complex list of constraints and heuristics. In this view we have placed, e.g., the production rules for enumerating the long list of possible inessential note patterns that enable the desirable bold clashes of simultaneous passing tones and suspensions. a constraint about not sounding the resolution of a suspension above the suspension, and a heuristic on following a suspension by another in the same voice (a “Bachian” clichC).14 The melodic-string view observes the sequence of individual notes of the different voices from a purely melodic point of view. The primitives of this view allow referencing the pitch and accidental of any note i of a voice v. This is the view where we have placed, e.g., a constraint about sevenths or ninths spanned in three notes, and a heuristic about continuing a linear progression. The merged melodic-string view is similar to the melodic-string view except that it observes the repeated pitches merged together. This view was used for recognizing and advising against certain bad melodic patterns that we feel are not alleviated even if there are repeating notes in the pattern. I4 We felt that enabling the bold clashes of multiple simultaneous inessential notes was indispensable for obtaining a “Bachian” melocid-harmonic Row. Not allowing multiple simultaneous inessential notes or style-specific modulations in the harmonization would have greatly simplified the problem domain, but then we would probably not obtain music. 162 KEMAL EBCIOeLU The time-slice view observes the chorale as a sequence of vertical time slices each of which has a duration of a small time unit (an eighth note), and imposes harmonic constraints. The primitives of this view allow referencing the pitch and accidental of a voice v at any time slice i, and whether a new note of voice v is struck at that time slice. We have placed, e.g., a constraint about consecutive octaves and fifths in this view. The Schenkerian-analysis view is based on our formal theory of hierarchical voice leading, inspired by Schenker [65,66] and also by Lerdahl and Jackendoff [45,46]. The core of this theory consists of a set of rewriting rules which, when repeatedly applied to a starting pattern, can generate the bass and descant lines of a chorale. The Schenkerian-analysis view uses our rewriting rules to find separate parse trees for the bass and descant lines of the chorale, employing a bottom-up parsing method, and using many heuristics for choosing (among the alternative possible actions at each parser step) the action that would hopefully lead to the musically most plausible parsing. Unlike Lerdahl and Jackendoff s theory, which is based on a hierarchy of individual musical events (e.g. chords, noteheads), our theory is based on a hierarchy of slurs, and is more in line with Schenker’s theory. The discussion of our voice-leading theory is beyond the scope of this paper, and the details can be found in [19]. The Schenkerian-analysis view observes the chorale as the sequence of steps of two nondeterministic bottom-up parsers for the descant and bass. This view reads the fill-in output. The primitives of this view allow referencing the output symbols of a parser step n, the new state that is entered after executing step n, and the action on the stack at parser step n. The rules and heuristics of this view belong to a new paradigm of automated hierarchical music analysis, and do not correspond to any rules that would be found in a traditional treatise. In this view we have placed, e.g., the production rules that enumerate the possible parser actions that can be done in a given state, a constraint about the agreement between the fundamental line accidentals and the key of the chorale, and a heuristic for proper recognition of shallow occurrences of the Schenkerian D-C-B-C ending pattern. The fill-in, time-slice, and melodic-string views are embedded in the same process, with fill-in as the master view among them. The order or scheduling of processes is cyclically chord skeleton, fill, Schenker bass, Schenker descant. Each time chord skeleton is scheduled, it adds a new chord to the chorale; each time fill-in is scheduled, it fills in the available chords and produces quarterbeats of the actual music until no more chords are available. Each time a Schenker process is scheduled, it executes parser steps until the parser input pointer is less than a lookahead window away from the end of the currently available notes for the descant or bass.15 When a process does not have any available inputs to enable it to execute any steps when it is scheduled, it simply schedules the next process in the chain without doing anything. The chorale melody is given as input to the program. “The lookahead window gradually grew bigger as our ideas evolved, and in the recent versions. for the sake of reducing module sizes, we have found it expedient to place the Schenker processes in a separate postprocessing program that reads its input from a file produced by the other views. Note that the technique of using a separate program for a particular process is not necessarily outside the paradigm of nondeterministic parallel processes; it is rather an optimization of a degenerate case of the same paradigm. AN EXPERTSYSTEM FOR HARMONIZINGCHORALES 163 There are currently a total of approximately 350 production rules, constraints, and heuristics in the chorale program. The rules and heuristics were found mainly from empirical observation of the chorales and personal intuitions, although we used a number of traditional treatises (such as [47] and [41, Vol. I]) as an anarchronistic, but nevertheless useful point of departure. The current version of the chorale program aims only to harmonize an existing chorale melody and assign an analysis to it. All parts of the chorale program are written in BSL, except for the graphics routines and the routine to read in and preprocess the chorale melody, which are written in c. In the VAX 11/780 version of the program, it used to take typically 15-60 minutes of CPU time to harmonize a chorale. In the present version, which has a larger knowledge base and some extremely difficult rules intended to increase the output quality, it typically takes about 3-30 minutes of IBM 3081 CPU time to harmonize a chorale, although a few chorales have required several hours. The program has presently been tested on about 70 chorales (consuming inordi- nate amounts of CPU time) and has reached an acceptable level of competence in its harmonization capability: we can say that its competence approaches that of a talented student of music who has studied the Bach chorales. The program has also produced good hierarchical voice-leading analyses of descant lines, but the Schenkerian-analysis knowledge base still reflects a difficult basic research project in music analysis, and is not as powerful as the harmonization knowledge base. We have also not been able to get any good parsings involving the basses so far. The CHORAL system takes an alphanumeric encoding of the chorale melody as input, and outputs the chorale score in conventional music notation, and the descant parse trees in Schenkerian slur-and-notehead notation. The output can be directed to a graphics screen, or can be saved in a file for later printing on a laser printer. The BSL compiler inserts a simple interactive interface in “(H F . . .)” constructs, which can explain the choices made at any step of a viewpoint, and other kinds of debugging tools are built into the program itself, including a graphic display of the progress of the composiuon. In the Appendix we present five examples of harmonizations produced by the program (all the chorale numbers in this paper are from [75]). The reader will notice that some notes are too high or too low for the range of some of the voices in these examples: this is because the program does not generate the chorale with the voices in their proper ranges, but it ensures that there exists a transposition interval that will bring all the voices to their proper ranges. Also note that the parallel fifths between the soprano and tenor that accompany the anticipation pattern at the ending of the second phrase of chorale no. 75 (Figure lo), and at the ending of the last phrase of chorale no. 68 (Figure 9), are allowable in the Bach chorale style; see, for example, no. 383 in [75]. Bach’s harmonizations of two of the same melodies, no. 128 and no. 48, are also given for comparison. Occasionally, the program’s harmo- nizations can be quite similar to Bach’s, as in no. 128 (Figures 6 and 7) but in other cases, the program has its own different, less austere style, as in no. 48 (Figures 11 and 12). Many more output examples, and the complete list of rules of the system in English (about 77 single-spaced book pages) can be found in [19]. As a concrete example of what type of knowledge is embodied in the program, and how such musical knowledge is expressed in BSL'S logiclike notation, we take a constraint from the chord-skeleton view. The following subformula asserts a famil- iar constraint about false relations (this is the most recent revision of this constraint; 164 KEMALERCIO~LU an earlier version was given in our previous publications): When two notes which have the same pitch name but different accidentals occur in two consecutive chords, but not in the same voice, and no single voice sounds these notes via chromatic motion, then the second chord must be a diminished seventh, or the first inversion of (a dominant seventh or a major triad), and the bass of the second chord must sound the sharpened fifth of the first chord and must be approached by an interval less than or equal to a fourth, or the soprano of the second chord must sound the flattened third of the first chord. In case the bass sounds the sharpened note of the false relation and moves by ascending major third (matching the pattern e-g# in a C-major-E-major chord sequence), then some other voice must move in parallel thirds or tenths with the bass (matching the pattern g-b).16 False relations are also allowed unconditionally between phrase boundaries, when there is a major-minor chord change on the same root. (The exception where the bass sounds the sharpened fifth of the first chord is commonplace; the less usual case where the soprano sounds the flattened third can be seen in the chorale “Herzlich thut mich verlangen “, no. 165. The case where there is a major-minor chord change on phrase boundaries can be seen in chorale no. 46 or no. 77. These exceptions are still not a complete list, but we did not attempt to be exhaustive.) The complexity of this rule is representative of the complexity of many of the production rules, constraints, and heuristics in the CHORAL system. We see the BSL code for this rule below: (A u bass ( c: = u soprano) (1 + u) (A v bass( < = v soprano) (1 + v) (imp(and( > n 0) (!= uv) (= = (mod(p1 u) 7) (mod(p0 v) 7)) (! = (al u) (a0 v)) (not(E w bass( < = w soprano) (1 + w) (and( = = (mod(p1 w) 7) (mod(p1 u) 7)) (= = (PO w) (Pl w)))))) (or (and(member chordtype (dimseventh domseventhl majorl)) (or(and(= = (a0 v) (1-t (al u))) (== vbass) (= = (mod( - (p0 v) rootl) 7) fifth) ( < = (abs( - (pl v) (p0 v)) fourth) (imp(thirdskipup(p1 v) (p0 v)) (E w tenor( -C = w soprano) (1 + w) (and( = = (mod( - (pl w) (pl v)) 7) third) (thirdskipup(p1 w) (PO w)))))) (and( = = (a0 v) (1 - (al u))) (= = v soprano) (= = (mod( - (p0 v) rootl) 7) third)))) “Both of these thirds are filled in with a passing note at the fill-in view. AN EXPERT SYSTEM FOR HARMONIZING CHORALES 165 (and( > fermatal 0) (= = root0 rootl) (= = chordtypel major0) (member chordtype minortriads))))))) Here, n is the sequence number of current chord; (pi v), i = 0, 1, . . . , is the pitch of voice v of chord n - i, encoded as 7* (octave number) + (pitch name); (ai v), i=O,l,..., is the accidental of voice v in chord n - i; and chordtypei and rooti, i=O,l,..., are the pitch configuration and root of chord n - i, respectively. fermatai, i = 0, 1,. . . , indicates the presence of a fermata over chord n - i when it is greater than 0. The notation p0, pl, etc. is an abbreviation system, obtained by an enclosing BSL “with” statement, that allows convenient and fast access to the most recent elements of the array of records representing the chord-skeleton view. (thirdskipup pi p2) is a macro which signifies that p2 is a third above pl. We repeat the constraint below in a more standard notation for clarity, using the conceptual primitive functions of the chord-skeleton view instead of the BSL data structures that implement them: (Vu 1 bass _< u I soprano)(Vv 1 bass I v I soprano) [[n > 0 & u # v & mod(p(n - l,u),7) = mod(p(n,v),7) & a(n - 1,~) # a(n,v) & not(3w (bass I w I soprano)[mod(p(n - l,w),7) = mod(p(n - l,u),7) & p(n - 1,~) = phw)ll [Ghordtype(n) E {dimseventh,domeseventhl,majorl} & [[a(n,v) = a(n - 1,~) + 1 & v = bass & mod(p(n,v) - root(n - 1),7) = fifth & abs(p(n - 1,~) - p(n,v)) s fourth & [thirdskipup(p(n - l,v),p(n,v)) =) (3w 1 tenor 2 w 5 soprano) [mod(p(n - 1,~) - p(n - l,v),7) = third & thirdskipup(p(n - l,w),p(n,w))]]J v [a(n,v) = a(n - 1,~) - 1 & v = soprano & mod(p(n,v) - root(n - 1),7) = third]]] [fermata(n - 1) > 0 & root(n) = root(n - 1) & chordtype(n - 1) = major0 & chordtype E minortriads]]] Before showing an example of a heuristic, it is appropriate to touch upon the significance of heuristics for music generation. It is a known fact that absolute constraints are not by themselves sufficient for musical results: Composers normally use much additional knowledge to guide their choices among the possible solutions. Our limited powers of introspection prevent us from exactly replicating the thought process of such choices in an algorithm; but there exist algorithmic approximations, based on large amounts of precise domain-specific heuristics, or preferences, that tend to give good results in practice (cf. [43]). The chorale program uses an extensive body of heuristics for selecting the preferred choice among the list of possibilities at each step of the program, as previously described in the section on the BSL generate-and-test paradigm. Examples of heuristics would be to continue a linear 166 KEMAL EBCIOGLU progression, or to follow a suspension by another one in the same voice. To exemplify the BSL code corresponding to a heuristic, we again take the chord-skeleton view. The following heuristic asserts that it is undesirable to have all voices move in the same direction unless the target chord is a diminished seventh. Here the construct (Em Q (ql q2 . . . ) (F Q)) is a macro which expands into (or (Fq,) (Fq,) . ..). th us allowing us to use the second-order concept of quantifica- tion over a set of predicates: (imp(and (> n 0) (EmQV 3 (A v bass( < = v soprano) (1 + v) (Q (~1 v) (PO v))))) (= = chordtype dimseventh)) We again provide the heuristic in a more standard notation, for clarification: [n > 0 & (3 E { < , > })(Vv Ibass 5 v I soprano)[Q(p(n - l,v),p(n,v))] 3 chordtype = dimseventh]. ON THE USE OF CONSTRAINTS AND HEURISTICS FOR MUSIC GENERATION It is worthwhile to discuss certain practical issues related to the use of constraints and heuristics for music generation. We will first explain the motivation behind the use of constraints and heuristics for algorithmic production of music. THE MOTIVATION BEHIND CONSTRAINTS AND HEURISTICS A composition is written incrementally, typically from left to right in a direct fashion for short pieces, or perhaps as a sequence of successively refined plans for large-scale works. At each stage of the composition, the composer either decides to add an item (e.g. a chord, a phrase, or a plan for a movement, assuming a traditional idiom) to the partial composition in the hope of achieving the best completion of the composition, or decides that the partial composition needs revising, and makes a sequence of erasures and changes in the previously written parts of the composition in order to make the composition ready for extension again. Given a partial composition x and an item y, the question where “x is acceptable, and one of the best ways to extend x is to add y to it” holds for (x,y) can be answered by a composer with a limited degree of accuracy and consistency; similarly, for a given acceptable partial composition x, the composer can find items y such that this question can be answered positively for (x,y). However. the set of pairs (x,y) for which the answer is yes, which can be called the extension set, is difficult to define with mathematical rigor. Moreover, the extension set does not remain constant between styles and historical periods, and evolves even during the course of the composition of a single piece. The general approach of this research was to select a relatively fixed style, the Bach chorale, attempt to approximate the extension set with a precise definition, and then use the precise definition in a AN EXPERT SYSTEM FOR HARMONIZING CHORALES 167 computer algorithm for generating music in that style.” Inspired by our own experience with a strict counterpoint program [17,18] and the recent AI research in expert systems, we have designed the present knowledge-based method for describ- ing the extension set, which appears to work, and succeeds in generating nontrivial music that is of some competence by educated musician standards. In the following sections we will discuss the general problems associated with the constraints and heuristics used in this knowledge-based method, and also describe the possible sources for constraints and heuristics. THE DIFFICULTY OF USING ABSOLUTE RULES TO DESCRIBE REAL MUSIC A major part of the knowledge of the chorale program is based on constraints, or absolute rules in other words. Absolute rules, such as those expressed by treatises on harmony, counterpoint, or fugue d ‘Pcole, assert, in a very inflexible manner, which pieces are acceptable, and which others are not. For artificial styles such as harmony, counterpoint, and fugue exercises, absolute rules are part of the usual musical knowledge and practice. However, some problems are encountered when we try to describe a real style of music, rather than artificial style, with absolute rules. The rules in the book do not work, and many treatises mention to what extent great composers break the rules [52,42]. Schenker [66] provides some modifications of traditional rules on fifths and octaves, so that the liberties taken by the masters are considered acceptable when they no long exist in a middleground reduction; unfortunately, Schenker’s rules do not meet the level of precision typically found in a traditional treatise. A number of treatises on composition attempt to describe the free compositional style [12,16,8] ([50,67] could also be considered in this category), but such treatises do not characterize the existing style of any master, and they often reflect a particular normative view of music. In general, prescribing rules for the music of a master is recognized to be undoable. Nevertheless, this fact alone does not imply that good approximations of a real style cannot be obtained with the aid of a judiciously chosen set of such rules: for example, [37], which describes real 16th-century counterpoint rather than school exercises, is a treatise in this direction. Moreover, absolute rules are a powerful software tool in an expert system; although they appear to impose stringent demands on the knowledge-base designer, in reality they are (in our opinion) conceptually clearer and easier to handle than assertions with numerical truth values [80,69,6] in an application as complex and as subjective as the present one. We therefore decided to take a constructive approach toward the use of absolute rules for describing a real style of music, namely the Bach chorales. HOW ABSOLUTE RULES CAN BE FOUND We will now discuss the sources from which absolute rules are obtained. A good source for finding absolute rules is the traditional harmony treatise. In the chorale program, we used a number of treatises, such as [47,14,41,4], as useful “Mechanizing the evolution of the extension set over time is a potentially more difficult problem that has not been attacked in the present research. 168 KEMAL EBCIOGLU FIGURE 1. Chorale no. 73 (Bach) points of departure, despite their anachronism. However, since treatises are usually tailored for school exercises rather than for real Bach chorales, rules from such books had to be amended to fit the actual chorales themselves. For example, the familiar rule about parallel fifths had to be amended to allow a diminished fifth followed by perfect fifth when the parts are moving by ascending step, because of the consistent occurrence of these fifths in the chorale style.” We see an example of such an occurrence in chorale no. 73 shown in Figure 1 (at the phrase ending, between alto and soprano).lg Unfortunately, if we try to make our rules comprehensive, such amendments seem never to reach an end. We would have liked to have absolute rules that would accept every chorale. However, attempting to do so results in the unwieldy prolifera- tion of allowable, conditional violations of some rules. Moreover, there are cases where the extenuating condition for the violation is hard to find. Consider the fifths by contrary motion in chorale no. 18 in Figure 2 (between tenor and soprano). We found it difficult to explain this liberty (except perhaps by the remote extenuating effect of the first inversion of the dissonant dominant seventh chord).*’ In certain cases, we therefore used our own judgement in deciding where to cut the list of conditional violations. Another source of rules is empirical observation and inductive reasoning on the chorales themselves. For example, most chorale phrases end on a chord with the root doubled, which suggests an implicit absolute rule. Such rules are also not without exception, and it is again impractical to codify the precise reasons for all the exceptions. To distinguish which exceptions are truly representative of the style, it is necessary to use musical judgement in order to make an educated guess as to where Bach did what he wanted to do and where he did what he had to do. For example, in chorale no. 100 given in Figure 3, this rule is violated by doubling the third in the phrase ending. The reason is obvious: doubling the root would have resulted in a parallel octave between the alto and bass, or some other unacceptable violation. “It is interesting to note that C. P. E. Bach [2] allows such fifths in the nonextremal parts, declaring them to be better than descending fifths where the first is diminished. He also allows quite a few other combinations of the diminished and perfect fifth, not often seen in the chorales. Reference [49] is another treatise which correctly points out that the ascending diminished-fifth-perfect-fifth sequence is legal in the Bach chorale style. lqA’All the Bach chorale examples in this article have been reprinted by permission from C. S. Terry (ed.), The Four-part Chords of J. S. Buch, copyright 1929 and 1964. Oxford University Press. “‘However, it appears that these fifths are not an oversight after all: they also occur in chorale no. 352 in the same context. They could be a license of the style when a descending fifth in the soprano is harmonized in this specific manner. AN EXPERT SYSTEM FOR HARMONIZING; CHORALES 169 FIGURE 2. Chorale no. 18 (Bach) Moreover, it is desirable to keep the cadence as it is because of the nice linear progression in the tenor. However, this exception is not a good candidate for inclusion in the program, since it would be marginal in loyalty to the style and would require complex extenuating conditions to be specified, to prevent the backtracking algorithm from using this licence in inappropriate contexts. So we overruled Bach in this case and declared that a phrase should end with the root doubled as an absolute rule, with exceptions allowing the fifth to be doubled in a IV’-V ending in the minor mode (see chorale no. 51 for an example), and the third to be doubled in a V-VI ending (commonplace). The arbitrariness of this constraint definition mechanism needs some elucidation. It would probably be easy to reduce the corpus of chorales to a tractable size and write constraints that accept all members of the corpus, thus making the method more scientific (it would probably be more difficult to do the same without reducing the corpus). However, we know by experience that the property of exact agreement of the constraints with the corpus per se would be of little help in improving the quality of the music produced by the knowledge base (Baroni and Jacoboni [3] make a similar observation). Moreover, we feel that regarding music knowledge-base design as more of an art, and giving full liberty to the knowledge-base designer’s goodwill and musical intuition in both the heuristics and the constraints, would produce more competent programs, without having to restrain the corpus of music that the knowledge-base designer would draw upon. We are not saying that it is undesirable to have a rule set that would exactly characterize a large musical corpus, similar to a theory that explains the outcomes of chemical experiments; however, musical pieces apparently do not enjoy the simplicity of some other natural phenomena, and for the time being we may have to stay with inexact rule sets rather than have none at all. FIGURE 3. Chorale no. 100 (Bach) 170 KEMAL EBCIO<:rLU THE SIGNIFICANCE OF HEURISTICS The second kind of difficulties faced by the music knowledge-base designer is related to finding adequate heuristics. The purpose of heuristics is to estimate, at each step, which among the possible ways of extending the partial chorale will lead to its best completion. Heuristics are very important, since programs without heuristics, which are based solely on absolute rules and random selection, tend to quickly get trapped in a very unmusical path, and generate gibberish instead of music.21 In the chorale program, we are using a natural extension of a heuristic technique we used in an early strict counterpoint program [17,18], which was very successful for its purpose. Note that in theory it would be possible to characterize any finite set of “best” solutions solely with absolute rules. In fact, a research effort for generation of Bach chorale melodies [3] has used the absolute-rule approach. However, heuristics have a different and more human-composer-like flavor of describing what constitutes a good solution, because heuristics, in contrast to constraints, are rules that are to be followed whenever it is possible to follow them. 22 The main advantage of heuristics over pure absolute rules and random search is the following: heuristics lead the solution path away from a large number of unmusical patterns; if there were no heuristics, unmusical patterns would probably be generated by the bundle, and would have to be painstakingly diagnosed and then carefully ruled out with potentially complex constraints. Thus, a system based on heuristics can get away with less constraints and/or less complex constraints than a similar system based on random search. However, in case the research goal itself is to make a fair measure- ment of the musical power of a set of absolute rules, then heuristics cannot be used, since heuristics strongly bias the solution path toward a particular style, whereas random search can produce a relatively unbiased selection among all the possible solutions that are accepted by the rule set. AN ALGORITHMIC PROBLEM WITH HEURISTIC ORDERING As described in the previous section on the operational details, heuristics are strictly prioritized in the chorale program, and tied to a backtracking scheme. This strict priority scheme is easy to understand and debug, and avoids dealing with problems associated with arbitrary numerical weighting schemes. It is also quite rich and expressive, because the prioritized heuristics have the generality of BSL formulas. “It should be noted that there are contexts where music generated by extremely naive random-num- ber generation methods [79], let alone absolute rules, is not necessarily gibberish; it may offer a refreshing sense of liberation from the traditional or modern constraints and cliches, and a sense of beauty from a sophisticated aesthetic viewpoint, in fact, a natural evolution of Western art music through the centuries. In this particular research we are obviously looking at the problem of computer music from a stubbornly traditional aesthetic point of view; in real life, we do not necessarily take such an approach. However, we feel that our present approach is useful, because answering the unanswered questions in computer generation of traditional music could also help to answer the many (nowadays unasked and) unanswered fundamental questions in the field of algorithmic composition. “In fact, it would not be desirable to always satisfy a heuristic such as continuing a linear progression, because a piece consisting merely of scales could ensue from such a practice. Heuristics are therefore only meaningful in conjunction with constraints that prevent them from being satisfied all the time. AN EXPERT SYSTEM FOR HARMONIZING CHORALES 171 However, there is an algorithmic problem associated with the stack-based back- tracking scheme and the heuristic ordering. At a given step, the heuristics may make an erroneous estimate; i.e., the item that the heuristics choose among the possibili- ties for adding to the chorale may not be on the path that leads to the best completion of the partial chorale. The reason such an error is possible is that heuristics typically depend only on a simple local property of the partial solution and on the item to be added to it. If the erroneous choice leads to a blind alley, the choice will eventually be undone by the backtracking mechanism. However, a locally good choice dictated by the heuristics may also later force a mediocre passage, which could have been avoided by a different, perhaps locally bad choice; or a locally good choice may force the program to miss a cliche or other “desirable” progression, which would not have been missed by a different, perhaps locally bad choice. Although such problems could be remedied by maintaining a priority queue of partial chorales, sorted by a numerical evaluation function [57,58,43], and/or by using heuristics with several levels of lookahead, we preferred to keep the concep- tual simplicity of BSL’S stack-based mechanism, and we used additional constraints in an attempt to provide remedies for these problems. In the cases where we understood the precise pattern that made a passage mediocre, we made mediocre passages either unconditionally forbidden or conditionally forbidden, via con- straints of the form “pattern x is not allowed” or “if pattern x could have been avoided, then it should have been avoided”, respectively. For the case where a locally good choice misses a future cliche opportunity, whereas a locally bad choice does not, we used a conditional backtracking scheme to provide a selective degree of heuristic lookahead: Whenever there is an opportunity for a cliche progression, the chorale program first prefers to generate that cliche and enters a cliche state; while the program is in that state, the cliche must be at least partially fulfilled; if this is not possible, the program will backtrack to the originating step, where it will not enter the same cliche state, and perhaps will choose what is best according to the local heuristic criteria. HOW HEURISTICS CAN BE FOUND Now we come to the problem of finding heuristics. One major source of heuristics is the preferences of general good counterpoint practice, such as moving by step rather than by skip, avoiding following a scalar motion with a skip in the same direction, etc., which a counterpoint treatise will tell us in some probably unalgorithmic recipe (e.g. [40]). The knowledge-base designer FIGURE 4. Chorale no. 22 (Bach) 172 KEMAL EBCIOtiLU must possess the minimal ability to make such preferences precise and algorithmic in a reasonable way, using his or her musical judgement. Another source of heuristics is the chorales themselves. Heuristics from this source are style-awareness heuristics, and roughly correspond to the informal knowledge acquired by a composer when he or she sets out to understand a style. These heuristics are developed by observing a very broad range of chorales. Examples are to follow a suspension with another in the same part, and to prefer certain recurring patterns-we can call them Bach chorale cliches. We see in chorale no. 22 (Figure 4) an example of the repeating suspension pattern in the alto in the first measure, and in the fourth measure we see a cliche chord progression, a cadence cliche in this case. The chorale program currently knows 11 such cliche progressions. However, getting such recurring patterns to be used is a different and less predictable matter within the extremely intense computation of chorale genera- tion, since whenever the use of a pattern is seemingly appropriate, it may result unexpectedly in e.g. a forbidden melodic motion in an inner voice. (Being more vulnerable to accusations of unmusicality, our program is more concerned with melodic motion in the inner parts than Bach is.) A third and valuable source is hand simulation of an algorithm in an attempt to generate specific chorales exactly as written by Bach. This exposes all details, causes one to find the plausible reasons underlying each choice, and allows postulating priorities for heuristics. For example, the heuristics behind the first two measures of chorale no. 210, Jesu meine Freude (Figure 5) can be explained as a concern to move by step and continue a linear progression in the bass and in the other parts, and to prefer a cadence cliche. The layout of the chords is affected by a preference to use triads rather than seventh chords and to double the root in triads. The inserted diminished seventh on the weak eighth beat of the third chord is explained as a desire at the fill-in view to change the plagal progression IV-I in the skeleton to one of the more desirable VII-I or V-I progressions. The reasons for the suspension in the first measure of the bass are a concern to hide the second inversion of a chord and a concern to continue eighth-note movement. We have made these concerns heuristics in the chorale program. We can see many interesting applications of these particular heuristics in very different contexts in the examples of [19]. Unfortunately, there are cases where we cannot find any plausible reason for choosing certain possibilities rather than others, or sometimes a choice that appears to be locally bad is made by Bach. Such situations tend to agree with the backtracking search model. However, because of the labor-intensive nature of such very detailed hand simulation, conclusive results for validating the back- tracking search model of composition can only be reached by drastically restricting the corpus. We were not primarily interested in validating a cognitive model for a FIGURE 5. Chorale no. 210: “Jesu Meine Freude” (Bach). AN EXPERT SYSTEM FOR HARMONIZING CHORALES 173 composer, so we did not push far in this direction. However, we feel that explicating the decisions made during such an algorithmic resynthesis of a piece could be an instructive future research direction to pursue in the field of music analysis, and is likely to yield profound results. EMOTIONAL CONTENT OF COMPUTER-COMPOSED MUSIC In this section, a final remark must be made about some common misconceptions about the “emotional content” of music generated by computers. Often it is taken for granted that mechanical music cannot have emotional content. Unfortunately, existing computer-generated compositions in traditional styles sometimes confirm this opinion. However, the factor responsible for the apparent lack of feeling is more often than not an inadequate program which lacks the knowledge base to characterize a sufficiently sophisticated style. In all cases of practical interest, the set of pieces in the desired style with the desired feelings is finite; thus there is no inherent theoretical problem in an algorithmic description of music with emotional content.23 A study by Meyer [51] ties emotion to concrete musical events, such as the delaying of expectations of chordal and melodic progressions. The whole burden is therefore on the expert-system designer, who must algorithmically encode the emotional content in rules and/or heuristics, where we are assuming that the set of desired solutions largely overlaps the set of solutions with emotional content. This is no small burden, however. In fact, actual composition of music in any decent style is invariably easier than characterizing precisely what that style is in terms of concrete attributes, and such characterization attempts appear to be limited to styles that are well understood. What is well understood is of course strictly dependent on the competence of the knowledge-base designer. However, each knowledge-base designer may also have a limit that applies to him or her: sometimes compositional ideas discovered after lengthy unconscious search are not well understood, and these ideas, like sufficiently hard proofs, unfortunately tend to be the most valuable ones. Thus, it is unknown to what extent human compositional ability can be algorithmi- cally replicated. However, there is no obstacle to establishing higher and higher standards in algorithmic composition, in fact, substantially higher than the existing ones. Moreover, large knowledge bases in an efficient computing environment have -‘Hofstadter [35,36], perhaps overly impressed by an older topic in recursive function theory, believes that works of art must be a productive set, i.e., given any algorithm, a work of art that is not generated by this algorithm can be found, or the algorithm can be shown to generate a non-work-of-art. For the case of music, we feel that the set of all “pieces” that can be encoded via digital recordings of some fixed sampling rate, and that take less than a reasonable time limit, is a satisfactory superset of the set of interesting music. The finiteness of this nevertheless huge set does not of course make the discouen, of a practical algorithmic description of music less difficult: it merely points out that productiveness is an incorrect model of the true difficulty. Also, even if we momentarily accept that we are dealing with an infinite set. Hofstadter’s choice of a productioe set (rather than, say, an WPVW~~ set) actually works against the point he wants to make: a productive set has an infinite recursively enumerable subset [64]. which by Hofstadter’s hypothesis would mean that there exists an algorithm which will produce infinitely many diRerent works of art, but never a non-work-of-art. Note, however, that the conjecture that art objects. like the true sentences of a sufficiently complex formal system, could be a productive set was indeed elegant in its own right when the repercussions of Giidel’s incompleteness theorem were strong [54]: thus, this particular stance of Hofstadter’s is marred primarily by its bad timing. 174 KEMALEBCIOtiLU an encouraging synergistic effect that sometimes transcends the naiveness of the individual rules and heuristics (43,441. CONCLUSIONS In this paper, we have described a knowledge-based heuristic search technique for generating tonal music, which seems to work, and which produces musical results. While it is imprudent to make claims about the accuracy of this algorithmic model for the human composition process, our work indicates that heuristic search, coupled with large, complex knowledge bases, can be effective for the purpose of music generation. Although a heuristic evaluation function was used in a very early program for generating simple serial music [29], research in algorithmic composition has since then concentrated mainly on nonheuristic search techniques, such as random selection of attributes of notes according to statistical distributions [79,34], or generation of music through terse formal grammars [38] or other mathematical artifacts [39] (these are mostly aoant-garde approaches; computer-generated tonal music has somehow failed to be popular among computer musicians, perhaps because of its reactionary overtones). However, the knowledge-based heuristic search technique that we described here, which is based on [17,18], is not at all restricted to tonal music: with the proper set of rules, it can certainly be used for more general algorithmic composition as well. We thus hope that our techniques can be of help to computer-music researchers who may be looking for alternative approaches to algorithmic composition. In this paper, we have also described a multiple-viewpoint knowledge representa- tion technique, which is analogous to the multiple levels of knowledge in the Hearsay-II expert system [23], but which is based on first-order predicate calculus, and which uses different primitive functions and predicates for representing each viewpoint. Another point of AI interest concerning the present research is the efficient compiled representation of musical knowledge in the CHORAL knowledge base, which can be seen as a step toward the goal of knowledge compilation [32], a goal which has been predicted to be an important one for the ambitious expert systems of the future. Finally, we have adopted a streamlined approach to the design of a complex expert system, and we have advocated the reduction of the semantic gap. Although a streamlined approach is not easy to defend with a “ logical” argument, computer-science experience with large software/hardware systems, as well as considerations of mathematical tractability, seems to suggest that it is a desirable approach. In this paper, we have also described a new logic-programming language BSL, which was designed to implement the CHORAL system. Logic programming is a clearly desirable means to implement large AI applications. It provides a fuller understanding of the declarative meaning of the knowledge in the program and the relationship of this declarative meaning to the actual execution of the program (this is not at all true for the OPS family and many other inference engines [32], where no adequate formalism exists for understanding the declarative meaning of the rules). Also, the knowledge-representation techniques in an expert system written in a logic-programming language can be more easily made to follow a formal discipline, since the properties of the desired outputs of the expert system are specified with AN EXPERT SYSTEM FOR HARMONIZING CHORALES 175 precise logic formulas. But we feel that logic-programming research should not confine itself to a narrow PROLOG-and-variants paradigm, since PROLOG is not the only medium to achieve such benefits of logic programming. In the context of the present research, we have designed and implemented BSL, a logic-programming language radically different from PROLOG. BSL has a sound formal basis: it allows the programmer to use the concepts of first-order logic, including universal and existential quantifiers, and is not limited to the clausal form of logic; moreover, it allows the benefits of an efficient Algol-class language for coping with substantial applications. We hope that our work with BSL will be of use to researchers who would like to use the concepts of logic in computation-intensive applications. FIGURE 6. Chorale no. 128 (Bach). FIGURE 7. Chorale no. 128 (CHORAL). 176 KEMAL. ERCIOC?LU FIGURE 7. Chorale no. (continued) (CHORAL). 128 AN EXPERT SYSTEM FOR HARMONIZING CHORALES 177 n h u UW FIGURE 8. Chorale no. 286 (CHORAL). 178 KEMAL EBCIOCLU UI’ I w FIGURE 8. Chorale no. (continued) (CHORAL). 286 FIGURE 9. Chorale no. 68 (CHORAL). AN EXPERT SYSTEM FOR HARMONIZING CHORALES 179 FIGURE 9. Chorale no. 68 (continued) (CHORAL). 1 , 4 I FIGURE 10. Chorale no. 75 (CHORAL). h Y ’ AN EXPERT SYSTEM FOR HARMONIZING CHORALES 181 FIGURE 11. Chorale no. 48 (Bach). FIGURE 12. Chorale no. 48 w (CHORAL). 182 KEMALERCIO<:LU FIGURE 12. Chorale no. 48 (continued) (CHORAL). APPENDIX This Appendix presents actual examples of harmonizations. The CHORAL harmo- nization of chorale no. 48 is shown in Figure 12, and Bach’s is shown in Figure 11 for comparison. The CHORAL harmonization of chorale no. 68 is shown in Figure 8, and that of no. 75 is shown in Figure 9. The CHORAL harmonization of chorale no. 128 is shown in Figure 10, and Bach’s is shown in Figure 11 for comparison. The CHORAL harmonization of chorale no. 286 is shown in Figure 8. I am indebted to my thesis advisor, the late Professor John Myhill, for encouraging me to study computer music and getting me interested in the mechanization of Schenkerian analysis. I would also like to thank Professor Pat Eberlein for her encouragement, which led to the receipt of a two-year NSF grant for this project. The research environment and the computing resources at IBM Yorktown Heights were invaluable during the later stages of this research. I am in particular thankful to John Darringer and Abe Peled for their support, to Jean-Louis Lassez for helpful discussions on logic programming, and to Josh Knight and Phil Perry for their unselfish help with the software problems I had in porting BSL and CHORAL to the IBM environment. I am also grateful to the referees for their useful comments on this paper. REFERENCES 1. Aho, A. V. and Ullman, J. V., Principles of Compiler Design, Addison-Wesley, 1977. 2. Bach, C. P. E., Essay on the True Art of Playing Keyboard Instruments (W. J. Mitchell, transl.), Norton, 1949. 3. Baroni, M. and Jacoboni, C., Verso una Grammatica delia Melodia, Univ. Studi di Bologna, 1976. 4. Bitsch, M., P&is d ‘Harmonie Tonale, Alphonse Leduc, Paris, 1957. 5. Bruynooghe, M. and Pereira, L. M., Revision of Top-Down Logical Reasoning through Intelligent Backtracking, Report 8/81, Centro di Informatica da Univ. Nova de Lisboa, Mar. 1981. 6. Buchanan, B. G. and Shortcliffe, E. H. (eds.), Rule Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, Addison-Wesley, 1984. 7. Cohen, J., Non-deterministic Algorithms, Comput. Surveys 11, No. 2 (June 1979). 8. Czerny, C., School of Practical Composition (John Bishop, transl.), Da Capo, New York, 1979. 9. Davis, R. and Ring, J., An Overview of Production Systems, in: Elcock and Michie (eds.), Machine Intelligence 8, Wiley, 1976. AN EXPERT SYSTEM FOR HARMONIZING CHORALES 183 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27 28. 29. 30. 31. 32. 33. 34. 35. de Bakker, J., Mathematical Theory of Program Correctness, North Holland, 1979. de Kleer, J. and Williams, B., Back to Backtracking: Controlling the ATMS, in: Proceedings of the Fifth National Conference on ArtiJicial Intelligence, 1986. D’Indy, V., Cows de Composition Musicale, Durand, Paris, 1912. Doyle, J., A Truth Maintenance System, Arttjkial Intelligence 12:231-272 (1979). Dubois, Th., Trait; d ‘Harmonie Theorique et Pratique, Heugel, Paris, 1921. Durand, E., Trait& de I’Accompagnement au Piano, Alphonse Leduc, Paris, n.d. (ca. 1890?). Durand, E., Trait& de Composition Musicale, Alphonse Leduc, Paris, nd. (ca. 1898?). Ebcioglu, K., Strict Counterpoint: A Case Study in Musical Composition by Computers (in English), M.S. Thesis, Dept. of Computer Engineering, Middle East Technical Univ., Ankara, Sept. 1979. Ebcioglu, K., Computer Counterpoint, in: Proceedings of the 1980 International Computer Music Conference (Queens College, New York), Computer Music Assoc., San Francisco, 1981. Ebcioglu, K., Report on the CHORAL Project: An Expert System for Harmonizing Four-Part Chorales, Research Report RC12628, IBM Thomas J. Watson Research Center, Yorktown Heights, N.Y., Mar. 1987. (This is a revised version of the author’s Ph.D. dissertation, An Expert System for Harmonization of Chorales in the Style of J. S. Bach, Technical Report TR 86-09, Dept. of Computer Science, S.U.N.Y. at Buffalo, Mar. 1986.) Ebcioglu, K., An Efficient Logic Programming Language and its Application to Music, in: Proceedings of the 4th ICLP, May 1987. Ebcioglu, K., A Compilation l’echnique for Software Pipelining of Loops with Condi- tional Jumps, in: Proceedings of the 20th Annual Workshop on Microprogramming (MICRO-20), Dec. 1987. Ellis, J. R., Bulldog: A Compiler for VLIW Architectures, MIT Press, 1986. Erman, L. D. et al., The Hearsay-II Speech Understanding System: Integrating Knowl- edge to Resolve Uncertainty, Comput. Surveys 12, No. 2 (June 1980). Fagin, B. and Dobry, T., The Berkeley PLM Instruction Set: An Instruction Set for Prolog, Report UCB/CSD 86/257, Computer Science Division (EECS), Univ. of Califor- nia at Berkeley, Sept. 1985. Feigenbaum, E. A., Themes and Case Studies in Knowledge Engineering, in: Donald Michie (ed.), Expert Systems in the Micro-electronic Age, Edinburgh U.P., 1979. Floyd, R., Nondeterministic Algorithms, J. Assoc. Comput. Mach. 14, No. 4 (Oct. 1967). Forgy, C. and McDermott, J., OPS: A Domain Independent Production System Lan- guage, in: Proceedings of the Fifth International Joint Conference in ArtiJicial Intelligence, 1977. Gergely, T. and Szots, M., Cuttable Formulas for Logic Programming, presented at 1984 International Symposium on Logic Programming, Feb. 1984. Gill, S., A Technique for the Composition of Music in a Computer, Comput. J. 6, No. 2 (July 1963). Harel, D., First Order Dynamic Logic, Lecture Notes in Comput. Sci., Springer-Verlag, 1979. Hayes-Roth, B., A Blackboard Architecture for Control, Art$cial Intelligence 26:251-321 (1985). Hayes-Roth, F., Waterman, D., and Lenat, D. B. (eds.), Building Expert Systems, Addison-Wesley, 1983. Hennessy et al., The MIPS Machine, in: Digest of Papers-Compcon Spring 82, Feb. 1982. Hiller, L., Music Composed with Computers: A Historical Survey, in: H. B. Lincoln (ed.), The Computer and Music, Cornell U.P., 1970. Hofstadter, D. R., Giidel, Escher, Bach: An Eternal Golden Braid, Basic Books, 1979. 184 KEMAL EBCIOGLU 36. 37. 38. 39. 40. 41. 42. 43. 44. 45. 46. 47. 48. 49. 50. 51. 52. 53. 54. 55. 56. 57. 58. 59. 60. 61. 62. 63. 64. 65. 66. Hofstadter, D. R., Metafont, Metamathematics, and Metaphysics, Technical Report 136, Computer Science Dept., Indiana Univ., Dec. 1982. Jeppesen, K., Counterpoint: The Polyphonic Vocal Style of the Sixteenth Century (Glen Hayden, transl.), Prentice-Hall, 1939. Jones, K., Compositional Application of Stochastic Processes, Comput. Music J. 5, No. 2 (1981). Kendall, G. S., Composing from a Geometric Model: Five-Leaf Rose, Comput. Music J. 5, No. 4 (1981). Koechlin, Ch., P&is des Regles de Contrepoint, Heugel, Paris, 1926. Koechlin, Ch., Trait& de I’Hurmonie, Vol. I, Editions Max Eschig, Paris, 1928; Vol. II, 1930; Vol. III, 1928. Koechlin, Ch. .&ude sur I ‘lkriture de la Fugue d ‘&Cole, Editions Max Eschig, Paris, 1933. Lenat, D. B., A.M.: An Artificial Intelligence Approach to Discovery in Mathematics and Heuristic Search, Report STAN-CS-76-570, Dept. of Computer Science, Stanford Univ., July 1976. Lenat, D. B., The Nature of Heuristics, Arttjkial Intelligence 19 (1982). Lerdahl, F. and Jackendoff, R., Toward a Formal Theory of Tonal Music, J. Music Theory 21 (1977). Lerdahl, F. and Jackendoff, R., A Generative Theory of Tonal Music, MIT press, 1983. Louis, R. and Thuille, L., Hurmonielehre, C. Griininger, Stuttgart, 1906. Our source is an unpublished Turkish translation by N. K. Akses. Lovelock, W., First Year Harmony, Hammond, London, n.d.; Third Year Harmony, 1956. McHose, A. I., The Contrapuntal Harmonic Technique of the 28th Century. Prentice-Hall, 1947. Messiaen, O., Technique de Man Languge Musicul, Alphonse Leduc, Paris, 1944. Meyer, L. B., Emotion and Meaning in Music, Univ. of Chicago Press, 1956. Morris, R. O., The Oxford Harmony, Oxford U.P., 1946. Myers, G. J., Advances in Computer Architecture, Wiley-Interscience, 1982. Myhill, J., Some Philosophical Implications of Mathematical Logic: Three Classes of Ideas, Rev. Metuphys. VI, No. 2, Dec. 1952. Newell, A. and Simon, H., GPS: A Program That Simulates Human Thought, in E. A. Feigenbaum and Feldman, (eds.), Computers und Thought, McGraw-Hill, 1963. Nicolau, A., Percolation Scheduling: A Parallel Compilation Technique, TR 85-678, Dept. of Computer Science Cornell Univ., May 1985. Nilsson, N., Problem Solving Methods in Artificial Intelligence, McGraw-Hill, 1971. Nilsson, N., Prmciples of Arttjicial Intelligence, Tioga, 1980. Patterson, D. A. et al., RISC-I: A Reduced Instruction Set VLSI computer, presented at Eighth Annual Symposium on Computer Architecture, May 1981. Pearl, J. (ed.), Special Issue on Search and Heuristics, Artificial Intelligence 21 (1983). Pereira, L. M. and Porto, A., Selective Backtracking for Logic Programs, Report l/80, Centro di Informatica da Univ. Nova de Lisboa, 1980. Radin, G., The 801 Minicomputer, in: Proceedings of the ACM Symposium on Architec- tural Support for Programming Languages and Operating Systems, Mar. 1982. Robinson, J. A., A Machine Oriented Logic Based on the Resolution Principle, J. Assoc. Comput. Much. 12 (1965). Rogers, H., Theory of Recursive Functions and Effective Computability, McGraw-Hill, 1967. Schenker, H., Five Graphic Analyses (Felix Saber, ed.), Dover, 1969. Schenker, H., Free Composition (Her Freie Satz) (E. Oster, trans. and ed.), Longman, 1979. ANEXPERTSYSTEM FORHARMONIZINGCHORALES 185 67. 68. 69. 70. 71. 72. 73. 74. 75. 76. 77. 78. 79. 80. Schillinger, J., The Schillinger System of Musical Composition, C. Fischer, New York, 1946. Schmidt, C. F. et al, The Plan Recognition Problem: An Intersection of Psychology and Artificial Intelligence, Artificial Intelligence 11, Nos. 1, 2 (Aug. 1978). Shortcliffe, E. H., Computer Based Medical Consultations: MYCIN, Elsevier, New York, 1976. Smith, D. C. and Enea, H. J., Backtracking in MIisp2, in: Proceedings of the Third International Joint Conference in Arttjicial Intelligence, 1973. Stallman, R. M. and Sussman, G. J., Forward Reasoning and Dependency-Directed Backtracking in a System for Computer-Aided Circuit Analysis, Arttjicial Intelligence 9 (1977). Stefik, M., Inferring DNA Structures from Segmentation Data, Artificial Intelligence 11 (1978). Sussman, G. J. and McDermott, D. V., From PLANNER to CONNIVER-A Genetic Approach, in: Proceedings of AFIPS 1972 FJCC, AFIPS Press, 1972, pp. 1171-1179. Sussman, G. J. and Steele, G. L., Constraints-A Language for Expressing Almost- Hierarchical Descriptions, Artificial Intelligence 14:1-39 (1980). Terry, C. S. (ed.), The Four-Voice Chorals ofJ. S. Bach, Oxford U.P., 1964. Turk, A. W., Compiler Optimizations for the WAM, in: Proceedings of the 3rd ICLP, 1986. Warren, S. H., Auslander, M. A., Chaitin, G. J., Chibib, A. C., Hopkins, M. E., and MacKay, A. L., Final Code Generation in the PL.8 Compiler, Report RC 11974, IBM T. J. Watson Research Center, 1986. Waterman, D. A., Adaptive Producton Systems, in: Proceedings of the Fourth Interna- tional Joint Conference in Artificial Intelligence, 1975. Xenakis, I., Musique-Architecture, Casterman, 1971. Zadeh, L. A., A Theory for Approximate Reasoning, in: J. E. Hayes, D. Michie, and L. I. Mikulich (eds.), Machine IntelIigence 9, Wiley, 1979. 
work_c366uj3kb5glveyyljzrpcezki	----	A decision-making framework for precision marketing A decision-making framework for precision marketing Zhen You a, Yain-Whar Si b, Defu Zhang a,⇑, XiangXiang Zeng a, Stephen C.H. Leung c, Tao Li a,d a Department of Computer Science, Xiamen University, Xiamen 361005, China b Department of Computer and Information Science, University of Macau, Macau, China c Department of Management Sciences, City University of Hong Kong, Hong Kong d School of Computer Science, Florida International University, Miami, FL, USA a r t i c l e i n f o Article history: Available online 20 December 2014 Keywords: Data mining Decision tree Forecasting Precision marketing Decision-making a b s t r a c t Precision marketing offers personalized customer service and is used to help enterprises increase their profits by means of high-efficiency marketing. This paper presents a novel decision-making framework for precision marking using data-mining techniques. First, this study presents a trend model to accurately predict monthly supply quantity; second, it uses a RFM (Recency, Frequency and Monetary) model to select attributes to cluster customers into different groups; third, it uses CHAID decision trees and Pareto values to identify important attribute values to distinguish different customer groups; and finally, it creates different supply strategies targeting each customer group. The objective of the proposed precision-making framework is to help managers identify the potential characteristics of different customer categories and put forward appropriate precision marketing strategies, which can greatly reduce inventory for every customer category. The real-world data from a company in China were collected and used in a case study to illustrate how to implement the proposed framework. This case study demonstrates that our proposed decision-making framework is efficient and capable of providing a very good precision marketing strategy for enterprises. � 2014 Elsevier Ltd. All rights reserved. 1. Introduction Due to the accelerated pace of economic globalization and increasing market competition, economic pressures and competi- tion have led enterprise managers to face the problem of choosing the right strategic decision-making policies for selling the right products to the right customers at the right time, such that the companies can increase their profits. Recently, it has been recog- nized that precision marketing has become a key means of gener- ating profit and is becoming increasingly important as customers become better informed about the products and their rights as con- sumers. The availability of customer data and transaction records provides better understanding of customers’ consumption behav- iors and preferences. In the increasingly competitive environment, enterprises have to create a decision-making model for precision marketing that provides appropriate strategies to manage the mar- ket positioning system for fulfilling their customers’ needs. The research motivation of this paper is stemmed from a real-world project. This project considers a marketing problem where the supplier or manufacturer provides different products for retail cus- tomers, of which some may sell well in some customer segments and some may not. Products that are not sold will be returned back to the supplier. Therefore, the supplier needs to find a good mar- keting strategy that minimizes goods in stock and satisfies the sup- plier, retailers and consumers. In recent years, the decision-making problems have received much attention due to a wide range of real-world applications. Many decision-making techniques have been proposed in literature. Chen and Wang (2009) presented multi-criteria optimization and compromise solutions. Saen (2010) developed a technique for order performance via similarity to ideal solution. Hsu, Chiang, and Shu (2010) developed a nonlin- ear programming for decision-making, and Lin, Chen, and Ting (2011) presented a Linear programming for decision making. When discussing the decision making in this context, it is important to pay attention to the role of artificial intelligence (AI). The decision support framework of AI is considered to be a major tool for obtaining information related to historical data col- lections using some artificial Intelligence technique such as genetic algorithm (GA), artificial colony optimization (Ghasab, Khamis, Mohammad, and Fariman, 2015), and other data-mining tech- niques (Chai, Liu, & Ngai, 2013). Nowadays, data-mining, which can extract useful customer information and discover the hidden customer’s behaviors from big data, has a great influence on http://dx.doi.org/10.1016/j.eswa.2014.12.022 0957-4174/� 2014 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. Tel.: +86 18959217108; fax: +86 0592 2580258. E-mail addresses: uzhen@foxmail.com (Z. You), fstasp@umac.mo (Y.-W. Si), dfzhang@xmu.edu.cn (D. Zhang), xzeng@xmu.edu.cn (X. Zeng), mssleung@cityu. edu.hk (S.C.H. Leung), taoli@cs.fiu.edu (T. Li). Expert Systems with Applications 42 (2015) 3357–3367 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.12.022&domain=pdf http://dx.doi.org/10.1016/j.eswa.2014.12.022 mailto:uzhen@foxmail.com mailto:fstasp@umac.mo mailto:dfzhang@xmu.edu.cn mailto:xzeng@xmu.edu.cn mailto:mssleung@cityu.edu.hk mailto:mssleung@cityu.edu.hk mailto:taoli@cs.fiu.edu http://dx.doi.org/10.1016/j.eswa.2014.12.022 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa guiding decision making and forecasting the effects of decisions. Guo, Yuan, and Tian (2009) used a hierarchical potential support vector machine (HPSVM) for supplier selection and improved the accuracy. Chang and Hung (2010) presented a rough set theory (RST) to analyze the rules of supplier selection and guide the deci- sion making. Chai et al. (2013) conducted a systematic review of literature about the application of decision-making techniques in supplier selection. They classified these techniques into three cat- egories: Multi-criteria decision making techniques (MCDM), math- ematical programming techniques (MP), and artificial intelligence techniques (AI). MCDM is a methodological framework that aims to provide decision makers a knowledgeable recommendation amid a finite set of alternatives (Chai et al., 2013). Wang, Wu, Wang, Zhang, and Chen (2014) developed two kinds of prioritized aggregation operators of IVHFLNs for MCDM. However, in practice, it is not an easy task to choose a good data mining tool since, each data-mining tool has its own advan- tages and disadvantages. For example, artificial neural networks (ANNs) involve too many hidden neurons and training parame- ters (Zhang, Jiang, & Li, 2005). Its disadvantages include the ‘‘black box’’ nature, greater computational burden and proneness to over-fitting. However, an advantage of ANN is that ANN has the ability to implicitly detect complex nonlinear relationships between dependent and independent variables. Decision trees are simple to use and easy to understand and they offer many advantages compared with other decision-making tools. How- ever, the disadvantages of decision trees include their instability and relatively low performance. Hence, there is no single best model for all the cases (Akın, 2015). Recently, researchers attempt to combine different models together by considering their respective advantages. Guo, Wong, and Li (2013) presented a multivariate intelligent decision-making mode which combined three different model to complete every phase of the decision making process. Tadić, Zečević, and Krstić (2014) developed a novel hybrid MCDM model that combines fuzzy decision making trial and evaluation laboratory model to provide support to deci- sion makers. Furthermore, Yan and Ma (2015) proposed a novel two-stage group decision-making approach to uncertain quality function deployment. To the best of our knowledge, the marketing problem consid- ered in this paper is still a new problem. The objective of this paper is to propose a decision-making framework for combining various data-mining algorithms to achieve precision marketing of real products. Real-world data that include historical monthly supply (quantity) and information of every customer were collected from a company in China. The goal is to find a model that can classify targeted customers and predict supply quantity and then provide a strategy for precision marketing, i.e., deciding the quantity of products that every store needs. Depending on the different char- acteristics and requirements of each phase of the marketing model, the decision-making framework uses four data-mining models/ algorithms, which are K-means algorithm, decision tree, Pareto ratio method, and RFM model, for decision-making. Overall, the purpose of this study is to generate, using data-mining techniques, a decision-making model for products’ precision marketing. The proposed decision-making framework is more accurate due to its integrated precision strategies which combine prediction models, clustering, and classifying model. Moreover, the proposed frame- work incorporates RFM model and Pareto ratio into the process of customer segmentation so that the generated strategies are more convincing. The rest of this paper is organized as follows. Section 2 introduces related works on common data-mining models and algorithms. Section 3 describes the methodology in our proposed framework briefly. Section 4 presents a case study and results of analyses. Finally, Section 5 concludes the paper. 2. Related works Since a single data-mining model may only be suitable for a specific problem, such as prediction or clustering or classification, in the proposed framework, we have combined four data-mining models or algorithms to derive a precision marketing strategy for enterprises. The literature on these four models or algorithms was reviewed below. 2.1. K-means algorithm Clustering is the process of grouping a set of physical or abstract items into classes of similar items where the groups are either meaningful or useful, or both. A well-known clustering algorithm is K-means, which was first proposed by MacQueen (1967). The accuracy of this algorithm depends on the initialization and the number of clusters (Mesforoush & Tarokh, 2013). The basic idea of K-means is to discover k clusters, such that the records within each cluster are similar to each other and distinct from the records in other clusters. K-means is an iterative algorithm: an initial set of clusters is defined and the clusters are repeatedly updated until no further improvement is possible (or the number of iterations exceeds a specified threshold). The K-means algorithm is widely used to pre-process data or for clustering because of its simplicity and efficiency (Mesforoush & Tarokh, 2013). K-means has been widely used to effectively identify the valuable customers and develop the related marketing strategies (Wei, Lee, Chen, & Wu, 2013; Mesforoush & Tarokh, 2013). In particular, Cheng and Chen (2009) use the RFM model and K-means to perform customer rela- tionship management, and the experimental results demonstrate that their proposed model is an effective method in customer value analysis. 2.2. Decision tree A decision tree is an efficient data mining algorithm with a strong explanatory capability (Zhang, Zhou, Leung, & Zheng, 2010). This study uses the Chi-squared Automatic Interaction Detector (CHAID) decision tree, a decision tree algorithm that is a type of database segmentation. The concept of CHAID was first published in Kass (1980). CHAID is used for prediction and classi- fication. Like other decision trees, CHAID’s advantages are that its output is highly visual and easy to interpret. A recent study indi- cates that CHAID is superior to judgment RFM for identification of likely responders (McCarty & Hastak, 2007). CHAID is similar to the RFM approach because it can identify the terminal nodes that will break even with respect to expected profit and costs. Therefore, CHAID was widely adopted used for segmenting cus- tomers and extracting rules that show the associations between the input and output variables (McCarty & Hastak, 2007; Chen & Wang, 2009; Mistikoglu et al., 2015). In addition, Coussement, Van den Bossche, and De Bock (2014) demonstrate that, when per- forming customer segmentation at different levels of uncertainty, CHAID outperforms RFM and logistic regression. Furthermore, CHAID can also assess its robustness to data accuracy problems. So in our study, we use CHAID to segment customers. 2.3. Pareto ratio The concept of Pareto ratio is from the ‘‘Pareto Principle’’ (also known as the 80–20 rule, the law of the vital few or the principle of factor sparsity) which states that for many events, roughly 80% of the effects come from 20% of the causes. This is a common rule of thumb in business. This distribution is relevant to entrepre- neurs and business managers, stating, for example, that 80% of the 3358 Z. You et al. / Expert Systems with Applications 42 (2015) 3357–3367 https://isiarticles.com/article/41312 
work_c3h37yn7zzcvxm3hom4avwnuza	----	NASA Technical Reports Server (NTRS) 
work_c3xo56fevbdmtf4nru5xrmau4q	----	Maintenance of belt conveyors using an expert system based on fuzzy logic Original Research Article Maintenance of belt conveyors using an expert system based on fuzzy logic D. Mazurkiewicz * Lublin University of Technology, Nadbystrzycka 36, 20-618 Lublin, Poland 1. Introduction New technological possibilities result in an extremely dynamic industrial development. Due to ever-increasing competition on the market, higher customer demands and pursuit of innovative products, manufacturers are compelled to imple- ment more and more advanced engineering solutions, particularly ones that lead to higher efficiency and reduced operational costs. In effect, machinery and technological devices are becoming more and more complex, which imposes higher demands on their users, including the necessity of applying suitable methods and techniques to ensure durability and reliability of frequently complex and elaborate production systems [1,2]. In recent years, conveyor belt transport systems have taken on a new significance due to numerous research studies on innovative design solutions for conveyor belts [3–10]. The a r c h i v e s o f c i v i l a n d m e c h a n i c a l e n g i n e e r i n g 1 5 ( 2 0 1 5 ) 4 1 2 – 4 1 8 * Tel.: +48 815384229. E-mail address: d.mazurkiewicz@pollub.pl a r t i c l e i n f o Article history: Received 13 November 2014 Accepted 19 December 2014 Available online 13 January 2015 Keywords: Computer-aided maintenance management systems Belt conveyor Fuzzy logic a b s t r a c t In recent years, conveyor belt transport systems have taken on a new significance due to numerous research studies on innovative design solutions. The application of these new developed solutions leads to considerable reduction in operational costs of transport systems, while ensuring their high reliability and service life at the same time. Nonetheless, there are still areas that pose challenge to both research and development. Typical chal- lenges are analyzed in this paper. The solution to the problems of conveyor transport maintenance can be the implementation of a system for estimation of technical condition of conveyor belt joints. It serves as a second level safety diagnostic system for transport. Besides real-time measurements, the system enables a long-term analysis of historic data for every single joint that makes up the conveyor belt loop, from the moment of its manufacture to the final operation. The effectiveness of a conveyor belt diagnostic system primarily depends on the use of a decision supporting system. With adequate inference rules applied, this system would increase the effectiveness and shorten the time of decision- making as well as verify generated signals. The above tasks can be performed by a suitable expert system that predicts values of the analyzed time series, using the predicted values and inference rules to verify any potential false alarm signals at the same time. The idea and algorithm of such an expert system were presented in this article as well. # 2015 Politechnika Wrocławska. Published by Elsevier Urban & Partner Sp. z o.o. All rights reserved. Available online at www.sciencedirect.com ScienceDirect journal homepage: http://www.elsevier.com/locate/acme http://dx.doi.org/10.1016/j.acme.2014.12.009 1644-9665/# 2015 Politechnika Wrocławska. Published by Elsevier Urban & Partner Sp. z o.o. All rights reserved. http://crossmark.crossref.org/dialog/?doi=10.1016/j.acme.2014.12.009&domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1016/j.acme.2014.12.009&domain=pdf http://dx.doi.org/10.1016/j.acme.2014.12.009 mailto:d.mazurkiewicz@pollub.pl http://www.sciencedirect.com/science/journal/16449665 http://www.elsevier.com/locate/acme http://dx.doi.org/10.1016/j.acme.2014.12.009 application of these new developed solutions leads, among others, to considerable reduction in operational costs of transport systems, while ensuring their high reliability and service life at the same time. Nonetheless, there are still areas that pose challenge to both research and development. As demonstrated in the studies [3,4,11], it is absolutely vital from a practical point of view that the service life of conveyor belt joints be made longer owing to the fact that these joints are crucial elements of the entire conveyor transport system. For instance, vulcanized and adhesive-bonded joints of textile- rubber belts are made up of lap joints, their theoretical durability for three separators amounting to only 67% of the belt's rated durability [4]. Moreover, the forces that affect the belt load and durability of both the belt and its joints vary during operation. For this reason, methods applied to this end, e.g. ones for determining non-stationary states, are often inaccurate, which results from lack of precise data and insufficient determination of boundary conditions. On the other hand, it is vital to optimize the belt tension force for various conveyor load conditions, as this should result in higher durability of both the belt and its joints. The application of a solution based on standard control algorithms seems, however, inadequate due to the nature of the problem as well as lack of both sufficient insight into the problem and measurement data. 2. Operation of an intelligent advisory system in considerable uncertainty conditions The solution to the above problems of conveyor transport maintenance can be the implementation of a system for estimation of technical condition of conveyor belt joints [11] that serves as a second level safety diagnostic system for transport. Besides real-time measurements, the system enables a long-term analysis of historic data for every single joint that makes up the conveyor belt loop, from the moment of its manufacture to the final operation. The monitoring device can also be converted into an autonomously function- ing instrument that can independently react to variable operational conditions and thus eliminate states that pose threat to the system's reliability, e.g. it can counteract belt failure by predicting the occurrence of operational parameters and their consequences. However, the vast number of data and information generated by the measuring system (which is the case when the monitoring covers a great deal of joints that are usually located on up to even several conveyors in a complex transport system) leads to a series of complications regarding interpretation of the collected data, which consequently hinders effective decision-making by the monitoring staff. In addition, given the specific nature of operation of conveyor belts (e.g. their exposure to high momentary loads), the measured discrete values of variations in length of an adhesive-bonded joint can generate false alarms that can result in unjustified stoppage of the transport system, unnecessary inspection of the joint or complicated and time-consuming examination of the generated signal by the operator. The system operator is then continuously required to take decisions, analyze a vast data set and observe changes therein, as well as predict the consequences of further changes in functional properties. This means taking decisions in uncertainty conditions [12,13], when we have no information about the probable occurrence of particular states (safe/ emergency/alarm) and can only predict values of the analyzed signals that have the form of a discrete time series. The conveyor transport system is an object that is affected by a number of factors. These factors are either imprecisely determined or difficult, if not impossible, to measure because of their random and incidental nature. This object is hard to describe owing to its dependence on numerous unpredicted variables that are often additionally difficult to be precisely determined. Failure-signaling symptoms can occur either once or many times; they can also have different intensity and nature. Another common problem here is lack of unambiguous information about the analyzed object and the required short time of a potential reaction. When equipping a comprehensive conveyor transport system with efficient and advanced monitoring and diagnostic systems, it is therefore necessary that sophisticated and intelligent tools be used to support the above. More specifically, the intelligent advisory system will be a dynamic system for the supervision of complex processes that are characterized by variable operating conditions and two time scales: micro- and macro-time scales [14], which is essential with regard to the entire life cycle of a single adhesive joint or a belt section that usually undergoes frequent relocations. 3. Inference supporting system based on fuzzy residuum evaluation in the diagnostics of conveyor belts The effectiveness of a conveyor belt diagnostic system primarily depends on the use of a decision supporting system. With adequate inference rules applied, this system would increase the effectiveness and shorten the time of decision- making as well as verify generated signals. The above tasks can be performed by a suitable expert system that predicts values of the analyzed time series, using the predicted values and inference rules to verify any potential false alarm signals at the same time. The prediction of variations in length of an adhesive-bonded joint involves classification and estimation of time series. Although all standard classification and estimation methods can, under certain conditions, be also employed for prediction, the most effective method for predicting time series involves the use of artificial neural networks [15–20]. A supervision system for adhesive-bonded joints of conveyor belts should therefore also include an intelligent expert system based on the popular knowledge representation and rule-based systems (Fig. 1). When measuring variations in length of belt joints by a computer-aided measurement system, it can be assumed that the observations are made at discrete and equal intervals of time t. For every joint we have data from the evaluation of its length DL in time moments t � 1, t � 2, . . ., t � n, which corresponds to the values of DLt�1; DLt�2; DLt�3; . . . ; DLt�n. The aim of the analysis is to determine values of elongation of the monitored conveyor belt joint DLtðwÞ, denoting the prediction in a moment t + w, i.e. in advance w, with the lowest possible mean deviations DLtþw � DLtðwÞ. When evaluating the potential a r c h i v e s o f c i v i l a n d m e c h a n i c a l e n g i n e e r i n g 1 5 ( 2 0 1 5 ) 4 1 2 – 4 1 8 413 https://isiarticles.com/article/46362 
work_c4lxulprjva3tf3xvg4q34lx3i	----	- 1 - Unsupervised Neural Models for Country and Political Risk Analysis Álvaro HerreroA, Emilio CorchadoB and Alfredo JiménezC A Department of Civil Engineering. University of Burgos C/ Francisco de Vitoria s/n, 09006, Burgos, Spain Tel: +34 947 25 9513 – Fax: +34 947 25 9000 ahcosio@ubu.es B Departamento de Informática y Automática, Universidad de Salamanca Plaza de la Merced s/n, 37008, Salamanca, Spain Tel: +34 616 449888 - Fax: +34 923 294514 escorchado@usal.es C Department of Economics and Business Administration, University of Burgos C/ Parralillos s/n, 09001, Burgos, Spain ajimenez@ubu.es Abstract. This interdisciplinary research project focuses on relevant applications of Knowledge Discovery and Artificial Neural Networks in order to identify and analyse levels of country, business and political risk. Its main goal is to help business decision-makers understand the dynamics within the emerging market countries in which they operate. Most of the neural models applied in this study are defined within the framework of unsupervised learning. They are based on Exploratory Projection Pursuit, Topology Preserving Maps and Curvilinear Component Analysis. Two interesting real data sets are analysed to empirically probe the robustness of these models. The first case study describes information from a significant sample of Spanish Multinational Enterprises (MNEs). It analyses data pertaining to such aspects as decisions over the location of subsidiary enterprises in various regions across the world, the importance accorded to such decisions and the driving forces behind them. Through a projection-based analysis, this study reveals a range of different reasons underlying the internationalization strategies of Spanish MNEs and the different goals they pursue. It may be concluded that projection connectionist techniques are of immense assistance in the process of identifying the internationalization strategies of Spanish MNEs, their underlying motives and the goals they pursue. The second case study covers several risk categories that include task policy, security, and political stability among others, and it tracks the scores of different countries all over the world. Interesting conclusions are drawn from the application of several business intelligence solutions based on neural projection models, which support data analysis in the context of country and political risk analysis. Keywords: neural visualization models, exploratory projection pursuit, unsupervised learning, country and political risk, business intelligence, knowledge extraction. *Manuscript http://ees.elsevier.com/eswa/viewRCResults.aspx?pdf=1&docID=8393&rev=0&fileID=75089&msid={DD799DFB-3CEC-499F-A2BF-C2F62D0DC871} - 2 - 1 Introduction Corporate decision-makers in a broad range of firms that seek overseas investment opportunities have understood for a number of years that in-depth knowledge of a foreign country’s economy is not enough in itself. It is also important to understand the dynamic forces that shape these countries’ politics. This is crucial for emerging markets, where market outcomes are influenced at least as much by policies as by economic factors. A company is considered a Multinational Enterprise (MNE) whenever it controls activities that generate added value in two or more politically independent geographic areas; in other words, whenever it coordinates and controls subsidiaries in one or more foreign countries (Durán, 2001). To that end, companies can make cross-border acquisitions of existing foreign enterprises, make greenfield investments or commit themselves to an alliance with other multinational enterprises or with a local partner to undertake joint activities. Subsidiaries or cross-border business units emerge in this way, which are totally or partially controlled by the parent company. Internationalization is a decision that enterprises must take with increasing frequency, as competition in many sectors due to globalization, obliges companies to enter international markets in the search for new markets and lower operating costs. Thus, the decision to become a MNE and the challenge of successfully undertaking such a transformation are more relevant than ever. Furthermore, the structural changes involved in economic development are systematically related to direct investments that are either received or made from a country or a geographic area (Lall, 1996), and at present this direct foreign investment is on the whole made by multinational enterprises; a good enough reason in itself for their study to assume a central role in academic research. Whereas, in the past, multinational enterprises were large companies with many resources that could be directed at minimizing information gaps, small and medium-sized enterprises, which nowadays also undertake international expansion, can not afford to expose their limited resources to such high levels of risk (Parks, 1996). Thus, a sound analysis of the risks that are inherent to undertaking activities at any particular location is of fundamental importance. Part of this study aims to analyse and establish whether Spanish MNEs take country risk (Avramovic, 1958; Saini & Bates, 1984) and political risk (Kobrin, 1979) into account, when taking decisions over their location and their presence in external markets. It also aims to determine whether the importance attributed to these variables is modified by the geographic region in which they intend to set up the subsidiary. Moreover, the selection of Spain, a country that over the past 30 years has become a member of the group of developed countries, will allow us to consider the possibility of applying generalizations to other recently developed countries (Galan et al., 2007). Issues surrounding the political risk analysis of Spanish MNEs calls for serious analysis. Such well-known political events as the revolutions in Nicaragua and Iran at the end of the 1970s caught many international banks and MNEs off guard. Other more recent examples include the present-day governments of Bolivia, Venezuela and even Argentina, whose policies have had an important impact on certain Spanish MNEs. - 3 - The large financial losses associated with the election of governments calling for greater economic independence from MNEs in those countries clearly demonstrated the importance of political events in host- country environments. The second case study describes and analyses the status of 158 countries based on ten risk categories which include security, political stability, and financial and tax policy, among others. Country and political risk are complex phenomena that encompass several aspects that can play key roles in investor decision-making processes. Using the same projection techniques as in the first case study, the aim of the second part is to analyse the situation of the countries included in the sample, taking into account the visualized data structure. These visually identified clusters, which comprise countries with a similar degree of institutional and economic development, provide an illustrative worldwide snapshot of risk levels; especially important in the context of a global crisis such as the present one. The cross-disciplinary field of computational finance relies on computer methods to make trading, hedging and investment decisions, and to manage the risk that is associated with such decisions. Machine learning (ML) is concerned with the design and development of algorithmic techniques that automatically extract rules and identify patterns from data, by applying computational and statistical methods. ML in particular, and Artificial Intelligence techniques in general, have been widely applied to solve different forecasting problems in Economics (Chen & Huang, 2007; Li & Sun, 2009). Business Intelligence plays a central role in producing up-to-date information for operative and strategic decision-making (Hannula & Pirttimäki, 2003). More recently, enterprises have started to apply Business Intelligence systems not only to support strategic decision-making but also a wide variety of other business activities (Elbashir et al., 2008). The identification of patterns that exist across dimensional boundaries in high dimensional datasets is a challenging task. Such patterns may become visible if changes are made to the spatial coordinates; however an a priori decision, as to which parameters will reveal most patterns, requires prior knowledge of the unknown patterns. Some techniques, such as intelligent data analysis (Chen & Yao, 2008) are aimed at revealing diverse non- trivial features or views of a large amount of data by studying the relations of the objects within a dataset. On the other hand, visualization techniques have been employed to analyse large datasets by providing a visual representation of the information contained in a dataset. These are considered viable approaches in the search for information, which they present on graphic display devices that highlight different characteristics and allow anomalies to be detected by the relevant decision-makers (Ahlberg & Shneiderman, 1994). Within this framework, Artificial Neural Networks (ANNs) are bio-inspired models that can be applied to different problems depending on the neural architecture to be used. They have been proven to successfully perform pattern recognition, information compression, dimensionality reduction, clustering, classification, data visualization, etc. As with other ML paradigms, the most interesting facet of ANN learning is its capacity to generalize after having learned/classified/identified the input patterns. In other words, the network can generalize its results onto a set of test patterns which have not been seen during learning. This study is based on the application of unsupervised learning within the framework of country, business and political risk. Thus, several ANNs are applied and compared in order to analyze the internal structures of two real- life cases studies. The empirical part of this research seeks to describe the principal characteristics of country and - 4 - political risk. To do so, two different datasets have been used. The first one (Spanish MNEs dataset) was generated to determine how country and political risk influence the location and the presence in foreign countries of Spanish MNEs, while controlling for the effects of enterprise and company-related variables. As a complementary case study, the Risk Briefing dataset (Risk Briefing, 2009) was chosen, providing a general overview of country and political risk all over the world. These two datasets are comprehensively described in this section. The study draws interesting conclusions and provides analysis on the use of these methods, which may help us to gain relevant knowledge about the institutional environment of the different possible locations of an investment, as well as valuable up-to-date information about the constantly-changing situation once the investment has been made. It is important to highlight that the impact of country and political risk is not only limited to the initial investment period but extends throughout the entire duration of the project. Cooperative Maximum-Likelihood Hebbian Learning (CMLHL) has been applied to the two case studies. The CMLHL projections were validated through a comparison with those generated by other dimensionality-reduction models such as Principal Component Analysis (PCA) (Hotelling, 1933; Oja, 1989; Pearson, 1901), SOM (Self- Organizing Map) (Kohonen, 1990), and CCA (Curvilinear Component Analysis) (Demartines & Herault, 1997)). Only the best results are presented (in terms of the projection), which were obtained after tuning the models. Several experiments were required to tune the SOM (SOM Toolbox for Matlab, 2009) to different options and parameters: grid size, batch/online training, initialization, number of iterations and distance criterion, among others. These parameters were tuned by following the strategies defined in previous SOM models (Kohonen, 1997; SOM Toolbox for Matlab, 2009). In the case of CCA, other parameters, such as initialization, epochs and distance criterion were tuned by following the criteria described in (Demartines & Herault, 1997). The study is structured as follows. Section 2 reviews previous work in the field of country and political risk analysis, while section 3 outlines the application of dimensionality reduction techniques for data analysis and describes the main neural projection models applied in this work. Section 4 describes the first of two case studies in this paper: the Spanish MNEs case study. This section also contains the experimental results obtained by applying the projection models, which are described in section 3 to this dataset. The second case study (Risk Briefing) is presented in section 5. Section 6 summarizes the conclusions and the future lines of research. 2 Previous Work Numerous works have sought to clarify the factors that are involved in decisions concerning the location of foreign investments, whether between developed countries, from developed countries to developing countries or, to a lesser extent, vice-versa from developing to developed countries. (Galan et al., 2007) contains an interesting table with a profuse bibliography of empirical studies that analyze the importance of certain investment-attraction factors. However, it may be said that the analysis of political risk as a fundamental factor in the location of direct investment has received far less attention than other factors, with the notable exceptions of (Marois, 1979, 1981) for French MNEs, (Rich & Mahmoud, 1990) for Canadian MNEs, (Mortanges & Allers, 1996) for Dutch MNEs, (Mutinelli & Piscitello, 1997) for Italian ones and (Noordin et al., 2006), the last-named work being one of the few to centre on multinational enterprises in a less-developed - 5 - countries; in this case Malaysia. Nevertheless, in certain cases some surveys report that as many as 100% of enterprises assessed the political risk to which their subsidiaries were exposed (Hashmi & Guvenli, 1992). In addition to other intra-firm operational strategies, such as raising the increase of within-firm sales (Feinberg & Gupta, 2009), the literature has traditionally associated a positive relation between Foreign Direct Investment (FDI) and political risk indices, on which higher scores indicate lower political risk levels. For instance, greater political restrictions imposed by the government add to the credibility of its commitments, which favours investments by foreign MNEs (Henisz & Zelner, 2001), whilst demands by stakeholder and pressure groups that run contrary to the interests of MNEs are given less attention (Henisz & Zelner, 2002). Also, greater assurances to ensure compliance with contracts, respect for property rights and greater economic freedom may attract more foreign investments (Bengoa & Sanchez-Robles, 2003; Kapuria-Foreman, 2007). Lastly, a lower corruption index score in the host country would have a positive relation with investment inflows, as perceived corruption levels would be lower (Cuervo-Cazurra, 2006, 2008; Wei, 1997). Nevertheless, results published in (De la Fuente et al., 2008; García-Canal & Guillén, 2008) show that, on some occasions, Spanish MNEs display a preferential bias towards countries with high corruption levels or towards those whose governments may exercise discretionality which would allow them to obtain competitive advantages over their competitors thanks to their negotiating skills gained from their experience of negotiating with governments in their country of origin, or the ease with which they might be able to benefit from corrupt systems. This has also been observed in American MNEs in the electrical sector (Holburn, 2001) that attempt to obtain competitive advantages by investing in countries where their political capabilities may be used to good effect (Henisz, 2003; Hillman & Hitt, 1999). This is also acknowledged by (Brouthers et al., 2008) who point out that some advantages may be gained in highly corrupt environments through corruption, as it may create resource-allocation efficiencies in countries with underdeveloped economic and legal regulations. These political capabilities are derived from the Resource-Based View of the firm (Wernerfelt, 1984) and relate to differences between the ability of firms to implement given strategies, depending on their tacit knowledge and skills. These differences tend to be developed in “factor-driven economies”, which is to say, in emerging economies where institutional development is more limited and transaction costs are still high due to fragile political stability and bureaucratic inefficiency (Wan, 2005). Thus, MNEs cultivate close relationships with local government to access the resources that such bodies control and allocate. Whereas the research work cited above tends to rely on variables elaborated by other scholars, MNEs assessing country and political risk traditionally use indices elaborated by private companies (Shell), consulting groups (Business International, Business Environment Risk Index, Political Risk Index or Frost and Sullivan Political Country Reports), specialized journals (The Economist, Institutional Investors or Euromoney) or rating agencies (Moody’s, Standard & Poors or SAP). In our study we offer two case studies, the first one uses three political risk variables frequently used in academic research (Index of Economic Freedom, Corruption Perceptions Index and Political Constraint Index) while the second case is based on data from one of the most important consulting groups (the Risk Briefing dataset from the Economist Intelligence Unit), thereby providing a comprehensive view of data from both academia and private companies. - 6 - Very few works have focused the identification of country and/or political risk from an Artificial Intelligence perspective. (Juliana & Heather, 2005) propose hierarchical clustering and ANN to predict country and sovereign risk rating within a macroeconomic context, approaching this task from a classification standpoint. On the contrary, and within a more microeconomic and “business-strategy” context, the present paper addresses the influence and role of political risk in the international expansion of Spanish MNEs from a visualization (through dimensionality reduction) standpoint. 3 Dimensionality Reduction Visualization for Data Analysis Projection methods project high-dimensional data points onto lower dimensions in order to identify "interesting" directions in terms of any specific index or projection. Such indices or projections are, for example, based on the identification of directions that account for the largest variance of a dataset (such as Principal Component Analysis (PCA) (Hotelling, 1933; Oja, 1989; Pearson, 1901) or the identification of higher-order statistics such as the skew or kurtosis index, which is the case of EPP (Friedman & Tukey, 1974). Having identified the interesting projections, the data is then projected onto a lower dimensional subspace plotted in two or three dimensions, which makes it possible to examine its structure with the naked eye. The remaining dimensions are discarded as they mainly relate to a very small percentage of the information or the dataset structure. In that way, the structure identified through a multivariable dataset may be visually analysed with greater ease. The combination of this type of technique together with the use of scatter plot matrixes constitutes a very useful visualization tool to investigate the intrinsic structure of multidimensional datasets, allowing experts to study the relations between different components, factors or projections, depending on the technique that is used. 3.1 Cooperative Maximum-Likelihood Hebbian Learning The standard statistical EPP method (Friedman & Tukey, 1974) provides a linear projection of a dataset, but it projects the data onto a set of basic vectors which best reveal the interesting structure in data; ‘interestingness’ is usually defined in terms of how far the distribution is from the Gaussian distribution. One neural implementation of EPP is Maximum-Likelihood Hebbian Learning (MLHL) (Corchado et al., 2004; Fyfe & Corchado, 2002), which identifies interestingness by maximising the probability of the residuals under specific probability density functions that are non-Gaussian. An extended version of this model is the Cooperative Maximum-Likelihood Hebbian Learning (CMLHL) (Corchado & Fyfe, 2003) model. CMLHL, which is based on MLHL (Corchado et al., 2004; Fyfe & Corchado, 2002) adds lateral connections (Corchado & Fyfe, 2003; Corchado et al., 2003) which have been derived from the Rectified Gaussian Distribution (Seung et al., 1998). The resultant net can find the independent factors of a data set but does so in a way that captures some type of global ordering in the data set. Considering an N-dimensional input vector ( x ), and an M-dimensional output vector ( y ), with ij W being the weight (linking input j to output i ), then CMLHL can be expressed (Corchado & Fyfe, 2003; Corchado et al., 2003) as: - 7 - 1. Feed-forward step: ixWy 1j jiji ∀= ∑ = N , . (1) 2. Lateral activation passing: ( ) ( )[ ]+−+=+ Aybτ(t)yty ii 1 . (2) 3. Feedback step: ∑ = ∀−= M i iijjj jyWxe 1 , . (3) 4. Weight change: ( ) p jjiij eesignyW ||..η=∆ . (4) Where: τ is the "strength" of the lateral connections between the output neurons, b the bias parameter, A a symmetric matrix used to modify the response to the data (Corchado & Fyfe, 2003), η is the learning rate and p a parameter related to the energy function (Corchado & Fyfe, 2003; Corchado et al., 2004; Fyfe & Corchado, 2002). The effect of the A matrix is based on the relation between the distances separating the output neurons. CMLHL has been previously applied as a dimensionality reduction technique in some other fields such as Computer Network Security (Corchado & Herrero, 2010; Herrero et al., 2009) or Knowledge Management (Herrero et al., 2010). 3.1.1 Fine Tuning The CMLHL fine-tuning process is based on the effect of changing the τ parameter. Experiments (Corchado & Fyfe, 2003) using the bars data set (Földiák, 1992), which add noise in a graduated manner across the outputs, have conclusively proven that the strength of the τ parameter directly affects the ability of the neural network to “gather” features together on the outputs. A low τ value allows the neural model to code both horizontal and vertical bars around a mode. If there is an increase in the τ value, the weak correlations between horizontal and vertical bars begin to have an impact on the learning. As the lateral connections become stronger, the bars are still learning one particular mode, but at the same time different orientations start to branch off. Finally, a separation between the two different orientations emerges, which is an interesting development as all the data presentations to the network consist of both horizontal and vertical bars. Further increases in the τ value force the network to learn only one orientation of bars. However, if the lateral connections are too strong, then the coding of the bars may be squashed into an area of the output space that is too small for all of the bars to be coded individually. The reason why one orientation of bars is suppressed is due to the pixel overlap between different orientations of bars. If the lateral excitation between the output neurons is strong enough, a single output neuron may be able to switch its preference from a horizontal to a vertical bar. (Corchado & Fyfe, 2003) consider that such switches in orientation identification were precursors to the creation of horizontal/vertical conceptualization in animals inhabiting a mixed environment. - 8 - 3.2 Self-Organizing Map The Self-Organizing Map (SOM) (Kohonen, 1990) was developed as a visualization tool for representing high dimensional data on a low dimensional display. It is also based on the use of unsupervised learning. However, it is a topology preserving mapping model rather than a projection architecture. The cerebral cortex of the human brain is one of the most complex biological systems. The different areas defined in this region are organized according to various sensory modalities: speech control, visual analysis, auditory control, etc. Each one of these areas consists of a large number of similar neurons that cooperate when carrying out their specific functions in which they have become specialized: auditory or hearing receptors, visualization, etc. Groups of neurons within each region respond jointly to excitations from the sensory cell they service (Patterson, 1998). There is a mapping of the features from sensory neurons to the associated spatial regions of the cortex. This biological feature mapping of the brain has been modelled reasonably well with ANNs. The computed SOMs are very similar to many brain maps as they also behave dynamically, in the same way, for example, as their magnification is adjusted in proportion to the occurrences of the stimuli (Kohonen, 1997). Thus the SOM (Baruque & Corchado, 2010) is a proper example of artificial topology preserving maps, where closer neurons are activated by similar inputs or stimuli. To mimic the biological brain maps, the SOM is composed of a discrete array of L nodes arranged on an N- dimensional lattice. These nodes are mapped into a D-dimensional data space while preserving their ordering. The dimensionality of the lattice (N) is normally smaller than that of the data, in order to perform the dimensionality reduction. The SOM can be viewed as a non-linear extension of PCA, where the global map manifold is a non- linear representation of the training data (Ritter et al., 1992). Typically, the array of nodes is one or two-dimensional, with all nodes connected to the N inputs by an N - dimensional weight vector. The self-organization process is commonly implemented as an iterative on-line algorithm, although a batch version also exists. An input vector is presented to the network and a winning node, whose weight vector c W is the closest (in terms of Euclidean distance) to the input, is chosen: ( ) i i Wc −= xminarg (5) The SOM is therefore a vector quantizer, and data vectors are quantised to the reference vector in the map that is closest to the input vector. The weights of the winning node and the nodes close to it are then updated to move closer to the input vector. There is also a learning rate parameter ( )η that usually decreases as the training process progresses. The weight update rule is defined as: [ ] )(, c icii NiWhW ∈∀−=∆ xη (6) When this algorithm is sufficiently iterated, the map self-organises to produce a topology-preserving mapping of the lattice of weight vectors to the input space based on the statistics of the training data. This neural model is applied here for comparative purposes as it is one of the most widely used unsupervised neural models for visualizing structure in high-dimensional data sets. - 9 - 3.3 Curvilinear Component Analysis Curvilinear Component Analysis (CCA) (Demartines & Herault, 1997) is a nonlinear dimensionality reduction method. Developed as an improvement on the SOM, it tries to circumvent the limitations inherent in previous linear models such as PCA. The principle of CCA is a self-organised neural network performing two tasks: a vector quantization of the submanifold in the data set (input space) and a nonlinear projection of these quantising vectors toward an output space, providing a revealing view of the way in which the submanifold unfolds. Quantization and nonlinear mapping are separately performed by two layers of connections: firstly, the input vectors are forced to become prototypes of the distribution using a vector quantization (VQ) method; then, the output layer builds a nonlinear mapping of the input vectors by considering Euclidean distances. In the vector quantization step, the input vectors ( i x ) are forced to become prototypes of the distribution by using competitive learning and the regularization method (Demartines, 1994) of vector quantization. Thus, this step, which is intended to reveal the submanifold of the distribution, regularly quantizes the space covered by the data, regardless of the density. Euclidean distances between these input vectors ( )( ) jiij xxdX ,= are considered, as the output layer has to build a nonlinear mapping of the input vectors. The corresponding distances in the output space are also used ( )( ) jiij yydY ,= . Perfect matching is not possible at all scales when the manifold is "unfolding", so a weighting function ( )( ) yij YF λ, is introduced, yielding the quadratic cost function: ( ) ( )∑∑ ≠ −= i ij yijijij YFYXE λ, 2 1 2 (7) Where: λ is a user-tuned parameter allowing an interactive selection of the scale at which the unfolding takes place. As regards its goal, the projection part of CCA is similar to other nonlinear mapping methods, in that it minimizes a cost function based on interpoint distances in both input and output spaces. Instead of moving one of the output vectors ( i y ) according to the sum of the influences of every other j y (as would be the case for a stochastic gradient descent), CCA proposes pinning down one of the output vectors ( i y ) "temporarily", and moving all the other j y around, disregarding any interactions between them. Accordingly, the proposed "learning" rule can be expressed as: ( ) ( )( ) ij ij ijijyijj Y yy YXYFty − −=∆ λα , ij ≠∀ (8) Where: ( )α is the step size that decreases over time. - 10 - 4 The Spanish MNEs Case Study The sample of enterprises on which the present study is based is made up of Spanish multinational enterprises of over 250 employees, which in December 2007 appeared on the list of the Spanish Institute of Foreign Commerce (ICEX - Instituto de Comercio Exterior), the web page of the Spanish Ministry of Industry, Tourism and Commerce (Network of Economic and Commercial Spanish Offices Abroad, 2009), and other foreign bodies concerned with foreign direct investment, which may be contacted through the ICEX, that compile directories of Spanish MNEs with investments in their countries. 4.1 The Dataset In total, the sample is formed of 166 Spanish MNEs, which have 1812 subsidiaries localized across the world, for which the data on the necessary variables was obtained for 1152 subsidiaries. Thus, the dataset is composed of 1152 observations relating to 12 variables, namely: Index of Economic Freedom, Corruption Perceptions Index, Political Constraints Index (POLCONV), total assets, employee numbers, Return on Equity (ROE), growth rate of sales, solvency ratio, number of foreign countries in which the MNE has its subsidiaries, Foreign Direct Investment/Gross Domestic Product (FDI/GDP), GDP growth, and total population. The first of these variables, was the mean average that the Index of Economic Freedom devised by the Heritage Foundation (Index of Economic Freedom, 2005) assigned to each country for the years 2004 and 2005. This index is made up of various variables that measure the independence of the judicial system, the ability of firms and individuals to ensure full compliance with contracts, corruption within the judicial system, the degree to which the government protects the property rights and the degree of freedom that exists for businesses, commerce and investment. The mean average of the 2004 and 2005 Corruption Perceptions Index devised by Transparency International (Corruption Perceptions Index, 2008) has been already included in the dataset. This index is used to measure corruption as perceived by business leaders and experts from each country. As complementary data, the dataset also includes the mean average of the last five years of Henisz’s Political Constraint Index (POLCONV) (Henisz, 1998). In this index, the number of independent authorities with a power of veto is taken into account, the score being modified in accordance with the possible alignments between authorities, such that they affect the actual constraints to which the government is subjected. Additional modifications are also made when some political authorities are neither totally aligned nor totally opposed, to take account of their composition when determining the degree of political constraint. By using several measures, we take into account different aspects of political risk, giving a better picture of the host country’s overall governance infrastructure (Slangen & van Tulder, 2009). Additionally, the dataset includes several firm characteristics such as: total assets, employee number, Return on Equity (ROE), growth rate of sales, solvency ratio and the number of foreign countries in which the MNE has its subsidiaries. Some country macroeconomic characteristics are also included: Foreign Direct Investment/Gross Domestic Product (FDI/GDP) as a measure of the openness of the destination economy, GDP growth as a measure of the attractiveness of the country and total population as a measure of the size. - 11 - The previously introduced dimensionality reduction techniques (see section 3) were applied to the Spanish MNEs dataset. The projections obtained by these neural projection models are set out and analyzed in the following sections. - 12 - 4.2 PCA Projection PCA was applied to the Spanish MNEs dataset. Fig. 1 presents the PCA projection onto the two first principal components. Fig.1 << File Fig.01.tif goes here>> 4.3 CMLHL Projection The following CMLHL projections (Figs. 2-4) reflect the different motivations driving Spanish MNEs to localize in countries within the different regions under analysis, which represent, to a great degree, the main host countries traditionally targeted by Spanish foreign direct investment. These projections were obtained by applying the following values to the different CMLHL parameters: number of iterations = 4.000, learning rate = 0.0208, p parameter = 1.2, and τ parameter = 0.11. Fig. 2 depicts all the possible combinations of the third first factor pairs obtained through CMLHL by means of a scatter plot matrix. Factor pairs under the diagonal provide no extra information. Fig.2 << File Fig.02.tif goes here>> The main results obtained by CMLHL (factor pairs 1-2 and 1-3 from Fig. 2) are analyzed in depth in this section. - 13 - 4.3.1 CMLHL Factor pair 1-2 Fig.3 << File Fig.03.tif goes here>> The groups identified in the CMLHL factor pair 1-2 analysis (Fig. 3) are described as follows. Group 1 (64 records) On the whole, this group is characterized by subsidiaries with a presence in European countries such as France, Ireland, Netherlands and Hungary. Subsidiaries may also be found in Latin-American countries - Brazil, Honduras, Colombia and Venezuela - though it is an almost token presence. In addition, the average size of these enterprises is smaller than the rest of the sample, and they have a varied sectoral composition. Thus, given the limited amount of available resources to launch their internationalization strategies, the enterprises in this group have sought to capitalize on the competitive advantages of the sectors to which they belong, acquired from their experience in the domestic market. Accordingly, they select specific markets and attempt to assure a solid competitive position. Group 2.1 (35 records) The most relevant European countries in this group are Germany and Austria, with a slight presence in Portugal of scant importance. Latin America is almost entirely represented by Argentina and Venezuela. In this case, the composition is less varied than in the preceding case, as these are subsidiary enterprises in the services sector, and here too the sizes of their parent enterprises are below average. It is noteworthy that the lowest scores on the Index of Economic Freedom for the sample are attributed to these countries, as once again they reveal a group that bases its internationalization strategy on competitive advantages, with little regard for the level of security offered by a high score on the Index of Economic Freedom, consistent with the results published in (De la Fuente et al., 2008; García-Canal & Guillén, 2008). Group 2.2 (63 records) Once again, Venezuela is the country to which the major part of direct Spanish investment is directed in this region, and other locations appear alongside it in the Dominican Republic, Peru and Uruguay. For its part, Europe is fundamentally represented by the UK, Czech Republic, Sweden and only slightly so by Portugal. This group has similar characteristics to group 2.1, as it is principally made up of smaller-sized enterprises in the services sector. The main difference lies in the location from which the enterprises can exploit their competitive advantages, which in this case tends to be in countries whose legal systems are based more on common law, as opposed to the preceding group, whose codified legal systems are influenced more by Roman law (civil law). Group 3.1 (19 records) This is a group with few members, in which only France, Greece and Guatemala are present. Enterprises of all sizes may be found, although predominantly from the manufacturing sector. They are not especially concerned by the degree to which ownership rights are protected or whether compliance with contracts are guaranteed; a fact that - 14 - is reflected by somewhat lower scores on the Index of Economic Freedom, showing a proactive use of political risk in the host countries. Group 3.2 (97 records) In this case, the most widely represented European countries are Germany, Belgium, Slovakia, Slovenia and Finland. Relatively well distributed, they have a similar presence in all Latin America locations spread across Argentina, Brazil, Chile, Colombia and El Salvador. In this case, the enterprises are of a larger-than-average size, and manufacturing enterprises predominate, as they do in the latter subgroup, although financial and banking enterprises are also found. These are enterprises which, given their greater volume of resources, launch their internationalization strategies in Germany and in the largest Latin-American countries, although they will also set up wherever they find an attractive growing market. It is the case at present of some countries that have recently joined the European Union. Group 3.3 (98 records) This subgroup, as is largely the case throughout Group 3, is made up of large enterprises, able to launch their internationalization strategy in the major economies of each region, but also wherever they come across a business opportunity. In this case, however, the locations are centred principally on Chile, although Argentina, Bolivia and Ecuador also stand out; and Belgium, within Europe, along with Germany and Austria, although it is one of the subgroups in which the European presence is the least significant. Group 4.1 (121 records) The presence in this subgroup of Poland and Estonia is very relevant, as they are the largest economies among all of the countries that joined the European Union in its expansion eastwards. With respect to the former, the great majority of Spanish subsidiary locations are found in this group, but with respect to the latter, all (although it should be said that there are not very many) the locations are found within this subgroup. Mexico, Panama and Costa Rica may be highlighted in Latin America, but behind the main destination that is Peru. The size of the enterprises is in line with sample average, with a slight bias in favour of the larger enterprises. As regards their sectors, service enterprises are the most abundant in absolute terms, but mining and construction enterprises are the protagonists if the data are analysed in relative terms. Group 4.2 (117 records) Within Europe, France, Italy, and Luxembourg all stand out, as well as, to a lesser extent, the Netherlands, Ireland and Denmark, whereas in Latin America, the presence of subsidiaries is evenly distributed between Argentina, Brazil, Chile, and Colombia. Costa Rica and Panama may be highlighted as countries which, despite being less attractive traditional destinations for Spanish investment because of the smaller size of their economies, have a similar number of subsidiaries as the other countries in this group. Average size is again slightly biased towards the larger enterprises, although the importance of this slight bias is minimized, if we bear in mind that - 15 - these subgroups have the highest number of enterprises. Once again, the role played by enterprises in the mining and construction sectors is relevant, although in this case they are accompanied by enterprises from the financial sector. It is worth recalling the significant investments made by the large Spanish banks in the principal countries of Latin America and in Central Europe, which are precisely the locations that are covered in this section. Group 4.3 (35 records) This Group is the smallest within Group 4, in which the European presence in Belgium, France and Greece predominates, whereas on the Latin American side there are locations in Argentina, Costa Rica, El Salvador and Venezuela, although of slight importance. There is an even distribution between large and small multinational enterprises, but it is very biased towards the services sector, which is a differentiating feature of this subgroup with respect to the preceding ones. Group 5.1 (68 records) Unquestionably, the determining feature of this subgroup is the presence of subsidiaries in Portugal - 52 of the 68 records -, together with a slight presence in Poland. The only Latin American locations are found in a further group that is characterized by its presence in Europe, Mexico and Peru. Undoubtedly, this subgroup is made up of enterprises that are searching for a location in countries with close cultural similarities, which is why they choose southern Europe and Latin America, while remaining attentive to the political constraints faced by the government, in an attempt to minimize the negative impact that unilateral decisions taken by the host government could have on the interests of the subsidiary enterprise, in line with the results obtained by (Henisz & Zelner, 2001, 2002). The fact that average enterprise sizes are smaller reinforces this hypothesis, as enterprises with fewer resources have to be more cautious when selecting their locations and to try to avoid mandatory contractual modifications or even nationalizations or expropriations in the worst case scenario; regardless of the individual economic sectors in which they are based, their presence is evenly distributed across them all. Group 5.2 (62 records) The presence in Europe is quite evenly distributed, with Belgium, France and Holland at the top of the list. However, the Latin American side is fundamentally composed of the large-scale economies of Argentina and, above all, Brazil. It is a widely distributed group with regard to its composition both by size and by sector, and is characterized by the "conservative" nature of its locations; a concept understood in terms of investments in neighbouring countries in the case of Europe, and in the larger economies with sizeable markets and a potentially broader consumer base, where they take an initiative to invest in a more distant location such as Latin America. In all cases, the most secure countries are chosen in view of corruption levels, economic freedoms, political restrictions and the cultural distance of these countries with respect to the other investment alternatives in each region, showing additional evidence of the relation found by (Bengoa & Sanchez-Robles, 2003; Cuervo-Cazurra, 2006, 2008; Kapuria-Foreman, 2007; Wei, 1997). - 16 - Group 5.3 (103 records) This is the most numerous subgroup within Group 5, in which both France and Italy stand out on the European side, and mainly Mexico and Peru, but also Brazil, Chile, Colombia, and El Salvador, in a subgroup in which the presence, to a greater or lesser extent, of practically all the Latin American countries may be underlined. Once again the composition of the sample is well distributed by size and by economic sector. The internationalization strategy followed in this subgroup is similar to the preceding group, in that the enterprises select countries within Europe that are very near to Spain, although it is more diversified with regard to the location choices in Latin America. Group 5.4 (59 records) Once again, this is a subgroup where the European presence outweighs the Latin American presence; 47 of the 59 records belong to subsidiaries localized in the ‘old continent’. The two countries with the greatest presence in this group are principally Germany, followed by Portugal. On the Latin American side, the presence in the majority of countries may be described as almost solely symbolic, except for the presence in Venezuela. Once again, this is a conservative strategy, as the two destinations appear that were missing among those countries that were fundamentally targeted by direct foreign investment from Spain. As in Group 5, enterprises from all sectors may be found, but they are characterized by a very small below-average size, which explains this tendency to prioritize investment security in predictable well-known institutional settings that are close to home. Group 6.1 (24 records) This is a small subgroup in which the UK, Czech Republic, Italy and Lithuania stand out on the European side, whereas on the Latin American side, practically all of the locations are in the Dominican Republic. Their sizes are somewhat above average. Following an analysis of the sectoral composition, it is worth stressing the prominence of fashion wholesalers and retailers; principally the multinationals with an interest in fashion. The strategy of this subgroup is not entirely defined. This is easily identifiable in the graph of the factor pair 1-2 projection, where it can be appreciated how its constituent elements are relatively spaced out in comparison with the distance between the elements of the other subgroups. Group 6.2 (57 records) This group is almost exclusively made up of locations in Latin America - 54 of a total of 57 records – which is why Mexico, Nicaragua, Panama and Paraguay, though no single European country may be highlighted. The enterprises in this subgroup are of a considerable size and are active in the mining, construction and services sectors. The principal Spanish multinationals (Telefónica, Repsol, ACS, Inditex, BBVA, BSCH) all have at least one subsidiary in this group that is characterized by very low corruption and political constraint indices. It is therefore evident that these locations respond to an internalization strategy that is guided by exploiting skilled negotiations with local authorities or indeed by benefiting from facilities offered to Spanish MNEs by the more corrupt systems that hold sway in these countries to consolidate positions with competitive advantages, at least in - 17 - the short term, in line with the results in (De la Fuente et al., 2008; García-Canal & Guillén, 2008; Holburn, 2001). However, in the long term, this can lead to problems relating to obligatory contractual modifications or expropriations (De la Fuente et al., 2008; García-Canal & Guillén, 2008). Group 6.3 (19 records) This is another sparsely populated group, in which Hungary, Greece and Guatemala figure prominently, with enterprises of an intermediary size that are in the wholesale and retail sectors, as in group 6.1. They have no clear internationalization strategy, which is evident from the high degree of dispersion between their constituent elements; they differ exclusively in those countries selected as investment destinations. Group 7.1 (3 records) This is a mini-group formed by 3 subsidiaries in small Eastern European countries such as Slovakia and Slovenia. All 3 belong to large Spanish MNEs (Iberdrola, Inditex and Mango) but they can almost be considered as special cases in the analysis. Group 7.2 (28 records) The principal destination of this subgroup is the UK, with 19 of the 28 records, while Portugal has much fewer. Only 3 records are found in Latin America and no one country may be highlighted. The average size is evenly distributed, but a high presence may be appreciated in the manufacturing sector, which seeks countries in this subgroup with high scores for the indices of corruption, political constraints and economic freedom. It differs from the preceding groups where this sector was prominent, but where the internationalization strategy paid less attention to political risk associated with investments. Group 7.3 (24 records) In this group, the subsidiaries are localized in Latvia, Portugal, Czech Republic and Sweden, along with the Dominican Republic solely on the Latin American side. They are large enterprises, with the presence of various well-known multinationals (Telefónica, Inditex, BBVA, BSCH), which take advantage of their greater volume of resources as was the case of groups 3.2 and 3.3, (it may be seen that they are found at a similar height in factor pair 1-2 projection – Fig. 3), to expand not only in the reference markets of each zone, but in other less accessible markets that nevertheless offer good business opportunities. Group 7.4 (22 records) This group is characterized by the massive presence of subsidiaries in the UK, which contrasts with the absence of locations in Latin America. This phenomenon is easily explainable if the large size of the subsidiaries and above all their sector is taken into account. Thus, there is predominance of financial entities in London, the financial heartland of the ‘Old Continent’, where they have set up operational bases for their international clients. - 18 - Group 7.5 (35 records) Finally, in this last group, locations of large multinationals across all sectors are found in Sweden, the UK and Uruguay, which seek countries where all the indices relating to economic freedoms and to protection of ownership rights are very high. General Conclusions The projection model employed in this study (CMLHL) has enabled us to distinguish different motivations that drive enterprises to localize in the regions under study, the vast majority of which represent the main destinations traditionally targeted by Spanish foreign investment. Thus, it has been observed how Group 1 is characterized by smaller enterprises that could not allow their resources to be misspent on risky internationalization strategies. Something similar took place in Group 2, in which small service sector enterprises are concentrated, and in which it is possible to distinguish two sub-groups according to the countries which manage them. Group 3 however, contains enterprises with very different characteristics, in which the manufacturing sector may be highlighted and large enterprises with complex internationalization strategies on account of the level of resources they can invest. For its part, Group 4, which contained the highest number of enterprises in the entire sample, showed the groupings of enterprises working in construction that is currently quite a controversial sector, which have targeted both developed economies with large markets and recent members of the European Union that have better forecasts for growth and profitability. Nonetheless, the existence of subgroups where the presence of Eastern European countries is so predominant, as in Group 4.1, demonstrates that these countries, despite having made significant economic and social progress, still constitute an investment destination, with specific characteristics which differ from the other European partners in the European project; findings that are in agreement with those of (Durán et al., 2008). Group 5, with all of its subgroups, has shown how some MNEs seek to minimize the risk associated with foreign investment by investing in countries that are culturally akin to their own, searching for better management and solutions to the problems that may arise. This group is the best example of the negative relation between political risk and FDI demonstrated by (Bengoa & Sanchez-Robles, 2003; Cuervo-Cazurra, 2006, 2008; Kapuria-Foreman, 2007; Wei, 1997). Group 6 is particularly interesting as it is made up of the flagships of Spanish foreign investment. Moreover, they demonstrate a degree of short-termism in their investment strategy in Latin America, seeking to gain an attractive competitive position at the start of the investment period, but unconcerned about the problems that are inherent to political risk. This proactive use of political risk is consistent with the results obtained by (De la Fuente et al., 2008; García-Canal & Guillén, 2008; Holburn, 2001). Finally, Group 7 constitutes a set of locations where most subgroups have been identified, and in which the high dispersion of their constituent elements may easily be appreciated. It is here that those enterprises with less obvious - 19 - internationalization strategies are found, on occasions similar to those of other groups, but in different countries or with enterprises from varied sectors. 4.2.2 CMLHL Factor Pair 1-3 Fig.4 << File Fig.04.tif goes here>> As a complementary projection to the factor pair 1-2 analysis (Fig. 3), Fig. 4 shows the CMLHL factor pair 1-3. The different groups identified in this figure are analysed in the following paragraphs. Group 1 (64 records) This is a group made up of large enterprises, which are diversified but with a slightly higher presence of the manufacturing sector, with subsidiaries in France, the Netherlands, and Hungary, and a scant Latin American presence, which is similar to group 1 in the factor pair 1-2 analysis. Groups 2, 3 and 4 (12, 421 and 122 records) These three groups are found very close to each other as may be seen in Fig. 4 showing the distribution of the factor pair 1-3 analysis, on which basis some shared approaches to the internationalization strategy conducted by these multinationals may be defined. Also, the large enterprises predominate in this case, with a presence across all sectors, and in a variety of both European and Latin American countries, due to the high number of individuals that make up this group. If subgroups were defined, as was done with the factor pair 1-2 analysis, closer sets would be observed in different zones in the upper, medium and lower part of the graph, but generally speaking these groups coincide with groups 2, 3 and 4 (composed by their respective subgroups) identified by the factor pair 1-2 analysis. Group 5 (169 + 86 + 57 records) Once again, a group with numerous enterprises of all sizes, from all possible sectors and country locations: as in the previous case, the subgroups in group 5 may be likened in a general way to those in the factor pair 1-2 analysis with regard to sectoral distribution and country location, although with some small differences with regard to size, as the bias towards small enterprises in the preceding analysis is not as clearly defined as in this case. Group 6 (109 records) The UK, the Czech Republic, and Italy are prominent in Europe, and likewise, Mexico, Panama and Nicaragua in Latin America. Once again, group 6 in this analysis jointly covers subgroups 6.1 and 6.2 in the preceding analysis, although in this case it does not include subgroup 6.3, the relevance of which, it may be recalled, was minor. Group 7 (23 + 45 records) This is a much less numerous group, in which the Dominican Republic, the UK and Portugal play the most important roles, with some contributions from Sweden, Latvia and Poland, which repeats the pattern analysed in - 20 - the 1-2 component, but without including strategies that are to a certain extent more "marginal" such as those of subgroups 7.1 and 7.5. Group 8 (34 records) Subgroup 7.5 which was not included beforehand does appear in this additional group that is contributed by the factor pair 1-3 analysis. Uruguay, Sweden and the UK all have a notable presence. General Conclusions In general terms, the factor pair 1-3 analysis adds weight to the conclusions drawn from the factor pair 1-2 analysis, as its groups contain more enterprises that cover the internationalization strategies detected earlier, although in summary form as it identifies far fewer subgroups. The analysis is therefore more general, which prevents the detection of slightly more heterogeneous groups, and less well-defined strategies, as was the case of subgroups 6.3 or 7.1, or certain subtle size-related biases. This analysis divided the former group 7 into two groups - 7 and 8 – but it assimilated the former subgroup 7.5 which appears at the extreme edge of the graph of the factor pair 1-2 analysis, which meant that this second analysis considered it to be a different group; although always with an importance well below groups 3, 4, 5 and 6. These constitute the principal groups in which Spanish foreign investment is concentrated, and where the internationalization strategies are found: on some occasions more traditional and conservative, on others more risk oriented or even, as shown by the bias towards countries with high levels of corruption, clearly motivated by short-termism (De la Fuente et al., 2008). 4.4 SOM Mapping The SOM mapping of the dataset is visualized by means of: � U-matrix (Fig. 5): distance matrix between the reference vectors of adjacent neurons of a two- dimensional lattice. � Numbered map (Fig. 6): neurons in the rectangular lattice have been labelled with a number. These numbers represent the amount of companies associated with each neuron. To ease the analysis of this result, neurons have been grouped according to their location in the map. Additionally, to make this mapping more intuitive, the neurons have been coloured following a topological ordering. These results are analyzed in the following paragraphs. The parameter values that generated this mapping were: random initialization, batch training, hexagonal lattice, and cut Gaussian neighbourhood function. The grid size was determined by means of a heuristic formula. - 21 - Fig.5 << File Fig.05.tif goes here>> Fig.6 << File Fig.06.tif goes here>> Group 1 (47 records) This group is made up of larger-than-average size firms from the wholesale and retail commercial sectors as well as from the financial services sector. Firms are concentrated in Belgium, followed by other countries such as Austria, UK and the Netherlands. Their strategies all reflect low levels of political risk and their locations are widely diversified in Latin America, where countries with low and high levels of political risk coexist alongside each other. Group 2 (233 records) This is the most numerous group in the SOM mapping, with firms of all sizes, although there is a slight predominance of larger-sized firms and the sample is very biased towards service sector firms. Their presence in Europe is widespread in countries that have traditionally been targeted by Spanish investment: France and Portugal, though a presence in Belgium, and Poland and Slovakia may once again be highlighted. Brazil, Chile and Venezuela are the main destinations for investment in Latin America. It is therefore a matter of companies that invest in markets which are geographically and culturally close to Europe and that prioritise other competitive advantages over and above the reduction of the risks that they encounter, as these remain within reasonable limits. Within Latin America, the firms prefer to opt for larger-sized markets with large numbers of potential consumers, higher income levels and lower levels of political risk. This is consistent with the majority of firms being found in the service sector, as their penetration into this region is driven by the search for new markets and not by a desire for cost reductions or greater efficiency. Group 3 (5 records) This is a group composed of only five firms, which may therefore be considered a special case that does not deserve any attention. Group 4 (88 records) Group 4 is composed in almost equal parts of firms of all sizes, in which service and financial firms predominate, but in which the participation, in relative terms, of the agricultural and food sectors may be highlighted. In Latin America, those countries with a degree of purchasing power such as Chile, Mexico, Peru and Venezuela stand out, where most of the firms in this group are active, whereas the scant presence in Europe is concentrated in Italy, which is somewhat logical if similarities between the foodstuffs consumed in both countries are taken into account. - 22 - Group 5 (57 records) This group is characterized by firms of a large size that, as in the previous group, are principally from the financial, agricultural and food sectors. However, their locations are radically different. The UK stands out in Europe, as the European financial heartland, whereas the target countries are smaller and less economically developed in Latin America; principally Costa Rica, Ecuador and Nicaragua. Group 6 (3 records) This is the least numerous group in the entire mapping, from which no conclusions may be drawn with respect to the characteristics of its constituent firms, which are considered special cases, in the same way as those in group 3. Group 7 (88 records) The principal feature of this group is the absence of large-sized firms, as the group is solely composed of firms in the first three quintiles of a population ordered from small to large, and fundamentally of firms in the first two: the two main sectors being principally services and manufacturing. The Latin American presence is composed almost exclusively of firms localized in Chile, whereas Portugal, the UK, the Netherlands and Belgium are prominent in Europe, and the presence of Nordic countries - Denmark and Finland - are also relevant. It therefore appears that the strategy of these firms consists in seeking out developed countries in which to carry out their businesses activities under the most stable possible conditions. Group 8 (214 records) This is the second most numerous group in the SOM mapping, with a below-average size, the principal references for which are the manufacturing and services sector. The presence in the American continent is very well distributed, both in large and in small countries, and with relatively low and high levels of political risk. It is also distributed in Europe, although there is a slight bias here towards countries with highly developed market economies, principally Germany, France, and the UK that are references in Europe and worldwide, and also Sweden. Group 9 (105 records) The bias towards large-sized firms in this group is one of the most pronounced. However, its sectoral distribution is much more diversified, with wholesale and retail firms from the financial sector, as well as from the transport, communications and public services sectors. Brazil, Colombia, Paraguay and Uruguay stand out in Latin America, whereas Greece and Eastern European countries such as Poland and Slovakia predominate in Europe. Thus, it is a matter of investments with a higher risk, in countries with higher levels of political risk, though these factors might be compensated by lower levels of competitiveness and by greater potential for growth. - 23 - Group 10 (36 records) This group is one of the most numerous in the mapping and is composed of large-sized firms. They belong to the agricultural, food and manufacturing sectors, with an almost total and notable absence of the financial and services sectors, which in principle are the most numerous. Their presence is concentrated on the one hand, in Bolivia, Brazil and Mexico and, on the other, mainly in Germany. Thus, the strategy of this small group of firms appears to centre on larger-sized markets. Group 11 (61 records) Large multinationals from the manufacturing and financial services sectors make up the greater part of this group. Italy and Mexico are the main destinations for foreign investments, clearly revealing an internationalization strategy based on minimization of the negative impact that a large cultural gap between the country of origin and the host country can have for the success of the investment. Group 12 (124 records) This group is composed of larger-than-average sized firms, but with a range of sizes, as there are various small- sized firms (first and second quintile), very few of a slightly larger-than-average size (fourth quintile) and many of a very much larger-than-average size (fifth quintile). As in group 10, the agricultural, food and manufacturing sectors predominate in the sample, although in this case such an acute absence of the other sectors is not evident, given the high number of records that make up the group. Ecuador, the Dominican Republic and Venezuela are prominent in Latin America, whereas France, Poland, Portugal and the UK are prominent in Europe. Two different investor strategies may be appreciated in this group: those that seek low salary costs for manufacturing purposes in Latin-American countries and Poland, and those that seek a market share in developed countries where they can successfully launch their products in European countries that are not so far away from Spain. Group 13 (69 records) A pronounced bias towards smaller-than-average size service sector firms may be seen in group 13. The Spanish MNEs in this group are found in small Latin American countries with higher levels of risk, namely El Salvador, Honduras, Peru and new members of the European Union from "Eastern Europe" such as Poland, Hungary and the Czech Republic. This exceedingly risky internationalization strategy is due to their smaller size thanks to which they can adapt with greater ease to the business setting and take advantage of lower sinking costs that may never be recouped in case of divestment. Group 14 (23 records) Finally, the last group is the smallest of the mapping (excluding the two special cases) and its composition is evenly distributed between small and large-sized firms. With the exception of one firm from the manufacturing sector and two from the telecommunications and public services sectors, the group is solely composed of financial and services-sector firms. A European presence is almost inexistent, as only two subsidiaries are found in Italy, - 24 - whereas the presence in Latin America is concentrated in Colombia, well in front of El Salvador, Ecuador and Mexico. 4.5 CCA Projection Fig. 7 presents the CCA projection. CCA was unable to display the inner structure of this dataset, and could only differentiate 2 main groups. The parameter values that generated this projection were: cityblock distance, lambda = 168,850 (default value), alpha = 0.5 and 5 epochs. Fig.7 << File Fig.07.tif goes here>> As can be seen in Fig. 7, CCA provides a main cloud containing most of the data, and simultaneously a smaller one above the main cloud. A full description of this projection is not provided as it is not given too much information about the internal structure of the dataset. 4.6 Discussion and Conclusions for the Spanish MNEs Case Study This case study has brought to light the different reasons for which Spanish MNEs internationalize and the way they pursue different goals, which may be appreciated in the above projections. These MNEs set up in a particular country according to their specific needs or those of their sector, as it is evident from the different groups identified by CMLHL. The projection of groups generated by the SOM has shown different internationalization strategies that are put into practice by Spanish MNEs, principally in accordance with their size, the sector to which they belong and the motivation that led them to choose a particular country in which to target their investment. However, it may be appreciated from a comparison between the results generated by this approach and those generated by CMLHL that the latter technique displays a more easily interpretable projection of the different groups, allowing more studied conclusions to be drawn based on a more homogeneous internal group structure, with clearly differentiated external structures. Thus, the SOM has on some occasions enabled risk-aversion strategies to be perceived and, on other occasions, versatile strategies that adapt to unfavourable business settings. However, some groups include subsidiaries localized in various countries with different strategies. In contrast, the internationalization strategies brought to light by CMLHL may be more clearly appreciated, even those that are more complex. Examples include certain strategies that lead companies to set up in countries with higher levels of political risk, where they can use their negotiating skills to gain competitive advantages from local government bodies with extensive discretional powers or in countries with corrupt systems, which is consistent with recent findings (De la Fuente et al., 2008; García- Canal & Guillén, 2008; Holburn, 2001) that had previously identified a proactive use of political risk in the internationalization strategy of firms under certain circumstances. Other company strategies target Eastern European countries because of their potential for growth and stability, thanks to their institutional development and recent entry into the European Union (Bevan & Estrin, 2004; Wheeler & Mody, 1992), and yet others are mainly guided by cost reductions and greater efficiency, amongst other considerations. Furthermore, the subgroups that - 25 - this technique has revealed, with similar strategies but with certain differentiating characteristics, enrich the conclusions that may be drawn and help to understand the complex internationalization strategies of Spanish firms. CMLHL (Figs. 2-4) generates a clearer and more widely spread mapping than PCA (Fig. 1), SOM (Figs. 5-6), and CCA (Fig. 7). It extracts more visual information, which in turn enables clearer and better conclusions to be drawn from the data. In brief, it may be concluded that this study has revealed a range of different reasons underlying the internationalization strategies of Spanish MNEs and the different goals they pursue. This may be appreciated from the different groups identified by CMLHL, which show enterprises localizing in a specific country according to their specific needs or those of their sector, as is evinced by the different subgroups. 5 The Risk Briefing Case Study The Risk Briefing dataset (Risk Briefing, 2009), which is maintained by the Economist Intelligence Unit, tracks the scores of 158 countries for ten categories of risk: security, political stability, government effectiveness, legal and regulatory, macroeconomic, foreign trade and payments, financial, tax policy, labour market, and infrastructure. These categories, updated once a quarter, provide an assessment of different aspects of country and political risk by covering, among others, the financial, macroeconomic, fiscal, labour and political environment. They range from 0 to 100, where 100 refers to the riskiest country in the selected category. As asserted by the authors, this dataset is designed to assist investors assess and plan in view of the risks of doing business around the world, providing comprehensive and timely analysis, forecasts, alerts, background studies, and data covering a wide range of risk factors. The dataset is generated by a global network of over 650 correspondents and analysts who monitor world events and assess their impact on business. As some of the risk categories are highly correlated, it was decided to remove six of them with a high correlation index. As a result, only four categories (security, macroeconomic, tax policy, and labour market) were included in the dataset for such case study. 5.1 PCA Projection Fig.8 << File Fig.08.tif goes here>> Fig. 8 presents the PCA projection onto the two first principal components. As may be seen, PCA is able to identify 11 different groups as described below. Group 1: 30 countries Low levels of below the average tax policy and labour market risk, and especially low levels of security risk are key features of this group. In this latter case, moreover, the scores are the lowest of the whole sample. These are characteristics of developed countries which are political and institutionally stable such as United States, France, Germany, Canada, Norway or Holland. However, due to the present global crisis and the multiple interlinkages of - 26 - the international economy, these countries suffer from very high levels of macroeconomic risk, definitely over the average. However, this group also includes a few new members of the European Union, such as Slovakia, Slovenia, Czech Republic, Latvia or Lithuania which are geographically close to other economies of this group and whose accession to this supranational body has lead to improvements in their economic and institutional development. However, it is still too early to affirm that the aforementioned countries share the very same features towards their risk profile and can be included within the same group with the other economies. Group 2: 15 countries Countries in this group, where Turkey, Argentina or Russia can be found, show slightly over the average security, tax policy and labour market risk levels. The institutional development and the political stability of the countries are lower than they were in the previous group. However, macroeconomic risk levels are lower because these economies are less integrated in the international economic context. Group 3: 43 countries Group 3 is formed by developing countries, some of them relatively advanced such as Mexico, Brazil, Vietnam, Malaysia, Egypt, Romania or Croatia, and others not so much like Armenia, Moldova, Namibia, Iran or Albania. Here, institutional development is deficient, which provokes high security risk levels. They are also politically unstable, with frequent changes in national policies that affect taxation and labour market and increase risk levels. However, their greater development compared to other developing economies and international linkages with other economies increase their relative exposure to the present global crisis rather more than might be expected at first, which explains their higher macroeconomic risk scores. It is worth highlighting that this group includes Italy and South Korea, countries that appear to fit less with other elements of the group, and more so with the other group of more developed countries. Group 4: 16 countries These countries have below average risk levels for all the different aspects of risk analysed in this paper. Scores are especially low in the security and tax policy risk categories, and closer to the average in the other two. This fact shows that countries are also developed but to a lesser degree than those of group 1 (i.e. Spain, Portugal, Chile, Australia or even to a lesser extent those more developed countries from the Middle East such as Bahrain, United Arab Emirates or Qatar). They represent a group of countries that are attractive for international investors seeking political stability in order to minimize the risk they face when exploiting the full potential of business opportunities in developed countries. Group 5: 32 countries The sixth group is made up by developing countries with few links to the international economy. This allows them to obtain lower macroeconomic risk levels in the context of a global crisis. One of the scarcer linkages comes from the oil market, since some of these countries are producers and exporters, which provides them with a - 27 - constant flow of income to finance their imports. Their levels of tax policy and labour market risk are above average. One of the reasons is that their governments wield highly discretional powers and face low restrictions (i.e. China, Libya, Morocco, Jordan, Kuwait, Bhutan or Syria). However, in order to maintain themselves in power and avoid international pressures, security risk levels are kept at an intermediate level, in an attempt to offer an image of a safe location for possible international investors. Group 6: 21 countries Group 6 is characterized by countries with a very low level of development such as Honduras, Kenya, Nigeria, Sudan, Bolivia, Colombia or Pakistan. Their governments are fragile and unstable, and offer few guarantees their commitments with regard to policies on taxation or the labour market will be upheld, increasing the risk associated with those two aspects. In addition, their economies depend heavily on other developed countries, for instance because they regularly buy their export products. Ever since the present global crisis has reduced international trade, exports have been dramatically reduced, which has increased macroeconomic risk. Group 7: 1 country (Iraq) This group can be considered a special case. As a country experiencing strong political instability, its situation differs radically from any other country in the sample. General Conclusions In conclusion, PCA has been able to differentiate, with an acceptable degree of precision, between the 158 countries in the sample, one group formed by the principal economies in the world and another one that also includes developed countries but which lags a little bit behind the first group. They are characterized by strong institutional development and high political stability, which translate into low security, labour market and tax policy risks. It has also identified groups of developing countries, differentiating between those with a higher exposure to the international crisis, either because they are emergent countries or because they are dependent on a developed economy that buys their exports, and those with low political stability and institutional development, which suffer higher levels of risk in several ways. However, in the present situation of international crisis, their macroeconomic risk level is low because they are hardly integrated in the international economy at all. It should not be overlooked that once the crisis abates, which may be expected in the near future, this risk will increase at the same time as it will decrease in the other countries. Therefore, it is advisable for these economies to institute measures that will make progress in economic, social and institutional development, in order to reduce the previously described risks and become attractive destinations for national and international investments. This way, higher standards of living will be achieved through higher income, wealth and lower unemployment. However, the groups identified by PCA offer, on certain occasions, controversial results when countries are included in the same group that have some kind of connection with the other individual countries but also have remarkable differences which suggests that they should not be catalogued within the same group. For instance in Group 1, where countries from Central and Eastern Europe, geographically close and recently integrated in the - 28 - European Union, are included in the same group as the principal economies in the world. The same may be said of Group 3 where South Korea and Italy are included in a group with much less developed countries. 5.2 CMLHL Projection The following CMLHL projections (Figs. 9-10) reflect the different groups of countries obtained with this neural model. These projections were obtained by applying the following values to the different parameters of CMLHL: number of iterations = 600, learning rate = 0.064, p parameter = 1.4, and τ parameter = 0.3. Fig. 9 depicts all the possible combinations of the third first factor pairs obtained through CMLHL by means of a scatter plot matrix. Factor pairs under the diagonal provide no extra information. Fig.9 << File Fig.09.tif goes here>> The main result obtained by CMLHL (factor pairs 1-2 from Fig. 9) is analyzed in depth in this section. - 29 - 5.2.1 CMLHL Factor Pair 1-2 Fig.10 << File Fig.10.tif goes here>> The groups identified in the CMLHL factor pair 1-2 analysis (Fig. 10) are described as follows. Group 1: 15 countries Some of the most important economies in the world are included in the first group, such as the United States or France, with other highly institutionally developed countries which include Denmark and Sweden. Comparing this case to group 1 in the PCA projection (Fig. 8), it can be seen that the CMLHL group is more homogeneous and selective, without the drawback of including additional economies such as the new members of the European Union, which despite the fact of being geographically close to other European economic dominant countries, still differ greatly with regard to risk levels. Group 2: 49 countries Group 2, which can be found very close to group 1, is also formed by developed countries, all of which have low levels of security, labour market and tax policy risk, but with slightly lower macroeconomic risk than the previous group. The composition is a little bit varied, ranging from Germany, Spain and Canada, to the new members of the European Union and even some of the most developed countries from Asia, namely Japan, and the United Arab Emirates. Anyway, the logic of this group is to invest in developed economies or those that are close to being so, with acceptable levels of economic, political and institutional development, but with a lower exposure to macroeconomic risks according to their scores. Group 3: 53 countries Group 3 shows many developing countries, with a clearly inferior development to those from groups 1 and 2, which translates into higher levels of security, labour market and tax policy risk, but also lower macroeconomic risk. It is worth pointing out that within the sample of developing countries, this group includes those relatively more developed countries, which is the case of Mexico, Russia, Brazil and Argentina. Group 4: 1 country (Iraq) This technique also considers this group to be a special case. As a country experiencing strong political instability, its situation differs radically from any other country in the sample. Group 5: 21 countries This group is characterized by developing countries that are less advanced than those of the previous group, such as Ecuador, Libya, Senegal, Congo, Bolivia, Chad or Ivory Coast. Therefore, risk levels are higher in all aspects except for the macroeconomic. This is due to their less important role in the international economy which implies less exposure to the global crisis. - 30 - Group 6: 18 countries The last of the groups comprises several African and Asian countries such as Kenya, Lesotho, Cambodia, Madagascar, Zimbabwe, Ghana or Gambia with a very high risk profile due to their low economic and institutional development and their political instability. However a remarkable difference with the countries of group 5 can be pointed out: while in the previous group macroeconomic levels were the lowest in the whole sample, in this case the levels are around the average, showing a relatively higher exposure to the crisis, probably due to the dependence of revenues coming from another economy through exports, as was also the case for PCA group 6. Group 7: 1 country (Iceland) The group can be considered a special case. CMLHL has been able to distinguish a country with unique characteristics. Despite the fact that the country only has 300,000 habitants, Iceland’s economy achieved spectacular growth over recent years, becoming an important finance centre in the world and one of the countries with the highest per capita income. However, it is also one of the countries that is suffering the consequences of the present crisis more than any other, with several multinational companies even leaving the country without any plans to return in the near future. The CMLHL projection (Fig. 10) shows this situation perfectly, as Iceland is not far away from group 2, showing a country that was on the road to becoming a solid developed country but which faces grave problems at present. In fact, security and tax policy risk are very low but the macroeconomic risk is extremely high. General Conclusions In conclusion, this technique has been able to differentiate 5 principal groups, two of which include developed countries and three include developing countries. Group 1 includes those developed countries that are more affected by the present crisis while group 2 includes those with a relatively lower exposure but which are also less strict about the requirements which define an economy as developed. Among the 3 groups that include the developing countries, group 3 includes those that are relatively more advanced, while groups 4 and 5 include those that are relatively less so. The former group is characterized by countries with a lower macroeconomic risk and the latter by countries with a higher level. In all of them, but especially in the last case, it is important once again to underline the need for structural measures in order to improve the situation, promoted by both local authorities and international bodies. CMLHL clustered the countries from the sample in fewer groups (excluding special cases), but they have a clearer, more refined and logical composition than those obtained by PCA (Fig. 8) where a group sometimes includes countries with a common characteristic but with also big differences which leads one to think that they should belong to different groups. This was the case for the new members of the European Union and also for Italy and South Korea. PCA (Fig. 8) included Iceland, a country where incredible growth rates were achieved in a short time ago, but which nowadays faces the disastrous consequences of a deep crisis, in Group 1, due to its low security and tax policy risk levels. However, CMLHL has been much more precise, differentiating this economy from the rest of countries in the sample. - 31 - 5.3 SOM Mapping As for the Spanish MNEs dataset, the SOM mapping of the Risk Briefing dataset is visualized by means of: � U-matrix (Fig. 11): distance matrix between the reference vectors of adjacent neurons in the two- dimensional lattice. � Numbered map (Fig. 12): neurons in the rectangular lattice were labelled with a figure that represented the number of companies associated with each neuron. To ease the analysis of this result, neurons were grouped according to their location on the map. Additionally, to make this mapping more intuitive, neurons were coloured following a topological ordering. These results are analyzed in the following paragraphs. The parameter values that generated this mapping were: random initialization, sequential training, hexagonal lattice, Gaussian neighbourhood function. The grid size was fixed to 9x6 by means of a heuristic formula. Fig.11 << File Fig.11.tif goes here>> Fig.12 << File Fig.12.tif goes here>> Group 1: 25 countries This group includes several countries with very low security and tax policy risk levels, around average labour market and very high macroeconomic risk levels such as Israel, Australia or Portugal. However, there are some others which do not seem to fit these characteristics. For example, security risks in Philippines, Georgia or Venezuela are extremely high, while macroeconomic risk in Burkina Faso or Qatar has lower macroeconomic risk than those developed economies Group 2: 13 countries Group 2 includes countries with slightly below average risk levels in all the different aspects analysed. This means that their institutional and economic development and their exposure to the crisis are intermediate, as happens in Jordan or Bulgaria. Again, there are some elements that, either because they have a much lower macroeconomic risk (Bangladesh and Senegal) or a much higher degree of development (Spain), do not seem to fit well with the rest of countries in the group. Group 3: 5 countries Group 3 is formed of the countries with the highest levels of security, labour market and tax policy risk. Iraq, Honduras or Yemen are locations where the risk that companies face is so high that investments are advised against unless they provide a very high rate of return. - 32 - Group 4: 23 countries In contrast, group 4 includes some of the most advanced economies such as France, Germany, Norway or Canada, where levels of risk are very low except for macroeconomic risk. Once again, the group is not completely homogeneous, since countries with different characteristic such as Iceland, Cuba or Botswana, are also included. Group 5: 10 countries High levels of macroeconomic risk are a key feature of group 5, derived from the high dependence from developed economies that are suffering the present crisis. However political and institutional development is not very advanced, which leads to high scores in the other risk aspects too. This is the case of Ecuador, Argentina or Kazakhstan. Strangely, it also includes two other countries with much lower security, tax policy and labour market risk: the United States and the UK. Group 6: 65 countries Group 6 has the most countries by a wide margin. It is made up of developing countries, both those which can be considered emerging economies such as Brazil, Turkey and China, and those that are less advanced, such as Sri Lanka, Zambia, Lesotho, Gambia and Ethiopia. Group 7: 17 countries Finally, the last group also includes developing countries –for instance, Burundi, Congo, Togo and Tanzania- with high scores in security, tax policy and market labour risks. However, in this case, their degree of development is the lower in the sample and their international linkages almost non existent, which leads to low exposure to the crisis and therefore to low macroeconomic risk levels. General Conclusions In conclusion, the SOM has offered overall results that are worse than the preceding two. The conclusions that may be drawn tend to point in the same direction, since groups are formed mainly either by developed or developing countries, depending on their exposure to macroeconomic risk and degree of economic, political, social and institutional development. However the SOM technique has offered results with several inconsistencies. Countries with different characteristics were included in the same group on certain occasions. Also, it was not possible to distinguish between different types of developing countries, since emerging and less advanced developing economies were included in the same group. Finally, special cases such as Iceland and Iraq were not isolated from the other countries, in spite of the very unusual situations they are undergoing. 5.4 CCA Fig. 13 presents the CCA projection of the Risk Briefing dataset. The parameter values that generated this projection were: standardized Euclidean distance, lambda = 160000, alpha = 0.4 and 7 epochs. Fig.13 << File Fig.13.tif goes here>> - 33 - The groups identified in the CCA projection can be described as follows: Group 1: 8 countries Group 1 includes global economic powers such as the United States, Japan, Canada and the UK which, as previously described, are characterized by high levels of institutional development and political stability, but also by a high exposure to the international crisis. However, this group also includes countries that do not share all those features, for instance Latvia and Costa Rica. Also in this group, the technique has not been able to differentiate the particular situation of Iceland, country that was starting to achieve a relatively advanced level of development but is now experiencing the worst consequences of the crisis. Group 2: 18 countries Group 2 also includes many developed countries, but with a relatively lower level of exposure to the crisis. Therefore, risk levels are very low in every single aspect except for the macroeconomic risk. In addition, this group is quite homogeneous, including countries such as Belgium, France, Finland, Germany, New Zealand, Norway and Switzerland. Group 3: 37 countries This group is the first that is mainly composed of emerging countries. However, it includes emerging countries with a more advanced development (i.e. Brazil, Turkey, Mexico or Russia) as well as others that lag some way behind such as Zambia, Uganda, Burundi or Sri Lanka. Group 4: 33 countries Different and heterogeneous countries are included in Group 4. More precisely, developed economies such as Australia, Spain, South Korea, Portugal and Italy are in the same group as emerging economies from Central and Eastern Europe (i.e. Poland, Slovakia or Czech Republic) and even some developing countries, such as Macedonia, Botswana or Azerbaijan. Group 5: 41 countries Countries with a low level of economic and institutional development and a high political instability, such as Ivory Coast, Gabon, Burkina Faso, Rwanda, Senegal, Togo, Namibia or Equatorial Guinea, comprise Group 5. Therefore, security, labour market and tax policy levels are high, while their low macroeconomic risk levels come from their low exposure to international risk given their almost non-existent connections with the international economy. Group 6: 20 countries As in the previous case, this group includes countries characterized by low levels of economic and institutional development, and high political instability, but with relatively higher levels of macroeconomic risk, resulting from - 34 - their dependence on a developed economy affected by the crisis. For instance, Honduras, Ecuador, Colombia or Pakistan share these features. Group 7: country (Iraq) This technique also identifies this group as a special case. As a country experiencing strong political instability, its situation differs radically from any other country in the sample. General Conclusions Once again, the CCA technique has been able to differentiate groups according to their level of economic and institutional development. The first two groups include the majority of the developed countries while developing countries comprise groups 4, 5 and 6. However, the composition of the groups has not been as homogeneous as expected, and group 3 is a good example of this, because it includes developed, emerging as well as developing countries. 5.5 Discussion and Conclusions for the Risk Briefing Case Study In the Risk Briefing case study, the four different techniques (CMHL, PCA, SOM and CCA) used to analyse several aspects of risk associated with possible locations for foreign direct investments, have offered similar results, showing the consistency and robustness of the conclusions that may be extracted from them. In general, countries are distributed into groups with similar levels of economic, political and institutional development. In fact, evidence suggests the existence of a different structure between developed and developing countries. Furthermore, within the group of developed countries, characterized by low levels of security, labour market and tax policy risk, two different sub-groups can be distinguished, according to their exposure to the present international crisis, which leads to different levels of macroeconomic risk. On the other hand, developing countries have also been included in different sub-groups depending on their relative degree of development, their dependence on other developed economies and their isolation from the international economy, which therefore have a lower level of macroeconomic risk due to their lower exposure to the crisis. However, the four projection techniques have offered results which differ in their accuracy and precision. On one hand, some of the sub-groups obtained with the CCA and SOM techniques are characterized by certain heterogeneity, since they include developed, emerging and developing countries. Moreover, both PCA and CMLHL offer much clearer and more homogeneous results. In fact, even though they all identified the particular situation of Iraq, as a country experiencing strong political instability, the CMLHL results picked out the special case of Iceland, showing the high level of detail and the precision of its results. - 35 - 6 General Conclusions and Future Work In general terms, it may be concluded that the projection models in general, and CMLHL in particular, are able to visualize and describe the situation of international companies and countries within the framework of political and country risk. This information is especially valuable to investors in a context of higher uncertainty derived from the global crisis. In fact, having access to reliable tools to provide clear, relevant and updated information to decision-makers can be an important source of competitive advantage both to managers in MNEs and policy- makers in countries. Future work will widen the international focus of the study to other areas and companies, for example from a different home country or to specific sectors where the institutional environment is more likely to have a greater impact on the location strategy (i.e. “regulated-sectors” such as banking or telecommunications) Also, other unsupervised neural models will be applied, for comparison purposes, to these interesting case studies. Acknowledgments. Alfredo Jiménez Palmero is grateful for the financial support from the Spanish Ministry of Science and Innovation through the FPU programme. This research has been partially supported through the Junta of Castilla and León under project BU006A08; the Spanish Ministry of Education and Innovation under project CIT-020000-2008-2 and CIT-020000-2009-12 . The authors would also like to thank the vehicle interior manufacturer, Grupo Antolin Ingenieria S.A., under project MAGNO2008 - 1028.- CENIT Project funded by the Spanish Government. - 36 - Figure Captions Fig. 1. PCA projection of the Spanish MNEs dataset. Fig. 2. CMLHL scatter plot matrix for the Spanish MNEs dataset. Fig. 3. CMLHL factor pair 1-2 for the Spanish MNEs dataset. Fig. 4. CMLHL factor pair 1-3 for the Spanish MNEs dataset. Fig. 5. SOM U-matrix of the Spanish MNEs dataset. Fig. 6. SOM coloured and numbered map of the Spanish MNEs dataset. Fig. 7. CCA projection of the Spanish MNEs dataset. Fig. 8. PCA projection of the Risk Briefing dataset. Fig. 9. CMLHL scatter plot matrix for the Risk Briefing dataset. Fig. 10. CMLHL factor pair 1-2 for the Risk Briefing dataset. Fig. 11. SOM U-matrix of the Risk Briefing dataset. Fig. 12. SOM coloured and numbered map of the Risk Briefing dataset. Fig. 13. CCA projection of the Risk Briefing dataset. - 37 - Appendix A: Country List and Codes for the Risk Briefing Dataset #1 Albania #2 Algeria #3 Angola #4 Argentina #5 Armenia #6 Aruba #7 Australia #8 Austria #9 Azerbaijan #10 Bahrain #11 Bangladesh #12 Belarus #13 Belgium #14 Belize #15 Benin #16 Bhutan #17 Bolivia #18 Bosnia and Herzegovina #19 Botswana #20 Brazil #21 Brunei #22 Bulgaria #23 Burkina Faso #24 Burundi #25 Cambodia #26 Cameroon #27 Canada #28 Cape Verde #29 Chad #30 Chile #31 China #32 Colombia #33 Congo #34 Costa Rica #35 Côte d'Ivoire #36 Croatia #37 Cuba #38 Cyprus #39 Czech Republic #40 Denmark #41 Dominican Republic #42 Ecuador #43 Egypt #44 El Salvador #45 Equatorial Guinea #46 Eritrea #47 Estonia #48 Ethiopia #49 Finland #50 France #51 Gabon #52 Gambia #53 Georgia #54 Germany #55 Ghana #56 Greece #57 Guatemala #58 Guinea #59 Guyana #60 Haiti #61 Honduras #62 Hong Kong #63 Hungary #64 Iceland #65 India #66 Indonesia #67 Iran #68 Iraq #69 Ireland #70 Israel #71 Italy #72 Jamaica #73 Japan #74 Jordan #75 Kazakhstan #76 Kenya #77 Kuwait #78 Kyrgyz Republic #79 Laos #80 Latvia #81 Lebanon #82 Lesotho #83 Libya #84 Lithuania - 38 - #85 Luxembourg #86 Macau #87 Macedonia #88 Madagascar #89 Malawi #90 Malaysia #91 Mali #92 Malta #93 Mauritania #94 Mauritius #95 Mexico #96 Moldova #97 Morocco #98 Mozambique #99 Myanmar #100 Namibia #101 Nepal #102 Netherlands #103 New Zealand #104 Nicaragua #105 Niger #106 Nigeria #107 Norway #108 Oman #109 Pakistan #110 Panama #111 Papua New Guinea #112 Paraguay #113 Peru #114 Philippines #115 Poland #116 Portugal #117 Qatar #118 Romania #119 Russia #120 Rwanda #121 São Tomé and Príncipe #122 Saudi Arabia #123 Senegal #124 Serbia #125 Seychelles #126 Singapore #127 Slovakia #128 Slovenia #129 South Africa #130 South Korea #131 Spain #132 Sri Lanka #133 Sudan #134 Swaziland #135 Sweden #136 Switzerland #137 Syria #138 Taiwan #139 Tajikistan #140 Tanzania #141 Thailand #142 Togo #143 Trinidad and Tobago #144 Tunisia #145 Turkey #146 Turkmenistan #147 Uganda #148 Ukraine #149 United Arab Emirates #150 UK #151 United States of America #152 Uruguay #153 Uzbekistan #154 Venezuela #155 Vietnam #156 Yemen #157 Zambia #158 Zimbabwe Table 1. Country List and Codes for the Risk Briefing Dataset. - 39 - References Ahlberg, C., & Shneiderman, B. (1994). Visual Information Seeking: Tight Coupling of Dynamic Query Filters with Starfield Displays. Paper presented at the SIGCHI conference on Human Factors in Computing Systems, Boston, Massachusetts, United States. Avramovic, D. (1958). Debt Servicing Capacity and Post War Growth in International Indebtedness. Baltimore, MD: Johns Hopkins Press. Baruque, B., & Corchado, E. (2010). A Weighted Voting Summarization of SOM Ensembles. Data Mining and Knowledge Discovery, (in press). Bengoa, M., & Sanchez-Robles, B. (2003). Foreign Direct Investment, Economic Freedom and Growth: New Evidence from Latin America. European Journal of Political Economy, 19(3), 529-545. Bevan, A. A., & Estrin, S. (2004). The Determinants of Foreign Direct Investment into European Transition Economies. Journal of Comparative Economics, 32(4), 775-787. Brouthers, L. E., Gao, Y., & McNicol, J. P. (2008). Corruption and Market Attractiveness Influences on Different Types of FDI. Strategic Management Journal, 29(6), 673-680. Corchado, E., & Fyfe, C. (2003). Connectionist Techniques for the Identification and Suppression of Interfering Underlying Factors. International Journal of Pattern Recognition and Artificial Intelligence, 17(8), 1447-1466. Corchado, E., Han, Y., & Fyfe, C. (2003). Structuring Global Responses of Local Filters Using Lateral Connections. Journal of Experimental & Theoretical Artificial Intelligence, 15(4), 473-487. Corchado, E., & Herrero, Á. (2010). Neural Visualization of Network Traffic Data for Intrusion Detection. Applied Soft Computing, ("Accepted with changes"). Corchado, E., MacDonald, D., & Fyfe, C. (2004). Maximum and Minimum Likelihood Hebbian Learning for Exploratory Projection Pursuit. Data Mining and Knowledge Discovery, 8(3), 203-225. Transparency International (2008) Corruption Perceptions Index. Retrieved 11/19/2008, from www.transparency.org. Cuervo-Cazurra, A. (2006). Who Cares about Corruption? Journal of International Business Studies, 37(6), 807-822. Cuervo-Cazurra, A. (2008). The Effectiveness of Laws Against Bribery Abroad. Journal of International Business Studies, 39(4), 634-651. Chen, S.-H., & Huang, Y.-C. (2007). Relative Risk Aversion and Wealth Dynamics. Information Sciences, 177(5), 1222-1229. Chen, Y., & Yao, Y. (2008). A Multiview Approach for Intelligent Data Analysis based on Data Operators. Information Sciences, 178(1), 1- 20. De la Fuente, J. M., Durán, J. J., & Jiménez, A. (2008). La importancia del riesgo país y político en la presencia exterior de las empresas multinacionales españolas (Working Paper). Paper presented at the XVIII ACEDE National Conference, León (Spain). Demartines, P. (1994). Analyze de données par réseaux de neurones auto-organizés. Institut National Polytechnique de Grenoble. Demartines, P., & Herault, J. (1997). Curvilinear Component Analysis: A Self-Organizing Neural Network for Nonlinear Mapping of Data Sets. IEEE Transactions on Neural Networks, 8(1), 148-154. Durán, J. J. (2001). Estrategia y economía de la empresa multinacional. Madrid: Pirámide. Durán, J. J., De la Fuente, J. M., & Jiménez, A. (2008). Political Risk as a Determinant of Investment by Spanish Multinacional Firms in Europe. Is There an East-West Structure? (Working Paper). Paper presented at the 34th EIBA Conference. Elbashir, M. Z., Collier, P. A., & Davern, M. J. (2008). Measuring the Effects of Business Intelligence Systems: The Relationship between Business Process and Organizational Performance. International Journal of Accounting Information Systems, 9(3), 135-153. Feinberg, S. E., & Gupta, A. K. (2009). MNC Subsidiaries and Country Risk: Internalization as a Safeguard against Weak External Institutions. The Academy of Management Journal (AMJ), 52(2), 381-399. Földiák, P. (1992). Models of Sensory Coding. PhD thesis, University of Cambridge. Friedman, J. H., & Tukey, J. W. (1974). A Projection Pursuit Algorithm for Exploratory Data-Analysis. IEEE Transactions on Computers, 23(9), 881-890. Fyfe, C., & Corchado, E. (2002). Maximum Likelihood Hebbian Rules. Paper presented at the 10th European Symposium on Artificial Neural Networks (ESANN 2002). Galan, J. I., Gonzalez-Benito, J., & Zuñiga-Vincente, J. A. (2007). Factors Determining the Location Decisions of Spanish MNEs: An Analysis based on the Investment Development Path. Journal of International Business Studies, 38(6), 975-997. García-Canal, E., & Guillén, M. F. (2008). Risk and the Strategy of Foreign Location Choice in Regulated Industries. Strategic Management Journal, 29(10), 1097-1115. Hannula, M., & Pirttimäki, V. (2003). Business Intelligence Empirical Study on the Top 50 Finnish Companies. Journal of American Academy of Business, 2(2), 593-601. Hashmi, M. A., & Guvenli, T. (1992). Importance of Political Risk Assessment Function in U.S. Multinational Corporations. Global Finance Journal, 3(2), 137-144. Henisz, W. J. (1998). The Institutional Environment for International Investment. Hass School of Business, University of California. Henisz, W. J. (2003). The Power of the Buckley and Casson Thesis: The Ability to Manage Institutional Idiosyncrasies. Journal of International Business Studies, 34(2), 173-184. Henisz, W. J., & Zelner, B. A. (2001). The Institutional Environment for Telecommunications Investment. Journal of Economics & Management Strategy, 10(1), 123-147. Henisz, W. J., & Zelner, B. A. (2002). Interest Groups, Institutional Structures and Electricity Investment: Mimeo. Herrero, Á., et al. (2009). Neural Projection Techniques for the Visual Inspection of Network Traffic. Neurocomputing, 72(16-18), 3649- 3658. Herrero, Á., et al. (2010). DIPKIP: A Connectionist Knowledge Management System to Identify Knowledge Deficits in Practical Cases. Computational Intelligence, 26(1), 26-56. Hillman, A. J., & Hitt, M. A. (1999). Corporate Political Strategy Formulation: A Model of Approach, Participation, and Strategy Decisions. The Academy of Management Review, 24(4), 825-842. Holburn, G. L. F. (2001). Regulatory Institutions and Firm Strategy: Theory and Evidence from the Electric Power Industry. University of California, Berkeley. - 40 - Hotelling, H. (1933). Analysis of a Complex of Statistical Variables into Principal Components. Journal of Education Psychology, 24, 417- 444. Heritage Foundation (2005) Index of Economic Freedom. Retrieved 11/19/2008, from www.heritage.org. Juliana, Y., & Heather, M. (2005). Comparison of Country Risk Models: Hybrid Neural Networks, Logit Models, Discriminant Analysis and Cluster Techniques. Expert Systems with Applications, 28(1), 137-148. Kapuria-Foreman, V. (2007). Economic Freedom and Foreign Direct Investment in Developing Countries. The Journal of Developing Areas, 41(1), 143-154. Kobrin, S. J. (1979). Political Risk - Review and Reconsideration. Journal of International Business Studies, 10(1), 67-80. Kohonen, T. (1990). The Self-Organizing Map. IEEE, 78(9), 1464-1480. Kohonen, T. (1997). Self-Organizing Maps (Vol. 30): Springer-Verlag New York, Inc. Lall, S. (1996). Learning from the Asian Tigers: Studies in Technology and Industrial Policy: Macmillan. Li, H., & Sun, J. (2009). Gaussian Case-Based Reasoning for Business Failure Prediction with Empirical Data in China. Information Sciences, 179(1-2), 89-108. Marois, B. (1979). Assessment and Management of Political Risk: Practice of French Firms: Annual Meeting of the Academy of International Business. London. Marois, B. (1981). Comment les Enterprises Francaises Gerent le Risque Politique. Revue Francaise de Gestion, May-August, 4-9. Mortanges, C. P. d., & Allers, V. (1996). Political Risk Assessment: Theory and the Experience of Dutch Firms. International Business Review, 5(3), 303-318. Mutinelli, M., & Piscitello, L. (1997). Differences in the Strategic Orientation of Italian MNEs in Central and Eastern Europe. The Influence of Firm-specific Factors. International Business Review, 6(2), 185-205. Spanish Ministry of Industry, Tourism and Commerce (2009) Network of Economic and Commercial Spanish Offices Abroad. Retrieved 11/28/2008, 2008, from http://www.oficinascomerciales.es. Noordin, B. A., Harjito, D. A., & Hazir, A. Y. (2006). Political Risk Assessment of Malaysian based Multinational Corporation. Problems and Perspectives in Management, 4(3), 91- 99. Oja, E. (1989). Neural Networks, Principal Components, and Subspaces. International Journal of Neural Systems, 1, 61-68. Parks, T. H. (1996). A Foreign Market Entry Decision Model: A Comparison of Theoretical and Actual Practices of Multinational Enterprises. Dissertation Abstracts International, 57(10), 4447. Patterson, D. W. (1998). Artificial Neural Networks: Theory and Applications: Prentice Hall PTR Upper Saddle River, NJ, USA. Pearson, K. (1901). On Lines and Planes of Closest Fit to Systems of Points in Space. Philosophical Magazine, 2(6), 559-572. Rich, G., & Mahmoud, E. (1990). Political Risk Forecasting by Canadian Firms. International Journal of Forecasting, 6(1), 89-102. The Economist Intelligence Unit (2009) Risk Briefing. Retrieved 04/11/2009, from http://viewswire.eiu.com/index.asp?layout=RKAllCountryVW3. Ritter, H., Martinetz, T., & Schulten, K. (1992). Neural Computation and Self-Organizing Maps; An Introduction: Addison-Wesley Longman Publishing Co., Inc. Saini, K. G., & Bates, P. S. (1984). A Survey of the Quantitative Approaches to Country Risk Analysis. Journal of Banking & Finance, 8(2), 341-356. Seung, H. S., Socci, N. D., & Lee, D. (1998). The Rectified Gaussian Distribution. Advances in Neural Information Processing Systems, 10, 350-356. Slangen, A. H. L., & van Tulder, R. J. M. (2009). Cultural Distance, Political Risk, or Governance Quality? Towards a More Accurate Conceptualization and Measurement of External Uncertainty in Foreign Entry Mode Research. International Business Review, 18(3), 276- 291. Laboratory of Computer and Information Science. Helsinki University of Technology (2009) SOM Toolbox for Matlab. Retrieved 24/06/2009, from http://www.cis.hut.fi/projects/somtoolbox/. Wan, W. P. (2005). Country Resource Environments, Firm Capabilities, and Corporate Diversification Strategies. Journal of Management Studies, 42(1), 161-182. Wei, S.-J. (1997). How Taxing is Corruption on International Investors? Review of Economics and Statistics, 82(1), 1-11. Wernerfelt, B. (1984). A Resource-based View of the Firm. Strategic Management Journal, 5(2), 171-180. Wheeler, D., & Mody, A. (1992). Institutional Investment Location Decisions, the Case of US Firms. Journal of International Economics, 33, 57–76. Fig. 1. PCA projection of the Spanish MNEs dataset. http://ees.elsevier.com/eswa/download.aspx?id=75103&guid=46ec9484-adf3-48d0-b782-39b3c07d49d3&scheme=1 Fig. 2. CMLHL scatter plot matrix for the Spanish MNEs dataset. http://ees.elsevier.com/eswa/download.aspx?id=75091&guid=1121d12a-f0cd-472f-a9b7-b5acfdfea39b&scheme=1 Fig. 3. CMLHL factor pair 1-2 for the Spanish MNEs dataset. http://ees.elsevier.com/eswa/download.aspx?id=75092&guid=8ce87d3c-4bee-459e-9ef2-444f182987f5&scheme=1 Fig. 4. CMLHL factor pair 1-3 for the Spanish MNEs dataset. http://ees.elsevier.com/eswa/download.aspx?id=75093&guid=6247d920-e781-4f98-9612-5b77a3ce136b&scheme=1 Fig. 5. SOM U-matrix of the Spanish MNEs dataset. http://ees.elsevier.com/eswa/download.aspx?id=75094&guid=d5c4c039-b9b8-49bd-8a07-2408843156de&scheme=1 Fig. 6. SOM coloured and numbered map of the Spanish MNEs datase http://ees.elsevier.com/eswa/download.aspx?id=75095&guid=c60744a3-ded6-4919-8899-bce7cdbed14a&scheme=1 Fig. 7. CCA projection of the Spanish MNEs dataset. http://ees.elsevier.com/eswa/download.aspx?id=75096&guid=b8708bba-8abc-49d3-9485-f3bfbd1ace60&scheme=1 Fig. 8. PCA projection of the Risk Briefing dataset. http://ees.elsevier.com/eswa/download.aspx?id=75097&guid=38df1260-2473-41e3-8859-10ed3cddbc1f&scheme=1 Fig. 9. CMLHL scatter plot matrix for the Risk Briefing dataset. http://ees.elsevier.com/eswa/download.aspx?id=75098&guid=5ab9cef1-c9de-4ec2-86b6-068dfe755f8a&scheme=1 Fig. 10. CMLHL factor pair 1-2 for the Risk Briefing dataset. http://ees.elsevier.com/eswa/download.aspx?id=75099&guid=60372a3e-b5ee-426a-8eb1-2cbe9824ccf7&scheme=1 Fig. 11. SOM U-matrix of the Risk Briefing dataset. http://ees.elsevier.com/eswa/download.aspx?id=75100&guid=8bc5e97f-e867-494f-b74c-6df78cf8f5e6&scheme=1 Fig. 12. SOM coloured and numbered map of the Risk Briefing data http://ees.elsevier.com/eswa/download.aspx?id=75101&guid=a6e7eb25-2889-464c-a969-9c3ce2353de7&scheme=1 Fig. 13. CCA projection of the Risk Briefing dataset. http://ees.elsevier.com/eswa/download.aspx?id=75102&guid=5c4db08c-2f6c-4eb0-8dca-3ea92a5184ca&scheme=1 
work_c4wisygkkjb6jiiafjsh5yovye	----	Towards expert systems for the selection of search keys Towards Expert Systems for the Selection of Search Keys* Raya Fidel Graduate School of Library and lnforma tion Science, University of Washington, Seattle, WA 98795 Intermediary expert systems are designed to mediate between end-users and complex information retrieval systems. However, since most of these expert systems are based on text analysis rather than on models of hum man searching, they cannot process requestrelated cri- teria, such as precision or recall requirements. Analysis of the searching behavior of human intermediaries re- vealed a routine for the selection of search keys-free- text or controlled vocabulary-along a decision tree. Examples of decision rules demonstrate that although further research is required, these rules can be auto- mated to significantly enhance the adaptability of inter. mediary expert systems. Introduction It is believed that end-users will very likely search their own requests online when search processes are simplified or made friendlier. The prevailing approach to providing easier and friendlier user-system communication is to de- velop “interface” or “intermediary” systems. Indeed, systems such as CITE [l] are already available for public access, and others, such as CONIT [2] or CANSEARCH [3], are being tested in experimental settings. Through such systems, users are freed from encoun- ters with many peculiarities in databases and search sys- tems, and yet can benefit from a large range of capabili- ties. In particular, users can enter a request in a loosely structured format, preferably in a natural language, sen- tence-like expression. An intermediary system processes the request terms, displays information to users, and asks for some sort of feedback. The information displayed may be in the form of a list of subject areas, databases, search keys (i.e., strings of characters that are entered to *This study was supported by the Graduate School of the University of Washington. Received March I, 1985; revised May 15, 1985; accepted June 13, 1985 be searched for occurrence in pre-defined fields), or ac- tual citations from which users are asked to make a selec- tion, possibly in ranked order. Interactions of this nature usually proceed until users terminate the session. Some intermediary systems are actually helper sys- tems: they support shortcuts in end-user training by pro- viding menu-driven interaction or online help programs. Typically, helper systems require end-users to make most of the decisions during a search process, or they drasti- cally simplify searching by reducing the number of op- tions to a minimum. Intermediary expert systems, on the other hand, can be quite powerful. Expert systems replicate the perfor- mance of an expert in a particular area by incorporating the knowledge of an expert with rules for making infer- ences on the basis of this knowledge. Systems knowledge may or may not be intended to model actual processes as they would be performed by human experts. This article illustrates how to model searching behav- ior of human intermediaries by demonstrating that for- mal rules for the selection of search keys can be extracted from human experts. These rules cannot be incorporated yet into a knowledge base of an intermediary expert sys- tem because they are incomplete and were derived from searching behavior that is limited in its subject area. The work presented here, however, clearly indicates that with more research a complete set of rules can be established. It also provides guidance for future exploration and points to various issues that could be readily considered by designers of intermediary expert systems. Modeling the Selection of Search Keys Search processes consist of three basic intellectual components: (1) definition of query structure; (2) selec- tion of search keys; and (3) evaluation of feedback. We focus here on the second component. In a database that JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE. 37(1):37-44,1986 CCC 0002-8231/86/010037-08$04.00 offers the capability, an intermediary system must exam- ine each term of a request and consider its representa- tion: as a controlled vocabulary key, as a free-text key, or as both. Expert systems vary in the degree of freedom they provide users in this selection: some dutifully search only those search keys designated by a user; some use search keys designated by a user to generate additional search keys; and some automatically generate search keys without user participation. Intermediary expert systems use mapping algorithms to generate search keys. Whether user support and in- volvement are high or low, a term may be mapped to a descriptor (a search key from a controlled vocabulary) through an exact or other kinds of match, or it may not be mapped to a descriptor at all. While some systems, such as CANSEARCH [3], use subject-specific ap- proaches, most existing intermediary systems, expert and helper, use mapping algorithms that are based on text characteristics, such as word-occurrence frequency or statistical associations. The most apparent drawback of text analysis al- gorithms is their lack of sensitivity to request-specific re- quirements that cannot be directly deduced from a re- quest statement. Consider, for example, a request about “the attitudes of anorexic students toward themselves during examinations periods.” One user is interested only in anorexic students, another wants to get all the in- formation available on the topic but is primarily inter- ested in students and is willing to look at material about student behavior during examinations periods. The sys- tem decides, say, to use anorexic students as a single search key, but may or may not suggest the term students as a search key, following its own algorithm. However, when experienced online searchers select search keys, they examine not only the degree of term- descriptor match, but they also consider other factors. Moreover, these additional factors are quite frequently essential to the success of a search. One wonders why the experience of human intermediaries has been neglected by system designers when it is an important source of knowledge. There may be two explanations. First, online search- ing behavior was not being investigated and thus could not contribute knowledge. Second, research and develop- ment in information storage and retrieval is familiar with text analysis because of the long and established experi- ence in automated indexing. Although methods and ap- proaches to automated indexing vary greatly, most of them rely on text analysis and are thus text-oriented rather then user-oriented. Only recently, in an experi- ment at the American Petroleum Institute, a first attempt has been made to develop an automated indexing system which models indexing behavior [4]. As a first step toward knowledge engineering in inter- mediary expert systems, I analyzed the online searching behavior of several human intermediaries and found that online searchers do indeed follow some rules [S]. In their selection of search keys, they utilize informal and some- times highly intuitive decision rules. Moreover, these rules can be detected, examined, and presented in a for- mal structure which can be processed by computers. This formal presentation incorporates knowledge of multiple experts and with further research can be used in a knowl- edge base for intermediary expert systems. The Study Method To examine online searching behavior, eight searchers were observed performing their regular, job-related searching [6]. Searchers who have been searching for more than two years were recruited from among informa- tion specialists in scientific areas, primarily in the life sci- ences. They were studied one at a time, and were asked to verbalize their thought processes during their searching to the degree that speaking out loud would not interfere with their performance. These verbalizations, including the creation of search strategy, were recorded. At the end of the observation period (approximately 10 to 15 searches), each searcher was interviewed to reveal and clarify information not accessible to observation. Data collected for analysis were about one hundred printed search protocols with transcriptions of verbalized thought processes and additional explanations from the final interviews. Each instance in which searchers had selected a search key was then identified and the reason for the specific choice was explicitly noted. Analysis of the first ten search protocols generated a preliminary list of condi- tions under which a particular selection was made. For example, a condition for choosing a free-text key is: to enter straightforwardly a specific concept which might not be a trustworthy descriptor. All the search protocols were then analyzed against this preliminary list of conditions. Each instance in which searchers had selected a search key was listed under the condition to which it applied. Instances whose condition could not be found suggested a new condition to be added to the list. This analysis revealed that most conditions were considered by most searchers, and only a few combi- nations reflect searchers’ individual idiosyncrasies. The list of conditions for the selection of search keys is presented in Figure 1 in the form of a decision tree. This set of decision rules is called here “the selection routine.” The selection routine specifies conditions which are necessary for a searcher to be able to select a particular type of search key. It describes the most commonly se- lected path, but there might be complications. As such, the selection routine is not deterministic: it cannot always accurately predict the selection of search keys unless other factors and their impact are known. This routine groups together similar conditions so it could be pre- sented in a decision tree, but it is not meant to represent a necessary sequence in the selection process. A list of the options in search key selection and the conditions necessary for each option is given in Table 1. 38 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 39 TABLE 1. A List of Options and the Associated Conditions. The Selection Routine OPTION CONDITIONS Use descriptors Add the next broader descriptor in the hierarchy Use generic descriptors in an inclusive mode Limit to retrieval by descriptors Limit to major descriptors Specify document type Use free-text terms Use free-text terms to probe indexing Use descriptors as free- text terms in other databases Use free-text terms for an inclusive search Use free-text terms to introduce uncommon types of search keys Descriptor Searching A term is a common term + it is mapped to a descriptor [A]. A term is a single meaning term + it is mapped to a descriptor + the descriptor is an exact match [D]. the concept has many synonyms [E2]. the concept is not clear to the searcher E31. the concept may not be explicitly mentioned [E4]. the descriptor is a partial match [F]. the descriptor is a broader term [.I]. A term is a single-meaning term + it is mapped to a descriptor + recall needs to be improved [E7]. A term is a single-meaning term + it is mapped to a descriptor -I- recall needs to be improved [E8]. A term is a single-meaning term + it is mapped to a descriptor + precision needs to be improved [E9]. A term is a single-meaning term + it is mapped to a descriptor + precision needs to be improved [ElO]. A term is a single-meaning term + it is mapped to a descriptor + precision needs to be improved [E12]. Free-Text Searching A term is a single-meaning + it is mapped to a descriptor + the concept is not “trustworthy” as an index term [El]. the descriptor is a broader term [HI. it cannot be mapped to a descriptor [K]. A term is a common term + it cannot be mapped to a descriptor [B]. A term is a single-meaning term + it cannot be mapped to a descriptor [L]. A term is a single-meaning term + it is mapped to a descriptor + a request needs to be searched on several databases [E13]. A term is a single-meaning term + it is mapped to a descriptor + the descriptor is a partial match [Cl. A term is a single-meaning term + it cannot be mapped to a descriptor [Ml. Other Combinations Use free-text terms in combination with descriptors Add free-text synonyms to descriptors Add truncated free-text terms to descriptors Add role indicators A term is a single-meaning term + it is mapped to a descriptor + the descriptor is a broader term [I]. A term is a single-meaning term + it is mapped to a descriptor + recall needs to be improved [ES]. A term is a single-meaning term + it is mapped to a descriptor + recall needs to be improved [E6]. A term is a single-meaning term + it is mapped to a descriptor + Change database precision needs to be improved [El I]. A term is a common term + it cannot be mapped to a descriptor [Cl. Before searchers decide how to represent a request term in a query formulation they must answer two central questions: (a) can a term be mapped to a descriptor, and (b) is it a “good” term for free-text retrieval. A searcher maps a term to a descriptor when she/he has decided that a particular descriptor best represents a request term, whether or not there is an exact match between the term and the descriptor. The second question is a little more complex and re- quires some explanation. Searchers consider a term to be a “good” term for free-text searching if it: (1) usually oc- curs in a particular context, (2) is uniquely defined, and (3) is specific in the concept it represents. Such a term will be called here a “single-meaning” term. On the other hand, a term that occurs in more than one context will be called a “common” term. For example, in the request about the attitude of anorexic students toward them- selves during examinations periods, terms such as ano- rexia and students are single-meaning terms. The term examination, on the other hand, is a common term; it can occur in a subject-related context (“the best way to take student examinations”), or in a descriptive capacity (“ex- amination of students’ responses”), or still further, it can be used very loosely to represent the concept of an inquiry of any kind. When a term is a common term, searchers do not have much choice in the selection of search keys: if it can be mapped to a descriptor, searchers almost always enter the descriptor as the search key [A] (i.e., option [A] in Figure 1. and Table 1.) because, by definition, it is not desirable to use a common term as a free-text search key. A common term that cannot be mapped to a descrip- tor almost always results in unsatisfactory retrieval. Searchers have no choice but to enter a free-text key, preferably in combination with other search keys, in or- der to retrieve citations, select some relevant ones and re- view their indexing in an attempt to find descriptors that might possibly be relevant [B]. For example, if the term examination cannot be mapped to a descriptor, one can devise a formulation (using the AND operator) that com- bines the terms students, anorexia, and the free-text term examination. Reviewing a sample of retrieved cita- tions, one may find that all the relevant citations include the descriptor Instructional Tests in their indexing, thus suggesting that this descriptor is an appropriate choice for the representation of the concept examinations. Such probing does not always further a search and searchers may then decide to select a different database: one which does allow the common term to be mapped to a descriptor [Cl. Single-meaning terms provide more options. If a sin- gle-meaning term cannot be mapped to a descriptor, searchers may enter a free-text term as the only search key [K], but they may also probe indexing of relevant ci- tations to make sure that no adequate descriptor is over- looked [L]. 40 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1966 In some cases, searchers may use a free-text key to search for a single-meaning term that cannot be mapped to a descriptor in a special way: they require that it occurs in a field other than the common ones, such as the jour- nal title field [Ml. Suppose a user is interested only in the psychological aspects of anorexic students, and suppose that the term psychology cannot be mapped to a descrip- tor. Searchers may predict that searching for the occur- rence of psychology in the text would retrieve a large number of irrelevant citations, and decide instead to re- trieve citations to articles whose authors are affiliated with organizations which include the stem psych in their titles, or articles that were published in sources whose ti- tles include this stem. The least problematic term is one that is single-mean- ing and also can be mapped to a descriptor. Such a term can be entered either as a controlled vocabulary, as a free-text key, or as both. Here, searchers are free to deal with other factors when selecting search keys. It is useful to show the conditions under which searchers select free- text keys, and those under which they choose to use de- scriptors. Selection of Free-Text Search Keys A single-meaning term can be mapped to a descriptor through an exact match, partial match, or a term might be mapped to a broader descriptor. When a single-mean- ing term is mapped to a descriptor through an exact match, searchers may use a descriptor [D], or elect to consider a variety of factors [El. Partial match usually implies mapping a request term to a narrower descriptor, or to a group of narrower de- scriptors. If suitable, searchers use a free-text key to in- clusively search concepts that are not grouped together by the hierarchy of the controlled vocabulary [G]. If, for ex- ample, the request term students is mapped to descrip- tors such as Foreign Students, College Students, or Un- dergraduates, and a descriptor Students does not exist, the free-text key can be used to retrieve information about any type of student. It should be noted that in some search systems use of the free-text key students also would retrieve citations that are indexed with descriptors which include the term. This is a source for constant con- fusion for searchers because the routine changes from one search system to another. When a single-meaning request term is mapped to a broader descriptor, searchers may prefer to preserve the specificity of the request and use free-text search keys [HI. If they are concerned with the precision of the set to be retrieved, they enter free-text terms in combination (using the AND operator) with the broader descriptor to which it is mapped [I]. Regardless of the degree of match between a single- meaning term and a descriptor, searchers may still prefer to use free-text search keys for three reasons. First, if re- call needs to be improved, searchers use both descriptors and free-text terms as search keys [ES]. For a further in- crease in recall, free-text keys are entered in a truncated form [E6]. Second, if searchers think that a particular descriptor may be assigned inconsistently by indexers, they consider the use of a free-text key to be more trust- worthy [El]. Suppose, for instance, a controlled vocabu- lary includes the descriptor Nutrition and also the de- scriptor Diet-each one to be assigned for distinct representation of a subject. When looking for nutrition- related problems of anorexic students, searchers may find the distinction confusing and thus assume that in- dexers are likely to be inconsistent in assigning these de- scriptors. To compensate for indexers’ errors, they may use both a descriptor and free-text keys, or only free-text keys. Lastly, if a request is to be searched on several data- bases, searchers may map single-meaning terms to de- scriptors in only a few of the relevant controlled vocabu- laries. Running the same query formulation against several databases, they then, in fact, search some of the terms as free-text keys in some of the databases [E13]. Selection of Controlled Vocabulary Search Keys The most straightforward use of a descriptor to repre- sent a single-meaning term is when a term is exactly matched with a descriptor and no other apparent con- straints exist. However, searchers may elect to enter a re- quest term as a descriptor when it is mapped to a descrip- tor through a partial match [F], in which case it is mapped to a narrower descriptor, or when the term is mapped to a broader descriptor [.I]. Such decisions de- pend on the nature of the request, and when searchers suspect that precision might not be satisfactory, they may combine free-text terms with a descriptor [I]. Even single-meaning terms may have some attributes that will make them unattractive for free-text searching. Regardless of the degree or nature of a term-descriptor match, searchers most often prefer to enter a descriptor when: (1) a term has many synonyms [E2]; (2) a concept and its use is not clear to a searcher [E3]; or (3) when a concept is likely to be implied rather then explicitly men- tioned in the searched text [E4]. The previously mentioned request about anorexic stu- dents provides a clear example of the last condition. The request concept attitudes toward themselves can indeed be entered and searched as a free-text phrase in most search systems. However, this concept can be expressed, directly or indirectly, in various other phrases, such as “students displayed attacks of self-hatred”, or “narcis- sism level dropped with time.” Searchers, then, consider descriptors such as Self Image, or Se&Esteem to be more reliable than free-text terms for searching. In addition to providing search keys for “problematic” terms, controlled vocabularies provide special means to improve precision and recall. When searchers perceive precision and/or recall to be unsatisfactory, they may de- cide to take advantage of these means and elect to use a descriptor to represent a single-meaning term. Although JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1988 41 these routines are quite well known, it might be helpful to mention them. When searchers decide that recall needs to be improved, they select a controlled vocabulary key because they can add the next broader descriptor in the hierarchy [E7], or use generic concepts in an inclusive mode [E8]. Controlled vocabularies readily suggest broader descriptors, and thus make it convenient to in- deed broaden a concept. Moreover, broadening the meaning of a concept is not always possible in free-text searching since the broader concept may be a common term. Inclusive searching-which is quite straightfor- ward in descriptor searching-facilitates retrieval of cita- tions indexed under a descriptor as well as those indexed under its narrower terms. Lastly, when searchers predict that precision may not be satisfactory, they may select a controlled vocabulary key because they can exercise various ways of adding weights to search keys such as: to limit to retrieval by de- scriptor only [E9]; to limit a descriptor to be a major de- scriptor [ElO]; to add role indicators [Eli]; or to specify document type [E12]. Discussion The selection routine clearly shows that the process of selecting search keys as performed by online searchers can be formalized into a decision tree. Moreover, several suggestions for improvements in existing systems are im- mediately apparent; others will require more research. First, the pattern of the selection routine illustrates the significance of decisions made during search key selec- tion to the success of a search. This pattern shows that when a term is “not adequate” for searching, i.e., it is a common term and/or it cannot be mapped to a descrip- tor, only a few options are available for searching. Only six of the twenty-five options in the selection routine are suggested for such terms, and some, such as the use of free-text terms to probe indexing, require a fair amount of creativity on the part of an intermediary. On the other hand, when a term is “good” for searching, i.e., it is a single-meaning term and it can be mapped to a descrip- tor, intermediaries can look at several options, as pre- sented in the check-list. In other words, only after termi- nological difficulties or peculiarities in representing request terms have been overcome for searching, can an intermediary consider additional factors that are essen- tial to the success of the search. Therefore, intermediary expert systems must be able to resolve terminological dif- ficulties before they can be equipped to deal with addi- tional factors. Second, using the selection routine, one can identify flaws in existing intermediary expert systems and at the same time propose methods to overcome these failings. While some flaws can be readily identified and possible remedies suggested already at this time, other issues re- quire additional research before their nature can be clearly defined. To demonstrate the ability of the selec- tion routine to illuminate such issues, a few examples are discussed below. One of the flaws in existing intermediary expert sys- tems that readily stands out is their inability to distin- guish between common and single-meaning terms. This distinction is important because if a term is a common term, experienced searchers almost always select it as a descriptor even if they have to change databases (unless they use it in a trial [B]). Yet, to my knowledge, no exist- ing system, provides safeguards to advise end-users against the use of common terms as free-text search keys. For example, when asked about “information retrieval, ” CITE suggested Information Systems, Information Ser- vices, Information Theory, and Information Centers, as descriptors, and retrieval, retrieving, and information as free-text search keys [ 11. Experienced searchers normally will avoid using these common terms as free-text keys, though end-users may prefer them because none of the descriptors exactly matches the original concept. The idea that some terms are not suitable for auto- mated processing is not new. In linguistics, homonymy and polysemy are specific cases of common terms which might be better described as ambiguous terms or terms that belong to more than one semantic domain. These concepts are essential to thesaurus construction. From the information science field, Fugmann, for example, differentiates between “individual” and “general” con- cepts, the latter being non-lexical concepts which are bet- ter searched with controlled vocabulary [7]. In addition, the assumption that some terms carry more information than others is a fundamental premise in automated in- dexing and abstracting. Control over common terms should require relatively modest modifications in intermediary expert systems. First, we have to devise a working definition of what con- stitutes a common term. For this purpose, it would be useful to test the hypothesis that terms which occur with high frequency in a database are also common terms. Suppose, however, that a system selects to define a com- mon term by, say, a consensus among three knowledge- able searchers who are highly experienced with a data- base. These searchers can then check each term in a thesaurus, including those that represent an entry or part of an entry and those that occur in a descriptor or in a lead-in entry, and determine which terms are common. Common terms can then be coded so they are not dis- played or used as free-text search keys. This method requires additional effort to identify com- mon terms that do not occur in a thesaurus or in other semantic networks. Some shortcuts can be devised, how- ever, such as the use of a number of thesauri. On the other hand, other methods to identify common terms may apply to all terms in a database whether or not they are listed in a thesaurus. For instance, a test can be con- ducted to discover *whether a correlation exists between the frequency in which a word occurs in a database and its adequacy for free-text searching. If a well-selected 42 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 sample of databases proves that common terms occur with high frequency in those databases and vice versa, then common terms can be singled out by frequency counts. Once a common term is defined, an intermediary sys- tem can interact with users. Suppose a common term cannot be mapped to a descriptor. By giving messages to users, a system can interrogate them to determine whether to select another database or whether to enter the term as a free-text search key to probe indexing. This interaction may reveal request-related requirements that are not reflected in a request statement, e.g., that the term is central to the request or that it always should be associated with another term. A second example of a flaw in existing intermediary expert systems is their failure to suggest the use of free- text terms for inclusive search [G]. As explained earlier, if the descriptor Students does not exist and a term is then mapped to descriptors such as University Students or Gifted Students through partial match, the free-text term students can be entered to search for any kind of student. This function can be easily automated. How- ever, one should be careful because some terms are not suitable to be entered as free-text keys for inclusive searching. The term attitudes is a case in point. If an exact match does not exist, the term attitudes can be mapped through a partial match to descriptors such as Employee Atti- tudes, Mother Attitudes, or Negative Attitudes. In searching for material about attitudes of students toward themselves, a user may find the last descriptor relevant, but the first two are a source for unwanted retrieval. To search the term attitude as a free-text key would magnify the problem. In this case, the user is better advised to scan all the descriptors to which the term is mapped and to select the relevant ones. Thus, we can designate for each word in a multi-words descriptor whether it can be automatically searched as a free-text key or whether it should be displayed to users when a partial match occurs. At this time, we do not have any scale based on systematic investigations that can de- termine which terms are suitable for inclusive searching in a free-text mode. We can, however, adapt a pragmatic approach and determine the status of each term by con- sulting with experienced online searchers. In the future, research in terminology may provide more rigorous crite- ria. A third example is the inadequacy of intermediary ex- pert systems for term analysis processes. Various statisti- cal approaches could be used to determine which attrib- utes of terms are significant for searching. For instance, the “trustworthiness” of a descriptor can be measured by the degree of consistency with which it is assigned. An extrapolation of the measurement of term consistency suggested by King & Bryant [8] could be used to measure trustworthiness of descriptors. A test database could be indexed by several indexers. A measurement for trust- worthiness could then be determined by the ratio between the number of documents to which a descriptor has been assigned by all indexers and the total number of docu- ments to which it has been assigned by any number of indexers (possibly weighted by the number of indexers se- lecting to assign a descriptor for each document). Here again, each descriptor that is not trustworthy can be flagged. During the search process, a system can then use free-text terms whenever it encounters a single-meaning term which is mapped to such a descriptor. The system may also convey its action to users. As these examples show, parts of the selection routine can be automated quite easily and with existing tech- niques. Other parts of the selection routine, such as, when searchers use a descriptor for a single-meaning term because they do not fully understand the concept it represents, may prove to be unsuitable for the design of intermediary expert systems. There are yet further decisions which can usefully be implemented after additional research is performed. Consider a situation in which a single-meaning term is mapped to a broader descriptor. There are three main options as shown in the decision tree (Figure 1.) at the points [HI, [I], and [J]. Now, suppose the term anorexia nervosa is mapped to the descriptor Appetite Disorders. An intermediary system may decide to enter anorexia nervosa as a free-text search key and possibly retrieve documents in which the subject is only mentioned, rather then discussed. Or, it can combine this free-text key with the descriptor Appetite Disorders, using the AND opera- tor, to retrieve articles that indeed discuss the disorder but may miss relevant documents that were not indexed under the broader descriptor. Or, the system can select to enter the descriptor, in which case relevant documents might still be missed while documents discussing appetite disorders other then anorexia nervosa probably will be re- trieved. No option is better than any other; it all depends on the nature of a request. In other words, one option is re- quired for one type of request and another is required for another type. We do not have yet a general typology of requests that we can use to support the selection of the best option. Further research in online searching behav- ior, however, can provide criteria to be used in automated systems. Statistical approaches may suggest some help. For ex- ample, one may statistically analyze user satisfaction rate with each of the options. Thus, even though we do not know explicitly which type of request requires which op- tion, we can implicitly detect which type is most common among a particular group of users by the option they find to be most satisfactory. We can then design an intermedi- ary expert system that first will always try the most com- monly satisfactory option, then ask for user’s reaction, and then utilize the next option if the first failed to pro- duce acceptable results. A much more promising approach suggests that online JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 43 searching behavior be investigated to reveal under which conditions each of the options is selected. We may find, for example, that when a term that is not central to a re- quest is mapped to a broader descriptor-and a user is primarily concerned with precision-searchers decide to combine a free-text search key with a descriptor. Trans- ferring this condition to automated systems will enable a system to decide when and how to interrogate users about request requirements that are relevant to the search pro- cess. More specifically, a system may proceed searching independently until it encounters a problematic term, such as one that is mapped to a broader descriptor, and then ask users for a specific kind of feedback. In sum- mary, it is not difficult to envision an intermediary expert system which would: (1) identify situations in which a re- quest statement is sufficient, and conversely those in which additional request criteria are needed for the search process to succeed, (2) list the relevant criteria so that users can provide the pertinent data, and (3) act upon data received to improve search results. Quite a powerful system! These few examples clearly demonstrate the benefits that could be gained from the study of searching behavior of human intermediaries, and from utilizing the experi- ence of human intermediaries in the design of intermedi- ary expert systems. The selection routine presented here is not sufficient to develop adaptive algorithms. Many issues, such as the nature of single-meaning and common terms or the con- ditions for the selection of a broader descriptor, need to be further investigated and rigorously defined. This rou- tine identifies problematic points in the search process and provides guidelines for research into searching be- havior that is relevant for automated systems. On the one hand, systems can interrogate users on request parame- ters-a subject users know best. On the other hand, they can select the most appropriate search keys-a decision casual end-users are not well enough informed to make. Based on searchers’ experience, intermediary expert sys- tems can become experts indeed. Acknowledgment The author wishes to thank Dagobert Soergel and Irene L. Travis for their helpful and insightful review of this article. References I. 2. 3. 4. 5. 6. 7. 8. Doszkocs, T. E. “Automatic vocabulary mapping in online search- ing.” International Classification. 10(2):78-83; 1983. Marcus, R. S. “An experimental comparison of the effectiveness of computers and humans as search intermediaries.” Journal of the American Societyfor Information Science. 34(6):381-404; 1983. Pollitt, A. S. “A ‘front-end’ system: An expert system as an online search intermediary.” Aslib Proceedings. 36(5):229-234; 1984. Brenner, E. H.; Lucey, J. H.; Martinez, C. L.; Meleka, A. 1984. “API’s machine aided indexing project.” Science Technology Li- braries. 5(1):49-62; 1984. Fidel, R. “Online searching styles: A case-study-based model of searching behavior.” Journal of the American Societyfor Informa- tion Science. 35(4):211-221; 1984. Fidel, R. “The case study method: A case study.” Library and In- formation Science Research. 6(3):273-288; 1984. Fugmann, R. “The complementarity of natural language and in- dexing languages.” International Classification. 9(3):140-144; 1982. King, D. W.; Bryant, E. C. The evaluation of information services and products. Washington, D.C.: Information Resources Press; 1971, p. 138. 44 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 Introduction Modeling the Selection of Search Keys FIG. 1. TABLE 1. The Selection Routine Discussion Acknowledgment References 
work_c5ljkxxrgja5rfhpbw3y2kkwnq	----	Improved multiclass feature selection via list combination Accepted Manuscript Improved multiclass feature selection via list combination Javier Izetta, Pablo F. Verdes, Pablo M. Granitto PII: S0957-4174(17)30467-0 DOI: 10.1016/j.eswa.2017.06.043 Reference: ESWA 11414 To appear in: Expert Systems With Applications Received date: 8 March 2017 Revised date: 30 June 2017 Accepted date: 30 June 2017 Please cite this article as: Javier Izetta, Pablo F. Verdes, Pablo M. Granitto, Improved multi- class feature selection via list combination, Expert Systems With Applications (2017), doi: 10.1016/j.eswa.2017.06.043 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. http://dx.doi.org/10.1016/j.eswa.2017.06.043 http://dx.doi.org/10.1016/j.eswa.2017.06.043 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Highlights • We introduce new SVM-RFE feature selection methods for multiclass problems • We use binary decomposition followed by strategies to combine lists of features • We discuss statistical approaches and voting theory methods • One-vs-One methods give better results than One-vs-All methods • The new K-First method is the more effective in selecting relevant features 1 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Improved multiclass feature selection via list combination Javier Izettaa, Pablo F. Verdesa, Pablo M. Granittoa,∗ aCIFASIS, French Argentine International Center for Information and Systems Sciences, UNR–CONICET, Bv. 27 de Febrero 210 Bis, 2000 Rosario, Argentina Abstract Feature selection is a crucial machine learning technique aimed at reducing the dimensionality of the input space. By discarding useless or redundant variables, not only it improves model performance but also facilitates its in- terpretability. The well-known Support Vector Machines–Recursive Feature Elimination (SVM-RFE) algorithm provides good performance with mod- erate computational efforts, in particular for wide datasets. When using SVM-RFE on a multiclass classification problem, the usual strategy is to decompose it into a series of binary ones, and to generate an importance statistics for each feature on each binary problem. These importances are then averaged over the set of binary problems to synthesize a single value for feature ranking. In some cases, however, this procedure can lead to poor selection. In this paper we discuss six new strategies, based on list combi- nation, designed to yield improved selections starting from the importances given by the binary problems. We evaluate them on artificial and real-world datasets, using both One–Vs–One (OVO) and One–Vs–All (OVA) strategies. Our results suggest that the OVO decomposition is most effective for feature selection on multiclass problems. We also find that in most situations the new K-First strategy can find better subsets of features than the traditional weight average approach. Keywords: ∗Corresponding author Email addresses: izetta@cifasis-conicet.gov.ar (Javier Izetta), verdes@cifasis-conicet.gov.ar (Pablo F. Verdes), granitto@cifasis-conicet.gov.ar (Pablo M. Granitto) Preprint submitted to Expert Systems with Applications July 5, 2017 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Feature Selection, Multiclass problems, Support Vector Machine 3 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 1. Introduction1 Many important problems in Machine Learning, as well as in in-silico2 Chemistry (Raies & Bajic, 2016), Biology, “high-throughput” technologies3 (Golub et al., 1999; Leek et al., 2010) or text processing (Forman, 2003;4 Uysal, 2016), share the property of involving much more features than mea-5 sured samples are available (Guyon & Elisseeff, 2003). The datasets associ-6 ated to these problems are, unsurprisingly, called “wide”. Usually, most of7 these variables carry a relatively low importance for the problem at hand.8 Furthermore, in some cases they interfere with the learning process instead9 of helping it, a scenario usually referred to as “curse of dimensionality”.10 Feature selection is an important pre–processing technique of Machine11 Learning aimed at coping with this curse (Kohavi & John, 1997). Its main12 goal is to find a small subset of the measured variables that improve, or at13 least do not degrade, the performance of the modeling method applied to the14 dataset. But feature selection methods do not only avoid the curse of dimen-15 sionality: they also allow for a considerable reduction in model complexity,16 an easier visualization and, in particular, a better interpretation of the data17 under analysis and the developed models (Liu et al., 2005).18 Several methods have been introduced in recent years, from general ones19 like Wrappers (Kohavi & John, 1997) and filters (Kira & Rendell, 1992) to20 very specific ones developed for SVM (Weston et al., 2000; Nguyen & De la21 Torre, 2010) and RVM (Mohsenzadeh et al., 2013, 2016) classifiers. Amongst22 other methods in the field (Hua et al., 2009), the well-known Recursive Fea-23 ture Elimination (RFE) algorithm provides good performance with moderate24 computational efforts (Guyon et al., 2002) on wide datasets. The original and25 most popular version of this method uses a linear Support Vector Machine26 (SVM) (Vapnik, 2013) to select the candidate features to be eliminated. Ac-27 cording to the SVM–RFE algorithm, the importance of an input variable28 i is directly correlated with the corresponding component (wi) of the vec-29 tor defining the separating hyperplane (w). The method is widely used in30 Bioinformatics (Guyon et al., 2002; Statnikov et al., 2005). Alternative RFE31 methods using other classifiers have also been introduced in the literature32 (Granitto et al., 2006; You et al., 2014).33 Typical feature selection algorithms are designed for binary classification34 problems, as the original version of RFE. Multiclass problems have received35 much less attention because of their increased difficulty. Also, because some36 classifiers involved in the selection process are designed to solve binary prob-37 4 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT lems. Most methods available for feature selection on multiclass problems38 are simple extensions of base methods. For example, RFE can be associated39 to a multiclass classifier like Random Forest (Breiman, 2001; Granitto et al.,40 2006).41 Although SVM was originally developed to deal only with binary prob-42 lems, it was extended to directly solve multiclass problems in different man-43 ners (Weston & Watkins, 1999; Crammer & Singer, 2001; Hsu & Lin, 2002),44 but with a modest success attributed mainly to the increased complexity of45 the solutions. On the other hand, in the last years several methods were46 developed to solve a multiclass problem using an appropriate combination of47 binary classifiers (Allwein et al., 2000; Hsu & Lin, 2002). The most usually48 followed strategy for multiclass SVM is known as “One–vs–One” (OVO). Ac-49 cording to this approach, a classification problem with c classes is replaced50 with M = c(c−1)/2 reduced binary ones, each one of them consisting of dis-51 criminating a pair of classes. In order to classify a new example, it is passed52 through all binary classifiers and the most voted class is selected. Another53 useful strategy is “One–vs–All” (OVA). In this second case, a problem with54 c classes is replaced with M = c reduced binary problems, each one of them55 consisting of discriminating a single class from all remaining ones.56 Therefore, the most usual approach to implement a multiclass SVM–RFE57 method is to directly apply the RFE algorithm over an OVO or OVA multi-58 class SVM (Ramaswamy et al., 2001; Duan et al., 2007; Zhou & Tuck, 2007).59 The pioneering work of Ramaswany et al. (Ramaswamy et al., 2001) pro-60 posed the OVA solution, but also compared results with the OVO strategy.61 Duan et al. (Duan et al., 2007) and Zhou et al. (Zhou & Tuck, 2007) devel-62 oped slight variations of the method, always considering both OVA and OVO63 implementations. Zhou et al. (Zhou & Tuck, 2007) also considered solutions64 to the RFE problem using a direct multiclass implementation.65 Interestingly, the solutions to the multiclass SVM–RFE problem that we66 have just described involve an important decision about the feature selection67 process which is usually neglected: they rank features by simply averag-68 ing components over the binary problems. For an input variable i they use69 < |wij| >j, the mean importance over all binary problems j, as the corre-70 sponding importance. As we discuss in the next section, this strategy can71 lead to sub–optimal selections in many cases. Once the original multiclass72 problem has been divided into multiple binary ones, the feature selection73 problem can be treated in a similar way. Then, a possible solution is to cast74 the multiclass feature selection problem as the problem of selecting candidate75 5 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT features from multiple lists (Jurman et al., 2008), each list corresponding to76 a different binary sub-problem.77 Similar solutions have been studied in related fields. In Bioinformatics,78 for example, Haury et al. (Haury et al., 2011) discussed the combination of79 multiple lists of genes from bootstraps of the same gene-expression dataset.80 Zhou and Dickerson (Zhou & Dickerson, 2014) and Zhou and Wang (Zhou &81 Wang, 2016) proposed the use of class–dependent features (different features82 for each binary problem) for biomarker discovery. Dittman et al. (Dittman83 et al., 2013) showed that combining multiple lists in binary classification84 problems can improve the feature selection results. In a short work in text85 categorization, Neumayer et al. (Neumayer et al., 2011) suggested that the86 combination of rankings generated by diverse methods can improve the re-87 sults of using a single method. Kanth and Saraswathi (Kanth & Saraswathi,88 2015) used class–dependent features for speech emotion recognition, but us-89 ing independent features for each class, not a final unique list.90 In this work we discuss in depth the use of combination of multiple lists in91 feature selection for multiclass classification problems. We first introduce a92 simple mathematical framework for multiple lists. Using this framework, we93 propose diverse strategies to produce improved selection of feature subsets94 with SVM-RFE. Also, we use some specifically–designed artificial datasets95 and real–world examples to evaluate them extensively, using both the OVO96 and OVA strategies.97 The rest of this article is organized as follows: in Section 2, we describe the98 feature selection methods introduced in this work. In Section 3 we evaluate99 these methods on diverse datasets and experimental setups. Finally, we draw100 our conclusions in Section 4.101 2. List combination methods for SVM–RFE102 The RFE selection method is a recursive process that ranks variables103 according to a given importance measure. At each iteration of the algo-104 rithm, the importance of each feature is calculated and the less relevant one105 is removed —in order to speed up the process, not one but a group of low106 relevance features is usually removed. Recursion is needed because the rel-107 ative importance of each feature can change substantially when evaluated108 over a different subset of features during the stepwise elimination process, in109 particular for highly correlated features. The inverse order in which features110 6 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT are eliminated is used to create a final ranking. Then, the feature selection111 process itself is reduced to take the first n features from this ranking.112 In the original binary version of SVM–RFE (Guyon et al., 2002), the113 projection of w (the normal vector to SVM’s decision hyperplane) in the114 direction of feature i, wi, is used as the importance measure. The method115 was efficiently extended to multiclass problems, employing the well-known116 OVO or OVA strategies to decompose the multiclass problem into a series117 of related binary ones (Ramaswamy et al., 2001; Duan et al., 2007; Zhou &118 Tuck, 2007). In both cases a set of M related binary problems is generated,119 each one solved by a vector wj. For each binary problem j, the importance120 of feature i is given by the corresponding component, wij.121 In order to obtain a unique importance for each feature in this setup, the122 simplest solution is to average the absolute value of the components |wij|123 over all related binary problems. We will call this method “Average” in the124 following. The Average solution is implemented, to the best of our knowledge,125 in all available RFE software packages, including the most popular amongst126 researchers (MATLAB, R and PYTHON platforms).127 However, the only real advantage of the Average strategy is its simplicity.128 Two main drawbacks of this approach should be taken into consideration but129 are usually ignored:130 1. The first issue can be called the flattening problem. Consider, for131 example, a feature e which is able to separate class j from all remaining132 classes, but is uninformative in other cases. Component wej will be133 large, but components wek with k 6= j will be small, giving a low value134 for < wej >j. Consider now another feature d which can give a modest135 help in separating any class from the others, obtaining always moderate136 values of wdj, and therefore giving a medium value for < wdj >j. The137 Average strategy will clearly rank the latter over the former, but in138 most scenarios it will be desirable to keep the first variable over the139 second.140 2. The second issue with the Average solution refers to relative scales.141 The length of vector wj is different for each binary problem, as it de-142 pends on the margin of the solution, which can change considerably for143 classes that are relatively close or far away in feature space. Averaging144 components of vectors of different lengths can lead to the selection of145 sub–optimal subsets.146 New strategies for feature selection able to overcome these drawbacks are 7 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Ranking List 1 List 2 . . . List M 1 f2 f3 . . . f1 2 f1 f7 . . . f3 3 f5 f2 . . . f6 ... ... ... . . . ... p f8 f4 . . . f7 Table 1: List of ranked features for each binary problem. needed. Here we propose to cast the problem as a selection of candidate features from multiple ranking lists (Jurman et al., 2008). We start by de- composing the multiclass problem into a set of M related binary problems (through the OVA or OVO strategies). The problem involves a set of p fea- tures, F = {f1, f2, . . . fp}. SVM–RFE produces a ranking (an ordered list) for each individual problem using the components wij. An example is shown in Table 1. This set of lists can be arranged in a matrix (Table 2) where each row shows the position of each feature in the ranking produced for the bi- nary problem shown on each column. We can now define a matrix of relative ranking positions as: ri,j = p−posi,j p = 1 − posi,j p , where ri,j is the relative ranking of feature fi in the list corresponding to bi-147 nary problem j, posi,j is the position of the same feature in the corresponding148 ranking (Table 2), and p is the total number of features in the problem. No-149 tice that the values of ri,j belong to the unit interval [0, 1] and depend linearly150 on the ranking position (a value of 1 − 1/p must be interpreted as the first151 position in the ranking).152 Important features should reflect in high values of ri,j for some i, j, mean-153 ing they are relevant to at least some of the binary classification problems.154 Two main strategies can be used to select those relevant features from this155 matrix and are discussed in the following.156 2.1. Methods based on relative ranking statistics157 The first strategy consists of measuring an appropriate statistic for each158 feature over all binary problems, and then using it to elaborate a final ranking159 of features. We selected the following four methods:160 8 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Features List 1 List 2 . . . List M f1 2 5 . . . 1 f2 1 4 . . . 7 f3 4 1 . . . 2 ... ... posi,j . . . ... fp posp,1 . . . . . . posp,M Table 2: Matrix showing the position of each feature in the ranking of each binary problem. Rows correspond to features and columns to binary problems. 2.1.1. Average-SD161 In this method, feature ranking is given by the average value of the relative162 position over all binary problems:163 Ri =< ri,j >j, where Ri is the ranking of feature fi in the final ordered list, used to select164 features in the multiclass problem. Ties are broken by the standard deviation165 (SD) of the relative position (higher is better). We show in the next section166 that features with higher SD are preferable over lower SD ones, because a167 larger SD means that the feature has some better–than–average rankings.168 Average-SD can be considered as the base strategy for multiple lists. It169 can overcome the relative scales problem on averaging weights, but is not170 expected to solve the flattening problem.171 2.1.2. Best Ranking172 In this second approach we rank every feature according to the best rel-173 ative ranking that it reaches over the set of binary problems:174 Ri = max(ri,j)j. Ties are broken by the mean value of the relative position over all prob-175 lems. A similar method has been used to select the winning class in multiple176 classifier systems (Ho et al., 1994). This strategy can be viewed as an ex-177 treme case, considering for each feature just one of the multiple rankings it178 receives and disregarding the rest. On the other hand, it is most aggressive179 in dealing with the flattening problem.180 9 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 2.1.3. 3Q-SD181 The third method orders features according to the 3rd quartile of the182 distribution of relative rankings:183 Ri = 3Q(ri,j)j, where the 3Q function returns the 3rd quartile of its argument. As in184 Average-SD, ties are broken by the SD. This approach is intermediate be-185 tween the two previous ones, searching for features that reach a high relative186 position, but also considering the full relative rankings distribution.187 2.1.4. K-First188 This method is adapted from a strategy to select relevant documents in189 information retrieval (Nuray & Can, 2006). The idea is to only consider190 features located in the top k positions of each individual list. We re-scale the191 relative rankings with a linear mapping reaching 0 for the k + 1 feature, and192 then take the average of this new relative importance:193 r′i,j = max(1 − posi,j k , 0) Ri =< r ′ i,j >j, where r′i,j is the re-scaled relative weight for feature fi and k is the number194 of features to be considered from each list (k < p). As in the Best Ranking195 method, ties are broken by the mean value of the original relative ranking,196 < ri,j >j. We discuss the set of parameter values k in the next section.197 This strategy is aimed at searching for features which are highly relevant for198 some of the problems, but is not limited to searching for the most relevant199 features —as the Best Ranking method is. It can potentially overcome both200 drawbacks of Average: relative scaling and flattening.201 2.2. Methods based on voting theory202 The second general strategy is related to voting theory (Saari, 2001;203 Young, 1988). In this setup we consider each binary problem as a voter,204 producing a ranking over a set of p candidates. Multiple methods were de-205 veloped over the years to solve the problem of combining elector preferences206 to find winner candidates —the most useful of them are known collectively207 as ”Condorcet Methods”. We focused on two popular procedures as selection208 methods for relevant features over multiple lists:209 10 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 2.2.1. Condorcet210 The most basic Condorcet method is known as Copeland’s method, or211 simply as Condorcet method (we will use the latter name in this work). It212 confronts each pair of features on every list (all binary problems), and then213 counts the number of wins minus the number of defeats for each feature214 (Young, 1988). A feature wins over another if it is ranked higher in the215 considered list. The global difference between wins and loses is used to216 rank features in the multiclass problem. Ties are broken by average relative217 rankings.218 2.2.2. Schulze219 This method, introduced by Schulze informally in 1997 and published220 later (Schulze, 2011), represents an improvement over previous Condorcet221 methods. It begins by counting wins and loses over each pair of features and222 all lists, storing these numbers in a pairwise preference matrix. Then a graph223 is constructed, with features as nodes and values in the matrix as weights.224 Finally, using a variant of the FloydWarshall algorithm, the strongest paths225 over the graph are selected for each pair of features, and their strengths226 are used to compare features. The strength of a path is defined as that227 of its weakest link (i.e., lower value in the matrix of preferences). A path228 between two nodes is valid if there is a sequence of strictly decreasing weights229 connecting them (Schulze, 2011). Features with more wins upon strength230 comparison are ranked first. The method is expected to perform better than231 basic Condorcet, but the computational load involved is significant.232 3. Evaluation on artificial datasets233 We first consider artificial classification problems in order to evaluate spe-234 cific aspects of the new methods and to be able to compare their capabilities235 in a controlled manner.236 3.1. Experimental setup237 As in previous works (Granitto et al., 2006), we strive to use an appro-238 priate computational setup for feature selection. We perform n = 100 times239 a random split (75%−25%) of each dataset in training and testing sets (the240 former are used to select features and train the classifiers, while the latter241 for model accuracy estimation). The testing sets are completely external to242 11 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT the feature selection process, thus providing unbiased estimates of classifica-243 tion errors for different number of features. The results of the n replicated244 experiments are then aggregated to yield mean error rate estimations and245 their corresponding SD.246 SVM–RFE was implemented using the OVO strategy unless specified247 otherwise. In both cases, we created the corresponding binary problems and248 produced a ranking of features for each of them. To create a ranking we used249 the standard SVM–RFE (linear kernel), as described by their authors (Guyon250 et al., 2002), eliminating 10% of the features at each iteration until there were251 less than 20 features left, when we slowed the procedure to eliminate 1 feature252 at each iteration. The fixed set of lists of ranked features were combined253 using the methods described before, producing a final list of features for each254 method under evaluation. Finally, for each method we fitted a multiclass255 SVM for a varying number of features, from 2 to p, using only the training256 data, and measured the classification error using the testing set. The C257 parameter was estimated in all cases using 5-fold cross validation of the258 training set.259 3.2. Artificial datasets260 We created three different multiclass datasets that provide diverse chal-261 lenges to our methods. In all cases, each class is sampled from a Gaussian dis-262 tribution with diagonal covariance matrix. For each dataset we can identify,263 by construction, a group of relevant features that can discriminate amongst264 classes and another group of irrelevant features containing Gaussian noise.265 All noisy features have the same mean (0) and SD (1) for all classes. Each266 dataset is composed of 3000 points evenly distributed among classes. The267 number of noisy features is fixed at 500.268 In the first dataset, called Artificial-1, there is a group of 5 features that269 is relevant for each class, i.e., class-specific features. The set of 5 features270 together shift the class center away from other classes. All relevant features271 have the same importance for the problem. The SD of the Gaussian distri-272 butions corresponding to relevant features are always set to 0.5.273 A different situation arises when there are sets of features which are rel-274 evant to some of the classes (more than one) but not for all of them. We275 created a second classification problem, Artificial-2, to evaluate this chal-276 lenge. The dataset has 8 classes and 25 relevant features, all sampled from277 Gaussian distributions with a SD of 0.5. The first 5 features are relevant for278 the first 3 classes of the problem. The following 5 are relevant for classes 4279 12 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Dataset Classes Relevant features Noisy features Artificial-1-3C 3 15 500 Artificial-1-4C 4 20 500 Artificial-1-5C 5 25 500 Artificial-1-8C 8 40 500 Artificial-1-16C 16 80 500 Artificial-2-8C 8 25 500 Artificial-3-3C 3 15 500 Artificial-3-8C 8 40 500 Artificial-3-16C 16 80 500 Table 3: Details of the artificial datasets used in this work. and 5 only, and are less relevant than the first 5. The rest of the features are280 relevant for a single class, 5 for each of the remaining 3 classes. These last281 features are less relevant than the first 10 features.282 Finally, we created a third problem, called Artificial-3, where all relevant283 features are equally useful for all classes at the same time. As in Artificial-1,284 there are 5 features for each class, all sampled from Gaussian distributions285 with a SD of 0.5.286 In all problems there is an overlap among classes, giving a nonzero Bayes287 error. We created five datasets for Artificial-1, with an increasing number288 of classes, and 3 datasets for Artificial-3 in the same way. Table 3 collects289 technical details of the datasets.290 291 3.3. Methodological setup292 3.3.1. K-First293 The K-First method is the only approach involving a parameter that294 needs to be set, k. The value of this parameter regulates the number of295 variables that receive a relative ranking. A very low value would make the296 method similar to Best Ranking, while a high one would turn the method297 into 3Q-SD (furthermore, k = p would convert the method into Average-SD).298 We evaluated several values of k (increasing fractions of p) over all ar-299 tificial datasets considered. Figure 1 shows the corresponding error curves300 as a function of the number of features selected by the method, for some301 representative problems. Error curves for all other artificial problems are302 13 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT similar to the reported ones (we include more figures in the Additional Ma-303 terial section). The vertical dotted lines show the correct number of relevant304 features for the problem, i.e. where the minimum of the curve should ideally305 be located. When possible, we show with a gray horizontal line the chance306 error level for the corresponding problem. The top row shows typical results307 for the Artificial-3 problem. In this case there are no class-specific features308 and, as a consequence, the results are almost independent of k. The bottom309 row shows results for class-dependent problems, Artificial-1 and 2. In this310 case the results clearly depend on k. We found that a value of 10% of p gives311 consistently good results in all artificial cases considered here, therefore we312 will use this value for the rest of the paper.313 3.3.2. Average-SD and 3Q-SD314 As noted before, these methods use the SD of the relative rankings as a315 breaking tie criterion, considering larger values of SD as better than smaller316 ones. This is based on the assumption that a large SD is associated with317 high rankings for some of the binary problems, and that such behavior is318 able to highlight class-dependent features over flat ones. In order to confirm319 this, we compared for both methods over a set of artificial problems the320 use of maximum versus minimum SD to break ties. Figure 2 shows the321 corresponding results for some representative cases. They are similar in all322 other cases (some of which are shown in the Additional Material section).323 As this figure shows, using maximum values always leads to equal or better324 performance than using minimum ones.325 3.3.3. OVA–SVM vs. OVO–SVM326 We applied both OVA and OVO strategies, combined with our feature327 selection methods, to all artificial datasets. We compared all results and328 found that the OVO strategy yields equal or superior performance in all329 cases. In Figure 3 we show some representative examples of this comparison,330 using the Artificial-1-8C and Artificial-2-8C datasets. On the left column331 we show OVA results, while the OVO case is depicted on the right column.332 We use the same scale for the corresponding panels. We also included the333 Bayes error for both datasets as dotted horizontal lines, and the true number334 of relevant features as a dotted vertical lines. More datasets are included in335 the Additional Material section.336 It is interesting to note that the two methods more directly aimed at find-337 ing class-relevant features (K-First and Best Ranking) are the ones showing338 14 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .6 3 0 0 .6 4 0 0 .6 5 0 0 .6 6 0 Features E rr o r ra te ● K First 2,5% K First 5% K First 10% K First 20% K First 40% ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .6 5 0 .7 0 0 .7 5 0 .8 0 0 .8 5 0 .9 0 Features E rr o r ra te ● K First 2,5% K First 5% K First 10% K First 20% K First 40% (a) (b) ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .7 5 0 .8 0 0 .8 5 0 .9 0 0 .9 5 Features E rr o r ra te ● K First 2,5% K First 5% K First 10% K First 20% K First 40% ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .5 0 .6 0 .7 0 .8 Features E rr o r ra te ● K First 2,5% K First 5% K First 10% K First 20% K First 40% (c) (d) Figure 1: Evaluation of different values of k for the K-First method. Each line shows average error rates as a function of the number of features selected by the corresponding method, with 1 SD error bars. (a) Artificial-3-3C (b) Artificial-3-8C (c) Artificial-1-16C (d) Artificial-2-8C (chance error 0.875). the bigger gains under the OVO strategy. Probably the OVO strategy can339 filter some of the noisy features more efficiently than OVA, as it considers340 significantly more lists of features (M = c(c − 1)/2 vs. M = c). After this341 comparison we will only use OVO-SVM to evaluate our new methods.342 3.4. Evaluation of the Methods on Artificial Datasets343 Figure 4 shows the results for 4 versions of the Artificial-1 problem and 2344 of the Artificial-3 problem. The remaining dataset from Artificial-1 is shown345 on Panel (c) of Figure 3. Results for Artificial-2 are shown on Panel (d) of346 the same figure. Additional datasets are included in the Additional Material347 section. Overall, artificial problems show that the K-First method is the348 15 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .5 5 0 .6 0 0 .6 5 0 .7 0 0 .7 5 0 .8 0 0 .8 5 Features E rr o r ra te ● Average−SD Max Average−SD Min 3Q−SD Max 3Q−SD Min ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .5 5 0 .6 0 0 .6 5 0 .7 0 0 .7 5 0 .8 0 0 .8 5 Features E rr o r ra te ● Average−SD Max Average−SD Min 3Q−SD Max 3Q−SD Min (a) (b) Figure 2: Comparison of tie breaking using maximum or minimum SD for Average-SD and 3Q-SD. Details are similar to Figure 1. Chance error of 0.875 for both panels. (a) Artificial-1-8C. (b) Artificial-2-8C. most efficient one in finding subsets of features with low classification error,349 followed closely by the Best-Ranking method. The Schulze method shows350 good performance in several datasets. The other 3 methods show similar351 results, though not as good as the first group.352 On the Artificial-3 datasets (all relevant features are useful for all classes)353 the differences among methods are clearly smaller than on the other 2 prob-354 lems (with class-dependent features). Differences in performance increase355 with the number of classes for the Artificial-1 dataset.356 Comparing the two methods based on voting theory, the low performance357 of Condorcet compared with Schulze is notorious. Taking a closer look at the358 method, we noticed that Condorcet produces a lot of ties in the rankings,359 which are broken using average positions. This produces a bias towards360 features with good global average values instead of features highly relevant361 for a few lists.362 Another interesting analysis that can be made with artificial datasets is363 the position occupied by the truly relevant features on the rankings produced364 by the different methods, as we know in advance which features are noisy365 and which ones are informative. A perfect method should rank all relevant366 features first, with all noisy features following.367 For each artificial problem, we analyzed the distribution of rankings given368 by each selection method to the set of relevant features and to the set of369 noisy features. We then computed some descriptive statistics of those two370 16 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 1 .0 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 1 .0 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze (a) (c) ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze (b) (d) Figure 3: OVA-SVM vs. OVO-SVM on two artificial problems. Details are similar to Figure 1. (a) RFE-OVA-SVM on Artificial-1-8C. (b) RFE-OVA-SVM on Artificial-2-8C. (c) RFE-OVO-SVM on Artificial-1-8C (d) RFE-OVO-SVM on Artificial-2-8C. distributions (Best, 1st. quartile, Mean, 3rd. quartile and Worst). In Table371 4 we show these statistics on dataset Artificial-1-8C, which is representative372 of the results obtained on the other versions of this problem. All six methods373 rank relevant features at the first positions and noisy features at the last ones,374 but there are important differences. Looking at the Mean and 3rd. Quartile375 of the distributions, it is clear that K-First, Best Ranking and Schulze, in that376 order, are the most accurate ones in ranking most of the features according377 to their global relevance. These results confirm that the low error rates on378 the figures discussed before are directly related to a better feature selection379 by those methods.380 In Table 5 we show the corresponding statistics for the Artificial-2-8C381 17 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Relevant 3Q-SD Av-SD Best Rank. K-First Condorcet Schulze Best 1 1 1 1 1 1 1st Q. 14 14 12 11 17 13 Mean 70 65 33 22 79 48 3rd Q. 98 86 41 31 110 61 Worst 436 476 357 495 474 410 Noisy 3Q-SD Av-SD Best Rank. K-First Condorcet Schulze Best 1 1 1 3 1 1 1st Q. 160 161 165 165 159 163 Mean 287 287 290 290 286 288 3rd Q. 415 415 415 415 415 415 Worst 540 540 540 540 540 540 Table 4: Statistics of the rankings given to relevant and noisy features by the diverse methods considered in this study for the Artificial-1-8C dataset. Values are rounded when needed. dataset. This is the most interesting problem, as it contains subsets of rele-382 vant features with diverse levels of relevance. In the table we separated the383 relevant features into 3 subsets. As it can be observed in the table, K-First384 is the most effective strategy in separating the subsets of relevant features,385 and not only relevant from noisy features.386 Finally, in Table 6 we show statistics for the Artificial-3-8C dataset —387 it is representative of all versions of this problem. As discussed before, all388 methods are almost equivalent on this dataset.389 A last comment is in order about the Best Ranking method. As it can390 be seen in the tables, it can give high rankings to noisy features more easily391 than the K-First method, as it bases the final ranking on a single value for392 each feature.393 4. Evaluation on real–world datasets394 We used 14 real–world datasets to evaluate our new methods. Details and395 origins of the datasets are collected on Table 7. We selected datasets from396 three different domains. The first 4 datasets collect mass spectrometry mea-397 surements of food products. All recorded peaks are present in the datasets.398 Some of the products under analysis can present class-specific features, re-399 flecting particularities of some products, such as origin or manufacturing400 18 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 1 to 5 3Q-SD Av-SD Best Rank. K First Condorcet Schulze Best 1 1 1 1 1 1 1st Q. 2 2 2 2 2 2 Mean 22 5 9 3 15 4 3rd Q. 28 5 17 4 18 5 Worst 481 62 45 7 85 18 6 to 10 3Q-SD Av-SD Best Rank. K First Condorcet Schulze Best 1 1 1 1 1 1 1st Q. 17 7 6 7 13 6 Mean 80 24 13 8 50 9 3rd Q. 158 30 19 9 58 11 Worst 523 241 44 14 453 70 11 to 25 3Q-SD Av-SD Best Rank. K First Condorcet Schulze Best 1 2 1 8 1 4 1st Q. 35 19 12 14 29 15 Mean 161 87 22 18 101 65 3rd Q. 265 116 30 22 185 70 Worst 524 500 57 29 524 501 Noisy 3Q-SD Av-SD Best Rank. K First Condorcet Schulze Best 1 3 3 10 1 8 1st Q. 142 147 151 151 142 148 Mean 269 273 275 275 274 274 3rd Q. 399 400 400 400 399 400 Worst 525 525 525 525 525 525 Table 5: Statistics of the rankings given to relevant and noisy features by the diverse methods considered in this study for the Artificial-2-8C dataset. The relevant features are divided into three subsets, and ordered according to their relevance by construction. Values are rounded when needed. 19 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Relevant 3Q-SD Av-SD Best Rank. K First Condorcet Schulze Best 1 1 1 1 1 1 1st Q. 11 11 11 11 11 11 Mean 27 29 41 26 28 29 3rd Q. 33 37 53 33 33 37 Worst 440 431 368 250 342 252 Noisy 3Q-SD Av-SD Best Rank. K First Condorcet Schulze Best 5 2 2 8 3 7 1st Q. 165 165 164 165 165 165 Mean 290 290 290 290 290 290 3rd Q. 415 415 415 415 415 415 Worst 540 540 540 540 540 540 Table 6: Statistics of the rankings given to relevant and noisy features by the diverse methods considered in this study for the Artificial-3-8C dataset. Values are rounded when needed. method. The following 6 datasets come from the UCI repository. These401 are more traditional datasets, with more samples than features and multiple402 classes, involving typical pattern recognition problems. Finally, we selected403 4 gene expression datasets from human tissues. These datasets were filtered404 by curators to obtain circa 1000 genes with high signal–to–noise ratio in each405 case.406 In order to compare our results against previous methods we implemented407 3 versions of MSVM-RFE, as described by Zhou & Tuck (2007). The first408 method uses the multiclass SVM developed by Crammer & Singer (2001).409 We will denote it ”Zhou C & S”. The second one uses the method of Weston410 & Watkins (1999), in the following denoted by ”Zhou W & W”. Finally, we411 implemented MSVM-RFE with the OVO decomposition of the traditional412 binary SVM. We will refer to this method as ”Zhou OVO”. Notice that it413 is equivalent to the Average Weights methodology, which is implemented by414 default in most available Machine Learning packages, as previously explained.415 On Figures 5 and 6 we show the results for eight datasets, while the416 remaining cases are shown in the Additional Material section. In general,417 differences in results for real world data are less notorious than for the Ar-418 tificial problems. For UCI and Mass-Spectrometry datasets, K-First is in419 general the method showing the best results in finding small subsets with420 reduced classification error, followed by Best Ranking and 3rd Quartile. In421 20 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT some problems, like Apples and Libra, differences are more notorious. On the422 gene expression datasets all methods show small differences, but in general423 the variants of MSVM-RFE exhibit better results than on the other domains.424 These datasets have been filtered by curators and as a consequence all fea-425 tures are informative. We believe that this improves the performance of426 averaging methods over methods that search for class-specific features.427 Error curves show the complete behavior of the methods as a function428 of the number of features, but occasionally diverse methods are more effi-429 cient in selecting a high number or just a few features. To produce a more430 concrete comparison, we measured for two fixed numbers of selected features431 (10 and 20) the proportion of runs on which method A shows a smaller error432 than method B, for each of the three domains under evaluation. The full433 resulting matrices are shown in the Additional Material section. From these434 matrices we computed a ranking for each method, counting the number of435 other methods that it excels. We show the corresponding results in Table 8.436 They confirm the information extracted from the error curves: on the UCI437 and Mass-Spectrometry domains K-first shows the best results, but Best438 Ranking and 3rd Quartile also have high rankings. On the gene expression439 domains the best results come from one of the MSVM-RFE methods.440 5. Computational burden441 We evaluated the burden of the 6 new methods as a function of the442 number of features and samples using the Artificial-1 dataset. In panel (a)443 of Figure 7 we show how the running time scales with the number of features444 in the problems, using a log-scale for times. We include the 3 versions of the445 method by Zhou et al. as a comparison. It is clear from this figure that all but446 the Schulze method scale almost linearly with the number of features, being447 Condorcet and 3rd Quartile the slowest methods. Schulze is cubic in the448 number of features, as it involves a variant of the FloydWarshall algorithm449 to find shortest paths in a graph. On panel (b) of the same figure we show450 the dependence on the number of samples in the dataset. All new methods,451 including Schulze, scale almost linearly with the number of samples. The452 two variants of Zhou’s method using direct multiclass SVMs show power–453 law scaling with the number of samples.454 21 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 6. Conclusions455 In this work we discussed in depth the use of combinations of lists of456 features (instead of averaging individual importances) in SVM–RFE for fea-457 ture selection on multiclass problems. Using an appropriate mathematical458 framework we introduced 6 different methods to produce the final ranking of459 features starting from a set of ranked features list produced by each binary460 problem. We evaluated them in a series of artificial and real world datasets.461 Our first conclusion is that the OVO strategy should be preferred over462 OVA for multiclass feature selection. Probably the higher number of binary463 problems in OVO helps in filtering out some noisy features that receive high464 rankings from just one or only a few binary problems, a similar beneficial465 effect to the use of ensembles in general.466 Our second conclusion is that, overall, the K-First method is the most467 consistent one in selecting subsets of relevant features that lead to smaller468 classification errors. The idea is well–known in the document retrieval liter-469 ature, only considers the top k values of each list, and adapts efficiently to470 feature selection. We showed with several artificial and real–world datasets471 that this new method is superior to the typical weights averaging that is472 implemented by default in all current Machine Learning libraries.473 Finally, two other methods also showed good results but present some474 drawbacks. The Best Ranking strategy is simple and efficient, but can lose475 performance on some problems, such as Artificial-2. Also, the use of a single476 value to characterize the behavior of a feature can give high rankings to noisy477 features by chance. The Schulze strategy, based on voting theory, shows a478 very good performance on some artificial datasets but does not compare well479 on real–world ones, and is by far the most complex and time–consuming480 strategy out of the six methods under evaluation.481 Overall, the new methods were designed for problems with class-specific482 features, which is where they show their best performance. As they employ483 the OVO strategy, they are also resistant to noisy features. Filtered domains,484 with lots of low-relevance features and little noise like our gene expression485 datasets, seem to represent a more challenging domain for our new methods.486 Work in progress includes a more extensive evaluation and the use of a487 penalty term to help discard correlated features.488 22 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT References489 Allwein, E. L., Schapire, R. E., & Singer, Y. (2000). Reducing multiclass490 to binary: A unifying approach for margin classifiers. Journal of machine491 learning research, 1 , 113–141.492 Breiman, L. (2001). Random forests. Machine learning , 45 , 5–32.493 Cappellin, L., Soukoulis, C., Aprea, E., Granitto, P., Dallabetta, N., Costa,494 F., Viola, R., Märk, T. D., Gasperi, F., & Biasioli, F. (2012). Ptr-495 tof-ms and data mining methods: a new tool for fruit metabolomics.496 Metabolomics , 8 , 761–770.497 Crammer, K., & Singer, Y. (2001). On the algorithmic implementation of498 multiclass kernel-based vector machines. Journal of machine learning re-499 search, 2 , 265–292.500 Del Pulgar, J. S., Soukoulis, C., Carrapiso, A., Cappellin, L., Granitto, P.,501 Aprea, E., Romano, A., Gasperi, F., & Biasioli, F. (2013). Effect of the502 pig rearing system on the final volatile profile of iberian dry-cured ham as503 detected by ptr-tof-ms. Meat science, 93 , 420–428.504 Dittman, D. J., Khoshgoftaar, T. M., Wald, R., & Napolitano, A. (2013).505 Classification performance of rank aggregation techniques for ensemble506 gene selection. In FLAIRS Conference.507 Duan, K.-B., Rajapakse, J. C., & Nguyen, M. N. (2007). One-versus-one and508 one-versus-all multiclass svm-rfe for gene selection in cancer classification.509 In European Conference on Evolutionary Computation, Machine Learning510 and Data Mining in Bioinformatics (pp. 47–56). Springer.511 Fabris, A., Biasioli, F., Granitto, P. M., Aprea, E., Cappellin, L., Schuhfried,512 E., Soukoulis, C., Märk, T. D., Gasperi, F., & Endrizzi, I. (2010). Ptr-tof-513 ms and data-mining methods for rapid characterisation of agro-industrial514 samples: influence of milk storage conditions on the volatile compounds515 profile of trentingrana cheese. Journal of mass spectrometry , 45 , 1065–516 1074.517 Forman, G. (2003). An extensive empirical study of feature selection metrics518 for text classification. Journal of machine learning research, 3 , 1289–1305.519 23 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Golub, T. R., Slonim, D. K., Tamayo, P., Huard, C., Gaasenbeek, M.,520 Mesirov, J. P., Coller, H., Loh, M. L., Downing, J. R., Caligiuri, M. A.521 et al. (1999). Molecular classification of cancer: class discovery and class522 prediction by gene expression monitoring. science, 286 , 531–537.523 Granitto, P. M., Furlanello, C., Biasioli, F., & Gasperi, F. (2006). Recursive524 feature elimination with random forest for ptr-ms analysis of agroindustrial525 products. Chemometrics and Intelligent Laboratory Systems , 83 , 83–90.526 Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature527 selection. Journal of Machine Learning Research, 3 , 1157–1182.528 Guyon, I., Weston, J., Barnhill, S., & Vapnik, V. (2002). Gene selection for529 cancer classification using support vector machines. Machine learning , 46 ,530 389–422.531 Haury, A.-C., Gestraud, P., & Vert, J.-P. (2011). The influence of feature532 selection methods on accuracy, stability and interpretability of molecular533 signatures. PloS one, 6 , e28210.534 Ho, T. K., Hull, J. J., & Srihari, S. N. (1994). Decision combination in535 multiple classifier systems. IEEE transactions on pattern analysis and536 machine intelligence, 16 , 66–75.537 Hsu, C.-W., & Lin, C.-J. (2002). A comparison of methods for multiclass538 support vector machines. IEEE transactions on Neural Networks , 13 ,539 415–425.540 Hua, J., Tembe, W. D., & Dougherty, E. R. (2009). Performance of feature-541 selection methods in the classification of high-dimension data. Pattern542 Recognition, 42 , 409–424.543 Jurman, G., Merler, S., Barla, A., Paoli, S., Galea, A., & Furlanello, C.544 (2008). Algebraic stability indicators for ranked lists in molecular profiling.545 Bioinformatics , 24 , 258–264.546 Kanth, N. R., & Saraswathi, S. (2015). Efficient speech emotion recognition547 using binary support vector machines & multiclass svm. In Computational548 Intelligence and Computing Research (ICCIC), 2015 IEEE International549 Conference on (pp. 1–6). IEEE.550 24 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Kira, K., & Rendell, L. A. (1992). The feature selection problem: Traditional551 methods and a new algorithm. In AAAI (pp. 129–134). volume 2.552 Kohavi, R., & John, G. H. (1997). Wrappers for feature subset selection.553 Artificial intelligence, 97 , 273–324.554 Leek, J. T., Scharpf, R. B., Bravo, H. C., Simcha, D., Langmead, B., Johnson,555 W. E., Geman, D., Baggerly, K., & Irizarry, R. A. (2010). Tackling the556 widespread and critical impact of batch effects in high-throughput data.557 Nature Reviews Genetics , 11 , 733–739.558 Lichman, M. (2013). UCI machine learning repository.559 Liu, H., Dougherty, E. R., Dy, J. G., Torkkola, K., Tuv, E., Peng, H., Ding,560 C., Long, F., Berens, M., Parsons, L. et al. (2005). Evolving feature selec-561 tion. IEEE Intelligent systems , 20 , 64–76.562 Mohsenzadeh, Y., Sheikhzadeh, H., & Nazari, S. (2016). Incremental rel-563 evance sample-feature machine: A fast marginal likelihood maximization564 approach for joint feature selection and classification. Pattern Recognition,565 60 , 835–848.566 Mohsenzadeh, Y., Sheikhzadeh, H., Reza, A. M., Bathaee, N., & Kalayeh,567 M. M. (2013). The relevance sample-feature machine: A sparse bayesian568 learning approach to joint feature-sample selection. IEEE transactions on569 cybernetics , 43 , 2241–2254.570 Monti, S., Tamayo, P., Mesirov, J. P., & Golub, T. R. (2003). Consensus clus-571 tering: A resampling-based method for class discovery and visualization of572 gene expression microarray data. Machine Learning , 52 , 91–118.573 Neumayer, R., Mayer, R., & Nørv̊ag, K. (2011). Combination of feature574 selection methods for text categorisation. In European Conference on In-575 formation Retrieval (pp. 763–766). Springer.576 Nguyen, M. H., & De la Torre, F. (2010). Optimal feature selection for577 support vector machines. Pattern recognition, 43 , 584–591.578 Nuray, R., & Can, F. (2006). Automatic ranking of information retrieval579 systems using data fusion. Information processing & management , 42 ,580 595–614.581 25 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Raies, A. B., & Bajic, V. B. (2016). In silico toxicology: computational582 methods for the prediction of chemical toxicity. Wiley Interdisciplinary583 Reviews: Computational Molecular Science, 6 , 147–172.584 Ramaswamy, S., Tamayo, P., Rifkin, R., Mukherjee, S., Yeang, C.-H., An-585 gelo, M., Ladd, C., Reich, M., Latulippe, E., Mesirov, J. P. et al. (2001).586 Multiclass cancer diagnosis using tumor gene expression signatures. Pro-587 ceedings of the National Academy of Sciences , 98 , 15149–15154.588 Saari, D. (2001). Chaotic elections!: A mathematician looks at voting . Amer-589 ican Mathematical Soc.590 Schulze, M. (2011). A new monotonic, clone-independent, reversal symmet-591 ric, and condorcet-consistent single-winner election method. Social Choice592 and Welfare, 36 , 267–303.593 Statnikov, A., Aliferis, C. F., Tsamardinos, I., Hardin, D., & Levy, S. (2005).594 A comprehensive evaluation of multicategory classification methods for595 microarray gene expression cancer diagnosis. Bioinformatics , 21 , 631–643.596 Uysal, A. K. (2016). An improved global feature selection scheme for text597 classification. Expert systems with Applications , 43 , 82–92.598 Vapnik, V. (2013). The nature of statistical learning theory . Springer science599 & business media.600 Weston, J., Mukherjee, S., Chapelle, O., Pontil, M., Poggio, T., & Vapnik, V.601 (2000). Feature selection for svms. In Proceedings of the 13th International602 Conference on Neural Information Processing Systems (pp. 647–653). MIT603 Press.604 Weston, J., & Watkins, C. (1999). Support vector machines for multi-class605 pattern recognition. In ESANN (pp. 219–224). volume 99.606 You, W., Yang, Z., & Ji, G. (2014). Feature selection for high-dimensional607 multi-category data using pls-based local recursive feature elimination. Ex-608 pert Systems with Applications , 41 , 1463–1475.609 Young, H. P. (1988). Condorcet’s theory of voting. The American Political610 Science Review , (pp. 1231–1244).611 26 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Zhou, N., & Wang, L. (2016). Processing bio-medical data with class-612 dependent feature selection. In Advances in Neural Networks (pp. 303–613 310). Springer.614 Zhou, W., & Dickerson, J. A. (2014). A novel class dependent feature se-615 lection method for cancer biomarker discovery. Computers in biology and616 medicine, 47 , 66–75.617 Zhou, X., & Tuck, D. P. (2007). Msvm-rfe: extensions of svm-rfe for mul-618 ticlass gene selection on dna microarray data. Bioinformatics , 23 , 1106–619 1114.620 27 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze (a) (b) ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .7 5 0 .8 0 0 .8 5 0 .9 0 0 .9 5 1 .0 0 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze (c) (d) ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .6 2 0 .6 4 0 .6 6 0 .6 8 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .6 5 0 .7 0 0 .7 5 0 .8 0 0 .8 5 0 .9 0 0 .9 5 Features E rr o r ra te ● Best ranking K first 3Q−SD Average−SD Condorcet Schulze (e) (f) Figure 4: Error curves for the six methods on some artificial datasets. Details are similar to Figure 1. (a) Artificial-1-3C (chance error 0.666) (b) Artificial-1-4C (chance error 0.75) (c) Artificial-1-5C (d) Artificial-1-16C (e) Artificial-3-3C (f) Artificial-3-8C 28 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Dataset F S C Description Apple 714 150 15 Mass spectrometry measurements over 15 varieties of Apple clones (Cappellin et al., 2012) Cheese 117 72 8 Mass spectrometry measurements over 8 va- rieties of Italian cheese (Fabris et al., 2010) Ham 1338 382 11 Mass spectrometry measurements over 11 varieties of Iberian Hams (Del Pulgar et al., 2013) Strawberry 232 233 9 Mass spectrometry measurements over 9 va- rieties of strawberries (Granitto et al., 2006) Multi-F 649 2000 10 Features of handwritten numerals extracted from a collection of Dutch utility maps (Lichman, 2013) Libras 90 360 15 Diverse hand movements from the Brazilian hands language (Lichman, 2013) Robot1 90 88 4 Robot Execution Failures Data Set, from UCI. Failures in approach to grasp position (Lichman, 2013) Robot3 90 47 4 Same as Robot1. Position of part after a transfer failure Robot4 90 117 3 Same as Robot1. Failures in approach to ungrasp position Robot5 90 164 5 Same as Robot1. Failures in motion with part Leukemia 985 248 6 Gene expression of Bone marrow samples with 6 subtypes of Leukemia (Monti et al., 2003) Lung 1000 197 4 Gene expression of lung tissues with 4 can- cer types (Monti et al., 2003) CNS 989 42 5 Gene expression of 5 tumor types of the cen- tral nervous system (Monti et al., 2003) Novartis 1000 103 4 Gene expression of tissue samples from 4 distinct cancer types (Monti et al., 2003) Table 7: Details on the 14 real–world datasets used in this work. Columns show the number of features (F), samples (S) and classes (C). 29 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 0 .6 0 0 .6 4 0 .6 8 0 .7 2 Features E rr o r ra te ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 0 .1 0 .2 0 .3 0 .4 0 .5 Features E rr o r ra te ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO (a) (b) ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 1000 0 .2 5 0 .3 5 0 .4 5 0 .5 5 Features E rr o r ra te ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 Features E rr o r ra te ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO (c) (d) Figure 5: Results on some real world datasets. Details are similar to Figure 1. (a) Apples. (b) Strawberry. (c) Ham. (d) Libras. Group 3Q-SD Av-SD Best K-First Cond. Schulze Z. C&S Z. W&W Z. OVO UCI-10 Feat 7 5 5 8 3 2 0 2 5 MS -10 Feat 0 3 7 8 2 1 5 4 6 GE -10 Feat 1 3 6 4 2 0 7 5 8 UCI-20 Feat 5 2 6 8 2 0 3 7 3 MS -20 Feat 0 2 7 8 3 1 4 5 6 GE -20 Feat 1 6 5 4 2 0 8 3 7 Table 8: Rankings of methods (higher is better) counting the number of times that one method outperforms another on each domain and number of selected features. 30 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 0 .3 0 0 .3 5 0 .4 0 Features E rr o r ra te ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 0 .3 0 0 .3 5 0 .4 0 0 .4 5 0 .5 0 Features E rr o r ra te ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO (a) (b) ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 1000 0 .0 0 0 .1 0 0 .2 0 0 .3 0 Features E rr o r ra te ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 5 10 20 50 100 200 500 1000 0 .1 0 .2 0 .3 0 .4 0 .5 Features E rr o r ra te ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO (c) (d) Figure 6: Results on some real world datasets. Details are similar to Figure 1. (a) Robot1. (b) Robot3. (c) Leukemia. (d) CNS. 31 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 2 5 1 0 2 0 5 0 2 0 0 5 0 0 Features T im e ● ● ● ● ● 100 200 300 400 500 ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO 2 5 1 0 2 0 5 0 Samples T im e ● ● ● ● 200 400 600 800 ● Best Ranking K first 3Q−SD Average−SD Condorcet Schulze Zhou C&S Zhou W&W Zhou OVO (a) (b) Figure 7: Comparison of running times for all methods evaluated in this work as a function of (a) the number of features and (b) the number of samples. Times (in seconds) are in log–scale. 32 
work_cdulhhfax5dwrkjqrb6ys52gae	----	G:/Data/Disk_C/Utenti/Francesco/Pac_F_Bianco/pubblicazioni/InProgress/ESWA_II/Manuscript.dvi Performance analysis of colour descriptors for parquet sorting Francesco Bianconia,∗, Antonio Fernándezb, Elena Gonzálezb, Stefano A. Saettaa aUniversità degli Studi di Perugia Dipartimento Ingegneria Industriale Via G. Duranti, 67 – 06125 Perugia (Italy) e-mail: bianco@ieee.org, stefano.saetta@unipg.it bUniversidade de Vigo Escuela de Ingenieŕıa Industrial Campus Universitario – 36310 Vigo (Spain) e-mail: {elena,antfdez}@uvigo.es Abstract In this paper we consider the problem of colour-based sorting hardwood parquet slabs into lots of similar visual appearance. As a basis for the development of an expert system to perform this task, we experimentally investigate and compare the performance of various colour descriptors (i.e.: soft descriptors, percentiles, marginal histograms and 3D histogram), and colour spaces (i.e.: RGB, HSV and CIE Lab). The results show that simple and compact colour descriptors, such as the mean of each colour channel, are as accurate as more complicated features. Likewise, we found no statistically significant difference in the accuracy attainable through the colour spaces considered in the paper. Our experiments also show that most methods are fast enough for real-time processing. The results suggest the use of simple statistical descriptors along with RGB data as the best practice to approach the problem. Keywords: Grading, Parquet, Sorting, Colour descriptors, Wood ∗Corresponding author. Preprint submitted to Expert Systems with Applications August 2, 2012 1. Introduction Wood is a widely used and greatly appreciated material. Countless are its applications in various industrial sectors including construction, interi- ors, furniture and shipbuilding. Like other products such as natural stone, ceramics, leather and similar, wood is mostly appreciated for its appearance; a feature that determines, to a great extent, its price. When wood is used for flooring, decking or façade cladding (in this case we usually refer to it as engineered wood), strict selection procedures are needed to assure satis- factory aesthetic results. To obtain beautiful and uniform surfaces, wood has to be carefully graded by fibre type and colour tone. In an increasingly globalised and competitive market, it is mandatory that wood products – particularly those of high range – be virtually extent of any defects. In an endeavour to meet such requirements and increase market shares, producers are trying to drastically improve their quality standards. In this context quality inspection plays a central role. As noted by Bombardier and Schmitt (2010), wood quality inspection involves two different and clearly separated problems: 1) detection, localiza- tion and classification of surface defects; and 2) sorting products into lots of similar appearance. In the parquet industry the two processes are usu- ally carried out sequentially and in this order. Both can be performed either manually or automatically. Technically speaking the first problem is referred to as grading and is related to detecting, measuring and counting superficial defects like knots, pockets, stains, veins, cracks, etc. Domain-specific stan- dards (DIN-EN-975-1, 2011; DIN-EN-975-2, 2004; DIN-1611, 2002) define different wood grades on the basis of the number and size of such defects along with procedures to their measurement and detection. As for the second problem, we can find it referred to as sorting (Lu et al., 1997), colour classification (Kurdthongmee, 2008), or, again, grading (Vienonen et al., 2002; Faria et al., 2008). To avoid confusion, throughout this paper we use the term grading to refer to the first problem and sorting for the second. When performed manually, grading is stressful and time consuming, though, in general, not particularly demanding, since defects are usually quite evident. In contrast, sorting products into groups of similar appear- ance is more subtle, since products of the same class may have differences in tone which can be very slight and difficult to detect even to a trained eye. In addition this process requires more than one slab to be observed at the same time. Subjective and environmental conditions, tiredness, boredom and other factors can significantly affect the outcome of the process. To this 2 Table 1: Summary list of methods for colour-based wood sorting. Reference Method Colour descriptor Colour space Arden (1991) Image-based Marginal histograms RGB Lu et al. (1997) Image-based 3D histogram RGB Vienonen et al. (2002) Spectrum-based Spectral histogram Spectrum Kurdthongmee (2008) Image-based One marginal histogram (hue) HSV Hrčka (2008) Image-based Mean + Standard deviation CIE Lab Faria et al. (2008) Image-based Approx. marginal histograms HSV, CIE Lab Schnabel et al. (2009) Spectrum-based Mean CIE Lab Bombardier and Schmitt (2010) Image-based Mean + Homogeneity CIE Lab + HSV Buchelt and Wagenführ (2012) Spectrum-based Mean CIE Lab we should add that recent studies showed how colour perception can be sig- nificantly influenced by age and socio-economical level of the subject (Kose, 2008). Our personal experience indeed confirms that different operators can produce very dissimilar results. Beginning with these considerations, it is therefore no surprise that the agreement between different operators can be as low as 60% (Roz̆man et al., 2006). As a consequence, manual inspection procedures can produce batches of products with significant variations of the visual appearance, causing sales returns and significant economical losses. Parquet producers are therefore more and more concerned with the de- velopment of computer vision systems capable of carrying out automatic quality control procedures. In this paper, in particular, we are concerned with the second problem, that of measuring and comparing the visual ap- pearance of parquet slabs in order to sort them according to suitable simi- larity criteria. More specifically we focus on the problem of colour sorting slabs of the same grade and quality. This consists of dividing a previously graded batch of parquet hardwood into different colour tones, among which differences in colour are usually very slight, yet noticeable. To this end we experimentally compare the effectiveness of a set of colour descriptors and spaces. We also discuss issues related to image acquisition and processing including computational time, which are of primary importance when it comes to designing and implementing practical, real-time solutions. The remainder of the paper begins with a brief survey of related research (Sec. 2). In the following sections we give a description of the materials (Sec. 3) and methods (Sec. 4) used in our research. In Sec. 5 we present the experimental activity followed by results (Sec. 6) and conclusions (Sec. 7). 3 2. Related research The first applications of computer vision to the wood industry date back to the 1980s (Conners et al., 1983; Sobey and Semple, 1989). Literature review shows that, since then, research has mostly focused on the prob- lem of grading. Studies have confirmed that it is possible to replace man- ual graders with automatic systems improving production effectiveness and product quality (Lycken, 2006; Kline et al., 2006). A preliminary step to automatic grading is defect detection and characterization. This has been typically approached using spectral features (Åström et al., 1999), spectral and X-ray features (Bond et al., 2002), colour features (Conners et al., 1992; Ciccotelli and Portala, 1992) and combinations of colour and texture features (Kyllönen and Pietikäinen, 2000; Gu et al., 2010). Other authors focused on how to deal with the grading problem once defects have been detected and located (Lycken, 2006; Castellani and Rowlands, 2009). For an up-to-date survey on automatic wood grading readers are referred to the work of Jabo (2011). A more recent application of computer vision to wood products is the automatic identification of wood types (Labati et al., 2009). This is about identifying the different types of wood that make up wood shipments; a procedure that has been used to detect illegally-traded timber (Hermanson and Wiedenhoeft, 2011). Herein we are concerned with the problem of sorting parquet hardwood into different colour tones. We therefore assume that hardwood has been already graded, either automatically or manually. A round-up of the meth- ods presented in this paragraph can be found in Tab. 1. Piuri and Scotti (2010) noted that approaches to colour-based sorting can be divided into two groups: image-based and spectrum-based processing systems. Both groups have pros and cons. Spectrum-based systems have the advantage of relying on device-independent data, but the disadvantage of a limited inspection area, which does not allow for full-field measurements. Conversely, image- based systems enable full-field inspection, but need colour calibration to produce device-independent data. Vienonen et al. (2002) described a spec- trophotometric system which takes four circular samples of approximately 20mm radius each from each parquet block. A 56-bin spectrum (from 275 to 965 nm) is computed from each block and used as feature vector. Clas- sification is based on two approaches: minumum distance classifier and a subspace classifier. Likewise, Buchelt and Wagenführ (2012) used a spec- trophotometer to evaluate colour differences of native wood surfaces. In their approach spectral data are converted into CIE Lab to estimate the 4 intra-class colour difference (∆E∗ ab ) of different species. In the same way, Schnabel et al. (2009) use spectrophotometric data and convert them into CIE Lab to model colour changes that wood undergoes during its lifetime. Methods based on image processing work with the output of industrial cameras, which is usually a set of RGB triplets. These can be either used ‘as is’ or converted into different colour spaces, such as HSV, CIE Lab, etc. In both cases the aim is to extract global statistical descriptors that char- acterize the colour content of the images. In the design of an expert system for wood sorting based on image processing, one has to deal with the choice of the right colour space and the appropriate descriptor. Related literature shows that various solutions have been proposed in the past. Lu et al. (1997) described a system based on 3D RGB histograms and minimum distance classifier for real-time colour sorting of edge-glued panel parts reporting an accuracy ranging from 83,0% to 99,1%. Kurdthongmee (2008) described an approach for colour-based classification of rubberwood boards for fingerjoint manufacturing in which a neural network is fed with a normalized histogram of hue (H). In a qualitative study Hrčka (2008) investigated the use of colour features to classify between common beech and European spruce, showing that colour coordinates in the CIE Lab space separate the two species rather well. Faria et al. (2008) employed both device-dependent (HSV) and device- independent (CIE Lab) colour coordinates for sorting three different types of wood, namely cherry tree, beech tree and oak. Their method employs a fuzzy classifier based on a bell membership function for each of the colour coordinates. More recently Bombardier and Schmitt (2010) used mean and homogeneity extracted from CIE Lab and HSV channels as colour features, and a fuzzy reasoning classifier as the building blocks of an expert system for wood colour recognition. This review of image processing-based methods shows that a wide vari- ety of approaches have been proposed in literature, but, at the same time, leaves the reader uncertain about which is the ‘best practice’ when it comes to designing and implementing an automatic sorting system. The results presented in the papers cited above look in fact rather scattered, inhomoge- neous and therefore difficult to compare to each other. Even more difficult is to reproduce the results presented in them, for data and algorithms used in the experiments are not available. Lastly, it is worth noticing that none of the above cited works provides a comparative analysis of methods on a statistical basis. As a consequence it is difficult to provide sensible answers to questions like: Which colour descriptor gives the best accuracy? Is there a colour space superior to the others? What is the trade-off between accu- racy and computational payload? In the following sections we try to answer 5 Figure 1: The imaging system. these questions on the basis of a comparative experimental analysis. 3. Materials We considered 14 classes of common parquet hardwood of different type, treatment and finish, as detailed in Tab. 3. The number of colour tones for class ranges from two to four, whereas the number of samples for tone ranges from six to eight. All the materials considered in the experiment are top-quality, therefore containing none to very few minor defects. Prelimi- nary subdivision of the samples of each class into the different colour tones (‘ground truth’) has been carefully performed by a pool of experienced work- ers. Images of the hardwood parquet specimens have been captured through an imaging system composed of a dome illuminator (Monster Dome Light 18.25”), an industrial CMOS camera equipped with a 6 mm fixed focal length lens and three pins to support the dome (Fig. 1). Characteristics and settings of the camera and lens are reported in Tab. 2. This solution leaves enough space for the specimen to pass below the dome and the camera, as if it were carried by a conveyor belt, a set-up which closely resembles the industrial conditions in which the system is supposed to operate. The camera is attached to the dome through a custom-designed support which permits relative rotation between the camera and the dome around the focal axis of the camera. The imaging system is patent pending (Bianconi et al., 2012a). The voltage of the illuminator has been set to 18 V and maintained 6 Table 2: Characteristics and settings of the camera and lens. Camera Model Edmund Optics 5012C LE Red gain 1.78 × Green gain 1.00 × Blue gain 1.33 × Gamma correction No Resolution 2560 × 1920 Debayering quality High Image format RGB24 (.bmp) Lens Model Pentax H614-MQ Focal length 6 mm (fixed) Aperture value 5.6 Focus ≈30 cm constant throughout the image acquisition procedure. This provides a light level of about 78600 lx at the center of the field of view. In order to avoid colour artifacts arising from image binning and/or undersampling, images have been taken at the native resolution of the camera (2560 × 1920 pix- els), which corresponds to a spatial resolution of approximately 180 dpi. Preliminary white balance has been carried out to adjust colour rendition. 4. Methods In the following two subsections we review the colour descriptors (Sec. 4.1) and spaces (Sec. 4.2) considered in this paper. In Sec. 4.3 we also discuss the problem of converting colour data from a device-dependent space into a device-independent one. 4.1. Colour descriptors Colour descriptors are statistical parameters that summarize the colour content of an image irrespectively of the spatial distribution. Consequently they are invariant to translation and rotation, and only slightly dependent on the viewing angle. By contrast, they are highly sensitive to changes in illumination. It is therefore mandatory that this be kept constant during the acquisition process, a condition that can be obtained, for instance, through the imaging device used in our experiments (Fig. 1). In real working condi- tions we can safely assume that similar devices can be adopted to produce invariable illumination conditions. 7 Table 3: Summary list of the hardwood parquet samples used in the exper- iments. Wood class Botanical name Treatment Tones Samples / tone Image resolution 1 Clorophora excelsa None 8 1200 × 600 2 Quercus petrea Painted 8 1500 × 500 3 Quercus petrea UV-treated, painted 8 1300 × 1000 4 Quercus petrea Painted 8 1400 × 480 5 Quercus petrea None 6 1300 × 480 6 Quercus petrea UV-treated, brushed 8 1400 × 1300 7 Quercus petrea Oil-treated, hand-planed, painted 7 1400 × 1300 8 Quercus petrea Oil-treated, hand-planed 8 1500 × 1300 9 Quercus petrea Thermo- treated 8 1500 × 1300 10 Quercus petrea Thermo- treated 8 1500 × 1300 11 Quercus petrea Brushed 8 1400 × 1300 12 Quercus petrea Oil-treated 8 2000 × 600 13 Tectona grandis None 8 1600 × 600 14 Tectona grandis None 8 1200 × 600 8 Table 4: Colour descriptors used in the experiments. Average extraction and classification time (in seconds) refer to the equipment described in Sec. 5.2 Color descriptor No. of features Avg. extraction time (a) Avg. classification time (b) (a + b) Mean 3 0.236 0.029 0.265 Mean + hom. 6 0.863 0.030 0.894 Mean + std 6 0.439 0.028 0.468 Mean + std + mom. 15 8.059 0.028 8.087 Quartiles 9 2.218 0.028 2.246 Quintiles 12 2.234 0.028 2.263 Marg. hist. (8 × 3) 24 0.542 0.029 0.571 Marg. hist. (16 × 3) 48 0.571 0.029 0.600 Marg. hist. (32 × 3) 96 0.610 0.029 0.639 3D hist. (83) 512 0.705 0.029 0.734 4.1.1. Soft colour descriptors López et al. (2008) introduced the term soft colour descriptors to refer to different combinations of the following simple statistical parameters: • Mean µc = 1 n n ∑ i=1 Ic,i (1) • standard deviation σc = 1 n − 1 √ √ √ √ n ∑ i=1 (Ic,i − µc) 2 (2) • k-th moment mc,k = n ∑ i=1 (Ic,i − µc) k hc (Ic,i) (3) In the above equations n is the number of pixels in the input image, Ic,i the intensity of the i-th pixel in the c-th colour channel, hc(x) the probability of the intensity value x in the c-th channel (usually estimated through a discrete histogram – hence the symbol hc, as in Eq. 4) and µc the average intensity value of the c-th channel. Throughout the paper we assume that Ic,i ∈ [0, 1] in any colour space. In our implementation this is obtained through appropriate normalization of the input data. Soft descriptors are easy to implement and fast to compute, therefore particularly suitable for real-time processing. Different combinations of soft 9 colour descriptors proved effective in applications ranging from grading of stoneware tiles (López et al., 2008) to sorting of recyclable paper (Rahman et al., 2011). In particular, mean values of the R, G and B channels alone showed surprisingly good performance in many tasks including automatic grading of ceramic tiles (Kukkonen et al., 2001) and banknote recognition (Garćıa-Lamont et al., 2012). Recently, Bombardier and Schmitt (2010) approached the problem of wood colour recognition by combining mean and homogeneity in the CIE Lab space. As for homogeneity, herein we used the following definition: homc = L ∑ l=1 hc [x (l)] x (l) (4) where x (l) represents the intensity value corresponding to the l-th bin of the histogram hc. In our implementation the value x (l) is measured at the centroid of each bin. Since bins’ edges are monotonically-increasing values in [0, 1], this approach ensures that Eq. 4 generates no division by zero. For this reason the definition of homogeneity used here is slightly different from the one proposed in the original references (Bombardier and Schmitt, 2010; Schmitt, 2007). In that case, in fact, a division by zero may occur when i = 0 (see Schmitt, 2007, Eq. 3.24). 4.1.2. Colour percentiles A percentile is the value that cuts the distribution of a random vari- able into two parts so that a given percent of observations fall below that value. Colour percentiles are generally computed from each colour channel. In wood grading they have been used both alone and in combination with textural features. Kauppinen (2000) employed colour percentiles alone in a non-segmenting method for grading parquet slabs. Niskanen et al. (2001) used colour percentiles in combination with either co-occurrence features or Local Binary Patterns for identification of knots in wood inspection. Re- cently Bianconi et al. (2012b) proved the method effective also in automatic classification of granite tiles. In the experiments presented herein, we used quartiles and quintiles of each colour channel. These are the values that cut the probability distribution of the intensity into equally-populated groups each containing 1/4 and 1/5 of the population, respectively. 4.1.3. Marginal histograms Marginal histograms estimate the colour content of an image through the probability distribution of colours as a function of each channel sepa- 10 rately, thus discarding any information about the other channels. Marginal histograms can be considered as projections of the 3D colour histogram (dis- cussed in the following subsection) into three one-dimensional subspaces. Herein we considered histograms composed of 8, 16 and 32 bins for each channel, giving 24, 48 and 96 features, respectively (see Tab. 4). Marginal colour histograms performed well in practical applications such as classifica- tion of printed colour paper (Pietikäinen et al., 1996), natural rocks (Lepistö et al., 2005) and generic colour textures (Bianconi et al., 2011a). In colour- based wood sorting their use has been propounded by Arden (1991) and, more recently, by Kurdthongmee (2008). 4.1.4. 3D colour histogram The 3D colour histogram estimates the joint probability distribution in the colour space. The method, which was originally proposed by Swain and Ballard (1991), consists of dividing the colour space in parts of equal volume and counting how many times each part is represented in the input image. Lu et al. (1997) used the 3D colour histogram for colour sorting edge- glued wooded panels. In our experiments we adopted the implementation proposed by Mäenpää and Pietikäinen (2004) in which the colour space is partitioned by dividing each colour channel into segments of equal length. Since we used eight subdivisions for each channel, the method generates 83 = 512 features. 4.2. Colour spaces Critical to the application studied herein is the choice of the right colour space. This can be either a device-dependent or a device-independent one. Device-dependent spaces are not directly related to how the human vision system perceives colours: they simply encode device-specific data at the device level (Kang, 2006). Device-dependent spaces include additive spaces, such as RGB and HSV, and subtractive spaces (not considered here). In contrast, device-independent colour spaces (also called colorimetric spaces) are directly related to the human vision system. Their main objective is in fact to define colour coordinates that are universally valid for the group of normal observers (Wyszecki and Styles, 1982). The basic colorimetric space is CIE XYZ. Any colour space that can be directly transformed into CIE XYZ is device-independent. A colour space is said uniform when the Euclidean distance between colours in that space is proportional to colour differences as perceived by humans. CIE Lab and CIE Luv are examples of device-independent and uniform colour spaces. 11 Related literature for a long time has debated about whether there is a colour space superior to the others. Device-independent and uniform colour spaces should be preferable – at least in principle – since they are inti- mately related to the way humans perceive colours. This assertion, how- ever, has not been clearly confirmed by the experiments. Comparison of different colour spaces for image classification has in fact led, so far, to con- tradictory or inconclusive results. Paschos (2001) found that perceptually uniform/approximately uniform colour spaces (CIE Lab and HSV, respec- tively) outperform RGB in many cases. By contrast, other authors found no significant difference among the colour spaces considered in their works: in a texture classification experiment Drimbarean and Whelan (2001) showed that none of the RGB, HSI, CIE XYZ, CIE Lab and YIQ proved sufficiently superior; likewise, Brunner et al. (1992) found no practically important dif- ferences in performance among RGB, HSV, CIE Lab, CIE Luv and YIQ for defect detection in Douglas-fir veneer. Finally, Qazi et al. (2011) recently showed that CIE Lab outperforms RGB and IHLS with textured images, but the trend is reversed for pure colour feature cues. In this paper we considered the same colour spaces studied by Paschos (2001), namely two device-dependent spaces (RGB and HSV) and one device- independent space (CIE Lab). We believe these represent sensible and viable choices for the problem studied herein. In practical applications the use of RGB data is dictated by the availability of such data as direct output of the imaging system. This avoids the computational overhead required to convert RGB data into other colour spaces. RGB, however, is not percep- tually uniform. HSV, in contrast, is approximately uniform and decouples colour data into an intensity (V channel) and a chromatic part (H and S channels). RGB colour data are converted to HSV through simple equations (Kang, 2006). Finally, CIE Lab is a device-independent and uniform colour space closely related to the human vision system. To obtain Lab colour coordinates from RGB, one needs to colour calibrate the image acquisition system. The procedure adopted hereis described in the following section. 4.3. Colour calibration Colour calibration consists of determining a function that maps device- dependent colour data into device-independent ones (León et al., 2006). The form of the function is established a priori. Most commonly this is a polynomial, but other methods, such as lookup tables (Po-Chieh, 1993) and neural networks (Schettini, 1995) have been proposed as well. In our experiments we adopted a simple linear model, which can be expressed in the following way: 12     L̂ ∗ â ∗ b̂ ∗ 1     =     M11 M12 M13 M14 M21 M22 M23 M24 M31 M32 M33 M34 0 0 0 1         R G B 1     (5) where L̂∗, â∗ and b̂∗ represent the estimated CIE Lab colour coordinates. This approach has some advantages: it is fast, easy to implement and re- quires the estimation of few parameters. Moreover, previous experiments showed that linear models perform as well as higher-degree models in colour calibration (Bianconi et al., 2011b). In order to determine the parameters of the model (the Mij coefficients), we need a set of reference colour patches of which both the device-dependent [R, G, B] T and the device-independent coordinates [L∗, a∗, b∗] T are known. The first are the raw output of the imaging system, the second can be measured through a colorimeter or any other colour measuring device. Most commonly device-independent colour data come with the reference patches. Here we used a standard calibra- tion set (X-Rite R© Color Checker) which contains 24 colour patches (Fig. 2). The corresponding device-independent data are available online (Babel- Color, 2012). Figure 2: The 24 reference colours of the X-Rite R© Color Checker. The unknown parameters are estimated through a least-squares proce- dure: M = arg min Mij ∈ R { R ∑ r=1 [ ( L̄ ∗ r − L̂ ∗ r )2 + (ā ∗ r − â ∗ r) 2 + ( b̄ ∗ r − b̂ ∗ r )2 ] } (6) where L̄∗, ā∗ and b̄∗ are the ‘true’ CIE Lab colour coordinates of the reference colour patches, R is the number of the patches (24, in this case) and M the matrix of Mij coefficients (Eq. 5). 13 5. Experiments In the experimental part we estimated the accuracy of the methods pre- sented in Sec. 4. The experiments have been designed to replicate the manual sorting process usually adopted in industry. In our experience this works as follows: As a first step one or more workers select one represen- tative wood sample for each tone of the class of wood which is going to be produced thereafter (in our experiments tones range from two to four per class). The samples are then passed to the worker in charge of the sorting process (let us refer to him as the ‘selector’), whom in general is given a short time to familiarize with the grades before the selection process begins. During the selection process the selector takes a position from which he can see both the tables to select (which are usually carried on a conveyor belt) and the samples of each grade. Whenever a table arrives, the selector diverts the slab to the storage bin corresponding to the sample that the specimen resembles most. In practice the whole process is clearly a supervised clas- sification task in which the samples represent the training set. Therefore, in order to comparatively evaluate the performance of the colour descrip- tors and spaces presented in the preceding sections, we submitted them to a supervised classification task. For each wood class we considered all the classification problems that can be generated by choosing one sample per tone for training while leaving the others for validation. This procedure pro- vides a deterministic estimation of the classification accuracy. The accuracy estimated this way is also very realistic, in our view, because the appraisal procedure matches the real working conditions quite well. Given T the number of tones for each wood class and S the number of samples for each tone (see Tab. 3), the classification accuracy a can be expressed in the following way: a = 1 PT (S − 1) P ∑ p=1 Cp (7) where P is the number of problems (P = ST ) and Cp the number of correctly classified samples in the p-th problem. T (S − 1) represents the number of samples to classify in each problem. 5.1. Classifier The selection of the appropriate classifier is always a crucial and difficult step in the design of an expert system. In this work two conditions limit this choice drastically: 1) the need for a classifier that works even with one 14 training sample only (see the considerations reported at the beginning of Sec. 5), and 2) the absence of tuning parameters, which may significantly modify the relative performance of the colour descriptors and spaces. This considered, we thought that the 1-NN classifier would be the most appropri- ate solution for the problem studied herein (in our implementation we used the Euclidean (L2) distance). Absence of tuning parameters, easiness of implementation and other desirable asymptotic properties make this classi- fication strategy particularly suitable for comparative purpose. Furthermore the 1-NN can work even with few training samples (as few as one, like in this case), whereas other classifiers, such as SVM, cannot. Finally, recent literature strongly supports the use of the 1-NN for performance compari- son of image analysis algorithms (Crosier and Griffin, 2010; Guo et al., 2010; Kandaswamy et al., 2011; Liu et al., 2012). 5.2. Implementation, execution and reproducible research Methods and algorithms have been implemented in Matlab R© R14. Im- age acquisition and classification have been performed in the Department of Industrial Engineering, at the University of Perugia, on a laptop PC equipped with Intel R© T2300, 1GB RAM, and Windows TM XP – 32 bits, Service Pack 3. For reproducible research purposes, all the data required to replicate the experiments (i.e.: source code, images and subdivisions into train and validation sets) are available in Ref. PG (2012)1. 6. Results Classification accuracies for each colour descriptor, wood class and colour space are reported in Tab. 5. On average the results confirm the effectiveness of the methods considered in the paper, with most methods attaining an accuracy close to 90%. For comparison purposes we have computed the 95% confidence intervals of the mean classification accuracy attained by each colour descriptor in each of the three colour spaces (Fig. 3). Since data (average accuracies) are not normally-distributed, confidence intervals for the means have been estimated through percentile bootstrap (Schmidheiny, 2012). The procedure consist of drawing a number B of bootstrap samples, computing the distribution of the mean over the bootstrap samples and deriving the lower and upper bounds of the confidence interval as the 2.5 and 97.5 percentiles of such distribution. 1To access the page: user = parquet, psw = sorting 15 Table 5: Average classification accuracy. Colour descriptor Wood class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Avg. RGB Mean 94.9 87.9 69.0 97.7 94.8 100.0 86.8 86.2 71.4 89.9 97.3 100.0 91.9 81.2 89.2 Mean + hom. 94.7 87.7 69.0 97.7 94.8 100.0 87.2 86.2 69.6 86.6 97.3 100.0 91.8 81.1 88.8 Mean + std 95.1 88.3 67.6 97.7 95.0 100.0 87.9 86.1 72.2 90.2 97.3 100.0 91.5 81.4 89.3 Mean + std + mom. 95.1 88.3 67.6 97.7 95.0 100.0 87.9 86.1 72.2 90.2 97.3 100.0 91.5 81.4 89.3 Quartiles 95.1 87.9 68.9 97.9 95.6 100.0 88.3 87.1 70.8 90.0 97.5 100.0 91.2 81.3 89.4 Quintiles 94.6 86.5 67.4 97.7 95.0 100.0 88.3 87.2 71.9 90.3 97.4 100.0 91.1 80.9 89.2 Marg. hist. (8 × 3) 91.2 83.5 67.1 97.0 90.4 100.0 85.3 86.9 57.9 83.8 97.9 100.0 84.7 85.5 86.5 Marg. hist. (16 × 3) 87.7 89.4 65.5 98.7 90.9 100.0 87.1 86.5 67.3 90.0 97.7 100.0 88.1 82.5 88.0 Marg. hist. (32 × 3) 94.0 87.6 64.6 98.4 94.4 100.0 91.6 87.6 68.6 89.5 98.1 100.0 89.1 80.5 88.9 3D hist. (83) 88.5 82.8 65.3 99.4 87.8 100.0 85.5 85.9 53.0 80.4 98.3 100.0 80.9 85.9 85.3 Avg. 93.1 87.0 67.2 98.0 93.4 100.0 87.6 86.6 67.5 88.1 97.6 100.0 89.2 82.2 88.4 HSV Mean 94.1 86.0 68.8 97.9 94.4 100.0 90.5 74.9 77.3 93.3 95.3 100.0 86.5 84.2 88.8 Mean + hom. 94.4 84.5 68.9 97.9 94.4 100.0 88.0 75.6 77.5 93.5 95.3 100.0 86.6 84.5 88.6 Mean + std 94.7 86.4 67.3 97.9 94.3 100.0 89.3 73.9 76.2 93.6 95.3 100.0 85.7 85.1 88.6 Mean + std + mom. 94.7 86.4 67.3 97.9 94.3 100.0 89.3 73.8 76.1 93.6 95.3 100.0 85.7 85.1 88.5 Quartiles 95.6 85.6 67.9 98.0 95.6 100.0 89.8 74.7 77.5 93.2 95.9 100.0 86.8 83.4 88.8 Quintiles 94.6 84.6 67.3 98.1 94.6 100.0 90.4 74.4 77.8 94.0 95.5 100.0 87.1 83.4 88.7 Marg. hist. (8 × 3) 90.6 88.7 66.5 100.0 87.0 100.0 89.1 74.1 63.6 88.5 97.3 98.4 74.4 84.3 85.9 Marg. hist. (16 × 3) 89.3 86.6 64.8 97.8 89.6 100.0 88.3 85.4 79.5 92.0 97.2 100.0 84.2 86.0 88.6 Marg. hist. (32 × 3) 95.2 86.2 64.1 97.3 94.4 100.0 93.3 83.8 77.9 92.4 98.5 100.0 83.6 80.2 89.1 3D hist. (83) 89.2 63.2 62.8 100.0 87.6 100.0 59.6 58.3 46.5 77.6 97.7 98.3 59.9 80.7 77.2 Avg. 93.2 83.8 66.6 98.3 92.6 100.0 86.8 74.9 73.0 91.2 96.3 99.7 82.1 83.7 87.3 Lab Mean 94.7 87.9 68.9 96.4 95.0 100.0 87.1 85.3 76.1 91.7 97.7 100.0 91.9 81.4 89.6 Mean + hom. 94.7 87.9 68.9 96.4 95.0 100.0 87.1 85.3 76.1 91.7 97.7 100.0 91.9 81.4 89.6 Mean + std 95.0 88.1 67.7 96.4 95.0 100.0 88.2 85.0 75.7 92.0 97.7 100.0 91.7 81.5 89.6 Mean + std + mom. 95.0 88.1 67.7 96.4 95.0 100.0 88.2 85.0 75.7 92.0 97.7 100.0 91.7 81.5 89.6 Quartiles 95.5 88.5 67.6 96.4 96.3 100.0 88.0 85.9 75.4 91.5 98.4 100.0 91.7 80.9 89.7 Quintiles 95.1 87.5 67.4 96.9 94.8 100.0 88.8 86.5 75.8 91.5 98.4 100.0 92.4 81.6 89.8 Marg. hist. (8 × 3) 87.7 65.3 65.0 86.4 94.6 100.0 79.7 87.4 70.4 88.3 90.1 100.0 78.6 61.9 82.5 Marg. hist. (16 × 3) 95.8 81.5 63.4 96.3 91.7 100.0 82.5 83.9 78.5 92.5 99.0 100.0 89.8 80.0 88.2 Marg. hist. (32 × 3) 96.2 87.7 62.6 90.0 95.4 100.0 92.4 79.1 81.3 94.9 99.8 100.0 91.1 80.0 89.3 3D hist. (83) 87.7 64.7 65.1 85.5 96.7 100.0 79.9 87.4 70.4 88.2 86.5 100.0 79.0 61.9 82.4 Avg. 93.8 82.7 66.4 93.7 94.9 100.0 86.2 85.1 75.5 91.4 96.3 100.0 89.0 77.2 88.0 16 60 70 80 90 100 Mean Mean + hom. Mean + std Mean + std + mom. Quartiles Quintiles Marg. hist. (8 × 3) Marg. hist. (16 × 3) Marg. hist. (32 × 3) 3D hist. (83) (a) RGB 60 70 80 90 100 Mean Mean + hom. Mean + std Mean + std + mom. Quartiles Quintiles Marg. hist. (8 × 3) Marg. hist. (16 × 3) Marg. hist. (32 × 3) 3D hist. (83) (b) HSV 60 70 80 90 100 Mean Mean + hom. Mean + std Mean + std + mom. Quartiles Quintiles Marg. hist. (8 × 3) Marg. hist. (16 × 3) Marg. hist. (32 × 3) 3D hist. (83) (c) CIE Lab Figure 3: Confidence intervals for the average classification accuracy (%) in the three colour spaces. Parameter B needs to be set by the user: Schmidheiny (2012) recommends using 1000 or more replications; herein we used 2000, as suggested by Wang (2001). We observe that in none of the three colour spaces there is a colour descriptor significantly superior to the others. Indeed Fig. 3 clearly shows large overlap among all the confidence intervals. It is interesting to notice that the simplest descriptor (mean of each colour channel) performs as good as the others. In contrast, the performance of 3D colour histograms is appreciably lower – though, as we mentioned above, this difference does not reach statistical significance. This result is logical and stems from the intrinsic structure of colour histograms (Bianconi et al., 2009): since colour differences between grades are very subtle (see Fig. 3) colours tend to spread over a limited portion of the colour space. Therefore, while different colours can be assigned to the same bin of the 3D histogram, many other bins remain empty. The same considerations apply to marginal histograms: here 17 we notice that accuracy improves as the number of subdivisions increases, as one would expect, since a finer subdivision of each colour axis allows for better discrimination between tones. Soft colour descriptors show very similar performance, suggesting that it is not much use adding higher-order statistical descriptors to the simple mean. Likewise, colour percentiles do not represent a significant improvement on soft colour descriptors. 80 90 100 RGB HSV CIE Lab Figure 4: Confidence intervals for the average classification accuracy (%) for each colour space. As for colour spaces, the results show that there is very little difference among RGB, HSV and CIE Lab: none of them proved significantly superior. These findings confirm the conclusions obtained by Brunner et al. (1992) and Drimbarean and Whelan (2001). These results suggest that RGB data are better employed with no changes or modifications: converting them into other colour spaces only adds unnecessary overhead without appreciable beneficial effects. We also notice that the general trend is very similar for the three colour spaces, showing no appreciable interaction effects between colour descriptor and colour space. Finally, it is worth mentioning that the overall processing time (feature extraction + classification) is rather contained, with most descriptors com- pleting the task in less than 1 sec. (see Tab. 4). We can see that most of the variability in computing time is due to feature extraction, whereas 1-NN classification requires almost the same amount of time for all descriptors. Furthermore, we should not forget that these figures have been obtained using a scripting language running on low cost hardware; therefore they largely overestimate the real computing time that can be achieved through dedicated hardware and optimized code. 18 7. Conclusions In this paper we have presented a performance analysis of different colour descriptors and spaces for sorting parquet slabs into classes of similar visual appearance. To the best of our knowledge, this is the first comprehensive study on the subject based on statistical analysis and reproducible experi- ments. The overall results show, on average, a low error classification process with about 90% accuracy. Most methods are also computationally inexpen- sive, therefore suitable for real-time processing. This outcome is satisfactory, considering that the results have been obtained with standard industrial equipment – specifically a single-sensor camera – and a very simple classifier (1-NN). We believe that the overall accuracy could be significantly increased by adopting higher-level imaging devices (i.e.: 3-sensor camera) and more sophisticated classifiers. The comparative analysis showed no significant dif- ference between the descriptors and colour spaces considered in the paper, therefore suggesting – from both standpoints of accuracy and computational efficiency – the use of simple statistical descriptors along with RGB data as the best practice. Acknowledgements This work was supported by the Spanish Government under projects TRA2011-29454-C03-01 and CTM2010-16573, and by the European Union within project Life09-ENV/FI/000568. References Arden, T., 1991. Color sorting of lumber. US patent no. 4992949. Åström, A., Åstrand, E., Johansson, M., 1999. Automatic parquet block sorting using real-time spectral classification. In: Real-Time Imaging IV. Vol. 3645 of Proceedings of SPIE. San Jose, California, USA, pp. 76–84. BabelColor, 2012. The ColorChecker (since 1976!). Available on- line at http://www.babelcolor.com/main_level/ColorChecker.htm# ColorChecker_data. Page visited on June 1, 2012. Bianconi, F., Fernández, A., González, E., Caride, D., Calviño, A., 2009. Rotation-invariant colour texture classification through multilayer CCR. Pattern Recognition Letters 30 (8), 765–773. 19 Bianconi, F., Harvey, R., Southam, P., Fernández, A., 2011a. Theoretical and experimental comparison of different approaches for color texture classification. Journal of Electronic Imaging 20 (4), 043006–1–17. Bianconi, F., Saetta, S., Sacchi, G., Asdrubali, F., Baldinelli, G., 2011b. Colour calibration of an artificial vision system for industrial applications: comparison of different polynomial models. In: Rossi, M. (Ed.), Colour and Colorimetry Multidisciplinary Contributions. No. 21 in Optics and Photonics Series Notebooks. Maggioli Editore, pp. 18–25. Bianconi, F., González, E., Fernández, A., Saetta, S. A., 2012a. Apparatus to acquire a plurality of superficial images of at least one body and re- lated method. IT patent application no. MI2012A001299. Filed on July 25, 2012. Bianconi, F., González, E., Fernández, A., Saetta, S. A., 2012b. Automatic classification of granite tiles through colour and texture features. Expert Systems with Applications 39 (12), 11212–11218. Bombardier, V., Schmitt, E., 2010. Fuzzy rule classifier: Capability for gen- eralization in wood color recognition. Engineering Applications of Artifi- cial Intelligence 23 (6), 978–988. Bond, B. H., Kline, D. E., Araman, P. A., 2002. Differentiating defects in red oak lumber by discriminant analysis using color, shape, and density. Wood and Fiber Science 34 (4), 516–528. Brunner, C., Maristany, A., Butler, D., VanLeeuwen, D., Funck, J., 1992. An evaluation of color spaces for detecting defects in Douglas-fir veneer. Industrial Metrology 2, 169–184. Buchelt, B., Wagenführ, A., 2012. Evaluation of colour differences on wood surfaces. European Journal of Wood and Wood Products 70, 389–391. Castellani, M., Rowlands, H., 2009. Evolutionary artificial neural network design and training for wood veneer classification. Engineering Applica- tions of Artificial Intelligence 22 (4-5), 732–741. Ciccotelli, J., Portala, J.-F., 1992. Applications of artificial vision in the wood industry. Industrial Metrology 2, 185–194. Conners, R. W., McMillin, C. W., Lin, K., 1983. Identifying and locating surface defects in wood: part of an automated lumber processing system. 20 IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI- 5 (6), 573–583. Conners, R. W., Cho, T.-H., Ng, C. T., Drayer, T. H., Araman, P. A., Brisbin, R. L., 1992. A machine vision system for automatically grading hardwood lumber. Industrial Metrology 2, 317–342. Crosier, M., Griffin, L. D., 2010. Using basic image features for texture classification. International Journal of Computer Vision 88, 447–460. DIN-1611, 2002. Sawn timber - appearance grading of softwood - part 1: European spruces, firs, pines, douglas firs and larches. Deutsches Institut für Normung. DIN-EN-975-1, 2011. Sawn timber - appearance grading of hardwoods - part 1: Oak and beech. Deutsches Institut für Normung. DIN-EN-975-2, 2004. Sawn timber - appearance grading of hardwoods - part 2: Poplars. Deutsches Institut für Normung. Drimbarean, A., Whelan, P. F., 2001. Experiments in colour texture analy- sis. Pattern Recognition Letters 22 (10), 1161–1167. Faria, J., Martins, T., Ferreira, M., Santos, C., 2008. A computer vision system for color grading wood boards using fuzzy logic. In: Proceedings of the IEEE International Symposium on Industrial Electronics (ISIE 2008). Cambridge, UK, pp. 345–350. Garćıa-Lamont, F., Cervantes, J., López, A., 2012. Recognition of mexi- can banknotes via their color and texture features. Expert Systems with Applications 39 (10), 9651–9660. Gu, I., Andersson, H., Vicen, R., 2010. Wood defect classification based on image analysis and support vector machines. Wood Science and Technol- ogy 44 (4), 693–704. Guo, Z., Zhang, L., Zhang, D., 2010. A completed modeling of local binary pattern operator for texture classification. IEEE Transactions on Image Processing 19 (6), 1657–1663. Hermanson, J. C., Wiedenhoeft, A. C., 2011. A brief review of machine vision in the context of automated wood identification systems. IAWA Journal 32 (2), 233–250. 21 Hrčka, R., 2008. Colour modelling as a tool for wood grading. In: Proceed- ings of COST E53 – Quality control for wood and wood products. Delft, The Netherlands, pp. 165–170. Jabo, S., 2011. Machine vision for wood defect detection and classification. Master’s thesis, Chalmers University of Technology, Göteborg, Sweden. Kandaswamy, U., Schuckers, S. A., Adjeroh, D., 2011. Comparison of texture analysis schemes under nonideal conditions. IEEE Transactions on Image Processing 20 (8), 2260–2275. Kang, H. R., 2006. Computational Color Technology. Spie Press. Kauppinen, H., September 2000. A two stage defect recognition method for parquet slab grading. In: Proceedings of the 15th International Conference on Pattern Recognition. Vol. 4. Barcelona, Spain, pp. 803–806. Kline, D., Surak, C., Araman, P., 2006. Automated hardwood lumber grad- ing utilizing a multiple sensor machine vision technology. Computers and Electronics in Agriculture 41 (1-3), 139–155. Kose, E., 2008. Modelling of colour perception of different age groups using artificial neural networks. Expert Systems With Applications 34 (3), 2129– 2139. Kukkonen, S., Kälviäinen, H., Parkkinen, J., 2001. Color features for quality control in ceramic tile industry. Optical Engineering 40 (2), 170–177. Kurdthongmee, W., 2008. Colour classification of rubberwood boards for fingerjoint manufacturing using a SOM neural network and image pro- cessing. Computers and Electronics in Agriculture 64 (2), 85–92. Kyllönen, J., Pietikäinen, M., November 2000. Visual inspection of parquet slabs by combining color and texture. In: Proceedings of the IAPR Work- shop on Machine Vision Applications. Tokyo, Japan, pp. 187–192. Labati, R., Gamassi, M., Piuri, V., Scotti, F., May 2009. A low-cost neural- based approach for wood types classification. In: Proc. of the International Conference on Computational Intelligence for Measurement Systems and Applications. Hong Kong, China, pp. 199–203. León, K., Mery, D., Pedreschi, F., León, J., 2006. Color measurements in L∗a∗b∗ units from RGB digital images. Food Research International 39 (10), 1084–1091. 22 Lepistö, L., Kunttu, I., Visa, A., 2005. Color-based classification of natu- ral rock images using classifier combinations. Lecture Notes in Computer Science 3540, 901–909. Liu, L., Fieguth, P., Clausi, D., Kuang, G., 2012. Sorted random projections for robust rotation-invariant texture classification. Pattern Recognition 45, 2405–2418. López, F., Valiente, J. M., Prats, J. M., Ferrer, A., 2008. Performance evalu- ation of soft color texture descriptors for surface grading using experimen- tal design and logistic regression. Pattern Recognition 41 (5), 1744–1755. Lu, Q., Conners, R., Kline, D., Araman, P., October 1997. A real-time algorithm for color sorting edge-glued panel parts. In: Proc. of the Inter- national Conference on Image Processing. Vol. 1. Barcelona, Spain, pp. 822–825. Lycken, A., 2006. Comparison between automatic and manual quality grad- ing of sawn softwood. Forest Products Journal 56 (4), 13–18. Mäenpää, T., Pietikäinen, M., 2004. Classification with color and texture: jointly or separately? Pattern Recognition 37 (8), 1629–1640. Niskanen, M., Silvén, O., Kauppinen, H., June 2001. Color and texture based wood inspection with non-supervised clustering. In: Proceedings of the 12th Scandivanian Conference on Image Analysis (SCIA 2001). Bergen, Norway, pp. 336–342. Paschos, G., 2001. Perceptually uniform color spaces for color texture anal- ysis: An empirical evaluation. IEEE transactions on Image Processing 10 (6), 932–937. PG, 2012. Parquet sorting. Code, data and results related to this paper. Available online at http://dismac.dii.unipg.it/parquet. Pietikäinen, M., Nieminen, S., Marszalec, E., Ojala, T., August 1996. Ac- curate color discrimination with classification based on features distribu- tions. In: Proceedings of the 13th International Conference on Pattern Recognition (ICPR’96). Vol. 3. Vienna, Austria, pp. 833–838. Piuri, V., Scotti, F., 2010. Design of an automatic wood types classifica- tion system by using fluorescence spectra. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews 40 (3), 358–366. 23 Po-Chieh, H., 1993. Colorimetric calibration in electronic imaging devices using a look-up-table model and interpolations. Journal of Electronic Imaging 2 (1), 53–61. Qazi, I., Alata, O., Burie, J., Moussa, A., Maloigne, C. F., 2011. Choice of a pertinent color space for color texture characterization using parametric spectral analysis. Pattern Recognition 44 (1), 16–31. Rahman, M. O., Hussain, A., Scavino, E., Basri, H., Hannan, M. A., 2011. Intelligent computer vision system for segregating recyclable waste papers. Expert Systems with Applications 38, 10398–10407. Roz̆man, D., Brezak, M., Petrovic, I., 2006. Parquet sorting and grading based on color and texture analyses. In: Proceedings of the IEEE Inter- national Symposium on Industrial Electronics. Vol. 1. Montreal, Canada, pp. 655 –660. Schettini, R., 1995. Colorimetric calibration of color scanners by back- propagation. Pattern Recognition Letters 16 (10), 1051–1056. Schmidheiny, K., 2012. The bootstrap. In: Short Guides to Microeconomet- rics. Universität Basel, available online at http://kurt.schmidheiny. name/teaching/bootstrap2up.pdf. Page visited on July 26, 2012. Schmitt, E., 2007. Contribution au système d’information d’un pro- duit “bois”. Appariement automatique de pièces de bois selon des critères de couleur et de texture. Ph.D. thesis, Faculté des Sciences & Techniques, Université Henri Poincaré Nancy - I, available on- line at: http://tel.archives-ouvertes.fr/docs/00/17/01/06/PDF/ Memoire_These_Emmanuel_SCHMITT_V2.pdf. Schnabel, T., Zimmer, B., Petutschnigg, A. J., 2009. On the modelling of colour changes of wood surfaces. European Journal of Wood and Wood Products 67, 141–149. Sobey, P., Semple, E., 1989. Detection and sizing visual features in wood using tonal measures and a classification algorithm. Pattern Recognition 22 (4), 367–380. Swain, M. J., Ballard, D. H., 1991. Color indexing. International Journal of Computer Vision 7 (1), 11–32. Vienonen, P., Asikainen, A., Eronen, J., 2002. Color grading of beech par- quet blocks by using spectral data. Forest Products Journal 52, 49–52. 24 Wang, F., 2001. Confidence interval for the mean of non-normal data. Qual- ity and Reliability Engineering International 17, 257–267. Wyszecki, G., Styles, W., 1982. Color Science. Concepts and Methods, Quantitative Data and Formulae. Second Edition. Wiley-Interscience. 25 
work_cfhcpwq6dzacfhsrew347ia2um	----	Corel Office Document Australian Journal of Basic and Applied Sciences, 5(6): 1126-1136, 2011 ISSN 1991-8178 A Knowledge Base System to Control Construction Problems in Rigid Highway Pavements 1Ahmed Mancy Mosa, 2Riza Atiq O.K. Rahmat, 2Mohd Raihan Taha, and 1Amiruddin Ismail 1Sustainable Urban Transport Research Centre (SUTRA), Faculty of Engineering and Built Environment, The National University of Malaysia, 43600 UKM Bangi, Selangor Darul Ehsan, Malaysia 2Department of Civil & Structural Engineering, Faculty of Engineering and Built Environment, The National University of Malaysia, 43600 UKM Bangi, Selangor Darul Ehsan, Malaysia Abstract: Problems that face the highway engineer in the domain of rigid highway pavements construction are almost complex and effected by different factors. Diagnosing of these problems and suggestion of suitable solutions require significant amount of engineering knowledge, which is almost difficult to be confined at all sites of construction. Therefore, construction problems in rigid highway pavements represent an excellent domain for development of a system that contains a knowledge base covers descriptions, causes, and solutions of these problems. In this paper, the knowledge isacquired from written sources in the form of textbooks, papers, journals, manuals, electronic files, and any other written references. This knowledge is documented, analysed, and represented in a form of dilated flowchart that can be used as a suitable tool to understand the problems and find the solutions. The flowchart, then, is converted to a computer program by coding the knowledge using VISUAL BASIC programming language. The developed program named KBS-CCPRP represents a knowledge base system to control construction problems in rigid highway pavements. Key words: Knowledge base system, construction problems, rigid pavements INTRODUCTION Transportation in the third world countries, basically, depends on road network rather than on other facilities such as railways, marines and airways. Therefore, it is important to construct roads of high quality, both as regards smoothness of ride and durability (Robert Hennes and Martin Ekse, 1990). Highway pavements are classified into flexible and rigid. Flexible pavement includes all types of bituminous mats and in general has little beam strength. It does not distribute load over the subgrade by its flexural resistance, but depends upon the shear strength of the base and surfacing. Rigid pavements includes plain and reinforced Portland- cement concrete pavements slabs which act like any elastic floor slab except that the positions of the distributed and concentrated forces are reversed from their conventional positions in a building frame (Robert Hennes and Martin Ekse, 1990). During construction process of rigid pavements, many problems and decisions present themselves; these problems are very similar whether the concrete quantity to be laid is small or large, they range from worrying to financially catastrophic by costly repairs. In general, these problems affect the desired quality and increase the initial construction cost (Stock, 1988; Ken Newman, 1986 ). Moreover, novice engineer cannot overcome such problems,suggest suitable solutions, avoid their causes, and prevent same problems that may occur in the rest parts of the work. Classification and computerization of these problems, their causes, instantaneous solutions, and preventive solutions can be very helpful in control, minimize, and prevent them. This paper presents the development of a knowledge base system that can be used by the novice engineers in the sites of rigid pavement construction to control the problems facing them. Moreover, the system can be adopted as an instructional tool throughout the repeated use by interesting highway engineers. Domain of the Study: Many problems can occur during construction of rigid pavement such as preparation of under-layers, materials selection, concrete producing, hauling, placing, compacting, finishing, and curing. This paper concerns with problems that occur between placing time and the time of opening the road to the traffic. This involves concrete placing, spreading, compacting, finishing, texturing, jointing, curing, protecting, testing, and final Corresponding Author: Ahmed Mancy Mosa, Sustainable Urban Transport Research Centre (SUTRA), Faculty of Engineering and Built Environment, The National University of Malaysia, 43600 UKM Bangi, Selangor Darul Ehsan, Malaysia 1126 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 checking. Methods of rigid pavement construction like semi-manual construction-methods to fully mechanized construction, methods by fix form paving, and slip form paving methods are involved in this study.This paper does not involve diagnosing the problems associated with concrete production and hauling to the site, as it considers that the concrete is to be supplied by ready mix plants that are provided with all information needed about the product quality. This paper, also, does not involve diagnosing the problems associated with construction and preparation of the roadbed, as the study considers that there is a ready roadbed previously constructed but that does not prevent testing the regularity of this ready roadbed due to importance of surface regularity on the performance of the concrete road. Classification of the Domain Problems: Through the extensive review of the specialized references and repeated analysis, the domain problems are classified depending on their forms, locations, effects, and other common features in order to be diagnosed by the inspectors visibly or depending on tests and measurement results. In addition, problems description, likely causes and solutions are stated. Domain problems are classified into categories according to the common features and conditions, each category includes number of problems. These categories and problems are stated below: A. Concreting under bad conditions: These are problems restricting the construction process like bad weather, machine stop, and discontinuity of concrete feeding. B. Bleeding: Appearance of water at the surface of the plastic placed concrete at any stage of construction. C. Cracking: These are cracks which appear at the concrete surface during plastic stage (when the concrete still fresh) or after hardening of the concrete (during curing process). D. Joint construction problems: These are problems related to the formation of the joint groove or to the sealing of the joint. E. Surface irregularity: Presence of spots of high or low levelsdiffers from the specified. F. Dusting of pavement surface: Development of a fine, powdered material that easily rubs off the surface of hardened concrete. G. Scaling of pavement surface: Hardened pavement surface scaling can lead to severe pavement abrasion. H. Honeycomb forms: Irregular air voids appears at the extracted testing cores or at the side faces of the slab after removing the forms. I. Tests results problems: These problems can be diagnosed during the final inspection phase depending on field or laboratory tests results and measurements. Interaction among the problems, their descriptions, causes, and solutions is expected due to their common features and unity of the subject, but without any conflict among them. Figure (1) illustrates the diagram of domain problems classification. The problems of the first category (Concreting under bad conditions) their descriptions, likely causes, and possible solutions are explained in the following articles. A. Concreting Under Bad Conditions: Bad conditions during pavement construction process can be expected, since this process is performed outside and on wide area of different weathers, topography, environments, facilities, resources, and other features. Table (1) demonstrate the probable causes of these problems. Hot weather placement generally has conditions that are more problematic especially when ambient air temperature above 30oC is critical (Frank McCullough and Terry Dossey, 1999; ACI 305.1-06, 2007). Cold weather can stop the concreting when the temperature is 5oC and fall since such weather has harmful effects on the concrete (ACI 306.1-90, 2002). The concrete temperature shall be measured before placing in accordance with ASTM C1064 or AASHTOT309. The problems can be explained in the following articles: A1. Rapid Slump Loss: Rapid decrease in the concrete fluidity that leads to rip or tear slab surface during finishing (ACI302.1R-04, 2004; ACI121R-04, 2004). P Instantaneous solutions (ACI325.12R-02, 2002; ACI305R-99, 1999): 1. Add small quantity of extra water to the concrete in order to substitute the lost water due to evaporation during concrete hauling and placing. 2. Sprinkle the slab surface with minimum quantity of water from a fine jet in order to aid the finishing process in the extreme cases. 1127 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 Table 1: Likely causes of the problems of concreting under bad conditions Concreting under bad conditions problems -------------------------------------------------------------------------------------------- Causes A1 A2 A3 A4 A5 A6 A7 A8 Hot Weather (T>=30oC, low humidity, windy) x x Bad planning or managing of the pavement construction x x x x x x Unexpected weather changes x x x x Insufficient or incorrect weather information x x x x Low air temperature (T=<5oC) x Bad maintenance operation for the machines x No maintenance record for the machines x Concrete plant failure x Inadequate materials at the plant site x Concrete plant is far from the construction site x The truck mixers fails or be restricted in a traffic jam x Bad communication among site, plant, and trucks operators x Instable or disturbed sub-base layer x Incorrect handling by unskilled staff or incorrect tools x The sub-base layer is not covered by membranes x Incorrect vibrating method or immersing vibrators deeply which x mix the sub-base material by the placed concrete Fig. 1: Construction problems in rigid pavement diagram P Preventive solutions (ACI224.1R-07, 2007; Walter and Price, 1982): 1. Supply mixtures of highest practical slump. 2. Place and finish the concrete during the coldest interval of the day as soon as delivered. 3. Shade the forms, reinforcement and subgrade or cool them with water prior to concrete placing. 1128 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 4. Reduce wind velocity using windbreaks. 5. Protect the concrete with temporary coverings during any delay between concreting operations. 6. Increase the humidity at the surface by fog spraying during concreting and until curing starts. A2. Concrete Setting Rate Increases Rapidly: Rapid concrete stiffening leads to difficulty in concrete spreading (ACI 305.1-06, 2007). P Instantaneous solutions (ACI 305.1-06, 2007; ACI305R-99, 1999): 1. Complete the concreting operation as soon as possible to avoid final setting of the concrete before finishing by increasing the labours and machinery efforts. 2. Instruct the concrete plant operators to reduce the delivered concrete to ensure stable proportioning between the delivered and placed concrete. P Preventive solutions: Adopt the steps stated in article A1 and use retarders of concrete hardening. A3. Raining During Construction: This problem may take place during any stage of the construction operation. Rainwater has very harmful effects on concrete since it resolves the concrete structure and eliminates its strength (Innovative Pavement Research Foundation, 2005). P Instantaneous solutions (Innovative Pavement Research Foundation, 2005; ACI330R-01, 2001): 1. Stop the construction operation. 2. Inform the concrete plant operators to stop the concrete feeding. 3. Perform construction joint for the last performed part. 4. Protect the performed parts from rainwater by covering them by suitable covers. Care shall be taken to avoid distorting the concrete surface by water or covers. 5. Resume the construction operation after rain stopping or if protection arrangements (like temporary shades) are implemented to protect the concrete from the rainwater entirely. 6. Remove and reconstruct the defected parts of the pavement. P Preventive solutions (Innovative Pavement Research Foundation, 2005): 1. Adopt correct and accurate weather information. 2. Adopt high-level management plan in addition to a plan for emergency cases. 3. Employ experienced staff. A4. Concreting During Cold Weather: This problem can appear in cold areas when the temperature is 5oC and falling. Low temperatures can affect risky damages in the concrete structure due to freezing action (ACI 306.1-90, 2002; O'Flaharty, 1988). P Instantaneous solutions (ACI 306.1-90, 2002; O'Flaharty, 1988; ACI212.3R-04, 2004): 1. In case of sudden decrease in temperature, adopt the actions of article A3. 2. If the concrete must be placed in cold weather the following actions can be implemented: a) Heat the aggregate and water to moderate temperatures to make the temperature of the concrete between 10oC and 25oC prior to placing. b) Produce concrete using rapid-hardening cement. c) Add calcium chloride to the mixture to accelerate the hydration process that produces large amount of heat and increases concrete temperature. P Preventive solutions: Adopt the steps of article A3 and use rapid-hardening Portland cement and calcium chloride to the mixture to accelerate the hydration (Innovative Pavement Research Foundation, 2005; ACI212.3R-04, 2004). A5. Concreting During Dusty Weather: Contamination of concrete with dust due to dusty weather that reduces strength and durability of the concrete. 1129 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 P Instantaneous solutions (ACI330R-01, 2001; ACI302.1R-04, 2005): 1. In case of slight dust, the problem can be neglected. 2. In case of considerable dust that can cause significant contamination in concrete, adopt the first five steps explained in article A3 and take additional samples from the contaminated parts to ensure their compliance to the requirements of strength and durability otherwise, these parts shall be removed and reconstructed. P Preventive solutions: Adopt the same steps as in article A3. A6. Paver Failure During Construction: This problem may take place during concrete placing that can interrupt the construction process (ACI330R-01, 2001; ACI304R-00, 2000). P Instantaneous solutions (ACI330R-01, 2001; ACI304R-00, 2000): 1. Stop the construction operation. 2. Inform the concrete plant operators to stop concrete feeding. 3. Perform construction joint for the last performed part. 4. Resume the construction operation when the machine is re-operated and its efficiency is ensured, or when a new machine is provided. P Preventive solutions (ACI121R-04, 2004; Standard Specification for Urban Infrastructure, 2002): 1. Adopt periodic maintenance to the construction machines and prepare detailed report for each one. 2. Adopt high-level management plan in addition to a plan for emergency cases. 3. Employ experienced staff. A7. Discontinuity of Concrete Feeding During Construction: Irregular, interrupted, or discontinuous concrete delivery to the site can interrupt the construction process and lead to other problems (ACI302.1R-04, 2004; ACI302.1R-04, 2005). P Instantaneous solutions (ACI330R-01, 2001; ACI304R-00, 2000): 1. Stop the construction operation. 2. Perform construction joint for the last performed part. 3. Contact the plant operators to solve the problem. 4. Resume the construction operation when uniform concrete feeding is provided. P Preventive solutions (ACI304R-00, 2000; Standard Specification for Urban Infrastructure, 2002): 1. Maintain the concrete plant and the truck mixers periodically and prepare detailed report for each one. 2. Provide the site with concrete from the nearest plants. 3. Deposit sufficient materials at the plant site. 4. Provide good communication among the site, plant, and trucks operators. 5. Adopt high-level management plan in addition to a plan for emergency cases. 6. Employ experienced staff. A8. Contamination of Concrete by Sub-base Material: Sub-base material can contaminate the concrete during placing, compacting, or any other handling step (ACI302.1R-04, 2004; ACI302.1R-04, 2005). P Instantaneous solutions: Remove the contaminated parts and substitute with new concrete (ACI311.4R-05, 2005). P Preventive solutions (ACI121R-04, 2004; ACI309R-05, 2005): 1. Ensure correct preparation for the sub-base layer prior to concreting. 2. Cover the sub-base layer with membranes. 3. Employ experienced staff and proper tools (avoid garden tools). 4. Ensure correct vibration process and avoid immersing the vibrators vertically or deeply. 1130 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 Fgi. 2: Concreting under bad conditions flow chart. 1131 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 Building of the System: The classified knowledge, which represents the core of the system, is prepared to be coded as a computer system through repressing it in a sophisticated flow cart. The flowchart illustrates the manner of chaining used in the developed system. This flowchart starts with the main menu that includes the main titles of the categories. Each main title leads to a brief description, thereafter, the flowchart path leads to an optional question, which leads to the main menu or the path continues forward to reach the problems in the specified category. In the same manner, the path continues until the results. This procedure in chaining, which iterates from the beginning of the flow chart to the end of it, is called forward chaining. This manner of chaining starts from the goals towards the conditions.The flow chart represents the basis to code the knowledge base in a computer environment. Figure (2) illustrates a part of the sophisticated flow chart. VISUAL BASIC programming language is used to develop the system KBS-CCPRP since it is very effective and flexible tool for development of the programs working under WINDOWS environment (Michael Halvorson, 2008). The source code version of the developed system is a multi-forms package involves 162 forms (files) that connected together in a tree structure. Each form includes a number of commands responsible of executing specific functions in the system. The forms and the commands are named with clear expressive names that are related to the function of each command. In addition, many remarks are included in the coding menus to simplify updating process. This version (with extension of .vbp) is prepared to be used by the knowledge engineer who is responsible of developing and updating the system, an executable version (with extension of .exe) is prepared to be used by the end user who is the highway engineer, and this version is protected and unchangeable. Figure (3) illustrates the structure of KBS-CCPRP. Fig. 3: KBS-CCPRP Structure Operating and Testing of the System: The system can be operated easily through clicking a button to execute any step using the mouse or the keyboard;the user can return to the last step anytime since, each form is provided with back button, also he can use the button (Close) to end the running of the system at any stage. Figure (4a-i) illustrates an example of KBS-CCPRP operating. The system is tested in a real environment through exposing it to operating by a number of civil engineers with different levels of experience, all of them emphasised that the KBS-CCPRP can be operated simply and friendly and it is helpful for novice engineers to overcome the domain problems. Fig. (4a): System operating. Press button to continue 1132 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 Fig. (4b): Review the flow diagram and press button to continue Fig. (4c): Select an option and press button to continue Fig. (4d): Review the menu and press button to continue 1133 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 Fig. (4e): Select an option and press button to continue Fig. (4f): Review the description and press button to review the causes Fig. (4g): Review the causes and press button to continue 1134 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 Fig. (4h): Review the instantaneous solutions and press button to continue Fig. (4i): Review the preventive solutions and press back or close. System Updating: The system is coded by a simple manner to ensure easy updating by the developer or any other knowledge engineer in this domain who can deal with VISUAL BASIC language by feeding the system with new information continuously. Conclusions: KBS-CCPRP can be used by the novice engineers to overcome the problems in the field of rigid pavements construction when no senior engineers are available, this can save cost, time, and efforts, and can prevent accumulation of the problems in the site during waiting for the senior engineer. The system can be used as an instructional tool in this domain through the repeated use of it; also, it can be used as an archive to document the domain knowledge in a classified form. The system can be verified by the experts in the domain of rigid pavements and can be updated continuously by feeding it with new experiences. REFERENCES ACI 305.1-06, 2007 .Specification for Hot Weather Concreting. American Concrete Institute International. ACI 306.1-90, 2002 .Standard Specification for Cold Weather Concreting. American Concrete Institute International. ACI302.1R-04, 2004 .Guide for Concrete Floor and Slab Construction. American Concrete Institute International. ACI121R-04, 2004 .Quality Management System for Concrete Construction. American Concrete Institute International. 1135 Aust. J. Basic & Appl. Sci., 5(6): 1126-1136, 2011 ACI325.12R-02, 2002 .Guide for Design of Jointed Concrete Pavements for Streets and Local Roads. American Concrete Institute International. ACI305R-99, 1999 .Hot Weather Concreting. American Concrete Institute International. ACI224.1R-07, 2007 .Causes, Evaluation, and Repair of Cracks in Concrete Structures. American Concrete Institute International. ACI330R-01, 2001. Guide for Design and Construction of Concrete Parking Lots. American Concrete Institute International. ACI212.3R-04, 2004. Chemical Admixtures for Concrete. American Concrete Institute International. ACI302.1R-04, 2005. Guide For Concrete Highway Bridge Deck Construction. American Concrete Institute International. ACI304R-00, 2000. Guide for Measuring, Mixing, Transporting, and Placing Concrete. American Concrete Institute International. ACI311.4R-05, 2005. Guide for Concrete Inspection. American Concrete Institute International. ACI309R-05, 2005. Guide for Consolidation of Concrete. American Concrete Institute International. Frank McCullough, B. and Terry Dossey, 1999 .Considerations for high Performance Concrete Paving. Journal of Transportation Research Board, 1684: 17-24. Ken Newman, 1986 .Common Quality inConcrete Construction. Journal of Concrete International, The Magazine of the American Concrete Institute (ACI), 8(3): 37-49. Innovative Pavement Research Foundation, 2005. Stabilized and Drainable Base for Rigid Pavement a Design and Construction Guide. Report IPRF-01-G-002-02-1(G). Michael Halvorson, 2008. Microsoft Visual Basic Step by Step. Redmond Washington. O’Flaharty, C A, 1988. Highway Engineering. Vol. 2, Third Edition, Hodder and Stoughton Limited, London, Britain. Robert, G. Hennes and Martin Ekse, 1990. Fundamentals of Transportation Engineering.Second Edition, McGraw-Hill Book Company, U.S.A. Standard Specification for Urban Infrastructure, 2002 .Rigid Pavement Construction. Works 5-1 Edition 1, Revision0. Stock, A.F., 1988 .Concrete Pavements. Elsevier applied science, London. Walter, H., Price, 1982. Control of Cracking of Concrete during Construction. Journal of Concrete International, 4(1): 40-43. 1136 
work_cfxf43jr2zgf3jrpnztzb4jqb4	----	Soft computing applications in the field of industrial and environmental enterprises Received: 14 June 2019 Accepted: 21 June 2019 DOI: 10.1111/exsy.12456 G U E S T E D I T O R I A L Soft computing applications in the field of industrial and environmental enterprises 1 | INTRODUCTION Artificial Intelligence (AI) has proved to be a promising field of study that can greatly contribute to the development of computational systems that successfully address an increasing number of challenges. From the very beginning, when the term was coined in 1956 during the Dartmouth Sum- mer Research Project, researchers from this field have shared a vision that computers can be made to perform intelligent tasks (Moor, 2006). A vast array of different techniques has been proposed so far to implement such intelligent systems, ranging from expert or knowledge‐based sys- tems (some of the earlier products of AI) to cutting‐edge proposals such us deep and machine learning. Among all the branches of the AI tree, in last decades, there has been significant progress in soft computing (Karray & De Silva, 2004). As pointed out by Zadeh (1994), while hard computing focuses on precision, certainty, and rigour, soft computing requires that computation, reason- ing, and decision making exploit a tolerance for imprecision and uncertainty wherever possible (Zadeh, 1994). Soft computing can be seen as a family consisting of many members, including evolutionary computation, fuzzy logic, and connectionist models among others (Pratihar, 2007). Its fascinating ability to simulate human intelligence has made it favourable in various scientific and technological disciplines (Chakraverty, Sahoo, & Mahato, 2019) such as computer science, mathematics, control theory, structural engineering, medical, and psychology. One of the domains where soft computing can arguably be of greatest utility is in the field of business. The current environment for businesses is increasingly complex and demanding. Companies are affected by multiple internal and external variables that have an impact on the various areas that make up the value chain, and this is exacerbated for those managing operations in multiple countries (i.e., exports, imports, and foreign direct investments). Decision takers within the companies can significantly benefit from and demand tools to analyse and extract the most knowl- edge from the large datasets generated by the interrelated activities that companies conduct. Soft computing can be of particular value for industrial and environmental companies. Compared to other sectors such as retailing, where indi- vidual‐level determinants are key to understand the buyer, industrial companies are characterized by a more critical role of operations efficiency in order to succeed in the market. Likewise, given the large volume of information that is now available for environmental firms, competition in this field is also more and more determined by a strategic and sophisticated use of data analytics. Overall, soft computing represents a potentially key source of competitive advantage sustainable over time that organizations, especially in the industrial and environmental field, can employ to outperform competitors, achieve higher returns, and ensure their own viability (Wernerfelt, 1984; Barney, 1991). By complementing broad analytical tools (such as Porter's five forces PESTLE or Blue Ocean; Porter, 1979; Kim & Mauborgne, 2014), soft computing may allow firms to materialize tangible resources from strategic advantages. This special issue compiles recent applications of soft computing techniques to the management of enterprise within industrial and environ- mental sectors. It is aimed at both researchers and practitioners from academia and industry who are engaged in deploying soft computing solu- tions for real‐life enterprise problems. Five papers are included in this special issue, covering a wide variety of case studies, ranging from photovoltaic solar systems to forest biomass estimation. For each different soft computing, techniques have been applied, namely, unsupervised neural networks, deep learning, regression models, multiagent systems, and case‐based reasoning. Thanks to the wide‐range panoramic view pre- sented by these complementary works, readers can get a feel for developing up‐to‐date soft computing systems to solve present problems in enterprise contexts. 2 | CONTENTS OF THE SPECIAL ISSUE In the paper by Jove, Casteleiro‐Roca, Quintián, Méndez‐Pérez, and Calvo‐Rolle (2019), a novel approach for fault detection in industrial pro- cesses, by means of unsupervised and projectionist techniques, is proposed. The proposed system is tested in a laboratory plant built with indus- trial equipment where anomalies have to be detected without any previous knowledge of the data. This system was designed to control the level Expert Systems. 2019;36:e12456. https://doi.org/10.1111/exsy.12456 © 2019 John Wiley & Sons, Ltdwileyonlinelibrary.com/journal/exsy 1 of 4 https://doi.org/10.1111/exsy.12456 https://doi.org/10.1111/exsy.12456 http://wileyonlinelibrary.com/journal/exsy http://crossmark.crossref.org/dialog/?doi=10.1111%2Fexsy.12456&domain=pdf&date_stamp=2019-08-08 2 of 4 GUEST EDITORIAL of liquid in a tank where the liquid is initially stored in a different tank placed at a lower level, and it is boosted by a three‐phase pump driven by a variable frequency driver. The objective tank has also two built‐in output valves; one of them is a proportional electric, and the other one is man- ual. In such a system, considered failures are water leaks from the main water tank to the lower tank through a valve. The correct operation of the plant meant that the output electrovalve had to be completely closed. The feasibility of the proposed method was checked by opening the electrovalve and simulating a water leak. In order to address such a problem, dimensionality reduction techniques are applied to support the visualization of the operation point in a two‐dimensional graph regardless of the number of variables. More precisely, data are analysed by means of Principal Component Analysis (PCA), Maximum Likelihood Hebbian Learning (MLHL), Beta Hebbian Learning algorithm (BHL), curvilinear Component Analysis (CCA), and ISOMAP. As the industrial plants usually work in one or a few working points, all the visualized data in an operation point should be near to each other. The data are projected into two‐dimension graph, and the user could define a contour in the data. With the defined contour, the proposed algorithm detects whether the working point is out of the previously defined working area. To validate the proposed system, two datasets with different levels of complexity (in terms of quantity and quality of information) have been used. The obtained results have been satisfactory, although the accuracy varies depending on the employed technique and the complexity of the dataset (the higher the complexity of the data, the less accurate the results are). Torres, Troncoso, Koprinska, Wang, and Martínez‐Álvarez (2019) propose a Deep Learning approach, based on feed‐forward neural net- works for big data time series, for predicting the electricity power generated by photovoltaic solar systems for the next day (at half‐hourly intervals). As it is well known, solar energy is highly variable as it depends on meteorological conditions (solar radiation, cloud cover, rainfall, and temperature). This dependency creates uncertainty about the amount of solar power that will be generated, which makes it difficult to integrate solar power into the electricity grid and electricity markets. Hence, an accurate prediction of the generated solar power is a task of utmost importance and relevance for both energy managers and electricity traders, in order to minimize uncertainty and ensure reliable electricity supply at acceptable cost. Day ahead predictions are one of the most common industry‐requested operational forecasts, as they are needed for operational planning, switching sources, programming backups, short‐term power purchases, and for planning of reserve usage and peak load matching. To perform such a short‐term forecast of big solar‐power time series data, a solution based on Deep Learning (DL) feed‐forward neural networks is conceived. To find a good combination of hyper‐parameters of the applied neural network, the H20 grid search method is applied. The proposed DL solution is compared to the Pattern Sequence‐based Forecasting (PSF) algorithm, which uses clustering and similarity of pat- terns, and a neural network‐based model with one hidden layer, used as a reference method. All these models are applied to Australian solar power data gathered during 2 years and compared in terms of the Mean Absolute Error and the Root Mean Squared Error. The accuracy of the applied methods is compared when using weather and weather forecast data as an additional input, taking into account different scenarios corresponding to different percentages of noise in the weather forecast data. Finally, the effect of varying the size of the historical window is also studied. In peer‐to‐peer lending, it is important to predict the repayment of the borrower to reduce the lender's financial loss. Such a topic is addressed in the paper by J.‐Y. Kim and Cho (2019) that proposes a deep dense convolutional network for repayment prediction in social lending. It maintains the borrower's semantic information and obtains a good representation by automatically extracting important low‐ and high‐level features simul- taneously. The proposed model classifies the loan status of a borrower by learning the pattern while maintaining and extracting discriminative fea- tures according to the loan status and improves the overall performance with a good representation. Authors assume that it would be easier to predict the loan status in the feature space if the network captures the inherent properties of the lending data of a borrower and generalizes them to other borrowers. Consequently, the feature space is learned by a DenseNet network that con- sists of several dense blocks and transition layers, in order to obtain a good representation from the social lending data of many borrowers. The lending data are then projected onto the learned feature space using the trained network and are predicted using the softmax classifier. Extensive experiments are conducted on data coming from the U.S. Lending Club. Torre‐Tojal, Lopez‐Guede, and Graña Romay (2019) propose the application of machine‐learning‐based predictive systems for the extrac- tion of biomass information from Light Detection and Ranging (LiDAR) data. LiDAR is a remote sensing tool that retrieves surface elevation measurements at high spatial resolutions, even in rough terrains and heavily forested areas. It is based on the emission of pulses by an infrared laser sensor emitting pulses and recording of the echoes returning to the sensor after being reflected by ground surfaces. Authors of this paper focus on the application of LiDAR for forest biomass estimation, what may have economic value owing to the use of biomass as a source of energy source or as the raw material for a wide variety of elaborate products such as construction materials, furniture, and paper. In order to predict the forest biomass from LiDAR data, three predictive machine learning approaches are applied, namely, Multiple Linear Regression, Random Forest, and Support Vector Regression. To test such models, they have been applied (under a fivefold cross‐validation schema) to the Pinus radiata species using public data gathered from a LiDAR flight conducted by the Basque Government (Spain) in 2012 and the dendometric data coming from the Fourth National Forestal Inventory that were gathered in 2011. The modelling performance results obtained in this study are comparable with those in other studies concerning plot‐level biomass estimations. GUEST EDITORIAL 3 of 4 Finally, a job‐offer recommender system for careers is proposed by Rivas, Chamoso, González‐Briones, Casado‐Vara, and Corchado (2019). It is a hybrid system whose architecture is based on a multiagent system that consists of (a) a Case‐Based Reasoning (CBR) system powered by a series of metrics that calculate the affinity between job offers and users and (b) an argumentation framework that extends the functionality of the CBR, thus making it possible for agents to have a dialogue. Individual solutions proposed by each one of the agents are first debated and settled before a single final suggestion is obtained. The proposed hybrid intelligent system is able to carry out analyses of user profiles and job offers found on the beBee social network. Con- sequently, it has been evaluated on a sample set of such users actively searching for employment in periods of 15 days during 4 months. As a result of the activity, a total of 118,882 recommendations have been made. The acceptance rate average of users who did interact with the pro- posed system is 75.81% out of the 42,478 user applications. There has been an evolution of the analysed rates during the 4 months in which the evaluation has been performed; the users' acceptance rate has increased from 73.72% to 78.6%, while the rejection rate has decrease from 26.28% to 21.4% in the last month. On the other hand, the interaction of companies has been also studied; their acceptance rate has increased from 21.74% to 24.03% during the 4 months. ACKNOWLEDGEMENTS The guest editors wish to thank Dr. Jon G. Hall (Editor‐in‐Chief of the Wiley‐Blackwell Journal Expert Systems: The Journal of Knowledge Engi- neering) for providing the opportunity to edit this special issue. We would also like to thank the referees who have thoroughly evaluated the papers and the editorial staff for their support. CONFLICT OF INTEREST None. ORCID Alfredo Jimenez https://orcid.org/0000-0001-7811-5113 Álvaro Herrero https://orcid.org/0000-0002-2444-5384 Alfredo Jimenez1 Álvaro Herrero2 1Department of Management, KEDGE Business School, Bordeaux, France 2Grupo de Inteligencia Computacional Aplicada (GICAP), Departamento de Ingeniería Civil, Escuela Politécnica Superior, Universidad de Burgos, Burgos, Spain Correspondence Álvaro Herrero, Grupo de Inteligencia Computacional Aplicada (GICAP), Departamento de Ingeniería Civil, Escuela Politécnica Superior, Universidad de Burgos, Av. Cantabria s/n, Burgos 09006, Spain. Email: ahcosio@ubu.es REFERENCES Barney, J. (1991). Firm resources and sustained competitive advantage. Journal of Management, 17(1), 99–120. https://doi.org/10.1177/ 014920639101700108 Chakraverty, S., Sahoo, D. M., & Mahato, N. R. (2019). Concepts of soft computing: Fuzzy and ANN with programming. Singapore: Springer. https://doi.org/ 10.1007/978‐981‐13‐7430‐2 Jove, E., Casteleiro‐Roca, J.‐L., Quintián, H., Méndez‐Pérez, J. A., & Calvo‐Rolle, J. L. (2019). A fault detection system based on unsupervised techniques for industrial control loops. Expert Systems, e12395. Karray, F. O., & De Silva, C. W. (2004). Soft computing and intelligent systems design: Theory, tools, and applications. New York: Pearson Education. Kim, J.‐Y., & Cho, S.‐B. (2019). Predicting repayment of borrows in peer‐to‐peer social lending with deep dense convolutional network. Expert Systems, e12403. Kim, W. C., & Mauborgne, R. A. (2014). Blue ocean strategy, expanded edition: How to create uncontested market space and make the competition irrelevant. Boston, Massachusetts: Harvard business review Press. Moor, J. (2006). The Dartmouth College artificial intelligence conference: The next fifty years. AI Magazine, 27(4), 87–87. Porter, M. E. (1979). The structure within industries and companies' performance. The Review of Economics and Statistics, 61(2), 214–227. https://doi.org/ 10.2307/1924589 Pratihar, D. K. (2007). Soft computing. Oxford, UK: Alpha Science International, Ltd. Rivas, A., Chamoso, P., González‐Briones, A., Casado‐Vara, R., & Corchado, J. M. (2019). Hybrid job offer recommender system in a social network. Expert Systems, e12416. https://orcid.org/0000-0001-7811-5113 https://orcid.org/0000-0002-2444-5384 https://orcid.org/0000-0001-7811-5113 https://orcid.org/0000-0002-2444-5384 https://doi.org/10.1177/014920639101700108 https://doi.org/10.1177/014920639101700108 https://doi.org/10.1007/978-981-13-7430-2 https://doi.org/10.1007/978-981-13-7430-2 https://doi.org/10.2307/1924589 https://doi.org/10.2307/1924589 4 of 4 GUEST EDITORIAL Torres, J. F., Troncoso, A., Koprinska, I., Wang, Z., & Martínez‐Álvarez, F. (2019). Big data solar power forecasting based on deep learning and multiple data sources. Expert Systems, e12394. Torre‐Tojal, L., Lopez‐Guede, J. M., & Graña Romay, M. M. (2019). Estimation of forest biomass from light detection and ranging data by using machine learning. Expert Systems, e12399. Wernerfelt, B. (1984). A resource‐based view of the firm. Strategic Management Journal, 5(2), 171–180. https://doi.org/10.1002/smj.4250050207 Zadeh, L. A. (1994). Fuzzy logic, neural networks, and soft computing. Communications of the ACM, 37(3), 77–85. https://doi.org/10.1145/175247.175255 https://doi.org/10.1002/smj.4250050207 https://doi.org/10.1145/175247.175255 
work_chipkkw5ijhnbi3ca6hpz7jozm	----	Perception granular computing in visual haze-free task Expert Systems with Applications 41 (2014) 2729–2741 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Perception granular computing in visual haze-free task 0957-4174/$ - see front matter Crown Copyright � 2013 Published by Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.11.006 ⇑ Corresponding author. Tel.: +86 01062600505. E-mail addresses: huhong@ict.ac.cn (H. Hu), pangl@ics.ict.ac.cn (L. Pang), tiandp@ics.ict.ac.cn (D. Tian), shizz@ics.ict.ac.cn (Z. Shi). Hong Hu, Liang Pang ⇑, Dongping Tian, Zhongzhi Shi Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Science, Beijing 100080, China a r t i c l e i n f o a b s t r a c t Keywords: Granular computing Leveled granular system Fuzzy logic Machine learning Haze free Brain-like computer In the past decade, granular computing (GrC) has been an active topic of research in machine learning and computer vision. However, the granularity division is itself an open and complex problem. Deep learning, at the same time, has been proposed by Geoffrey Hinton, which simulates the hierarchical structure of human brain, processes data from lower level to higher level and gradually composes more and more semantic concepts. The information similarity, proximity and functionality constitute the key points in the original insight of granular computing proposed by Zadeh. Many GrC researches are based on the equivalence relation or the more general tolerance relation, either of which can be described by some dis- tance functions. The information similarity and proximity depended on the samples distribution can be easily described by the fuzzy logic. From this point of view, GrC can be considered as a set of fuzzy logical formulas, which is geometrically defined as a layered framework in a multi-scale granular system. The necessity of such kind multi-scale layered granular system can be supported by the columnar organiza- tion of the neocortex. So the granular system proposed in this paper can be viewed as a new explanation of deep learning that simulates the hierarchical structure of human brain. In view of this, a novel learning approach, which combines fuzzy logical designing with machine learning, is proposed in this paper to construct a GrC system to explore a novel direction for deep learning. Unlike those previous works on the theoretical framework of GrC, our granular system is abstracted from brain science and information science, so it can be used to guide the research of image processing and pattern recognition. Finally, we take the task of haze-free as an example to demonstrate that our multi-scale GrC has high ability to increase the texture information entropy and improve the effect of haze-removing. Crown Copyright � 2013 Published by Elsevier Ltd. All rights reserved. 1. Introduction Lin (2012) pointed out that Granulation seems to be a natural methodology deeply rooted in human thinking. Many daily things are routinely granulated into sub-things (Lin, 2012). In the IEEE- GrC2006 conference of information about the granular computing (GrC), the outline of GrC is defined as a general computation theory for effectively using granules such as classes, clusters, subsets, groups and intervals to build an efficient computational model for complex applications with huge amounts of data, information and knowledge (Zadeh, 1997). Just as the scholars summarized in the IEEE-GrC2006 conference though the label is relatively recent, the basic notions and principles of GrC, though under different names, have appeared in many related fields, such as information hiding in programming, granularity in artificial intelligence, divide and conquer in theoretical computer science, interval computing, cluster analysis, fuzzy and rough set theories, neutrosophic com- puting, quotient space theory, belief functions, machine learning, databases and many others (Bargiela & Pedrycz, 2006). The above definition of GrC is too augmental and the subjects about classes, clusters, subsets, groups and intervals have already studied by arti- ficial intelligence and mathematics for a long time. What is really new point for GrC? We think that the new or main point of the GrC lies in the original insight of GrC proposed by Zadeh, in which there are three basic concepts that underlie human cognition: granulation, organization and causation. Informally, granulation involves decomposition of whole into parts; organization involves integration of parts into whole; and causation involves association of causes with effects. Granulation of an object A leads to a collection of granules of A, with a granule being a clump of points (objects) drawn together by indistinguishability, similarity, proximity or functionality (Zadeh, 1997). In this original insight of GrC, Zadel pointed out three important aspects about GrC: (1) the GrC is a main character of human cognition, (2) so called GrC is based on indistinguishability, similarity, proximity or functional- ity, (3) there is a close relationship among granulation, organiza- tion and causation. Based on these points, we think that it is necessary to find some key points of GrC in human cognition. There are two kinds of GrC research: perception-level and knowledge-level. A perception-level GrC does a series feature transformation and tries to find meta-knowledge implied in sam- ples; a knowledge GrC tries to process knowledge or structure http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2013.11.006&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.11.006 mailto:huhong@ict.ac.cn mailto:pangl@ics.ict.ac.cn mailto:tiandp@ics.ict.ac.cn mailto:shizz@ics.ict.ac.cn http://dx.doi.org/10.1016/j.eswa.2013.11.006 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 2730 H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 information based on metal-knowledge. In this paper we focus on the perception-level. Indistinguishability, similarity and proximity can be described by equivalence relation or tolerance relation, and these relations can be described by some kind distance functions. From the rele- vant literature, it is easy to see that many GrC researches focus on classification and clustering (Yao, 2000, 2001; Zhang & Zhang, 2003, 2004a). Zhang and Zhang (2003, 2004a, 2004b, 2005) use the quotient space theory to try to study indistinguishability and similarity. Yao (2001) extends the equivalent class to rough approximation set. The quotient space structure described by equivalence relation is used to probe the structure of granules such as classes, clusters, subsets, groups etc. In a more general way, Lin (1998, 2007) and Yao (1998, 1999) use binary relations and neigh- borhood systems to study indistinguishability and similarity respectively, the geometric concepts: partitions, covering and topology, and neighborhood can be described by binary relations in the algebra. Pedrycz, Hirota, Pedrycz, and Dong (2012) define granular on fuzzy sets and discuss several operations and their granular consistency (Pedrycz et al., 2012). Lin (2012) gives out a summary about the history of granular computing, he discusses all formal description of granular computing and some further directions e.g. GrC, databases and data mining, GrC and clouding computing etc. In fact, GrC should be discussed in the framework of human cognition from perception to pattern recognition and knowledge processing. Although the concept of granular comput- ing has been proposed more than ten years, only few people pay attention to this subject. In fact granular computing pays much more attention to the leveled computing of intelligence. Just as Yao (2006) pointed out: ‘‘Granules in the family are called focal elements of discussion at the level. Each level is represented by a plane. While granules at the same level are of similar nature, gran- ules at different levels may be very different. Consequently, we may use different vocabularies and languages for descriptions at different levels.’’ The leveled computation revealed by granular computing is very important for machine learning, e.g. the famous approach deep learning. The term deep learning gained much attraction in the mid-2000s after a publication by Bargiela and Pedrycz (2006), Castro (1995). Nowadays it becomes a huge wave of technology trend for big data and artificial intelligence. Deep learning simulates the hierarchical structure of human brain, pro- cesses data from lower level to higher level, and gradually com- poses more and more semantic concepts. These facts mean that deep learning has a close relationship with the granular computing. All the granular computing researches aforementioned, in gen- eral, neglect information transformation and feature abstraction, which are very important for deep learning. In this paper, we pro- pose a novel framework of granular system, which has ability to process information transformation and object-background sepa- ration. We take the haze-free task as an example to validate the ability of our granular system. These facts mean that the deep learning has a close relationship with the granular computing proposed by us. The main contributions of this paper include: (1) The basic notions and principles of GrC termed with differ- ent names, but they have appeared in many related fields, such as information hiding in programming, granularity in artificial intelligence, divide and conquer in theoretical com- puter science, interval computing, cluster analysis, fuzzy and rough set theories, neurotrophic computing, quotient space theory, belief functions, machine learning, databases, and many others, so old version of Granular computing is just an abstraction of old methods, in this paper we give a novel concrete model for granular computing which has a multi scale layered structure from feature abstraction to classification. (2) Our novel granular system has a close relation with deep learning, so it develops a new focus for deep learning. It is the first time that fuzzy logic is introduced for leveled fea- ture abstraction in deep learning. (3) Although fussy logic is often mentioned in granular comput- ing, for example,fuzzy logic and rough set technique are used by Lin (1999) for word computing, and Liu, Xiong, and Wu (2012) use fuzzy lattices in the classification based on hyperspherical granular computing (Liu et al., 2012). In this paper, fuzzy logic is not only used for describing granu- lar similar to hyperspherical granular in Liu et al. (2012), but also for feature abstraction and classification. For this pur- pose, we propose a novel and effective approach which is combined fuzzy logical designing, PSVM and back propagation. (4) The granular computing proposed by us gives a novel approach for the task of haze-free, the experiments’ result show that this approach is sound. The rest of this paper is organized as follows: In Section 2,we try to give out a formal definition of granular system and granular computing based on the a tolerance relation, which is described by fuzzy logical formula; in Section 3, we discuss the algorithm to design a granular system; in Section 4, we give out a concrete example of designing a granular system for haze-free task; at last Section 5 is the discussion and looking forward to the future. 2. Granular system based on tolerance relation The difference between a granular system based on equivalence relation and a granular system based on tolerance relation is that an equivalence relation will divide a space into nonoverlapping covering while a tolerance relation will create overlapping cover- ing of this space. Yiyu Yao proposed granular computing paradigm for concept learning in which two learning strategies are investigated. A global attribute-oriented strategy searches for a good partition of a uni- verse of objects and a local attribute–value-oriented strategy searches for a good covering (Yao & Deng, 2013). In this paper, granular computing is started from feature vectors e.g. images, not attributes. In order to simulation perception procession of our cognition, we define a set of multi-scale nested convex regions with a corresponding computing based on this set. There are two main purposes to build such a granular system based on tolerance relation: (1) Granular systems are designed to describe similarity and proximity of information, which can be described by toler- ance relation. Granular systems based on tolerance relation can be viewed as a topological structure built by topological bases on the topological space ðX; sÞ induced from a metric space ðX; disÞ by the metric dis. Granular systems based on tolerance relation can be used to describe domain stricture, which represents indistinguishability, similarity and prox- imity of examples.Classification is determined by the indis- tinguishability, similarity and proximity of information. There are two kind similarity among examples-static simi- larity and dynamical similarity. (a) If elements of classes are distributed in standard convex regions, we can use some kind distance function to describe classes distribution domains. In this case, simi- larity between two objects can be intuitively described by distance functions. If disðx; yÞ is a distance function H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 2731 in the n-dimensional space Rn and c is a point in Rn , the formula disðc; yÞ < r described a convex region D in Rn which takes c as its center. Every point y in this region is equal to y ¼ c þ e, here e is some kind noise, if D is just a ball, e can be viewed as white noise which has an amplitude less than r.We denote such kind similarity as static similarity. (b) The dynamical similarity is different from static similar- ity, if one object O1 continuously changes to another object O2, e.g. a tadpole continuously grow up to a frog, then O1 and O2 are dynamical similar, i.e. if all elements in a class A are dynamical similar, the distribution domain of A is a connected domain. Dynamical similarity will cause distribution domain become very complicate and have a nonlinear borderline. Although,a dynamical similarity may cause the inner class difference larger than the among classes difference, it can also be described by equivalence relation or tolerance relation. The difference of a granular system based equivalence relation and granular system based on tolerance relation is that an equivalence relation will divide a space into non overlapping covering and a tolerance relation will create overlapping covering of this space. The relation described by the formula disðx; yÞ < r is the special case of a tolerance relation. (2) The main purpose of information transformation in pat- tern recognition is to recognize or classify different objects from their mixture, so information transformation used in pattern recognition should be taken place in a granular system which describes static similarity and dynamical similarity. This is the main point we will dis- cuss in this paper. Now we try to use fuzzy logical formula based on distance func- tion to define granular systems. There are three distance axioms: ð1Þdisða;aÞ¼0;ð2Þdisða;bÞ¼disðb;aÞ;ð3Þdisða;bÞþdisðb;cÞP disða;cÞ. disðx;yÞ<r defines a tolerance relation and disðx;yÞ¼0 defines an equivalent relation, and every tolerance relation can be viewed as an abstract distance which does not obey ðd3Þ, so any granular systems based on equivalent relation and tolerance relation can be described by distance function in a geometrical way. Definition 1 (A simple fuzzy logical formula based on distance function). A simple fuzzy logical formula based on distance func- tion spða; cjdis; d; xÞ is denoted as: spða; cjdis; d; xÞ¼ mðdisða; cjxÞ; dÞ ð1Þ where disða; cjxÞ¼ disða � x; c � xÞ, x � x ¼ðx1x1; . . . ; xnxnÞ, and disðÞ is point to point distance function, and x is a weight vector for dimensions of feature vector space S; mðÞ can be viewed as a membership function of a fuzzy set, it is usually a continuous func- tion. For example mðdisða; cjxÞ; dÞ¼ ðd � disða; cjxÞÞ=d; disða; cjxÞ < d 0; disða; cjxÞ P d � ð2Þ where the distance function can be a set to set distance function, e.g. Hausdorff function. For the point to point distance function case, we define the r-cut set of spða; cjdis; d; xÞ as. sprða; cjdis; d; xÞspða; cjdis; d; xÞ > r (strong r-cut set) or sprða; cjdis; d; xÞ¼ spða; cjdis; d; xÞ P r (r-cut set). sprðx; cjdis; d; xÞ defines an open convex region and denoted as a granule, and sprðx; cjdis; d; xÞ defines a closed convex region. Definition 2 (leveled perception granular system based on tolerance relation Gsys). A granular system based on tolerance relations of distance function (granular system for short) is a set of granules, every granule is a 2-tuple fg; SFg, here g is a convex region which is described by tolerance relations of distance function, and SF is a set of fuzzy logical functions (denoted as ‘‘adjoint functions’’) which are computed from the convex region g, the outputs of all fuzzy logical functions in SF are denoted as ‘‘an adjoint vector’’ of this granule. The granules of Gsys have the following attributions. (1) Multi-scale leveled structure: The metric space X (e.g. a finite connected volume region in the n-dimensional real space Rn) is the only level 0 granule, the level 0 granule is denoted as GðcoeG0Þ¼ fX; SFg, where coeG0 is a coefficient set, coeG0 is usually empty, and the function SF is a set of fuzzy logical functions (denoted as adjoint functions). The convex region of a level 1 granule GðcoeG1Þ¼ fg1; SF1g is defined by the conjunction of finite number r-cut sets spr a; c11jdis 1 ; d11; x � � ; spr a; c12jdis 1 ; d12; x � � ; . . . ; spr a; c1k1jdis 1 ; � d1k1 ; xÞ, where coeG 1 is its coefficient set to define the convex region g1 , so g1 can also be written as g1ðcoeG1Þ and coeG1 ¼ c11; c12; . . . ; c1k1 2 X; d 1 1; d 1 2; . . . ; d 1 k1 ; dis1 n o . The first level of our granular system is denoted as C1ðXÞ¼ fGðcoeG1Þg. If the l level granules GðcoeGlÞ¼ fgl; SFlg have been defined, the l þ 1 level granules can be defined as GðcoeGlþ1Þ, which can be defined as the convex region created by the intersection of gl and finite number strong r-cut sets: spr a; clþ11 jdis lþ1 ; dlþ11 ; x � � ; spr a; clþ12 jdis lþ1 ; dlþ12 ; x � � ; � � � ; spr a; clþ1klþ1 � jdislþ1; dlþ1klþ1 ; xÞ. A level l þ 1 granule’s coefficient is coeGlþ1 ¼ fclþ11 ; c lþ1 2 ; . . . ; c lþ1 klþ1 2 X; dlþ11 ; d lþ1 2 ; . . . d lþ1 klþ1 ; dislþ1g, where klþ1 is denoted as the number of simple logical formulas in same level. The centerofGðcoeGlþ1Þ is clþ1i 2 GðcoeG lÞ, and its radius is dlþ1i 6 d l i; lim l !1d l i ¼ 0. The GðcoeG 1Þ is called as the ‘‘the father granule of GðcoeGlþ1Þ ’’. (2) Granular computing (GrC): A granular computing is described by a set of fuzzy logical formula upon above multi scale leveled structure of convex regions. The purpose of a granular computing (GrC) is to transfer fea- ture information and classy points in an input space X, so at least one level of a granular computing (GrC) outputs a fuzzy label for points in an input space X. If there are totally m classes, a fuzzy label L is a m dimensional fuzzy vector L ¼fl1; l2; . . . ; lmg; and P i¼1;...;mli ¼ 1. Castro (1995) proved that Fuzzy logic controllers using fuzzy rules are universal approximations, later, Li and Philip Chen (2000) shows a proof of the equality between a forward neu- ral circuit (or circuit) and a fuzzy logical inference. So it is not difficulty to prove that any continuous functions F : Rn�!½0; 1�n can be simulated by such kind nested layered granular computing with arbitrary small errors. A level l þ 1 adjoint function F lþ1 receives its input from the outputs of level l þ p; p > 1 adjoint functions, i.e. if a leveled granular system Gsys has k levels, GrC is taken from level k to 1, so the level of a GrC is upside down with the level of the Gsys. The 1st level GrC takes place in the smallest kth level granules’ convex regions of the Gsys. Two kinds layered computing can be taken place over a granular system. In the first kind layered computing,the adjoint feature vectors of larger scale level n granules are computed based on the 2732 H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 adjoint feature vectors of smaller scale n þ 1 level gran- ules,such kind layered computing has strictly nested struc- ture, (Fig. 1(a)), and is denoted as ‘‘nested layered computing’’. In the second kind layered computing, adjoint feature vectors of level n granules can be computed based on all adjoint feature vectors of smaller granules, which have level greater than n (Fig. 1(b)), such kind layered computing is denoted as ‘‘unnested layered computing’’. Nested layered computing is a special case of unnested layered computing. (3) Radiuses of convex regions: A granular system can have countable infinite or finite levels. The radiuses of gran- ules’convex regions decrease and tend to zero when the level goes to infinite. (4) Centers’ grid: The centers of granules will distribute in a so called center grid, we call the set of all centers of granules of level l þ 1 on the granule GðcoeGlÞ as the center grid of level l granule GðcoeGlÞ denoted as Gclþ1ðGðcoeGlÞÞ . We denote the set of all centers of level l þ 1 granules over X as Gclþ1ðXÞ and all centers of level l þ 1 granules over a level k < l granule GðcoeGkÞ as Gclþ1ðGðcoeGkÞÞ. The center grid is usually discrete, but it can also be a contin- uous set e.g. the whole metric space X. (5) Shape of granules: the shapes of granules is defined by their distance functions disðÞ . If disðÞ can be an abstract distance function, then a granule’s convex region can be an arbitrary convex region. Every level uses same distance function, so the granules in the same level have the same shape, but for different levels, granules’region may have different shapes. (6) The cover over a granule: In order to create a cover over GðcoeGlÞ, the elements in the centers’ grid Gclþ1 ðGðcoeGlÞÞ should be tight enough. Such kind cover can be formerly defined as: Clþ1ðGðcoeGlÞÞ¼ ^ 16i6klþ1 sprðx; clþ1i jdis lþ1 ; dlþ1i ; xÞ^ gðcoeG lÞjclþ1i 2 Gc lþ1ðGðcoeGlÞÞ; sprðx; clþ1i jdislþ1; d lþ1 i ; xÞ^ gðcoeG lÞ – / � � All level l þ 1 granules create a cover of the whole space X, and denoted as the level l þ 1 cover or the level l þ 1 layer of X, Clþ1ðXÞ¼ [ GðcoeGlÞ2ClðXÞ fClþ1ðGðcoeGlÞÞg . Radial Basis neural network (Haykin, 2008), which can be used to simulate continuous functions, is an example of two layers gran- ule system. Definition 3 (Hyper-granules and mini-granules). All level n þ 1 granules GðcoeGnþ1Þ, which are contained in a level n granule GðcoeGnÞ, denoted as ‘‘mini-granules’’, and the level n granule GðcoeGnÞ is denoted as a ‘‘hyper-granule’’. Fig. 1. Two kind layered compu After the theory of fuzzy logic was conceived by Zadeh (1965), many fuzzy logical systems have been presented, for example, the Zadeh system, the probability system, the algebraic system, and Bounded operator system, etc. According to universal approx- imation theorem (Haykin, 1994), in this paper, the extended Bounded operator is selected, which is denoted as ‘‘q-value Weighted Bounded operator’’. It is not difficult to prove that q-value weighted fuzzy logical formulas can precisely simulate any continuous functions F : Rn�!½0; 1�n with arbitrary small error, or vice versa, i.e. every GrC can be completed by a set of fuzzy logical functions of q-value weighted bounded operator with arbitrary small error. Definition 4 (Bounded Operator Fð�f ;�fÞ). Bounded product: x�f y ¼ maxð0; x þ y � 1Þ, and Bounded sum: x�f y ¼ minð1; x þ yÞ, where 0 6 x; y 6 1 . In order to simulate GrC, it is necessary to extend the Bounded Operator to Weighted Bounded Operator. The fuzzy formulas de- fined by q-value weighted bounded operators is denoted as q-value weighted fuzzy logical functions. Definition 5 (q-value Weighted Bounded operator Fð�f ;�fÞ). q-value Weighted Bounded product: p1�f p2 ¼ F�f ðp1; p2; w1; w2Þ ¼ maxð0; w1p1 þ w2p2 �ðw1 þ w2 � 1ÞqÞ ð3Þ q-value Weighted Bounded sum: p1�f p2 ¼ F�f ðp1; q2; w1; w2Þ¼ minðq; w1p1 þ w2p2Þ ð4Þ where 0 6 p1; p2 6 q. For association and distribution rules, we define: ðp1Df p2ÞHf p3 ¼FHf ðFDf ðp1;p2;w1;w2Þ;p3;1;w3Þ and p1Dfðp2Hf p3Þ¼ FDf ðp1; FHf ðp2; p3; w2; w3Þ; w1; 1Þ, Here Df ; Hf ¼�f or �f . We can prove that �f and �f follow the associative condition (see Appendix C) and x1�f x2�f x3 . . .�f xn ¼ min q; X 16i6n wixi ! ð5Þ x1�f x2�f x3 . . .�f xn ¼ max 0; X 16i6n wi xi � X 16i6n wi � 1 ! q ! ð6Þ For more above q -value weighted bounded operator Fð�f ;�fÞ fol- lows the Demorgan Law, i.e. ting over granular system. H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 2733 Nðx1�f x2�f x3 . . .�f xnÞ¼ q � min q; X 16i6n wi xi ! ¼ max 0; q � X 16i6n wi xi ! ¼ max 0; X 16i6n wiðq � xiÞ�ð X 16i6n wi � 1Þq ! ¼ Nðx1Þ�f Nðx2Þ�f Nðx3Þ � � ��f NðxnÞ: ð7Þ But for the q-value weighted bounded operator Fð�f ;�fÞ, the distribution condition is usually not hold, and the boundary condi- tion is hold only all weights equal to 1, for p1�f q ¼ F�f ðp1; q; w1;w2Þ¼maxð0;w1 p1 þð1�w1ÞqÞ and p1�f q ¼ F�f ðp1; q; w1; w2Þ¼ minðq; w1 p1 þ w2 qÞ. In this paper, we show that the task of haze-free can be completed by a common GrC based on fuzzy logical formulas of bounded fuzzy operator. 3. Hybrid designing of leveled perception granular system based on fuzzy logic and PSVM Owing to the limitation of the scope, in this paper only nested layered GrC is discussed. A nested layered GrC is defined by the in- put and output relation of a granular computing on a granular sys- tem. There are three kinds relations between nearby layers (layers k and k þ 1) of a nested GrC: (1) binary logic; (2) fuzzy logic; (3) alogical relation. Because fuzzy logic and binary logic are all created by the sig- moid function, so back propagation method can be used to mod- ify weights of all layers. In order to speed up the learning process, for a layered GrC, we combine logical designing with PSVM (Fung & Mangasarian, 2001), such kind novel approach is called as ‘‘Logical support vector machine (LPSVM)’’. For nested layered GrC, parameters in the binary logical layers can be di- rectly designated according to the binary relation; for the fuzzy logical layers, parameters can also be set according to these lay- ers’functions, but a suitable small adjustment by back propaga- tion is necessary, this is similar to the deep learning proposed by Geoffrey Hinton such that a many-layered neural network could be effectively pre-trained one layer at a time, treating each layer in turn as an unsupervised restricted Boltzmann machi- ne,then using supervised backpropagation for fine-tuning. For the non logical (alogical) layer, parameters should be learned based on samples according to the input and output relation function fiðx1; x2; x3; . . . ; xnÞ, we can use Back Propagation method or PSVM, to learn weights for fiðx1; x2; x3; . . . ; xnÞ . The designing strategy of LPSVM: � Step 1: Except for the alogical layer’s weights, designing the lay- ers’ weights according to the logical (binary or fuzzy) relations, for fuzzy logical relations, a suitable modification of weights maybe be necessary according to the task of this layer; � Step 2: Alogical layers’ weights are computed for the input layer to the last output layer. For an alogical layer i, if X is the input train set, computing the inner layers’ output from the 1st layer to the ði � 1Þth layer based on X; � Step 3: Using PSVM to compute the ith layer’s weights W i according to (8); � Step 4: Repeat the Step 2 to Step 4, until the output error is small enough. � Step 5: using back propagating approach to modify all layers weights. The weight vector W l of W i ¼ X0DU ð8Þ Where The weight vector W i of the node, U is computed by (9) and X and D are the problem data, i.e. X ¼ ½X1; . . . ; Xn�, and diagonal ma- trix D ¼ y1 0 0 .. . . . . .. . 0 . . . yn 2 64 3 75;ðXi; yiÞ is a training sample with Xi feature vector and target yi . U ¼ I m þ DðXX0 þ EE0ÞD � ��1 E ð9Þ Where m is a positive parameter selected for guarantee of a small magnitude kW ik; I is the identity matrix, and E is a vector with all elements are 1. 4. Granular system for visual task The columnar organization of our brain’s primary visual cortex strongly supports the granular system defined aforementioned. Many functions of the primary visual cortex are still unknown, but the columnar organization is well understood. The lateral geniculate nucleus (LGN) transfers information from eyes to brain stem and primary visual cortex (V1) (Mountcastle, 1997). Colum- nar organization of V1 plays an important role in the processing of visual information. Local similarity of information processing gives rise to Colum- nar organization has a granular structure. V1 is composed of a grid of ð1 1mm2Þ neural area of hypercolumns (hc) in our brain’s pri- mary visual cortex. Every hypercolumn contains a set of minicol- umns (mc), which have same focus. Each hypercolumn analyzes information from one small region (described by a distance func- tion) of the retina. Adjacent hypercolumns analyze information from adjacent areas of the retina, so the structure of a columnar organization can be described by a set of fuzzy logical formulas similar to a granular system. Hypercolumns (or supercolumns), minicolumns (mc) can be viewed as granules. Similar to the pri- mary visual cortex,in our granular system, there are two kind gran- ules:hyper-granule and mini-granules in some levels of our granular system. A hyper-granule contain a bundle of mini- granules. Definition 6 (Perception Granular system of Columnar Organization (COGsys)). A perception columnar organization is a special per- ception granular system, in which, there is at least one hyper- granule GðcoeGnþ1Þ such that all mini-granules included in it have same convex region, but different adjoint functions. In this paper, in order to simulating visual cortex, a granular system of columnar organization (COGsys) is designed for the haze-free task.In our Hybrid designing approach (LPSVM), we firstly design Leveled Granular Systems with the help of fuzzy lo- gic, and then we use PSVM to accomplish the learning for some concrete visual tasks. 4.1. The theory of image matting According to Levin, Lischinski, and Weiss (2008), image matting refers to the problem of softly extracting the foreground object from an input image and a trimap image. ‘‘Tripmap’’ means three kinds of regions, white denotes definite foreground region, black denotes definite background region and gray denotes undefined region. Formally, image matting methods take I as an input, which is assumed to be a composite of a foreground image F o and a back- ground B in a linear form and can be written as I ¼ aF o þð1 � aÞB . For the haze-free task, the fuzzy label of haze or non-haze is described by the parameter a . And the task of image matting tries to find a function Fo ¼ fFoðIÞ . Closed form solution assumes that a is a linear function of the input image I in a small window 2734 H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 w : ai aIi þ b;8i 2 w . Then to solve a spare linear system to get the alpha matte. Our GrC approach gets rid of the linear assump- tion between a and I. Instead, we try to introduce nonlinear rela- tion between a and I: aw ¼ FðW IwÞ ð10Þ here W Iw is the image block included in the small window w, and aw is its center pixel’s fuzzy label. We take color or texture in local win- dow as our input feature, and the trimap image as the target. After training, the neural fuzzy logical network will generate the result of alpha matte. In the application of alpha matting, our method can re- move the haze using dark channel prior as the trimap. 4.2. Leveled perception granular system for haze-free task In this section, we try to design a Perception Granular system of Columnar Organization (COGsys) for the haze-free task, here only nested layered GrC is needed. The recognition of our Leveled Fig. 2. A 4 layers’ structure of a granular sy Granular System (see Fig. 2) is started with the recognition orien- tation or simple structure of local patterns, then the trimap image is computed based on these local patterns. Eq. (11), which has a high ability to simulate fuzzy logic operator (see the detail in the appendix) is used to design GrC. The weight wi in Eq. (11) can be viewed as connections among granules. A nested layered GrC is de- fined by the input and output relation of a granular computing on a granular system. just as above mentioned, there are three ways to design weights of a layered GrC:according to the binary or fuzzy logical relation about this layered GrC and according to the input and output relation function fiðx1; x2; x3; . . . ; xnÞ from training samples. Ulþ1;i ¼ X k wlþ1;i;k � Ilþ1;k;i Olþ1;i ¼ sigmðUlþ1;i; T lþ1;i; kÞ ð11Þ where sigmðÞ is a sigmoid function Eq. (12), and Olþ1;i is the output of a level l þ 1 granule. stem for haze-background separation. H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 2735 sigmðx; t; kÞ¼ 1=f1 þ expf�k � ðx � tÞgg ð12Þ The Theorem 1 discussed in the Appendix A guarantee that above defined granule computing can simulate a boolean function with arbitrary small error. As the designing of Gsys contains two parts: (1) convex regions, (2) adjoint fuzzy logical functions for GrC. The following Gsys for haze-free task is a very simple convex regions can be described by distance function. The input space is just an image, which is a 5-dimensional space X ¼fðx; y; r; g; bÞg, here every example ðx; y; r; g; bÞ represents a pix- el of this image, ðx; yÞ is the pixel’s location and ðr; g; bÞ is pixel’s color value. the nested granular system is build on the image. A granular system is built upon images, with fuzzy logical formula spða; cjdis; d; xÞ here x ¼ð1; 1; 0; 0; 0Þ, and d ¼ 0 for level 1 GrC, and d > 1 for higher level GrCs. All levels’ centers are located on the whole image plane, so every centers grid is just the image plane and granules are overlapped. In the following pages, we focus on the designing of adjoint fuz- zy logical functions for GrC. If there are k levels in our Gsys, the kth level receives the input image I, and the first level granule outputs the result haze free im- age. The relation between input and output of a level-l granule is described by Eq. (11). The weights among granules can be designed by LPSVM, the weights of 1st and 2nd layers are designed by fuzzy logic, and the weights of the 3rd layer are designed by PSVM to learn the trimap image. For the sake of simplicity, in the following Gsys, we use the or- der of GrC level which is upside down with the granular system le- vel, and one layer may contain two GrC levels. The GrC of COGsys is formally defined as bellow: (1) The 1st layer – fuzzy logical layer Every hyper-granule (Fig. 3) in the 1st layer tries to change a 3 3 pixels’ image block Ib 3 3 into a binary 3 3 pixels’texture pattern. The input image is normalized. A hyper-granule HG ¼ðg; SFÞ in the 1st layer contains 3 3 mini-granules to focus a 3 3 small window, every mini-granule focuses only one pixel, so the convex region of a hyper-granule is described by disðx; cÞ� 0 . A hyper-granule completes the task of image processing. There are three kinds fuzzy logical functions in a hyper-granule’s SF : Fig. 3. pattern just a 3 every g (1) In a local image pattern recognition way (LIPW): the 1st processing directly transforms every pixel’s value to a fuzzy logical one by a sigmoid function. F1ðfIbg3 3Þ¼ fsigmðpi;jÞg3 3 ð13Þ Every 1st layer granule tries to change a local image into a binary texture . For a hyper granule is defined by a distance function disðx; cÞ < 3, which is 3 small window, a hyper-granule in the 1st layer contains 9 granules, and ranule focuses only one pixel. here pi;j; i; j ¼ 1; 2; 3 is the RGB pixel value in a small 3 3 window; (2) In a local Binary Pattern operator simulating way (LBPW). The 2nd processing is also completed by a sig- moid function; the difference is that every boundary pix- el’s value is fuzzy exclusive OR � with the center pixel’s value before sending it to a sigmoid function, F 2ðfIbg3 3Þ¼ ffðpi;jÞg3 3 ð14Þ Here fðpi;jÞ¼ sigmðpi;j � p2;2Þ when i; j – 2, and fðp2;2Þ¼ 0. F2 is similar to a Local Binary Pattern operator (LBP) mentioned in Ojala, Pietikäinen, and Harwood (1996) as a mean of sum- marizing local gray-level structure. The operator takes a local neighborhood around each pixel, thresholds the pixels of the neighborhood at the value of the central pixel and uses the resulting binary-valued image patch as a local image descrip- tor. It was originally defined for neighborhoods, giving 8 bit codes based on the 8 pixels around the central one. Such processing emphasizes the contrast of texture, and our exper- iments support this fact. (3) Hybrid LIPW and LBPW (LBIPW). The adjoint function F 3ð�Þ in LBIPW is same as F2ð�Þ in LBPW, except that fðp2;2Þ¼ p2;2 in F 3ð�Þ, while fðp2;2Þ¼ 0 in F 2ð�Þ . Every granule in a 1st layer’s granule has only one input weight wij in Fig. 3, which equals 1; when k !þ1, the coef- ficient k in Eq. (11) changes the outputs from fuzzy values to binary numbers. (2) The 2nd layer–binary logical layer Every 2nd layer mini- granules try to recognize a definite shape (see Fig. 4), so they share the same convex region with a 1st layer hyper-gran- ule, which focuses on the same small 3 3 window in an image, and can be described by disðx; cÞ < 2. If there are total q local small patterns, a hyper-granule in the 2nd layer con- tains q (in our system q ¼ 256 or 512) mini-granules of the 2nd layer, which have same receptive field, but with a differ- ent adjoint fuzzy logical function, which tries to recognize a definite shape from the output of a 1st layer hyper-granule. For example, the ‘‘\’’ shape in Fig. 4 can be described by a adjoint fuzzy logical formula (Eq. (15)). The ‘‘and’’ operator for 9 inputs in Eq. (15)can be created by a granule mc (see. Fig. 4). In Eq. (15), every pixel Pij has two states mij and mij . Suppose the unified gray value (or RGB value) of Pij is gij, and an image module needs a high value gij at the place of mij and a low value at mij . So the input for the gran- ule mc at mij is Iij ¼ gij, and at mij is Iij ¼�ð1:0 � gijÞ. A not gate mc0 is needed for Iij ¼�ð1:0 � gijÞ, here gij; i; j ¼ 1; 2; 3 is the output of a 1st layer hyper-granule. P¼m11 ^m12 ^m13 ^m21 ^m23 ^m31 ^m33 ^m22 ^m32 ð15Þ wij ¼ 1; if the jth bit of a binary pattern ¼ 1 �1; if the jth bit of a binary pattern ¼ 0 � ð16Þ where for LIPW and LBIPW, j ¼ 1; 2; 3; . . . ; 9; for LBPW, the cen- ter 1st-layer granule is useless, so j ¼ 1; 2; 3; . . . 8. There are three kinds hyper-granules in the 2nd-layer, which receive three differ- ent outputs of a 1st-layer’s hyper-granule, so a hyper-granule in the 2nd-layer may work in one of following three ways: 1. In the local image pattern recognition way (LIPW): every 2nd layer hyper-granule contains 512 2nd-layer’s mini-granules, and inputs of these 2nd-layer’s mini-granules come from a 1st-layer’s hyper-granule which works in LIPW way. Every 2nd-layer’s hyper-granule tries to classify the image block in this window into 512 binary texture patterns (BTP), e.g. eight important BTPs are shown in Fig. 5. The pixel value is ‘‘1’’ for Fig. 5. Every the 2nd layer’s granule contains 256 or 512 granule which corresponds to 256 or 512 modules in above picture. Fig. 4. A hyper-granule in the 2nd layer contains q granules which have same receptive field and try to recognize q definite small shapes. A ‘and’ granule is needed for every 2nd layer granule. 2736 H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 white and ‘‘0’’ for black. In this mode, 3 3 granules of the 1st layer output a 3 3 vector, i.e., a 3 3 fuzzy logical pattern of a BTP, which is computed by a sigmoid function. 2. In the local Binary Pattern operator simulating way (LBPW), a 2nd-layer’s hyper-granule contains 256 2nd-layer’s mini- granules which receive input from the output of a 1st-layer’s hyper-granule, which works in the way of LBPW. 3. In the hybrid LIPW and LBPW (LBIPW) way, a 2nd-layer’s hyper- granule contains 512 2nd-layer’s mini-granules which receive input from the output of a 1st-layer’s hyper-granule, which has 9-dimensions. In our system, a Gsys is built for every color channel R,G or B, so a hyper-granule in the 2nd layer has a 512 3 dimensions output or 256 3 dimensions output. As a binary logical layer, in order to recognize a binary pattern, an ‘and’ granule with index i is needed (see Fig. 4) for every 2nd- layer granule, and the weights of this ‘and’ granule to the 1st-layer granules are set as Eq. (16), the corresponding parameters in Eq. (11) are set as the threshold T i ¼ 5:1, and k ¼ 0:9. 4.2.1. The 3rd layer – alogical layer The convex region of this layer can also be described by disðx; cÞ < 2. The output of a hyper-granule in the 2nd layer, which has 3 256 or 3 512 dimensions, is transformed to the 3rd-layer granules to compute the similarity parameter or fuzzy value ai in Eq. (10), the weights of this layer is computed by psvm, the target is provided by so called dark channel prior which is computed by the approach mentioned in He, Sun, and Tang (2011). As all ai are optimised on the whole image, in this layer,the whole image is the only convex region. As the small windows focused by hy- per-granules in the 2nd-layer are overlapped, the focuses of 3rd- layer’s granules are also overlapped. 4.2.2. The 4th layer – fuzzy logical layer In this layer, a granule tries to remove the haze from original image. A granule in the 4th layer computes a pixel of a haze free image according to fuzzy logical equation Eq. (17) IiðxÞ¼ minfq; aiðxÞ � JiðxÞþð1 � aiðxÞÞ � Aig¼ JiðxÞ�f Ai ð17Þ where Ji is the haze free image, Ii is the original image, Ai is the global atmospheric light which can be estimated from dark channel prior, ai is the alpha matte generated by 3rd layer, and �f is the q-value Weighted Bounded sum with weights w1 ¼ aiðxÞ; w2 ¼ 1 � aiðxÞ, here q is max gray or RGB value of a pixel, and aiðxÞ and ð1 � aiðxÞÞ are weights. Although we can use back propagation approach to compute pixels’ value JiðxÞ given the haze image pixel value IiðxÞ based on Eq. (17), for the sake of simplicity, we directly use the Eq. (18) mentioned by He et al. (2011) to com- pute the haze free image. As every aiðxÞ is computed upon the whole image, the pixel of haze-free image is also computed upon whole image, so the whole image is also the convex region of this layer. JiðxÞ¼ IiðxÞ� Ai maxðaiðxÞ; a0Þ þ Ai ð18Þ where Ji is the haze free image, Ii is the original image, Ai is the glo- bal atmospheric light which can be estimated from dark channel prior, ai is the alpha matte generated by 3rd layer, and a0 is a threshold, a typical value is 1. 4.3. Experiments result The haze-free experiment result (1) The haze-free and texture information entropy Texture information can give out a rough measure about the effect of haze-freeing, we use the entropy of the texture his- togram to measure the effect of deleting haze from images. The entropy of the histogram is described in Eq. (19). Haze makes the texture of an image unclear, so theoretically speaking, haze removing will increase the entropy of the texture histogram. Entropy : H ¼� XG�1 i¼0 pðiÞlog2½pðiÞ� ð19Þ Fig. 6. The processing result of granular system for visual haze-free task. H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 2737 The pðiÞ denote the rate of each pattern in histogram. In general pðiÞ define in Eq. (20). Here patterns we use are the LBPs in Eq. (14), where Sigmðin � iCÞ¼ 1, if in > iC þ 10 else Sigmðin � iCÞ¼ 0 . Table 1 The tex Area2: Area Area Area Fig. 7. granule pðiÞ¼ hðiÞ=ðNMÞ; i ¼ 0; 1; . . . ; G � 1 ð20Þ In Fig. 6(a), (b), (c), (d) and (e) are the results of LBPW, LBIPW, LIPW, the linear mode (LMKH) by He et al. (2011), and the original image respectively. From Fig. 6, we can see that the texture structure in the waist of a mountain becomes vaguer from LBPW, LBIPW, LIPW to LMKH. For the sake of the 2nd kind processing in the 1st-layer’s granules pays much more attention to the contrast, LBPW has the highest ability to remove the haze, LBPW and LIPW are complemen- tary approaches, LBIPW, which is the cooperation of them, has a similar ability as the linear approach proposed by He et al. (2011). According to the results showed in the Table 1, which are about tex- ture information entropy of the image, we can see that the texture information entropy is increased after haze-free processing, so our approaches have higher ability to increase the texture information entropy than the linear approach proposed by He et al. (2011). ture information entropy of the image blocks (Area1: the waist of a mountain; right bottom corner) in the Fig. 6. LBPW LBIPW LIPW LMKH Original 1 5.4852 5.2906 5.1593 4.8323 1.0893 2 6.1091 10.3280 10.2999 9.1759 8.3718 The relation among the precision (rmse) of PSVM learning and k parameters in th . Theoretically speaking, LBPW is a pure texture processing, so LBPW has a highest value, LIPW is much more weaker than LBPW, LBIPW is the hybrid of LBPW and LIPW, so it has a average ability. The tex- ture information entropy of the Area1 correctly reflects this fact. But for the Area2, as it already has a clearest texture structure in the ori- ginal image, the deleting of haze may cause overdone. The texture information is over emphasized by LBPW in the Aera2, so it has a lowest texture information entropy and almost becomes a dark area. This fact means that overtreatment is more easier to appear in a non linear processing than a linear one in the haze-free task. (2) The effect about the degree of fuzzyness Just as the Theorem 1 mentioned above, the parameter k in Eq. (10) can control the fuzzyness of a granule, when the parameter k in Eq. (10) tends to infinite, a granule behaves from a fuzzy logical formula to a binary logical formula. This experiment is about the relation among the precision (rmse) of PSVM learning and k parameters in the first and second layer. LBPW is a pure texture processing and pays much more attention to the contrast of an image’s nearby pixels, a set of large k is necessary for a low rmse, which corre- sponds to binary logic; but LBIPW and LIPW aphe pear to prefer fuzzy logic for a set of small k when rmse is small. A possible explanation for this fact is that LBP proposed by Ojala et al. (1996) is binary, not fuzzy, and has a sound clas- sification ability for image understanding under binary pat- tern, but LBIPW and LIPW are not binary, they have fuzzy information at least for the center pixel of a 3 3 small window (Fig. 7). e first and second layer, the parameter k in Eq. (10) can control the fuzzyness of a Fig. 9. Simulating fuzzy logical and-or by changing thresholds of Eq. (11). The X- axis is the threshold value divided by 0.02, the Y-axis is errG. The real line is errAnd between I1 �f I2 and V i , and the dot line is the errOr between I1�f I2 and V i . Fig. 8. More result of granular system for visual haze-free task. 2738 H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 (3) The comparison between our approach and LMKH To illustrate the effect of our approach in haze-free task, we apply it on the other images and compare with LMKH (Fig. 8). Half of the result is better than the LMKH, the rest is as good as the LMKH by manual evaluation. 5. Discussion In this paper, we give out a concrete example to show that the theory of GrC can help us to design the brain-like computer. The experimental results show that LPSVM is a promising approach for designing of a granular system similar to a columnar organiza- tion for image haze-removing task. The concept of granular com- puting is proposed by Bargiela and Pedrycz (2006). Just as he said: a granule is a clump of objects (points) drawn together by indistinguishability, similarity, proximity or functionality. The nec- essary of granular computing to study the information transforma- tion in the pattern recognition lies in indistinguishability, similarity, proximity or functionality of sensed information Due to the local similarity in the information processing of pattern rec- ognition, multi-scale information processing is a common phe- nomenon in pattern recognition. In actual fact, the GrC based on the leveled granular system aforementioned can simulate all mul- ti-scale information processing with arbitrary small error. This fact is very important for the hot approach-deep learning. In this paper, we use a novel designing approach (LPSVM) to design a granular system similar to the structure of columnar organization of visual cortex, We demonstrate that fuzzy logic and machine learning can be hybrid and cooperated easily to design a granular system. This approach not only give out a novel concrete realization of abstract models for granular computing mentioned Lin (2012), but also gives a new focus for deep learning. For more,the corre- sponding of GrC can simulate multi-scale information processing for the task of haze-free of images, and our experiments show that H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 2739 our approach has some approvement for the task of haze-free com- paring to the approach proposed by the linear approach proposed by He et al. (2011). For further directions, although our LPSVM gives out a concrete example for designing a granular system for haze free task, many details of LPSVM should be studied in the task of pattern recogni- tion under the framework of deep learning, especially for the lay- ered feature abstraction in the task of pattern recognition. For more, we will extend the investigation by looking at other nested layered computing for more complex tasks. However, since layered computing has no feedback, which is important for many visual tasks in dynamical situations, we also plan to extend our layered granular computing to a more general one which allows for both feedback and dynamical regulation for the task of computer vision. Acknowledgments This work is partially supported by the National Program on Key Basic Research Project (973 Program) (No. 2013CB329502), the National Natural Science Foundation of China (Nos. 61072085, 61035003, 61202212, 60933004), the National High- tech R and D Program of China (863 Program) (No. 2012AA011003), the National Science and Technology Support Program (2012BA107B02) and the China Information Technology Security Evaluation Center (CNITSEC-KY-2012-006/1). Appendix A. Sigmoid function and Binary Logic Theorem 1. Suppose in Eq. (11), every wlik ¼ bk T i; bk > 0; T i > 0; 1 6 k 6 K, for more, C ¼fSiji ¼ 1; . . . ; Lg is a class of index sets, and every index set Si is a subset of f1; 2; 3; . . . ; Kg, then we have: (1) If fðx1; x2; . . . ; xkÞ¼ _ l¼1;...;L ð^xjiÞ jl2Sl is a disjunctive normal form (DNF) formula, and the class C ¼fSiji ¼ 1; . . . ; Lg is the class which has the following two characters: (1). for every Si; Sj 2 C; Si \ Sj – Sk 2 C for all k and i – j (this condition assures that fðx1; x2; . . . ; xkÞ has a simplest form); (2). every Si has the character P j2Sl bj > 1, where Si 2 C, and any index sets S0 R C have character P j2S0bj < 1, or if P j2S0bj > 1, there must be an index set Si 2 C such that S0 \ Si ¼ Si (this condition assures C is the largest), then the output described by Eq. (11) can simulate the DNF formula fðx1; x2; . . . ; xkÞ¼ _ l¼1;...;L ð^xjiÞ i2Sl with arbitrary small error, where xi ¼ zi, if the corresponding input Ii ¼ zi, or xi ¼ �zi if Ii ¼ 1 � zi . (2) If a neural cell described by Eq. (11) can simulate the Boolean formula fðx1; x2; . . . ; xkÞ with arbitrary small error, and ð^xiÞ i2Sl is an item in the disjunctive normal form of fðx1; x2; . . . ; xkÞ, i.e. fðx1; x2; . . . ; xkÞ¼ 1 at xj ¼ 1 for all j 2 Sl and xj ¼ 0 for all j R Sl, then P i2Sl bi > 1. (3) If a couple of index sets Sl1 and Sl2 can be found in the formula fðx1;x2;. . .;xkÞ¼ _ l¼1;...;k; ð ^ t2Sl xtÞ, such that ð ^ t12Sl1 xt1Þ^ð ^ t22Sl2 xt2Þ¼ zi ^ �zi ¼ false, then the output described by Eq. (11) can’t simulate the formula fðx1; x2; . . . ; xkÞ . Proof. (1) If It ¼ 1, for all t 2 Sl, and It ¼ 0, for all t R Sl, becauseP i2Sl bi > 1, then for the index set Sl is a subset of f1; 2; 3; . . . ; Kg, we have V i ¼ 1=½expð�kðUi � T iÞþ 1� ¼ 1=½expð�kð X 16k6K wikIk � T iÞÞþ 1� ¼ 1=½expð�kð X i2Sl bi � 1ÞT iÞÞþ 1�; so limk!þ1V i ¼ 1 ¼ fðx1; x2; . . . ; xkÞ. If It ¼ 1;8t 2 S0; It ¼ 0;8t R S0 and S0 R C, then according to the condition of this theorem: if P i2S0bi < 1, limk!þ1V i ¼ 0 ¼ fðx1; x2; . . . ; xkÞ; if P i2S0bi > 1, then there is an in- dex set Si 2 C such that S0 \ Si ¼ Si , then limk!þ1V i ¼ 1 ¼ fðx1; x2; . . . ; xkÞ. So when k !1, the error between output described by Eq. (11) and fðx1; x2; . . . ; xkÞ trends to 0. (2) If the output described by Eq. (11) can simulate the Boolean formula fðx1; x2; . . . ; xkÞ which is not a constant with arbi- trary small error, and for a definite binary input x1; x2; . . . ; xk , then the arbitrary small error is achieved when k trends to infinite and ðUi � T iÞ¼ P k2Sl wikIk � T i – 0 where Sl is the set of the labels and Ii ¼ 1, for all i 2 Sl , and Ii ¼ 0, for all i R Sl. The theorem’s condition supposes that every wik ¼ bkT i; bk > 0; T i > 0; 1 6 k 6 K, and x1; x2; . . . ; xk are bin- ary number 0 or 1, so if fðx1; x2; . . . ; xkÞ is not a constant, when fðx1; x2; . . . ; xkÞ¼ 0, there must be limk!þ1V i ¼ 0; and when fðx1; x2; . . . ; xkÞ¼ 1, it is necessary for limk!þ1V i ¼ 1. limk!þ1V i ¼ 0 needs that -kð P i2Sl biT i � T iÞ trends to minus infinite and limk!þ1V i ¼ 1 needs that -kð P i2Sl biT i � T iÞ trends to plus infinite. So if fðx1; x2; . . . ; xkÞ¼ 1 at xj ¼ 1 for all j 2 Sl and xj ¼ 0 for all j R Sl, in order to guarantee limk!þ1errfðx1;x2;...;xkÞðwi;1; wi;2; . . . ; wi;k; T iÞ ¼ 0; P i2Sl bi > 1 must be hold, here errfðx1;x2;...;xkÞðwi;1; wi;2; . . . ; wi;k; T iÞ is the error between output described by Eq. (11) and fðx1; x2; . . . ; xkÞ. (3) The third part of the theorem is based on the simple fact that for a single neuron V i is monotone on every input Ii which can be zi or 1 � zi. h Appendix B. Sigmoid function and fuzzy logic For more above granular computing can approximately simu- late Bounded operator. Bounded operator Fð�f ;�fÞ Bounded prod- uct p�f q ¼ maxð0; p þ q � 1Þ, Bounded sum p�f q ¼ minð1; p þ qÞ. Based on Eq. (11), the membrane potential’s fixed point under in- put Ik is Ui ¼ P k wikIk and the output at the fixed point is V i ¼ 1=ðexpð�Ui þ T iÞþ 1Þ. If there are only two inputs I1; I2ðI1; I2 2 ½0; 1�Þ in Eq. (11), we set w1 ¼ 1:0 and w2 ¼ 1:0, then Ui ¼ I1 þ I2 . Now we try to prove that the Bounded operator Fð�f ;�fÞ is the best fuzzy operator to simulate neural cells described by (3) and the threshold Ti can change the neural cell from the bounded oper- ator �f to �f by analyzing the output at the fixed point V i ¼ 1=ðexpð�Ui þ T iÞþ 1Þ. If C > 0 is a constant and Ui ¼ I1 þ I2 P C, then 1=ðexpð�C þ T iÞþ 1Þ6 V i < 1 . When Ui ¼ I1 þ I2 !þ1V i ! 1, so in this case, if C is large enough, V i 1. If -C 6 Ui ¼ I1 þ I2 6 C, then 1=ðexpðC þ T iÞþ 1Þ6 V i 6 1=ðexpð�C þ T iÞþ 1Þ, according to equation (a). We can select a T i, that makes jT i þ P1 j¼2ð�Ui þ T iÞ j =j! � P1 k¼2ð�1Þ k expð�kðUi T iÞÞj small enough, then V i I1 þ I2 . 2740 H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 V i ¼ 1=ðexpð�Ui þ T iÞþ 1Þ ¼ 1 � expð�Ui þ T iÞþ X1 k¼2 ð�1Þk expð�kðUi � T iÞÞ ¼ Ui � T i � X1 j¼2 ð�Ui þ T iÞ j =j! þ X1 k¼2 ð�1Þk expð�kðUi T iÞÞ ¼ Ui � T i � X1 j¼2 ð�Ui þ T iÞ j =j! þ X1 k¼2 ð�1Þk expð�kðUi T iÞÞ:ðaÞ So in this case, V i I1�f I2 ¼ minð1; I1 þ I2Þ. Similarly, if Ui ¼ I1 þ I2 !�1 V i ! 0. So when C is large enough and Ui ¼ I1 þ I2 6�C < 0, then V i 0. When -C 6 Ui ¼ I1 þ I2 6 C, if we select a suitable T i which makes T i þ P1 j¼2ð�Ui þ T iÞ j =j! � P1 k¼2ð�1Þ k expð�kðUi � T iÞÞ 1, then V i I1�f I2 ¼ maxð0; I1 þ I2 � 1Þ. Based on above analysis, the Bounded operator fuzzy system is suitable for GrC described by Eq. (11) when ai ¼ 1:0; w1 ¼ 1:0 and w2 ¼ 1:0. For arbitrary positive ai; w1 and w2, we can use corre- sponding q-value weighted universal fuzzy logical function based on Bounded operator to simulate such kind neural cells. If a weight w is negative, a N-norm operator NðxÞ¼ 1 � x should be used. Experiments done by scanning the whole region of ðI1; I2Þ in ½0; 1�2 to find the suitable coefficients for �f and �f show that above analysis is sound. We denote the input in Eq. (11) as ~x ¼ðI1; I2Þ. The ‘‘errOr’’ for �f and ‘‘errAnd’’ for �f are shown in Fig. 9 as the solid line and the dotted line respectively. In Fig. 9, the threshold T i is scanned from 0 to 4.1 with step size 0.01. The best T i in Eq. (4) for �f is 2.54 and the best T i in Eq. (4) for �f is 0, when a ¼ 1:0; w1 ¼ 1:0 and w2 ¼ 1:0. In this case the ‘‘errOr’’ and ‘‘errAnd’’ is less than 0.01. Our experiments show that suitable T i can be found. So in most cases, the bounded operator Fð�f ;�fÞ mentioned above is the suitable fuzzy logical framework for the neuron defined by Eq. (3). If the weight 0 < w1 and 0 < w2, we should use a q-value weighted bounded operator Fð�f ;�fÞ to rep- resent above neuron. Appendix C. Associative condition and Demorgan law of q- weighted bounded operator It is easily to see �f follows the associative condition and x1�f x2�f x3 . . .�f xn ¼ minðq; P 16i6nwixiÞ. For �f , we can prove the associative condition is hold also. The proof is listed as below: If w1p1 þ w2p2 �ðw1 þ w2 � 1Þq P 0, we have: ðp1�f p2Þ�f p3 ¼F�f ðF�f ðp1;p2;w1;w2Þ;p3;1;w3Þ ¼F�f ðw1p1þw2p2�ðw1þw2�1Þq;p3;1;w3Þ ¼maxð0;w1p1þw2p2�ðw1þw2�1Þqþw3 p3 �ð1þw3�1ÞqÞ ¼maxð0;w1p1þw2p2þw3p3�ðw1þw2þw3�1ÞqÞ; if w1p1 þ w2 p2 �ðw1 þ w2 � 1Þq < 0, we have ðp1�f p2Þ�f p3 ¼F�f ðF�f ðp1;p2;w1;w2Þ;p3;1;w3Þ ¼F�f ð0;p3;1;w3Þ¼maxð0;0þw3 p3 �ð1þw3 �1ÞqÞ ¼maxð0;w3 p3 �w3qÞ ¼ for06p36q0 ¼maxð0;w1 p1 þw2p2 þw3p3 �ðw1 þw2 þw3 �1ÞqÞ; so ðp1�f p2Þ�f p3 ¼ p1�fðp2�f p3Þ¼ maxð0; w1 p1 þ w2p2 þ w3 p3 �ðw1 þ w2 þ w3 � 1ÞqÞ. By inductive approach, we can prove that �f also follows the associative condition and x1�f x2�f x3 . . .�f xn ¼ maxð0; P 16i6n wixi �ð P 16i6nwi � 1ÞqÞ. For more if we define NðpÞ¼ q � p (usually, a negative weight wi corresponds a N-norm), above weighted bounded operator Fð�f ;�fÞ follows the Demorgan Law, i.e. Nðx1�f x2�f x3 . . .�f xnÞ¼ q � min q; X 16i6n wi xi ! ¼ max 0; q � X 16i6n wixi ! ¼ max 0; X 16i6n wiðq � xiÞ�ð X 16i6n wi � 1Þq ! ¼ Nðx1Þ�f Nðx2Þ�f Nðx3Þ . . .�f NðxnÞ: References Andrzej, B., & Pedrycz, W. (2006). The roots of granular computing. In GrC (pp. 806– 809). Castro, J. L. (1995). Fuzzy logic controllers are universal approximators. IEEE Transactions on Systems, Man and Cybernetics, 25(4), 629–635. Fung, G., & Mangasarian, O. L. (2001). Proximal support vector machine classifiers. In Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 77–86). ACM. Haykin, S. (1994). Neural networks: A comprehensive foundation. Prentice Hall PTR. Haykin, S. (2008). neural networks: A comprehensive foundation. Englewood cliffs: Prentive-Hall. He, K., Sun, J., & Tang, X. (2011). Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12), 2341–2353. Levin, A., Lischinski, D., & Weiss, Y. (2008). A closed-form solution to natural image matting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2), 228–242. Lin, T. Y. (1998). Granular computing on binary relations I: Data mining and neighborhood systems. Rough Sets in Knowledge Discovery, 1, 107–121. Lin, T. Y. (1999). Granular computing: Fuzzy logic and rough sets. Computing with words in information/intelligent systems (Vol. 1, pp. 183–200). Springer. Lin, T. Y. (2007). Neighborhood systems: A qualitative theory for fuzzy and rough sets. Berkeley: University of California. 94720. Lin, T. Y. (2012). Granular computing: Practices, theories, and future directions. In Computational complexity (pp. 1404–1420). Springer. Li, H.-X., & Philip Chen, C. L. (2000). The equivalence between fuzzy logic systems and feedforward neural networks. IEEE Transactions on Neural Networks, 11(2), 356–365. Liu, H., Xiong, S., & Wu, C.-a. (2012). Hyperspherical granular computing classification algorithm based on fuzzy lattices. Mathematical and Computer Modelling. Mountcastle, V. B. (1997). The columnar organization of the neocortex. Brain, 120(4), 701–722. Ojala, Timo, Pietikäinen, Matti, & Harwood, David (1996). A comparative study of texture measures with classification based on featured distributions. Pattern Recognition, 29(1), 51–59. Pedrycz, Adam, Hirota, Kaoru, Pedrycz, Witold, & Dong, Fangyan (2012). Granular representation and granular computing with fuzzy sets. Fuzzy Sets and Systems, 203, 17–32. Yao, Y. Y. (1998). Relational interpretations of neighborhood operators and rough set approximation operators. Information Sciences, 111(1), 239–259. Yao, Y. Y. (1999). Granular computing using neighborhood systems. In Advances in soft computing (pp. 539–553). Springer. Yao, Y. Y. (2000). Granular computing: Basic issues and possible solutions. In Proceedings of the 5th joint conference on information sciences (Vol. 1, pp. 186– 189) Citeseer. Yao, Y. Y. (2001). On modeling data mining with granular computing. In 25th Annual international computer software and applications conference, 2001. COMPSAC 2001 (pp. 638–643). IEEE. Yao, Y. Y. (2001). Information granulation and rough set approximation. International Journal of Intelligent Systems, 16(1), 87–104. Yao, Y. (2006). Granular computing for data mining. In Defense and security symposium (pp. 624105). International Society for Optics and Photonics. Yao, Y., & Deng, X. (2013). A granular computing paradigm for concept learning. In Emerging paradigms in machine learning (pp. 307–326). Springer. Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338–353. http://refhub.elsevier.com/S0957-4174(13)00906-8/h0015 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0015 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0020 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0020 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0020 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0025 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0030 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0030 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0035 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0035 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0035 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0040 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0040 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0040 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0045 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0045 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0050 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0050 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0055 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0055 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0060 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0060 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0065 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0065 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0065 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0070 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0070 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0070 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0075 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0075 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0080 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0080 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0080 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0085 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0085 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0085 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0090 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0090 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0095 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0095 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0100 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0100 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0100 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0105 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0105 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0110 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0110 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0115 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0115 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0120 H. Hu et al. / Expert Systems with Applications 41 (2014) 2729–2741 2741 Zadeh, L. A. (1997). Toward a theory of fuzzy information granulation and its centrality in human reasoning and fuzzy logic. Fuzzy Sets and Systems, 90(2), 111–127. Zhang, L., & Zhang, B. (2003). Theory of fuzzy quotient space (methods of fuzzy granular computing). Journal of Software, 14(4), 770–776. Zhang, L., & Zhang, B. (2004a). The quotient space theory of problem solving. Fundamenta Informaticae, 59(2), 287–298. Zhang, L., & Zhang, B. (2004b). The quotient space theory of problem solving. Fundamenta Informaticae, 59(2), 287–298. Zhang, L., & Zhang, B. (2005). Quotient space model based hierarchical machine learning. International conference on neural networks and brain, 2005. ICNN&B’05 (Vol. 1). IEEE, pp. xiv–xiv. http://refhub.elsevier.com/S0957-4174(13)00906-8/h0125 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0125 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0125 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0130 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0130 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0135 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0135 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0140 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0140 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0145 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0145 http://refhub.elsevier.com/S0957-4174(13)00906-8/h0145 Perception granular computing in visual haze-free task 1 Introduction 2 Granular system based on tolerance relation 3 Hybrid designing of leveled perception granular system based on fuzzy logic and PSVM 4 Granular system for visual task 4.1 The theory of image matting 4.2 Leveled perception granular system for haze-free task 4.2.1 The 3rd layer – alogical layer 4.2.2 The 4th layer – fuzzy logical layer 4.3 Experiments result 5 Discussion Acknowledgments Appendix A Sigmoid function and Binary Logic Appendix B Sigmoid function and fuzzy logic Appendix C Associative condition and Demorgan law of q-weighted bounded operator References 
work_chyoul2fuvf6jo3cuco3onjhom	----	Insight SFI Research Centre for Data Analytics Search Insight Menu About Who we are What we do Our structure People Work With Us Senior leadership Principal Investigators Funded Investigators Research and Operations Research Application Domains Demonstrators Research Challenges Core Scientific Expertise Publications Projects European Funded Projects Business Masterclasses Business Team Public Engagement EPE Committee Citizen Science News Latest News Media Queries Newsletter Spotlight on Research Contact Search Search for: Close search Close Menu AboutShow sub menu Who we are What we do Our structure PeopleShow sub menu Work With Us Senior leadership Principal Investigators Funded Investigators Research and Operations ResearchShow sub menu Application Domains Demonstrators Research Challenges Core Scientific Expertise Publications Projects European Funded Projects BusinessShow sub menu Masterclasses Business Team Public EngagementShow sub menu EPE Committee Citizen Science NewsShow sub menu Latest News Media Queries Newsletter Spotlight on Research Contact Insight Impact Ep.15 Happy Poddy’s Day! Listen to the final episode of The Insight Podcast Read On Insight Careers Insight is hiring! Check out our jobs page Read On Work with us Insight Impact Professor Barry O’Sullivan wins the prestigious Nerode Prize Read On Insight Business Visit our business team and read industry case studies Insight Research Research informed by our vision: Enabling Citizens. Smarter Societies. Insight People Find an expert. Work with us. Public Engagement School visits, events, citizen science, public education and more Latest News ‘People believe VR is all about games!’ Lukasz Porwol talks to Silicon Repu... 24/03/2021 Insight partners with global leader Avaya on Next Gen AI 24/03/2021 Prof Alan Smeaton talks Deepfakes with Newstalk’s Pat Kenny 16/03/2021 Ep.15 Happy Poddy’s Day! Listen to the final episode of The Insight Podcast 16/03/2021 Empowering Citizens Smarter Societies Find out more Insight in Numbers 0 + Researchers € 0 + Million in Funding 0 + Industry Partners 0 Research Institutions Stay up to date with our Newsletter Receive all the latest News & Insights* Leave this field empty if you're human: Privacy Statement Copyright Statement Data Protection Notice 
work_cjaafgvvdrcvhjnqlmkv42z6qm	----	col.dvi 1 Fault Detection and Fault Tolerance in Robotics Monica Visinsky, Ian D. Walker, and Joseph R. Cavallaro Department of Electrical & Computer Engineering Rice University Houston, TX 77251-1892 2 1 INTRODUCTION Abstract Robots are used in inaccessible or hazardous environments in or- der to alleviate some of the time, cost and risk involved in prepar- ing men to endure these conditions. In order to perform their expected tasks, the robots are often quite complex, thus increas- ing their potential for failures. If men must be sent into these environments to repair each component failure in the robot, the advantages of using the robot are quickly lost. Fault tolerant robots are needed which can effectively cope with failures and continue their tasks until repairs can be realistically scheduled. Before fault tolerant capabilities can be created, methods of de- tecting and pinpointing failures must be perfected. This paper develops a basic fault tree analysis of a robot in order to obtain a better understanding of where failures can occur and how they contribute to other failures in the robot. The resulting failure flow chart can also be used to analyze the resiliency of the robot in the presence of specific faults. By simulating robot failures and fault detection schemes, the problems involved in detect- ing failures for robots are explored in more depth. Future work will extend the analyses done in this paper to enhance Trick, a robotic simulation testbed, with fault tolerant capabilities in an expert system package. 1 Introduction In hazardous environments or environments which are not read- ily accessible to man, robots must be able to efficiently adapt to failures in both software and hardware in order to continue working until the problem can be realistically repaired. Before a robot can try to cope with a failure, however, it must first be able to detect and pinpoint the problem. This paper devel- ops a basic fault tree analysis of robots in order to obtain a better understanding of where failures can occur and how they contribute to other failures or limitations in each robot. The resulting flow chart style picture of failures in a robot can also be used to analyze the resiliency of the robot in the presence of specific faults. Once a failure has been detected, the robot can reorganize its view of its internal structure in such a way as to hide or isolate the fault so the robot can continue working. The focus of this research is on finding real-time fault detection and fault tolerance methods which maintain as much of the robot’s functionality as possible while not requiring that redundant or extra parts be added to the robot. 1.1 Previous Work 1.1.1 Redundancy Based Fault Tolerance Previous work on fault tolerance in robotics has concentrated on dealing with faults in one specific part of the robot (mechanical 1.1 Previous Work 3 failure in the motor, kinematic joint failure, etc.) with only to- ken thought going to the more critical, systemwide effect of the failures. Relatively little focus has been given to the question of how to detect failures in robots. Previous research tends to con- centrate on fault tolerance algorithms, especially those schemes which rely on duplicating parts such as joint motors [13, 16] for their fault tolerant abilities. These redundancy based schemes are similar to several computer fault tolerant algorithms which are also based on redundancy of parts. One common computer fault tolerant algorithm is Triple Modular Redundancy (TMR) in which three processors all work on the same problem and com- pare their results. If one of the processors is faulty and its result does not agree with the results of the other two processors, the faulty processor is voted out of the final decision and the correct result is passed on to the rest of the system. This fault toler- ance scheme fails, however, if more than one of the processors is faulty. Duplication of physical parts provides a backup in case the element performing the work fails. Redundancy of parts can also provide a useful means of checking to see if a component is in error. For the equivalent redundancy based robot fault tolerant al- gorithms, two motors have been placed in each robot joint to provide a backup in case one motor stalls, runs away, or be- gins free-spinning. The fault tolerant advantages of redundancy has also led to adding extra parallel structures, such as seven legs when only six are needed, in order to allow many differ- ent configurations in the presence of a failure. Previous work by Tesar, et al at UT, Austin [13] and independently by Wu [16] with Lockheed at Johnson Space Center have explored the aforementioned method of duplicating motors. Two motors in a joint work together so as to provide one output velocity for the joint. When one of the motors breaks, the other one takes over the faulty motor’s functions. The faulty motor must be isolated from the system or the second motor must be able to adjust its output to account for any transients introduced to the system by the failed motor. If the robot is performing a time- critical or delicate task, fault tolerance must allow the robot to get a run-away motor (which could cause drastic changes in the robot’s motion) under control quickly before any damage to the environment occurs. 1.1.2 Structure-Independent Fault Tolerance Many useful robots have already been created. In order to pro- vide fault tolerance for these robots without redesigning them, algorithms need to be developed that will utilize the advantages of the existing structure and not require the addition of extra motors, sensors, or other components to the robot. These algo- rithms should be easily adaptable to most robots regardless of the robot structure. 4 1 INTRODUCTION To avoid adding redundant parts for fault tolerance in comput- ers, algorithms have been developed which reconfigure the data or code in a computer system among the working parts after one component has failed. Some computer fault tolerant systems handle a fault by allowing a graceful degradation in function- ality or speed. The literature speaks of time redundancy [4] in which a computational cycle is lengthened so a fault-free part (or parts) will have enough time to handle the tasks of a faulty com- ponent. Other systems use set-switching or processor-switching schemes [4] for reconfiguration. In processor-switching, fault- free components can be collected to form a basic subpart of the configuration, such as a row of an array, until the full config- uration is achieved. This method may, however, require many extra interconnections between components. In software, check bits and error correction codes help insure that data is success- fully transmitted in the system and allow a reconstruction of the original data if a transmission line is faulty. For robotics, however, little work has been done in developing algorithms for accommodating a failure using only the available physical parts. Many robots exist which do not have redundant motors or extensive sensors in the joints. Duplicating motors in- creases the size of the robot, the cost involved in building it, and the weight and inertia which affect the robot controller. It would thus be cost effective to find fault tolerance schemes that do not rely on a specific robotic architecture but reorganize the robotic algorithms in the controller or utilize the self-motion capability of robots with redundant joints in order continue working. In order to develop these schemes, the advantages and capabilities of a general robot architecture must be researched. Maciejewski at Purdue University has quantified the effect of joint failure on the remaining dexterity of a kinematically redun- dant manipulator [9]. He calculates an optimal configuration of redundant arms to maximize the fault tolerance while minimiz- ing the degradation of the system in the event of a failure. His method currently only provides fault tolerance if the robot is near this initial configuration and can try and arrange its joints to mimic the fault safe configuration as close as possible. Robot controllers may further attempt to keep the robot arm in a con- figuration where the joints are arranged to stay away from any possible singularities or uncomfortable positions in case a joint fails during the operation. These studies do not rely on adding extra motors or other components to the robot, but they also do not explore what the robot controller must do in order to utilize the remaining dexterity to continue its tasks. This paper considers and analyzes systemwide failures (elec- tromechanical, computer software/hardware, etc.) and their inter-relationship via fault trees. We focus on developing fault detection and fault tolerance schemes using only the components normally available to the robot. Previous work in fault tolerance forms a subset of our analysis, and our structure has the addi- tional advantage of allowing the best results from more specific fault schemes to be embedded into our tree analysis. 1.2 Riceobot and Robo-MEDIC This paper begins by specifically analyzing the fault trees of the Rice University robot, the Riceobot, but the results apply to most robots. The fault tolerant algorithms developed from 1.3 Fault Detection Simulator and Trick 5 � � � �� � � � � � � � � � ��................ ......... .�� �� � � � � �� �� �� �� ...................................... “Trick” CPU Controller Real-Time Simulation and Control Software Error Detection and Robot Manipulator Operator Interface Intelligent Correction “Robo-MEDIC”Robot Manipulator Operator Figure 1: Robo-MEDIC Robotic Fault Tolerance Hardware/Software Environment. this analysis will be embedded into the CLIPS expert system environment [6]. This NASA-developed public domain software package is commonly used by government agencies and is run- ning on our computer systems. The resulting expert system package, Robo-MEDIC (Robot Ma- nipulator Error Detection and Intelligent Correction) will pro- vide diagnostic assistance to the operator and will interface with the control computer of the robot as shown in Figure 1. Robo- MEDIC will be able to use the fault trees as a flow chart of failures. Nodes in the trees will have some fault tolerant action associated with them that will allow the robot to take advantage of inherent backup or alternate paths charted by the fault tree. By maneuvering around the trees, Robo-MEDIC will perform fault tolerant recovery actions as a sequence of these smaller, simpler actions. 1.3 Fault Detection Simulator and Trick In addition to the fault tree analysis, we are examining fail- ures and testing fault detection schemes using a simulation of a generic four link, planar robot. We will be integrating the concepts derived from the simulator into Trick [1], a robotics software testbed developed at NASA Johnson Space Center by Leslie J. Quiocho and Robert Bailey. The Trick software package already contains the information to model the seven-joint Robotic Research Arms, the Space Shut- tle RMS, and the full Riceobot with base and two arms. Data modules provided by Trick allow the user to build customized robots with many different types of sensors, joints, and links. The software can currently model a joint failure by locking the joint. Our research is expanding the capabilities of the software to model runaway joints, fault detection schemes, and fault toler- ance algorithms. The flexibility of the software allows the failure analysis developed in this paper to be extended to a variety of different robots. 2 Robotic Fault Tree Analysis 2.1 Analysis Technique Fault Tree Analysis (FTA) is a deductive method in which failure paths are identified by using a fault tree drawing or graphical 6 2 ROBOTIC FAULT TREE ANALYSIS Table 1: Fault Tree Analysis Symbols Symbol Function AND gate All inputs required to produce output event. OR gate Any one input event causes the output event. Rectangle A malfunction which results from a combination of fault events through logic gates. Diamond A fault event for which the causes are left undeveloped. Circle A basic fault event. This includes component failures whose frequency and failure mode are known. Triangle A suppressed tree. The tree is detailed in another figure. representation of the flow of fault events [2]. FTA is a well- known analysis technique often used in industry for computer control systems and large industrial plants. Each event in the tree is a component failure, an external disturbance, or a system operation. The top event is the undesired event being analyzed and, in this research, is the failure of the entire robot. The events are connected by logic symbols to create a logical tree of failures. Some of the basic symbols are explained in Table 1. The explanation of the FTA technique in [2] promotes a top down development of the fault tree. The top event is broken down into primary events that can, through some logical com- bination, cause the failure at the top. This process is repeated to deeper levels until a basic event or an undeveloped event is reached. Some conditions or causes may be left undeveloped if the probability that they will occur is small enough to be ig- nored. 2.2 Failure Propagation/Probability Analysis Once the fault trees are built, the information available may be enhanced by a quantitative analysis of the failures. Failure rates are assigned to each input event and propagated up the tree based on the rules of the connecting logic gates. The output of an OR gate is the sum of the inputs. The resulting probability of the combined input events is greater than the probability of an individual input event. The output of an AND gate is the product of the inputs. The resulting probability of all the events occurring is thus less than the probability of any one occurring. The AND gates represent some form of redundant measurement or capability and are more desirable in the fault tree as the probability of a failure decreases through the combination of lower level events. A Markov or semi-Markov model approach to probability anal- ysis can also be developed based on the PAWS/STEM [3] and CARE III [12] reliability analysis packages. These packages can analyze simpler fault trees as well as Markov chains, but they are not necessarily optimized to handle the simpler structures 2.3 Fault Tree Pruning 7 [10]. These analysis tools were also not designed for robotics and thus would not take advantage of some of the commonality within robot structures. We will include some of the advan- tageous aspects of using Markov models in our CLIPS-based expert system, Robo-MEDIC. A quantitative analysis provides a measure of the overall chance of a complete failure for each robot. The structure provided by the fault trees organizes the probabilities appropriately for the robot system and provides a simple map of how the probabilities relate to each other. Using the trees, robots of significantly different origin and structure can be compared for fault tolerant abilities and survivability. The integrated Robo-MEDIC expert system will provide diagnostic capabilities by using the fault trees and will alert the operator of an impending failure. It can be used for off-line comparisons of robots or for suggesting possible corrective actions to an operator and the low-level robot controller during real operations. 2.3 Fault Tree Pruning A suggested drawback of FTA is that there is no way to ensure that all the causes of a failure have been evaluated [2]. The designer tends to pick out the important or most obvious events that would cause a given failure. However, the events that are not modeled normally have a low probability of occuring and can be ignored or treated as a basic event without overly biasing the analysis. Several failures may also be interconnected creating lateral branches or cycles in the fault tree. In some robots, one mo- tion at a joint may be coupled with another motion such that failure of either motion causes the failure of the other. It is also difficult to determine the relationships between some failures. For example, the failure of all the internal feedback sensors at the elbow joint of a robot may make the robot blind to the el- bow position. The elbow has not actually failed, but the robot is unable to detect the results of any commands sent to the elbow. Thus, the sensor malfunction does not contribute to an elbow failure specifically but may cause a failure of the entire robot. Relationships like these make the tree complex and difficult to understand. These problems can be overcome by working to simplify the tree. In the case of the coupled motions, the two failures can be considered as one with twice as likely a probabil- ity of occurring. 2.4 Riceobot Fault Trees To provide a foundation for the analysis of general robots, we have chosen to analyze the arm of the Rice University Riceobot. The arm has eight degrees of freedom: three motions in the shoulder (z translation, pitch, and yaw), two motions in the elbow (roll and pitch), and three motions in the wrist (roll, pitch, and yaw). The results obtained from the Riceobot apply to most general robots especially since the Riceobot has a wide variety of commonly used link, joint, and motor arrangements. Overall Robot Failure Several fault trees have been developed and a few are reproduced in the following pages. The top event is obviously the failure of the robot (Figure 2). The primary causes of a robot failure are 8 2 ROBOTIC FAULT TREE ANALYSIS Failure of Robot Failure of Computer System Power Failure External Failure of Shoulder Failure of Elbow Failure of Wrist Failure of Gripper Figure 2: Top Level Fault Tree for Entire Riceobot. power failure, computer system failure, or a combination of fail- ures of the joints. If the robot is fault tolerant, it can withstand the failure of several joints. By stablizing the faulty motion or joint in some manner (such as locking the joint), it is possible that the other motions can still provide some functional capabil- ity to the robot. This ability results in the AND gate combining the joint failures in Figure 2 and decreases the probability of a failure of the robot. Joint Failures The Riceobot has two directly driven motions: the shoulder z- direction motion and the wrist roll motion (Figure 3). The fault trees for these motions are quite simple since only the failure of External Gear-Train Failure Failure of Wrist Failure of Wrist Pitch/Yaw External Failure of Wrist Roll Link Breaks Chain/ Cable Loose External Chain/ Cable Snap Cable Guide Wheel Loose or Off Connector Failure Jacket Unscrews Cable Comes Free of Clamp Strip Threads Failure of Universal Joint Failure of Motor Failure of Motor Joint Link Becomes Stuck Joint Breaks Joint Clamp to Motor Loosens Figure 3: Sub-Level Fault Tree for Wrist System. 2.4 Riceobot Fault Trees 9 the motor plays an important role in the failure of the motion. The other motions of the Riceobot depend on some form of gear- train assembly to allow the spatially separated motor to drive the joint. Failure of the gear-train can be caused by basic events as simple as a loosening of the chain or cable. Motor and Sensor Failures The probability of a motor failure is dependent on the type of motor used. The Riceobot contains both brushless DC and step- per motors. Each motor also has a gear box which may fail due to gear slippage or wear. A power failure affects all motors as well as any other electrically driven parts in the robot, but each motor could lose power separately if its specific power cables break. A motor failure could conceivably also be the cause of a sensor failure when sensors are mounted on the shafts of the motors. Sensors are also affected by incorrect calibration and external noise or vibrations. (See Figure 4.) Computer System Failures The computer system of the Riceobot consists of three main parts: (1) amplifiers which read from the optical encoders and drive the motors, (2) servo control chips which store informa- tion about the different motors and convert the desired angles into currents for each motor, and (3) an on-board host computer which is programmed in C and computes the desired angles for the desired motions (Figure 5). These three parts each contain at least one board filled with TTL chips, capacitors, power tran- sistors, resistors, and other analog and digital circuit parts. A failure of any one of these parts may not cause a failure of the entire board; but if a board did fail, the robot would be unable to function. The robot cannot withstand the failure of all the servo controllers or all the amplifiers because it would no longer be able to communicate with the joints. External Gear-Train Failure Failure of Wrist Failure of Wrist Pitch/Yaw External Failure of Wrist Roll Link Breaks Chain/ Cable Loose External Chain/ Cable Snap Cable Guide Wheel Loose or Off Connector Failure Jacket Unscrews Cable Comes Free of Clamp Strip Threads Failure of Universal Joint Failure of Motor Failure of Motor Joint Link Becomes Stuck Joint Breaks Joint Clamp to Motor Loosens Figure 3: Sub-Level Fault Tree for Wrist System. 10 2 ROBOTIC FAULT TREE ANALYSIS Sensor Failure Local Power Lines Fail Incorrect Callibration Sensor Logic Board Failure External Damage/ Noise Failure of Motor Local Power Lines Fail Internal Motor Failure Gear Box Failure Gears Slip Wear on Gears Connection to Amp Failure Lose Cable Strip Connector for Cable Strip Faulty Figure 4: Sub-Level Fault Tree for Sensor System. Detecting a non-terminal failure in the computer system requires some form of testing circuitry or the ability to poll components to see if they are still alive. The IEEE standard 1149.1 Test Access Port may be incorporated into any VLSI chip on the boards and could be used for active testing. Radiation hard- ened circuits [15] should also be used in the computer system. Correction code bits can be used to check data transfers and could identify a bus failure if the bits were consistently wrong. 2.5 Derived Riceobot Fault Detection The qualitative analysis of these fault trees has proven useful in pointing out some of the limitations of the Riceobot in regards to fault detection and fault tolerance. With only one sensor at each joint, the Riceobot represents the worst case scenario for detecting sensor and joint failures. The only option available to the fault detection software is to compare the sensed angles with the calculated desired angles. After accounting for a predeter- mined threshold to mask any precision errors in the calculations or sensing equipment and possibly adjust for load effects, any difference between the sensed and desired values must be con- sidered the result of a failure. The computer is, however, unable to differentiate between a sensor malfunction and an actual joint failure due, for example, to a frozen motor. The computer must therefore shut down the joint and proceed with fault tolerance schemes based on a new model of the robot with fewer possible motions. The Riceobot’s fault detection capabilities could be improved by drawing on the vision system to determine the joint angles. With the computer’s calculations and the sensor reading, this third estimate of the angle would help distinguish between a sensor failure and a real joint failure. The Riceobot would still be able to function in the presence of one sensor failure. Using the vision system for this task, however, increases the load on the image processing software and may hinder the system’s ability 11 Sensor Failure Local Power Lines Fail Incorrect Callibration Sensor Logic Board Failure External Damage/ Noise Failure of Motor Local Power Lines Fail Internal Motor Failure Gear Box Failure Gears Slip Wear on Gears Connection to Amp Failure Lose Cable Strip Connector for Cable Strip Faulty Figure 4: Sub-Level Fault Tree for Sensor System. to perform its normal vision tasks. 3 Robot Fault Detection The fault trees give an idea of the interaction between failures in a system. The trees also provide a map of alternate paths for detecting faults or bypassing failures. In order to expand on this information and to show how modeling errors or other un- certainties affect fault detection, we need to simulate the robot and the fault detection algorithm. Because of the Riceobot’s lack of sensors and the complexity needed for its fault detec- tion algorithm, we are initially simulating fault detection using a computer modeled planar, four link robot [7]. The current program will need to be expanded extensively for the Riceobot and will be accomplished by implementing the fault detection routine in the CLIPS expert system as part of Robo-MEDIC. The four link robot is essentially just four cylinders placed end- to-end. All joints are rotational and move in the same plane. A simulated optical encoder and tachometer were added for each joint. The fault tree for this robot is relatively simple (Figure 6). We have not included the possibility of link breakage or global power failure in the simulation. The motors are in essence direct drive with no gear trains and fail only in a locked mode. This conditions pruned each joint subtree down from the complexity of the Riceobot trees to an easily simulated failure situation. It is interesting to note that it is the fault detection software which allows the joint to survive in the presence of a single sensor failure thus creating the AND gate under each motor failure in the tree. If both sensors at a joint fail, the host computer is blind to that joint and the fault detection routine forces a motor failure to prevent the joint from moving too far without computer supervision. Thus, the dual sensor failure subtree is a cause of the motor failure event for each joint. 12 3 ROBOT FAULT DETECTION Failure of Computer System Card Cage Failure External Host Computer Failure Amplifier Board Failure Servo-Control Board Failure Sensor Failures Failure of Components bonding,status indicators, etc., fail Sensor Failure (n) Sensor Failure (2) Sensor Failure (1) Failure of Components transistors, capacitors, or resistors, fail TTL chips, processors, ROM, RAM fail soldering, chip sockets, etc, fail bonding, status indicators, sockets fail Sensor Failures Failure of Components Figure 5: Sub-Level Fault Tree for Computer System. 3.1 Fault Detection Simulator The structure of the simulator is shown in Figure 7. The flow of information is from the simulated host computer through the robot and then the sensors to the fault detection program and finally back to the host computer. The host computer uses the desired angles computed by a planner routine and the estimated present position of the robot derived from the sensors to calcu- late the torque necessary to move each link to its desired desti- nation. The controller is a standard PD computed torque type controller. The robot routine then takes the calculated torques and determines the new position, velocity, and acceleration for each joint. The optical encoders estimate the positions by trun- cating the value of each angle based on each encoder’s precision. The tachometers pass the velocities through a first order filter based on a predetermined motor lag time. The sensors are mod- ules from the Trick simulation package and represent our initial efforts at integration with Trick. These estimates of the an- gles and velocities are passed into the fault detection procedure which checks for failures and either passes to the host what it considers good estimates of the position, velocity, and accelera- tion or signals the motor of a joint with two bad sensors to shut down. 3.2 Host Computer Model The simulated host computer uses the following dynamics equa- tion as a model for the robot: τ = [M (θ)]θ̈ + N (θ, θ̇), (1) where τ is the joint torque vector, [M ] is the inertia matrix, and N is the Coriolis and centrifugal torque vector. The [M ] matrix and N vector are computed based on the estimated angles from the sensors. Since the robot is planar, gravity is orthogonal to 3.2 Host Computer Model 13 Failure of Computer System Card Cage Failure External Host Computer Failure Amplifier Board Failure Servo-Control Board Failure Sensor Failures Failure of Components bonding,status indicators, etc., fail Sensor Failure (n) Sensor Failure (2) Sensor Failure (1) Failure of Components transistors, capacitors, or resistors, fail TTL chips, processors, ROM, RAM fail soldering, chip sockets, etc, fail bonding, status indicators, sockets fail Sensor Failures Failure of Components Figure 5: Sub-Level Fault Tree for Computer System. the plane of motion and there are no resultant gravity torques to consider. Friction is also ignored in this model. The PD controller for this model becomes: τ = [M (θ)]{θ̈d + [KP ](θd − θ) + [KD](θ̇d − θ̇)} + N (θ, θ̇). (2) The matrices [KP ] and [KD] are the position and derivative gains, respectively, and are used to control tracking and steady state errors by feedback control. For critical damping, the gains become: Failure of Joint 0 Failure of Robot Failure of Joint 1 Failure of Joint 2 Failure of Joint 3 Failure of Motor Failure of Motor Parts Failure of Tachometer Failure of Encoder Figure 6: Four Link, Planar Robot Fault Tree 14 3 ROBOT FAULT DETECTION θ e θ t Instigate Failures Fault Detection Encoders Tachs Robot (calc real accelerations, velocities, and angles) Host Computer (calc torques)Planner Report of Failures Failure Information fail motor signal τ θ d θ d θ d θ θ θ θ d θ d θ d θ r θ r Figure 7: Fault Detection Simulator Flow Chart [KD] = 2ω, [KP ] = ω 2, (3) where ω is the natural frequency input by the user. The natural frequency is typically set to 1 for most runs of the simulator. A higher natural frequency would increase the effects of the gains in the feedback control. 3.3 Robot Model The simulated robot is a procedure which takes the computed torque from the simulated host computer and determines the resulting four robot angle accelerations based on the equation: θ̈ = [M̂ ]−1τ − [M ]−1(N̂ ). (4) Here, [M̂ ] and N̂ are the inertia matrix and Coriolis and cen- trifugal torque vector as before but are now based on the real robot angles instead of the sensed angles. The matrices have also been injected with a small constant error to simulate the effects of a load on the robot. The joint angle, θ, and its first derivative are estimated by the equations: θ̇i = θ̇i−1 + (Δt)θ̈i, (5) θi = θi−1 + (Δt)θ̇i. (6) If a motor failure has occurred, θ̈ and θ̇ are set to zero to simulate the effects of a locked motor. The position thus remains con- stant. Only the locked motor is currently simulated, but other failure modes could result in runaway motors or free-spinning motors. The robot’s position, velocity, and acceleration calculated in this procedure are sent to the sensor routines. The robot position is also sent to the graphics simulator which displays the motion on the screen. This is the same graphics program used by the Trick simulation package. 3.4 Fault Detection Capabilities 15 3.4 Fault Detection Capabilities 3.4.1 Failure Modes If a sensor breaks and the failure goes undetected, the host com- puter will be performing its calculations using erroneous infor- mation. In this simulator, the encoders break in a frozen mode continuously reporting the last value read before the failure. The tachometers fail by continuously reporting zero velocities and thus constant positions. With these failure modes, the host sees the error between the sensed angle and the desired angle grow for the joint with the faulty sensor, and the control equa- tions increase the appropriate output torque to the robot to try and compensate for the error. The joint with the faulty sensor swings wildly off course because the host keeps trying harder and harder to get the broken sensor value to match the desired value. Since the calculations for all the joints are based on knowledge of where the other joints are located, all of the other output torques are also computed incorrectly and the joints all stray from their desired paths. When a motor fails, it locks in position and the joint is then unable to move. If a motor failure goes undetected, the sensors are still reading the correct information. In reality, the motor failure would probably result in a sensor failure as well, but the result would still be that both sensors are reporting a constant joint angle. The control equations try to push the broken joint closer to the desired value but the frozen motor does not respond to the torques. Since the sensors are still reporting the actual position of the joint, all the other calculations are based on cor- rect data and the other joints can continue with their normal motions. The plan must be modified, however, to get the end effector to its desired location. 3.4.2 Thresholds These two undetected failures reveal the importance of getting accurate sensor readings and of detecting a sensor failure quickly. A frozen motor is not as critical a failure in most cases and can be dealt with at a more leisurely pace. Since the sensors are not perfectly accurate, an acceptable threshold for the error between sensor reading and desired value must be chosen. Unfortunately, even during normal operation, the error between the actual angle and the desired angle can be relatively large especially at the beginning of a run before the controller has time to bring the error under control. Choosing the maximum error found during a failure-free run results in a threshold that is so large, it may take several time steps to notice the error from a broken sensor. By the time the failed sensor is detected, the robot controller has already been infected with the erroneous information and the robot is either off course or has damaged itself. Fortunately, the error between the angles recorded by the two sensors during normal operation is very small even after integrat- ing the tachometer reading to get the angular position. Fortu- nately, modeling errors and errors induced by unpredicted loads would affect both sensors in a similar manner. A tight threshold can be chosen for a comparison of the two sensed positions. If this threshold is exceeded, the fault detection software assumes that one of the sensors has failed and appropriately chooses one as the working sensor from which to take the recorded data. The larger thresholds from the typical error between the sensed and desired angles are still monitored, however. The large thresholds 16 3 ROBOT FAULT DETECTION provide a means of checking for a motor failure. The psuedocode for these checks is reproduced below. The angle θd and its derivatives are the desired values. The variables θt, θ̇t, and θ̈t are the values derived from the tachometer reading. The results based on the encoder are θe, θ̇e, and θ̈e. Finally, θ and its derivatives are the values sent to the robot controller. If ((encoder working) and (tachometer working)){ θ = θe, θ̇ = θ̇t, θ̈ = θ̈t If ((|θt − θe|) >= threshold){ if (encoder working){ tachometer = failed θ = θe, θ̇ = θ̇e, θ̈ = θ̈e }else{ encoder = failed θ = θt, θ̇ = θ̇t, θ̈ = θ̈t } }else{ if ((|θd − θt|) >= tachometer-threshold) tachometer = failed if ((|θd − θe|) >= encoder-threshold) encoder = failed } } If ((tachometer == failed) and (encoder != failed)){ if ((|θd − θe|) < encoder-threshold){ θ = θe, θ̇ = θ̇e, θ̈ = θ̈e }else{ encoder = failed motor = failed send stop motor signal to robot } } If ((encoder == failed) and (tachometer != failed)){ if ((|θd − θt|) < tachometer-threshold){ θ = θt, θ̇ = θ̇t, θ̈ = θ̈t }else{ tachometer = failed motor = failed send stop motor signal to robot } } Choosing which sensor has failed and which is still working when the tight tolerance is exceeded is the most difficult task. Intu- itively, one would expect the sensor with a reading closer to the desired value to be the working sensor and would switch to obtaining all the information from that sensor. However, ex- perience has shown that the fault detection software choses the correct sensor only when the desired values are increasing. If the desired angles are decreasing in value, it consistently picks the failed sensor as the working one. This problem is a result of the time it takes the controller to bring the errors under control and the failure modes for the sen- sors. Both the encoders and the tachometers fail by reporting a constant angular position either directly or by integration of a zero velocity. First, let us assume the sensors always read less than the desired value. If a sensor fails and gets stuck at a spe- cific value while the desired values are increasing, the error will 3.4 Fault Detection Capabilities 17 Table 2: Sensor Failure Detection Tests Ordering Encoder Fails Tachometer Fails θd < θt, θe largest error largest error (2) (2) θt, θe < θd smallest error smallest error (1,2) (1,2) θt < θd < θe largest error smallest error (1,2) θe < θd < θt smallest error largest error (1,2) grow and the fault detection routine should take the angle in- formation from the sensor that reads closer to the desired value. However, if the sensor fails while the desired values are decreas- ing, the desired values are approaching the failed value. The error starts decreasing and the surviving sensor is often the one whose absolute error is larger. (See Table 2 for listing of results.) The opposite relationships hold if both sensors are reading val- ues greater than the desired angle. A failed sensor would then have the smaller error during an increase in desired angles and the larger error during a decrease in desired angles. The various sensor failure situations that arise in the presence of increasing desired values and the relative size of the error for the surviving sensor are listed in Table 2. The cases in which the initial, naive algorithm is successful in choosing the survivor are marked with an “1”, and the cases in which the second algorithm is successful are marked with a “2”. The table represents half of the possible cases. The table for decreasing desired values would look similar with the equivalent number of successes for each algorithm as in Table 2. By checking whether the desired values are increasing or de- creasing and performing the appropriate comparisons to choose the surviving sensor, the algorithm’s success rate for the cases listed in the table increases from 50% to 75%. The algorithm still has a problem with the cases where the sensors are reading on opposite sides of the desired value. The fault detection pro- gram picks a failed tachometer as the survivor if the tachometer value is trailing a trend of decreasing desired values while the encoder is preceding the desired values. Similarly, the fault de- tection program should pick a failed encoder as the surviving sensor during a trend of increasing desired values if the encoder value is preceding the desired value by a large enough margin while the tachometer is trailing the desired value. We were un- able to demonstrate this second error as the encoder value was close enough to the desired value that when the encoder failed, the desired value passed the failed value in the next time step. The fault detection scheme was able to guess correctly in this situation. In general, our simple fault detection simulator is capable of de- tecting for each joint a single sensor failure, a single sensor failure followed by a motor failure, or a motor failure. The simulator will eventually detect a second sensor failure and will catch the single failures it has missed, but it has allowed enough erroneous sensor readings through to the controller that other joints have been knocked off course and fail as well. In order to improve the fault detection algorithm, we must switch from the hard-coded voting scheme presented above to other forms of analytical re- 18 4 CONCLUSIONS AND FUTURE WORK dundancy which use filters [14, 11], adaptive thresholds [8], or parity relations [5]. Willsky [14, 5] gives a good overview of the various methods of analytical redundancy. Most of the work in this area has been focused on failure detection in aircraft, power generation system and other mechanical systems. Unfor- tunately, the amount of uncertainty and modeling errors present in most robotic control systems makes several of these methods inaccurate. The generation of residuals using parity relations is one example of a method which would be unsuitable for robotic applications. [8]. 4 Conclusions and Future Work In this paper we have presented new results in fault tree analysis and fault detection for robot manipulators. This research sets the stage for significantly enhanced activity in fault tolerance for robotics. Once a failure can be detected and isolated, a fault tolerant expert system like Robo-MEDIC can proceed with the appropriate actions to make use of the existing robot structure, redundancy, and alternate paths. There already exist a variety of computer fault tolerance schemes which can provide a starting point for creating these structurally independent robotic fault tolerance algorithms. The fault tree analysis for the Riceobot has proven useful in pointing out some trouble spots for fault detection and fault tolerance. Even without a quantitative analysis, the importance of certain components and the severity of different failures are revealed in the fault trees. For robots, the good health of the internal sensors is shown to be extremely desirable. Erroneous data from even one sensor at a joint can cause the whole robot to deviate drastically from its course if the failure is not detected quickly. Without the sensors, the robot also loses much of its ca- pability to detect faults. Developing methods for early detection of sensor malfunctions thus has a high priority in this research. By simulating relatively simple fault detection situations, we are gaining a better understanding of how to satisfy this need for early detection. Our simulator has shown that to avoid false alarms due to modeling errors and noise we must implement large thresholds which let some failures go undetected for too long. Other relationships must be developed concerning the in- formation available in order to improve the fault detection algo- rithms. Once these schemes have been perfected, we will be able to embed the algorithms in an expert system and integrate the simulation into the Trick simulation package to create a more flexible fault detection and fault tolerance simulator. The analysis of the fault trees in this paper will be useful in creating fault tolerant algorithms. Through an analysis of the structures, fault tolerance schemes must be developed which will attempt to maintain the health of the internal nodes in the pres- ence of failures in their children. Detection schemes and fault tolerant algorithms will be tested on the Trick robotic software testbed. Acknowledgments This work was supported in part by the National Science Foun- dation under grants MIP-8909498 and MSS-9024391 and in part REFERENCES 19 by a Mitre Corporation Graduate Fellowship and an NSF Grad- uate Fellowship. References 1. Bailey, R. W. and Quiocho, L. J., Trick Simulation En- vironment Developer’s Guide, NASA JSC Automa- tion and Robotics Division, Houston, TX, Beta-release edi- tion, February 1991. 2. Bloch, H. P. and Geitner, F. K., An Introduction to Machinery Reliability Assessment, Van Nostrand Reinhold, New York, NY, 1990. 3. Butler, R. W. and Stevenson, P. H., “The PAWS and STEM Reliability Analysis Programs,” NASA Technical Mem- orandum 100572, NASA Langley Research Center, Hamp- ton, VA, March 1988. 4. Chean, M. and Fortes, J. A., “A Taxonomy of Reconfig- uration Techniques for Fault-Tolerant Processor Arrays,” IEEE Computer, 23(1):55–67, January 1990. 5. Chow, E. Y. and Willsky, A. S., “Analytical Redundancy and the Design of Robust Failure Detection Systems,” IEEE Transactions on Automatic Control, AC- 29(7):603–614, July 1984. 6. Giarratano, J. and Riley, G., Expert Systems: Prin- ciples and Programming, PWS-Kent Publishing Co., Boston, MA, 1989. 7. Hamilton, D. L., “A Simulation of Robot Motion Control,” Advisors: I. D. Walker and J. K. Bennett, Rice Univer- sity, Department of Electrical and Computer Engineering, Houston, Texas, April 1991. 8. Horak, D. T., “Failure Detection in Dynamic Systems with Modeling Errors,” AIAA Journal of Guidance, Control, and Dynamics, 11(6):508–516, November- December 1988. 9. Maciejewski, A. A., “Fault Tolerant Properties of Kinemat- ically Redundant Manipulators,” In 1990 IEEE Confer- ence on Robotics and Automation, pages 638–642, Cincinnati, OH, May 1990. 10. Martensen, A. L. and Butler, R. W., “The Fault-Tree Com- piler,” NASA Technical Memorandum 89098, NASA Langley Research Center, Hampton, VA, January 1987. 11. Merrill, W. C., DeLaat, J. C., and Bruton, W. M., “Ad- vanced Detection, Isolation, and Accommodation of Sensor Failures - Real-Time Evaluation,” Journal of Guidance, Control, and Dynamics, 11(6):517–526, November- December 1988. 12. Stiffler, J. J. and Bryant, L. A., “CARE III Phase II Re- port - Mathematical Description,” NASA Contractor Report 3566, NASA Langley Research Center, Hampton, VA, 1982. 13. Tesar, D., Sreevijayan, D., and Price, C., “Four-Level Fault Tolerance in Manipulator Design for Space Operations,” In NASA First International Symposium on Measure- ment and Control in Robotics, volume 3, page J3.2.1, Houston, TX, June 1990. 14. Willsky, A. S., “A Survey of Design Methods for Failure De- tection in Dynamic Systems,” Automatica, 12:601–611, 1976. 15. Winokur, P. S., “Radiation-Hardened Circuits for Robotics,” Proc. Fourth Topical Meeting on 20 REFERENCES Robotics and Remote Systems, pages 311–315, Febru- ary 1991. 16. Wu, E., Diftler, M., Hwang, J., and Chladek, J., “A Fault Tolerant Joint Drive Systems for the Space Shuttle Remote Manipulator System,” In 1991 IEEE Interna- tional Conference on Robotics and Automation, pages 2504–2509, Sacremento, CA, April 1991. 
work_ckevfqmvd5eidkjrn3rfy5b4eq	----	ReliAble dependency arc recognition Expert Systems with Applications 41 (2014) 1716–1722 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a ReliAble dependency arc recognition 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.08.070 ⇑ Corresponding author. Tel.: +86 13936137628. E-mail address: tliu@ir.hit.edu.cn (T. Liu). Wanxiang Che, Jiang Guo, Ting Liu ⇑ School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001, China a r t i c l e i n f o a b s t r a c t Keywords: Natural language processing Syntactic parsing Dependency parsing RADAR Binary classification We propose a novel natural language processing task, ReliAble dependency arc recognition (RADAR), which helps high-level applications better utilize the dependency parse trees. We model RADAR as a bin- ary classification problem with imbalanced data, which classifies each dependency parsing arc as correct or incorrect. A logistic regression classifier with appropriate features is trained to recognize reliable dependency arcs (correct with high precision). Experimental results show that the classification method can outperform a probabilistic baseline method, which is calculated by the original graph-based depen- dency parser. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction As a fundamental task of natural language processing, depen- dency parsing has become increasingly popular in recent years. It aims to find a dependency parse tree among words for a sentence. Fig. 1 shows an example of dependency parse tree for a sentence, where sbj is a subject, obj is an object, etc. (Johansson & Nugues, 2007). Dependency parsing are widely used: in biomedical text mining (Kim, Ohta, Pyysalo, Kano, & Tsujii, 2009), as well as in tex- tual entailment (Androutsopoulos & Malakasiotis, 2010), informa- tion extraction (Wu & Weld, 2010; Banko, Cafarella, Soderland, Broadhead, & Etzioni, 2007) and sentiment analysis (Meena & Pra- bhakar, 2007). The performance of dependency parsing has increased recently (Kübler, McDonald, & Nivre, 2009). However, when we migrate dependency parsing systems from laboratory demonstrations to high-level applications, even the best parser available today still encounter some serious difficulties. First of all, parsing performance usually dramatically degrades in real fields because of domain migration. Secondly, since every parser inevitably will make some mistakes during decoding, out- puts from any dependency parser are always fraught with a variety of errors. Thus, in some high-level applications which expect to use correct parsing results, it is extremely important to be able to pre- dict the reliability of the auto-parsed results. If these applications just use correct parsing results and ignore incorrect results, their performances may be improved further. For instance, if an entity relation extraction (a kind of information extraction) system, which depends on parsing results heavily (Zhang, Zhang, Su, & Zhou, 2006), only extracts relations from correct parsing sentences, then the system can extract more accurate relations and import less wrong relations through incorrect parsing results. Although some implied relations in those incorrect parsing sentences are missed, these missing relations may be extracted from other sen- tences that can be parsed correctly while zooming in the data to the whole Web. Most large-margin based training algorithm for dependency parsing output models that predict a single parse tree of the in- put sentence, with no additional confidence information about the correctness of it. Therefore, an interesting problem is how to judge a parsing result as correct or not. However, it is difficult to obtain a parse tree in which all sub-structures are parsed cor- rectly. CoNLL 2009 Shared Task results show that only about 40% English and 35% Chinese sentences can be parsed complete cor- rectly (Hajič et al., 2009b). Some previous studies have ad- dressed the problem to recognize reliable parsing results (Reichart & Rappoport, 2007; Dell’Orletta & Venturi, 2011; Kawa- hara & Kurohashi, 2010; Ravi, Knight, & Soricut, 2008). A parsing result is reliable when, the result is correct with high precision. However, all these studies focus on judging if the parsing results of a whole sentence are reliable or not, which can cause the fol- lowing problems: 1. The reliable parsing results may still include some wrong pars- ing sub-structures. Different applications need different key sub-structures, such as backbone structures that are keys for semantic role labeling (Gildea & Jurafsky, 2002) and branch structures are important for multiword expression (Sag, Bald- win, Bond, Copestake, & Flickinger, 2002). If these key sub- structures are parsed incorrectly, even though the whole sen- tence is parsed with a high reliability, the tiny errors will be still harmful to these given applications. This problem results in a low precision. http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2013.08.070&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.08.070 mailto:tliu@ir.hit.edu.cn http://dx.doi.org/10.1016/j.eswa.2013.08.070 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. An example of dependency parse tree. W. Che et al. / Expert Systems with Applications 41 (2014) 1716–1722 1717 2. The unreliable parsing results may include some useful sub- structures but should not be discarded totally. For instance, extracting entity relations is possible if the parse tree path is correct between two entities despite other parts of the sentence being incorrectly parsed. Discarding unreliable parsing sen- tences can result in a low recall. Therefore, dependency arcs, novel reliability measuring objects for dependency parsing are proposed. A reliable dependency hap- pens when a word can find its parent and label the dependency relation between them correctly with high precision. Once all reli- able dependency arcs in a sentence are found, the corresponding parse paths or sub-trees from them can be mapped out. These reli- able sub-structures can be used according to the needs of different applications. In paying attention to reliable parts and ignoring unreliable ones in a sentence, the precision of applications can be improved. Meanwhile, when the number of reliable sub-structures is more than that extracted from reliable whole sentences, higher recall can be obtained. The problem of ReliAble Dependency Arc Recognition (RADAR) can be regarded a binary classification problem. The positive exam- ples are the correctly predicted arcs and the others are the negative examples. Thus, the problem can be converted to find appropriate classifiers and proper features. Different from normal binary classi- fication problems, the data are not balanced for RADAR. For the state-of-the-art dependency parser, the LAS (Labeled Attachment Score) can achieve about 80% in Chinese data set and 90% in Eng- lish data set (Hajič et al., 2009b), which means that the ratio of the number of correct dependency arcs to the number of incorrect dependency arcs is 4:1 for Chinese and 9:1 for English. Aside from learning from the imbalanced data, how to evaluate RADAR is an- other issue. The normal evaluation methods based on accuracy are not suitable for the problem. If an incorrect dependency arc is rec- ognized as a correct arc, the cost is larger than the opposite sce- nario. In addition, the classification accuracy would not be a suitable evaluation metric in an imbalanced scenario. Therefore, there is a need to find more appropriate evaluation criteria. The rest of the present paper is organized as follows. Section 2 presents related work. Section 3 describes the proposed method. Section 4 discusses the present experimental setting and results. We conclude and set the direction of the future work in Sections 5 and 6 respectively. 1 When the size of a set is 1, the accuracy of a sentence can be predicted. 2. Related work To the best of our knowledge, Yates, Schoenmackers, and Etzi- oni (2006) was the first work to address explicitly the parsing reli- ability recognition problem. They detected erroneous parses using web-based semantics. In addition, an ensemble method based on different parsers trained on different data sampled from a training corpus to select high quality parsing results was proposed as well (Reichart & Rappoport, 2007). Dell’Orletta and Venturi (2011) was another study that detect reliable dependency parses with some heuristic features. Kawahara and Uchimoto (2008) classified sen- tences into two classes, reliable and unreliable, with a binary clas- sifier. Ravi et al. (2008) predicted the accuracy of a parser on sets of sentence by fitting a real accuracy curve with linear regression algorithm.1 However, all these works focused on recognizing reliable parsing results of whole sentences and caused corresponding prob- lems for some applications as discussed in Section 1. Although the parsing reliability recognition of whole sentences can be used in Active Learning (Settles, 2010) or Semi/Un-super- vised Learning (Goldwasser, Reichart, Clarke, & Roth, 2011), recog- nizing sub-structures reliability is also useful. For instance, some studies (van Noord, 2007; Chen, Kawahara, Uchimoto, & Zhang, 2008, 2009) used sub-trees or word pairs extracted from a large auto-parsed corpus to help the dependency parser. However, the confidence of a sub-tree or a word pair is only expressed by its count that appears in the corpus. Therefore, their methods may be biased toward frequently appearing sub-trees or word pairs, which may be incorrect, and penalizes the sparse but correct ones. The studies most relevant to ours are done by Atserias, Attardi, Simi, and Zaragoza (2010) and Avihai Mejer (2012). They both re- ported the similar problem with ours. Atserias et al. (2010) shows how to use the probability scores that a transition-based parser normally computes, in order to assigning a confidence score to parse trees. They assign such score to each arc and the active learn- ing application uses the worst. Another independent work done by Avihai Mejer (2012) describes several methods for estimating the confidence in the per-edge correctness of a predicted dependency parse. The best method they confirmed in their study is based on model re-sampling, which is inefficient. Our work differs in that we proposed a novel supervised approach which makes use of additional information as the features for learning models. 3. Method description This section introduces the dependency parsing model and a method to estimate the probability of each dependency arc. A bin- ary classification method to recognize reliable arcs follows. Besides a classifier, the classification method includes three sorts of fea- tures and a process to construct training data. 3.1. Graph-based dependency parsing Given an input sentence x = w1 . . . wn, a dependency tree is de- noted by d ¼fðh; m; lÞ : 0 6 h 6 n; 0 < m 6 n; l 2Lg, where (h, m, l) represents a dependency arc wh ? wm whose head word (or father) is wh and modifier (or child) is wm with a dependency label l, and L is the set of all possible dependency relation labels. The artificial node w0, which always points to the root of the sentence, is used to simplify the formalizations. Then, an optimal dependency tree d̂ is determined based on x: d̂ ¼ arg max d Scoreðx; dÞ Recently, graph-based dependency parsing has gained interest due to its state-of-the-art performance (Kübler et al., 2009). Graph- based dependency parsing views the problem as finding the highest scoring tree from a directed graph. Based on dynamic programming decoding, it can find efficiently an optimal tree in a huge search space. In a graph-based model, the score of a dependency tree is fac- tored into scores of small parts (sub-trees): Scoreðx; dÞ¼ w � fðx; dÞ¼ X p # d Scoreðx; pÞ where f(x, d) refers to the feature vector and w is the corresponding weight vector, p is a scoring part that contains one or more depen- dence arcs in the dependency tree d. 1718 W. Che et al. / Expert Systems with Applications 41 (2014) 1716–1722 The present paper devotes to realize RADAR in graph-based parsing algorithm. The most intuitive method is to use the proba- bility of a dependency arc to denote its reliability. However, the graph-based parsing method is based on a discriminative model but not a generative model, therefore, we only can obtain the score of each arc (Score(x, p)) but not its accurate probability. To over- come this difficulty, a study of Koo, Globerson, Carreras, and Collins (2007)’s idea is used to estimate the probability of the dependency arc. Before estimating the arc probability, the probability of a dependency tree d, which can be obtained by exponentiating and renormalizing the score Score(x, d), has to be computed: Pðdjx; MÞ¼ expðScoreðx; dÞÞ Zðx; MÞ Zðx; MÞ¼ X d02T ðxÞ expðScoreðx; d0ÞÞ where M is the dependence parsing model, and Z(x;M) is a normal- ization factor. T ðxÞ is the set of all possible parse trees for the sentence. Given the conditional distribution P(djx;M), the probability of a dependency arc (h, m, l) is: Pððh; m; lÞjx; MÞ¼ X d02T ðxÞ:ðh;m;lÞ2d0 Pðd0jx; MÞ Note that both probabilities require a summation over the set T ðxÞ, which is exponential in sentence size n. For the sake of simplicity, we used the k-best dependency tree list in place of T ðxÞ, where k is set to 1000 in the experiments. 3.2. RADAR as binary classification In a dependency parse tree, all dependency arcs are classified into two classes: positive (the modifier and head words of an arc are extracted correctly and the syntactic relation is correct.2) and negative (otherwise). RADAR can be regarded as a binary classifica- tion problem naturally. For the classification problem, three ques- tions need to be answered: 1. Which classifier is used? 2. What features are extracted? and 3. How is the training data constructed? 3.2.1. Classifier The logistic regression classifier (Bishop, 2006) is used in the experiments. The main reason to select logistic regression is that it can estimate the probability of each class, which will be used as the criteria for RADAR later. In addition, logistic regression is fast for training and effective for predicting. The present study used L2- regularized logistic regression, which solves: arg min w kwk2 2 þ Cþ Xnþ fijyi¼þ1g logð1 þ eð�yi w T xiÞÞþ C� Xn� fijyi¼�1g logð1 þ eð�yi w T xiÞÞ where (xi,yi) is the ith training sample and w is the feature weight vector. C+ and C� are the penalties of constraints violation for posi- tive and negative classes respectively, which can be used to handle the imbalanced data problem. A larger C� can be set to prevent the classifier from tending to recognize incorrect arcs as correct ones (the majority class). 3.2.2. Features Besides a classifier algorithm, features are also important for a classification system. To predict whether a dependency arc is cor- 2 This is equivalent to find the correct head word and dependency relation for a word. rect or not, we defined three sorts of features. Below is the detail description of the features and intuitions. 1. Text-based features relate only to the original text of a sentence. � The length of a sentence: Longer sentences are harder to be parsed correctly than shorter sentences. We group the lengths into three buckets: with a bucket LS (long-sentence) for length 40+, a single bucket MS (middle-sentence) for 16– 40, and a single bucket SS (short-sentence) for 1–15. The length thresholds are tuning on the development data. � Number of unknown words: Words that never appear in the training set are unknown words. Sentences containing more unknown words are more unlikely to be parsed correctly. Since the number of unknown words usually ranges in a small interval, we didn’t group them info buckets. Therefore, each number is assigned a single bucket. Note that the two features indicated above have the same value over the words in the same sentence. � Unknown word property: The boolean feature represents whether a word is an unknown word. The parser perfor- mance is worse for an unknown word. � Current word: wm (current word at position m). � Bigrams of POS tags: tm�1tm (POS tags at position m � 1 and m) and tmtm+1. � Trigram of POS tag: tm�1tmtm+1. 2. Parser-based features are extracted based on parsing results. � The length of a dependency arc: Number of words between wm and wh (head word). As with the text-based ‘‘Length of Sentence’’ feature, we also attempted to group this feature into buckets. However, it did not help. So we just assign a single bucket for a single length. � Word collocation: wmwh (current word at position m and its head word). � POS tag collocation: tmth (POS of current word and its head word). � Dependency relation: lm (dependency relation label of cur- rent word and its head word). We have also considered the direction of the dependency arc with respect to the word, but it brings no improvement. � Combination of POS tag and dependency relation: tmlm. � The InBetween POS features (form of a POS trigram): tmtbth (POS of current word, of its parent, and of a word in between) � The Surrounding POS features (form of a POS 4-gram): tm�1- m�1tmthth�1,tm�1tmthth+1, tm+1tmthth�1,tm+1tmthth+1. � Combined features. The combination of some of the features above can prove very helpful. 3. Comparison among Parsers � Agreement of different parsers: Intuitively, if two parsers based on different learning schemes agree on one depen- dency arc, it will probably be a reliable one. In this paper, we use a transition-based parser (Nivre, 2006) as the refer- ence parser. If the reference parser products the same label- ing (including the head and dependency label) for the current word with the basic parser, this feature is set TRUE; otherwise FALSE. 3.3. Training process In this section, we will introduce how to train a RADAR model. Fig. 2 shows the training process flow. The core problem is how to construct a reliable arc training corpus. Here, n-fold validation method is used. At step , the traditional dependency parsing training corpus was divided into n folds. At each time, n-1 folds were selected to train a dependency parser with graph-based mod- Table 1 Number of sentences and dependency arcs in training, development, and test set and their corresponding performance. Lang Data set #{sent} #{arcs} LAS (%) Chinese Train 55,496 1,025,054 81.37 Dev 1,500 29,091 81.19 Test 3,000 56,786 81.43 Fig. 2. The process flow of training a reliable arc classifier. W. Che et al. / Expert Systems with Applications 41 (2014) 1716–1722 1719 els. Then the leaving fold was parsed with the parser at step and the obtained automatic parsing dependency arcs were split into positive (correct) and negative (incorrect) classes according to their comparison with the gold parse trees. The above steps were repeated n times and the RADAR training corpus was constructed at step . Finally at step , the reliable arc classifier was trained with the training corpus. English Train 39,060 837,340 89.72 Dev 1,325 29,024 88.36 Test 2,389 50,291 90.03 Table 2 The contribution of different sort of features on the development data of CoNLL2009- en and CDT. Lang Feature sort AUC-PR (%) Decrease Chinese All Features 94.89 N/A –Text-based 94.81 0.07% –Parser-based 93.06 1.75% –Comparison 93.90 0.99% English All Features 97.97 N/A –Text-based 97.95 0.02% –Parser-based 97.35 0.62% –Comparison 97.73 0.24% Table 3 AUC-PR and AUC-ROC measurements on the test data. Lang Classification Probabilistic AUC-PR (%) AUC-ROC (%) AUC-PR (%) AUC-ROC (%) Chinese 94.65 81.42 93.59 79.38 English 98.56 89.29 97.68 85.50 4. Experiments 4.1. Data set In order to evaluate the effectiveness of the our approach, we conducted experiments on both English data and Chinese data. For English, we used the data of CoNLL 2009 Shared Task: Syn- tactic Dependency Parsing (Hajič et al., 2009a). However, when dealing with sentences longer than 60 words, the 1000-best pars- ing process became extremely slow and memory consumption. Therefore, we just discard those sentences in our experiments. For Chinese, we used the Chinese Dependency Treebank (CDT) as the experimental data set (Liu, Ma, & Li, 2006). CDT consists of 60,000 sentences from the People’s Daily in 1990s. The third and fourth columns of Table 1 show the numbers of sentences and dependency arcs (not including the punctuations) in training, development, and test set respectively. 4.2. Dependency parser and classifier An open source NLP toolkit, mate-tools,3 was selected as the dependency parser, which implemented a fast and the state-of- the-art graph-based dependency model (Bohnet, 2010). We modified the source code of mate-tools to output k-best results of each sen- tence. The probability of each dependency arc was calculated with the method introduced in Section 3.1. The last column of Table 1 shows the performance of mate-tools on the CoNLL and CDT data set. Note that the training data are constructed with a fourfold vali- dation as described in Section 3.3 and the LAS of the training data is the average of these folds. Finally, the parser for the development and test data was trained on the whole training data. According to the LAS value as we can see, for English data, the number of positive examples rates about 8 to 9 times as the number of negative exam- ples, while for Chinese data that is 4 to 5 times, which implies an imbalance. The maltparser (Nivre et al., 2007) with default parameter set- tings was used as the transition-based reference parser to obtain the Agreement feature.4 The liblinear (Fan, Chang, Hsieh, Wang, & Lin, 2008) was used as 3 http://code.google.com/p/mate-tools/. 4 http://maltparser.org/. 5 http://www.csie.ntu.edu.tw/�cjlin/liblinear/. the RADAR classifier,5 which implements the L2-regularized logistic regression algorithm with parameters C+ and C�. 4.3. Evaluation method Accuracy is perhaps the most commonly used evaluation meth- od for a classification problem. However, for imbalanced data, accuracy can yield misleading conclusions. For example, for a bin- ary classification problem with a 90:10 class distribution, the Accu- racy can easily reach 90% when we simply classify all the samples to the majority class. In practical applications, it is strongly desired that the most of used arcs are reliable (high precision). However, if http://code.google.com/p/mate-tools/ http://maltparser.org/ http://www.csie.ntu.edu.tw/~cjlin/liblinear/ http://www.csie.ntu.edu.tw/~cjlin/liblinear/ Fig. 3. P@N curves on test data set. Fig. 4. PR curves on the test data. 1720 W. Che et al. / Expert Systems with Applications 41 (2014) 1716–1722 studies only care about the precision of reliable arcs, the number of recognized reliable arcs can be restricted. This situation leads to a low recall of reliable arcs, which are useless in practical applica- tions. The F measure is another normal evaluation method that considers precision and recall at the same time. However, it is not a proper evaluation here either. The same with accuracy as in the previous example, the minority class has much less impact on the F score than the majority class. In practice, the top N most reliable dependency arcs are impor- tant. It can be evaluated with another information retrieval evalu- ation method, P@N. The N is different for different applications. For instance, for a search engine, the N is set to 10 or 20 usually, be- cause users usually only check those top results. However, in the RADAR problem, the N should be larger, while it is difficult to fix the N for different applications. In case of imbalanced data, ROC curve is usually thought as a suitable evaluation metric. However, it was suggested that the PR curves give more informative pictures of the performance of a learning algorithm (Davis & Goadrich, 2006). Therefore, in the cur- rent experiments, the PR curve is used as another evaluation meth- od. In addition, we also calculate the Area Under both the ROC Curve and PR Curve in order to give a quantitative comparison. The AUCCalculator-0.2 tool was used to calculate the AUC-PR and AUC-ROC.6 Larger AUC values indicate better performance. 6 http://mark.goadrich.com/programs/AUC/. 4.4. Contribution of features To evaluate the contribution of different sorts of features men- tioned in Section 3.2.2, we build a classifier with all of these three sorts of features at first. Then, we threw away each sort of fea- tures respectively to see the performance decreasing. Once a sort of thrown features drop down the performance more than an- other, it indicates that this sort of features is more important than another. Contributions are evaluated on the development set of English data and Chinese data respectively. Table 2 shows the re- sults. Note that in order to achieve optimal AUC-PR, the parame- ters of logistic regression classifier are tuned on the development data. The final settings are C+ = 1 and C� = 1.4. Since C� > C+, we can prevent negative examples from being easily recognized into positive class. From Table 2, we can see that all of the three sorts of features are useful for RADAR, because each removed sort of features de- crease the performance. Among these three sorts, the parser-based features contribute the most and text-based features are the least important. The reason is probably that text-based features do not look at the full parse tree, they just evaluate the hardness of cor- rectly parsing a sentence or a modifier. However, this is far from enough to decide the reliability of an arc. At the same time, some of the text-based features, such as the sentence length and the un- known word number, have the same value for all words in a sen- tence. As a consequence, they would be less indicative. In addition, The comparison features are also helpful. http://mark.goadrich.com/programs/AUC/ W. Che et al. / Expert Systems with Applications 41 (2014) 1716–1722 1721 4.5. Results Table 3 shows the AUC-PR and AUC-ROC measurements on the test data. We can see that both for English and Chinese data, the classification method outperforms the probabilistic baseline sys- tem. In order to understand the improvement better and more intuitively, we provide the P@N curve with different N in Fig. 3 and the Precision-Recall (PR) curve in Fig. 4. We can see in the figures that, for different Recall/N, the classification method is all better than the probabilistic baseline, especially when N is small or Recall is low. Based on our anal- ysis, there are two reasons for this. First, the k-best approaching for estimating the probability of an arc is an approximate one. It depends very much on the selection of k. Although a 1000- best estimation is good enough, it is not ideal. The second fac- tor is that the classification method can make use of more extra information (features) such as parsers’ agreement, as well as more powerful tools (classifiers), which help to defeat the prob- abilistic method. We would also note that the k-best decoding in graph-based parsing comes at a high cost in terms of both memory and time consumption, especially when a large k (e.g. k = 500,1000. . .) is set. Thus the classification approach is much more efficient and flexible than the probabilistic approach. 5. Conclusion This present paper proposes a novel natural language process- ing task, RADAR (ReliAble Dependency Arc Recognition). The per- formance of practical applications, such as information extraction and question answering, has potential to be improved further if the reliable arcs can be recognized correctly rather than only rec- ognizing reliable parse trees of whole sentences. We model RADAR as a binary classification problem with imbal- anced data. Then we can elastically use various features and clas- sifiers to achieve better performance. In this paper, we design three sorts of features to express reliability of arcs. A logistic regression model with adjustable C parameter was used as the classifier, which can classify large scale data with high speed and performance. Different from normal binary classification problem, RADAR cannot be evaluated with accuracy or F measure because of the problem of data imbalance. Instead, the P@N curve and PR curve, together with the use of associated area under the curve (AUC- PR, AUC-ROC) evaluate RADAR systems more suitably. The experimental results show that the classification method with appropriate features can outperform the arc probabilistic baseline method. In addition, we also evaluated the contributions of different sorts features. 6. Future work In the future, the work can be extended in the following ways: 1. The improvement of the performance of RADAR will be in two aspects, classifier and features. Besides logistic regression, more classifiers can be compared, which should not only have a high speed and accuracy, but also have output probability. RADAR also can be regarded as a ranking problem and can use various learning to rank algorithms. Besides classification methods, more combinatorial features and global features can be com- bined, such as the reliabilities surrounding the other and cur- rent arcs. Hence, it is no longer a simple binary classification model. In addition, another kind of global features counts from large scale unlabeled data, such as the number of arcs, which are parsed automatically or the Google Web 1T n-gram count information can be used. 2. Apply RADAR to applications, such as entity relation extraction based on a web corpus. These applications may recognize larger reliable sub-structures, such as parse paths. However, corre- sponding reliable arcs cannot be accumulated simply. Thus, rec- ognizing reliable sub-structures becomes an interesting problem. 3. It would be necessary to apply our approach to transition-based parser, to check whether it produces similar results. 4. The last future work is domain adaption problem, i.e. how to adapt the reliability model trained on one domain into another domain or even open domains. This is also an open question for natural language processing. Acknowledgment We gratefully acknowledge the support of the National Natural Science Foundation of China (NSFC) via Grant 61133012 and 61370164. The National Basic Research Program (973 Program) of China via Grant 2014CB340503. References Androutsopoulos, I., & Malakasiotis, P. (2010). A survey of paraphrasing and textual entailment methods. Journal of Artificial Intelligence Research, 38(1), 135–187. Atserias, J., Attardi, G., Simi, M., & Zaragoza, H. (2010). Active learning for building a corpus of questions for parsing. LREC. European Language Resources Association. Avihai Mejer, K. C. (2012). Are you sure? Confidence in prediction of dependency tree edges. In NAACL-HLT2012. Banko, M., Cafarella, M. J., Soderland, S., Broadhead, M., & Etzioni, O. (2007). Open information extraction from the web. In Proceedings of the 20th International Joint Conference on Artifical Intelligence. IJCAI’07 (pp. 2670–2676). San Francisco, CA, USA: Morgan Kaufmann Publishers Inc.. Bishop, C. M. (2006). Pattern recognition and machine learning (Information science and statistics). Secaucus, NJ, USA: Springer-Verlag New York, Inc.. Bohnet, B. (2010). Top accuracy and fast dependency parsing is not a contradiction. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010). Coling 2010 Organizing Committee, Beijing, China, August 2010. (pp. 89–97). Chen, W., Kawahara, D., Uchimoto, K., & Zhang, Y. (2008). Dependency parsing with short dependency relations in unlabeled data. In IJCNLP08. Chen, W., Kazama, J., Uchimoto, K., & Torisawa, K. (2009). Improving dependency parsing with subtrees from auto-parsed data. Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing – EMNLP ’09 (vol. 2). Morristown, NJ, USA: Association for Computational Linguistics, p. 570. Davis, J., & Goadrich, M. (2006). The relationship between precision-recall and roc curves. In Proceedings of the 23rd International Conference on Machine Learning. ICML ’06 (pp. 233–240). New York, NY, USA: ACM. Dell’Orletta, F., & Venturi, G. (2011). ULISSE: An unsupervised algorithm for detecting reliable dependency parses. In CoNLL11 (pp. 115–124). Fan, R.-E., Chang, K.-W., Hsieh, C.-J., Wang, X.-R., & Lin, C.-J. (2008). Liblinear: A library for large linear classification. Journal of Machine Learning Research, 9, 1871–1874. Gildea, D., & Jurafsky, D. (2002). Automatic labeling of semantic roles. Computational Linguistics, 28, 245–288. Goldwasser, D., Reichart, R., Clarke, J., & Roth, D. (2011). Confidence driven unsupervised semantic parsing. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (pp. 1486–1495). Portland, Oregon, USA: Association for Computational Linguistics. Hajič, J., Ciaramita, M., Johansson, R., Kawahara, D., Martí, M. A., Màrquez, L., et al. (2009a). The CoNLL-2009 shared task: Syntactic and semantic dependencies in multiple languages. In Proceedings of the 13th Conference on Computational Natural Language Learning (CoNLL 2009): Shared Task (pp. 1–18). Boulder, Colorado: Association for Computational Linguistics. Hajič, J., Ciaramita, M., Johansson, R., Kawahara, D., Martí, M.A., Màrquez, L., Meyers, A., Nivre, J., Padó, S., Štěpánek, J., Straňák, P., Surdeanu, M., Xue, N., & Zhang, Y. (2009). The CoNLL-2009 shared task: Syntactic and semantic dependencies in multiple languages. In CoNLL-2009. Johansson, R., & Nugues, P. (2007). Extended constituent-to-dependency conversion for English. In Proceedings of NODALIDA 2007, Tartu, Estonia. Kawahara, D., & Kurohashi, S. (2010). Acquiring reliable predicate-argument structures from raw corpora for case frame compilation. In LREC10 (pp. 1389– 1393). Kawahara, D., & Uchimoto, K. (2008). Learning reliability of parses for domain adaptation of dependency parsing. In IJCNLP08 (pp. 709–714). http://refhub.elsevier.com/S0957-4174(13)00691-X/h0005 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0005 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0010 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0010 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0015 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0015 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0015 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0015 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0020 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0020 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0025 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0025 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0025 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0025 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0030 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0030 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0030 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0035 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0035 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0035 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0040 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0040 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0045 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0045 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0045 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0045 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0045 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0050 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0050 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0050 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0050 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0050 1722 W. Che et al. / Expert Systems with Applications 41 (2014) 1716–1722 Kim, J.-D., Ohta, T., Pyysalo, S., Kano, Y., & Tsujii, J. (2009). Overview of BioNLP’09 shared task on event extraction. In Proceedings of the Workshop on Current Trends in Biomedical Natural Language Processing: Shared Task. BioNLP ’09 (pp. 1–9). Stroudsburg, PA, USA: Association for Computational Linguistics. Koo, T., Globerson, A., Carreras, X., & Collins, M. (2007). Structured prediction models via the matrix-tree theorem. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL) (pp. 141–150). Prague, Czech Republic: Association for Computational Linguistics. Kübler, S., McDonald, R. T., & Nivre, J. (2009). Dependency parsing. Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers. Liu, T., Ma, J., & Li, S. (2006). Building a dependency treebank for improving chinese parser. Journal of Chinese Language and Computing, 16, 207–224. Meena, A., & Prabhakar, T. V. (2007). Sentence level sentiment analysis in the presence of conjuncts using linguistic analysis. In Proceedings of the 29th European Conference on IR Research. ECIR’07 (pp. 573–580). Berlin, Heidelberg: Springer-Verlag. Nivre, J. (2006). Inductive dependency parsing (Text, Speech and Language Technology). Secaucus, NJ, USA: Springer-Verlag New York, Inc.. Nivre, J., Hall, J., Nilsson, J., Chanev, A., Eryigit, G., Kübler, S., et al. (2007). Maltparser: A language-independent system for data-driven dependency parsing. Natural Language Engineering, 13(2), 95–135. Ravi, S., Knight, K., & Soricut, R. (2008). Automatic prediction of parser accuracy. In EMNLP08 (pp. 887–896). Morristown, NJ, USA: Association for Computational Linguistics. Reichart, R., & Rappoport, A. (2007). An ensemble method for selection of high quality parses. In ACL07 (pp. 408–415). Sag, I. A., Baldwin, T., Bond, F., Copestake, A. A., & Flickinger, D. (2002). Multiword expressions: A pain in the neck for nlp. In Proceedings of the 3rd International Conference on Computational Linguistics and Intelligent Text Processing. CICLing ’02 (pp. 1–15). London, UK, UK: Springer-Verlag. Settles, B. (2010). Active learning literature survey. Tech. rep., University of Wisconsin-Madison. van Noord, G. (2007). Using self-trained bilexical preferences to improve disambiguation accuracy. In Proceedings of the 10th International Conference on Parsing Technologies – IWPT ’07 (pp. 1–10). Morristown, NJ, USA: Association for Computational Linguistics. Wu, F., & Weld, D. S. (2010). Open information extraction using wikipedia. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics. ACL ’10 (pp. 118–127). Stroudsburg, PA, USA: Association for Computational Linguistics. Yates, A., Schoenmackers, S., & Etzioni, O. (2006). Detecting parser errors using web- based semantic filters. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing – EMNLP ’06 (pp. 27). Morristown, NJ, USA: Association for Computational Linguistics. Zhang, M., Zhang, J., Su, J., & Zhou, G. (2006). A composite kernel to extract relations between entities with both flat and structured features. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th Annual Meeting of the Association for Computational Linguistics. ACL-44 (pp. 825–832). Stroudsburg, PA, USA: Association for Computational Linguistics. http://refhub.elsevier.com/S0957-4174(13)00691-X/h0055 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0055 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0055 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0055 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0060 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0060 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0060 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0060 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0060 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0065 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0065 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0070 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0070 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0075 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0075 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0075 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0075 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0080 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0080 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0085 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0085 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0085 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0090 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0090 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0090 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0095 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0095 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0095 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0095 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0100 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0100 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0100 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0100 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0105 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0105 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0105 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0105 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0110 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0110 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0110 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0110 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0115 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0115 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0115 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0115 http://refhub.elsevier.com/S0957-4174(13)00691-X/h0115 ReliAble dependency arc recognition 1 Introduction 2 Related work 3 Method description 3.1 Graph-based dependency parsing 3.2 RADAR as binary classification 3.2.1 Classifier 3.2.2 Features 3.3 Training process 4 Experiments 4.1 Data set 4.2 Dependency parser and classifier 4.3 Evaluation method 4.4 Contribution of features 4.5 Results 5 Conclusion 6 Future work Acknowledgment References 
work_cm2bnjuqhfdkbjahytaognrt2a	----	2 1 Improved orthogonal array based simulated annealing for design optimization Chan, K.Y. Kwong, C.K.* and Tsim, Y.C. Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon Hong Kong, PRC Abstract Recent research shows that simulated annealing with orthogonal array based neighbourhood functions can help in the search for a solution to a parametrical problem which is closer to an optimum when compared with conventional simulated annealing. Previous studies of simulated annealing analyzed only the main effects of variables of parametrical problems. In fact, both main effects of variables and interactions between variables should be considered, since interactions between variables exist in many parametrical problems. In this paper, an improved orthogonal array based neighbourhood function (IONF) for simulated annealing with the consideration of interaction effects between variables is described. After solving a set of parametrical benchmark function problems where interaction effects between variables exist, results of the benchmark tests show that the proposed simulated annealing algorithm with the IONF outperforms significantly both the simulated annealing algorithms with the existing orthogonal array based neighbourhood functions and the standard neighbourhood functions. Finally, the 2 improved orthogonal array based simulated annealing was applied on the optimization of emulsified dynamite packing-machine design by which the applicability of the algorithm in real world problems can be evaluated and its effectiveness can be further validated. Keywords: Orthogonal array, interaction analysis, neighbourhood functions, simulated annealing, design optimization Corresponding author: Dr. C.K. Kwong Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, PRC Tel: (852) 27666610 Fax: (852) 23625267 Email: mfckkong@polyu.edu.hk 3 1. Introduction Simulated annealing (SA) is a point-based stochastic optimization method, which explores iterationally from an initial solution to the optimum [4, 14]. Each iteration employs a neighbourhood function to generate a candidate solution by a randomized perturbation on a current solution. Therefore, design of neighbourhood functions plays an important role in developing an effective simulated annealing. The searching mechanism of SA has a very good convergent property [19] and SA has been widely applied in solving many hard optimization problems [32, 33]. However, it can be noted in many previous researches [1, 17, 26] that SA can find good or reasonable solutions, but in many cases it cannot search for a global optimum. The searching ability of SA improves in the earlier stages of the searching process, but it saturates or even terminates in later stages. Therefore, it is difficult to obtain any substantial improvements by examining neighbouring solutions in the later stage of the search. Vaessens et al. [34] put this searching method into the context of a local, or neighbourhood search. Also it has been noted in some previous researches that long computational time is commonly required for SA to search for an acceptable solution for solving hard optimization problems [29, 38, 39]. Various approaches have been proposed to improve the searching mechanism by modifying neighbourhood functions [8, 38, 39, 30], modifying criterion of accepting a new candidate solution [28], incorporating with other optimization methods [7, 21, 36] and parallelized computing [1]. A recent approach to improving the searching mechanism of SA has been proposed by introducing orthogonal arrays into neighbourhood functions of SA [9, 10, 12, 27]. The Orthogonal arrays exploit the neighbourhood of a current solution by 4 analyzing the main effect of variables in the current solution. The neighbourhood function, which uses the orthogonal array for exploiting solution spaces, is called orthogonal array based neighbourhood function (ONF) in this paper. It has been shown that ONF can speed up the search of SA and determine a more accurate candidate solution for electromagnetic problems [27], floorplanning problems [10] and controller design problems [9, 12] compared with other neighbourhood functions. However, we found that the exploitation of candidate solutions can be further improved by considering not only the main effects of variables but also the interaction effects between variables. It is due to the fact that strong interaction effects could exist between variables in many optimization problems. If strong interaction effects exist in localized features of a search space, poor results may be obtained by considering main effects only in variables [6, 23, 24, 25]. In this paper, an improved orthogonal array based neighbourhood function (IONF) for SA which considers both main effects in variables and interaction effects between variables is proposed. Application of the proposed simulated annealing algorithm with the IONF on the optimization of emulsified dynamite packing-machine design is also described. IONF employs the approach of interaction plots [22] to analyze interaction effects between variables. Interaction plots have been commonly used to analyze interaction effects between parameters in industrial systems [31, 13, 18, 20, 40]. The background of orthogonal arrays and ONF, as well as the proposed IONF are described in Section 2 and 3 respectively. Benchmark results of solving a set of hard benchmark problems [37, 11] using the three simulated annealing algorithms with employing ONF, IONF and the standard neighbourhood function respectively are shown in Section 4. Application of the improved simulated annealing on the optimization of 5 emulsified dynamite packing-machine design and further validation of the effectiveness of the algorithm are described in Section 5. 2. Orthogonal Experimental Design and Neighbourhood Function The use of orthogonal arrays in planning experiments and analyzing experimental data is briefly described in Section 2.1. In Section 2.2, background and limitations of the orthogonal array based neighbourhood functions (ONF), that has been applied in solving many hard optimization problems [9, 10, 12 and 27], are presented. 2.1 Orthogonal Experimental Design One major objective of an experimental design is to find the best combination of parameter levels for optimal performance of a system or a model. If an experimental design is based on the full factorial one and the number of parameters to be investigated is large, a large number of experimental runs always are required to be carried out. To reduce the number of experimental runs, a fractional factorial design is an alternative in which the experimental design can be based on orthogonal arrays [2]. An experimental plan based on an orthogonal array L2N+1(pN) involves a maximum N parameters and p levels in each parameter for 2N+1 experimental runs. If an orthogonal array L2N+1(3N) with 3 levels is considered, for j=1,2,…N and k = 1,2,3, main effect Mjk of the parameter j for level k is defined as: ∑ + = ×= 12 1 N t tFtyjk M (1) where yt denotes an objective function value of the combination corresponding to experiment t, and 6    = .not is experiment of parameter of level if 0 . is experiment of parameter of level if 1 ktj ktj t F For smaller-the-better type problems, the best level Best(j) of the j-th parameter is denoted as: ( )( )321 ,,minarg)( jjj MMMjBest = (2) where 'arg(min(..))' is a function that returns the index of the minimum value. For example, if the value of Mj2 is the smallest among the values of Mjk where k=1,2 and 3, then Best(j)=2. For larger-the-better type problems, the best level Best(j) of the j-th parameter is denoted as: ( )( )321 ,,maxarg)( jjj MMMjBest = (3) where ' arg(max(..))' is a function that returns the index of the maximum value. 2.2 Orthogonal Array based Neighbourhood Function (ONF) Ho et al [9, 10, 12] and Shu et al [27] proposed an orthogonal array based neighbourhood function (ONF) which aims at generating a candidate solution by using the combinations of the orthogonal array ( )NNL 312 + . ONF generate a candidate solution Q=ONF(P1) from P1, where ( )mSSP ,...,11 = is the current solution. First two temporary solutions, P2 and P3 are generated by perturbing P1 as follows: ( )1112 ,..., mSSP = and ( )2213 ,..., mSSP = , (4) where iiiiii SSSSSS −=+= 21 and , i=1,...,m. All iS are generated based on the Cauchy- Lorentz probability distribution [29]. Consequently, Q is produced by a combination of variables of P1, P2, and P3. 7 In ONF, P1, P2 and P3 are considered as level 2, level 1 and level 3 of an experimental design. To assign the m variables from P1, P2 and P3 into the N parameters in ( )NNL 312 + , the m variables are divided into N non-overlapping groups. Each non- overlapping group of variables is considered as a parameter in ( )NNL 312 + . The number of variables in each group is determined by a generated random integer li, i=1,2,…N, where ml N i i =∑ =1 (5) ONF then uses the t-th combination of ( )NNL 312 + to compute yt corresponding to the t-th experiment with t=1,...2N+1, and computes all main effects jkM with j=1,2,..,N and k=1,2,3 based on (1). Finally, it determines the best level of each parameter based on the computed main effects by (2) for smaller-the-better type problems or (3) for larger-the- better type problems. The algorithm of the ONF, Q=ONF(P1), is shown in the Appendix 1. ONF uses the analysis of main effects to determine the optimal levels of parameters which is the simplest approach to analyze experimental results [2, 22]. However, it is quite common that an interaction effect exists between two parameters in a function [6]. Further studies of interaction effects and main effects in a function have been done by [23, 24, 25] based on ‘analysis of variance (ANOVA)’. The parametrical effects of a function are analyzed using a total sum of squared deviations SS, which can be divided into SS of main effects and interaction effects as shown below: Total SS = SS of main effects + SS of interactions With the higher SS of interactions, the lack of provision of adequacy dealing with the potential interactions between parameters is a major weakness of ONF. To solve the 8 optimization problems where low interaction effects exist between parameters, ONF could work properly. However, if strong interaction effects exist between parameters in optimization problems, the optimal combination based on ONF may not be reproducible. 3 An Improved Orthogonal Array Based Neighbourhood Function In this paper, an improved orthogonal array based neighbourhood function (IONF) is proposed with the consideration of interaction effects between parameters. In the ONF, the children are produced by considering only the largest main effects of parameters. But, in IONF, the children are produced by considering both main effects of parameters and interaction effects between parameters. Interaction plots [22] are adopted to investigate the magnitudes of interaction effects between parameters. In the IONF, an interaction matrix MIij is generated to estimate the magnitudes of interaction effects between parameters i and j, where i, j=1,2,..N. It can be expressed as: ( )( ) 33 3nm,1for ;, × ≤≤= nmIMI ijij (6) where the numbers of rows and columns of the interaction matrix MIij are both equal to the number of levels of L2N+1(3N) which is 3. The elements of MIij, ( )nmIij , , which represents the average fitness of the thi parameter with level m and thj parameter with level n, are defined as: ( ) ( ) ( ) ∑ =           + ∑ =           +⋅ = N p nthj mthiNNL thp N p nthj mthiNNL thp pf nmIij 1 isparameter theand isparameter theof level the,312 ofn combinatio In the 1 isparameter theand isparameter theof level the,312 ofn combinatio In the , (7) 9 where fp denotes an objective function value of the combination corresponding to the p-th row of L2N+1(3N), 3,1 ≤≤ nm and [ ]    = otherwise. 0 true.isbracket theinsidestatement theif 1 condition Then interaction plots are used to investigate the magnitude of interaction effects between parameters i and j. The thr line of the interaction plot is defined as: ( ) ( ) ( )( ) 3. 2, 1, where,3,,2,,1 == rrIrIrI(r)Line ijijijij (8) Figure 1 shows that the lines cross indicating the existence of strong interactions. If strong interaction effects do not exist in all parameter pairs, only the main effects of parameters need to be studied. The candidate solution Q of the IONF is generated by the combination of the parameters with the largest main effects based on (2) for smaller-the- better type problems or (3) for larger-the-better type problems. In this case, the algorithm of IONF is identical to the one of ONF. However, if strong interaction effects exist in any one of the parameter pairs, the candidate solution Q is first generated by the best level combinations of the orthogonal array L2N+1(3N) with the optimal yt. For those parameters without strong interaction effects between each other, the level combinations in Q are replaced by the parameters with the largest main effects based on (2) or (3). The algorithm of the IONF, Q=IONF(P1), is given as follows: Algorithm Q=IONF(P1) Step 1) Generate P2 and P3 with ( )mSSP ,...,11 = based on (4). Step 2) Divide P1, P2 and P3 into N groups based on (5). Step 3) Represent levels 2, 1 and 3 of the j-th parameter of ( )NNL 312 + by the j-th group of P1, P2 and P3 respectively. 10 Step 4: Compute yt based on the t-th combination of ( )NNL 312 + of the t-th experiment. Step 5: Compute the main effects Mjk where j=1,...,N and k=1,2 and 3 based on (1). Step 6: Construct the interaction matrices ijMI by (6), where i, j=1,2,…N with ji ≠ . Step 7: Construct the interaction plot for ijMI where i, j=1,2,…N with ji ≠ . Step 8: Check whether the parameters i and j have a strong interaction effect of each other, where i, j=1,2,…N with ji ≠ . Step 9: If strong interaction effect exists in any one of the parameter pairs, goto Step 10, otherwise goto Step 13. Step 10: Form the candidate solution Q by the combination of the ( )NNL 312 + with the optimal yt. Step 11: For the parameter pair with no strong interaction effect, the level combinations in Q are replaced by the level combinations with the largest main effects based on (2) or (3). Step 12: Output Q as the resulting solution of IONF. Then goto step 15. Step 13: Determine the best level Best(j) on the j-th parameter based on (2) for smaller-the-better type problems and (3) for larger-the-better type problems. Step 14: The candidate solution Q is produced by the combinations of best levels of parameters. Step 15: Terminate the algorithm. 11 4 Benchmark results based on non-separable functions To evaluate the effectiveness of the proposed IONF, benchmark tests based on the three simulated annealing algorithms, which employ the standard neighbourhood function SNF [32], ONF [9, 10, 12, 27], and the proposed IONF respectively, were conducted to solve a set of selected parametrical benchmark functions. Those benchmark functions ( 61 ff − ) is shown in Table 1, in which interaction effects exist between variables. 41 ff − were collected from [12], while 65 ff − were collected from [37]. They cannot be decomposed into linear combinations of independent sub-functions since variables interact with each other and cannot be enumerated completely. They can be classified as good test suites for the algorithms since they are non-separable in which each sub-function contains at least two variables [35]. To evaluate the performance of the three neighbourhood functions (SNF, ONF and IONF), the simulated annealing algorithm used by [12, 27] was employed in this research which is shown below. 12 Algorithm of simulated annealing Begin i=1 Randomly generate a candidate solution s Repeat t=T0; I=I0 Repeat:- Q = neighbour_function(s) if f(Q)< f(s) then replace s with Q else generate a random number r if exp(-(f(s)-f(Q))/t)>r then replace s with Q endif endif t=t*CR until t<I*CR i=i+1 Until i=( pre-defined number of function evaluations) End. In the algorithm, a candidate solution s is randomly initialized first, and then it starts with the highest temperature (t=T0). The algorithm then modifies the candidate solution Q by using a neighbourhood function and judges if the yielded solution will be promoted for the next inner iteration. Then the algorithm reduces the temperature t to t=t*CR by a cooling coefficient (CR) at the end of each inner iteration. When the temperature reduces to I*CR, the solution Q results from the inner iteration. The simulated annealing algorithm stops iterating until i reaches the pre-defined number of functional evaluations. The three simulated annealing algorithms with SNF, ONF and IONF respectively are called standard simulated annealing (SSA), orthogonal simulated annealing (OSA) and improved orthogonal simulated annealing (IOSA). They have been implemented to solve the benchmark functions 1f to 6f . By referring to the previous research [9], the parameters used in OSA and IOSA are T0=50, I0=5 and CR=0.95. The orthogonal array 13 used in OSA and IOSA is also same as the one used in [9], which is L2N+1(3N), where ( ) ( ) 2/13 12log3 −= +pN and p is the number of variables of the optimization problem. Since only one functional evaluation is implemented in SNF, a functional evaluation is required in each iteration of SSA. However, 2N+1 functional evaluations are implemented in the ONF and IONF. Therefore, 2N+1 functional evaluations are required in each iteration of the OSA and IOSA. To make the number of functional evaluations of all the algorithms to be the same, the number of iterations used in SSA is larger than those used in both the OSA and IOSA. Therefore, value of CR used in SSA is larger than the one used in the OSA and IOSA. The parameters of SSA used are T0=50, I0=5 and CR=0.99. The pre-defined number of functional evaluations used in all the algorithms is 10 000. Tables 2 to 7 as shown in Appendix 2 are the statistical results based on the algorithms for solving the functions f1 to f6 respectively, where the dimensions, p= 20, p= 40, p= 60, p= 80 and p= 100, are used. The statistical results of means and standard deviations of the solutions over the 30 runs are shown as well. It can be seen from the tables that the mean values based on the IOSA are smaller than those based on the OSA or SSA. Also the standard deviations of IOSA are found to be the smallest compared with those of the other two simulated annealing algorithms in all benchmark functions. Therefore, in terms of quality and robustness of the solutions obtained, IOSA outperforms the other two simulated annealing algorithms in all the six benchmark functions. Based on Tables 2 to 7, Figures 2 to 7 can be generated, which show the means of the solutions based on the three simulated annealing algorithms respectively for the six benchmark functions. It can be found from the figures that the IOSA outperforms the 14 OSA and the SSA in all the six benchmark functions with various number of dimensions. This indicates that the IOSA can perform well in solving the problems where interaction effects exist between variables with small and large numbers of dimensions. 5 Application of IOSA on design optimization and further validation In weapon manufacturing, handling of dynamite is in powder state which is easy to explode during transportation or storage. Nowadays, dynamite is first emulsified into liquid state which is more safe in handling. A machine normally is required to perform the emulsification of dynamite which is commonly named as an emulsified dynamite packing machine. In this section, optimization of emulsified dynamite packing machine design based on the IOSA is illustrated. An optimization problem of the machine design formulated by [16] was adopted in this research which contains the following engineering requirements (i.e. X=X1, X2,…, X7) and customer requirements (i.e. Y=Y1, Y2,…, Y4): X1 – precision of molding of clip X2 – precision of dynamite packing X3 – control force of dynamite packing X4 – efficiency of dynamite packing X5 – hardness of pressing hammer X6 – noise of cam power transmission X7 – height of machine bed Y1 –quality of packing dynamite Y2 – efficiency of packing dynamite Y3 – packing noise Y4 – rigidity of machine Details of the optimization problem [16] are shown below. ( )rrrrr yyyyosc 1 4321 10.016.028.046.0 +++= (9) 15 subject to: 88.037.133.098.5 3211 +−−= xxxy (10) 54.025.196.045.2 5432 +++= xxxy (11) 00.120.4 63 += xy (12) 25.100.4 74 += xy (13) ( ) ( ) 222 2 1 5.013.1 ≤−+− ff (14) 43211 166.0217.0449.0464.0 yyyyf +−+= (15) 43212 695.0508.0100.0030.0 yyyyf −−−−= (16) 10083051510252050 7654321 ≤+++++++ xxxxxxx (17) 4,...,1 ,51 =≤≤ iyi (18) 7,...,1 ,10 =≤≤ jx j (19) where - xi (i=1,2,…7) is the level of attainment of Xi; yi (i=1,2,…4) is the value of customer satisfaction of Yi; - (9) is the objective function of deriving overall customer satisfaction (OCS); - (10) to (13) are the models of functional relationship between customer requirement Yi, i=1,…,4 and engineering requirements, Xj, j=1,…,7; - (14) to (16) are the constraints of product positioning; - (17) is the cost constraint that is subject to a budget with the fixed cost and the cost incurred for achieving each engineering requirement, Xj, j=1,…,7; - (18) and (19) are the ranges of values of the customer satisfactions and levels of attainment of the engineering requirements respectively. 16 The IOSA was used to determine the optimal setting of levels of attainment of the engineering requirements, x1, x2, x3, x4, x5, x6 and x7, by maximizing the overall customer satisfaction. The algorithm was implemented by using Matlab, in which a candidate solution is represented as: [ ]7654321 ,,,,,, xxxxxxxs = (20) The objective function used in the algorithm is defined by maximizing the optimization function (9) subject to the constraints. It is defined as: ( ) ( ) ( )( ) ( ) ( ) ( )            <<−− = otherwise 0or 0 if ,min max sf sfsfsfsf sf a cbcb (21) where ( ) ( )rrrrra yyyysf 1 4321 10.016.028.046.0 +++= ; (22) ( ) ( ) ( )[ ]22212 13.15.0 −+−−= ffsfb ; (23) ( ) ( )7654321 83051510252050100 xxxxxxxsfc −−−−−−−−= ; (24) ( ) ( ) 51 if 0 5 and 1 if 1 R ;R3 4 1 2    ≤≤ << =⋅−= ∑ = i ii ii i id y yy ysf (25) y1, y2, y3 and y4 can be found in (10), (11), (12) and (13) respectively; f1 and f2 can be found in (15) and (16) respectively. (23) is formulated to handle the constraints (14)-(16). (24) and (25) are formulated to handle the constraints (17) and (18) respectively. The constraint (19) can be dealt with by setting the ranges of solutions in the simulated annealing. The parameter setting used in the IOSA is the same as the one used in Section 4 except the number of functional evaluations. The variable r used in (25) was set to 1, 2, 3, 17 4 and 5. With a higher value of r, interaction effects between variables are higher. Since the simulated annealing algorithm is stochastic one, different solutions are obtained from different runs. Therefore, 30 test runs were performed in order to obtain the two statistics, means and variances of overall customer satisfaction with respect to the values of r. After assessing the interaction effects between the requirements as shown in the house of quality for the machine design [16], they were assessed as ‘low to medium’ and hence r value was set as 2. After running the IOSA for 30 times, the optimal solutions s corresponding to r=2 was found to be (0.8735, 0.0002, 0.8055, 0.5479, 0.9913, 0.3211, 0.2076). To further evaluate the effectiveness of the IOSA in solving real world optimization problems compared with other simulated annealing algorithms, SSA and OSA were also used to solve the same optimization problem using the same parameter setting and number of test runs. Table 8 shows the means and standard deviations of the runs with different values of r. It can be found from the table that the means of the overall customer satisfaction based on the IOSA for any value of r are larger than the ones based on the OSA and SSA, except r=1. It is because there is no interaction between variables of (9) while r=1. Therefore the overall mean of customer satisfaction based on the IOSA is nearly identical to that based on the OSA. Regarding the standard deviation, it can be found from Table 8 that the average standard deviations of the IOSA is smaller than the ones based on the OSA and SSA. The t-test was conducted to evaluate how significantly the IOSA is better than the other two simulated annealing algorithms in this optimization problem. The t-values between the IOSA and the other two simulated annealing algorithms are shown in Figure 18 8, from which it can be seen that all the t-values are higher than 1.675 except the t-value for OSA-IOSA while r=1. Based on the normal distribution table, if the t-value is higher than 1.675, the difference of performance between two algorithms is significant with a confidence level of 95.3%. Therefore it can be concluded that the performance of the IOSA is significantly better than the SSA and OSA. As explained before, while r=1, interaction effects does not exist in (9). There is no significant difference in performance between the OSA and IOSA. The significance of the difference increases as r increases. It can be explained that with a larger value of r, interaction effects between variables become stronger. On the other hand, computational times of generating solutions based on the IOSA were also compared with those based on the OSA and SSA. Figure 9 shows the computational times of the three algorithms with respect to different values of r. Execution of the algorithms is based on a Pentium 4 PC with 2.26 MHz. It can be seen from the figure that the times taken to search for solutions based on the IOSA are less than those based on the OSA and SSA for all values of r, except r=1. When r=1, the time taken based on the IOSA is identical to that based on the OSA, but is still less than that based on the SSA. 6 Conclusion and further work In this paper, an improved orthogonal array based neighbourhood function (IONF) for simulated annealing which considers both main effects of individual variable and interaction effects between variables is described. The proposed IONF is to make up the deficiency of the existing orthogonal array based neighbourhood functions (ONF), which 19 is lack of consideration of interaction effects between variables. Benchmark tests involving 6 parametrical benchmark functions were conducted based on the 3 simulated annealing algorithms, IOSA, OSA and SSA, which were incorporated with the IONF, ONF and SNF respectively. Results of the benchmark tests indicate that IOSA outperforms OSA and SSA in terms of quality and robustness of solutions. The IOSA was successfully applied to solve the design optimization problem of emulsified dynamite packing machines in which interaction effects exist between the requirements. Further validation tests based on the case of the machine design were conducted in order to evaluate the effectiveness of the IOSA in solving real world optimization problems compared with the OSA and SSA. Results of the validation tests also indicate that IOSA outperforms the other two simulated annealing algorithms in terms of quality and robustness of solutions. T-tests were also conducted. Results of the tests indicate that the IOSA outperforms the other two algorithms significantly. Besides, the times taken to search for solutions based on the IOSA are the smallest compared with those based on the other two algorithms. Further work would involve the study of interaction effects among three or more variables in the IOSA. On the other hand, the orthogonal arrays with considering interaction effects could also be investigated in particle swarm optimization. The resulting algorithm could be used to model those systems or processes which are highly dimensional and nonlinear. 20 Acknowledgments The work described in this paper is supported substantially by a grant from the Research Grants Council of Hong Kong, China (Project No. PolyU 5184/07). References 1. Aydin M.E. and Fogarty T.C. (2004). A distributed evolutionary simulated annealing for combinatorial optimisation problems, Journal of Heuristics, 10(3), 269-292. 2. Box G.E.P., Hunter W.G. and Hunter J.S. (1978). Statistics for Experimenters. John Wiley. 3. Bryne D.M. and Taguchi S. (1986). The Taguchi approach to parameter design, ASQC Quality Congress Transaction, Anaheim, 168. 4. Cerny V. (1985). Thermodynamical approach to the travelling salesman problem: an efficient simulation algorithm. Journal of Optimization Theory and Applications, 45, 41– 51. 5. Chatterjee S., Carrera C. and Lynch L.A. (1995). Genetic algorithms and travelling salesman problems. European Journal of Operational Research, 93, 490-510. 6. Davidor Y. (1991). Epistasis variance: a viewpoint on GA-hardness. In G.J.E. Rawlins (Ed.) Foundations of Genetic Algorithms, Morgan Kaufmann, San Mateo. 7. Fogel D.B. (1994). An introduction to simulated evolutionary optimization. IEEE Transactions of Neural Networks, 5(1), 3-14. 8. Gong G., Liu Y. and Qian M. (2001). An adaptive simulated annealing algorithm. Stochastic Processes and their Applications, 94, 95-103. 21 9. Ho S.J., Ho S.Y., and Shu L.S. (2004). OSA: Orthogonal simulated annealing algorithm and its application to designing mixed H2=H∞ Optimal Controllers. IEEE Transactions on Systems. Man and Cybernetics – Part A: Systems and Humans, 34(5), 588-600. 10. Ho S.Y., Ho S.J., Lin Y.K. and Chu W.C.C. (2004). An orthogonal simulated annealing algorithm for large floorplanning problems. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 12(8), 874-876. 11. Ho S.Y., Shu L.S. and Chen J.H. (2004). Intelligent evolutionary algorithms for large parameter optimization problems. IEEE Transactions on Evolutionary Computation, 8(6), 522-541. 12. Ho S.J., Shu L.S. and Ho S.Y. (2006). Optimizing fuzzy neural networks for tuning PID controllers using an orthogonal simulated annealing algorithm OSA. IEEE Transactions on Fuzzy Systems, 14(3), 421-434. 13. Kim J.D. and Choi M.S. (1995). Stochastic approach to experimental analysis of cylindrical lapping process. International Journal of Machines Tools Manufacturing, 35(1), 51-59. 14. Kirkpatrick S., Gelatt JR., and Vecchi M.P. (1983). Optimization by simulated annealing. Science, 220, 671-680. 15. Kratica J., Tosic D., Filipovic V. and Ljubic I. (2001). Solving the simple plant location problem by genetic algorithm. RAIRO Operations Research, 35, 127-142. 16. Kwong C.K., Chen Y. and Bai H. Integrating perceptual product mapping with QFD for new product design. (working paper) 22 17. Lin C.K.Y., Haley K.B. and Sparks C. (1995). A comparative study of both standard and adaptive versions of threshold accepting and simulated annealing algorithms in three scheduling problems. European Journal of Operational Research, 83, 330-346. 18. Lin Y.H., Tyan Y.Y., Chang T.P. and Chang C.Y. (2004). An assessment of optimal mixture for concrete made with recycled concrete aggregates. Cement and Concrete Research, 34, 1373-1380. 19. Locatelli M. (2000). Simulated annealing algorithms for continuous global optimization: convergence conditions. Journal of Optimization Theory and Applications, 104(1), 121-133. 20. Mohan N.S., Ramachandra A. and Kulkarni S.M. (2005). Influence of process parameters on cutting force and torque during drilling of glass fiber polyester reinforced composites. Composite Structures, 71, 407-413. 21. Moilanen A. (2001). Simulated evolutionary optimization and local search: introduction and application to tree search. Cladistics, 17, 12-15. 22. Phadke M.S. (1987). Quality engineering using robust design. New York: Prentic Hall. 23. Reeves C.R. and Wright C.C. (1995). An experimental design perspective on genetic algorithms. Foundation of Genetic Algorithms, 3, 7-22. 24. Reeves C.R. and Wright C.C. (1995). Epistasis in Genetic Algorithms: An Experimental Design Perspective. Proceedings of the 6th International Conference on Genetic Algorithms (pp. 217-224). 25. Reeves C.R. (1999). Predictive measures for problem difficulty. Proceedings of the 1999 Congress on Evolutionary Computation, 1, (pp. 736-742). 23 26. Ruiz-Torres A.J., Enscore E.E. and Barton R.R. (1997). Simulated annealing heuristics for the average flow-time and the number of Tardy jobs bi-criteria identical parallel machine Problem. Computers Industry Engineering, 33(1-2), 257-260. 27. Shu L.S., Ho S.Y. and Ho S.J. (2004). A novel orthogonal simulated annealing algorithm for optimization of electromagnetic problems. IEEE Transactions on Magnetics, 40(4), 1791-1795. 28. Szu H. and Hartley R. (1987). Fast simulated annealing. Physical Letters, 122, 157- 162. 29. Szu H. (1987). Nonconvex optimization by fast simulated annealing. Proceedings of the IEEE, 75(11), 1538-1540. 30. Tsallis C. and Stariolo D.A. (1996). Generalized simulated annealing. Physica A, 233, 395-406. 31. Unal R., Stanley D.O. and Joyner C.R. (1993). Propulsion system design optimization using the Taguchi Method. IEEE Transactions on Engineering Management, 40(3), 315- 322. 32. Van Laarhoven P.J.M., Aarts E.H. and Lenstra J.K. (1992). Job shop scheduling by simulated annealing. Operations Research, 40(1), 113-125. 33. Van Laarhoven P.J.M. and Aarts E.H.L. (1987). Simulated Annealing: Theory and Applications. D. Reidel Publishing Co. 34. Vaessens R.J.M., Aarts E.H.L. and Lenstra J.K. (1992). A local search template. Proceedings of Parallel Problem-Solving from Nature (pp. 65-74). 24 35. Whitley D., Mathias K., Rana S. and Dzubera J. (1995), Building better test function. Proceedings of the 6th International Conference on Genetic Algorithms (pp. 239-246). 36. Wong S.Y.W. (2001). Hybrid simulated annealing/genetic algorithm approach to short term hydro-thermal scheduling with multiple thermal plants. Electric Power Energy Systems, 23, 565-575. 37. Yao X., Lin Y. and Lin G. (1999). Evolutionary programming made faster. IEEE Transactions on Evolutionary Computation, 3(2), 82-102. 38. Yao X. (1991). Simulated annealing with extended neighbourhood. International Journal of Computer Mathematics. 40, 169-189. 39. Yao X. (1993). Comparison of different neighbourhood sizes in simulated annealing. Proceedings of 4th Australian Conference on Neural Networks (pp. 216-219). 40. Yin G.Z. and Jillie D.W. (1989). Orthogonal design for process optimization and its application in plasma etching. Taguchi Methods: Applications in World Industry, A. Bendell, J. Disney, W.A. Pridmore (ed.s), IFS Publications/Springer-Verlag, (181-198). 25 Appendix 1 Algorithm Q=ONF(P1) Step 1) Generate P2 and P3 with ( )mSSP ,...,11 = based on (4). Step 2) Divide P1, P2 and P3 into N groups based on (5). Step 3) Represent levels 2, 1 and 3 of the j-th parameter of ( )NNL 312 + by the j-th group of P1, P2 and P3 respectively. Step 4) Compute yt based on the t-th combination of ( )NNL 312 + as the t-th experiment. Step 5) Compute the main effect Mjk where j=1,...,N and k=1,2,3 based on (1). Step 6) Determine the best level Best(j) on the j-th parameter based on (2) for smaller-the-better problems and (3) for larger-the-better problems. Step 7) The candidate solution Q is produced by the combinations of best levels of parameters. Appendix 2 [Insert Tables 2 to 8 here] 26 Figure 1 Strong interaction effect exists between parameters i and j 10 20 30 40 50 60 70 80 90 100 110 10 -4 10 -3 10 -2 10 -1 10 0 function f1 dimension p m ea n of s ol ut io ns SSA OSA IOSA Figure 2 Results of function f1 Lineij(1) Lineij(3) Iij(3,3) Iij(3,2) Iij(3,1) Iij(2,1) Iij(2,2) Iij(2,3) Iij(1,2) Iij(1,3) Iij(1,1) Line with crosses Fitness value Lineij(2) Level 1 Level 2 Level 3 27 20 30 40 50 60 70 80 90 100 10 -7 10 -6 10 -5 10 -4 10 -3 10 -2 10 -1 m ea n of s ol ut io ns dimension p function f2 SSA OSA IOSA Figure 3 Results of function f2 28 20 30 40 50 60 70 80 90 100 10 2 10 3 10 4 10 5 10 6 m ea n of s ol ut io ns dimension p function f3 SSA OSA IOSA Figure 4 Results of function f3 29 20 30 40 50 60 70 80 90 100 0 1 2 3 4 5 6 7 8 9 m ea n of s ol ut io ns dimension p function f4 SSA OSA IOSA Figure 5 Results of function f4 30 20 30 40 50 60 70 80 90 100 10 0 10 1 10 2 10 3 m ea n of s ol ut io ns dimension p function f5 IOSA OSA SSA Figure 6 Results of function f5 31 20 30 40 50 60 70 80 90 100 10 2 10 3 10 4 10 5 10 6 10 7 m ea n of s ol ut io ns dimension p function f6 IOSA OSA SSA Figure 7 Results of function f6 t-values between the algorithms 0.1 1 10 100 1 2 3 4 5 r -values T- va lu es SSA - IOSA OSA - IOSA Figure 8 t-values between the algorithms 32 1 1.5 2 2.5 3 3.5 4 4.5 5 4 5 6 7 8 9 10 11 12 13 14 C om pu ta tio na l t im e (s ec on d) Value of r Computational times on the algorithms SSA OSA IOSA Figure 9 Computational time on solving the emulsified dynamite packing-machine design problem 33 Table 1 Selected test suites Parametrical benchmark functions Domain range ( )ix Minimum ( )                   +++= ∑ ∑ = − = + + P i P i ii ii xx xxNf 1 1 1 1 11 3 2 sinsin2min [3,13] P 0       +      −= ∏∑ == 1cos 4000 1 min 11 2 2 P i i P i i i x xf [-600,600] P 0 ( ) ( )∑ − = +     −+−= 1 1 222 13 1100min P i iii xxxf [-5.12,5.12] P 0 ( )             ∑ − ∑ −+= = =− P i i P i i N x N x eeef 1 1 2 20cos 2.0 4 2020min π [-30,30] P 0       += ∑ ∏ = = P i P i ii xxf 1 1 5 min [-10,10] P 0 2 1 1 6 min∑ ∑ = =         = P j j i ixf [-100,100] P 0 Table 2 Result of benchmark tests based on f1 N= 20 40 60 80 100 SSA mean 0.5701×10-3 0.6776×10-3 0.7053×10-3 0.1411×10-1 0.1522×100 SSA std 0. 6554×10-3 0.5836×10-3 0.6530×10-3 2.7248×10-3 0.1494 OSA mean 0. 4088×10-3 0.5686×10-3 0.6393×10-3 0.5546×10-2 0.0123×100 OSA std 0.1086×10-3 0.4831×10-3 0.5799×10-3 1.7035×10-3 0.9814×10-1 IOSA mean 0. 1284×10-3 0.1105×10-3 0.4091×10-3 0.1562×10-2 0.0085×100 IOSA std 0.0173×10-3 0.0854×10-3 0.4887×10-3 0.4785×10-3 0.4959×10-1 std – standard deviation 34 Table 3 Result of benchmark tests based on f2 N= 20 40 60 80 100 SSA mean 0. 3879×10-2 0.8310×10-4 0.0269 0.0549119 0.0610416 SSA std 0. 6680×10-2 1.9315×10-4 0.1473 0.2037 0.2206 OSA mean 0.0013×10-2 0.1989×10-4 0.2635×10-5 0.003210 0.02036 OSA std 0.4111×10-3 0. 2368×10-5 07335×10-5 1.7452×10-4 0.1010 IOSA mean 0.0002×10-2 0.0192×10-4 0.2286×10-6 1.410343×10- 6 1.25755×10- 6 IOSA std 0. 4765×10-4 0. 2503×10-5 0.5935×10-6 1.7535×10-6 1.0850×10-6 std – standard deviation Table 4 Result of benchmark tests based on f3 N= 20 40 60 80 100 SSA mean 1.61998×102 3.322986×104 2.01267×105 4.5998×105 7.9039×105 SSA std 1.02351×102 2.036671×104 7.195423×104 1.006506×105 1.37904×104 OSA mean 1.262328×102 4.853939×102 2.02662×103 4.04657×104 9.57764×104 OSA std 42.70728 1.99286×102 1.35091×103 2.61012×104 3.84730×104 IOSA mean 1.046856×102 4.09799×102 1.45172×103 1.6685×104 4.599238×104 IOSA std 30.61629 1.31064×102 1.095457×103 1.32495×104 1.923225×104 std – standard deviation Table 5 Result of benchmark tests based on f4 N= 20 40 60 80 100 SSA mean 5.78800 8.31013 8.18400 7.6318988 7.238934 SSA std 2.20272 0.69861 0.349668 0.24445 0.2487 OSA mean 0.9788 2.06892 3.93421 6.18335 6.170734 OSA std 0.1888 0.58577 0.75633 0.52325 0.3433 IOSA mean 0.8939 1.727303 3.60402 5.36611 5.60388 IOSA std 0.17321 0.343023 0.659474 0.57294 0.3425 std – standard deviation 35 Table 6 Result of benchmark tests based on f5 N= 20 40 60 80 100 SSA mean 6.10265 1.26377×102 2.53108×102 3.62345×102 4.673949×102 SSA std 4.66434 25.6713 23.0232 30.8571 30.6804 OSA mean 1.98888 8.42111 39.66437 1.36615×102 2.8576×102 OSA std 0.48300 1.70283 8.56296 28.47988 31.3114 IOSA mean 1.37129 7.59070 31.42920 1.17040×102 2.5174×102 IOSA std 0.362736 1.424418 9.39414 34.64055 28.96643 std – standard deviation Table 7 Result of benchmark tests based on f6 N= 20 40 60 80 100 SSA mean 5.19732×103 1.56938×105 1.11410×106 3.83418×106 8.24746×106 SSA std 1.1171×104 6.9159×104 2.73481×105 7.26947×106 1.09917×106 OSA mean 2.39963×102 1.52578×103 3.27441×104 3.42513×105 1.34091×106 OSA std 1.24722×102 2.3918×102 2.71025×104 1.22372×105 5.93870×105 IOSA mean 1.82887×102 1.4093×103 2.321762×104 2.87125×105 1.11755×106 IOSA std 1.0915×102 1.8154×102 1.28995×104 1.14011×105 3.04895×105 std – standard deviation Table 8 Results of validation tests based on the case of design optimization r= 1 2 3 4 5 SSA mean 4.1223 3.5952 4.0196 4.0130 3.9232 SSA std 0.1223×10-3 0.4756 0.1509×10-0 0.8520×10-1 0.1190×10-1 OSA mean 4.2372 4.0089 4.1333 4.1246 4.0010 OSA std 0.9903×10-6 0.4792 0.9413×10-1 0.1658×10-1 0.3740×10-1 IOSA mean 4.2373 4.3098 4.2372 4.2368 4.2373 IOSA std 0. 2679×10-5 0.4793 0.6072×10-1 0.1391×10-1 0.1264×10-1 std – standard deviation 5 Application of IOSA on design optimization and further validation 6 Conclusion and further work 
work_cmnxpvvmvfhmbamqm7uwgbfrqy	----	Evaluating health facility access using Bayesian spatial models and location analysis methods RESEARCH ARTICLE Evaluating health facility access using Bayesian spatial models and location analysis methods Nicholas J. TierneyID 1,2,3*, Antonietta Mira4,5, H. Jost Reinhold4, Giuseppe Arbia4,6, Samuel Clifford 1,2,7,8 , Angelo Auricchio 9,10,11 , Tiziano Moccetti 9 , Stefano Peluso 4,6 , Kerrie L. Mengersen 1,2 1 Queensland University of Technology, Department of Statistical Science, Mathematical Sciences, Science & Engineering Faculty, Brisbane, Queensland, Australia, 2 ARC Centre of Excellence for Mathematical and Statistical Frontiers (ACEMS), Brisbane, Queensland, Australia, 3 Department of Econometrics and Business Statistics, Monash University, Melbourne, Victoria, Australia, 4 Data Science Center, Institute of Computational Science, Università della Svizzera italiana, Lugano, Switzerland, 5 Department of Science and High Technology, Università dell’Insubria, Como, Italy, 6 Department of Statistical Sciences, Università Cattolica del Sacro Cuore, Milan, Italy, 7 Department of Infectious Disease Epidemiology, London School of Hygiene & Tropical Medicine, London, United Kingdom, 8 Centre for Mathematical Modelling of Infectious Diseases, London School of Hygiene & Tropical Medicine, London, United Kingdom, 9 Fondazione Ticino Cuore, Lugano, Switzerland, 10 Division of Cardiology, Fondazione Cardiocentro Ticino, Lugano, Switzerland, 11 Center for Computational Medicine in Cardiology, Università della Svizzera Italiana, Lugano, Switzerland * nicholas.tierney@gmail.com Abstract Background Floating catchment methods have recently been applied to identify priority regions for Auto- mated External Defibrillator (AED) deployment, to aid in improving Out of Hospital Cardiac Arrest (OHCA) survival. This approach models access as a supply-to-demand ratio for each area, targeting areas with high demand and low supply for AED placement. These methods incorporate spatial covariates on OHCA occurrence, but do not provide precise AED loca- tions, which are critical to the initial intent of such location analysis research. Exact AED locations can be determined using optimisation methods, but they do not incorporate known spatial risk factors for OHCA, such as income and demographics. Combining these two approaches would evaluate AED placement impact, describe drivers of OHCA occurrence, and identify areas that may not be appropriately covered by AED placement strategies. There are two aims in this paper. First, to develop geospatial models of OHCA that account for and display uncertainty. Second, to evaluate the AED placement methods using geospa- tial models of accessibility. We first identify communities with the greatest gap between demand and supply for allocating AEDs. We then use this information to evaluate models for precise AED location deployment. Methods Case study data set consisted of 2802 OHCA events and 719 AEDs. Spatial OHCA occurrence was described using a geospatial model, with possible spatial correlation PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 1 / 18 a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Tierney NJ, Mira A, Reinhold HJ, Arbia G, Clifford S, Auricchio A, et al. (2019) Evaluating health facility access using Bayesian spatial models and location analysis methods. PLoS ONE 14(8): e0218310. https://doi.org/10.1371/journal. pone.0218310 Editor: Tayyab Ikram Shah, The University of the South Pacific, FIJI Received: June 18, 2018 Accepted: May 31, 2019 Published: August 7, 2019 Copyright: © 2019 Tierney et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: Summary information of the key characteristics of the data are provided in the table in Appendix C in S1 File. The OHCA and AED data cannot be shared due to their sensitive nature. The Ticino Registry of Cardiac Arrest (TIRECA) (Mauri et al, 2016 (Reference 17 in manuscript), http://www. ticinocuore.ch/en) is a registry which contains private and sensitive information, and as part of establishing the registry it was required that they cannot be shared publicly. For further information on the data available in the registry, and to discuss http://orcid.org/0000-0003-1460-8722 https://doi.org/10.1371/journal.pone.0218310 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0218310&domain=pdf&date_stamp=2019-08-07 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0218310&domain=pdf&date_stamp=2019-08-07 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0218310&domain=pdf&date_stamp=2019-08-07 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0218310&domain=pdf&date_stamp=2019-08-07 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0218310&domain=pdf&date_stamp=2019-08-07 http://crossmark.crossref.org/dialog/?doi=10.1371/journal.pone.0218310&domain=pdf&date_stamp=2019-08-07 https://doi.org/10.1371/journal.pone.0218310 https://doi.org/10.1371/journal.pone.0218310 http://creativecommons.org/licenses/by/4.0/ http://www.ticinocuore.ch/en http://www.ticinocuore.ch/en accommodated by introducing a conditional autoregressive (CAR) prior on the municipality- level spatial random effect. This model was fit with Integrated Nested Laplacian Approxima- tion (INLA), using covariates for population density, proportion male, proportion over 65 years, financial strength, and the proportion of land used for transport, commercial, build- ings, recreation, and urban areas. Optimisation methods for AED locations were applied to find the top 100 AED placement locations. AED access was calculated for current access and 100 AED placements. Priority rankings were then given for each area based on their access score and predicted number of OHCA events. Results Of the 2802 OHCA events, 64.28% occurred in rural areas, and 35.72% in urban areas. Additionally, over 70% of individuals were aged over 65. Supply of AEDs was less than demand in most areas. Priority regions for AED placement were identified, and access scores were evaluated for AED placement methodology by ranking the access scores and the predicted OHCA count. AED placement methodology placed AEDs in areas with the highest priority, but placed more AEDs in areas with more predicted OHCA events in each grid cell. Conclusion The methods in this paper incorporate OHCA spatial risk factors and OHCA coverage to identify spatial regions most in need of resources. These methods can be used to help understand how AED allocation methods affect OHCA accessibility, which is of significant practical value for communities when deciding AED placements. Introduction To best serve the public, hospitals, medical centres, and emergency services should be in loca- tions where they can serve the most people in need. Recent research has evaluated various techniques for evaluating or identifying locations of External Defibrillators (AEDs) [1, 2]. This is an idea upon which this paper builds: exploring how to evaluate, locate, and improve AED locations. AEDs are a portable device that can be used by a layperson to provide advanced life support for out of hospital cardiac arrests (OHCAs) by bringing the device to the victim. OHCAs are a major public health problem affecting about 1 citizen per 1000 inhabitants in developed coun- tries [3–5]. OHCA survival decreases up to 10% for each minute of delay between collapse and treatment, but survival can be improved through delivery of cardiopulmonary resuscitation (CPR) from bystanders, and early defibrillation [6] through devices such as AEDs. Bystander response to OHCA events improve OHCA survival by performing CPR and providing advanced life support using AEDs [3, 4]. The impact of bystander AED use is increasing with technology such as smartphone applications being used to assist in responding to and treating OHCA events, improving survival [7]. AEDs need to be close to OHCA events so they can be used quickly in response to an OHCA event—ideally within 100m or a 2 minute walk to provide adequate coverage [2, 8, 9]. AEDs are currently placed following methodology prescribed by the American Heart Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 2 / 18 data access, please contact Mr. Claudio Benvenuti (claudio.benvenuti@ticinocuore.ch) at Fondazione Ticino Cuore. Simulated data, code for the analysis, and further links to the third party census, housing, geospatial, building map, and shapefile data sources used in the paper are made available at https://zenodo.org/record/3260038. Funding: One of the authors (NJT) wishes to acknowledge the receipt of an Australian Postgraduate Award from the Australian Government and QUT. Financial support from Fondazione Fratelli Agostino Enrico Rocca is acknowledged. The author Antonietta Mira was partially supported by SNF grant 105218_166504. The study was supported by a grant of the Swiss Heart Foundation “Lay people involvement in out- of-hospital cardiac arrest resuscitation: strategies for improvement and impact on survival”. Competing interests: The authors have declared that no competing interests exist. https://doi.org/10.1371/journal.pone.0218310 mailto:claudio.benvenuti@ticinocuore.ch https://zenodo.org/record/3260038 Association (AHA) and European Resuscitation Council (ERC), with AEDs being placed where OHCA events occur every 2 and 5 years, respectively. However, it is prohibitively expensive to place AEDs where all OHCA events occur. Alternative methods for AED place- ment need to be considered. Population-based strategies place AEDs in locations with high observed OHCA occurrence [10], but these may not generalise well to other areas. For example, golf courses had high OHCA incidence in one region, but not another [2, 10, 11]. This motivates the need for more sophisticated AED placement strategies. AED access can be improved by modelling access as a relationship between supply (AEDs locations), and demand, (OHCA events). This can be achieved using a method known as the two-step floating catchment area (henceforth 2SFCA), a special case of the gravity-based model, which has been used to measure healthcare access [12]. As the name suggests, the 2SFCA involves two steps. The first step calculates the supply-to-demand ratio for each supply point, which measures how well the supply meets the demand. The second step calculates the access for the spatial catchment. A modification of the floating catchment approach has been applied to identify priority regions for AED placement based on supply of AEDs and demand of OHCA events [1]. Here, the authors incorporated exponential weights to account for the step-wise decay of distance access in a given catchment, to give more weight to OHCA events that are closer to AEDs, rather than equally weighting them. Demand was measured by incorporating risk factors for OHCA occurrence, such as age, gender, income, and land use information. A Bayesian geospa- tial model was used to account for these factors, and estimate the count of OHCA events, which were then used to compute access in a geospatial region. Following this, the authors identified priority areas for AED placement based on the supply-to-demand ratio. Their approach is incomplete, however, as it does not identify exact AED locations, or provide meth- ods for precise placement. This means AEDs could be poorly placed within a region identified as priority. There is a missed opportunity to compare and combine information from priority regions with exact ideal AED locations. Mathematical optimisation strategies have improved AED access [2, 8, 9]. This approach identifies precise AED locations that optimally cover as many OHCA events in a set distance (e.g., 100m) as possible. These models are based on the Maximal Covering Location Problem, originally described by Church and Velle [13]. The approach has been applied to AED place- ments by Chan et al [8], where they were shown to be more effective than population-based approaches. This approach is referred to as a fixed location method, as AED locations are fixed and cannot be relocated. Recent research [14] has further validated fixed-location optimisation strategies, demonstrating that it is effective in rural and urban areas. Additionally, they dem- onstrated the potential efficacy of a cost effective relocation approach, where existing AEDs are relocated to cover more OHCA events, covering up to 50% of OHCA events. These two approaches for AED placement work at different scales and use differently struc- tured information. The first incorporates spatial covariates related to OHCA occurrence, and OHCA and AED locations to identify priority regions for AED placement, but does not iden- tify precise AED locations. The second approach identifies precise AED locations based on previous OHCA events, but does not consider other known risk factors for OHCA, such as age, population density, and gender [1, 15, 16] into the AED placements. It is currently unknown whether these optimisation approaches are allocating AEDs to regions needing access, only that they improve coverage. The AED placements also do not take into account other more global spatial predictors of OHCA occurrence. These two approaches could be used together to jointly evaluate the impact of AED placements and drivers of OHCA Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 3 / 18 https://doi.org/10.1371/journal.pone.0218310 occurrence. This provides a more comprehensive approach and identifies precise AED place- ments, priority areas for AED placements, and describes drivers of OHCA events. This paper has two aims. The first is to develop geospatial models of OHCA, accounting for and displaying uncertainty. The second is to evaluate the AED placement methods using geos- patial models of accessibility. To achieve our aims we identify communities with the greatest gap between demand and supply for allocating AEDs, and then use this approach to evaluate models for precise AED location deployment. This paper proceeds as follows, we first describe the case study data and data processing. Then, we explain how spatial effects are measured, how access is measured, and how we measure access and priority ranking. We then describe the optimal allocation of AEDs. Results are then presented, and discussed, along with ideas for future research and final conclusions. Materials and methods Ethics Data are collected and stored following Good Clinical Practice Guidelines and the relevant leg- islation governing the use of patient data. The investigation complied with the Declaration of Helsinki”s principles for physicians engaged in biomedical research involving human subjects. The Queensland University of Technology Human Research Ethics Committee assessed this research as meeting the conditions for exemption from HREC review and approval in accor- dance with section 5.1.22 of the Australian National Statement on Ethical Conduct in Human Research. Data Data in this study can be broken up into two different categories: (1) OHCA and AED related data, (2) and spatial and statistical data. Data: OHCA and AED related data. OHCA was data obtained from a cardiac arrest reg- istry based in Ticino, Switzerland, which has a population of 346,539 (as of 2013) and covers 2812 km 2 [17]. The data contains OHCA events from January 1st 2005 to December 31st 2015 for individuals older than 1 year. OHCA events in this registry are defined as events that occur outside of a hospital, where there is cessation of cardiac mechanical activity, and is confirmed by the absence of signs of circulation. OHCA events are geolocated as GPS latitude and longi- tude co-ordinates. A small number of cases matched to written addresses, which places events in the centre of a street or suburb. Data for AEDs contain GPS co-ordinates and availability: either public and available 24 hours a day 7 days a week, or time limited where AEDs are in a non-public structure. Locations for potential AEDs included GPS co-ordinates for every build- ing in Ticino, approximately 118,000 locations. These data were provided by the Federal Statis- tical Office, Federal Register of Buildings and Dwellings. Data: Spatial and statistical data. Shapefiles of the spatial polygons related to the 115 municipalities were made available from the Ufficio del catasto e dei riordini fondiari of Can- ton Ticino. General information on each municipality, including the financial strength of each area, was provided by USTAT, the Statistical Office of Canton Ticino. Financial strength was computed as a weighted average of indicators related to revenue, taxation, and the resident population. For complete details on the calculation of financial strength see Appendix A in S1 File. The land use, specifically, the number of hectares used for transport, industry and com- mercial use, buildings, recreation, and special urban use, was also available for each of these municipalities, and provided by the Federal Office of Statistics. Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 4 / 18 https://doi.org/10.1371/journal.pone.0218310 Overview of methods A brief overview of each step of the analysis is now provided. First, A grid is overlaid over the spatial area, the number of OHCA events in each grid cell are calculated, and each grid cell is assigned covariate information from the municipality it falls into. Next, a Bayesian geospatial model is fit to obtain an estimate of the expected number of OHCA events in a given grid cell. Following this a spatial access model is fit at each grid cell, providing each grid with an access score. The fixed location method is then used with the observed OHCA events to identify 100 optimal AED locations. Priority rankings are then created. Each of these steps is now explained in more detail. Data pre-processing One problem that must be tackled is that the data used in analyses are collected at different spatial scales. Specifically, OHCA events, AED locations, and building locations have GPS co- ordinates, and thus observed at the point level, while census information of the other variables are obtained at the spatial scale of municipalities. However, the municipality areas are too large for the purposes of the floating catchment analysis, so in the absence of other readily available information, a grid was generated to provide smaller resolution for analysis. This grid was defined with dimensions 0.01 degrees latitude and longitude, resulting in 3205 grid cells 1100m by 1100m (see Fig 1). The number of OHCA events that fall within each grid cell were then calculated. Covariate information for each grid cell was calculated from the municipality in which the grid cell centroid fell, and was divided by the number of cells falling within the borders of each municipality. For example, if ten grid cells fell within a municipality, the covar- iate values for each grid were given by dividing the municipality covariate value by ten. Fig 1 shows the grid created, and a map showing the number of OHCA events that lie within each spatial area. Modelling spatial effects Bayesian models provide simple and effective ways of analysing small effects, and allow infer- ences using probabilities of potential events and outcomes. A model is fit to estimate the num- ber of OHCAs in a given grid cell. The following covariates have been shown to be related to OHCA incidence [1, 15, 16], and so are included in the model: proportion of men, proportion of people aged over 65, financial strength of an area, and proportion of area used for transport, industry and commercial uses, buildings, recreation, and special urban use. As 85% of the spa- tial grid cells have zero OHCA events and many low counts (Fig 2), a Zero-Inflated Poisson (ZIP) model is used to account for these additional zeros. A ZIP model allows the number of events yki in the i th grid cell in the kth municipality to be zero with some probability p, or else to follow a Poisson distribution (which can also have zeros) with probability 1 − p: yki � 0; with probability p Poisson ðlkiÞ; with probability 1 � p ð1Þ ( The probability p is expressed as: p ¼ expðyÞ 1 þ expðyÞ ð2Þ and a prior is set on θ. In this case study θ was set to be y � Nð0:000001; 2Þ. Hence a priori, p has a mean of 0.5, and 2.5 th and 97.5 th percentiles of 0.05 and 0.95, respectively. Other Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 5 / 18 https://doi.org/10.1371/journal.pone.0218310 specifications of θ were considered (for example, y � Nð0:001; 2Þ, and y � Nð0:1; 1Þ), but these had very little impact on the parameter estimates. The parameter λki was modelled as a function of the risk factors x = (x1, . . ., xm) as in Eq 3. In this model, ui is a spatially structured random effect with an intrinsic conditional autore- gressive specification [18, 19], and vi is a spatially unstructured random effect, (see Eqs 4 and Fig 1. Map of the OHCA events, where the higher the number of OHCA events, the darker the bin colour. https://doi.org/10.1371/journal.pone.0218310.g001 Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 6 / 18 https://doi.org/10.1371/journal.pone.0218310.g001 https://doi.org/10.1371/journal.pone.0218310 5, respectively). logðlkiÞ¼ b0 þ XM m¼1 bmxmi þui þvi ð3Þ uijuj;tu � N 1 ni X i�j uj; 1 tuni ! ; i 6¼ j ð4Þ vi ¼ Nð0;svÞ ð5Þ Here τu is a precision parameter; ni is the number of neighbours of the i th grid cell, and the sub- script i * j refer to a cell i of the grid cell and to its neighbour j respectively. An alternative for- mulation that accounts for age and gender as an offset was also considered and is described in Appendix B in S1 File. Measuring accessibility and priority ranking An altered 2SFCA method is applied, which uses an exponential decay from the AED to OHCA distance, as described in [1]. Our adoption of a exponential decay function extends the 2SFCA method of [1] to rank geographical regions according to their accessibility. We believe that the continuous decreasing behaviour of the demand weight is a better approximation of the reality, relative to a step-wise decreasing function. Also, the chosen decay can be inter- preted in terms of demand or in terms of procedure effectiveness: after the threshold of 100m, we consider, based on literature considerations, the AED not effective to cover an OHCA that will likely divert on a different AED (therefore no demand for the first AED). This process has two steps, and follows the notation where i is a grid cell, j is an AED point location, and k is a municipality. The first step calculates the supply to demand ratio, Rj, for each AED location. The supply of Sj is defined as 1 as there is 1 AED, and demand is the weighted sum of the demand scores of an OHCA event from spatial area dkj within distance d0 (100m) of AED j. The choice of 100m is not arbitrary and is based on current literature, [2, 8, 9], and dictated by resuscitation guidelines [20]. To elaborate, without cardiopulmonary resuscitation and Fig 2. Distribution of non-zero OHCA counts in each grid cell from 2005 to 2015. https://doi.org/10.1371/journal.pone.0218310.g002 Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 7 / 18 https://doi.org/10.1371/journal.pone.0218310.g002 https://doi.org/10.1371/journal.pone.0218310 AED use, the probability of surviving an OHCA is reduced by 10-15% every minute of cardiac arrest. A health untrained person can run at 2 metres per second, so a distance of 500m will be covered in about 250 seconds or about 4 minutes. This means if an AED is 500 metres away, 4 minutes equals a probability to survive of slightly more than 50%, which is very undesirable. This is a further medical justification to optimize AED distribution to reduce access time to 1 to 2 minutes or up to a distance of 100 metres. Thus, the paper focuses only on bystander response of 100m. We report the results based on a different distance of 250m in Appendices F and G in S1 File, but this caused a mismatch between AED placements and OHCA predic- tions. Following this, we decide to focus on bystander response up to 100m. The ratio of supply to demand is then: Rj ¼ Sj P k2fdkj�d0g DkiGðdkj;d0Þ ð6Þ Here, the estimated demand score, Dki for the OHCA in grid cell i of municipality k, is weighted by a exponential function with a smooth decay, set to be equal to zero at d0: Gðdkj;d0Þ¼ e � 1 2 � dkj d 0 � �2 � e � 1 2 1 � e � 1 2 ; dkj � d0 0 dkj > d0 ð7Þ 8 >>>>>>< >>>>>>: An alternative, stepwise formulation was considered, but the results were effectively identi- cal to the exponential decay (see Appendices E, F, and G in S1 File). In the second step, access, Ai, aggregates this information, such that for each spatial area i is calculated by summing up the weighted supply-to-demand ratios for each AED that fall within spatial area i: Ai ¼ XJ j¼1 Rj ð8Þ Priority ranks are created by considering regions with predicted OHCA counts greater than 1 in a grid cell, then ordering by the lowest access score and highest predicted OHCA counts. To find the areas most in need of resources, we consider the top 20 priority areas and then rank them according to those that had the most AEDs placed. The optimal allocation method for AEDs AED locations are identified using the fixed location method optimisation and the observed OHCA events. This identifies a set of AED locations that maximize the number of OHCA events covered within a set distance of an AED. Possible AED locations are given by the data- base of buildings in Ticino. A number of new AEDs can be specified, for example, the top 100 AED locations covering the most historical OHCA events. More formally, the variables xj, and yi are binary, where xj is equal to 1 when OHCA j is covered, and 0 otherwise for j = 1, . . ., J OHCA events. Similarly, yi is 1 if an AED is placed in location i, and 0 otherwise, for possible AED locations i = 1, . . ., I. The matrix A has J rows of OHCA incidents, and I columns of potential AED locations. Here, aji is binary, and indicates Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 8 / 18 https://doi.org/10.1371/journal.pone.0218310 whether OHCA j is covered (within 100m) by location i. A ¼ ½aj;i� ¼ 0 1;1 1 1;2 � � � 0 1;I 1 2;1 0 2;2 � � � 1 2;I .. . .. . . . . .. . 0J;1 1J;2 � � � 1J;I 0 B B B B B B @ 1 C C C C C C A ð9Þ This optimisation model maximizes the total number of OHCAs covered by the configura- tion of the AEDs yi, max XJ j¼1 xj ð10Þ subject to the constraints: only N locations for AED placement are selected, XI i¼1 yi ¼ N ð11Þ and that when at least one AED i covers an OHCA j, and the corresponding AED yi is selected, then OHCA xj is covered. xj � XI i¼1 aijyi8j¼1...J ð12Þ OHCAs that are already covered by current AED placements are removed from the analy- sis, to improve only coverage of uncovered OHCAs. The fixed location model is performed for N = 100, to find the top 100 locations for AEDs. Computation. The R statistical and programming environment [21] was used for all anal- ysis and visualisation, along with the integrated design environment, RStudio [22]. Reproduc- ibility was ensured using the rmarkdown [23] and knitr [24] packages. The R package, maxcovr [25] was used to determine optimal AED locations, using lpSolve internally to as the linear programming solver for the fixed location method [26]. Data read in was per- formed using readr and readxl [27, 28], and data manipulation and results extraction used R packages dplyr, tidyr, purrr, and kableExtra [29–32]. Spatial and statistical analysis was performed using the packages simple features, sp, spdep and raster [33–36]. The ZIP model was fit using Integrated Nested Laplace Approximation (INLA), using option “zeroinflatedpoisson1” [37]. Results Among the 2802 OHCA cases in this study, over 70% of OHCA events occurred in those aged over 65 (Table 1). Additionally, the majority of events occurred in urban areas (Table 2). Table 1. Age distribution of OHCA patients. Age (Years) N Percent 0-15 3 0.11 15-64 742 26.48 65-105 2044 72.95 (Missing) 13 0.46 https://doi.org/10.1371/journal.pone.0218310.t001 Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 9 / 18 https://doi.org/10.1371/journal.pone.0218310.t001 https://doi.org/10.1371/journal.pone.0218310 The model performs well at predicting expected counts in a region (Fig 3). Fig 4 uses sam- ples from the posterior density of each covariate, and shows the 95% credible intervals, and the probability of the covariate being greater than zero in a grid cell. Note here that recreational and industrial areas, as well as the proportion of men, being older than 65, and the population density increase OHCA occurrence. Additionally, the proportion of buildings, urban areas, and financial strengths are associated with lower counts of OHCA events. The posterior mean, standard deviation, and residuals (observed count—posterior mean) are depicted in Fig 5. Areas with values greater than zero indicate that observed counts of OHCA are greater than model expected values. Areas with values less than zero indicate that observed counts of OHCA events are less than model expected values. Areas marked with darker colours indicate higher occurrence. The posterior mean values are similar to the observed values, and some areas have a high standard deviation at the edges of the map. The presence of possible spatial autocorrelation of the regression residuals is tested using the Mor- an’s I test [38, 39]. The global Moran’s I value is I = -0.03, with p = 0.989, indicating no signifi- cant overall spatial clustering in the residuals of the model. Areas with high difference in observed and predicted are mostly in rural areas. The top 10 grid cell locations with the most OHCA events are mostly in urban regions (Table 2). A table of the top 20 municipalities by model error is shown in Appendix B in S1 File. The analysis reveals that almost all areas (99.1%) had an access score of less than one. Prior- ity ranks were created and access scores of each spatial area are shown in Fig 5 (bottom right), where areas shown in grey indicate locations that have a predicted OHCA count of < 1. Tables 3 and 4 show the number of AEDs placed in the the top 20 least accessible areas. The top 10 priority areas have AEDs placed in them, but priority ranks 11 to 20 have only 5 AEDs placed, 4 of which are in rural areas. When examining the top 20 grid cell areas where AEDs are placed, the optimisation favours urban areas, which have higher model predicted OHCA counts. Discussion This paper presents a novel modelling framework that jointly prioritizes regions for AED placement, evaluates AED placement methods, and identifies covariates important for predict- ing OHCA events. This reconciles a missed opportunity in past research by [1], which did not identify exact AED locations in priority regions, and in [14], which did not combine exact AED locations with relevant spatial information important in predicting OHCA occurrence [1, 15, 16]. Combining information from geospatial models and the precise facility locations Table 2. Top 10 Municipalities by OHCA count, also indicating whether they are urban or rural areas. Municipality Urban / Rural Count Percent 69 Urban 489 17.45 66 Urban 130 4.64 78 Urban 127 4.53 13 Urban 126 4.50 36 Urban 70 2.50 51 Rural 67 2.39 7 Rural 64 2.28 49 Rural 60 2.14 82 Rural 60 2.14 14 Urban 59 2.11 https://doi.org/10.1371/journal.pone.0218310.t002 Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 10 / 18 https://doi.org/10.1371/journal.pone.0218310.t002 https://doi.org/10.1371/journal.pone.0218310 gives a more complete picture of the impact of facility locations. The geospatial model revealed an increase of OHCA occurrence for the covariates: recreational and industrial areas, and the proportion of men, proportion of the population older than 65, and the population density. Additionally, the proportion of buildings, urban areas, and financial strengths are associated with lower counts of OHCA events. A limitation of this study is that the currently available shapefiles are at a low spatial resolu- tion. This lead us to make the unrealistic assumption that municipalities have their covariates equally distributed within the surface area. Data at a higher spatial resolution could also help to glean more useful insight into the relationship between OHCA risk factors and AED place- ment. However, a main aim of this paper was to demonstrate finding factors related to OHCA Fig 3. Observed counts of OHCA events in each grid cell compared to the posterior mean counts in a grid cell, with a line of perfect fit. https://doi.org/10.1371/journal.pone.0218310.g003 Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 11 / 18 https://doi.org/10.1371/journal.pone.0218310.g003 https://doi.org/10.1371/journal.pone.0218310 occurrence, and using the floating catchment approach and AED allocation methods in evalu- ating priority areas and placement methodology. The distance to an existing OHCA event is crucial to survival, but is only a measure of past occurrence. Incorporating floating catchment area models with spatial modelling of relevant spatial risk factors for OHCA events helps us understand how AED deployment strategies work in spatial areas with higher OHCA occurrence risk factors (more men, higher population density and older populations). There may be situations with scarcity of services, where multi- ple OHCA events happen and exhaust available AEDs. To fully account for such a situation, additional modelling would need to be made where a “wrong decision” was made, where per- haps an agent made a false trip to where an AED was located previously. These have not been modelled yet, although scenarios have been considered in [2], which modelled scenarios with many active bystanders searching for AEDs to use for treatment, and bystanders found AEDs farther away from the bystander. Use of the binary coverage in the optimisation model means that areas are either covered or not. This means that an OHCA event 99m and 1m away from an AED location get the same weight, of 1, and an AED 101m away gets a weight of 0—it is marked as being not covered. Although accessibility score in this paper uses a exponential decay, an approach that considers continuous distance has not been considered in AED placement. Continuous distance could be incorporated, using what is known as the continuous coverage problem, described in [40, 41]. This could perhaps incorporate factors such as spatial risk or lives saved through AED placement, so that that spatial areas with higher OHCA risk could have more AEDs. Fig 4. Posterior mean, 95% credible intervals, and probability of effect being > 0 for the model covariates. https://doi.org/10.1371/journal.pone.0218310.g004 Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 12 / 18 https://doi.org/10.1371/journal.pone.0218310.g004 https://doi.org/10.1371/journal.pone.0218310 Fig 5. Top Left: Posterior mean estimates of the number of OHCA events in each spatial area, Top Right: Posterior Estimates of Standard Deviation of the number of OHCA events in each spatial area, Bottom Left: Error as measured by the observed—posterior mean of the number of OHCA events in each spatial area. Bottom Right: Access score areas with predicted OHCA count of less than 1 are ignored (grey). https://doi.org/10.1371/journal.pone.0218310.g005 Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 13 / 18 https://doi.org/10.1371/journal.pone.0218310.g005 https://doi.org/10.1371/journal.pone.0218310 Table 3. Top 20 priority ranks, access scores, model predictions, the number of AEDs added, and whether the areas were rural or urban. Priority Rank Access Model Predicted # AEDs Added Urban / Rural 1 0 32.22 3 rural 2 0 23.24 1 urban 3 0 21.33 2 rural 4 0 18.77 2 rural 5 0 17.51 1 urban 6 0 14.74 1 urban 7 0 14.14 1 rural 8 0 11.10 1 rural 9 0 10.40 1 rural 10 0 10.40 1 rural 11 0 10.30 1 rural 12 0 9.57 0 urban 13 0 9.00 1 urban 14 0 8.91 1 rural 15 0 8.58 0 rural 16 0 8.31 0 rural 17 0 8.21 0 urban 18 0 8.06 1 rural 19 0 7.83 0 rural 20 0 7.79 1 rural https://doi.org/10.1371/journal.pone.0218310.t003 Table 4. Top 20 number of AEDs added, along with priority ranks, access scores, model predictions, and whether the areas were rural or urban. Priority Rank Access Model Predicted # AEDs Added Urban / Rural 229 0.07 74.37 6 urban 261 0.18 43.60 5 urban 240 0.10 46.35 4 urban 1 0.00 32.22 3 rural 244 0.10 42.22 3 urban 284 0.50 44.79 3 urban 312 3.83 42.41 3 urban 3 0.00 21.33 2 rural 4 0.00 18.77 2 rural 209 0.02 24.99 2 rural 221 0.05 19.65 2 rural 223 0.06 32.32 2 rural 235 0.09 32.45 2 urban 237 0.09 19.75 2 urban 252 0.13 34.14 2 urban 255 0.15 25.86 2 rural 260 0.18 20.36 2 rural 281 0.49 33.89 2 rural 302 1.19 38.04 2 urban 2 0.00 23.24 1 urban https://doi.org/10.1371/journal.pone.0218310.t004 Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 14 / 18 https://doi.org/10.1371/journal.pone.0218310.t003 https://doi.org/10.1371/journal.pone.0218310.t004 https://doi.org/10.1371/journal.pone.0218310 Information on risk of death from an OHCA event, or perhaps some measures of accessibility, could also be included so that areas with greater risk are given more weight. In our work and in previous literature [1], the accessibility models only used the mean pos- terior values from the geospatial model. The Bayesian framework can be used to recover distri- butions of accessibility for each spatial region, which would account for uncertainty in accessibility, and allow for calculation of probabilities of spatial regions being ranked as the lowest access, and calculation of regions in the lowest/top 10% accessibility of a region. This could be useful when determining priority rankings, allowing for questions regarding not just the access score, but the certainty of access. In practical cases, there are many more rural areas with low access and high demand com- pared to urban areas, and so additional approaches are needed for servicing rural regions with low access. One response to this need is the use of drone deployed AEDs [42, 43]. These differ- ent uses of AED deployment could in the future be combined with current deployment meth- ods, potentially included as an additional cost to regular AEDs, and also constrained, according to areas with lower access scores. Deploying AEDs and drone delivery services is an expensive endeavour, so future methods could, in addition, incorporate the use of AED deployment methods, into a relocation-type model. This could, for example, explore the use of relocating several AEDs in particular regions to one central node containing an AED delivery drone. Alternative arrival vectors, such as those by an ambulance, may be considered for AED delivery. However, their modelling requires answering a different set of questions, concerning placement of ambulances and of ambulance stations, and accounting for additional external features such as traffic at a given time of day. Additionally, an ambulance may take up to 7 or 8 minutes to arrive on scene, which is beyond a reasonable survival time, which may mean other arrival vectors such as motorcycles may be considered. Future research could explore different distances (100m–1km) for different modes of trans- port, such as healthy runners, motorcycles, and ambulances. The arrival and approach of healthy runners and an ambulance should be addressed in a different framework that assesses the full combinations of these events. This might include situations such as race conditions, where someone starts at 100m, someone runs 400m, and an ambulance arrives. This is a com- plex analysis, and is partially discussed in Chan et al 2016 [2], where many different optimisa- tion formulations account for different scenarios of arrival. Furthermore, larger distances should account for available paths/routes instead of assuming a straight line of travel. Path dis- tances are typically challenging to calculate, but modern tools such as dodgr provide accessi- ble calculations of many-to-many pairwise distances for flow via networks, providing realistic routing through streets [44]. This paper creates priority rankings comprised of spatial risk factors and coverage informa- tion from OHCA and AED events. This information could be used to create a table of ranks and access scores that consider changes in access for different AED placements. It also pro- vides a different perspective for evaluating AED placement methodology, allowing for different questions and decisions to be made in AED placement. Supporting information S1 File. Appendix A, Definition of financial strength. Appendix B, Alternative formulation of spatial regression that accounts for age and gender in an offset term. Appendix C, Top 20 Municipalities by model error count, and their model error and predictions. Appendix D, Summary of key characteristics of OHCA data. Appendix E, Stepwise and Exponential Decay. Appendix F, Figure of different access models. Appendix G, Tables of exponential and stepwise Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 15 / 18 http://www.plosone.org/article/fetchSingleRepresentation.action?uri=info:doi/10.1371/journal.pone.0218310.s001 https://doi.org/10.1371/journal.pone.0218310 decay for priority and top AEDs. (PDF) Acknowledgments The authors acknowledge Prof. Rolf Krause and Prof Martin Weiser for their input into the optimisation modelling, the Federazione Cantonale Ticinese Servizi Ambulanze to give us access to the data, the Swiss Federal Statistical Office, GEWO, in the person of Mauro Nanini, for the valuable information on buildings locations in Ticino, the Cantonal Statistical Office of Canton Ticino (USTAT) for their advice and suggestions, Mr. Matthew Sutton at QUT for his help and input into the optimisation modelling, Moreover, we would like to thank Chantal Zengaffinen for her invaluable work in cleaning up data and performing preparatory work for geolocation of OHCAs. Author Contributions Conceptualization: Nicholas J. Tierney, Antonietta Mira, Samuel Clifford, Angelo Auricchio, Kerrie L. Mengersen. Data curation: Nicholas J. Tierney, H. Jost Reinhold, Angelo Auricchio, Tiziano Moccetti. Formal analysis: Nicholas J. Tierney, Antonietta Mira, Samuel Clifford, Kerrie L. Mengersen. Methodology: Nicholas J. Tierney, Antonietta Mira, Angelo Auricchio, Stefano Peluso. Resources: Angelo Auricchio, Kerrie L. Mengersen. Software: Nicholas J. Tierney. Supervision: Antonietta Mira, Giuseppe Arbia, Samuel Clifford, Kerrie L. Mengersen. Visualization: Nicholas J. Tierney, Kerrie L. Mengersen. Writing – original draft: Nicholas J. Tierney, Antonietta Mira, Giuseppe Arbia, Angelo Aur- icchio, Kerrie L. Mengersen. Writing – review & editing: Nicholas J. Tierney, Antonietta Mira, H. Jost Reinhold, Giuseppe Arbia, Samuel Clifford, Angelo Auricchio, Tiziano Moccetti, Stefano Peluso, Kerrie L. Mengersen. References 1. Lin B-C, Chen C-W, Chen C-C, Kuo C-L, Fan I-C, Ho C-K, et al. Spatial decision on allocating auto- mated external defibrillators (AED) in communities by multi-criterion two-step floating catchment area (MC2SFCA). Int J Health Geogr. 2016; 15: 17. https://doi.org/10.1186/s12942-016-0046-8 PMID: 27225882 2. Chan TCY, Demirtas D, Kwon RH. Optimizing the deployment of public access defibrillators. Manage- ment science. 2016; 62: 3617–3635. https://doi.org/10.1287/mnsc.2015.2312 3. Marenco JP, Wang PJ, Link MS, Homoud MK, Estes NA 3rd. Improving survival from sudden cardiac arrest: The role of the automated external defibrillator. JAMA. 2001; 285: 1193–1200. https://doi.org/10. 1001/jama.285.9.1193 PMID: 11231750 4. Weisfeldt ML, Becker LB. Resuscitation after cardiac arrest: A 3-phase time-sensitive model. JAMA. 2002; 288: 3035–3038. https://doi.org/10.1001/jama.288.23.3035 PMID: 12479769 5. Gräsner J-T, Lefering R, Koster RW, Masterson S, Böttiger BW, Herlitz J, et al. EuReCa ONE-27 nations, ONE europe, ONE registry: A prospective one month analysis of out-of-hospital cardiac arrest outcomes in 27 countries in europe. Resuscitation. 2016; 105: 188–195. https://doi.org/10.1016/j. resuscitation.2016.06.004 PMID: 27321577 6. Link MS, Atkins DL, Passman RS, Halperin HR, Samson RA, White RD, et al. Part 6: Electrical thera- pies. Circulation. American Heart Association, Inc. 2010; 122: S706–S719. Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 16 / 18 https://doi.org/10.1186/s12942-016-0046-8 http://www.ncbi.nlm.nih.gov/pubmed/27225882 https://doi.org/10.1287/mnsc.2015.2312 https://doi.org/10.1001/jama.285.9.1193 https://doi.org/10.1001/jama.285.9.1193 http://www.ncbi.nlm.nih.gov/pubmed/11231750 https://doi.org/10.1001/jama.288.23.3035 http://www.ncbi.nlm.nih.gov/pubmed/12479769 https://doi.org/10.1016/j.resuscitation.2016.06.004 https://doi.org/10.1016/j.resuscitation.2016.06.004 http://www.ncbi.nlm.nih.gov/pubmed/27321577 https://doi.org/10.1371/journal.pone.0218310 7. Caputo M, Muschietti S, Burkart C, Benvenuti C, Conte G, Regoli F, et al. Lay persons alerted by mobile application system initiate earlier cardiopulmonary resuscitation: A comparison with sms- based system notification. Resuscitation. 2017. https://doi.org/10.1016/j.resuscitation.2017.03.003 8. Chan TCY, Li H, Lebovic G, Tang SK, Chan JYT, Cheng HCK, et al. Identifying locations for public access defibrillators using mathematical optimization. Circulation. 2013; 127: 1801–1809. https://doi. org/10.1161/CIRCULATIONAHA.113.001953 PMID: 23553657 9. Sun CLF, Demirtas D, Brooks SC, Morrison LJ, Chan TCY. Overcoming spatial and temporal barriers to public access defibrillators via optimization. Journal of the American College of Cardiology. 2016; 68: 836–845. https://doi.org/10.1016/j.jacc.2016.03.609 PMID: 27539176 10. Becker L, Eisenberg M, Fahrenbruch C, Cobb L. Public locations of cardiac arrest. Implications for pub- lic access defibrillation. Circulation. 1998; 97: 2106–2109. https://doi.org/10.1161/01.cir.97.21.2106 PMID: 9626169 11. Engdahl J, Herlitz J. Localization of out-of-hospital cardiac arrest in göteborg 1994–2002 and implica- tions for public access defibrillation. Resuscitation. Elsevier; 2005; 64: 171–175. https://doi.org/10. 1016/j.resuscitation.2004.08.006 12. White N, Mengersen K. Predicting health programme participation: A gravity-based, hierarchical model- ling approach. Journal of the Royal Statistical Society: Series C (Applied Statistics). 2016; 65: 145–166. https://doi.org/10.1111/rssc.12111 13. Richard Church CRV. The maximum covering location problem. Papers in Regional Science. 1974; 32: 101–118. https://doi.org/10.1111/j.1435-5597.1974.tb00902.x 14. Tierney NJ, Reinhold HJ, Mira A, Weiser M, Burkart R, Benvenuti C, et al. Novel relocation methods for automatic external defibrillator improve out-of-hospital cardiac arrest coverage under limited resources. Resuscitation. 2018. https://doi.org/10.1016/j.resuscitation.2018.01.055 PMID: 29414670 15. Ong MEH, Tan EH, Yan X, Anushia P, Lim SH, Leong BS-H, et al. An observational study describing the geographic-time distribution of cardiac arrests in singapore: What is the utility of geographic informa- tion systems for planning public access defibrillation? (PADS phase i). Resuscitation. 2008; 76: 388– 396. https://doi.org/10.1016/j.resuscitation.2007.09.006 PMID: 17976889 16. Folke F, Lippert FK, Nielsen SL, Gislason GH, Hansen ML, Schramm TK, et al. Location of cardiac arrest in a city center: Strategic placement of automated external defibrillators in public locations. Circu- lation. Am Heart Assoc; 2009; 120: 510–517. 17. Mauri R, Burkart R, Benvenuti C, Caputo ML, Moccetti T, Del Bufalo A, et al. Better management of out- of-hospital cardiac arrest increases survival rate and improves neurological outcome in the swiss canton ticino. Europace: European pacing, arrhythmias, and cardiac electrophysiology: journal of the working groups on cardiac pacing, arrhythmias, and cardiac cellular electrophysiology of the European Society of Cardiology. 2016; 18: 398–404. https://doi.org/10.1093/europace/euv218 18. Besag J, York J, Mollié A. Bayesian image restoration, with two applications in spatial statistics. Annals of the Institue of Statistical Mathematics. Kluwer Academic Publishers; 1991; 43: 1–20. https://doi.org/ 10.1007/BF00116466 19. Arbia G. Spatial econometrics: Statistical foundations and applications to regional convergence. Springer Science & Business Media; 2006. 20. Ringh M, Hollenberg J, Palsgaard-Moeller T, Svensson L, Rosenqvist M, Lippert FK, et al. The chal- lenges and possibilities of public access defibrillation. Journal of internal medicine. 2018; 283: 238–256. https://doi.org/10.1111/joim.12730 PMID: 29331055 21. R Core Team. R: A language and environment for statistical computing [Internet]. Vienna, Austria: R Foundation for Statistical Computing; 2018. Available: https://www.R-project.org/. 22. RStudio. RStudio [Internet]. Boston, MA, USA: RStudio; 2017. Available: http://www.rstudio.org/. 23. Allaire J, Cheng J, Xie Y, McPherson J, Chang W, Allen J, et al. Rmarkdown: Dynamic documents for r [Internet]. 2017. Available: http://rmarkdown.rstudio.com. 24. Xie Y. Dynamic documents with R and knitr [Internet]. 2nd ed. Boca Raton, Florida: Chapman; Hall/ CRC; 2015. Available: http://yihui.name/knitr/. 25. Tierney N. Maxcovr: A set of tools for solving maximal coverage problem. [Internet]. 2017. Available: https://github.com/njtierney/maxcovr. 26. Berkelaar M, others. [Internet]. 2015. Available: https://CRAN.R-project.org/package=lpSolve. 27. Wickham H, Hester J, Francois R. Readr: Read rectangular text data [Internet]. 2017. Available: https:// CRAN.R-project.org/package=readxl. 28. Wickham H. Readxl: Read excel files [Internet]. 2016. Available: https://CRAN.R-project.org/package= readxl. Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 17 / 18 https://doi.org/10.1016/j.resuscitation.2017.03.003 https://doi.org/10.1161/CIRCULATIONAHA.113.001953 https://doi.org/10.1161/CIRCULATIONAHA.113.001953 http://www.ncbi.nlm.nih.gov/pubmed/23553657 https://doi.org/10.1016/j.jacc.2016.03.609 http://www.ncbi.nlm.nih.gov/pubmed/27539176 https://doi.org/10.1161/01.cir.97.21.2106 http://www.ncbi.nlm.nih.gov/pubmed/9626169 https://doi.org/10.1016/j.resuscitation.2004.08.006 https://doi.org/10.1016/j.resuscitation.2004.08.006 https://doi.org/10.1111/rssc.12111 https://doi.org/10.1111/j.1435-5597.1974.tb00902.x https://doi.org/10.1016/j.resuscitation.2018.01.055 http://www.ncbi.nlm.nih.gov/pubmed/29414670 https://doi.org/10.1016/j.resuscitation.2007.09.006 http://www.ncbi.nlm.nih.gov/pubmed/17976889 https://doi.org/10.1093/europace/euv218 https://doi.org/10.1007/BF00116466 https://doi.org/10.1007/BF00116466 https://doi.org/10.1111/joim.12730 http://www.ncbi.nlm.nih.gov/pubmed/29331055 https://www.R-project.org/ http://www.rstudio.org/ http://rmarkdown.rstudio.com http://yihui.name/knitr/ https://github.com/njtierney/maxcovr https://CRAN.R-project.org/package=lpSolve https://CRAN.R-project.org/package=readxl https://CRAN.R-project.org/package=readxl https://CRAN.R-project.org/package=readxl https://CRAN.R-project.org/package=readxl https://doi.org/10.1371/journal.pone.0218310 29. Wickham H, Francois R. Dplyr: A grammar of data manipulation [Internet]. 2017. Available: https:// CRAN.R-project.org/package=dplyr. 30. Wickham H. Tidyr: Easily tidy data with ‘spread()’ and ‘gather()’ functions [Internet]. 2017. Available: https://CRAN.R-project.org/package=tidyr. 31. Wickham H. Purrr: Functional programming tools [Internet]. 2017. Available: https://CRAN.R-project. org/package=purrr. 32. Zhu H. KableExtra: Construct complex table with’kable’ and pipe syntax [Internet]. 2018. Available: https://CRAN.R-project.org/package=kableExtra. 33. Pebesma E. Sf: Simple features for r [Internet]. 2017. Available: https://CRAN.R-project.org/package= sf. 34. Bivand RS, Pebesma E, Gómez-Rubio V. Applied spatial data analysis with r [Internet]. 2nd ed. NY, USA: Springer New York; 2013. Available: http://www.asdar-book.org/. 35. Bivand R, Piras G. Comparing implementations of estimation methods for spatial econometrics. Journal of Statistical Software, Articles. 2015; 63: 1–36. 36. Hijmans RJ. Raster: Geographic data analysis and modeling [Internet]. 2016. Available: https://CRAN. R-project.org/package=raster. 37. Rue H, Martino S, Chopin N. Approximate bayesian inference for latent gaussian models by using inte- grated nested laplace approximations. Journal of the Royal Statistical Society Series B, Statistical meth- odology. Blackwell Publishing Ltd; 2009; 71: 319–392. https://doi.org/10.1111/j.1467-9868.2008. 00700.x 38. Moran PAP. Notes on continuous stochastic phenomena. Biometrika. JSTOR; 1950; 37: 17–23. https:// doi.org/10.1093/biomet/37.1-2.17 39. Arbia G. A primer for spatial econometrics: With applications in R. Springer; 2014. 40. Wei R, Murray AT. Continuous space maximal coverage: Insights, advances and challenges. Comput- ers & operations research. 2015; 62: 325–336. https://doi.org/10.1016/j.cor.2014.04.010 41. Murray AT. Maximal coverage location problem. International regional science review. SAGE Publica- tionsSage CA: Los Angeles, CA; 2015. 42. Claesson A, Fredman D, Svensson L, Ringh M, Hollenberg J, Nordberg P, et al. Unmanned aerial vehi- cles (drones) in out-of-hospital-cardiac-arrest. Scand J Trauma Resusc Emerg Med. 2016; 24: 124. https://doi.org/10.1186/s13049-016-0313-5 PMID: 27729058 43. Pulver A, Wei R, Mann C. Locating AED enabled medical drones to enhance cardiac arrest response times. Prehosp Emerg Care. 2016; 20: 378–389. https://doi.org/10.3109/10903127.2015.1115932 PMID: 26852822 44. Padgham M. Dodgr: An r package for network flow aggregation. Transport Findings. Network Design Lab; 2019. Bayesian analysis of health facility access PLOS ONE | https://doi.org/10.1371/journal.pone.0218310 August 7, 2019 18 / 18 https://CRAN.R-project.org/package=dplyr https://CRAN.R-project.org/package=dplyr https://CRAN.R-project.org/package=tidyr https://CRAN.R-project.org/package=purrr https://CRAN.R-project.org/package=purrr https://CRAN.R-project.org/package=kableExtra https://CRAN.R-project.org/package=sf https://CRAN.R-project.org/package=sf http://www.asdar-book.org/ https://CRAN.R-project.org/package=raster https://CRAN.R-project.org/package=raster https://doi.org/10.1111/j.1467-9868.2008.00700.x https://doi.org/10.1111/j.1467-9868.2008.00700.x https://doi.org/10.1093/biomet/37.1-2.17 https://doi.org/10.1093/biomet/37.1-2.17 https://doi.org/10.1016/j.cor.2014.04.010 https://doi.org/10.1186/s13049-016-0313-5 http://www.ncbi.nlm.nih.gov/pubmed/27729058 https://doi.org/10.3109/10903127.2015.1115932 http://www.ncbi.nlm.nih.gov/pubmed/26852822 https://doi.org/10.1371/journal.pone.0218310 
work_cn4tyqxheranlecnhco2xwfcde	----	Microsoft Word - فهرست.doc DOI: 10.22094/JOIE.2018.242.1532 131 Three Hybrid Metaheuristic Algorithms for Stochastic Flexible Flow Shop Scheduling Problem with Preventive Maintenance and Budget Constraint Sadigh Raissi a,* , Ramtin Rooeinfar b ,Vahid Reza Ghezavati b a Islamic Azad University, South Tehran Branch, Tehran, Iran b School of Industrial Engineering, South Tehran Branch, Islamic Azad University, Tehran, Iran Received 20 2017; Revised 02 2018; Accepted 08 2018 Abstract Stochastic flexible flow shop scheduling problem (SFFSSP) is one the main focus of researchers due to the complexity arises from inherent uncertainties and also the difficulty of solving such NP-hard problems. Conventionally, in such problems each machine’s job process time may encounter uncertainty due to their relevant random behaviour. In order to examine such problems more realistically, fixed interval preventive maintenance (PM) and budget constraint are considered.PM activity is a crucial task to reduce the production efficiency. In the current research we focused on a scheduling problem which a job is processed at the upstream stage and all the downstream machines get busy or alternatively PM cost is significant, consequently the job waits inside the buffers and increases the associated holding cost. This paper proposes a new more realistic mathematical model which considers both the PM and holding cost of jobs inside the buffers in the stochastic flexible flow shop scheduling problem. The holding cost is controlled in the model via the budget constraint. In order to solve the proposedmodel, three hybrid metaheuristic algorithms are introduced. They include a couple of well-known metaheuristic algorithms which have efficient quality solutions in the literature. The two algorithms of them constructed byincorporationof the particle swarm optimization algorithm (PSO) and parallel simulated annealing (PSA) methods under different random generation policies. The third one enriched based on genetic algorithm (GA) with PSA. To evaluate the performance of the proposed algorithms, different numerical examples are presented. Computational experiments revealed that the proposed algorithms embedboth desirable accuracy and CPU time. Among them, the PSO-PSAП outperforms than other algorithms in terms of makespan and CPU time especially for large size problems. Keywords: Stochastic flexible flow shop; Budget constraint, Preventive maintenance; Genetic algorithm; Simulated annealing; Particle swarm optimization. 1. Introduction The flexible flow shop scheduling problem (FFSSP) consists of a flow manufacturing line with one or more parallel machine on some processing stages (or workstation) in series. Multiple products (or jobs) are produced in each stage. The objective function of FFSSP is on minimizing the total completions of all jobs (makespan or Cmax)(Hoogeveen et al., 1996; Gupta, 1998). This kind of issue often takes into account an NP-hard problem. This means that at a reasonable computational time, an optimal answer cannot be obtained. (Brucker and Kramer, 1995)have proved that a two-stage FFSSP remains NP-hard even there is only one machine on the first stage and two machines on the second stage.Most researchers have focused on the system which their relevant process time deploy a given deterministic value. However, in real circumstances, processing time of jobs at each stage is the main part of makespan. Most research focused on the system with deterministicprocessing time. However, in a real system, theprocessing time is arandom variable due to random behaviour of tool wearing, operator skill, material variability and so on (Koulamas and Kyparisis, 2000). According to Choi and Wang (2012) makespan estimationmight become invalid under differentcircumstances. Another important factor that effects on makespan is interruption caused by machine breakdowns (Fahmy and Sharif, 2009). One of the most important ways to deal with these types of random interruptions is on following preventive and scheduled tasks. Hence, unavailability of machines, due to preventive maintenance, could be incorporated as a set of constraints in the mathematical models. Taking into account such disruptions due to preventive maintenance and stochastic processing time makes the problem more real, but more complicated and has been kept on the research gap at yet.The significant contributions of this paper are to propose a more realistic scheduling model for such production and delivering an efficient solution methods.The proposed model integrates FFSSP with preventive maintenance (PM) and budget constraint under stochastic processing time. Hence,consideringqueues between consecutive stages is a respectable alternative. When a job is processed at upstream stage and all the downstream machines get to busy state or comes under *Corresponding author Email address: raissi@azad.ac.ir , Summer & Autumn 2019,Vol.12, Issue 2 Journal of Optimization in Industrial Engineering 131-147 November September September Sadigh Raissi et al. / Three Hybrid Metaheuristic Algorithms… 132 any PM activity, therefore the job should wait in the bufferswhichcause holding cost.Because the total budget in handhas a given level this leads to interrupt the operations and cause increasing of the makespan. Consequently, in the proposed model the jobs are scheduled in such a way that there is no interruption in their operations. By considering all of these affecting factors, the optimized makespan will be acquired by the model. Obviously, such FFSSP also ruins an NP-hard model. Therefore, three hybrid populations based metaheuristic algorithms are introduced as solution methods. The flexible flow shop scheduling problem has been studied extensively in the literature. Koulamas and Kyparisis(2000) developed a heuristic method for a two and a three-stage FFSSP with random processing time based on the makespan objective function. They proved the effect of the proposed heuristic by finding the solutions which were better than the available lower bounds. A Tabu search (TS) algorithm together with a procedure for a constructing a complete schedule to solve the FFSSP with limited buffers has been given by Wardono and Fathi(2004) that minimizes job completion time. Their algorithm acts based on the stage-oriented decomposition approach. Akrami et al. (2006) presented a mixed-integer linear mathematical programming for FFSSP. They assumed that there are limited buffers among stages. Two meta-heuristics, based on genetic algorithm (GA) and TS algorithm, were presented to optimality solve the model. A flow shop scheduling was investigated with limited intermediate buffers was studied by Wang and Tang (2009). They focused on the makespan minimizing as a performance criterion and they applied a TS algorithm to optimize the problem. In order to improve the diversity of the TS, a scatter search mechanism was applied. The computational results showed the high efficiency of their proposed hybrid metaheuristic. Al-Hinai and ElMekkawy(2011) proposed a flexible flow shop problem with random machine failures and a two stage hybrid GA was applied. The first stage was on minimizing the primary objective; the makespan, and the second stage was optimized the bi- objective function and integrates machines assignments and operations sequencing with the expected machine breakdown. Tran and Ng (2011)presented a water-flow algorithm to solve the FFSSP with intermediate buffers. They pooled the amount of precipitation and falling force to form flexible erosion capability. This work helped the erosion process of the algorithm to focus on exploiting promising regions strongly. They also utilized an improved procedure for constructing a complete schedule from a permutation that represents the sequence of jobs at the first stage of the scheduling problem. Kianfar et al. (2012) investigated a FFSSP with non-deterministic arrival of jobs and sequence dependent setup times. They used average tardiness of jobs as the objective function. To optimize the problem, they presented a novel dispatching rule and hybrid GA. A computer simulation model was also developed to evaluate the presented dispatching rule. The results showed that their proposed dispatching rule can lead to much better results in comparison with the traditional dispatching rules. Singh and Mahapatra (2012) proposed a novel PSO to solve FFSSP. An efficient mutation operator was embedded in PSO to prevent solutions from falling into the local optimums. The performance of the PSO was evaluated against GA by a set of test problems taken from the literature. According to the obtained results, the percentage deviation of the proposed PSO of the lower bound is equal to 2.961 and the same measure for GA is equal to3.559.In order to deal with uncertain job processing times in FFSSP’s, Choi and Wang (2012) presented a decomposition-based approach. The method combines two reactive approaches with a reactive- proactive approach. Lin and Ying (2013) proposed a hybrid algorithm based on the features of artificial immune systems and the annealing process of simulated annealing algorithms (SA) to optimize the FFSSP with limited buffer storage between stages. Almeder and Hartl(2013) studied the FFSSP with limited buffer storage in the metal–working industry. The partitioned the problem as a two stage problem. The first stage contains a single machine and its buffer. The semi-finished parts are stored in this buffer until a machine of the second stage is available. The second stage contains two parallel non- identical machines. They applied a hybrid approach based on discrete-event simulation and variable neighbourhood search to optimize the problem. The hybrid approach was led to an improvement between 3% and 10% compared with the current production plan of the company. By considering unrelated parallel machines, sequence- dependent setup times, probable reworks and different ready times, Rabiee et al. (2014) investigated the no-wait two-stage FFSSP. They proposed a novel hybrid algorithm based on imperialist competitive algorithm (ICA), SA, variable neighbourhood search (VNS) and GA. The performance of the proposed hybrid algorithm was evaluated against ICA, SA, VNS, GA and ant colony optimization (ACO) and high efficiency of their proposed hybrid algorithm was revealed. Arnaout (2014) tackled a rescheduling problem for the flow shop problem associated with stochastic processing and setup time. They proposed a new repair rule which and compared it with the existing algorithms. The results obtained from the computational experiments showed that the proposed repair rule can perform better those available algorithms in literature. Rahmani and Heydari (2014) studied FFSSP under uncertain processing times and unexpected arrivals of new jobs. They proposed a new approach to find robust schedules in this situation. Their approach was a proactive–reactive method which uses a two-step procedure. In order to consider stochastic processing time, Wang and Choi (2014) introduced a novel decomposition- based the Holonic approach (DBHA) to solve a FFSSP with stochastic processing time. Their proposed method was based on genetic algorithm control (GAC) and the shorting processing time contract net procontrol (SPT- CNP). Also, K-means clustering is utilized to divide machines into different clusters according to their Journal of Optimization in Industrial Engineering Vol.12, Issue 2, Summer & Autumn 2019, 131- 147 133 stochastic environments. The gained results showed that DBHA had better quality solutions against GAC and SPT- CNPA simulation based optimization approach for stochastic hybrid FFSSP in a real-world semiconductor back-end assembly facility was presented by Lin and Chen (2015). In their approach, the optimization strategy, based on genetic algorithm, was used to find the optimal assignment of the production line and machine type at each stage. The simulation model was used to evaluate the performance of solutions. They applied a real case study to prove the necessity of using simulation optimization approaches for practical applications. A novel algorithm was developed by Li and Pan (2015) to solve the hybrid flow shop with limited buffers. They combine two metaheuristic algorithms of artificial bee colony and TS and used makespan as the performance criterion. Their proposed algorithm was evaluated against several algorithms reported in the literature and the experimental results showed the high effectively and efficient performance of the proposed algorithm. Sangsawang et al. (2015) proposed two metaheuristic algorithms to optimize a two-stage re-entrant FFSSP. The first algorithm was a hybridization of GA and adaptive auto-tuning based on a fuzzy logic controller. The second one was a hybridization of particle swarm optimization (PSO) and Cauchy distribution. Experimental results revealed that both hybrid algorithms could present better solutions and are more powerful than the classical metaheuristics. The statistical analyses indicated that the hybrid PSO and hybrid GA can improve the best solutions in the literature by averages of 15.60% and 15.51%, respectively. Zabihzadeh and Rezaeian(2015) developed a mixed integer linear programming model for FFSSP. They assumed that there are some robots between stages for unloading, transferring and loading parts. Their proposed model has ability of determining the number of robots and jobs sequence. They applied two meta-heuristic algorithms; GA and ACO, to solve their model. Computational results showed that the GA is more efficient than ACO to optimize the model. Tang et al. (2016) presented a new model for dynamic FFSSP’s. By taking emergency maintenance, the proposed model minimized both objectives of energy consumption and makespan. Since they developed model was NP-hard, a multi-objective PSO algorithm to optimize the model. For finite capacity material requirement planning system in a flexible flow shop, Sukkerd and Wuttipornpun(2016) presented a hybrid GA and a Tabu search algorithm (HGATS). The results showed that the HGATS can outperform on comparing with the existing algorithms. Correspondingly, computational time of HGATS is acceptable when applied to real industrial systems. Rahmani and Ramezanian(2016) addressed a stochastic FFSSP in which new jobs arrive into the process as disruptions. They developed a mathematical model to minimize total weighted tardiness. A variable neighbourhood search algorithm was used to solve the model. The efficiency of the proposed algorithm evaluated through a set of test problems. González-Neira et al. (2016) presented a multi-criteria FFSSP in which criterion is quantitative and the other is qualitative. They assumed that job processing time deploys by a stochastic manner.The integral analysis method (IAM) was implemented to solve the relevant problem. In the IAM method, the problem first was introduced, then, cardinal analysis, ordinal analysis and integration analysis are done. Results showed that IAM method is able to select the alternatives with high efficiency in terms of both types of criteria. The most important criticism of scheduling problems is the gap between academic and practical problems. Even though the developed models have tried to be more realistic, but they fail to find the exact makespan in real life problems. This comes from the fact that the model cannot consider all the factors affected on makespan. As a result, there is a gap between obtained makespan by mathematical models and the makespan occurs in reality. In order to bridge this gap, this research presents an integrated mathematical model which is capable to consider all of the influential factors on makespan. The proposed model can simultaneously consider preventive maintenance, stochastic processing time and budget constraint. By integrating all of these subjects the model will reflect the real performance of a FFSSP. The rest of the paper is organized as follows: In section 2 the problem definition, mathematical model formulation and assumption of the model is presented. In section 3, the proposed hybrid metaheuristic algorithms in which three hybrid metaheuristic algorithms, including combining parallel SA with two types of PSO algorithms (PSO-PSAІ and PSO-PSAП), and a GA with parallel SA (GA-PSA) are specially explained. In section 4, computational results are presentedwhich actually compare the results of proposed metaheuristic algorithms. First, the small scale problem size is solved with CPLEX software. Then, the metaheuristic algorithms are used to find the near optimal solutions for the large scale of the problems. Analysis of the results and comparisons demonstrate the performance of the proposed solving methodologies on different problem sizes. Finally, conclusions and future researches are presented in the last section. 2. Mathematical Model Formulation In this section the problem under consideration is described. Consider a problem of J jobs and S working stages as shown in Figure 1. Each stage s has a number of 𝑁𝑠 identical parallel machines that operate in parallel. All jobs visit all stages from the first to the last stage. They are processed by one machine at each stage. Each job processed all the stages and every machine process maximum one job at a time. Thereis a buffer between two consecutive stages. The machines are under PM tasks. . Consequently, each machine could be available or unavailable at each time. All jobs are independent of each other and they are available at time zero. The processing time of jobs is considered to be stochastic. In order to Sadigh Raissi et al. / Three Hybrid Metaheuristic Algorithms… 134 model such randomly distributed processing time, the stochastic programing technique is applied in which each possible processing time is called as a scenario. The probability of occurrences of each scenario is shown by Pr(π). Therefore, the averaged processing time of each job at each stage is calculated according to its processing times in different scenarios and the probability of occurrences of each scenario. The objective is to set a sequence of jobs, allocating jobs to machines at each stage and determining the time between two consecutive maintenance activities in such a way that total completion time being minimized. Figure 2 shows an example of the Gantt chart of proposed stochastic flexible flow shop scheduling problem (SFFSSP) with 3 stages and 5 jobs. There is2, 3 and 2 machines in stages 1, 2 and 3, respectively. In machine 1 in stage 1, maintenance occurs after processing job 3. The other maintenance occurs in stages 2 and 3. The job 2 is processed in stage 1 from time 2 to 5 and is started in stage 2 from time 7. Therefore, job 2 is sent to the buffer of stage 1 and cause holding cost for each period of time, until machine 1 in stage 2 gets empty. Sometimes more than one job is placed in the buffers. For example, in the buffer of stage 2, from time 11 and 12, jobs 1 and 3 are placed in and increasing the holding cost for staying of each period of time in the buffer. Other constraints as well assumptions are listed as follows:  There is a set of jobs denoted J (j=1, 2,…, J) jobs which are available at time zero and no job may be cancelled before completion.  The set of L consecutive stages denoted by (s=1, 2,…, L),  Each stage is equipped with non-identical parallel machines denoted by (m=1, 2,…, 𝑁𝑠 ).  The set of R, denoted by (r=1,2,…, R) indicated the number of intermediate buffers.  All jobs have to process serially through all machines at each stage for only once time.  The interruption processing of each job on all machines at each stage is not acceptable. On the other hand, once a job is started, it must be processed to completion without any interruption either on or between machines.  All machines are continuously available at time zero, in another word; non-machines are failing at the starting time.  Each machine could process only one job at the same time, and each job has to visit each machine exactly once.  The preventive maintenance activities are performed on each machine at the fixed intervals (𝑃𝑚𝑠).  Once the preventive maintenance activity is carried out, there is no probability of a subsequent machine breakdown.  All jobsat each intermediate buffer havea same holding cost, but different at each stage. Fig.1.The framework of the flexible flow shop problem of this study 2.1. Notations To present the model using mathematical terms, consider the following notations. Indices s Indexofstages{s= 1,2,…, L} m Index of machines at s {m= 1, 2, …,𝑁𝑠} j Index of jobs{d, j= 1, 2,…, J} r Index of intermediate buffers{r= 1,2,…, R} h Index of job sequence{h, u=1 ,2,…,𝐾𝑚} n Index of maintenance activity which is done on machine j {n= 1,2,…,𝑉𝑚𝑠} 𝜋Index of probabilistic scenarios{𝜋 = 1, 2,…,𝛱} Parameters 𝑝𝑗𝑠(𝜋)The processing time of job j at stage s in scenario π 𝑃𝑟⁡(𝜋)The probability of occurrences scenario π ℎ𝑟𝑠𝑗 Holding cost of job jinintermediate buffer ofr at stage s 𝜇𝑚𝑠The repair rate of machine m at stage s 𝜆𝑚𝑠The failure rate of machine m at stage s 𝐷𝑚𝑠The duration time of maintenance activity of machinematstages �̅�𝑚𝑠The mean number of jobs that are performed on machine m at stage s 𝛽The minimum of availability of the system M An arbitrary large position number B Total budget Dependent decision variables 𝑆𝑗𝑠Starting time for the processing of job j at stage s 𝐶𝑗𝑠Completion time of job j at stage s 𝑃𝑚𝑠The time between two consecutive maintenance activities on machine m at stage s 𝑚𝑚𝑠The number of maintenance activities on machine m at stage s 𝑑𝑟𝑠𝑗 Waiting time of job j atintermediate bufferr at stage s 𝑍𝑗𝑟(𝑡) {1 ⁡if⁡job⁡𝑗⁡is⁡in⁡intermediate⁡buffer⁡𝑟⁡at⁡time⁡𝑡0 otherwise 𝑇𝑚𝑠The completion time of the last PM action on machine m at stage s 𝐴𝑚𝑠(𝑡)The availability of machine m at stage s at timet 𝐴𝑠(𝑡)The unavailability of stage s at time t 𝐴𝑠𝑦𝑠(𝑡)The unavailability of system at time t W 0 or 1 Journal of Optimization in Industrial Engineering Vol.12, Issue 2, Summer & Autumn 2019, 131- 147 135 Independent decision variables 𝑋𝑗ℎ𝑚𝑠 {1 ⁡if⁡job⁡𝑗⁡is⁡processed⁡in⁡sequence⁡ℎ⁡by⁡machine⁡𝑚⁡at⁡stage⁡𝑠0 ⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡otherwise 𝑌𝑛𝑚𝑠 {1 ⁡if⁡𝑛⁡th⁡maintenance⁡activity⁡on⁡machine⁡𝑚⁡at⁡stage⁡𝑠⁡is⁡executed0 otherwise 2.2 Mathematical model In our proposed model,availability of machine min stage s at time t under PM activities could be calculated according to Villemeur(1991) by Eq. 1. 𝐴𝑚𝑠(𝑡) = 𝜇𝑚𝑠𝜇𝑚𝑠+𝜆𝑚𝑠 + 𝜆𝑚𝑠𝜇𝑚𝑠+𝜆𝑚𝑠 𝑒𝑥𝑝[−(𝜇𝑚𝑠 + 𝜆𝑚𝑠)(𝑡 − 𝑇𝑚𝑠)]⁡⁡⁡⁡⁡⁡(1)re we considered system configuration as a parallel-series system in which machines ateach stage are parallel and stages are a series. A stage is said to be unavailable if all of its machines are unavailable. Therefore, the unavailability of stage s is calculated by Eq. 2 and unavailability of the total system by Eq. 3. 𝐴𝑠(𝑡) = ∏(1 − 𝐴𝑚𝑠(𝑡))𝑁𝑠𝑚=1 ⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(2) 𝐴𝑠𝑦𝑠(𝑡) = 1 − ∏(1 − 𝐴𝑠(𝑡))⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(3)𝐿𝑠=1 Therefore, the proposed model as an extension to basic model was taken fromGonzález-Neira et al. (2016) will be as follows. 𝑀𝑖𝑛⁡⁡{𝑚𝑎𝑥⁡(𝐶𝑗𝑠)}⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(4) St: ∑ ∑ 𝑋𝑗ℎ𝑚𝑠𝐾𝑚ℎ=1 𝑁𝑠 𝑚=1 = 1⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡∀𝑗 ∈ 𝐽;𝑠 ∈ 𝐿⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(5) ∑ ∑𝑋𝑗ℎ𝑚𝑠𝐽𝑗=1 𝑁𝑠 𝑚=1 = 1⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡∀𝑠 ∈ 𝐿;ℎ ∈ 𝐾𝑚⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(6) 𝐶𝑗𝑠 ≥ 𝑆𝑗𝑠 + ∑𝑃𝑟⁡(𝜋)𝑝𝑗𝑠(𝜋)𝛱𝜋 ⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(7) 𝐶𝑗𝑠−1 − 𝑀 (1 − 𝑍𝑗𝑟(𝑡)) ≤ 𝑡 ≤ 𝑆𝑗𝑠 + 𝑀 (1 − 𝑍𝑗𝑟(𝑡)) ∀𝑗 ∈ 𝐽;𝑠 ∈ 𝐿;∀𝑟 ∈ 𝑅|𝑠 = 𝑟⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(8) 𝑆𝑗𝑠 + (1 − 𝑋𝑗ℎ𝑚𝑠)𝑀 ≥ 𝐶𝑑𝑠 − (1 − 𝑋𝑑𝑢𝑚𝑠)𝑀⁡⁡⁡⁡ ∀𝑗,𝑑 ∈ 𝐽⎹⁡𝑗 ≠ 𝑑; ⁡ℎ,𝑢 ∈ 𝐾𝑚⎹⁡𝑢 < ℎ,𝑚 ∈ 𝑁𝑠;𝑠 ∈ 𝐿⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(9) ∑∑ 𝑑𝑟𝑠𝑗 ℎ𝑟𝑠𝑗𝑅𝑟=1 𝐽 𝑗=1 ≤ 𝐵⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(10) 𝑑𝑟𝑠𝑗 ≥ 𝑆𝑗𝑠+1 − 𝐶𝑗𝑠⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(11) (𝑛𝑃𝑚𝑠𝑌𝑛𝑚𝑠 − 𝐶𝑗𝑠)𝑋𝑗ℎ𝑚𝑠 ≥ −𝑀⁡(1 − 𝑊)⁡ ∀𝑗 ∈ 𝐽;∀𝑚 ∈ 𝑁𝑠;∀𝑠 ∈ 𝐿; ℎ ∈ 𝐾𝑚;0 ≤ 𝑛 ≤ 𝑉𝑚𝑠⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(12) (𝐶𝑗𝑠−𝑝𝑗𝑠(𝜋) − 𝑛𝑃𝑚𝑠𝑌𝑛𝑚𝑠 − 𝐷𝑚𝑠)⁡𝑋𝑗ℎ𝑚𝑠 ≥ −𝑀(𝑊)⁡⁡⁡⁡⁡⁡⁡ ⁡⁡⁡⁡0 ≤ 𝑛 ≤ 𝑉𝑚𝑠⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(13) 𝑃𝑚𝑠 = ∑ ∑ 𝑋𝑗ℎ𝑚𝑠 ∑ Pr(𝜋) 𝑝𝑗𝑠(𝜋)𝛱𝜋𝐾𝑚ℎ=1𝐽𝑗=1 𝑚𝑚𝑠 ∀𝑚 ∈ 𝑁𝑠;∀𝑠 ∈ 𝐿⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(14) 𝑇𝑚𝑠 ≤ 𝑛𝑃𝑚𝑠⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(15) 1 − 𝐴𝑠𝑦𝑠(𝑡) ≥ 𝛽⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(16) 𝐶𝑗𝑠,𝑑𝑟𝑠𝑗 ≥ 0⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡∀𝑗 ∈ 𝐽; ∀𝑠 ∈ 𝐿⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(17)⁡ 𝑋𝑗ℎ𝑚𝑠 ∈ {0,1}∀𝑗 ∈ 𝐽;∀𝑚 ∈ 𝑁𝑠;∀𝑠 ∈ 𝐿; ℎ ∈ 𝐾𝑚⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(18) 𝑌𝑛𝑚𝑠 ∈ {0,1}∀𝑚 ∈ 𝑁𝑠;∀𝑠 ∈ 𝐿;𝑛 ∈ 𝑉𝑚𝑠⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(19) Eq. (4) indicates the objective function. The constraint (5) ensures that the assignment of each job to one and only one machine at each stage. Constraints set (6) determines that the position of a machine sequence is takes by only one job at each stage. Constraints set (7) states that it is not allowed starting processing jobs at the next stage unless they have completed processing at the previous stage. Constraint sets (8) and (9) indicate that no interference should be taken among jobs on a common machine at any stage if the machine is available. On the other words, the difference between the processing times of any two jobs assigned to the same machine should not have any overlap.The constraint set (10) guarantees that the holding costs should be less than the available budget. Constraints set (11) determines the waiting time of jobs in each buffer.Constraints set (12) and (13) ensure that there are no overlap among operations and maintenance task. Constraint sets (14)-(16) are related to control the system availability.Finally, constraint set (17)-(19) controls the decision variable types. Sadigh Raissi et al. / Three Hybrid Metaheuristic Algorithms… 136 Fig. 2. An example of proposed SFFSSP Gantt chart of this study 3. Solution Methodologies As mentioned earlier, the proposed modelof SFFSSP is an NP-hard problem and solving this problem with exact methods in reasonable CPU time is impossible, so we should use heuristic or metaheuristic algorithms to solve it(Pinedo& Chao, 1999). Pinedo (2002) proved that some exact methods have been developed for solvingSFFSSP’s, but they are not suitable for more than 20 jobs and 10 machines. Also, heuristic algorithms are often may be trapped onsome local solutions, but metaheuristic algorithms can be described as a master strategy that guides and modifies subordinate heuristics to explore the solutions beyond the local optimally (Osam et al., 1996).We describe threehybridmetaheuristic algorithms, including combining parallel SA with two types of PSO algorithms (named PSO-PSAІ and PSO-PSAП), and a GA with parallel SA (named GA-PSA) which have good solution quality in the literature(Singh and Mahapatra, 2012; Kianfar et al., 2012; Rabiee et al., 2014; Tang et al., 2016; Sukkerd and Wuttipornpun, 2016). Theproposed metaheuristic algorithms are explained in the nextsubsections on details. 3.1. Hybrid PSO-PSA algorithms PSO algorithm is an evolutional solution method performed on a population of candidate solutions called particles. These particles move around in the search-space according to simple routine. The particles movements are guided by the best found positions in the search-space, which are continually updated as better positions are found by the particles. At each iteration, the in position of a particle (X vector) is updated by calculating the velocity (Vel vector) using the differences between the current position of the particle and the two following vectors (Kennedy and Eberhart, 1995):  The best position experienced by the particle in all previous iterations. This is called the particle best (Pbest).  The best position experienced by all particles in population in all previous iterations. This is called the global best (gbest). Generally, a PSO is a continuous algorithm inherently, while SA is a discrete one. Experiments show that combining PSO with a discrete algorithm such as SA creates better performances. Poli, Kennedy and Blackwell (2007) presented a review on the variation and the hybridization of the PSO. We proposed two types of hybrid PSO-PSA algorithms named PSO-PSAІ and PSO- PSAП to have both advantages of these methods. The basic idea of the hybrid PSO-PSA algorithm is running the PSO algorithm and improving the best results by applying parallel SA (PSA). In order to have variety in the proposed algorithm, we consider every bad, normal and good solutions could be selected with same chances.Combining PSO and PSA decreases the probability to be trapped in the local optimal solutions. Also, by introducing a suitable neighbourhood formation structure, the search process is enhanced and finds the near global optimum solution. The essential components of our proposed PSO-PSAІ and PSO-PSAП are similar, except in generating initial solutions as described below. The pseudo-code for our PSO-PSA algorithm is shown Figure 6. 3.1.1. Initial solution of PSO-PSAІ The random generation policy is used to generate the initial population of PSO-PSAІ. 3.1.2. Initial solution of PSO-PSAП In PSO-PSAП a new procedure is applied to generate initial solutions. First, we consider SFFSSP with relaxed conditions in which the binary constraints of model are Journal of Optimization in Industrial Engineering Vol.12, Issue 2, Summer & Autumn 2019, 131- 147 137 relaxed. In other word the 𝑋𝑗ℎ𝑚𝑠is considered to be a real number between 0 and 1. Next, the relaxed model was solved by CPLEX. The optimum values of decision variables are numbers between 0 and 1 which are looked as a probability distribution. These values are used as primal inputs of initial solutions for PSO-PSAП. For example, if is X2312 = 0.8 then we set X2312 = 1 with the 0.8 and X2312 = 0 with the 0.2. The initial solutions obtained by this methodare distributed in the high quality of solution space. In order to perform a comprehensive search the low quality of solution space must be investigated. Therefore, some initial solutions are generated, unlike the obtained probability distributions. For example, if X2312 = 0.8, then we set X2312 = 1 with the 0.2 and X2312 = 0 with the 0.8, the two groups if initial solutions are merged and formed initial solutions of PSO- PSAП. The input parameters of both proposed PSO-PSAІ and PSO-PSAП are: the population size (Npop), the number of successive iterations in which the best solution does not change (M), the cognition learning factor (C1), The social factor (C2) the inertial weight (w), the population size (Npop), the number of internal loop (In loop), and the temperature decreasing rate (α). The other components of proposed PSO-PSAІ and PSO-PSAП are the same and described as follows: 3.1.3. The solution structure of PSO-PSA The solution is represented by two sub-matrixes, each of them are associatedto a special area of decision variable. The first sub-matrix presents the sequence of jobs contains S*J matrix where S is the number of stations and J is the number of jobs. An enhanced version of random key representation, proposed by Norman and Bean (1999) is used to show the sequence of jobs which is capable to preserve the solution feasibility. In this way, each job at each station is assigned a real number between(1,𝑁𝑠). The integer part is the number determines the machine number to which the job is assigned and the fractional part is used to sort the jobs assigned to that machine. For example, consider a problem with 4 jobs, 5 stations and 4 machines at each station. An example of a solution for this problem is presented in Figure 3. According to Figure 4, in station 2, the jobs 1 and 4 are both processed on machine 3, and the job 2 is processed on machine 1, and the job 3 is processed on machine 2. Also the order of jobs to be scheduled on machine 3 is job 1 followed by job 4.The second sub-matrix is related to the number of maintenance activity on machines. It is a S*M matrix where S is the number of stations and M is the number of machines at each station. Each cell of this matrix is filled with a random number between 1 and the maximum number of allowable maintenance activity. Figure 4 is an example of a problem with 5 stations and 4 machines at each station. As shown in Figure 4, three, one, three and four maintenance activities are performed on machines 1, 2, 3 and 4 in station 1, respectively. Therefore, the decision variables 𝑌111 , 𝑌211 and 𝑌311will be equal to one. Fig.3. An example of representation of solution (Sequence of jobs) Fig. 4. An example of representation of solution (Maintenance activity) 3.1.4. Particle movements For particle movements, the following formulas are used to update the velocity and position vectors of a particle: 𝑉𝑒𝑙𝑖(𝑘 + 1) = 𝑤 ∗ 𝑉𝑒𝑙𝑖(𝑘) + 𝐶1 ∗ 𝑟1 ∗ (𝑃𝑏𝑒𝑠𝑡𝑖 − 𝑋𝑖(𝑘)) +𝐶2 ∗ 𝑟2 ∗ (𝑔𝑏𝑒𝑠𝑡 − 𝑋𝑖(𝑘)) (19) 𝑋𝑖(𝑘 + 1) =⁡𝑋𝑖(𝑘) +⁡𝑉𝑒𝑙𝑖(𝑘 + 1) (20) In equation (19), Veli(k) is the velocity of particle i in the k th iteration and Xi(k) states the position of particle i in iteration k. Also, Pbesti is the vector for the best known position of particle i and gbest is the best position vector of all particles in the whole population. 𝑤 is called the inertia weight that determines the impact of the current velocity of a particle on its velocity at the next iteration. The parameters C1 and C2 are acceleration coefficients which have constant values to determine the impact of Pbest and gbestvalues in defining the velocity, respectively. r1 and r2 are two random numbers uniformly distributed in [0,1]. 3.1.5. Local search improvements One of the challenges of the algorithms is trapped in local optimal solutions. In the other words, it is possible that the near optimum solution which has found yet, is selected and the algorithm accepts this solution as a final optimum solution and stops. The hybrid algorithms are used to combine the base algorithm with different strategies to improve the algorithm performances. We utilized medium radius local search for the proposed hybrid PSO-PSA algorithms. In this method, the local searches are not used on each swarm, and we used the double change technique. If the improvements on the convergence of fitness function are achieved, we replace it to the previous swarm, but, if no improvements have been seen, we accept this by Bultzen probability. Sadigh Raissi et al. / Three Hybrid Metaheuristic Algorithms… 138 3.1.6. Initial temperature A suitable initial temperature is the one that results in an average increase of acceptance probability near to one. The value of initial temperature will clearly depend on the scaling of fitness and, hence, it should be problem- specific. Therefore, we first generate a large set of random solutions, then a standard division of them is calculated and is used to determine the initial temperature in the way that the acceptance probability of primary generations reach to 0.999. Consequently, the initial Tk is set to 1000based on some preliminary examinations. 3.1.7. Cooling Schedule The performance of this algorithm also depends on the cooling schedule, which is essentially the temperature updating function. In the proportional decrement scheme, temperatures at the k and k+1 steps of the outer loop, 𝑇𝑘and 𝑇𝑘+1, are related by: 𝑇𝑘+1 = 𝛼𝑇𝑘⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(21) Where 𝛼 is cooling rate and is obtained by some experiment. 3.1.8. Stopping criteria To limit the number of iterations of PSO-PSA algorithms, some convergence experiment was performed and the best criterion was applied as follows: PSO-PSA will be stopped when the best solution does not change after a pre-determined number of successive iterations (M). Also, the PSA is allowed to search in a temperature level, for In-loop iterations. The optimum value of M and In-loop is determined by Taguchi experiments. The pseudo-code of the proposed PSO-PSA is described in Figure 5. Procedure of hybrid PSO-PSA algorithms P: initial particle with random positions and velocities Untiltermination condition is met Do For each particle ido Update the velocity of particle i Update the position of particle i Evaluate particle i Update Pbest and gbest Endfor Start PSA with some of the best and the worst chromosomes Set repetition counter k = 0 Untiltermination condition is met Do Set repetition counter M = 0 UntilM= in_loopDo Generate a neighbour solution: ω2 Calculate Δω1, ω2= f(ω) - f(ω0) If Δω1, ω2 ≤ 0, then ω1 = ω2 If Δω1, ω2 >0, then ω1 = ω2 with probability exp (-Δω1, ω 2 / tk) m = m + 1 End Tk+1= α Tk k = k + 1 End Transfer the improved solution to the particles End Fig. 5. Pseudo-code of hybrid PSO-PSA 3.2. Hybrid GA-PSA algorithm (GA-PSA) Genetic algorithm (GA) has no ability to search effectively to find the best global optimum solution. Also, this algorithm isn’t a capable to complete local searches on solutions. Therefore, we can combine the power of GA in global search with simulated annealing (SA) local searches to address the global optimum solution. This hybrid algorithm which combines GA and SA has both advantages of these algorithms and help to improve solution performances.In this hybrid algorithm, first the GA generates initial solutions with crossover and mutation operators. Next, some of these solutions have been selected for the parallel SA (PSA) as initial solutions. Then, the parallel local search process on the selected solution starts. To have variety in the proposed algorithm, we consider every bad, normal and good solutions could be selected with same chances. The PSA procedure in the same applied in PSO-PSA. The other components of PSO-SA are as follows: 3.2.1. Initialization The input of GA-PSA is the population size (Npop), the number of successive iterations in which the best solution does not change (M), the crossover probability (Pc), and the similarity coefficient (SC), and the mutation probability (Pm) the number of internal loops (In loop), and the temperature decreasing rate (α) are first initialized. Then, to generate an initial population, a random generation policy is utilized in this step. Since the solutions obtained by a metaheuristic algorithm are sensitive to their parameter values, a statistical procedure based on the Taguchi parameter tuning method is used to tune the parameters. 3.2.2. Selection operator One of the most key elements of a GA is the selection operator which is used to select chromosomes (parents) which lead to generate new chromosomes (offspring). The proposed selection operator is roulette wheel selection method in which parent chromosomes are probabilistically selected based on their fitness function value. The better chromosomes are selected with the highest probability. Using the roulette wheel selection each chromosome in the population occupies a slot with a slot size proportional to the chromosome fitness. When the wheel is randomly spun, the chromosome corresponding to the slot where the wheel stopped is selected as the first parent. This process is repeated to find the second parent. Clearly, since better chromosomes have larger slots, they have better chances to be chosen in the selection process. Journal of Optimization in Industrial Engineering Vol.12, Issue 2, Summer & Autumn 2019, 131- 147 139 3.2.3. The chromosome structure The chromosome structure of the GA-PSA is just like that used in “solution structure” in PSO-PSA. 3.2.4. Chromosome evaluation In order to evaluate chromosomes, each chromosome is simulated for 30 times. Then, the obtained results are averaged and considered as the chromosome fitness function. In the other word, by changing the stochastic input parameters, the fitness function of the chromosome will change. Therefore, the fitness function of each chromosome is not under deterministic value and show the stochastic nature of the model. 3.2.5. The crossover operator As defined in the previous section, the chromosomes have two parts, so a crossover operator is applied in these two parts. For each part of the chromosome a single-point crossover is applied. For the first part of the chromosome, a cross point between (1, J) is generated (J is the number of jobs). Then, the crossover operator is applied according to Figure 8. The same procedure is applied to generate the second of the offspring chromosomes. 3.2.6. The mutation operator The mutation operator is used in only some iterations. The similarity coefficient (SC) determines if mutation is applied or not. We can calculate the SC as follows: 𝑆𝐶𝑎𝑏 = ∑ ∑ ∑ ∑ 𝜕(𝑋𝑗ℎ𝑚𝑠(𝑎), 𝑋𝑗ℎ𝑚𝑠(𝑏))𝐿𝑠=1𝑁𝑠𝑚=1𝐾𝑚ℎ=1𝐽𝑗=1 𝑆 × 𝐽 (22) Where 𝑋𝑗ℎ𝑚𝑠(a)⁡and⁡𝑋𝑗ℎ𝑚𝑠(𝑏) are decision variables in chromosomes a and b. For comparing the similarity between two chromosomes, we consider the similarity of each gene that can be expressed as follows: 𝜕(𝑋𝑗ℎ𝑚𝑠(𝑎), 𝑋𝑗ℎ𝑚𝑠(𝑏)) = {1⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡𝑖𝑓⁡𝑋𝑗ℎ𝑚𝑠(𝑎) = 𝑋𝑗ℎ𝑚𝑠(𝑏)⁡⁡0⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒⁡ (23) The average similarity coefficient of the population is calculated as follows: 𝑆𝐶̅̅̅̅ = ∑ ∑ 𝑆𝐶𝑎𝑏𝑁𝑏=𝑎+1𝑁−1𝑎=1 (𝑁2) (24) WhereN is the number of chromosomes in the population. Considering a pre-defined threshold similarity coefficient, the specified mutation operator will be automatically applied. Two swapping types are used in proposed GA- PSA. Swapping type 1 is used to define the neighbourhood N(s) in local search (PSA) and swapping type 2 is used as mutation operator of GA.  Swapping type 1 The mutation operator is applied on both two parts of chromosomes. To this aim, a column of each chromosome sub matrix is randomly selected and is inversely arranged. Figure 6 shows an example of the mutation operator. Fig.6. An example of mutation operator of swapping type 1 (Sequence of jobs)  Swapping type 2 The swapping type 2 works which select two rowsof each chromosome sub matrix is randomly selected and swapped (see Figure7). Fig. 7. An example of mutation operator of swapping type 2 (Sequence of jobs) Sadigh Raissi et al. / Three Hybrid Metaheuristic Algorithms… 140 Fig. 8. An example of crossover operator (Sequence of jobs) 3.2.7. Stopping criteria The stopping criteria in the GA-PSA is similar to the ones described for PSO-PSA. Figure 9 shows the pseudo-code of the proposed GA-PSA algorithm. Procedure of hybrid GA-PSA algorithm P: generate the initial population (Npop) Untiltermination condition is metDo Evaluate the chromosomes Apply the crossover operator Apply the mutation operator Start PSA with some of the best and the worst chromosomes For all the selected chromosomes Do Set repetition counter k = 0 Untiltermination condition is met Do Set repetition counter M= 0 UntilM= in_loopDo Generate a neighbour solution: ω2 Calculate Δω1, ω2= f(ω) - f(ω0) If Δω1, ω2 ≤ 0, then ω1 = ω2 If Δω1, ω2 >0, then ω1 = ω2 with probability exp (-Δω1, ω 2 / tk) m = m + 1 End Tk+1= α Tk k = k + 1 End End Transfer the improved solutions to the genetic population and combine all the generated populations End Fig. 9. Pseudo-code of GA-PSA 4. Computational Evaluation All experimental results have been carried out on an ASUS laptop with a 2.4 GHz, core i5 processor using a 4GB of RAM. All metaheuristic algorithms have been implemented in MATLAB Software (Version 7.10.0.499, R2010a) and the linear programming models have been solved using CPLEX 12. Also, the Minitab 17 software has been used to apply theTaghuchi design of experiment method for parameters tuning.Two sets of test problems are applied to solve the model. These test problems are generated based on the data given in Table 1. Here, we considered 5 scenarios, each of which has equal probability to be occurring.Three machines workat each station, considering the number of jobs and the numbers of stations, the two test sets aregeneratedin small and large problem sizes. First, 18small sized test problems are generated and the solutions, obtained by the algorithms, are evaluated against global optimum amounts.Furthermore, 60 large sizedtest problems are applied to compare the performance of the algorithms. We considered two kinds of total budget called hereinafter by type 1 and 2. Both budget types are associated to the number of jobs. The budget types 1 and 2 are respectively calculating by multiplying of number of jobs in 1200 and 1000. Therefore, the budget type 2 is more strict than type 1. Table 1 Factor and their levels Factors Levels Number of jobs (j) [4, 5, 6, 20, 50, 60, 100] Number of stations (s) [2, 3, 4, 6, 9] Holding cost ($/sec) Uniform [50, 100] Process times (sec) Uniform [1, 99] The minimum of availability 0.95 The repair rate of machine (sec) Uniform [10, 50] The failure rate of machine (sec) Uniform [150, 450] Number of scenarios 5 4.1. Parameter tuning As mentioned before, the initial parameters of hybrid PSO-PSA algorithms (including both PSO-PSAІ and PSO-PSAП) are the population size (Npop), the number of successive iterations in which the best solution does not change (M), the cognition learning factor (C1), The social factor (C2) the inertial weight (w), the population size (Npop), the number of internal loop (In loop), and the temperature decreasing rate (α). Also, the population size (Npop), the number of successive iterations in which the best solution does not change (M), the crossover probability (Pc), and the similarity coefficient (SC), and the mutation probability (Pm) the number of internal loops (In loop), and the temperature decreasing rate (α) were used in GA-PSA.To investigate the influence of those parameters on the performance of the algorithms, we implement the Taguchi’s method in the design of experiments (DOE) (Montgomery, 2005).In the Taguchi’s method, the results are converted into an estimator called single to noise ratio (S/N). The S/N ratio shows both the mean and the variation in the response variables. To minimize the objective function the S/N ratio is calculated as the following formula: 𝑆 𝑁⁄ = −10log ⁡(1 𝑛⁄ ∑ 1𝑦𝑖2𝑛𝑖=1 )⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(23) Which, n and yiindicate number of replications and process response value at i’th replication. We chose the Journal of Optimization in Industrial Engineering Vol.12, Issue 2, Summer & Autumn 2019, 133- 149 141 orthogonal array L27 for both PSO-PSA and GA-PSA. In Taguchi’s design of experiments, the relative percentage deviations (RPD’s) transform into S/N ratio, we figure out the response value of each parameter in the proposed algorithms.After testing the different parameter levels and analysing the obtained results, the better initial parameter levels were gained. The initial and optimumparameter levels of each proposed hybrid algorithm are shown in Tables 2.According to the parameter values in Table 2, we illustrate the trend of each factor level of PSO-PSA and GA-PSA are in Figures 10 and 11, respectively. Also, the results of each test problem solved by the proposed algorithms are shown in Figure 12. The statistical results of the objective function (makespan)and CPU time for the small and large sized problems are presented in Tables 5 and 6.Additionally, we utilized the relative percentage deviation (RPD) to test the performances of applied metaheuristics as below: 𝑅𝑃𝐷 = 𝐴𝑙𝑔𝑠𝑜𝑙 − 𝐿𝑠𝑜𝑙𝐿𝑠𝑜𝑙 ⁡⁡⁡⁡⁡⁡⁡⁡𝑖 = 1,…,𝑛⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(32) Where, Algsolis the objective function value for a given algorithm, Lsolis the best value of the objective function between algorithms and n is the number of small size or large size problems. Table 2 Initial and optimum parameter levels of hybrid algorithms Algorithms Parameters Factor levels Optimum amount 1 2 3 PSO-PSA Npop PSO-PSA 100 200 300 200 M 10 20 30 20 C1 1 1.5 2 1.5 C2 0.7 1 1.5 1 Ω 0.3 0.4 1 0.4 In loop 5 10 15 10 α 0.87 0.91 0.95 0.91 GA-PSA Npop GA-PSA 100 200 300 200 M 10 20 30 20 Pc 0.85 0.9 0.95 0.9 SC 0.6 0.7 0.8 0.7 Pm 0.005 0.01 0.015 0.01 In loop 5 10 15 10 α 0.87 0.91 0.95 0.91 Fig. 10. Factor level trend of PSO-PSAalgorithm Fig. 11. Factor level trend of GA-PSAalgorithm Sadigh Raissi et al. / Three Hybrid Metaheuristic Algorithms… 142 Table 5 Solutions found by the algorithms of the small-sized test problems CPLEX PSO-PSAП PSO-PSAІ GA-PSA Problem # Budget type NOJ (j) NOS (s) Makespan CPU time Gap (%) Makespan CPU time Makespan CPU time Makespan CPU time 1 1 2 4 2 205 210 22 22 0 0 205 210 22 23 205 229 26 37 205 233 29 39 2 1 2 209 217 35 36 0 0 209 218 37 35 216 225 52 47 219 228 55 49 3 1 2 3 214 240 95 91 0 0 214 241 97 94 215 248 109 117 227 253 115 122 4 1 2 218 223 155 150 0 0 218 224 156 152 227 237 160 182 249 250 163 186 5 1 2 4 232 236 185 184 0 0 235 238 191 187 243 246 305 287 247 249 321 328 6 1 2 249 250 235 234 0 0 249 252 238 236 257 264 398 382 260 262 415 402 7 1 2 5 2 237 241 645 647 0 0 238 245 651 650 248 252 1039 1077 251 255 1100 1138 8 1 2 253 262 822 830 0 0 256 268 831 833 265 273 1005 1015 270 276 1176 1165 9 1 2 3 270 280 882 890 0 0 274 282 890 893 280 292 996 1016 284 298 1034 1086 10 1 2 289 304 918 922 0 0 292 308 923 931 302 317 987 1083 305 321 1150 1154 11 1 2 4 335 348 1012 1048 0 0 337 350 1024 1055 346 355 1056 1088 350 369 1149 1047 12 1 2 326 351 943 950 0 0 329 353 951 958 336 363 1008 1032 339 368 1062 1094 13 1 2 6 2 349 362 1083 1116 0 0 352 366 1099 1128 361 374 1125 1176 367 379 1139 1182 14 1 2 351 366 1043 1055 0 0 355 371 1074 1094 363 381 1133 1156 369 385 1273 1258 15 1 2 3 350 372 1103 1144 24.54 28.55 354 376 1127 1185 367 392 1176 1192 369 395 1201 1282 16 1 2 347 360 1078 1123 29.17 31.47 351 363 1112 1175 358 371 1185 1211 363 374 1246 1267 17 1 2 4 361 367 1197 1255 68.35 66.92 364 371 1225 1292 373 379 1312 1332 377 383 1424 1487 18 1 2 358 394 1205 1270 74.18 78.53 363 398 1243 1293 370 384 1294 1345 376 399 1431 1481 Average 1 2 268.28 299.06 703.22 720.39 10.90 51.37 288.61 301.89 716.17 734.11 296.22 310.11 798.11 820.33 301.50 315.39 860.17 875.94 Gap (%) = CPLEX optimality gap Fig. 12. The performance of proposed hybrid algorithmsof the large-sized test problems in terms of average makespan 0 2000 4000 6000 8000 10000 12000 14000 16000 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 97 10 0 10 3 10 6 10 9 11 2 11 5 11 8 A v e r a g e m a k e sp a n Test problem PSO-PSAП PSO-PSAІ GA-PSA Journal of Optimization in Industrial Engineering Vol.12, Issue 2, Summer & Autumn 2019, 133- 149 143 Table 6 Solutions found by the algorithms of the large-sized test problems Problem # Budget type NOJ (j) NOS (s) PSO-PSAП PSO-PSAІ GA-PSA Makespan CPU time Makespan CPU time Makespan CPU time 1 1 2 20 3 958 1005 2231 2277 937 996 2265 2309 1009 1055 2319 2393 2 1 2 941 994 2226 2202 959 1002 2261 2278 992 1038 2318 2373 3 1 2 964 1032 2230 2282 988 1042 2278 2266 1031 1086 2302 2285 4 1 2 984 1034 2236 2224 981 1034 2279 2273 1035 1088 2333 2383 5 1 2 952 993 2227 2265 985 1071 2291 2281 993 1039 2329 2382 6 1 2 6 1127 1179 2385 2463 1138 1197 2450 2408 1161 1231 2483 2479 7 1 2 1152 1223 2397 2467 1131 1174 2438 2464 1201 1276 2508 2579 8 1 2 1080 1147 2395 241 1097 1142 2428 2426 1119 1165 2469 2486 9 1 2 1188 1252 2373 2434 1149 1223 2443 2406 1245 1297 2471 2532 10 1 2 1164 1218 2380 2379 1170 1228 2431 2426 1104 1159 2468 2464 11 1 2 9 1174 1210 2417 2490 1178 1234 2458 2524 1216 1278 2529 2531 12 1 2 1120 1190 2416 2490 1155 1225 2471 2450 1178 1249 2521 2480 13 1 2 1197 1260 2433 2476 1233 1304 2443 2441 1257 1307 2471 2457 14 1 2 1143 1215 2421 2437 1145 1213 2465 2495 1169 1233 2500 2549 15 1 2 1187 1250 2417 2412 1233 1293 2476 2485 1236 1309 2511 2521 16 1 2 50 3 2617 2706 2903 2990 2684 2793 3021 3030 2767 2912 3197 3137 17 1 2 2652 2758 2911 2945 2730 2881 3028 3058 2813 2936 3134 3216 18 1 2 2562 2693 2910 2956 2639 2753 3074 3124 2697 2799 3251 3339 19 1 2 2635 2740 2908 2952 2697 2813 3010 3013 2768 2916 3159 3130 20 1 2 2589 2676 2901 2936 2680 2789 3050 3042 2757 2858 3180 3270 21 1 2 6 2839 2992 3010 3097 2920 3035 3141 3085 3055 3218 3300 3378 22 1 2 2815 2922 3009 3085 2918 3036 3115 3075 2994 3106 3240 3233 23 1 2 2620 2725 3010 3103 2698 2801 3116 3146 2770 2873 3255 3319 24 1 2 2843 2996 3031 3019 2933 3035 3199 3286 3050 3169 3352 3444 25 1 2 2568 2687 3022 3073 2637 2727 3184 3260 2750 2860 3300 3269 26 1 2 9 2827 2943 3181 3178 2911 3024 3324 3427 3004 3153 3491 3571 27 1 2 2825 2989 3186 3225 2898 3064 3318 3399 3009 3187 3516 3460 28 1 2 3109 3296 3164 3137 3214 3373 3288 3273 3325 3478 3399 3465 29 1 2 2827 2948 3181 3203 2931 3043 3280 3369 3051 3210 3390 3338 30 1 2 3087 3245 3193 3292 3178 3322 3344 3380 3299 3489 3517 3476 31 1 2 60 3 4117 4388 3578 3690 4525 4793 3834 3926 4900 5726 4070 4084 32 1 2 3861 4059 3596 3616 4194 4466 3837 3763 4468 4823 4035 4125 33 1 2 4026 4327 3594 3470 4304 4684 3847 3786 4752 5026 4097 4179 34 1 2 3828 4043 3586 3670 4192 4452 3876 3968 4507 4910 4134 4243 35 1 4048 3589 4424 3839 4787 4112 Sadigh Raissi et al. / Three Hybrid Metaheuristic Algorithms… 144 2 4412 3694 4675 3840 5185 4207 36 1 2 6 4406 4650 3709 3692 4712 5107 3999 3922 5205 5487 4221 4180 37 1 2 4331 4711 3729 3764 4658 4993 4025 3975 5161 5487 4246 4230 38 1 2 4123 4479 3701 3788 4403 4675 4030 4128 4860 5256 4351 4483 39 1 2 4250 4513 3700 3809 4570 4916 3912 3988 5072 5392 4240 4342 40 1 2 4236 4552 3722 3745 4641 4982 3916 3943 5060 5322 4157 4256 41 1 2 9 4668 4927 3954 4002 5008 5350 4221 4189 5415 5760 4494 4638 42 1 2 4443 4749 4024 3961 4793 5051 4248 4177 5155 5592 4499 4553 43 1 2 4249 4607 3947 4001 4626 4958 4156 4232 4943 5218 4431 4511 44 1 2 4705 4857 3958 3937 5156 5283 4227 4232 5302 5673 4449 4477 45 1 2 4569 4985 3949 4037 4761 4912 4226 4168 5018 5325 4554 4537 46 1 2 100 3 9395 10037 4845 4805 10169 11081 5459 5445 11358 12081 6083 6024 47 1 2 9416 10112 4857 5003 10344 11193 5313 5335 11573 12201 6022 5913 48 1 2 9712 10299 4799 4866 10521 11197 5446 5561 11681 12717 5957 5860 49 1 2 9279 9762 4811 4962 10184 10869 5301 5307 10896 11522 5982 5906 50 1 2 9137 9880 4824 4756 9879 10651 5423 5410 11125 11700 6040 6227 51 1 2 6 9938 10733 4968 5022 11013 11748 5629 5732 11839 12457 6274 6258 52 1 2 9725 10589 4972 4932 10744 11570 5587 5544 11916 12867 6240 6367 53 1 2 9747 10378 4920 5059 10919 11581 5415 5340 11947 12926 6167 6248 54 1 2 10115 10917 4937 4948 11005 11661 5566 5489 12176 13142 6305 6327 55 1 2 9844 10468 5017 4972 10732 11580 5479 5498 11919 12673 5978 6091 56 1 2 9 10733 12093 5428 5366 11540 12239 6568 6677 12597 13491 7390 7630 57 1 2 11007 11613 5799 5817 12329 13058 6558 6643 13316 14096 7855 7953 58 1 2 10873 11434 5849 5997 12069 12826 6557 6438 13537 14717 7462 7577 59 1 2 11654 12466 5892 6020 13002 13705 6694 6782 14018 15217 7583 7467 60 1 2 10650 10987 5824 5742 11537 12587 6502 6697 12880 13879 7893 7834 Average 1 2 4547.18 4842.08 3579.70 3617.55 4900.02 5211.50 3850.98 3867.98 5307.42 5656.53 4158.90 4193.93 NOJ (j) =number of jobs NOS (s) =number of stations After solving all test problems, the results showed that in the all test problems, PSO-PSAПhas better performances in terms of makespan and CPU time. The comparisons of makespan show that the PSO-PSAП provides better solution quality with difference average RPD (ARPD) of 5.24, and 12.02 in comparison to PSO-PSAІ, and GA- PSA, respectively. Also, we can see that in the comparison of CPU times, PSO-PSAП give better results in terms of difference ARPD of 6.04, and 13.42 against PSO-PSAІ, and GA-PSA. Figures 13 and 14 show the mean plot of the CPU time and makespan of the proposed metaheuristic algorithms, respectively.According to computational results, it can be inferred that in small- sized test problems, all the algorithms approximately have the same computational effort. Therefore, it can be proved that all the metaheuristic algorithms are able to reach optimal/near optimal solutions.But, as the sizes of the problems are increased, the difference between algorithms is more revealed so that PSO-PSAП overcomes all other algorithms. However, the PSO-PSAІ, and GA-PSA are highly affected by the problem size so that by increasing it, the CPU time is exponentially increased. Figure 15 depicts the 95% confidence intervals of makespan and Figure 16 showsthe 95% confidence intervals for RPD among the proposed algorithms. Journal of Optimization in Industrial Engineering Vol.12, Issue 2, Summer & Autumn 2019, 133- 149 145 Fig. 13. Means plot of CPU time of hybrid algorithms Fig. 14. Means plot of makespan of hybrid algorithms 5. Conclusion In this study, we developed the stochastic flexible flow shop scheduling problem (SFFSSP) considering fixed interval preventive maintenance (PM) and budget constraint. This new type of problem which is the main contribution of the research is presented an integrated mathematical model which is capable to consider all of the influential factors on makespan. In theproposed SFFSSP, there is a buffer between each stage, if all machines are busy or under PM action, the job can wait in the buffer. The budget constraint controls the holding costs of total jobs in all buffers should not be greater than available budget. The proposed model considers not only the preventive maintenance, but also stochastic processing time and budget constraint. By integrating all of these subjects the model will reflect the real performance of a SFFSSP. Since the problem was strongly NP-hard, three hybrid algorithms, including PSA with two types of PSO algorithms (PSO-PSAІ and PSO-SAП), and a GA (GA- PSA) were proposed to solve the model, which have suitable quality solutions in the literature.To compare the computational results, we tested two types of test problems containing 18 small sized test problems and 60 large sized test problems. As the results showed, the PSO- PSAП algorithm provided better quality solutions in both makespan and CPU time among the test problems. Also, the higher performance of proposed PSO-PSAП algorithm with respect to other algorithms is more revealed in the large sized test problems. The presented model is still open to considering other options, such as sequence dependent setup time, machine random breakdown, and the problem of job availability at time zero. Also, it might be exciting in working on bi- objective SFFSSP’s, which the other objective function could be minimizes the maximum tardiness.Another research direction could be incorporating different transportation types to transport jobs between each stage. Additionaleffort can try to solve model by developing a new solution methodology such as a new hybrid algorithm or a new population-based algorithm can be investigated.We assumed that each job has a same holding cost at each intermediate buffer, but it is different at each stage. Another aspect deserving future efforts is to consider that the holding costs of each job are different at any intermediate buffer. Fig 15. The 95% confidence intervals of makespanof the small- sized test problems Fig 16. The 95% confidence intervals of makespanof the large- sized test problems References Akrami, B., Karimi, B., Hosseini, S.M.M., (2006). Two metaheuristic methods for the common cycle 0 5 10 15 20 25 30 20 50 60 100 R P D NOJ PSO-PSAП PSO-PSAІ GA-PSA 0 5 10 15 20 25 20 50 60 100 R P D NOJ PSO-PSAП PSO-PSAІ GA-PSA 0 50 100 150 200 250 300 350 CPLEX PSO-PSAП PSO-PSAІ GA-PSA R P D 0 2 4 6 8 10 12 14 16 PSO-PSAП PSO-PSAІ GA-PSA R P D Sadigh Raissi et al. / Three Hybrid Metaheuristic Algorithms… 146 economic lot sizing and scheduling in flexible flow shops with limited intermediate buffers: the finite horizon case.Applied Mathematics and Computation, 183, 634- 645. Al-Hinai, N., ElMekkawy, T.Y., (2011). Robust and stable flexible job shop scheduling with random machine breakdowns using a hybrid genetic algorithm.International Journal of Production Economics, 132, 279-291. Almeder, C., Hartl, R.F., (2013). A metaheuristic optimization approach for a real-world stochastic flexible flow shop problem with limited buffer.International Journal of Production Economics, 145, 88-95. Arnaout, J.P., (2014). Rescheduling of parallel machines with stochastic processing and setup times.Journal of Manufacturing Systems, 33 (3), 376-384. Brucker, P., Kramer, B., (1995). Shop scheduling problems with multiprocessor tasks on dedicated processors. Annals of Operations Research, 50, 13– 27. Choi, S.H., Wang, K., (2012). Flexible flow shop scheduling with stochastic processing times: A decomposition-based approach.Computers & Industrial Engineering, 63, 362-373. Fahmy, S.A., Sherif, A., Balakrishnan, S., ElMekkawy, T.Y., (2009). A generic deadlock-free reactive scheduling approach.International Journal of Production Research, 47 (20), 5657-5676. González-Neira, E.M., García-Cáceres, R.G., Cabellero- Villalobos, J.P., Molina-Sánchez, L.P., Montoya- Torres, J.R., (2016). Stochastic flexible flow shop scheduling problem under quantitative and qualitative decision criteria.Computer and Industrial Engineering, 101, 128-144. Gupta J.N.D., (1988). Two stage hybrid flow shop scheduling problem. Journal of Operational Research Society, 39(4), 359-64. Hoogeveen, J.A., Lenstra, J.K., Veltman, B., (1996). Minimizing the makespan in a multiprocessor flow shop is strongly NP-hard. European Journal of Operational Research, 89, 172-175. Kennedy, J.,Eberhart,R.C.,(1995). Particle swarm optimization.Proceeding of IEEE International Conference on Neural Network, Piscataway: IEEE 4, 1942-1948. Kianfar, K., FatemiGhomi, S.M.T., OroojlooyJadid, A., (2012). Study of stochastic sequence-dependent flexible flow shop via developing a dispatching rule and a hybrid GA.Engineering Applications of Artificial Intelligence, 25, 494-506. Koulamas, C., Kyparisis, G.J., (2000). Scheduling on uniform parallel machines to minimize maximum lateness.Operations Research Letters, 24(6),175-179. Li, J.Q., Pan, Q.K., (2015). Solving the large-scale hybrid flow shop scheduling problem with limited buffers by a hybrid artificial bee colony algorithm. Information Sciences, 316:487–502. Lin, SW., Ying, K.C., (2013). Minimizing makespan in a blocking flowshop using a revised artificial immune system algorithm.Omega, 2 (4), 383-389. Lin, J.T., Chen, C.M., (2015). Simulation optimization approach for hybrid flow shop scheduling problem in semiconductor back-end manufacturing.Simulation Modeling Practice and Theory, 51, 100-114. Montgomery, D.C. (2005). Design and analysis of experiments. Arizona, John Wiley and Sons. Norman, B.A., Bean, J.C., (1999). A genetic algorithm methodology for complex scheduling problems. Naval Research. Logistics, 46: 199-211. Osman, I.H., Kelly, J.P., (1996). Metaheuristics: an overview. Metaheuristics: theory and applications. Boston: Kluwer Academic Publisher, 1-21. Pinedo, M., Chao, X., (1999). Operations scheduling with applications inmanufacturing and services. New York: McGraw-Hill. Pinedo, M., (2002). Scheduling theory, algorithms, and systems. Englewood Cliffs, New Jersey: Prentice- Hall. Chapter 2. Poli, R., Kennedy, G., Blackwell, T., (2007). Particle swarm optimization: an overview. Swarm Intelligence, 1(3), 33-57. Rabiee, M., Sadeghi Rad, R., Mazinani, M., Shafaei, R., (2014). An intelligent hybrid metaheuristic for solving a case of no-wait two-stage flexible flow shop scheduling problem with unrelated parallel machine.International Journal of Advanced Manufacturing Technology, 71: 1229-1245. Rahmani, D., Heydari, M., (2014). Robust and stable flow shop scheduling with unexpected arrivals of new jobs and uncertain processing times. Original Research Article, 33 (1), 84-92. Rahmani, D., Ramezanian, R., (2016). A stable reactive approach in dynamic flexible flow shop scheduling with unexpected disruptions: A case study. Computers & Industrial Engineering, 98, 360-372. Sangsawang, C., Sethanan, K., Fujimoto, T., Gen, M., (2015). Metaheuristics optimization approaches for two-stage reentrant flexible flow shop with blocking constraint.Expert Systems with Applications, 42, 2395–2410. Singh, M.R., Mahapatra, S.S., (2012). A swarm optimization approach for flexible flow shop scheduling with multiprocessor tasks.International Journal of Advanced Manufacturing Technology, 62: 267-277. Sukkerd, W., Wuttipornpun, T., (2016). Hybrid genetic algorithm and tabu search for finite capacity material requirement planning system in flexible flow shop with assembly operations. Computers & Industrial Engineering, 97, 157- 169. Tang, D., Dai, M., Salido, M.A, Giret, A., (2016). Energy-efficient dynamic scheduling for a flexible flow shop using an improved particle swarm optimization. Computers in Industry, 81, 82–95. Journal of Optimization in Industrial Engineering Vol.12, Issue 2, Summer & Autumn 2019, 133- 149 147 Tran, T.H., Ng, K.M. (2011). A water flow algorithm for flexible flow shop scheduling with intermediate buffers.Journal of Scheduling, 14, 483-500. Villemeur, A., (1991). Reliability, availability, maintainability and safety assessment. USA: Wiley. [32] Wang, X., Tang, L., (2009). A Tabu search heuristic for the hybrid flowshop scheduling with finite intermediate buffers. Computers & Operations Research, 36 (3), 907–918. Wang, K., Choi, S.H., (2014). A holonic approach to flexible flow shop scheduling under stochastic processing times.Computers & Operations Research, 43, 157-168. Wardono, B., Fathi, Y., (2004). A Tabu search algorithm for multi-stage parallel machine problem with limited buffer capacities. European Journal of Operational Research, 155 (2), 380-401. Zabihzadeh, S.R., Rezaeian, J., (2015). Two metaheuristic algorithms for flexible flow shop scheduling problem with robotic transportation and release time. Applied Soft Computing, 40, 319-330. This article can be cited: Raissi, S., Rooeinfar, R. & Ghezavati, V. R. (2019) Three Hybrid Metaheuristic Algorithms for Stochastic Flexible Flow Shop Scheduling Problem with Preventive Maintenance and Budget Constraint Journal of Optimization in Industrial Engineering. 12 (2), 131-147. http://www.qjie.ir/article_543744.html DOI: 10.22094/JOIE.2018.242.1532 http://www.qjie.ir/article_543744.html 
work_cocjcgde7vgvpbfuwmkltp7vky	----	Topological constraints and robustness in Liquid State Machines Expert Systems with Applications 39 (2012) 1597–1606 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Topological constraints and robustness in liquid state machines Hananel Hazan ⇑, Larry M. Manevitz Department of Computer Science, University of Haifa, Mount Carmel, Haifa 31905, Israel a r t i c l e i n f o a b s t r a c t Keywords: Liquid State Machine Reservoir computing Small world topology Robustness Machine learning 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.06.052 ⇑ Corresponding author. Tel.: +972 4 8288337; fax: E-mail addresses: hhazan01@cs.haifa.ac.il (H. Ha (L.M. Manevitz). The Liquid State Machine (LSM) is a method of computing with temporal neurons, which can be used amongst other things for classifying intrinsically temporal data directly unlike standard artificial neural networks. It has also been put forward as a natural model of certain kinds of brain functions. There are two results in this paper: (1) We show that the Liquid State Machines as normally defined cannot serve as a natural model for brain function. This is because they are very vulnerable to failures in parts of the model. This result is in contrast to work by Maass et al. which showed that these models are robust to noise in the input data. (2) We show that specifying certain kinds of topological constraints (such as ‘‘small world assumption’’), which have been claimed are reasonably plausible biologically, can restore robustness in this sense to LSMs. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Processing in artificial neurons typically is a-temporal. This is because the underlying basic neuronal model, that of Pitts and McCulloch (1943) is a-temporal by nature. As a result, most appli- cations of artificial neural networks are related in one way or an- other to static pattern recognition. On the other hand, it has long been recognized in the brain science community that the McCul- lough–Pitts paradigm is inadequate. Various models of differing complexity have been promulgated to explain the temporal capa- bilities (amongst other things) of natural neurons and neuronal networks. However, during the last decade, computational scientists have begun to pay attention to this issue from the neurocomputation perspective as well, e.g. Fern and Sojakka (n.d.), Jaeger (2001a, 2001b, 2002), Lukosevicius and Jaeger (2009) and Maass, Natschläger, and Markram (2002a, 2002b, 2002d), and investiga- tions as to the computational capabilities of various models are being investigated. One such model, the Liquid State Machine (LSM) (see Fig. 1) (Maass et al., 2002a), has had substantial success recently. The Liquid State Machine is a somewhat different paradigm of compu- tation. It assumes that information is stored, not in ‘‘attractors’’ as is usually assumed in recurrent neural networks, but in the activity pattern of all the neurons which feed-back in a sufficiently recur- rent and inter-connected network. This information can then be recognized by any sufficiently strong classifier such as an Adaline ll rights reserved. +972 4 8288181. zan), manevitz@cs.haifa.ac.il (Widrow & Hoff, 1960), Back-Propagation, SVM1 or Tempotron (Gutig & Sompolinsky, 2006). (The name ‘‘liquid state’’ comes from the idea that the history of, e.g. timings of rocks thrown into a pond of water, is completely contained in the wave structure.) Moreover, the ‘‘persistence of the trace’’ (or as Maass put it, the ‘‘fading memory’’ (Lukosevicius & Jaeger, 2009)) allows one to recognize at a temporal distance the signal that was sent to the liquid; and sequence and timing effects of inputs. The Liquid State Machine is a recurrent neural network. In its usual format (Lukosevicius & Jaeger, 2009; Maass et al., 2002a), each neuron is a biologically inspired artificial neuron such as an ‘‘integrate and fire’’ (LIF) neuron or an ‘‘Izhikevich’’ style neuron (Izhikevich, 2003). The connections between neurons define the dynamical process, and the recurrence connections define what we call the ‘‘topology’’ in this paper. The properties of the artificial neurons, together with these recurrences, results in any sequence of history input being transformed into a spatio-temporal pattern activation of the liquid. The nomenclature comes from the fact that one can intuitively look at the network as if it was a ‘‘liquid’’ such as a pond of water, the stimuli are rocks thrown into the water, and the ripples on the pond are the spatio-temporal pattern. In the context of LSM the ‘‘detectors’’ are classifier systems that receive as input a state (or in large systems a sample of the ele- ments of the liquid) and are trained to recognize patterns that evolve from a given class of inputs. Thus a detector could be a SVM or an Adaline (Widrow & Hoff, 1960), perceptron (Pitts & McCulloch, 1943), or three level back propagation neural networks, etc. 1 SVM = support vector machine. http://dx.doi.org/10.1016/j.eswa.2011.06.052 mailto:hhazan01@cs.haifa.ac.il mailto:manevitz@cs.haifa.ac.il http://dx.doi.org/10.1016/j.eswa.2011.06.052 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. Liquid State Machine framework. 2 http://www.lsm.tugraz.at/csim/. 3 http://www.cri.haifa.ac.il/neurocomputation. 1598 H. Hazan, L.M. Manevitz / Expert Systems with Applications 39 (2012) 1597–1606 The term detector is standard in the LSM community and date back to Maass et al. (Jaeger, 2001a; Lukosevicius & Jaeger, 2009; Maass, 2002; Maass & Markram, 2004; Maass et al., 2002b) the idea is that the ‘‘detectors’’ are testing whether the information for classification resides in the liquid; and thus are not required to be biological. In this way, it is theoretically possible for the detectors to recognize any spatio-temporal signal that has been fed into the liquid; and thus the system could be used for, e.g. speech recognition, or vision, etc. This is an exciting idea and, e.g. Maass and his colleagues have published a series of papers on it. Amongst other things, they have recently shown that once a detector has been sufficiently trained at any time frame, it is resilient to noise in the input data and thus it can be used successfully for generalization (Bassett & Bullmore, 2006; Fern & Sojakka, n.d.; Maass et al., 2002b). Furthermore, there is a claim that this abstraction is faithful to the potential capabilities of the natural neurons and thus is explan- atory to some extent from the viewpoint of computational brain science. Note that one of the underlying assumptions is that the detector works without memory; that is the detector should be able to classify based on instantaneous static information; i.e. by sampling the liquid at a specific time. That this is theoretically pos- sible is the result of looking at the dynamical system of the liquid and noting that it is sufficient to cause the divergence of the two classes in the space of activation. Note that the detector systems (e.g. a back propagation neural network, a perceptron or a support vector machine (SVM)) are not required to have any biological plausibility; either in their de- sign or in their training mechanism, since the model does not try to account for the way the information is used in nature. Despite this, since natural neurons exist in a biological and hence noisy environ- ment, for these models to be successful in this domain, they must be robust to various kinds of noise. As mentioned above, Maass et al. (Lukosevicius & Jaeger, 2009; Maass, Legenstein, & Markram, 2002; Maass et al., 2002b; Maass & Markram, 2004) addressed one dimension of this problem by showing that the systems are in fact robust to noise in the input. Thus small random shifts in a temporal input pattern will not affect the LSM’s ability to recognize the pattern. From a machine learning perspective, this means that the model is capable of generalization. However, there is another component to robustness; that of the components of the system itself. In this paper we report on experiments performed with various kinds of ‘‘damage’’ to the LSM and unfortunately have shown that the LSM with any of the above detectors is not resistant, in the sense that small damages to the LSM neurons reduce the trained classifiers dramatically, even to essentially random values (Hazan & Manevitz, 2010; Manevitz & Hazan, 2010). Seeking to correct this problem, we experimented with differ- ent architectures of the liquid. The essential need of the LSM is that there should be sufficient recurrent connections so that on the one hand, the network maintains the information in a signal, while on the other hand it separates different signals. The models typically used are random connections; or those random with a bias to- wards ‘‘nearby’’ connections. Our experiments with these topolo- gies show that the network is very sensitive to damage because the recurrent nature of the system causes substantial feedback. Taking this as a clue, we tried networks with ‘‘hub’’ or ‘‘small world’’ (Albert & Barabási, 2000; Barabási, 2000; Barabási & Albert, 1999) architecture. This architecture has been claimed (Achard, Salvador, Whitcher, Suckling, & Bullmore, 2006; Bassett & Bullmore, 2006; Varshney, Chen, Paniagua, Hall, & Chklovskii, 2011) to be ‘‘biologically feasible’’. The intuition was that the hub topology, on the one hand, inte- grates information from many locations and so is resilient to dam- age in some of them; and on the other hand, since such hubs follow a power rule distribution, they are rare enough that damage usu- ally does not affect them directly. This intuition was in fact borne out by our experiments. 2. Materials and methods We simulated the Liquid State Machine with 243 integrate and fire neurons (LIF) in the liquid following the exact set up of Maass and using the code available at the Maass laboratory software ‘‘A neural Circuit SIMulator’’.2 To test variants of topology we re- implemented the code, available at our website.3 The variants of the topologies implemented are described in the paper below as are the types of damages. Input to the liquid was at 30% of the neu- rons, the same input at all locations in a given time instances. The detectors of the basic networks were back propagation networks with three levels with 3 neurons in the hidden level and one output neuron. In most experiments, the input was given by the output of all non-input neurons of the liquid (i.e. 170 inputs to the detector). In some experiments (see section below) the inputs to the detector were given over 20 time instances and so the detector had 3400 http://www.lsm.tugraz.at/csim/ http://www.cri.haifa.ac.il/neurocomputation H. Hazan, L.M. Manevitz / Expert Systems with Applications 39 (2012) 1597–1606 1599 inputs. The networks were tested with 20 random temporal binary sequences of length 45 chosen with uniform distribution. The experiments were repeated 500 times and statistics reported. 3. Theory/calculations As discussed in the introduction, in a system, there are two sources of potential instability. First is the issue of small variants in the input. Systems need to balance the need of separation with generalization. That is, on the one hand, one may need to separate inputs with small variations into separate treatment, but on the other hand, small variants may need to be treated as ‘‘noise’’ or generalization of the trained system. For the LSM, as is typically presented in the literature, it is understood, e.g. from the work of Lukosevicius and Jaeger (2009) and Maass (2002)) that the LSM and its variants do this successfully in the case of spatio-temporal signals. The second issue concerns that of the sensitivity of the system to small changes in the system itself, which we choose to call ‘‘damages’’ in this paper. This is very important if, as is the case for LSM, it is supposed to be explanatory for biological systems. Our experiments therefore are based on simulating the LSM with temporal sequences and calculating how resistant they are to two main kinds of such damages. The damages chosen for inves- tigation were: (1) at each time instance a certain percentage of neurons in the liquid would refuse to fire regardless of the internal charge in its state; (2) at each time instance a certain percentage of neurons would fire regardless of the internal charge, subject only to the limitation of the refractory period. Since the basic results (see below) showed that the standard variants of LSM were not robust to these damages at various small levels, we considered topological differences in the connectivity of the LSM. 3.1. First experiments: LSMs are not robust 3.1.1. The experiments To test the resistance of standard LSM to noise, we (i) down- loaded the code of Maass et al. from his laboratory site4 and then implemented two kinds of damage to the liquid and (ii) re-imple- mented the LSM code so that we could handle variants. These mod- els use a kind of basic neuron that is of the ‘‘leaky integrate and fire’’ (LIF)5 variety and in Maass’ work, the neurons are connected randomly but with some biologically inspired parameters: 20% inhibitory and a connectivity constraint giving a preference to geo- metrically nearby neurons over more remote ones. (For precise details on these parameters, see: neural Circuit SIMulator4 and Maass and Markram (2002).) External stimuli to the network were always sent to 30% of the neurons, always chosen to be excitatory neurons. Initially, we experimented with two parameters: (i) the percentage of neurons damaged; (ii) the kinds of damages. The kinds were either transforming a neuron into a ‘‘dead" neuron; i.e. one that never fires or transforming a neuron into a ‘‘generator’’ neuron, i.e. one which fires as often as its refractory period allows it, regardless of its input. We did experiments with different kinds of detectors: Adaline (Widrow & Hoff, 1960), Back-Propagation, SVM and Tempo- tron (Gutig & Sompolinsky, 2006). Classification of new data could then be done at any of the sig- nal points. We ran experiments as follows: we randomly chose twenty temporal inputs; i.e. random sequences of 0s and 1s of length 45, corresponding to spike inputs over a period of time; and trained an LSM composed of 243 integrate and fire neurons 4 A neural Circuit SIMulator: http://www.lsm.tugraz.at/csim/. 5 LIF = leaky integrate and fire. as in the liquid (Maass & Markram, 2002) to recognize ten of these inputs and reject the other ten. Each choice of architecture was run 500 times varying the precise connections randomly. We tested the robustness of the recognition ability of the network with the fol- lowing parameters: – The neurons in the network were either leaky integrate and fire neurons (Maass, 2002) or Izhikevich (Izhikevich, 2003) style neurons. – The average connectivity of the networks was maintained at about 20% chosen randomly in all cases although with different distributions. – The damages were either ‘‘generators", i.e. the neurons issued a spike whenever their refractory period allowed it; or they were ‘‘dead’’ neurons that could not spike. – The degree of damage was systematically checked at 0.1%, 0.5%, 1%, 5%, and 10% in randomly chosen neurons. The results shown in tables throughout the paper are in per- centages, over the (500) repeated tests. One hundred percent indi- cates that all the 20 vectors of one test, over 500 repetitions of the test were fully recognized correctly. Fifty percent indicates that only half the vectors over 500 times were recognized. (This corre- sponds to a chance baseline). The graphs presented below show the full distribution of all the tests and the results over all the kinds of damages and all varieties of topologies. As expected, they dis- tribute as Gaussian, but note that the average success rate varies from a baseline of 10 successes (50%) for random guessing (see Fig. 2) to as high as almost 20 (98%) for generalization in certain cases and 88% for some of the damages. 3.2. Second experiments: modifications of the LSM 3.2.1. Different kinds of basic neurons In attempts to restore the robustness to damage, we experi- mented with the possibility that a different kind of basic neuron might result in a more resilient network. Accordingly, we imple- mented the LSM with various variants of ‘‘leaky integrate and fire neurons’’, e.g. with history dependent refractory period (Manevitz & Marom, 2002) and by using the model of neurons due to Izhikevich (2003). The results under these variants were qualitatively the same as the standard integrate and fire neuron. (The Izhikevich model produces a much more dense activity in the network and thus the detector was harder to train but in the end the network was trainable and the results under damage were very similar.) Accordingly, we report only results with the standard integrate and fire neuron as appears, e.g. in Maass’ work (Maass, 2002). Fig. 2. Results of identification of random vectors on an untrained LSM with uniform random connections. This is a baseline. The result is a Gaussian distribution around 10 vectors. http://www.lsm.tugraz.at/csim/ 1600 H. Hazan, L.M. Manevitz / Expert Systems with Applications 39 (2012) 1597–1606 3.2.2. Allowing detectors to have memory In trying to consider how to make the model more robust to damage, we investigated the fact that the detector has no memory. Perhaps, if we allow the detector to follow the development of the network for a substantial amount of time, both in training and run- ning, it would be more robust. To check this, we took the most ex- treme other case; we assumed that the detector system in fact takes as input a full time course of 20 iterations of the output neu- rons of the liquid. This means that instead of a NN with input of 170; we had one with 20 times 170 time course inputs. It seemed reasonable that (i) with so much information, it should be rela- tively easy to train the detector; (ii) one could hope that damage in the liquid would be local enough that over the time period, the detector could correct for it. In order to test this, we re-imple- mented the LSM detector to allow for this time entry. Our detector was trained and tested as follows. There were 170 output units. At a ‘‘signal point’’ each of them was sampled for the next 20 iterations and all of these values were used as a single data point to the detector. Thus the detector had 170 times 20 inputs. We chose separate detector points typically at intervals of 50. We then used back propagation on these data points. This means that eventually the detector could recognize the signal at any of Fig. 3. Histogram of connection distributions when the output connections were randomly selected chosen according to a power-law. Note that the input histogram is different than the output histogram. Fig. 4. Maass LSM (a) normal operation; (b) with 10% dead damage; (c) with 10% the ‘‘signal points’’; after training there was no particular impor- tance to the choice of separation of the signal points except that there was no overlap between the data points. While we did not control for any connections between the intervals of data points (i.e. 50, and we also checked other time intervals) and possible nat- ural oscillations in the network, we do not believe there were any. As anticipated, there was no significant trouble in training the net- work to even 100% of recognition of the training data. The ‘‘detectors’’ were three level neural networks, trained by back-propagation. We also did some experiments with the Tempo- tron (Gutig & Sompolinsky, 2006); and with a simple Adaline detector (Widrow & Hoff, 1960). Training for classification could be performed in the damage-less environment successfully with any of these detectors. Then we exhaustively ran tests on these possibilities. In all of these tests, following Maass (2002), Maass and Markram (2002) and Maass et al. (2002a), we assumed that approximately 20% of the neurons of the liquid were of the inhib- itory type. The architecture of the neural network detector was 204 input neurons (which were never taken from the neurons in the LSM which were also used as inputs to the LSM) 100 hidden level neurons and one neuron for the output. Results running the Maass et al. architecture are presented in Fig. 4 and Table 4 and can be compared with a random connected network of 10% average connectivity, see Table 2. The bottom line (see the results section) was that even with low amounts damage and under most kinds of connectivity, the net- works would fail; i.e. the trained but damaged network loss of function was very substantial and in many cases could not perform substantially differently from a random selection. 3.3. Third experiments: changing the architecture Our next approach, and ultimately the successful one, was to experiment with different architectures. The underlying intuition is that the recurrent nature of the liquid results in feedback of information making the network dynamics too sensitive to changes in the network. Since one can look at ‘‘damages’’ as noise. One can easily discern the large change in the reaction of the network. H. Hazan, L.M. Manevitz / Expert Systems with Applications 39 (2012) 1597–1606 1601 instantaneous changes in the architecture, it seems reasonable to design architectures that can somehow ‘‘filter’’ out minor changes. The liquids were varied in their topologies in the following ways: 1. Random connectivity. Each neuron in the network is connected to 20% of the other neurons in a random fashion. (i) In the ori- ginal Maass topology the connections are chosen with a larger bias for nearby neurons (see Maass, 2002; Maass et al., 2002a; Maass, Natschläger, & Markram, 2002c). This is the literature standard and is what is usually meant as LSM. (ii) We also tested a network without such bias; i.e. the connections are chosen to 20% of the other neurons randomly and uniformly. The results presented below showed that these architectures are not robust. 2. Reducing the connectivity to 10% and 5% in the above arrange- ment. The intuition for this was that with lower connectivity, the feed-back should be reduced. The results presented below show that this intuition is faulty and that these networks are even less robust than the above (see Tables 1, 2, 5 and 6). 3. Implementation of ‘‘Hub’’ topologies in either input connectiv- ity or output. The intuition here is that the relative rarity of ‘‘hubs’’ results in their damage being a very rare event. But when they are not damaged, they receive information from many sources and can thus filter out the damage thus alleviat- ing the feedback in the input case. In the output hub case, the existence of many hubs should allow the individual neurons to filter out noise. The construction of hubs was done in various fashions: a. Hand design of a network with one hub for input. See Appendix A for a full description of this design. b. Small world topologies. Since small world topologies follow power law connectivity, they produce hubs. On the other hand such topologies are thought to emerge in a ‘‘natural’’ fashion (Albert & Barabási, 2000; Barabási, 2000; Barabási & Albert, 1999; Varshney et al., 2011) and appear in real neuronal systems (Albert & Barabási, 2000; Bassett & Bullmore, 2006), see Fig. 3. Note however, that in our context there are two directions to measure the power law: input and output connectivity histograms for the neurons. We checked the following variants: Table 1 Five percentage uniform random connectivity without memory input to the detector.a Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 100% 55% 53% 52% 51% 49% Noisy neurons 100% 63% 54% 55% 51% 50% Dead and noisy 100% 55% 52% 52% 50% 50% Generalization 100% 93% 88% 80% 75% 78% a For all the tables that are shown in this paper, 50% is the baseline of random classification. Table 2 Ten percentage uniform random connectivity without memory input to the detector. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 100% 56% 53% 51% 51% 49% Noisy neurons 100% 73% 58% 54% 51% 52% Dead and noisy 100% 59% 54% 52% 52% 51% Generalization 100% 100% 93% 88% 83% 81% i. Input connectivity is power law. That is we assign a link from a uniformly randomly chosen neuron to a second neu- ron chosen randomly according to a power law. In this case the input connectivity follows a power law; while the output connectivity follows a Gaussian distribution. ii. Output connectivity is power law. That is we reverse the above. In this case the input connectivity is Gaussian while the output connectivity is power law. iii. Replacing ‘‘Gaussian’’ with ‘‘uniform’’ in case (i) above. iv. Replacing ‘‘Gaussian’’ with ‘‘uniform’’ in case (ii) above. v. We also tried choosing a symmetric network with power law connectivity (i.e. for both input and output.) Note that in this case, the same neurons served as ‘‘hubs’’ both for input and output. vi. Finally, we designed an algorithm to allow distinct input and output power law connectivity. In this case the hubs in the two directions are distinct. Algorithms 1 and 2 below accom- plish this task. Algorithm 1 Generate a random number between min and max value with Power law distribution, Input: min,max, size, How_many_numbers, counter Arry = array, Magnify = 5 for i = 1 to How_many_numbers index = random(array.start,array.end) end_array = array.end candidate = array[index] AddCells(array, Magnify); for t = 0 to Magnify array[end_array + t] = candidate end for shuffle(array) output_Array[i] = candidate counterArry[candidate]++ end for shuffle(counterArry) Output output_Array,counterArry Algorithm 2 Create the connectivity matrix for the liquid network using the Algorithm 1 as an Input weight_Matrix use algorithm 1 to creart (arraylist, counterArry) counter = 0 for i=1 to counterArry.lenght for t=1 to counterArry[i] weight_Matrix[i, arraylist[counter]]=true counter++ end for end for One problem with the various algorithms for designing power law connectivity is that under a ‘‘fair’’ sampling, the network might not be connected. This means that such a network actually has a lower, effective connectivity. We decided to eliminate this problem by randomly connecting the disconnected components (either from an input or output perspective) to another neuron chosen randomly but proportionally to the connectivity. (This does not guarantee connectivity of the graph, but makes it unlikely, so that the effective connectivity is not substantially affected.) Table 4 Twenty percentage connectivity under Maass’s distribution preferring local connections. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 90% 60% 52% 51% 50% 50% Noisy neurons 90% 78% 57% 52% 52% 52% Dead and noisy 90% 54% 52% 53% 50% 50% Generalization 90% 96% 93% 93% 84% 84% Table 5 Five percentage uniform random connectivity with memory input to the detector. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 100% 55% 53% 53% 51% 50% Noisy neurons 100% 63% 54% 54% 53% 51% Dead and noisy 100% 56% 53% 52% 51% 51% Generalization 100% 93% 87% 80% 75% 79% Table 6 Ten percentage uniform random connectivity with memory input to the detector.a Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 100% 58% 55% 53% 49% 50% Noisy neurons 100% 74% 59% 57% 54% 50% Dead and noisy 100% 61% 54% 55% 50% 50% Generalization 100% 96% 92% 85% 82% 82% a For all the tables that shown in this paper, 50% is the baseline of random classification. Table 7 Twenty percentage uniform random connectivity with memory input to the detector. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 100% 63% 55% 52% 50% 50% Noisy neurons 100% 87% 67% 61% 54% 52% Dead and noisy 100% 68% 57% 52% 50% 49% Generalization 100% 98% 97% 95% 89% 86% Table 8 Maass’s distribution like in Table 4 but with memory input to the detectors. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 100% 61% 53% 49% 49% 50% Noisy neurons 100% 79% 60% 55% 51% 49% Dead and noisy 100% 64% 55% 52% 51% 52% Generalization 100% 100% 96% 93% 84% 85% 1602 H. Hazan, L.M. Manevitz / Expert Systems with Applications 39 (2012) 1597–1606 4. Results 4.1. First experiments: LSM is not robust First, there was not much difference between the detectors; so eventually we restricted ourselves to the back-propagation detector. (Note that none of units of the liquid input were accessed by the detectors were allowed to be input neurons of the liquid.) It turned out that while the detector is able to learn the randomly chosen test classes successfully, if there is sufficient average connectivity (e.g. 20%), almost any kind of damage caused the detector to have a very substantial decay in its detecting ability (see Table 3). Note that even with lower connectivity, which has less feedback, the same phenomenon occurs. See Table 1 (5% connectivity) and Table 2 (10% connectivity). When the network is connected randomly but with bias for geo- metric closeness as in Maass’ distribution, the network is still very sensitive (although a bit less so). Compare Table 4 to Table 3. After our later experiments, we returned to this point (see con- cluding remarks, below). In Fig. 4 we illustrate the difference in reaction of the network by a raster (ISI) display. Note that with 10% damage, it is quite evident to the eye that the network di- verges dramatically from the noise free situation. In Tables 1–4 one can see this as well with 5% noise for purely random connec- tivity. Actually, with low degrees of damage the detectors under even the Maass connectivity (see Table 4) show dramatic decay in recognition although not to the extremes of random connectiv- ity. These results (see Tables 1–4) were robust and repeatable under many trials and variants. Accordingly, we conclude that the LSM, either as purely defined with random connectivity, or, as implemented in Maass et al. (2002a) cannot serve as a biologically relevant model. 4.2. Second experiments: varying the neurons and allowing the detectors to have memory 4.2.1. Variants of neurons (history dependent refractory period and izhikevich) The results under these variants were qualitatively the same as the standard integrate and fire neuron. (The Izhikevich model pro- duces a much more dense activity in the network and thus the detector was harder to train but in the end the network was train- able and the results under damage were very similar.) Accordingly, we report only results with the standard integrate and fire neuron as appears, e.g. in Maass’ work. 4.2.2. Detectors with memory input The ‘‘detectors’’ in our experiments were either three level neu- ral networks, trained by back-propagation, the Tempotron (Gutig & Sompolinsky, 2006); or with a simple Adaline detector (Widrow & Hoff, 1960). Training for classification could be performed in the damage-less environment successfully with any of these detectors. We exhaustively ran tests on these possibilities; including damage degree and kinds and detector types. Tables 5–8 show the results with different uniform connectivity in the liquid when there is memory input to the detector. Table 8 Table 3 Twenty percentage uniform random connectivity without memory input to the detector. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 99% 60% 53% 51% 51% 50% Noisy neurons 99% 86% 65% 58% 52% 50% Dead and noisy 99% 65% 55% 53% 50% 51% Generalization 99% 100% 97% 94% 87% 84% Fig. 5. Histographs of correctness results in LSM networks with 20 time interval input, different amounts of ‘‘dead’’ neuron damage, average connectivity of 20% with a uniform random distribution on the connections. Fig. 6. Histographs of correctness results in LSM networks with 20 time interval input, different amounts of ‘‘noise generator’’ neuron damage, average connectivity of 20% with a uniform random distribution on the connections. Fig. 7. Histographs of correctness results in LSM networks with one hub distribu- tion with different amounts of ‘‘noise generator’’ neuron damage. Fig. 8. Histographs of correctness results in LSM networks with different amounts of ‘‘dead’’ neuron damage with one hub distribution. H. Hazan, L.M. Manevitz / Expert Systems with Applications 39 (2012) 1597–1606 1603 shows similar result (like in Table 4) for the Maass connectivity with memory input to the detector. Histographs of sample results with 5% and 10% damage for the neural network detectors are pre- sented in Figs. 5–13 . (Since the results for the other detectors were similar, we did not run as many tests on them) Note, Figs. 5–13 re- fer to the various kinds of hub architectures with the memory in the detector. In all of these tests, following Maass, we assumed that approx- imately 20% of the neurons of the liquid were of the inhibitory type. The architecture of the neural network detector was 204 in- put neurons (which were never taken from the neurons in the LSM which were also used as inputs to the LSM � 30 times) 3 hidden le- vel neurons and one neuron for the output. For 20% connections the Maass et al. architecture without memory in the detector as presented in Table 4 can be compared with a uniform random con- nected network of 20% average connectivity without memory in the detector in Table 3, and can be compared as well with the Maass topology with memory in Table 8 and with uniform random of 20% connectivity with memory in Table 7. Note that Table 1 can be also compared to Table 5 and Table 2 can be compared to Table 6. Since this paper is about robustness Figs. 5 and 6 present the full distribution of the experiments of Table 7 under these conditions with different degrees of damage. Note that with damage over 1%, the histogram deteriorates dramatically. The bottom line of all these comparisons is that decreasing con- nectivity and adding memory to the detector slightly increases the robustness performance with low amounts of damage, but even with low amounts of damage and under all our variants of random connectivity, the networks would fail. That is, the trained but dam- aged network loss of function was very substantial and in many cases could not perform substantially differently from a random classification (see Figs. 5 and 6). 4.3. Third experiments: varying the architecture of the network 4.3.1. Hand chosen one-hub topology Since the Maass et al. topology and the uniform random distri- bution topology showed a high level of vulnerability to any small amount of damage, and since adding memory to the detector helped only marginally to recover from damage in the liquid; we Table 9 One hub network with memory input to the detector. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 100% 95% 88% 85% 76% 67% Noisy neurons 100% 97% 91% 86% 70% 62% Dead and noisy 100% 96% 89% 86% 75% 68% Generalization 100% 100% 97% 97% 96% 95% started to create different topologies to test the robustness of the liquid with the same parameters as those set by Maass et al. (see Jaeger, 2001a; Lukosevicius & Jaeger, 2009; Maass, 2002; Maass et al., 2002a, 2002c, 2002d; Natschläger, Maass, & Markram, 2002a; Natschläger, Markram, & Maass, 2002b). One of those topologies is the hub topology that is described in detail in Appen- dix A. In this case, one can see from Table 9, Figs. 7 and 8, that the robustness was substantially increased. However, under the con- struction as presented in Appendix A, there can appear substantial disconnected components in the liquid. Moreover, in results not presented in this paper, the signal has weaker persistence, i.e. detectors are able to recognize the signals in a substantially smal- ler time window. 4.3.2. Small world topologies The general connectivity in the human brain has been held to have some small world properties (Achard et al., 2006) Algorithm 1 is designed to obtain a hub topology using a more ‘‘natural’’ algo- rithm thus creating a topology that will be robust to damage in the liquid and to be more ‘‘natural’’ in its construction. One of the properties of the small world is the power-law distribution. The results of the small world with a power-law distribution (see Table 10, Figs. 9 and 10) were, however, very similar to the Maass topology and to the uniform random topology in the Table 10 Small world with a power-law distribution with memory input to the detector. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 100% 55% 51% 51% 50% 51% Noisy neurons 100% 79% 58% 53% 50% 51% Dead and noisy 100% 58% 51% 50% 48% 50% Generalization 100% 100% 97% 93% 90% 89% Fig. 9. Histographs of correctness results in LSM networks with different amounts of ‘‘dead’’ neuron damage with small world topology obtained with a power law distribution. Fig. 10. Histographs of correctness results in LSM networks with different amounts of ‘‘noise generator’’ neuron damage for small world topology obtained with a power-law distribution. Table 11 Small world with a double power-law distribution with memory input to the detector. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 96% 95% 87% 83% 74% 69% Noisy neurons 96% 99% 93% 88% 72% 64% Dead and noisy 96% 97% 89% 84% 70% 66% Generalization 96% 99% 99% 98% 97% 97% Table 12 Small world with a double power-law distribution without memory input to the detector. Damage Non 0.1% 0.5% 1% 5% 10% Dead neurons 62% 83% 67% 61% 56% 53% Noisy neurons 62% 91% 75% 66% 54% 55% Dead and noisy 62% 86% 69% 65% 52% 55% Generalization 62% 100% 96% 95% 93% 91% Fig. 12. Histographs of correctness results in LSM networks with different amounts of ‘‘dead’’ neuron damage with small world topology obtained with a double power law distribution. Fig. 13. Histographs of correctness results in LSM networks with different amounts of ‘‘noise generator’’ neuron damage for small world topology obtained with a double power-law distribution. 1604 H. Hazan, L.M. Manevitz / Expert Systems with Applications 39 (2012) 1597–1606 robustness from damage in the liquid. On the other hand, they had improved generalization capability (see Table 10). Looking closer at the distribution, as can be seen from Fig. 3, Algorithm 1 actually creates a power-law distribution in terms of total connections, but when we separate the connections to input and output connections, we see that while the output has a power law distribution, the input connections have a roughly random uniform distribution. 4.3.3. Small world topologies with double power-law distribution Accordingly, using Algorithms 1 and 2 we created a double power-law distribution (using the reverse order for input connec- tions and output connections as in Fig. 11). The robustness and the generalization ability was much improved The best results were with a double-power law where the distributions are over distinct neurons and these are the results presented here in Tables 11 and 12 and Figs. 12 and 13 . Fig. 11. Connection distribution of small-world with double power-law. 5. Discussion In this work, we looked at the robustness of the LSM paradigm and by experimenting with temporal sequences showed that the basic structural set up in the literature is not robust to two kinds of damages; even at small levels of damages. We also investigated this for various degrees of connectivity. While lowering the average degree of connectivity resulted in decreased sensitivity in all architectures to some extent, the bot- tom line is that decreased connectivity is ineffective. In addition, it became evident that lowering the connectivity also decreases the strength the network has in representability and, importantly, Fig. 14. A graphical summary of the results presented in this paper. The ‘‘standard’’ LSM topologies either uniform or in Maass’s original papers are not robust; but small world topologies show an improvement, which is most marked in the case of a two-way power law distribution. Fig. A1. Hub topology. H. Hazan, L.M. Manevitz / Expert Systems with Applications 39 (2012) 1597–1606 1605 in the persistence of the signal. (That is, a low degree of connectiv- ity causes the activity to die down quickly because of the lack of feedback. Thus the network is bounded in time and cannot recognize an ‘‘older’’ input signal.) Thus we see, as is to be expected from the analysis in Jaeger (2001a, 2001b, 2002) and Maass et al. (2002a) that a higher connectivity gives a larger set of ‘‘filters’’ that separate signals, but on the other hand makes it more sensitive to changes. In any case, even with low connectivities, the random topology was not robust; nor was the Maass topology. (While not at random levels of identification, as we have seen, e.g. in Tables 1 and 2 it suf- fered very substantial decays with even small amounts of damages. In addition, other experiments (not shown here) with connectivi- ties below 15–20%, show that the networks do not maintain the trace for very long.) We also investigated some variants in the kinds of neurons. It seems that the LSM (or ‘‘reservoir computing’’ concept) does not change much vis a vis robustness to internal noise based on these choices. We did see substantial improvement when supplying a window of time input to the detector rather than an instant of time. How- ever, alone this was not sufficient. The major affect was changing the topology of connectivity to accommodate the idea of hubs, power law and small world con- nectivity. Under these topologies, with the best result occurring when we have power law histogram of both input and output con- nectivity to the neurons with separate neurons as hubs in both directions, the liquids are robust to damages. 6. Conclusions We have shown experimentally that the basic LSM is not robust to ‘‘damages’’ in its underlying neurons and thus without elabora- tion cannot be seen to be a good fit for a model for biological com- putation. (We mention (data not shown here) that this result holds even if training is continued while the network is suffering damage.) However, choosing certain power law topologies of the connec- tivity can result in more robust maintenance of the pertinent infor- mation over time. A graphical summary of the results for robustness under different topologies is given in Fig. 14. In the papers (Bassett & Bullmore, 2006; Varshney et al., 2011), a distribution was chosen for biological reasons to allow preference for close neurons. This distribution is superior to the totally ran- dom one, but is still not sufficiently robust. Choosing a power law distribution and being careful to making the assignments dif- ferently for in and out connectivity proved to be the best. Since this is thought of as a potentially biological arrangement (Barabási & Albert, 1999; Bassett & Bullmore, 2006); LSM style networks with this additional topological constraint can, as of this date, be consid- ered sufficiently biological. Other distributions may also work. Acknowledgements We want to acknowledge the distinction and support provided by Sociedad Mexicana de Inteligencia Artificial (SMIA) and the 9th Mexican International Conference on Artificial Intelligence (MICAI- 2010) in order to enhance, improve, and publish this work. We thank the Caesarea Rothschild Institute for support of this research. The first author thanks Prof. Alek Vainstein for support in the form of a research fellowship. A short version of this work was presented in the MICAI-2010 meeting (Manevitz & Hazan, 2010) whom we thank for inviting us to write this extended ver- sion. We also thank the Maass laboratory for the public use of their code. Appendix A The architecture one hub was made as the following: 1606 H. Hazan, L.M. Manevitz / Expert Systems with Applications 39 (2012) 1597–1606 � Divide all the neurons (240) to groups; the size of each group is randomly chosen between 3 and 6 neurons in one group. Each neuron in the entire group connects to 2 of his neighbors in the same group. � Choose 1=4 of the groups to be hubs and the rest of the groups we will call them the base. � For 20% connection (that is 11,472 connections) 90% of the con- nections are from the base groups to the hub groups, 7% are from the hub group to base group and 3% are connections between the hub groups. To accomplish that: – Choose (10324 times) random neurons from the base groups and connect each one with a randomly neuron from a hub group. – Randomly choose (803 times) a randomly neuron and connect it to a randomly chosen neuron from the base neurons. – Connect (345 times) randomly one of the neurons from the hub neurons to anther neuron but from a different group (see Fig. A1). References Achard, S., Salvador, R., Whitcher, B., Suckling, J., & Bullmore, E. (2006). A Resilient, low-frequency, small-world human brain functional network with highly connected association cortical hubs. The Journal of Neuroscience, 26(1), 63–72. doi:10.1523/JNEUROSCI.3874-05.2006. Albert, R., & Barabási, A.-L. (2000). Topology of evolving networks: Local events and universality. Physical Review Letters, 85(24), 5234–5237. Retrieved from <http:// www.ncbi.nlm.nih.gov/pubmed/11102229>. Barabási, G. B. A.-L. (2000). Competition and multiscaling in evolving networks. cond-mat/0011029. Retrieved from <http://arxiv.org/abs/cond-mat/0011029>. Barabási, A.-L., & Albert, R. (1999). Emergence of scaling in random networks. Science, 286(5439), 509–512. doi:10.1126/science.286.5439.509. Bassett, D. S., & Bullmore, E. (2006). Small-world brain networks. The Neuroscientist, 12(6), 512–523. doi:10.1177/1073858406293182. Fern, C., & Sojakka, S. (n.d.). Pattern recognition in a bucket. Retrieved from <http:// citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.97.3902>. Gutig, R., & Sompolinsky, H. (2006). The tempotron: A neuron that learns spike timing-based decisions. Nature Neuroscience, 9(3), 420–428. doi:10.1038/ nn1643. Hazan, H., & Manevitz, L. M. (2010). The liquid state machine is not robust to problems in its components but topological constraints can restore robustness. In IJCCI (ICFC-ICNC) (pp. 258–264). Izhikevich, E. M. (2003). Simple model of spiking neurons. IEEE Transactions on Neural Networks, 14(6), 1569–1572. doi:10.1109/TNN.2003.820440. Jaeger, H. (2001a). The ‘‘echo state’’ approach to analysing and training recurrent neural networks (No. GMD Report 148). German National Research Center for Information Technology. Retrieved from <http://www.faculty.iu-bremen.de/ hjaeger/pubs/EchoStatesTechRep.pdf>. Jaeger, H. (2001b). Short term memory in echo state networks (No. GMD Report 152). German National Research Center for Information Technology. Retrieved from <http://www.faculty.iu-bremen.de/hjaeger/pubs/STMEchoStatesTechRep.pdf>. Jaeger, H. (2002). Adaptive nonlinear system identification with echo state networks. Retrieved from <http://www.faculty.iu-bremen.de/hjaeger/pubs/esn_NIPS02>. Lukosevicius, M., & Jaeger, H. (2009). Reservoir computing approaches to recurrent neural network training. Computer Science Review, 3(3), 127–149. doi:10.1016/ j.cosrev.2009.03.005. Maass, W. (2002). Paradigms for computing with spiking neurons. In J. L. van Hemmen, J. D. Cowan, & E. Domany (Eds.). Models of neural networks. Early vision and attention (Vol. 4, pp. 373–402). New York: Springer. Maass, W., Legenstein, R. A., & Markram, H. (2002). A new approach towards vision suggested by biologically realistic neural microcircuit models. In Proceedings of the 2nd workshop on biologically motivated computer vision. Lecture notes in computer science. Springer. Retrieved from papers/lsm-vision-146.pdf. Maass, W., & Markram, H. (2002). Temporal integration in recurrent microcircuits. In M. A. Arbib (Ed.), The handbook of brain theory and neural networks (2nd ed.. Cambridge: MIT Press. Maass, W., & Markram, H. (2004). On the computational power of circuits of spiking neurons. Journal of Computer and System Sciences, 69(4), 593–616. doi:10.1016/ j.jcss.2004.04.001. Maass, W., Natschläger, T., & Markram, H. (2002a). Computational models for generic cortical microcircuits. In J. Feng (Ed.), Computational neuroscience: A comprehensive approach. CRC-Press. Retrieved from papers/lsm-feng-chapter- 149.pdf. Maass, W., Natschläger, T., & Markram, H. (2002b). Real-time computing without stable states: A new framework for neural computation based on perturbations. Neural Computation, 14(11), 2531–2560. Retrieved from papers/lsm-nc-130.pdf. Maass, W., Natschläger, T., & Markram, H. (2002c). A model for real-time computation in generic neural microcircuits. In Proceedings of NIPS 2002 (Vol. 15, pp. 229–236). Retrieved from papers/lsm-nips-147.pdf Maass, W., Natschläger, T., Markram, H. (2002d). A fresh look at real-time computation in generic recurrent neural circuits, Tech. Report, Institute for Theoretical Computer Science, TU Graz, Graz, Austria. Manevitz, L., & Hazan, H. (2010). Stability and topology in reservoir computing. In G. Sidorov, A. Hernández Aguirre, & C. Reyes García (Eds.), Advances in soft computing. Lecture notes in computer science (Vol. 6438, pp. 245–256). Berlin/ Heidelberg: Springer. Retrieved from <http://dx.doi.org/10.1007/978-3-642- 16773-7_21>. Manevitz, L. M., & Marom, S. (2002). Modeling the process of rate selection in neuronal activity. Journal of Theoretical Biology, 216(3), 337–343. Retrieved from <http://www.ncbi.nlm.nih.gov/pubmed/12183122>. Natschläger, T., Maass, W., & Markram, H. (2002). The ‘‘Liquid Computer’’: A novel strategy for real-time computing on time series. Special Issue on foundations of information processing of TELEMATIK, 8 (1), 39–43. Retrieved from papers/lsm- telematik.pdf. Natschläger, T., Markram, H., & Maass, W. (2002). Computer models and analysis tools for neural microcircuits. In R. Kötter (Ed.), A practical guide to neuroscience databases and associated tools. Boston: Kluwer Academic Publishers. Retrieved from papers/lsm-koetter-chapter-144.pdf. Pitts, W., & McCulloch, W. S. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biology, 52(1–2), 99–115. discussion 73-97. Varshney, L. R., Chen, B. L., Paniagua, E., Hall, D. H., & Chklovskii, D. B. (2011). Structural Properties of the Caenorhabditis elegans Neuronal Network. PLoS Computational Biology, 7(2), e1001066. doi:10.1371/journal.pcbi.1001066. Widrow, B., & Hoff, M. (1960). Adaptive switching circuits. 1960 {IRE} {WESCON} Convention Record, Part 4 (pp. 96–104). {IRE}. Retrieved from <http://isl- www.stanford.edu/~widrow/papers/c1960adaptiveswitching.pdf>. http://dx.doi.org/10.1523/JNEUROSCI.3874-05.2006 http://www.ncbi.nlm.nih.gov/pubmed/11102229 http://www.ncbi.nlm.nih.gov/pubmed/11102229 http://arxiv.org/abs/cond-mat/0011029 http://dx.doi.org/10.1126/science.286.5439.509 http://dx.doi.org/10.1177/1073858406293182 http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.97.3902 http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.97.3902 http://dx.doi.org/10.1038/nn1643 http://dx.doi.org/10.1038/nn1643 http://dx.doi.org/10.1109/TNN.2003.820440 http://www.faculty.iu-bremen.de/hjaeger/pubs/EchoStatesTechRep.pdf http://www.faculty.iu-bremen.de/hjaeger/pubs/EchoStatesTechRep.pdf http://www.faculty.iu-bremen.de/hjaeger/pubs/STMEchoStatesTechRep.pdf http://www.faculty.iu-bremen.de/hjaeger/pubs/esn_NIPS02 http://dx.doi.org/10.1016/j.cosrev.2009.03.005 http://dx.doi.org/10.1016/j.cosrev.2009.03.005 http://dx.doi.org/10.1016/j.jcss.2004.04.001 http://dx.doi.org/10.1016/j.jcss.2004.04.001 http://dx.doi.org/10.1007/978-3-642-16773-7_21 http://dx.doi.org/10.1007/978-3-642-16773-7_21 http://www.ncbi.nlm.nih.gov/pubmed/12183122 http://dx.doi.org/10.1371/journal.pcbi.1001066 http://isl-www.stanford.edu/~widrow/papers/c1960adaptiveswitching.pdf http://isl-www.stanford.edu/~widrow/papers/c1960adaptiveswitching.pdf Topological constraints and robustness in liquid state machines 1 Introduction 2 Materials and methods 3 Theory/calculations 3.1 First experiments: LSMs are not robust 3.1.1 The experiments 3.2 Second experiments: modifications of the LSM 3.2.1 Different kinds of basic neurons 3.2.2 Allowing detectors to have memory 3.3 Third experiments: changing the architecture 4 Results 4.1 First experiments: LSM is not robust 4.2 Second experiments: varying the neurons and allowing the detectors to have memory 4.2.1 Variants of neurons (history dependent refractory period and izhikevich) 4.2.2 Detectors with memory input 4.3 Third experiments: varying the architecture of the network 4.3.1 Hand chosen one-hub topology 4.3.2 Small world topologies 4.3.3 Small world topologies with double power-law distribution 5 Discussion 6 Conclusions Acknowledgements Appendix A References 
work_cogaeulogjaplho23qnrwh6ynm	----	Microsoft Word - SALMon - JCIS.doc (*) Artículo publicado en Expert Systems with Applications, 42(19), 2015. (I.F.: 2.24) Monitoring the service-based system lifecycle with SALMon* Marc Oriol, Xavier Franch, Jordi Marco Universitat Politècnica de Catalunya, Barcelona (Spain), GESSI research group. {moriol,franch}@essi.upc.edu, jmarco@cs.upc.edu 1 Resumen Los Sistemas Basados en Servicios (SBS) son sistemas software altamente dinámicos compuestos por un conjunto de servicios web provenientes de distintos, y posible- mente heterogéneos, proveedores. En contraste con otros sistemas software tradicio- nales, el comportamiento dinámico de los SBS requiere de información actualizada sobre la calidad de servicio (QoS) para poder actuar i administrar correctamente las actividades en las distintas fases del ciclo de vida de los SBS (p.e., selección de servi- cios, despliegue, evaluación de niveles de acuerdo de servicio –SLA–, y adaptación). La necesidad de proveer esta información referente a la QoS ha resultado en distintas soluciones tecnológicas construidas alrededor de un componente software de monito- rización (abreviadamente, monitor). Sin embargo, los monitores actualmente disponi- bles en el estado del arte han sido diseñados para apoyar sólo un subconjunto de las actividades del ciclo de vida del SBS. Para cubrir esta brecha de investigación, presentamos SALMon, una plataforma de monitorización de servicios versátil que provee información acerca de la QoS según la forma y enfoque adecuado para las distintas actividades del ciclo de vida. Para ello, hemos identificado los requisitos de las distintas actividades mediante (1) el análisis de las necesidades descritas en la literatura y (2) reuniones y entrevistas con miembros de distintos grupos de investigación que trabajan en dichas actividades. A partir de estos requisitos, hemos desarrollado SALMon. Uno de los principales obstáculos es que los requisitos de monitorización para cada tarea son distintos, y en algunos casos, pueden ser contradictorios. Esto nos ha llevado a diseñar SALMon como un sistema modular, capaz de conectar los distintos módulos de manera que permita configurar la forma final del monitor según las necesidades de cada proyecto. Los principales avances y características que SALMon provee son los siguientes:  Combinación de estrategias de monitorización pasiva y testing online.  Combinación de estrategias de configuración basado en transformación de mo- delos y mediante invocación directa.  Extensibilidad mediante el añadido de nuevos atributos de calidad.  Bajo acoplamiento con las tecnologías de los servicios monitorizados.  Alta interoperabilidad. Para validar nuestra propuesta, hemos integrado SALMon con plataformas que cu- bren las actividades del ciclo de vida anteriormente descritas. Para ello, hemos reali- zado un conjunto de colaboraciones con otras universidades y centros de investiga- ción, donde SALMon ha sido integrado con sus herramientas. En particular:  WeSSQoS [1] (Selección de servicios): Sistema experto capaz de tratar con dis- tintas estrategias de selección para servicios web según la QoS requerida.  FCM [2] (Despliegue): Sistema experto para la toma de decisiones en cuanto a la mejor estrategia de despliegue en un sistema de federación de clouds.  SALMonADA [3] (Validación de SLAs): Una plataforma capaz de identificar y reportar violaciones de SLAs con la descripción de las causas de violación.  MAESoS [4], PROSA [5], PROTEUS [6], CARE [7] (Adaptación): Plataformas que dan soporte a SBS adaptativos, capaces de definir y ejecutar adaptaciones en respuesta a un mal funcionamiento o fallo de los servicios del sistema. 2 Agradecimientos Este trabajo ha sido parcialmente financiado por el gobierno español mediante los proyectos CICYT TIN2010-19130-C02-01 y TIN2013-44641-P; y por la Unión Eu- ropea mediante la red de excelencia de servicios S-Cube, número de contrato 215483. 3 Referencias [1] O. Cabrera, M. Oriol, X. Franch, J. Marco, L. López, O. Fragoso, R. Santaolaya. Open framework for web service selection using multimodal and configurable tech- niques. In Computación y Sistemas (CyS). 2014 [2] A. Kertész, G. Kecskemeti, M. Oriol, P. Kotcauer, S. Acs, M. Rodríguez, O. Mercè, A. Cs. Marosi, J. Marco, X. Franch. Enhancing Federated Cloud Management with an Integrated Service Monitoring Approach. In JGC, 11(4), 2012. [3] C. Müller, M. Oriol, X. Franch, J. Marco, M. Resinas, A. Ruiz-Cortés, M. Rodrí- guez. Comprehensive Explanation of SLA Violations at Runtime. In TSC, 7(2), 2013. [4] X. Franch, P. Gruenbacher, M. Oriol, B. Burgstaller, D.Dhungana, L. López, J. Marco, J. Pimentel. Goal-driven Adaptation of Service-Based Systems from Runtime Monitoring Data, in Procs. of COMPSACW, 2011. [5] O. Sammodi, A. Metzger, X. Franch, M. Oriol, J. Marco, K. Pohl. Usage-based online testing for proactive adaptation of service-based applications, in Procs. of COMPSAC, 2011. [6] E. Schmieders, A. Micsik, M. Oriol, K. Mahbub, R. Kazhamiakin. Combining SLA prediction and cross layer adaptation for preventing SLA violations, in Procs of. 2nd WoSS, 2012. [7] M. Oriol, N. Qureshi, X. Franch, A. Perini, J. Marco. Requirements monitoring for adaptive service-based applications, in Procs. of REFSQ, 2012. 
work_cozpjz3benauvg56gi3my4mj74	----	Home Page: World Allergy Organization Journal Skip to Main Content Login to your account Email/Username Password ShowForgot password? Remember me Don’t have an account? Create a Free Account If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password Email* If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password Cancel Close Home Articles & Issues Back Current Issue List of Issues For Authors Back About Open Access  Author Information Researcher Academy  Submit Your Manuscript  Journal Info Back About the Journal Contact Us Editorial Board New Content Alerts Submit Your Manuscript  World Allergy Organization  Go search All Content Article Title Authors Keywords Abstract Article Title, Abstract, Keywords Advanced SearchSave search Please enter a term before submitting your search. Ok Log in Register Log in Issue Highlights World allergy organization anaphylaxis guidance 2020 The usage, quality and relevance of information and communications technologies in patients with chronic urticaria: A UCARE study Immunopathological features of air pollution and its impact on inflammatory airway diseases (IAD) Food protein-induced allergic proctocolitis in infants: Literature review and proposal of a management protocol Assessment of Google Trends terms reporting allergies and the grass pollen season in Ukraine Previous slide Pause slide Next slide Most Read Most Cited World Allergy Organization Anaphylaxis Guidance 2020 Victoria Cardona, Ignacio J. Ansotegui, Motohiro Ebisawa, Yehia El-Gamal, Montserrat Fernandez Rivas, Stanley Fineman, Mario Geller, Alexei Gonzalez-Estrada, Paul A. Greenberger, Mario Sanchez Borges, Gianenrico Senna, Aziz Sheikh, Luciana Kase Tanno, Bernard Y. Thong, Paul J. Turner, Margitta Worm DOI: https://doi.org/10.1016/j.waojou.2020.100472 Vol. 13, Issue 10 Published online: October 30, 2020 Open Access COVID-19 vaccine-associated anaphylaxis: A statement of the World Allergy Organization Anaphylaxis Committee Paul J. Turner, Ignacio J. Ansotegui, Dianne E. Campbell, Victoria Cardona, Motohiro Ebisawa, Yehia El-Gamal, Stanley Fineman, Mario Geller, Alexei Gonzalez-Estrada, Paul A. Greenberger, Agnes S.Y. Leung, Michael E. Levin, Antonella Muraro, Mario Sánchez Borges, Gianenrico Senna, Luciana K. Tanno, Bernard Yu-Hor Thong, Margitta Worm On behalf of the WAO Anaphylaxis Committee DOI: https://doi.org/10.1016/j.waojou.2021.100517 Vol. 14, Issue 2 Published online: February 3, 2021 Open Access Inborn errors of immunity with atopic phenotypes: A practical guide for allergists Riccardo Castagnoli, Vassilios Lougaris, Giuliana Giardino, Stefano Volpi, Lucia Leonardi, Francesco La Torre, Silvia Federici, Stefania Corrente, Bianca Laura Cinicola, Annarosa Soresina, Caterina Cancrini, Gian Luigi Marseglia, Fabio Cardinale On behalf of the Immunology Task Force of the Italian Society of Pediatric Allergy and Immunology (SIAIP) DOI: https://doi.org/10.1016/j.waojou.2021.100513 Vol. 14, Issue 2 Published online: February 23, 2021 Open Access Enhancing innate immunity against virus in times of COVID-19: Trying to untangle facts from fictions Désirée Larenas-Linnemann, Noel Rodríguez-Pérez, Alfredo Arias-Cruz, María Virginia Blandón-Vijil, Blanca E. Del Río-Navarro, Alan Estrada-Cardona, José E. Gereda, Jorge A. Luna-Pech, Elsy Maureen Navarrete-Rodríguez, Ernesto Onuma-Takane, César Fireth Pozo-Beltrán, María Isabel Rojo-Gutiérrez DOI: https://doi.org/10.1016/j.waojou.2020.100476 Vol. 13, Issue 11 Published online: October 9, 2020 Open Access COVID-19, asthma, and biological therapies: What we need to know Mário Morais-Almeida, Rita Aguiar, Bryan Martin, Ignacio J. Ansotegui, Motohiro Ebisawa, L. Karla Arruda, Marco Caminati, Giorgio Walter Canonica, Tara Carr, Geoffrey Chupp, Jonathan Corren, Ignacio Dávila, Hae-Sim Park, Nicola A. Hanania, Lanny Rosenwasser, Mario Sánchez-Borges, J. Christian Virchow, Anahí Yáñez, Jonathan A. Bernstein, Luis Caraballo, Yoon-Seok Chang, Manana Chikhladze, Alessandro Fiocchi, Sandra N. González-Diaz, Luciana Kase Tanno, Michael Levin, Jose António Ortega-Martell, Giovanni Passalacqua, David B. Peden, Philip W. Rouadi, James L. Sublett, Gary W.K. Wong, Eugene R. Bleecker DOI: https://doi.org/10.1016/j.waojou.2020.100126 Vol. 13, Issue 5 Published online: May 16, 2020 Open Access 2020 Cited articles About Us About the Journal The official journal of the World Allergy Organization, the World Allergy Organization Journal provides the global scientific and medical communities with up-to-date, quality content focused on the epidemiology of allergy, asthma, anaphylaxis, and clinical immunology. As an open access journal, WAOjournal can direct world-wide attention to the most important characteristics of allergy, asthma, and anaphylaxis, including current diagnostic and therapeutic options. In keeping with the mission of its sponsoring organization, the Journal is also committed to focusing on the epidemiologic consequences of climate change and the subsequent results in food habits and environmental pollution. This dissemination of quality information in an open access format will improve patient care on a global basis. More About the World Allergy Organization The World Allergy Organization (WAO) is an international umbrella organization whose members consist of 99 regional and national allergology and clinical immunology societies from around the world. By collaborating with member societies, WAO provides direct educational outreach programs, symposia and lectureships to members in nearly 100 countries around the globe. Meet the Editors Alessandro Fiocchi, MD "It is worth remembering that we are all dedicated to the care of patients with allergic diseases, and the Journal is dedicated to promoting any activity that enhances our capabilities in this area and extends new knowledge." Erika Jensen-Jarolim, MD "Allergy is a global clinical problem, but the answer to it lies in the allergen molecules. Only when understanding the mechanisms will we have a chance to prevent, diagnose and treat allergies in our patients in an optimal way." Current Issue April 2021 Volume 14, Issue 4 Past Issues New Content Alerts Submit a Manuscript Access this journal on ScienceDirect Visit ScienceDirect to see if you have access via your institution. List of Issues Open Access Journal Learn more about open access. Journal Metrics 3.506 2019 Impact Factor 1.043 SCImago Journal Rank (SJR) 5.4 2019 CiteScore Tweets by JournalWao PlumX Metrics Check out our current top social media articles via PlumX Metrics to see how readers are engaging with recently published research. Home Articles & Issues Current Issue List of Issues For Authors About Open Access Author Information Researcher Academy Submit Your Manuscript Journal Info About the Journal Contact Us Editorial Board New Content Alerts Submit Your Manuscript World Allergy Organization We use cookies to help provide and enhance our service and tailor content. To update your cookie settings, please visit the Cookie Preference Center for this site. Copyright © 2021 Elsevier Inc. except certain content provided by third parties. The content on this site is intended for healthcare professionals. Privacy Policy   Terms and Conditions   Accessibility   Help & Contact 
work_cp3ptydqtnh3vgmfprtemfs5u4	----	Microsoft Word - psg chapter_internet.doc A. Fred, J. Filipe, M. Partinen, T.Paiva, "PSG-Expert: An Expert System for the Diagnosis of Sleep Disorders". In European Neurological Network, T. Paiva and T. Penzel (Eds), IOS Press, no 78, series Studies in Health Technology and Informatics, pp 127-147, 2000. PSG-EXPERT – AN EXPERT SYSTEM FOR THE DIAGNOSIS OF SLEEP DISORDERS Ana FRED Instituto de Telecomunicações / Instituto Superior Técnico Av. Rovisco Pais, 1049-001 Lisboa, Portugal Email: afred@lx.it.pt Joaquim FILIPE Escola Superior de Tecnologia / Instituto Politécnico de Setúbal R. Vale de Chaves, Estefanilha, 2910 Setúbal, Portugal Markku PARTINEN Sleep Disorders Clinic and Research Center Helsinki, Finland Teresa PAIVA Centro de Estudos Egas Moniz / Hspital de Santa Maria Lisboa, Portugal Abstract. This paper describes PSG-EXPERT, an expert system in the domain of sleep disorders exploring polysomnographic data. The developed software tool is addressed from two points of view: (1)- as an integrated environment for the development of diagnosis-oriented expert systems; (2)- as an auxiliary diagnosis tool in the particular domain of sleep disorders. Developed over a Windows platform, this software tool extends one of the most popular shells – CLIPS (C Language Integrated Production System) with the following features: backward chaining engine; graph-based explanation facilities; knowledge editor including a fuzzy fact editor and a rules editor, with facts-rules integrity checking; belief revision mechanism; built-in case generator and validation module. It therefore provides graphical support for knowledge acquisition, edition, explanation and validation. From an application domain point of view, PSG-Expert is an auxiliary diagnosis system for sleep disorders based on polysomnographic data, that aims at assisting the medical expert in his diagnosis task by providing automatic analysis of polysomnographic data, summarising the results of this analysis in terms of a report of major findings and possible diagnosis consistent with the polysomnographic data. Sleep disorders classification follows the International Classification of Sleep Disorders. Major features of the system include: browsing on patients data records; structured navigation on Sleep Disorders descriptions according to ASDA definitions; internet links to related pages; diagnosis consistent with polysomnographic data; graphical user-interface including graph-based explanatory facilities; uncertainty modelling and belief revision; production of reports; connection to remote databases. 1. INTRODUCTION The problem of medical diagnosis can be stated as follows: given a set of symptoms (clinical data) and signals, or test results (tests performed on the patient), appraise pathological situations, identifying which diseases justify the particular findings. A variety of computer assisted techniques have been proposed to help solving the diagnostic problem [1,7,12,15]. Commonly known as expert systems, approaches, undertaking distinct formalisms for modeling domain knowledge, include: rule-based or production systems, frame-based systems, semantic networks, neural networks, and Bayesian belief networks. Explanatory mechanisms are more easily implemented with rule-based systems and graphical models making these preferred solutions to the diagnostic problem. Since the pioneering work by Feigenbaum (with the mass spectrogram interpreter DENDRAL, 1971), and the MYCIN project (a computer system that diagnosis bacterial infections of the blood and prescribes suitable drug therapy) by Shortliffe in 1976, the theory of expert systems has progressively evolved, accompanied by the appearance of expert systems development tools. This paper describes PSG-EXPERT, an expert system in the domain of sleep disorders. This auxiliary diagnosis system aims at assisting the medical expert in his diagnosis task by providing automatic analysis of polysomnographic data, summarising the results of this analysis in terms a report of major findings and possible diagnosis consistent with the polysomnographic data. Sleep disorders classification follows the International Classification of Sleep Disorders - ICSD [5]. The development of knowledge based systems to be used as ancillary diagnosis systems, usually undertakes a software engineering methodology based on the spiral model. Using this paradigm a series of increasingly refined prototypes are produced until an end-user version is finalised and delivered. Each development cycle comprises three main phases, namely: knowledge acquisition and refinement; software implementation; and system validation. Although the first two phases are of major importance, validation plays a decisive role in systems quality and acceptance, having also a great impact in the development time. Yet, many development tools or methodologies [2] fail to give much attention to this phase or completely overlook it. The first part of the paper (section 3) presents essentially the technical aspects of the PSG-EXPERT, as an integrated environment, which provides an adequate graphical support for every phase in the expert system development cycle, from knowledge acquisition and edition to reasoning explanation and knowledge-base validation, with emphasis on the last phase. The PSG-EXPERT environment constitutes a specialization of a general Artificial Intelligence technology for our application domain. Section 4 addresses the tool from an application domain perspective, as an auxiliary diagnosis environment for sleep disorders. Adopting a users point of view, the versatility and distinctive features of the system are illustrated through a set of interaction examples. 2. SYSTEM ARCHITECTURE AND FUNCTIONALITY Many tools (shells) have been proposed to assist the development team of expert systems in the design and implementation phase, in order to achieve fast prototyping. These shells are often text-based, with the exception of very expensive tools [8] [13][14]. Some of the more flexible and expensive tools were developed for dedicated environments such as Lisp-machines, although nowadays these machines have almost disappeared, due to their specificity and high cost [1]. Therefore, the AI community has been watching with some degree of expectation, the development of an academic tool, called CLIPS [4] that was created at the NASA’s Johnson Space Centre. CLIPS has the advantage of being C-based, instead of LISP-based and it is a very low-cost expert system development tool. However, there are some difficulties, concerning mainly two problems: the lack of a backward- chaining module (CLIPS has a forward-chaining inference engine based on the well-known OPS5 architecture) and the lack of an explanation facility (either text-based or graphic- based). Besides their particular difficulties and/or limitations as discussed above, existing tools neglect the important phase of system validation. Some actually provide incipient rule integrity checking mechanisms, but none seems to address performance evaluation issues. On the contrary, PSG-EXPERT includes a dedicated software module that has been developed in order to facilitate validation in a distributed environment, where several experts ought to be consulted at different physical locations. This software module is inspired in modern organisational workflow and groupware methodologies [3]. Several levels of validation can be defined: overall system evaluation concerning functionality, robustness, user-friendliness; knowledge integrity and correctness; diagnosis performance. All these levels are approached by the aforementioned software module, hereafter designated by CASE-BUILDER. PSG-EXPERT is based on and extends the CLIPS expert system development tool with a set of features which fill in a gap that hampered the development of CLIPS-based expert systems since its creation, in 1984. Besides having a backward chaining mode and an explanation facility, which facilitate the development of diagnosis-oriented applications, PSG-EXPERT completely encapsulates CLIPS, under a Windows-based graphical user interface (GUI). The new tool includes graph-based graphical explanation facilities; a knowledge editor including a fuzzy fact editor and facts-rules integrity checking; a belief revision mechanism; a built-in case generator and a validation module. The graphical interface was developed in Delphi, a Windows rapid application development (RAD) tool. The interface connects to CLIPS using pipes for data transfer and semaphores for synchronisation. Figure 1 summarises the system architecture and its main features. It comprises three main conceptual modules, namely: knowledge acquisition, diagnosis and validation. Each conceptual module corresponds to a different software product that can be used Figure 1. The PSG-EXPERT architecture. Acquisition Knowledge Editor Inference Extended-CLIPS inference engine Explanation Graph-based reasoning visualisation Validation Case Building and Assessment Knowledge Base Diagnosis independently of the other two. The knowledge acquisition facility is implemented through a knowledge editor which enables the easy introduction and modification of knowledge, including facts, rules and criteria, as described in section 2.2. The diagnosis module includes the knowledge base, the inference engine and a graphic-based explanatory facility. Easy case building and assessment is possible through the validation module. The underlying knowledge representation scheme is described next. 2.1 Knowledge Representation PSG-EXPERT is a rule-based expert system development tool. Besides the typical if- then rules, PSG-EXPERT has a special construct – the Criterion – that represents an expression of the form “if at least n conditions are satisfied then...”. A criterion can actually be rewritten as a set of if-then rules, but the abbreviation has been considered very useful by the domain experts because it matches the way their knowledge is commonly expressed. The structure of a criterion is shown in figure 2, together with the set of if-then rules it replaces. Criterion If at least 2 of the premises P1, P2, P3 Then conclusion C1 If P1 and P2 then C1 If P1 and P3 then C1 If P2 and P3 then C1 Figure 2. A Criterion in PSG-EXPERT is a new knowledge structure that replaces a set of if-then rules. Similarly to CLIPS, PSG-EXPERT has an uncertainty management mechanism based upon certainty factors manipulation [1]. With respect to criteria, special combination formulas had to be used in order to process certainty factors, so that the resulting value is the same as if the set of equivalent if-then rules were used. Any rule entered by the knowledge engineer can be used in forward and backward chaining, unlike some expert system development tools, such as GoldWorks [13], which require a specification of the reasoning direction in which the rules are going to be used. An extension to CLIPS was developed in order to support backward chaining. In diagnosis it is often necessary to combine rules in favour of a certain conclusion with others which are against it. Moreover, there may be rules that completely preclude a diagnosis in spite of a lot of evidence in favour. This diagnosis-specific framework is provided in PSG-EXPERT by means of a built-in rule that combines the facts in favour and against a given conclusion, therefore establishing a model structure, and a task structure. This gives a certain guidance for knowledge acquisition and representation, unlike generic expert system development tools which leave all the structuring decisions to the expert. This model-based development methodology is characteristic of 2nd generation expert systems [2]. 2.1.1 Facts Facts in PSG-EXPERT actually represent the grounded beliefs contained in the knowledge base, being either crisp or fuzzy, and follow from the CLIPS fact representation scheme. A fact can be asserted in the knowledge base as a simple fact or as a structured fact. Structured facts are similar to frames in other representation languages. The constituents of structured facts are slots, with a name and a value. A default value may be associated with each slot. Similarly to the frame paradigm there is an inheritance mechanism, which propagates default values defined in the frame structures to their instances. In order to accommodate the ”conditions in favour” and “conditions precluding” philosophy, the system embodies a special type of structured facts: final diagnoses (disorders). From the user point of view these facts have the form disorder <name> is <value> with <value> being one of the following: • likely – meaning that some condition in favour of a given diagnosis was satisfied; • precluded – there are some conditions met that preclude the given diagnosis; • present – the result of combining likely and precluded intermediate results by an internal rule. 2.2 The Knowledge Acquisition Facility The system provides a user-friendly interface for the introduction, revision and maintenance of the knowledge-base which enables the application of adequate knowledge elicitation principles [6]. A structured environment for the edition of facts (crisp and fuzzy), rules and criteria was designed that performs basic rules-facts integrity checking. Figure 3 illustrates this module by showing a typical screen for the edition of fuzzy variables. Figure 3. The Knowledge Editor main screen. 2.3 The Diagnosis Task Diagnosis is a knowledge-intensive task. It requires both a knowledge-base including all relevant facts and rules, as well as an inference machine capable of logical reasoning and an adequate human-machine interface in order to justify the derived conclusions [1,12]. 2.3.1 The Inference Engine As indicated previously, PSG-EXPERT extends the forward engine provided by CLIPS and supports backward chaining. This allows a goal-directed type of reasoning which is often necessary for diagnosis purposes. The backward chaining engine has been implemented as a modification of the basic forward chaining engine, adding the capability to manipulate a special structured fact called GOAL and a set of special facts called RULES that are actually the translation of forward chaining rules into a special structured fact syntax, to be used by the backward chaining engine. The translation is internal and transparent to the knowledge engineer. 2.3.2 Dealing with Missing Data There are two ways of handling missing data under the PSG-EXPERT environment, that must be selected before triggering the diagnosis mechanism: (i)- assume normality; (ii)- ask the user. When the first option is selected, default values are set for each un-instantiated fact, as defined through the knowledge editor. While reasoning according to the “ask user” option, only those facts whose knowledge is essential to prove/disprove a given conclusion are prompted to the user through an adequate interface that enables questioning the system why that piece of information is being asked. 2.3.3 The Explanation Facility PSG-EXPERT provides a graphical interface for explaining the reasoning process followed during either backward chaining or forward chaining. If, during the backward chaining process, the user is asked to provide the value of some variable he/she can ask the traditional “why” question, which the system will answer showing the rule which it is trying to prove. After both backward chaining or forward chaining the user can ask “how” the conclusions were deduced [6]. In that case a reasoning graph is presented. Given the diagnosis orientation of this tool the reasoning graph produced by the explanation mechanism has some special nodes that represent diagnoses (or disorders – in the medical domain). The criteria used in the reasoning are not expanded into a set of if- then rules in the explanation graph because criteria reflect the way experts actually reason. Therefore, to differentiate criteria from if-then rules, different symbols are used in the graph. Facts are also shown explicitly in the reasoning chain using different symbols. Lines connecting all these elements show the conceptual links that have been established during the reasoning performed by the expert system. The user can then, not only visualise the conceptual graph representing the reasoning process followed by the expert system during the diagnosis, but also inspect the details of the different elements of the graph (facts, rules, criteria and final diagnoses). 2.4 The Validation Facility Validation is probably one of the most difficult parts of an expert system development life-cycle [2]. PSG-EXPERT includes a tool, called CASE-BUILDER, to facilitate the system validation by domain experts. This is mainly a set of organised screens through which the experts can provide all the relevant data – filling in the values of available variables – in order to describe a case or situation. Additionally, the same experts are required to indicate a diagnosis for each case they have described. Case data or sample problems are critical to the evaluation of a knowledge-based system. Test cases provide the major vehicle for testing and demonstrating the capabilities of a knowledge-based system. Unlike tracing, which is directed toward determining how the system arrived at its conclusions, test cases are directed toward determining whether the system can reach the proper conclusions. Testing consists of feeding data into the system and comparing the system conclusions with what was expected. These data can be of three types: (i) very specific cases which are required to test that specific situations are handled correctly. These data can be “manufactured”; (ii) other cases are required to permit the evaluation of the application’s handling of typical or common situations. Such data could be collected from real problems presented to human experts for solution; (iii) still other cases might be generated somewhat randomly to represent the breadth of situations the application may have to handle in the future. The apparently best source of cases is historical data. This ensures that cases are realistic, representative of real problems and they are not contrived. However, using historical data often has three significant disadvantages: (i) the data are rarely complete; (ii) the historical solution might not have been adequate – e.g. not analysed by an expert; (iii) historical data are frequently on paper rather than in electronic form. Generating new cases from scratch, directly into electronic format, has the advantage that data can be created fairly inexpensively and can be tailored to test particular components of an application. This approach may however have two serious drawbacks: (i) solutions must be created for each test case generated. The expert must indicate the solution – the diagnosis can not be “generated” as the input data for the case was. Furthermore, (ii) since these cases can be “unrealistic”, the expert may not be eager to invest the time and personal effort to develop the solutions. The conclusion is that, unfortunately, generating case data is much easier than generating the solutions that go with those data. Therefore the case-data problem should be viewed as the case-solution problem; the ability to generate solutions and not data is what tends to limit the case data that can be developed for an application. CASE-BUILDER provides the advantages of case generation described above, minimising the two main problems that have been pointed out. Firstly, this software tool provides not only the means for data input but also an easy and fast way to identify both the solution and the respective justification. It also allows experts to provide advice “on the side” which can later on be used by the development team to refine the system knowledge base. Secondly, since the data is developed by an expert, it seldom is “unrealistic”. CASE-BUILDER provides a mechanism that allows multi-expert validation of a knowledge system. The first step in this process consists of an expert entering the data and providing his/her own diagnosis. Then, using CASE-BUILDER mechanisms for case communication, the case that was simulated by the first expert circulates amongst a group of experts who can provide their own evaluation of the case but without each of them being able to see the others’ diagnoses. All the communication is made using a special e-mail utility embedded in CASE-BUILDER. Sending cases only requires a button click. After all the experts have analysed a case it then is sent to the development team, who can use the case solutions to assess the accuracy of the system and to improve it if possible, by adopting some of the experts’ recommendations. Figure 4. Data entry screen in Case-Builder. Although this tool is specific for supporting the validation of expert systems made for diagnosis purposes, it is not specific of the medical domain. All examples given here are though extracted from an application on the medical domain. The list of variables presented in CASE-BUILDER, and their respective possible values, is easily configurable through a database, avoiding the need for modification and recompilation of the program every time a new variable is added or removed. Figure 4 shows the data entry screen. Each case must be assigned one or more diagnosis consistent with the data (see figure 5). Introduction of the diagnosis is made by clicking on the Disorders icon on the left column and selecting one of the available diagnosis, on the right side of the screen. There are also dedicated spaces where the expert can enter free text to explain the diagnosis (Justifications) and to include additional comments considered of interest (Observations) useful for the knowledge engineers further develop the knowledge base. Furthermore, cases thus entered are saved in a local database, being immediately accessible in the diagnosis module. The expert can hence simulate particular situations to be evaluated by the system and then analyse its reasoning in these case-studies, a process particularly useful and efficient for knowledge evaluation and revision. This provides a means for evaluating diagnosis consistency both in between experts and by the system. Figure 5. Diagnosis entry screen in Case-Builder. 3. DIAGNOSING WITH PSG-EXPERT: THE USER’S PERSPECTIVE The previously described specialized environment was developed to be used as the technical basis for developing an expert system for the diagnosis of sleep disorders using mainly polysomnographic and clinical data [9,10,11]. Figure 6: Inputs and outputs of the PSG Expert. Figure 6 gives a schematic description of the expert system environment. Biological signals are automatically scored by a core of signal processing algorithms. The resulting polysomnographic data and additional clinical data are stored in a neurological database. Given a particular recording, the system identifies the set of possible sleep disorders consistent with the data and produces a report similar to what is produced nowadays by human experts. The major features of the system are as follows: • Browsing of patients information and viewing of polysomnographic data records, from either local or remote databases (see fig. 7). • Structured navigation on Sleep Disorders descriptions according to ASDA definitions. User Signal Processing Data Base Report EEG EOG ... EMG TIB SL TST REML SPT SWSL …... • Internet link to related html pages. • Diagnosis of sleep disorders consistent with polysomnographic data, supporting: • Fuzzy rule-based reasoning operating both in data-driven (forward chaining) and goal-driven (backward chaining) modes and including a belief revision mechanism • HOW and WHY explanation of the reasoning process • Graph-based graphical explanation facilities • Handling of missing data and uncertainty • Production of a report summarising the major findings on data records. • Built-in CaseBuilder facilities for the introduction of new patient records into the database or to data exchange and system validation through transparent email communications. The final version of this expert system was submitted to a validation phase, using the CASE-BUILDER tool described in the previous section. 3.1 Polysomnographic Signals and Polysomnographic Data Biosignals most frequently used to describe human sleep are: the electrooculogram (EOG) - used to monitor eye movements; the electroencephalogram (EEG), two or more channels, from which brain electrical activity is assessed; the electromyogram (EMG) - chin and legs - used to detect limb movements and muscle tonus; respiration related signals, such as the respiratory flux and effort, O2 saturation, or snoring; and the electrocardiogram (ECG). Analysis of these signals enables the discrimination between the three distinct physiological states: wakefulness, rapid-eye-movement sleep (REM) and slow-wave or non-REM sleep. The global features of sleep dynamics, i.e. sleep macrostruture, have been summarized in a graphical representation, known as the hypnogram, of the evolution of sleep along the night as a sequence of sleep stages identified according to standard criteria established by Rechstschaffen and Kales. From this data global parameters characterizing sleep quality, sleep structure and cyclicity have been defined. It is recognized, however, that reliable diagnosis cannot be achieved based only on this type of global and limited information. Within each stage the EEG exhibits a rich variety of patterns of transient and background activity, that compose the sleep EEG microstructure. Parameters derived from EEG spectral analysis and time domain analysis have therefore been included as important information to consider for diagnosis purposes. Non EEG activity derived parameters complete the set of variables obtained from the processing of the above polysomnographic signals used for diagnosis of sleep disturbances. The types of data incorporated by the PSG-EXPERT are variables are organized into the following categories: clinical history; hypnogram data; sleep parameters; spectral data; EEG time related activity; non EEG activity. Figure 4 showed a typical screen exhibiting the structured presentation of the above types of information, presented in the phase of data entry through the CASE-BUILDER sub-module. Figure 7 depicts another example concerning the visualization of a patient’s record. Figure 7: Screen for visualization of patient data concerning respiratory parameters. 3.2 Application Case-Study The case-study concerns the analysis of a patient case which data are partly in a database. Two diagnostic strategies are possible: testing a specific disorder (hypothesis) – backward chaining reasoning - or finding all possible diagnosis consistent with the given– forward chaining reasoning. When forward chaining is performed no additional data is used; however, in the case of backward chaining reasoning, the system may ask for some additional data, directly correlated with the diagnostic hypothesis under analysis, which may be entered by the doctor, as needed. Figure 7 shows the explanation graph produced after a diagnosis. In this case the system performed a forward chaining reasoning and it identified a disorder (obstructive sleep apnea) with the certainty factor that is indicated on the upper-left area of the figure. Figure 5. Observing the diagnosis result and the reasoning process. The complete reasoning chain is shown in a scrollable window, on the right part of the screen. When the user clicks on a graph element, detailed information about it can be seen at the bottom left part of the screen. 5. CONCLUSIONS This paper presented technical aspects of PSG-EXPERT, an integrated environment for the development of diagnosis-oriented knowledge based systems applied to sleep disorders. This system provides graphical support for every phase of the development cycle, from knowledge acquisition and edition to reasoning explanation and knowledge-base validation. Using the Delphi language and encapsulating CLIPS, PSG-EXPERT is a Windows based software system that extends the later shell with the following features: backward chaining engine; graph-based explanation facilities; knowledge editor including fuzzy facts and rules editing with rules-facts integrity checking; belief revision mechanism; and a built- in case generator and validation module. In addition to the “if-then” structures, common to all rule-based systems, knowledge representation supports a new type of concept –the criterion – that enables a more compact and natural way of expressing rules of the form: “if at least n conditions are satisfied then...”. Furthermore, the system is designed with built-in structuring decision rules that permits the combination of evidence in favour and precluding a given diagnosis, thus establishing a model structure and a task structure. Unlike usual commercial or public domain shells, a key feature of this system is the validation module, an interactive subsystem that enables fast creation and interchange of commented cases that serve the purpose of diagnosis consistency and performance evaluations. This enhancement over traditional systems promotes the use of groupware techniques to expert system development and validation. Concerning its application as an auxiliary diagnosis system in the domain of sleep disorders, versatility of the system was partially illustrated through a set of interaction examples. Distinctive features of the system include: the exploration of objective polysomnographic data in the reasoning process, complementing and/or confirming clinical data; representation of expert knowledge in terms of natural language like diagnosis rules; a graphical interface exposing the reasoning process; distributed environment for data storage and retrieval; internet links to related pages; uncertainty modelling and belief revision; production of reports; a case building and validation module promoting groupware techniques; a global user-friendly interface supporting three main types of system usage – design and knowledge maintenance, diagnosing and validation. REFERENCES [1] Buchanan, B. and E. Shortliffe (Eds), Rule-Based Expert Systems, Addison-Wesley, 1984. [2] Burnet, E. and N. Scharenborg, The KADS method: Industrial Applications and Prospects for Expansion, in the 12th International Conference Avignon’92, 1992. [3] Brenner, W. Agents as Tools of the Information Society, in Intelligent Software Agents, Brenner, Zarnekow and Wittig (Eds), Springer-Verlag, 1998. [4] CLIPS Version 6.04, CLIPS Reference Manual, Software Technology Branch, Lyndon B. Johnson Space Centre, 1997. [5] Diagnostic Classification Steering Committee, Thorpy M. J. (Chairman), International Classification of Sleep Disorders: Diagnostic and Coding Manual (Revised), Rochester, Minnesota: American Sleep Disorders Association, 1997. [6] Diaper D.(Ed.), Knowledge Elicitation: Principles, Techniques and Applications, Ellis Horwood, Ltd, 1989. [7] Duda, R., and .P. E. Hart (Eds), Pattern Classification and Scene Analysis, John Wiley & Sons, 1973. [8] EXSYS Professional, User’s Guide, EXSYS Inc., 1994. [9] Filipe, J., Fred A., Paiva, T., Partinen M. A Second-Generation Expert System for the Diagnosis of Sleep Disorders using Polysomnographic Data, in Proc. Int’l Conf. on Information Systems Analysis and Synthesis, pp. 296-300, 1996. [10] Filipe, J., Fred A., Fernandes M. GDOS – A Graphical Diagnostic-Oriented Expert System Development Tool, in Proc. Int’l Conf. on Enterprise Information Systems, pp. 211-218, 1999. [11] Fred, A. and Filipe, J. GMED – A Case-Study of a Specialized Graphical Medical Diagnosis Shell Application, in Proceedinds of IASTED International Conference on Artificial Intelligence and Soft Computing,, 1999. [12] Giarratano, J. and G. Riley (Eds), Expert Systems: Principles and Programming, PWS Publishing Company, Boston, 1994. [13] GoldWorks, User’s Guide, Gold-Hill Inc., 1993. [14] KEE – The Knowledge Engineering Environment, Intellicorp, 1987. [15] Russel, S. and P. Norvig (Eds), Artificial Intelligence: A Modern Approach, Prentice-Hall, Inc., 1995. 
work_cq2kwqg6oneihh37tsxgs2xlnm	----	PII: 0166-218X(88)90005-4 Discrete Applied Mathematics 19 (1988) 45-63 North-Holland 45 GRAPH EMBEDDING IN SYNCHEM2, AN EXPERT SYSTEM FOR ORGANIC SYNTHESIS DISCOVERY * Joseph D. BENSTOCK** and Donald J. BERNDT Department of Computer Science, State University of New York at Stony Brook, Stony Brook, NY 11794, USA Krishna K. AGARWAL Department of Computing and Information Sciences, Trinity University, San Antonio, TX 78284, USA Received 15 October 1985 Revised 30 April 1986 Graph embedding (subgraph isomorphism) is an NP-complete problem of great theoretical and practical importance in the sciences, especially chemistry and computer science. This paper presents positive test results for techniques to speed embedding by modeling graphs with subroutines, precalculating edge tables, turning recursion into iteration, and using search- ordering heuristics. The expert system SYNCHEM~ searches for synthesis routes of organic molecules without the online guidance of a user, and this paper examines how embedding information helps to imple- ment the central operations of SYNCHEM.2: selection, application, and evaluation of chemical reac- tions. The paper also outlines the architecture of SYNCHEM~, analyzes the computational time complexity of embedding and related problems in graph isomorphism and canonical chemical naming, and suggests topics and techniques for further research. 1. Introduction Graph embedding algorithms determine whether a given guest graph can be embedded within a given host graph; some embedding algorithms in addition store the mapping(s) of the isomorphism(s) between the guest and subgraphs of the host. Such algorithms are of great theoretical and practical importance in a number of the sciences. We present positive test results for techniques to speed embedding by modeling graphs with subroutines, precalculating edge tables, turning recursion into iteration, and using search-ordering heuristics. As a central process in the expert System SYNCHEM~ at SUtiY/Stony Brook, an AI project under the direction of H. Gelernter, embedding assists in organic syn- thesis discovery, an important and difficult problem in the natural sciences for * This research was supported by Eastman Kodak Co. and NASA Research Grant #NAG 3-651. ** Present address: Computer Resources International A/S, Vesterbrogade lA, DK-1620 Copenhagen V, Denmark. 0166-218X/88/$3.50 0 1988, Elsevier Science Publishers B.V. (North-Holland) 46 J.D. Benstock et al. which SYNCHEM~ has generated results of interest to chemists and biochemists [7,19,20]. SYNCHEM searches for synthesis routes for a given organic molecule without the online guidance of a user, and uses embedding information to help select, apply, and evaluate chemical reactions. Various algorithms for substructure search, molecular pattern matching, pretransform and posttransform testing, and heuristic subgoal evaluation will be presented and discussed in the context of SYNCHEM2. Embedding is an NP-complete problem related to important problems in graph isomorphism and canonical chemical naming that currently have unknown com- putational time complexity and that bear upon the deep questions of containment in the polynomial hierarchy of problem complexity classes, including P and NP. We will analyze the time complexity of these problems and the relationships between them. Graph embedding is a central operation in all the following applications in- volving pattern matching and feature detection: canonical chemical nomen- clature and indexing, chemical substructure searching, and synthesis discovery [2,7, 10,33,36,39,42,43], aspects of developmental biology [ 141, and modeling memory recall in cognitive psychology [7,32]. Applications are found in electrical network design and optimization [43], image processing involving template match- ing and feature recognition [7,8, 11, 16, 17,34,35,37,38,40], and speech processing using templates and networks of phonemes [7]. Embedding techniques find wide- spread use in the following subfields of computer science: database management and information retrieval [6,7,8,11,30], program transformation (including syn- thesis, optimization, and automated programming) [2,5,7, lo], computer learning and knowledge representation [7,15], analogical reasoning using graph homo- morphism [12,22], state representation and pattern recognition in game trees [2,6, lo], as well as study of programming languages, data and control flow, and concurrency [9, 141. The paper consists of six sections: (1) Introduction, (2) Architecture of SYNCHEM& (3) Using embedding in SYNCHEM~, (4) Computational time complexity of embed- ding and related problems, (5) Techniques to speed embedding, (6) Summary and directions for further research. 2. Architecture of SYNCHEM Other introductions to the SYNCHEM project are found in [7,19,20]. 2. I. Control strategy Problem reduction by retrosynthesis The problem reduction strategy SYNCHEM~ uses for finding synthesis rOUteS for an input organic molecule (called the tirrget molecule) is to select reactions that Graph embedding in SYNCHEM~ 47 retrosynthesize it into compounds that are simpler in the sense that they are com- mercially available or are expected to be more readily synthesizable. These generated compounds are referred to as subgoal conjuncts in artificial intelligence and precur- sors in chemistry. Problem representation using AND/OR graphs The search space of the problem is represented by an AND/OR Problem Solving Graph (PSG) as in Fig. 1, in which a molecule is represented by a compound node, that when selected for development, spawns alternative subgoal nodes [7,3 11. Com- pound nodes and subgoal nodes alternate in the PSG and are connected by reaction edges. Subgoal nodes (whose children are one or more reacting compounds) are by definition ‘OR’ nodes because the establishment of any subgoal node solves the parent compound. Compound nodes, on the other hand, are ‘AND' nodes because all sibling compound nodes must be established to solve their parent subgoal. Available compounds are ‘terminal’ nodes in the PSG, and normally are not further developed. A cycle of subgoal generation is completed when every feasible reaction schema from the knowledge base that meets the dynamic selection criteria has been applied to a particular compound. The local tactic of SYNCHEM~ is to progressively reduce the ‘distance’ between leaf nodes of the PSG and a list of available com- pounds by selecting and applying reactions until all leaf nodes represent available compounds. SG #I SG #2 SG #3 SG #4 I CvC C A I CvC C 0 0 0 O-H O-H 0 I I C - C C c=o I 0 0 C - c jc C C \\ \ c Cl O-H O-H c=c C Fig. I. A Problem Solving Graph. Adapted from [24]. The target compound at the root is developed into three ‘OR’ subgoal nodes, SC #2, 3, and 4. Subgoal nodes have one or more reacting ‘AND' com- pound nodes as children. The goal of SYNCHEM~ is to develop all leaf nodes into available compounds. Note that chemical convention omits depiction of hydrogen atoms on carbons when they clutter draw- ings; thus all carbons have four edges, although in drawings some edges may be implicit. 48 J.D. Benstock et al. Heuristic evaluation functions An exhaustive search of the problem space is computationally infeasible (‘com- binatorially explosive’) for any but the most simple syntheses. As the search gener- ates and evaluates subgoals, it must correspondingly limit the growth of the PSG and prune it. Thus the search is guided by figures of merit, which are heuristically computed values used to predict the success of program alternatives. The three types of merit, compound, reaction, and subgoal, are integers in the range [0, . . . , 1001. Merits are computed using information about the current chemical environment of a given node in the PSG to adjust the default values from the knowledge bases [20,24]. Compound merits are used to predict the likelihood of successfully completing a pathway from a particular compound all the way down to available compounds and reflect the maximum subgoal merit of subgoal children. (Available compounds are assigned a merit of 100 on the scale [O, . . . , 1001.) Reaction merits, working in the opposite direction, estimate the cost of successfully reacting compounds to solve their parent subgoal. Reaction merits are a function of the dynamic parameters ease, yield, and confidence, which are also rated [0, . . . , 1001. The default parameter values associated with each reaction are estimates of usefulness under standard con- ditions, supplied by chemists who have worked on the knowledge bases. The subgoal merit of each subgoal is a dynamic evaluation function of its constituent child compound merits, its reaction merit, and after updating, the subgoal merits of descendants. After each cycle of subgoal generation, affected subgoal and com- pound merits throughout the PSG are updated by recomputing them iteratively, starting from newly generated subgoals and moving up the PSG to the target com- pound at the top. The merit of a subgoal is generally somewhat worse than the merit of its worst unsolved child compound. Effort distribution The Effort Distribution algorithm dynamically apportions search time for each compound node so that the program can develop the search space rationally. Search status flags for compound nodes (open, solved, just solved, stuck, available, un- solvable) are updated and distributed throughout the PSG after each cycle of subgoal generation [24]. SYNCHEM~ has the capability of allowing the user to control the degree to which the search concentrates on syntheses that are either inexpensive, short, diverse, different from those in the literature, or have well-understood chemistry, etc. 2.2. Knowledge bases The knowledge bases consist of four available-compounds libraries (two of which are currently active), and four reaction libraries. Graph embedding in SYNCHEM~ 49 Compound representation and libraries Davis, Boivie, and Agarwal have designed SLINGS (SYNCHEM Linear Input Graphs) for SYNCHEM~ [ 131, a compact linear canonical representation for molecules, for user input/output as well as internal storage and communication be- tween program modules. Compounds are also represented by the Extended Topo- logical Representation (XTR), a data structure that includes a connection matrix and that computes and stores additional information on demand, such as atomic degree, number of cycles, and so on. SYNCHEM~ has a knowledge base of about 10,000 available compounds, most of which are from the Aldrich Catalog-Handbook of Organic and Biochemicals. Reaction representation and libraries SYNCHEM~ selects reactions from about 1200 reactions in its knowledge base by determining the attributes of a compound and using these attributes as keys to reac- tions in the knowledge base. Attributes are characteristic chemical features, func- tional groups, or patterns of infeasible structures. The main reaction library is divided into 30 currently active chapters; each chapter is associated with a unique attribute and a chapter screening test. A reaction with more than one attribute is arbitrarily assigned to a chapter. Reactions within chapters are arranged in order of decreasing probable utility. Reactions are stored in the reaction library as reaction schemata, which may be viewed as graph rewriting rules. A reaction schema consists of a goalpattern graph and its retrosynthetic derivative, the subgoalpattern graph, which have isomorphic node sets. The goal pattern or the subgoal pattern may be disconnected. To deter- mine whether a reaction may be applied to a compound, i.e., a reaction schema ap- guest Y H ?-C-Cl + H-Cl ?-b-H + Cl-Cl I!l k!l HH HH H-&k-Cl + H-Cl H-&-H + Cl-Cl IG ifi host Fig. 2. A reaction schema and application. The guest graph in the reaction schema on top embeds in the host graph, showing that the reaction schema can be applied to the host graph, as on the bottom. The left and right sides of the reaction schema are called the goal pattern and subgoal pattern respectively. The fragment variable nodes labeled ‘?’ correspond to subgraphs, called fragments, in the host and its retrosynthetic derivative. 50 J.D. Benstock et al. plied to a host graph, SYNCHEM~ considers the goal pattern graph as the guest graph and searches for embeddings of it within the host graph. Fig. 2 shows a guest graph that embeds in a host graph, which indicates that the reaction schema can be applied to the host graph. 3. Using embedding in SYNCHEM~ 3.1. Definition of chemical graph embedding Because a goal pattern graph stands for a class of compounds, it is convenient to use fragment variable nodes to stand for classes of compound substructures. Definition. A fragment variable node (or variable node, for short) is a guest graph node that, according to its label, maps to exactly one of the following attribute classes: any element at all, any element except hydrogen, any halogen atom, or an alkyl carbon. Similarly, a variable edge maps to either a single, double, or resonant bond. Definition. A constant node is the opposite of a variable node and is labeled by a definite element. Definition. A chemical graph embedding of a guest graph in a host graph is a l-to-l mapping of guest nodes into host nodes such that: (1) adjacency is preserved, (2) node and edge labels are preserved, (3) stereochemical orientation is preserved, and (4) images of variable nodes are part of connected subgraphs in the host correspond- ing to their label and containing no images of constant nodes [39]. These subgraphs in the host and its retrosynthetic derivative are called fragments and are an instance of the attribute class the labeled variable node represents. In- tuitively, fragments are what is ‘left over’ in the host graph after images of constant nodes are accounted for. See Fig. 3. \ ,c=c’ $1 $2 H y H Fig. 3. Variable nodes and fragments. From [39]. The guest graph embeds in the host graph. The hydrogen and carbon atoms in the guest map to the obvious atoms in the host; the two variable nodes labeled ‘$’ both correspond to the same fragment in the host graph, an eight node subgraph constituting the rest of the host. ‘$N’, where N is an integer, is represented as R, in chemistry literature, a molecular structure variable. Graph embedding in SYNCHEM~ 51 3.2. Selecting reactions Attribute determination In order for the program to know which types of reactions to select for a par- ticular compound, it must first determine the attributes of that compound. It uses SUBSRCH, a procedure designed by Boivie, for finding all attributes within a com- pound from a set of predefined substructures. SUBSRCH operates by conducting a binary search of the attribute table and comparing table entries to neighborhoods recursively ‘grown’ around each compound atom. SUBSRCH operates by examining the neighborhood of each atom in a compound, using breadth-first search (BFS, see Section 5.3), and identifying each neighborhood as: (1) part of a potential attribute warranting further search, (2) an instantiation of an attribute, or (3) a substructure not warranting further search, unlike any in the table (in which case SUBSRCH backtracks). The SUBSRCH procedure is more effi- cient than general embedding algorithms for locating sets of predefined substruc- tures and identifying attributes that are: (1) sensitive and that interfere with reactions, (2) readily available or readily synthesizable, or (3) chemically invalid 1101. Screening reactions Once SUBSRCH has determined the attributes of a compound, SYNCHEM conducts screening pretransform tests on them to select reactions from the reaction library. Pretransform tests for a reaction use chemist-specified combinations of conditions of the following four types: the compound (can’t/must) have (any/all) of the at- tributes on a list associated with the reaction. An example of a pretransform test would be that a Grignard reaction on lactones or aldehydes can’t have any carboxyl or nitrile groups, and must have any secondary or tertiary alcohol group. 3.3. Applying reactions If all the pretransform tests for a reaction on a compound are satisfied then SYNCHEM~ attempts the following embedding using Sanders’s MATCH algorithm: it takes the goal pattern graph of the reaction schema as the guest graph, and the graph representing the compound as the host graph, and generates maps for all embeddings of the guest within the host [39]. The difference between SUBSRCH and MATCH is that, while SUBSRCH locates a ‘root’ node for each instance of an attribute within the host, MATCH determines the exact location and scope of the guest graph within the host graph, generating a l-to-l map from nodes of the guest graph to nodes of the host graph. Because a reaction can take place at several sites in a molecule, but sometimes such chemistry is redundant, MATCH has the option of generating only one embedding on any given procedure call. MATCH is called recursively with a guest graph and a host graph and the most recently matched pair of nodes. It then finds all possible extensions of existing par- 52 J.D. Benstock et al. tial embeddings using a dual depth-first search (DDFS), and returns the maps of all complete embeddings. The pseudocode for the MATCH algorithm is given below. Note that a frond or back edge is an edge that completes a cycle in a graph traversal 1411. The Dual Depth-First Search MATCH Algorithm. Node a in the guest already matches node a in the host and MATCH attempts to extend the embedding to nodes b and p. procedure MATCH (a, a: nodes); begin for each unmarked neighbor b of a do begin mark b; for each unmarked neighbor /3 of a do begin if (node labels of b and /3 match and fronds of b and ,8 match and edge labels of edges (a, b) and (a, /I) match) then begin mark /3; image(b) : = p; if (all nodes of guest graph are marked) then record embedding eke Cd MATCH(b, p); unmark /I; image(b) : = 0; end; (if nodes, fronds, and edges match} end; {for each unmarked neighbor of a do} unmark b; end; {for each unmarked neighbor of a do} end; (procedure MATCH} If the guest graph of the reaction can be embedded within the host graph represent- ing the compound, then SYNCHEM~ applies the reaction to the host graph using the procedure SUBGENR, producing a subgoal graph consisting of one or more com- pounds, as in Fig. 2. 3.4. Evaluating reactions After applying a reaction, SYNCHEM~ canonically names the generated compounds and conducts a series of posttransform tests, using SUBSRCH and MATCH, to evaluate the application to the particular compound in the chemical environment existing at that node in the PSG. The default values for the ease, yield, and confidence parameters of the reaction are adjusted at this stage and combined to calculate the adjusted reaction merit. Posttransform tests can cause a subgoal to be rejected (e.g., Graph embedding in SYNCHEM~ 53 due to an invalid attribute having excessive bond strain), the reaction procedure to be modified (e.g., changing the reagent), or protection procedures to be specified for sensitive groups. In each case, the specific site of the embedding and the current chemical environment are considered. Posttransform tests may be concerned with, for example, electronic and stereochemical effects, steric hindrance, proximity of functional groups to the reac- tion site, ring structures and sizes, alkynes in a ring, allenes in a small ring, relative stabilities of carbonium ions, leaving groups, migratory aptitudes, the largest at- tribute in the molecule, symmetry properties of the goal and subgoal molecules, etc. 1241. After posttransform testing, compound merits of newly generated compounds are evaluated using several criteria, calling SUBSRCH to identify significant attributes and heuristic procedures to estimate compound merit. Compound merit is computed in a procedure that considers the number and bonding of carbon atoms, the ratio of number of functional groups to number of carbon atoms, ring system complexities, and metastable oxidation states. In general, simpler compounds are rated pre- ferentially. 4. Computational time complexity of embedding and related problems 4. I. The polynomial hierarchy A program for a deterministic Turing machine (DTM) consists of: (1) a finite state read/write head; (2) an infinite tape of cells; (3) a finite set Z of input tape symbols; (4) a finite set Q of states, including a start state q. and two halt states qy and qN, and (5) a transition function that directs the tape head in a given state at a given cell to read the symbol in the cell, to write a symbol over it (perhaps the same one), and to move either left or right. A program for a nondeterministic Turing machine (NTDM) has a transition function that directs the tape head in a given state at a given cell to read the cell, and to nondeterministically choose a pair, consisting of a symbol to write and a head movement, from a set. A nondeterministic program ‘guesses’ a solution of a decision problem by choosing the right such pair at each step, and then ‘verifies’ whether it is actually a solution. An example of this concept is the following informal description of an NDTM for graph isomorphism: it guesses a mapping between the vertex sets of two graphs, and then proceeds to verify whether the mapping is l-to-l, onto, and adjacency-preserving. An input string XEZ* is a finite sequence of symbols and we say (for a DTM or an NDTM) that program M accepts x iff it begins at the first tape cell, reads the string x, and halts in the accepting state qy. The language accepted or recognized by program M is L, = {x: XEZ* and M accepts x}. 54 J.D. Benstock et cd There is a natural correspondence between recognizing languages and solving deci- sion problems. A polynomial time program is one in which the number of computation steps is bounded by a polynomial function of the input length. The complexity class P is defined to be P = {L: 5’ a polynomial time DTM program recognizing L}. The class NP is defined to be NP = {L: 2 a polynomial time NTDM program recognizing L}. The class co-NP is defined to be co-NP = {L: 5’ a polynomial time NTDM program recognizing Z* - L}. The P = NP problem is open because, while verifying solutions can be done in polynomial time for many problems, it is not known whether guessing solutions can also be done in polynomial time for those problems. To rate the relative complexity of languages, it is useful to define a polynomial transformation between languages L, and L, as a function f computable by a recursive (always halting) polynomial time DTM program such that XE L, iff f(x) E L,. We then write L, I, L,, and say that L, many-one Turing reduces to L,. This expresses the relation that if the ‘harder’ language L, is in P, then so is L,. ‘I,’ is a partial order. L1 and L, are polynomial time equivalent iff L, &, L, and Lz<, L,. We then write L, =, L,. The class NP-complete (NPC) consists of the ‘hardest’ problems in NP: NPC={L: LENP, and VL’ENP,L’S,L}. A polynomial algorithm found for an NPC problem would be translatable to a polynomial algorithm for any NP problem, by direct construction using the reduc- ing polynomial transformation. It is felt that there is a strong possibility that NPC problems are intractable, with optimal solutions taking an exponential amount of time, although good approximate solutions may sometimes be found in polynomial time. We can show a given problem to be NPC by showing that it contains a known NPC problem as a subcase with a recursive (always halting) polynomial time transformation reducing the known problem to the given one. (Intuitively, if the new, harder, problem were polynomially solvable, then its known NPC subcase would be polynomially solvable as well). Ladner showed that if P # NP, then there exists a dense hierarchy of problems strictly between P and NPC [ 181. The class NP- Incomplete is defined to be NPI = NP - (P U NPC). Fig. 4 depicts the containments in the polynomial hierarchy under this assumption that P#NP, in which case NPI is nonempty. Accounts of Turing machines and complexity classes are found in [3,18,21]. Graph embedding in SYNCHEM 55 4.2. Time complexity of subgraph isomorphism Embedding is NP-complete for chemical and general graphs. Compounds can be modeled by graphs with labeled edges and nodes. Thus the problem of determining whether one compound is ‘part’ of another can be restated as that of determining, in a pair of labeled graphs, whether a graph G can be embedded in a graph G’, i.e., whether G is isomorphic to a subgraph of G’. This is the labeled variant of the Subgraph Isomorphism problem (SGI). To see that SGI is NPC, we use the standard technique of noting that it is NP (because it is possible to verify a solution in polynomial time), and that it contains a known NPC problem, CLIQUE, as a subcase. Intuitively, this shows that SGI is in the class NPC because it is ‘at least as hard’ as its subcase, known to be ‘at least as hard’ as any problem in NP. CLIQUE asks whether a graph contains a complete subgraph of size ?k. CLIQUE is a subcase of SGI because an algorithm for SGI can be iteratively called as a subroutine to attempt to embed the complete graphs of size ?k in CLIQUE’s input graph. That iteration takes polynomial time. We conclude that CLIQUES, SGI, and that SGI is NPC. A polynomial reduction from unlabeled graphs to graphs with only one label shows that SGII, labeled SGI. Thus the chemical application we are interested in is modeled by labeled SGI, which is NPC. 4.3. Relationship of SGI to graph isomorphism and chemical naming The Graph Isomorphism problem (GI) asks whether a graph G is isomorphic to a graph G’. (SGI asks whether there exists a 1-I map between the graphs, while GI asks whether there exists an onto l-l map between the graphs.) The isomorphisms of a graph onto itself are called automorphisms or symmetries. In synthesis discovery embedding, nontrivial automorphisms and interactions between goal pat- tern, subgoal pattern, and host graphs often generate redundant compounds and waste the substantial time required to canonically name, store, evaluate, and develop them. Agarwal discusses several heuristics for eliminating redundant maps (some before generation and some after) [2], but notes that the heuristics bear fur- ther unification and extension. Because heuristics to speed an algorithm can them- selves suffer from combinatorial explosion, applications must be time analyzed carefully. SGI and GI algorithms can be used as subroutines to help solve the chemical canonical naming problem, which is of central concern whenever dealing with a chemical database. The graph naming problem (GRAPHNAME) determines a canonical name for a compound. SYNCHEM~ uses the SLING representation and algorithms [13]. The Morgan algorithm is another popular and effective system [29]. 4.4. Time complexities of graph isomorphism and chemical naming The graph isomorphism problem is of considerable theoretical interest because, 56 J.D. Benstock et al. NP Fig. 4. The polynomial hierarchy of complexity classes, assuming Pf NP. From [18, p. 1541. although it models many fundamental problems, its exact complexity is not current- ly known. A polynomial algorithm for SGI could be easily modified to polynomially solve GI, so GI<, SGI, but GI has not been shown to be NPC. The problems Ladner showed to be NPI under the assumption P # NP are ‘artificially constructed’ diagonalized problems, but determination of graph isomorphism and integer primality are two naturally occurring candidates for membership in NPI. Further- more, since there is no known efficient test for determining non-isomorphism of graphs, only failure of isomorphism tests, it is not known whether GI is in co-NPC. Thus the complexity of GI bears on the deep questions of containment among the complexity classes P, NP, co-NP, and the rest of the polynomial hierarchy. Indications that GI may be in the class NPI are: (1) The failure of current NPC proof techniques to show the membership of GI in NPC, as opposed to the thousands of NP problems that have been shown to be in NPC. (2) The polynomial time equivalence of the existence and enumeration GI pro- blems, which is at variance with some NPC problems [18] (although the polynomial equivalence of existence and enumeration variants of many NPC problems is cur- rently unknown [23,44]). The enumeration variant of GI asks how many isomor- phisms exist between a pair of graphs. Graph isomorphism can be tested in polynomial time for restricted classes of graphs, such as trivalent or planar [18,25]. Luks showed in 1980 that isomorphism for graphs of bounded degree can also be tested in polynomial time [27]. An in- teresting question is whether bounding the degree of guest and/or host graphs makes SGI also solvable in polynomial time, and if it does, whether Luks’s 0(n6) algorithm could be made practical or improved for chemical graphs of the size cur- rent chemistry programs may encounter (about 50 vertices). Chemical graphs naturally have bounded degree, which is usually four (for elements such as carbon), and certainly less than eight. Luks’s technique exemplifies the group theoretic approach to GI, which seeks to determine a set of generating permutations for the automorphism group of a graph. The other main approach, exemplified by the work of Miller [28], is topological, and seeks to embed a graph onto a surface of minimal genus and then dissect the surface into planar components. The relationship between the two approaches is not known. There are a number of combinatorial and group-theoretic problems (some Graph embedding in SYNCHEM~ 51 concerning group generating sets) that are polynomial time equivalent to GI and we say that they constitute the class of isomorphism complete problems [25]. A polynomial time solution to GRAPHNAME would polynomially solve GI, but an interesting fact is that it is not currently known whether a polynomial time solu- tion to GI could be extended to polynomially solve GRAPHNAME [25]. Thus we know that GI I ,,, GRAPHNAME, but not whether GRAPHNAME I, GI. 5. Techniques to speed embedding 5.1. Modeling graphs with subroutines Unrolling loops and making fewer array references and procedure calls can significantly shorten execution time of a program. Benstock has developed a tech- nique to do this that models or mirrors each guest graph with a short subroutine in source code. A simple algorithm creates this code out of an adjacency represen- tation for a guest graph. It apportions a ‘for’ loop in source code for each node of a guest graph, so that some guest graph information is implicit in the body of that loop. The technique is similar to that of transforming the labeled graphs of transi- tion diagrams into code for lexical analyzers of compilers, as in [5]. There is a l-to-l map from a graph node to a ‘for’ loop that searches the node’s neighbors to extend the embedding. The following operations are saved: procedure calls and array references for node, edge, and frond matching are either eliminated (in the frequent situation that any label will match, or that no fronds exist) or reduced to references to constants. The pseudocode below for the mirroring technique shows the ‘for’ loop of one node with the ‘for’ loop of its successor (in DFS traversal) nested within it. Pseudocode for Mirroring Technique for each unmarked neighbor p of CY do begin if (node labels of b and /3 match and fronds of b and p match and edge labels of edges (a, b) and (a, /3) match) then begin mark /I; image(b) :=/3; a :=/I; for e-f unmark p; image(b) : = 0; CY : = image(a); end; {if nodes, fronds, and edges match} end; {for each unmarked neighbor of (x do} 58 J.D. Benstock et al. The mirroring technique reduced execution time of MATCH by a factor of 174 when attempting to embed the perester pattern graph of Fig. 6 in each of the five host graphs of Fig. 5. It reduced execution time by a factor of 35 when embedding a benzene ring in itself. The mirroring technique is suitable for stable databases re- quiring fast pattern matching and with storage enough for a subroutine correspond- ing to each pattern in the database. 5.2. Eliminating recursion and precalculating tables We have developed an embedding algorithm GREMBED (for Graph Embed) that turns the recursion of MATCH into iteration, and uses a precalculated edge table of the guest graph to control the search, instead of using direct DFS of the guest graph. It is efficient in the context of SYNCHEM~ to precalculate edge tables a single time rather than searching guest graphs anew each time (with marking, unmarking, querying, and backtracking operations) since there are only about 1200 reaction schemata, and edge tables of their guest graphs would be consulted repeatedly [l]. These modifications made GREMBED about two to three times faster than MATCH, in our tests which used the five host graphs shown in Fig. 5 and the five guest graphs shown in Fig. 6. We attempted embeddings, using MATCH and GREMBED, for each of the 25 possible pairs of host and guest graphs. The total execution times in milliseconds for finding all embeddings are shown below in Table 1, along with the MATCH/GREMBED ratios of those times. They are grouped by host and guest graph. The average ratio (speedup factor) is 2.6. The figures for finding just one embedding (existence) are shown in Table 2. The average ratio for this case is 3.3. Both algorithms were written in VAX/VMS PL/I and run on a MicroVAX-II. For simplicity, these tests did not consider stereochemistry and the heuristics for reduc- ing the generation of redundant graphs. y=c c~c-c-c-c=c-c=o H-O 0 C=O (1) (2) 0 0: E-o-o-c-c 0 $-O-H CJ (4) (5) Fig. 5. Five host graphs used to compare MATCH and GREMBED. Unlabeled vertices represent carbon atoms by convention. Hexagonal rings with a circle in the center represent benzene rings where each of the six edges is labeled as a resonant bond. A carbon in a benzene ring has room outside the ring for only one single bond. (1) arbitrary molecule, (2) 3-vinyl-2-hepten-6-ynal, (3) OO-ethyl hydrogen monoperox- phthalate, (4) 3-chlorophthalide, (5) furoyl. Graph embedding in SYNCHEM 59 (6) (7) (8) ??? ? I-2 I I? ? ? 0 I ; ? 0 ? ??? (9) (10) Fig. 6. Five guest graphs used to compare MATCH and GREMBED. (6) alkene, (7) lactone, (8) perester, (9) furan-3-carboxaldehyde, (10) cyclohexene. There are 8 and 32 embeddings of alkene and cyclohexene respectively in the first host graph, 16 embeddings of alkene in the second host, 6 embeddings of perester in the third host, no embeddings in the fourth host, and 16 and 2 embeddings of alkene and furan respectively in the fifth host. We also compared GREMBED and the canonical naming algorithm of SYNCHEM~ for testing graph isomorphism, and found GREMBED to be about 24 times faster on the average when testing each of the host graphs of Fig. 5 for isomorphism with itself. GREMBED was 70 times faster than canonical naming for the molecule below. Table 1. Times (in milliseconds) using MATCH and GREMBED to find all embeddings of the 25 host-guest pairs. Summed and grouped by host and guest graph, and showing the MATCH/GREMBED ratios of the times. Name MATCH GREMBED Ratio arbitrary molecule 3.vinyl-2-hepten-6-ynal OO-ethyl hydrogen monoperoxphthalate 3-chlorophthalide furoyl 1055 410 2.57 665 210 3.17 825 315 2.62 520 200 2.60 315 160 1.97 alkene 740 360 2.06 perester 540 160 3.38 lactone 465 205 2.27 furan 525 175 3.00 cyclohexene 1110 395 2.81 60 J.D. Benstock et al. Table 2. Times (in milliseconds) using MATCH and GREMBED to find the first embedding (i.e., determine existence of embedding) of the 25 host-guest pairs. Summed and grouped by host and guest graph, and showing the MATCH/GREMBED ratios of the times. Name MATCH GREMBED Ratio arbitrary molecule 3-vinyl-2-hepten-6-ynal OO-ethyl hydrogen monoperoxphthalate 3chlorophthalide furoyl 885 255 3.47 665 210 3.17 795 260 3.06 535 200 2.68 325 105 3.10 alkene 575 200 2.88 perester 525 165 3.18 lactone 480 176 2.74 furan 515 165 3.12 cyclohexene 1130 275 4.11 This indicates that GI algorithms are to be preferred to GRAPHNAME algorithms in applications in which approximately one hundred graphs or fewer are to be pair- wise compared and duplicates not named. 5.3. Search-ordering heuristics Comparisons between search-ordering heuristics should analyze, among other factors, the following graph parameters: the expected number of embeddings (so we can seek to primarily accept or reject matches), the cycle structure and general graph topology, and the absolute and relative sizes of the guest and host graphs. For the sake of efficiency, attempts should be made to match distinguishing structural features, such as attributes and fronds, as soon as possible in an embedding search. Time might be saved in SYNCHEM~ embedding searches if substructures in the guest graph were matched as units to their images in the host graph, using informa- tion already gathered by SUBSRCH. This attribute matching would also be suitable for applications with guest graphs that had many predefined substructures. MATCH is passed matched initial root nodes by SUBSRCH, and we have found it to be highly sensitive to the choice of pattern graph root node. We have found that choosing a guest root node within a distinctive region of the guest (allowing rapid rejection of nonembeddings) can speed embedding by a factor of as much as four. Breadth-first search (BFS) is the graph traversal method that visits the root, then all nodes at distance one from the root, then all nodes at distance two, etc., and never revisits nodes. Thus it crosses every edge exactly twice, counting backtracking upon finding an already visited node. Depth-first search (DFS) is the graph traversal method that visits the unvisited node of maximum depth at each step. BFS tends to waste storage space, since desired nodes within the graph lie too far from the root for most applications, and BFS also tends to manifest embedding failures more Graph embedding in SYNCHEM~ 61 slowly than DFS. Embedding using DFS is usually faster than BFS, perhaps because it covers cycles and other structural features sooner. BFS is faster than DFS is in some instances, however, and further investigation is needed to determine the condi- tions under which this obtains [ 1,261. Note that a computation to determine whether DFS or BFS is preferable in a given instance would add its own costs to the algorithm. One possible heuristic to increase efficiency visits nodes in the order of highest- degree-first [I]. This, in effect, causes the algorithm to traverse chains and cycles of carbons early. We found that this particular traversal heuristic made no signifi- cant difference in embedding times. A fragment code is a set of predefined symbols standing for a molecular substruc- ture such as a ring or functional group. Time might be saved if SYNCHEM~ used a fragment code to represent substructures, as is the practice in many chemical databases. However, we feel this refinement would have only a small effect, which might even be negated by the added time spent expanding the fragment symbols, e.g., as must be done when rings are broken. 6. Summary and directions for further research Graph embedding is a widely applied process that needs heuristic and theoretical improvements to push upward the threshold of combinatorial explosion. Our mir- roring technique, which models graphs with procedures, makes embedding faster by a factor between one and two orders of magnitude. The mirroring technique is suitable for stable databases needing fast pattern matching. We have eliminated recursion and used a precalculated edge table in the general embedding algorithm of SYNCHEM~, which made embedding faster by a factor of about three in 25 test cases. It will be useful to determine the relative contributions of recursion elimina- tion and the precalculated table, and the reasons why BFS is occasionally faster than DFS. In addition to graph-theoretical heuristics, a chemical context allows chemical heuristics to speed embedding searches and a wide range of possible suitable heuristics is open for investigation, including the suggested ‘attribute matching’. A statistical profile of SYNCHEM~ execution is needed to determine the distribution and structure of chemical graphs most frequently input to MATCH. A preliminary count indicates that embeddings determined by SUBSRCH are rejected by MATCH 60% of the time in typical sYNcm342 runs. Further investigation is needed into other possible embedding heuristics and their costs, as well as the costs of determining when they are favored. Towards this end, we have developed an embedding program and a driver program above it that can accept arbitrary traversal sequences, and can output traversal and embedding infor- mation at variable levels of detail. A particularly useful investigative tool would be a graphics program that displayed traversals through graphs at variable speeds using 62 J.D. Benstock et al. highlighting or color, allowing for concrete experience of, and experimentation with, different embedding strategies on different classes of graphs. Acknowledgements We wish to thank Prof. H. Gelernter and Dr. G. Miller for suggestions and helpful discussions, and Noglle Benstock for assistance with the diagrams. References [l] K.K. Agarwal, Graph transformation and canonization algorithms, Ph.D. Thesis, State Univ. of NY at Stony Brook (1976). [2] K.K. Agarwal, D.P. Agrawal and M.A. Lassner, Subgraph identification using associative techni- ques with applications to information science and chemical structure investigation, NSF Report, Wayne State Univ., College of Engineering, Detroit, MI (1982). [3] A.V. Aho, J.E. Hopcroft and J.D. Ullman, The Design and Analysis of Algorithms (Addison- Wesley, Reading, MA, 1974). [4] A.V. Aho, J.E. Hopcroft and J.D. Ullman, Data Structures and Algorithms (Addison-Wesley, Reading, MA, 1983). [5] A.V. Aho, R. Sethi and J.D. Ullman, Compilers: Principles, Techniques, and Tools (Addison- Wesley, Reading, MA, 1986). [6] R.B. Banerji, Artificial Intelligence: A Theoretical Approach (North-Holland, Amsterdam, 1980). [7] A. Barr, P.R. Cohen and E.A. Feigenbaum, eds., The Handbook of Artificial Intelligence (William Kaufmann, Los Altos, CA, 1982). [8] H.G. Barrow, A.P. Ambler and R.M. Burstall, Some techniques for recognizing structures in pic- tures, in: S. Watanabe, ed., Frontiers of Pattern Recognition (Academic Press, New York, 1972) l-30. [9] L. Bit, Processing of semantic nets on dataflow architectures, Artificial Intelligence 27 (1985) 219-227. [lo] R.H. Boivie, Heuristic search guidance in SYNCHEMZ, Ph.D. Thesis, State Univ. of NY at Stony Brook (1977). [11] J.M. Brayer and K.S. Fu, Some multidimensional grammar inference methods, in: C.H. Chen, ed., Pattern Recognition and Artificial Intelligence (Academic Press, New York, 1976) 29-60. [12] J.G. Carbonell and S. Minton, Metaphor and common sense reasoning, in: J.R. Hobbs and R.C. Moore, eds., Formal Theories of the Common Sense World (Ablex, Norwood, NJ, 1985) 405-426. [13] H.W. Davis, Computer representation of the stereochemistry of organic molecules, in: Inter- disciplinary Research Series 23 (Birkhauser, Basel, 1976). [14] H. Ehrig, M. Nagl and Cl. Rozenberg, eds., Graph-Grammars and Their Application to Computer Science, 2nd International Workshop, 1982, Haus Ohrbeck, Germany, Lecture Notes in Comp. Sci. Vol. 153 (Springer, Berlin, 1983). [15] N.V. Findler, ed., Associative Networks: The Representation and Use of Knowledge by Computers (Academic Press, New York, 1979). [16] E.C. Freuder, Structural isomorphism of picture graphs, in: C.H. Chen, ed., Pattern Recognition and Artificial Intelligence (Academic Press, New York, 1976) 248-256. [17] K.S. Fu, Syntactic Methods in Pattern Recognition (Academic Press, New York, 1974). [IS] M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide to the Theory of NP- Completeness (Freeman, San Francisco, CA, 1979). Graph embedding in SYNCHEMS? 63 [19] H.L. Gelernter, N.S. Sridharan, A.J. Hart, S.C. Yen, F.W. Fowler and H.-J. Shue, The discovery of organic synthesis routes by computer, in: Topics in Current Chemistry 41(l) (Springer, Berlin, 1973) 114-130. [20] H.L. Gelernter, A.F. Sanders, D.L. Larsen, K.K. Agarwal, R.H. Boivie, G.A. Spritzer and J.E. Searleman, Empirical explorations of SYNCHEM, Science 197 (1977) 1041-1049. [21] J.E. Hopcroft and J.D. Ullman, Introduction to Automata Theory, Languages, and Computation (Addison-Wesley, Reading, MA, 1979). [22] E.B. Hunt, Artificial Intelligence (Academic Press, New York, 1975). [23] D.B. Johnson and S. Kashdan, Lower bounds for selection in X+ Y and other multisets, J. ACM 25(4) (1978) 556-570. [24] M.A. Jones, The organization and design of SYNCHEMZ, Tech. Report No. 82/038, State Univ. of NY at Stony Brook (1982). [25] C.M. Hoffmann, Group-Theoretic Algorithms and Graph Isomorphism, Lecture Notes in Comp. Sci. 136 (Springer, Berlin, 1982). [26] M.A. Lassner, Graph embedding algorithms and their applications, Ph.D. Thesis, Wayne State University, College of Engineering (1981). [27] E.M. Luks, Isomorphism of graphs of bounded valence can be tested in polynomial time, in: Pro- ceedings 21st IEEE Symp. on Foundations of Comp. Sci., Rochester (1980) 42-49. [28] G.L. Miller, Isomorphism of k-contractible graphs. A generalization of bounded valence and bounded genus, Information and Control 56(1/2) (1983) l-33. [29] H.L. Morgan, The generation of a unique machine description for chemical structures ~ a techni- que developed at Chemical Abstracts Service, J. Chemical Documentation 5(2) (1965) 107-I 13. [30] R.R. Muntz, The WELL system: a multi-user database system based on binary relationships and graph-pattern-matching, Information Systems 3 (1978) 99-115. [31] N.J. Nilsson, Problem-solving Methods in Artificial Intelligence (McGraw-Hill, New York, 1971). [32] D.A. Norman and D.E. Rumelhart, Explorations in Cognition (Freeman, San Francisco, CA, 1975). [33] L.J. O’Korn, Algorithms in the computer handling of chemical information, in: R.E. Christof- fersen, ed., Algorithms for Chemical Computation (American Chemical Society Symposium Series Vol. 46, Washington, DC, 1977) 122-148. [34] T. Pavlidis, Linear context-free graph grammars, J. ACM 19(l) (1972) 1 l-22. [35] J.F. Pfaltz and A. Rosenfeld, Web grammars, in: 1st Proc. IJCAI, Washington, DC (May 1969) 609-619. [36] R.C. Read and D.G. Corneil, The graph isomorphism disease, J. Graph Theory 1 (1977) 339-363. [37] A. Rosenfeld, Picture automata and grammars: An annotated bibliography, in: Proc. Symp. Com- put. Image Process. Recognition 2 (Univ. of Missouri at Columbia, Aug. 1972). [38] A. Rosenfeld, Picture processing: 1977, in: Computer Graphics and Image Processing 7 (1978) 21 l-242. [39] A.F. Sanders, Some applications of graph theory, Ph.D. Thesis, State Univ. of NY at Stony Brook (1976). [40] A.C. Shaw, Picture graphs, grammars and parsing, in: S. Watanabe, ed., Frontiers of Pattern Recognition (Academic Press, New York, 1972) 491-510. [41] R.E. Tarjan, Depth-first search and linear graph algorithms, SIAM J. Comput. l(2) (1972) 146-160. [42] R.E. Tarjan, Graph algorithms in chemical computation, in: R.E. Christoffersen, ed., Algorithms for Chemical Computation (American Chemical Society Symposium Series Vol. 46, Washington, DC, 1977) l-20. [43] A. Tucker, Applied Combinatorics (Wiley, New York, 1980). 1441 L.C. Valiant, The complexity of computing the permanent, Theoret. Comp. Sci. 8 (1979) 189-202. 
work_ctmauypowbdg5nrhmjqlmxutk4	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy Modelling situation awareness for Context-aware Decision Support Yu-Hong Feng, Teck-Hou Teng, Ah-Hwee Tan * Intelligent Systems Centre and School of Computer Engineering, Nanyang Technological University, Nanyang Avenue, Singapore 639798, Singapore Abstract Situation awareness modelling is popularly used in the command and control domain for situation assessment and decision support. However, situation models in real-world applications are typically complex and not easy to use. This paper presents a Context-aware Decision Support (CaDS) system, which consists of a situation model for shared situation awareness modelling and a group of entity agents, one for each individual user, for focused and customized decision support. By incorporating a rule-based inference engine, the entity agents provide functions including event classification, action recommendation, and proactive decision making. The implemen- tation and the performance of the proposed system are demonstrated through a case study on a simulated command and control application. � 2007 Elsevier Ltd. All rights reserved. Keywords: Situation awareness; Context awareness; Decision support; Intelligent agents 1. Introduction The concept of situation awareness is well established in the field of human factor studies in complex environments. According to Endsley (1988), situation awareness refers to ‘‘the perception of the elements in the environment within a volume of time and space, the comprehension of their mean- ing and the projection of their status in the near future’’. In the domain of command and control, wherein a designated commander is required to exercise authority and direction over various forces in order to achieve the given goals, a clear picture of the current situation and an accurate pro- jection of the future states are essential for effective deci- sion making. As such, situation awareness modelling has become an important component of command and control decision support systems. Over the years, guidelines and processes have been developed for building situation awareness models from a goal-oriented perspective (Ends- ley, 2001). However, to the best of our knowledge, none has incorporated the notion of context awareness for pro- viding customized situation awareness. Context awareness was introduced by Schilit and Thei- mer (1994) to develop software that adapts according to its locations of use, the collection of nearby people and objects, as well as changes to those objects over time. With technology advancements and the growing interest in mobile and wearable computation devices, context aware- ness has become one of the major research areas in those fields. The definition of context has also expanded from physical attributes to include device characteristics as well as user-specific factors, such as profiles and preferences (Dey, 2001; Harter, Hopper, Steggles, Ward, & Webster, 1999; Moran & Dourish, 2001). Whereas situation awareness focuses on the modelling of a user’s environment so as to help the user to be ‘‘aware of his current situation’’, context awareness is about exploiting the context of a user and helping the user to have a more effective interaction with the system by actively changing the system’s behavior according to the user’s cur- rent context or situation. In the domain of command and control, individual users require specific sets of situation awareness. Also, the same piece of information may have different meanings and usages for different people in the 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.09.061 * Corresponding author. E-mail addresses: yhfeng@ntu.edu.sg (Y.-H. Feng), teng0032@ntu. edu.sg (T.-H. Teng), asahtan@ntu.edu.sg (A.-H. Tan). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 36 (2009) 455–463 Expert Systems with Applications Author's personal copy same environment. Therefore, the system must know the context of the current user and forward focused contextual information to the user. Building a single situation model and presenting the full set of information to all users is not only inefficient but highly confusing. On the other extreme, building separate situational awareness models for individual users is not only inefficient but more impor- tantly, raises the issue of consistency. In this paper, we present a system known as CaDS (for Context-aware Decision Support) that incorporates a shared situation awareness model but provides individual human operators with customized views and services through a group of entity agents. The entity agents, one for each individual user, communicate with the situation model and extract information of relevance for presenta- tion to their respective users in accordance to the user con- text. To reduce the cognitive load of the human operators, the entity agents perform event classification and action recommendation in a proactive manner. We have applied CaDS to a simulated command and control domain and evaluated its performance based on several mission scenar- ios. Our experimental results show that the entity agents are able to perform event classification and action recom- mendation with a high level of accuracy and thereby reduce the cognitive load of the human operators significantly. The rest of the paper is organized as follows. Section 2 presents the overall system architecture. Section 3 discusses the various components in the shared situation awareness model. Section 4 presents the design and implementation of the entity agents for context-aware decision support. Section 5 illustrates the various system functionalities through a case study on the command and control applica- tion. Section 6 describes our evaluation methodology and reports the experimental results. Section 7 discusses related work. The final section concludes and outlines the future work. 2. System overview Referring to Fig. 1, the CaDS architecture incorporates a situation model managing the shared situation awareness model and a group of entity agents, supported by an under- lying game simulation engine. The game simulation engine GECCO (for Game Environment for Command and Con- trol Operations) GECCO (2004) is a publicly available gen- eric platform for creating and playing real-time multi- player strategy games. The situation awareness model is based on the Endsley’s three layer structure (Endsley, 1995), but with the addition of a terrain model for the command and control domain. Assuming the existence of a sensor network, incoming data are first represented at the Perception layer (level 1) of the situation model. The Comprehension layer (level 2) then interprets the data and provides assessment of the current situation. To support anticipatory decision making, the Projection layer (level 3) predicts future states based on the understanding of the current situation. At present, only some basic projection functions have been implemented for illustration purpose. All three layers of the situation aware- ness model function concurrently and iteratively. Each entity in the environment is represented by an entity agent. Each agent has a set of goals and strategies, which form the basis of the entity’s behaviors. The goals and strategies are in turn translated into executable struc- tures like missions, plans and actions. Together with the physical attributes such as location and strengths, all these constitute the context based on which an agent provides a customized set of views and services of the shared situation awareness model to its user. Given a user context, an entity agent pro-actively scans the situation awareness model for relevant information in the environment. It then highlights important events to the user according to their significance in the given context. Our current implementation of the entity agents is based on the production rule representation and a logical reasoning mechanism. 3. Situation awareness modeling We describe the various levels of the situation awareness model, including the terrain model, in the following sections. 3.1. Terrain model Representing the knowledge of the physical world, ter- rain information forms the basis of all three layers of situ- ation awareness as well as context-aware decision support. As such, the situation model must be able to monitor the dynamic changes in the terrain, which in turn may result in significant events in the situation awareness model. The terrain model consists of two key elements, namely nodes and links. A node represents a critical point or loca- tion on the map with the necessary descriptive attributes. A link corresponds to a connection or path between a given pair of nodes, with the necessary attributes describing the condition of the connection as well as a weight attribute providing distance information. The terrain model further incorporates a shortest path algorithm for identifying the optimal route between any two nodes in the terrain. To facilitate the process of building and editing the ter- rain model, an interactive software tool named TerrainFig. 1. The CaDS system architecture. 456 Y.-H. Feng et al. / Expert Systems with Applications 36 (2009) 455–463 Author's personal copy Builder (Fig. 2) has been developed in-house. Using the notions of nodes and links, a terrain model can be easily constructed from a two-dimensional map by clicking on the appropriate positions on the map. 3.2. SA Level 1: Perception The key elements in the level 1 of the situation model are entities and events. An entity represents an object in a sit- uation which has attributes such as identities, capabilities, and so on. Virtually everything in the environment can be an entity. As shown in Fig. 3, the Entity class is a general description of an object in a situation with the basic and necessary attributes, such as identity, trajectory, and child- Entities. The trajectory is a sequence of movement made by the entity. The Event class is a data structure encapsulating all the relevant information of a physical occurrence in a situa- tion. These informations are decomposed into five ele- ments, i.e. the five W’s, namely When, Where, Who, What and Why. An event injection triggers the system to reassess the relevant entities’ attributes and their relations with the others. These changes in turn result in a new state of situation awareness and may eventually lead to critical decisions. 3.3. SA Level 2: Comprehension The comprehension level makes sense out of the data provided by the perception level and integrates them into meaningful pieces of information. Consider a movement scenario as an example, in which a space vehicle is tasked to move from one location to another and to avoid aliens along the way. In this scenario, a list of high level functions is essential for providing situation comprehension for indi- vidual entities. An example list of such functions is given below. • Distance: Calculating the distance from the current loca- tion to the destination location. • Speed required: Calculating the speed required to reach the destination on time. • Fuel required: Calculating the fuel required to reach the destination based on the current consumption rate. • Enemy on the route: Scanning the planned route for enemy. • Enemy around the route: Scanning the areas around the planned route for potential threats.Fig. 2. The Terrain Builder. Entity ConceptualEntityPhysicalEntity -position : [double, double] -condition : String -volitional : boolean MovableEntity -capabilities : String -speed : int -fuel : int -identity : String -trajectory : Trajectory -childEntities : Entity Event -when : String -where : String -who : String -what : String -why : String Relation -relationType : String 1…* 0…* 1…* 0…* Fig. 3. The schema of the situation model. Y.-H. Feng et al. / Expert Systems with Applications 36 (2009) 455–463 457 Author's personal copy • Enemy behind: Checking whether there is any enemy coming up from behind. • Zone of alert: Evaluating the conceptual proximity of enemy in terms of zones. By exploiting the relative location information provided by the enemy detection functions and the proximity informa- tion by the zone of alert functions, an entity is in a better position to formulate its strategies and plans. For instance, if an enemy is on the planned route and in zone 3 (meaning far away), an entity might choose to continue on this route as the enemy is still relatively far away and it may eventu- ally go out of its route. However, if the enemy is in zone 1 (meaning near), the user might choose to re-plan his route. 3.4. SA Level 3: Projection The ability to foresee the future is especially crucial in command and control. Based on the understanding of the current situation, it is possible to predict the future states to a certain extent. For illustration, we have imple- mented three types of projection functions: Enemy location projection, terrain projection and goal status projection. Enemy location projection is responsible for predicting the future locations and the corresponding timing of a hos- tile entity. The system keeps a finite history of the time– space information on the enemy. Statistical inference is then performed over these historical data in order to esti- mate the entity’s expected location at a particular time. Data points far back in the history are either ignored or given a lower priority in estimating the new locations. Whenever there is a change in terrain data, terrain pro- jection is conducted to predict the possible future changes, based on the existing knowledge of terrain changes. For example, if a road is blocked due to a heavy rain and our prior knowledge informs that it usually takes up to three hours for a blocked road to clear, the system would then expect to observe another terrain change to unblock the road within three hours. As mentioned, an important agenda of an entity is goal attainment. Thus an entity needs to assess the goal status at a regular time interval. Using the present information, such as the planned route, the entity’s fuel level and its current speed, the projection layer of the situation awareness model is able to compute the expected time to accomplish the goals. All these projected situations then form a part of the situation awareness. 4. Context-aware decision support The primary role of an entity agent is to provide context- aware decision support in a command and control setting so that decisions can be made in a timely and effective manner. To this end, we have developed entity agents with a range of context-aware capabilities, including event classification, action recommendation and decision mode selection. Refer- ring to Fig. 4, each entity agent communicates with the situation model and employs a rule-based inference engine to provide decision support to its user according to the user context model. The user context used by the agents consists of goals, plans and physical attributes, such as location and capabilities. An agent’s goal defines the target states of the corresponding entity. An agent’s plan contains the planned sequences of actions to achieve these goal states. 4.1. The inference engine As the activities of an entity should be driven by its goals, its attention must be directed accordingly. Using an iterative process, the attainment of goals must be con- tinuously monitored and assessed. Based on the user con- text, especially the goals and plans, an entity agent should thus constantly look out for situations in the envi- ronment that may affect its goal attainment. Our current implementation of the entity agents is based on a production rule based engine known as DROOL (Rupp, 2004), which is publicly available and fully open- source. Being one of the established ways to implement sit- uation awareness, production rules provide us with a sim- ple and comprehensible specification for recording the heuristics articulated by the domain experts. Based on the Forgy’s Rete algorithm, the rules in Drools can be written in Java, Python and Groovy via the use of XML. Each rule in Drools consists of a list of parameters, the IF conditions and the THEN consequence. In addition, a salience attribute can be assigned to each rule, as a form of priority in the activation queue. The working memory is stateless and it contains all the knowledge or facts. Facts can be asserted, modified and retracted from the working memory. Rules have to be loaded into the working memory before firing. A rule firing (execution) can modify the con- tent of the working memory and therefore possibly trigger other rules to fire. 4.2. Event classification To reduce the cognitive load of human operators, infor- mation should be presented to a user only if it is of rele- vance and significance to the user. In our system, a user corresponds to an entity in the simulated world. For exam- ple, an explorer commander is represented as an explorer Fig. 4. The entity agent model. 458 Y.-H. Feng et al. / Expert Systems with Applications 36 (2009) 455–463 Author's personal copy entity in the simulation. Therefore, the explorer com- mander should only receive relevant and significant infor- mation with respect to his/her context. Specifically, information should be presented according to their impact on the goals of an entity. Depending on how an event may affect the goal attainment status of the entity, it is presented to the user in one of the three message levels: Events that do not affect goals are posted as Infor- mation, events that may affect goals are issued as Warning, and events that endanger goals are flagged as Alert. The event classification rule module assigns an appropri- ate message level to each incoming event that would be pre- sented to the user according to the situation assessed. A typical event classification rule for an explorer1 is as shown below. The TerrainRouteBlocked rule checks that if the sit- uation type is Terrain and the event will affect the current route of the entity, the message level is set to Alert. RULE name = TerrainRouteBlocked IF SituationType==Terrain AffectCurrentPath is TRUE THEN MsgLevel = Alert 4.3. Option generation and evaluation For effective decision making in a given situation, a human operator needs to gain a good level of situation awareness by assessing his current situation. With the situ- ation awareness, he can then consider the options of actions that can be performed and decide on the best options available. An entity agent facilitates this process by monitoring the current situation and recommending possible courses of actions to its user when the need arises. The generate choice rule module is responsible for deduc- ing choices or options for a particular situation. Currently, there is a total of five action choices available for any given situation, namely Resume, Increase Speed, Decrease Speed, Reroute and Wait. The rules determine if a choice is applica- ble in a given situation and generate a list of appropriate choices from the set of actions. For instance, the following rule illustrates the generation of a choice. Under the situation type of Terrain, if the rule detects that the current route is blocked and an alternate route is available, a Reroute choice is created and inserted into the choice list. RULE name = SituationT1 IF SituationType==Terrain CurrentPathIsBlocked is TRUE AlternateRouteAvailable is TRUE THEN action = Reroute Given the current situation, the entity agent goes through the complete set of rules, each of which checks for specific conditions and generates a choice when appro- priate. The result of this process is a list of action choices ready for ranking. The rank choice rule module assesses each of the avail- able choices for a given situation with a score between 0 and 1. Generally, the scores are calculated as a function of minimal changes required and payoff in terms of goal attainment. Thus, the option with the minimal change is preferred among those with the same payoff. For example, in the movement scenarios, a Reroute action is a high change option, whereas Resume is a low change action. Therefore, given the same level of payoff, Resume is pre- ferred over Reroute. 4.4. Mode selection To exploit their proactive capabilities, we further allow entity agents to take an action automatically when appro- priate. The guiding principle is that when a decision is of high confidence and low significance, the action can be car- ried out without waiting for an explicit instruction from the human operator. Currently, there are three decision modes defined in the system. In the Auto mode, an entity agent automatically chooses the best option available for the current situation. The Recommend mode is often used when the selected option is of high consequence, wherein the agent raises a dialog prompt with a ranked list of the available options for the user. The Don’t Know mode is used when the system does not recognize the current situation and/or the system cannot identify an appropriate action. The mode selection rule module evaluates the decision mode for the current situation according to the ranked list of choices. If the agent cannot identify an action choice that can attain the assigned goal, the decision mode is set to Don’t Know. If the top choice is of high consequence, the Recommend mode is used to prompt the user for a final decision. Otherwise, the decision mode is set to Auto, under which the agent will carry out the top action choice auto- matically without the need for user intervention. The following is a sample rule used for mode selection. This rule says that the decision mode is set to Auto if the situation type is GoalStatus and the agent has found that it is impossible to attain the goal with the given resource. Consequently, the agent will report to the headquarter automatically that the entity is unable to attain its assigned goal. RULE name = GoalStatusReportToHQ IF DecisionAssist is TRUE SituationType==GoalStatus GoalAttainment is FALSE THEN DecisionMode = Auto1 Different entities can be associated with different rules. Y.-H. Feng et al. / Expert Systems with Applications 36 (2009) 455–463 459 Author's personal copy 5. Case study: Mission on mars Our implementation of CaDS is based on GECCO (GECCO, 2004), which is capable of supporting various scenarios, including war strategy games, fire fighting and rescue missions. A simple exploration and movement mis- sion is used in our study to demonstrate the various capa- bilities described. Given a fictional mission on Mars, a space vehicle Explorer ER100 is tasked to explore the unknown areas on Mars. With a sensor network deployed on the explored regions, the areas are continually monitored and activities are reported to the situation model. Consider a scenario in which the Explorer ER100 is required to return to its space station before running out of fuel. Along the way, the explorer may encounter aliens which will threaten his survival and thus the attainment of the goal. Another threat that may affect goal attainment is the changes in the terrain situation. For example, an onset of bad weather may prevent it from reaching the destination on time. In our simulated environment, there is another entity called Headquarter that oversees the Explorer’s mission and makes critical decisions only when necessary. During the simulation, two types of events are injected into the virtual environment: terrain events and entity events. A terrain event changes the terrain model, for instance, by blocking a particular node. An entity event leads to changes in a selected entity, either in terms of its attribute values or activities. As the situation changes upon each event injection, the situation model is constantly updated accordingly. In each cycle, the agent re-assesses the situation, gains a new understanding and possibly pro- jects future changes. As shown in the CaDS graphical interface (Fig. 5), all entities are displayed as image icons overlayed on the ter- rain map. Each entity is represented and assisted by an entity agent so that the sensed data in the situation aware- ness model are processed and filtered and only relevant information are presented to the user. For each entity, its goals, missions and plans are displayed accordingly on the main panel, together with other important attributes, such as the fuel level and the health index. Each entity agent communicates with its user through a message panel (in the bottom right corner), in which incoming events are displayed according to their significance, and recommenda- tions are provided to the user as the situation arises. The message panel and the graphical terrain map together enhance the user awareness of the situation in a context- aware manner. The display on the interface automatically switches into the corresponding perspective when a differ- ent entity is selected. Besides classifying and displaying events in the appro- priate message levels, the entity agent pro-actively assesses the situation for available action options and ranks them accordingly. If an action is required of which the entity agent has no authority to perform, the agent raises a choice dialog window to the user, with a ranked list of the actions and its recommendations. As shown in Fig. 6, an alien is spotted along the planned route of ER100. The agent detects the threat, assesses the various options, and recom- Fig. 5. The context-aware graphical interface of CaDS. 460 Y.-H. Feng et al. / Expert Systems with Applications 36 (2009) 455–463 Author's personal copy mends Reroute as the best solution. As Reroute is a major decision, the action is posted as an recommendation to the user together with a detailed description of the analysis and reasoning process. 6. Performance evaluation In this section, we present an experimental study in which a human operator is asked to gauge the accuracy of the entity agent acting for Explorer ER100 in terms of event classification and action recommendation. The human operator is to determine, given a specific situation, if the system classifies an incoming event or recommends an action in an appropriate manner. In addition, we esti- mate the effectiveness of the ER100 agent in terms of the reduction in the cognitive load of its commander. During the course of simulations, events can be gener- ated based on either implicit or explicit triggers. Implicit triggers are generated from the natural evolution of the simulation (e.g., the programmed responses of an entity towards the others), whereas explicit triggers are the results of external event injection. We use the following two key parameters in generating the test scenarios: (i) Initial posi- tion: The initial position of ER100 determines its route selection and therefore influences the subsequent evolution of the simulation; and (ii) Terrain change: Any variation in the terrain data, for example the blocking of a road and the clearing of a blocked road, may induce the entity into mak- ing the necessary adjustment to its current plan. Each distinct initial position of ER100 marks a unique scenario setting. In each scenario, multiple situations are created and decisions are made as the scenario plays out. A total of 73 situations based on five scenarios have been collected. The goal of ER100 remains unchanged through- out all the scenarios. 6.1. Event classification A primary role of entity agents is to classify incoming events into three levels of significance, namely Information, Warning and Alert. We define two performance measures for this function, namely the accuracy of event classifica- tion and the reduction in the cognitive load for the human operators. Based on the test situations, a human operator is asked to gauge the accuracy of the system in terms of event clas- sification. When the message level assigned by CaDS to an event is different from that of the operator, the specific event classification is deemed as inappropriate. As shown in Table 1, the entity agent achieves a high accuracy of 100.0% in classifying events. This high level of performance is achievable as the task of classifying events is relatively straightforward and our rules are sufficiently rich to cover most types of events. Context-based classification of events contributes to the reduction of cognitive load on the human operators. Instead of having to attend to each and every incoming events, a user now only needs to pay attention to events classified as Warning and Alert. To evaluate the system’s performance in quantitative terms, we define the cognitive load reduction (CLR) index for an event type c formally Fig. 6. An illustration of context-based decision support. Table 1 The prediction accuracy in event classification Number of predictions Number of correct Prediction accuracy (%) Information 30 30 100.0 Warning 34 34 100.0 Alert 9 9 100.0 Total 73 73 100.0 Y.-H. Feng et al. / Expert Systems with Applications 36 (2009) 455–463 461 Author's personal copy as Rc ¼ 1 �ðN c=NÞ, where Nc is the number of event c requiring attention and N is the total number of incoming events. As shown in Table 2, the entity agent has been effec- tive in reducing the cognitive load of the ER100 com- mander by 52.1% for Warning and 89.0% for Alert. 6.2. Action recommendation In this set of the experiments, we compare the recom- mendations made by the entity agent against the preferred action choices of the human operator. As observed in Table 3, among the actions chosen, the entity agent has a prediction accuracy of 50.0%, 75.0% and 100.0% for Increase Speed, Resume and Reroute, respectively. Fig. 7 shows an instance of the incorrect action recommenda- tions. The resume action was suggested to the human oper- ator when in fact the reroute option would be the more appropriate one. As the scenario evolved, the Explorer ER100 was observed to sustain attack by the alien. We note that not all the action choices are used during the experiments. This indicates that the specified rules have not been able to cover the space of the available actions adequately, especially on the aspect of controlling speed. Potentially, more rules can be added to cover this defi- ciency. However, the level of system complexity increases as the decision space gets more extensive with the increased numbers of rules and situation parameters. In particular, as different initial conditions may give rise to a variety of sit- uations as the scenario evolves, the task of addressing every single possible situation is a great challenge. 7. Related work Intelligent agents have been used for situation awareness modelling in a number of prior work. Urlings, Tweedale, Sioutis, and Ichalkaranje (2003) employ BDI (Beliefs– Desires–Intentions) agents to enhance situation awareness of human–machine team in a military environment. Based on the JACK agent development environment and the Unreal Tournament (UT) game engine, they present a multi-agent framework. Three types of agents are described, including a commander agent for supporting the human commander, a communication agent for com- munication between the commander agent and the human commander, and troop agents for exploring the environ- ment and for executing the orders. Although their com- mander agent is similar to our entity agents, Urling et al. are focusing on situation awareness in single agent and they Table 2 The CLR indices for warning and alert events Total number of events Number of warnings Number of alerts 73 35 8 CLRI 52.1% 89.0% Table 3 The prediction accuracy in action recommendation Action choice Number of predictions Number of matches Prediction accuracy (%) Increase speed 2 1 50.0 Decrease speed 0 0 NA Resume 20 15 75.0 Reroute 23 23 100.0 Wait 0 0 NA Total 45 39 86.7 Fig. 7. An incorrect prediction by CaDS. 462 Y.-H. Feng et al. / Expert Systems with Applications 36 (2009) 455–463 Author's personal copy do not employ the concept of shared situation awareness. In addition, although a high level description of the com- mander agent is provided, there is no concrete account on the specific decision support functions implemented and their evaluation. So and Sonenberg (2004) incorporate the notion of situ- ation awareness into a computational model for proactive agent behavior. Their model is based on a rule-based knowledge representation and a forwarding reasoning mechanism. While they develop a meta-control strategy for directing an agent’s attention in sensing and delibera- tion, their focus again is on single agent and the approach is to build a situation awareness model for each agent. More recently, Thangarajah, Padgham, and Sardina (2006) also adopt the notion of ‘situation’ for decision making in intelligent agents. Instead of using the Endsley’s three-level hierarchy, they treat situations as a conceptual entity which arguably allows a richer semantics specifica- tion. For modelling the agents’ reasoning process, Than- garajah et al. extend a BDI language called CAN for rule-based specification of the agents’ behavior. Although they see the need of organizing situation objects according to their relevance to individual agents for efficiency reason, their system again does not have the notion of shared situa- tion awareness and context-aware decision support. 8. Conclusions We have shown how context-awareness can be exploited in situation assessment and how context-aware decision support can be used to reduce the cognitive load of human operators in the command and control domain. Compared with traditional situation awareness models, our system provides customized views and ser- vices out of a shared situation awareness so as to sup- port focused and effective situation awareness based on context. More importantly, we illustrate that a context- based inference engine can be used to support a myriad of proactive decision support functions that facilitate commanders to operate in complex and time critical operations. Whereas Endsley has also advocated a goal-oriented approach to designing situation models, we provide a computational model as well as its imple- mentation for exploiting goal-based contextual informa- tion to achieve user-specific situation awareness. While the current decision support engine based on pro- duction rules has served the purpose of knowledge acquisi- tion and the proof of concepts, the task of defining the necessary heuristics based on a bounded definition of the command and control application and responding to each and every new development can be tedious. Instead of using rule-based specification, machine learning capabili- ties can be incorporated to aid in expanding the scope of the decision-making module. In fact, the task of incorpo- rating learning for situation awareness and context-based decision support has been identified as our next research target. Ultimately, we are moving towards a cognitive deci- sion system that will support both direct rule-based knowl- edge specification as well as learning based knowledge acquisition based on human-agent interaction. Although our experimentation thus far is based on a rel- atively simple scenario of explorer movement and alien avoidance, the framework theoretically can be applied to the general domain of command and control, including battlefield modelling and tactical warfare planning. These will form part of our future work. Acknowledgements The reported work is supported in part by a research grant from the Intelligent Systems Centre. The authors thank Jun Jin and Chun-Yip Lam for contributing to the development of the Mission On Mars simulator. References Dey, A. K. (2001). Understanding and using context. Personal and Ubiquitos Computing, 5, 4–7. Endsley, M. R. (1988). Design and evaluation for situation awareness enhancement. In Proceedings of the human factors society 32nd annual meeting, Santa Monica, CA (Vol. 1, pp. 97–101). Endsley, M. R. (1995). Toward a theory of situational awareness in dynamic systems. Human Factors, 37, 32–64. Endsley, M. R. (2001). Designing for situation awareness in complex systems. In Proceedings of the second international workshop on symbiosis of humans, artifacts and environment, Kyoto, Japan. Gecco (2004). Game environment for command and control operations. Available from:<http://www.csc.kth.se/tcs/gecco/>. Harter, A., Hopper, A., Steggles, P., Ward, A., & Webster, P. (1999). The anatomy of a context-aware application. In Proceedings of the 5th annual ACM/IEEE international conference on mobile computing and networking, Seattle, WA (pp. 59–68). Moran, T. P., & Dourish, P. (2001). Introduction to this special issue on context-aware computing. Human Computer Interaction, 16, 87–95. Rupp, N. A. (2004). An introduction to the drools project. Available from: <http://www.theserverside.com/tt/articles/article.tss?l=Drools/>. Schilit, B., & Theimer, M. (1994). Disseminating active map information to mobile hosts. IEEE Network, 8, 22–32. So, R., & Sonenberg, L. (2004). Situation awareness in intelligent agents: Foundations for a theory of proactive agent behavior. In Proceedings of IEEE/WIC/ACM international conference on intelligent agent technology (pp. 86–92). Thangarajah, J., Padgham, L., & Sardina, S. (2006). Modelling situations in intelligent agents. In Proceedings of fifth international joint confer- ence on autonomous agents and multiagent systems, Hakodate, Japan (pp. 1049–1051). Urlings, P., Tweedale, J., Sioutis, C., & Ichalkaranje, N. (2003). Intelligent agents and situation awareness. In Proceedings of the 7th international conference on knowledge-based intelligent information and engineering systems, LNCS 2774 (pp. 723–733). Y.-H. Feng et al. / Expert Systems with Applications 36 (2009) 455–463 463 
work_cvclentnwreufbsg3hbh34iyly	----	PII: 0888-613X(87)90012-0 Architecture of an Expert System for Composite Document Analysis, Representation, and Retrieval E d w a r d A . F o x a n d R o b e r t K. France Department o f Computer Science, Virginia Tech ABSTRACT The C O D E R (COmposite D o c u m e n t Expert~extended~effective Retrieval) pro- j e c t is a multi-year effort to investigate h o w best to apply artificial intelligence methods to increase the effectiveness o f information retrieval systems handling collections o f composite documents. To ensure system adaptability and to allow controlled experimentation, C O D E R has been designed as a distributed expert system. The use o f individually tailored specialist experts, coupled with standardized blackboard modules f o r communication and control and external kno wledge bases f o r maintenance o f f a c t u a l world knowledge, allows f o r quick prototyping, incremental development, and flexibility under change. The system as a whole is being implemented under U N I X as a set o f MU-Prolog and C modules communicat- ing through pipes and T C P / I P sockets. KEYWORDS" information retrieval, artificial intelligence, distributed ex- pert system, knowledge bases, blackboard architecture, lexicon con- struction I N T R O D U C T I O N As the world's pool o f information, particularly o f machine-readable text, rapidly expands, it becomes increasingly necessary to engage the help o f Project funded in part by grants from the National Science Foundation (IST-8418877), the Virginia Center for Innovative Technology (INF-85-016), and by an AT&T equipment contribution. An earlier version o f this paper was presented at the Third Annual USC Computer Science Symposium; Knowledge-Based Systems: Theory and Applications, Columbia, South Carolina, March 31-April 1, 1986. Address correspondence to Edward A. Fox, Department o f Computer Science, Virginia Tech, Blacksburg, Virginia 24061. International Journal of Approximate Reasoning 1987; 1 : 151-175 © 1987 Elsevier Science Publishing Co., Inc. 52 Vanderbilt Ave., New York, NY 10017 0888-613X/87/$3.50 151 152 Edward A. Fox and Robert K. France computers to control and manipulate it. Initial attempts at computer-aided information storage and retrieval (ISR) made centralized databases accessible to a large community of users [1] but focused principally on performance and have achieved only moderate levels o f effectiveness [2]. Today, end users often prefer to search for themselves [3], using gateways, front ends, intermediaries, and interfaces [4], or aided by powerful microcomputers attached to optical disk stores. These end users need more effective and adaptable tools such as have been investigated by the research community [5]. CODER (COmposite Document Expert~extended~effective Retrieval) is a research system intended to address these needs through the mechanisms of knowledge-based and goal- directed artificial intelligence (AI) techniques. Problem Description The CODER system is aimed at investigating issues of meaning representa- tion and the effective matching of user needs with relevant (passages of) documents. Although the SMART system has been evolving for more than 25 years with similar objectives [6], recent experience with reimplementing a modem version [7] and with using its latest form [8] suggests than an AI-based architecture would make further development and experimentation much easier. Questions in key subject areas that could then be studied (along with references to related work) include: COMPOSITE DOCUMENTS 1. Can composite documents that include text, factual information, and references to other documents [9] be effectively analyzed [10] so that entire documents or appropriately sized passages [11] can be retrieved? 2. Can document analysis and modeling improve with findings about abstract document structure [12], message composition [13], office modeling of documents and other objects [14, 151, document formatting [16], and related standards [17]? EFFECTIVE RETRIEVAL 3. Can effective retrieval methods suitable for the growing number of full text databases [18] be developed [2] using automatic techniques [19]? 4. Because the overlap between results of different retrieval methods is small [20], can overall effectiveness increase by use of several? Can retrieval systems be tailored to different users' understandings of relevance? 5. Can rule-based processing allow more effective combination of biblio- graphic information about documents [21] with other factual and content components than has been achieved with statistically based approaches [2217 Expert System/Composite Document Analysis 153 . Can a heuristic approach allow selection for each query of the most appropriate search strategy (e.g., choice between clustered versus inverted file searching, according to findings in [23]), the best retrieval approach (e.g., extended Boolean [24] versus vector space [25] versus probabilistic [26, 27]), and the fastest method for identifying good documents [28]? AI METHODS 7. Is the logic programming paradigm in general [29, 30] and the Prolog language in particular [31] mature enough to use for natural language analysis [32], the rule-based processing commonly used in expert system development [33], and general AI programming [34] in a large, complex system? 8. Is the blackboard model [35] of a distributed expert information- providing mechanism [36] suitable for an ISR system? KNOWLEDGE REPRESENTATION 9. In light o f the many knowledge representation schemes suggested for information retrieval [37], can computationally tractable ones [38] be developed [39]? 10. In particular, are frames [40] useful in representation and reasoning [41] about document content in a way that can aid information retrieval [42]? 11. Can temporal data be suitably represented and used [43] in retrieval? COMPUTATIONAL LINGUISTICS 12. Can linguistic analysis aid information retrieval [44], not only through improved analysis o f queries [45] but also in document analysis [46] through skimming [47, 48] or far more robust and detailed analysis [49, 50] o f more than a constrained sublanguage [51]? 13. Can machine-readable dictionaries [52] support expansion of the small lexicons used to date in text analysis systems [53]? HUMAN-COMPUTER INTERFACE 14. Does current knowledge about human-computer interaction [54] and information retrieval [55] allow development of interfaces that can adapt to individual needs and preferences? 15. Can information retrieval systems satisfy some of the needs for tutoring systems by making books [56], encyclopedias [57], and other reference works more accessible? 16. Does a graphics-based interface [58] where problem formulation, query construction, term expansion, feedback, browsing, and profile-based filtering are all interwoven in a highly interactive human-computer dialogue [59] lead to more effective and pleasant retrieval? 154 Edward A. Fox and Robert K. France Related W o r k Several research investigations are related to the CODER project. The earliest use o f expert system methods in the retrieval area was probably in the CONIT system [60]. The closest contemporary effort is development o f I3R by Thompson and Croft [61]. I3R differs by being coded as a monolithic system in Lisp, interfaced with a database system, aimed to explore retrieval methods that access a statistically analyzed document collection, and implemented using fewer but more complex experts. Yet like CODER, I3R is built around a blackboard that coordinates retrieval processing. The RUBRIC system, which uses a rule-based approach whereby queries become small knowledge bases [62], incorporates a variety o f techniques for combining evidence [63] that have been included in CODER. TOPIC is also o f interest, as it attempts to parse documents to condense their content and identify important concepts [46]. The more ambitious Project Minstrel, applying retrieval and AI methods o f office modeling, is based on a comprehensive knowledge representation facility [15]. Although CODER incorporates many ideas from other research efforts, it is unique in its aim and scope. CODER provides a unified paradigm for the entire process of information storage, representation, and retrieval based on a tailor- made encoding o f knowledge (see [64] and the section on knowledge administration below) and a flexible architecture designed to support the storage and manipulation o f that knowledge. This article concentrates on the architec- tural issue; the interested reader is referred to France and Fox [64] for more details on how knowledge is used. A R C H I T E C T U R E CODER is organized as an integrated system for document entry, analysis, storage, retrieval, and display. It should be adaptable as a standalone system, as a server for interactive or batch entry o f documents or queries, or as an intelligent intermediary to another database system. The following discussion relates to the most comprehensive case: standalone implementation. In keeping with design principles o f modularity and object-oriented program- ming, CODER is made up o f four different types o f objects differentiated by their use o f knowledge (see Figure 1). Experts are specialists in particular restricted domains pertinent to the tasks at hand. They communicate with each other only through blackboards, which serve as holding areas for session-specific knowledge. Each blackboard is the external knowledge source for a strategist that also has a local knowledge base of planning rules to coordinate the activities o f experts in the community. External knowledge bases store information o f common interest to several experts. Because they deal only with factual world knowledge, they require only limited inference abilities. Finally, there are Expert System/Composite Document Analysis 155 Expert Internal Knowledge Base BIackboard / Strategist Complex BLACKBOARD ( ) ) ) ) ) Strategist Internal Knowledge Base External Knowledge Base Resource Manager [ Figure 1. CODER Object Classes r e s o u r c e m a n a g e r s mediating between low-level machine structures and the abstract representations used by the rest o f the system. These may be implemented in a procedural language, as they do not require special knowledge or inference capabilities., The internal structure o f C O D E R is shown in Figure 2. The central region or " s p i n e " includes external databases for documents and terms, along with the knowledge administration complex. The resources o f the spine are shared by two expert communities, one for document analysis and one for retrieval. F r o m an external perspective the system wraps so that users (shown at either end o f the figure) are inside the system; they can both enter documents and retrieve them, possibly in an integrated fashion. Each user communicates with a resource manager specialized to his or her preferred interface, which in turn communicates with a group o f translation specialists to effect a two-way dialogue between the user and the rest o f the system. The interaction o f these specialists with each 156 Edward A. F o x and Robert K. France " ii i / / \ \ a ~ *.-.. E r , o go © e - ~e © e4 ° 4 Expert System/Composite Document Analysis 157 other and with other experts is mediated by a blackboard. Each expert community m a y also reference additional external knowledge bases, such as the user model base. Attached to each blackboard and coordinating all the activities o f the subsystem is a strategist. The overall operation o f C O D E R is shown in Figure 3. Because one or more parts o f C O D E R can be assigned to separate processors, it is logical to view the system as made up o f groupings o f modules needed for c o m m o n functions. F o r example, one user might be entering new documents so the system can analyze and store them, while other users are searching and retrieving documents. In both these cases, state information about the progress o f the s y s t e m ' s services for a user is maintained entirely on the blackboard involved. Finally, at the DOCUMENT ENTRY & ~1 \ \ ANALYSTS SESSION Figure 3. DOCUMENT RETRI EYAL SESSION OO¢~.tt, tEN'r COMMON EXPERTS & " FACT BASES" DOCUMENT RETRI EYAL SESSION Overview of System Operation in a Distributed Environment 158 Edward A. Fox and Robert K. France center of the figure are the shared experts and external knowledge bases involved in supporting these tasks. The sections that follow provide more detailed information about the various CODER components. Knowledge Administration Knowledge in CODER is partitioned both horizontally between the two subsystems and among the modules of each subsystem, and vertically along what Sterling [65] refers to as the "logical levels of problem solving." The top level of this second division is the goal-oriented planning knowledge that guides the session strategists. In the current CODER implementation this strategic knowledge is encoded in rules for recognizing and reacting to stages in the problem tasks. Actual steps in the problem solution are carried out by the experts in each community using tactical knowledge of how to accomplish their designated tasks. Finally, the characteristics of the problem universe are represented as world knowledge in the external knowledge bases. Strategic and tactical knowledge are stored locally in the modules that use them. The same is not true for world knowledge. Facts about the world provide the premises from which the experts reason about their tasks, and facts and hypotheses about the world make up the problem-state descriptions that inform the strategists' decisions. Thus, the factual representation language used in CODER to encode world knowledge also serves as a lingua franca for communication among the modules that make up each subsystem community. This language, defined in Figure 4, is itself made up of three levels. Elementary data types include distinct sets of names for entities of different sorts, as well as such familiar primitives as character, integer, and atom. Frames provide a facility for building definite descriptions of entities according to prototypical descriptions drawn from a tangled hierarchy of classes. And relations are predications on those entities, either ascribing accidental properties to them or describing relations among them. Relations are familiar to AI programmers by analogy to Prolog predicates. CODER relations differ from Prolog predicates in having specified type signatures (arity and types on arguments) and algebraic attributes (whether they are transitive, symmetric, and so forth). Elementary data types are also familiar; restricted data types are defined through a type of restriction polymorphism [66]. The semantics of frames, however, may require some explanation. The subsumption relation given in Figure 5a defines the inheritance hierarchy that can be specified for frame types. The frame with no slots subsumes all other frames, and two frames are equivalent if they subsume one another. Figure 5b defines the matching of frame objects, in terms of the types of the various slots and their values. Frame object A matches frame object B if every filled slot of A matches a filled slot of B, where elementary objects match Expert System/Composite Document Analysis 159 relation ::= relation_name (argument) + argument ::= relation I frame I elementary..object I quantifier argument frame ::= frame_name (slot_name s l o t _ f i l l e r ) * s l o t _ f i l l e r ::= frame I elementary_object I quantifier s l o t _ f i l l e r elementary_object ::= (quantifier) (primitive_object r e s t r i c t i o n ) quantifier ::= I | = r t J f I l e t J r I integer I n o n _ e m p t y _ l i s t _ o f I n o n _ . e m p t g . . s e t . . o f Figure 4. Definition of Factual Knowledge Representation Formalism s u b s u m e s ( a n c e s t o r _ f r a m e , d e s c e n d e n t _ f r a m e ) - s l o t _ l i s t ( a n c e s t o r _ f r a m e , a n c _ l i s t ) , s l o t _ l i s t ( d e s c e n d e n t _ f r a m e , d e s c _ l i s t ) , V x ( x ~ a n c _ l i s t D 3 y ( y e d e s c _ l i s t A n a m e ( x ) = n a m e ( y ) A s u b s u m e s ( t y p e ( x ) , t y p e ( y ) ) )). Figure 5a. Semantics of Frames: Frame Subsumption m a t c h ( f r a m e l , frame2) - s l o t _ l i s t ( t y p e ( f r a m e 1 ), l i s t l ) A s l o t _ l i s t ( t y p e ( f r a m e 2 ) , l i s t 2 ) A V x (x • l i s t l A h a s _ v a l u e ( f r a m e l , x, v) D 3y (y • l i s t 2 A name(x)=name(y) A has_value(frame2, y, r) A match(v, r) )). m a t c h ( e l t l , e l t 2 ) - e l t l = elt2. Figure 5b. Semantics of Frames: Frame Matching 160 Edward A. Fox and Robert K. France if and only if they are equal. This asymmetric matching is based only on filled slots, so objects o f differing types can still match. Matching o f frames is computationaUy inexpensive and mirrors whether the two objects, or two descriptions o f the same object, are indeed similar. By defining suitable frame types, the knowledge administrator can describe the various entities to be handled in a particular CODER system. Experts can then instantiate objects o f these types and store them on the blackboard or in external knowledge bases. Consistent use o f this formalism is maintained system-wide by the modules o f the knowledge administration complex (see Figure 6), which include type managers for each component o f the language. For example, the frame type manager keeps track o f the classes suitable for describing entities and the Constructor Plan~ler set_of nov-set ore. / / Ralatton Type Haneger Is-relation arity signature reflexive transitive sgmmotrio antisgmmotrio Relation Object Hana-'~er / ~ nov..relation J I argument arguments isomorphio list_el nov_list at©. t Fra is-eli densriptton subtypes snpertgpos reeker T I ElernentartJ Data Object Manager Frame Tupe Manager is_frame slot-list subframos super frames SUbsumes Frame Object Manager nov--frame has._slot_value set._slot_value removo__slot_value matohtng_frames equal_frames Figure 6. Internal Structure of the CODER Knowledge Administration Complex (arrows indicate dependencies) Expert System/Composite Document Analysis 161 attributes appropriate to each class; the relation type manager keeps track o f the relations existing among entities and the characteristics o f each relation. New types are added only by an external system administrator rather than by the modules of the system: deciding what sorts o f types should be recognized by the system is part o f the knowledge engineering involved in constructing CODER. Knowledge objects, by contrast--concrete expressions in the representation language--are created, modified, and destroyed dynamically during the opera- tion o f the system. External Knowledge Bases Whereas the knowledge administration complex aids with control o f the types o f knowledge representations employed in a particular system, the external knowledge bases (EKBs) provide storage and access to large numbers o f facts. The document knowledge base, the lexicon, and the user model base are all EKBs that each maintain knowledge about a particular class o f objects as specific statements o f fact. The functionality o f external knowledge bases is specified by the operations shown in Figure 7. Formally, propositions entered into a fact base are required to be ground instances o f logical relations known to the system; that is, to involve neither unbound variables nor meta-terms. These propositions are added to an external knowledge base as single statements but may be retrieved in either of two ways. The knowledge base may be queried with a skeletal fact, that is, a fact containing one or more variables, and will return the set o f all facts in the knowledge base that match the skeleton. Alternately, the knowledge base may be queried with an object (an elementary datum, a frame, or a relation) and will return the set o f all facts involving that object. In addition, a knowledge base may be queried about the number o f facts that match an object or a skeletal fact: e a t e r Figure 7. The Functionality of an External implementation independent internal structure. I t , , f a c t s _ v i t h _ v a l u e f a o t s _ v i t h _ f r a m e I r a o t s _ v i t h _ r e l f a o t s _ m a t o h i n g f r a m e s _ m a t c h i n g n u m _ v i t h _ v a l u e n u m _ v i t h _ f r a m e n u m _ v i t h . . r e l --7-1/ Knowledge Base. An EKB has no 162 Edward A. Fox and Robert K. France this information can be used by the querying entity for statistical purposes or simply to avoid receiving excessively large sets o f facts. The lexicon maintains knowledge about terms in the language. It can be conceptually divided into two parts, one o f general linguistic knowledge and the other o f specialized world knowledge particular to the collection o f documents employed. Although knowledge from both conceptual halves may be recalled for a given request, tagging the knowledge in this way promotes portability by allowing knowledge o f general use to be decoupled from the pragmatics of a given document collection and reused in other applications. Construction o f the current CODER lexicon following these principles is highlighted in Figure 8. The initial loading o f facts portrayed at the top of the figure is from one large machine-readable dictionary [67]. Table 1 describes the various relations initially derived from the more than 8 0 , 0 0 0 headwords present. Further analysis such as of the definitions (see c _ D E F ) should lead to additional relations that would be more directly usable for parsing. The document knowledge base maintains facts about the documents as assertions relating a document (passage) and a knowledge structure. A simple attached resource manager provides storage and retrieval for raw document text. Together these modules constitute the document database. Finally, there is a File of Entr~s Lexieal md • Semanfl©s Relations P.O.S.'s Jtnd Category Information • .. ( Deflnlt~ ... Tex~ ) Oenera] Lingutst~ Knovledge Plorphologj©al Variant / Ilia--Specialized Information Vm'ld Knowledge . • Protot~pteal use Information [ of ArHfle|al Figure 8. PIV'~RSe, Co-ool~lrrehce, H t e r ~ , Information Construction of the CODER Lexicon I~owledge obta~ed dur~g Do©ument Inpu~ Expert System/Composite Document Analysis 163 T a b l e 1. Relations abstracted f r o m the Collins Dictionary of the English Language [70] tapes c_ABBREV c_ALSO_CALLED c_CATEGORY c_COMPARE c _ D E F c _ D E F _ N U M c_HEADWORD c_MORPH c_NLAST c_PAST c_PLURAL c_POS c_RELADJ c_SAMP c_SINGULAR c_SYLL c_USAGE c _ U S A G E _ N U M c _ V A R _ S P E L L c _ V A R _ S Y L L Abbreviation of headword Headword also commonly called this Category (semantic label) of headword Compare to another headword and sense(s) Definition of headword Number of (up to) 80-character blocks of definition Headword entry Morphological variant of headword (including part of speech) Rest (e.g., first/middle name) of proper noun headword Past form of headword Plural of headword (sometimes just the ending) Part of speech Related adjective to headword Example of headword in context Singular form of headword (sometimes just the ending) Syllabification of headword Usage notes providing guidance on usage of headword Number of (up to) 80-character blocks in usage note Variant spelling(s) (if any) Syllabification of variant spelling(s) user model base o f facts about individual users. These include reports o f occurrences during a single session and general statements about the user, such as the type o f information that has proven relevant in the past, background knowledge particular to or supplied by the user, and c o m m o n characteristics o f relevant documents. This body o f knowledge about users, and the bodies o f knowledge discussed above, inform the s y s t e m ' s response in intelligently analyzing and retrieving documents for a particular individual. Expels Conceptually, an expert is a specialist in a certain restricted domain pertinent to the task at hand. Experts are designed to be implemented in relative isolation f r o m one another: no expert has knowledge o f the internal behavior o f the other experts in the community, and all experts communicate with the community strictly through the given blackboard. Part o f the specification o f an individual expert is the set o f predicates that it m a y view in a given blackboard area and the (possibly overlapping) set o f predicates that it m a y post back. Obviously, there must be agreement a m o n g the expert implementors on the structure and bounds o f those predicates i f the experts are to w o r k together. What each expert does 164 Edward A. Fox and Robert K. France with those predicates, however, and what internal knowledge and processes it uses to produce new hypotheses are left to the implementor of the individual expert. Each expert can therefore be built in the way that best takes advantage of the characteristics of its particular domain of expertise. An expert has only two requirements for its operation: it should be knowledge driven, and it must recognize the appropriate commands from the strategist scheduler. The first is philosophical in nature: it is part of the CODER design that the complexities of the system tasks be realized in the knowledge required for their execution rather than the process of execution itself. For experts this implies that expertise be represented as explicit knowledge, separate from whatever engine manipulates it. The knowledge in the expert, moreover, is constrained by system design to be on a higher level than factual knowledge: either rules for finding and manipulating factual knowledge in an external knowledge base or facts that relate to classes of objects in the problem universe. The second requirement is pragmatic: for the strategist to schedule their activity properly, experts must go through a canonical cycle of operations. The typical CODER expert consists of a communications interface, a local knowledge base, and an inference engine (see Figure 9). The interface provides for communication with the blackboard and optionally with external knowledge sources such as resource managers or EKBs. The local knowledge base contains the particular expertise necessary for the proper execution of the expert's tasks; the inference engine is chosen to best execute those tasks. Possible engines include both forward-chaining and backward-chaining rule interpreters, frame- based classification engines, and pattern-matching engines. These engines could then be associated with different rule bases, classification trees, and similarity measures, respectively, to produce specialized experts in a variety of disciplines. External Call Manager I Inference I Engine post viev Ir"r' abort a D s v o r s attompt-hyp attend_to_area attend_to_quest Figure 9. Canonical CODER Expert Showing Internal Structure and Functionality of Interface with Blackboard/Strategist Complex Expert System/Composite Document Analysis 165 Some examples o f expert k n o w l e d g e bases and inference techniques are given in Table 2. R e c e n t research supports the view that it is possible to build engines that c o v e r a b r o a d range o f p r o b l e m s without falling into the computational trap o f general inference (see [38] f o r a formal analysis o f this effect). This m e t h o d o f p r o b l e m d e c o m p o s i t i o n is particularly well suited to the tasks o f d o c u m e n t analysis and retrieval, w h e r e the relevance o f a given d o c u m e n t to a given information need is influenced b y m a n y factors. Assigning an expert to a T a b l e 2. K n o w l e d g e bases and inference types f o r some sample experts Expert Local Knowledge Inference Name Base Engine Date Mappings from different natural Forward chaining representations for dates into a (rule based) canonical internal representation Bibio. Ref. Different types o f biblographic entities Classificational and clues to recognize them; lexical conventions for representations of biblio. entities in text and in bibliographies Doe. Type Different types o f documents (both hard Classificational and soft types) and clues to recognize them and their component fields Declension, conjugation and case- changing rules for English; irregular morphological variants (or how to find them in the lexicon) Related Term Methods and metrics for navigating Relational relation networks in the lexicon Cluster Heuristics for finding document Special purpose Morphology Hard coded passages in the database that are similar to those identified by the user as close to current needs Methods to transform a fact-based representation o f a search request to a p-norm representation and conduct a search Methods and metrics for identifying documents in the database that share linguistic substructures with the retrieval request P-Norm Hard coded Linguistic Relational 166 Edward A. Fox and Robert K. France small area o f specialization and decoupling it from the remainder o f the system, however, has additional advantages. First, the development o f the expert is separated from that o f the surrounding system. Interaction problems, normally a plague o f AI systems, are thereby kept to a minimum. In addition, the experts are kept small, so problems o f rule interaction within the expert are minimized. Furthermore, tasks that are found to require too much complexity can be further subdivided according to the areas o f expertise required to solve them, until they are reduced to manageable size. B lackboard/ Strategist C o m p l e x A blackboard is an area for communication among experts [35]. This communication takes place through posting and reading hypotheses in special- ized subject areas. In CODER blackboards (see Figure 10), a specialization of this process provides a means for asking and answering questions, which are Specialist A [ - competent to perform I certain tasks in (or [ between) certain [ - competent to perform / certain tasks in (or I / ~ - ~ between) certain ~ / / • t Y - competent to per f o r m [ ~ oertain tasks in (or r between) oortain [ Blackboard Posting Areas Priorit W Posting Areas: ~ Ouestion and Answer Area Pending H~Jpothesis Area ~)~ Subject Posting Areas: ~ Subject Area I Subject Area 2 Subject Area N Blackboard Strategist (Planner) - - maintains a model of each area specialist. - - schedules specialist activit W. - - maintains eonsistenc~ of blackboard postin 9 areas. - - selects consistent set of hwpothosos for pendtng area. Translation Expert 1 Translation Expert M Figure 10. Blackboard/Strategist Complex Showing Mapping of Experts in the Immediate Community to Blackboard Areas Expert System/Composite Document Analysis 167 contained in a separate area o f the blackboard. The importance o f this type o f communication was noted convincingly by Belkin and colleagues [36]. In addition, the CODER blackboards provide a special area, maintained by the blackboard strategist, containing a small set o f consistent hypotheses o f high certainty. This pending hypothesis area is available for read access by all experts and thus, indirectly, by the outside world. It provides an instantaneous picture o f the " c o n s e n s u s " o f the blackboard: what the system as a whole hypothesizes about the problem under consideration at any moment. A hypothesis is a higher-order knowledge structure built on the factual knowledge forms supported by the knowledge administration complex. It consists o f five parts: 1. The fact hypothesized. 2. The identifier o f the expert hypothesizing it. 3. The confidence that the expert has in it (which can be assigned by different methods for different types o f hypotheses, according to whatever knowledge aggregation scheme is appropriate for the set o f constraints and knowledge sources at hand). 4. The hypothesis identifier. 5. The dependencies on other hypotheses. This latter information, apart from aiding selection o f the set o f pending hypotheses, allows truth maintenance functions to be performed within the blackboard subject areas. If an expert withdraws a hypothesis, for instance, or radically changes the associated confidence level, this information makes it possible to schedule tasks to reconsider the dependent hypotheses. Monitoring the blackboard for this class o f event is one function o f the blackboard strategist, shown in the lower part o f Figure 11. Because the logic task scheduler's rules governing truth maintenance are independent o f the particular predicates involved in the facts hypothesized, this function is independent o f the application domain o f the blackboard community. The strategist also monitors the blackboard for domain-specific events and conditions that trigger new processing. These categories o f function are kept separate in the strategist, so the truth maintenance function can be transported to other tasks. Both task schedulers in the strategist have been designed as rule interpreters, as neither the strategies involved in truth maintenance nor those involved in information analysis or retrieval are yet well understood. The final component o f the strategist is a dispatcher o f the tasks identified by the other components. On the basis o f the mix o f tasks scheduled by the truth maintenance and task oriented components, it attempts to make optimal use o f all available machine resources by issuing appropriately timed commands to the experts in the blackboard community. This allows different groups o f experts to be active at different phases in the community task, but also allows experts outside the currently active group to be called up to answer a question or to reconsider a hypothesis. In an ideal environment with one processor per expert, 168 Edward A. Fox and Robert K. France C Question / Answer Area C Pending Hupothes|s Ar.a ) C Subject Posting Area Subject Posting Area Post & Retract Commands Posting Area Manager Maintains Blackboard Areas | Logic Task Scheduler Interprets Generic Truth- Maint. Rules o - i L °°-*'*k Handler Scheduler Associates Interprets Questions with Application- Answer Sources Specific Rules Done & Cheokpo Reports .ules for I I Wh|oh Exports I I Rules eor • c a n Ansver Conslsten¢~ I I I i Evaluating and ~ ~ ~ Reaoting to Problem Phases Wake, A b o r t , Attend etc. Commands Task Posting Area History Figure 11. Internal Structure of Blackboard/Strategist Complex N B L A C K B 0 A R D 5 T R A T F G I $ T such dispatching duties would be kept at a minimum, and the normal cycle o f activities o f each expert would ensure highly parallel processing. I M P L E M E N T A T I O N S T A T U S A N D F U T U R E W O R K Early work on the CODER system concentrated on design and on preparation o f the knowledge to be loaded into the external knowledge bases. A test collection was needed that would allow investigation o f the many questions o f interest, so the decision was made to collect all issues o f A I L i s t Digest, an electronic mail publication distributed from the D A R P A Internet to many Expert System/Composite Document Analysis 169 networks, beginning with the first one edited by Kenneth Laws in 1984. As o f May 1987, roughly 13 megabytes o f data, including over 6500 messages by many different authors in widely differing formats, have been collected. To provide domain knowledge relevant to this collection as a supplement to the general English lexicon, The Handbook o f Artificial Intelligence has also been obtained in machine-readable form. Queries and relevance judgments on this test collection have been captured using the SMART system. Experimentation in natural language processing is best supported by a large, comprehensive lexicon. The most efficient construction approach was to reformat machine-readable dictionaries and to parse entries into suitable structures. Because the G.&C. Merriam Company and the Longman Group Limited both refused to provide their dictionaries, it was decided to use four separate dictionaries obtained from the Oxford Text Archive [68-71] so that the resulting lexicon could be made freely available to other researchers. Initial efforts focused on the largest o f these, the Collins Dictionary o f the English Language. Development o f CODER is taking place in the UNIX ~ environment. Pipes and TCP/IP sockets [72] allow intermodule and intermachine communication. Thus, procedural modules like interface managers can be coded in C or C+ + , a dialect supporting the class-object paradigm [73]. MU-Prolog was selected as the AI implementation language, as it includes a clause indexing facility for medium-size collections o f facts or rules, and two types o f database support [74]. The first scheme uses hashing, and the second employs a two-level superimposed coding scheme [75] that performs well for partial matches [76] and can easily support large Prolog databases [77]. MU-Prolog also has tools for information hiding, interfacing with the UNIX operating system, and reducing dependence on rule ordering and extra-logical operations. Implementation o f the modules o f the CODER system began early in 1986. At the end o f the summer o f 1986, the knowledge administration and blackboard/ strategist complexes were nearly complete, the communication routines were well underway, interface managers using CURSES and SUN-Windows pack- ages had been prototyped (see Figure 12), a p-norm search routine had been tested, and a first version o f the document-type specialist developed. Further coding according to the detailed specifications given in France [78] should lead to a working prototype in 1987. Subsequent efforts will aim first at demonstrating the feasibility o f using CODER for document analysis and retrieval and at comparing different approaches to see if CODER will indeed simplify experimentation regarding the application o f AI methods to ISR problems. Much o f this work, however, will require development o f a more refined lexicon, specification o f heuristics regarding appropriate retrieval methods for particular types o f queries, and Trademark of AT&T Bell Laboratories. 170 Edward A. Fox and Robert K. France Blackboard T r a n s l a t i o n I n t e r f a c e I'lanager E x p e r t s Retrieval Black- board \ _F Feed@sck Expert Quer~ Parsing Expert Figure 12. Retrieval Subsystem User Interface thorough integration o f user models into a truly interactive system for satisfying information needs. It is hoped that initial success with CODER will be followed by a long period o f productive research and experimentation that will contribute to the human knowledge base about ISR. A C K N O W L E D G M E N T S Numerous students enrolled during the last two years in Virginia Tech's two- quarter graduate sequence on information storage and retrieval have played a role in the development o f CODER. For an MS project, Robert Wohlwend made an initial version o f the Prolog lexicon from the typesetting tapes o f one major dictionary. Research assistants Mary Beth Weaver and Qi Fan Chen are currently involved in system implementation. Pat Cooper and Joy Weiss have provided secretarial support. The Oxford Text Archive and the Melbourne University Department o f Computer Science have supplied tapes with dictionaries and Prolog interpreters, respectively. The staff o f the SUMEX Computer Project at the Stanford University Medical Center, with permission o f the authors and William Kaufman, provided the machine-readable version o f the Handbook o f Artificial Intelligence. Kenneth Laws has helped assure the project o f a complete backfile o f AIList Digest issues. Expert System/Composite Document Analysis 171 References 1. Neufeld, M. L., and Cornog, M., Database history: From dinosaurs to compact discs, J. A m . Soc. Inf. Sci. 37(4), 183-190, 1986. 2. Blair, D. C., and Maron, M. E., An evaluation of retrieval effectiveness for a full- text document-retrieval system, Commun. A C M 28(3), 289-299, 1985. 3. Ojala, M., Views on end-user searching, J. A m . Soc. Inf. Sci. 37(4), 197-203, 1986. 4. Williams, M. E., Transparent information systems through gateways, front ends, intermediaries, and interfaces, J. A m . Soc. Inf. Sci. 37(4), 204-214, 1986. 5. Fox, E. A., Information retrieval: Research into new capabilities, in CD-ROM: The N e w Papyrus (S. Lambert and S. Ropiequet, Eds.), Microsoft Press, Redmond, WA, 143-174, 1986. 6. Salton, G., The SMART system 1961-1976: Experiments in dynamic document processing, Encyclopedia o f Library and Information Science 28, 1-36, 1980. 7. Fox, E. A., Some considerations for implementing the SMART Information Retrieval System under UNIX, TR 83-560, Dept. of Computer Science, Cornell Univ., 1983. 8. Buckley, C., Implementation of the SMART Information Retrieval System. TR 85- 686, Dept. of Computer Science, Cornell Univ., 1985. 9. Fox, E. A., Composite document extended retrieval: An overview, Research and Development in Information Retrieval, Eighth Annual Int. A C M SIGIR Conference., Montreal, 42-53, 1985. 10. Fox, E. A., Analysis and retrieval of composite documents, A S I S "85, Proceedings o f the 48th A S I S Annual Meeting, 54-58, 1985. 11. O'Connor, J., Answer-passage retrieval by text searching, J. A m . Soc. Inf. Sci. 31(4), 227-239, 1980. 12. Kimura, G. D., A Structure Editor and Model for Abstract Document Objects, Dissertation, Tech. Report No. 84-07-04, Univ. of Washington, Dept. of Computer Science, 1984. 13. Babatz, R., and Bogen, M., Semantic relations in message handling systems: Referable documents, Paper presented at IFIP Working Group 6.5 Symposium, 1985. 14. Horak W., and Kronert, G., An object-oriented office document architecture model for processing and interchange of documents, Proceedings o f the Second A CM- SIGOA Conference on Office Information Systems, 152-160, 1984. 15. Harper, D. J., Dunnion, J., Sherwood-Smith, M., and van Rijsbergen, C. J., Minstrel-ODM: A basic office data model, Inf. Proc. & Mgmt. 22(2), 83-107, 1986. 172 Edward A. Fox and Robert K. France 16. Peels, A., Janssen, N., and Nawijn, W., Document architecture and text formatting, A C M Trans. Office Inf. Sys. 3(4), 347-369, 1985. 17. Rauch-Hindin, W., Upper level OSI protocols near completion, Mini-Micro Systems (18)9, 53-66, 1986. 18. Tenopir, C., Full-text databases, A R I S T 19, 215-246, 1984. 19. Salton, G., Another look at automatic text-retrieval systems, Commun. A C M 29(7), 648-656, 1986. 20. Katzer, J., et al., A study of the overlap among document representations, Inf. Tech.: Res. & Dev. 1(4), 261-274, 1982. 21. Bichteler J., and Eaton HI, E. A., The combined use of bibliographic coupling and cocitation for document retrieval, J. A m . Soc. Inf. Sci., 31(4), 278-282, 1980. 22. Fox, E. A., Extending the Boolean and Vector Space Models of Information Retrieval and P-Norm Queries and Multiple Concept Types, Dissertation, Cornell Univ., University Microfilms Int., Ann Arbor, Mich., 1983. 23. Voorhees, E. M., The Effectiveness and Efficiency of Agglomerative Hierarchic Clustering in Document Retrieval, Dissertation, TR 85-705, Dept. of Computer Science, Cornell Univ., 1985. 24. Salton, G., Fox, E. A., and Wu, H., Extended boolean information retrieval, Commun. A C M 26(I1), 1022-1036, 1983. 25. Salton, G., Wong, A., and Yang, C. S., A vector space model for automatic indexing, Commun. A C M 18(11), 613-620, 1975. 26. Robertson, S. E., and Sparck Jones, K., Relevance weighting of search terms, J. A m . Soc. Inf. Sci. 27(3), 129-146, 1976. 27. Van Rijsbergen, C. J., Information Retrieval, 2nd ed., Butterworths, London, 1979. 28. Buckley, C., and Lewit, A. F., Optimization of inverted vector searches, Research and Development in Information Retrieval, Eighth Annual International A C M SIGIR Conference, Montreal, 97-110, 1985. 29. Kowalski, R. A., Logic f o r Problem Solving, Elsevier North-Holland, New York, 1979. 30. Genesereth, M. R., and Ginsberg, M. L., Logic programming, Commun. A C M 28(9), 933-941, 1985. 31. Clocksin, W. F., and Mellish, C. S., Programming in Prolog, 2nd ed., Springer- Verlag, New York, 1984. 32. Pereira, F., Logic for Natural Language Analysis, Tech. Note 275, SRI Interna- tional, 1983. 33. Helm, A. R., Marriott, K., and Lassez, C., Prolog for expert systems: An Expert System/Composite Document Analysis 173 evaluation, Proceedings o f the Expert Systems in Government Symposium, 284- 293, 1985. 34. Bobrow, D. G., If Prolog is the answer, what is the question? Or what it takes to support AI programming paradigms, I E E E Trans. Software Eng. 11(11), 1401- 1408, 1985. 35. Nii, H. P., Blackboard systems: The blackboard model o f problem solving and the evolution o f blackboard architectures, A I Mag. 7(2), 38-53, 1986. 36. Belkin, N. J., Hennings, R. D., and Seeger, T., Simulation o f a distributed expert- based information provision mechanism, Inf. Tech.: Res. Dev. Applications 3(3), 122-141, 1984. 37. Smith, L. C., and Warner, A. J., A taxonomy o f representations in information retrieval system design, in Representation and Exchange o f Knowledge as a Basis o f Information Processes (H. J. Dietschmann, Ed.), North-Holland, New York, 31-49, 1984. 38. Levesque, H. J., A fundamental tradeoff in knowledge representation and reasoning, Proceedings o f the Fifth CSCSI National Conference, London, Ontario, 141- 152, 1984. 39. Patel-Schneider, P. F., Brachman, R. J., and Levesque, H. J., ARGON: Knowledge Representation Meets Information Retrieval, Proceedings o f the First Conference on Artificial Intelligence Applications, December 1984, Denver, CO. Washington, D.C.: IEEE Computer Society Press; 280-286, 1984. 40. Minsky, M., A framework for representing knowledge, The Psychology o f Computer Vision, (P. Winston, Ed.), McGraw-Hill, New York, 1975. 41. Fikes, R., and Kehler, T., The role o f frame-based representation in reasoning, Commun. A C M 28(9), 904-920, 1985. 42. Patel-Schneider, P. F., Small can be Beautiful in Knowledge Representation, Proceedings o f the IEEE workshop on Principles o f Knowledge-Based Systems. Denver, CO, 11-16, 1984. 43. Zarri, G. P., An outline o f the representation and use o f temporal data in the RESEDA system, Inf. Tech.: Res. Dev. Applications 2(2/3), 89-108, 1983. 44. Sparck Jones, K., and Kay, M., Linguistics and Information Science, Academic Press, New York, 1973. 45. Sparck Jones, K., and Tait, J. I., Automatic search term variant generation, J. Doc. 40(1), 50-66, 1984. 46. Hahn, U., and Reimer, U., Heuristic text parsing in " T o p i c " : Methodological issues in a knowledge-based text condensation system, in Representation and Exchange o f Knowledge as a Basis o f Information Processes (H. J. Dietschmann, Ed.), North-Holland, New York, 143-163, 1984. 47. DeJong, G., An overview o f the FRUMP system, in Strategies f o r Natural 174 Edward A. Fox and Robert K. France Language Processing, (W. G. Lehnert and M. H. Ringle, Eds.), Lawrence Erlbaum, Hillsdale, N.J., 149-176, 1982. 48. Mauldin, M. L., Thesis proposal: Information retrieval by text skimming, unpublished manuscript, Dept. of Computer Science, Carnegie-Mellon Univ., Pittsburgh, Penn., 1986. 49. Riesbeck, C. K., Realistic language comprehension, in Strategies f o r Natural Language Processing, (W. G. Lehnert and M. H. Ringle, Eds.), Lawrence Edbaum, Hillsdale, N.J., 435-454, 1982. 50. Selfridge, M., Integrated processing produces robust understanding, Comp. Ling. 12(2), 89-106, 1986. 51. Sager, N., Sublanguage grammars in science information processing, J. A m . Soc. Inf. Sci. 26(1), 10-16, 1975. 52. Amsler, R. A., Machine-readable dictionaries, A R I S T 19, 161-209, 1984. 53. White, C., The linguistic string project dictionary for automatic text analysis, Proceedings o f the Workshop on Machine-Readable Dictionaries, SRI, Menlo Park, Calif., 1983. 54. Borgman, C. L., Psychological research in human-computer interaction, A R I S T 19, 33-64, 1984. 55. Sewell, W., and Teitelbaum, S., Observations of end-user online searching behavior over eleven years, J. A m . Soc. Inf. Sci. 37(4), 234-245, 1986. 56. Weyer, S. A., The design of a dynamic book for information search, Int. J. Man- Machine Stud. 17(1), 87-107, 1982. 57. Weyer, S. A., and Borning, A. H., A prototype electronic encyclopedia, A C M Trans. on Office Info. Syst. 3(I), 63-68, 1984. 58. Frei, H. P., and Jauslin, J. F., Graphical presentation of information and services: A user oriented interface, Inf. Tech.: Res. Dev. 2(1), 23-42, 1983. 59. Oddy, R. N., Information retrieval through man-machine dialogue, J. Doc. 33(1), 1-14, 1977. 60. Yip, M.-K., An Expert System for Document Retrieval, MS Thesis, M.I.T., 1979. 61. Thompson, R. H., and Croft, W. B., An expert system for document retrieval, Proceedings o f the Expert Systems in Government Symposium, IEEE, 448-456, 1985. 62. McCune, B. P., et al., RUBRIC: A system for rule-based information retrieval, IEEE Trans. Software Eng. SE-11(9), 939-945, 1985. 63. Tong, R. M., et al., A rule-based approach to information retrieval: Some results and comments, AAAI-83: Proceedings of the National Conference on Artificial Intelligence. Washington, DC, 411-415, 1983. 64. France, R. K., and Fox, E. A., Knowledge structures for information retrieval: Expert System/Composite Document Analysis 175 Representation in the CODER project, Proceedings of the Second IEEE Expert Systems in Government Conference, McLean, Va., 135-141, 1986. 65. Sterling, L., Logical levels of problem solving, Proceedings o f the Second International Logic Programming Conference, Uppsala Univ., Uppsala, Sweden, 231-242, 1984. 66. Cardelli, L., and Wegner, P., On understanding types, data abstraction, and polymorphism, A C M Computing Surveys 17(4), 471-522, 1985. 67. Wohlwend, R. C., Creation of a Prolog Fact Base from the Collins English Dictionary, MS Report, Dept. of Computer Science, Virginia Tech, Blacksburg, Va., 1986. 68. Cowie, A. P., and Mackin, R., Oxford Dictionary of Current Idiomatic English. Volume 1: Verbs with Prepositions & Particles, Oxford UP, Oxford, England, 1975. 69. Cowie, A. P., Mackin, R., and McCaig, I. R., Oxford Dictionary of Current Idiomatic English. Volume 2: Phrase, Clause & Sentence Idioms. Oxford U.P., Oxford, England, 1983. 70. Hanks, P. (Ed.), Collins Dictionary of the English Language, William Collins, London, 1979. 71. Homby, A. S., (Ed.), Oxford Advanced Dictionary of Current English, Oxford, U.P., Oxford, England, 1974. 72. Leffler, S. J., Fabry, R. S., and Joy, W. N., A 4.2BSD interprocess communication primer, in ULTRIX-32, Supplementary Documents, vol. III, 3-5-3-28, 1984. 73. Stroustrup, B., The C+ + Programming Language, Addison-Wesley, Reading, Mass., 1985. 74. Naish, L., MU-Prolog 3.2db Reference Manual, Dept. of Computer Science, Univ. of Melbourne, Melbourne, Australia, 1985. 75. Sacks-Davis, R., and Ramamohanarao, K., A two level superimposed coding scheme for partial match retrieval, lnfo, Syst. 8(4), 273-280, 1983. 76. Sacks-Davis, R., Performance of a multi-key access method based on descriptors and superimposed coding techniques, Info. Syst. 10(4), 391--403, 1985. 77. Ramamohanaro, K., and Shepherd, J., A Superimposed Codeword Indexing Scheme for Very Large Prolog Databases, Tech. Report 85/17, Dept. of Computer Science, Univ. of Melbourne, Melbourne, Australia, 1985. 78. France, R. K., An Artificial Intelligence Environment for Information Retrieval Research, MS Thesis, Dept. of Computer Science, Virginia Tech, Blacksburg, Va., 1986. 
work_cxgbziaonnb2fonutq7ug27pza	----	MULTI-AGENT COOPERATION FOR PARTICLE ACCELERATOR CONTROL Paul Skarek, László Zsolt Varga1 CERN, CH-1211 Geneva 23, Switzerland Phone: +41-22-767 3522, Fax: +41-22-767 9145, email: paul.skarek@cern.ch ABSTRACT We present practical investigations in a real industrial controls environment for justifying theoretical DAI (Distributed Artificial Intelligence) results, and we discuss theoretical aspects of practical investigations for accelerator control and operation. A generalized hypothesis is introduced, based on a unified view of control, monitoring, diagnosis, maintenance and repair tasks leading to a general method of cooperation for expert systems by exchanging hypotheses. This has been tested for task and result sharing cooperation scenarios. Generalized hypotheses also allow us to treat the repetitive diagnosis-recovery cycle as task sharing cooperation. Problems with such a loop or even recursive calls between the different agents are discussed. Keywords: Distributed AI, industrial control, diagnosis, cooperating expert systems, multi-agent systems. 1. INTRODUCTION Quite exhaustive theoretical studies exist in DAI, but it seems that there is not enough feedback from practice. In this paper we present practical investigations for applying and justifying the theoretical DAI results in a real industrial controls environment, and, conversely, we discuss the theoretical aspects of practical findings in these applied investigations made for accelerator control and operation. The results presented here are partly based on the research carried out at CERN during the ESPRIT-II Project ARCHON™ [1]. The CERN Proton Synchrotron accelerator environment, the motivation to apply DAI methods and, related to it, the results of the ESPRIT-II Project ARCHON on ARchitecture for Cooperating Heterogene- ous ON-line systems, are given in chapter 2. Chapter 3 describes different cooperation techniques applied to accelerator control and accelerator operation, some cooperation scenarios between diagnostic agents and examples taken from a control program called SETUP [2] to bring the accelerator into working con- dition after a shutdown period or after some component breakdowns. The analysis made in chapter 4 leads to a general method of exchanging hypotheses by cooperating expert systems which has been tested. This 1. on leave from KFKI-MSZKI, Budapest, POB 49, 1525 Hungary, and supported by OTKA F4064 result sharing cooperation amounts to exchanging diagnostic knowledge at different levels (partial results) and of different certainty. The definition of a generalized hypothesis allows for a unified view of control, monitoring, diagnosis, maintenance and repair tasks such that agents implementing these tasks simply exchange hypotheses. This permits us to treat the repetitive diagnosis-recovery cycle as task shar- ing cooperation. Problems with such a loop or even recursive calls between the different agents are dis- cussed. From another area of accelerator control concerning the different aspects of diagnosing timing faults we show that feedback from one diagnostic agent to another can be considered as similar to machine learning in a multi-agent system by rule compilation from simulation runs in another agent. 2. BACKGROUND CERN is a European research institute, financed by 19 member states and employing some 3000 staff members. In addition, 5000 visiting physicists, engineers and computer experts from nearly 400 research centres and universities in 40 countries use its facilities: accelerators and experimental areas. Particle accelerators are necessary to provide physicists with beams for their experiments. The CERN accelerator complex is one of the world’s most sophisticated high energy research tools. The CERN Pro- ton Synchrotron (PS) is the heart of CERN’s accelerators and experimental facilities, and acts also as injector for CERN’s bigger accelerators, the Super Proton Synchrotron and the huge LEP (Large Electron Positron rings). Accelerator operation and maintaining the underlying control system are complex tasks and difficult to survey. The necessary knowledge is naturally distributed, as is the control system which contains many systems to solve common problems. These facts suggest the application of DAI methods and in particular solutions emerging from multi-agent cooperation. Given the need to apply DAI techniques, CERN joined the ARCHON project as an application part- ner providing a large accelerator control system and two expert systems as a test-bed for the development and evaluation of the methodologies and software produced by this collaboration. ARCHON [3], [4] can be considered as a cooperation shell for combining semi-autonomous systems to work on an implicitly defined common goal, in our case, the correctly operating accelerator. In ARCHON, and this was one of the very important design objectives, the different agents can be pre-existing systems. Problems, insights and experiences gained whilst deploying ARCHON technology in real applications with the two on-line running examples (respectively IBERDROLA, an electricity utility in Spain, and CERN) are described in [5]. Details of applying the ARCHON technology can also be found in [6] and [7]. The ARCHON system architecture [3] defines an agent as an independent system (Intelligent Sys- tem) plus its “ARCHON Layer”, the cooperation layer to be added. The Intelligent Systems are systems dealing with a certain field of application. They are first of all expert systems, but other systems like data- base and control systems are also included: so that one could talk about Industrial Systems in general. The ARCHON Layers of each agent provide the functionality for cooperation, so that they can com- municate to coordinate their tasks if necessary, similarly to human operators in a control room who decide when and how to cooperate and take the necessary actions. Via these ARCHON Layers the Intelligent Systems now control not only themselves but also the way they interact with each other: this is the essence of cooperation provided by an ARCHON Layer. It is not only the distribution of data and control, but rather the distribution of the control of all data available. Each ARCHON Layer has two I/O channels according to its twofold functionality: controlling the Intelligent System and controlling cooperation, i.e. one connecting to the underlying domain system and the other to communicate to the other agents in the community. The ARCHON Layer consists of several modules like the one for the task control and status check of the Intelligent System and another one that reasons about other agents and decides when and how to cooperate, and supervises cooperation by situa- tion assessment and planning. The acquaintance and self models are also part of the ARCHON Layer. 3. MULTI-AGENT COOPERATION SCENARIOS Cooperating Diagnostic Expert Systems The two diagnostic expert systems written at CERN and then used for the ARCHON test-bed were CODES [8], which stands for COntrol system Diagnosis Expert System and BEDES [9], [10] for BEam Diagnosis Expert System. The first aspect in running an accelerator is addressed by CODES and related to fault finding and repair tasks in the complex control system of these accelerators. The main problem lies in the complexity and the implicit connectivity of the processes to be diagnosed, with its numerous inter- connected hardware and software modules. CODES was implemented using the expert system toolkit KEE™ from IntelliCorp to write a generic shell for diagnostic reasoning and includes device-centred model-based reasoning and on-line access to a database describing the parts and modules of the control system and their connectivity. BEDES addresses another aspect of the work done in the control room of an accelerator, the work of the operators to set up a certain particle beam, keep stable running conditions or change to a different kind of beam. BEDES was written to help the operators to detect and to treat mal- functions in the transfer line between LINAC II and the PS Booster Accelerator (PSB)1, and in the proc- ess to inject the beam into the PSB. The operators act upon the accelerator exclusively via the control system. Different sections of the accelerator are controlled from different general purpose consoles (work stations) in the same control room. From here stems the idea of introducing the notion of cooperation between expert systems and cooperating heterogeneous systems in general. Just as different operators coordinate their work on a com- mon goal by cooperative communication, the software in the different ARCHON Layers provides this cooperation. A general method has been developed at CERN for cooperation which was applied to expert systems that create hypothesis trees for their diagnosis. This method uses essentially result sharing cooperation, by which it speeds up the diagnostic process and gives more detailed results. The method works better if hypotheses are in a tree structure, because the additional information of relations between the hypotheses provides more information for cooperation. The two expert systems use the same diagnostic shell: Hypotheses are created by some high level symptoms, error codes coming from the control system, by the verification method of a hypothesis just being evaluated, etc. They are kept in an agenda which is aware of a tree structure of the hypotheses and they are selected from this agenda by certain criteria related to the importance of a hypothesis to be able to lead to a result. Child hypotheses in the tree structure usually correspond to subparts and more specific fault suspicions. In more detail, a hypothesis is an object which comprises, among other attributes, the fol- lowing principal information: Suspected Entity: which object in the control system is suspected (e.g. a focusing quadrupole); State of the Entity: what is the suspected erroneous state (e.g. switched off); and Verification Method: how to verify or abandon this particular hypothesis. The verification method includes both procedural data (e.g. to be collected from the database or from the process via the control system itself), and declarative data (i.e. the diagnostic rules). The cooperation is based on the exchange of hypotheses. This has the advantage that it fits naturally into the diagnostic structure of the expert systems and can be applied to existing expert systems - as shown in our case - with only slight modifications [11]. Although the hypothesis exchange as a means for cooperation was developed mainly for the cooperation of diagnostic expert systems and a larger applica- bility was not among the design goals like in the design of the Knowledge Query and Manipulation Lan- guage (KQML) and the Knowledge Interchange Format (KIF), hypothesis exchange proves to be a generally applicable method as we will see later in this paper. 1. The LINAC II is a linear accelerator which injects into the PS Booster, a pre-accelerator boosting the energy and the beam qualities before injecting the beam into the main PS accelerator. The two expert system may cooperate in several ways. In a simple task sharing cooperation when BEDES finds something that indicates a possible error in the control system, it sends this indication to CODES. In a result sharing cooperation CODES runs ahead of BEDES and makes sure that BEDES can complete its diagnosis. The result of CODES, sent to BEDES, enables BEDES to arrive at a better or more detailed diagnosis. The result sharing cooperation is also possible in the other direction: repeated messages from BEDES redirect the attention of CODES which is working in parallel to BEDES in its expanded hypothesis tree. Details of these scenarios are described in [5], [11], and [12]. These scenarios have been implemented and tested as part of the ARCHON project. The basic elements of this cooperation method are the following: sending hypothesis to each other based on the acquaintance models to check if they are of interest at all, translating received hypothesis to be understood by the receiving agent if necessary, inserting hypotheses into the agenda, and influencing the reasoning mechanism inside the expert systems by changing the “importance” of this hypothesis. The cooperating expert systems can give more details and more accurate results than the stand-alone systems which means that information can be presented to the system’s operators in a more coherent and accurate manner. In stand-alone operation, BEDES can tell that the beam is lost and suspects the operator having set wrong values or used wrong archived values and then CODES can be asked if the control system is not faulty. The DAI system, on the other hand, can immediately present the information that the efficiency of the beam went down because a control module is stopped and a control value cannot be set. Cooperation Between Accelerator Setup And Diagnosis The SETUP is a collection of programs to bring an equipment or interface module or a collection of them into a correct working state. It is quite natural to think of the SETUP as an integrated part of a gen- eral fault finding system. In relation to a CODES-agent (CODES stands here for any diagnostic expert system for analysing control system faults) the SETUP would represent a “recovery”-agent: on finding an error, CODES can cooperate with the SETUP for error recovery, and - in the other direction - the SETUP agent, in case of errors in its recovery actions, can ask CODES for more details. One could even think of an extended arbitration via an ARCHON Layer of finding the most efficient recovery procedure if there is a dynamic choice between several possibilities. There are also good reasons to include BEDES, the beam diagnosis expert system, in such a cooper- ation: as we have seen, BEDES provides a different, additional view of the diagnosis and therefore a more powerful diagnosis for the SETUP than CODES alone, and in some cases, BEDES needs directly recov- ery actions provided by a SETUP. The diagnostic modules can cooperate with the SETUP modules in other ways as well. Besides help- ing the diagnostic module with the capability of recovery, the SETUP module can contribute to the diag- nosis process of the diagnostic module. The diagnosis requires some tests to be executed (in the verification methods of the hypotheses). In a simple case it reads an alarm code from the appropriate equipment driver, but by calling the SETUP the diagnostic module can execute complex tests and receive high level information. These scenarios have been proposed as improvements to the present SETUP sys- tem [13]. Cooperation Between Different Aspects Of Timing Diagnosis The timing system in accelerator control has two main functions: to provide real-time control inde- pendent of the operator, and the time-sharing and sequencing of different operation modes, i.e. different beams through the chained accelerators. The timing diagnosis can be related to three levels of the timing system: (1) Programming the schedule of operation modes mentioned above, (2) generating the master timing, and (3) the derived timings on different levels and generated in the various timing modules of the control system, depending on the changing operation modes. On the operation modes programming level the diagnostic and checking system (as described in the paper I. Campos, J. Lewis, P. Skarek, L.Z. Varga; Rule-Based Consultant for Accelerator Beam Schedul- ing, to be submitted to: ICALEPS’95, Chicago; 1995.) should detect if the schedule can be executed by the accelerators. On the master timing level the diagnostic system should help to detect why the master timing generator select certain operation modes, because this is not uniquely determined by the planned schedule, but also by the state of the equipment. On the detailed timing level the diagnostic system should be able to detect whether the correct signals are generated as it is requested by the master timing genera- tor. The different aspects of timing diagnosis are interrelated. Some of the restrictions expressed in the rules of the operation modes programming level are partially based on the results of the master timing and detailed timing level analysis. For example if the master timing generator can never select a certain com- bination in a schedule then this combination can be ruled out immediately at the programming level. Automating the feedback from the lower level diagnostic agents to the operation modes programming level is similar to machine learning. The agent responsible for the programming level instructs the master timing level to execute a cer- tain schedule and assigns to this schedule a credibility factor (in the sense: suitable for execution). This credibility factor is - let us say - 1 in the beginning. If the master timing level is unable to execute parts of the schedule, it reports this to the programming level which reduces the factor each time it detects a prob- lem, and if it goes below a certain level, the programming level can take action to add a rule to the pro- gramming level diagnostic system to inhibit this type of schedule. The programming level must be intelligent enough to identify those characteristics of the schedule which cause the master timing level to fail. This identification can be done by comparing the schedule that failed several times with accepted schedules, and by finding those characteristics which are present only in a schedule which fails. Currently the first two agents (operation modes programming and master timing simulation) are implemented and running; the proposed feedback and the detailed timing diagnosis are in the design stage. The reason for this rule generation from numerous simulation runs is an argument for speed since the master timing level would detect the impossibility anyway, but the simulation can take a very long time. If the rules in the first agent are rather complete, the simulation run can become superfluous. 4. ANALYSIS AND GENERALIZATION We have seen that diagnostic expert systems were made to cooperate by hypothesis exchange. Now we are going to extend our view of hypotheses in such a way that the major tasks in accelerator control can be related to these hypotheses to constitute a sound basis for cooperation. Generalized Hypotheses We define a generalized hypothesis as a triplet consisting of an object, the suspected element, with a certain fault attribute (what is suspected) and a verification method or a reference to it (what to do and how to proceed in the diagnosis). In diagnosis one usually talks about the Hypothesis-and-Test strategy. We have combined these activities into the generalized hypothesis, which is a knowledge package or a knowledge source, but in a wider sense than used in the blackboard paradigm. The tasks behind the “verification-methods” need not be locally available but can be distributed in the agents’ community. This conceptual unit called generalized hypothesis represents the “Diagnostic Knowledge” and exchanging hypotheses between knowledge based systems in the form of cooperation should be seen as the exchange of formalized diagnostic knowledge. It amounts to a Hypothesis-and-Test strategy in a distributed but conceptually very compact form. Another attribute of a hypothesis refers to the “State of the Hypothesis”, which can have values like: “not yet evaluated”, “proved”, “denied”, etc. It corresponds to different types of diagnostic knowledge, and may also be seen as different steps towards a final result. The mechanism of our cooperation scheme is exchanging hypotheses based on the state of the hypothesis and possibly on the availability of the veri- fication method. We give some simple example of hypotheses from the accelerator domain in TABLE I. It is worth- while noting that the methods can create other hypotheses, child hypotheses in the hypotheses tree. Hypotheses can be “at different levels”, e.g. talking about different objects in a part/subpart hierarchy. The different levels in this abstraction correspond to different levels of the partial result as present in the result sharing cooperation via a Functionally Accurate/Cooperative [14] type strategy. Usually one distinguishes in diagnostic reasoning between symptoms (observed, initial data to start the reasoning) from which one derives hypotheses on several levels. If one finally can identify the fault, then this corresponds to a final verified hypothesis, which in turn is now able to explain the symptoms. For us conceptually and in our implementation these are all hypotheses on different levels of details as the examples in TABLE II show. Our definition for a hypothesis even allows more: the unification of the concepts of control, diagno- sis, maintenance and repair/recovery actions and represents therefore a handy way of implementing coop- eration between agents executing these tasks. We recall that the SETUP program system brings the accelerator into an initial state (normal work- ing conditions) after a shutdown or simply recovers from a failure in one or several modules of the control system. It initializes the equipment one by one and sets the correct values in the correct order. The SETUP is just an example of a controls task. Diagnosis finds out why a certain control or accelerator components are not working as they are supposed to do and suggests possible recovery actions from the fault. Cooper- ation between the SETUP, i.e. a control and the corresponding diagnosis, is quite natural. If a control action cannot be executed, one needs to diagnose the reason, which in turn will suggest a repair/recovery action to correct the fault. This recovery action is again a control which might fail, and diagnosis should give as answer: why, and so on. This type of cooperation implements the task sharing operation. The interwoven play between control and diagnosis can be seen on FIGURE 1. TABLE I - Examples For Hypotheses. Object Attribute Method The particle beam is lost Some measurement procedures (including data acquisitions) A crate controller is in bypass A rule set or/and testprogram to verify it TABLE II - Symptoms And Faults Seen As Hypotheses. Symptom Hypothesis Fault Object: a magnet an electronic module a microprocessor inthat module Attribute: is uncontrollable is faulty has a corruptedsoftware table FIGURE 1. - Control And Diagnosis Table III shows how the different control tasks are represented as hypotheses. The generalized hypotheses integrate the different activities in accelerator control. Any agent can create such hypotheses, and the hypotheses will be treated by those agents which have the capability to execute a verification method of these hypotheses. In this way the cooperative problem solving can be viewed as problem solv- ing in a tree of these generalized hypotheses, where the hypotheses are evaluated by different agents, always by agents, who have the best knowledge base and capabilities for that. The hypotheses can be dis- tributed to the appropriate agent by the ARCHON Layer. With that unified view for an hypothesis, coop- eration is reduced to the standard operation of exchanging hypotheses for all kinds of control and diagnostic tasks. Handling Of Infinite Cycles In the scenarios of control/recovery and diagnosis agents, repetitive mutual calls are possible and these calls may end up in infinite loops or even recursive calls of infinite depth. It may occur that the request from one agent to another agent may generate another request to the other agent which may regen- erate the first request and these cycles are infinitely repeated. In a loop, we have repeated calls from a diagnosis agent to a recovery agent (SETUP), but if the recovery is successful, the whole system is all right immediately, i.e it recovers immediately from the first, the original problem. In recursion any recovery is related just to the last diagnosed problem. This latest diagnosis can now therefore try the corresponding remedy proposed. Thus we have to work back through the different levels of depth to the moment where the original recovery works. This need not, however, to be the case: it is possible to think of a diagnosis at any level which has several possibilities for a recovery, and it would now try an alternative. This backtracking in a tree of recovery possibilities complicates the interaction between control and diagnosis agents even more. The detection of an infinite loop or recursion is difficult, because an agent cannot reliably establish from its local information if the other agent’s action is a response to a message sent to it, or if there is another reason. Therefore it cannot detect alone the presence of infinite cycles with certainty. Since infi- nite cycles must be avoided, they have to be detected and stopped with the help of a joint agreement of a group of agents. A seemingly easy way to stop the infinite cycle would be that if an agent detects that it TABLE III - The Generalization Of Hypotheses. Object Attribute Method CONTROL in general: Element to control (to set) Control value (Set point value) Equipment access RECOVERY Faulty element Fault Repair and recoveryaction DIAGNOSIS Faulty element Suspected fault Verification methods MONITORING Element that can gowrong Possible fault Checks (if it has really happened) CONTROL DIAGNOSIS (Monitoring is just anticipated diagnosis) If problems occur, then suggested RECOVERY/REPAIR repeatedly executes the same action, it stops the repeated action. However this may not be appropriate in the case of a database type of agent for example, which has to answer to similar requests repeatedly, and these requests may not be part of an infinite cycle. An infinite cycle detection protocol is needed to stop such a cycle. Such a protocol must be activated when one of the agents suspects that an infinite cycle has been entered, i.e. when it sends out several requests and/or data of the same type, receives several requests and/or data of the same type and there is a causal relation between the incoming and outgoing request and/ or data. An infinite cycle detection protocol could work basically in the following way: If agent A0 suspects that a request related to a message X0 sent to agent A1 ends in an infinite cycle, it sends a message “Agent A0 suspects an infinite cycle related to request X0” to agent A1. If agent A1 receives this message and agent A1 also suspects an infinite cycle, agent A1 finds out which messages are generated as a conse- quence of X0. Let us assume that these messages are X10, X11, etc. and they are sent to agent A10, A11, etc., respectively. Agent A1 then sends a message to each of agents A1i with the content: “Agent A0 and A1 suspect an infinite cycle related to request X0 and X1i”. These messages are propagated by the agents in the similar way as A1 has done it. If in the end one of these messages gets back to one of the agents that has sent out such a message, there is a loop of agents where each agent suspects an infinite cycle and it indicates that there is an infinite cycle. Now the agents can stop this cycle. The basic infinite cycle detection algorithm described above can be included in the concept of the generalized hypotheses. The generalized hypotheses can contain, as attributes, references to the requests and data exchanged in cooperation and the causal actions of the agents is represented by the creation of descendent hypotheses. If one keeps track of the creation dependencies of hypotheses, the information that is needed in the above infinite cycle detection protocol is immediately included in the generalized hypotheses. Each hypothesis should thus contain references to its direct and indirect ancestors. This was already included into the hypotheses used in the ARCHON test-bed scenarios. If an agent sees that a chain of descendent hypotheses returns to the agent several times, it can detect the infinite cycle. 5. SUMMARY This paper has presented examples of applying the methods and results of DAI to an industrial con- trols environment, in particular in the accelerator field, and it has demonstrated the validity of several DAI concepts and techniques. These investigations are among the first experiences in creating operational DAI systems by transforming existing intelligent systems of a real industrial environment into members of a multi-agent community. The generalization of the fault hypothesis concept has two major advantages: firstly, a unified view of control, diagnosis and recovery, which leads to an easy mapping of different tasks in accelerator control to cooperating agents, and secondly a straightforward method of transforming autonomous intelligent systems into a cooperating multi-agent system by exchanging hypotheses. Although some translation between the internal representation of the intelligent systems and the hypothesis representation may be needed, hypothesis exchange is a simple way to exchange knowledge in a multi-agent system. These ideas have been illustrated by several cooperation scenarios and a suggestion to avoid the problem of infinite cycles of mutual agent calls and an example of relating knowledge exchange between agents to machine learning. The development of new intelligent systems related to timing and alarm infor- mation handling as well as their transformation into a multi-agent community are under study. 6. ACKNOWLEDGEMENTS The work described here is partly based on the concepts of the ARCHON architecture, which was developed in the ESPRIT II project P-2256 (see Ref. [1]). We acknowledge the continuing support of the former group leader of the PS controls group, Fabien Perriollat, during the ESPRIT collaboration and its follow up. 7. REFERENCES. 1. ARCHON (ESPRIT Project No.2256) whose partners were: Atlas Elektronik, JRC ISPRA, Framentec- Cognitech, Labein, Queen Mary & Westfield College, IRIDIA, Iberdrola, EA Technology, Amber, Technical University of Athens, University of Amsterdam, CAP VOLMAC, CERN and University of Porto. 2. G. Daems, V. Filimonov, V. Homutnikov, F. Perriollat, Yu. Riabov, P. Skarek; A Knowledge Based Control Method: Application to Accelerator Equipment Setup, presented at the Intern. Conf. on Accelerator and Large Experimental Physics Control Systems, ICALEPCS-93, Berlin, Germany, Oct. 18- 22; 1993. 3. Th. Wittig, ed.; ARCHON: An Architecture for Multi-Agent Systems, Ellis Horwood, ISBN 0-13- 044462- 6.; 1992. 4. N.R. Jennings, T. Wittig; ARCHON: Theory and Practice, In: N.M. Avouris and L. Gasser (eds): Distributed Artificial Intelligence: Theory and Praxis, Kluwer Academic Publishers, ISBN 0-7923-1585- 5, Dordrecht; 1992. 5. N. R. Jennings, J. Corera, I. Laresgoiti, E. H. Mamdani, F. Perriollat, P. Skarek, L. Z. Varga; Using ARCHON™ to develop real-world DAI applications for electricity transportation management and particle accelerator control, to appear in IEEE Expert - Special Issue on Real World Applications of DAI; 1995. 6. D. Cockburn, N. R. Jennings; ARCHON: A Distributed Artificial Intelligence System for Industrial Applications, in Foundations of Distributed Artificial Intelligence (eds. G. M. P. O’Hare and N. R. Jennings) Wiley, 1995. 7. N. R. Jennings, J. M. Corera, I. Laresgoiti; Developing Industrial Multi-Agent Systems, (Invited Paper), First International Conference on Multi-Agent Systems (ICMAS’95), San Francisco, CA., June 12-14, 1995. 8. E. Malandain, S. Pasinelli, P. Skarek; A Fault Diagnosis Expert System for the CERN PS, Proc. Europhysics Conf. on Control Systems for Experimental Physics, Villar-sur-Ollon, Switzerland, Sept.28 - Oct.2, 1987, Proceedings in: CERN 90- 08 (ed. B. Kuiper), 217-220; 1987. 9. E. Malandain; An Expert System in the Accelerator Domain, 1st Intern. Workshop on Softw. Eng., AI and Expert Systems for High Energy and Nuclear Phys., March 19-24, 1990, IN2P3, Lyon Villurbanne, France, and CERN/PS 90- 45 (OP); 1990. 10. E. Malandain, P. Skarek; An Expert System in the Accelerator Domain, IASTED, Proc. of the Symposium EXPERT SYSTEMS Theory & Application, Switzerland, June 16-18, 1987, M.H. Hamza (Ed.), ACTA Press, Anaheim, CA, 100-105; 1987. 11. N.R. Jennings, L.Z. Varga, R.P. Aarnts, J. Fuchs and P. Skarek; Transforming Standalone Expert Systems into a Community of Cooperating Agents, Engineering Applications of Artificial Intelligence, Vol.6, No.4, Aug., 317-331; 1993, and ESPRIT ARCHON Technical Report No. 42/2-93, and CERN/PS 93-15 (CO); 1993. 12. F. Perriollat, P. Skarek, L.Z. Varga; Report on the CERN Application Study, ARCHON Public Deliverable 1060; 1993. 13. P. Skarek, L.Z. Varga; The Integration of the SETUP - Diagnosis and Recovery in Accelerator Control, CERN/PS/CO/Note 94-34, CERN, Geneva, 9 June; 1994. 14. V.R. Lesser, D.D. Corkill; Functionally Accurate, Cooperative Distributed Systems, IEEE Tr. on Systems, Man, and Cybernetics, Vol. SMC-11, No.1, January, 81-99; 1981. 
work_cxofegzdavaaxo4tk7bhxh6jii	----	ed265686.tif.pdf DOCUMENT RESUME ED 265 686 EC 181 375 AUTHOR Hofmeister, Alan M.; Ferrara, Joseph M. TITLE Expert Systems and Special Education. SPONS AGENCY Special Education Programs (ED/OSERS), Washington, DC. PUB DATE [84] GRANT G008400650 NOTE 16p. PUB TYPE Reports - Descriptive (141) EDRS PRICE MF01/PC01 Plus Postage. DESCRIPTORS *Artificial Intelligence; *Computer Managed Instruction; ''Disabilities; Educational History; Elementary Secondary Education; *Special Education IDENTIFIERS *Expert Systems ABSTRACT The application of artificial intelligence to the problems a education is examined. One of the most promising areas in artificial ictelligence is expert systems technology which engages the user in a problem-solving diaglogue. Some of the characteristics that make expert systems "intelligent" are identified and exemplified. The rise of expert systems is reviewed, and selected present and potential applications of expert systems to the field of learning disabilities are presented, such as the development of an instructional prescription based on assessment information, the classification of students based on assessment information, and the selection of appropriate behavior management strategies based on classroom observational data. (CL) *********************************************************************** Reprzductions supplied by EDRS are the best that can be made from the original document. ********************t************************************************** " U.S. =PANTOMIME Of EDUCATION NATIONAL INST!TUTI OF EDUCATION EDUCATIONAL RESOURCES INFORMATION t MEN (MCI EKING document his been recroduced is naiad from the person or wasninsion originating it. O Minor charges have bun mode to improve reproduction suably. Points of view or opinions staled in this docu- ment do not necswarily represent official NIE position or policy. 6044 Expert Systems Expert Systems And Special Education Alan M. Hofmeister and Joseph M. Ferrara Utah State University BEST COPY AVAILABLE Running Head: EXPERT SYSTEMS 1 Technical paper prepared with assistance of funds from the Special Education Programs, Department of Education Grant # G008400650. Project Officer: Jane Hauser "t 2 Expert Systems 2 AI Expert Systems and Special Education Artificial intelligence (AI) is an extensive field that has attracted the interests of a range of researchers and product developers. The field has been broadly defined as follows: Artificial Intelligence !AI) is the part of computer science concerned with designing intelligent computer systems, that is, systems that exhibit the characterist:cs we associate with intelligence in human behavior - understanding language, learning, reasoning, solving problems, and so on. (Barr & Feigenbaum, 1981, p. 3) The Rise of the Expert System The early years of AJ. research were associated with efforts to model broad-based human reasoning. These efforts were characterized by a belief that it was possible to identfy a few basic laws of reasoning, couple those laws with the speed and memory of powerful computers and produce practical "intelligent" applications. These general-purpose approaches to problem solving proved to be too weak. When applied to specific problems, they were neither that intelligent nor that practical. In reaction, a number of researchers de-emphasized general information processing. For them, the structure and nature of specific knowledge bases became a major area of 3 Expert Systems 3 interest. The terms knowledge engineering, and knowledge programming reflected this change in emphasis. As a result of this increased interest in knowledge bases, a number of knowledge engineer's efforts have achieved some success in replicating human expertise in specific areas. These developments have been termed "expert systems ". An expert system typically engages the user in a dialogue. This dialogue, in many ways, parallels the conversation a person might have with a consultant who has expertise in the specific area. The system interrogates the user, asking questions that will pinpoint the problem and the information needed by the expert system to suggest a solution. The expert system attempts to solve the problem using the information supplied by the user and rule based procedures that are specific to the knowledge area (Stefik, Aikins, Balzer, Benoit, Birnbaum, Hayes Roth, Sacerdoti, 1983). Some expert systems will pose solutions and assign a confidence factor or probability level to the solution. If the information supplied by the user is incomplete, a solution with a reduced confidence factor may be presented. Generally, the procedures used by expert systems have been developed after examination of examples of problem solving. These examples may be generated by one expert or a group of experts. The examples are studied to identify the underlying rules used by the experts in the problem solving or to verify the rules supplied by the experts. These Expert Systems 4 underlying rules are combined to form the problem solving processes used by the expert system. It is not unusual for an expert system to use several hundred rules. Expert System Examples Expert systems have effectively generated rules and solved problems in a number of areas. PROSPECTOR, for example, is an AI program which is used in the field of mineral exploration. PROSPECTOR interprets soil and geological data and predicts the probable location of mineral deposits. In an experiment testing PROSPECTOR's effectiveness, users correctly predicted the location of a molybdenum deposit worth one hundred million dollars. Another AI program, MACSYMA was designed to solve a variety of complicated mathematical problems. Scientists and engineers access MACSYMA through a telephone network. Research chemists employ DENDRAL. Using mass spectra/ and nuclear magnetic resonance information. DENDRAL can identify a substance's potential molecular structure. MYCIN, another well-known expert system. has led to educational applications (Davis, Buchanan, and Shortliffe. 1975). This program allows the user to feed in information on the characteristics of bacterial cultures and the patient's present symptoms. It then deduces the bacterial disease. In its initial form, this intelligent data base was used as a diagnostic tool by the physician. This same intelligent data base was then included in an intelligent 5 Expert Systems 5 CAI program, NEOMYCIN (Clancy & Letsinger, 1981), designed to be used to teach the diagnosis of bacterial diseases. When one examines the research and development being conducted in AI in different western countries and in the different areas of society, one void becomes very obvious. Although there are rather extensive developments in industry, medicine, and the defense department, there appears to be a comparative dearth of research efforts to apply AI to problems in public education in the United States (Sleeman & Brown, 1982; Roberts & Park, 1983). There are several reasons why educators have not been active in this area. First, the technical and personnel resources necessary for the development of artificial intelligence (AI) products have, until recently, been rare and expensive. One report estimates that 'fewer than half a dozen doctoral theses appear each year in some aspect of building knowledge systems" (Stefik et al., 1983). Second, the long-term efforts necessary for AI product development did not fit the funding patterns for educational research. In referring to examples of effective expert systems, Winston (1979) noted that no one should look at these systems without understanding the "veers of team effort that have gone into translating the basic strategies into working, useful systems" (p. 237). 6 Expert Systems 6 Expert System Development Tools Until recently AI research and development activities were not widespread. Today, however, research in artificial intelligence is no longer limited to a few basic researchers working in well-endowed laboratories. The successful commercial and industrial application of a number of expert systems has led to the development of flexible AI tools which allow individuals with limited training in computer science to build expert systems. Prior to the development of these tools, methods of knowledge programming were complex. Mastery of these methods usually required several years of study. The existence of development tools has reduced programmer training time. In addition, advances in hardware development have yielded more advanced and more affordable equipment. Although sophisticated AI development tools are typically designed to operate on large computers and/or spe:dalized machines designed specifically for AI development, there are AI tools which can operate on microcomputers (Teknowledge, 1984; Needle, 1984; Export Software International, 1984). Even the Apple IIe computer is being groomed for AI developrent and use. The December 1983, issue of A+ (an independent magazine for Apple computer users) carried a notice that by 1985, hardware and software which would allow for the development of AI programs on the Apple Ile would be available. Although the a Expert Systems . 7 development of complex expert systems is still clearly not a project for a lov power microcomputer, the existence of these scaled down systems appears to place expert systems development within the reach of public education. The computer equipment needed to develop an expert system is often far more expensive than the equ:Ipment needed to use a finished system. An extended version of Interlisp, a programming environment designed so support artificial intelligence applications, has been made available for personal computers and a number of expert systems originally developed on more expensive machines have been made operational on microcomputers (Sleeman and Brown, 1982). At least one AI development tool designed for large mainframe computers has taken advantage of the fact that the completed expert system can run on microcomputers. The Knowledge Engineering System (KES) is designed to download programs developed on large mainframe machines to microcomputers (Magl4c.ro, 1984). In commenting on the reason for the recent dramatic increase of interest in expert systems, John Seely Brown (1984) observed: The answer rests not in the intellectual arena but in the recent dramatic advances in hardware.. . . Artificial-intelligence systems that required dedicated, million-dollar mainframes five years ago now can run on machines that cost only 8 Expert Systems 8 $25,000. For the first time we have cost- effective delivery engines for expert systems, a major change. (p.82) It is our assessment that low-cost versions of most popular minicomputers will be available within the next few years. Further, we believe that a wide variety of Expert System development tools for those machines as well as existing microcomputers will be marketed. Potential Applications in Special Education Hayes-Roth, Waterman, and Lenat, (1983) have documented the generic categories of knowledge engineering. Their listing places an emphasis on prediction, interpretation, diagnosis, remediation, planning, monitoring, and instructional tasks. A review of this listing indicates that a potential exists to match knowledge engineering approaches with problem areas in special education. The categories such as diagnosis, planning, and instruction are highly related to important special education activities. In special education, potential problems might include: (1) the development of an instructional prescription based on assessment information, (2) the classification of a child into one of the special education categories (learning disabled, mentally retarded, etc.) based on assessment information, and (3) the selection of an appropriate behavior management strategy based on classroom observational data. Most situations where consultant help Expert Systems 9 has value represent potential areas for the development of expert systems in,special education. When expert systems are used fgr direct instruction, the terms "Intelligent Computer Assisted Instruction' (ICA') or "Intelligent Tutoring Systems" (ITSs) are sometimes applied to the products. Expert system product development efforts can have at least three beneficial effects on the field of special education. First, an expert system teamed with a powerful small computer, can make low cost, computer consultant services available to classroom teachers. A second benefit is the training value of the 41. "intelligent knowledge base" generated by the development of the expert system. This knowledge base is a model of reality and can be used in the training of human experts. This training value has important implications in human service programs where simulation based training can reduce the threats presented to special education students by beginning instructors and diagnosticians. Hofmeister (1984) developed an expert syrtem *CLAS.LD" to provide a second opinion regarding the accuracy of tho classification "learning disabled". To use the system, the user brings the psychological and educational information used by the assessment team to the computer and responds to a series of questions posed by the computer. The system operates on a high powered microcoAputer. It is under 10 Expert Systems 10 study for its value as a consultant and as a clinical training resource where graduate students can test their diagnostic and classification skills against the expert system. The system uses Utah and federal regulations related to public law 94-142 in its problem solving processes. The expert knowledge base was built on the opinions of several nationally recognized authorities in learning disabilities. A third benefit of AI product development is more subtle but just as important. In order to develop an expert system, knowledge engineers oust organize and analyze the existing knowledge within a subject area. Expert system development could, in this way, accelerate the clarification and expansion of knowledge in Special Education. The knowledge generation and clarification activities associated with AI product development have research implications of considerable value. At least two well- validated AI programs, DEBUGGY (Brown & Burton, 1978) and The Computer Guided Diagnosis of Learning Disabilities Prototype ('olbourn, 1982; Colbourn & Mcleod. 1983). have significant implications for special education. In DEBUGGY, the knowledge generation aspect was significant. The developers of DEBUGGY added considerably to our knowledge of student errors in arithmetic. With DEBUGGY, the user is trained to identify error patterns in students' attempts at arithmetic problems. 11 Expert Systems 11 Colbourn (1982), deve.s.oped a prototype expert system which provided the user with a diagnostic report which could then be used in the development of a remedial program. The performance of this prototype program suggests that using an expert system for special education diagnosis is clearly feasible. After comparing computer-guided diagnoses with those of humans on 22 files, Colbourn and McLeod (1983) concluded: In general, the results of the evaluation were encouraging; the expert system's diagnoses were accurate. Furthermore, because of the system's speed at analyzing error patterns, its diagnostic reports included more information than these of the human diagnosticians. This was particularly ncticeable with regards to the analysis of phonics skills. (p. 37) Summary. During a presentation to congress, Sell (1983) made the observation that "Too much computer software is simply electronic page turning, and it has little advantage over a well-illustrated book" (p. 4). Artificial Intelligence represents a dramatic shift from 'electronic page turning". It has been called the "fifth generation' or "second computer age," and it is characterized, not so much by increases in the power of hardware, but by dramatic increases in the sophistication of software. Feigenbaum 12 Expert Systems 12 and McCorduck (1983) observed that it is, . . . the important computer revolution, the transition 'from information processing to knowledge processing, from computers that calculate and store data to computers that reason and inform. Artificial intelligence is emerging from the laboratory and is beginning to take its place in human affairs. (p. 1) We believe that artificial intelligence, particularly expert systems, can and should play an important role in the future of Special Education. This contribution would occur through (1) the increase in availability of expert system consultant services, (2) the use of the problemsolving models and associated software and knowledge bases for staff training, and (3) the clarification and expansion of knowledge bases as a result of the expert system development process. 13 Expert Systems 13 References Barr, A., & Feigenbaum, E. A. (Eds.). The handbook of artificial intelligence, (Vol. 1). Los Altos, CA: William Kaufman, Inc., 1981. Bell, T. H. Computers and education. Statement made to subcommittee on Investigations and oversights, Rouse Committee on Science and Technology, U.S. Rouse of Representatives, Washington, D.C., September 28, 1983. Brown, J. S., & Burton, R. R., Diagnostic models for procedural bugs in basic mathematical skills. Cognitive Science, 1978, 2, 155-192. Clancey, W.J., & Letsinger, R. NEOMYCIN: Reconfiguring a rule-based expert system for application to teaching. Proceedings of the seventh international joint conference on artificial intelligence, University of British Columbia, Vancouver, B.C., Canada, 1981. Colbourn, M. J. (1982). Computer-guided diagrams of learning disabilities: A prototype. M. Ed. Thesis.. Department for the Education of Exceptional Children, University of Saskatchewan. (ERIC Document Reproduction Service No. il."D 222 032) Colbourn, M. J. & Mcleod J. (1983). Computer- guided educational diagnosis: a prototype expert system. Journal of Special Education Technology, 6, 30-39. Davis, R., Buchanan, B., & Shortliffe, E. Production rules as a representation for a knowledge-based consultation 14 Expert Systems 14 program. Report STAN-CS-75-519, Memo AIM-266. Computer Science Department, Stanford University, 1975. Export Software International. (1984). Expert-Ease Users Manual. Edinburgh, U.K.: Author. Feigenbaum, E. A., & McCorduck, P. The fifth generation. Reading, Mass.: Addison-Wesley, 1983. Hayes-Roth, F., Waterman, D.A., Lenat, D. B. Building expert systems. Reading, Mass.: Hofmeister, A. M. (1984). "CLAS.LD". Addision-Wesley, 1983. An classifying learning disabilities. expert system for Computer program for the IBMPC. Utah Statc University, Logan, Utah 84321. Magliero, A. (1984). Review of Expert System Tools For a Prototipe_Prolect. (Available from Software Architecture and Engineering, Arlington, VA). Needle, D. (1984, August). From the News Desk. Info World, p. 9. Roberts, F. C., & Park, 0. Intelligent computer-assisted instruction: An explanation and an overview Educa- tional Technology, 23(12), 7-12. Sleeman, D., & Brown, J. S. Intelligent tutoring systems. London: Academic Press, 1982. Stefik, M., Aikins, J., Balzer, R., Benoit, J., Birnbaum, L., Hayes-Roth, F., & Sacerdoti, E. The architecture of expert systems. In F. Hayes-Roth, D.A. Waterman, & D. B. Lenat (Eds.), Building expert systems. Reading, Expert Systems 15 Mass.: Addison-Wesley, 1983. Teknowledge Inc. (1984). M.1 Product Description. Palo Alto, CA: Author. Winston, P. A. & Prendergast, K. A. (1984). The AI Business: Commercial Uses of Artificial Intelligence (p. 82). Cambridge, Mass: MIT. Winston, P. H. Artificial intelligence. Reading, Mass.: Addison-Wesley, 1979. 
work_cxp44guqhraubbygiwash2hh4q	----	A Study on the Conception of Generic Fuzzy Expert System for Surveillance (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 218 | P a g e www.ijacsa.thesai.org A Study on the Conception of Generic Fuzzy Expert System for Surveillance Najar Yousra Dep. of Informatics, Higher Institute of Informatics Tunis, Tunisia Ketata Raouf National Institute of Applied Sciences and Technologies. Research unit on Intelligent Control Design and Optimisation of Complex System (ICOS) Tunisia Ksouri Mekki National School of engineering in Tunis, Tunisia Research Unit LASC, ENIT Abstract—This paper deals with using fuzzy logic to minimize uncertainty effects in surveillance. It studies the conception of an efficient fuzzy expert system that had two characteristics: generic and robust to uncertainties. Analyzing distance between variables optimal and real values is the main idea of the research. Fuzzy inference system decides, then, about significant variables state: normal or abnormal. A comparison between three proposed fuzzy expert systems is presented to highlight the effect of membership number and type. Beside, being generic this system could also be applied in three fields: industrial surveillance, camera surveillance and medical surveillance. To expose results in these fields, matlab is used to realize this approach and to simulate systems responses which revealed interested conclusions. Keywords—Generic Fuzzy expert system; surveillance; uncertainty'error analysis ;three tanks; ECG I. INTRODUCTION Ambiguous environments constitute an enormous problem for decision makers. As a matter of fact, uncertainties affect decision making especially for surveillance in many fields. These uncertainties are the result of many sources such as: nonlinearities, non exhaustive mathematical models, non effectiveness of sensors/detection equipments and qualitative knowledge representation. The most common methodologies that had dealt with this issue are traditional tools as probability theory, error interval analysis and especially fuzzy theory [1]. The first link between fuzzy theory and decision making was introduced in [2]. It was based on the fact that according to a criterion good solutions are fuzzy sets. Besides, the best solution set is obtained from their intersections [3]. The most popular fuzzy sets approach, in decision-making, is the maximum ranking solutions. This method is natural when interpreting the fuzzy sets as flexible constraints. While uncertainty affects several domains and has many facets (randomness, fuzziness etc.), fields and applications concerned with this issue are, especially in the last decade, growing proving the efficiency of fuzzy logic use. Fuzzy Expert Systems (FES) are expert system that uses fuzzy sets to reason [5]. In another words, FES are intelligent tools capable of making decisions dealing with ambiguous data. A recent research [4] had proved that, in 2010, that the number of published papers adopting fuzzy systems approaches is the most important. Besides, the same article confirms that industrial and medical applications and especially diagnosis is the most growing application field of these techniques. II. FUZZY EXPERT SYSTEM OVERVIEW Fuzzy expert system or fuzzy inference system is composed of three units: fuzzification, inference engine and defuzzification. It treats qualitative data with vague and fuzzy descriptions. The application of fuzzy expert system touched many fields especially industrial and medical surveillance. Recent works used FES in fault diagnosis applications. [11] had realized a diagnosis application which based on FES to identify failures in power system by analyzing amplitude and signal orientation then by classifying abnormalities. In the same spirit, many applications in power system had been developed with FES such as [29][14][8]. Detecting episodes of poor water quality is realized by fuzzy inference system in [15]. In the same domain these works using FES treated aluminum electrolysis [16] and detecting failures in computer [17]. Research results also in developing decision making applications using fuzzy expert systems in medical diagnosis. [18] presents a fuzzy application to analyze diabetic state. In another hand, using fuzzy logic many systems take decision: [19] about hypertension state, [20] about lever state, [21] about state of prostate cancer, [22] about heart state, [23] about breast cancer. TABLE I. FES APPLICATION FES Application Fields Publications about SEF Industrial diagnosis [45], [30], [31], [32], [33], [28], [34], [35], [36], [37], [38], [39], [40], [41], [42], [6], [43] , [44], [7], [11], [14], [8], [15], [10], [50],[58] Medical diagnosis [47], [52], [46] ,[48], [49] ,[18], [19], [20],[21],[53],[23], [54], [19], [45] Video surveillance [54],[60] Economic domain [55], [56], [51] Civil domain [34], [10], [27] Software domain [57] (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 219 | P a g e www.ijacsa.thesai.org Fig. 1. Fuzzy Expert System publications per domain (50 publications) Fig. 2. Fuzzy Expert System publications per year (2007-2013) This state of art on FES between 2007 and 2013 took into account 50 new publications. We conclude on the importance of fuzzy inference systems in decision making issue. In fact, uncertainty is a matter that affects all kind of field which explains the applications diversification. This research confirms the conclusions made in [4] about the most significant application field which is diagnosis (industrial and medical). We explain these facts by the need of decision aid systems in diagnosis and by the abundance of fuzzy data in these environments. Results generated by different expert systems are robust against vagueness and uncertainty and a certitude coefficient is usually calculated to enhance the effectiveness and the interpretation of outputs. We remark also that each application had its particular inputs and outputs. Thus, the developed fuzzy expert systems are specific for each treated problem. This point had been the key of our research issue. The abundance of articles in this issue indicated efficient results in industrial and medical diagnosis and surveillance. However, other fields are concerned with fuzzy systems. [52] used genetic algorithm to build rule base, [43] evaluated the performance of software based on certain characteristics, [24] developed FES to evaluate the state of public discharge land, [25] realized a multi agent system and FES decided about the role of each agent, [26] used FES in supply chain localization, [27] used FES in travelling domain, [28] used FES in renewable energy. III. RESEARCH ISSUE Our aim is to propose a generic fuzzy expert system that could be applied in several domains to monitor the state of significant variables characterizing the studied situation. In this optic, we should first determinate these variables and fixe their optimal and desired values. Then, the proposed FES is responsible of deciding whether the variable is behaving in optimal trajectory. We gave a special attention to research of Evsukoff [7] that presents a FES based on the analysis of significant variables residual and their variation. It was applied in industrial fault detection where partial decisions are made about variables (normal-OK or alarm-AL). Fig. 3. Evsukoff fuzzy inference system In an earlier work [58], we’ve proposed a modified version of Evsukoff FES which minimized rules number and raised robustness against uncertainties by weighting rules with triangular functions instead of fixed values. In the same way and applied in control, [59] analyzed the error end its variation using fuzzy expert system with seven membership functions for each input. Figure 4 illustrates its inference system. Fig. 4. Panda inference system for error analysis After studying different points of view, we are trying to determinate the most appropriate fuzzy expert system to be adopted and to be a generic tool for monitor a situation and for helping in decision making about its state. We are proposing a system which gives partial conclusion about each variable by analyzing its residual. Also, we are studying in this work the effects of raising the number of membership function in FES. Finally, we should prove that the approach is generic by applying same FES in different fields. IV. CONCEPTION OF FES FOR SURVEILLANCE Surveillance is, generally, assured through two steps: detection of anomalies and their diagnosis. The FES we are proposing is responsible of detecting abnormal situation. Figure 5 is schema block that defines the inputs and the outputs of the system. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 220 | P a g e www.ijacsa.thesai.org Fig. 5. Schema bloc of FES In fact, It has two inputs: residue/error ( r ) and residue derivative ( dr ). The residue is considered in this case as distance separating variables actual/real values from desired/optimal ones. Then, we could define second input as residue derivative that could inform about residue evolution. The output, in another hand, is certitude factor ( cf ) that evaluates the state of concerned variable. Matlab is used to develop and to simulate FES because it had fuzzy logic toolbox. Fig. 6. Matlab Schema bloc of r and dr calculation A. Fuzzification Input variables could be qualified as table 2 indicates. Each variable could be presented by linguistic terms (3-5-7). Three scenarios are to be considered: TABLE II. FUZZIFICATION R AND DR Prop. Number of Membership functions Symbolic labels Types of membership functions Sc 1 3 A(r)/B(dr)= {N,Z,P} N (negatif) Z (zero) P (positif) Trapezoïdal/ Triangular Sc 2 5 A(r)/B(dr)= {NB,NS,Z,PS, PB} NB (negatif big) NS (negatif small) Z (zero) PS (positif small) P (positif big) Trapezoïdal/ Triangular Sc 3 7 A(r)/B(dr)= {NB,NM,NS,Z,PS ,PM,PB} NB (negatif big) NM (negatif moyen) NS (negatif small) Z (zero) PS (positif small) PM (positif moyen) P (positif big) Trapezoïdal/ Triangular r and dr are variables ranging respectively in sets of symbolic labels A(r) and B(dr). The terms describe qualitative value of magnitude of both residue and its variations. Fuzzification of the two inputs could adopt three scenarios with: 3 membership functions, 5 membership functions, 7 membership functions. TABLE III. FUZZIFICATION OF CF Number of MF Linguistic Terms MF types 6 C(cf)= { AL0, AL0.2, AL0.4, AL0.6, AL0.8, A} AL0 AL0.2 AL0.4 AL0.6 AL0.8 AL1 triangular The linguistic variable cf is output variable ranging in sets of symbolic labels C(cf) = {AL0, AL0.2, AL0.4, AL0.6, AL0.8, AL1}.( as table 3 ). B. Defuzzification In this system output calculation, a crisp value is required. Thus, the defuzzification operation is requisite. In this approach, the gravity centre is the method adopted to get the crisp value traducing the severity of generated alarm, from the output membership function. C. Inference engine Linguistic model relating variables r and dr to variable D is written as rule base, relating the terms of A(r) and B(dr) to those of C(cf) in n rules : If r is Ak and dr is Bl then cf is Cs )1( In our research, we are studying the effect of raising the number of symbolic labels describing the linguistic variable. We suppose these hypotheses: H1: Raising symbolic labels enhance robustness against uncertainty. H2: It could lead to the augmentation of rule number which affects negatively time response. We are working within three scenarios depending on the number of membership functions representing variables. For each case, an inference system relating inputs to outputs is proposed. - FES1 : r (5 MF) and dr (3MF): In this case, residual associated with five symbolic labels and dr with three symbolic labels. The number of rules n=15. Table 4 is an illustration of the inference. TABLE IV. FES1: INFERENCE SYSTEM r dr N Z P NB AL1 AL1 AL0.8 NS AL1 AL0.6 AL0.4 Z AL0.2 AL0 AL0.2 PS AL0.4 AL0.6 AL1 PB AL0.8 AL1 AL1 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 221 | P a g e www.ijacsa.thesai.org - FES2 : r (5 MF) and dr (5MF) In this case, residual associated with five symbolic labels and dr with five symbolic labels. The number of rules n=25. Table 5 is an illustration of the inference. TABLE V. FES2: INFERENCE SYSTEM r Dr NB NS Z PS PB NB AL1 AL1 AL1 AL0.8 AL0.6 NS AL1 AL0.6 AL0.2 AL0.6 AL0.4 Z AL0.4 AL0.2 AL0 AL0.2 AL0.4 PS AL0.4 AL0.6 AL1 AL0.6 AL1 PB AL0.6 AL0.8 AL1 AL1 AL1 - FES3: r (7MF) and dr (7MF) In this case, residual associated with seven symbolic labels and dr with seven symbolic labels. The number of rules n=49. Table 6 is an illustration of the inference. TABLE VI. FES3: INFERENCE SYSTEM r dr NB NM NS Z PS PM PB NB AL1 AL1 AL1 AL1 AL0.8 AL0.8 AL0.6 NM AL1 AL1 AL0.8 AL0.6 AL0.6 AL0.4 AL0.4 NS AL0.8 AL0.6 AL0.4 AL0.4 AL0.2 AL0.2 AL0 Z AL0.4 AL0.2 AL0 AL0 AL0 AL0.2 AL0.4 PS AL0 AL0.2 AL0.2 AL0.4 AL0.4 AL0.6 AL0.8 PM AL0.4 AL0.4 AL0.6 AL0.6 AL0.8 AL1 AL1 PB AL0.6 AL0.8 AL0.8 AL1 AL1 AL1 AL1 - Comparative study and results To compare and define conclusions, we fix the universe of discourse inputs and output: r is in [-2 2], dr is in [-8 8]. We should mention that these intervals depend on studied situations and variables. Assuming that the construction of three inference system obeyed to the same logic which is: When residual is zero the variable is normal. Otherwise, negative or positive values are synonyms of abnormality. Derivative magnitude informs about residual evolution. Next figures 7, 8 and 9 illustrate 3 d response of three FES. Fig. 7. FES1: 3D cf(r,dr) Fig. 8. FES2: 3D cf(r,dr) Fig. 9. FES3: 3D cf(r,dr) We could notice that the three systems have the same evolution: cf is around zero when residual is null and it raised to reach 1 when residual absolute value rises. However, when number of MF is important system response is more slow and soft. Let’s study with precision the three fuzzy expert systems for minimal and for important variations of residual. For minimal variations of residual, we remark that the three systems don’t reach zero even when cf is equal to 0. Minimal value is 0.06: this fact is justified by the uncertainty of measures and of information. Around zero the third system is more precise and the confidante zone, where the variable is normal, is larger than the other FES. TABLE VII. FES1: CF FOR SMALL VARIATIONS OF RESIDUES r dr -0.2 -0.1 -0.05 0 0.05 0.1 0.15 0.2 -8 0,27 0,33 0,32 0,32 0,23 0,24 0,26 0,27 0,23 0,2 0,19 0,17 0,12 0,06 0,12 0,17 0,19 0,2 0,2 0,19 0,17 0,12 0,06 0,12 0,17 0,19 0,2 0,2 0,19 0,17 0,12 0,06 0,12 0,17 0,19 0,2 0,2 0,19 0,17 0,12 0,06 0,12 0,17 0,19 0,2 0,2 0,19 0,17 0,12 0,06 0,12 0,17 0,19 0,2 0,21 0,24 0,22 0,19 0,16 0,24 0,26 0,27 0,24 0,23 0,27 0,26 0,24 0,23 0,32 0,32 0,33 0,27 -6 -4 -2 0 2 4 6 8 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 222 | P a g e www.ijacsa.thesai.org TABLE VIII. FES2: CF FOR SMALL VARIATIONS OF RESIDUES TABLE IX. FES3: CF FOR SMALL VARIATIONS OF RESIDUES In this case, we could conclude on the fact that first and second systems are more suitable for critical situations where little variations of residues are significant for system safety. However, the third system could be best used in non critical situations. TABLE X. FES1: CF FOR IMPORTANT VARIATIONS OF RESIDUES TABLE XI. FES2: CF FOR IMPORTANT VARIATIONS OF RESIDUES TABLE XII. FES3: CF FOR IMPORTANT VARIATIONS OF RESIDUES For important variation of residual, three systems are reaching their maximum value which is 0.93. This coefficient is synonym of evident abnormal situation for the considered variable. To highlight these results, FES proposed must be applied on systems from different fields. The next section is reserved to three applications of FES for surveillance: industrial and medical one. V. STUDY CASE The aim of this work is to propose generic fuzzy expert system that could monitor several kinds of situations. A. Industrial application : Three tanks system The system under consideration is a pilot plant of the research unit: System analysis and command located in ENIT (National Engineer Institute of Tunisia). The considered system is composed of three interconnected cylindrical tanks, two pumps, six valves, pipes, water reservoir in the bottom, measurement of liquid levels and other elements. The pumps pump water from the bottom reservoir to the top of the left and right tanks. Fig. 10. Three tanks model - System Modeling While tanks 1, 2 and 3 are identical with cross section S and maximum fluid level lmax, Drain tank is characterized with cross section Sd and maximum fluid level ldmax. Tanks 1 and 3 are coupled with tank 2 by two AON (all-or-none) valves with cross section Sn and outflow coefficients. Two proportional valves EV1 and EV2 directly connected to a pump, with highest possible flow rate denoted qmax supply tanks 1 and 2. Three sensors are installed to measure the three levels l1, l2 and l3. The experimental plant that is equipped with sensors r dr -0.2 -0.1 -0.05 0 0.05 0.1 0.15 0.2 -8 0,45 0,38 0,27 0,26 0,10 0,26 0,27 0,33 0,4 0,4 0,3 0,2 0,17 0,06 0,17 0,2 0,3 0,4 0,4 0,3 0,2 0,17 0,06 0,17 0,2 0,3 0,4 0,4 0,3 0,2 0,17 0,06 0,17 0,2 0,3 0,4 0,4 0,3 0,2 0,17 0,06 0,17 0,2 0,3 0,4 0,4 0,3 0,2 0,17 0,06 0,17 0,2 0,3 0,4 0,4 0,31 0,24 0,26 0,15 0,26 0,24 0,34 0,42 0,4 0,33 0,27 0,32 0,22 0,32 0,27 0,38 0,45 -6 -4 -2 0 2 4 6 8 r dr -0.2 -0.1 -0.05 0 0.05 0.1 0.15 0.2 -8 0,52 0,39 0,31 0,22 0,22 0,20 0,17 0,24 0,33 0,44 0,29 0,22 0,13 0,12 0,13 0,17 0,24 0,38 0,4 0,24 0,17 0,06 0,06 0,06 0,17 0,24 0,4 0,4 0,24 0,17 0,06 0,06 0,06 0,17 0,24 0,4 0,4 0,24 0,17 0,06 0,06 0,06 0,17 0,24 0,4 0,38 0,24 0,17 0,13 0,12 0,13 0,22 0,29 0,44 0,35 0,24 0,17 0,18 0,18 0,18 0,28 0,36 0,49 0,33 0,24 0,17 0,20 0,22 0,22 0,31 0,39 0,52 -6 -4 -2 0 2 4 6 8 r dr -2 -1.5 -1 0 0.5 1 1.5 2 -8 0,93 0,92 0,93 0,92 0,93 0,82 0,8 0,7 0,6 0,92 0,71 0,71 0,50 0,44 0,54 0,7 0,6 0,5 0,93 0,71 0,6 0,4 0,2 0,4 0,6 0,5 0,4 0,2 0,19 0,17 0,12 0,06 0,12 0,17 0,19 0,2 0,4 0,4 0,4 0,47 0,5 0,47 0,4 0,49 0,57 0,4 0,5 0,6 0,71 0,93 0,71 0,6 0,71 0,93 0,5 0,6 0,7 0,72 0,92 0,71 0,71 0,71 0,92 0,6 0,7 0,8 0,82 0,93 0,92 0,93 0,92 0,93 -6 -4 -2 0 2 4 6 8 r dr -2 -1.5 -1 0 0.5 1 1.5 2 -8 0,93 0,93 0,92 0,93 0,93 0,87 0,82 0,80 0,8 0,92 0,77 0,71 0,71 0,71 0,65 0,62 0,60 0,6 0,93 0,79 0,71 0,64 0,6 0,55 0,5 0,44 0,4 0,2 0,19 0,17 0,12 0,06 0,12 0,17 0,19 0,2 0,3 0,34 0,37 0,40 0,42 0,46 0,47 0,47 0,44 0,4 0,44 0,5 0,55 0,6 0,64 0,71 0,79 0,93 0,6 0,60 0,62 0,65 0,71 0,71 0,71 0,77 0,92 0,8 0,80 0,82 0,87 0,93 0,93 0,92 0,93 0,93 -6 -4 -2 0 2 4 6 8 r dr -2 -1.5 -1 0 0.5 1 1.5 2 -8 0.94 0.93 0.92 0.93 0.94 0.80 0.80 0.75 0.60 0.93 0.93 0.83 0.76 0.65 0.65 0.55 0.51 0.45 0.83 0.71 0.63 0.60 0.50 0.40 0.40 0.29 0.29 0.70 0.54 0.44 0.35 0.35 0.24 0.19 0.24 0.20 0.40 0.25 0.17 0.07 0.06 0.07 0.17 0.25 0.40 0.20 0.24 0.19 0.24 0.35 0.35 0.44 0.54 0.70 0.45 0.51 0.55 0.65 0.65 0.76 0.83 0.93 0.93 0.6 0.75 0.80 0.80 0.94 0.93 0.92 0.93 0.94 -6 -4 -2 0 2 4 6 8 D r a in T a n k T a n k 1 T a n k 2 T a n k 3 P u m p 2 E V 1 2 E V 3 2 E V b 1 E V b 3 l1 l3 ld q 1 2 q 3 2 q b 1 q 2 q b 3 1 l2 P u m p 1 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 223 | P a g e www.ijacsa.thesai.org and actuators, communicates via data acquisition system with a personal computer. Because the modeled system presents much non linearity, we’ve tried to model each component in separate block. All individual parts models were incorporated into single block in Matlab/Simulink environment. The block has 11 inputs: 2 float signals controlling the pumps ( 1 and 2 ) and 9 Boolean signals controlling the valves (EV12, EV32 and EV2 ). Besides it has 3 float signals outputs from water level heights.[58] Fig. 11. L1/l2/l3 optimal variations - FES application and results: Measurable variables that could inform about system state are summarized in table 13. After fixing studied variables, residues and their variations are calculated. In this system we are installing three FES associated with all variables. In this section, we suppose that studied system is having an actuator failure in the tank 1 that leaks. Tank1 leaking is happening at the 70 th second and its value is equal to 0.02. We are studying in the following paragraphs three FES responses to simulated failure. Figure 12 is an illustration of certitude coefficient calculated for tank1 level by both 3 proposed FES. TABLE XIII. MEASURABLE VARIABLES It is remarkable that they respond with the same shape and variations while they obey to the same logic. Alarm is generated immediately when cf is superior to 0.2. Figure 13 illustrates that: FES1 generates alarm at 70.27 s time, FES2 at 70.275 s and FES3 at 70.29 s. However, FES1 is having most important maximum value (cf reaches 0.625 ) and FES 2 and FES 3 reaches 0.6 for maximum cf value. These results mean that an alarm is generated in appropriate time for variable l1 that is immediately related with leak. Fig. 12. Cf(t) (-:FES1, -: FES2, -:FES3) Fig. 13. Cf(t) maximum and minimum values(-:FES1, -: FES2, -:FES3) Looking at first alarm generation time and alarm maximum severity factor, fault localization is evident. 0 100 200 300 400 500 600 700 800 900 1000 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 t(s) h 1 ( m ) / h 3 ( m ) 0 50 100 150 200 250 300 350 400 450 500 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 t(s) h 2 ( m ) Variables Name Normal values Range residues Associate anomalies Tank1 level l1 h1 optimal variations [-1 1] f1: Tank1 leak f2: EV12 failed f3:Pump1 failed Tank2 level l2 l2 optimal variations [-1 1] f4:Tank2 leak f5:EV12 failed f6:EV23 failed Tank3 level l3 l3 optimal variations [-1 1] f7:Tank3 leak f8:EV23 failed f9:Pump3 failed (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 224 | P a g e www.ijacsa.thesai.org Variable l1(t) that monitors Tank 1 is the most and the first affected one. Consequently, this decision support system guides human operator in his decision. To diagnosis the current situation, a normalized fault signature depending on calculated certitude coefficient is proposed. Having a vector r(k) (k in {1,2,3}) describing respectively residues of the three tanks levels, we define a normalized vector rn(k) : n A failure signature matrix could indicate about the incidence of failures on residues: In this study case, we conclude that the generic fuzz expert system applied to industrial field provides a decision support system. It detects failures and generates alarms with severity or certitude factors in one hand. In another hand, it helps locating failure origin root based on certitude factor and first alarm generation time. Industrial field supposes that n (number of variables characterizing system) FES has to be installed. These FES are running in real time while the studied process is functional which makes the first FES the more suitable to be applied because of its time response and time execution. TABLE XIV. MEASURABLE VARIABLES B. Medical application : ECG analysis An electrocardiogram (ECG) is a simple and commonly performed test that records the electrical activity of the heart. An ECG is used to measure the rate and rhythm of the heart. It is a useful investigation in screening for heart disease and for those people who have a cardiovascular disorder. An ECG can show the presence of any damage to the heart, although not all heart conditions can be detected by an ECG. Fig. 14. ECG signal Table 14 summarizes the inputs of this module which are deduced from the signal in figure 14 of ECG. The table gives also normal values of the inputs. Limits and thersholds for normal values are those of an athlete [61]. Fig. 15. ECG monitoring system The calculation of residue is as the folowing equations shows : If v ϵ [lmin – lmax] then r = 0 If v < lmin or then r = v- lmin v > lmax then r = v-lmax Considering that ti is the actual date of analysis, ti-1is the last one, so we could calculate: dr i = [ri - ri-1]/ti - ti-1 TABLE XV. VARIABLES CHARECHTERIZING ECG First alarm generation time Alarm max certitude/s everiy Alarm persistence mean time Fault localization l1 FES1 :70.27s FES2 :70.275s FES3:70.29s 62% 60% 60% 1.04 s Failure root l2 FES1 :75s FES2 :76s FES3:77s 26% 19% 23% 0.06 s Failure propagation l3 FES1 :101s FES2 :102s FES3:104s 21% 22% 22% 0.052s Failure propagation (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 225 | P a g e www.ijacsa.thesai.org - FES application and results While ECG is normal, the three fuzzy expert systems response is illustrated in table 15. TABLE XVI. RESULTS WITH NORMAL ECG V a r ia b le s R e sid u e R e sid u e s d e r iv a tiv e c f S E F 1 c f S E F 2 c f S E F 3 HR=110 0 0 0.063 0.063 0.063 QRS=0.1 0 0 0.063 0.063 0.063 QTc=0.4 0 0 0.063 0.063 0.063 Pa=2 0 0 0.063 0.063 0.063 Pd=0.02 0 0 0.063 0.063 0.063 Qa=2 0 0 0.063 0.063 0.063 Qd=0.035 0 0 0.063 0.063 0.063 Axis=35 0 0 0.063 0.063 0.063 Ra=3 0 0 0.063 0.063 0.063 ST=0.51 0 0 0.063 0.063 0.063 Sa= 0.5 0 0 0.063 0.063 0.063 ECG Normal - cf 94 % 94 % 94 % Time response 0.0180 0.021 0.029 Precision 0.063 0.063 0.063 TABLE XVII. RESULTS WITH ABNORMAL ECG The whole system output is in this case the maximum between eleven calculated cf. If system output is superior then 0.2, ECG is abnormal. In the same way, ECG diagnosis could be done using heart anomalies signatures. In this case signature is equal to [0 1 1 0 0 0 0 0 0 0 0 ] which is equivalent to “Complete bundle brunch block” VI. CONCLUSION This research work aims to study fuzzy expert system for monitoring. These support decision systems are very used in several applications and fields. A state of art has proved that developed fuzzy expert systems are usually specific to studied process whether in variables choice or in inference logic. Proposing generic fuzzy expert system for surveillance independently of its application had been the subject of this paper. The main idea is to characterize concerned application with measurable variables. Fuzz expert system is installed to monitor each one by analyzing the distance between variable and its optimum behavior (error or residue). Three generic fuzzy expert systems were proposed. The number of membership functions describing error and its variations differentiates between different proposed systems. Two criteria had been discussed time response and incertitude minimization. Increasing membership functions improves precision and describes better each variable variation. However, system time response rises which is annoying especially in real time. When applying FES in industrial and medical diagnosis, results confirms that it provides decision support systems that detects abnormal situations and affects certitude coefficient enhancing uncertainty. VII. FUTURE WORKS Diagnosis process which is in our approach based on anomalies signatures could be improved by using artificial V a ria b le s N a m e N o rm a l V a lu e s fo r a n a th le te R a n g e o f r a n d d r A sso c ia te d A n o m a lie s o fl E C G Heart rate HR [45 – 150] r [-4 4] dr [-2 2] Bradycadie Tachycardie QRS QRS QRS<0.12s r [-0.1 0.1] dr [-2 2] Complete bundle brunch block QTc QTc 0.34>QTc> =0.48s r [-0.1 0.1] dr [-2 2] Short QTc interval Long QTc interval P wave amplitude Pa Pa<2.5mm r [-1 1] dr [-2 2] Right atrial enlargement P wave duration Pd Pd<0.04s r [-0.1 0.1] dr [-2 2] left atrial enlargement Q wave amplitude Qa Qa<3 mm r [-1 1] dr [-2 2] Pathologic Q wave Q wave duration Qd Qd<0.04s r [-0.1 0.1] dr [-2 2] Pathologic Q wave Axis Ax - 90°<Ax<12 0° r [-4 4] dr [-2 2] -Left axis deviation -Right ventricular hypertrophy R wave amplitude Ra Ra<5 mm r [-1 1] dr [-2 2] Pathologic R wave ST segment ST ST<1 mm r [-1 1] dr [-2 2] ST segment depression T wave amplitude Sa Sa<1 mm r [-1 1] dr [-2 2] T wave inversion V a r ia b le s c f S E F 1 c f S E F 2 c f S E F 3 HR=160 0.063 0.063 0.063 QRS= 0.18 0.600 0.611 0.685 QTc=0.483 0.2758 0.334 0.319 Pa=2 0.063 0.063 0.063 Pd=0.02 0.063 0.063 0.063 Qa=2 0.063 0.063 0.063 Qd=0.035 0.063 0.063 0.063 Axis=35 0.063 0.063 0.063 Ra=3 0.063 0.063 0.063 ST=0.51 0.063 0.063 0.063 Sa= 0.5 0.063 0.063 0.063 ECG Normal - cf 40 % 39 % 32 % Time response 0.0190 0.018 0.025 ECG certitude coefficient 0.600 0.611 0.685 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 226 | P a g e www.ijacsa.thesai.org neural networks (ANN). Multi layer perceptron (MLP) is a suitable tool for anomalies classification that had been used in many applications. REFERENCES [1] Dubois D., Prade H., “An Introduction to fuzzy systems,” Clinica Chimica Acta, Vol. 270, pp. 3-29, 1998. [2] Bellman R. E., Zadeh L., “Decision making in a fuzzy environment,” Manage Sc, Vol. 17, pp. 141-146, 1970. [3] Dubois D., Fargier H., Prade H.“ Refinements of the maximum approach to decision making in a fuzzy environment,” Fuzzy Sets and Systems, Vol. 81, pp. 3-29, 1996. [4] Sahin S., Tolun M.R., Hassanpour R., “Hybrid expert systems: A survey of current approaches and application,” Expert Systems with Applications, vol. 39, p. 4609–4617, 2012. [5] Siler W., Buckley J., “Fuzzy Expert Systems and Fuzzy Reasoning,” John Wiley & Sons, Inc., New Jersey, 2005. [6] Venkatasubramanian V., Rengaswamy R., S.N. Kavur, Y. Kewen, “A review of process fault detection and diagnosis Part II: Qualitative Models and search strategie,” Computers and Chemical Engineering, Vol. 27, pp. 313-326, 2003. [7] A. Evsukoff, S., Gentil, J., Montmain, “Fuzzy reasoning in co-operative supervision systems,” Control Engineering Practice, Vol. 8, pp. 389- 407, 2000. [8] Németh B., Laboncz S., Kiss I., Csépes G, “Transformer condition analyzing expert system using fuzzy neural system,” IEEE International Symposium on Electrical Insulation (ISEI), Canada, 2010. [9] Yun-Seong L., Hyung-Chul K., Jun-Min C., Jin-O, “A New Method for FMECA Using Expert System And Fuzzy Theory,” PMAPS, 2010. [10] H.C.W. Lau, R.A. Dwight, “A fuzzy-based decision support model for engineering asset condition monitoring – A case study of examination of water pipelines,” Elsevier -Expert Systems with Applications, vol. 38, pp. 13342–13350, 2011. [11] Abdelazeem A., Abdelsalam A., Eldesouky Abdelhay, A., Sallam, “ Characterization of power quality disturbances using hybrid technique of linear Kalman filter and fuzzy-expert system,” Elsevier -Electric Power Systems Research, vol. 83, pp. 41– 50, 2012. [12] Mohsen Naderpour, Jie Lu, “A Fuzzy Dual Expert System for Managing Situation Awareness in a Safety Supervisory System,” WCCI 2012 IEEE World Congress on Computational Intelligence, pp. 10 -15, 2012. [13] Petr Chalupa, Jakub NOVÁKovak, Vlademir Bonbal, “ Detailed Simulink Model of Real Time Three Tank System,” Recent Researches in Circuits, Systems, Communications and Computers, pp. 161-166, 2010. [14] Abdelaziz, A.Y., Mekhamer, S.F., Nada, M.H., “A fuzzy expert system for loss reduction and voltage control in radial distribution systems,” Elsevier - Electric Power Systems Research, vol. 80, p. 893–897, 2010. [15] Angulo C., Cabestany J., Rodríguez P., Batlle M., González A., de Campos S., “Fuzzy expert system for the detection of episodes of poor water quality through continuous measurement,” Elsevier -Expert Systems with Applications, vol. 39, pp. 1011–1020, 2012. [16] Dan-yang C., Shui-ping Z., Jin-hong L., “Variable universe fuzzy expert system for aluminum electrolysis,” Elsevier -Trans. Nonferrous Met. Soc, vol. 21, p. 429-436, 2011. [17] Krivoulya G., Dudar Z., Kucherenko D., Mehana S., “Fuzzy Expert System for Diagnosis of Computer Failures,” CADSM’2009, 24-28 February, Polyana-Svalyava, UKRAINE, p. 225-230, 2009. [18] Chang-Shing Lee, Mei-Hui Wang, “A Fuzzy Expert System for Diabetes Decision Support Application,” IEEE Transaction on Systems Man and Cybernetics - PART B: Cybernetic, Vol. 41, N°. 1, pp. 139- 153, 2011. [19] Azian Azamimi Abdullah, Zulkarnay Zakaria and Nur Farahiyah Mohammad, “Design and Development of Fuzzy Expert System for Diagnosis of Hypertension,” Second International Conference on Intelligent Systems, Modelling and Simulation (ICISMS 2011), pp. 113- 117, 2011. [20] M. Neshat, M. Yaghobi, M. B. Naghibi, A. Esmaelzadeh, “Fuzzy Expert System Design for Diagnosis of liver disorders,” International Symposium on Knowledge Acquisition and Modeling (ISKAM 2008), pp. 252-256, 2008. [21] M.J.P. Castanho, F. Hernandes, A.M. De Ré, S. Rautenberg, A. Billis, “Fuzzy expert system for predicting pathological stage of prostate cancer,” Elsevier -Expert Systems with Applications, 2012. [22] Ali Adeli & Adeli Neshat, “A Fuzzy Expert System for Heart Disease Diagnosis,” Proceedings of the international multi conference of engineering and computer scientists, hong kong, mars 17-19 -2010. [23] Ali Keles, Aytürk Keles, Ugur Yavuz, “Expert system based on neuro- fuzzy rules for diagnosis breast cancer,” Expert Systems with Applications, vol. 38, pp. 5719–5726, 2011. [24] I.M. Dokas, D.A. Karras, D.C. Panagiotakopoulos, “Fault tree analysis and fuzzy expert systems: Early warning and emergency response of landfill operations,” Environmental Modelling & Software, vol. 24, pp. 8–25, 2009. [25] M.H. Fazel Zarandi, P. Ahmadpou, “Fuzzy agent-based expert system for steel making process,” Expert Systems with Applications, vol. 36, pp. 9539–9547, 2009. [26] Chou S., Chang Y., Shen, C., “A fuzzy simple additive weighting system under group decision-making for facility location selection with objective/subjective attributes,” European Journal of Operational Research, vol. 198, pp. 132-145, 2008. [27] Dell’Orco M., Circella G., Sassanelli D., “A hybrid approach to combine fuzziness and randomness interval choice prediction,” European Journal of Operational Research, vol. 195, pp. 648-658, 2008. [28] Doukas H. C., Andreas B. M. Psarras J. E., “Multi-criteria decision aid for the formulation of sustainable technological energy priorities using linguistic variables,” European Journal of Operational Research, vol. 182, pp. 844-855, 2007. [29] Deyin Ma., Yanchun, L. Xiaoshe, Z. Renchu G. Xiaohu S, “Multi-BP expert system for fault diagnosis of power system,” Engineering Applications of Artificial Intelligence, vol. 26, pp. 937–944, 2013. [30] O.C. Pires, C. Palma, I. Moita, J.C. Costa, M.M. Alves and E.C. Ferreira, “A fuzzy logic based expert system for diagnosis and control of an integrated wastewater treatment,” 4th Mercosur Congress on Process Systems Engineering, Brasil, 2013. [31] John Farrell, Abraham Kandel, “A fuzzy rule-based expert system for marine bioassessment,” Fuzzy Sets and Systems, vol. 89, pp. 27 -34, 1997. [32] G. Hong, X. Chen, X. Xue, S. Zhang, “Expert Systems for Fault Diagnosis Integrating Neural Network and Fuzzy Inference,” International Conference of Information Technology, Computer Engineering and Management Sciences, pp. 245-249, 2011. [33] G. Molnárka, “Management of Uncertainty in Visual Examination Procedure in Building Diagnostics with Fuzzy Expert System,” ISCIII 2009: 4th International Symposium on Computational Intelligence and Intelligent Informatics, pp. 21–25, Egypt, 2009. [34] LI Jie, SHEN Shi-tuan, “Research on the Algorithm of Avionic Device Fault Diagnosis Based on Fuzzy Expert System,” Chinese Journal of Aeronautics, Vol. 20, pp. 223-229, 2007. [35] Liu Xiaobo, Li Jianping, “Fault Diagnosis of Fan Based on Fuzzy Neural Expert System,” Second International Conference on Intelligent Computation Technology and Automation, 2009. [36] Chi-Jen Lin a, Wei-Wen Wu, “A causal analytical method for group decision-making under fuzzy environment,” Expert Systems with Applications, vol. 34, pp. 205–213, 2008. [37] Jacky Montmain, Sylviane Gentil, “Dynamic causal model diagnostic reasoning for online technical process supervision,” Automatica, vol. 36, pp. 1137-1152, 2000. [38] Bodapati Nageswararao & B. Jeyasurya, “Fuzzy-expert system for voltage stability monitoring and control,” Electric Power Systems Research vol. 47 pp. 215–222, 1998. [39] Burak Ozyurt, Abraham Kandel, “A hybrid hierarchical neural network- fuzzy expert system approach to chemical process fault diagnosis,” Fuzzy Sets and Systems, vol. 83 pp. 11- 25, 1996. [40] P. Baraldi, M. Librizzi, E. Zio, L. Podofillini, V.N. Dang, “Two techniques of sensitivity and uncertainty analysis of fuzzy expert (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 8, 2013 227 | P a g e www.ijacsa.thesai.org systems,” Expert Systems with Applications, vol. 36, pp. 12461–12471, 2009. [41] Palluat N., Racoceanu D., Zerhouni N., “A neuro-fuzzy monitoring systemApplication to flexible production systems,” Computers in Industry, vol. 57, pp. 528–538, 2006. [42] Yvonne Power & Parisa A. Bahri, “A two-step supervisory fault diagnosis framework,” Computers and Chemical Engineering, vol. 28, pp. 2131–2140, 2004. [43] Ling Wang, Jian Chu, Jun Wu, “Selection of optimum maintenance strategies based on a fuzzy analytic hierarchy process,” Int. J. Production Economics, vol. 107, pp. 151–163, 2007. [44] Zhang Quan, Chen Nanyu, Huang Jun, Meng Zhijun, “Application of Expert System Fuzzy BP Neural Network in Fault Diagnosis of Piston Engine,” International Conference on Computer Science and Electronics Engineering, pp.604-607, 2012. [45] M. Kalpana, A.V Senthil Kumar, “Fuzzy Expert System for Diabetes using Fuzzy Verdict Mechanism,” Int. J. Advanced Networking and Applications, Vol. 03(02), pp. 1128-1134, 2011. [46] Mahdi Jampour, Mohsen Jampour, Maryam Ashourzadeh, Mahdi Yaghoobi, “A Fuzzy Expert System to Diagnose Diseases with Neurological Signs in Domestic Animal,” 2011 Eighth International Conference on Information Technology: New Generations, 2011. [47] Novruz ALLAHVERDI & Tevfik AKCAN, “A Fuzzy Expert System Design for Diagnosis of Periodontal Dental Disease,” 5 th international conference on application of information and communication technologies, 2011. [48] Oana GEMAN, “A Fuzzy Expert Systems Design for Diagnosis of Parkinson's Disease,” Proceedings of the 3rd International Conference on E-Health and Bioengineering - EHB 2011. [49] Djam, X.Y. and Y.H. Kimbi, “Fuzzy Expert System for the Management of Hypertension,” Pacific Journal of Science and Technology, Vol. 12(1), pp. 390-402, 2011. [50] Miguel L. J., Blazquez L. F., “Fuzzy Logic based decision making for fault diagnosis in a DC motor,” Engineering applications of Artificial Intelligence, vol. 18, pp. 432-450, 2005. [51] Baron L., Achche S. and Balazinski M., “Fuzzy decision support system knowledge base generation using genetic algorithm,” International Journal of approximative reasoning, vol. 28, pp. 125-148, 2011. [52] M. J. P. Castanho, L. C. de Barros, Akebo Yamakami, Lae´rcio Luis Vendite, “Fuzzy expert system: An example in prostate cancer,” Applied Mathematics and Computation, vol. 202, pp. 78–85, 2008. [53] Azian Azamimi Abdullah, Zulkarnay Zakaria and Nur Farahiyah Mohammad, “Design and Development of Fuzzy Expert System for Diagnosis of Hypertension,” Second International Conference on Intelligent Systems, Modelling and Simulation (ICISMS 2011), pp. 113- 122217, 2011. [54] Péter L. Venetianer, Hongli Deng, “Performance evaluation of an intelligent video surveillance system – A case study,” Computer Vision and Image Understanding, vol. 114, pp. 1292–1302, 2010. [55] Beuthe M., Eeckhoudt L., Scannella G., “A practical multicriteria methodology for assessing risky public investments,” Socio-Economic Planning Sciences, vol. 34, pp. 121-139, 2000. [56] Kunch, P. L., Fortemps, P., “A fuzzy decision support system for the economic calculus in radioactive waste management,” Information science, vol. 142, pp. 103-116, 2002. [57] Wang J., & Lin Y., “A fuzzy multicriteria group decision making approach to select configuration items for software development,” Fuzzy sets and systems, vol. 134, pp. 343-363, 2003. [58] Najar Yosra, Ketata Raouf, Ksouri Mekki, “Fuzzy Expert System for Residual Analysis,” Journal of Information Organization, Volume 3 Number 1, pp. 23-35, 2013. [59] Anup Kumar Panda, Suresh Mikkili. “FLC based shunt active filter (p– q and Id–Iq) control strategies for mitigation of harmonics with different fuzzy MFs using MATLAB and real-time digital simulator,” Electrical Power and Energy Systems, vol. 47, pp. 313–336, 2013. [60] Alexandros Iosifidis, Anastasios Tefas, Nikolaos Nikolaidia, Ioannis Pitas, “Multi-view human movement recognition based on fuzzy distances and linear discriminant analysis,” Computer Vision and Image Understanding, vol. 116, pp. 347–360, 2012. [61] Jonathan A Drezner, “Standardised criteria for ECG interpretation in athletes: a practical tool,” Br J Sports Med, Vol 46, 2012. 
work_cyju22gu4nbqnkszvac5a5m3oy	----	Simple knowledge structures for perceptual classification expert systems Behavior Research Methods. Instruments. & Computers 1990. 22 (2). 208-2/1 Simple knowledge structures for perceptual classification expert systems RICHARD BLOCH and SOREL BOSAN College of William and Mary and Eastern State Hospital Williamsburg, Virginia A new approach is described for the development and structure of perceptual knowledge-based systems. This systematic method for acquiring perceptual knowledge for use in knowledge-based systems and for representing perceptual knowledge within the systems has revealed that two broad classes of perceptual activity can each be characterized by a single logical operator. The number of rules necessary to accomplish particular perceptual tasks can also be estimated. An effort to develop techniques for automating the differential diagnosis of movement disorders led to an examination of methods for acquiring perceptual knowl- edge in a form that could be incorporated readily into a knowledge-based system. The potential for automated differential diagnosis of involuntary movement disorders such as Parkinson's disease, Huntington's chorea, tardive dyskinesia, or cerebral palsy revolved around the ability of a computer to acquire information about the parameters of movement and to be able to distinguish between nor- mal and various forms of abnormal movement. To acquire data about parameters of movement such as velocity, direction, and extent requires the ability to "see" the movement either through computer vision or through analog-to-digital conversion of voltage changes proportional to the movement. Once movement data is gathered, by an image processor, through psychophysio- logical measures such as EMG, or from accelerometers, the data must be reduced and interpreted if the movements are to be classified as normal or abnormal, or as one par- ticular disordered movement type. The same data acquisi- tion and interpretation steps are necessary for any percep- tual classification problem, whether it involves sexing baby chickens, industrial quality-control monitoring for prod- ucts such as ice cream or perfume, or voice recognition. There are at least two approaches to automated clas- sification: neural networks and expert systems. Neural networks model human learning at the neuronal level, making and strengthening connections as a function of the association of input characteristics with response out- comes. Successful implementations generally require many repeated exposures to the stimulus situations that must be discriminated. But with stimulus types that are relatively rare, this can be problematic. Expert systems, on the other hand, incorporate the knowledge of human experts into knowledge structures that enable the computer to utilize the information to make Sorel Bosan is now at Vanderbilt University. Nashville. TN. Cor- respondence may be addressed to Richard Bloch, Eastern Stale Hospi- tal, Drawer A, Williamsburg, VA 23187. judgments. Knowledge acquisition for expert-system de- velopment has traditionally relied on verbal interviews, protocol analysis, and observation of selected experts. A knowledge engineer translates the experts' verbal descrip- tions of, and behaviors toward, decision-making variables into a knowledge-based structure, such as rules, and com- bines the rules in decision-making structures to emulate the expert's decisions. In knowledge domains where this approach can be used, the knowledge acquisition has proven to be a bottleneck for the development of expert systems (Hayes-Roth, Waterman, & Lenat, 1983). Cur- rent research to reduce the bottleneck is focused on the use of automated interviews (Gruber & Cohen, 1987) and on interfaces that gather knowledge directly from the ex- pert, bypassing the translation of the expert's knowledge by the knowledge engineer (Kitto & Boose, 1987). Despite current efforts to improve knowledge-acquisition methods, these methods are quite difficult to apply to per- ceptual domains. An alternative approach has been de- scribed and tested (Bosan, 1989), which utilizes interactive simulations to acquire knowledge. In this technique, ex- perts are given a computer-controlled simulation of the stimulus situation of interest, with values of key features of the stimulus determined by the expert's use of inter- active controls. Data obtained from an expert on each dimension represents that expert's judgment that, after ap- propriate adjustment to match a selected stimulus condi- tion, the final adjusted simulation value represents the ex- pert's estimation of that dimension for that stimulus condition. With repeated trials, the extent of adjustment variability can be indexed. Thus, for example, in adjusting the tempo of movement of a computer-controlled graphi- cal display of a particular body part to match the stimu- lus condition Parkinsonian tremor, an expert neurologist finishes changing the tempo adjustment when it matches his or her knowledge of Parkinsonian tremor tempo. Repeated adjustments for tremor tempo would provide a basis for establishing a distribution of scores that could be standardized to provide their associated probabilities. Rules are developed as a function of the number of dimensions and the number of conditions to be discrimi- Copyright 1990 Psychonomic Society, Inc. 208 nated. There are two ways to generate rules. One way is to create a single rule for each stimulus type. Thus, in the example above, there would be a single rule for Parkinsonian tremor, with clauses for every dimension, such as tremor tempo. Because each simulated dimension would be adjusted several times, the rule could have a range of true values that would be as inclusive or exclusive as desired. Each discriminated condition would have the same number of clauses in its single rule, and each clause would have the same structure. To extend the example begun above, in discriminating between Parkinsonian tremor (PT) and other movement conditions, each condi- tion would have a single rule that had the following form: IF: [value of dimension(tempo) in test stimulus] 2: [lower boundary of simulated condition(PT) dimension(tempo)] and :s [upper boundary of simulated condition(PT) dimension(tempo)] AND [value of dimension(2) in test stimulus] 2: [lower boundary of simulated condition(PT) dimen- sion(2)] and :S [upper boundary of simulated condition(PT) dimension(2)] AND [value of dimension(N) in test stimulus] 2: [lower boundary of simulated condition(PT) dirnen- sion(N)] and :S [upper boundary of simulated condition(PT) dimension(N)] THEN: [test stimulus] = [condition(PT)] IF: [value of dimension(tempo) in test stimulus] 2: [lower boundary of simulated condition(2) dimen- sion(tempo)] and :S [upper boundary of simu- lated condition(2) dimension(tempo)] AND [value of dimension(N) in test stimulus] 2: [lower boundary of simulated condition(2)] and :S [upper boundary of simulatedcondition(2) dimension(N)] AND THEN: [test stimulus] = [condition(N)] Although each rule has the same number and form of clauses, the rule structure has no particular relationship to the conditions being discriminated. The rule structure represents a general form for containing perceptual judg- ment ranges set by human experts. Also note that each clause is ANDed within a stimulus condition. For a rule to be true, each of the dimension clauses within the rule for the stimulus condition must be true. This is also a con- stant that applies across stimulus conditions and represents all classification/diagnostic problems. The use of a range could be based on the statistical dis- tribution of the experts' judgments. In the initial test of this method, the range was arbitrarily determined by the mean plus and minus two standard deviations. The range PERCEPTUAL EXPERT SYSTEMS 209 could, however, be based on a variety of other statistical functions. With this single rule-per-stimulus-condition format, a conflict resolution strategy may be required to make a final decision. If, in the example above, more than one rule were true, this strategy would be used to determine the output. Furthermore, this strategy could be represented as an additional rule. This rule could decide that if a test stimulus yields more than one stimulus condition to be true, or if no stimulus condition rules are found to be true, then the test stimulus condition is unknown or is a stimu- lus condition other than those simulated. With a single rule for each stimulus condition, the num- ber of rules equals the number of discriminated condi- tions plus the final conflict resolution rule. Given the capa- bilities of today's rulebase shells, the limiting factor in developing rulebases that classify dozens, or even hundreds, of perceptual stimuli becomes the time required for expert input. If, as with neural networks, this method of acquiring and organizing knowledge is scalable, the multiple smaller perceptual rulebases might be able to be combined without requiring expert reexamination of per- ceptual simulations. An alternative to single rules per stimulus condition is to break each clause in the single rule example into its own rule. This would mean that for every stimulus con- dition being classified or discriminated, there would be the same number of rules as there are dimensions being adjusted. Using the same example of Parkinsonian tremor, this approach would produce a rule base in the form of: IF: [value of dimension(tempo) in test stimulus] 2: [lower boundary of simulated condition(PT) dimension(tempo)] and :S [upper boundary of simulated condition(PT) dimension(tempo)] THEN: [test stimulus dimension(tempo)] = [PT dimen- sion (tempo)] IF: [value of dimension(2) in test stimulus] ~ [lower boundary of simulated condition(PT) dimen- sion(2)] and :S [upper boundary of simulated condition(PT) dimension(2)] THEN: [test stimulus dimension(2)] = [PT dimension(2)] IF: [value of dimension(N) in test stimulus] 2: [lower boundary of simulated condition(condN) dimen- sion(N)] and :S [upper boundary of simulated condition(condN) dimension(N)] THEN: [test stimulus dimension(N)] = [condition(condN) dimension(N)] Again, all the rules within the rulebase would have the same basic form, even though they were relevant to differ- ent dimensions of different stimulus conditions. Additional rules would be necessary to make the expert system's final judgment. For each stimulus condition, there would be one rule that combined all the dimensions with a logical AND such as: 210 BLOCH AND BOSAN IF: [test stimulus dimension(tempo)] = [PT dimen- sion(tempo)] AND [test stimulus dimension(2)] = [PT dimension(2)] AND AND [test stimulus dimension(N)] [PT dimen- sion(N)] THEN: [test stimulus] = [PT] These combination rules would all use the logical AND f?r each di.mension, independent of what was being clas- sified or diagnosed. In short, the logical AND, whether within a single rule per condition, or between single rules per condition x dimension combination, characterizes all perceptual classification and diagnostic problems. Again there is one additional conflict resolution strategy rule that is needed to handle the situation in which none of the rules combining dimensions for a stimulus condi- tion are found true or more than one is found true. Thus, when utilizing single rules per dimension, the total num- ber of rules required to classify N stimulus conditions be- comes a function of the formula: no. of rules = (no. of stimulus conditions) x (no. of dimensions) + I. Just as the logical AND applies to classification and di- agnosis problems in using the interactive simulation method to acquire knowledge for perceptual expert sys- tems, it is proposed that the logical OR applies to per- ceptual detection problems. In this context, detection is composed of a search strategy for scanning in space or time for a stimulus event. It is the search rather than the event recognition that is central to the detection process. For example, in scanning a surface for an imperfection, gaze is moved from one location to the next, whereas in listening for a sound, both spatial orientation to potential sources and repeated samplings over time occur. The scanning process essentially ends with the detection of a stimulus event, such as would be signaled by an orient- ing response. The use of interactive simulations to acquire knowledge for a detection expert system is not as well understood as their use to acquire knowledge for classification problems. It can be proposed that the dimensions adjusted by the expert involve parameters or dimensions of the scanning process such as the order in which the segments of the whole scan area are scanned; the size of the in- dividual scan segments; the rate of scanning each segment; and the degree of change from background required to stop scanning and cause an assessment as to whether a stimulus is the sought-after condition. In sum, these dimensions amount to a detection strategy. The rules generated in detection problems are also sug- gested to be simple and systematically structured. Just as with classification problems, each complete search activity within a sensory modality can have a single rule, such as: IF: search segment(l) stimulus change OR search segment(2) = stimulus change OR OR search segment(N) = stimulus change THEN: stimulus change detected Another example would be: IF: search segment size(l) = stimulus change OR search segment size(2) = stimulus change THEN: stimulus change detected Again, it should be noted that in the class of detection activities, logical ORs rather than ANDs link the rule clauses. The number of clauses within a rule is a function of the number of search segments. If the assessment of the stimulus change is not considered part of the detec- tion process, each sensory modality may require a single rule. Thus, the number of rules may be a function of the number of sensory modalities scanned. If more than a sin- gle modality is scanned, the rule structure would become: IF: search segment(l) modality(l) = stimulus change OR search segment(2) modality(l) = stimulus change OR search segment(N) modality(l) = stimulus change THEN: stimulus change modality(1) detected Each additional modality would have a similar rule. The combination rule to decide if an event was detected could again use the logical OR to produce a decision that an event was either detected or not detected; for example: IF: stimulus change modality(l) detected OR stimulus change modality(2) detected OR THEN: stimulus change detected. Returning to the original goal of developing a knowledge- based system capable of discriminating between different movement disorders, it becomes clear that most percep- tual tasks begin with an embedded detection problem. In order to classify movement disorders, movement must first be detected. Once it is detected, the classification process begins. The design of an expert system capable of this sequence of activities would, thus, require a scan- ning strategy based on one or more experts' scanning strategies, and classification rules that are tested once the scanning process produces a stimulus detection. The clas- sification process would determine if the detected stimulus was, for example, a movement or not, and then, if it was judged a movement, would classify the movement type. In this way, detection and classification rules are the basic building blocks of perceptual behavior, and they en- able perceptual behavior to be modeled. Complex per- ceptual expert systems become conceivable as a combi- nation of relatively simple detection and classification problems with consistent rule structures and various se- quences of OR and AND decision structures. Through the building of sequences of detection and classification tasks, the potential is created to develop knowledge-based systems that could detect and classify various objects; for example: 1. Barriers in a robotics environment could be detected and classified in terms of what types of barrier or what types of avoidance behavior should follow. 2. Movement in a security monitoring environment could be detected and classified as resulting from human or other causes. 3. Manufactured objects on an assembly line could be examined for defects; once detected, they would be classi- fied as to whether they were correctable or noncorrectable. In summary, the first systematic method for acquiring perceptual knowledge for use in knowledge-based sys- tems, and for representing perceptual knowledge within knowledge-based systems, has pointed toward surprisingly simple and consistent rule structures. The process has re- vealed that two broad classes of perceptual activity can each be characterized by a single logical operator. In the case of classification and diagnostic activities, the logi- cal operator is an AND, whereas in detection activities, PERCEPTUAL EXPERT SYSTEMS 211 the operator is an OR. It has been proposed that complex perceptual activities can be broken down into a sequence of steps, each step utilizing one of these activity classes, and one of the two operators. The number of rules neces- sary to accomplish particular perceptual tasks can also be estimated, with classification and diagnostic activities be- ing quite efficient in the number of rules necessary for rudimentary discriminations. This represents a new ap- proach to the development and structure of perceptual knowledge-based systems and holds promise for a wide variety of industrial and medical problems. REFERENCES BOSAN, S. (1989). Interactive simulation for knowledge acquisition. Unpublished MA thesis, College of William & Mary, Williams- burg, VA. GRUBER, T. R., &c COHEN, P. R. (1987). Design for acquisillon: Principles of knowledge-system design to facilitate knowledge ac- quisition. International Journal of Man-Machine Studies, 26, 143- 159. HAYES-ROTH, F., WATERMAN, A. D., &c LENAT, D. B. (1983). Build- ing expert systems. Reading, MA: Addison-Wesley. KITTO, C. M., &c BOOSE, J H. (1987). Choosing knowledge acquisi- tion strategies for application tasks. IEEE Proceedings, 8, 96-103. 
work_d37jg3ke6repxmxwbmn2psye6a	----	The development of an expert system for managerial evaluation of internal controls Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) MANAGERIAL EVALUATION OF INTERNAL CONTROLS 103INTELLIGENT SYSTEMS IN ACCOUNTING, FINANCE AND MANAGEMENT Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002 /isaf.246 Copyright © 2004 John Wiley & Sons, Ltd. * Correspondence to: Clyde W. Holsapple, School of Management, Carol M. Gatton College of Business Administration, University of Kentucky, Lexington, KY 40506-0034, USA. E-mail: cwhols@pop.uky.edu THE DEVELOPMENT OF AN EXPERT SYSTEM FOR MANAGERIAL EVALUATION OF INTERNAL CONTROLS CHULEEPORN CHANGCHITa AND CLYDE W. HOLSAPPLEb* a College of Business, Texas A&M University-Corpus Christi, USA b School of Management, Carol M. Gatton College of Business Administration, University of Kentucky, USA SUMMARY Both practitioners and researchers have devoted significant effort to the study of decision aids, especially expert systems, to assist auditors in internal control evaluations. In addition to being used as a decision aid, researchers have long contended that expert systems could be used to train non-expert users. Even though the professional accounting literature makes it clear that responsibility for maintaining an effective internal control system rests with management rather than auditors, the focus to date has been on expert systems aimed at assisting/training auditors, not an organization’s management. In contrast, this study focuses on management as users of an expert system for internal control evaluation. We describe the development process, explain how the resulting system was evaluated, and discuss results of that evaluation. These results suggest that such a system gives a new way to help managers increase effectiveness and efficiency of a critical organizational process: the evaluation of internal controls. Copyright © 2004 John Wiley & Sons, Ltd. 1. INTRODUCTION A number of reviews have been published dealing with the nature of expertise in auditing (Marchant, 1989; Bonner and Pennington, 1991; Fedorowicz et al., 1992; Libby, 1995; Lieb and Gillease, 1996; Weber, 1999). Among these reviews, several have examined experienced auditors’ internal control knowledge (Eining and Dorr, 1991; Libby, 1995). The study and evaluation of internal controls is a problem involving the expertise of well-trained auditors. The internal control evaluation process is characterized by the use of heuristic rules to determine how well the client’s controls support specific assertions for specific accounts (Gadh et al., 1993). Both practitioners and researchers have devoted significant effort to the study of audit decision aids to assist auditors in internal control evaluations (Bailey et al., 1985; Boritz, 1985; Gal, 1985; Cummings and Apostolou, 1987; Cummings et al., 1988; O’Leary and Watkins, 1989; Brown and Phillips, 1990; Murphy, 1990; Graham et al., 1991; Vinze et al., 1991; Messier, 1995). Several researchers suggest that audit judgment can be improved through the use of decision aids (Messier, 1995; Brown and Jones, 1998). Among these decision aids, expert systems have evolved to become practical tools used in aiding decision making (Bonczek et al., 1981; Holsapple and Whinston, 1987; O’Leary, 1988; Biggs and Morrison, 1990; Wong and Monaco, 1995). 104 C. CHANGCHIT AND C. W. HOLSAPPLE Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) In addition to being used as a decision aid, researchers have long contended that expert systems could also be used to train non-expert users (Biggs et al., 1987: Borthick and West, 1987; Ege and Sullivan, 1990; Gal and Steinbart, 1992; Pei and Reneau, 1992; Bonner and Walker, 1994; Steinbart and Accola, 1994; Odem and Dorr, 1995). Prior studies found that subjects who practiced making decisions with the aid of an expert system were better and quicker at reaching decisions than subjects who practiced without the support of the expert system (Oz, 1989; Eining and Dorr, 1991; Fedorowicz et al., 1992). Expert systems for internal control evaluation are generally developed using either professional literature for a knowledge source, or interview or protocol analysis to acquire knowledge from human experts (Meservy et al., 1986; Holsapple and Raj, 1994). The resultant system is able to emulate the results of an auditor’s internal control evaluation process. The professional accounting literature makes it clear that the responsibility for maintaining an effective internal control system rests with management rather than with the auditors (COSO, 1992; Arens and Loebbecke, 1994). However, the focus to date has been on expert systems aimed at assisting or training auditors, not an organization’s management. Most of these systems are designed to help auditors determine the extent of other tests they will perform in conducting an audit. If an auditor determines that the client’s internal control systems are designed properly and are functioning as designed, he or she can reduce direct testing of account balances accordingly (Gadh et al., 1993). Although the literature does not report on expert systems that can facilitate the transfer of internal control knowledge to managers, there are several reasons why such systems deserve to be invest- igated. First, establishment and supervision of internal control systems are the responsibility of management, not external auditors (COSO, 1992). Second, managers have more immediate and detailed insight into operations (including their internal control facets) than external auditors. Third, decisions made by managers often have internal control implications, but they may not be recog- nized or adequately considered due to managers’ insufficient internal control knowledge. Fourth, evaluation of internal control systems ideally should be treated as a continuous process, allowing weaknesses to be prevented or detected as soon as possible. Moreover, in the wake of recent financial scandals, the Sarbanes–Oxley Act of 2002 was enacted to emphasize the importance of vigilant and effective corporate governance participants—including the board of directors, the audit committee, management, internal auditors, external auditors, and governing bodies (e.g. Securities and Exchange Commission, American Institute of Certified Public Accountants)—can play in preventing and detecting financial statement fraud (Rezaee, 2004). The act requires public companies to validate the accuracy and integrity of their financial management. The processes and documentation required for compliance are rigorous: companies must have estab- lished procedures for meeting their reporting obligations, and CEOs and CFOs must personally certify that their own company’s statements are complete and accurate. Having said this, we do not imply that internal control evaluations should be performed by management instead of auditors. Rather, we contend that ongoing attention to internal control issues by management would be a useful complement to ordinary auditing activity. However, this requires that managers possess or have easy access to internal control evaluation knowledge. The expert system described in this paper is aimed at transferring such knowledge to managers. The remainder of this paper begins with an overview of system objectives. This is followed by a description of the development process, including knowledge acquisition, rule set specification, and expert testing. We discuss how the resulting system was evaluated in experimental sessions with practicing managers. The results give strong evidence that an expert system can be valuable in both directly supporting managers’ internal control decisions and training them to make such decisions. Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) MANAGERIAL EVALUATION OF INTERNAL CONTROLS 105 The major contribution of this work is its exploration of a new way to increase the effectiveness and efficiency of a critical organizational process: the evaluation of internal controls. 2. OBJECTIVES This study is unique in its focus on managers as users of an expert system for internal control evaluations. Even though there is evidence in some problem domains that the use of expert systems can facilitate the transfer of expertise to novice subjects, we cannot automatically assume that this is feasible or effective for the transfer of internal control evaluation knowledge to experienced managers. Managers come from diverse backgrounds (marketing, operations, technical, etc.) and do not necessarily have an accounting educational background (Viator and Curtis, 1998). Differences in educational background have been directly linked to differences in knowledge structure (Curtis and Viator, 2000). Studies have shown that a mismatch between knowledge structure and task structure can have a detrimental effect on performance (Pei et al., 1994; Nelson et al., 1995), and differences in knowledge structure are systematically associated with the quality of performance in reviewing internal control systems (Curtis and Viator, 2000). Given the variations in managers’ educational backgrounds, work experiences, and knowledge structures, constructing an expert system for effec- tive internal control evaluation by managers (as compared with external auditors or accounting student novices) may not be feasible. Moreover, real managers may resist learning about internal controls, or they may feel uncomfortable using an expert system to do so. A manager may provide incorrect inputs, not understand system outputs, find it difficult to work with an expert system, or resist considering the advice it gives. The central objective of this study is to explore whether the foregoing reservations can be over- come. This exploration has two aspects: building an expert system aimed at helping managers evaluating internal control weaknesses and demonstrating that practicing managers can use this system in a way that results in the transfer of internal control knowledge to them. Because the targeted users are managers, who tend not to be well versed in internal control evaluation concepts, the required features of the system are different from those of expert systems developed for auditors. Specifically, the system was designed to meet the following objectives: • The system should aid users in understanding why internal control is important. It is divided into five sections based on internal control objectives. Each section begins with an introduction to the basic concepts of such control. • The system should aid users in understanding how a particular problem is solved. For instance, several diagrams are incorporated into the interface in order to help users capture the concept of how a weakness is detected so that the learning is enhanced. • The system should be easy to understand and use by persons experienced with business operations but who are unfamiliar with internal control concepts. For example, the system prompts users to enter data relevant to detecting internal control weaknesses (e.g. the name of the person who re- ceives cash, the name of the person who records cash receipts), and reports each weakness found. • The system should report each weakness found, not just the adequacy of the overall organization’s internal control system. • The system should provide the reason why each weakness found is considered to be an internal control weakness. 106 C. CHANGCHIT AND C. W. HOLSAPPLE Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) 3. DEVELOPMENT PROCESS Expertise has been defined as the knowledge about a particular domain, understanding of domain problems, and skill in solving such problems. The nature of expertise includes the ability to (1) solve the problem, (2) explain the result, (3) learn, (4) restructure knowledge, (5) break rules, (6) deter- mine relevance, and (7) degrade gracefully (Davis and Lenat, 1982). An expert’s knowledge has both public and private aspects. Public knowledge includes the facts, theories, and definitions as found in texts and journals referenced by those studying in the domain, whereas much private knowledge is in the form of rules of thumb referred to as heuristics (Hayes-Roth et al., 1983). Heuristics allow experts to make educated guesses when necessary, to recognize promising ap- proaches to problems, and to deal effectively with errorful or incomplete data (Meservy, 1985). The review and evaluation of internal accounting controls is a critical step in every financial audit and is an area in which the auditor exhibits substantial expertise (Meservy, 1985). To meet the objectives outlined in Section 2, it is necessary to determine the reasoning that auditors use in evaluating a client’s internal control system, formalize and represent that reasoning knowledge as rule sets stored in a knowledge system, and then test the usefulness of the resultant expert system. The knowledge incorporated into this expert system formalizes the expertise of an auditor experi- enced in evaluating an internal control system of the sales and collection cycle in medium-size merchandising organizations. Typically, the strengths and weaknesses of an internal accounting control system are evaluated by determining control objectives, identifying controls and faults from a description of the system, and then combining the controls and faults into an overall evaluation of the sufficiency with which each control objective has been met. The expert from whom the knowledge for this expert system was acquired is a partner in a major international accounting firm. He has more than 10 years of experience in the area of internal control evaluation and demonstrated significant interest in the research project. The tool selected for developing the expert system was the GURU integrated artificial intelligence environment, version 3.01 (Holsapple and Whinston, 1987; Micro Data Base Systems, 1991). Beyond the rule management facilities typically found in expert system shells, GURU provides integral facilities for relational database management, forms management, program management, and spreadsheet management. These are integrated into a single problem-processing system in such a way that any of these knowledge handling techniques can be used independently or several can be used interdependently. In this particular expert system application, the techniques used included rule management, program management, forms management, and database management. For in- stance, conclusions of some rules invoked program modules which, in turn, invoked operations on form specifications. Once the source of expertise was identified and the development tool was selected, knowledge acquisition began. 3.1. Knowledge Acquisition Method Knowledge acquisition plays an important role in expert system development. It is evident that the quality of the resulting system is dependent on the quality of the knowledge originally elicited (Bolger and Wright, 1994; Moody et al., 1998). The knowledge acquisition process began with the question: What are the processes that auditors are using to arrive at internal accounting control evaluations? The task of reviewing and evaluating internal accounting was then analyzed, resulting in the identification and characterization of important aspects of the task. Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) MANAGERIAL EVALUATION OF INTERNAL CONTROLS 107 Knowledge for the expert system was acquired via a 6 month series of interviews with the expert. The expert was asked to identify all potential weaknesses that might occur in the sales and collection cycle of a medium-size merchandising organization. The expert was asked further to describe, in detail, the techniques and processes that were used to discover each of these weaknesses in a client’s internal control system. The rationale for each heuristic was also acquired in an attempt to develop an expert system that would be able to emulate both the expert’s knowledge and their reasoning behavior. Figure 1 shows how the interview process was conducted. The interview questions were divided into two streams. The first line of questions was asked to examine what descriptive knowledge was needed by the expert auditor in order to evaluate clients’ internal control systems. Such questions were then incorporated into the presentation system of the expert system for eliciting the users’ inputs (which comprised the expert system’s language system). The second stream of questions was asked to understand what reasoning knowledge the expert auditor employed in drawing inferences from the descriptive knowledge. This reasoning knowledge was then incorporated into the knowledge system of the expert system as rule sets. Interestingly, consistent with the finding of Frederick (1991), during the interview process, the expert first identified each transaction required in sales and collection for a merchandizing organization schematically (by flows of transactions). All the potential weaknesses that could occur in each of these transactions were then identified. The expert then regrouped these weaknesses taxonomically by objective. Figure 2 presents a resultant model for assessing internal controls in the sales and collection-cycle functions of a medium-size merchandising organization. Figure 1. A diagram of the interview process 108 C. CHANGCHIT AND C. W. HOLSAPPLE Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) Figure 2. Model of internal controls for a sales and collection cycle in a merchandising organization 3.2. Rule Set Specification In order to construct the expert system’s rule sets, all variables needed were identified in each of the basic internal controls grouped by objective. For example, in order to identify all variables needed for checking the control for adequate segregation of duties, the expert started by describing all departments involved in sales and collection cycles of a merchandising organization. Each func- tion required in each department was then identified and represented as a variable. Logical variables (Yes/No) were also used to check on the existence of certain controls. For instance, in order to check whether the functions of recording cash/cheques and the function of receiving cash/cheques were performed by the same person or persons who have a family relationship, the variables needed were identified and represented as follows: • The function of recording cash/cheques was represented as RCQ. • The function of receiving cash/cheques was represented as ECQ. • A logical variable was used to check whether the persons who perform these functions are related. Once all variables needed were identified, knowledge acquired from the expert was represented in sets of rules. These rule sets were also grouped according to the basic internal controls by objective. Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) MANAGERIAL EVALUATION OF INTERNAL CONTROLS 109 Figure 3. A diagram for checking the existence of proper authorization of sales orders, sales invoices, credit memo, and changes in payment conditions memos 3.3. System Outputs Several diagrams were devised to facilitate learning by the system’s users. For example, each restriction between functions was analyzed and represented first as a diagram describing all require- ments for adequate segregation of duties. The diagrams were then incorporated into the expert system as form specifications to provide users with an idea of what duties would be detected by the subsequent rule sets. Figure 3 presents an example of a diagram for detecting whether there is a proper authorization for sales orders, sales invoices, credit memos, and changes in payment conditions memos. The screen displaying this diagram is followed by a menu that prompts the user to select the person required to approve a specific document. For instance, Figure 4 presents a menu for selecting the person required to approve a sales invoice. In order to ensure that the user does not select a menu option accidentally, the next screen prompts the user for verification of his/her selection. Figure 5 presents an example of a verification screen. By processing sets of rules, the system infers a recommendation of potential internal control weaknesses. This recommendation identifies significant internal control weaknesses discovered in the situation being evaluated and indicates resulting exposures that could occur in a user’s organi- zation. For example, if the user does not identify the right person(s) required to approve sales invoices, then a weakness in internal control is reported, along with reasons why the expert considers this circumstance to be a weakness. Figure 6 presents an example of the weakness found for improper authorization of sales invoices. 110 C. CHANGCHIT AND C. W. HOLSAPPLE Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) Figure 4. A menu for selecting the person who is required to approve sales invoices Figure 5. A verification of user’s selection Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) MANAGERIAL EVALUATION OF INTERNAL CONTROLS 111 Figure 6. A report of a weakness found for improper authorization of sales invoices Figures 7 and 8 present additional examples of recommendations for weaknesses found in the internal control system. Such weaknesses indicate an inadequate segregation of duties by: • Having the same person doing both the functions of preparing sales invoices and recording sales invoices. • Having a relationship between the person who records cash /cheques and the person who receives cash /cheques. 3.4. Evaluating the System’s Validity Validation is often considered the cornerstone of expert system evaluation (Back, 1994). It is the process of analyzing the knowledge and decision-making capabilities of the expert system (O’Leary, 1988). Comparison of an expert’s recommendations with those of the expert system is a major criterion for evaluating that system’s validity. Test cases are developed and used to examine whether the expert system offers sufficiently good and timely advice compared with the human expert. For the expert system described here, test cases were generated from the manipulation of several cues for detecting potential weaknesses in an internal control system over the sales and collection cycle. These cues were obtained from a review of auditing texts, accounting texts, and input from accounting professors and experienced auditors. The expert was asked to evaluate each test case and detect its potential internal control weaknesses. Reasons for each potential weakness were also requested. Then, the prototype expert system was used to detect the potential weaknesses and offer reasons for such weaknesses as well. The results were then compared. 112 C. CHANGCHIT AND C. W. HOLSAPPLE Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) It turned out that the expert and expert system had some discrepancies in detecting the internal control weaknesses in the first case and the last (i.e. the fourth) case. The expert could identify only eight weaknesses in the first case and only nine weaknesses in the last case (instead of 10 as the expert system did). Where there were discrepancies, the expert reconsidered his responses and agreed that the expert system’s responses were indeed correct. Interestingly, this illustrates that an expert system can sometimes be useful even to an expert (e.g. to double check the expert’s reason- ing). After all refinements were made, the resulting system consisted of 256 rule sets and 952 additional files (i.e. form specifications, database tables, program fragments). 3.5. Evaluating the System’s Utility Beyond validity, it is important to assess an expert system’s utility with respect to its objective. Usefulness can be evaluated via empirical testing. To examine the impact of using an expert system to facilitate the transfer of the auditor’s internal control evaluation knowledge to management, an experiment was conducted to assess the effectiveness of managers using it. The experiment was conducted in a conference room and a laboratory room of a college, providing an isolated and controlled environment for the study. The research model used is illustrated in Figure 9. Two decision-aid treatments were used as values of the independent variable: expert system (ES) and no decision aid (NDA). Three response variables were measured as follows: 1. Effectiveness—accuracy of decision making was examined as a measure of the system’s effectiveness. Figure 7. A report of a weakness found for inadequate segregation of duties: having the same person perform both functions of preparing and recording sales invoices Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) MANAGERIAL EVALUATION OF INTERNAL CONTROLS 113 Figure 8. A report of a weakness found for inadequate segregation of duties: the functions of receiving cash / cheque and recording cash /cheque are performed by persons who are related 2. Participant perception of the task—post-experiment questionnaires were used to measure the participants’ perceptions about the task. 3. Participant satisfaction with performance—post-experiment questionnaires were used to measure the participants’ satisfaction with performance. The following hypotheses were tested to examine the utility of the expert system in facilitating transfer of an auditor’s internal control evaluation knowledge to management. Figure 9. Research model 114 C. CHANGCHIT AND C. W. HOLSAPPLE Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) H1. The improvement in accuracy scores of participants trained with the expert system (partici- pants in the ES group) is higher than the improvement in accuracy scores of participants in the NDA group. H2a. Participants in the ES group are more satisfied with their accuracy than participants in the NDA group. H2b. Participants in the ES group are more satisfied with their speed than participants in the NDA group. H3a. Participants in the ES group perceive that performing tasks requires less effort than partici- pants in the NDA group perceive. H3b. Participants in the ES group perceive that performing tasks is more interesting than partici- pants in the NDA group perceive. Sixty-five practitioners voluntarily participated in this experiment. All played managerial roles in their respective organizations. No participant had either specific background in internal control concepts or prior hands-on experience with an expert system. Table I shows the industries repre- sented by the participants. The experiment involved four sessions, as presented in Table II. Each session made use of one of four cases. Each case presented an eight-page, single-spaced narrative description of a scenario Table II. Experimental tasks in each session ES group NDA group Session II Participants detect internal control weaknesses in case D with the ES Participants detect internal control weaknesses in case D without the ES Session III Participants detect internal control weaknesses in case C with the ES Participants detect internal control weaknesses in case C without the ES Session I Participants detect internal control weaknesses in case B without the ES Participants detect internal control weaknesses in case B without the ES Session IV Participants detect internal control weaknesses in case A without the ES Participants detect internal control weaknesses in case A without the ES Table I. Participant demographics Industry No. of subjects Consultant 12 Legal 8 Manufacturing 8 Government 11 Insurance 3 Banking 2 Construction 2 Distribution 2 Mining 2 Transportation 2 Education 3 Finance 3 Others 7 Total 65 Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) MANAGERIAL EVALUATION OF INTERNAL CONTROLS 115 involving an organization with some internal control weaknesses. To prevent any bias in designing the case studies, the expert who participated in developing the prototype expert system was not allowed to participate in generating case studies. Three experienced auditors and three managers were asked to pilot test the case studies to validate their contents, as well as ensure that each case had the same level of difficulty. Ten MBA students performed pre-tests to validate the clarity of the cases. Prior to the experiment, the four case studies were randomly assigned to the sessions. Participants played the role of a manager trying to detect the potential weaknesses of the internal control system described in a case study. Ten internal control weaknesses were intentionally incor- porated into each case. The maximum time allowed for each session was 2 h. For each session, an accuracy score was calculated and the decision time was recorded. Questionnaires were distributed at the end of each session to measure participants’ satisfaction with their performances, as well as their perceptions about the task (see Appendix A). Accuracy of decision making is regarded as a measure of the system’s effectiveness. Two expe- rienced auditors and two accounting professors evaluated the list of internal control weaknesses previously described. Each evaluator assigned a score to each weakness using a scale of 0 to 10, based on the degree of importance of each weakness (0 being the least important, 10 being very important). The average of the scores for each weakness (i.e. total divided by four) was assigned as its importance. In arriving at an accuracy score, all points related to correctly identified weaknesses are added. In order to prevent the subjects from trying to detect weaknesses by guessing, they were informed at the beginning of the experiment that one-third of all points for inaccurately identified weaknesses would be subtracted as the penalty for guessing. A non-response for a potential weakness results in neither addition nor subtraction. In testing the hypotheses, we examined the deviations of participants’ performances between session I and session IV. In both sessions, participants had to perform tasks without the expert system support. The deviations examined are attributed to the learning effects that occurred while performing tasks in sessions I to III. There were two types of potential learning effect that we expected here. First, there is the learning effect that could occur from participants having used the expert system; this could happen only for the ES group. Second, there is the learning effect that could occur from participants having performed four cases consecutively; the degree of this type of learning effect could be expected to be comparable for the ES group and NDA group. To verify homogeneity of internal control evaluation knowledge between the ES and NDA groups, an analysis of variance (ANOVA) was performed on the first session results for the experimental group (ES group) versus the control group (NDA group). The ANOVA was performed to test for differences in the Accuracy Score between the groups. At an α level of 0.05, no significant differ- ence was found between the groups. 4. EXPERIMENTAL FINDINGS The experimental data were analyzed as a completely randomized design with a one-way treatment structure using the ANOVA technique. Treatments had a grouping structure and fixed effects. Each participant was an experimental unit. The F-test was used to test hypotheses about group means for each response variable: (i) accuracy score (H1), (ii) participants’ satisfaction with performance (H2a and H2b), and (iii) participants’ perceptions of the task (H3a and H3b). Hypothesis H1 was tested by comparing the deviations of performances of each participant in session IV versus session I 116 C. CHANGCHIT AND C. W. HOLSAPPLE Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) between the expert system group and the NDA group. Hypotheses H2a, H2b, H3a, and H3b were tested by comparing participants’ responses from the questionnaires given at the end of session IV. Regarding hypothesis H1, session I versus session IV results show that participants could detect internal control weaknesses significantly more accurately after being trained with the expert system ( µ I = 4.575 versus µ IV = 16.073) than before. Even though there is also an increase in the accuracy score for the NDA group ( µ I = 8.584 versus µ IV = 12.250), this increase was not nearly at the same significance level as in the experimental group. It is clear that the increase in accuracy in the ES group (having practiced with the expert system) is much greater than the increase in accuracy in the NDA group. The accuracy score in the experimental group is more than three times better after being trained with the expert system, whereas the improvement in the NDA group is less than 50%. A t-test conducted on the improvement score between two groups (11.498 for the ES group and 3.666 for the NDA group) also confirms that the improvements in performance of participants in the ES group were significantly higher than the improvements in performance of participants in the NDA group from an accuracy standpoint ( P = 0.0035). The results are presented in Table III. The major findings for testing hypotheses H2a, H2b, H3a, and H3b are summarized in Table IV. The participants’ answers to the post-experiment questionnaires reveal the attitudes of participants in both groups toward the task as follows: Table IV. Results of testing hypotheses H2a, H2b, H3a, and H3b Response Variable H0 Session IV P-value a ES group NDA group Mean of participants’ attitude on the difficulties of the task H2a 3.65 2.93 0.085 (1: very difficult; 7: very easy) Mean of participants’ attitude about their satisfaction with H2b 3.94 3.00 0.05 the accuracy in answering the case study (1: very unsatisfactory; 7: very satisfactory) Mean of participants’ attitude about their satisfaction with the H3a 4.04 2.87 0.025 speed in answering the case study (1: very unsatisfactory; 7: very satisfactory) Mean of participants’ attitude on how interesting was the task H3b 4.18 2.87 0.025 (1: very boring; 7: very interesting) a Represents significance of at least α = 0.05. Table III. Results of testing hypothesis H1 Mean of accuracy score Deviation between P-value between (range: −20 to 38) Sessions I and IV ES and NDA groupsc Session Ia Session IVb ES group (49 participants)d 4.575 16.073 11.498 0.0035 NDA group (15 participants) 8.584 12.250 3.666 a Participants perform case B without the expert system. b Participants perform case A without the expert system. c Represents significance of at least α = 0.01. d One participant had an emergency call and left while performing case II. His data were taken out. Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) MANAGERIAL EVALUATION OF INTERNAL CONTROLS 117 • Participants in the ES group were more satisfied with their perceived accuracy than participants in the NDA group ( µ ES = 3.94 versus µ NDA = 3.00, P = 0.05). • Participants in the ES group were more satisfied with their perceived speed than participants in the NDA group ( µ ES = 4.04 versus µ NDA = 2.87, P = 0.025). • There was no significant difference in participants’ attitude between the ES group and the NDA group on the difficulty of the task. That is, expert system exposure and training does not appear to affect perceptions of task difficulty. These findings might be attributed to the following: 1. Because participants in the ES group were allowed to use the expert system in sessions II and III, they may lose some confidence after the expert system was removed in session IV. 2. The increase in participants’ perceptions in the NDA group may be the result of learning that may have accrued from having the participants perform three consecutive cases. • Participants in the ES group perceived that the task was more interesting than participants in the NDA group perceived ( µ ES = 4.18 versus µ NDA = 2.87, P = 0.025). • Answers to open-ended questions also reveal satisfaction with the use of the expert system. Most of the participants were interested in the availability of the system and the idea that they could learn how to evaluate internal controls. A number of these participants even asked whether the system was commercially available. 5. CONCLUSIONS, LIMITATIONS, AND DIRECTION FOR FUTURE RESEARCH This research is an initial investigation of the use of an expert system to train practicing managers on making internal control evaluations. The major finding of this research is that it is feasible to transfer an auditor’s internal control evaluation knowledge to management via the use of an expert system. However, as with any computerized system, the expert system must be used with care. It is important to note that the objective of the system is not to replace the auditor’s work. An auditor still plays the traditional role. The main purpose of the system is to provide managers with a practical applied understanding of internal controls. Having this knowledge can help managers maintain an effective internal control system, thus providing more reliable accounting data and better safeguard- ing of assets. It could let organizations save time and money by allowing weaknesses in internal control systems to be detected and solved more quickly. The communication between management and auditors relating to the importance of internal controls might also be improved. Generalizability of the results in this study may be constrained in several respects. First, this research concentrates only on the evaluation of controls commonly found in the sales and collection cycle. Second, it investigates internal control systems commonly found in the merchandising indus- try. It is not expected to handle novel (uncommonly different) accounting systems. In addition, because this is just an initial study, there may still be some limitations concerning the design of system’s user interface (e.g. the inability to go back directly to change or correct an answer in the previous screen). These limitations point to directions in which the research presented here can be extended by future investigations. The future development of this system might use an Internet-compatible tool so that access to the system can be significantly increased. Researchers might try to incorporate additional transac- tion cycles, industries, or other auditing functions into the expert system. They could develop and study another expert system by acquiring the expertise from an internal auditor instead of the external auditor. Another direction is to acquire expertise from multiple auditors and then conduct experiments similar to the one reported here. Finally, researchers may investigate the feasibility of 118 C. CHANGCHIT AND C. W. HOLSAPPLE Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) integrating this kind of expert system with a company’s databases so that users are not required to input as much information. APPENDIX A: POST-EXPERIMENT QUESTIONNAIRES Please answer the following questions. Please keep in mind that the following questions refer ONLY to the last case study you have done. 1. On a scale of 1 (very unsatisfactory) to 7 (very satisfactory), how satisfied were you with your accuracy in answering the case study? 1 2 3 4 5 6 7 very unsatisfactory slightly neutral slightly satisfactory very unsatisfactory unsatisfactory satisfactory satisfactory 2. On a scale of 1 (very difficult) to 7 (very easy), how difficult was it to do the case study? 1 2 3 4 5 6 7 very difficult slightly neutral slightly easy very difficult difficult easy easy 3. On a scale of 1 (very unsatisfactory) to 7 (very satisfactory), how satisfied were you with your speed in answering the case study? 1 2 3 4 5 6 7 very unsatisfactory slightly neutral slightly satisfactory very unsatisfactory unsatisfactory satisfactory satisfactory 4. On a scale of 1 (very boring) to 7 (very interesting), how interesting was the task you performed? 1 2 3 4 5 6 7 very boring slightly neutral slightly interesting very boring boring interesting interesting REFERENCES Arens AA, Loebbecke JK. 1994. Auditing: An Integrated Approach, 6th edn. Prentice Hall: New Jersey. Back B. 1994. Validating an expert system for financial statement planning. Journal of Management Information Systems 10(3): 157–177. Bailey Jr AD, Duke GL, Gerlach J, Ko C, Meservy R, Whinston AB. 1985. TICOM and the analysis of internal controls. Accounting Review 60(2): 186 –211. Biggs SF, Morrison TA. 1990. What accountants need to know about expert systems. The CPA Journal 60: 98 –101. Biggs SF, Messier Jr WF, Hansen JV. 1987. A descriptive analysis of computer audit specialists’ decision- making behavior in advanced computer environments. Auditing: A Journal of Practice & Theory 6(1): 1–21. Bolger F, Wright G. 1994. Assessing the quality of expert judgement: issues and analysis. Decision Support Systems 11: 1–24. Bonczek RH, Holsapple CW, Whinston AB. 1981. Foundations of Decision Support Systems. Academic Press: New York. Bonner SE, Pennington N. 1991. Cognitive processes and knowledge as determinants of auditor expertise. Journal of Accounting Literature 10: 1–50. Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) MANAGERIAL EVALUATION OF INTERNAL CONTROLS 119 Bonner SE, Walker PL. 1994. The effects of instruction and experience on the acquisition of auditing know- ledge. The Accounting Review 69(1): 157–178. Boritz JE. 1985. The effect of information presentation structures on audit planning and review judgments. Contemporary Accounting Research 8(1): 193–218. Borthick AF, West OD. 1987. Expert systems—a new tool for the professional. Accounting Horizons 1(1): 9–16. Brown CE, Phillips ME. 1990. Expert system for management accountants. Management Accounting 7: 18 –20. Brown DL, Jones DR. 1998. Factors that influence reliance on decision aids: a model and an experiment. Journal of Information Systems 12(2): 75 – 94. COSO. 1992. Internal Control-Integrated Framework, Committee of Sponsoring Organizations of Treadway Commission. AICPA: New York. Cummings BK, Apostolou NG. 1987. Expert systems in auditing: an emerging technology. Internal Auditing 3(2): 3–10. Cummings BK, Lauer J, Baker R. 1988. Expert systems in internal auditing. Internal Auditing 4(1): 49 – 64. Curtis MB, Viator RE. 2000. An investigation of multidimensional knowledge structure and computer auditor performance. Auditing: A Journal of Practice and Theory 19(2): 83–103. Davis R, Lenat DB. 1982. Knowledge-based Systems in Artificial Intelligence. McGraw-Hill: California. Ege G, Sullivan WG. 1990. Expert system for management accountants. Management Accounting 7: 18–20. Eining MM, Dorr PB. 1991. The impact of expert system usage on experiential learning in an auditing setting. Journal of Information Systems 5(1): 1–16. Fedorowicz J, Oz E, Berger PD. 1992. A learning curve analysis of expert system use. Decision Sciences 23(3): 797–818. Frederick DM. 1991. Auditor representation and retrieval of internal control knowledge. The Accounting Review 66(2): 240–258. Gadh VM, Krishnan R, Peters JM. 1993. Modeling internal control and their evaluation. Auditing: A Journal of Practice & Theory 12(Supplement): 113–129. Gal GF. 1985. Using auditor knowledge to formulate data model constraints: an expert system for internal control evaluation. PhD dissertation, Michigan State University. Gal GF, Steinbart PJ. 1992. Interface style and training task difficulty as determinants of effective computer- assisted knowledge transfer. Decision Sciences 23: 128–143. Graham LE, Damens J, Ness GV. 1991. Developing Risk AdvisorTM: an expert system for risk identification. Auditing: A Journal of Practice & Theory 10(1): 69 – 96. Hayes-Roth F, Waterman D, Lenat D. 1983. Building Expert Systems. Addison-Wesley: Reading, MA. Holsapple CW, Raj VS. 1994. An exploratory study of two KA methods. Expert Systems 11(2): 77–88. Holsapple CW, Whinston AB. 1987. Business Expert Systems. Irwin: Homewood, IL. Libby R. 1995. The role of knowledge and memory in audit judgment. In Judgment and Decision-Making Research in Accounting and Auditing, Ashton RH, Ashton AH (eds). Cambridge University Press: New York. Lieb EB, Gillease JK. 1996. Du Pont uses a decision support system to select its audit portfolio. Interfaces 26(3): 10 –21. Marchant GA. 1989. Analogical reasoning and hypothesis generation in auditing. Accounting Review 64(3): 500 –513. Meservy RD. 1985. Auditing internal control: a computational model of the review process. PhD dissertation, University of Minnesota. Meservy RD, Bailey Jr AD, Johnson PE. 1986. Internal control evaluation: a computational model of the review process. Auditing: A Journal of Practice & Theory 6(3): 44 –74. Messier Jr WF. 1995. Research in and development of audit decision aids. In Judgement and Decision-Making Research in Accounting and Auditing, Ashton RH, Ashton AH (eds). Cambridge University Press: New York. Micro Data Base Systems. 1991. Getting Started with GURU. Micro Data Base Systems, Inc.: Lafayette, IN. Moody JW, Blanton JE, Cheney PH. 1998. A theoretically grounded approach to assist memory recall during information requirements determination. Journal of Management Information Systems 15(1): 79 – 98. Murphy DS. 1990. Expert system use and the development of expertise in auditing: a preliminary investigation. Journal of Information Systems 4(2): 18 –35. Nelson MW, Libby R, Bonner SE. 1995. Knowledge structure and the estimation of conditional probabilities in audit planning. Accounting Review 70(1): 27–47. 120 C. CHANGCHIT AND C. W. HOLSAPPLE Copyright © 2004 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 12, 103 –120 (2004) Odem MD, Dorr PB. 1995. The impact of elaboration-based expert system interfaces on de-skilling: an epis- temological issue. Journal of Information Systems 9(1): 1–17. O’Leary DE. 1988. Methods of validating expert systems. Interfaces 18(6): 72–79. O’Leary DE, Watkins PR. 1989. Review of expert systems in auditing. Expert Systems Review for Business & Accounting 9(2): 3–22. Oz E. 1989. A study of improvement in decision-making skills through the use of expert systems. Unpublished PhD dissertation, Boston University. Pei BKW, Reneau JH. 1992. The effect of memory structure on using rule-based expert systems for training: a framework and an empirical test. Decision Sciences 21: 263–286. Pei BKW, Steinbart PJ, Reneau JH. 1994. The effect of judgment strategy and prompting on using rule-based expert systems for knowledge transfer. Journal of Information Systems 8(1): 21– 42. Rezaee Z. 2004. Causes, consequences, and deterrence of financial statement fraud. Critical Perspectives on Accounting in press. Steinbart PJ, Accola WL. 1994. The effect of explanation type and user involvement on learning from and satisfaction with expert systems. Journal of Information Systems 8(1): 1–17. Viator RE, Curtis MB. 1998. Computer auditor reliance on automated and non-automated controls as a function of training and experience. Journal of Information Systems 12(1): 19–30. Vinze AS, Karan V, Murthy US. 1991. A generalizable knowledge-based framework for audit planning expert systems. Journal of Information Systems 5(2): 78–91. Weber R. 1999. Information Systems Control and Audit. Prentice-Hall: New Jersey. Wong BK, Monaco JA. 1995. Expert system applications in business: a review and analysis of the literature (1977–1993). Information & Management 29: 141–152. 
work_d7yiarpyufhghlqmh2ogsmbrne	----	2013_09_05.pub Management Systems in Production Engineering   2013, No 1 (9), pp 26‐30       Abstract:  The fundamental goal of modern informa on technology is to support decision making process in organiza ons, for both  rou ne and highly complex problems, based on heterogeneous data sources. The main objec ve of the paper is to pre‐ sent the construc on concept of knowledge base for an expert system designed to support the decision making in car‐ diac surgery. A knowledge‐based medical expert system is designed to support decision‐making process in cardiac surge‐ ry where knowledge from medical guidelines, risk scales, and registries is applied to reason the medical procedure in a  par cular case scenario.   DECISION MAKING PROCESS IN CARDIAC SURGERY – CONCEPT OF BUILDING AN EXPERT SYSTEM   INTRODUCTION  The extent of contemporary medical knowledge results  from an ongoing development as well as scien fic and tech‐ nical progress in healthcare. Medicine is s ll regarded as an  art of science, however this point of view has significantly  developed over the past ages. Uniform therapeu c indica‐ ons  and  treatment  approach have  introduced  standardi‐ zed  medical  care.  Con nuous  healthcare  policing  has  im‐ proved  treatment  outcomes  and  overall  pa ent  safety,  decreased  the  incidence  of  malprac ce,  and  introduced  cost‐effec veness and economic approach into healthcare.  Progress and  improvement of medical care requires obse‐ rva on of guidelines and medical standards by healthcare  professionals. It is s ll expected though, that clinicians will  maintain their prac ces in accordance with novel techniqu‐ es,  improved  methods,  and  treatment  approach.  Medical  professionals are supported in their clinical daily rou ne by  increased delivery of scien fic papers, new versions of gui‐ delines  and  recommenda ons,  and  updated  risk  scales.  Although  informa on  received  is  complimentary,  uniform  presenta on requires complex data processing and integra‐ on of informa on from various sources and areas of medi‐ cine.   Decisions made in clinical medicine result in a series of  events in the whole environment of medical center. Selec‐ ted medical therapy requires use of par cular medical equi‐ pment  and  device.  In  some cases,  as a  consequence,  this  may cause temporary altera ons of center‘s ac vity, inclu‐ ding  hiring  addi onal  staff, extra  costs  and  spending.  The  primary goal of this approach  is pa ent survival and op ‐ mal recovery, and in case of ac ve employees – return to  normal  work.  Unfavorable  medical  decisions  are  usually  undertaken because of lack of extensive and current medi‐ cal knowledge, mistaken analysis of various factors or due  to ignorance of medical checklist. It is conceivable that pro‐ perly designed search engine would mine through available  registries  and  present  comparable  case  study  for  assess‐ ment whereas not every clinical scenario could have been  previously  described  in  medical  guidelines.  Each  pa ent  undergoing  medical  procedure  is  at  risk  of  temporary  or  long‐term disability or even death. Every medical complica‐ on  generates  an  organiza onal  hurdle  for  the  medical  facility and is associated with increased costs of treatment.  Medical  treatment  costs  that  exceed  standard  therapy  (based on diagnosis related groups – DRG) directly influen‐ ce the hospital economic balance. Pa ent rights are well‐ defined  in  legal  acts  and  give  way  to  compensa on  de‐ mands against the staff and medical center.   This complex decision‐making process could have been  supported by applied systems of ar ficial intelligence based  on  available  sources  of  knowledge.  The  system  would  re‐ place  neither  human  decisions  nor  responsibility,  but  it  would help to iden fy the best procedure, especially in the  most  complicated  cases.  In  this  paper  we  present  one  of  the  most  technologically  advanced  medical  specialty,  car‐ diac  surgery,  with  brief  descrip on  of  the  sources  of  knowledge  available  and  the  characteris cs  of  knowledge  engineering  and  methods  applied  to  generate  an  expert  system.  Cardiac surgery – overview of the knowledge sources Cardiac surgery is an invasive (surgical) medical special‐ ty designed for treatment of heart diseases. Despite clear  anatomical  borders  for  cardiac  interven ons,  surgeons  must  be  proficient  in  other  medical  special es,  such  as:  thoracic  and  vascular  surgery,  cardiology,  intensive  care,  transplanta on, etc. Cardiac surgery influences wide areas  of socio‐economic life of the society: circulatory disease is a  leading mortality cause in Poland. Medical costs associated  with cardiac interven ons are of the highest among medi‐ Tomasz HRAPKOWICZ, Marcin MARUSZEWSKI   CompassMedica Sp. z o. o.   Anna KEMPA, Krzysztof MICHALIK, Bogna ZACNY   University of Economics in Katowice   Key words: expert system, cardiac surgery, knowledge base, decision‐making process     Management Systems in Produc on Engineering 1(9)/2013                                                                                                               27                      T. HRAPKOWICZ et al. — Decision making process in cardiac surgery – concept of building an expert system  cal special es. Decision‐making in cardiac surgery is a com‐ plex process that involves a whole team of specialists: car‐ diologists,  cardiac  surgeons,  anesthesiologists  and  others.  Various criteria are taken into account in this process inclu‐ ding  individual  professional  experience,  opinion  of  other  experts  in  the  field,  and  most  importantly  the  unbiased  sources  designed  to  support  decision‐making  in  specific  medical cases: medical guidelines, standards, and risk sca‐ les.   Guidelines  are  elementary  tools  used  by  cardiac  sur‐ geons  in  medical  assessment.  They  are  issued  based  on  limited sources of  informa on, which are mainly historical  (previously published) and therefore do not present up‐to‐ date state‐of‐the‐art  in cardiac surgery. The sources of  in‐ forma on in cardiac surgery include: results of randomized,  prospec ve, mul center clinical trials and other,  including  meta‐analyses, as well as expert opinions invited to par ci‐ pate  in  work  groups  to  establish  the  rules  of  conduct  in  specific types of clinical scenarios. The above‐men oned lay  ground for medical guidelines (recommenda ons) for par ‐ cular clinical diagnoses, are limited to most common cases  and do not cover every possible diagnos c and therapeu c  scenario. Another  limita on of the guidelines  is extrapola‐ on of uniform therapeu c assump ons for en re pa ent  popula on whereas source data have been obtained from  selected  groups  of  pa ents  in  individual  clinical  studies.  Guidelines in cardiac surgery are published by a whole ran‐ ge of mul disciplinary scien fic socie es: European Society  for Cardiovascular Surgery, European Associa on for Cardio ‐thoracic  Surgery,  Society  of  Vascular  Surgery,  American  Heart  Associa on,  European  Society  of  Cardiology,  and  others.  Medical  standards  warrant  uniform  procedures  and  increase safety at every stage of medical approach. Of note,  although very helpful, current standards of care do not re‐ fer  to  the  most  complicated  medical  cases.  Standards  of  care rather define the scope of possible treatment strate‐ gies for specific scenarios in medicine. Although the overall  goal is to apply medical standards to avoid common errors  in  the  therapeu c  process,  these  are  typically  limited  to  specific country, region or even medical facility, and cannot  be universally u lized in every cardiac surgery center.  Healthcare professionals also exploit medical registries  that  combine  well‐organized  clinical  facts  and  data  inclu‐ ding cases where medical treatment was conducted in ac‐ cordance  with  the  guidelines  as  well  as  scenarios  not  de‐ scribed  by  recommenda ons  where  therapy  was  selected  based on clinical experience of the medical team. In addi‐ on,  each  center  may  develop  local  registries  and  allow  medical judgment based on expert consensus provided the  presented  case  is  beyond  medical  guidelines.  The  three  most characteris c limita ons of each database are accura‐ cy, expense, and analysis. For each database full‐ me dedi‐ cated  staff,  adequate  resources,  and  a  system  to  ensure  data  integrity are essen al. Of these, dedicated and com‐ mi ed staff is the most important and expensive. The value  of people responsible for data collec on and input cannot  be  overemphasized:  they  require  educa onal  support  by  the surgeons to ensure that clinical informa on is accurate‐ ly and consistently interpreted for data entry. The value of  a database is dependent on valid data [14].  Risk  scales  are  helpful  in  uniform  groups  of  pa ents,  par cularly to jus fy treatment strategy in extremely risky  procedures. Risk scales have been designed to predict early  mortality  and  morbidity  of  the  reference  class  of  pa ent  popula on,  and  refer  to  large  databases  origina ng  from  volunteering units, without external valida on of the data‐ set in many cases. Most common risk scales in cardiac sur‐ gery are EuroSCORE and STS. Even moderate levels of error  can lead to substan al inaccuracy in es mates of mortality  rates and in some circumstances these inaccuracies can be  gross,  especially  at  the  low  mortality  rates  that  are  now  prevalent in cardiothoracic surgery [4].  High  cost,  complexity  and  compe on  among  cardiac  centers also bring the need for a computer support that will  provide  unbiased  informa on  on  best  op mal  treatment  for the par cular case, surgical strategy and planning – inc‐ luding op mal bed assignment, local risk stra fica on and  cost  associated  with  complica ons,  cost‐effec veness  and  procedure cost evalua on. Integra on of all available medi‐ cal knowledge into an expert computer applica on will help  assess the risk of malprac ce or human error during pa ent  evalua on for surgical treatment.   Characteristics of the knowledge engineering Knowledge engineering is an experimental science. Ma‐ ny methods for knowledge representa on have been deve‐ loped; knowledge bases were created in various areas. It is  observed, however, that the source of progress in knowled‐ ge engineering are experiments and research on represen‐ ta on of specialized prac cal knowledge, verified  in par ‐ cular applica ons, such as medicine. Specificity of problems  knowledge engineering is facing might be contained in the  simplified  statement  that  the  knowledge  from  each  field  can be poten ally formalized, but a significant bo leneck of  such ac vity is the complexity of the given field, the degree  of  knowledge  openness  and  the  structure  of  knowledge  sources.  The  more  complex  and  dependent  on  sca ered  sources  the  field  is,  the  higher  challenge  it  makes  for  knowledge engineers. Simplified models, making a certain  fragment of the given whole, seem to work normally, whe‐ reas working out the founda ons of a complex system gives  rise to a number of difficul es. Knowledge engineering pro‐ poses methods that can support this laborious process, but  it does not present ready solu ons for the par cular area –  since  these  must  be  created  through  coopera on  among  knowledge engineers and experts in the given discipline.   At first knowledge engineering comprised mostly tasks  connected with knowledge acquisi on for expert systems.  Today  it can be presented as a field connected with crea‐ ng knowledge bases and using seman c technologies for  knowledge processing by computer systems [5]. The issues  falling within the scope of knowledge engineering comprise  methods  for  knowledge  acquisi on  (including  acquiring  knowledge  from  experts,  knowledge  discovery  in  databa‐ ses,  text  documents),  ways  of  knowledge  representa on,  knowledge formaliza on, methods for knowledge analysis,  knowledge processing [13].   Current state of development in knowledge engineering  for cardiac surgery worldwide can s ll be defined as pione‐ ering.  Studies  on  related  disciplines  have  been  published  for cardiology [3, 6, 12, 16], infec ous diseases [1, 7], and  oncology [2, 10]. Knowledge used by cardiac surgeons mo‐ stly occurs  in the form of a text document and databases  (registries). Only few opera ve risk scales are available as  interac ve Internet forms and applica ons accessible to the  users of mobile devices. There have not been published any  interac ve forms of knowledge base for treatment guideli‐ nes for e.g. ischemic heart disease.   28                                                                                                               Management Systems in Produc on Engineering 1(9)/2013                               T. HRAPKOWICZ et al. — Decision making process in cardiac surgery – concept of building an expert system    It is conceivable that expert systems compu ng medical  standards,  guidelines  as  well  as  risk  scales  and procedure  costs would support medical professionals in op mal treat‐ ment selec on and planning.  MATERIALS AND METHODS  Prior to design of knowledge base to support decision  making process in cardiac surgery a formal model and rese‐ arch methodology has to be developed. This is achieved by  solving the following steps:    Loca on  and  organiza on  of  knowledge  sources  in  cardiac surgery,   Working  out  methods  for  acquiring  and  gathering  knowledge on the basis of located knowledge sources  and experiences of experts in cardiac surgery,   Working  out  a  formal  model  for  field  knowledge  in  cardiac surgery using a rule‐based formalism.  Development of expert system requires studies on the  possibility of knowledge arrangement and formaliza on  in  such highly complex domain like cardiac surgery. A unique  research  team  consis ng  of  specialists  in  cardiac  surgery  (medical experts) and knowledge engineering will design a  research  model  including  the  process  of  field  knowledge  acquisi on and update of guidelines and recommenda ons.  Later, the hypotheses on feasibility of field knowledge for‐ maliza on  in  cardiac  surgery  for  the  future  applica on  in  decision‐making process will be verified. At current state of  knowledge it is feasible to develop a formal model of field  knowledge  in cardiac surgery based upon medical guideli‐ nes  and  standards.  The  formal  model  would  develop  knowledge  base  to  perform  meta‐analysis  of  medical  re‐ commenda ons  and  assess  risk  associated  with  selected  medical  procedures.  It  also  seems  possible  to  develop  a  formal model of field knowledge in cardiac surgery for spe‐ cific geographic region with applica on of medical registry  data mining. The formal model of specific knowledge would  compare guidelines and recommenda ons with actual clini‐ cal daily prac ce in the requested area. The more specific,  i.e.  limited to the country, region or medical center  is the  registry  database,  the  more  accurate  informa on  can  be  provided on what sort of treatment strategy is being offe‐ red to the pa ent that seeks medical care. Development of  a formal model of economic knowledge that would include  procedure costs and es mated length of hospital stay and  support  decision‐making  process  from  the  perspec ve  of  cost‐effec veness in cardiac surgery would be possible.  Within the framework of the  ini al research described  in a publica on by the authors of the project [11] literature  analysis was made in order to assess the state of knowled‐ ge  in the studied area, which allowed to show the useful‐ ness  and  originality  of  the  problem  chosen  for  research.  Also an analysis of some sources of knowledge was perfor‐ med  (including  the  guidelines  of  the  European  Society  of  Cardiology) to acquire and formalize knowledge on surgical  treatment of aor c valve disease. The final set of rules com‐ prised a small subarea of problems, which ul mately need  to  be  analyzed.  However,  this  simple  model  has  already  dis nguished  early  problems  (concerning  the  acquisi on,  structuraliza on  of  knowledge),  significant  from  the  point  of view of formalizing knowledge in cardiac surgery. There‐ fore many different experimental models have to be deve‐ loped to build a consistent formal model.   Scien fic technique and research process can be descri‐ bed through the planned work stages:   Analysis and reviews of literature understood as dee‐ pening of analysis in respect of the state of science in  the studied area. As part of preliminary research, the  analyses  have  already  been  made  which  have  allo‐ wed  the  usefulness  and  originality  of  the  problem  intended to research,    Adapta on of the results of other authors’ studies –  thorough  analysis  of  other  scien fic  works  is  also  supposed to allow the evalua on of possibili es for  adapta on of some solu ons, for instance in the con‐ structed knowledge model the authors consider using  the  medical  terms  described  by  the  current  ontolo‐ gies, e.g. the medical terms classifica on SNOMED,    Loca on and organiza on of sources of knowledge in  cardiac  surgery  –  the  work  on  the  analysis  of  knowledge sources will be carried out throughout the  process of working on the project, while in the ini al  phase a detailed analysis will be made concerning the  structure  of  knowledge  contained  in  the  available  sources.  An  important  problem  is  the  precision  of  knowledge sources in the form of guidelines. Guideli‐ nes  not  always  show  clearly  enough  which  compo‐ nent condi ons (determining the way of conduct) are  of priority importance in rela on to the other. It may  turn  out  that  in  an  untypical  situa on,  the  recom‐ menda ons of guidelines are difficult to use. In such  cases  it  may  be  necessary  to  locate  and  use  other  sources of knowledge,    Adapta on  of  the  results  of  other  authors’  studies  will  also  include  methodologies  describing  the  pro‐ cess of knowledge acquisi on, gathering and formali‐ za on. As a result of cri cal analysis of the results of  research  within  this  scope,  conduc ng  experiments  for chosen approaches and based on this, developing  the authors’ own adapta on of the methodology  in  this area is planned,   The above men oned works will allow making a for‐ mal  model  for  field  knowledge  in  cardiac  surgery  using  rule‐based  formalism.  It  was  assumed  that  a  par cular  disease  en ty  will  be  the  developmental  unit  of  the  model.  For  each  disease  en ty,  the  cor‐ rectness of rules will be checked (in respect of cohe‐ rence,  contradic on,  coinciding  of  rules),  as  well  as  their  completeness.  A er  this  stage,  the  knowledge  model will be verified concerning the facts by experts  – doctors being members of the research team – if it  includes all the schemes of surgical treatment requi‐ red at the given unit.   In  the  course  of  knowledge  formaliza on  for  cardiac  surgery, it is of great importance that there is a possibility  of developing expert system prototypes for the next frag‐ ments of a knowledge base coded on the basis of the deve‐ loped  formal  model.  Prototypes  will  be  worked  out  using  the field‐independent tool used to build expert systems, PC ‐Shell from the ar ficial intelligence package Aitech Sphinx,  which  was  already  used  in  medicine,  for  instance  in  psy‐ chiatry for diagnosis of affec ve disorders, and the work in  this field is con nued [9]. PC‐Shell is a system of hybrid ar‐ chitecture,  i.e.  combining  different  methods  for  problem  solving  and  knowledge  representa on.  One  of  interes ng  proper es of the PC‐Shell system is a built‐in, fully integra‐ ted, simulator of a neural network.  Another essen al cha‐ racter of the PC‐Shell system is its blackboard architecture,  allowing the division of a large knowledge base into smaller  modules,  orientated  thema cally  –  which  will  be  par cu‐   Management Systems in Produc on Engineering 1(9)/2013                                                                                                               29                      T. HRAPKOWICZ et al. — Decision making process in cardiac surgery – concept of building an expert system  larly useful in the designed system, on account of substan‐ al sca ering of knowledge sources.    The methods of the study used within the framework of  the one of research hypothesis – possibility for expanding  the knowledge base with knowledge from medical registers  documen ng actual medical events – will be subordinated  to  rules  of  conduct  in  the  course  of  performing  tasks  on  data explora on, such as a reduc on of feature space di‐ mension,  classifica on,  data  clustering  or  associa on  di‐ scovery. A er a detailed iden fica on the character of data  that  is  supposed  to  be  added  to  the  knowledge  base  –  a  selec on of data mining algorithms will be made, including  ar ficial immune systems [8], confirmatory Factor Analysis  [15]. The decision about choosing an algorithm will be ta‐ ken based on measures calculated in the course of verifica‐ on and used methods for results visualiza on.   To verify the medical registers (making sta s cal analy‐ sis – ini al stage of data explora on), the authors intend to  use tools provided by the WEKA  library (Waikato Environ‐ ment for Knowledge Analysis) worked out by the University  of Waikato in New Zealand as an open source so ware un‐ der GNU General Public License, which makes a useful tool  for a circle of scholars dealing with machine learning, data  analysis and data mining. Moreover, using such tools as the  R program is an cipated. R is a programming language and  development environment, which allows making sta s cal  calcula ons  and  data  visualiza on.  The  source  code  R  is  also  published  under  GNU  GPL  license,  and  the  program  itself provides a wide range of sta s cal techniques, at the  same  me allowing  independent modifica on of the used  algorithms, which at such a novel project will be a signifi‐ cant asset.   RESULTS  The  ini al  expert  system  prototype  has  begun  proper  func on a er inser on of the rules from the European So‐ ciety  of  Cardiology  guidelines  for  valvular  heart  disease  (aor c valve surgery). For the need of PC‐Shell expert sys‐ tem, isolated sec ons have been separated: lexical (facets)  and the rules of procedure (rules). Facets describe a ribu‐ tes,  possible  values,  and  op onal  ques on  wording.  The  course  of  consulta ons  with  the  user  is  characterized  by  the level of details of ques ons, which depends on the wor‐ ding of the rules. For example, in the case of query about  the range of systolic le  ventricular dysfunc on rule checks  the value of le  ventricular ejec on frac on, which means  that the system asks for the exact numerical value (Fig. 1).  Fig.  1  The  ques on  about  the  value  of  le   ventricular  ejec on  frac on    Whereas  in the assessment of the degree of the valve  calcifica on and  increased peak flow the rule checks only  whether it occurs: the system asks the user if this situa on  happens at all but not wishes to ascertain about the exact  value of the flow rate (Fig. 2).  Fig. 2 The ques on about the presence of aor c calcifica on    The result of a sample consulta on shows Fig. 3. Proto‐ type system has confirmed indica ons for surgery and pre‐ sented a scheme of inference.   Fig. 3 The result of an example consulta on  CONCLUSION  Successful performance of the project requires transpa‐ rent  and  con nuous  exchange  of  expecta ons,  thoughts,  and  needs  between  two  independent  study  par cipants:  computer  scien sts  skilled  in  knowledge  engineering  and  medical professionals (cardiac surgeons) who feed the sys‐ tem with their proficiency in medical guidelines, standards,  rules of conduct, and clinical rou ne. There is no research  model available to adapt for the proposed project: no me‐ dical guidelines have been translated into logic‐based rules,  medical  registries  are  not  assembled  as  knowledge  base,  nor have economic results of medical centers been associa‐ ted  with  risk  modeling.  Extensive  and  specialist  field  knowledge in cardiac surgery is going to be translated into  knowledge base for the first  me in this project. Design of  the  formal  model  that  will  address  the  above‐described  problem  requires  crea on  of  unique  methodology  where  characteris cs of scien fic domains involved will be reflec‐ ted both in terms of knowledge engineering as well as com‐ plexity  of  the  domain.  The  research  methodology  will  ac‐ commodate study results on verifica on of knowledge engi‐ neering methods applied for the first  me in cardiac surge‐ ry. Adequacy of applied research methods will be verified  and  unique  methodology  is going  to  be developed  in  this  project.   Provided  the  well‐defined  sources  of  informa on,  the  system knowledge is going to be sufficient to perform deep  and unbiased deduc on. Thorough and updated knowledge  installed in this universal system will enable a real life case  conference, where recommended mode of treatment along  with associated risks and medical costs are going to be pre‐ sented for the inquired medical case. Independent structu‐ re  elements  will  provide  various  data  source  installa on.  This will create unique expert tool available and applicable  30                                                                                                               Management Systems in Produc on Engineering 1(9)/2013                               T. HRAPKOWICZ et al. — Decision making process in cardiac surgery – concept of building an expert system    to  individual requirements (medical providers, pa ents,  in‐ surers, students, etc.) worldwide.  REFERENCES  [1]  Adlassnig  K.P.,  Blacky  A.,  Koller  W.:  Ar ficial‐ intelligence‐based  hospital‐acquired  infec on  control.  Stud  Health  Technol  Inform.  Vol.  149,  2009,  pp.  103‐ 110.  [2]  Avci  E.:  A  New  Expert  System  for  Diagnosis  of  Lung  Cancer:  GDALS_SVM.  J  Med  Syst.  Vol.  36,  2012,  pp.  2005‐2009.  [3]  Dassen W. R., Karthaus V. L., Talmon J. L., Mulleneers  R. G., Smeets J. L., Wellens H. J.: Evalua on of new self‐ learning  techniques  for  the  genera on  of  criteria  for  differen a on of wide‐QRS tachycardia in supraventri‐ cular tachycardia and ventricular tachycardia. Clin Car‐ diol. Vol. 18, 1995, pp. 103‐108.  [4]  Gallivan S., Stark J., Pagel C., Williams G., Williams W.:  Dead  reckoning:  can  we  trust  es mates  of  mortality  rates in clinical databases? European Journal of Cardio‐ thoracic Surgery. Vol. 33, 2008, pp. 334‐340 .  [5]  Gołuchowski J. (red.): Wprowadzenie do inżynierii wie‐ dzy. Difin, Warszawa 2011.  [6]  Heden B., Edenbrandt L., Haisty Jr. W. K, Pahlm O.: Ar ‐ ficial neural networks for the electrocardiographic dia‐ gnosis  of  healed  myocardial  infarc on.  Am  J  Cardiol.  Vol. 74, 1994, pp. 5‐8.  [7]  Kopecky D, Adlassnig K. P, Prusa A. R, Hayde M, Hayas‐ hi  Y,  Panzenböck  B,  Rappelsberger  A.,  Pollak  A.:  Knowledge‐based genera on of diagnos c hypotheses  and  therapy  recommenda ons  for  toxoplasma  infec‐ ons in pregnancy. Med Inform Internet Med. Vol. 32 (3), 2007, pp. 199‐214.  [8]  Kempa  A.:  Specyfika  sztucznych  systemów  immunolo‐ gicznych.  Studia  i  materiały  Polskiego  Stowarzyszenia  Zarządzania  Wiedzą,  Polskie  Stowarzyszenie  Zarządza‐ nia Wiedzą. Nr 37, 2011.  [9]  Kwiatkowska M., Michalik K., Kielan K.: Computa onal  Representa on  of  Medical  Concepts:  A  Semio c  and  Fuzzy Logic Approach, [in:] So  Compu ng in Humani‐ es  and  Social  Sciences.  Springer‐Verlag.  Heildelberg  2012.  [10]  Makal  J.,  Nazarkiewicz  A.,  Oniśko  A.,  Orzechowski  P.:  System  ekspertowy  do  wspomagania  diagnozy  łagod‐ nego przerostu prostaty. Pomiary Automatyka Roboty‐ ka. Nr 7‐8, 2004.  [11]  Maruszewski M., Kempa A., Michalik K., Zacny B., Hrap‐ kowicz T.: Koncepcja bazy wiedzy systemu ekspertowe‐ go dla kardiochirurgii [w:] Sobczak A. (red.): Technolo‐ gie  informatyczne  w  administracji  publicznej  i  służbie  zdrowia. Zarządzanie cyfrową transformacją organizacji  publicznych, Roczniki Kolegium Analiz Ekonomicznych,  2012.  [12]  Rabelo Júnior A., Rocha A. R., Oliveira K., Souza A., Xi‐ menes  A.,  Andrade  C.,  Onnis  D.,  Olivaes  I.,  Lobo  N.,  Ferreira N., Werneck V.: An expert system for diagnosis  of acute myocardial  infarc on with ECG analysis. Ar f  Intell Med. Vol. 10(1), 1997, pp. 75‐92.  [13]  Sołdek J., Pechmann P.: Wiedza – istota, pozyskiwanie i  wykorzystanie w przedsiębiorstwie, w: Sołdek J. (red):  Metody  Informatyki  Stosowanej.  Kwartalnik  Komisji  Informatyki  Polskiej  Akademii  Nauk  Oddział  w  Gdań‐ sku. Tom 11, nr 1, 2007, s. 23‐38.   [14]  Williams  W.  G.:  Uses  and  Limita ons  of  Registry  and  Academic  Databases.  Seminars  in  Thoracic  and  Car‐ diovascular Surgery: Pediatric Cardiac Surgery Annual.  Vol. 13, Issue 1, 2010, pp. 66‐70.  [15]  Zacny  B.:  Konfirmacyjna  analiza  czynnikowa  w  bada‐ niach ekonomicznych, [w:] Zeliaś A. (red.): XX Semina‐ rium  Ekonometryczne  im.  Profesora  Zbigniewa  Paw‐ łowskiego. Materiały z XXXVIII Konferencji Statystyków,  Ekonometryków i Matematyków Akademii Ekonomicz‐ nych Polski Południowej. AE w Krakowie, Kraków 2004,  s. 359‐369.  [16]  Yang T. F., Devine B., Macfarlane P. W.: Ar ficial neural  networks  for  the  diagnosis  of  atrial  fibrilla on.  Med  Biol Eng Comput. Vol. 32, 1994, pp. 615‐619.  dr n. med. Tomasz Hrapkowicz  dr n. med. Marcin Maruszewski   CompassMedica Sp. z o. o.,   ul. Karłuszowiec 9, 42‐600 Tarnowskie Góry, Poland  e‐mail: t.hrapkowicz@compassmedica.com  m.maruszewski@compassmedica.com    dr Anna Kempa  dr Krzysztof Michalik  dr Bogna Zacny  University of Economics in Katowice  Faculty of Informa cs and Communica on  ul. 1 Maja 50, 40‐287 Katowice, Poland  e‐mail: kempa@ue.katowice.pl  krzysztof.michalik@ue.katowice.pl  zacny@ue.katowice.pl  << /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Dot Gain 20%) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Error /CompatibilityLevel 1.4 /CompressObjects /Tags /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /CMYK /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams false /MaxSubsetPct 100 /Optimize false /OPM 1 /ParseDSCComments true /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo true /PreserveFlatness true /PreserveHalftoneInfo false /PreserveOPIComments true /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Apply /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true ] /NeverEmbed [ true ] /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 300 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 300 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.50000 /EncodeColorImages false /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /ColorImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 300 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 300 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages false /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /GrayImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 1200 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 1200 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 >> /AllowPSXObjects false /CheckCompliance [ /None ] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile () /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description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> /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000410064006f006200650020005000440046002065876863900275284e8e9ad88d2891cf76845370524d53705237300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef69069752865bc9ad854c18cea76845370524d5370523786557406300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /CZE <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> /DAN <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> /DEU <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> /ESP <FEFF005500740069006c0069006300650020006500730074006100200063006f006e0066006900670075007200610063006900f3006e0020007000610072006100200063007200650061007200200064006f00630075006d0065006e0074006f00730020005000440046002000640065002000410064006f0062006500200061006400650063007500610064006f00730020007000610072006100200069006d0070007200650073006900f3006e0020007000720065002d0065006400690074006f007200690061006c00200064006500200061006c00740061002000630061006c0069006400610064002e002000530065002000700075006500640065006e00200061006200720069007200200064006f00630075006d0065006e0074006f00730020005000440046002000630072006500610064006f007300200063006f006e0020004100630072006f006200610074002c002000410064006f00620065002000520065006100640065007200200035002e003000200079002000760065007200730069006f006e0065007300200070006f00730074006500720069006f007200650073002e> /ETI <FEFF004b00610073007500740061006700650020006e0065006900640020007300e4007400740065006900640020006b00760061006c006900740065006500740073006500200074007200fc006b006900650065006c007300650020007000720069006e00740069006d0069007300650020006a0061006f006b007300200073006f00620069006c0069006b0065002000410064006f006200650020005000440046002d0064006f006b0075006d0065006e00740069006400650020006c006f006f006d006900730065006b0073002e00200020004c006f006f0064007500640020005000440046002d0064006f006b0075006d0065006e00740065002000730061006100740065002000610076006100640061002000700072006f006700720061006d006d006900640065006700610020004100630072006f0062006100740020006e0069006e0067002000410064006f00620065002000520065006100640065007200200035002e00300020006a00610020007500750065006d006100740065002000760065007200730069006f006f006e00690064006500670061002e000d000a> /FRA <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> /GRE <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a stvaranje Adobe PDF dokumenata najpogodnijih za visokokvalitetni ispis prije tiskanja koristite ove postavke. Stvoreni PDF dokumenti mogu se otvoriti Acrobat i Adobe Reader 5.0 i kasnijim verzijama.) /HUN <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> /ITA <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> /JPN <FEFF9ad854c18cea306a30d730ea30d730ec30b951fa529b7528002000410064006f0062006500200050004400460020658766f8306e4f5c6210306b4f7f75283057307e305930023053306e8a2d5b9a30674f5c62103055308c305f0020005000440046002030d530a130a430eb306f3001004100630072006f0062006100740020304a30883073002000410064006f00620065002000520065006100640065007200200035002e003000204ee5964d3067958b304f30533068304c3067304d307e305930023053306e8a2d5b9a306b306f30d530a930f330c8306e57cb30818fbc307f304c5fc59808306730593002> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020ace0d488c9c80020c2dcd5d80020c778c1c4c5d00020ac00c7a50020c801d569d55c002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /LTH <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> /LVI <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> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken die zijn geoptimaliseerd voor prepress-afdrukken van hoge kwaliteit. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <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> /POL <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> /PTB <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> /RUM <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> /RUS <FEFF04180441043f043e043b044c04370443043904420435002004340430043d043d044b04350020043d0430044104420440043e0439043a043800200434043b044f00200441043e043704340430043d0438044f00200434043e043a0443043c0435043d0442043e0432002000410064006f006200650020005000440046002c0020043c0430043a04410438043c0430043b044c043d043e0020043f043e04340445043e0434044f04490438044500200434043b044f00200432044b0441043e043a043e043a0430044704350441044204320435043d043d043e0433043e00200434043e043f0435044704300442043d043e0433043e00200432044b0432043e04340430002e002000200421043e043704340430043d043d044b04350020005000440046002d0434043e043a0443043c0435043d0442044b0020043c043e0436043d043e0020043e0442043a0440044b043204300442044c002004410020043f043e043c043e0449044c044e0020004100630072006f00620061007400200438002000410064006f00620065002000520065006100640065007200200035002e00300020043800200431043e043b043504350020043f043e04370434043d043804450020043204350440044104380439002e> /SKY <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> /SLV <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> /SUO <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> /SVE <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> /TUR <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> /UKR <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> /ENU (Use these settings to create Adobe PDF documents best suited for high-quality prepress printing. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.) >> /Namespace [ (Adobe) (Common) (1.0) ] /OtherNamespaces [ << /AsReaderSpreads false /CropImagesToFrames true /ErrorControl /WarnAndContinue /FlattenerIgnoreSpreadOverrides false /IncludeGuidesGrids false /IncludeNonPrinting false /IncludeSlug false /Namespace [ (Adobe) (InDesign) (4.0) ] /OmitPlacedBitmaps false /OmitPlacedEPS false /OmitPlacedPDF false /SimulateOverprint /Legacy >> << /AddBleedMarks false /AddColorBars false /AddCropMarks false /AddPageInfo false /AddRegMarks false /ConvertColors /ConvertToCMYK /DestinationProfileName () /DestinationProfileSelector /DocumentCMYK /Downsample16BitImages true /FlattenerPreset << /PresetSelector /MediumResolution >> /FormElements false /GenerateStructure false /IncludeBookmarks false /IncludeHyperlinks false /IncludeInteractive false /IncludeLayers false /IncludeProfiles false /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe) (CreativeSuite) (2.0) ] /PDFXOutputIntentProfileSelector /DocumentCMYK /PreserveEditing true /UntaggedCMYKHandling /LeaveUntagged /UntaggedRGBHandling /UseDocumentProfile /UseDocumentBleed false >> ] >> setdistillerparams << /HWResolution [2400 2400] /PageSize [612.000 792.000] >> setpagedevice 
work_dbs6u7zxgvhfxduglseygwz2yy	----	PII: 0898-1221(87)90228-8 Comput. Math. Applic. Vol. 14, No. 9-12. pp. 793-802, 1987 009%4943/87 $3.00+0.00 Printed in Great Britain Pergamon Journals Ltd E X P E R T S Y S T E M S F O R C A N C E R C H E M O T H E R A P Y S. GAGLIO Institute of Electrical Engineering, University of Palermo, Italy M. GENOVESI, C. RUGGmRO a n d G . SPrNELLI Department of Communications, Computer and System Science, Genoa University, Italy C. NICOLINI Chair of Biophysics, School of Medicine, Genoa University, Italy G . BONADONNA a n d P. VALAGUSSA National Cancer Institute, Milan, Italy Abstract--In the present paper we describe BREASTCAN and NEWCHEM, two expert systems for the characterization of optimal adjuvant cancer therapies. The purpose of BREASTCAN is to support physicians in the postoperative breast cancer therapy, on the basis of currently used therapy protocols. It was developed in Prolog and positively validated, referring to the chemotherapies used by oncologists for some patients in the National Cancer Institute in Milan. NEWCHEM is a system oriented to the development of new cancer therapies, based on pharmaco-cell kinetic modeling and the newest molecular knowledge about neoplastic process. The system is being built and at first it will be validated by experiments on mice. Our aim with NEWCHEM is to extend our knowledge base and our rules to incorporate also all the most advanced knowledge at the molecular and cellular level, both theoretical and experimental, to make readily accessible to the health-community a system which will be really as expert as the present state of the art allows. 1. I N T R O D U C T I O N T h e use o f A I (artificial intelligence) techniques, with its logical-deductive m e c h a n i s m s , t h a t r e p r o - duce the h u m a n r e a s o n i n g w i t h o u t the restriction o f h u m a n capabilities in m e m o r y a n d association, is p a r t i c u l a r l y suitable f o r t h o s e medical p r o b l e m s , in which y o u m u s t r e g a r d a g r e a t n u m b e r o f p r o g n o s e s a n d their effects a n d interactions. This is the case o f c a n c e r c h e m o t h e r a p y , used f o r 10 y e a r s with e n c o u r a g i n g results. I n the p r e s e n t p a p e r we describe B R E A S T C A N a n d N E W C H E M , t w o e x p e r t systems f o r the c h a r a c t e r i z a t i o n o f o p t i m a l a d j u v a n t c a n c e r therapies. T h e y are b o t h i m p l e m e n t e d o n a V A X 11/750 a t the D e p a r t m e n t o f C o m m u n i c a t i o n s , C o m p u t e r a n d System Science o f the U n i v e r s i t y o f G e n o a , with the c o - o p e r a t i o n o f the C h a i r o f Biophysics o f School o f Medicine o f the G e n o a U n i v e r s i t y a n d the N a t i o n a l C a n c e r I n s t i t u t e in Milan. T h e p u r p o s e o f B R E A S T C A N is t o s u p p o r t physicians in the p o s t o p e r a t i v e b r e a s t c a n c e r t h e r a p y , o n the basis o f c u r r e n t l y used t h e r a p y p r o t o c o l s . I t was d e v e l o p e d in P r o l o g a n d positively validated, referring t o the c h e m o t h e r a p i e s used b y oncologists f o r s o m e p a t i e n t s in the N a t i o n a l C a n c e r I n s t i t u t e in Milan. I n B R E A S T C A N the k n o w l e d g e is represented using production rules, t h a t are s t r u c t u r e s w h o s e f o r m is I F . . . T H E N , and frames, c o m p l e x d a t a structures f o r m o d e l i n g s t e r e o t y p e d situations. E v e r y f r a m e describes a c e r t a i n s i t u a t i o n o r hypothesis. N E W C H E M is a s y s t e m o r i e n t e d t o the d e v e l o p m e n t o f new c a n c e r therapies, b a s e d o n p h a r m a c o - c e l l kinetic m o d e l i n g a n d the newest m o l e c u l a r k n o w l e d g e a b o u t neoplastic process. T h e s y s t e m is being built a n d a t first it will b e v a l i d a t e d b y e x p e r i m e n t s o n mice. I t is c o n s t i t u t e d b y a set o f frames, a deduction rule system, a n agenda-based c o n t r o l system, a n hierarchical planning system, a d a t a b a s e a b o u t drugs, a qualitative reasoning system, m a t h e m a t i c a l m o d e l s f o r s i m u l a t i o n a n d o p t i m a l c o n t r o l o f p h a m a c o - e n z y m e ( M i c h a e l i s - M e n t e n ) a n d p h a r m a c o - c e l l processes. 2. B R E A S T C A N : A N E X P E R T S Y S T E M F O R P O S T O P E R A T I V E B R E A S T C A N C E R T H E R A P Y C h e m o t h e r a p y as a d j u v a n t t r e a t m e n t in p a t i e n t s with o p e r a b l e b r e a s t c a n c e r a n d histologically positive axillary n o d e s h a s b e e n used f o r a b o u t 10 years, achieving results t h a t s u p p o r t the use o f a d j u v a n t c h e m o t h e r a p y in clinical practice [1-3], 793 794 s. GAGLIO et al. Numerous clinical trials have been designed and activated, employing adjuvant chemotherapy alone or combined with endocrine or immunotherapy. Both study design and results o f adjuvant protocols are influenced by a great number o f prognostic variables (e.g. subgroups o f patients with different size o f primary tumor (T) and extent o f axillary node (N) involvement, presence or absence o f hormone receptors, quantities o f drugs to be administered and intervals between each drug and treatment cycles, combinations o f drugs, optimal treatment duration) whose significance and possible interactions need to be assessed, as well as by a great variety o f clinical problems often originating from chemotherapy during the postoperative course o f breast cancer (e.g, treatment discontinuation because o f toxicity, continuous low drug dosage, different forms o f salvage regimen, occurrence o f general medical problems not related to adjuvant therapy). It is therefore felt that AI techniques [4] could make valuable contributions in this field. In particular it would be worth while to set up an expert system to assist physicians in the selection and performance o f optimal treatment--maximizing therapeutic effectiveness and limiting toxicity. The use o f the expert system should facilitate an optimal decision making at each treatment cycle by taking into account all foreseeable medical variables. An expert system for oncology protocol management, focused on the treatment o f Hodgkin's Disease and the non-Hodgkin's lymphomas, has been recently set up [5]. The present paper describes an expert system for adjuvant breast cancer chemotherapy--based on few widely accepted clinical protocols [2]--which has been implemented to assist clinicians treating patients by standard chemotherapy. 2.1. Outline o f B R E A S T C A N B R E A S T C A N is written in Prolog [6]--a programming language that has been used in many areas o f AI research--that has been chosen because it allows to quickly write clear, concise and readable programs, and provides an efficient built-in deduction system. B R E A S T C A N is written employing production rules--that is general statements about objects and their relationships--and frames--that is data structures that model stereotype situations. Production rules encode single pieces o f expert knowledge and are written as Prolog clauses. They specify particular solutions to a given problem according to given conditions. As an example, the following B R E A S T C A N rule rule(4,3,ctx,2,Time),300,Atruth,10):-- f (look__for (platelets,2,Time, Platelets,Truth 1 )), Platelets < 1 0 0 0 0 0 , f (look__for (leukocytes,2,Time, Leukocytes,Truth2)), Leukocytes < 3800, f (Iook__for(n__delays,2,Time, Ndelays, 10) ), Ndelays = 2, min(Truthl ,Truth2,Atruth). establishes that the suggested dose o f Ctx (cyclophosphamide) is 300 mg/m 2 if platelets count is less than 100,000, leukocytes count is less than 3800, and two consecutive delays have already occurred in drug administration. A rule is characterized by a truth value which is combined with truth values resulting from the satisfaction o f antecedent conditions to provide a measure o f evidence o f the conclusion. Production rules have been grouped, in BREASTCAN, into frames. Each frame corresponds to a given situation or hypothesis. One frame (named "initial") contain the rules establishing whether the patient is eligible for chemotherapy Or not; another frame (named "registration") manages the description o f the patient pathology prior to chemotherapy, to be used for adjuvent therapy choice and to be recorded for subsequent statistical studies that the user may want to carry out, and recommends one treatment protocol, each treatment protocol specifying one chemotherapy cycle. The user can then select one treatment protocol (accepting the system recommendation or not). At every stage the user can obtain explanations from the system, that can describe the pathway followed to reach any stage o f consultation. Frames are organized in a tree-like structure as shown in Fig. 1. Expert systems for cancer c h e m o t h e r a p y 795 + . . . . . . . . . . . . I N I T I A L [ FRAME DESCRIPTOR + . . . . . . . . . . . . REGISTRATION [ SUBFRAME POINTERS + . . . . . . . . . . . . FLOW-SHEET I S T E P S J + . . . . . T I l I R U L E S + . . . . . . . . . . . . PRIMARY TREATMENT - - + . . . . . T2 l I DEFAULT RULES I + ..... Tn I TRIGGERS + . . . . . . . . . . . . FOLLOW-UP I COMPLETION ACTIONS l + ..... FTI I TREATMENT I U T I L I T I E S + . . . . . . . . . . . . UPON . . . . . . . . . . + . . . . . FT2 FAILURE I ITEM DESCRIPTORS + . . . . . FTn Fig. 1 Fig. 2 The treatment protocols are subframes of one frame (named "treatment") that takes into account the stage o f the treatment, the possible presence of pathologies developed during it, and suggests (activating one of the subframes) the drugs to be administered and their quantities, and the date of the next drug administration. 2.2. Control structure The control system activates and executes frames. A frame can be in one of the following states: --inactive; --active (that is ready for execution); - - c u r r e n t (in execution); --completed. The control system provides a number of procedures not available directly in Prolog, designed to assist consultation. Frames are selected for instantiation and execution according to an "applicability value", that evalutes the agreement between the specific situation considered and the situation to which the frame relates; this is often accomplished comparing, for each frame, the postsurgical conditions characterizing the frame with the patient conditions. 2.3. Frame structure and execution Frames consist of a sequence of steps, rules and facts. Steps consist of groups of statements that are executed sequentially, and steps are executed sequentially. Some steps are conditioned, that is their execution depends on a condition to be tested: if the condition is not true, the step is not executed and the control passes to the following step. The steps of a frame encode the procedural knowledge ("what to do") of the stereotype situation modelled by the frame. The rules of a frame encode the declarative knowledge (what is true and "how to do" things within the given situation). The facts are pairs item--value that encode what is known about particular situations (measures and information inferred by the system. Each value is provided with a descriptor containing the way the value has been determined (asked, deduced, default) and certainty factor, which expresses the extent of certainty associated with the value. Certainty factors are propagated and modified in the course of rule application as described in Buchanan and Shortliffe [7]. The structure o f a frame is shown in Fig. 2. The frame descriptor contains some general information about the frame, as its name, the number of steps, its possibility, and the condition for activation. Subframe pointers indicate the subframes of the frame. Triggers are forward rules that perform particular actions when critical values of the parameters are added to the set of items. Completion actions are performed when the frame enters the completed status, for instance the activation of another frame. 796 S. GAGLIO et al. Utilities are auxiliary procedures used within the steps. Item descriptors encode the constraints to be satisfied by the values of the items, the questions to be asked o f the user concerning the items, and take care of data integrity (for instance, menopause age cannot be greater than the current age of the patient). When, during the execution of a frame, the value of an item is needed, the system first consults its data base (the pairs item--value in the frames); if it does not find it, it tries to apply the rules of the frame to deduce it; finally, if it fails, it asks the user the value; when the answer is not even provided by the user, the system uses default rules and values. 2.4. Interfacing with the user BREASTCAN offers the possibility to obtain indications about the line of reasoning that leads backwards from the current answer by the system. This capability is useful in many respects: realizing how the system comes to a conclusion enables users to make the most o f the consultative advice and also helps users in difficult, "non-standard" situations, in which they may wish to violate a rule; moreover, it makes it easier for the user to make changes in the program. When starting consultation about a patient for the first time, the user is asked questions that enable the system to decide whether chemotherapy is appropriate for that patient. If the answer is no the system gives indications about the reasons of this conclusion and some advice about that particular case. At the end of this stage the user can choose whether he wishes to carry on or not (irrespective of the conclusions of the system about the eligibility of the patient). The following stage, relating to the description o f the patient pathology prior to chemotherapy, consists of six parts; at the end of each of them the data entered by the user in that section are displayed and the user is enabled to change values that he may have entered by mistake. The system now suggests one or more treatments and the user can choose which treatment he wishes to perform (following the suggestions by the system or not). 2,5. Assessment and further developments BREASTCAN has been proven capable to closely reproduce the actual decision taken over the period of 6 months by medical oncologists at the National Cancer Institute of Milan, Italy, in the treatment of a few patients, taken as test-cases in a preliminary retrospective confirmation of the system validity. Our next step is then to extend this study to a large number o f patients and to incorporate into the decision-making process all other possible strategies suggested by other alternative treatment protocols from worldwide clinical trials on the breast cancer. 3. N E W C H E M : AN EXPERT SYSTEM FOR THE T R E A T M E N T OF D I S S E M I N A T E D C A N C E R Since early 1974 a comprehensive program was developed at the Biophysics Division o f Temple University in Philadelphia (U.S.A.), aiming to develop a rational basis for the chemotherapy of cancer [8]. This approach resulted in a widely interdisciplinary effort which, by a constant feedback between experimentation on animals and theoretical simulations, has been able to predict and explain multiple drug-action and interaction at molecular [9] and cellular [10] level for a variety o f normal and cancer tissues. Through the aid o f optimal control theory [11] such a complex modeling can furthermore suggest the optimal treatment, with the sequence of rest-periods and drug administration at the proper timing and dosage. Recent preliminary results have indeed been quite encouraging on the treatment of lung metastates from melanoma B-16 in mice [12]. There appears to be however a basic premise for the successful utilization of the sophisticated pharmaco-enzyme and pharmaco-cell kinetic modeling, namely the detailed molecular knowledge of the drug-tissue metabolism and o f the physico-chemical properties o f both normal and cancer cell in the various functional state (cycling, non-cycling and in varying stage o f differentiation). Fortunately, modern biophysical techniques [13] permit such a characterization at single cell level with high accuracy and frequently on real-time. The transfer of such a complex and multifold approach to routing medical practice appears however nearly prohibitive, for its highly analytical Expert systems for cancer chemotherapy 797 ÷ . . . . . . . . . . . . . ÷ I AGENDA I ÷ . . . . . . . . . ÷ ÷ . . . . . . . . . . . . . ÷ + . . . . . . . . . . . ÷ I l I I I I I RULES I I TASK - I I I PLANNER I I I i T i i + . . . . . . . . . + I TASK - 2 I + . . . . . . . . . . . * I i . . . I I i . . . . . . . . . . . . . . . I I I . . . . . . . . . v I v v ÷ . . . . . . . . . . . . ÷ ÷ . . . . . . . . . . . . . . . . . ÷ + . . . . . . . . . . . . . ÷ I I < - - I I t QUALITATIVE I I DEDUCTION t TASK I I I • ->I I l<--I REASONING l I SYSTEM ) PROCESSOR I I I I I--> I - I SYSTEM I ÷ . . . . . . . . . . . . + . . . . . . . . . . . . . . . . . . I + . . . . . . . . . . . . . + ÷ . . . . . . . ÷ I I v ÷ . . . . . . . . . . + + . . . . . . . . . . . ÷ + . . . . . . . . . . . + I I I I I I I TASK I I DRUGS l I USER I I I I i I I I RULES I I DATA BASE I I INTERFACE I I I I I I I ÷ . . . . . . . . . . + + . . . . . . . . . . . + + . . . . . . . . . . . + + . . . . . . . . . . + + . . . . . . . . . . . . . . . . . + I I I I +-->1 FRAMES i < . . . . + 1 MATHEMATICAL I I I I I + . . . . . . . . . . + I MODELS I I I ÷ . . . . . . . . . . . . . . . . . ÷ Fig. 3 nature and for the large number of parameters involved (constantly expanding due to the progress in cell and molecular biology). We have then recently considered to use AI to bridge such a gap, by developing an ad hoc expert system named N E W C H E M to guide our present animal experimentation but with the aim to improve in the near future human cancer treatment. 3. I. The structure o f N E W C H E M The structure of N E W C H E M is shown in the diagram of Fig. 3. It is based on many paradigms of AI, since it puts together into one integrated system, a set of components, which are: a production rule-based system, a frame stucture, an agenda-based control, a planner, a data base about drugs, a qualitative reasoning system, a blackboard system. All those components cooperate to produce a plan (i.e. a sequence of actions) to be carried out by the user to achieve a given goal. We adopt such a composite structure, because we believe that intelligent behaviour and accurate answers can be only obtained using specific techniques for the different kinds of knowledge and strategies to be embodied by the system. This is in agreement with current trends in expert system research. For instance, more recent systems like C E N T A U R [14], ONCOCIN [5] and ABEL [15] use multiple representations of knowledge. Furthermore, N E W C H E M uses mathematical models for pharmaco-cell kinetics [12], pharmaco-enzyme kinetics [9], and optimal control theory [11], written in F O R T R A N , which are queried when the system needs precise answers concerning simulation and optimization of parameters at various levels of the decision-making process. The system is written in Franz Lisp and uses the PEARL package [15] for efficient storage and retrieval, and the FLAVORS package for object oriented programming. The interaction between N E W C H E M and the experimenter is as follows: (a) The system asks the experimenter to provide some data describing the status of a subject to be treated. (b) The system produces an optimal treatment plan together with a set of justifica- tions for its choices and the expected results. 798 s. GAGLIO et aL (c) The experimenter can propose changes in the plan and ask about possible consequences. (d) A new treatment plan is proposed whenever the results are different from the expected ones. In the following a short description of each component of the system is given. 3.2. Frames Frames in N E W C H E M are associated with different points of view of the problem. They collect specific data and knowledge about inferences and actions concerning the specific domain. Data are related to different aspects of the case being examined. Frames are associated respectively with primary tumor, metastates, different tissues and organs (like bone marrow and small intestine). Different frames are also associated with discriptions respectively at the cellular level and the molecular level. For each aspect, a frame acts as a blackboard in which all data are collected together with an indication about how they have been obtained and a measure o f uncertainty. Consistency checking among data is also performed within the frame. Frames also contain action blocks, which are sequences of tasks to be performed to fill the data entries and for different operations of the whole system. Tasks may refer, for instance, to the activation of other frames, to queries to the declarative knowledge base and to the mathematical model, and to planning activity. A particular action block is the activation block, to be executed at the activation o f the frame. The way a task is executed is based on a set of task rules, that for each task select the appropriate operations, depending on the data known so far (the known states o f the subject). As an example, in the frame "treatment", a rule for the task "select a treatment plan" is the following: (Task-rule (Task "select a treatment plan") (Context treatment) (Condition (isknown general__strategy)) (Conclude~action '(progn (enter__date) (enter__new__data) (deduce general__strategy) (ask__user) (deduce strategy) (ask__user) (deduce treatment) (ask__user) (find plan) (ask_.user)))) Some data within a frame can be inferred from other data by means o f backward production rules. For instance, the rule (Rule (Ruleno 45) (Context treatment) (Condition (and ( > =subject performance~tatus 50) (for ?organ in (kidney, heart, liver, lungs) (same ?organ sufficiency satisfactory)) (same__strategy kill__tumor)) ) (ConcludLvalue_treatment_type drug__combination)) gives treatment__type the value drug__combination if the performance status of the subject is > = 50, the sufficiency o f kidney, heart, liver and lungs is satisfactory and the strategy is to kill the tumor. Expert systems for cancer chemotherapy 799 W i t h an item (or d a t a entry) in a frame, o n e can associate a " t r i g g e r " , which p r o m o t e s a p p r o p r i a t e actions when a critical value for the item is obtained. Items can have multiple values, d e t e r m i n e d by m ean s o f different rules. Different values have, however, associated different scores representing their uncertainty. At a given time a task can be active or inactive. W h e n a fram e enters the active status, the set o f tasks, m a k i n g its activation block, is loaded in the ag en d a o f the system, waiting to be executed. G r o u p s o f frames are organized in a tree-like fashion w h en ev er some o f t h em can be considered m o r e specific points o f view. In such a case a hierarchical relation " s u b f r a m e " holds between pairs o f frames. 3.3. The agenda and the task processor T h e c o n t r o l o f the system uses an agenda o f waiting tasks a n d a task processor. T h e tasks in the agenda belong to active frames. F o r each o f them the following d a t a are maintained: • T h e n a m e o f the task. • T h e frame to which it belongs. • A priority value. • A justification. At the end o f the execution o f a task, a new task is selected acco rd i n g to its priority. T h e task processor is a p r o b l e m solver that behaves in the following ways: • F o r each task, it searches the associated task rules (within the fram e the tasks belong to) to find a set o f possible actions. Each action is a L I S P co d e to which it is assigned a score, expressing h o w m u c h it is a p p r o p r i a t e fo r the c u r r e n t situation. • It selects the action with the highest score • It executes the selected action. Actions m a y require to insert new tasks in the agenda. F o r each o f t h e m a p ri o ri t y value is c o m p u t e d and the c u r r e n t task is inserted in the list o f its justifications. 3.4. The deduction system T h e d e d u c t i o n system is invoked whenever in the system a value fo r an item is required. T h e d e d u c t i o n system first explores the frame to which the item belongs t o see if a value is already k n o w n . I f this is n o t the case, it looks f o r rules to deduce it. Rules are first searched fo r in the c u r r e n t f r a m e (the f r a m e o f the c u r r e n t task), then in the frames which are ancestors o f the c u r r e n t one in the subframe hierarchy, and, finally, in the fram e to which the item belongs. I f n o rule is f o u n d , the value is asked o f the user or a default value, if specified, is used. T h e d e d u c t i o n system provides for an item, which is an attribute associated with an object, a list o f P E A R L structures o f the following kind: (Item (Object lisp) (Attribute lisp) (Context lisp) (Time lisp) (Value lisp) (Uncertainty integer) ) E a c h structure expresses a value for the item at a given time with an associated u n c e r t a i n t y value. 3.5. Planner T h e task o f the p l a n n e r is to p r o d u c e a plan fo r the t r e a t m e n t o f the subject, i.e. a partially o r d e r e d set o f actions to be p e r f o r m e d by the experimenter. T h e o rd eri n g o f the actions refers to a t e m p o r a l sequence, in the sense that some actions m u st be p e r f o r m e d b efo re others. T h e p l a n n e r is i n v o k e d by the task executor an d receives as i n p u t an initial p l an t o refine. Refinement is p e r f o r m e d by substituting a given action in the set with a n o t h e r partially o r d e r e d 800 S. OAOLIO e t al. ;Task_plan RS : ;IF the type of tumor is metastatized, ; its progress istanzla reduced and ; the state of the treatment istanzia in progress ;THEN change only one drug (ci Task_plan R3 (Name First_strategy) (Number 3) (Context General_strategy) (Type father) (Tax_father nil) (Tax_child nil) (Condition (prog (Date Listglob) (setq Date (daymonthyear)) (setq Listglob (list 'cell_descr 'cell_descr 0)) (return (fu_and (same 'type 'tumor 'metastatized Date Listglob) (same 'progress 'tumor 'reduced Date Listglob) (same 'status 'treatment 'in_progress Date Listglob))))) (Action (conclude 'First_strategy 'change_only_drug 9.2)) (Description (prog () (printterpr " I find that the TYPE of TUMOR is METASTATIZED (CERTAINTY FACTOR : ~f) ,, (same 'type 'tumor 'metastatized (daymonthyear) (list 'cell_descr 'cell_descr 0)) 5 I) (printterpr " the PROGRESS of TUMOR is REDUCED (CERTAINTY FACTOR : %f) ,, (same 'progress 'tumor 'reduced (daymonthyear) (list 'cell_descr 'cell_descr e)) fi 2) (printterpr " and the STATUS of TREATMENT is IN PROGRESS (CERTAINTY FACTOR : ~f) . (same 'status 'treatment 'in_progress (daymonthyear) (list 'cell_descr 'cell_descr 0)) 5 2) (cprinttab " THEN I recommend to CHANGE ONLY ONE DRUG " nil 5)))) (insertdb R3) Fig. 4 set o f more specific actions. This is accomplished using refinement rules, which are context- dependent production rules, that replace a given action in their left part with the set o f actions specified in their fight part. Consistency tests are also performed on the overall plan to resolve interactions and conflicts. This planning is similar to the hierarchical planning performed by N O A H [17]. The actions specified in a plan refer to different kinds o f intervention on the subject, mainly surgery, radiology and administration o f drugs. Some constraints, describing the particular modalities o f the intervention (for instance, the properties o f the drugs to be administered) are associated with these actions. These constraints are used to select, at the end o f the planning activity, the most appropriate intervention through a query to the declarative knowledge base--for instance, the specific drug needed. This selection is made only when the plan is completely refined, in order to avoid an early choice that can compromise the whole plan. This is what is called a least commitment strategy. Constraints can also be propagated within the plan as in M O L G E N [18]. The plan is associated as a value with the item plan in the frame treatment. An example o f a rule o f the Planner is shown in the Fig. 4. 3.6. The data base about drugs The data base about drugs o f the system contains data concerning all drugs that are to be used during the treatment considered by the system and their properties. Drugs are represented as structured types o f LISP objects, stored in an internal form as blocks o f memory, regarding logical groupings o f betorogeneous data as slots and slot fillers. One main structure defines the way in which all drugs are characterized, while each drug is represented as an individual instance o f this structure. The main structure can be regarded as a reference model o f a drug in which the greatest number o f attributes and properties are available, Expert systems for cancer chemotherapy 801 and each individual drug is represented filling up all known attributes o f the main structure and specifying properties to the greatest possible detail. This knowledge representation is implemented in P E A R L an AI language allowing to create hierarchically defined slot filler representations and to handle them efficiently, by easy inserting and fetching. P E A R L is implemented in LISP, so the knowledge base can be directly addressed within a LISP system as the one described in the previous sections. Specific rules contained in the system directly address attributes o f the drugs represented in P E A R L in the declarative data base. The main structure adopted is implemented in P E A R L as follows: (drug (name symbol) (type symbol) (main___attribute symbol) (class symbol) (toxicity symbol) (action struct)) According to this main structure, each specific drug considered by the system is represented as follows: (drug (name cyclophosphamide) (type alchilant) (main__attribute polifunctional) (action (type inhibits) (object mitosis))) 3. 7. The qualitative reasoning system The qualitative reasoning system uses qualitative process theory, as developed by Forbus [19], to reason about processes at a cellular level. This component is used within the whole system to perform qualitative simulations in order to "roughly" predict the consequence o f specific actions. In qualitative process theory physical parameters are characterized by a quantity space in which only intervals o f values are considered. Usually, what interests is if a quantity is positive, negative or null, and if it is increasing, decreasing or stationary. Furthermore, physical situations and processes are described symbolically, in a language based on predicate calculus. F o r a given process, the objects involved, the conditions in which the process takes place, the relations holding among the parameters, and the causes o f change (influences) are specified. The qualitative reasoning sytem is implemented in an object oriented style o f programming which is provided by the "flavors package" o f Franz Lisp. As an example the inhibition o f D N A replication by a substance can be described as follows: (make_instance 'process ':individuals '(DNA cell ?S) ': preconditions '(and (substance ?S) (in hibitor__D N/k~synthesis ?S)) ': influences '(I- (Am (synthesis DNA)) (Am (concentration ?S)))) The set o f possible sequences o f processes that follow a given action on a cell or on a cell population (for instance, an increase in concentration o f a substance) is the result o f a qualitative simulation. This result is useful during the planning process in order t o take care o f the possible consequences o f the actions that have been selected. 802 S. GAGLIO et al. 3.8. T h e m a t h e m a t i c a l m o d e l T h e m o d e l i n g c o n c e r n s v a r i o u s levels o f d e c i s i o n - m a k i n g a n d a t t e m p t s t o p r o v i d e a n i n t e g r a t e d a n d realistic a p p r o a c h t o o p t i m a l c a n c e r t r e a t m e n t . E a c h a s p e c t o f this m o d e l i n g h a s b e e n extensively d e a l t in p r e v i o u s p u b l i c a t i o n [5, 8, 9 - 1 1 , 18], a n d is b a s e d o n n u m e r o u s b i o p h y s i c a l p a r a m e t e r s , w h i c h a r e r e v i e w e d in details in a f o r t h c o m i n g b o o k [13]. E a c h m o d e l a c t s s y n e r g i s t i c a l l y w i t h i n a p r o p e r m u l t i p l e time scale, n a m e l y : - - a t t h e level o f s - m i n - h , a m o d e l ( D N A M E T ) f o r D N A e n z y m a t i c s y n t h e s i s d e t e r m i n e s m u l t i p l e D N A a n t i m e t a b o l i t e s i n t e r a c t i o n t o i n d u c e e i t h e r differential s y n c h r o n y o r differential killing b e t w e e n n o r m a l a n d c a n c e r tissues; a t t h e level o f h - d a y s , a cell kinetics m o d e l d e t e r m i n e s d r u g - t i s s u e s t o s u g g e s t ( S I V F I T ) o p t i m a l s h o r t t e r m s d r u g p r o t o c o l s ( t y p e , d o s a g e a n d timing); - - a t t h e level o f w e e k s - m o n t h s , a n o p t i m a l c o n t r o l - b a s e d m o d e l ( S I V F I T 2) p r e d i c t s l o n g - t e r m t i m i n g o f t r e a t m e n t , w h e r e rest p e r i o d s a n d c h a n g e s in t h e t y p e o f d r u g s a r e d e t e r m i n e d t a k i n g i n t o c o n s i d e r a t i o n l o n g - t e r m effects like d r u g resistance, d u e t o gene a m p l i f i c a t i o n a n d / o r cell m u t a t i o n . 4. C O N C L U S I O N T w o e x p e r t s y s t e m s in O n c o l o g y h a v e b e e n d e s c r i b e d . T h e first o f t h e m , B R E A S T C A N , w h i c h f o l l o w s a t r a d i t i o n a l a p p r o a c h , is b a s e d o n l y o n t h e t r a d i t i o n a l i n p u t d e r i v e d f r o m clinical e x p e r i e n c e w i t h e m p i r i c a l trials, w h i c h a r e c a r r i e d o u t o n r a n d o m i z e d p a t i e n t s w i t h the p r e v a i l i n g c h e m o t h e r a p e u t i c p r o t o c o l s . O u r a i m w i t h t h e s e c o n d o n e , N E W C H E M is i n s t e a d t o e x t e n d o u r k n o w l e d g e b a s e a n d o u r rules t o i n c o r p o r a t e a l s o all t h e m o s t a d v a n c e d k n o w l e d g e a t t h e m o l e c u l a r a n d c e l l u l a r level, b o t h t h e o r e t i c a l a n d e x p e r i m e n t a l , t o m a k e r e a d i l y accessible t o the h e a l t h - c o m m u n i t y a s y s t e m w h i c h will be really as e x p e r t as the p r e s e n t state o f t h e a r t allows. R E F E R E N C E S 1. G. Bonadonna and P. Valagussa, Chemotherapy of breast cancer: current views and results. Int. J. Radiat. Oncol. Biol. Phys. 9, 279-297 (1983). 2. S. E. Jones and S. E. Salmon, Adjuvant systemic therapy o f cancer IV. Grune & Stratton, New York (1984). 3. G. Bonadonna and P. Valagussa, Adjuvant systemic therapy for resectable breast cancer. J. Clin. Oncol. (in press). 4. J. C. Kunz, E. H. Shortliffe, B. G. Buchanan and E. A. Feigenbaum, Comparison of techniques of computer-assisted decision-making in medicine. Modeling and Analysis in Biomedicine, pp. 335-367. World Scientific, Singapore (1984). 5. E. H. Shortliffe, A. C. Scott, M. B. Bischoff, A. B. Campbell, W. Van Melle and C. D. Jacobs, ONCOCIN: An expert system for oncology protocol management. Proc. 7th Int. Joint Conf. on Artificial Intelligence, Vancouver, B.C., pp. 876-881 (1981). 6. W. F. Clocksin and C. S. Mellish, Programming in Prolog. Springer, New York (1981). 7. B. G. Buchanan and E. H. Shortliffe (Eds), Rule-Based Expert Systems: The MYC1N Experiments o f the Stanford Heuristic Programming Project. Addison-Wesley, Reading, Mass (1984). 8. C. Nicolini, The principles and methods of cell synchronization in cancer chemotherapy. Biophys. Biochem. Acta 458, 243-282 (1976). 9. C. Nicolini, A. Belmont, M. Grattarola, C. Moore and E. Milgram, Pharmaco-enzyme kinetics simulations of experimental interactions among multiple antineopolastic drugs. Biochem. Pharm. 28, 2891-2908 (1979). 10. S. Zietz, C. Desaive, S. Lessin and C. Nicolini, Modeling to determine dose dependence of drug and cell kinetic parameters. Comput. Biomed. Res. 13, 297-305 (1980). 11. S. Zietz and C. Nicolini, Mathematical approaches to optimization of cancer chemotherapy. Bull. Math. Biol. 41, 305-324 (1979). 12. C. Nicolini, S. Zietz, M. Tansy and S. Lessin, Modeling and analysis in cancer treatment. Modeling and Analysis in Biomedicine, (Ed. C. Nicolini) pp. 291-333. World Scientific, Singapore. 13. C. Nicolini, Biophysics and Cancer. Plenum, New York (1985). 14. J. S. Aikins, Prototipical knowledge for expert systems. Artificial Intelligence 20, 163-210 (1983). 15. R. S. Patil, P. Szolovits and W. B. Schwartz, Modeling knowledge of the patient in acid-base and electrolyte disorders. Artificial Intelligence in Medicine (Ed. P. Szolovits) pp. 191-226. Westview Press, Boulder, Co. (1982). 16. M. Deering, J. Faletti and R. Wilensky, Using the PEARL AI package. Technical Report Department of EECS, University of California, Berkeley (1982). 17. E. D. Sacerdoti, NOAH: Planning in a hierarchy o f abstraction spaces. Artificial Intelligence 5, 115-135 (1974). 18. M. Stefik, Planning with constraints. Artificial Intelligence 16, 111-140 (1981). 19. K. D. Forbus, Qualitative process theory. Artificial Intelligence 24, 85-168 (1984). 
work_dbv35sagorckbdvaob7uqlwxma	----	Probabilistic approaches to rough sets Y.Y. Yao Department of Computer Science, University of Regina, Regina, Saskatchewan, Canada S4S 0A2 E-mail: yyao@cs.uregina.ca Abstract: Probabilistic approaches to rough sets in granulation, approximation and rule induction are reviewed. The Shannon entropy function is used to quantitatively characterize partitions of a universe. Both algebraic and probabilistic rough set approx- imations are studied. The probabilistic approximations are defined in a decision-theoretic framework. The problem of rule induction, a major application of rough set theory, is studied in probabilistic and information-theoretic terms. Two types of rules are analyzed: the local, low order rules, and the global, high order rules. Keywords: rough set approximations, granular computing, decision-theoretic rough set model, rule induction, high order rules, belief functions 1. Introduction As a recently renewed research topic, granular computing is an umbrella term to cover any theories, methodologies, techniques and tools that make use of granules (i.e. sub- sets of a universe) in problem solving (Zadeh, 1979, 1997; Yao, 2000; Lin et al., 2002). The basic guiding principle of granular computing is to ‘exploit the tolerance for imprecision, uncertainty and partial truth to achieve tractability, robustness, low solution cost and better rapport with reality’ (Zadeh, 1997). This principle offers a more practical philosophy for real-world problem solving. Instead of searching for the optimal solution, one may search for good approximate solutions. The theory of rough sets provides a special and concrete model of granular computing (Pawlak, 1982, 1991). Three related issues of granulation, approximation and rule induction are central to studies of rough sets. Granulation of a universe involves the decomposition of the universe into families of subsets, or the clustering of elements into groups. It leads to a collection of granules, with a granule being a clump of points (objects) drawn together by indistinguishability, similarity, proximity or functionality (Zadeh, 1997). Granulation may produce either a single-level flat structure or a multi-level hierarch- ical structure (Yao, 2001a). The theory of rough sets uses equivalence relations to represent relationships between elements of a universe. An equivalence relation induces a single-level granulation, namely, a partition of the universe. A natural consequence of granulation is approximation. With respect to an equivalence relation, some subsets cannot be exactly expressed in terms of the equivalence classes and must be approximately represented by a pair of lower and upper approximations. By extending the approximations of subsets to a family of subsets, it is possible to study the approximation of a partition. Rule induction deals with finding relationships between concepts. With each granule, or a family of granules, representing instances of a certain concept, one can study rule induction in set-theoretic terms (Yao, 2001b). Approx- imations of subsets and families of subsets offer insights and methods for rule induction (Pawlak, 1991). Although mainstream research in rough set theory has been dominated by algebraic and non-probabilistic studies, probabilistic approaches have been applied to the theory ever since its inception (Wong & Ziarko, 1987; Pawlak et al., 1988; Yao et al., 1990; Yao & Wong, 1992; Düntsch & Gediga, 2001). More specifically, many authors implicitly used a probabilistic approach by counting the number of elements of a set. On the other hand, there is still a lack of systematic study of probabilistic approaches in a unified and general framework. The main objective of this paper is to provide a critical analysis and review of probabilistic and information- theoretic approaches to rough sets. With respect to three related issues of granulation, approximation and rule induction, we focus the discussion on probability-related measures. Such a comprehensive study provides a solid basis for further study of probabilistic rough set theory. 2. Approximation space and information granulation The underlying notion for granulation in rough sets is equivalence relations or partitions. Let U be a finite and nonempty universe. A binary relation E � U � U on U is called an equivalence relation if it is reflexive, symmetric and transitive. A partition of U is a collection of nonempty and pairwise disjoint subsets of U whose union is U. Each subset in a partition is also called a block. There is a one-to- Article ______________________________________ Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 287 one relationship between equivalence relations and parti- tions. For an equivalence relation E, the equivalence class ½x�E ¼ fy 2 UjyExg ð1Þ consists of all elements equivalent to x, and is the equivalence class containing the element x. The family of equivalence classes U=E ¼ f½x�Ejx 2 Ug ð2Þ is a partition of the universe U. On the other hand, given a partition p of the universe, one can uniquely define an equivalence relation Ep: xEpy , x and y are in the same block of the partition p ð3Þ In this paper, we will use equivalence relations and partitions interchangeably. The pair apr ¼ (U, E) is called an approximation space, indicating the intended applica- tion of the partition U = E for approximation (Pawlak, 1982). The partition U = E is commonly known as the quotient set and provides a granulated view of the universe. A coarse-grained view of the universe may arise in several ways. For instance, the equivalence relation is derived based on available knowledge. Due to a lack of informa- tion or vague information, some distinct objects cannot be differentiated (Pawlak, 1982). That is, the available information only allows us to talk about an equivalence class as a whole instead of many individuals. In some situations, it may only be possible to observe or measure equivalence classes. It may also happen that a coarse- grained view is sufficient for a particular problem (Zhang & Zhang, 1992; Zadeh, 1997). In the granulated view, equivalence classes are the basic building blocks and are called elementary (equivalence) granules. They are the smallest nonempty subsets that can be defined, observed or measured. From elementary granules, we can construct larger granules by taking unions of elementary granules. It is reasonable to assume that one can define, observe and measure these granules through the information and knowledge on the equivalence granules. The set of all definable granules, denoted by s(U = E), consists of the empty set +, the entire universe U, and unions of equivalence classes. The system s(U = E) is closed under set complement, intersection and union. It is a sub- Boolean algebra of the Boolean algebra formed by the power set 2 U of U and a s-algebra of subsets of U generated by the family of equivalence classes U = E. In addition, U = E is the basis of the s-algebra s(U = E). Each partition represents one granulated view of the universe. Granulated views induced by all partitions form a partition lattice. The order relation of the lattice is defined as follows. A partition p1 is a refinement of another partition p2, or equivalently p2 is a coarsening of p1, denoted by p1 � p2, if every block of p1 is contained in some block of p2, or equivalently each block of p2 is a union of some blocks of p1. In terms of equivalence relations, we have U = E1 � U = E2 if and only if E1 � E2. Given two partitions p1 and p2, their meet p14p2 is the largest partition which is a refinement of both p1 and p2, and their join p13p2 is the smallest partition which is a coarsening of both p1 and p2. The meet has all nonempty intersections of a block from p1 and a block from p2 as its blocks. The blocks of join are the smallest subsets which are exactly a union of blocks from both p1 and p2. In terms of equivalence relations, given two equivalence relations E1 and E2, the meet of U = E1 and U = E2 is defined by the equivalence relation E1 \ E2, and the join is defined by the equivalence relation (E1 [ E2)n, the transitive closure of relation E1 [ E2. The finest partition is given by {{x}jx 2 U} consisting of singleton subsets from U, and the coarsest partition is {U}. The partition lattice clearly shows the structure of different granulations of the universe. It can be used to search for a suitable level of granulation for problem solving (Zhang & Zhang, 1992). Many machine learning algorithms using rough sets are based on the search of the partition lattice (Yao & Yao, 2002). Information-theoretic measures can be used to quantify the degree of granularity of each partition (Lee, 1987; Düntsch & Gediga, 2001; Yao, 2003b). With respect to a partition p ¼ {A1, A2, . . . , Am}, we have a probability distribution Pp ¼ jA1j jUj ; jA2j jUj ; . . . ; jAmj jUj � � ð4Þ where j � j denotes the cardinality of a set. The Shannon entropy function of the probability distribution is defined by HðpÞ ¼ HðPpÞ ¼ � Xm i ¼ 1 jAij jUj log jAij jUj ð5Þ The entropy reaches the maximum value log jUj for the finest partition consisting of singleton subsets of U, and it reaches the minimum value 0 for the coarsest partition {U}. In general, for two partitions with p1 � p2, we have H(p1) � H(p2). That is, the value of the entropy correct- ly reflects the order of partitions with respect to their granularity. Additional support for using the entropy as a measure of generality can be seen as follows. We can re-express equation (5) as HðpÞ ¼ log jUj � Xm i ¼ 1 jAij jUj logjAij ð6Þ The first term is a constant independent of any partition. The quantity log jAij is commonly known as the Hartley measure of information of the set Ai. It has been used to measure the amount of uncertainty associated with a finite set of possible alternatives, namely the nonspecificity 288 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 inherent in the set (Klir & Folger, 1988). The function log jAij is a monotonic increasing transformation of the size of a set. It may be used to measure the granularity of the set. Large sets result in higher degrees of granularity than small sets. The second term of the equation is basically an expectation of granularity with respect to all subsets in a partition. It follows that we can use the following function as a measure of granularity for a partition: GðpÞ ¼ Xm i ¼ 1 jAij jUj log jAij ð7Þ In contrast to the entropy function, for two partitions p1 and p2 with p1 � p2, we have G(p1) � G(p2). The coarsest partition {U} has the maximum granularity value log |U|, and the finest partition {{x}|x 2 U} has the minimum granularity value 0. 3. Rough set approximations In this section we discuss approximations of sets and approximations of probabilities, as well as probabilistic approximations of sets. 3.1. Approximations of sets Consider an approximation space apr ¼ (U, E). The set of definable subsets is given by s(U = E). For a subset A � U, the greatest definable set contained in A is called the lower approximation of A, written apr U=E ðAÞ, and the least definable set containing A is called the upper approxima- tion of A, written aprU=EðAÞ. The subscript U = E indicates that the approximations are defined with respect to the partition U = E. When no confusion arises, we simply drop U = E. Lower and upper approximations can be expres- sed as aprðAÞ ¼ [ fXjX 2 sðU=EÞ; X � Ag aprðAÞ ¼ \ fXjX 2 sðU=EÞ; X Ag ð8Þ In terms of equivalence classes, lower and upper approx- imations can be expressed by aprðAÞ ¼ [ ½x�E � A ½x�E aprðAÞ ¼ [ ½x�E \ A 6¼ ; ½x�E ð9Þ The lower approximation aprðAÞ is the union of equiva- lence classes which are subsets of A. The upper approxima- tion aprðAÞ is the union of equivalence classes which have a nonempty intersection with A. One may interpret apr; apr: 2U ! 2Uas two unary set- theoretic operators, called approximation operators. The system ð2U; :; apr; apr; \; [Þ is called a Pawlak rough set algebra (Yao, 1996). It is an extension of the set algebra (2 U , :, \ , [ ). Properties of approximation operators, pertinent to our discussion, are summarized below: (i) aprðAÞ ¼ :aprð:AÞ; aprðAÞ ¼ :aprð:AÞ; (ii) aprðAÞ ¼ aprðAÞ ¼ A; for A 2 sðU=EÞ; (iii) aprðAÞ � A � aprðAÞ; (iv) aprðA \ BÞ ¼ aprðAÞ \ aprðBÞ; aprðA [ BÞ ¼ aprðAÞ [ aprðBÞ; (v) aprðA [ BÞ aprðAÞ [ aprðBÞ; aprðA \ BÞ � aprðAÞ \ aprðBÞ: Property (i) shows that lower and upper approximations are dual to each other. Property (ii) indicates that the lower and upper approximations of a definable set are the set itself. By property (iii), a set lies within its lower and upper approximations. Property (iv) states that the lower appro- ximation distributes over intersection, and the upper approximation distributes over union. Property (v) shows the sub-distributivity of approximation operators. Many probability-related measures on approximations have been proposed and studied. Pawlak (1982, 1991) suggested an accuracy measure of rough set approximation given by aðAÞ ¼ japrðAÞj japrðAÞj ¼ PðaprðAÞjaprðAÞÞ ð10Þ It may be interpreted as the probability that an element belongs to the lower approximation, given that the element belongs to the upper approximation. This measure can also be expressed in terms of the well-known Marczewski– Steinhaus metric (Yao, 2001a). Measures of quality of lower and upper approximations are given respectively by Pawlak (1991): qðAÞ ¼ japrðAÞj jUj ¼ PðaprðAÞÞ qðAÞ ¼ japrðAÞj jUj ¼ PðaprðAÞÞ ð11Þ They are referred to as rough probability by Pawlak (1984) and have been used by many authors (Grzymala-Busse, 1987; Wong & Lingras, 1989; Yao & Lingras, 1998; Düntsch & Gediga, 2001). The relationship between the accuracy and quality of approximations can be expres- sed as aðAÞ ¼ qðAÞ qðAÞ ð12Þ The accuracy measure can be re-expressed as aðAÞ ¼ japrðAÞj japrðAÞj ¼ japrðAÞj jUj � japrð:AÞj ð13Þ Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 289 which suggests that the accuracy measure also depends on the lower approximation of :A. Based on this observation, Gediga and Düntsch (2001) suggested use of the function gðAÞ ¼ japrðAÞj jAj ¼ PðaprðAÞjAÞ ð14Þ as a measure of the precision of deterministic approxima- tion of A. Using the same argument, we suggest that the quality of non-deterministic approximation of A can be measured by gðAÞ ¼ jAj japrðAÞj ¼ PðAjaprðAÞÞ ð15Þ In this case, we have aðAÞ ¼ gðAÞgðAÞ ð16Þ The two measures q and g monotonically increase as aprðAÞ approaches A when different partitions are used. On the other hand, q and g show the opposite direction of changes. For two partitions p1 and p2 with p1 � p2, we have (vi) apr p1 ðAÞ apr p2 ðAÞ; aprp1ðAÞ � aprp2ðAÞ: By combining these with (iii), we obtain apr p2 ðAÞ � apr p1 ðAÞ � A � aprp1ðAÞ � aprp2ðAÞ ð17Þ As expected, a finer partition induces a tighter approxima- tion. All of the measures correctly reflect this observation, as shown by the following properties: (1) ap1ðAÞ � ap2ðAÞ; (2) q p1 ðAÞ � q p2 ðAÞ; qp1ðAÞ � qp2ðAÞ; (3) g p1 ðAÞ � g p2 ðAÞ; gp1ðAÞ � gp2ðAÞ: That is, for two partitions with p1 � p2, we obtain the same qualitative evaluation by all those measures, namely, p1 is the same or better than p2. For an arbitrary pair of partitions, the pair of q and g, or the pair of q and g, produce the same qualitative evaluation, which may be different from the one given by the a accuracy measure (Gediga & Düntsch, 2001). The approximation of a subset can be easily extended to the approximation of a family of subsets. Consider two partitions pA ¼ {A1, A2, . . . , An} and pB ¼ {B1, B2, . . . , Bm} We construct an approximation space aprpA ¼ ðU; EpAÞ using the partition pA. Each equivalence class of pB is approximated by aprpA ðBiÞ and aprpAðBiÞ. By extending the accuracy and quality measure to the approximation of partition, Pawlak (1991) suggested the following quantities: apAðpBÞ ¼ �mi ¼ 1j aprpAðBiÞj �mi ¼ 1j aprpAðBiÞj g pA ðpBÞ ¼ �mi ¼ 1j aprpAðBiÞj jUj ð18Þ Furthermore, g pA can be expressed as, respectively, the expectation, a weighted sum, and the sum of g, a and q on individual equivalence classes (Gediga & Düntsch, 2001): g pA ðpBÞ ¼ Xm i ¼ 1 jBij jUj g pA ðBiÞ ¼ Xm i ¼ 1 PðBiÞ gpAðBiÞ g pA ðpBÞ ¼ Xm i ¼ 1 j aprpAðBiÞj jUj apAðBiÞ ð19Þ ¼ Xm i ¼ 1 PðaprpAðBiÞÞapAðBiÞ g pA ðpBÞ ¼ Xm i ¼ 1 q pA ðBiÞ The overall measure apAðpBÞ cannot be similarly expressed. It is reasonable to use as an alternative overall measure a0pAðpBÞ ¼ Xm i ¼ 1 apAðBiÞ ð20Þ In fact, apAðpBÞ and a0pAðpBÞ represent two different averaging methods: one is the application of a measure to the pooled results, and the other is the average of the measurements on the individual results. Given a subset B � U, we can partition the universe as {B, :B}. By applying the partition based measure g pA , Düntsch and Gediga (2001) suggested the following measure of the approximation quality of pA with respect to B: g0 pA ðBÞ ¼ g pA ðpBÞ ¼ qpAðBÞ þ qpAð:BÞ ð21Þ This measure is the ratio of correct classification of either B or :B based on the partition pA. 3.2. Approximations of probabilities In an approximation space apr ¼ (U, E), suppose a set function is defined on s(U = E). One can extend the func- tion to nondefinable subsets through the lower and upper approximations. Many authors have studied the approxi- mation of probabilities in the framework of rough sets, which leads to belief functions (Pawlak, 1984; Grzymala- Busse, 1987; Skowron, 1989, 1990; Wong & Lingras, 1989; Skowron & Grzymala-Busse, 1994; Yao & Lingras, 1998). A belief function is a mapping from 2 U to the unit interval [0, 1] and satisfies the following axioms. (F1) Bel(+) ¼ 0. (F2) Bel(U) ¼ 1. (F3) For every positive integer n and every collection A1, . . . , An � U, BelðA1 [ A2 . . . [ AnÞ � X i BelðAiÞ � X i < j BelðAi \ AjÞ . . . þ ð�1Þnþ1BelðA1 \ . . . \ AnÞ 290 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 Axioms (F1) and (F2) may be considered as normaliza- tion conditions. Axiom (F3) is a weaker version of the commonly known additivity axiom of probability func- tions. It is referred to as the axiom of superadditivity. The dual of a belief function, called a plausibility function Pl, is defined by PlðAÞ ¼ 1 � Belð:AÞ ð22Þ For any subset A � U, Bel(A) � Pl(A). Consider first a simple case where a probability function P on 2 U is defined based on the counting of elements in a set (Grzymala-Busse, 1987; Skowron & Grzymala-Busse, 1994); namely, for A � U, PðAÞ ¼ jAj jUj ð23Þ Clearly, we have qðAÞ ¼ PðaprðAÞÞ � PðAÞ � PðaprðAÞÞ ¼ qðAÞ ð24Þ While aprðAÞ and aprðAÞ are approximations of the set A, qðAÞ and qðAÞ are the approximations of the probability of the set A. It can easily be verified by using properties (i)–(iv) that the qualities of lower and upper approximations are a pair of belief and plausibility functions. Suppose P is a probability function defined on sðU=EÞ. It is not defined for subsets of U which are not members of sðU=EÞ. One can extend P to 2U in two standard ways by defining functions Pn and P n , traditionally called the inner measure and the outer measure induced by P. For an arbitrary subset A � U, we define PnðAÞ ¼ supfPðXÞjX 2 sðU=EÞ; X � Ag ¼ PðaprðAÞÞ PnðAÞ ¼ inffPðXÞjX 2 sðU=EÞ; X Ag ¼ PðaprðAÞÞ ð25Þ Pawlak (1984) referred to the pair (Pn(A), P n (A)) as the rough probability of A. The inner and outer probabilities Pn and P n are a pair of belief and plausibility functions (Wong & Lingras, 1989; Fagin & Halpern, 1991; Yao & Lingras, 1998). 3.3. Probabilistic rough set approximations Algebraic rough set approximations may be considered as qualitative approximations of a set. The extent of over- lap between a set and an equivalence class is not considered. By incorporating the overlap, many authors have intro- duced and studied probabilistic rough set approximations (Wong & Ziarko, 1987; Pawlak et al., 1988; Ziarko, 1993; Pawlak & Skowron, 1994). Most proposals introduced cer- tain parameters based on intuitive arguments. The decision- theoretic rough set model provides a solid basis for probabilistic approximations (Yao et al., 1990; Yao & Wong, 1992). We first briefly review the Bayesian decision-theoretic framework (Duda & Hart, 1973). Let O ¼ {o1, . . . , os} be a finite set of s states, and let A ¼ {a1, . . . , am} be a finite set of m possible actions. Let P(ojjx) be the conditional probability of an object x being in state oj given that the object is described by x. Let l(aijoj) denote the loss, or cost, for taking action ai when the state is oj. For an object with description x, suppose an action ai is taken. Since P(ojjx) is the probability that the true state is oj given x, the expected loss associated with taking action ai is given by RðaijxÞ ¼ Xs j ¼ 1 lðaijojÞPðojjxÞ ð26Þ The quantity R(ai|x) is called the conditional risk. Given a description x, a decision procedure is a function t(x) that specifies which action to take. For every x, t(x) chooses one action from a1, . . . , am. The overall risk R is the expected loss associated with a given decision procedure. Since R(t(x)|x) is the conditional risk associated with action t(x), the overall risk is defined by R ¼ X x RðtðxÞjxÞPðxÞ ð27Þ where the summation is over the set of all possible descrip- tions of objects. One can obtain an optimal decision procedure by minimizing the overall risk. If t(x) is chosen so that R(t(x)jx) is as small as possible for every x, the overall risk R is minimized. The Bayesian decision procedure can therefore be formally stated as follows. For every x, compute the conditional risk R(aijx) for i ¼ 1, . . . , m defined by equation (26), and then select the action for which the conditional risk is the minimum. If more than one action minimizes R(aijx), any tie-breaking rule can be used. The Bayesian decision procedure can be applied to define probabilistic rough set approximations. Given a subset A � U, we can form a set of two states O ¼ {A, :A} indicating that an element is in A and not in A, respectively. We use the same symbol to denote both a subset A and the corresponding state. In the non-probabilistic rough set model, with respect to A, we divide the universe U into three disjoint regions, the positive region POS(A), the negative region NEG(A) and the boundary region BND(A): POSðAÞ ¼ aprðAÞ NEGðAÞ ¼ U � aprðAÞ BNDðAÞ ¼ aprðAÞ � aprðAÞ ð28Þ In developing a probabilistic rough set model, with respect to three regions, the set of actions is given by A ¼ {a1, a2, a3}, where a1, a2 and a3 represent the three actions in classifying an object, deciding POS(A), deciding NEG(A) and deciding BND(A), respectively. The symbol [x]E, the equivalence class containing x, is also used to represent a description of x. The required conditional probabilities are defined by the rough membership functions (Pawlak & Skowron, 1994) mAðxÞ ¼ j½x�E \ Aj j½x�Ej ¼ PðAj½x�EÞ ð29Þ Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 291 Let l(aijA) denote the loss incurred for taking action ai when an object in fact belongs to A, and let l(aij:A) denote the loss incurred for taking the same action when the ob- ject does not belong to A. The expected loss R(aij[x]E) associated with taking the individual actions can be expressed as Rða1j½x�EÞ ¼ l11PðAj½x�EÞ þ l12Pð:Aj½x�EÞ Rða2j½x�EÞ ¼ l21PðAj½x�EÞ þ l22Pð:Aj½x�EÞ Rða3j½x�EÞ ¼ l31PðAj½x�EÞ þ l32Pð:Aj½x�EÞ ð30Þ where li1 ¼ l(ai|A), li2 ¼ l(ai|:A) and i ¼ 1, 2, 3. The Baye- sian decision procedure leads to the following minimum- risk decision rules. (P) If R(a1j[x]E) � R(a2j[x]E) and R(a1j[x]E) � R(a3j[x]E), decide POS(A). (N) If R(a2j[x]E) � R(a1j[x]E) and R(a2j[x]E) � R(a3j[x]E), decide NEG(A). (B) If R(a3j[x]E) � R(a1j[x]E) and R(a3j[x]E) � R(a2j[x]E), decide BND(A). Tie-breaking rules should be added so that each element is classified into only one region. Since P(Aj[x]E) þ P(:Aj[x]E) ¼ 1, the above decision rules can be simplified such that only the probabilities P(Aj[x]E) are involved. We can classify any object in the equivalence class [x]E based only on the probabilities P(Aj[x]E), i.e. the rough mem- bership values, and the given loss function lij (i ¼ 1, 2, 3; j ¼ 1, 2). Consider a special kind of loss function with l11 � l31 < l21 and l22 � l32 < l12. That is, the loss of classifying an object x belonging to A into the positive region POS(A) is less than or equal to the loss of classifying x into the boundary region BND(A), and both of these losses are strictly less than the loss of classifying x into the negative region NEG(A). The reverse order of losses is used for classifying an object that does not belong to A. For this type of loss function, the minimum-risk decision rules (P), (N), (B) can be written as (P) if P(Aj[x]E) � g and P(Aj[x]E) � a, decide POS(A); (N) if P(Aj[x]E) � b and P(Aj[x]E) � g, decide NEG(A); (B) if b � P(Aj[x]E) � a, decide BND(A); where a ¼ l12 � l32 ðl31 � l32Þ � ðl11 � l12Þ g ¼ l12 � l22 ðl21 � l22Þ � ðl11 � l12Þ b ¼ l32 � l22 ðl21 � l22Þ � ðl31 � l32Þ ð31Þ By the assumptions l11 � l31 < l21 and l22 � l32 < l12, it follows that a 2 (0, 1], g 2 (0, 1) and b 2 [0, 1). A loss function should be chosen in such a way as to satisfy the condition a � b. This ensures that the results are consistent with rough set approximations. Namely, the lower approximation is a subset of the upper approxima- tion, and the boundary region may be nonempty. When a>b, we have a>g>b. After tie-breaking, we obtain the decision rules (P1) if P(Aj[x]E) � a, decide POS(A); (N1) if P(Aj[x]E) � b, decide NEG(A); (B1) if b < P(Aj[x]E) < a, decide BND(A). When a ¼ b, we have a ¼ g ¼ b. In this case, we use the decision rules (P2) if P(Aj[x]E)>a, decide POS(A); (N2) if P(Aj[x]E) < a, decide NEG(A); (B2) if P(Aj[x]E) ¼ a, decide BND(A). For the second set of decision rules, we use a tie-breaking criterion so that the boundary region may be nonempty. The standard and other probabilistic rough set models can be easily derived by choosing different loss functions. Consider the loss function l12 ¼ l21 ¼ 1 l11 ¼ l22 ¼ l31 ¼ l32 ¼ 0 ð32Þ There is a unit cost if an object belonging to A is classified into the negative region or if an object not belonging to A is classified into the positive region; otherwise there is no cost. In this case, we have a ¼ 1 > b ¼ 0, a ¼ 1 � b and g ¼ 0.5. According to decision rules (P1), (N1), (B1), we obtain the standard rough set approximations (Pawlak, 1982, 1991). Another loss function is given by l12 ¼ l21 ¼ 1 l31 ¼ l32 ¼ 0:5 l11 ¼ l22 ¼ 0 ð33Þ A unit cost is incurred if the system classifies an object belonging to A into the negative region or an object not belonging to A into the positive region; half of a unit cost is incurred if any object is classified into the boundary region. There is no cost for other cases. It follows that a ¼ b ¼ g ¼ 0.5. By using decision rules (P2), (N2), (B2), we obtain the probabilistic rough set approximation proposed by Pawlak et al. (1988). Suppose a loss function with l11 � l31 < l21 and l22 � l32 < l12 satisfies the conditions l12 � l32 � l31 � l11 ðl12 � l32Þðl32 � l22Þ ¼ ðl31 � l11Þðl21 � l31Þ ð34Þ We have a ¼ 1 � b � 0.5. This leads to the variable preci- sion rough set model (Ziarko, 1993). 4. Probabilistic measures for rule induction An important application of rough sets is data analysis and rule induction (Wong & Ziarko, 1986; Pawlak, 1991; Grzymala-Busse, 1992; Tsumoto, 1998). This section reviews probabilistic and information-theoretic measures used in rule induction algorithms (Yao & Zhong, 1999; Yao, 2003b). 292 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 4.1. Information tables An information table provides a convenient way to des- cribe a finite set of objects by a finite set of attributes (Pawlak, 1991). In this paper, we use an extended information table by adding binary relations on attribute values, and two languages (Yao, 2001b). Formally, an information table can be expressed as S ¼ ðU; At; Lv; Lr; fVaja 2 Atg; fRaja 2 Atg; fIaja 2 AtgÞ where U is a finite nonempty set of objects, At is a finite nonempty set of attributes, Lv is a language dealing with values of objects, Lr is a language dealing with relations of objects, Va is a nonempty set of values for a 2 At, Ra is a nonempty set of binary relations on Va, a 2 At, and Ia: U- Va is an information function. Each information function Ia is a total function that maps an object of U to exactly one value in Va. An information table represents all available information and knowledge. That is, objects are only perceived, observed or measured by using a finite number of properties. In the language Lv, anatomic formula is givenby (a, R, v), where a 2 At, R 2 Ra and v 2 Va. If f and c are formulae, then so are :f, f ^ c, f _ c, f - c and f � c. The semantics of the language Lv can be defined in the Tarski style through the notions of a model and satisfiability. The model is an information table S, which provides an interpretation for symbols and formulae of Lv. The satisfiability of a formula f by an object x, written x �s f or in short x � f if S is understood, is given by the following conditions: (m1) x � (a, R, v) iff Ia(x) Rv, (m2) x � :f iff not x � f, (m3) x � f 4c iff x � f and x � c, (m4) x � f 3c iff x � f or x � c, (m5) x � f-c iff x � :f3c, (m6) x � f � c iff x � f-c and x � c-f. If f is a formula, the set ms(f) defined by mSðfÞ ¼ fx 2 Ujx � fg ð35Þ is called the meaning of the formula f in S. If S is understood, we simply write m(f). The following proper- ties hold: (a) m(a, R, v) ¼ {x 2 U|Ia(x) R v}, (b) m(:f) ¼ :m(f), (c) m(f4c) ¼ m(f) \ m(c), (d) m(f3c) ¼ m(f) [ m(c), (e) m(f-c) ¼ :m(f) [ m(c), (f ) m(f � c) ¼ (m(f) \ m(c)) [ (:m(f) \ :m(c)). The meaning of a formula f is therefore the set of all objects having the property expressed by the formula f. In other words, f can be viewed as the description of the set of objects m(f). Thus, a connection between formulae of Lv and subsets of U is established. When the relation R is chosen to be the equality relation ¼ , we obtain the conventional decision logic language (Pawlak, 1991). With the introduction of language Lv, we have a formal description of concepts. A concept definable in an infor- mation table is a pair (f, m(f)), where f 2 Lv. More specifically, f is a description of m(f) in S, the intension of concept (f, m(f)), and m(f) is the set of objects satisfying f, the extension of concept (f, m(f)). A concept (f, m(f)) is said to be a sub-concept of another concept (c, m(c)), or (c, m(c)) a super-concept of (f, m(f)), if m(f) � m(c). A concept (f, m(f)) is said to be the smallest nonempty concept in S if there does not exist another proper nonempty sub-concept of (f, m(f)). Two concepts (f, m(f)) and (c, m(c)) are disjoint if m(f) \ m(c) ¼ |. If m(f) \ m(c) 6¼ |, we say that the two concepts have a nonempty overlap and hence are related. The language Lr is defined in a similar manner to Lv, except that an atomic formula is given by (a, R), where R 2 Ra and a 2 At. Semantics of formulae of Lr are inter- preted by pairs of objects in U. That is, (m10) (x, y) � (a, R) iff Ia(x) R Ia(y). For formula f, the set ms(f) defined by mSðfÞ ¼ fðx; yÞ 2 U � Ujðx; yÞ � fg ð36Þ is called the meaning set of f in S. If S is understood, we simply write m(f). A pair (x, y) 2 m(f) is said to satisfy the expression f. Similarly, the formula f can be viewed as the description of the set of object pairs m(f), and each object pair in m(f) as an instance of the concept given by f. 4.2. Two types of rules Knowledge derivable from an information table is com- monly represented in the form of rules. Roughly speaking, rules show the connections between attributes, which are normally characterized by the problem of determining the values of one set of attributes based on the values of another set of attributes. Depending on the meanings and forms of rules, we can classify rules in many ways. A clear classification of rules is useful for an understanding of the basic tasks of machine learning and data mining. Rules can be classified into two groups in terms of their directions, one-way and two-way connections, and further classified into two levels in terms of their applicability, local and global connections (Yao & Zhong, 1999; Yao, 2001b, 2003b). A one-way connection shows that the values of one set of attributes determine the values of another set of attributes, but does not say anything about the reverse. A two-way connection is a combination of two one- way connections, representing two different directions of connection. A local connection is characterized by a rule showing the relationship between one specific combination Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 293 of values on one set of attributes and one specific com- bination of values on another set of attributes. A global connection is characterized by a rule showing the rela- tionships between all combinations of values on one set of attributes and all combinations of values on another set of attributes. For clarity, we only consider one-way connections and the equality relation on attribute values, as was commonly done in rough sets. In this case, a local one-way connection is expressed by a rule of the form, using formulae of Lv, ða; ¼ ; vaÞ ) ðb; ¼ ; vbÞ ð37Þ where a, b 2 At, and va 2 Va, vb 2 Vb. It can be more conveniently expressed as, for x 2 U, IaðxÞ ¼ va ) IbðxÞ ¼ vb ð38Þ The rule is commonly paraphrased as ‘if the value of an object is va on an attribute a, then its value is vb on another attribute b’. A global one-way connection is expressed by a rule of the form, using formulae of Lr, ða; ¼ Þ ) ðb; ¼ Þ ð39Þ where a, b 2 At, or conveniently as, for (x, y) 2 U � U, IaðxÞ ¼ IaðyÞ ) IbðxÞ ¼ IbðyÞ ð40Þ That is, ‘if two objects have the same value on an attribute a, then they have the same value on another attribute b’. Functional dependence in a database is an example of such global rules. The formulation of rules using atomic formulae can be easily extended to any formulae of languages Lv and Lr. A local rule states knowledge about one object. A local one- way rule shows that, if the object has a specific value on one set of attributes, then it will have a specific value on another set of attributes. On the other hand, a global rule states knowledge about a pair of objects. A global one-way rule suggests that, if a pair of objects have the same value on one set of attributes, then they will have the same value on another set of attributes. Based on this observation, a global rule is also called a high order rule, while a local rule is called a low order rule (Yao, 2003a). 4.3. Interpretation of rough set theory in information tables The abstract theory of rough sets can be explained by using an information table. Such an interpretation is useful for rule induction. Let W ¼ {W1, . . . , Wn} � At be a set of attributes in an information table. We form a family of elementary formulae FW ¼ f^ni ¼ 1ðWi; ¼ ; wiÞjwi 2 VWig of the lan- guage Lv. For simplicity, let VW ¼ VW1 � . . . � VWn. We also express the family of elementary formulae by FW ¼ {W¼ wjw2 VW}. The family of nonempty meaning sets form a partition of the universe, namely pW ¼ fmð^ni ¼ 1ðWi; ¼ ; wiÞÞ 6¼ |jwi 2 VWig ð41Þ It is referred to as the partition induced by the set of attributes W. For the same set of attributes, we can construct a formula ^ni ¼ 1ðWi; ¼ Þ of the language Lr. The meaning of the formula EW ¼ mð^ni ¼ 1ðWi; ¼ ÞÞ ¼ fðx; yÞ 2 U � UjIWi ðxÞ ¼ IWiðyÞ; 1 � i � ng ð42Þ is an equivalence relation on U. Similarly, another set of attributes Z ¼ {Z1, . . . , Zm} defines another partition pZ and the corresponding equivalence relation EZ. Rough set approximations of a single subset are relevant to the induction of local or low order rules. Consider a formula ^mj ¼ 1ðZj; ¼ ; zjÞ. In terms of attributes in W, we can obtain various local rules of the following format: ^ni ¼ 1ðWi; ¼ ; wiÞ ) ^ m j ¼ 1ðZj; ¼ ; zjÞ ð43Þ Let fW ¼ ^ni ¼ 1 ðWi; ¼ ; wiÞ and cZ ¼ ^mj ¼ 1 ðZj; ¼ ; zjÞ. With respect to the three regions of rough set approxima- tions, we can construct three classes of rules. (I) Positive region: certain positive rules m(fW) � m(cZ), fW-cZ. (II) Boundary region: uncertain positive rules m(fW) 6� m(cZ) and m(fW) \ m(cZ) 6¼ +2, fW ) cZ. (III) Negative region: certain negative rules m(fW) \ m(cZ) ¼ +, fW - :cZ. Certain rules can be considered as the degenerate cases of uncertain rules. Since certain rules can be interpreted using the logical connective -, we use the same symbol. All three classes of rules express the relationship between two concepts in terms of their meaning sets. Probabilistic measures introduced earlier can be used to quantify the uncertainty of rules. For example, a measure from the rough membership function jmðfWÞ \ mðcZÞj jmðfWÞj ð44Þ can be used to measure the accuracy of one rule. Other measures associated with approximation can be used to show the characteristic of a set of rules. For instance, the measure suggested by Gediga and Düntsch (2001) g pW ðmðcZÞÞ ¼ j apr pW ðmðcZÞÞj jmðcZÞj ¼ P fW 2 FW ;mðfW Þ � mðcZÞ jmðfWÞj jmðcZÞj ð45Þ is the ratio of objects correctly classified by all certain positive rules to the objects satisfying the condition cZ. Similarly, the accuracy of approximation apW ðmðcZÞÞ is the ratio of objects correctly classified by all certain positive 294 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 rules to the objects classified by both certain and uncertain positive rules. Approximation of one partition based on another parti- tion can be summarized by the following rule: ^ni ¼ 1ðWi; ¼ Þ ) ^ m j ¼ 1ðZj; ¼ Þ or simply W ) Z ð46Þ The measures apW ðpZÞ and gpW ðpZÞ can be used to measure the strength of the global, high order rule. A more detailed probabilistic and information-theoretic analysis of low and high order rules is given in the following sections (Yao & Zhong, 1999; Yao, 2003b). 4.4. Probabilistic measures for low order rules Suppose f and c are two formulae of the language Lv. For a rule f ) c, its characteristics can be summarized by the following contingency table: � c :c Total f a b a þ b :f c d c þ d Total a þ c b þ d a þ b þ c þ d ¼ jUj a ¼ jmðf ^ cÞj b ¼ jmðf ^ :cÞj c ¼ jmð:f ^ cÞj d ¼ jmð:f ^ :cÞj Different measures can be defined to reflect various aspects of rules. The generality of f is defined by GðfÞ ¼ jmðfÞj jUj ¼ a þ b jUj ð47Þ which indicates the relative size of the concept f. Obvi- ously, we have 0 � G(f) � 1. A concept is more general if it covers more instances of the universe. A sub-concept has a lower generality than its super-concept. The quantity may be viewed as the probability of a randomly selected element satisfying f. The absolute support of c provided by f is the quantity ASðf ) cÞ ¼ ASðcjfÞ ¼ jmðcÞ \ mðfÞj jmðfÞj ¼ a a þ b ð48Þ The quantity 0 � AS(c|f) � 1 states the degree to which f supports c. It may be viewed as the conditional probability of a randomly selected element satisfying c given that the element satisfies f. In set-theoretic terms, it is the degree to which m(f) is included in m(c). Clearly, AS(c|f) ¼ 1 if and only if m(f) 6¼ | and m(f) � m(c). That is, a rule with the maximum absolute support 1 is a certain rule. The change of support of c provided by f is defined by CSðf ) cÞ ¼ CSðcjfÞ ¼ ASðcjfÞ � GðcÞ ¼ a a þ b � a þ c jUj ð49Þ Unlike the absolute support, the change of support varies from � 1 to 1. We can consider G(c) to be the prior probability of c and AS(cjf) the posterior probability of c after knowing f. The difference of posterior and prior probabilities represents the change in our confidence regard- ing whether f is actually related to c. For a positive value, we may say that f is positively related to c; for a negative value, we may say that f is negatively related to c. The change of support relative to c is given by RCSðc ) cÞ ¼ CSðcjfÞ GðcÞ ¼ ASðcjfÞ GðcÞ � 1 ¼ Gðc ^ fÞ GðcÞGðfÞ � 1 ¼ jUjjmðcÞ \ mðfÞj jmðcÞjjmðfÞj � 1 ¼ ajUj ða þ cÞða þ bÞ � 1 ð50Þ It is interesting to note that the first term in the relative change of support is related to the probabilistic indepen- dence of c and f. The generality G(c) is related to the satisfiability of c by all objects in the database, and AS(f ) c) is related to the satisfiability of c in the subset m(f). A high AS(f ) c) does not necessarily suggest a strong association between f and c, as a concept c with a large G(c) value tends to have a large AS(f ) c) value. The change of support CS(f ) c), or the relative change of support RCS(f ) c), may be more accurate. 4.5. Information-theoretic measures for high order rules Recall that a set of attributes W induces a partition pW of the universe. Let PðwÞ ¼ PðmðW ¼ wÞÞ ¼ jmðW ¼ wÞj jUj ð51Þ Shannon’s entropy function of pW, simply written as H(P(W)), is given by HðPðWÞÞ ¼ EPðWÞ½�logPðWÞ� ¼ � X w 2 VW PðwÞ logPðwÞ ð52Þ where EP(W)[ � ] denotes the expected value with respect to the probability distribution of W. For two sets of attributes Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 295 W and Z, their joint entropy is defined by HðZ; WÞ ¼ � X z 2 VZ X w 2 VW pðz; wÞ log pðz; wÞ ð53Þ The conditional entropy H(Z|W) is defined as the expected value of subpopulation entropies H(Z|w) with respect to the probability distribution P(W): HðZjWÞ ¼ X w 2 VW PðwÞHðZjwÞ ¼ � X w 2 VW PðwÞ X z 2 VZ PðzjwÞ logPðzjwÞ ¼ � X z 2 VZ X w 2 VW Pðz; wÞ logPðzjwÞ ¼ EPðZ;WÞ½� log PðZjWÞ� ð54Þ Conditional entropy is nonnegative and nonsymmetric, namely H(ZjW)� 0 and in general H(ZjW) 6¼ H(WjZ). Con- ditional entropy can also be expressed by HðZjWÞ ¼ HðZ; WÞ � HðWÞ ð55Þ It measures the additional amount of information provided by Z if W is already known. The probability P(z) is the generality of the granule m(Z ¼ z). The function �log P(z) is a monotonic decreasing transformation of P(z). As the expected values of �log P(z), the entropy function is related to the granularity of the partition pZ. The probability P(zjw) is a measure for the local rule W ¼ w ) Z ¼ z. As an expected value, the conditional entropy H(Z|W) provides a measure for the global rule W ) Z. It may be viewed as an inverse measure of global one-way association of two sets of attributes (Pawlak et al., 1988): IC1ðW ) ZÞ ¼ HðZjWÞ ð56Þ A normalized version is given by (Pawlak et al., 1988) IC2ðW ) ZÞ ¼ HðZjWÞ logjVZj ð57Þ For an attribute Z, conditional entropy can be used to select important attributes for discovering a one-way asso- ciation W ) Z. Measures IC1 and IC2 can be used to rank attributes in an increasing order. If one prefers to rank attributes in a decreasing order, the following correspond- ing direct measures of one-way association can be used: C1ðW ) ZÞ ¼ logjVZj � HðZjWÞ ð58Þ C2ðW ) ZÞ ¼ 1 � HðZjWÞ logjVZj ð59Þ In these measures, the attribute entropy H(Z) may be used in place of log jVZj. We obtain the following measures: C3ðW ) ZÞ ¼ HðZÞ � HðZjWÞ ¼ IðZ; WÞ ð60Þ C4ðW ) ZÞ ¼ 1 � HðZjWÞ HðZÞ ¼ IðZ; WÞ HðZÞ ð61Þ Measure C3 is in fact the mutual information between W and Z. It is commonly referred to as information gain and is widely used in machine learning (Quinlan, 1986). Like the change of support for local rules, C3 may be viewed as changes of entropy for global rules. Similarly, C4 may be viewed as a relative change of entropy for global rules. 5. Conclusion While nonprobabilistic studies of rough sets focus on algebraic and qualitative properties of the theory, prob- abilistic approaches are more practical and capture quan- titative properties of the theory. The granularity of a partition can be quantified by information-theoretic measures. Existing measures of accuracy and quality of approximations can be quantified by probability-related measures. The probabilistic and information-theoretic approaches are particularly useful in rule induction, an important application of rough set theory. Most of the measures discussed in this paper are based on simple counting of the number of elements of a set. Furthermore, we have restricted our discussion to granula- tions by equivalence relations or partitions. It should be pointed out that the argument can be easily extended to more general probability functions and general granulation structures. References DUDA, R.O. and P.E. HART (1973) Pattern Classification and Scene Analysis, New York: Wiley. DÜNTSCH, I. and G.R. GEDIGA (2001) Rough information analysis, International Journal of Intelligent Systems, 16, 121–147. FAGIN, R. and J.Y. HALPERN (1991) Uncertainty, belief, and probability, Computational Intelligence, 7, 160–173. GEDIGA, G. and I. DÜNTSCH (2001) Rough approximation quality revisited, Artificial Intelligence, 132, 219–234. GRZYMALA-BUSSE, J.W. (1987) Rough-set and Dempster–Shafer approaches to knowledge acquisition under uncertainty— a comparison, Manuscript, Department of Computer Science, University of Kansas. GRZYMALA-BUSSE, J.W. (1992) LERS—a system for learning from example based on rough sets, in Intelligent Support: Handbook of Applications and Advances of the Rough Set Theory, R. Slowinski (ed.), Dordrecht: Kluwer Academic, 3–18. KLIR, G.J. and T.A. GOLGER (1988) Fuzzy Sets, Uncertainty, and Information, Englewood Cliffs, NJ: Prentice Hall. LEE, T.T. (1987) An information-theoretic analysis of relational databases—part I: data dependencies and information metric, IEEE Transactions on Software Engineering, SE-13, 1049–1061. LIN, T.Y., Y.Y. YAO and L.A. ZADEH (eds) (2002) Data Mining, Rough Sets and Granular Computing, Heidelberg: Physica-Verlag. PAWLAK, Z. (1982) Rough sets, International Journal of Computer and Information Sciences, 11, 341–356. PAWLAK, Z. (1984) Rough probability, Bulletin of Polish Academy of Sciences, Mathematics, 32, 607–615. 296 ________________________________________________________________ Expert Systems, November 2003, Vol. 20, No. 5 PAWLAK, Z. (1991) Rough Sets, Theoretical Aspects of Reasoning about Data, Dordrecht: Kluwer Academic. PAWLAK, Z. and A. SKOWRON (1994) Rough membership func- tions, in Advances in the Dempster–Shafer Theory of Evidence, R.R. Yager, M. Fedrizzi and J. Kacprzyk (eds), New York: Wiley, 251–271. PAWLAK, Z., S.K.M. WONG and W. ZIARKO (1988) Rough sets: probabilistic versus deterministic approach, International Journal of Man–Machine Studies, 29, 81–95. QUINLAN, J.R. (1986) Induction of decision trees, Machine Learning, 1, 81–106. SKOWRON, A. (1989) The relationship between the rough set theory and evidence theory, Bulletin of Polish Academy of Sciences, Mathematics, 37, 87–90. SKOWRON, A. (1990) The rough set theory and evidence theory, Fundamenta Informaticae, 13, 245–262. SKOWRON, A. and J. GRZYMALA-BUSSE (1994) From rough set theory to evidence theory, in Advances in the Dempster–Shafer Theory of Evidence, R.R. Yager, M. Fedrizzi and J. Kacprzyk (eds), New York: Wiley, 193–236. TSUMOTO, S. (1998) Automated extraction of medical expert system rules from clinical databases on rough set theory, Information Sciences, 112, 67–84. WONG, S.K.M. and P.J. LINGRAS (1989) The compatibility view of Shafer–Dempster theory using the concept of rough set, Methodologies of Intelligent Systems, 4, 33–42. WONG, S.K.M. and W. ZIARKO (1986) Algorithm for inductive learning, Bulletin of the Polish Academy of Sciences, Technical Sciences, 34, 271–276. WONG, S.K.M. and W. ZIARKO (1987) Comparison of the probabilistic approximate classification and the fuzzy set model, Fuzzy Sets and Systems, 21, 357–362. YAO, J.T. and Y.Y. YAO (2002) Induction of classification rules by granular computing, Proceedings of the Third International Conference on Rough Sets and Current Trends in Computing, LNAI 2475, Berlin: Springer, 331–338. YAO, Y.Y. (1996) Two views of the theory of rough sets in finite universes, International Journal of Approximation Reasoning, 15, 291–317. YAO, Y.Y. (2000) Granular computing: basic issues and possible solutions, Proceedings of the 5th Joint Conference on Information Sciences, Association for Intelligent Machinery, 186–189. YAO, Y.Y. (2001a) Information granulation and rough set approximation, International Journal of Intelligent Systems, 16, 87–104. YAO, Y.Y. (2001b) Modeling data mining with granular computing, Proceedings of the 25th Annual International Computer Software and Applications Conference, 638–643. YAO, Y.Y. (2003a) Mining high order decision rules, in Rough Set Theory and Granular Computing, M. Inuiguchi, S. Hirano and S. Tsumoto (eds), Berlin: Springer, 125–135. YAO, Y.Y. (2003b) Information-theoretic measures for knowledge discovery and data mining, in Entropy Measures, Maximum Entropy and Emerging Applications, Karmeshu (ed.), Berlin: Springer, 115–136. YAO, Y.Y. and P.J. LINGRAS (1998) Interpretations of belief functions in the theory of rough sets, Information Sciences, 104, 81–106. YAO, Y.Y. and S.K.M. WONG (1992) A decision theoretic framework for approximating concepts, International Journal of Man–Machine Studies, 37, 793–809. YAO, Y.Y. and N. ZHONG (1999) An analysis of quantitative measures associated with rules, Proceedings of the Third Pacific- Asia Conference on Knowledge Discovery and Data Mining, LNAI 1574, Berlin: Springer, 479–488. YAO, Y.Y., S.K.M. WONG and P.J. LINGRAS (1990) A decision- theoretic rough set model, in Methodologies for Intelligent Systems, Vol. 5, Z.W. Ras, M. Zemankova and M.L. Emrich (eds), New York: North-Holland, 17–24. ZADEH, L.A. (1979) Fuzzy sets and information granularity, in Advances in Fuzzy Set Theory and Applications, N. Gupta, R. Ragade and R. Yager (eds), Amsterdam: North-Holland, 3–18. ZADEH, L.A. (1997) Towards a theory of fuzzy information granulation and its centrality in human reasoning and fuzzy logic, Fuzzy Sets and Systems, 19, 111–127. ZHANG, B. and L. ZHANG (1992) Theory and Applications of Problem Solving, Amsterdam: North-Holland. ZIARKO, W. (1993) Variable precision rough set model, Journal of Computer and System Sciences, 46, 39–59. The author Yiyu Yao Yiyu Yao is a professor of computer science in the Department of Computer Science, University of Regina, Regina, Saskatchewan, Canada. His research interests are information retrieval, Web intelligence, data mining, fuzzy sets, rough sets and granular computing. He received a PhD degree from the University of Regina, Canada. He has published over 100 journal and conference papers. He is a member of the editorial boards of the Web Intelligence and Agent Systems journal (IOS Press). He has served and is serving as a program co-chair of three international conferences, and as a program committee member in over 20 international conferences. He is a member of IEEE and ACM. Expert Systems, November 2003, Vol. 20, No. 5 ________________________________________________________________ 297 
work_dc54mmloafb7te6x6fbwljgm2a	----	PII: 0888-613X(88)90068-0 ABSTRACTS 1987 NAFIPS Workshop A F u z z y S e t - T h e o r e t i c A p p r o a c h t o D e c i s i o n M a k i n g S a l w a A m m a r Quantitative Methods Department, School o f Management, Syracuse University, Syracuse, New York 13244 In many complex decision situations the data available may not be sufficient to define a decision-making problem in an exact and objective form. The processing o f such data involves approximating and subjectively assessing the information necessary to describe the decision situation. The ability to capture the vague nature o f the information may be the key to solving an ill-defined problem. This paper presents quantitative models that accommodate the imprecision resulting from the vagueness and the subjectivity in the assessment o f decision situations. The tools o f quantification used to represent this imprecision are fuzzy set theory and operations. A new interpretation o f the decision-making process in a fuzzy environment is motivated. Guidelines by which the " b e s t " decision alternative can be determined are presented. I n t e g r a t i n g E x p e r t S y s t e m s U s i n g F u z z y N u m b e r s M i c h a e l J . B a l d w i n VEDA, Inc., 1002 Black River Boulevard, Suite 1, Rome, New York 13440 Real-time processing constraints o f operational expert avionics systems are driving a trend to concurrent processing o f several systems. This paper demonstrates an algorithmic technique to integrate expert systems where the systems are allowed to use fuzzy numbers as output. One approach to solving the integration problem is examined; each alternative is considered as a possible target, each j u d g e as one o f four expert planners, and each criterion as a consideration that is examined by at least one expert. The experts are allowed to use fuzzy numbers to quantify each target's suitability within each criterion. Four on-board experts are allowed to use fuzzy numbers because there is difficulty assigning specific real numbers to alternatives. The expert opinions are integrated in a manner that also considers the mission. Separate sets o f criteria weights are used, where each set provides a " m i s s i o n c o n t e x t " under which the analyst operates. The result is a ranking o f subsets o f targets, starting with the most desirable. Other analysts operating in parallel provide support for other aircrew knowledge needs. This approach minimizes the information pipeline to the pilot and exploits the p o w e r o f concurrent processing. G i v e n sufficient computing power, the design is one possible method for providing real-time data analysis. International Journal of Approximate Reasoning 1988; 2:95-112 © 1988 Elsevier Science Publishing Co., Inc. 52 Vanderbilt Ave., New York, NY 10017 0888-613X/88/$3.50 95 
work_dcjnnv2pm5dtbbugqybghxj57q	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 221003289 Params is empty 221003289 exception Params is empty 2021/04/06-03:05:49 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221003289 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:49 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_ddphwyykzngl5e4d6go2eoadsu	----	Provided by the author(s) and University College Dublin Library in accordance with publisher policies. Please cite the published version when available. Title Stability of topic modeling via matrix factorization Authors(s) Belford, Mark; MacNamee, Brian; Greene, Derek Publication date 2017-01 Publication information Expert Systems with Applications, 91 : 159-169 Publisher Elsevier Item record/more information http://hdl.handle.net/10197/9198 Publisher's statement þÿ�T�h�i�s� �i�s� �t�h�e� �a�u�t�h�o�r ��s� �v�e�r�s�i�o�n� �o�f� �a� �w�o�r�k� �t�h�a�t� �w�a�s� �a�c�c�e�p�t�e�d� �f�o�r� �p�u�b�l�i�c�a�t�i�o�n� �i�n� �E�x�p�e�r�t� �S�y�s�t�e�m�s� with Applications. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Expert Systems with Applications (98, (2018)) DOI:10.1016/j.eswa.2017.08.047 Publisher's version (DOI) 10.1016/j.eswa.2017.08.047 Downloaded 2021-04-06T02:06:12Z The UCD community has made this article openly available. Please share how this access benefits you. Your story matters! (@ucd_oa) © Some rights reserved. For more information, please see the item record link above. https://twitter.com/intent/tweet?via=ucd_oa&text=DOI%3A10.1016%2Fj.eswa.2017.08.047&url=http%3A%2F%2Fhdl.handle.net%2F10197%2F9198 Stability of Topic Modeling via Matrix Factorization Mark Belford∗, Brian Mac Namee, Derek Greene Insight Centre for Data Analytics, University College Dublin, Ireland Abstract Topic models can provide us with an insight into the underlying latent structure of a large corpus of documents. A range of methods have been proposed in the literature, including probabilistic topic models and techniques based on matrix factorization. However, in both cases, standard implementations rely on stochastic elements in their initialization phase, which can potentially lead to different results being generated on the same corpus when using the same parameter values. This corresponds to the concept of “instability” which has previously been studied in the context of k-means clustering. In many applications of topic modeling, this problem of instability is not considered and topic models are treated as being definitive, even though the results may change considerably if the initialization process is altered. In this paper we demonstrate the inherent instability of popular topic modeling approaches, using a number of new measures to assess stability. To address this issue in the context of matrix factorization for topic modeling, we propose the use of ensemble learning strategies. Based on experiments performed on annotated text corpora, we show that a K-Fold ensemble strategy, combining both ensembles and structured initialization, can significantly reduce instability, while simultaneously yielding more accurate topic models. Keywords: Topic modeling, Topic stability, LDA, NMF 1. Introduction Topic models aim to discover the latent semantic structure or topics within a corpus of documents, which can be derived from co-occurrences of words across the documents. Popular approaches for topic modeling have involved the application of probabilistic algorithms such as Latent Dirichlet Allocation (Blei et al., 2003). More recently, Non-negative Matrix Factor- ization approaches (Lee and Seung, 1999) have also been suc- cessfully applied to identify topics in unstructured text (Arora et al., 2012; Kuang et al., 2015). The standard formulations of both the LDA and NMF algo- rithms include stochastic elements in their initialization phase, prior to an optimization phase which produces a local solution. This random component can affect the final composition of the topics found and the rankings of the terms that describe those topics. This is problematic when seeking to capture a defini- tive topic modeling solution for a given corpus and represents a fundamental instability in these algorithms – different runs of the same algorithm on the same data can produce different out- comes. This problem has been widely studied in the context of partitional clustering algorithms such as k-means, which tends to converge to one of numerous local minima, depending on the choice of starting condition (Bradley and Fayyad, 1998). It has long been recognized as a significant drawback of such algorithms and a substantial number of works exist which at- ∗Corresponding author Email addresses: mark.belford@insight-centre.org (Mark Belford), brian.macnamee@ucd.ie (Brian Mac Namee), derek.greene@ucd.ie (Derek Greene) tempt to address the issue (e.g. Pena et al. (1999),Kuncheva and Vetrov (2006)). In the case of topic modeling, instability can manifest itself in two distinct aspects. The first can be observed when exam- ining the topic descriptors (i.e. the top terms representing each topic) over multiple runs. The term rankings may change con- siderably, where certain terms may appear or disappear com- pletely between runs. Secondly, issues of instability can also be observed when examining the degree to which documents have been associated with topics across different runs of the same al- gorithm on the same corpus. In both cases, such inconsistencies can potentially alter our interpretation and perception of a given topic model. Also, it is clear that any individual run should not be treated as a “definitive” summary of the underlying topics present in the data. Generally speaking, in the comparative evaluation of topic modeling approaches, researchers tend to focus on either the coherence of the topic descriptors (Newman et al., 2010) or the extent to which the topics accurately coincide with a set of ground truth categories or human annotations (Kuang et al., 2015). However, few researchers have considered the evalu- ation of different approaches from the point of view of their stability across multiple runs. In this paper we quantitatively assess the extent to which standard randomly-initialized NMF and LDA algorithms are unstable with respect to the topics that they produce on a di- verse collection of text corpora. To do this we propose mea- sures that capture the two distinct aspects of instability outlined above. We then focus on addressing the issue in the context of matrix factorization, exploring the use of strategies that involve Preprint version September 7, 2017 improved initialization and ensemble learning. In particular, we propose a new combined approach, motivated by the tradi- tional concept of k-fold cross-validation, which can yield stable results while also often producing more accurate and coherent models1. The rest of the paper is structured as follows. In Section 2 we provide an overview of relevant work in topic modeling and the more general area of cluster analysis. In Section 3 we discuss the problem of topic model instability in more detail, describing three new measures to quantify instability in topic models. In Section 4 we propose ensemble approaches to ad- dress the issue, which are subsequently evaluated on ten differ- ent text corpora in Section 5. Finally in Section 6 we conclude the paper with ideas for future work. 2. Related Work 2.1. Topic Modeling Topic models attempt to discover the hidden thematic struc- ture within an unstructured collection of text without relying on any form of training data. These models date back to the early work on latent semantic analysis (LSA) by Deerwester et al. (1990), who proposed applying SVD to decompose a document-term matrix to uncover the associations between terms and concepts in the data. In basic terms, a topic model con- sists of k topics, each represented by a ranked list of strongly- associated terms (often referred to as a “topic descriptor”). Each document in the corpus can also be associated with one or more of these topics to varying degrees. Considerable research on topic modeling has focused on the use of probabilistic methods, where a topic is viewed as a probability distribution over words, with documents being mixtures of topics (Steyvers and Griffiths, 2007). The most widely-applied probabilistic topic modeling approach has been LDA (Blei et al., 2003). Different approximation methods have been proposed for LDA inference, including variational infer- ence and Markov chain Monte Carlo (MCMC). Such approx- imation algorithms can converge to different local maxima on the same data (Zhao et al., 2015). The most commonly-used implementation, provided by the Mallet software package (Mc- Callum, 2002), relies on fast Gibbs sampling, where the initial state is determined by a user-specified random seed. Alternative algorithms, such as Non-negative Matrix Fac- torization (Lee and Seung, 1999), have also been effective in discovering topics in text corpora (Arora et al., 2012; Kuang et al., 2015). NMF is an unsupervised approach for reducing the dimensionality of non-negative matrices. When working with a document-term matrix A, the goal of NMF is to approximate this matrix as the product of two non-negative factors W and H, each with k dimensions. The rows of the factor H can be in- terpreted as k topics, defined by non-negative weights for each of the m terms in the corpus vocabulary. Ordering each row provides a topic descriptor, in the form of a ranking of the terms relative to the corresponding topic. The columns in the matrix 1See https://github.com/derekgreene/topic-ensemble W provide membership weights for all documents with respect to each of the k topics. One of the advantages of NMF over tra- ditional LDA methods is that there are fewer parameter choices involved in the modeling process, while it also has a tendency to identify more coherent topics than LDA (O’Callaghan et al., 2015). NMF is commonly initialized by assigning random non- negative weights to the entries in the factors W and H. By ap- plying an optimization process, such as alternating least squares (Lin, 2007), the factors are iteratively improved to reduce the approximation error until a local minimum is reached. As a result, the values in the initial pair of factors will have a signifi- cant impact on the values in the final factors (i.e. the topic-term and topic-document weights), even after a large number of iter- ations have been performed. Alternative initialization schemes for NMF have focused on increasing the accuracy of the final factors by using a more structured process, such as seeding us- ing a prior clustering algorithm (Wild et al., 2004). Another approach, Non-negative Double Singular Value Decomposition (NNDSVD) (Boutsidis and Gallopoulos, 2008), chooses initial factors based on a sparse SVD approximation of the original data matrix. This has been shown to be particularly effective on sparse data, such as text (O’Callaghan et al., 2015). In its ba- sic form, NNDSVD contains no stochastic element and should technically converge to the same pair of factors each time, al- though this depends on the underlying SVD implementation be- ing used. 2.2. Stability in Cluster Analysis Partitional clustering algorithms, such as k-means and k- medoids, have an inherent stability problem. That is to say, if we run the same algorithm on the same data or data drawn from the same source repeatedly, we frequently achieve different re- sults between each run. This variation can either be due to poor random seeds leading to convergence to different local minima (Pena et al., 1999), or as a result of perturbations in the data (Ben-Hur et al., 2002). One widely-adopted approach for dealing with the issue is to adopt a better cluster initialization strategy that is either fully deterministic or at least produces less variation than random initialization, while simultaneously yielding more useful clus- terings. A popular initialization approach proposed by Arthur and Vassilvitskii (2007), referred to as k-means++, involves choosing an initial seed item at random as the first cluster cen- ter and then choosing each subsequent cluster center with a probability proportional to its squared distance from the items nearest existing cluster centers. To further improve the result- ing clustering, this process can be repeated for several different initial seed items. While this strategy is not deterministic, it does tend to yield more consistent results across multiple runs. Researchers have also proposed fully deterministic strategies, where initial cluster centers are determined based on embed- ding methods such as PCA (Su and Dy, 2004), or coming from the prior application of another algorithm such as hierarchical clustering (Celebi and Kingravi, 2012). 2 2.3. Ensemble Methods An alternative strategy for reducing instability in unsuper- vised learning is to use ensemble clustering techniques, which are based on the premise that combining large, diverse sets of clusterings can produce a more stable and accurate solution (Strehl and Ghosh, 2002). Ensemble approaches are usually divided into two different stages. Firstly, a collection of base clusterings are generated (i.e. the ensemble members), typically by repeatedly applying an algorithm such as k-means with ran- dom initialization to the full dataset or to random samples of the data (Minaei-Bidgoli et al., 2004). Secondly, an integration function is applied to combine the base clusterings into a sin- gle consensus clustering. One of the most common integration strategies utilized is to leverage information from the ensem- ble regarding the level of “co-association” between all pairs of items. The underlying idea behind this is that items that are fre- quently assigned together in different clusterings will naturally belong to the same underlying group (Strehl and Ghosh, 2002). The resulting consensus clustering represents an approximation of the “average” clustering from among the ensemble members. Hadjitodorov et al. (2006) demonstrated a trade-off between di- versity and quality in cluster ensembles, and proposed a number of measures to quantify diversity for such ensembles. Work regarding the optimality and consistency of solutions produced by clustering and biclustering algorithms has been previously carried out in other domains (Bertoni and Valentini, 2005; Pio et al., 2015). However, in the general area of ma- trix factorization, there has been only some initial work on the use of ensemble approaches. This includes using a hierarchi- cal scheme to combine multiple factorizations in the study of protein networks (Greene et al., 2008) and the generation of ensembles of factorizations via a boosting-like approach (Suh et al., 2016). Recent work has also looked at using stability as a means of identifying an appropriate number of topics in a given corpus when applying NMF to text data (Greene et al., 2014). However, these studies have not investigated the extent of the problems introduced by instability in the context of topic modeling. 3. Stability of Topic Modeling In this section we introduce three new measures for assess- ing the stability of a collection of topic models, and use these to demonstrate how standard NMF and LDA approaches can be prone to produce unstable results when applied to text corpora. 3.1. Overview As discussed in Section 2, standard implementations of topic modeling approaches, such as LDA and NMF, commonly em- ploy stochastic initialization prior to optimization. As a result, the models they produce can vary quite considerably between different runs. Regardless of whether we are applying proba- bilistic or non-probabilistic algorithms, we can observe that this variation manifests itself in two ways: in relation to term-topic associations, or document-topic associations. In the former, the ranking of the top terms that describe a topic can change significantly between runs. In the latter, documents may be strongly associated with a given topic in one run, but may be more closely associated with an alternative topic in another run. In more extreme cases, a consequence of both manifestations is that topics can “appear” or “disappear” across different runs of the algorithm. This presents a challenge for domain experts who seek to gain a reliable insight into a particular corpus of documents. Depending on the topics resulting from a given algorithm run, their interpretation of the data may change con- siderably. However, the implications of this variation are rarely discussed in the topic modeling literature, particularly in the context of matrix factorization. 3.2. Measuring Stability To actually quantify the level of stability/instability present in a collection of topic models {M1, . . . ,Mr} generated over r runs on the same corpus, we propose three measures which reflect both aspects of topic model stability as described above. These measures are general in the sense that they can be ap- plied to models generated using either probabilistic or matrix factorization algorithms. 3.2.1. Descriptor Set Difference If the topics present in two topic models are similar, we should naturally expect that the prominent terms appearing in the topic descriptors in both models will be similar. Formally, if we represent each topic in a single model Mi with a number of top-ranked terms t, we can calculate the descriptor set as the union of top terms across all k topics, which we denote Ti. By measuring the symmetric difference between the descriptor sets for two different models, we can broadly gauge the similarity of the two models. This is useful as we can capture the variance at the descriptor level as terms may appear and disappear between runs. Formally, given two topic models Mi and Mj , each con- taining k topics represented by their top t terms, we calculate the descriptor set difference as: DSD(Mi,Mj) = Ti4Tj t×k (1) A value of 0 indicates identical descriptor sets (i.e. no differ- ence), while a value of 1 indicates that the topic descriptors for the two models share no common terms at all. Given a collec- tion of r topic models, we can calculate the Average Descriptor Set Difference (ADSD): ADSD = 1 r × (r −1) r∑ i,j,i 6=j DSD(Mi,Mj) (2) This produces a value ∈ [0,1], where a value closer to 0 for Eqn. 2 is indicative of a more stable collection of models. 3.2.2. Topic-Term Stability While the ADSD gives an overall measure of the difference between two models, it does not account for cases where topics are “mixed” across different runs of the algorithm (i.e. the same 3 terms appear in different topics across different runs). There- fore, we propose a measure that compares the similarity be- tween two topic models based on a pairwise matching process at the topic level. This is important as topics may appear and disappear between different runs and also helps to capture the variance at the individual topic level. First, given a pair of individual topics represented by their top t terms, we can measure the similarity between them based on the Jaccard Index: Jac(Ri,Rj) = |Ri ∩Rj| |Ri ∪Rj| (3) where Ri denotes the top t ranked terms for the i-th topic (i.e. its topic descriptor). We can use the above to build a measure of the agreement between two complete topic models, each con- taining k topics. We construct a k × k similarity matrix S, such that the entry Sxy indicates the agreement between the x-th topic in the first model and the y-th topic in the second model, as calculated using Eqn. 3. We then find the best match between the rows and columns of S (i.e. the topics from the first model and the second model). The optimal permutation π may be found in O(k3) time by solving the minimal weight bipartite matching problem using the Hungarian method (Kuhn, 1955). From this, we can produce a Term Stability (TS) score: TS(Mi,Mj) = 1 k k∑ x=1 Jac(Rix,π(Rix)) (4) where π(Rix) denotes the topic in model Mj matched to Rix in model Mi by the permutation π. Values for the above take the range [0,1], where a comparison between two identical k-way topic models will result in a score of 1. For a collection of r topic models {M1, . . . ,Mr}, we can calculate the Average Term Stability (ATS): ATS = 1 r × (r −1) r∑ i,j,i 6=j TS(Mi,Mj) (5) where a score of 1 indicates that all pairs of topic descriptors matched together across the r runs contain the same top t terms. 3.2.3. Partition Stability The second manifestation of topic model instability relates to document-topic associations. To measure the extent to which the associations between a document and one or more topics varies across different runs, for each run we can look at the dominant topic for every document. That is, we convert the document-topic associations (the probabilities in the case of LDA or the W factor weights in the case of NMF) into a dis- joint partition by taking the maximum value for each document. We can then compare the similarity between the partitions gen- erated in two runs using standard clustering agreement mea- sures. Utilizing this information allows us to observe if the dominant topic for each document changes frequently between runs. One widely-used such measure is Normalized Mutual In- formation (NMI) (Strehl and Ghosh, 2002), which quantifies the level of agreement between two partitions Pi and Pj : NMI(Pi,Pj) = I(Pi,Pj)√ H(Pi)H(Pj) (6) where I(Pi,Pj) is the mutual information between the assign- ments in the two partitions and H(Pi) is the entropy of the as- signments in Pi alone. We can compute the overall level of agreement between a set of r partitions generated by r runs of an algorithm on the same corpus as the mean Pairwise Normalized Mutual Infor- mation (PNMI) for all pairs: PNMI = 1 r × (r −1) r∑ i,j,i 6=j NMI(Pi,Pj) (7) where Pi is the partition produced from the document-topic as- sociations in model Mi. If the partitions across all models are identical, PNMI will yield a value of 1. 3.3. Examples of Instability We now provide examples of how the measures proposed above can reflect the problem of instability, using a corpus of news articles from the New York Times as an example (the nytimes-2003 corpus described later in Section 5). Firstly, to illustrate the issue of term instability, we con- sider topic models generated for r = 100 runs of randomly- initialized NMF and LDA, with a fixed number of topics k = 7 (corresponding to the number of annotated categories in the data). Fig. 1 shows the ADSD scores for each algorithm, as the number of top terms in the topic descriptors increases from 10 to 100. We can observe that, even with this relatively relaxed measure, there exists substantial variation in the terms appear- ing in the models for both algorithms. If we represent each topic using 10 terms, the ADSD score is as high as 0.47 for LDA and 0.31 for NMF. Even when we extend the topic descriptors to contain 100 terms, which we might expect to capture the bulk of the key terms for the topics in this corpus, the mean differ- ence across runs is 0.23 and 0.19 respectively. 0.00 0.10 0.20 0.30 0.40 0.50 10 20 30 40 50 60 70 80 90 100 A D S D Number of Top Terms NMF LDA Figure 1: Plot of the Average Descriptor Set Difference (ADSD) for 100 runs of NMF and LDA on the nytimes-2003 corpus, for increasing numbers of top terms representing each topic. 4 Table 1: Top-ranked terms for topics related to “sport”, generated across five runs of randomly-initialized NMF and LDA applied to a news corpus. # LDA Top 10 Terms 1 game, team, season, play, coach, games, points, players, against, football 2 game, season, team, coach, play, games points, league, football, players 3 game, season, coach, team, football, league, giants, play, jets, players 4 game, season, team, yankees, games, play, mets, nets, left, league 5 game, team, season, play, games, players, coach, yankees, time, against # NMF Top 10 Terms 1 game, season, team, yankees, games, nets, play, points, players, coach 2 game, team, season, nets, points, games, coach, play, knicks, players 3 game, team, season, nets, points, games, play, coach, knicks, giants 4 game, nets, team, season, coach, points, knicks, jets, giants, play 5 game, team, season, nets, points, games, play, knicks, coach, kidd We can explore this term instability further by inspecting the topic-term stability of the topics. As an example, Table 1 refers to five separate runs of a related topic (corresponding to the category “sport” in the ground truth for this corpus) for NMF and LDA. For both algorithms it is clear that the order- ing of the top terms for this topic can change considerably, and it is also possible for terms to completely disappear between different runs. Using the same corpus, we can also explore the partition stability afforded by both NMF and LDA. Fig. 2 plots the distri- butions for the NMI agreement scores between all pairs of par- titions corresponding to the set of 100 models produced by each algorithm. Two partitions with identical document assignments would yield an NMI score of 1. However, only 0.4% of all pairs of partitions achieve this for NMF. In the case of LDA, of the 4950 unique pairs of partitions, none achieve perfect agreement and only 0.3% achieve an NMI score ≥ 0.9. When we aver- age the agreement scores over all runs, the overall PNMI scores for the two algorithms are 0.78 and 0.66 respectively, indicating there is considerable variation in the outputs of both algorithms 0% 5% 10% 15% 20% 25% 30% 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 P er ce nt ag e of P ai rs (a) NMF 0% 5% 10% 15% 20% 25% 30% 35% 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 P er ce nt ag e of P ai rs (b) LDA Figure 2: Distribution of pairwise NMI agreement scores for document parti- tions resulting from 100 runs of (a) NMF, (b) LDA on the nytimes-2003 corpus. Table 2: Top-ranked documents for topics related to “sport”, generated across five runs of randomly-initialized NMF and LDA applied to a news corpus. # LDA Top 10 Documents 1 s4310, s5376, s4247, s5262, s6055, s1493, s3167, s5670, s4972, s6636 2 s4441, s6267, s4247, s3521, s8146, s5262, s4708, s4681, s4460, s8937 3 s0113, s3521, s4708, s1698, s5299, s4972, s8577, s8351, s5855, s6834 4 s6267, s0113, s5376, s9894, s3521, s5262, s9116, s4681, s4708, s8937 5 s6267, s0113, s9894, s4247, s5262, s4681, s2056, s4972, s8577, s6636 # NMF Top 10 Documents 1 s5995, s6558, s3547, s9993, s5281, s8029, s2484, s5114, s1227, s2934 2 s6558, s5995, s8029, s5457, s5843, s2484, s9993, s1227, s5193, s2068 3 s8029, s6558, s5995, s5457, s5843, s2484, s1227, s9993, s5193, s2068 4 s5457, s8029, s5843, s6558, s9993, s5193, s2068, s7687, s2484, s9924 5 s8029, s6558, s5995, s5457, s5843, s2484, s1227, s9993, s5193, s2265 across these runs. Again, manually inspecting the top-ranked documents in the topics for these models reveals the extent of the variation. Table 2 lists the identifiers of the top ten documents assigned to each of the topics related to “sport” selected from five runs of NMF and LDA in Table 1. We observe that, similar to the case of the top terms, the ordering of documents is also subject to the same inherent instability. 3.4. Redundant Stability While the examples above demonstrate that the production of robust, reliable topic models is important, it is also necessary to emphasize that stability should not be the sole requirement for a useful topic modeling algorithm. As observed by Ben-Hur et al. (2002) in the context of partitional clustering, in some situations stability can simply be indicative of an algorithm’s tendency to converge to a given local solution, regardless of the quality of that solution. In the context of NMF, we could initial- ize the factors W and H in a deterministic way with arbitrary non-negative values. However, this “redundant stability” is un- likely to provide a useful model. Therefore, in the next section we propose techniques that yield solutions that are not only sta- ble but also accurate – i.e. the topics are semantically coherent and provide a useful insight into the content of the corpus. 4. Methods We now propose ensemble methods for topic modeling via matrix factorization, which can be utilized to address the is- sue of stability, while also potentially producing more accurate topic models for a corpus of unstructured text. 4.1. Basic Ensemble Method We apply ensemble learning for topic modeling in the form of two layers of matrix factorization. Fig. 3 shows an overview of the method, which can naturally be divided into two steps, similar to existing strategies in ensemble clustering (Strehl and Ghosh, 2002): 1. Generation: Create a set of base topic models by exe- cuting r runs of NMF applied to the same corpus, repre- sented as a document-term matrix A. 5 NMF Original Corpus Base Models Topic-Term Matrix NMF Ensemble Topic Model Generation Step Integration Step Figure 3: Overview of a general ensemble strategy for topic modeling with matrix factorization, consisting of two steps: generation and integration. 2. Integration: Transform the base topic models to a suit- able intermediate representation, and apply a final run of NMF to produce a single ensemble topic model, which represents the final output of the method. We now discuss both of these steps in more detail. 4.1.1. Ensemble Generation Unsupervised ensemble procedures typically seek to encour- age diversity with a view to improving the quality of the infor- mation available in the integration phase (Topchy et al., 2005). Therefore, we create a diverse set of r base topic models (i.e. the topic term descriptors and document assignments will differ from one base model to another). Here we encourage diver- sity by relying on the inherent instability of NMF with random initialization – we generate each base model by populating the factors W and H with values based on a different random seed, and then applying NMF to A. In each case we use a fixed pre- specified value for the number of topics k. After each run, the H factor from the base topic model (i.e. the topic-term weight matrix) is stored for later use. 4.1.2. Ensemble Integration Once we have generated a collection of r factorizations, in the second step we create a new representation of our corpus in the form of a topic-term matrix M. The matrix is created by stacking the transpose of each H factor generated in the first step, as illustrated in Fig. 4. Here each factor consists of k top- ics {t1, . . . , tk} and m terms {w1, . . . ,wm}. We construct this topic-term matrix as we may often expect to see similar topics appearing between different runs. However, they may not be identical with respect to their terms, and we wish to leverage this variance. It is important to note that this process of com- bining the factors is order independent. This results in a matrix M where each row corresponds to a topic from one of the base topic models, and each column is a term from the original cor- pus. Each entry Mij holds the weight of association for term i in relation to a single topic from a base model. Once we have created M, we apply the second layer of NMF to this matrix to produce the final ensemble topic model. The reasoning behind applying NMF a second time to these topic descriptors is that they explicitly capture the variance be- tween the base topic models. To improve the quality of the resulting topics, we generate initial factors using NNDSVD ini- tialization (Boutsidis and Gallopoulos, 2008). As an input pa- rameter to NMF, we specify a final number of k′ topics, which is typically set to be the same as the value k used in the gen- eration step. The resulting H factor provides weights for the terms for each of the k′ ensemble topics – the top-ranked terms in each column can be used as descriptors for a topic. To pro- duce weights for the original documents in our corpus, we can “fold” the documents into the ensemble model by applying a projection to the document-term matrix A: D = A ·H T Each row of D now corresponds to a document, with columns corresponding to the k′ ensemble topics. An entry Dij indicates the strength of association of document i in ensemble topic j. 4.2. K-Fold Ensemble Method While the basic ensemble generation approach described in Section 4.1.1 does yield a diverse set of base topic models, the use of random initialization means that some of these models H1 t1 tk H2 t1 tk t1 tk Hr w1 w2 w3 w4 w5 w6 wm Figure 4: Illustration of the stacking process where the H factors from r topic models, each consisting of k topics and m terms, are combined to create a single topic-term matrix. 6 1. Construct full document-term matrix A for the corpus. 2. For p rounds: 1. Randomly divide corpus into f folds. 2. For each of the f folds: (a) Exclude the current fold. (b) Apply NMF with NNDSVD initialization to doc- uments in A from the other (f −1) folds to gen- erate k topics. 3. From the resulting p × f models, construct a topic-term matrix M by stacking the transpose of each H factor. 4. Apply NMF with NNDSVD initialization to M to gener- ate k′ topics, where the factor H gives the final topic-term associations. 5. Compute D = A · H T to find the final document-term as- sociations. Figure 5: Summary of K-Fold ensemble topic modeling method. will correspond to poor local minima with low accuracy. Fur- thermore, given the number of possible initial factors that could be generated in this way, there is still potential for several runs of the complete ensemble process to yield somewhat different final results. Therefore, we consider the use of improved ini- tialization to generate more accurate base models, while also using a more structured strategy to create the models in order to reduce variability. This strategy is based on traditional k- fold cross-validation as performed in evaluation in supervised learning. In our case, we randomly divide the corpus of documents into f folds of equal size. Each of the f folds is excluded in turn, and we apply NMF with NNDSVD initialization to the documents from the remaining (f−1) folds, yielding f models. To reduce variability, we repeat the process for p rounds using different splits of the data, yielding a total of p×f topic mod- els, each generated on a large subsample of the corpus. This collection of base topic models is then integrated as described in Section 4.1.2 to produce a final topic model. A full summary of this approach is given in Fig. 5. 5. Evaluation In this section we comprehensively assess the problem of instability in topic modeling for standard NMF and LDA ap- proaches on a diverse collection of corpora, and examine the extent to which superior initialization and ensemble methods can improve the stability of NMF-based approaches, while also yielding accurate models. 5.1. Datasets For our experiments, we use a diverse set of ten corpora, including both high-quality long texts and user-generated con- tent. All of these corpora have human annotated “ground truth” topical categories, allowing us to evaluate model accuracy. Six of these datasets consist of news articles from individual main- stream news sources (BBC, The New York Times, The Guardian, and The Irish Times), categorized by subject matter. Two more datasets consist of pages from a 2014 Wikipedia dump, catego- rized by their associated WikiProject. These eight datasets were previously used for topic modeling evaluations (Greene et al., 2014). We also include the popular 20-newsgroups dataset, where the ground truth categories correspond to individual news- groups (e.g. “comp.graphics”, “comp.windows.x”,“rec.autos”). To evaluate performance on social media data, we include a newly-collected corpus in our experiments, known as the 20- topics dataset, which consists of 4,170,382 tweets from 1,200 prominent Twitter accounts. These accounts have been manu- ally assigned to 20 different categories (e.g. “aviation”, “health”, “tech”). Each document in the corpus corresponds to the con- catenation of the tweets posted by a single user for a given week during the period March 2015 to February 2016. The corpus contains 40,498 such “user documents”. A detailed summary of all corpora used in our experiments is provided in Table 3. 5.2. Experimental Setup When pre-processing the corpora, terms appearing in < 20 documents are filtered. We use a single list of common English stop-words for all datasets. LDA operates on bag-of-words text representations, and so was applied to the raw frequency val- ues. For NMF, the same documents were transformed to log- based Term Frequency-Inverse Document Frequency (TF-IDF) vectors, and document length normalization was subsequently applied to produce the final document-term matrix. In our experiments, we compare five different topic model- ing approaches: 1. Standard LDA with random seeding, using the popular Mallet implementation with Gibbs sampling (McCallum, 2002). 2. NMF with random initialization, using the fast alternat- ing least squares variant proposed by (Lin, 2007) and pro- vided by the sckit-learn toolkit (Pedregosa et al., 2011). 3. NMF with non-random NNDSVD initialization, also im- plemented in sckit-learn. 4. Basic ensemble topic modeling for matrix factorization with random initialization, as described in Section 4.1. 5. K-Fold ensemble topic modeling for matrix factorization combined with improved initialization, as described in Section 4.2. For these approaches, there are a number of common and dis- tinct parameters which need to be specified: Common parameters: For all approaches, the number of top- ics k is set to correspond to the number of ground truth categories for each dataset. NMF parameters: For NMF with both random and NNDSVD initialization, the maximum number of iterations is set to 100 by default. For the random case, a different random seed is used for each run to populate values in the initial 7 Table 3: Details of the ten corpora used in our experiments, including the total number of documents n, number of terms m, and number of categories k̂ in the associated “ground truth” annotations. Corpus n m k̂ Description bbc 2,225 3,121 5 General news articles from the BBC from 2003. bbc-sport 737 969 5 Sports news articles from the BBC from 2003. guardian-2013 6,520 10,801 6 Corpus of news articles published by The Guardian during 2013. irishtimes-2013 3,246 4,832 7 Corpus of news articles published by The Irish Times during 2013. nytimes-1999 9,551 12,987 4 A subset of the New York Times Annotated Corpus from 1999. nytimes-2003 11,527 15,001 7 A subset of the New York Times Annotated Corpus from 2003. wikipedia-high 5,738 17,311 6 Subset of 2014 Wikipedia dump, where articles are assigned labels based on their high level WikiProject. wikipedia-low 4,986 15,441 10 Subset of 2014 Wikipedia dump, where articles are labeled with fine-grained WikiProject sub-groups. 20-newsgroups 18,662 9,954 20 Collection of posts from 20 different internet newsgroups. 20-topics 40,498 32,464 20 Tweets from user accounts associated with 20 different topical areas. factors W and H. This process is repeated for r = 100 runs. LDA parameters: The LDA algorithm has two additional hy- perparameters. We use the Mallet default values, with α = 5 and β = 0.01. The maximum number of iterations is set to 1000. For each run, a different random seed is used to initialize the Gibbs sampling process. This pro- cess is repeated for r = 100 runs. Basic ensemble: For the first ensemble approach, we integrate a collection of 100 members, generated via random ini- tialization. The final number of topics k′ is set to be the same as the number of ground truth categories for each dataset. This entire process is repeated 20 times to allow us to assess stability. K-Fold ensemble: For the second ensemble approach, we ap- ply p = 10 rounds of f = 10 folds, thus also yielding a collection of 100 ensembles members for integration, with k′ determined as above. Again this entire process is repeated 20 times. 5.3. Model Stability To assess the stability of a collection of models generated by each algorithm, we use the term-based measures ADSD (Eqn. 2) and ATS (Eqn. 5) using the top t = 10 terms for each topic, and the document-level PNMI measure (Eqn. 7). Results for these measures are shown in Tables 4, 5, and 6 respectively. Across all three measures, we observe that the NNDSVD NMF and K- Fold approaches clearly yield the most stable results. Both of these methods produce models with perfect stability for the ma- jority of our datasets - i.e. they yield models in which the topic- term and document-topic associations remain the same. As ex- pected, the randomly-initialized approaches perform the worst due to their inherent instability caused by stochastic elements as can be identified by their standard deviation scores. While the basic ensemble approach yields high stability for the smaller corpora, we do see some variation between different runs at the term and document level for the larger corpora. Here, the ran- dom initialization used when generating the ensemble members is still leading to variation in the results at the final ensemble in- tegration phase, even with 100 ensemble members. However, the more structured nature of the generation phase for the K- Fold approach effectively negates this problem. It is interesting to observe that, for the 20-newsgroups and 20-topics datasets, which contain noisier user-generated con- tent and a larger number of underlying topics, the K-Fold en- semble approach yields higher levels of stability than NNDSVD- initialized NMF. This suggests that combining the subsampling element of the ensemble process with a structured NNDSVD initialization produces a more reliable solution. It is important to note that the widely-used implementa- tion of NNDSVD provided by the sckit-learn toolkit, as used in our experiments, relies on an approximate truncated singular value decomposition method involving randomization, in order to make it applicable to large data matrices. While the resulting decompositions are often identical, this is not always the case. Computing a full SVD would eliminate the instability, with the trade-off that the running time requirements for decomposing a large, high-dimensional document-term matrix would increase dramatically. 5.4. Model Quality The primary focus of our work is on model stability. But as noted in Section 3.4, stability without meaningful and coher- ent topics is unlikely to be useful. Therefore we consider the quality of the models produced by each of the five methods, in terms of accuracy and coherence. Partition accuracy: In either the case of NMF or LDA, we can convert a model’s document-topic associations into a disjoint partition. We can then compare this partition with the corresponding disjoint “ground truth” categories for each corpus using NMI (Strehl and Ghosh, 2002). Topic coherence: Coherence refers to the overall quality and the semantic relatedness of the terms appearing in a topic 8 Table 4: Model stability — Comparison of Average Descriptor Set Difference (ADSD) scores for five topic modeling approaches. Corpus LDA NMF NNDSVD Ensemble K-Fold bbc 0.14 ± 0.15 0.15 ± 0.25 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 bbc-sport 0.29 ± 0.14 0.21 ± 0.21 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 guardian-2013 0.15 ± 0.16 0.18 ± 0.20 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 irishtimes-2013 0.35 ± 0.13 0.42 ± 0.22 0.00 ± 0.00 0.01 ± 0.01 0.01 ± 0.01 nytimes-1999 0.23 ± 0.18 0.36 ± 0.22 0.00 ± 0.00 0.21 ± 0.23 0.00 ± 0.00 nytimes-2003 0.47 ± 0.13 0.31 ± 0.18 0.00 ± 0.00 0.01 ± 0.01 0.00 ± 0.00 wikipedia-high 0.26 ± 0.12 0.21 ± 0.13 0.00 ± 0.00 0.01 ± 0.01 0.00 ± 0.00 wikipedia-low 0.21 ± 0.06 0.12 ± 0.09 0.00 ± 0.00 0.06 ± 0.07 0.00 ± 0.01 20-newsgroups 0.31 ± 0.06 0.32 ± 0.09 0.18 ± 0.08 0.12 ± 0.05 0.05 ± 0.04 20-topics 0.23 ± 0.06 0.23 ± 0.09 0.07 ± 0.07 0.11 ± 0.06 0.01 ± 0.01 Table 5: Model stability — Comparison of term stability scores for five topic modeling approaches, calculated using Average Term Stability (ATS). Corpus LDA NMF NNDSVD Ensemble K-Fold bbc 0.86 ± 0.14 0.88 ± 0.21 1.00 ± 0.00 1.00 ± 0.00 1.00 ± 0.00 bbc-sport 0.68 ± 0.15 0.80 ± 0.20 1.00 ± 0.00 1.00 ± 0.00 1.00 ± 0.00 guardian-2013 0.83 ± 0.18 0.83 ± 0.19 1.00 ± 0.00 1.00 ± 0.00 1.00 ± 0.00 irishtimes-2013 0.61 ± 0.14 0.64 ± 0.18 1.00 ± 0.00 0.99 ± 0.01 0.99 ± 0.01 nytimes-1999 0.65 ± 0.19 0.72 ± 0.17 1.00 ± 0.00 0.82 ± 0.19 1.00 ± 0.00 nytimes-2003 0.53 ± 0.12 0.76 ± 0.14 1.00 ± 0.00 0.99 ± 0.01 1.00 ± 0.00 wikipedia-high 0.79 ± 0.15 0.77 ± 0.13 1.00 ± 0.00 0.99 ± 0.01 1.00 ± 0.00 wikipedia-low 0.79 ± 0.10 0.88 ± 0.08 1.00 ± 0.00 0.94 ± 0.07 0.99 ± 0.01 20-newsgroups 0.65 ± 0.07 0.66 ± 0.09 0.81 ± 0.08 0.86 ± 0.05 0.93 ± 0.05 20-topics 0.69 ± 0.08 0.78 ± 0.08 0.94 ± 0.07 0.89 ± 0.06 0.99 ± 0.01 Table 6: Model stability — Comparison of document stability scores for five topic modeling approaches, based on Pairwise Normalized Mutual Information (PNMI). Corpus LDA NMF NNDSVD Ensemble K-Fold bbc 0.86 ± 0.09 0.89 ± 0.10 1.00 ± 0.00 1.00 ± 0.00 1.00 ± 0.00 bbc-sport 0.74 ± 0.08 0.87 ± 0.09 1.00 ± 0.00 1.00 ± 0.00 1.00 ± 0.00 guardian-2013 0.84 ± 0.07 0.88 ± 0.07 1.00 ± 0.00 1.00 ± 0.00 1.00 ± 0.00 irishtimes-2013 0.73 ± 0.07 0.81 ± 0.07 1.00 ± 0.00 0.99 ± 0.00 1.00 ± 0.00 nytimes-1999 0.69 ± 0.09 0.67 ± 0.14 1.00 ± 0.00 0.83 ± 0.13 0.98 ± 0.01 nytimes-2003 0.66 ± 0.06 0.78 ± 0.07 1.00 ± 0.00 0.97 ± 0.01 0.99 ± 0.00 wikipedia-high 0.86 ± 0.07 0.86 ± 0.04 1.00 ± 0.00 0.99 ± 0.00 1.00 ± 0.00 wikipedia-low 0.89 ± 0.04 0.89 ± 0.03 1.00 ± 0.00 0.96 ± 0.04 1.00 ± 0.00 20-newsgroups 0.63 ± 0.02 0.69 ± 0.04 0.80 ± 0.05 0.89 ± 0.04 0.96 ± 0.03 20-topics 0.84 ± 0.03 0.84 ± 0.03 0.97 ± 0.03 0.97 ± 0.02 1.00 ± 0.00 descriptor. While a range of measures have been pro- posed in the literature, we employ a widely-used mea- sure, Normalized Pointwise Mutual Information (NPMI) (Bouma, 2009), which uses term co-occurrence counts from the full corpus to measure the average coherence of the topics in a given model, based on the top t terms in their descriptors. In our evaluations we use t = 10 terms. With regards to the NPMI coherence of the topics produced, Table 7 shows that our proposed basic ensemble and K-Fold approaches perform the best. However, it should be noted that in most cases the differences in the average coherence scores are small. The most noticeable gap is between the LDA ap- proach and the other NMF-based approaches, which may re- flect the tendency of LDA to produce more generic and less semantically-coherent terms (O’Callaghan et al., 2015). To provide a clearer measure of model quality, we next consider the quality of the five methods by evaluating the par- tition accuracy with respect to the ground truth labels. Ta- ble 8 shows the means and standard deviations of the NMI scores for the methods on the ten corpora. Here we see that the best-performing algorithms are NNDSVD-initialized NMF and the K-Fold ensemble approach, although this varies with the dataset. The randomly-initialized algorithms exhibit con- siderable variation in the quality of the models they produce, as indicated by the standard deviation scores, and are worse on average than the ensemble and SVD-based methods, with the 9 Table 7: Model quality — Comparison of topic coherence scores for five topic modeling approaches, based on Normalized Pointwise Mutual Information (NPMI). Corpus LDA NMF NNDSVD Ensemble K-Fold bbc 0.09 ± 0.01 0.15 ± 0.01 0.16 ± 0.00 0.16 ± 0.00 0.16 ± 0.00 bbc-sport 0.11 ± 0.01 0.14 ± 0.01 0.14 ± 0.00 0.15 ± 0.00 0.14 ± 0.00 guardian-2013 0.11 ± 0.01 0.15 ± 0.01 0.16 ± 0.00 0.16 ± 0.00 0.16 ± 0.00 irishtimes-2013 0.08 ± 0.01 0.12 ± 0.01 0.13 ± 0.00 0.13 ± 0.00 0.13 ± 0.00 nytimes-1999 0.10 ± 0.01 0.12 ± 0.01 0.12 ± 0.01 0.12 ± 0.01 0.12 ± 0.01 nytimes-2003 0.10 ± 0.03 0.12 ± 0.01 0.12 ± 0.01 0.10 ± 0.02 0.12 ± 0.01 wikipedia-high 0.15 ± 0.01 0.23 ± 0.01 0.24 ± 0.00 0.23 ± 0.00 0.24 ± 0.00 wikipedia-low 0.19 ± 0.01 0.24 ± 0.01 0.25 ± 0.00 0.24 ± 0.00 0.25 ± 0.00 20-newsgroups 0.12 ± 0.01 0.15 ± 0.01 0.16 ± 0.01 0.15 ± 0.00 0.16 ± 0.00 20-topics 0.19 ± 0.01 0.30 ± 0.01 0.30 ± 0.00 0.29 ± 0.00 0.29 ± 0.00 Table 8: Model quality — Comparison of document partition accuracy scores for five topic modeling approaches, based on Normalized Mutual Information (NMI). Corpus LDA NMF NNDSVD Ensemble K-Fold bbc 0.81 ± 0.06 0.79 ± 0.04 0.82 ± 0.00 0.79 ± 0.00 0.80 ± 0.00 bbc-sport 0.68 ± 0.05 0.80 ± 0.06 0.83 ± 0.00 0.85 ± 0.00 0.85 ± 0.00 guardian-2013 0.77 ± 0.04 0.82 ± 0.04 0.83 ± 0.00 0.84 ± 0.00 0.84 ± 0.00 irishtimes-2013 0.68 ± 0.04 0.72 ± 0.04 0.77 ± 0.00 0.76 ± 0.00 0.77 ± 0.00 nytimes-1999 0.64 ± 0.04 0.53 ± 0.05 0.49 ± 0.00 0.51 ± 0.02 0.51 ± 0.00 nytimes-2003 0.60 ± 0.03 0.58 ± 0.04 0.55 ± 0.00 0.56 ± 0.01 0.58 ± 0.00 wikipedia-high 0.69 ± 0.02 0.72 ± 0.01 0.72 ± 0.00 0.73 ± 0.00 0.74 ± 0.00 wikipedia-low 0.84 ± 0.02 0.87 ± 0.03 0.86 ± 0.00 0.89 ± 0.02 0.88 ± 0.00 20-newsgroups 0.48 ± 0.01 0.47 ± 0.02 0.49 ± 0.01 0.48 ± 0.01 0.48 ± 0.01 20-topics 0.84 ± 0.02 0.83 ± 0.02 0.85 ± 0.01 0.87 ± 0.01 0.88 ± 0.00 exception of the case where LDA is applied to the two New York Times corpora. 5.5. Statistical Significance To further investigate the differences between the algorithms, we performed a series of statistical tests on the results presented in the previous section. We carried out a non-parametric Fried- man’s Aligned Rank test (Garcı́a et al., 2010) for each of the five measures previously reported (ADSD, ATS, PNMI, NPMI, and NMI) to test for the presence of statistically significant dif- ferences in the results amongst the five algorithms and across the ten datasets. These tests returned p-values of 0.000004 (ADSD), 0.000002 (ATS), 0.000003 (PNMI), 0.00002 (NPMI), and 0.118 (NMI) respectively. This indicates that statistically significant differences, at the 1% confidence level, exist in the results achieved by the different algorithms for each measure, except for partition accuracy (NMI). To determine if there was a statistically significant differ- ence between our proposed K-Fold algorithm and the other topic modeling approaches with respect to the four remaining mea- sures, we performed a series of Friedman’s Aligned Rank Pair- wise post hoc tests (Garcı́a et al., 2010), with the K-Fold ap- proach used as a control. The results from these tests are re- ported in Table 9. It is interesting that there is a statistically significant difference, at the 1% confidence level, between our proposed approach and the two randomly initialized topic mod- eling algorithms across all measures, which along with the pre- viously reported performance of the algorithm suggests that the K-Fold approach produces more stable and higher quality topic models. There is no statistical difference between our proposed K-Fold approach, the ensemble and NNDSVD, which may in- dicate that these three measures are similar due to producing more deterministic solutions and this notion is further strength- ened due to their similar performance regarding the measures previously reported. It is also interesting to note that there is a statistical difference between the LDA and the K-Fold approach with regards to coherence, likely due to LDA based approaches generating topics with a lower coherence (O’Callaghan et al., 2015). 5.6. Computational Expense While the timing of algorithms can vary depending on the hardware and implementation utilized, it is useful in this case to obtain an estimate of how much longer the proposed ensem- ble approaches take with respect to traditional topic modeling algorithms. Each topic modeling approach was run 100 times and the average times are reported in Table 10. The experi- ments were carried out on a machine with 12 2.4GHz cores and 128GB of RAM. These running times are impacted due to numerous factors, including the number of documents in the corpus, the dimensionality of the corresponding document- term matrix, and the number of topics k selected for the corpus. As expected, it is clear that the two ensemble approaches take considerably longer to run than the other algorithms. This is naturally due to the underlying nature of their generation step, where 100 iterations of NMF have to be generated. While the 10 Table 9: Posthoc p-values based on Friedman’s Aligned Rank Pairwise test for each of the four measures that were statistically significant, while using the K-Fold approach as a control algorithm. * p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.001, **** p ≤ 0.0001 Measure LDA NMF NNDSVD Ensemble K-Fold ADSD **** 0.00001 **** 0.00002 0.645 0.273 NA ATS **** 0.000002 **** 0.0001 0.713 0.250 NA PNMI **** 0.000002 **** 0.00007 0.576 0.304 NA NPMI **** 0.00001 0.111 0.921 0.452 NA Table 10: Comparison of average running time in seconds for the five topic modeling approaches. Corpus LDA NMF NNDSVD Ensemble K-Fold bbc 40.74 0.22 0.30 22.68 30.00 bbc-sport 33.34 0.08 0.08 7.90 8.56 guardian-2013 231.33 1.98 1.94 198.89 189.87 irishtimes-2013 82.75 0.85 0.86 86.76 84.69 nytimes-1999 370.85 2.95 3.74 296.18 375.97 nytimes-2003 454.36 6.11 4.83 613.00 538.57 wikipedia-high 667.88 5.12 3.31 513.68 319.96 wikipedia-low 627.28 6.42 4.07 644.70 380.85 20-newsgroups 229.29 14.16 15.15 1421.44 1527.62 20-topics 1802.16 72.07 115.02 7226.22 6793.77 LDA Mallet implementation is widely used for topic modeling, we observe that it is considerably slower in the majority of our experiments, in fact it is frequently slower than our proposed K-Fold approach which utilizes a more structured but slower initialization step. 5.7. Discussion So far we have discussed the two main criteria for evaluat- ing topic modeling algorithms separately. These are the evalu- ation of model quality, by examining topic coherence and par- tition accuracy, and the evaluation of model stability by exam- ining term stability and document stability. However, it is im- portant to note that the output of both criteria should be consid- ered together – our evaluations highlight that some topic mod- eling approaches perform well with respect to one criterion, while performing poorly with respect to the other. An exam- ple of this can be seen in the results produced by NNDSVD for the nytimes-1999 dataset. For the term and document stability, perfect stability scores are achieved (Tables 5 and 6). How- ever, when we take into account partition accuracy for the same dataset (Table 8), the quality of this solution is not as good as it initially appears. The NNDSVD initialization actually per- forms the worst with regards to accuracy in this case, with a low NMI score of 0.49. Similarly, while randomly-initialized LDA out-performs all other approaches on both New York times cor- pora in terms of partition accuracy (Table 8), the corresponding stability scores are consistently poor across all measures (Ta- bles 4–6). Following the discussion of “redundant stability” in Section 3.4, these results raise an interesting problem in that, while we strive to produce the most stable results as possible, it may also be the case that these results are of poor quality from a model quality standpoint. Among the two newly-proposed approaches, it is interesting to observe that the basic ensemble approach does not perform as well as the K-Fold approach, even though they are based on a similar ensemble process. In the case of the former, we aim to promote diversity when generating our base ensemble members by using randomly-initialized NMF, as motivated by previous work in both supervised and unsupervised ensemble learning (Brown et al., 2005; Kuncheva and Hadjitodorov, 2004). How- ever, for larger datasets, the stochastic nature of this approach tends to cause the final results to contain a degree of variance across different runs of the overall ensemble. In contrast, by combining structured document subsampling with NNDSVD initialization to generate each ensemble member, the K-Fold approach exhibits very little instability across the 20 runs in our experiments, as indicated in the results in Tables 4–6. These findings correspond to those of Hadjitodorov et al. (2006), who demonstrated that a moderate level of diversity leads to useful ensembles in cluster analysis. Overall, among the techniques considered in our experi- ments, the K-Fold ensemble approach produces the best models when taking into account both quality and stability. While the observed NMI scores were lower for certain datasets, it still performs better than the other methods with respect to half of our corpora. This strategy also appears to handle noisy user- generated data well, in comparison to alternative techniques. 6. Conclusions While topic modeling methods such as LDA and NMF are widely applied in a range of domains to analyze unstructured text, researchers often do not consider the effect that random initialization has on models produced by these methods. In this paper we have demonstrated that, for both methods, this 11 can result in significant variations in the topics produced over multiple runs over the same corpus. This effect is manifested both at the term and document level, which can potentially lead to different human interpretations of the underlying thematic structure of the data. To address the issue of instability in the context of NMF, we have investigated the extent to which improved algorithm initialization and ensemble strategies can produce more stable models, which are almost potentially more accurate and insight- ful. We compared the performance of these approaches with re- gards to five different metrics that measure stability, accuracy, and coherence of topics. Our results indicate that a new K-Fold ensemble approach afforded the most stable and accurate set of models, although initializating NMF based on a SVD approx- imation of the document-term matrix can also provide a clear improvement over standard NMF and LDA methods. One concern that arises in the application of ensemble learn- ing techniques in general relates to scalability. While the en- semble techniques described in this paper can be naturally par- allelized, there is considerable scope for reducing the compu- tation time required to generate the ensemble. A potentially promising idea to investigate in this context relates to the con- cept of snapshot ensembles from supervised learning (Huang et al., 2017), where a single algorithm run is allowed to con- verge to several local minima during the optimization process, each providing a contribution to the overall ensemble. Such an approach might also be used to yield more stable topic models via matrix factorization and reduce the computational expense. Acknowledgement. This research was supported by Science Foundation Ireland (SFI) under Grant Number SFI/12/RC/2289. References Arora, S., Ge, R., and Moitra, A. (2012). Learning topic models – Going be- yond SVD. In Proc. 53rd Symposium on Foundations of Computer Science, pages 1–10. IEEE. Arthur, D. and Vassilvitskii, S. (2007). k-means++: The advantages of careful seeding. In Proc. 18th Annual ACM-SIAM symposium on Discrete algo- rithms, pages 1027–1035. Society for Industrial and Applied Mathematics. Ben-Hur, A., Elisseeff, A., and Guyon, I. (2002). A stability based method for discovering structure in clustered data. In Proc. 7th Pacific Symposium on Biocomputing, pages 6–17. Bertoni, A. and Valentini, G. (2005). Random projections for assessing gene expression cluster stability. In Proc. IEEE International Joint Conference on Neural Networks (IJCNN’05), volume 1, pages 149–154. Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent dirichlet allocation. Journal of Machine Learning Research, 3:993–1022. Bouma, G. (2009). Normalized Pointwise Mutual Information in Collocation Extraction. In Proc. International Conference of the German Society for Computational Linguistics and Language Technology, GCSL ’09. Boutsidis, C. and Gallopoulos, E. (2008). SVD based initialization: A head start for non-negative matrix factorization. Pattern Recognition. Bradley, P. S. and Fayyad, U. M. (1998). Refining initial points for k-means clustering. In Proc. 15th International Conference on Machine Learning (ICML’98), volume 98, pages 91–99. Brown, G., Wyatt, J., Harris, R., and Yao, X. (2005). Diversity creation meth- ods: a survey and categorisation. Information Fusion, 6(1):5–20. Celebi, M. E. and Kingravi, H. A. (2012). Deterministic initialization of the k-means algorithm using hierarchical clustering. International Journal of Pattern Recognition and Artificial Intelligence, 26(07):1250018. Deerwester, S. C., Dumais, S. T., Landauer, T. K., Furnas, G. W., and Harsh- man, R. A. (1990). Indexing by latent semantic analysis. Journal of the American Society of Information Science, 41(6):391–407. Garcı́a, S., Fernández, A., Luengo, J., and Herrera, F. (2010). Advanced nonparametric tests for multiple comparisons in the design of experiments in computational intelligence and data mining: Experimental analysis of power. Information Sciences, 180(10):2044–2064. Greene, D., Cagney, G., Krogan, N., and Cunningham, P. (2008). Ensemble Non-negative Matrix Factorization Methods for Clustering Protein-Protein Interactions. Bioinformatics, 24(15):1722–1728. Greene, D., O’Callaghan, D., and Cunningham, P. (2014). How Many Topics? Stability Analysis for Topic Models. In Proc. European Conference on Machine Learning (ECML’14), pages 498–513. Springer. Hadjitodorov, S. T., Kuncheva, L. I., and Todorova, L. P. (2006). Moderate diversity for better cluster ensembles. Information Fusion, 7(3):264–275. Huang, G., Li, Y., Pleiss, G., Liu, Z., Hopcroft, J. E., and Weinberger, K. Q. (2017). Snapshot ensembles: Train 1, get M for free. CoRR, abs/1704.00109. Kuang, D., Choo, J., and Park, H. (2015). Nonnegative Matrix Factorization for Interactive Topic Modeling and Document Clustering. In Partitional Clustering Algorithms, pages 215–243. Springer International Publishing. Kuhn, H. W. (1955). The hungarian method for the assignment problem. Naval Research Logistics Quaterly, 2:83–97. Kuncheva, L. I. and Hadjitodorov, S. T. (2004). Using diversity in cluster en- sembles. In Proc. IEEE International Conference on Systems, Man and Cybernetics, volume 2, pages 1214–1219. IEEE. Kuncheva, L. I. and Vetrov, D. P. (2006). Evaluation of stability of k-means cluster ensembles with respect to random initialization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(11):1798–1808. Lee, D. D. and Seung, H. S. (1999). Learning the parts of objects by non- negative matrix factorization. Nature, 401:788–91. Lin, C.-J. (2007). Projected Gradient Methods for Nonnegative Matrix Factor- ization. Neural Computation, 19(10):2756–2779. McCallum, A. K. (2002). Mallet: A machine learning for language toolkit. URL http://mallet.cs.umass.edu/ Minaei-Bidgoli, B., Topchy, A., and Punch, W. F. (2004). Ensembles of parti- tions via data resampling. In Proc. International Conference on Information Technology: Coding and Computing (ITCC’04). Newman, D., Lau, J. H., Grieser, K., and Baldwin, T. (2010). Automatic Eval- uation of Topic Coherence. In Proc. Annual Conference of the North Amer- ican Chapter of the Association for Computational Linguistics, HLT ’10, pages 100–108. O’Callaghan, D., Greene, D., Carthy, J., and Cunningham, P. (2015). An anal- ysis of the coherence of descriptors in topic modeling. Expert Systems with Applications (ESWA), 42(13):5645–5657. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., et al. (2011). Scikit- learn: Machine learning in Python. Journal of Machine Learning Research, 12(Oct):2825–2830. Pena, J. M., Lozano, J. A., and Larranaga, P. (1999). An empirical comparison of four initialization methods for the k-means algorithm. Pattern Recogni- tion Letters, 20(10):1027–1040. Pio, G., Ceci, M., Malerba, D., and D’Elia, D. (2015). ComiRNet: a web- based system for the analysis of miRNA-gene regulatory networks. BMC Bioinformatics, 16(9):S7. Steyvers, M. and Griffiths, T. (2007). Probabilistic topic models. Handbook of latent semantic analysis, 427(7):424–440. Strehl, A. and Ghosh, J. (2002). Cluster ensembles - a knowledge reuse frame- work for combining multiple partitions. Journal of Machine Learning Re- search, 3:583–617. Su, T. and Dy, J. (2004). A deterministic method for initializing k-means clus- tering. In Proc. 16th IEEE International Conference on Tools with Artificial Intelligence, pages 784–786. IEEE. Suh, S., Choo, J., Lee, J., and Reddy, C. K. (2016). L-EnsNMF: Boosted Local Topic Discovery via Ensemble of Nonnegative Matrix Factorization. In Proc. IEEE 16th International Conference on Data Mining. Topchy, A., Jain, A., and Punch, W. (2005). Clustering ensembles: Models of consensus and weak partitions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(12):1866–1881. Wild, S., Curry, J., and Dougherty, A. (2004). Improving non-negative ma- trix factorizations through structured initialization. Pattern Recognition, 12 37(11):2217–2232. Zhao, W., Chen, J. J., Perkins, R., Liu, Z., Ge, W., Ding, Y., and Zou, W. (2015). A heuristic approach to determine an appropriate number of topics in topic modeling. BMC Bioinformatics, 16(13):1. 13 
work_df4os2mi6zcshcikihumzaazfm	----	1471-2105-10-284.fm HAL Id: hal-02661323 https://hal.inrae.fr/hal-02661323 Submitted on 30 May 2020 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. CASSIOPE: an expert system for conserved regions searches Virginie Lopez Rascol, Anthony Levasseur, Olivier Chabrol, Simona Grusea, Philippe Gouret, Etienne Danchin, Pierre Pontarotti To cite this version: Virginie Lopez Rascol, Anthony Levasseur, Olivier Chabrol, Simona Grusea, Philippe Gouret, et al.. CASSIOPE: an expert system for conserved regions searches. BMC Bioinformatics, BioMed Central, 2009, pp.284. �hal-02661323� https://hal.inrae.fr/hal-02661323 https://hal.archives-ouvertes.fr BioMed CentralBMC Bioinformatics ss Open AcceResearch article CASSIOPE: An expert system for conserved regions searches Virginie Lopez Rascol†1, Anthony Levasseur*†2,3, Olivier Chabrol1, Simona Grusea1, Philippe Gouret1, Etienne GJ Danchin4,5,6 and Pierre Pontarotti1 Address: 1EBM UMR 6632.LATP, 3 place V. Hugo - 13 331 Marseille cedex 03 France, 2UMR-1163, INRA de Biotechnologie des Champignons Filamenteux, IFR86-BAIM. ESIL, 163 avenue de Luminy, CP 925, 13288 Marseille Cedex 09, France, 3Universités Aix-Marseille I & II, UMR-1163, case 925, 163 avenue de Luminy, 13288 Marseille Cedex 09, France, 4UMR 1301, INRA, F-06903 Sophia-Antipolis, France, 5UMR 6243, CNRS, F- 06903 Sophia-Antipolis, France and 6UMR 1301, UNSA, F-06903 Sophia-Antipolis, France Email: Virginie Lopez Rascol - virginie.lopez-rascol@univ-provence.fr; Anthony Levasseur* - anthony.levasseur@univ-provence.fr; Olivier Chabrol - Olivier.Chabrol@univ-provence.fr; Simona Grusea - gruseas@yahoo.com; Philippe Gouret - philippe.gouret@univ- provence.fr; Etienne GJ Danchin - etienne.danchin@sophia.inra.fr; Pierre Pontarotti - pierre.pontarotti@univ-provence.fr * Corresponding author †Equal contributors Abstract Background: Understanding genome evolution provides insight into biological mechanisms. For many years comparative genomics and analysis of conserved chromosomal regions have helped to unravel the mechanisms involved in genome evolution and their implications for the study of biological systems. Detection of conserved regions (descending from a common ancestor) not only helps clarify genome evolution but also makes it possible to identify quantitative trait loci (QTLs) and investigate gene function. The identification and comparison of conserved regions on a genome scale is computationally intensive, making process automation essential. Three key requirements are necessary: consideration of phylogeny to identify orthologs between multiple species, frequent updating of the annotation and panel of compared genomes and computation of statistical tests to assess the significance of identified conserved gene clusters. Results: We developed a modular system superimposed on a multi-agent framework, called CASSIOPE (Clever Agent System for Synteny Inheritance and Other Phenomena in Evolution). CASSIOPE automatically identifies statistically significant conserved regions between multiple genomes based on automated phylogenies and statistical testing. Conserved regions were searched for in 19 species and 1,561 hits were found. To our knowledge, CASSIOPE is the first system to date that integrates evolutionary biology-based concepts and fulfills all three key requirements stated above. All results are available at http://194.57.197.245/cassiopeWeb/ displayCluster?clusterId=1 Conclusion: CASSIOPE makes it possible to study conserved regions from a chosen query genetic region and to infer conserved gene clusters based on phylogenies and statistical tests assessing the significance of these conserved regions. Source code is freely available, please contact: Pierre.pontarotti@univ-provence.fr Published: 10 September 2009 BMC Bioinformatics 2009, 10:284 doi:10.1186/1471-2105-10-284 Received: 23 February 2009 Accepted: 10 September 2009 This article is available from: http://www.biomedcentral.com/1471-2105/10/284 © 2009 Rascol et al; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 1 of 10 (page number not for citation purposes) http://www.biomedcentral.com/1471-2105/10/284 http://creativecommons.org/licenses/by/2.0 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=19740451 http://194.57.197.245/cassiopeWeb/displayCluster?clusterId=1 http://194.57.197.245/cassiopeWeb/displayCluster?clusterId=1 http://www.biomedcentral.com/ http://www.biomedcentral.com/info/about/charter/ BMC Bioinformatics 2009, 10:284 http://www.biomedcentral.com/1471-2105/10/284 Background Comparative genomics and the reconstruction of ances- tral genomes provide landmarks better understand the biological rules governing evolution. The most obvious way to make progress in ancestral genome reconstruction is to compare the organizational structure of conserved genomic regions in a large number of informative species [1]. Hypotheses can then be formulated to account for such conserved genomic regions: • Observed conserved regions are due to chance and are not biologically significant. • Conserved regions result from a common ancestral region through inheritance. • Conserved regions are due to evolutionary conver- gence with possible selective pressure. CASSIOPE is able to reject the null hypothesis (conserved regions due to chance) in favor of one of the two alterna- tives, but cannot distinguish between them (ancestral regions or convergence). In literature reports, conserved regions are frequently defined through BLAST [2] or align- ment by similarity search to determine putative "ortholo- gous" genes [3,4]. Furthermore, the significance of the observed conserved gene clustering has to be statistically tested to reduce the risk of false positives. Several tools and databases (Phig's [5]), Ensembl [6]) provide informa- tion on conserved regions across different species but even when they do use phylogenetic methods, there is no sta- tistical processing assessing the significance of the con- served regions. In contrast, a few methods seek conserved regions using statistics [7,8] but do not offer a phylogeny-based distinc- tion between orthologs and paralogs. Thus, in [8], GRIMM-synteny computes conserved regions based on gene markers ("orthologous" sequences or "orthologous" alignments that users have to input) and a distance threshold. There are two requirements for identifying biologically significant conserved regions: • Identification of conserved markers (orthologs or paralogs between different species) • Identification of significantly conserved clusters of these markers Currently, there is no software available that detects con- served regions by providing a phylogenetic determination of conserved markers together with and a score for their significance. Furthermore, those resources that are availa- ble pre-compute conserved regions on a limited number of species, eliminating the possibility or running searches using custom-input regions. In short, biologists today find themselves needing one set of tools to identify orthologous or paralogous markers and then another set of tools to evaluate the significance of observed conserved regions. The lack of software able to provide automated output of statistically-estimated information on conserved regions at several-genome scale together with the growing amount of genomic data being filed prompted us to automate comparative analysis based on conserved genes clusters, through the CASSIOPE project. The CASSIOPE project proposes new methodology using evolutionary biology-based concepts. First, orthologs and paralogs are detected via phylogenetic analysis. Several approaches not based on phylogenetic analysis claim to find orthology. However, the clustering requires a com- plete genome and, in the case of lineage-specific differen- tial paralog loss, provides spurious data that contradict the identification of orthologs and paralogs based on phy- logeny. Secondly, chromosomal regions from different species that are inherited from a common ancestor have a higher probability of containing homologs than under neutrality. This probability has to be rigorously calculated to give a score on the evolutionary history of the species compared. These evolutionary concepts have been embedded in CASSIOPE. CASSIOPE deploys the following core technologies: • Data-processing system: the computer system is a modular system with several agents (virtual machines that work on specific tasks) deployed in conjunction with an expert system that communicates with every agent and takes rule-based decisions to answer initial biological questions. The rulesets of the expert system can be updated, removed or added, just as a human scientist would. • Data flexibility: searches can be run for newly- sequenced regions or genomes. The comparative data is initially pooled and computed, and then recom- puted when saved data is older than one month. • Detection of orthologous genes by robust phyloge- netic reconstruction. • Statistical score to assess significance of conserved regions. • Reverse-search feature making it possible to extend the initial searches. Page 2 of 10 (page number not for citation purposes) BMC Bioinformatics 2009, 10:284 http://www.biomedcentral.com/1471-2105/10/284 The stability and reliability of CASSIOPE were determined using several tests on large genomic regions (containing several hundred genes). Methods As previously stated, the search for biologically-relevant conserved genomic regions requires phylogenetic orthol- ogy assessment, statistical testing, and as many available genomes as possible. The breakthrough offered by CASSI- OPE is that it integrates all three of these key steps in a sin- gle automated process: • Phylogeny: orthologous/paralogous genes are deter- mined by phylogenetic methods (using Figenix, [9]). Phylogenetic information also allows reconstruction of the evolutionary history, and therefore more accu- rate ancestral genome reconstruction. • Statistical test: we apply a statistical test for each con- served region to assess the significance of the observed conserved gene clusters [for a detailed description see [10]], taking multigenic families and paralogs into account. • Reverse search: when conserved regions between a region A1 of species A and a region B1 of species B are found, the reverse conserved regions are subsequently checked, i.e. after finding conserved regions from A1 to B1, a reverse search is performed from B1 in order to screen for the conserved region in species A. This step will either find part of the same region A1 or expand this region or a different new region (whether or not it is on the same chromosome). This reciprocal process thereby finds multi-directional conserved regions from all species to all species (all-against-all), enables to confirm conserved regions, boundary these regions, and highlight translocations, duplications and other evolutionary phenomena at chromosome level. Orthologous conserved regions Here we describe the algorithm used in CASSIOPE with, as starting material, specific genomic region Ra in species A located on chromosome Ca: (1) For each gene from Ra, we search for orthologous genes in multiple species using phylogenetic methods. If {Ga1,..., Gan} are genes of Ra, then for each Gai 1 <i <n, we obtain: {Os1,..., Osp} a list of orthologous genes of gene i in sev- eral species sj 1 <j <p (2) Once phylogenetic reconstruction of all genes from this region Ra has allowed the determination of orthologous genes, we classify them as a function of i. species ii. chromosome iii. number of orthologs on this chromosome ≥ 3. Thus, several clusters in different species can be obtained. (3) We complete all clusters by including the non- orthologous genes found inside clusters delimited in step (2). (4) We assess statistical support for conserved regions between start region and each new region found in previous steps using a "binomial test" [10] We test the null hypothesis H0 and the alternative hypoth- esis H1: H0: Conserved genes clustering between start region and identified end region results from a random process. H1: Conserved genes clustering between start region and orthologous end region does not result from a random process. p: probability of observing a number k of orthologous genes (randomly pulled out of n genes) all mapped to the same region. Therefore, p is the probability that each gene contained in the start region will have one (or several) orthologs in the end region. q: probability that orthologs are localized elsewhere in the target genome (q = 1-p). If Ra is start region and Rb a putative "target" conserved region found in species B, then: p = number of genes in Rb/genes in genome B. n = number of genes in species A which have at least one orthologous gene in species B (found by phylogeny) ∀ ∈ = = −k P X k C p qn k k n kΩ, ( ) α = − =∑1 0 p X k k ( ) Page 3 of 10 (page number not for citation purposes) BMC Bioinformatics 2009, 10:284 http://www.biomedcentral.com/1471-2105/10/284 k = number of genes in Rb that are orthologous to genes in Ra. If a given gene from Ra has multiple orthologous genes in species B (due to duplication after speciation of A and B), the probability of finding an orthologous gene in Rb is higher than for one-to-one orthologous genes. Conse- quently, in the case of multiple orthologs, k is corrected as follows: If gi, 1 <i <n is a gene in Ra And if li = number of orthologous genes (found from gi) in the species B genome And if mi = number of othologous genes (found from gi) that fall in region Rb Then (5) Reverse search: we restart the algorithm with all the newly found regions as starting regions using the following logical rules: unless the new region is included in or overlaps a region that has been already studied, we restart with this new region. The algorithm stops when all the conserved sites have been calculated. Paralogous conserved regions CASSIOPE is also able to identify paralogous conserved regions (conserved genes clusters in the same genome) exactly in the same manner as it finds orthologous con- served regions. Working with a fixed time-window on duplication events and with phylogenetic trees, paralogous genes are detected by finding matching duplication nodes, i.e. nodes that correspond to the defined time-window. For instance, the search for paralogous genes that appear after duplication of vertebrate nodes eliminates duplication nodes that contain species other than vertebrates. Paralogous genes are then clustered and "non-matching" clusters (less than three genes) are removed. For each cluster, conserved regions are computed. The sta- tistical test is the same as for orthologous regions. Where: p = number of genes in Rb/genes in B genome n = number of genes in species A that are at least one par- alog in species B (found by phylogeny) k = number of paralogous genes in Rb Results and Discussion System Expansion of the panel of genomes available for compar- ison will allow us to construct higher resolution models of genome evolution. However, the vast amounts of data involved make it impossible to manually identify all con- served regions among a large number of species. Another key requirement is regular updates on conserved region data, as genome assembly and annotation can be refined over time as new genomes are released. CASSIOPE is a modular system superimposed on a multi-agent frame- work. The expert system controls three slaves. This is a centralized system where one of the "agents" has a global view of the process and drives the non-intelligent slaves (Figure 1). Each virtual machine has a specific task and all the machines work together to address user queries. The process developed in CASSIOPE (Figure 2) involves sev- eral tasks, such as phylogenetic reconstruction or consult- ing web databases, and each task is independent from the others. The expert system contains a full ruleset allowing decision-taking on information received. The whole proc- ess is contained in the expert system and uses other agents to obtain the information required. a) Agents Expert System This is the core of CASSIOPE. The Expert System commu- nicates with different agents to answer the following ques- tion: which genomic regions are significantly conserved? It receives queries and tries to find the required informa- tion in database. If information is incomplete, unavaila- ble or outdated (> 1 month), the system deduces the questions it has to ask to the other agents. One example rule, taken from reverse search mode, con- cerns when to restart with new region R. If the question for R has already been solved: R is included in another solved region or R overlaps another solved region then the Expert System does not restart with R k mi lii = ⎢ ⎣ ⎢ ⎢ ⎥ ⎦ ⎥ ⎥∑ ∀ ∈ = = −k P X k C p qn k k n kΩ, ( ) α = − =∑1 0 p X k k ( ) Page 4 of 10 (page number not for citation purposes) BMC Bioinformatics 2009, 10:284 http://www.biomedcentral.com/1471-2105/10/284 else it restarts with R. Each time it registers conserved regions, an e-mail is sent to the user indicating the species-pairs involved and the associated significance score. At the same time, the expert system creates reports for each cluster found, indicating the different steps performed and the elements used for deductions. Tree Agent CASSIOPE uses the FIGENIX computer platform [9] to assess phylogenetic relationships (orthology/paralogy). FIGENIX allows phylogenetic trees to be built via specific pipelines. In CASSIOPE, the CassiopePhyloM pipeline is used to reconstruct the desired trees. From the query sequence, a dataset of putative homologous sequences is first constructed using BLASTp using protein sequences from ENSEMBL. CASSIOPE filters the raw dataset to elim- inate potentially non-homologous sequences (E-value threshold: 10-4), disturbing alignments, and duplicates. The next step uses CLUSTALW to produce an alignment that is then modified to eliminate large gaps. Since phylo- genetic analysis is achieved at domain level, we next detect these domains with HMMPFAM. For each domain align- ment (extracted from the original alignment), a bias cor- rection phase is run to eliminate: i) non-monophyletic "repeats", ii) sequences with divergent composition, which is done using the amino-acid composition test in TREE-PUZZLE software (with an alpha risk set to 5%), and iii) sites not under neutral evolution SHIFT-FINDER. Indeed phylogenetic reconstruction methods are not tol- erant to sites highly divergent to neutral evolution and molecular clock. Sites not respecting this rule potentially produce errors in trees' reconstruction and thus have to be masked. Once the domains have been "purified", and after congruent domain selection using the HOMPART test in the PAUP package, a new alignment is built by merging preserved parts of the domains' alignments. This new alignment is then used to generate three phylogenetic trees using NJ, ML (with TREE-PUZZLE) and MP (with PAUP package) methods. By comparing the topologies of CASSIOPE multi-agent system, showing all the agents and the communications (blue) between them, together with non-system elementsFigure 1 CASSIOPE multi-agent system, showing all the agents and the communications (blue) between them, together with non-system elements. Pink: Expert System; Violet: persistence agent and Postgres database; Green: tree agent and FIGENIX platform; Yellow: Web agent. Page 5 of 10 (page number not for citation purposes) BMC Bioinformatics 2009, 10:284 http://www.biomedcentral.com/1471-2105/10/284 Page 6 of 10 (page number not for citation purposes) The diagram depicts how processes and agents are overlaidFigure 2 The diagram depicts how processes and agents are overlaid. Agents are pink, green and yellow, as represented in Fig 1. Arrows represent communication channels between the expert system and the other agents. The persistence agent is not represented as it is used throughout the process: (input). The user gives the two boundary genes of the target region - (1) The region is completed by all genes present in the Ensembl database - (2) The phylogenetic tree is computed for each gene - (3) Orthologous genes are calculated (forester) - (4) Each orthologous gene is located on the chromosome in its species - (5) Orthologous genes are clusterized if they are in the same chromosome and the same species. If the cluster contains fewer than three genes, it is removed - (6) Clusters are completed with genes contained in the region and that are not orthologous genes - (7) A score is calculated for each cluster: if the conserved site is not significant, then the cluster is removed - (8) The system restarts with new regions. BMC Bioinformatics 2009, 10:284 http://www.biomedcentral.com/1471-2105/10/284 these trees with the PSCORE command ("Templeton win- ning sites" test) from the PAUP package and the KISHINO-HASEGAWA test from the TREE-PUZZLE pack- age, these trees are fusioned to produce a unique consen- sus tree. This consensus protein tree can then be compared with a reference species tree (the tree of life from NCBI) to deduce proteins orthologous to the query sequence [see Additional file 1]. FIGENIX is accessible online http://figenix2.up.univ- mrs.fr/Figenix/index.jsp; however, the Tree Agent runs tasks directly by querying FIGENIX without interface. Database Agent All features are registered in a local database (managed with postgresql http://www.postgresql.org). Each time the Expert System searches for information, it asks the Data- base Agent whether this information already exists and whether it is out of date, i.e. > 30 days old. For instance, when phylogenic trees are screened for a given gene G, the database agent checks whether G is already included in orthologs list of another gene, taking the FIGENIX task identifier and asking FIGENIX to send the corresponding gene tree (this avoids re-computing existing trees). This database can also be consulted using a "telescope viewer" interface (see section 3.2) that allows user-friendly visualization of results. Web Agent We have developed an agent that searches for information on remote sources on the Web. As a source of genomic and proteomic sequence data, we chose the Ensembl [6] database as it is a non-redundant and frequently updated database allowing retrieval of gene locations on different chromosomes at the base pair (bp) level, and that can be accessed via an API from a JAVA library (ENSJ). b) Telescope viewer CASSIOPE enables users to view all the conserved sites registered in the database. Whenever CASSIOPE receives new conserved region queries, the results are saved and made visible using the telescope viewer. This viewer has three levels: • Chromosome level: for a given species, a chromo- some can be chosen showing regions that are con- served in other species and their identifiers. • Region level: by clicking on "chromosome region" in the last level, a region on the chromosome appears in front of boxes. Each box represents conserved regions on each chromosome of each species. It allows a glo- bal view of species sharing the conserved regions and of their distribution in each species. • Gene level: by clicking on a box in other species, con- served sites appear between two regions with all the genes of the two regions together with homologous between-gene links, making it possible to explore gene distribution in both regions. Ensembl link and FIGE- NIX task identifier are available for each gene. c) CASSIOPE Parameters CASSIOPE allows users to modify several of the parame- ters used at different steps in order to fine-tune the soft- ware: CASSIOPE. global parameters • Completed: Boolean parameter -Either the region is defined by both extreme genes (Web Agent completes it) or the region is defined by user, who give his own boundary genes region • Scope: this parameter is used to focus on a species subset (using TAX ID) • "SourceURI": to select the databases to be used by the Web Agent (at the present time, the software can only use Ensembl as a data source, but adaption to other databases is possible) • Ortholog or paralog region search: to determine type of conserved sites researched by the user • Reverse search: to restart the process • Range: to choose time-windows for duplication • Distance between two genes: the orthologous clus- ters found are sometimes large with long non-orthol- ogous genes gaps but can still be significantly conserved. To specify conserved region, it is possible to split the orthologous cluster if the distance between two orthologous genes is >x bps. Phylogeny parameters FIGENIX is Web-interfaced with configurable phylogeny parameters. Therefore, working through the CASSIOPE interface, users can configure FIGENIX at the same time, with: • Pipeline model: FIGENIX contains several pipelines for phylogenetic reconstruction, any of which can be selected by the user. (We recommend using__CassiopePhyloM__ as it is the most up-to-date and appropriate pipeline for reliably finding orthologs and paralogs). Page 7 of 10 (page number not for citation purposes) http://figenix2.up.univ-mrs.fr/Figenix/index.jsp http://figenix2.up.univ-mrs.fr/Figenix/index.jsp http://www.postgresql.org BMC Bioinformatics 2009, 10:284 http://www.biomedcentral.com/1471-2105/10/284 • Scope: to focus on a species subset or eliminate cer- tain species from the phylogenetic analysis (which could be different from the CASSIOPE scope). • "NoDuplicationRange": allows sequences that belong to a given taxonomic group to be considered as orthologs. For instance, in phylogenies using individ- ual genes, human and dog sometimes appear more closely related than human and mouse, and some- times human and mouse are closer than human and dog; to prevent a duplication node between these three species being systematically deduced, we use the NoDuplicationRange parameter by indicating [9615, 10090], which are the NCBI Taxonomic Identifiers (TaxIDs) for dog and mouse, respectively. • Database: database used for the initial homology search (Ensembl) • Tree of life: FIGENIX needs a reference tree of life to infer duplication and speciation nodes on tree fusions. By default, the NCBI tree of life is used, but users can set their own species tree. Case Study In this section, we describe an example analysis using CASSIOPE on a test-region that was previously known to have conserved regions in vertebrate species. Biological data The example region chosen contains 283 genes and is localized to chromosome 9 of the human genome from base pair 129,045,207 (ENSG00000136895: Ensembl gene identifier) to base pair 140,191,570 (ENSG00000159247). This region is an MHC-like paralo- gous regions [11,12]. It provides an interesting test-case as it is known to be conserved through vertebrate evolution. Conserved regions were searched for in all of the 19 fol- lowing species: Mammals: Homo sapiens, Monodelphis domestica, Bos taurus, Mus musculus, Rattus norvegicus, Canis lupus familiaris, Equus caballus, Pongo abelii, Pan troglodytes, Macaca mulatta Birds: Gallus gallus Teleost fish: Takifugu rubripes, Danio rerio, Tetraodon nigro- viridis, Oryzias latipes Insects: Drosophila melanogaster, Anopheles gambiae Nematodes:Caenorhabditis elegans Yeast: Saccharomyces cerevisiae Computing features The parameters used in this case-study are reported in Table 1. CASSIOPE was run on a bi-processor machine with 2 Gb of RAM. All the agents ran on the same computer. How- ever, as this software is based upon modular architecture, agents could be distributed over several computers. Output (results) CASSIOPE ran for 11 days and found 1,561 conserved sites in all 19 species using reverse-search parameters. CASSIOPE launched the computation of 3,736 phyloge- nies via FIGENIX. Most of the execution time was spent computing phylogenetic trees, which is well-known as a time-intensive step. In order to considerably accelerate the process, CASSIOPE could also run a fast computation by using only NJ method. From the MHC-like paralogous region on human chromosome 9 (the start region of the process), 37 conserved regions were found (Figure 3). CASSIOPE then automatically computed conserved regions in all 19 species [see Additional file 1, figure s6], leading to a total of 1,561 hits. Comparing the results obtained from Ensembl data with the results obtained from the CASSIOPE process, CASSIOPE found the same regions and/or more regions for some species, as well as new conserved regions with species that were not com- puted in Ensembl databases. Detailed results are given in figure 3. The Ensembl results are available at: http:// www.ensembl.org/Homo_sapiens/ cytoview?l=9:129045207-140191570;h=, syntenic regions link. • For instance CASSIOPE and Ensembl found the same regions for Bos taurus, Canis lupus familiaris, Equus caballus, Monodelphis domestica, Mus musculus, Rattus norvegicus Pan troglodytes, Pongo abelii and Gallus gallus. • In the case of Macaca mulatta CASSIOPE found one supplementary region compared to Ensembl. • For the others species, only CASSIOPE identified conserved regions. It seems clear that the conserved regions in Ensembl are not yet fully computed. Furthermore, CASSIOPE also statistically assessed the sig- nificance of conserved regions, providing scores. We clearly identified a teleost-specific duplication, as was previously proposed [13]: corresponding to one region on one chromosome, we obtained several conserved regions on two chromosomes. For one chromosome in vertebrate genomes, there are two chromosomes in teleost genomes. Using the "telescope viewer", exploration of all regions Page 8 of 10 (page number not for citation purposes) http://www.ensembl.org/Homo_sapiens/cytoview?l=9:129045207-140191570;h= http://www.ensembl.org/Homo_sapiens/cytoview?l=9:129045207-140191570;h= http://www.ensembl.org/Homo_sapiens/cytoview?l=9:129045207-140191570;h= BMC Bioinformatics 2009, 10:284 http://www.biomedcentral.com/1471-2105/10/284 Page 9 of 10 (page number not for citation purposes) Table 1: Global parameters CASSIOPE global parameters Completed Yes Scope No Source URI http://ensembldb.ensembl.org Ortholog/paralog Ortholog Reverse search Yes Range All Phylogeny parameters Pipeline model __CassiopePhylo+M__ Scope No NoDuplication Range [9615, 10090] Database Ensembl Tree of life cassiope1 [see Additional file 1, figure s4] Cluster parameters Distance between 2 genes 10,000,000 bps Conserved regions from a human MHC-like region (283 genes)Figure 3 Conserved regions from a human MHC-like region (283 genes). Each box represents species possessing one or sev- eral conserved regions with the start region. Regions are represented on the corresponding chromosome in the species. The right-hand-side boxes show Ensembl results (regions are delimited by red boxes), and the left-hand-side boxes show CASSI- OPE results. If Ensembl results are lacking, the box is left empty. CASSIOPE found more conserved regions than Ensembl. The presence of duplication in teleost fish genomes is shown. http://ensembldb.ensembl.org BMC Bioinformatics 2009, 10:284 http://www.biomedcentral.com/1471-2105/10/284 Publish with BioMed Central and every scientist can read your work free of charge "BioMed Central will be the most significant development for disseminating the results of biomedical researc h in our lifetime." Sir Paul Nurse, Cancer Research UK Your research papers will be: available free of charge to the entire biomedical community peer reviewed and published immediately upon acceptance cited in PubMed and archived on PubMed Central yours — you keep the copyright Submit your manuscript here: http://www.biomedcentral.com/info/publishing_adv.asp BioMedcentral highlights each gene with its detected orthologous genes, phylogenies and links to the Ensembl database (for fur- ther information on genes and access to the actual sequence, [see Additional file 1]). Conclusion CASSIOPE is a reliable and flexible tool that provides access to up-to-date information based on the compari- son of multiple genomes. It allows the study of conserved regions starting from a specific query genetic region. Infer- ence of conserved gene clusters is based on phylogenetic reconstruction, which adds an historical dimension, and on statistical assessment of the significance of the infer- ence. Finally, the built-in telescope viewer makes it easy for users to explore the results, and promptly gives an idea on the conserved regions and gene organization of each pair of species compared. CASSIOPE could be extended to other applications. In the test case reported here, CASSI- OPE was only used for a global view of identified con- served regions. The CASSIOPE project is expected to give insights on genome structure and evolution and to be use- fully applied in biomedical and agricultural fields (e.g., identification of QTLs or disease-gene identification). For instance, pairing information regarding the conservation of genomic regions with functional information regarding diseases could point to candidates for genes involved in pathologies. Authors' contributions AL and VLR contributed equally to this work. AL, EGJD and PP designed and integrated the biological and evolu- tionary concepts in the system. VLR and OC carried out the software design. PG developed the first informatic sys- tem. OC and VLR have developed the final version of CASSIOPE (including structural improvements) on a soft- ware architecture proposed by PG. AL and PP wrote the manuscript. All the authors read and approved the final manuscript. Authors declare to have no competing finan- cial or other interest in relation to this work. Additional material Acknowledgements This work was supported by the ANR EvolHHuPro and the ANR E-TRI- CEL. Authors are grateful to Thomas Sorger for critical review of the man- uscript. We wish to thank the reviewers for their interesting suggestions that helped to improve the quality of the manuscript. References 1. Lopez Rascol V, Pontarotti P, Levasseur A: Ancestral animal genomes reconstruction. Curr Opin Immunol 2007, 19:542-546. 2. Altschul SF, Gish W, Miller W, Myers EW, Lipman DJ: Basic local alignment search tool. J Mol Bio 1990, 215:403-410. 3. Brudno M, Malde S, Poliakov A, Do CB, Couronne O, Dubchak I, Bat- zoglou S: Global alignment: Finding rearrangements during alignment. Bioinformatics 2003, 19:i54-62. 4. Darling ACE, Mau B, Blattner FR, Perna NT: Mauve: Multiple align- ment of conserved genomic sequence with rearrangements. Genome Research 2004, 14:1394-1403. 5. Dehal PS, Boore JL: A phylogenomic gene cluster resource: the Phylogenetically Inferred Groups (PhIGs) database. BMC Bio- informatics 2006, 7:201. 6. Hubbard TJP, Aken BL, Ayling S, et al.: Ensembl 2009. Nucleic Acids Res 2009, 37:D690-D697 [http://www.ensembl.org]. 7. Hoberman R, Sankoff D, Durand D: The statistical significance of max-gap clusters. Comparative Genomics 2004:55-71. 8. Bourque G, Pevzner PA, Tesler G: Reconstructing the genomic architecture of ancestral mammals: lessons from human, mouse, and rat genomes. Genome Research 2004, 14:507-516. 9. Gouret P, Vitiello V, Balandraud N, Gilles A, Pontarotti P, Danchin EGJ: FIGENIX: Intelligent automation of genomic annota- tion: expertise integration in a new software platform. BMC Bioinformatic 2005, 6:198 [http://figenix2.up.univ-mrs.fr/Figenix/ index.jsp]. 10. Danchin EGJ, Pontarotti P: Statistical evidence for a more than 800-million-year-old evolutionarily conserved genomic region in our genome. J Mol Evol 2004, 59:587-597. 11. Abi-Rached L, Gilles A, Shiina T, Pontarotti P, Inoko H: Evidence of en bloc duplication in vertebrate genomes. Nat Genet 2002, 31:100-105. 12. Vienne A, Shiina T, Abi-Rached L, Danchin E, Vitiello V, Cartault F, Inoko H, Pontarotti P: Evolution of the proto-MHC ancestral region: more evidence for the plesiomorphic organisation of human chromosome 9q34 region. Immunogenetics 2003, 55:429-436. 13. Jaillon O, Aury JM, Brunet F, et al.: Genome duplication in the tel- eost fish Tetraodon nigroviridis reveals the early vertebrate proto-karyotype. Nature 2004, 431:946-957. Additional file 1 User manual and detailed results provided by CASSIOPE. The data provided describe the CASSIOPE web page and results obtained by using CASSIOPE Click here for file [http://www.biomedcentral.com/content/supplementary/1471- 2105-10-284-S1.doc] Page 10 of 10 (page number not for citation purposes) http://www.biomedcentral.com/content/supplementary/1471-2105-10-284-S1.doc http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=17702562 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=17702562 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=12855437 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=12855437 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15231754 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15231754 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=16608522 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=16608522 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=19033362 http://www.ensembl.org http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15059991 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15059991 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15059991 http://figenix2.up.univ-mrs.fr/Figenix/index.jsp http://figenix2.up.univ-mrs.fr/Figenix/index.jsp http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15693615 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15693615 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15693615 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=11967531 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=11967531 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=14530884 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=14530884 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=14530884 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15496914 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15496914 http://www.biomedcentral.com/ http://www.biomedcentral.com/info/publishing_adv.asp http://www.biomedcentral.com/ Abstract Background Results Conclusion Background Methods Orthologous conserved regions Paralogous conserved regions Results and Discussion System a) Agents Expert System Tree Agent Database Agent Web Agent b) Telescope viewer c) CASSIOPE Parameters CASSIOPE. global parameters Phylogeny parameters Case Study Biological data Computing features Output (results) Conclusion Authors' contributions Additional material Acknowledgements References 
work_dgj4alh4ljbztkeddln4v3m72u	----	Improving direct mail targeting through customer response modeling Expert Systems With Applications 42 (2015) 8403–8412 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Improving direct mail targeting through customer response modeling Kristof Coussement a,∗, Paul Harrigan b,1, Dries F. Benoit c,2 a IESEG School of Management – Université Catholique de Lille (LEM, UMR CNRS 9221), Department of Marketing, 3 Rue de la Digue, F-59000 Lille, France b The University of Western Australia – UWA Business School, M263, 35 Stirling Highway, Crawley, 6009, Australia c Faculty of Economics and Business Administration, Ghent University, Tweekerkenstraat 2, B-9000 Ghent, Belgium a r t i c l e i n f o Keywords: Direct marketing Direct mail Response modeling Database marketing a b s t r a c t Direct marketing is an important tool in the promotion mix of companies, amongst which direct mailing is crucial. One approach to improve direct mail targeting is response modeling, i.e. a predictive modeling approach that assigns future response probabilities to customers based on their history with the company. The contributions to the response modeling literature are three-fold. First, we introduce well-known statisti- cal and data-mining classification techniques (logistic regression, linear and quadratic discriminant analysis, naïve Bayes, neural networks, decision trees, including CHAID, CART and C4.5, and the k-NN algorithm) to the direct marketing community. Second, we run a predictive benchmarking study using the above classifiers on four real-life direct marketing datasets. The 10-fold cross-validated area under the receiver operating char- acteristics curve is used as evaluation metric. Third, we give managerial insights that facilitate the classifier choice based on the trade-off between interpretability and predictive performance of the classifier. The find- ings of the benchmark study show that data-mining algorithms (CHAID, CART and neural networks) perform well on this test bed, followed by simplistic statistical classifiers like logistic regression and linear discrimi- nant analysis. It is shown that quadratic discriminant analysis, naïve Bayes, C4.5 and the k-NN algorithm yield poor performance. © 2015 Elsevier Ltd. All rights reserved. 1. Introduction The move from mass-marketing to mass-customization is no bet- ter reflected than in the area of direct marketing, and in particular direct mail. Marketers no longer distribute their messages to a mass market, nor do they distribute based on basic demographic character- istics; rather they distribute and optimize different messages to dif- ferent segments that are developed based on past behavior (Jonker, Piersma, & Van den Poel, 2004; Rowe, 1989; Wierich & Zielke, 2014). Still, the need to improve the effectiveness of direct mail campaigns is a persistent issue in many industries (Guido, Prete, Miraglia, & De Mare, 2011; Mahdiloo, Noorizadeh, & FarzipoorSaen, 2014). Before sending direct mail, a key dilemma for marketers is which customers to target. In an effort to answer this question, marketers tend to use response modeling. Response modeling identifies cus- tomers that are likely to respond better to the marketing campaign based on their past response behavior. ∗ Corresponding author. Tel.: +33 320545892. E-mail addresses: k.coussement@ieseg.fr (K. Coussement), paul.harrigan@uwa.edu.au (P. Harrigan), dries.benoit@ugent.be (D.F. Benoit). 1 Tel.: +61 8 6488 1979. 2 Tel.: +32 9 264 3552. The above perfectly fits in the philosophy underpinning one-to- one marketing communications seen in the customer relationship management (CRM) domain (Mahdiloo et al., 2014). CRM is a strategic approach to marketing underpinned by relationship marketing the- ory (Morgan & Hunt, 1994), which has been defined as “a compre- hensive strategy and process that enables an organization to iden- tify, acquire, retain and nurture profitable customers by building and maintaining long-term relationships with them” (Sin, Tse, & Yim, 2005, p. 1266). At the heart of CRM is data on customers. The in- creasing power of CRM technologies enables more and more sophis- ticated data collection, storage and analysis techniques. The ability to draw powerful analyses from customer data makes CRM – and thus response modeling – a critical success factor in today’s rapidly chang- ing environment (Danaher & Rossiter, 2011; Kumar, 2008; Ngai, Xiu, & Chau, 2009). The focus of this paper is on customer response modeling. The contributions of our research study are three-fold. First, we will in- troduce the most popular response modeling methods to the di- rect marketing community. In particular, we review a range of popular classification algorithms borrowed from the statistical and data-mining community (logistic regression, linear and quadratic discriminant analysis, naïve Bayes, neural networks, decision trees (CHAID, CART and C4.5) and the k-NN algorithm). Second, we com- plement the existing response modeling literature by integrating and http://dx.doi.org/10.1016/j.eswa.2015.06.054 0957-4174/© 2015 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2015.06.054 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.06.054&domain=pdf mailto:k.coussement@ieseg.fr mailto:paul.harrigan@uwa.edu.au mailto:dries.benoit@ugent.be http://dx.doi.org/10.1016/j.eswa.2015.06.054 8404 K. Coussement et al. / Expert Systems With Applications 42 (2015) 8403–8412 contrasting all classification algorithms into a framework that aims to benchmark their predictive capabilities in discriminating responders from non-responders in four real-life direct mail companies. Third, managerial insights on classifier choice are given to the direct mar- keting community taken into account the comprehensibility and pre- dictive performance of the response model. This paper is structured as follows. The next section introduces the direct marketing field and its links with customer response modeling. Following that the range of classification algorithms are introduced and explained. We then describe the evaluation met- ric used and further explain the characteristics of the datasets and the experimental setting. Finally, we present the results and their implications. 2. Direct marketing and response modeling Direct marketing is defined as the ‘interactive system of marketing which uses one or more advertising media to affect a measurable re- sponse and/or transaction at any location’ (Direct Marketing Associa- tion 2009). Direct marketing is big business. It is projected that direct marketing expenditures in the US will grow to $196 billion in 2016, with direct mail forming part of this growth. Direct mail is targeted at customers that are most likely to be enticed by particular offers, as opposed to a traditional mass marketing approach whose promo- tional activities are addressed to customers and prospects indistinctly (Guido et al., 2011; Mahdiloo et al., 2014; Risselada et al. 2014). Direct mail is not being killed off by the Internet; rather it is being used as a complementary channel (Danaher & Rossiter, 2011). Winterberry Group confirms that direct mail is still on the rise (Conlon, 2015). In 2014, direct mail spending grew with 2.7% in the United States compared to the projected 1.1% growth. Moreover, the market ana- lysts project a 1% growth increase in direct mail spending for 2015, equivalent to $45.7 billion of the $156.8 billion representing the total direct and digital spending projection for 2015. The reason by Win- terberry group is that direct mail costs will stay steady, and thus they expected that the projected 1% growth to come from volume increases. Continued growth will be predicated upon the levels of return on investment of direct mail campaigns, which significantly depends on marketers being able to use specialized targeting techniques to come up with the right set of customers to contact (Lamb, Hair, & McDaniel, 1994). Thus, the importance in knowing which customers are more likely to respond to a certain mailing is of paramount importance to mar- keters. Determining or predicting those customers who have a high probability to respond to a specific mailing based on their past be- havior is called the customer response modeling (Bose & Chen, 2009; Mahdiloo et al., 2014). Response modeling is part of the classification literature stream. Classification is the procedure where customers are predicted to belong to predefined groups or target classes based on their historical customer information (Blattberg, Kim, & Neslin, 2008). Typically, a response model is estimated on a training set in which both the independent variables, describing and profiling a par- ticular customer, and the dependent (response) variable, whether the customer responded on a certain mailing, are observed. Then, the es- timated model on the training data is applied to a new set of cus- tomers that are not used during training (the test set). The result is a response probability for each customer in the test set, dependent on his or her past behavior. Managerially speaking, depending on the direct mail campaign budget, the company is able to target the top x% of customers with the highest response probability given by the response model. The next section of this paper will introduce and describe the range of statistical and data-mining algorithms that can be used in customer response modeling. 3. Classification algorithms The essence of one-to-one marketing communication is provid- ing the right customers with marketing messages that they can eas- ily act on (Ryals, 2005) This means that ‘prediction and targeting are both key to decision making underlying direct marketing campaigns’ (Zahavi & Levin, 1997, p.35). Therefore, understanding which tech- niques yield the best predictive capabilities is vital for direct mar- keters (Bose & Chen, 2009; Rada, 2005). With increased efficiencies and effectiveness, marketers could reduce mailing costs (Barwise & Farley, 2005), increase conversion rates (Kaefer, Heilman, & Ramenof- sky, 2005), and increase customer retention (Watjatrakul & Drennan, 2005). Our literature review reveals that existing literature utilizes sev- eral statistical and data-mining classification algorithms in various research setups to separate responders from non-responders. How- ever, we complement the academic literature by presenting and in- tegrating the most popular classifiers into one predictive bench- mark study over multiple response datasets, while summarizing the managerial implications for managers. Several statistical classifica- tion methods to predict customer responses have been proposed and utilized, such as logistic regression, discriminant analysis and naïve Bayes (Baesens, Viaene, Van den Poel, Vanthienen, & Dedene, 2002; Berger & Magliozzi, 1992; Coussement, Van den Bossche, & De Bock, 2014; Cui, Wong, & Zhang, 2010; Deichmann, Eshghi, Haughton, Sayek, & Teebagy, 2002; Kang, Cho, & MacLachlan, 2012; Lee, Shin, Hwang, Cho, & MacLachlan, 2010). These techniques can be very pow- erful, but each algorithm also makes several stringent, but different, assumptions on the underlying distribution between the indepen- dent variables and the dependent variable. To counter this, more ad- vanced data-mining algorithms have been proposed for discriminat- ing between responders and non-responders, such as artificial neu- ral networks (Baesens et al., 2002; Chen, Hsu, & Hsu, 2011; Curry & Moutinho, 1993; Zahavi & Levin, 1997), decision tree-generating tech- niques (Buckinx, Moons, Van den Poel, & Wets, 2004; Chen, Hsu, & Chu, 2012; Haughton & Oulabi, 1997; McCarty & Hastak, 2007; Rada, 2005) and k-NN learners (Govindarajan & Chandrasekaran, 2010; Kang et al., 2012). The following sections review the most popular response models by describing their functioning, and by discussing their merits and drawbacks. 3.1. Logistic regression Logistic regression (LOG) is a well-known and industry-standard classification technique for predicting a dichotomous dependent vari- able such as respond/do not respond to a mailing (Coussement et al., 2014; Suh, Noh, & Suh, 1999). Besides applications in direct marketing, it is an often used technique in a variety of predictive busi- ness settings like customer segmentation (McCarty & Hastak, 2007), churn prediction (Neslin, Gupta, Kamakura, Lu, & Mason, 2006), cus- tomer choice modeling (West, Brockett, & Golden, 1997) and many others. Moreover, logistic regression has several advantages (Hosmer & Lemeshow, 2000). For a given training set with N labeled training examples (xi,yi)} with i = 1, 2, … , N with input data xi є Rn and corresponding binary target labels yi є {0, 1}, the logistic regression tries to estimate the probability P(y = 1|x) given by P(y = 1|x) = 1 1 + exp(−(w0 + wx)) (1) with x є Rn being equal to an n-dimensional input vector, w to the pa- rameter vector and w0 to the intercept. The parameters w0 and w are usually estimated using a maximum likelihood procedure (Hosmer & Lemeshow, 2000). https://isiarticles.com/article/39889 
work_dizww225ubapncnapx4igjacrq	----	IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 69 - NARROWLY- BOUNDED TURF IRRIGATION SYSTEM SELECTION USING "EXPERT SYSTEM" APPROACH *Bedair, O. M. ABSTRACT Water scarcity in turf and landscape areas and increased water losses by using improper irrigation for narrow turf, or steep slopes cause design problems. Proper irrigation system selection for strips, islands, and areas near buildings, sidewalk, and steep areas is very important to obtain good turf quality, minimum operation, costs and water losses. The objective of this study includes an expert system (ES) approach to assist proper choice for narrowly-bounded turf strips to get the best appearance and quality for different site conditions for qualifying resources. Results were validated by consultation with domain experts and knowledge available from literature, published research and pertinent experimentation to accommodate case studies using prepared and modified decision table. Six case studies (four actual: A, B, C, and D, and two virtual sites: E and F) including extreme site conditions were used to test the proposed expert system for the proper selection of irrigation system from choices of irrigation systems. "Corvid Exsys" software program was designed to assist proper irrigation selection according to site conditions. The proposed ES and "Corvid Exsys" software were evaluated and tested on actual and virtual case studies, to assisted proper irrigation system selection for narrowly- bounded and sloped areas. The results of proposed ES on selected case studies and "Corvid Exsys" software outcomes were discussed. Results showed that the proposed software assist proper selection of irrigation system according to site conditions and resources including extreme conditions. The results showed that the subsurface drip irrigation system (SDI) gained the highest score by using case studies (B, C, D, and E) and in case study (A, and F), multi-stream spray gained the highest score compared with other systems. Keywords: Irrigation systems, Turf strips irrigation, Expert system, narrowly- bounded turf. *Lecturer, Ag. Eng. Dep., Fac. Ag., Ain Shams U., Cairo, Egypt. Misr J. Ag. Eng., 35 (1): 69 - 90 IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 70 - 1. INTRODUCTION n today's urban environments, there are many types of confined areas. They include a wide range of medians, islands or narrow strips of land near buildings, or sidewalks. A narrow strip of turf leads to significant quantities of water runoff, so careful planning and design can make these areas manageable. Awady, 2016, showed that Expert System (ES) approach can be efficiently used for best selection of appropriate system choice among different situations. Whittaker et al., 1987, reported that decision-making requires the use of expert knowledge; judgment and experience. Major steps and components that are involved in the complex process of building and developing of an (ES) were defined as: identification and software design; conceptualization formalization; implementation and validation of the program. Hillel, 2008, reported that the effects of wind, sprinkler overlap, and evapotranspiration can lead to application disuniformities which, in turn, can lead to excess water application. Drip irrigation has been used in horticultural operations since the middle of the 20 th century. There have been some investigations of the viability of using sub-surface drip irrigation (SDI) to irrigate turf grass ( Zoldoske et al., 1995; Leinauer and Makk, 2005; Johnson and Leinauer, 2004; Devitt and Miller, 1988; Ferguson, 1994) some of the benefits of SDI over conventional irrigation are that it operates at lower volumes and flow rates, puts water directly into the root zone, and is thus less susceptible to lessees from wind and evapotranspiration. Toro, 2006, reported that narrow or irregularly shaped areas, including turf, less than 8 feet in width in any direction, shall be irrigated with SDI or low volume water irrigation system. SDI saves water with minimal water loss due to mist, evaporation, runoff or wind drift. In addition, the amount of chemicals and fertilizer required to maintain the health of landscape is decreased, since these are applied directly at the root zone. 2. MATERIAL AND METHODS The aim of this study is to assist in proper irrigation system selection for narrowly- bounded turf strips based on E.S. I IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 71 - 2.1 Engineering and hydraulic characteristics of turf irrigation systems: The performance and required data of turf irrigation systems, under investigation, are presented in the following table (1): 2.2 Actual site conditions and hypothetical areas under investigation, including engineering and hydraulic criteria of irrigation system: Study areas were conducted in the Cairo governorate, Egypt. The latitude and longitude of the site are 30.0206857 N, 31.4419913 E for "Lake View" residential, 30.032517 N, 31.530409 E for "Mountain View" residential, 30.025772 N, 31.445462 E for "Villa", and 30.213209 N, 31.683102 E for "Stalla, Misr- El Gdida" residential district respectively. Fig. (1): Locations of actual cases under study. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 72 - Fig. (2): Actual narrowly-bounded strip and sloped areas photos. Table (2) shows site conditions and hypothetical areas under investigation including engineering and hydraulic criteria of irrigation systems. Four different representative sites and conditions including extreme cases with different resources and conditions, named case A, B, C, D (actual sites), and two case E, F(virtual sites) are shown in table(2). 2.3 Procedure for the selection of the proper turf irrigation system: Decision table was prepared using methodology of Awady, 2016, with a committee of irrigation consultants and expertise, irrigation engineers and technical irrigation workers, to illustrate system choices and qualifier conditions to assist proper turf irrigation system selection for narrow strip areas. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 73 - *Inhabited compound with turf green areas scattered among runway strips and buildings. The areas differ from as wide as 620000m 2 to as narrow as 1.5-3.5m bounded small areas of only 6200m 2 . # According to Central Laboratory of Agricultural Climate (CLAC). Each case study had scores of confidence for each system, which reflect the suitability to the circumstances, table (3). Committee of five irrigation experts, four irrigation engineers in addition to three technicians and four project owners carried out consultations during seven meetings for about two hours/each to put an appropriate decision table for the proper selection of turf irrigation system, according to site conditions. The derived decision table is validated in actual and virtual case studies to test, IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 74 - and compare and agreement with the outcomes from derived decision table for all site conditions including extreme site conditions. "Corvid Exsys" software program was used during consultant's committee ES, as tool for designers and technical users according to site conditions and resources in narrowly bounded, steep sloped areas. Allotted weight of each qualifier was suggested according to experts' judgment by sorting qualifiers based on their importance and effectiveness on irrigation system selection. Each qualifier was given weight based on its effect on irrigation system selection among affected qualifier. This weight was multiplied times score for particular irrigation system (based on experts judgment) to get final score used to judge irrigation system appropriateness according to site conditions. The qualifiers are represented in table (3), with final scores of irrigation systems under investigation, as discussed in the following: 1. Strip width suitability: was classified into three categories named (narrow- moderate- and wide), SDI system proves to be the most proper system for narrow strip areas compared to other systems. The lowest score was given to circular spray system in narrow strips. 2. Feasibility: Feasibility was categorized into three levels, according to density of narrow strip areas within the project area. The levels are named (dense plot- moderate density, and sparse plot). The highest score of 1.125 was given to SDI at dense plot level, and the lowest score of 0.25 was given to circular spray at the same level. 3. Water saving: water saving had weight of 1.0 ( referring to experts judgment) reflecting the effect of this qualifier on system selection multiplied times the score weight of water saving qualifier ( according to experts opinion) to get the score shown in table.3 for irrigation systems under investigation, reflecting degree of system ability to save water. The more water saving gets highest score, and vice versa. For example, sub-surface drip (SDI) system had a score of 0.9 and circular spray system had a score of 0.45. 4. Area consolidation: was classified into three categories named (low- medium, and high) based on the ratios between small to project areas. The highest scores of 0.675, 0.675 were given to circular spray system and SDI system at high and low areas consolidation category, respectively. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 75 - Table.3: Irrigation system scores based on experts judgment according to field conditions and allotted weights.  Allotted weight (according to experts' judgment), Max.= 6.0 *Favorability of watering pattern to strip area. 5. Water quality: Clear water was given an equivalent score for all systems under choice. Low quality water was given the highest score for the highest clogging resistant systems. 6. Land slope: was classified based on land leveling into three categories, named (even land- moderate level – and uneven land). The IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 76 - highest score was given to low precipitation rate system to eliminate water losses due to runoff and deep percolation, specially with increasing land slope. SDI got score of 0.475 and multi-stream got a score of 0.35 for even land. . 7. Wind losses: wind losses were classified into three categories, named (low- medium- and high) due to wind speed. Wind losses increased by increasing speed and vise versa. The system resisting wind (wind losses) got the highest score and vise verse. Subsurface drip irrigation (SDI) had a score of 0.4 at high wind losses (high wind speed), multi-stream system score was 0.22, or medium category. Fig.3: Qualifiers weight percent used for irrigation system appropriateness judgment. 2.4. Irrigation system choices and qualifiers: Irrigation systems under investigation included: 1. Circular Spray 2. Rectangular spray 3.Multi-stream 4. Sub surface drip. Score weights are given according to experts judgment as they affect irrigation- system selection among alternative actual conditions. Other virtual scores were collected to different choices according to different qualifiers. There assumption was based on experience and judgment of IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 77 - domain field engineering experts or extracted from literatures, such as (Omer et al., 2014). The highest irrigation system score gained from decision table was evaluated by experts to reach satisfaction results and agreement for outcomes to assist a proper selection according to site conditions. 2.5. Validation and study cases: Six case studies (four actual named A, B, C, and D in addition to two virtual sites named E and F) including extreme conditions were used to test the proposed ES for the proper selection of irrigation system from alternative choices of irrigation systems under investigation. Table 2 represents cases under study including conditions of each case. Cases were exposed to consultation with domain expert for validation of decision table. Each case of irrigation system was weighed under each suggested case. The manipulation of decision table was done using “Corvid Exsys” program, ver. 6.1.0 under windows 7 which was designed to include all qualifiers and site conditions, as shown in tables(3). The highest score represents the most system appropriateness for the case study. The planning of “Corvid Exsys” program is shown in the following steps, to assist irrigation engineers, irrigation technicians, and owners to choose the proper irrigation system. Fig. (4) represents a model structure for description of variables, key factors and qualifiers in order to determine the irrigation technology and its attributed techniques (choosing the best of irrigation system for turf) under diverse physical landscape resources and situations. 2.6. Verification, validations, and evaluation: The accuracy of proposed ES was tested through a sequence of steps to evaluate ES results as follows: 1. Collection of site date, parameters, and affecting factors required to select the proper irrigation system. 2. Validation was carried out to test ES logic results in particular case- study to satisfy the requirements of irrigation engineers and owners. 3. Procedure to compare ES results in different site conditions, including extreme conditions and determine the degree of confidence by using IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 78 - proposed ES selection of proper irrigation system for different case studies. Fig.4: Diagram of factors affecting selection of proper irrigation system according to site conditions. 3. RESULTS AND DISCUSSION 3.1 Proper irrigation system selection in different case studies: The results of testing decision table on selected case-study under investigation, in addition to software program-application on one case study, including software operation steps are discussed in the following: Figure (5) shows the weight score ratio for alternative choices of proper turf irrigation systems according to qualifier based on derived ES and actual site conditions named A, B, C, and D including wide variety of site conditions. The result shows that, in case "B", SDI system got the highest weight ratio of 4.53 with increases of 14.34 %, 54.4%, and 51.4% on Multi-stream, Rectangular spray and Circular spray, respectively, according to site conditions shown in table 2 in the "Material and Methods" section. It is clear that strip-shape factor is the most affecting factor on total score weight ratio gained for investigation, according to site condition. Case" A", illustrates weight ratio gained for irrigation systems under investigation. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 79 - The highest weight ratio was given to Multi-stream spray with increments of 12.9%, 14.2%, and 3.6 % for SDI, Rectangular spray, and Circular spray, respectively. It is clear that SDI and rectangular spray systems were given an approximately equal weight ratio. That is due to site conditions. The most important affecting factor on total score weight, given to irrigation systems under selection, can be arranged in descending order as follows: Strip width suitability; feasibility; water saving; area consolidation; water quality; land slope and wind losses. Meanwhile, in case "C", the highest weight ratio of 4.49 was given to SDI and the F ig . (5 ): E ffe c t o f a c tu a l site a re a s o n w e ig h t sc o re ra tio fo r d iffe re n t irrig a tio n sy ste m . IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 80 - minimum of 2.53 was given to rectangular spray, according to site conditions shown in table 2 in the "Material and Methods" section. It is clear that the most affecting factor for total weight ratio goes to strip width that adds about 43.7% of total score to SDI. Case "D", shows that the highest score of 5.32 goes to SDI with an increase of 21.8%, 51.9%, and 44.7% for Multi-stream, Rectangular spray and Circular spray, resp. according to site conditions represented in table 2 in the "Material and Methods" section. Qualifiers had equal effects on total weight ratio given to each system under investigation. Figure (6) shows the weight score ratio for alternative choices of proper turf irrigation systems, according to qualifiers based on derived ES and virtual site conditions named E and F including wide variety of site condition. F ig . (6 ): E ffe c t o f v irtu a l site a re a s o n w e ig h t sc o re ra tio fo r d iffe re n t irrig a tio n . sy ste m . IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 81 - The results show that: in case "E", the SDI system recorded the highest score of 4.16 with increases of 22.35%, 44.23%, and 34.13% over Multi- stream, Rectangular spray and Circular spray irrigation systems, resp., based on site conditions in table 2 in the "Material and Methods" section. It is clear that the most affecting factor on total weight ratio is with strip width with an average weight of about 31%. Meanwhile, case "F", Multi- stream system, recorded the highest score of 3.98 with increases of 4%, 14%, and 4.6% over SDI, rectangular spray, and circle spray irrigation systems, resp. The factors affecting total weight ratio for irrigation system under investigation had the same effect among irrigation systems due to site conditions. Figure (7) shows that by decreasing strip width gives SDI priority among irrigation systems under investigations. For wide strip, it is clear that, multi-stream system had the second weight ratio of 1.2 after SDI with weight ratio of 1.28. Moderate strip-width for SDI gets the highest weight ratio among other alternative systems. Figure (8) shows that for dense plot, SDI system gets the highest score, and in moderate and sparse cover, for circular spray gets the highest score. Figure (9) shows that SDI gets the highest weight ratio in low and medium categories of area consolidation. However, for high area consolidation category, Circular spray system got the highest score Fig. (7): Effect of suitability of strip width for irrigation systems under investigation on ES weight ratio. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 82 - Fig. (8): Effect of quality feasibility for irrigation systems under investigation on weight ratio. Fig. (9): Effect of area consolidation for irrigation systems under investigation on score weight ratio. Figure (10) shows that for the best water quality all irrigation systems under investigation had the same score weight ratio. For other water qualities, the highest score goes to Circular and Rectangular spray systems. Figure (11) shows that the highest score was given to SDI for all wind loss categories and the lowest score was given to the Rectangular spray for all wind categories. Figure (12) shows that the highest score was given to SDI for all land slope category and the lowest score was given to the Rectangular spray for all categories. It is clear that from fig. (12), that SDI gained the highest score for all land slopes and the lowest score was for Rectangular spray, for all land slopes. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 83 - Fig.(10):Effect of water quality on score weight ratio for irrigation systems under investigation. Fig. (11): Effect of wind losses on weight ratio for irrigation systems under investigation. Fig.(12): Effect of land slope on weight ratio for irrigation systems under investigation. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 84 - 3.2 ES to assist choices for turf irrigation: A new and simple ES was developed and tested to validate represented cases to test integrity of system. All cases were tested in the ES program. Results of representative cases are illustrated in the following. Fig. (13) shows the results of tested ES for case (C), to choose the proper irrigation system for site ( c) . Fig. (13): ES inputs to the choices of the irrigation system under the stated site- conditions. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 85 - Fig. (14): Printout of the ES program for choice of the irrigation system under site conditions. (Final output screen: The output takes in consideration water saving. Notice that the highest score revealed the proper irrigation system.) 4. CONCLUSION Irrigation system selection for strips, green islands, and areas near buildings, sidewalks, and sloped areas is very important to obtain a good turf quality, minimum operation, installation costs and water losses. All is for the choice of the proper irrigation system in narrowly- bounded turf strip. Hereby, the aim of this investigation was to build, verify and validate an ES program for the choice in narrowly-bounded turf strips to get the best appearance and quality for different site conditions for a set of qualifying resources.Six case studies (four actual: named A, B, C, and D, in addition to two virtual cases named E and F), including extreme site conditions, were used to test the proposed ES for the proper selection of irrigation system from alternative choices of irrigation under IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 86 - investigation.The results of proposed ES on selected case studies and "Corvid Exsys" software outcomes are discussed. Results show that the proposed ES and "Corvid Exsys" software can be efficiently used to assist proper selection of irrigation system according to site conditions and resources, including extreme conditions. Results obtained indicate the following:  The subsurface drip irrigation system saves water throughout the year as compared to using other irrigation systems in narrowly bounded turf strips.  The results showed that the output of the (ES) gave high score in sub surface drip irrigation system in some case studies (B, C, and D, E) and high score in Multi-stream spray system in other case studies (A, and F). 5. REFERENCES Awady, M. N., 2016. Computer applications in agricultural Engineering (In Arabic), Ch.7 on ES and their application in AE projects, Col. A. E., Azhar U., 86-100. Corvid Exsys, Inc., 2016, Knowledge automation expert system technology, © Copyright Exsys Inc. 2011-2016, 6565 America’s Parkway. NE Suite 200, Albuquerque, NM 87110 U.S.A.www.exsys.com Devitt, D.A. and W.W. Miller, 1988. Subsurface drip irrigation of bermudagrass with saline water.Appl. Agr. Res., Vol. 3, No. 4: 133- 143. Ferguson, K.R. 1994. Subsurface drip irrigation for turf. Proc. 1994 Int. Irrig. Show. Nov. 5-8. Hillel, D. 2008. 40 Years of drip irrigation. CSA News. 53, (9): 3-7. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 87 - Hunter, 2017. Product catalog. San Marcos, Calif.: Hunter Indust. Inc.: 48-74. Johnson, C and B. Leinauer, 2004. Effect of salinity level and irrigation type on the establishment rate and winter survival of turfgrass. http://www.environmentalturf.com/pdf/newmexico1.pdf. (Accessed, October 17, 2009). Atlanta, GA. Leinauer, B. and J. Makk, 2005. Effect of irrigation type and root zone material on irrigation efficiency, Turfgrass quality, and water use on putting greens in the Southwest. USGA Turfgrass and Env. Res. Summary. Omer, M. M.; El-Giendy, A. M. and Arafa, Y.E., 2014, Turf irrigation management under arid ecosystem condtions based on a developed expert systems, Misr. J. Agric. Eng., conf. (19):129-143. Rain Bird. 2017. Rain Bird Landscape Irrigation Products 2016-2017 Catalog. Azusa, Calif.: Rain Bird Co. Toro Solutions. 2006. Toro Company. Bloomington, MN. http://toro.com/watermgmt/brochures/solutions.pdf. (Accessed October 17, 2009). Whittaker, A. D.; D. D. Jones; R. H. Thieme and J. R. Barrett, 1987. Guidelines for getting started with expert systems, Agric. Eng., July/Aug.: 24-27. Zoldoske, D.F.; S. Genito and G.S. Jorgenson, 1995. Subsurface drip irrigation (SDI) on turfgrass: A U. Exp. Irri. Notes: C. for Irri. Tech., Fresno, CA. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 88 - 6. Appendix (I) Fig. (i) Program steps used to develop an ES to assist the choice of a proper irrigation system according to qualifier conditions. IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 89 - Fig.(i): Interface of “Corvid Exsys” program. العربيالملخص نظام خبير باضتخذامالمحذودة الخضراء لشرائح المططحاثنظام ري اختيار *أضامت محمذ أحمذ بذير نحذسةح الوسطحاث الخضشاء الضْقت ًالولششائوناسب الغْش اخخْاس نظام الشُ ّؤدٍ لشُالوْاه الوسخخذهت إجوالِ% هن 05الوْاه ّصل الَ فِالَ فقذ كبْش ًالجزس ةًساجالوً لزلك ّيذف البحث الَ بناء نظام خبشة باالسخعانت بخبشاء الوجال الوخخصصْن حلك الوساحاث . الوسطحاث سُحصوْن ًحنفْز ًحشغْل ًكزلك فنْْن حشكْب نظن ًهينذسِهن االسخشاسّْن ششائح الوسطحاث الخضشاء الضْقت سُاالخخْاس الوناسب لنظام فِللوساعذة ًالخضشاء .لظشًف الوٌقع ًالجزس ًفقا ًالونحذسة ًالوجاًسة مصر. -مذرش الهنذضت السراعيت، زراعت عين شمص، القاهرة* IRRIGATION AND DRAINAGE Misr J. Ag. Eng., January 2018 - 90 - ًالوسخطْل ًالذائشُبالشش الشرارٍ الشًُنظام السطحِبالخنقْط ححج الشُّعخبش نظام ششائح الوسطحاث الخضشاء سُ نظام خخْاسالالوخاحت كبذائل الشُهن أنظوت ًالشعاعِ ًالجزس.ًالونحذسة ًالوجاًسة الضْقت ًهٌقعْن A ًB ًC ًDىوا هٌاقع فعلْت 6حن حقْْن ًاخخباس نظام الخبشة الوقذم علَ الخبشاء. بآساءشولج كل الظشًف الشائعت ًالوخطشفت ًرلك باالسخعانت E ًFىوا افخشاضْْن لخسيْل Windows 7ّعول هن خالل ”Corvid Exsys“حن حطٌّش بشناهج حاسب الَ اسخخذام نظام الخبشة الوقذم ًحقْْوو ًاخخباسه بٌاسطت خبشاء الوجال للحصٌل علَ االخخْاس ًفقا للظشًف الحقلْت. الشُاالهثل لنظام رلك الظشًف فًِفقا لظشًف الوٌقع بوا الخبشة الوقذمًأظيشث النخائج هنطقْت نخائج نظام Aالوٌاقع فِ عالْتدسجت ثقت (SDI) السطحِخنقْط ححج بال الشُالوخطشفت، فقذ أعطَ نظام ًB ًC ًD الشعاعِبالشش الشُبْنوا أعطَ نظام (Multi-stream) أعلَ دسجت ثقت . E ًFللوٌاقع 
work_dmzaumeknjfqrivu467qln6fxu	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220983317 Params is empty 220983317 exception Params is empty 2021/04/06-03:05:25 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220983317 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:25 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_do6p5ndkibcaro722dg3dx77ry	----	SOUTHERN JOURNAL OF AGRICULTURAL ECONOMICS DECEMBER, 1989 XLAYER: AN EXPERT SYSTEM PROVIDING MANAGEMENT ADVICE TO COMMERCIAL LAYER MANAGERS Ed Schmisseur and John Pankratz Abstract Expert or knowledge-based system pro- XLAYER, an expert/knowledge-based mi- grams contain domain-specific knowledge and crocomputer program, was designed and de- use complex inferential reasoning to reach veloped to diagnose layer management prob- conclusions human experts would reach if faced lems and recommend expert remedial man- with a comparable problem (Hart; Weissand agement advice. The program also provokes Kulikowski). The basic knowledge contained management action by calculating the eco- in these programs consists of rules, facts, re- nomic loss attributed to major management lationships, reasons, and heuristics obtained problems. It analyzes data generated by a from human experts who can solve problems commercially marketed layer performance fi- in a particular domain of expertise. Expert nancial microcomputer program and has dem- systems embody techniques for solving prob- onstrated the ability to emulate poultry man- lems, manipulating stored knowledge, coping agement experts in the diagnoses of 80 indi- with uncertain information, and explaining how vidual layer management problems. The pro- an inference is proceeding or a conclusion has gram provides scarce expert poultry manage- been reached (Simons). ment advice to poultry layer managers regard- Expert or knowledge-ased system program less of size and scale of operation. developments are a new area of applied poul- try science research. The results of one effort to apply expert system technology to poultry Key words: expert system, knowledge-based layer management are reported here. system, rule-based, economic loss, layer management.l, lr m. PROBLEM-SOLVING FOUNDATIONS Expert or knowledge-based computer pro- The practical and empirically verifiable di- grams are becoming an important operations agnostic rules of poultry layer management and management component of many Fortune as practiced by a cooperating poultry manage- 500 companies (Simons; Herrod) and are being ment consulting firml served as the problem- actively explored to assist smaller businesses solving paradigm used in the XLAYER pro- (Harmon et al.). They are beginning to emerge gram. These domain-dependent methods were as a field of research, development, and use in developed by consulting firms and have proven commercial agriculture (McKinion and useful over a period of years through practical Lemmon; Michalski et al.; Boulanger; Roach consulting experience. Collectively, these diag- et al.; Fermanian et al.; Spahr and Puckett; nostic rules implicitly seek to maximize layer Levins and Varner; Oltjen et al.). They also flock profitability by identifying problems represent the next logical step in the progres- which impact production performance, costs, sion of computer-supported information and and returns. They utilize problem-solving tech- decision-aid systems for farm managers first niques and data manipulations which explic- initiated by agricultural researchers in the itly and rigorously critique these separate but early 1950s. (Schmisseur and Doluschitz). interrelated areas of layer management. I The management expertise contained in the XLAYER program represents that of Mel Gehman, President and Founder of Heritage PMS, Inc., RD #3 Box 458-B, Annville, Pennsylvania 17003. Ed Schmisseur is an Associate Professor, and John Pankratz is a Courtesy Appointment, Department of Agricultural and Resource Economics, Oregon State University. Approved for publication as Technical Paper No. 8907 of the Oregon Agricultural Experiment Station. This project was financed in part by the Department of Agricultural and Resource Economics and the Oregon Agricultural Experi- ment Station, Oregon State University and Heritage PMS, Inc., of Annville, PA. Copyright 1989, Southern Agricultural Economics Association. 183 The general problem-solving approach em- financial performance, and egg gradeout. Flock bodied in the XLAYER program is relatively performance data contain livability, egg pro- straightforward. It attempts confirmation of duction, and nutrition information such as mor- all management problems contained in the know- tality, body weight, house temperature, egg ledge base by matching a list of management case weight, shell thickness, metabolizable problem symptoms with indicators of substan- energy, and protein, and the content of spe- dard flock performance. When a critical num- cific amino acids in the layer ration. Financial ber of matches occurs, a management problem performance indicators report itemized income is confirmed and a management recommen- and expense categories and net profit. Egg dation is issued. The search for management gradeout data show egg size and grade, qual- problems begins with the evaluation of pro- ity, quantity, and price information. duction decisions followed by decisions about The XLAYER program uses approximately egg marketing, and then production costs. 100 different data observations generated by The diagnostic rules, general problem- the Layer Performance program. These data, solving approach, and search strategy embod- automatically stored in a file called "Record," ied in the XLAYER program were solicited represent both actual layer performance and from the cooperating consultants. A series of equivalent performance potentials or stan- interviews with these experts dards. Performance potentials or standards are 1) identified important management problems based on standards published by commercial impacting layer flock profitability; genetic companies and breeder farms. They 2) dictated the data and the heuristics needed reflect the different layer strains, age of birds, to diagnose these problems; molting phase, and environmental conditions. 3) established the analytical or step-by-step Other data required by the XLAYER pro- procedure followed to determine problems; and gram, but not produced by Layer Perform- 4) ascertained what management recommen- ance, include information on pullet flock his- dation would be issued. tory, layer house equipment, marketing ar- After each interview, new insights about the rangements and prices, and criteria by which diagnostic process were encoded in the flock performance are compared to perform- XLAYER program and then tested using test ance standards. Pullet history information in- case data. In subsequent interviews, previ- cludes pullet weights, feed consumption, uni- ously discovered problem-solving rules were formity and shank index information at se- confirmed, new problem-solving capabilities lected weeks of rearing, and general pullet were added, and the depth of the program's rearing conditions. This information is manu- diagnostic capabilities was increased. ally entered and permanently stored in a file called "Flock Profile" when a pullet flock starts DATA REQUIREMENTS production. The XLAYER program was developed to Information about layer house equipment analyze weekly layer flock performance data includes the type of feeding, egg gathering, generated by a commercially-marketed poul- and manure handling system in the layer try management microcomputer program, house. These data, also manually entered, are Layer Performance. 2 This integration was permanently stored in a file called "Housing/ important since the consulting firm developed Equipment Profile." Layer Performance to collect and calculate Information about marketing and prices in- data useful in their diagnoses of layer man- eludes egg selling/processing arrangements, agement problems. Furthermore, many of egg prices by egg grade or case weight, and XLAYER's data needs could be obtained di- feed ingredient prices. These data are speci- rectly from data files produced by Layer Per- fled in a file called "Price Profile." Price infor- formance rather than through additional key- mation is manually updated weekly. board entries. The various criteria by which flock perform- The major types of data produced by Layer ance are compared to flock performance stan- Performance and utilized by the XLAYER dards are contained in a file called "Evalu- program are illustrated in Figure 1. The Layer ative Criteria." These criteria were obtained Performance program produces a report which from the cooperating consulting firm. They are provides weekly data on three major areas of based on experience in analyzing flock per- poultry management-flock performance, formance records. 2Layer Performance is a registered trademark of Heritage PMS, Inc., RD #3 Box 458-B, Annville, Pennsylvania 17003. 184 Layer Report Flock Egg Performance Gradeout Egg Income Profit Size/Price Production Grade Livability Nutrition Expenses Quality Figure 1. Heritage's Layer Performance Report. PROGRAM OVERVIEW infer management problems based on other known information. At each query point, us- The XLAYER program: ers can ask the XLAYER program to explain 1) identifies major management problems its reasons for the query. Also, at the end of judged to be impacting layer flock profitability; the consultation, users can query the program 2) calculates associated economic losses at- to explain how it arrived at its conclusions. tributed to major management problems; and After the XLAYER program has completed 3) issues specific management recommenda- its analysis, an "Executive Management Re- tions designed to eliminate the management port" is compiled. This report includes a brief problem. It is written using a commercially- description of each management problem, its marketed expert development and delivery associated economic loss to provoke manage- shell microcomputer program, M.1.3 The ment action and facilitate quick partial bud- XLAYER program' s code was written in the geting of the decision recommendation, and "PROLOG"-like production rule language con- specific management recommendations. If no tained in the M.1 program. The shell program management problems exist, the XLAYER requires an IBM-PC compatible microcom- program generates a brief report indicating puter equipped with a graphics card and a this finding. minimum of 320K RAM. The XLAYER program's relationship to its DIAGNOSTIC FRAMEWORK data needs and users is illustrated in Figure 2. Users execute and communicate with the A generalization of XLAYER's diagnostic XLAYER program by interacting with M.1. framework is presented in Figures 3 through The M.1 program automatically calls the 5. These diagnostics are contained in some 400 XLAYER program and immediately begins individual production rules. Diagnostics per- the consultation. During consultation, the tain to egg production, egg blend price, vari- XLAYER program automatically seeks rele- able production costs, and fixed costs. The di- vant flock performance, performance stan- agnostic framework, or fault tree, illustrates dards, and other needed data. Should data not the complexity of the inference process. It also be found, the natural language user interface shows the path of impact of management of M.1 queries the user. Users then must re- decisions made in each specific area of pro- spond. A response of "unknown" is acceptable duction. as the XLAYER program has the ability to Management decisions about layer nutrition Layer Performance Report <-1 aRecords & Standards ^ —Pro |Flock Profile U se r— Housing/Equipment Profile M.1 M. 4- |Price Profile XLAYER[_ ItXLAYER IEvaluative Criteria A- I' Questions I . | Management Report Figure 2. XLAYER's Relationship to Users and Data. 3 M.1 is a registered trademark of TEKNOWLEDGE, Inc., P.O. Box 10119, Palo Alto, California 94303. 186 depict one of the complex inferences within plementation problem can manifest itself in the XLAYER program. Nutritional decisions several performance indicators. impact egg production directly and also indi- When a methionine problem is confirmed, rectly through feed consumption (Figure 3). another XLAYER rule is invoked which actu- But, nutrition decisions also indirectly impact ally issues the management alert and pre- egg blend price through both small eggs and scribes a remedy. This rule is: egg quality (Figure 4) and indirectly impact variable production cost through feed costs f methionine and display ([Verify suspected (Figure 5). Specifically, egg quality is directly methionine imbalance in the layer ration. affected by shell quality, which in turn is im- Methionine intake should be within 20 per- pacted by nutrition decisions, and feed cost is cent of your flock's potential intake. If con- directly affected by feed consumption, which firmed reformulate the rations protein and in turn is impacted by nutrition. Similar com- methionine content. plex inferences exist for management decisions then problem. affecting environmental conditions in the layer However, before the primary rule can be house, disease, pullet quality, and stress. confirmed, either low feed consumption and The complexity of the diagnostics contained low production or small egg sizes must also be in the XLAYER program is more easily seen confirmed from other production rules con- in its diagnostic rules. For example, the pri- tained in the XLAYER program. mary rule confirming a methionine supplemen- The low feed consumption, low production, and tation problem exhibits a complex inference. small egg sizes diagnostic rules are relatively This rule and others presented also provide simple. The low feed consumption rule is: insights into the basic analytical process or problem-solving approach employed in the iffeedperhenay = FPHD XLAYER program. The rule, taken directly and goalfeedperhenday = GFPHD from the XLAYER code, is: and critical_feed_perhen_day = CFPHD and FPHD < GFPHD * CFPHD if ((low_feed_consumption and low_production) then low_feed_consumption. or (low_feed_consumption and small_egg_sizes ois unkn( onsumption and smThis rule confirms low feed consumption whenis unknown) and methionineperhenday = MPHD the amount of feed consumed per hen day is and goalmethionineperhenday = GMPHD less than the goal or standard for feed con- and uppertwidebounds p UW B sumption per hen day adjusted downward by and louer wide bounds = LWB coefficient of tolerance for feed consumption. and MPHD > GMPHD * UWB The low production rule is: or MPHD < GMPHD * LWB) if eggs_per_hen_housed = EPHH or (small_egg_sizes and goal_eggs_perhen_housed = GEPHH and methionine_per_hen_day = MPHD and critical_eggs_per_hen_housed = CEPHH and goal_methionine_per_hen_day = GMPHD and EPHH < GEPHH * CEPHH and upper_moderate_bounds = UMB and egg_blend_price = EBP and lower_moderate_bounds = LMB and starting_number_hens = SNH and MPHD > GMPHD * UMB and (((GEPHH - EPHH) * SNH)/12) * EBP = or MPHD < GMPHD * LMB) PRODL then methionine. and display( [' The rule's interpretation is: Economic losses attributed to low produc-The rule's interpretation is:The rule '£ mterpretat ' .11n tion are: ',tab(6), $(PRODL)] )1) if low feed consumption and low produc- i r tion exists and the methionine level fed per then lowproducton. hen day exceeds either an upper or lower It confirms low production when egg produc- bound, tion per hen housed is less than the standard 2) or small egg sizes exists without either for production adjusted downward by a coeffi- low feed consumption or low production and cient of tolerance for egg production, and it the methionine level fed per hen day ex- calculates the associated economic loss. This ceeds either a narrower upper or lower loss is based on the difference between the bound, egg production standard and actual egg pro- then a methionine supplementation problem duction per hen housed adjusted by the total is confirmed. It shows that a methionine sup- number of hens housed. 187 Egg Production Quality Conditions Consumption 1 Disease Pullet Disease Fright Nutrition Parasites Quality Parasites Environmental Ration Mechanical Environmental Cannibalism Conditions Formulation Equipment Conditions Prolapse Figure 3. Egg Production Diagnostic Framework of XLAYER. Egg Blend Price Market Small Practices Egg Quality Eggs I Shell Interior Quality Quality Equipment Stress Disease Candling Stress Off Taste Damage Error Discoloring mnvroma ImIn1ral [7Handling & EnvironmentalNutrition Disease Equipment Conditions Absorption Problems Low Feed Parasites Pullet Figure 4. Egg Blend Price Diagnostic Consumption Nutrition Diseases Quality Framework of XLAYER. High Variable Cost Medication High Energy Miscellaneous Labor Wage Labor Wage Costs Costs Feed Costs Costs Rate Hours Costs Feed Ingredient Consumption Costs i Interest & RepairI | + | +LDepreciation Costs Environmental Feather Conditions Wasteage Loss Nutritional Theft Figure 5. Variable and Fixed Cost Diagnostic Framework of XLAYER. The small egg sizes rule is: in new grains gradually even if the cost if caseweight = CW per pound is higher. Gradually move to and goal_caseweight = GCW the lower cost grain substitute.'] ) and critical_case_weight = CCW then problem. and CW < GCW * CCW and CaverageWblendrice = ABP Notedly absent from the diagnostic frame- and dozen_eggs_graded = DEG work are interrelationships between such fac- and potential blend_price = PBP tors as nutrition, disease, and stress. Although and (PBP - ABP) * DEG = SMALLEGGS these interrelationships are theoretically pos- and display( [' sible, the cooperating consultants did not re- Economic loss attributed to small egg sizes quire analyzing these type of interrelationships is approximately,',tab(4), $(SMALL in their consulting work and did not include EGGS), per week.'] ) them in the XLAYER program. They argued then small_egg_sizes. in today's skillfully managed commercial poul- try flocks, management problems attributed This rule confirms small egg sizes when egg to these type of interrelationships are rare. case weight is less than a case weight stan- dard adjusted downward by a tolerance level PROGRAM OUTPUTS for egg case weight. It also calculates the as- T X p sociated economic loss due to small egg sizes . The XLAYER program possesses the abil based on the difference between the egg blend t to dagnose ultiple management problems price received and the potential egg blend andrecommendmanagementactions formore price multiplied by the number of eggs graded than 80 individual layer production manage- Another type of diagnostic rule contained in ment problems. This capability is exhibited in the XLAYER program reflects a complex in- a management report displayed on screen and/the XLAYER program reflects a complex in- or printed. The report identifies management ference requiring the user to respond to a prte. The reort idifis management query about current production conditions. A problems, their associated economic loss, and simple example of this type of rule is: specifc management recommendations. An example management report appears in Fig- if (low_feed_consumption and low_production) ure 6. Specific management problems and rec- or (low_feed_consumption and ommendations related to housing and equip- small_egg_sizes) ment management, nutrition, diseases, eco- and grain_change = yes nomics, marketing practices, and general then change_grain. management practices are contained in the As this rule is being confirmed and the pre- program. conditions of either low feed consumption and Approximately 37 percent of the XLAYER low production or low feed consumption and program's diagnostics are about layer nutri- small egg sizes exist, the user is automatically tion Nutritional diagnostics include such queried about possible recent changes in the things as excessive or restrictive metabolizable type of grain used in the layer ration. The energy, calcium, and sodium supplementation, query involves two separate rule statements methionine, methionine-cystine, lysine, thre- onine, and trytophane amino acid imbalance, question(grain_change)=' improper vitamin and/or trace mineral mix Have you significantly changed the type supplementation, and poor egg shell quality of grain used in your layer ration?'. induced by improper ration formulation. legalvals(grain_change) = [yes,no]. Some 23 percent of the XLAYER program's The first of these two query rules issues a diagnostics directly address general manage- prespecified question to the user, while the ment practices. Improperly trained egg can- second provides acceptable answers and error dlers, a poorly managed poultry waste sys- checking capability to the query. tem, cage overcrowding or under-utilization, If the user responds to the questions by an- sudden changes in layer house environment, swering yes, another XLAYER rule is invoked excessive egg handling time, and egg or feed which states the management problem and of- theft represent some of the type of general fers a remedy. This rule is: management problems considered by the XLAYER program. if changegrain and display( [' Housing/equipment diagnostics account for A major change in the grain ration is about 18 percent of the XLAYER program's suspected to be causing production prob- management expertise. The XLAYER pro- lems. Reformulate your ration and phase gram identifies problems like improperly ad- 191 justed or operating mechanical egg gathering PROGRAM TESTING/VALIDATION and feeder equipment, high or low layer house The XLAYER program has been subject to temperature, low or high water consumption controlled testing with numerous test case due to malfunctioning waterers or poor well data. During initial case testing, the XLAYER water quality, limited lighting, and inadequate program's diagnostic capabilities were further ventilation. veno dti crlastion. 1 p enhanced, and its depth of knowledge was in- Economic diagnostics, representing 14 per- cresed. More than 300 individual cases, rep- cent of the XLAYER program's capability, resenting a flock's complete weekly data set, address traditional economic performance were subjected to analysis by the XLAYER problems. These include such things as high program. These cases, although carefully con- repair and maintenance, pullet rearing, labor, structed, confirmed that the program could interest, and energy costs. successfully diagnose each, as well as various Disease and marketing problems represent practical combinations, of the more than 80 the remaining diagnostic powers of the pro- layer management problems included in the gram. High broker fees, inappropriate class XLAYER program weight sales contract, respiratory diseases, Currently, it is being intensively field tested parasites, infectious coryza, fowl cholera, and with flock data from both the East and West MG disease are some of the disease and mar- coasts. In these tests weekly flock perform- keting diagnostics addressed by the XLAYER ane data are being analyzed by the XLAYER program. program and its diagnostics are compared to PROGRAM LIMITATIONS Pthose independently developed by the poultry management experts being emulated by the The XLAYER program has the potential to XLAYER program. In these limited field tests be a useful management tool, yet it has limita- with some 122,000 hens in various flock sizes, tions. Major limitations relate to the program's representing approximately 70 flock weeks of knowledge base. XLAYER's knowledge base production data, the XLAYER program has is static since the program does not possess exhibited the promising ability to perform on the ability to learn from experience. As a re- a level consistent with that of its counterpart sult, XLAYER's knowledge base must be con- poultry management experts. That is, in ap- tinually updated as new diagnostic techniques proximately 90 percent of the weekly tests, are developed and possibly new genetic bird the XLAYER program's diagnostics were strains are introduced. The magnitude of this identical to those independently prescribed by limitation is unknown. It may be mitigated the poultry management experts. because new knowledge can be added to ex- Additional field testing and more intensive pert system programs much more easily than validation with more participating flocks is modifying code in algorithmic programs. continuing. Field testing is scheduled to be XLAYER's knowledge base is also limited completed by the summer of 1990. to the diagnostic knowledge of the cooperat- ing consultants. No measure of their poultry PROGRAM AVAILABILITY management diagnostic skills has been made. The XLAYER program will be made avail- The magnitude of this limitation is also un- able to poultry managers by the cooperating known. consulting firm after extensive field testing is Economic losses attributed to low production are: $425 per week. Verify suspected high metabolizable energy level in layer ration. If confirmed, reformulate ration to reduce metabolizable energy level to your flock's standard recommended in the Layer Performance Financial Report. Also note a methionine-cystine imbalance in the layer ration. Methionine-cystine should be within 20 percent of your flock's potential intake. If confirmed, reformulate the ration's protein content and methionine-cystine supplementation. Water intake appears high. Figure 6. XLAYER's Management Report: An Example. 192 completed in the summer of 1990. The cost of gram indicates that it can substitute for scarce the program has not been established. Pro- and relatively expensive human, poultry layer gram code, excluding Layer Performance and management expertise. Its successful adoption M.1, can be obtained by contacting the authors and widespread use should increase the pro- of this article. duction efficiency of layer operations and im- rCONCLUSIONS prove producer profits. The XLAYER program also has shown that The present level of performance of the it can provide valuable management informa- XLAYER program suggests that expert sys- tion in a readily useable form to a range of tem programs have potential applied applica- flock sizes. Relative to economies of size is- tion in poultry layer management. Further- sues, this expert system program appears to more, it has been demonstrated that an ex- be size neutral. That is, it equally benefits dif- pert system program can be linked to a stand- ferent sizes of operation. If any unequal bene- alone agricultural, computerized decision- fit should occur, it is hypothesized benefits support program and can automatically and favor smaller operations because of the econo- routinely access information generated by this mies associated with making available scarce program to diagnose management problems and expensive human expertise in the more and provide insightful management advice. economical form of an expert system micro- The level of performance of the XLAYER pro- computer program. REFERENCES Boulanger, A.G. "The Expert System PLANT/cd: A Case Study in Applying the General Pur- pose Inference System ADVICE to Predicting Black Cutworm Damage in Corn." M.S. Thesis, Computer Sci. Dept., Univ. Ill., Champaign-Urbana, IL, 1983. Fermanian, T.W., R.S. Michalski, and B. Katz. "An Expert System to Assist Turfgrass Manag- ers in Weed Identification." In Proc. 1985 Summer Computer Conference. July 22-24, Chicago, IL, American Society of Agricultural Engineers, St. Joseph, MI, 1985. 499-502. Harmon, P., R. Maus, and W. Morrissey. Expert Systems Tools and Applications. New York: John Wiley and Sons Inc., 1988. Hart, A. Knowledge Acquisitionfor Expert Systems. New York: McGraw-Hill Book Co., 1986. Herrod, R.A. "Industrial Applications of Artificial Intelligence." In Proc. 1987 IEEE Western Conference on Computer Systems. June 2-4, Anaheim, CA, IEEE Computer Society Press, Washington, DC, 1987. 196-202. Levins, R.A., and M.A. Varner. "An Expert Diagnostic Aid for Reproduction Problems in Dairy Cattle." Comput. Electron. in Agric., 2(1987):47-56. McKinion, J.M., and H.E. Lemmon. "Expert Systems for Agriculture." Comput. Electron. in Agric., 1(1985):31-40. Michalski, R.S., J.H. Davis, V.S. Bisht, and J.B. Sinclair. "A Computer-based Advisory System for the Diagnosis of Soybean Diseases in Illinois." Plant Diseases, 67(1983):459-63. Oltjen, J.W., L.G. Burditt, and G.E. Selk. "Expert System for Culling Management of Beef Cows." Okla. Agric. Exp. Stn. Rpt., MP-119, 1987.85-88. Roach, J.W., R.S. Virkar, M.J. Weaver, and C.R. Drake. "POMME: A Computer-based Consul- tation System for Applied Orchard Management Using PROLOG." Expert Syst., 2(1985):56-69. Schmisseur, E., and R. Doluschitz. "Expert System Insights: Future Decision Tools for Farm Managers." J. Am. Soc. Farm Managers and Rural Appraisers, 51(1987):51-57. Simons, G.L. Expert Systems and Micros. Manchester, England: The National Computing Centre Limited, 1985. Spahr, S.L., and H.B. Puckett. "Electronics and Livestock Production." Illinois Review, 28(1986):22-25. Weiss, S.M., and C.A. Kulikowski. A Practical Guide to Designing Expert Systems. Totowa, NJ: Rowman and Allanheld, 1984. 193 194 
work_dodmms5bzrctzhd32bmag6ieti	----	Constrained Dynamic Rule Induction Learning Fadi Thabtaha, Issa Qabajehb, Francisco Chiclanac a. Applied Business and Computing, NMIT, Auckland, New Zealand b. School of Computer Sciences and Informatics, De Montfort University, Leicester, UK c Centre for Computational Intelligence, Faculty of Technology, De Montfort University, Leicester, UK Abstract One of the known classification approaches in data mining is rule induction (RI). RI algorithms such as PRISM usually produce If-Then classifiers, which have a comparable predictive performance to other traditional classification approaches such as decision trees and associative classification. Hence, these classifiers are favourable for carrying out decisions by users and hence they can be utilised as decision making tools. Nevertheless, RI methods, including PRISM and its successors, suffer from a number of drawbacks primarily the large number of rules derived. This can be a burden especially when the input data is largely dimensional. Therefore, pruning unnecessary rules becomes essential for the success of this type of classifiers. This article proposes a new RI algorithm that reduces the search space for candidate rules by early pruning any irrelevant items during the process of building the classifier. Whenever a rule is generated, our algorithm updates the candidate items frequency to reflect the discarded data examples associated with the rules derived. This makes items frequency dynamic rather static and ensures that irrelevant rules are deleted in preliminary stages when they don’t hold enough data representation. The major benefit will be a concise set of decision making rules that are easy to understand and controlled by the decision maker. The proposed algorithm has been implemented in WEKA (Waikato Environment for Knowledge Analysis) environment and hence it can now be utilised by different types of users such as managers, researchers, students and others. Experimental results using real data from the security domain as well as sixteen classification datasets from University of California Irvine (UCI) repository reveal that the proposed algorithm is competitive in regards to classification accuracy when compared to known RI algorithms. Moreover, the classifiers produced by our algorithm are smaller in size which increase their possible use in practical applications. Keywords: Classification, Data Mfining, Prediction, PRISM, Rule Induction, Online Security 1. Introduction Data mining, which is based on computing and mathematical sciences, is a common intelligent tool currently used by managers to perform key business decisions. Traditionally, data analysts used to spend a long time gathering data from multiple sources and little time was spent on analysis due to the limited computing resources. Though since the rapid development of computer networks and the hardware industry, analysts nowadays are spending more time on examining data, seeking useful concealed information. In fact, after the recent development of cloud computing, data collection, noise removal, data size and data location are no longer obstacles facing analysts. Data analysis or data mining is concerned about finding patterns from datasets that are useful for users, particularly managers, to perform planning (Thabtah and Hamoud, 2014). One of the known data mining tasks that involve forecasting class labels in previously unseen data based on classifiers learnt from training dataset is classification. Normally, classification is performed in two steps: Constructing a model often named the classifier from a training dataset, and then utilising the classifier to guess the value of the class of test data accurately. This type of learning is called supervised learning since while building the classifier the learning is guided toward the class label. Common applications for classification are medical diagnoses (Rameshkumar et al., 2013), phishing detection (Abdelhamid et al., 2014), etc. There have been many different classification approaches including decision trees (Witten and Frank, 2005), Neural Network (NN) (Mohammad et al., 2013), Support Vector Machine (SVM) (Cortes and Vapnik, 1995), Associative Classification (AC) (Thabtah, et al., 2004), rule induction (RI) (Holt, 1993) and others. The latter two approaches, i.e. AC and RI, extract classifiers which contain “If-Then” rules so this explains their wide spread applicability. However, there are differences between AC and RI especially in the way rules are induced as well as pruned. This article falls under the umbrella of RI research. PRISM is one of the RI techniques which was developed in (Cendrowska, 1987) and slightly enhanced by others, i.e. (Stahl and Bramer, 2008) (Elgibreen and Aksoy, 2013) (Stahl and Bramer, 2014). This algorithm employs separate-and-conquer strategy in knowledge discovery in which PRISM generates rules according to the class labels in the training dataset. Normally for a class, PRISM starts with an empty rule and keeps appending items to the rule’s body until this rule reaches zero error (Definition 8- Section 2.2). When this occurs, the rule gets induced and the training data samples connected with the rule are discarded. The algorithm continues building other rules in the same way until no more data connected with the current class can be found. At this point, the same steps are repeated for the next-in-line class until the training dataset becomes empty. One of the obvious problems associated with PRISM is the massive numbers of rules induced, which normally results in large size classifiers. This problem is attributed to the way PRISM induces the rules where it keeps adding items to the rule’s body until the rule becomes 100% accurate despite the low data coverage. In other words, PRISM does not mind inducing many specific rules, each covering a single data sample, rather producing a rule, say, with 90% accuracy covering 10 data samples. This excessive learning limits the applicability of PRISM as a decision making tool for managers in application domains and definitely overfits the training dataset. This is since managers normally prefer a summarised set of rules that they are able to control and comprehend rather a larger high maintenance set of rules. In fact, there should be a trade-off between the number of rules offered and the predictive accuracy performance of these rules. This paper investigates shortcomings associated with PRISM algorithm. Specifically, we look into three main issues: 1) The search space reduction: When constructing a rule for a particular class, PRISM has to evaluate the accuracy of all available items linked with that class in order to select the best one that can be added to the rule’s body. This necessitates large computations when the training data has many attribute values and can be a burden especially when several unnecessary computations are made for items that have low data representation (weak items). A frequency threshold that we call (freq) can be employed as a pre-pruning of items with low frequency. It prevents these items from being part of rules, and therefore the search space gets minimised. 2) PRISM only generates a rule when its error is zero, which may cause overfitting the training dataset. We want to derive good quality rules, not necessarily with 100% accuracy, to reduce overfitting and increase data coverage. We utilise rule’s strength parameter (Rule_Strength) to separate between acceptable and non-acceptable rules in our classifier. 3) When removing training data linked with a rule, we ensure that other items which have appeared in the removed data are updated. In particular, we amend the frequency of the impacted items. This indeed maintains the true weight of the items rather the computed frequency from the initial input dataset. In response to the above raised issues, we are developing in this article, a new dynamic learning method based on RI that we name enhanced Dynamic Rule Induction (eDRI). Our algorithm discovers the rule one by one per class and primarily uses a freq threshold to limit the search space for rules by discarding any items with insufficient data representation. For each rule, eDRI updates items frequency that appeared within the deleted training instances of the generated rule. This indeed gives a more realistic classifier with lower numbers of rules leading to a natural pruning of items during the rule discovery phase. Lastly, the proposed algorithm limits the use of the default class rule by generating rules with accuracy <100%. Often these rules are ignored by PRISM algorithm since they don’t hold zero error. These rules are only used during class prediction phase instead of the default class rule and when no 100% accuracy rule is able to classify a test data. This paper is structured as follows: Section 2 illustrates the classification problem and its main related definitions. Section 3 critically analyses PRISM and its successors, and Section 4 discusses the proposed algorithm and its related phases besides a comprehensive example that reveals eDRI’s insight. Section 5 is devoted to the data and the experimental results analysis, and finally, conclusions are provided in Section 6. Input: Training dataset T Output: A classifier that consists of If-Then rules Step 1: For each subset Ti in T that belong to wi Do Step 1.1 Make an empty rule, rj: If Empty then wi Step 1.2: Calculate Attx in Ti p(wi = i| Attx); Step 1.3: Append the Attx with the largest p(wi = i| Attx) to the body of rj Step 2: Repeat steps 1.1-1.3 until rj has 100% accuracy or no longer can be improved Step 2.1: Generate rj and insert it into the classifier Step 3: Discard all data examples from Ti that match rj’s body Step 4: Continue producing rules until Ti is empty or no rule with accepted error can be found Step 5: Repeat steps 1-4 until no more data subsets of can be wi found Step 6: IF(Ti does not contain any data examples of class wi ) Generate the classifier. Fig. 1 PRISM Pseudocode 2. The Classification Problem and Definitions Given a training dataset T, which has x distinct columns (attributes) Att1, Att2, … ,Attn, one of which is the class, i.e. cl. The cardinality of T is |T|. An attribute may be nominal which means it takes a value from a predefined set of values or continuous. Nominal attributes values are mapped to a set of positive integers whereas continuous attributes are preprocessed by discretising their values using any discretisation method. The aim is to make a classifier from T, e.g. clAttClassfiier → : , which forecasts the class of previously unseen dataset. Our classification method employs a user threshold called freq. This threshold serves as a fine line to distinguish strong items ruleitems<item, class> from weak ones based on their computed occurrences in T. Any ruleitem that its frequency passes the freq is called as a strong ruleitem, otherwise it is called weak ruleitem. Below are the main terms used and their definitions. Definition 1: An item is an attribute plus its values name denoted (Ai, ai). Definition 2: A training example in T is a row consisting of attribute values (Aj1, aj1), …, (Ajv, ajv), plus a class denoted by cj. Definition 3: A ruleitem r has the format<body, c>, where body is a set of disjoint items and cis a class value. Definition 4: The frequency threshold (freq) is a predefined threshold given by the end user. Definition 5: The body frequency (body_Freq) of a ruleitem r in T is the number of data examples in T that match r’s body. Definition 6: The frequency of a ruleitem r in T (ruleitem_freq) is the number of data examples in T that match r. Definition 7: A ruleitem r passes the freq threshold if, r’s |body_Freq|/ |D| ≥ freq. Such a ruleitem is said to be a strong ruleitem. Definition 8: A rule r expected accuracy is defined as |ruleitem_freq|/ |body_Freq|. Definition 9: A rule in our classifier is represented as: clbody → , where body is a set of disjoint attribute values and the consequent is a class value. The format of the rule is: 121 ... claaa n →∧∧∧ 3. Literature Review PRISM is a key algorithm for building classification models that contain simple yet effective easy to understand rules. This algorithm was developed in 1987 based on the concept of separate and conquer where data examples are separated using the available class labels. For each class (wi) data samples, an empty rule (If nothing then w1) is formed. The algorithm computes the frequency of each attribute value linked with that class and appends the attribute value with the largest frequency to the current rule’s body. PRISM terminates the process of a building a rule when that rule has an accuracy 100% according to definition (8). When the rule is generated, the algorithm continues producing the remaining possible rules from w1’ s data subset until the data subset becomes empty or no more attribute value can be found with an acceptable accuracy. At that point, PRISM moves to the second class data subset and repeats the same steps. The algorithm terminates when the training dataset is evaluated and when this happens the rules formed make the classifier. Figure 1 below depicts PRISM algorithm major steps. Below are the main pros and cons associated with PRISM algorithm based on the comprehensive review that we have done, PRISM Pros 1) The simplicity of rules generation in which only the rule’s accuracy parameter is computed to decide on the rule significance. 2) The classifier contains easy to understand rules which can empower decision makers particularly in domain applications that necessitate interpretations. 3) Easy implementation especially with the available computing resources. Actually, PRISM can be easily implemented in different versions: local, online, distributed and in environments that require parallelism. 4) The predictive power of its classifier can be seen as acceptable when contrasted to other classic data mining approaches such as search methods, decision trees, neural networks, associative classification and many others. PRISM Cons 1) In the original PRISM algorithm, there is no clear search space reduction mechanism for candidate items and therefore for large dimensional datasets, the expected numbers of candidate items can be huge. This may limits its use for certain applications. 2) Noisy datasets that contain incomplete attributes and missing values. This can been as a major challenge for PRISM since no clear mechanisms for handling noise is presented. Currently, adopted approaches from information theory are used to handle noise (Bramer, 2000) (Stahl and Bramer, 2014). 3) Handling numeric attributes. No build-in strategy is available in PRISM to discretise continuous attributes. 4) No clear rule pruning methodology is present in the original PRISM. This may lead to the discovery of large numbers of rules and lead to combinatorial explosion (Abdelhamid and Thabtah, 2014). There is a high demand on pruning methods to cut down the number of rules without hindering the overall performance of the classifiers. 5) Conflicting rule: There is no clear mechanism on how to resolve conflicting rules in PRISM. Currently the choice is random and favours class labels with the largest frequency linked with the item rather than keep multiple class labels per item. 6) Breaking tie among items frequency while building a rule: When two or more items have the same frequency, PRISM looks at their denominator in the expected accuracy formula. Yet, sometimes these items have similar accuracy and denominators which makes the choice random. Arbitrary selection without scientific justification can be seen as a biased decision and may not be useful for overall algorithm’s performance. One of the major challenges faced by PRISM algorithm is noisy training datasets. These are datasets that contain missing values, data redundancy, and incomplete data examples. A modified version has been developed called N- RISM to handle noisy datasets besides focusing on maximising the prediction accuracy (Bramer, 2000). Experimental tests using eight datasets from the UCI (Lichman, 2013) have showed consistent performance of N-PRISM when compared with classic PRISM particularly on classification accuracy obtained from noisy and noise-free datasets. It should be noted that the differences between the original PRISM and N-PRISM are very minimal. Modified PRISM methods were developed based on a previously design pre-pruning method called J- Pruning by (Bramer, 2002) (Stahl and Bramer, 2012). The purpose was to reduce overfitting the training dataset. J-Pruning is based on the information theory test performed typically in decision tree by measuring the significance of removing an attribute from the rule’s body. The algorithm of (Stahl and Bramer, 2012) was proposed to address rule pruning issue during the process of building rules for each class label. Experimental results using sixteen UCI datasets showed decrement on the number of items per rule. In 2015, (Othman and Bryant, 2015) have investigated instance reduction methods as a paradigm for rule pruning in RI. They claimed that minimising the training dataset to the subset needed to learn the rules is one way to shrink the search space. The authors have applied three common instance reduction methods namely DROPS, ENN and ALLKNN on limited number of datasets to evaluate their impact on the resulting classifiers. The limited results obtained can be seen as a promising direction for using instance reduction methods as pre-pruning phase in RI. The database coverage pruning (Liu et al., 1998) and its successors (Abdelhamid, et al., 2014) (Thabtah, et al., 2011) were one of the major breakthroughs in associative classification that can be employed successfully in RI as late or post pruning methods. These pruning methods necessitate a positive data coverage per rule with a certain accuracy so the rule can be part of the classifier. A new version of the database coverage pruning was developed by (Ayyat, et al., 2014) as a rule prediction measure. The authors have proposed a method that considers the rule rank as a measure of goodness besides the number of similar rules sharing items in their body. These two parameters play a critical role in differentiating among available rules especially in predicting the test data class labels. Recently, (Elgibreen and Aksoy, 2013) investigated a common problem in PRISM and RULES (Pham and Soroka, 2007) family of algorithms which is the tradeoff between training time and classification accuracy during constructing the rule based classifiers. Moreover, the same article also highlighted performance deterioration of RI methods when applied to incomplete datasets. The result is a new covering algorithm that utilises “Transfer Knowledge” approach to fill in missing attribute values specifically the target attribute in the training dataset before mining kicks in. The “Transfer Knowledge” is basically building up a knowledge base based on learning curves via agents from different environments and from previous learning experience and then using this knowledge base to fill in incomplete data examples (Ramon et al., 2007). Experimental results against eight datasets revealed that the improved covering algorithm consistently produced competitive classifiers when compared with other RI algorithms such as RULES-IS and PRISM. One of the major obstacles facing the data mining algorithm is the massive amount of data that are stored and scattered in different geographical location. Decision makers are striving to have an “on the fly” mining approach that processes very huge database simultaneously hoping to improve planning decisions. Most of the research works on parallelisation in classification are focused on decision trees. However, RI may present simple classifiers to decision makers. The big data problem have been investigated by amending PRISM algorithm to handle parallelisation (Stahl and Bramer, 2012). The authors developed a strategy for parallel RI called Parallel Modular Classification Rule Induction (PMCRI). This strategy is a continuation of an early work by the same authors in 2008 which resulted in parallel PRISM (P-PRISM) (Stahl and Bramer, 2008). P-PRISM algorithm was disseminated to overcome PRISM’s excessive computational process of testing the entire population of data attribute inside the training dataset. The parallelism involves sorting the attribute values based on their frequency in the input dataset and the class attribute. This means the algorithm will need only this information for processing the rules and hence holding these data rather than the entire input data reduces computing resources such as processing time and memory. The attribute values and their frequency are then distributed to clusters and rules are generated from each cluster. All rules are finally merged to form the classifier. Experiments using a replication of the Diabetes and Yeast datasets from UCI repository have been conducted. The results show that P-PRISM as well as the parallel RI strategy scales well. However, a better approach to evaluate the parallelisation is to utilise unstructured real data such as Bioinformatics or text mining where dimensionality is huge and number of instances vary rather structured datasets. (Stahl and Bramer, 2014) developed a PRISM based method for ensemble learning in classification. Usually ensemble learning is an approach in classification used to improve the classifier’s predictive accuracy by deriving multiple classifiers using any learning approach such as Neural Network, decision trees, etc, and then merging them to form a global classifier. Since PRISM often suffers from overfitting problem to decrease this risk, the authors have built multiple classifiers based on PRISM using the ensemble learning approach. The results derived from fifteen UCI datasets revealed that the ensemble learning model based on PRISM was able to generate results comparable with classic PRISM algorithm. Moreover, a parallel version of the new model has also been implemented by the authors and tested with respect to training time. Results on run time showed that the parallel version scales well during the process of constructing the classifier. PRISM successors have focused mainly on improving the algorithm scalability or reducing overfitting. We have seen approaches such as P-PRISM that showed promising research direction toward parallel data mining using RI. Other approaches such as Ensemble Learning based PRISM was able to minimise overfitting the training data by generating ensemble classifiers that later on are integrated together to make a final classifier. Lastly, early pruning has been introduced to further minimise the search space by the introduction of J-pruning. There are needs to further cut down the irrelevant candidate items during forming the rules in PRISM. This indeed will have a positive impact on the algorithm’s efficiency in mining the rules and building the classifier. We think that post pruning is essential in PRISM and should increase its chance of being used as a data mining solution and decision making tool in practical applications. Yet since the classifiers produced by this family of algorithms is large in size this limits its usage. One promising solution to shrink the size of the classifier is by eliminating rules overlapping in the training example and having rules to cover larger data samples. This solution can be accomplished by having live items frequency during the mining phase since PRISM relies primarily on item’s frequency to build rules. In particular, PRISM algorithm often employs the rule’s accuracy as a measure to generate the rule especially when the rule’s accuracy reaches 100%. This normally results in low data coverage rules. We want each candidate item that can be part of a rule body to be associated with its true frequency in the training dataset that usually changes when a rule is generated. Recall that when a rule is outputted by PRISM, all training data linked with it are discarded and this may affect items appearing in those discarded rows. When each item has its true data representation while building the classifier this indeed decreases the search space and all insufficient candidate items will be deleted rather stored as in PRISM. The overall benefit will be that of having fewer number of rules classifiers. These classifiers can be seen as a real decision support system since they hold concise knowledge that are maintainable and usable by decision makers. 4. Enhanced Dynamic Rule Induction Algorithm The proposed algorithm (Figure 2) has two primary phases: rule production and class assignment of test data. In phase (1), eDRI produces rules from the training dataset that have accuracy>=Rule_Strength. This parameter is similar to the confidence threshold used in associative classification (Thabtah, et al., 2004) that aims to generate near perfect rules besides rules with 100% accuracy as classic PRISM. The proposed learning method ensures the production of rules that have zero error as well as rules that survive the Rule_Strength parameter. The algorithm terminates building up the classifier when no more rules have an acceptable accuracy or the training dataset becomes empty. When this occurs, all rules get merged together to form the classifier. There is another parameter utilised by eDRI algorithm in phase one to minimise the search space of items. This parameter is called freq and it is similar to the minimal support threshold used in the association rule. The freq parameter is primarily utilised to differentiate between items that are frequent (have high number of occurrences in the training dataset) from those which are not frequent. This surely eliminates items which have low frequency, i.e. <freq threshold, as early as possible and thus saving computing resources as well as ensuring only significant items (frequent items) can be part of any rule’s body. The infrequent items are kept in PRISM in the hope of producing 100% accuracy rules, which indeed can be seen a major deficiency. All rules are induced in phase (1). Whenever a rule is built, training data examples linked with it are deleted, and the frequency of the waiting candidate items to be added which appeared in the discarded data, are instantly updated. The update involves decrementing their frequencies. This can be seen as a quality assurance measure of not relying on the original frequency of items computed initially from the training dataset. Rather we have a dynamic frequency per item that is continuously changing whenever a rule is generated. Having said this, eDRI is a RI algorithm that does not allow items inside rules to share training data examples, hence ensuring live classifiers that are not dependant on a static training datasets, instead a dynamic dataset that lessens every time a rule is formed. In phase (2), rules derived are used to guess the type of the class for test cases. Our algorithm assumes that the attributes inside the training dataset are nominal and any continuous attribute must be discretised before the inducing the rules starts. 4.1 Rule Production and Classifier Building Input: Training dataset T, Minimum Rule_Strength, Minimum Frequency (freq) thresholds Output: A classifier that consists of If-Then rules Step 1: For each Attx in T Do Step 1.1: Calculate Attx in Ti p(wi = i| Attx); Step 1.2: Append the Attx with the largest accuracy (wi = i| Attx) to the body of rj Step 2: Repeat steps 1.1-1.2 until rj has either a) 100% accuracy b) or no longer can be improved and it has strength >= Rule_strength Step 2.1: Generate rj Step 3: Discard all data examples from T that contained rj’s body Step 3.1: Update the frequency of all impacted candidate items to reflect step 3 Step 3.2: Continue producing rules from Ti until all remaining unclassified items have frequency <freq threshold or no more data in Ti can be found Step 4: Generate the classifier. Fig. 2 eDRI Pseudocode The proposed algorithm scans the training dataset to record items plus their frequencies. Then it creates the first rule by appending the item that when added to the current rule achieves the best accuracy. Any item with frequency less than the freq threshold is ignored. The algorithm continues adding items to the current rule’s body until the rule becomes with 100% accuracy. For any rule that cannot reaches 100%, our algorithm checks whether its accuracy is greater than the Rule_Strength threshold. If the rule passes the Rule_Strength it will be created otherwise it will be deleted. Two noticeable differences between our algorithm and PRISM successors in building rules: a) In eDRI, no item is added to the rule’s body unless it has the minimum frequency requirement. Otherwise the item gets ignored and will not be part of any rule b) In eDRI, there is a possibility of creating rules that has accuracy less than 100% whereas PRISM only allows the generation of perfect rules. In eDRI, whenever a rule is generated, the following applies 1. The rule’s data samples in the training dataset are discarded 2. Before building the next in-line rule, our algorithm updates items frequencies which have appeared in the removed data samples. The proposed algorithm continues creating rules for the current class until no more items with sufficient frequency can be found. At this point, eDRI moves to the next class and repeats the same process until the training dataset becomes empty. The above rule discovery procedure keeps items rank dynamic specifically since an item frequency with the class is continuously amended whenever a rule is derived. The dynamism provides a distinguishing advantage for the algorithm in determining items which become weak during constructing the classifier. This minimises the search space for candidate items and should provide smaller in size classifiers. As matter fact, the proposed algorithm develops a pruning procedure that reduces overfitting and results in rules with larger data coverage than PRISM. To clarify, PRISM keeps adding items to the rule’s body regardless the number of data examples that are covered by the rule. The focus of PRISM is maximising rule’s accuracy even when the discovered rule covers one data example. This obviously may overfit the training data and results in large numbers of specific rules. A simple run of PRISM algorithm over the “Weather.nominal” (14 data samples) dataset from the UCI repository revealed two rules covering a single data example each. In other words, 33.33% of the PRISM classifier (2 rules out of six) covers just two data examples. This result if limited show how PRISM continues training without providing a red flag when to stop rule learning. For eDRI, these two rules are basically discarded since they don’t hold enough data representation. Since PRISM algorithm creates only rules with 100% accuracy, there is no rule preference procedure. On the other hand, eDRI classifier may contain rules with good accuracy not necessarily 100% and hence our algorithm utilises a rule sorting procedure to differentiate among rules. The rule’s strength and frequency are the main criteria employed to sort rules. When two or more rules have the same strength then eDRI uses the rule’s frequency as a tie breaker. Finally, whenever two or more rules have identical strength and frequency then the algorithm prefers rules with less number of items in their body. Table 1: Sample training dataset from (Witten and Frank, 2005) outlook temperature humidity windy play covering rule sunny hot high FALSE no Rule 2 sunny hot high TRUE no Rule 2 overcast hot high FALSE yes Rule 1 rainy mild high FALSE yes Rule 3 rainy cool normal FALSE yes Rule 3 rainy cool normal TRUE no Default Rule: NO overcast cool normal TRUE yes Rule 1 sunny mild high FALSE no Rule 2 sunny cool normal FALSE yes Rule 3 rainy mild normal FALSE yes Rule 3 sunny mild normal TRUE yes Default Rule: NO overcast mild high TRUE yes Rule 1 overcast hot normal FALSE yes Rule 1 rainy mild high TRUE no Default Rule: NO 4.2Test Data Prediction Our classifier consists of two types of rules a) Rules with 100% accuracy: Primary rules (higher rank) b) Rules with good accuracy, i.e. < 100%: Secondary rules (lower rank) Whenever a test data is about to be classified, eDRI goes over the rules in the classifier in top down fashion starting with the primary rules. The first rule that has items identical to the test data classifies it. In other words, if the rule’s body is contained with the test data then this rule’s class is assigned to the test data. If no primary rules match the test data then eDRI moves into the lower rank rules to forecast the test data class. This procedure limits the use of the default class rule which may increase the numbers of misclassifications. It should be noted that when no rules are in the classifier match the test data then the default class rule is fired. 4.4 Example on the Proposed Algorithm and PRISM This section provides a thorough example to distinguish the learning process of PRISM and our proposed algorithm, In particular, we show how rules are induced. The dataset displayed in Table 1 is used for the purpose of comparison. Assume that the freq threshold is set to 3 and the Rule_Strength to 80%. eDRI picks the item that has the largest accuracy when linked with a class after calculating the frequency of all items in Table 2.The largest accuracy item, i.e. (4/4), is associated with “Outlook=overcast” and it is linked with class “YES”. Our algorithm generates the below rule since it achieves 100% accuracy and removes its training data (Red text of Table 1) besides updating the frequency of all items appeared in the removed data as shown in Table 2. All items highlighted in red in Table 2 failed to pass the freq threshold therefore they are ignored. Our algorithm has substantially minimised the search space by keep only strong items (the one not highlighted red in Table 2). Whereas, PRISM keeps these items aiming to seek for specific rules. RULE (1) If “Outlook=overcast” then YES (4/4) After R1 is created, three items linked with class “YES” will be discarded since they become weak and these are “Temperature=cool”, “Humidity=high” and “Windy=true”. Their frequency is highlighted in red within the third column of Table 2. eDRI starts building a new rule from the remaining data examples excluding the removed data of RULE (1). The largest accuracy (item+class), i.e. 80%, is linked with “Humidity=high, NO” according to the updated frequency and accuracy of Table 2 (Columns 4 and 5). So the algorithm starts making the second rule below RULE (2) If “Humidity=high” then NO (4/5) Since RULE (2) is not yet 100% accurate, all instances associated with it are separated in Table 3 to seek another possible item that can maximise the accuracy. eDRI computes the frequency of items of Table 3 as shown in Table 4. Based on Table 4, the highest rule’s accuracy, i.e. 3/3, is linked with “Outlook=sunny” so this item is added to the current rule’s body, and the rule’s accuracy becomes 100%. Hence, we generate Rule (2) below, remove all training data connected with it (the yellow text of Table 1), and update the frequency and accuracy for the remaining impacted items. RULE (2) If “Humidity=high and Outlook=sunny” then NO (3/3) After creating RULE (2), four items have been ignored since they become weak (their frequencies are highlighted in red in Table 1- Column 6). Actually, there are no longer items linked with class “NO” simply because none of the remaining ones survive the freq threshold. The first two rules cover successfully seven examples in Table 1 and only seven data examples are left. eDRI calculates again the accuracy of possible remaining items. Item “Windy=False, YES” is the largest accurate item with accuracy 100% (4/4) hence a the below rule is formed. Table 2: Candidate items frequency and accuracy during rule generation process During Rule 1 Generation After generating Rule (1) After generating Rule (2) After generating Rule (3) Candidate Item+Class Original Frequency in the training data Rules Accuracy Freq. status Rules Accuracy Freq. status Rules Accuracy Freq. status Rules Accuracy Outlook=sunny, NO 3 60% 3 60% 0 Outlook= rainy, NO 2 Temperature= hot, NO 2 Temperature= mild, NO 2 Temperature= cool, NO 1 Humidity= high, NO 4 57.15% 4 80% 1 Humidity= normal, NO 1 Windy= true, NO 3 50.00% 3 75% 2 Windy= false, NO 2 Outlook=overcas t, YES 4 100% 0 Outlook=sunny, YES 2 40.00% Outlook= rainy, YES 3 60.00% 3 60% 3 60% 0 Temperature= hot, YES 2 0 Temperature= mild, YES 4 66.67% 4 66.67 4 80% 2 Temperature= cool, YES 3 75.00% 2 Humidity= high, YES 3 42.85% 1 Humidity= normal, YES 6 85.71 4 80% 4 80% 1 Windy= true, YES 3 50.00% 1 2 Windy= false, YES 6 75.00% 4 66.67 4 100% 0 Table 3: Data samples associated with item “Humidity=high” outlook temperature humidity windy play sunny hot high FALSE no sunny hot high TRUE no rainy mild high FALSE yes sunny mild high FALSE no rainy mild high TRUE no Table 4: Updated accuracy of items computed from Table 3 Item+ NO Frequency Rule's Accuracy Outlook=sunny 3 100% Outlook=rainy 2 50% Temperature= hot 2 100% Temperature= mild 3 67% Windy= true 2 100.00% Windy= false 3 67% Windy= false 3 67% RULE (3) If “Windy=false” then YES (4/4) This rule covers four data examples so they will be deleted (highlighted in green in Table 1) and three examples are left unclassified since non of the remaining items pass the freq threshold hence a default class rule is generated based on the largest frequency class connected with the unclassified instances (Default class rule is “NO” 2/3). The classifier of eDRI contains 3 rules and one default class rule as follows RULE (1) If “Outlook=overcast” then YES (4/4) RULE (2) If “Humidity=high and Outlook=sunny” then NO (3/3) RULE (3) If “Windy=false” then YES Otherwise “YES” Otherwise “NO” (2/3) In the above example, eDRI was able to induce three possible rules and a default class whereas PRISM produced six possible rules and a default class as shown below. Actually, most of PRISM rules cover very limited data examples (1 or 2 rows) whereas our algorithm rules cover at least three data examples. This makes a difference especially in large datasets or datasets with high dimensionality as we will see in the experimental section. The fact that our algorithm derived less rules by 42.85% than PRISM from a dataset of just 14 examples is an evidence if limited on the power of the new rule induction procedure proposed. Prism rules --------------- 1) If outlook = overcast then yes 2) If humidity = normal and windy = FALSE then yes 3) If temperature = mild and humidity = normal then yes 4) If outlook = rainy and windy = FALSE then yes 5) If outlook = sunny and humidity = high then no 6) If outlook = rainy and windy = TRUE then no Otherwise “YES”. 5. Data and Experimental Results 5.1 Settings For the rule discovery phase, all algorithms have utilised 10-fold cross validation in the experiments. This procedure has been adopted to compute the error rate during constructing the classifiers since it reduces overfitting and it is commonly used in classification. All experiments have been conducted on a computing machine with 1.7 GHz processor. We have implemented our algorithm in WEKA (Witten and Frank, 2005) for fair testing so WEKA environment was used for conducting all experimental results. WEKA is an open source Java platform that was developed at the University of Waikato, New Zealand. It contains different implementations of data mining and machine learning methods for tasks including classification, clustering, regression, association rules and feature selection. We tested the applicability of eDRI algorithm and in general RI on two sets of data: real data related to website security (Mohammad, et al, 2015) and sixteen UCI datasets (Lichman, 2013). The aim of the experiments is to show the pros and cons of our algorithm when compared to other RI algorithms. We have chosen three known RI algorithms (PRISM, OneRule (Holt, 1993), Conjunctive Rule (Witten and Frank, 2005)) and one hybrid decision tree algorithm called PART (Frank and Witten, 1998) for fair comparison. The choice of these algorithms are based on two facts: 1) They produce If-Then classifiers 2) They utilise different learning strategies to induce the rules It should be noted that the majority of PRISM successors such as (P-PRISM, N-PRISM) focus on treating noisy datasets and parallelism as described earlier in Section 3. Thus, their rule learning strategy is the same as the one in PRISM so we use PRISM WEKA’s version for comparison. For eDRI, the frequency threshold has been set to 1% and the Rule_Strength to 50%. These values have been obtained after running a number of warming up experiments where the results show a balance between the classifier's size and the classification accuracy. Below are the main evaluation measures used to analyse the experimental results 1) Classifiers error rate (%) 2) Classifier size measured in the available numbers of rules 3) The number of instances covered by the rule and the available number of items in the rule's body (rule's length). We have introduced two measures named Average Rule Length and Weighted Average Rule Length described later in this section along with their mathematical notations. 4) The number of data examples scanned to generate the classifier. We want to measure the increase and decrease in the search space by eDRI when compared to PRISM. 5) Time taken to build the classifiers. 5.2 Security Data Results Phishing is one of social engineering techniques used to exploit unawareness of people (Abdelhamid, et al., 2014). It allows hackers to get an advantage of the weaknesses in the web by demanding confidential information from users, such as usernames, passwords, financial account credentials and credit card details. This is often performed by inserting links and objects within emails that direct users to fake websites, which look like the original website. Often, victims of phishing may lose their bank details and other private information to the phishy email senders (Phishers). As a matter of fact, phishing costs financial institutions as well as online users hundreds of millions in monetary damages annually. We consider a real dataset related to website phishing in the first sets of experiments. The data has been collected using a PHP script of (Mohammad, et al., 2012) and contains thirty features plus the class. The dataset features have been considered in previous research articles which supports our selection, i.e. (Abdelhamid, et al., 2014). The phishing dataset is of type binary classification since there is two class labels (Phishy, Legitimate). We have utilised over 11000 website examples from different sources including Yahoo directory, Millersmiles and Phishtank archives. A sample of ten data examples associated with sixteen features are depicted in Table 5. The first row of the table shows the sample features and the last column in the table is the class attribute (Phishy (-1) / Legitimate (1)). Some features are given two possible values (-1 for phishing and 1 for legitimate) and others ternary values (-1 for phishing, 1 for legitimate and 0 for suspicious). Features used are not limited to “URL of Anchor”, “URL Length”, “Having_IP_Address”, “Prefix_Suffix”, “Iframe”, “Right”Click”, etc. More details on the complete set of features can be found in (Mohammad, et al., 2015. Table 5: Sample of ten websites data related to sixteen phishing features having_IP_ Address URL_ Length Shortining Service at- Symbol Double Slash Prefix_ Suffix Sub Domain SS Lfinal Domain _registr ation Favicon Port … Class -1 1 1 1 -1 -1 -1 -1 -1 1 1 … -1 1 1 1 1 1 -1 0 1 -1 1 1 … -1 1 0 1 1 1 -1 -1 -1 -1 1 1 … -1 1 0 1 1 1 -1 -1 -1 1 1 1 … -1 1 0 -1 1 1 -1 1 1 -1 1 1 … 1 -1 0 -1 1 -1 -1 1 1 -1 1 1 … 1 1 0 -1 1 1 -1 -1 -1 1 1 1 … -1 1 0 1 1 1 -1 -1 -1 1 1 1 … -1 1 0 -1 1 1 -1 1 1 -1 1 1 … 1 1 1 -1 1 1 -1 -1 1 -1 1 1 … -1 having_IP_Address URL_Length Shortining_Service at-Symbol Double_Slash Prefix_Suffix Sub_Domain SSLfinal Domain_registration Favicon Port … Class -1 1 1 1 -1 -1 -1 -1 -1 1 1 … -1 1 1 1 1 1 -1 0 1 -1 1 1 … -1 1 0 1 1 1 -1 -1 -1 -1 1 1 … -1 1 0 1 1 1 -1 -1 -1 1 1 1 … -1 1 0 -1 1 1 -1 1 1 -1 1 1 … 1 -1 0 -1 1 -1 -1 1 1 -1 1 1 … 1 1 0 -1 1 1 -1 -1 -1 1 1 1 … -1 1 0 1 1 1 -1 -1 -1 1 1 1 … -1 1 0 -1 1 1 -1 1 1 -1 1 1 … 1 1 1 -1 1 1 -1 -1 1 -1 1 1 … -1 Table 6: Best nine features detected by information gain filtering method Feature name IG IG Rank Score SSLfinal_State 1 0.4994 URL_of_Anchor 2 0.4770 Prefix_Suffix 3 0.1230 web_traffic 4 0.1145 having_Sub_Domain 5 0.1090 Links_in_tags 6 0.0470 Request_URL 7 0.0460 SFH 8 0.0370 Domain_registeration_length 9 0.0360 Fig. 4 Number of rules generated by PRISM and eDRI algorithms from the security data Fig. 3 The One error rate(%) of the considered algorithms on the phishing dataset Figure 3 shows the error rate (%) results of all considered algorithms against complete 31-features and the top 10-features after preprocessing the original dataset using information gain (IG) feature selection method (Table 6). The reason for selecting these features is since they are the only ones that pass preprocessing phase by scoring above the minimum score of IG. Figure 3 demonstrated a good and consistent performance in regards to error rate by the RI algorithms considered. Precisely, eDRI achieved competitive error rate to PRISM on the complete phishing data and outperformed all RI algorithms on the top ten features set. Surprisingly PRISM outperformed eDRI on the 31-features dataset by 2.47% yet derived 626 more rules. On the other hand and for the selected ten features by IG, eDRI outperformed PRISM and the rest of the RI algorithms. To be more specific, eDRI achieved 7.94%, 3.92% and 3.92% lower error rate than PRISM, Conjunctive Rule and OneRule algorithms. These figures give a clear evidence that the proposed algorithm’s learning strategy has improved the classifier predictive power at least of the reduced features set of the security data. Actually, eDRI ability to limit the use of the default class has a positive effect on its accuracy. Moreover, pruning useless and redundant rules enhanced the performance by ensuring only effective rules are utilised in prediction. These rules unlike those of PRISM and its successors cover larger portions of the training dataset. One notable result on the phishing dataset is that PART decision tree algorithm produced the least error rate. PART has achieved less error than eDRI since it adopts information theory pruning (Entropy) to further discards rules. In addition this algorithm employs Reduced Error Pruning (REP) to prune partial decision trees which normally degrade error rate. We believe that if eDRI utilises Entropy for post pruning we could end up with further improved classifiers. We looked at the number of rules resulted by our algorithm and PRISM, which is shown in Figure 4. This figure illustrates that eDRI substantially minimised the classifier size. In fact, PRISM has derived 626 and 326 more rules than eDRI from the 31-features dataset and 10-features dataset where many of its rules are covering very limited data samples hence overfitting the training dataset. We looked further into the classifier produced by eDRI from the complete phishing dataset and observed that out of the 44 rules generated, there are 22 rules have error greater than zero, making them represent 50% of eDRI’s classifier. Despite that these rules don’t hold an accuracy of 100%, they are new useful knowledge that PRISM classifier don’t hold. This is a good indication that generating rules not necessarily perfect (100% accuracy) not only minimises the classifier size but also covers larger numbers of training data examples per rule. Hence the overall performance of the classifier gets enhanced besides improving human controllability. In other words, having a smaller classifiers is a definite advantage for managers and decision makers since they are able to govern less amount knowledge during the decision making process. These classifiers work fine for applications such as medical diagnoses where general practitioners can enjoy a concise set of rules for daily diagnoses of their patients. We have investigated the relationship between the rule length in the classifier and the number of instances each rule has covered. We created new simple measures called the average rule length (ARL) and weighted average rule length (WARL) that are showing in the below equations: ARL = (# !" !"#$% !" !"#$%& ! ∗ !"#$ !"#$%& !) !"#$% # !" !"#!" WARL = 𝑟𝑖!𝑠 𝑙𝑒𝑛𝑔𝑡ℎ ∗ # !" !"#" !"#$%&! !"#$%$ !" !" !"#$ !" !!! !"#$%$%& !"#"$%# ! !!! For instance, assume that we have the following two rules: Rule 1 ( if a=windy b = mild then yes) (length 2), and this rule covers 3 data instances Rule 2 ( if a=sunny then no) (length 1), and this rule covers 97 data instances ARL = (1+2)/2 = 1.5 WARL = 2 * 3/100 + 1 * 97/100 = 1.03 For the dataset we consider and based on PRISM’s classifier, the ARL and WARL computed from 10- features and 31-features datasets are (7.19, 4.23) and (10.09, 3.99) respectively. Whereas and based on eDRI’s classifier, the ARL and WARL computed from 10-features and 31-features datasets are (4.67, 3.18) and (6.48, 3.96) respectively. The WARL figures demonstrate how the training attributes have been utilised to build the classifier. For instance, PRISM needed more features to make the rules than eDRI especially from the 10-features dataset. As a matter of fact, PRISM was excessively searching for more items to add per rule during rule building process. This explains the many specific rules (rules associated with large number of items) derived by PRISM which each classifies few training data examples. We assume that the above rationale is behind a larger WARL of PRISM. On the other hand and for the ARL computed, it seems that also eDRI has less number of items linked with each rule and this explains that our algorithm prefers general rules with lower length than PRISM. These general rules may contain several specific rules and this can be another reason behind the lower numbers of rules in eDRI’s classifiers. Since eDRI has been implemented in WEKA, we have investigated the number of times our algorithm pass over the training data examples when compared to PRISM. Hence, we recorded the number of rows that both algorithms have to scan while building the rules. It turns out that PRISM has passed over 34,465,178 training examples during the process of creating 670 rules. Whereas eDRI scanned 9,388,208 data examples resulting in 44 rules. These results evidently show that the proposed algorithm has substantially reduced the search space for rules. The fact that PRISM has to pass over many more millions of data examples explains the poor rule discovery mechanism employed by this algorithm, which overfits the training data, aiming to induce perfect yet low data coverage rules. This mechanism requires overtraining by repetitively splitting data examples whenever an item is appended to a rule in order to increase the rule’s accuracy. The outcome guarantees the production of rules having zero error but it is definitely time consuming and computing resource demanding. The frequency threshold employed in eDRI was able to successfully discard many items that have low data representation. This obviously decreased the search space of items and therefore the classifier size is reduced. Moreover, allowing rule’s which are Table 7: UCI datasets characteristics besides the considered algorithms error rate generated from the UCI datasets Dataset # of classes # of attributes # of instances PRISM eDRI PART Conjunctive Rule OneRule Contact lenses 3 5 24 45.84 33.33 16.66 37.50 29.16 Vote 2 17 435 6.67 6.43 4.96 4.96 4.37 Weather 2 5 14 64.28 35.71 42.85 35.71 57.14 Labor 2 17 57 25.35 22.80 17.54 24.56 31.57 Glass 7 10 214 51.40 47.66 39.25 53.73 48.59 Iris 3 5 150 12.04 10.00 4.66 39.33 4.00 Diabetes 2 9 768 39.01 27.08 26.56 34.37 26.43 Segment Challenge 7 20 1500 14.66 13.08 11.77 70.93 47.60 Zoo 7 11 101 57.42 6.93 7.92 40.59 93.06 Tic-Tac 2 10 958 4.32 8.45 5.84 31.00 30.06 Pima 2 7 768 43.74 23.56 23.30 25.65 25.26 Breast cancer 2 10 286 41.50 32.51 30.41 33.56 34.26 German_credit 2 16 1000 36.20 28.80 30.70 30.60 28.90 Autos 2 15 690 21.30 17.86 14.20 14.49 14.49 Soybean 19 36 683 14.63 10.10 8.49 73.79 66.47 Unbalanced 2 33 856 4.20 3.03 1.63 1.40 1.40 not 100% accurate to be generated decreases the number of data examples that eDRI passes over. This is since eDRI is not always eager to generate rules that are perfect rather it prefers rule’s with large data coverage but with acceptable accuracy, and this was obvious in its classifier that contains 22 non-perfect rules. 5.3 UCI Data Results Different datasets from the UCI data repository have been used to generalise the performance of the proposed algorithm. The datasets have been selected based on different characteristics including the number of features (attributes), the size, the number of classes, and the type of the attributes as shown in Table 7. All second sets experiments have been conducted in WEKA. Table 7 also displays the error rate of the classifiers generated by PRISM, OneRule, Conjunctive Rule, PART and eDRI algorithms. We also show the average error rate of all considered algorithms in Figure 5. Based on the average error rate results of the classifiers, eDRI has derived on average the highest predictive RI classifiers but PART which is a decision tree algorithm maintained the best classifiers. In the figure, it is clear that eDRI algorithm performed on average well when compared to OneRule, PRISM and Conjunctive Rule algorithms. More specific, on average eDRI gained lower error rate by 9.70%, 14.05%, 13.56% than PRISM, Conjunctive Rule and OneRule algorithms respectively. This gain has been resulted from the dynamic rules production by this algorithm which surely keeps the accurate rules which lead to improvement of the predictive power. On the other hand, decision tree algorithm has higher classification accuracy on overage than our algorithm. To be more precise, PART has on average 2.54% slightly higher accuracy than eDRI on the sixteen UCI datasets used in the experiments. This can be attributed to the Reduced Error Pruning (REP), which are used by this algorithm during the process of constructing the classifier. The fact that eDRI is competitive to PART and produced on average higher predictive classifiers than its own kind is an achievement. We further analysed Table 7 to seek detailed performance per dataset by the considered algorithms. Based on the table, the won-tie-lost records of eDRI against PRISM, Conjunctive Rule, OneRule and PART are 14-0-2, 12-1-3, 10-0-6 and 3-0-13. In other words, the proposed algorithm consistently outperformed PRISM and the other considered RI algorithms on the UCI datasets. This is evidently show good predictive performance by the proposed algorithm on small, medium and large datasets which makes it a favourable tool for decision making. PRISM performance on the UCI datasets has improved if compared to the real data from the security domain. This could be since the security data features are highlight correlated with the class attribute and this forces PRISM in searching deeply within the dataset for several specific rules. This explains the large rule’s length vs. instances covered for PRISM. The classifier size of both PRISM and eDRI are shown in Figure 5. The figure clearly demonstrates that our algorithm’s pruning method has significantly minimised the number of rules and no deterioration on the classifier’s accuracy has been observed. The average number of rules derived by PRISM and eDRI from the sixteen UCI datasets are 117.18 and 33.31 respectively. This drastic decrease in the classifier size of eDRI is attributed to a) the discarding of weak items during building the rule and b) the production of rules not necessarily having 100%. These two distinctive advantages have resulted in concise classifiers of eDRI if compared to PRISM. We think that the dynamic update of candidate items cuts down the search space of the items and therefore less numbers of candidate items are presented. So removing the overlapping of the training instances among rules has a good effect on the classifiers size. In particular, eDRI ensures that all candidate items frequencies are amended on the fly whenever a rule gets produced, this minimises the available numbers of candidate items for the next rules, and hence rules generated quicker and usually classifies more data. The time taken to construct the classifiers by PRISM, PART and our algorithm has been recorded to evaluate the efficiency factor. Figure 7 depicts the results in ms for the considered algorithms and against the UCI datasets we consider. The figure obviously illustrates that eDRI utilises less time to build the classifier than PART and PRISM due to there is no need to keep splitting training data samples in order to maximise each rule’s accuracy. This has definite advantage at least over PRISM algorithm, which repetitively adds items to the rule aiming to reach 100% accuracy. This repetitive addition of items results in three major disadvantages: • The resulting rules covering very small data samples. This may lead to the production of several redundant rules. • More data samples need to scanned to identify the highest frequent items to be added to the current rule and eventually overfitting the training dataset. There is no clear stopping condition for rule learning in PRISM • More training time is wasted Figure 7 also demonstrated that PART is slower than eDRI in building the classifier due to the multiple pruning procedures invoked by PART. We investigated the search space used by PRISM and eDRI to determine the major differences between both algorithms in constructing the classifier. Hence, we recorded the number of data samples (rows) scanned by both algorithms in determining items frequency and accuracy while building the rules. Table 8 illustrates our finding. For most of the datasets considered, PRISM had to scan more data samples to come up with the final classifier except for the “Labor” and “Vote” datasets. For instance and for the “Segment Challenge” dataset which consists of 20 attributes and 1500 instances, PRISM has to pass over 7,147,542 rows whereas eDRI scanned 938,263 rows. This is a clear sign that the proposed algorithm had substantially reduced the search space for rules in its classifier’s building process. The fact that eDRI on average reduced the search space by around four times less than PRISM is a definite bonus. This is since it stops building rules early when any rule has an acceptable error rate. Table 8:# of scanned data samples during building the classifier by PRISM and eDRI on the UCI datasets Datasset PRISM eDRI Contact lenses 1503 168 Vote 212805 223797 Weather 1160 196 Labor 13796 19507 Glass 325951 58613 Iris 20613 15265 Diabetes 1979939 418680 Segment Challenge 7,147,542 938,263 Zoo 72,852 21,011 Tic-Tac 509,652 120,526 Pima 682,505 358,451 Fig. 5 Average error rate (%) for the considered algorithms on the UCI datasets Fig. 6 The classifier size of PRISM and eDRI algorithms on the datasets Fig. 7 The classifier building time in ms for PRISM and eDRI algorithms on the datasets 6. Conclusions Rule Induction (RI) is a promising classification approach in data mining that attracted researchers due to its simplicity. This article deals with major shortcomings associated with PRISM, which is considered one of the known RI algorithms. Specifically, we investigated the problem of generating rules with limited data coverage by PRISM and discarding good quality rules covering larger training data portions, hence outputting massive classifiers. A new modified PRISM based algorithm that we call Enhanced Dynamic Rule Induction (eDRI) has been proposed and built into the WEKA environment to deal with the abovementioned deficiencies. Our algorithm minimises the search space for rules by cutting down items with low data representation and stopping learning when any rule meets the rule’s strength threshold. These improvements guarantee smaller sized classifiers that don’t overfit the training dataset and maintain their predictive performance. The classifiers derived by the proposed algorithm are of high benefit to managers particularly when taking key business decisions since they can serve as a rich knowledge base inside a decision making tool. This is due to the classifiers being easy to understand, having high predictive accuracy, and they are controllable as well as robust (flexible to be amended without drastic change). Moreover, eDRI can be employed as a prediction module embedded inside a decision support system in various domain applications that necessitate both simplicity of the outcome and good accuracy, such as medical diagnoses systems. Since general practitioners usually have busy clinics with tight allocated time per patient they favour decision support systems with a concise set of knowledge similar to our classifiers. Results using a real dataset related to security as well as general classification datasets from the UCI repository have been conducted utilising a number of RI and decision tree algorithms. The results revealed that the proposed algorithm is highly competitive with respect to error rate, search space, classifier size and processing time when compared to PRISM, OneRule, PART and Conjunctive Rule algorithms. Moreover, eDRI consistently produced fewer number of rules than PRISM on both the UCI and the security datasets, which increases its applicability. Nevertheless, PART decision tree algorithm outperformed eDRI in accuracy and on average has generated fewer numbers of rules for the majority of the datasets used. Hence, pruning using Entropy can be seen as advantageous for eDRI as a post pruning procedure. In addition, the current version of eDRI has no embedded discretisation method for continuous attributes. In the near future, we would like to build eDRI as part of a decision support system for medical diagnoses to further evaluate its effectiveness and suitability in practical applications. In addition, eDRI will be embedded soon by researchers in Huddersfield University as an anti-phishing tool. This will raise awareness of public employees about phishing because of eDRI's outcome simplicity that can be understood by different users. Furthermore, eDRI can be utilised in web and mobile accessibility to evaluate websites and mobile applications for visually impaired users based on discovering correlations among features related to these user categories. We also intend in the near future to employ information theory pruning in particular Entropy to increase the predictive power of eDRI besides further decreasing its classifier size. References [1] Abdelhamid N., and Thabtah F. (2014) Associative Classification Approaches: Review and Comparison. Journal of Information and Knowledge Management (JIKM), 13, 1450027. Sept 2014. Worldscinet. [2] Abdelhamid N., Ayesh A., Thabtah F. (2014) Phishing detection based associative classification data mining. Expert Systems with Applications 41 (13) Pages 5948–5959, Oct 2014. [3] Ayyat Susan, Lu J., Thabtah F. (2014) Class Strength Prediction Method for Associative Classification. Proceedings of the IMMM 2014, The Fourth International Conference on Advances in Information Mining and Management, pp. 5-10. Paris, France, July 2014. Best Paper Award. [4] Bramer, M.A. (2000) Automatic induction of classification rules from examples using N-PRISM”, Research and Development in Intelligent Systems XVI, Springer-Verlag, pp. 99–121, 2000. [5] Bramer A. (2002) An information-theoretic approach to the pre-pruning of classification rules, in: B. N. M Musen, R. Studer (Eds.), Intelligent Information Processing, Kluwer, 2002, pp. 201{212}. [6] Cendrowska, J. (1987) PRISM: An algorithm for inducing modular rules. International Journal of Man-Machine Studies, Vol.27, No.4, 349-370. [7] Cortes, C., and Vapnik, V. (1995). Support-Vector Networks. Machine Learning, 20 (3), 273 – 297. [8] Elgibreen H. A., Aksoy M. S. (2013) Rules – Tl: A Simple And Improved Rules Algorithm For Incomplete And Large Data. Journal of Theoretical and Applied Information Technology, pp. 28-40. [9] Frank, E., and, Witten, I. (1998) Generating accurate rule sets without global optimisation. Proceedings of the Fifteenth International Conference on Machine Learning, (p. . 144–151). Madison, Wisconsin. [10] Holte, R.C. (1993). Very Simple Classification Rules Perform Well on Most Commonly Used Datasets. Machine Learning, 11, pp 63-90. [11] Lichman, M. (2013). UCI Machine Learning Repository [http://archive.ics.uci.edu/ml]. Irvine, CA: University of California, School of Information and Computer Science. [12] Liu, B., Hsu, W., and Ma, Y. (1998) Integrating classification and association rule mining. Proceedings of the Knowledge Discovery and Data Mining Conference- KDD, 80-86. New York. [13] Mohammad RM, Thabtah F, McCluskey L (2013) Predicting phishing websites based on self-structuring neural network. Neural Computing and Applications, 1-16 (2013). [14] Mohammad R. Thabtah F. McCluskey L., (2015) Phishing websites dataset. Available: https://archive.ics.uci.edu/ml/datasets/Phishing+Websites Accessed January 2016. [15] Othman O. M. and Bryant C. H. (2015) Pruning Classification Rules with Instance Reduction Methods. International Journal of Machine Learning and Computing, Vol. 5, No. 3, June 2015. [16] Pham D. T. and Soroka A. J. (2007) An Immune-network inspired rule generation algorithm (RULES-IS). In Third Virtual International Conference on Innovative Production Machines and Systems, Whittles Dunbeath, 2007. [17] Rameshkumar, K., Sambath, M. and Ravi, S. (2013) Relevant association rule mining from medical dataset using new irrelevant rule elimination technique. Proceedings of ICICES , 300-304. [18] Ramon J., EDRIessens K., and Croonenborghs T. (2007) Transfer Learning in Reinforcement Learning Problems Through Partial Policy Recycling: Machine Learning. ECML 2007. vol. 4701, J. Kok, J. Koronacki, R. Mantaras, S. Matwin, D. Mladenic, and A. Skowron, Eds., ed: Springer Berlin / Heidelberg, 2007, pp. 699-707. [19] Stahl, F., Bramer, M.A. (2014) Random Prism: an alternative to Random Forests. Research and Development in Intelligent Systems XXVIII, 2011, pp 5-18, Springer. [20] Stahl F., Bramer M. (2012) Computationally efficient induction of classification ruleswith the PMCRI and J- PMCRI frameworks, Knowledge-Based Systems 35, (2012) 49–63. [21] Stahl F., and Bramer M. (2008) P-Prism: A Computationally Efficient Approach to Scaling up Classification Rule Induction. Artificial Intelligence in Theory and Practice II, IFIP – The International Federation for Information Processing Volume 276, 2008, pp 77-86. [22] Thabtah F., Hammoud S (2014) Parallel Associative Classification Data Mining Frameworks Based Mapreduce. To Appear in Journal of Parallel Processing Letter. March 2015. World Scientific. [23] Thabtah, F., Cowling, P., and Peng, Y. (2004) MMAC: A new multi-class, multi-label associative classification approach. Proceedings of the Fourth IEEE International Conference on Data Mining (ICDM ’04), 217-224. [24] Thabtah F., Hadi W., Abdelhamid N., Issa A. (2011) Prediction Phase in Associative Classification. Journal of Knowledge Engineering and Software Engineering. Volume: 21, Issue: 6(2011) pp. 855-876. WorldScinet. [25] Witten I. H. and Frank E. (2005). Data Mining: Practical Machine Learning Tools and Techniques. eDRI-EWSA-29-3 eDRI-EWSA-29-3.2 eDRI-EWSA-29-3.3 
work_dsbwecuosjf2xookbqblodxboe	----	Microsoft Word - IFAC90WithFiguresReformatted.doc PROCESS CONTROL USING A REAL TIME EXPERT SYSTEM Proc. IFAC 1990, Estonia, USSR pp. 234-239 (reformatted) R. Moore, H. Rosenof, and G. Stanley * Gensym Corporation, 125 CambridgePark Drive, Cambridge, Massachusetts 02140, U.S.A. Abstract. The challenges and status of the application of expert system technology in process plants is considered. In particular the problems of knowledge representation include the need to represent dynamic qualitative knowledge, dynamic analytic knowledge and the deep structure of the process. The application of inference in real-time requires paradigms which use metaknowledge to focus the inferencing resources of the expert system. Finally the application of truth maintenance requires a temporal model of the time dependence of the truth of data and inferred results. A structure which includes these considerations is presented. Keywords. Artificial intelligence, Expert systems, Real time computer systems, Process control INTRODUCTION The work done to connect expert systems to on-line processes goes back many years (Astrom 1984, Kawakami 1984, Lusk and Stratton 1983, Yamada and Motoda 1983). The early activity toward developing a real-time expert system technology (Cartwright and Ruskin 1986, Kramer and Palowitch 1985, Moore and co-workers 1984, Sauers and Walsh 1983) was intended for the area of on-line process diagnosis. Some early real-time designs were subsequently placed on-line in refineries and other sites (Moore and Kramer 1986). Subsequent work (Astrom 1989, Hofmann and co- workers 1989, Moore and co-workers 1987, Wolfe 1987) has extended the working definition of "real-time expert system" to include temporal reasoning based on integrated heuristic and analytic knowledge, and a full object-oriented representation integrated with deep knowledge of the structure of the plant. This paper presents the current working definition of a real- time expert system, based on 200 site installations. The many technology-leading organizations involved have provided significant contribution to the concepts. Some public references to this are (Arzen 1989a), (Aynsley, Peel and Morris 1989), (Byrd, Fisher and Mallett 1989), (Merris 1989), (Nevins 1989), and (Noren 1988). Current installations include applications in process monitoring and control, manufacturing, spacecraft telemetry, robotics, environmental control, network management and others. The most frequent implementations are in the chemical, petrochemical, aerospace and nuclear power industries. KNOWLEDGE REPRESENTATION Engineers and other plant experts know many types of process behavior, ranging from the analytical models of reactions to the lessons of past experience. To represent this in an expert system requires a knowledge representation that integrates several kinds of knowledge, including not only rules, but also differential equation models such as material balances, energy balances and reaction kinetics. A paradigm which only allows heuristics would, in the case of a process plant, only represent part of the expert knowledge of plant behavior. In this view, what is required is an integration of heuristic and analytic forms of knowledge. The object-oriented paradigm is used, with a hierarchical frame structure. Each object has behavior, which is directed or represented by combinations of analytic (model-based) and heuristic (rule-based) statements. In the object-oriented paradigm, behaviors may be defined for a class, and be inherited by each member of the class or subclass. However, with the object oriented paradigm, a member of a class can inherit some behaviors and not other, more specific, behaviors. For the most efficient representation, behavior is defined generically, with as much abstraction as possible, so that a class hierarchy is defined with behaviors at the highest class possible. Generic Representation of Behavior In many plants there are sets of objects which have related knowledge. For example, there may be many chemical reactors, all of which behave according to their individual states and inputs, but all with temperature, pressure and reaction relationships functionally defined by the combination of these and generic relationships. All the reactors may need inference analysis of a similar sort. The generic knowledge at the various class levels is combined with the parameters, states, inputs and specific behaviors of the specific instance to determine the overall object behavior, such as; d/dt( the temperature of any reactor) =(the net-energy- input of the reactor - the net-energy-output of the reactor + the reactor-heat of the reactor) * (the .............. and; the net-energy-input of any process-equipment = the sum over the inputs of the process-equipment of (........) Representation of the Plant Structure The process experts reason about the behavior using their knowledge of plant structure, knowing how the connected process units affect each other. In some applications the connections may change dynamically, as when alternate scenarios are being evaluated. To represent this, it is useful to use a graphical form of knowledge representation, so experts can define the plant structure and interactions by connecting objects on a computer workstation screen. The expert system can interpret this deep structure as part of the reasoning paradigm. The inference engine actively interprets the schematic representation of plant structure, reasoning about up-stream and down-stream causes and effects. To allow this, the object behaviors are defined in terms of data at object ports, rather than naming the specific objects providing the interactions, such as; .....the integrated-heat-flux of any reactor = the sum over the inputs of the reactor of......... and; if the pressure of any reactor > the maximum-allowed- pressure of the reactor then invoke overpressure rules for every process-equipment connected to the reactor. Temporal Knowledge Representation Temporal knowledge is typically very important, as the expert may be more concerned with the direction of dynamic behavior than with the current values of data. The differential equation models are one analytic representation of temporal knowledge. Other forms, which may be incorporated into heuristic or analytic behavior include relationships between data or events over time, such as; If the temperature of any tank as of 2 minutes ago > .....then.... or functions of data over time, such as; if the rate of change per minute of the temperature of any reactor over the last 2 minutes >.....then invoke safety rules for the reactor Integration of Knowledge Representations The overall knowledge representation is done using a combination of heuristic and analytic object behavior, including temporal knowledge, some of which is inherited from higher classes and some of which is specific, combined with the deep structure of the connectedness of the objects as dynamically interpreted during operation of the expert system, Fig. 1. Hofmann (1989) shows how this knowledge representation can be used to rapidly develop large applications using block diagrams of objects and their interactions as the user interface. We can say that now there is an increasing realization that a full control strategy requires not only parameter identification, state estimation and control , but that the validity of the data and process models must be checked even before the data and models are used in state estimation and subsequent control. As systems get more complex, there is a higher and higher probability of some failure occurring. The systems must be fault tolerant, allowing best available information to be used in flexible ways. For large system implementations, a productive human interface for the development and maintenance of the system, as is described here, becomes a necessary condition for practical implementation. Fig. 2. shows how the structured natural language parser can be used for building the knowledge base. REASONING IN REAL TIME First attempts at using expert systems for process control involved taking a "snap-shot" of plant data and using a static expert system paradigm to perform inference. Static expert systems use pattern matching of a set of "facts" and a set of rules, using some combination of forward and backward chaining to combine the inferences. With no time constraints, this might be practical. However with a time constraint, we need more efficient inference paradigms. Code improvements and computer performance improvements can help. For example the knowledge base can be manually partitioned to cut down on the search space. This is conceptually like having separate libraries or separate expert systems. Another efficiency derives from truth-maintenance. If only a few data are changing at a time, the truth- maintenance techniques of unwinding obsolete conclusions and only updating conclusions dependent on the new values is useful. However in most real-time applications a considerable amount of data is changing, at a rapid rate. Even with the code improvements and the continuing computer performance improvements, the static expert system approaches lead to slow performance on even small prototypes of a few hundred rules and a few hundred rapidly changing data. Real-Time Reasoning Paradigms A fundamentally different inference approach is appropriate for real-time problems. One approach that a human expert uses in a real-time situation is to maintain a peripheral awareness across the domain, watching for performance exceptions, and then focusing on areas of interest, using knowledge appropriate to the task. The G2 inference engine operates similarly. This requires various types of metaknowledge, so the expert system can invoke the appropriate knowledge. The metaknowledge may involve the problem type, the objects involved including connected objects, and other knowledge- about-knowledge. A retrieval facility allows the inference engine to invoke the requested types of knowledge, such as; if.....then invoke safety rules for any reactor where the temperature of the reactor > .... The initiation of a particular reasoning process may be from several causes. The most obvious is an event, which in the process plant is typically an alarm, such as; whenever the level-alarm of any tank receives a value and when the alarm is not return-to-normal then .... In process plants, one of the more promising roles of the expert system is to detect problems earlier, before they lead to an alarm condition. This requires that some reasoning be done continuously or periodically, regardless of the absence of "events" in the plant. This reasoning typically involves multiple measurements and dynamic models, analytic or heuristic, which in combination represent the way an expert would interpret the data. The inference engine continually scans some of the knowledge which the expert has specified for awareness of conditions which are not yet at alarm limits. If a safety- threatening condition occurs, the inference engine uses metaknowledge to determine which knowledge to invoke, thus focusing on the area of interest. Performance issues One benefit of the metaknowledge approach is that very large knowledge bases can be run in real time. Since many types of problems and behaviors are represented in the knowledge base, it can get quite large, with thousands of objects and rules. However G2 does not consume computer time looking for patterns. Rather it focuses attention on the knowledge needed. Thus a knowledge base may contain thousands of rules and objects, most of which are not consuming much CPU time, since nothing interesting is happening with them at the moment. In static expert systems, truth maintenance involves changing inferences when data changes. In real-time problems there may be an additional requirement to change inferences even if no new data is available, since measurements cannot be assumed to remain valid indefinitely. One way to express this temporal validity information is to attach an expiration time to each value maintained by the inference engine, and propagate this when inference is carried forward. Generally, when a conclusion is based on several time-sensitive variables, the earliest of their respective expiration times will be carried forward. Expiration times can be propagated forward through multiple levels of inference, but there are also ways to limit this propagation. This requires that data types be more complex, including validity information as well as values. The inference engine takes advantage of the truth maintenance structure to achieve further efficiency. If a high level conclusion is not currently of interest, then there is no need to forward-chain from the newest data values to update the conclusion. If the high level conclusion becomes of interest, then the validity information determines whether the conclusion needs updating, via backward chaining, or whether the latest value is still valid. This reasoning process can be overridden by the expert specifying that certain data should always drive forward chaining. The overall result is a considerable performance improvement over static paradigms, and it seems similar to the way a human expert deals with a dynamic domain. The forward propagation of validity intervals is under control of the developer. An example where the expiration times would not be carried forward would be in identification applications, where individual data may contribute to the learning of the parameter values of a model, and the "learned values" might have longer validity than each individual datum that entered into the identification. The Scheduler A real-time scheduler maintains the overall operation. Each task is a small one, perhaps a few milliseconds, and then the scheduler is in control of the next task. Some tasks are done periodically, such as integration of the state variable models and scanning of rules which are looking for problems. Other tasks are scheduled for the next available time-slot, such as invoking rules through metaknowledge. Still other tasks are scheduled for specific times, such as procedural checking of a batch operation. The scheduler also manages the data acquisition, by scheduling the data servers, and the data output which may be setpoint changes or direct control actions. The scheduler has facilities to consider the handling of asynchronous data when making an evaluation, which is a frequent problem in distributed applications. In most conventional continuous control theory (especially in state- variable systems), it is assumed that an entire vector of measurements is attained at a particular instant. All calculations are then done instantaneously as well. Neither data acquisition nor calculations are really done so instantly and synchronously in the real world. G2 is designed to handle every input, output, and calculation asynchronously and with priorities. This allows generality, allows true distributed processing, and insures that the computer never rests when there are tasks to be done. Ensuring Coherence The asynchronous data arrival time is available through its time stamp. As noted already, this time stamp can be propagated through subsequent calculations as a default option. More subtly, consider the following rule, where X is a variable that can change quite rapidly; If X> 10 and Y > 10 then conclude that ... G2 may be able to get a value of X quite quickly from a data acquisition system or internal calculation. Suppose, however, that it takes a long time to get Y, perhaps from a slow data acquisition system, or from a large external simulation. By the time Y arrives, the value of X may have expired due to a short validity interval assigned to a rapidly-changing variables. G2 will automatically get a new value for X in this case. G2 ensures that a coherent data set is acquired - that is, at one point in time, all variables have the indicated values, within the time resolution indicated by validity intervals. This data set coherence becomes an important issue as larger distributed systems are considered, and no single measurement vector representing values at a single point in time is available. Similar considerations exist with rule actions. They can be carried out "in parallel", or "in order". The difference becomes especially apparent if one of the rule conclusions changes the value of one of the antecedent variables either directly or indirectly. For "parallel" rule action execution, the values of the antecedent's variables are held constant until all actions can be completed. "In order" execution allows sequential execution. Note that the inference engine is required to support "loops" between rule antecedents and consequents. While this is forbidden in many static expert systems, it is essential if the real-time expert system is going to deal with closed-loop systems, which do in fact have "loops". The provisions of validity intervals, data coherence and other aspects described above make this possible. Acting Within a Time Limit Tasks have priorities, and within each basic interval the tasks are executed by priority. Priorities can be changed, and whole sections of knowledge can be disabled or enabled under rule control as an overload strategy. The scheduler has active metering, accessible to rules, so that flexible overload strategies can be constructed. Finding an answer within some specific time interval is facilitated by a default option. If a rule cannot be satisfied within a specified interval, another rule can be invoked to take a backup action. A typical cause might be the required data not being available. An alternative is to guarantee the "best" answer within a finite time-out period, such as; If the first of the following that has a current value ( the temperature-sensor of the reactor, the model-inferred- temperature of the reactor, the manual-backup- temperature of the reactor, the default-temperature of the reactor) > 200 then....... In the above expression, if no valid value is currently available in the first element of the list, a data acquisition is started. In the meantime, the second element is checked. If it has no current valid value, the calculations for the second one are started, and so on. At the end of the time-out period, the "best" answer (closest to the head of the list) is taken, and any remaining outstanding calculations are cancelled. KNOWLEDGE VALIDATION Consistency and completeness of knowledge are significant issues in the process domain, particularly as safety may be involved. The knowledge structure of G2, which allows dynamic models to be attached to objects, provides a framework for simulation testing as a step toward judging the consistency and completeness of the knowledge. Failure scenarios are a powerful way of validating a knowledge base. Where a dynamic model already exists, a failure scenario is created by introducing a problem such as a failed sensor or controller, or a physical problem in some component such as a leak or breakdown or overheating. The dynamic model will then propagate the effects of the failure, and one gets an opportunity to learn what the knowledge base will do in this situation. Many fault conditions, which would otherwise be untested until actual plant emergencies, can be simulated in this way. Provisions for locking a knowledge base after validation provides security. At the same time, experimentation with new or modified knowledge bases can be done concurrently with on-line operation of the validated knowledge base. This is possible because the knowledge base is stored as an ASCII file, which can be transported to other G2's and operated on by other G2's. The abstraction of data serving allows multiple versions of the knowledge base to operate on the same CPU or on networked CPU's, accessing the same on-line data, so experimentation with alternative knowledge bases can be done concurrently with normal operation. METHODOLOGY EXAMPLE We consider a typical fault diagnosis using G2. The first step is detecting the problem, which may be indicated by an event (alarm). However the more desirable situation is to detect problems before they develop into "events". To allow this, generic rules can scan across the domain looking for problems. This typically involves a combined analysis using multiple measurements. Analytic and heuristic models define expected behavior. The expert system then compares expected performance with measured behavior. This is similar in concept to the Kalman filtering or state observer technology, except that heuristic as well as analytic behavior may be the basis for the expected behavior. The next step is understanding the problem. For this purpose we use several tools for diagnosis of anomalous behavior. These include invoking diagnosis of connected objects, distinguishing between measurement errors and domain faults, invoking diagnostic knowledge using metaknowledge, and focus on problems. Reacting quickly to problems is facilitated by testing hundreds of rules per second, with sometimes multiple focuses of attention. The expert system can give advice or take action. APPLICATION STATUS As of this writing, the on-line applications of G2 include telemetry monitoring of flight data, manufacturing processes, manufacturing assembly, a biochemical process, a variety of chemical and petrochemical processes, closed-loop robotics and closed-loop process control. Research prototypes are underway in network monitoring, financial transaction monitoring, office workflow planning, closed life support systems, semiconductor manufacturing and a number of nuclear reactor safety projects. Some of the applications involve implementation of conventional multivariable and advanced control, using the combination of heuristic and analytic methodology, and taking advantage of the framework to facilitate implementation. Other applications are extending beyond what is reasonably possible with conventional systems, to implement measurement validation across a plant with 5000 objects and the equivalent of 7500 rules in one case. Other applications are directed toward large-scale fault-tolerant systems, which can use heuristics and the facilities for real-time truth maintenance to provide backup and "best available solution" behavior. REFERENCES Arzen, K-E. (1989a). Knowledge-Based Control Systems: Aspects on the Unification of Conventional Control Systems and Knowledge-Based Systems. Proc. 1989 American Control Conference, Pittsburgh, Vol. 3, pp 2233-2238. Arzen, K-E. (1989b). An Architecture for Expert System Based Feedback Control. submitted to Automatica. Astrom, K. J. and J. J. Anton (1984). Expert Control. IFAC World Congress, Budapest. Astrom, K. J. Towards Intelligent Control - Keynote Speech to the 1988 American Control Conference. IEEE Control Systems Magazine, 9:3, 60-64. Aynsley, M., D. Peel, and A. J. Morris (1989). A Real Time Knowledge Based System for Fermentation Control. Proc. 1989 American Control Conference, Pittsburgh, 3, 2239-2244. Byrd, J. S., J. J. Fisher, and W. R. Mallett (1989). Expert Robots for Process Environments. Robotics and Autonomous Systems. New York: Elsevier Science Publishers. Cartwright, C. and P. Ruskin (1986). Musing on the needs of real-time AI toolkits. Expert System User, V2, N-9, London, 30. Hofmann, A. G., G. M. Stanley, and L. B. Hawkinson (1989). Object-Oriented Models and Their Application in Real-Time Expert Systems. Proc. 1989, Society for Computer Simulation International Conference, San Diego. Kawakami, J. (1984). Overview of Artificial Intelligence Applied to Control Technology in Japan and in Hitachi. Hitachi Research Laboratory, private communication, Japan. Kramer, M. A. and B. L. Palowitch, Jr. (1985). Expert Systems and Knowledge-Based Approaches to Process Malfunction Diagnosis. Proc. AIChE Annual Meeting, Chicago. Lusk, E. and R. Stratton (1983). Automated Reasoning in Man-Machine Control Systems. Proc. Ninth Annual Advanced Control Conference, West Lafayette, 41-48. Merris, D. (1989). Gensym real time expert systems move into applications arena. Flexible Automation, Feb. 1989, 4. Moore, R. L., L. B. Hawkinson, C. G. Knickerbocker, and L. M. Churchman (1984). A Real-Time Expert System for Process Control. Proc. First Conference on Artificial Intelligence Applications (IEEE), Denver, 569-576. Moore, R. L., and M.A. Kramer (1986). Expert Systems in On-Line Process Control. Proc. Third International Conference on Chemical Process Control, Asilomar, 839-867. Moore, R. L., L. B. Hawkinson, M. Levin, A. G. Hofmann, B. L. Matthews, M. H. David (1987). Expert System Methodology for Real-Time Process Control. Proc. 10th World Congress on Automatic Control, IFAC, Munich, 6, 274-281. Noren, C. S. (1988). Rapid Prototyping Network Management Systems. IEEE Milcom '88, San Diego. Rosenof, Howard (1989). Real-Time Expert Systems in Process Control. EPRI Conference on Expert Systems Applications for the Electric Power Industry, Orlando. Sauers, R. and R. Walsh (1983). On the Requirements of Future Expert Systems. Proc. Eighth IJCAI , Karlsruhe, 110- 115. Stanley, G. (1989). The G2 Real-Time Expert System. Third International Conference on Expert Systems and the Leading Edge in Production and Operations Management, Hilton Head. Wolfe, A. (1987). An Easier Way to Build a Real-Time Expert System. Electronics, March 5, 1987. Yamada, N., and H. Motoda (1983). A Diagnosis Method of Dynamic System using the Knowledge on System Description. Proc. Eighth IJCAI, Karlsruhe, 225-229. * Robert Moore may be contacted at http://www.calventuretech.com/contactus.html * Howard Rosenof may be contacted at http://www.linkedin.com/pub/howard-rosenof/5/229/866 * Greg Stanley may be contacted at http://gregstanleyandassociates.com/contactinfo/contactinfo.htm Fig. 1. Examples of heuristic and analytic knowledge using deep structure. Fig. 2. Structured natural language for building the knowledge base. 
work_ducj2hy2njcavhcab2qnqp4mee	----	JESTEC TEMPLATE Journal of Engineering Science and Technology Vol. 12, No. 11 (2017) 2909 - 2921 © School of Engineering, Taylor’s University 2909 PROTOTYPE WEB-BASED EXPERT SYSTEM FOR FLEXIBLE PAVEMENT MAINTENANCE ABDALRHMAN MILAD 1, 2 *, NOOR EZLIN AHMED BASRI 1 , HASSAN M. ABDELSALAM 1 , RIZA ATIQ ABDULLAH BIN O.K RAHMAT 1, 2 1 Department of Civil and Structural Engineering, Faculty of Engineering and Built Environment, Universiti Kebangsaan Malaysia, 43600 UKM Bangi, Selangor DE, Malaysia 2 Sustainable Urban Transport Research Centre (SUTRA) *Corresponding Author: xyz_5001@yahoo.com.sg Abstract The paper describes the development of a prototype web-based expert knowledge system that can be used to maintain flexible pavement within a tropical region. This prototype system provides the advantages of using existing web-based expert system technology. Currently, deterioration of asphalt pavement layers is one of the biggest problems in Malaysia and requires maintenance to ensure that the roads remain open and able to guarantee the regularity, punctuality, and safety of all transport services. According to this process, the knowledge collection that has acquired and the date concerning to domain expert system of the development web-based system was launched with knowledge representation IF and THEN rules and coded by PHP programming. The web pages that support the user interface are created using a framework consisting of HTML, CSS, and J-Query. The prototype web-based expert system uses the knowledge of a pavement maintenance expert, or a specialist in pavement problem remediation, to emulate a portion of their professional reasoning abilities, which it can then use to assist with the maintenance of existing roads and enhance the efficiency and accuracy of the professional engineers tasked with the assessment of all available remedies. Thus, the system increases the performance level of the engineers in analysing, discerning and customising the information that will assist decision makers throughout the project, so the probability that the right decision and treatment are implemented at the right time is increased. Keywords: Prototype system, Road maintenance, Expert system, Remedies, Techniques. 2910 A. Milad et al. Journal of Engineering Science and Technology November 2017, Vol. 12(11) 1. Introduction In Malaysia, substantial amounts of time and money are spent on a yearly basis for the maintenance and rehabilitation of major and minor roads with evidence of distress, such as cracking and rutting [1]. However, there exists a deficit of skilled individuals with specialised training and expertise in the area of pavement evaluation [2]. This lack of adequate labour has resulted in a situation, wherein maintenance workers lack the craft of maintaining urban and rural road infrastructure, and possess limited knowledge of road construction or pavement maintenance. As a result, these workers require consultation from experts before selecting appropriate treatments during the maintenance of distressed asphalt pavement [3]. The maintenance of distressed pavement, and the process of returning it to a serviceable state is one of the most challenging problems faced by pavement engineers and executive decision makers within the highway sector [4]. A diagnosis of pavement distress requires a significant amount of engineering judgment, and individuals with the skills to make such judgements are in scarce supply, particularly when it comes to the maintenance of flexible pavements in Malaysia and other tropical regions. At the same time, pavement maintenance is one of the most cost-effective, socially and environmentally sustainable ways to achieve deteriorating roadway systems, when utilised in a suitable way [5]. Moreover, pavement maintenance is one of the most substantial components of a complete road network, and should be granted due significance for priority analysis. Priority Abbreviations CSS Cascading Style Sheets DC Density of Crack FHA Federal Highway Administration GUI Graphical User Interface HTML Hyper Text Mark-up Language ISA International Standard Atmosphere IT Internet Technology ITD Idaho Transportation Department Jquery JavaScript Library NACA National Advisory Committee for Aeronautics PC Pavement Condition PHP Programming Hypertext Pre-processor PWBKES Prototype Web-Based Knowledge Expert System RF Road Functional Class RPS Rapid Prototype System RT Remedy Techniques SC Severity of Crack SQL Structured Query Language TRB Transportation Research Board UWM University of Wisconsin-Madison VCTIR Virginia Centre for Transportation Innovation and Research WHO World Health Organization WSDOT Washington State Department of Transportation Prototype Web-Based Expert System for Flexible Pavement Maintenance . . . . 2911 Journal of Engineering Science and Technology November 2017, Vol. 12(11) analysis is a multi-criteria process for determining the appropriate classification list of candidate sections for maintenance and is based on several factors [6]. Road maintenance and rehabilitation of pavement models have been developed throughout the world, where an expert system is adapted to sustain the pavement maker of choice. Prediction of future pavement performance of a road network is a crucial step in a pavement management system equipped with a corresponding annual budget. Researchers used soft computing, i.e., fuzzy logic, neural network, and so on in enhancing the pavement management systems [7 - 9] and pavement deterioration modelling [10]. The development of web-based expert systems for assistance in remediation of flexible pavement deterioration has received increasing attention in the literature over the last few years [11]. Currently, hundreds of millions of users can access several billion documents on the Web, and even larger datasets reside within organisations’ intranets and Web-accessible databases (the so-called Deep Web) [12]. As the amount of available data continues to grow rapidly, it becomes increasingly difficult for users to organise, find, access, and maintain the information required [13]. By involving pavement engineers and other officials due to the lack of easily available systems that allow access to critical information relevant to flexible pavement maintenance technologies and may assist an engineer or a planner in deciding which technologies are potentially applicable to their projects [14]. When utilised in an intelligent manner, pavement maintenance is a cost-effective and environmentally effective technique for managing a deteriorating roadway system and is necessary for long-term planning of anticipated maintenance requirements needed to satisfy the users of the roadway systems [5]. 2. The Importance Factors Considered in the Prototype System This study focuses on the following factors that contribute to distress type:  the observed distress;  the cracking severity;  the cracking density;  the road function class;  the climate region; These factors were considered by the system in the development and selection of pavement treatments. Each factor was divided into particular categories, identified by existing documents and through interviews with selected experts in asphalt pavement. Most of the interviews were conducted with selected asphalt pavement experts, as it was considered inappropriate to gather group and class input from every worker. 3. Development of ESTAMPSYS Prototype At the point in time when pavement condition surveys are conducted, the knowledge engineer requires a minimum amount of information in order for 2912 A. Milad et al. Journal of Engineering Science and Technology November 2017, Vol. 12(11) him/her to make wise decisions about the level of need for rehabilitation, and the strategy by which to achieve the rehabilitation. These data requirements identify firstly the pavement distress types, as either single distress or combined distress, and secondly the category of the problem. The distress types are categorised according to their causal mechanisms and assigned a distress severity that measures the level of severity of each observed distress type. The distress severity measures the degree of deterioration, the amount of distress density and the relative percentage of the area of the project affected by each combination of distress type and severity. In addition, consideration is given to the degree to which the road is functional, as this may be impacted by the regional climate. A technically sound engineering condition survey must address each one of these needs, although the parameters of each category may vary from agency to agency. Table 1 shows an example of a distress type description, its associated severity, and density groups that can be used as criteria to choose remediation techniques that will correct the condition. Table 1. Criteria of pavement maintenance for transverse cracking. Type of Distress Evaluation RF RT PC SC DC Single Transverse Cracking Low Rare Minor Routing and seal >6mm Corrective Major Routing and seal >6mm Corrective Intermittent Minor Routing and seal >6mm Corrective Major Routing and seal >6mm Corrective Frequent Minor Burn and Seal Corrective Major Clean and seal Corrective Extensive Minor Crack filling Corrective Major Crack filling Corrective 3.1. The approach to resolving engineering problems Within the domain of engineering, problems and solutions are identified from two types of knowledge. The first is deterministic knowledge, characterised by a body of well-accepted and proven information that is available to engineers working in the field. The second is heuristic knowledge or personal information developed by individual engineers characterised by beliefs, opinions, and rules-of-thumb [15]. Typically, challenging engineering problems cannot be solved based merely on deterministic knowledge; there are two reasons for this. First, the problem may be so severe that the available deterministic knowledge is incomplete. Second, occasionally, the solution to an engineering problem cannot be classified as entirely correct or fully incorrect. In many cases, the knowledge engineer must select the best option from among a set of alternatives such that it is cost-effective and “good enough.” As these decisions must be made on-location, the engineer must be adaptable to the domain, and he/she must possess excellent evaluation Prototype Web-Based Expert System for Flexible Pavement Maintenance . . . . 2913 Journal of Engineering Science and Technology November 2017, Vol. 12(11) skills and good judgment acquired from previous experience in solving similar issues. Additionally, this selection process also requires that the engineers possess excellent technical skills. 3.2. Knowledge acquisition The process by which an expert knowledge base is acquired is a critical component of knowledge engineering. The knowledge acquisition process consists of gathering relevant information about a domain, usually from an expert who already possess substantial knowledge in highway pavement construction and maintenance, and who then aggregates further knowledge obtained from documented sources and human expertise Sources of knowledge can be partitioned into two classes. The first class consists of documented knowledge such as professional journals, manuals, and books written in the field of pavement maintenance. The information extracted from these sources becomes the knowledge foundation and is used principally by the proposed web-based expert system. Written sources in the domain of flexible pavement maintenance consist of studies that address various subjects and structures with documented exposure to pavement maintenance activities, as shown in Table 2. The second class of knowledge consists of undocumented sources that are integrated within expert systems by domain experts. The selection of domain experts is one of the most important components of any expert system development. Domain experts must be both knowledgeable of and have sufficient experience (theoretical, practical, or a combination) in their field. Moreover, the multiplicity of knowledge sources and types increases the complexity of knowledge acquisition. In addition, the time requirements for knowledge acquisition increase the difficulty in creating the knowledge base. Table 2. The majority of sources knowledge acquisition. No Title Publisher Year Version 1 Treatment for Virginia Pavements VDOT 2015 2 Pavement Patching Practices TRB 2014 3 Pavement Rating Manual ITD 2011 4 Pavement Distress Identification Manual FHA 2006-2009 5 Pavement Surface Evaluation and Rating UWM 2001 6 Asphalt Pavement Maintenance WSDOT 2000 7 Pavement Surface Condition Field Rating WSDOT 1999 3.3. Selection of building tool The development of a web-based expert system is largely dependent upon the design of the end-user interface (GUI) and server-side programming language employed (i.e., the web-based program in which the knowledge-based rules are encoded). The end-user interface manages all user input and, therefore, is designed for simplicity and ease of use. In this paper, the web pages that support the user interface are designed using a framework consisting of HTML, CSS, and J-Query. Additionally, a responsive web interface, for facilitating browsing from mobile devices, is ensured by utilising the Bootstrap framework. On the server side, the data extracted from the web page is processed by a proxy or agent process. This 2914 A. Milad et al. Journal of Engineering Science and Technology November 2017, Vol. 12(11) processing is executed on the server side using PHP, a high-level programming and scripting language that is easy-to-use and does not require the extensive knowledge of object oriented programming necessary when using C++, C# or Java. The relational database used to store the acquired knowledge is MySQL, a proprietary, non-standard implementation of entry level SQL. The expert system supports a GUI, thus making it more accessible for users with minimal expertise in data management. Users are not aware of the dependency on the MySQL database and need not have any knowledge of SQL. Novice users can, therefore, learn and improve their knowledge in flexible pavement maintenance. 3.4. Knowledge representation The structure of the ESTAMPSYS prototype, as shown in Fig. 1, consists of the relationships between the main components of the expert system and includes working memory, the inference engine, the knowledge base, and the user interface. Moreover, the expert system consists of two main parts, described as follows.  The development environment includes the part that is used to inject the expert’s knowledge into the expert system environment.  The consultation environment consists of the part that is used by non-expert users to gain knowledge. Fig. 1. Architecture of web-based expert system. 3.5. Graphical user interface Graphical User Interface (GUI) design focuses on interactions with users obtaining expert knowledge. The GUI provides screens, boxes, action buttons and other primitive input/output elements that are easy to access, understand, and use so as to facilitate the user actions. Making expert systems as user-friendly possible and avoiding the creation of a complicated design makes expert systems attractive to users in many fields. Expert systems must enable users to identify their problems while minimising any confusion or frustration that the user may encounter. The ESTAMPSYS web-based expert system provides a variety of different and useful user functions. The primary category of the system is displayed as toolboxes within the user interface. Toolboxes for flexible pavement distress are provided to assist end-users. Figures 2 and 3 show a screenshot of the main menu. In this section, the function of the toolboxes for single distress and combined distress, problems, and solutions are discussed and include practical examples. Prototype Web-Based Expert System for Flexible Pavement Maintenance . . . . 2915 Journal of Engineering Science and Technology November 2017, Vol. 12(11) Fig.2. A screenshot of the main menu. Fig. 3. A screenshot of the problems. 3.6. Toolbox for problems selection with solution At this stage, the toolbox for problems asks the user to select parameters based on the deterioration of pavement, as shown in Fig. 4, and presents a list of five categories, which were described previously, representing different criteria 2916 A. Milad et al. Journal of Engineering Science and Technology November 2017, Vol. 12(11) relating to a single instance of pavement distress. On this page, users can select the problems they have faced and the amount of distress observed, as it relates to severity, density, functional road class, and climate users of the expert system. Descriptions of such problems can help both users and engineers know the cause of problems and the effect these problems on the maintenance of roads. For example, if single pavement distress is observed, its severity is low, density is rare, the functional road class is minor, and climate is tropical rain, then the next page will inform the user that Crack Routing and Sealing is the appropriate solution, as shown in Fig. 5. Fig. 4. A screenshot of problems and solutions. Prototype Web-Based Expert System for Flexible Pavement Maintenance . . . . 2917 Journal of Engineering Science and Technology November 2017, Vol. 12(11) Fig. 5. A screenshot of solution. 2918 A. Milad et al. Journal of Engineering Science and Technology November 2017, Vol. 12(11) 4. Evaluation of the System Evaluation of expert systems is a very complicated task. It is a fundamental one, however, if the expert system is to be put in practice [16, 17]. Moreover, evaluations are useful in determining whether an expert system is meeting its intended aims. The evaluation methods are classified into two groups. Ten entrants were chosen and divided among two groups to test the satisfaction reported by the ESTAMPSYS users. The first group included five domain experts, while the second group included five computer engineers. The users evaluated ESTAMPSYS, assigning a mean value above five if they were satisfied. Evaluation deals with building the correct expert system by ensuring that the system precisely integrates human expertise. Therefore, the satisfaction of experts represents a significant factor in the evaluation process. The group of experts within the domain of pavement engineering produced a system verification mean value of (4.4000) i.e. (4.4000/5) = (88%), which represents the percentage of the sample that agrees that the system is functioning correctly. The group of computer scientists produced a system evaluation mean value of (4.5750), i.e., (4.5750/5) = (91.5%), and represents the percentage of the sample that agrees that the system is working correctly. The data from our questionnaires was analysed via an independent t-test over the sample space such that the value of the t statistic was (-1.297) with (8) degrees of freedom, so the p-value of this test is (0.231). Thus there is no statistical difference between these groups at a value of alpha=0.05. In this study, no significant difference between the groups was found for any questions, as shown in Table 3. Moreover, Fig. 6 explains the result questionnaire with the factors of evaluation, learnability of the ability of system application to authorise end-users learning how to work with the prototyping, overall assessment, quickness in running, lack of bugs, and ease to use. Table 3. Expert responses for evaluation statistically by t-test. Group N Mean SD t DF P value Pavement Engineering 5 4.400 0.231 -1.297 8 0.231 Computer science 5 4.575 0.190 Fig. 6. Results of prototype system evaluation. Prototype Web-Based Expert System for Flexible Pavement Maintenance . . . . 2919 Journal of Engineering Science and Technology November 2017, Vol. 12(11) 5. Maintain and Update the System In this step, the maintenance and update stage of the web-based expert system is considered. The first stage is to discover bugs as well as problems that may appear during the web-based running of the system with adaptation to user requests. The second stage is to ensure that the system is up to date and possessing the most accurate as well as recent knowledge concerning the domain of application in the new integration of new modified knowledge. The modifications to the knowledge-base are documented and attached to the design document accordingly; these amendments may reflect some changes to the implementation. Furthermore, the required modifications are made to the development version of the system, and all modifications are documented and attached to the user manual of the next release of the complete system [18]. To achieve the second stage, an arrangement for periodical meetings with domain experts are required, where domain knowledge is reviewed, and the latest updates in the field are discussed, and the requisite knowledge is acquired and augmented into the knowledge-base. Prototyping can be maintained by a knowledge engineer or by any qualified user, competent in PHP programming, under the supervision of a highway pavement engineer. The source code of prototyping includes remarks to simplify the update operation, especially when the user carrying out the update is not the developer (the knowledge engineer). 6. Conclusions This paper explains the development of a prototype web-based expert system to manage flexible pavement maintenance. The expert system, named ESTAMPSYS, implements suggestions for remedies that occur from pavement damage. We also document the effects resulting from the suggestions with the intent to assist practising engineers in identifying required repairs. The ESTAMPSYS inference engine analyses the input data parameters and recommends a set of solutions for each problem, such that the specific recommended solutions depend on the severity and density of the damaged pavement. If more than one solution is recommended, the inference engine requests further data, consisting of factors or conditions of the damage to allow the inference engine to identify the optimal solution. The extracted knowledge is not infallible but provides good to excellent documentation about pavement damage, particularly their effects, their causes, preventive actions, and remedies. The expert system can help road maintenance workers improve their professional ability to evaluate available treatments. The following conclusions were obtained in this study:  Single pavement distress problems were identified and classified according to multiple sources, such as manuals, journals and field experts.  Prototype, a web-based expert system for flexible pavement maintenance in tropical regions, has been developed by computerization of acquired knowledge using PHP programming.  Web pages that support the user interface were created using a framework consisting of HTML, CSS, and J-Query. ESTAMPSYS possesses an advanced knowledge base consisting of single distress module.  The graphical interface of the system accepts input to guide the user to agree on an input data selection to avoid user mistakes. 2920 A. Milad et al. Journal of Engineering Science and Technology November 2017, Vol. 12(11)  ESTAMPSYS was verified, validated and evaluated. ESTAMPSYS verification required unit testing.  Updating ESTAMPSYS to include new experiences is an easy operation because the system includes help facilities within the source code. Acknowledgement The authors would like to thank the Sustainable Urban Transport Research Centre (SUTRA), UKM for their assistance to provide all the facilities for this research. References 1. Hassim, S.; Teh, K.T.; Muniandy, R.; Omar, H., and Hassan, A. (2007). A prototype expert system for the selection of road construction materials. The Journal of Engineering Research, 4 (1), 1-10. 2. Goh, A.T.C. (1993). Advisory expert system for flexible pavement design. Artificial Intelligence in Engineering, 8(1), 47-56. 3. Zaniewski, J.; and Mamlouk, M. (1999). Pavement preventive maintenance: Key to quality highways. Transportation Research Record: Journal of the Transportation Research Board, 1680, 26-29. 4. Shah, Y.U.; Jain, S.S.; Tiwari, D.; and Jain, M.K. (2013). Development of overall pavement condition index for urban road network. Procedia-Social and Behavioral Sciences, 104, 332-341. 5. Kang, M.M. (2014). Development of an expert system for sustainable pavement preservation strategies (ES 2 P 2 S). Doctoral dissertation, The University of Wisconsin-Madison. 6. Shah, Y.U.; Jain, S.S.; and Parida, M. (2014). Evaluation of prioritization methods for effective pavement maintenance of urban roads. International Journal of Pavement Engineering, 15(3), 238-250. 7. Moazami, D.; and Muniandy, R. (2010). Fuzzy inference and multi-criteria decision making applications in pavement rehabilitation prioritization. Australian Journal of Basic and Applied Sciences, 4(10), 4740-4748. 8. Chou, J.S. (2009). Web-based CBR system applied to early cost budgeting for pavement maintenance project. Expert Systems with Applications, 36(2), 2947-2960. 9. Sandra, A.K.; Vinayaka Rao, V.R.; Raju, K.S.; and Sarkar, A.K. (2007). Prioritization of pavement stretches using fuzzy MCDM approach - A case study. In Soft Computing in Industrial Applications, ASC 39, 265-278. 10. Ortiz-García, J.J.; Costello, S.B.; and Snaith, M.S. (2006). Derivation of transition probability matrices for pavement deterioration modeling. Journal of Transportation Engineering, 132(2), 141-161. 11. Milad, A.; Basri, N.E.A., Borhan, M. N.; and Rahmat, R.A.A.O. (2016). A review of web based expert systems for flexible pavement maintenance. Jurnal Teknologi, 78(6), 139-147. Prototype Web-Based Expert System for Flexible Pavement Maintenance . . . . 2921 Journal of Engineering Science and Technology November 2017, Vol. 12(11) 12. Davies, J.; Lytras, M.; and Sheth, A.P. (2007). Guest editors' introduction: semantic-web-based knowledge management. IEEE Internet Computing, 11(5), 14-16. 13. Davies, J.; and Weeks, R. (2004, January). QuizRDF: Search technology for the semantic web. In Proceedings of the 37 th IEEE Annual Hawaii International Conference on System Sciences, 8. 14. Douglas, S.C. (2012). A web-based information system for geoconstruction technologies and performance of stone column reinforced ground. Ph.D Thesis, Iowa State University. 15. Lu, S.C.Y. (1986, December). Knowledge-based expert system: A new horizon of manufacturing automation. In Proceedings of Knowledge-Based Expert Systems for Manufacturing in the Winter Annual Meeting of ASME, Anaheim, California , 11-23. 16. Aguilar, R.M.; Muñoz, V.; Noda, M.; Bruno, A.; and Moreno, L. (2008). Verification and validation of an intelligent tutorial system. Expert Systems with Applications, 35(3), 677-685. 17. Mosqueira-Rey, E.; and Moret-Bonillo, V. (2000). Validation of intelligent systems: a critical study and a tool. Expert Systems with Applications, 18(1), 1-16. 18. Abdelhamid, Y., Hassan, H., and Rafea, A. (1997). A proposed methodology for expert system engineering. In 5th International conference on artificial intelligence applications. Egyptians Computer Society, Cairo, Egypt. 
work_dui2nfpcc5g5zd7r7f6mnwznvy	----	Comparison of multilabel classification models to forecast project dispute resolutions Comparison of multilabel classification models to forecast project dispute resolutions Jui-Sheng Chou ⇑ Department of Construction Engineering, National Taiwan University of Science and Technology, 43 Sec. 4, Keelung Rd., Taipei 106, Taiwan a r t i c l e i n f o Keywords: Data mining Multilabel classification Forecasting Dispute resolutions Public–private partnership Procurement management a b s t r a c t Early forecasting of project dispute resolutions (PDRs) provides decision-support information for resolv- ing potential procurement problems before a dispute occurs. This study compares the performances of classification and ensemble models for predicting dispute handling methods in public–private partner- ship (PPP) projects. Model analyses use machine learners (i.e., Support Vector Machines (SVMs), Artificial Neural Networks (ANNs), and Tree-augmented Naïve (TAN) Bayesian), classification and regression-based techniques (i.e., Classification and Regression Tree (CART), Quick, Unbiased and Efficient Statistical Tree (QUEST), Exhaustive Chi-squared Automatic Interaction Detection (Exhaustive CHAID), and C5.0), and combinations of these techniques that performed best for a set of PPP data. Analytical results exhibit that the combined technique of QUEST + CHAID + C5.0 has the best classification accuracy at 84.65% in pre- dicting dispute resolution outcomes (i.e., mediation, arbitration, litigation, negotiation, administrative appeals or no dispute occurred). Moreover, as the dispute category and phase in which the dispute occurs are known during project execution, the best classification model is the CART model, with an accuracy of 69.05%. This study demonstrates effective classification application for early PDR prediction related to public infrastructure projects. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction Taiwan has legally supported PPP projects for more than ten years. The National PPP Taskforce of the Taiwan Public Construc- tion Commission (TPCC) is generally responsible for nationwide policies and in some cases provides advice about provisions for individual projects. Engineering departments and local govern- ments are typically responsible for PPP project delivery. To achieve effective control of diverse projects under current workloads and to design proactive dispute management strategies, knowledge of possible PPP project dispute resolutions before disputes occur is essential to providing the governmental PPP Taskforce with infor- mation about future countermeasures. Additional preparation is generally beneficial once a dispute occurs by reducing the effort, time, and cost to multiple parties during dispute settlement. PPP projects involve devoted stakeholders, including a pro- moter (government), private investors, and financial institutions. Due to the high risks associated with the construction industry, re- peated challenges for stakeholders can result in project delays, budget overruns, and poor construction quality during the imple- mentation, construction, operating, and transfer phases. Although numerous studies (Abednego & Ogunlana, 2006; Cheung, 1999; Cheung, Suen, & Lam, 2002; Gebken & Edward Gibson, 2006; Jones, 2006) demonstrated that an efficient, effective, and fair dispute resolution process is essential for PPP project success, this study fo- cuses on identifying warnings of potential dispute resolutions prior to project initiation. The proposed classification methods provide governmental authorities with the information needed to design proactive measures during project preparation and the phase when a dispute occurs. Many PPP projects initiated during the last decade have failed due to disputes occurring in the build, operate, and transfer (BOT) phases. According to the TPCC, the dispute rate was 23.6% during 2002–2009 (PCC, 2011). The most common processes for handling disputes are mediation/negotiation and non-mediation procedures. Non-mediation procedures include arbitration, litigation, and administrative appeals. In Taiwan, up to 84% of PPP projects are set- tled by mediation or negotiation within only 1–9 months after disputes occur (PCC, 2011). Notably, arbitration or litigation costs all parties considerably more time and money when a mediated agreement cannot be reached. This study applies multilabel classification models to early pre- dict PPP dispute likelihood and potential resolutions, thereby alle- viating the future adverse effects of disputes on project delivery, operation, and transfer from a governmental perspective. First, this study acquired historical dispute data for PPP projects started dur- ing 2002–2009 to establish functional relationships between pro- ject characteristics and their corresponding dispute resolutions. Differing from conventional construction project disputes, PPP 0957-4174/$ - see front matter � 2012 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2012.02.103 ⇑ Tel.: +886 2 2737 6321; fax: +886 2 2737 6606. E-mail address: jschou@mail.ntust.edu.tw Expert Systems with Applications 39 (2012) 10202–10211 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2012.02.103 mailto:jschou@mail.ntust.edu.tw http://dx.doi.org/10.1016/j.eswa.2012.02.103 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa project disputes may occur during the building phase, as well as during the operating, renting, or transfer phases. Thus, a second modeling phase with only dispute cases was implemented to iden- tify the possibility of dispute resolutions under a set of known pro- ject attributes, dispute items, and the phase in which a dispute occurs. The rest of this paper is organized as follows. Section 2 thor- oughly reviews artificial intelligence literature and its application in predicting construction claims and litigation outcomes. Section 3 then presents the research methodology and evaluation meth- ods, respectively, providing a theoretical basis for classification models adopted in subsequent investigations. Section 4 describes the project dispute database and compares model performance based on classification techniques. Conclusions are finally drawn in Section 5, along with recommendations for future research. 2. Literature review of dispute forecasting In response to extremely large upfront investment costs, recent public infrastructure and building construction projects have been financed via PPP (Clifton & Duffield, 2006). Nevertheless, disputes between PPP participants usually occur unexpectedly and may involve many issues, including surety bond issue, sub-contractor qualifications, licenses, permits, investment scale, resident rights, government guarantees, excessive profits, operating period, taxa- tion, and default loan commitment (Jones, 2006). Disagreements among parties typically jeopardize a project plan via a time-con- suming dispute resolution process that damages a government’s reputation as it relates to PPP projects and reduces the willingness of investors to participate in future projects. When a dispute or claim occurs, the local government usually resorts to adjudication by the central governmental authority – the TPCC in this case – when initial negotiations fail to resolve conflicts. Once agreed upon by all interested parties, an impartial committee may be the next option for dispute mediation (Jones, 2006; Keith, 1997). The timing of committee formation, operating functions, and method implementation must be defined contractu- ally before project execution. Since mediation sometimes fails to resolve disputes, arbitration or litigation becomes the only option based on existing laws. However, as stakeholders lack confidence in the current arbitra- tion system, litigation is the primary resolution process for most Taiwanese PPP projects (PCC, 2010). Given that not all disputes or claims require a costly and time-consuming dispute resolution process, method that provides early warnings by predicting dis- pute and its handling methods is needed, such that governments and investors can enact dispute-prevention measures during pub- lic construction projects. Hence, management personnel would benefit when the TPCC has a decision-support tool for estimating dispute likelihood and for outlining how disputes would be re- solved before project start. Several studies have attempted to minimize the number of con- struction litigation cases by predicting the likely court rulings. Arditi, Oksay, and Tokdemir (1998), for example, trained a network using Illinois appellate court data, and achieved a 67% prediction accuracy for litigation outcomes (Arditi et al., 1998). They argued that if parties in a dispute know with some certainty how a case will be resolved in court, the number of disputes could be reduced markedly. Artificial intelligence (AI) techniques have achieved excellent prediction accuracy with the same dataset; a prediction accuracy of 83.33% was achieved with a case-based reasoning (Arditi & Tokdemir, 1999b), 89.95% was achieved with boosted decision trees (Arditi & Pulket, 2005), and 91.15% was attained with inte- grated prediction modeling (Arditi & Pulket, 2010). These studies used AI to enhance outcome prediction in conventional construc- tion procurement litigation. Furthermore, Chau (2007) found that, excluding the above case studies, AI techniques are rarely applied in the legal field (Chau, 2007). Thus, Chau utilized AI techniques based on particle swarm optimization to predict construction litigation outcomes, a field in which relatively new data mining (DM) techniques are rarely applied. The network developed by Chau achieved a prediction accuracy rate of 80%, much higher than chance. Nevertheless, Chau suggested using additional case factors related to cultural, psycho- logical, social, environmental, and political characteristics in future work. For construction disputes triggered by change order of the con- struction process and design, Chen (2008) developed a K Nearest Neighbour (KNN) pattern classification scheme that identifies po- tential lawsuits based on a nationwide study of US court records (Chen, 2008). Chen indicated that the KNN approach has an 84.38% classification accuracy. Chen and Hsu (2007) further ap- plied a hybrid artificial neural network-case based reasoning (ANN–CBR) model to a disputed change order dataset to obtain early-warning information. Their classifier achieved a prediction rate of 84.61% (Chen & Hsu, 2007). Although many studies have used CBR and its variations to identify similar dispute cases as references for dispute settlements, Cheng, Tsai, and Chiu (2009) further refined and improved the con- ventional CBR approach by combining fuzzy set theory with a new similarity measurement that integrates Euclidean distance and co- sine angle distance (Cheng et al., 2009). Their model successfully extracted the knowledge and experience of experts from 153 con- struction dispute cases collected manually from multiple sources. Generally, all related work focused on either specific change or- der disputes or conventional contracting projects. Characteristics and environments for construction projects under the PPP strategy, however, differ markedly from the contractor-owner relationships and require insightful analyses via AI or DM techniques with exploratory modeling performance comparisons to assist govern- ment agencies in predicting likely dispute outcomes before disputes occur. Since a dispute always involves numerous complex and inter- connected factors that are difficult to rationalize, using DM tech- niques is now among the most effective methods for identifying hidden relationships between available or accessible attributes and dispute resolution methods (Arditi & Pulket, 2005, 2010; Arditi & Tokdemir, 1999a; El-Adaway & Kandil, 2010; Kassab, Hegazy, & Hipel, 2010; Pulket & Arditi, 2009). Identifying these variables will provide practitioners with an improved understanding of the com- plexity of PPP project disputes. The DM- and AI-based approaches are related to computer sys- tem programs that attempt to resolve problems intelligently by emulating human brain processes. As AI technology enhances the ability of computer programs to handle tasks at which humans re- main superior (Haykin, 1999), AI techniques are typically applied to solve prediction and classification problems. Researchers in var- ious scientific and engineering fields have recently combined dif- ferent AI models to enhance their efficacy. Numerous studies have demonstrated that hybrid AI schemes generate promising results in many industries (Adeodato, Arnaud, Vasconcelos, Cunha, & Monteiro, 2011; Andrawis, Atiya, & El- Shishiny, 2011; Arditi & Pulket, 2010; Chen, 2007; Chou, Chiu, Farfoura, & Al-Taharwa, 2011; Chou, Tai, & Chang, 2010; Kim & Shin, 2007; Lee, 2009; Li et al., 2005; Min, Lee, & Han, 2006; Nandi et al., 2004; Wichard, 2011; Wu, Tzeng, & Lin, 2009; Wu, 2010). Notably, since selecting the most appropriate combination of AI models is both difficult and time-consuming, such that further at- tempts are not worthwhile unless prediction performance is improved significantly. J.-S. Chou / Expert Systems with Applications 39 (2012) 10202–10211 10203 https://isiarticles.com/article/18194 
work_dvbjk5b4traj7hnli4juvqavvu	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220969455 Params is empty 220969455 exception Params is empty 2021/04/06-03:05:08 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220969455 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:08 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_dwwzpxhm4naixfw5gkmce7vciu	----	AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 1 EXPERT SYSTEMS WITH APPLICATIONS An International Journal AUTHOR INFORMATION PACK TABLE OF CONTENTS . XXX . • Description • Audience • Impact Factor • Abstracting and Indexing • Editorial Board • Guide for Authors p.1 p.2 p.2 p.2 p.2 p.5 ISSN: 0957-4174 DESCRIPTION . Expert Systems With Applications is a refereed international journal whose focus is on exchanging information relating to expert and intelligent systems applied in industry, government, and universities worldwide. The thrust of the journal is to publish papers dealing with the design, development, testing, implementation, and/or management of expert and intelligent systems, and also to provide practical guidelines in the development and management of these systems. The journal will publish papers in expert and intelligent systems technology and application in the areas of, but not limited to: finance, accounting, engineering, marketing, auditing, law, procurement and contracting, project management, risk assessment, information management, information retrieval, crisis management, stock trading, strategic management, network management, telecommunications, space education, intelligent front ends, intelligent database management systems, medicine, chemistry, human resources management, human capital, business, production management, archaeology, economics, energy, and defense. Papers in multi- agent systems, knowledge management, neural networks, knowledge discovery, data and text mining, multimedia mining, and genetic algorithms will also be published in the journal. Reproducibility Badge Initiative and Software Publication Reproducibility Badge Initiative (RBI) is a collaboration with Code Ocean (CO), a cloud based computational reproducibility platform that helps the community by enabling sharing of code and data as a resource for non-commercial use. CO verifies the submitted code (and data) and certifies its reproducibility. Code submission will be verified by the Code Ocean team for computational reproducibility by making sure it runs, delivers results and it is self-contained. For more information please visit this help article. Note that an accepted paper will be published independently of the CO application outcome. However, if the paper receives the Reproducibility badge, it will be given additional exposure by having an attached R Badge, and by being citable at the CO website with a DOI. We invite you to convert your open source software into an additional journal publication in Software Impacts, a multi-disciplinary open access journal. Software Impacts provides a scholarly reference to software that has been used to address a research challenge. The journal disseminates impactful and re-usable scientific software through Original Software Publications which describe the application of the software to research and the published outputs. For more information contact us at: software.impacts@elsevier.com Benefits to authors https://help.codeocean.com/en/articles/1120151-code-ocean-s-verification-process-for-computational-reproducibility https://www.journals.elsevier.com/software-impacts https://www.journals.elsevier.com/software-impacts AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 2 We also provide many author benefits, such as free PDFs, a liberal copyright policy, special discounts on Elsevier publications and much more. Please click here for more information on our author services. Please see our Guide for Authors for information on article submission. If you require any further information or help, please visit our Support Center AUDIENCE . Knowledge Engineers, Managers, Systems Engineers, Automation and Control Engineers, Electronic and Electrical Engineers. IMPACT FACTOR . 2019: 5.452 © Clarivate Analytics Journal Citation Reports 2020 ABSTRACTING AND INDEXING . Cambridge/Computer and Information Abstracts Web of Science Research Alert Current Contents - Engineering, Computing & Technology Science Citation Index Expanded Scopus INSPEC EDITORIAL BOARD . Editor-in-Chief Binshan Lin, Louisiana State University in Shreveport, Shreveport, Louisiana, United States of America Founding Editor Jay Liebowitz, University of Maryland Global Campus, Adelphi, Maryland, United States of America Associate Editors Rodrigo Agerri, University of the Basque Country UPV/EHU, HiTZ Center - Ixa, Donostia-San Sebastian, Spain Information Extraction, Sequence Labelling, Natural Language Processing Stuart Barnes, King’s College London, King’s Business School, London, United Kingdom Sharing Economy, Consumer Behaviour, Sustainability, Digital Innovation, Machine Learning Vincent Charles, Pontifical Catholic University of Peru CENTRUM Graduate Business School, Lima, Peru Management Science, Artificial Intelligence, Data Science, Big Data Kwai-Sang Chin, City University of Hong Kong Department of Systems Engineering and Engineering Management, Hong Kong, Hong Kong Decision Support Systems, Management Intelligence, Quality and Engineering Management Sumali Conlon, University of Mississippi, University Park, Mississippi, United States of America Business intelligence, Expert and Intelligent Systems, Machine learning, Data science, Natural language processing, Intelligent systems and AI in Business & Erik Cuevas, University of Guadalajara University Center of Exact Sciences and Engineering, Guadalajara, Mexico Research topics/areas, Optimization, Metaheuristic methods, Evolutionary Computation, Swarm Intelligence, Image processing PIERPAOLO D'URSO, University of Rome La Sapienza Department of Social and Economic Sciences, Roma, Italy fuzzy clustering, clustering of spatial data, clustering of time series, clustering of spatio-temporal data, clustering of complex structures of data, imprecise data science Reggie Davidrajuh, University of Stavanger, Stavanger, Norway Discrete-event systems, Petri Nets, Modeling, Simulation, Performance Evaluation, Algorithms Zhi-Hong Deng, Peking University School of Electronics Engineering and Computer Science, Beijing, China Machine learning, Natural language processing, Data mining Debora Di Caprio, University of Trento Department of Economics and Management, Trento, Italy Knowledge-based Systems, Algorithm, Statistic Computing, Fuzzy system, Soft computing, Intelligent Systems and AI in Business & Social Science https://www.elsevier.com/authors/author-services https://www.elsevier.com/journals/expert-systems-with-applications/0957-4174/guide-for-authors https://service.elsevier.com/app/home/supporthub/publishing/ AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 3 Giorgio Maria Di Nunzio, University of Padua Department of Information Engineering, Padova, Italy Interactive Machine Learning, Visual Data Mining, Information Retrieval Theory, Big Data, Linked Open Data, Digital Geolinguistics Sofia Balula Dias, CIPER, Faculdade de Motricidade Humana, Universidade de Lisboa, Cruz Quebrada, Portugal Intelligent Learning Management Systems, Artificial Intelligence, Serious Games, Blended-Learning, Fuzzy Logic, Deep Learning, Higher Education, Exergames, Affective learning, Collaborative learning, Online learning Elisabetta Fersini, University of Milan-Bicocca Department of Informatics Systems and Communication, Milano, Italy Machine Learning, Natural Language Processing, Data Science Iztok Jr Fister, University of Maribor, Maribor, Slovenia Computational intelligence, Data mining, Sport, Swarm intelligence C. Bryan Foltz, The University of Tennessee at Martin, College of Business and Global Affairs, Martin, Tennessee, United States of America Information Systems Security, Privacy, Web Development Francisco Guijarro, Polytechnic University of Valencia, Valencia, Spain Business Intelligence, Trading Rules, Portfolio Management, Valuation Jin-Kao Hao, University of Angers, Angers, France Combinatorial Optimization, Metaheuristics, Large Scale Optimization, Applications Li-An Ho, Tamkang University, Taipei, Taiwan Knowledge Management, Project Management, Information Management, Human Resources Management, Human Capital, Instructional Design Alessio Ishizaka, NEOMA Business School, Mont St Aignan, France Multi-Criteria Decision Analysis, Supply Chain Management, Decision Support Systems, Operations Management, Decision Analysis Dragi Kocev, Jozef Stefan Institute, Ljubljana, Slovenia Ensemble Learning, Structured Output Prediction, Multi Label Classification, Hierarchical Multi Label Classification, Semi-supervised Learning, Feature Selection Salim Lahmiri, Concordia University, John Molson School of Business, Department of Supply Chain and Business Technology Management, Montréal, Quebec, Canada Data science, Data analytics, Signal processing, Intelligent systems, Pattern recognition Voon-Hsien Lee, Tunku Abdul Rahman University - Perak Campus, Kampar, Malaysia Business intelligence, Intelligent Systems and AI in Business & Social Science, Quality Management, Supply Chain Management Chunping Li, Tsinghua University, School of Software, Beijing, China Artificial Intelligence, Data Analysis and Data Mining, Machine Learning and Automated Reasoning Chinho Lin, National Cheng Kung University Department of Industrial and Information Management and Institute of Information Management, Tainan, Taiwan Knowledge Management, Strategic Management, Supply Chain Management, Total Quality Management Peter Love, Curtin University, Perth, Western Australia, Australia Infrastructure, Safety, Resilience Engineering, Asset Management, Information Systems Evaluation, Quality Management, Computer vision, Engineering Informatics, Deep Learning, Machine Learning Yannis Marinakis, School of Production Engineering & Management, Chania, Greece Swarm Intelligence, Vehicle Routing Problem, Supply Chain Management, Metaheuristic Algorithms, Optimization Yalda Mohsenzadeh, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States of America Rui Neves, University of Lisbon Higher Technical Institute, Lisboa, Portugal Algorithms applied to the stock market Luiz Eduardo de Oliveira, Federal University of Parana, Curitiba, Brazil Pattern Recognition, Machine Learning, Image Processing Keng-Boon Ooi, UCSI University, Kuala Lumpur, Malaysia Artificial neural networks for machine learning, decision support systems, data mining Joanna Paliszkiewicz, Warsaw University of Life Sciences, Warszawa, Poland Trust management, knowledge management, social media, leadership, trust Zbigniew Pastuszak, Maria Curie-Sklodowska University in Lublin Faculty of Economics, Lublin, Poland Business intelligence, Knowledge-based Systems, Data science, Decision support systems, prediction systems and early warning systems, Data driven optimization, Intelligent systems and AI in business & social science Guojie Song, Peking University, Beijing, China Data mining, Social network analysis, Intelligent transportation system Ying Tan, Peking University School of Electronics Engineering and Computer Science, Beijing, China AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 4 Fields of specialization- Deep Neural Networks, Deep Learning, Swarm Intelligence, Fireworks Algorithms, Data Mining, Financial Predication Joao Manuel R.S. Tavares, University of Porto Faculty of Engineering, Porto, Portugal Biomechanics, Medical Imaging, Computational Vision, Gait &, Posture, Biodevices, Image Processing &, Analysis, NDT Manoj Kumar Tiwari, National Institute of Industrial Engineering, Mumbai, India Decision Support Models, Simulation Model, Computational Intelligence, Machine Learning, Supply Chain and Logistics, Manufacturing Systems, , Conrad Tucker, Carnegie Mellon University, Pittsburgh, Pennsylvania, United States of America Machine learning , Intelligent systems and AI in engineering, Expert and Intelligent Systems Julien Velcin, Lumière University Lyon 2, Lyon, France Artificial Intelligence, Machine Learning, Data Mining, Natural Language Processing, Social Media Analysis, Digital Humanities Ling Wang, Tsinghua University Department of Automation, Beijing, China Computational Intelligence, Scheduling, Meta-heuristics, Soft Computing, Intelligent Optimization Ping Wang, James Madison University, Department of Computer Information Systems & Business Analytics, Harrisonburg, Virginia, United States of America Supply Chain Management Models and Analysis, Structural Equation Modeling, Multivariate Analysis, Classification Problems June Wei, University of West Florida College of Business, Pensacola, Florida, United States of America Business Intelligence, Knowledge-Based Systems Petros Xanthopoulos, Stetson University School of Business Administration, Deland, Florida, United States of America Business Analytics, Machine Learning, Operations Research Hua Xu, Tsinghua University Department of Computer Science and Technology, Beijing, China Natural Language Processing, Intelligent Optimization Algorithms, Intelligent Information Processing, Multimodal Information Processing and Mining Steve Yang, Stevens Institute of Technology Wesley J Howe School of Technology Management, Hoboken, New Jersey, United States of America Microstructure, Behavioral Finance, Portfolio Optimization, Algorithmic Trading, Financial Systemic Risk Editorial Advisory Board Pornthep Anussornnitisarn, Kasetsart University, Bangkok, Thailand Dianne Berry, University of Reading, Reading, United Kingdom Bernt Bremdal, University of Oslo, Oslo, Norway Francisco Javier Cantu, Monterrey Institute of Technology and Higher Education, Monterrey, Mexico Kimiz Dalkir, McGill University, Montreal, Quebec, Canada Randall Davis, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States of America Edward Feigenbaum, Stanford University, Stanford, California, United States of America Roar Fjellheim, Computas AS, Oslo, Norway Mark Fox, University of Toronto, Toronto, Ontario, Canada Olli-Pekka Hilmola, LUT University - Kouvola Unit, KOUVOLA, Finland Jae Kyu Lee, Korea Advanced Institute of Science and Technology College of Business, Seoul, South Korea Dusan Lesjak, International School for Social and Business Studies, Celje, Slovenia Ana Gabriela Maguitman, National University of the South, Bahia Blanca, Argentina Shawn Mankad, Weill Cornell Medicine, New York, New York, United States of America Vladimir Milacic, University of Belgrade, Belgrade, Serbia Arcot Desai Narasimhalu, Singapore Management University, Singapore, Singapore Daniel O'Leary, University of Southern California, Los Angeles, California, United States of America Zbigniew Pastuszak, Maria Curie-Sklodowska University in Lublin Faculty of Economics, Lublin, Poland Cezar Scarlat, Polytechnic University of Bucharest, Bucuresti, Romania Randall Shumaker, University of Central Florida, Orlando, Florida, United States of America James Slagle, University of Minnesota, Minneapolis, Minnesota, United States of America Ching Y. Suen, Concordia University, Montréal, Quebec, Canada Jie Tang, Tsinghua University, Beijing, China Efraim Turban, University of Hawai'i at Manoa, Honolulu, Hawaii, United States of America Ismail Burhan Turksen, University of Toronto, Toronto, Ontario, Canada Jan Vanthienen, KU Leuven, Leuven, Belgium Shouhong Wang, University of Massachusetts Dartmouth, Dartmouth, Massachusetts, United States of America Mark Zielinski, Harvard Medical School, Boston, Massachusetts, United States of America AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 5 GUIDE FOR AUTHORS . INTRODUCTION EXPERT SYSTEMS WITH APPLICATIONS is a refereed international journal whose focus is on exchanging information relating to expert and intelligent systems applied in industry, government, and universities worldwide. The thrust of the journal is to publish papers dealing with the design, development, testing, implementation, and/or management of expert and intelligent systems, and also to provide practical guidelines in the development and management of these systems. Types of Paper The journal will publish papers in expert and intelligent systems technology and application in the areas of, but not limited to: finance, accounting, engineering, marketing, auditing, law, procurement and contracting, project management, risk assessment, information management, information retrieval, crisis management, stock trading, strategic management, network management, telecommunications, space education, intelligent front ends, intelligent database management systems, medicine, chemistry, human resources management, human capital, business, production management, archaeology, economics, energy, and defense. Papers in multi-agent systems, knowledge management, neural networks, knowledge discovery, data and text mining, multimedia mining, and genetic algorithms will also be published in the journal. Contact Details for Submission Authors are requested to submit their original papers using the submission website. To submit your paper online, please go to https://www.editorialmanager.com/eswa/Default.aspx. Authors interested in online submission are requested to go to the website and upload their manuscript and its associated artwork. An electronic (PDF) proof is generated and the reviewing process is carried out using that PDF. The PDF file may be edited after acceptance to follow journal standards. Authors and editors send and receive all correspondence by email via the website and no paper correspondence is performed. For the name and address of the editor-in-chief, please refer to the list of editors in each issue of the journal or on the journal's homepage. Submission checklist You can use this list to carry out a final check of your submission before you send it to the journal for review. Please check the relevant section in this Guide for Authors for more details. Ensure that the following items are present: One author has been designated as the corresponding author with contact details: • E-mail address • Full postal address All necessary files have been uploaded: Manuscript: • Include keywords • All figures (include relevant captions) • All tables (including titles, description, footnotes) • Ensure all figure and table citations in the text match the files provided • Indicate clearly if color should be used for any figures in print Graphical Abstracts / Highlights files (where applicable) Supplemental files (where applicable) Further considerations • Manuscript has been 'spell checked' and 'grammar checked' • All references mentioned in the Reference List are cited in the text, and vice versa • Permission has been obtained for use of copyrighted material from other sources (including the Internet) • A competing interests statement is provided, even if the authors have no competing interests to declare • Journal policies detailed in this guide have been reviewed • Referee suggestions and contact details provided, based on journal requirements AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 6 For further information, visit our Support Center. BEFORE YOU BEGIN Ethics in publishing Please see our information pages on Ethics in publishing and Ethical guidelines for journal publication. Declaration of competing interest All authors must disclose any financial and personal relationships with other people or organizations that could inappropriately influence (bias) their work. Examples of potential conflicts of interest include employment, consultancies, stock ownership, honoraria, paid expert testimony, patent applications/ registrations, and grants or other funding. Authors should complete the declaration of competing interest statement using this template and upload to the submission system at the Attach/Upload Files step. Note: Please do not convert the .docx template to another file type. Author signatures are not required. If there are no interests to declare, please choose the first option in the template. This statement will be published within the article if accepted. More information. Submission declaration and verification Submission of an article implies that the work described has not been published previously (except in the form of an abstract, a published lecture or academic thesis, see 'Multiple, redundant or concurrent publication' for more information), that it is not under consideration for publication elsewhere, that its publication is approved by all authors and tacitly or explicitly by the responsible authorities where the work was carried out, and that, if accepted, it will not be published elsewhere in the same form, in English or in any other language, including electronically without the written consent of the copyright- holder. To verify originality, your article may be checked by the originality detection service Crossref Similarity Check. Preprints Please note that preprints can be shared anywhere at any time, in line with Elsevier's sharing policy. Sharing your preprints e.g. on a preprint server will not count as prior publication (see 'Multiple, redundant or concurrent publication' for more information). Use of inclusive language Inclusive language acknowledges diversity, conveys respect to all people, is sensitive to differences, and promotes equal opportunities. Content should make no assumptions about the beliefs or commitments of any reader; contain nothing which might imply that one individual is superior to another on the grounds of age, gender, race, ethnicity, culture, sexual orientation, disability or health condition; and use inclusive language throughout. Authors should ensure that writing is free from bias, stereotypes, slang, reference to dominant culture and/or cultural assumptions. We advise to seek gender neutrality by using plural nouns ("clinicians, patients/clients") as default/wherever possible to avoid using "he, she," or "he/she." We recommend avoiding the use of descriptors that refer to personal attributes such as age, gender, race, ethnicity, culture, sexual orientation, disability or health condition unless they are relevant and valid. These guidelines are meant as a point of reference to help identify appropriate language but are by no means exhaustive or definitive. Please note that changes to the list of contributors are not permitted after the article has been accepted. Authors' affiliations must be the institutions where the research presented in the article took place. Please note that changes to the author affiliations are not permitted once the corrected proof is published online. Author contributions For transparency, we encourage authors to submit an author statement file outlining their individual contributions to the paper using the relevant CRediT roles: Conceptualization; Data curation; Formal analysis; Funding acquisition; Investigation; Methodology; Project administration; Resources; Software; Supervision; Validation; Visualization; Roles/Writing - original draft; Writing - review & editing. Authorship statements should be formatted with the names of authors first and CRediT role(s) following. More details and an example Changes to authorship Authors are expected to consider carefully the list and order of authors before submitting their manuscript and provide the definitive list of authors at the time of the original submission. Any addition, deletion or rearrangement of author names in the authorship list should be made only https://service.elsevier.com/app/home/supporthub/publishing/ https://www.elsevier.com/about/policies/publishing-ethics https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/declaration-of-competing-interests https://service.elsevier.com/app/answers/detail/a_id/286/supporthub/publishing/ https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/editors/perk/plagiarism-complaints/plagiarism-detection https://www.elsevier.com/editors/perk/plagiarism-complaints/plagiarism-detection https://www.elsevier.com/about/policies/sharing/preprint https://www.elsevier.com/about/policies/sharing https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/authors/journal-authors/policies-and-ethics https://www.elsevier.com/authors/journal-authors/policies-and-ethics/credit-author-statement AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 7 before the manuscript has been accepted and only if approved by the journal Editor. To request such a change, the Editor must receive the following from the corresponding author: (a) the reason for the change in author list and (b) written confirmation (e-mail, letter) from all authors that they agree with the addition, removal or rearrangement. In the case of addition or removal of authors, this includes confirmation from the author being added or removed. Only in exceptional circumstances will the Editor consider the addition, deletion or rearrangement of authors after the manuscript has been accepted. While the Editor considers the request, publication of the manuscript will be suspended. If the manuscript has already been published in an online issue, any requests approved by the Editor will result in a corrigendum. Article transfer service This journal is part of our Article Transfer Service. This means that if the Editor feels your article is more suitable in one of our other participating journals, then you may be asked to consider transferring the article to one of those. If you agree, your article will be transferred automatically on your behalf with no need to reformat. Please note that your article will be reviewed again by the new journal. More information. Copyright Upon acceptance of an article, authors will be asked to complete a 'Journal Publishing Agreement' (see more information on this). An e-mail will be sent to the corresponding author confirming receipt of the manuscript together with a 'Journal Publishing Agreement' form or a link to the online version of this agreement. Subscribers may reproduce tables of contents or prepare lists of articles including abstracts for internal circulation within their institutions. Permission of the Publisher is required for resale or distribution outside the institution and for all other derivative works, including compilations and translations. If excerpts from other copyrighted works are included, the author(s) must obtain written permission from the copyright owners and credit the source(s) in the article. Elsevier has preprinted forms for use by authors in these cases. Elsevier supports responsible sharing Find out how you can share your research published in Elsevier journals. Role of the funding source You are requested to identify who provided financial support for the conduct of the research and/or preparation of the article and to briefly describe the role of the sponsor(s), if any, in study design; in the collection, analysis and interpretation of data; in the writing of the report; and in the decision to submit the article for publication. If the funding source(s) had no such involvement then this should be stated. Open access Please visit our Open Access page for more information. Elsevier Researcher Academy Researcher Academy is a free e-learning platform designed to support early and mid-career researchers throughout their research journey. The "Learn" environment at Researcher Academy offers several interactive modules, webinars, downloadable guides and resources to guide you through the process of writing for research and going through peer review. Feel free to use these free resources to improve your submission and navigate the publication process with ease. Language (usage and editing services) Please write your text in good English (American or British usage is accepted, but not a mixture of these). Authors who feel their English language manuscript may require editing to eliminate possible grammatical or spelling errors and to conform to correct scientific English may wish to use the English Language Editing service available from Elsevier's Author Services. Submission Our online submission system guides you stepwise through the process of entering your article details and uploading your files. The system converts your article files to a single PDF file used in the peer-review process. Editable files (e.g., Word, LaTeX) are required to typeset your article for final publication. All correspondence, including notification of the Editor's decision and requests for revision, is sent by e-mail. https://www.elsevier.com/authors/article-transfer-service https://www.elsevier.com/about/policies/copyright https://www.elsevier.com/about/policies/copyright/permissions https://www.elsevier.com/__data/assets/word_doc/0007/98656/Permission-Request-Form.docx https://www.elsevier.com/authors/journal-authors/submit-your-paper/sharing-and-promoting-your-article https://www.elsevier.com/journals/expert-systems-with-applications/0957-4174/open-access-options https://researcheracademy.elsevier.com/ https://webshop.elsevier.com/language-editing-services/language-editing/ https://webshop.elsevier.com/language-editing-services/language-editing/ AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 8 Additonal Information Papers submitted to Expert Systems with Applications, may also be posted on The Computer Science Preprint Server http://www.compscipreprints.com. Such posting on The Computer Science Preprint Server is in conformity with Elsevier copyright policy and in no way conflicts with submission to Expert Systems with Applications. PREPARATION Peer review This journal operates a single anonymized review process. All contributions will be initially assessed by the editor for suitability for the journal. Papers deemed suitable are then typically sent to a minimum of two independent expert reviewers to assess the scientific quality of the paper. The Editor is responsible for the final decision regarding acceptance or rejection of articles. The Editor's decision is final. Editors are not involved in decisions about papers which they have written themselves or have been written by family members or colleagues or which relate to products or services in which the editor has an interest. Any such submission is subject to all of the journal's usual procedures, with peer review handled independently of the relevant editor and their research groups. More information on types of peer review. Use of word processing software It is important that the file be saved in the native format of the word processor used. The text should be in single-column format. Keep the layout of the text as simple as possible. Most formatting codes will be removed and replaced on processing the article. In particular, do not use the word processor's options to justify text or to hyphenate words. However, do use bold face, italics, subscripts, superscripts etc. When preparing tables, if you are using a table grid, use only one grid for each individual table and not a grid for each row. If no grid is used, use tabs, not spaces, to align columns. The electronic text should be prepared in a way very similar to that of conventional manuscripts (see also the Guide to Publishing with Elsevier). Note that source files of figures, tables and text graphics will be required whether or not you embed your figures in the text. See also the section on Electronic artwork. To avoid unnecessary errors you are strongly advised to use the 'spell-check' and 'grammar-check' functions of your word processor. Article structure Subdivision - numbered sections Divide your article into clearly defined and numbered sections. Subsections should be numbered 1.1 (then 1.1.1, 1.1.2, ...), 1.2, etc. (the abstract is not included in section numbering). Use this numbering also for internal cross-referencing: do not just refer to 'the text'. Any subsection may be given a brief heading. Each heading should appear on its own separate line. Introduction State the objectives of the work and provide an adequate background, avoiding a detailed literature survey or a summary of the results. Material and methods Provide sufficient details to allow the work to be reproduced by an independent researcher. Methods that are already published should be summarized, and indicated by a reference. If quoting directly from a previously published method, use quotation marks and also cite the source. Any modifications to existing methods should also be described. Theory/calculation A Theory section should extend, not repeat, the background to the article already dealt with in the Introduction and lay the foundation for further work. In contrast, a Calculation section represents a practical development from a theoretical basis. Results Results should be clear and concise. Discussion This should explore the significance of the results of the work, not repeat them. A combined Results and Discussion section is often appropriate. Avoid extensive citations and discussion of published literature. http://www.compscipreprints.com https://www.elsevier.com/reviewers/what-is-peer-review https://www.elsevier.com/reviewers/what-is-peer-review https://www.elsevier.com/authors/journal-authors/submit-your-paper AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 9 Conclusions The main conclusions of the study may be presented in a short Conclusions section, which may stand alone or form a subsection of a Discussion or Results and Discussion section. Essential title page information • Title. Concise and informative. Titles are often used in information-retrieval systems. Avoid abbreviations and formulae where possible. • Author names and affiliations. Where the family name may be ambiguous (e.g., a double name), please indicate this clearly. Present the authors' affiliation addresses (where the actual work was done) below the names. Indicate all affiliations with a lower-case superscript letter immediately after the author's name and in front of the appropriate address. Provide the full postal address of each affiliation, including the country name and, the e-mail address of each author. • Corresponding author. Clearly indicate who will handle correspondence at all stages of refereeing and publication, also post-publication. Ensure that phone numbers (with country and area code) are provided in addition to the e-mail address and the complete postal address. Contact details must be kept up to date by the corresponding author. • Present/permanent address. If an author has moved since the work described in the article was done, or was visiting at the time, a 'Present address' (or 'Permanent address') may be indicated as a footnote to that author's name. The address at which the author actually did the work must be retained as the main, affiliation address. Superscript Arabic numerals are used for such footnotes. Highlights Highlights are optional yet highly encouraged for this journal, as they increase the discoverability of your article via search engines. They consist of a short collection of bullet points that capture the novel results of your research as well as new methods that were used during the study (if any). Please have a look at the examples here: example Highlights. Highlights should be submitted in a separate editable file in the online submission system. Please use 'Highlights' in the file name and include 3 to 5 bullet points (maximum 85 characters, including spaces, per bullet point). Abstract A concise and factual abstract is required. The abstract should state briefly the purpose of the research, the principal results and major conclusions. An abstract is often presented separately from the article, so it must be able to stand alone. For this reason, References should be avoided, but if essential, then cite the author(s) and year(s). Also, non-standard or uncommon abbreviations should be avoided, but if essential they must be defined at their first mention in the abstract itself. Keywords Immediately after the abstract, provide a maximum of 6 keywords, using American spelling and avoiding general and plural terms and multiple concepts (avoid, for example, 'and', 'of'). Be sparing with abbreviations: only abbreviations firmly established in the field may be eligible. These keywords will be used for indexing purposes. Abbreviations Define abbreviations that are not standard in this field in a footnote to be placed on the first page of the article. Such abbreviations that are unavoidable in the abstract must be defined at their first mention there, as well as in the footnote. Ensure consistency of abbreviations throughout the article. Acknowledgements Collate acknowledgements in a separate section at the end of the article before the references and do not, therefore, include them on the title page, as a footnote to the title or otherwise. List here those individuals who provided help during the research (e.g., providing language help, writing assistance or proof reading the article, etc.). Formatting of funding sources List funding sources in this standard way to facilitate compliance to funder's requirements: Funding: This work was supported by the National Institutes of Health [grant numbers xxxx, yyyy]; the Bill & Melinda Gates Foundation, Seattle, WA [grant number zzzz]; and the United States Institutes of Peace [grant number aaaa]. https://www.elsevier.com/authors/journal-authors/highlights AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 10 It is not necessary to include detailed descriptions on the program or type of grants and awards. When funding is from a block grant or other resources available to a university, college, or other research institution, submit the name of the institute or organization that provided the funding. If no funding has been provided for the research, please include the following sentence: This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Math formulae Please submit math equations as editable text and not as images. Present simple formulae in line with normal text where possible and use the solidus (/) instead of a horizontal line for small fractional terms, e.g., X/Y. In principle, variables are to be presented in italics. Powers of e are often more conveniently denoted by exp. Number consecutively any equations that have to be displayed separately from the text (if referred to explicitly in the text). Footnotes Footnotes should be used sparingly. Number them consecutively throughout the article. Many word processors can build footnotes into the text, and this feature may be used. Otherwise, please indicate the position of footnotes in the text and list the footnotes themselves separately at the end of the article. Do not include footnotes in the Reference list. Artwork Electronic artwork General points • Make sure you use uniform lettering and sizing of your original artwork. • Embed the used fonts if the application provides that option. • Aim to use the following fonts in your illustrations: Arial, Courier, Times New Roman, Symbol, or use fonts that look similar. • Number the illustrations according to their sequence in the text. • Use a logical naming convention for your artwork files. • Provide captions to illustrations separately. • Size the illustrations close to the desired dimensions of the published version. • Submit each illustration as a separate file. • Ensure that color images are accessible to all, including those with impaired color vision. A detailed guide on electronic artwork is available. You are urged to visit this site; some excerpts from the detailed information are given here. Formats If your electronic artwork is created in a Microsoft Office application (Word, PowerPoint, Excel) then please supply 'as is' in the native document format. Regardless of the application used other than Microsoft Office, when your electronic artwork is finalized, please 'Save as' or convert the images to one of the following formats (note the resolution requirements for line drawings, halftones, and line/halftone combinations given below): EPS (or PDF): Vector drawings, embed all used fonts. TIFF (or JPEG): Color or grayscale photographs (halftones), keep to a minimum of 300 dpi. TIFF (or JPEG): Bitmapped (pure black & white pixels) line drawings, keep to a minimum of 1000 dpi. TIFF (or JPEG): Combinations bitmapped line/half-tone (color or grayscale), keep to a minimum of 500 dpi. Please do not: • Supply files that are optimized for screen use (e.g., GIF, BMP, PICT, WPG); these typically have a low number of pixels and limited set of colors; • Supply files that are too low in resolution; • Submit graphics that are disproportionately large for the content. Color artwork Please make sure that artwork files are in an acceptable format (TIFF (or JPEG), EPS (or PDF) or MS Office files) and with the correct resolution. If, together with your accepted article, you submit usable color figures then Elsevier will ensure, at no additional charge, that these figures will appear in color online (e.g., ScienceDirect and other sites) in addition to color reproduction in print. Further information on the preparation of electronic artwork. https://www.elsevier.com/authors/author-schemas/artwork-and-media-instructions https://www.elsevier.com/authors/author-schemas/artwork-and-media-instructions https://www.elsevier.com/authors/author-schemas/artwork-and-media-instructions AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 11 Figure captions Ensure that each illustration has a caption. Supply captions separately, not attached to the figure. A caption should comprise a brief title (not on the figure itself) and a description of the illustration. Keep text in the illustrations themselves to a minimum but explain all symbols and abbreviations used. Text graphics Text graphics may be embedded in the text at the appropriate position. If you are working with LaTeX and have such features embedded in the text, these can be left. See further under Electronic artwork. Tables Please submit tables as editable text and not as images. Tables can be placed either next to the relevant text in the article, or on separate page(s) at the end. Number tables consecutively in accordance with their appearance in the text and place any table notes below the table body. Be sparing in the use of tables and ensure that the data presented in them do not duplicate results described elsewhere in the article. Please avoid using vertical rules and shading in table cells. References Citation in text Please ensure that every reference cited in the text is also present in the reference list (and vice versa). Any references cited in the abstract must be given in full. Unpublished results and personal communications are not recommended in the reference list, but may be mentioned in the text. If these references are included in the reference list they should follow the standard reference style of the journal and should include a substitution of the publication date with either 'Unpublished results' or 'Personal communication'. Citation of a reference as 'in press' implies that the item has been accepted for publication. Web references As a minimum, the full URL should be given and the date when the reference was last accessed. Any further information, if known (DOI, author names, dates, reference to a source publication, etc.), should also be given. Web references can be listed separately (e.g., after the reference list) under a different heading if desired, or can be included in the reference list. Data references This journal encourages you to cite underlying or relevant datasets in your manuscript by citing them in your text and including a data reference in your Reference List. Data references should include the following elements: author name(s), dataset title, data repository, version (where available), year, and global persistent identifier. Add [dataset] immediately before the reference so we can properly identify it as a data reference. The [dataset] identifier will not appear in your published article. References in a special issue Please ensure that the words 'this issue' are added to any references in the list (and any citations in the text) to other articles in the same Special Issue. Reference management software Most Elsevier journals have their reference template available in many of the most popular reference management software products. These include all products that support Citation Style Language styles, such as Mendeley. Using citation plug-ins from these products, authors only need to select the appropriate journal template when preparing their article, after which citations and bibliographies will be automatically formatted in the journal's style. If no template is yet available for this journal, please follow the format of the sample references and citations as shown in this Guide. If you use reference management software, please ensure that you remove all field codes before submitting the electronic manuscript. More information on how to remove field codes from different reference management software. Users of Mendeley Desktop can easily install the reference style for this journal by clicking the following link: http://open.mendeley.com/use-citation-style/expert-systems-with-applications When preparing your manuscript, you will then be able to select this style using the Mendeley plug- ins for Microsoft Word or LibreOffice. Reference style Text: Citations in the text should follow the referencing style used by the American Psychological Association. You are referred to the Publication Manual of the American Psychological Association, Seventh Edition, ISBN 978-1-4338-3215-4, copies of which may be ordered online. https://citationstyles.org https://citationstyles.org https://www.mendeley.com/reference-management/reference-manager/ https://service.elsevier.com/app/answers/detail/a_id/26093/ https://service.elsevier.com/app/answers/detail/a_id/26093/ https://apastyle.apa.org/products/publication-manual-7th-edition AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 12 List: references should be arranged first alphabetically and then further sorted chronologically if necessary. More than one reference from the same author(s) in the same year must be identified by the letters 'a', 'b', 'c', etc., placed after the year of publication. Examples: Reference to a journal publication: Van der Geer, J., Hanraads, J. A. J., & Lupton, R. A. (2010). The art of writing a scientific article. Journal of Scientific Communications, 163, 51–59. https://doi.org/10.1016/j.sc.2010.00372. Reference to a journal publication with an article number: Van der Geer, J., Hanraads, J. A. J., & Lupton, R. A. (2018). The art of writing a scientific article. Heliyon, 19, Article e00205. https://doi.org/10.1016/j.heliyon.2018.e00205. Reference to a book: Strunk, W., Jr., & White, E. B. (2000). The elements of style (4th ed.). Longman (Chapter 4). Reference to a chapter in an edited book: Mettam, G. R., & Adams, L. B. (2009). How to prepare an electronic version of your article. In B. S. Jones, & R. Z. Smith (Eds.), Introduction to the electronic age (pp. 281–304). E-Publishing Inc. Reference to a website: Powertech Systems. (2015). Lithium-ion vs lead-acid cost analysis. Retrieved from http://www.powertechsystems.eu/home/tech-corner/lithium-ion-vs-lead-acid-cost-analysis/. Accessed January 6, 2016 Reference to a dataset: [dataset] Oguro, M., Imahiro, S., Saito, S., & Nakashizuka, T. (2015). Mortality data for Japanese oak wilt disease and surrounding forest compositions. Mendeley Data, v1. https://doi.org/10.17632/ xwj98nb39r.1. Reference to a conference paper or poster presentation: Engle, E.K., Cash, T.F., & Jarry, J.L. (2009, November). The Body Image Behaviours Inventory-3: Development and validation of the Body Image Compulsive Actions and Body Image Avoidance Scales. Poster session presentation at the meeting of the Association for Behavioural and Cognitive Therapies, New York, NY. Journal abbreviations source Journal names should be abbreviated according to the List of Title Word Abbreviations. Video Elsevier accepts video material and animation sequences to support and enhance your scientific research. Authors who have video or animation files that they wish to submit with their article are strongly encouraged to include links to these within the body of the article. This can be done in the same way as a figure or table by referring to the video or animation content and noting in the body text where it should be placed. All submitted files should be properly labeled so that they directly relate to the video file's content. In order to ensure that your video or animation material is directly usable, please provide the file in one of our recommended file formats with a preferred maximum size of 150 MB per file, 1 GB in total. Video and animation files supplied will be published online in the electronic version of your article in Elsevier Web products, including ScienceDirect. Please supply 'stills' with your files: you can choose any frame from the video or animation or make a separate image. These will be used instead of standard icons and will personalize the link to your video data. For more detailed instructions please visit our video instruction pages. Note: since video and animation cannot be embedded in the print version of the journal, please provide text for both the electronic and the print version for the portions of the article that refer to this content. Data visualization Include interactive data visualizations in your publication and let your readers interact and engage more closely with your research. Follow the instructions here to find out about available data visualization options and how to include them with your article. Supplementary material Supplementary material such as applications, images and sound clips, can be published with your article to enhance it. Submitted supplementary items are published exactly as they are received (Excel or PowerPoint files will appear as such online). Please submit your material together with the article and supply a concise, descriptive caption for each supplementary file. If you wish to make changes to supplementary material during any stage of the process, please make sure to provide an updated file. Do not annotate any corrections on a previous version. Please switch off the 'Track Changes' option in Microsoft Office files as these will appear in the published version. https://www.issn.org/services/online-services/access-to-the-ltwa/ https://www.sciencedirect.com https://www.elsevier.com/authors/author-schemas/artwork-and-media-instructions https://www.elsevier.com/authors/tools-and-resources/data-visualization AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 13 Research data This journal encourages and enables you to share data that supports your research publication where appropriate, and enables you to interlink the data with your published articles. Research data refers to the results of observations or experimentation that validate research findings. To facilitate reproducibility and data reuse, this journal also encourages you to share your software, code, models, algorithms, protocols, methods and other useful materials related to the project. Below are a number of ways in which you can associate data with your article or make a statement about the availability of your data when submitting your manuscript. If you are sharing data in one of these ways, you are encouraged to cite the data in your manuscript and reference list. Please refer to the "References" section for more information about data citation. For more information on depositing, sharing and using research data and other relevant research materials, visit the research data page. Data linking If you have made your research data available in a data repository, you can link your article directly to the dataset. Elsevier collaborates with a number of repositories to link articles on ScienceDirect with relevant repositories, giving readers access to underlying data that gives them a better understanding of the research described. There are different ways to link your datasets to your article. When available, you can directly link your dataset to your article by providing the relevant information in the submission system. For more information, visit the database linking page. For supported data repositories a repository banner will automatically appear next to your published article on ScienceDirect. In addition, you can link to relevant data or entities through identifiers within the text of your manuscript, using the following format: Database: xxxx (e.g., TAIR: AT1G01020; CCDC: 734053; PDB: 1XFN). Mendeley Data This journal supports Mendeley Data, enabling you to deposit any research data (including raw and processed data, video, code, software, algorithms, protocols, and methods) associated with your manuscript in a free-to-use, open access repository. During the submission process, after uploading your manuscript, you will have the opportunity to upload your relevant datasets directly to Mendeley Data. The datasets will be listed and directly accessible to readers next to your published article online. For more information, visit the Mendeley Data for journals page. Data in Brief You have the option of converting any or all parts of your supplementary or additional raw data into a data article published in Data in Brief. A data article is a new kind of article that ensures that your data are actively reviewed, curated, formatted, indexed, given a DOI and made publicly available to all upon publication (watch this video describing the benefits of publishing your data in Data in Brief). You are encouraged to submit your data article for Data in Brief as an additional item directly alongside the revised version of your manuscript. If your research article is accepted, your data article will automatically be transferred over to Data in Brief where it will be editorially reviewed, published open access and linked to your research article on ScienceDirect. Please note an open access fee is payable for publication in Data in Brief. Full details can be found on the Data in Brief website. Please use this template to write your Data in Brief data article. MethodsX You have the option of converting relevant protocols and methods into one or multiple MethodsX articles, a new kind of article that describes the details of customized research methods. Many researchers spend a significant amount of time on developing methods to fit their specific needs or setting, but often without getting credit for this part of their work. MethodsX, an open access journal, now publishes this information in order to make it searchable, peer reviewed, citable and reproducible. Authors are encouraged to submit their MethodsX article as an additional item directly alongside the revised version of their manuscript. If your research article is accepted, your methods article will automatically be transferred over to MethodsX where it will be editorially reviewed. Please note an open access fee is payable for publication in MethodsX. Full details can be found on the MethodsX website. Please use this template to prepare your MethodsX article. https://www.elsevier.com/authors/tools-and-resources/research-data https://www.elsevier.com/authors/tools-and-resources/research-data/data-base-linking https://www.elsevier.com/authors/tools-and-resources/research-data/data-base-linking#repositories https://www.elsevier.com/authors/tools-and-resources/research-data/mendeley-data-for-journals https://www.journals.elsevier.com/data-in-brief/about-data-in-brief/video-discover-the-benefits-of-publishing-your-research-data https://www.elsevier.com/journals/data-in-brief/2352-3409/open-access-journal https://www.journals.elsevier.com/data-in-brief https://www.elsevier.com/__data/assets/word_doc/0004/215779/Datainbrief_template.docx https://www.journals.elsevier.com/methodsx https://www.journals.elsevier.com/methodsx https://www.elsevier.com/__data/assets/word_doc/0020/203528/MethodsX-Article-Template.docx AUTHOR INFORMATION PACK 6 Apr 2021 www.elsevier.com/locate/eswa 14 Data statement To foster transparency, we encourage you to state the availability of your data in your submission. This may be a requirement of your funding body or institution. If your data is unavailable to access or unsuitable to post, you will have the opportunity to indicate why during the submission process, for example by stating that the research data is confidential. The statement will appear with your published article on ScienceDirect. For more information, visit the Data Statement page. AFTER ACCEPTANCE Online proof correction To ensure a fast publication process of the article, we kindly ask authors to provide us with their proof corrections within two days. Corresponding authors will receive an e-mail with a link to our online proofing system, allowing annotation and correction of proofs online. The environment is similar to MS Word: in addition to editing text, you can also comment on figures/tables and answer questions from the Copy Editor. Web-based proofing provides a faster and less error-prone process by allowing you to directly type your corrections, eliminating the potential introduction of errors. If preferred, you can still choose to annotate and upload your edits on the PDF version. All instructions for proofing will be given in the e-mail we send to authors, including alternative methods to the online version and PDF. We will do everything possible to get your article published quickly and accurately. Please use this proof only for checking the typesetting, editing, completeness and correctness of the text, tables and figures. Significant changes to the article as accepted for publication will only be considered at this stage with permission from the Editor. It is important to ensure that all corrections are sent back to us in one communication. Please check carefully before replying, as inclusion of any subsequent corrections cannot be guaranteed. Proofreading is solely your responsibility. Offprints The corresponding author will, at no cost, receive a customized Share Link providing 50 days free access to the final published version of the article on ScienceDirect. The Share Link can be used for sharing the article via any communication channel, including email and social media. For an extra charge, paper offprints can be ordered via the offprint order form which is sent once the article is accepted for publication. Both corresponding and co-authors may order offprints at any time via Elsevier's Author Services. Corresponding authors who have published their article gold open access do not receive a Share Link as their final published version of the article is available open access on ScienceDirect and can be shared through the article DOI link. AUTHOR INQUIRIES Visit the Elsevier Support Center to find the answers you need. Here you will find everything from Frequently Asked Questions to ways to get in touch. You can also check the status of your submitted article or find out when your accepted article will be published. © Copyright 2018 Elsevier | https://www.elsevier.com https://www.elsevier.com/authors/tools-and-resources/research-data/data-statement https://www.elsevier.com/authors/journal-authors/submit-your-paper/sharing-and-promoting-your-article/share-link https://www.sciencedirect.com/ https://webshop.elsevier.com/article-services/article-offprints/ https://service.elsevier.com/app/home/supporthub/publishing https://service.elsevier.com/app/answers/detail/a_id/29155/supporthub/publishing/kw/status+submitted+article/ https://service.elsevier.com/app/answers/detail/a_id/5981/kw/5981/p/13783/supporthub/publishing https://service.elsevier.com/app/answers/detail/a_id/5981/kw/5981/p/13783/supporthub/publishing 
work_dxkjxmdd4rcyna6xb27nflpnky	----	LUND UNIVERSITY PO Box 117 221 00 Lund +46 46-222 00 00 An Expert System for Frequency Response Analysis Larsson, Jan Eric 1990 Document Version: Publisher's PDF, also known as Version of record Link to publication Citation for published version (APA): Larsson, J. E. (1990). An Expert System for Frequency Response Analysis. (Technical Reports TFRT-7469). Department of Automatic Control, Lund Institute of Technology (LTH). Total number of authors: 1 General rights Unless other specific re-use rights are stated the following general rights apply: Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal Read more about Creative commons licenses: https://creativecommons.org/licenses/ Take down policy If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim. https://portal.research.lu.se/portal/en/publications/an-expert-system-for-frequency-response-analysis(ce0368c2-ad12-4677-b712-49253b60f2cd).html 
work_dysnlxuejre6janvqnq2i2u33e	----	Microsoft Word - 08.doc 1 Configuration Irrigation schedule Based on Expert Systems and Operations Research Iman Hassen*, Mohamed Rasmy**, Ahmed Rafea*** *Central Lab for Agriculture Expert System, ARC Email: iman @ mail.claes.sci.eg **Faculty of Computer and Information, Cairo University Email: m_h_rasmy @ hotmail.com ***Computer Science Dept., AUC Email: rafea@aucegypt.edu Abstract: This paper presents configuration design knowledge to solve irrigation scheduling problem in agriculture domain and a general approach for configuration management knowledge. It also introduces a Configuration Agriculture Irrigation Schedule (CAIS) tool that produces the configuration irrigation schedule design according to the CommonKADS methodology. A configuration that satisfy the user requirements, and that additionally satisfy some optimality criteria is required. By integrating Dynamic programming and inference knowledge, with demand including different objective, that it is used in generating the inference steps (PPSM) to select the suitable components of the inference steps minimizing the number of components and/or maximizing the accuracy of the result of each inference step. Operations Research techniques greatly improved the understanding of the interactions of a large number of components and enhanced the decision selection under uncertain outcomes. Copyright ® 2004 IFAC Keywords: Configuration, Operations research, Expert systems, Knowledge-based systems, Knowledge acquisition, Artificial intelligence, Dynamic programming. 1. INTRODUCTION Recent approaches integrate both the Operations research (OR) and Expert Systems (ES) together under a unify framework. Although the benefits of applying OR and ES technology can be significant, developing real world extremely time consuming task. Modeling an agriculture domain requires incremental modifications to candidate schedules need to be designed and implemented to generate a configuration design includes these two approaches. Configuration design is a form of design where a set of pre-defined components is given and an assembly of selected components is sought that satisfies a set of requirements and obeys a set of constraints (Mittal, S., 1989 & Yu, B,1995). Sometimes an optimality criterion may be given that defines an ordering upon possible solutions. Following Chandrasekaran (1990) one could argue that a specification of the technology to be applied in the design process is also part of the input. In addition to the problem specification—knowledge is needed about the design domain (domain theory), additional requirements and/or constraints that are not explicitly given as input but that are implicitly given in regulations, common sense etc., optimality criteria and optionally certain preferences. Preferences are design choices which are not the outcome of the design process itself, but which are given beforehand and constrain the space of design options. Such design choices may be changed when no solution can be found within the constrained search space. We propose Configuration Agriculture Irrigation Schedule (CAIS) tool according to the CommonKADS methodology (Wielinga, 1994) which could overcome some of the limitation found in previous research efforts In the following section, a Configuration Agriculture Irrigation Schedule (CAIS) components are described, in section three the processes of CAIS tool are presented, the conclusion is in section four. 2. CONFIGURATION AGRICULTURE IRRIGATION SCHEDULE (CAIS) COMPONETS A Configuration Agriculture Irrigation Schedule (CAIS) tool consist of three main components, namely Agriculture Schedule ontology, OR library, and expert systems library. This tool produce one intermediate result called 'irrigation knowledge base' through the Knowledge Acquisition process, and it generate a configuration irrigation schedule design through the generation process as depicted in figure 1. The following subsections describe each component of the three main components in addition to the intermediate result and the final product of the tool. dt 2.1 AGRICULTUURE SCHEDULE ONTOLOGY An ontology for a knowledge based system is an explicit specification for the objects, concepts and other entities that are presumed to exist in some area of interest as well as the relationships that hold among them (Gruber T.R., 1993). In CAIS Agriculture schedule ontology classified into five parts: resources, irrigation domain ontology (activity), irrigation schedule (product), user requirement (demand), and constraint. Resources are an entity that supports or enables the execution of activities (Stephen F. 1997). Resources are generally in finite supply and their availability constraints when and how activities execute. In CAIS, we conceptualize the resource as consisting of objects (i.e. soil, water, climate, farm, previous crop, and plantation) that are elementary entities such as integers and string (i.e. soil salinity, soil texture). The activity is a set of processing steps required to produce or provide the product (irrigation schedule). It is the set of activities that fulfill the demand. The Sequences of activity are expansion, Et0, EtCrop, SHWC, Water Requirement, and Frequency/Interval to achieve the user requirement. An irrigation domain ontology (domain model) defines declaratively the set of terms and relations of a domain independent of any problem solving method. The domain model contains static information about the application environment. 2.2 OPERATIONS RESEARSH LIBRARY Operations Research library is a collection of algorithms on dynamic programming, linear programming, and pert/cpm techniques that used in irrigation schedule. The dynamic programming is used in generating the inference steps (PPSM) to select the suitable components of the inference steps minimizing the number of components and/or maximizing the accuracy of the result of each inference step. A dynamic programming formulation that used in generating inference steps described as follows: ft(st) = max/min (r(dt, st) + [ft (s t+1)]) Subject To S t+1 = gt (dt, st) St � St, dt � Dt, t = 1, 2, 3,….n Where ft(st) is the maximum expected total return r(dt, st), t=1,2,…….n finite stages of tasks, st is a finite state of a Primitive Problem Solving Method, dt, is a decision for each task, The state of the system in t+1 is described by the generation function gt (dt, st). The objective function can be changed to minimize the components as follows: ft(st) = minimum component of PPSM (inference step) when it is in PPSM (state) s with n more tasks (stages) to generate the inference step. Another dynamic programming formulation that used in a dialy irrigation task described as follows: ft(st)= max[(Yt*Pt - (Cw*qt*It + It* CE * CL) + ft (s t+1)] t=1,2,3,4 ………….T Subject To Ot = Et0(D, T, R) (1) Et = EtCrop(O) (2) User requireme Knowledge Engineer Domain Expert Irrigation Knowledge Base Knowledge Acquisition PSM Library Task Library PPSM Library Dynamic Pro. Linear Prog. PERT/CPM Operations Research Library Meta D.M. Re. Library generation Configuration Irrigation Schedule Design Agriculture Schedule Ontology Irrigation Library Expert System. Library Operations Research Library Operations Research Library Operations Research Library Figure 1: Generating Configuration Irrigation Schedule Design d=0,1 St = SHWC (3) Wt = Et + Et-1 (4) St + Rt – Wt = qt (5) E0 = 0 (6) It = (7) Where ft(st) is the maximum profit of expected yield in ton, t=1,2,….n finite stages of plant age per day, Yt is the expected yield at certain stages, Pt is the price per ton, Cw is water cost; wt is the water requirement , It is an irrigate variable taking the value 1 for irrigate or 0 if not., CE represents Energy costs for each irrigation, CL denote labor costs. Ot is a value of Et0 inference step that determined by current date, daily temperature and relative humidity that denoted by D, T, R respectively. Et is a value of EtCrop inference step that determined by plant age and Et0 that denoted by t and ot respectively. St is a value of SHWC soil available water inference step that determined by plant age. R is a rain, and qt is a dummy variables to take decision in {0, 1}, 1 means start irrigation today, 0 means wait until to next day. The whole describtion of those formulations will be explained in section 3. 2.3 EXPERT SYSTEMS LIBRARY Expert system library is a collection of problem solving method (PSM), Tasks, primitive problem solving method (PPSM), Meta domain knowledge representation, and irrigation libraries. The roles that need to be filled by a PSM prescribe what domain knowledge is expected from the expert. The domain knowledge is, therefore, no longer something that is acquired and represented independently of how it will be used in concrete problem solving. However, it is still explicitly and separately represented, giving the advantage of maintainability and systematic ness in knowledge acquisition. This library contains nine PSMs irrigation schedule The task reflects the general problem solving method for scheduling. A problem solving method can be defined as a knowledge-use-level characterization of how a task might be solved. It can also be seen as an abstract model of how to solve certain problem. In irrigation schedule design, a generic problem solving method called “propose and revise� is based on mathematics model and heuristic model. Irrigation Task library contains 18 tasks for irrigation schedule for crop. Irrigation is the main task of the irrigation generates design. The goal of the irrigation task is to get irrigation schedule of the case description. Its inputs are the case description. Its output is the irrigation schedule of the case description. 2.4 IRRIGATION KNOWLEDGE BASE The irrigation knowledge base is the result of the knowledge acquisition process. It contains all the related knowledge for the crop, Evapotranspiration (Et0) method used, EtCrop methods used, Available water in soil methods used, and water requirement methods used. 2.5 IRRIGATION SCHEDULE DESIGN Irrigation schedule design is the final product to satisfy the user requirement. There three main components of the irrigation schedule design: task knowledge, inference knowledge, and domain knowledge as described in the following subsections. 2.5.1 IRRIGATION TASK KNOWLEDGE The design of task knowledge consists of two parts: the task definition and the task body. The task definition describes the main system goal, as well as the input and the output roles. The task body describes the task type (composite or primitive), subtasks (if any), additional roles (which are not included in either input roles or output roles), and the control over these subtasks. Figure 2 shows a generated part of task knowledge of irrigation. It shows that the irrigation task contains only one task, namely: 'Propose mathematic irrigation schedule'. The input is case-description and the output is proposing irrigation schedule. Type of task is composite. The irrigation schedule subtasks involves two Transfer Tasks, one Primitive Task, and one composite task namely: 'Get case description', 'Display irrigation schedule', 'Initialize irrigation parameters', and 'Compute propose irrigation schedule' respectively. There are also two Task Static Primitive (inference step) are expand and irrigation units. 0 qt > 0 1 and Et = 0 qt <= 0, task: Irrigation schedule , task-definition: goal: the main goal of the irrigation is to determine the water requirement during cultivation. input: case-description output: irrigation schedule task_body type: Composite sub_tasks: Propose mathematic irrigation schedule control_structure: Propose mathematic irrigation schedule. task: Propose mathematic irrigation schedule task-definition: goal: Generating an propose irrigation schedule input: case-description output: Propose irrigation schedule task_body type: Composite sub_tasks: Get case description , determine plant parameters, Initialize irrigation parameters, Compute propose irrigation schedule, irrigation units, Display irrigation schedule control_structure: Get case description , case description, expansion model -> expanded case description, Initialize irrigation parameters, Compute propose irrigation schedule, water requirement, unit model -> irrigation unit, Display irrigation schedule. Figure 2: task knowledge of irrigation 2.5.2 INFERENCE KNOWLEDGE The inference describes how domain knowledge is used in elementary reasoning steps (or inferences). The inference knowledge is described using two views namely: inference structure and inference specification (Breuker J. et al., 1994). The design of inference knowledge consists of two main parts namely: inference structure and inference specification. An inference structure diagram is used to describe this type of knowledge. Three types of components are used in this diagram, namely: inference steps, dynamic roles, and static roles. An Inference specification contains five parameters: name, static role, input role, output role, and specification. A name of inferences represent the role that these inferences play in solving the problem, static roles indicate which domain model should be accessed, input and output roles are supposed to be a part of the overall working memory of the problem solver and are thus not directly linked to specific domain model. The first inference expand (see figure 3) uses the expansion model to derive a new data using the available data included in case description role. 2.5.3 DOMAIN KNOWLEDGE The domain knowledge layer consists of two parts: the set of ontology related to the application domain and domain models. Domain ontology is considered as the most fundamental constituent of domain knowledge. It defines the language of the domain, i.e., the terms used to describe the domain. This language consists of a number of constructs such as concepts, expressions, and relations (Gruber T.R., 1992). The word “ontology” can be defined as �what exist in the real life application". Generally speaking, the scheduling ontology should be mapped to five classes of concepts: resources, irrigation domain ontology (activity), irrigation schedule (product), user requirement (demand), and constraint. A domain model is defined as a coherent collection of expressions about a domain that represents a particular viewpoint defined in ontology. The domain model may therefore embody certain assumptions that are specific for the ontology that is used. Domain model have an internal structure for knowledge representation. Irrigation schedule domain model includes 9 domain models. The irrigation domain model are expansion model, Evapotranspiration(Et0) model, EtCrop Model, Water Requirement model , SHWC model, Frequency_interval_model, and Constraint Model. Each domain has its classification (i.e Evapotranspiration (Et0) model classify into Et0 Penman method, Eto modified Penman method, Et0 Penman Mountance method, Et0 Hargrve method, ET0 Blaney method, ET0 Radiation method, Et0 Penman Monteith method, ET0 Pan evaporation method, Et0 Reference based on Sector, Eto Reference based on Governor, and Et0 Reference based on Directora.). Each domain model classification contains a parts which describe the knowledge representation for this method. For example, Et0 Hargrve method consists of four parts: 1)Eto Hargerve for current month, 2) Eto Hargerve for next month, 3) Smooth factor, and 4) adaptive Eto Hargerve 3. CONFIGURATION AGRICULTURE IRRIGATION SCHEDULE (CAIS) PROCESSES There are two main processes in the CAIS tool, the first one is a knowledge acquisition, and the second is generation as shown in figure 1. The first process produces the irrigation knowledge base, and the second one produce the irrigation schedule design. The next two subsections describe those processes. 3.1 KNOWLEDGE ACQUISITION The irrigation knowledge base is the result of the knowledge acquisition process. In the knowledge acquisition process, the knowledge engineer collecting the user requirement and eliciting the knowledge using the agriculture schedule ontology and the expert system library by using a special knowledge acquisition tool that consists of five parts crop parameters, Evapotranspiration (Et0) method, EtCrop methods, Available water in soil methods, and water requirement methods. After the knowledge engineer collecting the knowledge relate to crops, the domain expert uses this tool to verify the crop parameters. 3.2 GENERATION The generation starts with the requirements definition where several questions are used to guide the generations of the irrigation schedule design for any crop. The criteria for generating design describe by the property (Attribute) and the configuration design for irrigation system is used to collect data on: o Generation model (e.g. mathematical, heuristic...) Name :Expand Goal :Derive parameters used in the system according to the available data involved in the case description role Static Role : Expansion model Input Role :Case description Output Role:Expanded case description Spec :soil.type : relation Plant age : function Figure 3: inference knowledge of irrigation o Schedule type(daily, weekly, each-10-day, monthly) o The objective of irrigation management (i.e. maximize net return, minimize irrigation costs, maximize yield, optimally distribute a limited water supply) o Scheduling problem solving containing propose and revise (i.e., propose, revise) o Events for reactive scheduling (i.e., forest, wind, rain...) o Disease (i.e., fungal...) o Yield response to water (i.e. forecasting expected yield) o Minimize the component of design (yes, no) o Evaporation method (Et0 Penman, Eto modified Penman, Et0 Penman Mountance, Et0 Hargrve, ET0 Blaney, ET0 Radiation, ET0 Pan evaporation, Et0 Reference Sector, Eto Reference Governor) The process generation is decomposed into four processes namely: generate tasks, generate inference, generate domain model, and generate ontology to produce the irrigation schedule design that contains four parts: tasks, inference, domain model, and ontology as shown in figure 4. The next subsections describe each one of them. 3.2.1 GENERATE TASK Irrigation schedule task has sub-tasks. The set of sub- tasks of a specific method have some sort of part-of relationship with the method. Each sub-task might itself be realized by different methods. Thus, the task-method decomposition is recursive until primitive inferences (or primitive tasks in our terminology) are reached, calling the mechanism of generating task by the "irrigation task schedule". There are four type of tasks: Primitive, Transfer, static Primitive, and composite task. In case of composite task, the generate task mechanism calls itself recursively to get the sub of problem solving method or primitive problem solving method. The irrigation scedule task contains two PSMs are 'Propose and revise irrigation schedule mathematical model' and 'Propose and revise irrigation schedule heuristic method'. It must select only one PSM based on the user requirement. The mechanism generate a set of task using PSM library, task library, OR library, and user requirement. Consultring the task library and match each selected task stored in the dynamic list of generated tasks to write its knowledge in the task knowledge. 3.2.2 GENERATE INFERENCE By using one of the OR library called dynamic programming to select inference step. The list of primitive tasks that generated in the priviouse section are considered as a stages. Given a sequence of task (stages) t1 through tn to be executed in linear order, where n is the number of task for specific PSM [T = {ET0, EtCrop, PAWC, Interval, Water Requirement,…..}], where ti � T. Primitive Problem Solving Method (Inference Step) is used to describe the natural property of stages (states), Each task contains more than one PPSM, it must select only PPSM (Inference Step). The finite PPSM of the system is denoted by st [sEt0 = { Et0 Penman method, Eto modified Penman method, Et0 Penman Mountance method, Et0 Hargrve method, ET0 Blaney method, ET0 Radiation method, Et0 Penman Monteith method, ET0 Pan evaporation method, Et0 Reference based on Sector, Eto Reference based on Governor, and Et0 Reference based on Directorate.}]. The decision at each task (stage) is � PSM Library Task Library� PPSM Library Meta D.M. Re. Library Generate task User requireme nt� Knowledge� Engineer Domain Expert Agriculture Schedule Ontology Irrigation Library Generate inference Generate domain model Knowledge Acquisition Set of tasks Task Knowled ge Inference Knowled Set of inference Domain Models Irrigation Knowledge Base Ontology OR Library (Dynamic Programming) Generate ontology. Figure 4: Detail Process of Generating Configuration Irrigation Schedule Design Function domain_model_ generation. Begin get List_of_inference from inference_DB(PPSM) While(not empty List_of_inference) Begin get name_PPSM from List_of_inference get static_role of name_PPSM store static_role in domain_model_DB(Domain_Model) End{ While } Get production_schedule.system = System get List_of_DM from domain_model_DB(Domain_Model, System) While(not empty List_of_DM) Begin get Domain_model from List_of_DM get System from List_of_DM getDomain_model_represention(System,Domain_model, _, Representation, Case Representation Relation: Procedure generate_relation_DM(table name(Item), input_output, position text, text) Table: Procedure enerate_table_DM(Domain_model, System) Function: Procedure generate_function_DM( Domain_model, System) End{Case} End{ While } Figure 5: Domain Model Algorithm denoted by Dt. When the process comes to PPSM (states) of some task, it can make different decision to determine the PPSM of the next task. The value of Dt(St) is used to denote the decision made of PPSM St of task T. The state generation inference knowledge of the CAIS system in t+1 is described by the function gt(Dt, St) and the return in each period rt(dt, st). The objective function is dynamic according to the user requirement. (i.e. Minimize the component of PPSM result; maximize the accuracy of PPSM result,…). Each PPSM (state) has the path (arcs) to represent the return value of the path and next PPSM (state). We add for each PPSM (state) the knowledge base concerning with each return value based on the objective function. Constraints are the most important part of every scheduling problem. There are two type of constraint. First, a set of constraints on the value of resources. Second, constraint model, a set of constraint on the propose irrigation schedule according to objective function to adaptive the irrigation schedule to satisfy on of the following objective maximize net return, minimize irrigation costs, maximize yield and optimally distribute a limited water supply. 3.2.3 GENERATE DOMAIN MODEL Generate domain model uses irrigation knowledge base, meta domain model representation, irrigation library, and set of inference as input and produces the domain model. Figure 5 depict the algorithm that used to generate the domain model. 4. CONCLUTION Integrating Operations research and Expert Systems under a unify framework in agriculture domain is very important task specially in scheduling system like irrigation system to generate a configuration design includes these two approaches. This paper presents configuration design knowledge to solve irrigation scheduling problem in agriculture domain and a general approach for configuration management knowledge. We present a Configuration Agriculture Irrigation Schedule (CAIS) tool that produces the configuration irrigation schedule design according to the CommonKADS methodology. This configuration satisfes the user requirements, and some optimal required criteria. By integrating Dynamic programming and inference knowledge, including different objective, to select the suitable components and minimize the number of components and maximize the accuracy of the result. 5. REFERENCES Breuker, J., and Van de Velde, W. (1994). CommonKADS Library for Expertise Modeling: Reusable Problem Solving Components. IOS-press, Amsterdam, August 1994. Chandrasekaran, B. (1990). Design problem solving: A task analysis. AI Magazine, Winter issue, pages 59–71. Gruber T.R., (1993). A translation approach to portable ontology specifications, Knowledge Acquisition. 5 199–220. Gruber T.R. (1992). Ontolingua: a mechanism to support portable ontologies. Technical Report KSL 91-66 (revision), Knowledge Systems Laboratory, Stanford University, March 1992. Mittal, S. & Frayman, F. (1989). Towards a generic model of configuration tasks. In Proceedings of the 11th IJCAI, pages 1395–1401, San Mateo, CA, Morgan Kaufman. Smith, S.F., O. Lassila and M. Becker (1996). A Scheduling Ontology: Informal Concept Descriptions , ICLL Working Paper, January, 1996. Stephen F. Smith and Marcel Becker (1997). An Ontology for Constructing Scheduling Systems, AAAI Spring Symposium on Ontological Engineering, Stanford, CA, March, 1997 (AAAI Press). Yu, B. and MacCallum, K. (1995). Modelling of Product Configuration Design and Management by Using Product Structure Knowledge. Int. Workshop on Knowledge Intensive CAD, Finland, 1995. Wielinga, Bob J. (1994) Expertise Model Definition Document, ESPRIT Project P5248 KADS-II, Document Id.: KADS-II/M2/UvA/026/5.0, University of Amsterdam, 1994. 
work_dywup6up6rc7tp7op74xhhlw4q	----	Carnegie Mellon University research repository - Browse Browse Discover research from Carnegie Mellon University Follow AllCategoriesgroupsSEARCH Hide footerAboutHow to DepositPreparing DataDMP ToolDeposit SupportContactDeposit AgreementTermsPrivacyToolsFAQsDisclaimerSitemap figshare. credit for all your research. 
work_e33jb3yaejg5xkpmnufychoy6i	----	Microsoft Word - 41_PE_03_14_183-186_sebok PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 3/2014 183 Milan ŠEBÖK1, Miroslav GUTTEN1, Matej KUČERA1, Daniel KORENČIAK1, Tomasz N. KOŁTUNOWICZ2 University of Žilina, Zilina, Slovakia (1), Lublin University of Technology, Lublin, Poland (2) Nondestructive diagnostics of electrical systems and equipments Abstract. This paper analyses the problem of thermal sensors, it examines the principles, functions and application for a non-contact temperature measurements. The knowledge in measurements of infrared radiance allows us to use the methods of thermovision diagnostics more effectively and to localise the disturbance which determines the quality of connection in distribution of electric energy. For technical testing of electrical systems and equipment thermography is an important diagnostic method for determining of failure of electrical systems and equipment as well as for detection of worsened condition of these systems. Streszczenie. W artykule przedstawiono problematykę termicznych sensorów, zasady badań uwzgledniające bezdotykowe pomiary temperatury, ich funkcje oraz wnioski. Dotychczasowa wiedza o pomiarach w podczerwieni pozwala na efektywniejsze wykorzystanie metod termowizyjnych celem lokalizacji zakłóceń związanych z jakością połączeń odpowiadających za dystrybucję energii elektrycznej. Termografia jest nieniszczącą metodą diagnostyczną pozwalającą na badania instalacji elektrycznych i sprzętu w celu określenia uszkodzeń urządzeń i systemów elektrycznych, a także do wykrywania pogarszania się ich stanu. (Nieniszcząca diagnostyka systemów elektrycznych i urządzeń). Keywords: Thermovision, radiation, calculation, thermogram. Słowa kluczowe: Termowizja, promieniowanie, kalkulacja, termogram. doi:10.12915/pe.2014.03.41 1. Introduction In measurement of electrical equipment and wires we deal with warming of contacts, switches, power cables, clamps, contacts of fuses. In electrical substation, temperature of each object is measured, focusing on the expansion joints, junctions, bends and coats drivers. Thermovision is used to measure warming of connections and clamps in electrical machines, as well as to the measurement of electrical equipments in the internal and external electrical distributing systems. Fundamental for a non-destructive diagnostics of electrical equipments using thermovision, is the ability to record and to transform infrared radiation (heating) to the form of real thermal images of objects, for a detection of a failure (defect) [6]. Infrared radiation is generated as a result of physical processes that take place in the object of radiation; moving atoms, molecules, vibration in crystal lattice, and transition of electrons from one energy level to another. The basic source of infrared radiation is elevated temperature of the source of radiation [3]. Radiation of hot sources acts (in respect of surrounding conditions), like visible light. To display temperature fields we can use visualization techniques used in optics. The only differences are materials used for elements of visualization systems, size of values which are derived from the wavelength of radiation, and also sensitivity of sensors for recording the signal [6]. The surface of the measured object in a state of thermodynamic equilibrium emits electromagnetic radiation and the radiated power depends on the thermodynamic temperature and properties of the surface object. Fig.1. Termogram and real picture of transformer For thermovision diagnostics of infrared radiation in the inside distribution of electric energy, we need to take into account many important factors affecting the accuracy of measurement. Results of the measured values of specific electric contact are often biased by measurement defects. In determining the classification of degrees to correct the defects, it is necessary to correct measured values due to disruptive effects of other objects [1]. 2. Theory Thermal Sensors are equipment, which transform input physical quantity (temperature) to electric signal. Transformed signal can be used in automatic control systems for data records elaboration and archivation [2]. Infra-radiation, which is absorbed to the active part of a sensor, increases its temperature. Temperature changes in the sensitive part of a sensor are represented by a relative slow process. In Fig.2 is a model – by construction of the thermal sensor [5]. Fig.2. Model – Simple construction of thermal sensor: 1 - sensitive part of sensor with temperature TD, active area of sensor S and thickness d, 2 - thermal bridge with thermal conductivity G, 3 - base with temperature TA If the sensitive part of sensor is characterised with heat capacitance C, absorbtance α and it is conducted surround area with temperature TA, with thermal conductivity G and active sensor is radiated with radiance current Φ(t), that changes with time, then if holds: if Φ(t) = 0 then TD = TA (1) if Φ(t) > 0 then TD > TA The state of the thermal balance occurs when the absorbed energy is equal conductive thermal energy of the thermal bridge [5]. By the following differential equation is described: G 3 21S Φ(t) TA 184 PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 3/2014 T0 ΔTm ΔT(t) 0 t (2)       tTGtT dt d Ct  , where: ∆T = TD -TA is temperature increase of sensitive part of sensor, when this part is heated with absorbed input radiance current αΦ(t). Let the radiance input current change in time (3) Φ(t) = Φ0 + Φm e jωt The result for the temperature increase is (4) ∆T(t) = T0 + ∆Tm e j(ωt-φ) The first term is the d.c. part and the second term is the harmonic part this equation [6]. On the basis of formulations (3), (4) we can design electrical (substitute) structure of thermal sensor Fig.3. Fig.3. Electrikal structure of thermal sensor Temperature increase ∆T is delayed against input radiance current Φ(t). For amplitude ∆Tm and shift movement period φ we write (5)  22 CG T mm      and        G C arctg   Fig.4. Input radiance current and delay of the temperaturel of thermal sensor 3. Radiation Heating is defined by the relationship α/ε, where α is the absorption coefficient of energy and ε is the emission coefficient (emissivity) of the measured body [1]. Ratio of intensity radiation of actual body and ideal black body at the same temperature is defined by spectral coefficient of emissivity [8]: (6)    TH TH T , , ),( 0        It is clear that the coefficient of spectral emissivity is equal to the spectral absorption coefficient. By Kirchhoff’s law the black body is an ideal emitter. Plank defines the spectrum of black body radiation [8]: (7) 1 2),( 52    kT hc e hc d TdH     Plank’s law is a function of spectral distribution of values. Win’s law clearly defines the shift of visible and invisible body radiation (when it is heated) to the side of the shorter waves. Stefan-Boltzmann’s law, as an integration of Planck's law by λ, defines an integral radiant flux density of black body at the temperature T [3]: (8)   4 0 /),( TddTdHH T     where: σ = 5,67.10-8 W/m2K4 – Boltzmann constant. Flux density of blackbody radiation (Fig.5) on the range of wavelengths λa, λb is received by integrating Planck equation by λ [6]. (9)   4/),( TddTdHH b a T      Fig.5. Flux density of black body radiation Derivation of Planck’s equation by temperature dT gives the change of spectral flux density emitted from black body as a function of temperature [6]: (10)       d dH eT ekhc T ddH kThc kThc . 1 )/(/( )/(2 )/(     Real objects generally do not behave as black bodies. No-black bodies absorb only a part of α(λ)Φ (incident radiation), part of the reflected radiation ε(λ)Φ and part τ(λ)Φ is a transient radiation. If the system is in thermodynamic equilibrium, according to conservation of energy, the reflected and transient energy is equal to the energy absorbed. Emissivity ε(λ) (coefficient of radiation), compensates absorption coefficient α(λ) then ε(λ) = α(λ). It follows that: (11) 1)()()(   The result of object temperature measurement T0, which is registered in the spectral range of wavelengths ∆λ ΔT(t) C αΦ(t) G αΦ0 Φm Φ(t) Φ0 0 t PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 3/2014 185 (surface density of radiant flux), is the registered radiant flux density Hreg [6]: (12)             ddTdH ddTdH ddTdHH ff aareg          ),()( /),()( /),()( 00 When an object is transparent τ(λ) = 0 and if T0 is much larger than Ta, the first part of the equation is very small. In this case the task is easier and it is essential to know ε0(λ). Difficulties arise when the body is surrounded by other objects, which have high temperature and these temperatures are higher than the examined object. In this case, its own radiation depends on the T0 and ε0 affected by reflected radiation error caused by parasitic (surrounding) objects with a temperature Te and emissivity εe. For measurements of this type it is necessary to know ε0 and T0 parameters and the number of equations, which are equal to number of unknown parameters. Radiation of measured object is formed by the sum of two parts; its own H1 radiation and parasitic H2 radiation in the infrared spectral range: (13)               1 1 1 /),()( /),()( /),()()( 2 00 1       ddTdHH ddTdH ddTdHSH ee eee where: S - geometric parameter which depends on the distance of two objects and on their surfaces. 4. Experimental In measurement of electrical equipment and wires we deal with warming of contacts, switches, power cables, clamps, contacts of fuses. In electrical substation temperature of each object is measured, focusing on the expansion joints, junctions, bends and coats drivers. A circuit containing conductor was set up in the laboratory. Fig.6. Thermogram of breaker The conductor was divided into three parts connected to the staples. The three phase breaker was derived from harmonic current. The thermogram of the circuit is on the Fig. 6. The measured circuit was load by the current from 5% to 100% of the conductor's nominal current load [7]. The measurements were taken under the following conditions:  clean and well closed terminal,  loose terminal (terminals were loose and conductors were connected only through their asymmetric position),  loose terminal and staple and conductor connection was tained with oil and sand. Measured results are graphicaly illustrated on the Fig. 7. Staple connected to the conductor cant be warmer than free conductor, if it is not damaged. On the picture we can see the negative area of the staple temperature increase compared with the closed conductor. The result of the smaller staple temperature is caused by the bigger load. Fig.7. Relation of the staple temperature increase and current load The final data of the staples temperature increase compared to the conductors are illustrated on the Fig.7 which means ΔT = TSV – TL, (TSV is the terminal temperature and TL is conductor temperature) and they were obtained by the temperature measurements of the electric joints and conductors. Fig.8 Termogram high voltage breakers On the Fig. 8 we can see the thermogram of measured object at a temperature T1 and emissivity ε1 which we want to know (radiant 1. and 3. phase with temperature T1 and T3) and from the other side we can see parasitic object with temperature Te, which is larger than T1 and T3 (radiant phase 2). Emissivity εe of parasitic object is high and the distance from measured object d is small. The temperature value Te and emissivity εe is unknown. The thermal camera distinguishes this different temperature of objects, i.e. temperature, which would have absolutely black body in this spectral range. The result of calculated equation is the temperature of 2. phase (parasite object) Te = 352.65 K. Value of calculated 186 PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 3/2014 temperature Te is near to measured temperature Te = 352.65 K. The size of radiation flux density of parasite object BR2 (εe = 0.96 and temperature Te = 352.65 K) is: (14)     1 /),(   ddTdHH eee Then the radiant flux density of the 1 and 3 phase measured object is: (15)     reflectionddTdHS radiationddTdHH eee           /),()1( /),( 00 If S = 1 then we have result calculated temperature of measured object (1. phase of breaker) with T1 = 293.15 K, and (3. phase of breaker) with T3 = 294.45 K For Measured data: F1: T1 = 327.45, K = 54.30ºC, ε1= 0.96 F2: Te = 352.65, K = 79.50ºC, εe= 0.96 F3: T3 = 338.19, K = 65.04ºC, ε3= 0.96 Following values were calculated: F1: T1 = 303.25, K = 30.10ºC, ε1 = 0.85 F2: Te = 348.45, K = 75.30ºC, εe = 0.92 F3: T3 = 306.19, K = 38.04ºC, ε3 = 0.89 5. Conclusion Comparing the results of calculated and measured values; we see that real measured temperature values are influenced by parasite object. The differences between the calculated and measured values are illustrated on the graph (Figs. 9, 10, 11). On the graph we see measured and calculated temperature differences of breaker phase F3 at the current load. Measured temperature of breaker (phase F1 and F3) is higher than calculated because close parasite object breaker F2 influences its temperature. As we can see on the graph (Fig. 6) temperature differences depend on the value of current load (In). Experimental measurements and mathematical calculations show the advantage of thermovision application on the illustrated diagnostics of the conductors and staples temperature increase. It is necessary to consider the parasitic effects of warming measured electrical systems. Fig.9. Measured and calculated data (phase F1) Fig.10. Measured and calculated data (phase F2) Fig.11. Measured an calculated data (phase F3) This work was supported by the Grant Agency VEGA from the Ministry of Education of Slovak Republic under contract 1/0624/13. Dr. T.N. Koltunowicz is a participant of the project: “Qualifications for the labour market – employer friendly university”, cofinanced by European Union from European Social Fund. REFERENCES [1] B e n k o I . , Determination of the Infrared surface Emisivity, Budapest, 1990 [2] P e n d e r C . W . , R o u x J . A . , Microcomputer System for controling and Infrared Scaning Camera. Microcomputer in optical System, USA,1999 [3] T o t h , D . , Infrared System Helps with Energy Efficiency, USA, 1995 [4] K l a b a c k a E . , Surface modifications for Thermovision Measurement, ČVUT, Prague 2001 [5] L y s e n k o V . , Detectors for noncontact temperature measurement, Prague 2005 [6] Š e b ö k M . , G u t t e n M . , K uč e r a M . , Diagnostics of electric equipments by means of thermovision, Przeglad Elektrotechnicny, 87 (2011), n. 10, 313-317 [7] K ú d e l čí k J . , B u r y P . , D r g a J . , K o pča n s k ý P . , Z á v i š o v á V . , T i m k o M . , Structure of transformer oil-based magnetic fluids studied using, Journal of Magnetism and Magnetic Materials,326 (2013), n.1, 75-80 [8] Š i m k o M . , C h u p áč M . , The theoretical synthesis and design of symmetrical delay line with surface acoustic wave for oscillators with single-mode regime of oscillation, Przeglad Elektrotechniczny, 88 (2012), n.12A, 347-350 Authors: Dr. Milan Šebök, Ph.D. (Eng), Prof. dr. hab. Miroslav Gutten, Ph.D. (Eng), Dr. Matej Kučera, Ph.D. (Eng), Dr. Daniel Korenčiak, Ph.D. (Eng), Department of Measurement and Application Electrical Engineering, Faculty of Electrical Engineering, University of Žilina, 1, Univerzitná Str, 01026 Žilina, E-mail: milan.sebok@fel.uniza.sk, gutten@fel.uniza.sk; Dr. Tomasz N. Kołtunowicz, Ph.D. (Eng), Faculty of Electrical Engineering and Computer Science, Department of Electrical Apparatus and High Voltages Technology, Lublin University of Technology, 38a, Nadbystrzycka Str., 20-618 Lublin, Poland, E-mail: t.koltunowicz@pollub.pl. 0 10 20 30 40 50 60 5 15 25 40 50 60 70 80 90 10 0 measured calculated In, % T, ºC 0 20 40 60 80 100 5 15 25 40 50 60 70 80 90 10 0 measured calculated In, % T, ºC 0 10 20 30 40 50 60 70 5 15 25 40 50 60 70 80 90 10 0 measured calculated In, % T, ºC 
work_e42y4jqinjft5e4hy5lzm4utsy	----	PII: S0957-4174(99)00028-7 Intelligent query agent for structural document databases q M.F. Jiang, S.S. Tseng*, C.J. Tsai Department of Computer and Information Science, National Chiao Tung University, Hsinchu 300, Taiwan, ROC Abstract Querying a database for document retrieval is often a process close to querying an answering expert system. In this work, we apply the expert system techniques to the intelligent query agent establishment and regard the structural document database as the expertise which can be the objective of the knowledge acquisition. A new knowledge representation, named Structural Documents (SDs), is proposed to be the base of our model, and a transformation process from the raw data to the format of a database is applied. Based on the SDs, more suitable results could be inferenced rapidly by inference engine, and the flow of inference is also described. For implementation, an intelligent Chinese information retrieval system for personnel regulations by integrating knowledge-based and full-text searching techniques is proposed. In our experiments, the structural information of the documents can be acquired from the database using the knowledge extraction module. By observing the operating process of users, we found the query process of users are simplified. q 1999 Published by Elsevier Science Ltd. All rights reserved. Keywords: Information retrieval; Document database; Intelligent agent; Expert system; Hierarchy clustering 1. Introduction Nowadays, database systems are useful in business and industrial environments due to their multiple applications. A lot of database systems are built for storing documents, say document databases, and deserves more attention. Recently, due to the rapid growth of the Internet and hardware perfor- mance, research relating to digital library has become an important issue. A major portion in the digital library is the electronic books which have the advantage of saving a lot of retrieval time. There has been a recent trend to publish electronic books excepting hard copy. Especially in the professional field, the reference manuals are usually preserved as the document databases in order to increase the convenience to query. If there is a database system, which can be modeled as shown in Fig. 1, users could formulate the query to retrieve documents, and reformulate the query when they see the results, and so on, until satisfied with the answer (Navarro & Baeza-Yates, 1997). The query effectiveness depends upon user’s knowledge about the query language. In order to improve the convenience of query for the traditional database system, we wanted to design a query agent which can be used to transform user’s demand into a query format, coordinate with a user’s request and revise the query interactively. For example, querying the personnel regulations for civil servant in Taiwan seems to be very sophisticated, and a lot of access time is required. Building a query agent is one of the solutions to such problems. Moreover, we hope that the agent provided an adaptability for different document databases. Therefore, we built an intelligent query agent (IQA), which is illustrated in Fig. 2, to assist users generating suitable queries and adjust the queries according to user’s demand (Riecken, 1994). Since querying a database for document retrieval is often a process close to querying an answering expert system (ES) (Celentano, Fugini, & Pozzi, 1995), in this work, we apply the ES techniques to the IQA establishment. In building IQA by the ES approach, we are concerned about the construction of the knowledge base, including the knowl- edge representation and the method of knowledge acquisi- tion. A document database consists of a large number of elec- tronic books. In addition to the electronic form of content stored in the database, some structural information, i.e. the chapter/section/paragraph hierarchy, may also be embedded in the database. Classical information retrieval usually allows little structuring (Navarro & Baeza-Yates, 1997), since it retrieves information only on data. However, the structural information is useful in querying the document database, for instance, most people always read books with a chapter-oriented concept. In accordance with the Expert Systems with Applications 17 (1999) 105–113PERGAMON Expert Systems with Applications 0957-4174/99/$ - see front matter q 1999 Published by Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 7 - 4 1 7 4 ( 9 9 ) 0 0 0 2 8 - 7 www.elsevier.com/locate/eswa q This work was supported by the National Science Council of the Republic of China under Grant No. NSC88-2213-E-009-078. * Corresponding author. Tel: 1 886-3-5712121, ext. 56658; fax: 1 886-3-5721490. E-mail address: sstseng@cis.nctu.edu.tw (S.S. Tseng) document structure, including the index and the table of content, it can be regarded as the expertise and also the objective of the knowledge acquisition. In order to elicit the knowledge embedded in the docu- ment structure, we first proposed a new knowledge repre- sentation, named Structural Documents (SD) (Jiang, Tseng, & Tsai, 1999), to be the basis of IQA. Second, our idea is to design a process of transforming the documents into a set of structural documents, which merge two documents with similarity greater than the given threshold into one struc- tural document. Based on this idea, we developed a cluster- ing-based approach to construct the SD in this work. Moreover, baed on the SD, more suitable results could be inferenced rapidly by the inference engine, and the flow of inference is also described. Besides, the architecture of the IQA and whole structural document database system is also proposed. As we know, a sound legal system and complete regula- tions are usually of great importance for a government by law. Right now, the personnel regulations for civil servant in Taiwan seem to be very sophisticated. Although several kinds of reference books about personnel regulations have been provided for the general public to inquiry, they are not easy to use and a lot of access time is required. Therefore, we design an intelligent Chinese information retrieval system for personnel regulations by integrating ES- based and full-text searching techniques. Our system may help users to retrieve the required personnel regulations. As mentioned above, a transformation process from the raw data to the format of a database is first applied. The embedded knowledge of the resulting data are then elicited by applying clustering techniques. By this way, the semantic indexes of the raw data can be established, and suitable results may be obtained. In our experiments, the structural information of the documents can be acquired from the database using the knowledge extraction module. By obser- ving the operating process of users, we found that the query processes of users are simplified. The organization of the rest of paper is described as follows. Reviews of related work are first described in Section 2. The knowledge extractor process we proposed is presented in Section 3. Section 4 presents the inference process of the IQA. Section 5 demonstrates the implemen- tations of this approach. Finally, the conclusion and future work are discussed in Section 6. 2. Related works An ES is a knowledge-base system to solve problem at the level of a human expert (Giarratano & Riley, 1993), which generally consists of three modules: user interface, knowledge base and inference engine. User interface helps users to easily communicate with the ES, knowledge base contains the knowledge which forms the basis of inference and inference engine generates some conclusions according to the knowledge bases. Generally, the process of building the knowledge base, which is called knowledge acquisition, is interviewing experts by knowledge engineers. However, it may cost a lot of time, as the domain expert may not have any sense of the computer techniques or the knowledge engineer is not familiar with the domain knowledge (Hwang & Tseng, 1990). In order to achieve the goal of decreasing the inter- vention of experts and the goal of smoothing the knowledge acquisition, we established an adaptive process to transfer domain knowledge to the format of a knowledge base. The approach undertaken in IQA is based on the assump- tion that the structure of reference books is a hierarchy structure. Let us take the book’s structure, shown in Fig. 3, as an example. In Fig. 3, there are nine chapters, where Chapter 1 consists of three sections and some section may be composed of paragraphs. As most of the readers usually refer the books following the tree-like hierarchy, consider- ing the book’s structure seems to influence in query constructing. Intuitively, the documents in the same chapter seem to have very likely higher similarity rather than the documents in other chapters. Similarly, for the documents in the same chapter, the documents in the same section have very likely higher similarity rather than the documents in other sections. So, the information of the structural hierar- chy may be useful in retrieving document databases. Besides, the semantic meaning of the documents in the M.F. Jiang et al. / Expert Systems with Applications 17 (1999) 105–113106 Fig. 1. The model of the database system. Fig. 2. An intelligent query agent for the document database system. Fig. 3. The chapter structure of a book. book is expressed in the contents including words, phrases, etc. Therefore, we propose a new knowledge representation mixing contents and the structure of document database in building the IQA. 3. System structure In this section, we will first introduce our system model and then the major modules of the system will be described in detail. 3.1. System overview Fig. 4 shows the overview of the system, which consists of three components: IQA; query transformer (QT) module and knowledge extractor (KE) module; and a traditional database system. IQA and QT are used to transform the user’s demand into an accessible query for the database system, and the KE module is used to extract the knowledge from the database and then store the extracted knowledge into the knowledge base of IQA. Based on the knowledge stored in the format of SD, which is the new knowledge representation proposed in this work, the IQA module infers and revises the suitable query for user’s demand, and the QT module transforms the result of IQA into an accessible query following the query syntax supported by the database system. 3.2. IQA and KE modules The IQA module consists of user interface, inference engine and knowledge base. The user interface helps users communicate easily with the IQA using a web-based approach. The inference engine implies the results based on knowledge and the process is addressed in Section 4. Knowledge base stores the knowledge for inference, uses a new knowledge representation called SD, which is formally defined in Definition 3.1. As mentioned above, the knowledge is extracted from the database by the KE module. The flow of the knowledge extracting is described as follows as shown in Fig. 5. The content of the document database is first transformed into sets of words by partition and transformation. Besides, the index of the book are further considered as the knowledge source to identify the set of keywords for each document in the partition and transformation step. By computing each pair of documents, the similarity matrix for documents are obtained, and used as the basis for the subsequent clustering to find a similar pair of documents. After clustering, the hierarchy of struc- tural documents can be obtained. In this figure, the two kinds of existing resources, includ- ing the index and the table of contents for the books are used as the knowledge source in the KE module. All the three procedures in the KE module described in detail are as follows. 3.2.1. Partition and transformation The KE module is capable of processing English or Chinese document database. As there is no obvious word boundary in the Chinese text, an identification process for identifying each possible disyllabic word from target data- base is needed. After the disyllabic words are identified from the sentence by applying the association measure, each document can be transferred to a set of keywords. That is, for two different Chinese characters, ci and cj, the association (Liang, 1995; Sproat and Shih, 1990) of the two M.F. Jiang et al. / Expert Systems with Applications 17 (1999) 105–113 107 Fig. 4. Overview of the system. Fig. 5. Flowchart of knowledge extracting. characters can be computed by: log2 freq�CiCj� n freq�Ci� n freq�Cj� n ; where the value of the function freq( ) is the occurrence frequency of any character or two successive characters and n is the number of all the characters in the target data- base. Since the association formula is originally designed for a large corpus, the obtained associations may not good enough to identify words. Therefore, further manual turning may be required. After the disyllabic words are identified from the sentence by applying the association measure, each document can be transferred to a set of words. For example, the sentence may be segmented into and . It should be noted here that there is no need to identify words in English text, and so the above segmented method should be skipped for English text. In our approach, the indexes of the book are also used to identify the set of keywords for each document. After the execution of the above two steps, each document is trans- formed into a set of keywords and words. Therefore, the size of both sets varies depending on the document itself, because the words are identified by the association measure and the keywords are identified by comparing with the index of the reference books. 3.2.2. Similarity measure To measure the similarity between two documents, we use the following heuristics. The similarity between two documents in the same chapter is higher than that in two different chapters, and the similarity between two docu- ments in the same section is higher than that in two different sections. Without losing the generality, we assumed the whole reference book to be divided into a three-tier hierar- chy, including chapter, section and paragraph. Based upon these heuristics, the Hierarchy Dependence (HD) between two documents can be easily computed by the following procedure: Step 1: If two documents are not in the same chapter, HD ← 0 and stop. Step 2: If two documents are not in the same section HD ← �1=s�; where s is the total number of sections in this chapter, and stop. Step 3: If two documents are not in the same paragraph, HD ← 0:5 1 �1=p� , where p is the total number of para- graphs in this section, and stop. Step 4. HD ← 1 . Let the two documents be denoted as Di and Dj. In the KE module, the similarity of Di and Dj, denoted by S�i; j�, is computed by the following formula: S�i; j�� �1 2 d� p same�i; j� 1 d p HD�i; j�; where same�i; j� means the number of words and keywords which appear both in documents Di and Dj. The value is normalized by dividing the total number of keywords in documents Di and Dj, HD�i; j� means the hierarchy depen- dence of documents Di and Dj and d is an adaptive weight value with 0 # d # 1. The default value of d is 0.5, which can be adjusted by the number of chapters for a given book. For example, when the number of chapters is near to 1 or n for a book divided into n documents, we set d as the value closed to 0, as there is not much meaning in the structure. Example 3.1. Let Di � { }, j Di j � 4, and Dj � { }, j Dj j � 3. Therefore, the number of the words which appear both in documents Di and Dj is 2, and we have same�i; j� �j Di > Dj j =�j Di j 1 j Dj j� � 2=�4 1 3�: Example 3.2. Let Di � {‘intelligent’, ‘query’, ‘agent’}, j Di j � 3, and Dj � {‘database’, ‘query’,}, j Dj j � 2. Therefore, the number of the words which appear both in documents Di and Dj is 1, and we have same�i; j� � 0:2. By computing the similarity measure of all two different documents Di and Dj, a similarity matrix [Sij] can be formed by letting Sij � S�i; j�. To simplify our further discussion, we assign Sij � 0 for i $ j. The matrix becomes an upper trian- gular matrix and the diagonal elements are 0’s. 3.2.3. Clustering and structuring documents Clustering is an important step in the KE module for the purpose of structuring documents. To transfer the docu- ments into a set of structural documents, the algorithm for similarity matrix updating in hierarchy clustering (Jain & Dubes, 1988) is used. In hierarchy clustering, a lot of approaches are considered in measuring similarity of clus- ters, including the single-link and the complete-link meth- ods. It seems that the clustered hierarchy generated by the complete-link method is more balanced than the one gener- ated by the single-link method. In this work, the complete- link method is implemented. To represent the knowledge in clustering process, the definition and notation of SD are formally defined. Definition 3.1. The SD with level l is defined recursively as follows: • A document Di is a SD with level 0, denoted as SDi which also can be denoted as �0Di�0. • A pair of SD, denoted as �lSDi; SDj�l, with the greatest M.F. Jiang et al. / Expert Systems with Applications 17 (1999) 105–113108 similarity measure, which is greater than a threshold u, among all the different pairs is also a SD with level l, where l � Maximum�m; n� 1 1, m is the level number of SDi and n is the level number of SDj. Definition 3.2. The similarity S 0 between the structural documents SDi and SDj is defined as follows: 1. If SDi � �Di� and SDj � �Dj�, the similarity S 0�SDi; SDj� � S�i; j�. 2. The similarity between �SDi; SDj� and SDk is defined as S 0�SDk;�SDi; SDj�� � Q{S 0�SDk; SDi�; S 0�SDk; SDj�}, where the operator Q means ‘minimum’ for the complete-link method, or ‘maximum’ for the single- link method. Assume we have n documents, and then an n × n similar- ity matrix can be generated by our method. First, each docu- ment is assigned to be a SD, i.e. SDi ��Di� for document Di. Moreover, the similarity matrix for documents is trans- ferred to the initial similarity matrix for SDs. The elements of the similarity matrix [S 0ij] is the similarity measure of the structural documents SDi and SDj, i.e. S 0 ij � Sij. Let C � {SDi} be the set of all SDi. Based upon the Johnson’s algo- rithms (1967) for hierarchy clustering (Jain & Dubes, 1988), we propose the following procedure for updating similarity matrix. Algorithm 3.1. Step 1: Find the most similar pair of structural documents in the current similarity matrix, say pair {p; q}, where, S0p;q � Maximum{S 0i;j, for any i; j}. Step 2: Merge structural documents SDp and SDq into a new single structural document (SDp, SDq). Step 3: Delete the structural documents SDp and SDq from the set C, and insert the new structural document �SDp; SDq� into the set, i.e. C ← C 2 {SDp} 2 {SDq} 1 {�lSDp; SDq�l}, where l � Maximum�m; n� 1 1, m is the level number of SDp and n is the level number of SDq. Step 4: Update the similarity matrix by deleting the rows and columns related to structural documents SDp and SDq, and adding a row and a column corresponding to the new structural document �SDp; SDq�. Step 5: If there are no two structural documents with similarity greater than u, stop. Otherwise, go to Step 1. After clustering, the hierarchy of structural documents can be obtained according to the set C. 3.3. An example Let u � 0:1, assume we have six documents, {D1; D2; D3; D4; D5; D6}, and the similarity matrix is: D1 D2 D3 D4 D5 D6 D1 D2 D3 D4 D5 D6 0 0:8 0:4 0:5 0:4 0:1 0 0:5 0:6 0:3 0:2 0 0:7 0:8 0:1 0 0:9 0:3 0 0:2 0 2 6666666666664 3 7777777777775 Before clustering, each document is assigned to be a structural document, i.e. SDi ��0Di�0 for document Di. The set of all SDi is written as C � {SD1; SD2; SD3; SD4; SD5; SD6}, and the initial simi- larity matrix for SDs is generated as: SD1 SD2 SD3 SD4 SD5 SD6 SD1 SD2 SD3 SD4 SD5 SD6 0 0:8 0:4 0:5 0:4 0:1 0 0:5 0:6 0:3 0:2 0 0:7 0:8 0:1 0 0:9 0:3 0 0:2 0 2 6666666666664 3 7777777777775 In the first iteration, the structural documents, SD4 and SD5, are merged into (SD4, SD5), as the elementary value of the 4th row and the 5th column is the maximum. The set of all SDi is written as C � {SD1; SD2; SD3; �1SD4; SD5�1; SD6}. After the similarity matrix is updated, we get a new similarity matrix after Step 4: 1 2 3 4; 5 6 1 2 3 4; 5 6 0 0:8 0:4 0:4 0:1 0 0:5 0:3 0:2 0 0:7 0:1 0 0:2 0 2 6666666664 3 7777777775 In the second iteration, we merge the structural docu- ments SD1 and SD2 into (1SD1, SD2)1, as the elementary value of the 1st row and 2nd column is the maximum. We have the new C � {�1SD1; SD2�1; SD3;�1SD4; SD5�1; SD6}. After the similarity matrix is updated, the new similarity M.F. Jiang et al. / Expert Systems with Applications 17 (1999) 105–113 109 matrix is transformed to: 1; 2 3 4; 5 6 1; 2 3 4; 5 6 0 0:4 0:3 0:1 0 0:7 0:1 0 0:2 0 2 6666664 3 7777775 Similarly, we merge the structural documents SD3 and (1SD4, SD5)1 into (2SD3, (1SD4, SD5)1)2 in the third iteration, since the elementary value of the 2nd row and 3rd column is the maximum. We have the new C � {�1SD1; SD2�1; �2SD3; �1SD4; SD5�1�2; SD6} and the new similarity matrix is transformed to: 1; 2 3; 4; 5 6 1; 2 3; 4; 5 6 0 0:3 0:1 0 0:1 0 2 664 3 775 In the last two iterations, we have the new C � {�3�1SD1; SD2�1;�2SD3; �1SD4; SD5�1�2�3; SD6} and {�4�3�1SD1; SD2�1;�2SD3; �1SD4; SD5�1�2�3; SD6�4}. According to the set C, the hierarchy of SDs can be obtained as Fig. 6. 4. Inference process based on the structural documents After the knowledge extracting process, in the previous section, the SD C have been built for the retrieving docu- ments. Based on the SD, more suitable results could be rapidly inferenced by an inference engine. The flow of the inference process is described in Fig. 7. The query result R is generated by some searching engine. However, in most cases, the result may be not satisfied very well for the user’s demand, likely more or less. Two algorithms are designed to solve the problem. In this section, the detail of inference process will be illustrated. First, we state the notations and definitions for the following algorithms. To easily represent the set of documents, a data structure bit_map is defined as: bit_map � b1b2…bn; where bk � 1, if document Dk belongs to some set of docu- ments, and bk � 0, otherwise. Based on the data structure bit_map, for a given query result, a set of documents generated by some searching engine can be written as: R � b1b2…bn; such that bk � 1, if document Dk belongs to the query result, and bk � 0, otherwise. In the format of SDs, a pair of structural documents �SDi; SDj� means the similarity between SDi and SDj is greater than the other SDs. For any structural document SDj, the most similar structural document SDk can be found when SDi, a pair of structural documents �SDj; SDk�, has been found. To easily describe the above idea, the definitions of immediate successor and successor are illustrated as following. Definition 4.1. A structural document SDi is said to be an immediate successor of structural documents SDj, if SDi � �lSDj; SDk�l or SDi ��lSDk; SDj�l for some structural document SDk, and level l. Definition 4.2. A structural document SDi is said to be a successor of a structural document SDj, if there exists an immediate successor SDk of SDj, such that SDi is also a successor of SDk or SDi is an immediate successor of SDj. The key process, that is to find the immediate successor, is to directly scan the structural document C, and described in the following algorithm. M.F. Jiang et al. / Expert Systems with Applications 17 (1999) 105–113110 Fig. 6. The hierarchy of structural documents according to the clustering result. Fig. 7. The flowchart of the inference process. Algorithm 4.1 ((Find_immediate_successor)). Input: C, SDi Output: The immediate successor of SDi Step 1: Find SDi in the set C. Step 2: Check the next element of SDi with two cases possibly existing. Case 1: The next element of SDi is “,”, Find the structural document SDj, for some j . 0, in the next element of “SDi,” and return (k SDi, SDj)k, for k � MAX(i, j) 1 1. Case 2: The next element of SDi is “)l” for some l, Find “(l” in the preceding element of SDi and return �lSD; SDi�l, for some structural document SDj. In the Step 2 of Algorithm 4.1, there are two cases possi- bly existing. In Case 1, output ��SDi; SDj� for some SDj [ C and in Case 2, output ��SDj; SDi� for some SDj [ C. By Definitions 3.1 and 4.1, it can be easily seen that output is the immediate successor of SDi. Now, a function bit_set is defined to convert the structural document SDi into a string of bits, such that the kth bit is 1, if the document Dk belongs to SDi, and is 0, otherwise. Example 4.1. Assume the amount of all documents is 5, bit_set����D1�; �D2��;�D3���� “11100”. The ith cover for all n documents is defined, following the definition of bit_map: Ci � b1b2…bn Algorithm 4.2. Input: Any structural document SDi, number k Output: Ck; Ck11; …; Ck1m, where the number m is the minimum integer such that R # Ck1m Procedure Cover�k; SDi� SDj ← Find_immediate_successor (C, SDi) Ck ← bit_set�SDj) If NOT_EQUAL(R, AND (R, Ck)) Then Call Cover�k 1 1; SDj� End_Procedure Example 4.2. Assume we have six documents, and the hierarchy of SDs are shown in Fig. 6 and C � {���SD1; SD2�;�SD3;�SD4; SD5���; SD6�}. Now, for a user’s query, database return the result, {D2; D3; D4} and R � 011100. When Cover�2; �D3�� is called, we get C2 � 001110 and C3 � 111110. Algorithm 4.3. / p The users want to get a set of docu- ments which can be deleted from the query result. p / Input: The query result R, C Output: The set of documents which can be deleted from the query result. Procedure Less Select a Di according to R SDj ← �Di� C1 ← {Di} Call Cover�2; SDj� Case COUNT_ONE(AND�R; Ck21��- COUNT_ONE(AND(R; SUB�Ck; Ck21��) , 0: RETURN AND(R; Ck21) . 0: RETURN AND(R; SUB�Ck; Ck21�) � 0: RETURN AND(R; Ck21) or AND(R; SUB�Ck; Ck21�) determined by user End_Case End_Procedure M.F. Jiang et al. / Expert Systems with Applications 17 (1999) 105–113 111 Fig. 8. The architecture of CPRCS. Fig. 9. The main window of the CPRCS. Algorithm 4.4. / p The users want to get a set of documents, which could offer more information. p / Input: The query result R, C Output: A set of documents, which could offer more information. Procedure More Select a Di according to R SDj ← �Di� C1 ← {Di} Call Cover�2; SDj� While COUNT_ONE(AND�R; Ck�� ± 0 do Case COUNT_ONE(AND(R; Ck21�)- COUNT_ONE(AND(R; SUB�Ck; Ck21��) , 0: R ← AND(R, SUB(Ck; Ck21�) and Call More . 0: R ← SUB(R, AND(R; SUB�Ck; Ck21��), k ← k 2 1 � 0: RETURN AND(R; Ck21) or AND(R; SUB�Ck; Ck21�) determined by user and stop. End_Case RETURN Ck End_While End_Procedure Example 4.3. Assume we have six documents, and the hierarchy of structural documents are shown in Fig. 6 and C � {���SD1; SD2�;�SD3; �SD4; SD5���; SD6�}. Now, for a user’s query, database return the result, {D3; D4; D6}. For the result, we have the corresponding structural documents {SD3; SD4; SD6} generated by the IQA. When the user feels insufficiency about the result, the IQA may generate the set {SD3; SD4; SD5 ; SD6} and asks database to return docu- ment {D5}. When the user feels insufficiency again, the IQA may generate the set { SD1 ; SD2 ; SD3; SD4; SD5; SD6} and the documents {D1; D2} are returned. On the other hand, if the user feels the amount of result {D3; D4; D6} is too many, IQA generate the set {SD3; SD4} and deletes the document {D6} from the query result. If the user feels the amount of result is too many again, the IQA generates the set {SD4}, and delete the document {D3} from the query result. Example 4.4. For another user’s query, database return the result, {D3; D4; D5}. The user feels the amount of result is too many, IQA generates the set {SD4; SD5} and deletes the document {D3} from the query result. Example 4.5. For another user’s query, database return the result, {D2; D3; D4}. The user feel the amount of result is too many, IQA generates the set {SD3; SD4} and deletes the document {D2} from the query result. 5. Implementation As we know, a sound legal system and complete regula- tions are usually of great importance for a government by law. Right now, the PR for civil servant in Taiwan seem to be very sophisticated. Although several kinds of reference books about personnel regulations have been provided for the general public to inquiry, they are not easy to use and a lot of access time is required. Therefore, improvement of the methods of inquiry and annotation of the personnel regula- tions has become an important issue. Besides, by comparing the experimental results between the traditional database model and our new model, we may verify the performance of the model based on SDs for the PR document databases. We implemented a database prototype, named CPRCS (Chinese Personnel Regulations Consultation System), for PR documents and used the IQA to assist the querying process. The CPRCS architecture is followed by the system structure of Fig. 4 and shown in Fig. 8. In CPRCS, users operate the system through the web pages interface, which can be easily promoted to be M.F. Jiang et al. / Expert Systems with Applications 17 (1999) 105–113112 Fig. 10. The amount of structural documents of similarity. For instance, the amount of structural documents is 224, when the similarity between any two documents in the same SD is greater than 0.7. extended. Two different modes, with or without IQA, are provided for users as they please. By observing the operating process of users, we found the query process of users are simplified. Fig. 9 is the main window of the system. Furthermore, the relation between the chapter hierarchy and the amount of the SDs is explained in the following experiment. There are 365 documents in the target database for the experiment. All the documents are divided into 11 chapters and the total amount of sections is 171. After the processing of the KE module, the amount of SDs of different similarity is shown in Fig. 10. The figure shows the merging situation in the clustering process. The results observed from the figure are discussed in the following way. When the similarity value is between 0 and 0.2, the amount of structural documents is about 11, which is the amount of chapters. Similarly, when the simi- larity value is between 0.4 and 0.7, the amount of structural documents is about 171, which is the total amount of sections. The situation says that the hierarchy of structural documents is similar to the chapter/section hierarchy for the book. The structural information of the documents can be acquired from the database using the KE module. 6. Conclusion Classical information retrieval usually allows little struc- turing. However, the structural information is useful in querying the document database, for instance, most people always read books with chapter-oriented concept. In accor- dance with the document structure, including the index and the table of content, it can be regarded as the expertise and also can be the objective of the knowledge acquisition. Since querying a database for document retrieval is often a process close to querying an answering expert system, in this work, we apply the ES techniques to the IQA establish- ment. In building IQA by the ES approach, we are concerned about the construction of the knowledge base, including the knowledge representation and the method of knowledge acquisition. Therefore, a new knowledge repre- sentation, named SDs, is defined to construct the acquisition model for IQA, and proposed the KE model to transform the data of database into the knowledge storing in IQA. For comparing the convenience of IQA, an intelligent retrieval system, CPRCS, is implemented. By observing the operat- ing process of users, we found the that query processes of users are simplified. Besides, the experimental result has shown the structural information of the documents can be acquired from the database using the KE module. Future research will focus on several areas. First, better similarity measurements are necessary for increasing the performance of clustering. The analysis for the influence of different clustering methods is not covered in this work. The general formula proposed by Jain and Dubes (1988) includes most of the commonly hierarchical clustering method which is the basis for future work. Another signifi- cant focus on this work is to extend the model based on SD, and the goal is to allow any different kinds of documents that can be merged into the same knowledge structure in order to increase the practicability of the system. References Celentano, A., Fugini, M. G., & Pozzi, S. (1995). Knowledge-based docu- ment retrieval in office environments: the Kabiria system. ACM Trans- actions on Information System, 13 (3), 237–268. Giarratano, J., & Riley, G. (1993). Expert systems, 2. PWS Publishing Company. Hwang, G. J., & Tseng, S. S. (1990). EMCUD: A knowledge acquisition method which captures embedded meanings under uncertainty. Inter- national Journal of Man Machine Studies, 33, 431–451. Jain, A. K., & Dubes, R. C. (1988). Algorithms for clustering data, Engle- wood Cliffs, NJ: Prentice Hall pp. 58–86. Jiang, M.F., Tseng, S.S., Tsai, C.J. (1999). Discovering structure from document databases, The Third Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD-99, Beijing, China. Liang, T. (1995). The Study of Character-based Signature Methods in Chinese Text Retrieval, PhD thesis, National Chiao Tung University, Taiwan. Navarro, G., & Baeza-Yates, R. (1997). Proximal nodes: a model to query document database by content and structure. ACM Transactions on Information Systems, 15 (4), 400–435. Riecken, D. (1994). Intelligent agent. Communications of ACM, 37 (7), 18– 21. Sproat, R., & Shih, C. (1990). A statistical method for finding word bound- aries in Chinese text. Computer Proceedings of Chinese and Oriental Languages, 4 (4), 336–351. M.F. Jiang et al. / Expert Systems with Applications 17 (1999) 105–113 113 
work_e45ah6d3gffmzaq4ar5j4tdvua	----	Investigating the post-complaint period by means of survival analysis Investigating the post-complaint period by means of survival analysis Bart Larivière*, Dirk Van den Poel Department of Marketing, Ghent University, Hoveniersberg 24, 9000 Ghent, Belgium Abstract Firms increasingly view each contact with their customers as an opportunity that needs to be managed. The primary purpose of this article is to gain a better understanding of the customers’ post-complaint period. Specific focus is placed on the impact of effective complaint handling on actual customer behavior throughout the time, whereas previous research has mainly focused on time-invariant or intentional measures. Survival analysis techniques are used to investigate the longitudinal behavior of complainants after their problem recovery. The proportionality assumption is tested for each explanatory variable under investigation. In addition, the impact for each variable is estimated by means of survival forests. Survival forests enable us to explore the evolution over time of the effects of the covariates under investigation. As such, the impact of each explanatory variable is allowed to change when the experiment evolves over time, in contrast to ‘proportional’ models that restrict these estimates to be stationary. Our research is performed in the context of a financial services provider and analyses the post-complaint periods of 2326 customers. Our findings indicate that (i) it is interesting to consider complainants since they represent a typical and rather active customer segment, (ii) furthermore, it is beneficiary to invest in complaint handling, since these investments are likely to influence customers’ future behavior and (iii) survival forests are a helpful tool to investigate the impact of complaint handling on future customer behavior, since its components provide evidence of changing effects over time. q 2005 Elsevier Ltd. All rights reserved. Keywords: Data mining; Customer relationship management; Consumer complaint behavior; Actual customer behavior; Proportionality; Survival forests 1. Introduction According to Keaveney (1995), the two major reasons why customers switch service providers are: (1) core service failures and (2) unfavorable service encounters with the company’s personnel. When customers face a problem they may respond by exiting (Zswitching to another provider), loyalty (Zstaying with the supplier anticipating that ‘things will get better’) or voicing (Zcomplaining to the firm or word-of-mouth to third-parties) (Levesque & McDougall, 1996; Tax, Brown, & Chandrashekaran, 1998). Unfortu- nately, it is only the tip of the iceberg that complains to the company (Stephens & Gwinner, 1998) since dissatisfied customers tend to remain passive when experiencing a problem (Bougie, Pieters, & Zeelenberg, 2003). Customers who do not complain to the firm when dissatisfied are of special concern to management for several reasons. First, the company loses the opportunity to rectify the problem (Fornell & Wernerfelt, 1988; Levesque & McDougall, 1996) and to restore the customer’s satisfaction level (Smith, Bolton, & Wagner, 1999). Second, the firm’s reputation can be damaged due to the negative word-of-mouth to friends, family or other people external to the customer’s social circle, e.g. via newspapers (Bougie et al., 2003; Singh, 1988) which might result in the loss of prospects as well as current customers (Stephens & Gwinner, 1998). Third, the firm is deprived of valuable information about its products and services (Fornell & Wernerfelt, 1987) that is likely to improve the bottom-line performance and to prevent similar problems in the near future. On the other hand, customers who complain and receive a proper response to their service failures are more likely to stay (e.g. Conlon & Murray, 1996), to buy new products (e.g. Maxham III & Netemeyer, 2003), to pay price premiums (e.g. Zeithaml, Berry, & Parasuraman, 1996), to engage in favorable word-of-mouth and to recommend the company’s services to others (e.g. Maxham III, 2001; Maxham III & Netemeyer, 2002). Furthermore, they show higher share- of-wallet behavior (e.g. Bowman & Narayandas, 2001) as well as higher commitment and trust towards the company (e.g. Tax et al., 1998). Finally, they are less vulnerable Expert Systems with Applications 29 (2005) 667–677 www.elsevier.com/locate/eswa 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.04.035 * Corresponding author. Tel.:C32 9 264 35 24; fax: C32 9 264 42 79. E-mail address: bart.lariviere@ugent.be (B. Larivière). http://www.elsevier.com/locate/eswa to switch (e.g. Bougie et al., 2003) and less likely to spread negative word-of-mouth to friends (e.g. Blodgett, Granbois, & Walters, 1993), or third-parties, such as other customers (e.g. Zeithaml et al., 1996). In sum, there is overwhelming evidence from previous research that successful complaint handling results in favorable customer outcomes. Addition- ally, in their study Fornell and Wernerfelt (1988) state that the return on investment in complaint management is likely to reach a 400 percent level. When considering the consumer complaint behavior (CCB) literature, Stephens and Gwinner (1998) argue that much of the research is dominated by studies trying to understand why customers complain. In their paper, they provide an exhaustive list of investigated antecedents, including individual characteristics, attitudes, situational factors, the cost of complaining, etc. It is only since the last decade that literature has caught up by investigating the consequences of complaint handling (cf. previous para- graph). However, current knowledge is limited in providing insights regarding behavioral intentions or self-reported actual behavior measures resulting from critical incident technique studies in which the respondents are requested to think about their latest service switch (e.g. Keaveney, 1995). As is well known, data on actual behavior are often unavailable. Generally, the only data available are the self- reported intentions of the individuals who completed a post- complaint questionnaire. Nevertheless, many authors argue that intentions are not always translated into subsequent behavior, since respondents typically do not have perfect information about changes that may occur in the future that may affect their behavior (Young, DeSarbo, & Morwitz, 1998). Unlike previous research, we investigate the impact of complaint handling on customers’ actual behavior instead of intended behavior (Zperceptual information). As a consequence, our research setting implies the need to link complaint data with complainants’ behavioral information that is stored in transactional databases. In this study, we decided to investigate the complainants ‘next-buy’ decision. We believe that an effective purchase reflects actual retention behavior (Larivière & Van den Poel, 2004). In contrast to the studies that have investigated intended repeat purchasing behavior by questioning items such as ‘In the near future, I intend to buy new products’, we consider an actual product opening as a real and executed consequence of such an intention. The variable ‘next-buy’ expresses whether the customer has bought a new product during the observed period of analysis. The variable is operationalized as a time-varying dependent variable, in which the right-censoring situation is taken into account; that is, customers who have not bought a new product by the end of the observed period of analysis might do so in the future (that is, right-censoring). Furthermore, we explicitly test whether the impact of complaint handling varies over time by means of survival forests, meaning that we allow for changes in the impact of complaint handling components on the customers’ next-buy decision. In the context of complaints, it is plausible to assume that some effects, such as receiving compensation, fade out after a while. As such, we cannot use conventional ‘proportional’ models that assume stationary effects of the covariates throughout the observed window of observation. In sum, we contribute to the existing CCB literature by presenting a framework of actual customer behavior in which we account for the right-censoring situation, and we explicitly test for the time-varying impact of explanatory variables by questioning the proportionality assumption. The rest of this paper is organized as follows. In Section 2 we elucidate both the methodological underpinnings of the proportionality assumption and the survival forests technique. In Section 3, we present the data set and the explanatory variables under investigation. The study results and its implications are reported in Section 4. Section 5 concludes the paper and outlines some directions for further research. 2. Methodology In this study we apply survival analysis techniques because our dependent variable is characterized by both a binary classification (‘buy’ or ‘no buy’) and a duration indicator for that purchasing (or censoring) event. First, we present the methodological underpinnings related to the proportionality assumption. Next, we elaborate on the survival forests technique that produces time-varying covariate estimates. 2.1. Testing the proportionality assumption Survival analysis is a class of statistical methods modeling the occurrence and timing of events (in this case: the complainant’s next-buy decision). Survival data have the following form: fðcn; tn; xnÞ; n Z 1; .; Ng (1) where n represents the index to the 2326 (N) complainants under investigation in this study; cn is the status (or binary classification) indicator which represents whether the complainant repurchased within the period of analysis; tn is the duration indicator and represents the time to the event or the censoring time (that is, for the customers who did not experience the event of buying within the period of analysis); xn is the vector of covariates for each customer n, and refers to the complaint handling and control explanatory variables in this study (cf. Section 3.2). The goal of survival analysis is to trace the effects of the covariates on the times to the event; or in this study: the impact of complaint handling on the duration to repurchase. The field of survival analysis is dominated by the Cox proportional hazard model (Stare, Harrell, & Heinzl, 2001). B. Larivière, D. Van den Poel / Expert Systems with Applications 29 (2005) 667–677668 https://isiarticles.com/article/20879 
work_e5cubmxbwrg7feb4cmck73ne2y	----	Can Machine Learning Offer Anything to Expert Systems? Machine Learning, 4, 251-254 (1989) © 1989 Kluwer Academic Publishers, Boston. Manufactured in The Netherlands. Can Machine Learning Offer Anything to Expert Systems? BRUCE G. BUCHANAN Professor of Computer Science, Medicine, and Philosophy, University of Pittsburgh, Pittsburgh, PA 15260 Today's expert systems have no ability to learn from experience. This commonly heard crit- icism, unfortunately, is largely true. Except for simple classification systems, expert systems do not employ a learning component to construct parts of their knowledge bases from libraries of previously solved cases. And none that I know of couples learning into closed- loop modification based on experience, although the SOAR architecture [Rosenbloom and Newell 1985] comes the closest to being the sort of integrated system needed for continuous learning. Learning capabilities are needed for intelligent systems that can remain useful in the face of changing environments or changing standards of expertise. Why are the learn- ing methods we know how to implement not being used to build or maintain expert systems in the commercial world? Part of the answer lies in the syntactic view we have traditionally taken toward learning. Statistical techniques such as linear regression, for example, are "knowledge-poor" proce- dures that are unable to use knowledge we may bring to the learning task. However, learning is a problem-solving activity. As such, we can and should analyze the requirements and functional specifications, the input and output, and the assumptions and limitations. If there is one thing we have learned in AI it is that complex, i.e., combinatorially explosive and ill-structured, problems often can be tamed by introducing knowledge into an otherwise syntactic procedure. There is not a single model for learning effectively, as is confirmed by the proliferation of methods, just as there is not a single model of diagnosis or of scheduling. If the machine learning community is going to make a difference in the efficiency of building commercial expert systems, we need not only more methods but a better understanding of the criteria for deciding which method best suits which learning problem. No matter what model one adopts, there is a significant chance of success in coupling a learning program with an expert system. And there will be significant benefits to the machine learning community in terms of research problems that crystallize out of the solution of applications. The workhorse of knowledge acquisition for commercial expert systems is still knowledge engineering. It is labor intensive and prone to error because it involves discussions among persons with very different backgrounds. We are slightly better at teaching knowledge engi- neering, and practicing it, than we were a decade ago. It has the outstanding virtue that it works. Partly because it is slow, it also allows more deliberate examination of conceptual frameworks, of assumptions, and of mistakes. But it is too slow: it stands in the same rela- tion to knowledge acquisition as do paper and pencil to calculation. 5 252 E.G. BUCHANAN Speeding up the knowledge engineering process has led to the development of customized editors of the SALT variety [Marcus and McDermott 1989]. These are being used success- fully and provide some assistance to knowledge base developers. But they are not learning programs; they are editors. As an aside, though, they allow us to meet McCarthy's 1958 challenge [McCarthy 1958] of building programs that can accept advice before we build systems that learn on their own. In a customized editor we see elements of expert systems. They use knowledge of a generic task such as diagnosis to guide the acquisition and debug- ging of a knowledge base. With our increased experience in building systems over the last decade, we should now be able to integrate the techniques of TEIRESIAS [Davis 1982] and ROGET [Bennett 1985] for even more powerful editors. As we move closer to automated knowledge acquisition and away from knowledge engi- neering, we must understand better how to convey the built-in assumptions of the programs that assist in constructing knowledge bases. With expert systems, there is a potential danger that the designers of programs will have different conceptual frameworks in mind than the users. In the case of learning programs, the same potential danger arises in the mismatch of frameworks between the designers of the learning programs and the designers of the expert systems. We in the machine learning community have not adequately addressed this question, nor would we think of it unless there were a pressing need to apply machine learning techniques to problems of interest to another community. Inductive learning has received considerable attention since the 1950s, with several ap- proaches now in our growing toolkit of programs that can assist in knowledge acquisition. Some programs assume the data are correct, that the descriptions and classifications of examples are error-free; some of the same ones assume that the examples are all available at once, and hereafter remain available; others assume that classification rules are categorical; most assume that the rules to be learned are one-step conjunctive rules. Successful applica- tions of ID3 and C4 [Quinlan, et al. 1986] and their cousins have provided knowledge bases for working expert systems whose task is to classify. These, together with a few other isolated cases of an expert system being built with the assistance of a learning program, have kept alive the promise of payoff from the 35-year investment in learning research. Explanation-based learning (EBL) is now receiving even more attention than induction or similarity-based learning. Here, more than with induction, the problem-solving nature of acquiring new knowledge is apparent. (See, for example, [Mitchell 1986].) Its main assumption is that enough theory exists to provide a rationalization of why one instance is or is not a prototypical member of a class. What else? Does it have to be a formal, deduc- tive rationalization, or will an empirical argument suffice? (See [Smith, et al. 1985] for an argument that a weaker, empirical model will suffice.) Unfortunately, EBL has not found any applications in building expert systems, thus it is still demonstrated only in laboratory prototypes. Case-based and analogical reasoning are still pipe dreams when matched against the harsh standards of robustness of commercial applications. Some of the problems stem from the dilemma that we can find some mappings from any previous case or previously studied domain to the present one: we almost have to know the answer in order to find the case or the analogy that lets us program a machine to find the answer. Analogical reasoning involves at least two problem-solving steps: finding a useful analogous domain (or object or process) and finding relevant mappings between the analog and the original thing. It 6 CAN MACHINE LEARNING OFFER ANYTHING TO EXPERT SYSTEMS? 253 is not a robust method in the sense that it still produces too many false positive results. Moreover, we don't see many suggestions for making these methods more selective. Their power seems to lie in offering suggestions when we have run out of other ideas. What about neural nets? They certainly are popular because the idea of getting something for nothing has always held great appeal. They have shown some success in knowledge-poor tasks, such as character recognition and other perceptual tasks. Expert systems, by their very nature, are knowledge-intensive, however, and thus are less amenable to learning with syntactic reinforcement methods. Once a neural net is tuned, there is no way to understand why it succeeds on some cases and fails on others except to say that the weights on some nodes are higher than on others. Moreover, their construction is not free; considerable effort must be invested in laying out the structure of the network before examples can be presented. Only a few of the techniques in the literature are immediately useful, and these have their limits. Part of our research charge needs to include understanding the scope and limits of different methods and determining when they are applicable. Expert systems provide a strong rationale for continued funding of research on machine learning, but they also serve to sharpen our understanding of problems. Expert systems offer a focus for development of new machine learning methods and better understanding of old ones. Of course, we need basic research as well as demand-driven research. At the moment, however, there is an imbalance in the amount of work on learn- ing in domains where we do not need to learn and on techniques with crippling assumptions. Let us attempt to understand the context of the commercial, military, and medical needs. In research on machine learning, as on other problem-solving methods, new wrinkles— perhaps new opportunities—will arise in experimenting with real, complex knowledge bases and applications. Using an expert system as a testbed offers a tough test of success. The commercial world of expert systems at large seems unconvinced that machine learning has anything to offer yet. I strongly disagree. Inductive learning, at least, is already in limited use, and present methods can be extended to make them more useful. Some of the issues that need to be resolved in order to make inductive methods a mainstay of commercial knowledge acquisi- tion are already modestly well understood: learning in the context of noisy and uncertain data, exploiting an existing partial theory, representing the objects of learning in a form other than classification rules, and tailoring the learning to the specific context in which the learned information will be used. These are partly issues of utility, but they are impor- tant research problems as well. Machine learning is ready for development now, with atten- dant benefits to us in crystallizing current and new research problems. References Bennett, J.S. 1985. ROGET: A knowledge-based system for acquiring the conceptual structure of a diagnostic expert system. J. Automated Reasoning 1, 49-74. Davis, R. 1982. TEIRESIAS: Applications of meta-level knowledge. In R. Davis and D. Lenat (Eds.). 1982. Knowledge-based systems in artificial intelligence. New York: McGraw-Hill. 7 254 B.G. BUCHANAN Marcus, S. and McDermott, J. 1989. SALT: A knowledge acquisition language for propose-and-revise systems. Artificial Intelligence 39, 1-37. McCarthy, J. 1968. Programs with common sense. Proc. Symposium on the Mechanisation of Thought Processes. Reprinted in M. Minsky (Ed.), Semantic Information Processing. Cambridge, MA: MIT Press, 1968. Mitchell, T.M., Keller, R., and Kedar-Cabelli, S. 1986. Explanation-based generalization: A unifying view. Machine Learning 1, 47-80. Quinlan, J.R., Compton, P.J., Horn, K.A., and Lazarus, L. 1986. Inductive knowledge acquisiton: A case study. Proceedings Second Australian Conference on Applications of Expert Systems. In J.R. Quinlan (Ed.), Applications of Expert Systems. Maidenhead: Academic Press (in press). Rosenbloom, P.S., and Newell, A. 1985. The chunking of goal hierarchies: A generalized model of practice. In R. Michalski, J. Carbonell, and T. Mitchell (Eds.), Machine learning: An artificial intelligence approach (Vol. 2). Los Altos, CA: Morgan-Kaufmann. Smith, R.G., Winston, H., Mitchell, T., and Buchanan, B.G. 1985. Representation and use of explicit justification for knowledge base refinement. In Proceedings of IJCA185. Los Altos, CA: Morgan-Kaufmann. 8 
work_e5xlnx3wejcfhbmov2bysmprpa	----	06 April 2021 POLITECNICO DI TORINO Repository ISTITUZIONALE Multi-document summarization based on the Yago ontology / Baralis E.; Cagliero L.; Fiori A.; Jabeen S.; Shah S.. - In: EXPERT SYSTEMS WITH APPLICATIONS. - ISSN 0957-4174. - 40:17(2013), pp. 6976-6984. Original Multi-document summarization based on the Yago ontology Publisher: Published DOI:10.1016/j.eswa.2013.06.047 Terms of use: openAccess Publisher copyright (Article begins on next page) This article is made available under terms and conditions as specified in the corresponding bibliographic description in the repository Availability: This version is available at: 11583/2509895 since: ELSEVIER Multi-document summarization based on the Yago ontology Elena Baralis, Luca Cagliero∗, Saima Jabeen Dipartimento di Automatica e Informatica, Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129, Torino, Italy Alessandro Fiori IRC@C: Institute for Cancer Research at Candiolo, Str. Prov. 142 Km. 3.95 10060 - Candiolo (TO) - Italy Sajid Shah Dipartmento di Elettronica e Telecomunicazioni, Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129, Torino, Italy Abstract Sentence-based multi-document summarization is the task of generating a succinct summary of a document collection, which consists of the most salient document sentences. In recent years, the increasing availability of semantics- based models (e.g., ontologies and taxonomies) has prompted researchers to investigate their usefulness for improving summarizer performance. However, semantics-based document analysis is often applied as a preprocessing step, rather than integrating the discovered knowledge into the summarization process. This paper proposes a novel summarizer, namely Yago-based Summarizer, ∗Corresponding author. Tel.: +39 011 090 7084. Fax: +39 011 090 7099. Email addresses: elena.baralis@polito.it (Elena Baralis), luca.cagliero@polito.it (Luca Cagliero), saima.jabeen@polito.it (Saima Jabeen), alessandro.fiori@ircc.it (Alessandro Fiori), sajid.shah@polito.it (Sajid Shah) Preprint submitted to that relies on an ontology-based evaluation and selection of the document sentences. To capture the actual meaning and context of the document sentences and generate sound document summaries, an established entity recognition and disambiguation step based on the Yago ontology is integrated into the summarization process. The experimental results, which were achieved on the DUC’04 benchmark collections, demonstrate the effectiveness of the proposed approach compared to a large number of competitors as well as the qualitative soundness of the generated summaries. Keywords: Document Summarization, Text Mining, Entity Recognition 1. Introduction Discovering the most salient information hidden in textual Web docu- ments is often a challenging task. In fact, the huge volume of electronic documents that users could retrieve from the Web is commonly very diffi- cult to explore without the help of automatic or semi-automatic tools. To tackle this issue, a particular attention has been paid to the development of text summarization tools. Summarizers focus on generating a succinct representation of a textual document collection. Specifically, sentence-based multi-document summarizers generate concise yet informative summaries of potentially large document collections, which consist of the most representa- tive document sentences. A significant research effort has been devoted to tackling the summa- rization problem by means of general-purpose information retrieval or data mining techniques. For example, clustering-based approaches (e.g., [47, 48]) 2 adopt clustering algorithms to group document sentences into homogeneous clusters and then select the most authoritative representatives within each group. In contrast, graph-based approaches (e.g., [36, 51, 52]) first generate a graph-based model in which the similarity relationships between pairs of sen- tences are represented. Next, they exploit popular indexing strategies (e.g., PageRank [9]) to identify the most salient sentences (i.e., the most authorita- tive graph nodes). However, in some cases, the soundness and readability of the generated summaries are unsatisfactory, because the summaries do not cover in an effective way all the semantically relevant data facets. A step beyond towards the generation of more accurate summaries has been made by semantics-based summarizers (e.g., [12, 13]). Such approaches combine the use of general-purpose summarization strategies with ad-hoc linguistic analysis. The key idea is to also consider the semantics behind the doc- ument content to overcome the limitations of general-purpose strategies in differentiating between sentences based on their actual meaning and context. Ontologies are formal representations of the most peculiar concepts that are related to a specific knowledge domain and their corresponding rela- tionships [5]. Ontologies find application in several research contexts, among which user-generated content analysis [20], e-learning platform development [25], and video and image analysis [45]. In recent years, the attention of the re- search community has been focused on both learning meaningful ontologies that contain salient document keywords [8] and improving the performance of the document summarization process by integrating ontological knowl- edge [21, 24, 34, 35]. For example, ontologies have been used to identify the document concepts that are strongly correlated with a user-specified 3 query [24, 34] or to map the document content to non-ambiguous ontological concepts [21, 35]. However, most the previously proposed approaches perform the semantics-based analysis as a preprocessing step that precedes the main summarization process. Therefore, the generated summaries could not en- tirely reflect the actual meaning and context of the key document sentences. In contrast, we aim at tightly integrating the ontology-based document anal- ysis into the summarization process in order to take the semantic meaning of the document content into account during the sentence evaluation and selection processes. With this in mind, we propose a new multi-document summarizer, namely Yago-based Summarizer, that integrates an established ontology-based entity recognition and disambiguation step. Specifically, a popular ontological knowledge base, i.e., Yago [41], is used to identify the key document concepts. The same concepts are also evaluated in terms of their significance with respect to the actual document context. The result of the evaluation process is then used to select the most representative docu- ment sentences. In such a way, the knowledge that is inferred from Yago is tightly integrated into the sentence evaluation process. Finally, a variant of the Maximal Marginal Relevance (MMR) evaluation strategy [10] is adopted to iteratively choose the best subset of informative yet non-redundant sen- tences according to the previously assigned sentence ranks. To demonstrate the effectiveness the proposed approach, we compared its performance on the DUC’04 benchmark document collections with that of a large number of state-of-the-art summarizers. Furthermore, we also performed a qualitative evaluation of the soundness and readability of the generated summaries and a comparison with the results that were produced 4 by the most effective summarizers. This paper is organized as follows. Section 2 compares our approach with the most recent related works. Section 3 presents and thoroughly describes the Yago-based Summarizer system. Section 4 experimentally evaluates the effectiveness and usefulness of the proposed approach, whereas Section 5 draws conclusions and presents future developments of this work. 2. Related work A significant research effort has been devoted to summarizing document collections by exploiting information retrieval or data mining techniques. Two main summarization strategies have been proposed in literature. Sentence- based summarization focuses on partitioning documents in sentences and generating a summary that consists of the subset of most informative sen- tences (e.g., [11, 29, 47]). In contrast, keyword-based approaches focus on detecting salient document keywords using, for instance, graph-based in- dexing [28, 51, 52] or latent semantic analysis [16]. Since sentence-based approaches commonly generate humanly readable summaries without the need for advanced postprocessing steps, our summarizer relies on a sentence- based approach. Summarizers can be further classified as constraint-driven if they entail generating a summary that satisfy a set of (user-specified) con- straints [2]. For example, sentences that are pertinent to a user-specified query can be selected [30, 33]. Unlike [2, 30, 33] our summarizer relies on a constraint-less approach. Most of the recently proposed (constraint-less) sentence-based summariz- ers exploit one of the following general-purpose techniques: (i) clustering, (ii) 5 graph mining, (iii) linear programming, and (iv) itemset mining. Clustering- based approaches (e.g., [47, 48]) group document sentences into homogeneous clusters and then select the best representatives (e.g., the centroids or the medoids [32]) within each cluster. While the authors in [47] propose a static summarization framework, the work that was first presented in [48] addresses the problem of incremental summary update: whenever a set of documents is added/removed from the initial collection, the previously generated summary is updated without the need for recomputing the whole clustering model. In parallel, some attempts to cluster documents rather than sentences have also been made [37, 7]. For example, MEAD [37] analyzes the cluster centroids and generates a pseudo-document that includes the sentences with the high- est tf-idf term values [27]. Then, the sentence selection process is driven by a score that considers (i) the sentence similarity with the centroids, (ii) the sentence position within the document, and (iii) the sentence length. A sim- ilar approach has also been adopted to summarize articles coming from the biological domain [7]. To tailor the generated summaries to the most relevant biological knowledge biologists are asked to provide a dictionary that is used to drive the sentence selection process. Similarly, this work also considers the document context to improve the summarization performance. Unlike [7], it exploits an ontological knowledge base, rather than a plain-text dictionary, to drive the sentence evaluation and selection process. Graph-based approaches to sentence-based summarization (e.g., [36, 44, 46, 51, 52]) generate a graph in which the nodes represent the document sentences, whereas the edges are weighted by a similarity measure that is evaluated on each node pair. Popular indexing strategies (e.g., PageRank [9], 6 HITS [23]) are exploited to rank the sentences based on their relative author- itativeness in the generated graph. In parallel, other approaches formalize the sentence selection task as a min-max optimization problem and tackle it by means of linear programming techniques [1, 3, 17, 42]. Still others analyze the underlying correlations among document terms by exploiting (i) frequent itemset mining techniques [6], (ii) probabilistic approaches [12, 13], or (iii) the Singular Value Decomposition (SVD) [40]. Ontologies have already been exploited to improve the document sum- marization performance. Specifically, they have been used to (i) identify the concepts that are either most pertinent to a user-specified query [24, 34] or most suitable for performing query expansion [31], (ii) model the context in which summaries are generated in different application domains (e.g., the context-aware mobile domain [18], the business domain [50], the disaster management domain [26]), and (iii) enrich existent ontological models with textual content [8]. Some attempts to consider the text argumentative struc- ture into account during the summarization process have also been made. For example, in [35] the authors propose to identify and exploit salient lexical chains to generate accurate document summaries. The summarizer proposed in [21] exploits Support Vector Machines (SVMs) to maps each sentence to a subset of taxonomy nodes. Similarly, in [4] a rhetorical role is assigned to each sentence by a stochastic CRF classifier [39], which is trained from a collection of annotated sentences. Unlike [4, 21] our approach does not rely on classification models. Furthermore, since our summarizer evaluates the sentence relevance regardless of the underlying document structure, our approach is, to some extent, complementary to the ones that have previously 7 been proposed in [4, 35]. 3. Yago-based Summarizer Yago-based Summarizer is a novel multiple-document summarizer that exploits the Yago ontological knowledge base [41] to generate accurate doc- ument summaries. Consider a collection of textual documents D={d1, . . . , dN}, where each document di ∈ D is composed of a set of sentences si1, . . ., siM. The summa- rizer generates a summary S={sij} 1 ≤ i ≤ N, 1 ≤ j ≤ M. The summary includes a worthwhile subset of sentences that are representative of the whole collection D. Figure 1 outlines the main Yago-based Summarizer steps, which are briefly summarized below. • Entity recognition and disambiguation. This step analyzes the input document collection with the goal of identifying the most relevant concepts and their corresponding context of use. To this aim, the Yago knowledge base is used to map the words that occur in the document sentences to non-ambiguous ontological concepts, called entities. To discriminate between multiple candidate entities for the same word combination, it adopts an entity relevance score that considers both the popularity and the contextual pertinence of each candidate entity in the analyzed document collection. • Sentence ranking. To include in the summary only the most per- tinent and semantically meaningful document content, sentences are evaluated and ranked according to the previously assigned entity scores. 8 Figure 1: The Yago-based Summarizer. • Sentence selection. To generate a summary of the document collec- tion an iterative procedure is applied to select the top-ranked sentences that are least similar to the previously selected ones. 3.1. Entity recognition and disambiguation Entity recognition and disambiguation are established document analysis tasks that aim at mapping the natural text to a set of non-ambiguous onto- logical concepts [32]. Yago-based Summarizer exploits the Yago ontological knowledge base [41], which relies on the Wikipedia free encyclopedia [49], to support the entity recognition and disambiguation process. The Yago analytical procedures have been called through the AIDA Web Service [22]. Consider a sentence sij that is composed of a collection of (possibly re- peated) words w1, w2, . . ., wZ. The goal is to map words wk, 1 ≤ k ≤ Z 9 to Yago ontological concepts, i.e., the entities. Note that an entity may be associated either with a single word or with a combination of words. The entity recognition step recognizes the entities that are associated with noun, dates, times, or numbers. As a clarifying example, consider the sentence reported in the left-hand side of Figure 2. Figure 2: Entity Recognition and Disambiguation example. Each of the three underlined word combinations Mercury, Solar System, and Sun is associated with at least one candidate entity in Yago. Note that not all the sentence words match at least one Yago entity. For example, the word Innermost has no matching entity. Furthermore, some words have many candidate entities, meaning that a word could have different mean- ings in different contexts. For example, Mercury could be associated with the candidate entities Mercury(Element) and Mercury(Planet), which cor- respond to the well-known chemical element and planet, respectively. Note also that for each entity Yago provides (i) a popularity score, which reflects 10 its frequency of usage (e.g., 241 for Mercury(Element)), (ii) a list of related keywords (e.g., Chemistry, Liquid) for its corresponding context of use, and (iii) the number of incoming and outcoming Wikipedia links [49]. Entity recognition for times, date, and numbers is based on regular ex- pressions and returns a single entity. For example, the expression March, 1st 2012 corresponds to the date 01/03/2012. Conversely, the entity recog- nition procedure for nouns could return many candidate entities. Hence, in the latter case a disambiguation step is applied in order to select, among the candidate entities, the most appropriate one. To tackle this issue, each candidate entity is weighted by a relevance score, which considers both its popularity and pertinence to the analyzed document context. Specifically, the rank entityRank(eq) of an entity eq with respect to a word wk that occurs in the document di ∈ D, is defined as follows: entityRank(eq) = θ · popularity(eq) (1) + φ · sim(cxt(eq), cxt(di)) + (1 − θ − φ) · coh(eq, D) where θ, φ ∈ [0, 1] are user-specified parameters that weigh the importance of each summation term, popularity(eq) is the Yago popularity score that is associated with the candidate entity eq, sim(cxt(eq),cxt(di)) is the similarity between the context of use of the candidate entity and the document di, and coh(eq, D) is the coherence of eq with respect to the whole document collection. By following the indications reported in [22], we set the values of θ and φ to 0.34 and 0.47, respectively. A thorough assessment of the 11 entity recognition system performance on real data is also reported in [19]. The first summation term is a popularity score, which indicates the global frequency of occurrence of the concept in the knowledge base. For instance, in Yago Mercury-Planet has, on average, a higher popularity score than Mercury-Element (307 against 241). However, the relevance of an entity within a document also depends on its context of use. Hence, the second summation term indicates the pertinence of the entity to the document. Specifically, it measures the cosine distance [32] between the context of the candidate entity eq, i.e., the list of contextual keywords that are provided by Yago (e.g., Chemistry, Liquid for the candidate entity Mercury-Element in Figure 2), and the context of the word wk at the document level (i.e., the list of words that co-occur with wk in di). Roughly speaking, the more contextual keywords match the document content the higher the pertinence of the candidate entity is. Finally, since the recognized entities are likely to be correlated each other, the last summation term measures the coherence of the candidate entity with respect to all of the other recognized candidate entities that correspond to any word in D. Since coherent entities are likely to share many Wikipedia links, similar to [22], we evaluate the entity coherence within the document collection D as the number of incoming Wikipedia links that are shared by eq and all of the other candidate entities that have been recognized in D. The entity scores will be used to drive the summary generation process, as discussed in the following sections. 3.2. Sentence ranking Yago-based Summarizer exploits the semantic knowledge that has been 12 inferred at the previous step to evaluate and rank the document sentences according to their significance in the document collection. To this aim, a rank is associated with each document sentence. The sentence rank reflects the relevance of the entities associated with its corresponding sentence words. Let sij be an arbitrary sentence and E(s i j) the set of entities (nouns, date, times, or numbers) that are associated with any word wk ∈ sij. The sij’s rank is computed as follows: SR(sij) = ∑ eq∈E(sij) EntityScore(eq) |E(sij)| (2) where EntityScore(eq) is defined by EntityScore(eq) =   γ if eq is a date, time, or number entity, γ + EntityRank(eq) if eq is a named entity (3) γ is a user-specified parameter that is used to privilege the sentences that contain many recognized entities. ∑ eq∈E(sij) EntityScore(eq) is the summa- tion of the entity ranks of all of the entities that are mapped to any word in sij (see Definition 1). Note that the sentences that do not contain any rec- ognized Yago entity have minimal sentence rank (i.e., 0), because they are not likely to contain any semantically relevant concept. In contrast, because of the γ correction, the sentences that contain only dates, times, or numbers are considered to be, on average, more relevant than the former ones, but less relevant than those that also contain named entities. The impact of the user-specified parameter γ on the summarization performance is discussed in Section 4. 13 3.3. Sentence selection Given a sentence ranking, the selection step focuses on generating the out- put summary of the document collection by including only the most represen- tative sentences. To achieve this goal, Yago-based Summarizer adopts a vari- ant of an established iterative re-ranking strategy, called Maximal Marginal Relevance (MMR) [10]. MRR has first been introduced in the context of query-based summary generation. At each algorithm iteration, it picks out the candidate sentence that is characterized by (i) maximal relevance with respect to the given query and (ii) minimal similarity with respect to the pre- viously selected sentences. Since our approach is not query-based, we adapt the former selection strategy to the problem under analysis. Specifically, Yago-based Summarizer selects, at each iteration, the top-ranked sentence with minimal redundancy with respect to the already selected sentences. At each iteration the former optimization problem can be formulated as follows: maximize {sij} α · SR(sij) − (1 − α) · sim(s i j, s̄ r t ) subject to α ∈ [0, 1] sij /∈ S s̄rt ∈ S (4) where S is the output summary that possibly includes some of the document sentences s̄rt, α is a user-specified parameter, and {sij} is the set of candidate sentences not yet included in the summary. The sentence ranking is evaluated using the expression reported in Formula 2. Furthermore, the similarity 14 sim(sij, s̄ r t) between the pair of sentences s i j and s̄ r t is evaluated using the cosine similarity [32] and takes value zero when the summary is empty (i.e., at the first algorithm iteration). The impact of the entity relevance and the similarity score is weighted by the α parameter. Specifically, the higher the value of α is, the more important the entity relevance score is with respect to the similarity score. Therefore, setting relatively high α values could yield informative but partially redundant summaries. Conversely, for lower α values the sentence relevance is partially neglected in behalf of a lower summary redundancy. 4. Experimental results We performed a variety of experiments to address the following issues: (i) a performance comparison between Yago-based Summarizer and many state-of-the-art summarizers on document benchmark collections (see Sec- tion 4.2), (ii) a qualitative comparison between the summaries generated by our approach and those produced by two representative competitors (see Sec- tion 4.3), and (iii) an analysis of the impact of the main system parameters on the Yago-based Summarizer performance (see Section 4.4). All the experiments were performed on a 3.0 GHz 64 bit Intel Xeon PC with 4 GB main memory running Ubuntu 10.04 LTS (kernel 2.6.32-31). The source code for Yago-based Summarizer is available, for research purposes, upon request to the authors. A detailed description of the experimental evaluation context is given below. 15 4.1. Evaluation context We evaluated the Yago-based Summarizer performance on the task 2 of the Document Understanding Conference (DUC) 2004, which is the lat- est benchmark contest that were designed for generic English-written multi- document summarization [47]. The analyzed DUC’04 collections have been provided by the contest organizers [15]. They consist of a large variety of English-written articles which range over different subjects. According to their subject, articles were preliminary clustered in 50 document groups. Each homogeneous collection contains approximately 10 documents. Fur- thermore, for each collection at least one golden summary is given by the DUC’04 organizers. Participants to the DUC’04 contest had to submit their own summaries and compare them with the reference (golden) ones. The more similar the generated summaries are to the reference models, the more accurate the summarization process is. To perform an analytical comparison between the summarizers’ perfor- mance on the task 2 of DUC’04 we used the ROUGE toolkit [27], which has been adopted as official DUC’04 tool for performance evaluation1. ROUGE measures the quality of a summary by counting the unit overlaps between the candidate summary and a set of reference summaries (i.e., the golden summaries). The summary that achieves the highest ROUGE score could be considered to be the most similar to the golden summary. To perform a fair comparison, before using the ROUGE toolkit we normalized the generated summaries by truncating each of them at 665 bytes (we round the number 1The provided command is: ROUGE-1.5.5.pl -e data -x -m -2 4 -u -c 95 -r 1000 -n 4 -f A -p 0.5 -t 0 -d -a 16 down in case of straddled words). Several automatic evaluation scores are implemented in ROUGE. As previously done in [6, 47], we will report only the ROUGE-2 and ROUGE-4 representative scores [27]. Similar results were achieved for the other ROUGE scores. 4.2. Performance comparison on the DUC’04 collections We compared the Yago-based Summarizer performance on the DUC’04 benchmark collections with that of: (i) the 35 summarizers submitted to the DUC’04 conference, (ii) the 8 summaries generated by humans and provided by the DUC’04 system (beyond the golden summaries), (iii) two widely used open source text summarizers, i.e., the Open Text Summarizer (OTS) [38] and TexLexAn [43], (iv) a recently proposed itemset-based summarizer [6], named ItemSum (Itemset-based Summarizer), and (v) a baseline version of Yago-based Summarizer, namely Baseline, which adopts an established term relevance evaluator, i.e., the tf-idf score [27], rather than the ontology-based entity rank evaluator (see Definition 1). For the DUC’04 competitors we considered the results that were pro- vided by the DUC’04 system [15]. Specifically, for the top-ranked DUC’04 summarizer, i.e., CLASSY [13], we considered its most effective version (i.e., peer65). Similarly, for the other competitors we tuned the algorithm param- eters to their average best value by following the indications that were given by the respective authors. For Yago-based Summarizer we set, as standard configuration, γ to 0.3 and α to 0.9. In Section 4.4 we analyze more in detail the impact of both parameters on the Yago-based Summarizer performance. Table 1 summarizes the results that were achieved by Yago-based Summa- rizer, Baseline-tf-idf, ItemSum, OTS, TexLexAn, the 8 humanly generated 17 summaries, and the 10 most effective summarizers presented in the DUC’04 contest. To validate the statistical significance of the Yago-based Summarizer performance improvement against its competitors we performed the paired t-test [14] at 95% significance level for all of the evaluated measures. Ev- ery statistically relevant worsening in the comparison between Yago-based Summarizer and the other approaches is starred in Table 1. Table 1: DUC’04 Collections. Comparisons between Yago-based Summarizer and the other approaches. Statistically relevant differences in the comparisons between Yago-based Summarizer (standard configuration) and the other approaches are starred. Summarizer ROUGE-2 ROUGE-4 R Pr F R Pr F TOP RANKED DUC’04 PEERS peer120 0.076* 0.103 0.086* 0.014* 0.019 0.016 peer65 0.091* 0.090* 0.091* 0.015* 0.015 0.015* peer19 0.080* 0.080* 0.080* 0.010* 0.010* 0.010* peer121 0.071* 0.085* 0.077* 0.012* 0.014* 0.013* peer11 0.070* 0.087* 0.077* 0.012* 0.015* 0.012* peer44 0.075* 0.080* 0.078* 0.012* 0.013* 0.012* peer81 0.077* 0.080* 0.078* 0.012* 0.012* 0.012* peer104 0.086* 0.084* 0.085* 0.011* 0.010* 0.010* peer124 0.083* 0.081* 0.082* 0.012* 0.012* 0.012* peer35 0.083* 0.084 0.083* 0.010* 0.011* 0.011* DUC’04 HUMANS A 0.088* 0.092* 0.090* 0.009* 0.010* 0.010* B 0.091* 0.096 0.092 0.013* 0.013* 0.013* C 0.094 0.102 0.098 0.011* 0.012* 0.012* D 0.100 0.106 0.102 0.010* 0.010* 0.010* E 0.094 0.099 0.097 0.011* 0.012* 0.012* F 0.086* 0.090* 0.088* 0.008* 0.009* 0.009* G 0.082* 0.087* 0.084* 0.008* 0.008* 0.007* H 0.101 0.105 0.103 0.012* 0.013* 0.012* OTS 0.075* 0.074* 0.074* 0.009* 0.009* 0.009* texLexAn 0.067* 0.067* 0.067* 0.007* 0.007* 0.007* ItemSum 0.083* 0.085* 0.084* 0.012* 0.014* 0.014* Baseline 0.092* 0.091* 0.092* 0.014* 0.014* 0.014* Yago-based Summarizer 0.095 0.094 0.095 0.017 0.017 0.017 Yago-based Summarizer performs significantly better than ItemSum, OTS, TexLexAn, and Baseline for all of the analyzed measures. Hence, the ontology- based sentence ranking and selection strategies appear to be more effective than traditional information retrieval techniques (e.g., the tf-idf-based sen- tence evaluation [27]) for summarization purposes. Although, in some cases, 18 peer120 performs best in terms of ROUGE-2 and ROUGE-4 precision, Yago- based Summarizer performs significantly better than all the 35 DUC’04 com- petitors in terms of ROUGE-2 and ROUGE-4 F1-measure (i.e., the harmonic average between precision and recall). Hence, the summaries that were gen- erated by Yago-based Summarizer are, on average, the most accurate and not redundant ones. Compared to the 8 humanly generated summaries, Yago-based Summa- rizer significantly outperforms 3 out of 8 and 8 out of 8 competitors in terms of ROUGE-2 and ROUGE-4 F1-measure, respectively. In contrast, CLASSY (peer65) performs significantly better than 2 out of 8 and 6 out of 8 humans in terms of ROUGE-2 and ROUGE-4 F1-measure, respectively. Similarly, peer120 performs worser than all the humans in terms of ROUGE-2 and outperforms 7 out of 8 humans in terms of ROUGE-4. Hence, Yago-based Summarizer placed, on average, better than CLASSY and peer120 with re- spect to the humans. 4.3. Summary comparison We conducted a qualitative evaluation of the soundness and readability of the summaries that were generated by Yago-based Summarizer and the other approaches. Tables 2 reports the summaries that were produced by Yago- based Summarizer, the top-ranked DUC’04 summarizer, i.e., CLASSY [13] (Peer-65), and a commonly used open source summarizer OTS [38] on a representative DUC’04 collection, which relates the activities and the main achievements of the Yugoslav war crime tribunal. The summary that was generated by Yago-based Summarizer appears to be the most focused one, because it covers all the main document topics, 19 Table 2: Summary examples. Method Summary Yago-based Summarizer Yugoslavia must cooperate with the U.N. war crimes tribunal investigating alleged atrocities during the wars in Croatia and Bosnia, international legal experts meeting in Belgrade said Sunday. The Yugoslav war crimes tribunal Monday acquitted a Muslim military com- mander of war crimes against Bosnian Serb prisoners in 1992, but convicted three underlings in the first U.N. case dealing with anti-Serb atrocities. American and allied forces in Bosnia on Wednesday arrested a Bosnian Serb general who was charged with genocide by the international war crimes tri- bunal in a recent secret indictment. CLASSY Some of the closest combat in the half year of the Kosovo conflict, to the point of fighting room to room and floor to floor, occurred near this village six weeks ago, in the days before 21 women, children and elderly members of the Delijaj clan were massacred by Serbian forces, their mutilated bodies left strewn on the forest floor. In its first case to deal with atrocities against Serbs during Bosnia’s civil war, a U.N. war crimes tribunal on Monday convicted three prison officials and guards, but acquitted a top military commander who oversaw the facility. Hundreds of people gathered at Sarajevo airport on Saturday to welcome Zejnil Delalic, who was cleared of war crimes charges earlier this week after spending 980 days in jail of the international war crimes tribunal in The Hague. OTS The Yugoslav war crimes tribunal Monday acquitted a Muslim military com- mander of war crimes against Bosnian Serb prisoners in 1992, but convicted three underlings in the first U.N. case dealing with anti-Serb atrocities. The Yugoslav war crimes tribunal cleared Zejnil Delalic, a Muslim, of re- sponsibility for war crimes committed against Serb captives at a Bosnian government-run prison camp under his command. court convicted camp com- mander Zdravko Mucic, a Croat, of 11 war crimes and grave breaches of the Geneva Conventions because he oversaw guards who murdered nine Serbs and tortured six. Indeed, the conflict between Serbian forces bent on keeping Kosovo in Serbia and guerrillas fighting for the independence of its heavily ethnic Albanian population first drew international attention with the massacre of the Jasari clan in early March by Serbian units at Prekaz, in central Kosovo. i.e., (1) the role of the Yugoslav war crime tribunal, (2) the acquittal of the Muslim military commander, and (3) the arrest of the Bosnian Serb general. In contrast, OTS and CLASSY cover, to some extent, only the topic (2). On the other hand, both OTS and CLASSY select other contextual sentences about the Kosovo war, which are very general and not representative of the key document message. Hence, the corresponding summaries are deemed to be partially redundant. 20 4.4. Parameter setting The setting of the user-specified α and γ parameters could affect the Yago- based Summarizer performance significantly. Hence, we thoroughly analyzed their impact on the Yago-based Summarizer ROUGE scores. Figures 3(a) and 3(b) plot the ROUGE-2 and ROUGE-4 F1-measure that were achieved by Yago-based Summarizer on the DUC’04 collection by varying the value of γ in the range [0,1] and by setting α to its best value (0.9), respectively. In contrast, Figures 4(a) and 4(b) plot the ROUGE-2 and ROUGE-4 F1-measure scores by setting the best γ value (0.3) and by varying α in the range [0,1]. 0.089 0.09 0.091 0.092 0.093 0.094 0.095 0.096 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 F 1 -m e a s u re γ (a) ROUGE-2 F1-measure. 0.015 0.0155 0.016 0.0165 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 F 1 -m e a s u re γ (b) ROUGE-4 F1-measure. Figure 3: Impact of γ on the Yago-based Summarizer performance. α=0.9. DUC’04 collections. When increasing the value of the γ parameter, the sentences that include many unrecognized words are on average penalized (see Formula 1). Based on the results that are reported in Figure 3, the best performance results were achieved by setting γ=0.3. Furthermore, the results remain relatively stable when γ ranges between 0.2 and 0.5. Since the EntityRank score of 21 0.05 0.06 0.07 0.08 0.09 0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 F 1 -m e a s u re α (a) ROUGE-2 F1-measure. 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 F 1 -m e a s u re α (b) ROUGE-4 F1-measure. Figure 4: Impact of α on the Yago-based Summarizer performance. γ=0.3. DUC’04 collections. many of the recognized entities fall in the same value range, it means that redoubling the score of the recognized named entities with respect to the time/date/number entities yields good summarization performance. In con- trast, setting γ out of the above value range implies giving an under- or over-emphasis to the least interesting entities. The α parameter allows the user to decide to which extent the similarity between the already selected sentence is relevant compared to the entity- based sentence rank for sentence selection. The higher the value of α is, the more important the ontology-based sentence rank becomes with respect to the similarity with the previously selected sentences (see Formula 4). Since the analyzed documents contain a limited amount of redundancy, Yago-based Summarizer achieves averagely high ROUGE scores by setting high α values (i.e., α > 0.7). With the DUC’04 collections, the best performance results were achieved by setting α=0.9. Note that, with such configuration setting, Yago-based Summarizer disregards, to a large extent, the impact of the sim- 22 ilarity score with respect to the ontology-based sentence ranking. However, when coping with document collections that contain a larger number of repe- titions, the user should set lower α values in order to achieve a good trade-off between summary relevance and redundancy. 5. Conclusions and future works In recent years, semantics-based document analysis has shown to improve the performance of document summarization systems significantly. However, since most of the related approaches perform semantics-based analysis as a preprocessing step rather than integrating the ontological knowledge into the summarization process, the quality of the generated summaries remains, in some cases, unsatisfactory. This paper proposes to improve the performance of state-of-the-art sum- marizers by integrating an ontology-based sentence evaluation and selection step into the summarization process. Specifically, an established entity recog- nition and disambiguation step based on the Yago ontology is used to identify the key document concepts and evaluate their significance with respect to the document context. The same results are then exploited to select the most representative document sentences. The experimental results show that Yago-based Summarizer performs better than many state-of-the-art summarizers on benchmark collections. Furthermore, a qualitative comparison between the summaries generated by Yago-based Summarizer and the state-of-the-art summarizers demonstrate the usefulness and applicability of the proposed approach. Future developments on this work will address the application of the 23 proposed summarization approach to multilingual document collections and the use of entropy-based sentence selection strategies to further improve the compactness of the generated summaries. References [1] Alguliev, R. M., Aliguliyev, R. M., & Hajirahimova, M. S. (2012). Gendocsummclr: Generic document summariza- tion based on maximum coverage and less redundancy. Ex- pert Systems with Applications, 39 , 12460 – 12473. URL: http://www.sciencedirect.com/science/article/pii/S0957417412006641. doi:10.1016/j.eswa.2012.04.067. [2] Alguliev, R. M., Aliguliyev, R. M., & Isazade, N. R. (2013). Cdds: Constraint-driven document summarization models. Expert Systems with Applications, 40, 458 – 465. URL: http://www.sciencedirect.com/science/article/pii/S0957417412009049. doi:10.1016/j.eswa.2012.07.049. [3] Alguliev, R. M., Aliguliyev, R. M., & Isazade, N. R. (2013). Mul- tiple documents summarization based on evolutionary optimization algorithm. Expert Systems with Applications, 40, 1675 – 1689. doi:10.1016/j.eswa.2012.09.014. [4] Atkinson, J., & Munoz, R. (2013). Rhetorics-based multi-document summarization. Expert Systems with Applications, 40, 4346 4352. URL: http://www.sciencedirect.com/science/article/pii/S0957417413000304. doi:10.1016/j.eswa.2013.01.017. 24 [5] Baader, F., Calvanese, D., McGuinness, D. L., Nardi, D., & Patel- Schneider, P. F. (Eds.) (2003). The Description Logic Handbook: The- ory, Implementation, and Applications. Cambridge University Press. [6] Baralis, E., Cagliero, L., Fiori, A., & Jabeen, S. (2012). Multi-document summarization exploiting frequent itemsets. In In Proceedings of the ACM Symposium on Applied Computing (SAC 2012). [7] Baralis, E., & Fiori, A. (2010). Summarizing biological literature with biosumm. In CIKM (pp. 1961–1962). [8] Baxter, D., Klimt, B., Grobelnik, M., Schneider, D., Witbrock, M., & Mladenic, D. (2009). Capturing document semantics for ontology gener- ation and document summarization. Semantic Knowledge Management: Integrating Ontology Management, Knowledge Discovery, and Human Language Technologies, (pp. 141–154). [9] Brin, S., & Page, L. (1998). The anatomy of a large-scale hypertextual web search engine. In Proceedings of the seventh international conference on World Wide Web 7 (pp. 107–117). [10] Carbonell, J., & Goldstein, J. (1998). The use of mmr, diversity-based reranking for reordering documents and produc- ing summaries. In Proceedings of the 21st annual international ACM SIGIR conference on Research and development in infor- mation retrieval SIGIR ’98 (pp. 335–336). New York, NY, USA: ACM. URL: http://doi.acm.org/10.1145/290941.291025. doi:10.1145/290941.291025. 25 [11] Carenini, G., Ng, R. T., & Zhou, X. (2007). Summarizing email con- versations with clue words. In World Wide Web Conference Series (pp. 91–100). [12] Conroy, J., Schlesinger, J., Kubina, J., Rankel, P., & OLeary, D. (2011). Classy 2011 at tac: Guided and multi-lingual summaries and evalua- tion metrics. In TAC’11: Proceedings of the The 2011 Text Analysis Conference. [13] Conroy, J. M., Schlesinger, J. D., Goldstein, J., & O’Leary, D. P. (2004). Left-Brain/Right-Brain Multi-Document Summarization. In DUC 2004 Conference Proceedings. [14] Dietterich, T. G. (1998). Approximate statistical test for comparing supervised classification learning algorithms. Neural Computation, 10. [15] Document Understanding Conference (2004). HTL/NAACL workshop on text summarization. [16] Dredze, M., Wallach, H. M., Puller, D., & Pereira, F. (2008). Generating summary keywords for emails using topics. In Proceedings of the 13th international conference on Intelligent user interfaces IUI ’08 (pp. 199–206). New York, NY, USA: ACM. URL: http://dx.doi.org/10.1145/1378773.1378800. doi:10.1145/1378773.1378800. [17] Filatova, E. (2004). A formal model for information selection in multi- sentence text extraction. In In Proceedings of the International Confer- ence on Computational Linguistics (COLING (pp. 397–403). 26 [18] Fortes, G. L. F., de Lima Jos Valdeni, Loh, S., & de Oliveira, J. P. M. (2006). Using ontological modeling in a context-aware summarization system to adapt text for mobile devices. In P. P. Chen, & L. Y. Wong (Eds.), Active Conceptual Modeling of Learning (pp. 144–154). Springer volume 4512 of Lecture Notes in Computer Science. [19] Hachey, B., Radford, W., Nothman, J., Honnibal, M., & Curran, J. R. (2013). Evaluating entity linking with wikipedia. Artif. Intell., 194 , 130–150. URL: http://dx.doi.org/10.1016/j.artint.2012.04.005. doi:10.1016/j.artint.2012.04.005. [20] Hamasaki, M., Matsuo, Y., Nishimura, T., & Takeda, H. (2009). Ontol- ogy extraction by collaborative tagging. In WWW 2009. ACM press. [21] Hennig, L., Umbrath, W., & Wetzker, R. (2008). An ontology-based approach to text summarization. In Web Intelligence/IAT Workshops (pp. 291–294). IEEE. [22] Hoffart, J., Yosef, M. A., Bordino, I., Frstenau, H., Pinkal, M., Spaniol, M., Taneva, B., Thater, S., & Weikum, G. (2011). Robust disambigua- tion of named entities in text. In EMNLP (pp. 782–792). ACL. [23] Kleinberg, J. M. (1999). Authoritative sources in a hyperlinked environ- ment. J. ACM , 46 , 604–632. [24] Kogilavani, A., & Balasubramanie, B. (2009). Ontology enhanced clus- tering based summarization of medical documents. International Jour- nal of Recent Trends in Engineering, 1 . 27 [25] Lau, R. Y. K., Song, D., Li, Y., Cheung, T. C. H., & Hao, J.-X. (2009). Toward a fuzzy domain ontology extraction method for adap- tive e-learning. IEEE Trans. on Knowl. and Data Eng., 21, 800–813. doi:http://dx.doi.org/10.1109/TKDE.2008.137. [26] Li, L., Wang, D., Shen, C., & Li, T. (2010). Ontology-enriched multi- document summarization in disaster management. In F. Crestani, S. Marchand-Maillet, H.-H. Chen, E. N. Efthimiadis, & J. Savoy (Eds.), SIGIR (pp. 819–820). ACM. [27] Lin, C.-Y., & Hovy, E. (2003). Automatic evaluation of summaries using n-gram co-occurrence statistics. In Proceedings of the North American Chapter of the Association for Computational Linguistics on Human Language Technology - Volume 1 (pp. 71–78). [28] Litvak, M., & Last, M. (2008). Graph-based keyword ex- traction for single-document summarization. In Proceedings of the Workshop on Multi-source Multilingual Information Extrac- tion and Summarization MMIES ’08 (pp. 17–24). Strouds- burg, PA, USA: Association for Computational Linguistics. URL: http://dl.acm.org/citation.cfm?id=1613172.1613178. [29] Mittal, J. G. V., Goldstein, J., Mittal, V., Carbonell, J., & Kantrowitz, M. (2000). Multi-document summarization by sentence extraction. In In Proceedings of the ANLP/NAACL Workshop on Automatic Summa- rization (pp. 40–48). [30] Mohamed, A., & Rajasekaran, S. (2006). Improving query-based sum- 28 marization using document graphs. In Signal Processing and Informa- tion Technology, 2006 IEEE International Symposium on (pp. 408–410). doi:10.1109/ISSPIT.2006.270835. [31] Nastase, V. (2008). Topic-driven multi-document summarization with encyclopedic knowledge and spreading activation. In EMNLP (pp. 763– 772). ACL. [32] Pang-Ning, T., Michael, S., & Vipin, K. (2005). Intoduction to data mining. [33] Park, S., & Cha, B. (2008). Query-based multi-document summa- rization using non-negative semantic feature and nmf clustering. In Networked Computing and Advanced Information Management, 2008. NCM ’08. Fourth International Conference on (pp. 609–614). volume 2. doi:10.1109/NCM.2008.246. [34] Ping, C., & M., V. R. (2006). A query-based medical information sum- marization system using ontology knowledge. In CBMS (pp. 37–42). IEEE Computer Society. [35] Pourvali, M., & Abadeh, M. S. (2012). Automated text summariza- tion base on lexicales chain and graph using of wordnet and wikipedia knowledge base. CoRR, abs/1203.3586 . [36] Radev, D. R. (2004). Lexrank: Graph-based lexical centrality as salience in text summarization. Journal of Artificial Intelligence Research, 22 , 2004. 29 [37] Radev, D. R., Jing, H., Stys, M., & Tam, D. (2004). Centroid-based summarization of multiple documents. Information Processing and Man- agement, 40, 919 – 938. [38] Rotem, N. (2011). Open text summarizer (ots). retrieved from http://libots.sourceforge.net/ in july 2011. [39] Saravanan, M., & Ravindran, B. (2010). Identifica- tion of rhetorical roles for segmentation and summariza- tion of a legal judgment. Artif. Intell. Law, 18, 45– 76. URL: http://dx.doi.org/10.1007/s10506-010-9087-7. doi:10.1007/s10506-010-9087-7. [40] Steinberger, J., Kabadjov, M., Steinberger, R., Tanev, H., Turchi, M., & Zavarella, V. (2011). Jrc’s participation at tac 2011: Guided and multilingual summarization tasks. In TAC’11: Proceedings of the The 2011 Text Analysis Conference. [41] Suchanek, F. M., Kasneci, G., & Weikum, G. (2007). Yago: a core of semantic knowledge. In Proceedings of the 16th international confer- ence on World Wide Web WWW ’07 (pp. 697–706). New York, NY, USA: ACM. URL: http://doi.acm.org/10.1145/1242572.1242667. doi:10.1145/1242572.1242667. [42] Takamura, H., & Okumura, M. (2009). Text summarization model based on the budgeted median problem. In Proceeding of the 18th ACM con- ference on Information and knowledge management (pp. 1589–1592). 30 [43] TexLexAn (2011). Texlexan: An open-source text summarizer. re- trieved from http://texlexan.sourceforge.net/ in july 2011. URL: http://texlexan.sourceforge.net/. [44] Thakkar, K., Dharaskar, R., & Chandak, M. (2010). Graph-based algo- rithms for text summarization. In Emerging Trends in Engineering and Technology (ICETET), 2010 3rd International Conference on (pp. 516 –519). doi:10.1109/ICETET.2010.104. [45] Town, C. (2006). Ontological inference for image and video analysis. Machine Vision and Applications, 17, 94–115. URL: http://dx.doi.org/10.1007/s00138-006-0017-3. 10.1007/s00138- 006-0017-3. [46] Wan, X., & Yang, J. (2006). Improved affinity graph based multi- document summarization. In In Proceedings of HLT-NAACL, Com- panion Volume: Short Papers (pp. 181–184). [47] Wang, D., & Li, T. (2010). Document update summarization using incremental hierarchical clustering. In Proceedings of the 19th ACM in- ternational conference on Information and knowledge management (pp. 279–288). [48] Wang, D., Zhu, S., Li, T., Chi, Y., & Gong, Y. (2011). Integrating document clustering and multidocument sum- marization. ACM Trans. Knowl. Discov. Data, 5 , 14:1– 14:26. URL: http://doi.acm.org/10.1145/1993077.1993078. doi:http://doi.acm.org/10.1145/1993077.1993078. 31 [49] Wikipedia (2013). Wikipedia website. last access: 01/03/2013. URL: http://www.wikipedia.org. [50] Wu, C.-W., & Liu, C.-L. (2003). Ontology-based text summarization for business news articles. In N. C. Debnath (Ed.), Computers and Their Applications (pp. 389–392). ISCA. [51] Yang, Z., Cai, K., Tang, J., Zhang, L., Su, Z., & Li, J. (2011). Social context summarization. In Proceedings of the 34th inter- national ACM SIGIR conference on Research and development in Information Retrieval SIGIR ’11 (pp. 255–264). New York, NY, USA: ACM. URL: http://doi.acm.org/10.1145/2009916.2009954. doi:http://doi.acm.org/10.1145/2009916.2009954. [52] Zhu, J., Wang, C., He, X., Bu, J., Chen, C., Shang, S., Qu, M., & Lu, G. (2009). Tag-oriented document summa- rization. In Proceedings of the 18th international conference on World wide web WWW ’09 (pp. 1195–1196). New York, NY, USA: ACM. URL: http://doi.acm.org/10.1145/1526709.1526925. doi:http://doi.acm.org/10.1145/1526709.1526925. 32 
work_e6tit3rbvvdrhf7l63fvslnwxm	----	[PDF] Online modulation recognition of analog communication signals using neural network | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1016/j.eswa.2006.04.015 Corpus ID: 17309878Online modulation recognition of analog communication signals using neural network @article{Gldemir2007OnlineMR, title={Online modulation recognition of analog communication signals using neural network}, author={Hanifi G{\"u}ldemir and Abdulkadir Seng{\"u}r}, journal={Expert Syst. Appl.}, year={2007}, volume={33}, pages={206-214} } Hanifi Güldemir, Abdulkadir Sengür Published 2007 Computer Science Expert Syst. Appl. In this paper, a neural network based online analog modulation recognition of communication signals is presented. The proposed system can discriminate between amplitude modulation (AM), frequency modulation (FM), double sideband (DSB), upper sideband (USB), lower sideband (LSB) and continuous wave (CW) modulations. A matlab graphical user interface (GUI) is designed to see the intercepted signal, its power spectral density, frequency and modulation type on the screen of the personnel computer… Expand View via Publisher researchgate.net Save to Library Create Alert Cite Launch Research Feed Share This Paper 6 CitationsBackground Citations 1 Methods Citations 2 View All Figures, Tables, and Topics from this paper figure 1 table 1 figure 2 table 2 figure 3 table 3 figure 4 table 4 figure 5 table 5 figure 6 figure 7 figure 8 View All 13 Figures & Tables Artificial neural network Graphical user interface Spectral density Biological Neural Networks Instantaneous phase Analog signal Signal-to-noise ratio MATLAB Computer simulation Simulation Delta-sigma modulation Carrier frequency Neural Network Simulation USB Extraction DNA Breaks, Double-Stranded Hearing Loss, High-Frequency Least significant bit Control table Verification of Theories 6 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency Modulation recognition of RF carriers using neural networks Carmen-Silvia Oprina, Otilia Cangea, M. Dima Computer Science 2016 8th International Conference on Electronics, Computers and Artificial Intelligence (ECAI) 2016 PDF Save Alert Research Feed Trends in Automatic Modulation Classification for Advanced Data Communication Networks P. Thakur, S. Madan, Mamta Madan 2015 2 Save Alert Research Feed A New Approach of Adaptive Network-Based Fuzzy Inference System Modeling in Laser Processing-A Graphical User Interface (GUI) Based Sivarao, P. Brevern, N. M. Eltayeb Computer Science 2009 14 PDF View 1 excerpt, cites methods Save Alert Research Feed Development of Man-Machine Interface using Matlab: an adaptive network-based fuzzy inference system modeling for laser machining Sivarao, Rizal, A. Tajul, Taufik Engineering 2009 PDF View 1 excerpt, cites methods Save Alert Research Feed ANFIS Modeling of Laser Machining Responses by Specially Developed Graphical User Interface Subramonian Sivarao Engineering 2009 6 PDF View 1 excerpt, cites background Save Alert Research Feed Sayısal modülasyonlu haberleşme işaretlerinden Wigner-Ville zaman-frekans dağılımlarına dayalı öznitelik çıkarımı Abdulkadir Şengür Art 2011 Save Alert Research Feed References SHOWING 1-10 OF 14 REFERENCES SORT BYRelevance Most Influenced Papers Recency An automatic modulation recognition algorithm for spectrum monitoring applications C. Dubuc, D. Boudreau, F. Patenaude, R. Inkol Computer Science 1999 IEEE International Conference on Communications (Cat. No. 99CH36311) 1999 34 Save Alert Research Feed A testbed for automatic modulation recognition using artificial neural networks S. C. Kremer, J. Shiels Computer Science CCECE '97. Canadian Conference on Electrical and Computer Engineering. Engineering Innovation: Voyage of Discovery. Conference Proceedings 1997 29 View 1 excerpt, references methods Save Alert Research Feed Procedure for automatic recognition of analogue and digital modulations E. Azzouz, A. Nandi Computer Science 1996 96 View 2 excerpts Save Alert Research Feed Identification algorithm of upper sideband and lower sideband SSB signals Y. O. Al-Jalili Computer Science Signal Process. 1995 28 View 1 excerpt, references background Save Alert Research Feed Automatic modulation classification using zeroç crossing Shue-Zen Hsue, S. Soliman Mathematics, Computer Science 1990 150 View 1 excerpt, references background Save Alert Research Feed Identification of the modulation type of a signal Y. Chan, L. G. Gadbois, P. Yansouni Physics, Computer Science ICASSP '85. IEEE International Conference on Acoustics, Speech, and Signal Processing 1985 62 View 1 excerpt Save Alert Research Feed Automatic classification of high frequency signals Friedrich K. Jondral Computer Science 1985 85 View 1 excerpt Save Alert Research Feed Modulation recognition using artificial neural networks A. Nandi, E. Azzouz Computer Science Signal Process. 1997 146 View 1 excerpt, references background Save Alert Research Feed A general approach to the automatic classification of radiocommunication signals L. Vergara-Dominguez, J. Páez-Borrallo, J. P. García, B. R. Mezcua Computer Science Signal Process. 1991 78 Save Alert Research Feed Automatic modulation recognition using time domain parameters J. Aisbett Engineering 1987 82 View 2 excerpts, references methods Save Alert Research Feed ... 1 2 ... Related Papers Abstract Figures, Tables, and Topics 6 Citations 14 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_e77yzhiezbgvzom6xvhg4tcdui	----	SWSNL: Semantic Web Search Using Natural Language Ivan Habernala,∗, Miloslav Konoṕıkb aDepartment of Computer Science and Engineering, Faculty of Applied Sciences, University of West Bohemia, Univerzitńı 8, 306 14 Plzeň, Czech Republic b NTIS – New Technologies for the Information Society, Faculty of Applied Sciences, University of West Bohemia in Pilsen, Univerzitńı 8, 306 14 Plzeň, Czech Republic Abstract As modern search engines are approaching the ability to deal with queries expressed in natural language, full support of natural language interfaces seems to be the next step in the development of future systems. The vision is that of users being able to tell a computer what they would like to find, using any number of sentences and as many details as requested. In this article we describe our effort to move towards this future using currently available technology. The Semantic Web framework was chosen as the best means to achieve this goal. We present our approach to building a complete Semantic Web Search using Natural Language (SWSNL) system. We cover the complete process which includes preprocessing, semantic analysis, semantic interpretation, and executing a SPARQL query to retrieve the results. We perform an end-to-end evaluation on a domain dealing with accommodation options. The domain data come from an existing accommodation portal and we use a corpus of queries obtained by a Facebook campaign. In our paper we work with written texts in the Czech language. In addition to that, the Natural Language Understanding (NLU) module is evaluated on another domain (public transportation) and language (English). We expect that our findings will be valuable for the research community as they are strongly related to issues found in real-world scenarios. We struggled with inconsistencies in the actual Web data, with the performance of the Semantic Web engines on a decently sized knowledge base, and others. Keywords: semantic search, natural language understanding, semantic web, statistical semantic analysis 1. Introduction Keyword-based search engines have proven to be very efficient on collections of unstructured textual content, e.g. web pages [1]. However, if users want to find information in a structured content, e.g. in a database, the basic key- word search might not be expressive enough [2]. A simple, yet sufficient solution on the Web can be provided by a form-based user interface which typically combines key- words with some other restrictions, according to the spe- cific domain structure [3]. Form-based interfaces are user friendly in the sense that they do not require the user’s prior knowledge of the underlying data structures. The structure is typically shown as multiple forms or menus that allow further specification of the user request. Never- theless, it is obvious that from the user’s point of view, a simple keyword input is a more straightforward approach than filling forms. A different approach to retrieving sought information consists in using a domain-specific query language. How- ever, as pointed out by [4], querying structured data using ∗Corresponding author. Tel: +420 377632456 Email addresses: habernal@kiv.zcu.cz (Ivan Habernal), konopik@kiv.zcu.cz (Miloslav Konoṕık) a domain-specific query language, e.g. SQL or SPARQL [5], can be complicated for casual users. Most of the exist- ing query languages use a precise syntax and the users are expected to be familiar with the back-end data structures. Evidently, this allows exact queries to be formulated by domain experts but on the other hand this is a big obsta- cle for casual users. One step beyond the above-mentioned traditional ap- proaches are Natural Language Interfaces (NLI). The key feature of such interface is that users can search for the required information by posing their queries using natural language (NL). It is assumed that posing NL queries is more precise than the keyword approach and at the same time more natural than the form-based approach. Recent studies [6, 7] indeed confirmed that NLI offers a better user experience than the form-based search for casual users. In the article we present a realistic approach to design- ing and developing a complete Semantic Web Search using Natural Language (SWSNL) system. The system builds on the Semantic Web paradigm and thus uses ontologies as a main vehicle for storing the domain structure and all the data (the Knowledge Base, KB). Furthermore, ontolo- gies, due to their ability to precisely capture the semantics, are also used to describe the meaning of user queries that Preprint submitted to Expert Systems With Applications November 14, 2012 are expressed in natural language (NL). The system can be seen as a search engine which accepts NL queries and performs the search in the back-end KB. Our SWSNL system also has a few aspects that distin- guish it from other related work. Firstly, we allow users to formulate their NL queries using more than a single sentence. Secondly, the Natural Language Understanding (NLU) module incorporates a stochastic semantic analysis model and does not require any syntactic parser prepro- cessing. The NLU module is trained from annotated data and it is language independent. Thirdly, the ontology for describing natural language queries is different from (inde- pendent of) the KB ontology. This independence enables switching between various KBs without affecting the NLU module. The system was thoroughly tested and evaluated on three different domains. Primarily, the end-to-end evalua- tion was performed on the accommodation options domain using actual data acquired from the Web as well as the cor- pus of actual NL queries in the Czech language. The other two domains, the Czech public transportation domain and a subset of the English ATIS dataset, were used to verify the portability to different domains and languages. The initial impulse to build the presented system was an idea to enrich an existing form-based application with a newly developed NLI. The expected benefits were tar- geted into better user experience and into testing the NLI approach in practice. The selected domain of accommo- dation options followed the idea as our preliminary survey which revealed a promising potential of such a domain. The selection of the Czech language as the domain lan- guage was not a random choice. The decision was meant to verify the technology in a different language than En- glish. The rest of the article is organized as follows. Section 2 outlines the context of the related work. In section 3 we introduce our system in a schematic overview. Sec- tion 4 describes the target domain together with the tech- niques required to deal with actual Web data sources. A natural language corpus is also presented. Section 5 de- scribes the formalism for capturing natural language query semantics. Section 6 proposes a statistical semantic model for analysing natural language queries and section 7 deals with the semantic interpretation of a semantic annotation in order to perform a search in a particular KB. Section 8 thoroughly describes the evaluation of our SWSNL sys- tem. Open issues and future work are then discussed in section 9. 2. Related Work Before we present the related work, we shortly discuss the terminology of the current research into Natural Lan- guage Interfaces. Typically, the term Natural Language In- terfaces (NLI) is used when a system can be accessed using (written) natural language. Such a system mostly operates on structured information and, given the natural language question, it tries to find the correct answer. The family of NLI systems can be divided into various sub-classes. A Natural Language Interfaces to Databases (NLIDB) sys- tem holds information in a relational database. The prin- ciples of NLIDB have been adapted to the Semantic Web resulting into the Natural Language Interfaces to Knowl- edge Bases (NLIKB). In this case of NLI, the information is stored in the form of ontology, which plays the funda- mental role in the Semantic Web. We define our research as Semantic Web Search Us- ing Natural Language (SWSNL). Although NLIKB cov- ers a task similar to our approach, most of the existing NLIKB systems are designed to evaluate rather complex logical queries over the Knowledge Base. On the contrary, SWSNL can be viewed as a special case of a search engine that retrieves a set of results according to a natural lan- guage query, operating in the Semantic Web field. Further- more, our SWSNL system does not belong to the family of question answering systems since these two fields have a very different motivation. Whereas a question answering system tries to answer a user’s NL questions, the purpose of our system is to retrieve search results according to a user’s NL queries. 2.1. Overview of recent NLIKB systems A very recent system called PowerAqua [8] is an ontolo- gy-based NLI system which surpasses traditional systems by managing multiple ontology sources and high scalabil- ity. Since its NL processing module remains the same as in the previous AquaLog system [9], we will review the AquaLog system. AquaLog is a portable NLIKB system which handles user queries in a natural language (English) and returns answers inferred from a knowledge base. The system uses GATE1 libraries (namely the tokenizer, the sentence splitter, the POS tagger, and the VP chunker). ORAKEL [10] is an ontology-based NLI system. It ac- cepts English factoid questions and translates them into first-order logic forms. This conversion uses full syntax parsing and a compositional semantics approach. ORAKEL can be ported into another domain but such porting re- quires a domain expert to create a domain-dependent lex- icon. The lexicon is used for an exact mapping from nat- ural language constructs to ontology entities. A possible drawback of ORAKEL’s approach is that the system can neither handle ungrammatical questions nor deal with un- known words. The FREyA system [11] is an NLIKB system that com- bines syntactic parsing with ontology reasoning. It derives parse trees of English input questions and uses heuris- tic rules to find a set of potential ontology concepts (for mapping from question terms to ontology concepts) using GATE and the OntoRoot Gazetteer [12]. The primary source for question understanding is the ontology itself. If the system encounters ambiguities, a clarification dialogue 1http://gate.ac.uk 2 is offered to the user. The potential ontology concepts re- trieved from the question analysis are then transformed into SPARQL. The system was tested on Mooney: Geog- raphy dataset [13]. A portable NLIKB system called PANTO [14] is based upon the off-the-shelf statistical parser Standford Parser and integrates tools like WordNet and various metrics al- gorithms to map the NL question terms to an interme- diate representation called QueryTriples.2 This seman- tic description is then mapped onto the OntoTriples that are connected to entities from the underlying ontology. This step involves a set of 11 heuristic mapping rules. Finally, OntoTriples are represented as SPARQL queries. The main idea of transforming a NL question into triples in PANTO is based upon the empirical observation that two nominal phrases from a parse tree are expected to be mapped onto a triple in the ontology. The system was eval- uated on the Mooney dataset and the output of PANTO was compared to the manually generated SPARQL queries. The NLP-Reduce system [15] does not involve any ad- vanced linguistic and semantic tools and depends on match- ing the query words to the KB instances. Its core part is a query generator which is responsible for creating SPARQL query given the words and the lexicon extracted from the KB. The major strength of the system is its good porta- bility as it does not depend on any complex NLP query processing. The ontology-based NLI system called QACID intro- duced in [16], covers a cinema/movie domain. Its tar- get language is Spanish. It consists of two main compo- nents, the user query formulation database, and textual- entailment engine. Whereas the former component serves mainly for development and system training purposes, the latter is intended for an unknown query processing. The core of the QACID system is a database of query formu- lations. The database contains a set of 54 clusters, each cluster represents one type of question and it has a rep- resentative query pattern which was derived from a set of training data. Each cluster is also associated with one SPARQL query. As pointed out in the QACID evaluation, the system is not able to answer unknown ontology con- cepts and therefore it fails if the user poses a query using terms that are not present in the lexicon. A domain-restricted NLI system called QUETAL [17] exploits a robust semantic analysis on a hybrid NLP ar- chitecture. The system was developed to answer NL ques- tions on two domains: the Nobel prizes and information about the Language Technology Institute. For these two domains, the ontologies were converted from existing on- tologies and merged with other more general ontologies (e.g. SUMO for the Nobel prizes). Question analysis is performed by the HPSG (Head-driven Phrase Structure Grammar [18]) syntactic and semantic parser with the sup- 2A mixed terminology (triplets vs. triples) appears in the litera- ture. In this paper we use triples as proposed in the RDF standard, http://www.w3.org/TR/rdf-concepts/#section-triples. port of the Heart of Gold architecture3 which provides shallow Named Entity Recognition (NER). The HPSG parser uses grammars for English and German and the output of the parser is a formalism called Minimal Re- cursion Semantics which is a flat, non-recursive semantic representation format [19]. This question representation is then transformed into the proto queries that are suitable for querying a particular KB. A domain-independ system QuestIO [20] extracts the domain lexicalizatin from the KB. The natural language query is transformed into SeRQL and then executed against the KB. The system was tested on 36 questions collected from the GATE user mailing list and several queries over a geographic ontology. Instead of supporting full NL questions, the QUICK (QUery Intent Constructor for Keywords) system guides users in constructing a semantic query from keywords [21]. After typing in the keywords, the system analyses them ac- cording to the underlying ontology and allows the user to choose the required semantic query in a clarification dia- logue. The performance was tested using two ontologies (a movie database and a song database) and 175 queries consisting of 2-5 keywords. Although QUICK is not a fully-fledged NLI system, it can be viewed as a trade-off between a simple keyword search and the NL question sys- tems. Another keyword-based information extraction and retrieval system based on ontology in the soccer domain is presented in [22]. The system is evaluated only on 10 simple queries. A good overview of other state-of-the art keyword-based semantic search systems is shown in [23]. 2.2. Existing Ontologies In Accommodation Domain The Travel ontology [24] is a core part of Travel Guides4. The goal of this system is to solve the problem of dis- tributed tourist sources and to help users find a vaca- tion package quickly. The ontology thoroughly models the tourist domain but it is not populated with instances of actual accommodation options. Another travel ontology developed by Knublauch in 2004 [25] was intended as an example how to put Semantic Web techniques into practice. The ontology covers the travel domain, however, it was not designed according to any existing travel information source. Accessing touristic ontology using NLI is the goal of [26]. The system is evaluated on a set of 20 simple NL queries. The author of [27] deals with a related task, namely with a personalised hotel search using Semantic Web technologies. However, his system is form-based and does not involve NLI. It is also worth mentioning that ontologies are nowa- days used by some companies as a backbone in tourism 3Heart of Gold is an open source middleware for combining shal- low and deep NLP components http://heartofgold.dfki.de/ 4http://sites.google.com/site/ontotravelguides/ 3 portals and services, such as by Seekda5 [28, 29] or Mon- deca6. 2.3. Evaluation in Related Work Whereas in established NLP branches standard test datasets together with evaluation criteria are available, there has been a lack of such dataset in the NLIKB field. Some typical problems related to the evaluation are: • There is no widely accepted agreement on what is a correct result. This is a very subjective expression as it stands for e.g. a recall of the correct query [10], a number of answered questions [9], or even a number of queries generated as an output [14]. • Each system operates on a different ontology and the ontologies vary in their size and complexity. As pointed out by [30], more attention should be paid to creating a standard evaluation set. Recently, various evaluation campigns have been initiated as a part of a se- mantic tool benchmarking project called Seals7 (Semantic Evaluation At Large Scale). In the field of NLIKB, [6] and [7] evaluate different approaches to semantic search for expert and casual users. The Mooney Geoquery was chosen as testing ontology. The dataset contains informa- tion about U. S. geography, such as cities, rivers, moun- tains, countries, etc. Originally, the dataset was in Prolog (consisting of about 1000 Prolog facts), later it was con- verted into OWL by [31]. The question corpus consists of 880 user questions with relation to U. S. geography and was collected from undergraduate students of the authors of [13]. A deep quantitative analysis of this dataset was performed by [31] and the authors of [10] point out some interesting real-world issues of this dataset. 3. SWSNL System Architecture Overview Our SWSNL system is a Java-based application con- sisting of various components. Figure 1 shows the archi- tecture of the system together with the data-flow of an example query. On the input, the user makes a query in natural lan- guage (NL). There are no restrictions on the query in terms of its length, as the user can express the query using key- words, a single sentence or the whole paragraph. We eval- uated our system on two languages, Czech and English. The details of the NL queries are described in section 4.1, some examples of actual queries can be found in Appendix A. The query is analysed by the Natural Language Un- derstanding (NLU) component. This component produces a KB-independent semantic representation of NL query. 5http://www.seekda.com/ 6http://www.mondeca.com/Clients 7http://www.seals-project.eu/ The NLU component is comprised of three basic blocks, the NL query preprocessing, Named Entity Recognition (NER) and Semantic Analysis. The details of the NLU module are presented in section 6. The module uses a sta- tistical model and thus require an annotated corpus of NL queries. The semantic interpretation module transforms the KB- independent query annotation into a query language which is then executed against the KB. The KB is created from the existing data, as described in section 4.2. As a result, a list of KB instances with their properties is displayed to the user. 4. Target Domain The target domain sets constraints for the demonstra- tion of practical and theoretical values of the developed NLI system. Even in cases where the system is domain independent (most of the current NLI systems) a suitable domain will show both the potential and the limitations of the system. As was pointed out in the introduction, we focus on the Czech language in the first place. Due to its high degree of inflection and relatively free word order, it makes the task more challenging. The system can be, however, adapted to English which will be proven later in the evaluation. We decided not to develop our system using any dataset from the existing related work. Firstly, there is no such dataset in the Czech langauge which meets requirements for performing end-to-end evaluation, namely availability of a KB that contains the desired domain information, a corpus of NL queries, and correct “answers” from the KB for each query. Secondly, the existing English corpora limit the queries only to single clauses or sentences. Thirdly, the widely-used Mooney Geoquery [13] corpus is, from our point of view, a typical corpus for the Question Answering task whereas our system is an replacement/enrichment of and existing form-based UI. In the rest of this section, the process of collecting an NL query corpus and creating a KB will be depicted. 4.1. Natural Language Query Corpus Whereas all existing NLI systems presented in the re- lated work section expect queries to be single clauses or single sentences, we decided to go beyond the state of the art by allowing users to formulate their queries using one or more sentences to express their needs more precisely and more naturally. We decided to use a voluntary-based way and started a Facebook campaign which can target a broad audience. The motivation stems from our previous experience with obtaining data for NLI systems [32] which showed us that obtaining data only from i.e. a group of students can neg- atively affect the data (in terms of a higher number of e.g. off-topic or silly queries). The Facebook page contains a clear description of the “virtual” SWSNL system which says: “There is a search 4 Figure 1: The SWSNL system architecture. engine that operates on the accommodation domain in the Czech Republic. Try to search for an accommodation op- tion that you really want by describing it using natural language instead of keywords. You are not limited to a sin- gle sentence.” A few diverse examples followed. At that time, we had no back-end database or knowledge base, thus we did not limit the users to asking queries that can be found in any existing accommodation search portal. On one hand, we did not restrict the user queries to comply with a structure or content of a particular domain model. On the other hand, there was a risk that the KB would not contain the answers to the collected NL queries. We asked each user to pose one to three queries. We gathered 68 queries in total. This was fewer than expected but the whole campaign was based upon spon- taneity and our intention was not to push anyone into participation. Note that the participants were not asked to find answers to their queries. The main goal was to collect just the queries in this step. Examples of the user queries can be found in Appendix A. Table 1 shows some statistics of the NL query cor- pus. Although there was no limit to the query length, the distribution of query lengths (in terms of the number of sentences used) illustrates that most of the users used single sentence and two sentence-long queries. However, the average query length shows that the queries are rel- atively long, which is also important from another point of view—the queries contain a lot of information (see the examples in Appendix A). This makes the corpus quite atypical compared to other existing corpora. Corpus statistics Queries consisting of 1 sentence 37 Queries consisting of 2 sentences 23 Queries consisting of 3 sentences 8 Average query length (words) 24.9 Average number of sentences per a query 1.54 Number of unique participating users 42 Average number of queries per user 1.54 Table 1: Various NL query corpus statistics. 4.2. Knowledge Base and Domain Ontology In order to evaluate a fully functional SWSNL pro- totype application, the underlying KB is essential. The KB is a module that stores complete knowledge about the covered domain. In the Semantic Web, ontologies are the main vehicle for domain modeling. Henceforth, we will use KB in terms of an ontology populated with instances. As a source website for data mining, the portal hotelypenziony.cz was chosen. The site provides a large database of various types of accommodation options (from cottages to luxury hotels). Almost all entries have some location information, such as GPS coordinates or an exact address. Furthermore, each accommodation page contains some information about facilities, prices, etc. but these fields are not mandatory. Some entries are enhanced with textual descriptions, like e.g. a site or a restaurant de- scription. The entire site was downloaded using an in-house craw- ler written in Groovy. A pattern-based information extrac- 5 tion tool was developed in order to extract the structured content from the crawled web pages. The extracted data were then transformed into the corresponding ontology in- stances and properties. The ontology is stored in OWL8 format. The domain ontology structure was designed accord- ing to the extracted information, see Figure 2. The struc- ture is more or less “flat”. It has one important class Accommodation with many object and data properties. The reason for this design stems from the structure of the website (the information source), where one accommoda- tion page is equivalent to one instance of the ontology class Accommodation.9 A transitional relation isSubpartOf (see Figure 2) al- lows capturing a hierarchy of Location instances, e.g. Street is a sub-part of City, etc. However, the prop- erty is not functional. This means that it is possible that one location is a sub-part of more than one other location. Typically, this is the case of <c isSubpartOf a> & <c isSubpartOf t> where c is an instance of City, a is an in- stance of District and t is an instance of TouristRegion. The task of creating such hierarchy only from the given address and the GPS coordinates can be found in [33]. Some existing ontologies for the accommodation do- main were presented in section 2.2. Although some of these ontologies are more complex in terms of number of classes and the relations among them, we decided not to re-use any of these ontologies in our system. The reason was that the structure of the source domain (the website) and the structure of the existing ontologies were so different, that we simply did not see any benefit in trying to fit our data into such different ontologies. 4.2.1. Qualitative Analysis of the Knowledge Base The total number of all instances (individuals) in the KB is 32,990 and the total number of all triples is 281,686. The total number of instances of the Accommodation class is 8641. An accommodation instance contains two types of prop- erties. We will distinguish them as text properties and structured properties. Structured properties. These properties are both da- tatype and object properties (in OWL terminology) with exactly specified semantics. This kind of prop- erty contains structured information such as number values (e.g. various capacities or prices) or facilities (e.g. room facilities, sports activities). Text properties. Text properties are four datatype prop- erties (in OWL terminology), namely accommDesc, 8http://www.w3.org/TR/owl2-overview/ 9The inheritance relation between RoomType and (PrivateDouble, PrivateSingle, ...) was an authors’ design choice. Another way to capture the same information would be e.g. a property numberOfBeds. siteDesc, conferenceDesc and restaurant. These prop- erties are mostly long text fields extracted from the accommodation website. They contain an unstruc- tured description of the accommodation option, e.g. restaurant information, description of the site, etc. Furthermore, these properties do not have an exactly specified semantic meaning, as opposed to structured properties. The structured properties are crucial for searching us- ing a query language. On the contrary, if a piece of infor- mation is available only in text properties, it can hardly be searched for using a query language. In our KB, about 60% of all accommodation instances have no additional property, except for the obligatory contact information. This means that we cannot say anything specific about these accommodation options. The system only knows that a certain accommodation option exists in a certain place but no additional information is offered. However, almost all the user queries (see the examples in Appendix A) relate to a more specific type of accommodation op- tions, with e.g. certain properties, prices, room types, etc. This means that almost 40% of accommodation options cannot be found when users ask for additional accommo- dation requirements. Note that this lack of structured properties is caused by the source web data, not by the ontology structure. 4.2.2. Ambiguity in the Knowledge Base The main advantage of a well-designed ontology is its precise semantics. It allows complex semantic constructs to be used in order to infer new facts and query the KB using a query language. In reality, systems must often deal with existing data since the amount of work needed to create and populate an ontology from scratch would be tremendous. Also our KB is completely created from existing data. As the original web data were not designed to respect precise domain se- mantics, a sort of semantic ambiguity or inconsisteny was transferred to the KB. Some notable problems are: (1) The accommodation types are quite ambiguous. There is no clear difference between e.g. Hut and Cottage, among others. (2) Lots of facilities are duplicated in some sense. For example, the KB contains four similar instances of a parking lot (namely parkingLot, parkingLotWithLights, parkingGarage, and securedParkingGarage). Note that we left all the data unchanged because it was not feasible to check and correct all the inconsistencies manually. 4.3. Gold Data Preparation The correct answer (result) to a NL query is a set of accommodation instances that satisfy the user’s require- ments. Many NL queries contain sorts of subjective ex- pressions, e.g. “cheap”, “nice”, “close to something”, etc. It is necessary to decide in advance how to interpret these expressions consistently. The same interpretation must also be used later by the SWSNL system. 6 Figure 2: Structure of the accommodation ontology. Expression Interpretation near a city all places in the district in which the city is located close to (a place) the distance is less than 1 km cheap any room of the accommodation option has a price lower than 1000 CZK cheapest returns only one cheapest accommodation option Table 2: Semantic interpretation of some NL expressions Table 2 presents some examples of the semantic inter- pretation. The values (prices and distances) were set after a consensus was reached in our research team, however, this interpretation definitely depends on each user’s point of view. We deal only with such expressions that can be quantified or expressed using the KB semantics. 4.3.1. Assigning the Correct Results For this task we combined manually constructed struc- tured and fulltext search queries. Structured search. For each NL query a corresponding SPARQL query was manually constructed. We put as many constraints from the NL query as possible into the SPARQL query. For example, if the user asks for particular facilities (e.g. parking, internet), we add all of them in the SPARQL query. This en- sures that only those accommodation instances that satisfy the constraints are selected from the KB. How- ever, adding such strict criteria has the effect. First, the results that partially satisfy the requirements are discarded.10 No ranking of results is involved. Fulltext search. As presented in section 4.2.1, each ac- commodation instance can have a few text descrip- tions (labels). Such a description usually contains information which might be available in the struc- tured properties as well (e.g. description of hotel facilities in the text and then again as structured properties). These text fields can be indexed and searched through using a fulltext search. Finally, the whole process of assigning the correct re- sults to each NL query was performed as follows: • If the NL query can be completely expressed using SPARQL, create such a query and add the results to the correct results. • If the NL query contains other requirements that cannot be expressed using SPARQL, filter the struc- tured search results with a fulltext search and check the returned results manually. • Moreover, for each NL query create the least restric- tive SPARQL query (only with location and accom- modation type) and filter the results using a fulltext 10This can be solved using OPTIONAL in SPARQL, however, at this stage we focused on results that satisfy all the requirements. 7 search. The fulltext keywords are manually selected according to the NL queries requirements. The re- sults must be checked manually. The gold data (in CSV format) as well as the KB (in OWL format) are licensed under CC-BY-SA11 and are publicly available.12 5. Semantic Description of NL Queries Some traditional approaches how NL query semantics can be expressed are e.g. frames, FOPL (first-order pred- icate logic), or semantic trees, among others. Related re- search [32, 34], based upon semantic trees, showed that describing the semantics of NL queries should follow a particular structure, e.g. using a schema for adding con- straints to the possible semantic trees. Ontologies offer such a mechanism by allowing precise semantic relations between instances to be defined. We will now discuss the term semantic annotation. In the Semantic Web, this term has multiple meanings, e.g. webpages are annotated according to a particular ontol- ogy (taxonomy) or named entities can be annotated with related instances from an ontology for furher named en- tity disambiguation [35], [36], [37]. Our usage comes from NLU research where sentences (or queries) can be anno- tated with some additional semantic information. Hence- forth, the term semantic annotation will be used in the sense of a semantic description of an NL query. On the Semantic Web, ontologies serve as a tool for storing domain information (KB). However, in our pro- posal the semantic annotation of NL queries based upon ontologies has no direct relation to the ontologies for stor- ing domain information (KB). This means that our seman- tic annotation is completely independent of a particular KB. Our semantic annotation uses ontologies exclusively to describe NL query semantics. The motivation for this separation was to enable e.g. switching between various KBs without affecting the NLU module or adapting the NLI to different language and keep the original KB. 5.1. Ontology for NL Query Semantics The domain-independent query ontology is shown in Figure 3. Each word is represented as an instance of the Word ontology class with some additional morphological properties and its position within the sentence. All Word instances are chained using the hasNextWord property. An instance of the Sentence class covers a whole particular NL query. The ontology also supports marking named entities. An instance of NamedEntity is connected to its content 11Creative Commons Attribution-ShareAlike 3.0 Unported Li- cense, http://creativecommons.org/licenses/by-sa/3.0/ 12http://liks.fav.zcu.cz/swsnl/ Figure 3: Domain independent NL query ontology (sentence ontol- ogy). words using a property startsWithWord, its lenght is stored in the numberOfWords property.13 Traditionally, named entities are used to label expres- sions with a special meaning (e.g. cities, persons, coun- tries, etc.). In this work, we extended the usage of a named entity to all words/word groups that carry impor- tant semantic information in the NL query. For example, we added the following domain named entities: relative lo- cation (covering phrases such as “near”, “in the center”, etc.), relative price (e.g. “cheap”), or additional require- ments (e.g. “fitness”, “internet”, etc.), and others. Thus the named entities in our approach can represent both “traditional” entities (e.g. place, accommodation type, fa- cilities, etc.) as well as certain domain-specific knowledge (e.g. relative location or price) and key phrases. We be- lieve that such arbitrary extension of the meaning of the term named entity is still acceptable. Named entities can also form a taxonomy as shown in Figure 3. This taxonomy can be extended in the ontology for a particular domain. For modeling the taxonomy of named entities, we use the inheritance of classes in OWL. For each query, instances of the domain-dependent query ontology classes (triples) are created. A complete example of an annotated query can be found in Appendix B. The NL query ontologies are stored in the OWL format and were developed using Protégé14. 5.2. NL Query Corpus Annotation The whole NL query corpus from section 4.1 was an- notated in the above-mentioned fashion. The annotated corpus is required for training the supervised statistical Natural Language Understanding module as will be shown later in section 6. The corpus was annotated by two independent anno- tators. For this task we developed an annotation tool 13This marking style was our design choice; we are aware of various different approaches to named entity annotation, such as using offsets in GATE. 14Protégé is a free, open source ontology editor and knowledge base framework. http://protege.stanford.edu/ 8 Figure 4: Annotation ontology for queries in the accommodation domain. The dotted lines represent the is-a relation. that provided a balanced trade-off between usability and robustness. The editor hides the complexity of the un- derlying triples and the sentence ontology.15 The inter- annotator agreement was measured on the number of match- ing triples16 from the NL query annotations. The inter- annotator agreement was found to be κ = 0.68, 95% CI (0.504, 0.848). 5.3. Formal Definition of Semantic Annotation Let S = w1 . . .wW be a sentence S (a query) consisting of W words w1...W and let T(Subj,Pred,Obj) be a triple of a subject Subj, a predicate Pred and an object Obj17. Then the semantic annotation Sem of a sentence S is an unordered set of M triples Sem(S) = {T1,T2, . . .TM}. (1) Within Sem(S), the triples can be connected. For- mally, two triples Tx(Subj1,Pred1,Obj1),Ty(Subj2,Pred2,Obj2) (2) can share their subjects or objects, so that Obj1 = Subj2 or Subj1 = Subj2, forming de facto a directed graph. 15This annotation task could be also achieved using other existing tools such as ontology-based gazetteer in GATE or CA Manager [38]. 16This is explained in detail later in section 8.1.2 17Please note that Subj, Pred and Obj are just names of the first, second and third position in the triple. They may or may not be related to the syntactic role of words in a sentence. Named Entities. Let Nτ be a Named Entity (NE) instance with type τ. Then Nτj,k = N τ span(wj . . .wk) (3) is NE of type τ which spans the words wj to wk, where 1 ≤ j < k ≤ W . This means that NE is associated with the words from the sentence. Let C be a semantic concept. C is not associated with any words. Both Nτ and C can be parts of a triple so that each triple has one of the following forms: T(Subj,Pred,Obj) = (4)  T(C,Pred,Nτj,k) if subj is C and obj is NE T(Nτ1j,k,Pred,N τ2 o,p) if subj and obj are NE T(Ca,Pred,Cb) if subj and obj are C where j,k,o,p are the NE spans, a,b are used to de- termine possibly different concepts C and τ1,τ2 can be different NE types. 6. Semantic Analysis of Natural Language Queries The main task of the Natural Language Understanding component is to create a semantic description of previously unseen NL queries. Our NLU component is based upon a statistical model and supervised learning. Furthermore, the component uses both off-the-shelf and newly developed preprocessing tools. All queries are preprocessed using tokenizer, part-of- speech tagger and lemmatizer from the Prague Depen- dency Treebank tokenizer and morphological tagger [39]. 6.1. Statistical model Using the formalism from section 5.3, the task of a semantic analysis is to find an appropriate semantic anno- tation Sem(S) given a sentence S. Using probability, we can formulate the problem as follows: P(Sem(S)|S) = P(T1,T2, . . .TM|S). (5) We approximate the probability by treating the triples as independent, thus P(T1,T2, . . .TM|S) ≈ M∏ P(Tm|S) (6) where P(Tm|S) = P(Tm(Subj,Pred,Obj)|w1 . . .wW ). (7) Furthermore, we limit our model to the first two triple types from Equation 4. This means that all triple ob- jects are NE and some triple subjects are NE, formally T = T(C,Pred,Nτ ) and T = T(Nτ1,Pred,Nτ2 ). This limitation is based upon our observation that the other two types of triples are not currently present in the cor- pus. 9 In our model the triple probabilities are approximated as P(T(C,Pred,Nτ )|S) ≈ (8) P(T(C,Pred,Nτ )|Nτj,k,wj−1) ·P(N τ j,k|S) and P(T(Nτ1,Pred,Nτ2)|S) ≈ (9) P(T(Nτ1j,k,Pred,N τ2 o,p)|N τ1 j,k,N τ2 o,p,wj−1,wo−1) ·P(Nτ1j,k|S) ·P(N τ2 o,p|S), where w is the lemma and P(Nτj,k|S) is obtained from named entity recognizer. Given a training corpus S of sentences and their an- notations S = {Sn,Sem(Sn)}, the probabilities are esti- mated using Maximum Likelihood Estimation (MLE). 6.1.1. Named Entity Recognition Named entity recognition is a crucial part of the query preprocessing. To achieve reasonable performance, we in- corporate three various NER approaches. Maximum Entropy NER. The maximum entropy-based NER (MaxEntNER) introduced in [40] was later extended and trained on a corpus from the Czech News Agency. The parameters of the corpus are very different from our NL query corpus since the Czech News Agency corpus consists of newspaper articles. However, the MaxEntNER is used to identify a few general named entities, such as Place and City. LINGVOParser. A shallow semantic parser based upon handwritten grammars, LINGVOParser [41, 42] is used to identify named entities with complex structure, such as Date, Currency or Number. String Similarity Matching. This NER approach (called WordSimilarityTrainableNER) was used as an ad-hoc tool to deal with the named entities that are not recognized by the two previous NERs. These are usually the domain- dependent entities. Recall that in our semantic annota- tion, the set of named entity types is much broader than in general NER because we use the named entities to cover all semantically important word groups (e.g. named enti- ties like AdditionalRequirements or AccommodationType). The details about string similarity matching will be ex- plained later in section 7.1. OntologyNER. One of the key features of our system is that the NLU module is independent of the back-end KB. However, some systems from the related work (e.g. [11] or [20]) use the KB as a valuable source for recognizing various domain entities, such as places or various facil- ities. Therefore we also developed OntologyNER which exploits the knowledge stored in the KB. This vocabulary- based NER search for the named entity candidates using instance labels from the KB. The search uses string sim- ilarity matching which will be explained later in section 7.1. 7. Semantic Interpretation and Search After the NL query is automatically annotated with its semantic description, it must be interpreted in order to find the desired results. Since our semantic represen- tation of an NL query (see section 5) is independent of a particular KB, it is transformed into an ontology query language (SPARQL) to perform a search in the KB. As an input, a semantic annotation of an NL query is given. As an output, a SPARQL query for the back-end KB is produced. The transformation consists of a set of domain-specific heuristic hand-written rules. Each rule processes partic- ular triples from the NL query annotation and outputs a corresponding SPARQL snippet. An example in Fig- ure 5 demonstrates how the semantic annotation of the NL query is transformed into the corresponding SPARQL query. 7.1. Matching Entities to Ontology Instances Some named entitites recognized in the NL query, such as City or Place, are related to their corresponding ontol- ogy instances. For example, if the NL query search for a certain accommodation option in a particular location (i.e., City[”Praha”] ), the resulting SPARQL query must restrict the search to that entity instance (i.e., ?Accommoda- tion p:isLocatedIn p:cPraha .). Therefore, named en- tity disambiguation is necessary to map the recognized named entities to KB instances. The mapping between named entities and the KB in- stances is based upon string similarity (or string distance). In order to select the best string metrics we manually created a small subset of the ontology instances paired with the named entities (50 pairs) and measured the accu- racy of various types of string metrics. The best accuracy about 96% was achieved with the Jaro-Winkler distance [43]. Only a selected subset of the named entity types is matched using this technique, namely all instances of the ontology class Place and Facility. Using this string similarity-based named entity disam- biguation, three types of errors may occur. First, the named entity is expressed in such a way that the string similarity between the words and the KB instance is very low. Second, the desired named entity has no correspond- ing instance in the KB. These two types of errors are quite obvious. The third type of error is caused by the ambi- guity within the KB. In some cases, two very similar in- stances have a very similar description. For example, the KB contains more than 30 instances with Praha (Prague) in their labels (referring to e.g. District, City or a city 10 Figure 5: An example of tranforming an annotated NL query into SPARQL. The NL query semantic annotation sample contains two triples in this example, namely (Accommodation, hasRoomTypeRequirement, RoomTypeRequirement[”jednol̊užkové”]) and (RoomTypeR- equirement[”jednol̊užkové”], hasNumberOfRoomsReq, Number[”dva”]) where RoomTypeRequirement and Number represent named entities. These two triples are translated into the two corresponding SPARQL snippets on the right-hand side. Notice that the NL query annotation is created according to the NL query ontology (see Figure 4), whereas the SPARQL query is bound to the particular back-end KB (see Figure 2). part). Currently, we do not deal with this kind of ambigu- ity. However, it could be easily solved i.e. by choosing the most likely instance learned from the training data or by showing a clarification dialogue (the case of [11] or [9]). 7.2. Interpretation of Named Entities with Numbers and Dates The named entitity types Number and Date are not related to any ontological instances (they do not directly represent any instance from the KB). The LINGVOPar- ser tool used to recognize this type of named entities (see section 6.1.1) also has built-in support for the semantic in- terpretation of the content of the named entities, namely numbers and dates. The tool utilizes hand-written seman- tic grammars for translation into a computer representa- tion (e.g. integers or dates). This allows interpretations of these entities to be integrated into the resulting SPARQL query. 7.3. Semantic Reasoning and Result Representation Given all the mechanisms introduced in the previous sections, it is possible to formulate the SPARQL query. Once the query is created, it can be passed to an ontol- ogy reasoner. Currently, various OWL reasoner Java im- plementations are available. They also vary in their OWL expressivity, which means what subset of the OWL seman- tics they can handle. To select a suitable reasoner for our system, a simple test was carried out. The most important OWL feature we use is the transitivity among location in- stances (see section 4.2). Thus, a simple SPARQL query was executed to obtain all instances in a particular loca- tion and its sub-locations. Table 3 outlines the results of this test. We tested reasoners from the Jena package as well as the Pellet reasoner. Only reasoners that sup- port transitive properties are listed. Given the measured performance among reasoners, the Pellet reasoner was se- lected for our system. The performance of the reasoners will be discussed later in section 9. Reasoner class Total time PelletReasoner 15m 14s OWLMicroReasoner 55m 5s OWLFBRuleReasoner 5h 15m 26s Table 3: Performance comparison of various OWL reasoners. The result of the semantic search is a set of accom- modation instances. The system displays all the retrieved accommodation instances together with their properties. 8. Evaluation In this section, we thoroughly evaluate our SWSNL system. The section is structured as follows. First, we focus on the evaluation of the Natural Language Under- standing module (introduced in section 6). Various com- binations of NER are also tested in this part of the eval- uation. Second, we present an end-to-end evaluation of our system. We describe the baseline approach and com- pare our system results with the baseline results. Third, we evaluate our NLU module on another two corpora and languages to demonstrate its portability capabilities. Our main corpus (the NL Query corpus on the accom- modation domain, see section 4.1) consists of 68 queries. We split the data into training and test set using the leave- one-out cross-validation [44]. In this method, each fold of the cross validation has only a single test example and all the rest of the data is used in training. We use this method because the available data is very small. 8.1. Natural Language Understanding Module Evaluation The NLU module (section 6) works in two stages: the Named Entity Recognition and the ontology-based statis- tical model for the semantic analysis. Its evaluation will be split into two sections, too. We use standard metrics, namely the precision (p), the recall (r) and the F-measure 11 Entity type \ NER WSim MaxEnt Onto LinPar AccommodationType 0.90 0.0 0.0 0.0 Place 0.37 0.45 0.22 0.0 Date 0.0 0.22 0.0 0.28 NumberOfPersons 0.62 0.0 0.0 0.0 RelativeLocationReq 0.70 0.0 0.0 0.0 City 0.33 0.70 0.35 0.0 Price 0.72 0.13 0.0 0.0 AdditionalReq 0.57 0.0 0.04 0.0 Number 0.32 0.10 0.0 0.37 RelativePrice 0.87 0.0 0.0 0.0 RoomTypeReq 0.55 0.0 0.0 0.0 Table 4: F-measure performance of various NER on different entity types. Some NER names are abbreviated; WSim is WordSimilari- tyTrainableNER and LinPar is LINGVOParserNER. (Fm): p = tp tp + fp , r = tp tp + fn , Fm = 2pr p + r , (10) where tp are true positives, fp are false positives, and fn are false negatives [45]. 8.1.1. Evaluation of NER The system uses four types of NER, namely the Max- EntNER, the WordSimilarityTrainableNER, the LINGVO- Parser, and the OntologyNER (see section 6.1.1). The cor- rect result (true positive) is when the named entity from the annotated corpus exactly matches the entity returned by NER, in terms of both the NE type and the word span. The results are shown in Table 4 and can be interpreted as follows. The second column (WSim) displays the results for the WordSimilarityTrainableNER. This NER is based on lexical similarity between the training examples and the unknown word(s). This NER performs best mostly for domain-dependent entities, such as AdditionalReq or Ac- commodationType, because these entities are expressed us- ing similar words in the data. We assume that this ap- proach is sufficient for domain-dependent entities. Our assumption that the MaxEntNER, even though it is trained on a completely different corpus, would yield acceptable results for general named entities, such as cities or places, is validated by the results in the third column (MaxEnt). This NER outperforms the rest of the NERs in City by almost 100% relatively. The poor results obtained by OntologyNER in the fourth column have the following explanation. This NER ex- ploits the lexical similarity between KB instances and the word(s) being recognized. The results for named entities City and Place are thus similar to the results from the sec- ond column (WordSimilarityTrainableNER). However, the performance of the OntologyNER is surpassed by the Max- EntNER, as it involves many features for entitity recogni- tion, such as the local context, word shape, etc. Finally, the last column shows that for complex named entities in natural language, such as dates or numbers, the entity-specific hand-crafted NER LINGVOParser would be the best choice. CombinedNER. For each entity type we selected the best performing NER from Table 1 and created the Combined- NER. This NER will be used in the overall evaluation and the semantic analysis evaluation in the next sections. How- ever, we expect that with larger training set it would be reasonable to use a more clever NER combination, e.g. a linear interpolation model with combination parameters set by the Expectation-Maximization algorithm. 8.1.2. Semantic Analysis Evaluation The result of the semantic analysis is a semantic an- notation of a NL query (recall section 5.3). However, the annotation of a NL query can also be seen as a set of connected triples. For each NL query, we compute preci- sion, recall and F-measure using the set of the gold triples (the human annotation from the NL query corpus) and the triples returned by the semantic analysis module.18 A true positive result triple is such a triple that matches the gold triple as follows: • The predicates are the same. • The objects/subjects have the same ontology class, i.e. <City, pred, obj> and <City, pred, obj> match while <City, pred, obj> and <Place, pred, obj> do not. • Furthermore, if the objects/subjects are named en- tities, their positions in the sentence and their con- tents are the same, i.e. <subj, pred, City(Velké1, Popovice2)> and <subj, pred, City(Popovice2)> do not match. The results of semantic analysis module (with the Com- binedNER incorporated) are shown in Table 5. The most important factor that influences the performance is the small size of the NL query corpus. Only 67 training ex- amples19 are not sufficient to train the semantic model properly. Moreover, the very important part of the Com- binedNER for recognizing non-general named entities, the WordSimilarityTrainableNER, is also completely trained from the data. p r Fm CombinedNER + Semantic Model 0.66 0.34 0.45 Table 5: The results of the semantic analysis with the CombinedNER 18We also considered some other measures developed for seman- tic annotation evaluation, but none of them was suitable for our purposes. BDM [46] computes semantic similarity between two an- notations of the same token in a document. The tree edit distance [47] expects the semantic annotation to form a tree. 19Note that we still use the leave-one-out cross-validation which ensures that the training and test data are distinct. 12 Figure 6: The semantic model performance depending on the train- ing set size. In the following test, we evaluated the semantic ana- lysis model’s dependency on the training data size. Using the same NL query corpus, we simulated 100% accuracy of the NER module (called MockNER) and changed the proportions of the training and test sets. Given the results from Figure 6, we can draw the following conclusion. Even with small training data, most of the returned triples are correct (p → 1). However, to obtain all triples of the semantic annotation of a single NL query, a larger training set is probably required (r is growing more slowly). 8.2. End-to-end Evaluation Each query from our NL query corpus has its own cor- responding set of correct search results (gold results), as described in section 4.3.1. This allows us to evaluate the complete SWSNL system. We performed three experi- ments. First, the baseline fulltext-based search system was evaluated. Second, we tested our SWSNL system. Third, we performed the same test with a simulation of 100% correct Semantic Analysis module. For each query the system returns a set of accommoda- tion instances. There is no ranking of the results because the KB querying is purely boolean which means that the single result either matches or not and no further discrim- ination is given. Thus, for each NL query we compute the precision, the recall, and the F-measure of the returned re- sult set according to the gold set. The final result is then the average of the results for each query (the Macro F-1 measure). 8.2.1. Baseline System As a baseline for the end-to-end evaluation we used the fulltext search. The crawled documents from the source website were indexed using Apache Lucene.20 In the pre- processing, the best available Czech stemmer [48] was used, 20http://lucene.apache.org/core/ p r Fm (a) Baseline fulltext 0.02 0.42 0.03 (b) SWSNL 0.25 0.70 0.37 (d) Correct semantic annotation 0.73 0.84 0.78 Table 6: The end-to-end results of (a) the baseline fulltext search system, (b) the search using our SWSNL system, and (c) the search using the correct semantic annotation of NL queries. the stop-words were filtered and the content was lower- cased. In this experiment, each NL query was treated as a bag of words with stop-words removed. Other standard measures used in Information Retrieval (IR) for evaluating ranked results are Mean Average Pre- cision, Precision at k (P@k) and R-Precision (R-Prec), see e.g. [45]. These metrics are suitable only for ranked results as retrieved e.g. by a fulltext search. Since we compare the baseline with our SWSNL system which does not pro- duce ranked results, it is not reasonable to measure the baseline using these metrics. The baseline system results are shown in Table 6, row (a). The results show very clearly that the keyword-based search performs poorly for such complicated tasks. The relatively high recall means that there are many hits in the retrieved set. The low precision is caused by the fact that the search model is not discriminative (i.e. it does not filter the results according to a certain location, etc.). However, this might serve as an approximation of results that would be returned by the current state-of-the-art general-purpose search engines. The parameter N for selecting only the top N retrieved documents was set empirically in order to maximize the F-measure. 8.2.2. SWSNL System Results Table 6, row (b) shows the overall results of the SWSNL system. The most important conclusion, given these re- sults, is that our semantic web search significantly out- performs the fulltext search. Whereas the fulltext-based search simply fails in this complex search task, our SWSNL system is able to yield acceptable results. The obtained F-measure of 36% may seem low but high numbers are not common in challenging IR tasks. 8.2.3. SWSNL System Results With Simulation of Correct Semantic Analysis We also tested the upper bounds of our SWSNL ap- proach by simulating the 100% correct Semantic Analy- sis module. The numbers in Table 6, row (c), confirm that some results cannot be found by the purely struc- tured search. This topic was already discussed in section 4.3 where the process of creating the gold results was pre- sented. In more detail, some of the results in the gold data were manually assigned to the queries according to their textual description. Such results are, however, not cur- rently retrieved by our SWSNL system, since the search operates only on structured properties (cf. section 4.2.1). The second reason that affects the performance in this test 13 is that some errors and ambiguity are encountered during mapping named entities to KB instances (as already dis- cussed in section 7.1). 8.3. Evaluation of NLU Module on Other Domains and Languages Although we put our main effort into the end-to-end evaluation on the accommodation domain in the Czech language, we also tested our NLU component on two dif- ferent corpora in order to prove its usability across do- mains and languages.21 In the following section, only the NLU module is evaluated as there exist no databases/KBs for these datasets for an end-to-end evaluation. 8.3.1. Czech Public Transportation Queries Corpus The Czech Public Transportation Queries Corpus (Con- nectionsCZ) is a NL query corpus in the Czech language. It contains 238 queries in the public transportation do- main. Most of the queries ask for a certain connection from one place to another, for train types, travel times, etc. The corpus is a subset of a larger corpus used in [32]. This type of corpus follows the fashion of corpora for Spoken Language Understanding (SLU) or human-com- puter spoken dialogue, such as ATIS [49], etc. Since our current work deals with search and Semantic Web tech- nologies, we are not convinced that our SWSNL system would be the best approach for this domain and task. Nev- ertheless, this evaluation serves as proof that our semantic model is domain independent and can perform well if a larger training corpus is available. In order to port the corpus into our system, we created an ontology for an NL query in this particular domain (see section 5). The domain-dependent query ontology is shown in Figure 7. The queries were then annotated according to this ontology. We conducted the two following experiments. In the first one, we used the simulation of 100% NER. Figure 8 (a) illustrates the improvement in the semantic analysis as the training set size grows. In the second experiment, we used the same NER con- figuration as in the accommodation domain (namely the Combined NER, see section 8.1.1). The results are shown in Figure 8 (b). The improvement in the performance is caused by the larger training size for both the NER and the semantic model. 8.3.2. ATIS Corpus Subset The ATIS (Air Travel Information Service) [49] cor- pus contains travel information queries in English and it has been widely used in NLU research. It is one of the commonly used corpora for testing of semantic analysis systems, used for evaluation in e.g. [50], [51], [52] and [53]. 21Both corpora can be found at http://liks.fav.zcu.cz/swsnl/. Figure 7: The ontology of the queries in the public transportation domain. The dotted lines represent the is-a relation. We extracted a small subset of the corpus (348 sen- tences) and annotated them using the same query ontology as for the Czech Public Transportation Queries Corpus, see previous section and Figure 7. Since our NER components are able to deal with the Czech language only, we performed only one test with the simulation of 100% NER on the ATIS corpus. The test is intended to evaluate the performance of our semantic model on a different language. The results are shown in Figure 8 (c). We decided to test our system on the ATIS corpus just to get an insight into the possibilities of porting the seman- tic model to another language. This should not serve as a comparison with other state-of-the-art semantic parsers, for two reasons. Firstly, our semantic annotation is differ- ent from the original frame-based ATIS annotation. Sec- ondly, we use a small subset of the whole corpus, which also negatively affects the performance. The complete cor- pus was not used because a manual annotation according to our semantic representation was required and it was not feasible to annotate the complete corpus. The obtained results illustrate that our semantic anal- ysis model can be used across different languages. No tweaking is required to adapt the system to English. The achieved performance (about 0.73 F-measure) is satisfac- tory, given the fact that the system was primarily devel- oped on a different domain and language. 8.3.3. Final Remarks to Evaluation We will also comment on the fact that our system was not evaluated on the already-mentioned Mooney Geoquery dataset (see section 2.3). Supported by our knowledge of question answering [54] the Mooney Geoquery is a typi- 14 Figure 8: The evaluation of the NLU module on different domains and languages w.r.t. the increasing sizes of the training data. (a) Semantic analysis on the Czech Public Transportation Queries Corpus with a simulation of a 100% correct NER. (b) Semantic analysis on the Czech Public Transportation Queries Corpus with the Combined NER. (c) Semantic analysis on the English ATIS corpus with a simulation of a 100% correct NER. cal factoid question-answering corpus. The answers may be lists of instances (e.g. ”what states border florida ?”), numbers (e.g. ”how long is the rio grande river ?”), or an aggregated result (e.g. ”how many square kilometers in the us ?”). On contrary, our system is developed as a replacement of a form-based UI for information retrieval. The types of queries in our corpus are significantly differ- ent from the Mooney Geoquery queries. In our case, the user wants to find a list of instances that fulfill his or her requirements instead of finding some facts about entities. 9. Conclusion This final section outlines the major contributions of this article as well as open issues and future work. A Semantic Web Search using Natural Language is, beyond all doubt, a challenging task. In order to develop a real- world system that could be deployed for a public use, many theoretical and practical problems must be resolved. We will examine these problems from various perspectives and also propose solutions to some of them. 9.1. Performance Issues One of the most critical issues which actually prevents our system from being tested in the real Web environment is the performance of the ontology reasoning. As already shown in Table 3 on page 11, we tested several OWL rea- soners and their ability to deal with transitive properties. The Pellet reasoner requires approximately 15 minutes on average to answer a single SPARQL query with transi- tive relations. We suspect that the reason for this is the size of our KB (281,686 triples, see section 4.2.1). However, to the best of our knowledge, there is no room for improve- ment without changing the back-end Semantic Web tools completely. Possible solutions. First, another ontology storage engine can be used, e.g. Sesame22. Second, it is possible to do materialisation first and then querying (as is OWLIM23). 9.2. Problems Caused by Actual Web Data Many practical issues related to the KB content were pointed out in section 4.2.1. The most important problem is the missing structured data. In other words, the data from the source Web site are incomplete. However, this will be the case in most of the systems that use public Web sources for populating their KBs. It is not feasible either to check or even to correct the data manually. Another problem is caused by the inconsistent source data. In our scenario, various KB instances have the same label which makes the mapping from the named entities to ontology instances (NE disambiguation) much harder (see section 7.1). Possible solutions. Better data post-processing after the information extraction step. Nevertheless, a huge manual effort would be required. 9.3. Future Work There are four directions that are worth exploring fur- ther: 22http://www.openrdf.org/ 23http://www.ontotext.com/owlim 15 Larger NL query corpus. The main limitation of our cur- rent semantic model performance is the small size of the corpus. Thus, it would be beneficial to obtain more data in e.g. a bootstrapping manner when the prototype is de- ployed on a testing server and is accessible by the public. Integrating fulltext search. A combination of structured search and fulltext search is a promising future task, based upon our preliminary research. Better Named Entity Recognition. The results definitely show that the NER component is crucial in the semantic search. The better the NER is, the better the semantic model performs and the more precise the search results are. Performance Improvement. Replacing the back-end with more scalable Semantic Web tools or with a relational database. 9.4. Notes to Related Work We would also like to discuss some of our findings re- garding the related work. Since our goal was to design a fully functional real-world system, many details must have been taken into account. However, in some related work these details were ignored od hidden, from our point of view. Firstly, many systems operate either on small on- tologies [10, 9] or simple domains [55, 17] therefore it can- not be assumed how would these approches perform on a decent-scale ontologies. Secondly, the natural language queries used for development and testing of some systems are mostly collected by the research teams themselves or by their students [17, 56, 13, 14, 57, 26]. In our opinion, a different way of obtaining queries should be considered in a real-world scenario to avoid any potential skewness in the data. Thirdly, many systems rely on full syntax processing of the queries and a heuristic processing of the semantics in their NLU component [9, 58, 10, 11]. Although syntactic parsing and using rules for NLU have been a widely used approach, however, in some cases this may not a feasible solution, e.g. when a syntactic parser is not available for that particular language. Finally, an exhaustive evalua- tion and a realistic discussion of the limitations is missing in some systems [59, 14]. 9.5. Main Contributions This article describes an attempt to develop a fully- functional semantic search system with a natural language interface. The following list summarizes the main contri- butions. The complete accommodation web portal was converted to the OWL ontology which allows a geospatial search us- ing an automatically-created hierarchy of locations. We also identified many practical problems that are likely to arise when a system must deal with noisy data obtained from the Web. A corpus of natural language queries in the Czech lan- guage was collected using a social media-based approach. We proposed a statistical semantic model for semantic analysis of natural language queries. It is a language- independent model which is trained from labeled data and does not depend on syntactic parsing. We created an evaluation dataset on which our system was tested. This dataset may be a valuable resource for any related research on natural language interfaces operat- ing on ontologies or as a front-end to information retreival systems. We offer this dataset and two other corpora un- der a licence which allows both research and commercial purposes. Acknowledgments This work was supported by grant no. SGS-2010-028 and by the European Regional Development Fund (ERDF), project “NTIS - New Technologies for Information Soci- ety”, European Centre of Excellence, CZ.1.05 /1.1.00/02.0090. Appendix A. Query Examples Here are a few examples of NL queries in Czech with their English translation from the NL query corpus as in- troduced in 4.1. • Ráda bych se ubytovala v Karľstejně s výhledem na hrad a královské vinice, nejlépe s polopenźı, parková- ńım a úschovnou kol, pro 2 lidi od pátku do neděle. — I’d like to have accommodation for two people in Karľstejn with a good view of the castle, including half-board, a parking lot, and a bike to hire from Friday till Sunday. • Jsme dva a chceme strávit jarńı prázdniny v maleb- ném hotýlku nebo apartmánu v Beskydech, bĺızko mo- dré sjezdovky, předpokládáme lyžárnu, vyžadujeme polopenzi. Bazén, pingpong, wellness a podobné le- grácky potěš́ı, ale nejsou nezbytné. — Two persons want to spend the spring holidays in a cute hotel or apartment in the Beskydy Mountains. We require a ski-locker room and full board. Swimming pool, a ping-pong table and some wellness will be a plus but that’s not necessary. • Hledám levné v́ıkendové ubytováńı pro 3 odoby v Jin- dřichově Hradci nebo bĺızkém okoĺı, v bĺızkosti p̊ujčov- ny j́ızdńıch kol nebo s možnost́ı uskladněńı vlastńıch, pokoj nejlépe s krbem a možnost́ı připojeńı na inter- net, vlastńı sociálńı zař́ızeńı podmı́nkou. — I’m look- ing for cheap weekend accommodation for 3 persons in the area of Jindřich̊uv Hradec, near to a bike- rental store or with a bike locker room. Room in- cluding a fireplace, Internet and a private bathroom. 16 • Hledám hotel na pořádáńı mezinárodńı konference dne 25.10.. Hotel muśı mı́t sál pro minimálně 100 lid́ı, wifi pokryt́ı, možnost uspořádáńı večeře, a mini- málně 10 dv̊ujl̊užkových pokoj̊u. Upřednostňuji hotel s vlastńı prezentačńı technikou — I’m looking for a hotel suitable for hosting an international conference on October 25. We require a large hall (100 people at least), Wi-Fi, banquet, conference equipment. 10 double-bed rooms are minimum. Appendix B. Complete NL Query Annotation Ex- ample Please refer to Figure B.9. References [1] M. L. Wilson, B. Kules, m. c. schraefel, B. Shneiderman, From keyword search to exploration: Designing future search inter- faces for the web, Foundations and Trends in Web Science 2 (1) (2010) 1–97. [2] E. Demidova, P. Fankhauser, X. Zhou, W. Nejdl, Divq: di- versification for keyword search over structured databases, in: Proceedings of the 33rd international ACM SIGIR conference on Research and development in information retrieval, SIGIR ’10, ACM, New York, NY, USA, 2010, pp. 331–338. [3] H. He, W. Meng, C. Yu, Z. Wu, Constructing interface schemas for search interfaces of web databases, in: A. Ngu, M. Kitsure- gawa, E. Neuhold, J.-Y. Chung, Q. Sheng (Eds.), Web Informa- tion Systems Engineering – WISE 2005, Vol. 3806 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2005, pp. 29–42. [4] E. Kaufmann, A. Bernstein, How useful are natural lan- guage interfaces to the semantic web for casual end-users?, in: Proceedings of the 6th international The semantic web and 2nd Asian conference on Asian semantic web conference, ISWC’07/ASWC’07, Springer-Verlag, Berlin, Heidelberg, 2007, pp. 281–294. [5] E. Prud’hommeaux, A. Seaborne, SPARQL Query Language for RDF, W3C Recommendation (2008). [6] K. Elbedweihy, S. N. Wrigley, F. Ciravegna, Evaluating se- mantic search query approaches with expert and casual users, in: 11th International Semantic Web Conference (ISWC2012), Boston, USA, 2012. [7] K. Elbedweihy, S. N. Wrigley, F. Ciravegna, D. Reinhard, A. Bernstein, Evaluating semantic search systems to identify future directions of research, in: R. Garćıa-Castro, N. Lyndon, S. N. Wrigley (Eds.), Second International Workshop on Eval- uation of Semantic Technologies, no. 843 in CEUR Workshop Proceedings, 2012, pp. 25–36. [8] V. Lopez, M. Fernández, E. Motta, N. Stieler, PowerAqua: Sup- porting users in querying and exploring the Semantic Web, Se- mantic Web (2011) 1–17. [9] V. Lopez, V. Uren, E. Motta, M. Pasin, AquaLog: An ontology- driven question answering system for organizational semantic intranets, Web Semantics 5 (2) (2007) 72 – 105. [10] P. Cimiano, P. Haase, J. Heizmann, M. Mantel, R. Studer, To- wards portable natural language interfaces to knowledge bases – the case of the orakel system, Data and Knowledge Engineering 65 (2) (2008) 325 – 354. [11] D. Damljanovic, M. Agatonovic, H. Cunningham, Natural lan- guage interfaces to ontologies: Combining syntactic analysis and ontology-based lookup through the user interaction, in: L. Aroyo, G. Antoniou, E. Hyvönen, A. ten Teije, H. Stucken- schmidt, L. Cabral, T. Tudorache (Eds.), The Semantic Web: Research and Applications, Springer Berlin / Heidelberg, 2010, pp. 106–120. [12] Cunningham et al., Text Processing with GATE (Version 6), university of Sheffield Department of Computer Science (April 2011). [13] L. R. Tang, R. J. Mooney, Using multiple clause constructors in inductive logic programming for semantic parsing, in: Proceed- ings of the 12th European Conference on Machine Learning, Springer-Verlag, London, UK, 2001, pp. 466–477. [14] C. Wang, M. Xiong, Q. Zhou, Y. Yu, Panto: A portable nat- ural language interface to ontologies, in: Proceedings of the 4th European conference on The Semantic Web: Research and Applications, ESWC ’07, Springer-Verlag, Berlin, Heidelberg, 2007, pp. 473–487. [15] E. Kaufmann, A. Bernstein, L. Fischer, NLP-Reduce: A ”näıve” but Domain-independent Natural Language Interface for Query- ing Ontologies, in: 4th European Semantic Web Conference (ESWC 2007), 2007, pp. 1–2. [16] Óscar Ferrández, R. Izquierdo, S. Ferrández, J. L. Vicedo, Ad- dressing ontology-based question answering with collections of 17 Figure B.9: This example displays a complete semantic annotation of a query from the accommodation corpus. The query can be translated as ”I would like one double-bed room and two four-bed rooms including breakfast somewhere in Pec pod Sněžkou or Harachov from 25. 12. till 1. 1.”. We added the translation to the corresponding tokens for easy understanding. 18 user queries, Information Processing & Management 45 (2) (2009) 175 – 188. [17] A. Frank, H.-U. Krieger, F. Xu, H. Uszkoreit, B. Crysmann, B. Jörg, U. Schäfer, Question answering from structured knowl- edge sources, Journal of Applied Logic 5 (1) (2007) 20 – 48. [18] C. Pollard, I. A. Sag, Information-based syntax and semantics: Vol. 1: fundamentals, Center for the Study of Language and Information, Stanford, CA, USA, 1988. [19] A. Copestake, D. Flickinger, C. Pollard, I. Sag, Minimal re- cursion semantics: An introduction, Research on Language and Computation 3 (2005) 281–332. [20] D. Damljanovic, V. Tablan, K. Bontcheva, A text-based query interface to owl ontologies, in: N. Calzolari, K. Choukri, B. Mae- gaard, J. Mariani, J. Odjik, S. Piperidis, D. Tapias (Eds.), Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), European Language Re- sources Association (ELRA), Marrakech, Morocco, 2008, pp. 361–375. [21] G. Zenz, X. Zhou, E. Minack, W. Siberski, W. Nejdl, From keywords to semantic queries–incremental query construction on the semantic web, Web Semantics 7 (3) (2009) 166 – 176. [22] S. Kara, Özgür Alan, O. Sabuncu, S. Akpınar, N. K. Cicekli, F. N. Alpaslan, An ontology-based retrieval system using se- mantic indexing, Information Systems 37 (4) (2012) 294 – 305. [23] B. Fazzinga, T. Lukasiewicz, Semantic search on the web, Se- mantic Web 1 (1-2) (2010) 89–96. [24] D. Damljanovic, V. Devedžic, Semantic Web and E-Tourism, in: M. Khosrow-Pour (Ed.), Encyclopedia of Information Science and Technology, Second Edition, IGI Global, 2009, pp. 3426– 3432. [25] H. Knublauch, Ontology-Driven Software Development in the Context of the Semantic Web: An Example Scenario with Pro- tege/OWL, in: D. S. Frankel, E. F. Kendall, D. L. Mcguinness (Eds.), 1st International Workshop on the Model-Driven Se- mantic Web (MDSW2004), 2004. [26] J. Ruiz-Mart́ınez, D. Castellanos-Nieves, R. Valencia-Garćıa, J. Fernández-Breis, F. Garćıa-Sánchez, P. Vivancos-Vicente, J. Castejón-Garrido, J. Camón, R. Mart́ınez-Béjar, Accessing touristic knowledge bases through a natural language interface, in: D. Richards, B.-H. Kang (Eds.), Knowledge Acquisition: Approaches, Algorithms and Applications, Springer Berlin / Heidelberg, 2009, pp. 147–160. [27] D. Yoo, Hybrid query processing for personalized information retrieval on the semantic web, Knowledge-Based Systems 27 (0) (2012) 211 – 218. [28] J. Scicluna, N. Steinmetz, Modelling e-tourism services and bundles, in: R. Law, M. Fuchs, F. Ricci (Eds.), Information and Communication Technologies in Tourism 2011, Springer Vi- enna, 2011, pp. 403–415. [29] J. Scicluna, N. Steinmetz, M. Zaremba, Service bundling with seekda! dynamic shop, in: U. Gretzel, R. Law, M. Fuchs (Eds.), Information and Communication Technologies in Tourism 2010, Springer Vienna, 2010, pp. 369–380. [30] D. Damljanović, K. Bontcheva, Towards enhanced usability of natural language interfaces to knowledge bases, in: V. Devedžić, D. Gašević (Eds.), Web 2.0 & Semantic Web, Springer US, 2009, pp. 105–133. [31] P. Cimiano, M. Minock, Natural language interfaces: What is the problem? – a data-driven quantitative analysis, in: H. Ho- racek, E. Métais, R. Muñoz, M. Wolska (Eds.), Natural Lan- guage Processing and Information Systems, Springer Berlin / Heidelberg, 2010, pp. 192–206. [32] I. Habernal, M. Konoṕık, Semantic annotation for the lingvose- mantics project, in: Proceedings of the 12th International Con- ference on Text, Speech and Dialogue, TSD ’09, Springer- Verlag, Berlin, Heidelberg, 2009, pp. 299–306. [33] I. Habernal, Semantic web search using natural language, Ph.D. thesis, University of West Bohemia, Faculty of Applied Sciences (2012). [34] I. Habernal, M. Konoṕık, On the way towards standardized semantic corpora for development of semantic analysis systems, in: SEMAPRO 2010: The Fourth International Conference on Advances in Semantic Processing, IARIA, 2010, pp. 96–99. [35] A. Kiryakov, B. Popov, I. Terziev, D. Manov, D. Ognyanoff, Semantic annotation, indexing, and retrieval, Web Semantics 2 (1) (2004) 49 – 79. [36] P. Cimiano, S. Handschuh, S. Staab, Towards the self- annotating web, in: Proceedings of the 13th international con- ference on World Wide Web, WWW ’04, ACM, New York, NY, USA, 2004, pp. 462–471. [37] L. Ding, T. Finin, A. Joshi, R. Pan, R. S. Cost, Y. Peng, P. Red- divari, V. Doshi, J. Sachs, Swoogle: a search and metadata en- gine for the semantic web, in: Proceedings of the thirteenth ACM international conference on Information and knowledge management, CIKM ’04, ACM, New York, NY, USA, 2004, pp. 652–659. [38] D. Damljanovic, F. Amardeilh, K. Bontcheva, CA manager framework: creating customised workflows for ontology pop- ulation and semantic annotation., in: Y. Gil, N. F. Noy (Eds.), K-CAP, ACM, 2009, pp. 177–178. [39] J. Hajič, J. Panevová, E. Hajičová, J. Panevová, P. Sgall, P. Pa- jas, J. Štěpánek, J. Havelka, M. Mikulová, Prague dependency treebank 2.0, linguistic Data Consortium, Philadelphia (2006). [40] M. Konkol, M. Konoṕık, Maximum entropy named entity recog- nition for czech language, in: Proceedings of the 14th inter- national conference on Text, speech and dialogue, TSD’11, Springer-Verlag, Berlin, Heidelberg, 2011, pp. 203–210. [41] I. Habernal, M. Konoṕık, Active tags for semantic analysis, in: Proceedings of the 11th international conference on Text, Speech and Dialogue, TSD ’08, Springer-Verlag, Berlin, Heidel- berg, 2008, pp. 69–76. [42] M. Konoṕık, I. Habernal, Hybrid semantic analysis, in: Pro- ceedings of the 12th International Conference on Text, Speech and Dialogue, TSD ’09, Springer-Verlag, Berlin, Heidelberg, 2009, pp. 307–314. [43] W. E. Winkler, String comparator metrics and enhanced de- cision rules in the Fellegi-Sunter model of record linkage, in: Proceedings of the Section on Survey Research (American Sta- tistical Association), 1990, pp. 354–359. [44] B. Liu, Web Data Mining: Exploring Hyperlinks, Contents, and Usage Data, 2nd Edition, Springer-Verlag New York, Inc., Se- caucus, NJ, USA, 2011. [45] C. D. Manning, P. Raghavan, H. Schütze, Introduction to Infor- mation Retrieval, Cambridge University Press, New York, NY, USA, 2008. [46] D. Maynard, W. Peters, Y. Li, Evaluating evaluation metrics for ontology-based applications: Infinite reflection., in: LREC, European Language Resources Association, 2008. [47] K.-C. Tai, The tree-to-tree correction problem, J. ACM 26 (3) (1979) 422–433. [48] L. Dolamic, J. Savoy, Indexing and stemming approaches for the czech language, Information Processing & Management 45 (6) (2009) 714 – 720. [49] D. Dahl, M. Bates, M. Brown, W. Fisher, K. Hunicke- Smith, D. Pallett, C. Pao, A. Rudnicky, E. Shriberg, J. Garo- folo, J. Fiscus, D. Danielson, E. Bocchieri, B. Buntschuh, B. Schwartz, S. Peters, R. Ingria, R. Weide, Y. Chang, E. Thayer, L. Hirschman, J. Polifroni, B. Lund, G. Kawai, T. Kuhn, L. Norton, ATIS3 Test Data, linguistic Data Con- sortium, Philadelphia (1995). [50] Y. He, S. Young, Semantic processing using the hidden vector state model, Computer Speech & Language 19 (1) (2005) 85– 106. [51] E. Iosif, A. Potamianos, A soft-clustering algorithm for au- tomatic induction of semantic classes, in: Interspeech-07, Antwerp, Belgium, 2007, pp. 1609–1612. [52] M. Jeong, G. G. Lee, Practical use of non-local features for sta- tistical spoken language understanding, Computer Speech and Language 22 (2) (2008) 148–170. [53] C. Raymond, G. Riccardi, Generative and discriminative algo- rithms for spoken language understanding, in: Interspeech-07, Antwerp, Belgium, 2007, pp. 1605–1608. 19 [54] I. Habernal, M. Konoṕık, O. Rohĺık, Question Answering, in: C. Jouis, I. Biskri, J.-G. Ganascia, M. Roux (Eds.), Next Gen- eration Search Engines: Advanced Models for Information Re- trieval, IGI Global, 2012, in press. [55] V. Tablan, D. Damljanovic, K. Bontcheva, A natural language query interface to structured information, in: Proceedings of the 5th European semantic web conference on The semantic web: research and applications, ESWC’08, Springer-Verlag, Berlin, Heidelberg, 2008, pp. 361–375. [56] J. Ruiz-Mart́ınez, D. Castellanos-Nieves, R. Valencia-Garćıa, J. Fernández-Breis, F. Garćıa-Sánchez, P. Vivancos-Vicente, J. Castejón-Garrido, J. Camón, R. Mart́ınez-Béjar, Accessing touristic knowledge bases through a natural language interface, in: D. Richards, B.-H. Kang (Eds.), Knowledge Acquisition: Approaches, Algorithms and Applications, Vol. 5465 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2009, pp. 147–160. [57] J. Lim, K.-H. Lee, Constructing composite web services from natural language requests, Web Smantics 8 (1) (2010) 1–13. [58] N. Stratica, L. Kosseim, B. C. Desai, Using semantic templates for a natural language interface to the CINDI virtual library, Data & Knowledge Engineering 55 (1) (2005) 4 – 19. [59] M. Gao, J. Liu, N. Zhong, F. Chen, C. Liu, Semantic mapping from natural language questions to OWL queries, Computa- tional Intelligence 27 (2) (2011) 280–314. 20 
work_e7wnts4lbfbvvo6v36fgzrvade	----	NASA Technical Reports Server (NTRS) 
work_eahbqr7l4bf3jeq6c3vrqq7gf4	----	[PDF] An expert system for control chart pattern recognition | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1007/S00170-011-3799-Z Corpus ID: 16009797An expert system for control chart pattern recognition @article{Bag2012AnES, title={An expert system for control chart pattern recognition}, author={Monark Bag and Susanta Kumar Gauri and S. Chakraborty}, journal={The International Journal of Advanced Manufacturing Technology}, year={2012}, volume={62}, pages={291-301} } Monark Bag, Susanta Kumar Gauri, S. Chakraborty Published 2012 Engineering The International Journal of Advanced Manufacturing Technology This paper focuses on the design and development of an expert system for on-line detection of various control chart patterns so as to enable the quality control practitioners to initiate prompt corrective actions for an out-of-control manufacturing process. Using this expert system developed in Visual BASIC 6, all the nine most commonly observed control chart patterns, e.g., normal, stratification, systematic, increasing trend, decreasing trend, upward shift, downward shift, cyclic, and mixture… Expand View on Springer researchgate.net Save to Library Create Alert Cite Launch Research Feed Share This Paper 34 CitationsHighly Influential Citations 1 Background Citations 9 Methods Citations 4 View All Figures and Tables from this paper figure 1 table 1 table 2 figure 2 figure 3 table 3 figure 4 figure 5 figure 6 View All 9 Figures & Tables 34 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency A modelling-oriented scheme for control chart pattern recognition Héctor De La Torre Gutiérrez Computer Science 2017 PDF Save Alert Research Feed METHODOLOGY FOR DESIGNING A CONTROL CHART PATTERN RECOGNIZER IN MONITORING METAL STAMPING OPERATION Norasulaini Binti Abdul Rahman, I. Masood Engineering 2015 1 PDF View 2 excerpts, cites background Save Alert Research Feed Development of fuzzy logic-based statistical process control chart pattern recognition system J. Chen, Yan Liang Engineering 2016 9 Save Alert Research Feed Control chart pattern recognition using an integrated model based on binary-tree support vector machine Cang Wu, Fei Liu, Bo Zhu Engineering 2015 17 Save Alert Research Feed Effectiveness of Different Preprocessing Techniques on Classification of Various Lengths of Control Charts Patterns K. Lavangnananda, Suthruthai Waiwing Computer Science 2015 2 PDF View 1 excerpt, cites background Save Alert Research Feed A hybrid method for estimating the process change point using support vector machine and fuzzy statistical clustering M. S. Kazemi, K. Kazemi, M. Yaghoobi, H. Bazargan Mathematics, Computer Science Appl. Soft Comput. 2016 13 Save Alert Research Feed Pattern recognition for manufacturing process variation using ensembled artificial neural network Teni, Muhammad Hafizzuddin Engineering 2015 PDF Save Alert Research Feed Control Chart Pattern Recognition in Metal-Stamping Process Using Statistical Features-Ann Norasulaini Binti Abdul Rahman, I. Masood, Mohd Nasrull Abdol Rahman, N. F. Nasir Computer Science 2017 1 View 1 excerpt, cites background Save Alert Research Feed Quality control in hard disc drive manufacturing using pattern recognition technique I. Masood, Victor Bee Ee Shyen Engineering 2016 1 Save Alert Research Feed An Effective and Novel Neural Network Ensemble for Shift Pattern Detection in Control Charts M. Barghash Computer Science, Medicine Comput. Intell. Neurosci. 2015 5 PDF View 1 excerpt, cites methods Save Alert Research Feed ... 1 2 3 4 ... References SHOWING 1-10 OF 20 REFERENCES SORT BYRelevance Most Influenced Papers Recency Out-of-control pattern recognition and analysis for quality control charts using LISP-based systems J. A. Swift, J. Mize Engineering 1995 78 Save Alert Research Feed A study on the various features for effective control chart pattern recognition Susanta Kumar Gauri, S. Chakraborty Engineering 2007 46 Save Alert Research Feed Feature-based recognition of control chart patterns Susanta Kumar Gauri, S. Chakraborty Engineering, Computer Science Comput. Ind. Eng. 2006 82 Save Alert Research Feed An Integrated Neural Network and Expert System Tool for Statistical Process Control D. Pham, E. Oztemel Engineering 1995 26 Save Alert Research Feed Recognition of control chart patterns using improved selection of features Susanta Kumar Gauri, S. Chakraborty Computer Science Comput. Ind. Eng. 2009 83 Save Alert Research Feed IntelliSPC: a hybrid intelligent tool for on-line economical statistical process control R. Guh Computer Science 1999 45 Save Alert Research Feed An expert system approach to quality control E. Paladini Computer Science 2000 32 PDF Save Alert Research Feed Feature-based control chart pattern recognition D. Pham, M. A. Wani Engineering 1997 115 Save Alert Research Feed Quality control expert systems: a review of pertinent literature Tsuang Kuo, A. Mital Engineering, Computer Science J. Intell. Manuf. 1993 26 Save Alert Research Feed An expert systems model for implementing statistical process control in the health care industry E. G. Tsacle, N. Aly Engineering 1996 15 View 1 excerpt, references background Save Alert Research Feed ... 1 2 ... Related Papers Abstract Figures and Tables 34 Citations 20 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_eapnqloeu5dg5m3dkrhyhvqwgy	----	Explanation-Based Learning With Analogy for Impasse Resolution Matt Timperleya,∗, Maizura Mokhtarb, Gareth Bellabyc, Joe Howed aUCLAN Energy, University of Central Lancashire, UK bRolls-Royce University Technology Centre, University of Sheffield, UK cSchool of Computing, Engineering and Physical Sciences, University of Central Lancashire, UK dThornton Energy Institute, University of Chester, UK Abstract This paper proposes an algorithm for the inclusion of analogy into Explanation- Based Learning (EBL). Analogy can be used when an impasse is reached to extend the deductive closure of EBL’s domain theory. This enables the gen- eration of control laws, via EBL, for hardware which is not catered for in the domain theory. This advantage addresses a problem which represents a dearth in the current literature. Integrated Modular Avionics (IMA) literature has thus far been concerned with the architectural considerations. This paper seeks to address the impact of hardware changes on the controllers within an IMA architecture. An algorithm is proposed and applied to control an aviation plat- form with an incomplete domain theory. Control rules are generated when no deductive explanations are possible, which still reflect the intent of the domain theory. Keywords: Explanation-Based Learning, Impasse Resolution, Analogy 1. Introduction Modularity in aviation has been gaining interest (Wilson & Preyssler, 2009) because of both the additional future-proofing inherent in having easily replace- able modules and the flexibility of role that this engenders (Committee on Ma- terials, Structures, and Aeronautics for Advanced Uninhabited Air Vehicles,5 Commission on Engineering and Technical Systems, 2000). Modularity makes changing the hardware of platforms easier in order to fit specific missions, as advocated in (Lopez et al., 2008). Changes to a hardware platform may re- quire changes to the control software. The impact of changes in the constituent ∗Corresponding author Email addresses: MTimperley@uclan.ac.uk (Matt Timperley), M.Mokhtar@sheffield.ac.uk (Maizura Mokhtar), GJBellaby@uclan.ac.uk (Gareth Bellaby), J.Howe@chester.ac.uk (Joe Howe) Preprint submitted to Expert Systems with Applications May 26, 2016 hardware of a platform on the software control systems, which operate on said10 platform, forms a gap in the existing literature, as do coping strategies for this scenario. Modular Avionics are aviation electrical systems which are integrated using a modular design paradigm. The components forming a modular aviation plat- form can be easily swapped (Faculty & Kahn, 2001). An Integrated Modular15 Avionics (IMA) architecture involves the interconnection of physically separate hardware components which share computing hardware, whilst remaining logi- cally separate (Lopez et al., 2008). The interest in modularity in aviation hardware, and the current gap in the literature highlights an open problem. It would be advantageous for software to20 adapt to control new items of hardware with a minimum of extra work. This objective can be satisfied in this work by the generation of control rules for an item of hardware which lies outside the original deductive closure of the domain theory. The foundations of generating fuzzy rules from Explanation-Based Learning25 (EBL) explanation structures are in (Timperley, 2015). An assumption em- ployed was that an appropriate, general, domain theory had been elicited. In this paper it is assumed that the domain theory is incomplete for the intended purpose. This new assumption leads to an impasse being reached when attempt- ing to form an explanation. As explanations are used as the basis of new fuzzy30 rules, no control laws will be derived. EBL is focussed on generalisation, usually through the introduction of vari- ables. This facet of EBL is what lends itself to modular control. In this work EBL can be further extended to support the generalisation of predicates, using analogies as supporting evidence.35 The aim of this paper is to incorporate analogical learning into Explanation- Based learning (EBL). Analogy is proposed as a way to increase the deductive closure of the domain theory upon reaching an impasse. This allows a controller to generate rules for hardware outside of its original design, and therefore not mentioned in its domain theory. Such a situation can occur when the hardware40 of the controlled platform changes. 1.1. Analogy A previous work has augmented EBL to reach a deductive solution by com- bining two incomplete domain theories (Hirowatari & Arikawa, 1994). Analogy is used as the basis for linking predicates within two incomplete domain theories.45 The combination of domain theories leads to a deductive solution, in contrast to what is proposed in this work. Augmentation of EBL with analogical capabili- ties is an area which merits further work. This paper builds upon the previous work by allowing analogy to extend an existing domain theory. Analogy can be employed to continue explanation upon reaching an impasse.50 It is likely that a new component will share both similarities to other compo- nents and an energy management strategy. This transfer of knowledge between situations can be achieved by derivational analogy (Carbonell, 1983). Anal- ogy has been used to map different situations to one another, in order to apply 2 prior knowledge to a new situation (Huhns & Acosta, 1988); thereby performing55 transfer learning. This paper aims to use EBL and analogy in order to generate useful control rules for hardware which is outside the design of the controller. The objectives of this paper are to: (i) Introduce an algorithm which augments EBL with analogy to perform transfer learning; and (ii) demonstrate the the augmented60 EBL algorithm can generate rules which are useful but lie outside the deductive closure of the domain theory. The objectives will be demonstrated by applying the augmented EBL algorithm to control a hardware platform where no rules can be generated deductively i.e. without transfer learning. If the generated rules control the platform in a manner which reflects the intent of the domain65 theory then the second objective will be satisfied. This paper is further divided into 5 sections. Section 2 compares the com- bination of EBL and analogy presented in this paper to existing methods. The advantages of combining EBL and analogy are described in Section 3, which also provides the proposed algorithm and satisfies the first objective of this paper.70 In order to satisfy the second objective of this paper the algorithm is applied to pitch control. The experiment described in Section 4 uses a domain theory for throttle control which cannot produce rules for pitch deductively. Analogy is used to generate rules for pitch. The result of applying these rules is given in Section 5. The paper concludes with Section 6.75 2. Related Work Transfer learning can be applied as an alternative to an impasse, such as in (Burstein, 1986). In other applications of analogical reasoning (e.g. (Carbonell, 1983), (Klenk & Forbus, 2009)) a known solution to a similar problem is found and mutated to be applicable to the current problem.80 Case Based Reasoning (CBR) was the method of analogical reasoning in (Carbonell, 1983) and (Klenk & Forbus, 2009). CBR uses similarities between previous solutions to various problems to solve the one at hand, the target (de Mántaras & Plaza, 1997). Closely similar solutions are modified to solve the target problem. Similarity metrics are used to guide searches for similar85 solutions. In this work a similarity between terms is used as the basis for analogy. If comparing the technique proposed with that of derivational analogy (Car- bonell, 1983), the analogue is a concept rather than a situation. Another differ- ence is, the presented technique, rather than tailoring a previous solution to a90 new situation, generalises a previous line of reasoning within a solution, to more concepts. Applying analogy to only a single line within an example is similar to when students refer to a specific line of a previous example to justify some reasoning rather than the example as a whole (VanLehn & Jones, 1998). This technique also shares some conceptual similarities with (Ishikawa & Terano,95 1996). Both this work and that of Könic et al (Könik et al., 2009) map between concepts in order to apply knowledge from one situation to another. There are 3 some differences between this work and (Könik et al., 2009). This technique is proposed to derive links between previously unrelated concepts. These links100 are embodied in a new, more abstract, concept definition. This technique has a different method of application, as an alternative to failure. Both works use a goal and explanation, as a bias, to constrain the possible analogies. This technique also maps between potentially overlapping concepts rather than like terms.105 This type of analogy, from information in (Falkenhainer, 1987), could be stated as using similarity-based generalisation in order to perform a form of analogical reasoning. The analogical reasoning is restricted to generalising in- formation to hold to more concepts than when derived. The algorithm forms a new concept from two existing definitions, keeping the parts specific to both and110 introducing variables where variation exists. In this regard the algorithm is sim- ilar to EGGS (Mooney et al., 1986) which performs generalisation by applying the aggregation of substitutions within an explanation to the whole explanation. In this work an analogy is seen as evidence for predicate generalisation or swapping. The generalisation can either replace the target predicate or be taken115 as evidence for swapping the source for the target. Analogies could be generated using different techniques in order to generate evidence for predicate swapping. One such method which is capable of deriving analogies from natural language is (Cambria et al., 2015). It may be possible to use such techniques and either apply them to domain theories. However, if domain theories are lacking in data120 richness then the system in (Cambria et al., 2015) could be applied to the expert knowledge from which the domain theory was designed. The algorithm proposed differs from (Hirowatari & Arikawa, 1994) by in- creasing slightly the deductive closure of the program in order to apply analogi- cal reasoning in more restricted form. This increase in deductive closure is based125 on the assumption that the addition of a new concept increases the number of facts that can be deduced within the system. An application of analogical learning with EBL (Hirowatari & Arikawa, 1994) was able to construct explanations from incomplete domain theories where the combination of domains made a deductive explanation possible. This work pro-130 poses that less exact analogues can be used to derive relations between terms with different predicates. Rather than establishing a link between the same terms in different domains, the emphasis is on relating different terms within the same domain. In this paper the algorithm proposed goes further and ex- tends a single incomplete domain theory using internal similarities between two135 predicates. This is achieved by deriving an abstraction which encompasses both predicates. 3. Abstraction Derivation This paper proposes the derivation of an analogical link between two de- ductively separate concepts. This link is formed by the generation of a new140 concept, which requires commonalities between the two target concepts to be 4 satisfied. The new concept is an abstraction of the two target concepts. Mem- bership of this abstraction implies that concepts are analogous in the manner given by the abstract concept definition. This could be thought of as weakening the preconditions on the target concepts to form an abstraction.145 An advantage of EBL can be exploited during the definition of an analogi- cal similarity; training examples encountered can provide an inductive bias for disregarding terms that are not relevant. During explanation an impasse may be reached when no deductive option remains, but an inductive leap is possible. Inductive leaps are realised by re-150 placing one concept in a rule with another. This gives rise to a new rule which relates previously unrelated concepts, linked by analogy. These analogous con- cepts are not deductively entailed. The difference in predicates prohibits their mutual unification to explain the same goal. Additions to the domain theory which embody the similarities between the two concepts are used as a basis for155 replacing one predicate with another. Rather than generalising a rule by abstracting specific variable bindings from terms, as in EBL (Ellman, 1989), two concept definitions are linked by a more abstract class so that the terms themselves can be replaced. This is similar to assuming that the predicates themselves could be bound differently. The160 abstract class can be formed from commonalities between otherwise disparate concept definitions. This class embodies an analogy and is expressed as a new rule. The algorithm will now be presented. 3.1. Forming an Abstract Definition The two concepts which are used to form an abstraction are the source and165 the target. The source is the definition of the predicate which causes an impasse. The target is a concept searched for, using the parameters of the source as a search bias. Mapping establishes the similarities between two identified concepts, taken from the model of analogy presented in (Gentner, 1983). An approach to achieve170 this within EBL is proposed. The approach for forming a new, more abstract, definition from two target concept definitions is: 1. Collect like terms in the leaf nodes of both concepts and reduce the number of occurrences to one by introducing variables and forming a substitution. Do this for both, or all, concepts being considered.175 2. Discard any terms that do not appear in all concept definitions. When considering many concept definitions, one may discard from consideration concepts that lack common terms rather than the terms themselves. As long as a relationship between at least two concepts is established, this is permissible.180 3. Use these common terms to form the precedents for a new concept def- inition. This definition captures some similarity between the concepts considered which is assumed to be absent from the initial domain theory. This approach has been formalised in Algorithms 1 and 2. Algorithm 1 se- lects the analogy with the most terms in it, corresponding to matches between185 5 any combination of definitions for both source and target terms. This represen- tation calculates all the possible abstractions between each definition of both concepts, sorted in descending order of complexity. In practice, the most com- plex abstraction is taken and assumed to show a closer connection between two concepts. Algorithm 2 forms the abstraction between two concepts.190 Algorithm 1 Impasse(term source) Require: source is a ground atom targets ← rules containing predicate(s) with arguments matching the source. SourceDefs ← rules previously derived that imply the source, stored within the rule-base. // Collect analogies between each definition of the source and target. These analogies are stored in "analogies" which is sorted by number of terms (de- scending) in the abstraction. for all target in targets do targetDefs ← rules previously derived that imply the source, stored within the rule-base. for all targetDef in targetDefs do for all sourceDef in sourceDefs do Abstract(source, target) store target in analogies along with its abstraction (targetDef). end for end for end for add the target, topmost from analogies, under source in the explanation struc- ture. Terms are discarded which are not common to both source and target def- initions, as these are assumed to be irrelevant to whatever similarity is being captured. This is to facilitate the capture of analogues that have the most in common. More general analogues that make a weaker claim of similarity between concepts could also be useful but could more readily lead to over-195 generalisation. Also, it may be computationally wasteful to derive very general analogues since these could potentially be applied often to little effect. Over-generalisation and over-specialisation have been the topic of previous works such as the EGGS algorithm (Mooney et al., 1986). The generalisation used to form the abstraction is similar to the EGGS algorithm. However, there200 are some differences: it is applied across two explanations, like predicates are aggregated within a single explanation (source or target definitions), and the generalised terms are added to a new explanation rather as replacements to an existing explanation. This algorithm will be illustrated using an example. 3.2. An Example205 The analogical EBL algorithm can be applied to an energy management example. Consider a battery powered system being altered to be powered by a 6 Algorithm 2 Abstract(term source, term target) Require: source and target predicates are ordered. get the first terms from source and target concepts. for all predicates in source do if current predicates from source and target concepts differ then get next source predicate. else while the next predicate in the source matches the current one do note the arguments, by arity, from source term. get the next term from source. end while while the next predicate in the target matches the current one do note the arguments, by arity, from target. get the next term from target. end while for all arguments from source for this predicate. do if there is a corresponding argument in target then add the current predicate, with the matching argument, to the ab- straction. else // this should be limited to once per predicate add the current predicate, with an introduced variable as the argu- ment, to the abstraction. end if end for end if end for return the abstraction. 7 fuel cell. The domain theory and goal are given in Table 1. Table 1: A sample input for EBL Goal can_supply(fuel_cell_1, ASI_demand) Training Data max_output(fuel_cell_1, 50) demand(ASI, ASI_demand) Domain Theory limited(X, charge) + waste(X, heat) + requires(X, charge) + produces(X, en- ergy) → battery(X) fuel(X, hydrogen) + waste(X, heat) + waste(X, water) + requires(X, hydrogen) + produces(X, energy) → fuel_cell(X) battery(X) + max_output(X,O) > demand(C,Y) → can_supply(X,C) fuel_cell(fuel_cell_1) battery(battery_1) The EBL algorithm reaches an impasse when attempting to explain the goal: battery(fuel_cell_one). This is shown in Figure 1. The boxed area in Figure210 1 is the point an impasse is reached. Single lines denote implications. The double line is where the algorithm generates an abstraction to link the source and target concepts. Figure 1: Example explanation structure starting with battery oriented rules and using ab- straction to substitute battery for source. Single lines depict implications and can be read that the lower line implies the statement above. A box shows where the proposed technique is being applied with the double line representing an abstract similarity. The algorithm uses the unbound goal which caused an impasse as the source concept. In this case: battery(X). Target concepts are identified as statements215 taking the form X(fuel_cell_1): In this case, fuel_cell(fuel_cell_1). The rules which imply source and target concepts are taken as their defini- tions. In this case the source definition is: limited(X, charge) + waste(X, heat) + requires(X, charge) + produces(X, energy) → battery(X). The target defini- tion is: fuel(X, hydrogen) + waste(X, heat) + waste(X, water) + requires(X,220 8 hydrogen) + produces(X, energy) → fuel_cell(X). Terms which appear in both definitions form part of the abstract definition, such as produces(X, energy). Predicates which appear in both definitions but with different arguments require the introduction of a variable, such as waste(X, W). Introducing a variable captures a weaker similarity between the two con-225 cepts and is akin to weakening the pre-conditions, compared to either target concept. The combination of these common factors yields the definition of the source abstraction: requires(X, F) ∧ waste(X, W) ∧ produces(X, energy) → source(X). The derived abstraction sits in support of using the target predicate in place230 of the source. Swapping the predicate may result in the generation of rules when reaching an impasse. In this example the rule for the battery can be expanded to apply to a fuel cell. This is shown in Figure 1. Two approaches to applying the analogy are: Firstly, the derived concept can replace battery in the explanation and a more general rule can be derived:235 source(X) + max_output(X,O) > demand(C,Y) → can_supply(X,C). Secondly, the source predicate could be replaced with fuel_cell and the analogy sits in support of this decision. The second approach is adopted in this paper. An alternative to implicitly linking two terms and replacing one with the other is possible. Both terms could be replaced with the abstraction, which240 would be a more explicit analogical link. However, this could lead to over- generalisation. By storing a new rule which includes the abstraction, more than the two concepts used to derive it could now be applicable. It could be that analogies only hold in certain cases; the impact of this can be lessened by using each abstraction in only the situation which caused its derivation. Further work245 could be conducted in studying the over-generalisation issue with regards to this technique. Further nuances of this algorithm merit discussion. 3.3. Discussion One disadvantage of applying an analogy is the utility problem. The Util- ity problem is defined as "The difficulty in ensuring that an acquired concept250 enhances the performance of the application system" (DeJong, 2004) and is discussed more in (Steven, 1990). Approaches to mitigating this should be em- ployed, such as including a priori knowledge relating to utility in the domain theory. Another approach is the empirical testing and removal of analogies that increase the branching factor significantly without deriving useful rules.255 Applying an analogy will have a cost, not just in matching but by increasing the branching factor of the problem. This is due to augmenting the domain theory; more possibilities result in longer searches. It may be that no useful in- formation is gained by permuting a rule to apply to a new situation via analogy. There is no deductive evidence that the similarities between two concepts mean260 that a given rule applies to both. As there is a cost and potentially no gain, since the leaps are inductive, this technique should not be overused. There may be reason to think that similarity with nodes higher in an explanation structure are better for judging applicability. This is described in (Carbonell, 9 1983). There may also be reasons to prefer more specific rules (Huhns & Acosta,265 1988). However, when no deductive option is available this technique may give an alternative to backtracking or failure. Choosing to derive an analogy could be likened to strategically weakening preconditions, given that an analogy is a subset of atoms of a concept. It may be possible, at any point, to derive many potential analogies between270 concepts. However, since not all of these will be useful a conservative approach is to only adopt an analogy as a last resort to failure. If an analogy is adopted during an explanation to prevent failure, then the validity of the analogy could be assessed by whether the new rule can be successfully applied. The criteria for judging the utility of the rules may well be domain independent.275 There may be more than one possible analogy which can be derived. Biases could be used to narrow down the list of potential definitions, ideally to one. One bias could be: A definition which makes more claims than another is stronger. Other biases could be derived by considering lexical similarity or using meta- data labelling of concepts. The bias employed in this paper is to simply take the280 analogy that makes the claim with the largest number of similarities. Analogues are computed for the potential matches and the strongest is chosen. The source definition is given by backward-chaining from the goal which causes an impasse. The target is inferred to exist and must be searched for. This search is potentially expensive though only a subset of the domain theory285 could fit the pattern. Another use for the derived analogy is when forming new rules, the analogy can be considered first and a general rule formed initially. The rule could also be annotated with any concepts considered during application for which it does not hold, like a caveat. This would follow from the assumption that these analogies290 are a simple representation of a potentially complex relationship and can be both correct and incorrect based on factors not modelled in the system. Any analogy proposed by this method may be flawed since it is an attempt to infer some commonality and relationship that is not explicitly reasoned. The more this is applied, the greater the chance of concluding a fallacy. Therefore295 analogies should initially be proposed and should never be assumed to be cor- rect. The argument is similar to that in (Dejong, 2006); these analogies capture information which can be both correct and incorrect depending on factors that the system is unaware of. To discard them too readily risks missing important relational information but they themselves may only capture an aspect of the300 real relationship. If an analogy never leads to useful rules that hold when en- countered in reality it should be discarded. Since it may be impossible to know this, a probability threshold, or similar technique, could indicate which rules to discard. Alternatively, the comparison of the impact of a given analogy on the goals of the system could be used.305 There is an issue with assuming an abstract similarity exists. Not all ana- logical inferences are equally likely (Davies, 1987) and where the similarity is based on features, not all will be relevant. This is partially ameliorated by EBL disregarding features not relevant to the example, using an inductive bias. When trying to derive an abstract similarity between two concepts, even using310 10 an example as an inductive bias, it is likely that not all features in the example are relevant to the analogue being derived. Using an example as an inductive bias, along with empirical validation, could help to disregard spurious analogies. Commonality in language is important to this method as it requires either shared terms or analogous terms. However, commonality in language may not be315 an unreasonable assumption. This is because the knowledge was originally rep- resented using an altered subset of a shared language between experts. Therefore it may be that lexical similarity can be used heuristically to establish or evaluate an analogue. Note that multiple applications of analogy could build a common set of terms320 from other concepts, whilst discarding language that isn’t commonly employed. The common terms could be the result of previous analogical reasoning. The effect of this technique would be most obvious where the domain theory uses disparate terms. In this type of domain, these analogy concepts can be used to transform nodes into a more common language.325 There is a danger of applying this technique too readily since over-application would be costly and may bring about no real benefit. It is important to select when to use analogy since each application represents an inductive leap and weakens the otherwise deductive proofs EBL produces. It is important to bear these issues in mind during knowledge engineering.330 The algorithm enables the expansion of a rule for the battery to consider a fuel cell, for example. However, whilst this looks like a sensible application of a rule in an abstract way it is likely that over-generalisation can occur. Alterna- tively, rules could be generated which would ideally be prohibited with a more complete domain theory. It is therefore important to only propose an analogy335 when the result is a useful rule. The utility of an analogy is not apparent at the time of derivation therefore the individual rules may need to be evaluated post generation. The algorithm is applied in simulation in order to explore these issues further. 4. Experiment Design340 An experiment is proposed in order to show that analogy can allow EBL to generate fuzzy rules for a piece of hardware that is not catered for by the domain theory. The generation of fuzzy rules which reflect the intent of the system would satisfy this objective i.e. rules which reduce energy consumption by controlling previously unseen hardware.345 The domain theory determines what rules can be generated. In this exper- iment it is designed for throttle control. However, the goal has had variables bound such that an impasse will be reached when attempting to generate rules for throttle control. This is because the variable binding height_to_go identifies a piece of hardware relating to pitch, rather than the throttle control. Analog-350 ical reasoning is required in order to generate control laws for pitch. Rules derived by this algorithm extend the deductive closure of the domain theory. The generation of any rules which reduce energy consumption fulfil the intent of the original domain theory. The rules generated in this chapter control the 11 pitch of the aircraft, which is not possible using the given domain theory without355 analogy. Therefore rules which reduce energy consumption generated in this experiment fulfil the second objective of this paper. The system being tested is the same as in (Timperley, 2015). EBL was used to generate fuzzy rules which were executed in order to control the throttle of an aviation platform. The generation of useful rules, specialised from a body360 of general expert knowledge, was demonstrated by said controller requiring less energy to complete a set flight. This paper uses the same experimental setup as the previous work. The approach presented in this paper has been implemented and integrated with the X-Plane simulator. X-Plane has previously been used in high-fidelity365 aerospace research, such as (Cameron et al., 2011), as part of a broader simula- tion suite. Pitch changes are executed by the FLCHG system in X-Plane. This means that the pitch controls themselves are not used. It is the throttle which controls climb and descent. The throttle setting is calculated based on the height to the370 next checkpoint being positive or negative. Controlling the pitch via the joystick results in a single large alteration which is then compensated for, when possible, by the autopilot. Because of this, it is the height to the next checkpoint which is controlled in this paper. Alteration of the required climb indirectly affects the throttle and therefore the pitch.375 An experiment was conducted in order to evaluate the approach presented in this paper. A series of replicable flights were devised to test the controller. Each flight begins with a take-off, followed by a series of checkpoints to be hit. The run ends when the final checkpoint is reached. The checkpoints are placed such that climbing, descending and turning all feature. The autopilot flies a set380 route, input into a Flight Management System (FMS). There are two sources of variation between the flights within a data set. The first is the throttle control value, which is altered by the proposed algorithm. This is allowed to vary over its entire value range [0-1]. The second source of variation is the simulator. The time of day is not bound which may affect385 temperature and therefore air pressure. This can alter the amount of energy perform certain activities, such as take off. The weather was set to be ‘clear’ which limits the variability of the weather significantly which in turn limits the effect on the simulated flights. Weather includes the winds encountered and other sources of pressure variation. Limiting these effects limits the variation in390 forces encountered during. This maintains a level of consistency in the energy requirements of each flight. The remaining variables controlled by the simulator are bounded by keeping the route fixed. The domain theory persistent between flights. A data set uses a single domain theory which is updated in each flight and passed to the next.395 The domain theory only contains rules to determine state, reason about relative orders of magnitude, and reduce the throttle of the aircraft. There are no rules to control pitch. The domain theory is constructed in a way which could support analogical reasoning. The characteristics required of such a domain theory are: additional contextual information which can act as definitions for400 12 source and target concepts. There should also be sufficient generalisation that source predicates can be replaced. For example, consider a system input being identified by the statement veloc- ity(input_value). Searches for X(input_value) would identify predicates with a particular input value. However, if the same system input is identified by405 reporting(velocity, input_value) then target concepts can be identified by the name of the input, rather than the current value. The inputs and the output are abstracted such that predicate replacement can occur. Additionally, the output is no longer hard coded into a parameter. This could aid the flexibility of the domain theory in a deductive sense. By410 abstracting the inputs and output, the relationship is made to apply to different sources. Changes were made to the domain theory to support the formation of specific analogies. Rules which have shared terms have been added to the domain theory to support the generation of analogies. These rules embody the additional con-415 textual information about concepts. This additional information, unnecessary to the intended model, may be easier than eliciting a minimal domain theory. It is worth noting that extra information may not always need to be added in order to form analogies, but can be used to restrict the analogies which are possible. This can be a helpful tool in reducing unwanted rules from being derived.420 The domain theory is given in Table 2. The table also includes the goal used to derive rules. Goal reduce(height_to_go,Y) Domain Theory: Rules numbers bigger(X,Y) → smaller(Y,X) !same(X,Y) → different(X,Y) !bigger(Y,X) → biggest(X) !smaller(Y,X) → smallest(X) !same(X,Y) ∧ bigger(X,Z) ∧ bigger(Z,Y) → bigger(X,Y) reporting(I,X) ∧ bigger(X,Y) → more_than(I,Y) reporting(I,X) ∧ smaller(X,Y) → less_than(I,Y) reporting(I,X) ∧ same(X,none) → none(I) more_than(I,none) → some(I) reporting(I,X) ∧ same(X,Y) → is(I,Y) reporting(I,X) ∧ started(I,Y) ∧ same(X,Y) → unchanged(I) reporting(I,X) ∧ started(I,Y) ∧ !same(X,Y) → changed(I) reporting(I,X) ∧ smaller(Y,X) → one_smaller(I,Y) reporting(I,X) ∧ bigger(Y,X) → one_bigger(I,Y) reduction strategy reporting(I,X) ∧ intend(I,N) ∧ bigger(X,N) → reduce(I,N) 13 throttle_command(I) ∧ height_to_go(H,I) ∧ velocity_sensor(C,I) ∧ report- ing(H,P) ∧ reporting(C,E) ∧ same(P,E) ∧ one_smaller(I,N) → intend(I,N) throttle_command(I) ∧ height_to_go(H,I) ∧ velocity_sensor(C,I) ∧ report- ing(H,P) ∧ reporting(C,E) ∧ smaller(P,E) ∧ one_smaller(I,N) → intend(I,N) state some(parking_brake) ∧ less_than(throttle,medium) ∧ none(altitude) → state(idle) none(parking_brake) ∧ less_than(throttle,medium) ∧ none(altitude) → state(taxi) none(parking_brake) ∧ more_than(throttle,small) ∧ none(altitude) → state(takeoff) current(takeoff) ∧ none(parking_brake) ∧ some(altitude) → state(flight) analogy controls(engine_speed) ∧ affects(air_speed) ∧ affects(acceleration) ∧ pro- vides(thrust) ∧ reflects(power) → throttle_command(throttle) controls(elevators) ∧ affects(altitude) ∧ affects(air_speed) ∧ pitches(aircraft) → pitch_command(height_to_go) reflects(motion) ∧ affected_by(throttle) → velocity_sensor(velocity, throttle) reflects(climb) ∧ affected_by(pitch) ∧ affected_by(throttle) → climb_sensor(climb_rate, height_to_go) change_in(y) ∧ affected_by(throttle) ∧ affected_by(pitch) → height_to_go(height_to_go, throttle) change_in(x) ∧ affected_by(throttle) ∧ affected_by(pitch) → ground_speed(groundspeed, height_to_go) Statements numbers bigger(full,huge) bigger(huge,large) bigger(large,medium) bigger(medium,small) bigger(small,tiny) bigger(tiny,none) same(X,X) analogy throttle_command(throttle) pitch_command(height_to_go) height_to_go(height_to_go, throttle) velocity_sensor(velocity, throttle) Table 2: Analogical Domain Theory The parameter values in the domain theory are used to identify simulator inputs. During explanation these cause variables to be bound to the current425 reading for the identified input. Examples of these parameter values include: 14 throttle, velocity_sensor, height_to_go, parking_break, altitude. Some bound values in the analogy section refer to abstract notions that do not refer to system inputs; pitch is one of these. These represent disconnected background infor- mation derived from the expert which add context but aren’t directly related430 to rule generation without analogy. They can be altered to altered to influence the abstractions which can be generated. The use of a single goal for this experiment also serves to limit the possi- ble analogies. This leaves the domain theory with an inherently hierarchical structure.435 It is worth noting that, in the current implementation, analogies are not added to the domain theory as new rules. The (sub)goal causing an impasse is swapped for the target with the larger correspondence i.e. the one which formed the most complex derived class. Additionally, the analogous nodes are not included in the fuzzy rule derivation process. These nodes are omitted because440 the nature of the link between the source and target terms is assumed to be irrelevant to the fuzzy rule. Instead the analogy is viewed more as evidence, or an operational node in support of the swapped node. This allows the explanation to continue without adding goals which would not have been present in a non- analogical explanation structure, minus the term that has been replaced.445 When attempting to use EBL to derive a new fuzzy rule an impasse will be reached, because there is no deductive explanation with the given domain theory for pitch control. These impasses can be avoided by forming analogies. The analogies in this case replace the inputs being compared and the output, but maintain the same overall structure as the throttle derivation from the450 previous work. The output is swapped from the throttle to the height to the next checkpoint. The inputs were also altered. The velocity sensor is swapped for the height to go to the next checkpoint. The height to go is swapped for ground speed. Swapping the inputs follows a simple rationale. When the ground speed is455 smaller than the height to go to the next checkpoint, reduce the climb rate. The climb rate could be calculated as the gradient: δy δx . A larger gradient, and steeper climb, would come about when the change in x (δy) is larger than the change in y (δx). Therefore reduction occurs when ground speed is smaller than the height to go to the next checkpoint.460 Analogy is required for the generation of all fuzzy rules during this exper- iment. No rules can be generated by the system initially, analogy is used to swap the output and inputs during rule generation. The domain theory re- mains largely the same as in the previous work. Some small additions are made to facilitate the formation of certain analogies. By swapping the output and465 inputs a general strategy is transferred to control an item of hardware which lies outside the original domain theory. This can be seen as further relaxing the restrictions that were present in the rule generation method proposed in the previous work. This has been facilitated by swapping the inputs and outputs referenced in a rule derivation.470 Inputs are fuzzified in the manner proposed in the previous work. Each measurement is fuzzified and interpolated and is therefore relative to itself. 15 This is because the range over which the linguistic values are interpolated is determined by past values. The rules generated are dependent on the number of fuzzy sets and refresh rate of inputs. Both of these factors influence the states475 which are considered by EBL, as well as the granularity of each rule. The analogies formed, by the proposed algorithm, in order to swap the inputs and output are given below. The derived concept is labelled as the target concept. • affected_by(throttle_ratio_all)∧ reflects(R)→ climb_sensor(climb_rate,480 height_to_go) • affected_by(pitch)∧change_in(C)→ground_speed(groundspeed, height_to_go) • affects(A) ∧ controls(C)→ pitch_command(height_to_go) Note that the output is taken as an input. This output is also written to through the switching mechanism. Having one input being written to by two485 systems means that timing can affect the rules learned, since the input will appear to have one value or another, depending on when it is sampled. This is one motivation for using fairly large data sets and averaging. It is worth noting that in this experiment the autopilot is always operated at 10Hz and the controller is initially operated at the same rate.490 In this experiment the outputs of forming an analogy are any fuzzy rules generated by employing said analogy. As no rules can be formed deductively, all rules will be formed in this experiment by analogy. Any rules generated can, depending on their impact on the flight, sit in support of the technique. The aim of the experiment is to show the possibility that the technique can be495 used to generate useful rules after knowledge engineering has taken place. The results of the experiment are presented in the next section. 5. Results and Discussion A set of 50 flights was conducted with a set series of waypoints. The goal used in this experiment leads to the generation of rules which control pitch.500 However, no rules can be generated for pitch control, using the domain theory designed for throttle control, without resolving impasses. In order to satisfy the first objective of this paper rules must be generated after reaching impasses. The second objective can be satisfied by demonstrating the production of useful rules.505 There are two sources of variation between the flights within a data set. The first is the throttle control value, which is altered indirectly by analogically derived rules via the height_to_go to the next checkpoint. The height_to_go is used by the autopilot to determine throttle but itself relates to pitch. The second source of variation is the simulator. The time of day is not510 bound which may affect temperature and therefore air pressure. This can alter the amount of energy perform certain activities, such as take off. The weather was set to be ‘clear’ which limits the variability of the weather significantly which 16 Figure 2: An average, over 50 flights, of the altitude of the aircraft. This is without Analogy. The shaded areas indicate the standard deviation around the mean. The inputs were sampled at 10Hz. in turn limits the effect on the simulated flights. Weather includes the winds encountered and other sources of pressure variation. Limiting these effects limits515 the variation in forces encountered during. This maintains a level of consistency in the energy requirements of each flight. It may be possible to further limit the weather and time variability using X-Plane. The remaining variables controlled by the simulator are bounded by keeping the route fixed. The domain theory persistent between flights. A data set uses a single domain theory which is520 updated in each flight and passed to the next. There are two reasons to consider the average altitude of the flights within a dataset. Firstly, the consistency of the dataset can be shown by considering the variation in the dataset without analogy, where no rules can be generated. Secondly, by comparing the average altitudes of each dataset, with and without525 analogy, significant impacts that the generated rules have on flight can be seen. This establishes that the generated rules impact the system. After which, it remains to demonstrate that the rules positively impact energy conservation. The mean altitude for all 50 flights, along with the standard deviation, are plotted in Figure 2. The small standard deviation implies that the route followed530 was consistent, with small differences, across all 50 flights. The standard deviation shows the areas most affected by the derived rules. Large standard deviations can be caused by differences in aircraft behaviour at the same point on the time axis. As can be seen, the effect is concentrated at the start of a climb and in the later section of flight (from about 5.5 minutes to535 7 minutes). The controller’s effect being concentrated at the start of a steep climb may be a result of the inputs being measured relative to themselves, rather than a scale that applies to both. The rate of increase in either input may be partially 17 Figure 3: The number of rules generated by the end of each flight is given. Rules generated in one run are not guaranteed to be executed on the same run. This graph is for the data set collected using analogy and the inputs were sampled at 10Hz. hidden. When both inputs are increasing then a small change, relative to its own540 scale, might be significant for an input. However, it is possible such an increase is not significant to the relationship being appealed to in the domain theory. Additionally, the baseline throttle control is rather binary. A binary throttle controller leads to a high ground speed when climbing, even when the climb is steep. Investigation of an interpolated fuzzy relationship for the comparison of545 two inputs could prove an interesting area for future work. The initial section of the steep climb is reduced drastically, then returns to a steep climb. This is less granular control than was exhibited when applying the domain theory to throttle control (Timperley, 2015). The loss of granular control is related to the throttle output only being effected in a binary manner,550 based on the finer control of another output (height to go). This means that a reduction in pitch may lack the finer control required to approach optimality. The effect of the rules, from these two examples, appears to be a general reduction in extreme levels of throttle output. However, more fluctuations are present. Fluctuations could cause an increase in the distance travelled without555 saving sufficient energy to have a net positive effect. Also, some of the larger plateaus have fluctuations which are small. These fluctuations come from the effect of the fuzzy rules competing with the auto-throttle. This shows that the effect of the fuzzy controller may be truncated by competition with the auto- throttle. The degree of competition with the auto-throttle is affected by the560 switch timings. The number of rules generated, by run, is shown in Figure 3. A full listing of the rules generated is given in Table 3. Each of these rules required analogical reasoning during derivation. They satisfy the first objective of the paper. 18 Run Input Output ground speed climb rate Height to go Height to go 1 large large small tiny 1 medium medium medium small 1 huge huge medium small 1 huge huge huge large 1 small small medium small 1 medium medium tiny none 1 medium medium small tiny 1 small small tiny none 1 huge huge large medium 1 huge huge tiny none 1 small small huge large 1 huge huge small tiny 1 large large medium small 1 large large large medium 1 none small medium small 1 small small small tiny 2 medium medium huge large 2 large large tiny none 2 medium medium large medium 3 large large huge large 11 medium medium medium none 12 small small large medium 13 medium medium small none 17 medium medium huge small 37 large large medium tiny Table 3: Table of the fuzzy rules generated, by run, with analogy. The controller was operated at 10Hz. 565 The sharp increase of rules in the final run may indicate that more rules may have been generated given a larger number of runs. The fuzzy rules, gener- ated by analogical EBL, affected the remaining battery charge. Battery charge measures the mainstay of the energy available to the platform. A more efficient flight results in a higher remaining charge at the end of a run. If the controller570 generates useful rules then a more efficient flight will be indicated by remaining charge. The generation of useful rules via analogy satisfies the second objective of the paper. The remaining battery charge is plotted by run in Figure 4. Figure 4 shows that some rules had a positive effect on efficiency but that the gains are situational. The rules are generated by analogy as a first attempt575 to extend an incomplete domain theory. Each rule represents a shallow under- standing of a deeper concept. The rules should ideally use empirical data about 19 Figure 4: The battery charge remaining at the end of each run for flights both with (red) and without Analogy (blue). Inputs sampled at 10Hz. individual rule impact, as well as specific combinations of rules in different or- ders. This data could be used to further specialise the rule to better reflect a deeper understanding of the concept it represents.580 Rules impact is situational and examples of a negative impact can be seen in Figure 4. This is likely due to the large reduction in altitude between about 5.5 and 7 minutes. This results in a longer path taken to reach the final checkpoint. A longer route in this case costs more energy than is saved by reducing altitude. The fuzzy rules are all part of a reduction strategy, which sits in opposition585 to the auto-throttle ascending to a new checkpoint. The switch settings alter the balance between these two factors. The effect of the switch timings is further examined in a series of flights using different sampling rates for the controller. Sets of 50 flights were again conducted. In each case the switch setting for the fuzzy controller was altered. The remaining charge, both with (red) and590 without Analogy (blue) is given, as well as the number of rules generated per flight. The aim is to elicit any general trends between the switch settings, the efficacy and the number of rules generated. The following switching rates are presented: 6.6Hz (every 0.15 seconds), 5Hz (every 0.20 seconds), and 4Hz (every 0.25 seconds). The figures for each result set are given below:595 • The number of rules generated for a 6.6Hz controller is given in Figure 5. The remaining battery charge is shown in Figure 6. • The number of rules generated for a 5Hz controller is given in Figure 7. The remaining battery charge is shown in Figure 8. • The number of rules generated for a 4Hz controller is given in Figure 9.600 The remaining battery charge is shown in Figure 10. 20 Figure 5: The number of rules generated by the end of each flight. Analogical controller enabled, inputs sampled at 6.6Hz. Figure 6: The battery charge remaining at the end of each run. Inputs sampled at 6.6Hz. 21 Figure 7: The number of rules generated by the end of each flight. Analogical controller enabled, inputs sampled at 5Hz. Figure 8: The battery charge remaining at the end of each run. Inputs sampled at 5Hz. 22 Figure 9: The number of rules generated by the end of each flight. Analogical controller enabled, inputs sampled at 4Hz. Figure 10: The battery charge remaining at the end of each run. Inputs sampled at 4Hz. 23 The ’height to go’ input, which is also the output, is written to by both the auto throttle and fuzzy controller. Rule generation is guided by the input value. This balance may simply have exposed the EBL algorithm to more situations than other sampling rates.605 More rules were generated when the controller operated at the highest fre- quencies. Rule generation was most prolific at 6.6Hz. Operating the controller at a higher rate gives it a higher resolution. Rules can be executed which reduce pitch, which prevents the aircraft reaching a state which necessitates the gen- eration of more rules. With a large pitch several rules need to be generated in610 order to result in a reduction in pitch. This is because a change is only affected by the autopilot once the output becomes negative, so several rules are required to bring a large pitch to a negative value. By keeping the pitch at smaller values, fewer rules are required. This may point to higher rates of controller execution being beneficial.615 Some rules may come from seeing an output from a previous controller exe- cution, rather than the value from the autopilot. The input may be insufficiently reduced to cause a change in pitch and the algorithm may produce further rules as a consequence. Not all rules generated are guaranteed to affect the throttle on their own, since the auto throttle is specified as being concerned with the620 sign of the value. Therefore several rules may need to be generated in turn, and executed in concert, to result in an effect. This could complicate a study of the impact of individual rules in this case. The remaining charge was generally improved by the 4Hz controller, but a significant loss in efficiency was evident in later runs. This may imply that625 rules generated earlier in this case were of a greater utility. This implies that even when useful rules are generated from an analogy, invalid rules may also be generated. An algorithm for adding caveats to an analogy in order to prevent detrimental rules from being learned could be a fruitful aspect of further work. This would likely require clearer data on the individual impact of each rule,630 both alone and in conjunction with others. Both cases need to be considered, as multiple rules may combine to produce a positive effect where only one would not. 6. Conclusion An approach for incorporating analogy into EBL has been presented. The635 technique could be used to further generalise rules in a system by linking pre- viously separate concepts. These links are formed by deriving an abstraction. Doing this may imply that whatever reasoning held in deriving the original rule analogously holds for the new hardware. Analogy is an example of transfer learning.640 Transfer learning can be applied as an alternative to an impasse. This tech- nique can operate where an impasse is reached due to a mismatch between outer predicates with the same form. The reason for addressing this type of impasse is that EBL, by generalising, already avoids some potential impasses due to differing parameters.645 24 By applying this technique EBL can make inductive leaps when deductive explanation fails. However, this entails all of the drawbacks of making inductive leaps. Given that the potential costs are weighed against the possibility of no gains, the technique should be applied conservatively. However, flights which performed significantly below the baseline could imply650 that the technique requires further work. Each rule should ideally be evaluated for its individual impact, but where the derivations are only possible by analogy it is likely that there may be no information for evaluation. Deviation from known mission parameters or principles could be possible. Further work into eliciting a general evaluation strategy for analogies would be beneficial. It would655 be particularly beneficial to glean more from the underlying correlation than just a connecting analogy. Each analogy holds only in certain circumstances. For this reason, formation of analogy may be considered the start of an ongoing learning task. Some rules were generated which satisfy the aim of this paper, by generating660 rules for a piece of hardware which was not within the original deductive closure of the system. However, the nature of the results illustrate that rules other than ones with a positive impact on energy management were produced. This means that the intent of the domain theory was not realised in every rule generated. The discrimination of useful rules should be the focal point of further work.665 7. Acknowledgements The authors would like to thank EPSRC for funding this work. References Burstein, M. H. (1986). A Model of Learning by Incremental Analogical Reason- ing and Debugging. Machine Learning: An Artificial Intelligence Approach670 (Vol. 2), . Cambria, E., Fu, J., Bisio, F., & Poria, S. (2015). AffectiveSpace 2: Enabling affective intuition for concept-level sentiment analysis. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence (pp. 508–514). Austin, Texas: AAAI Press.675 Cameron, N., Webster, M., Jump, M., & Fisher, M. (2011). Certification of a Civil UAS: A Virtual Engineering Approach. In 2011 AIAA Modelling Simulation and Technologies Conference and Exhibit (pp. 1–15). Portland, Oregon: AAAI Press. Carbonell, J. G. (1983). Derivational Analogy and Its Role in Problem Solving.680 In AAAI-83. Committee on Materials, Structures, and Aeronautics for Advanced Uninhab- ited Air Vehicles, Commission on Engineering and Technical Systems, N. R. C. (2000). Uninhabited Air Vehicles : Enabling Science for Military Systems. Washington, DC, USA: National Academies Press.685 25 Davies, T. R. (1987). A logical approach to reasoning by analogy. In In IJCAI- 87 (pp. 264–270). Morgan Kaufmann. DeJong, G. (2004). Explanation-Based Learning. In A. Tucker (Ed.), Computer Science Handbook. (2nd ed.). Dejong, G. (2006). Toward Robust Real-World Inference: A New Perspective690 on Explanation-Based Learning. Ellman, T. (1989). Explanation-based learning: a survey of programs and per- spectives. ACM Comput. Surv., 21, 163–221. doi:http://doi.acm.org/10. 1145/66443.66445. Faculty, T. A., & Kahn, D. (2001). The Design and Development of a Modular695 Avionics System. Falkenhainer, B. (1987). An examination of the third stage in the analogy process: verification-based analogical learning. In Proceedings of the 10th international joint conference on Artificial intelligence - Volume 1 IJCAI’87 (pp. 260–263). San Francisco, CA, USA: Morgan Kaufmann Publishers Inc.700 Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7, 155–170. Hirowatari, E., & Arikawa, S. (1994). Incorporating Explanation-Based Gener- alization with Analogical Reasoning. Bulletin of Informatics and Cybernetics, 26, 13–33.705 Huhns, M. N., & Acosta, R. D. (1988). Argo: a system for design by analogy. doi:10.1109/64.21891. Ishikawa, T., & Terano, T. (1996). Analogy by abstraction: Case retrieval and adaptation for inventive design expert systems. Expert Systems with Applications, 10, 351–356.710 Klenk, M., & Forbus, K. (2009). Domain transfer via cross-domain analogy. Cognitive Systems Research, 10, 240–250. Könik, T., O’Rorke, P., Shapiro, D., Choi, D., Nejati, N., & Langley, P. (2009). Skill transfer through goal-driven representation mapping. Cognitive Systems Research, 10, 270–285.715 Lopez, J., Royo, P., Barrado, C., & Pastor, E. (2008). Modular avionics for seamless reconfigurable UAS missions. doi:10.1109/DASC.2008.4702748. de Mántaras, R. L., & Plaza, E. (1997). Case-Based Reasoning: an overview. AI Commun., 10, 21–29. Mooney, R. J., Bennett, S. W., & Urbana, A. (1986). A DOMAIN INDE-720 PENDENT EXPLANATION-BASED GENERALIZER. In AAAI-86 (pp. 551–555). Philadelphia, Pennsylvania: AAAI Press. 26 http://dx.doi.org/http://doi.acm.org/10.1145/66443.66445 http://dx.doi.org/http://doi.acm.org/10.1145/66443.66445 http://dx.doi.org/http://doi.acm.org/10.1145/66443.66445 http://dx.doi.org/10.1109/64.21891 http://dx.doi.org/10.1109/DASC.2008.4702748 Steven, M. (1990). Quantitative results concerning the utility of explanation- based learning. Artificial Intelligence, 42, 363–391. Timperley, M. (2015). The Integration of Explanation-Based Learning and Fuzzy725 Control in the Context of Software Assurance as Applied to Modular Avionics. Ph.D. thesis University of Central Lancashire. VanLehn, K., & Jones, R. M. (1998). Learning Physics Via Explanation-Based Learning of Correctness and Analogical Search Control,. Wilson, A., & Preyssler, T. (2009). Incremental certification and Integrated730 Modular Avionics. Aerospace and Electronic Systems Magazine, IEEE, 24, 10–15. doi:10.1109/MAES.2009.5344176. 27 http://dx.doi.org/10.1109/MAES.2009.5344176 Introduction Analogy Related Work Abstraction Derivation Forming an Abstract Definition An Example Discussion Experiment Design Results and Discussion Conclusion Acknowledgements 
work_edr3aeu24nfxvhl7arsm5m7nou	----	Automatic Interchange of Knowledge between ontologies Knowledge accumulation through automatic merging of ontologies Adolfo Guzmán-Arenas and Alma-Delia Cuevas Centro de Investigación en Computación (CIC), Instituto Politécnico Nacional (IPN) and Instituto Tecnológico de Oaxaca (México) a.guzman@acm.org; almadeliacuevas@gmail.com Abstract. In order to compute intelligent answers to complex questions, using the vast amounts of information existing in the Web, computers have (1) to translate such knowledge, typically from text documents, into a data structure suitable for automatic exploitation; (2) to accumulate enough knowledge about a certain topic or area by integrating or fusing these data structures, taking into account new information, additional details, better precision, synonyms, homonyms, redundancies, apparent contradictions and inconsistencies found in the incoming data structures to be added; and (3) to perform deductions from that amassed body of knowledge, most likely through a general query processor. This article seeks to solve point (2) by using a method (OM, Ontology Merging), with its algorithm and implementation, to fuse two ontologies (coming from Web documents) without human intervention, producing a third ontology, taking into account the inconsistencies, contradictions and redundancies between them, thus delivering an answer close to reality. Results of OM working on ontologies extracted from Web documents are shown. Key words. Ontology fusion; Ontology alignment; Artificial Intelligence; Knowledge representation; Semantic Web. 1 Introduction In this post-industrial age, most treasured resources are (a) knowledge and (b) efficient processing of such knowledge; akin to land, cattle and slaves in the middle age. By (a) we mean not data, but information stored in an way amenable for computer deductions. Plenty of information, to be sure. By (b) we mean that the computer (and not just people) should be able to make deductions and infer (non trivial) consequences, and provide answers, through suitable deductive mechanisms or algorithms. We find ontologies (§2.1) suitable for this knowledge storage. Internet and searchers such as Google (or on-line encyclopedias such as Wikipedia) are able to provide large quantities of knowledge, thus partially fulfilling (a). Nevertheless, this knowledge (contained in a long list of perhaps relevant documents 1 ) must be manually processed (each of them has to be read by a person), missing goal (b). Current situation is similar to the librarian of a large library, who answers a petition from a user by supplying her a large list of books, some of which may contain the desired answer (or more often, from the careful reading of several books, the user obtains her answer). The desired situation would be a librarian (now a wise man) who replies with the right answer to a question about a topic that his library covers, since he already read and digested every book in his shelves. Since the computer must acquire the information in point (a) by gradual accumulation of much new knowledge (concepts, relations, typical values…) into the information it previously digested, the machine has to identify by itself redundancies, repeated information, small and big contradictions, synonyms and homonyms, among others cases. This paper presents OM (for ontology merging), an algorithm that (automatically) merges ontologies A and B yielding ontology C, thus addressing goal (2) in the abstract [Neither goal (1) nor (3) are addressed]. 2 C is consistent with the knowledge of A and B. Also, the accumulation of knowledge requires that a fair amount of ontologies about the same topic should be merged, which calls for repeated use of OM. The correct merge must consider not only the syntax of the word and phrases describing the concept (these are called the definition of a concept, see Figure 3), but the semantics of the concept thus represented, too [its “meaning” as described by its relations to other concepts in the source ontologies, its similarities to other concepts, superficial 1 Most of the irrelevant documents result because the search is specified through words, not through concepts. It is a syntactic search. 2 A tool addressing goal (1) may take the form of a good parser that uses its previous knowledge about the subject (most likely, another ontology) and translates the input text into an ontology representing the knowledge found in it. A tool for (3) looks like a general question-answerer or a deductive program that matches a question (written perhaps as a graph [or an ontology] with free variables, which will be bound to the desired values or answers) with the comprehensive ontology. See, for instance, [2]. Neither tools (1) or (3) are the subject of this paper. mailto:a.guzman@acm.org mailto:almadeliacuevas@gmail.com likeness, homophones, etc.]. 3 Therefore, the representation to be used (described in §2.2) should be able to characterize these relations. CyC [18] also sought to build a common-sense ontology, but by manual methods. Recent previous work is described in §1.4. Section 2 explains the parts of OM, Section 3 describes how OM achieves the knowledge accretion, and Section 4 shows some results. Figure 1. Part of an ontology A (refer to Figure 4) about wine, constructed from text in [25]. Only some nodes are shown 3 When reading a phrase such as “Clyde is an elephant,” human beings immediately assume that Clyde has a trunk, four gray feet, and so on (previous knowledge). “Elephant” has a lot of meaning to us. OM is incapable of doing that. Its only available knowledge is in ontologies A and B. Of course, if B (or A) stores the information that elephants have a trunk, then OM can reflect in C that Clyde has a trunk, too. If A and B both say that the earth is flat, so it will say C. OM is incapable of finding what holds in real life, what is true or right. Also, for useful knowledge accumulation (amassing true facts) in A, OM should be fed with trusted ontologies, and in the right order. For instance, in C = OM(OM(OM(OM(A0, A1), A2), A3), A4), the first ontology, A0, should be very basic and well formed, much like the first facts a child begins to learn in his(her) life. 1.1 Knowledge accumulation and extraction by people A. Accumulation into comprehensive documents. Comprehensive documents that describe in detail knowledge in a given area are now produced by informed writers, who express this knowledge in text documents (or in mathematical notation) in a precise and clear way. What they do is to articulate as clearly as possible their knowledge, accumulated through years of experience, deductions and experiments. Unfortunately, being these documents in a natural language, the extraction of knowledge from them by other persons is done by reading and understanding them, not by automated means. That is, they are useful for human deduction, not for machine deduction. B. Knowledge extraction from (A) by the reader. An user extracts from a comprehensive document (as described in (A) information to add to his/her knowledge, by reading and understanding it. In this way, his/her knowledge increases. Another way for the reader to increase his/her knowledge in a given topic is to find documents that may contain part of the information he/she seeks, looking for them in the library, in an encyclopedia, or in the Web. Internet contains a lot of information in billions of documents, amount them there are Web sites, texts documents, doorway services, music, images, maps, etc. When accessed through searchers (Google, CiteSeer,…), they recover only a small part of the available information because the access is in syntactic form (through labels, words and phrases); that is, through lexicographic and syntactic comparisons. Also, the answer is a large list of documents that are not always appropriated. Thus, each of them has to be read and discarded or digested (understood), a manual process for the reader. 1.2 Knowledge accumulation and extraction by computer C. Accumulation into comprehensive ontologies. If we believe that ontologies may be a useful way to represent knowledge for human and computer knowledge extraction, then a comprehensive ontology of a given subject may be produced by the tool of point (1) in the abstract, applied to the corresponding comprehensive text document ( §1.1.A). Lacking such document, the tool of point (1) in the abstract could be used to build an ontology for each of the many documents found as explained in §1.1.B, and then merging such ontologies using OM or some further improvement to OM. The result would be a comprehensive ontology useful for computer extraction of information and knowledge. D. Knowledge extraction from (C). Whatever information the user decides to extract from a comprehensive ontology could be mined by posing his/her query in suitable form to such ontology, using the tool of point (3) in the abstract. 1.3 Where does OM fit? OM may render possible the automatic accumulation of knowledge by computer, if enough specific ontologies were available. 4 A computer that knows a lot about a given topic, and that stores such knowledge in a way amenable to automatic processing (deduction, inferences), will be a formidable knowledge assistant. See also §3.3 “Future Work.” 1.4 Current work A. Building comprehensive ontologies. There have been projects [18] that seek to build a comprehensive ontology by hand, or a list of synonyms (collections of English words and phrases), notably WordNet [7]. B. Ontology representation languages. There are several languages used to represent ontologies. DAML + OIL (Darpa Agent Markup Language + Ontology Inference Layer) [4] is a tag language that provides semantic information to Web resources. As a proposition for ontology presentation, OIL uses formal semantics and logical reasoning to describe the meaning of terms and the implicit information deduction. OIL has a well defined syntax, based on DTD 5 and XML Schema 6 . The Resource Description Framework RDF/XML [15] is strongly related to XML; therefore, useful to use it in conjunction with XML. It is exploited mainly to represent metadata. OWL [1] has all the features of RDF and has been adopted by the W3C consortium as an standard language to represent ontologies in the semantic Web. The Knowledge 4 Relatively few ontologies exist now, since they have to be manually constructed. 5 DTD or Document Type Definition is as definition of an SGML document, an older version of XML, that specifies restrictions on its structure. 6 It is an XML base alternative for DTDs, describing the structure of an XML document. See: w3shools.com/schema/default.asp. Interchange Format KIF [9] is a language designed for knowledge exchange among different computer systems, created by different programmers at different times, languages, etc. OCML, Operational Concept Modeling Language [6], is a language useful to construct knowledge models, allowing the specification and operation of functions, relations, classes, instances and rules. The Open Knowledge Base Connectivity OKBC [20] is a protocol to access knowledge bases stored in knowledge representation systems, for instance an object oriented data base. Our OM notation (§2.2) was designed not to produce still another definition language, but as a fast way to provide ontologies the needed features that will enable OM’s work. Thus, if a given knowledge composition present in the source document was needed to be characterized in the ontology, the OM notation could be modified to allow its easy depiction. The advantage of this approach was a moldable representation language where features could be quickly added. The disadvantage is that we will need perhaps, in the future, to write a translator from OM notation to an standard ontology language. C. Accumulation of knowledge through ontologies. There are computer aided methods to identify and join the knowledge in two ontologies. These methods help the user to complete the work. PROMPT [8], Chimaera [16], OntoMerge [6], IF-Map and ISI [21] require that user solve problems presented during the fusion. Methods like FCA-Merge [19] use Formal Concept Analysis for ontology representation, forcing these to be necessarily consistent. But the majority of ontologies in Web present inconsistencies, either among different parts of the same ontology, or when one is compared with another. A recent advance in ontology union is HCONE-merge [13], which extracts from the semantic data base WordNet [7] intermediary information for the fusion, requiring less user support. 2 The parts of OM OM fuses ontologies represented in a notation, called simply “OM notation” (§2.2 and Figure 3). The examples in this paper show ontologies that were manually created from real documents extracted from the Web, and written in such notation. Then OM fused them (automatically, that is). The same ontologies were also manually fused. The manual result was then manually compared against the automatic result. Table 1 shows some of these comparisons. OM has some built-in (mostly syntactic) initial knowledge about words and word phrases (§2.3). OM rests heavily in two functions previously defined elsewhere: COM (§2.4), which compares two concepts in different ontologies, and conf (§2.5), which finds the confusion (a kind of dissimilarity) among two concepts belonging to the same hierarchy. 2.1 Ontology definition In Philosophy, ontology (which is called “being theory”) is a general analysis of the being, what is it, how is it and how is it possible. In Computer Science, ontology is a data structure, a data representation tool to share and reuse knowledge between Artificial Intelligence systems which share a common vocabulary. Thus, an ontology is a set of definitions, classes, relations, functions and other objects of speech [5]. Mathematically, it does not exist a satisfactory theory which characterizes formally the ontology definition. However, some intents have appeared, such as [12]. Formally [5], ontology is a pair O = (Ç, R) Where: Ç is a set of nodes (representing concepts), some of which are relations. R is a set of restrictions, of the form (r; c1; c2;…; ck) between the relation r and concepts c1,…, ck. Lower c is used to refer to each concepts of set Ç, while a semicolon separates members in the restrictions. For example: (cut; scissors; sheet), (print; printer; document; ink). In these examples, the concepts that are relations too, are: cut and print. The first element of a restriction is a relation. The restrictions are not limited to have two members besides the relation. Therefore, an ontology is a hypergraph with Ç the set of nodes and R the set of hyper-relations (which we call restrictions). For clarity, in this paper ontologies are represented as graphs, where nodes are concepts and edges are restrictions. Figure 1 shows part of an ontology about wine. Figure 2 is another ontology about wine. Notice how they describe wine from different view points. Figure 2. Ontology B about wine (refer to Figure 4 and Table 1), partially shown. A solid dot represents a partition. Thus, Table Wines are partitioned by color into Red wine, White wine and Pink wine. The relation character partitions Table wines into Sweet wine and Dry wine. Constructed from [24] 2.2 OM notation Until now, ontologies have been proposed to solve Semantic Web problems, facilitating computers to better understand Web sites, for instance. The ontologies should express the required semantics, and the language in which these ontologies are written should be powerful enough for this purpose. The purpose of OM notation is to facilitate automatic merging. It represents ontologies through a structural design with XML-like labels, identifying the concepts and their restrictions. Examples appear in Figure 3 and Figure 4. Figure 3. Representation of an ontology in OM Notation. The label of each concept (such as thing) comes after <concept>; the language of the concept’s definition (such as English) goes between <language> and </language>; the definition of the concept (a word or a world phrase such as concrete_object, physical_object) goes between <word> and </word>; the restrictions of the concept (such as eats of concept human being) go between <relation> and </relation>. The description of a concept ends in </concept>. Nested concepts (such as thing and physical_object) indicate that physical_object is subordinate (or hyponym) of thing, the precise meaning of this subordination is indicated by <subset> thing </subset>. The restrictions expressed by nesting are called implicit restrictions (currently, they are member of, part of, subset, part*). The other restrictions, such as eats, are called explicit. Explicit restrictions of a concept are known in other works as relations, properties or attributes of the concept. Notice how human being is partitioned by age. A complete description of the OM notation appears in [5] In OM Notation, a restriction can be n-ary; a restriction relates concepts (the first concept in a restriction is a relation). For example, (Color; Zebra; White; Black) is a restriction on concepts Color, Zebra, White and Black. The restrictions can be loosely considered as properties or characteristics of the node or concept where they are defined. Example: the restriction eats in Figure 3. Other restrictions exist, such as hyponym (subordinate) restrictions, that are expressed through nested concepts. Example: plant is a subset of physical object. The purpose of OM notation was not to introduce yet another language; but to provide the ways and means to represent the nuances of knowledge that OM required to perform an effective automatic accumulation. 2.2.1 Contributions of OM notation Most important are: (a) Ability to represent partitions. A partition of a set is a collection of subsets such that any two of them are mutually exclusive, and all are collectively exhaustive. OM can represent partitions, while current languages (§1.4.B) may not. For instance, not only male person and female person are subsets of person, they are a partition of person. The gender of a person will tell us to which of the partitions male person and female person that person belongs. See in Figure 2 how categories partitions wine into fortified wines, sparkling wines and table wines. Another example: if John is a person and person has a partition called age with subsets child, young person and adult, from the age of John (a number) the system can classify John as a child, a young person or an adult. (b) A concept can be a relation, too. Often, ontologies are represented as a graph O = (C, R) consisting of two disjoint sets: C (nodes, or concepts) and R (edges, or relations).7 Two disadvantages of this visually oriented approach are:  all the relations are binary;  a concept can not be a relation. In OM notation relations are also concepts (Cf. §2.1). With this improvement it is possible8 to know more about the relation that restricts the concepts. For example, one can say (mother of; Mary Ball Washington; George Washington) to indicate that Mary is George Washington’s mother, but mother of can be a concept that contains more information elsewhere in the ontology, for instance, related to child of by the relation inverse,9 and it is a subset of relation relative. 2.3 Initial knowledge used by OM OM has built-in some (mostly syntactic) initial knowledge about words and word phrases. These are: A.- Articles and linking words like in, the, to, this, and, or, etc. are ignored in the definition of the concept. B.- Words that change or reject concepts in the relation’s definition, such as: except, without. For example: Poppy without Petiole. Which means that the concept Petiole does not form part of the concept Poppy. C.- Hierarchies of concepts. 10 The hierarchy represents a taxonomy of related terms. It is used to compute the confusion (§2.5). We exploit it to detect synonyms and “false inconsistencies” due to degree of detail. Additional sources of initial knowledge (not yet incorporated into current OM) could be: D.- OM could use as part of its initial knowledge the previous knowledge already amassed. See formula in footnote 3. OM does not do this, yet. Other potential sources of initial knowledge are mentioned in §4.2. 2.4 COM finds the closest concept to CA in another ontology B Let A, B be ontologies. The algorithm COM takes a concept CAA and finds its most similar concept CBB. It has been described elsewhere [10]; we explain it here briefly for completeness. PACA stands for the parent of A, 11 and PBB stands for the parent 11 of CB. To find the similarity between nodes of an ontology, OM uses COM, which employs the definitions (words) and restrictions on those concepts according to four cases: Case A: CA finds its match CBB and its parent PA also finds its corresponding PBB (§2.4.1). Case B: PA finds its match PBB but CA can not find its corresponding CBB (§2.4.2). Case C: CA finds its corresponding CB but its parent PA can not find a suitable PBB (§2.4.3). Case D: CA can not find its corresponding CB, and neither PA can find its corresponding PBB (§2.4.4). 2.4.1. Case A: CA finds its match CB and PA also finds its match PB Concepts CB and PB are looked in B, seeing if the definition of some concept in B matches the majority of the words and word phrases of the definition of CA, and the definition of a parent of this CB matches the majority of the words and word phrases of the definition of PA, a parent of CA. If needed, grandparents of CA and CB are considered, too. Finding a match, COM returns two values: (1) The CB found in B; that is, the most similar concept to CA; (2) The similarity value vs, a number between 0 (no similarity) and 1 (completely similar). 7 Earlier ontology representations were even more restrictive: an ontology had to be a tree, a graph without cycles. In these, a concept could not have two ancestors (two parents): Mexico could not be both a nation and an emerging market. 8 But not required. Thus, a relation represented by just its name is a “shallow” relation of which little is known (except its name). 9 Another example: the relation to light (that is, to illuminate) may also be a concept, if one wishes to add more properties to this action (for instance, that it spreads in 3 dimensions). A different concept that can also be represented is light (an electromagnetic radiation). OM does not consider them equivalent, even if somebody gave them the same name. OM has machinery to identify homonyms (§3.6). 10 A hierarchy H of an element set E (a set whose elements are non-numeric constants) is a tree with root E and where each node it is either an element of E or it is a set whose sons form a partition of such node. 11 Or for a parent of it, since a concept can have several parents. Each possible parent is considered in turn. Figure 4. Ontologies A, B, C about wines, partially shown. The result of the fusion is ontology C. How OM produced the result is described in §3. Accuracy of C is found in table 1. Example: Suppose hammer  A has the definition (hammer; mallet) and its parent tool  A has definition (tool; utensils; appliance). mallet  B has the definition (mallet; hammer), while its parent implement has definition (implement; utensils; device; tool). In this case hammer  A matches mallet  B, since its parent tool matches implement. COM returns mallet as the most similar concept to hammer  A, and vs=1. Example: suppose we want to find the most similar concept to lemon  A. Let us have (parent of; lemon; fruit)A, (parent of; lemon; citric)B, (parent of; orange; citric)B, (parent of; tangerine; citric)B, (parent of; citric; fruit)B, (parent of; apple; fruit)B, (parent of; pineapple; fruit)B. When looking for the most similar concept to lemon  A, its parent fruit  A does not match with citric  B, but with the grandparent of lemon  B, namely fruit  B. Thus, COM returns lemon  B. 2.4.2. Case B: The parent PB similar to PA is found, but none of PB’sons matches CA If no CB is found, then COM is called recursively with PA as parameter to see who is PA’s parent. If the parent of PA is the root of A (OARoot), COM finishes without success. Otherwise, each son of PB is sequentially examined. If some son CB’ of PB has the majority of its restrictions and values similar (through recursive use of COM) to those of CA, such CB’ is the one looked for, and COM returns CB’B as the concept most similar to CA. If no son CB’ of PB has such property, then we search for such CB’ among the sons of the parent of PB, that is, among the brothers of PB. If needed, such CB’ is sought among the sons of the sons of PB, that is, among the grandsons of PB. The successful CB’ is the desired answer. If no CB’ can be found after all this search, then COM returns “a new son of PB,” meaning that the most similar node in B is a son of PB, which does not yet exist in B. In this case OM will just add CA as an additional son to PB in the resulting ontology C. This is how synonyms are handled by OM. For instance, we want to find in B the most similar concept to maize  A. Suppose (parent of; maize; cereal)A, (parent of; corn; cereal)B, (parent of; wheat; cereal)B, (parent of; sorghum; cereal)B. We see that cereal  A matches cereal  B, but no maize  B is found. Then COM looks successively at corn, wheat, sorghum, observing if most of the properties of one of these sons (size of seed, shape, color…) of cereal match the corresponding properties of maize. The answer is corn, which is returned by COM. If (parent of; corn; cereal)B were absent from B, then COM will return “a new son of cereal  B” as the most similar value to maize  A. 2.4.3. Case C: CA finds its corresponding CB but its parent PA can not find a suitable PBB If CB is found but no PB is found, then the grandfather of CBB and great-grandfather of CBB are tested to see if they are similar to PA. If some of them is, that is enough and CB is returned as the value of COM (this is Case A). If this does not happen, COM verifies if the majority of the restrictions (relations and its values) of CA are compatible with those of CB; also, do most of the sons of CA coincide (through recursive use of COM) with respective sons of CB? Such being the case, CB is the most similar concept to CA, and COM finishes. If just part of the properties of CB are similar to those of CA, the answer is “probably CB” and the algorithm finishes. If no one or few properties nor sons are similar, then the answer “doesn’t exist” is returned. This case considers concepts that have the same definition but different meaning, for instance, if (parent of; bill; invoice)A, while (parent of; bill; currency)B, then COM returns as the most similar concept to billA the value “it does not exist” in B. But it will also find that Mexico  A is most similar to Mexico  B, where (parent of; Mexico; nation)A and (parent of; Mexico; emerging market)B, since most properties of both Mexico’s coincide. In C, Mexico (represented by a single concept) will be both son of nation and of emerging market. Another example: if A contains (parent of; tangerine; citric), and B contains restrictions (parent of; orange; citric), (parent of; lemon; citric), (parent of; grapefruit; citric), then COM may return “probably orange  B” as the most similar concept to tangerine  A, depending on how many restrictions in both fruits were similar. 2.4.4. Case D: CA can not find its corresponding CB, and neither PA can If neither CB nor PB can be found in B, then the answer is “doesn’t exist” and COM ends. This situation is common when we are talking about two different ontologies, for example: A is an ontology about Information Systems while B is an ontology about Natural Resources. This is a very easy case for OM: if A knows nothing about dinosaurs, and B knows a lot, then all of B’s dinosaur knowledge is added to C, since it finds no contradictions or equivalences in A. But beware that A may have some knowledge that is also shared with B’s dinosaur knowledge, for instance the concepts size, leg, skin, feather, wing, fang, herbivore, etc. These shared concepts are fused, not just added to C. 2.5. conf finds the confusion among two concepts placed in a hierarchy The function conf has been explained elsewhere [14]; we give here a short introduction, for completeness. conf(r, s) measures the confusion (a kind of error) when object r is used instead of the intended or real object s. What is the capital of Germany? Berlin is the correct answer; Frankfurt is a close miss, Madrid a fair error, and sausage a gross error. What is closer to a cat, a dog or an orange? Can we measure these errors and similarities? Can we retrieve objects in a database that are close to a desired item? Yes, by arranging these symbolic (that is, non numeric) values in a hierarchy. 10 The confusion appears when a value of a restriction in CA is incompatible, contradicts or is inconsistent with respect to the corresponding value of the equivalent restriction in CB, Example: (form; Earth; round)A and (form; Earth; flat)B. conf gives a value between 0 (total contradiction) and 1 (total compatibility), not just 1 or 0. Let r and s be two values (nodes) in a hierarchy with height 12 h. The function CONF(r, s) called absolute confusion that result by using r instead of s (the intended value) is: CONF(r, r) = CONF(r, s) = 0 when s is one of the r antecessor; CONF(r, s) = 1 + CONF (r, father_of(s)) otherwise. CONF(r, s) is the number of descending links as the hierarchy is traveled from r to s, the intended value. CONF is not a symmetric function. Its value can be altered by simply adding additional nodes between or levels in the hierarchy. To avoid this, we define the relative confusion, or simply confusion. Definition. The function conf(r, s) that results by using r instead of s is: h srCONF srconf ),( ),(  The function conf(r, s) is the absolute confusion CONF(r, s) divided by h, the height of the hierarchy. Example: If a hierarchy has (part of; San Pablo Guelatao; Ixtlan Mountains), (part of, Ixtlan Mountains; Oaxaca), (part of; Oaxaca; Mexico), then CONF(San Pablo Guelatao, Mexico) = 0; CONF (Mexico, San Pablo Guelatao) = 2. 3. Automatic merging of ontologies by OM OM merges two ontologies A and B into a result C. At least the roots of A and B coincide. OM was developed at CIC-IPN in a doctoral dissertation [5]. It is a totally automatic algorithm because it does not need user intervention and it is strong because it can join ontologies with inconsistencies (see Table 1). The general form for obtaining C is: C = A  {cB’ , rB | cB’ , rB = ext(rA, rB)  cA  A } The resulting ontology C is the original A increased by some concepts and restrictions of B that the function ext gets. Where: CA is a concept of ontology A, rA are the restrictions of cA that exist in A, rB are restrictions of cB that exist in B and cB is the most similar concept in B to cA (as found by COM, §2.4);  means ontology joining. It is a carefully joining, different to set union. ext(rA, rB) is the algorithm that completes the restrictions rB which are not in A with those cB’ (which are in B) that do not contradict knowledge from A. That is, for each node cA  A, all its restrictions in A are retained in C, and only some restrictions in B of cB (the concept most similar in B to cA) are added to C, as well as their “target” concept cB’. For instance, let cA = Gottfried Wilhelm Leibniz with most similar concept cB = Leibniz  B. If the restriction (rB cB cB’) = (was born in; Leibniz; Leipzig) is in B, and if it does not contradict knowledge from A, then restriction (rA cA cB’) = (was born in; Gottfried Wilhelm Leibniz; Leipzig) is added to cA in C. More examples below. This extraction must be carefully, so as not to introduce inconsistencies, contradictions or redundant information in C. ext is large, and it is explained in §3.2 to 3.7. When applying ext to each concept cAA, OM adds to C the additional 12 The height of a tree is the number of ridge that exist in the way of its root to the most distant leaf. knowledge (about cB) in B not present in A. Notice that as new nodes from B are added to C, these are also searched for additional restrictions in B. So that the quantifier  cAA in the above formula should really say  CCOLD. Past ontologies generated by OM could be available to OM as built-in knowledge to be used in subsequent knowledge accumulations (Cf. §2.3.D). This will greatly enhance its merging ability. 3.1 General description of OM to fuse ontologies An overview of OM is: 1. Copy ontology A into C. 2. Starting from root concept cRoot in C, and through each concept in C (including those recently added as partial result of the fusion): I. Look in B for its most similar concept cB (using COM of §2.4). II. If such cB exists in B, then a) its new restrictions are added to C; b) redundant restrictions are eliminated(§3.2); c) some collections of subsets are promoted to partitions (§3.3); d) new partitions are added to the result C (§3.5); e) identical partitions are detected (§3.3); f) synonymous are verified (§3.4); g) some inconsistencies are detected and solved (using Confusion theory, §2.5); h) homonyms are detected and handled as two different concepts (§3.6); i) some contradictions are detected and solved (§3.7), so that inconsistency does not creep into the result. III. If no such cB is found in B, then take the following concept cC ( traveling C in depth-first priority) and go to I. If in step II the contradiction between some node in A and another node in B could not be solved, then the value of A prevails, and it is conserved in the result C. Thus, if A says that the Earth is flat, and B that it is round, and no further knowledge is available to detect the contradiction, A prevails. More in §3.7. 3. Stop when all the concepts of C are analyzed. 3.2 Elimination of redundant restrictions OM purges redundant restrictions from the resulting fusion, as these are created. Example: In Figure 5, ontology A has the knowledge (part; Apparatus; Organ), while B knows (subset; Apparatus(System); Thing). Thus, the result C should have both of these restrictions. Nevertheless, OM compares the predecessors of Apparatus in A against the predecessors of Apparatus (System) in B, and discovers that restriction (subset; Apparatus; Thing) can be derived now in C through the recently introduced restriction (part; Apparatus; Organ). Therefore, (subset; Apparatus (System); Thing) is deleted from C. 3.3 Handling subsets and partitions If an ontology has a partition of a given node into subsets, whereas the other ontology has some of these subsets present, then the partition prevails in the result C. Subsets outside those forming the partition still go to C. Example: height  A partitions person  A into short person, medium height person, tall person. In B, person has subsets physician, tall person, engineer, medium height person, plumber, short person. Then person  C will have the partition short person, medium height person, tall person, and the subsets physician, engineer, plumber. If one ontology partitions a node in a way (example: age partitions person  A into young, mature, old), and the other ontology partitions the same node in a different way (gender partitions person  B into male, female), then both partitions appear in C. A partition appearing in just one of the ontologies goes to the result (§3.5). Two partitions on the same relation (say, age partitions person  A into young, mature, old, while age partitions person  B into juvenile, adult, senile) may be merged if the corresponding subsets are indeed equivalent; otherwise, both partitions go to C. More about partition handling in [5]. Example: in Figure 6 the concept Human body system in A finds its most similar concept in B to be Human body system, too, which has the partition System. OM observes that the parts of this partition correspond to known parts of Human body system in A. Thus, System is kept as partition in C. Figure 5. In ontology C the concept Apparatus (System) has the restrictions (part; Apparatus (System); Organ) and (subset; Apparatus (System); Thing). This last restriction is redundant and OM eliminates it. Also, in C the concept Locomotive has the definition (Locomotive Apparatus, Locomotive System) and the restrictions (part; Locomotive; Organ) and (part; Locomotive; Apparatus (System)). This last restriction is expunged from C by OM, which deems it redundant . Figure 6. The partition System of the concept Human body system in B is added to resulting ontology C, replacing the less precise knowledge that Muscular System, Skeletal System, Cardiovascular System, Articular System and Nervous System as just part of Human body system. That is, for A some other system besides those five could still be part of Human body system, while for B (or C) this is not possible 3.4 Identification of synonyms and their treatment Figure 7 shows an example with concepts Locomotive system in A and Locomotive Apparatus in B. Through COM, OM finds that for concept Locomotive system in A its most similar concept in B is Locomotive Apparatus, by syntactic comparison of their respective definitions “Locomotive System/Locomotive Apparatus” and “Locomotive Apparatus.” Therefore, both concepts are merged into a single concept in C, which inherits all the restrictions of Locomotive system  A and Locomotive Apparatus  B. Figure 7. Ontologies A and B where Locomotive system and Locomotive Apparatus concepts are detected as synonyms This same process of identification of synonymy is applied when the concepts are relations. 3.5 Copy of a new partition The example in Figure 8 shows ontology A with the concept Esplacnology having a partition that Esplacnology in B lacks. Therefore, the partition is added to the resulting ontology. Figure 8. The partition systems is copied into the resulting ontology C 3.6 Homonyms –detection and proper handling Fluke in A and fluke in B are detected as different concepts, since they have different parents, different descendants, and different properties. Both are copied to C (Figure 9), as two different concepts with the same name. Figure 9. Fluke in ontologies A and B are homonyms, thus C has two different concepts with the same name 3.7 Detecting contradictions and resolving some of them In addition to the detection of some contradictions shown already in §§3.3, 3.4, 3.6, OM has another powerful way to detect and solve some types of contradictions. Many relations can have several values (say, my_friend_is), but some relations (say, my_father_is) can only have one value. Thus, Lincoln can not have two fathers. A person can not be born in two different places. He or she can not be buried in two different sites. How does OM handle the case where A says (was born in; Lincoln; Kentucky) while B says (was born in; Lincoln; USA)? First, it checks that the relation was_born_in is indeed single-valued. If it is, it checks that the relation was_born_in  A is the same as the relation was_born_in  B (it could be just homonyms, with different meaning, as in §3.6). Then, it looks at USA and Kentucky. Are they perhaps synonyms, just as maize and corn (§3.4)? Failing all of this, OM resorts to conf (§2.5) to find the confusion between USA and Kentucky, and the confusion between Kentucky and USA, and it finds that one of them, conf(Kentucky, USA), is 0. Kentucky can be used instead of USA in this case; so, OM keeps the most specific value [the value r that forced conf(r, s) to be 0], Kentucky in this case, in the result C. When trying to merge (place of birth; Gottfried Wilhelm Leibniz, Germany)A with (place of birth; Gottfried Wilhelm Leibniz, Leipzig)B (Figure 10), the ontology in Figure 11 is of no use for conf (it does not mention Leipzig, a city of Saxony). Then, OM does not add (place of birth; Gottfried Wilhelm Leibniz, Leipzig) to C, but it keeps the current knowledge form A, that is (place of birth; Gottfried Wilhelm Leibniz, Germany). Thus, if OM can not solve the contradiction between knowledge in A and knowledge in B, A prevails. OM refuses to add to C the unresolved contradicting knowledge coming from B. The unfortunate outcome is that once C gets something wrong, it will not change it regardless of enough posterior evidence to the contrary. §4.2.E suggests how to defeat this. Figure 10. Ontology A says that (place of birth; Gottfried Wilhelm Leibniz; Germany) while B says (place of birth; Gottfried Wilhelm Leibniz; Leipzig). Contrary to our intuition, OM selected (place of birth; Gottfried Wilhelm Leibniz; Germany) as the result to keep in C. The text in §3.7 explains why 4 Results and discussion We now present some results of the automatic accumulation or merging of ontologies A and B by OM. These two ontologies were merged to produce the result C (another ontology). Each source pair A and B describe the same topic; both must have the same root concept. A and B were manually obtained from different documents in the Web. Then, C was automatically produced by OM. To check this resulting ontology, the same source ontologies were manually fused, this fusion being by definition “correct.” The error and efficiency of the result C was computed by the following formulas, where T is the total number of restrictions and concepts in the correct answer: Error = (number of wrong restrictions and concepts in C) / T Efficiency = 100 * (number of correct restrictions and concepts in C) / T. The first column of Table 1 presents the source ontologies. The time that OM took to fuse them appears in column 2. The slowest fusion is for One hundred years of solitude because it has more restrictions and OM verifies carefully elements of each restriction. The second column shows details for concept accumulation; the third column, for accumulation of restrictions. These columns also compare the result of OM against the correct result. Figure 11. Parts of Germany 4.1 Discussion Recent ontology fusion methods perform the alignment (suggesting which concepts in A should match with which of B) by comparing names and labels of such concepts and of their neighbors. Fusion (to decide which of the suggested concepts is indeed the right one to be fused) is done with the help of a person. HCONE-Merge [13] uses the data base WordNet [7] to find the meaning of the concepts in the alignment, alleviating the user chores. OM uses the definition of the concepts, their neighborhood and their characteristics. These are verified through a recursive process (each characteristic [a relation] can be also a concept, and it is thus analyzed as such). This recursive process can be interpreted like a semantic search or semantic analysis of the concepts. Instead of relying on a person to do the fusion of the aligned concepts, OM relies on COM (§2.4) and the confusion conf (§2.5), thus achieving automatic fusion. As an alternative to the subjective way of manually checking the results of OM to gauge its efficiency and error rate, perhaps a better way to verify them would be to use tool (3) of the abstract. With that question-answerer at hand, the user could pose “interesting questions” to the result C. The answers could be verified (manually) to see if indeed they are right. Again, we can not escape from subjectivity. A Ph. D. thesis in progress [2] seeks to generate this question-answerer, but for heterogeneous data bases, not for ontologies. Table 1. Behavior of OM in some real examples Source Ontologies Results for concept accumulation Results for accumulation of restrictions error % Effic Wine, Figure 4 [24, 25]. The 25 concepts of B were added and fused correctly to the 25 of A, getting a total of 46 concepts in C. 25A + 25B = 46C. (2 sec). 31 restrictions of B were added and fused correctly into 27 of A, getting a total of 55 restrictions in C. 31B + 27A = 55C. 0 100 Neurotransmis sion (A) and Schizophrenia. (B) [26]. 56A + 26B = 77C. The manual method gave 79 (2 of 79 concepts were missing) (2 sec). 79A + 51B = 127C. The manual method gave 129 (2 of 129 were not copied). 0.019 98 Human anatomy. 48B + 28A = 59C (2 sec) [22, 23]. 66B + 30A = 104C (New relations were deduced after fusion and added). 0 100 Continent. 54B + 50A = 66C (3 sec) [27, 28]. 40B + 34A = 46C. 0 100 Self- inconsistent ontologies. 5B + 6A = 9C. There were 5 original inconsistencies (3 concepts in A and 2 in B were wrongly classified). C had the same 5 original inconsistencies (1 sec). 3B + 4A = 7C. There were 2 inconsistencies (2 restrictions that were not defined, one on each ontology). The result C had 1 inconsistency, OM solved the other. 13 0 100 One hundred years of solitude. 126B + 90A = 141C. The manual method gave 149 (8 of 149 concepts were missed). (10 minutes). [29, 30] 283B + 231A = 420C. The manual method gave 432 (12 of 432 were not copied). 0.034 96.5 Solar System [31, 32]. 60B + 79A = 125C. (Everything was correct) (4 sec.) 45B + 56A = 59C. 0 100 Oaxaca (5 min). 117B + 234A = 309C. Manual method gave 310 (1 of 310 was not copied). 43B + 61A = 96C [33, 34]. 0.002 99.7 OM needs to be tried on large, extensive ontologies. Unfortunately, few large ontologies are publicly available now. Of them, most are “shallow” in that they do not have enough restrictions to provide meaning to the nodes; they still rely heavily on the name associated to the node to characterize it (what we call its definition, see caption to Figure 3). With respect to the repetitive accumulation of knowledge in order to achieve a large resulting ontology (§1.2 and footnote 4), we have not done any practical test, yet. Lack of enough ontologies on the same topic to accumulate is the main reason. Moreover, when carried out, this accumulation must be done under supervision (the user should double check that no inconsistency creeps into the result C, specially in the early stages of knowledge accumulation 14 ) and in the right order (Cf. footnote 3). The first ontologies to be assimilated should be basic, simple, accurate and solid, since they will be the foundation of further knowledge buildup. Is this automated assimilation or accumulation of knowledge the entrance to “intelligence by computers” or “machine learning,” or is it just another way to process knowledge? Is the computer capable to learn by reading documents? These are some of the questions that could be answered (by somebody else, we prefer to keep working  ). 13 In this case it is difficult to know what is the correct (manual) answer, since the source ontologies are self-inconsistent to begin with. 14 As OM accumulates more knowledge, it will be more resistant to inconsistency creeping into its knowledge, since it will know more. 4.2 Recommendations for future work Of course, the parser and the question answerer, points (1) and (3) in the abstract, should be tackled. Also, some additions to enrich OM and therefore improve the fusion process could be: A. Use Wordnet, Wordmenu, etc. to improve the disambiguation of terms by increasing its knowledge of synonyms and near by concepts; B. Annotate the text to be parsed by assigning to each word its correct sense. This is word disambiguation, as in [3]. C. Add built-in knowledge (or better yet, knowledge explicitly represented in OM notation) about the relation part of, to understand that a boat is composed of prow, stern, deck, rudder, propeller… (using perhaps WordMenu). As a generalization, if additional knowledge about relations could be conveniently added in OM notation, then OM could reason “more deeply” and better about these relations and their arguments in the corresponding restrictions. For instance, it could construct (deduct) new restrictions. This is opposite to defining this additional knowledge about relations by formal means, by logic equations, by a mathematical notation, or by a programming language, all of them opaque to OM. D. Currently, OM does not know that things can change. It is not aware of the passage of time. It gets confused when A says that George Bush was governor of Texas, while B says that he was president of USA, since a person can not have both jobs simultaneously (OM ignores that the jobs were held sequentially). We notice that some documents do not mention when something happened, for instance “he left Santa Cruz seminary.” Other phrases give a partial knowledge of time, such as: “he visited most of the Mexican republic carrying national documents after that he had been moved from his position office” or similar form. Other writings mention things that happen all the time; permanently, such as “Rouse is Benito’s sister,” or “Juarez was born in Oaxaca.” A relevant work is [17]. E. The same for events that occur in space (different places at the same time). How is it possible that it is both raining in Mexico and not raining in Mexico? (Because Mexico is large). F. For arguments that are single valued (such as the place of birth), OM keeps just one value, as expected. If it could keep more than one (the “right” or “current” answer and other “less likely” answers), then OM could change its mind if enough evidence arrives, something that actually it can not do (Cf. §3.7). A way to do this is to use the theory of inconsistency, now under development [11]. Acknowledgments. Work herein reported was partially supported by Project 43377 (CONACYT). The authors recognize the partial support of CONACYT and SNI. Prof. Sergei Levachkine has been providing useful suggestions. References URLs were last time consulted on November 17, 2008. 1. Bechnofer, S., van Harmelen, F., Hendler, J., Horrocks, I., McGuinness, D.L., Patel-Schneider, P. F., Andrea, L. (2004) OWL Web Ontology Language Reference. W3C Recommendation. http://www.w3.org/TR/2004/REC-owl- ref-20040210/ 2. Botello, A. Inferring relations between autonomous data bases. CIC-IPN. Ph. D. thesis in progress. In Spanish. 3. Colorado, F. Mapping words to concepts: disambiguation. M. Sc. thesis in progress. CIC-IPN. In Spanish. 4. Connolly, C., Harmelen, F., Horrocks, I., McGuinnes, D. L., Patel-Schneider, P. F. Andrea Stein L. (2001) DAML+OIL Reference description. W3C Note 18. http://www.w3.org/TR/2001/NOTE-daml+oil-reference- 20011218 5. Cuevas, A. Ontology union using semantic properties. (2006) Ph. D. thesis, CIC-IPN. In Spanish. 6. Dou, D., McDermott, D., and Qi, P. (2002) Ontology Translation by Ontology Merging and Automated Reasoning. In Proc. EKAW Workshop on Ontologies for Multi-Agent Systems. 7. Fellbaum, C. (1999) WordNet An Electronic Lexical Database. Library of Congress Cataloging in Publication Data. 8. Fridman, N., and Musen, M. (2000) PROMPT: Algoritm and Tool for Automated Ontology Merging and Alignment. In Proc. Seventeenth National Conference on Artificial Intelligence. pp 450-455, Austin, TX, USA. 9. Genesereth, M. Knowledge Interchange Format. Draft proposed American National Standard NCITS.T2/98-004. 10. Guzman, A., and Olivares-Ceja, J. (2006) Measuring the understanding between two agents through concept similarity. Expert Systems with Applications 30, 4 577-591, Elsevier. 11. Jimenez, A. Characterizing and pondering logical properties on qualitative values organized into hierarchies. Ph. D. thesis in progress. CIC-IPN. In Spanish. http://www.w3.org/TR/2004/REC-owl-ref-20040210/ http://www.w3.org/TR/2004/REC-owl-ref-20040210/ http://www.w3.org/TR/2001/NOTE-daml+oil-reference-20011218 http://www.w3.org/TR/2001/NOTE-daml+oil-reference-20011218 12. Kalfoglou, Y., and Schorlemmer, M. (2002) Information-Flow-based Ontology Mapping. Proceedings of the 1 st International Conference on Ontologies, Databases and Application of Semantics (ODBASE’02), Irvine, CA. 13. Kotis, K., and Vouros, G., Stergiou, K. Towards Automatic of Domain Ontologies: The HCONE-merge approach. Elsevier’s Journal of Web Semantic (JWS) 4:1, pp 60-79. 14. Levachkine, S., and Guzman, A. (2007) Hierarchy as a new data type for qualitative variables. Expert Systems with Applications 32, 3, 899-910. 15. Manola, F., Miller, E. RDF Primer. W3C Recommendation. 10 February 2004. http://www.w3.org/TR/2004/REC- rdf-primer-20040210/ 16. McGuinness, D., Fikes, R., Rice, J., and Wilder, S. (2000) The Chimaera Ontology Environment Knowledge. In Proceedings of the Eighth International Conference on Conceptual Structures Logical, Linguistic, and Computational Issues (ICCS 2000). Darmstadt, Germany. 17. Puscasu, G., Ramirez Barco P, et al. (2006) On the identification of temporal clauses. LNAI 4293, 911-921. 18. Reed, S. L., and Lenat, D. (2002) Mapping Ontologies into CyC. In proceeding of AAAI Workshop on Ontologies and the Semantic Web, Edmonton, Canada. 19. Stumme, G., Maedche, A. (2002) Ontology Merging for Federated ontologies on the Semantic Web. In: E. Franconi, K. Barker, D. Calvanese (Eds.): Proc. Intl. Workshop on Foundations of Models for Information Integration (FMII’01), Viterbo, Italy, 2001. lNAI, Springer (In press). 20. WEB Onto. http://137.108.64.26:3000/webonto?ontology=AKTIVE-PORTAL-ONTOLOGY&name= TECHNOLOGY&type=CLASS References in Internet Internet references [21, 24, 25] are in English; the rest are in Spanish. 21. Loom. http://www.isi.edu/isd/LOOM/LOOM-HOME.html 22. Human anatomy (A). From text in: http://es.wikipedia.org/wiki/Anatom%C3%ADa_humana 23. Human body (B) was built from text in: http://www.salonhogar.net/CuerpoHumano/Introd_Cuerpo_humano.htm 24. Wine (B). From text in: http://www.studyworld.com/newsite/ReportEssay/CreativeWriting/PersonalEssays%5CClassification_of_Wine- 40444.htm. 25. Wine (A). From text in http://en.wikipedia.org/wiki/Wine 26. We thank Paola Nery www.geocities.com/paolaneriortiz/ because she manually converted two documents into ontologies: neurotransmisor (A): es.wikipedia.org/wiki/Neurotransmisor and esquizofrenia (B): www.nimh.nih.gov/publicat/spSchizoph3517.cfm. 27. Continent (B) was produced from text in: http://es.wikipedia.org/wiki/Redefinici%C3%B3n_de_planeta_de_2006. 28. Continent (A) was obtained from text in http://es.wikipedia.org/wiki/Continente 29. 100 years (A) was from text in http://html.rincondelvago.com/cien-anos-de-soledad_gabriel-garcia-marquez22.html (document no longer available) 30. 100 years (B) was obtained from text in http://www.monografías.com/trabajos10/ciso/ciso.shtml 31. Solar System (B) comes from text in: http://www.solarviews.com/span/solarsys.htm. 32. Solar System (A) was formed from text in http://es.wikipedia.org/wiki/Sistema_Solar 33. Oaxaca (A) was obtained from text in www.oaxaca-mio.com/atrac_turisticos/infooaxaca.htm 34. Oaxaca (B) was extracted from text found in www.elbalero.gob.mx/explora/html/oaxaca/geografia.html http://www.w3.org/TR/2004/REC-rdf-primer-20040210/ http://www.w3.org/TR/2004/REC-rdf-primer-20040210/ http://www.isi.edu/isd/LOOM/LOOM-HOME.html http://es.wikipedia.org/wiki/Anatom%C3%ADa_humana http://www.salonhogar.net/CuerpoHumano/Introd_Cuerpo_humano.htm http://www.studyworld.com/newsite/ReportEssay/CreativeWriting/PersonalEssays%5CClassification_of_Wine-40444.htm http://www.studyworld.com/newsite/ReportEssay/CreativeWriting/PersonalEssays%5CClassification_of_Wine-40444.htm http://en.wikipedia.org/wiki/Wine http://www.nimh.nih.gov/publicat/spSchizoph3517.cfm http://es.wikipedia.org/wiki/Redefinici%C3%B3n_de_planeta_de_2006 http://es.wikipedia.org/wiki/Continente http://html.rincondelvago.com/cien-anos-de-soledad_gabriel-garcia-marquez22.html http://www.monografías.com/trabajos10/ciso/ciso.shtml http://www.solarviews.com/span/solarsys.htm http://es.wikipedia.org/wiki/Sistema_Solar http://www.oaxaca-mio.com/atrac_turisticos/infooaxaca.htm http://www.elbalero.gob.mx/explora/html/oaxaca/geografia.html 
work_ehnrcr6edngixmx6qgbvy6fore	----	Extracting expertise from experts: Methods for knowledge acquisition Extracting expertise from experts: Methods for knowledge acquisition Abstract: Knowledge acquisition is the biggest bottleneck in the development if expert systems. Fortunately, the process of translating expert knowledge to a,form suitable f o r expert system development can benefit f r o m methods developed by cognitive science to reveal human knowledge structures. There are two clusses of t h i w investigative methods, direct and indirect. W e provide reviews, criteria f o r ’ use, and literature sources f o r all principal methods. Direct methods discussed are: inret-views, questionnaires, observation of tusk perjbrmance, protocol analysis, interruption analysis, closed curves. and inferential flow analysis. Indirect methods include: multidimensional scaling, hierarchical clustering, Kenera1 weighted network.s, ordered trees, and repertory grid analysis. J U DlTH REITMAN OLSON Cruduute School of Business Administrurion The University of Mic,higur Ann Arbor Michigan IJSA HENRY H. RUETER Vwtor. Research Inc Ann Arbor Michigun U S A 1. Introduction Expert systems are here to stay. Although they are out of fashion in the research laboratories, they are nonetheless growing in popularity inside business and industry. Armies of small expert systems are being built to solve routine, medium- difficult problems in areas such as diagnosis of engine failures, tax planning, and feasibility analysis of cases i n union disputes about seniority. The literature on how to build expert systems is burgeoning. People flock to seminars and tutorials at professional conferences because they need training on how to build systems for very practical ends. I t has long been recognized that the system development process for expert systems and other artificial intelligence applications is dif- ferent from that for standard software. Whereas standard software development has a small frac- tion of the time spent in planning and a lot of coding, the majority of the time in expert system development is in planning - in deciding what knowledge should be encoded into the system. The bottleneck in the development of expert sys- tems is in extracting the knowledge from the ex- pert, that is, in knowledge acquisition I I ] and 121. Even books about how to build expert sys- tems have very little to offer about knowledge ac- quisition. Expert system developers have relied on two standard methods for getting the knowledge out 3f the expert and into the system: I ) The leveloper, or knowledge engineer i n this citse. :ngages in intense interview with the expert o r 2) the knowledge engineer becomes a n s x p e r l iim/herself, relying on introspection t o articulate .he requisite knowledge. The knowledge gineer then encodes the knowledge into a I;m guage of choice, encoding facts into OHJECT- ATTRIBUTE-VALLIE triples and infercnces into IF-THEN rules. When run, thc inferencc en- gines grind away at these knowledge structure\ working either with facts (e.g. the features that the particular case under diagnosis exhibits) j i w . ward to the goal, or working with the goal ( e . g one of several known diagnosis categories) huc X - ward to the facts. triples and forward and backward \card1 strategies are a small subset of the knowledge structures and search strategies human expert\ have. Expertise is primarily a skill o f r e c o g n - tion, of ‘seeing’ old patterns in thc new ptoblern. Chess experts, for example, have thc* same limited abilities as novices t o hold intorination for analysis and look ahead only 3 limited num- ber of steps. They excel because they h;rvc hundreds of thousands of chess configuration\ i n their memories, and can quickly encode the cur- rent situation into constellations o f previously - seen chess -patterns. The choice of candidate ‘good moves’ for the expert is thus restricted 10 ii small set of known good movcs that fit the pa- terns, whereas the novice has no such expert p a l - tern knowledge to filter out bad c;indidatil\ 131, 141 and [ S l . There is also evidence to suggest that c t p c r t a see more richly encoded patterns than novices\ do. They have organized the concepts in thcir knowledge bases with much more depth and with many more central associations t h i t n novices. For example, expert Algol prograiniiicw had much more structure i n the rclation\hrp\ among concepts held i n nieinory thaii novice\ did. And, experts’ organizations wcre hip.11ly similar, whereas novices’ structures weie s a t . tered, based on a variety of irrelevant prior i t s sociations 161. Not only d o experts have in1orrri;rtiori o r - ganized in a highly structured manner. hut they use a variety of kitids of knowledge structure\. Some things are stored in simple /isr.s. e.9. tht. months of the year and the days o f thc week. Some information fits a stored f u l d c better. i n l o r mation such as calendar appointments and the pe-- riodic table. Some information is stored ;I\ ;I tlow diagram, such as a ducisiorr ti’(’(’, for c x - Simple OBJECT-ATTRIBI.ITE- V A 1.1 1 E I52 Expert Sy\tem\, Augu\t 1987 V o l 4 , N o . 3 ample, representing the routing of telephone mes- sages to the people who can handle them. There is information stored in hierarchies of relation- ships, nested categories or clusters, such as animal taxonomies. Networks store richly con- nected language associations. Such information as room arrangements or maps may be stored as physical space. And, some information may be stored about a device’s internal components and how they are causally related as a physical model, commonly referred to as a mental model. Experts may hold what they know about objects in a myriad of different representations, each suitable for a particular kind of reasoning or retrieval. Not only do experts see problem situations dif- ferently, they may also search differently through intermediate problem states. Recent work in medical reasoning, for example, has shown that experts primarily work forward through the problem (beginning with a ‘good’ representation), and that novices work backward from the goal. Furthermore, from work on plan- ning, we see that experts also plan ahead loosely, filling in details when the situation dictates, and that they move back and forth from abstract to detailed entities when evaluating tactics and strategies [7J and [8]. Typically the knowledge engineer has only OB- JECT-ATTRIBUTE-VALUE and IF-THEN rules to use to encode what the expert knows. If in acquiring the expert’s knowledge, the knowledge engineer focuses only on knowledge that can be easily encoded in these forms, then significant expertise will be missed. The process of translating expert knowledge to this form will benefit greatly from a knowledge of 1) the fact that many different expert structures and proces- ses are possible, and 2) which tools are ap- propriate to uncovering these various structures. The purpose of this paper is to convey to the knowledge engineer the myriad of methods used in cognilive science research for revealing expert knowledge structures and processes. Some of these methods may be useful in the developing expert systems. The ultimate goal, however, is to alert the knowledge engineer to the fact that simple, intense, time-consuming interviews of experts leading to the coding of OBJECT-AT- TRIBUTE-VALUE triples and IF-THEN rules may be both misleading and costly as well as dif- ficult. Experts have rich structures and reasoning abilities. Truly representing the expert’s exper- tise depends on knowledge of these structures and abilities. 2. Methods for knowledge acquisition There are two classes of methods for revealing what experts know. ‘Direct methods’ ask the ex- pert to report on knowledge he/she can directly articulate. This set of methods includes inter- views, questionnaires, simple observation, think- ing-out-loud protocols, interruption analysis, drawing closed curves, and inferential flow analysis. In contrast, ‘indirect methods’ d o not rely on the expert’s abilities to articulate the in- formation that is used; they collect other be- haviors, such as recall or scaling responses from which the analyst can make inferences about what the expert must have known in order to respond the way he/she did. ‘Indirect methods’ include multi-dimensional scaling, hierarchical clustering, general weighted networks, ordered trees from recall, and repertory grid analysis. 2.1 Direct methods 2.1.1 Interviews. Interviews are the most com- mon method for eliciting knowledge from the ex- pert. In conversation, the expert reveals the ob- jects he/she thinks about, how they are related or organized, and the processes helshe goes through in making a judgment, solving a problem, or designing a solution. There are simple guidelines that can be followed to make interviewing effi- cient. 1 . Enlist the expert’s cooperation. Interviewing an unwilling expert dooms the project to failure. There is a number of reasons why an expert may not be cooperative. First, the expert may know that he/she performs the task in simple, intuitive ways that, if revealed, would reduce the esteem others hold for himher. Second, the expert may not know how he/she performs that task and may be reluctant to express the uncertainty, thinking that experts are expected to be rational and ar- ticulate. And third, the expert may believe that if h i s h e r expertise can be captured in a computer, hisfher job will be eliminated. On the other hand, the expert may be flattered to think that the com- pany is willing to invest the time and money to clone the expertise, to allow many more problems to be solved with the knowledge this one person has gained. 2. Ask free-form questions at the start, narrow- ing in spec$icity as the interview process progresses. The goal at first is to discover the vocabulary the expert uses and to allow h i m h e r to begin to articulate the inferences drawn and ~~~ ~ ~ Expert Systems, August 1987. Vol. 4 , N o . 3 . 153 the relationships seen. Asking, "How do you do your task?" can be followed with, "Recall the last case ..." The interviewer should note the order in which the expert addresses topics, the relative importance of the way evidence is weighed. 3. Do not impose your own understanding on the expert. Allow the expert to talk, even if the cur- rent topic seems tangential to the main purpose. Do not interrupt. Ask about what you do not un- derstand, but do not impose your own, naive bias on what the expert is saying. 4 . Limit the sessions t o coherent tasks, recogniz- ing fatigue and attentiond limits. The inter- viewer should be aware of the effort that goes into answering questions about expertise, and make sessions tolerably long. However, breaking in the middle of a thought or problem is un- natural and disturbing. It is almost impossible to 'take up where we left off' if major questions are not settled on a problem or task at the end of the previous session. The interview questions should center on the expert's knowledge of objects, relationships, and inferences. Good questions to ask might be: "What kinds of things do you like to know "What facts or hypotheses do you try to estab- "What are the factors that influence how you "What type of values can this object have? "Does this factor depend on other factors? If "Is this factor needed for solving all problems about when you begin to ponder the problem?" lish when thinking about a problem?" reason about a problem?" What range of values is permissible?" so, which ones?" in the domain or for just some?"' It is probably best to begin knowledge acquisi- tion with an interview, to identify a set of objects and their relationships (noting especially whether the expert is thinking of these objects in special relationships like lists, tables, physical spaces, etc). However, the specificity and com- pleteness of what can be obtained soon reaches a limit in the free form style. To alleviate this problem, often the interviewer can mix in ques- tions of other styles, such as focusing on a par- ticular case in detail, called the 'critical incident technique'. Focusing on a case elicits particular descriptions, rules and objects, which can be ex- From [9] 2mined for their generality i n later sessions By isking for 'symptoms' and 'characteri\tics' one Aicits features, while asking tor 'evidcncc' :licits inferences. 2.1.2 Questionnaires. Interviews have a distinct advantage in that they can elicit unforeseen infor mation. Interviews are free form; experts L ~ I I generate information in the order they wi\h i n the detail they wish. Experts are in control Inter- views, however, are very time-conwming. Ques- tionnaires, on the other hand, have the advantape of being a very efficient way to gather infornia- tion. Furthermore, the expert can fill O U I que+ tionnaires in a leisurely and relaxed atmospherc Questionnaires can be particularly useful i n d i \ - covering the objects of the domain, in uncover ing relationships, and perhaps in determining un- certainties, if the expert system attaches uncer- tainty to its conclusions. The questionnaires suggested for ube here drc not the kind used in survey research. Instead, they consist of cards or pieces of paper on which are printed some standard, but open -ended ques- tions. Figures 1 and 2 illustrate cards uwd t o elicit variables (in Figure 1 the object 'sale\' is defined) and relationships (in Figure 2 the relationship between 'sales', 'quota', and 'ha\e' I-. Figure 1. Questionnaire cardfor variahk elicitation [ 9 ) Figure 2. Questionnaire cardfor relationship elicitation [91 154 Expert Systems, August 1987. Vol. 4 , N o . 3. is drawn.) Figure 3 illustrates the use of questionnaires to elicit uricertainties about particular inferences an expert has reported. Questionnaires are par- ticularly appropriate for eliciting uncertainty in- formation, since normal verbal responses from people are not very reliable in revealing this kind ~ of information. People are not good at estimating probabilities; they overestimate low ones and un- derestimate high ones [ 101. Consequently, simp- ly asking for probabilities in an interview is not effective. Eliciting probability estimates using pre- formatted response scales can yield much more accurate estimates. There are two preferred formats: 1) the bar on which the expert indicates a point to reflect uncertainty, and 2) a five point verbal scale on which the expert marks the work most closely associated with hisher certainty. The five-point scale is one taken from Meister’s compilation of verbal scales possessing equal spacing and reliability of measurement 11 1 I. Figure 3. Response scale for uncertainty elicitation [9] 2.1.3 Observation of the taskperformunce. Often the best way to discover how an expert makes a judgment, diagnosis, or design decision is to watch the expert work at a real problem. In this situation, the knowledge engineer has several ways of discovering the objects, relationships, and inferences that the expert is using. The first decision that must be made is how to record the expert’s performance. One possibility is simply to watch, take notes, and try to follow the ex- pert’s thinking process on the fly. A second pos- sibility is to videotape the process for later review with the expert. In choosing between these methods, remember that the first method suffers from time pressure and observer bias, while the second relies on the expert’s less than perfect ability to recall the reasons underlying hisher performance. 2.1.4 Protocol analysis. A close cousin of simple observation is protocol analysis. Like observa- tion, above, the expert engages in normal task be- havior with particular typical problems. In addi- tion to video-recording the session and annotat- ing the behaviors after the fact, the knowledge engineer asks the expert to ‘think out loud’ while performing the task. The expert is to answer the questions, “What are your goals?“ “What are your methods?“ “What are you seeing?“ In- ference is then drawn from a transcript of this session about the objects, their relationships, and the inferences the expert was drawing moment by moment. The advantage of this method over the an- notated silent task performance of the observa- tion method is that there is no delay between the act of thinking of something and reporting it. But protocol analysis is not appropriate for all kinds of tasks. Ericsson and Simon [I21 carefully detail the kinds of tasks for which thinking-out- loud protocol might be acceptable, useful kind of data. To summarize, those tasks for which ver- balization is a natural part of thinking are those for which we can take thinking-out-loud as data. That is, if verbal information is produced while someone makes inferences to himherself, or in identifying salient features of the objects in the situation, then the information from the protocols is acceptable data. However, there are other kinds of tasks, those for which some idiosyncratic language is used in the process (e.g. in composing music, composers often have a special language for the parts of the piece they are writing or the section of the style they are in- stantiating currently), for which the process of thinking-out-loud and explaining might be dis- torted or even wrong. And, of course, there are tasks for which there is no natural verbalization; perceptual-motor tasks are examples of these. Verbalization of perceptual-motor tasks makes someone attend to aspects not normally attended to, and the attention required to report on the process usurps resources normally devoted to the task itself. Once obtained, protocols must be analyzed. The goal in obtaining a protocol lies in identify- ing the kinds of objects the expert sees, the relationship that exists between the protocols, and the kinds of inferences drawn from the relationships seen. For example, in the protocol in Figure 4, the problem solver is trying to find a solution to the ‘cryptarithmetic problem’ DONALD+GERALD=ROBERT. In cryptarith- metic, each letter can be mapped into one digit such that the arithmetic operations apply. The Expert Systems, August 1987. Vol. 4 , N o . 3 . 155 IS6 problem solver has been given the fact that D=5. In the protocol, the analyst looks for the change in focus of attention: it moves from working for- wards ("each D is 5; therefore, T is zero ... Now do I have any other Ts? No, but I have another D"). Later it reveals some working backwards (...Since R is going to be an odd number and D is 5 , G has to be an even number). One sees a shift in goals and subgoals, the kinds of things the problem solver is paying attention to, and the kinds of inferences made [ 131. Figure 4. Cr~ptarithmetic psoblem [I31 2.1.5 Iiiteiwption analysis. One way of preserv- ing the natural thought process of the problem solver is to let himher proceed without thinking aloud. But, when the process gets to a point where the observer can no longer understand the expert's thought processes, the observer inter- wpts. At that point, the observer asks in detail why the expert did what he/she did, trying t o cap- :ure at the moment the focus of attention and the kinds of inferences drawn for the features noticed [14]. This process can be very instructive about the process observed, but, or course, oncc a process has been interrupted, there is very littlc chance that it can be restarted. This procedure is likely to give most of its value after the expert system has been coded in its prototype stage, and the expert's performance is being compi4recl to that of the system. 2.1.6 Drawing closed curves. The previcius methods attempt to reveal the contents ot the thought processes during the solution of existing problems. They highlight the vocabulary the ex- pert uses to identify the objects and their relation- ships; they highlight the kinds of inferences drawn. These methods are free of assumptions about the form of the relationships among the items, be they lists or tables or networks o r physi- cal space. In contrast, the method of drawing closed curves is a specialized method for indicat- ing the relationships among those ob.jects t h a t can be assumed to be encoded i n a p f t j s i c i t f space representation. In the method of drawing closed curves, the ex- pert is asked to indicate which of a collection of physical ob- jects 'go together', to draw the related object4 in a closed curve. This technique is applicable to any spatial representation, such as a typeset formula, an x-ray or CAT scan, or a position on a game board. For example, a Go expert drew closed ciirveb around the stones on a position in the chesh-like game of Go [ 5 ] . Figure S illustrates several aspects of his responses. Four positions are dis- played, A-D. Inside each stone is a number which represents the ordinal position in which that stone was placed on the board in a Iecall task. Note that the order matches the closed cur.- ves to a remarkable degree; all stones of a chunk are recalled before moving on to another chunk. Furthermore, on the right side of the Figure are shown three successive recall trials of the same position. As above, the numbers indicate the order in which the stones were placed in recall trial. It is noteworthy that groups of stones, indi- cated by closed curves on a separate occasion, are recalled consistently together before other groups are recalled. This regularity of behavior suggests the validity and reliability of the inlor- mation contained in the originally drawn closed Expert Systems, August 1987. Vol. 4 , No. 3 curves. 2.1.7 Inferential flow analysis. A variant on the interview is the method called inferential flow analysis. In this method, answers to particular questions about causal relations are used to build up a causal network among concepts o r objects in the domain of expertise. Salter [IS] used this technique to uncover laymen's models of economics. This technique begins with a list of some of the key objects in the domain of expertise. In Sal- ter's case, this list contained such items as busi- ness borrowing, personal savings rate, produc- tivity, etc. Salter then asked an interviewee a series of pointed questions about the relationship between two of the objects. For example, he would ask, "What is the relationship between savings rate and inflation?" Answers revealed the linkages among items between these two key objects and the direction of the relationships. For example, the person might respond, "If inflation goes up, savings rates will g o down, because savings interest rates are lower than the amount one can save by buying now instead of later." These two items are then linked in an inverse relationship, in which purchasing is an interven- ing variable. Responses to a set of questions should uncover some consistency in the relationships between in- tervening variables. Each time an item is men- tioned i n an answer it is linked with the other items in the answer, with the links labelled posi- tive or negative as indicated. Linked items are joined into one all-inclusive network of rela- tions. At the first mention of a link, a standard weight is given to the link, e.g. 0.50. With each succeeding mention the link is raised in strength, some proportion of the remaining strength be- tween the current value and 1.00. The resulting relations are displayed as a network, such as the one displayed in Figure 6 overleaf. Although this technique appears somewhat ad hoc, the resulting networks have been shown to be both stable and consistent with other sets of behaviors [ 151. This technique is simple to apply and powerful as a tool for displaying to the ex- pert aspects of the expertise heishe has un- covered to that point. This display can be used ef- fectively as a stimulus for further interviews. 2 . 2 Indirpct methods All of the previously described methods ask the expert directly what heishe knows. They rely on the availability of the information t o both intro- spection and articulation. Of course, it is not al- Figure 5. Closed curves drawn by Go expert [S] ways the case that the expert has access t o the details of h i s h e r knowledge or mental process- ing. In fact, it is not uncommon for experts to perceive complex relationships or come to sound conclusions without knowing exactly how they did it. In these cases, indirect knowledge elicita- tion methods are required. In all the following methods, experts are not asked to express their knowledge directly. In- stead, they are given a variety of other tasks, e.g. to rate how similar these two objects are, o r t o recall all these objects several times from several starting points. From the results, the analyst then Expert Systems, August 1987. Vol. 4 , No. 3 . 157 Figure 6. Inferentia1,flow network infers underlying structure among the objects rated or recalled. All the indirect methods dis- cussed here have been validated in experimental studies that have convincingly demonstrated their psychological validity. These different techniques make different as- sumpfions about the form of the underlying repre- sentation, whether it is physical space, lists, net- Figure 7,. Similarity judgment matrix works, or tables, etc. It is important to use only those methods for which the assumptions fit the analysts’ best guess as to what form the expert’s underlying representation is in. This information can be gleaned from initial interviews with the expert as well as from careful questioning and noting of object names and/or notations the ex- pert makes. 2.2.1 Multidimensional scaling. Multidinm- sional scaling (MDS) is a technique that should be used only on data that are assumed to have come from stored representations of p h y . s i c ~ . i / 1 2 - dimensional space [ 161. The subject provides similarity judgments on all pairs of objects o r concepts in the domain of inquiry. These judg- ments are assumed to be symmetric and graded; i.e. A is as similar to B as B is to A, ant1 the similarities are assumed to take on a variety of continuous values, not just 0 or 1 . The scaling technique produces a layout of the items in space. All objects of a particular target domain are paired with all other objects in a set of queries to the expert: ”How similar are A and B?” These similarity judgments are arrayed in a half-matrix such as that shown in Figure 7. The matrix in Figure 7 is part of a matrix that com- pares pairs of common farm and wild animals. This matrix is then the input to an analysis program which searches for the best placement of these objects in space of user-specified dimen- sion. Each dimensional solution has a ‘stress’ as.- sociated with it, a measure of the deviation from a perfect fit. The analyst looks for solutions with low ‘stress’. Those using fewer dimensions are then plotted. (For higher dimensions, some of the more illuminating two dimensional projections can be drawn). The analyst then examines the plots to judge the ‘best’ placement of the axes and a plausible labelling for them. In Figure 8, the solution to a fuller similarity matrix of the form in Figure 7, the two axes might be recog- nized as ‘size’ (the abscissa) and ‘ferocity’ (the ordinate). I 158 Expert Systems, August 1987. Vol. 4, No. 3 . Figure 8. Multidimensional sculing solution in two dimensions This technique is good for producing a diagram that the expert may then inspect and describe in more detail. It can reveal interesting clusters of objects, neighbor relations, and outliers. One dif- ficulty with this technique, however, is the tedium of collecting the required pair-wise similarity judgments; for n objects, n(n-1)/2 judg- ments are required, a number that quickly grows to the hundreds and thousands with more than a few objects. Furthermore, it is difficult for the analyst to find the dimensionality with the best ‘stress’ value, and then perceive the best place- ment of the axes and the axes’ names. Using the technique is straightforward; interpreting the results is not. 2.2.2 Johnson hierarchical clustering. Like MDS, hierarchical clustering begins with a half- matrix of similarity judgments. The assumptions for this technique, however, are in direct con- tradiction to those for multidimensional scaling. Whereas multidimensional scaling assumes sym- metric distances and graded properties, hierarchi- cal clustering assumes merely that an item is or is not a member of a cluster. Judgments are as- sumed to be a function of the number of nested clusters two items have in common, or the ‘height’ at which two items become members of the same superordinate category. Items cannot at once satisfy assumptions for both multidimen- sional scaling and hierarchical clustering [ 171. Johnson hierarchical clustering is a fairly simple, straightforward algorithm that starts with the half-matrix of distances and ends with a hierarchical representation of the items. In broad strokes, pairs of items that are the closest in the matrix are joined to a single cluster, a new matrix drawn with this cluster serving as a new ‘item’. This new matrix is examined again for that pair of ‘items’ that is closest together. These are joined as if the next new ‘item’, and a new matrix drawn. Each time a new matrix is drawn, interitem distances among unclustered items are copied from the original matrix to the new; dis- tances between items and clusters are calculated as either the minimum distance of all cluster items to the item, the maximum, or the average. For example, using the matrix in Figure 7, the items that are the closest are COW-SHEEP- GOAT. The value in the re-written matrix for the distance between this cluster and PIG, for ex- ample, using the minimum joining algorithm, as- signs ‘2’ to the distance, the minimum of 2, 3 and 2 (for PIG-GOAT, PIG-COW, and PIG- SHEEP, respectively). Figure 9 shows the full rewritten matrix with the COW-SHEEP-GOAT cluster now serving as an ‘item’. In this matrix, PIG joins COW- SHEEP-GOAT (at 2) as does DOG to RABBIT. This matrix is rewritten into Figure 10. In Figure 10, HORSE joins the ((COW-GOAT-SHEEP)- PIG) cluster at 3, and the two clusters join at 6. The completed hierarchy is shown in Figure 1 1 ~ cow Sheep- Goat Pig Horse Dog Rabbit cov- Sheep- Goat Pig Horse Dog Rabbit Figure 9. Similarity mutrix of Fig. 7 with Cow-Sheep-Goat cluster An advantage of hierarchical clustering is that it can be done with paper and pencil. Unfor- tunately, it begins with a distance matrix that is just as tedious to collect as that for multi- dimen- sional scaling. Furthermore, without some theoretical justification for choosing a particular joining algorithm (the minimum, maximum, or average) one must choose arbitrarily; unfor- tunately, different algorithms can produce remarkably different hierarchies. In this sense, the analysis is somewhat subjective. 1 Unfortunately, some people routinely do a multi-dimensional scaling solution in two dimensions (arbitrari- ly) and then indicate the nested curves found in Johnson hierarchical clustering by drawing closed curves around chunk elements. This cannot be done if one adheres to the underlying assumptions of both techniques. Expert Systems, August 1987. Vol. 4 , N o . 3 . 159 CGS-Pig Horse Dog-hbbit CSG-Pig Horse Dog-Ubbit Figure 10. Similarity matrix of F i g . 9 afterfurther clusterin2 Figure 11. Final cluster hierarchy f o r animal exuniple 2.2.3 General weighted networks. Like the preceding two techniques the expert gives sym- metric distance judgments on all possible pairs of objects. These distances are assumed to arise from the expert traversing a network of associa- tions, a network in which there is a single primary path between every two items, and, for some of them, a differently encoded, secondary path between them as well. A recent investiga- tion using networks is [ 191. From the distance matrix, a minimal connected network (MCN) is first formed. This network is formed by connecting the most closely linked items, such as COW-SHEEP in Figure 7.’ In Figure 12. M C N and MEN f o r animals exump!e Figure 12, the solid lines show the resulting net- work for these items. Secondly, additional links are added to this tree and the resulting structure called the mrnim,il :laborated network (MEN). Here, we add a l i n k d and only if it is shorter than the links currently in the network between those two node\ ‘The dashed lines in Figure 12 are those atfditioniil links appropriate to the MEN. These two struc- tures are then examined for 1) dominating concepts - those that have 3 large number of connections to many other nodes, and 2) members of cycles - those items that are fully linked into circles. In Figure 12, SHEEP is a dominating concept, the one with the most primary links, HORSE i s somewhat less dominating, because although it has many links, most are from the elaborated net- work. Figures 13 and 14 show the results 0 1 an ex- ploration of the MCN and MEN for expert and novice pilots, rating a set of terms having to do with ‘split plane concepts’ [20]. Figure 13 I \ - lustrates the network for an expert; Figure 14 shows the network for the same concepts held by a novice. Several things were noted in the study: Experts’ structures were simpler than stu- dents’. Elaborated links connected integrated larger conceptual structures. Experts could easily identify link relations, using such terms as ‘Affects’, ‘Is-a’. ‘Desirable’, ‘Acceptable’, etc. The fact that the experts were so clearly dif- ferent from the novices suggests that this techni- que can reveal significant aspects of expertise, aspects that clearly should be encoded into an ex- pert system. 2.2.4 Ordered trees f i o m recall. Ordered trees come from work by Reitman and Kueter ( 2 I 1 in their exploration of how memory organizations differ among experts and novices in a particular domain. Unlike the indirect methods described above, ordered trees begin not with a distance matrix but with recall trials. The technique as- sumes that objects belong to a cluster o r not. similar to the assumption of hierarchical cluster- ’ SHEEP is taken as the most central item i n thi\ ’tied’ cluster because it is, on average, closer to all other items than either GOAT or COW, I60 Expert Systems, August 1987. Vol. 4, No. 3 Figure 13. M C N and MEN f o r expertfighter pilots 1201 Figure 14. MCN and MEN@ novice jighter pilots 1201 Expert Systems, August 1987. Vol. 4 , No. 3 . ing. Unlike hierarchical clustering, however, this technique is built upon a model of how the data are produced by the subject: it assumes that people recall all items from a stored cluster before recalling items from another cluster. This assumption builds on data from people recalling from known (learned) organizations. Regularities found over a set of recall orders are assumed to reflect organization in memory. Figure 15 illustrates four orders of the seven animal names used in previous examples. The ex- pert/subject is asked to recall object names ten to twenty times; to encourage variety, on some of the trials he/she is told which item to begin with. These recall trials are then examined for regularities. All sets of items that are recalled together are identified as chunks, the chunks written into a lattice (ordered inclusion covering relationship). This lattice is then re-drawn into an ordered tree structure, such as that in Figure 16, where arrows (either uni-directional or bi- directional) are drawn over the chunk elements that were recalled consistently in a particular order (or, in the case of a bi-directional, one order and its reverse). This analysis can be done by hand, but because it is tedious and open to perceptual error, a computer analysis is best. A program can also perform certain advanced analyses in addition, such as calculating an index of organization and looking for ‘outlier’ trials, without which the tree structure reveals sig- nificantly more structure. Figure 15. Four animal recall trials with low-level chunks marked This technique has been used in a variety of studies of expert-novice differences. In [6], for Figure 16. tfigher-order animal rec ull( h n X ! andfinal order tree example, novice, intermediate, and expert Algol- W programmers were asked to recall Algol keywords many times from many starting point5 while their performance orders were recorded Experts differed remarkably from the n o v t Experts showed much more organiiation. ,rnd the similarity among the expert 5tructures wa\ far greater than that among the novices In 12 1 1 , furthermore, the pauses between recall\ of wc- cessive items was accounted for by the number of chunk boundaries crossed in the inferred memory organization. There has been a vmety of studies that have used this technique t o reveal organization in different domains of e x p e r t i ~ , all showing a convergence among expert5 in their organization of the concepts in memory Figures 17 and 18 show the ordered tree5 t o r one expert and one novice, respectively, in the study of Algol knowledge. The expert clearly u n derstands the function of these 5pecial word\, grouping the words concerning loops (WHILE- DO, FOR-STEP), the logic item5 (AND-OK, TRUE-FALSE), and the string representation items (STRING-BITS, LONG-SHORT, REAL,) The novice, on the other hand, grouped the short words (AND-OR-OF-FOR), grouped the condi- tional words (THEN-WHILE, ELSE-IF) in an order not standard to programming, dnd clustered words into a small scenario connected with do with “long and short bits ot stnng” (BITS-STRING-LONG-SHORT). Clow ex- amination of the resulting ordered trees can reveal aspects of what the expert ‘sees’ in a \itua- 162 Expert Systems, August 1987. Vol. 4 , No. 3 . Figure 17. Ordered tree for expert Algol programmer [ 6 ] tion in his/her domain of expertise. 2.2.5 Repertory grid analysis. This technique, the last lo be presented in this review, is the most complete. It includes an initial dialog with the ex- pert, a rating session, and analyses that both cluster the objects and the dimensions on which the iterris were rated. Essentially, it is a free- form recall and rating session in which the analyst makes inferences about the relationships among objects and the relatedness of the dimen- sions the expert pays attention to. Repertory grid analysis is a technique whose origins are in personal construct theory in clini- cal psychology [22]. Used in the clinical setting, it was intended to reveal to the patients the kind of attributes they normally or abnormally attend to in their emotional lives. Boose [231 and [24] has adapted it for the explicit development of rules in expert systems. The initial session begins with an open inter- view of the expert, asking him/her to name some objects in the domain of expertise. After a small set is generated, the analyst picks three of these objects and asks, “What trait distinguishes‘ any two of these objects from the third?” The expert names a dimension in whatever vernacular is natural. He/she then indicates which are ‘high’ on this trait, and which are ‘low’. The analyst records the dimension and assigns a scale value (e.g. 1-3) to the three objects. The analyst then picks three other objects and asks the same ques- tion about what trait distinguishes two of these from the third. This process of asking for salient dimensions continues with a significant number of triples, enough so that the analyst is satisfied that he/she has uncovered the major dimensions of similarity and dissimilarity. Having collected a ‘grid’ with objects at the top and dimensions across the left border, the analyst then asks the expert to fill in all the missing values. That is, all objects are then to be rated on all dimensions. Figure 19 is an example grid that rates quality of students on nine different elicited dimensions, where each student is rated on a three point scale for each of the dimensions that Figure 18. Ordered tree f o r novice Algol programmer [6] Expert Systems, August 1987. Vol. 4, No. 3 . 163 Figure 19. Rating gridfor seven students l2.51 the expert generated (taken from [25]). Two further analyses use this grid: clustering of its objects (in this case, students), and cluster- ing of its dimensions. Johnson hierarchical clustering (described above) is used, so a dis- tance matrix is required. For clustering objects, the distance metric is straightforward: for each pair of objects, count the absolute difference be- tween the scores each object was given on each Figure 20. Between-object (student) distance matrix Figure 21. Student hierarcAy jimension. Thus, for objects E l and E2 in our e x - ample, the distance is: 2+1+2+2+2+2+2+2+2 = 17 Figure 20 illustrates the distance matrix that arises from this calculation on the ratings f o r Figure 19. The resulting hierarchy, using the minimum joining method, is shown in Figure 2 1 . This display should be given to the expert f o r fur- ther comment and analysis. The second analysis done on the original grid examines the similarity of the dimensions. Here the definition of distances is not so straightfor- ward. Since no judgment was ever made as trr which end of a scale received a ‘3’ and which ;I ‘ l ’ , there may be cases where very \iruilar dimensions are highly correlated but were as- signed opposite scale values. Therefore, part of the calculation of the distance matrix involves ‘flipping’ appropriate dimensions. This is tiont: in the following manner: first, the full distance matrix is calculated as the actual (not absolute) difference between the ratings of all objects on two dimensions. Above the diagonal are the dis- tances between two dimensions the way they are written; below the diagonal, the second dimen- sion (the ‘row’ dimension) is reverse in scale. That is, comparing C1 and C 2 across all ob.iect\ produces a score of (CI-C2) =2+1+2+1+2+1+1= 10 Next, the C2 values would be ‘flipped’, e a c h scale value 3 translated as a 1, each I a\ a 3 Cnn- sequently, comparing C1 with C2’ (llipped) I 64 Expert Systems, August 1987. Vol. 4 , No. 3 . would result in: Figure 22 is the full matrix of the distances among all pairs and ‘flipped’ pairs of dimensions from the grid in Figure 19. since Johnson hierar- chical clustering cannot use assymetric distan- ces, a symmetric half-matrix is needed. To trans- late this into a half-matrix, we select the highest of the two relevant cells, (Cl-C2) and (Cl-C2’). This resulting half-matrix is shown in Figure 23; the hierarchical clustering solution is shown in Figure 24. Figure 22. Asymmetric hetween-concept(rating scale) distances Boose’s expertise transfer system further ex- amines these clustered dimensions to find those that are highly correlated, those that imply one another, and those that are super-ordinate to each other. Rules are obtained in a second interview with rhe expert, wherein the traits or dimensions elicited in the first phase are reviewed and named. ”he program then generates production rules. Some rules reflect the correlation between dimensions, some combine dimensional values to predict an end category. Boose [24] reports that several hundred prototype systems have been created using this technique. It seems to be particularly applicable to classification type problems, where features of a new object (case) are observed and the object sorted into one of a known set of categories. The system has several advantages: it produces similarity matrices with a procedure that is much Figure 23. Symmetric hetween-concept(ratinR scale) distances Figure 24. Rating scale hierarchy less tedious than directly rating the similarity of all pairs. Furthermore, it has been used to com- bine the expertise from two different experts in the same domain, and it has been used to com- bine two experts in different aspects of the same general domain. Individual repertory grids can also be used as a basis for discussing or negotiat- ing disagreements among experts. In this situa- tion, individual grids are generated by each ex- pert; verbal discussion ensues from both experts viewing both grids and clustering solutions. Similarly, grids could be used as an aid in trans- fering expertise from an expert tutor to a novice. - Expert Systems, August 1987. Vo[. 4, No. 3. - I65 166 The novice grid and the expert grid could be compared, and mismatches determine the focus of additional instruction. 3. General discussion Experts have stored rich representations of facts, objects and their attributes, as well as a set of in- ference rules that connect constellations of facts for use in problem-solving situations. The methods collected in this review differentially il- luminate these various kinds of knowledge. Table 1 categorizes the twelve methods accord- ing to whether they illuminate objects, their relationships, or inference rules. Clearly, in the open, free-form format of the direct techniques, with the exception of drawing closed curves and inferential flow analysis, the knowledge engineer has a chance of finding any kind of information. These techniques are not specifically designed to elicit one particular kind of information over another. Drawing closed curves explicitly il- luminates the relationships among objects in the problem space; inferential flow analysis, as indi- cated in its name, displays the inference chains experts may use to reach conclusions. All of the direct techniques, however, suffer from the fact that experts cannot always say what they know or how they solve a particular problem. If knowledge engineers confine their knowledge acquisition techniques to these, they run the risk of excluding important kinds of in- formation from their expert systems. The indirect techniques, however, are more limited in what they can reveal. All of the in- direct techniques illuminate particular aspects of the relationships among the objects in the domain of expertise. The repertory grid analysis can also produce inferences based on the correla- tions among attributes of objects. The indirect techniques, however, involve as- sumptions about the underlyingform of the repre- sentation of objects and their relations. Table 2 aligns the indirect techniques with each of the forms listed in the introduction: lists, tables, categorical hierarchies, inferential flows (decision trees), networks, physical space, and physical models. The two direct techniques, drawing closed curves and inferential flow analysis, are included here because they similar- ly make assumptions about the format of the un- derlying representation. For each technique, the primary assumed form is indicated with an ‘X’, and those forms that can additionally be revealed are marked with a ‘+’. An additional category of form, ‘undefined OBJECT-ATTRIBUTE- VALUE’, is included to account for the primary iata elicited in repertory grid analy\is. All of these techniques add richness to the ba\e i f information about expertise that can be :licited from interviews. Care should be taken, iowever, with the amount of faith that I \ placed In the validity of the displays of knowledge each >f these techniques produces. For example. protocol analysis, if applied to a task that I \ not normally verbalized, may distort the problem solving process and reveal what the expert t h r n k \ Table 1. Kinds of information methods cur1 r e v d Interviews X X X Questionnaires X X X Obscrvriion X X X Protocol Anrlyrir X X X Interruption X X X Closed Curves Inferential Flow MDS Hierarchies GWN X X X X X Ordered Trec X Repertory Grid X X X Table 2. Kinds ojstructures the method,s can ~ h o u Lists Tables Hierarchies Flow Networks Physicrl Physica Space Model MDS + + X -Y X OWN + + + X + Ordered T r m + X Rep. Grid x + Closed cwu X + Inter. Flow X + + he/she should report rather than what he/\he nor- mally uses. The indirect technique\ are more resistant to these distortions because they dtr not suggest to the expert/subject the ‘right’ way to respond. Since indirect techniques are based on ah\ump- tions and underlying theories, they can be abu\ed Expert Systems, August 1987. V o l . 4 , No. 3 . to the extent that their basic assumptions are not met by the data. Multidimensional scaling, for example, can be applied to situations for which there is no reason to assume an underlying n- dimensional space. Johnson hierarchical cluster- ing is too often used to indicate the ‘clusters’ of points on a multidimensional scaling solution. This co-representation is a direct violation of the underlying assumptions; the data cannot simul- taneously satisfy both sets of assumptions. General weighted networks assume an underly- ing network; ordered trees assume a tree struc- ture in which some clusters have prescribed or- ders of recall imposed on them. Repertory grid analysis assumes that objects are stored with their values on bi-polar dimensions, and con- nected in networks of dimensional values. Only one technique makes explicit assumptions about how the expert produces the data. Ordered trees assume that items are stored in nested clusters and that all items of a cluster are recalled before all items of any other cluster. The remainder of the techniques rely loosely on the ability of the subject/expert to make a single valued ‘similarity’ judgment from whatever the stored form is. Just as a statistician makes judgments about the suitability of a data set to the assumptions of a proposed analysis, the knowledge engineer must make judgments of the suitability of a method for knowledge elicitation to the kinds of knowledge the expert is assumed to possess. There is a number of ways these techniques can be misapplied for scientific discovery of mental organizations. However, if used as exploratory analyses of the way specific experts may store and use knowledge, these techniques can bring a great deal of information to the knowledge en- gineer. Using these techniques, knowledge en- gineers can hope to uncover more of what ex- perts know than is currently accessible through interviewing or introspection. 4. References [ l ] P. Harmon and D. King, Expert Systems: Art$cial Intelligence in Business, John Wiley and Sons, New York, 1985. [2] D.A. Waterman, A Guide to Expert Sys- tems, Addison- Wesley, Reading, Mas- sachusetts, 1986. [3] W.G. Chase and H.A. Simon, ‘The mind’s eye in chess’, in W.G. Chase (ed.) Visual Information Processing, Academic Press, New York, 1973. [4] A.D. deGroot, Thought and choice in chess, Mouton Co., Paris, 1965. [5] J.S. Reitman, ‘Skilled perception in Go: Deducing memory structures from inter- response times’, Cognitive Psychology, 8, 1976, pp. 336-356. [6] K.B. McKeithen, J.S. Reitman, H.H.Rueter and S.C. Hirtle, ‘Knowledge organization and skill differences in com- puter programmers’, Cognitive Psych- [7] V.L.. Pate1 and G.J. Groen, ‘Knowledge based solution strategies in medical reasoning’, Cognitive Science, 10, 1 , 1986. [8] B. Hayes-Roth and F. Hayes-Roth, ‘A cognitive model of planning’, Cognitive Scitxce, 3, 1979, pp. 275-310. [9] C.W. Holsapple and A.G. Whinston, Manager’s Guide to Expert Systems, Dow-Jones-Irwin, New York, 1986. ology, 13,1981, pp. 307-325. [lo] C.H. Coombs, R.M. Dawes and A. Tversky, Mathematical Psychology: An Elementary Introduction, Prentice-Hall, Englewood Cliffs, New Jersey, 1970. [ l l ] D. Meister, Behavioral Analysis and Measurement Methods, John Wiley and Sons, New York, 1985. [12] K.A. Ericcson and H.A. Simon, Protocol analysis: Verbal reports as data, MIT Press, Cambridge, Massachusetts, 1984. [13] A. Newell and H.A. Simon, Human Problem Solving, Prentice-Hall, Englewood Cliffs, New Jersey, 1972. [I41 R. Schweickert, A.M. Burton, N.K.Taylor, E.N. Corlett, N.R. Shadbolt and A. Hedgecock, A comparison of knowledge elicitation techniques f o r ex- pert systems: A case study in lighting f o r industrial inspection, Technical Report, Department of Psychology, Purdue University, 1986. [ 151 W.J. Salter, ‘Tacit theories of economics’, Proceedings of the 5th Annual Con- ference of the Cognitive Science Society, Rochester, New York, 1983. [16] R.N. Shepard, A.K. Romney and S.B. Nerlove, Multidimensional Scaling: Theory and Applications in the Be- havioral Sciences, Volume I, Seminar Press, New York, 1972. Expert Systems, August 1987. Vol. 4, No. 3 . 167 [17] E.W. Holman, ‘The relation between hierarchical and Euclidean models for psychological distances’, Psychometrika, 37,4, 1972, pp. 417423. 11 81 S.C. Johnson, ‘Hierarchical clustering schemes’, Psychometriku, 32, 1967, pp. [19] R.W. Schvaneveldt, F.T. Durso and D.W. Dearholt, Pathfinder: Scaling with net- work structures, Memoranda in Computer and Cognitive Science, Report No. MCCS-85-9, Computer Research Laboratory, New Mexico State Univer- sity, 1985. [20] R.W. Schvaneveldt, M. Anderson, T.J. Breen, N.M. Cooke, T.E. Goldsmith, F.T. Durso, R.G. Tucker and J.C. DeMaio, Structures of memory for criticalflight in- formation, Air Force Human Resource Laboratory, Technical Report 81 -46, 1982. 241-254. About the authors [21] J.S. Reitman and H.H. Rueter, ‘OrganiLa- tion revealed by recall orders and con- firmed by pauses’, Cognitive Psvcho/o,qy, 12,4, 1980, pp. 554-581. 1221 G.A. Kelley, The Psychology oj l ’ p r x i n d Constructs, Norton, New York, 1955. [23] J.H. Boose, ‘A knowledge acquisition program for expert systems based on per- sonal construct psychology’, Intrrnutional Journal of Man Machine Studirs, 23, 1985, pp. 495-525. [24] J.H. Boose, Expertise Trun.sjer Jhr f h p t ~ r t System Design, Elsevier, New York, 19x6. [25] A. Hart, Knowledge Acquisition j i ) r Et-- pert Systems, McGraw-Hill, New York, 1986. Judith Reitman Olson Judith Olson is an Associate Professor of Computer and Infor- mation Systems at the Business School at the University of Michigan, and an Adjunct Associate Professor in the Michigan Psychology Department. She is the Director of the Human-Com- puter Interaction Laboratory, a collection of researchers from computer science, computer and information systems, organiza- tional behavior, industrial and operations engineering, technical communication, and psychology. Her work focuses on the design of interfaces for multi-media document preparation and mailing, the analysis of office work with the goal of suggestion computer- based support 01 automation, and knowledge acquisition for ex- pert systems. Henry H. Rueter Dr Rueter is a Senior Systems Analyst at Vector Research, Inc. in Ann Arbor, Michigan, and a Visiting Scholar at the Human- Computer Interaction Laboratory at the University of Michigan. His current research interests include training and expert sys- tems. One of the methods described in this paper, the Reitman- Rueter ordered tree analysis, was the subject of Dr Rueter’s PhD dissertation at Michigan. Prior to his current position, Dr Rueter received the MS degree in mathematics and the MA de- gree in psychology from Georgia State University. His under- graduate degree was a BS in chemistry from the University of Georgia 168 Expert Systems, August 1987. Vol. 4 , No. 3 . 
work_eitdkhvy5fbkrew5k5ay456a44	----	European Journal of Operational Research 70 (1993) 1-15 North-Holland Invited Review The use of mathematical programming with artificial intelligence and expert systems Richard D. McBride and Daniel E. O'Leary School of Business, University of Southern California, Los Angeles, CA 90089-1421, USA Abstract: Researchers have developed artificially intelligent (AI) and expert systems (ES) to assist in the formulation, solution and interpretation of generic mathematical programs (MP). In addition, re­ searchers also have built domain-specific systems either modeled around a mathematical program or which include a mathematical program module. In these systems, the specificity of the domain allows researchers to extend the interpretation or formulation beyond that available from the generic set of assumptions about mathematical programming. Further, researchers have begun to investigate the use of mathematical program formulations of expert systems. The purpose of their research has been to, e.g., understand the complexity of the expert systems and also to examine the feasibility of mathematical programming as an alternative solution methodology for those expert systems. This paper surveys and extends some of that literature that integrates AIlES and MP, and elicits some of the current research issues of concern. Keywords: Expert systems; Mathematical programming; Artificial Intelligence 1. Introduction The purpose of this paper is to investigate the interface between Mathematical Programming (MP) and Expert Systems (ES) and Artificial In­ telligence (AI). There are at least four facets of that interface. First, AIlES can be used gener­ ally to facilitate the use of MP. For example, AIlES can be used to formulate and interpret MP. Second, MP can be used to ensure that AIlES systems generate good solutions. MP ap­ proaches generate optimal solutions that might be used to solve parts of problems faced by Correspondence to: D.E. O'Leary, School of Business, Univer­ sity of Southern California, Los Angeles, CA 90089-1421, USA. general AIlES problems. Third, in some cases AIlES can be formulated as MP. This may have benefits such as understanding about the com­ plexity of the AIlES system or getting solutions faster. Fourth, it may be that formulation of parts of MP algorithms as ES I AI problems could facil­ itate the computational quality of MP algorithms. 1.1. Coupling ES I AI and MP Winston (1984, p. 2) noted that "one central goal of Artificial Intelligence is to make comput­ ers more useful". In that sense, there is substan­ tial need for artificial intelligence in the use of mathematical programming. Often the formula­ tion and interpretation of mathematical programs is too time consuming or too costly to impact 0377-2217/93/$06.00 © 1993 Elsevier Science Publishers B.V. All rights reserved http:0377-2217/93/$06.00 2 R.D. McBride, D.E. O'Leary / Integrating MP and AI / ES decision making. Further, the use of much MP software requires substantial expertise in opera­ tions research and the particular software. Thus, often times operations research just is not used (e.g., Fabozzi and Valente, 1976). Given this per­ spective there is substantial need for AI in opera­ tional research to make mathematical program­ ming 'useful'. This can be done by using AIlES to formu­ late, solve and analyze MP, in order to facilitate use of MP. Both domain-independent and do­ main dependent approaches are discussed in this paper. In addition, AIlES can be used in other facets of using MP models, such as model man­ agement. 1.2. Using MP in AlI ES models Not only can AIlES assist in the use of MP models, but the reverse can also be true. AIjES models typically employ heuristic approaches to solve problems. However, in some cases various subproblems might be effectively and efficiently solved as MP problems. This would guarantee optimal solutions to those portions for which optimal solutions could be developed. 1.3. Formulating ES I Al as MP There is another side to the relationship be­ tween AIlES and MP. MP has been established for a number of years. As a result, many facets of the complexity of particular problems have been established. For example, it has been established that the traveling salesman problem is a very complex problem (NP-complete). Thus, if it could be shown that an AIlES issue is equivalent to a traveling salesman problem then that would indi­ cate that it is an equally complex problem. Further, many special formulations of MP problems have been made to solve specific prob­ lems, for example, shortest path problems or min­ imum cutset problems. Thus, if there is a map­ ping that allows us to view AIjES as mathemati­ cal programs, then we can make use of those formulations to solve analogous problems. In addition, assuming that we can establish a mapping between MP and ES then those situa­ tions when it may be preferable to solve the problem from an MP perspective or an ES per­ spective could be explored. 1.4. Formulating MP solution algorithms with AlI :1 ES Probably the most neglected aspect of coupling AIlES into or with MP is integrating AIlES into making MP algorithms more effective. When re­ searchers first were unsuccessful in their ability to duplicate the results of Karmarkar's well-known algorithm for linear programs, it was rumored that part of the success of Karmarkar's algorithm was Karmarkar. It was rumored that Karmarkar 'assisted' the algorithm with his expertise at criti­ cal points in the solution process. Whether or not this was true is not the issue to be discussed here. However, it does raise the question: (How) can expertise be integrated into algorithms to assist the algorithms? If so, then AIlES can be used to improve existing algorithms. Since little has been developed on this question, it will not receive further discussion in this paper. 1.5. Plan of this paper This paper proceeds in the following manner. Section 2 investigates some of the issues involved with coupling of AIlES and MP. In addition, that section provides a brief review of some AI concepts. Section 3 discusses intelligent systems using MP that are domain independent. Some research issues are discussed as a summary for that sec­ tion. Section 4 examines the implications of build­ ing intelligence into domain-dependent applica­ tion system that either are built around mathe­ matical programming systems or mathematical programs that are built into AIlES. This section examines some of the implications of computeriz­ ing the knowledge of the operational research expert and those implications of using a mathe­ matical programming approach as opposed to a heuristic approach. It also discusses the unique aspects of development and implementation re­ quired for establishing the example systems with the embedded mathematical programs. Section 5 discusses MP representations of AIlES, and their use to better understand AIlES applications and the potential use of those MP formulations rather than AIlES approaches to solve those problems. Section 6 provides a sum­ mary of the paper. R.D. McBride, D.E. O'Leary / Integrating MP and AI / ES 3 1. Coupling AI / ES and MP Although most of the research done to date has focused on expert systems and MP this paper does not limit itself to ES. Instead, a broad-based approach is used, focusing on a number of com­ ponents of AI. The purpose of this section is to briefly review some key concepts in AIlES and the nature of coupling AIlES and MP. 2.1. Background: AI and ES Newell and Simon (1972, p.6) have defined AI as " ... the part of computer science devoted to getting computers (or other devices) to perform tasks requiring intelligence". Two areas of AI appear to be of substantial use in coupling with MP, expert systems and case-based reasoning. ES are a branch of AI that have received substantial attention. There are a number of defi­ nitions of ES. The most narrow of those defini­ . tions focus on the need for the system to be a rule-based system ('If A, then B'), based on knowledge gathered from an expert, that func­ tions at the level an expert would function. Other defulitions allow alternative forms of knowledge representation (e.g., frames), other sources of knowledge acquisition (e.g., text books), and other levels of performance (e.g., 'satisfactory'). Case-based reasoning (CBR) involves the pro­ cess of making decisions based on specific exam­ ples of what has occurred in the past, rather than a set of rules. Previous cases or plans are stored for use in solving future problems. In addition, means of adapting previous decision making· problems are saved. By making previous solutions available to decision makers, the decision maker can anticipate variables of concern and alterna­ tive solutions. In addition, past mistakes can be avoided, while short-cuts can be made available. As noted by Hammond [1988, p.17], the ideas behind case-based planning rise out of the simple principle: If it worked, use it again, and a corollary; if it works, don't worry about it. The refinements of the basic idea come out of a second, equally simple principle: If it didn't work, remember not to do it again, to which is added: If it doesn't work, fix it. 2.2. Insight, not numbers MP, in general, and linear and integer pro­ gramming, in particular, have been some of the most successful operations research methods to solve a wide range of constrained optimization problems. Although much of the research and many of the headlines are aimed at generating faster algorithms, as noted by Geoffrion (1976), "the purpose of mathematical programming is insight, not numbers". In order to build such insight into the systems that use MP, there have been at least two basic trends. First, some researchers have investigated AIlES .approaches to have the system formulate, solve, debug and interpret output from the sys­ tem, while assuming relatively generic domains (domain independent). Second, other researchers have made more detailed assumptions about the domain in which the application is based (domain dependent). These assumptions generally allow the researcher to develop systems that formulate programs based on general inputs from the user and they allow the researcher to be more specific in the investigation of output and interpretation of the meaning of that output. 2.3. Shallow versus deeply coupled systems Kitzmiller and Kowalik (1987) distinguish be­ tween shallow and deeply coupled systems. A coupled system is any system linking both nu­ meric and symbolic processing. Such a distinction may be useful in the analysis of coupling of AIlES and MP. In the case of shallow systems, the MP is treated as a 'black box'. The AIlES portions of the system have little knowledge about what goes on in the use of MP. On the other hand, in deeply coupled systems, the AIlES por­ tion of the system has extensive knowledge about MP. For example, it may be able to formulate or interpret the system. As will be seen later in the paper, most of the systems developed so far coupling AIlES and MP have been deeply coupled. Since most of these systems are research systems designed to explore the relationship between AIjES and MP this is not unexpected. In those systems cost and development time typically are not a factor. How­ ever, this seems to ignore the potentially cost effective approach of shallow coupling. As a re­ r f 4 R.D. McBride, D.E. O'Leary / Integrating MP and AI / ES suit, it is probably not unusual that the one shallow coupled system found in the litterature was a system developed for an actual business setting. Deeply coupled systems likely are more robust (Kitzmiller and Kowalik, 1987), with respect to the use of the system's results. This occurs be­ cause the AIjES portion is aware of the limita­ tions of the use of MP and what MP requires. Unfortunately, these systems are not often 'user­ proofed' with respect to those limitations of MP. If the system formulates, solves and interprets the solution, the user may not know if the problem is an appropriate use of the system. 2.4. Other implications of coupling AI j ES and MP There are at least two implications of coupling a mathematical program in an AljES system. First, coupling MP and AIjES indicates that the expertise of the operation research analyst can be captured in a computer program. This suggests that the design, formulation, interpretation and management of those MP can be formulated as a program. Initially, it was unclear as to the ability of developers to accomplish that task. As seen in Sections 3 and 4, there should be no question regarding that task. This also indicates that there probably is no more need to develop prototypes simply to test the ability to develop programs of this sort. Second, building MP components into an AIjES indicates that the developer expects an advantage with MP. Those advantages could in­ clude that an analytic approach rather than a heuristic approach yields a faster solution time or better solutions. Whether or not those advan­ tages can be realized is an empirical issue, and is discussed later in the paper. 2.5. Development methodologies for coupled AI j ES and MP A prototyping approach typically is promul­ gated for expert systems (e.g., Hayes-Roth et aI., 1983) However, researchers recently have sug­ gested the use of more traditional software engi­ neering processes (Bull et aI., 1987). The develop­ ment of coupled MP and AIjES systems has not pointed in any new directions or in anyone direction. Most of the systems appear to have been developed using a prototyping approach, although that is not clear. In addition, the possi­ ble existence of unique development methodolo­ gies deriving from such coupled hybrid systems also is not clear. 3. Domain-independent analysis of mathematical programs The most general approach to deriving insight from a mathematical program is to make no assumptions about the specific domain, in effect, draw only on the knowledge of mathematical programming and mathematical programming technical knowledge for system intelligence. Re­ search in this area can focus on how to accom­ plish particular tasks in this process, what tasks can or should be automated, and to what extent these efforts can remain domain-independent. Systems have been developed with intelligence to • choose which algorithm is required to solve the particular problem and solve the problem, • formulate problems as mathematical pro­ grams, and • interpret and debug problems. 3.1. Algorithm choice and solution Schittkowski (1985, p.2) developed a system that the author called " ... the first implementa­ tion towards an expert system for mathematical programming". Various options are available in the system for the formulation of various linear and nonlinear functions. The intelligence of the system comes from its ability to choose a suitable linear or nonlinear mathematical programming algorithm to solve the program. The system then writes a FORTRAN source program that will solve the problem. That program is then executed and the numerical results are stored in a database, available for further processing, retrieval or mod­ ification. 3.2. Problem formulation AI jES can be used to assist in the formulation of MP. The research that has been done to-date ,'--­ R.D. McBride, D.E. O'Leary / Integrating MP and AI / ES 5 has focused on general aspects of the formulation process. For example, there must be a process to generate variables and constraints in an intelli­ gent manner in order to facilitate problem forma­ tion. As will be seen latter, if additional assump­ tions are made about the domain, then greater model specificity can be obtained in the formula­ tion process, since the domain dictates certain problem features. There have been at least two approaches toward the development of intelligent systems designed to assist users in the general formulation of MP. LPFORM In a sequence of papers, Murphy and Stohr (1986), Stohr (1988), Murphy, Stohr and Ma (1988), Murphy, Stohr and Asthana (1989), and Ma, Murphy and Stohr (1989a,b) describe an intelligent system, LPFORM, designed to formu­ late linear programming problems. As will be seen latter in this paper, LPFORM also facili­ tates domain-dependent problem formulation. As noted in Stohr (1988) there are three major com­ ponents to that system. First, LPFORM supports three generic classes of inputs: row orientation ('the total tons of prod­ uct produced in each factory has to be less than certain limits'), activity orientation ('we have buy­ ing and selling activities in each of our ware­ houses') and transportation structures. These classes of inputs are used to define different objects of interest. Second, LPFORM allows the definition of problems using icons to represent objects being modeled, such as warehouses, transportation flows, etc. This object-oriented approach to rep­ resenting knowledge allows hierarchical arrange­ ment of objects so that lower level objects inherit properties of higher level objects. Once an object is defined and related to other objects, inheri­ tance limits the number of properties that must be defined at the level of individual objects. This hierarchal structure also is used to structure the graphics used in LPFORM. An algebraic repre­ sentation is generated as an artifact of the formu­ lation of the problem in terms of its objects. Third, LPFORM also provides for a relational database for data required in the problem. Fur­ ther, LPFORM allows the storage and recall of model components that have been stored previ­ ously. Users can recall all models using a given resource or all with a common activity. There is more discussion on LPFORM in Section 4. NETSYS McBride and O'Leary (1993) discuss another system aimed at the generic formulation of gener­ alized network models, a class of mathematical programming problems. Many problems can be formulated as generalized networks: transporta­ tion, shortest path, assignment and transshipment models. Generalized network models can represent machine efficiencies, sewage treatment, and many other processes. When network multipliers are interpreted as transforming one good to another, then generalized networks can be used to model manufacturing, blending, production, and a broad range of other processes. Since the scope of the system is limited to network problems, this allows the system to per­ form formulation, data editing, etc., based on the assumption of a generalized network structure. The structure assumption also allows the system to perform a number of infeasibility checks, and present information on those infeasibilities, so that the user can address those problems. NETSYS also uses an object-oriented ap­ proach that is facilitated by the development en­ vironment, Microsoft's Windows. The system uses network models, built by creating node sets, and taking cross products and 1-1 mappings with subsets of the node sets to create the arc sets. This approach requires minimal user effort to build substantial models. The system does a minor amount of interpre­ tation of the solution. The system also provides a sensitivity analysis of the problem. 3.3. Model interpretation and debugging In one of the first systems to integrate AIlES and MP, Greenberg (1983) described a system called ANALYZE. That system was designed for the linear programming expert, to assist in the investigation of solutions of linear programs. The system has the ability to investigate issues such as feasibility, redundancy and sensitivity analysis at a general level. Greenberg (1983, 1985) describes a later ver­ sion of ANALYZE as forming (Greenberg, 1983, 6 RD. McBride, D.E. O'Leary / Integrating MP and AI / ES p.333) " ... one fundamental component of an intelligent mathematical programming system". In that revised version, an English language dis­ course version of the earlier system is developed and extended. Greenberg's (1983) system is de­ signed for use after the system has been formu­ lated, however, it can be used during model im­ plementation. As a result, it might be used in conjunction with those systems discussed above for formulation. Such a system is dependent on knowing the structure of the constraints to ensure that it can debug and interpret the system. For example, it needs to know the type of equations that it can expect. Given those expectations and a user sup­ plied set of tables that correspond to those equa­ tions and variables, the system uses that structure to analyze and 'interpret' the resulting tableau. 3.4. Implementation characteristics The implementation of these systems has cer­ tain generic concerns including: the use of graph­ ics and knowledge representation. Graphic interfaces One of the most comprehensive discussions of the use of graphics in any of the above systems is Ma, Murphy and Stohr (1989a). The primary fo­ cus of that graphic interface is to provide the user with a different representation that allows them to depict their problems in a graphic, rather than mathematical form. Important empirical issues are 'does it make a difference' and 'what are the best approaches'. Ma (1988) has initiated addi­ tional research in this area. Knowledge representation Consistent with the Kitzmiller and Kowalik (1987), the dominant form of knowledge repre­ sentation appears to be objects. Generally the user chooses or generates an object that meets their needs and the system generates equations that correspond to that object. In addition to their correspondence to graphic structures, it seems that the hierarchical nature of objects is a very convenient approach to store knowledge for such systems. Further, recently developed pro­ gramming vehicles, such as WINDOWS, are ob­ ject-oriented, furthering the ability to develop such systems. 3.5. Model management An important issue is the reusability of previ­ ous model parts or templates. Generally, this is referred to as model management. An approach that employs AI for model man­ agement would likely use a case-based approach to capture and retain knowledge about previous models that have been generated. That knowl­ edge could be used to assist in generating new models or modifying existing models. In a domain-independent environment, there are a number of characteristics that can be used to structure the cases. These include: who devel­ oped the application, when it was developed, who was the client of the application, size of the model, and many other characteristics. In addi­ tion, hierarchical relationships within a given characteristic can also be pursued. The use of case-based reasoning to manage models has received limited attention. As a re­ sult, additional research is needed to focus on what are the appropriate characteristics used to capture previous models, how well the case-based approach works in the choice of previous MP models, and a variety of other systems issues. This model management issue has behavioral and design aspects that are facilitated by investi­ gation of a case-based approach. From a behav­ ioral perspective, to what extent would users and designers employ parts of previous systems? What do expert model designers do? From a design perspective, what is the best way to store, recall and use previous model templates? In addition, what characteristics distinguish those templates that receive frequent re-use - are they domain­ dependent? Further, how do we store parts of models for further use? Whenever there are more than a few core models or core users such issues become very important. 3.6. Selected additional research issues There are additional research issues in both the behavioral analysis of the integration of A1 /ES and MP and in the design of these sys­ tems. This section summarizes some of those issues. Although the goal of these systems is to help people solve problems, little, if any empirical work has been done to examine the extent to " 7 R.D. McBride, DE O'Leary / Integrating MP and AI/ ES which the models developed aid in problem solu­ tion. Related work on the impact of using domain dependent systems is discussed in the next sec­ tion. Virtually all of the systems discussed in this section at some point began to draw on domain knowledge. Focusing only on MP structures is a focus primarily on syntax (e.g., Greenberg, 1987). Generally, the introduction of semantics occurs with the investigation of a specific domain. Un­ fortunately, there is little research that studies 'how far' we can go in the development of these systems without assuming domain knowledge. Murphy, Stohr and Ma (1988) is an exception. In that paper, there are some formal results that indicate what can be accomplished with syntacti­ cal knowledge and what requires semantic knowl­ edge. Although the development of systems pro­ vides an empirical frontier of the ability to model between semantic and syntactical, such theoreti­ cal results provide important guidelines to future system development efforts. 4. Domain-specific analysis of mathematical pro­ grams There have been a number of domain-specific systems that have integrated MP and AIlES. In most situations, it appears that the focus is the MP and the AIlES has been designed to facili­ tate the development or interpretation or use of the system. In at least one case discussed below, the AIlES was the focus, but MP was used to provide a better solution to one component of the AIIES system. Domain knowledge can facilitate the integra­ tion of AIlES in MP. The domain provides the opportunity for increased specificity in the analy­ sis of the input or output of the program. For example, if it is known what the output variables represent, then it is possible to develop rules to analyze the output, including the solution and the dual variables, just as a human analyst might do. 4.1. Applications There have been a number of domain-specific applications that employ at least a linear pro­ gramming module. These applications range from production models to finance models to strategy models to personnel models. It appears that the most research in this area has been done in production and related models. Some of the ap­ plications discussed in this section are domain­ based applications of some of the literature' dis­ cussed in the previous section on domain-inde­ pendent models. Most of the models in this section appear to be deeply coupled models. That is, most of these models are intelligent formulators and inter­ preters of MP for domain-specific problems. Production management Binbasioglu and Jarke (1986, p.215) made heavy use of domain specific knowledge to gener­ ate a prototype system for the " ... conceptual development and symbolic (as contrasted to nu­ meric) formulation of the model" for production management problems. The model combines the use of syntactic knowledge about linear program­ ming and semantic knowledge about the produc­ tion management. The system is a prototype that has not been deployed for actual use. The focus of the prototype was to bypass the potentially lengthy and ambiguous development process typi­ cally involved in operational research projects. As a result, the system was designed so that the manager could formulate a model directly. Knowledge in the system is represented as frames. Accordingly, the system allows for unique characterization of individual products and re­ sources, and hierarchies in the relationships of machines used in the production of different products. The system has a linear programming knowl­ edge base and an application knowledge base. The linear programming knowledge uses two dif­ ferent knowledge bases. One contains knowledge about naming and selecting variables, problem types and other issues. Another contains knowl­ edge about obtaining and interpreting parameter values. The system's knowledge about the domain covers a large range of problems, including the allocation of resources to products, relative com­ position and contribution of products, etc. Widget production management In the discussion of LPFORM, a number of different models are discussed to exemplify the concepts and gain additional meaning through semantic knowledge. Stohr (1988) discusses a pro­ 4; F*' 8 RD. McBride, D.E. O'Leary / Integrating MP and Al / ES duction management problem that produces wid­ gets and then ships them to warehouses. Murphy et aI. (1988) discuss a problem of shipping grain, Ma et aI. (1989a) investigate capital capacity, and Ma et aI. (1989b) investigate a problem of trans­ porting energy. In order to formulate the problem, the user needs to specify the basic structure of the prob­ lem, using the graphic interface. In the widget example, the user specifies a relationship be­ tween Factories and Warehouses and between Raw Materials and Widgets. At the basis of these applications is a 'definition of production' activity provided to the system. This definition includes a declaration of inputs and outputs, upper bounds lower bounds, coefficient definitions, activity name, and other information. In the case of wid­ get production that means defining raw materials as inputs, widgets as outputs, activity name as 'product mix'. Then the user needs to specify the "'fu%~~\'\~\.\,.~,weviously established sets Facto­ ries, Warehouses, Widgets and Raw lVlaterl'fu."s. Production scheduling via transportation model Using the version of the transportation model developed by Bowman (1956) for production scheduling, O'Leary (1986) designed an expert system to analyze the output from that mathemat­ ical program. That analysis exploited three as­ pects of the formulation. First, the transportation problem provides knowledge that the right-hand side. quantities are sources and requirements, while the solutions to the problem define the amount of a shipment from a source to a destina­ tion. Second, the context of production schedul­ ing further defines those constants as forecasted sales and necessary production. Third, the further detail of production scheduling establishes the sources of production and the costs as inventory, regular time wages and overtime wages. All of this domain information provides additional in­ sights over and above the notion that the problem is a generic linear program. Thus, this informa­ tion can be used to interpret the MP solutions, using a rule-based ES. <;.~" t\'Ib: Qlanning and interface to strateirc plan­ ning Lee and Lee (1987) develop a system that analyzes the results from a short-term product mix model with a rule-based system to determine if the short-term results and long-term objectives are congruent. The output of the analysis by the rules can include a set of changes that should be made to the linear programming model to have consistency in the short- and long-term goals. Often linear programming is criticized for its focus only on the long run or only on a given objective. This type of application can be use to mitigate many of those criticisms, by introducing additional analysis into the results to answer some of those criticisms. Intelligent aggregate planning Lee and Kang (1988) developed a rule-based system to assist in the consideration of additional qualitative factors in aggregate production plan­ ning problems (e.g., Holt, Modigliani, Muth and Simon, 1960). In particular, issues such as em­ ployee morale and customer goodwill are not directly included in such models. Thus, a rule­ based system was developed to facilitate consid­ erauon ot 'Th~'e\"h~~'\."",\t.~w.. the context of the MP. The rule-based system evaluates the minimum cost solution in terms of its impact on employee morale and customer goodwill. If any of the qual­ itative goals related to these issues is unsatisfac­ tory, then trade-offs can be made via new or different constraints for the MP. Additional extensions and discussions of re­ lated models are provided by Lee (1991). For example, similar approaches are discussed for scheduling crude oil delivery. In addition, an ap­ proach is discussed for formulating such MP (Lee et aI., 1989). Aggregate production planning Donnelly et al. (1991) develop a system, AG­ PLAN-LP, that combines a mathematical pro­ gramming model with heuristics to generate a production plan. Holt, Modigliani, Muth and Si­ mon (1960) formulated the production planning problem as a quadratic program. Although this means that an optimal solution can be developed for the formulated problem, the disadvantages of U~\l\~ such an approach include that management may not be comfortable with the m'Afuematics and the model requ\1.'~"1. ~\. '<\.~~tOximation, which may not accurately reflect the actual pro­ cess. -------------------------------------------.....-.,. 9 RD. McBride, D.E. O'Leary / Integrating MP and Al / ES In a test of the quality of using an ES /MP approach, Donnelly et aL (1991) compared an expert system based entirely on heuristics to AG­ PLAN-LP. The traditional expert system has roughly 500 rules distributed in four modules (data acquisition, data modification, initial pro­ duction plan and final production plan). AG­ PLAN-LP was created by removing the 25 rules in the initial production plan module, and substi­ tuting them with 40 rules and a link to a linear programming model. The authors do a detailed analysis of the qual­ ity of the two systems. Apparently, not only was AGPLAN-LP able to generate more inexpensive solutions than the expert system, but it also was able to generate solutions that were more accept­ able to the planner. They also found that AG­ PLAN-LP had more acceptable solutions than those generated solely by a linear program. Cash management In a sequence of papers, McBride, O'Leary and Widmeyer (1989a,b, 1990), discuss an intelli­ gent system for using linear programming to solve cash management problems. Like the Binbasioglu and Jarke (1986) system, CASHMANAGER is designed for a domain user not the operation research expert. Knowledge in the system is stored in frames and rules. With CASHMANAGER, the user generates a problem by choosing from the set of available financial instruments represented as templates, with corresponding sets of equations. The user must only specify time frame over which the system is operative and the costs and returns associated with the instrument. The system was developed to formulate and solve the problem of choosing the optimal portfolio. In addition, the system then can assist the user in understanding the results, e.g., by analyzing the dual variables. CASHMANAGER does more than just a con­ ceptual formulation of the problem: it also inves­ tigates the formulation for feasibility. In addition, since the system is aimed at relatively naive users, the user is integrated into the process of making sure that the system is feasible. Infeasibilities are displayed to the user in a manner that relates to the instruments and the corresponding dollar flows. The system assumes an underlying generalized network structure to the cash management prob­ lem. That approach allows the generation of inte­ ger solutions in a timely manner. Debt advisory system Dempster and Ireland (1988, 1989) have devel­ oped a system called MIDAS. The system is an intelligent debt management decision support system, which supports planning for a Canadian utility. As noted by Dempster and Ireland (1988, pAlS), the conceptual design of the system " ... provides for selected models appropriate to the decision process; configures and solves the models on user request; explains, interprets and refines their results interactively and assists the user in evaluating alternative borrowing plans". In terms of the specific mathematical program­ ming module this system is relatively unique, since it focuses on a stochastic programming model. Once output is received from the model it is refined by the system based on rules developed jointly with an experience corporate treasurer and professional underwriters. The knowledge in MIDAS falls into four main categories: corporate knowledge (goals, planning constraints, evaluation criteria); financial! eco­ nomic knowledge (debt sources, types of debt available, future rate expectations); modeling knowledge (variables and their explanations, mathematical functions); and system and input/ output objects (to handle I/O functions). As with a number of other systems, the frames allow for a parsimonious way to store knowledge, because of the hierarchical relationships that can be estab­ lished. Manpower allocation Berger and Huttinger (1986) describe an appli­ cation that is a combination of linear program­ ming and expert systems to solve manpower allo­ cation issues. The ES is used to formulate the MP. That system has three distinct modules. The first module ranks tasks in terms of cost-benefit ratios, where cost is measured in person-years and benefit is measured as the fraction of an organization's objective that is accomplished by the task. The expert system uses rules and heuris­ tics to convert the input allocations to a set of constraints. Then the linear programming module allocates manpower, subject to the constraint set generated by the ES. 10 RD. McBride, D.E. O'Leary / Integrating MP and AI / ES Developing financial statements Back (1992) developed an expert system that integrated a rule-based approach and a MP ap­ proach to assist in the development of financial statements. In Finland the development of finan­ cial statements is a complex process with many interacting rules and choices to be made. It gen­ erally is so complex that even experts are likely to make the wrong decision. Thus, potentially this system can provide the user with the appropriate decision support to help generate the financial statements. Back (1992) is one of the few researchers to do behavioral experiments to determine whether the use of the system that she developed actually made a difference. Her analysis found that the use of the system led to correct answers in 94 out of 96 questions whereas traditional support, using a text book to determine if answers were correct, led to only 53 correct answers out of 120 possible problems. The system makes a difference. Auditing judgement Kellyet al. (1987), Willingham and Ribar (1988) and Ribar (1988) describe one of the largest financial-based expert systems developed to date. The system was designed to assist public account­ ing auditors in their analysis of the collectability of loans. As developed by Peat, Marwick, Main and Co., 'Loan Probe' has over 3000 rules. As noted in Willingham and Ribar (1988), it also has an embedded integer programming module. The integer programming module was devel­ oped in an iterative manner. As noted in Willing­ ham and Ribar (1988, p. 183), "Initially, we attempted to solve this problem by investigating the methods used by practitioners. This attempt resulted in a large, that is, approxi­ mately 1000 rule subsystem which generally yielded only a good first approximation. This approach was eventually abandoned and the problem reformulated as an integer programming problem". The user does not know that the system contains an integer programming module. Instead, the re­ sults from the programming module are directly integrated into the system. This system is different from the others dis­ cussed in this section since the focus is not on using an ES to either formulate or interpret the results of the MP. Instead, MP is used to solve a subproblem in and ES, which then feeds into the rest of the system. This system's use of MP is probably best described as shallow, yet very cost effective. It is likely that more applications of this type will be generated if developers of ES do not limit themselves to a particular type of problem solution approach, whether that is rules or linear programming. 4.2. Advantages and disadvantages There are some potential advantages and dis­ advantages of using a coupled AIlES and MP approach in application domains. Replacing the operational research analyst Coupling a mathematical programming mod­ ule in an AIlES replaces some of the role that the operational research analyst might play. In­ stead of the analyst gathering information from the user, formulating the problem, solving the problem and interpreting the solution, systems have been developed to perform those tasks. Thus, the time and the cost required to develop a model, within the domain of the system, and a solution likely will be substantially smaller. However, there also may be some drawbacks. Because it is a computer system, it is likely to be locked into a limited number of views of the world. Typically, such systems are rigid and sel­ dom are able to provide unique or creative solu­ tions to unusual problems. The environment may change causing a change in the optimal decision process, yet current generation AI and expert systems rarely are designed to change in response to those environmental changes. The systems dis­ cussed apparently do not have any such sensing mechanisms to note when there is a change in the environment. Thus, an important research issue is how to develop systems that know when they are likely to not be useful and how to develop systems that change in response to environmental changes. Optimality The use of optimization models, with rule­ based approaches, has certain advantages and disadvantages. In the case of linear programs, algorithms typically rapidly find 'optimal solu­ 1..___ 11 R.D. McBride, D.E. 0 'Leary / Integrating MP and Al / ES tions'. However, in the case of integer or nonlin­ ear programs, depending on the processing capa­ bilities, optimal solutions may take substantial amounts of time and intermediate solutions may not be feasible. Further, although MP provides an optimal so­ lution, it may suffer from being a single dimen­ sional optimality. It is only optimal for the given objective function, and in the case of integer solutions, heuristics aimed to move it slightly away from that solution may provide substantially worse solutions. Further, the solution derived by an optimiza­ tion model is 'optimal' for the world for which it has been constructed. This is a limited view of the world, where precision at the 12th decimal place is important. Thus, systems like those of Lee and Kang (1988) that assist in the analysis of qualita­ tive factors might be extended to account for other 'realism factors.' Additionally, it could be important to study the use of multiple criterion decision making (MCDM) MP and AIlES. Feasibility The use of mathematical programs introduces the problem of feasibility. For example, one po­ tential problem with the Donelly et al. (1991) approach is that the system just stops whenever there is an infeasible solution to the linear pro­ gram. As noted by other researchers (e.g., Mur­ phy and Stohr, 1986), this is not a desirable approach for supporting decision making. Thus, research needs to be done on how users react to infeasibilities and different system responses, how relatively naive users can be asked to fix such infeasibilities and what intelligence the system needs to attack those infeasibilities. Explanation As noted in Swartout (1981, p.815), "to be acceptable, expert programs must be able to ex­ plain what they do and justify their actions in terms understandable to the users". With MP there generally is no 'explanation' as to 'why' the model chose certain results. Even when an expla­ nation can be generated it is limited to expression in terms of the MP structure, e.g., constraints. Although Kosy and Wise (1984) provide a means of interpreting some fundamental accounting equations, the use of MP leads to substantially more difficult explanation situations. ES have a tradition of 'explanation' with solu­ tion. If MP are to be coupled with AIlES then in order to maintain that tradition, research needs to be done in the area of explanation of why·one MP formulation was arrived at and why one solu­ tion is preferred over another. Some research has been done by Greenberg (1983, 1985, 1987); how­ ever, additional research is necessary. 4.3. Additional issues in systems design and devel­ opment: Knowledge acquisition, system develop­ ment, testing Coupling of AIlES and MP leads us to ques­ tion whether or not there are any unique require­ ments or approaches for the design and develop­ ment of such coupled systems. Knowledge acquisition There has been little investigation into the necessity of alternative forms of knowledge acqui­ sition for systems that are either shallowly or deeply coupled AIlES and MP systems. How­ ever, there has been some initial research into knowledge acquisition from experts for generic analytical representation (Fischhoff, 1989). Un­ fortunately, that study did not deal explicitly with the case of MP. As a result, knowledge acquisi­ tion in coupled systems of the type discussed in this paper remains a research area. Development Piaget (1973) noted that "all mathematical ideas begin by a qualitative construction before acquiring a metrical character". Piaget also noted that by moving too rapidly from the qualitative to the quantitative can impact the ability to under­ stand the mathematics underlying a given pro­ cess. In terms of systems development this can mean that unless there is an initial explicit effort to include a mathematical program in the overall system, that the mathematical programming na­ ture of subsystems is likely to be iteratively dis­ covered. As noted above, Loan Probe was devel­ oped in concert with those notions. Two rule­ based versions were developed before the mathe­ matical programming-based version was devel­ oped. Testing There are at least two basic testing questions that need to be addressed. First, does the ap­ 12 R.D. McBride, D.E. O'Leary / Integrating MP and AI / ES proach with a coupled AIlES and MP system do as well as or better than only an AIlES approach or only an MP approach. This is the critical problem of establishing a bench mark discussed by, e.g., Hayes-Roth (1989) and others. Second, what unique approaches must be used in the verification and validation of coupled AIlES and MP systems (Cohen and Howe, 1989). In the first case where there are MP and non-MP versions of the system, as in the case of the Loan Probe or the aggregate production plan­ ning problem, previous versions offer a unique benchmarks to verify and validate the output from the coupled system. For example, solutions provided by non-MP versions can be used to test MP versions, where ideally, later versions with the MP are at least as good, based on a single criteria, as the heuristic-based system. In the second case, there are substantial bod­ ies of literature in the areas of verification and validation of AIlES (e.g., O'Leary, 1987). Those efforts usually are aimed at the overall perfor­ mance of the system or take advantage of unique knowledge base structures. However, structural analysis of knowledge bases rarely considers more than a single form of knowledge representation at a time. Although there have been investigations into verification proceedures for different representations (e.g, rules), there has been little research into hybrid types of systems such as the coupled AIlES and MP systems discussed in this paper. Systems that couple MP and AIlES add additional complexity to the testing process, since they include multiple forms of knowledge. This coupled form of knowl­ edge appears to have received only limited re­ search to this point (e.g., O'Leary, 1988), and is an area where research is ongoing. 5. Using MP instead of an AIlES formulation Probably the most recent use of MP and AIlES is to use MP as an alternative formulation to the AIlES formulation typically either for solution purposes or to better understand charac­ teristics of the AIlES formulation (e.g., Jersolow, 1988). In these cases the MP is formed either based on a network representation of the knowl­ edge base or as a set of logical propositions. 5.1. Network representations Many forms of knowledge representation are either formulated as a networks or can be repre­ sented as networks. For example, the knowledge representation format, semantic networks, are network or tree representations of knowledge (e.g., Rich, 1983); frames are linked in either a network or tree format (e.g., Winston, 1984); in­ fluence diagrams (e.g., Howard, 1989) are Bayesian network representations of knowledge; and rules can be represented as a network (e.g., O'Leary, 1990). For example the rule 'if a, then b' can be viewed as a variable Xa,b' Since optimization problems in networks can be represented as MP, this indicates that opti­ mization issues in the areas of knowledge bases can be investigated using MP formulations. This allows the introduction into AIlES, of a wide range of problems that have already been solved in MP. For example, Glover and Greenberg (1988) and O'Leary (1990) develop MP formulations for verification of rule-based knowledge bases, e.g., detecting circular reasoning. MP formulations also can be used to study characteristics of problems in knowledge bases. For example, O'Leary (1990) formulates the problem of ordering rules in a knowledge base as a traveling salesman problem. This suggests that level of complexity of a problem in knowledge basis can be found by finding equivalent MP problems. 5.2. Logic operators Dhar and Ranganathan (1990) view knowledge as a set of logical propositions. Since Dantzig (1963) showed how logical propositions can be modeled in an integer programming approach, the extension from knowledge bases to MP is clear. For example, the clause 'Xl or not orx 2 X3' can be written as the following inequality: Xl + (1 -x 2 ) +x 3 ~ 1, where the values of true and false are denoted as 1 and O. This approach is discussed in more detail in Hooker (1988). Dhar and Ranganathan (1990) provide an empirical comparison between the expert system approach and the integer program­ ming approach. ~-------------------~--~~~--------.I'L.__-­ 13 R.D. McBride, D.E. O'Leary / Integrating MP and AI / ES Dhar and Ranganathan (1990) found that us­ ing an integer programming approach yielded unpredictable decision times, varying from a few minutes to a few days. In contrast, the expert system's time was less volatile. Consistently the solution time required between one and two hours on a SUN-3. In addition, in contrast to the inte­ ger programming approach, if the expert system could not find a solution, it generated a partial solution, if there was one. The integer program­ ming approach consistently was unable to find feasible solutions after many hours of running time. This has two difficulties: length of time and the impact of not have any feasible solution after all that time. This can prove additionally inconve­ nient when alternative plans or approaches need to be established. The integer programming approach typically generated solutions that were judged at least partially inappropriate by a human expert for three primary reasons: single objective limita­ tions, compiled knowledge limitations and global optimization limitations. The objective function in a linear program has a single goal. However, as researchers in multiple criteria decision making argue, human experts try to respond to multiple goals. In addition, single objectives, by their very nature of excluding other objectives, sometimes lead to unsatisfactory solu­ tions. Some steps can be taken to account for multiple objectives, such as objectives as addi­ tional constraints. Compiled knowledge refers to the sensitivity of MP to cost coefficients. As noted by Dhar and Ranganathan (1990, p.332), "the basic problem with the cost coefficients is that they incorporate a lot of compiled knowledge about preferences, flexibility and trading criteria which makes the behavior of the system somewhat unpredictable". Further, in some cases, knowledge may be diffi­ cult to express in the structure of constraints. For examples, anomalies in processes that are used in some situations and not in others can be difficult to model. Global optimization limitations refers to the basic tendency of MP to move to the lower bounds or upper bounds on the variables and the general inability to explain why a particular solution was chosen. Thus, in a comparison of MP and AIjES it appears in somecases the local optimality ap­ proach of ES is better received by users of the system and by those for on whom implementation of the system depends. Although MP provides a global optimum, that optimum often is for a different representation of the problem than the one for which an ESIAI approach could be for­ mulated. 6. Summary This paper has surveyed some of the coupled AIlES and MP systems. The general capabilities of these systems range from formulating mathe­ matical programming problems from limited raw data, to interpreting output from the systems, to extrapolating changes in the mathematical pro­ grams based on the output to deciding how to solve these systems. This approach to decision making has some apparent advantages. To a certain extent, these systems are aimed at replacing operational re­ search analysts, in order to improve the cost and time required to develop a solution to a decision problem that includes a constrained optimization problem. Further, mathematical programs exploit the decision making technology to the advantage of the user of such systems. However, this approach is not without limita­ tion. Such systems may lock decision makers into decision models that are inappropriate. Similarly, systems that include a mathematical program­ ming module are likely to furnish few explana­ tions for why a given solution was pursued. The development of systems that include such a module may be part of a gradual movement from the qualitative to the quantitative that oc­ curs with understanding a process. However, by integrating multiple forms of knowledge process­ ing (numeric and symbolic) into a given system testing of the resulting system may be compli~ cated further. Finally, implementation of such systems is facilitated by software that allows the user to access such models outside the basic software. Acknowledgement The authors would like to acknowledge the comments of Bruce Murtagh and the referees on an earlier version of this paper. 14 RD. McBride, DE O'Leary I Integrating MP and AI I ES References Asthana, A. (1989), "LPGRAPH: An iconic language for formulating linear programs", Unpublished Ph.D. Disser­ tation, Stern School of Business, New York University. Back, B. (1992), 'Assisting inexperienced accountants in de­ veloping financial statements', International Journal of In­ telligent Systems in Accounting, Finance, and Management 1/3, 155-162. Berger, M., and Huttinger, J. (1986), "The Expert System ­ Linear Programming approach (ESLP)", paper presented at the ORSA/TIMS meeting, October, 1986. Binbasioglu, M., and Jarka, M. (1986), "Domain specific DSS tools for knowledge-based model building", Decision Sup­ port Systems 2/3, 213-223. Bowman, E. (1956), "Production scheduling by the trans­ portation method of linear programming", Operations Re­ search 4, 100-103. Bull, M., Duda, R., Port, D., and Reiter, J. (1987), "Applying software engineering principles to knowledge-base devel­ opment", in: Proceedings of the First Annual Conference on Expert Systems in Business, Learned Information, NJ, 27­ 38. Cohen, P., and Howe, A. (1989), "Toward AI research methodology: Three case studies in evaluation", IEEE Transactions on Systems, Man, and Cybernetics 19/3, 634­ 646. Dantzig, G. (1963), Linear Programming and Extensions, Princeton University Press, Princeton, NJ. Dempster, M., and Ireland, A. (1988) "A financial expert Decision Support System", in: G. Mitra (ed.), Mathemati­ cal Models for Decision Support, NATO ASI Series, Vol. F8, Springer-Verlag, Berlin. Dempster, M., and Ireland, A. (1989), "Object-oriented model integration in MIDAS", Proceedings of the Twenty-Second Annual International Conference on System Sciences, Vol­ ume III, 612-619. Dhar, V., and Ranganathan, N. (1990), "Integer programming vs. Expert Systems: An experimental comparison", Com­ munications of the ACM 33/3,323-336. Donnelly, R., Saniga, E., Chester, D., and Pohlen, M. (1991), "Expert Systems in aggregate production planning", Un­ published Paper. Fabozzi, F., and Valente, J. (1976) "Mathematical Program­ ming in American companies: A sample survey", Inter­ faces 7, 93-98. Fischhoff, B. (1989), "Eliciting knowledge for analytical repre­ sentation", IEEE Transactions on Systems, Man, and Cy­ bernetics 19/3, 448-461. Geoffrion, A. (1976), "The purpose of mathematical program­ ming is insight, not numbers", Interfaces 7, 81-92. Glover, F., and Greenberg, H. (1988), "Logical testing for rule-based management", Annals of Operations Research 12, 199-216. Greenberg, H. (1983), "A functional description of ANA­ L YZE: A computer assisted analysis system for linear programming models", ACM Transactions on Mathemati­ cal Software, 9, 18-56. Greenberg, H. (1985), "The fifth generation of mathematical programming systems; Towards an intelligent MPS", Un­ published Paper, presented at the TIMS College of Prac­ tice of Management Science, 1985. Greenberg, H. (1987), "A natural language discourse model to explain linear programming models and solutions", Decision Support Systems 3, 333-342. Hammond, K., "Case-based planning: Viewing planning as a memory task", in: J. Kolodner, Case-Based Reasoning, Proceedings of a Workshop, Morgan Kaufman, Palo Alto, CA. Hayes-Roth, F. (1989), "Towards benchmarks for knowledge systems and their implications for data engineering", IEEE Transactions on Knowledge and Data Engineering 111, 101-10. Hayes-Roth, F., Waterman, D., and Lenat, D. (1983), Build­ ing Expert Systems, Addison-Wesley, Reading, MA. Holt, C, Modigliani, F., Muth, J., and Simon, H. (1960), Planning Production Inventories and Work Force, Prentice­ Hall, Englewood Cliffs, NJ. Hooker, J. (1988), "A quantitative approach to logical infer­ ence", Decision Support Systems 4/1,45-96. Howard, R. (1989), "Knowledge maps", Management Science, 35/8,903-921. Jeroslow, R. (1988), "Alternative formulations of mixed inte­ ger programs", Annals of Operations Research 12, 241-276. Kelly, K., Ribar, G., and Willingham, J. (1987), "Interim report on the development of an Expert System for the auditor's loan loss evaluation", in: Proceedings of the Touche Ross I University of Kansas Audit Symposium 1987, 167-188. Kitzmiller, C, and Kowalik, J. (1987), "Coupling symbolic and numeric computing in knowledge-based systems", AI Mag­ azine Summer, 87-90. Kosy, D., and Wise, B. (1984), "Self-explanatory financial planning models", in: Proceedings of the National Confer­ ence on Artificial Intelligence, Austin, Texas, 1984, 176-181. Lee, J., (1991), "Integration of Artificial Intelligence with optimization", Unpublished paper, Department of Man­ agement Science, Korea Advanced Institute of Science and Technology, Seoul, Korea. Lee, J., and Kang, B. (1988), "Intelligent production planning analysis using the post model analysis approach", in: E. Turban and P. Watkins, Applied Expert Systems, North­ Holland, Amsterdam, 87-106. Lee, J., and Lee, H. (1987), "Interaction of strategic planning and short-term planning; An intelligent DSS by the post­ model analysis approach", Decision Support Systems 3, 141-154. Lee, J., Chu, S., Shim, S., and Kim, M. (1989), "A knowledge-based formulation of linear programming mod­ els using UNlK-OPT", 2nd International Workshop on Artificial Intelligence in Economics and Management, North Holland. Ma, P. (1988), "An intelligent approach towards formulating linear programs", Unpublished Ph. D. Dissertation, New York University. Ma, P., Murphy, F., and Stohr, E. (1989a), "A graphics interface for linear programming", Unpublished Paper, Center for Research on Information Systems, New York University, February, 1989. Ma, P., Murphy, F., and Stohr, E. (1989b), "Representing • RD. McBride, D.E. O'Leary / Integrating MP and AI/ ES 15 knowledge about linear programming formulations", An­ fUlls of Operations Research 21, 149-172. McBride, R. (1988), "NETSYS - A generalized network modeling system", Working Paper, University of Southern California, School of Business, August, 1988. McBride, R., and O'Leary, D. (1993), "A modeling system for fmancial applications", Advances in Mathematical Pro­ gramming and Financial Planning 3, 119-136. McBride, R., O'Leary, D., and Widmeyer, G. (1989a) "A system for supporting cash management decisions", in: Proceedings DSS-89, Sponsored by The Institute of Man­ agement Sciences, San Diego, CA, June, 1989. McBride, R., O'Leary, D., and Widmeyer, G. (1989b) "A knowledge-based system for cash management with man­ agement science expertise", Annals of Operations Research 301-316. McBride, R., O'Leary, D., and Widmeyer, G. (1990), "CASH MANAGER: A knowledge-based DSS for cash manage­ ment", Advances in Mathematical Programming and Fi­ nancial Planning 2,89-106. Murphy, F., and Stohr, E. (1986), "An intelligent system for formulating linear programs", Decision Support Systems 2/1,39-47. Murphy, F., Stohr, E., and Asthana, A. (1989), "Reprcsenta­ tion schemes for mathematical programming models", Un­ published Paper, Center for Research on Information Systems, New York University, September, 1989. Murphy" F., Stohr, E., and Ma, P. (1988), "Composition rules for building linear programming models from component models", Unpublished Paper, Center for Research on Information Systems, New York University, July, 1988. Newell, A., and Simon, H. (1972) Human Problem Solving, Prentice-Hall, Englewood Cliffs, NJ. O'leary, D. (1986), "Expert Systems in mathematical pro­ gramming", Presented at the Symposium for Artificial Intelligence in Engineering, October, 1985; published in: B. Silverman and W. Hutzler (eds.), Artificial Intelligence for Military Applications, Operations Research Society of America. O'Leary, D. (1987), "Validation of Expert Systems", Decision Sciences 18/3, 468-486. O'Leary, D. (1988) "The use and validation of optimization methods in Expert Systems", Unpublished paper; pre­ sented at EURO IX, TIMS XXVI, Paris, July 6-8, 1988. O'Leary, D. (1990), "Measuring and managing complexity in knowledge-based systems: A network and mathematical programming approach", Presented at the New York ORSA/TIMS Meeting, October 1989; published in: D. Brown and C. White (eds.), Artificial Intelligence and Op­ erations Research, Sponsored by ORSA/TIMS, K1uwer Academic Publishers, Dordrecht. Piaget, J. (1973), To Understand is to Invent, Grossman, New York. Ribar, G. (1988), "Development of an Expert System", Expert Systems Review 1/3, 3-8. Rich. E. (1983), Artificial Intelligence, McGraw-Hili, New York. Swartout, W. (1981), "Explaining and justifying expert con­ sulting programs", in: Proceedings of International Joint Conference on Artificial Intelligence, 815-823. Schitttkowski, K. (1985), "EMP: A software system for mathe­ matical programming", Unpublished Paper, Mathematis­ ches Institut, Univeritlit Bayreuth, Bayreuth, Germany. Stohr, E. (1988), "Automated support for formulating linear programs", Unpublished Paper, Center for Research on Information Systems, New York University, February, 1988. Willingham, J., and Ribar, G. (1988), "Development of an expert audit system for loan loss evaluation", in: Auditor Productivity in the Year 2000, 1987 Proceedings of the Arthur Young Professor's Roundtable, Arthur Young, Res­ ton, VA. Winston, P. (1984), Artificial Intelligence, Addison-Wesley, Reading, MA. 
work_ej6l72na6rhrfd7levjj2olbt4	----	DIGITAL SIGNAL PROCESSOR CONTROL ALGORITHM FOR PWM INVERTER A. A. Heggo ABSTRACT Recent developments in power electronic device technology promises faster switching capability at high power. The hybrid pulse width modulation method which requires two of the four switches in a full bridge inverter are used, and enable pulse pattern operation at high frequency. The use of two level switching instead of three level switching enables the use of higher frequency for given computation time delay. The proposed control scheme is implemented using bi-polar junction transistors (BJT) controlling an inverter t o produce a very low frequency (THD) sinusoidal output voltage. Simulation and experimental results are presented to verify the performance. 1 INTRODUCTION The 111 bridge inverter in Fig. (1) is widely employed in various applications such as motor drives and active filter [1][2]. The inverter comprises switching poles S1, S2, S3, S4, S5 and S6. The switching poles are commonly controlled by a variety of PWM techniques 131, and by pulse- shified square-wave drives [4]. Application of switching devices such as IGBT achieve very high switching frequency PWM inverters with improved performance [5]. With the availability of high frequency switching devices the instantaneous feedback control (IFC) was presented [6]. The advantages of this technique are high transient response and the disadvantage is that relatively large harmonic amplitudes occur for frequencies near the average switching frequency. By using a microprocessor, a digital feedback approach such as' a microprocessor deadbeat control was proposed [7][8]. The PWM inverter system is converted into a discrete time system and a state feedback output deadbeat control is applied. Manuscript received from Dr; A. A. Heggo on : 1111 111998 Accepted on: 13/6/1999 Engineering Research Bulletin, Vol 22,No 3, 1999 Minufiya University, Faculty of Engineering, Shebin El-Kom , Egypt, ISSN 11 10-1 180 Digital signal processor @SPs) are now applied for the control of power electionics and drive systems. DSPs are much faster (ten or one hundred times) than a microprocessor [9]. Simplification of control hardware and corresponding reduction of cost are the principal advantages ofDSP control [lo]. A digital controller was used to implement the control algorithm and provide switching signal t o the power circuit. The proposed control model is implemented using DSPs controlling an inverter to produce a very low'total harmonic distortion (THD) in sinusoidal output voltage. Fig. 1 - System description 2 CONTROL ALGORITHM A modified algorithm of the deadbeat controlled PWM inverter is suitable for stabilised power supply systems. Two levels are used in the pulse pattern. Given a computation time delay, twice the switching fi-equency can be adapted resulting in lower THD sinusoidal output. This work presents the deduction of a new discrete time state equation and the proposed deadbeat control algorithm for PWM inverters. Simulation results are obtained and presented. , 2.1 Deadbeat Control This technique depends upon the digital feedback closed loop. This means measuring of output, and controls the inverter switches to generate the required PWM pattern to produce a low THD sinusoidal output voltage. The digital control algorithm is designed to control pulse width such that the output voltage equals the sinusoidal reference at every sampling instant, Fig. (2). So the output voltage will be in phase and very close to sinusoidal reference. Any deviation of the output voltage from the reference due to a load disturbance or non- linear circuits is controlled within one sampling interval Ts. (a) Sampling Instant 1 2 3 4 5 Fig. 2 - Sampling Time The PWM pattern is determined at every sampling instant digitally by DSPs based on the output measurements and the references. 2.2 State Model Reference I error Vin Vout + Inverter Filter Load Digital AID Controller Converter Fig. 3 - Block Diagram of Digital Control Inverter As shown in Fig. (3) the deadbeat controller for PWM inverter is considered. The inverter, L C filter and resistive load represent the plant of closed loop digital feedback system. L '~ Load R v o u t Fig. 4 - State Model The circuit shown in Fig. (4) is modelled as a second-order system with state vector fli], where Vis load voltage and i is capacitor current. The state equation becomes: where: 2.3 Two Level scheme The basic circuit for the deadbeat controlled PWM inverter with a two level switching pattern is shown in Fig. (5). Positive half cycle Negative half cycle Fig. 5 - Two Level Deadbeat Controlled PWM Inverter Pattern The continuous time domain state ( I ) can be written as: i= A X + B U where: x = state vector u = scalar input A = non-singular matrix Then the closed form circuit is: where x(fJ is the initial state vector at t = to ifthe input u is constant for f , i t i t 1 then Eq.(4) becomes: using Eq.(5), the discrete-time system equation with the input is derived as follows: and by expansion the terms: AAT/2 AAT A ~ ( A T / ~ ) ~ e x ] + - + 2 2 e A T ~ l + A T + (AT )2 2 then Eq.(8) becomes: X/(K + I ) TI = eAT X(KI] - 2eATI2 BEAT + ( + A'z + (I I ) 2 6 where (K + I ) = t;, and KT = 1, this is a discrete-time system of Eq.(3) Rewriting Eq.(l 1) gives: where: f, is corresponding to element of eAT gi is corresponding to element of 2 e A T n ~ ~ hi is corresponding to element of (T + A f / 2 + A ~ / ~ ) B E V(K), i(X) and AT@) represent the values of voltage, the current and the sampling time: t = K T Therefore, by analysis of Eq.(l2) gives: 3 COMPUTER SIMULATION The following circuit parameters were used in computing the pulse width AT(K), and to obtain waveforms of voltage and current. Line Load Two level scheme THDIfimdamental 0.889l15.6 The simulation results are shown in Figs. (6), (7) and (8) 4 CONCLUSION A two deadbeat controller is used to minimise the total harmonic distribution of the digital inverter. This scheme will lead to the following advantages: - provide more computation time for the same interval switching number - provide a maximum voltage equal to the dc busbar voltage during each switching interval, Ei + E The obtained results show the waveforms of output voltages during various types of modulation. REFERENCES Hue, C., Two-Level Switching Pattern In Deadbeat DSP Controlled PWM Inverter, IEEE Translation of Power Electronic, Vol. 10, No. 3, May 1995. Goodnough, F., Mos Controlled thyristor turns off IMW in 2ms, Electron Design, November 1988. Gokhale, K. P., Kawamers, A,, and HoR, R. G., Deadbeat Microprocessor Control Of PWM For Sinusoidal Output Waveform Systems, IEEE Trans. Ind. Appliket, Vol. 1A23, No. 5, pp 901-909, September/October 1987. Enjeti, P. N., Programmed Pwm Technique To Eliminate Harmonics - A Critical Evaluation, IEEE IAS C o g Rec., pp 418-430, 1988. Patel, H. and Hoft, R. G., Generalised Technique Of Harmonics Elimination And Voltage Control In Thyristor, Part I - Harmonic Elimination, IEEE Trans. Ind. Appli., Vol. 1 IA-9, No. 3, pp 3 10-317, MayIJune 1974. Kawamura, A., Haneyoshi, T. and Hoft, R. G., Deadbeat Controlled PWM Inverter With Parameter Estimation Using Only Voltage Sensor, IEEE, Power Electron, Vol. 13, No. 2, pp 118 - 125. 7. Hue, C. and Hofi, R. G., Deadbeat Controlled Pwm Inverter With Two Level Pulse Pattern For Ups, by MSEE project - University of Massouri, Columbia, November 1990. 8. Truxal, Feedback Control System And Synthesis. 9. Heuldewoth, J. and Grant, D. A,, The Use Of Harmonic Distortion To Increase The Output Voltage Of Three-Phase PWM Inverter, IEEE proc. Vol. 136, PLB No. 4, July 1989, pp 189-194. 10. Heumann, K., Rapp, G. and Jung, M., Comparative Study Of New Power Transistor With Respect To High Frequency Inverter Application, EPE 89, 1989, pp 99-104. 6 NOMENCLATURE AT sampling time A, B matrix of state variable equation AID analogue / digital converter C filter capacitance DSP digital signal processor f output frequency I capacitor current K integer, positive number 0, 1,2,3 L filter inductive N no. of sampling PWM pulse width modulation R load resistance T interval time V load output voltage V;, input voltage (5- b) CURRENT WAVICI;ORhl Fig.6 The voltage and current waverorms under the "15 Pulse Modulation Mode" - Pul :e acSc Fig.7 The voltage and current waveforms under the "9 Pulse Modulation Mode" , , , . (8 - b ' C U R I < E N T \VAVI3FOllh] Fig.8 The voltage and current waveforms under the "5 Pulse Modulation Mode" 
work_ejke54pkqfa7ngcaso2iwvuw3u	----	NASA Technical Reports Server (NTRS) 
work_elhlxrp4mjdxfdm56dtpp5kjku	----	Semantic Web Service Discovery Using Natural Language Processing Techniques Jordy Sangersa, Flavius Frasincara, Frederik Hogenbooma,∗, Vadim Chepeginb aErasmus University Rotterdam, P.O. Box 1738, 3000 DR, Rotterdam, The Netherlands bTie Kinetix, P.O. Box 3053, 2130 KB, Hoofddorp, The Netherlands Abstract This paper proposes a Semantic Web Service Discovery framework for finding Semantic Web services by making use of natural language processing techniques. The framework allows searching through a set of semantic Web services in order to find a match with a user query consisting of keywords. By specify- ing the search goal using keywords, end-users do not need to have knowledge about semantic languages, which makes it easy to express the desired semantic Web services. For matching keywords with semantic Web service descriptions given in WSMO, techniques like part-of-speech tagging, lemmatization, and word sense disambiguation are used. After determining the senses of relevant words gathered from Web service descriptions and the user query, a matching process takes place. The performance evaluation shows that the three proposed matching algorithms are able to effectively perform matching and approximate matching. Keywords: Natural language processing, Semantic Web services, WSMO, Web service discovery 1. Introduction With the emergence of Web services and the Service Oriented Architecture (SOA), business process components are more and more being decoupled. Using Web services in SOA creates a wide network of services that collaborate in order to implement complex tasks. Currently, Web services are commonly described via narrative Web pages containing information about their operations in natural languages. These Web pages contain plain text with no machine interpretable structure and therefore cannot be used by machines to automatically process the descriptive information about a Web service. ∗Corresponding author; tel: +31 (0)10 408 8907; fax: +31 (0)10 408 9031 Email addresses: jordysangers@hotmail.com (Jordy Sangers), frasincar@ese.eur.nl (Flavius Frasincar), fhogenboom@ese.eur.nl (Frederik Hogenboom), vadim.chepegin@tieglobal.com (Vadim Chepegin) Preprint submitted to Expert Systems with Applications January 24, 2013 To promote the automation of Web service discovery (Keller et al., 2005) and composition (Hikimpour et al., 2005), a number of different semantic languages (Martin et al., 2004; de Bruijn et al., 2005; Vitvar et al., 2007) have been created that allow describing the functionality of services in a machine interpretable form, while original Web service descriptions contained only information about the data types and bindings as a description of a Web service functionality. The semantic Web service descriptions use ontologies to describe the behavior of a Web service by applying reasoning over Web service semantics. In this way, the semantics described in ontologies enable systems to interpret what a Web service is doing, stimulating automatic Web service discovery and compo- sition. The ontologies, however, are created by humans and therefore contain natural language. This allows humans to understand the concepts defined, but a system, in contrary to humans, can only understand on- tology concepts and their relationships to a limited extend. Natural Language Processing (NLP) techniques can therefore help in better defining the context of a Web service. When using one holistic ontology, machines can discover and compose Web services automatically based on the semantics defined. Using one holistic ontology is, however, hardly reachable and therefore it is impossible to reason based only on formal logic. NLP techniques can help overcome the ambiguity problems between different ontologies that are being used by semantic Web service descriptions. Last, service composition should be driven by people who know business processes and not by techni- cians. Thus, end users must be able to discover these Web services based on keywords written in human language. Therefore, a discovery mechanism must be developed in such a way that a bridge between key- words written in a natural language, on one hand, and Web service descriptions provided using semantically enhanced languages, on the other hand, can be created. In this paper, the Semantic Web Service Discovery (SWSD) framework is proposed, which enables end users to search, using keywords, for existing Web services described by means of a Semantic Web language for service annotation. This process consists of several steps including: • Extraction of information from semantic descriptions to create the context of a Web service; • NLP for disambiguating words’ meanings and establishing a context for a set of words; • Matching the users search context with a Web service context by means of a similarity measure. The result of this process will be a ranked list of Web services that match the users search criteria. This context-based matchmaking mechanism provides flexibility by not only searching for exact word matches, 2 but also by looking for synonyms found in a popular lexical database as WordNet (Miller et al., 1990) and making use of its extensive network of relationships between different words (e.g., synonyms, hypernyms, etc.). This work is based on our previous efforts (Sangers et al., 2012), in which we propose a linguistic approach for semantic Web service discovery, yet in our current endeavours we provide more details on the proposed approach and, additionally, we present a genetic algorithm-based method to learn the feature weights. The remainder of this paper is structured as follows. Section 2 discusses related technologies for describing and searching Web services. Next, Section 3 proposes the SWSD framework for discovery of Web services based on keywords. A possible implementation of the framework is described in Section 4. Subsequently, Section 5 evaluates different matching algorithms between the user input and a Web service description. Last, Section 6 concludes the paper and discusses future work. 2. Related Work This section reviews state-of-the-art languages and tools for service discovery. First, a comparison is made between different types of Web service description languages in Section 2.1. Since this paper covers the discovery of semantic Web services, three semantic Web service description languages are described and compared with each other in Section 2.2. Last, several approaches for discovering Web services are discussed in Section 2.3. 2.1. Web Service Description Languages Because Web service discovery depends to a large extent on how Web services are described, an overview of the different types of languages used for describing Web services is first given. These lan- guages differ in models and formalisms used for describing Web services and which Web service properties they cover. This range of languages can go from a simple piece of plain text describing the Web service, to large and complex semantic descriptions of the Web service’s behavior by means of ontologies. Figure 1 shows an overview of the most widely used languages for describing a Web service. Distinction among four description layers is made in order to demonstrate an increasing role of semantics in such languages. Each layer can contribute to the process of matching a service with a given query. Thus, it is necessary to be able to make use of the information encoded into each layer in order to have a good overview on the capability of a service. 3 Figure 1: Web service description languages. To make machines use Web services automatically, their interfaces are commonly defined in languages such as Web Service Description Layer (WSDL) (Christensen et al., 2001), Web Application Description Layer (WADL) (Hadley, 2009), or hRESTs (Kopecký et al., 2009). These languages describe the bind- ings, the operations and their associated input and output data types, and the end points of a Web service. These descriptions enable humans and applications to understand where, when, and what a Web service is expecting from the user on a syntactical level. The top layer depicted in Figure 1 consists of semantic Web service description languages such as Web Service Modeling Ontology (WSMO) (de Bruijn et al., 2005) and OWL-S (Martin et al., 2004). These lan- guages employ ontologies for describing the behavior of a Web service. Concepts, attributes, and relations from existing ontologies and logical expressions can be used to state conditions and effects of a Web service. In order to provide a bridge between the syntactical languages such as WSDL and hRESTs to the semantically enriched languages such as WSMO and OWL-S, middle layer languages were defined. Mi- croWSMO (Kopecký et al., 2009), SA-REST (Lathem et al., 2007), and Semantic Annotations for WSDL (SAWSDL) (Farrell and Lausen, 2007) link the concepts from the semantic descriptions with the data types for the input and output of a Web service, or with its operations, and provide methods to transform data types to semantic concepts and the other way around. The latter two layers are not widely used yet. Therefore most Web service discovery systems use only the descriptions written in languages like WSDL. However, since semantics provide more information about 4 Web services, they can easily be used for Web service discovery. In our view, both types of languages can contribute to the discovery of Web services. 2.2. Semantic Web Service Description Languages Table 1 lists the core differences between three semantic Web service description languages, OWL-S, WSMO and WSMO-Lite. They mainly differ in the syntax they have, the different parts they consist of, and which kind of reasoning they use to describe the logic behavior of a Web service. OWL-S builds on Web Ontology Language (OWL) (Bechhofer et al., 2004) and consists of different ontologies to describe a Web service. The language is defined in order to enable three major tasks which could not be fulfilled with older technologies (WSDL). These are automatic Web service discovery, auto- matic Web service invocation, and automatic Web service composition and interoperation. Because OWL-S uses ontologies to describe Web services, these services and their behavior become machine interpretable and thus tasks such as discovery and composition can be automated. OWL-S makes use of three different ontologies: a Service Profile, which states what the Web service does, a Service Model, which describe how the Web service performs the tasks, and a Service Grounding, which describes how to access the Web service. In recent work, Farrag et al. (2012) introduced an algorithm for mapping WSDL descriptions to OWL-S descriptions by making use of an ontology-based approach. WSMO is a framework for describing Web services and consists of four top-entities: Ontologies, Web services, Goals, and Mediators. Ontologies provide the terminology used by other WSMO elements. Web services describe the capabilities, interfaces, and internal working of Web services. Goals represent user desires, and Mediators provide bridges between different Ontologies, Web services, or Goals to overcome interoperability problems. WSMO uses a specific designed language called WSML (de Bruijn, 2008) and can contain powerful logical formulae to describe the different WSMO elements. It also contains a ground- ing feature to link concepts with WSDL data types so that automatic invocation can be achieved. Table 1: Core differences between OWL-S, WSMO and WSMO-Lite. OWL-S WSMO WSMO-Lite Syntax OWL WSML RDF/XML Parts Service Profile, Service Model, Service Grounding Ontology, Web Service, Goal, Mediator Ontology, Web Service Reasoning Logic language Logic and rule language Logic language 5 WSMO-Lite (Vitvar et al., 2007) was created because of the need for a simple semantic Web service de- scription language. It is therefore a lightweight set of semantic service descriptions written in RDFS (Brick- ley and Guha, 2004) that can be used for annotations of various WSDL elements using SAWSDL annota- tion mechanism. WSMO-Lite only makes use of Ontologies and Web service descriptions and contains no grounding information, which makes it dependent of SAWSDL. The behavior of a Web service is only im- plicitly described by defining just preconditions and effects, so no specific how-questions can be answered. Although these three languages have different properties, they all describe the semantics of a Web service. They provide information about the context domain, behavior and usage of a Web service. Because they use elements from predefined ontologies, those ontologies have to be used together with the semantic Web service descriptions for the discovery and matching of the services. 2.3. Web Service Discovery Engines We can distinguish between two types of approaches for Web service discovery, i.e., discovery based on clustering operation parameters on the one hand, and rich semantics on the other hand. This section continues by elaborating on existing or proposed Web service discovery systems for each of these two different approaches. 2.3.1. Web service discovery based on clustering operation parameters One approach for Web service discovery is by searching for similarities among different WSDL service descriptions, enabling searching for substitutable and composable Web services as similar operations and services can be discovered based on operation parameters. Web service operation semantics can be extracted and employed for discovery purposes. Woogle (Dong et al., 2004) is a Web service search engine that employs clustering techniques for group- ing operation parameters, and for a given query, it searches for similar and/or composable Web service operations. For this, Woogle automatically defines the underlying semantics of WSDL descriptions based on the operation parameters and uses these semantics to match operations. However, if independent ontolo- gies which define the Web service semantics exist, the behavior of a Web service can be known without investigating parameter names and is therefore preferable to use. Operation parameter clustering techniques are also employed in Seekda! (Semantic Technology Insti- tute, 2009), which extracts service semantics from WSDL files, enabling runtime exchange of similar and composable services. Seekda! is part of the Service-Finder (Cefriel et al., 2009) framework, which is a 6 platform for service discovery where information about services is gathered from various sources like Web pages and blogs. The information is automatically added to a semantic model using automatic service an- notation, realizing flexible discovery of services. Service-Finder uses service semantics for discovery and composition, but gathers this information dynamically and hence does not take into account predefined semantics. 2.3.2. Web service discovery using rich semantics Another approach for semantic Web service discovery is the use of predefined ontologies. By identify- ing semantic similarities between ontologies, related semantic Web services can be discovered. To identify semantic similarities, Mediation Spaces can be used (Dietze et al., 2009). These Mediation Spaces me- diate on data-level as well as semantic-level for discovery of related semantic Web services according to ontologies, other semantic Web services, or WSMO Goals. GODO (Gomez et al., 2004) does not search for similar Web services, but instead employs a goal-driven approach. GODO has a repository with WSMO Goals and analyzes a user-described goal (in natural lan- guage). The WSMO Goal with the highest match will be sent to WSMX (DERI Galway, 2008), which is an execution environment for WSMO service discovery and composition. The WSMX environment subse- quently searches for a WSMO Web service that is linked to the given WSMO Goal via WSMO Mediators and returns the WSMO Web service. Although this approach makes good use of the capabilities of the WSMO framework, it cannot be applied for other semantic languages like OWL-S and WSMO-Lite, as they do not have goal representation elements. Bener et al. (2009) proposed an architecture that performs semantic matching of Web services based on both input and output descriptions, as well as preconditions and effects. In the semantic Web service matchmaking process, OWL-S descriptions are assumed (in contrast to the WSMO descriptions which are the focus of our research), and the matchmaking is done at the conceptual level using SWRL rules. Similar to the previous approach, Yang et al. (2008) make use of rules for matching semantic Web services. However, in contrast to the work of Bener et al. (2009), the authors take into account the service context, which is described by means of a context profile in OWL-S. Contexts are gathered through a Java Expert System Shell (JESS)-enabled context elicitation system featuring an ontology-based context model formally describing and acquiring contextual information. Service matching is done based in inputs and outputs and can result in an exact match, a plug-in match, and a subsumed match. 7 Luna et al. (2013) proposed BAX-SET PLUS, which is a multi-agent taxonomy-based method for categorization, search, and retrieval, of semantic Web services described in OWL-S. Here, user-selected concepts from a taxonomy are matched against concepts contained in OWL-S service descriptions. An important difference is that, in contrast to the work presented by the authors, we extract concepts from a natural language query instead of from an existing taxonomy that is browsed by the user. Moreover, our work focuses on WSMO service descriptions in WSML, instead of the OWL-S specifications focused on by Luna et al. (2013). Also, for all aforementioned approaches, no deep linguistic analysis, like proposed in our paper, is performed. Recently, Paulraj and Swamynathan (2012) proposed a method for content-based semantic Web ser- vice discovery. In contrast to the general approach where user queries are matched against OWL-S inputs, outputs, preconditions, and effects (IOPE), the framework allows users to submit free text as input. This alleviates the restrictions put on user queries in that they must be of the same format as that of the IOPEs present in OWL-S. In their work, nouns are extracted from text that is initially unstructured. These nouns are subsequently used for service discovery, after a disambiguation process that makes use of the Word- Net lexical database for determining the meaning of the nouns. Although there are similarities with the work presented in our paper, it should be noted that, similar to many semantics-based Web service discov- ery frameworks, Paulraj and Swamynathan (2012) focus specifically on OWL-S. In contrast, we focus on WSMO and aim for a more universal approach that can be utilized in various semantic Web service de- scription languages (with a few adaptations). Although both in (Paulraj and Swamynathan, 2012) and our work, inputs are initially unstructured, in our current endeavours we provide a means for ranking the results, while Paulraj and Swamynathan (2012) do not implement any form of result ranking. 3. Semantic Web Service Discovery Framework The SWSD framework comprises a keyword-based discovery process for searching Web services that are described using a semantic language. The search mechanism incorporates NLP techniques in order to establish a match between a user search query and a semantic Web service description. Logics-based specifications that are defined in the Web service descriptions are not taken into consideration, but the definitions of concepts stated in other imported ontologies are exploited. In this way, the framework is able to establish a broader search field by also using related concepts from the ontologies for identifying the context in which the Web service is operating. 8 Section 3.1 covers the global architecture of the SWSD framework and a short description of the main components. Section 3.2 describes how the SWSD framework reads semantic Web service descriptions and which elements from those descriptions are being used for the discovery. In Section 3.3 is presented how the context of a semantic Web service is established. Last, Section 3.4 describes two different algorithms to match a user query with a semantic Web service using the query and service contexts. 3.1. Framework Architecture As an input, the SWSD framework requires a set of Web services that are described in semantic lan- guages (e.g., WSMO (de Bruijn et al., 2005), WSMO-Lite (Vitvar et al., 2007), or OWL-S (Martin et al., 2004)). The descriptions are subsequently analyzed, resulting in the extraction of words that could rep- resent the context of the Web services (i.e., the names of the operations, and nouns and verbs stated in non-functional descriptions of concepts or conditions). Next, the extracted words are disambiguated, as multiple senses can be assigned to the same words. Last, the disambiguated words are matched with the disambiguated words from the search query, resulting in a ranked list of Web services. Figure 2 describes the architecture of the SWSD framework. The process is subdivided into three major tasks, i.e., Service Reading, Word Sense Disambiguation (WSD), and Match Making. Service Reading comprises parsing a semantic Web service description, extracting names and non-functional descriptions of used concepts. The WSD task subsequently determines the senses of a set of words. Last, the Match Making step determines the similarity between the different sets of senses, which is ultimately used for ranking the analyzed Web services. 3.2. Semantic Web Service Reader To enable a search engine to look through Web service descriptions written in different languages, it has to have several different Web service description readers, one for each language. In the case of semantically described Web services, the readers must be able to parse a description and extract concepts, attributes, and relations from WSMO, WSMO-Lite, OWL-S, etc. Therefore the first step in the process of searching for semantically described Web services is to implement readers for different languages and formats. A semantic Web service reader must be able to extract various elements out of a Web service description and its used ontologies. In the case of a WSMO Web service, names and non-functional descriptions of elements such as the capabilities and their subelements, conditions and effects, help in understanding the context of the Web service. Thereby, by extracting concepts out of the definedBy statement of those 9 Figure 2: SWSD architecture. elements and searching their non-functional descriptions in an ontology, the context of the Web service can be determined in a more accurate manner than just by using service names. The non-functional descriptions are written in natural language and thus contain a human description of the specified element. By extracting these descriptions, one can thus establish the context of the Web service operations. Before extracting words from a Web service description, the description has to be parsed. Different languages can mean different syntaxes and therefore different parsers are needed. For WSMO a WSML parser like WSMO4J (EU IST and FIT-IT, 2008) can help in this process and for OWL-S and WSMO-Lite, which are written using an OWL (Bechhofer et al., 2004) or RDF (Klyne and Carroll, 2004) language, a parser like Sesame (openRDF.org, 2009) or Jena (HP Labs Semantic Web, 2009) can be used. 10 To find useful words for WSD, the element names and non-functional descriptions must be split into different words. In the case of element names, simply splitting the words when a case transition has occurred is enough, since in most cases they are written in “camel notation” (e.g., HotelBookingWebService). Each word in the sentences found in the non-functional descriptions, must be tagged with the right Part-of-Speech (POS) in order to be useful. Using a POS-tagger, nouns and verbs from sentences are extracted and used for the WSD. 3.3. Word Sense Disambiguation A user can represent its goal by defining a sets of words (nouns and verbs). Because many words have associated multiple meanings (e.g., mouse can be used to represent either an animal or a computer device), disambiguation of word senses helps in finding the correct context. The disambiguation will be applied to the set of words gathered from the user input and from a semantic Web service description, resulting in two sets of disambiguated senses that can be employed for matching. As non-supervised WSD allows disambiguation of words without user interference, we use a variant of the SSI algorithm (Navigli and Velardi, 2005) for retrieving the senses out of a set of words, which is defined as follows: selectedS ense(word) = arg max s j∈senses(word) ∑ sci∈I sim(s j, sci) . (1) The algorithm disambiguates a word (word) based on a previously disambiguated set of words and their related senses. Per word sense s j, a similarity with the senses from the context (sci) is calculated, and the sense with the highest similarity is chosen. Subsequently, the word and its chosen sense are added to context I and the process is iterated until there are no ambiguous words left. Naturally, at the start of the process, a context is not yet established, and hence I = ∅. Hence, we fill context I with all monosemous words, i.e., words that have only one sense, as these do not require any disambiguation. Subsequently, the iterative process of disambiguating polysemous words can be initiated, always targeting the least ambiguous words (i.e., with the least amount of senses) first. For each of the available senses, the algorithm is simulated as if the sense was used as the starting context. Each time a new sense is added to the context, the similarity between the new sense and the context is stored. The sense which creates the highest sum of similarity measures during its simulation run is used for the context initialization. Because previous studies have shown that the method of Jiang and Conrath (Jiang and Conrath, 1997) performs better than other Semantic distance measures (Budanitsky and Hirst, 2001), we employ this 11 method for computing the sense similarities. The measure makes use of the Information Content (IC), which depends on the probability of the occurrence of a particular sense in a large corpus: IC(sense) = 1 log P(sense) . (2) Formula (3) gives the similarity between two senses. Besides the IC, also the Least Common Subsumer (LCS ) between two concepts is used, representing the first ancestor concept that subsumes both senses when going in a bottom-up direction through a hypernym tree of senses. The IC of the LCS is computed and compared with the IC of the two senses: sim(si, s j) = 1 IC(si) + IC(s j) − 2 × IC(LCS (si, s j)) . (3) 3.4. Sense Matching After the disambiguation of all words gathered from the user input and a semantic Web service descrip- tion, one is left with several different sets of senses. For the matching process, each word in the user query is assumed to be equally important and hence the user input contains one set of senses. However, a Web service description can contain words that better represent the context of the Web service than other words. Therefore, after the WSD, several sets of senses, each having a different weight for the matching process, are computed for a Web service. This section starts with a high level view of the matching between the user input and the different levels of information extracted from a Web service description. Next, the Jaccard matcher for matching sets of words or senses is described. Last, a matching approach based on similarities between words or senses is explained. 3.4.1. Level Matching Not only the disambiguated senses are matched, but also words that are not appearing in the used lexicon, as these words can represent important names or concepts for the discovery of Web services. Hence, for matching user input with a semantic Web service description, the user input contains a set of ambiguous words wsu and a set of senses ssu. The Web service description on the other hand contains multiple sets of words mwsw and multiple sets of senses mssw. Because the Web service description, presented in the next section, provides n sets containing words and senses, each having a different importance for the matching process, the final similarity between the user input and the Web service input is computed as a weighted 12 average of the similarities between each set of words mwswi ∈ mwsw and senses msswi ∈ mssw from the Web service description and the set of words and senses from the user input: f inalS im(ssu,mssw,wsu,mwsw) = n∑ i=1 wi × levelS im(ssu,msswi,wsu,mwswi) . (4) Here, the weights wi are established through optimization techniques and sum up to 1 in order to make sure that the final similarity between the user query and a Web service description has a range between 0 and 1. For each set of words and senses from the Web service description, the system employs two mea- sures, i.e., one for sense matching (defined in Formulas (6) and (8)), and one for the matching of non- disambiguated words (see Formulas (7) and (11)). These measures range between 0 (no match) and 1 (exact match) and are combined into a single measure using a weighted average: levelS im(ssu, ssw,wsu,wsw) = wsense × senseS im(ssu, ssw) + wword × wordS im(wsu,wsw) , (5) which is used n times in Formula (4), where msswi = ssw and mwswi = wsw. The weights corresponding to the sense similarity and the word similarity are, as with the final similarity, established by means of optimization techniques and must sum up to 1. 3.4.2. Jaccard Matching The Jaccard matcher, which is employed for matching sets of words or senses, makes use of the Jaccard Index (Jaccard, 1901). This method is often used for computing set similarity and thus compares different sense sets: senseS im(ssu, ssw) = |ssu ∩ ssw| |ssu ∪ ssw| . (6) Subsequently, we calculate a similarity coefficient by dividing the number of senses appearing in both sets by the total number of senses in both sets. This results in a percentage of exact matching items and can also be applied for matching the words that could not be disambiguated: wordS im(wsu,wsw) = |wsu ∩ wsw| |wsu ∪ wsw| . (7) 3.4.3. Similarity Matching Normally, for the calculation of similarity values, only perfect matching items are used. In order to additionally take into consideration partial matches, the similarity matcher makes use of a similarity-based approach for matching different sets of senses or non-disambiguated words, expressing close relatedness with values approaching 1, and non-relatedness with values close to 0. 13 Sense set similarities are calculated using the same similarity function as in WSD. The similarity be- tween the user set of senses ssu and a Web service set of senses ssw is computed as follows: senseS im(ssu, ssw) = ∑ su∈ssu senseS core(su, ssw) |ssu|+ |ssw| + ∑ sw∈ssw senseS core(sw, ssu) |ssu|+ |ssw| , (8) where the average of the similarity between each sense su from the user set of senses, and the Web service set of senses is computed. The average of the similarity between each sense sw from the Web service set of senses, and the user set of senses is added to that to provide a symmetric match. The similarity between a sense sa and a set of senses ssb is calculated by taking the maximum similarity between the sense and one of the senses sb from the other set, i.e.: senseS core(sa, ssb) = max sb∈ssb senseNorm(sa, sb) . (9) The similarities calculated with Formula (3) range between 0 and infinity, yet we prefer a range between 0 and 1 for quantifying the matching degree. Hence, we employ a logarithmic function in order to transform the values of the similarity: senseNorm(sa, sb) = 1 − e−sim(sa,sb) . (10) Here, exact similar senses will have 1 as resulting similarity and a total mismatch between senses will result in 0. As we are unable to determine the senses of some words, we fall back to non-semantic measurements of similarities for matching the sets of non-disambiguated words. For this, we employ the Levenshtein distance metric (Levenshtein, 1966), which calculates the total number of operations that need to be done in order to transform one word to another. The similarity between two sets of words is subsequently calculated similarly to the comparison of two sense sets, yet now the Levenshtein distance is applied instead of the similarity function from WSD. We calculate the similarity between the user set of words wsu and a Web service set of words wsw as: wordS im(wsu,wsw) = ∑ wu∈wsu wordS core(wu,wsw) |wsu|+ |wsw| + ∑ ww∈wsw wordS core(ww,wsu) |wsu|+ |wsw| , (11) where the similarity between a word and a set of words, wordS core(wa,wsb) is computed as: wordS core(wa,wsb) = max wb∈wsb wordNorm(wa,wb) , (12) and the Levenshtein distance is used in the following way for comparing two words wa and wb: wordNorm(wa,wb) = max(0,1 − 2 × levenshtein(wa,wb) maxLength(wa,wb) ) . (13) 14 Here, maxLength(wa,wb) denotes the number of tokens of the longest word that is being compared. If this formula returns a negative value, which means that the amount of changes that should be made exceed half of the length of the largest word, a value of 0 will be used to indicate a total mismatch. 4. Semantic Web Service Discovery Engine This section describes the SWSD engine, which is an implementation of the SWSD approach and allows users to search for Web services on an existing repository by defining a set of keywords. The steps required for this implementation are closely related to the steps stated for the SWSD framework stated in the previous section. However, this is one possible implementation of the framework, using specific languages and external libraries to provide a discovery engine. Section 4.1 introduces the SWSD engine as an implementation of the SWSD framework. Section 4.2 describes how the SWSD engine reads WSMO descriptions and which elements from those descriptions are being used for the service discovery. How the context of a semantic Web service is established is presented in Section 4.3. Last, Section 4.4 describes how matching algorithms are used for computing the similarity between the user input and semantic Web service information. 4.1. SWSD Engine At the moment, the SWSD engine can only apply a search on Web services which are annotated using the WSMO (de Bruijn et al., 2005) framework as WSMO is a powerful framework that can contain very wide information about Web services. So, only a WSMO Web service reader and a WSMO Ontology reader have been implemented, but based on the modularity of the implementation, the engine can be extended with readers that can parse other semantic Web service languages. To read the WSMO files, WSMO4J (EU IST and FIT-IT, 2008) is used. WSMO4J is an API and a reference implementation for building semantic Web Services and Semantic Business Process applications based on WSMO. By using WSMO4J, WSMO files can be parsed, read, and written. For the WSD, Word- Net (Miller et al., 1990) is used through two WordNet API’s, in order to find senses belonging to words. The Java WordNet Library (JWNL) (Walenz and Didion, 2011) is used for the morphological analysis of each word, and the MIT Java WordNet Interface (JWI) (Finlayson, 2012) is subsequently employed for retrieving WordNet synsets. Next, JWordNetSim (The University of Sheffield, 2009) is used to calculate the similarity between two WordNet senses because it contains an implementation of the Jiang and Conrath formula, which we proposed to use in the SWSD framework. For the part-of-speech tagging, the Stanford 15 Figure 3: Result presentation example. parser (The Stanford Natural Language Processing Group, 2009) is used. The overall implementation is made in Java, due to the availability of external packages written in this language for WSD and WSMO parsing and reading. Figure 3 shows the user interface of the SWSD engine. The user can fill in a comma-separated list of words representing his goal and click the search button. The system will then check if each word given by the user can be a noun, verb, both noun or verb, or something else. This is needed since the WSD process can only be used for a set of words containing the same POS tag. For the words that do not have a clear POS, the user will be asked to select the appropriate POS to use for the search. After that, the system will do the search and propose a list of Web services, which are ranked by their similarity with the user input. Each item in the list contains the name of the Web service, its non-functional description, its similarity score and a normalized similarity score that has a range from 0 to 1 (computed by dividing the similarity score with the highest similarity score). 16 namespace { _"http://api.google.com/GoogleSearch#", dc _"http://purl.org/dc/elements/1.1#" } webService GoogleSearchWebService nonFunctionalProperties dc#description hasValue "Web service that searches for Web sites containing words that matches given search words." endNonFunctionalProperties importsOntology _"http://api.google.com/http_api.google.com_GoogleSearch#GoogleOntology.wsml" capability DoSearch postcondition nonFunctionalProperties dc#description hasValue "If there is an input with at least a query, the output must be a GoogleSearchResult. Some values in the input, must also be returned in the output." endNonFunctionalProperties definedBy ?input[hasQuery hasValue ?query] memberOf GoogleSearchParameterSet and ?query memberOf Query implies exists ?output (?output memberOf GoogleSearchResult and ?output[hasSearchQuery hasValue ?query]) . Name and non-functional description of the Web service are extracted. Nouns and verbs are used for WSD. Identifier of the ontology where concepts are defined. Non-functional descriptions of properties of the capabilities are extracted. Nouns and verbs are used for WSD. Concepts uses from ontologies are extracted and searched in their ontologies. Figure 4: A WSMO Web service information extraction example. namespace { _"http://api.google.com/GoogleSearch#", dc _"http://purl.org/dc/elements/1.1#" } ontology SimpleGoogleOntology concept GoogleSearchResult nonFunctionalProperties dc#description hasValue "Result of a Google search method." endNonFunctionalProperties hasSearchQuery ofType (1) Query nfp dc#description hasValue "The query for the search request." endnfp concept Query hasText ofType (1) _string nfp dc#description hasValue "String containing keywords representing the search query." endnfp Name and non-functional description of the concept are extracted. Nouns and verbs are used for WSD. Name and non-functional description of the attribute are extracted. Nouns and verbs are used for WSD. Concepts related via a property are also searched in the ontology. Figure 5: A WSMO ontology information extraction example. 4.2. Semantic Web Service Reader For reading WSMO files, an implementation of the WSMO4J API is used, resulting in an engine that can extract different types of information from WSMO Web services and WSMO ontologies. Because WSMO Web services and ontologies use different structures, two different readers must be used. Figure 4 shows an example WSMO Web service describing the Google Search Web service. The parts surrounded by rectangles are extracted by the system to establish the context of the Web service. First the name and non-functional description of the Web service are analyzed for extracting relevant words (nouns/verbs). Then the words from the non-functional descriptions of the properties of the capabilities are extracted and the reader will search in the logical formulae, written after each definedBy statement, for concepts used from external ontologies, which are stated after the importsOntology statement. 17 After reading the WSMO Web service file, the found concepts are used to search in ontologies to find their description. Figure 5 shows an example WSMO Ontology containing some concepts used in the Web service description from Figure 4. Based on a full identifier, the reader can search for a concept. If a concept is found, the non-functional definition, attributes and related concepts can be used for WSD. In total the engine creates seven different levels of information about the Web service. Each of these levels has a different amount of importance to the matching process. Those amounts are expressed in weights, summing up to 1. The different levels and their associated weights are: • Non-functional description and name of the Web service, 0.210, (direct relation, from Web service); • Non-functional descriptions of properties of capabilities of the Web service, 0.167, (direct relation, from Web service); • Non-functional descriptions and names of concepts used by Web service, 0.191, (direct relation, from ontology); • Non-functional descriptions and names of attributes of concepts used by the Web service, 0.008, (indirect relation, from ontology); • Non-functional descriptions and names of concepts related via attributes with concepts used by the Web service, 0.153, (indirect relation, from ontology); • Non-functional descriptions and names of superconcepts of the concepts used by the Web service, 0.158, (semi-direct relation, from ontology); • Non-functional descriptions and names of subconcepts of the concepts used by the Web service, 0.113, (semi-direct relation, from ontology). The weights are established by making use of a Genetic Algorithm, which is a specific technique to find approximate solutions for optimization problems. In our case, we want to find a set of weights which opti- mize the search results. Genetic Algorithms mimic an evolution of individuals in a certain population. Each individual is represented with a chromosome consisting of data that can be recombined and mutated during next generations. We choose to represent the individuals using a chromosome consisting of nine variables, the seven weights used for stating the importance of the information gathered from a Web service, and two weights for setting the trade-off between the sense similarity and the lexical representation similarity. 18 From each generation, the best individuals will survive and create an offspring. To define which indi- viduals are good enough to survive, a fitness function is used. In our case, this is a function that takes the average precision values for each query used in our two data sets described in Section 5. Last, the genetic algorithm has been executed with a population size of 100 and the number of generations set to 50. As can be seen from the weights, the information directly related to the Web service is the most im- portant information for the matching process. This information consists of the name and non-functional descriptions of the Web service, its capabilities, and concepts used. Less important for the matching are the non-functional descriptions and names of the super- and subconcepts of the used concepts by the Web service. Least important and thus given a low weight is the information that is indirectly related to the Web service description. This information is represented by attributes and concepts related to the concepts used by the Web service. Using these weights, the information that is semantically closest to the core Web service description will have the most impact for the matching process. The names and non-functional descriptions of the entities returned from these two readers will then go through a NLP step. Nouns and verbs are extracted from the non-functional descriptions using the Stanford POS-tagger and words are split if they consist of case-transitions. A sentence like the non-functional description of the Google Search Web service presented in Figure 4 (“Web service that searches for Web sites containing words that matches given search words.”) will generate a set of nouns {Web, service, sites, words} and verbs {searches, containing, matches, given}. The name of the Google Search Web service, GoogleSearchWebService, will be split into the set of nouns {Google, Web, Service} and verbs {Search}. So instead of using compound words and whole sentences, the system only uses nouns and verbs extracted from them (as only these can be found in WordNet). 4.3. Word Sense Disambiguation For WSD the system makes use of two different API’s of WordNet. By using them, words can be found in WordNet by giving a lexical representation of the word and its POS. These words will have different senses and each sense has its own identifier, a WordNet synset. The system has to find the synsets based on a set of lexical representations of words employed by users or a Web service description. For establishing a starting context, WordNet will be used to find monosemous words. If no monosemous word is found, the word with the least synsets will be used to simulate the best starting context. For each synset, the similarity of the other words will be computed as if this synset was the starting context. Each time a new synset is added to the context, the similarity between the new synset and the context is stored. 19 The synset which creates the highest sum of similarity measures during its simulation is used as the real starting context. Once having a context, the SSI algorithm will be used to find the synsets of the other words. The similarity is measured using an implementation of the Jiang and Conrath method (Greenwood, 2007). It can handle pairs of WordNet synsets as input and will return the similarity between them. For each word, the synset with the highest similarity to the context, will be added to the context. 4.4. Sense Matching The synsets resulting from the disambiguation process must be matched in order to get a final similarity measure. Because the information in a Web service description has different levels of importance in the matching process, several sets of synsets belonging to a Web service will come out after the disambiguation phase. Thus one set of synsets coming from the user input must be matched with several sets of synsets coming from a Web service. Each set from the Web service will have a weight representing the value of its information for the matching process. For example, the synsets found from the non-functional description of the Web service are more important during the matching, than the names of related concepts. These weights are determined using a Genetic Algorithm and must sum up to 1 in order to get a final match with the range [0 . . . 1]. To overcome the fact that words that are not present in WordNet are not used in the matching process, every lexical representation of the extracted words from the user input, which are not in WordNet, is being compared to the lexical representations of the extracted words from a Web service description that are not in WordNet. To provide a flexible matching, the words that are not in WordNet are first being lemmatized before they are matched. This means that for example the word searching will also match the word searched. By combining the synset similarity value and the lexical representation similarity, using a weighted average, we determine a final similarity value. The weights have been optimized by the genetic algorithm described in Section 4.2 and set to 44/100 for the synset similarity and 56/100 for the lexical representation similarity. The value of the synset similarity is higher than the lexical representation similarity as they provide for a richer information comparison. 5. Evaluation This section covers the evaluation of different matching algorithms that can be used for semantic Web service discovery. The algorithms described in Section 3.4 are implemented in the SWSD engine and are 20 evaluated by using a set of predefined queries and sets of preferred Web services related to one of the queries. Section 5.1 explains how the testing is done. In Section 5.2 the results of those tests are presented. Section 5.3 concludes which algorithms provides the best matching for discovery of semantic Web services. 5.1. Experimental Setup For testing the algorithms provided in Section 3.4, we make use of a repository of 35 WSMO Web ser- vice descriptions stored in WSML format. As for WSMO Web service descriptions there are no existing data sets readily available such as for instance the OWLS-TC 4.0 OWL-S service retrieval test collection (Klusch et al., 2010), we have manually built 35 WSMO service descriptions. For evaluation purposes, we have defined 61 test queries in order to analyze the outcomes of each of the algorithms. These queries represent possible sets of keywords a user could use when searching for a Web service. For each query, we have defined a list of preferred Web services (present in the repository of the SWSD engine) that should be returned with a high ranking, as they all are significantly related to the queries. Besides the algorithms described in Section 3.4, a simple matching algorithm that does not make use of natural language processing is added for the evaluation. The algorithm employs the Jaccard matching algorithm only for lexical representations of the query keywords and the extracted words. The algorithm makes use of the different information levels of a Web service and can therefore be used for testing whether there is an added value in using NLP in the process of semantic Web service discovery. The testing is done with usage of precision and recall metrics. For every query, the precision and recall values for each of the algorithms are computed according to the list of Web services they provide using that particularly query as an input. This list is a ranked list, based on the similarity values calculated by the used matching algorithm, of all the Web services that are in the repository and will be compared with the list of preferred Web services stated for the query. The precision and recall values are computed by traversing the provided list of Web services, according to a query and matching algorithm, from top to bottom. If a Web service y j is stated as preferred, then the Web service is classified as 1. If it is not preferred, the Web service is classified as 0. Using this classification, the precision pi and recall ri for a position i in the list of provided Web services can be calculated as follows, where k is the number of preferred Web services: 21 pi = ∑i j=1 y j i , (14) ri = ∑i j=1 y j k . (15) Figure 6 shows the calculation of precision and recall for a given list of classified Web services. For each preferred Web service – classified as 1 and visualized with a blue dot – the precision and recall values are calculated. Each time a Web service that is not stated as preferred (classified as 0 and visualized with a yellow dot) is found, the precision will drop. The PR-graph can show when it takes a long time before another preferred Web service is found in the list. This is the case when the precision drops heavily. 5.2. Experimental Results To test the performances of the three matching algorithms, we divide our 61 predefined queries into two types. In total, 33 queries are used for measuring the matching performance of the algorithms for searches for Web services that are present in the repository. The remaining 28 queries are used for measuring the performance of queries for Web services that are not present in the repository. In the latter set of queries, a number of similar Web services from the repository are defined in order to test performance of similar Web service discovery when a specified service does not exist in the repository. Both sets of queries are divided into two parts, a training set and a test set. The training set, consisting of 40 queries, is used for training the weights that are used by the matching algorithms, while the test set, having 21 queries, is used for testing the performance of the matching algorithms. Testing with 21 queries and three matching algorithms generates 63 PR-graphs. Visualizing this amount of graphs is not very insightful and hence we create PR-graphs consisting of average precision values for the recall points. This enables us to relate all the different algorithms at once. However, the testing is done with Figure 6: Precision and recall calculation example. 22 lists of preferred Web services that can vary in the number of Web services they consists of. For testing, lists that contain two to five preferred Web services have been used. Because these variations in number of Web services cause different recall values, average precision values was only be calculated for queries that have the same amount of preferred Web services. Hence, we analyze eight different PR-graphs that visualize the performances of the different match- ing algorithms. The four PR-graphs that are depicted in Figure 7, show the average results for the exact matching tests. For each of the four different numbers of preferred returned Web services (n) a PR-graph is created. The four PR-graphs that are shown in Figure 8, show the average results for the approximate matching tests. From the different PR-graphs that are shown in Figure 7, we can make two observations. First, in most cases, the Jaccard algorithm shows a higher precision for most recall values than the simple and the similarity algorithm. Second, all algorithms require about the same precision for providing a full recall. This means that in order to provide all the preferred Web services to the user, the algorithms are required to display about the same amount of Web services. However, according to the fact that the Jaccard algorithm provides a higher precision for a lower recall, the Jaccard algorithm provides at least some of the preferred Web services in an earlier stage to the user than the others. It can therefore be seen as the best algorithm to discover exact matching Web services. From the different PR-graphs that are shown in Figure 8, we can make the observation that the similarity algorithm performs overall better for discovery of similar Web services than the Jaccard and the simple matching algorithm, as in most of the cases the PR-graph lines of the similarity algorithm are above the lines of the Jaccard algorithm and the simple matching algorithm. 5.3. Summary In order to test the performance of the three matching algorithms explained in Section 3.4, we have performed 61 tests, of which 31 tests were done to measure the performance of the algorithms according to discovery of exact matching Web services, and 28 tests were done to measure the performance of the algorithms according to discovery of similar Web services. Each test consisted of a set of keywords used as a query and a set of Web services that were preferred to be provided to the user as early as possible. Based on our evaluation, we conclude that the Jaccard matching algorithm is the best method for discovering exact matching Web services, whereas the similarity matching algorithm is the best method for discovering similar Web services. 23 n=2 n=3 n=4 n=5 Figure 7: PR-Graphs for discovery of exact matching services. n=2 n=3 n=4 n=5 Figure 8: PR-Graphs for discovery of similar services. 24 6. Conclusions and Future Work The SWSD framework proposes a keyword-based discovery process for searching Web services that are described using semantically enriched annotations done by means of semantic languages for service description. It makes an intensive use of natural language processing techniques and a WordNet-based similarity measure for matching keywords. By using an approach with two different similarity functions, one for lexical similarities and one for semantic similarities, standard lexical matching algorithms as well as semantic matching algorithms, using ontologies, can be applied for discovery of semantic Web services. The SWSD engine can search for WSMO Web services based on user search keywords. A matching score is computed based on the similarity between the words in the user query and a Web service description. Experiments have been done to test the performance of three different matching algorithms. The Jaccard matching algorithm performs best for discovering exact matching Web services, while matching using a similarity approach gives the best results for finding similar Web services. As a future work, the SWSD engine could be extended in such a way that it has the ability to read more annotation formats, e.g., WSMO-Lite. In this case, not only WSMO Web services can be discovered, but also Web services described using other semantic languages than WSMO. Also, additional information about a Web service (e.g., signature of its operations, messages being exchanged, etc.) could be retrieved from its WSML specification. With this information, the context of the Web service can be described in more detail. References Bechhofer, S., van Harmelen, F., Hendler, J., Horrocks, I., McGuinness, D. L., Patel-Schneider, P. F., Stein, L. A., 2004. OWL Web Ontology Language Reference – W3C Recommendation 10 February 2004. From: http://www.w3.org/TR/owl-ref/. Bener, A. B., Ozadali, V., Ilhan, E. S., 2009. Semantic Matchmaker with Precondition and Effect Matching Using SWRL. Expert Systems with Applications 36 (5), 9371–9377. Brickley, D., Guha, R., 2004. RDF Vocabulary Description Language 1.0: RDF Schema – W3C Recommendation 10 February 2004. From: http://www.w3.org/TR/rdf-schema/. Budanitsky, A., Hirst, G., 2001. Semantic Distance in WordNet: An Experimental, Application-Oriented Evaluation of Five Mea- sures. In: Workshop on WordNet and Other Lexical Resources at 2nd Meeting of the North American Chapter of the Association for Computational Linguistics (NAACL 2001). Association for Computational Linguistics, pp. 29–34. Cefriel, Seekda!, Ontoprise, University of Sheffield, 2009. Service-Finder. From: http://www.service-finder.eu/. Christensen, E., Curbera, F., Meredith, G., Weerawarana, S., 2001. Web Services Description Language (WSDL) 1.1 – W3C Note 15 March 2001. From: http://www.w3.org/TR/wsdl. 25 de Bruijn, J., 2008. D16 The WSML Specification – WSML Working Draft 2008-08-08. From: http://www.wsmo.org/TR/ d16/. de Bruijn, J., Bussler, C., Domingue, J., Fensel, D., Hepp, M., Keller, U., Kifer, M., Konig-Ries, B., Kopecky, J., Lara, R., Lausen, H., Oren, E., Polleres, A., Roman, D., Scicluna, J., Stollberg, M., 2005. Web Service Modeling Ontology (WSMO) – W3C Member Submission 3 June 2005. From: http://www.w3.org/Submission/WSMO/. DERI Galway, 2008. Web Service Execution Environment. From: http://www.wsmx.org/. Dietze, S., Gugliotta, A., Domingue, J., 2009. Exploiting Metrics for Similarity-based Semantic Web Service Discovery. In: IEEE 7th International Conference on Web Services (ICWS 2009). IEEE Computer Society, pp. 327–334. Dong, X., Halevy, A., Madhavan, J., Nemes, E., Zhang, J., 2004. Similarity Search for Web Services. In: 30th International Conference on Very Large Data Bases (VLDB 2004). Vol. 30. pp. 372–383. EU IST, FIT-IT, 2008. WSMO4J API. From: http://wsmo4j.sourceforge.net/. Farrag, T. A., Saleh, A. I., Ali, H. A., 2012. Towards SWSs Discovery: Mapping from WSDL to OWL-S Based on On- tology Search and Standardization Engine. IEEE Transactions on Knowledge and Data Engineering. To appear (DOI: 10.1109/TKDE.2012.25). Farrell, J., Lausen, H., 2007. Semantic Annotations for WSDL and XML Schema – W3C Recommendation 28 August 2007. From: http://www.w3.org/TR/sawsdl. Finlayson, M., 2012. JWI: The MIT Java WordNet Interface. From: http://projects.csail.mit.edu/jwi/. Gomez, J. M., Rico, M., Garcia-Sanchez, F., Bejar, R. M., Bussler, C., 2004. GODO: Goal Driven Orchestration for Semantic Web Services. In: 1st Workshop on Web Services Modeling Ontology Implementations (WIW 2004). Vol. 113. CEUR Workshop Proceedings. Greenwood, M., 2007. JWordNetSim. From: http://nlp.shef.ac.uk/result/software.html. Hadley, M. J., 2009. Web Application Description Language – W3C Member Submission 31 August 2009. From: http://www. w3.org/Submission/wadl/. Hikimpour, F., Sell, D., Cabral, L., Domingue, J., Motta, E., 2005. Semantic Web Service Composition in IRS-III: The Structured Approach. In: 7th IEEE International Conference on E-Commerce Technology (CEC 2005). IEEE Computer Society, pp. 484– 487. HP Labs Semantic Web, 2009. Jena. From: http://jena.sourceforge.net/. Jaccard, P., 1901. Étude Comparative de la Distribution Florale dans une Portion des Alpes et des Jura. Bulletin de la Société Vaudoise des Sciences Naturelles 37, 547–579. Jiang, J., Conrath, D., 1997. Semantic Similarity Based on Corpus Statistics and Lexical Taxonomy. In: International Conference Research on Computational Linguistics (ROCLING X). pp. 19–33. Keller, U., Lara, R., Lause, H., Polleres, A., Predoiu, L., Toma, I., 2005. Semantic Web Service Discovery – WSMX Working Draft - October 3, 2005. From: http://www.wsmo.org/TR/d10/v0.2/d10.pdf. Klusch, M., Kapahnke, P., Fries, B., Khalid, M. A., Vasileski, M., 2010. OWLS-TC 4.0 – Release 21 September 2010. From: http://semwebcentral.org/projects/owls-tc/. Klyne, G., Carroll, J. J., 2004. Resource Description Framework (RDF): Concepts and Abstract Syntax – W3C Recommendation 10 February 2004. From: http://www.w3.org/TR/rdf-concepts/. Kopecký, J., Vitvar, T., Fensel, D., Gomadam, K., 2009. D12v0.1 hRESTS & MicroWSMO – CMS WG Working Draft 10 March 26 2009. From: http://cms-wg.sti2.org/TR/d12/v0.1/. Lathem, J., Gomadam, K., Sheth, A. P., 2007. SA-REST and (S)mashups: Adding Semantics to RESTful Services. In: 1st Interna- tional Conference on Semantic Computing (ICSC 2007). IEEE Computer Society, pp. 469–476. Levenshtein, V. I., 1966. Binary Codes Capable of Correction Deletions, Insertions, and Reversals. Soviet Physics Doklady 10 (8), 707–710. Luna, J. A. G., Pardo, I. D. T., Builes, J. A. J., 2013. Recent Progress in Data Engineering and Internet Technology. Vol. 156 of Lecture Notes in Electrical Engineering. Springer, Ch. BAX-SET PLUS: A Taxonomic Navigation Model to Categorize, Search and Retrieve Semantic Web Services, pp. 359–364. Martin, D., Burstein, M., Hobbs, J., Lassila, O., McDermott, D., McIlraith, S., Narayanan, S., Paolucci, M., Parsia, B., Payne, T., Sirin, E., Srinivasan, N., Sycara, K., 2004. OWL-S: Semantic Markup for Web Services – W3C Member Submission 22 November 2004. From: http://www.w3.org/Submission/OWL-S/. Miller, G. A., Beckwith, R., Fellbaum, C., Gross, D., Miller, K., 1990. Introduction to WordNet: An On-Line Lexical Database. International Journal of Lexicography 3 (4), 235–244. Navigli, R., Velardi, P., 2005. Structural Semantic Interconnections: a Knowledge-Based Approach to Word Sense Disambiguation. In: IEEE Transactions on Pattern Analysis and Machine Intelligence. Vol. 27. IEEE Computer Society, pp. 1075–1086. openRDF.org, 2009. Sesame. From: http://www.openrdf.org/. Paulraj, D., Swamynathan, S., 2012. Advanced Computing, Networking and Security. Vol. 7135 of Lecture Notes in Computer Science. Springer, Ch. Content Based Service Discovery in Semantic Web Services Using WordNet, pp. 48–56. Sangers, J., Frasincar, F., Hogenboom, F., Hogenboom, A., Chepegin, V., 2012. A Linguistic Approach for Semantic Web Service Discovery. In: Casillas, J., Martı́nez-López, F. J., Corchado, J. M. (Eds.), First International Symposium on Management Intelligent Systems (IS-MiS 2012). Vol. 171 of Advances in Intelligent Systems and Computing. Springer, pp. 131–142. Semantic Technology Institute, 2009. Seekda! From: http://seekda.com/. The Stanford Natural Language Processing Group, 2009. Stanford Log-linear Part-Of-Speech Tagger. From: http://nlp. stanford.edu/software/tagger.shtml. The University of Sheffield, 2009. Pure Java WordNet Similarity Library. From: http://nlp.shef.ac.uk/result/software. html. Vitvar, T., Kopecky, J., Fensel, D., 2007. WSMO-Lite: Lightweight Semantic Descriptions for Services on the Web. In: 5th IEEE European Conference on Web Services (ECOWS 2007). IEEE Computer Society, pp. 77–86. Walenz, B., Didion, J., 2011. JWNL: Java WordNet Library. From: http://sourceforge.net/projects/jwordnet/. Yang, S. J. H., Zhang, J., Chen, I. Y. L., 2008. A JESS-Enabled Context Elicitation System for Providing Context-Aware Web Services. Expert Systems with Applications 34 (4), 2254–2266. 27 
work_emoco2eizzerpjy6bp5hn3agmu	----	An integrated fuzzy AHPﾖELECTRE methodology for environmental impact assessment An integrated fuzzy AHP–ELECTRE methodology for environmental impact assessment Tolga Kaya a,⇑, Cengiz Kahraman b a Istanbul Technical University, Department of Management Engineering, 34367 Macka, Istanbul, Turkey b Istanbul Technical University, Department of Industrial Engineering, 34367 Macka, Istanbul, Turkey a r t i c l e i n f o Keywords: Multicriteria Environmental impact assessment AHP ELECTRE Fuzzy dominance relation a b s t r a c t Environmental impact assessment (EIA) is the assessment of the possible impact that a proposed plan or project may have on the environment, together consisting of the natural, social and economic aspects. The aim of this paper is to propose an environmental impact assessment methodology based on an inte- grated fuzzy AHP–ELECTRE approach in the context of urban industrial planning. In the proposed meth- odology the criteria weights are generated by a fuzzy AHP procedure. The fuzzy set theory is a perfect means for modeling uncertainty or imprecision arising from human mental phenomena. The usage of fuzzy sets in describing uncertainties and vagueness in different environmental factors simplifies the complex structure of EIA. A fuzzy outranking methodology, fuzzy ELECTRE is used to assess the environ- mental impact generated by the six different industrial districts which were predicted to shape the future industrial structure of Istanbul metropolitan area. Finally, a fuzzy dominance relation (FDR) methodology is used to rank the alternatives from the most risky to the least. A sensitivity analysis is also provided. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Environmental impact assessment (EIA) is the process of identi- fying, predicting, evaluating and mitigating the biophysical, social, and other relevant effects of development proposals prior to major decisions being taken and commitments made (International Asso- ciation for Impact Assessment (IAIA), 1999). EIA is an important procedure for ensuring that the likely effects of new development on the environment are fully understood and taken into account before the development is allowed to proceed. EIA enables envi- ronmental factors to be given due weight, along with economic or social factors, when planning applications are being considered. If properly carried out, it benefits all those involved in the planning process. For the planning authority and other public bodies with environmental responsibilities, environmental impact assessment provides a basis for better decision making. Decision making in environmental problems can be complex and multifaceted due to the inherent trade-offs between sociopo- litical, environmental, ecological, and economic factors. Moreover, environmental decisions may involve many different stakeholders with different priorities or objectives. The selection of appropriate remedial strategies for contaminated sites, land use planning, and regulatory processes often involves multiple criteria such as the distribution of costs and benefits, environmental impacts for dif- ferent populations, safety, ecological risk, or human values. Consid- erable research in the area of multicriteria decision making (MCDM) has made available practical methods for applying scien- tific decision theoretical approaches to complex multicriteria prob- lems (Kiker, Bridges, Varghese, Seager, & Linkovjj, 2005). Analytical hierarchy process (AHP) (Ramanathan, 2001; Tran et al., 2002), ELECTRE (Joerin & Musy, 2000; Rogers & Bruen, 1998), PROM- ETHEE (Al-Rashdan, Al-Kloub, Dean, & Al-Shemmeri, 1999; Vaillan- court & Waaub, 2002), and multiple attribute utility theory (MAUT) (Prato, 2003; Store & Kangas, 2001) are among the MCDM tech- niques which are widely used in environmental decision making issues. Environmental assessments often have to deal with attributes difficult to define and components that may involve both quantita- tive and qualitative factors. EIA is a complex issue not only because of its wide scope but also because of the wide range of attributes that bear on its assessment. In terms of scope, an assessment may cover subjective and vague decisions that may affect various interest groups or stakeholders each with their own demands and needs. In view of these difficulties, methods based on fuzzy lo- gic may be quite useful in undertaking difficult assessment proce- dures. The fuzzy set theory (Zadeh, 1965) was introduced to express the linguistic terms in decision-making process in order to resolve the vagueness, ambiguity and subjectivity of human judgment. Fuzzy methods may be very useful in dealing with com- plex and ill defined environmental decision making problems (Mendoza & Prabhu, 2003). 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.01.057 ⇑ Corresponding author. E-mail address: kayatolga@itu.edu.tr (T. Kaya). Expert Systems with Applications 38 (2011) 8553–8562 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.01.057 mailto:kayatolga@itu.edu.tr http://dx.doi.org/10.1016/j.eswa.2011.01.057 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa The ELECTRE method for choosing the best action from a given set of actions was devised in 1965. The acronym ELECTRE stands for ‘‘ELimination Et Choix Traduisant la REalite (ELimination and Choice Expressing the REality)’’. ELECTRE is a well known MCDM method that has a history of successful real world applications. It has been applied in past to various types of decision-making situ- ations. ELECTRE requires an input of criteria evaluations for the alternatives, called decision matrix, preference information, ex- pressed as weights, thresholds, and other parameters (Sevkli, 2009). All the ELECTRE-type methods involve two major proce- dures: the modeling of preferences with outranking relations, fol- lowed by an exploitation procedure. The outranking relation of Ak ? Al says that Ak outranks Al, if Ak is at least as good as Al on a majority of criteria and this result is not significantly based on any of the other criteria. ELECTRE methods can operate with one or several crispy or fuzzy outranking relations (1990; Benayoun, Roy, & Sussmann, 1966; Montazer, Saremi, & Ramezani, 2009; Roy, 1985; Shanian & Savadogo, 2006). In fuzzy ELECTRE, linguistic preferences can easily be converted to fuzzy numbers. For the determination of the relative importance of evaluation criteria, fuzzy AHP can be used since it is based on pairwise comparisons and allows the utilization of linguistic vari- ables. From paired comparisons a relative scale of measurement is derived. Although the pairwise comparison approach is demand- ing in terms of solicited input from the experts, it offers maximum insight, particularly in terms of assessing consistency of the ex- perts’ judgment. This technique is ideal for closer examination of a selected set of EIA criteria in urban industrial planning context. In this study, an integrated fuzzy AHP–ELECTRE methodology is proposed for EIA. In the proposed methodology, the weights of the assessment criteria are determined by a fuzzy AHP procedure. With the proposed methodology, six urban districts which were predicted to be among the developing industrial areas according to the Istanbul 2006 Metropolitan Plan will be evaluated in terms of EIA. Finally, using a fuzzy dominance relation (FDR) approach, the most risky industrial districts will be determined in order to provide a base for urban industrial rehabilitation strategies. The rest of the paper is organized as follows: In Section 2, a lit- erature review about multiple attribute EIA is briefly given. In the third section, an integrated fuzzy AHP–ELECTRE methodology is presented. In Section 4, following the determination of the selec- tion criteria and alternatives, the proposed methodology is applied to an urban industrial EIA problem. In Section 5, a sensitivity anal- ysis is provided. Finally, in the last section, conclusions and sugges- tions for further study are given. 2. Literature review EIA is a process of having the ultimate object of providing deci- sion makers with an indication of the likely consequences of their actions. In other words, EIA is an assessment of the impacts of a planned activity on the environment. It is a significant, anticipa- tory, environmental management tool. International debate fo- cuses on its enhancement to meet the challenges of sustainable development and demands for scientifically robust integrated and participative decision-making (Morris & Therivel, 2003; Petts, 1999; Wathern, 2001). There is a vast MCDM literature which deals with EIA problems: Marttunen and Hamalainen (1995) presented the decision analysis interview-method and applied the methodology to the assessment of the environmental impacts of two water development projects. Geldermann, Spengler, & Rentz, 2000 proposed using PROMETHEE for environmental assessment and provided a case study in iron and steel making industry. Spengler, Geldermann, Hahre, Siever- dingbeck, and Rentz (1998) developed a multiple criteria based decision support system for environmental assessment of recycling measures in the iron and steel making industry. Al-Rashdan et al. (1999) used PROMETHEE for EIA and ranking the environmental projects in Jordan. Janssen (2001) provided an overview on the use of multicriteria analysis in EIA in the Netherlands. Ramanathan (2001) used AHP for capturing the perceptions of stakeholders on the relative severity of different socio-economic impacts, which will help the authorities in prioritizing their environmental man- agement plan and allocating the budget available for mitigating adverse socio-economic impacts. Seppala, Basson, and Norris (2002) provided an overview of the commonly used MCDM methods and discuss life cycle impact assessment (LCIA) in relation to them. Bojorquez-Tapia, Sanchez- Colon, and Martınez (2005) combined two MCDM techniques, sto- chastic AHP and compromise programming, to ascertain the envi- ronmental impacts of and to rank two alternative sites for Mexico City’s new airport. Payraudeau and van der Werf (2005) provided an analysis of six main types of methodologies (environmental risk mapping, life cycle analysis, environmental impact assessment, multi-agent system, linear programming and agro-environmental indicators) to illustrate the variety of methods used in EIA. Wang, Table 1 EIA evaluation criteria proposed by Liu and Lai (2009). EIA aspects EIA criteria Environmental pollution Air pollution Water pollution Soil pollution Noise pollution Solid waste pollution Ecological alteration Terrestrial Aquatic Socioeconomic disturbance Economics Society Culture Fig. 1. Membership function of TrFN ~m. Table 2 Fuzzy evaluation scale for the weights. Linguistic terms Fuzzy score Absolutely strong (AS) (5/2, 3, 7/2, 4) Very strong (VS) (2, 5/2, 3, 7/2) Fairly strong (FS) (3/2, 2, 5/2, 3) Slightly strong (SS) (1, 3/2, 2, 5/2) Equal (E) (1, 1, 1, 1) Slightly weak (SW) (2/5, 1/2, 2/3, 1) Fairly weak (FW) (1/3, 2/5, 1/2, 2/3) Very weak (VW) (2/7, 1/3, 2/5, 1/2) Absolutely weak (AW) (1/4, 2/7, 1/3, 2/5) 8554 T. Kaya, C. Kahraman / Expert Systems with Applications 38 (2011) 8553–8562 https://isiarticles.com/article/6239 
work_enf4y3q56fcfvebnjcr3eyu4da	----	Rudarsko-geološko-naftni zbornik Vol. 25 str. 107-113 Zagreb, 2012. UDK 504.064:622 Pregledni rad UDC 504.064:622 Review Language/Jezik:Hrvatski/Croatian EKSPERTNI SUSTAVI ZA ZAŠTITU OKOLIŠA S PRIMJENOM U RUDARSTVU EXPERT SYSTEMS FOR ENVIRONMENTAL PROTECTION IN MINING DAMIR RUMENJAK Ministarstvo zaštite okoliša i prirode, Zagreb,Ulica Republike Austrije 14 Ključne riječi: ekspertni sustavi, okolišni indikatori, mjere preferencije, ljestvice, neizrazita logika, lingvističke varijable, agregacija Key words: expert systems, environmental indicators, measures of pref- erence, scales, fuzzy logic, linguistic variables, aggregation Sažetak Ekspertni sustavi primjenjuju se u različitim djelatnostima, s ciljem donošenja odluka na inteligentan i transparentan način. U poslovnim djelatnostima ekspertni sustavi usmjereni su k donošenju operativnih i ekonomskih odluka. U zaštiti okoliša, uvažavajući posebnosti različitih poslovnih djelatnosti, u izboru odluka koriste se i ostali aspekti od- lučivanja, koji osim ekonomsko-operativnih razloga, vrednuju i druge razloge izbora odluka. Neizrazita logika je logički i matematički aparat, koji se preporuča za primjenu u takvim ekspertnim sustavima. U radu se prikazuju neke značajke i pravila konstrukcije ekspertnih sustava za pri- mjenu u zaštiti okoliša u rudarstvu, a koji primjenjuju neizrazitu logiku. Abstract Expert systems are used in various activities with the main goal of decision-making in intelligent and transparent way. In business de- cision-making, such systems are directed to pure operational and eco- nomical decisions. In environmental protection they have to accept the specificity of business sector for which they have been intended, but also use environmental reasons and evaluate them differently than eco- nomic. Fuzzy logic is logical and mathematical mean recommended for such expert systems. Some characteristics and rules for construction of systems based on fuzzy logic are provided in the work. Uvod: Pojam ekspertnih sustava Ekspertni sustavi su sustavi povezivanja znanja i pra- vila primjene tog znanja u različitim područjima, kao i ostalih dijelova sustava bitnih za širu primjenu tog znanja. Osnovna primjena ekspertnih sustava je primjena u odlu- čivanju, odnosno izboru ili opravdanju pojedinih varijanti tehnoloških postupaka (ili dijelova postupaka) kada se radi o gospodarenju resursima, npr. u eksploataciji mi- neralnih sirovina. U širem smislu, ekspertni sustavi pripadaju umjetnoj inteligenciji (Jackson, 1998). Na slici 1. je opisan jednostavni ekspertni sustav s dijelovima i interakcijama bitnih dijelova i korisnika. Najvažniji dijelovi takvog sustava su baza znanja i pravila po kojima se to znanje primjenjuje, i koje čine ekspertnu školjku sustava. Karakteristika baze znanja ekspertnih sustava su stal- na promjena u skladu s neprekidnim razvojem i promje- nom znanja u određenim područjima. Pravila se rjeđe mi- jenjaju. Na slici 1. označena je interakcija između ostalih dijelova sustava, prije svega ekspertnih timova koji dopri- nose ekspertnom znanju (koje vode koordinatori koji moraju sagledavati cijeli sustav) te pravila i znanja, ali i korisnika koji utječu na konač- no prihvaćanje (formalizaciju) znanja, pogotovo kada se ekspertni sustavi koriste u administrativnim postupcima, kao što je procjena utjecaja na okoliš. Korisnici, preko povratnih informacija, najčešće djeluju na cjelinu (školj- ku), dok je modifikacija unutar školjke najčešće stvar ek- spertnih timova i koordinatora. Naravno, ovakva podjela uloga nije ograničenje u razvoju ekspertnih sustava, kao niti moguće interakcije koje se između dijelova sustava prema slici 1. još mogu uspostaviti. Rud.-geol.-naft. zb., Vol. 25, 2012. D. Rumenjak: Ekspertni sustavi za zaštitu okoliša...108 Detaljnija podjela pravila u ekspertnim sustavima provodi se na pravila u sustavu i meta-pravila, kojima se omogućuje primjena u situacijama kada to nije moguće riješiti pravilima samog sustava. Okolišni indikatori i mjere preferencije Ekspertno znanje u ekspertnim sustavima izražava se veličinama (varijablama) određene vrste. U zaštiti oko- liša varijable koje se primjenjuju su okolišni indikatori. Teorija okolišnih indikatora je područje koje se razvija, dajući odgovarajuće podijele okolišnih varijabli koje se primjenjuju u izgradnji ekspertnih sustava. Indikatori se mogu povezivati u tzv. zajedničke (skupne) indikatore. U Patiero-Fernandez et al. (2005) za povezivanje indikato- ra onečišćenja voda u rudarstvu predlažu se izrazi: (1) (2) gdje su: Pvoda, : zbirni indikator onečišćenja vode, gwvoda,pH: indikator kiselosti, gwvoda,i: indikatori oneči- šćenja voda, IA/proiz.zat var a.uk.: zbirni indikator korištenja zemljišta, Isub.: indikator u ovisnosti od eksploatacije (podzemno/površinsko rudarstvo), IA,j : vrsta korištenog zemljišta, IA,Mineral ;površina eksploatacije i rekultivacije, IA,T : površina jalovišta, IA,D: površina otkrivke, N: skupina indikatora korištenja prirode, S: skupina indikatora kori- štenja zemljišta za naselja, W: skupina indikatora korište- nja zemljišta za ostale ekonomske aktivnosti, M: skupina ostalih indikatora. X.Zhu et al. (2007) daje sljedeći izraz za povezivanje okolišnih indikatora za sanaciju i rekultivaciju lokacija obuhvaćenih rudarskim aktivnostima: , (3) gdje su: : vrijednost indikatora sanacije i rekultivacije, : indikatori sanacije i rekultivacije. Rud.-geol.-naft. zb., Vol. 25, 2012. 2 Rumenjak, D: Ekspertni sustavi za zaštitu okoliša… Okolišni indikatori i mjere preferencije Ekspertno znanje u ekspertnim sustavima izražava se veličinama (varijablama) određene vrste. U zaštiti okoliša varijable koje se primjenjuju su okolišni indikatori. Teorija okolišnih indikatora je područje koje se razvija, dajući odgovarajuće podijele okolišnih varijabli koje se primjenjuju u izgradnji ekspertnih sustava. Indikatori se mogu povezivati u tzv. zajedničke (skupne) indikatore. U Patiero-Fernandez et al. (2005) za povezivanje indikatora onečišćenja voda u rudarstvu predlažu se izrazi: )( ,, ,,, ,,...var./ TADA MWSNj MineralAjAsubukazatproizA IIIIII    (1)    pHgwmgwP pHvodai i ivodavoda    66, 15 1 , (2) gdje su: vodaP , : zbirni indikator onečišćenja vode, pHvodagw , : indikator kiselosti, ivodagw , : indikatori onečišćenja voda, ..var./ ukazatproizAI : zbirni indikator korištenja zemljišta, .subI : indikator u ovisnosti od eksploatacije (podzemno/površinsko rudarstvo), jAI , : vrsta korištenog zemljišta, MineralAI , ;površina eksploatacije i rekultivacije, TAI , : površina jalovišta, DAI , : površina otkrivke, N: skupina indikatora korištenja prirode, S: skupina indikatora korištenja zemljišta za naselja, W: skupina indikatora korištenja zemljišta za ostale ekonomske aktivnosti, M: skupina ostalih indikatora. X.Zhu et al. (2007) daje sljedeći izraz za povezivanje okolišnih indikatora za sanaciju i rekultivaciju lokacija obuhvaćenih rudarskim aktivnostima: , (3) gdje su: : vrijednost indikatora sanacije i rekultivacije, : indikatori sanacije i rekultivacije. Okolišni indikatori su ujedno i preferencije za odlučivanje. Povezivanjem indikatora izrazima (1), (2) i (3), određuje im se već odgovarajuća mjera preferencije koja je potrebna za odlučivanje u ekspertnim sustavima. Kao zajednička mjera preferencije (MP) najčešće se definira korisnost (eng. utility), koja se prikazuje kao funkcija. Za korisnost kao funkciju postavlja se aksiom (Russel & Norvig, 2010): Ako su A i B preferencije za dvije varijable i ako je , onda vrijedi vrijedi: , gdje su i mjere preferencije za varijable A i B. Na temelju tog aksioma razvijaju se ostale relacije za matematičku obradu mjera preferencije koje nisu uvijek (ili nisu u pravilu) linearne funkcije. Uz korisnost, u novijim pristupima u odlučivanju u zaštiti okoliša sve se više primjenjuju i druge mjere preferencije (dobrobit, blagostanje), koje se i odgovarajuće matematički definiraju (Dasgupta, 2001). Mjere preferencije (korisnost, kao i ostale) u ekspertnim sustavima mogu se prikazati različitim ljestvicama, o kojima je moguće govoriti kao o dva osnovna tipa ljestvica; ordinalnim (nominalnim) i kardinalnim ljestvicama (tablica 1.). Kardinalne ljestvice koriste konvencionalan matematički aparat i statističke procedure, dok je operacije kod ordinalnih potrebno odgovarajuće (konsenzualno) definirati. Izuzetak je neizrazita logika, čije je aksiomatski aparat dobro razvijen konvencionalnom matematikom. Pravila u sustavu Znanje Razvoj eksp. sustava (uži ekspertni timovi, koordinatori) Korisnici (administrativni postupci, sudjelovanje javnosti, projektiranje) Ekspertna školjka Slika 1. Ekspertni sustav s dijelovima i interakcijama M eta-pravila Figure 1. Expert system with parts and interactions Rud.-geol.-naft. zb., Vol. 25, 2012. 2 Rumenjak, D: Ekspertni sustavi za zaštitu okoliša… Okolišni indikatori i mjere preferencije Ekspertno znanje u ekspertnim sustavima izražava se veličinama (varijablama) određene vrste. U zaštiti okoliša varijable koje se primjenjuju su okolišni indikatori. Teorija okolišnih indikatora je područje koje se razvija, dajući odgovarajuće podijele okolišnih varijabli koje se primjenjuju u izgradnji ekspertnih sustava. Indikatori se mogu povezivati u tzv. zajedničke (skupne) indikatore. U Patiero-Fernandez et al. (2005) za povezivanje indikatora onečišćenja voda u rudarstvu predlažu se izrazi: )( ,, ,,, ,,...var./ TADA MWSNj MineralAjAsubukazatproizA IIIIII    (1)    pHgwmgwP pHvodai i ivodavoda    66, 15 1 , (2) gdje su: vodaP , : zbirni indikator onečišćenja vode, pHvodagw , : indikator kiselosti, ivodagw , : indikatori onečišćenja voda, ..var./ ukazatproizAI : zbirni indikator korištenja zemljišta, .subI : indikator u ovisnosti od eksploatacije (podzemno/površinsko rudarstvo), jAI , : vrsta korištenog zemljišta, MineralAI , ;površina eksploatacije i rekultivacije, TAI , : površina jalovišta, DAI , : površina otkrivke, N: skupina indikatora korištenja prirode, S: skupina indikatora korištenja zemljišta za naselja, W: skupina indikatora korištenja zemljišta za ostale ekonomske aktivnosti, M: skupina ostalih indikatora. X.Zhu et al. (2007) daje sljedeći izraz za povezivanje okolišnih indikatora za sanaciju i rekultivaciju lokacija obuhvaćenih rudarskim aktivnostima: , (3) gdje su: : vrijednost indikatora sanacije i rekultivacije, : indikatori sanacije i rekultivacije. Okolišni indikatori su ujedno i preferencije za odlučivanje. Povezivanjem indikatora izrazima (1), (2) i (3), određuje im se već odgovarajuća mjera preferencije koja je potrebna za odlučivanje u ekspertnim sustavima. Kao zajednička mjera preferencije (MP) najčešće se definira korisnost (eng. utility), koja se prikazuje kao funkcija. Za korisnost kao funkciju postavlja se aksiom (Russel & Norvig, 2010): Ako su A i B preferencije za dvije varijable i ako je , onda vrijedi vrijedi: , gdje su i mjere preferencije za varijable A i B. Na temelju tog aksioma razvijaju se ostale relacije za matematičku obradu mjera preferencije koje nisu uvijek (ili nisu u pravilu) linearne funkcije. Uz korisnost, u novijim pristupima u odlučivanju u zaštiti okoliša sve se više primjenjuju i druge mjere preferencije (dobrobit, blagostanje), koje se i odgovarajuće matematički definiraju (Dasgupta, 2001). Mjere preferencije (korisnost, kao i ostale) u ekspertnim sustavima mogu se prikazati različitim ljestvicama, o kojima je moguće govoriti kao o dva osnovna tipa ljestvica; ordinalnim (nominalnim) i kardinalnim ljestvicama (tablica 1.). Kardinalne ljestvice koriste konvencionalan matematički aparat i statističke procedure, dok je operacije kod ordinalnih potrebno odgovarajuće (konsenzualno) definirati. Izuzetak je neizrazita logika, čije je aksiomatski aparat dobro razvijen konvencionalnom matematikom. Pravila u sustavu Znanje Razvoj eksp. sustava (uži ekspertni timovi, koordinatori) Korisnici (administrativni postupci, sudjelovanje javnosti, projektiranje) Ekspertna školjka Slika 1. Ekspertni sustav s dijelovima i interakcijama M eta-pravila Figure 1. Expert system with parts and interactions Slika 1. Ekspertni sustav s dijelovima i interakcijama Figure 1. Expert system with parts and interactions Okolišni indikatori su ujedno i preferencije za odlu- čivanje. Povezivanjem indikatora izrazima (1), (2) i (3), određuje im se već odgovarajuća mjera preferencije koja je potrebna za odlučivanje u ekspertnim sustavima. Kao zajednička mjera preferencije (MP) najčešće se definira korisnost (eng. utility), koja se prikazuje kao funkcija. Za korisnost kao funkciju postavlja se aksiom (Russel & Norvig, 2010): Ako su A i B preferencije za dvije varijable i ako je , onda vrijedi vrijedi: , gdje su i mjere preferencije za varijable A i B. Na temelju tog ak- sioma razvijaju se ostale relacije za matematičku obradu mjera preferencije koje nisu uvijek (ili nisu u pravilu) li- nearne funkcije. Uz korisnost, u novijim pristupima u od- lučivanju u zaštiti okoliša sve se više primjenjuju i druge mjere preferencije (dobrobit, blagostanje), koje se i odgo- varajuće matematički definiraju (Dasgupta, 2001). Mjere preferencije (korisnost, kao i ostale) u eksper- tnim sustavima mogu se prikazati različitim ljestvicama, o kojima je moguće govoriti kao o dva osnovna tipa ljestvi- ca; ordinalnim (nominalnim) i kardinalnim ljestvicama (tablica 1.). Kardinalne ljestvice koriste konvencionalan matematički aparat i statističke procedure, dok je operaci- je kod ordinalnih potrebno odgovarajuće (konsenzualno) definirati. Izuzetak je neizrazita logika, čije je aksiomatski aparat dobro razvijen konvencionalnom matematikom. 109 Rud.-geol.-naft. zb., Vol. 25, 2012. D. Rumenjak: Ekspertni sustavi za zaštitu okoliša... Tablica 1. Ljestvice mjera preferencije koje se koriste u ekspertnim sustavima Table 1. The scales of measures of preferences used in expert systems LJESTVICA OSOBINE LJESTVICE DOPUSTIVE MATEMATIČKE I LOGIČKE OPERACIJE DOPUSTIVA STATISTIČKA PROCEDURA VREDNOVANJA REZULTATA Nominalna Podjela prema klasifikaciji Jednostavna supstitucija ili konsenzualno ekspertno definiranje operacija (simbolička logika i aritmetika) Informacijska statistika Ordinalna Podjela u terminima važnosti Relacije ekvivalencije s drugim monotonim rastućim ili padajućim funkcijama ili transformacije članova za kvantifikaciju u kardinalnu ljestvicu Neparametarska statistika Intervalna Podjela u terminima jednakih razlika Linearne transformacije članova ljestvice Parametarska statistika Racionalna (omjerna) Podjela u terminima jednakih omjera Množenje i djeljenje s konstantama ili s drugim vrijednostima iz ljestvice Parametarska statistika Lingvistička na temelju neizrazitih skupova Podjela prema pripadajućim područjima i funkcijama udjela Neizrazita aritmetika i neizrazita logika Neizrazita statistika U tablici 1. su također date statističke procedure vred- novanja rezultata, a primjer statističke procedure izbora lokacije odlagališta otpada ordinalnom ljestvicom prika- zan je u Rumenjak (2006). Za transformaciju ordinalne u kardinalnu ljestvicu mogu se koristiti nelinearne transformacije, kao npr. T = Mi M i ,....1, 21 = − , (4) gdje je M broj članova ljestvice. Osim ljestvica mjera preferencije, u ekspertnim susta- vima moguća je i primjena ekonomskih mjera preferenci- je (novčana vrijednost okoliša. Takve mjere preferencije mogu biti od važnosti za povezivanje s poslovnim odlu- čivanjem. Izbor odluke u ekspertnom sustavu prikazan je u tablici 2. Razvoj ekspertnih sustava za zaštitu okoliš u rudar- stvu u Hrvatskoj Od 2000. nadalje, donošenjem zakonskih propisa koji- ma se je bolje regulirao način provedbe procjene utjecaja na okoliš, u Hrvatskoj se češće koriste ekspertne metode prihvatljivosti za okoliš. Daljnji razvoj propisa iz zaštite okoliša i usuglašavanje sa zakonodavstvom EU-a, proši- rili su potrebu za takvim pristupom. Te metode, od kojih se za potrebe procjene utjecaja na okoliš pojavio relativno veliki broj, mogu se nazvati i ekspertnim sustavima, iako često nemaju sve osobine takvih sustava. Za primjenu u rudarstvu prepoznato je nekoliko metoda (Rumenjak, 2007). Sve sadrže različite mjere preferencije (ordinalna ili razmjerna ljestvica, tablica 1.) te definiraju način izbo- ra varijanti (agregacije rezultata). Tablica 2. Izbor odluke kroz vrednovanje mjera preferencije okolišnih varijabli, , s kriterijima Cj i varijantama ai Table 2. Choice of decision through evaluation of measures of preference for environmental indicator , with criteria Cj and alternatives ai varijante Kriteriji C1 ..... ..... ..... ..... .... ....... ..... ...... ..... ...... ..... ..... .... ...... ..... ...... ..... ....... ..... ..... Rud.-geol.-naft. zb., Vol. 25, 2012. D. Rumenjak: Ekspertni sustavi za zaštitu okoliša...110 Metoda ranjivosti s matricom interakcija (Marušič, 1999) u primjeni je u susjednoj Sloveniji, iz koje je potekla te je pronašla primjenu u određenom broju postupaka pro- cjena utjecaja na okoliš za eksploataciju mineralnih sirovi- na u Hrvatskoj. Definirana je ordinalnom ljestvicom od 1 do 5. Broj indikatora koji se primjenjuje varira, ali najčešće se kreće oko 20. Agregacija rezultata se može provoditi preko glavnih dijelova tehnološkog procesa i /ili kriterija. Metoda oblikovanja i prenamjene predložena je 2005. (Krasić, 2005), posebno za rudarstvo. Definira 40 indi- katora, s ordinalnim ili racionalnim ljestvicama od 1 do 3, podijeljenih u tri osnovne skupine: prirodni, ekonom- sko-operativni i okolišni. Do izbora varijante eksploata- cije dolazi se usporedbom, kao i u metodi ranjivosti, s prethodno definiranom ljestvicom rezultata. Također su u primjeni neke nespecijalizirane metode (sustavi) multikriterijalne analize s ordinalnim/omjernim ljestvicama (često u rasponu od 1 do 10), koje se kreiraju u zavisnosti od slučajeva u primjeni. Glavni problemi kod takvih metoda su loše definirane ljestvice i nedostatak po- trebnog konsenzusa oko definiranja ljestvica i indikatora, što rezultira u niskoj vjerodostojnosti takvih metoda. Izbor oko 30 indikatora temeljem teorije indikatora održivog razvitka, koji su prikladni za primjenu u zaštiti okoliša u rudarstvu daje se u Möllerherm et al. (2005). Ekspertni sustav brze matrične metode (RIAM) (Pa- stakia,1998) nije među onim sustavima koji su posebno namijenjeni za korištenje u zaštiti okoliša u rudarstvu, ali je razmatran i predložen u varijanti neizrazite logike (Ru- menjak, 2004). To je veoma osjetljiva višeljestvična me- toda, koja istovremeno koristi 5 različitih ljestvica mjera preferencije te široki spektar indikatora održivog razvitka. Zbog svoje osjetljivosti prikladna je za svaki kriterij koji dolazi u obzir u zaštiti okoliša. Agregacija se kod ove me- tode provodi na drugačiji način od prethodno navedenih metoda, usporedbom frekvencija područja rezultata za va- rijante odlučivanja. Ekspertni sustavi s neizrazitom logikom Neizrazita logika (eng. fuzzy logic) je logička i ma- tematička disciplina koja se u svojim operacijama služi neizrazitim skupovima. Neizrazite skupove karakterizira funkcija udjela ( , tj. pripadnosti određenom skupu, koja je . Na temelju neizrazitih skupova definiraju se lingvi- stičke varijable, koje povezuju lingvistička svojstava (kroz koje se izražava ekspertno znanje) s funkcijama udjela neizrazitih skupova. Lingvistička varijabla je, kao skup, i neizraziti broj, na koji je moguće primijeniti od- govarajuće definirane matematičke operacije. Svojstva ili preferencije koja lingvističke varijable (LV), kao mjere preferencije, opisuju u ekspertnim susta- vima, mogu biti različita. Osim prihvatljivosti za okoliš, moguće je definirati i povoljnost i osjetljivost (svojstva posebno primjenljiva za prirodu). U podsustavima eksper- tnih sustava moguće je koristiti i druge posebne varijable, npr. intenzitet različitih utjecaja, fleksibilnost tehnoloških rješenja itd., koje su najčešće dio nekog podsustava u ek- spertnom sustavu (Rumenjak, 2009). Razvijena lingvistička varijabla, za indikator okoliša za koji se određuje prihvatljivost za okoliš (Rumenjak, 2007), prikazana je na slici 2. Na lingvističkim varijablama, kao neizrazitim skupovima, moguće je defi nirati odgovarajuće pod-, moguće je defi nirati odgovarajuće pod-moguće je defi nirati odgovarajuće pod-će je defi nirati odgovarajuće pod-e je defi nirati odgovarajuće pod- je defi nirati odgovarajuće pod-je defi nirati odgovarajuće pod- defi nirati odgovarajuće pod-definirati odgovarajuće pod- odgovarajuće pod-odgovarajuće pod-će pod-e pod- pod-pod- skupove (particije), koje se može iskoristiti u daljnoj raz- (particije), koje se može iskoristiti u daljnoj raz-particije), koje se može iskoristiti u daljnoj raz-), koje se može iskoristiti u daljnoj raz-koje se može iskoristiti u daljnoj raz- se može iskoristiti u daljnoj raz-se može iskoristiti u daljnoj raz- može iskoristiti u daljnoj raz-može iskoristiti u daljnoj raz-že iskoristiti u daljnoj raz-e iskoristiti u daljnoj raz- iskoristiti u daljnoj raz-iskoristiti u daljnoj raz- u daljnoj raz-u daljnoj raz- daljnoj raz-daljnoj raz- raz-raz- radi ekspertnog sustava. Particijama LV proširuje se dalje lingvistička ljestvica, koja je inače zadane s 2 temeljne particije: prihvatljivost i neprihvatljivost za okoliš. Izražavanje neizvjesnosti u neizrazitoj logici U neizrazitoj logici neizvjesnost se izražava koncep- tom neizrazitosti, odnosno funkcijama udjela neizrazitih skupova. Rud.-geol.-naft. zb., Vol. 25, 2012. Rumenjak, D: Ekspertni sustavi za zaštitu okoliša… 5 Razvijena lingvistička varijabla, za indikator okoliša za koji se određuje prihvatljivost za okoliš (Rumenjak, 2007), prikazana je na slici 2. Na lingvističkim varijablama, kao neizrazitim skupovima, moguće je definirati odgovarajuće podskupove (particije), koje se može iskoristiti u daljnoj razradi ekspertnog sustava. Particijama LV proširuje se dalje lingvistička ljestvica, koja je inače zadane s 2 temeljne particije: prihvatljivost i neprihvatljivost za okoliš. Izražavanje neizvjesnosti u neizrazitoj logici U neizrazitoj logici neizvjesnost se izražava konceptom neizrazitosti, odnosno funkcijama udjela neizrazitih skupova. Način izražavanja neizvjesnosti preko neizrazitih skupova prikazan je na sl. 3., za lingvističku varijablu povoljnosti za okoliš (Rumenjak, 2007), s prikazom različitih vrijednosti funkcija udjela μ za istu vrijednost indikatora i: Mjera preferencije (MP) u lingvističkoj varijabla prikazana na sl. 2 i 3 prikazuje se funkcijom udjela i položajem na ljestvici mjere preferencije (tzv. položajna ljestvica), najčešće u zatvorenom intervalu . Nelinearnost Nelinearnost je značajka pojava i procesa u okolišu (Pravdić, 2008.), koji su često nelinearni, u dinamičkom i statičkom značenju. Važno je napomenuti da nelinearnost ima podjednaku važnost i za fazu odlučivanja kod izbora varijanti, kao i za monitoring utjecaja kod eksploatacije, a posebno u fazi sanacije rudarskih radova, (Rumenjak et al, 2008). Nelinearnost se u neizrazitoj logici prikazuje preko gustoće preklapanja particija lingvističkih varijabli (Ross, 1995), slika.4. Primjer nelinearnosti u ekspertnim sustavima u rudarstvu je važno pitanje prihvatljivosti trajanja eksploatacije na površinskim kopovima, koje se često postavlja u administrativnim postupcima kao što je procjena utjecaja na okoliš. U nekim izvorima ističe se da je duža eksploatacija neprihvatljiva za okoliš (Craig et al. 2011), a to je često i stav javnosti u procjenama utjecaja na okoliš i administracije, koja takvu percepciju dužine trajanja eksploatacije koristi kao ograničavajući faktor za odobravanje eksploatacije. Međutim, teorijska razmatranja pitanja „blagostanja“, kao mjere preferencije za okoliš, navode da je duže trajanje bilo koje aktivnosti s obzirom na korištenje resursa u okolišu, prihvatljivije za okoliš (Dasgupta, 2001). 1 0 Slika 3. Način prikazivanja neizvjesnosti za indikator okoliša (i) u lingvističkoj varijabli povoljnosti za okoliš s tri particije (podjele) neizrazitog skupa (1), (2) i (3) i μ (1) (2) (3) Figure 3. Presenting of uncertainity for environmental indicator (i)in linguistic variable favourable for environment with three partitions (1), (2) i (3) Mjera indikatora utjecaja, i Slika 2. Razvijena lingvistička varijabla prihvatljivost za okoliš s dvije particije - prihvatljivost i neprihvatljivost za okoliš Neprihvatljivost za okoliš μ 1 0 Figure 2. Developed linguistic variables acceptability for environment with two partitions- acceptability and non- acceptability for environment Rud.-geol.-naft. zb., Vol. 25, 2012. Rumenjak, D: Ekspertni sustavi za zaštitu okoliša… 5 Razvijena lingvistička varijabla, za indikator okoliša za koji se određuje prihvatljivost za okoliš (Rumenjak, 2007), prikazana je na slici 2. Na lingvističkim varijablama, kao neizrazitim skupovima, moguće je definirati odgovarajuće podskupove (particije), koje se može iskoristiti u daljnoj razradi ekspertnog sustava. Particijama LV proširuje se dalje lingvistička ljestvica, koja je inače zadane s 2 temeljne particije: prihvatljivost i neprihvatljivost za okoliš. Izražavanje neizvjesnosti u neizrazitoj logici U neizrazitoj logici neizvjesnost se izražava konceptom neizrazitosti, odnosno funkcijama udjela neizrazitih skupova. Način izražavanja neizvjesnosti preko neizrazitih skupova prikazan je na sl. 3., za lingvističku varijablu povoljnosti za okoliš (Rumenjak, 2007), s prikazom različitih vrijednosti funkcija udjela μ za istu vrijednost indikatora i: Mjera preferencije (MP) u lingvističkoj varijabla prikazana na sl. 2 i 3 prikazuje se funkcijom udjela i položajem na ljestvici mjere preferencije (tzv. položajna ljestvica), najčešće u zatvorenom intervalu . Nelinearnost Nelinearnost je značajka pojava i procesa u okolišu (Pravdić, 2008.), koji su često nelinearni, u dinamičkom i statičkom značenju. Važno je napomenuti da nelinearnost ima podjednaku važnost i za fazu odlučivanja kod izbora varijanti, kao i za monitoring utjecaja kod eksploatacije, a posebno u fazi sanacije rudarskih radova, (Rumenjak et al, 2008). Nelinearnost se u neizrazitoj logici prikazuje preko gustoće preklapanja particija lingvističkih varijabli (Ross, 1995), slika.4. Primjer nelinearnosti u ekspertnim sustavima u rudarstvu je važno pitanje prihvatljivosti trajanja eksploatacije na površinskim kopovima, koje se često postavlja u administrativnim postupcima kao što je procjena utjecaja na okoliš. U nekim izvorima ističe se da je duža eksploatacija neprihvatljiva za okoliš (Craig et al. 2011), a to je često i stav javnosti u procjenama utjecaja na okoliš i administracije, koja takvu percepciju dužine trajanja eksploatacije koristi kao ograničavajući faktor za odobravanje eksploatacije. Međutim, teorijska razmatranja pitanja „blagostanja“, kao mjere preferencije za okoliš, navode da je duže trajanje bilo koje aktivnosti s obzirom na korištenje resursa u okolišu, prihvatljivije za okoliš (Dasgupta, 2001). 1 0 Slika 3. Način prikazivanja neizvjesnosti za indikator okoliša (i) u lingvističkoj varijabli povoljnosti za okoliš s tri particije (podjele) neizrazitog skupa (1), (2) i (3) i μ (1) (2) (3) Figure 3. Presenting of uncertainity for environmental indicator (i)in linguistic variable favourable for environment with three partitions (1), (2) i (3) Mjera indikatora utjecaja, i Slika 2. Razvijena lingvistička varijabla prihvatljivost za okoliš s dvije particije - prihvatljivost i neprihvatljivost za okoliš Neprihvatljivost za okoliš μ 1 0 Figure 2. Developed linguistic variables acceptability for environment with two partitions- acceptability and non- acceptability for environment Slika 2. Razvijena lingvistička varijabla prihvatljivost za okoliš s dvije particije - prihvatljivost i neprihvatljivost za okoliš Figure 2. Developed linguistic variables acceptability for envi- ronment with two partitions- acceptability and non- acceptability for environment Slika 3. Način prikazivanja neizvjesnosti za indikator okoliša (i) u lin- gvističkoj varijabli povoljnosti za okoliš s tri particije (podjele) neizra- zitog skupa (1), (2) i (3) Figure 3. Presenting of uncertainity for environmental indicator (i)in linguistic variable favourable for environment with three partitions (1), (2) i (3) Način izražavanja neizvjesnosti preko neizrazitih sku- pova prikazan je na sl. 3., za lingvističku varijablu povolj- nosti za okoliš (Rumenjak, 2007), s prikazom različitih vrijednosti funkcija udjela μ za istu vrijednost indikatora i: Mjera preferencije (MP) u lingvističkoj varijabla pri- kazana na sl. 2 i 3 prikazuje se funkcijom udjela i položa- 111 Rud.-geol.-naft. zb., Vol. 25, 2012. D. Rumenjak: Ekspertni sustavi za zaštitu okoliša... jem na ljestvici mjere preferencije (tzv. položajna ljestvi- ca), najčešće u zatvorenom intervalu . Nelinearnost Nelinearnost je značajka pojava i procesa u okolišu (Pravdić, 2008.), koji su često nelinearni, u dinamičkom i statičkom značenju. Važno je napomenuti da nelinearnost ima podjednaku važnost i za fazu odlučivanja kod izbora varijanti, kao i za monitoring utjecaja kod eksploatacije, a posebno u fazi sanacije rudarskih radova, (Rumenjak et al, 2008). Nelinearnost se u neizrazitoj logici prikazuje preko gustoće preklapanja particija lingvističkih varijabli (Ross, 1995), slika.4. Primjer nelinearnosti u ekspertnim sustavima u ru- darstvu je važno pitanje prihvatljivosti trajanja eksploa- tacije na površinskim kopovima, koje se često postavlja u administrativnim postupcima kao što je procjena utjecaja na okoliš. U nekim izvorima ističe se da je duža eksplo- atacija neprihvatljiva za okoliš (Craig et al. 2011), a to je često i stav javnosti u procjenama utjecaja na okoliš i administracije, koja takvu percepciju dužine trajanja ek- sploatacije koristi kao ograničavajući faktor za odobrava- nje eksploatacije. Međutim, teorijska razmatranja pitanja „blagostanja“, kao mjere preferencije za okoliš, navode da je duže trajanje bilo koje aktivnosti s obzirom na ko- rištenje resursa u okolišu, prihvatljivije za okoliš (Das-Das- gupta, 2001). Rud.-geol.-naft. zb., Vol. 25, 2012. 6 Rumenjak, D: Ekspertni sustavi za zaštitu okoliša… Opravdano je stoga pretpostaviti da funkcija prihvatljivosti trajanja eksploatacije, u zavisnosti od vremena eksploatacije, ima oblik prikazan na slici. 4. Aproksimacija nelinearnosti ove funkcije lingvističkim varijablama može se provesti lingvističkom varijablom s tri ili više particija, koje se odgovarajuće preklapaju u zavisnosti od karakteristika nelinearne funkcije. Particije mogu biti različito definirano trajanje eksploatacije, kao što je to prikazano u tablici 3. Primjer ekspertnog sustava s lingvističkim varijablama za primjenu u rudarstvu Dijelovi ekspertnog sustava, koji se temelje na konstrukciji lingvističkih varijabli (LV) prema slici.2 prikazani su na primjeru “metoda prenamjene i preoblikovanja”, tablica 3. Sustav koji je ovdje prikazan, dobiven je transformacijom ordinalne ljestvice mjera preferencije polazne metode u kardinalnu ljestvicu, primjenom nelinearne transformacije preko izraza (4) i sadrži okolišne indikatore prenamjene prostora površinskog kopa. Kao okolišni indikator, u sustav je uključeno vrijeme trajanja eksploatacije. Značajka primjera prikazanog u tablici 3. je ta što je on dio šireg ekspertnog podsustava, što je u ovom slučaju vidljivo iz broja članova kardinalne ljestvice (M =20), koji predstavljaju broj Domene particije prihvatljivosti (I, II, III) Slika 4. Aproksimacija nelinearnosti na funkciji prihvatljivosti trajanja eksploatacije za okoliš s tri particije LV preklapanjem neizrazitih podskupova – particija Prihvatljivost za sastavnicu okoliša I II τ (vrijeme trajanja eksploatacije) III Figure 4. Approximation of non-linearity applied on function of acceptability for environment with three partitions overlapped on the axis of the function 1 0 Slika 5. Podjela lingvističke varijable prihvatljivosti trajanja eksploatacije s obzirom na nelinearnost funkcije prihvatljivosti na četiri particije, s položajnom ljestvicom mjera preferencije MP i 0 1 MP = μ Figure 5. Division of linguistic variable - duration of exploatation, considering non-linearity of function with four partitions and positional scale MP Opravdano je stoga pretpostaviti da funkcija prihvat- ljivosti trajanja eksploatacije, u zavisnosti od vremena eksploatacije, ima oblik prikazan na slici. 4. Aproksima- cija nelinearnosti ove funkcije lingvističkim varijablama može se provesti lingvističkom varijablom s tri ili više particija, koje se odgovarajuće preklapaju u zavisnosti od karakteristika nelinearne funkcije. Particije mogu biti ra- zličito definirano trajanje eksploatacije, kao što je to pri- kazano u tablici 3. Primjer ekspertnog sustava s lingvističkim varijabla- ma za primjenu u rudarstvu Dijelovi ekspertnog sustava, koji se temelje na kon- strukciji lingvističkih varijabli (LV) prema slici.2 prika- zani su na primjeru “metoda prenamjene i preoblikova- nja”, tablica 3. Sustav koji je ovdje prikazan, dobiven je transformacijom ordinalne ljestvice mjera preferencije polazne metode u kardinalnu ljestvicu, primjenom neli- nearne transformacije preko izraza (4) i sadrži okolišne indikatore prenamjene prostora površinskog kopa. Kao okolišni indikator, u sustav je uključeno vrijeme trajanja eksploatacije. Značajka primjera prikazanog u tablici 3. je ta što je on dio šireg ekspertnog podsustava, što je u ovom sluča- ju vidljivo iz broja članova kardinalne ljestvice (M =20), koji predstavljaju broj svih članova podsustava. Aproksi- macija particija lingvističkih varijabli je provedena line- arnim (trokutastim) funkcijama udjela. Na temelju ove metode, a u skladu s konceptom ne- linearnosti prikazanom na slici 4., moguće je provesti daljnju podjelu (particiju) lingvističke varijable modelom prikazanom na slici 5., povećavajući broj particija (Cox, 1999). Rud.-geol.-naft. zb., Vol. 25, 2012. D. Rumenjak: Ekspertni sustavi za zaštitu okoliša...112 Tablica 3. Lingvističke varijable prihvatljivosti za okoliš, s pripadnim ljestvicama mjera preferencije (MP) i funkcijama udjela kao dio šireg eksper- tnog sustava Table 3. Linguistic variables acceptability for environment, with appropriate scales of measure of preference (MP) and membership values as a part of expert system Lingvističke varijable Opisna (nominalna) ljestvica Ordinalne vrijednosti po metodi Kardinalna ljestvica M=20 Funkcije udjela Particija prihvatljivosti Particija neprihvatljivosti 1.1. Društveni značaj prenamjene velik, ekskluzivni sadržaji; - 4 - 0.18 1 0 relativno velik, vrlo vrijedni sadržaji; - - 3 - - 0.13 0.7 0.4 mali, manje korisni sadržaji; - 2 - 0.08 0.4 0.7 nikakav, nema posebne namjene - 1 - 0.03 0.2 0.9 1.2. Prilagodljivost projektnog rješenja oblikovanog prostora budućoj namjeni izvrsna, u potpunosti prilagodljivo; - 4 - 0.18 1 0 prihvatljiva, prilagodljivo s malim dodatnim zahvatima; - - 3 - - 0.13 0.7 0.4 loša, prilagodljivo s velikim dodatnim zahvatima; - - 2 - - 0.08 0.4 0.7 neprihvatljiva, nema posebne namjene ili prilagodba u potpunosti mijenja projektna rješenja. - 1 - 0.03 0.2 0.9 Trajanje eksploatacije preko 30 god. - 4 - 0.18 1 0 do 30 god. - 3 - 0.13 0.7 0.4 do 20 god. - 2 - 0.08 0.4 0.7 do 10 god. - 1 - 0.03 0.2 0.9 Utjecaj buduće namjene na okoliš povoljan, ne narušava eko- sustav; - 4 - 0.18 1 0 uvjetno povoljan, povremeno može utjecati na eko-sustav; - 3 - 0.13 0.7 0.4 nepovoljan, stalno može utjecati na eko-sustav; - - 2 - - 0.08 0.4 0.7 izrazito nepovoljan, nema posebne namjene ili se izravno utječe na eko-sustav. - - - 1 - - 0.03 0.2 0.9 Agregacija rezultata Način na koji se dolazi do mjera preferencije te sam izbor varijanti ili odlučivanje u ekspertnim sustavima naziva se još i agregacijom. Agregacija u ekspertnim su- stavima odvija se prema pravilima, koja su dio samog sustava, a ponekad i određenih pravila izvan sustava (me- ta-pravila). Agregacija ovisi o obrascu na kojem se ek- spertni sustav postavlja, npr. utilitarna (poslovni sustavi) ili neutilitarna agregacija (zaštita okoliša). „Trade-off“ agregacija je određeni kompromis izme- đu utilitarne i neutilitarne koncepcije odlučivanja. Za ek- spertne sustave koji koriste neizrazitu skupove (brojeve) moraju se primijeniti pravila neizrazite aritmetike. Opći izraz za proračun neizrazitom aritmetikom pri- mjenom ekstenzijskog principa na zbroj mjera preferen- cije koji se prikazuju neizrazitim brojevima, primjenom intervalne aritmetike (Buckley & Eslami, 2002) je : (5) gdje je : funkcija udjela za rezultat agregacije koji je neizraziti skup, ; član neizrazitog skupa mje- ra preferencije , : funkcija udjela člana skupa preferencije u terminima neizrazite logike, gornja najmanja granica (< max za kontinuirane funkcije). Agregacija približnim zaključivanjem u neizrazitoj logici daje mogućnost primjene logičkih pravila u iz- boru odluke. Opći slučaj približnog zaključivanja (eng. approximate reasoning), koje se služi neizrazitim sku- povima može se prikazati kao proces koji se sastoji od kompozicije pravila i operatora te dekompozicije rezulta- ta zaključivanja do ekvivalentne vrijednosti (Cox, 1999): 113 Rud.-geol.-naft. zb., Vol. 25, 2012. D. Rumenjak: Ekspertni sustavi za zaštitu okoliša... [ ] [ ] )( ),()( 1 Cdecomp yxxcompy RL PP n i c M ⇐ℜ •⇐ = µµµ (6) gdje je član : [ ] [ ]),( yxxcomp RL PP µµ • izraz za kompoziciju unutar pravila, n i M 1= ; agregacija rezutata, )(Cdecomp⇐ℜ ; dekompozicija (defazifikacija, engl. defuzzification) do ekvivalentne vrijednosti neizrazitog skupa. Odluka o izabranoj varijanti u neizrazitoj logici dono- si se temeljem funkcije odlučivanja : { { }          = ∈ ∗ )(),....(),(minmax)( 21 aaaa rCCCAaD µµµµ (7)  r j jCD 1= = gdje su { }[ ])(),....(),( 21 aaa rCCC µµµ : pojedine vrijedno- sti udjela kriterija po varijantama, Dµ *)(a : funkcija udjela izabrane varijante koja se još naziva funkcijom odlučivanja, Cj: kriterij odlučivanja, μCj: funkcija udjela varijante po kriteriju dobivena iz skupa kriterija, tj. nji- hovih udjela, D; model odlučivanja. Rezultat se prikazuje funkcijom udjela i položajnom ljestvicom. Razvijeni model (7) temeljit će se utilitarnoj ili neutili- tarnom obrascu odlučivanja, u ovisnosti od matematičke definicije vrednovanja mjera preferencije po kriteriju od- lučivanja, Cj (Rumenjak, 2009). Zaključak Ekspertni sustavi za izbor odluka zahtijevaju se zbog složenosti problematike odlučivanja u zaštiti okoliša, koja se mora povezati s poslovnim odlučivanjem. Ekspertni sustavi grade se temeljem indikatora i mjera preferencije. Indikatori su baza ekspertnog znanja, dok se odlučivanje provodi temeljem mjera preferencije, koje se definiraju kao funkcije okolišnih indikatora. U procjeni utjecaj na okoliš već je našao primjenu jedan broj ekspertnih su- stava, dok su neki samo predloženi. Neizrazita logika i lingvističke varijable, na kojoj se ekspertni sustavi mogu temeljiti, u informacijskom, preferencijalnom te u pogle- du postavljanja pravila i agregacije rezultata ispunjavaju sve zahtjeve ekspertnih sustava u zaštiti okoliša. Literatura Buckley J.J., Eslami E. (2002): An Introduction to Fuzzy Logic and Fu- zzy Sets, Physica-Verlag, pp.63, Heidelberg Cox E. (1999): The Fuzzy Systems Handbook, AP Professional, pp 290 (1), pp. 682 (2), San Diego Craig J.R., Vaughan D.J., Skiner B.J. (2011): Earth Resources and the Environment, Prentice-Hall, pp. 82, Boston Dasgupta, P. (2001): Human Well- Being and the Natural Environment, Oxford University Press, Oxford, pp. 91 (1), pp.240(2) Jackson, P. (1998): Introduction to Expert Systems, Adison-Wesley, pp. 2, Harlow Krasić D. (2005): Doktorski rad, Rudarsko-geološko-naftni fakultet, Zagreb Marušič J. (1999): Okoljevarstvene presoje v okviru prostorskega na- črtovanja na ravni občine - Modeli v prostorskem načrtovanju, 3. zvezak, Geoinformacijski centar R. Slovenije, str.10, Ljubljana Möllerherm S., Martens P.N. Patiero-Fernandez J.B. (2005): Deve- lopment of a Sustainability Evaluation System, Second Internatio- nal Conference Sustainable Development Indicators in the Mineral Industry - SDIMI 2005, 18-20 May, P.419, Aachen Pastakia C.M.R (1998): The Rapid Impact Assessment Matrix (RIAM)- A New Tool for Environmental Impact Assessment, u Jensen K. ed., Environmental Impact Assessment - Using the Rapid Impact Asse- ssment Matrix (RIAM), Olsen&Olsen, 8 pp, Fredensburg Patiero-Fernandez J.B., Martens P.N., Möllerhaim S. (2005): Deve- lopment of Environmental Indicators for the Raw Mineral Industry, Second International Conference Sustainable Development Indi- cators in the Mineral Industry - SDIMI 2005, 18-20 May, P.731, Aachen Pravdić V. (2008.): Review of the book Graham Harris: Seeking Sustai- nability in an Age of Complexity, Kem. ind. 57 (5) 277-279, Zagreb Ross, T.J. (1995). Fuzzy Logic for Engineering Applications, Mc-Graw Hill, pp. 267, New York Rumenjak D.(2009): Doktorski rad, Rudarsko-geološko-naftni fakultet u Zagrebu Rumenjak D. (2006), Metodologija izbora varijanti u procjeni utjecaja na okoliš i cost-benefit analizi, IX. Međunarodni Simpozij gospo- darenja otpadom. 15-18 November, P.755, Zagreb Rumenjak D., Rajković D., Salopek B. (2007): Some Possibilities for Construction of Linguistic Variables for Sustainable Decision Ma- king, 3rd International Conference on Sustainable Development Indicators in the Mineral Industry - SDIMI 2007, 17-20 June, P. 211, Milos Island Rumenjak, D. D. Rajković and B. Salopek (2008), Some solutions for fuzzy approximate reasoning in sustainable development decision- making and monitoring phase using environmental indicators, 1 st International Conference on Indicators for Land Rehabilitation and Sustainable Development, 22-23 October, P. 190, Bejing Rumenjak D., Salopek B, Rajković D. (2004): Change of Decision Making Principle in Environmental Impact Assessment Applied on Screening Matrices, International Conference on Advances in Min- eral Resources and Environmental Geotechnology (AMIREG), 7-9 June, P.751, Chania Russel S., Norvig P. (2010): Artificial Intelligence - A Modern Approach, Prentice Hall, pp. 613, Boston X.Zhu, D.Li, Komnistas K. (2007): Development of an indicator based revegetation SDSS in areas affected by mining activities, 3rd Inter- national Conference on Sustainable Development Indicators in the Mineral Industry - SDIMI 2007, 17-20 June, P.267, Milos Island, Rud.-geol.-naft. zb., Vol. 25, 2012. D. Rumenjak: Ekspertni sustavi za zaštitu okoliša...114 EXPERT SYSTEMS FOR ENVIRONMENTAL PROTECTION IN MINING Expert systems are used in various activities with the main goal of decision-making in intelligent and transpa- rent way. In business decision-making, such systems are directed to pure operational and economical decisions. In environmental protection they have to accept the specifi- city of business sector for which they have been intended, but also use environmental reasons and evaluate them differently than economic. The expert systems connect rules and knowledge as it is shown for simple case appropriate for mining, in Fig.1. The knowledge in expert systems for environmen- tal protection are often expressed by values known as environmental indicators, of which some examples for mining are given by expressions containing appropriate aggregation formulas (1), (2) and (3). To use such indica- tors in evaluation procedures, the measure of preferences (MP) must be joined to them, of which the most common is utility. The others MPs are possible (welfare, well-be- ing), but are mostly still under development. For practical purposes, MPs are represented by scales of type shown in Table 1. The evaluation procedures (decision-making) with such scales are shown in Table 2. Two basic types of scale are reco- gnizable, ordinal and cardinal, and transformation from ordinal to cardinal scales is given by expression (4). In Croatia some expert systems for mining are alre- ady in use. Method of Vulnerability of Environment with Matrix of Interactions is relatively widespread in Croatia and in neighbouring Slovenia, where it originated. It de- fines scales from 1 to 5. The number of indicators used is usually about 20 but could vary. The aggregation co- uld be performed either over main parts of technological operations and/or over criteria. Method of Reshaping and Reclamation (MRR), designed especially for mining is proposed in 2005. It defines 40 indicators in ordinal/ratio ranges from 1 to 3 or sometimes 5, divided in three main groups: natural, economical-technical and ecological. The aggregation is similarly done as in previous method, with predefined range values. Rapid Impact Assessment Ma- trix or RIAM method is not yet amongst those used in mineral industry, but its use has already been considered and recommended as that suitable for expert systems for mining. The recent development of expert systems for mi- ning is the use of fuzzy variables (linguistic variable or LV) of which general one is given in Fig. 2, representing properite of acceptability for environment. The LVs are connecting together indicators and measures of preferen- ces in the same variable. The well known feature of LVs is representation of uncertainty by more partitions of lin- guistic variables, Fig.3. The second is non-linearity, Fig.4, which is characteristics of many processes in nature and environment, and is achieved by overlapping of partitions of such variables. The case of duration of exploitation is given for non-linearity by the example of LV in Fig.5. The importance of duration of exploitation as environmental variable and its course of influence is still under debate in literature and administrative procedures like environmen- tal impact assessment. Here, as result of such debates, non-linearity for the variable is supposed- The variables for the system proposed on those con- siderations given in Table 3, are Social significance of reshaping and reclamation according to future use, Flexi- bility of project solutions of reclamation to future uses on the location, Duration of exploitation, and Impacts of fu- ture uses on environment. They are originally from MRR method but here adapted to fuzzy logic. The aggregation of the results of systems based on fuzzy logic could be provided in utilitarian way, using expression (5) or non- utilitarian way more suitable for environmental decisi- on-making, using (6) and (7). Both approaches could be combined to achieve the desirable effects for the system. 
work_epdrsr6kbfaptbjgtkit66ji6i	----	Microsoft Word - 01 A knowledge Conversion _김진성.doc International Journal of Fuzzy Logic and Intelligent Systems, vol. 11, no. 1, March 2011, pp. 1-7 DOI : 10.591/IJFIS.2011.11.1.001 1 A knowledge Conversion Tool for Expert Systems Jin S. Kim1 School of Business Administration, Jeonju University 303 Cheonjam-Ro, Wansan-Gu, Jeonju, Jeonbuk 560-759, South Korea Abstract Most of expert systems use the text-oriented knowledge bases. However, knowledge management using the knowledge bases is considered as a huge burden to the knowledge workers because it includes some troublesome works. It includes chasing and/or checking activities on Consistency, Redundancy, Circulation, and Refinement of the knowledge. In those cases, we consider that they could reduce the burdens by using relational database management systems-based knowledge management infrastructure and convert the knowledge into one of easy forms human can understand. Furthermore they could concentrate on the knowledge itself with the support of the systems. To meet the expectations, in this study, we have tried to develop a general-purposed knowledge conversion tool for expert systems. Especially, this study is focused on the knowledge conversions among text-oriented knowledge base, relational database knowledge base, and decision tree. Key Words: Decision tree, Expert system, Inference, Intelligent system, Knowledge base, Relational database. 1. Introduction The two core components in developing an expert system (ES) are the inference engine (IE)/mechanism and knowledge base (KB) construction. Where, the inference mechanism was the fundamental technology to improve the total performance of the ES. Therefore, various inference mechanisms were proposed by the researchers after the proposition of the MYCIN [3]. However, the speed of inference of IE and efficient management of the KB were still remained as a tackling point of ES development [1, 2]. In addition, the acquisitions and managing of knowledge were regarded as a huge burden to the ES managers and knowledge workers because the processes include various complex routines. The routines include Checking of Consistency / Redundancy / Unlimited Circulation of the inference (rules), Refinement of the rules, Conversions of the rules, and etc. To reduce the burdens, this study is focused on the Conversions of the rules and the Checking of Unlimited Circulation of the Inference Rules. Since the ES have been widely used in the domains where mathematical models were could not be easily built [1, 2], human expert's knowledge was stored frequently in the KB as a form of inference rule. Then the rules and knowledge were frequently expressed by the OAV (Object-Attribute-Value) typed IF-THEN rule. However, the human expert's knowledge was sometimes conflict, redundant, and linked in an unlimited circulation. To minimize the errors may come from the human expert knowledge this study proposes a knowledge conversion tool to transform the traditional IF-THEN rules into a Decision Tree (DT). By using the DT, ES managers may reduce their burdens in managing KB. At the first step, the relational database management system (RDBMS) was used to construct KB and transform/manage it. The advantages of using RDBMS-based KB are could be summarized as follows: ·In anytime and anywhere, by using the communication network, the ES managers or knowledge workers can share the knowledge stored in the RDBMS-based KB [4]. ·Knowledge management in RDBMS is more efficient than the text-based such works. ES managers can execute SQL- based various knowledge management routines such as Create, Search, Retrieve, Update, Sort, Refinement, View/Group by using the Keywords or Logical Conditions and etc. with the RDBMS-based KB [2]. The remainder of this paper is organized as follow: The research methodology was proposed in Section 2. The prototype system and the results of implementation were presented in Section 3. The conclusions and future works are finally given in Section 4. 2. Research Methodology 2.1 Knowledge Elicitation To construct the KB, a knowledge elicitation routine is required. The routine is commonly regarded as a major obstacle in the process of ES development. Generally, the knowledge used in a KB construction could be acquired through one of two ways: either Manual or Automatic acquisition [5]. Traditionally, knowledge workers gathered the knowledge through the interactions with the human experts. However, the method was turn out to be a bottleneck, causing delays in the ES development [7]. One of the efficient knowledge acquisition and/or extraction mechanisms to overcome the limitations, Data mining (DM) was introduced. The DM is motivated by the need of new Manuscript received Jan. 11, 2011; revised Jan. 27, 2011; Accepted Feb. 11, 2011. 1 Corresponding Authors International Journal of Fuzzy Logic and Intelligent Systems, vol. 11, no. 1, March 2011 2 techniques to help analyzing, understanding or visualizing the huge amount of data gathered from business and scientific applications [6]. This study is based on those two knowledge acquisition routes. Control Knowledge Doma in Knowledge Doma in Ontology Doma in Models Library Knowledge Crea tor Intermedia te knowledge structure Knowledge Ba se Genera tor Knowledge Tra nsformer acquire & modify Doma in Expert Knowledge Elicitation used by ES Generator Knowledge Expresser IF-THEN Rules Decision Tree Cognitive Ma p New Ca ses Inference Result Former Ca ses use by generate transform transformation use by Interact with Inference Engine RDB-KBText-KB #n#n Knowledge #1 RDB #1 used by :: Restore & Revise Knowledge Maintenance RDBMS:- RDBMS-ba sed Knowledge Ma intena nce : Redundancy, Sort, Keyword Search, Relations, etc. Logic:- Logic-ba sed Knowledge Ma intena nce : Unlimited Circulation, Consistency, Economic, etc. Fig. 1 Research methodology Entire processes of our proposed methodology were graphically presented in Fig. 1. It includes six main components namely: Knowledge Elicitation, Library, ES Generator, Knowledge Expresser, Inference Engine, and Knowledge Maintenance. The former version of the methodology was proposed by Kim [1] and Rafea et al. [4]. However, the Fig. 1 shows the expanded/revised components and architecture. The detailed activities of the components are as follows: ·Library: Library has both reusable domain knowledge and control knowledge such as domain ontology, domain models, and control knowledge. It is used to create a new knowledge and refine the old KB with it. ·Knowledge Elicitation: The main functions of knowledge elicitor are to create, maintain, and restore knowledge elicited from the external input, fetch the relevant knowledge components from the library, and transform the knowledge into an appropriate knowledge structure. Especially, the detailed maintenance is executed in the Knowledge Maintenance component. ·ES Generator: Automatically, it generates an executable knowledge, which corresponds to the intermediate knowledge generated above. It contains knowledge creator, knowledge transformer, and knowledge base generator. During the knowledge conversion, ES Generator uses the RDBMS to restore and revise her KBs. The text-oriented knowledge was stored in Text-KB and RDB- oriented transformed knowledge was stored in RDB-KB. ·Knowledge Expresser: It supports the three knowledge expression methods such as, IF-THEN Rules, DT and Cognitive Map. It also could help end-users to understand the structure of knowledge. ·Knowledge Maintenance: It has two main procedures RDBMS-based knowledge maintenance and Logic-based knowledge maintenance. RDBMS-based knowledge maintenance supports the DB-oriented maintenance activities such as Duplication Check, Sort, Keyword Search, Checking Relationships and etc. Then the Logic-based knowledge maintenance support the activities: Checking of Unlimited Circulation, Consistency, Economics of the Rules and etc. The Fig. 1 shows all the components used in our proposed research methodology. 2.2 Knowledge Conversion This study supports three knowledge conversion and/or transformation methods. In the Fig. 2, we can see the procedures used in the conversions from IF-THEN rules to DT. The detailed knowledge inference mechanism and other procedures were introduced in Kim's study [2]. j = IFsStart, CurrrentIFLevel = CurrentTHENLevel = CurrentTopIFLevel = CurrentIFNodeCount = CurrentTHENNodeCount = TempToCount = 0 RootLevel = 1 For i =0 to number of rules For i = 0 to MaxLevels (=30) For i = 0 to MaxLevels (=30) Levels(i) = New Class Level While (Not IsDBNull (IF value in DB)) StoredRule = False StoredNode = CompareWithStoredNodes with IF value Set CurrentIFLevel and CurrentNodeCount IF Not StoredNode IFLevel.Nodes(CurrentIFNodeCount).NodeValue = IF Value IFLevel.NodeCount = IFLevel.NodeCount + 1 Yes Levels(i) = New Class Level IFLevel.Nodes.RuleNos(RuleCount) = i IFLevel.Nodes.RuleCount = IFLevel.Nodes.RuleCount + 1 IF CurrentTopIFLevel < CurrentIFLevel Yes No No IF i= stored Rule#Stored Rule = True CurrentRuleCount = IFLevel.Nodes.RuleCount Yes No A B A knowledge Conversion Tool for Expert Systems 3 CurrentTopIFLevel = CurrentIFLevel IF RootLevel < CurrentIFLevel RootLevel = CurrentIFLevel Yes C j = j+1 While (Not IsDBNull (IF value in DB)) A B For i =0 to number of rules While (Not IsDBNull (IF value in DB)) StoredNode = CompareWithStoredNodes with THEN value Set CurrentTHENLevel, CurrentNodeCount (Stored or Increased) IF Not StoredNode Nodes(CurrentNodeCount).NodeValue = THEN Value THENLevel.NodeCount = THENLevel.NodeCount + 1 Yes THENLevel.Nodes.RuleNos(RuleCount) = i RuleCount = RuleCount + 1 IF RootLevel < CurrentTHENLevel RootLevel = CurrentTHENLevel Yes No No C TempToCount = IFLvel.Nodes.ToCount IFLevel.Nodes.ToNodes(TempToCount) = New ToNode( ) IFLevel.Nodes.ToNodes(TempToCount).ToLevel = CurrentTHENLevel IFLevel.Nodes.ToNodes(TempToCount).ToNode = CurrentTHENNodeCount IFLevel.Nodes.ToCount = IFLevel.Nodes.ToCount + 1 IF i= stored Rule# No Yes Stored Rule = True CurrentRuleCount = THENLevel.Nodes.RuleCount Fig. 2 Conversions from IF-THEN rules to DT 2.3 Checking of Unlimited Circulation of Inference Rules Through the Knowledge Conversion process all the inference rules were stratified. Using the DT, after the process, we can check the unlimited circulation of the rules. To find out the circulation, we just need to check the Level number and Node numbers of every rule. Table 1 shows the pseudo code for the mechanism. The checking mechanism is so simple because the previous procedures offered important information required in checking circulation. Table 1 Pseudo code for the checking of circulation For i=0 to Root/Top Level Current Level = i For j = 0 To Node count in Levels i For k = 0 To Count of connection If Level of target node < Current level Then Circulation = True End If Next k Next j Next i If Not Circulation Then There's no circulation End 3. Implementation and Results 3.1 Open Knowledge Base To implement our proposed mechanism, we developed a prototype ES development tool K-Expert (ver. 1.0). The tool was developed by using the Visual Studio 2010 in a Windows 7 environment. The main KBs were constructed by MS-Access database. The first sample data set used in the implementation is the Classification of Animals. The Fig. 3 shows the main window of the K-Expert and process of Open knowledge base. Fig. 3 The knowledge base stored in a MS-Access DB table The above table shows the inference rules and variables, and the below table contains the names of the variable and the alternative questions for the inference rules shown in the above table. The alternative questions and values can replace the variables when the rule is executed. 3.2 Knowledge Conversion In the text box at the bottom of Fig. 4, we can see the information of Decision Tree (level #, Node#, Rule#, and Target (to) Node#). International Journal of Fuzzy Logic and Intelligent Systems, vol. 11, no. 1, March 2011 4 Fig. 4 The Levels and Nodes in the Decision Tree The tool has drawn the DT information into a graphical expression. Table 2 shows the descriptions of the information. Table 2 The information used to draw the DT Level 0: '' The number of the Level = 0 (Node0) Animal has hair '' The node number = 0 '' The node value is 'Animal has hair' Rules: 1. '' The rule number = 1 To-Node: 1-0. " The target (to) node number = 0 in Level 1 (Node1) Animal gives milk Rules: 2.16. To-Node: 1-0. (Node2) Animal has feathers Rules: 3. To-Node: 1-1. (Node3) Animal flies Rules: 4. To-Node: 1-1. (Node4) Animal lays eggs ..... Fig. 5 shows the whole structure of the decision tree. On that figure, the level 0 is shown at the top of the tree and level 4 is shown at the bottom of it. Where, the users could change the positions of each node and the figure shows the changed shape. 3.3 Checking of Unlimited Circulation Fig. 6 shows the process of Checking of Unlimited Circulation. The result of the process is shown in the textbox at the bottom of the window. The sentence below shows that the system found one circulation. ---------------------------------------------------------------------------- Circulation Level# = 1, Node# = 0, Animal is mammal ---------------------------------------------------------------------------- Fig. 5 A Graphical representation of the Decision Tree The meaning of the sentence above is as follows: ---------------------------------------------------------------------------- "One circulation is found at the Node #1 in the Level #1. Then the value of the node is 'Animal is mammal'." ---------------------------------------------------------------------------- The Fig. 7 shows the circulation (unlimited loop) on the decision tree. Where the first link came from #1 Node (Animal gives milk) in the second link came from Level 0, and #0 Node (Animal is mammal) in the Level 1. Therefore, the relationship of the two nodes could be represented as a circulation or unlimited loop. Fig. 6 Checking of Circulation A knowledge Conversion Tool for Expert Systems 5 Fig. 7 The circulation on the decision tree 3.4 Application of Classification Problem In this section, to verify the usefulness of our proposed expert systems development tool, we used the second sample data set titled as 'Blood Transfusion Data Set' gathered from the Blood Transfusion Service Center in Hsin-Chu City, Taiwan. This is a traditional classification problem [8]. The result of the classification was used as a target group in a marketing project. Table 3 shows the sample data set and brief descriptions. Table 3 Blood Transfusion Data Set *) Field descriptions R (Recency): months since last donation F (Frequency): total number of donation M (Monetary): total blood donated in c.c. T (Time): months since first donation D (Donation): 1 stands for donating blood; 0 stands for not donating blood R, F, M, T, D 4,4,1000,4 ,0 2 ,7,1750,14 ,1 1 ,12,3000,35 ,0 2 ,9,2250,22 ,1 5 ,46,11500,98 ,1 4 ,23,5750,58 ,0 0 ,3,750,4 ,0 2 ,10,2500,28 ,1 1 ,13,3250,47 ,0 2 ,6,1500,15 ,1 ... The Fig. 8 shows the result of Knowledge Elicitation. In this study the SPSS Modeler 13 was used to support the procedure. Totally 11 inference rules were generated. Then the Table 4 shows the text-oriented KB preserving the inference rules shown in the Fig. 8. Fig. 8 Knowledge Elicitation Table 4 Text-oriented KB (inference rules) Inference rules Rule 1 if Time > 51.500 and Frequency < 6.500 then 0.000 Rule 2 if Recency > 9.500 then 0.000 Rule 3 if Monetary > 4625.000 and Frequency < 21.000 then 1.000 Rule 4 if Monetary > 4625.000 and Recency < 5.500 and Frequency < 21.000 then 1.000 ... Rule 10 if Frequency > 18.500 and Recency < 5.500 then 1.000 Rule 11 if Frequency > 18.500 then 1.000 Fig. 9 shows the RDB converted from the inference rules in Table 4, and Fig. 10 shows its graphical representation of the decision tree. International Journal of Fuzzy Logic and Intelligent Systems, vol. 11, no. 1, March 2011 6 Fig. 9 The RDB-KB Fig. 10 Decision Tree of the Blood Fusion data set The Fig. 11 shows the Economics of the Rules. The inference rules #6, #7, #8, #9, and #10 would be replaced be the rule #11 because the shortest rule #11 can replace the other rules (#6- #11). Fig. 11 The Economics of the rules If there is a Circulation, it would be discovered by our proposed mechanism. However, there is no circulation because the rules in Fig. 8 were automatically generated by the system. After the construction of the initial KB, the human expert could engage into the revision of the KB. In that case, our proposed mechanism could help the expert to revise the KB without the concerns with unexpected errors. 4. Conclusions In this study, we proposed a RDBMS-based knowledge conversion tool for ES development. The two main functions we have focused on this study are the Conversions of the rules and the Checking of Unlimited Circulation of the Inference Rules. Including those functions, the tool has six main components Library, Knowledge Elicitation, ES Generator, Knowledge Expresser, Inference Engine, and Knowledge Maintenance. To show the processes executed by the tool, we developed the prototype system K-Expert. The tool was developed by using the Visual Studio 2010 and MS-Access Database. Through the study, it is expected that our proposed mechanism would have significant effects on the knowledge management in using and/or development of ES. Especially, it could help to reduce the burden of knowledge management. The further research topics still remain as follows. First, to help the knowledge managers efficiently, we need to develop the other reasonable Knowledge Maintenance routines. Second, the other intelligent decision support mechanisms such as fuzzy logic, cognitive map, and neural networks would be improve the reasoning ability of ES. References [1] J. S. Kim, "Prediction of User's Preference by using Fuzzy Rule & RDB Inference: A Cosmetic Brand Selection," International Journal of Fuzzy Logic and Intelligent Systems, vol.5, no.4, pp.353-359, 2005 [2] J. S. Kim, "RDB-based Automatic Knowledge Acquisition and Forward Inference Mechanism for Self-Evolving Expert Systems," Journal of Fuzzy Logic and Intelligent Systems, vol.13, no.6, pp.743-748, 2003. [3] B. G. Buchanan and E. H. Shortliffe, Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Projects, Addison-Wesley, Reading, MA, 1984. [4] H. Yan, Y. Jiang, J. Zheng, B. Fu, S. Xiao, and C. Peng, "The Internet-based Knowledge Acquisition and Management Method to Construct Large-scale Distributed Medical Expert Systems," Computer Methods and Programs in Biomedicine, vol.74, no.1, pp.1-10, 2004. [5] A. Rafea, H. Hassen, and M. Hazman, "Automatic Knowledge Acquisition Tool for Irrigation and Fertilization Expert Systems," Expert Systems with Applications, vol.24, no.1, A knowledge Conversion Tool for Expert Systems 7 pp.49-57, 2003. [6] M.S. Chen, J. Han and P.S. Yu, "Data mining: An Overview from a Database Perspective," IEEE Transactions on Knowledge and Data Engineering, vol.8, no.6, pp.866-883, 1996. [7] I. Hatzilygeroudis and J. Prentzas, "Integrating (rules, neural networks) and Cases for Knowledge Representation and Reasoning in Expert Systems," Expert Systems with Applications, vol.27, no.1, pp.63-75, 2004. [8] I-C Yeh, K-J Yang, and T-M Ting, "Knowledge Discovery on RFM Model using Bernoulli Sequence," Expert Systems with Applications, vol.36, no.3, pp.5866-5871, 2009. Jin S. Kim Associate Professor of MIS in Jeonju University Research Areas: Fuzzy theory, artificial intelligence, intelligent decision support, expert systems, and semantic web E-mail : kimjs@jj.ac.kr 
work_eqp7tonlgvcolfqw4r6j5ps5bi	----	An expert system to predict protein thermostability using decision tree Expert Systems with Applications 36 (2009) 9007–9014 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a An expert system to predict protein thermostability using decision tree Li-Cheng Wu a, Jian-Xin Lee b, Hsien-Da Huang c, Baw-Juine Liu d, Jorng-Tzong Horng a,b,e,* a Institute of System Biology and Bioinformatics, National Central University, Taiwan b Department of Computer Science and Information Engineering, National Central University, Taiwan c Institute of Bioinformatics, National Chiao-Tung University, Taiwan d Computer Science and Information Engineering, Yuan Ze University, Taiwan e Department of Bioinformatics, Asia University a r t i c l e i n f o Keywords: Expert system Machine learning Bioinformatics Protein thermostability Decision Tree 0957-4174/$ - see front matter � 2008 Elsevier Ltd. A doi:10.1016/j.eswa.2008.12.020 * Corresponding author. Address: Department of Central University, Taiwan. E-mail address: horng@db.csie.ncu.edu.tw (J.-T. Ho a b s t r a c t Protein thermostability information is closely linked to commercial production of many biomaterials. Recent developments have shown that amino acid composition, special sequence patterns and hydrogen bonds, disulfide bonds, salt bridges and so on are of considerable importance to thermostability. In this study, we present a system to integrate these various factors that predict protein thermostability. In this study, the features of proteins in the PGTdb are analyzed. We consider both structure and sequence features and correlation coefficients are incorporated into the feature selection algorithm. Machine learning algo- rithms are then used to develop identification systems and performances between the different algorithms are compared. In this research, two features, (E + F + M + R)/residue and charged/non-charged, are found to be critical to the thermostability of proteins. Although the sequence and structural models achieve a higher accuracy, sequence-only models provides sufficient accuracy for sequence-only thermostability prediction. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction Protein stability is intimately connected with protein folding and all proteins have to be folded into their final active state to be active and stable. In biotechnology, chemical reactions need to be performed at high temperatures to decrease reaction time. However, many proteins are not very stable when heated. Research is needed that helps proteins to remain active and stable at high temperatures, which will overcome many limitations to their industrial applications (Huang et al., 2004). There are four thermo- stability classes of protein, psychrophilic, mesophilic, thermophilic and hyperthermophilic, separated by denaturation temperature of 20 �C, 50 �C and 80 �C (Vieille & Zeikus, 2001). Current research has shown that the properties of proteins such as number of ion pairs, and salt bridges are considerably related to protein thermostability (Vieille & Zeikus, 2001). Research has shown that single point mutation, when related to thermostability, may allow thermosta- bility prediction (Capriotti, Fariselli, & Casadio, 2004; Capriotti, Fariselli, & Casadio, 2005; Saraboji, Gromiha, & Ponnuswamy, 2006). Instead of analyzing mutant proteins, research is needed that investigates the global relationships between thermostability, protein structure and sequence properties. Recent investigations have shown that features such as amino acid composition and ion pairs are related to the thermostability. ll rights reserved. Computer Science, National rng). However, a general model of how these features are related to thermostability remains a scientific challenge. Our previous work (Huang et al., 2004) has identified a general model relating ther- mostability and ion pairs. Since structured folding is a complex question, practically a model of thermostability based on sequence information would be more useful than one based on other types of analysis. Our goal is to develop a predictive model for thermostability based on sequence and structural features. Given a protein se- quence/structure, the system aims to classify proteins into three different thermostability classes: mesophilic, thermophilic and hyperthermophilic. We aim to achieve high accuracy, sensitivity, specificity and precision. Higher accuracy may be available if both specific protein structure and sequence information are available, but a general model that will accept sequence-only information is also constructed. 2. Related works The optimal growth temperature of an organism indicates the temperature at which the organism’s growth is most rapid. The prokaryotic Growth Temperature Database (PGTdb1) is a database of prokaryotic thermostability information. We adapted the thermo- stability information from the PGTdb for this work. The PDB2 is the 1 http://pgtdb.csie.ncu.edu.tw/. 2 http://www.rcsb.org/pdb/. mailto:horng@db.csie.ncu.edu.tw http://pgtdb.csie.ncu.edu.tw/ http://www.rcsb.org/pdb/ http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 9008 L.-C. Wu et al. / Expert Systems with Applications 36 (2009) 9007–9014 single worldwide repository for the processing and distribution of 3- D structure data of large molecules including proteins and nucleic acids (Berman et al., 2000). Protein sequence and structure informa- tion are collected through the PDB database. Prediction of protein thermostability based on single mutation points has shown that prediction of thermostability is possible (Capriotti et al., 2004, 2005; Farias, van der Linden, Rego, Araujo, & Bonato, 2004; Saraboji et al., 2006). Previous research (Haney, Stees, & Konisky, 1999; Vieille & Zeikus, 2001) has shown that many special properties of proteins are related to protein thermo- stability. Usually, protein stability is intimately connected with the protein three-dimensional structure. It is believed that sequence determines the final folding of protein although three-dimensional free folding of an arbitrary protein sequence is still a scientific challenge. The thermostability of proteins is directly connected to its structural stability. Thus, the features maintaining the structure folding are often considered as significant factors in protein ther- mal stability. It has been demonstrated that there is a relationship between thermostability and sequence information. Amino acid composition and intrinsic propensity are thought to play important roles in thermal stability (Vieille & Zeikus, 2001). For instance, the protein’s content of hydrophobic amino acids is reasonably related to its thermal stability. Many mutant rules, such as Trp ? Tyr, Cys ? Ile, have been proposed to enhance thermostability (Grom- iha, Oobatake, & Sarai, 1999). In addition to the amino acid compo- sition, research has also found that there is a relationship between sequence pattern and optimal growth temperature. The pattern [EdH] and [EdT] are favored by mesophilic proteins (Liang, Huang, Ko, & Hwang, 2005). Sometimes sequence analysis can focus on a special pattern and its corresponding structure. The sequence pat- tern is transformed into a favored secondary structure by the data- base records giving a good linear relationship between local structure and melting temperature (Chan et al., 2004). There are various structure features that show a relationship with thermostability. Secondary structure, such as a-helices, and intrahelical interactions within the protein have been analyzed and it has been concluded that high a-helical stability is necessary for protein thermostability (Petukhov, Kil, Kuramitsu, & Lanzov, 1997). Hydrogen bonds, ion pairs, hydrophobic interactions and disulfide bonds are thought to be the major forces affecting protein tertiary structure. Hydrogen bonds are non-covalent bonds be- tween donor and acceptor atoms and have been widely discussed in protein structure research. Some studies have indicated that an increase in hydrogen bonds contributes to thermostability (Vogt, Woell, & Argos, 1997), and that side-chain–side-chain hydrogen bonding plays the major role in this (Ragone, 2001). Ion pairs are another feature that has been widely discussed with respect to thermostability. Researchers have found that �1.1 ions pairs are lost for every 10 �C fall in thermostability per subunit (Gianese, Bossa, & Pascarella, 2002). When the distance between the ion pair is subdivided into <4 Å, 4–6 Å and 6–8 Å, it was found that the 6–8 Å group seem to play a more significant role in ther- mostability than the other two (Szilagyi & Zavodszky, 2000). Other possible properties, such as a hydrophobic core (Haney et al., 1999), electrostatic interactions and dielectric response (Dominy, Minoux, & Brooks, 2004), aromatic clusters on protein surface (Kannan & Vishveshwara, 2000) and B values reported in high-res- olution X-ray crystal structure (Parthasarathy & Murthy, 2000), are all considered to be related to thermostability. An important restriction is that high-resolution and a high-quality structure are needed to calculate such tertiary structure properties. Researchers usually utilize homology tools to produce theoretical protein structures when the real structure is unknown, and then compare those properties across the different thermostability classes. In addition to structural and sequence features, combination of features may also act as a key to thermostability (Farias et al., 2004). Research (Farias et al., 2004) has shown that the ratio between glutamic acid plus lysine and glutamine plus histidine gives a higher thermostability. This means that combination of amino acid (E + K)/(Q + H) may be important to protein thermo- stability. The above research shows that almost all properties of a protein are related to the thermostability based on observations that com- pare homologous protein pairs or several protein mutants. A global view of how these features are related to thermostability is needed. In this study, we incorporate several machine learning approach such as Naïve Bayes (Huang et al., 2004), SVM and neural network together with k-NN (Baumgartner et al., 2004) into an investiga- tion of the relationship between protein features and thermostability. 3. Materials and methods 3.1. System flow We have developed a data-mining system to analysis the rela- tionship between protein features and thermostability. The system builds a model based on given protein features and thermostability information. The system then takes the inputted protein features and tries to predict the thermostability of a given protein as one of three classes, mesophilic, thermophilic and hyperthermophilic. The system incorporates feature selection to eliminate low-con- tributing features and incorporates several machine learning ap- proaches. Fig. 1 shows the system flow. 3.2. Materials 3.2.1. Sampling data The optimal growth temperature of an organism indicates the temperature at which the organism’s growth is most rapid. Re- search on the source organism’s optimal growth temperature can be used in place of experimental thermostability data when com- paring proteins. Most proteins of an organism are reasonably ac- tive and quite stable at the optimal growth temperature of the organism. Two databases, PGTdb and PDB, are used in our research. We identified 5487 proteins for which there are both optimal growth temperature information in PGTdb and structural information in PDB. Fig. 2 show the data distribution for these proteins with re- spect to the temperature. In total, there are 41 different optimal temperatures for the pro- teins. Fig. 2 shows that the number of proteins with an optimal temperature of 37 �C is far more than others. Most studied pro- karyotic organisms are mesophilic and therefore over-sampling at this temperature in the PGTdb is not unexpected. As a result, the distribution is unbalanced and therefore any prediction will fo- cus on proteins with an optimal temperature of 37 �C, In other words, a prediction model that predicts every protein to have an optimum temperature of 37 �C will achieve high accuracy, but will fail to reach the goal of linking features and thermostability. In or- der to balance the temperature difference sample size, we carried out a Z-test to pick the temperatures that contain significantly more proteins then other temperatures. There were seven temper- atures that exceeded the Z-test limit, which indicated a barrier of 125 proteins. We select 125 proteins randomly by computer from each of these seven temperatures to balance the data. As a result, 1810 proteins from different temperatures form our sample data- set. The dataset consists of 878 mesophilic, 580 thermophilic and 352 hyperthermophilic proteins. Fig. 1. System flow. 3 NBC (http://www.cs.pdx.edu/~timm/dm/nbc.html). L.-C. Wu et al. / Expert Systems with Applications 36 (2009) 9007–9014 9009 3.2.2. Candidate features Several candidate features are considered to be globally related to the thermostability of proteins. We categorize these features into four categories as follows. (A) Primary structure: amino acid composition of the protein sequence. (B) Secondary structure: the amount of helix structure found within the protein and the number of atoms making up the helices. (C) Tertiary structure: Ion pairs, hydrogen bonds, disulfide bonds and accessible surface area (ASA). (D) Extended properties: Normalize the value obtained from (A), (B) and (C) above and the ratio of various other features such as polar/nonpolar. 3.2.3. Data generation Several tools were used to calculate the secondary and tertiary structural features of each protein as needed. Helix packing pair (Dalton, Michalopoulos, & Westhead, 2003) was used to calculate the helix properties of the proteins. HBPLUS v3.0 (McDonald & Thornton, 1994) was used to calculate the hydrogen bonds in the proteins using the default parameters and with the distance be- tween donor and accept atoms set at <3 Å for strong hydrogen bonds. EDPDB v03c (Matthews, 1995) was used to calculate the disulfide bonds and the accessible surface areas (ASA) of the pro- teins using the default parameters. In total, 111 features were cre- ated for each protein. 3.3. Method 3.3.1. Feature selection The total number of features was 111. Some features will con- tribute in only a minor way to the thermostability, but others will have a major effect. Clearly, not all of these candidate features are critical related to protein thermostability. Inputting the whole dataset will therefore result in the data analysis process screening a lot of noisy information and this will give rise to weak results. Therefore, we adapted the correlation coefficient method in order to carry out feature selection. The use of correlation coefficients is similar to the way they are normally used in statistical (Wacker- ly & Scheaffer, 1996) and genetic analysis (Baumgartner et al., 2004). In statistics, correlation coefficients are used to measure- ment the linear relation between two random variables. Let the two random variables, X and Y, have n pair elements each. Under these circumstances, the correlation coefficient (r) between X and Y is defined as follow: r ¼ Pn i¼1ðxi � �xÞðyi � �yÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPn i¼1ðxi � �xÞ 2Pn i¼1ðyi � �yÞ 2 q where (xi, yi) is the ith pair and �x�y is the mean of X and Y, respectively. Here we use correlation coefficients to measure the significance of features. We calculate the correlation between optimal growth temperature and each feature. After calculating all absolute values for the correlation coefficients, we extracted the features with |r| = 0.3. 3.3.2. Naïve Bayes Naïve Bayes is a data-mining and machine learning tool base on the Bayes theorem and assumes that each variable (property) is independence of each other. Compared with other machine learn- ing approach, Naïve Bayes usually shows great speed and high accuracy when analyzing a large dataset. Let c denote a class, x denote a piece of evidence available to the machine learning algo- rithm. The conditional probability P(c|x) represents the probability of class c when evidence x occurs. By Bayes’ rule: PðcjxÞ¼ pðc \ xÞ pðxÞ ¼ pðcÞpðxjcÞ pðxÞ where p(x) is the prior probability of evidence x, p(c) is the prior probability of class c, p(x|c) is the probability of occurence that the evidence x is found under class c. For each test data with an un- known class label, we can assign a class label ci which has maxi- mum P(ci|x). We used the implementation of the Naïve Bayes classifier (NBC) by Tim Menzies.3 http://www.cs.pdx.edu/~timm/dm/nbc.html Fig. 2. Distribution of data. 9010 L.-C. Wu et al. / Expert Systems with Applications 36 (2009) 9007–9014 3.3.3. Decision Tree Decision Tree is a flow-chart-like tree structure (Han and Kam- ber, 2005). The topmost node is the root node, each internal node denotes an attribute test, each branch represents an outcome of the test, and each leaf node represents classes. In order to classify an unknown dataset, the attribute values are tested against the Decision Tree. A path is traced from the root to a leaf node that holds the class prediction for the test data, Different to Naïve Bayes and Neural Network, the output model built by Decision Tree is interpretable and can easily be converted to classification rules. We use Decision Tree tool C4.54 to training our dataset. 3.3.4. Neural Network Neural Network is a machine learning approach that was in- spired by biological nervous systems. Usually, the network consists of three major layers, input layer, hidden layer, output layer. Each layer consists of several perceptions unit (nerves), and sometimes the hidden layer consists of more than one layer when solving a complex problem. We use a neural network tool (SNNS5) to design a network using the standard backpropagation algorithm, 10,000 cycles, learning rate 0.2, weight range [�1, 1] and one hidden layer .The amount of units in input layer is equal to the dimension of feature vector after feature selection. The number of units in hidden layer is double that of the input layer. 3.3.5. Model evaluation An evaluation process is needed to measure the performance of models created by machine learning approach. We used 10-fold cross-validation to evaluate the model. Thus, 1810 proteins are separated into 10 subsets randomly and then each subset is taken as test data in turn. Four performance indices, accuracy, specificity, sensitivity and precision, were used during our evaluation. Accu- 4 C4.5 (http://www2.cs.uregina.ca/~hamilton/courses/831/notes/ml/dtrees/c4.5/ tutorial.html). 5 Stuttgart Neural Network Simulator (http://www-ra.informatik.uni-tuebin- gen.de/SNNS/). racy is the percentage of all correct decisions made by classifica- tion algorithm. Sensitivity is the proportion of the data in class A that is also classified into class C. Specificity represent the propor- tion of data not in class C that are not classified into class C either. Precision is the proportion of data predicted in class C that is really in class A. For example, each protein in data set has two thermostability class labels after the K-fold process. One is the protein’s real ther- mostability class (Creal), and another is the thermostability class (Cpredict) assigned by the model built by machine learning. If we fo- cus on thermostability class mesophilic, then there is a subset la- beled REAL that is made up of proteins whose Creal = mesophilic. After the K-fold process, another subset os produced labeled PRE- DICT, which consists of protein whose Cpredict = mesophilic. Let the four subsets TP, FP, TN and FN refer to True Positive, False Po- sitive, True Negative, and False Negative, respectively, then TP is consists of proteins whose both Creal and Cpredict are mesophilic. FP is consists of proteins whose Cpredict = mesophilic, but Creal is not mesophilic. TN is consists of proteins whose neither Creal nor Cpredict are mesophilic. FN is consists of proteins whose Creal = mesophilic, but Cpredict is not mesophilic. REAL is the union of TP and FN. PREDICT is the union of FP and TP Then the four performance indices are as follows: Accuracy ¼ TP þ TN TP þ TN þ FP þ FN Sensitivity ¼ TP TP þ FN Specificivity ¼ TN TN þ FP Precision ¼ TP TP þ FP http://www2.cs.uregina.ca/~hamilton/courses/831/notes/ml/dtrees/c4.5/tutorial.html http://www2.cs.uregina.ca/~hamilton/courses/831/notes/ml/dtrees/c4.5/tutorial.html http://www-ra.informatik.uni-tuebingen.de/SNNS/ http://www-ra.informatik.uni-tuebingen.de/SNNS/ Table 1 Top 20 features with the highest correlation coefficients. Feature Correlation coefficient Glutamate (E)/residue 0.561 Charged/non-charged 0.496 Non-charged/residue 0.496 Charged/residue 0.489 Basic/residue 0.458 (E + F + M + R)/residue 0.437 Lysine (K)/residue 0.421 L.-C. Wu et al. / Expert Systems with Applications 36 (2009) 9007–9014 9011 To minimize the bias within such a random process, we re- peated the above procedure one hundred times and used the aver- age of the 100 results to measure our system. Fig. 3 show the training and predicting processes. A model was created using a machine learning approach. The model contains several interpretable or non-interpretable rules depend on the training algorithm. From a given feature set for a protein, the mod- el will calculate the thermostability class of protein following these rules. Glutamine (Q)/residue 0.404 Threonine (T)/residue 0.388 Acidic/residue 0.363 Isoleucine (I)/residue 0.350 Glutamate (E) 0.328 Serine (S)/residue 0.325 Asparagine (N)/residue 0.286 Lysine (K) 0.285 Alanine (A)/residue 0.282 Glutamine (Q) 0.280 Ion pair (GLU_ARG_24) 0.330 Ion pair (GLU_HIS_24) 0.370 Ion pair (GLU_LYS_24) 0.380 4. Results 4.1. Correlation coefficient We considered the 111 proteins features and the correlation coefficients showed that some feature only contributed in a minor way to thermostability. Table 1 shows the top 20 sequence features and their correlation coefficients. A full table of all correlation coef- ficients is given in the Appendix. For this study, we selected a correlation coefficient threshold of 0.3, which gave a subset of eleven sequence features and three structural features. The sequence features were made up of gluta- mate (E)/residue, charged/non-charged, charged/residue, basic/ residue, (E + F + M + R)/residue, lysine (K)/residue, glutamine (Q)/ residue, threonine (T)/residue, acidic/residue, isoleucine (I)/residue and serine (S)/residue. The structural features consisted of the three strong ion pair formed by glutamate/arginine, glutamate/his- tidine and glutamate/lysine. 4.2. Cross validation result 4.2.1. Three test cases For this study, we designed three test cases to examine the models built by machine learning. We constructed three datasets that divided proteins into different temperature types, MT, TH and MTH. The MT dataset divided the protein thermostability by 50 �C. Using this division, we tried to identify the difference be- Fig. 3. Training and pred tween mesophilic proteins and thermophilic proteins in the MT dataset. Thus, hyperthermophilic proteins are treated as thermo- philic proteins in the MT dataset. Similarly, the TH dataset divided the protein thermostability by 80 �C and mesophilic proteins are treated as thermophilic proteins in the TH dataset. In the MTH dataset, the data is divided into three classes by 50 �C and by 80 �C and we attempted to identify all three different thermosta- bility classes during the same process. 4.2.2. Comparing the three machine learning approaches Table 2 show the accuracy results for mesophilicity based on the eleven features. This table reveals that the Decision Tree and Neural Network approaches gave better performance than Naïve Bayes. Overall, Decision Tree gave the highest performance overall because Neural Network gave a significantly poorer performance with class TH. icting system flow. Table 2 Accuracy results based on the 11 sequence-only features. Classifier MT TH MTH Naïve Bayes 0.76 0.78 0.77 Decision Tree 0.87 0.84 0.85 Neural Network 0.88 0.75 0.81 Table 3 Accuracy results based on all 14 features. Classifier MT TH MTH Naïve Bayes 0.76 0.78 0.75 Decision Tree 0.87 0.84 0.86 Neural Network 0.85 0.82 0.74 Table 5 Detailed results for the four indices based on all 14 features using the MT dataset. Mesophilic Thermophilic AVG. (%) STD. AVG. (%) STD. Accuracy 86.8 0.005 86.8 0.005 Sensitivity 85.8 0.007 87.7 0.010 Specificity 87.7 0.010 85.8 0.007 Precision 87.5 0.009 86.1 0.006 Table 6 Detailed results for the four indices based on 11 features using the TH dataset. Thermophilic Hyperthermophilic AVG. (%) STD. AVG. (%) STD. Accuracy 84.3 0.016 84.3 0.016 Sensitivity 81.1 0.024 87.6 0.018 Specificity 87.6 0.018 81.1 0.024 Precision 86.8 0.018 82.3 0.019 Table 7 Detailed results for the four indices based on all 14 features using the TH dataset. Thermophilic Hyperthermophilic AVG. (%) STD. AVG. (%) STD. Accuracy 83.5 0.016 83.5 0.016 Sensitivity 80.5 0.021 86.5 0.019 Specificity 86.5 0.019 80.5 0.021 Precision 85.6 0.019 81.7 0.017 Table 8 Detailed results for the four indices based on 11 features using the MTH dataset. Mesophilic Thermophilic Hyperthermophilic AVG. (%) STD. AVG. (%) STD. AVG. (%) STD. Accuracy 87.0 0.0066 83.6 0.0075 89.3 0.0056 Sensitivity 87.3 0.0090 73.1 0.0144 73.2 0.0181 Specificity 86.8 0.0086 88.5 0.0084 93.2 0.0057 Precision 86.2 0.0080 75.0 0.0143 72.2 0.0171 Table 9 Detailed results for the four indices based on all 14 features for the MTH dataset. Mesophilic Thermophilic Hyperthermophilic AVG. (%) STD. AVG. (%) STD. AVG. (%) STD. Accuracy 87.7 0.006 81.2 0.007 89.6 0.006 Sensitivity 85.2 0.009 75.8 0.015 69.7 0.023 Specificity 90.0 0.007 83.7 0.010 94.3 0.006 Precision 88.9 0.007 68.9 0.013 74.5 0.020 9012 L.-C. Wu et al. / Expert Systems with Applications 36 (2009) 9007–9014 Table 3 show the accuracy result for mesophilicity based on all 14 sequence and structural features. Again, like the results shown in Table 2, Decision Tree gave the best result overall. 4.3. Decision Tree detailed results In the above section, Decision Tree gave good performance and the accuracy was generally better than 84%. In order to analyze the stability of the model, we repeat the cross-validation process 100 times and calculate the mean value and standard deviation of these 100 results. Using the MT database, we can see from Table 4 that the mean values (AVG.) and standard deviations (STD.) for the four perfor- mance indices based on 100 results using 11 features. Firstly, accu- racy, sensitivity, specificity and precision for mesophilicity and thermophilicity are all higher than 86%. Secondly, the standard deviations of all indices are 0.01 or lower. This means that our pre- diction system built by Decision Tree is both stable and gives good performance. The features were then increased to 14 and Table 5 shows the mean values and standard deviations for the four performance indices based on 100 results. Compared to Table 4, the result still show good performance, which is all better than 85%. For the TH dataset, the same format results are show in Tables 6 and 7. For 11 features system, all these indices are higher than 81% and the standard deviations are all lower than 0.024; thus the per- formance and stability are still good. For the 14 features system, all the indices are higher than 80% and the standard deviations are all lower than 0.021. The same approach was used for the MTH dataset and Table 8 shows the results for the 11 features system, while Table 9 shows the results for the 14 features system. For 11 features, only preci- sion for thermophilicity (65.0%), sensitivity for hyperthermophilic- ity (63.0%) and precision for hyperthermophilicity (69.4%) are lower than 70%. For mesophilicity, all four indices are better than 83%, while specificity for hyperthermophilicity is even higher than 93%. The standard deviations for the indices for three of the ther- mostability classes are all lower than 0.03. For 14 features, only precision in thermophilicity (68.9%) and sensitivity in hyperther- mophilicity (69.7%) are lower than 70%, and all indices have im- proved compared with 11 features. Thus, based on standard Table 4 Detailed results for the four indices based on 11 features using the MT dataset. Mesophilicity Thermophilicity AVG. (%) STD. AVG. (%) STD. Accuracy 87.3 0.005 87.3 0.005 Sensitivity 86.1 0.008 88.5 0.010 Specificity 88.5 0.010 86.1 0.008 Precision 88.2 0.009 86.4 0.007 deviation and error classification, we can say that our model shows stable performance under cross-validation. In addition to these four indices, we also carried out a special test to evaluate the model’s consistency comparing the model cre- ated by the MT dataset and that created by the TH dataset. For each protein, we use both models to predict the thermostability. Incon- sistencies between the models may occur if the MT model specifies a protein is mesophilic while the TH model specifies it is hyper- thermophilic. We calculate the percentage of proteins that show such inconsistency. Using two random models, the result should be an average of inconsistency of 25%. The output inconsistencies based on 11 features and 14 features were 1.8% and 1.6%, respec- tively. This meant that the models created by MT and TH datasets showed a low level of inconsistency. Table 10 New protein sequence dataset prediction results. Index MT TH MTH Accuracy 0.67 0.74 0.68 Sensitivity 0.56 0.82 0.54 Specificity 0.77 0.51 0.80 Precision 0.68 0.82 0.70 Appendix A All features and their correlation coefficients. Features Correlation coefficient Alanine (A) �0.117 Arginine (R) 0.171 Asparagine (N) �0.197 Aspartic acid (D) �0.126 Cysteine (C) �0.009 Glutamic acid (E) 0.328 Glutamine (Q) �0.280 Glycine (G) �0.082 Histidine (H) �0.049 Isoleucine (I) 0.177 Leucine (L) 0.101 Lysine (K) 0.285 Methionine (M) �0.042 Phenylalanine (F) �0.013 Proline (P) �0.002 Serine (S) �0.181 Threonine (T) �0.216 Tryptophan (W) �0.195 Tyrosine (Y) �0.049 Valine (V) 0.139 Length 0.014 Acidic 0.161 Basic 0.218 Polar �0.001 Non-polar 0.032 Cyclic �0.057 Acyclic 0.030 Aliphatic 0.037 Aromatic �0.075 Hydrophobic 0.032 Hydrophobic (+G) 0.014 Hydrophilic �0.001 Charged 0.193 Non-charged �0.056 Alanine (A)/residue �0.282 Arginine (R)/residue 0.246 Asparagine (N)/residue �0.286 Aspartic acid (D)/residue �0.233 Cysteine (C)/residue 0.003 Glutamic acid (E)/residue 0.561 Glutamine (Q)/residue �0.404 Glycine (G)/residue �0.183 Histidine (H)/residue �0.092 Isoleucine (I)/residue 0.350 Leucine (L)/residue 0.158 Lysine (K)/residue 0.421 Methionine (M)/residue �0.040 Phenylalanine (F)/residue �0.022 L.-C. Wu et al. / Expert Systems with Applications 36 (2009) 9007–9014 9013 4.4. Prediction results Here we applied new protein datasets to our model. First, 11415 proteins were downloaded from SWISS-Pro via PGTdb. Among these, there were 9959 proteins with an optimal growth tempera- ture lower than 50 �C, 869 proteins with one between 50 �C and 80 �C and 990 proteins with one higher than 80 �C. In order to make this test fair, we also used sampling of this data set. After sampling, the new dataset consisted of a total of 2943 proteins made up of 1377 mesophilic, 752 thermophilic and 814 hyper- thermophilic proteins. Table 10 shows the performance result. These proteins had only sequence information available and therefore the structural fea- tures aspect of the analysis could not be applied. This dataset is made up of many difference sequences, and the results showed a lower performance index than with the previously used dataset, especially that including structural information. The model based on the TH dataset showed highest sensitivity, which suggests that the TH model is better at detecting hyperthermophilic proteins. 5. Discussion and conclusions A large amount of useful thermostability data on protein was obtained from PGTdb (Huang et al., 2004) and thermostability pre- diction based on this data was carried out. Several important fea- tures need to be highlighted from this research and the comparison of the various different machine learning techniques. Important structural and sequence features are highlighted. Those features with a higher correlation coefficient absolute values are important not only to our prediction but will also be useful to pro- tein engineering. After comparing the difference methods, Decision Tree showed the best performance. Four performance indices, accuracy, sensitivity, specificity and precision were obtained for each dataset and overall high accuracy was achieved. The predic- tion results from the test data made up of 2943 protein from SWISS-Prot showed that an accuracy of better then 67% could be obtained when doing real predictions. The sequence and structure features together gave a better per- formance than sequence features alone, which confirms previous studies, which have shown that structural feature contribute to ther- mostability. The sequence and structural model increased not only the four indices of accuracy, sensitivity, specificity and precision, but also reduced their standard deviations. Thus the sequence and structural model gives higher stability than the sequence alone model. Two non-standard features were used in our research. (E + F + M + R)/residue and charged/non-charged get high linear correlation coefficients with optimal growth temperature. In Far- ias’s research (Farias et al., 2004), a special ratio (E + K)/(Q + H) was also considered a significant factor when distinguishing ther- mophilic and mesophilic proteins. In our research, charged amino acids consisted of aspartic acid, glutamic acid, lysine, arginine and histidine. Non-charged amino acids are the remaining 15 ami- no acids. Although the biological significance of these two features has not been mentioned in previous research, the features (E + F + M + R)/residue and charged/non-charged were important factors in this work and need to be consider as new key features when studying protein thermostability. Our final prediction system is based on Decision Tree. Decision Tree is an interpretable machine learning approach, which means that the output from Decision Tree consists of many explicable rules. Thus, hundreds of such rules were made available by this re- search. Each rule shows a series of conditions for determining pro- tein thermostability. Future works will be related to how to use these rules to alter protein thermostability and this has been initiated. (continued on next page) Appendix A (continued) Features Correlation coefficient Proline (P)/residue �0.027 Serine (S)/residue �0.325 Threonine (T)/residue �0.388 Tryptophan (W)/residue �0.235 Tyrosine (Y)/residue �0.121 Valine (V)/residue 0.249 Acidic/residue 0.363 Basic/residue 0.458 Polar/residue �0.087 Non-polar/residue 0.087 Cyclic/residue �0.185 Acyclic/residue 0.185 Aliphatic/residue 0.081 Aromatic/residue �0.195 Hydrophobic/residue 0.087 Hydrophobic (+G)/residue 0.004 Hydrophilic/residue �0.087 Charged/residue 0.489 Non-charged/residue �0.489 Acidic/basic �0.109 Basic/acidic 0.124 Polar/nonpolar �0.121 Nonpolar/polar 0.052 Cyclic/acyclic �0.171 Acyclic/cyclic 0.188 Charged/non-charged 0.496 Non-charged/charged �0.393 Hydrophobic/hydrophilic 0.052 Hydrophilic/hydrophobic �0.121 Hydrophobic (+G)/hydrophilic 0.012 Hydrophilic/hydrophobic (+G) �0.074 (E + F + M + R)/residue 0.437 ASA 0.029 Disulfide bond 0.017 Hydrogen bond �0.038 Strong hydrogen bond �0.019 Hydrogen bond/residue �0.034 Strong hydrogen bond/residue �0.015 Helix 0.034 Atom in helix 0.063 Helix/residue 0.050 Atom in helix/residue 0.091 Ion pair (ASP_ARG_24) 0.013 Ion pair (ASP_HIS_24) 0.025 Ion pair (ASP_LYS_24) �0.008 Ion pair (GLU_ARG_24) 0.330 Ion pair (GLU_HIS_24) 0.370 Ion pair (GLU_LYS_24) 0.380 All ion pair_24 0.224 Ion pair (ASP_ARG_46) 0.233 Ion pair (ASP_HIS_46) 0.245 Ion pair (ASP_LYS_46) �0.067 Ion pair (GLU_ARG_46) �0.201 Ion pair (GLU_HIS_46) �0.174 Ion pair (GLU_LYS_46) 0.033 All ion pair_46 0.082 Ion pair (ASP_ARG_68) 0.064 Ion pair (ASP_HIS_68) �0.042 Ion pair (ASP_LYS_68) �0.057 Ion pair (GLU_ARG_68) �0.035 Ion pair (GLU_HIS_68) 0.181 Ion pair (GLU_LYS_68) 0.194 All ion pair_68 0.180 9014 L.-C. Wu et al. / Expert Systems with Applications 36 (2009) 9007–9014 References Baumgartner, C., Bohm, C., Baumgartner, D., Marini, G., Weinberger, K., Olgemoller, B., et al. (2004). Bioinformatics, 20(17), 2985–2996. Berman, H. M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T. N., Weissig, H., et al. (2000). Nucleic Acids Research, 28(1), 235–242. Capriotti, E., Fariselli, P., & Casadio, R. (2004). Bioinformatics, 20(Suppl. 1), I63–I68. Capriotti, E., Fariselli, P., & Casadio, R. (2005). Nucleic Acids Research, 33(Web Server issue), W306–310. Chan, C. H., Liang, H. K., Hsiao, N. W., Ko, M. T., Lyu, P. C., & Hwang, J. K. (2004). Proteins, 57(4), 684–691. Dalton, J. A., Michalopoulos, I., & Westhead, D. R. (2003). Bioinformatics, 19(10), 1298–1299. Dominy, B. N., Minoux, H., & Brooks, C. L. 3rd, (2004). Proteins, 57(1), 128–141. Farias, S. T., van der Linden, M. G., Rego, T. G., Araujo, D. A., & Bonato, M. C. (2004). In Silico Biology, 4(3), 377–380. Gianese, G., Bossa, F., & Pascarella, S. (2002). Proteins, 47(2), 236–249. Gromiha, M. M., Oobatake, M., & Sarai, A. (1999). Biophysical Chemistry, 82(1), 51–67. Haney, P. J., Stees, M., & Konisky, J. (1999). Journal of Biological Chemistry, 274(40), 28453–28458. Huang, S. L., Wu, L. C., Huang, H. D., Liang, H. K., Ko, M. T., & Horng, J. T. (2004). Applied Bioinformatics, 3(1), 21–29. Huang, Shir-Ly, Wu, Li-Cheng, Huang, Hsien-Da, Liang, Han-Kuen, Ko, Ming-Tat, & Horng, J.-T. (2004). Applied Bioformatics, 3(1), 21–29. Huang, S. L., Wu, L. C., Liang, H. K., Pan, K. T., Horng, J. T., & Ko, M. T. (2004). Bioinformatics, 20(2), 276–278. Han, J., & Kamber, M. (2005). Data mining concepts and techniques. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc. Kannan, N., & Vishveshwara, S. (2000). Protein Engineering, 13(11), 753–761. Liang, H. K., Huang, C. M., Ko, M. T., & Hwang, J. K. (2005). Proteins, 59(1), 58–63. Matthews, X. Z. a. B. W. (1995). Journal of Applied Crystallography, 28, 624–630. McDonald, I. K., & Thornton, J. M. (1994). Journal of Molecular Biology, 238(5), 777–793. Parthasarathy, S., & Murthy, M. R. (2000). Protein Engineering, 13(1), 9–13. Petukhov, M., Kil, Y., Kuramitsu, S., & Lanzov, V. (1997). Proteins, 29(3), 309–320. Ragone, R. (2001). Protein Science, 10(10), 2075–2082. Saraboji, K., Gromiha, M. M., & Ponnuswamy, M. N. (2006). Biopolymers. Szilagyi, A., & Zavodszky, P. (2000). Structure with Folding and Design, 8(5), 493–504. Vieille, C., & Zeikus, G. J. (2001). Microbiology and Molecular Biology Reviews, 65(1), 1–43. Vogt, G., Woell, S., & Argos, P. (1997). Journal of Molecular Biology, 269(4), 631–643. Wackerly, D. D. W. M., III, & Scheaffer, R. L. (1996). Mathematical statistics with applications. New York: Duxbury Press. An expert system to predict protein thermostability using decision tree Introduction Related works Materials and methods System flow Materials Sampling data Candidate features Data generation Method Feature selection Naïve Bayes Decision Tree Neural Network Model evaluation Results Correlation coefficient Cross validation result Three test cases Comparing the three machine learning approaches Decision Tree detailed results Prediction results Discussion and conclusions Appendix A References 
work_eqqnsswvlng5fdijjwfahbe72q	----	NASA Technical Reports Server (NTRS) 
work_ese5a35olbgtnazmyggs7tkx3u	----	Journal of Acquired Immune Dej’icicienc~ Syadmmes and Human Retrovirology 15:356-362 0 1997 Lippincott-Raven Publishers, Philadelphia Application of an Expert System in the Management HIV-Infected Patients *Michael J. Pazzani, tDarry1 See, j-Edison Schroeder, and iJeremiah Tilles of Departments of *Information and Computer Sciences and :Medicine, University of California, Irvine, Orange, California, U.S.A Summary: A rule-based expert system, Customized Treatment Strategies for HIV (CTSHIV), which encodes information from the literature on known drug-resistant mutations was developed. Additional rules include ranking and weighting based on antiviral activities, redundant mechanisms of action, overlapping toxicities, relative levels of drug-resistance, and proportion of drug-resistant clones in the HIV quasispe- ties. Plasma was obtained from HIV-infected patients and the RNA was extracted. Segments of the HIV pol gene encoding the entire protease, reverse transcriptase, and integrase proteins were amplified by reverse transcriptase-polymerase chain reaction (using a total of three primer pairs) and cloned. Sequencing was performed on five clones from each of two patients. When the patient’s RNA sequencing data were entered into the expert program, and the information was downloaded directly into the CTSHIV program, the five most effective two, three, and four drug regimens coupled with an explanation for their choice were displayed for each patient. Thus, the CT- SHIV system couples efficient genetic sequencing with an expert program that rec- ommends regimens based on information in the current medical literature. It may serve as a useful tool in the design of clinical trials and in the management of HIV-infected patients. Key Words: AIDS-Expert program-PCR-Cloning-Sequencing- Antiretrovirals. Many compounds have been reported to inhibit repli- cation of the human immunodeficiency virus (HIV) both in vitro and in vivo (1:2). In fact, there are already 11 antiretroviral agents licensed for use in the United States, and many more are currently undergoing preclinical and clinical evaluation. The approved agents include 2’,3’ dideoxynucleotide analog, reverse transcriptase inhibi- tors (azidothymidine [AZT], didanosine, dideoxycyti- dine, starvudine [d4T], and lamivudine [3TC]); non- nucleoside reverse transcriptase inhibitors (NNRTIs; Nevirapine, Delavirdine) and protease inhibitors (Sa- quinavir, Ritonavir, Indinavir, and Nelfinavir). Multiple clinical trials have demonstrated that treatment with these agents can result in improvement in markers such as CD4 count and viral load (3-5). Other studies have Address correspondence and reprint requests to Dr. Jeremiah Tilles, Department of Medicine, University of California, Irvine, Building 53, Route 81, 101 South City Drive, Orange, CA 92868, U.S.A. Manuscript received August 7, 1996; accepted May 12, 1997. suggested that antiretroyiral therapy can improve clinical outcomes such as mortality, occurrence of opportunistic infections, and progression to AIDS (6,7). Combination therapy seems to be superior to monotherapy (8). Unfortunately, current therapeutic regimens result in suppression but not eradication of HIV. Furthermore, antiretroviral therapy is limited by both drug toxicity and the invariable development of resistance. A variety of studies have demonstrated that resistance is associated with specific mutations in the HIV pol gene (9-12). If it were possible to directly monitor the occurrence of such mutations in the HIV RNA sequence of a given patient, specific alternative therapies might be considered before a surrogate marker (e.g., CD4 count or viral load) could be expected to even reflect a failure of the current regi- men. However, detection of drug resistance may only be indicated for those patients with a rise in viral load on therapy based on studies in which continued suppression of HIV RNA was associated with a wild-type genome 356 (7). Furthermore, the clinical application of sequence in- formation in an individual patient may not currently be feasible because of the time and labor required for the sequencing process and subsequent data management. Fortunately, computer technology can be applied to solve the problem of bulky sequence data management. Rule-based expert systems (13,14) are computer pro- grams that declaratively represent knowledge of a spe- cialized problem and facts about a specific case? and draw inferences about the case. In the program devel- oped for the current study (Customized Treatment Strat- egies for HIV [CTSHIV]), the rules encode information on drug-resistant mutations of HIV and characteristics of current antiviral agents, the facts are the sequences of the HIV genomes obtained from a specific individual, and the inference to be drawn is a set of recommended drug regimens. The knowledge of a specific problem may be represented as a set of weighted if-then rules of the form: IF <antecendent> THEN <consequence>. For example, one such rule in CTSHIV is: IF Leucine is encoded by codon 4 1 of the RT portion of the pol gene, then do not use AZT (weight = 0.3). The rule is derived from the literature, which reports that a methionine-to- leucine mutation at codon 41 confers approximately fourfold resistance to AZT (15). The rules are weighted from 0.1 (low priority) to 1.0 (high priority) based on level of resistance or support in the literature. To draw an inference about a particular conclusion (e.g., which drugs should not be used), the expert system finds rules whose consequence matches that conclusion, and attempts to determine whether the antecedent of the rule is true. If the antecedent can be shown to be true, then the conclu- sion is asserted to be true. There are two ways in which an antecedent may be true. First, the antecedent may match a fact that is known. For example, in the rule just stated, this would check to see whether there is a leucine encoded by codon 41 of the RT portion of the pol gene. Second, the antecedent may match the consequence of another rule, and the antecedent of that rule can then be shown to be true. The current study reports a process by which the pro- tease, reverse transcriptase, and integrase segments of the HIV pol gene can be cloned using three primer sets and the sequences downloaded into an expert system that recommends multiple ranked two-, three-, and four-drug regimens based primarily on the occurrence of known mutations coding for specific drug resistance. METHODS Cloning and Sequencing Blood was obtained from 10 HIV-infected individuaIs after informed consent was obtained using a form approved by the Institutional Re- view Board at UC Irvine, All patients had CD4 counts <500/mm3, two had never received antiretroviral therapy, and the others had received a variety of regimens, Standard reverse transcription-polymerase chain reaction (RT-PCR) was performed using published primer sets for pro- tease (16) and RT (17) and primers developed in our laboratory for integrase (upstream 5’CAAGTAGATAAATTAG TCAGTGCTG- GAATCAGG-3’: downstream 5’.CCTAGCTTTCCCTGAAACATA- CATATGGTG’IT3’). PCR was performed on all 10 patients to dem- onstrate reproducibility of the primer sets used. Because cloning and sequencing of PCR products is well standard- ized, PCR product from only two patients was sequenced. Briefly. a portion of PCR product (calculated to give an approximate vector insert ration of 1: 1) was Iigated into a T/A vector (Invitrogen). Two micro- liters of the ligation reaction was transferred to a tube containing 50 p,l of competent Eschen’chia coli cells. The bacteria were transformed (18), plated onto Luria-Bertani (LB) agar plates containing 50 pg/ml of kanamycin. and grown at 37°C for 72 h. For each plate, a single white colony was grown overnight in 5 ml of liquid LB broth containing 50 kg/ml of kanamycin. The plasmid DNA was purified and extracted using a Qiagen plasmid preparation kit (Chatsworth, CA). For the two patients, five clones per PCR product were sequenced using a dye- termination process (19) with an automated ABI sequencer. One was asymptomatic and had a CD4 count of 285/mm3 without prior antiret- roviral therapy. The other patient had a history of Pneumucvstis carinii pneumonia and mild wasting syndrome, had received multiple antiret- roviral regimens, and currently had a CD4 count of 55 cells/mm’. Development of Rules for the Expert Program Base substitutions coding for either in vitro and/or resistance to specific drugs (20-24) were identified in the literature and included as rules for the expert system. The rules were weighted based on the degree of resistance. Additional rules ranked available therapies based on HIV antiviral activity, overlapping toxicities, and mechanisms of action. Currently, there are 48 rules in the program (Table 1). Development of the Expert Program The CTSHIV system is implemented in first-order computer learning (FOCL-1-2-3) (14), a backward chaining expert system for the Mac- intosh computer. The facts needed to determine which drugs to exclude are downloaded in CTSHIV. The recommendation process goes through six distinct phases: 1. Exclusion of antiretroviral agents based on sequence information. The medical literature contains an increasing number of studies on the relationship between mutations and drug resistance. 2. Inclusion of drugs with enhanced antiretroviral activity associated with specific genetic sequences (25). 3. Exclusion of antiviral combinations with overlapping toxicities, redundant mechanisms of action, or unproven efficacy and/or safety (e.g., NNRTIs used in combination with protease inhibi- tors). 4. Weighting the aforementioned combinations from 0.1 (low pri- ority) to 1.0 (high priority) based on either: a. level of resistance conferred (phase 1; 2-fold or >lOO-fold resistance were assigned a value of 0.1 or 1.0, respectively; intermediate resistance was assigned a value between 0.1 and 1.0 [22,23]), or b. strength of supporting evidence in the literature (phases 2-3 125,261). EXPERT SYSTEM IN MANAGEMENT OF HIV-INFECTED PATIENTS 357 JournaI of Acquired Immune Dejkienc)- Syndromes and Human Retroviroiog~, Vol. IS, No. 5, 1997 3.58 M. J. PAZZANI ET AL. TABLE 1. Complete set qf current rules qf the Customized Treatment Strategies for HIV expert program 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. AZT codon 41 RT (CTT CTC TTA CTA CTG TTG) [0.3] (15) AZT codon 41 RT (CM CTC TTA CTA CTG TTG) and codon 215 RT (TTC TAC) [l.O] (31) AZT codon 67 RT (AAT AAC) and codon 70 RT (AGA AGG CGG CGC) and codon 215 RT (TTC TAC) and codon 219 RT (CAA CAG) [0.8] (32) AZT codon 41 RT (CTT CTC TTA CTA CTG TTG) and codon 67 RT (AAT AAC) and codon 70 RT (AGA AGG CGG CGC) and codon 215 RT (TTC TAC) [I.01 (32) AZT codon 67 RT (AAT AAC) IO.61 (33) AZT codon 70 RT (AGA AGG CGG CCC) [0.6] (33) AZT codon 215 RT (TTC TTT) [0.7] (33) AZT codon 219 RT (CAA CAG GAA GAG) [O.l] (33) ddl codon 74 RT (GTA GTG GTT) [0.4] (20) (ddl ddC) codon 184 RT (GTA GTG GTT) [0.2] (34) (AZT ddl d4T ddC) codon 151 RT (ATG) [l.O] (35) AZT codon 210 RT (TGC) and codon 215 (TTC ‘ITT) [0.9] (36) (AZT ddl d4T ddC) codon 69 RT (GGA GGT GGG) [0.6] (37) 3TC codon 184 RT (GTA GTG GTT) [l.O] (34) 3TC codon 184 RT (ATA ATC ATT) [0.6] (34) (Nevirapine Delavirdine) codon 103 RT (AAT AAC) [0.5] (38) Nevirapine codon 106 RT (GCA GCT GCC) [0.9] (39) Nevirapine codon 108 RT (ATA ATC ATT) [0.5] (40) (Nevirapine Delavirdine) codon 181 RT (TGT TGC) [0.9] (41) Nevirapine codon 181 RT (ATA ATC ATT) LO.71 (42) Nevirapine codon 188 RT (TGT TGC) [0.5] (40) Nevirapine codon 190 RT (GCA GCT CCC) [OS] (39) Delavirdine codon 236 RT (CM CTC TTA CTA CTG TTG) W.61 (43) Delavirdine codon 69 RT (GGA GGT GGG) [0.2] (37) Nevirauine codon 69 RT (CGA GGT GGG) 10.41 (37) Ritonaiir codon 8 Pro (CAA CAG AAA AkG) [0.4]‘(44) Ritonavir codon 32 Pro (ATA ATC ATT) [0.3] (45) (Ritonavir Saquinavir) codon 48 Pro (GTA GTG GTT) [0.6] (46) (Ritonavir Indinavir) codon 82 Pro (GCA GCT GCC TTC TTT) LO.61 (47) 30. Ritonavir codon 10 Pro (ATA ATC ATT) [0.4] (47) 31. Ritonavir codon 54 Pro (GTA GTG GTT) [0.4] (47) 32. 33. 34. 35. 36. 37. 38. 39. 40. 41. 42. 43. 44. 45. 46 47 48. Indinavir codon 46 Pro (ATA ATC ATT CTT CTC TTA CTA CTG TTG) [0.6] (48) Indinavir codon 82 Pro (ACA ACT ACC) [0.6] (48) (Ritonavir Indinavir Saquinavir) codon 84 Pro (GTA GTG GTT) 10.61 (47) Indinavir codon 32 Pro (ATA ATC ATT) [0.2] (48) (Indinavir Ritonavir Nelfinavir) codon 71 Pro (GTA GTG GTT) [O.l] (47,49) Saquinavir codon 90 Pro (ATG) [0.6] (50) Saquinavir codon 48 Pro (GTA GTG GTT) and codon 90 Pro (ATG) [l.O] (50) Neltinavir codon 30 Pro (AAT AAC) [0.8] (49) USE Nevirapine if codon 236 RT (CTT CTC TTA CTA CTG TTG) [0.5] (43) Do not use Nelfinavir with Indinavir, Ritonavir, or Saquinavir t1.01 (51) Do not use Indinavir with Ritonavir or Saquinavir [l .O] (51) Do not use both Delavirdine and Nevirapine together [0.9] (8) Do not use both ddc and ddl together [ 1.01 (51) Do not use either Nelfinavir, Indinavir, Ritonavir, or Saquinavir with either Nevirapine or Delavirdine [l.O] (51) The expert program will receive five sequences from each patient. If all five cause a rule to fire, then multiply the weight of rule by 1.0; if four of five cause the rule to fire, multiply the weight by 0.8; if three of five cause the rule to fire, multiply the weight by 0.6; if two of five cause the rule to fire, multiply the weight by 0.4; and if one of five cause the rule to fire, multiply the weight by 0.2. Display five two-drug regimens, five three-drug regimens. and five four-drug regimens. If more regimens are permissible, delete the following drugs in order of preference: ddC, ddl Nevirapine, d4T, Saquinavir, Delavirdine, AZT, 3TC, and Nelfinavir. When more than one two, three, or four drug regimens are displayed, rank each regimen by providing a value for each drug within the regimen based on the sum of the weights accumulated from the rules. Rank in descending order the regimens with the lowest total value. In case of equal value: rank in the following order: Indinavir, Ritonavir. Nelfinavir. 3TC, AZT, Delavirdine, Saquinavir, d4T, Nevirapine, ddl. ddC. The rules were developed from information in the medical literature and integrated into the FOCL expert system shell. Most of the rules are exclusionary, based on drug-resistant mutations. For example, rule 1 is interpreted as, “Do not use AZT if the nucleotides at codon 41 of the reverse transcriptase portion of the polymerase gene are ‘CCT CTC TTA CTA CTG or TTG.’ ” The parentheses containing the names of drugs or codons represent an “or” between the elements within them. Weights are in brackets, scaled from 0.1 (lowest priority) to 1.0 (highest priority), Exclusionary rules are weighted by the degree of drug resistance conferred. Drug interaction rules are weighted by the severity of overlapping toxicities. Disallowed drug combinations are weighted by the relative lack of supporting evidence in the literature for their use. Each weighted rule is referenced. AZT, azidothymidine: ddl, didanosine; ddC, dideoxycytidine: d4T, starvudine; 3TC, lamivudine. 5. Adjusting the weight of each conclusion by the proportion of clones causing the rule to fire. 6. Identifying and ranking combinations of candidate drugs based on antiviral activity. RESULTS Recommended regimens for the study patients: HIV nucleic acid was successfully amplified from all 10 pa- tients using each of the three primer pairs. Complete sequences of five clones from two patients were down- loaded into the CTSHIV program and selected regimens choice @a&e 2). Ali five clones frdm patient 1 (CD4 were disulaved in coniunction with exDlanations for their count of 285/mm’, no prior antiretrovirals) had identical sequences, including 14 and 6 amino acid substitutions for RT and Pro, respectively, compared with HXB2 (Fig. 1). None of the changes resulted in the firing of a rule. In contrast, there was great heterogeneity among the five clones from patient 2 (CD4 count of 55/mm3, multiple antiretroviral regimens in the past). There was an average of 47 and 22 amino acid changes for RT and Pro, re- spectively, firing eight exclusionary rules in the expert program. EXPERT SYSTEM IN MANAGEMENT OF HIV-INFECTED PATIENTS TABLE 2. Protocols recommended for patients I and 2 Protocols recommended for patient 1 The following protocols with tu‘o drugs are recommended: Indinavir 3TC Indinavir AZT Ritonavir 3TC Ritonavir AZT 3TC AZT The following protocols with three drugs are recommended: Indinavir 3TC AZT Ritonavir 3TC AZ7 Ritonavir 3TC Saquinavir Ritonavir AZT Saquinavir 3TC AZT Delavirdine The following protocols with four drugs are recommended: Indinavir 3TC AZT D4T Rotonavir 3TC AZT Saquinavir Ritonavir 3TC AZT D4T Ritonavir 3TC Saquinavir D4T Ritonavir AZT Saquinavir D4T Protocols recommended for patient 2 The following protocols with two drugs are recommended: 3TC Delavirdine 3TC d4T 3TC Nevirapine Delavirdine d4T 3TC ddC The following protocols with three drugs are recommended: 3TC Delavirdine d4T 3TC d4T N&rapine 3TC Delavirdine ddC 3TC d4T ddC 3TC N&rapine ddC The following protocols with four drugs are recommended: 3TC Delavirdine d4T ddC 3TC d4T Nevirapine ddC 3TC Delavirdine d4T ddl 3TC d4T Nevirapine ddl 3TC d4T ddC Indinavir 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.4 0.6 By rule I, AZT reduced by 0.12 because two sequences had TTG at codon 41 of the RT gene. Kellam P, Boucher C, Larder BA. Fifth mutation in human immunodeficiency virus type 1 reverse transcriptase contributes to the development of high-level resistance to zidovudine. Proc Nail Acad Sci U.SA 1992:89: 1934-8. By rule 28, Ritonavir reduced by 0.24 because two sequences had GTG at 0.6 codon 48 of the Pro gene. Boucher C. Rational approaches to resistance: Using 5aquinavir. AIDS 1996;1O(suppl l):S15-9. By rule 5, AZT reduced by 0.6 because five sequences had AAC at codon 67 of the RT gene. Cal&do A, Savara A, An D, DeVore K, Kaplan J, D’Aquila R. Effects of zidovudine-selected human immunodeficiency virus type 1 reverse transcriptase amino acid substitutions on processive DNA synthesis and viral replication. J Vi& 1996:70:2146-53. By rule 34, Indinakir, Saquinavir, Ritonavir reduced by 0.6 because five sequences had GTG at codon 84 of the Pro gene. Schmidt J, Ruiz L, Clotet B, et al. Resistance-related mutations in the HIV-1 protease gene of patients treated for I year with the protease inhibitor ritonavir (ABT-538). AIDS 1996:10:995-9. By rule 38, Saquinavir reduced by 0.4 because two sequences had GTG at codon 48 of the Pro gene and ATG at codon 90 of the Pro gene. Jacobsen H, Haenggi M, Ott M, Duncan I, Andreoni M, Vella S, Mous 1. Reduced sensitivity TO saquinavir: An update on genotyping from phase III1 trials. Anrivirai Res 1996;29:95-7. By rule 9, ddl reduced by 0.4 because five sequences had GTA at codon 74 of the RT gene. St Clair M, Martin J, Tudor W. Resistance to ddl and sensitivity to AZT induced by a mutation in HIV-1 reverse transcriptase. Science 1991;253:1557-9. By rule 6, AZT reduced by 0.6 because five sequences had AGA at codon 70 of the RT gene. Caliendo A, Savara A, An D, DeVore K, Kaplan J, D’Aquila R. Effects of zidovudine-selected human immunodeficiency virus type 1 reverse transcriptase amino acid substitutions on processive DNA synthesis and viral replication. J Viral 1996;70:2146-53. By rule 37. Saquinavir reduced by 0.36 because three sequences had ATG at codon 90 of the Pro gene. Jacobsen H, Haenggi M, Ott M, Duncan I, Andreom M, Vella S, Mous J. Reduced wnsitivity to saquinavir: An update on genotyping from phase I/I1 trials. Antiviral Rer 1996;29:95%7. DISCUSSION Antiretroviral therapy is useful in slowing down the progression of HIV infection. However, all therapeutic regimens studied to date eventually fail, resulting in clinical deterioration of the patient and worsening of sur- rogate markers such as a rise in CD4 count, an increase in viral load, or both. The current standard of practice indicates a change in therapy when this occurs. However, guidelines as to specific changes in antiretrovirals is lacking, and the possibility of appropriately changing therapy for specific mutants before surrogate marker de- terioration is possible. The time to onset of clinically important mutations is highly variable. Cross-resistance occurs between agents in the same class or with similar mechanisms of action (21). Finally, drug-specific resis- tance mutations also occur in wild-type strains (22). Thus, basing changes in therapy on drug sensitivity would seem to have merit. The results of the current study demonstrate the feasibility of such an approach using an expert program to recommend treatment regi- mens. The program also encodes information about level of drug resistance, overlapping toxicities, and other fac- tors. Learning algorithms such as FOCL (14) can be applied to the program to modify it based on outcome measures from patients such as CD4 count, viral load, and progression to AIDS. Old rules can be reweighted or eliminated. New rules can be added based on information from the literature or the development of new antiretro- viral agents. The HIV strains in an individual patient are not ge- nomically homogenous. The current system allows de- tection of a resistant mutation if present in at least one of five clones. If detection of crucial mutations in a patient found less frequently than 20% is later thought to be important, hybridization techniques (27) could be ap- plied to capture those present in as few as 1 of 20 strains. The CTSHIV system uses primers from conserved re- gions, yielding detectable PCR products for each of the 10 patients tested. Furthermore, it contains an expert sys- tem to effectively manage the sequence data and recom- mend specific therapeutic regimens based on genetic se- quences and information from the medical literature. Nucleotide sequences could be aligned using software programs (28), reducing the chance of downloading in- determinate foci. However, the current system incorpo- Journal of Acquired Immune Deficiency Syndromes and Human Retrovimlogy, Vol. 15, No. 5, 1997 360 M. J. PAZZANI ET AL. position 1 11 21 31 41 51 61 71 HXB2 CCTCAGGTCA CTCTTTGGCA ACGACCCCTC GTCACAATAA AGATAGGGGG GCAACTAAAG GAAGCTCTAT TAGATACAGG Patient 1 clone1 ---- __A___ -___ -___ __ _____ __._. -- _G______ ___.______ __________ __________ __________ clone2 ______A___ ____ ____._ ____ ______ -_ _G______ ___.______ __________ __________ __________ clone3 ---- __A___ __________ _______.__ ___G___-__ _.________ ____.__-__ __________ -_________ clone4 ---- __A___ ____.___-_ __________ ---G--_--u _._.___.__ __________ __________ _______-__ clone6 ______A.__ ____ ______ _______.__ __ _G______ __________ __________ __________ __________ Patient 2 clone1 ---- __A___ ____C_____ _______G__ ___T______ G-C_______ ______G___ ____________________ clone2 ---- __A___ ____ C _____ _______G__ ___T______ G_C____A__ ______G___ _________________-__ clone3 ---- __A___ ____C_____ __________ ___T___.__ __C____A__ ____._G___ ____________________ clone4 ---- __A___ ____C_____ _______G__ ___T______ G_C____A__ ______G___ ____________________ clone5 ---- __A___ _-__C__--_ __________ ___T.__.__ __C____A__ ____._G___ ____________________ 81 91 101 111 121 131 141 151 HXB2 AGCAGATGAT ACAGTATTAG AAGAAATGAG TTTGCCAGGA AGATGGAAAC CAAAAATGAT AGGGGGAATT GGAGGTlTTA Patient 1 clone, __________ ------____ __________ -__-___.__ _A________ -._-_-_--. ___.______ ___--__--- clone2 ___.______ -----_---. ___.__.__. ---------_ _A________ __________ __________ __________ clone3 ___.______ ___.______ ___.______ -__-___-._ _A__________________ __________ __________ cbne4 __._______ _____.____ ___._.____ _______.__ _A________ _______-__ __________ -__-__---_ clone5 __________ ____-_____ __________ ___-__..__ _A________ ___-___.__ __________ __________ Patient 2 clone, _ _ _ _ _ _ _ _ _ _ - _ - _ - - - - - - T_______G_ __________ __________ __________ __________ --.--.---- c,,3ne2 __________ .___._____ T_______G_ __________ __________ __________ __________ __________ clone3 __________ __________ _.________ ____________________ __________ __T_______ __________ clone4 ___ .._____ ____-__--_ T_______G_ __________ __________ ______.___ __________ __________ clone5 __..______ __________ __________ ______.___ __________ __________ __T_______ _.________ 161 171 161 191 201 211 221 231 HXB2 TCAAAGTAAG ACAGTATGAT CAGATACTCA TAGAAATCTG TGGACATAAA GCTATAGGTA CAGTATTAGT AGGACCTACA Patient 1 clone1 _________.__________ ____-_____ ____.._-_. ---- __C___ ____,.__I^ __A__ ____ _“_I__.___ clone2 ____________________ __________ ____-__-______ __C___ _________. __A_______ __________ clone3 ____________________ ___--..-.- --------__ .--_ __C___ __________ __A-______ __________ clone4 _________.__________ __________ -______-__ .--- __C___ ___________.A_______ __________ clone6 ___ ._._ __- __________ ___-______ ____--_-__ ---- __C___ ____._____ __A_______ ___.______ Patient 2 clone, __________CA________ __________ ___ __T___A__________ -A___TC__C__________ __________ ,3,3-,e2 __________CA________ _-_____-_- _C___T___,+__________ _A__.TC__C __________ __________ clone3 __________CA________ -___-__-_- _-__ _T____ __________ _A____C___ _______.__ __________ clone4 __________ CA________ __________ _____T___A __________ _,t,___TC_-C __________ __________ c,on*5 __________ CA________ __________ _____T____ __________ _A____C___ __________ __________ 241 251 261 271 261 291 HXBP CCTGTCAACA TAATTGGAAG AAATCTGTTG ACTCAGATTG GTTGCACTTT AAATTTT Patient 1 c,o,,e, __________ _.____________.__.__ __________ _G________ _______ clone2 __________ ____________________ __________ _G________ _______ clone3 __________ ____________________ __________ _G________ _______ clone4 __________ ______________-_____ ._________ _G________ _______ clone5 __________ __________-___-_____ __________ _G________ _______ Patient 2 clone, _________ =_________________A__ ________G_ ________C_ _______ clone2 _____ ___. G____________________ _______.G_ ________C_ _______ .-lone3 _________G____________________ ________G_ ________C_ ____.__ clone4 _________o_______ ______ _-- _A__ ________G_ ________C_ _______ clone5 _._ ______ 0 ._.._____________A__ ________G_ ________C_ _______ FIG. 1. Complete sequence of the protease portion of the human immunodeficiency virus polymerase gene (297 bp) for each of five clones from two patients. Patient 1 has never received antiretroviral therapy, and has a current CD4 count of 285/mm3. Patient 2 has received multiple combinations of azidothymidine, didanosine, lamivudine, and Saquinavir. The sequences are compared with that of strain HXBP (Los Alamos data bank). The sequences that cause a rule to fire are in bold type. J~~mal of Acquired Immune Deficiency Syndromes and Human Retrovirology, Vol. 15, No. 5, 19%’ EXPERT SYSTEM IN MANAGEMENT OF HIV-INFECTED PATIENTS 361 rates sequencing of multiple clones. Although our pre- liminary studies suggested that such a process both re- duces the display of indeterminate residues and enhances the opportunity to identify drug-resistant strains repre- senting a small proportion of the total number of quasi- species in an individual patient compared with sequenc- ing the heterogenous mixture of strains, it also adds sig- nificantly to the time and cost. In this pilot study, every effort was made to yield data as cleanly as possible. However, if the system were to be used in the everyday management of patients, sequencing of uncloned, ampli- fied products would be more practical and may yield comparable results. Specialty Laboratories (Santa Monica, CA, U.S.A.) is currently offering genotyping of heterogenous PCR product commercially. Alternatively, rapid sequencing of the polymerase and protease genes in development by Affymetrics (Sunnyvale, CA, U.S.A.) and Applied Biosystems (Foster City, CA, U.S.A.) using fluorescent primer-based cycle sequencing (29) may ren- der the use of the CTSHIV expert system more feasible. The CTSHIV system is limited by an incomplete un- derstanding of the correlation between genomic muta- tions conferring drug resistance and clinical and surro- gate marker outcomes. Furthermore, the expert program can only incorporate rules from the current literature and recommend licensed drugs. A study to correlate surro- gate marker response and clinical outcomes with the use of the program is in progress. In addition, a study by the CPCRA is under way to assess the clinical utility of genotyping (Volberding P, personal communication, February 1997). Studies with large numbers of patients will ultimately be required to validate the CTSHIV sys- tem and determine whether it may assist in the everyday management of HIV-infected individuals and in the de- sign, implementation, and interpretation of outcomes in clinical trials (30). Currently, however, there is no evi- dence that the system will perform better than clinical judgment. Acknowledgments: The technical assistance of Tonya Clark is appreciated. The authors thank Douglas Richman, M.D. for reviewing and revising the expert rules. The sequencing was performed by Erik Avajani and Clifford Brunk. This work was supported by the California Universitywide AIDS Research Program through the California Collaborative Treatment Group (CCTG). REFERENCES 1. DeClercq E. Antiviral therapy for human immunodeficiency virus infections. Clin Microbial Rev 1995;8:200-39. 2. Frost S, McLean A. Quasispecies dynamics and the emergence of drug resistance during zidovudine therapy of HIV infection. AIDS 1994;8:323-32. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13 14 15. 16. 17. 18. 19. 20. 21. 22. Katlama C, Ingrand D, Loveday C, et al. Safety and efficacy of lamivudine-zidovudine combination therapy in antiretroviral-naive patients. MMA 1996;276: 118-25. Holodniy M, Katzenstein D, Winters M, et al. Measurement of HIV virus load and genotypic resistance by gene amplification in asymptomatic subjects treated with combination therapy. / Acquir Immune Defic Syndr 1993;6:366-9. Kahn J, Lagakos S, Richman D, et al. A controlled trial comparing zidovudine with didanosine in human immunodeficiency virus in- fection: The NIAID AIDS Clinical Trials Group. N Engl J Med 1992;327:581-7. Schooley RT. Ramirez-Ronda C, Lange JM, et al. Virologic and immunologic benefits of initial combination therapy with zidovu- dine and zalcitabine or didanosine compared with zidovudine monotherapy: Wellcome Resistance Study Collaborative Group. J infect Dis 1996;173:1354-66. Zhu QY, Scarborough A, Polsky B: Chou TC. Drug combinations and effect parameters of zidovudine, stavudine, and nevirapine in standardized drug-sensitive and resistant HIV type 1 strains. AIDS Res Hum Retroviruses 1996;12:507-17. Lange JM. Triple combinations: Present and future. J Acquir Zm- mw-ze Defic Syndr Hum Retrovirol 1995;1O(suppl l):S77-82. Nielsen C, Bruun L, Mathiesen L. Pedersen C, Gerstoft J. Devel- opment of resist;mce to zidovudine (ZDV) and didanosine (ddl) in HIV from patients in ZDV, ddl and alternating ZDV/ddl therapy. AIDS 1996;10:625-33. Larder B. Kellam P, Kemp S. Zidovudine resistance predicted by direct detection of mutations in DNA from HIV-infected lympho- cytes. AIDS 1991;5:13744. Larder B, Kemp S. Multiple mutations in HIV-l reverse transcrip- tase confer high-level resistance to zidovudine (AZT). Science 1989;246:1155-8. Frenkel L. Wagner M, Atwood S, Cummins T, Dewhurst S. Spe- cific, sensitive, and rapid assay for HIV pol mutations associated with resistance to zidovudine and didanosine. J Infect Dis 1995; 33:342-7. Jackson P. Introduction to expert systems. Reading, MA: Addison- Wesley, 1990. Pazzani M, Kibler D. The utility of knowledge in inductive leam- ing. Machine Leatnlizg 1992;9:57-94. Kellam P, Boucher C, Larder BA. Fifth mutation in human immu- nodeficiency virus type 1 reverse transcriptase contributes to the development of high-level resistance to zidovudine. Proc Nat1 Acad Sci USA 1992;89:1934-8. Fontenant G, Johnston K, Cohen J, Gallaher W, Robinson J, Luftig R. PCR amplification of HIV-l proteinase sequences directly from lab isolates allows determination of five conserved domains. Viral 1992;190:1-10. Gu Z, Gao Q, Fang H, et al. Identification of a mutation at codon 65 in the IKKK motif of reverse transcriptase that encodes human immunodeticiency virus resistance to 2’,3’-dideoxycytidine and 2’,3’-dideoxy-3’.thiacytidine. Antimicrob Agents Chemother 1994;38:275-81. Clark J. Novel non-templated nucleotide addition reactions cata- lyzed by procaryotic and eucaryotic DNA polymerases. Nucleic Acid Res 1988;20:9677-86. Giesecke H, Obermaier B, Domdey H. Rapid sequencing of the Sendai virus 6.8 kb large (L) gene through primer walking with an automated DNA sequencer. J Viral Methods 1992;38:47-60. St Clair M, Martin J, Tudor W. Resistance to ddl and sensitivity to AZT induced by a mutation in HIV- 1 reverse transcriptase. Science 1991;253:1557-9. Gu Z, Gao Q, Li X, Parniak M, Wainberg M. Novel mutdon in the human immunodeficiency virus type 1 reverse transcriptase gene that encodes cross-resistance to 2’,3’-dideoxyinosine and 2’,3’- dideoxycytidine. J Viral 1992;66:7128-35. Gu Z, Gao Q, Fang H, Pamiak M, Brenner B, Wainberg M. Iden- tification of novel mutations that confer drug resistance in the Jounud ofAcquired hmune Deficiency Syndromes and Human Retrovirology, Vol. 15, IVO. 5, 1997 M. J. PAZZANI ET AL. human immunodeficiency virus polymerase gene. Leukemia 1994; 8S1:5166-9. 23. Richman D, Shih C, Lowy 1: Rose J, Prodanovich P, Goff S, Griffin J. Human immunodeficiency virus type 1 mutants resistant to nonnucleoside inhibitors of reverse transcriptase arise in tissue culture. Proc Nat1 Acad Sci USA 1991;88:11241-5. 24. El-Farrash M, Kuroda M, Kitazaki T, Masuda T. Kato K, Hatanaka M, Harada S. Generation and characterization of a human immu- nodeficiency virus type 1 (HIV-l) mutant resistant to an HIV-l protease inhibitor. J Viro[ 1994;68:23339. 25. Dueweke T, Pushkarskaya T, Poppe S, et al. A mutation in reverse transcriptase of bis(heteroaryl)piperazine-resistant human immu- nodeficiency virus type 1 that confers increased sensitivity to other non-nucleoside inhibitors. Proc Nutl Acad Sci USA 1993;90: 4713-7. 26. Hammer S, Kessler H, Saag M. Issues in combination antiretrovi- ral therapy: a review. J Acquir Immune Defic Syndr 1994;7 Suppl 2:S2435. 27. Holodniy M, Mole L, Margolis D, et al. Determination of human immunodeficiency virus RNA in plasma and cellular viral DNA genotypic zidovudine resistance and viral load during zidovudine- didanosine combination therapy. J Viral 1995;69:3510-6. 28. Michael N, Davis KE, Loomis-Price L, VanCott T, Burke D, Red- field R, Birx D. V3 seroreactivity and sequence variation: Tracking the emergence of V3 genotypic variation in HIV-l-infected pa- tients. AIDS 1996;10:121L9. 29. Luehrsen KR, Marr L, van der Knaap E, Cumberledge S. Analysis of differential display RT-PCR products using fluorescent primers and GENESCAN software. Biotechaiques 1997;22:168-74. 30. Ohno-Machado L, Parra E, Henry S, Tu S, Musen M. AIDS 2: A decision-support tool for decreasing physician’s uncertainty re- garding patient eligibility for HIV treatment protocols. Proceed- ings of the 17th Annual Symposium on Computer Applications in Medical Care, Washington, DC, 1993:429-33. 3 1. Boucher C, O’Sullivan E, Mulder A, et al. Ordered appearance of zidovudine resistance mutations during treatment of 18 human immunodeflciency virus-positive patients. J Infect Dis 1992;165: 105-10. 32. Gurusinghe A, Land S, Birch C, et al. Reverse transcriptase mu- tations in sequential HIV-l isolates in a patient with AIDS. J Med Virol 1995;46:23843. 33. Caliendo A: Savara A, An D, DeVore K, Kaplan J, D’Aquila R. Effects of zidovudine-selected human immunodeficiency virus type 1 reverse transcriptase amino acid substitutions on processive DNA synthesis and viral replication. J Viral 1996;70:2146-53. 34. Quan Y, Gu Z, Li X, Li Z, Morrow C, Wainberg M. Endogenous reverse transcription assays reveal high-level resistance to the tri- phosphate of (-)2’-dideoxy-3’-thiacytidine by mutated M184V hu- man immunodeficiency virus type 1. J Viral 1996;70:5642-5. 35. Iversen AK, Shafer RW, Wearly K, Winters MA, Mullins JI, Che- sebro B, Morgan TC. Multidrug-resistant human immunodeficien- cy virus type 1 strains resulting from combination antiretroviral therapy. J Viral 1996;70: 1086-90. 36. Hooker D, Tachedjian G, Soloman A, et al. An in viva mutation from leucine to tryptophan at position 210 in human immunode- 37. 38. 39. 40. 41. 42. 43. 44. 45. 46. 47. 48. 49. 50. 51. ficiency virus type 1 reverse tmnscriptase contributes to high-level resistance to 3’.azido-3’-deoxythymidine. J Vi& 1996;70:8010-8. Kew Y, Saloman H, Olsen L, Wainberg M, Prasad V. The nucleo- side analog-resistant E89G mutant of human immunodeficiency virus type 1 reverse transcriptse displays a broader cross-resistance that extends to nonnucleoside inhibitors. Anfimicrob Agents Che- mother 1996;40:17114. Zhu Q, Scarborough A, Polsky B, Chou T. Drug combinations and effect parameters of zidovudine, stavudine, and nevirapine in stan- dardized drug-sensitive and resistant HIV type 1 strains. AIDS Res Hum Retroviruses 1996;12:507-17. Richman D, Havlir D, Corbeil J, et al. Nevirapine resistance mu- tations of human immunodeficiency virus type 1 selected during therapy. / Viral 1994;68:1660-6. Richman DD. Resistance of clinical isolates of human immunode- ficiency virus to antiretroviral agents. Antimicrob Agents Che- mother 1993;37:1207713. Richman D. Shih CK, Lowy I, Rose J. Prodanovich P, Goff S, Griffin J. Human immunodeflciency virus type 1 mutants resistant to nonnucleoside inhibitors of reverse transcriptase arise in tissue culture. Proc Nat1 Acad Sci USA 1991;88:11241-5. Havlir, Eastman S, Gamst A, Richman D. Nevirapine-resistant human immnodeficiency virus: Kinetics of replication and esti- mated prevalence in untreated patients. J Viral 1996;70:7894-9. Dueweke TJ, Pushkarskaya T, Poppe SM. et al. A mutation in reverse transcriptase of bis(heteroary1) piperazine-resistant human immunodeticiency virus type 1 that confers increased sensitivity to other non-nucleoside inhibitors. Proc Nat1 Acad Sci USA 1993;90: 4713-7. Ho DD, Toyoshima T, MO H, et al. Characterization of human immunodeficiency virus type 1 variants with increased resistance to a CZ-symmetic protease inhibitor. J Viral 1994;68:201&2020. Kaplan AH, Michael SF, Wehbie RS, et al. Selection of multiple human immunodeficiency virus type 1 variants that encode viral proteases with decreased sensitivity to an inhibitor of viral prote- ase. Proc Nat1 Acad Sci USA 1994;91:5597-601. Boucher C. Rational approaches to resistance using saquinavir. AIDS 1996;1O(suppl l):S15-9. Schmidt J, Ruiz L, Clotet B, et al. Resistance-related mutations in the HIV-l protease gene of patients treated for 1 year with the protease inhibitor ritonavir (ABT-538). AIDS 1996;10:995-9. Condra _I, Holder D, Schlief W, et al. Genetic correlates of in vivo viral resistance to indinavir, a human immunodeficiency virus type 1 protease inhibitor. J Viral 1996;70:8270-6. Patick A, Duran M, Cao Y, et al. Genotypic analysis of HIV-l variants isolated from patients treated with the protease inhibitor Neltinavir, alone or in combination with d4T or AZT and 3TC [Abstract]. Presented at the Fourth Conference on Retroviruses and Opportunistic Infections, Washington, DC, 1997. Jacobsen H, Haenggi M, Ott M, Duncan I, Andreoni M, Vella S, Mous J. Reduced sensitivity to saquinavir: An update on genotyp- ing from phase I/II trials. Antiviral Res 1996;29:9557. Taburet A, Singlas E. Drug interactions with antiviral drugs. Clin Pharmacokinet 1996;30:385-401. J*umal of Acquired Immune Deficiency Syndromes and Human Retrovirology, Vol. 15, No. 5, 1997 
work_ev5sudcnnrf5ddyn4x5riblkw4	----	<4D6963726F736F667420576F7264202D20C8EDF4EEF0ECE0F2E8E7E0F6E8FF20EEE1F0E0E7EEE2E0EDE8FF203230313320B9322E646F6378> 122 ИСПОЛЬЗОВАНИЕ ЭКСПЕРТНЫХ СИСТЕМ ПРИ СОСТАВЛЕНИИ РАБОЧИХ УЧЕБНЫХ ПЛАНОВ И.Ю. Пикалов, В.А. Кудинов Кафедра методики преподавания информатики и информационных технологий Курский государственный университет ул. Радищева, 33, Курск, Россия, 305000 В работе дается понятие экспертной системы, приводится ее состав и области применения. Описаны возможности экспертных систем при работе в режиме ввода знаний и режиме консульта- ции при разработке рабочих учебных планов и выборе студентами дисциплин по выбору. Выявлены преимущества использования экспертной системы при разработке основных образовательных про- грамм высшего профессионального образования. Ключевые слова: экспертная система, основная образовательная программа, рабочий учеб- ный план, компетенции. В настоящее время специалистам все чаще приходится иметь дело с нефор- мализованными задачам, т.е. задачами, в которых трудно или невозможно создать точно математическое описание или построить алгоритм решения. Для решения таких задач используются экспертные системы. Под экспертной системой (ЭС) будем понимать систему искусственного интеллекта, использующую знания из сравнительно узкой предметной области для решения возникающих в ней задач, причем так, как это делал бы эксперт-человек, т.е. в процессе диалога с заинтере- сованным лицом, поставляющим необходимые сведения по конкретному вопро- су [1]. По качеству и эффективности решения экспертные системы не уступают ре- шениям эксперта-человека. Решения экспертных систем обладают прозрачностью, т.е. могут быть объяснены пользователю на качественном уровне. Это качество экспертных систем обеспечивается их способностью рассуждать о своих знаниях и умозаключениях. Экспертные системы способны пополнять свои знания в ходе взаимодействия с экспертом [2]. В основе экспертной системы находится обшир- ный запас знаний о конкретной проблемной области. В большинстве случаев эти знания организуются как некоторая совокупность правил, которые позволяют де- лать заключения на основе исходных данных или предположений. Основу экс- пертных систем составляют база знаний, машина логического вывода и модуль извлечения знаний [1]. Коммерческое внедрение экспертных систем началось в 1980-х гг., с тех пор непрерывно происходит их беспрецедентное развитие и распространение. В на- стоящее время экспертные системы широко применяются в бизнесе, науке, техни- ке, сельском хозяйстве, на производстве, в медицине, в видеоиграх и в других направлении человеческой деятельности. В действительности трудно представить себе такую область, в которой бы в наши дни не использовались экспертные сис- темы [3]. Уже в работе Нейлора [4] описано создание экспертных систем с исполь- Пикалов И.Ю., Кудинов В.А. Использование экспертных систем при составлении рабочих... 123 зованием языка программирования высокого уровня, что значительно расширяет круг создателей таких систем. Особое значение экспертные системы приобретают в тех областях, где важ- но, чтобы принимаемые решения о стратегии и тактике развития региона были тщательно продуманы и обоснованы. Это относится и к подготовке высококва- лифицированных кадров в различных областях деятельности, что невозможно без качественно разработанных основных образовательных программ (ООП), со- ответствующих современному уровню развития науки и потребностям региона. Это становится особо актуально в связи с переходом на Федеральные государ- ственные образовательные стандарты (ФГОС) высшего профессионального обра- зования (ВПО) 3-го поколения. На наш взгляд, можно выделить четыре основных направления, которые яв- ляются наиболее значимыми при переходе на стандарты третьего поколения: — усиление межпредметных связей; — воспитательная составляющая; — связь с работодателями при составление и реализации ООП; — практическая направленность. Конкретные виды профессиональной деятельности, к которым готовится вы- пускник, должны определять содержание его основной образовательной програм- мы, разрабатываемой высшим учебным заведением совместно с объединениями работодателей. Одной из главных целей освоения образовательных программ по ФГОС ВПО третьего поколения являются компетенции, сформированные уча- щимся в ходе обучения, при этом под термином компетенция понимается спо- собность применять знания, умения и личностные качества для успешной дея- тельности в определенной области. Содержание компетенций, которые планируется формировать в процессе обу- чения в вузе, определяет состав дисциплин и содержание их программ, а, следо- вательно, поможет составить рабочий учебный план направлений подготовки, определиться со специализацией бакалаврских и магистерских программ. Разра- ботчики ООП должны действовать в тесном сотрудничестве с экспертами из всех областей знаний, которые тем или иным образом затрагиваются в ней. Нам кажет- ся, что для реализации эффективной работы по созданию рабочих учебных планов ООП потребуется экспертная система. Как правило, ЭС может функционировать в двух режимах: 1) режим ввода знаний — в этом режиме эксперт посредством редактора базы знаний вводит известные ему сведения о предметной области в базу знаний ЭС; 2) режим консультации — пользователь ведет диалог с ЭС, сообщая ей све- дения о текущей задаче и получая рекомендации ЭС. При работе в режиме ввода знаний экспертная система должна иметь следу- ющие возможности. 1. Составлять банк формируемых компетенций. Для этой работы должны быть привлечены как научно-педагогические работники и Учебно-методическое объединение учебного заведения, так и объединения работодателей. Для решения данной задачи мы сначала предлагаем предприятиям и организациям области, Вестник РУДН, серия Информатизация образования, 2013, № 2 124 а также другим возможным работодателям поработать с экспертной системой с целью формирования банка компетенций, востребованных на рынке труда, ко- торые нужно сформировать у бакалавра или магистра соответствующего направ- ления подготовки. Это можно сделать через сайт университета или факультета. При этом можно работать как с самой экспертной системой, так и опосредованно, например, заполнив анкету, а результаты анкетирования в систему будет вносить специальный работник. В банк формируемых компетенций войдут все компетен- ции, которые указаны в ФГОС ВПО, и добавятся те компетенции, которых нет в указанном списке, но достижение которых, по мнению предприятий, будет же- лательным. 2. Сопоставлять знания, умения и навыки сформированным в системе ком- петенциям. Для этого с системой должны работать специалисты из тех предмет- ных областей, к которым относятся выбранные компетенции. 3. Каждой компетенции поставить в соответствие одну или несколько учеб- ных дисциплин. 4. Обучить систему следующим действиям: а) под выбранные компетенции предлагать нужную дисциплину для данного направления подготовки (специальности); б) указывать какую компетенцию (или какие компетенции) формирует вы- бранная дисциплина; в) посмотреть какие знания умения и навыки формируются указанной ком- петенцией (так называемый паспорт компетенции); г) под указанные знания, умения и навыки предлагать возможные компе- тенции. В режиме консультации можно будет использовать систему при решении следующих задач: — закреплять конкретные компетенции за направлением подготовки или специальностью; — в стандарте указываются учебные дисциплины, обязательные для изучения, а также перечень дисциплин для разработки примерных программ. Можно по- смотреть, какие компетенции этими дисциплинами формируются, и сразу ото- бражать это в матрице компетенций и учитывать при дальнейшей разработке ООП; — под выбранные компетенции формировать перечень и порядок изучения учебных дисциплин при формировании рабочего учебного плана; — посмотреть паспорт любой компетенции. Таким образом, система будет полезна бакалаврам (или магистрантам) при выборе конкретного профиля направления подготовки, программы магистерской подготовки или специализации, а также для выбора своей собственной траектории обучения. Это можно сделать и через дополнительный набор компетенций, кото- рые обучающийся хочет сформировать, и через набор знаний, умений и навыков, и даже исходя из той области, где он хочет работать. В понятие траектории обуче- ния мы вкладываем набор дисциплин по выбору, последовательности их изучения и составление индивидуального плана изучения образовательной программы. Пикалов И.Ю., Кудинов В.А. Использование экспертных систем при составлении рабочих... 125 Использование экспертной системы при составлении рабочего учебного плана имеет ряд преимуществ. Во-первых, со временем система будет расширяться и на- капливать соответствие компетенции и учебные дисциплины, что позволит исполь- зовать ее не только при составлении рабочего учебного плана одного конкрет- ного направления подготовки, но и для множества образовательных программ. Во-вторых, можно составлять рабочие учебные планы направлений подготовки без повторного привлечения экспертов из различных предметных областей. В-третьих, данную систему можно использовать не только разработчикам ООП направления подготовки или специальности, но и студентам при составлении индивидуального плана обучения и выборе дисциплин по выбору. В-четвертых, систему легко мож- но будет переобучать в соответствии с меняющимися требованиями рынка труда. ЛИТЕРАТУРА [1] Брукинг А., Джонс П., Кокс Ф. и др. Экспертные системы. Принципы работы и приме- ры. — М.: Радио и связь, 1987. [2] URL: http://www.mari-el.ru/mmlab/home/AI/7_8/ [3] Джарратано Д., Райли Г. Экспертные системы: принципы разработки и программиро- вание. — М.: Вильямс, 2007. [4] Нейлор К. Как построить свою экспертную систему. — М.: Энергоатомиздат, 1991. LITERATURA [1] Bruking A., Dzhons P., Koks F. i dr. Jekspertnye sistemy. Principy raboty i primery. — M.: Radio i svjaz’, 1987. [2] URL: http://www.mari-el.ru/mmlab/home/AI/7_8 [3] Dzharratano D., Rajli G. Jekspertnye sistemy: principy razrabotki i programmirovanie. — M.: Vil’jams, 2007. [4] Nejlor K. Kak postroit’ svoju jekspertnuju sistemu. — M.: Jenergoatomizdat, 1991. USING EXPERT SYSTEMS IN DRAWING UP A WORKING CURRICULUM I.Y. Pikalov, V.A. Kudinov Chair of a technique of teaching of informatics and information technologies Kursk state university Radishchev str., 33, Kursk, Russia, 305000 The concept of an expert system is given, its structure and range of application are presented in the article. The possibilities of expert systems at a mode of knowledge input and a mode of consulting at ela- boration and choosing disciplines at students’ option are described. The advantages of the expert system use at elaboration of higher education are revealed. Key words: expert system, principal educational program, curriculum, competence. 
work_evgdqfryuze4lkand5nsb47sqq	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220980593 Params is empty 220980593 exception Params is empty 2021/04/06-03:05:22 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220980593 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:22 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_evgugtyf3ne7dkovektow4doia	----	[PDF] Fuzzy expert systems: an application to short-term load forecasting | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1049/IP-C.1992.0066 Corpus ID: 53663668Fuzzy expert systems: an application to short-term load forecasting @inproceedings{Hsu1992FuzzyES, title={Fuzzy expert systems: an application to short-term load forecasting}, author={Y. Hsu and K. Ho}, year={1992} } Y. Hsu, K. Ho Published 1992 Engineering An expert system using fuzzy set theory is presented for short-term load forecasting. Since most statistical methods for short-term load forecasting rely heavily on weather variables and statistical models, errors may appear in the forecasted hourly loads due to uncertainties in weather variables and statistical models. Thus, to have better accuracy, the operators in many utilities try to update the forecasted loads in real time using the records of the past few hours and their heuristic rules… Expand View via Publisher ntur.lib.ntu.edu.tw Save to Library Create Alert Cite Launch Research Feed Share This Paper 97 CitationsHighly Influential Citations 4 Background Citations 19 Methods Citations 23 View All Figures and Tables from this paper table 1 figure 1 figure 2 table 2 figure 3 table 3 table 4 figure 5 table 5 figure 6 figure 7 View All 11 Figures & Tables 97 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency Short term load forecasting using hybrid adaptive fuzzy neural system: The performance evaluation Daud Mustafa Minhas, Raja Rehan Khalid, G. Frey Computer Science 2017 IEEE PES PowerAfrica 2017 8 Save Alert Research Feed Short-term load forecasting by a neuro-fuzzy based approach R. Liang, Ching-Chi Cheng Engineering 2002 65 Save Alert Research Feed Combined regression-fuzzy approach for short-term load forecasting R.-H. Liang, C. Cheng Engineering 2000 28 Save Alert Research Feed Short term electric load forecasting using a neural network with fuzzy hidden neurons S. Kuusisto, M. Lehtokangas, J. Saarinen, K. Kaski Computer Science Neural Computing & Applications 2005 13 View 1 excerpt, cites methods Save Alert Research Feed Short-term electric load forecasting based on a neural fuzzy network S. Ling, F. Leung, H. Lam, P. Tam Computer Science IEEE Trans. Ind. Electron. 2003 97 PDF View 1 excerpt, cites methods Save Alert Research Feed Neuro-Fuzzy Approach Based Short Term Electric Load Forecastig B. K. Chauhan, A. Sharma, M. Hanmandlu Engineering 2005 IEEE/PES Transmission & Distribution Conference & Exposition: Asia and Pacific 2005 10 View 1 excerpt, cites methods Save Alert Research Feed Study of Load Forecasting Techniques using Fuzzy Logic A. Tale, Arvind Singh Gusain, Jyotirmoy Baguli, Rubina Sheikh, Altaf Q. H. Badar Computer Science 2017 6 PDF Save Alert Research Feed Short term load forecasting using fuzzy neural networks A. Bakirtzis, J. Theocharis, S. Kiartzis, K. J. Satsios Engineering 1995 218 Save Alert Research Feed Daily load forecasting with a fuzzy-input-neural network in an intelligent home S. Ling, F. Leung, P. Tam Computer Science 10th IEEE International Conference on Fuzzy Systems. (Cat. No.01CH37297) 2001 5 PDF Save Alert Research Feed Short-term load forecasting methods: A review A. Srivastava, A. Pandey, D. Singh Engineering, Computer Science 2016 International Conference on Emerging Trends in Electrical Electronics & Sustainable Energy Systems (ICETEESES) 2016 74 View 1 excerpt, cites methods Save Alert Research Feed ... 1 2 3 4 5 ... References SHOWING 1-10 OF 15 REFERENCES SORT BYRelevance Most Influenced Papers Recency Short term load forecasting of Taiwan power system using a knowledge-based expert system Kun-Long Ho, Y. Hsu, +4 authors Kung-Keng Chen Computer Science 1990 250 Save Alert Research Feed An expert system based algorithm for short term load forecast S. Rahman, R. Bhatnagar Engineering 1988 465 Save Alert Research Feed A regression-based approach to short-term system load forecasting A. Papalexopoulos, T.C. Hesterberg Mathematics Conference Papers Power Industry Computer Application Conference 1989 748 PDF Save Alert Research Feed Short-Term Load Forecasting Using General Exponential Smoothing W. R. Christiaanse Engineering 1971 392 Save Alert Research Feed ALFA: automated load forecasting assistant K. Jabbour, J.F.V. Riveros, D. Landsbergen, W. Meyer Engineering 1988 102 Save Alert Research Feed Short-term load forecasting G. Gross, F. Galiana Engineering Proceedings of the IEEE 1987 837 PDF Save Alert Research Feed The role of fuzzy logic in the management of uncertainty in expert systems L. Zadeh Mathematics 1983 1,245 PDF Save Alert Research Feed Weather sensitive electric demand and energy analysis on a large geographically diverse power system &#2014;Application to short term hourly electric demand forecasting R. Thompson Engineering IEEE Transactions on Power Apparatus and Systems 1976 33 Save Alert Research Feed Fuzzy Set Theory - and Its Applications H. Zimmermann Computer Science 1985 6,327 Save Alert Research Feed Practical Application of Weather Sensitive Load Forecasting to System Planning J. Davey, J. J. Saacks, G. W. Cunningham, K. Priest Engineering 1973 18 Save Alert Research Feed ... 1 2 ... Related Papers Abstract Figures and Tables 97 Citations 15 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_ew2dceh4x5g5heioqvpyjyw2ge	----	Title of the Paper: The Chemical Thermodynamic Module of The Expert System for Thermodynamics (“TEST”) Web Application Subrata Bhattacharjee, subrata@thermo.sdsu.edu Christopher P. Paolini, paolini@engineering.sdsu.edu Computational Thermodynamics Laboratory Mechanical Engineering Department San Diego State University, San Diego, CA 92084 Abstract Accessible from www.thermofluids.net, the TEST web portal is a freely accessible thermodynamics web-based software package for engineering education. It combines several modules – multi-media problems and examples, an online solution manual for educators, traditional thermodynamic charts and tables, fifteen chapters of animations to illustrate thermodynamic systems and fundamental concepts, and a suite of thermodynamic calculators called daemons for solving thermodynamic problems and pursuing what-if scenarios. Designed as ‘browsable’ Java applets, these daemons cover a wide range of topics - property evaluation of working substances, energy, entropy, and exergy analysis of generic open and closed systems, IC engines, gas and vapor power cycles, refrigeration, HVAC, combustion, chemical equilibrium, and gas dynamics. In this paper, the chemistry module comprising of combustion and equilibrium analysis daemons for both open steady systems and unsteady processes is presented. The combustion daemons can be used for the analysis of both premixed and non-premixed reactants modeled with a perfect gas, ideal gas, or a solid/liquid model. Using the intuitive graphical user interface of the reaction panel, a user can quickly set up standard reactions such as theoretical reactions, reactions with a known amount of excess or deficient air, and reactions with known dry products analysis. The balanced equations are then used in the state panel and the states of fuel, oxidizer, and products are fully or partially calculated based on the known properties at the ports of a steady state reactor or at the beginning and final states of a process. Finally, in the device (or process) panel, known information about the device (or process) is supplied and the daemon updates the entire analysis to produce important combustion parameters such as the adiabatic flame temperature, heat transfer, and entropy generation in the device or the process. Once a solution is achieved, what-if studies can be performed by changing the fuel, the air fuel ratio, or any other controllable parameter by simply changing the parameter and calculating a single update button. The equilibrium daemons handle gas mixtures and condensed phases and builds upon the combustion daemons. Reactants are entered by selecting from a list of species and specifying their moles. From a list of plausible products, the user can select product species. In the state panel, known thermodynamic properties such as pressure and temperature are specified. The products composition is then calculated by minimizing the Gibbs function of the products mixture. These simple-to-use daemons, comparable in their accuracy to established codes such as NASA CEA1 and STANJAN2, are used to determine equilibrium constant, equilibrium http://www.thermofluids.net/ composition, and equilibrium temperature of methane combustion as a function of equivalent ratio. Introduction The advantages of web-based educational resources are obvious. Without any need for installation, students can access the latest version of software from any computer platform. Developers can update their software in real time and instructors can assign students exercises that require use of web-based software without having to rely on IT support personnel to make the software explicitly available on systems in university computer laboratories. Requiring only a standard web browser to operate, students can access these tools using devices other than laptops such as high-end mobile phones. Advances in Java programming, JavaScript, CSS, and other browser technologies can enhance3 the quality of education in a classroom. With the recent introduction of Web Service technology, data and software can be seamlessly integrated with any type of client application4. This can allow development of very powerful web applications that can exchange information with the server only when necessary, thereby giving the application a look and feel of a locally installed application while providing access to rich databases and calculation algorithms. In the last couple of decades, a number of software applications for thermodynamic analysis have been developed. The ASME provides a comprehensive database for thermodynamics analysis5. However, EES6 by F-Chart Software can be used to solve large sets of equations with built in functions for thermodynamic and transport properties of many substances. TPX7 is an Excel plug-in where the core thermodynamic state can be evaluated by entering two independent properties. Thermoptim8 is a Java based software for analyzing thermodynamic cycles and can be run over the web. A graphical menu driven installable program is GasCAD9. Another installable program is CyclePad10, which allows users to construct and analyze a wide variety of thermodynamic cycles. However, none of these resources cover the entire range of thermodynamic topics covered in a typical engineering thermodynamics textbook. For solving combustion and equilibrium problems, educators frequently rely on STANJAN2 or, in the case of advanced analysis in a graduate class room, NASA CEA1,11, both written in FORTRAN. There are other applications such as GasEQ12, REACT13, REACT! 14, SOLGASMIX15, FactSage16 to name a few. Most of these applications must be locally installed on a PC and run only under the Windows operating system. Recently, Paolini and Bhattacharjee developed a platform independent equilibrium calculation tool 17 that uses NASA CEA thermochemical data. It incorporates gas and condensed species and can calculate equilibrium composition along with other state properties. The Expert System for Thermodynamics or (“TEST”), freely accessible from the portal www.thermofluids.net, was launched18 to make thermodynamic calculations more accessible and user friendly. More than 17,000 students, educators and professionals have registered to use TEST. A visual manual for TEST has been published by Prentice Hall19. Some of the daemons have been reviewed and included in various public domain educational web sites such as the MERLOT project20. A comparison of various multimedia materials for statistical and thermal http://www.thermofluids.net/ physics ranked TEST in the top category21. Other articles that refer to TEST can be found in the references22, 23,24, 25, 21,26,27,28. In this article, the chemical module of TEST is presented with examples of practical problem solving. The module can be used for balancing a reaction, determining the equivalence ratio, evaluating adiabatic flame temperature, analyzing a combustion chamber for complete reactions, determining equilibrium composition, equilibrium flame temperature, and analyzing open and closed systems with equilibrium or complete chemistry. Through a few button clicks, the user friendly graphical user interface will produce equilibrium composition or temperature in a combustion chamber as the equivalence ratio is varied over a wide range. Comparisons will be made with other established chemical analysis applications. Figure 1. Chapters 13 and 14 of the Problems and Examples sections deal with combustion and chemical equilibrium. Combustion and Equilibrium Problems TEST can be used as courseware for thermodynamics education and can complement leading thermodynamics textbooks 29, 30, 31. While there are a few excellent online coursewares available32,33, TEST brings together pedagogic discussions, animations, examples and problem sets, and online solvers for the entire range of topics covered by engineering thermodynamics. All of TEST’s resources are accessible from the task bar shown in Figure 1. The problems and examples link displays fifteen chapters. Chapter 13 deals with combustion and chapter 14 with phase and chemical equilibrium. Each problem is linked to a detailed manual solution and instructions for obtaining a TEST solution. While manual solutions to examples are accessible to all, only registered educators are allowed access to manual solutions in the Problems pages. Clicking on the solution link splits the screen into two frames with the solution displayed in the lower frame. Sometimes problem schematics, displayed in the right margin, are linked to animations, images, or tables and charts, which are also displayed in the lower frame. Each chapter is divided into several sections, such as the three sections – Reactions, Open Systems, Closed Systems - shown in Figure 1. Thus, problem 13-2-24 refers to the 24th problem in section 2 of chapter 13. Combustion Tables and Charts The Map link on the task bar (see Figure 1) brings up a map of the daemons offered by TEST. Fifty charts and tables are linked from the Tables & Charts link shown in 2. Several of those can be useful for a reaction analysis. Tables D.4 through D.15 are thermodynamic tables for twelve combustion species in SI units and Tables D.4E through D.15E are the corresponding tables in English units (see 3). Each table displays Figure Figure o ,f kh and kM for the species in the header and lists properties k ( )kh T , ( )ku T , and ( )0ks T in increments of the independent variable temperature T . Tables G.1 and G.1E list ( )0ks TkM , o,f kh , o ,f kg , and for twenty-nine species including several fuels in their condensed phase. Tables G.2 and G.2E list lower and higher heating values of several fuels. Table G.3 lists equilibrium constants in ln ( )pK form for several stoichiometric reactions as a function of temperature. The use of Cascading Style Sheets (CSS) makes these tables easily readable and the HTML format makes it easy to switch between tables or switch units. State Daemons TEST is a web-based suite of applications for analyzing engineering thermodynamic systems. Java applets designed to solve a particular type of problem are called daemons in TEST. The state module34, for example, contains a number of daemons to evaluate system and flow states of various working substances. There are twenty-six customized daemons for evaluation of states of different working substances. These working substances include 50 common solids and liquids, 65 phase-change fluids, 60 gases, and several gas mixtures. Figure 2. The Map link brings up a clickable map of different classes of TEST daemons. Figure 3. Traditional tables and charts for combustion and equilibrium analysis in both SI and English units are easily accessible. To illustrate how the state daemons can be used to look up thermodynamic properties, suppose we would like to evaluate specific enthalpy and specific entropy of carbon dioxide at 2525 K and 1 MPa. Interpolating ( )kh T and ( )0ks T from Table D.6, we obtain ( ) ( ) ( ) ( ) IGIG IG o ,, 2000 298k k f k k kh p T h T h h h⎡ ⎤= = + −⎣ ⎦ ( ) ( ) 2 IG CO 1 MPa, 2525 K 393, 522 123, 465 0 270, 057 kJ/kmolh⇒ = − + − = − (1) ( ) ( )IG 0 0 , lk k p s p T s T R p = − n ( ) 2 IG CO 1000 kJ 1 MPa, 2525 K 323.417 8.314 ln 304.35 101 kmol K s⇒ = − = ⋅ (2) Although the ideal gas (IG) daemon can be used to evaluate a complete thermodynamic state of a gas, the enthalpy calculated does not include the enthalpy of formation which is displayed in the message panel instead. The n-IG mixture daemon, located at the Daemons> States> System> n- IG Model page and shown in Figure 4, is more suitable for evaluating properties of pure gases and their mixtures. Launch the daemon by selecting the n-IG (ideal gas mixture of n species) model in the system state page linked from the TEST map. To evaluate a state, create an ideal gas mixture – in this case pure CO2 – by adding 1 kg of CO2. Make sure the enthalpy of formation is included in the calculation by clicking the Yes radio button. Enter p1 and T1 in appropriate units. The expert system behind the daemon will not allow any further properties to be entered since two thermodynamic properties are sufficient for finding an equilibrium state. Press the Enter key or the Calculate button and the complete state is Figure 4. Properties of carbon dioxide evaluated at 1 MPa, 2525 K by the n-IG system state. calculated and displayed. The green background signifies input and the cyan background indicates a computed property. The property symbols are also color coded – red for material properties such as the molar mass MM, blue for all thermodynamic properties, green for extrinsic properties, and black for system properties34. While the ideal gas table displays only ( )kh T , ( )ku T , and ( )0ks T , the state daemon produces the complete set of thermodynamic properties – v , u , h , , , and . To obtain mole based values, switch to the I/O panel which serves as a calculator. Evaluate the expressions =MM1*h1 s g pc and =MM1*s1 to obtain and 270, 071 kJ/kmol− ( )304.46 kJ/ kmol K⋅ , respectively, which agree with the results of equation (1) and (2). The daemon is superior to the table, not only by eliminating interpolation, but by producing the complete state. Moreover, any combination of independent properties can be entered to produce a state. Given entropy and pressure or specific volume, for instance, the manual procedure is not trivial to obtain the temperature. The daemon, on the other hand, can calculate the state instantly, even from unusual combinations such as and . pc g igure 5). The open ess in a d into alancing Reactions o illustrate how the combustion da let us analyze a combustion chamber operating on > e, e Combustion Daemons The combustion daemons are organized in the system tree (see the TEST map in 2) under both open steady systems and closed process categories (see aemons are used for analyzing steady flow combustion chambers and reaction chambers, while the closed process daemons are meant for analyzing a reacting proc closed chamber. In each branch of the TEST Map, the combustion daemons are separate two types – non-premixed, where fuel and oxidizer are initially separated, and premixed, where the fuel and oxidizer are mixed before ignition occurs. Finally, for each type of daemon, two material models – IG (ideal gas) and PG (perfect gas) models – can be selected. Figure 5. Different combustion daemons are used to analyze open steady systems and closed processes. Figure F steady d B T emons work, in a steady state, described by problem 13-2-24 in Figure 1. Launch the non-premixed IG combustion daemon located in Daemons> Systems > Steady> Specific> Combusti Non-Premixed> IG-Model page. The daemon has four different panels – Reaction, State, Devic and I/O Panels – and the Reaction Panel is displayed by default as shown in > Open Figure 6. Select octane(L) in the fuel block. Standard reactions such as a theoretical reaction with air or pure oxygen can be instantly obtained by selecting the appropriate action from the action menu (se Figure 6). For 20% excess air, enter 1.2λ = and then select Excess/Deficient Air from the action menu. Products composition for complete combustion along with left over oxygen and nitrogen are displayed in the products block. The amount of air, 71.43 kmol per kmol of fuel, is the air- fuel ratio on a mole basis. Note that the equivalence ratio, φ , is also reported next to λ . Air ca be decomposed into a mixture of oxygen and nitrogen by selecting the Air->O2, N2 from the action menu. Note that for a reaction with deficient air, the same procedure can be followed with 1 n λ < . To determine the air-fuel ratio on a mass basis, select the mass radio button that appears above the Reaction Panel tab and the molar amount of each species is converted to m Select Normalize from the action menu to express the reaction in terms of 1 kg of fuel. The amount of air, 18.11 kg, then represents the mass based air-fuel ratio. The theoretical air fuel ratio, likewise, can be obtained as 151 kg of air/kg of fuel or 59.52 kmol of air/kmol of fuel. ass. i eaction between liquid octane and 20% excess air is bala using the three steps descri n this no mmon type of reaction encountered in com nvolves products analysis. Given n e or F gure 6. R ther co nced bustion i bed i image. The product composition for complete combustion is automatically selected. A a dry product analysis, a reaction can be balanced by selecting the fuel, oxidizer, and products species, and then entering the known amounts of products components until there are as many unknowns as atoms. For example, suppose a volumetric dry products analysis in the combustio of octane with air produces (see Problem 13-1-19) 10.02% CO2, 0.88% CO, 5.62% O2, and 83.48% N2. To solve this problem, select octane as the fuel, and enter CO2, CO, and O2 amounts in kmol in the products block. This leaves the fuel, air, and nitrogen amounts as unknowns and three atom balance equations. Select Balance Equation from the action menu and then normaliz to produce a molar air fuel ratio of 77.63. With the molar theoretical air fuel ratio already calculated as 59.52, the percent theoretical air used can be calculated as 77.63/59.52 = 1.30 130%. Analysis of Steady-State Reactors n e evaluated with the mass flow rates obtained f ows an The State Panel picks up the composition of fuel, oxidizer, and products from the Reaction Panel. Returning to the combustion of liquid octane with air (see Figure 1), we can express the reaction in terms of 1 kg of fuel. Using a scaling factor of 0.002 and selecting Multiply by Scaling Factor from the action menu, the reaction is expressed i terms of 0.002 kg of octane. In the State Panel, evaluate state-1 for fuel, state-2 for oxidizer, and state-3 for products. A state number must be associated with fuel, oxidizer, and products mixtures. Enter the known pressure and temperature at each state and the states ar rom the Reaction Panel. The Device Panel sh image of the combustion chamber with the mass, energy, and entropy equations simplified for an open steady system. After selecting state-1 (fuel) as the i1 (inlet-1) state, state-2 (oxidizer) as the i2 (inlet-2) state, and state-3 (products) as the e (exit) state, observe in Figure 8 how the image of the system adjusts as the port states are loaded. Enter the known amount of external work and press the Calculate button to solve the mass, energy, and entropy equations. Because the states are already completely known, the heat transfer and entropy generation rates are calculated and displayed. The negative amount, -33.38 kW, indicates that heat is being rejected by the engine at this rate. Figure 7. Evaluate three states for fuel, oxidizer, and products. igure 8. The device panel of the open steady non-premixed combustion daemon showing the solution Problem o determine the adiabatic flame temperature, set T3 as an unknown in state-3, and enter Qdot and Wdot_ext as zero in the Device Panel. The Super-Calculate button can be used to update all F to 13-2-24 in Figure 1. T solutions and T3 is calculated and posted as state-3. A detailed output is also generated by the Super-Calculate operation. For a parametric study – say, how the adiabatic flame temperature changes with the equivalence ratio – change lambda, select Excess/Deficient Air from the actio menu, and then click Super-Calculate. For theoretical combustion ( 1 n λ = ), the detailed output shown in is Figure 9 with the adiabatic flame temperature for a complete theoretical reaction calculated as 2396 K. The fuel can also be used as a parameter. To do so, de-select octane in the fuel block and select methane. From the action menu, select Theoretical Air to produce the balanced complete reaction. Now, press the Super-Calculate button to produce a T3 of 2329 K as the adiabatic flame temperature. Figure 9. Adiabatic flame temperature for theoretical combustion of liquid octane obtained from the open-steady non-premixed combustion daemon. a fuel can be accomplished by configuring a theoretical reaction nd specifying standard atmospheric conditions for fuel, oxidizer, and products. The heating e g he hemical exergy can also be obtained by simply evaluating the or reversible work in the IO Panel as Evaluation of heating values of a value is simply the heat transfer calculated in the device panel for 1 kg of fuel. For example, for methane, configuring a theoretical reaction on the basis of 1 kg of fuel, the heat transfer can b calculated as 50,024 kW, that is, the heating value is 50,024 kJ/kg since the daemon interprets the fuel mass as the mass flowing into the chamber every second. This must be the lower heatin value since the water in the products is assumed to be in the vapor phase. To obtain the higher heating value, simply de-select H2O in the products block and replace it with H2O(L), as shown in Figure 10. Balance the reaction by selecting Balance Reaction and then pre per-Calculate button. The higher heating value in the device panel is then shown as 55,511 kJ/kg. These values can be verified from t tabulated values of higher and lower heating values of different fuels in Table G.1. The approximate value of the fuel c expression f ss the Su Figure 10. To obtain a higher heating value, de-select H2O and select H2O(L) from the products menu. Select Balance Reaction from the action menu and then press Super- Calculate to update all results. =mdot1*g1+mdot2*g2-mdot3*g3 f flect a single inlet and a single exit for the premixed reactor. Balancing a reaction requires selecting the The above expression evalu imation because the ontribution of mixing to the exergy is ignored in this result. n in the previous examples. The orresponding PG daemon differs only in one aspect – it assumes constant specific heat values of ce omes from the fact that the Reaction Panel now has only two blocks – one for reactants and one ates to 50,869 kJ/kg for methane, an approx c So far, we have used the non-premixed open-steady IG daemo c for each component. Analysis with the PG daemons may not be accurate, especially at high temperatures, but may serve a useful purpose in establishing the importance of variable specific heat and for verification of manual solutions. For example, the adiabatic flame temperature methane calculated with the PG model is 2792 K as opposed to 2329 K from the IG daemon. The premixed daemons work almost the same way as the non-premixed daemons. The differen c Figure 11. The Process Panel set up for calculating the adiabatic temperature in a closed chamber. or products. Therefore the State Panel and the Device Panel are also modified to re reactants and products and entering amounts of ( )n a− components, where n is the total number of species that appear in the reaction and a is the number of atoms. For the theoretical combustion of methane, the reaction can be balanced by entering 1 kg of me ane and 17.201 kg of air as reactants, selecting the products of complete combustion, and then selecting Bal Reaction from the action menu. In the State Panel, evaluate the reactants and products states. The heating value, 55,511 kJ/kg, calculated in the Device Panel is identical to the one determined b the non-premixed daemon. However, the reversible work (i.e. =mdot1*g1 – mdot2*g2) th ance y calculated in the I/O Panel as 50757 kJ/kg is slightly lower because the fuel in the reactants now has a lower partial pressure than in the non-premixed case. The entropy generation rate in the reactor is calculated as part of the device variables and ported in the Device Panel (see Figure 8). The boundary temperature T_B in the Device Panel osed Combustion Processes rium daemons for analyzing unsteady combustion rocesses executed in a closed system. The fuel and oxidizer can be premixed in a closed , K, and e . In ss cal Equilibrium Analysis hemical equilibrium analysis, which ibrium c composition, n steady systems and closed processes. re at a given mperature and pressure. The entropy maximum principle for an isolated system at equilibrium re refers to the temperature of any external thermal energy reservoir with which the heat interaction occurs. In most situations, it can be left at the default value of 298 K, the temperature of standard surroundings. Analysis of Cl TEST offers separate combustion and equilib p chamber or separated into two sub-chambers in the beginning of the process. As an example, consider the combustion of 1 kg of liquid octane with a theoretical amount of air, 15.096 kg obtained from any non-premixed daemon. The reactants are kept at 1 atm, 25 o C in a rigid chamber and ignited. To analyze the resulting process, launch the premixed closed-process IG combustion daemon located in the Daemons> Systems> UnsteadyProcess> Specific> Combustion> Premixed> IG Model page. Configure the reaction in the Reaction Panel for theoretical combustion of octane. Evaluate the reactants state with p1 = 1 atm, T1 = 298 partially evaluate the products state with Vol2 = Vol1 (for a constant-volume process). In th Process Panel, load state-1 as the beginning state (b-state) and state-2 as the final state (f-state), and enter Q = W = 0. Click Calculate and then Super-Calculate to obtain the final state. The final pressure and temperature are found in state-2 as 1065 kPa and 2911 K, respectively processes where a boundary work transfer is involved, work must be evaluated manually (or using an expression such as ‘=p1*(Vol2- Vol1)’) and entered in the Proce Panel. Chemi Figure 12. The equilibrium daemons for analyzing open- teady closed-process be found in the same pages s reactors can TEST offers several daemons for as the corresponding combustion daemons. c can be used for evaluation of equil and energy and entropy analysis of ope There are two approaches for evaluating the equilibrium composition of a mixtu onstants, determination of equilibrium te can be shown to reduce to the minimization of the system’s Gibbs function if the temperature and pressure remain constant. Assuming the mixture to consist of n given species, the mixture’s Gibbs function can be expressed as a function of n unknown mole fractions subject to an atom balance constraint. The traditional numerical practice used for solving optimization problems35,36 of this type is the method of Lagrange multipliers using an iterative Newton-Raphson technique for solving the resulting set of nonlinear equations. Both NASA CEA1 and STANJAN2 employ the method of element potentials. Both of these codes are written in FORTRAN and cannot be easily incorporated into other codes. The second approach, used mostly for calculating a manual solution, uses equilibrium constants that relate mole fractions of products and reactants derived from the Gibbs function minimization principle. A number of simultaneous stoichiometric equations are selected and their equilibrium constants are obtained from a table such as Table G.3 at a given temperature. By selecting as many reactions as the number of unknowns, n a− tials. Front- on, located (where a is the number of different atoms), a set of simultaneous algebraic non-linear equation can be constructed. For systems with few species (say, the dissociation of oxygen at high temperature), a manual solution of these equations can be obtained which results in the composition of the equilibrium mixture. Paolini and Bhattacharjee17 modified the C s EA algorithm and developed an object-oriented Java ode f olving for an equilibrium distribution using the method of element potenc or s sti end Java applets with easy-to-use graphical user interfaces were developed to run this equilibrium code through a web browser. These applets are enhanced and integrated with the chemical module of TEST in this work. The open-steady equilibrium daemon, located in Daemons> Systems> Open> Steady> Specific> ombu on> Chemical Equilibrium page, and the closed-process equilibrium daemC in Daemons> Systems> Closed> Process> Specific> Combustion> Chemical Equilibrium page, have the same look and feel as the corresponding non-premixed combustion daemon with the addition of a new panel called the Composition Panel. To illustrate the use of the equilibrium daemons, suppose we are to evaluate the equilibrium constant, ( )ln pK , of the stoichiometric reaction 2 2CO CO 0.5O+ at 3000 K, which is listed in Table .3 s -1.111. As depicted in Figure 13, launch the open-steady equilibrium daemon and switch to the Composition Panel. To configure eaction, select CO2, change its molar amount to 1 kmol, and press G a the stoichiometric r Figure 13. Determination of the equilibrium constant for the dissociation reaction 3000 K 2 2 at CO CO 0.5O+ . the Enter amount o nte amount o key to register the input. Notice that the products table on the right gets populated by all atomically feasible species. Select CO and O2 in products table, leaving their amounts unknown. In the State Panel, enter p1 = 100 kPa, T1 = 3000 K, select the Products radio button, and press the Calculate button. The stoichiometric reaction is balanced and the log of the equilibrium constant ( )ln pK is reported in the Composition and I/O Panel as -1.108. If this is the only possible reactio equilibrium products will contain all three species. To determine the equilibrium composition of 1 kmol of CO2 at 100 kPa and 3000 K r key to register the input. Notice that the products table on the right gets populated by all atomically feasible species. Select CO and O2 in products table, leaving their amounts unknown. In the State Panel, enter p1 = 100 kPa, T1 = 3000 K, select the Products radio button, and press the Calculate button. The stoichiometric reaction is balanced and the log of the equilibrium constant ( )ln pK is reported in the Composition and I/O Panel as -1.108. If this is the only possible reactio uilibrium products will contain all three species. To determine the equilibrium composition of 1 kmol of CO2 at 100 kPa and 3000 K , select as any proba 2, CO, O2, O, and C(graphite), for example, in the products The combination of the Reaction Panel along with the equilibrium solver creates a powerful tool r equilibrium energy and entropy analysis. In the next example, adiabatic flame temperature is mbustion chamber at 1 atm and 298 K. Do a perature T depends on the equivalence ratio eq , select as any proba 2, CO, O2, O, and C(graphite), for example, in the products The combination of the Reaction Panel along with the equilibrium solver creates a powerful tool r equilibrium energy and entropy analysis. In the next example, adiabatic flame temperature is mbustion chamber at 1 atm and 298 K. Do a perature T depends on the equivalence ratio n, the ble species – CO, O2, CO n, the ble species – CO, O2, COmm table. Since the State Panel is already configured, use the Super-Calculate button to update all results. The equilibrium composition is displayed in the Composition and I/O Panel as shown in Figure 13. Once the reactants (in this case 1 kmol of CO2) are configured, it becomes easy to perform what-if studies. For example, to learn how an increase in pressure would affect the f CO, change p1 to 1 MPa and click Super-Calculate. Shown in Figure 14, the amo of CO decreases from 0.454 kmol to 0.245 kmol, as Le Chatelier's Principle would predict fo the dissociation of CO2. table. Since the State Panel is already configured, use the Super-Calculate button to update all results. The equilibrium composition is displayed in the Composition and I/O Panel as shown in Figure 13. Once the reactants (in this case 1 kmol of CO2) are configured, it becomes easy to perform what-if studies. For example, to learn how an increase in pressure would affect the f CO, change p1 to 1 MPa and click Super-Calculate. Shown in Figure 14, the amo of CO decreases from 0.454 kmol to 0.245 kmol, as Le Chatelier's Principle would predict fo the dissociation of CO2. unt r unt r igure 14. Determination of the equilibrium composition of carbon dioxide at 100 kPa and 3000 K. F fofo calculated using different models – complete reaction with PG model, complete reaction with IG model, and equilibrium calculations with eight different species in the products. Example: Equilibrium Flame Temperature calculated using different models – complete reaction with PG model, complete reaction with IG model, and equilibrium calculations with eight different species in the products. Example: Equilibrium Flame Temperature Suppose methane and air enters an adiabatic co parametric study of how the adiabatic flame tem Suppose methane and air enters an adiabatic co parametric study of how the adiabatic flame tem φafaf . Vary φ over the range 0.5 through 1.5. Use TEST daemons with (a) the PG mixture model (complete combustion), (b) IG mixture model (complete combustion), and (c) IGE mixture model (equilibrium combustion) with the following eight components in the products: 2CO , CO , 2H O , OH , 2O , 2N , NO , 2NO . To solve this example problem using TEST, la th -s y -p ixed PG com tion daemon. On a separate tab, open the corresponding IG combustion daemon, and on a third browser tab, launch the open-steady equilibrium daemon In the Reaction Panel of each daemon, select methane in the fuel block, enter a value of 1 unch e . open tead non rem bus λ = (inverse of the equivalence ratio) for a theoretical mixture, and select Excess/Deficient Ai the action menu. The balanced reaction is displayed. For the combustion daemons, the procedure has been already outlined. In the State Panel, evaluate state-1 and state-2 as the fuel and oxidizer states from the known pressure and temperature. Evaluate state-3 for the products, partially, by entering a value for p1 only. In the Device Panel, import the fuel, oxidizer, and products states, enter Qdot = Wdot_ext = 0, and then click Super-Calculate. The adiabatic flame temperature for the theoretical reaction is given in state-3 as T3. In the equilibrium daemon, the procedure is simila r from r. In the Composition Panel, the reactants e amounts are already imported from the reaction panel. Select the eight products species. In th State Panel, evaluate the reactants state, state-1, for the given pressure (1 atm) and temperature (298 K). For the products state, state-2, select the Products radio button and enter the pressure (=p1). In the Device Panel, load the inlet and exit states, and enter Qdot = Wdot_ext = 0. Click Super-Calculate and the equilibrium products composition, as well as the complete products state, state-2, is calculated and displayed. The temperature at state-2 is the adiabatic equilibrium flame temperature. After compiling the three adiabatic a temperatures in a spread sheet, to obtain the adiabatic temperatures at different equivalence ratio, change λ to a new value in the reaction panel of each daemon and press the Super- Calculate button. Results compiled in this manner are plotted in Figure 15. As expected, the PG mode predicts the flame temperature due to the assumption of constant specific heat. Results from the equilibrium composition have been compared with NASA CEA with remarkable agreement. The IG model produces excellent agreement with equilibrium calculations for ultra-lean combustion. However, as the equivalence ratio increases, the discrepancy between the two m The mole or mass fraction of any species can be similarly plotted. Conclusion l over odels quickly increases as can be seen in Figure 15. llection of thermodynamic databases, software, and courseware, accessible from 1500 2000 2500 3000 0.5 0.7 0.9 1.1 1.3 1.5 T em p er at u re , K Equivalent Ratio TEST-PG Model TEST-IG Model TEST-IGE Model Figure 15. Adiabatic flame temperature for methane and air combustion as a function of equivalence ratio. Prediction from the equilibrium computation (IGE model) is the most accurate. TEST is a co www.thermofluids.net. TEST is freely accessible from any academic institution and is currentl used in more than 100 institutions around the world. The thermodynamic calculators in TEST, called daemons, have been developed using an object-oriented paradigm in Java so that code can y http://www.thermofluids.net/ be reused. All TEST daemons have been thoroughly tested under different versions of IE, Firefox, and Safari running on Windows, MacOS, and Linux platforms. The only browser p in required to run TEST daemons is version 4 (or greater) of the Java Platform. The chemical module of TEST can be used to access thermodynamic data, two chapters of problems and examples on combustion and chemical equilibrium, and several daemons for the analysis of and closed reacting systems. The combustion daemons can be used for balancing a reaction and carrying out energy and entropy analysis of a combustion chamber for a given reaction. The equilibrium daemons, which build upon the combustion daemon, preserve the simplicity of th combustion daemon and yet extend their usefulness by calculating the equilibrium composition of the products as part of the analysis. The capabilities of these equilibrium daemons are comparable to established applications such as NASA CEA37,1,11 and STANJAN2. Acknowledgement lug- open e NSF through the CyberInfrastructure CI-TEAM Grant References 1. Gordon S, McBride, BJ. Computer Program for Calculation of Complex Chemical 2. ilibrium Analysis: ersity; 3. Lee HP. Comparison between traditional and web-based interactive manuals for laboratory 4. Paolini CP, Bhattacharjee, S. A Web Service Infrastructure for Thermochemical Data. J. 5. mics, Combustions and Chemical Balance e/THERMO/MASTER.HTML" We gratefully acknowledge support from 0753283. Equilibrium Compositions and Applications – I. Analysis 1994. Reynolds WC. The Element-Potential Method for Chemical Equ Implementation in the Interactive program STANJAN. Palo Alto, CA: Stanford Univ 1986. Technical Report A-3391. based subjects. The International Journal of Mechanical Engineering Education. 2002; 30(4). Chem. Inf. Model. 2008; 48(7): 1511-1523. ASEE. Engineering Database – Thermodyna Analysis. Available at: HYPERLINK "http://www.mecheng.asme.org/databas http://www.mecheng.asme.org/database/THERMO/MASTER.HTML . Ac 21, 2008. cessed September 6. F-Chart Software. Engineering Equation Solver. EES - Engineering Equation Solver. Available at: HYPERLINK "http://www.fchart.com/ees/ees.shtml" http://www.fchart.com/ees/ees.shtml . Accessed January 31, 2009. Goodwin D. Thermodynamic Properties for Excel. Available at: H7. YPERLINK "http://kr.caltech.edu/me/software/tpx" http://kr.caltech.edu/me/software/tpx . Studies Cf E. Applied Thermodynamics Software - Thermoptim. Available at: 8. -HYPERLINK "http://www-cenerg.cma.fr/eng/software/thermoptim" http://www cenerg.cma.fr/eng/software/thermoptim . ProMetric Technologies. GasCad and Rea9. lGas. Available at: HYPERLINK "http://www.prometrictech.com/demodown.html" http://www.prometrictech.com/demodown.html . 10. Group QR. CyclePad. Available at: HYPERLINK /aboutcp.html" "http://www.qrg.ils.nwu.edu/projects/NSF/Cyclepad http://www.qrg.ils.nwu.edu/projects/NSF/Cyclepad/aboutcp.html . McBride BJ, Gordon, S. Computer Program for the Calculation of C11. omplex Chemical n 12. Morley C. GasEQ: A Chemical Equilibrium Program for Windows. October 27, 2007. Equilibrium Compositions and Applications – II. Users Manual and program Descriptio 1996. Available at: HYPERLINK "http://www.gaseq.co.uk/" http://www.gaseq.co.uk/ . Ramette RW. REACT: Exploring Practical Thermodynamic and Equilibrium Calcul13. ations. J. 14. Chemical Equilibrium Calculator. 1992. Available at: Chem. Educ. Software. 1995; 8B1. UnitOOPS Software. REACT! – A HYPERLINK "http://www.yale.edu/yaleche/chemeng/job/rxtinfo.htm" http://www.yale.edu/yaleche/chemeng/job/rxtinfo.htm . Accessed October 27, 2007. Eriksson G. Thermodynamic studies of high temperature equilibria. XII. SOLGASMI15. X, a . 16. Bale C, Chartrand, P, Degterov, S, et al. FactSage Thermochemical Software and Databases. 17. Paolini CP, Bhattacharjee, S. An Object-Oriented Online Tool for Solving Generalized ical 18. Bhattacharjee S. TEST - The Expert System for Thermodynamics. Paper presented at: ce & 19. ermodynamics: Prentice Hall; 2002. computer program for calculation of equilibrium compositions in multiphase systems. Chem Scr. 1975; 8: 100-103. Calphad. 2002; Vol. 26: 189-228. Chemical Equilibrium Problems. Paper presented at: 2008 ASME International Mechan Engineering Congress and Exposition (IMECE08), October 31 – November 6, 2008; Boston, Massachusetts, USA. Proceedings of the 2003 American Society for Engineering Education Annual Conferen Exposition, 2003; Nashville, Tennessee. Bhattacharjee S. The Expert System for Th 20. The MERLOT project. Available at: HYPERLINK "http://www.merlot.org" http://www.merlot.org . 21. Benedict M, Debowska, E, Jodl, HJ, et al. Report and Recommendation on Available he ; 22. M, Yap, C, Mannan, MA. A WWW-based Course Structure for Teaching a ucation. 23. Grove R. Internet-based expert systems. Expert Systems. December 2002; 17(3): 129-135. 25. lassroom via use of TEST™ Multimedia Material for Statistical and Thermal Physics. Paper presented at: Proc. of t 10th Workshop on Multimedia in Physics Teaching and Learning (EPS - MPTL 10), 2005 Berlin. Ridman Thermodynamics and heat Transfer Course. International Journal of Engineering Ed 2001; 17(17): 176-186. 24. Al-Nais MO, Aichouni, M. Web-Based Approach Applied To Technical Courses To Enhance Engineering Education Quality. Paper presented at: The Third Forum on Engineering Education, 2003; United Arab Emirates (UAE). Kumpatyal SK. Learning enhancement in Thermodynamics C software in design projects and laboratory. Paper presented at: The 2002 American Society for Engineering Education Annual Conference & Exposition, 2002; Montréal, Quebec, Canada. 26. Aichoouni M, Al-Nais, M. Interactive Demonstrations of Statistical Quality Control for on 27. s calculations for VxOy grown 28. Web Application for nted at: 29. 30. Moran MJ, Shapiro, HN. Fundamentals of Engineering Thermodynamics. 6 edition ed: 31. Sonntag RE, Borgnakke, C, Wylen, GJ. Fundamentals of Thermodynamics. 5th edition ed: 32. Urieli I. Chapter 11: Combustion. Engineering Thermodynamics - A Graphical Approach. apt.7_11/Chapter11.html" Engineering Students Using Computer Based Tools. Paper presented at: The Third Forum Engineering Education, 2003; United Arab Emirates (UAE). Kresse G, Surnev, S, Ramsey, MG, Netzer, FP. First-principle on Pd(1 1 1). Surface Science. October 2001; 492(3): 329-344. Paolini C, Bobba, K, Surana, P, Bhattacharjee, S. A Java Based Performing Chemical Equilibrium Analysis in Thermodynamics Courses. Paper prese 36th ASEE/IEEE Frontiers in Education Conference, October 28–31, 2006; San Diego, CA. Boles MA, Cengel, YA. Thermodynamics: an engineering approach. 6th edition ed. Boston: McGraw-Hill; 2008. Wiley; 2007. John Wiley & Sons; 1997. February 1, 2009. Available at: HYPERLINK "http://www.ent.ohiou.edu/~thermo/Applied/Ch http://www.ent.ohiou.edu/~thermo/Applied/Chapt.7_11/Chapter11.html . A 2, 2009. ccessed February 33. Spakovszky ZS. 16. Unified: Thermodynamics and Propulsion. Unified Engineering. 2009. /www/FALL/thermodynamics/notes/" Available at: HYPERLINK "http://web.mit.edu/16.unified http://web.mit.edu/16.unified/www/FALL/thermodynamics/notes/ . A 2009. ccessed February 2, 34. Bhattacharjee S, Paolini, CP. Property Evaluation in The Expert System for Thermodynamics 35. ction Equilibrium Analysis: Theory and Algorithms: 36. Van Zeggeren F, Storey, SH. The computation of chemical equilibria. London: Cambridge 37. enter. NASA Computer Program – Chemical Equilibrium with ("TEST") Web Application. Journal of Computer Coupling of Phase Diagrams and Thermochemistry (CALPHAD). 2008. Smith WR, Missen, RW. Chemical Rea John Wiley & Sons; 1982. University Press; 1970. NASA Glenn Research C Applications. Available at: HYPERLINK "http://www.grc.nasa.gov/WWW/CEAWeb/" http://www.grc.nasa.gov/WWW/CEAWeb/ . Accessed November 7, 2007. 
work_ewvk6lznwbfn3j56wv6u32zmjy	----	BlockU: Extended usage control in and for Blockchain O R I G I N A L A R T I C L E BlockU: Extended usage control in and for Blockchain Yasar Khan1 | Toqeer Ali2 | Megat Fariz1 | Fernando Moreira3 | Frederico Branco4,5 | José Martins4 | Ramiro Gonçalves4 1Malaysian Institute of Information Technology, University Kuala Lumpur, Kuala Lumpur, Malaysia 2Faculty of Computer and Information System, Islamic University of Madinah, Madinah, Saudi Arabia 3IJP, REMIT, Universidade Portucalense & IEETA, Universidade de Aveiro, Aveiro, Portugal 4CSIG, INESC TEC and University of Trás-os- Montes e Alto Douro, Vila Real, Portugal 5Departamento de Informática e Matemáica, Instituto Politécnico de Bragança, Bragança, Portugal Correspondence Frederico Branco, INESC TEC and University of Trás-os-Montes e Alto Douro, Quinta de Prados, 5001-801 Vila Real, Portugal. Email: fbranco@utad.pt Abstract An electronic business transaction among untrusted bodies without consulting a mutually trusted party has remained widely accepted problem. Blockchain resolves this problem by introducing peer-to-peer network with a consensus algorithm and trusted ledger. Blockchain originally introduced for cryptocurrency that came with proof-of-work consensus algorithm. Due to some performance issues, scientists brought concept of permissioned Blockchain. Hyperledger Fabric is a permissioned Blockchain targeting business-oriented problems for industry. It is designed for effi- cient transaction execution over Blockchain with pluggable consensus model; how- ever, there is limitation of rapid application development. Hyperledger introduced a new layer called Hyperledger Composer on top of the Fabric layer, which provides an abstract layer to model the business application readily and quickly. Composer pro- vides a smart contract to extend the functionality and flexibility of Fabric layer and provides a way of communication with other systems to meet business requirements. Hyperledger Composer uses role-based access control (RBAC) model to secure access to its valuable assets. However, RBAC is not enough because many business deals require continuous assets monitoring. Our proposed model, BlockU, covers all possible access control models required by a business. BlockU can monitor assets continuously during transactions and updates attributes accordingly. Moreover, we incorporate hooks in Hyperledger Composer to implement extended permission model that provides extensive permission management capability on an asset. Subse- quently, our proposed enhanced access control model is implemented with a minimal change to existing Composer code base and is backward compatible with the current security mechanism. K E Y W O R D S fabric, Hyperledger Composer, permissioned Blockchain, UCON 1 | INTRODUCTION Before Blockchain, performing a trusted transaction in an untrusted network was either not possible or it was only possible using complex methods such as remote attestation (Ismail, Syed, & Musa, 2014; Jaeger, Sailer, & Shankar, 2006). With the advent of Blockchain, it became very simple to perform a secure transaction between two unknown and untrusted systems. Blockchain was initially introduced in 2008 by Satoshi Nakamoto for Bitcoin transactions. Bitcoin provided a distributed ledger with no trusted or central party, offering secure Bitcoin transactions among peer-to-peer parties. The order of transactions is validated using a consensus algorithm, that is, a proof-of-work algorithm (Nakamoto, Received: 19 September 2018 Revised: 15 July 2019 Accepted: 15 November 2019 DOI: 10.1111/exsy.12507 Expert Systems. 2020;1–12. wileyonlinelibrary.com/journal/exsy © 2020 John Wiley & Sons, Ltd 1 https://orcid.org/0000-0002-5731-6650 https://orcid.org/0000-0001-8434-4887 https://orcid.org/0000-0002-7787-6305 https://orcid.org/0000-0001-8698-866X mailto:fbranco@utad.pt http://wileyonlinelibrary.com/journal/exsy http://crossmark.crossref.org/dialog/?doi=10.1111%2Fexsy.12507&domain=pdf&date_stamp=2020-01-20 2008). Blockchain has been useful in meeting the needs of today's businesses, and it has the potential to radically change future business applica- tions and models. According to a recent report by Gartner (2018), Blockchain's value may reach $3.1 trillion by 2030. Each transaction among two untrusted parties involves a trusted party, such as a bank or a certification authority. To perform a transaction between unknown parties, many blockchain-based frameworks have been proposed (Abdeen, Ali, Khan, & Yagoub, n.d.; Ali, Ali, Alsaawy, Khalid, & Musa, 2019; Crosby, Pattanayak, Verma, & Kalyanaraman, 2016; Khan, Zuhairi, Ali, Alghamdi, & Marmolejo-Saucedo, 2019; Kwon, 2014; Wood, 2014) and are categorized as public and private blockchain frameworks. One of them is the Hyperledger project supported by IBM and other renown indus- tries, for example, Intel, Cisco, and SAP. Hyperledger Fabric provides a decentralized operating system for next-generation business applications (Androulaki et al., 2018; Cachin, 2016). It overcomes performance issues that were a rather big concern in public blockchain systems. Hyperledger Fab- ric implements smart contracts called chaincodes among various distributed nodes; they contain the details of a contract, for example, the selling policy of a car, and are disclosed to all nodes that represent different stakeholders (e.g., buyer, seller, and manufacturer). These nodes or peers are called endorsers and are defined in the endorsement policy of the chaincode (Androulaki et al., 2018). Once the contract is finalized, only the transaction is stored on the distributed ledger. This process is conducted by the Orderer node in Hyperledger Fabric. The Orderer node maintains the order of trans- actions, for example, which transaction comes first and then second and so on in the ledger. The Orderer node in Hyperledger Fabric uses practical Byz- antine fault tolerance as a consensus algorithm instead of a proof-of-work algorithm, which is used in Bitcoin (Castro & Liskov, 1999). Hyperledger Fabric provides various APIs for application developers to design and develop their own applications based on their business networks (Cachin, 2016). However, writing applications in Hyperledger Fabric is to an extent different due to the provided APIs. In addition, although Hyperledger Fabric uses role-based access control (RBAC) security model, it has been shown in the past that RBAC cannot perform attribute updates or the fined-grained usage control of assets (Park & Sandhu, 2004). RBAC is a very popular security model that has shaped the security domain for decades. This model is mainly used in operating systems like Windows, Linux, and Solaris and databases like Oracle. There are many permissioned blockchains like Ethereum, Hyper- ledger Fabric, Swatooh, Iroha, and many more. However, to the best of our knowledge, all these frameworks are using traditional access control models such discretionary access control, mandatory access control model, and RBAC. We have opted for Hyperledger Fabric, introduced by IBM in 2017 (Dhillon, Metcalf, & Hooper, 2017), for this research because it is open source and is widely used in different business applications. Hyperledger Com- poser provides application development APIs very easily and rapidly. Hyperledger Composer also implements smart contracts based on clients, assets, users, and access control lists. In this paper, we proposed a detailed usage control model that is more effective than the existing RBAC model that is cur- rently implemented in Hyperledger Composer. Our newly proposed access control model has the capability to design a variety of business scenarios. In this research, we formally constructed a usage control model for Hyperledger Composer to clarify examples based on the model. In addi- tion, we described a detailed permission model that enables the members of a business network and the owner of assets to specify usage-based constraints while performing transactions. The new model is tailored specifically to Hyperledger Composer. The model has the capability to spec- ify various types of policies depending on business needs, for example, allowing a specific asset to be accessed only for a specific time, with a spe- cific computational power, or revoking an asset access from a user based on the occurrence of some defined condition. Finally, we implemented an extended usage control model on the smart contract implementation of Hyperledger Composer, which uses a blockchain to record every trans- action on the distributed ledger. 1.1 | Use case In a blockchain business network, many business deals rely on resource/asset usage control. Businesses share resources instead of buying them, and although based on the usage of assets, we can grant and revoke access to resources; this is called access decision continuity. Renting/ accessing a resource for a specific time is common, such as when renting a car or releasing health records. In the example of renting a car, multiple parties are involved including the owner of the car, the renter of the car, the bank, and vehicle-tracking entities. The assets involved in this trans- action are the car and the money. The renting authority rents the car for a specific time frame and zone the user can drive it in. The renter can order a car online, and the car will be reserved for them. The renter pays money to rent the car, and the asset's ownership is exchanged for a spe- cific duration. For example, a car may be rented for 24 hr, and the driver can only drive the car in New York City. The vehicle-tracking entities track the car closely using defined GPS coordinates. If the car exits the approved zone, alerts are sent to the owner and the renter. After a few warnings, if the car is still not in the approved range, the ownership of the car may be revoked. The car will be locked or tracked further by gov- ernment authorities. Similarly, if the car is not returned in the agreed amount of time, the owner can send alerts to the renter. The owner can also request the bank to deduct money from the renter's account. In this case, the bank can check the policy defined between the two parties and transfer money from the renter's account to the owner's account. If the renter account does not have enough money to comply, the bank can send a message to the vehicle-tracking system to lock the car unless the owner responds and pays the money. Once the money is paid, access can be granted again. This business model cannot be implemented in Hyperledger Composer, as an extended usage control model is needed. The rest of the paper is organized as follows: Section 2 elaborates the background of blockchain studies, and Section 3 explains the UCON model in detail, Section 4 presents the proposed model, whereas Section 5 explains its implementation, and Section 6 concludes the paper, whereas Section 7 presents future directions of the studies. 2 KHAN ET AL. 2 | BLOCKCHAIN BACKGROUND In this section, brief background studies are provided related to Blockchain, Hyperledger, and Hyperledger Fabric. Moreover, certain access con- trol types that involve usage access control are also elaborated. 2.1 | Blockchain Blockchain is a distributed ledger that was initially introduced for Bitcoin transactions. Each block has a block number that is generated using min- ing techniques, the details of the transaction, the previous block, and the newly generated block hash. Using a previous hash to generate a new block creates a chain of blocks known as a blockchain. There are two types of blockchains, that is, public and private blockchains. Using a blockchain, two untrusted networks can easily perform a transaction in a secure manner, as a blockchain can maintain a distributed ledger and record every transaction on the ledger nodes. A distributed ledger uses mining or a consensus protocol to solve double spending problems and malicious transactions. Once a block is created, any small change in the transaction or other details of the block are detected and further accepted by the rest of the nodes in the network. 2.2 | Hyperledger Hyperledger is a permissioned blockchain technology with various private blockchain projects. The one we opted for in these studies is Hyperledger Fabric. Hyperledger Fabric is a pluggable and generic smart contract framework that is implemented on top of blockchain to design any kind of application, including cryptocurrency applications (though it was not developed specifically for crypto- currency applications). Apart from Hyperledger Fabric, Hyperledger also offers Hyperledger Burrow, Hyperledger Sawtooth, Hyper- ledger Indy, and Hyperledger Iroha for various purposes. The following section further elaborates on the components of Hyperledger Fabric. 2.3 | Hyperledger Fabric Initially, blockchain was designed as a public ledger, which means that every transaction performed by the blockchain network should be visible to all nodes. However, cases exist in which two or more parties do not wish to disclose their transaction details but would still like to benefit from blockchain technology. Therefore, IBM started a generic private blockchain implementation known as Hyperledger Fabric. In a private network, a new participant cannot be added unless it receives an invitation to become part of the business network. 2.3.1 | Execute-order-validate architecture Hyperledger Fabric uses an execute-order-validate architecture (Androulaki et al., 2018), whereas traditional blockchains use an order-execute architecture. In an order-execute architecture, once a transaction is generated, it requires mining to generate a hash. In addition, it is generated by peers or minors. Once one of the peers can generate an acceptable hash, it disseminates it to the rest of other peer nodes. Further, every peer on the local system validates the block based on the value of its previous block hash. This order-execute architecture suffers from a few problems, such as sequential execution, nondeterministic code, confidential execution, fixed trust, and hard-coded consensus. To overcome these shortcom- ings, Hyperledger Fabric offers a permissioned, blockchain-backed smart contract that uses an execute-order-validate architecture with a chaincode, an enforcement policy, and a Hyperledger Fabric network. In Hyperledger Fabric, a chaincode implements the smart contract. The chaincode runs in the execution phase of Hyperledger Fabric's archi- tecture. Using the application, one can design their own logic in any programming language and perform a transaction with any other application in the network. Once a transaction is performed, it is stored on a distributed ledger. The endorsement policy can specify an endorser for a specific transaction. Hyperledger Fabric's use of an enforcement policy is a novel concept in permissioned-based smart contracts. In a public smart contract, there is no privacy, meaning that all transactions are visible to everyone on the network. In addition, a transaction details need to be specified exchanged between two parties to successfully execute a transaction. In contrast, smart contracts have endorsers that only enable the transaction to be visible to select parties. However, once the endorsers agree on certain parameters, a final transaction is propagated and validated based on the consensus algorithm. KHAN ET AL. 3 Every blockchain-based system involves a network of nodes with various roles, including Hyperledger Fabric. However, the nodes in the net- work have identities in permissioned blockchain-based systems, including client nodes, peer nodes, and ordering service nodes. Client nodes initi- ate transactions and send them further to the execution phase. Peer nodes perform certain tasks like maintaining the ledger and validating the transactions between endorsing nodes. Of these nodes, some endorsing nodes respond to the transaction proposals requested by the client. Ordering service nodes first complete the transactions of clients via the chaincode, and then they complete the transactions of the endorsers. Ordering service nodes handle the consensus and maintain the order by which transactions are updated on the ledger. 3 | UCON FOR HYPERLEDGER COMPOSER The first and most popular model for usage control is UCON, which was proposed by Park and Sandhu (2004) and was later formalized by Zhang et al. (2005). The following discussion is taken mostly from this formalization. This model is based on two core concepts: decision continuity and attribute mutability. Decision continuity refers to the concept that a decision to grant access is not a singular operation but rather a continuous action that is car- ried out in parallel to subjects' access. Once access has been granted, UCON's reference monitor tracks the usage and can perform several useful tasks based on it. For this purpose, the second aspect of UCON, attribute mutability, is of paramount value. Attribute mutability enables changes to a subject or object's attributes because of accessing or trying to access an object. UCON allows policy writers to define certain types of attribute mutability (Figure 1). UCON includes the following states: initial, requesting, denied, accessing, revoked, and end. The model puts each (s, o, r) triple in one of these states, where s is the subject, o is the object, and r is the right. Consider the following: • t1 = (s1, o1, r1) state(t1) = accessing. Subject s1 is currently exercising right r1 on object o1; the reference monitor tracks this usage as soon as s1 tries to access o1. This is the non- bypassability feature of the reference monitor, which ensures that access is mediated by the security policy enforced by the reference monitor. Immedi- ately after trying to access o1, the state of the triple is changed to requesting. The reference monitor can perform one of the following actions at this stage: • permitAccess: Allow s1 to access o1; in this case, the state changes to accessing. • denyAccess: Disallow access, which changes the state to denied. One of these two decisions must be made by the reference monitor, and they are, of course, mutually exclusive. The reference monitor can also perform the following action: • preUpdate: Update the attributes of the subject or the object. This is a pre-update operation and is highly useful for keeping track of how many times a subject has tried to access an object. Hence, repeated attempts by a subject to access a restricted object can be detected by the refer- ence monitor, after which the subject would be penalized for misbehaviour. If the subject is denied access, the session finishes, and the triple must return to the initial state before performing any restricted action. How- ever, if the state changes to accessing, the subject is free to access the object. At any point in time during access, the reference monitor might revoke the granted permission. The following four operations can potentially occur: FIGURE 1 UCON model states 4 KHAN ET AL. • endAccess: The subject voluntarily gives up the object. In this case, the reference monitor does not have to perform any task. • revokeAccess: Because of the usage or due to changes in the operating environment, the reference monitor revokes access. In this case, the subject is no longer able to access the object, and the state of the system changes to revoked. • onUpdates: During access, the reference monitor might update some attributes to keep track of the usage. This is a clean method for gauging usage. • postUpdates: The reference monitor can perform updates to a subject or object's attributes based on whether the subject voluntarily gave up the object or if it was taken back forcefully. In the following section, we formally define UCON's access and usage control. 4 | BLOCKU: EXTENSION OF USAGE CONTROL Usage control must not be limited to granting access to a right only once; instead, it should constantly keep track of the usage of objects and decide continuously whether to allow a subject access to objects or not. If a subject exceeds the allowed quota or if another constraint is violated during the use of an object, the subject's right should be revoked. This simple but powerful idea enables various policy specifications and thus extends the expressiveness of access control systems immensely. With this new capability, it is possible for an organization to specify a policy that dictates that an object can be accessed only within the premises of its offices and within a certain time frame. It also allows for the specification of constraints, such as allowed usage time per session and the revocation of usage after a specified limit. 4.1 | Model definition Hyperledger Composer's architecture is based on assets, participants, transactions, and conditions. According to our proposed usage control for Hyperledger Composer, we defined the models based on the new requirements. These models are extensions to Hyperledger Composer and can aid in the development of business applications based on blockchains. Definition 1: Participants and assets. We denote the set of participants in an active business network with P and the set of all assets within a transac- tion with A. Assets in a specific transaction can be accessed by the participant association function (ζ: A ! P). Each asset is associated with a unique participant. Definition 2: Permissions. We denote the set of all permissions with S. Permissions are the collection of labels that allow access to an asset. A permis- sion requires Ps: A ! A to provide access s ∈ S, which p ∈ P requires the user to have. Definition 3: Access association with assets. The access association function (ψ: P ! 2s) returns the set of permissions that belong or were given to par- ticipant p ∈ P at runtime. 4.2 | Extension to the core Composer model To define extended usage control permissions over assets, we introduce the concept of the participant state as well as updates to the operation, condition, and policies. The participant state decides which set of attributes belongs to the participant, including arbitrary characteristics about participants that are related to their permissions. Definition 4: Participant state t: γ (P) ! com (γ(P)) is a function that maps a participant's attributes to their values. This is a dictionary of attributes that is associated with each participant. γ is a function that provides the set of attributes that belongs to a participant, and the set of all possible values for an attribute (a) is denoted with com(a). These attributes can be updated via the attribute update function. KHAN ET AL. 5 4.2.1 | Pre-authorization models Before a participant accesses an asset in our Hyperledger Composer extension, we added a usage decision function. Definition 5: According to the above additional usage functions, our model has the following components: • P, A, S, ATT(S), ATT(A), befA (participants, assets, permissions, participant attributes, asset attributes, and asset attributes before authoriza- tion), and • granted (p,a,s) = befA((ATT(p)), ATT(a), S). Similarly, an additional step before an update is sometimes required, which is defined in our model as follows: Definition 6: befUpdate(ATT(P)) and befUpdate(ATT(a)) are optional operations to update the attributes of participants and assets, respectively. 4.2.2 | Ongoing-authorization model In the ongoing-authorization model, the composer requests asset usage without any pre-decision making. However, the authorization of assets is monitored constantly. In this model, each participant is initially allowed to use an asset; however, this model does remove their access if certain requirements are not met. In Hyperledger Composer, such a model is very useful when assets are accessed for a long duration of time. This func- tionality can be further extended into four different models: • the A0 model (no pre-updates) • the A1 model (with pre-updates) • the A2 model (with continuous updates) • the A3 model (with post-updates) Definition 7: In the A0 model, P, A, S, ATT(S), and ATT(A) remain the same as in the pre-authorization model. • onA (continuous authorization) • allowed (p,a,s) ) true • stopped (p,a,s) ( ¬ conA(ATT(p),ATT(a),S) Definition 8: The A1 model is like the A0 model except it updates using the following pre-update processes: preUpdate(ATT(p)) and onUpdate(ATT(a)), which are optional procedures to perform update operations on ATT(p) and ATT(a), respectively. Definition 9: The A2 model is also similar to the A0 model except it adds the following continuous-update processes: onUpdate(ATT(p)) and onUpdate(ATT(a)), which are optional procedures to perform update operations on ATT(p) and ATT(a), respectively. Definition 10: The A3 model adds the following post-update processes to the A0 model: postUpdate(ATT(s)) and postUpdate(ATT(o)), which are optional procedures to perform update operations on ATT(s) and ATT(o), respectively. 4.2.3 | Pre-obligation models Hyperledger Composer introduced a new obligation feature in its architecture. Using this feature allows a participant to access an asset if and only if they provide information during the authentication process. Hyperledger Composer verifies this information using participants. However, the existing architecture of Hyperledger Composer only supports first-time authentication. Our model can check this information continuously. 6 KHAN ET AL. Definition 11: The pre-obligation model has the following components: • P,A,S,ATT(P) and ATT(A) are the same as in the pre-authorization model • OBS, OBO, and OB (obligation subjects, obligation objects, and obligation actions, respectively) • preB and preOBL (pre-obligation predicates and pre-obligation elements, respectively) • preOBL ⊆ OBS × OBO × OB • preFulfilled: OBS × OBO × OB ! {true, false} • getPreOBL: P × A × S ! 2preOBL, a function to select pre-obligations for a requested usage • preB(P,A,S) = ^ (obsi, oboi, obi) ∈ getPreOBL(P,A,S) preFulfilled(obsi, oboi, obi) • preB(P,A,S) = true by definition if getPreOBL(P,A,S) = φ • allowed(P,A,S) ) preB(P,A,S). Definition 12: The preB1 model is identical to the befB0 model except it adds the following pre-update processes: preUpdate(ATT(P)) and preUpdate(ATT(a)), which are optional procedures to change certain attributes because of pre-obligations. Definition 13: The preB3 model is identical to the befB0 model except it adds the following post-update processes: • postUpdate(ATT(p)) and postUpdate(ATT(a)), which are optional procedures to change certain attributes because of pre-obligations. 4.2.4 | Continuous-obligation models The continuous-obligation model is like the pre-obligation model except it requires the verification of obligations periodically or con- stantly. To implement this model, we introduce a time parameter (T), which depends on a time- or element-based period; for example, we can ask a participant to submit information every hour or to provide information after a specific amount of resource usage. Based on mutability issues, the continuous-obligation model can implement the following four models: the B0 model, which includes an ongoing- obligation predicate instead of a pre-obligation predicate; the B1 model; the B2 model; and the B3 model. The B1, B2, and B3 models are the same as the B0 model except that they add pre-updates, ongoing updates, and post-updates, respectively. Definition 14: The B0 model has the following components: • P,A,S, ATT(S), ATT(A), OBS, OBO, and OB are the same as in the befB model • T, which represents a set of time or event elements • onB and onOBL, which represent ongoing obligation predicates and ongoing obligation elements, respectively • onOBL ⊆ OBS × OBO × OB × T • getOnOBL: P,A,S ! 2OnBL, which is a function to select ongoing obligations for a requested usage • onFulfilled: OBS × OBO × OB × T ! true, false • onB(P,A,S) = ^(obsi, oboi, obi,ti) ∈ getOnOBL (P,A,S) onFulfilled(obsi, oboi, obi, ti) • onB (P,A,S) = true by definition if getOnOBL (P,A,S) = θ • allowed(P,A,S) ) true • stopped(P,A,S) ( ¬onB(P,A,S). Definition 15: The B1 model is identical to the B0 model except it adds the following pre-update processes: preUpdate(ATT(p)) and preUpdate(ATT(a)), which are optional procedures to change certain attributes as a result of pre-obligations. Definition 16: The B2 model is identical to the B0 model except it adds the following ongoing update processes: onUpdate(ATT(p)) and onUpdate(ATT (a)), which are optional procedures to change certain attributes as a result of pre-obligations. KHAN ET AL. 7 Definition 17: The B3 model is identical to the B0 model except it adds the following post-update processes: postUpdate(ATT(p)) and postUpdate(ATT (a)), which are optional procedures to change certain attributes as a result of pre-obligations. 4.2.5 | Pre-condition model Some conditions define environmental variables through which we can enforce certain policies; for example, employees of an organization can use its resources only when they are in the premises of the organization. Such restrictions are not directly associated with any assets in Hyperledger Com- poser. In the pre-condition model, such conditions are checked each time the assets are accessed. The pre-condition model introduces the pre-condition predicate (preC), which must be evaluated before the requested rights are exercised. The following definitions formalize the preC model. Definition 18: The preC model has the following components: • P,A,S, ATT(S), and ATT(O) are the same as in the preA model • preCON (a set of pre-condition elements) • getPreCON: P × A × S ! 2 preCON • preConChecked: preCON ! true, false • preC(P,A,S) = ^preConi∈getPreCON(P,A,S)preConChecked(preConi) • allowed(P,A,S) ) preC(P,A,S). 4.2.6 | Ongoing-conditions model In many cases, environmental restrictions must be satisfied while rights are in active use. This can be supported using the onC0 model. In the onC0 model, usage is allowed without any decision process at the time of the request. However, there is an ongoing condition predicate to check certain environmental statuses repeatedly throughout usage. As mentioned earlier, the onC0 model is intrinsically immutable. The following defi- nitions formalize the onC0 model. Definition 19: The onC0 model has the following components: • P,A,S, ATT(S), and ATT(A) are the same as in the preA model • onCON (a set of ongoing conditions) • getOnCON: P × A × S ! 2 onCON; onConChecked: onCON ! true, false • onC(P,A,S) = onConi∈getOnCON(P,A,S) onConChecked(onCon i) • allowed(P,A,S) ) true • stopped(P,A,S) ( ¬ onC(P,A,S). Herein, we extended the core Hyperledger Composer model with the above additional models. This enriches future applications with a vari- ety of permissions based on business needs. 5 | UCON INCORPORATION IN HYPERLEDGER FABRIC Hyperledger Composer is a high-level layer in Hyperledger Fabric's architecture. Hyperledger Composer uses a generic chaincode to communicate with Hyperledger Fabric. Hyperledger Composer provides some extra layers and features that are not directly part of Hyperledger Fabric. To implement our proposed model, we extended Hyperledger Composer with extra features and introduced new transaction calls to be saved on the blockchain's trusted ledger and state database. To achieve this, we added hooks in the current JavaScript transaction processor, connecting us to the extra models we defined earlier. Similarly, we introduced automatic blockchain transactions that are stored on the blockchain. Whenever an asset or participant is added to the system using Hyperledger Composer's API, our proposed model automatically creates different transactions that are stored on the blockchain and state database. The purpose of these transactions is to monitor the activity of the participants while they are accessing and using any asset. Our model automatically responds to the global configuration that is set to monitor asset usage. 8 KHAN ET AL. To incorporate the extended version of the access control, we updated three sections of the Composer as depicted in Figure 2. • Access control: In ACL module, we introduced new tags that support our UCON policies. These tags are declared in the development phase of any BNA. • Composer runtime: The second module that affected our proposed solution is the Composer runtime. Composer runtime is basically connected with Fabric, which is the backbone of the blockchain infrastructure. The newly defined access control models are linked with the Composer runtime, so it can be hooked to the application execution. • JavaScript transactions processor: In the third module of UCON incorporation, we created a BlockU reference monitor that continuously moni- tors the assets and records all transactions in the blockchain against a participant. An asset can have multiple attributes, and we can define sep- arate transactions for each of the attributes; similarly, we can create various BlockU rule for each attribute (for instance, a vehicle asset that contains attributes, such as millage, price, manufacture, and current state). Our newly introduced functions can retrieve an attribute—for exam- ple, millage—via the transaction asset vehicle millage. There are two types of restrictions on the usage of an asset. One constraint is developed in the business logic of the application; for example, a person cannot buy an asset if his payment is less than the product amount; such kind of constraints are also handled in the chaincode transaction functions. rule SampleConditionalRule { description: "Description of the ACL rule" participant(m): "org.example.SampleParticipant" operation: ALL //CREATE, UPDATE, DELETE or ALL resource(v): "org.example.SampleAsset" condition: (v.owner.getIdentifier() == m.getIdentifier()) action: ALLOW // ALLOW, DENY // BlockU Tags blockU: onOBL //befA, onA, befOBL, onOBL, preCON, onCON blockUcondition: (v.blockUReferenceMonitor.usage(m.getIdentifier(), DATETIME) < 10) } FIGURE 2 Proposed usage control framework for permissioned Blockchain KHAN ET AL. 9 Our proposed model verifies each request against the generated policy before allowing or denying participants to use the assets (Figure 3). The model adds this failure request as a new transaction on the blockchain that can be seen by a different network and the system administrator who can monitor and respond to participants as and when needed. It also helps to resolve business disputes. This detailed permission model enables members of a business network and owners of assets to specify usage-based constraints while performing transactions. The new model is tailored to Hyperledger Composer. As mentioned earlier, attribute mutability includes three states in the UCON engine, that is, the first “pre-update,” the sec- ond “upon update,” and the third “post-update.” Pre-update attributes are set before accessing an asset. In our use case, we set the start time to zero initially, ensuring that after 24 hr, the asset would be revoked. Once an asset is moved to the accessing state, the mutable attribute is updated accordingly. We also set the warnings to zero to ensure that once the number of warnings reaches the allotted time, the access would be revoked. Each state change of UCON is stored in a transparent ledger, in case of a dispute or if the bank would like to deduct money from the renter. Our sec- ond condition involves constantly monitoring the defined zone. In the case of the car rental, we set the current coordinates of the car as a pre- update attribute to constantly monitor the car's location. In addition, we set the warning attribute. In case the car travels out of range, the user receives five warnings. If the user does not respond to these warnings, access to the car is revoked, and all transactions are locked. 6 | CONCLUSION In the recent past, permissioned Blockchain got acceptance due to the various issues in public Blockchain, such as performance, security, and other business requirements. Among these Blockchain frameworks, the Hyperledger project gained popularity due to its modular, open source architecture and huge industry support. The research done can be adopted by any permissioned Blockchain framework; however, the proof of concept is demonstrated in Hyperledger Composer. The study contributes in extending the permissions model of Hyperledger Composer by intro- ducing BlockU. BlockU covers various business applications that need continuous access to their assets by a participant. Many business use cases cannot be implemented using the RBAC model. Usage control models are difficult to implement; thus, extending permissions are necessary to meet the needs of current and future business applications. The proposed model allows developers and end users to define their own access poli- cies for their shared assets rather than using fixed policies developed and instantiated at the start of the application. They can also define asset's FIGURE 3 Extended usage control model for Blockchain 10 KHAN ET AL. access rules violation policy and set environmental constraints to the usage of their assets. Such violation policies can trigger consequent responses, which helps organizations with real-time monitoring and tracking the participant's activities. 7 | FUTURE WORK To identify future work, further investigation can be made to improve the policy execution engine and to deploy it on the underlying Fabric archi- tecture. Similarly, we aim to develop a new blockchain history track for our proposed model policies, which will be separated from the application level chaincode. A new system can be introduced, that is, system chaincode in the fabric architecture; it will have its own private data section where the proposed policies shall be stored, and its core functionality will be the execution engine logic. Similarly, a high-level language interface is required for the user to write their policies, which ultimately can be transformed to the core Hyperledger Composer engine. CONFLICT OF INTERESTS None. ORCID Toqeer Ali https://orcid.org/0000-0002-5731-6650 Frederico Branco https://orcid.org/0000-0001-8434-4887 José Martins https://orcid.org/0000-0002-7787-6305 Ramiro Gonçalves https://orcid.org/0000-0001-8698-866X REFERENCES Abdeen, M. A. R., Ali, T., Khan, Y., & Yagoub, M. C. E. (n.d.). Fusing identity management, HL7 and Blockchain into a global healthcare record sharing architecture. Ali, J., Ali, T., Alsaawy, Y., Khalid, A. S., & Musa, S. (2019). Blockchain-based smart-IoT trust zone measurement architecture. Proceedings of the International Conference on Omni-Layer Intelligent Systems, 152–157. Androulaki, E., Barger, A., Bortnikov, V., Cachin, C., Christidis, K., De Caro, A., … Muralidharan, S. (2018). Hyperledger fabric: A distributed operating system for permissioned blockchains. Proceedings of the Thirteenth EuroSys Conference, 30. Cachin, C. (2016). Architecture of the hyperledger blockchain fabric. Workshop on Distributed Cryptocurrencies and Consensus Ledgers, 310, 4. Castro, M., & Liskov, B. (1999). Practical Byzantine fault tolerance. OSDI, 99, 173–186. Crosby, M., Pattanayak, P., Verma, S., & Kalyanaraman, V. (2016). Blockchain technology: Beyond bitcoin. Applied Innovation, 2(6–10), 71. Dhillon, V., Metcalf, D., & Hooper, M. (2017). The hyperledger project. In Blockchain enabled applications (pp. 139–149). Berkeley, CA: Springer. Gartner. (2018). Gartner says global artificial intelligence business value to reach $1.2 trillion in 2018. Retrieved from https://www.gartner.com/en/ newsroom/press-releases/2018-04-25-gartner-says-global-artificial-intelligence-business-value-to-reach-1-point-2-trillion-in-2018 Ismail, R., Syed, T. A., & Musa, S. (2014). Design and implementation of an efficient framework for behaviour attestation using n-call slides. Paper presented at the Proceedings of the 8th International Conference on Ubiquitous Information Management and Communication, p. 36. Jaeger, T., Sailer, R., & Shankar, U. (2006). PRIMA: Policy-reduced integrity measurement architecture. Paper presented at the Proceedings of the Eleventh ACM Symposium on Access Control Models and Technologies, pp. 19–28. Khan, M. Y., Zuhairi, M. F., Ali, T., Alghamdi, T., & Marmolejo-Saucedo, J. A. (2019). An extended access control model for permissioned blockchain frame- works. Wireless Networks, 25, 1–12. Kwon, J. (2014). Tendermint: Consensus without mining. Draft v. 0.6, Fall, 1, 11. Nakamoto, S. (2008). Bitcoin: A peer-to-peer electronic cash system. Park, J., & Sandhu, R. (2004). The UCON ABC usage control model. ACM Transactions on Information and System Security (TISSEC), 7(1), 128–174. Wood, G. (2014). Ethereum: A secure decentralised generalised transaction ledger. Ethereum Project Yellow Paper, 151(2014), 1–32. Zhang, X., Parisi-Presicce, F., Sandhu, R., & Park, J. (2005). Formal model and policy specification of usage control. ACM Transactions on Information and Sys- tem Security (TISSEC), 8(4), 351–387. AUTHOR BIOGRAPHIES Yasar Khan received an M.S. degree in Computer Sciences from the Institute of Management Sciences, Peshawar, Pakistan, in 2016. After completion of his Master Degree in Computer Sciences in 2008, he joined Security Engineering Research Group, Institute of Management Sci- ences, Peshawar in 2009 as Android System and Architecture Developer,where he worked on many areas of Security Engineering such as Integrity Measurement Architecture, Android Permission Extension, Usage Control Engine and on Cross-Domain Access Control Manage- ment. He is currently doing his Ph.D. in Information Technology at University of Kuala Lumpur, Malaysia. His current research interests include Blockchain, IoT, malware analysis, machine learning and Cross Domain Access Control Management. KHAN ET AL. 11 https://orcid.org/0000-0002-5731-6650 https://orcid.org/0000-0002-5731-6650 https://orcid.org/0000-0001-8434-4887 https://orcid.org/0000-0001-8434-4887 https://orcid.org/0000-0002-7787-6305 https://orcid.org/0000-0002-7787-6305 https://orcid.org/0000-0001-8698-866X https://orcid.org/0000-0001-8698-866X https://www.gartner.com/en/newsroom/press-releases/2018-04-25-gartner-says-global-artificial-intelligence-business-value-to-reach-1-point-2-trillion-in-2018 https://www.gartner.com/en/newsroom/press-releases/2018-04-25-gartner-says-global-artificial-intelligence-business-value-to-reach-1-point-2-trillion-in-2018 Toqeer Ali Syed is currently working as an Associate Professor in Islamic University of Madinah, KSA. He also served for two years as a Senior Lecture at University Kuala Lumpur. He is a professional, teacher and a researcher working in the field of System Security, Deep Learning, Blockchain Applications and Cloud Computing. He has been actively involved in research and teaching activities since last 10 years. He received a PhD degree in Engineering Technology (IT) from University Kuala Lumpur in 2014. He completed his Masters in Computer Sci- ences. He has published his research findings in several forums of international repute. Megat F. Zuhairi is an Associate Professor at University Kuala Lumpur, Malaysian Institute of Information Technology where he has been since 2004. He currently serves as Deputy Dean Student Development and Campus Lifestyle. He was an Engineer at Marconi (M) Bhd., where he worked 2 years after he completed his BEng (Hons) in Electrical & Electronics from UNITEN. Generally, he teaches computer network sub- jects but every other semesters he also teaches electrical and electronics. He was a Cisco Certified Network Associate (CCNA) between 2007 and 2009 and now serves as Instructor for various courses for Cisco Network Academy. His research interests lie in the area of computer data networking, and wireless mobile ad hoc communications. Over the span of 4 years he has published more than 25 journals and conference publications including 2 books. In recent years, he has focused on wireless sensor network and data acquisition and analysis within the con- text of Internet of Things. He has served on roughly 15 conference programme committees as reviewer and technical members. Fernando Moreira is the Director of the Science and Technology Department. He is graduated in Computer Science (1992), M.Sc. in Elec- tronic Engineering (1997) and PhD in Electronic Engineering (2003), both at Faculty of Engineering of the University of Porto, and Habilitation (2018). He is a member of the Science and Technology Department at Portucalense University since 1992, currently as Full Professor and a visiting professor at the University of Porto Business School. He teaches subjects related to undergraduate and post-graduate studies. He supervises several PhD and M.Sc. students. He is a (co-)author more than 150 scientific publications with peer-review on national and interna- tional journals and conferences. He serves as a member of the Editorial Advisory Board for several journals and books. He organized several special issues from JCR journals. He has already regularly served as a member of Programme and Scientific committees of national and inter- national conferences. He was the MSc in Computation coordinator during the last ten years. He holds editorial experience, and he is co-editor of several books. He is associated with NSTICC, ACM and IEEE. His principal research areas are mobile computing, ICT in Higher Education, Mobile learning, Social business and Digital transformation. He was awarded Atlas Elsevier Award, April 2019. Frederico Branco is Assistant Professor at the University of Trás-os-Montes and Alto Douro and Senior Research of INESC TEC research cen- ter. He has published over 70 articles in journals and event proceedings. He is also involved in several academic works, as dissertation and thesis supervisor and degree projects responsible, and is continuously participating in several research projects. His professional career is also directly related with the industry, with particular focus in various planning and implementation projects of Information Systems, with particu- lar attention on agri-food and services sectors. Currently holds several functions of senior management in the areas of Operations,Information Systems and Quality Management. José Martins is currently an Invited Assistant Professor at the University of Trás-os-Montes e Alto Douro (Department of Engineering), Invited Assistant at the Polytechnic Institute of Bragança and Senior Researcher at INESC TEC research center. He has published over 90 arti- cles in indexed journals and event proceedings focusing the Information Systems, Management Information Systems, Software Engineering and Human-Computer Interaction topics. Currently he is supervisor for several Master Degree dissertations and PhD thesis. During his research career, he has participated in several research projects and is currently a member of various research projects aimed at merging infor- mation systems and technologies with other fields of study. Throughout his professional career José has also worked as an information sys- tems and technologies senior consultant where he directly participated in several international projects. At the present time José Martins dedicates most of his time to his lectures and to his research activities where he tries to understand the variables and (in)direct impacts of ICT adoption at individual and firm levels, and how IS solutions can be idealized, specified and developed in order to fully address their audience needs and requirements. Ramiro Gonçalves is an Associate Professor with Habilitation at University of Trás-os-Montes e Alto Douro, in Vila Real, Portugal, and a Senior Researcher at INESC TEC Associated Laboratory, Porto, Portugal. Ramiro is currently the Executive Director of the PhD degree in Informatics and has around 200 publications (including book chapters, Scientific Citation Index journal articles, as well as publications in refer- eed conference proceedings). How to cite this article: Khan Y, Ali T, Fariz M, et al. BlockU: Extended usage control in and for Blockchain. Expert Systems. 2020;1–12. https://doi.org/10.1111/exsy.12507 12 KHAN ET AL. https://doi.org/10.1111/exsy.12507 BlockU: Extended usage control in and for Blockchain 1 INTRODUCTION 1.1 Use case 2 BLOCKCHAIN BACKGROUND 2.1 Blockchain 2.2 Hyperledger 2.3 Hyperledger Fabric 2.3.1 Execute-order-validate architecture 3 UCON FOR HYPERLEDGER COMPOSER 4 BLOCKU: EXTENSION OF USAGE CONTROL 4.1 Model definition 4.2 Extension to the core Composer model 4.2.1 Pre-authorization models 4.2.2 Ongoing-authorization model 4.2.3 Pre-obligation models 4.2.4 Continuous-obligation models 4.2.5 Pre-condition model 4.2.6 Ongoing-conditions model 5 UCON INCORPORATION IN HYPERLEDGER FABRIC 6 CONCLUSION 7 FUTURE WORK CONFLICT OF INTERESTS REFERENCES 
work_exsdmxannvandeuxlialp6kkxq	----	PII: 0198-9715(91)90007-Z A decision support and expert system for retail planning Citation for published version (APA): Borgers, A. W. J., & Timmermans, H. J. P. (1991). A decision support and expert system for retail planning. Computers, Environment and Urban Systems, 15(3), 179-188. https://doi.org/10.1016/0198-9715(91)90007-Z DOI: 10.1016/0198-9715(91)90007-Z Document status and date: Published: 01/01/1991 Document Version: Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers) Please check the document version of this publication: • A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website. • The final author version and the galley proof are versions of the publication after peer review. • The final published version features the final layout of the paper including the volume, issue and page numbers. Link to publication General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal. If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement: www.tue.nl/taverne Take down policy If you believe that this document breaches copyright please contact us at: openaccess@tue.nl providing details and we will investigate your claim. Download date: 06. Apr. 2021 https://doi.org/10.1016/0198-9715(91)90007-Z https://doi.org/10.1016/0198-9715(91)90007-Z https://research.tue.nl/en/publications/a-decision-support-and-expert-system-for-retail-planning(21b313cb-2a5e-47b3-ba04-50e0bf304fe8).html Cornput.. Environ. and Urban Systems, Vol. 15, pp. 179-lWl991 Printed in the USA. All rights reserved. 0198-9715191 $3.00 + .oo Copyright 0 1991 Pergamon Press plc A DECISION SUPPORT AND EXPERT SYSTEM FOR RETAIL PLANNING Aloys Borgers and Harry Timmermans Urban Planning Group Department of Architecture and Urban Planning, Eindhoven University of Technology ABSTRACT. An important issue in Dutch retail planning practice is prediction of the likely impacts of new retail developments on consumer demand and assessment of the feasibility of new projects. In practice, a planning team has to decide which retail plan out of a number of alternative plans has to be implemented. A decision support system for retail planning can facilitate this decision. The decision support system we are proposing consists of three modules. The first module is used to manage the required data. The second module is used to model consumer choice behaviour. This module also contains the program GIANT, which can be considered as an expert system to assist the analyst in constructing and analysing conjoint-based simulator systems. The third module summarizes and evaluates the predicted impacts of the alternative retail plans. 1. AIMS AND NATURE OF RETAIL PLANNING IN THE NETHERLANDS Physical planning has traditionally played an important role in Dutch society. The lack of space, rapid urbanization and the open nature of its society have increased the need to control land development very carefully. Over the years a detailed set of plans and regulations has been approved which regulates future land use. Such plans require the approval of democrati- cally elected councils while the public can also participate in the planning process at various stages. Consequently, there has been a strong tendency during the last 20 years to inform plan- ning proposals with the results of scientific analyses. For example, Dutch law states that a gen- eral survey should be conducted before municipalities can start the preparation of their plans. In practice, the importance of scientific research in the planning process differs considerably between planning sectors. Retail planning is one sector of urban and regional planning with a traditionally strong input of research in the planning and design process. Until a few years ago, municipalities were obliged by law to conduct some kind of retail planning research whenever a new plan was developed. Although this situation has changed somewhat, and this law no longer exists due to the economic crisis and consequent deregulation, every-day practice still Requests for reprints should be sent to Aloys Borgers, Department of Architecture and Urban Planning, Eindhoven University of Technology, P.O. Box 513,560 MB Eindhoven, The Netherlands. 179 780 A. Borgefs and t-i. Timmarmans engages in a considerable amount of applied research, especially when the envisaged changes in the retail structure are substantial. Perhaps the most important issue in this respect is to predict the likely impacts of new retail developments on consumer demand and assess the feasibility of new projects. The results of such studies are also used to decide on the amount of space and location required by new retail facilities. Traditionally, such decisions were made on the basis of a set of normative ratios between the amount of retail space and the number of inhabiter of the plan area, adjusted according to the type of residential zone involved. It soon became evident, however, that such simple procedures were inadequate to justify the retail plans. Especially the larger scale devel- opments and the peripheral shopping centres had impacts that were beyond previous experi- ence. Starting in the late 197Os, these developments stimulated the use of more complex research strategies to assess the viability of new retail developments on the one hand and the likely impacts of these new developments on existing shopping centres on the other hand. It resulted in an official government report and a series of guidelines for the Dutch provinces and municip~i~es, which suggested in a very detailed manner the kind of research these planning agencies had to conduct to substantiate their decision making. The research effort typically involves a large-scale in-home interview to determine consumer orientation patterns for different types of goods, an inventory of available floor space in the various shopping centres, disaggregated according to type of firm, and an interview with the entrepreneurs. Recently, information about the supply side of the retailing system has been stored in a national geographical information system, which covers most of the country and is developed specifically for retail planning purposes. The consumer interview is used to con- struct mathematics models of consumer shopping behaviour which predict the likely impacts of policy decisions related to building programs, transportation and retail developments, on the level of retail turnover in the various shopping centres of the study area. The assessment of these effects is typically accomplished by comparing the predicted retail turnover per square foot of floor space with a normative planning standard, derived from a national survey, some- times adjusted to local circumstances. If changes in retail structure are predicted not to have substantial negative impacts on the level of turnover per square foot floor space, that is, the predicted level of turnover is not lower than the norm, it is generally concluded that the pro- posed changes in retail structure are justified. 2. MODELLING CONSUMER CHOICE BEHAVIOUR To predict the likely effects of policy measures on the retail system, the effects of those mea- sures on consumer choice behaviour need to be predicted. For this purpose, it is common prac- tice to construct a mathematical model of consumer choice behaviour. In such a model, con- sumer choice ~haviour is explicitly related to a number of features or attributes of the shop- ping centres, including the distance between residential and shopping centre location. After deciding which type of consumer model will be used and which attributes are to be incorporat- ed in the consumer model (model specification), the exact relationships between choice behaviour and attributes have to be determined (model calibration). Model calibration finally results in a number of parameters which can be considered as weights for the selected attributes. Next, the performance of the model has to be investigated. Only if there is a good fit between observed and predicted choice behaviour, can the model be used to predict consumer choice behaviour at future moments in time, conditional upon the assumption that consumer choice behaviour is stable over time. To predict the likely effects of policy measures on con- sumer behaviour, the policy measures have to be translated in terms of the attributes of the shopping centres used in the model. Using the adjusted attribute scores, consumer choice behaviour under new circumstances can be predicted. A Decision Support and Expert System for Retail Planning 181 In the context of Dutch retail planning, the models used are generally based on observed consumer choice behaviour. In several regions, the conventional spatial interaction model has been used. Further, the multinomial logit model has found ample application. Recently, indi- vidually based choice simulator systems based on conjoint measurements have been devel- oped. In contrast to spatial interaction and random utility models which are derived from observed behaviour, the simulator system is based on individuals’ expressed or intended behaviour given a number of hypothetical choice alternatives. 2.1 Models Based on Observed Choice Behaviour Spatial interaction models (Wilson, 1971) are based on observed flows of consumers between residential zones and shopping centres. In practice, the origin-constrained version of this model is used to predict the distribution of consumers over the alternative shopping cen- tres, given their residential zone (see, e.g., Roy & Anderson, 1988). According to this model, the number of consumers visiting a particular shopping centre depends on the (relative) attrac- tiveness of the shopping centre and the (relative) distance between the residential zone and shopping centre. The attractiveness of shopping centres is generally represented by some surro- gate variable such as the amount of floor space or number of outlets per shopping centre. Sometimes, the model is generalized to incorporate more explanatory variables in the attrac- tiveness component. In contrast to spatial interaction models which are based on aggregate flows of consumers, random utility models (e.g., Ben-Akiva & Lerman, 1985) are based on the observed choice behaviour of individual consumers. The well-known multinomial logit model (McFadden, 1974) is the most frequently used model which can be derived from random utility theory. According to this model, the probability that a consumer chooses a particular shopping centre depends on the relative utility of that centre. The utilities of the shopping centres are assumed to be individual-specific and depend on characteristics of the shopping centres as perceived by each individual. Further, random utility models take into account individual choice sets: only those shopping centres a consumer is familiar with can be chosen by that consumer. From a theoretical point of view, a random utility model is preferable to a spatial interaction model. In practice, however, predicting the likely effects of policy measures using a random utility model requires the specification of two additional submodels. First, a submodel to con- struct individual choice sets is required. To determine the set of available shopping centres for each individual consumer, logistic regression analysis may be used (see, e.g., van der Heijden & Timmermans, 1984). The second submodel is necessary to relate attribute scores as per- ceived by the individual to objective characteristics of the shopping centres. Curve fitting pro- cedures may be used to determine such relationships. 2.2 Models Based on Expressed or Intended Choice Behaviour In a transportation planning context, this approach has become known as the stated prefer- ence approach. The approach has gained increasing popularity in retailing, marketing, recre- ation, transportation, housing, migration, and other activity based studies (see, e.g., Golledge & Timmermans, 1988). The model system involves generating hypothetical shopping centres by combining shopping centre characteristics, some of which may be beyond the domain of expe- rience, and then requesting consumers to express some measure of preference for the resulting set of shopping opportunities. Individual preference structures may then be derived from these data and choices may be simulated by implementing some decision rule, for example, that a consumer will choose the alternative which received the highest preference. Since these prefer- 18.2 A. Borgers and H. Timmermans ence structures relate to hypothetical shopping centres, the effects of retail change on individu- al choice behaviour and turnover levels of the shopping centres in the study area may be pre- dicted by generating new alternatives; these reflect the changes and enable choice behaviour to be simulated by applying the derived preference structures and decision rules to the resulting choice sets. More specifically, the approach involves the following steps (see Timmermans, 1984). First, some measurement model is specified. Next, the hypothetical choice alternatives are generated. This step involves identifying the attributes which are considered important in influencing the choice behaviour of interest, defining these attributes in terms of attribute levels, and then com- bining these into hypothetical choice alternatives according to some experimental design. The choice of design is largely dictated by the chosen measurement model because complex mea- surement models cannot be estimated on the basis of simple experimental designs. Subsequently, subjects are requested to express their overall evaluation or preference for the resulting set of choice alternatives. These preference measures are then decomposed into the so-called part-worth utilities associated with the various attribute levels, given the a priori specified measurement model. The validity of the decomposition may be assessed by calculat- ing some goodness of fit measure. Having derived the preference structures, the next step involves specifying a decision rule which predicts an individual’s choice behaviour given the position of the choice alternatives on his or her subjective preference scale. In general, different types of decision rules may be considered. The first rule states that an individual will invari- ably choose the alternative included in his choice set which received the highest preference score. The second rule is not deterministic and states that choice probabilities are systematical- ly related to preferences. The simulation works as follows. First, the shopping centres are described in terms of the attributes included in the preference experiment. Next, for each individual separately, a choice set is identified using some submodel. The consumer’s preference score for each shopping cen- tre is then calculated by applying his part-worth preference function to the description of the shopping centres in terms of attribute levels. Given these preference scores, actual behaviour is simulated by implementing the decision rule. These predicted individual choices are then aggregated to yield an estimate of aggregate retail turnover. Impacts are simulated by defining retail plans in terms of the attribute levels included in the conjoint measurement model. This results in adjusted attribute levels of the shopping centres involved and hence in changes in preferences and subsequent choice behaviour. This process is repeated iteratively for some fixed time horizon. Recently, Louviere and Woodworth (1983) developed an alternative approach. Basically, their approach involves measuring choices directly rather than constructing choices from pref- erence ratings. Essentially, this approach amounts to estimating random utility models from experimental design data, that is, choices among hypothetical alternatives. These choice designs differ from preference designs in that the hypothetical choice alternatives should now be placed into choice sets. This can be done in a number of ways (Louviere & Timmermans, 1987). Because one now deals with choices per se rather than preferences, the estimated choice model can be used directly to simulate consumer choice behaviour. 3. A DECISION SUPPORT SYSTEM FOR RETAIL PLANNING The application of choice models to problems of retail planning commonly involves research institutes or universities in collecting the data, estimating the model and running a fixed num- ber of scenarios. Retail planning processes usually involve a number of representatives of vari- ous parties (local government, local retailers, national chains, Ministry of Economic Affairs, consumer groups, etc.). In practice, this planning group proposes a number of scenarios to be A Decision Support and Expert System for Retail Planning 183 considered and the analyst then runs his model to assess the impacts of such scenarios. His findings are then used to support the decision-making process as an input to the discussions or negotiations or as a way of formulating additional scenarios. Due to this formalized nature of the planning process, it can often take several weeks before the next meeting is held. Perhaps more problematic though, is that many research contracts include limits on the number of runs executed, which will often frustrate the planning process in that additional analyses require additional funding. Moreover, to assess the impacts of plans of local retailers which emerge later, the model has to be used several times. Hence, a need was felt to develop user-friendly computer software that can be considered as a decision support system for retail planning. We will now describe some computer software that has been developed by the authors. The software consists of three main modules. First, the information base includes software to man- age the necessary information. Second, the modelbase contains software to model consumer choice behaviour. Finally, the third part contains software to summarize the results of different scenarios and evaluate those scenarios on a series of criteria or system performance indicators which will assist the planning team in the decision-making process. Each module runs quite separately and uses the output of previous modules as input. 3.1 Module 1: Database The database module enables users to input and manage the required data. This module can be used to simulate policy measures by changing, deleting or introducing shopping centres in future time periods, or by changing the number of inhabitants per residential zone or even changing the number of residential zones. The required data contains the following information. The Supply Side This part of the information base includes data about the location, size and type of shops. To calculate normative turnover figures for each shopping centre, normative turnover-to-floor space levels, disaggregated to type of shop are required. In addition, characteristics of shopping centres like the amount of parking facilities, type of shopping centre, and so on, are required. The Transportation Network To calculate travel times between residential zones and shopping centres, a transportation network has to be specified, including any future alterations. Alternatively, a matrix of travel times might be specified. Population Figures and Expenditures To predict the turnover in each shopping centre in the future, the development of the number of inhabitants per residential zone and their expenditures need to be known. In addition, the outflow of expenditures to shopping centres in external zones and the inflow of expenditures of consumers living in external zones must be assessed. Choice Behaviour To estimate the parameters of the model, data is required concerning consumer choice behaviour. The required data depends on the type of model used. If a spatial interaction model is used, a so-called observed interaction matrix should be constructed. Each cell of this matrix contains the number of times a particular shopping centre is being visited by consumers living in a particular residential zone. In the case where a random utility model is used, much more information is required. First, for each respondent, the set of available shopping centres has to be identified. Second, the number of times a respondent visits each available shopping centre is 184 A. Borgers and H. Timmermans required. Third, the attribute scores of the available shopping centres as perceived by the respondent are needed. In the case of the simulator system, depending on whether a preference task or a choice task is being applied, individual preferences or intended choices are required. To assess the applicability of the simulator system in practice, observed choice behaviour is required as well. 3.2 Module 2: Mode/base This part of the decision support system contains software to calibrate a consumer choice model and to predict consumer choice behaviour using the calibrated model. For each type of model, a separate computer program is available. In addition, a program to determine the good- ness-of-fit between observed (or intended) choice behaviour and choice behaviour predicted by the model is available. Spinmo The SPINMO (Spatial INteraction MOdelling) program (Borgers & Timmermans, 1988) uses the scores of the shopping centres on one or more attributes and a matrix of distances between residential zones and shopping centres to model the observed flows of consumers. Several distance functions can be specified, for example, a power function, an exponential function or a combination of both. SPINMO minimizes the sum of squared deviances between observed and predicted flows by means of a gradient search and/or a sequential search method. Caldis The CALDIS (CALibration of DIScrete choice models) program (Borgers, 1985) was devel- oped to calibrate several types of discrete choice models. The PC-version of CALDIS, howev- er, only contains the multinomial logit model. The program estimates the weights of the attributes given the number of times each individual chooses each available shopping centre and the attribute scores as perceived by the individuals. The program uses the log-likelihood as an optimization criterion. Like the SPINMO program, CALDIS uses a gradient and/or a sequential search method to find the optimal point in the parameter space. After predicting individual choice frequencies, both observed and predicted individual choice behaviour can be aggregated into flows between residential zones and shopping centres. Giant The GIANT (Generation, Implementation & ANalysis of Treatments) system (Timmermans & Borgers, 1989), part of which is still being developed, can be considered as an expert system to assist the analyst in constructing and analyzing conjoint-based simulator systems. The design of experiments, which is required in using such a modelling approach, demands some expert-knowledge which is often not available. This program aims at filling this gap. It should be emphasized that the system is thus not specifically developed for retail planning applica- tions. It is a program to support the formulation of decompositional preference and choice models and analyze human decision making and choice behaviour in general. The program may thus be of use in all possible kinds of planning and policy-making contexts which relate to preferences and choice behaviour. The program is menu-driven. It allows the user (a) to construct designs and generate experi- mental tasks, (b) to input and control subjects’ responses to such experimental tasks, (c) to esti- mate preference functions and choice models from these experimental designs data, and (d) to simulate actual choice behaviour using the estimated preference functions or choice models. The most difficult part in developing these models is the construction of experimental designs that allow estimating the required model. GIANT allows the user to address this prob- A Decision Support and Expert System for Retail Planning 185 lem at four levels of sophistication and expertise. The inexperienced user can choose from a number of typical problems for which the modelling approach may be used. Depending upon the chosen option, the program then raises some additional questions and finally constructs the required designs. This option might also be used by an experienced user to deal with standard problems. At the next, more sophisticated level the system can be used to answer a series of questions and the system then determines whether a suitable experimental design can be con- structed. A third option allows the user to use an experimental design from the data base of the program. Finally, the program contains an option to construct one’s own design. Evidently, this option requires sufficient knowledge on the part of the user, but the system provides a number of analytic modules to assist the user in understanding the properties of the constructed design. Several standard problems are incorporated in the system design. One such problem is that of conjoint preference measurement. This option should be chosen if one wishes to measure individual preferences for hypothetical choice alternatives. The program enables the user to construct the hypothetical choice alternatives so that individual overall preference measures can be decomposed into part-worth utilities for the attribute levels. Another standard problem concerns the estimation of the impacts of possible modifications in existing alternatives. In this case, the program generates a number of treatments. The proposed modifications and their levels are varied over the treatments so that the effect of each modification can be assessed. Still another problem concerns the prediction of market shares in so-called variable markets. Choice sets varying in size and composition are constructed to provide insight into expressed choice behaviour. To deal with such problems, several decisions have to be made. In the most simple mode, decisions are made by the system which confronts the user only with those options that leave some freedom of choice. All decisions that are logically required to deal with a certain problem are made automatically. In the other modes, these decisions have to be made by the user. The system only warns against inconsistencies and improper solutions and provides some addition- al information. The most important decision in this respect is whether the user wishes to implement a prefer- ence design or a choice design because this decision dictates many of the subsequent decisions, schemes of analyses and simulations. If a choice task is selected, the user has to decide whether to use a paired comparison design or a multiple comparison design. A paired comparison design has often some advantages, but a multiple comparison design should be chosen whenev- er one wishes to examine extended choice sets. The system has several options to construct such multiple comparison designs. In particular, it allows the construction of single fractional factorial designs, split plot designs, foldover designs, and randomized designs. If required, the statistical properties of the design are displayed and the user is guarded against making faults. The system also allows the user to include a base alternative to each choice set and to construct either fixed or varying choice set designs. Having generated a design, the program can be used to output the experimental tasks. The user is requested to name the choice alternatives, attributes, and attribute levels assumed to affect consumer behaviour. In addition, the system permits inputting additional texts introduc- ing the experimental task or specifying the measurement scale. A simple text-processing rou- tine is written to help the user in describing the design and the experimental task. Important for many applied research projects is that the system allows the user to randomize the choice sets, choice alternatives, attributes and/or attribute levels across subjects. This is a very time-con- suming task if the designs are constructed by hand and therefore this option is often omitted in practice at the risk of introducing different kinds of biases in the measurement. The next main part of the computer program allows the user to input the responses to the constructed experimental tasks. Again, several options are available. Ranking, ratings, single choice and allocation data may be input, and whenever possible, the input data are automatical- 186 A. Borgers and H. Timmermans ly checked. The system contains also a number of routines to convert data. Ratings may be converted into rankings; rankings may be exploited to generate choice data; allocation data can be converted into single choice data. Once the response data are available the preference or choice model can be estimated. Several estimation procedures may be used, depending upon the scale properties and the kind of model one wishes to use (for details, see Timmermans, 1984). A necessary step in the esti- mation involves the construction of a design matrix. That is, the design has to be coded in terms of a series of indicator variables. GIANT allows a choice among four coding schemes: dummy coding, effect coding, orthogonal polynomials and a scheme suggested by Louviere (1988). Part of the coding depends upon the model specification and the problem studied. This part of the coding is handled automatically. The coded data can then be used in routines that analyze the data and estimate the specified model. To assist the user in interpreting the results of the analysis, the program generates a couple of summary statistics such as mean coeffi- cients, graphical plots, analyses of monotonicity and tabulations of goodness-of-fit measures. Finally, the system can be used to perform the simulation and assess the likely impacts of designated policy options or plans. The system allows the use of deterministic choice rules, various probabilistic choice rules, and several direct choice models. Information about market shares is provided. GOF Before using the specified model to predict the likely impacts of retail plans, the validity of the model has to be determined. This can be done by comparing predicted flows of consumers between residential zones and shopping centres with observed flows of consumers. The pro- gram GOF calculates several Goodness-Of-Fit measures (see Timmermans & Borgers, 1985) between predicted and observed flows of consumers and between the predicted and observed number of consumers for each shopping centre. 3.3 Module 3: Evaluation In practice, several alternative plans concerning the spatial retail system are usually consid- ered. The planning team has to decide which plan is most desirable. Given the likely impacts of each plan on several criteria, the selection process can be supported by some multi-criteria evaluation system. The evaluation system uses a number of system performance indicators as criteria (van der Heijden, 1986). The indicators compare the impacts of a specific plan to the impacts of the so-called trend scenarios (when no policy measures are realized). Alternatively, the impacts of a specific plan may be compared to normative estimates. In addition to the indi- cators, user determined criteria can be used to evaluate these alternative retail plans. The evaluation technique implemented belongs to the category of mixed data evaluation techniques (see Voogd, 1983). The advantage of mixed data multicriteria evaluation techniques is their ability to process both quantitative and qualitative criteria. Although the system perfor- mance indicators always imply quantitative criteria, the user may wish to add criteria of a qual- itative nature. In addition to a matrix of scores of each retail plan on the criteria, a vector of weights has to be specified. These weights reflect the importance attached to the various crite- ria. In practice, the weights have to be specified by the planning team. The system performance indicators can be distinguished in terms of three types. Indicators of retailers’ interest refer to the economic functioning of the shopping centres. Indicators of con- sumer interest concern the availability of shopping facilities for consumers living in the resi- dential zones. Finally, indicators of public interest concern items such as the equality of distri- bution with respect to the accessibility of consumers to shopping centres and the over and under allocation of expenditures to shopping centres. A Decision Support and Expert System for Retail Planning 187 indicators of Retailers’ Interest It is important for the planning team to get insight into the development of the turnover-to- floor space ratio given some plan as compared to the development according to the trend. The turnover-to-floor space (T-t-F) figures are considered as a surrogate measure of efficiency. The first indicator of retail interest is determined by calculating the ratio T-t-F(plan)/T-t-F&end) for each shopping centre at each time period. A ratio less than 1.0 indicates a negative impact for the retailers in the specified shopping centre in the specified time period. By calculating the mean value over all shopping centres and all time periods, an average indicator for the plan under consideration results. Another indicator of retailers interest is based upon the number of times the T-t-F(plan) is lower than the T-t-F(trend) for a particular shopping centre over all time periods. Averaging over all shopping centres and dividing this average value by the number of time periods gives a second indicator for the proposed plan. Indicators of Consumers’ interest For consumers, the available shopping opportunities are important, Therefore, indicators of consumers’ interest are based on the available amount of floor space and number of shops for consumers living in a particular residential zone. To determine which shopping centres are available, some (residential zone specific) distance threshold has to be specified. The available shopping opportunities for consumers living in a specific residential zone according to the pro- posed plan can be compared to the available shopping opportunities according to the trend by calculating the ratio between both figures. An overall indicator can be determined by averaging the ratios over all residential zones and time periods. Such overall indicators can be calculated for the availability of floor space, floor space per branch, number of shops, number of shops per branch, number of different branches, and so on. Changes in the local access of the shopping centres can be measured by summing the expen- ditures in the nearest shopping centre over the residential zones and time periods. Dividing this sum of expenditures by the sum of expenditures in the nearest shopping centres for the trend gives another indicator of consumer interests. Of course, this indicator can be disaggregated according to various branches of shopping types. lndica tors of Public Interest In general, minimizing travel distances is considered as a planning objective. For a plan under consideration, the average distance travelled by consumers can be calculated. The first indicator of public interest equals the ratio of the average distance under the conditions of the proposed plan and the average distance for the trend. Another general planning objective concerns equali- ty in the retail system. Policy measures should pursue equality between retailers and between consumers. Equality between retailers can be assessed by calculating some measure of deviation in the vector containing the averaged turnover-to-floor space ratio for each shopping centre. Equality between consumers can be determined by calculating a measure of deviation in the vector containing the averaged amount of available shopping facilities per residential zone. 4. EPILOGUE In this paper, we have described a decision support and expert system for retail planning. The computer package, parts of which are still being developed, contains a database manager, a modelbase and an evaluation module. So far, the system is being used both in a scientific envi- ronment to get more insight into spatial choice behaviour, and in an applied planning context. In two main cities of the Netherlands, the so-called PEARL system (Borgers & Timmermans, 1987) is used to predict the effects of retail plans proposed by the local govem- 188 A. Borgers and H. Timmermans ment or by local retailers. PEARL is an abridged version of the decision support system described above. The modelbase contains software to predict consumer choice behaviour at the aggregate level. The models were calibrated at our department using observed consumer choice behaviour. As yet, the evaluation module only relates predicted turnover-to-floor space figures to normative turnover-to-floor space figures. To predict the effects of a new retail plan, the plan has to be converted into new attribute scores (floor space, price setting, distances, etc.) for the shopping centres under consideration. These data have to be entered by means of the data manager. After predicting consumer choice behaviour, the system produces and compares predicted turnover-to-floor space figures to normative figures. Otherwise, the residential building programs with shopping centres. system can be used to evaluate the effects of alternative respect to the economic functioning of the associated REFERENCES Ben-Akiva, M., & Lerman, S. R. (1985). Discrete choice analysis: Theory and application to travel demand. Cambridge, MA: The MIT Press. Borgers, A. (1985). CALDIS: A computer program for the calibration of discrete choice models. Urban Planning Group, University of Technology, Eindhoven, The Netherlands. Borgers, A., & Timermans, H. (1987). PEARL: A computer program to predict the effects of alternative retail lay- outs (in Dutch). Urban Planning Group, University of Technology, Eindhoven, The Netherlands. Borgers, A., & Timmermans, H. (1988). SPINMO: A computer program for spatial interaction modelling. Urban Planning Group, University of Technology, Eindhoven, The Netherlands. Golledge, R., & Timmennans, H. (Eds.). (1988). Behavioural modelling in geography and planning. London: Croom Helm. Louviere, J. (1988). Analyzing decision making: Metric conjoint analysis. Newbury Park, UK: Sage Publications. Louviere, J., & Timmermans, H. (1987). A review of some recent advances in decompositional preference and choice models. Paper presented at the 5th European Colloquium on Quantitative Theoretical Geography, Bardoneccia, Italy. Louviere, J., & Woodworth, G. (1983). Design and analysis of simulated consumer choice or allocation experiments: an approach based on aggregate data. Journal of Marketing Research, 20.350-367. McFadden, D. (1974). Conditional logit analysis of qualitative choice behaviour. In P. Zarembka (Ed.), Frontiers in econometrics (pp. 105-142). New York: Academic Press. Roy, J., & Anderson, M. (1988). Assessing impacts of retail development and redevelopment. In P. Newton, M. Taylor, & R. Sharpe @is.), Desktop planning: Microcomputer applications for infrastructure and services planning and management (pp. 172-179). Melbourne, Australia: Hargreen Publishing Company. Timmermans, H. (1984). Decompositional multiattribute preference models in spatial choice analysis. Progress in Human Geography,& 189-221. Timmermans, H., & Borgers, A. (1985). On the assessment of model performance in spatial choice analysis. In G. Bahrenberg & J. Deiters (Eds.), Methodology, models and methods in regional science (pp. 33-63). Osnabrilk, Germany: OSG. Timmermans, H., & Borgers, A. (1989). GIANT A computer program for the generation, implementation and analysis of treatments. Urban Planning Group, University of Technology, Eindhoven, The Netherlands, van der Heijden, R. (1986). A decision support system for the planning of retail facilities: Theory, methodology and application. PhD dissertation, Urban Planning Group, University of Technology, Eindhoven, The Netherlands. van der Heijden, R., & Timmermans, H. (1984). Modelling choice-set generating processes via stepwise logit regres- sion procedures: Some empirical results. Environment and Planning A, 16,1249-1255. Voogd, H. (1983). Multicriteria evaluations for urban and regional planning. London: Pion. Wilson, A. (1971). A family of spatial interaction models and associated developments. Environment and Planning A, 3, l-32. 
work_eyrgzf5g7neibnpi4o2sfe6bii	----	poe.dvi Genetic wavelet packets for speech recognition Leandro D. Vignolo∗, Diego H. Milone, Hugo L. Rufiner Research Center for Signals, Systems and Computational Intelligence, Departamento de Informática, Facultad de Ingenieŕıa y Ciencias Hı́dricas, Universidad Nacional del Litoral, CONICET, Argentina Abstract The most widely used speech representation is based on the mel-frequency cepstral coefficients, which incorporates biologically inspired characteristics into artificial recognizers. However, the recognition performance with these features can still be enhanced, specially in adverse conditions. Recent ad- vances have been made with the introduction of wavelet based representations for different kinds of signals, which have shown to improve the classification performance. However, the problem of finding an adequate wavelet based representation for a particular problem is still an important challenge. In this work we propose a genetic algorithm to evolve a speech representation, based on a non-orthogonal wavelet decomposition, for phoneme classification. The results, obtained for a set of spanish phonemes, show that the proposed genetic algorithm is able to find a representation that improves speech recog- nition results. Moreover, the optimized representation was evaluated in noise conditions. Key words: Phoneme classification, genetic algorithms, wrappers, wavelet packets ∗Corresponding author. Research Center for Signals, Systems and Computational Intelligence, Departamento de Informática, Facultad de Ingenieŕıa y Ciencias Hı́dricas, Universidad Nacional del Litoral, Ciudad Universitaria CC 217, Ruta Nacional No 168 Km 472.4, TE: +54(342)4575233 ext 191, FAX: +54(342)4575224, Santa Fe (3000), Argentina. Email address: ldvignolo@fich.unl.edu.ar (Leandro D. Vignolo) URL: http://fich.unl.edu.ar/sinc (Leandro D. Vignolo) Preprint submitted to Expert Systems with Applications September 18, 2012si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. 1. Introduction Automatic speech recognition systems need a pre-processing stage to make phoneme key-features more evident, in order to obtain significant im- provements in the classification results [1]. This task was first addressed by signal processing techniques like filter-banks, linear prediction coding and cepstrum analysis [2]. The most popular feature representation currently used for speech recognition is built from the mel-frequency cepstral coeffi- cients (MFCC) [3], which are based on a linear model of voice production together with the codification on a psycho-acoustic scale. However, due to the degradation of recognition performance in the presence of additive noise, many advances have been conducted in the development of alternative fea- ture extraction approaches. In particular, techniques like perceptual linear prediction [4] and relative spectra [5] incorporate features based on the hu- man auditory system and provides some robustness in ASR. Also, significant progress has been made with the application of different artificial intelli- gence techniques in the field of speech processing [6]. Besides, the utilization of wavelet based analysis for speech feature extraction has recently been studied [7, 8, 9, 10]. The multi-resolution analysis associated with discrete wavelet transform (DWT) can be implemented as a filter bank decomposition (or filter bank schemes) [11]. Wavelet packet transform (WPT) is a generalization of the DWT decomposition which offers a wider range of possibilities for signal rep- resentation in the time-scale plane [12]. To obtain a representation based on this transform, usually, a particular orthogonal basis is selected among all the available basis. Nevertheless, in phoneme classification applications there is not evidential benefit on working with orthogonal basis. Moreover, it is known that the analysis performed at the level of the auditory cortex is highly redundant and, therefore, non-orthogonal [13]. Without this restric- tion the result of the full WPT decomposition is a highly redundant set of coefficients, from which a convenient representation for the problem in hand can be selected. Many approaches addressing the optimization of wavelet decompositions for feature extraction have been proposed. For instance, in [14] an automatic extraction of high quality features from continuous wavelet coefficients ac- cording to signal classification criteria was presented. In [15], an approach based on the best basis wavelet packet entropy method was proposed for elec- troencephalogram classification. Also, a method for the selection of wavelet 2si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. family and parameters was proposed for the phoneme recognition task [16]. Similarly, the use of wavelet based decompositions has also been proposed as a tool for the development of robust features for speaker recognition [17, 18]. Another interesting approach was proposed in [19], in which a novel approach for generating the wavelet that maximizes the discrimination capability of ECG beats using particle swarm optimization. Also, the use of evolutionary computation techniques in order to optimize over-complete decompositions for signal approximation was proposed in [20]. Furthermore, the use of a genetic algorithm to optimize WPT based features for pathology detection from speech was proposed in [21], where an entropy criterion was minimized for the selection of the wavelet tree nodes. Similar approaches propose the optimization of wavelet decomposition schemes using evolutionary compu- tation for denoising [22, 23] and image compression [24]. Besides, different approaches have been proposed for the optimization of wavelet based repre- sentations using swarm intelligence [25, 26]. Many other studies also rely on evolutionary algorithms for feature selection [27, 28, 29] and the optimization of speech representations [30, 31, 32]. However, the flexibility provided by the full WPT decomposition has not yet been fully exploited in the search for a set of features to improve speech recognition results. When this search is not restricted to non-redundant representations, there is a large number of non-orthogonal dictionaries to be explored, leading to a hard combinatorial problem. Here we propose a new approach to optimize over-complete decomposi- tions from a WPT dictionary, which consists in the use of a genetic algorithm (GA) for the selection of wavelet based features. In order to evaluate the so- lutions during the search, the GA uses a learning vector quantization (LVQ) classifier. Some preliminary results with this strategy were presented in [33]. The methodology, referred to as genetic wavelet packets (GWP), relies on the benefits provided by evolutionary computation to find a better signal representation. This feature selection scheme, known as wrapper [34, 35], is widely used as it allows to obtain the good solutions in comparison with other techniques [36]. The organization of this paper is as follows. In Section 2, brief descriptions of the properties of WPT and GA are presented. Next, our wrapper method for the selection of the WPT components is described. The following section discusses the obtained recognition results for a set of spanish phonemes. Finally, the general conclusions and future work are presented. 3si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. 2. Materials and methods 2.1. Wavelet and wavelet packet transforms In contrast with sine and cosine bases, wavelet bases are simultaneously localized in time and frequency. This feature is particularly interesting in the case of signals which present both stationary and transient behaviors. Wavelets can be defined, in a simplified manner, as a function of zero mean, unitary norm and centered in the neighborhood of t = 0 [37]: ψ(t) ∈ L2(R); ∫ ∞ −∞ ψ(t)dt = 0; ‖ψ(t)‖ = 1. (1) A family of time-frequency atoms is obtained by scaling and translating the wavelet function: ψu,s(t) = 1√ s ψ ( t − u s ) , (2) with u,s ∈ R. This way, the continuous wavelet transform of the signal x(t) is defined as the inner product with this family of atoms Wx (u,s) = 〈x(t),ψu,s(t)〉 = +∞ ∫ −∞ x(t) ψ∗u,s (t) dt. (3) The discrete dyadic wavelet transform (DWT) of x[n] ∈ RN is obtained by discretizing translation and scaling parameters in (3), as u = m and s = 2j . A fast implementation of the DWT based multiresolution analysis exists [11], which uses low-pass and high-pass filters to decompose a signal to detail (dj[n]) and approximation (aj [n]) coefficients. Since the filter outputs contain half the frequency components of the original signal, both approximation and detail can be sub-sampled by two, maintaining the number of samples. This process is iteratively repeated for the approximation coefficients, increasing the frequency resolution on each decomposition step. As result a binary decomposition tree is obtained, where each level corresponds to a different scale j [38] dj+1[m] = √ 2 ∞ ∑ n=−∞ g[n − 2m]aj [n], (4) aj+1[m] = √ 2 ∞ ∑ n=−∞ h[n − 2m]aj[n], (5) 4si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. here g[n] and h[n] are the impulse responses of the high-pass and low-pass filters associated with the wavelets and scaling functions, respectively. The WPT could be considered as an extension of the DWT which pro- vides more flexibility on frequency band selection. With the same reasoning above, details (high frequency components) can be decomposed as well as approximations (low frequency components). In a similar way to the DWT, the full wavelet packets decomposition tree is obtained by c 2p j+1[m] = √ 2 ∞ ∑ n=−∞ g[n − 2m]cpj [n], (6) c 2p+1 j+1 [m] = √ 2 ∞ ∑ n=−∞ h[n − 2m]cpj [n], (7) where j is the depth of the node and p indexes the nodes in the same depth, every c p j with p even is associated to approximations and every c p j with p odd is associated to details. The wavelet packet analysis allows to represent the information contained in a signal in a more flexible time-scale plane, by selecting different sub-trees from the full decomposition (Figure 1). For the selection of the best tree it is possible to make use of the knowledge about the characteristics of the signal and to obtain an efficient representation in the transform domain. For the case of signal compression the criteria is based on “entropy” measures, method named as best orthogonal basis [39]. Another possibility, closer to the classification problem, is to use the local discriminant basis algorithm, which provides an appropriate orthogonal basis for signal classification [40]. These criteria are based on the assumption that an orthogonal basis is convenient. Nevertheless, for the case in study there is not evidence on the convenience of a non-redundant representation. Moreover, the redundancy often provides additional robustness for classification tasks in adverse conditions [31]. Be- cause of this, a method which explores a wider range of possibilities should be studied. 2.2. Genetic wavelet packets Genetic algorithms are meta-heuristic optimization methods motivated by the process of natural evolution [41]. A classic GA consists of three kinds of operators: selection, variation and replacement [42]. Selection mimics the natural advantage of the fittest individuals, giving them more chance to 5si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Figure 1: Wavelet packets tree with six decomposition levels for a 256 samples signals. Figure 2: General scheme of the proposed wrapper method. reproduce. The purpose of the variation operators is to combine informa- tion from different individuals and also to maintain population diversity, by randomly modifying chromosomes. The number of individuals in the cur- rent population that are substituted by the offspring is determined by the replacement strategy. The information of a possible solution is coded in a chromosome and its fitness is measured by an objective function, which is specific to a given problem. Parents, selected from the population, are mated to generate the offspring by means of the variation operators. The popula- tion is then replaced and the cycle is repeated until a desired termination criterion is reached. Once the evolution is finished, the best individual in the population is taken as the solution for the problem [43]. Genetic algorithms are inherently parallel, and one can easily benefit from this to increase the computational speed [33]. In this case, the objective function needs to evaluate the signal represen- tation suggested by a given chromosome, providing a measure of the class separability. Therefore, the fitness function was defined as a phoneme clas- sifier, based on the optimized learning vector quantization (O-LVQ) tech- 6si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Figure 3: Frequency band integration scheme (half tree). nique [44]. This classifier uses a set of reference vectors (codebook ) which are adapted using a set of training patterns in order to represent the distribu- tion of classes. O-LVQ was chosen because it requires much less processing time than those used in state-of-the-art speech recognizers, based on hidden Markov models (HMM). Although, after the optimization, for the validation of the evolved representation an HMM based classifier was also used. For the evaluation of every individual in the population, the classifier is trained and tested on a phoneme corpus, and the recognition rate is used as the fitness value. The scheme of the proposed wrapper method is shown in Figure 2. The GA uses roulette wheel selection method, the classic mutation and one-point crossover. Also, an elitist replacement strategy was applied, which maintains the best individual to the next generation. The feature extraction scheme was designed for signals of 256 samples length, this is 32 ms frames at 8 kHz sampling frequency. And the iterative process of filtering and decimation was performed to obtain six decomposi- tion levels, obtaining a full wavelet packet tree consisting of 1792 coefficients. Then, in order to reduce the dimensionality of the search space, the coeffi- cients inside each frequency band were “integrated” by groups. This means that each band was subdivided into groups, and an energy coefficient for each group was obtained by e s j = ∑ ∀ck∈G s j |ck|2, (8) where esj is the energy coefficient for integration group j in scale s, G s j , and ck is the k-th coefficient belonging to this group. Figure 3 illustrates the integration scheme for the first half of the WPT decomposition tree, while the other half is integrated in the same manner. Each small square represents 7si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Table 1: Integration scheme applied to the WPT tree for a 256 sample signal, which reduces from 1792 wavelet coefficients to 208 integration coefficients. Level 1 2 3 4 5 6 Nodes 2 4 8 16 32 64 Integration groups per node 8 8 4 2 1 1 Wavelet coefficients per group 16 8 8 8 8 4 Integration coefficients per level 16 32 32 32 32 64 a single component, in the first row (level 0) this is a sample of the temporal signal and for the other rows (levels 1 to 6) each of them correspond to a single wavelet coefficient. White and gray zones delimit the different integration groups and the tree nodes in each decomposition level are indicated with thick line rectangles. Table 1 shows the number of components and coefficients in each integration group. This integration scheme was heuristically designed, considering the most relevant frequency bands in speech and their temporal resolutions. After band integration, a normalization was applied: if wp[k] is the k- th energy coefficient corresponding to the pattern p, then the normalized coefficient will be given by ŵp[k] = wp[k] max i {wi[k]} . (9) A canonical evolution model with binary chromosomes was used, in which every individual represents a different selection of the WPT band-integrated coefficients. Each gene in a chromosome represents one normalized coef- ficient, and its value indicates whether that coefficient should be used to parametrize the signals (Figure 4). Once the data base is processed, each feature vector is composed by the normalized and band-integrated WPT co- efficients. These labeled patterns are used to train and test the O-LVQ based classifier. When a particular individual is evaluated, each feature vector is reduced to the subset of coefficients indicated by the chromosome. The selection of individuals should be done considering the set of coeffi- cients represented by each chromosome. The chromosomes which codify the best signal representations, those which allow better classification results, should be assigned high probability. As the codification could be redundant and no restriction is imposed for coefficients combination, the GA initial- ization consists on a random settling of the genes in the chromosomes. All 8si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Figure 4: Codification example with a 80 genes chromosome and the corresponding WPT tree. Dark boxes represent the used coefficients and the white boxes represent those that are discarded. Algorithm 1: Optimization for GWP. Obtain full WPT for each phoneme example in the corpus using (6) and (7) Apply the band-integration scheme to WPT coefficients using (8) Normalize the integrated coefficients for each pattern according to (9) Initialize the GA population Evaluate population (Algorithm 2) repeat Select parents (roulette wheel) Create a new population from selected parents Replace population Evaluate population until stopping criteria is met the steps involved in the GWP feature selection strategy are summarized in Algorithm 1, while the details for the population evaluation are shown in Algorithm 2. 2.3. Phoneme corpus The speech data used for experimentation is a subset of the Albayzin geographic corpus [45, 46], named Minigeo. This subset consists of 600 ut- terances, spoken by twelve different speakers (six men and six women) which where between fifteen and fifty five years old. This speech data has been phonetically segmented using a speech recognition system based on hidden Markov models [47]. This way the temporal localization of every phoneme within each utterance was obtained. The extracted phonetic corpus was par- titioned in three groups, a training set and a testing set to be used during evolution, and a third data set which was used only for the validation of best 9si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Algorithm 2: Evaluate population. for each individual in the population do Re-parameterize training/test patterns according to the chromosome Train the LVQ based classifier on the training set Test the LVQ based classifier on the test set Assign classification rate as the current individual’s fitness solutions after the feature selection process. The experiments included the phonemes /a/, /e/, /i/, /o/, /u/, /b/, /d/, /p/ and /t/ from Spanish. The five vowels were included because of their obvious importance in the language, while the four occlusive phonemes have been included because of their similar characteristics, that make a set particularly difficult to distinguish [48]. Even though it can be suggested that all phonemes should be included, our hypothesis is that this strategy simplifies the task of the GA while still allows to find features useful in continuous ASR. In order to avoid adding additional complexity to the search of the GA, every sample used in the optimization consisted in a single speech frame of 32 ms length, which was extracted from the center of the phoneme utterance. 3. Results and discussion 3.1. Genetic optimization of wavelet packets In [49], various wavelet families have been tested in order to find which one is the most convenient for signal classification. In this work the tests included the most widely used families, among which we can mention Meyer, Daubechies, Symmlets, Coiflets y Splines [50]. As result, the 4th order Coiflet family was chosen to be used on the following experiments. For the first experiment, a codebook of 117 vectors (13 per phoneme) was used within the LVQ classifier and the initial learning rate was set on 0.02. The classifier training was made in 6 epochs with 1449 patterns and 252 patterns were left for testing. For the GA, the population size was set on 100 individuals, while crossover and mutation probabilities were set on 0.9 and 0.05, respectively. The performance of the best solution found was 57.94% of correct classifications. As the LVQ codebook initialization has a random component, repeated evaluations for the same individual may result on different fitness values. 10si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Table 2: Summary of obtained fitness and validation results. Strategy Convergence Fitness Validation (# generations) average STD Random LVQ initialization 26 57.94% 53.69% 2.33% Fixed LVQ initialization 216 57.78% 56.67% 2.9% Generational LVQ initialization 355 64.07% 59.16% 2.91% Then, in order to obtain a good estimation of the performance for the best individual, after the evolution, it was evaluated ten times with a validation data set In this process we used 2637 training patterns and 450 test pat- terns which were not used during the evolution. As result, an average of 53.69% correct classifications with a standard deviation of 2.33% was ob- tained. Also, in order to analyze the effect of the random LVQ initialization on the evolution, two different alternatives were considered. In the first case, the randomness was eliminated from the codebook initialization. In this case, the GA was able to find an individual with a fitness (classification rate) of 57.78%. With the validation procedure, described earlier, an average clas- sification rate of 56.67% was obtained. Even though an improvement was obtained, it is possible that the evolution was biased by this fixed codebook initialization. Then, another strategy was considered, in which a fixed initial- ization sequence was used for each generation. In order to allow individuals evaluated with different conditions (initializations) to coexist in the same generation a generational gap of 10 individuals was used, maintaining more information from one generation to the next. This means that, in addition to the best individual which is preserved by the elitist strategy, another 10 individuals are chosen by the selection algorithm to be maintained to the next generation. In this case the best solution achieved 64.07% correct clas- sifications, and an average classification rate of 59.16% was obtained with the validation data. Table 2 summarizes these results, showing that this last strategy allowed to improve the generalization capability. Table 3 shows a confusion matrix obtained from validation results. As this matrix shows, the /t/ phoneme is mainly (61.51%) classified as /p/. This error turn up because the experimental data was taken from the cen- tral part of the samples and the plosive phonemes, like /p/ and /t/, have their most particular attributes at the beginning (the phoneme plosion). A similar problem happens with phoneme /d/, and this might be solved for all plosive phonemes by considering their context (i.e. a number of precedent 11si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Table 3: Confusion matrix obtained from the validation of the best GA solution, giving 59.16% average classification rate . /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ 84.85 00.30 00.00 11.82 01.21 01.21 00.30 00.30 00.00 /e/ 02.73 76.06 01.82 05.15 03.64 03.33 01.51 03.64 02.12 /i/ 00.00 08.18 86.97 00.00 00.30 02.73 00.61 00.30 00.91 /o/ 25.15 10.61 00.00 42.42 14.24 04.24 02.73 00.61 00.00 /u/ 08.79 00.00 01.51 08.48 59.39 14.24 00.61 05.15 01.82 /b/ 00.30 02.42 00.00 04.54 09.70 62.12 06.06 13.03 01.82 /d/ 10.30 31.82 09.09 07.57 04.54 04.54 10.61 17.27 04.24 /p/ 00.00 00.00 00.00 00.00 00.00 03.03 02.42 78.18 16.36 /t/ 00.00 00.00 00.00 00.30 00.00 04.54 01.82 61.51 31.82 and posterior frames). 3.2. Comparative analysis In order to compare this results, the same classifiers were trained with other state-of-the-art speech features: the classic MFCC [3], the alterna- tive cepstral features based on Slaney’s filter-bank [51], and a representation based on the standard DWT. It is worth pointing out that, as the speech data used in these experiments was sampled at 8 kHz, the original parameters of the filter-bank proposed by Slaney were modified by following the reason- ing in [52]. Table 4 shows the results obtained with the above mentioned representations and using a validation data set consisting of 450 patterns, which were not used for the optimization. As it can be seen, the best aver- age classification rate was obtained with the GWP. This shows that by the genetic optimization of the full WPT decomposition it is possible to improve the classification results, in contrast to the classic cepstrum based represen- tations. On the other hand, it can be noticed that the phoneme /d/ seems to be the most difficult for all representations, except for DWT. Even though the state-of-the-art speech recognizers use HMM for acoustic modeling, the significant improvements provided by GWP when using an LVQ based classifier should not be disregarded. These results clearly show that a more efficient class separation is provided in the GWP features space. It should be taken into account that this straightforward classifier was used as objective function to guide evolution, and only the central frame was considered for each phoneme pattern during the optimization. 12si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Table 4: Classification results obtained with an LVQ based classifier. Phoneme Cepstral coefficients Wavelets MFCC(13) Slaney(18) DWT(256) GWP(104) /a/ 82.80 70.40 69.39 84.85 /e/ 77.40 61.60 54.54 76.06 /i/ 90.00 61.60 84.54 86.97 /o/ 46.20 28.40 54.54 42.42 /u/ 31.60 21.40 31.51 59.39 /b/ 45.20 39.60 58.48 62.12 /d/ 08.80 09.00 59.69 10.61 /p/ 55.60 45.20 04.85 78.18 /t/ 48.60 58.40 31.21 31.82 Average 54.02 43.96 49.86 59.16 3.3. Evaluation with hidden Markov models In this work we have raised the hypothesis that, by using a simple classifier in the optimization, the class separability would be maximized and this could also be beneficial to a more sophisticated classifier, like HMM [53]. In order to verify this, the performance of an HMM based classifier was evaluated for each of the representations in Table 4. This classifier is based on a continuous HMM, using Gaussian mixtures with diagonal co-variance matrices for the observation densities, as common in ASR [54]. We used a three state model with mixtures of four gaussians, constructed with the tools provided in the HMM Toolkit (HTK) [47]. These tools use the Baum-Welch algorithm [55] to train the HMM parameters, and the Viterbi algorithm [53] to search for the most likely state sequence. This classifier was evaluated in a ten-fold cross-validation process with random partitions, each of which consisted of 2484 and 621 patterns for training and testing, respectively. It is important to point out that, because of the nature of HMM, in the evaluation of this classifier all the successive frames composing a phoneme were used. While, for the LVQ based classifier, only the central frame was used for a particular phoneme. For the features based on DWT the training of the HMM classifier could not converge, which is mainly because the gaussian mixtures are not able to adequately model the probability densities of these coefficients [56]. An- other problem for training the models with DWT coefficients is due to the high dimensionality of this representation. Then, a post-processing based on principal component analysis (PCA) [57] was applied in order to obtain a rep- resentation of lower dimensionality, with probability densities more similar to 13si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Table 5: Classification results obtained with an HMM based classifier. Phoneme Cepstral coefficients Wavelets MFCC(13) Slaney(18) DWT+PCA(134) GWP(104) /a/ 59.70 58.26 49.71 54.21 /e/ 67.55 64.21 45.50 60.30 /i/ 59.00 63.49 62.02 76.82 /o/ 31.74 27.53 27.38 34.78 /u/ 37.68 51.02 37.82 58.99 /b/ 43.76 26.08 42.61 30.59 /d/ 30.72 16.52 22.45 26.81 /p/ 36.96 40.00 36.37 49.44 /t/ 71.46 60.00 59.43 53.33 Average 48.74 45.24 42.60 49.48 gaussians. The best result for DWT+PCA was obtained when preserving the 99% of the variance, giving a representation of 134 dimensions. Even though our optimized representation is also based on wavelets, and the same prob- lem could be expected, no post-processing was necessary for GWP. Thus, we can assume that the band integration, besides reducing dimension, produced coefficients with probability densities more appropriate to gaussian mixture modeling. Table 5 shows the phoneme classification results obtained with HMM, comparing GWP and other state-of-the-art speech representations. The op- timized representation is the same from Table 4, which was evolved using an LVQ-based classifier. It can be noticed that the best results are those ob- tained by means of the optimized representation and MFCC, similar to the case of the validation with LVQ (Table 4). Even though the fitness was mea- sured by means of an LVQ based classifier in the optimization, the evolved representation provided satisfactory results when using HMM. This means that the optimized representation captures information which is relevant for the discrimination, regardless of the type of classifier. Moreover, using this low-cost classifier, we have successfully saved significant computational time in the optimization. It should be taken into account that if we had used an HMM based classifier and, therefore, considered all the possible frames within a phoneme example, the evaluation of each individual would have taken approximately ten times more. It is also important to note that the proposed GWP representation, which yielded the best classification result when using an LVQ based classifier, pro- vides relatively lower performance with HMM. This is because, as explained 14si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. before, the probability densities of the coefficients provided in wavelet based representations are not entirely suitable for gaussian mixture models [56]. Then, different alternatives remains to be explored besides band integration and PCA post-processing, in order to obtain coefficients more suitable to gaussian mixture models. Also, it should be considered that the dimension- alities of the wavelet based representations are much higher than those of the cepstral representations, which makes the training of the classifier more difficult. Despite the previous considerations, the results from Tables 4 and 5 show that the proposed method is useful for the optimization of wavelet based rep- resentations. Moreover, the results obtained with the GWP features shown that using this evolutionary methodology it is possible to improve the per- formance of the classical representations in ASR. 3.4. Evaluation in noise conditions In order to evaluate the robustness of the optimized representation, white noise was added to the original utterances. The tests were made at several noise levels, and the mismatch training and test (MMTT) condition was considered, which means that the classifiers were trained with clean signals and tested with noisy signals. In general, the input of an ASR system would consist on speech signals with different SNR to those in the training set. For this reason, the evaluation of the recognition performance in MMTT conditions is more realistic than the case where the SNR is the same in both training and test sets. Each test consisted in a ten-fold cross-validation process with random partitions, which consisted of 2484 and 621 patterns for training and testing, respectively. The process of training and testing was repeated for these ten partitions and results were averaged, ensuring that the resultant accuracy would not be biased because of a particular data partition. Figure 5 shows the average results, as well as the estimated standard deviations, showing that the GWP improves the MFCC in all cases. At 0 and 5 dB SNR the DWT+PCA representation gives better results, however, for most of the noise conditions its performance is noticeably worse than GWP. It can be noticed that, on average, the results given by GWP are significantly better when compared to the other representations. In order to evaluate the statistical significance of these results, we estima- ted the probability that GWP is better than each of the reference representa- tions for the given tasks. To perform this test, statistical independence of the 15si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. 0 dB5 dB10 dB15 dB20 dB30 dB40 dBclean 10 15 20 25 30 35 40 45 50 55 SNR C la ss ifi ca tio n r a te [ % ] GWP DWT+PCA MFCC Slaney Figure 5: Classification results obtained with an HMM based classifier evaluated in MMTT conditions. classification errors for each phoneme has been assumed, and the binomial distribution of the errors was approximated by means of a Gaussian distri- bution. Table 6 shows the statistical significance of the classification rates obtained with the HMM based classifier: results improved by GWP with statistical significance higher than 97% are indicated with the superscript △. It can be noticed that most of the classification improvements obtained with the proposed optimized representation are statistically significant. For instance, the probability that GWP performs better than the cepstral coef- ficients obtained with Slaney’s filterbank is higher than 0.99 for all SNRs. Table 7 shows more detailed information about the classification perfor- mance comparing MFCC and GWP, for the case of 40 dB SNR in MMTT conditions. In these confusion matrices, the rows correspond to the actual phoneme and the columns to the predicted phoneme, while the percentages of correct classification are on the diagonal. These matrices show coincidences between the phonemes which are most confused with MFCC and the ones that are confused with GWP. For example, in both cases phoneme /t/ was confused with /p/ and vice versa, which is reasonable as these two plosive 16si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Table 6: Statistical significance of classification results obtained with an HMM based classifier evaluated in MMTT conditions. Superscript △ indicates that the statistical significance of the improvement of GWP is higher than 97%. SNR Cepstral coefficients Wavelets MFCC(13) Slaney(18) DWT+PCA(134) GWP(104) clean 48.74 45.24△ 42.60△ 49.48 40 dB 48.58△ 45.15△ 43.40△ 50.19 30 dB 47.38 43.82△ 41.18△ 47.77 20 dB 40.42△ 43.64△ 42.23△ 46.72 15 dB 39.12△ 39.19△ 41.08△ 42.93 10 dB 33.77△ 23.45△ 36.10 36.11 5 dB 23.32△ 12.55△ 27.78 25.05 0 dB 12.84△ 11.17△ 22.34 16.42 Table 7: Confusion matrix obtained from the validations with MFCC and GWP in MMTT conditions and 40 dB SNR. MFCC GWP /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ 61.7 08.8 00.2 11.9 02.8 04.5 09.6 00.3 00.3 57.8 08.3 00.0 13.2 11.6 05.1 04.1 00.0 00.0 /e/ 11.0 66.4 06.5 04.8 03.9 03.7 03.8 00.0 00.0 09.1 61.0 08.4 04.4 01.2 02.8 11.7 00.3 01.2 /i/ 00.2 24.6 60.9 02.6 06.0 03.8 02.0 00.0 00.0 00.2 06.8 77.5 00.7 03.1 00.6 09.4 00.7 01.0 /o/ 15.5 17.1 03.2 32.3 13.6 13.3 03.5 00.3 01.2 08.6 10.3 01.2 35.7 29.3 05.5 05.7 02.6 01.3 /u/ 04.8 04.2 07.4 19.6 38.0 15.9 09.0 00.9 00.3 02.3 01.8 02.9 16.1 57.1 13.8 04.1 02.1 00.0 /b/ 04.7 01.3 03.4 09.9 05.9 45.2 18.6 09.9 01.3 04.5 00.3 00.0 11.5 25.8 29.7 17.0 06.2 05.1 /d/ 06.1 37.0 00.9 03.1 02.8 07.4 27.8 05.7 09.4 06.1 23.2 09.4 05.4 07.7 05.4 26.8 07.4 08.7 /p/ 00.0 00.4 00.0 00.0 01.3 06.0 16.7 34.1 41.6 00.0 00.0 00.0 00.2 00.7 03.2 10.0 52.5 33.5 /t/ 00.0 00.0 02.3 00.2 00.2 02.6 07.0 16.8 71.0 00.0 01.0 00.7 00.4 00.9 00.6 12.0 30.7 53.6 Average: 48.58% Average: 50.19% consonants share many spectral features. In a similar way vowels /o/ and /u/, which are close in the formants map, are quite confused in both cases. Similarly, Table 8 shows the confusion matrices obtained in the classification with MFCC and GWP in the case of 15 dB SNR and MMTT conditions. It can be noticed that the vowels /a/ and /u/ are mostly misclassified with MFCC, but they are significantly better distinguished with GWP. The op- timized features also introduced an important improvement for the vowel /i/, which is confused with the phoneme /d/ when using MFCC. It is also interesting to note that the phonemes which have their classification rates most affected by noise when using MFCC and GWP do not match. Though, when the noise level is increased, the number of confusions between phonemes /t/ and /d/ increases for both MFCC and GWP. Similarly, the number of confusions between phonemes /t/ and /p/ is also increased in both cases. Another interesting remark is that, for both representations, phoneme /t/ is 17si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. Table 8: Confusion matrix obtained from the validations with MFCC and GWP in MMTT conditions and 15 dB SNR. MFCC GWP /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ /e/ /i/ /o/ /u/ /b/ /d/ /p/ /t/ /a/ 04.5 31.0 00.0 20.1 00.0 03.2 32.8 08.1 00.3 43.8 17.3 00.5 18.1 09.8 03.4 07.3 00.0 00.0 /e/ 00.0 80.6 03.3 00.0 00.0 00.3 15.8 00.0 00.0 04.2 60.3 17.7 03.5 01.3 00.9 11.0 00.0 01.2 /i/ 00.0 19.1 53.1 00.0 00.2 00.2 24.4 00.0 03.2 00.0 11.0 82.5 01.2 03.4 00.3 01.7 00.0 00.0 /o/ 00.0 25.2 01.8 26.7 01.3 09.6 24.6 05.4 05.5 07.1 13.6 00.8 31.9 29.6 02.5 12.5 00.2 02.1 /u/ 00.0 08.4 11.6 16.8 03.6 27.8 29.4 01.3 01.0 02.3 04.4 06.1 19.1 57.5 05.2 05.4 00.0 00.0 /b/ 00.0 03.7 04.1 03.5 00.4 12.0 56.7 08.7 11.0 01.5 09.3 08.6 18.1 29.9 09.6 14.9 01.6 06.7 /d/ 00.0 29.6 01.2 00.6 00.0 02.8 49.0 03.4 13.6 04.5 27.7 12.3 02.1 11.4 00.9 27.5 02.2 11.5 /p/ 00.0 00.5 00.0 00.0 00.0 00.0 05.1 48.1 46.4 00.0 00.5 06.8 01.6 08.7 01.6 19.4 12.0 49.4 /t/ 00.0 00.0 00.0 00.0 00.0 00.0 13.3 12.2 74.5 00.0 01.9 06.7 00.0 03.2 00.7 15.7 10.6 61.3 Average: 39.12% Average: 42.93% better classified when the SNR is 15 dB than when it is 40 dB. These results show that the classification performance of the classical representation can be improved. This suggest that, by means of this GWP methodology, ad- ditional robustness against white noise can be provided to a state-of-the-art ASR system. In order to perform a qualitative analysis of the optimized representation, the tiling of the time-frequency plane was constructed from the selected de- composition using the criteria proposed in [58]. The result is shown in Fig- ure 6, where each decomposition level is depicted separately for an easier interpretation. Each ellipse represents a selected group from the integration scheme (Table 1), then their widths and time localizations are determined by the time-frequency atoms corresponding to the integration group (Figure 3). Therefore, each element in the tiling represents a time-frequency atom that was obtained by adding the original wavelet atoms, according to the integration scheme. This explains why the atoms for levels 1 and 2 are the same time width, as the number of coefficients integrated in the groups for level 1 are twice the number of coefficients in the groups for level 2 (Figure 3). The same explanation applies for the width of the atoms in levels 5 and 6. Note that the whole time-frequency tiling is obtained by the superposi- tion of these six sub-figures, yielding great overlapping between atoms from different decomposition levels. A first observation is that the optimization of the WPT-based decomposition led to a highly redundant non-orthogonal representation, which has been able to exploit this redundancy in order to increase robustness against additive noise. However, the optimized repre- sentation uses only 50% of the total of the coefficients obtained from the integration of the whole WPT-tree. This also suggest that it could be pos- 18si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. sible to achieve still further redundant and robust representations. It is also interesting to note that there are some selected atoms concentrated at the center of the time axis, which could be related to how the phonemes were sampled from the speech corpus, as only the frames extracted from the center of each utterance were considered for the optimization. Also, there are some atoms concentrated at the side parts of the time axis, which could be related to the plosive phonemes. 4. Conclusion and future work A wrapper optimization strategy has been proposed, taking advantage of the benefits provided by evolutionary computation techniques, in order to carry on the search for an advantageous wavelet-based speech representa- tion. The results, obtained in the classification of a group of nine phonemes from spanish, shown that the optimized representation provides important improvements in comparison to the classical features. This suggests that the task of a classifier is simplified when using this optimized representation, due to a better class separation in the features space. Therefore, the proposed strategy provides an alternative feature set for speech signals, which allows to improve the classification results in the presence of noise. In future work we would design more specific genetic operators, so that more information about the problem in hand could be incorporated to the search. Pursuing the objective of finding a representation more suitable for HMM with gaussian mixture modeling, one interesting idea is to incorpo- rate some measure about the gaussianity of the probability densities of the GWP coefficients to the fitness function. Besides, we would study different alternatives to the proposed band integration scheme and the use of tempo- ral information, in regard to successive speech frames, like first and second derivatives. In order to obtain a representation which allows to improve the results in a continuous speech recognition system, future experiments will include more phonemes in the data-sets used for the optimization. Also, the robustness of the representation will be evaluated in comparison with different state-of- the-art robust representations, considering different noise types. References [1] L. Rabiner, B. Juang, Fundamentals of Speech Recognition, Prentice Hall, NJ, 1993. 19si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. 0 5 10 15 20 25 30 0 1 2 3 4 Time [msec] F re q u e n cy [ K H z] Selected atoms for level 1 0 5 10 15 20 25 30 0 1 2 3 4 Time [msec] F re q u e n cy [ K H z] Selected atoms for level 2 0 5 10 15 20 25 30 0 1 2 3 4 Time [msec] F re q u e n cy [ K H z] Selected atoms for level 3 0 5 10 15 20 25 30 0 1 2 3 4 Time [msec] F re q u e n cy [ K H z] Selected atoms for level 4 0 5 10 15 20 25 30 0 1 2 3 4 Time [msec] F re q u e n cy [ K H z] Selected atoms for level 5 0 5 10 15 20 25 30 0 1 2 3 4 Time [msec] F re q u e n cy [ K H z] Selected atoms for level 6 Figure 6: Tiling of the time-frequency plane obtained for the optimized decomposition. For a better visualization, each decomposition level was schematized separately. 20si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. [2] L. Rabiner, R. Schafer, Digital Processing of Speech Signals, Prentice Hall, NJ, 1978. [3] S. V. Davis, P. Mermelstein, Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences, IEEE Transactions on Acoustics, Speech and Signal Processing 28 (1980) 57–366. [4] H. Hermansky, Perceptual linear predictive (plp) analysis of speech, J Acoust Soc Am 87 (4) (1990) 1738–1752. doi:10.1121/1.399423. [5] H. Hermansky, N. Morgan, Rasta processing of speech, IEEE Trans Speech Audio Process 2 (1994) 578–589. doi:10.1109/89.326616. [6] A. Hassanien, G. Schaefer, A. Darwish, Computational intelligence in speech and audio processing: Recent advances, in: X.-Z. Gao, A. Gaspar-Cunha, M. Kppen, G. Schaefer, J. Wang (Eds.), Soft Com- puting in Industrial Applications, Vol. 75 of Advances in Intelligent and Soft Computing, Springer Berlin / Heidelberg, 2010, pp. 303–311, 10.1007/978-3-642-11282-9-32. [7] N. Nehe, R. Holambe, DWT and LPC based feature extraction methods for isolated word recognition, EURASIP Journal on Audio, Speech, and Music Processing 2012 (1) (2012) 7. doi:10.1186/1687-4722-2012-7. URL http://asmp.eurasipjournals.com/content/2012/1/7 [8] S. Patil, M. Dixit, Speaker independent speech recognition for diseased patients using wavelet, Journal of the Institution of Engineers (India): Series B 93 (2012) 63–66, 10.1007/s40031-012-0010-3. URL http://dx.doi.org/10.1007/s40031-012-0010-3 [9] E. Avci, Z. H. Akpolat, Speech recognition using a wavelet packet adap- tive network based fuzzy inference system, Expert Systems with Appli- cations 31 (3) (2006) 495 – 503. doi:10.1016/j.eswa.2005.09.058. [10] J.-D. Wu, B.-F. Lin, Speaker identification using discrete wavelet packet transform technique with irregular decomposition, Expert Systems with Applications 36 (2, Part 2) (2009) 3136 – 3143. [11] M. Vetterli, C. Herley, Wavelets and filter banks: Theory and design, IEEE Trans. Signal Proc. 40 (10) (1992) 2207–2232. 21si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. [12] N. Hess-Nielsen, M. V. Wickerhouser, Wavelets and Time-Frequency Analisys, Proceedings of the IEEE 84 (4) (1996) 523–540. [13] R. Munkong, B.-H. Juang, Auditory perception and cogni- tion, Signal Processing Magazine, IEEE 25 (3) (2008) 98–117. doi:10.1109/MSP.2008.918418. [14] S. Ray, A. Chan, Automatic feature extraction from wavelet coefficients using genetic algorithms, in: Proceedings of the 2001 IEEE Signal Pro- cessing Society Workshop, Neural Networks for Signal Processing XI, 2001, pp. 233–241. [15] D. Wang, D. Miao, C. Xie, Best basis-based wavelet packet entropy feature extraction and hierarchical eeg classification for epileptic detec- tion, Expert Systems with Applications 38 (11) (2011) 14314 – 14320. doi:10.1016/j.eswa.2011.05.096. [16] H. Rufiner, J. Goddard, A method of wavelet selection in phoneme recog- nition, in: Circuits and Systems, 1997. Proceedings of the 40th Midwest Symposium on, Vol. 2, 1997, pp. 889 –891 vol.2. [17] P. Kumar, M. Chandra, Hybrid of wavelet and mfcc features for speaker verification, in: Information and Communication Technolo- gies (WICT), 2011 World Congress on, 2011, pp. 1150 –1154. doi:10.1109/WICT.2011.6141410. [18] V. Tiwari, J. Singhai, Wavelet Based Noise Robust Features for Speaker Recognition, International Journal of Signal Processing 5 (2) (2011) 52 – 64. [19] A. Daamouche, L. Hamami, N. Alajlan, F. Melgani, A wavelet optimiza- tion approach for ecg signal classification, Biomedical Signal Processing and Control. In Press. doi:10.1016/j.bspc.2011.07.001. [20] A. R. Ferreira da Silva, Approximations with evolutionary pursuit, Sig- nal Processing 83 (3) (2003) 465–481. [21] R. Behroozmand, F. Almasganj, Optimal selection of wavelet- packet-based features using genetic algorithm in pathological assess- ment of patients’ speech signal with unilateral vocal fold paraly- sis, Computers in Biology and Medicine 37 (4) (2007) 474 – 485. doi:10.1016/j.compbiomed.2006.08.016. 22si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. [22] A. R. Ferreira da Silva, Wavelet denoising with evolutionary al- gorithms, Digital Signal Processing 15 (4) (2005) 382 – 399. doi:10.1016/j.dsp.2004.11.003. [23] E.-S. El-Dahshan, Genetic algorithm and wavelet hybrid scheme for ecg signal denoising, Telecommunication Systems 46 (2011) 209–215, 10.1007/s11235-010-9286-2. [24] R. Salvador, F. Moreno, T. Riesgo, L. Sekanina, Evolutionary Ap- proach to Improve Wavelet Transforms for Image Compression in Em- bedded Systems, EURASIP Journal on Advances in Signal Processing 2011 (2011). doi:10.1155/2011/973806. [25] W. Zhao, C. E. Davis, Swarm intelligence based wavelet coefficient feature selection for mass spectral classification: An application to proteomics data, Analytica Chimica Acta 651 (1) (2009) 15 – 23. doi:10.1016/j.aca.2009.08.008. [26] A. Daamouche, F. Melgani, Swarm intelligence approach to wavelet design for hyperspectral image classification, Geoscience and Remote Sensing Letters, IEEE 6 (4) (2009) 825 – 829. doi:10.1109/LGRS.2009.2026191. [27] W. Pedrycz, S. S. Ahmad, Evolutionary feature selection via structure retention, Expert Systems with Applications 39 (15) (2012) 11801 – 11807. doi:10.1016/j.eswa.2011.09.154. [28] S. Chatterjee, A. Bhattacherjee, Genetic algorithms for feature selection of image analysis-based quality monitoring model: An application to an iron mine, Engineering Applications of Artificial Intelligence 24 (5) (2011) 786 – 795. doi:10.1016/j.engappai.2010.11.009. [29] Y.-X. Li, S. Kwong, Q.-H. He, J. He, J.-C. Yang, Genetic algorithm based simultaneous optimization of feature subsets and hidden markov model parameters for discrimination between speech and non-speech events, International Journal of Speech Technology 13 (2010) 61–73, 10.1007/s10772-010-9070-4. [30] L. D. Vignolo, H. L. Rufiner, D. H. Milone, J. C. Goddard, Evolution- ary Splines for Cepstral Filterbank Optimization in Phoneme Classifica- 23si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. tion, EURASIP Journal on Advances in Signal Processing Volume 2011, doi:10.1155/2011/284791. [31] L. D. Vignolo, H. L. Rufiner, D. H. Milone, J. C. Goddard, Evolutionary Cepstral Coefficients, Applied Soft Computing 11 (4) (2011) 3419 – 3428. doi:10.1016/j.asoc.2011.01.012. [32] L. Vignolo, H. Rufiner, D. Milone, J. Goddard, Genetic optimization of cepstrum filterbank for phoneme classification, in: Proceedings of the Second International Conference on Bio-inspired Systems and Signal Processing (Biosignals 2009), INSTICC Press, Porto (Portugal), 2009, pp. 179–185. [33] L. Vignolo, D. Milone, H. Rufiner, E. Albornoz, Parallel implementation for wavelet dictionary optimization applied to pattern recognition, in: Proceedings of the 7th Argentine Symposium on Computing Technology, Mendoza, Argentina, 2006. [34] S. Durbha, R. King, N. Younan, Wrapper-based feature subset selection for rapid image information mining, Geoscience and Remote Sensing Letters, IEEE 7 (1) (2010) 43 –47. doi:10.1109/LGRS.2009.2028585. [35] H.-H. Hsu, C.-W. Hsieh, M.-D. Lu, Hybrid feature selection by com- bining filters and wrappers, Expert Systems with Applications 38 (7) (2011) 8144 – 8150. doi:10.1016/j.eswa.2010.12.156. [36] R. Kohavi, G. H. John, Wrappers for feature subset selection, Artificial Intelligence 97 (1-2) (1997) 273 – 324. doi:10.1016/S000437029700043X. [37] S. Mallat, A Wavelet Tour of signal Processing, 3rd Edition, Academic Press, 2008. [38] S. Mallat, A theory of multiresolution of signal decomposition: the wavelet representation, IEEE Trans. Pattern Anal. Machine Intell. 11 (7) (1989) 674–693. [39] R. Coifman, M. V. Wickerhauser, Entropy-based algorithms for best basis selection, IEEE Transactions on Information Theory 38 (2) (1992) 713–718. 24si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. [40] N. Saito, Local feature extraction and its applications using a library of bases, Ph.D. thesis, Yale University, New Haven, USA, director-Ronald R. Coifman (1994). [41] S. N. Sivanandam, S. N. Deepa, Introduction to Genetic Algorithms, Springer London, Limited, 2008. [42] H. Youssef, S. M. Sait, H. Adiche, Evolutionary algorithms, simulated annealing and tabu search: a comparative study, Engineering Applica- tions of Artificial Intelligence 14 (2) (2001) 167 – 181. doi:10.1016/S0952- 1976(00)00065-8. [43] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs, Springer-Verlag, 1992. [44] T. Kohonen, Improved versions of learning vector quantization, in: Proc. of the Int. Joint Conf. on Neural Networks, San Diego, 1990, pp. 545– 550. [45] A. Echeverŕıa, J. Tejedor, D. Wang, An evolutionary confidence measure for spotting words in speech recognition, in: Y. Demazeau, F. Dignum, J. Corchado, J. Bajo, R. Corchuelo, E. Corchado, F. Fernández-Riverola, V. Julián, P. Pawlewski, A. Campbell (Eds.), Trends in Practical Ap- plications of Agents and Multiagent Systems, Vol. 71 of Advances in Intelligent and Soft Computing, Springer Berlin / Heidelberg, 2010, pp. 419–427. [46] A. Moreno, D. Poch, A. Bonafonte, E. Lleida, J. Llisterri, J. Marino, C. Nadeu, Albayzin speech database design of the phonetic corpus, Tech. rep., Universitat Politecnica de Catalunya (UPC), Dpto. DTSC (1993). [47] S. Young, G. Evermann, D. Kershaw, G. Moore, J. Odell, D. Ollason, V. Valtchev, P. Woodland, The HTK book, HTK version 3.1, Cambridge University (2001). URL http://htk.eng.cam.ac.uk [48] A. Quilis, Tratado de Fonoloǵıa y Fonética Españolas, Biblioteca Románica Hispánica, Editorial Gredos, Madrid, 1993. 25si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. [49] H. Rufiner, Comparación entre análisis onditas y fourier aplicados al reconocimiento automático del habla, Master’s thesis, Universidad Autónoma Metropolitana, Iztapalapa (1996). [50] I. Daubechies, Ten Lectures on Wavelets, Society for Industrial and Applied Mathematics, Philadelphia, PA, USA, 1992. [51] M. Slaney, Auditory Toolbox, Version 2, Technical Report 1998-010, Interval Research Corporation, Apple Computer Inc. (1998). [52] T. Ganchev, N. Fakotakis, G. Kokkinakis, Comparative evaluation of various MFCC implementations on the speaker verification task, in: Proceedings of the SPECOM-2005, 2005, pp. 191–194. [53] X. D. Huang, Y. Ariki, M. A. Jack, Hidden Markov Models for Speech Recognition, Edinburgh University Press, 1990. [54] K. Demuynck, J. Duchateau, D. Van Compernolle, P. Wambacq, Im- proved Feature Decorrelation for HMM-based Speech Recognition, in: Proceedings of the 5th International Conference on Spoken Language Processing (ICSLP 98), Sydney, Australia, 1998. [55] F. Jelinek, Statistical Methods for Speech Recognition, MIT Press, Cam- brige, Masachussets, 1999. [56] D. H. Milone, L. E. D. Persia, M. E. Torres, Denoising and recognition using hidden Markov models with observation distributions modeled by hidden markov trees, Pattern Recognition 43 (4) (2010) 1577 – 1589. doi:10.1016/j.patcog.2009.11.010. [57] C. M. Bishop, Pattern Recognition and Machine Learning, 1st Edition, Springer, 2007. [58] M. Lewicki, Efficient coding of natural sounds, Nature Neuroscience 5 (4) (2002) 356–363. 26si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) L . D . V ig no lo , D . H . M il on e & H . L . R uf in er ; "G en et ic w av el et p ac ke ts f or s pe ec h re co gn it io n" E xp er t S ys te m s w it h A pp li ca ti on s, 2 01 2. 
work_eytfreollragxo26ezexsnsqqa	----	BioMed Central BMC Medical Informatics and Decision Making ss Open AcceResearch article Context-sensitive autoassociative memories as expert systems in medical diagnosis Andrés Pomi*1 and Fernando Olivera2 Address: 1Sección Biofísica, Facultad de Ciencias, Universidad de la República, Iguá 4225, 11400 Montevideo, Uruguay and 2Departamento de Biofísica, Facultad de Medicina, Universidad de la República, General Flores 2125,11800 Montevideo, Uruguay Email: Andrés Pomi* - pomi@fcien.edu.uy; Fernando Olivera - folivera@fmed.edu.uy * Corresponding author Abstract Background: The complexity of our contemporary medical practice has impelled the development of different decision-support aids based on artificial intelligence and neural networks. Distributed associative memories are neural network models that fit perfectly well to the vision of cognition emerging from current neurosciences. Methods: We present the context-dependent autoassociative memory model. The sets of diseases and symptoms are mapped onto a pair of basis of orthogonal vectors. A matrix memory stores the associations between the signs and symptoms, and their corresponding diseases. A minimal numerical example is presented to show how to instruct the memory and how the system works. In order to provide a quick appreciation of the validity of the model and its potential clinical relevance we implemented an application with real data. A memory was trained with published data of neonates with suspected late-onset sepsis in a neonatal intensive care unit (NICU). A set of personal clinical observations was used as a test set to evaluate the capacity of the model to discriminate between septic and non-septic neonates on the basis of clinical and laboratory findings. Results: We show here that matrix memory models with associations modulated by context can perform automatic medical diagnosis. The sequential availability of new information over time makes the system progress in a narrowing process that reduces the range of diagnostic possibilities. At each step the system provides a probabilistic map of the different possible diagnoses to that moment. The system can incorporate the clinical experience, building in that way a representative database of historical data that captures geo-demographical differences between patient populations. The trained model succeeds in diagnosing late-onset sepsis within the test set of infants in the NICU: sensitivity 100%; specificity 80%; percentage of true positives 91%; percentage of true negatives 100%; accuracy (true positives plus true negatives over the totality of patients) 93,3%; and Cohen's kappa index 0,84. Conclusion: Context-dependent associative memories can operate as medical expert systems. The model is presented in a simple and tutorial way to encourage straightforward implementations by medical groups. An application with real data, presented as a primary evaluation of the validity and potentiality of the model in medical diagnosis, shows that the model is a highly promising alternative in the development of accuracy diagnostic tools. Published: 22 November 2006 BMC Medical Informatics and Decision Making 2006, 6:39 doi:10.1186/1472-6947-6-39 Received: 15 May 2006 Accepted: 22 November 2006 This article is available from: http://www.biomedcentral.com/1472-6947/6/39 © 2006 Pomi and Olivera; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 1 of 11 (page number not for citation purposes) http://www.biomedcentral.com/1472-6947/6/39 http://creativecommons.org/licenses/by/2.0 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=17121675 http://www.biomedcentral.com/ http://www.biomedcentral.com/info/about/charter/ BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 Background The extreme complexity of contemporary medical knowl- edge together with the intrinsic fallibility of human rea- soning, have led to sustained efforts to develop clinical decision support systems, with the hope that bedside expert systems could overcome the limitations inherent to human cognition [1]. Despite the foundational hopes have not been fulfilled [2], the unaltered and increasing necessity for reliable automated diagnostic tools and the important benefit to society brought by any success in this area make every advance valuable. To further the research on computer-aided diagnosis begun in the 1960s, models of neural networks [3] have been added to the pioneering work on artificial-intelli- gence systems. The advent of artificial neural networks with ability to identify multidimensional relationships in clinical data might improve the diagnostic power of the classical approaches. A great proportion of the neural net- work architectures applied to clinical diagnosis rests on multilayer feed-forward networks instructed with back- propagation, followed by self-organizing maps and ART models [4,5]. Although they perform with significant accuracy, this performance nevertheless remained insuffi- cient to dispel the common fear that they are "black- boxes" whose functioning cannot be well understood and, consequently, whose recommendations cannot be trusted [6]. The associative memory models, an early class of neural models [7] that fit perfectly well with the vision of cogni- tion emergent from today brain neuroimaging techniques [8,9], are inspired on the capacity of human cognition to build semantic nets [10]. Their known ability to support symbolic calculus [11] makes them a possible link between connectionist models and classical artificial- intelligence developments. This work has three main objectives: a) to point out that associative memory models have the possibility to act as expert systems in medical diagnosis; b) to show in a sim- ple and straightforward way how to instruct a minimal expert system with associative memories; and c) to encourage the implementation of this methodology at large scale by medical groups. Therefore, in this paper we address – in a tutorial approach – the building of associative memory-based expert systems for the medical diagnosis domain. We favour a comprehensive way and the possibility of a straightforward implementation by medical groups over the mathematical details of the model. Methods Context-dependent autoassociative memories with overlapping contexts Associative memories are neural network models devel- oped to capture some of the known characteristics of human memories [12,13]. These memories associate arbi- trary pairs of patterns of neuronal activity mapped onto real vectors. The set of associated pairs is stored superim- posed and distributed throughout the coefficients of a matrix. These matrix memory models are content-address- able and fault-tolerant, and are well known to share with humans the ability of generalization and universalization [14]. In the attempt to overcome a serious problem of these classical models – their impossibility to evoke different associations depending on the context accompanying a same key stimulus- Mizraji [15] developed an extension of the model that performs adaptive associations. Context- dependent associations are based on a kind of second order sigma-pi neurons [16], and showed an interesting versatility when they were incorporated in modules employed to implement chains of goal-directed associa- tions [17], disambiguation of complex stimuli [18], logi- cal reasoning [19,20], and multiple criteria classification [21]. A context-dependent associative memory M acting as a basic expert system is a matrix where di are column vectors mapping k different diseases (the set {d} is chosen orthonormal), and sj(i) are column vectors mapping signs or symptoms accompanying the i disease (also an orthonormal set). The sets of symptoms corresponding to each disease can overlap. The Kronecker product (�) between two matrices A and B is another matrix defined by A � B = a(i, j)·B (2) denoting that each scalar coefficient of matrix A, a(i, j), is multiplied by the entire matrix B. Hence, if A is nxm dimensional and B is kxl dimensional, the resultant matrix will have the dimension nkxml. Note that if d are n-dimensional and s are k-dimensional vectors, the memory is a rectangular nxnm matrix. Also, the memory M can be viewed as resulting from the Kro- necker product (�) enlargement of each element of a nxn square autoassociative matrix di diT by a row column rep- resenting the sum of corresponding signs and symptoms: M d d si i j T j ii k = ⊗ ( )∑∑ = ( ) ( )1 1 Page 2 of 11 (page number not for citation purposes) BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 By feeding the context-sensitive autoassociative module M with signs or symptoms, the system retrieves the set of possible diseases associated with such set of symptoms, or a single diagnosis if the criteria suffice. At resting conditions the system is grounded in an indif- ferent state g. If each disease was instructed only one time, in the mathematics of the model this implies the priming of the memory with a linear combination in which every disease has an equal weight where and I is the nxn identity matrix. From (4) it is evident that, after the priming, the context-dependent memory becomes a classical memory associating symp- toms with diseases. If a set of sufficient concurrent signs and symptoms is presented to the waiting memory (σ = ∑s), after iteration, a final diagnosis results. It is important to point out that if the sets {sj(i)} corre- sponding to each disease were disjoint sets, then any sin- gle symptom s j(i) would be patognomonical and sufficient to univocally diagnose di. Otherwise, the output will be a linear combination of possible diseases, each one weighed according to the scalar product between the set of actual symptoms (σ) and the set of symptoms corre- sponding to each different disease: , σ > di. See Figure 1 and its legend. Forcing the sum of scalar products to unity, this output provides a probabilistic mapping of the possible diseases associated with the clinical presenta- tion. NUMERICAL EXAMPLE How to instruct the memory Let us illustrate how to instruct the memory with a mini- mal numerical example. Consider the set of three diseases and its characteristic symptoms shown in Figure 2. The first task is to codify the sets of signs and diseases with orthogonal vectors, for which we will use the following orthogonal matrices. According to the table and equation (1), we instruct the memory by adding a matrix for each disease. For the first disease we have d1d1T � (s1 + s3 + s4)T In the same way we will have two other matrices for the other diseases. The sum of the three matrices constitutes the memory M. How the system works See also Figure 1 and its legend. Time step 1 Initial state of the system: Indifferent vector gT = (d1 + d2 + d3) = [1 1 1] A first clinical data (s3) arrives: s3 T = [0.5 0.5 -0.5 -0.5] Preprocessing of input vectors is performed: h = g � s3 hT = [0.5 0.5 -0.5 -0.5 0.5 0.5 -0.5 -0.5 0.5 0.5 -0.5 -0.5] Resulting associated output: Mh (a linear combination of possible diagnoses) M d d si i T i k j T j i = ⊗ ( ) = ∑ ∑ 1 3 ( ) M g I d g d s d sn n i i i j j i T i i j j i T( ) , ( ) ( ) ( ) ( ) ⊗ = < > = ( )× ∑ ∑ ∑ ∑ 4 g di i = ∑ < ∑∑ sj j ii ( ) Diseases Signs & symptoms 1 0 0 0 1 0 0 0 1 0 ⎡ ⎣ ⎢ ⎢ ⎢ ⎤ ⎦ ⎥ ⎥ ⎥ .55 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 * − − − − − − ⎡ ⎣ ⎢ ⎢ ⎢ ⎢ ⎤ ⎦ ⎥ ⎥ ⎥ ⎥ d d d s s s s1 2 3 1 2 3 4 1 0 0 0 0 0 0 0 0 1 5 0 5 0 5 0 5 1 5 0 5 0 5 0 5 0 0 0 0 0 0 0 0 ⎡ ⎣ ⎢ ⎢ ⎢ ⎤ ⎦ ⎥ ⎥ ⎥ ⊗ − = = − [ . . . . ] . . . . 00 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ⎡ ⎣ ⎢ ⎢ ⎢ ⎤ ⎦ ⎥ ⎥ ⎥ M = − − − ⎡1 5 0 5 0 5 0 5 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 5 0 5 0 5 0 5 . . . . . . . .⎣⎣ ⎢ ⎢ ⎢ ⎤ ⎦ ⎥ ⎥ ⎥ output( )1 1 1 0 = ⎡ ⎣ ⎢ ⎢ ⎢ ⎤ ⎦ ⎥ ⎥ ⎥ Page 3 of 11 (page number not for citation purposes) BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 Resulting probabilistic map (each coefficient of the out- put vector is divided by the sum of them all): Time step 2 A new symptom (s2) arrives: s2 T = [0.5 -0.5 0.5 -0.5] Preprocessing of input vectors is performed: h = output(1) � s2 hT = [0.5 -0.5 0.5 -0.5 0.5 -0.5 0.5 -0.5 0 0 0 0] Resulting associated output (Mh): Resulting probabilistic map: Final result The system has arrived to an only final diagnosis that cor- responds to disease 2. REAL DATA APPLICATION – diagnosing late-onset neonatal sepsis Late-onset sepsis (invasive infection occurring in neonates after 3 days of age) is an important and severe problem in infants hospitalized in neonatal intensive care units (NICUs) [22]. The clinical signs of infection in the new- born are variable, and the earliest manifestations are often subtle and nonspecific. In the presence of a clinical suspi- cion of sepsis an early and accurate diagnosis algorithm would be of outstanding value but is not yet available [23]. In a recent retrospective study that included 47 neonates with clinical diagnosis of suspected sepsis, Mar- tell and collaborators [24] assessed a group of clinical and laboratory variables – surgical history, metabolic acidosis, hepatomegalia, abnormal white blood cell (WBC) count, hyperglycemia and thrombocytopenia-determining their sensitivity, specificity, likelihood ratio and post-test prob- prob( ) . .1 0 5 0 5 0 = ⎡ ⎣ ⎢ ⎢ ⎢ ⎤ ⎦ ⎥ ⎥ ⎥ output( )2 0 1 0 = ⎡ ⎣ ⎢ ⎢ ⎢ ⎤ ⎦ ⎥ ⎥ ⎥ prob( )1 0 1 0 = ⎡ ⎣ ⎢ ⎢ ⎢ ⎤ ⎦ ⎥ ⎥ ⎥ A module with diagnostic abilitiesFigure 1 A module with diagnostic abilities. The neural module receives the input of two vectors: one representing the set of pos- sible diseases up to the moment and the other vector corresponding to a new sign, symptom or laboratory result. The action of the neurons that constitute the neural module can be divided into two sequential steps: the Kronecker product of these two entries and the association of this stimulus with an output activity pattern. This output vector is a linear combination of a nar- rower set of disease vectors that can be reinjected if a new clinical data arrives or can be processed to obtain the probability attributable to each diagnostic decision. signs & symptoms diseases neural module M output probabilistic map Page 4 of 11 (page number not for citation purposes) BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 ability. Sepsis was defined as a positive result on one or more blood cultures in a neonate with clinical diagnosis of suspected sepsis. A prevalence of 34% was found for their NICU. We instructed a context-dependent autoassociative mem- ory according to equation (3) with data published in [24] in order to evaluate its capacity to recognize patients with or without sepsis. As a test-set, we used 15 cases of sus- pected neonatal sepsis coming from the same NICU (per- sonal observations of one of us-AP). From equation (3) it is clear that the different clinical presentation of the indi- vidual cases are added up and resumed in the vector ( ) representing the characteristic signs of each ill- ness condition. We trained the memory instructing two terms di corresponding to the two final diagnoses of con- firmed sepsis and absence of sepsis. M = [septic] [septic]T � [attributes _ septic]T + [healthy] [healthy]T � [attributes _ healthy]T The column vectors used for the septic and healthy condi- tions were [1 0]T and [0 1]T respectively. The column vectors with the attributes corresponding to the septic and non-septic patients were generated from the available data as follows. For each one of the variables studied in [24] (see Figure 3) we reconstructed the values of the true positive (TP), false positive (FP), true negative (TN) and false negative (FN) number of patients: TP = sensitivity × E FP = (sensitivity/LR) × NE TN = specificity × NE FN = N - (TP+FP+TN). These values became the coefficients of the two thirteen- dimensional column vectors [attributes_septic] and [attributes_healthy]. This procedure is shown in Figure 4. Finally, after normalization, the vectors used for the instruction of the memory M are [attributes_septic]T = [0.0604 0.4225 0.0604 0.4225 0.1509 0.3320 0.0604 0.0604 0.3621 0.0604 0.4225 0.0604 0.4225] [attributes_healthy]T = [0.0142 0.4248 0.0142 0.4248 0.0566 0.3823 0.0354 0.0177 0.3859 0.0283 0.4106 0.0283 0.4106]. The memory M resumes the cumulated experience in sus- pected late-onset sepsis of this particular NICU through the clinical presentations of one year hospitalized neonates. To test our system we presented to the memory a set of 15 personal clinical observations of neonates with the clini- cal diagnosis of suspected sepsis hospitalized in the same NICU. We coded the thirteen attributes with the canonical basis vectors (the columns of a 13-dimensional identity matrix). For example, the presence of metabolic acidosis was represented with [0 01 0 0 0 0 0 0 0 0 0 0]T and the absence of acidosis with the vector [0 0 01 0 0 0 0 0 0 0 0 0]T. For each patient of the test set we added the vectors corresponding to the confirmed presence or absence of any sign. These 15 vectors representing the clinical presen- tation of the neonates with the diagnosis of suspected sep- sis are shown in Figure 5. The classification of each patient was obtained as follows: sj T j i( ) ∑ Three diseases with their corresponding signsFigure 2 Three diseases with their corresponding signs. The set of signs and symptoms associated with three different diseases instructed in the memory of the numerical example according to equation (1). Signs & symptoms s1 s2 s3 s4 d1 d2 D is ea se s d3 Page 5 of 11 (page number not for citation purposes) BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 i) The vector with the clinical presentation is presented to the memory M. The output, [result_vector], is a linear combination of the vectors septic [1 0]T and healthy [0 1]T: [result_vector] = M * ([indifferent_vector] � [clinical pres- entation]) The [indifferent_vector] is the sum of septic and healthy vectors: [1 1]T. ii) A diagnosis results from the evaluation of the coeffi- cients of the two-dimensional [result_vector]. If the first coefficient is greater than the second the case is classified as sepsis. If the second coefficient is the largest the patient is classified as non-septic. Results A context-dependent memory model acting as a minimal expert system In this work we show a minimal, context-dependent, memory nucleus able to support diagnostic abilities. Our expert system consists of an autoassociative memory with overlapping contexts and feedback loop that makes the output able to be reinjected into the memory at the next time step (Figure 1). A memory M acting as a basic expert system is a matrix (equation 3) where the di are column vectors mapping k different dis- eases (the set {d} is chosen to be orthonormal), sj(i) are column vectors mapping signs and symptoms accompa- nying the i disease (also an orthonormal set), and � is the Kronecker product [25]-see Methods-. Note that if d are n- dimensional vectors (n ≥ k), and s are m-dimensional, then didiT are square symmetric matrices, and the memory M is a rectangular matrix of dimensions nxnm. The instruction of the expert The cognitive functioning shown by this kind of neural network model is based on the establishment of context- dependent associations. The instruction of the expert therefore consists in the instruction of the memory that stores these associations. Each disease is instructed to the memory together with its characteristic signs and symptoms (these can include the results of laboratory exams, imaging studies, etc). For this to be done, the first step is to code each disease to be instructed with a different orthonormal vector. The same must be done with the set of signs, symptoms and para- clinical results that could accompany that set of diseases, also coding them with different column vectors of any orthonormal basis of adequate dimension. Once the signs and symptoms corresponding to each dis- ease have been identified and expressed as orthogonal vectors, the construction of the memory can commence. According to equation (1) this instruction consists in the superposition (the addition) of different rectangular matrices, each one corresponding to a different disease. The instruction of the memory can be developed along two different paths. a. Learning from the textbook. In this case, the expert is instructed according to the updated aca- demic knowledge of each disease. One first disease is taken, which is coded by the column vector di, and the outer product of this vector is made by itself (a square matrix is constructed that contains this autoassociation). At the same time, all the signs and symptoms characteris- tic of this disease are identified and the vectors coding them are added up ( ). Finally the Kronecker product between the square matrix and the transpose of the vector- sum is performed. An analogous procedure is accom- plished for any pathology. Each new resulting rectangular matrix of dimension nxnm is added to the previous ones already stored in memory M (a minimal numerical exam- M d d si i T i k j T j i = ⊗ = ∑ ∑ 1 ( ) , sj j i( ) ∑ Data from the study of Martell and collaborators [24]Figure 3 Data from the study of Martell and collaborators [24]. The total number of neonates with suspected late- onset sepsis was N = 47 (septic: E = 16; non-septic: NE = 31). Metabolic acidosis was defined for patients with ade- quate ventilation; WBC count was considered normal within 5.000–25.000; thrombocytopenia was defined for platelet counts <40.000. sensibility specificity LR Surgery 0,13 0,96 3,25 Acidosis 0,13 0,96 3,25 Hepatomegalia 0,31 0,88 2,58 Leucocytosis 0,13 0,92 1,60 Leucopenia 0,13 0,96 3,25 Hyperglycemia 0,13 0,92 1,62 Thrombocytopenia 0,13 0,92 1,62 Page 6 of 11 (page number not for citation purposes) BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 ple is presented in section Methods-How to instruct the memory-). b. Learning by experience. This is a case-based way of instructing the memory. It allows the expert to pro- gressively capture the prevalence of the different diseases in a community. Once finalized the previous instruction, the memory is fed with the actual clinical findings of each particular patient assisted by the physician, attributing this particular constellation of signs and symptoms to the corresponding final diagnosis. The matrices resulting from new patients are progressively added to the memory. This type of representation implies two essential distinc- tions from the previous learning-from-the-textbook mem- ory. Pathologies are not equally weighed in the memory but their representations depend on the frequency of pres- entation of cases in the population. In addition, for each disease the different symptoms also are not equally weighed: those corresponding to the more frequent clini- cal presentations will be strengthened. Medical queries Once the training phase is finalized, the system is ready to be used. The presentation of a first sign or symptom initi- ates a medical query. The availability of a new clinical or laboratory finding causes the expert to advance one more step in its diagnostic decision. Although we have many new signs and symptoms, in order to obtain a progressive narrowing of the set of possible diagnoses they must be presented to the expert one per time. At each step, the new data are entered into the memory along with the set of possible diagnoses until that moment. Finally, if the whole set of signs and symptoms available until the moment is sufficient, the system will arrive to a unique diagnosis. We then follow the system operation. The starting point is when the first clinical data appears. The vector corre- sponding to this symptom is multiplied by means of the Kronecker product times the vector that represents the set of possible diagnoses (in the starting point it is an indif- ferent vector). If the memory was instructed with equally- weighed pathologies the indifferent vector is the sum of Attributes' vectors for the septic and non-septic groupsFigure 4 Attributes' vectors for the septic and non-septic groups. For each variable, true positive (TP), false positive (FP), true negative (TN) and false negative (FN) were calculated: TP = sensitivity × E; FP = (sensitivity/LR) × NE; TN = specificity × NE; FN = N - (TP+FP+TN). Presence or absence of the sign [attributes_septic] [attributes_healthy] (+) TP FPSurgery (–) FN TN (+) TP FPMetabolic acidosis (–) FN TN (+) TP FPHepatomegalia (–) FN TN TP (a1) FP (b1) TP (a2) FP (b2) Leucocytosis Leucopenia Normal WBC count E – (a1+a2) NE – (b1+b2) (+) TP FPHyperglycemia (–) FN TN (+) TP FPThrombocytopenia (–) FN TN Page 7 of 11 (page number not for citation purposes) BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 all the vectors of diseases stored in the memory. If, on the contrary, the memory was instructed on the basis of indi- vidual cases, the indifferent vector will be the same linear combination of the vector diseases stored (the weight of each disease corresponds to the one of its frequency of presentation). The resulting column vector is now multi- plied by the memory matrix. The exit vector contains either a univocal diagnosis (if the clinical data are suffi- cient) or a certain linear combination of vectors corre- sponding to several diseases. If a unique diagnosis was not arrived at, when one has a new sign or symptom, its cor- responding vector will enter the memory after making its Kronecker product by the exit vector of the previous step. The process is repeated and stops when a final diagnosis is reached or when new clinical data is not available (see the continuation of the numerical example in section Meth- ods-How the system works-). Even if at a certain state a final diagnosis has not been reached, the outcome of the system nevertheless repre- sents a probabilistic mapping of the possible diagnoses, each one with its respective probability in agreement with the data available until the moment. In order to obtain such a map in a direct way it is convenient to choose as disease vectors the columns of an identity matrix of suita- ble dimension. In that case, in each exit vector the posi- tions of the coefficients different from zero mark the different possible diagnoses. Applying a normalization to this exit vector in such a way that the sum of their compo- nents is one, the value of each coefficient different from zero represents the probability of each one of those diag- noses. Otherwise, these probabilities can be obtained by multiplying the exit vector by the orthonormal matrix that codifies the diseases. A reduced model for the diagnosis of late-onset neonatal sepsis The system described in section Methods classified the patients of the test-set (N = 15) as follows (S = sepsis; NS = non-septic): 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 S NS NS S S S S S NS NS S S S S S Test setFigure 5 Test set. For each one of the 15 neonates of the test set, the thirteen-dimensional column vector has ones in the coefficients corresponding to the position of the confirmed presence or absence of the signs already shown in Figure 4. The final diagnosis is shown at the bottom of the column (S: sepsis; NS: no sepsis). Case number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 0 0 0 1 1 0 0 0 0 0 0 1 0 0 0 1 1 1 0 0 1 1 1 1 1 1 0 1 1 1 1 1 0 0 1 0 0 0 0 0 0 0 1 0 1 0 0 1 1 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 1 1 1 0 0 0 0 1 1 0 0 0 1 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 1 0 0 0 0 1 1 0 0 0 1 0 0 0 0 0 1 1 0 0 0 0 0 1 1 0 1 1 0 1 0 0 0 1 1 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 1 1 0 1 0 0 0 0 1 1 0 0 0 1 0 1 0 0 0 1 1 1 1 0 0 1 0 1 0 1 0 1 1 1 0 0 0 0 1 1 0 0 0 1 0 Illness condition S NS NS S S S S S NS NS S S S NS S Presence of a sign Absence of a sign Page 8 of 11 (page number not for citation purposes) BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 Comparing this classification with the actual illness con- dition of the patients-shown in Figure 5 – it results that only patient 14 was misdiagnosed. The 2 × 2 table shown in Figure 6 resumes the behaviour of our diagnostic sys- tem. The sensitivity was 100% and the specificity 80%. The likelihood ratio (LR = TP/FP) was 5. Using this set of variables as input data, the performance of the system in the classification task can be evaluated as very good. It reached a high accuracy ((TP+TN)/N) of 93.3%, and a Cohen's kappa index of 0.84. (Kappa = (Accuracy - A_chance)/(1-A_chance), where A_chance is the accuracy expected by chance. A_chance = (<TP>+<TN>)/N) where <TP> = (TP+FP)(TP+FN)/N and <TN> = (FN+TN)(FP+TN)/N). Discussion and conclusions We have shown here that context-dependent associative memories could act as medical decision support systems. The system implies the previous coding of a set of diseases and its corresponding semiologic findings in individual basis of orthogonal vectors. The model presented in this communication is only a minimal module able to evalu- ate the probabilities of different diagnoses when a set of signs and symptoms is presented to it. This expert system based on an associative memory shares with programs using artificial intelligence a great capacity to quickly narrow the number of diagnostic possibilities [1]. Also, it is able to cope with variations in the way that a disease can present itself. A clear advantage of this system is that the probability assignment to the different diagnostic possibilities in any particular clinical situation does not have to be arbitrarily assigned by the specialist, but is automatically provided by the system, in agreement with the acquired experience. In this sense, this neural network model is akin to statisti- cal pattern recognition [1]. However, neither programs based on simple matching strategies nor most used neural network models are able to explain to the physician how they have reached their conclusions. On the contrary, the operation of this system, that unveils the underlying asso- ciative structure of human cognition, is transparent. Obvi- ously, it must be understood that this is not the unique mechanism involved in human decision making. The rel- evant properties of this associative memory model are summarized in Figure 7 in comparison to other neural network models and rule-based artificial intelligence sys- tems. Beginning with a textbook-instructed memory, the system evolves accommodating (superimposing in the memory) new manifestations of disease gathered over time. This process of continued network education based on empir- ical evidence leads to databases representative of the dif- ferent patient populations with its own geo- demographical characteristics. This model can be easily improved in various directions. The functioning of the system described up to now can be considered a passive phase (in the sense that it consists on an automatic evaluation of the available information). By adding another module to the system, consisting of a sim- ple memory that associates diseases with the set of its find- ings, the expert can enhance its diagnostic performance. Remaining two or three different diagnostic hypothesis Associative memory classificationFigure 6 Associative memory classification. The 15 cases tested (actual sepsis: 10; non-septic: 5) were classified as positive (S) or negative (NS) by the neural network. TP = true positive; FP = false positive; TN = true negative and FN = false negative. A total of 11 neonates were tested positive, and 4 negative. The sensitivity was 100% and the specificity 80%. Actual illness condition Septic Non-septic S TP = 10 FP = 1 11 NS FN = 0 TN = 4 4 System classification 10 5 Page 9 of 11 (page number not for citation purposes) BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 within the previous passive phase of diagnosis refine- ment, this new module can be fed with the vectors map- ping each one of these diseases to elicit its associated set of clinical findings. The set of absent features supporting one or the other disease determines what information must be sought next. Another important expansion of the expert allows giving up the strong assumption that all the findings correspond to a unique disease. Our context-dependent memory stops and gives a null vector when contradictory data are proportioned. To prevent such behaviour, a module akin to a novelty filter could be interposed within the recursion with the following properties: if a vector with only zero coefficients arrives, this module associates the whole set of diseases, avoiding lying aside relevant diagnoses and concurrent pathologies. However, this theme needs fur- ther investigation: as for almost every expert system [26], the clustering of findings and their attribution either to only one disease or to several disorders is a major chal- lenge. The primary implementation of a reduced version of the model with the aim of classifying septic or non-septic neonates showed the highly satisfactory capacity of the model to be applied to real data. We conclude that con- text-sensitive associative memory model is a promising alternative in the development of accuracy diagnostic tools. We expect that its easy implementation stimulate groups of medical informatics to develop this expert sys- tem at real scale. Competing interests The authors declare that they have no competing interests. Authors' contributions AP conceived the application of the model to medical diagnosis, drafted the manuscript and carried out the implementation with real data of neonates with suspected late-onset sepsis. FO participated in the elaboration of the numerical examples, computational programs and the discussion of the model. Both authors read and approved the final manuscript. Acknowledgements We thank Dr. Eduardo Mizraji for useful comments and Dr. Julio A. Hernández for revision and improvement of the manuscript. References 1. Szolovits P, Patil RS, Schwartz WB: Artificial Intelligence in med- ical diagnosis. Annals of Internal Medicine 1988, 108:80-87. 2. Schwartz WB, Patil RS, Szolovits P: Artificial Intelligence in med- icine: Where do we stand? New England Journal of Medicine 1987, 316:685-688. 3. Arbib MA, Ed: The Handbook of Brain Theory and Neural Networks Cam- bridge, MA: MIT Press; 1995. 4. Cross SS, Harrison RF, Lee Kennedy R: Introduction to neural networks. The Lancet 1995, 346:1075-1079. 5. Lisboa PJG: A review of evidence of health benefit from artifi- cial neural network in health intervention. Neural Networks 2002, 15:11-39. 6. Baxt WG: Application of artificial neural networks to clinical medicine. The Lancet 1995, 346:1135-1138. 7. Kohonen T: Associative Memory: A System-Theoretical Approach New York: Springer-Verlag; 1977. 8. Friston KJ: Imaging neuroscience: Principles or maps? Proc Natl Acad Sci USA 1998, 95:796-802. 9. McIntosh AR: Towards a network theory of cognition. Neural Networks 2000, 13:861-870. 10. Pomi A, Mizraji E: Semantic graphs and associative memories. Physical Review E 2004, 70:066136. 11. Mizraji E: Vector logics: the matrix-vector representation of logical calculus. Fuzzy Sets and Systems 1992, 50:179-185. 12. Anderson JA, Cooper L, Nass MM, Freiberger W, Grenander U: Some properties of a neural model for memory. AAAS Sympo- sium on Theoretical Biology and Biomathematics 1972 [http://www.phys ics.brown.edu/physics/researchpages/Ibns/ Cooper%20Pub040_SomePropertiesNeural_72.pdf]. Milton, WA. Leon N Cooper Publications Outstanding characteristics of different modelsFigure 7 Outstanding characteristics of different models. AM: associative memory; ANN: artificial neural network; AI: artificial intelligence. AM model Other ANN models Rule-based AI systems Dealing with non linear information + + - Narrowing diagnostic possibilities + + + Automatic assignment of probabilities + - - Explaining capacity + - + Page 10 of 11 (page number not for citation purposes) http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=3276267 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=3276267 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=3821801 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=3821801 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=11958484 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=11958484 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=9448243 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=11156197 http://www.physics.brown.edu/physics/researchpages/Ibns/Cooper%20Pubs/040_SomePropertiesNeural_72.pdf http://www.physics.brown.edu/physics/researchpages/Ibns/Cooper%20Pubs/040_SomePropertiesNeural_72.pdf http://www.physics.brown.edu/physics/researchpages/Ibns/Cooper%20Pubs/040_SomePropertiesNeural_72.pdf BMC Medical Informatics and Decision Making 2006, 6:39 http://www.biomedcentral.com/1472-6947/6/39 Publish with BioMed Central and every scientist can read your work free of charge "BioMed Central will be the most significant development for disseminating the results of biomedical researc h in our lifetime." Sir Paul Nurse, Cancer Research UK Your research papers will be: available free of charge to the entire biomedical community peer reviewed and published immediately upon acceptance cited in PubMed and archived on PubMed Central yours — you keep the copyright Submit your manuscript here: http://www.biomedcentral.com/info/publishing_adv.asp BioMedcentral 13. Cooper LN: Memories and memory: a physicist's approach to the brain. International J Modern Physics A 2000, 15:4069-4082 [http:/ /journals.wspc.com.sg/ijmpa/15/1526/S0217751X0000272X.html]. 14. Cooper LN: A Possible Organization of Animal Memory and Learning. In Proceedings of the Nobel Symposium on Collective Proper- ties of Physical Systems Edited by: Lundquist B & S. New York: Aca- demic Press; 1973. 15. Mizraji E: Context-dependent associations in linear distrib- uted memories. Bulletin Math Biol 1989, 51:195-205. 16. Valle-Lisboa JC, Reali F, Anastasía H, Mizraji E: Elman topology with sigma-pi units: An application to the modelling of verbal hallucinations in schozophrenia. Neural Networks 2005, 18:863-877. 17. Mizraji E, Pomi A, Alvarez F: Multiplicative contexts in associa- tive memories. BioSystems 1994, 32:145-161. 18. Pomi-Brea A, Mizraji E: Memories in context. BioSystems 1999, 50:173-188. 19. Mizraji E, Lin J: A dynamical approach to logical decisions. Com- plexity 1997, 2:56-63. 20. Mizraji E, Lin J: Fuzzy decisions in modular neural networks. Int J Bifurcation and Chaos 2001, 11:155-167. 21. Pomi A, Mizraji E: A cognitive architecture that solves a prob- lem stated by Minsky. IEEE on Systems, Man and Cybernetics B (Cybernetics) 2001, 31:729-734. 22. Stoll BJ, Hansen N, Fanaroff AA, Wright LL, Carlo WA, Ehrenkranz RA, Lemons JA, Donovan EF, Stark AR, Tyson JE, Oh W, Bauer CR, Korones SB, Shankaran S, Laptook AR, Stevenson DK, Papile L-A, Poole WK: Late-Onset Sepsis in Very Low Birth Weight Neonates: The Experience of the NICHD Neonatal Research Network. Pediatrics 2002, 110:285-291. 23. Rubin LG, Sánchez PJ, Siegel J, Levine G, Saiman L, Jarvis WR: Evalu- ation and Treatment of Neonates with Suspected Late- Onset Sepsis: A Survey of Neonatologists' Practices. Pediat- rics 2002, 110(4):e42. 24. Perotti E, Cazales C, Martell M: Estrategias para el diagnóstico de sepsis neonatal tardía. Rev Med Uruguay 2005, 21:314-320 [http://www.rmu.org.uy/revista/2005v4/art11.pdf]. 25. Van Loan CF: The ubiquitous Kronecker product. Journal of Computational and Applied Mathematics 2000, 123:85-100. 26. Szolovits P, Pauker SG: Categorical and probabilistic reasoning in medicine revisited. Artificial Intelligence 1993, 59:167-180. Pre-publication history The pre-publication history for this paper can be accessed here: http://www.biomedcentral.com/1472-6947/6/39/prepub Page 11 of 11 (page number not for citation purposes) http://journals.wspc.com.sg/ijmpa/15/1526/S0217751X0000272X.html http://journals.wspc.com.sg/ijmpa/15/1526/S0217751X0000272X.html http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15935616 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15935616 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=15935616 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=7919113 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=7919113 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=10400268 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=12165580 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=12165580 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=12165580 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=12359815 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=12359815 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list_uids=12359815 http://www.rmu.org.uy/revista/2005v4/art11.pdf http://www.biomedcentral.com/1472-6947/6/39/prepub http://www.biomedcentral.com/ http://www.biomedcentral.com/info/publishing_adv.asp http://www.biomedcentral.com/ Abstract Background Methods Results Conclusion Background Methods Context-dependent autoassociative memories with overlapping contexts NUMERICAL EXAMPLE How to instruct the memory How the system works Final result REAL DATA APPLICATION - diagnosing late-onset neonatal sepsis Results A context-dependent memory model acting as a minimal expert system The instruction of the expert Medical queries A reduced model for the diagnosis of late-onset neonatal sepsis Discussion and conclusions Competing interests Authors' contributions Acknowledgements References Pre-publication history 
work_eyuc5thn3vajbhkgbinle2hjg4	----	Discussion related to ”Wang, C.-H., & Lu J.-Z. (2009). A hybrid genetic algorithm that optimizes capacitated vehicle routing problem. Expert Systems with Applications, 36(2), 2921-2936” Eneko Osaba, Roberto Carballedo, Fernando Diaz, Asier Perallos Deusto Institute of Technology (DeustoTech), University of Deusto, Av. Universidades 24, Bilbao 48007, Spain Abstract This paper presents a discussion arisen after reading ”A hybrid genetic algorithm that optimizes capacitated vehicle routing problem”, by Chung-Ho Wang and Jiu-Zhang Lu, (2009). Expert System with Applications (35)(pp. 2921-2936). The discussed paper presents a hybrid genetic algorithm applied to the Capacitated Vehicle Routing Problem (CVRP). When the authors present the results obtained by the technique, they claim to have overcome the best-known solution in two instances of Christofides and Eilon CVRP Benchmark. This statement can create confusion and controversy, for several reasons that we will explain and clarify in this short communication. Keywords: Capacitated Vehicle Routing Problem, Optimization, Benchmark, Best Known Solution 1. Discussion The discussed paper (Wang & Lu (2009)) presents a hybrid genetic algorithm solving the well-known Capacitated Vehicle Routing Problem (CVRP). To prove the quality of the proposed technique, the authors carry out the good practice of applying their approach to a benchmark, in this case, the CVRP Benchmark of Christofides and Eilon (Diaz (2012)). At the time of analysing the results obtained by their hybrid genetic algorithm, they introduce the following affirmation: ”Optimization results for the E-n30-k3 problem with 29 distribution points and the E-n51-k5 problem with 50 distribution points both exceed the current best known value” This statement may be a source of controversy and discussion if it is not explained properly. In this case, we will focus only on the instance E-n30-k3 or, called in other way, Eil30. In the benchmark where we can find this instance, we can see that the optimal solution for this instance is 534 or 538.958, depending on whether the distances are Email addresses: e.osaba@deusto.es (Eneko Osaba), roberto.carballedo@deusto.es (Roberto Carballedo), fernando.diaz@deusto.es (Fernando Diaz), perallos@deusto.es (Asier Perallos) Preprint submitted to Expert System With Applications April 19, 2013 calculated with rounded integers or decimal values. This optimal solution is taken as a reference in many works, as Prasad (2011); De CT Gomes & Von Zuben (2002); Campos & Mota (2000); Chen et al. (2006); Fukasawa et al. (2006) or Baldacci et al. (2004), and has never been exceeded in any of them. Thus, the optimal solution recognized by the scientific community is composed by 3 routes, as follows: Route 1 : 1 - 19 - 15 - 16 - 13 - 7 - 17 - 9 - 14 - 8 - 12 - 11 - 10 - 23 - 18 - 1 Route 2 : 1 - 26 - 28 - 27 - 29 - 25 - 25 - 24 - 6 - 21 - 1 Route 3 : 1 - 22 - 3 - 4 - 1 - 5 - 2 - 20 - 1 Despite this, the discussed paper offers a solution for this instance which greatly improves the best known value. This solution has a value of 508.14, and it is composed by 4 routes, as follows: Route 1 : 1 - 19 - 24 - 11 - 12 - 13 - 9 - 15 - 10 - 18 - 8 - 14 - 17 - 16 - 1 Route 2 : 1 - 21 - 4 - 5 - 6 - 2 - 7 - 25 - 26 - 30 - 28 - 29 - 27 - 20 - 1 Route 3 : 1 - 3 - 23 - 1 Route 4 : 1 - 22 - 1 Viewing this solution, the question is: Is this solution really better than the optimal known one? The answer to this question is in the objective function used. To obtain the best known solution (534 or 538.958) an objetive function that prioritizes the minimiza- tion of the routes is used, which leaves the minimization of the distance as the second objective. So, if the first objective is to reduce the number of vehicles used, the solution found by the authors of the discussed paper would not overcome the current best value, since their solution needs one more vehicle. Anyhow, the authors of the paper we are dealing use an objective function which priority is uniquely to minimize the distance. Thus, the solution to improve is not the one that the benchmark offers as the best value (534 or 538.958). In this case, it would have to be compared with a solution obtained with the same objective function. This solution is the one that we show below, which has been found previously in other works, such as Juan et al. (2010). It is composed of 4 routes and it has a value of 503 or 505.011: Route 1 : 1 - 19 - 24 - 11 - 12 - 13 - 9 - 15 - 10 - 18 - 8 - 14 - 17 - 16 - 1 Route 2 : 1 - 20 - 7 - 2 - 25 - 26 - 30 - 28 - 29 - 27 - 1 Route 3 : 1 - 21 - 4 - 5 - 6 - 3 - 23 - 1 Route 4 : 1 - 22 - 1 So, in conclusion, we can say that the statement introduced by the authors of the discussed paper is not true, since they have not made the comparison with the appropiate solution. Anyway our intention with this paper is not to accuse the authors of performing a bad practice. In fact, this confusion has occurred previously in the literature, for example, in De Backer et al. (1997) and Juan et al. (2010). Our main objective, related with our previous work Osaba & Carballedo (2012), is to help the creation of reliable and detailed benchmarks. In this case, in the Christofides and Eilon benchmark, the two optimal solutions for the E-n30-k3 should be shown, specifying the objective function 2 used to obtain each of them. This way, researchers will not have doubts to compare their solutions, and confusions like that happens in the discussed paper could be avoided. References Baldacci, R., Hadjiconstantinou, E., & Mingozzi, A. (2004). An exact algorithm for the capacitated vehicle routing problem based on a two-commodity network flow formulation. Operations Research, 52 , 723–738. Campos, V., & Mota, E. (2000). Heuristic procedures for the capacitated vehicle routing problem. Computational Optimization and Applications, 16 , 265–277. Chen, A.-l., Yang, G.-k., & Wu, Z.-m. (2006). Hybrid discrete particle swarm optimization algorithm for capacitated vehicle routing problem. Journal of Zhejiang University-Science A, 7 , 607–614. De Backer, B., Furnon, V., Prosser, P., Kilby, P., & Shaw, P. (1997). Local search in constraint pro- gramming: Application to the vehicle routing problem. In Constraint Programming 97, Proceedings of Workshop on Industrial Constraint-Directed Scheduling. De CT Gomes, L., & Von Zuben, F. J. (2002). A neuro-fuzzy approach to the capacitated vehicle routing problem. In Proceedings of the International Joint Conference on Neural Networks. IJCNN’02 (pp. 1930–1935). volume 2. Diaz, B. (2012). Vrp web. http://neo.lcc.uma.es/radi-aeb/Web-VRP, . Fukasawa, R., Longo, H., Lysgaard, J., Aragão, M. P. d., Reis, M., Uchoa, E., & Werneck, R. F. (2006). Robust branch-and-cut-and-price for the capacitated vehicle routing problem. Mathematical programming, 106 , 491–511. Juan, A., Faulin, J., Jorba, J., Riera, D., Masip, D., & Barrios, B. (2010). On the use of monte carlo simulation, cache and splitting techniques to improve the clarke and wright savings heuristics. Journal of the Operational Research Society, 62 , 1085–1097. Osaba, E., & Carballedo, R. (2012). A methodological proposal to eliminate ambiguities in the compar- ison of vehicle routing problem solving techniques. In International Joint Conference on Computa- tional Intelligence, IJCCI 2012 (pp. 310–313). Prasad, P. (2011). Optimization of capacitated vehicle routing problem by nested particle swarm opti- mization. American Journal of Applied Sciences, 8 , 107–112. Wang, C.-H., & Lu, J.-Z. (2009). A hybrid genetic algorithm that optimizes capacitated vehicle routing problems. Expert Systems with Applications, 36 , 2921–2936. 3 
work_eyuq3jyx6fazveygjt2ntiutji	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220996790 Params is empty 220996790 exception Params is empty 2021/04/06-03:05:41 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220996790 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:41 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_ez3sps3eubcgxmhkxgye5viov4	----	Microsoft Word - 01_2010.docx Edinburgh Research Explorer Are rating agencies' assignments opaque? Evidence from international banks Citation for published version: Bellotti, T, Matousek, R & Stewart, C 2011, 'Are rating agencies' assignments opaque? Evidence from international banks', Expert Systems with Applications, vol. 38, no. 4, pp. 4206-4214. https://doi.org/10.1016/j.eswa.2010.09.085 Digital Object Identifier (DOI): 10.1016/j.eswa.2010.09.085 Link: Link to publication record in Edinburgh Research Explorer Document Version: Peer reviewed version Published In: Expert Systems with Applications Publisher Rights Statement: © Bellotti, T., Matousek, R., & Stewart, C. (2011). Are rating agencies' assignments opaque? Evidence from international banks. Expert Systems with Applications, 38(4), 4206-4214. 10.1016/j.eswa.2010.09.085 General rights Copyright for the publications made accessible via the Edinburgh Research Explorer is retained by the author(s) and / or other copyright owners and it is a condition of accessing these publications that users recognise and abide by the legal requirements associated with these rights. Take down policy The University of Edinburgh has made every reasonable effort to ensure that Edinburgh Research Explorer content complies with UK legislation. If you believe that the public display of this file breaches copyright please contact openaccess@ed.ac.uk providing details, and we will remove access to the work immediately and investigate your claim. Download date: 06. Apr. 2021 https://doi.org/10.1016/j.eswa.2010.09.085 https://doi.org/10.1016/j.eswa.2010.09.085 https://www.research.ed.ac.uk/portal/en/publications/are-rating-agencies-assignments-opaque-evidence-from-international-banks(7b0b13c5-998a-46d3-b2b6-8a29cc8b7582).html CENTRE FOR EMEA BANKING, FINANCE & ECONOMICS Are Rating Agencies’ Assignment Opaque? Evidence from International Banks Tony Bellotti, Roman Matousek and Chris Stewart No. 01/10 Working Paper Series 2 Are Rating Agencies’ Assignments Opaque? Evidence from International Banks Tony Bellotti 1 Roman Matousek 2 Chris Stewart 3 Abstract We compare the ability of ordered choice models and support vector machines to model and predict international bank ratings. Although support vector machines can identify significant determinants we argue that ordered choice models are more reliable for this. Our findings suggest that ratings reflect a bank’s financial position, the timing of rating assignment and a bank’s country of origin. Accounting for country effects substantially improves predictive performance. We find that support vector machines can produce considerably better predictions of international bank ratings than ordered choice models due to the formers ability to estimate a large number of country dummies unrestrictedly. Keywords: International banks, ratings, support vector machines, ordered choice models, country effects JEL Classification: C25, C51, C52, G21. 1 University of Edinburgh Business School, William Robertson Building, 50 George Square, Edinburgh EH8 9JY. Email: Tony.Bellotti@ed.ac.uk. 2 Corresponding author: Centre for EMEA Banking, Finance and Economics, London Metropolitan Business School, London Metropolitan University, 84 Moorgate, London, EC2M 6SQ. Tel: 020-7320-1569. E-mail: r.matousek@londonmet.ac.uk. 3 London Metropolitan Business School, London Metropolitan University, 84 Moorgate, London, EC2M 6SQ. Tel: 020-7320-1651. E-mail: c.stewart@londonmet.ac.uk. 3 1. Introduction Ratings agencies’ exclusive position may be justified on the grounds of reducing asymmetric information between investors and companies (Portes, 2008). However, the current global financial crisis has severely damaged the reputation of ratings agencies (RAs) that mispriced credit risk through their ratings assignments. A number of relatively financially sound banks, according to ratings assignments, were forced to close or be bailed out by governments. This raises the question of how RAs determine bank ratings. Ratings are ordinal measures that should not only reflect the current financial position of sovereign nations, firms, banks, etc. but also provide information about their future financial positions. There has been extensive research in predicting bond ratings using ordered choice models, non-parametric techniques and artificial intelligence methods, for example, Altman and Saunders, (1998), Kamstra et al (2001), Huang et al., (2004), Kim (2005) and Lee (2007). However, we are not aware of any research that model bank ratings, although Morgan (2002) attempts to identify the determinants of the difference in two separate RAs’ bank rating assignments using (ordered) logit regressions. Morgan’s work is motivated by the inherently opaque nature of banks in terms of those outside of banks, including the RAs, assessing the risks taken by inherently opaque banks. Within this context, we seek to shed light upon how ratings agencies determine the risks of banks. Thus, we employ financial variables, in addition to country risk (which we model using country specific dummy variables), as determinants of bank ratings using both ordered choice models and support vector machines (SVMs). The main challenge in modelling ratings is to increase the probability of correct classifications. Therefore, our comparison of SVMs with ordered choice models for predicting individual bank ratings as produced by Fitch Ratings (FR) is a significant 4 contribution to current research in this field. In doing so, we model the bank ratings assigned by FR using both ordered choice models and SVMs with the aim of shedding light upon their determination and comparing the two modelling methodologies. We consider three sets of determinants of ratings with the first set being financial variables. Secondly, we examine whether bank ratings are systematically determined by the timing of the rating. Thirdly, we incorporate country indicator variables to capture country-specific variations in ratings under the rationale that a bank’s rating is related to the country in which it is based. Accounting for country (fixed) effects within the context of modelling bank ratings, is an additional contribution of our study – we demonstrate that it substantially enhances the predictive accuracy of our models. We also assess the predictive power of our models and compare the performances of ordered choice models and SVMs. The organization of the paper is as follows. Section 2 provides a brief literature review. Section 3 describes the data and the methods applied while Section 4 discusses the principal empirical findings. Section 5 then considers our models’ predicted ratings. Section 6 concludes and provides policy recommendations and suggestions for further research. 2. Brief Literature Review The ability of RAs to assign ratings correctly has extensively been questioned (Altman and Saunders, 1998, Levich et al, 2002, Altman and Rijken, 2004, Amato and Furfine, 2004, Portes, 2008). One of the most frequent arguments about the prediction abilities of RAs is that they could provide misleading information since the analysis is backward looking rather than forward looking. In addition, the low transparency of ratings assignments contributes to the concern over the accuracy of ratings. Further, RAs do not 5 have, and cannot have, superior information to market participants about uncertainty and the degree of insolvency (illiquidity) of companies. A prediction of the financial soundness of banks, corporations and sovereign countries has been of central importance for analysts, regulators and policy makers. Research focusing on the prediction of bank failures by applying Early Warning Systems (EWS) has also been extensive – see, for examples, Mayer and Pifer (1970), Altman and Saunders (1998), Kolari et al (1996) and Kolari et al (2002). There has been widespread research in predicting bond ratings using multi-variate discriminant analysis as well as probit and logit models (Altman & Saunders, 1998). Kamstra et al (2001) have recently demonstrated that predictive accuracy can be improved by combining several forecasting methods to predict bond ratings. Kim (2005) used non- parametric techniques designed to capture the dynamic relationship between input and output variables. Huang et al (2004) and Lee (2007) show that artificial intelligence methods do not provide superior predictions of bond ratings to standard ordered choice methods. Comparing ordered logit/probit regressions with SVM is a valid way of addressing the main challenge in modelling ratings, which is to increase the probability of correct classifications. However, we are not aware of any previous studies that compare ordered choice models and SVMs in terms of modelling and predicting individual bank rank ratings, which is the aim of this paper. 3. Data and Methodology FR, as one of the largest rating companies for the banking industry around the world, releases four types of ratings; legal ratings, long term and short-term (security) ratings and individual ratings. 6 We focus on individual ratings because they assess the financial position of a bank itself. As stated by FR the rating is closely linked with financial performance (financial variables). The individual rating provided by FR is divided into five categories according to the performances of rated banks and further subdivided to give a total of nine rating categories. 4 Using data on 681 international banks’ ratings between 2000 and 2007 collected from BankScope, we estimate models of the determinants of these ratings, denoted i Y . This variable is ordinal and has up to nine ranked categories that are assigned integer values from 1 to 9, such that lower values indicate a lower rating. The sample size falls as higher-order lagged explanatory factors are added to the model and this can cause all banks in a particular category to be excluded from the sample. In our application the number of categories is 8 (ratings 1 to 8) because 4 lags are included in all of our models. The eight rating categories (with assigned values in brackets) are: E (1), D/E (2), D (3), C/D (4), C (5), B/C (6), B (7), A/B (8) – there is no data on banks with an A rating in our sample. Since all models include a fourth lag of at least one financial variable the sample size used in estimation is reduced to 359 – 360 observations. The average numerical ratings range from 5.83 in 2002 to 4.31 in 2005 suggesting a general decline in ratings over time. We assess whether ratings have declined through time in our modelling. We apply ordered choice estimation techniques and SVMs to this rating data. The SVM approach has a number of advantages over other machine learning algorithms such as 4 The standard classification of the individual rating is A, B, C, D and E. A further graduation among these five ratings is used, that is, A/B, B/C, C/D and D/E. The grade A says that the bank is in an impeccable financial position with a consistent record of above average performance. The B rating defines a bank as having a sound risk profile without any significant problems. The bank’s performance generally has been in line with, or in a better position than, that of its peers. The C rating includes banks which have an adequate risk profile but possess one troublesome aspect, giving rise to the possibility of risk developing, or which have generally failed to perform in line with their peers. The D rating includes banks which are currently under-performing in some notable manner. Their financial conditions are likely to be below average and their profitability is poor. These banks have the capability of recovering using their own resources, but this is likely to take some time. Finally, the E rating includes banks with very serious problems which either require or are likely to require external support. 7 neural networks (NN): (1) it has a solid theoretical foundation in statistical learning theory (Vapnik 1995), (2) SVM finds a single global minima whereas NN may find local rather then global minima, (3) computational complexity does not increase with the size of the input space as it does with NN, and (4) SVM includes a regularization term to control model complexity and therefore avoid over-fitting. The last point is a critical advantage SVM has over standard methods too, such as logistic regression, which enables it to implement complex non-linear models such as polynomial and Gaussian models along with handling a large number of covariates (features) such as multiple country dummy variables (codes). 3.1 Ordered Choice Estimation Techniques The ordered dependent variable model assumes the latent variable form below (see Greene 2008). The model has both cross-sectional and time-series dimensions, however, the latter is referenced to the year a rating was made rather than the calendar year. Further, there is only one time-series observation for each bank’s dependent variable, although lags are available on the explanatory factors. As such, the model is not a pooled data specification. Rather it is a cross-sectional model with time-series dynamics in the explanatory variables, thus: it K k itkk K k itkk K k itkk K k itkkit uXXXXY ++++= ∑∑∑∑ = − = − = − = − 1 4,4 1 3,3 1 2,2 1 1,1 * ββββ (1) where 4 ,3 ,2 ,1 , , =− lX litk is the l th lag on the k th explanatory variable for the i th bank in period t, it u is a stochastic error term, and * it Y is the unobserved dependent variable that is related to the observed dependent variable, it Y , (assuming eight categories) as follows: 8 * 7 * 1-j 1 * if 8 7 ,...,3 ,2 , if if 1 itit jitit itit YY jYjY YY <= =≤<= ≤= λ λλ λ (2) where 1 λ , 2 λ ,…, 7 λ are unknown parameters (limit points) to be estimated with the coefficients (the kl β s). Our interest is primarily confined to the general direction of correlation between the dependent and independent variables. Therefore, we use the sign of kl β to provide guidance on whether the estimated signs of coefficients concur with our a priori expectations. This is instead of looking at the marginal effects which indicate the direction of change of the dependent variable (for each value of the dependent variable) to a change in litk X −, . For ordered choice models these marginal effects are difficult to interpret. Greene (2008) suggests that probit and logit models yield results that are very similar in practice. 3.2 Support Vector Machines Since the SVM classifier is binary it is not immediately suitable for the bank rating problem. Multiple bank ratings can be modelled using multiple SVM classifiers in a “one-against-one” or directed acyclic graph approach (Huang et al 2004). However this requires many SVMs, which further increases model complexity and computation time. A simpler method is to use SVM regression to model ratings directly. An advantage that SVM regression has in contrast to classical regression techniques such as OLS is that the loss function is ε -insensitive meaning that differences between the predicted and true value less than ε are not treated as errors. Since we are interested in predicting integers, predictions are 9 rounded to the nearest integer and so are indifferent to the fractional part of the prediction; for example, if the target rating is 2, then predictions of 2.1 or 2.3 are equally valid. Hence we can set 5.0≤ε to represent this indifference. This property allows SVM to be a more sensitive model for bank ratings. SVM regression is expressed formally as follows. Given n observations x i , y i( ) where x i is a vector of covariates and y i is a real number outcome, then SVM regression constructs a linear model bxwy +⋅=ˆ by solving the quadratic optimization problem: min w,ξ ,ξ * 1 2 w ⋅ w + C ξ i + ξ i * i=1 n ∑       subject to y i − w ⋅ x i − b ≤ ε + ξ i y i − w ⋅ x i − b ≥ ε + ξ i * ξ i ,ξ i * ≥ 0      (3) This is essentially a least absolute value regression method with the ε -insensitive loss function:    −− ≤− = otherwiseˆ ˆ if0 )ˆ,( ε ε yy yy yyL (4) and a regularization term to minimize the magnitude (complexity) of w. The relative importance of the two optimization goals of fitting the data and reducing model complexity is controlled by the parameter C. Lower values give greater emphasis to reducing model complexity whilst larger values of C give greater emphasis to model fit. By reducing the complexity of the model we make it less likely that the model over-fits the in-sample training data and so should yield better performance on out-of-sample test data. 10 This representation is in primary form, but it can be transformed into dual form using Lagrange multipliers. In dual form non-linear models can be implemented efficiently using kernel methods. Typical kernels are polynomial and Gaussian kernels (Vapnik 1995). However, for our study we used the simple linear model since we want to extract and report coefficient estimates w to compare with the ordered choice model. We also do not use complex kernels since we want to avoid over-fitting on the in-sample data set. We apply the SVM method to bank ratings data and compare its in-sample predictive performance with that of the current standard method for modelling ratings; ordered choice models. Various combinations of ε and C are considered in the SVM application. For these experiments we used LIBSVM, a popular implementation of SVM by Chang and Lin (2001). 3.3 Covariates and Modelling We consider three sets of covariates to model bank ratings, being financial variables, the year in which the rating was made, [denoted it time ] and country effects. The financial variables that we consider are as follows. The ratio of equity to total assets [denoted it Equity ], the ratio of liquid assets to total assets [ it Liquidity ] the natural logarithm of total assets [ ( ) it Assetsln ] and the net interest margin [NI_Margin], ititit OEAOIANOA −= (where it OIA is the ratio of operating income to total assets and it OEA is the ratio of operating expenses to assets), the ratio of operating expenses to total operating income [ it OEOI ] and the return on equity [ it ROAE ]. 5 5 The following three further variables were also considered for inclusion in the model: the ratio of operating expenses to assets [ OEA ], the ratio of operating income to assets [ OIA ] and the return on assets [ ROAA ]. These were excluded from the model because they would cause a high degree of multicollinearity and their effects could be captured in other ways. That is, the effects of OEA and OIA are captured by the variable OEAOIANOA −= while ROAA is a close substitute of ROAE (which it is highly correlated with). The highest pairwise simple correlations amongst the explanatory factors involve these variables. Specifically, the 11 The first to fourth lagged values of the financial variables are considered as potential determinants of bank ratings. We do not include current values of these seven variables because they may contain information that was unknown at the time the rating was made. For example, if a bank’s rating was decided in January 2007 then the value of any explanatory factor measured over the whole of 2007 would be unknown when the rating was made. Models could not be estimated when the lag length exceeded four. Therefore, models are estimated from one up to four lags of these variables. Finally, we incorporate country indicator variables to capture country-specific variations in ratings. 89 country dummy variables (there are banks from 90 countries in total) account for country-specific effects (capturing, for example, country risk). The 89 individual country dummy variables were all entered simultaneously in the SVM application. However, for the ordered choice models these country dummy variables could not all be entered simultaneously because such a model could not be estimated. Therefore, these country dummies are combined in to a single index of indicators, following Hendry (2001). 6 To obtain an initial index, 1 I , we calculate the average rating of each country in our whole sample where m δ denotes the average rating for the mth country. A dummy variable for each country is constructed such that it is unity for that country’s observation and zero otherwise. The initial index is then constructed as: ∑ − = = 1 1 1 M m mm DI δ , where m D denotes the dummy for the m th country and M is the total number of countries. This index was checked for appropriateness by estimating a single ordered choice model (we used the probit form in out application) that included the country index plus one individual country’s dummy. If the simple correlation coefficients for the each pairing (calculated using a common sample) are the following: OEA and OIA , 0.98; ROAA and NOA , 0.89; OIA and NOA , 0.84; OEA and NOA , 0.72; OIA and ROAA , 0.71; ROAA and ROAE , 0.62; ROAA and OEA , 0.60. The simple correlation coefficients of pairs of variables retained in the model are all well below 0.5 (most are substantially lower than this), which helps to ensure that the reported regressions do not suffer from severe multicollinearity. 6 Hendry’s analysis is within the context of modelling inflation using time-series data. Hendry and Santos (2005) discuss the potential advantages of using such an index. 12 latter was significant at the 5% level the value of this dummy’s coefficient was incorporated into the country index. This was repeated for all ninety countries, that is, ninety distinct regressions that contained only two variables (the country index and a particular country’s dummy) were estimated. After all the coefficients of the individual country dummies that were significant in these ninety regressions had been incorporated into the index this step was repeated until no individual country dummies were significant at the 5% level (when included in a regression with the country index). 7 The weights used in the resulting country index (denoted it Country ) are reported in Table 1. Models were then constructed using the country specific terms and the other explanatory factors (financial variables and time term). For the ordered choice models a cross-sectional variant of the general-to-specific method was employed to produce the favoured model. When more than one model could be chosen the favoured parsimonious model was selected upon the basis of the lowest SBC. Both general and parsimonious models are reported. Regarding the SVM procedure we estimate a general model with all variables included and, based upon bootstrapped confidence intervals, a parsimonious SVM is selected. That is, those covariates that are individually significant according to the confidence intervals are selected for the parsimonious model – no joint tests of significance for sets of variables are conducted. The production and use of bootstrapped confidence intervals within the SVM approach is an innovation of this study. An advantage of the SVM approach over the use of ordered choice models is that it allows all of the individual country dummies to be included simultaneously and their coefficients to be freely estimated in an unrestricted model. For comparative purposes we report results on predictive accuracy for SVMs that use the country 7 The adjustments made to the initial country index based upon the average of each country’s rating was to first, add 0.118 to the weights in the index for Argentina, Benin, Iran, Jamaica, Kenya, Lebanon, Mohgolia, Nigeria and Tunisia, and, second, subtract 7.379766 from the weight in the index for Bangladesh. 13 index employed in the ordered choice models. We also report predictive performance measures for both SVMs and ordered choice models that exclude country effects in an attempt to determine the importance of accounting for country effects when modelling bank ratings. Finally, the predictive performance of ordered choice models that incorporate a country index based upon the weights obtained from the general SVM is also reported. 4. Empirical Results: determinants of ratings The ordered logit and probit regression results for the determinants of bank ratings with four lags of the explanatory variables are given in Table 2. For all specifications we report a general model (including all lags of the variables) and one parsimonious specification. For the ordered choice models the favoured parsimonious specification only includes individually (according to z-statistics) and jointly (according to a likelihood ratio test, denoted LR statistic) significant variables. In all cases the restrictions placed on the general model to obtain the parsimonious model cannot be rejected according to a likelihood ratio test [LR(general→*)]. Whilst these generally are exclusion restrictions we also consider combining 2−itLiquidity and 3−itLiquidity into the difference variable, 322 −−− −=∆ ititit LiquidityLiquidityLiquidity , given that they have approximately equal and opposite signs. Upon this basis the favoured model includes 2−∆ itLiquidity for both probit and logit forms. The favoured parsimonious models will yield more efficient inference relative to the general model and are, therefore, used for inference. The same models are favoured for the probit and logit forms. The favoured parsimonious models include the following statistically significant effects with an unambiguous direction of correlation. The variable time has a negative effect 14 on bank ratings: the more recently the bank’s rating was made the lower the rating will be, ceteris paribus. Equity (capital adequacy) has a positive effect on a bank’s rating: a more capitalised bank has a higher rating. The natural log of assets also has a positive effect on bank ratings: banks with a larger size of assets have a higher rating. OEOI has a negative correlation with a bank’s rating. The return on assets, ROAE , has a positive impact upon ratings. All of these effects are consistent with prior beliefs. Country has a positive coefficient indicating that country specific effects affect a bank’s rating: a bank in a less stable/developed/rich economy appears to have a lower rating. For example, Canada, Ireland, Norway and Sweden are in the group of countries with the highest country specific rating while Bangladesh has the lowest country specific rating (interestingly Andorra is ranked in the top band of the country index). This finding confirms our hypothesis that a bank’s country of origin plays an important role in assigning individual ratings, capturing constant country specific effects (rather like fixed-effects in a panel data model) that are not explained by the financial variables. Both the second and third lags of Liquidity are significant and their coefficients are of approximately equal and opposite sign in the general models. Hence, it is the second lag of the change in liquidity, 2−∆ tLiquidity , rather than its level, that appears to be important (in the parsimonious specification) and it has a plausible positive effect upon bank ratings. That is, a bank whose liquidity increased two periods ago has a higher rating. We note that this effect would not have been revealed had we not allowed for sufficient lags in the dynamic specification. We believe that allowing for such lags is a strength of our investigation relative to analyses that do not consider such dynamics. NI_Margin is not significant, thus it appears that NI_Margin does not determine bank ratings. 15 Finally, the second lag of NOA is significant. We are cautious of interpreting this as supportive of a significant effect upon rating because 2−tNOA has a theoretically implausible negative sign. This apparent and unexpected correlation may be due to a Type I error (of which there is a 5% chance given our chosen significance level). Under the heading ‘Undifferenced specifications’ in Table 3 are the SVM estimated weights with two sets of bootstrap confidence intervals at the 80% level. We do not report the 95% confidence intervals because only two non-country covariates are significant using this level of significance (and one has an unexpected sign), being ( ) 1 ln −tAssets and 3−tLiquidity , and, in addition, up to twenty country dummies are significant. 8 To obtain a number of significant covariates closer to that obtained with the ordered choice models we employed a broader 80% confidence interval to produce our reported parsimonious model. The following eight non-country variables are included in the parsimonious model (their weights’ signs are given in parentheses): time (negative), ( ) 1 ln −tAssets (positive), 1−tOEOI (negative), 2−tOEOI (negative), 1−tROAE (positive), 2−tLiquidity (positive), 3−tLiquidity (negative) and 3−tNOA (negative). All these weights’ signs are plausible except for 3−tLiquidity and 3−tNOA . However, the weight on 3−tLiquidity is approximately of equal magnitude to that on 2−tLiquidity suggesting that the difference of this variable is important, which is consistent with the findings from the ordered choice model. Nevertheless, the 95% confidence intervals are broad and disappointing and using 80% confidence intervals simply to secure more acceptable results is arbitrary. We believe that these disappointing confidence intervals may be due to multicollinearity between the lags of covariates leading to unstable SVM estimates. Therefore, we applied an alternative “difference” model with undifferenced covariates used for lag 1 and the first three lags of 8 With bootstrapped confidence intervals 13 country dummies are significant while 20 of the country indicators are significant according to the confidence intervals based upon the normal distribution. 16 differenced variables. The weights of the model, with two sets of 95% confidence intervals, are reported in Table 3 under the heading ‘Differenced specifications’. These results seem more satisfactory because there are a greater number of significant covariates in the ‘difference specification’ compared to the ‘undifferenced specification’ using the 95% confidence interval, and broadly corroborate the results produced by the ordered choice model. The reported parsimonious model, based on the 95% confidence intervals, includes the following variables: 1−tEquity (positive), ( ) 1ln −tAssets (positive), 1−tOEOI (negative), 1−tROAE (positive) and 2−∆ tLiquidity (positive) plus 11 country dummy variables. All of the weights on the included variables have theoretically plausible signs that reinforce our satisfaction with this ‘differenced specification’. Overall SVMs used with bootsrapped confidence intervals applied at standard levels can be employed to determine the significant variables in the same way as ordered choice models. Upon this basis a parsimonious specification can be chosen. To determine if SVMs select parsimonious models that are as good as those obtained using ordered choice models we need to compare the predictive accuracy of these specifications. The predictive accuracy of these models is considered in the next section. 5: Empirical Results: predictive performance The percentage of correct predictions of the general and parsimonious specifications for both ordered choice models and SVMs are reported in Table 4 and Table 5, respectively. 9 From Table 4 (row entitled With Country) we see that there are between 50.1% and 52.1% correct predictions for the ordered choice models including the country variable (predictions 9 This prediction is calculated using the same sample employed to estimate the data. It is a fit measure rather than providing an assessment on out-of-sample data. We did not reserve any data for out-of-sample evaluation in order to maximise the period that could be used for estimation and, therefore, maximise efficiency of estimation, which is especially important for ordered choice models. 17 are obtained from the models reported in Table 2). 10 The percentages of correct predictions for these models excluding the country variable are reported in the row entitled No Country of Table 4 for comparative purposes. 11 The parsimonious model is obtained by applying the general-to-specific method with all variables except for Country included in the general model. The estimated percentage correct predictions for these regressions are between 34.0% and 36.6%. The predictive accuracy is substantially greater (by at least over 13.5 percentage points) for models incorporating the Country variable compared to those that do not. The regressions including this country index also have much larger pseudo 2 R s than those that do not and the country index is highly significant in all models in which it is included. This further demonstrates the importance of modelling country effects for predicting international bank ratings. The percentages of correctly predicted bank ratings obtained from the undifferenced SVM with all country dummies included simultaneously and estimated unrestrictedly for various combinations of C and ε are reported in Table 5 in the section headed Undifferenced Model with Unrestricted Dummies. The predictive accuracy of the general SVM is between 48.5% and 62.4% (the majority of SVM predictions exceed 57%) which is substantially better than obtained from the ordered choice models. 12 If such performance can be repeated out-of-sample this would suggest the adoption of SVMs would provide greater predictive accuracy than ordered choice 10 These percentage of correct predictions are similar for probit and logit specifications, if the latter, in general, produce slightly more accurate predictions. 11 To save space we do not report these estimated models, however, these results are available from the authors on request. 12 To place this predictive performance in context we note that when there are nine (eight) rating categories the expected accuracy of predictions by chance is 11.1% (12.5%). Hence, the best performing SVM can increase the predictive accuracy by over fifty percentage points. 18 models that are currently used as standard for this purpose. 13 This is important given that prediction is the primary purpose of such models. The predictive performance of the parsimonious undifferenced SVM with unrestricted dummies is between 42.49% and 52.09%. This performance is substantially worse than the general SVM version of this model and no better (and often worse) than that of the ordered choice models (reported in the row No Country of Table 4). The implication is that model reduction within the SVM method has lead to important variables being excluded suggesting that one might be cautious when using the SVM methodology to identify the significant determinants of ratings. This contrasts with the ordered choice method where there was no systematic difference in the predictive performance of general and parsimonious models. We report the predictive accuracy of the difference SVM (the estimation results are given in Table 3) in the section headed Differenced Model with Unrestricted Dummies of Table 5. The predictive accuracy is between 47.4% and 61.8% for the general differenced SVM and in the range of 43.2% and 47.4% for its parsimonious counterpart. The maximum predictive accuracy of the differenced SVM (being 61.8%) is slightly less than that of the undifferenced SVM (62.4%). It is usually understood that multicollinearity is a problem for accurate estimation of coefficients rather than prediction and these results show that this is the case for our SVM models: there is no problem for prediction, only for reporting the weight estimates and their associated confidence intervals. Similar to the undifferenced SVM the predictive accuracy of the parsimonious differenced SVM is substantially lower (with a maximum of 47.4%) than for its general counterpart. This reinforces our earlier findings and demonstrates that for the SVM method using as many covariates as possible is best for achieving good predictive results, even if some of those covariates are not statistically significant in the conventional sense. 13 The in-sample predictive performance of the general undifferenced SVMs is at least as good as the ordered choice models with 5.0<ε and C > 0.5 and is best when 25.0=ε and C = 2 (with 62.4% correct predictions). This suggests that choice of these parameters is important in the selection of the SVM used for prediction. 19 The predictive accuracy of undifferenced SVMs that exclude the country dummy variables are reported in the section headed Undifferenced Model with No Country Dummies in Table 5. The percentages of correct predictions are between 34.8% and 39.0% (29.0% - 31.8%) for the general (parsimonious) undifferenced SVM without country dummies. This is a substantially worse predictive performance relative to when country dummies are included in SVMs. This confirms the findings obtained from the ordered choice models’ predictive performance and further emphasises the importance of accounting for country effects when modelling and predicting bank ratings. SVMs with undifferenced variables are re-estimated using the single country index (employed in the ordered choice models) to capture country effects instead of the individual country dummies and their prediction accuracy is reported in the section headed Undifferenced Model with Single Country Index of Table 5. The predictive accuracy is substantially lower for the SVMs using the single country index (between 48.7% and 53.2% for the general model and 49.0% and 52.6% for the parsimonious specification) than for the SVMs with unrestricted dummies but are similar to those produced by the ordered choice models (that also use this country index). Hence, the superior predictive performance of SVMs over ordered choice models seems to be because the former can estimate the coefficients of the country dummies simultaneously and unrestrictedly, whereas the ordered choice models cannot. To explore this issue further we calculate the percentage of correct predictions from ordered choice models that include a new country index constructed using the weights obtained from the general undifferenced SVM with unrestricted dummies. 14 The weights for this model are reported in Table 6. Predictive performance is reported for ordered choice models that include this new country index in the row entitled With Country SVM of Table 4. 14 The simple and Spearman rank correlation coefficients are 0.793 and 0.862, respectively, for the correlation between the original country index and the one based upon the SVM weights. Hence, whilst they are highly correlated there are still notable differences between the two indexes. 20 The original index is replaced with the new index in the general models and the general-to- specific methodology is applied to obtain parsimonious models. The predictive performance measures fall in the range of 53.8% to 57.8% with the parsimonious model securing the highest accuracy. It is noticeable that all of the ordered choice models that use the index based upon SVM weights have a superior predictive accuracy than the best performing ordered choice model that employs the original index. However, these ordered choice models can still be substantially outperformed by the undifferenced SVM that estimates the country dummies unrestrictedly. This further highlights the advantage that the SVM has in terms of its ability to freely estimate the coefficients of all of the country dummies. 6. Conclusions Using data on banks from around the world we compare two different methodologies, ordered choice models and SVMs, for predicting and identifying the significant determinants of bank ratings. The ordered choice models unambiguously identify the following significant determinants of ratings. Banks with a greater capitalisation ( Equity ), larger assets [ ( )Assetsln ], and a higher return on assets ( ROAE ) have higher bank ratings. Further, if a bank’s liquidity ( )Liquidity∆ increased two periods ago it ratings will rise. Conversely, the greater is a bank’s ratio of operating expenses to total operating income ( OEOI ) and the more recent is the date that the rating is made ( time ) the lower is the rating of the bank. However, we also find that net operating income to total assets ( )NOA has an unexpected negative influence on bank ratings. In addition, there is strong evidence that a bank’s country of origin has a significant influence on bank ratings. Inclusion of this country effect 21 substantially raises the ability of an ordered choice model to accurately predict international bank ratings relative to models that exclude country effects. The SVM results confirm the importance of country effects as significant determinants of ratings: when they are excluded the predictive performance of the model considerably deteriorates relative to a model including country effects. However, although the SVM method could be adapted to identify the significant variables it suggested few significant determinants using a 95% level. Using an 80% confidence interval unsurprisingly raised the number of significant variables. Given the arbitrary and unconventional choice of confidence level required to secure this result we considered whether the breadth of the confidence intervals was due to multicollinearity among the lags of the covariates. Using an SVM that incorporated variables in both difference and level form yielded more significant determinants using a conventional 95% confidence interval that all had expected signs on the weights. These significant variables are, Equity , ( )Assetsln , ROAE , Liquidity∆ and OEOI which are the same as for the ordered choice models except NOA and time , which were not indicated as significant by the differenced SVM. However, the predictive performance of the parsimonious SVMs were considerably lower than that of the general SVMs suggesting that the model reduction method applied to SVMs leads to important determinants being excluded from the model and, therefore, not being identified. For this reason we are cautious to present the variables identified by the SVM method as the only significant determinants of ratings. Thus, based primarily on the results obtained from the parsimonious ordered choice models, we conclude that ratings reflect a bank’s financial position (as measured by various financial variables), the timing of when the rating was made and a bank’s country of origin. Regarding the timing effect, FR may have applied more prudent views and policies as a reaction to critiques of their role during the financial turbulence of the late 1990s. We have 22 therefore identified a set of determinants that are revealing in how ratings agencies determine the risks of inherently opaque banks, especially given the high predictive accuracy achieved by our models. We have also found that SVMs can produce substantially better in-sample predictions of international bank ratings than the standard method currently used for this purpose, ordered choice models. This appears to be due to the SVM’s ability to estimate a large number of country dummies’ coefficients unrestrictedly, which was not possible with the ordered choice models due to the small sample size. Given that prediction is the primary purpose of modelling ratings, this is an important result. In this paper we have only considered in-sample predictions due to the fact that using some of the data for an out-of- sample data set would have severely restricted the (training) in-sample to less than the 360 observations. However, consideration of the relative out-of-sample predictive performance of SVMs and ordered choice models, requiring more observations than were available here, would be a desirable avenue for further research. Additionally we deliberately did not use SVMs with non-linear kernels to model the data in this exercise since we did not have sufficient data for an independent validation set to optimize non-linear model parameter settings. However we expect that using different kernels may improve performance. We intend to pursue these lines of research with a larger data set. 23 References Altman, E. I., and Saunders, A. (1998). Credit risk measurement: Developments over the last 20 years. Journal of Banking and Finance, 21, 1721–1742. Altman, E., and Rijken, H., A. (2004). How Rating Agencies Achieve Rating Stability, Journal of Banking & Finance, 28, 2679-2714. Altman, E., and Rijken, H., A. (2006). A Point in Time Perspective on Through-the Cycle Ratings, Financial Analysts Journal, 62, 54-70. Altman, E., Bharath, S., and Saunders, A. (2002). Credit ratings and the BIS capital adequacy reform agenda. Journal of Banking & Finance 26, 929-951. Altman, E., Saunders, A., (2001). An analysis and critique of the BIS proposal on capital adequacy and ratings. Journal of Banking & Finance 25, 25–46. Amato, J. D. and Furfine, C.H. (2004). Are Credit Ratings Procyclical?, Journal of Banking and Finance 29, 2641–2677. Chang C-C and Lin C-J (2001) LIBSVM : a library for support vector machines. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm Greene W. H., (2008). Econometric Analysis, Pearson, Prentice Hall, 6 th edition. Hendry D. F. and Santos C. (2005). “Regression models with data-based indicator variables” Oxford Bulletin of Economics and statistics, 67, 5, 571 – 595. Hendry D. F., (2001). “Modelling UK Inflation, 1875 – 1991” Journal of Applied Econometrics, 16, 255 – 275. Huang, Zan, Hsinchun Chen, Chia-Jung Hsu, Wun-Hwa Chen, Soushan Wu. (2004). Credit rating analysis with support vector machines and neural networks: a market comparative study, Decision Support Systems, 37, pp. 543-558, Kamstra, M., Kennedy, P., Suan, T.K. (2001). Combining bond rating forecasts using logit. The Financial Review, 37, pp.75-96. Kim, S. K. (2005). Predicting bond ratings using publicly available information, Expert Systems with Applications, 29, pp.75-81 Kolari, J. D. Glennon, H. Shin and M. Caputo, (2002). Predicting large US commercial bank failures, Journal of Economics and Business, 54, pp. 361–387. Lee, Y.C. (2007). Application of support vector machines to corporate credit rating prediction, Expert Systems with Applications 33, pp. 67–74. Levich, R., Majnoni, G., Rinhart, C. (2002). Ratings, Rating and Agencies and the Global Financial System, Kluwer Publishing. Meyer, P., & Pifer, H. (1970). Prediction of bank failures. Journal of Finance, 25, 853–868. Morgan D. P. (2002). Rating Banks: Risk and Uncertainty in an Opaque Industry. The American Economic Review, 92 (4), 874–888. Pinto, A. R. (2006). Control and Responsibility of Credit Rating Agencies in the United States, American Journal of Comparative Law, 54, 341-356. Portes, R. (2008). Ratings agency reform, Vox, 22 January,2008, <http://www.voxeu. org/index.php?q=node/887> 24 Vapnik V (1995). The Nature of Statistical Learning Theory. Springer NY 25 Table 1: Ordered Choice Models’ Country Index Weights Country Weight Country Weight ANDORRA 7.00 TAIWAN 4.09 CANADA 7.00 COLOMBIA 4.00 IRELAND 7.00 COSTA RICA 4.00 NORWAY 7.00 LITHUANIA 4.00 SWEDEN 7.00 MALTA 4.00 USA 6.82 MOROCCO 4.00 SWITZERLAND 6.75 PERU 4.00 SPAIN 6.71 TURKEY 3.76 NETHERLANDS 6.50 EL SALVADOR 3.75 SAUDI ARABIA 6.43 INDONESIA 3.75 AUSTRIA 6.00 INDIA 3.54 CZECH REPUBLIC 6.00 EGYPT 3.50 ESTONIA 6.00 HUNGARY 3.50 HONG KONG 6.00 BULGARIA 3.40 ICELAND 6.00 LATVIA 3.33 JORDAN 6.00 ARGENTINA 3.12 SAN MARINO 6.00 BENIN 3.12 SOUTH AFRICA 6.00 IRAN 3.12 KOREA REP. OF 5.82 JAMAICA 3.12 FRANCE 5.71 KENYA 3.12 UNITED KINGDOM 5.64 LEBANON 3.12 ITALY 5.55 MONGOLIA 3.12 CHILE 5.50 NIGERIA 3.12 GERMANY 5.50 TUNISIA 3.12 KUWAIT 5.50 KAZAKHSTAN 3.00 SLOVENIA 5.50 ROMANIA 3.00 GREECE 5.43 RUSSIAN FEDERATION 2.90 BERMUDA 5.33 PHILIPPINES 2.83 QATAR 5.25 VENEZUELA 2.75 UNITED ARAB EMIRATES 5.25 VIETNAM 2.67 BAHRAIN 5.17 GEORGIA REP. OF 2.50 AUSTRALIA 5.00 PAKISTAN 2.50 MACAU 5.00 CHINA-PEOPLE'S REP. 2.44 OMAN 5.00 UKRAINE 2.21 PANAMA 5.00 ALBANIA 2.00 TRINIDAD AND TOBAGO 5.00 ARMENIA 2.00 MEXICO 4.89 AZERBAIJAN 2.00 JAPAN 4.86 BOSNIA-HERZEGOVINA 2.00 ISRAEL 4.67 MACEDONIA (FYROM) 2.00 SLOVAKIA 4.50 NIGER 2.00 BRAZIL 4.38 SERBIA 2.00 CYPRUS 4.33 SRI LANKA 1.80 THAILAND 4.33 DOMINICAN REPUBLIC 1.75 POLAND 4.29 BELARUS 1.60 MALAYSIA 4.25 BANGLADESH -6.38 Table 1 notes. The country index for the ordered choice models is constructed as the sum of the products of the country weights (given in the column headed weight) and the individual country dummy variables. 26 Table 2: Bank ratings ordered choice regressions Logit specifications Probit specifications Variables General model Parsimonious model General model Parsimonious model Country 1.975 (12.938) 1.904 (14.391) 1.078 (12.81) 1.039 (13.618) time –0.298 (–2.114) –0.249 (–2.147) –0.182 (–2.50) –0.155 (–2.407) 1−tEquity –0.001 (–0.014) 0.005 (0.18) 1−tLiquidity 0.437 (0.295) 0.446 (0.55) ( ) 1ln −tAssets 0.840 (2.762) 0.524 (6.900) 0.522 (2.97) 0.290 (7.072) NI_Margin 1−t –0.034 (–0.239) –0.028 (–0.47) 1−tNOA 6.520 (0.663) 1.954 (0.38) 1−tOEOI –0.528 (–2.094) –0.508 (–3.449) –0.296 (–2.01) –0.310 (–3.840) 1−tROAE 0.021 (1.367) 0.033 (3.595) 0.015 (1.98) 0.020 (3.948) 2−tEquity 0.032 (0.611) 0.014 (0.49) 2−tLiquidity 5.180 (2.805) 2.877 (2.75) ( ) 2ln −tAssets –0.009 (–0.013) –0.161 (–0.41) NI_Margin 2−t 0.108 (0.697) 0.066 (0.89) 2−tNOA –27.913 (–1.553) –10.376 (–2.360) –15.194 (–1.55) –5.824 (–2.081) 2−tOEOI –1.377 (–2.091) –1.812 (–3.132) –0.749 (–2.03) –1.005 (–3.128) 2−tROAE 0.016 (1.616) 0.008 (1.30) 3−tEquity 0.066 (0.868) 0.081 (6.7346) 0.044 (1.32) 0.051 (6.909) 3−tLiquidity –5.481 (–2.548) –2.914 (–2.67) ( ) 3ln −tAssets –0.356 (–0.361) –0.120 (–0.25) NI_Margin 3−t –0.067 (–0.850) –0.039 (–0.92) 3−tNOA 4.089 (0.410) 2.978 (0.57) 3−tOEOI –0.302 (–0.592) –0.265 (–0.84) 3−tROAE 0.001 (0.233) 0.000 (0.05) 4−tEquity –0.002 (–0.039) –0.004 (–0.20) 4−tLiquidity 0.205 (0.141) –0.184 (–0.26) ( ) 4ln −tAssets 0.105 (0.188) 0.079 (0.29) NI_Margin 4−t 0.035 (1.091) 0.024 (1.30) 4−tNOA 0.392 (0.100) –0.398 (–0.19) 4−tOEOI –0.012 (–0.094) –0.030 (–0.37) 4−tROAE 0.002 (0.0665) 0.004 (2.599) 0.001 (0.74) 0.002 (2.121) 2−∆ tLiquidity 5.311 (4.157) 3.029 (4.125) Limit Points λ1 8.049 (5.315) 6.797 (5.643) 4.505 (5.593) 3.866 (5.988) λ2 11.618 (7.1275) 10.316 (7.916) 6.304 (7.402) 5.649 (8.205) λ3 14.158 (8.507) 12.822 (9.384) 7.685 (8.675) 7.012 (9.703) λ4 16.123 (9.335) 14.786 (10.239) 8.735 (9.749) 8.058 (10.543) λ5 18.456 (10.211) 17.098 (11.195) 10.023 (10.367) 9.333 (11.483) λ6 20.121 (10.877) 18.740 (11.968) 10.934 (11.029) 10.232 (12.228) λ7 22.618 (11.803) 21.175 (12.946) 12.301 (12.034) 11.567 (13.309) Fit Measures % correct 52.089 50.278 50.139 50.278 Pseudo 2 R 0.385 0.381 0.373 0.368 SBC 2.993 2.681 2.926 2.729 LR statistic 536.110 [0.000] 532.230 [0.000] 519.016 [0.000] 514.638 [0.000] LR(general→*) NA 6.993 [0.997] NA 7.368 [0.995] Observations 359 360 359 360 Table 2 notes. The dependent variable is a bank’s rating which has eight categories that correspond to the integer values in the range of 1 to 8 and yields seven limit points, 7 ,...,2 ,1 , =iiλ (the intercept is not separately identified from the limit points). Z-statistics (in parentheses) are based upon Huber-White standard errors and the percentage of correct predictions (% correct) use the category with the highest probability to give the predicted rating. Also reported are the Pseudo 2 R , Schwartz’s information criterion, SBC, and likelihood ratio tests for the model’s explanatory power, LR Statistic, and the deletion of variables from the general model to obtain the parsimonious model, LR(general→*). Probability values are given in square parentheses. All regressions were estimated using E-Views 6.0 and STATA 10. 27 Table 3: Bank ratings SVMs Undifferenced specifications Differenced specifications Variables General model Parsimonious model Variables General model Parsimonious model time –0.096 [–0.24, –0.01] * –0.091 [–0.01, 0.01] time –0.102 [–0.29, 0.03] 1−tEquity 0.016 [–0.01, 0.06] 1−tEquity 0.050 [0.03, 0.08] * 0.050 [0.03, 0.07] * 1−tLiquidity 0.369 [–0.13, 1.25] 1−tLiquidity –0.133 [–0.08, 0.95] ( ) 1ln −tAssets 0.371 [0.20, 0.64] * 0.350 [0.28, 0.39] * ( ) 1ln −tAssets 0.402 [0.25, 0.46] * 0.434 [0.36, 0.52] * NI_Margin 1−t 0.010 [–0.08, 0.05] NI_Margin 1−t –0.039 [–0.11, 0.01] 1−tNOA –0.444 [–0.79, 0.05] 1−tNOA –0.459 [–1.12, 0.25] 1−tOEOI –0.266 [–0.53, –0.10] * –0.183 [–0.43, –0.08] * 1−tOEOI –0.776 [–1.77, –0.14] * –0.364 [–1.13, –0.02] * 1−tROAE 0.010 [0.01, 0.02] * 0.009 [0.002, 0.02] * 1−tROAE 0.020 [0.003, 0.04] * 0.014 [–0.01, 0.03] 2−tEquity 0.017 [–0.02, 0.05] 1−∆ tEquity -0.036 [–0.08, 0.02] 2−tLiquidity 0.793 [0.01, 1.47] * 1.055 [0.28, 1.62] * 1−∆ tLiquidity 0.300 [–0.71, 1.46] ( ) 2ln −tAssets 0.242 [–0.27, 0.55] ( ) 1ln −∆ tAssets -0.034 [–0.30, 0.52] NI_Margin 2−t 0.007 [–0.09, 0.08] ∆ NI_Margin 1−t 0.050 [–0.07, 0.12] 2−tNOA –0.508 [–8.01, 11.32] 1−∆ tNOA 0.008 [–0.54, 0.41] 2−tOEOI –0.494 [–1.04, –0.03] * –0.977 [–1.31, –0.57] * 1−∆ tOEOI 0.511 [–0.08, 1.42] 2−tROAE 0.008 [–0.05, 0.01] 1−∆ tROAE -0.010 [–0.03, 0.02] 3−tEquity 0.011 [–0.03, 0.06] 2−∆ tEquity -0.005 [–0.05, 0.06] 3−tLiquidity –1.272 [–2.00, –0.47] * –1.251 [–1.82, –0.47] * 2−∆ tLiquidity 1.404 [0.23, 2.55] * 1.165 [–0.09, 2.30] ( ) 3ln −tAssets –0.475 [–0.90, 0.15] ( ) 2ln −∆ tAssets 0.218 [–0.43, 0.75] NI_Margin 3−t –0.039 [–0.08, 0.08] ∆ NI_Margin 2−t 0.043 [–0.10, 0.11] 3−tNOA –0.654 [–1.07, –0.08] * –0.235 [–1.14, 0.55] 2−∆ tNOA 0.155 [–0.46, 0.74] 3−tOEOI –0.151 [–0.79, 0.17] 2−∆ tOEOI 0.041 [–0.53, 1.00] 3−tROAE 0.000 [–0.01, 0.01] 2−∆ tROAE -0.002 [–0.03, 0.004] 4−tEquity 0.007 [–0.03, 0.03] 3−∆ tEquity -0.009 [–0.04, 0.05] 4−tLiquidity –0.036 [–0.88, 0.50] 3−∆ tLiquidity -0.195 [–1.54, 0.93] ( ) 4ln −tAssets 0.259 [–0.18, 0.63] ( ) 3ln −∆ tAssets -0.210 [–0.88, 0.40] NI_Margin 4−t –0.012 [–0.05, 0.01] ∆ NI_Margin 3−t 0.018 [–0.03, 0.07] 4−tNOA –0.015 [–0.62, 0.65] 3−∆ tNOA -0.639 [–1.66, 0.45] 4−tOEOI 0.005 [–0.18, 0.08] 3−∆ tOEOI -0.018 [–0.22, 0.26] 4−tROAE 0.002 [–0.001, 0.005] 3−∆ tROAE -0.002 [–0.01, 0.003] Table 3 notes. The dependent variable is the bank rating (which is never differenced). Confidence intervals are reported from a bootstrap (1,000 samples) in square brackets []. These are 80% confidence intervals for the undifferenced specifications while 95% level intervals are employed for the difference specifications. An asterix, * , indicates a variable that is significant according to the reported confidence intervals. Models also include dummy variables for countries but these are not shown to save space. 28 Table 4: Percentage Correct Predictions: Ordered Choice Models Logit Probit General Parsimonious General Parsimonious With Country 52.1% 50.3% 50.1% 50.3% No Country 35.9% 36.6% 34.0% 34.1% With Country SVM 56.8% 57.8% 53.8% 53.8% Table 4 notes. The predicted rating for each observation is chosen upon the basis of the category with the highest probability. The row entitled With Country refers to the percentage of correct predictions obtained from the models reported in Table 2. The row entitled No Country reports the correct predictions for models developed using the general-to-specific method where the country variable is excluded from the general model. The row entitled With Country SVM denotes the predictive accuracy of ordered choice models that use the country index constructed with the weights obtained from the SVM’s general model (reported in Table 6) and the parsimonious models are developed applying the general-to-specific method. 29 Table 5: Percentage Correct Predictions from SVMs Undifferenced Model with Unrestricted Dummies General Parsimonious ε ε 0.05 0.25 0.5 0.05 0.25 0.5 C 0.25 53.2% 53.8% 48.5% 48.2% 46.2% 44.3% 0.5 56.8% 59.3% 51.0% 50.1% 49.3% 46.8% 1 57.9% 61.3% 53.2% 50.4% 51.8% 48.8% 2 58.2% 62.4% 56.3% 51.5% 52.1% 47.6% 4 59.9% 61.6% 53.2% 51.3% 50.7% 49.0% 8 58.5% 61.3% 53.2% 51.5% 50.7% 48.8% 12 58.2% 61.6% 52.6% 51.5% 50.7% 48.2% Differenced Model with Unrestricted Dummies General Parsimonious ε ε 0.05 0.25 0.5 0.05 0.25 0.5 C 0.25 52.4% 54.3% 47.4% 46.5% 47.1% 44.3% 0.5 56.3% 59.3% 51.5% 46.8% 46.8% 45.1% 1 57.4% 60.2% 52.4% 46.8% 47.4% 45.4% 2 57.7% 61.8% 54.3% 47.4% 47.4% 44.6% 4 58.8% 60.7% 53.8% 46.5% 47.1% 44.3% 8 59.1% 61.6% 53.8% 46.5% 46.0% 44.0% 12 58.5% 59.6% 54.9% 46.5% 46.0% 43.2% Undifferenced Model with No Country Dummies General Parsimonious ε ε 0.05 0.25 0.5 0.05 0.25 0.5 C 0.25 37.6% 37.9% 34.8% 29.5% 29.5% 29.0% 0.5 38.7% 38.4% 35.4% 30.9% 31.8% 29.5% 1 38.12% 38.4% 36.5% 30.9% 30.9% 30.6% 2 39.3% 38.2% 36.8% 29.2% 29.5% 30.1% 4 38.4% 38.4% 38.2% 30.1% 29.5% 29.5% 8 38.7% 39.0% 38.4% 29.0% 29.5% 29.8% 12 38.4% 38.4% 38.2% 29.0% 29.2% 29.2% Undifferenced Model with Single Country Index General Parsimonious ε ε C 0.05 0.25 0.5 0.05 0.25 0.5 0.25 52.9% 52.6% 49.6% 51.8% 51.3% 50.1% 0.5 53.2% 52.4% 49.3% 51.5% 51.5% 49.6% 1 52.9% 52.9% 48.7% 52.4% 52.1% 49.0% 2 52.6% 52.9% 50.1% 52.4% 51.5% 49.6% 4 52.9% 52.9% 51.8% 52.1% 51.5% 50.7% 8 52.9% 52.1% 50.7% 52.6% 50.7% 51.3% 12 52.6% 52.1% 51.3% 52.4% 50.4% 51.8% Table 5 notes. the statistics are the percentage of correct predictions obtained from SVMs. All sections except for the second section are based upon SVMs using undifferenced variables. The section headed Undifferenced Model with Unrestricted Dummies refers to SVMs that allow all country dummies to be estimated unrestricted – the regressions from which the predictions are made are reported in Table 3 under the heading Undifferenced specifications. The section headed Undifferenced Model with No Country Dummies reports predictive accuracy using SVMs that exclude all country terms. The section headed Undifferenced Model with Single Country Index provides results based upon SVMs including the single country index constructed from the ordered choice regressions (the weights are reported in Table 1) and no individual country dummy variables. The section headed Differenced Model with Unrestricted Dummies 30 refers to SVMs that allow all country dummies to be estimated unrestricted and include differenced variables – the regressions from which the predictions are made are reported in Table 3 under the heading Differenced specifications. 31 Table 6: Undifferenced General SVM Model’s Country Index Weights Country Weight Country Weight USA 2.283 KENYA 0.000 ANDORRA 2.000 LATVIA 0.000 UNITED KINGDOM 2.000 MACEDONIA (FYROM) 0.000 SWITZERLAND 1.705 MALTA 0.000 BERMUDA 1.684 NETHERLANDS 0.000 CANADA 1.281 NORWAY 0.000 KOREA REP. OF 1.066 PHILIPPINES 0.000 CZECH REPUBLIC 1.065 SAN MARINO 0.000 SAUDI ARABIA 1.050 SERBIA 0.000 FRANCE 1.019 SOUTH AFRICA 0.000 SLOVENIA 0.929 SPAIN 0.000 PANAMA 0.644 SWEDEN 0.000 UNITED ARAB EMIRATES 0.616 TAIWAN 0.000 KUWAIT 0.612 TUNISIA 0.000 MACAU 0.608 HUNGARY -0.009 OMAN 0.519 ALBANIA -0.015 LITHUANIA 0.452 MALAYSIA -0.026 AUSTRIA 0.435 POLAND -0.093 CYPRUS 0.412 INDONESIA -0.133 MEXICO 0.401 ROMANIA -0.170 SLOVAKIA 0.382 MOROCCO -0.179 BAHRAIN 0.337 ISRAEL -0.185 QATAR 0.307 THAILAND -0.251 EL SALVADOR 0.297 NIGER -0.266 TRINIDAD AND TOBAGO 0.297 PERU -0.320 BRAZIL 0.249 TURKEY -0.396 JAPAN 0.246 AZERBAIJAN -0.463 GERMANY 0.228 KAZAKHSTAN -0.473 JORDAN 0.108 RUSSIAN FEDERATION -0.539 GEORGIA REP. OF 0.034 INDIA -0.550 MONGOLIA 0.003 COLOMBIA -0.619 ARMENIA 0.000 UKRAINE -0.763 AUSTRALIA 0.000 JAMAICA -0.869 BENIN 0.000 VIETNAM -0.969 BOSNIA-HERZEGOVINA 0.000 VENEZUELA -0.984 BULGARIA 0.000 NIGERIA -1.072 CHILE 0.000 ARGENTINA -1.083 COSTA RICA 0.000 LEBANON -1.134 EGYPT 0.000 BELARUS -1.205 ESTONIA 0.000 PAKISTAN -1.416 GREECE 0.000 BANGLADESH -1.452 HONG KONG 0.000 SRI LANKA -1.611 ICELAND 0.000 DOMINICAN REPUBLIC -1.696 IRELAND 0.000 IRAN -2.141 ITALY 0.000 CHINA-PEOPLE'S REP. -2.188 Table 6 notes. The estimated weights for the individual country dummy variables are obtained from the SVM model where the variables are not differenced and the model includes 32 all possible variables (general model). The other estimation results for this model are reported in Table 3. 
work_f2kze3heqfe65gmxy6kg7fbuda	----	Annoyance accumulation modeling in a community noise annoyance expert system Annoyance accumulation modeling in a community noise annoyance expert system Dick Botteldooren Acoustics group, Department of Information Technology, Universi~ of Gent, Belgium Abstract: When the state of the sound environment in a region is monitored by tracing the percentage of people that is annoyed by noise, simulation models for community noise annoyance are needed. Building such simulation tools is not an easy task because information is scarce and relations between quantities involved is uncertain. Tberefor an expert system that makes optimal use of the scare information is constructed. Modifier fuzzyness is fully taken into account. Combination of annoyance caused by different sources is an important part of the system. The probability approximation is used in this paper to incorporate perceptual annoyance models for combined community noises. The uncertainty that still exists in this field does not pose any problems to the system. FUZZY AND ANNOYAN~ Annoyance is a vague concept. Often vague or fuzzy concepts are context dependent. In annoyance models this is sometimes referred to as the principle of compromise (1). Modifiers (e.g. highly, moderately) show a similar fuzzy ness. In annoyance surveys one often gets around the fuzzyness in modifiers by using a discrete numerical scale. In the expert system proposed here, both numerical and verbal scales can be used, as long as consistency is maintained in all calculations. It is advantageous to cover the whole range of modifiers in the calculations including for example zero on a numerical scale or “not annoyed” on a verbal scale. THE EXPERT SYSTEM The expert system designed for community noise annoyance simulation, consists of an environment that allows to take into account uncertainty and vagueness in all calculations, Uncertainty on quntities is represented by a probability distribution. Interfaces to Gaussian and level set descriptions are provided. The latter approach is preferred for input in many cases. It consists in fixing a number of uncertainty intervals for each input (e.g. 6870, 95?0 and l~~o certainty intervals). Standard operations such as adding, multiplying, matrix multiplication or interpolation are defined for uncertain variables. The key feature of the system is “consensus”. It allows to combine in a formal way, different calculations or different expert opinions to obtain a possibly better approximation from the scarce pieces of uncertain information. Details on the expert system were reported in (2). ANNOYANE ACCUSATION In community noise annoyance simulation, accumulation of annoyance caused by different sound sources is an important step in the process. Several perceptual models backed up by laboratory experiments exist, but no detailed data based on annoyance reports is available (1). In particular the principle of compromise poses major problems. In this work a general approach based on probability theory is introduced. Assume two sources of annoyance A and B. Let the subscript denote the quantifier of annoyance. The probability that a single individual in a population is annoyed at level i is assumed to be equal to the percentage of individuals in that population that is annoyed at level i. If the total annoyance is called H, than the probability that an individual is annoyed to level i is given by P(Hk) = ~ P(Hk[Ai, ~j)~(At, ~j), (1) a,j where I indicates conditional probability. The only requirement for this equation to be valid is that all levels of annoyance are covered by the subscripts. The advantage of using equation 1 lies in the fact that the conditional probability will probably be less dependent on exposure levels. This is a basic assumption of most perceptual models. P(Ai, Bj ) gives the probability of simultaneous theoretical annoyance in the absence of the other source. It therefor does not take into account masking or inhibition. In some cases P(Ai, Bj ) can be obtained directly from an exposure model in other cases it is useful to further expand it as P(Ai/Bj)P(Bj). In the absence of geographical clustering of noise sources independence can be assumed, which leads to P(Aa, Bj) = PA. me probability tensor P(Hk lAi, Bj) 1137 TABLE 1. Comparison of annoyance accumulation and sound level summation for total annoyance calculation model moderately annoyed higtiy annoyed 68% low mostprobable 68% high 68%1ow mostprobable 68% high level summation 27 34 48 13 18 25 annoyance accumulation 41 53 66 19 26 34 includes the influence of the presence of the other source, the summation principle and the compromise principle. At this moment there is not enough knowledge available about the probabilities ~(~k 1A,, Bj ). Therefor each element of the tensor is implemented as an uncertain value. This allows to incorporate knowledge as vague as: “if a person is hi~hlv annoyed by sound of source A and highlv annoved by sound of source B, then he will QUITE LIKELY be highlv snnoved by the combination of both sources’!or “if a person is not annoved by sound of source A end hi~hlv ennoved by sound of source B, then he will PROBABLYbe less highly annoved by the combination of both sources” where the fuzzy quantifiers are underlined and the uncertain element of~(HklAi, Bj) is written in capital letters. In community noise annoyance calculations, introducing a secondary condition can often be very helpfulto accu- ratelyquantify F(HklAi, Bj)andto verify the independence ofexposure, Agoodexample are well localized sources (e.g. airports) where aregion can beidentified to which the impact ofone sourceis limited. As an example assume that P(T) istheprobability that aperson lives within anareawhere both sources Aand Bcanbeimpoflant. In~ source A is not present. Total annoyance can than be split as P(~&) = P(Hk IT)P(T) + P(Hk l~)P(~) , or after similar processing as Eqs 1 p(~~) = ~ P(H~lAi, BjlT)p(Ai, BjlT)p(T) + ~p(HklAo, BjlT)p(BjlT)P(T) (2) ~,~ j The first part of the equation describes real accumulation, while the second part only includes the compromise principle. EXAMPLE Within the limited space of this written paper, only one extreme example is given to illustrate the technique. Traffic noise annoyance is assessed in two different ways. First total noise exposure is determined (in a sample taken at random) and multiplied by average dose-response relations to yield a distribution of the population over different annoyance levels. Secondly the exposure to local traffic noise is obtained as well as exposure to highway traffic noise. Both exposure lead to a distribution of the population over annoyance levels. Finally annoyance of both sources is accumulated using the extreme assumptions of exposure independence, a strongest component perceptual model and no context dependence. In this case the tensor p(~k /Ai, Bj ) is reduced to Hnot Hmode.ateiu Hhighlv An Am Ah An Am Ah An Am Ah Bnl OOBn OIO Bnool (3) BmOOOBml10 Bmool BhOOOBh OOO Bhlll without any uncertainty. Certainty intervals (68%) and most probable value for percentage of people highly and moderately annoyed are given in table 1 for both approaches. REFERENCES 1. B. Berghrnd and M.E. Nilsson, “Empirical issues concerning annoyance models for combined community noises;’ Proceedings of Infemoise 97, Budapest, Hungary, 1053-1058, 1997 2. J. De Poorter and D. Botteldooren, “Vagueness and Uncertainty in Community Noise Annoyance Modeling;’ Proceedings Of Internoise 97, Budapest, Hungary, 1207-1210, 1997 1138 
work_f4qlez74vzabda5cn74pnukejy	----	Towards expert systems for the selection of search keys Towards Expert Systems for the Selection of Search Keys* Raya Fidel Graduate School of Library and lnforma tion Science, University of Washington, Seattle, WA 98795 Intermediary expert systems are designed to mediate between end-users and complex information retrieval systems. However, since most of these expert systems are based on text analysis rather than on models of hum man searching, they cannot process requestrelated cri- teria, such as precision or recall requirements. Analysis of the searching behavior of human intermediaries re- vealed a routine for the selection of search keys-free- text or controlled vocabulary-along a decision tree. Examples of decision rules demonstrate that although further research is required, these rules can be auto- mated to significantly enhance the adaptability of inter. mediary expert systems. Introduction It is believed that end-users will very likely search their own requests online when search processes are simplified or made friendlier. The prevailing approach to providing easier and friendlier user-system communication is to de- velop “interface” or “intermediary” systems. Indeed, systems such as CITE [l] are already available for public access, and others, such as CONIT [2] or CANSEARCH [3], are being tested in experimental settings. Through such systems, users are freed from encoun- ters with many peculiarities in databases and search sys- tems, and yet can benefit from a large range of capabili- ties. In particular, users can enter a request in a loosely structured format, preferably in a natural language, sen- tence-like expression. An intermediary system processes the request terms, displays information to users, and asks for some sort of feedback. The information displayed may be in the form of a list of subject areas, databases, search keys (i.e., strings of characters that are entered to *This study was supported by the Graduate School of the University of Washington. Received March I, 1985; revised May 15, 1985; accepted June 13, 1985 be searched for occurrence in pre-defined fields), or ac- tual citations from which users are asked to make a selec- tion, possibly in ranked order. Interactions of this nature usually proceed until users terminate the session. Some intermediary systems are actually helper sys- tems: they support shortcuts in end-user training by pro- viding menu-driven interaction or online help programs. Typically, helper systems require end-users to make most of the decisions during a search process, or they drasti- cally simplify searching by reducing the number of op- tions to a minimum. Intermediary expert systems, on the other hand, can be quite powerful. Expert systems replicate the perfor- mance of an expert in a particular area by incorporating the knowledge of an expert with rules for making infer- ences on the basis of this knowledge. Systems knowledge may or may not be intended to model actual processes as they would be performed by human experts. This article illustrates how to model searching behav- ior of human intermediaries by demonstrating that for- mal rules for the selection of search keys can be extracted from human experts. These rules cannot be incorporated yet into a knowledge base of an intermediary expert sys- tem because they are incomplete and were derived from searching behavior that is limited in its subject area. The work presented here, however, clearly indicates that with more research a complete set of rules can be established. It also provides guidance for future exploration and points to various issues that could be readily considered by designers of intermediary expert systems. Modeling the Selection of Search Keys Search processes consist of three basic intellectual components: (1) definition of query structure; (2) selec- tion of search keys; and (3) evaluation of feedback. We focus here on the second component. In a database that JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE. 37(1):37-44,1986 CCC 0002-8231/86/010037-08$04.00 offers the capability, an intermediary system must exam- ine each term of a request and consider its representa- tion: as a controlled vocabulary key, as a free-text key, or as both. Expert systems vary in the degree of freedom they provide users in this selection: some dutifully search only those search keys designated by a user; some use search keys designated by a user to generate additional search keys; and some automatically generate search keys without user participation. Intermediary expert systems use mapping algorithms to generate search keys. Whether user support and in- volvement are high or low, a term may be mapped to a descriptor (a search key from a controlled vocabulary) through an exact or other kinds of match, or it may not be mapped to a descriptor at all. While some systems, such as CANSEARCH [3], use subject-specific ap- proaches, most existing intermediary systems, expert and helper, use mapping algorithms that are based on text characteristics, such as word-occurrence frequency or statistical associations. The most apparent drawback of text analysis al- gorithms is their lack of sensitivity to request-specific re- quirements that cannot be directly deduced from a re- quest statement. Consider, for example, a request about “the attitudes of anorexic students toward themselves during examinations periods.” One user is interested only in anorexic students, another wants to get all the in- formation available on the topic but is primarily inter- ested in students and is willing to look at material about student behavior during examinations periods. The sys- tem decides, say, to use anorexic students as a single search key, but may or may not suggest the term students as a search key, following its own algorithm. However, when experienced online searchers select search keys, they examine not only the degree of term- descriptor match, but they also consider other factors. Moreover, these additional factors are quite frequently essential to the success of a search. One wonders why the experience of human intermediaries has been neglected by system designers when it is an important source of knowledge. There may be two explanations. First, online search- ing behavior was not being investigated and thus could not contribute knowledge. Second, research and develop- ment in information storage and retrieval is familiar with text analysis because of the long and established experi- ence in automated indexing. Although methods and ap- proaches to automated indexing vary greatly, most of them rely on text analysis and are thus text-oriented rather then user-oriented. Only recently, in an experi- ment at the American Petroleum Institute, a first attempt has been made to develop an automated indexing system which models indexing behavior [4]. As a first step toward knowledge engineering in inter- mediary expert systems, I analyzed the online searching behavior of several human intermediaries and found that online searchers do indeed follow some rules [S]. In their selection of search keys, they utilize informal and some- times highly intuitive decision rules. Moreover, these rules can be detected, examined, and presented in a for- mal structure which can be processed by computers. This formal presentation incorporates knowledge of multiple experts and with further research can be used in a knowl- edge base for intermediary expert systems. The Study Method To examine online searching behavior, eight searchers were observed performing their regular, job-related searching [6]. Searchers who have been searching for more than two years were recruited from among informa- tion specialists in scientific areas, primarily in the life sci- ences. They were studied one at a time, and were asked to verbalize their thought processes during their searching to the degree that speaking out loud would not interfere with their performance. These verbalizations, including the creation of search strategy, were recorded. At the end of the observation period (approximately 10 to 15 searches), each searcher was interviewed to reveal and clarify information not accessible to observation. Data collected for analysis were about one hundred printed search protocols with transcriptions of verbalized thought processes and additional explanations from the final interviews. Each instance in which searchers had selected a search key was then identified and the reason for the specific choice was explicitly noted. Analysis of the first ten search protocols generated a preliminary list of condi- tions under which a particular selection was made. For example, a condition for choosing a free-text key is: to enter straightforwardly a specific concept which might not be a trustworthy descriptor. All the search protocols were then analyzed against this preliminary list of conditions. Each instance in which searchers had selected a search key was listed under the condition to which it applied. Instances whose condition could not be found suggested a new condition to be added to the list. This analysis revealed that most conditions were considered by most searchers, and only a few combi- nations reflect searchers’ individual idiosyncrasies. The list of conditions for the selection of search keys is presented in Figure 1 in the form of a decision tree. This set of decision rules is called here “the selection routine.” The selection routine specifies conditions which are necessary for a searcher to be able to select a particular type of search key. It describes the most commonly se- lected path, but there might be complications. As such, the selection routine is not deterministic: it cannot always accurately predict the selection of search keys unless other factors and their impact are known. This routine groups together similar conditions so it could be pre- sented in a decision tree, but it is not meant to represent a necessary sequence in the selection process. A list of the options in search key selection and the conditions necessary for each option is given in Table 1. 38 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 39 TABLE 1. A List of Options and the Associated Conditions. The Selection Routine OPTION CONDITIONS Use descriptors Add the next broader descriptor in the hierarchy Use generic descriptors in an inclusive mode Limit to retrieval by descriptors Limit to major descriptors Specify document type Use free-text terms Use free-text terms to probe indexing Use descriptors as free- text terms in other databases Use free-text terms for an inclusive search Use free-text terms to introduce uncommon types of search keys Descriptor Searching A term is a common term + it is mapped to a descriptor [A]. A term is a single meaning term + it is mapped to a descriptor + the descriptor is an exact match [D]. the concept has many synonyms [E2]. the concept is not clear to the searcher E31. the concept may not be explicitly mentioned [E4]. the descriptor is a partial match [F]. the descriptor is a broader term [.I]. A term is a single-meaning term + it is mapped to a descriptor + recall needs to be improved [E7]. A term is a single-meaning term + it is mapped to a descriptor -I- recall needs to be improved [E8]. A term is a single-meaning term + it is mapped to a descriptor + precision needs to be improved [E9]. A term is a single-meaning term + it is mapped to a descriptor + precision needs to be improved [ElO]. A term is a single-meaning term + it is mapped to a descriptor + precision needs to be improved [E12]. Free-Text Searching A term is a single-meaning + it is mapped to a descriptor + the concept is not “trustworthy” as an index term [El]. the descriptor is a broader term [HI. it cannot be mapped to a descriptor [K]. A term is a common term + it cannot be mapped to a descriptor [B]. A term is a single-meaning term + it cannot be mapped to a descriptor [L]. A term is a single-meaning term + it is mapped to a descriptor + a request needs to be searched on several databases [E13]. A term is a single-meaning term + it is mapped to a descriptor + the descriptor is a partial match [Cl. A term is a single-meaning term + it cannot be mapped to a descriptor [Ml. Other Combinations Use free-text terms in combination with descriptors Add free-text synonyms to descriptors Add truncated free-text terms to descriptors Add role indicators A term is a single-meaning term + it is mapped to a descriptor + the descriptor is a broader term [I]. A term is a single-meaning term + it is mapped to a descriptor + recall needs to be improved [ES]. A term is a single-meaning term + it is mapped to a descriptor + recall needs to be improved [E6]. A term is a single-meaning term + it is mapped to a descriptor + Change database precision needs to be improved [El I]. A term is a common term + it cannot be mapped to a descriptor [Cl. Before searchers decide how to represent a request term in a query formulation they must answer two central questions: (a) can a term be mapped to a descriptor, and (b) is it a “good” term for free-text retrieval. A searcher maps a term to a descriptor when she/he has decided that a particular descriptor best represents a request term, whether or not there is an exact match between the term and the descriptor. The second question is a little more complex and re- quires some explanation. Searchers consider a term to be a “good” term for free-text searching if it: (1) usually oc- curs in a particular context, (2) is uniquely defined, and (3) is specific in the concept it represents. Such a term will be called here a “single-meaning” term. On the other hand, a term that occurs in more than one context will be called a “common” term. For example, in the request about the attitude of anorexic students toward them- selves during examinations periods, terms such as ano- rexia and students are single-meaning terms. The term examination, on the other hand, is a common term; it can occur in a subject-related context (“the best way to take student examinations”), or in a descriptive capacity (“ex- amination of students’ responses”), or still further, it can be used very loosely to represent the concept of an inquiry of any kind. When a term is a common term, searchers do not have much choice in the selection of search keys: if it can be mapped to a descriptor, searchers almost always enter the descriptor as the search key [A] (i.e., option [A] in Figure 1. and Table 1.) because, by definition, it is not desirable to use a common term as a free-text search key. A common term that cannot be mapped to a descrip- tor almost always results in unsatisfactory retrieval. Searchers have no choice but to enter a free-text key, preferably in combination with other search keys, in or- der to retrieve citations, select some relevant ones and re- view their indexing in an attempt to find descriptors that might possibly be relevant [B]. For example, if the term examination cannot be mapped to a descriptor, one can devise a formulation (using the AND operator) that com- bines the terms students, anorexia, and the free-text term examination. Reviewing a sample of retrieved cita- tions, one may find that all the relevant citations include the descriptor Instructional Tests in their indexing, thus suggesting that this descriptor is an appropriate choice for the representation of the concept examinations. Such probing does not always further a search and searchers may then decide to select a different database: one which does allow the common term to be mapped to a descriptor [Cl. Single-meaning terms provide more options. If a sin- gle-meaning term cannot be mapped to a descriptor, searchers may enter a free-text term as the only search key [K], but they may also probe indexing of relevant ci- tations to make sure that no adequate descriptor is over- looked [L]. 40 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1966 In some cases, searchers may use a free-text key to search for a single-meaning term that cannot be mapped to a descriptor in a special way: they require that it occurs in a field other than the common ones, such as the jour- nal title field [Ml. Suppose a user is interested only in the psychological aspects of anorexic students, and suppose that the term psychology cannot be mapped to a descrip- tor. Searchers may predict that searching for the occur- rence of psychology in the text would retrieve a large number of irrelevant citations, and decide instead to re- trieve citations to articles whose authors are affiliated with organizations which include the stem psych in their titles, or articles that were published in sources whose ti- tles include this stem. The least problematic term is one that is single-mean- ing and also can be mapped to a descriptor. Such a term can be entered either as a controlled vocabulary, as a free-text key, or as both. Here, searchers are free to deal with other factors when selecting search keys. It is useful to show the conditions under which searchers select free- text keys, and those under which they choose to use de- scriptors. Selection of Free-Text Search Keys A single-meaning term can be mapped to a descriptor through an exact match, partial match, or a term might be mapped to a broader descriptor. When a single-mean- ing term is mapped to a descriptor through an exact match, searchers may use a descriptor [D], or elect to consider a variety of factors [El. Partial match usually implies mapping a request term to a narrower descriptor, or to a group of narrower de- scriptors. If suitable, searchers use a free-text key to in- clusively search concepts that are not grouped together by the hierarchy of the controlled vocabulary [G]. If, for ex- ample, the request term students is mapped to descrip- tors such as Foreign Students, College Students, or Un- dergraduates, and a descriptor Students does not exist, the free-text key can be used to retrieve information about any type of student. It should be noted that in some search systems use of the free-text key students also would retrieve citations that are indexed with descriptors which include the term. This is a source for constant con- fusion for searchers because the routine changes from one search system to another. When a single-meaning request term is mapped to a broader descriptor, searchers may prefer to preserve the specificity of the request and use free-text search keys [HI. If they are concerned with the precision of the set to be retrieved, they enter free-text terms in combination (using the AND operator) with the broader descriptor to which it is mapped [I]. Regardless of the degree of match between a single- meaning term and a descriptor, searchers may still prefer to use free-text search keys for three reasons. First, if re- call needs to be improved, searchers use both descriptors and free-text terms as search keys [ES]. For a further in- crease in recall, free-text keys are entered in a truncated form [E6]. Second, if searchers think that a particular descriptor may be assigned inconsistently by indexers, they consider the use of a free-text key to be more trust- worthy [El]. Suppose, for instance, a controlled vocabu- lary includes the descriptor Nutrition and also the de- scriptor Diet-each one to be assigned for distinct representation of a subject. When looking for nutrition- related problems of anorexic students, searchers may find the distinction confusing and thus assume that in- dexers are likely to be inconsistent in assigning these de- scriptors. To compensate for indexers’ errors, they may use both a descriptor and free-text keys, or only free-text keys. Lastly, if a request is to be searched on several data- bases, searchers may map single-meaning terms to de- scriptors in only a few of the relevant controlled vocabu- laries. Running the same query formulation against several databases, they then, in fact, search some of the terms as free-text keys in some of the databases [E13]. Selection of Controlled Vocabulary Search Keys The most straightforward use of a descriptor to repre- sent a single-meaning term is when a term is exactly matched with a descriptor and no other apparent con- straints exist. However, searchers may elect to enter a re- quest term as a descriptor when it is mapped to a descrip- tor through a partial match [F], in which case it is mapped to a narrower descriptor, or when the term is mapped to a broader descriptor [.I]. Such decisions de- pend on the nature of the request, and when searchers suspect that precision might not be satisfactory, they may combine free-text terms with a descriptor [I]. Even single-meaning terms may have some attributes that will make them unattractive for free-text searching. Regardless of the degree or nature of a term-descriptor match, searchers most often prefer to enter a descriptor when: (1) a term has many synonyms [E2]; (2) a concept and its use is not clear to a searcher [E3]; or (3) when a concept is likely to be implied rather then explicitly men- tioned in the searched text [E4]. The previously mentioned request about anorexic stu- dents provides a clear example of the last condition. The request concept attitudes toward themselves can indeed be entered and searched as a free-text phrase in most search systems. However, this concept can be expressed, directly or indirectly, in various other phrases, such as “students displayed attacks of self-hatred”, or “narcis- sism level dropped with time.” Searchers, then, consider descriptors such as Self Image, or Se&Esteem to be more reliable than free-text terms for searching. In addition to providing search keys for “problematic” terms, controlled vocabularies provide special means to improve precision and recall. When searchers perceive precision and/or recall to be unsatisfactory, they may de- cide to take advantage of these means and elect to use a descriptor to represent a single-meaning term. Although JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1988 41 these routines are quite well known, it might be helpful to mention them. When searchers decide that recall needs to be improved, they select a controlled vocabulary key because they can add the next broader descriptor in the hierarchy [E7], or use generic concepts in an inclusive mode [E8]. Controlled vocabularies readily suggest broader descriptors, and thus make it convenient to in- deed broaden a concept. Moreover, broadening the meaning of a concept is not always possible in free-text searching since the broader concept may be a common term. Inclusive searching-which is quite straightfor- ward in descriptor searching-facilitates retrieval of cita- tions indexed under a descriptor as well as those indexed under its narrower terms. Lastly, when searchers predict that precision may not be satisfactory, they may select a controlled vocabulary key because they can exercise various ways of adding weights to search keys such as: to limit to retrieval by de- scriptor only [E9]; to limit a descriptor to be a major de- scriptor [ElO]; to add role indicators [Eli]; or to specify document type [E12]. Discussion The selection routine clearly shows that the process of selecting search keys as performed by online searchers can be formalized into a decision tree. Moreover, several suggestions for improvements in existing systems are im- mediately apparent; others will require more research. First, the pattern of the selection routine illustrates the significance of decisions made during search key selec- tion to the success of a search. This pattern shows that when a term is “not adequate” for searching, i.e., it is a common term and/or it cannot be mapped to a descrip- tor, only a few options are available for searching. Only six of the twenty-five options in the selection routine are suggested for such terms, and some, such as the use of free-text terms to probe indexing, require a fair amount of creativity on the part of an intermediary. On the other hand, when a term is “good” for searching, i.e., it is a single-meaning term and it can be mapped to a descrip- tor, intermediaries can look at several options, as pre- sented in the check-list. In other words, only after termi- nological difficulties or peculiarities in representing request terms have been overcome for searching, can an intermediary consider additional factors that are essen- tial to the success of the search. Therefore, intermediary expert systems must be able to resolve terminological dif- ficulties before they can be equipped to deal with addi- tional factors. Second, using the selection routine, one can identify flaws in existing intermediary expert systems and at the same time propose methods to overcome these failings. While some flaws can be readily identified and possible remedies suggested already at this time, other issues re- quire additional research before their nature can be clearly defined. To demonstrate the ability of the selec- tion routine to illuminate such issues, a few examples are discussed below. One of the flaws in existing intermediary expert sys- tems that readily stands out is their inability to distin- guish between common and single-meaning terms. This distinction is important because if a term is a common term, experienced searchers almost always select it as a descriptor even if they have to change databases (unless they use it in a trial [B]). Yet, to my knowledge, no exist- ing system, provides safeguards to advise end-users against the use of common terms as free-text search keys. For example, when asked about “information retrieval, ” CITE suggested Information Systems, Information Ser- vices, Information Theory, and Information Centers, as descriptors, and retrieval, retrieving, and information as free-text search keys [ 11. Experienced searchers normally will avoid using these common terms as free-text keys, though end-users may prefer them because none of the descriptors exactly matches the original concept. The idea that some terms are not suitable for auto- mated processing is not new. In linguistics, homonymy and polysemy are specific cases of common terms which might be better described as ambiguous terms or terms that belong to more than one semantic domain. These concepts are essential to thesaurus construction. From the information science field, Fugmann, for example, differentiates between “individual” and “general” con- cepts, the latter being non-lexical concepts which are bet- ter searched with controlled vocabulary [7]. In addition, the assumption that some terms carry more information than others is a fundamental premise in automated in- dexing and abstracting. Control over common terms should require relatively modest modifications in intermediary expert systems. First, we have to devise a working definition of what con- stitutes a common term. For this purpose, it would be useful to test the hypothesis that terms which occur with high frequency in a database are also common terms. Suppose, however, that a system selects to define a com- mon term by, say, a consensus among three knowledge- able searchers who are highly experienced with a data- base. These searchers can then check each term in a thesaurus, including those that represent an entry or part of an entry and those that occur in a descriptor or in a lead-in entry, and determine which terms are common. Common terms can then be coded so they are not dis- played or used as free-text search keys. This method requires additional effort to identify com- mon terms that do not occur in a thesaurus or in other semantic networks. Some shortcuts can be devised, how- ever, such as the use of a number of thesauri. On the other hand, other methods to identify common terms may apply to all terms in a database whether or not they are listed in a thesaurus. For instance, a test can be con- ducted to discover *whether a correlation exists between the frequency in which a word occurs in a database and its adequacy for free-text searching. If a well-selected 42 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 sample of databases proves that common terms occur with high frequency in those databases and vice versa, then common terms can be singled out by frequency counts. Once a common term is defined, an intermediary sys- tem can interact with users. Suppose a common term cannot be mapped to a descriptor. By giving messages to users, a system can interrogate them to determine whether to select another database or whether to enter the term as a free-text search key to probe indexing. This interaction may reveal request-related requirements that are not reflected in a request statement, e.g., that the term is central to the request or that it always should be associated with another term. A second example of a flaw in existing intermediary expert systems is their failure to suggest the use of free- text terms for inclusive search [G]. As explained earlier, if the descriptor Students does not exist and a term is then mapped to descriptors such as University Students or Gifted Students through partial match, the free-text term students can be entered to search for any kind of student. This function can be easily automated. How- ever, one should be careful because some terms are not suitable to be entered as free-text keys for inclusive searching. The term attitudes is a case in point. If an exact match does not exist, the term attitudes can be mapped through a partial match to descriptors such as Employee Atti- tudes, Mother Attitudes, or Negative Attitudes. In searching for material about attitudes of students toward themselves, a user may find the last descriptor relevant, but the first two are a source for unwanted retrieval. To search the term attitude as a free-text key would magnify the problem. In this case, the user is better advised to scan all the descriptors to which the term is mapped and to select the relevant ones. Thus, we can designate for each word in a multi-words descriptor whether it can be automatically searched as a free-text key or whether it should be displayed to users when a partial match occurs. At this time, we do not have any scale based on systematic investigations that can de- termine which terms are suitable for inclusive searching in a free-text mode. We can, however, adapt a pragmatic approach and determine the status of each term by con- sulting with experienced online searchers. In the future, research in terminology may provide more rigorous crite- ria. A third example is the inadequacy of intermediary ex- pert systems for term analysis processes. Various statisti- cal approaches could be used to determine which attrib- utes of terms are significant for searching. For instance, the “trustworthiness” of a descriptor can be measured by the degree of consistency with which it is assigned. An extrapolation of the measurement of term consistency suggested by King & Bryant [8] could be used to measure trustworthiness of descriptors. A test database could be indexed by several indexers. A measurement for trust- worthiness could then be determined by the ratio between the number of documents to which a descriptor has been assigned by all indexers and the total number of docu- ments to which it has been assigned by any number of indexers (possibly weighted by the number of indexers se- lecting to assign a descriptor for each document). Here again, each descriptor that is not trustworthy can be flagged. During the search process, a system can then use free-text terms whenever it encounters a single-meaning term which is mapped to such a descriptor. The system may also convey its action to users. As these examples show, parts of the selection routine can be automated quite easily and with existing tech- niques. Other parts of the selection routine, such as, when searchers use a descriptor for a single-meaning term because they do not fully understand the concept it represents, may prove to be unsuitable for the design of intermediary expert systems. There are yet further decisions which can usefully be implemented after additional research is performed. Consider a situation in which a single-meaning term is mapped to a broader descriptor. There are three main options as shown in the decision tree (Figure 1.) at the points [HI, [I], and [J]. Now, suppose the term anorexia nervosa is mapped to the descriptor Appetite Disorders. An intermediary system may decide to enter anorexia nervosa as a free-text search key and possibly retrieve documents in which the subject is only mentioned, rather then discussed. Or, it can combine this free-text key with the descriptor Appetite Disorders, using the AND opera- tor, to retrieve articles that indeed discuss the disorder but may miss relevant documents that were not indexed under the broader descriptor. Or, the system can select to enter the descriptor, in which case relevant documents might still be missed while documents discussing appetite disorders other then anorexia nervosa probably will be re- trieved. No option is better than any other; it all depends on the nature of a request. In other words, one option is re- quired for one type of request and another is required for another type. We do not have yet a general typology of requests that we can use to support the selection of the best option. Further research in online searching behav- ior, however, can provide criteria to be used in automated systems. Statistical approaches may suggest some help. For ex- ample, one may statistically analyze user satisfaction rate with each of the options. Thus, even though we do not know explicitly which type of request requires which op- tion, we can implicitly detect which type is most common among a particular group of users by the option they find to be most satisfactory. We can then design an intermedi- ary expert system that first will always try the most com- monly satisfactory option, then ask for user’s reaction, and then utilize the next option if the first failed to pro- duce acceptable results. A much more promising approach suggests that online JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 43 searching behavior be investigated to reveal under which conditions each of the options is selected. We may find, for example, that when a term that is not central to a re- quest is mapped to a broader descriptor-and a user is primarily concerned with precision-searchers decide to combine a free-text search key with a descriptor. Trans- ferring this condition to automated systems will enable a system to decide when and how to interrogate users about request requirements that are relevant to the search pro- cess. More specifically, a system may proceed searching independently until it encounters a problematic term, such as one that is mapped to a broader descriptor, and then ask users for a specific kind of feedback. In sum- mary, it is not difficult to envision an intermediary expert system which would: (1) identify situations in which a re- quest statement is sufficient, and conversely those in which additional request criteria are needed for the search process to succeed, (2) list the relevant criteria so that users can provide the pertinent data, and (3) act upon data received to improve search results. Quite a powerful system! These few examples clearly demonstrate the benefits that could be gained from the study of searching behavior of human intermediaries, and from utilizing the experi- ence of human intermediaries in the design of intermedi- ary expert systems. The selection routine presented here is not sufficient to develop adaptive algorithms. Many issues, such as the nature of single-meaning and common terms or the con- ditions for the selection of a broader descriptor, need to be further investigated and rigorously defined. This rou- tine identifies problematic points in the search process and provides guidelines for research into searching be- havior that is relevant for automated systems. On the one hand, systems can interrogate users on request parame- ters-a subject users know best. On the other hand, they can select the most appropriate search keys-a decision casual end-users are not well enough informed to make. Based on searchers’ experience, intermediary expert sys- tems can become experts indeed. Acknowledgment The author wishes to thank Dagobert Soergel and Irene L. Travis for their helpful and insightful review of this article. References I. 2. 3. 4. 5. 6. 7. 8. Doszkocs, T. E. “Automatic vocabulary mapping in online search- ing.” International Classification. 10(2):78-83; 1983. Marcus, R. S. “An experimental comparison of the effectiveness of computers and humans as search intermediaries.” Journal of the American Societyfor Information Science. 34(6):381-404; 1983. Pollitt, A. S. “A ‘front-end’ system: An expert system as an online search intermediary.” Aslib Proceedings. 36(5):229-234; 1984. Brenner, E. H.; Lucey, J. H.; Martinez, C. L.; Meleka, A. 1984. “API’s machine aided indexing project.” Science Technology Li- braries. 5(1):49-62; 1984. Fidel, R. “Online searching styles: A case-study-based model of searching behavior.” Journal of the American Societyfor Informa- tion Science. 35(4):211-221; 1984. Fidel, R. “The case study method: A case study.” Library and In- formation Science Research. 6(3):273-288; 1984. Fugmann, R. “The complementarity of natural language and in- dexing languages.” International Classification. 9(3):140-144; 1982. King, D. W.; Bryant, E. C. The evaluation of information services and products. Washington, D.C.: Information Resources Press; 1971, p. 138. 44 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-January 1986 Introduction Modeling the Selection of Search Keys FIG. 1. TABLE 1. The Selection Routine Discussion Acknowledgment References 
work_f5c7couvdvce7lirhnifeq6hka	----	Meysam Rahmani Katigari, Haleh Ayatollahi, Mojtaba Malek, Mehran Kamkar Haghighi EVIDENCE-BASED MEDICINE 80 February 15, 2017|Volume 8|Issue 2|WJD|www.wjgnet.com Fuzzy expert system for diagnosing diabetic neuropathy Meysam Rahmani Katigari, Haleh Ayatollahi, Mehran Kamkar Haghighi, Department of Health Information Management, School of Health Management and Information Sciences, IRAN University of Medical Sciences, Tehran 1996713883, Iran Mojtaba Malek, Division of Endocrinology and Metabolism, Research Center of Firooz­gar Hospital, IRAN University of Medical Sciences, Tehran 159374­7811, Iran Author contributions: All authors contributed to this manu­ script. Supported by The Iran University of Medical Sciences, No. 54­1. Conflict-of-interest statement: There are no conflicts of interest arising from this work. Data sharing statement: No further data are available. Open-Access: This article is an open­access article which was selected by an in­house editor and fully peer­reviewed by external reviewers. It is distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY­NC 4­.0) license, which permits others to distribute, remix, adapt, build upon this work non­commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non­commercial. See: http://creativecommons.org/ licenses/by­nc/4­.0/ Manuscript source: Invited manuscript Correspondence to: Haleh Ayatollahi, Assistant Professor in Medical Informatics, Department of Health Information Management, School of Health Management and Information Sciences, IRAN University of Medical Sciences, No. 6, Shahid Yasami St., Vali­e­Asr St., Tehran 1996713883, Iran. ayatollahi.h@iums.ac.ir Telephone: +98­21­88794­301 Fax: +98­21­88794­301 Received: July 27, 2016 Peer-review started: July 29, 2016 First decision: September 8, 2016 Revised: October 12, 2016 Accepted: December 1, 2016 Article in press: December 2, 2016 Published online: February 15, 2017 Abstract AIM To design a fuzzy expert system to help detect and diagnose the severity of diabetic neuropathy. METHODS The research was completed in 2014 and consisted of two main phases. In the first phase, the diagnostic parameters were determined based on the literature review and by investigating specialists’ perspectives (n = 8). In the second phase, 244 medical records related to the patients who were visited in an endocrinology and metabolism research centre during the first six months of 2014 and were primarily diagnosed with diabetic neuropathy, were used to test the sensitivity, specificity, and accuracy of the fuzzy expert system. RESULTS The final diagnostic parameters included the duration of diabetes, the score of a symptom examination based on the Michigan questionnaire, the score of a sign examination based on the Michigan questionnaire, the glycolysis haemoglobin level, fasting blood sugar, blood creatinine, and albuminuria. The output variable was the severity of diabetic neuropathy which was shown as a number between zero and 10, had been divided into four categories: absence of the disease, (the degree of severity) mild, moderate, and severe. The interface of the system was designed by ASP.Net (Active Server Pages Network Enabled Technology) and the system function was tested in terms of sensitivity (true positive rate) (89%), specificity (true negative rate) (98%), and accuracy (a proportion of true results, both positive and negative) (93%). CONCLUSION The system designed in this study can help specialists Submit a Manuscript: http://www.wjgnet.com/esps/ DOI: 10.4239/wjd.v8.i2.80 World J Diabetes 2017 February 15; 8(2): 80-88 ISSN 1948-9358 (online) 81 February 15, 2017|Volume 8|Issue 2|WJD|www.wjgnet.com Rahmani Katigari M et al . Expert system for diagnosing diabetic neuropathy and general practitioners to diagnose the disease more quickly to improve the quality of care for patients. Key words: Expert systems; Fuzzy logic; Artificial inte- lligence; Diabetes mellitus; Diabetes complications; Diabetic neuropathies © The Author(s) 2017. Published by Baishideng Publishing Group Inc. All rights reserved. Core tip: In this study, an expert system was designed for diagnosing diabetic neuropathy. This system can help specialists to diagnose the disease more quickly by using the most common diagnostic parameters. Even general practitioners can use this system in remote areas to improve the quality of care for patients with diabetes. With it, patients will no longer need to undertake complex procedures, and the care plan can be applied at the right time. Rahmani Katigari M, Ayatollahi H, Malek M, Kamkar Haghighi M. Fuz­z­y expert system for diagnosing diabetic neuropathy. World J Diabetes 2017; 8(2): 80­88 Available from: URL: http:// www.wjgnet.com/194­8­9358/full/v8/i2/80.htm DOI: http:// dx.doi.org/10.4­239/wjd.v8.i2.80 INTRODUCTION One of the biggest challenges currently experienced by healthcare organizations is the increasing burden of chronic diseases posing serious threats to public health in developing countries[1]. Diabetes is one of the world’s most common and costly chronic diseases, and the number of patients suffering from diabetes has been showing an increasing trend in many countries[2]. This can be attributed to population growth, aging, urbanization, prevalence of obesity, and a sedentary lifestyle[2,3]. Long-term complications of diabetes develop gradually and might be disabling or life-threatening - for example, vascular and tissue injuries caused by the progression of diabetes can lead to serious complications, such as retinopathy, nephropathy, cardiovascular dis- ease, cerebrovascular disease, peripheral vascular disease, metabolic disease, and diabetic foot ulcer[4,5]. However, the most common complication of diabetes is impairment of the peripheral neural system, which is known as diabetic neuropathy and a major problem with different signs and symptoms. Compared with other diabetes complications, it is one of the first reasons for hospitalizing patients with diabetes[6]. The severity of pain, decreased or lack of sensation, increased risk of foot ulceration, and amputation are the consequences of diabetic neuropathy[7]. Diabetic peripheral neuropathy is usually seen in more than 10% of patients with type II diabetes. Early diagnosis and treatment is the first step to reduce the incidence of foot ulcers and amputations[8]. The main cost of this disease is related to organ amputation. The risk of lower extremity amputation in patients is significantly high in case of this disease. Nevertheless, almost 85% of amputations are preventable by early detection of the disease, early intervention, good control of diabetes, and patient education[9]. Moreover, several studies show that neuropathy may negatively affect the quality of life for patients with diabetes[10,11]. Owing to the high prevalence of neuropathy among patients with diabetes, it is necessary to conduct annual screening and further evaluation as well as to devise a plan for managing the disease. However, one of the major problems associated with the diagnosis of diabetic neuropathy is the lack of a reliable clinical scale for grading the severity of the disease[12]. A variety of methods are used to detect peripheral neuropathy. These include the nerve conduction velocity test, the vibration perception threshold, the monofilament test, the clinical neuropathy examination, the Toronto clinical scoring system, and the Michigan neuropathy screening instrument (MNSI)[13]. Other than clinical examination, laboratory tests, such as haemoglobin A1c level, fasting blood sugar, and oral glucose tolerance test, along with risk factors like age, sex, renal disease, and smoking need to be considered[14]. It is notable that the boundary between illness and health is not clear in diabetic neuropathy, and it is difficult to express clinical diagnosis as the lack of or the existence of the disease. Since the disease develops on a continuous basis, two-valued logic cannot be used to express this continuity anymore[6]. Therefore, new methods for diagnosing the disease have been considered[15]. Among these methods, special attention has been paid to the development of information technology applications, decision support systems, and fuzzy expert systems[16,17]. The fuzzy expert system is a new version of expert systems that uses fuzzy logic for data processing. In a fuzzy expert system, the inference is conducted by a set of membership functions and fuzzy rules rather than by the rules of two-valued logic[18]. The Fuzzy expert systems are used to describe uncertain phenomena because real-world phenomena are much more complex than an exact and absolute description[19,20]. The ability to implement human science through specific linguistic concepts and fuzzy rules, non-linearity, adaptability of these systems, and the level of accuracy are the most important features of these systems[21]. Although fuzzy expert systems have been designed for different purposes in the healthcare setting, only a few studies have focused on the use of these systems with regard to the diagnosis of diabetic neuropathy[22]. MATERIALS AND METHODS Objective To design a fuzzy expert system to categorize the severity of diabetic neuropathy based on clinical exa- 82 February 15, 2017|Volume 8|Issue 2|WJD|www.wjgnet.com minations and results of laboratory tests. Setting, design, and sample size This study was completed in 2014. The study consisted of two main phases. In the first phase, the parameters required for the diagnosis of diabetic neuropathy were determined on the basis of the literature review[23,24]. These parameters formed a questionnaire to investigate specialists’ views about the importance of each of them. In the second phase, the system was tested by using real data. In the first phase, eight endocrinologists participated in the study. Owing to the limited number of specialists, no sampling method was applied in this phase. In the second phase, 244 medical records were identified from a database located in an endocrinology and metabolism research centre. These records were related to those patients who visited the centre during the first six months of 2014 and who were primarily diagnosed with diabetic neuropathy. Methods for data collection and distribution The questionnaire was distributed among the specialists by one the researchers (MRK), and their views on the importance of the diagnostic parameters were investi- gated. In second phase, a form was used to extract the required data from the medical records. Development of the questionnaire As noted before, the questionnaire was designed based on the literature review[23,24]. It comprised two parts: The first part included the specialists’ demographic information, such as age, gender, and work experience; the second part was designed based on a five-point Likert scale (5 = very important, 4 = important, 3 = relatively important, 2 = less important, 1 = unimportant) and consisted of 15 questions to identify the degree of importance of each diagnostic parameter. The face and content validity of the questionnaire was approved by experts in the field of endocrinology. Its reliability was confirmed by using the test-retest method (α = 0.9). Statistical analysis A data analysis was performed by using SPSS (version 20.0) software, and parameters with a mean score of less than three were excluded to facilitate the process of writing fuzzy rules. To test the system, the sensitivity, specificity, and accuracy of the fuzzy expert system were measured and compared with the final diagnosis recorded in the database. Cohen’s kappa coefficient and the receiver operating characteristic (ROC) curve were used to report data. Participants and recruitment Before conducting the research, the approval of an institutional review board was obtained. In the first phase, the target population comprised endocrinologists working in an endocrinology and metabolism research centre. They were contacted by one of the researchers (MRK) and the research facilitator (MM), and were invited to take part in the study. Their participation in the research was completely voluntary. Regarding the medical records, patient identities were excluded and only the required data was extracted so that it can be used in the process of evaluation. RESULTS Participants As noted before, the first part of the questionnaire included the participants’ demographic information. According to the results, most of the participants were men (n = 5, 62.5%) aged between 30-50 years. The highest frequency (n = 3, 37.5%) was related to the age group of 46-50 years and the specialists with more than 16 years of work experience. Diagnostic parameters for diagnosing diabetic neuropathy The second part of the questionnaire was related to the diagnostic parameters required for diagnosing diabetic neuropathy. This part included the duration of diabetes, the symptom assessment based on MNSI, the sign examination based on MNSI, and the related laboratory tests. Table 1 presents the specialists’ views in relation to the importance of the aforementioned diagnostic parameters. As Table 1 shows, from the specialists’ point of view, the most important diagnostic parameters were the duration of diabetes (4.88 ± 0.35), the glycolysis haemoglobin level (4.50 ± 0.75), and the score of the sign examination based on the Michigan questionnaire (4.38 ± 0.51). The lowest degree of importance (2.13 ± 0.83) was related to the amount of phosphorus in blood. After determining the diagnostic parameters of diabetic neuropathy, the semantic network of the expert system was drawn (Figure 1). Designing a fuzzy expert system As can be seen in the above figure, the ultimate goal, namely diagnosing diabetic neuropathy, is shown in the centre, and the diagnostic parameters are in the leaf nodes. In order to design the fuzzy expert system, all input variables were fuzzified based on membership functions. The system had seven input variables: The duration of diabetes, the score of the symptom examina- tion based on the Michigan questionnaire, the score of the sign examination based on the Michigan ques- tionnaire, the glycolysis haemoglobin level, fasting blood sugar, blood creatinine, and albuminuria. The system also had one output variable, which was the severity of diabetic neuropathy. The rules of the expert system were written based on the semantic network, consulting a specialist, and giving the same weight to all rules. The inference engine of the system was designed by using the Mamdani inference method. Figure 2 provides Rahmani Katigari M et al . Expert system for diagnosing diabetic neuropathy 83 February 15, 2017|Volume 8|Issue 2|WJD|www.wjgnet.com an overview of the fuzzy inference architecture of the system. Finally, the graphical user interface of the expert system was designed by using Active Server Page. Network Enabled Technology (ASP.NET). It is an open- source server-side web application framework designed for web development to produce dynamic web pages (Figure 3). The input variables, such as the duration Degree of importance Unimportant (1) Less important (2) Relatively important (3) Important (4) Very important (5) Mean ± SD Duration of diabetes 0 0 0 1 (12.5%) 7 (87.5%) 4.88 ± 0.35 Symptom assessment based on MNSI 0 0 1 (12.5%) 5 (62.5%) 2 (25%) 4.13 ± 0.64 Sign examination based on MNSI 0 0 0 5 (62.5%) 3 (37.5%) 4.38 ± 0.51 HbA1c 0 0 1 (12.5%) 2 (25%) 5 (62.5%) 4.50 ± 0.75 CBC 1 (12.5%) 3 (37.5%) 4 (50%) 0 0 2.38 ± 0.74 FBS 0 0 0 6 (75%) 2 (25%) 4.25 ± 046 ESR 1 (12.5%) 3 (37.5%) 3 (37.5%) 1 (12.5%) 0 2.52 ± 092 Oral GTT 1 (12.5%) 4 (50%) 1 (12.5%) 2 (25%) 0 2.50 ± 1.06 Albuminuria 0 1 (12.5%) 1 (12.5%) 4 (50%) 2 (25%) 3.88 ± 0.99 TSH 2 (25%) 1 (12.5%) 3 (37.5%) 2 (25%) 0 2.63 ± 1.18 B12 Vitamin 2 (25%) 1 (12.5%) 1 (12.5%) 4 (50%) 0 2.88 ± 1.35 BUN 1 (12.5%) 3 (37.5%) 3 (37.5%) 1 (12.5%) 0 2.38 ± 0.91 BCr 0 1 (12.5%) 2 (25%) 5 (62.5%) 0 3.50 ± 0.75 Calcium 2 (25%) 1 (12.5%) 4 (50%) 1 (12.5%) 0 2.50 ± 1.06 Phosphorus 2 (25%) 3 (37.5%) 3 (37.5%) 0 0 2.13 ± 0.83 Table 1 The degree of importance of the diagnostic parameters for diagnosing diabetic neuropathy from the specialists’ perspectives BCr: Blood Creatinine; BUN: Blood urea nitrogen; TSH: Thyroid-stimulating hormone; GTT: Glucose tolerance test; ESR: Erythrocyte sedimentation rate; MNSI: Michigan Neuropathy Screening Instrument; HbA1c: Hemoglobin A1c; CBC: Complete blood count; FBS: Fasting blood sugar. Feeling in Legs and feet Score of history MNSI questionnaire Score of physical assessment Duration of diabetes Diagnosis of diabetic neuropathy Principal lab test Albumin Fasting blood sugar Glycosylated hemoglobin A1C Creatinine Monofilament Vibration perception at great toe Ankle reflexes Ulceration Appearance of feet Figure 1 The semantic network of the expert system. MNSI: Michigan Neuropathy Screening Instrument. Rahmani Katigari M et al . Expert system for diagnosing diabetic neuropathy 84 February 15, 2017|Volume 8|Issue 2|WJD|www.wjgnet.com of diabetes, the results of laboratory tests, and scores obtained from the Michigan questionnaire, could be entered into the system manually either in the textual or in the numerical format based on the user’s choice. The output variable, namely the severity of the disease, which was shown as a number between zero and 10, had been divided into four categories: absence of the disease, (the degree of severity) mild, moderate, and severe. Figure 4 shows the risk of diabetic neuropathy based on the scores obtained from the Michigan question- naire. According to Figure 4, by increasing the scores Duration-of-diabete (2) History-MNSI (3) Physical-examination-MNSI (3) Glycosylated-hemoglobin-A1c (3) Fasting-blood-sugar (3) Urine-albumin (2) Creatinine (2) Degree of DN-editRule9 (mamdani) 76 rules Degree-of-diabetic-neuropathy (4) System degree of DN-EditRule9: 7 inputs, 1 outputs, 76 rules Figure 2 An overview of the fuzzy inference architecture of the system. Figure 3 The graphical user interface of the fuzzy expert system. Rahmani Katigari M et al . Expert system for diagnosing diabetic neuropathy 85 February 15, 2017|Volume 8|Issue 2|WJD|www.wjgnet.com obtained from the Michigan questionnaire, the severity of diabetic neuropathy will increase accordingly. System function evaluation The system was tested by using real data. In total, the records of 244 patients with diabetic neuropathy were identified. However, 31 records were excluded due to the incompleteness of clinical data. The remaining records (n = 213) included 118 patients who were diagnosed with diabetic neuropathy, while diagnosis was ruled out for the rest (n = 95). The system function was tested in terms of sensitivity (true positive rate), specificity (true negative rate), and accuracy (proportion of the true results, both positive and negative), which were 89%, 98%, and 93%, respectively. Finally, the system’s output was compared with the final diagnoses made by the specialists and recorded in the patients’ records. These diagnoses were made by using the nerve conduction velocity test, the vibration perception threshold, the monofilament test, and the clinical neuropathy examination. The comparison was conducted by using the Kappa coefficient and the K value was 0.6. According to Landis and Koch, a Kappa value between 0.4 and 0.75 shows a fair to good agree- ment[25]. Therefore, the system designed in this study showed a fair to good level of similarity between the system’s function and the specialists’ diagnoses. The ROC curve presents the results of testing the system (Figure 5). As can be seen in the above figure, the ROC curve is ideal. It is close to the high point of the square that represents an appropriate function of the system. DISCUSSION As mentioned before, one of the most common long- term complications of diabetes mellitus is diabetic neuropathy. In order to control this complication, it is important to diagnose it both accurately and timely[10]. Although there are a variety of methods to detect the disease, it is difficult to diagnose it at the very early stage[13]. Therefore, the use of IT applications, such as fuzzy expert systems, is suggested. In the present study, seven diagnostic parameters- the duration of diabetes, the symptom assessment, the sign examination based on the MNSI, the glycolysis haemoglobin level, fasting blood sugar, blood creatinine, and albuminuria-were considered as input variables, and the severity of diabetic neuropathy was considered as an output variable. These variables were selected based on the specialists’ perspectives and the literature review. Similarly, the knowledge and experience of four experts in the field of diabetic neuropathy was investigated in the study conducted by Picon et al[22] to determine the diagnostic parameters and to design a knowledge-based system. In their research, four inputs variables included symptom, the sign assessment based on the Michigan questionnaire, the glycolysis haemoglobin level, and the duration of diabetes. The output of the system classified the severity of diabetic neuropathy in three categories: Mild, moderate, and severe. In contrast with the study of Picon et al[22] the number of input variables increased History-MNSI Physical-examination-MNSI D eg re e- of -d ia be ti c- ne ur op at hy 8 7 6 5 4 3 10 5 0 0 2 4 6 8 10 Figure 4 The risk of diabetic neuropathy based on the scores of the Michigan Neuropathy Screening Instrument questionnaire. MNSI: Michigan Neuropathy Screening Instrument. Rahmani Katigari M et al . Expert system for diagnosing diabetic neuropathy 86 February 15, 2017|Volume 8|Issue 2|WJD|www.wjgnet.com in the current research and laboratory test results were included to improve the accuracy of diagnosis. Similarly, Neshat et al[26]’s study considered six input variables and one output variable to diagnose liver disorders. To diagnose heart ailments, Adeli et al[27] used 12 input variables and considered the diagnosis of heart diseases as the output variable. In the present study, values between zero and 10 were considered for the output variable, which was the severity of diabetic neuropathy. An increase in the value of output variable showed the level of severity for diabetic neuropathy. In the current study, the fuzzy sets and membership functions for each of the seven input variables and the output variable were finalized after consulting a specialist. This approach can help eliminate the rules that could be covered by other rules, and finally, 76 rules were used to design the system. Similarly, DoostHoseini et al[28] consulted doctors to reduce the number of rules to an appropriate number. In another study, Zolnoori et al[29] developed a fuzzy expert system for diagnosing asthma. Given that the patients’ records were incom- plete, an indirect approach was used to develop the system’s knowledge base. In this approach, the resear- chers reviewed books and scientific papers, and also conducted structured and unstructured interviews with doctors and patients. Having analysed the data, the most important variables useful for diagnosing asthma were identified. In the present study, the system interface was designed by using ASP.NET rather than matrix labora- tory (MATLAB). In fact, web-based applications have more flexibility and can be used by multiple users at the same time. Ease of use is another feature of these systems, which, in turn, can increase the work efficiency. In this study, the output of the system was divided into four categories: The absence of the disease, mild, moderate, and severe. In contrast, Picon et al[22] classified the severity of neuropathy into three categories: Mild, moderate, and severe. Moreover, the specificity and sensitivity of the system were not reported in their study. In the current study, the specificity of the system was 98%, which shows a high level of system performance. Also, there was a relatively good agreement between the system’s function and the diagnoses recorded by the specialists. Although other methods of diagnosis were not considered in the current study, the specificity and sensitivity of the system highly suggested that such a system could help physicians to diagnose the disease more quickly by using parameters like results of laboratory tests. In the current study, the main aim was to develop an expert system for diagnosing diabetic neuropathy. Therefore, the clinical effectiveness of the system was not evaluated due to resource restrictions. Conducting evaluation studies after implementing the system in the actual healthcare setting would help determine the impact of the system on the health status of patients. In conclusion, an expert system was designed for diagnosing diabetic neuropathy in this study. As dia- betic neuropathy is a chronic disease that may have serious consequences, early diagnosis of the disease is important to control it. The system designed in the current study could help specialists to diagnose the disease more quickly by using the most common dia- gnostic parameters. General practitioners can use such a system in remote areas to improve the quality of care for patients with diabetes. With it, the disease can be diagnosed more easily and quickly. There is no need to undertake complex procedures, and the care plan can be applied at the right time. Further research is suggested to increase the number of variables to improve the accuracy, sensitivity, and specificity of the system. Moreover, the feasibility of using this method in daily clinical practice and its impact on the efficiency and Diagonal segments are produced by ties 1 - Specificity 0.0 0.2 0.4 0.6 0.8 1.0 1.0 0.8 0.6 0.4 0.2 0.0 S en si ti vi ty ROC curve Figure 5 The receiver operating characteristic curve. Rahmani Katigari M et al . Expert system for diagnosing diabetic neuropathy 87 February 15, 2017|Volume 8|Issue 2|WJD|www.wjgnet.com cost-effectiveness compared to those of other methods need to be investigated in future studies. COMMENTS Background One of the major problems associated with the diagnosis of diabetic neuropathy is the lack of reliable clinical scale for grading the severity of the disease. A variety of methods, such as the nerve conduction velocity test, the vibration perception threshold, and the monofilament test, are used to detect the peripheral neuropathy. In addition to clinical examination, laboratory tests and risk factors of the disease such as age, sex, renal disease, and smoking need to be considered. Research frontiers Since the disease usually develops on a continuous basis, two-valued logic cannot be used to express this continuity any more. Therefore, new methods for diagnosing the disease have been considered. Among these methods, the development of information technology applications, decision support systems, and fuzzy expert systems have received special attention. Innovations and breakthroughs In order to diagnose diabetic neuropathy, clinical examinations as well as results of laboratory tests like the haemoglobin A1c level, fasting blood sugar, and the oral glucose tolerance test should be considered. In this study, information technology was used to design a fuzzy expert system to diagnose the severity of diabetic neuropathy based on clinical examinations and laboratory tests. Applications The system designed in the current study can help specialists to diagnose the disease more quickly by using the most common diagnostic parameters. General practitioners, too, can use it in remote areas to improve the quality of care for patients with diabetes. With it, the disease can be diagnosed more easily and quickly. There is no need to undertake complex procedures, and the care plan can be applied at the right time. Terminology The fuzzy expert system is a new version of expert systems that uses fuzzy logic for data processing. A fuzzy expert system is used to describe uncertain phenomena because the real-world phenomena are much more complex than an exact and absolute description. The most common complication of diabetes is impairment of the peripheral neural system, which is known as diabetic neuropathy. Peer-review This is interesting and important paper for diagnosis of diabetic complications. The paper is well-written and focused. REFERENCES 1 Nolte E, McKee M. Caring for people with chronic conditions: A health system perspective. England: Open University Press, 2008: 1­24­5 2 Seino Y, Nanjo K, Tajima N, Kadowaki T, Kashiwagi A, Araki E, Ito C, Inagaki N, Iwamoto Y, Kasuga M, Hanafusa T, Haneda M, Ueki K. Report of the committee on the classification and diagnostic criteria of diabetes mellitus. J Diabetes Investig 2010; 1: 212­228 [PMID: 24­84­34­35 DOI: 10.1111/j.204­0­1124­.2010.00074­.x] 3 Mehri A, Hasani­Ranjbar S, Larijani B, Abdollahi M. A systematic review of efficacy and safety of Urtica dioica in the treatment of diabetes. Int J Pharmacol 2011; 7: 161­170 [DOI: 10.3923/ ijp.2011.161.170] 4­ Simon GE, Katon WJ, Lin EH, Ludman E, VonKorff M, Ciechanowski P, Young BA. Diabetes complications and depression as predictors of health service costs. Gen Hosp Psychiatry 2005; 27: 34­4­­351 [PMID: 16168795 DOI: 10.1016/j.genhosppsych.2005.04­. 008] 5 Svensson M, Eriksson JW, Dahlquist G. Early glycemic control, age at onset, and development of microvascular complications in childhood­onset type 1 diabetes: a population­based study in northern Sweden. Diabetes Care 2004­; 27: 955­962 [PMID: 1504­7655 DOI: 10.2337/diacare.27.4­.955] 6 Gordois A, Scuffham P, Shearer A, Oglesby A, Tobian JA. The health care costs of diabetic peripheral neuropathy in the US. Diabetes Care 2003; 26: 1790­1795 [PMID: 12766111 DOI: 10.2337/ diacare.26.6.1790] 7 Larejani B, Zahedi F. Epidemiology of diabetes mellitus in Iran. Iranian Journal of Diabetes and Metabolism 2001; 1: 1­8 8 Liu F, Bao Y, Hu R, Zhang X, Li H, Zhu D, Li Y, Yan L, Li Y, Lu J, Li Q, Zhao Z, Ji Q, Jia W. Screening and prevalence of peripheral neuropathy in type 2 diabetic outpatients: a randomiz­ed multicentre survey in 12 city hospitals of China. Diabetes Metab Res Rev 2010; 26: 4­81­4­89 [PMID: 20661939 DOI: 10.1002/dmrr.1107] 9 Cheer K, Shearman C, Jude EB. Managing complications of the diabetic foot. BMJ 2009; 339: b4­905 [PMID: 19955124­ DOI: 10.1136/bmj.b4­905] 10 Vinik AI, Park TS, Stansberry KB, Pittenger GL. Diabetic neuro­ pathies. Diabetologia 2000; 43: 957­973 [PMID: 10990072 DOI: 10.1007/s0012500514­77] 11 Said G. Diabetic neuropathy­­a review. Nat Clin Pract Neurol 2007; 3: 331­34­0 [PMID: 1754­9059 DOI: 10.1038/ncpneuro0504­] 12 Bostani A, Homayuonfar H. The Relationship between NCS Findings and Toronto Clinical Scoring System of Neuropathy in Diabetic Polyneuropathy. Journal of Kermanshah University of Medical Sciences 2006; 88: 2358­2367 13 Rahman M, Griffin SJ, Rathmann W, Wareham NJ. How should peripheral neuropathy be assessed in people with diabetes in primary care? A population­based comparison of four measures. Diabet Med 2003; 20: 368­374­ [PMID: 127524­85 DOI: 10.104­6/ j.14­64­­54­91.2003.00931.x] 14­ Ziegler D, Rathmann W, Dickhaus T, Meisinger C, Mielck A. Neuropathic pain in diabetes, prediabetes and normal glucose tolerance: the MONICA/KORA Augsburg Surveys S2 and S3. Pain Med 2009; 10: 393­4­00 [PMID: 19207236 DOI: 10.1111/ j.1526­4­637.2008.00555.x] 15 Toloie Ashlaqi A, Mohsen Taheri S. Designing an expert system for suggesting the blood cancer treatment. Journal of Health Administration 2010; 13: 4­1­50 16 Wei HY, Lu CS, Lin TH. Exploring the P2 and P3 ligand binding features for hepatitis C virus NS3 protease using some 3D QSAR techniques. J Mol Graph Model 2008; 26: 1131­114­4­ [PMID: 18024­210 DOI: 10.1016/j.jmgm.2007.10.005] 17 Riazi H, Larijani B, Langariz­adeh M, Shahmoradi L. Managing diabetes mellitus using information technology: a systematic review. J Diabetes Metab Disord 2015; 14: 4­9 [DOI: 10.1186/s4­0200­015­ 0174­­x] 18 Riahi-Madvar H, Ayyoubz­adeh SA, Khadangi E, Ebadz­adeh MM. An expert system for predicting longitudinal dispersion coefficient in natural streams by using ANFIS. Expert Syst Appl 2009; 36: 8589­8596 [DOI: 10.1016/j.eswa.2008.10.04­3] 19 Keshwani DR, Jones DD, Meyer GE, Brand RM. Rule­based Mamdani­type fuz­z­y modeling of skin permeability. Appl Soft Comput 2008; 8: 285­294­ [DOI: 10.1016/j.asoc.2007.01.007] 20 Saritas I, Oz­kan IA, Allahverdi N, Argindogan M. Determination of the drug dose by fuz­z­y expert system in treatment of chronic intestine inflammation. J Intell Manuf 2009; 20: 169­176 [DOI: 10.1007/s1084­5­008­0226­x] 21 Sadoughi F, Sheikhtaheri A, Meidani Z, Shahmoradi L. Manage­ ment information system (concepts, structure, development and evaluation). Tehran: Jafari, 2011 22 Picon AP, Ortega NR, Watari R, Sartor C, Sacco IC. Classification of the severity of diabetic neuropathy: a new approach taking uncertainties into account using fuz­z­y logic. Clinics (Sao Paulo) 2012; 67: 151­156 [PMID: 2235824­0 DOI: 10.6061/clinics/2012(02)10] 23 Crawford F, Anandan C, Chappell FM, Murray GD, Price JF, Sheikh A, Simpson CR, Maxwell M, Stansby GP, Young MJ, COMMENTS Rahmani Katigari M et al . Expert system for diagnosing diabetic neuropathy 88 February 15, 2017|Volume 8|Issue 2|WJD|www.wjgnet.com Abbott CA, Boulton AJ, Boyko EJ, Kastenbauer T, Leese GP, Monami M, Monteiro­Soares M, Rith­Najarian SJ, Veves A, Coates N, Jeffcoate WJ, Leech N, Fahey T, Tierney J. Protocol for a systematic review and individual patient data meta­analysis of prognostic factors of foot ulceration in people with diabetes: the international research collaboration for the prediction of diabetic foot ulcerations (PODUS). BMC Med Res Methodol 2013; 13: 22 [PMID: 234­14­550 DOI: 10.1186/14­71­2288­13­22] 24­ Boulton AJ, Vinik AI, Arez­z­o JC, Bril V, Feldman EL, Freeman R, Malik RA, Maser RE, Sosenko JM, Ziegler D. Diabetic neuropathies: a statement by the American Diabetes Association. Diabetes Care 2005; 28: 956­962 [PMID: 15793206 DOI: 10.2337/diacare.28.4­.956] 25 Landis JR, Koch GG. A review of statistical methods in the analysis of data arising from observer reliability studies (Part II). Stat Neerl 1975; 29: 151­161 [DOI: 10.1111/j.14­67­9574­.1975.tb00259.x] 26 Neshat M, Yaghobi M, Naghibi M, Esmaelz­adeh A. Fuz­z­y Expert System Design for Diagnosis of liver disorders. Knowledge Acquisition and Modeling; International Symposium on Az­ad University of Mashhad Iran: IEEE, 2008: 252­256 [DOI: 10.1109/ kam.2008.4­3] 27 Adeli A, Neshat M. A fuz­z­y expert system for heart disease diagnosis. Proceedings of international multi conference of engineers and computer scientists. Hong Kong: Citeseer, 2010 28 DoostHoseini E, Hassanpour­ez­atti M, Navidi H, Abachi T. A Fuz­z­y expert system for prescribing atorvastatin optimum dose. Koomesh 2011; 13: Pe4­3­Pe4­9 29 Zolnoori M, Zarandi MH, Moin M. Application of intelligent systems in asthma disease: designing a fuz­z­y rule­based system for evaluating level of asthma exacerbation. J Med Syst 2012; 36: 2071­2083 [PMID: 21399914­ DOI: 10.1007/s10916­011­9671­8] P- Reviewer: Ido Y, Pastromas S S- Editor: Kong JX L- Editor: A E- Editor: Wu HL Rahmani Katigari M et al . Expert system for diagnosing diabetic neuropathy © 2017 Baishideng Publishing Group Inc. All rights reserved. Published by Baishideng Publishing Group Inc 8226 Regency Drive, Pleasanton, CA 94588, USA Telephone: +1-925-223-8242 Fax: +1-925-223-8243 E-mail: bpgoffice@wjgnet.com Help Desk: http://www.wjgnet.com/esps/helpdesk.aspx http://www.wjgnet.com 80 WJDv8i2-Back Cover 
work_f5ozgouflzd4no5ka6mm4wsjlm	----	NASA Technical Reports Server (NTRS) 
work_f7v4i2gs45es7gspubjuwpo2vq	----	None 
work_fcg3oviwu5hq5cr44uhgenroxq	----	NASA Technical Reports Server (NTRS) 
work_fhuo7yphnzdujox5ixsw6b6lfa	----	Variable neighborhood search heuristics for a test assembly design problem Jordi Pereiraa,b,∗, Mariona Vilàb aDepartamento de Ingenieŕıa Industrial, Universidad Católica del Norte, Av. Angamos 0610, Antofagasta, Chile bEscola Universitària d’Enginyeria Tècnica Industrial de Barcelona, Universitat Politècnica de Catalunya. C. Comte Urgell, 187 1st. Floor, 08036, Barcelona, Spain Abstract Test assembly design problems appear in the areas of psychology and educa- tion, among others. The goal of these problems is to construct one or multiple tests to evaluate some criteria. This paper studies a recent formulation of the problem known as the one-dimensional minimax bin-packing problem with bin size constraints (MINIMAX BSC). In the MINIMAX BSC, items are initially divided into groups and multiple tests need to be constructed using a single item from each group, while minimizing the difference among the tests. We first show that the problem is NP-Hard, which remained an open question. Second, we propose three different local search neighborhoods derived from the exact resolution of special cases of the problem, and combine them into a Variable Neighborhood Search (VNS) metaheuristic. Finally, we test the proposed al- gorithm using real-life-based instances. The results show that the algorithm is able to obtain optimal or near-optimal solutions for instances with item pools with up to 60.000 items. Consequently, the algorithm is a viable option to de- sign large-scale tests, as well as to provide tests for online small-sized situations such as those found in e-learning platforms. Keywords: Bin Packing, Test Assembly Design, Test Splitting, Variable Neighborhood Search. 1. Introduction Test assembly design studies the problem of selecting the optimal subset of items (questions) among those available to define one or several tests (question- naires) subject to specific constraints and objectives. These problems appear in multiple areas, specifically in test design (van der Linder, 2005) both for edu- ∗Corresponding author Email addresses: jorgepereira@ucn.cl, jorge.pereira@upc.edu (Jordi Pereira), mariona.vila.bonilla@upc.edu (Mariona Vilà) Preprint submitted to Expert Systems with Applications January 19, 2015 cation (Porter et al., 2013; Sun et al., 2008) and psychology studies (Veldkamp, 2005). Based on the specific objective of the test, as well as the characteristics of the questions and questionnaires, different problems and formulations may appear. One of these formulations arises when considering the problem as a10 packing problem, specifically as a one dimensional Bin Packing Problem (BPP) with additional constraints (see Dyckhoff (1990) for a classification of packing problems, and Wäscher et al. (2007) for an update of the previous classification). In the BPP formulation, the questions and questionnaires are interpreted as the items and the bins of the BPP, respectively, and the weight of the items corresponds to some metric (e.g., the difficulty of the question) that needs to be considered during the test design. The objective of the problem is to find disjoint subsets of all the items so that the sum of weights (i.e., the total difficulty) of each subset is as evenly matched as possible. In this work we consider the formulation introduced in Brusco et al. (2013).20 This formulation is known as the one-dimensional minimax bin-packing problem with bin size constraints (MINIMAX BSC). To solve the problem, the authors propose a mixed zero-one integer linear programming model that is then solved using the CPLEX commercial software. For large real-life problems, the authors propose a Simulated Annealing (SA) algorithm that outperforms the quality of the solutions provided by CPLEX. While Brusco et al. (2013) do not address the complexity of the problem, the analysis of the results showed that all of the algorithms proposed in the paper were able to optimally solve large problems with 2 questionnaires, leading the authors to hypothesize that the special case of 2 questionnaires may be solvable30 in polynomial time (Brusco et al., 2013, p. 623). Furthermore, the quality of the solutions provided by the SA deteriorates as the number of questionnaires in- creases, which may identify possible improvements if different solution methods were used. 1.1. Contributions of this work In this paper, we address the complexity of the MINIMAX BSC and demon- strate that the MINIMAX BSC is NP-Hard when two questionnaires are to be constructed and strongly NP-Hard when more than two questionnaires are re- quired. We also show that the case with two sets of questions can be optimally solved in polynomial time and that the case with two questionnaires can be40 efficiently solved as a subset sum knapsack problem (Kellerer et al., 2004). Based on these special solvable cases, we derive a new constructive heuristic and three local search neighborhoods. These methods are then combined in a Variable Neighborhood Search (VNS) metaheuristic (Mladenović & Hansen, 1997) to provide a combined approach to solve the problem. VNS is a popu- lar metaheuristic approach (Mladenović & Hansen, 1997; Hansen et al., 2010) based on the use and exploration of different local neighborhoods of an incum- bent solution and a mechanism to escape from local optima by restarting the search on random neighbor solutions. Both mechanisms provide a method to systematically explore the solution space.50 2 The results of the method clearly outperform those provided by previous algorithms, such as the SA method proposed in Brusco et al. (2013). More specifically, the proposed method is able to routinely reach a BPP-based lower bound in instances with a large number of questions per questionnaire (i.e., over ten questions per questionnaire) with characteristics of real-life problems, and still manages to obtain small deviations from the theoretical bound in other cases. 1.2. Paper outline The remainder of the paper is structured as follows. In Section 2, we study the problem, describe its mathematical formulation, address the issue of com-60 plexity and identify special cases that can be efficiently solved. We also outline some of the work in the literature devoted to test design and other related prob- lems, such as the BPP. In Section 3, we describe the proposed local searches and a constructive heuristic. We also propose their combination in a VNS-based method. In Section 4, we describe the results of a computational experiment to assess the quality of the proposed algorithm. Finally, in Section 5, we provide conclusions of the present work and identify some future areas of research. 2. Problem description 2.1. Literature review Test assembly design was early identified as a combinatorial optimization70 problem. According to Van der Linden (van der Linden, 1998), an initial work in the field (Birnbaum, 1968) proposed a three step process to design and con- struct tests: (1) identification of the goal of the test under construction; (2) identification of a target function in accordance with the goal; and (3) selection of items based on the target function and the fulfillment of some constraints. These steps clearly mimic the formulation and resolution of any other combina- torial optimization problem. These three steps are preceded by a preliminary one in which the item pool is constructed. Item pool construction requires the estimation of the information function of the items, a step that is performed using and Item Response Theory80 (IRT), see Veldkamp (2013) for a basic description and a critique of IRT, and Lu (2014) for a description of an estimation method. The information function is then discretized into specific ability levels of interest (such as the pass/fail ability level for an examination test) that provide coefficients for the objective and/or the constraints of the combinatorial optimization model. The present paper only considers the item selection process. Even when we limit our attention to the selection step, the literature of the field is very broad. We can divide previous work into three groups: (1) offline single test construction, in which the objective is to design a single test, which is com- pletely constructed before its use; (2) online single test construction, in which90 the objective is to design a single test, which is sequentially constructed using the information provided by the answers to previous items; and (3) multiple 3 test construction, in which multiple tests are constructed together in order to evaluate examinees with different, but equivalent, test forms. Note that a method to solve any of these problems can be used to solve other problems after some modifications, and consequently we offer a review of recent procedures for each of these three methods. Offline single test construction is known in the literature as static test gen- eration (Nguyen & Fong, 2013), or automated test assembly (Veldkamp, 2013). A general reference on the area (van der Linder, 2005) provides a general in-100 teger programming formulation for the problem. The proposed objective is to maximize the minimum amount of discrimination index on any ability level of interest, while satisfying some constraints on the different requirements of the test. Veldkamp (2013) considers the model in van der Linder (2005) under ro- bustness considerations on coefficients provided during item pool construction. The effect of using a robust formulation is then verified by constructing a test using a 306 items pool, and comparing the solutions provided by both models. The robust solution is shown to be less sensitive to errors in the evaluation of the information function. Nguyen & Fong (2013) identifies the problem as a multidimensional knap-110 sack problem (Kellerer et al., 2004), in which the test tries to maximize the discrimination degree of the test (how good the question is at recognizing user proficiency) under constraints on number of questions, time to perform the test, average difficulty and number of questions per topic. The model is then solved using a branch and cut based procedure and the solutions are compared to previous methods found in the literature on instances with up to 50.000 items. Online single test construction corresponds to the construction of the test along with its resolution and it is usually known in the literature as CAT (com- puterized adaptive test). Due to the variable nature of the problem, the methods are usually constructive and item selection makes use of greedy rules. Further-120 more, CATs try to take into account that multiple tests are to be generated and, thus some randomization is introduced to avoid excessive use of a subset of questions. In He et al. (2014), four different item-selection methods are evaluated. These methods take into account multiple particularities, such as side con- straints, content specifications, time requirements, or item formats, among oth- ers. The heuristics are compared in terms of their use of items (how good are they providing multiple different tests) and their accuracy to measure the information function. Edmonds & Armstrong (2009) considers a hybrid between offline and online130 single test construction, denoted as Multiple Stage adaptive Test (MST). The hybrid divides items into groups, each belonging to a different stage of the test. When an item is to be selected, the method randomly takes an item from the corresponding group by taking into account the stage and the level shown by the examinee on previous items. The definition of each group is modeled and solved using a commercial solver (CPLEX) for a 1.336 items pool. While IRT is the predominant method to evaluate the information provided by the items, alternatives exist. In Smits & Finkelman (2014) a health-oriented 4 alternative is put forward. The objective is now to minimize the time required to administer a diagnosis. Consequently, the method minimizes the number of140 questions required to identify the expected total score of the test. The method applies an ordinal regression on the previous answers in order to stop admin- istering questions once a required degree of precision on the final result of the test has been reached. The applicability of the model is then tested on a real data set provided by a screener for depression. Multiple test construction, or parallel test design, is used when multiple in- terchangeable tests need to be constructed. An example in which multiple tests are needed is the evaluation of different candidates at different time frames, or the assignment of different tests to different students in order to avoid cheating. A basic multiple test construction method builds tests sequentially until the150 desired number of tests have been generated. Chen et al. (2012) proposes one of these sequential methods. The method first divides the item pool into groups (referred as cells) containing questions with similar characteristics. Then, a specified number of items from each group are randomly selected in order to generate different tests in each execution. Chen (2015) provides an improved method that incorporates a content-balancing element similar to those found on CAT test construction methods. Both works use a real 540 items pool to test their proposed method. Chang & Shiu (2012) proposes a method to construct multiple tests based on CLONALG, an algorithm based on the clonal selection principle in a biological160 immune system. The authors report results with a 4.000 items pool used to construct 5 parallel tests. Hwang et al. (2008) proposes a tabu search approach that generates multiple tests subject to time constraints and identical number of questions per test. The method was tested in simulated 50.000 items pools, which represents the largest sized item pool used in the literature. The previous references try to focus on constructing a predefined number of tests while minimizing deviations among the tests and/or maximizing the dis- crimination power of each test. Another possible objective aims at maximizing the utilization of the item pool, by constructing the maximum number of tests170 under some constraints on the repeated use of each item. Songmuang & Ueno (2011) consider a mixed problem in which both the number of constructed tests and quality of each individual test are considered. The problem is then solved using a Bee Algorithm. The computation is per- formed using parallelization and multiple tests are obtained for item pools with up to 20.000 items. Ishii et al. (2014) considers the maximization of the number of tests using a maximum clique formulation. Their work builds upon the study by Belov & Armstrong (2006), in which the problem is assimilated to a set packing problem and solved using a constraint programming approach. The proposal of Ishii180 et al. (2014) constructs a graph in which a maximal subset of tests that share none or few items (the clique) is sought. The method first constructs multiple tests with the required conditions like the discrimination power and number of items used (see also Belov, 2008). The tests define the vertices of the graph and 5 edges among vertices are included if the number of shared items is deemed small. Once the graph has been built, a classical method to ascertain the maximum clique is used. A different approach, which also links test assembly design to a classical combinatorial optimization problem was proposed in Brusco et al. (2013). In this case, the problem is formulated as a constrained Bin Packing Problem190 (BPP) and it is named the one-dimensional minimax bin-packing problem with bin size constraints (MINIMAX BSC). The authors consider that the item pool has been previously divided into groups of items with similar characteristics, as seen in van der Linden & Boekkooi-Timminga (1988), Chen et al. (2012) or Chen (2015), and then they consider the generation of multiple tests each containing an item from each group. As pointed out by van der Linden & Boekkooi-Timminga (1988) and Brusco et al. (2013), the division of items into groups has multiple advantages as it can be used to obtain tests with similar composition as well as to enforce constraints which may be difficult to specify. The objective of the problem is then to200 obtain tests with similar discrimination index on the ability level of interest. The authors solve the problem using a commercial software (CPLEX) and a Simulated Annealing approach on item pools with up to 6.000 items. In this paper, we extend the work of Brusco et al. (2013) by considering its complexity and by providing a new efficient solution method for the MIN- IMAX BSC. The computational experiment shows that we are able to solve instances with up to 60.000-item pools within less than a minute in a modern commodity computer, and the solutions obtained show very small optimality gaps when compared to a theoretical lower bound on the maximum discrimi- nation index. For smaller sized instances, the method provides solutions of the210 same quality within seconds. These results open the door to their use in on- line single test construction methods using the shadow test approach (He et al., 2014). The shadow test approach is a online single test construction method in which a complete test is constructed in each CAT decision step and the final item used is selected from the constructed test. Additionally, this work also verifies the applicability of the proposed method on different conditions to those proposed in Brusco et al. (2013). According to Brusco et al. (2013), items are grouped according to their discrimination indexes. We conduct an experiment in which items are randomly grouped (so we can consider that they have been grouped by subject topic or some other220 constraint). The results of the additional experiment show that the method is still capable of obtaining optimal or near-optimal solutions, highlighting the possible application of this method to other multiple test construction problems in which the number of tests is known in advance and the constraints can be adapted into the multiple initial groups method required by the MINIMAX BSC formulation. As the problem under study is a special case of the BPP, we also provide a brief bibliography on the methods available for packing problems. We refer the reader to the classifications of packing problems provided in Dyckhoff (1990); Wäscher et al. (2007) and the references included in these papers for a more230 6 extensive and comprehensive literature review. We will only highlight that mul- tiple approaches have been proposed for the resolution of the one-dimensional case, both exact (i.e., usually based on column generation Kallrath et al., 2014; Brandão & Pedroso, 2014) and heuristic, such as genetic algorithms (Quiroz- castellanos et al., 2015), variable neighborhood search (Fleszar & Hindi, 2002; M’Hallah et al., 2013) or hyperheuristics (López-Camacho et al., 2014) methods. Our proposal belongs to the second group and its more distinguishing fea- ture is the use of special neighborhood structures, rather than the common swap or exchange neighborhoods. The proposed neighborhood structures allow us to explore larger neighborhoods and to find the best move in these neighborhoods240 within very short running times. The neighborhood structures are then com- bined in a Variable Neighborhood Search (VNS) method, and allow us to obtain optimal or near-optimal solutions in very short running times. While the pro- posed method is not directly applicable to other Bin Packing problems, it offers an example on how to apply the resolution of special cases of the problem into the definition of effective heuristics. 2.2. Mathematical formulation, lower bounds and special cases The MINIMAX BSC considers the problem of sorting a number of questions into questionnaires of equal length (i.e., each is composed of the same number of questions). The questions are initially divided into different homogeneous250 sets, and a solution corresponds to the assignment of one question from each set to each questionnaire so that all questions are assigned, and some measure of fairness among the resulting questionnaires is optimized (e.g., minimizing the maximum difference between the total difficulty of the questions assigned to two different questionnaires). These constraints (e.g., multiple sets and assigning a question of each set to each questionnaire) are particularly common in applications pertaining to the splitting of a set of items to form parallel tests of equal length. For example, in the research of van der Linden & Boekkooi-Timminga (1988), a first stage divides the questions into homogenous sets. Then, a second stage tries to find a260 solution in which questionnaires may only use one question from each set. While the first and second stage problem could be solved as a single problem, the first stage allows for the inclusion of additional constraints, (e.g., the grouping of questions with comparable difficulty or similar content) and avoids the selection of highly unbalanced solutions (e.g., one test is only composed of easy and difficult questions, while another contains a combination of moderately difficult questions). The formulation proposed in Brusco et al. (2013) is described as follows: T disjoint sets (1 ≤ t ≤ T) of questions exist, and each set is composed of B questions (1 ≤ r ≤ B) with weights wrt (1 ≤ r ≤ B, 1 ≤ t ≤ T). These270 weights represent the discrimination index of the question (i.e., the percentage of recipients that are capable of correctly answering the question). The questions need to be grouped into B (1 ≤ b ≤ B) groups (i.e., questionnaires) in such a way that each questionnaire contains exactly one item from each one of the sets; and the total weight of the questions assigned to a questionnaire identifies the 7 Figure 1: Example of a MINIMAX BSC instance and one of its possible solutions. The instance has five sets and three questions per set. The questions must be divided into three questionnaires, which are pictured in the solution. Each question is named with a lowercase letter, and its weight wrt is indicated above (e.g., question c has a weight of 3). In the solution pictured, questionnaires are identified with uppercase letters, and the total weight of the questionnaires can be calculated by summing the weights of all the questions assigned to each questionnaire (e.g., questionnaire A, which contains questions a, e, h, l and o, has a total weight of 29). The value that needs to be optimized is the maximum total weight for any questionnaire. In the solution pictured, the objective value is 32, which corresponds to questionnaire B. difficulty of the questionnaire. The objective is then to obtain questionnaires with similar difficulty. An example of both an instance and a solution for this problem is detailed in Figure 1. One of the possible methods to achieve the previous objective and the one used in the MINIMAX BSC is to minimize the maximum difficulty of any ques-280 tionnaire; this is known as a minimax objective and it is similar to the ob- jective of some Bin Packing related problems, like the unrelated parallel ma- chine scheduling problem with maximum completion objective (Dell’ Amico & Martello, 1995). Equations (1) to (5) are a valid integer programming (IP) formulation for the problem, which uses binary decision variables xrbt to identify the assignment of different questions to the questionnaires. The variable is equal to 1 when item r from set t is assigned to group b, and 0 otherwise. Finally, a real-valued variable Z is used to identify the maximum weight among the questionnaires. [MIN]Z (1) subject to:290 Z ≥ B∑ r=1 T∑ t=1 wrt ·xrbt, 1 ≤ b ≤ B; (2) B∑ r=1 xrbt = 1, 1 ≤ b ≤ B, 1 ≤ t ≤ T ; (3) 8 B∑ b=1 xrbt = 1, 1 ≤ r ≤ B, 1 ≤ t ≤ T ; (4) xrbt ∈{0, 1}, 1 ≤ r ≤ B, 1 ≤ b ≤ B, 1 ≤ t ≤ T (5) Equation (1) corresponds to the objective function (i.e., the minimization of the weight of the questionnaire with the largest sum of weights). Z is calculated in the constraint set (2). Equation set (3) verifies that each questionnaire has one question from each set, while equation set (4) checks that each question of each set is assigned to one questionnaire. Finally, constraint set (5) defines the domain of the variables. Note that constraint set (4) corresponds to the traditional BPP constraint set, forcing each item to be assigned to one bin, while constraint set (3) corresponds to the set of additional constraints over the300 traditional BPP constraint set. While the mathematical formulation can be used to obtain the optimal solu- tion to small-sized problems and to obtain lower bounds on the optimal solution of larger-sized ones, alternative methods exist in some cases: 1. Let W denote the sum of all of the weights of the questions (W =∑ i≤b≤B,1≤t≤T wrt). Then, equation (6) provides a valid lower bound for Z. This lower bound is used in the computational experiments to verify the optimality of the solutions. ZLB = dW/Be. (6) 2. When T=2, the optimal solution can be obtained as follows: first, order the questions of one set in non-decreasing order of weights and the ques-310 tions of the other set in non-increasing order of weights; let (a1,a2, ...,aB) and (a′1,a ′ 2, ...,a ′ B) be such sequences. Then, build a solution so that the b-th questionnaire (1 ≤ b ≤ B) is composed by questions {ab,a′b}. Claim 1. The previous algorithm provides an optimal solution. Proof The claim is verified using exchange arguments. Let us consider a solution constructed with the previous method, and consider the case when the sum of weights from the first questionnaire equals the objective function. Then, to reduce the objective value, one of the questions in (a2, ...,aB) must be assigned together with question a′1; however, because the weight of a1 is smaller than or equal to the weights of the questions in (a2, ...,aB), the objective value320 would not improve. Hence, the current solution is optimal. Now suppose that the sum of weights from the second questionnaire equals the value of the objective function. Then, to reduce the objective value, a1 can only be assigned together with a′2, effectively reducing the sum of weights of the second questionnaire. However, the best possible assignment for a′1 among the remaining questions is thus a2 and wa′1 + wa2 ≥ wa2 + wa′2. Hence, the exchange would increase the objective value. The previous exchange argument can be extended to the remaining question- naires. Note that we must exchange the questions between a questionnaire and an earlier questionnaire, but this would not improve the global solution unless330 9 the newly exchanged question is exchanged with the question of an earlier ques- tionnaire. Eventually, the question from the first or the second questionnaire is to be considered, and the previous arguments state that the exchange is not going to improve the solution, thus proving this claim. � A computer implementation to solve the MINIMAX BSC with T = 2 re- quires the ordering of two sets with complexity of O(B · logB). This is a clear improvement over the B! permutations that would need to be explored if a brute force method were used. 2.3. Complexity of the problem In this section, we first reduce the Partition Problem to the MINIMAX BSC340 problem in which B = 2. Then, we show that the 3-Partition Problem reduces to a general MINIMAX BSC, ascertaining the theoretical difficulty of the problem. Finally, we identify the relationship between the MINIMAX BSC problem and the subset sum knapsack problem (Kellerer et al., 2004), which will be used in the proposed VNS as one of the neighborhood evaluation methods. Theorem 1. The MINIMAX BSC problem with B=2 is NP-Hard. Proof. We use a reduction for the Partition problem that is known to be NP-Complete (Garey & Johnson, 1979, p. 223). Problem PARTITION. Given a finite set A and a size s(a) ∈ Z+ for each a ∈ A, is there a subset A′ ⊂ A such that ∑ a∈A′ s(a) = ∑ a∈A′−A s(a)?350 To ease the explanation of the reduction procedure, we assume that A is an ordered set, and thus, the t-th element of A can be referred to as at. Reduction: Given an instance of the problem PARTITION, the correspond- ing MINIMAX BSC instance with B = 2 is constructed as follows: Let T be the cardinality of A (T = |A|), and let each set (1 ≤ t ≤ T) be composed by two items with weights wt1 = at and wt2 = 0. The optimum objective value for this instance is equal to ∑ a∈A s(a)/2 if and only if the answer to the original instance of the Problem PARTITION is “Yes” � Theorem 2. The general MINIMAX BSC problem is strongly NP-Hard.360 Proof. We use a reduction for the 3-Partition problem that is known to be strongly NP-Complete (Garey & Johnson, 1979, p. 96). Problem 3-PARTITION. Given a set A of 3 ·m elements, a bound U ∈ Z+, and a size s(a) ∈ Z+ for each a ∈ A such that U/4 < s(a) < U/2 and such that∑ a∈A s(a) = m · U, can A be partitioned into m disjoint sets A1,A2, ...,Am such that, for 1 ≤ i ≤ m, ∑ a∈Ai s(a) = U? Reduction: Given an instance of problem 3-PARTITION, the corresponding instance of MINIMAX BSC is constructed as follows: B = m and T = 3 · m. Then, let wt1 = at, wtb = 0 for 2 ≤ b ≤ B and 1 ≤ t ≤ T . The optimum objective value for this instance of the MINIMAX BSC is equal370 to U if and only if the answer to the original instance of Problem 3-PARTITION is “Yes” � The reduction of the MINIMAX BSC with B = 2 to the Partition Problem allows us to solve this special case using efficient methods found in the literature 10 for the subset sum knapsack problem. While the subset sum knapsack problem is NP-Hard (Garey & Johnson, 1979, p. 65), the problem is known to be easily solvable in many practical circumstances. The subset sum knapsack problem can be defined as follows: let there be n (1 ≤ i ≤ n) items, each with weight wi, and a maximum capacity W . The objective is to find the subset of items such that the total weight is maximized380 without surpassing the maximum capacity W . The mathematical formulation of the problem using a set of variables xi ∈ {0, 1}, (1 ≤ i ≤ n) corresponds to maximizing ∑ 1≤i≤n wi ·xi subject to ∑ 1≤i≤n wi ·xi ≤ W . In this case, the subset sum knapsack problem considers the optimal as- signment of questions to one of the questionnaires (i.e., implicitly assigning the remaining questions to the other questionnaire). Let wi be the additional dif- ficulty associated to selecting the most difficult question in the set i, which implies wi = |wi1 −wi2| for each i (1 ≤ i ≤ T). Let xi ∈{0, 1}, 1 ≤ i ≤ T be a binary variable that equals 1 if the most difficult question from set i is assigned to the questionnaire under construction and 0 if the easy question from set i is390 assigned instead. Finally, let W be b ∑ i≤i≤T |wi1 −wi2|/2c. Claim 2. The optimal solution to the previously defined subset sum knap- sack problem provides an optimal assignment of questions to questionnaires. Proof. To prove this claim, first note that each question must belong to one questionnaire; then, the unselected question of each set defines the second questionnaire. Let Z1 and Z2 be the total difficulty of each questionnaire, which can be obtained using (7) and (8): Z1 = T∑ i=1 xi · |wi1 −wi2| + T∑ i=1 MIN(wi1; wi2) (7) Z2 = T∑ i=1 (1 −xi) · |wi1 −wi2| + T∑ i=1 MIN(wi1; wi2) (8) Then, the ideal weight for the non-constant part of Z1 is b ∑ i∈B |wi1 − wi2|/2c, which would lead to a non-constant part of Z2 equal to d ∑ i∈B |wi1 −400 wi2|/2e. Hence, if Z1 decreases, Z2 increases. As the subset sum knapsack problem maximizes Z1, and the maximum possible value for Z1 is b ∑ i∈B |wi1− wi2|/2c, the optimal subset sum problem minimizes the non-constant part of (8), effectively minimizing the MINIMAX BSC problem.� Note that during the previous proof, the definition of the problem imposes w.l.o.g. that Z1 ≤ Z2. 3. Description of the algorithm The Variable Neighborhood Search (VNS) approach is a metaheuristic used as a framework to develop heuristics for specific problems. The primary con- cepts of the VNS approach are based on the following observations (Hansen410 et al., 2010): (1) local optimality with respect to a neighborhood does not im- ply local optimality in other neighborhood structures; (2) an optimum is a local 11 optimum for any neighborhood structure; and (3) usually a local minimum for one neighborhood is similar to those in other neighborhoods. Based on these observations, VNS uses multiple local searches, each with a different neighborhood structure. The algorithm starts from an initial solution and applies different local searches until a locally optimal solution to each of these neighborhoods is found. This technique is usually known as a variable neighborhood descent (VND). While VND improves upon the use of traditional local search methods, it420 may still become trapped in a local (i.e., non-global) optimum. To escape from these optima and to diversify the search, an additional mechanism must be incorporated. This mechanism is known as a shake operator and introduces a random movement from the current solution to a neighbor solution in one of the neighborhood structures. The resulting algorithm is known as a general VNS. This paper proposes a general VNS for the MINIMAX BSC; the remaining subsections describe the primary components of the implementation. Subsection 3.1 describes a greedy constructive procedure, which is used to obtain an initial solution and also during the “shaking” procedure. Subsection 3.2 describes the neighborhood structures used in this implementation and the methods used430 to reach the best neighbor for each neighborhood structure. Subsection 3.3 documents the ”shaking” procedure, and Subsection 3.4 details the conditions used to terminate the search. 3.1. Greedy Initialization heuristic While the definition of a good initial solution is not an important aspect of an effective VNS heuristic, the quality of the initial solution helps reduce the time required to reach the first local optimum. Therefore, we propose the use of a fast O(T ·B · logB) heuristic, with an absolute performance guarantee equal to R = max1≤t≤T rt (rt = max1≤r≤Bwrt − min1≤r≤Bwrt). Algorithm 1 depicts the algorithm.440 Algorithm 1. Greedy heuristic. for each set 1 ≤ b ≤ B do Zb=0 end for for each set 1 ≤ t ≤ T do Order the questions of the set into a non-decreasing order of the weights. Order the questionnaires into a non-increasing order of accumulated weights. for each t, 1 ≤ t ≤ T do Assign the t-th question to the t-th questionnaire. end for end for An example depicting the process of Algorithm 1 can be found in Figure 2. Note that the method is based on considering the questionnaires of a partial solution as single questions and finding the best assignment combining these 12 Figure 2: Illustration of the constructive heuristic. The instance is illustrated in the first row of the figure. First, set 1 is assigned to the questionnaires at random, as shown in the second row of the figure. Then, set 2 is assigned following the heuristic logic described as follows: the question with the lowest weight is assigned to the partially filled questionnaire with the highest weight; the question with the second lowest weight is assigned to the questionnaire with the second highest weight; etc. until all of the questions are assigned, as shown in the third row of the figure. Afterwards, the process is repeated for the following sets until all questions from all sets have been assigned. The last row in the figure shows the final solution. 13 questions with an additional set, which is equivalent to finding the optimal solution of a MINIMAX BSC problem when T = 2 (see Subsection 2.2). To analyze the computational complexity of the algorithm, first note that the computationally expensive part of the algorithm is ordering the questions and questionnaires. This step is executed T times, and each ordering has O(B · logB) complexity. Hence, the final computational complexity of the algorithm is O(T ·B · logB).450 Claim 3. Algorithm 1 provides a solution with an absolute performance guarantee of R. Proof. First, we show that the maximum difference between then accumu- lated weights of any two questionnaires is equal to R. To simplify the explana- tion, let d denote the maximum difference in the accumulated weights between two questionnaires. Initially, d is set to 0. Accordingly, after the first assignment of questions to questionnaires, R ≥ d holds. During subsequent assignments, d is bounded by max{d,rt} with rt equal to the range of the assigned set (if rt ≤ d, then d−rt is bounded by d, and if rt ≥ d, then rt − d is bounded by rt). Because both460 d ≤ R and rt ≤ R hold, the difference between any two groups will always be smaller than R. Second, note that for any valid solution, the minimum total weight of any questionnaire (min1≤b≤B{Zb}) is a lower bound on the optimal solution. Be- cause R is a bound on the difference between any two questionnaires obtained using Algorithm 1, the difference between the solution provided by Algorithm 1 and a lower bound for the problem is also bounded by R. This construct proves the claim that Algorithm 1 provides a solution that has an absolute performance guarantee equal to R.� The solution offered by this constructive method depends on the order in470 which the sets are considered. A preliminary computational experiment showed that the experimental efficiency of Algorithm 1 is improved when the sets are assigned in non-increasing values of rt order. Accordingly, the initial solution of the VNS uses this ordering. 3.2. Neighborhood structures Neighborhood structures are the most important part of the definition of a successful VNS implementation. In this work, we propose three different neighborhood structures that try to improve the incumbent solution in different ways. The search in each of the neighborhoods is conducted by solving special cases of the problem (i.e., a MINIMAX BSC with either T = 2 or B = 2) until480 the solution is locally optimal for all of the explored neighborhoods. We now proceed to define each of the three neighborhoods: Neighborhood 1. This neighborhood considers the reassignment of the questions of one set to the questionnaires. This neighborhood corresponds to the exploration of all possible solutions that can be obtained by removing all the questions of a set from the incumbent solution and their reassignment to the questionnaires. For all of the sets, the 14 following three steps are performed: (1) remove the questions that correspond to that set from the questionnaires; (2) explore the reassignment of the extracted questions to the partially filled questionnaires; and (3) assign the questions to490 the questionnaires according to the least expensive solution obtained in step (2). Step (2) can be time consuming, as all possible permutations of the assign- ments of questions to questionnaires need to be considered in the worst case. Note that the weight of the questionnaires in the partial solution can be con- sidered as a unified contribution to the total weight of the final questionnaires. Consequently, each partially filled questionnaire can be considered as a single question from a different set, and the optimal reassignment of the extracted questions can be obtained by solving the MINIMAX BSC with T = 2 (see Subsection 2.2). This neighborhood is implemented in a first descent fashion, and as such, the500 ordering of sets may modify the final solution. Therefore, to reduce the effect of a specific ordering, whenever neighborhood 1 is explored, a random order of sets is generated and the sets are considered in accordance with the proposed order. Algorithm 2 depicts the final algorithm. Algorithm 2. Neighborhood structure 1 search. Define a random order of the T sets repeat until no further improvement is possible for each t (1 ≤ t ≤ T) do Remove the questions from ordered set t from the questionnaires Solve the MINIMAX BSC problem with T = 2. if the new solution improves upon the incumbent solution then Update the incumbent solution end if end for end repeat An example of this neighborhood is depicted in Figure 3. Note that the neighborhood considered in Brusco et al. (2013) can be seen as a special case of this neighborhood. Their neighborhood studies the reassign- ment of two questions from the same set to their respective questionnaires, which is a limited version of neighborhood 1. Therefore, the limited neighborhood is not used in the final VNS implementation.510 Neighborhood 2. This neighborhood considers the division of the solution into two partial solutions that are then optimally combined to possibly form an improved solution. Let T ′ and T ′′ be a partition of the sets in T . For any given solution S, let S′ and S′′ be the partial assignment of the questions to questionnaires defined in S for the questions of subsets T ′ and T ′′, respectively. Then, consider each questionnaire in S′ and S′′ as a single question (i.e., all questions assigned to a partial questionnaire are to be assigned together when the partial questionnaires are combined). Based on the previous definition, neighborhood 2 corresponds 15 Figure 3: Example of the application of the neighborhood 1. The figure depicts an initial solution in the first row (taken from the example in Figure 1), from which the first neigh- bourhood is created. First, all of the questions that correspond to set 3 are removed from the solution (second row in the figure). Underneath the partial solution, the total weight of each questionnaire is depicted in a box (for example, the total weight for the partial questionnaire A is 22). Then, the questions corresponding to set 3 are optimally reassigned (third row) providing an improved solution. 16 to the exploration of all possible partitions of S into S′ and S′′, as well as all of520 their respective assignments combining S′ and S′′ to obtain a new solution. The previous neighborhood definition has two issues that must be addressed to obtain an efficient implementation: 1. The number of partitions in the previous definition is not bounded by a polynomial function to the number of sets in the instance, and thus only a limited number of partitions can be considered. The implementation uses the following technique; each time neighborhood 2 is explored, we define a random order of the sets and we explore the t−1 (1 ≤ t < T) partitions in which T ′ is composed of the t first sets of questions according to the random ordering, and T ′′ is composed of the remaining sets of questions.530 2. Once T ′ and T ′′ have been defined, the optimal combination of the partial questionnaires in S′ and S′′ can be solved as a MINIMAX BSC with T = 2, which is solvable in polynomial time (see Subsection 2.2). Algorithm 3 shows the final algorithm. Algorithm 3. Neighborhood structure 2 search. Define a random order of the T sets repeat until no further improvement is possible for each t (1 ≤ t ≤ T) do Obtain the MINIMAX BSC problem with T = 2 with partitions T ′ and T ′′ Solve the MINIMAX BSC problem with T = 2 if the new solution improves upon the incumbent solution then Update the incumbent solution end if end for end repeat An example of an improving move using neighborhood 2 is depicted in Figure 4. Note that neighborhood 1 is a special case of neighborhood 2 if multiple orderings are considered. The previous statement holds if at least one of the considered orderings contains each set as its first or last studied set. Then, the exploration of neighborhood 2 using all of these orderings subsume neighborhood540 1 as the algorithm considers the reassignment of each individual set of questions in addition to the joint reassignment of sets of questions. Neighborhood 3. This neighborhood explores the optimal reassignment of questions to two questionnaires. While neighborhood 1 considers changes of different sets of questions and neighborhood 2 considers changes of different groups of sets of questions, neigh- borhood 3 studies the effect of changing the questions assigned to two ques- tionnaires. Neighborhood 3 is thus defined as the exploration of all possible solutions in which the reassignment of questions to exactly two questionnaires may vary.550 17 Figure 4: Example of the application of the neighborhood 2. The initial solution (first row, from the example in Figure 1) is divided by selecting two groups of sets (one is composed of sets 3 and 5 and is highlighted by boxing their respective questions; the other group corresponds to the remaining sets). The last row depicts the solution obtained after optimally solving the related MINIMAX BSC problem with T = 2. 18 The previous definition naturally leads to multiple possible partial solutions, each requiring the exploration of a combinatorial problem (i.e., all possible as- signments of questions to two questionnaires must be investigated). Conse- quently, an efficient exploration of the neighborhood must be based on some observations that may lead to an effective search for improving solutions. The following observations are used in the proposed neighborhood search method: 1. To improve the solution, some questions assigned to the questionnaire with the largest sum of weights must be exchanged. 2. The optimal assignment of two questionnaires can be found by solving a subset sum knapsack problem (see Section 2.3).560 Combining both observations, we can establish that one of the questionnaires considered in the subset sum knapsack problem must be one of the question- naires in which constraint (2) in the mathematical model is fulfilled in equality (i.e., the questionnaire with the largest weight, or one of them if several exist). As in the previous neighborhoods, it is important to define the order in which the questionnaires are inspected. While the objective in the previous cases was to avoid a bias towards some sets, we opt for an ordering in this case that maximizes the probability of finding an improved solution; specifically, the questionnaires are ordered in non-decreasing order of the sum of weights. Then, for each questionnaire and in the order defined above, the algorithm solves a570 MINIMAX BSC problem that combines the said and the last questionnaire in the list. Algorithm 4 illustrates this procedure. Algorithm 4. Neighborhood structure 3 search. Order the questionnaires in non-decreasing order to their respective weights for each r (1 ≤ r ≤ B) do Solve the MINIMAX BSC problem using the r-th and B-th questionnaires if the new solution improves upon the incumbent solution then Update the incumbent solution end if end for An example of neighborhood 3 is depicted in Figure 5. 3.3. Shaking procedure The shaking procedure randomly chooses exchanges defined by one of the considered neighborhoods. In this case, we opt for the application of a neigh- borhood structure similar to neighborhood 1 with a variable parameter k that controls the distance from the current partial solution. For any given k (2 ≤ k < T), the proposed procedure removes the questions associated to k distinct sets from the incumbent solution. Once these questions580 are removed, the questions of each set are introduced in the solution using the selection rule specified by neighborhood structure 1, which considers the partial 19 Figure 5: Example of the application of the neighborhood 3. The first row depicts an initial solution (from the example in Figure 1). A new instance of the problem is created by consid- ering the questions assigned to questionnaires A and B; questionnaire B is selected because it is the questionnaire with the highest weight. The new instance is depicted in the second row and corresponds to an instance with 5 sets and 2 questions per set. The third row depicts the optimal reassignment of questions to questionnaires A and B that, with questionnaire C (which remains unchanged) provide an improved solution. 20 solution as a single set of questions and finds the optimal assignment of a new set of questions. Note that the consecutive introduction of multiple sets of questions can also be achieved using algorithm 1; thus, the shaking procedure can be described as the elimination of k sets from the solution and their reintroduction into the current solution using the constructive greedy heuristic; however, a discrepancy exists because the sets are reintroduced in random order rather than in the order proposed in Subsection 3.1. The shaking procedure highlights the relationship between the constructive procedure and the two first neighborhood590 structures, because all three structures are based on the optimal solution to a MINIMAX BSC instance with T = 2. The values of parameter k are controlled as follows: the initial value is set to k = 2 because k = 1 in the reintroduction scheme would return the solution to the same local optimum and thus it does not modify the solution. Whenever a local optimum for each of the three neighborhoods is reached, the shaking procedure is restarted, and k is increased by one. Whenever k = T −1, the shaking procedure generates a random solution which can be considered as restarting the search; k is also restarted in this circumstance (i.e., k is set equal to 2).600 3.4. Termination condition and outline of the complete algorithm The termination condition is linked to a run time limit or the verification of an optimal solution (i.e., when the incumbent solution is equal to the lower bound defined in Subsection 2.2). The complete algorithm is depicted in Algo- rithm 5. Algorithm 5. VNS method. Find a heuristic solution using Algorithm 1, and initialize k. Init: repeat until termination condition is met Search neighborhood structure 1 until a locally optimal solution is found. Search neighborhood structure 2 until a locally optimal solution is found. If the solution was improved during the search, goto Init. Search neighborhood structure 3 until a locally optimal solution is found. If the solution was improved during the search, goto Init. end repeat Note that the application of the neighborhoods in the predefined order was decided based on run time considerations: the less-time-consuming algorithm is the first to be applied, while the most-time-consuming is the last to be applied. 4. Computational Experiments The previous algorithm was implemented in C and compiled using gcc v.4.9.0610 and run on computer with a 2.9 GHz Intel Core i5 processor and 8 GB of 21 RAM running the Mac OS X 10.9 operating system. To solve the subset sum knapsack problems, the expknap implementation (Pisinger, 1995) available at http://www.diku.dk/-pisinger/codes.html was used. In addition to the algorithms described in this paper, we also implemented the SA algorithm proposed in Brusco et al. (2013). We did not report results for the IP formulation because Brusco et al. (2013) have already verified that this formulation is inefficient for large size instances. The parameters of the SA were identical to those proposed in Brusco et al. (2013). Furthermore, if the SA terminates before the specified running time620 limit, the algorithm is restarted with a different randomly generated initial solution. This VNS implementation is parameter-free, and thus no parameter tuning was required. Both algorithms were run for at most 600 s, which was measured as wall time. The two algorithms were initially tested on a set of instances constructed by Brusco et al. (2013). The characteristics of both sets correspond to those found in previously simulated data found in the literature (Veldkamp, 2005; van der Linder, 2005; van der Linden & Boekkooi-Timminga, 1988). Instances with varying number of questions (e.g., 300, 600, 3000 and 6000 questions) were generated. These questions were divided into different num-630 ber of questionnaires (e.g., B = 2, 3, 4, 5, 10, 20, 30, 60, 120, 300, 600 and 1200 questionnaires). For each number of questions, the abovementioned num- ber of questionnaires is only used if the number of questions is at least five times bigger than the number of questionnaires (i.e., each questionnaire will be formed by five or more questions). Finally, the weights of the questions in each instance were generated using the following method, which was proposed in Br- usco et al. (2013), and based on van der Linden & Boekkooi-Timminga (1988). The method uses formula (9) to compute the weight of each question based on two parameters πrt and ρrt, which represent the difficulty (i.e., percentage of examinees who obtained the correct answer) and the discrimination effect of the640 question, respectively. wrt = .5 ·πrt + .5 ·πrt · (1 −πrt) ·ρrt (9) To generate sets of instances with similar weights, the method initially ob- tains the difficulty π1t and discrimination effect ρ1t from the first question of each set t. These values are drawn from a uniform distribution with interval [.3, .8] for the difficulty effect (π) and from the interval [.25, .60] for the discrimination effect (ρ). For the remaining questions πrt and ρrt (2 ≤ r ≤ B), the values are randomly drawn from uniform distributions with interval [π1t − .1,π1t + .1] and [ρ1t − .1,ρ1t + .1]. These weights are then used in combination with (9) to generate the final weights. Ten instances per combination of number of questions, questionnaires and650 distribution of weights were constructed to create a total of 400 instances. Note that while the weights were obtained from uniform distributions and thus are real values, the method works with integer-valued weights to avoid rounding issues. To achieve this objective, we decided to multiply the obtained values by 22 a constant (106) and round to the nearest integer. Consequently, this algorithm is working using real values with six digits of precision. Table 1 provides a summary of the experimental results. For each number of questions and questionnaires, the table provides a comparison between the heuristic method depicted in Subsection 3.1, the VNS method proposed in this paper and the SA algorithm proposed by Brusco et al. (2013). For each al-660 gorithm, we provide the average absolute gap between the lower bound (LB), calculated using equation (6), and the upper bound (UB) provided by the al- gorithm (that is, UB −LB). For the metaheuristic approaches, the number of optimal solutions (out of 10 instances) and the running time required to reach the best solution (measured in wall time) are also provided. An analysis of the table shows that the gap between the solutions provided by both metaheuristics and the lower bound are small in all cases. Furthermore, both methods are able to dramatically improve the initial solutions provided by the heuristic with a performance guarantee. Furthermore, the VNS was able to obtain the optimal solution for most670 of the instances, including all of the instances in which the ratio between the questions and questionnaires was higher than 15 within negligible running times (i.e., within one second). Hence, we can conclude that those instances in which the ratio of questions per questionnaire is larger than or equal to 15 are to be considered easy for the proposed method. The running time increased for the instances in which the algorithm was not able to verify optimality, but were still reduced. Note that table 1 reports the running time to reach the best solution for all methods rather than the total computing time required. According to the termination condition, if the optimal solution is not verified, both algorithms680 continue until the time limit is reached. Based on the results provided above, we test the algorithm with larger in- stances (e.g., with T=30.000 and 60.000, and different values of B ranging from 2 to 12.000). Similar results to those reported in table 1 were obtained, and the VNS method was able to optimally solve all instances with a ratio of questions to questionnaires larger than 10 within a maximum of 5 seconds of computing time. Therefore, we limit this report to the results for those groups of instances in which the VNS fails to verify the optimal solution for every instance with the given ratio of questions per questionnaire (“challenging” instances). Table 2 reports these results.690 The results from table 2 show that the algorithm is still effective for larger- sized instances. Furthermore, the number of verified optimal solutions increases due to the existence of more alternatives to assign the questions. Consequently, we conclude that for the VNS method, increasing the number of questions does not affect the quality of the solutions, while the running times increase in ac- cordance with the requirement of solving larger problems, specifically larger knapsack instances. The results provided in tables 1 and 2 indicate that the proposed VNS offers high quality results, but do not identify which components of the algorithm are responsible for the good performance. Therefore, an additional experiment700 23 Table 1: Results of the proposed algorithms. For each combination of characteristics, we provide the number of optimal solutions verified by each of the metaheuristic algorithms, the average absolute gap (i.e., the average difference between the solution and the lower bound, provided by equation 6) for each algorithm and the average time required to reach the best solution by the metaheuristic methods; for the SA algorithm, we report the time required for replication that provided the solution under consideration. Questions B Heuristic VNS VNS VNS SA SA SA Gap Opt Gap t Opt Gap t 300 2 152.8 10 0 0.0 10 0 0.0 300 3 2092.4 10 0 0.0 5 0.7 0.0 300 4 3990 10 0 0.0 0 11.4 0.1 300 5 5184.9 10 0 0.0 0 25.6 0.1 300 10 10424.8 10 0 0.0 0 107.6 0.1 300 20 14789.5 9 0.1 0.1 0 164.4 0.0 300 30 14425 0 9 0.1 0 237.3 0.0 300 60 26446.1 0 201.7 0.0 0 368.7 0.0 600 2 23.9 10 0 0.0 10 0 0.0 600 3 1253.9 10 0 0.0 4 0.6 0.2 600 4 2598.8 10 0 0.0 0 4.4 0.2 600 5 4292.7 10 0 0.0 0 12.9 0.2 600 10 12199.9 10 0 0.0 0 50.6 0.1 600 20 14419.5 10 0 0.0 0 85.8 0.1 600 30 17964.2 10 0 0.0 0 116.6 0.1 600 60 13006.9 0 6.6 0.2 0 183.4 0.1 600 120 28216.4 0 133.9 0.1 0 290.3 0.0 3000 2 14.3 10 0 0.0 10 0 0.0 3000 3 453 10 0 0.0 10 0 1.4 3000 4 2093.4 10 0 0.0 5 0.6 1.2 3000 5 2614.6 10 0 0.0 0 2.1 1.0 3000 10 9293.1 10 0 0.0 0 10.2 0.6 3000 20 13998 10 0 0.0 0 19.5 0.4 3000 30 15172.7 10 0 0.0 0 23.9 0.3 3000 60 13931.3 10 0 0.0 0 35.5 0.2 3000 120 17909.3 10 0 0.1 0 58.5 0.2 3000 300 5742.3 0 2.5 1.1 0 120 0.1 3000 600 30084.8 0 42.2 0.6 0 174.9 0.2 6000 2 3.7 10 0 0.0 10 0 0.0 6000 3 252.6 10 0 0.0 10 0 2.7 6000 4 1297.2 10 0 0.0 8 0.2 2.5 6000 5 2510.4 10 0 0.0 4 0.8 2.1 6000 10 8054.3 10 0 0.0 0 5.2 1.3 6000 20 10717.1 10 0 0.0 0 9.1 0.8 6000 30 14209.7 10 0 0.0 0 11.1 0.6 6000 60 13686.2 10 0 0.0 0 17.3 0.5 6000 120 11105.7 10 0 0.0 0 29.8 0.3 6000 300 7902.1 10 0 0.1 0 59.7 0.3 6000 600 5324.9 0 1.6 0.9 0 90.3 0.4 6000 1200 29043.9 0 25.7 1.1 0 290.9 0.4 24 Table 2: Results for larger instances. Only those combinations of the number of questions and questionnaires that are deemed to be ”challenging” are reported. For each combination of characteristics, we provide the same results that in table 1. Questions B Heuristic VNS VNS VNS SA SA SA Gap Opt Gap t Opt Gap t 30000 3000 4208.8 6 4 60.8 0 351.1 2.6 30000 6000 24234.8 0 11.8 10.3 0 2155 2.4 60000 6000 12963.3 3 0.7 24.6 0 1055.3 4.8 60000 12000 38837.5 0 4 48.3 0 2757 5.7 geared towards evaluating the contribution of each of the specific components was conducted. The experiment considers different configurations of the VNS algorithm in which some of the neighborhoods or the shaking procedure are not used. While specific neighborhood can be removed from the algorithm without modifying its global structure, it is necessary to include some kind of restarting procedure to escape from local optimality, continue the search and effectively use the allotted computing time. Consequently, when the shaking procedure is not used, it is substituted by the greedy initialization procedure proposed in Subsection 3.1 in which the order of the sets is randomly determined before executing algorithm710 1. A total of 14 different configurations of the VNS algorithm were tested (all possible combinations of use/not use for each of the three neighborhoods and the shaking procedure, with the exception of the two combinations that do not use any of the three neighborhoods) on the large and “challenging instances (the 40 instances reported in table 2). Each configuration was run for at most 600 seconds of wall time, and three values were recorded: (1) the running time required to reach the best solution; (2) the absolute gap between the solution found by the configuration and the best known solution for the instance; and (3) the absolute gap between the solution found by the configuration and the720 theoretical bin packing lower bound. Furthermore, and in order to verify if the restarting procedure was efficient, we also run the procedure as a simple descent method with all of the neighborhood structures and recorded the same information. Table 3 summarizes the results obtained for the experiment. The examina- tion of table 3 highlights the effectiveness of neighborhood 3 (gaps are much smaller if neighborhood 3 is used). Note that the gap reduction is accompanied by larger computational times. The shaking procedure, as well as neighborhood 1 and neighborhood 2 do not affect the solution quality but they do seem to reduce the running time required to reach the best solution. Also note that the730 gap difference among the VNS configurations using neighborhood 3 is minimal. If we compare the single descent method with the VNS algorithm, the re- sults from table 3 show that the VNS is capable of providing improvements on the solution provided by the simple descent method, at the cost of larger computational times. Please note that the gap reduction is small as the solu- 25 Table 3: Results for different configurations of the proposed algorithms. For each combination of use/not use of each of three neighborhoods (columns N1, N2 and N3) and shaking procedure (column “Shake”), the average absolute gap between the solution found by the configuration and the theoretical bin packing lower bound (column Gapbound); the average absolute gap between the solution found by the configuration and the best known solution for the instance (column Gapbest); and the average running time required to reach the best solution (column t) are reported. Additionally, the results of the simple descent algorithm (no restart used) are also reported (row with “Shake” entry equal to “Descent”). N1 N2 N3 Shake Gapbound Gapbest t No No Yes No 4.4 0.5 36.3 No No Yes Yes 4.4 0.6 39.5 No Yes No No 17.8 13.9 26.5 No Yes No Yes 26.6 22.7 16.1 No Yes Yes No 4.2 0.3 31.7 No Yes Yes Yes 4.3 0.4 34.0 Yes No No No 25.3 21.5 6.0 Yes No No Yes 16.7 12.8 2.8 Yes No Yes No 4.2 0.3 32.5 Yes No Yes Yes 4.3 0.4 22.7 Yes Yes No No 10.4 6.5 22.6 Yes Yes No Yes 12.3 8.4 21.1 Yes Yes Yes No 4.2 0.3 34.9 Yes Yes Yes Yes 4.0 0.1 21.6 Yes Yes Yes Decent 6.6 2.7 10.0 26 tions provided by the descent heuristic are already very good and show close to nonexistent optimality gaps. In addition to the previous qualitative analysis, we conducted an ANOVA test on two response variables (Gapbest and running time) and a MANOVA test using both response variables (see Christensen, 2011, for a general reference on740 ANOVA and MANOVA). The tests reported differences among VNS configura- tions, but an analysis of the residuals showed severe violations of the normality assumptions and heteroscedasticity issues. As a consequence, we chose to reex- amine the results using a non-parametric MANOVA alternative. The permutational MANOVA using distance matrices proposed in Anderson (2001) was selected. The method does not require traditional assumptions; it is based on a geometric interpretation of the analysis of variance; and it uses permutations in order to obtain p-values. Furthermore, the code is readily available on the R statistical package (Oksanen et al., 2015). The test was run using four factors (the three neighborhood searches, and the shaking procedure)750 and a blocking factor (the instances) that tries to block the variability caused by specific characteristics of each instance. The results show that neighborhood structures 1 and 3 are significant with a p-value < 0.001, and the shaking procedure is significant with a p-value < 0.01. Neighborhood structure 2 is not significant, leading us to conclude that it could be removed from the algorithms without modifying its results (both in terms of running time and solution quality). If we combine the p-values with the analysis of table 3, we conclude that the different components of the proposed VNS method are able to improve the solution obtained by the final algorithm (neighborhood structure 3); to reduce760 the required running time to reach the aforementioned solution(neighborhood structure 1 and shaking procedure); or do not significatively modify running time or solution quality. A final experiment was performed to evaluate the behavior of the proposed method if the sets were based on some other constraints (e.g., length, time re- quired to answer the question) and not on similar objective function weights. A new instance set in which the weights of the questions are uniformly distributed is considered, see for example Nguyen & Fong (2013). For these instances, each weight is randomly selected form a uniform distribution on the interval [.1, .9] maintaining the remaining characteristics from the instances.770 The results, see table 4, show that the performance of the VNS method is significantly better than its SA counterpart. Hence the VNS depends less on a grouping of questions by similar weights. The relation between the number of questions and questionnaires to ascertain the difficulty of the instance still holds. However, the instances are harder to solve for both metaheuristics, and larger absolute gaps appear for those instances in which optimality is not verified (even if they are still small). Note that the heuristic constructive method is clearly affected by changes in structure and generates significantly worse initial solutions. As similar results in terms of optimally solved instances are obtained, we conclude that the distribution of weights among the questions in each set780 does not greatly affect the performance of the VNS method. 27 Table 4: Results for the instances generated according to a uniform distribution as in Nguyen & Fong (2013). For each combination of characteristics, we provide the number of optimal solutions verified by each algorithm; the average absolute gap, which is measured as the average difference between the solution and the lower bound using equation 6; and the average time required to reach the best solution; for the SA algorithm, we report the time required for replication that provided the solution under consideration. Questions B Heuristic VNS VNS VNS SA SA SA Gap Opt Gap t Opt Gap t 300 2 761.3 10 0 0.0 10 0 0.0 300 3 11312.1 10 0 0.0 0 17.5 0.0 300 4 33248.2 10 0 0.0 0 94.4 0.1 300 5 52034.2 10 0 0.0 0 263.6 0.1 300 10 121717.5 10 0 0.0 0 803.7 0.0 300 20 151533.7 0 2.7 0.1 0 1303.5 0.0 300 30 152080.8 0 77.6 0.1 0 1863.4 0.0 300 60 276051.8 0 1715.4 0.0 0 2853.1 0.0 600 2 227.1 10 0 0.0 10 0 0.0 600 3 7769.4 10 0 0.0 0 6.8 0.2 600 4 24895.5 10 0 0.0 0 42.7 0.2 600 5 41638.3 10 0 0.0 0 117.8 0.2 600 10 133000 10 0 0.1 0 355.5 0.1 600 20 151155.8 10 0 0.1 0 710.3 0.1 600 30 198745.9 10 0 3.2 0 873.9 0.1 600 60 137782.5 0 55.6 0.4 0 1441.3 0.0 600 120 318178 0 1090.6 0.1 0 2331.8 0.0 3000 2 156.3 10 0 0.0 10 0 0.0 3000 3 3727 10 0 0.0 3 0.7 1.6 3000 4 15952.3 10 0 0.0 0 8.7 1.2 3000 5 25568.9 10 0 0.0 0 23.9 1 3000 10 92353.9 10 0 0.0 0 77.1 0.6 3000 20 138538.2 10 0 0.1 0 141.1 0.4 3000 30 162414.9 10 0 0.2 0 182.9 0.3 3000 60 179236.1 10 0 0.3 0 287 0.2 3000 120 217096.8 10 0 0.5 0 442.7 0.2 3000 300 76249 0 20 3.8 0 987.5 0.1 3000 600 365729 0 358.5 0.2 0 1325.4 0.2 6000 2 29.5 10 0 0.0 10 0 0.1 6000 3 2497.9 10 0 0.0 6 0.4 3.6 6000 4 10685.3 10 0 0.0 0 5.2 2.6 6000 5 31974.9 10 0 0.0 0 11.4 2.1 6000 10 72561.1 10 0 0.1 0 40.2 1.3 6000 20 125779.9 10 0 0.1 0 70.2 0.8 6000 30 175586.3 10 0 0.1 0 91.7 0.6 6000 60 192118.3 10 0 0.3 0 134.7 0.4 6000 120 177719.9 10 0 0.5 0 229.5 0.3 6000 300 108666 10 0 1 0 482.5 0.3 6000 600 53802.5 0 11.4 7.9 0 667.9 0.4 6000 1200 366379.5 0 209.3 0.8 0 2186 0.5 28 5. Conclusions In this paper, we have investigated a Bin Packing-based formulation of the test assembly design problem known as the one-dimensional minimax bin- packing problem with bin size constraints (MINMAX BSC). We first show that the problem in NP-Hard when the number of questionnaires is 2 and strongly NP-Hard if more questionnaires are considered. Second, we develop a con- structive heuristic with an absolute performance guarantee, and we derive three neighborhood structures for the problem that we can explore by optimally solv- ing special cases of the problem. Third, we combine these procedures into a790 variable neighborhood search in order to efficiently search the solution space. Finally, the quality of the proposed procedures is tested in computational exper- iments on instances with up to 60.000 item pools, up to 12.000 questionnaires, and different discrimination effect structures (see Brusco et al., 2013; Nguyen & Fong, 2013). The proposed method shows significant improvements over previous algorithms proposed in the literature (Brusco et al., 2013), and the optimality of the solutions provided for instances with a 60.000 items pool and up to 3.000 tests has been verified. Our complexity results follow the line of previous works (like Belov & Arm- strong, 2006; Nguyen & Fong, 2013; Ishii et al., 2014) in which other formula-800 tions of the test assembly design problem were identified as classical strongly NP-hard combinatorial optimization problems. Our approach differs when we consider the resolution considerations, as Belov & Armstrong (2006); Nguyen & Fong (2013); Ishii et al. (2014) apply classical algorithms of the identified optimization problem to the resolution of the test design problem, while our Bin Packing-based formulation significantly differs from the classical problem, thus forcing us to develop specially tailored procedures for its resolution. The proposed algorithm corresponds to a neighborhood exploration heuris- tic. If compared to previous neighborhood-based approaches for test assembly design (Hwang et al., 2008; Brusco et al., 2013), which are based on traditional810 swap or exchange neighborhoods, our neighborhoods are based on the optimal exploration of special structures. These neighborhoods are derived from theo- retical results for the MINMAX BSC, and successfully use the structure of the problem to reach high quality results within very reduced running times. The quality of the solutions provided by the method is demonstrated by our experimental results with instances larger than those previously considered in the literature (see Hwang et al., 2008; Nguyen & Fong, 2013). Please note that the results are not directly comparable as there are strong differences among different formulations for the test assembly design problem. Nevertheless, we can make comparisons with a Bin Packing lower bound, which shows the high820 quality of the solutions found. The proposed method routinely finds the optimal solution when the ratio between items per questionnaire is above 10, and obtains very low gaps if the aforementioned ratio is 10 or below. These results apply under varying conditions, such as when the items are grouped according to similar weights, as in Brusco et al. (2013), or when the items are randomly grouped. Results for randomly grouped items consider cases in which additional 29 constraints are to be enforced (He et al., 2014; Ishii et al., 2014; Nguyen & Fong, 2013), or cases in which a cell approach (Chen et al., 2012; Chen, 2015) is used to maintain similar characteristics not only among questionnaires, but also among the item composition of each questionnaire.830 In addition to the quality of the results, we would like to highlight the reduced running times required to solve the instances, which are below the 60 seconds mark even for the largest item pools. These results lead us to conjecture that the proposed procedure could also be used for on-line test construction methods as in van der Linden & Veldkamp (2004) or He et al. (2014). The possibility of efficiently solving large-sized instances leads us to con- clude that any further research should address the inclusion of additional char- acteristics into the formulation. Among these characteristics, we would like to highlight the joint evaluation of multiple ability levels of the recipients on a single unified test (van der Linder, 2005; Chang & Shiu, 2012; Nguyen & Fong,840 2013) or the inclusion of additional constraints, such as incompatibilities among questions of different sets or limiting the number of questions with specific at- tributes (Hwang et al., 2008; Yao, 2014; Yao et al., 2014). These additional characteristics are usually present in many real situations and would impose challenging changes to both the bin-packing-based formulation as well as to the neighborhood structures proposed in this work. References Anderson, M. J. (2001). A new method for non-parametric multivariate analysis of variance. Austral Ecology, 26 , 32–46. Belov, D. I. (2008). Uniform test assembly. Psychometrika, 73 , 21–38. doi:10.850 1007/s11336-007-9025-0. Belov, D. I., & Armstrong, R. D. (2006). A constraint programming ap- proach to extract the maximum number of non-overlapping test forms. Computational Optimization and Applications, 33 , 319–332. doi:10.1007/ s10589-005-3058-z. Birnbaum, A. (1968). Some latent trait models and their use in inferring an examinees ability. In F. Lord, & M. Novick (Eds.), Statistical theories of mental test scores chapter Some laten. (pp. 385–479). Reading, MA: Addison- Wesley. Brandão, F., & Pedroso, J. a. P. (2014). Fast pattern-based algorithms for cut-860 ting stock. Computers and Operations Research, 48 , 69–80. URL: http://dx. doi.org/10.1016/j.cor.2014.03.003. doi:10.1016/j.cor.2014.03.003. Brusco, M. J., Köhn, H. F., & Steinley, D. (2013). Exact and approximate meth- ods for a one-dimensional minimax bin-packing problem. Annals of Opera- tions Research, 206 , 611–626. URL: http://link.springer.com/10.1007/ s10479-012-1175-5. doi:10.1007/s10479-012-1175-5. 30 http://dx.doi.org/10.1007/s11336-007-9025-0 http://dx.doi.org/10.1007/s11336-007-9025-0 http://dx.doi.org/10.1007/s11336-007-9025-0 http://dx.doi.org/10.1007/s10589-005-3058-z http://dx.doi.org/10.1007/s10589-005-3058-z http://dx.doi.org/10.1007/s10589-005-3058-z http://dx.doi.org/10.1016/j.cor.2014.03.003 http://dx.doi.org/10.1016/j.cor.2014.03.003 http://dx.doi.org/10.1016/j.cor.2014.03.003 http://dx.doi.org/10.1016/j.cor.2014.03.003 http://link.springer.com/10.1007/s10479-012-1175-5 http://link.springer.com/10.1007/s10479-012-1175-5 http://link.springer.com/10.1007/s10479-012-1175-5 http://dx.doi.org/10.1007/s10479-012-1175-5 Chang, T. Y., & Ke, Y. R. (2013). A personalized e-course composition based on a genetic algorithm with forcing legality in an adaptive learn- ing system. Journal of Network and Computer Applications, 36 , 533–542. URL: http://dx.doi.org/10.1016/j.jnca.2012.04.002. doi:10.1016/j.870 jnca.2012.04.002. Chang, T.-Y., & Shiu, Y.-F. (2012). Simultaneously construct IRT-based paral- lel tests based on an adapted CLONALG algorithm. Applied Intelligence, 36 , 979–994. URL: http://link.springer.com/10.1007/s10489-011-0308-x. doi:10.1007/s10489-011-0308-x. Chen, P.-h. (2015). A sampling and classification item selection ap- proach with content balancing. Behavior research methods, . doi:10.3758/ s13428-014-0452-4. Chen, P.-H., Chang, H.-H., & Wu, H. (2012). Item Selection for the Devel- opment of Parallel Forms From an IRT-Based Seed Test Using a Sampling880 and Classification Approach. Educational and Psychological Measurement, 72 , 933–953. doi:10.1177/0013164412443688. Christensen, R. (2011). Plane Asnwers to complex questions. The theory of linear models. (4th ed.). Springer. Dell’ Amico, M., & Martello, S. (1995). Optimal Scheduling of tasks on identical parallel processors. ORSA Journal of Computing, 2 , 191–200. doi:10.1287/ ijoc.7.2.191. Dyckhoff, H. (1990). A typology of cutting and packing problems. Eu- ropean Journal of Operational Research, 44 , 145–159. URL: http: //www.sciencedirect.com/science/article/pii/037722179090350K.890 doi:10.1016/0377-2217(90)90350-K. Edmonds, J., & Armstrong, R. (2009). A mixed integer programming model for multiple stage adaptive testing. European Journal of Operational Re- search, 193 , 342–350. URL: http://linkinghub.elsevier.com/retrieve/ pii/S0377221707010752. doi:10.1016/j.ejor.2007.10.047. Falkenauer, E. (1996). A hybrid grouping genetic algorithm for bin packing. Journal of Heuristics, 2 , 5–30. URL: http://link.springer.com/10.1007/ BF00226291. doi:10.1007/BF00226291. Fleszar, K., & Hindi, K. S. (2002). New heuristics for one-dimensional bin- packing. Computers & Operation Research, 29 , 821–839. doi:10.1016/900 S0305-0548(00)00082-4. Garey, M. R., & Johnson, D. S. (1979). Computers and Intractability. A guide to the theory of NP-Completeness. New York: Freeman. 31 http://dx.doi.org/10.1016/j.jnca.2012.04.002 http://dx.doi.org/10.1016/j.jnca.2012.04.002 http://dx.doi.org/10.1016/j.jnca.2012.04.002 http://dx.doi.org/10.1016/j.jnca.2012.04.002 http://link.springer.com/10.1007/s10489-011-0308-x http://dx.doi.org/10.1007/s10489-011-0308-x http://dx.doi.org/10.3758/s13428-014-0452-4 http://dx.doi.org/10.3758/s13428-014-0452-4 http://dx.doi.org/10.3758/s13428-014-0452-4 http://dx.doi.org/10.1177/0013164412443688 http://dx.doi.org/10.1287/ijoc.7.2.191 http://dx.doi.org/10.1287/ijoc.7.2.191 http://dx.doi.org/10.1287/ijoc.7.2.191 http://www.sciencedirect.com/science/article/pii/037722179090350K http://www.sciencedirect.com/science/article/pii/037722179090350K http://www.sciencedirect.com/science/article/pii/037722179090350K http://dx.doi.org/10.1016/0377-2217(90)90350-K http://linkinghub.elsevier.com/retrieve/pii/S0377221707010752 http://linkinghub.elsevier.com/retrieve/pii/S0377221707010752 http://linkinghub.elsevier.com/retrieve/pii/S0377221707010752 http://dx.doi.org/10.1016/j.ejor.2007.10.047 http://link.springer.com/10.1007/BF00226291 http://link.springer.com/10.1007/BF00226291 http://link.springer.com/10.1007/BF00226291 http://dx.doi.org/10.1007/BF00226291 http://dx.doi.org/10.1016/S0305-0548(00)00082-4 http://dx.doi.org/10.1016/S0305-0548(00)00082-4 http://dx.doi.org/10.1016/S0305-0548(00)00082-4 Hansen, P., Mladenović, N., & Moreno Pérez, J. A. (2010). Variable neighbour- hood search: methods and applications. Annals of Operations Research, 175 , 367–407. URL: http://link.springer.com/10.1007/s10479-009-0657-6. doi:10.1007/s10479-009-0657-6. He, W., Diao, Q., & Hauser, C. (2014). A Comparison of Four Item-Selection Methods for Severely Constrained CATs. Educational and Psychological Mea- surement, 74 , 677–696. URL: http://epm.sagepub.com/cgi/doi/10.1177/910 0013164413517503. doi:10.1177/0013164413517503. Hwang, G. J., Chu, H. C., Yin, P. Y., & Lin, J. Y. (2008). An innovative parallel test sheet composition approach to meet multiple assessment criteria for national tests. Computers and Education, 51 , 1058–1072. doi:10.1016/ j.compedu.2007.10.006. Ishii, T., Songmuang, P., & Ueno, M. (2014). Maximum Clique Algorithm and Its Approximation for Uniform Test Form Assembly. IEEE Transactions on Learning Technologies, 7 , 83–95. URL: http://ieeexplore.ieee.org/ articleDetails.jsp?arnumber=6701350. doi:10.1109/TLT.2013.2297694. Kallrath, J., Rebennack, S., Kallrath, J., & Kusche, R. (2014). Solving920 real-world cutting stock-problems in the paper industry: Mathematical ap- proaches, experience and challenges. European Journal of Operational Re- search, 238 , 374–389. URL: http://dx.doi.org/10.1016/j.ejor.2014. 03.027. doi:10.1016/j.ejor.2014.03.027. Kellerer, H., Pferschy, U., & Pisinger, D. (2004). Knapsack Problems. Springer. van der Linden, W. J. (1998). Optimal Assembly of Psychological and Educational Tests. Applied Psychological Measurement, 22 , 195– 211. URL: http://apm.sagepub.com/content/22/3/195.abstract. doi:10. 1177/01466216980223001. van der Linden, W. J., & Boekkooi-Timminga, E. (1988). A Zero-One Program-930 ming Approach to Guiliksen’s Matched Random Subtests Method. Applied Psychological Measurement, 12 , 201–209. URL: http://apm.sagepub.com/ content/12/2/201.abstract. doi:10.1177/014662168801200210. van der Linden, W. J., & Veldkamp, B. P. (2004). Constraining Item Exposure in Computerized Adaptive Testing With Shadow Tests. Journal of Educational and Behavioral Statistics, 29 , 273–291. doi:10.3102/10769986029003273. van der Linder, W. J. (2005). Linear Models for Optimal Test Design. Springer. López-Camacho, E., Terashima-Marin, H., Ross, P., & Ochoa, G. (2014). A unified hyper-heuristic framework for solving bin packing prob- lems. Expert Systems with Applications, 41 , 6876–6889. URL: http:940 //linkinghub.elsevier.com/retrieve/pii/S0957417414002668. doi:10. 1016/j.eswa.2014.04.043. 32 http://link.springer.com/10.1007/s10479-009-0657-6 http://dx.doi.org/10.1007/s10479-009-0657-6 http://epm.sagepub.com/cgi/doi/10.1177/0013164413517503 http://epm.sagepub.com/cgi/doi/10.1177/0013164413517503 http://epm.sagepub.com/cgi/doi/10.1177/0013164413517503 http://dx.doi.org/10.1177/0013164413517503 http://dx.doi.org/10.1016/j.compedu.2007.10.006 http://dx.doi.org/10.1016/j.compedu.2007.10.006 http://dx.doi.org/10.1016/j.compedu.2007.10.006 http://ieeexplore.ieee.org/articleDetails.jsp?arnumber=6701350 http://ieeexplore.ieee.org/articleDetails.jsp?arnumber=6701350 http://ieeexplore.ieee.org/articleDetails.jsp?arnumber=6701350 http://dx.doi.org/10.1109/TLT.2013.2297694 http://dx.doi.org/10.1016/j.ejor.2014.03.027 http://dx.doi.org/10.1016/j.ejor.2014.03.027 http://dx.doi.org/10.1016/j.ejor.2014.03.027 http://dx.doi.org/10.1016/j.ejor.2014.03.027 http://apm.sagepub.com/content/22/3/195.abstract http://dx.doi.org/10.1177/01466216980223001 http://dx.doi.org/10.1177/01466216980223001 http://dx.doi.org/10.1177/01466216980223001 http://apm.sagepub.com/content/12/2/201.abstract http://apm.sagepub.com/content/12/2/201.abstract http://apm.sagepub.com/content/12/2/201.abstract http://dx.doi.org/10.1177/014662168801200210 http://dx.doi.org/10.3102/10769986029003273 http://linkinghub.elsevier.com/retrieve/pii/S0957417414002668 http://linkinghub.elsevier.com/retrieve/pii/S0957417414002668 http://linkinghub.elsevier.com/retrieve/pii/S0957417414002668 http://dx.doi.org/10.1016/j.eswa.2014.04.043 http://dx.doi.org/10.1016/j.eswa.2014.04.043 http://dx.doi.org/10.1016/j.eswa.2014.04.043 Lu, H.-y. (2014). Application of Optimal Designs to Item Calibration. Plos One, 9 , 1–8. doi:10.1371/journal.pone.0106747. M’Hallah, R., Alkandari, A., & Mladenovic, N. (2013). Packing unit spheres into the smallest sphere using VNS and NLP. Computers and Operations Re- search, 40 , 603–615. URL: http://dx.doi.org/10.1016/j.cor.2012.08. 019. doi:10.1016/j.cor.2012.08.019. Mladenović, N., & Hansen, P. (1997). Variable neighborhood search. Computers & Operations Research, 24 , 1097–1100. URL: http:950 //www.sciencedirect.com/science/article/pii/S0305054897000312. doi:10.1016/S0305-0548(97)00031-2. Nguyen, M. L., & Fong, A. C. M. (2013). Large-Scale Multiobjective Static Test Generation for Web-Based Testing with Integer Programming. IEEE Transac- tions on Learning Technologies, 6 , 46–59. URL: http://www.computer.org/ csdl/trans/lt/2013/01/tlt2013010046.html. doi:10.1109/TLT.2012.22. Oksanen, J., Blanchet, F. G., Kindt, R., Legendre, P., Minchin, P. R., O’Hara, R. B., Simpson, G. L., Solymos, P., M., H., H., S., & Wagner, H. (2015). vegan: Community Ecology Package. URL: http://cran.r-project.org/ package=vegan.960 Pisinger, D. (1995). An expanding-core algorithm for the exact 0- 1 knapsack problem. European Journal of Operational Research, 87 , 175–187. URL: http://www.sciencedirect.com/science/article/pii/ 0377221794000133. doi:10.1016/0377-2217(94)00013-3. Porter, A., Polikoff, M. S., Barghaus, K. M., & Yang, R. (2013). Construct- ing Aligned Assessments Using Automated Test Construction. Educational Researcher, 42 , 415–423. URL: http://edr.sagepub.com/content/42/8/ 415.abstract. doi:10.3102/0013189X13503038. Quiroz-castellanos, M., Cruz-reyes, L., Torres-jimenez, J., S, C. G., Fraire, H. J., & Alvim, A. C. F. (2015). A grouping genetic algorithm with controlled970 gene transmission for the bin packing problem. Computers & Operation Re- search, 55 , 52–64. URL: http://dx.doi.org/10.1016/j.cor.2014.10.010. doi:10.1016/j.cor.2014.10.010. Smits, N., & Finkelman, M. D. (2014). Variable length testing using the ordi- nal regression model. Statistics in Medicine, 33 , 488–499. doi:10.1002/sim. 5936. Songmuang, P., & Ueno, M. (2011). Bees Algorithm for Construction of Multiple Test Forms in E-Testing. IEEE Transactions on Learning Tech- nologies, 4 , 209–221. URL: http://ieeexplore.ieee.org/lpdocs/epic03/ wrapper.htm?arnumber=5560630. doi:10.1109/TLT.2010.29.980 33 http://dx.doi.org/10.1371/journal.pone.0106747 http://dx.doi.org/10.1016/j.cor.2012.08.019 http://dx.doi.org/10.1016/j.cor.2012.08.019 http://dx.doi.org/10.1016/j.cor.2012.08.019 http://dx.doi.org/10.1016/j.cor.2012.08.019 http://www.sciencedirect.com/science/article/pii/S0305054897000312 http://www.sciencedirect.com/science/article/pii/S0305054897000312 http://www.sciencedirect.com/science/article/pii/S0305054897000312 http://dx.doi.org/10.1016/S0305-0548(97)00031-2 http://www.computer.org/csdl/trans/lt/2013/01/tlt2013010046.html http://www.computer.org/csdl/trans/lt/2013/01/tlt2013010046.html http://www.computer.org/csdl/trans/lt/2013/01/tlt2013010046.html http://dx.doi.org/10.1109/TLT.2012.22 http://cran.r-project.org/package=vegan http://cran.r-project.org/package=vegan http://cran.r-project.org/package=vegan http://www.sciencedirect.com/science/article/pii/0377221794000133 http://www.sciencedirect.com/science/article/pii/0377221794000133 http://www.sciencedirect.com/science/article/pii/0377221794000133 http://dx.doi.org/10.1016/0377-2217(94)00013-3 http://edr.sagepub.com/content/42/8/415.abstract http://edr.sagepub.com/content/42/8/415.abstract http://edr.sagepub.com/content/42/8/415.abstract http://dx.doi.org/10.3102/0013189X13503038 http://dx.doi.org/10.1016/j.cor.2014.10.010 http://dx.doi.org/10.1016/j.cor.2014.10.010 http://dx.doi.org/10.1002/sim.5936 http://dx.doi.org/10.1002/sim.5936 http://dx.doi.org/10.1002/sim.5936 http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=5560630 http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=5560630 http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=5560630 http://dx.doi.org/10.1109/TLT.2010.29 Sun, K.-T., Chen, Y.-J., Tsai, S.-Y., & Cheng, C.-F. (2008). Creating IRT-Based Parallel Test Forms Using the Genetic Algorithm Method. Applied Measure- ment in Education, 21 , 141–161. URL: http://www.tandfonline.com/doi/ abs/10.1080/08957340801926151. doi:10.1080/08957340801926151. Vanderbeck, F. (1999). Computational study of a column generation algorithm for bin packing and cutting stock problems. Mathematical Programming, 86 , 565–594. doi:10.1007/s101079900096. Veldkamp, B. P. (2005). Encyclopedia of Social Measurement. El- sevier. URL: http://www.sciencedirect.com/science/article/pii/ B0123693985004473. doi:10.1016/B0-12-369398-5/00447-3.990 Veldkamp, B. P. (2013). Application of robust optimization to auto- mated test assembly. Annals of Operations Research, 206 , 595–610. URL: http://link.springer.com/10.1007/s10479-012-1218-y. doi:10. 1007/s10479-012-1218-y. Wäscher, G., Hauß ner, H., & Schumann, H. (2007). An improved typology of cutting and packing problems. European Journal of Operational Research, 183 , 1109–1130. URL: http://www.sciencedirect.com/science/article/ pii/S037722170600292X. doi:10.1016/j.ejor.2005.12.047. Yao, L. (2014). Multidimensional CAT Item Selection Methods for Domain Scores and Composite Scores With Item Exposure Control and Content Con-1000 straints. Journal of Educational Measurement, 51 , 18–38. doi:10.1111/jedm. 12032. Yao, L., Pommerich, M., & Segall, D. O. (2014). Using Multidimensional CAT to Administer a Short , Yet Precise , Screening Test. Applied Psychological Measurement, 38 , 614–631. doi:10.1177/0146621614541514. 34 http://www.tandfonline.com/doi/abs/10.1080/08957340801926151 http://www.tandfonline.com/doi/abs/10.1080/08957340801926151 http://www.tandfonline.com/doi/abs/10.1080/08957340801926151 http://dx.doi.org/10.1080/08957340801926151 http://dx.doi.org/10.1007/s101079900096 http://www.sciencedirect.com/science/article/pii/B0123693985004473 http://www.sciencedirect.com/science/article/pii/B0123693985004473 http://www.sciencedirect.com/science/article/pii/B0123693985004473 http://dx.doi.org/10.1016/B0-12-369398-5/00447-3 http://link.springer.com/10.1007/s10479-012-1218-y http://dx.doi.org/10.1007/s10479-012-1218-y http://dx.doi.org/10.1007/s10479-012-1218-y http://dx.doi.org/10.1007/s10479-012-1218-y http://www.sciencedirect.com/science/article/pii/S037722170600292X http://www.sciencedirect.com/science/article/pii/S037722170600292X http://www.sciencedirect.com/science/article/pii/S037722170600292X http://dx.doi.org/10.1016/j.ejor.2005.12.047 http://dx.doi.org/10.1111/jedm.12032 http://dx.doi.org/10.1111/jedm.12032 http://dx.doi.org/10.1111/jedm.12032 http://dx.doi.org/10.1177/0146621614541514 Introduction Contributions of this work Paper outline Problem description Literature review Mathematical formulation, lower bounds and special cases Complexity of the problem Description of the algorithm Greedy Initialization heuristic Neighborhood structures Shaking procedure Termination condition and outline of the complete algorithm Computational Experiments Conclusions 
work_fisyw7slizc6tkjozwij2kimli	----	Editorial: Recent Advances on Knowledge Discovery and Business Intelligence Paulo Cortez1 Manuel Filipe Santos1 1 ALGORITMI Research Centre, Department of Information Systems, University of Minho, 4800-058 Guimarães, Portugal Email: pcortez@dsi.uminho.pt, mfs@dsi.uminho.pt 1 Introduction Information Technology (IT) is advancing rapidly. In the last years, IT costs have reduced while data storage, communication and processing capabilities have in- creased. In effect, vast datasets are becoming commonplace. All this data hold valuable knowledge such as trends and patterns, which can be used to improve decision making and optimize chances of success. These datasets are often highly complex for a manual analysis due to the volume, variety and velocity proper- ties of data. Following these advances, there has been a growing interest in using IT to provide useful knowledge, extracted from humans or raw data, for decision support. This interest was approached under different perspectives through the last decades [Carbonell et al., 1983, Ein-Dor and Segev, 1993, Turban et al., 2010, Piatetsky-Shapiro, 2011]: • Machine Learning (ML), since the 1960s; • Decision Support Systems (DSS), since the 1970s; • Expert Systems (ES), since the 1980s; • Analytics, Business Intelligence (BI), Data Mining (DM), Enterprise Infor- mation Systems (EIS), Knowledge Discovery (KD) in data, since the 1990s; and • Big Data and Data Science, since the 2000s. In particular, ES were originally focused on using symbolic information and in- ference in order to mimic human specialists [Buchanan, 1986]. In most cases, these ES were based on expert-driven knowledge that was extracted from the specialists (e.g., by using interviews). However, since the 1990s, there has been a shift towards the use of data-driven models and increase of DM, KD and BI fields. These data- driven models were built from past data in order to produce informed predictions or descriptive knowledge that could be useful for supporting decisions. Following 1 this trend, data-driven knowledge, used either solely or complemented by expert- driven knowledge, turned into a key element of modern ES, highlighting the im- portance of the KD and BI areas. KD is an Artificial Intelligence subfield that is related with the extraction of useful and understandable knowledge from raw data [Fayyad et al., 1996], while BI is an umbrella term that represents several technolo- gies (e.g., Data Warehouses, KD and Dashboards) to access past data and support decision-making [Turban et al., 2010]. Since the 1990s, there has been a rapid expansion of the KD and BI fields in terms of both research contributions and their use in real-world applications. This special issue, entitled ‘Recent Advances on Knowledge Discovery and Business Intelligence’ focuses on novel KD contributions that present a valuable impact in several BI domain applications. It consists of extended versions of papers from the 3rd Knowledge Discovery and Business Intelligence (KDBI) thematic track of the 16th Portuguese Conference on Artificial Intelligence (EPIA 2013) that was held in Angra do Heróısmo, Açores, Portugal. A total of 18 papers were submitted to the 3rd KDBI thematic track, from which the best 9 papers were invited for this special issue. Each extended paper was reviewed by a minimum of three reviewers (related with both the 3rd KDBI thematic task and Expert Systems journal), and passed through two rounds of reviews. Finally, the best four papers were accepted, corresponding to an acceptance rate of 22%, when considering the initial 3rd KDBI submitted papers, and 44%, when considering the extended invited papers. The next section briefly introduces this special issue accepted papers. 2 Contents of the special issue In the first paper ‘Contrast set mining in temporal databases’, Magalhães and Azevedo (2014) extend the Rules for Contrast Sets (RCS) technique for handling temporal data mining tasks. The goal of a Contrast Set is to discover attribute-value pairs that differ in two or more groups. The presented extension proposes a set of temporal patterns that can capture significant differences in Contrast Sets discovered through time. Two distinct real-world problem domains, related with Portuguese labor and NBA basketball data, were used to demonstrate the capabilities of the RCS extension for a temporal analysis, revealing interesting patterns. In ‘Feature selection for clustering categorical data with an embedded modeling approach’, Silvestre et al. (2014) present a novel approach that simultaneously clusters categorical data and selects relevant features. The approach is based on a Gaussian mixture model, where the Minimum Message Length (MML) criterion is used to guide the selection of the relevant features and a modified Expectation- Maximization (EM) algorithm estimates the model parameters. The usefulness of the proposed approach was illustrated on synthetic datasets and real-world data related with the perceived quality of life from the European Official Statistics (EOS). In the work ‘Eigenspace method for spatiotemporal hotspot detection’, Fanaee-T and Gama (2014) deal with hotspot detection, which aims at the identification of unexpected subgroups, in space-time data. The proposed EigenSpot method tracks changes in space-time occurrences and it is potentially useful for surveillance systems (e.g. bio-surveillance). The experiments held adopted synthetic and real data. The 2 latter data was related with brain tumor cases in 32 subregions of the New Mexico State, United States, from 1973 to 1991. The obtained results have shown that EigenSpot compares favorably against the state-of-the-art STScan method. In the last paper, entitled ‘Re-sampling strategies for regression’, Torgo et al. (2014) present a general re-sampling approach for regression tasks, which can be used for forecasting rare values of continuous target variables. Their approach is based on modifications of two popular re-sampling classification techniques (under-sampling and SMOTE). An extensive set of experiments was conducted using 18 datasets from several domains (e.g., automotive industry), highlighting the advantages of the proposed re-sampling methods. All these papers represent the best contributions to the 3rd ‘Knowledge Discovery and Business Intelligence’ track of EPIA 2013. We hope that the contributions of this special issue enrich both KD and BI, promoting the interaction between both fields. Acknowledgments Many individuals contributed to this special issue. We would like to thank Dr. Jon G. Hall, Editor-in-Chief of Expert Systems, for his interest and support. We also wish to thank the other KDBI 2013 track (of EPIA) co-organizers, Lúıs Cav- ique, João Gama and Nuno Marques, without their help this issue would not have been possible. Finally, we thank the authors, who contributed with their papers, and the reviewers (from the KDBI 2013 program committee and also others), who contributed with their detailed comments and valuable opinions. This work was sup- ported by FCT - Fundação para a Ciência e Tecnologia within the Project Scope: PEst-OE/EEI/UI0319/2014. References [Buchanan, 1986] Buchanan, B. G. (1986). Expert systems: working systems and the research literature. Expert systems, 3(1):32–50. [Carbonell et al., 1983] Carbonell, J., Michalski, R. and Mitchell, T. (1983). Ma- chine Learning: A Historical and Methodological Analysis. Artificial Intelligence Magazine, 4(3):69–79. [Ein-Dor and Segev, 1993] Ein-Dor, P. and Segev, E. (1993). A classification of information systems: Analysis and interpretation, Information Systems Research, 4(2):166–204. [Fanaee-T and Gama, 2014] Fanaee-T, H. and Gama, J. (2014). Eigenspace method for spatiotemporal hotspot detection. Expert systems, pages –. [Fayyad et al., 1996] Fayyad, U., Piatetsky-Shapiro, G., and Smyth, P. (1996). Ad- vances in Knowledge Discovery and Data Mining. MIT Press. 3 [Magalhães and Azevedo, 2014] Magalhães, A. and Azevedo, P. (2014). Contrast set mining in temporal databases. Expert systems, pages –. [Piatetsky-Shapiro, 2011] Piatetsky-Shapiro, G. (2011). Poll Results: What do you call analyzing data? http://www.kdnuggets.com/2011/09/ the-top-name-for-data-mining.html. [Silvestre et al., 2014] Silvestre, C., Cardoso, M. and Figueiredo, M. (2014). Feature selection for clustering categorical data with an embedded modeling approach. Expert systems, pages –. [Torgo et al., 2014] Torgo, L., Branco, P., Ribeiro, R. and Pfahringer, B. (2014). Re-sampling strategies for regression. Expert systems, pages –. [Turban et al., 2010] Turban, E., Sharda, R., Aronson, J., and King, D. (2010). Business Intelligence, A Managerial Approach. Prentice-Hall, 2nd edition. The authors Paulo Cortez Paulo Cortez is Associate Professor with Habilitation at the Department of Infor- mation Systems, University of Minho, Portugal. He is also coordinator of Infor- mation Systems and Technologies (IST) research group of ALGORITMI Centre. His research interests include: Business Intelligence (Decision Support, Data Min- ing and Forecasting); and Artificial Intelligence (Computational Intelligence, Neural Networks, Evolutionary Computation and Applications). Currently, he is associate editor of the Expert Systems and Neural Processing Letters journals and participated in 9 R&D projects (principal investigator in 2). He is co-author of more than eighty indexed (ISI, Scopus) publications in international journals (e.g., Decision Support Systems, Applied Soft Computing) and conferences (e.g., IEEE IJCNN). Web-page: http://www3.dsi.uminho.pt/pcortez Manuel Filipe Santos Manuel Filipe Santos is Associate Professor at the Department of Information Sys- tems, University of Minho, Portugal, and leader of the Intelligent Data Systems group of Centro Algoritmi, with interests in the fields of: Business Intelligence, Data Mining and Learning Classifier Systems. He participated in several R&D projects (principal investigator in 3). He is also co-author of several publications in international journals (e.g., Artificial Intelligence in Medicine) and conferences (e.g., GECCO). 4 
work_fk3j5lujubeqvghgryjz6fmxey	----	doi:10.1016/j.eswa.2004.08.003 DTD 5 ARTICLE IN PRESS Expert system methodologies and applications—a decade review from 1995 to 2004 Shu-Hsien Liao* Department of Management Sciences and Decision Making, Tamkang University, No. 151, Yingjuan Rd, Danshuei Jen, Taipei 251, Taiwan, ROC Abstract This paper surveys expert systems (ES) development using a literature review and classification of articles from 1995 to 2004 with a keyword index and article abstract in order to explore how ES methodologies and applications have developed during this period. Based on the scope of 166 articles from 78 academic journals (retrieved from five online database) of ES applications, this paper surveys and classifies ES methodologies using the following eleven categories: rule-based systems, knowledge-based systems, neural networks, fuzzy ESs, object- oriented methodology, case-based reasoning, system architecture, intelligent agent systems, database methodology, modeling, and ontology together with their applications for different research and problem domains. Discussion is presented, indicating the followings future development directions for ES methodologies and applications: (1) ES methodologies are tending to develop towards expertise orientation and ES applications development is a problem-oriented domain. (2) It is suggested that different social science methodologies, such as psychology, cognitive science, and human behavior could implement ES as another kind of methodology. (3) The ability to continually change and obtain new understanding is the driving power of ES methodologies, and should be the ES application of future works. q 2004 Published by Elsevier Ltd. Keywords: Expert systems; Expert system methodologies; Expert system applications; Literature survey 1. Introduction Expert systems (ES) are a branch of applied artificial intelligence (AI), and were developed by the AI community in the mid-1960s. The basic idea behind ES is simply that expertise, which is the vast body of task-specific knowledge, is transferred from a human to a computer. This knowledge is then stored in the computer and users call upon the computer for specific advice as needed. The computer can make inferences and arrive at a specific conclusion. Then like a human consultant, it gives advices and explains, if necessary, the logic behind the advice (Turban & Aronson, 2001). ES provide powerful and flexible means for obtaining solutions to a variety of problems that often cannot be dealt with by other, more traditional and orthodox methods. Thus, their use is proliferating to many sectors of our social and technological life, where their applications 0957-4174/$ - see front matter q 2004 Published by Elsevier Ltd. doi:10.1016/j.eswa.2004.08.003 * Tel.: C886-2947-2044; fax: C886-2945-3007. E-mail address: michael@mail.tku.edu.tw. are proving to be critical in the process of decision support and problem solving. As a part of ES research, this paper surveys the development of ES through a literature review and classification of articles from 1995 to 2004 as a basis, exploring the ES methodologies and applications during that period. The reason for choosing this period is that the Internet was opened to general users in 1994 and this new era of information and communication technology has played important roles, not only in the field of ES, but also in the ability to collect data from online database. This literature survey started on March 2003 and it was based on a search in the keyword index and article abstract for ‘ES’ on the Elsevier SDOS, IEEE Xplore, EBSCO (electronic journal service), Ingenta, and Wiley InterScience online database, for the period from 1995 to 2004, in which 10,439 articles were updated and found on June 2004. After topic filtering, there were 166 articles from 78 journals related to the keyword ‘ES applications’, 98 of which were connected to the methodology of keyword ‘ES methodology’. Based on the scope of 166 articles on ES applications, this paper surveys and classifies ES methodologies using eleven Expert Systems with Applications xx (2004) 1–11 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa Table 1 Rule-based systems and their applications Rule-based systems/applications Authors State transition analysis Ilgun, Kemmerer, and Porras (1995) Psychiatric treatment Goethe and Bronzino (1995) Production planning Hamada, Baba, Sato, and Yufu (1995) Advisory system Kose et al. (1995) Teaching Chan, Ma, Chan, and Chen (1995) Electronic power planning Rahman and Hazim (1996) Automobile process planning Sabourin and Villeneuve (1996) Hypergraph representation Ramaswamy, Sarkar, and Chen (1997) System development Mulvaney and Bristow (1997) Knowledge verification/vali- Wu and Lee (1997) Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–112 DTD 5 ARTICLE IN PRESS categories: rule-based systems, knowledge-based systems, neural networks, fuzzy ESs, object-oriented methodology, case-based reasoning (CBR), system architecture develop- ment, intelligent agent (IA) systems, modeling, ontology, and database methodology together with their applications for different research and problem domains. The rest of the paper is organized as follows. Sections 2–12 present the survey results of ES methodologies and applications based on the above categories, respectively. Section 13 presents some discussion, extending to sugges- tions for future development of ES methodologies and applications. Finally, Section 14 contains a brief conclusion. dation Alcohol production Guerreiro et al. (1997) DNA histogram interpretation Marchevsky, Truong, and Tolmachoff (1997) Knowledge base maintenance Higa and Lee (1998) Scheduling strategy Zupan and Cheng (1998) Management fraud assessment Deshmukh and Talluru (1998) Knowledge acquisition Wu and Chen (1999) Communication system fault diagnosis Leon, Mejias, Luque, and Gonzalo (1999) Bioseparation Lienqueo, Salgado, and Asenjo (1999) Material processing design Kim and Im (1999) Resource utilization McCoy and Levary (2000) Biochemical nanotechnology Wasiewicz, Janczak, Mulawka, and Plucienniczak (2000) Probabilistic fault diagnosis Leung and Romagnoli (2000) Agriculture planning Plant and Vayssieres (2000) Load scheduling Croce et al. (2001) Apiculture Mahaman et al. (2002) Agricultural diagnostic/advisory Mahaman, Passam, Sideridis, and Yialouris (2003) Geoscience Soh, Tsatsoulis, Gineris, and Bertoia (2004) Sensor control Valenzuela, Bentley, and Lorenz (2004) Tutoring system Hatzilygeroudis and Prentzas (2004) Knowledge representation Hatzilygeroudis and Prentzas (2004) 2. Rule-based systems and their applications A rule-based ES is defined as one, which contains information obtained from a human expert, and represents that information in the form of rules, such as IF–THEN. The rule can then be used to perform operations on data to inference in order to reach appropriate conclusion. These inferences are essentially a computer program that provides a methodology for reasoning about information in the rule base or knowledge base, and for formulating conclusions. Applications of rule-based systems on ESs are including: state transition analysis, psychiatric treatment, production planning, advisory system, teaching, electronic power planning, automobile process planning, hypergraph rep- resentation, system development, knowledge verification/ validation, alcohol production, DNA histogram interpre- tation, knowledge base maintenance, scheduling strategy, management fraud assessment, knowledge acquisition, knowledge representation, communication system fault diagnosis, bioseparation, material processing design, resource utilization, biochemical nanotechnology, proba- bilistic fault diagnosis, agriculture planning, load schedul- ing, apiculture, tutoring system, geoscience, and sensor control. The methodology of rule-based systems and their applications are categorized in Table 1. 3. Knowledge-based systems and their applications The most common definition of KBS is human-centered. This highlights the fact that KBS have their roots in the field of artificial intelligence (AI) and that they are attempts to understand and initiate human knowledge in computer systems (Wiig, 1994). The four main components of KBS are usually distinguished as: a knowledge base, an inference engine, a knowledge engineering tool, and a specific user interface (Dhaliwal & Benbasat, 1996). On the other hand, the term KBS includes all the organizational information technology applications that may prove helpful to manage the knowledge assets of an organization, such as ESs, rule- based systems, groupware, and database management systems (DBMS) (Laudon & Laudon, 2002). Some of these applications which are implemented by knowledge-based systems include the following: medical treatment, personal finance planning, engineering failure analysis, waste management, production management, thermal engineering, decision support, knowledge manage- ment, knowledge representation, power electronics design, framed buildings evaluation, financial analysis, chemical incident management, automatic tumor segmentation, business game, climate forecasting, agricultural manage- ment, steel composition design, strategic management, environmental protection, wastewater treatment, decision making and learning, isokinetics interpretation, chemical process controlling, therapy planning, plant process control, outage locating planning, concurrent system design, case validation, chip design, agriculture planning, power trans- mission protection, crop production planning, tropospheric chemistry modeling, planar robots, and urban design. The methodology of knowledge-based systems and their appli- cations are categorized in Table 2. Table 2 Knowledge-based systems and their applications Knowledge-based systems/ applications Authors Medical treatment Alonso-Amo, Perez, Gomez, and Montes (1995) Personal finance planning Dirks, Kingston, and Haggith (1995) Engineering failure analysis Graham-Jones and Mwllor (1995) Waste management Wei and Weber (1996) Production management Dawood (1996) Thermal engineering Afgan and Carvalho (1996) Decision support Keefe and Preece (1996) Knowledge management Dutta (1997) Knowledge representation Mitra and Basu (1997) Framed buildings evaluation Lu and Simmonds (1997) Power electronics design Fezzani, Piquet, and Foch (1997) Financial analysis Matsatsinis, Doumpos, and Zopounidis (1997) Chemical incident management Finch and Lees (1997) Automatic tumor segmentation Clark et al. (1998) Business game Duan, Edwards, and Robins (1998) Climate forecasting Rodionov and Martin (1999) Agricultural management Girard and Hubert (1999) Steel composition design Manohar, Shivathaya, and Ferry (1999) Strategic management Volberda and Rutges (1999) Environmental protection Gomolka and Orlowski (2000) Wastewater treatment Baeza, Ferreira, and Laufuente (2000) Decision making and learning Mockler, Dologite, and Gartenfeld (2000) Isokinetics interpretation Alonso, Fuertes, Martinez, and Montes (2000) Chemical process controlling Barrera-Cortes, Astruc, and Tufeu (2001) Physical therapy planning Tunez, Aguila, and Marin (2001) Plant process control Acosta, Gonzalez, and Pulido (2001) Outage locating planning Liu and Schulz (2002) Concurrent system design Mills and Gomaa (2002) Case validation Knauf, Gonzalez, and Abel (2002) Chip design Bourbakis, Mogzadeh, Mertoguno, and Koutsougeras (2002) Agricultural planning Cohen and Shoshany (2002) Power transmission protection Orduna, Garces, and Handschin (2003) Crop production planning Edrees, Rafea, Fathy, and Yahia (2003) Tropospheric chemistry model- ing Saunders, Pascoe, Johnson, Pilling, and Jenkin (2003) Urban design Xirogiannis, Stefanou, and Glykas (2004) Planar robots Sen, Minambres, Garrido, Almansa, and Soto (2004) Table 3 Neural networks and their applications Neural networks/applications Authors Fault diagnosis Wang, Qu, Liu, and Cheng (2004), Yang, Han, and Kim, (2004) Optimal power flow Decision making Alarm processing system Inference mechanisms Diagnostic system Machine learning Fu (1998) Power load forecasting Facility layout design Process control Knowledge learning Gold mining process design Robotic systems Parameter setting Waste treatment Biomedical application Mitigation processes control Engineering ceramics Acoustic signal diagnosing Li, Tasi, Tasi, and Chiu (2004) Crude oil distillation Liau et al. (2004) Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–11 3 DTD 5 ARTICLE IN PRESS 4. Neural networks and their applications An artificial neural network (ANN) is a model that emulates a biological neural network. This concept is used to implement software simulations for the massively parallel processes that involve processing elements inter- connected in network architecture. The artificial neuron receives inputs that are analogous to the electrochemical impulses that the dendrites of biological neurons receive from other neurons. The output of the artificial neuron corresponds to signals sent out from a biological neuron over its axon. These artificial signals can be changed similarly to the physical changes occurring at neural synapses (Turban & Aronson, 2001). Some of the applications that are implemented by neural networks are the following: fault diagnosis, optimal power flow, decision making, alarm processing system, inference mechanisms, diagnostic system, machine learning, power load forecasting, facility layout design, process control, knowledge learning, gold mining process design, robotic systems, parameter setting, waste treatment, engineering ceramics, mitigation processes control, acoustic signal diagnosing, crude oil distillation, and biomedical appli- cation. The methodology of neural networks and their applications are categorized in Table 3. 5. Fuzzy expert systems and their applications Fuzzy ESs are developed using the method of fuzzy logic, which deals with uncertainty. This technique, which uses the mathematical theory of fuzzy sets, simulates the process of normal human reasoning by allowing the computer to behave less precisely and logically than conventional computers. This approach is used because decision-making is not always a matter of black and white, true or false; it often involves gray areas and the term may be. Accordingly, creative decision-making processes can be characterized unstructured, playful, contentious, and ram- bling (Jamshidi, Titli, Zadeh, & Boverie 1997). Some applications implemented by fuzzy ESs are such as: power load forecasting, online scheduling, chemical Table 4 Fuzzy expert systems and their applications Fuzzy expert systems/ applications Authors Power load forecasting Kim, Park, Hwang, and Kim (1995) Online scheduling Chang and Thia (1996) Chemical process fault diagnosis Ozyurt and Kandel (1996) Ecological planning Zhu, Band, Dutton, and Nimlos (1996) Power system diagnosis Cho and Park (1997) Control systems Bugarin and Barro (1998) Uncertainly reasoning Pan, DeSouza, and Kak (1998) Knowledge integration Lee, Han, Song, and Lee (1998) Fault diagnosis Lee et al. (2000) and Soliman, Rizzoni, and Kim (1999) Power system classification Dash, Mishra, Salama, and Liew (2000) Fault detection EI-Shal and Morris (2000) Demand evaluation Benson and Asgarpoor (2000) Wastewater treatment Carrasco, Rodriguez, Punal, Roca, and Lema (2004), Punal, Rodriguez, Car- rasco, Roca, and Lema (2002), and Punal et al. (2001) Machinability data selection Wong and Hamouda (2002, 2003) Water supply forecast Mahabir, Hicks, and Fayek (2003) Radiography classification Liao (2003) On-line analytic processing Leung, Lau, and Kwong (2003) Hotel selection Ngai and Wat (2003) Dryer tool integration Lababidi and Baker (2003) Medical diagnosis Meesad and Yen (2003) and Sendelj and Devedzic (2004) Pooled flood frequency analysis Shu and Burn (2004) Medical consultation system Boegl, Adlassnig, Hayashi, Rothenfluh, and Leitich (2004) Job matching Drigs et al. (2004) Performance indexing Padilla-Medina and Sanchez-Marin (2004) Computer security Reznik and Dabke (2004) Gesture recognition Frantti and Kallio (2004) Table 5 Object-oriented methodology and their applications Object-oriented methodology/ applications Authors Industry diagnosis Batanov and Cheng (1995) Knowledge representation Vranes and Stanojevic (1995) Electronic power capacity planning Deb (1995) Knowledge learning Menzies (1997) Power system maintenance Kawahara, Sasaki, Kubokawa, Asahara, and Sugiyama (1998) Knowledge engineering Geymayr and Ebecken (1998) Manufacturing information network Lau, Tso, and Ho (1998) Syntactic programming Depradine (2003) Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–114 DTD 5 ARTICLE IN PRESS process fault diagnosis, ecological planning, control sys- tems, uncertainly reasoning, knowledge integration, fault diagnosis, power system classification, fault detection, demand evaluation, wastewater treatment, machinability data selection, water supply forecast, radiography classifi- cation, on-line analytic processing, hotel selection, dryer tool integration, pooled flood frequency analysis, medical consultation system, job matching, performance indexing, computer security, gesture recognition, and medical diag- nosis. The methodology of fuzzy ESs together with their applications is categorized in Table 4. 6. Object-oriented methodology and their applications Object-oriented methodology combines into one object data together with the specific procedures that operate on this data, where the object combines data and program code. Instead of passing data to procedures, programs send a message for an object to perform a procedure that is already embedded in it. Then, the same message may be sent to many different objects, but each will implement that message differently. An object’s data are encapsulated from other parts of the system, so each object is an independent software building block that can be used in many different systems without changing the program code. Some applications implemented by object-oriented methodology include the following: industry diagnosis, knowledge learning, manufacturing information network, power system maintenance, knowledge engineering, syn- tactic programming, and knowledge representation. The methodology of object-oriented methodology and their applications are categorized in Table 5. 7. Case-based reasoning and their applications The basic idea of CBR is to adapt solutions that were used to solve previous problems and use them to solve new problems. In CBR, descriptions of past experience of human specialists, represented as cases, are stored in a database for later retrieval when the user encounters a new case with similar parameters. The system searches for stored cases with problem characteristics similar to the new one, finds the closest fit, and applies the solutions of the old case to the new case. Successful solutions are tagged to the new case and both are stored together with the other cases in the knowledge base. Unsuccessful solutions also are appended to the case base along with explanations as to why the solutions did not work (Kolonder, 1994). Some of the applications implemented by CBR include the following: manufacturing process design, knowledge management, power system restoration training, ultrasonic inspection, medical planning, medical application, fault diagnosis, e-learning, and knowledge modeling. These CBR and their applications are categorized in Table 6. 8. Modeling and their applications Modeling methodology becomes an interdisciplinary methodology of ES in order to build formal relationships with logical model design in different knowledge/problem Table 6 Case-based reasoning and their applications Case-based reasoning/applications Authors Manufacturing process design Takahashi, Oono, and Saitog (1995) Knowledge management Noh, Lee, Kim, Lee, and Kim (2000) Power system restoration training Islam and Chowdhury (2001) Ultrasonic inspection Jarmulak, Kerckhoffs, and Veen (2001) Medical planning Abidi and Manickam (2002) Medical application Montani and Bellazzi (2002) Knowledge modeling Gardan and Gardan (2003) Fault diagnosis Yang et al. (2004) E-learning Fu and Shen (2004) Table 8 System architecture and their applications System architecture/applications Authors Material evaluation and selection Mahmoud and AL-Hammad (1996) Computer aided design Tucho, Sierra, Fernandez, Vijande, and Moris (2003) and Vranes and Stanojevic (1999) Ergonomics design Gilad and Karni (1999) ISO system implementation Khan and Hafiz (1999) Corporate recovery decision sup- port Collier, Leech, and Clark (1999) Concurrent engineering Reidsema and Szczerbicki (2001) Military application Liao (2001) Training simulator Lopez, Flores, and Garcia (2003) Ferryboat configuration Shaalan, Rizk, Abdelhamid, and Bahgat (2004) Liquid retaining structure design Chau and Albermani (2004) Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–11 5 DTD 5 ARTICLE IN PRESS domains. Furthermore, modeling technology can provide quantitative methods to analyze data to represent or acquire expert knowledge with inductive logic programming or algorithms so that AI, cognitive science and other research fields could have broader platforms to implement techno- logies for ES development. The applications implemented by modeling are such as: process control, medical analysis, management decision- making, software evaluation, medical system validation, assembly task planning and simulation, transport terminal design, project allocation, and endometrial hyperplasia classification. The methodology of modeling and their applications are categorized in Table 7. 9. System architecture and their applications System architecture of an ES is similar to an architectural sketch of a house. It gives users a general idea of what the system will look like and how it is going to implement systems. The architecture shows the general capabilities of the system, the users’ interfaces, system functions, system (data) flow, system management, DBMS, necessary proto- col, and specific programming language, such as blackboard architecture, CommonKADS, etc. Once the system Table 7 Modeling and their applications Modeling/applications Authors Process control Medical analysis Peng, Xiao, Nie, Wang, and Wang (1996) Management decision-making Mookerjee and Mannino (1997) Software evaluation Vlahavas, Stamelos, Refanidis, and Tsoukias (1999) Medical system validation Martin-Baramera et al. (2000) Assembly task planning and simulation Zha and Lim (2000) Endometrial hyperplasia classi- fication Morrison et al. (2002) Transport terminal design Abacoumkin and Ballis (2004) Project allocation Cheung, Hong, and Ang (2004) architecture design and implementation are completed, users can manipulate and control system functions on the system architecture. Some of the applications implemented by system architecture illustrate the following: material evaluation and selection, computer aided design, ergonomics design, ISO system implementation, corporate recovery decision support, concurrent engineering, military application, train- ing simulator, liquid retaining structure design, and ferry- boat configuration. These System architectures and their applications are categorized in Table 8. 10. Intelligent agents and their applications An IA is a computer program that helps a user with routine computer tasks. This is a new technology, and as such there are several definitions, database capabilities, and different applications in autonomous programs. Several of the names used to describe IAs are include software agents, wizards, and multi-agent (Turban & Aronson, 2001). Some of the applications implemented by IAs are such as: tutoring systems, system analysis and design, electronic service maintenance, carbon contamination rules, know- ledge representation, adaptive systems, air pollution control, building architecture design, agricultural decision support, industry simulation, and knowledge engineering on the WWW platform. The methodology of IAs together with their applications is categorized in Table 9. 11. Ontology and their applications Ontology is a system of vocabulary, which is used as a fundamental concept for describing the task/domain know- ledge to be identified. This vocabulary is used as a communication basis between domain experts and know- ledge engineers. Accordingly, a reusable task/domain model can be represented and a computer program code is Table 9 Intelligent agents and their applications Intelligent agents/applications Authors Tutoring systems Cruces and Arriaga (2000) Supply chain Gjerdrum, Shah, and Papageorgiou (2001) System analysis and design Gruer, Hilaire, Koukam, and Cet- narowicz (2002) Electronic service maintenance Yu, Iung, and Panetto (2003) Carbon contamination rules Vegh (2003) Knowledge representation Qian, Li, Jiang, and Wen (2003) Industry simulation Aldea et al. (2004) Adaptive systems Juuso (2004) Air pollution control Zhou, Huang, and Chan (2004) Building architecture design Alibaba and Ozdeniz (2004) Olive oil automation Bonstre et al. (2004) Agricultural decision support Thomson and Willoughby (2004) Knowledge engineering on the WWW platform Shaalan, EI-Badry, and Rafea (2004) Table 11 Database methodology and their applications Database methodology/ applications Authors Power system planning Park and Lee (1995) Geography planning Kirkby (1996) Geographical information system Filis, Sabrakos, Yialouris, Sideridis, and Mahaman (2003) Sedimentary rock interpretation Abel, Silva, Ros, Mastella, and Camp- bell (2004) Traditional Chinese medicine diagnosis Wang et al. (2004) Medical expert system Yan et al. (2004) Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–116 DTD 5 ARTICLE IN PRESS generated in that ontology for knowledge acquisition, reuse, heuristic learning. Some applications implemented by ontology include the following: medical decision support, knowledge reuse, preventive control, landscape assessment, knowledge acquisition, and chess heuristic pruning. These ontology and their applications are categorized in Table 10. 12. Database methodology and their applications A database is a collection of data organized to efficiently serve many applications by centralizing the data and minimizing redundant data (McFadden, Hoffer, & Prescott, 2000). A DBMS is the software that permits an organization to centralize data, manage them efficiently, and provide access to the stored data by application programs (Laudon & Laudon, 2002). However, some large databases make knowledge discovery computationally expensive because some domains or background knowledge, hidden in the database may guide and restrict the search for important knowledge. Therefore, modern database methodologies need to process large volumes, multiple hierarchies, and different data formats to discover in-depth expert Table 10 Ontology and their applications Ontology/applications Authors Medical decision support Tu, Eriksson, Gennari, Shahar, and Musen (1995) Knowledge reuse Takaoka and Mizoguchi (1996) Preventive control Thukaram and Parthasarathy (1997) Landscape assessment Martinez-Bejar, Ibanez-Cruz, Compton, and Cao (2001) Knowledge acquisition Ruiz-Sanchez, Valencia-Garcia, Fernandez-- Breis, Martinez-Bejar, and Compton (2003) Chess heuristic pruning Montani and Bellazzi (2002) Knowledge modelling Gardan and Gardan 2(003) knowledge from large databases, such as data mining and searching approach. Some applications implemented by database method- ology present as the following: power system planning, geography planning, geographical information system, sedimentary rock interpretation, traditional Chinese medi- cine diagnosis, and medical ES. The methodology of database methodology and their applications are categor- ized in Table 11. 13. Discussions, limitations, and suggestions 13.1. Discussions ES methodologies and applications are a broad category of research issues on ES Some specific methodologies and methods are presented as examples for exploring the suggestions and solutions to specific ES problem domains. Therefore, methodologies and applications of ES are attracting much attention and efforts, both academic and practical. From this literature review, we can see that ES methodologies and applications developments are diversi- fied due to their authors’ backgrounds, expertise, and problem domains. This is why some authors can appear in the literature on different methodologies and applications. On the other hand, some methodologies have common concepts, and types of methodology. For example, rule- based systems and knowledge-based systems, or fuzzy logic versus ANN methodology. However, a few authors work in different methodologies and applications. This indicates that the trend of development on methodology is also diversified due to author’s research interests and abilities in the methodology and problem domain. This may indicate that the development of ES methodologies is directed toward expertise orientation. Furthermore, some applications have a high degree of overlap in different methodologies. For example, teaching/ training, knowledge acquisition, knowledge representation, knowledge learning, fault diagnosis/detec- tion, medical applications, production planning, system design/development, modeling, process control, decision Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–11 7 DTD 5 ARTICLE IN PRESS making, waste treatment, resource management, biomedical application, robotic systems, forecasting, ecological plan- ning, agriculture planning, geoscience, power system planning, chemical application, industry planning, manage- ment issues, and knowledge reuse, are all topics of different methodologies, which implement ES in a common problem domain. This indicates that those applications are the major trend of ES development, and many methodologies focus on these problems. This may direct development of ES applications toward problem domain orientation. In this paper, most of the articles discussed were from different categories including agriculture, agronomy, auto- mation, biochemistry, biology, chemistry, computer science, biology, ecology, education, energy, engineering, entomology, environmental sciences, genetics, geochemis- try, geology, geosciences, health care sciences, hematology, hydrology, materials, mathematics, mechanics, medical, military, operation research/management sciences, onto- logy, plant science, remote sensing, robotics, and water resources, which retrieved from Elsevier SDOS, IEEE Xplore, EBSCO (electronic journal service), Ingenta, and Wiley InterScience online database. We do not conclude that ES methodologies and applications are not developed in other science fields. However, we would like to see more ES methodologies and applications of different research fields published in order to broaden our horizon of academic and practice works on ES. 13.2. Limitations Firstly, a literature review for the broad category of ES methodologies and applications is a difficult task due to the extensive background knowledge needed for collecting, studying, and classifying these articles. Although limited in background knowledge, this paper makes a brief literature review of ES from 1995 to 2004 in order to explore how ES methodologies and applications have developed in this period. Indeed, the categorization of methodologies and their applications is based on the keyword index and article abstract in this research. Some other articles may have implemented similar ES methodologies in their applications without an ES index, so this paper might not find these reference sources. Therefore, the first limitation of this article is the author’s limited knowledge in presenting an overall picture of this subject. Secondly, although 166 articles from 78 academic journals (five online databases) cited in this paper, there are other academic journals listed in the science citation index (SCI) engineering index (EI), and the social science citation index (SSCI), as well as other academic journals/ magazines, practical articles and reports are not included in this survey. These would have provided more complete information to explore the development of ES methodo- logies and applications. Thirdly, non-English publications are not considered in this survey to determine the effects of different cultures on the development of ES methodologies and applications. We believe that ES methodologies and applications in addition to those discussed in this article have also been developed and published in other areas and languages. 13.3. Suggestions (1) Other social science methodologies. In this article, the definition of ES methodology is not complete because other methodologies, such as social science method- ologies, were not included in the survey. However, qualitative questionnaires and statistical methods are another research technology to solve problems in social studies. For example, cognitive science, psychology and human behavior are used to implement different methods for exploring specific human expert problem. Therefore, other social sciences methodologies may include ES methodology categories in future works. (2) Integration of methodologies. ES is an interdisciplinary research topic. Thus, future ES developments need integration with different methodologies, and this integration of methodologies and cross-interdisciplin- ary research may offer more technologies to investigate ES problems. (3) Change is a source of future ES development. The change due to social and technical reasons may either enable or inhibit ES methodologies and application development. This means that inertia, stemming from the use of routine problem solving procedures, stagnant knowledge sources, and following past experience or knowledge may impede changes in terms of learning and innovation for individuals and organizations. Therefore, to continue creating, sharing, learning, and storing knowledge on different methodologies and application domains may also become a source of ES development. 14. Conclusions This paper is based on a literature review of ES methodologies and applications from 1995 to 2004 using a keyword index and article title search. We conclude that ES methodologies are tending to develop towards expertise orientation and that ES applications development is a problem-oriented domain. It is suggested that different social science methodologies, such as psychology, cognitive science, and human behavior could implement ES as another kind of methodology. Integration of qualitative, quantitative and scientific methods and integration of ES methodologies studies may broaden our horizon on this subject. Finally, the ability to continually change and obtain new understanding is the power of ES methodologies, and will be the ES application of future works. Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–118 DTD 5 ARTICLE IN PRESS References Abacoumkin, C., & Ballis, A. (2004). Development of an expert system for the evaluation of conventional and innovative technologies in the intermodal transport area. European Journal of Operational Research, 152, 410–419. Abel, M., Silva, A. L., Ros, L. F., Mastella, L. S., & Campbell, J. A. (2004). PetroGrapher: managing petrographic data and knowledge using an intelligent database application. Expert Systems with Applications, 26, 9–18. Abidi, S. S. R., & Manickam, S. (2002). Leveraging XML-based electronic medical records to extract experiential clinical knowledge. Inter- national Journal of Medical Informatics, 68, 187–203. Acosta, G., Gonzalez, C. A., & Pulido, B. (2001). Basic tasks for knowledge-based supervision in process control. Engineering Appli- cations of Artificial Intelligence, 14, 441–455. Afgan, N. H., & Carvalho, M. G. (1996). Knowledge-based expert system for fouling assessment of industrial heat exchangers. Applied Thermal Engineering, 16, 203–208. Aldea, A., Banares-Alcantara, R., Jimenez, L., Moreno, A., Martinez, J., & Riano, D. (2004). The scoipe of application of multi-agent systems in the process industry: three case studies. Expert Systems with Applications, 26, 39–47. Alibaba, H. Z., & Ozdeniz, M. B. (2004). A building elements selection system for architects. Building and Environment, 39, 307–316. Alonso, F., Fuertes, J. L., Martinez, L., & Montes, C. (2000). An incremental solution for developing knowledge-based software: its application to an expert system for isokinetics interpretation. Expert Systems with Applications, 18, 165–184. Alonso-Amo, F., Perez, A. G., Gomez, G. L., & Montes, C. (1995). An expert system for homeopathic glaucoma treatment (SEHO). Expert Systems with Applications, 8, 89–99. Baeza, J. A., Ferreira, E. C., & Laufuente, J. (2000). Knowledge-based supervision and control of wastewater treatment plant: a real-time implementation. Water Science and Technology, 41, 129–137. Barrera-Cortes, J., Astruc, J. P., & Tufeu, R. (2001). Knowledge base specification to automate the fluid critical point of fluids. Applied Artificial Intelligence, 15, 453–470. Batanov, D. N., & Cheng, Z. (1995). An object-oriented expert system for fault diagnosis in the ethylene distillation process. Computers in Industry, 27, 237–249. Benson, S. J., & Asgarpoor, S. (2000). A fuzzy expert system for evaluation of demand-side management alternatives. Electronic Machines and Power Systems, 28, 749–760. Boegl, K., Adlassnig, K. P., Hayashi, Y., Rothenfluh, T. E., & Leitich, H. (2004). Knowledge acquisition in the fuzzy knowledge representation framework of a medical consultation system. Artificial Intelligence in Medicine, 30, 1–26. Bonstre, A., Ors, R., & Peris, M. (2004). Advanced automation of a flow injection analysis system for quality control of olive oil through the use of a distributed expert system. Analytica Chimica Acta, 506, 189–195. Bourbakis, N. G., Mogzadeh, a., Mertoguno, S., & Koutsougeras, C. A. (2002). A knowledge-based expert system for automatic visual VLSI reverse-engineering: VLSI layout version. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 32, 428–437. Bugarin, A. J., & Barro, S. (1998). Reasoning with truth values on compacted fuzzy chained rules. IEEE Transactions on Systems, Man, and Cybernetics—Part B: Cybernetics, 28, 34–46. Carrasco, E. F., Rodriguez, J., Punal, A., Roca, E., & Lema, J. M. (2004). Diagnosis of acidification states in an anaerobic wastewater yreatment plant using a fuzzy-based system. Control Engineering Practice, 12, 59–64. Chan, Y. F., Ma, H. K., Chan, H. Y., & Chen, T. Y. (1995). Teaching family planning with expert system. Computers Education, 24, 293–298. Chang, C. S. M., & Thia, B. S. (1996). Online rescheduling of mass rapid transit systems: fuzzy expert system approach. IEE Proceedings Electronic Power Applications, 143, 307–316. Chau, K. W., & Albermani, F. (2004). Hybrid knowledge representation in a blackboard KBS for liquid retaining structure design. Engineering Application of artificial Intelligence, 17, 11–18. Cheung, Y., Hong, G. M., & Ang, K. K. (2004). A dynamic project allocation algorithm for a distributed expert system. Expert Systems with Applications, 26, 225–232. Cho, H. J., & Park, J. K. (1997). An expert system for fault section diagnosis of power systems using fuzzy relations. IEEE Transactions on Power Systems, 12, 342–348. Clark, M. C., Hall, L. O., Goldgof, D. B., Velthuizen, R., Murtagh, R., & Silbiger, M. S. (1998). Automatic tumor segmentation using knowl- edge-based techniques. IEEE Transactions on Medical imaging, 17, 187–201. Cohen, Y., & Shoshany, M. (2002). A national knowledge-based crop recognition in Mediterranean environment. International Journal of Applied Earth Observation, 4, 75–87. Collier, P. A., Leech, S. A., & Clark, N. (1999). A validated expert system for decision making in corporate recovery. International Journal of Intelligent System in Accounting, Finance and Management, 8, 75–88. Croce, F., Delfino, B., Fazzini, P. A., Massucco, S., Morini, A., Silvestro, F., & Sivieri, M. (2001). Operation and management of the electronic system for industrial plants: an expert system prototype for load- scheduling operator assistance. IEEE Transactions on Industry Applications, 37, 701–708. Cruces, A. L. L., & Arriaga, F. D. (2000). Reactive agent design for intelligent tutoring systems. Cybernetics and Systems: An International Journal, 31, 1–47. Dash, P. K., Mishra, S., Salama, M. M., & Liew, A. C. (2000). Classification of power system disturbances using a fuzzy expert system and a fourier linear combiner. IEEE Transactions on power Delivery, 15, 472–477. Dawood, N. N. (1996). A strategy of knowledge elicitation for developing an integrated bidding/production management expert system for the precast industry. Advances in Engineering Software, 25, 225–234. Deb, A. K. (1995). Object-oriented expert system estimated line ampacity. IEEE Computer Application in Power, 12, 30–35. Depradine, C. (2003). Expert system for extracting syntactic information from Java code. Expert Systems with Applications, 25, 187–198. Deshmukh, A., & Talluru, L. (1998). A rule-based fuzzy reasoning system for assessing the risk of management fraud. International Journal of Intelligent Systems in Accounting, Finance and Management, 7, 223–241. Dhaliwal, J. S., & Benbasat, I. (1996). The use and effects of knowledge- based system explanations: theoretical foundations and a framework for empirical evaluation. Information Systems Research, 7, 342–362. Dirks, S., Kingston, J. K. C., & Haggith, M. (1995). Development of a KBS for personal financial planning guided by pragmatic KADS. Expert Systems with Applications, 9, 91–101. Drigs, A., Kouremenos, S., Vrettos, S., & Kouremenos, D. (2004). An expert system for job matching of the unemployed. Expert Systems with Applications, 26, 217–224. Duan, Y., Edwards, J. S., & Robins, P. C. (1998). Experiences with EXGAME: an expert system for playing a competitive business game. International Journal of Intelligent System in Accounting, Finance, and Management, 7, 1–19. Dutta, S. (1997). Strategies for implementing knowledge-based systems. IEEE Transactions on Engineering Management, 44, 79–90. Edrees, S. A., Rafea, A., Fathy, I., & Yahia, M. (2003). NEPER: a multiple strategy wheat expert system. Computers and Electronics in Agricul- ture, 40, 27–43. EI-Shal, S. M., & Morris, A. S. (2000). A fuzzy system for fault detection in statistical process control of industrial processes. IEEE Transactions on Systems, Man, and Cybernetics—Part C: Applications and Reviews, 30, 281–292. Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–11 9 DTD 5 ARTICLE IN PRESS Fezzani, D., Piquet, H., & Foch, H. (1997). Expert system for the CAD in power electronics—application to UPS. IEEE Transaction on Power Electronics, 12, 578–586. Filis, I. V., Sabrakos, M., Yialouris, C. P., Sideridis, A. B., & Mahaman, B. (2003). GEDAS: an integrated geographical expert database system. Expert Systems with Applications, 24, 25–34. Finch, D. J., & Lees, P. F. (1997). A hybrid knowledge-based system for chemical incident management. Expert Systems with Applications, 12, 349–361. Frantti, T., & Kallio, S. (2004). Expert system for gesture recognition in terminal’s user interface. Expert Systems with Applications, 26, 189–202. Fu, Y., & Shen, R. (2004). GA based CBR approach in Q&A system. Expert Systems with Applications, 26, 167–170. Gardan, N., & Gardan, Y. (2003). An application of knowledge based modeling using scripts. Expert systems with Applications, 25, 555–568. Gilad, I., & Karni, R. (1999). Architecture of an expert system for ergonomics analysis and design. International Journal of Industrial Ergonomics, 23, 205–221. Girard, N., & Hubert, B. (1999). Modeling expert knowledge with knowledge-based systems to design decision aids. Agricultural Systems, 59, 123–144. Gjerdrum, J., Shah, N., & Papageorgiou, L. G. (2001). A combined optimization and agent-based approach to supply chain modeling and performance assessment. Production Planning and Control, 12, 81–88. Goethe, J. W., & Bronzino, J. D. (1995). An expert system for monitoring psychiatric treatment. IEEE Engineering in Medicine and Biology, November/December, 776–780. Gomolka, Z., & Orlowski, C. (2000). Knowledge management in creating hybrid systems for environmental protection. Cybernetics and Systems: An International Journal, 31, 507–529. Graham-Jones, P. J., & Mwllor, B. G. (1995). Expert and knowledge- based systems in failure analysis. Expert Systems with Applications, 2, 137–149. Gruer, P., Hilaire, V., Koukam, A., & Cetnarowicz, K. (2002). A formal framework for multi-agent systems analysis and design. Expert Systems with Applications, 23, 349–355. Guerreiro, M. A., Andrietta, S. R., & Maugeri, F. (1997). Expert system for the design of an industrial fermentation plant for the production of alcohol. Journal of Chemical Technological Biotechnology, 68, 163– 170. Hamada, K., Baba, T., Sato, K., & Yufu, M. (1995). Hybridizing a genetic algorithm with rule-based reasoning for production planning. IEEE Expert, 35, 60–67. Hatzilygeroudis, J., & Prentzas, J. (2004a). Integrating (rules, neural networks) and cases for knowledge representation and reasoning in expert systems. Expert Systems with Applications, 27, 63–75. Hatzilygeroudis, J., & Prentzas, J. (2004b). Using a hybrid rule-based approach in developing an intelligent tutoring system with knowledge acquisition and update capabilities. Expert Systems with Applications, 26, 477–492. Higa, K., & Lee, H. G. (1998). A graph-based approach for rule integrity and maintainability in expert system maintenance. Information and Management, 33, 273–285. Ilgun, K., Kemmerer, R. A., & Porras, P. A. (1995). State transition analysis: a rule-based intrusion detection approach. IEEE Transactions on Software Engineering, 21, 181–199. Islam, S., & Chowdhury, N. (2001). A case-based Windows graphic package for the education and training of power system restoration. IEEE Transaction on Power Systems, 16, 181–192. Jamshidi, M. A., Titli, A., Zadeh, L., & Boverie, S. (1997). Applications of fuzzy logic: Towards high machine intelligent quotient systems. Upper Saddle River, NJ: Prentice Hall. Jarmulak, J., Kerckhoffs, E. J. H., & Veen, P. P. (2001). Case-based reasoning for interpretation of data from non-destructive testing. Engineering Applications of Artificial Intelligence, 14, 401–417. Juuso, E. K. (2004). Integration of intelligent systems in development of smart adaptive systems. International Journal of Approximate reason- ing, 35, 307–337. Kawahara, K., Sasaki, H., Kubokawa, J., Asahara, H., & Sugiyama, K. (1998). A proposal of a supporting expert system for outage planning of electronic power facilities retaining high power supply reliability. IEEE Transactions on Power Systems, 13, 1453–1462. Keefe, R. M., & Preece, A. D. (1996). The development, validation and implementation of knowledge-based systems. European Journal of Operational Research, 92, 458–473. Khan, M. K., & Hafiz, N. (1999). Development of an expert system for implementation of ISO quality systems. Total Quality Management, 10, 47–59. Kim, H. S., & Im, Y. T. (1999). An expert system for cold forging process design based on a depth-first search. Journal of Materials Processing Technology, 95, 262–274. Kim, K. H., Park, J. K., Hwang, K. J., & Kim, S. H. (1995). Implementation of hybrid short-term load forecasting system using artificial neural networks and fuzzy expert systems. IEEE Transactions on Power Systems, 10, 1534–1539. Kirkby, S. D. (1996). Integrating a GIS with an expert system to identify and manage dryland salinization. Applied Geography, 16, 289–302. Knauf, R., Gonzalez, A. J., & Abel, T. (2002). A framework for validation of rule-based systems. IEEE Transactions on Systems, Man, and Cybernetics—Part B: Cybernetics, 32, 281–291. Kolonder, J. L. (1994). Case-based reasoning. New York: Morgan Kaufmann. Lababidi, H., & Baker, C. G. J. (2003). Web-based expert system for food dryer selection. Computers and Chemical Engineering, 27, 997–1009. Lau, H. C., Tso, S. K., & Ho, J. K. L. (1998). Development of an intelligent task management system in a manufacturing information network. Expert Systems with Applications, 15, 165–179. Laudon, K. C., & Laudon, J. P. (2002). Essential of management information systems (5th ed). Englewood cliffs, NJ: Prentice Hall. Lee, H. J., Park, D. Y., Ahn, B. S., Park, Y. M., Park, J. K., & Venkata, S. S. (2000). A Fuzzy expert system for the integrated fault diagnosis. IEEE Transactions on Power Delivery, 15, 833–845. Lee, K. C., Han, J. H., Song, Y. U., & Lee, W. J. (1998). A fuzzy logic- driven multiple knowledge integration framework for improving the performance of expert systems. International Journal of Intelligent System in Accounting, Finance, and Management, 7, 213–222. Leon, C., Mejias, M., Luque, J., & Gonzalo, F. (1999). Expert system for the integrated management of a power utility’s communication system. IEEE Transactions on Power Delivery, 14, 1208–1212. Leung, D., & Romagnoli, J. (2000). Dynamic probabilistic model-based expert system for fault diagnosis. Computers and Chemical Engineer- ing, 24, 2473–2492. Expert Systems with Applications, 25 313–330. Leung, R. W. K., Lau, H. C. W., & Kwong, C. K. (2003). An expert system to support the optimization of ion plating process: an OLAP-based fuzzy-cum-GA approach. Li, W., Tasi, Y. P., Tasi, & Chiu, C. L. (2004). The experimental study of the expert system for diagnosing unbalance by ANN and acoustic signals. Journal of Sound and Vibration, 272, 69–83. Liao, S. H. (2001). A knowledge-based architecture for implementing military geographical intelligence system on Intranet. Expert Systems with Applications, 20, 313–324. Liao, T. W. (2003). Classification of welding flaw types with fuzzy expert systems. Expert Systems with Applications, 25, 101–111. Liau, L. C. K., Yang, C. K., & Tasi, M. T. (2004). Expert system of a crude oil distillation unit for process optimization using neural networks. Expert Systems with Applications, 26, 247–255. Lienqueo, M. E., Salgado, J. C., & Asenjo, J. A. (1999). An expert system for selection of protein purification process: experimental validation. Journal of Chemical Technology and Biotechnology, 74, 293–299. Liu, Y., & Schulz, N. N. (2002). Knowledge-based system for distribution system outage locating using comprehensive information. IEEE Transactions on Power Systems, 17, 451–456. Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–1110 DTD 5 ARTICLE IN PRESS Lopez, M. A. A., Flores, C. H., & Garcia, E. G. (2003). An intelligent tutoring system for turbine startup training of electrical power plant operators. Expert Systems with Applications, 24, 95–101. Lu, X., & Simmonds, S. H. (1997). KBES for evaluating R.C. framed buildings using fuzzy sets. Automation in Construction, 6, 121–137. Mahabir, C., Hicks, F. E., & Fayek, A. R. (2003). Application of fuzzy logic to forecast seasonal runoff. Hydrological Processes, 17, 3749–3762. Mahaman, B. D., Harizanis, P., Filis, I., Antonopoulou, E., Yialouris, C. P., & Sideridis, A. B. (2002). A diagnostic expert system for honeybee pests. Computers and Electronics in Agriculture, 36, 17–31. Mahaman, B. D., Passam, H. C., Sideridis, A. B., & Yialouris, C. P. (2003). DIARES-IPM: a diagnostic advisory rule-based expert system for integrated pest management in Solanaceous crop systems. Agricultural Systems, 76, 1119–1135. Mahmoud, M. A. A., & AL-Hammad, A. (1996). An expert system for evaluating and selection of floor finishing materials. Expert Systems with Applications, 10, 281–303. Manohar, P. A., Shivathaya, S. S., & Ferry, M. (1999). Design of an expert system for the optimization of steel compositions and process route. Expert Systems with Applications, 17, 129–134. Marchevsky, A. M., Truong, H., & Tolmachoff, T. (1997). A rule-based expert system for the automatic classification of DNA ‘ploidy’ histograms measured by the CAS 200 image analysis system. Cytometry, 30, 39–46. Martin-Baramera, M., Sancho, J. J., & Sanz, F. (2000). Controlling for change agreement in the validation of medical expert systems with no gold standard: PNEUMON-IA and PENOIR revisited. Computers and Biomedical Research, 33, 380–397. Martinez-Bejar, R., Ibanez-Cruz, F., Compton, P., & Cao, T. M. (2001). An easy-maintenance, reusable approach for building knowledge-based systems: application to landscape assessment. Expert Systems with Applications, 20, 153–162. Matsatsinis, N. F., Doumpos, M., & Zopounidis, C. (1997). Knowledge acquisition and representation for expert systems in the field of financial analysis. Expert Systems with Applications, 12, 247–262. McCoy, M. S., & Levary, R. R. (2000). A rule-based pilot performance model. International Journal of System Science, 31, 713–729. McFadden, F. R., Hoffer, J. A., & Prescott, M. B. (2000). Modern database management (5th ed). New York: Prentice-Hall. Meesad, P., & Yen, G. G. (2003). Combined numerical linguistic knowledge representation and its application to medical diagnosis. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 33, 206–222. Menzies, T. (1997). Object-oriented patterns: lessons from expert systems. Software Practice and Experience, 27, 1457–1478. Mills, K., & Gomaa, H. (2002). Knowledge-based automation of a design method for concurrent systems. IEEE Transactions on Software Engineering, 28, 228–255. Mitra, R. S., & Basu, A. (1997). Knowledge representation in MICKEY: An expert system for designing microprocessor-based systems. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 27, 467–479. Mockler, R. J., Dologite, D. G., & Gartenfeld, M. E. (2000). Talk with the experts: learning management decision-making using CAI. Cybernetics and Systems: An International Journal, 31, 431–464. Montani, S., & Bellazzi, R. (2002). Supporting decisions in medical applications: the knowledge management perspective. International Journal of Medical Informatics, 68, 79–90. Mookerjee, V. S., & Mannino, M. V. (1997). Sequential decision models for expert system optimization. IEEE Transactions on Knowledge and Data Engineering, 9, 675–687. Morrison, M. L., McCliggage, W. G., Price, G. J., Diamond, J., Sheeran, M. R. M., & Mulholland, K. M. (2002). Expert system support using a Bayesian belief network for the classification of endometrial hyperpla- sia. Journal of Pathology, 197, 403–414. Mulvaney, D., & Bristow, C. (1997). A rule-based extension to the CCC language. Software Practice and Experience, 27, 747–761. Ngai, E. W. T., & Wat, F. K. T. (2003). Design and development of a fuzzy expert system for hotel selection. Omega , 275–286. Noh, J. B., Lee, K. C., Kim, J. K., Lee, J. K., & Kim, S. H. (2000). A case- based reasoning approach to cognitive map-driven tacit knowledge management. Expert Systems with Applications, 19, 249–259. Orduna, E., Garces, F., & Handschin, E. (2003). Algorithmic-knowledge- based adaptive coordination in transmission protection. IEEE Trans- actions on Power Delivery, 18, 61–70. Ozyurt, B., & Kandel, A. (1996). A hybrid hierarchical neural network- fuzzy expert system approach to chemical process fault diagnosis. Fuzzy Sets and Systems, 83, 11–25. Padilla-Medina, J. A., & Sanchez-Marin, J. S. (2004). An adaptive fuzzy expert system to evaluate human visual performance. Fuzzy Sets and Systems, 142, 321–334. Pan, J., DeSouza, G. N., & Kak, A. C. (1998). FuzzyShell: a large-scale expert system shell using fuzzy logic for uncertainty reasoning. IEEE Transactions on Fuzzy Systems, 6, 563–581. Park, Y. M., & Lee, K. H. (1995). Application of expert system to power system restoration in local control center. Electrical Power and Energy Systems, 17, 407–415. Peng, C., Xiao, S., Nie, Z., Wang, Z., & Wang, F. (1996). Applying Bayes’ theorem in medical expert systems. IEEE Engineering in Medicine and Biology, May/June, 76–79. Plant, R. E., & Vayssieres, M. P. (2000). Combining expert system and GIS technology to implement a state-transition model of oak woodlands. Computers and Electronics in Agriculture, 27, 71–93. Punal, A., Rodriguez, J., Carrasco, E. F., Roca, E., & Lema, J. M. (2002). Expert system for the on-line diagnosis of anaerobic wastewater treatment. Water Science and Technology, 45, 195–200. Punal, A., Rodriguez, J., Franco, A., Carrasco, E. F., Roca, E., & Lema, J. M. (2001). Advanced monitoring and control of anaerobic wastewater treatment plants: diagnosis and supervision by a fuzzy-based expert system. Water Science and Technology, 43, 191–198. Qian, Y., Li, X., Jiang, Y., & Wen, Y. (2003). An expert system for real- time fault diagnosis of complex chemical processes. Expert Systems with Applications, 24, 425–432. Rahman, S., & Hazim, O. (1996). Load forecasting for multiple sites: development of an expert system-based technique. Electronic Power Systems Research, 39, 161–169. Ramaswamy, M., Sarkar, S., & Chen, Y. S. (1997). Using directed hypergraphs to verify rule-based expert systems. IEEE Transactions on Knowledge and Data Engineering, 9, 221–237. Reidsema, C., & Szczerbicki, E. (2001). A blackboard database model of the design planning process in concurrent engineering. Cybernetics and Systems: An International Journal, 32, 755–774. Reznik, L., & Dabke, K. P. (2004). Measurement models: application of intelligent methods. Measurement, 35, 47–58. Rodionov, S. N., & Martin, J. H. (1999). An expert system-based approach to prediction of year-to-year climatic variations in the north Atlantic region. International Journal of Climatology, 19, 951–974. Ruiz-Sanchez, J., Valencia-Garcia, R., Fernandez-Breis, T., Martinez-Bejar, R., & Compton, P. (2003). An approach for incremental knowledge acquisition from text. Expert Systems with Applications, 25, 77–86. Sabourin, L., & Villeneuve, F. (1996). OMEGA, an expert CAPP system. Advances in Engineering Software, 25, 51–59. Saunders, S. M., Pascoe, S., Johnson, A. P., Pilling, M. J., & Jenkin, M. E. (2003). Development and preliminary test results of expert system for the automatic generation of tropospheric VOC degradation mechan- isms. Atmospheric Environment, 37, 1723–1735. Sen, M. D. L., Minambres, J. J., Garrido, A. J., Almansa, A., & Soto, J. C. (2004). Basic theoretical results for expert systems, application to the supervision of adaptation transients in planar robots. Artificial intelligence, 152, 173–211. Sendelj, R., & Devedzic, V. (2004). Fuzzy systems based on component software. Fuzzy Sets and Systems, 141, 487–504. Shu-Hsien Liao / Expert Systems with Applications xx (2004) 1–11 11 DTD 5 ARTICLE IN PRESS Shaalan, K., EI-Badry, M., & Rafea, A. (2004a). A multiagent approach for diagnostic expert systems via the internet. Expert Systems with Applications, 27, 1–10. Shaalan, K., Rizk, M., Abdelhamid, Y., & Bahgat, R. (2004b). An expert system for the best weight distribution on ferryboats. Expert Systems with Applications, 26, 397–411. Shu, C., & Burn, D. H. (2004). Homogeneous pooling group delineation for flood frequency analysis using a fuzzy expert system with genetic enhancement. Journal of Hydrology, 291, 132–149. Soh, L. K., Tsatsoulis, C., Gineris, D., & Bertoia, C. (2004). ARKTOS: An intelligent system for SAR sea ice image classification. IEEE Transactions on Geoscience and Remote Sensing, 42, 229–248. Soliman, A., Rizzoni, G., & Kim, Y. W. (1999). Diagnosis of an automotive emission control system using fuzzy inference. Control Engineering Practice, 7, 209–216. Takahashi, M., Oono, J. I., & Saitog, K. (1995). Manufacturing process design by CBR with knowledge ware. IEEE Expert, December, 74–80. Takaoka, Y., & Mizoguchi, R. (1996). Identification of ontologies to reuse knowledge for substation fault recovery support system. Decision Support Systems, 18, 3–21. Thomson, A. J., & Willoughby, I. (2004). A web-based expert system for advising on herbicide use in Great Britain. Computers and Electronic in Agriculture, 42, 43–49. Thukaram, B. D., & Parthasarathy, K. (1997). An expert system for power system voltage stability improvement. Electrical Power and Energy Systems, 19, 385–392. Tu, S. W., Eriksson, H., Gennari, H., Shahar, Y., & Musen, M. A. (1995). Ontology-based configuration of problem-solving methods and gener- ation of knowledge-acquisition tools: application of PROTÉGÉ-II to protocol-based decision support. Artificial Intelligence in Medicine, 7, 257–289. Tucho, R., Sierra, J. M., Fernandez, J. E., Vijande, R., & Moris, G. (2003). Expert tutoring system for teaching mechanical engineering. Expert Systems with Applications, 24, 415–424. Tunez, S., Aguila, I. M., & Marin, R. (2001). An expertise model for therapy planning using abductive reasoning. Cybernetics and Systems: An International Journal, 32, 829–849. Turban, E., & Aronson, J. E. (2001). Decision support systems and intelligent systems, sixth Edition (6th ed). Hong Kong: Prentice International Hall. Valenzuela, L. M., Bentley, J. M., & Lorenz, R. D. (2004). Expert system for integrated control and supervision of dry-end sections of paper machines. IEEE Transactions on Industry Applications, 40, 680–691. Vegh, J. (2003). The ‘carbon aontamination’ rule set implemented in an ‘embedded expert system’. Journal of Electron Spectroscopy and Related Phenomena, 133, 87–101. Vlahavas, I., Stamelos, I., Refanidis, I., & Tsoukias, A. (1999). ESSE: an expert system for software evaluation. Knowledge-Based Systems, 12, 183–197. Volberda, H. W., & Rutges, A. (1999). FARSYS: a knowledge-based system for managing strategic change. Decision Support Systems, 26, 99–123. Vranes, S., & Stanojevic, M. (1995). Integrated multiple paradigms within the blackboard framework. IEEE Transactions on Software Engineer- ing, 21, 244–262. Vranes, S., & Stanojevic, M. (1999). Design knowledge representation in Prolog/Rex. Engineering Applications of Artificial Intelligence, 12, 221–228. Wang, X., Qu, H., Liu, P., & Cheng, Y. (2004). A self-learning expert system for diagnosis in traditional Chinese medicine. Expert Systems with Applications, 26, 557–566. Wasiewicz, P., Janczak, T., Mulawka, J. J., & Plucienniczak, A. (2000). The inference based on molecular computing. Cybernetics and Systems: An International Journal, 31, 283–315. Wei, M. S., & Weber, F. (1996). An expert system for waste management. Journal of Environmental Management, 46, 345–358. Wiig, K. M. (1994). Knowledge management, the central management focus for intelligent-acting organization. Arlington: Schema Press. Wong, S. V., & Hamouda, A. M. S. (2002). A fuzzy logic based expert system for machinability data-on-demand on the Internet. Journal of Materials Processing Technology, 124, 57–66. Wong, S. V., & Hamouda, A. M. S. (2003). The development of an online knowledge-based expert system for machinability data selection. Knowledge-based Systems, 16, 215–229. Wu, T. P., & Chen, S. M. (1999). A new method for constructing membership functions and fuzzy rules from training examples. IEEE Transactions on Systems, Man, and Cybernetics—Part B: Cybernetics, 29, 25–37. Wu, C. H., & Lee, S. J. (1997). Enhanced high-level Petri nets with multiple colors for knowledge verification/validation of rule-based expert systems. IEEE Transactions on Systems, Man, and Cybernetics—Part B: Cybernetics, 27, 760–773. Xirogiannis, G., Stefanou, J., & Glykas, M. (2004). A fuzzy cognitive map approach to support urban design. Expert Systems with Applications, 26, 257–268. Yan, H., Jiang, Y., Zheng, J., Fu, B., Xiao, S., & Peng, C. (2004). The internet-based knowledge acquisition and management method to construct large-scale distributed medical expert system. Computer Methods and Programs in Biomedicine, 74, 1–10. Yang, B. S., Han, T., & Kim, Y. S. (2004). Integration of ART-Kohoene neural network and case-based reasoning for intelligent fault diagnosis Yu, R., Iung, B., & Panetto, H. (2003). A multi-agents based E-maintenance system with case-based reasoning decision support. Engineering Applications of artificial Intelligence, 16, 321–333. Expert Systems with Applications, 26, 387–395. Zha, X., & Lim, S. Y. E. (2000). Assembly/disassembly task planning and simulation using expert Petri nets. International Journal of Production Research, 15, 3639–3676. Zhou, Q., Huang, G. H., & Chan, C. W. (2004). Development of an intelligent decision support system for air pollution control at coal-fired power plants. Expert Systems with Applications, 26, 335–356. Zhu, A. X., Band, L. E., Dutton, B., & Nimlos, T. J. (1996). Automated soil inference under fuzzy logic. Ecological Modeling, 90, 123–145. Zupan, B., & Cheng, M. K. (1998). Optimization of rule-based systems using state space graphs. IEEE Transactions on Knowledge and Data Engineering, 10, 238–254. Expert system methodologies and applications-a decade review from 1995 to 2004 Introduction Rule-based systems and their applications Knowledge-based systems and their applications Neural networks and their applications Fuzzy expert systems and their applications Object-oriented methodology and their applications Case-based reasoning and their applications Modeling and their applications System architecture and their applications Intelligent agents and their applications Ontology and their applications Database methodology and their applications Discussions, limitations, and suggestions Discussions Limitations Suggestions Conclusions References 
work_fm6xz6iburcpxgxdmde7yho7jq	----	Mapping Preferences into Euclidean Space Oscar Luacesa, Jorge Dı́eza,∗, Thorsten Joachimsb, Antonio Bahamondea,b aUniversidad de Oviedo, Artificial Intelligence Center, Gijón, Asturias, Spain bCornell University, Department of Computer Science, Ithaca, NY, USA Abstract Understanding and modeling human preferences is one of the key problems in applications ranging from marketing to automated recommendation. In this paper, we focus on learning and analyzing the preferences of consumers regarding food products. In particular, we explore machine learning methods that embed consumers and products in an Euclidean space such that their relationship to each other models consumer preferences. In addition to pre- dicting preferences that were not explicitly stated, the Euclidean embedding enables visualization and clustering to understand the overall structure of a population of consumers and their preferences regarding the set of products. Notice that consumers’ clusters are market segments, and products clusters can be seen as groups of similar items with respect to consumer tastes. We explore two types of Euclidean embedding of preferences, one based on inner products and one based on distances. Using a real world dataset about con- sumers of beef meat, we find that both embeddings produce more accurate models than a tensorial approach that uses a SVM to learn preferences. The ∗Corresponding author: Tel: +34 985 182 588 Email addresses: oluaces@uniovi.es (Oscar Luaces), jdiez@uniovi.es (Jorge Dı́ez), tj@cs.cornell.edu (Thorsten Joachims), abahamonde@uniovi.es (Antonio Bahamonde) Preprint submitted to Expert Systems with Applications July 15, 2015 reason is that the number of parameters to learned in embeddings can be considerably lower than in the tensorial approach. Furthermore, we demon- strate that the visualization of the learned embeddings provides interesting insights into the structure of the consumer and product space, and that it provides a method for qualitatively explaining consumer preferences. Addi- tionally, it is important to emphasize that the approach presented here is flexible enough to allow its use with different levels of knowledge about con- sumers or products; therefore the application field is very wide to grasp an accurate understanding of consumers’ preferences. Keywords: Preference Learning, matrix factorization, graphical representations, learning to order, visualization 1. Introduction In 1927, Thurstone (1927) presented a law of comparative judgment to approach to qualitative comparisons from a psychological point of view. Ac- cording to this law, users react differently to each item, and they identify the degree of compatibility with the quality to be compared. The difference of these degrees define the discriminal process between pairs of items. From a Machine Learning perspective, there are two main ways to ap- proach preferences. They can be represented by a real-valued function, which assigns a utility value to the object, or by a preference relation, which com- pares two different items; see (Hüllermeier and Fürnkranz, 2013). In the first approach, the degrees of compatibility (usually called utilities) are considered as the target output that can be learned by means of ordinal regression or 2 classification methods. This is a suitable approach when it is possible to assume that users assign those utilities depending exclusively on the item being assessed. However, in some cases there is a batch effect; that is, the assessment of an item depends on the batch of items included in the same comparison. When this is the case, it is more suitable to use the second approach and consider preferences as a binary relation. The goal here is to learn the relative ordering of items given by the user instead of the utility itself. This is the approach used by Herbrich et al. (1999), Joachims (2002), Bahamonde et al. (2007) and Rendle et al. (2009). In this paper we are concerned with learning preferences expressed by consumers of a kind of products. Consumers typically assign the utilities only as a way to express relative preferences instead of absolute values. Therefore, the datasets that we are going to use are collections of pairwise comparisons, called preference judgments, that represent the discriminant process of one consumer between two products. In addition to modeling the products, we also explicitly model individual consumers. To represent the interaction of consumers and products we propose several factorization approaches and a tensorial approach. A desirable property of factorization approaches is that they entail an embedding of both consumers and products in a common Euclidean space where the utility can be expressed in geometric (or graphical) terms. The con- sequence is that, as a side effect, learning preferences with these approaches provides a setting for visualization of clusters in both consumers and prod- ucts. In the following sections we introduce a common framework to learn fac- 3 torizations and SVM tensorial models. The purpose is to discuss the char- acteristics of these approaches according only to their mathematical formu- lations. In all cases, the objects involved in preferences (consumers and products) can be represented by a combination of a binary identification code or by vectors of feature-values. Notice that, for instance, in food products, the features of consumers or the products are not always available. Moreover, if a food industry is planing to launch a new product there is a reduced set of options that they want to test, and they can be just represented by an identification code. On the other hand, when there is a selected panel of singular consumers, they can be unequivocally identified with a label. After the formal presentation of the methods, we show the results of an exhaustive experimentation using a real world dataset of consumers of beef meat. First we compare the results of factorization methods with those achieved by SVM. In the datasets used in this paper, factorization methods outperform SVM, probably this is a general fact. One reason is that the number of parameters to be learned is smaller in the factorization approaches. Additionally, the formal models learned in all cases are quite similar and they all capture the possible interactions of both the features of consumers and products. The contributions of the paper are the following: (i) it presents a com- mon framework for different approaches to learn preferences using matrix factorization, (ii) the paper illustrates the formal presentation with a real world problem of consumers and (food) products, (iii) the last section shows a favorable comparison of factorization and SVM tensorial approaches, (iv) 4 the paper emphasizes the graphical and geometrical possibilities of the fac- torization methods as a tool to analyze the complex relationships of users and items. 2. Related Work Learning preferences has been studied with different approaches. A recent Special Issue of the Machine Learning Journal (Hüllermeier and Fürnkranz, 2013) includes some interesting approaches and application fields. From a conceptual point of view, the aim is to learn an ordering relation from some pairwise comparisons. Thus, the learning task can be read as a binary classification task using SVMs, see for instance (Herbrich et al., 1999; Joachims, 2002). In this paper we adopt a more general strategy, we explicitly optimize a loss function with regularization. For instance, it is possible to use the lo- gistic loss as in the learner proposed by Rendle et al. (2009). The algorithm presented in that paper was derived from a Bayesian analysis of the ranking problem of a user for a set of items. The algorithm is called Bayesian Person- alized Ranking (BPR) and was devised as a method to solve the maximum posterior estimation. We present a general setting that includes, at the same time, a factoriza- tion framework and a tensorial approach: a SVM that uses tensor products to model the interactions of consumer and item representations. Tensorial representations were already used, for instance, in (Basilico and Hofmann, 2004; Rendle and Schmidt-Thieme, 2010; Pahikkala et al., 2012). We report a comparison between factorization methods and this SVM approach in the 5 experimental section using a real dataset. Factorization algorithms were previously used in recommender systems in some of the best ranked systems of the Netflix prize; see for instance (Koren et al., 2009). Many other papers propose matrix factorization for solving specific problems in recommender systems; see for instance (Ocepek et al., 2015). A software library to do factorization machines with a wide variety of options is presented in (Rendle, 2012). Other similar implementations can be found in (Chen et al., 2011; Agarwal and Chen, 2009; Bayer, 2015). Let us remark that the goal of this paper is not to present another implemen- tation. We use a quite straightforward SGD (Stochastic Gradient Descent) implementation whose main advantage is that it is the same for 3 different approaches. We want to underscore that the differences arise only from the formulation of the approach, but not from any implementation issue. Ad- ditionally, we are interested in discussing the necessity of using the features of the items and consumers involved in the preferences judgments. This is a central point in learning preferences and the approach presented in this paper is quite suitable for this purpose. We may use all the convenient information in a real word application. Another interesting use of factorizations is presented in (Weston et al., 2010, 2011). The target is information retrieval, and so the aim was to optimize the ranking of labels attached to queries (images or music). In this case the output is an ordered set of labels. As was said in the introduction, we are concerned with the graphical properties of the model learned from preferences. Both consumers and items 6 are located in an Euclidean space with one specific aim. This is the case, for instance in (Moore et al., 2012; Chen et al., 2012) where the purpose was to build an embedding of songs. The proximity was learned from a collection of playlists. Once the map is built, playlists are generated using the relative distances of songs. To model proximity, the authors use Gaussian distributions, and the same tool was used also to add tags to the songs. In this paper we do not assume any distribution of the representation of data in the Euclidean space. Another paper related to the work reported here is (Xing et al., 2002). Here, the authors learn a metric for Euclidean points that represents sim- ilarities and dissimilarities. The metric is given by a positive semi-definite matrix that is the solution of a convex optimization problem. In our case the factorization eases the learning process since the number of parameters to be estimated may be significantly fewer. Moreover, the preference learn- ing tasks only have dissimilarity examples; there are no similarity cases to guide the induction. On the other hand, in (Xing et al., 2002) there is no reduction of dimensionality neither visualization purposes: the objective is to find clusters. A quite similar approach can be found in (Parameswaran and Weinberger, 2010), in this case to learn a metric for Multi-Task Learning. To learn metrics is also the aim of Peltonen et al. (2003). The purpose is to reduce the dimensionality from visualization. The source data are collections of labelled data for classification tasks. The proposals are extensions of the so-called Self-Organizing Maps (SOM). However, notice that our purpose is not only to learn a metric, but to learn preferences while represent in a metric space both consumers and products. 7 Finally, the visualization method presented here is in fact a supervised learning algorithm, like supervised PCA (Koren and Carmel, 2004; Yu et al., 2006; Du et al., 2015) for instance. The difference is that our approach explicitly incorporates the loss function and the definition of similarity that we want to obtain at the end of the process. And, of course, the method presented here is devised for learning preferences. 3. Formal Framework Let us consider the following dataset D = {(x1,f(x1)), . . . , (xn,f(xn))}. (1) Here we assume that f is an unknown real function on the space from where inputs x ∈ Rm are drawn. The aim is to find a new function g of input data x, that depends also on some parameters θ, such that the variations of f can be predicted by the variations of g. The function g will have an analytical definition that makes it straightforward to compute g on any input. In symbols, the aim of g is to maximize the probability Pr ( f(x) > f(x′) ⇐⇒ g(x,θ) > g(x′,θ) ) . (2) In the following, as usual in this context, we will call g a utility function. To learn g we define the following ordering induced by D Dor = { ( xi,xj; [[f(xi) > f(xj)]] ) : i,j = 1, . . . ,n}. (3) The symbol [[p]] stands for the value 1 when the predicate p is true, and −1 otherwise. In the next subsections we present an approach to learn g from this binary classification task. 8 Formally, the learning process of the parameters θ of g starts with the dataset Dor (Eq. 3). Soon we shall see that we may use only the examples of the positive class, D+or = {(xi,xj) :f(xi) > f(xj), i,j = 1, . . . ,n}. (4) Notice that, in fact, we do not need the function f in our approach. In practice, f is hidden and we do not have access to it; otherwise, the dataset (Eq. 1) could be seen as a regression task. Roughly speaking, the dataset D+or is the set of pairs where an explicit ordering has been registered. Each pair is formed by the better and the worse objects. Usually these pairs are called preference judgments, see (Joachims, 2002). We adopted a margin maximization approach, detailed in the next sec- tion, to learn from such a dataset in order to include the hypothesis learned by SVMs as in (Herbrich et al., 1999; Joachims, 2002; Bahamonde et al., 2007; Basilico and Hofmann, 2004; del Coz et al., 2005; Dı́ez et al., 2005, 2006). This could also be solved using a probabilistic approach. 4. Maximum Margin Approach As usual, we assume that all these examples are independently and iden- tically drawn (i.i.d.) from an unknown distribution. Thus, using a maximum margin approach, the parameters θ should minimize Loss(θ,D+or) = ∑ (xi,xj)∈D+or max ( 0, 1 −g(xi,θ) + g(xj,θ) ) . (5) Following Rendle et al. (2009), margin maximization can be done using an SGD algorithm (Robbins and Monro, 1951) with a regularization term 9 for the parameter θ, r(θ). Thus, the optimal value, θ∗ is given by θ∗ = argmin θ Loss(θ) + νr(θ) (6) The idea is to ensure that the difference of the utilities in a preference judgment is at least 1. g(xi,θ) −g(xj,θ) ≥ 1, (xi,xj) ∈ D+or Of course, this is equivalent to g(xj,θ) −g(xi,θ) < −1, (xi,xj) ∈ D−or where D−or is the subset of Dor (Eq. 3) with negative classes. The consequence is that we may get rid of the negative part since it is redundant. The corresponding optimization with this loss function can be solved with Algorithm 1 that implements this approach using an L2 regularization term. The updating step due to (xi,xj) is done by: θ ← θ −γ [∂(Loss(θ))ij ∂θ + ν ∂r(θ) ∂θ ] . (7) That is, θ ← θ + γ [∂g(xi,θ) ∂θ − ∂g(xj,θ) ∂θ −ν ∂r(θ) ∂θ ] (8) if 1 −g(xi,θ) + g(xj,θ) > 0. Additionally, to ensure numerical stability, following Weston et al. (2010, 2011), we use a parameter R (a radius) such that the size of θ is always smaller or equal than R. 10 Algorithm 1 SGD algorithm to learn a utility function that maximizes the margin as defined in (Eq. 5) using an L2 regularization Input: D+or; {(Eq. 4)} Input: γ > 0 {learning rate}; ν > 0 {regularization parameter}; Input: R > 0 {radius}; assign random values to the components of θ; repeat fetch random (xbetter,xworse) ∈ D+or; if 1 > g(xbetter,θ) −g(xworse,θ) then θ ← θ + γ [ ∂g(xbetter,θ) ∂θ − ∂g(xworse,θ) ∂θ −ν∂r(θ) ∂θ ] ; if ∥∥θ∥∥ > R then θ ← R∥∥θ∥∥θ; end if end if until stop criterion 5. Factorization and Tensorial Approaches In the last section, inputs were described by a generic vector x and the aim was to emphasize the ordering of these vectors according to f values. Now we are going to get into the structure of inputs as the concatenation of two different vectors, the representation of consumers and items or products (we prefer products to use p instead of i for short in equations). Thus, in the following, we are going to assume that each input data can be split in two parts: x = (c,p). 11 In this section, we introduce three possible definitions of the utility func- tion g. They have in common that rest on the interaction of the vectorial representation of consumers and products. 5.1. Mapping Consumers and Products: Matrix Factorization In this subsection, we are going to consider an embedding of both con- sumers and products in a common Euclidean space. Then, the function g (Eq. 2) will be defined in terms of the mappings in the common space. We assume that consumers are described by vectors in a Euclidean space of dimension |Con|, while products are given by vectors with |Prod| compo- nents. We are going to represent them in a common space of dimension k using two linear maps given respectively by matrices W and V . R|Con| −→ Rk, c W c, (9) R|Prod| −→ Rk, p V p. (10) Let us remark that, as usual, we are considering vectors as column matrices. In this context, the parameter θ to be learned is the set of matrices W , V . We are trying to solve the optimization problem W∗,V ∗ = argmin W,V ( Loss(W ,V ,D+or) + νr(W ) + νr(V ) ) . (11) Notice that there are different options to define the interaction of con- sumers and products. Next, we present two of them. 12 5.1.1. Inner Products The first alternative is to formalize the interactions by the following inner product of the mappings of consumers and products in Rk. gin(x) = gin(c,p) = 〈W c,V p〉 (12) =(W c)TV p = cTW TV p = |Con|∑ r=1 |Prod|∑ s=1 ( W TV ) rs ( cpT ) rs =(V p)TW c = |Con|∑ r=1 |Prod|∑ s=1 ( V TW ) rs ( pcT ) rs . It is interesting to realize that the utility function gin is given by a linear combination of all products formed by one component from the consumer description, and one component of the description of the product. The type of equation is different if we add one constant component (with value 1 for instance) to the vectorial representation of consumers and prod- ucts; that is, cT ← [cT 1]; pT ← [pT 1]. (13) In this case, the utility function can be thought as follows, gin(c, p) = ∑ r,s αrscrps + ∑ r βrcr + ∑ s δsps + τ, (14) for some real coefficients αr,s,βr,δs and τ. That is, the utility is a polynomial of degree 2 where the monomials of degree 2 are always built by the product of one component of the representation of consumers and other from the products. If we compare the equations (Eq. 12, 14), we appreciate that the coeffi- cients of the polynomial that defines gin are factorized in two matrices, as was mentioned in the introduction. 13 The partial derivatives needed to implement this approach in Algorithm 1 are the following: ∂gin(c,p) ∂W = ∂(V p)T (W c) ∂W = V pcT , ∂gin(c,p) ∂V = ∂(W c)T (V p) ∂V = W cpT . On the other hand, we use the square of the Frobenius norm as the regularization summand. r(W ) = ∥∥W∥∥2 F = Tr(W T W ), r(V ) = ∥∥V ∥∥2 F = Tr(V T V ). Therefore, the regularization derivatives are ∂Tr(W TW ) ∂W = 2W , ∂Tr(V TV ) ∂V = 2V . (15) The Frobenius norm of matrices is also used to measure the size of the pa- rameters in Algorithm 1. 5.1.2. Euclidean Closeness The second option that we explore for defining g is the interaction given by the closeness. In symbols, we define gcl(c,p) = − ∥∥W c−V p∥∥2 = − ∥∥W c∥∥2 −∥∥V p∥∥2 + 2〈W c, V p〉 = − ∥∥W c∥∥2 −∥∥V p∥∥2 + 2gin(c,p). (16) Notice that, comparing with the utility function defined in (Eq. 12), now we add more summands to the equation. The new utility, gcl, includes the 14 weighted sum of all monomials of degree 2 formed with variables taken from the description of consumers (c) or products (p). Of course, to guarantee this, we need to add one constant component (with value 1 for instance) to the vectorial representation of consumers and products, see (Eq. 13). The derivatives needed to implement the learning algorithm are the fol- lowing: ∂gcl(c,p) ∂W = − ∂(W c)T (W c) ∂W +2 ∂gin(c,p) ∂W = −W (2ccT )+2V pcT, ∂gcl(c,p) ∂V = − ∂(V p)T (V p) ∂V +2 ∂gin(c,p) ∂V = −V (2ddT )+2W cpT. We use the same regularization than in the case of the utility defined in terms of the inner product. The advantage of this definition of g is that the visual semantics is more easy to appreciate. The Euclidean representation of consumers and products are closed or further according with the preferences. The inner product is a simpler equation, but it is harder to visualize. 5.2. Tensor Product The full description of the utility functions presented in the last subsection (Eq. 14) can be seen as a particular case of a linear function in the tensor product of consumers and products. In symbols, g⊗(c, p) = 〈w,c⊗p〉, (17) where w is a vector in the Euclidean space of dimension |Con|×|Prod|. If we use a pair of indexes to refer to the components of w, the previous equation 15 can be written as g⊗(c, p) = ∑ r,s wrscrps. (18) Once more, this expression includes all the terms of (Eq. 14) provided that a constant component is included in vectors c and p. It is important to emphasize that the number of parameters to be learned in this approach is considerably more than in the factorization cases provided that the value of k (the dimension of the Euclidean space is small). Again, we may learn these parameters, the components of w, using the Algorithm 1. For this purpose, we only need to compute the derivative ∂g⊗(c, p) ∂w = c⊗p, (19) and the derivative of the regularization summand, that is given by ∂r(w) ∂w = 2w. (20) The Algorithm 1 learns the parameter w of (Eq. 17) using a SGD. The size of the parameter is the Euclidean norm of w. This approach is then equivalent to a Support Vector Machine (SVM) used to learn to rank. In the experimental section we will denote this learning algorithm as SVM⊗. 6. Experimental Results In this section we report a set of experiments carried out to show the performance of the proposals of this paper. First we present some imple- mentation details of Algorithm 1. Then we introduce the datasets used in the experiments to report the accuracy obtained with each utility function (let us recall that in all cases the learning algorithm is the same). Finally, 16 we show some graphical representations obtained as side effect of the learn- ing process to illustrate the visualization possibilities of the factorization approaches when used to learn consumer preferences. 6.1. Implementation Details The implementation of the Algorithm 1 was done using Pegasos as a model, see (Shalev-Shwartz et al., 2011). Thus, the learning rate follows the equation γ = γ0 1 + γs(n− 1) . To avoid many parameters to be adjusted, we fixed γ0 = 1, and γs = 0.01. The radius (Section 4) was also fixed: R = 1. As usual, n is the ordinal of the iteration. To update the model learned by the algorithm we used a mini batch strategy, averaging the updates every time that 10% of the training examples were processed. The only adjustable parameter in the algorithm was the regularization parameter. We made an internal grid search to determine the best option in the set ν ∈{10i : i = −1, . . . ,−10} using a 2-fold cross validation repeated 3 times on the training set. Finally, the algorithm stops when the size of the difference of parameter θ in two consecutive iterations is smaller than 10−6 or the number of iterations is 5000. 17 Task |D+or| Acceptability 3084 Flavor 3080 Tenderness 3313 Table 1: Sizes of the datasets used in the experiments. |D+or| stands for the number of Preference Judgments (Eq. 4). The number of consumers is 392 and there are 307 items 6.2. Datasets The dataset used in this paper comes from a study carried out to deter- mine the features that entail consumer acceptance of beef meat from seven Spanish breeds (Gil et al., 2001; Sañudo et al., 2004; del Coz et al., 2005; Dı́ez et al., 2005, 2006; Bahamonde et al., 2007). Each piece of meat was de- scribed by: the weight of the animal, ageing time, breed, 6 physical features describing its texture and 12 sensory characteristics rated by 11 different experts (132 ratings). The dataset has 307 different items. In each testing session, 4 or 5 pieces of meat were tested and a group of consumers were asked to rate (on a scale of 1 to 10 points) three different aspects: tenderness, flavor and overall acceptance. The number of consumers involved in this panel was 392. The features of consumers are just sex, age and job. The preferences expressed by consumers were represented in a dataset of preference judgments like D+or where each input x is the concatenation of the feature description of the item and of the user. We only considered pairs where the preferences of the consumer were strictly different for two items. Thus, in each dataset the number of preference judgments is slightly different. 18 Table 1 reports the number of preference judgments for each learning task. The data was preprocessed. The discrete features were binarized in the whole dataset. On the other hand, the continuous features were standardized in each training set; the mean and standard deviation of training data were used to standardize the test set. 6.3. Results and Discussion To estimate the accuracy of the utility functions learned, we used cross validation in the D+or versions of acceptability, flavor and tenderness. In addition to feature descriptions of consumers and items, we added a binary identification of them. That is to say, each object (consumer or item) includes in its description a vector of dimension the number of objects; in that vector all components are 0 but the one with index the ordinal of the object that has value 1. To check the role played by these identifiers, we considered two different versions of each dataset: with and without identifiers. In preference learning, sometimes we do not have any feature description of items or consumers, then we can only use such identifiers. To ease of reading, let us put a simple example. If we have only 3 con- sumers, their representations can be the following. With id codes the con- sumers are presented by consumer1 = (1, 0, 0,sex1,age1,job1), consumer2 = (0, 1, 0,sex2,age2,job2), consumer3 = (0, 0, 1,sex3,age3,job3). 19 gcl gin Dataset SVM⊗ 2 10 100 2 10 100 Acceptability no ID 28.2 28.6 26.4 25.4 31.7 26.8 26.5 with ID 21.9 26.9 18.7 16.2 27.5 18.8 16.1 Flavor no ID 30.7 35.2 29.9 29.7 36.2 28.5 28.6 with ID 24.6 32.9 21.4 19.9 31.5 20.4 18.1 Tenderness no ID 25.5 26.4 24.3 25.0 28.5 23.9 23.9 with ID 21.1 24.3 17.5 15.6 26.4 18.2 16.2 Table 2: Percentages of misclassified preference judgments estimated with 10-fold cross validation using internal grid search for the parameters of the learners. Columns labeled by 2, 10 and 100 report the scores of factorizations obtained with that value of k. The testing was carried out in each fold while training was performed in the remaining 9 Without identification codes, we drop the first three binary components and each consumer will represented only by their sex, age and job. Of course, analogous representations can be used for products. The first block of experiments used a 10-fold cross validation. Systems were trained using 9 folds and the test was performed on the remaining fold. The scores are reported in Table 2. We observe that the performance of the SVM that uses the tensor product (SVM⊗) is worse than the performance of the factorization methods (gin and 20 gcl gin Dataset 2 10 100 2 10 100 Acceptability no ID 39.0 38.3 38.7 37.9 38.0 38.0 with ID 38.7 37.4 37.4 37.7 37.3 36.6 Flavor no ID 43.5 43.4 43.0 43.7 43.1 42.5 with ID 43.1 42.6 42.9 43.2 41.3 40.3 Tenderness no ID 36.5 35.0 35.2 35.3 34.8 35.2 with ID 35.9 34.7 34.1 35.5 34.6 34.5 Table 3: Percentages of misclassified preference judgments estimated with 10-fold cross validation using internal grid search for the parameters of the learners. Columns labeled by 2, 10 and 100 report the scores of factorizations obtained with that value of k. The training was carried out in each fold while testing was performed in the remaining 9 gcl) that are really quite similar. Additionally, the influence of the dimension of the Euclidean space k (Eq. 9) is dramatic in factorization systems. Greater values of k provide better results. In all cases, the scores of the tensorial version are somewhere in the middle of the factorization scores with k = 2 and k = 10. In all cases the use of identifiers improves considerably the scores. In some case the difference is 10 points better with identifiers than without them. The reason is that some items or consumers in test sets were also 21 known in training stage. But this is the case in many sensory data studies. Sometimes the number of options that a food industry is considering for a new product is the whole set of items both in training and in test. On the other hand, if we want to model the assessments of a selected panel of consumers, they must be present in training and in test examples. In the experiments reported in Table 2, 90% of preference judgments are in the respective training set; therefore, most of the consumers and items in each respective test set appear in the training set too. To check the effect of the appearance of already known objects, and also to check the effect of the number of training examples, we performed two additional experiments. However, in this case we used only factorization systems since the perfor- mance of the tensorial systems was very poor. In this way, first we report the experiments carried out with 10-fold cross validation by using each fold as training set and the remaining 9 as test. The results are shown in Table 3. The results are substantially worse, as the number of training examples is very small. Nevertheless, the impact of the identifiers of consumers and items is beneficial in all cases but one, although the increase in accuracy is smaller than in the experiments reported in Table 2. On the other hand, again we realize that higher values of k give rise to better performance. Finally, in Table 4 we report an intermediate setting. Now we use only two folds; therefore, half of the items and consumers in the test set already appeared in the training set. As expected, the results are better than those of Table 3, but worse than the scores shown in Table 2. In this case, for k = 100, the error is mostly below 25% with identifiers, and around 30% without identifiers. The role of k is again of paramount importance. 22 gcl gin Dataset 2 10 100 2 10 100 Acceptability no ID 33.0 30.9 31.5 31.2 31.5 31.6 with ID 30.6 26.1 24.7 31.3 28.4 23.6 Flavor no ID 37.1 33.3 33.4 37.6 33.9 32.1 with ID 34.9 29.1 28.7 36.3 28.2 24.5 Tenderness no ID 29.4 27.8 28.3 28.9 29.1 28.9 with ID 27.1 25.7 24.1 28.3 24.4 23.7 Table 4: Percentages of misclassified preference judgments estimated with 2-fold cross validation using internal grid search for the parameters of the learners. Columns labeled by 2, 10 and 100 report the scores of factorizations obtained with that value of k 6.4. Visualization of Preferences The graphical possibilities of factorization methods, in addition to good prediction scores, provide also some interesting applications. In particular, visualization is very natural when the Euclidean space has up to 3 dimen- sions. But another application is clustering in order to find groups of con- sumers with similar tastes or collections of items with similar appreciations by consumers. In this subsection we illustrate these applications in sensory data analy- sis. To create the subsequent visualizations we used all available data with 23 identities, applying Algorithm 1 with gcl and k = 2. The idea is to obtain pictures where the proximity of one item and one consumer is the utility that represents the preference. The resubstitution error in acceptability is 15.27%, in flavor is 18.47%, and in tenderness is 13.91%. In the graphs, the small dots represent consumers located in R2 according to their ratings of acceptability (Figure 1), flavor (Figure 2) and tenderness (Figure 3) respectively. According to the literature about sensory preferences of beef meat, (Gil et al., 2001; Sañudo et al., 2004; del Coz et al., 2005; Dı́ez et al., 2005, 2006; Bahamonde et al., 2007), the most important features that explain the preferences of consumers are ageing and intramuscular fat (intrafat for short). These are discrete features. Ageing has 3 different values: 1, 7 and 21 days. And intrafat was discretized to obtain 3 options: low, medium and high. Thus, in the same graph of consumers, we represented the average item with each value of these important features. This is a kind of tag in the sense used in (Chen et al., 2012; Moore et al., 2012) of the feature values. The left part of Figure 1 is a Voronoi diagram of the space where seeds are the centroids of the Euclidean representation of the possible values of ageing. The lowest ageing values, 1 and 7 have centroids very close, and what it is really interesting is the split between the low (1 or 7 days) and high (21 days). In the right hand side of the figure, the centroids of items with medium or high intrafat are near and provide a clear split with consumers that prefer low intrafat values. Notice that the split due to ageing and intrafat are almost 24 1 7 21 −1.2 −1.0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 −2.0 −1.8 −1.6 −1.4 −1.2 −1.0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 medium low high −1.2 −1.0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 −2.0 −1.8 −1.6 −1.4 −1.2 −1.0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 Figure 1: Consumers represented according to their ratings in acceptability. Voronoi diagram whose seeds are the centroids of items with different values of ageing (left) and intrafat (right) the same. That is to say, consumers that like meat with 21 days of ageing also prefer meat with low intrafat values. In Figure 2 the position of consumers is different with respect to the previous picture, now the feature rated by consumers is flavor. In this case the relevancy of ageing (left hand side of the figure) is clear. Consumers mostly prefer the flavor of meat after 21 days of ageing. Notice that the relative position of the centroids is increasing from left to right. According to intrafat, flavor divides consumers in those that prefer low or medium (their centroids are quite near) and those that prefer the flavor of meat with high intrafat. There are two market segments according to intrafat when the flavor is the target feature. Finally, Figure 3 depicts consumers located in R2 according to their rat- 25 1 7 21 −1.4 −1.2 −1.0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 medium high low −1.4 −1.2 −1.0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 Figure 2: Consumers represented according to their ratings in flavor. Voronoi diagram whose seeds are the centroids of items with different values of ageing (left) and intrafat (right) ings of tenderness. In this case, the centroids are clearly separated. When considering centroids of ageing values we appreciate that 21 is the value more associated with tenderness, this is a well known fact since the ageing is closely related with physical measures of softness in meat. 7. Conclusions We have presented factorization approaches to learning and visualizing preferences of consumers about a kind of products. The models learned are more accurate than existing tensorial approaches that typically use a SVM. The framework presented in this paper includes at the same time factoriza- tion and tensorial methods; both cases use the same learning algorithm with a different equation as the goal to optimize. Then, the accuracy of the model 26 7 1 21 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 −1.2 −1.0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 high medium low −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 −1.2 −1.0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 Figure 3: Consumers represented according to their ratings in tenderness. Voronoi diagram whose seeds are the centroids of items with different values of ageing (left) and intrafat (right) can be explained in terms of the number of parameters to learn. Factoriza- tion models are obtained with two embeddings and need substantially less parameters than tensorial approaches. Additionally, embeddings can be seen as Euclidean representations of both consumers and products. The closeness of these representations have a straightforward semantics. Hence, consumers’ clusters can be seen as market segments, and products clusters are groups of similar items with respect to consumer tastes. As in any other knowledge-based system, we observed that the available knowledge about consumers and products is of prime importance. If the identifiers of consumers and products are included, the accuracy of the hy- pothesis learned is dramatically improved. However, if only identifiers are 27 included, there is a drawback that must be considered: no predictions can be made for new (unknown) consumers and/or products. This is the main limi- tation of the method, although the method presented here is flexible enough to be able to use the available knowledge. The overall approach presented in this paper can be extended to other application fields. The requirements include situations where the the interac- tion of two vectors determines a class or an amount endowed with some kind of ordering. This is the case of recommender systems or in general matrix completions, well-known applications of embeddings or matrix factorizations. What we emphasize here is the graphical properties of the Euclidean repre- sentations. Then, it is possible to learn similarities of objects with respect to their behavior with a class. The applications include direct marketing and fraud detection. To check the validity of the proposal we used a set of experiments carried out with real data of sensory analysis of beef meat according to consumer preferences. Factorization methods outperform tensorial SVM. On the other hand, the Euclidean representations obtained in these datasets emphasize the relevance of some well-known traits involved in consumer preferences. The software used in the experiments can be downloaded from this1 web- site. 1We will provide a link to download the implementation in the final version of the paper 28 Acknowledgments The research reported here is supported in part under grant TIN2011- 23558 from the Ministerio de Economı́a y Competitividad, Spain. The paper was written while A. Bahamonde was visiting Cornell University with grants Movilidad Campus de Excelencia Internacional (Universidad de Oviedo) and from Programa Nacional de Movilidad de Recursos Humanos (Ministerio de Educación, Cultura y Deporte, Spain). References Agarwal, D., Chen, B.-C., 2009. Regression-based latent factor models. In: Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, pp. 19–28. Bahamonde, A., Dı́ez, J., Quevedo, J., Luaces, O., del Coz, J., 2007. How to learn consumer preferences from the analysis of sensory data by means of support vector machines (SVM). Trends in Food Science & Technology 18 (1), 20–28. Basilico, J., Hofmann, T., 2004. A joint framework for collaborative and content filtering. In: Proceedings of the 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, pp. 550–551. Bayer, I., 2015. fastfm: A library for factorization machines. arXiv preprint arXiv:1505.00641. 29 Chen, S., Moore, J., Turnbull, D., Joachims, T., 2012. Playlist prediction via metric embedding. In: Proceedings of the 18th ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining. ACM, pp. 714–722. Chen, T., Zheng, Z., Lu, Q., Zhang, W., Yu, Y., 2011. Feature-based ma- trix factorization. Tech. rep., Apex Data & Knowledge Management Lab, Shanghai Jiao Tong University. arXiv:1109.2271. del Coz, J. J., Bayón, G. F., Dı́ez, J., Luaces, O., Bahamonde, A., Sañudo, C., 2005. Trait selection for assessing beef meat quality using non-linear SVM. In: Advances in Neural Information Processing Systems 17 (NIPS ’04). pp. 321–328. Dı́ez, J., Del Coz, J., Bahamonde, A., Sañudo, C., Olleta, J., Macie, S., Campo, M., Panea, B., Albert́ı, P., 2006. Identifying market segments in beef: Breed, slaughter weight and ageing time implications. Meat science 74 (4), 667–675. Dı́ez, J., del Coz, J., Sañudo, C., Albert́ı, P., Bahamonde, A., 2005. A kernel based method for discovering market segments in beef meat. Proceedings of the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases: ECML/PKDD 2005, 462–469. Du, C., Zhe, S., Zhuang, F., Qi, Y., He, Q., Shi, Z., 2015. Bayesian maximum margin principal component analysis. In: Twenty-Ninth AAAI Conference on Artificial Intelligence. 30 Gil, M., Serra, X., Gispert, M., Angels Oliver, M., Sañudo, C., Panea, B., Olleta, J. L., Campo, M., Oliván, M., Osoro, K., Garćıa-Cachan, M., Izquierdo, M., Espejo, M., Mart́ın, M., Piedrafita, J., 2001. The effect of breed-production systems on the myosin heavy chain 1, the biochemical characteristics and the colour variables of longissimus thoracis from seven spanish beef cattle breeds. Meat Science 58 (2), 181–188. Herbrich, R., Graepel, T., Obermayer, K., 1999. Large margin rank bound- aries for ordinal regression. Advances in Neural Information Processing Systems, 115–132. Hüllermeier, E., Fürnkranz, J., 2013. Editorial: Preference Learning and Ranking. Machine Learning, 1–5. Joachims, T., 2002. Optimizing search engines using clickthrough data. In: Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, pp. 133–142. Koren, Y., Bell, R., Volinsky, C., aug. 2009. Matrix Factorization Techniques for Recommender Systems. Computer 42 (8), 30 –37. Koren, Y., Carmel, L., 2004. Robust linear dimensionality reduction. Visu- alization and Computer Graphics, IEEE Transactions on 10 (4), 459–470. Moore, J., Chen, S., Joachims, T., Turnbull, D., 2012. Learning to embed songs and tags for playlist prediction. In: Proceedings ISMIR. Ocepek, U., Rugelj, J., Bosnić, Z., 2015. Improving matrix factorization rec- ommendations for examples in cold start. Expert Systems with Applica- tions. 31 Pahikkala, T., Airola, A., Stock, M., De Baets, B., Waegeman, W., 2012. Efficient regularized least-squares algorithms for conditional ranking on relational data. Machine Learning, 1–36. Parameswaran, S., Weinberger, K. Q., 2010. Large margin multi-task metric learning. In: Lafferty, J., Williams, C., Shawe-Taylor, J., Zemel, R., Cu- lotta, A. (Eds.), Advances in Neural Information Processing Systems 23, NIPS. pp. 1867–1875. Peltonen, J., Klami, A., Kaski, S., 2003. Learning metrics for information visualization. In: Proceedings of the Workshop on Self-Organizing Maps (WSOM’03). pp. 213–218. Rendle, S., 2012. Factorization Machines with libFM. ACM Transactions on Intelligent Systems and Technology (TIST) 3 (3), 57. Rendle, S., Freudenthaler, C., Gantner, Z., Schmidt-Thieme, L., 2009. BPR: Bayesian Personalized Ranking from Implicit Feedback. In: Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence. AUAI Press, pp. 452–461. Rendle, S., Schmidt-Thieme, L., 2010. Pairwise Interaction Tensor Factoriza- tion for Personalized Tag Recommendation. In: Proceedings of the third ACM International Conference on Web Search and Data Mining. ACM, pp. 81–90. Robbins, H., Monro, S., 1951. A stochastic approximation method. The An- nals of Mathematical Statistics, 400–407. 32 Sañudo, C., Macie, E., Olleta, J., Villarroel, M., Panea, B., Albertı, P., 2004. The effects of slaughter weight, breed type and ageing time on beef meat quality using two different texture devices. Meat Science 66 (4), 925–932. Shalev-Shwartz, S., Singer, Y., Srebro, N., Cotter, A., 2011. Pegasos: Pri- mal estimated sub-gradient solver for SVM. Mathematical Programming 127 (1), 3–30. Thurstone, L. L., 1927. A law of comparative judgment. Psychological Review 34 (4), 273. Weston, J., Bengio, S., Hamel, P., 2011. Multi-tasking with joint semantic spaces for large-scale music annotation and retrieval. Journal of New Music Research 40 (4), 337–348. Weston, J., Bengio, S., Usunier, N., 2010. Large Scale Image Annotation: Learning to Rank with Joint Word-Image Embeddings. Machine Learning 81 (1), 21–35. Xing, E. P., Jordan, M. I., Russell, S., Ng, A. Y., 2002. Distance metric learning with application to clustering with side-information. In: Advances in neural information processing systems. pp. 505–512. Yu, S., Yu, K., Tresp, V., Kriegel, H.-P., Wu, M., 2006. Supervised prob- abilistic principal component analysis. In: Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data min- ing. ACM, pp. 464–473. 33 Introduction Related Work Formal Framework Maximum Margin Approach Factorization and Tensorial Approaches Mapping Consumers and Products: Matrix Factorization Inner Products Euclidean Closeness Tensor Product Experimental Results Implementation Details Datasets Results and Discussion Visualization of Preferences Conclusions 
work_fmu36dlyd5h77k73ph7i57ew4a	----	Informe 1 Model-based Expert System to Automatically Adapt Milling Forces in Pareto Optimal Multi-Objective Working Points Luis Rubio* 1 , Manuel De la Sen 2 , Andrew P. Longstaff 1 and Simon Fletcher 1 1 Centre for Precision Technologies, University of Huddersfield, Queensgate, Huddersfield West Yorkshire, HD1 3DH, United Kingdom. *Corresponding author: e-mail: L.R.Rodriguez@hud.ac.uk, Telephone: +44(0)1484471805. 2 Institute of Research and Development of Processes, University of the Basque Country, Barrio Sarriena s/n, 48940, Leioa, Bizkaia, Spain . Abstract – The objective of this paper is to present an open and modular expert rule-based system in order to automatically select cutting parameters in milling operations. The knowledge base of the system presents considerations of stability, machine drives efficiency and restrictions while adaptively controlling milling forces in suitable working points. Moreover, a novel classical cost function has been conceived and constructed to Pareto-optimise cutting parameters subjected to multi-objective purposes, namely: tool-life, surface roughness, material remove rate and stability rate parameter. Different Pareto optimal front solutions can be obtained modulating the weighting factors of the cost function. Additional rules have been added in order to manually and/or automatically modulate this cost function. Furthermore, a database which relates weighting factors, cutting conditions and cost function variables is produced for learning purposes. Chatter detection and suppression system automatically feedback to the system to take into account non-modelled disturbances. Finally, since the knowledge of the system is basically obtained from mathematical models, the possibility of combining experience and knowledge from expert engineers and operators is included. In this way, best practice from mathematical modelling and expert engineers and operators is joined in one system obtaining a full, automated system combining the best of each world. As a result, the expert rule-based system selects Pareto optimal cutting conditions for a broad range of milling processes, sorting out automatically different problems such as chatter vibrations, incorporating model reference adaptive control (MRAC) of forces. This procedure is intuitive, being executed in the same way as a human expert would do and it provides the possibility to interact with expert engineers and operators in order to take into account their experience and knowledge. Finally, the expert system is designed in modular form allowing incorporating new functionalities in rule based forms to them or just adding new modules to improve the performance of the milling system. Keywords: Milling Processes, Expert System, Optimization, Manufacturing Automation, Adaptive Forces Control 1. Introduction Despite various attempts at optimization, the selection of cutting parameters for CNC milling operations is still largely driven by machine operators on the shop-floor. They use their experience and/or handbooks in order to program adequate cutting parameters [Balakrishnan and DeVries, 1982, Liang, et al. 2004]. Normally, those parameters are selected in an intuitive way and/or using machining handbooks leading to programming cutting parameters under safety upper limits in order to prevent vibrations and process malfunctions. In this sense, automation techniques are being introduced in manufacturing environments to computerize and mailto:L.R.Rodriguez@hud.ac.uk 2 achieve more accurate solutions due to increase competitive markets. As a first solution adaptive controllers substituted fixed gain controllers in order to behave better under sensible changes in the depth of cut [Koren, 1989]. One step ahead is the adaptive optimization of the cutting parameters using intelligent techniques. Those techniques present multi-function optimization through, for example, neural networks, genetic algorithms and other bio-inspired techniques [Surech, et al. 2003, Cus & Balic,2003, Zuperl & Cus 2003, Abellan et al. 2008, Wong and Hamouda 2003b, Zuperl, et al. 2005]. Nevertheless, those methodologies provide intrinsic mathematical non-linear functions, which learn in hidden “black boxes” from a series of examples, given useful solutions but reducing the transparency of the interfaces. Furthermore, expert systems have also been developed to cope with this problem. In manufacturing processes, expert systems propose two steps. First, a knowledge base about the system is addressed, and then, some pseudo-heuristic rules are used, extracted from knowledge or experience in order to infer a solution, often using Fuzzy Logic (FL) or reasoning. Some versatile approaches in the literature are, for example; Wong and Hamouda, (2003a), where an online knowledge based fuzzy expert system for machinability data selection using Fuzzy Logic as reasoning mechanism in capturing the knowledge of machining operators is developed. Also in Vidal, et al. 2005 computer aided process planning is used for choosing the manufacturing route in metal removal processes and their cutting parameters. On the other hand, Zhang and Lu, (1992) discussed an expert system for economic evaluation of machining planning operation through integration of the manufacturing and management systems, trying to plan technical issues tied to the economical ones such as amortization of the machine, taking into account labour, materials and working capacity. In Morgan, et al. (2007), the milling system is fully diagnosed using FL, providing sources of machining problems and corrective actions. Finally, in Iqbal, et al. (2007), tool life is enhanced and work-piece surface finish improved using experimental data and if-then rules through analysis of variance techniques and numeric optimization. Moreover, some theoretical and experimental work has been carried out in order to get optimal machining parameters subject to dimensional precision and surface quality, tool 3 life expectancy and production times [Vivanco, et al. 2004,Chien & Chou, 2001 and references therein], in different machining processes and using different materials. The term expert system was originally used to denote systems using a significant amount of expert information about a particular domain in order to solve problems within that domain. Due to the important role of knowledge in such systems, they have also been called knowledge- based systems. However, since the terminology has been applied to so many diverse systems, it has essentially evolved into two different uses of the term. First, the term is often used to describe any system constructed with special kinds of “expert-systems” programming languages and tools, including production systems, rule-based systems, frame-based systems, “blackboard” architectures, and programming languages such as LISP or Prolog. Nevertheless, another important feature is that, since expert systems are usually non-deterministic, a large number of modules may be candidates for activation at any given moment. Thus, a criterion is needed to determine how to select which of the applicable modules must be executed next, and what to do after selection. This second type is the more appropriate job of an expert system in the sense that it is a system that “reasons” about the problem in much the same way as humans do. Despite the fashion of using bio-inspired optimization methods, this work proposes classical optimization methods which allow assembling self-learning and self-adaptive algorithms in a modular way to search for different Pareto optimal solutions. For this purpose, it is used the calculus of optimal working points associated with methods to calculate the stability of the dealt system. Then, the proposed expert rule-based system deals with stability issues and more in- depth analysis can be added. Furthermore, this paper makes an attempt to manage the milling system through reasoning of the possible states in spite of using blackboard architectures. In this paper, the milling system is described from the point of view of an expert system, but not in the traditional sense where knowledge is extracted from expert engineers’ and/or operators’ experience. Instead, the dynamic behaviour of the system obtained from theoretical models is used to develop the expert system. Those models of the cutting process consist of the relative 4 compliance between the tool and the work-piece, given by the modal parameters of the used tool. The advantage of this methodology is that the use of mathematical models gives universal solutions in contrast to linguistic solutions which give particular solutions. Moreover, the proposed expert system can be supplemented by semantic-based solutions added in parallel to the modelled expert system, giving the potential to act with the best qualities of each. While this dual method is attractive, care has to be taken of the increase in complexity and cost to develop models of the system and parameterize their constants. This is mitigated in the case of milling since it is a well-known study in literature [Budak & Altintas, 1998, Balanchandran, 2001] and it is possible to take advantage of this extensive literature. The paper is scheduled as follows. A brief introduction of the dynamic delay differential equation which governs milling systems is initially given. The transfer function which relates forces and programmed feed rates is then explained. This is followed by the development of seven modules which are composed by rules. The first module gives advice about robustness of the system and allowable input cutting parameters; the second one introduces model reference adaptive control of milling forces; the third covers constraints of the spindle and feed drives capabilities; the fourth suggest initial cutting parameters; the fifth model is composed of a novel cost function which measures the performance of the system giving Pareto optimal cutting parameters depending on the milling process to carry out and, the possibility to program different cutting parameters corresponding to different Pareto optimal fronts automatically through the modification of the weighting factors of the cost function; module 6 gives automatic feedback to the system due to non-modelled facts; finally module 7 provides the possibility to interact with expert engineers or operators to input to the system their experience and knowledge. Results clarify the developed work and conclusions and discussion will end the paper. 5 2. System description Milling processes are well characterized as mechanical systems which are particularly sensibility to acquiring vibrations. In this section, the milling process is modelled as a second order differential equation, which is excited by forces whose inherent terms excite the modal parameters of the system. This fact results in the conversion of resultant energy into vibrations of the system. Those vibrations are generated under certain cutting conditions depending on the process being carried out, clamping of the workpiece, tool and workpiece materials, etc. [Budak & Altintas,1998, Landers & Ulsoy, 1993]. In this frame of mind, the standard milling system responds to a second order differential equation excited by the cutting forces, ( ) ,[ Budak & Altintas,1998, Landers & Ulsoy, 1993], ̈( ) ̇( ) ( ) ( ) (1) where ( ) { ( ) ( )} are the relative displacements between the tool and the workpiece in the X-Y plane, ( ) { ( ) ( )} , and and are the modal mass, damping and stiffness matrices, all of them represented in two dimensions. The milling cutting force is represented by a tangential force proportional with the instantaneous chip thickness, and a radial force which is expressed in terms of the tangential force [Budak & Altintas, 1998, Balanchandran, 2001] ( ) ( ) and ( ) ( ) (2) where and , the tangential and radial specific cutting constants which are dependent on the tool material for any geometry, , the axial depth of cut and, ( ) , the chip thickness, obtaining the cutting forces in Cartesian coordinates [Balanchandran, 2001]. The most critical variable in the equation of motion, the chip thickness, ( ), consists of a static part and a dynamic one. The static’s is proportional to the feed rate and it is attributed to the rigid body motion of the cutter. The dynamic one models two subsequent passes of the tool through the same part of the work-piece. The phase shift between two consecutive passes of one tooth on 6 the working-piece can be seen in figure 1. It is widely modelled as [Budak & Altintas, 1998, Balanchandran, 2001], ( ) [ ( ) ( )] [ ( ) ( )] (3) where is the feed rate, the immersion angle and is a delayed term defined as , is the number of teeth and the spindle speed in . The equations of motion (1, 2 and 3) correspond with a second order delay differential equation. It can be solved numerically [Budak & Altintas, 1998, Insperger & Stepan, 2002] or analytically [Stepan, 1989]. X Y xk yk xb yb sN j ju jv rjF tjF jtooth  )( 1 jtooth )( 2 jtooth )( jtoothbyleft marksvibration   )( 1  jtoothbyleft marksvibration )( 2  jtoothbyleft marksvibration Workpiece directionfeed  Figure 1: Cross-sectional view of a milling tool [Altintas, 2000]. On the other hand, the transfer function of the system, in chatter and resonant free zones, can be separated as a series decomposition of the transfer function which relates the resultant force and the actual feed delivered by the drive motor, which models the deflection of the tool, and the transfer function which represents the Computerized Numerical Control (CNC). Then, a continuous transfer function which relates both signals, measured resultant force and the actual feed delivered by the drive motor can be showed as a first order dynamic [Altintas, 2000], ( ) ( ) ( ) ( ) (4) 7 where ( ⁄ ) is the resultant cutting pressure constant, ( ) is the axial depth of cut, ( )is a non-dimensional immersion function, which is dependent on the immersion angle and the number of teeth in cut, is the number of teeth in the milling cutter, ( ⁄ ) the spindle speed and ⁄ . At the same time, the relationship between the machine tool control, the CNC and, the motor drive system can be approximated as a first order system within the range of working frequencies [Altintas, 2000]. This transfer function relates the actual, , and the command, , feed velocities, ( ) ( ) ( ) (5) where represents an average time constant. The combined transfer function of the system is given by [Altintas, 2000], ( ) ( ) ( ) ( )( ) (6 )with ( ⁄ ) ⁄ . 3. Knowledge base identification Milling processes basically consist of three possible phases: roughing, medium and finishing the surface [Juneja, et al. 2003]. The selection of optimal cutting parameters, for each phase, plays an important role in manufacturing. In roughing phases, large amounts of material are removed with the emphasis on speed, but the tendency of the system to propagate and amplify vibrations limits the process. In medium phases, a trade-off between removal rate and form generation of the workpiece is balanced. Finally, finishing cuts demand more accurate control programs, adequate cutting conditions and selection of the milling process to achieve accurate products. Further to the above constraints of the system, the life of the tool can represent an important performance index to take into account when selecting cutting conditions. One reason is that an increase in the cost of tooling will decrease the benefits of rapid production. Another reason is that wear changes the geometry of the tool, which can either lead to degradation in cutting accuracy and speed or more frequent tool changes which increases production time [Kalpakjian & Schimidt, 2000]. 8 Of significant and increasing importance for many components is the surface finish. Roughness is one measure of the texture of the surface and is another indicator of quality of the final product [Baek, et al. 1997] that is affected by both tool wear and vibrations. The main source of vibrations in milling arises from the regenerative effect [Altintas, 2000, Balanchandran, 2001] known as chatter vibrations which are typically modelled by the equations (1), (2) and (3). They can be solved in the time domain and frequency domain. Frequency domain outputs the well- known stability charts, which gives stability frontiers in the cutting space parameter [Altintas, 2000, Balchandran, 2001]. Other sources of vibrations are couple-mode [Tlusty and Koenigsberger, 1970] and forced vibrations [Sutherland & Andrew, 1968]. Furthermore, controlling the forces is required in mechanical systems in order to govern the system under variation in system parameters. Keeping the forces under a prescribed safe upper limit prevents the system from having deflections on the tool and avoids tool damage and breakage [Altintas, 1992]. Conversely, the bigger the permitted force is, the more material can be removed in a single pass. From this we can see that there is a trade-off between deflection of the tool and material remove rate when programming the reference forces [Zuperl. et al, 2005, Rubio, et al. 2007a and 2007b, Altintas, 2000]. Finally, since milling processes can be defined as the relative movement between feeding a workpiece while rotating a multi-tooth tool, the feed and spindle servo-drives restrictions are considered in order to give more robustness and efficiency to the system. Then, the knowledge base presented in this section is transcribed as a series of rules which operates in an open and modular way. In this way, the expert system can be applied in modules adapting it to every machine and conditions. 3.1 Module I: Stability robustness and allowable input space parameter In this section, some robustness considerations are discussed, after which, the allowable input space parameter is declared. For this purpose, the following algorithmic methodologies are taken into consideration and programmed in the expert system: 9 Rule 1: Stability robustness. This rule deals with the regions (axial depth of cut and spindle speed pairs) where there exist uncertainty in the stability due to the model of chatter vibrations. Rule 1.1: Stability region inexactitude The stability lobes are calculated from a linear approximation [Budak & Altintas, 1998, Altintas, 2000], so that the nominal stability frontier and its neighbourhood are inaccurate as stable regions of the real nonlinear problem. For this reason and in order to calculate secure stability lobes, an accurate stability margin is prescribed. The Laplace transform imaginary axis is translated a value to the left part, allowing minimization of the approximations when stability charts are calculated in the frequency domain using equations (1)-(3); it is supposed that the chatter vibrations happen at ( ) instead of at when the stability border line is calculated. By examining the influence of in the stability lobes (figure 2), it is concluded that large values are necessary in order to get appreciable stability margins in lobes. Figure 2: Influence of in stability lobes. Rule 1.2: Influence of the axial depth of cut in robustness The expression of the value of the axial depth of cut which limits stable and unstable zones is multiplied by a factor, , , aimed at improving the robustness of the system when 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 x 10 4 0 1 2 3 4 5 6 spindle speed (rpm) de pt h- of -c ut (m m ) delta=0 delta=5 10 lobes charts are calculated using equations (1)-(3). Then, a refinement on this parameter, , allows better control capacity in the spindle speed, but limits the amount of material to be removed. From figures 2 and 3, it can be concluded that has more influence in stability charts than , in absolute terms. Figure 3: Influence of in stability charts. Rule 2: Allowable cutting space parameter. Better knowledge about feasible cutting input space parameters allows programming more successful cutting conditions. This rule gives the steps followed in this paper to calculate the allowable cutting space parameter. Rule 2 has been split into two rules: Rule 2.1: Border line data extraction This first sub-rule calculates the value of the pair, axial depth of cut and spindle speed, which compose the border line between stable and unstable zones, equations (1)-(3), satisfying rule 1. The purpose of this rule is to identify the border line between stable and unstable cutting parameters. Rule 2.2: Broad input cutting parameter space Since it is a requirement to work in a stable and robust region where the system will not be influenced by chatter vibrations, the set of input parameters is given by the pairs axial depth of 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 x 10 4 0 1 2 3 4 5 6 spindle speed (rpm) de pt h- of -c ut (m m ) alpha=1 alpha=0.95 11 cut and spindle speed which are in the stable region of the lobes satisfying rule 1. Then, the pairs which are below the border line according to rule 1 are calculated. 3.2 Module II: Model reference adaptive control of milling forces A model reference adaptive control scheme is proposed to control the forces of the system and considerations of the sampling time are added. Rule 3: Considerations about controlling the forces The control of forces in milling machines occupies an extensive amount of literature [Altintas, 2000, Lauderbaugh & Ulsoy, 1998A and 1998B, Peng, 2004, Rubio et al. 2007a and 2007b]. The reasons for this are that keeping forces below prescribed safety upper bounds avoids spindle, tool and/or work-piece deflections which deteriorate resultant geometric accuracy and, can cause the breakage of the tool or damage machine components [Altintas, 1992,Campomanes & Altintas, 2003]. Then, the upper limit of the total cutting force, , that results from machining operation must not exceed the allowed cutting force, , that the tool can resist. In this paper, an adaptive control is proposed to program the feed rate according to the transfer function of equation (6). The use of adaptive controllers is due to the variation of cutting and intrinsic parameters and, possible external perturbations in the system during machining. Since the transfer function of the system is continuous and machine process is controlled at each spindle speed period, the zero order hold ( ) equivalent of ( ) is considered, using a recursive least square estimator to update the estimation parameter vector [Rubio et al. 2007a and 2007b]. Rule 4: Sampling time selection and computer resources maximization The selection of the sampling time is typically selected as ⁄ , in . Since this rule is intended to give a compromise between reconstructing clear digital output signals and 12 optimizing computer resources during steady state operation, the following algorithmic methodology is proposed: If and remain constant ⁄ where predefined by the designer, else . 3.3 Module III: Drives constraints Milling, basically, consists of feeding a workpiece relative to a rotating multi-tooth tool. Therefore, there are two main types of drives, namely, the spindle motor and feed drives. Rule 5: Spindle power consumption Power draw from spindle motor and the efficiency of the motor are machine tool specific values where the net power can be calculated to ensure that the machine in question can cope with the cutter mass and operation. Therefore, the power drawn from the spindle motor constrains the machining efficiency [Maeda, et al. 2005]. The cutting power is found from equation 7 [Altintas, 2000, Maeda, et al. 2005] as, ∑ ( ) (7) where is the tool diameter, the spindle speed and the tangential cutting force. The cutting power required for the spindle motor is the maximum value among the instantaneous power in one tooth period, ( ) (8) [Altintas, 2000, Maeda, et al. 2005]. Rule 6: Feed drive restrictions Feed drive motor must have enough torque to accelerate the table and workpiece and to cover frictions and forces acting in the feeding directions of the table. For simplicity, in this case it will suppose that feed drives are limited by its upper maximum feed, given by feed drive system or force upper. In this case it is represented by which can be affected by, for example, the weight and size of the workpiece, depending upon machine configuration. 13 3.4 Module IV: Initial cutting parameter selection This module gives assistance in the selection of the initial cutting parameters which can potentially lead to saving machining times. Rule 7: Initial cutting parameter space selection This rule provides guidance to propose initial cutting parameters: Rule 7.1: Initial cutting parameters subjected to constraints The admissible input parameter-space subject to cutting force control and spindle motor power availability and feed drive constraints is obtained. The initial cutting parameter space is given by the triple- ( ) so as fulfill rule 1, , where is the power available in the spindle motor and where is given by feed drive system or force restrictions. Rule 7.2: Selection of initial spindle speed Spindle speeds associated with the modal frequencies are selected as initial candidate working points. Those spindle speeds are associated with the areas in the stability lobes where deeper axial depth of cuts can be programmed. 4. Module V: Cost function definition and rules to inference with it This section defines a classical cost function to Pareto optimise multi-purpose objectives. 4.1 Cost function definition A novel cost function has been conceived to allow an inference engine to carry out the selection of suitable cutting parameters. The tool cost model for a single milling process can be calculated using the following equation, (9): ( ( )) 14 The cost function has four terms. Each term is composed of a weighting factor ( ), a normalisation factor ( ) and the function which delimits the process efficiency. These functions are: the life of the tool, ; the material remove rate, ; the surface finish, ; and the robustness of the system , . The tool cost function is designed to be directly proportional to the life of the tool, material remove rate and robustness of the system and inversely proportional to surface roughness. So, optimal solutions will maximise , and while minimising . These parameters play an important role when selecting cutting parameters since they are usually used as benchmark indices in industries to measure the performance of the system. They are defined as following: 4.1.1 Life of the tool (TOL) is a measure of the length of time a cutting tool will cut effectively. According to some studies [Wong & Hamouda, 2003 ,Ginta, et al. 2009, Alauddin, et al. 1997], an increase in the cutting speed, feed rate and axial depth of cut will decrease the tool life. In this paper, the Taylor Equation for Tool Life Expectancy, a model typically used in literature, is used to evaluate in the expert system. This model is represented by the equation [Wong & Hamouda, 2003, Ginta, et al. 2009]: (10) where is a model constant, and , are model parameters and and , the cutting speed ( ⁄ ), axial depth of cut ( ) and feed per tooth ( ⁄ ) . 4.1.2 Material or metal remove rate (MRR) The measures the amount of material removed from the workpiece. Its definition is, , (11) where is the axial depth of cut ( ), the radial depth of cut ( )and the feed velocity ( ). 15 4.1.3 Surface roughness (SURF) The variations of the surface roughness are widely used criteria for the assessment of the surface quality. Some research works use the empirical relationship of the equation (12), [Wong & Hamouda, 2003, Kalpakjian, 2000]. This approach is adopted in this paper: (12) where and are the cutting velocity ( ), the feed velocity ( ) and axial depth of cut ( ) , and is a model constant and, and surface roughness model parameters. 4.1.4 Robustness of the system (ROS) The robustness of the system is a non-conventional measure to evaluate how close cutting parameters are to the border line in the stability lobe diagrams. For this purpose, a non- dimensional parameter is defined by: ( {( ) ( ) }) (13) where is the length of the axial depth of cut to the lobe and is the length of the spindle speed to the lobe, considering them to be non-dimensional. Finally, the weighting factors, have the restriction that the sum of the parameters is the unity, i.e. ∑ . Their declaration depends on process constraints. Normalization factors, , equalize the magnitude order of each term in the cost function. They are defined as: (14) where , represents each term of the cost function of the equation (9), which eventually, can be represented as ∑ (15) 16 4.2 Rules of inference with the cost function The following rules are proposed in order to modulate the cost function: Rule 8: Initial weighting factors selection. The weighting factors can be selected by an expert engineer or operator who can judge the importance of each term in the cost function with respect to the phase of the operation to be carried out. The initial weighting factors are selected in a heuristic way and only experience and expertise in their programming will make the expert system more efficient. Rule 9: Declaration of normalized factors In the definition of the normalization factors, , the pairs ( ) are computed as the maximum and minimum values of each term in the cost function, being the number of potential terms in the cost function. Rule 10: Pareto optimal cutting parameter selection The Pareto optimal concept is introduced here in order to clarify the selection of weighting factors. This concept requires optimisation of different control objectives, known as multi- objective optimization, and in general there is no single optimal solution, instead a set of possible solutions exits called a Pareto optimal front [Su and Hou, 2008, Zhang et al. 2011]. Which solution is chosen from the front depends on the weights that are given to the different objectives. Rule 10.1: Coarse selection The selected cutting parameters will be the values of and corresponding to the minimum value of the cost function according to selected values of the parameters. It can be expressed mathematically as follows, ( ) { ( ( ) ( ) ( ) ( ) )}, (16) obtaining the 3-tuple of candidate input cutting parameter, ( ). 17 Rule 10.2: Fine selection In order to have a more accurate solution, the cutting parameters are searched with a finer integration step around the point where the cost function gives its minimum value. In this case, the cutting parameter space is given by a 3-tuple ( ( ) ( ) ( ) ) around , for , where is the number of points to be considered, according to rules 1 and 2. The procedure for obtaining the required cutting parameters is the same as used in Rule 12.1 through equation (16) for the above defined new cutting parameters. Mathematically, it is expressed as, ( ){ ( ( ( )) ( ( )) ( ( )) ( ( )) )} (17) obtaining the 3-tuple of refined candidate input cutting parameter, ( ). Rule 11: Automatic modification of weighting factors In order to achieve certain process or machine tool requirements in the cost function variables, the parameters are automatically redefined as in equation (15): ∑ , which represents the proposed cost function of the equation (9). Then, for a given operation, if ̇ , and then if ∑ ∑ ( ) ( ) , ← ( ̅ ̅ ) , else if ∑ ∑ ( ) ( ) , ← ( ̅ ̅ ) end. Rule 12: Automatic modification of weighting factors through the gradient ascent method The steepest ascendant method is based on the estimation of a parameter vector in order to maximize a cost function [Passino, 2004]. Here, it is applied as, from equation (9), ∑ , then ] ( ), until | ] ( )| , . Rule 13: Re-normalisation of the weighting factors. After using rules 13 and 14, the weighting factors need to be normalized again to fulfil the constraint ∑ . Then, two possibilities are considered: 1. ← ∑ ← ̅⁄ ⇒ ∑ . 18 2. ← and ← for and . Rule 14: Re-parameterization of the parameters The parameters and the corresponding cost function variables are saved and stored. Then, a database is created with those values, which relates parameters to process variables in order to retain previous knowledge. In this way, the system can incorporate non responsive data mining algorithms, semantic rules or fuzzy rule based system to improve performance. Moreover, the system is open to the heuristic interaction with operators and expert engineers. This information can be further used for programming initial parameters according to production requirements, training novel operators and searching for more accurate cutting parameters. 5. Feedback to the expert system 5.1 Module VI: Automatic feedback and monitoring Due to non-modelled facts such as tool wear or run-out the predicted models cannot give accurate predictions about chatter vibrations when the tool is, for example, worn or run out. To consider this fact an automatic chatter detection and suppression system has been incorporated to the system. Rule 15: Automatic chatter vibration detection Despite having taken into account robustness margins against unstable situations against chatter, tool wear and/or other non-linear phenomena in the process can lead to vibrations. To accommodate this eventuality, a chatter detection algorithm has been integrated to the system in order to take action if chatter conditions are experienced due to unexpected conditions. Some approaches for on-line detection of chatter can be found in [Soliman & Ismail, 1997, Li et al., 1997 and Gradisek, et al. 1998]. In our case, the following algorithmic methodology has been taken into consideration [Campomanes & Altintas, 2003]: If (8), then there exists chatter vibration. 19 Where is the maximum uncut chip thickness during a dynamic time domain simulation and max,sh is the maximum uncut chip thickness during a time domain simulation in which the work-piece and cutter remain rigid. The threshold of instability is selected to be 1.25 as in Campomanes & Altintas, (2003). Rule 16: Automatic chatter vibration suppression In the case where chatter is detected, an automatic algorithm to suppress chatter has been embedded into the system. Since regenerative chatter is related to the interaction of the closed loop between two independent entities, the structural dynamics of the machine and the dynamics of the process, it is necessary to influence one of them to achieve a more stable and robust system. The most promising among the methods of influencing the cutting process is to control the spindle speed on-line, which means influencing the dynamics of the process. This can be achieved in two ways, by selecting the spindle speed or modulating it. In the expert system a spindle speed is selected automatically as Smith and Tlusty (1992) proposed. Then, in the case where chatter vibration is detected, an automatic variation of the spindle speed to deal with chatter problem is suggested. The strategy consist of detecting the chatter frequency and readjusting spindle speed in such a way that the new tooth passing frequency equals a multiple value of the chatter frequency. 5.2 Module VII: Expert Human Machine Interface Since expert systems normally codify expert engineers and/or operators experience and knowledge in order to create decision support systems (DSS), this last module incorporates the possibility of adding to the system this experience and knowledge in order to achieve the best of two possibilities. Moreover, monitoring vibrations and other parameters such as system forces, feeds, MRR, TOL and SURF is suggested to support the interaction between expert engineers, operators and the system. In this way, this interaction will be more reliable. 20 Rule 17: Monitoring signals In order to facilitate the human machine interface, it is proposed to monitoring the most important signals to provide better information to the expert system. For example, there are some sources of vibrations which can appear when machining. For example, forced vibrations, mode coupled, vibrations due to run-out, and so on [Wiercigroch and Budak, 2001]. For this reason, it is necessary to monitory the frequencies which appear in the system in order to identify them, and have the chance to deal with them. In this case, straightforward Fourier Frequency Transform (FFT) is presented for this purpose. Also, the most signal it is possible to monitor in the system the better performance will be achieved, then the above proposed signals are taken into account in this section in order to have better interaction with Rule 18 proposal. Rule 18: Expert Human Machine Interface This last rule incorporates the possibility of introducing expert engineers’ and/or operators’ experience and knowledge in form of, for instance, if/then rules in order to create decision support systems using data mining techniques for knowledge discover and reasoning techniques for supporting the system with possible decisions and advises [Gilbert et al. 2010, Luau et al. 2012, Alonso et al. 2012]. Figure 4 pictures the schematic representation of the DSS. E X P E R I E N C E Self-Learning System Data Mining Techniques K N O W L E D G E Self-Adaptive System Reasoning Techniques D - E S C Y I S S T I E O M N - Expert engineers and operators Figure 4: Suggested scheme of the DSS for dealing with expert engineers and operators experience and knowledge 21 In figure 4, it can be seen that experience can become knowledge through a self-learning system composed by data mining techniques. Also, knowledge can be introduced directly by expert engineers and operators. A decision system is extracted from knowledge through a self-adaptive system using reasoning techniques. The so-called “self-learning” and “self-adaptive” system is suggested to do it automatically. This rule can be added to the whole system or to each module independently. Milling Process determination (up, down, immersion, etc.) Modal characteristics, diameter, number of teeth of the tool Process restrictions (Maximum power available in the spindle motor, chatter threshold..) Expert system inputs Selected Cutting Parameters Expert System Expert system OutputsKnowledge Base Module V: Rules to cost function inference Initial weighting factors selection Declaration of the standardazing factors Change requirements?? Finer selection of the cutting parameters Adequate MRR, TOL, SURF and ROS ? Program Cutting Parameters and record data Weighting factors modification Automatic weighting factors modification Steepest method for automatic weighting factor modification Weighting factors renormalization 1. 2. Heuristic modification of the weighting factors 3. COST FUNCTION Transfer function cutting constant and time constants Feedback module VI Expert Interaction Module VII Yes No M o d u l E I Allowable cutting space parameters Robustness in calculate the lobes M o d u l e II Forces digital controller Automatic sampling time selection M o d u l e III Feed velocity restriction Efficiency of the spindle power consumption M o d u l e IV Initial cutting parameters selection Chatter detection system Chatter suppression system Monitoring Expert HMI Figure 5: Schematic Representation of the Expert System. Finally, figure 5 depicts a schematic representation of the expert system composed of five main blocks. The first one gives the inputs to the expert system, representing the milling process determination, modal characteristics, tool diameter and number of teeth and tool and workpiece material properties and, the transfer function which relates output force and programmed feed rates and process restrictions. The second block represents the knowledge base of the system and is composed for the first four modules, all of them related to the milling process. The third block describes the inference of the system with the cost function and concentrates on the modulation of it with some expert rules. This block, it also re-parameterises those parameters according to knowledge base insights. The fourth block shows the outputs. It, also, reflects the feedback to the system, in case of not achieving adequate objectives or of changing system requirements. Finally, the fifth block of the figure 5 represents the modules VI and VII. The 22 feedback module detects and suppresses chatter on-line in the case where non-modelled facets appears in in the milling process due, for instance, to tool wear. The expert interaction module provides the facility to interact automatically with experience and knowledge obtained from expert engineers and operators. The proposed rules are summarized in table 1. Rule Knowledge base 1. Since stability lobes are calculated from a linear approximation a stability region inexactitude is added. Also, a robustness factor which influences the axial depth of cut is taken into account. 2. Cutting space parameters extracted from lobes and restrictions in spindle power availability and cutting force controllers. 3. Considerations about controlling the forces; using model reference adaptive control for keeping forces under prescribed upper limit. 4. Sampling period considerations to maximize computer resources. 5. Spindle power consumption restrictions. 6. Feed drive limitations 7. Giving appropriate the initial cutting space parameter to decrease searching time. Rules to inference with the cost function 8. Initial weighting factors selection; the selection of right initial weighting factors plays an important role in achieving good solutions in short time. 9. Declaration of standardizing factors. 10. Selection of cutting parameters criterion; coarse and fine criterion. 11. Automatic rule of weighting factors modification. 12. Automatic weighting factors modification through gradient descendent method. An alternative way to modify automatically the weighting factors is proposed. 13. Renormalization of the weighting factors if they are re-programmed automatically through rules 11 and 12 14. Re-parameterization of weighting factors. 23 Feedback and expert HMI rules 15. Chatter vibration detection algorithm is added to be more reliable. 16. Chatter suppression algorithm to lead the system to stable and reliable cutting conditions. 17. Monitoring signals 18. Interaction with expert engineers and operators Table 1: Schematic representation of the expert rules 6. Case study For the validation of this method, an end mill has been chosen with the modal characteristics in the X and Y directions corresponding to table 2, with three tooth and 30 millimeter diameter. The work-piece is a rigid aluminium block whose specific cutting energy is and the proportionally factor is taken to be . ( ⁄ ) ( ) k ( ) X 603 3.9 5.59 Y 666 3.5 5.715 Table 2 Tool modal parameters. Other parameters belonging to the expert system are shown in table 3. They are the stability margin factor, , and the stability margin factor for the axial depth of cut, . Stability margin factors Life of the tool model parameters Surface roughness model parameters Table 3: Cost function model parameters. 0.05 0.95 1677110 -3.02 -0.54 -1.14 6088 0.4638 1.129 0.4461 24 Moreover, the parameters of the tool-life model ( ) and the parameters corresponding to the surface roughness model ( ), which have been considered constants in the simulations for simplicity purposes, are listed in table 3. Regarding to the model reference adaptive control, the transfer function of the equation (6); the cutting pressure of the transfer function has been selected to be constant and equal to ⁄ in all range of cutting parameters, the CNC time constant, and, ⁄ . The continuous model reference system of the adaptive control is chosen to be a typical continuous second order plant with and , where is the sampling period. In this work, it is desirable for the reference force to be maintained at 600 N. The analytical tests for cutting parameter selection were conducted within a known range of spindle speeds, starting at the spindle speed corresponding to its natural frequency and its sub- multiples. The integration in the axial depths of cuts is designed to start at its minimum value in the stability border line divided by a natural number designed by the engineer, with increments of the same size. The spindle motor is supposed to guarantee 2745.3 Watts of power at maximum value and, the feed motor is able to produce 25 ⁄ of feed velocity. The resulting cutting parameters are those which lead to a Pareto optimal value of the cost function depending on the chosen weighting factors. The conducted milling process is a full immersion face up-milling operation. Table 4 and figure 6 gather the scenarios which are specified in the current example: 1. The first point, (specified as point 1 in the table 4 and in figure 6), is associated with programming of initial parameters. In this case, the initial cutting parameters are programmed, intuitively, to preserve the tool, ensuring enough stability margin and minimizing the surface roughness, while the will not be important at first sight. Then, the following weighting factors are programmed and . They are equivalent to programming the following cutting parameters, ( ) The associated values of the material 25 removal rate, life of the tool, robustness of the system and surface roughness which are achieved working under those cutting parameters are and . And, the values which measure the chatter instability and the spindle power consumption are . Finer selection of cutting parameters can be achieved through rule 10.2. In this case, this rule leads to program the following cutting parameters, ( ). This second Pareto optimal solution improves every factor of the cost function except the robustness of the system, as it can be seen comparing previous them with values showed in table 4. ( ) ( ) ( ⁄ ) ( ⁄ ) ( ) ( ) ( ) 1. 1800 0.3693 4.24 0.0470 45.83 0.8810 1.18 1.0355 1696.5 2. 2412 0.3609 5.58 0.0604 19.6340 0.9753 1.26 1.0224 2273.3 3. 2010 1.5124 17.96 0.8149 3.2051 8.2557 0.0022 1.1157 1894.4 , . , . , . Table 4: Weighting factors and their associated cutting parameters and cost function values. 2. Secondly (specified as point 2. in the table 4 and in figure 6), a situation is assumed in which it is required to increase the productivity. In this case, the can be delimited between two values bigger than the current (point 1.) MRR, for example, . Using rule 12, this situation leads to program the following cutting parameters ( ), which gives the performance indexes and . This working point can be summarized in programming the following weighting factors and , using rule 13.b. In this 26 case, the values which measure the chatter instability and the spindle power consumption are . 3. The third situation (point 3. in the table 4 and in figure 6) that has been taken into consideration for increasing the removed material from the workpiece is tuning the - values manually, by engineer or operator experience, or, automatically, through rules 11 and 12. For example, the following values could be programmed and , which boost the increased of , leads to the following cutting parameters: ( )and the following cost function variables values and . Table 4 gathers the three described cases. It shows the programmed cutting parameters, the values of the variables of the cost function in that working point, the values which measure the chatter instability and the spindle power consumption and their corresponding weighting factors at those working points. Figure 6 pictures the represented cutting parameters in table 3 on the stability charts, pairs spindle speed and axial depth of cut for the cases 1, 2 and 3, and the programmed feed velocity for the three cases, represented by changes in the feed velocity. It can be observed that the control signal (feed velocity) is smooth and feasible except in the transient response and when cutting parameters change and the pairs of spindle speed and axial depth of cuts are both under the border line in the stable zone. The control signal has a peak at the transitory due to selection of initial conditions which are not close to the real values. It also experiments not smooth transition when changed from working point 2 to 3. The frequency response of the points 1 and 3 gives the tooth pass frequency and its harmonics in each case, as with case 2 which has not been considered in the graphs. The expert system is able to move around the cutting parameter space subjected to different production ‘states’ or requirements. This fact can be achieved by the programming of easy and intuitive parameters which leads to giving adequate cutting parameters to the milling 27 system according to production requirements. In this way, the expert system provides an intuitive but intelligent planning of cutting parameters giving a fast solution if changeable situations happen. Finally, through rule 14, the obtained parameters are stored subjected to learning and adaptive skills to help improve proven solutions. Figure 6: Programmed cutting parameters in lobes chart, programmed feed rates and frequencies. 7. Conclusions and discussion The paper describes, in an intuitive way but with mathematical thoroughness, the construction of an expert rule based system for cutting parameter selection in the milling processes. The approach consists of a series of rules split into 7 modules. Each module can interact independently leading to a universal system in the sense that it can be applied to every machine. The first module gives robustness to the system and presents the rough allowable input cutting parameters. The second module includes the model reference adaptive control of milling forces functionality, keeping the forces of the system under prescribed upper limit in spite of variations in system parameters. Moreover, the possibility of manipulating the sampling time of the system 1.5 2 2.5 3 3.5 4 4.5 0 1 2 3 spindle speed (krpm) d e p th -o f- c u t (m m ) 0 20 40 60 80 100 120 140 160 180 200 0 10 20 30 A m p lit u d e i n m m /s e c samples Control law : command f eed rate 0 100 200 300 400 500 600 700 800 0 0.2 0.4 0.6 0.8 1 x 10 -4 frecuency (Hz) a m p li tu d e 0 100 200 300 400 500 600 700 800 0 0.2 0.4 0.6 0.8 1 x 10 -3 frecuency (Hz) a m p li tu d e 1. 2. 3. 28 is added to preserve computer resources if necessary. The third module covers the spindle and feed drive motors constraints. The fourth module gives initial computational input parameters subjected to constraint of the motors and suggests potential initial spindle speed candidates. The fifth module presents a novel multi-objective cost function to evaluate the performance of the system. It has been devised from first principles and depends on the material to be removed, tool life, surface roughness and a stability margin. Weighting factors give the importance of the each term in the cost function. Each term is then modulated by a weighting factor, where the most important term is associated with the largest weighting factor to obtain Pareto optimal cutting parameters. Other Pareto optimal fronts can be obtained by two methods. First, by refining the searching cutting parameters around the selected one and, secondly, by modulating the weighting factors automatically if new production requirements are required or by interacting with system engineers or operators in order to take advantage of their experience. This information is stored in a database in order to register and modify previous data. Module 6 gives automatic feedback to the system if chatter vibrations are experienced due to non-modeled nonlinear effects such as wear or run out of the tool. Finally, module 7 proposes to monitor key parameters to better inter-actuation with expert engineers and operators and a scheme to infer with them providing in this way the best qualities of traditional expert systems and model based expert systems. Nevertheless, since the presented expert system is based on mathematical models of the system, its outputs selection is dominated by the modal parameters of the tool, tool and workpiece material properties, the linear transfer function which gives the relationship between the resultant force and the feed velocity, stability robustness constants, the determination of the process and the process restrictions. The developed expert rule based system suggests an open and modular architecture which is able to automatically execute orders and reason about the milling problem in a comparable way to humans reasoning, giving performances of possible programmable cutting conditions through the cost function with the possibility of incorporating new functionalities. Those performance indices are then used to select appropriate cutting parameters for the operation which leads to 29 the maximum productivity while respecting stability, spindle and feed motor consumptions and control constraints. Initial candidate cutting conditions are determined from the knowledge of the system, optimising, in this way, the process of training the system. Furthermore, the expert system is able to modify the values of the weighting factors automatically or interacting with operators with an easy interface if new constraints are required or they change, obtaining different Pareto optimal solutions. Finally, the system has automatic feedback if chatter vibrations occur due to non-modelled parameters such as wear or run out the tool. The developed expert rule based system uses classical mathematical approaches to calculate cutting conditions. In the simplest implementation the operator of the machine just has to tune to the system with four simple and intuitive values in order to program cutting parameters to fulfill full process requirements. A simulation example which shows the behaviour of the system is presented. As an added benefit, the integrated database, which is provided to introduce experience based or knowledge based rules if necessary or, the proposed control scheme which is able to update with information from external sources of experience and knowledge, providing an open and modular architecture with learning and adaptable skills. Finally, the expert system scheme can be extrapolated to any kind of system, given an intuitive, open control architecture, where learning and adaptive resources can be added according to multi-objective purposes. The open and modular design of the expert system allows new functionalities to be incorporated as and when new research allows. Acknowledgments The authors gratefully acknowledge the UK’s Engineering and Physical Science Research Council (EPSRC) funding of the EPRSC Centre for Innovative Manufacturing in Advance Metrology (Grant Ref: EP/I033424/1). 30 References ALAUDDIN, M., EL BARADIE, M.A. & HASHMI, M.S.J. ,(1997). PREDICTION OF TOOL LIFE IN END MILLING BY RESPONSE SURFACE METHODOLOGY. JOURNAL OF MATERIALS PROCESSING TECHNOLOGY, 71 (3), 456-465. ALONSO, F., MARTINEZ, L., PEREZ, A., VALENTE, J.P. (2012). COOPERATION BETWEEN EXPERT KNOWLEDGE AND DATA MINING DISCOVERED KNOWLEDGE: LESSONS LEARNED. EXPERT SYSTEMS WITH APPLICATIONS, 39, 7524-7535. ALTINTAS, Y., (1992). PREDICTION OF CUTTING FORCES AND TOOL BREAKAGE FROM FEED DRIVE CURRENT MEASUREMENTS, JOURNAL OF ENGINEERING FOR INDUSTRY, 114, 386-392. ALTINTAS, Y. (2000). MANUFACTURING AUTOMATION. CAMBRIDGE UNIVERSITY PRESS. AVELLAN, J.V., ROMEROS, F., SILLER, H.R., ESTRUD, A. & VILA, C.,(2008). ADAPTIVE CONTROL OPTIMIZATION OF CUTTING PARAMETERS FOR HIGH QUALITY MACHINING OPERATIONS BASED ON NEURAL NETWORKS AND SEARCH ALGORITHMS. ADVANCE IN ROBOTICS, AUTOMATION AND CONTROL, I-TECH VIENNA AUSTRIA. BAEK, D.K., KO, T.J. & KIM, H.S. , (1997). A DYNAMIC SURFACE ROUGHNESS MODEL FOR FACE MILLING, PRECISION ENGINEERING, 20, 171-178. BALANCHANDRAN, B., (2001). NON-LINEAR DYNAMICS OF MILLING PROCESSES. PHILOSOPHICAL TRANSACTIONS OF THE ROYAL SOCIETY, LONDON A, 359, 793-819. BALAKRISHANA, P. AND DEVRIES, M.F., (1982). A REVIEW OF COMPUTERIZED MACHINABILITY DATABASE SYSTEMS, PROCEEDINGS OF THE NAMRC-X, 348-356. BUDAK, E. & ALTINTAS, Y., (1998). ANALYTICAL PREDICTION OF CHATTER INSTABILITY IN MILLING-PART I: GENERAL FORMULATION & PART II: APPLICATION TO COMMON MILLING SYSTEMS. JOURNAL OF DYNAMIC SYSTEMS, MEASUREMENT AND CONTROL, 120, 22-36. CAMPOMANES M.L. & ALTINTAS, Y., (2003). AN IMPROVED TIME DOMAIN SIMULATION FOR DYNAMIC MILLING AT SMALL RADIAL IMMERSIONS. JOURNAL OF MANUFACTURING SCIENCE AND ENGINEERING, 126, 416-422. CHIEN, W.T. & CHOU, C.Y., (2001). THE PREDICTIVE MODEL FOR MACHINABILITY OF 304 STAINLESS STEEL, JOURNAL OF MATERIALS PROCESSING TECHNOLOGY, 118 (2), 442-447. CUS, F. & BALIC, J., (2003). OPTIMIZATION OF CUTTING PROCESS BY GENETIC ALGORITHM APPROACH. ROBOTICS AND COMPUTER-INTEGRATED MANUFACTURING, 19 (1-2), 113-121. 31 GRADISEK, J., GOVEKAR E. AND GRAVEC, I., (1998). USING COARSE-GRAINED ENTROPY RATE TO DETECT CHATTER IN CUTTING, JOURNAL OF SOUND AND VIBRATION, 214, 941-952. GILBERT, K., SANCHEZ-MARRE, M. AND CODINA, V., (2010). CHOOSING THE RIGHT DATA MINING TECHNIQUE: CLASSIFICATION OF METHODS AND INTELLIGENT RECOMMENDATION. IN THE PROCEEDINGS OF THE 2010 INTERNATIONAL CONGRESS ON ENVIRONMENTAL MODELLING AND SOFTWARE, 1933-1940. GINTA, T.L., NURUL AMIN, A.K.M., MOHD RADZI, H.C.D. & LAJIS, M.A., (2009). TOOL LIFE PREDICTION BY RESPONSE SURFACE METHODOLOGY IN END MILLING TITANIUM ALLOY TI- 6AL-4V USING UNCOATED WC-CO INSERTS. EUROPEAN JOURNAL OF SCIENTIFIC RESEARCH, 28 (4), 533-541. INSPERGER, T. & STEPAN, G., (2002). SEMI-DISCRETIZATION METHOD FOR DELAYED SYSTEMS. INTERNATIONAL JOURNAL OF SEMI-DISCRETIZATION METHODS IN ENGINEERING, 55 (5), 503- 518. IQBAL, A., HE, N., LI, L. & DAR, N. U., (2007). A FUZZY EXPERT SYSTEM FOR OPTIMIZING PARAMETERS AND PREDICTING PERFORMANCE MEASURES IN HARD-MILLING PROCESS. EXPERT SYSTEMS WITH APPLICATIONS, 32, 1020-1027. JUNEJA, B.L., SEKHON G.S. & SETH, N., (2003). FUNDAMENTALS OF METAL CUTTING AND MACHINE TOOLS. NEW AGE INTERNATIONAL. KALPAKJIAN, S. & SCHMIDT, S.R., (2000). MANUFACTURING ENGINEERING AND TECHNOLOGY.NEW YORK: PRENTICE HALL, UPPER SADDLE RIVER. KOREN, Y., (1989). ADAPTIVE CONTROL SYSTEMS FOR MACHINING. AMERICAN SOCIETY OF MECHANICAL ENGINEERS, MANUFACTURING REVIEW, 2 (1), 6-15. LANDERS, R.G. & ULSOY, A.G., (1993). CHATTER ANALYSIS OF MACHINING SYSTEMS WITH NON-LINEAR FORCE PROCESSES. ASME INTERNATIONAL MECHANICAL ENGINEERING CONGRESS AND EXPOSITION, 58, 183-190. LAUDERBAUGH, L.K. & ULSOY, A.G., (1998A). DYNAMIC MODELLING FOR CONTROL OF THE MILLING PROCESSES, JOURNAL OF ENGINEERING FOR INDUSTRY, 110, --. LAUDERBAUGH, L.K. & ULSOY, A.G., (1998B). MODEL REFERENCE ADAPTIVE CONTROL IN MILLING, JOURNAL OF ENGINEERING FOR INDUSTRY, 111, ---. LIANG, S.Y., HECKER, R.L. & LANDERS, R.G., (2004). MACHINING PROCESS MONITORING AND CONTROL: THE STATE OF THE ART. ASME J. OF MANUFACTURING SCIENCE AND ENGINEERING, 126 (2), 297-310. 32 LI, X.Q., WONG Y.S. AND NEE, A.Y.C., (1997). TOOL WEAR AND CHATTER DETECTION USING THE COHERENCE FUNCTION OF TWO CROSSED ACCELERATIONS. INTERNATIONAL JOURNAL OF MACHINE TOOLS AND MANUFACTURE, 37, 425-435. LIAU, S., CHU, P., HSIOA, P-Y, (2012), DATA MINING TECHNIQUES AND APPLICATIONS – A DECADE REVIEW FROM 2000-2011, EXPERT SYSTEMS WITH APPLICATIONS, 39, 11303-11311. MAEDA, O., CAO, Y. AND ALTINTAS, Y., (2005). EXPERT SPINDLE DESIGN SYSTEM. INTERNATIONAL JOURNAL OF MACHINE TOOLS AND MANUFACTURE, 45, 537-548. MORGAN, G., CHENG, R.Q., ALTINTAS, Y. & RIDGWAY, K., (2007). AN EXPERT TROUBLESHOOTING SYSTEM FOR THE MILLING PROCESS. INTERNATIONAL JOURNAL OF MACHINE TOOLS AND MANUFACTURE, 47, 1417-1425. PASSINO, K.M., (2004). BIOMIMICRY FOR OPTIMIZATION, CONTROL AND AUTOMATION. LONDON: SPRINGER. PENG, Y.H., (2004). ON THE PERFORMANCE ENHANCEMENT OF SELF-TUNING ADAPTIVE CONTROL FOR TIME VARYING MACHINING PROCESS. INTERNATIONAL JOURNAL OF ADVANCE MANUFACTURING TECHNOLOGY, 24, 395-403. RUBIO, L., DE LA SEN M. & BILBAO-GUILLERMA, A., (2007A). INTELLIGENT ADAPTIVE CONTROL OF FORCES IN MILLING PROCESSES. PROCEEDINGS OF THE 2007 IEEE MEDITERRANEAN CONFERENCE ON CONTROL AND AUTOMATION, ATHENS, GREECE, 1-6. RUBIO, L., DE LA SEN M. & BILBAO-GUILLERMA, A., (2007B). DISCRETE-TIME CONTROL OF MILLING FORCES USING FRACTIONAL ORDER HOLDS BY ON-LINE ADJUSTMENT OF THE CORRECTING GAIN. PROCEEDINGS OF THE 2007 IEEE MULTI-CONFERENCE ON SYSTEMS AND CONTROL, SINGAPURE, 1330-1335. SOLIMAN, E. AND ISMAIL, F., (1997). CHATTER DETECTION BY MONITORING SPINDLE DRIVE CURRENT. INTERNATIONAL JOURNAL OF ADVANCED MANUFACTURING TECHNOLOGY, 13, 27- 34. STEPAN, G., (1989). RETARDED DYNAMICAL SYSTEMS; STABILITY AND CHARACTERISTICS FUNCTIONS”, NEW YORK : JOHN WILEY AND SONS. SU, C., HOU, T., USING MULTI-POPULATION INTELLIGENT GENETIC ALGORITH TO FIND THE PARETO-OPTIMAL PARAMETERS FOR A NANO-PARTICLE MILLING PROCESS, EXPERT SYSTEMS WITH APPLICATIONS, 34, 2502-2510. SURECH, R., BASAVARAJAPPA, S., GAITONDE, V. N. & SAMUEL, G.L., (2012). MACHINABILITY INVESTIGATIONS ON HARDENED AISI 4340 STEEL USING COATED CARBIDE INSERTS. INTERNATIONAL JOURNAL OF REFRACTORY METALS AND HARD MATERIALS, IN PRESS. 33 SUTHERLAND, I.A. & ANDREW, C., (1968). FORCED VIBRATION AND CHATTER IN HORIZONTAL MILLING: AN INVESTIGATION USING A STRUCTURAL MODEL, PROCEEDINGS OF THE INSTITUTION OF MECHANICAL ENGINEERS. TLUSTY, J. & KOENIGSBERGER, F., (1970). MACHINE TOOL STRUCTURES, PERGAMON PRESS. VIDAL, A., ALBERTI, M., CIURANA, J. & CASADESÚS, M., (2005). A DECISION SUPPORT SYSTEM FOR OPTIMIZING THE SELECTION OF PARAMETERS WHEN PLANNING MILLING OPERATIONS. INTERNATIONAL JOURNAL OF MACHINE TOOLS AND MANUFACTURING. 45, 201-210. VIVANCO, J., LUIS, C. J., COSTA, L. & ORTIZ, J.A., (2004). OPTIMAL MACHINING PARAMETERS SELECTION IN HIGH SPEED MILLING OF HARDENED STEELS FOR INJECTION MOULDS. JOURNAL OF MATERIALS PROCESSING TECHNOLOGY, 155-156, 1505-1512. WIERCIGROCH, M. AND BUDAK, E.,(2001). SOURCES OF NONLINEARITIES, CHATTER GENERATION AND SUPPRESSION IN METAL CUTTING, PHIL. TRANS. R. SOC, LOND. A, 395, 663- 693. WONG, S.V. & HAMOUDA, A.M.S. ,(2003A). THE DEVELOPMENT OF AN ON-LINE KNOWLEDGE- BASED EXPERT SYSTEM FOR MACHINABILITY DATA SELECTION, KNOWLEDGE-BASED SYSTEMS, 16, 215-219. WONG, S.V. AND HAMOUDA, A.M.S. ,(2003B). MACHINABILITY DATA REPRESENTATION WITH ARTIFICIAL NEURAL NETWORK, JOURNAL OF MATERIALS PROCESSING TECHNOLOGY, 16, 538- 544. ZHANG, G. & LU, S. C-Y. ,(1992). AN EXPERT SYSTEM APPROACH FOR ECONOMIC EVALUATION OF MACHINING OPERATION PLANNING. ARTIFICIAL INTELLIGENCE APPLICATIONS IN MANUFACTURING. ZHANG, Y. , WU X., XING Z., HU, W., (2011).ON THE GENERATING INTERPRETABLE AND PRECISE FUZZY SYSTEMS BASED ON PARETO MULTI-OBJECTIVE COOPERATIVE CO- EVOLUTIONARY ALGORITHM. APPLIED SOFT COMPUTING, 11, 1284-1294. ZUPERL, U. & CUS, F. ,(2003). OPTIMIZATION OF CUTTING CONDITIONS DURING CUTTING USING NEURAL NETWORKS. ROBOTICS AND COMPUTER-INTEGRATED MANUFACTURING, 19 (1-2), 189- 199. ZUPERL, U., CUS, F. & MILFELNER, M. ,(2005). FUZZY CONTROL STRATEGY FOR AN ADAPTIVE FORCE CONTROL IN END-MILLING. JOURNAL OF MATERIAL PROCESSING TECHNOLOGY, 1472- 1478. 
work_fn2casszyfgapaec5cjxvik37u	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 221000222 Params is empty 221000222 exception Params is empty 2021/04/06-03:05:45 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221000222 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:45 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_fnploejenbc2raktl46esxm44i	----	PII: 0022-247X(91)90214-K JOURNAL OF MATHEMATICAL ANALYSIS AND APPLICATIONS 159, 55&594 (1991) On the Scoring Approach to Admissibility of Uncertainty Measures in Expert Systems IRWIN R. GOODMAN Naval Ocean Systems Center, San Diego, California 92152 HUNG T. NGUYEN AND GERALD S. ROGERS Department of Mathematical Sciences, New Mexico State University, Las Cruces, New Mexico 88003 Submitted by L. Zadeh Received March 19, 1990 This paper arose from our need to rigorize, clarify, and address fully the results of Lindley’s paper (Scoring rules and the inevitability of probability, Znternat. Statist. Rev. SO, (1982), l-26). Herein, we develop a calculus of admissibility in a game theoretic setting. In the case of an additive aggregation function, it is shown that decomposable measures, such as those used in fuzzy logics, are admissible. Also, the problem of the admissibility of the Dempster-Shafer belief functions is investigated via the concept of random sets. It is shown that the class of admissible measures in a scoring framework depends on the assumptions concerning the aggregation function in use. In particular, for nonadditive aggregation functions, an admissible measure may not be transformable to a finitely additive probability measure. 0 1991 Academic Press, Inc. 1. INTRODUCTION With the advent of Artificial Intelligence and the development of expert systems, a number of schools of thought has arisen concerning how uncer- tainties in complex real-world situations are to be modeled. In addition to probability in all of its variants [34], approaches to uncertainty modeling now include the Dempster-Shafer theory of belief functions [26], Zadeh fuzzy logic [32], and nonmonotonic logics [20]; a survey of these approaches is in [28]. Roughly speaking, the choice of a set-function to model the uncertainty involved in a problem at hand is related to the pragmatic aspects or, at a deeper level, to the semantic nature of the type of uncertainty under consideration. Lindley [16, 173 proposed a simple but novel approach, extending DeFinetti’s original considerations on coherence of uncertainties to a more 550 0022-247X/91 $3.00 Copyright 0 1991 by Academic Press, Inc. All rights of reproduction in any form reserved. SCORING APPROACH TO ADMISSIBILITY 551 general setting, for judging the usefulness of different competing uncertainty measures. DeFinetti’s original work [6] may be viewed as a two-person zero-sum game, played between player I, “nature” or “the master of ceremonies,” and player II, “decision-maker” or “you” or “bookie” as denoted variously in the literature [8, 13, 221. DeFinetti’s uncertainty game has great appeal: it is determined by the cumulative amount of the bets. The concept of admissibility in DeFinetti’s game is in fact a type of uniform local admissibility (see also [ 141) which is commonly expressed as the coherence axiom. DeFinetti’s chief result is that the only coherent uncertainty measures are finitely additive conditional probability measures. Lindley’s main contribution to the situation was to investigate DeFinetti’s game by replacing the squared loss functions by a more general score function. For the most part, DeFinetti and Lindley assumed addition for the overall loss function (or aggregation function). Lindley’s chief results are as follows: (i) If an uncertainty measure p is admissible with respect to a score function f, then p can be transformed into a finitely additive conditional probability measure via a known transform depending on f, say Pf; (ii) Within the class of score functions f such that Pf is increasing, the necessary condition in (i) is also sufficient. Roughly speaking, an admissible uncertainty measure has to be a func- tion of a probability measure, i.e., one cannot avoid probability ! However, note that such an admissible uncertainty measure need not be a probability measure ! (See also the axiomatic work of Cox [S].) (iii) As implications from (i), the Dempster-Shafer belief function, Zadeh’s possibility measure, confidence values and significant statements are all inadmissible ! The purpose of our work is threefold: (a) To analyze Lindley’s results and implications, for which we recast Lindley’s somewhat informal arguments and concepts totally within a game theoretic setting. It is pointed out in this paper that DeFinetti in his earlier [6] more restricted work and later, Lindley [ 161 in his generalization of DeFinetti’s efforts, both tacitly assumed: (I) “measure-free” conditional events exist independent of any par- ticular choice of probability measure, but are compatible with the usual evaluation of conditional probability. (II) The usual (unconditional) event indicator function can be extended to be well defined upon conditional events. 552 GOODMAN,NGUYEN, AND ROGERS (III) A natural conjunctive chaining relation holds between condi- tional events. As a consequence of the above assumptions, in carrying out the analysis here, basically two cases are considered apropos to choosing an uncertainty measure in the DeFinetti-Lindley uncertainty game: (1) all finite sequences of conditional events, where each such sequence possesses a common antecedent-this includes as a special case all finite sequences of (uncondi- tional) events, by identifying unconditional events as conditional ones having a universal antecedent-and (2) all finite sequences of conditional events with possibly differing antecedents-this case obviously includes the first as a special case. The structure of uncertainty games is rigorously spelled out in Section 2. In Section 3, a calculus of admissibility, from an analytic viewpoint, is developed for games with arbitrary aggregation functions. In Section 4, uncertainty games with an additive aggregation function are considered in detail together with a Bayesian analysis. (b) To show, contrary to Lindley’s conclusions outlined in (iii) above, that there are rather large classes of nonadditive uncertainty measures, such as belief functions and decomposable measures in fuzzy logics, which are admissible. Also, Zadeh’s max-possibility measures are shown to be uniform limits of admissible measures (Sections 5 and 6). (c) To study the effects of the assumption of additive score functions, we present, in Section 7, various examples of non-additive aggregation functions. These illustrate the fact that the class of admissible measures in a scoring framework depends heavily on the nature of the aggregation functions. In particular, there exist aggregation functions such that admissible measures cannot be transformed into finitely additive probability measures (as opposed to the case of the additive aggregation function in Lindley’s work). In summary, by formalizing Lindley’s work within a general and rigorous game theory framework, we develop a calculus of admissibility which can be used to compare competing uncertainty measures in Artificial Intelligence. We shed light on controversial conclusions concerning the inadmissibility of some well-known uncertainty measures and on the position of the “inevitability of probability.” 2. STRUCTURE OF UNCERTAINTY GAMES In this section, we will formalize Lindley’s scoring approach in a game theory setting. Since the scoring approach involves concepts such as SCORING APPROACH TO ADMISSIBILITY 553 “conditional events,” uncertainty measures, score functions, aggregation functions (implicitly), and admissibility, we need to define these terms rigorously. 2.1. Conditional Events Let Q be a set and d a Boolean (or 0~ ) algebra of subsets of 52. Elements of d are called events. The set complement of A in 52 is denoted by A’; the intersection of A, B in 52 is AB; and their union is A u B. For A E&‘, the indicator function has values ifoEA if cuEA’ In certain forms, it is simpler to identify A with I, so that A = 1 if A “occurs” and A = 0 if A “does not occur.” For A, BE d, the “measure-free” conditional event A ) B is defined by DeFinetti [6] (see also [22]) as the restriction of I, to B, i.e., (AlB)(o)= 0 i 1 if weAB if UEA’B (undefined) if o E B”. Except when B = Sz and (A IQ) is identified with A, these conditional events are not elements of d. The above definition implies the invariant form (A 1 B) = (ABI B). Assume, also the fundamental homomorphic-like forms compatible with any fixed antecedent conditional probability (A I B)” = (A’1 Bh (AuC)lB)=(AlB)u(CIB), MCI B) = (-4 I @(Cl B). Although DeFinetti recognized the potential use of measure-free condi- tional events, in obtaining his key results a formal calculus of relations was not developed. (Again, see [6, especially Vol. 1, Chap. 4, Vol. 2, pp. 266 et passim to 3333.) However, DeFinetti and Lindley implicitly recognized the natural conjunctive chaining relation among conditional events mentioned earlier. (Specifically, see the remark at the end of Section 2.3 and Theorems 554 GOODMAN, NGUYEN, AND ROGERS 3.23 and 4.21’.) For a general treatment of conditional events, see [24] or cw Goodman and Nguyen have derived a full calculus of operators and relations extending the unconditional counterparts for boolean algebras of events to the conditional case. In addition, a wide variety of desirable mathematical properties of these entities have been proven to hold based upon a minimal set of elementary assumptions, including the tacitly assumed conjunctive chaining relation mentioned above. In the context of uncertainty modeling, the uncertainty of an event A is, in general, assigned on the basis of additional information, another event B. But, a priori, conditional uncertainty measure p need not be a probability so that an algebra of measure-free conditional events has to be investigated as a domain for p. Now let d be the class of all conditional events, i.e., d= ((A(B BE&~. By the identification of (A 10) with A, we obtain d c 2. For any set X, and n 2 1, X” denotes the product space Xx Xx . . . x X (n times). We will use the notation A’, to denote the space of all finite n-tuples {a,: f, = (x,, . . . . x,,), xi E A’}. Unless otherwise indicated, “xi E x” and the like will always mean xi, . . . . x, for a generic n. For Al, = (A 1) . . . . A,,)EJP, we identify A, with its indicator function A,(o) defined as (A,(o), . . . . A,(w))E (0, l}“. For example, as(o) = (1, 1, 0, 0, 0,O) indicates that A,, A2 occurred but A3, Ah, A,, A, did not occur. Similarly, we identify with the function a,: B + (0, 1, u}” having values -&w = ((4 I F,)(o), .*a, (J% I J-n)(o)) E (0, 1, q. 2.2. Uncertainty Games We proceed now to formalize a special class of games called uncertainty games. Roughly speaking, these are triples (A,, A,, L) in which A I is a set of configurations (or realizations) of finite collections of (conditional) events, A, is a collection of “set’‘-functions representing “uncertainty measures,” and L is a real-valued loss (or penalty) function. Specifically, A, = Jm x Q. This set A I is regarded as the space of all possible “moves” or “pure strategies” of player I. Next, fix, once for all, four real numbers, a2 < a, < a, <a,; let A,= {,u: P: d+ [a,, aJ} = [a*, a3]“l. Each element of A, is a map, SCORING APPROACH TO ADMISSIBILITY 555 which assigns a number, describing its uncertainty, to each (conditional) event. A2 is regarded as the space of “moves” of player II. Consider now the choice of loss function. This is carried out in two stages as the composition (aggregation) of other (score) functions. As in Lindley’s paper, we call any function f: [a,, a3] x { 0, 1, U> -+ R (the real numbers), a score function if the following are satisfied: (i) For each Jo (0, l}, f(., j): [a,, a31 -+ If3 satisfies the following “regularity” conditions: f( ., j) is continuously differentiable, with a unique global minimum in [a,, a,] at aj, (decreasing over [a,, ai] and increasing over Cq, ~~1); (ii) .f(x, 24) = 0, Vx E [a,, ax]. For example, we may interpret the score function as follows: player I has selected A E & and o E Q and player II has selected p E AZ, then the score for player II is &(A), 1) if A happens to occur, f(p(A), 0) if A does not occur. Now we need to extend f as a map from the space {(a,, 2,): n B 1, jZ.,E [a,, a31n, i,E (0, 1, 24}“> to the space IR,: for each n 2 1, J?,, = (x,, . . . . x,), i, = (tl, . . . . t,), fcL~,) = (f(XlY t,), *..9 fkY, 47)). Similarly, each uncertainty map (or measure) p: d -+ [a,, a,] is extended to a map from ZZ& to [al, a3100 as follows: for each n 2 1, i,=(A 1, . . . . A,) Ed”, c((A^,) = (AA I), . ..t AA,)) E [a,, dn. An obvious way of combining individual scores &(A ;), Ai( (i’ 1, 2, . ..) n) to obtain the total score is using addition on R, i.e., take Lf, +(a,, w, p) = Cr= 1 f(p(Ai), Ai(o Thus the loss function L,., + depends on two functions: f (score) and + (aggregation), This special case will be referred to as the additive aggregation case. In general, by an aggregation function, we mean a mapping ~9: IR, --t R such that (a) II/ is continuously differentiable in all of its arguments; (b) 1+9 is increasing in each of its arguments. (c) $(O,,) = 0, Vn 2 1, where 0, is the zero vector in R”. Note that the additive aggregation function is generated by ordinary addi- tion on R: II/ = + is equivalent to the sequence of functions (g,, n 2 l), where g,: IR’ + R, g(x,, . . . . xn) = CF= 1 xi. Similarly, we identify an aggrega- tion function $ with the sequence (+,, n 3 l), where (I/, is the restriction of II/ to R”. Note also that, while the additive aggregation function is 409’15912-IX 556 GOODMAN, NGUYEN, AND ROGERS symmetric, there is no a priori reason to impose such a condition on arbitrary aggregation functions. Now, given a score functionf and an aggregation function II/, we define the loss function Lf,, as follows: (Note that fand ,u are used in the extended sense.) The triple (A 1, AZ, Lf,$) is called an uncertainty game and is denoted by Gf,,. In DeFinetti’s game [6], a, = a2 = 0, a, = a3 = 1, and fb,A= {b”-i” ; ;I$ndefined); here $ = + . In Lindley’s extension of DeFinetti’s game, uj, j = 0, 1,2, 3, are not restricted to [0, 11, $ = + , and f is an arbitrary score function. 2.3. Subgames To formalize various concepts of admissibility in an uncertainty game Gr,$, we first introduce the concept of subgame. Suppose we are interested only in some given finite collection of condi- tional events, say a, then we need to look only at the subgame where and For example, we can view A,E$ as a set a,= {A,, . . . . A,} se. Then [a*, a31An= {p: {A,, . . . . A,} --* [az, a,]} so that each p in [a,, ujlAn is the restriction of a p in [a,, a3]“. In a subgame, the finite collection of conditional events in a is specified; if player II chooses p to express an uncertainty about these conditional events, then the overall loss would be Lf,ti,,dw ~1. The subgame Gf,+.,i is regarded as a game with partial infor- mation, namely player II does know that a is to be considered. For example, if a=((ElF), (E”(F)) and $= +, ~(u)=((EIP)(~), (EC1 F)(o)), and the set of configurations of a giving rise to non-zero losses (f (x, u) z 0) is {(EIF), 11, ((E”IF), 01, ((EIF), O), ((E”IF), l,}. SCORING APPROACH TO ADMISSIBILITY 557 For is {(AO), (0, l)}, 2 = (x1, x,), where x1 = p(EI F), .x2 = p(E’I F), we have two overall scores: f(xr, l)+f(Xz, 0) (if EIF occurs) and f(X,? 0) +f(x*, 1) (if E” 1 F occurs). It should be noted that so far only those finite sequences of events have been considered which have a common antecedent. Relevant to this, denote for any Fe ZZZ’, d-fF= {(EIF): EEL}. The reason for this is that if one wished to determine the possible indicator evaluation combinations among any sequence such as ((E, I F,), (E, / F2), (E, I F3)), until recently, no standard technique existed for dealing with this issue. However, Goodman, Nguyen, and Walker [ 121 and Schay [24] have proposed independent nonstandard syntactic approaches for treating this and related problems. Only one such situation will be considered in this paper, namely the fundamental natural conjunctive chaining relation, (El FG)(F( G) = (EFI G), which is obviously true for all corresponding conditional probability evaluations AEI =I .PV’I (3 =PPI G) for p(G) > 0. More details of this will be seen in Theorem 3.2.3 et passim. 2.4. Equivalent Reduced Forms of Games and Subgames The space /1 r = J&, x 52 in the game G/,@ is infinite, in general, but for each R E Jm, the space of configurations of a, namely a(Q) is finite, a sub- set of 10, 1, u}lA^‘, where Ial denotes the “dimension” of 8; e.g., if A E d”, then IAl = n. On the other hand, the domain of Lf,$ involves a, but since Lf,JA, ., p) is constant on each (A)-’ (t), t E A(Q), 8 can be replaced by the finite d-partition (of Q) generated by a. Specifically, for each a E Jm, say A = ((E, IF,), . . . . (E, I F,)), consider the finite collection of events %(A^)= {E;F,, E;F;, F;, i= 1, . . . . n}. Let z(a) denote the canonical parti- tion of Q generated by +?(A) (which reduces to A^ = { Ei} when all F; = Q. Then, rc(a)= {B,,j= 1, 2, . . . . 23n}, where each B, is of the form nr=, 02, Dk E %‘(A^), sk = 1 or c; 0: = Dk, 0: is the set complement of D,. Also, each D, is a union of the Bis. (See [23, p. 12-153.) Note that because the E’s and F’s are not necessarily distinct, the cardinality of z(a), say m = ITCH, is most often ~2~“, but at least 2. The cardinality of rc(a) may be less than n = I Al as we now demonstrate. 558 GOODMAN, NGUYEN, AND ROGERS Let Then a = (EF, E”F, F”, F). SF@) = {Dl = EF, D, = E”F, D, = F”, D, = F}. Only the following configurations of “occurrences” can arise so that m=l7c(A)l=3<4=lAj: 1 0 0 1 for B, =EF, 0 1 0 1 for B, = E”F, 0 0 0 1 for B,=F”. Thus, B’BcBC=EF 1 2 3 BCBIBC=ECF 1 2 3 B;B;B,=F” B1B;B;uB;B2B;=F. Of~ourse,m=3>2when~=((ElF),(Ec~F))for~(~)={EF,EcF,FcJ= W ). The (equivalent) reduced form of GY,$ is G:, = (A:, AZ, L&l, where where n:={(a,B):a~~~,B~.(a)}, I I a Lj!&k 4 ~0 = W(P(A), A(B))), a(B) = t E (0, 1, u}‘“’ -=-B={o:&)=t}, i.e., /i(B) = &II) for any choice of w in B. Similarly, the (equivalent) reduced form G,,,, A is G&A = (Ata, Az,a, L&A), where ny,=?T(A) SCORING APPROACH TO ADMISSIBILITY 559 and 3. ANALYTIC STUDY OF ADMISSIBILITY In this section, we will introduce various concepts of admissibility for Gf,$ and then develop analytic techniques for (weak) local admissibility with arbitrary aggregation functions $. This also extends Lindley’s results in the case of an additive aggregation function, namely, giving sufficient conditions for (weak) local admissibility. For related works in Statistics, see [4, 143. For the concept of Pareto optimality, see [2, 271. 3.1. Concepts of Admissibility In this subsection, we spell out relevant forms of admissibility in the reduced form of the game First, ,LL E A, is (ordinary) admissible with respect to GTIL if there is no VIZ,~, such that Lz$(A, B, v) <,!,;+.(A, B, p) for all (A, B) E /1: with the strict inequality for at least one (a, B). Similarly with respect to the subgame Gzti,~, PE/~*,A = [a,, as]” is a-admissible (A-AD) if there is no v E AZ,2 such that for all BE ~(2) with strict inequality for at least one B. More generally, let 8 belong to the power set 9(Jm). Then p E A2 is b-admissible with respect to _Gf:+ if p is a-admissible for all 2 E 8’. (Note that A2,~ E A,.) When d = SZ!~, p is uniformly admissible. It is easy to see that uniform admissibility implies ordinary admissibility. Continuing with the subgame where 2 is fixed, we note that p(A) E IR” for which there is the usual topology based on the Euclidean norm 11 /I. Thus we may have a neighborhood of p WP,~)= {vEA~,A: Ilv(~)--~(~)ll <r}. Then p E Ax,2 is A-focal admissible (A-LAD) if there is some N(p, r) such that (1) does not hold for all VEN(~, r). For the last two concepts of admissibility, it will be convenient to introduce the following notation. 560 GOODMAN, NGUYEN, AND ROGERS With /i = (A 1, . . . . 4), let {B,, . . . . B,) be a listing of the elements of z(a). The set of values Lj&i(B,, PL)= W(P(~~), A,(B,)), . . . ..fWL)~ -MB,))), . . . . L&-i(L PL)= Icl(f(A~~), A,Mn)), . . ..f(~(An). 4,(&))) can be thought of as the value of a transformation L;$,A: [a*, u31n + Iw” at the point ,(a) = @(A,), . . . . ,u(A,)). For simplicity, we drop the extra symbols and write, in general, L(a) L(i)= ; [ I JLO) for 2 = (x,, . . . . X,)E [a,, a,]“=D. By the natures of $ andf, L can be arranged in the partitioned form L(l) il 1 L(2) ’ where L(‘) is constant on D but L(‘) (of length k) is not constant in any neighborhood in D. Of course, L”’ may not appear. Then p E n2,~ or equivalently i E D, is a-weak/y admissible (A-WAD) if there is no 9 ED such that L”‘(9) cs L(‘)(Z), where cs is the strong (Pareto) order in real Euclidean space Rq: P = (PI 9 **., P,) cs (v, 3 . . . . vq) = v^ if pj < vj for all j = 1, 2, . . . . q. If pj < vi for all j= 1, 2, . . . . q we write simply /i Q 9; when the inequality holds for only some j, we write fi <,,, t. It is easy to see that A-AD is stronger than A-WAD; for, if there is a 9 ED such that L”)(p) cs L(‘)(a), then since, if LC2) appears, it is constant on D. Moreover, the admissibility considered informally by Lindley turns out to be “weak-local” and will be considered further in Section 4. Finally, 2 is weak-local admissible (A-WLAD) if for each 9 in R” with (lyJ/ = 1, and each c( > 0 there is an r = r(Z, $, a) > 0 such that there is no t E (0, r) for which L”‘(?i++g)-L”‘(.?),< -atl,. (2) Here 1, is a k by 1 vector all of whose components are 1. SCORING APPROACH TO ADMISSIBILITY 561 Note that -at 1, cs 0 in Rk. It is convenient to refer to such j as a direction. Then, locally, z? cannot be “beat linearly in any direction.” Of course, A-WAD implies &WLAD: if (2) holds for some j, a, t, then L”‘(i + tf) <s L”‘(3). We close this section with a theorem which gives global and local equivalences under the additional hypothesis that L is convex, that is, each component of L is convex. THEOREM 3.1.1. Let L be convex. Then (i) a admissibility is equivalent to A local admissibility; (ii) A-WAD is equivalent to A-WLAD. Proof Since it is obvious that admissibility implies local admissibility, we consider only the converses. (i) In terms of L, .? is not A-admissible if there is a i, in D such that L(9) <w L(Z). (3) Convexity of L means that for each t in (0, l), L@+t(j-a))<(l-t)L(i)+tL(g). Combining this with (3), we obtain L(Z+ t(j-a))<, L(i). For any r > 0, choosing t = r/(r + jl j - ill ) makes /) t(j - a)\\ < r whence .? is not J-LAD. (ii) If i is not A-WAD, there is a j such that L(‘)(p) cs L(‘)(i). Then a=min(Lj”(f)-Lj’)(j),j= l(l)k} >O and L~l~(~)<L~l~(~)-cclk. Combining this with convexity, we obtain for all t E (0, l), L”‘(.? + t(j- a)) = L”‘(( 1 - t)i + tj) d (1 - t) L”‘(2) + tL’l’( j) <(l-t)L’i’(i)+t(L(‘)-Ml,) =L”‘(i)-atlk. Therefore, L”‘(~++(+c?))-L”‘(~)< -atl,<,O for all tE(0, l), in particular, for t =s/llj-a\l < 1, t(p--i) =si, where llill = 1 and 2 is not a-WLAD. 1 All of the above admissibility concepts-for unrestricted d-can be 562 GOODMAN,NGUYEN, AND ROGERS modified in the obvious way for restrictions such as requiring only finite sequences of conditional events, each having a common antecedent. 3.2. Some Characterizations of A^- WLAD Since WLAD involves only the nonconstant L(l), we simplify by writing this as L. By regularity conditions on $ and f, L is differentiable. We denote the k by n matrix of partial derivatives at f as J= (8Li/8xj); we also take 2, 9 as column vectors. Then we have the following. THEOREM 3.2.1. i is WLAD if and only if there is no p such that Jjj cs 0. Proof If there is a 9 such that Ji, <s 0, then for the direction i = c//II 311, Jz^ <s 0. Differentiability yields For each component Lj there is an aj such that Lj(? + tjg - L(i) < - ajtj with tj in a neighborhood of zero. Take t in the intersection of these neighborhoods and a = min { c1i, . . . . ak}. Then L(Z + t.f) -L(g) < - at l,, which makes 2 not WLAD. Conversely, if 2 is not WLAD, then there is a direction 9 and an c1> 0 such that for all r so there is a t, < r for which L(2+ t,jq- L(i)< -E&l,. Since t,+O as r-+0, J$<,O. 1 Recall that for 2 = (A,, . . . . A,), PE [a,, a31A is identified with P = (X,) ..*, x,) E D = [a,, asIn. In view of Theorem 3.2.1, the analytic study of weak local admissibility of p or 2 is reduced to finding conditions on J(a) so that there is no 9 E Iw” for which J(Z). $ cs 0. It is well known that the solutions of a system like Ji, = i depend heavily on the rank p of J. Also, the columns of J and the rows of J, 9 and i can be permuted without changing the rank or character of the solutions; these can also be partitioned. In the following, we assume that this has been done so that when the rank is p, the system is where J, is p by p and nonsingular, C is p by n-p and R is k-p by p. Of course, if p = n, the columns involving C do not appear and if p = k, the rows involving R do not appear. The following theorem contains several results relevant to this work. SCORING APPROACH TO ADMISSIBILITY 563 THEOREM 3.2.2. (a) Zf p =n = k, then i is not WLAD. (Indeed, if J is non-singular, then the system Jy = i has a solution j for every i, in particular 2 <,s 0.) (b) Zf n = k and .? is WLAD, then det J= 0. (cl ,%? is WLAD zf and only zf there is no i E IV’ such that J, [ 1 RJI i<,O. (d) Zf p = 1, then 2 is WLAD if and only if the vector has both positive and negative components or at least a zero component. (e) In case k = n = 3, j? is WLAD if and only if det J = 0 and either (i) p = 1 and J, [ 1 RJ, has a zero component or contains both positive and negative components or (ii) p=2 and R<O. Partial Proof Indeed, if det J= 0, and p = 1 with the above specified structure of J1 [ 1 RJ, ’ then i is WLAD as in (d); if det J= 0 and p = 2, then if 3 is not WLAD, there will be j E [w* such that i.e., J, j = f cS 0 and Ri cS 0 which is only possible if R < 0. Conversely, suppose that i is WLAD. Then first, det J = 0 as in (b); thus p = 1 or 2 since J z 0. If p = 1, and i is WLAD, we have the above specified structure of 564 GOODMAN,NGUYEN, AND ROGERS by (d). If p = 2, and 2 is WLAD, we have by (c): there is no 9 E I@ such that JI [ 1 RJ, f<,O, i.e., there is no G cS 0 such that R6 -q 0, and hence R 6 0. 1 COROLLARY 3.2.1 (Corresponds to [16, Lemma 11). Let a= {(EIF)}. Then p(EIF)=x is WLAD ifand only ifxe [a,, a,]. Proof: Here J= Yw(4 1 )I f’k 1) 1 IC/‘(f(X? 0)) f’k 0) * Obviously, for XE [az, a31 - [a,, a,], the components of J are nonzero with the same sign, and hence x is not WLAD. For XE [a,, a,], J has at least a zero component, or two nonzero components with opposite signs, and thus x is WLAD. Remark. In view of Corollary 3.2.1, from now on, the range of uncer- tainty measures is restricted to [a,, al]. COROLLARY 3.2.2 (Corresponds to [ 16, Lemma 23). Let A^ = {(El F), (E’IF)}, then Z= (x, y), x=,u(E(F), y=p(E’JF), is WLAD ifand only if det J=O. Proof: Here J= [ u-f(x, l), f(Y, 011 .0x? 1) cm, 11, f(Y, 011 S'(Y, 0) ti’Cf(X~ Oh f(Y, 1 )I f’CT 0) Ic/‘Cf(& 01, f(Y, 111 S'(Y, 1) 1 . The condition is necessary by Theorem 3.2.2 (b). Suppose det J=O, then p = 1. The sulkiency follows by Theorem 3.2.2 (d). 1 COROLLARY 3.2.3 (New Result). For a = {(E, 1 F), (E2 IF), (E, u E, IF)} with E, E, = 0; set x = p(E, ) F), y = p(E, I F), z = p(E, u E, I F) with x, y, z E [a,, a,], 2 = (x, y, z). Take t,b = + . Then the nonzero losses are fk l)+f(Y9o)+f(z, 11, .mo)+f(Y, l)+f(z, l), f(x,O)+f(y,O)+ f(z, 0), and hence f’(x, 1) f’(Y, 0) f’k 1) f’k 0) f'(v, 1) f’k 1) . S’(4 0) f'(Y, 0) f’k 0) 1 SCORING APPROACH TO ADMISSIBILITY 565 With respect to the game G, +,A, the following are equivalent (i) 2 is A-WLAD, (ii) det J= 0. Proof. (i) j (ii) by Theorem 3.2.1. (ii) j (i). Since det J = 0, the rank p of J is either 1 or 2. (a) If p = 1, then the first column of J is f ‘(a,, 1) [ 1 0 with f’(a,, 1) < 0 when x = a, 0 and is with f’(a, , 0) > 0 when x = a, ; for XE (a,, a,), this column has both positive and negative components. Hence 1 is &WLAD. (b) If p = 2, permute the second and third columns of J to obtain the partition where J = f’(x, 1) f’k 1) l C f’h 0) f’k 1) 1 and RJ, = (f’(x, 0), f’(z, 0)). Then det J,=f’(x, l)f’(z, l)-f’(x,O)f’(z, l)>O when z #a,. In this case, we have R = 0) .I-‘(% 1) -S’k 0) f’k 0) f’k 1) f’k 0) -S’(x, 0) f’(Z, 1) det J, > det J, 1 with f’(x, O)[f'(z, 1) - f’(z, 0)] 6 0. Look at f’(x, 1) f’(z, 0) -f’(x, 0) f’(z, 1). (*) 566 GOODMAN, NGUYEN, AND ROGERS By hypothesis, det J= 0 and this is equivalent to P,(z) = P,-(x) + P/(y), where P,-: [ao, a,] -+ [0, l] is given by f ‘(x9 0) pf(x) =f’(x, 0) -f’(x, 1)’ Thus Pf(x) <P,(z) which in turn implies that (*) < 0. When z = a,, the first and third columns of J become Consider J, [ 1 RJI ’ where 0 Id 1 0 . f'(Z130) J= f’(%l) 0 ’ 1 f’(40) f’(&,O) 1 (**I and RJ1 = (f'(x, 0), 0). We have det J1 =fl(x, 1) f'(a,, O)<O for xfu,. In this case R- -( f'(x3 l)f'(ulJvO det J, with f'(x, i)f'(u,, O)<O. Finally, when x = z = a,, the first and third columns of J are Let 0 f'(u,,O) f'(u,,O) I and RJ, = (0,O). Then, det J1 = [f ‘(a,, O)]* > 0, and R = (0,O). 1 The following result is actually an application of Corollary 3.2.3. However, as it is basic for most of the rest of our work, we state it as a theorem. From now on, PE [a,, u,]~. We say that: p is g2-WLAD if p is a-WLAD for 2 of the form { (E( F), (EC ( 8’)); SCORING APPROACH TO ADMISSIBILITY 567 .H is &&WLAD if p is A-WLAD for a of the form ((E, I F), L% I F), (El u J% I F)} with El 4 = fzr. Also, P,-(x) is always given by (* * ) above. THEOREM 3.2.2’ (Equal Antecedent Counterpart of [16, Lemma 51). Suppose $ = + . Then the following are equivalent. (i) p is &‘* and &- WLAD, (ii). Prop: ~4~4 [0, l] is afinitely add‘ ltive probability measure, for all FE&, where ,Q1’F=d ((E(F): EEL}. Proof. (i)= (ii). Since p is gZ-WLAD, we have, for any E, FE&‘, p(EI F) = x, ,u(EC 1 F) = y, det J= 0, where J= [ f'h 1) f'(Y,O) f'(-%O) f'(Y, 1) 1 (1) then, P,(x) + P,-(y) = 1. In particular, for F= Q, we have Pfop(E) + P,+(EC) = 1. Next, since p is &s-WLAD, for any E,, E,, FE d with E, E, = a, we have, det J = 0, where, x = p(E, 1 F), y = p(E, I F), z = p(E, u E, 1 F) and so that Pr(z) = Pf(x) + Pr( y), by computation. In particular, for F = Q, PpdE, u Ed = Pp/.@,) + P,oP(&). Thus Q = P,o p: &- + [0, l] is a finitely additive probability measure, for all FE&. (ii) =z- (i). First note that p: & + [0, 11, (ii) means that the restric- tion of p to &F (still denoted by p) is such that Pfop is a finitely additive probability on JZ?~, say QF = P,o p, for all FE .r4. Thus which means det J = 0, where J is as in (1). The rank p of J is therefore 1. Since x E [a,, a,], each column of J contains a zero component or has 568 GOODMAN,NGUYEN, ANDROGERS both positive and negative components, and hence (x, y) is a-WLAD, VA = {(EJF), (E’)F)}, i.e., p is $-WLAD. The fact that p is &WLAD follows from Corollary 3.2.3, since for any E,, E,, FE -01, with E, E2 = a, the condition </o/W, u E, I F) = f’,” ~(4 IF) + +P(E, IF) is equivalent to det J= 0, where J is as in (2). 1 In the proof of Theorem 3.2.2’, it might be tempting to conclude that the conditional probability is uniquely compatible with each QF. But this is not necessarily so. In fact Aczel Cl, pp. 321-3241 required additional properties before proving such as relation. These properties include continuity and monotonicity of functional forms in both antecedent and consequent probabilities. The appropriate strengthening of Theorem 3.2.2’ to account for possibly varying conditional event antecedents utilizes the following property: For any P E La,, a, 19”, call p &-WLAD, if p is A-WLAD for A^ of the form ((El FG), (FI G), (EF( G)), assuming the natural conjunctive chaining relation that both DeFinetti and Lindley implicitly assumed (see earlier discussions), (El FGWI G) = (4 G); all E, F, G E RI, and interpreted via DeFinetti’s conditional event indicator function. THEOREM 3.2.3 (Corresponds to [16, Lemma 51). The following statements are equivalent under the above assumption and for II/ = + : (i) p is g2;, 6”, &WLAD. (ii) Pro ,u: SCI~ + [O, I] is finitely additive conditional probability measure for each FE &, i.e., assuming p(F) > 0, for all E E ~4, (+PMEI F)) = U’p/WlF) = f’,MWYPfMF)). Proof: Follows a similar format as for the proof of Theorem 3.2.2’, where now, in addition to Eqs. (1) and (2) holding when (i) is assumed, one has due to p being G$‘,-WLAD, 1) f’(h 1) f’(w 1) J= 0) f’(o, 1) f’(w, 0) 1 , (3) f’(u, 0) S’(w, 0) where u=p(EJFG), v=p(FIG), w=p(EF(G). 1 SCORING APPROACH TO ADMISSIBILITY 569 4. GAMES WITH AN ADDITIVE AGGREGATION FUNCTION AND BAYESIAN ANALYSIS This section is devoted to a detailed investigation of games with an additive aggregation function; to be complete, we consider also their mixed extensions. Under an additional assumption on the score functions, we establish the equivalences among different concepts of admissibility, including Bayesian decision functions and DeFinetti’s coherence measures. We also show that nonatomic probability measures are not “coherent” in games with improper score functions. 4.1. Mixed Extensions of Uncertainty Games Consider the game G,+ in its equivalent reduced form (A :, A 2, Lf+) with Because of the nature of /i:, mixed strategies or prior probability measures on A: will be defined as follows. Player I will first pick an A^ E J& according to some probability measure 4 on (s&,, SJ), where 3 is some a-algebra of subsets of s&, and then depending upon a, pick a configuration of occurrences of A^ (as a finite collection of conditional events, equivalently) an element of the partition n(A), according to some probability measure rA on the power class 9y7c(A)) of 7c(A). Now, since ~(a) is finite, each probability measure on 9(x(a)) is identified by its probability density function on X(A), i.e., 6: {B,,j= 1, . . . . m} -+ [0, 11, f QB,)= 1, j= 1 where Bis are some listing of the elements of n(a), and m = In(a Next, each such 8 generates a probability measure P, on (Sz, &) such that PdBj) = e(Bj), Vj = 1, . . . . m. Indeed, for each j = 1, 2, . . . . m, let Pj be a probability measure on B,, i.e., on the a-algebra trace {A Bj: A E &} with Pj( Bj) = 1. Define, for A E A’, P,(A) = CJ’= , O( Bj) Pj(ABj). Note that PO(A_) is completely determined by 8, Vi = 1, . . . . n, where a = (A r, . . . . A,), n= IAJ. Then, since each Aims? in general, the probability measure P, on d is extended to d via the condi- tional probability operation. For a rigorous treatment of this extension, see c121. 570 GOODMAN, NGUYEN, AND ROGERS If (X, 9) is a measurable space, we will denote by P(X, X) the collection of all probability measures on it. The collection of probability density func- tions on z(a) is denoted by P(rc(a)). Thus, we have, by identification, P(7Q)) c P(Q, a). The space of prior probability measures on Ai+ is The expected loss of p E AZ, with respect to a prior (q( .), z.( ., +)) is so that the mixed extension of G;, is Similarly, the mixed extension of the subgame CT,,, is where ~~,+,a(& PL) = Cj$’ Lf,S,A(Bj, P) NBj). Now, we view the subgame GTti,~=(n(A), [a,, uSI’, L!$,J) as a statistical game coupled with the random variable X whose distribution P, depends on BE n(a); when X is constant, on each BE n(a), the risk is L(B, p)+ On the other hand, since ~(2) is finite, standard results from decision theory (for finite games) hold for Gzti,A (see, e.g., [7, 31). Then the risk set L(D) is closed and bounded since D is compact in R” and L is continuous. For convenience, we state below some definitions and basic results. (i) ,U~E [a,, al] A is Bayes with respect to a prior distribution r on .(A) if which is the minimum Bayes risk; E, denotes the expectation with respect to t. We write ,nL, for the Bayes uncertainty measure with respect to r. SCORING APPROACH TO ADMISSIBILITY 571 (ii) r0 is a least favorable prior if infE,,L(.,~)=supinfE,L(.,~) P 7 p which is the lower value of the game. (iii) %?L ~cz,, a,ld is complete if for each ALE [a,, a,]“--%‘, there is v E %? which is “better” than p, i.e., L( ., VI cw L( ., P) (A-AD). V is minimal complete if %’ is complete but no proper subclass of 5%’ is complete. (iv) For admissibility of Bayes rules in G~,.A, we have (a) If pL, exists uniquely up to equivalence (p - v if and only if L( ., p) = L( ., v)), then pL, is admissible (A-AD). (b) If the prior r is strictly positive (i.e., r(B) > 0, VBE n(a)), then pz exists, and is A-AD. (v) If p is A-AD, then p = pr for some r. (vi) The class of Bayes rules is complete, and the class of admissible Bayes rules is minimal complete. (vii) Let 6r~&; then p is said to be b-Bayes if p is Bayes with respect to Gzi,~ for all a E 8’. In particular, p is uniform Buyes if d = s&. Note that p is Bayes with respect to G,!, when the prior is in P(/l,), i.e., of the form r = q( .), r.( ., .). 4.2. Equivalences of Various Concepts of Admissibility In the rest of this section, we consider $ = +. THEOREM 4.2.1 (Equal Antecedent Form of [ 16, Theorem 21). Consider the game Gl + with f such that PY is increasing. Let p E [a,, alId. The following are equivalent. (i) u is uniformly admissible, w.r.t. all finite equal antecedent condi- tional event sequences. (ii) p is & and &-weak local admissible. (iii) u is untform Bayes, w.r.t. all finite equal antecedent conditional event sequences. tttve probability measure, for all PEvL Prop: ~4~ [0, l] is a finitely add’ Proof (i) =S (ii). Obvious by definition. W/l 5912. I9 572 GOODMAN,NGUYEN, ANDROGERS (i) =z. (iii). Follows by standard results of the game GT +, 2, namely if p is A-AD, then ,u is Bayes (with respect to some prior z on .(a)), and conversely, if p is Bayes (p = ,u~), then since ,u~ is unique, pL, is R-AD. The uniqueness of e, (up to equivalence) can be seen as follows. Let ZE P’(a(A)), a = ((EiJFi), i= 1, . . . . n); we have E7L(a, PI= i C fCP(Ai)3 Ai(B)I T(B) i=l Ben(A) =,$, {fCP(ai)9 ll C r(B) +fCPCAih Ol C BE QlW BE Q2(G Q,(i)= {BE7C(A) : BGEiFi) Then Qz(i) = (BE X(A) I BEiF,= a}. since 2 r(B)=l- c t(B). BE PAi) BE Ql(i) Therefore, p(Ai) = P;l[&EQ,cil r(B)] is uniquely determined by r. (ii) * (iv). By Theorem 3.2.2’. (ii)* (iii). Assume (ii). Then (iv) holds. Since Cecnc~) r(B)= 1, r = Prop can be taken as a prior for x(a). Then p = ~1~ and hence (iii) holds. Conversely, when (iii) holds, (i) also holds, and, a fortiori, (ii). 1 THEOREM 4.2.1’ (Corresponds to [16, Theorem 21). Make the same assumptions as in Theorem 4.2.1. Then Theorem 4.2.1 holds with the folIowing strengthening. (1) Omit the constraint “w.r.t. all finite equal antecedent conditional event sequences” in both (i) and (iii). (2) Add the property that p is gh-weak local admissible to (ii). (3) Replace s&., for each FE d by simply d in (iv). SCORING APPROACH TO ADMISSIBILITY 573 Proof: Obvious by inspection of the proof of Theorem 4.2.1 and the role that &WLAD plays in making stronger Theorem 4.2.1 (iv). 1 Remark. For any a E J&, relative to CT + ,A, (i) A least favorable prior rOg p(rr(A)) can be obtained by minimizing the objective function Pf, +,,it59 Pu)’ 1 C fCPtA,), Ai(B)I t(B) i=l Bcsn(A) subject to the constraint CBEkCAI z(B) = 1. (ii) Using the proof of the uniqueness of pr in the above theorem, we can obtain a minimax uncertainty measure p0 E n 2,A, where inf sup P.f, +,A(? PL) = sup PEA2.A’ rEP(Tc(a)) rtP(n(A)) Pr, +,A(? PO). (iii) G T +,A has the value V,(f, a), where Vo(f9 4 = sup reP(n(A)) Pr. +,,.&r, /%I) = P/, +,A(%3 I%J. THEOREM 4.2.2. Consider the game GT + with PJ increasing. Suppose that f is not a proper score function (i.e., P,(x) E x) andf is twice differentiable. Then no nonatomic conditional probability measure p on d can be Gf +- uniformly admissible. Proof. Assume the contrary, i.e., the nonatomic probability p on d is uniformly admissible with respect to GT + . By Theorem 4.2.1, we should have, by the nonatomicity of p, <r(t) + P,( 1 - t) = 1, VtE [O, 11. (1) By hypothesis here, it can be shown Pf(XY) = P,(x) Pf(Yh vx, y E I3 11. (2) Differentiate (2) with respect to x and then separately with respect to y, to obtain YV,, (XY) = (Pf)’ (x) pf(YL X(Pf)’ bYI = PfbNPJ (Y). Simple division yields x[log PJx)]’ = y[log P,(y)]‘, so that x[log PJX)] = c (a constant), and hence I’,(x) = xc. Then (1) forces c to be 1. 1 574 GOODMAN, NGUYEN, AND ROGERS Remarks. (i) To accomodate this situation, we will consider a concept of admissibility in the wide sense of Section 5. (ii) The condition (iv) in Theorem 4.2.1 is the coherence principle of DeFinetti. The equivalences in Theorem 4.2.1 explain the concept of “reasonable” uncertainty measures, from a decision viewpoint, and lead to the discoveries of other “reasonable” measures which need not be probability measures. In view of Theorem 4.2.1, the coherence principle can be used as a definition of admissibility. (iii) Relative to the invariance property of the probability transform Ps, we present the following. Consider two games G,, + , G, + (say, with the same aj, j = 0, 1,2, 3), with Pf, P, increasing. Let p (resp. v) be uniformly admissible with respect to G,, ( resp. G, + ). It is not true in general Pfo,urPgoV. (*I Indeed, let Q,, Q, be two probability measures on &, with Q, # Q2, take ,u=P+Q, and v = Pi’ 0 Q,. For (*) to hold, one needs to consider the uniform admissibility of the (vector-valued) uncertainty measure (,u, v) with respect to the joint uncertainty game GcJ,,,, + = (/ii, ,4:, Lcf,,,, +), where L (/, g), + (A 03 PL, VI = Lf, + (A a, P) + L,, + (4 0, VI. 5. ADMISSIBILITY OF POSSIBILITY MEASURES This section is devoted to the study of admissibility of a class of uncer- tainty measures called decomposable measures and its implications for fuzzy logics. 5.1. Decomposable Measures and Fuzzy Logics Since the concept of (Zadeh) possibility measures and the techniques in fuzzy logics might not be familiar to all, we first present some background (see, e.g., [33, 32, 111). Roughly speaking, fuzzy logics differ from ordinary two-valued logic by their semantic evaluations of logical connectives. For our purpose here, we will focus on the evaluations of the connective “or” which corresponds to the main properties of associated uncertainty measures. A function (operator) T: [O, l] x [0, l] + [0, l] is called a t-conorm (see, e.g., [25]) if T is associative, commutative, and nondecreasing in each argument; also T(x, 0) =x and T(l, X) = 1, VXE [0, 11. A t-conorm T is said to be Archimedean if T is continuous and T(x, x) > x, Vx E (0, 1). SCORING APPROACH TO ADMISSIBILITY 575 Some examples of t-conorms are x if y=O T(x, Y)’ Y if x=0 (not continuous) 1 otherwise T(x, Y) = Max(x, Y) (not Archimedean) T(x, y) = Min(x + y, 1) (Archimedean). For the related concept of copulas in Statistics, see, e.g., [9, 10, 191. An Archimedean t-conorm T has the following representation [ 181: T is an Archimedean l-conorm if and only if there exists an increasing, con- tinuous function g (called the additive generator or generator of T) which maps [0, l] -+ [0, + co] with g(0) = 0, and such that vx, YE lx, 11, T(x> Y) =g*Mx) +gb)L where the pseudo-inuerse g*: [0, + co] + [0, l] is detined by g*(x)= 1 i g-‘(x) if XE CO, g(l)1 if x>g(l). For example, (i) For p>O, Tp(x, Y) = W’ + Y’ - x~Y~)“~ =g;(g,(x) + g,(y)), where g,(x) = - (l/p) log( 1 - xp), g;‘(x) = (1 - eppx)“P =g,*(x); note that g,( 1) = + co here. (ii) For p B 1, T,(x, JJ) = [Min(xP -I- yp, l)]“” has generator g,(x) = xp, 1 x’IP g*(x)= 1 if xE[O, l] if x>l with g,( 1) = 1 here. Since a t-conorm T is associative and commutative, we can extend T to T: co, 11, + co, 11, where T(x) = x, by convention, T(x I, -*-, xx) = Tb,, T(x,, . . . . x,)1, n 2 2. The representation of an Archimedean t-conorm T becomes W,, x2, . . . . x,) =g* 0 1. 576 GOODMAN, NGUYEN, AND ROGERS Now, let (52, J&‘) be a measurable space. A mapping p from & to some interval of R, say [a,, a,], is called a decomposable measure if there exists T: [a,, a31 x [a,, a31 + [a,, a,] such that, for A, BE d with AB= a, p(A u B) = T(p(A), p(B)) (see, e.g., [31]). When such a Texists, it is called the composition law of p. As a generic example of a decomposable measure, begin with any fuzzy set membership function 01: D --, [0, l] and any t-conorm T. Then, if 0 is finite, one can make the extension of a as ,u,,~: P(Q) + [0, 11, where for any AcSZ, ,ua, T(A) = T(a(o) : w E A). (+I P is clearly decomposable. Conversely, given any decomposable measure pa’G.r.t. some t-conorm T, for any finite A as above, Eq. (+ ) holds with h,T=k In fuzzy logics, composition laws are t-conorms (with [a,, a3] = (0, 11). Note that probability measures are decomposable measures with T(x, y) = Min(x + y, 1); this is an Archimedean t-conorm with generator g(x)= -log(l-x) and g*(x)=g-‘(x)=1-e-“, since g(l)= +oo. Indeed, let P be a probability measure on d, and A, BE d with AB = 0. Since P(A) + P(B) = P(A u B) < 1, we have P(A u B) = P(A) + P(B) = Min(P(A) + P(B), 1). Furthermore, in the case of probability measures, the generator g is such that g( 1) = + co (and hence g* =g-r); the corresponding t-conorm T is called a strict (Archimedean) t-conorm. For a t-conorm T, a T-possibility measure is defined to be a map from d to [0, 1 J with T as composition law. For example, (Zadeh) max- possibility measure is a decomposable measure with T(x, y) = Max(x, y). Let B be discrete and a: Q + [0, 11. Let T be an Archimedean t-conorm with generator g. We denote by ,u~,~ the T-possibility measure defined as follows. For A E Q, finite, p,,T(A)= T(a(o), wEA)=g* [I &44)]. A For A countably infinite, PL,,r(A)=g* [z da(w))]. A Note that CA g(a(w)) < + CCL 5.2. General Admissibility under Additive Aggregation In order to discuss Lindley’s conclusions about the inadmissibility of uncertainty measures, we consider G, + . In view of the results of Sections SCORING APPROACH TO ADMISSIBILITY 577 3 and 4, we have to consider the concept of admissibility of a given uncer- tainty measure in a wide sense. Specifically, an uncertainty measure I-L E [a,, aIlJ is said to be general admissible if there is a game Gf, + such that p is uniformly admissible with respect to that game. In this sense, any probability measure is general admissible by taking the score functionf to be a proper score function ! But for a proper score function f, the associated probability transform P,-(x) = x, Vx E [0, 11, is increasing, and hence I’;’ exists. So we require, in general admissibility, the existence of a score function f such that Py ’ exists. It is seen that with respect to a game G’,.+ with Pf increasing, p is general admissible if and only if P,o p is a limtely additive probability. It is this equivalent property that we will use as a definition for general admissibility. In discussing possibility measures on discrete spaces, we can even consider a stronger concept of admissibility, namely general admissibility in the a-additive sense, i.e., Prop is a a-additive probability measure. It will be shown in this subsection that operations in fuzzy logics are, in general, “admissible,” and even (Zadeh)-max possibility measures are (uniform) limits of admissible decomposable measures. For related works in Statistics see, e.g., [lS]. Throughout this subsection, Q is a discrete space (finite or countably infinite), d = P(Q) and restrict d to all &,,, FE d. For CL: P(Q) + [0, 11, we write p( (0)) = p(w), so that the restriction of p to singletons is regarded as a function, still denoted as p, from Q to [0, 11. We also omit, from now on, the qualification “w.r.t. all finite equal antecedent conditional event sequences.” THEOREM 5.2.1. Let p’: 8(Q) + [0, 11. Then the following are equivalent: (i) p is general admissible in the a-additive sense. (ii) p is a decomposable measure with composition law T being an Archimedean t-conorm with generator g such that g( 1) = 1, and ; gtAw)) = 1. (*I ProojI (i) * (ii). Let f be a score function such that tr is increasing and such that P,.op is a o-additive probability measure. For A E 9(Q), we have P,WW = 2 P/tA~)), A and hence p(A) = PF ‘(CA P&(o)). By taking g = Pr, we have g( 1) = 1, and we see that p is decomposable with Ttx, Y) = P?tP,tx) + Prty)). 578 GOODMAN,NGUYEN,ANDRoGERS Of course, (ii)*(i). Since p is T-decomposable on a discrete space, we have VA E LqQ), p(A)= T(P(~), WEA). Take P,-= g (noting that we can solve for f), we have PpPc(~)=dm4~)? meA) = g [g* (; g(cw)] =; dP(W)) since, by hypothesis CA g(p(o)) < C, g@(w)) = 1 = g( 1). Thus Pro ,u is a a-additive probability measure. 1 Remark. In view of the above theorem, we see that if p is T-composable (or equivalently, if p is generated by its restriction to singletons and T) and if p is general admissible in the o-additive sense, then p = 6,,, the Dirac (probability) measure at o0 E R when v(wO) = 1. The following result provides a necessary and sufficient condition for p to be general admissible in the o-additive sense when sup, p(o) < 1. THEOREM 5.2.2. Let CC 52 -+ [0, l] with Q countably infinite, such that sup, U(O) < 1 (a f 0). Then a necessary and sufficient condition for the existence of a T-possibility measure ,u, generated by c1 and some T, which is general admissible in the o-additive sense, is that VXE(O, 11, a-‘[x, 1]={o:x~fx(o)~1} (**) is finite. In particular, if S is finite and sup, M(W) < 1, then there exists T such that pa,= is general admissible in the o-additive sense. Proof: (a) Necessity. If there is X,,E [O, l] such that cr-‘C.-C,, l] is infinite, then lim 1 g(Ao))= + ~0, n-t +m 4 where A,cc~[x,,, 11. For, although A, is finite, and A,~c~[x,, l] for each n > 1, & g(p(o))>g(x,) [A,(. Thus (*) of Theorem 5.2.1 will not hold. (b) Sufficiency. In view of Theorem 5.2.1, we need to show only the existence of a generator g such that g( 1) = 1 and Cn g(a(w)) = 1. Then we SCORING APPROACH TO ADMISSIBILITY 579 can take pd, r(A) = g* [CA g(cl(w))], where T is the Archimedean t-conorm with generator g. Let O<x,=sup,a(w)<l. Let {01>0)={~:01(~)>O}=U~~~K,, where K, = {o : l/n < a(o) d l/(n - l)}, n > 2. Let n, k 2 such that l/n, < x0< l/(n,-- 1). By (**), cr-‘[l/n,,, 1) is finite; thus (a(u) :coEKno} = 1 Xl 9 x2, . ..1 x,~} for some n, . We can assume 1 1 -<Xl<X2< ‘.. <x,*<--- n0 no- 1’ Note that (**) implies that sup, a(o) is attained at w’, where a(d) = x,, = x0. (Indeed, if a(w”) > a(d) =x,, , then 0” E K,,. and hence contradicts the definition x,, = M$x a(w)). Since [0, 1 ] = [0, l/no] u (l/n,, 11, we first construct g on [0, l/n,]. For n an, and r>O, define g,(O) = 0 and Then l/m < l/n =z. g,( l/m) < g,( l/n) and lim, _ o. g,( l/n) = 0. Thus we construct a continuous, increasing function on [0, l/n,] by extending g, continuously on each [l/n, l/(n - l)], for n > no, (say by joining g,(l/n) and g,( l/(n - 1)) by a straight line). On [l/n,, 11, we proceed as follows. Let a(r) = 1 s,(a(o)). G Since x0 E K,, K&= {o : a(w) < l/n,}; i.e., for OE K’n,, a(o) e [0, l/n,], which is the domain of g, defined above. We have OGa(r)= +c” 1 g,(a(o)) n=n,,+l K. < y (lK,l)g, < +f 1 n=ng+ 1 n=no+l (n- 1)’ Thus lim, _ o1 u(r) = 0. 580 GOODMAN, NGUYEN, AND ROGERS Let r be large enough so that where C,=a-‘(xi)cK,,,, Vj= 1, . . ..n.. There exist real numbers yj, j = 1, . ..) n such that max(,(,),g,(~))<y,<y2~ . . . <y,<l and u(r)+ t ICjlYj=l j=l (see the lemma below). Thus, on [l/no, 11, define g,(xi) = yi, j= 1, . . . . n,, for r large enough. Extend g, by continuity to [l/n,, x1] and on each CXj, xi+ 11, where xnL+ I= 1, and g,( 1) = 1. The condition (*) of Theorem 5.21 is satisfied. Indeed C g,(a(u))=j!l ICjl gAXj)= 5 lcjl Yj K&l j=l so that z g,(a(u)) = $ g,(a(o)) = C g,(a(o)) no KW =a(r)+ !f lC,j yj= 1. j=l Finally, take pa, *(A) = G(a(o)), w E A = g-l& g(cr(o))], where g =g, for r large enough, g*=g-‘, since VA SD, C,, g(a(o)) < &, g(a(o)) = 1 =g( 1 ), and T is the Archimedean t-conorm with generator g. 1 LEMM.4. Let nj, j = 1, . . . . k, be k positive integers. Let ql, q2 E W + such that 0 G go = Max(q,, q2) < 1 1 +ci”=1 “i’ Then there exist real numbers yl, . . . . yk such that (i) 90<.h<y2< ... <yk<L (ii) q, =~~=I njyj= 1. SCORING APPROACH TO ADMISSIBILITY 581 Proof: Let yj = q. + j/c, where c= [ 5 jnj][l--q,q, 2 nj]p’>o, j=l j= 1 Thus qo<y,<y,< ... < yk. Now y, < 1 if and only if q. CT=, jnj = k( 1 - q1 - q. c,“= 1 nj) < xi”= 1 jnj, if and only if k(l-q,)<(l-qO)~,k=, jnj+kq,C~=lnj=dk. But dk2 ek ‘(l-qo)~fzi j+kqo~~=l l=(l-qo)(k(k+1)/2)+k2qo. Now (1 -q,)k<e, if and only if k+l l-q,<-- 2 (l-qd+kq, if and only if 0 G ~(1+4cJ+ql~ which is true here since qo, q, 2 0, and k 2 1. For (ii) ql+ 2 nj.Yj=ql+ i nj(qo+j/C) j= 1 j=l k =ql+qo C n,+ ,I=1 If p is T-decomposable with T(x, y) = Max(x, y) then p is not general admissible even in the additive sense. This is due essentially to the fact that Max(x, y) is not an Archimedean t-conorm. However, Max(x, y) can be approximated by Archimedean ones. For example, let T,(x, y)= [Min(xP+ yp, l)]“P, pa 1. For each p> 1, T, is an Archimedean t-conorm with generator g,(x) = xp, g,( 1) = 1. As usual, we extend T, to n arguments, as T,(x,, x2, . . . . x,) = T,(x,, T,(x,, . . . . -4). Then for each fixed n, T,(x,, x2, . . . . x,) + Max(x,, x2, . . . . x,) as p + + cc, uniformly in (x,, x2, . . . . x,). Indeed, since we have Max(x,, x2, . . . . x,1 d T,(x,, x2, . . . . x,) < Min[Max(x,, . . . . xn)n’Ip, 11 0~ T,(x,, x2, . . . . x,) = Max(x,, x2, . . . . x,) d nlip - 1. Thus, if Q is finite, and CL: Y(Q) + [0, l] has Max(x, y) as composition law, i.e., VA C Q, p(A) = My P(W), 582 then GOODMAN,NGUYEN, AND ROGERS uniformly in A. Now, it is easy to see that, in Theorem 52.1, if we require only general admissibility in the additive sense (which is equivalent to uniform admissibility) the condition (*) can be weakened to Thus, assuming in addition that En p(w) < 1, for all Vpa 1, v,(A) = T,(p(o), o E A) is general admissible in the additive sense, since Therefore, in this case, Max-possibility measures are uniform limits of general admissible T,-possibility measures. More generally, for Sz countably infinite, and a: D + [0, l] such that sup, a(o) < 1, the max-possibility measure generated by tl is Pa(A) = sup a(o) A and this can be approximated by admissible measures, Specifically, THEOREM 5.2.3. Let l2 be countably infinite and a: Sz + [O, l] such that sup, a(o) -C 1. Suppose that there are non-negative reals a, b such that Then (i) For A EP(Q), such that Aa-‘(t, p=(A)=lim,,, +m p,,=,(A), the limit being uniform in all such A with (Al in,, for any fixed positive integer n,; also t, = sup, u.(o). (ii) For Aa-‘(t,)# 0, SCORING APPROACH TO ADMISSIBILITY 583 Proof. (i) We are now going to construct generators g,, for suf- ficiently large p, such that kp(A)=gpl [I g,(a(w))] A are general admissible in the o-additive sense. More specifically, not only is g, such that g,( 1) = 1 and C, g,(a(o)) = 1, but precisely g,(x) = xp on [0, t,], where t, > t,, t, = sup a(0) CT (>O by assumption) and C, = a-‘(t,). First Vp > 0, define g,(x) = xp on [O, t,]. We will extend gP to [0, l] by defining a value for g,(tl) such that tz<gp(t,)< 1, set g(l)= 1 (with extension by continuity as usual) and obtain C g,(a(w)) = 1 R Let a(Q)= {ti,j= 1, 2, . ..} and Cj=aP1(tj). where @p,3= C ap(w)=+f lC$/lT=+f C IC,l t; c;c; j=3 n=2 l/n<r,<l/(n--1) = L 1 ICj(t,P + y 1 1 IC,l f,“. 1/2<1,<1 n=3 l/ncr,<l/(n-I) By (*), the first term is at most, ~22~ times (a finite sum of tp), O<r,<l; this sum tends tozero asp-+ +co. In the second term c c Icjl ‘75 n = 3 l/n < ‘, < ll(n - I) 584 GOODMAN, NGUYEN, AND ROGERS Therefore, this second term is bounded by +O” (n+ l)b a +fnb(n-l)p. 1 np n=3 n=2 +m 3nb1 <UC - ( > -==a(3/2)b +f n-(~-b). n=2 2 np n=2 Thus lim, _ + o. @P,3 = 0. Let p be large enough so that @P,3 < $ and f; < l/2() C1) + ) CJ ). Define g,(t,)=(l--~,,)/IC,I. We have t$‘<g,(t,)< 1 if and only if if and only if @p,,3+w11 + IGot;< 4c,I+@p,* which is true by construction ( GPp, 2 > 0 since f2 > 0). Also, by construction, we have 1 = Qp,2 + ICll &#l) =c gpke4). 52 Thus, if Aa-‘( $3, = T,(a(w), w E A) with 7’,(x, y) = [Min(xP + yP, l)]““. Hence lim, ~ + o. ~~,~$A)=Max(a(o), weA} for any finite A. (ii) For Aa-‘(t,)#@, we have t, = Max{a(o), ok A} < T,(cr(w), o~,4) < 1 and hence I/4z,Tp(4-P12(~)l G 1 -t,. I Remarks. (i) Let (Sz, ~44) be an arbitrary measurable space, and ,u: d + [a,, a,] be general admissible in the a-additive sense with Prop = Q, where Q is a discrete probability measure on (0, ,pP). Then p has the same representation as in the discrete case. Indeed, let B0 CO, countable such that Q(sZ,) = 1. Then VA E &, SCORING APPROACH TO ADMISSIBILITY 585 (ii) A continuous analog is as follows. Let (~2, ~4) = (R, B), ho: R + CO, 11, Fe(x) = Q(( - co, xl) and F,: R --) [a,, a,], F,,(x) = p(( - 00, xl). Since p = P; ’ 0 Q, and P; ’ is increasing, /J is a nondecreasing set-function; hence F, is nondecreasing, and A E 93, p(A) = 9;’ [j,, d(P,o F,(x))]. If, in addition, P,- is differentiable, then P(A) = PF1 j P;-(F,(x)) dl;,(x) . A 1 6. ADMISSIBILITY OF DEMPSTER-SHAFER BELIEF FUNCTIONS Dempster-Shafer belief functions have become very popular in recent years for modeling aspects of expert systems and combination of evidence problems in AI. The purpose of this section is to respond to Lindley’s comments about the inadmissibility of belief functions [16]. For simplicity let C2 be a finite set. A belief function Be1 on P(a), the power class of G, can be defined as follows. Let m: Y’(n) -+ [0, l] such that m(0) = 0, Cscn, m(A) = 1. Then, Bel(B)= 1 m(A). AGE m is the probability allocation for Bel. By the Mobius inversion formula, m can be recovered from Bel, m(A)= c (-l)'"-"I Bel(B), BCA where JA( denotes the cardinality of A. Note that if p: B(G) -+ [0, l] is such that p(sZ) = 1 and VAGQ, 2 (-l)‘“-B’/@)>O Bc_A then p is a belief function. If we think of “sets” as “points,” then m plays the role of a probability mass function, and Be1 is a “cumulative distribution function.” Since Sz is finite, we have Bel(A)= P(Xe.tT(A)), VAEQ, where X is a random set, defined on some probability space (a, 9, P), and taking values in P(s2) with “density” m, i.e., P(B:X(O)=A)=m(A). 586 GOODMAN,NGIJYEN, ANDROGERS Note that Bel(A) + Bel(A”) < 1. We extend Be1 to conditional events s’(Q) as follows. For E, FEY(G), define Bel(E)F)=P(XcEjXGF) when P(Xs F) > 0. For more information on belief functions, we refer the reader to 126, 21, 11, 303. As in Section 5, we consider Gf,, with Pr increasing. By the nature of belief functions, we consider [a,, a,] = [O, 11. Note also that the existence of f such that Pr is increasing is equivalent to that of a (surjective) continuous, increasing function h: [O, 1 ] + [O, 11. Thus, in view of Theorem 4.2.1, p E [0, 1 ] d is general admissible if and only if there is such an h for which h 0 p is a finitely additive probability measure. In discussing the admissibility of belief functions on Sz finite, & = P(Q), it should be noted that the range of a belief function is not the whole inter- val [O, 1 ] ! As we will see, as in the case of fuzzy logics, some classes of belief functions are admissible while others are not. Thus, if DeFinetti’s coherence principle is viewed as a rational way of choosing “reasonable” uncertainty measures, the following analysis will provide criteria for selecting “good” belief functions. First, a simple condition of inadmissibility. THEOREM 6.1. Let (0, d) be a measurable space and p E [0, 1 ] d. Zf there is E, FE d such that I@) = P(F) and AEC) Z ,W”)> then p is not general admissible. Proof. If /J were general admissible, then there would be an h: [0, l] + [O, 11, (surjective) continuous, increasing, such that h 0 ,u is a finitely additive probability measure. Thus and hence h 0 ,u(E”) = h 0 p(FC) which contradicts the hypothesis, since h-’ exists. 1 SCORING APPROACH TO ADMISSIBILITY 587 EXAMPLES. (a) 0 = (ml, 02, w3, m4, w5, w6}. Let E = {wl, w2}, F= {03, w4}. Let m: P(Q) + [0, l] with m(w1) = m(w3) =p, > 0 m(w2) = m(w4) =p2 > 0 m({w,,w,})=m({w,,o,))=p,>O ~({w,~w,~%))=P4~O m({w,Y%))=P,>O ~({%~4))=P,~O m(A)=0 for all other subsets and Pl+Pz+P3+P4+P5+!76=f. We have Bel(E)=Bel(F)=p,+p,+p,. Bel(Ec)=p,+p,+p,+p4+p,, but Bel(F”) =pl +pZ +p3 # Bel(E”). (b) Consider the degenerate belief function focused on A, Bel,(B) = 1 if AcB and zero otherwise. Let B, A # @ and B,A = @. Then Bel,(B,) = Bel,(B,) = 0, but Bel,(B;)=O# 1 =Bel,(B”,). The hypothesis of Theorem 6.1 expresses the fact that p(E”) is not a function of p(E). Thus, an uncertainty measure p is not general admissible if there is no rp: [0, l] + [0, l] such that YE, FE d, AE” I -F) = cpbWl F)). (*) A typical example is the Max-possibility measure (which explains its inadmissibility mentioned in Section 5). The relation (*) always holds for probability measures, but as we have just seen, (*) might fail in the case of belief functions. The Theorem 6.1 provides a necessary condition for general admissibility: If p is general admissible, then necessarily, (*) must hold. The following result provides sufficient conditions for general admissibility of belief functions. THEOREM 6.2. Let l2 be finite and d = P(Q). 409/159/2-20 588 GOODMAN, NGUYEN, AND ROGERS (i) Zf Bel: J? + [O, l] (extended to d as mentioned previously) is such that VE, FE s#‘, with Bel(F) > 0, Bel(E’IF)=q[Bel(ElF)], (**) where cp: [0, 11 + [O, 1 ] is differentiable, then Be1 is general admissible. (ii) For each, r B 1 and Q: STJ’ + [0, 1 ] a finitely additive probability measure, Qr is a belief function on zl which is general admissible. Proof: (i) The proof of (i) follows from [S]: let E,, E,, FE& with Bel(F) > 0. Using the conditional extension of Bel, we have where A: [0, l]* + [0, 11, A(x, y) = xy. We also have Bel(F( F) = 1. Thus, together with (**), Bel’ is a finitely additive probability measure on d for some positive real r. Since (.)‘: [0, l] -+ [0, l] is surjective, continuous and increasing, Be1 is general admissible. (ii) For r > 1 and Q a finitely additive probability measure in 52 (finite), Q’ is a belief function if VAGQ, m,(A)= c (-l)lA-BI Q’(B)>O. BEA For JAI =0, i.e., A=@, we have m,(@)=O. For IAl = 1, say, A = (or>, we have m,(~~l~)=Q’(~~l~)~O. For IAJ =2, say A= {CD,, co,}, mr({~~~~2~)=Q'({ o,,w,>,-Q'((ol>,-Q'((~2}) =CQ({o,>,+Q({o,>,lr-Q'({o,>> -QW4)N. Since the function u(tl) = CZ” + (1 - a)l is such that u(O) = u( 1) = 1 and u( . ) is convex with minimum at LI = i, u(f) < 1, we have SCORING APPROACH TO ADMISSIBILITY 589 For IAl=3, say A={w,~~,w~}, let a=Q{w,), b=Q({w2}), c=Q((c+}). Then m,(A)=(a+b+c)‘-(a+h)‘-(a+~)‘-(b+c)’+ u’ + b’ + c’. Thus m’(A) 3 0 if (a+b+c)‘>(u+b)‘-a’+(u+c)‘-c’+(b+c)’-b’. (*) Now, if r = n >, 1 is integral, then the right hand side of (*) is uibJck, where S = {(i, j, k) : i+ j + k = n and not all the i, j, k are positive}. Thus (* ) holds since a, b, c 3 0 and Sc{(i,j,k):i+j+k=n}. For r > 1, real, we rewrite (*) as (u+b+c)‘+a’+b’+c’~(u+b)‘+(u+c)‘+(b+c)’. (**) Let u(r), w(r) be the left and right hand sides of (** ), respectively. Observe that u(r) and w(r) are convex functions on [ 1, + cc), since Ofu+b+c<l. Also u(l)=w(1)=2(u+b+c) and lim u(r)= lim w(r)=O. r- +cr ” +r Thus (**) holds since (**) holds for any integer r > 1. The above argument applies to the case IAl > 4 as well. 1 Remarks. (i) Of course if Be1 = Q’ where Q is a finitely additive probability measure and r > 1, then Be1 is general admissible by Theorem 6.2 (i): VE, FE .ra2, Bel(E’ 1 F) is a function of Bel(E 1 F), namely cp(x)=(l -x)‘. (ii) We have seen that, in the scoring approach to admissibility of uncertainty measures, if the aggregation function is taken to be addition (as in Lindley’s work), then well-known measures such as Max-possibility measures (which are consonant plausibility functions) and degenerate belief functions are not general admissible (i.e., they are incoherent in DeFinetti’s sense). Although, we have shown that, in this case, there exist admissible T-possibility measures and admissible belief functions, it is useful to consider arbitrary aggregation functions in order to set up a general framework for studying the question of admissibility, i.e., a general concept 590 GOODMAN, NGUYEN, AND ROGERS of coherence. From a logical viewpoint, this is precisely the concern of AZ researchers as well as statisticians dealing with applications of belief functions to statistical inference (e.g., [30, p. 1061071). Section 7 will give some insights in this direction. 7. UNCERTAINTY GAMES WITH NONADDITIVE AGGREGATION FUNCTIONS In this section, some specific nonadditive aggregation functions are considered. In Example 1 it is shown that a simple nonadditive form of aggregation leads to a corresponding uncertainty game for which uniformly admissible measures are not transformable to probabilities; this is in opposition to Lindley’s games with an additive aggregation function. Example 2 presents a situation in which a nonadditive aggregation function has a general additive form. In this case, uniformly admissible measures have probability-like characterizations. In Example 3, we present some specialized cases of Example 2. EXAMPLE 1. Consider a = ((E, 1 F), (E, ( F), (E, u Ez 1 t;)), with E,E,=@. Let x = WI If’), Y = AE2 I J’), z = WI u E2 IF). We will discuss the weak local admissibility of 2 = (x, y, z) with respect to the game GJti,~, where the game is specified as follows. Recall that $: [w, + [w is specified by the sequence Ic/, : BY --t 58, n > 1. For our purpose here, it suffices to consider ti3: R3 + R. We take $Ju, u, w) = uu + w. For the score function f, we take a, = 0, a, = 1 and f(x, 0) =x2, f(x, 1)=(x-1)2 on CO, 11, We are going to show that there is no transform h: [0, l] -+ [0, l] such that h(z) = h(x) + h(Y) when (x, y, z) is WLAD. As a consequence, if p is uniformly admissible, p need not be transformable to a finitely additive probability measure. With the notation of Section 3, we have [ (x-l)y2 (x-1)2y z-l J=2 x(y-1)2 x2+ 1) z-l . XY2 X’Y Z 1 To find 2 = (x, y, z) - WLAD, we use the sufficient condition given in Theorem 3.2.2 (e). SCORING APPROACH TO ADMISSIBILITY 591 First, det J = 0 gives x*+-y*-xy(x+y) ‘(” ‘)= 2(x+y)-3xy- 1 for xy#O, 1. (*) ForO~x~l,y=l,wehavez=1.For~=(x,1,l)withO~x~l,Jhas rank p = 2 with R = (0, 0), and so (x, 1, 1) is WLAD. If h: [0, l] -+ [0, I] is such that h(z) =h(x)-th(y) for any (x, y, z) - WLAD, then, in particular, h(l)=h(x)+h(l), VXE(O, l), implying that h(x) = 0, Vx E (0, 1). From (*), we see that when x= y= i, we have z= 1. For x= ($, i, I), and p = 2. For with and RJ, = ($, -a), we have R = ( - LO), so that f = (4, 4, 1) is WLAD, and hence Now it can be seen that i = (0, 0,O) is WLAD, since P’L and contains one zero component. Thus h(O)= h(O)+ h(O), implying that h(O) = 0. Therefore, h c 0 on [0, 11. 592 GOODMAN,NGUYEN, AND ROGERS EXAMPLE 2. Consider a as in Example 1. Let f be an arbitrary score function and ti3(& u, w) = 4%(401(~) + e(u) + %(W))> where ‘pi: R + R, j = 0, 1,2, 3 are four surjective, increasing, continuously differentiable functions. This is a generalized form of the additive aggrega- tion function. Note that 11/3(~, v, w) = UD + w in Example 1 is not of this form. Large classes of functions of several variables can be represented, or approximated, by general forms of addition of simple argument functions as above; see, e.g., the survey paper of Sprecher 1291 on the work of Kolmogorov and others in considering the related 13th-problem of Hilbert. If $ = (x, y, z) is WLAD then det J= 0, implying that p~p,of(4 = Pqp,&) + p,,of(Y)7 where Pa,&) = dCf(c 011 f’(c 0) cp: Cf(k O)l f’(c 0) - dCf(& 1 )I f’(t, 1)’ EXAMPLE 3. As a special case of Example 2, one can consider uncer- tainty games with symmetric aggregation functions as follows. Let g: R + R be a (surjective), increasing, continuously differentiable function. For n 2 1, define as IClg,ntXIY ae.9 xn)=g-l ,i gtxi) L > The game Gs,*, can be identified with G,,, + , where II/, is specified by $g,n, n > 1. For example, for a fixed p > 1, take gpw = tP for t>O. Note that Max(x,, x2, . . . . However, II/ = Max yields x,) is a limiting case of tigptn when p + + co. essentially only trivial admissible uncertainty measures and hence is not a good candidate for being a viable nonadditive aggregation function. SCORING APPROACH TO ADMISSIBILITY 593 ACKNOWLEDGMENTS We thank D. V. Lindley and L. A. Zadeh for their kind discussion on an earlier draft of this work. REFERENCES 1. J. ACZEL, “Lectures on Functional Equations and Their Applications,” Academic Press, New York, 1966. 2. J. P. AUL~N, A Pareto minimum principle, in “Differential Games and Related Topics” (H. W. Kuhn and G. P. Szego, Eds.), pp. 147-175, North-Holland, Amsterdam, 1971. 3. D. BLACKWELL AND M. A. GIRSHICK, “Theory of Games and Statistical Decisions,” Dover, New York, 1979. 4. L. D. BROWN, A heuristic method for determining admissibility of estimators-with applications, Ann. Sfatisi. 7, No. 5 (1979), 96&994. 5. R. T. Cox, “The Algebra of Probable Inference,” The Johns Hopkins Press, Baltimore. 1961. 6. B. DEFINETTI, “Theory of Probability,” Vols. 1 and 2, Wiley, New York, 1974. 7. T. S. FERGUSON, “Mathematical Statistics,” Academic Press, New York, 1967. 8. D. A. FREEDMAN AND R. A. F?JRVES, Bayes’ method for bookies, Ann, Math. Statist. 40 (1969), 1177-1186. 9. C. GENES AND R. J. MACKAY, Copules ArchimCdiennes et families de lois bidimension- nelles dont les marges sent donnkes, Canad. J. Statist. 14 (1986), 145-159. 10. C. GENEST AND R. J. MACKAY, The joy of copulas: bivariate distributions with uniform marginals, Amer. Statist. 40 (1986), 28&283. 11. I. R. GOODMAN AND H. T. NGUYEN, “Uncertainty Models for Knowledge-Based Systems,” North-Holland, Amsterdam, 1985. 12. I. R. GOODMAN, H. T. NGUYEN, AND E. A. WALKER, ‘*Conditional Inference and Logic for Intelligent Systems: A Theory of Measure-Free Conditioning,” North-Holland, Amsterdam, 1991. 13. D. HEATH AND W. SUDDERTH, On finitely additive priors, coherence, and extended admissibility, Ann. Sfafisr. 6 (1978), 333-345. 14. A. KOZEK, Towards a calculus for admissibility, Ann. Statist. 10, No. 3 (1982), 825-837. 15. R. B. LATTA, Composition rules for probabilities from paired comparisons, Ann. Sfafisf. 7, No. 2 (1979), 349-371. 16. D. V. LINDLEY, Scoring rules and the inevitability of probability, Intern. Statist. Reo. 50 (1982), l-26. 17. D. V. LINDLEY, The probability approach to the treatment of uncertainty in Artilicial Intelligence and expert systems, Statist. Sci. 2, No. 1 (1987), 17-24. 18. C. H. LING, Representation of associative functions, Puhl. Math. Debrecen 12 (1965), 189-2 12. 19. A. W. MARSHALL AND I. OLKIN, Families of multivariate distributions, J. Amer. Stofisf. Assoc. 83, No. 403 (1988), 834-841. 20. D. MCDERMOTT AND J. DOYLE, Non-monotonic logic, I, Artificial Intelligence 13, No. 1 (1980), 41-72. 21. H. T. NGUYEN, On random sets and belief functions, J. Math. Anal. Appl. 65, No. 3 (1978), 531-542. 22. E. REGAZZINI, DeFinetti’s coherence and statistical inference, Ann. Sratisf. 15, No. 2 (1987). 845-864. 594 GOODMAN, NGUYEN, AND ROGERS 23. A. RENYI, “Foundations of Probability,” Holden-Day, San Francisco, 1970. 24. G. SWAY, An algebra of conditional events, J. Math. Anal. Appl. 24, No. 2 (1968), 334344. 25. B. SCHWEIZER AND A. SKLAR, “Probabilistic Metric Spaces,” North-Holland, Amsterdam, 1983. 26. G. SHAFER, “A Mathematical Theory of Evidence,” Princeton Univ. Press, Princeton, NJ, 1976. 27. S. SMALE, Global analysis and economics, I, Pareto optimum and a generalization of Morse theory, in “Dynamical Systems” (M. M. Peixoto, Ed.), pp. 531-544, Academic Press, New York, 1973. 28. D. J. SPIEGELHALTER, A statistical view of uncertainty in expert systems, in “Artificial Intelligence and Statistics” (W. A. Gale, Ed.), pp. 17-56, Addison-Wesley, Reading, MA, 1986. 29. D. A. SPRECHER, A survey of solved and unsolved problems on super-positions of functions, J. Approx. Theory 6 (1972), 123-134. 30. L. A. WASSERMAN, “Some Applications of Belief Functions to Statistical Inference,” Ph.D. thesis, Univ. of Toronto, Ontario, 1987. 31. S. WEBER, I-decomposable measures and integrals for archimedean f-conorms I, J. Math. Anal. Appl. 101 (1984), 114-138. 32. L. A. ZADEH, The role of fuzzy logic in the management of uncertainty in expert systems, Fuzzy Sets and Systems 11 (1983), 199-227. 33. L. A. ZADEH, Possibility theory and soft data analysis, in “Mathematical Frontier of the Social and Policy Sciences” (L. Cobb and R. M. Thrall, Eds.), pp. 69-129, Westview, Boulder, CO, 1981. 34. T. L. FINE, “Theories of Probability,” Academic Press, New York, 1973. 
work_fntrwpzutnd5dni2qafymzmlca	----	Knowledge acquisition for fuzzy expert systems Knowledge Acquisition for Fuzzy Expert Systems Gwo-Jen Hwang Computer Center, National Chiao Tung University, Hsinchu, 300, Taiwan, Republic of China Expert systems have been successfully applied to a wide variety of application domains. To achieve better performance, researchers have tried to employ fuzzy logic to the development of expert systems. However, as fuzzy rules and membership functions are difficult to define, most of the existing tools and environments for expert systems do not support fuzzy representation and reasoning. Thus, it is time-consuming to develop fuzzy expert systems. In this article we propose a new approach to elicit expertise and to generate knowledge bases for fuzzy expert systems. A knowledge acquisition system based upon the approach is also presented, which can help knowledge engineers to create, adjust, debug, and execute fuzzy expert systems. Some control techniques are employed in the knowledge acquisition system so that the concepts of fuzzy logic could be directly applied to conventional expert system shells; moreover, a graphic user interface is provided to facilitate the adjustment of membership functions and the display of outputs. The knowledge acquisition system has been integrated with a popular expert system shell, CLIPS, to offer a complete development environment for knowledge engineers. With the help of this environment, the development of fuzzy expert systems becomes much more convenient and efficient. 0 1995 John Wiley & Sons, Inc. I. INTRODUCTION In recent years, expert systems have been applied to a wide variety of application domains, which not only showed the utility of this approach, but also revealed some problems of applying it. Part of the problems in employing conventional expert systems may be due to the use of binary logic to express domain knowledge, whose expressions may not be natural or straightforward and, hence, the expertise is not completely represented. For example, there are only two possible sides (true or false) in binary logic and a clear boundary must be indicated for making choices, while in the real world, it is difficult to judge things to be absolutely true or false. The problem becomes more compli- cated if linguistic hedges (e.g., very, gradually, moderately, m o r e or less, etc.) are used in expressing expertise.’ A good solution to this problem is to employ the concept of F u z z y S e t and Fuzzy L o g i c in expert systems; that is, fuzzy expert systems. For example, let E represent the set of “the people who are tall.” In conventional binary logic, INTERNATIONAL JOURNAL OF INTELLIGENT SYSTEMS, VOL. 10, 541-560 (1995) 0 1995 John Wiley & Sons, Inc. CCC 0884-8 173/Y5/06054 1-20 542 HWANG we must clearly specify whether an element belongs to E or not. However, in fuzzy logic, each element belongs to the set with a degree of membership. A person of 160-cm height belongs to E with the degree of membership 0.0; while the persons of 170-cm, 175-cm, and 180-cm heights may have degrees of membership 0.5, 0.75 and 1 .O, respectively. Moreover, when considering the linguistic hedges (e.g., very, gradually, moderately, inore or l e s s , etc.), all we need to do is to compute the degrees of memberships according to some hedge functions. For example, let the Fu,,(x) = x', where x is degree of memberships without qualifier. If a person is said t o be T A L L with degree 0.8, according to F v e V ( x ) , we conclude that the person is VERY T A L L with degree 0.64. However, the development of a fuzzy expert system is time-consuming. The difficulty may be due to the lack of methodologies and tools to help in eliciting knowledge for constructing fuzzy rules and in defining the membership functions for fuzzy values. Moreover, most existing expert system shells do not support fuzzy reasoning, which increases the cost needed for developing fuzzy expert systems. T o cope with these problems, we propose a knowledge acquisition system that not only elicits knowledge from domain experts, but also generates knowledge bases for fuzzy expert systems. We have integrated the environment with a popular expert system shell, CLIPS, to enable fuzzy reasoning. Some experimental results showed that the approach is desirable and is worth further study. In this article, we propose a knowledge acquisition environment that con- sists of a knowledge acquisition unit, a graphic membership function developer, a fuzzy interface for expert systems shells, and a set of control rules for fuzzy reasoning. The environment has been integrated with a well-known expert system shell to perform fuzzy reasoning. Finally, some test results of an experi- ment on a practical medical diagnosis problem are shown to evaluate the new approach, and a conclusion is given. 11. KNOWLEDGE ACQUISITION FOR FUZZY EXPERT SYSTEMS One critical bottleneck to building expert systems is the elicitation of knowl- edge from domain experts. Many knowledge engineering tools (e.g., ETS,? N ~ o E T S , ~ AQUINAS,4 MOLE,' NICOD,' ICONKAT,' K S S 0 , 8 ASK,9 SMORE,'O ALT," RuleCon,'* KITTEN,I3 etc.) and interviewing technique^,'^,'^ have been developed to cope with this problem. In this article, we propose a knowledge acquisition environment, Knowledge Acquisition for Fuzzy Expert Systems (KAFES), which is capable of eliciting expertise from domain experts and generating knowledge bases for fuzzy expert systems. The environment is integrated with a well-known expert system shell, CLIPS, to enable fuzzy reasoning. It consists of four components: ( 1 ) Knowledge acquisition unit: a revised repertory grid mechanismI6 is used t o ( 2 ) Graphic membership function developer: a tool to help knowledge engineers enable the elicitation of fuzzy knowledge. adjust membership functions of fuzzy values. KNOWLEDGE ACQUISITION 543 Domain Knowledge Engineers Figure 1. The structure of KAFES. (3) Fuzzy interface for expert system shells: the interface performs the fuzzification and defuzzification operations, which facilitate fuzzy reasoning on conventional expert system shells. Currently, it is available for CLIPS, a shell for the develop- ment and delivery of expert systems developed by the Software Technology Branch of the Information System Directorate at NASA/Johnson Space Center. (4) A set of control rules for CLIPS, which enables fuzzy reasoning on CLIPS. The whole structure of the environment is depicted in Figure 1. A. Knowledge Acquisition Unit Many existing systems employ the repertory grid test, which was originally invented and widely used by George Kelly in clinical psychology16 when inter- viewing experts. In a repertory grid, a set of elements is listed across the top of the grid. Various combinations of three elements are then presented t o the expert, and the question, “Think of an important trait or characteristic, which makes two of the elements different from the third,” is asked. The trait given by the expert is listed on the left-hand side of the grid, and its opposite is listed on the right-hand side. Each pair of a trait and its opposite is called a construct. After the set of constructs is ready, the expert is asked t o fill the grid with ratings. A 1 -+ 5 rating mechanism is often used in filling the grid; i.e., each rating is an integer ranging from 1-5. A 1 is at the left-hand pole, a 5 is a t the right-hand pole, while the rest fall between the poles. An example of a repertory 544 C J 1 4 5 1 4 c ; c 4 1 1 1 5 1 c ; H W A N G Figure 2. An example of a repertory grid. grid with five elements (El, E 2 , E,, E4, and E,) and four constructs ( C l , C2, C3, and C,) is given in Figure 2. Each rating expresses the relationship between a construct and an element, which does not provide sufficient information for building fuzzy expert systems. Consider the following repertory grid: Dangue Measles . . . fever Low 5 2 . . . High . . . . . . . . . From the ratings of the grid, we are only aware of the information, “For Dangue Fever, a patient is likely to have a high temperature,” and, “For Measles, a patient’s temperature is usually not high.” It is quite possible that the expert wants to emphasize the expressions by adding some linguistic hedges: “ F o r Dangue Fever, a patient’s temperature should be VERY high.” “ F o r Measles, a patient’s temperature is usually M O R E OR LESS low.” From the examples, we find it difficult t o apply conventional repertory grids to the development of fuzzy expert systems, which is d u e to the following reasons: (1) Repertory grids do not explicitly hold information for linguistic hedges. (2) Repertory grids do not hold information for helping the determination of mem- bership functions. Thus, it would be more desirable to adapt the grid so that it is appreciate for fuzzy logic instead of using fuzzy logic with repertory grids. To cope with these problems, we propose thefuzzy table to hold the information for generating fuzzy rules and membership functions. The procedure for constructing a fuzzy table is similar to that of a repertory grid, which consists of three steps shown as follows: Step 1: Elicit all the elements (candidate solutions o r actions) from the KNOWLEDGE ACQUISITION 545 expert. The elements (e.g., G , , GZ, G,, G4, and G,) provided by the expert are placed across the top of the table: Step 2: Elicit attributes (fuzzy variables) from the expert. Each time, the expert is asked for an attribute so that three elements can be partially or fully distin- guished. The fuzzy variables are listed on the left-hand side of the table. For each fuzzy variable, three fuzzy values are specified and listed on the right- hand side according to the following procedure: KAFES: Please select a set of fuzzy values for fuzzy variable V,: 1. LOW/ MIDDLE/HIGH. 2. SHORT/MIDDLE/TALL. 3 . LIGHTINORMALI HEAVY, 4. SMALL/MIDDLE/BIG. 0. Other (user-defined). EXPERT: 1. KAFES: Please select a set of fuzzy values for fuzzy variable V,: 1. LOW/ MIDDLEIHIGH. 2. SHORTIMIDDLEITALL. 3 . LIGHT/NORMAL/ HEAVY. 4. SMALL/MIDDLE/BIG. 0. Other (user-defined). EXPERT: 2. KAFES: Please select a set of fuzzy values for fuzzy variable V,: 1. LOW/ MIDDLE/HIGH. 2. SHORTIMIDDLEITALL. 3 . LIGHT/NORMAL/ HEAVY. 4. SMALL/MIDDLE/BIG. 0. Other (user-defined). EXPERT: 0. KAFES: Please indicate the lower bound of the fuzzy values. EXPERT: SLOW. KAFES: Please indicate the middle of the fuzzy values. EXPERT: NORMAL. KAFES: Please indicate the upper bound of the fuzzy values. EXPERT: FAST. An example of a fuzzy table with fuzzy variable and values is given as follows: (3, (3, G3 G4 G.5 Vl v2 v3 v4 LOW ; MIDDLE ; HIGH SHORT ; MIDDLE ; TALL SLOW ; NORMAL ; FAST LOW ; MIDDLE ; HIGH Step 3: Fill all of the [element, attribute] entries of the grid. Each entry consists of two parts: a rating to indicate the most desirable fuzzy value and the degree of certainty for giving that rating. It is possible that the expert selects the most desirable fuzzy value, but still lacks confidence about his (or her) choice. A 7-scale rating mechanism is employed in the fuzzy table. The ratings are integers ranging from - 3 to + 3 , which represent the possible combinations 546 HWANG of linguistic hedges and fuzzy values. Consider the ratings of fuzzy variable V, in the table; 3 means V E R Y HIGH, 2 means HIGH, 1 means MORE OR LESS HIGH, 0 means MIDDLE, - I means MORE OR LESS LOW, -2 means LOW, and - 3 means VERY LOW. The degree of certainty for each rating is described by “S”or “ N ” , which represent VERYSUREand NOTVERYSURE, respectively. For example: GI G2 G3 G4 GS 31s OIN -31s 11s 31s LOW ; MIDDLE ; HIGH 21s 1/S 31s 21s O / S SHORT ;MIDDLE ;TALL v3 21s 21s 31s 1lN 21s SLOW ; NORMAL ; FAST V, - l l N ‘21s -21s -31s 31s LOW ; MIDDLE ; HIGH Vl v2 Each column of the fuzzy table will be translated to a fuzzy rule. The truth value for the output of i - th rule is computed by x 0.8 + 0.2. # of “S” TRUTH = ( # of “S,! + # of $“,I) For example, the first column of the above fuzzy table is translated to the following rule: I F V, i s V E R Y H I G H , and V2 i s TALL, and V, i s FAST, and V, is MORE OR LESS LOW THEN I n s e r t n e w f a c t G , TRUTH= 0 . 9 5 . Initially, experts are asked to provide numerical information for fuzzy val- ues, and some default functions are employed. Let fuzzy values of afuzzy variable be represented as LS (left side), MS (middle side), and RS (right side). Assuming that the numerical values with membership degrees 1 .O for LS, MS, and RS are a, p , and y, respectively, we have the following membership function^:'^ KNOWLEDGE ACQUISITION 547 , Low Middle High V1 250 300 350 Figure 3. Graphic representation of fuzzy values for V , . For V, in the above table, if the expert gives numerical value 250 for fuzzy value LOW (LS), 300 for MIDDLE (MS), and 350 for HIGH (RS), membership functions given in Figure 3 are derived. The degrees of memberships of linguistic hedges are computed by the follow- ing hedge functions: 1. F(X),Or = 1 - X 2. F(X)vERy = X 2 - xO.5 3 . F(X)MORE-OR-LESS - 7 where X is the degree of memberships without considering linguistic hedges. B. Graphic Membership Function Developer Graphic membership function developer is used to help knowledge engi- neers to create and adjust membership functions for input facts. It provides a graphic view, so that knowledge engineers can adjust the shape of membership function curves directly to get the desirable results. 548 HWANG Defuuificanon Figure 4. Fuzzy reasoning by applying the fuzzy interface. C. Fuzzy Interface The knowledge representation and inference process in a fuzzy expert system is quite different from those in a conventional expert system. However, as there already exist many expert system shells and knowledge engineering tools for conventional expert systems, it would be very convenient if we could build fuzzy expert systems based upon the existing shells or tools. We selected one of the most popular expert system shells, C Language Integrated Production System (CLIPS) to work with the knowledge acquisition system. CLIPS is a tool for the development and delivery of expert systems developed by the Software Technology Branch of the Information System Directorate at NASA/ Johnson Space Center. The fuzzy interface is capable of accepting input values from users, per- forming fuzzification operations, invoking the inference of CLIPS, and perform- ing defuzzification operations on the outputs from CLIPS. If the knowledge engineer or the expert is not satisfied with the system output, the graphic developer is helpful in adjusting the membership functions to achieve better performance. The process of fuzzy reasoning by applying the fuzzy interface is shown in Figure 4. After performing the inference process, all the system outputs are collected by the fuzzy interface to make the final conclusion. The centroid defuzzification is employed to derive a more reasonable result, which will be shown by graphic representation as well as text display to facilitate the modification and debugging of the fuzzy expert systems. D. Control Rules for Fuzzy Reasoning A significant difference between fuzzy rules and conventional rules is the use of linguistic hedges, such as NOT, V E R Y , NOT V E R Y , M O R E O R L E S S , NOT M O R E O R L E S S , etc. For example, two rules with truth value 1.0 are given as follows: KNOWLEDGE ACQUISITION 549 ( d e f r u l e Domain-Rule, (LEVEL ? gg) ? f f l < - (TEMP I S H I G H ? d d l & : (> ? d d l ?gg) ) ? f f 2 < - (FAN I S NOT FAST ?dd2&: (> ?dd2 ?gg) ) => ( a s s e r t (DEL-FACT FAN I S FAST) ) ( b i n d ?dd (min ? d d l ?dd2) ) ( p r i n t o u t t 'FAN I S FAST ' ? d d c r l f ) ( a s s e r t (NEW-FACT FAN I S FAST ? d d ) ) ) ( d e f r u l e Domain-Rule, (LEVEL ?gg) ? f f l < - (TEMP I S VERY H I G H ? d d l & : (> ? d d l ?gg) ) ? f f 2 < - (FAN I S NOT FAST ?ddZ&: (> ?dd2 ?gg) ) => ( a s s e r t (DEL-FACT FAN I S FAST) ) ( b i n d ?dd (min ? d d l ?ddZ) ) ( p r i n t o u t t 'FAN I S VERY FAST ' ?dd c r l f ) ( a s s e r t (NEW-FACTFAN ISVERYFAST ? d d ) ) ) For an input value (e.g., 25"C), the fuzzy interface first performs fuzzifica- tion operations to produce the corresponding fuzzy fact (e.g., temperature is NORMAL 0.95), and then transfers it to CLIPS for inference. The problem is, how could those fuzzy facts match the premise patterns of rules if linguistic hedges exist. For example, the fact (TEMPERATURE IS HIGH 0.6) should be able to match the patterns (TEMPERATURE IS VERY HIGH) with degree 0.36, (TEMPERATURE IS HIGH) with degree 0.6, (TEMPERATURE IS LOW) with degree 0.4, (TEMPERATURE IS NOT VERY HIGH) with degree 6.4, etc. To cope with this problem, some control rules are used to generate a set of facts with linguistic hedges to replace the original fuzzy fact. One such control rule is shown as follows: ( d e f r u l e G e n e r a t e - L i n g u i s t i c - H e d g e - F a c t s ( d e c l a r e ( s a l i e n c e 9 9 7 ) ) ? f f < - (NEW-FACT ? O O I S $?CCl ? C C ~ ?dd) => ( r e t r a c t ? f f ) ( i f (member NOT $ ? c c l ) t h e n ( b i n d ? d d ( - 1 ? d d ) ) ) ( i f (member VERY $ ? c c l ) t h e n ( b i n d ? d d ( * * ? d d O . 5 ) ) ) ( i f (member MORE-OR-LESS $?ccl) t h e n ( b i n d ? d d ( * * ?dd 2 ) ) ) (assert (?oo IS?ccZ?dd) ( ? o o I S N O T ? c c 2 = ( - l ? d d ) ) (?ooISVERY?cc2 = ( * * ? d d Z ) ) ( ? O O I S NOTVERY ? C C Z = ( - 1 ( * * ?dd 2 ) ) ) (?OOISMORE-OR-LESS?CCZ=(** ?ddO. 5 ) ) ( ? O O ISNOTMORE-OR-LESS ?CCZ = ( - 1 ( * * ? d d o . 5 ) 550 HWANG For example, let the input fact be (temperature is 18"C), the following fuzzy facts will be produced by fuzzification operations: (temperature is HIGH 0.5) (temperature is MIDDLE 0.8) (temperature is LOW 0.0) The fact (temperature is LOW 0.0) is not removed immediately, since some of its corresponding facts with linguistic hedges will have nonzero degrees. To enable fuzzy inference with linguistic hedges, the following fuzzy facts are generated by the control rules to replace the three original facts: (temperature is HIGH 0.5) (temperature is NOT HIGH 0.5) (temperature is VERY HIGH 0.25) (temperature is NOT VERY HIGH 0.75) (temperature is MORE-OR-LESS HIGH 0.707) (temperature is NOT MORE-OR-LESS HIGH 0.293) (temperature is MIDDLE 0.8) (temperature is NOT MIDDLE 0.2) (temperature is VERY MIDDLE 0.64) (temperature is NOT VERY MIDDLE 0.36) (temperature is MORE-OR-LESS MIDDLE 0.894) (temperature is NOT MORE-OR-LESS MIDDLE 0.106) (temperature is LOW 0.0) (temperature is NOT LOW 1.0) (temperature is VERY LOW 0.0) (temperature is NOT VERY LOW 1 .O) (temperature is MORE-OR-LESS-LOW 0.0) (temperature is NOT MORE-OR-LESS LOW 1 .O) 111. EVALUATION OF THE NEW APPROACH Some experiments concerning the diagnosis of Acute Exanthema were made to evaluate the new approach. The earlier knowledge base was constructed by applying the EMCUD method," and 47 rules are generated for 14 diseases. I n Appendix A, it can be seen that there are lots of redundant rules although the test results (in Table I) seem to be desirable. With the new approach, 14 fuzzy rules are generated (as given in Appendix B) with the same test results. IV. CONCLUSIONS In this article, we propose a knowledge acquisition system, KAFES, for constructing fuzzy expert systems. We have applied KAFES to a medical application. According to our experience, it can be seen that the use of fuzzy concepts to represent expertise significantly reduces the redundant rules in the knowledge bases while achieving a promising performance. Moreover, it seems KNOWLEDGE ACQUISITION 55 1 Table I. Test results of the knowledge base with redundant rules.a Case number 1 2 3 4 5 6 7 8 9 1 0 1 1 1 2 1 3 Physician 12 3 3 1 2 1 1 4 2 6 5 5 3 1 Without redundant rules 12 X X X X X 14 X 6 X X 3 1 With redundant rules 12 3 3 1 2 1 1 4 2 6 5 5 3 1 Case number 14 15 16 17 18 19 20 21 22 23 24 25 Physician 6 6 1 2 5 8 9 1 4 1 3 4 1 2 1 4 Without redundant rules X X 12 5 X 9 14 13 4 1 2 14 With redundant rules 6 6 1 2 5 8 9 1 4 1 3 4 1 2 1 4 aThe codes of diseases and their translations are: I-Measles, 2-German measles, 3-Chickenpox, 4-Smallpox, 5-Scarlet fever, 6-Exanthem subitum, 7-Fifth disease, 8-Meningococcemia, 9-Rocky MI. spotted fever, 10-Typhus fevers, I I-Infectious mononucleosis, 12-Enterovirus infections, 13-Drug eruptions, and 14-Eczema herpeticum. to be more natural and straightforward to represent knowledge of experts by fuzzy rules. Currently, we are trying to enhance the system by applying multime- dia techniques. Some related studies, including the learning algorithm for the adjustment of membership functions, the improvement of inference methodolo- gies, and the representation of multimedia knowledge bases, are in progress. APPENDIX A: THE REPERTORY GRID FOR ACUTE EXANTHEMA D, - Measles D 2 - German measles D3 - Chickenpox D4 - Smallpox D , - Scarlet D6 - Exanthem subitum D , - Fifth disease D8 - Meningococcemia D9 - Rocky Mt. spotted fever D l 0 - Typhus fevers D,, - Infectious mononucleosis D , , - Enterovirus infections D , , - Drug eruptions D , , - Eczema herpeticum Koplik’s spots 3-4 days of fever 1-2 days of fever Headache Flushed cheeks Vomiting Lymphadenopathy Red Maculopapular Deep lesion Centripetal Circumoral pallor Petechiae Splenomegaly Chills Sore throat Desauamation D l D 2 D 3 D 4 D , D6 Ol 1 3 3 3 3 3 3 N o Koplik’s spots 1 5 3 3 4 1 5 Not 3-4days 3 5 4 4 2 3 5 Not 1-2days 3 3 1 1 3 3 3 No headache 3 3 3 3 3 3 1 No flushedcheeks 5 5 5 5 1 5 5 Novomiting 3 1 3 3 3 3 3 N o Lymphadenopathy 1 2 2 2 1 2 1 N o t r e d 3 3 5 1 3 3 3 Superficial 3 3 1 5 3 3 3 Centrifugal 5 5 5 5 1 5 1 No circumoral pallor 5 5 5 3 5 5 5 N o petechiae 3 3 3 3 3 3 3 N o Splenomegaly 3 5 3 1 3 3 5 N o chills 3 3 3 1 3 3 3 No sore throat 1 5 3 3 1 3 3 No desauamation 552 HWANG DS D9 D I O D I l D 1 2 D 1 3 D14 Koplik's spots 3 3 3 3 3 3 3 3-4days of fever 4 I 1 4 4 4 5 1-2 days of fever 2 4 2 2 1 2 5 Headache 3 1 1 3 3 3 3 Flushed cheeks 3 3 3 3 3 3 3 Vomiting 1 5 5 5 5 5 5 Lymphadenopathy 3 3 3 3 3 3 3 Red Maculopapular 2 2 2 2 2 2 2 Deep lesion 3 3 3 3 3 3 3 Centripetal 3 5 1 3 3 3 3 Circum. pallor 5 5 5 5 5 5 5 Petechiae 1 1 1 3 5 5 3 Splenomegal y 3 3 3 1 3 3 3 Chills 3 1 1 3 3 3 5 Sore throat 3 3 3 1 3 3 5 Desquamation 3 3 3 3 3 3 3 No Koplik's spots Not 3-4 days Not 1-2 days No headache No flushed cheeks No vomiting No Lymphadenopathy Not red Superficial Centrifugal N o circumoral pallor No petechiae No Splenomegaly N o chills No sore throat No desquamation Rules generated from the Repertory Grids by EMCUD: RULE,: I F THEN RULE,, ,: I F THEN RULE,,,: I F THEN RULE,,,: I F THEN RULE,: I F (Koplik'sspots = T) A (daysoffever 5 4 ) A (daysoffever 2 3 ) A (vomiting = Fj A (Maculopapularcolor = brickred) A (circumoralpallor = Fj A (petechiae = F) A (desquamation = T ) Disease = Measles CF = 1 . 0 (Koplik'sspots = Tj ((daysoffever > 4 ) V (daysoffever < 3 ) ) A (vomiting = Fj A (Maculopapularcolor = brickredj A (circumoralpallor = Fj A (petechiae = Fj A (desquamation = T ) Disease = Measles CF = 0.8 (Koplik'sspots = T ) A (daysoffever 5 4 ) A (daysoffever 2 3 ) A (vomiting = F) A (Maculopularcolor = brickred) A (circumoralpallor = F ) A (petechiae = T) A (desquamation = T ) Disease = Measles CF = 0 . 9 (Koplik'sspots = T) A (daysoffever 5 4 ) A (daysoffever 2 3 ) A (vomiting = F) A (Macu1opapularcolor = bri ckred) A (circumoralpallor = F) A (petechiae = F) A (desquamation = F) Disease = Measles CF = 0.8 (daysoffever = 0 ) A (vomi ting = F) A (Lymphadenopathy = T ) A (Maculopapularcofor = pink) A (circumoralpal- l o r = F ) A KNOWLEDGE ACQUISITION 553 THEN RULE2,1: IF THEN RULE2,2: IF THEN RULE,,,: IF THEN RULE,,,: IF THEN RULE,: IF THEN RULE,,,: IF THEN RULE,.,: IF THEN (petechiae = F ) A (chills = F ) A (desquamation = F ) Disease = German Measles CF = 1.0 (daysoffever > 0 ) A (vomiting = F ) A (Lymphade- nopathy = T ) A (Maculopapularcolor =pink) A (circumoralpal- lor = F ) A (petechiae = F ) A (chills = F ) A (desquama- tion = F ) Disease = German Measles CF = 0.8 (daysoffever = 0) A (vomiting = F) A (Lymphade- nopathy = T ) A (Maculopapularcolor = pink) A (circumoralpal- lor = F ) A (petechiae = T ) A (chills = F ) A (desquama- tion = F) Disease = German Measles CF = 0.9 (daysoffever = 0) A (vomiting = F) A (Lymphade- nopathy = T ) A (Maculopapularcolor = pink) A (circumoralpal- lor = F ) A (petechiae = F) A (chills = T ) A (desquama- tion = F ) Disease = German Measles CF = 0.8 (daysoffever = 0) A (vomiting = F ) A (Lymphade- nopathy = T ) A (Maculopapularcofor = pink) A (circumoralpal- l o r = F ) A (petechiae = T ) A (chills = F ) A (desquama- tion = T ) Disease = German Measles CF = 0.8 (daysoffever 5 1) A (daysoffever > 0) A (head- ache = T ) A (vomiting = F ) A (Maculopapufarcolor E {pink, red}) A (lesiondegree = superficial) A (distribution = centripetal) A (circumoralpallor = F ) A (petechiae = F ) Disease = Chickenpox CF = 1.0 ((daysoffever > 1) v (daysoffever = 0)) A (headache = T ) A (vomiting = F ) A (Maculopapularcolor E {pink, red}) A (lesiondegree = superficial) A (distribution = centripetal) A (circumoralpallor = F ) A (petechiae = F ) Disease = chickenpox CF = 0.8 (daysoffever 5 1) A (daysoffever > 0) A (head- ache = F ) A (vomiting = F ) A (Maculopapularcolor E {pink, red}) A (lesiondegree = superficial) A (distribution = centripetal) A (circumoralpallor = F ) A (petechiae = F ) Disease = Chickenpox CF = 0.8 HWANG 554 RULE,,,: IF THEN RULE,: IF THEN RULE,,,: I F THEN RULE,,,: IF THEN RULE,,,: IF THEN RULE,: IF THEN RULE,,,: IF (daysoffever I 1) A (daysoffever > 0) A (head- ache = T ) A (vomiting = F) (Maculopapularcolor E {pink, red}) A (lesiondegree = superficial) A (distribution = centripetal) A (circumoralpallor = F ) r\ *petechiae = T) Disease = Chickenpox CF = 0 . 9 (daysoffever < 4 ) A (daysoffever 2 3 ) A (head- ache = T ) A (vomiting = F ) A (Maculopapularcolor E {pink, red}) A (lesiondegree = deepseated) A (distribution = centrifugal) A (circumoralapplor = F ) A (chills = T ) A (sore- throat = T ) Disease = Smallpox CF = 1.0 ( (daysoffever 2 4 ) v (daysoffever < 3) ) A (headache = T ) A (vomiting = F ) A (Maculopapularcolor E {pink, red}) A (lesiondegree = deepseated) A (distribution = centrifugal) A (circumoralpallor = F ) A (chills = T ) A (sore- throat = T ) Disease = Smallpox CF = 0. 8 (daysoffever < 4 ) A (daysoffever 2 3) A (head- ache = F ) A (vomiting = F) A Maculopapularcolor E {pink, red}) A (lesiondegree = deepseated) A (distribution = centrifugal) A (circumoralpallor = F ) A (chills = T ) A (sore- throat = T ) Disease = Smallpox C F = 0.8 (daysoffever < 4 ) A (daysoffever 2 3) A (head- ache = T ) A (vomiting = F ) A (Maculopapularcolor E {pink, red}) A (lesiondegree = deepseated) A (distribution = centrifugal A (circumoralpallor = F ) (chills = T ) A (sore- throat = F ) Disease = Smallpox CF = 0 . 9 (daysoffever 5 2 ) A (daysoffever > 0 . 5 ) A (vom- iting = T ) A (Maculopapularcolor = red) A (circumoralpallor = T ) A (petechiae = F ) A (desquamation = T ) Disease = Scarlet fever CF = 1.0 ((daysoffever > 2 ) v (daysoffever 5 0 . 5 ) ) A (vomiting = T ) A (Maculopapularcolor = red) A (circumoralpall or = T ) A (petechiae = F ) A (desquamation = T) KNOWLEDGE ACQUISITION 555 RULE,, 2 : RULE, : RULE,, : RULE,, 2 : RULE, : RULE,, : RULE, : RULE,, : RULE, : THEN IF THEN IF THEN IF THEN IF THEN IF THEN IF THEN IF THEN IF THEN IF Disease = Scarlet fever CF = 0.8 (daysoffever 5 2 ) A (daysoffever > 0 . 5 ) A (vom- iting = T ) A (Maculopapularcolor = red) A (circumoralpallor = T ) A (petechiae = T ) A (desquamation = T ) Disease = Scarlet fever CF = 0.9 (daysoffever c: 4) A (daysoffever 2 3) A (vomit- ing = F ) A (Maculopapularcolor = pink) A (circumoralpal- lor = F ) A (petechiae = F ) Disease = Exanthem subitum CF = 1. 0 ( (daysoffever > 4) v (daysoffever < 3 ) ) A (vom- iting = F ) A (Maculopapularcolor = pink) A (circumoralpal- l o r = F ) A (petechiae = F ) Disease = Exanthem subitum CF = 0.8 (daysoffever 5 4) A (daysoffever 2 3) A (vomit- ing = F ) A (Maculopapularcolor = pink) A (circumoralpal- lor = F ) A (petechiae = T ) Disease = Exanthem subitum CF = 0 . 9 (daysoffever = 0 ) A (flushedcheeks = T ) A (vom- iting = F ) A (Macul opapul arcol o r = red) A (circumoralpall or = T ) A (petechiae = F) A (chills = F) Disease = Fifth disease CF = 1. 0 (daysoffever = 0 ) A (flushedcheeks = T ) A (vom- iting = F ) A (Macul opapul arcol or = red) A (circumoralpall or = T ) A (petechiae = T ) A (chills = F) Disease = Fifth disease CF = 0 . 8 (daysoffever < 1) A (daysoffever > 0 ) A (vomit- ing = F ) A (Maculopapularcolor E {pink, red}) A (circumoral- pallor = F ) A (petechiae = T ) Disease = Meningococcemia CF = 1.0 ( (daysoffever > 1) v (daysoffever = 0 ) ) A (vom- iting = T ) A (Maculopapularcolor E {pink, red}) A (circumoral- pallor = F ) A (petechiae = T ) Disease = Meningococcemia CF = 0.8 (daysoffever 5 4) A (daysoffever 2 3) A (head- ache = T)A (vomiting = Fj A (Maculopapularcolor E {pink, red} j A (distribution = centrifugal) A (circumoralpal- lor = F j A 556 HWANG THEN RULE,, ,: IF THEN RULE,,: IF THEN RULE,,,,: IF THEN RULE,,: IF THEN RULE,,,,: IF THEN RULE,,,,: IF THEN RULE,,,,: IF THEN RULE,,: IF (petechiae = T) A (chills = T) Disease = Rocky Mt. spots fever CF = 1.0 (daysoffever 5 4 ) A (daysoffever 2 3) A (head- ache = F ) A (vomiting = F ) A (Maculopapularcolor E {pink, red}) A (distribution = centrifugal) A (circumoralpal- lor = F ) A (petechiae = T) A (chills = T ) Disease = Rocky Mt. spots fever CF = 0.8 (daysoffever 5 4) A (daysoffever 2 3) A (head- ache = T ) A (vomiting = F ) A (Maculopapularcolor E {pink, red}) A (distribution = centripetal) A (circumoralpal- lor = F ) A (petechiae = T ) A (chills = T) Disease = Typhus fevers CF = 1.0 (daysoffever 5 4) A (daysoffever 2 3) A (head- ache = F ) A (vomiting = F) A (Maculopapularcolor E {pink, red}) A (distribution = centripetal) A (circumoralpal- l o r = F ) A (petechiae = T) A (chills = T) Disease = Typhus fevers CF = 0.8 (daysoffever I 1) A (daysoffever > 0 ) A (vomit- ing = F ) A (Maculopapularcolor E {pink, red}) (circumoral- pallor = F ) A (Splenomegaly = T ) A (sorethroat = T) Disease = Infectious mononucleosis CF = 1.0 ((daysoffever > 1) v (daysoffever = 0)) A (vom- iting = F ) A (Maculopapularcolor E {pink, red}) A (circumoral- pallor = F ) A (Splenomegaly = T) A (sorethroat = T ) Disease = Infectious mononucleosis CF = 0.8 (daysoffever 5 1) A (daysoffever > 0) A (vomit- ing = F ) A (Maculopapularcolor E {pink, red}) A (circumoral- pallor = T ) A (Splenomegaly = T ) A (sorethroat = T) Disease = Infectious mononucleosis CF = 0.8 (daysoffever 5 1) A (daysoffever > 0) A (vomit- ing = F ) A (Maculopapularcolor E {pink, red}) A (circumoral- pallor = F ) A (Splenomegaly = T) A (sorethroat = F ) Disease = Infectious mononucleosis CF = 0 . 9 (daysoffever 5 2) A (daysoffever 2 1) A (vomit- ing = F ) A (Maculopapularcolor = pink) A (circumoralpal- lor = F ) A (petechiae = F ) K N O W L E D G E ACQUISITION 557 THEN RULE,,,,: I F RULE,, : RULE,, , 1 RULE,, : RULE,, , : RULE, : RULE, : RULE, : RULE, : RULE, : RULE, : RULE, : RULE, : THEN I F THEN I F THEN I F THEN I F THEN I F THEN I F THEN I F THEN I F THEN I F THEN I F THEN I F THEN I F Disease = Enterovirus infections CF = 0.8 (daysoffever 5 2) A (daysoffever 2 1) A (vomit- ing = F ) A (Macul opapul arcol o r = pink) A (circumoralpal- l o r = F ) A (petechiae = T) Disease = Enterovirus infections CF = 0.6 (daysoffever 5 1) A (daysoffever > 0 ) A (vomit- ing = F)A (Maculopapularcolor = pink) (circumoralpall or = F ) A (petechiae = F ) Disease = Drug eruptions CF = 0.8 (daysoffever 5 1) A (daysoffever > 0 ) A (vomit- ing = F ) A (Maculopapularcol or = pink) (circumoralpall or = F ) A (petechiae = T) Disease = Drug eruptions CF = 0 . 6 (daysoffever = 0) A (vomiting = F ) A (Maculopapularcolor = pink) A (circumoralpal- l o r = F ) A (chills = F) A (sorethroat = F) Disease = Eczema herpeticum CF = 0.8 (daysoffever = 0) A (vomiting = F ) A (Maculopapularcolor = pink) A (circumoralpal- l o r = F ) A (chills = F) A (sorethroat = T) Disease = Eczema herpeticum CF = 0.6 (WBC f low) Disease = Measles CF = 0.2 (WBC # low) Disease = Exanthem subitum CF = 0.2 ( WBC # low) A ( WBC f normal ) Disease = German measles CF = 0.2 ( Whil eBallCel lamount i5000) WBC = low CF = 1.0 ( Whi 1 eBal1 Cell amount 2 5000) A (WhileBallCellamount 5 10000) WBC = normal CF = 1.0 (WhileBallCellamount > 10000) WC = highCF = 1.0 APPENDIX B: FUZZY RULES (Koplik’s spots is definite) A (fever is seri- ous) A (vomiting is not apparent)A (Maculopapular color is brick red) A (circu- moral pallor is not apparent) A (petechiae is not apparent) A (desquamation is apparen t 1 Disease = Measles CF = 1.0 (fever is none) A (petechiae is not apparent)A (vomiting is not apparent) A (Lymphadenopathy is definite) A 558 HWANG THEN RULE,: I F THEN RULE,: I F THEN RULE,: I F THEN RULE,: I F THEN RULE,: I F THEN RULE,: I F THEN RULE,: I F THEN ( M a c u l o p a p u l a r c o l o r i s p i n k ) A ( c i r c u m o r a l p a l - l o r i s n o t a p p a r e n t ) A ( c h i l l s i s n o t a p p a r e n t ) A ( d e s q u m a t i o n i s n o t a p p a r e n t ) D i s e a s e = German M e a s l e s C F = 1 . 0 ( f e v e r i s l i g h t ) A ( h e a d a c h e i s a p p a r e n t ) A ( v o m i t i n g i s n o t a p p a r e n t ) A ( M a c u l o p a p u l a r c o l o r i s more or l e s s r e d ) A ( l e s i o n d e g r e e i s s u p e r f i c i a l ) A ( d i s t r i b u t i o n i s c e n t r i p e t a l ) A ( c i r c u m o r a l p a l l o r i s n o t a p p a r e n t ) A ( p e t e - c h i a e i s n o t a p p a r e n t ) D i s e a s e = C h i c k e n p o x C F = 1.0 ( f e v e r i s s e r i o u s ) A ( h e a d a c h e i s a p p a r e n t ) A ( s o r e t h r o a t i s a p p a r e n t ) ( v o m i t i n g i s n o t a p p a r e n t ) A ( M a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) A ( l e s i o n d e g r e e i s d e e p s e a t e d ) A [ d i s t r i b u t i o n i s c e n t r i f u g a l ) A ( c i r c u m o r a l p a l l o r i s not a p p a r e n t ) A ( c h i l l s i s a p p a r e n t ) A D i s e a s e = S m a l l p o x C F = 1. 0 ( f e v e r i s m o r e o r l e s s s e r i o u s ) A ( v o m i t i n g i s a p p a r e n t ) A ( M a c u l o p a p u l a r c o l o r i s r e d ) A ( c i r c u m o r a l p a l - l o r i s a p p a r e n t ) A ( p e t e c h i a e i s n o t a p p a r e n t ) A ( d e s q u a m a t i o n i s a p p a r e n t ) D i s e a s e = S c a r l e t f e v e r C F = 1 . 0 ( f e v e r i s s e r i o u s ) A ( v o m i t i n g i s n o t a p p a r e n t ) A ( M a c u l o p a p u l a r c o l o r is p i n k ) ( c i r c u m o r a l p a l l o r i s n o t a p p a r e n t ) A ( p e t e - c h i a e i s not a p p a r e n t ) D i s e a s e = E x a n t h e m subitum C F = 1.0 ( f e v e r i s n o n e ) A ( f l u s h e d cheeks i s a p p a r e n t ) A ( v o m i t i n g i s not a p p a r e n t ) A [ M a c u l o p a p u l a r color i s r e d ) A ( c i r c u m o r a l p a l - l o r i s a p p a r e n t ) A ( p e t e c h i a e i s n o t a p p a r e n t ) A [ c h i l l s i s not a p - p a r e n t ) D i s e a s e = F i f t h d i s e a s e C F = 1.0 ( f e v e r i s l i g h t ) A ( v o m i t i n g i s a p p a r e n t ) A ( p e t e c h i a e i s a p p a r e n t ) ( M a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) A ( c i r c u - m o r a l p a l l o r i s n o t a p p a r e n t ) A D i s e a s e = M e n i n g o c o c c e m i a C F = 1. 0 ( f e v e r i s s e r i o u s ) A ( h e a d a c h e i s a p p a r e n t ) A ( v o m i t i n g i s not a p p a r e n t ) A ( M a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) A ( d i s t r i b u t i o n i s c e n t r i f u g a l ) A ( c i r c u m o r a l p a l - l o r i s not a p p a r e n t ) A ( p e t e c h i a e i s a p p a r e n t ) A ( c h i l l s i s a p p a r e n t ) D i s e a s e = R o c k y Mt. s p o t s f e v e r C F = 1 . 0 KNOWLEDGE ACQUISITION 559 RULE,,: RULE,, : RULE,, : RULE,,: RULE,, : I F THEN I F THEN I F THEN I F THEN IF THEN ( f e v e r i s s e r i o u s ) A ( h e a d a h c e i s a p p a r e n t ) A ( v o m i t i n g i s n o t a p p a r e n t ) A ( M a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) A ( d i s t r i b u t i o n i s c e n t r i p e t a l ) A ( c i r c u m o r a l p a l - l o r i s n o t a p p a r e n t ) A ( p e t e c h i a e i s a p p a r e n t ) A ( c h i l l s i s a p p a r e n t ) D i s e a s e = T y p h u s f e v e r s C F = 1 . 0 ( f e v e r i s l i g h t ) A ( v o m i t i n g i s n o t a p p a r e n t ) A ( M a c u l o p a p u l a r c o l o r i s p i n k o r r e d ) A ( c i r - c u m o r a l p a l l o r i s not a p p a r e n t ) A ( S p l e n o m e g a l y i s a p p a r e n t ) A ( s o r e t h r o a t i s a p - p a r e n t ) D i s e a s e = I n f e c t i o u s m o n o n u c l e o s i s C F = 1 . 0 ( f e v e r i s m o r e o r l e s s s e r i o u s ) A ( v o m i t i n g i s not a p p a r e n t ) A ( M a c u l o p a p u l a r c o l o r is more or l e s s p i n k ) A ( c i r c u m o r a l p a l l o r i s n o t a p p a r e n t ) ( p e t e c h i a e i s n o t a p p a r e n t ) D i s e a s e = Enterovirus i n f e c t i o n s C f = 0 . 8 ( f e v e r i s l i g h t ) A ( v o m i t i n g i s n o t a p p a r e n t ) A ( p e t e c h i a e i s not a p p a r e n t ) ( M a c u l o p a p u l a r color i s more or l e s s p i n k ) ( c i r - c u m o r a l p a l l o r i s not a p p a r e n t ) / D i s e a s e = Drug e r u p t i o n s C F = 0 . 8 [fever i s none) A ( v o m i t i n g i s not a p p a r e n t ) A ( M a c u l o p a p u l a r c o l o r i s more o r l e s s p i n k ) A ( c i r c u m o r a l p a l l o r i s not a p p a r e n t ) ( c h i l l s i s not a p p a r e n t ) A ( s o r e t h r o a t i s n o t a p p a r e n t ) D i s e a s e = Eczema h e r p e t i c u m C F = 0 . 8 This study was supported by the National Science Council, Republic of China, under the contract number: NSC82-0408-E-009-049. References 1. G.J. Klir and T.A. Folger, Fuzzy Sets, Uncertainty, andInformation, Prentice-Hall, New Jersey 1988. 2 . J . Boose, “A knowledge acquisition program for expert systems based on personal construct psychology,” Int. J . Man-Mach. S t u . , 23, 495-525. 3. J. Boose and J.M. Bradshaw, “NeoETS: capturing expert system knowledge in hierarchical rating grids,” IEEE Expert S y s t e m in Government Symposium, 1986. 4. J. Boose and J.M. Bradshaw, “Expertise transfer and complex problems: using AQUINAS as a knowledge acquisition workbench for knowledge-based systems,” Int. J . Man-Mach. S t u . , 26, 3-28 (1987). 5. L. Esheman, D. Ehret, J . McDermott, and M. Tan, “MOLE: a tenacious knowledge acquisition tool,” Int. J . Man-Mach. S t u . , 26, 41-54 (1987). 6. K.M. Ford, F.E. Petry, J.R. Adams-Webber, and P.J. Chang, “An approach to knowledge acquisition based on the personal construct systems,” IEEE Trans. K n o w / . and D a t a E n g . , 3(1), 78-88 (1991). 7. K . M . Ford, A. Canas, J. Jones, H. Stahl, J. Novak, and J.R. Adams-Webber, HWANG “ICONKAT: an integrated constructivist knowledge acquisition tool,” Int. J . Knowl. A c q . , 3 , 215-236 (1991). 8. B.R. Gaines, “An overview of knowledge acquisition and transfer,” Int. J . 9. S.T. Gruber, The Acquisition of Strategic Knowledge, Academic Press Inc., San Diego, 1988. 10. G. Kahn, S . Nowlan, and J . McDermott, “MORE: an intelligent knowledge acquisi- tion tool,” International Joint Conference of Art$cial Intelligence, 31, 495-5 15 (1985). 11. S. Marcus, “Taking backtracking with a grain of SALT,” Int. J . Man-Mach. S t u . , 12. R.M. O’Bannon, “An intelligent aid to assist knowledge engineers with interviewing experts,” IEEE Western Conference on Expert S y s t e m , 1987. 13. M.L.G. Shaw and B.R. Gaines, “KITTEN: knowledge initiation and transfer tools for experts and novices,” Int. J . Man-Mach. S t u d . , 27, 251-280 (1987). 14. J.M. Bradshaw, K.M. Ford, J.R. Adams-Webber, and J.H. Boose, “Beyond the repertory grid: new approaches to constructivist knowledge acquisition tool develop- ment,” Int. J . Intell. Sys., 8, 287-333 (1993). 15. L . E . Wood and J.M. Ford, “Structuring interviewings with experts during knowl- edge elicitation,” Int. J . Intel. Sys., 8, 71-90 (1993). 16. G.A. Kelly, The Psychology of Personal Constructs, Norton, New York, 1955. 17. J . Giarratano and G . Riley, Expert Systems: Principles and Programming, PWS- KENT Publishing Company, Boston, 1989. 18. G.J. Hwang and S.S. Tseng, “EMCUD: a knowledge acquisition method which captures embedded meanings under uncertainty,” Int. J . Man-Mach. Stu., 33, Man-Mach. Stti., 26, 453-472 (1987). 26, 383-398 (1987). 431-451 (1990). 
work_fpkizchfqbclvbn2xcyfpjutpi	----	Exploring expert system success factors for business process reengineering Ž .J. Eng. Technol. Manage. 15 1998 179–199 Exploring expert system success factors for business process reengineering Youngohc Yoon a, Tor Guimaraes b,), Aaron Clevenson c a Department of Information Systems, Virginia Commonwealth UniÕersity, Richmond, VA 23284-4000, USA b J.E. Owen Chair of Excellence, Tennessee Technological UniÕersity, CookeÕille, TN 38505, USA c E.I. DuPont de Nemours, One Kingwood Place, Suite 215, Kingwood, TX 77339, USA Abstract Ž .Business process reengineering BPR has become the buzzword representing dramatic changes to the business processes of organizations trying to quickly preempt or react to market opportuni- ties and competition. Much of the changes are enabled by computer-based technology such as Ž .expert systems ES providing a unique opportunity to study significant implementations of the technology within a relatively short time. Eight ES implementation success factors proposed in the literature were empirically tested in this study in terms of their direct and indirect importance to the benefits from using ES in BPR. Sixty-two ES applications within E.I. Dupont de Nemours dealing with business process changes significant enough to be called BPR were used. Despite the relatively small sample size, four of the eight success factors were corroborated: user satisfaction with the ES, the difficulty of the business problem addressed, the degree of user involvement in the ES implementation process, and characteristics of the ES shells. q 1998 Elsevier Science B.V. All rights reserved. Ž .Keywords: Expert systems; Knowledge-based systems KBS ; Business process reengineering; Benefits; ES success factors 1. Introduction Ž .An expert system ES is a computer system employing expert knowledge to attain high levels of performance in solving the problems within a specific domain area. By encapsulating expert knowledge and experience, ES enables organizations to support ) Corresponding author. Tel.: q1-615-372-3385; fax: q1-615-372-6249; e-mail: tg5596@tntech.edu. 0923-4748r98r$19.00 q 1998 Elsevier Science B.V. All rights reserved. Ž .PII S 0 9 2 3 - 4 7 4 8 9 8 0 0 0 1 1 - 3 ( )Y. Yoon et al.r J. Eng. Technol. Manage. 15 1998 179–199180 important decision making and improve organization productivity. Today, ES have demonstrated their potential, gained credibility, and are being widely used to solve a Žvariety of business problems in industry and government Hayes-Roth and Jacobstein, .1994 . Due to its capability to effectively improve the way an organization does business, many ES applications have been used to change a variety of business processes ŽAcorn and Walden, 1993; Hamscher, 1994a; Lazarus et al., 1993; McManus and .Garland, 1993; Nguyen and Czerwinsksi, 1993; Pierson and Gallant, 1993 . There are many examples of ES applications used to dramatically change old business processes. In one case, SMART at Compaq has enhanced the effectiveness of its customer support staff by distributing problem-solving knowledge to the entire Ž .support staff at the point it is needed Acorn and Walden, 1993 . Moreover, Quick- Source has enabled Compaq to empower its customers with expert knowledge which Žallows them to solve advanced network printer problems entirely on their own Nguyen .and Czerwinsksi, 1993 . With the use of ES technology, Compaq has reengineered its customer support strategy and implementation. In another organization, MPSA and COLES are two interrelated systems which offer automated expert support to AT & T’s scheduling and customer service staffs by improving and expanding the utility of Ž .existing legacy mainframe systems McManus and Garland, 1993 . At Bell Atlantic, SNNS was developed to provide automated support for the sales-service negotiation Ž .process Carr et al., 1994 . This system has significantly improved sales efficiency and optimized customer contact time. Due to their effectiveness in improving existing business processes, ES have been recognized as important implementation vehicles for business process reengineering Ž . Ž .BPR Friedenberg and Rice, 1994; Hamscher, 1994b . However, in many cases ES development and implementation has failed for a variety of reasons, including instances Žwhen systems were rejected by their target user communities Coats, 1988; Keyes, .1989b; Sloane, 1991 . Thus, a central issue deals with the factors affecting ES success or failure under the time constraints created by BPR projects. This also becomes increas- ingly important as managers contemplate investing company resources, time and effort into establishing the infrastructure for adopting or expanding the use of ES technology. The purpose of this study is to empirically test the various factors affecting ES success when the system is used for BPR, as contrasted with the use of ES for merely automating existing business processes or making relatively small changes. Much of the research on MIS implementation research has been focused on identify- ing the factors which appear to be conducive to either success or failure of other types of Ž .computer-based information systems Guimaraes et al., 1992; Liang, 1986 . In the Žcontext of ES, many success factors have been reported Ignizio, 1991; Keyes, 1989b; .Prerau, 1990; O’Neal, 1990; Turban, 1992b; Yoon et al., 1995 . Although most of the findings have been on the basis of personal opinion, some recent studies have empiri- Ž .cally tested ES success factors Yoon et al., 1995 . The independent variables in this study, the factors related to ES success reported in the literature, were categorized into Ž . Ž .seven groups: problem importance, problem difficulty, developer s skill, end-user s characteristics, shell characteristics, user involvement, and management support. Some of these success factors are not unique to ES implementations. The vast majority of the Ž . Ž .studies deal with transaction processing systems TPS , decision support systems DSS , https://isiarticles.com/article/434 
work_fqlqk53qyrc3le4yaygrmbhaoy	----	Automatic recognition of industrial tools using artificial intelligence approach Expert Systems with Applications xxx (2013) xxx–xxx Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Automatic recognition of industrial tools using artificial intelligence approach Tomasz Les a, Michal Kruk c, Stanislaw Osowski a,b,⇑ a Warsaw University of Technology, Warsaw, Koszykowa 75, Poland b Military University of Technology, Warsaw, Kaliskiego 2, Poland c University of Life Sciences, Warsaw, Nowoursynowska 165, Poland a r t i c l e i n f o a b s t r a c t Keywords: Pattern recognition Diagnostic features Data mining Classification Industrial tools 0957-4174/$ - see front matter � 2013 Elsevier Ltd. A http://dx.doi.org/10.1016/j.eswa.2013.02.030 ⇑ Corresponding author at. Warsaw University of Te wa 75, Poland. Tel.: +48 222347235. E-mail address: sto@iem.pw.edu.pl (S. Osowski). Please cite this article in press as: Les, T., et al. A tions (2013), http://dx.doi.org/10.1016/j.eswa.20 The paper presents an automatic approach to the recognition of the industrial tools on the basis of their image registered by the camera. The solved tasks include: the automatic localization of the tools in the image, preprocessing of the image (binarization, noise removal, filling the holes, normalization, etc.), generation of the numerical descriptors, verification of descriptors in the role of the diagnostic features, selection of the features and final stage of classification of the tools. The efficiency of the developed system has been verified on the example of exemplary set of industrial tools and the results of this verification are presented and discussed in the paper. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction Machine vision systems are being increasingly used for sophis- ticated applications such as recognition and classification of differ- ent processes. An important milestone in the development of ‘‘intelligent’’ inspection systems has been the rapid growth of com- puting power in recent years, coupled with the idea that we could successfully emulate the low-level mechanisms of the brain. Thanks to the methodology offered by the artificial intelligence methods, they are able to solve difficult computational problems arising in the task of object recognition. The natural human visual system has the ability to recognize an object despite changes in the object’s position in the input field, its size, or its orientation. For many industrial applications involving classification of elements under recognition, the machine vision systems must also have this ability. Many industrial applications of machine vision allow objects to be identified and classified by their boundary contour or silhouette (Kim & Nam, 1995; Nabhani & Shaw, 2002). An example of it is a robot recognition of industrial parts and tools in the factor environment. In most assembly or sorting lines several different types of tools of the shapes known in advance are handled. Their recognition belongs to the problem of object recognition or pattern matching (Duda, Hart, & Stork, 2003). This work presents and investigates the application of machine vision methods in the processing of single camera multi-positional images for 2D object recognition and classification. The boundary contour information was chosen as the basis for representing the ll rights reserved. chnology, Warsaw, Koszyko- utomatic recognition of indust 13.02.030 chosen industrial tools at their recognition. The appropriately preprocessed shape of the object may form the set of diagnostic features, used as the input information to the classifier, responsible for the final recognition of the object. In our approach to the recognition/classification of industrial components we divide the process into two major stages. The first one is the extraction of the component of interest from the regis- tered image and characterization of it by the numerical features well representing its shape. In the second stage the numerical descriptors, called also the diagnostic features, are put to the input of automatic classifier, responsible for the final recognition. Many recent solutions of classifiers, including artificial neural networks may perform the role of final recognition. In this work we will ap- ply two most efficient solutions: the random forest of decision trees and Support Vector Machine (SVM). The numerical experiments will be done on the set of simple industrial tools differing by the shape, size, location and orienta- tion in the input field. The preprocessing steps will extract the component from the image, convert them to the normalized fea- tures independent on their location, orientation and size. After the selection process we get the most important features, which are used as the input information to the classifier, performing the final recognition. 2. The preprocessing stage The image of the input field may contain the interesting object represented in general by RGB color components and some smaller elements representing the noise. The main task of this stage is to localize the object and preprocess it in the way enabling to get its numerical characterization. rial tools using artificial intelligence approach. Expert Systems with Applica- http://dx.doi.org/10.1016/j.eswa.2013.02.030 mailto:sto@iem.pw.edu.pl http://dx.doi.org/10.1016/j.eswa.2013.02.030 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa http://dx.doi.org/10.1016/j.eswa.2013.02.030 2 T. Les et al. / Expert Systems with Applications xxx (2013) xxx–xxx 2.1. Binarization of the image The binarization of the image is the process of converting the image into the binary one, in which each pixel has one of two pos- sible intensity levels: either one or zero. The first step is to convert the color into gray levels. This is done for each pixel in the position (x, y) by using standard relation (Gonzalez & Woods, 2011) fðx; yÞ¼ 0:3Rðx; yÞþ 0:59Gðx; yÞþ 0:11Bðx; yÞ ð1Þ where R, G and B represent the three color characterization of the pixel. In the second stage the 256 gray levels are converted to a black and white image by choosing the appropriate threshold value, and classifying all pixels with values above this threshold as white, and all other pixels as black. The important problem is to select the correct threshold. It depends on the lighting conditions at the image registration as well as on the type of the image. The threshold value may be selected on the way of many trials or by applying the adap- tive method, for example the Otsu method (Otsu,1979). 2.2. Localization of an object The next step of processing is the localization of the tool in the input field image. The basic assumption is that the tool occupies the largest area in the field. Therefore we scan the whole field searching for the largest compact region. This algorithm is called the grain growth (Gonzalez & Woods, 2011; Soille, 2003) and is re- lied on finding all pixels forming compact region. The algorithm starts form the middle of the image looking for the black pixel. Next we move to all possible black pixels adjacent to the already found, increasing in this way the grain. The largest found compact region composed of only black pixels is taken as our object of interest. All other pixels of the image are treated as the background and changed to the white color. The last step is to fill the holes in the discovered object. This is done by applying the morphological operation of opening and closing (Soille, 2003). Fig. 1 illustrates the process of grain growth applied for discov- ering the shape of a grinder. The first image (a) illustrates 1% advancement, the second (b) 40% of advancement of grain growth, the third (c) 100% of grain growth and the last one (d) the shape of the tool after filling the holes. 2.3. Definition of numerical descriptors of the image After recognizing the shape of the object the next step is to de- fine the numerical descriptors characterizing the object. In our solution we have relied on the geometrical descriptors. The basis of this description is formed by the geometrical characterization of the object, especially the real and convex areas of the object, the real and convex perimeters, as well as box counting dimension (Schroeder, 2006). The real area of the object is treated as the total number of black pixels forming the object. The perimeter is the number of boundary pixels of the object. The convex parameters (area and perimeter) have been defined using the concept of convex envelope fixed on (a) (b) Fig. 1. The illustration of the process of grain growth for discovering the grinder in the in (d) the final shape after filling the holes. Please cite this article in press as: Les, T., et al. Automatic recognition of indust tions (2013), http://dx.doi.org/10.1016/j.eswa.2013.02.030 the boundary points of the shape of the object (Cormen, Leiserson, Rivest, & Clifford, 2009). The box counting dimension belongs to one of the measures fre- quently used in fractal theory. It is a method of analyzing data of complex patterns by breaking the object into smaller and smaller pieces called boxes, and analyzing the pieces at each smaller scale. The aim of this process is to examine how observations of detail change with scale. In principle it measures how the length of the complex curve is changing when the measurement is performed with the increased accuracy (Schroeder, 2006). Let us assume that the object under characterization is placed on the surface covered with the set of regular cubes (squares in 2-dimensional space) of the size e. Then we count the number of cubes that contain any fragment of the object. This number is evi- dently dependent on the size of e. Let us denote it as N(e). Changing the size e we get different values of N(e), generally increasing when e is decreased. In practice instead of e we use the number s2 of elementary cubes, each of the size e, fixed on the considered analyzed squared area. The value of s is inversely proportional to e. In this way the definition of the box counting dimension d can be written in the form d ¼ lim s!1 logðNðsÞÞ logðsÞ ð2Þ The value d, interpreted as the slope of the curve log(N(s)) with re- spect to log(s), represents the box-counting dimension, characteriz- ing the complexity of the curve. To use the box-counting notion we have to define only one parameter – the number of boxes, depen- dent on s. This number defines also the size of the elementary bin- ary matrix s � s. We check if any fragment of the curve enters each box. If yes, we put one in proper location of the matrix, in other case zero. In this way we get the number N(s). Using the relation (2) we get the value of the box counting dimension. The box counting dimension can form the direct numerical descriptor, since according to the definition it represents the relative value. On the other hand from the analysis of the images it is evident that the basic geometrical parameters (area and perimeter) could not be used as the direct descriptors, since their values change with the size of the object. Instead we have use some derivatives of them, defined in a relative way. Let us denote the directly mea- sured parameters by the following symbols: A – real area of the object Ac – convex area of the object P – perimeter of the object Pc – convex perimeter of the object B – box counting dimension On the basis of them we have defined the following descriptors, being the potential diagnostic features (Kruk, Osowski, & Koktysz, 2007) � Circularity ratio Fc ¼ 4pA P ð3Þ (c) (d) put field: (a) 1% of advancement, (b) 40% of advancement, (c) 100% of advancement, rial tools using artificial intelligence approach. Expert Systems with Applica- http://dx.doi.org/10.1016/j.eswa.2013.02.030 T. Les et al. / Expert Systems with Applications xxx (2013) xxx–xxx 3 � Convex circularity ratio Please tions ( Fcc ¼ 4pAC PC ð4Þ � Compactness factor Fh ¼ P2 A ð5Þ � Convex compactness factor Fch ¼ P2c Ac ð6Þ � Corrugation factor Fce ¼ Pc P ð7Þ � Shredding factor Fr ¼ A Ac ð8Þ � The ratio of B and P Fbp ¼ B AP ð9Þ � The ratio of B and Pc Fbpc ¼ B Pc ð10Þ � The ratio of B and A Fba ¼ B A ð11Þ � The ratio of B and Ac Fbac ¼ B Ac ð12Þ These 10 descriptors and the box counting dimension form the set of 11 potential diagnostic features used by the classifier to recog- nize the objects. 2.4. Selection of the features The automatically generated descriptors may have different impact on the classification process (Duda et al., 2003; Guyon & Elisseeff, 2003). Some of them may occupy the same values for different classes and some represent the noise from the point of view of pattern recognition. Good feature should be characterized by the stable values for samples belonging to the same class and at the same time they should differ significantly for different classes. Thus the important problem in the classification and machine learning is to find out the features of the highest importance for the problem solution. Observe that the elimination of some features leads to the reduction of the dimensionality of the feature space and improvement of performance of the classifier in the testing mode for the data not taking part in learning. cite this article in press as: Les, T., et al. Automatic recognition of indust 2013), http://dx.doi.org/10.1016/j.eswa.2013.02.030 From many known techniques of feature selection (Duda et al., 2003; Guyon & Elisseeff, 2003) like principal component analysis, analysis of correlation existing among features, correlation be- tween the features and the classes, application of feature ranking by applying the linear SVM, the analysis of mean and variance of the features belonging to different classes, we have chosen the last one. The variance of the features corresponding to members of one class should be as small as possible. On the other hand, to distin- guish between different classes, the positions of means of feature values for the data belonging to different classes should be sepa- rated from each other as much as possible. Both requirements are combined together to form the discrimination coefficient SAB(f) defined for the feature f at recognition of two objects belonging to different classes A and B SABðfÞ¼ jcAðfÞ� cBðfÞj rAðfÞþ rBðfÞ ð13Þ In this definition cA and cB are the mean values of the feature f in the class A and B, respectively. The variables rA and rB represent the standard deviations determined for the respective class. The large value of SAB(f) indicates good potential separation ability of the fea- ture f for these two classes. On the other side its small value means that this particular feature is not good for the recognition between classes A and B. The set of descriptors of highest values of discrim- ination coefficients form the optimal set of features. 3. Classification systems The selected set of features is the diagnostic information put to the input of the classifiers. To get the best results of pattern recog- nition we have to apply the most efficient classifiers. According to the actual knowledge to the best solutions of the classifiers belong the Support Vector Machine (Schölkopf & Smola, 2002; Vapnik, 1998) and Breiman random forest of the decision trees (Breiman, 2001). These two systems of classifiers implemented in Matlab (Matlab user manual with toolboxes, 2012) will be applied and studied in the paper. Basically the SVM is a linear machine, working in the high dimensional feature space formed by the non-linear mapping of the N-dimensional input vector x into a L-dimensional feature space (L > N) through the use of a kernel function K(x, xi). It is known as an excellent classifier of good generalization ability (Vapnik, 1998). The learning problem of SVM is formulated as the task of separating the learning vectors into two classes of the destination values either di = 1 (one class) or di = �1 (the opposite class), with the maximal separation margin. The separation margin formed in the learning stage according to the assumed value of the regularization constant C provides some immunity of this classifier to the noise, inevitably contained in the testing data. The great advantage of SVM is the unique formulation of learn- ing problem leading to the quadratic programming with linear constraints, which is easy to solve. The SVM of the Gaussian kernel has been used in our application. The hyperparameters r of the Gaussian function and the regularization constant C are usually ad- justed by repeating the learning experiments for the set of their predefined values and choosing the best one at the validation data sets. The typical values of these parameters for the normalized in- put data are c � 1 and C � 1000. To deal with a problem of many classes the one against one or one against all approaches working on a principle of the majority voting (Schölkopf & Smola, 2002) are usually applied. In our solution we have applied the one against one approach, since this approach lead to the better total results of recognition. rial tools using artificial intelligence approach. Expert Systems with Applica- http://dx.doi.org/10.1016/j.eswa.2013.02.030 4 T. Les et al. / Expert Systems with Applications xxx (2013) xxx–xxx The Breiman (Breiman, 2001) random forest is an ensemble of many decision trees and outputs the class pointed by the majority of the individual trees. The method combines ‘‘bagging’’ idea and the random selection of features for each node in order to construct a collection of decision trees with controlled variation. Each tree in the forest is constructed in a way providing the highest degree of independence. Assume that the number of training cases is p, and the number of input variables in the classifier be N. Denote by m the number of input variables to be used to determine the decision at a node of the tree; m should be less than N. Choose a training set for the tree by choosing n times (with replacement) from all p available train- ing cases. Use the rest of the cases to estimate the error of the tree, by predicting their classes. For each node of the tree, choose ran- domly m variables on which to base the decision at that node. Determine the best split based on these m variables in the training set. Allow each tree to be fully grown and not pruned. The class prediction of a new sample is done by pushing it down the set of trees. Each tree assigns the label of the training sample in the terminal leave it ends up in. The same procedure is iterated over all trees in the ensemble, and the majority of votes of all trees in the forest dictates the class. 4. The results of numerical experiments 4.1. Data base In the numerical experiments we have recognized 9 classes of the industrial tools. The recognized classes include: class 1 – setsquares, class 2 – magnifying glasses, class 3 – hammers, class 4 – boxes, class 5 – screwdrivers, class 6 – pliers, class 7 – drills, class 8 – discs for grinders, class 9 – grinders. The single exemplary representatives of each class are presented in Fig. 2. In the first set of experiments of the pattern recognition we have represented each class by 33 individuals, differing signifi- cantly from each other. Fig. 3 presents the exemplary set of tools Fig. 2. The examples of tools used for recognition in the numerica Please cite this article in press as: Les, T., et al. Automatic recognition of indust tions (2013), http://dx.doi.org/10.1016/j.eswa.2013.02.030 representing the class of hammers. As it is seen they differ signif- icantly by the shape, size, proportion of parameters, orientation, position in the viewing field, etc. In the second set of experiments we have checked how our automatic system is resistive to different types of noise disturbing the individual image. This time we have assumed that the system recognizes one chosen representative of each tool, represented in a different scale, various positions in the viewing field, at the pres- ence of some cracks, traces of dust of different shape, existence of blurs, etc. The number of representatives of each class was also equal 33. The exemplary set of grinder’s discs used in experiments are presented in Fig. 4. 4.2. The results of the first set of experiments The aim of the first experiments was to check how the devel- oped system is resistive to the changes of the shape of the tools, such as that presented in Fig. 3 for hammers. In the first stage of preprocessing, after extraction of the shapes of each tool, the most important was the generation and selection of the diagnostic fea- tures defined in Section 2.3. We have checked the discriminative strength of each descriptor for all combinations of classes (36 pairs of 2-class problems). For each pair of classes the value of Fisher dis- criminant coefficient has been calculated and then averaged over all pairs. The cumulative diagnostic value of the descriptor fk (k = 1, 2, . . . , 11) is defined as SðfkÞ¼ 1 36 X11 i;j¼1 i–j SijðfkÞ ð14Þ Fig. 5 presents the distribution of the values of S(fk) for the succeed- ing features, averaged over all combinations of classes.As it is seen the highest cumulative value represents shredding factor, which is over two times better than the next one (the convex compactness factor). The least discriminative values are associated with all four ratios of box counting dimensions with respect to the geometrical l experiments (each tool is represented by a single example). rial tools using artificial intelligence approach. Expert Systems with Applica- http://dx.doi.org/10.1016/j.eswa.2013.02.030 Fig. 3. The set of 33 images of hammer used in the experiments. T. Les et al. / Expert Systems with Applications xxx (2013) xxx–xxx 5 parameters. However the box counting dimension B itself is reason- ably good. To check the quality of such values of the discriminative mea- sures of the diagnostic features we have mapped the data into a 2-dimensional system represented by two best descriptors. Fig. 6 presents the distribution of data in two dimensional space formed by Fr and Fch (two best descriptors). We can see the interesting distribution of the data points. The representatives of classes are placed in different regions of the plane, usually in the compact sets. However, some classes interlace with the other, making their recognition impossible. It is evident, that the recognition of classes using only two diagnostic features is impossible. We have to use richer set of descriptors. To get their optimal quantity we have made introductory experiments of rec- ognition at application of varying number of the features arranged in the decreasing order of their discriminative ability. As the clas- sifier we have applied the Gaussian kernel SVM working in one against one mode. Two third of the data has been used in learning and the rest (1/3 of data) in testing. As a results of these experi- Please cite this article in press as: Les, T., et al. Automatic recognition of indust tions (2013), http://dx.doi.org/10.1016/j.eswa.2013.02.030 ments we have found the optimal number of descriptors is equal 5 (Fr, Fch, Fh, Fce and B). Next we have performed the 10 fold cross validation experi- ments of recognition using these 5 diagnostic features as the input information for the Gaussian kernel SVM operating in one against one mode. The data was divided into 10 parts. Nine parts were used in learning and the last one in testing. In each experiment the testing part was exchanged with one of the learning parts. The results of testing were averaged. The mean value of the relative testing error was equal 8.42%. Table 1 depicts the confusion matrix of the recognition, presented in the form of recognition results con- cerning the testing data not taking part in learning in all 10 cross validation experiments. The same experiments have been repeated using the random forest as the classifier. Two third of data points were used in learn- ing and 1/3 of testing. The best results of recognition have been ob- tained by assuming m = 3 variables used to determine the decision at a node. The best results of recognition obtained at application of 30 decision trees in the forest was this time worse than that at rial tools using artificial intelligence approach. Expert Systems with Applica- http://dx.doi.org/10.1016/j.eswa.2013.02.030 Fig. 6. The distribution of data points in the coordinate system represented by two most important features. Fig. 4. The exemplary set of noisy images of discs of grinder used in the second type of experiments. Fig. 5. The averaged cumulative values of the Fisher discriminant measure for the succeeding features in the first set of experiments. 6 T. Les et al. / Expert Systems with Applications xxx (2013) xxx–xxx Please cite this article in press as: Les, T., et al. Automatic recognition of industrial tools using artificial intelligence approach. Expert Systems with Applica- tions (2013), http://dx.doi.org/10.1016/j.eswa.2013.02.030 http://dx.doi.org/10.1016/j.eswa.2013.02.030 Table 1 The confusion matrix of recognition of 9 classes of industrial tools. Classes 1 2 3 4 5 6 7 8 9 1 (Setsquares) 33 0 0 0 0 0 0 0 0 2 (Magnifying glasses) 0 29 0 0 0 1 3 0 0 3 (Hammers) 0 2 30 0 0 1 0 0 0 4 (Boxes) 0 0 0 29 0 0 0 4 0 5 (Screwdrivers) 0 1 0 0 31 1 0 0 0 6 (Pliers) 0 1 1 0 0 30 0 0 1 7 (Drills) 1 2 1 0 0 1 28 0 0 8 (Discs for grinders) 0 0 0 3 0 0 0 30 0 9 (Grinders) 0 0 0 0 0 1 0 0 32 T. Les et al. / Expert Systems with Applications xxx (2013) xxx–xxx 7 application of SVM. The mean value of the relative testing error was equal 10.46%. Fig. 8. The distribution of data points in the coordinate system represented by (a) two most important features: shredding and corrugation factors and (b) two least significant descriptors. 4.3. The results of the second set of experiments The second set of experiments aimed on checking the sensitiv- ity of the developed system on the existence of the noise and other disturbances existing in the image of the tools. This time the origi- nal shape of the tool was unique and the disturbances have been introduced artificially by us. The exemplary set of noisy images of discs of grinder used in the second type of experiments was shown in Fig. 4. The other tools have been also disturbed in a sim- ilar fashion. Similarly to the previous experiments for each pair of classes we have calculated the value of Fisher discriminant coefficients and then averaged over all pairs. The cumulative diagnostic value of the succeeding descriptors arranged in the decreasing order are presented in Fig. 7. Once again the highest cumulative value repre- sents shredding factor, which is more than three times higher than the next one (the compactness factor). The least discriminative val- ues are associated with the box counting dimension and its ratio to the geometrical parameters. To check the discriminative quality of two best descriptors we have presented the distribution of classes in a 2-dimensional system (Fig. 8(a)). As it is seen most of the rep- resentatives of the classes form the compact clusters placed very close to each other. Only single samples of different classes are interlaced with each other (one sample of class 9 placed close t class 2, the representatives of class 1 very close to class 4). To check how reliable is application of Fisher measure of discrimination ability of descriptors we have compared the data distribution at application of only two worst features (B and Fba). It is shown in Fig. 8(b). As it is seen this time the representatives Fig. 7. The averaged cumulative values of Fisher discriminant for the succeeding features in the second set of experiments. Please cite this article in press as: Les, T., et al. Automatic recognition of indust tions (2013), http://dx.doi.org/10.1016/j.eswa.2013.02.030 of all classes are interlaced with each other making the recognition of classes very hard. The experiments of recognition have been made by applying the Gaussian kernel SVM working in one against one mode and ran- dom forest of decision trees at m = 4, found to be the best. Te experiments have been conducted in the same way as in the first case (two third of the data used in learning and the rest in testing, experiments arranged in a 10-fold cross validation). The best re- sults have been obtained at application of the first four descriptors: Fr, Fce, Fh, Fch used as the diagnostic features. The averaged results of testing obtained at application of SVM and random forest have been compared. This time the mean value of the relative testing error was much smaller and equal 1.12% for SVM and only 0.67% for random forest. Table 2 shows the Table 2 The confusion matrix of recognition of 9 classes of industrial tools in the second set of experiments. Classes 1 2 3 4 5 6 7 8 9 1 (Setsquares) 33 0 0 0 0 0 0 0 0 2 (Magnifying glasses) 0 33 0 0 0 0 0 0 0 3 (Hammers) 0 0 32 0 0 1 0 0 0 4 (Boxes) 0 0 0 33 0 0 0 0 0 5 (Screwdrivers) 0 0 0 0 33 0 0 0 0 6 (Pliers) 0 0 0 0 0 33 0 0 0 7 (Drills) 0 0 0 0 0 0 33 0 0 8 (Discs for grinders) 0 0 0 0 0 0 0 33 0 9 (Grinders) 0 1 0 0 0 0 0 0 32 rial tools using artificial intelligence approach. Expert Systems with Applica- http://dx.doi.org/10.1016/j.eswa.2013.02.030 8 T. Les et al. / Expert Systems with Applications xxx (2013) xxx–xxx confusion matrix of the pattern recognition at using the random forest as the classifier system. They are presented in the form of recognition results of testing data not taking part in learning in all 10 cross validation experiments. As expected we have got much smaller error of recognition, since the shape of each tool was this time known in advance and the only differences among the samples have been caused by the effect of noise, artificially introduced to the original images. 5. Conclusions The paper has presented the automatic approach to the recogni- tion of the industrial tools by using the methods based on the arti- ficial intelligence and advanced image processing. The boundary contour information was chosen as an effective method of repre- senting the industrial components under recognition. The starting point of the recognition procedure is the camera image of the tools. The succeeding stages of developed procedure included: the automatic localization of the tools in the image, pre- processing of the image such as the binarization, noise removal, filling the holes, and normalization, generation of the numerical descriptors characterizing the shape of the tools, verification of descriptors in the role of the diagnostic features, selection of the features and final stage of classification of the tools. Two different solutions of the final classifier system have been tried. One was based on Support Vector Machine and the second applied the ran- dom forest of the decision trees. Both belong to the most efficient automatic classification systems. The efficiency of the developed system has been verified on the example of 9 exemplary sets of the industrial tools. The tools Please cite this article in press as: Les, T., et al. Automatic recognition of indust tions (2013), http://dx.doi.org/10.1016/j.eswa.2013.02.030 differing significantly with a shape, size and location have been considered first in the recognition process. The numerical results of this verification have been presented and discussed in the paper. The next experimental set assumed recognition of the known in advance shape of the tools under the noisy environment. Experi- mental results show the good performance of the proposed system both in a noise-free and the noisy environment. References Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32. Cormen, T. H., Leiserson, C. E., Rivest, R., & Clifford, C. S. (2009). Introduction to algorithms. Boston: MIT Press. Duda, R. O., Hart, P. E., & Stork, P. (2003). Pattern classification and scene analysis. New York: Wiley. Gonzalez, R. C, & Woods, R. E (2011). Digital image processing. Upper Saddle River: Prentice Hall. Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature selection. Journal of Machine Learning Research, 3, 1157–1182. Kim, H., & Nam, K. (1995). Object recognition of one-DOF to tools by back- propagation neural net. IEEE Transaction on Neural Networks, 6, 484–487. Kruk, M., Osowski, S., & Koktysz, R. (2007). Numerical characterization of the images of prostate cancer for recognition of Gleason scale. Przeglad Elektrotechniczny, 83, 227–230. Matlab user manual with toolboxes (2012). Natick: MathWorks. Nabhani, F., & Shaw, T. (2002). Performance analysis and optimization of shape recognition and classification using ANN. Robotics and Computer-Integrated Manufacturing, 18, 177–185. Otsu, N. (1979). A threshold selection method from grey-level histograms. IEEE Transaction on Systems, Man, and Cybernatics, 9, 62–66. Schölkopf, B., & Smola, A. (2002). Learning with kernels. Cambridge MA: MIT Press. Schroeder, M. (2006). Fractals, Chaos, power laws. New York: W.H. Freeman and Company. Soille, P. (2003). Morphological image analysis, principles and applications. Berlin: Springer. Vapnik, V. (1998). Statistical learning theory. New York: Wiley. rial tools using artificial intelligence approach. Expert Systems with Applica- http://dx.doi.org/10.1016/j.eswa.2013.02.030 Automatic recognition of industrial tools using artificial intelligence approach 1 Introduction 2 The preprocessing stage 2.1 Binarization of the image 2.2 Localization of an object 2.3 Definition of numerical descriptors of the image 2.4 Selection of the features 3 Classification systems 4 The results of numerical experiments 4.1 Data base 4.2 The results of the first set of experiments 4.3 The results of the second set of experiments 5 Conclusions References 
work_fs5diuqqzjehbijmzgwngicnyq	----	[22]2010-109(김상훈,정상용).hwp 149조명․전기설비학회논문지 제24권 제10호, 2010년 10월 Genetic Algorithm과 Expert System의 결합 알고리즘을 이용한 직구동형 풍력발전기 최적설계 (Optimal Design of Direct-Driven Wind Generator Using Genetic Algorithm Combined with Expert System) 김상훈 * ․정상용 ** (Shang-Hoon Kim․Sang-Yong Jung) Abstract In this paper, the optimal design of a wind generator, implemented with the hybridized GA(Genetic Algorithm) and ES(Expert System), has been performed to maximize the AEP(Annual Energy Production) over the whole wind speed characterized by the statistical model of wind speed distribution. In particular, to solve the problem of calculation iterate, ES finds the superior individual and apply to initial generation of GA and it makes reduction of search domain. Meanwhile, for effective searching in reduced search domain, it propose Intelligent GA algorithm. Also, it shows the results of optimized model 500[kW] wind generator using hybridized algorithm and benchmark result of compare with GA. Key Words：Direct-Driven Wind Generator, Optimization, GA(Genetic Algorithm), ES(Expert System), CBR(Case Based Reasoning), AEP(Annual Energy Production), Mutation Rate * 주저자：동아대학원 전기공학과 석사과정 ** 교신저자：동아대학교 전기공학과 조교수 Tel：051-200-6945, Fax：051-200-6947 E-mail：whiteland1@nate.com 접수일자：2010년 8월 26일 1차심사：2010년 8월 31일 심사완료：2010년 10월 1일 1. 서 론 최근 환경오염 및 지하자원의 고갈로 인해 에너지 잔량이 앞으로 얼마 남지 않은 추세로 자연환경에서 얻을수있는친환경에너지와대체에너지원을개발 하여신재생에너지가각광받고있는추세이다. 특히 풍력발전과 태양 발전의 경우 에너지원이 무한하고 영구적이며 환경오염을 유발시키지 않아 큰 관심을 받고 있는 추세이다. 풍력발전기는크게기어타입(Geared Type)과기어 리스타입(GearlessType)으로나눌수있다. 풍력발전 은발전을위한토크가터빈의크기에따라비례하게 되므로 효율적인 발전을 위해서는 저속의 고토크 발 전기사양을요구하게된다. 따라서이러한저속고토 크의 입력을 증속기를 통하여 적정크기의 발전기로 입력함으로서원하는출력을얻을수있다. 하지만증 속기의설치로인한발전기나셀의중량증가, 증속기 내부의기어에의한소음및진동발생,정기적인증속 Journal of the Korean Institute of IIIuminating and Electrical Installation Engineers (2010) 24(10)：149～156 24-10-22논문 DOI : 10.5207/JIEIE.2010.24.10.149 150 Genetic Algorithm과 Expert System의 결합 알고리즘을 이용한 직구동형 풍력발전기 최적설계 Journal of KIIEE, Vol.24, No.10, October 2010 기의 유지보수 등의 단점이 발생하게 된다[1]. 기어리스 타입의 직구동형 풍력발전기는 블레이드 와 발전기사이 증속기를 제거하고 동일 축으로 연결 하여 동기속도로 운전한다. 증속기를 제거함으로서 중량및노이즈, 유지보수비용이감소되며토크밀도 가매우높은장점이있다. 특히표면부착형영구자석 동기발전기(SPMSG)는 저속에서 높은 토크분포와 효율을 보이게 되는데 이는 저속으로 운전하는 직구 동형 풍력발전기의 특성에 매우 적합한 특성을 가진 다[2]. 한편근대의전기기기설계법을보게되면전문가의 경험이나 과거의 사례를 통하여 그 모델을 근본으로 하여 사용자의 요구사양에 맞추어 설계를 수행하는 것이보통이었다. 그러나최근의전기기기설계의경 향을 살펴보면 경험에 의존하던 설계를 최적화 기법 을 적용하여 개발하려는 추세이며 이를 위한 최적화 기법또한급진적으로발전하고관련논문도많이제 출되고 있다. 최적화 기법은 크게 확률론적 방법(Stochastic Method)과 결정론적 방법(Deterministic Method)로 나눌수 있다. 확률론적 방법은 여러 가지 경우에 대 해 확률적으로 탐색해가는 방법으로 모든 경우의 수 를고려한다는점에서그해의신뢰성이높은편이나, 그러한해를얻기위해서는매우많은지점에서의해 석을필요로하게되어최적화수행시간이매우오래 걸린다는 단점이 있다. 결정론적방법은어떠한지점에서주위를탐색하여 개선된값이있다면그방향을선택하여최적해탐색 지점을이동해가는방식으로최적화수행시간이매우 빠른장점이있다. 하지만전범위를탐색하기보다는 주변을탐색하는방법이므로국부해로수렴될가능성 이 높고 해의 신뢰성 또한 떨어지는 편이다. 본 논문에서는 연간 에너지 생산량을 최대화 하는 설계를 수행하기 위해 최적화 알고리즘과 유한요소 해석법을 연계하여 전 풍속영역에 대해 최적화를 수 행하였다. 한편신뢰성높은결과를얻기위해확률론 적방법인GA를사용하였다. 하지만앞서언급하였듯 이 확률론적 방법과 유한요소 해석법을 통해 신뢰도 높은최적화결과를얻기위해서는매우많은연산시 간을소요하게된다. 이를해결하기위해유전알고리 즘(GA)에 앞서 언급하였던 전문가의 경험이나 예전 사례를 기본으로 하여 변형설계를 수행하는 전문가 시스템(ES)을결합하여정확도를유지하면서빠른수 렴에도달하는혼합알고리즘을제시한다. 특히사례 기반의 전문가 시스템은 데이터베이스로부터 유사도 분석을 통하여 목적함수와 유사한 후보를 선출한 다 음이로부터GA의설계변수탐색범위를축소시켜최 적값을 빠르게 탐색하는 알고리즘을 개발하였다. 2. 풍력발전기의 특성 단위시간에대한공기의부피변화 ·와질 량변화 ··를 고려한 순수 풍력량 는 다음과 같다.         (1) 여기서  : 공기밀도[㎏/m3], A : 블레이드통과면적 [m 2 ] 풍력발전기로들어온에너지는터빈의요우, 피치제 어를통해최대효율을갖도록제어된입력이들어오 게된다. 풍력터빈의출력(Pi)은블레이드의피치컨 트롤을 거쳐 그림 1과 같이 주속비와 출력 계수 ()를각각고려한수식 (1)의풍력에너지(Pw)가 발전기입력(Ps)으로 전달되게 된다[3].     (2) 풍력 터빈에 의한 입력(Ps)은 발전기 입력 Pi와 같 으므로발전기출력은식 (3)과(4)로계산할수있다.    · (3)     (4) 여기서    : 주속 비, :터빈 속도[rad/s], R: 블레이드 직경, p: 극수, : 영구자속에 의한 상당 쇄교자속 151 김상훈․정상용 조명․전기설비학회논문지 제24권 제10호, 2010년 10월 이렇게계산된발전기입력전류를통하여원하는파 라미터들을 유한요소 해석을 통해 계산해 낸다. 그림 1. 풍력 터빈의 능력곡선 Fig. 1. Characteristic of Wind turbine 3. 풍속의 확률 분포에 따른 AEP 계산 풍력은발전기설치지점의지리적요건, 날씨, 계절, 주변요인에 의해 크게 변하게 되므로 실제로 풍력의 가변성을 예측하여 발전기의 출력을 계산하는 것은 매우어렵다. 개략적으로나마운전풍속영역을고려하 는방법으로세가지정도의방법이있다. Grauers[4] 는 정격풍속에서 발전기의 손실을 기준으로 각 풍속 에서의 손실 비례계수를 구할 때 풍속분포를 고려하 는 방법을 사용하고 있으나 계산법이 복잡하여 실용 적이지 못하다. Inoue[5]은 풍속 확률분포 함수로서 Weibull function을사용하고있으나, 이경우에는발 전기가 설치되는 지역의 풍속분포에 대한 자세한 데 이터(shape factor와 scale factor 등)가 필요로 하기 때문에 비교적 신뢰성이 높으나 적용범위가 매우 좁 은 편이다. 본 논문에서는 운전풍속영역을 고려하는 방법으로발전기가설치되는지역의평균풍속에대한 데이터만으로도적용이가능한Rayleigh함수를풍속 확률밀도 함수로 사용하여 풍력발전기의 연간에너지 생산량(Annual EnergyProduction, AEP)을산정하였 다. 풍속의 확률분포 식은 다음과 같다.                (5) : 지역내평균풍속, : 풍속v에따른풍속확 률 평균풍속에 따른 확률 분포는 그림 2와 같다. 그림 2. 풍속의 확률분포 Fig. 2. Destribution of wind speed 최적화목표는풍속의평균분포를통하여연간에 너지생산량의최대화를목적함수로두고최적설계를 하였다. 일년간특정풍속이부는유효시간H(v)는다 음과 같이 계산할 수 있다.  ··∆ (6) : 일년의총시간(약8760h), ∆ : 풍속한구간 의 크기 특정 풍속의 유효시간이 정해지면, 각 풍속별 전력 총량이계산되며, 궁극적으로연간운전시간을곱하 여다음과같은연간총에너지생산량을구하게된다.       [Wh] (7) 여기서  ×  4. GA와 ES를 결합한 최적화 알고리즘 4.1 유전알고리즘(Genetic Algorithm) 유전 알고리즘은 자연계의 생명체 중 환경에 잘 152 Genetic Algorithm과 Expert System의 결합 알고리즘을 이용한 직구동형 풍력발전기 최적설계 Journal of KIIEE, Vol.24, No.10, October 2010 적응한개체가좀더많은자손을남길수있다는자 연선택과정과자연계의생명체의설계도와같은유 전자의변화를통해서좋은방향으로발전해나간다 는 자연 진화의 과정을 모방하여 컴퓨터로 모의 수 행을 하는 최적화 알고리즘의 하나이다. 즉 실세계 의문제를풀기위해잠재적인해들을컴퓨터상에서 코딩된 개체로 나타내고, 여러 개의 개체들을 모아 개체군을 형성한 뒤, 세대를 거듭하면서 이들의 유 전 정보를 서로 교환하거나 새로운 유전 정보를 부 여하면서적자생존의법칙에따라모의진화를시킴 으로써, 주어진 문제에 대한 최적의 해를 찾는 계산 모델이다[6]. 이러한방법의장점은목적함수의미분과정이나특 별한수학적연산을필요로하지않는다는것이다. 또 한, 점에의한탐색이아니라개체들이모여이루는개 체군에 의한 병렬적인 탐색이라는 점에서 기존의 최 적화알고리즘과는차이가있다. 한세대의개체군에 속한개체들은진화를거듭하면서, 이전세대까지축 적된 정보를 서로 교환하고 새로운 영역으로의 탐색 을시도한다. 탐색의방향이나영역이초기값에의해 서 결정되지 않고 세대마다 확률적으로 결정되므로 지역최소점에빠질가능성이적어전역최적화가가 능한 알고리즘이다[7]. 4.2 전문가 시스템(Expert System) 전문가 시스템(ES)은 A.I의 한 종류로서 문제해결 에 있어서 전문가들이 작성해 놓은 지식이나 규칙을 적용하는 방법이다. 지식베이스에는 크게 사례 기반 과규칙기반두가지종류가있다. 규칙기반은과거 의 경험을 통하여 특정 상황에서 유의사항이나 절대 적진리등을규칙으로정하여 “만약..라면 ..하는” 방 법으로 해답에 접근해가는 방식이다. 또 한가지 사 례기반추론기법(CBR)은 한마디로 어떤 문제를 해결 하기 위해 과거에 사용했던 구체적인 경험을 바탕으 로 새로운 문제를 해결하는 방법이라고 할 수 있다. 사례기반추론을 이용한 방법은 규칙기반보다는 단순 하며, 특히문제영역이잘정형화되지않는분야에서 는 좋은 접근법이라 할 수 있다[8]. 그림 3. 사례기반 시스템의 개념도 Fig. 3. Conceptual Diagram of CBR 4.3 GA와 ES의 결합 알고리즘 GA는 세대기반의 확률론적 전역 탐색법으로서 일 반적으로최적화에사용되는수학적인미분계산이필 요하지 않아 활용 범위가 매우 광범위 하다. 그러나 이러한방대한활용성및전역탐색기법의단점이그 렇듯이 넓은 탐색범위를 조건에 맞춰 검색해야 하는 부담때문에소요시간이매우오래걸리는것이단점 이다. 여기에사례기반의엔진을결합하게되면과거 설계사례 및 경험에서 우수한 인자들을 선별하여 각 인자들의 변수범위를 분석하여 탐색범위를 제한하게 되고그범위내에서초기세대를생성하게되기때문 에초기세대의해석값이기존의탐색에의한초기세 대보다월등히뛰어난성능을보이게된다. 또한변이 율의범위또한크게범위를벗어나지않고최적화된 범위내에서탐색을수행하기때문에빠른속도로최 적값에수렴할수있을것이다. 그림4은결합알고리 즘의 개념을 나타내고 있다. 혼합알고리즘의최적화순서를살펴보게되면최초 로 사용자가 원하는 설계사양을 입력하게 되면 혼합 알고리즘은과거설계사례들과목표설계사양간의유 사도를분석을수행한다. 유사도분석을수행함에있 어 설계 사양과 값이 가까울수록 파라미터에 가중치 를 두어 차후 임의의 파라미터를 선출하여 설계변수 의범위지정시좀더높은확률로관여할수있도록 가중치설정한다. 유사도분석을통하여설계의기초 153 김상훈․정상용 조명․전기설비학회논문지 제24권 제10호, 2010년 10월 가 될 다수의 유사모델을 선출하여, 두 번째선별의 과정을 거쳐 사용자가 지정한 개수의 후보를 선출한 다. 선출된각후보들로부터설계변수와동일한파라 미터들을추출해낸다. 추출된파라미터들은이전단계 의가중치판단에의해값들의보정과정을통하여사 용자가원하는세부정밀도를만족하는변수범위를지 정하게된다. 또한보정된값을사용하여GA의초기 세대를생성한다. 검증을위해Test Function에적용 하여그생성된값의해석을수행한결과, 최적값에비 슷한 결과가 생성되었다. 그림 4. ES+GA의 개념도 Fig. 4. Conceptual Diagram of ES+GA 이와같이최적화된변수범위와초기세대를사용한 최적화과정은그수렴속도가매우빠를것이며성능 또한 기준 이상의 능력을 보여줄 수 있을 것이다. 4.4 지능형 유전알고리즘(Intelligent GA) 기존의GA는초기에설정한횟수만큼세대의재생 산과 계산을 반복하게 되므로 탐색범위가 좁아져 최 적값에빠르게수렴했다하더라도정해진계산횟수만 큼반복을해야알고리즘의수행을마치게된다. 특히 유한요소 해석법은 비용함수계산시간이 오래 걸리게 되므로 똑같은 파라미터의 반복된 계산은 결과에 영 향을 미치지 않을 뿐만 아니라 탐색시간도 늘어나게 되는단점이있다. 따라서본논문에서는재생산된개 체의 계산에 앞서 이전에 계산된 개체의 파라미터를 비교하여 똑같은 값을 가질 경우 비용함수 계산절차 를 생략하고 이전의 계산결과로 대체하는 방법을 적 용하였다. 그림5에는Intelligent GA의흐름도와개념 을 간략히 나타내었다. 그림 5. Intelligent GA의 Flow Chart Fig. 5. Flow Chart of Experted GA Intelligent GA의요는앞서언급했듯이ES에의한 탐색범위의축소및빠른수렴효과로인해파라미터 가 완전히 똑같은 개체가 생성될 확률이 매우 높다. 이렇게같은파라미터에따라자동설계되어있는모 델을해석수행하였을경우동일한결과값을나타내 게 될 것이다. 따라서이러한중복현상을방지하기위해재생산과 정후각개체의파라미터를모델링에적용하기전개 체의 염색체를 이전 계산한 염색체 데이터뱅크와 대 조하여같은염색체를가진개체를찾는다. 만약검색 결과같은개체가있는경우데이터뱅크에저장되어 있는결과값을복사하여그개체의연산을종료한다. 즉 불필요한 유한요소 계산을 건너뛰고 다음 개체의 계산을수행하게되는것이다. 반대로검색결과같은 염색체를 가진 개체가 없다면 모델링과 유한요소 해 석을 수행한다. 혼합알고리즘과Intelligent GA의성능을검증하기 위해기존의유전알고리즘과Intelligent GA알고리즘, 혼합알고리즘3가지경우에대하여단순함수의수행, 비교한 결과를 표 1에 나타내었다. 결과에따르면기존의GA의경우0.1012[min]의시 154 Genetic Algorithm과 Expert System의 결합 알고리즘을 이용한 직구동형 풍력발전기 최적설계 Journal of KIIEE, Vol.24, No.10, October 2010 간이 걸린 반면 같은 파라미터를 건너뛴 Intelligent GA의경우0.07154분이걸려약30[%]의개선효과를 나타내었다. 또한혼합알고리즘의경우기존GA에서 반복된 파라미터의 생성횟수가 5223회에 인 것에 반 해11920회의동일파라미터생성으로범위가좁아질 수록그 현상이증가하는 현상을 볼 수 있었다. 이로 인한최적화시간또한60[%] 이상개선된효과를얻 을 수 있었다. 표 1. 단순 함수(x2+y2=0)에 대한 GA와 Intellignet GA, ES+iGA의 결과 비교 Table 1. COMPARISON THE RESULT OF 3 METHODS(GA, iGA and ES+iGA) GA iGA ES+iGA Duplicated 5223 5223 11920 Time 0.1012[min] 0.07154 0.0452[min] 5. 최적화 알고리즘의 시험함수 적용 제시한 알고리즘과 일반적인 GA의 성능을 벤치마 크하기위해Branin함수와Shubert함수에적용하여 비교분석 하였다. 시뮬레이션을 수행한 컴퓨터 성능 (2.7[GHz], 3.25G Memory) 때문에 수렴속도가 매우 빨라비용함수계산시시간지연(0.1[μs])을주어결과 데이터 변화를 관찰할 수 있게 하였다. 그림 6. Branin 함수 Fig. 6. Branin Function 그림 7. Shubert 함수 Fig. 7. Shubert Function 표 2. Branin 함수에 대한 GA와 ES+iGA의 결과 비교 Table 2. COMPARISON RESULTS OF BRANIN FUNCTION OF GA AND ES+iGA GA ES+iGA Global Minimum Cost 0.39791 0.39789 Time 0.1052[min] 0.0411[min] 표 3. Shubert 함수에 대한 GA와 ES+iGA의 결과 비교 Table 3. COMPARISON RESULTS OF BRANIN FUNCTION OF GA AND ES+iGA GA ES+iGA Global Minimum Cost -180.1840 -186.7260 Time 0.1068[min] 0.0471[min] 위의결과에서나타내듯이두시험함수모두제안된 알고리즘의 경우가 더 높은 신뢰성과 빠른 수렴속도 를보임을알수있다. 특히다봉성함수인Shubert함 수의경우는알려진최적값에더가까운값을제시하 여 신뢰성 있는 결과를 제시하였다. 6. 풍력발전기 최적설계 제안한최적화알고리즘을통하여500[kW]급직구 동형풍력발전기의설계를수행하였다. 설계목표사 양은 다음과 같다 155 김상훈․정상용 조명․전기설비학회논문지 제24권 제10호, 2010년 10월 표 4. 풍력발전기 설계 사양 Table 4. SPECIFICATION OF WIND POWER GENERATOR 정격 출력(Ps) 500[kW] 풍속 사양 Cut-in 풍속 3.5[m/s] 정격 풍속 13.5[m/s] Cut-out 풍속 26[m/s] 발전기 형식 SPMSG 터빈 직경 39[m] 운전 속도 0～32[rpm] 블레이드 회전 면적 1207[m2] 제어방법 Pitch control 그림 8은 설계변수를 나타내고 있으며, 자극각도 (), 고정자치폭(), 회전자요크두께(), 고정자 슬롯높이() 등을 설계변수로 선정하였다. 그림 8. 표면부착형 영구자석 동기기 설계변수 Fig. 8. Design variable of SPMSG 표 5. 후보군의 설계치수 Table 5. Variables of candidates 용량 ([kW]) X1 X2 X3 X4 직경 ([m]) 극/슬롯 740 2.77 10.55 22.470 70.35 42 100/300 500(1) 2.027 8.940 21.098 63.775 39 100/300 500(2) 2.5 12.5 13 58 39 100/300 450 2.272 10.412 10.153 53.25 36 100/300 설계목표를ES에입력하여데이터베이스에서유사 도검색을통하여4개의 후보를선출하였다. 선정된 용량은 각각 740[kW](1개 후보), 500[kW](2개 후보), 450[kW](1개후보)를베이스로하여설계변수를추출, Intelligent GA에적용하였다. 각후보군에대한설계 변수를 표 5에 정리하였다. 그림 9에서 나타내듯이 최적화의 수렴과정을 살펴 보면 제안된 알고리즘의 경우 최적점에 매우 가까운 지점에서 시작되는 것을 알 수 있다. 그림 9. GA와 ES+GA의 수렴과정 Fig. 9. Convergence of GA and ES+iGA 표 6. 설계 결과 Table 6. RESULT OF OPTIMIZATION Symbol GA ES+GA Designed Model Generator 1233.0512[MWh] 1234.5222[MWh] Convergence Time 58.8152[H] (3528.91[min]) 46.793[H] (2807.57[min]) Parameter X1 2.0277 1.817 X2 8.9407 8.883 X3 21.098 16.428 X4 63.7750 65.132 설계파라미터의 변화추이를 살펴보면 자석의 높이 는동일하게가져간상태에서최적설계결과폭이줄 었기 때문에 자석량이 줄어들어 제작 비용이 감소하 게되었다. 또한D 2 L을고려하지않은상태에서AEP 의 최대화를 목적함수로 하였기 때문에 치높이에 의 한모터의직경이늘어나는양상을보였다. 또한회전 자 요크는 대폭 줄어드는 경향을 나타내었다. 설계 결과를 살펴보면 표 6에서 나타나듯이 GA에 비해혼합알고리즘이발전량은1.5[MWh] 정도상승 156 Genetic Algorithm과 Expert System의 결합 알고리즘을 이용한 직구동형 풍력발전기 최적설계 Journal of KIIEE, Vol.24, No.10, October 2010 하였고 수행시간은 20[%] 정도 감소하였다. 이는 곧 최대연간에너지생산량을위한ES를결합한GA기 반의 풍력발전기 최적설계는 빠른 수렴성과 개선된 결과를 나타내고 있음을 알 수 있다. 7. 결 론 본논문에서는풍력발전기의연간평균풍속을고려 하여 풍속별 확률분포를 사용하여 연간 에너지 생산 량을최대로하는유한요소해석을통한설계법을제 시하였다. 또한 막대한 시간이 걸리는 GA의 최적화 알고리즘과전문가시스템을결합하여신뢰성을유지 하면서 빠른 수렴을 해내는 새로운 알고리즘을 개발 하였다. 이상의방법을사용하여500[kW]급풍력발전 기의 연간 최대 에너지 생산모델을 목적함수로 하는 최적설계를 수행한 결과, 기존의 GA와 비교하여 20[%]정도개선된수렴속도를보여주었고, 목적함수 또한 1.5[MWh]정도 개선된 결과를 나타내었다. 본논문에서제안한혼합형알고리즘은수학적인정 보를요구하지 않기때문에의료, 항공, 화학, 신소재 등의 추상적인 학문분야에서도 널리 사용될 수 있을 것이라 확신한다. 향후 규칙기반 시스템을 적용하여 목표 사양에 부합될 뿐만 아니라 기본적인 요구사양 도 만족시키는 지식기반 최적화 알고리즘을 개발할 것이다. 본 연구는 교육과학기술부와 한국산업기술진흥원의 지역 혁신인력양성사업과 2009학년도 동의대학교 교내연구비에 의해 연구되었음(과제번호2009AA161). References [1] MADS를 이용한 직접구동형 풍력발전기 최적설계, 박지 성(Ji-Seong Park), 한국조명․전기설비학회, 조명․전기 설비학회논문지, 제23권 제12호 2009.12, pp. 48～ 57(10pages). [2] Yicheng Chen, Pragasen Pillay, Khan.A., “PM wind generator comparison of different topologies,” Proc. of 39th IAS Annual Meeting Conference, Vol. 3, No. 3-7, pp 1405-1412, October.2004. [3] J. F. Manwell, J. G. McGowan, A. L. Rogers, “Wind Energy Theory, Design and Application,” John Wiley & Sons, 1st Ed., 2002. [4] Anders Grauers, “Design of Direct-driven Permanent- magnet Generators for Wind Turbines,” Ph.D Thesis Chalmers University, October 1996. [5] Inoue, A., Hasan Ali Mohd., Takahashi, R., Murata, T., Tamura, J., Ichinose and M., Kazumasa Ide, “A Calculation Method of the Total Efficiency of Wind Generator,” Proc. of PEDS2005, Vol. 2, No. 28-01, pp. 1595-1600, Nov. 2005. [6] 유전알고리즘과 신경회로망을 이용한 선형유도전동기의 최적설계, 김창업(Chang-Eob Kim), 한국조명․전기설비학 회, 조명․전기설비학회논문지, 제17권 제5호 2003.9, pp. 29～35. [7] 문병로, “유전 알고리즘”, 2003. [8] 직류서보모터의 고효율 수요관리를 위한 전문가 시스템, 김광헌(金垙憲), 한국조명․전기설비학회, 조명․전기설 비, 제8권 제3호 1994.6, pp. 71～79(9pages). ◇ 저자소개 ◇──────────── 김상훈(金相勳) 1982년 6월 13일생. 2009년 동의대 전기 공학과 졸업. 2009년～현재 동아대학원 전기공학과 석사과정. 정상용(鄭相龍) 1973년 9월 20일생. 1997년 서울대 공대 전기공학부 졸업. 1999년 동 대학원 졸업 (석사). 2003년 동 대학원 졸업(박사). 2003～2006년 현대자동차 연구개발본부 선임연구원. 현재 동아대학교 전기공학과 조교수. 
work_fspmmk7yhrgyjk26nsl2as3t7m	----	[PDF] Application of fuzzy TOPSIS in evaluating sustainable transportation systems | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1016/j.eswa.2011.04.005 Corpus ID: 14349198Application of fuzzy TOPSIS in evaluating sustainable transportation systems @article{Awasthi2011ApplicationOF, title={Application of fuzzy TOPSIS in evaluating sustainable transportation systems}, author={A. Awasthi and S. Chauhan and H. Omrani}, journal={Expert Syst. Appl.}, year={2011}, volume={38}, pages={12270-12280} } A. Awasthi, S. Chauhan, H. Omrani Published 2011 Computer Science Expert Syst. Appl. Sustainable transportation systems are the need of modern times. [...] Key Method Fuzzy TOPSIS is used to generate aggregate scores for sustainability assessment and selection of best alternative. In step 3, sensitivity analysis is performed to determine the influence of criteria weights on the decision making process. A numerical illustration is provided to demonstrate the applicability of the approach. The strength of the proposed work is its practical applicability and the ability to generate good quality…Expand View via Publisher isiarticles.com Save to Library Create Alert Cite Launch Research Feed Share This Paper 168 CitationsHighly Influential Citations 10 Background Citations 32 Methods Citations 41 View All Topics from this paper The Quality of Life Numerical analysis Aggregate data 168 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency Selection of sustainable urban transportation alternatives using an integrated intuitionistic fuzzy Choquet integral approach G. Büyüközkan, O. Feyzioglu, Fethullah Göçer Computer Science 2018 29 Save Alert Research Feed Selection of Alternative Fuels for Sustainable Urban Transportation under Multi-criteria Intuitionistic Fuzzy Environment Sathi Mukherjee Mathematics 2017 11 Save Alert Research Feed An integrated Decision-Making Approach for Road Transport Evaluation in a Sustainable Supply Chain Mana Andarkhora, Amirhossein Azadnia, Saeid Gholizadeh, Pezhman Ghadimi Computer Science 2019 Highly Influenced PDF View 16 excerpts, cites background Save Alert Research Feed Estimation of Performance Indices for the Planning of Sustainable Transportation Systems A. Paz, Pankaj Maheshwari, P. Kachroo, Sajjad Ahmad Engineering, Computer Science Adv. Fuzzy Syst. 2013 42 PDF Save Alert Research Feed Investigating ideal-solution based multicriteria decision making techniques for sustainability evaluation of urban mobility projects A. Awasthi, H. Omrani, P. Gerber Computer Science 2018 19 Save Alert Research Feed Feasibility stage screening for sustainable energy alternatives with a fuzzy multi-criteria decision analysis protocol D. Wood Modeling Earth Systems and Environment 2021 View 1 excerpt, cites methods Save Alert Research Feed Sustainable transport fleet appraisal using a hybrid multi-objective decision making approach Chunguang Bai, B. Fahimnia, Joseph Sarkis Computer Science Ann. Oper. Res. 2017 27 Save Alert Research Feed Comparison of fuzzy-based and AHP methods in sustainability evaluation: a case of traffic pollution-reducing policies R. Rossi, Massimiliano Gastaldi, Gregorio Gecchele Economics 2013 25 PDF View 1 excerpt Save Alert Research Feed Transportation sustainability index in dairy industry – Fuzzy logic approach I. Djekic, N. Smigic, Ruzica Glavan, J. Miočinović, I. Tomašević Business 2018 15 Highly Influenced PDF View 3 excerpts, cites background and methods Save Alert Research Feed Sustainability Evaluation of Transportation Policies: A Fuzzy-Based Method in a “What to” Analysis R. Rossi, Massimiliano Gastaldi, Gregorio Gecchele Business 2014 6 PDF View 1 excerpt, cites background Save Alert Research Feed ... 1 2 3 4 5 ... References SHOWING 1-10 OF 49 REFERENCES SORT BYRelevance Most Influenced Papers Recency Sustainability Assessment at the Transportation Planning Level: Performance Measures and Indexes C. Jeon, A. Amekudzi, R. Guensler Engineering 2008 84 Save Alert Research Feed Sustainability: an ill-defined concept and its assessment using fuzzy logic Yannis A. Phillis, Luc A. Andriantiatsaholiniaina Computer Science 2001 331 Save Alert Research Feed FUZZY COMPREHENSIVE ASSESSMENT, FUZZY CLUSTERING ANALYSIS AND ITS APPLICATION FOR URBAN TRAFFIC ENVIRONMENT QUALITY EVALUATION Y. Tao, Yang Xin-miao Engineering 1998 50 Save Alert Research Feed The application of knowledge management to support the sustainable analysis of urban transportation infrastructure T. El-Diraby, B. Abdulhai, K. Pramod Engineering 2005 9 View 1 excerpt, references background Save Alert Research Feed Analysis of reservoir water quality using fuzzy synthetic evaluation Ruei-Shan Lu, S. Lo, J.-Y. Hu Computer Science 1999 151 Save Alert Research Feed MULTI-CRITERIA APPROACH FOR THE SELECTION OF ALTERNATIVE OPTIONS FOR ENVIRONMENTALLY SUSTAINABLE TRANSPORT SYSTEM IN DELHI S. Yedla, R. Shrestha Engineering 2003 206 View 1 excerpt, references methods Save Alert Research Feed A problem-structuring model for analyzing transportation-environment relationships F. Ülengin, Özgür Kabak, Sule Önsel, Burç Ülengin, E. Aktas Computer Science Eur. J. Oper. Res. 2010 50 PDF Save Alert Research Feed EFECT – evaluation framework of environmental impacts and costs of transport initiatives D. Tsamboulas, G. Mikroudis Engineering 2000 70 View 1 excerpt, references background Save Alert Research Feed Sustainable Urban transportation: Performance indicators and some analytical approaches J. Black, Antonio Páez, P. Suthanaya 2002 214 PDF Save Alert Research Feed Evaluating Urban Sustainability Using Land-Use Transport Interaction Models K. Spiekermann, M. Wegener Economics 2004 25 PDF View 1 excerpt, references background Save Alert Research Feed ... 1 2 3 4 5 ... Related Papers Abstract Topics 168 Citations 49 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_ftxkkedfivdpdbnwya6qnvm66m	----	Symbiotic filtering for spam email detection Clotilde Lopes a Paulo Cortez a,∗ Pedro Sousa b Miguel Rocha b Miguel Rio c aDep. of Information Systems/Algoritmi, University of Minho, 4800-058 Guimarães, Portugal bDep. of Informatics/CCTC, University of Minho, 4710-059 Braga, Portugal cDep. of Electrical Engineering, University College London, WC1E 7JE UK Abstract This paper presents a novel spam filtering technique called Symbiotic Filtering (SF) that aggregates distinct local filters from several users to improve the overall perfor- mance of spam detection. SF is an hybrid approach combining some features from both Collaborative (CF) and Content-Based Filtering (CBF). It allows for the use of social networks to personalize and tailor the set of filters that serve as input to the filtering. A comparison is performed against the commonly used Naive Bayes CBF algorithm. Several experiments were held with the well-known Enron data, under both fixed and incremental symbiotic groups. We show that our system is competitive in performance and is robust against both dictionary and focused con- tamination attacks. Moreover, it can be implemented and deployed with few effort and low communication costs, while assuring privacy. Key words: Anti-spam filtering, Naive bayes, Collaborative filtering, Content-based filtering, Word attacks 1 Introduction Unsolicited bulk email, widely known as spam, has become a serious prob- lem for network administrators and for Internet users in general. According to MAAWG (2009), it accounts for 89% to 92% of all email messages sent and it consumes resources (i.e. time spent reading messages, bandwidth, CPU and disk) that are far away from negligible. Spam is also an intrusion of privacy ∗ Corresponding author. E-mail pcortez@dsi.uminho.pt; tel.: +351 253510313; fax: +351 253510300. Preprint submitted to Elsevier 27 November 2009 and used to spread malicious content (e.g. online fraud or viruses). The cost of sending these emails is, however, very close to zero. Internet connections are cheap and with the advent of botnets (e.g. Bobax worms), criminal orga- nizations have access to potentially millions of infected computers and thus send emails from what has to be regarded as legitimate users (Ramachandran and Feamster, 2006; Kanich et al., 2008). In the last years, the most popular anti-spam solutions have been based in Content-Based Filtering (CBF) (in particular, Bayesian filtering) (Garriss et al., 2006; Guzella and Caminhas, 2009). These class of algorithms use mes- sage features (e.g. word frequencies) for statistically discriminating email into legitimate (ham) and illegitimate (spam) messages. However, CBF presents some drawbacks. Often, there is a large gap between the high-level concept (e.g. spam image) and the low-level message features (e.g. bit colors). Also, CBF tends to give weak performances for new users, as it requires a large number of representative examples. Moreover, spammers can mix spam with normal words (often not visible to the final user), in what is known as dictio- nary or focused attacks (Nelson et al., 2008). When users flag these messages as spam, their training set is contaminated and the CBF performance is heavily reduced. As an alternative, Collaborative Filtering (CF) is a distinct anti- spam strategy, where information (e.g. IP or message fingerprints) is shared about spam messages (Zhong et al., 2008). Yet, pure CF often suffers from several issues (e.g. first-rater, sparsity of data and privacy, see Section 2). To solve the CBF drawbacks, we propose a novel distributed approach, termed Symbiotic Filtering (SF), that combines features from CBF and CF. Symbiosis is a close interaction among different entities and this phenomenon is present not only in biological species but also in business enterprises (Alocilja, 2000). We prefer the term symbiotic rather than collaborative or cooperative, since in this case each individual may have a distinct goal, as spam by definition is a personal concept. The idea is to promote a cooperation among distinct entities (e.g. email users) interested on personalized filtering (e.g. spam detection). Rather than exchanging messages, these entities will share information about what each local filter has learned (e.g. Bayesian model). The aim of SF is to foster mutual relationships, where all or most members benefit. Under SF, a given user is interested in improving filtering at a personal level. The Internet is used to gather collaborators among these (high number of) users. High group dynamics are expected, as members may join or leave the collaboration, and also there are privacy issues regarding what can be shared. SF is different from the centralized CBF-CF works (e.g. (Yu et al., 2003)) since SF data and models are distributed through different entities. Hence, there are issues of user management (e.g. adding or removing a user), privacy, security and motivation (e.g. each user should benefit from the collaboration). The main contributions of this paper are: i) we propose the new SF concept 2 that combines filters from distinct entities in order to improve local filtering while assuring privacy; ii) we apply SF to spam detection and compare it with a local CBF filter (i.e. Naive Bayes); iii) the spam detection performance is measured under distinct scenarios, to test the effect of using fixed and incremental symbiotic groups and also to access the robustness to dictionary and focused attacks. This paper is structured as follows. Section 2 presents the related work. Next, we introduce the individual and symbiotic filtering methods (Section 3). The results are presented in Section 4. Finally, closing conclusions are drawn (Section 5). 2 Related Work Several solutions have been proposed to fight spam, which can fall into three main categories (Garriss et al., 2006; Méndez et al., 2008): designing New Mail Protocol Systems (NMPS), Collaborative Filtering (CF) and Content-Based Filtering (CBF). Examples of the first approach are: digitally signing mail, where recipients authenticate the sender’s address and mail content (Wong, 2005); requiring the sender to “pay” (e.g. give a small fee or solve a com- putational puzzle) for each message sent (Loder et al., 2004) or limiting the amount of email any sender may send (Walfish et al., 2006). Yet, none of these solutions is currently adopted in a massive fashion and we are still far from a worldwide acceptance of a NMPS. Also, the proposal of payment systems did not take into account the effect spamming botnets, where a large number ma- chines are controlled for malicious messaging (Ramachandran and Feamster, 2006; Kanich et al., 2008). CF is based on sharing information about spam messages and it can be based on lists (e.g. blacklists with IP addresses of known spammers), or di- gest/fingerprints extracted from spam messages. Often, DNS-based Blackhole Lists (DNSBLs) work in a centralized fashion, being vulnerable to Denial-of- Service (DoS) attacks. Also, they may blacklist legitimate users, with a false negative rate of about 50%, and spammers that use BGP spectrum agility techniques are rarely listed in DNSBLs (Ramachandran and Feamster, 2006). Several CF systems based on social networks have also been proposed. For example, Kong et al. (2005) suggest a system where users manually identify spam and then publish a digest through her/his social network. Garriss et al. (2006) propose the propagation of whitelists among socially connected users. Zhong et al. (2008) introduce a large-scale privacy CF based on digests. CBF filters use a text classifier, such as the popular Naive Bayes algorithm (used by the Thunderbird client), that learns to discriminate spam from mes- sage features (e.g. common spam words). At the present time, CBF is the most used anti-spam solution (Garriss et al., 2006). Current research relies 3 mainly on improving individual classifier performance, by a better preprocess- ing (Méndez et al., 2008) or enhancement of the learning algorithm (Chang et al., 2008; Guzella and Caminhas, 2009). Ensembles that combine distinct spam classifiers have also been proposed (Hershkop and Stolfo, 2005). Both CF and CBF have drawbacks. CF often suffers from first-rater, sparsity of data and privacy problems. The first issue is due to the difficulty of classify- ing emails that have not been rated before, the second problem is present when users rate few messages and the last problem depends on what is shared. For example, while presenting a better privacy digest protocol (when compared with previous CF solutions), the approach of Zhong et al. (2008) is still vul- nerable to privacy breaches. In addition, people have personal views of what is spam and CF often discards this issue (Gray and Haahr, 2004). On the other hand, CBF frequently suffers from lack of sufficient training messages and is highly vulnerable to contamination attacks (as explained in the previous section). By fusing the CF and CBF views there is a potential for a better personalized filtering. However, the number of studies that unify CBF and CF is scarce and mainly focused towards recommendation systems that run at centralized systems (Yu et al., 2003). Garg et al. (2006) describe a spam strategy based on sharing of email filters among collaborating users. The authors point out that exchanging filters requires less communication than CF systems, as the need to share filters is relatively rare when compared with exchanging digests each time an email is received. Yet, they fail to acknowledge motivational (i.e. each user should benefit from the collaboration), temporal (i.e. how to syn- chronize several distinct filters), privacy and security issues regarding filter sharing. More recently, Lai et al. (2009) presented a collaborative approach to exchange spam rules. However, this approach was designed for rule sharing at the server level, demanding secure channels between all these servers. Also, the authors only explored simple attributes (e.g. message length) and not word content. Moreover, explicit (and simple) human understandable rules are re- quired, thus approaches such as Neural Networks or Support Vector Machines are not suitable. In contrast, our SF approach is more flexible, since it can address any type of CBF filter. Also, it is suited for the Web 2.0 paradigm, where users can exchange filters from their social networks. The SF approach should also be more robust to contamination when compared to pure CBF, since it aggregates responses from several (possible unknown) users and tar- geting a specific victim may be easy but contaminating the whole symbiotic group is not. 4 3 Filtering Methods 3.1 Naive Bayes filtering We will address only textual content (i.e. word frequencies) of email mes- sages. This popular approach (e.g. Thunderbird filter) has the advantage of being generalizable to wider contexts, such as spam instant messaging (spim) detection. While different data mining algorithms can be adopted for spam filtering, such as Support Vector Machines (Cheng and Li, 2006), we will use the simpler Naive Bayes (NB), which is widely adopted by anti-spam filtering tools (Metsis et al., 2006; Guzella and Caminhas, 2009). As both individual and symbiotic strategies will be compared using the same learning algorithm, we believe that most of the results presented in this paper can be extended to other text classifiers. We will also adopt the preprocessing proposed in (Kosmopoulos et al., 2008): (1) The word frequencies are extracted from the subject and body message (with the HTML tags previously removed). Each message j is encoded into a vector xj = (x1j, . . . ,xmj), where xij is the number of occurrences of token Xi in the text. (2) The feature selection is applied, which consists in ignoring any words when xij < 5 in the training set and then selecting up to the 3000 most relevant features according to the Mutual Information (MI(Xi)) crite- rion: MI(Xi) = ∑ c∈{s,¬s} p(Xi|c) log( p(Xi|c) p(Xi)p(c) ) (1) where c is the message class (s - spam or ¬s - ham), p(Xi|c) is the probability of finding token Xi in emails from class c, p(Xi) and p(c) are the proportions of Xi terms and c class examples present in the data. (3) Each xij value is transformed into: x ′ ij = log(xij + 1) (TF transform), x′′ij = x ′ ij ·log(k/ ∑ k δik) (IDF transform) and x ′′′ ij = x ′′ ij/ √∑ l(xlj) 2 (length normalization), where δik is 1 if the token i exists in the message k and 0 otherwise. The NB computes the probability that a document j is spam (s) for a filter trained over Du email data from user u, according to: p(s|xj,Du) = α ·p(s|Du) m∏ i p(Xi|s,Du) (2) 5 where α is normalization constant that ensures that p(s|x,Du)+p(¬s|x,Du) = 1, p(s|Du) is the p(s) of dataset Du. The p(Xi|s,Du) estimation depends on the NB version. In this work, we will use the multi-variate Gauss NB (as implemented in the R tool, see Section 4) (Metsis et al., 2006): p(Xi|c,Du) = 1 σi,c √ 2π exp(− (x′′′ij −µi,c)2 2σ2i,c ) (3) where µi,s and σi,s are the mean and standard deviation estimated from the c = s or c = ¬s messages of Du. In (Nelson et al., 2008), it has been shown that local spam filters are vulnerable to dictionary and focused contamination attacks. The former attack is used to reduce the CBF efficiency, leading the victim to read spam, while the latter can be used to prevent the victim from reading an important email. Both attacks can be achieved by sending spam messages mixed with normal words. Once the victim labels these messages as spam, the training set is contaminated and the filter will be affected the next time it is retrained. A dictionary aggression consists in sending a large amount of normal words, while the focused assault assumes that the attacker has some knowledge of a specific message that the victim will receive in the future (e.g. a competing offer for a given contract). 3.2 Symbiotic filtering In our proposed SF, the individual predictions can be combined by using a collaborative ensemble of the local filters. To tackle the concept drift (i.e. the learning tasks changes through time) nature of spam (Fdez-Riverola et al., 2007), the ideal symbiotic combination function should be dynamic. To achieve this we propose a hierarchical learning, where the outputs of the distinct fil- ters are used as the inputs of another (meta-level) learner. Hence, each user has a local meta-learner that is responsible for aggregating the distinct filter responses. This meta-learner is dynamically trained to get a high accuracy on the user past data, thus it assigns different weights to the CBF filters through time (see Figure 6). While several algorithms could be used for this hierarchi- cal learning (e.g. SVM), we will adopt the same NB described in the previous section. The rationale is that NB is commonly adopted by anti-spam solutions, thus incorporating SF into these tools would be simpler by reuse of code. We assume that each user u trains a local filter θu,t over her/his Du training data. Filters can be trained asynchronously and L filters will be available for each user at time t: {θ1,t, . . . ,θL,t}. The Symbiotic NB (SNB) meta-model 6 spam probability is given by: p(s|xj,D′u) = α ·p(s|D′u) ∏L i=1 p(θi,t|s,D′u) p(θi,t|c,D′u) = 1 σi,c √ 2π exp(−(p(s|xj,θi,t,D ′ u)−µi,c)2 2σ2 i,c ) (4) where D′u is the SNB training set and p(s|xj,θi,t,D′u) is the probability given by the filter θi,t, as computed in Equation 2. To reduce memory and compu- tational requirements, we allow that D′u ⊆ Du, where M = |D′u| denotes the most recent messages from u mailbox. It should be noted that any token from xj that is not considered by θi,t will simply be discarded by the filter from user i. Similarly, any input attribute from θi,t that is not included in xj will be set to 0. While sharing models is less sensitive than exchanging email messages, there are still privacy issues to be considered. For instance, if user A has access to the filter of user B, then A may feed a given token (or set of tokens) into the model and thus know some probability that such token was classified by B as spam or ham. Our privacy solution resides in an anonymous exchange of the filters, which can occur under a centralized server or a Peer-to-Peer (P2P)-like application. Under the first option, all users register into a centralized and secure service. This service could be implemented by large companies or email providers (e.g. Gmail or Hotmail), when all emails are stored at a given server. For scalabil- ity, user profiles could be defined (e.g. country or profession) and clustering algorithms could be used to group users with similar interests. Another vari- ant would be the definition of social networks, where users could choose their “friends”. These systems could, for example, be implemented by social net- working websites (e.g. Facebook or MySpace). Alternatively, when the mes- sages are stored locally at the client side, the server would be responsible for a blind exchange of the filters, using secure transfers (left of Figure 1). To exchange the filters, a standard format should be adopted, such as the Pre- dictive Model Markup Language (PMML) (Grossman et al., 2002), which is compatible with a large number of data mining tools. In should be noted that exchanging filters requires less communication costs. For example, a filter built from a millions of emails can be described by a few hundreds or thousands of bytes (depending on the filter algorithm used) (Garg et al., 2006). When a new email is received, the user can easily compute the SF, as a copy of all filters is available locally. The above systems have the disadvantage of having to depend and trust in a centralized service. As an alternative, the use of a P2P-like distribution scheme is also possible to be adopted (right of Figure 1). Under this solution all peers may donate, store and fetch filters among each other. This approach 7 anonymousanonymous filter filter user A user C user B anonymous filter filter anonymous filter A filter B user C user Buser A server secure Fig. 1. Anonymous exchange of filters by using a secure server (left) or a P2P-like application (right) could be implemented as a trusted and secure plug-in of an email client (e.g. Thunderbird). The filter sharing process among the peers could work similarly to the explained in the secure server scheme (i.e. using PMML). Further pri- vacy increase could be achieved if each individual does not know the symbiotic group composition. However, for some scenarios it may also be attractive that the symbiotic group composition could be assessed by all the participants in a given group. Yet, even in such scenarios it might be very difficult to “guess” who created each particular model, as in SF there will be typically a large number of users that dynamically may join or leave the collaboration. 4 Experiments and Results 4.1 Spam data To evaluate SF, ideally there should be real mailboxes collected from distinct users (possibly from a social network) during a given time period. Yet, due to logistic and privacy issues, it is quite difficult to obtain such data (in par- ticular personal messages) and make it public. Hence, we will use a synthetic mixture of real spam and ham messages, in a strategy similar to what has been proposed in (Metsis et al., 2006; Zhong et al., 2008). The ham messages will come from the Enron email collection, which was originally used for a global evaluation of filters by merging all Enron user messages into a single corpus. In particular, we will use the cleaned-up form provided by (Becker- mann et al., 2004) and we will select the five Enron employees with the largest mailboxes collected during the same time period: kaminski-v (kam), farmer-d (far), beck-s (bec), lokay-m (lok) and kitchen-l (kit). Since these employees worked at the same organization, it is reasonable to assume that they would know each other, i.e. belong to a social network. We will also use the spam collection of Bruce Guenter (http://untroubled.org/spam/), which is based in spam traps (i.e. fake emails published in the Web), during the years of 2006 and 2007 (our dataset was built in 2008). Only messages with Latin character sets were selected, because the ham messages use this type of character coding and non-Latin mails would be easy to detect. Also, since this collection con- 8 tains several copies of the same messages (due to the use of multiple traps), we removed duplicates by comparing MD5 signatures of the body messages. We propose a mixture algorithm that is based on the time that each message was received (date field, using the GMT time zone). By preserving the tem- poral order of the emails, we believe that a more realistic mixture is achieved than the sampling procedures adopted in (Metsis et al., 2006) or (Zhong et al., 2008). Since the Enron data is from a previous period (see Table 1), we first added 6 years to the date field of all ham messages. Let St denote a spam mes- sage received at time t, Si,f = (Sti,Sti+1, . . . ,Stf ) the time ordered sequence of the Bruce Guenter spam, Hu,i,f and Su,i,f the sequences of ham and spam messages for user u from time ti to tf . For a given time period t ∈ (ti, . . . , tf ), the algorithm randomly selects |S′i,j| spam messages from Si,j. Then, Su,i,j is set by sampling messages from S′i,j with a probability of P for each message selection. The size of S′i,j (cardinality) is given by: |S′i,j| = R· ∑L i=1 |Hu,i,j| P ·L (5) where L denotes the total number of users available at the time period and R is the overall (i.e. including all user and time data) spam/ham ratio. Since the time periods are different for each user (Table 1), four time sequences (i.e. ti and tj values) were used by the algorithm (Figure 2). Table 1 The S-Enron corpus main characteristics user ham spam time spam size size period /ham kam 4363 2827 [12/05,05/07] 0.6 far 3294 2844 [12/05,05/07] 0.9 bec 1965 2763 [01/06,05/07] 1.4 lok 1455 2202 [06/06,05/07] 1.5 kit 789 623 [02/07,05/07] 0.8 The mixture is affected by the R and P parameters. Since a high number of ex- periments is addressed in this work, we will fix these parameters to reasonable values. While the global spam/ham ratio is R = 1, the individual ratios range from 0.6 to 1.5. Also, the spam/ham ratios fluctuate through time (as shown in Figure 3). On the other hand, the probability of spam selection affects the percentage of common spam between users. If two users have similar profiles (e.g. email exposure), then they should receive similar spam. We assume that this scenario is expected for the Enron employees and thus set P = 0.5. Under this setup and for a given time period, any 2 users will receive around 50% of 9 Date (Year) 2006 2007 kam far bec lok kit L=2 L=3 L=4 L=5 A B C Fig. 2. Time view of the S-Enron mailboxes similar spam, 3 users will share around 25% of spam and so on. The result- ing corpus is named S-Enron and it is publicly available in its raw form at: http://www3.dsi.uminho.pt/pcortez/S-Enron. 0 10 20 30 40 50 60 0 .5 1 .0 1 .5 2 .0 batch (x100 emails) ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Fig. 3. Evolution of the spam/ham ratio for the user far and scenario C 4.2 Evaluation As explained in Section 3.2, spam detection suffers from concept drift (Fdez- Riverola et al., 2007). Features such as the amount of spam received, the ham/spam ratio and even the content itself evolve through time. Hence, to evaluate spam filters, we will adopt the more realistic incremental retraining evaluation procedure, which periodically trains and tests filters. Under this procedure, a mailbox is split into batches b1, . . . ,bn of K adjacent messages (|bn| may be less than K) (Metsis et al., 2006). For i ∈ {1, . . . ,n − 1}, the filter is trained with Du = b1 ∪ . . .∪bi and tested with the messages from bi+1 (Figure 4). This procedure is more realistic than the simple 50% train/test split adopted in (Zhong et al., 2008). The predicted class for a probabilistic filter is given by s if p(s|xj,Du) > D, where D ∈ [0.0, 1.0] is a decision threshold. For a given D and test set, it is 10 K |b | 1 2 3 test set test set test set training set tr. set training set n−1 training set ts. set ...... runs b b b 1 2 3 batch b n ... mailbox KKK n Fig. 4. The incremental retraining procedure possible to compute the true (TPR) and false (FPR) positive rates: TPR = TP/(TP + FN) FPR = FP/(TN + FP) (6) where TP , FP , TN and FN denote the number of true positives, false posi- tives, true negatives and false negatives. The receiver operating characteristic (ROC) curve shows the performance of a two class classifier across the range of possible threshold (D) values, plotting FPR (x-axis) versus TPR (y-axis) (Fawcett, 2006). The global accuracy is given by the area under the curve (AUC= ∫ 1 0 ROCdD). A random classifier will have an AUC of 0.5, while the ideal value would be 1.0. Since the cost of losing normal email (FP) is much higher than receiving spam (FN), D is usually set to favor points in the low false-positive region of the ROC. Thus, we will also adopt the Normalized AUC (NAUC), which is the AUC area in the section FPR ≤ r, divided by r (Chang et al., 2008). Typically, the target FPR rate (r) is close to 0.0. We will also compute the relative Gain of the symbiotic performance over the local filter: Gain= ξSNB/ξNB − 1, where ξ is the evaluation metric (i.e. AUC or NAUC) and SNB and NB are the symbiotic and individual filters. With the incremental retraining procedure, one ROC will be computed for each bi+1 batch and the overall results will be presented by adopting the vertical aver- aging ROC (i.e. according to the FPR axis) algorithm presented in (Fawcett, 2006). Statistical confidence will be given by a paired t-student test, at the 95% confidence level (Flexer, 1996). 4.3 Experimental setup All experiments were conducted in the R environment, an open source and high-level programming language for data analysis (R Development Core Team, 2008). In particular, the NB algorithm described in Section 3.1 is implemented by the naiveBayes function of the e1071 R package, while the text prepro- cessing uses several functions from the tm package (Feinerer et al., 2008). 11 During the all experiments, we set K = 100 (a reasonable value also adopted in (Metsis et al., 2006)). For the SNB, we used a similar number for the hierarchical training set size, i.e. M = 100. This small value has the advantage of reducing memory requirements (the user only needs 100 messages in his mailbox) and some initial experiments with larger values of M revealed no gain in performance. The AUC, NAUC and Gain values will be shown in percentage. All NAUC values will be computed with r = 0.01 (1%). 4.4 Fixed symbiotic group Two distinct scenarios will be tested, according to the time periods A and B of Figure 2. Given the S-ENRON corpus characteristics, in this work we will explore a small number of fixed symbiotic users: L = 5 for A and L = 3 for B. The incremental retraining method (Section 4.2) was applied to both sce- narios, by considering all messages within the corresponding time period. Thus, the number of kam, far and bec batches (n) will be different for A and B. The obtained results are summarized as the mean of all test sets (bi+1, i ∈ {1, . . . ,n − 1}) and shown in Table 2 and Figure 5. The best values are in bold, while underline denotes a statistical significance (i.e. p-value<0.05). In Figure 5, bars denote 95% t-student confidence intervals and only the most interesting region of FPR is shown for the ROC curves. For the first scenario (A), the symbiotic strategy outperforms the local filter for all users and metrics, except for lok and NAUC. A similar behavior occurs for the B setting, where SNB is better than NB except for kam and AUC. As false positives have higher costs in spam detection, the NAUC results are particularly important. Thus, it is interesting to notice that there is a high AUC improvement (i.e. Gain) given by the symbiotic method in several cases (kam, far and kit for A and bec for B). To demonstrate the SNB dynamics, Figure 6 shows the first two consecutive graphs of the SNB input importances under scenario II. Each edge represents the influence (in %) of the NB filter (the origin) in the symbiotic model (the destination), as measured by applying a sensitivity analysis procedure (Kewley et al., 2000). The text in bold (e.g. b2) denotes the last batch used to train the NB classifier. For example, the first SNB model of user far (left graph) uses a NB filter from kam that was trained using 200 messages (Dkam = b1 ∪ b2). 12 Table 2 The results for scenarios A and B scenario A AUC NAUC user n NB SNB Gain NB SNB Gain kam 16 62.1 95.6 54 0.8 74.4 9804 far 13 93.5 95.1 2 15.6 58.6 276 bec 9 91.5 94.0 3 53.5 65.8 23 lok 9 91.4 95.2 4 79.9 75.6 -3 kit 15 74.6 95.3 28 18.2 71.7 294 scenario B AUC NAUC user n NB SNB Gain NB SNB Gain kam 70 94.7 94.3 -0.4 54.0 73.0 35 far 60 89.4 91.7 2.6 54.3 66.9 23 bec 48 83.5 93.4 11.9 23.8 74.3 212 4.5 Incremental symbiotic group A more realistic scheme is adopted for the time period C, where users join the symbiotic group in an incremental fashion, at different time stages according to Figure 2. Thus, L will grow from 2 to 5. The results are presented in Table 3. As expected, the symbiotic strategy (SNB) clearly favors newcomers, which have small mailboxes and thus benefit from the collaboration. In effect, the NAUC differences are quite large, such as in bec and lok for L = 4 and L = 5; and kit for L = 5. For demonstration purposes, the ROC curves are plotted for kam, far and bec, when L = 3 and L = 5 (Figure 7). However, the results show that even “veteran” users benefit from the symbiotic relation when the number of users grow. For instance, the kam and far NAUC results for L = 5 improve, with a gains of 28% and 16%, respectively. 4.6 Contamination attacks We will repeat the experiments of Section 4.4, by considering only scenario A and user bec to test the effects of mailbox contamination. The dictionary assault is simulated by replacing the first 10 spam emails at batch 4 from bec by the GNU aspell (http://aspell.net/) English dictionary (version 6.0, with 138599 tokens). In Figure 8, gray lines denote the behavior of NB and SNB without the attack (i.e. results of Section 4.4), black lines show the per- formance under the attack and the dot-dashed vertical line shows when the 13 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 kam (A) ● ● ● ● ● ● ● ● ● ● ● ● SNB NB 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 far (A) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 bec (A) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 lok (A) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 kit (A) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 kam (B) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 far (B) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 bec (B) ● ● ● ● ● ● ● ● ● ● ● Fig. 5. ROC curves for scenarios A and B bec kam far bec kam far b 1 b 1 b 1 b 2 b 3 b 3 b 2 b 2 b 2 b 2 31% 33% 33% 31% 41% 33% 26% 100% 32% 34% 34% 33% 33% 33% 33% b 38% 1 b 2 b 2 b 1 b 2 b 1 Fig. 6. Examples of SNB input importances for B attack starts. The filters of bec are not trained with contaminated messages at batch 4, yet these messages appear in the test set and thus the NB and SNB performances suffer a moderate decay. The true effect of the attack is only visible at batch 5, where local CBF is highly affected. Only 10 messages were replaced and yet the filter detection capability is reduced to a random classi- fier (since AUC=0.5) through all remaining batches. In contrast, the symbiotic 14 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 kam (L=3) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 kam (L=5) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 far (L=3) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 far (L=5) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 bec (L=3) ● ● ● ● ● ● ● ● ● ● ● 0.00 0.02 0.04 0.06 0.08 0.10 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 bec (L=5) ● ● ● ● ● ● ● ● ● ● ● Fig. 7. ROC curves for users kam, far and bec (L=3 at left and L=5 at right) method is only initially affected, since as time goes by the performance gets closer to the no attack scenario. Also, the remaining symbiotic users maintain their spam detection capabilities, as shown by the far results, which is a rep- resentative example. This behavior is explained by the SNB algorithm, which simply discards a given filter if it does not help to predict the recent past M messages of the user. Hence, this experiment shows that SF is robust also to saboteurs, i.e. if a particular user intentionally feeds the group with a random or bad filter then this filter will be simply ignored. 0 .5 0 .6 0 .7 0 .8 0 .9 1 .0 bec test set batch (x100 e−mails) A U C 2 3 4 5 6 7 8 9 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● SNB NB SNB (attack) NB (attack) 0 .5 0 .6 0 .7 0 .8 0 .9 1 .0 far test set batch (x100 e−mails) A U C 4 5 6 7 8 9 10 11 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● SNB SNB (attack) Fig. 8. The effect of the dictionary attack The dictionary attack can be solved by performing a rollback (i.e. returning 15 Table 3 The results for scenario C L=2 AUC NAUC user n NB SNB Gain NB SNB Gain kam 4 94.4 88.1 -3 30.1 48.4 61 far 3 89.9 82.1 -9 60.7 54.3 -11 L=3 AUC NAUC user n NB SNB Gain NB SNB Gain kam 17 91.5 87.4 -5 47.5 59.5 25 far 16 86.6 87.0 0.4 44.6 59.1 33 bec 14 80.4 87.6 9 58.2 65.8 13 L=4 AUC NAUC user n NB SNB Gain NB SNB Gain kam 39 95.4 96.8 1.5 75.0 75.6 0.8 far 32 88.7 94.7 6.8 53.1 72.5 37 bec 28 84.5 96.9 15 13.0 78.6 504 lok 29 73.1 96.2 32 6.1 71.5 1076 L=5 AUC NAUC user n NB SNB Gain NB SNB Gain kam 15 98.4 98.0 -0.5 61.3 78.2 28 far 13 89.6 95.5 6.5 61.8 71.9 16 bec 8 85.2 97.4 14 1.8 70.3 3848 lok 9 63.6 97.3 53 0.7 78.3 10992 kit 15 74.6 97.9 31 18.2 71.8 295 to the previous filter) or using the RONI defense (Nelson et al., 2008), which rejects training examples that have a large negative impact in spam detection. Yet, focused assaults are more difficult to prevent and finding a defense is still an open problem (Nelson et al., 2008). We believe SF is an interesting solution due to the same rationale presented for the dictionary aggression, i.e. the combination of multiple filters should overcome the limitations of a single model contamination. A new set of experiments was devised, using again scenario A and bec mailbox. During a given run, a legitimate message was randomly selected, from batches 6 to 9, as the target text. We assume that the attacker is confident about the 16 target content and thus can guess 50% of the target words. At batch 4, 10 spam emails were replaced by the contaminated messages. We repeated this procedure during 20 runs. The effect of this attack on spam is minimal and thus we will only show the effect on the target ham emails. Figure 9 plots the filter spam probability (y-axis) for each target message. The obtained probability for each run is plotted along the x-axis (total of 20 runs). Since all target messages are ham, a robust filter should present low spam probabilities, near the zero horizontal axis. The results show that local filter (NB) is much more vulnerable to focused attacks than the symbiotic strategy (SNB). The spam probability mean values of NB and SNB are 0.69 and 0.32 (the differences are statistically significant). For example, when using a decision threshold of D = 0.5, 14 (of 20) messages are classified by NB as spam, while this number lowers to 6 for SNB. Even if D is raised to 0.999, NB predicts 13 spam emails and SNB only detects 5. 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 runs sp a m p ro b a b ili ty f o r ta rg e t te xt ( h a m ) 1 5 10 15 20 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● SNB NB Fig. 9. The effect of the focused attack 5 Conclusions Email has long been one of the most important and widely used Internet ap- plications. However, spam emerged quickly after email itself and nowadays it accounts for the majority of the email traffic. Thus, several anti-spam tech- niques were developed. This paper proposes a novel Symbiotic Filtering (SF) approach, which combines several features from Collaborative Filtering (CF) and Content-Based filtering (CBF). It takes the advantage of use of social networks, where users with the same or related interests have the opportunity to form mutually beneficial alliances with the aim of enhancing spam detec- 17 tion techniques. Instead of sharing messages or digests, the idea is to share filters. To combine the individual probabilities, the proposed solution uses the concept of hierarchical learning, where a meta-learner is dynamically trained to improve its accuracy. After describing the proposed model, we compared the effectiveness of the SF versus CBF, using for that purpose a realistic mixture of real spam and ham messages. Promising results were obtained by the SF, which outperformed the local filtering for a small number of users (from 3 to 5). Moreover, we have shown that SF is more robust to word attacks (e.g. dictionary or focused as- saults). Furthermore, we proposed several deployment scenarios for SF, under a secure server or P2P settings. There is a continuous race between spammers and anti-spammers and a local classifier that is currently perfect will be eventually defeated. We believe that a stronger protection is achieved by adopting a dynamic cooperation of filters from distinct users. In future work, we intent to apply SF to other personalized filtering scenarios, such as Web page blocking (e.g. sensitive content). Also, we will explore scalability issues. Under a large group, this could be achieved by adopting user selection algorithms (e.g. clustering user profiles). Acknowledgments This work is supported by FCT grant PTDC/EIA/64541/2006. References Alocilja, E. (2000). Principles of Biosystems Engineering. Erudition Books, Massachusetts, USA. Beckermann, R., McCallum, A., and Huang, G. (2004). Automatic catego- rization of email into folders: benchmark experiments on Enron and SRI corpora. Ir-418, University of Massachusetts Amherst. Chang, M., Yih, W., and Meek, C. (2008). Partitioned Logistic Regression for Spam Filtering. In 14th ACM SIGKDD int. conference on Knowledge discovery and data mining, pages 97–105. Cheng, V. and Li, C. (2006). Personalized Spam Filtering with Semi- supervised Classifier Ensemble. In IEEE/WIC/ACM International Con- ference on Web Intelligence. Fawcett, T. (2006). An introduction to ROC analysis. Pattern recognition letters, 27(8):861–874. Fdez-Riverola, F., Iglesias, E. L., Dı́az, F., Méndez, J. R., and Corchado, J. M. 18 (2007). Applying lazy learning algorithms to tackle concept drift in spam filtering. Expert Systems with Applications, 33(1):36–48. Feinerer, I., Hornik, K., and Meyer, D. (2008). Text Mining Infrastructure in R. Journal of Statistical Software, 25(1-54). Flexer, A. (1996). Statistical Evaluation of Neural Networks Experiments: Minimum Requirements and Current Practice. In Proceedings of the 13th European Meeting on Cybernetics and Systems Research, volume 2, pages 1005–1008, Vienna, Austria. Garg, A., Battiti, R., and Cascella, R. (2006). May I borrow your filter? Exchanging filters to combat spam in a community. In Advanced Infor- mation Networking and Applications, 2006. AINA 2006. 20th International Conference on, volume 2. Garriss, S., Kaminsky, M., Freedman, M., Karp, B., Mazières, D., and Yu, H. (2006). RE: reliable email. In Proceedings of the 3rd conference on Networked Systems Design and Implementation (NSDI), pages 297–310, San Jose, CA. USENIX Association Berkeley, CA, USA. Gray, A. and Haahr, M. (2004). Personalised, Collaborative Spam Filtering. In 1st Conference on E-Mail and Anti-Spam CEAS. Grossman, R., Hornick, M., and Meyer, G. (2002). Data Mining Standards Initiatives. Communications of ACM, 45(8):59–61. Guzella, T. and Caminhas, W. (2009). A review of machine learning ap- proaches to Spam filtering. Expert Systems with Applications, 36:10206– 10222. Hershkop, S. and Stolfo, S. (2005). Combining Email Models for False Positive Reduction. In 11th ACM SIGKDD int. conference on Knowledge discovery and data mining, pages 21–24. Kanich, C., Kreibich, C., Levchenko, K., Enright, B., Voelker, G., Paxson, V., and Savage, S. (2008). Spamalytics: An Empirical Analysis of Spam Mar- keting Conversion. In Computer and Communications Security Conference (CCS’08), pages 27–31. ACM. Kewley, R., Embrechts, M., and Breneman, C. (2000). Data Strip Mining for the Virtual Design of Pharmaceuticals with Neural Networks. IEEE Trans Neural Networks, 11(3):668–679. Kong, J., Boykin, P., Rezaei, B., Sarshar, N., Roychowdhury, V., Rothenstein, B., Damian, I., Paramonov, P., Lyuksyutov, S., Barrat, A., et al. (2005). Let your cyberalter ego share information and manage spam. eprint arXiv: physics/0504026. Kosmopoulos, A., Paliouras, G., and Androutsopoulos, I. (2008). Adaptive Spam Filtering Using Only Naive Bayes Text Classifiers. In CEAS 2008 - Fifth Conference on Email and Anti-Spam. Lai, G., Chen, C., Laih, C., and Chen, T. (2009). A collaborative anti-spam system. Expert Systems with Applications, 36:6645–6653. Loder, T., Van Alstyne, M., and Wash, R. (2004). An economic answer to unsolicited communication. In proceedings of the 5th ACM conference on Electronic Commerce, pages 40–50. ACM New York, NY, USA. 19 MAAWG (2009). Email Metrics Program: The Network Operators’ Perspec- tive. Report #10 – third and fourth quarter 2008, Messaging Anti-Abuse Working Group, S. Francisco CA, USA. Méndez, J., Cid, I., Glez-Peña, D., Rocha, M., and Fdez-Riverola, F. (2008). A Comparative Impact Study of Attribute Selection Techniques on Näıve Bayes Spam Filters. In Springer, editor, 8th Industrial Conference on Data Mining, volume LNAI 5077, pages 213–227. Metsis, V., Androutsopoulos, I., and Paliouras, G. (2006). Spam Filtering with Naive Bayes – Which Naive Bayes? In Third Conference on Email and Anti-Spam (CEAS), pages 125–134. Nelson, B., Barreno, M., Chi, F., Joseph, A., Rubinstein, B., Saini, U., Sutton, C., Tygar, J., and Xia, K. (2008). Exploiting Machine Learning to Subvert Your Spam Filter. In 1st Usenix Workshop on Large-Scale Exploits and Emergent Threats, pages 1–9. ACM Press. R Development Core Team (2008). R: A language and environment for statis- tical computing. R Foundation for Statistical Computing, Vienna, Austria, ISBN 3-900051-00-3, http://www.R-project.org. Ramachandran, A. and Feamster, N. (2006). Understanding the Network- Level Behavior of Spammers. In ACM, editor, SIGCOMM’06, pages 291– 302. Walfish, M., Zamfirescu, J., Balakrishnan, H., Karger, D., and Shenker, S. (2006). Distributed Quota Enforcement for Spam Control. In Proceedings of the 3rd conf. on Networked Systems Design and Implementation (NSDI), pages 281–296, San Jose, CA. USENIX Association Berkeley. Wong, M. (2005). Sender authentication: What to do. White Paper, July. Yu, K., Schwaighofer, A., Tresp, V., Ma, W., and Zhang, H. (2003). Collab- orative Ensemble Learning: Combining Collaborative and Content-Based Information Filtering via Hierarchical Bayes. In 19th Int. Conf. on Uncer- tainty in Artificial Intelligence (UAI), pages 353–360. ACM. Zhong, Z., Ramaswamy, L., and Li, K. (2008). ALPACAS: A Large-scale Privacy-Aware Collaborative Anti-spam System. In IEEE INFOCOM, pages 556–564. 20 
work_fu4sy3h4kvastowjmwh2xx3zeu	----	General and specific formalization approach for a Balanced Scorecard: An expert system with application in health care General and specific formalization approach for a Balanced Scorecard: An expert system with application in health care Holger Kunz ⇑, Thorsten Schaaf Charité Universitätsmedizin Berlin, Institute of Medical Informatics, Hindenburgdamm 30, 12203 Berlin, Germany a r t i c l e i n f o Keywords: Health care Balanced Scorecard Formalization Aggregation Utility function a b s t r a c t The Balanced Scorecard (BSC) presents the essentials of strategic and performance management in clear, straightforward manner which is also usable in health care. If a BSC for a clinical department is an agree- ment, the first question to consider is the method by which it can be ascertained whether a strategy has been accomplished. There are many different techniques like AHP (analytic hierarchy process) and fuzzy systems to calculate indices. However, how does a formalized mathematical groundwork looks like that integrates current approaches and is still general enough to incorporate future expert systems with applications? The purpose of this paper is the formalization of BSC evaluation by respecting current research. The for- malized expert system was implemented in an information system for health care management. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction and motivation Any hospital is confronted by substantial society changes due to a variety of factors such as increasing life expectancy and popu- lation ageing, economical pressure and competition, limited resources as well as shrinking and tight budgets, new governmen- tal deregulations and liberalization. In order to cope with the changing nature of this environment, which is also aggravated by deep structural reorganization, it is important to accomplish suc- cess goals with a tool of strategic management such as a Balanced Scorecard. In order to support strategy and performance management, Robert Kaplan and David Norton’s ‘‘Balanced Scorecard” concept (Kaplan, 1992) may be used. It is a tool of strategic management, strategic communication and performance management, providing frequently measured performance and regular reviewing and refinement strategy with an ongoing evaluation process of clinical indicators. A BSC design and implementation process can be sepa- rated into four stages: (1) translating the vision and gaining con- sensus; (2) communicating the objectives, setting the goals, and linking strategies; (3) setting targets, allocating resources, and establishing milestones; (4) and feedback and learning (Stewart & Bestor, 2000). Types of performance indicators and factors are termed as ‘‘perspectives”. These perspectives as in the original def- inition are differentiated into a financial, a customer, and a process and innovation perspective. The simply monitoring of key financial indicators, which have often been historical in nature and concen- trated almost exclusively on lagging indicators are hiding the key drivers. Examples of Balanced Scorecards in health care can be found for a burn centre (Wachtel, Hartford, & Hughes, 1999), a kidney transplant (Colaneri, 1999), an ambulant treatment (Curtright, Stolp-Smith, & Edell, 2000), an electronic patient record (Gordon & Geiger, 1999), a children’s hospital (Meliones, 2000), dialysis (Peters, 1999), anesthesiology (Zbinden, 2002), cardiology (Chang, 2002), behaviour therapy (Santiago, 1999), information strategy of Canadian NHS (Protti, 2002), and indicator system for Dutch health system (Asbroek et al., 2004). These articles give a brief overview about BSCs and their development and application in healthcare management. Strategic management is an externally oriented philosophy of managing an organization that links strategic thinking and analysis to organizational action (Ginter, Swayne, & Duncan, 2002). The strategic process can be presented as a model where key elements include an environment scanning, strategy formulation, strategic implementation as well as a strategic evaluation (Wheelen & Hun- ger, 2004). The following points should be noted in connection with the practical implementation of clinical Balanced Scorecards: 1. Which indicators are relevant for the clinical Balanced Scorecard? The proposed strategic management system should enable hos- pital managers, not just to be informed about financial indica- tors but also about specific requirements that are relevant to health care. Indicators consist of various measures that may 0957-4174/$ - see front matter � 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.07.127 ⇑ Corresponding author. E-mail addresses: holger.kunz@bdvb.de (H. Kunz), thorsten.schaaf@charite.de (T. Schaaf). Expert Systems with Applications 38 (2011) 1947–1955 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2010.07.127 mailto:holger.kunz@bdvb.de mailto:thorsten.schaaf@charite.de http://dx.doi.org/10.1016/j.eswa.2010.07.127 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa be expressed or implied. Not all the indicators are of equal importance and the healthcare manager has to seek so to classify them according to their importance. However, this is beyond the scope of this paper which focuses on the evaluation. 2. How to evaluate indicators of a clinical Balanced Scorecard? The main focus of this study is how to present the multitude of indi- cators in an adequate way for managing purposes. First of all, it is necessary to answer the utilization of each individual indica- tor. This means to which degree is a specific target for an indi- cator accomplished. Secondly, how can these indicators be aggregated in a general way on the foundation of an appropriate metric into a single performance index for each perspective and on BSC level these perspective indices into one entire BSC index. Although many publications (Asbroek et al., 2004; Colaneri, 1999; Curtright et al., 2000; Chang, 2002; Gordon & Geiger, 1999; Meliones, 2000; Peters, 1999; Protti, 2002; Santiago, 1999; Wachtel et al., 1999; Zbinden, 2002) describe Balanced Scorecards in health care only few like (Tarantino, 2003; Griffith & White, 2005) show how an overall index can be computed. This study therefore aims to formalize the computation of an index both for entire BSC and for each perspective. Therefore it uses concepts of decision theory that give a profound base for formalization. The method part of this paper shows a general definition of an indicator system for each perspective in a BSC for a clinical appli- cation and how a utility function can be defined. In addition to util- ity functions a formalization of a general approach as utility values can be aggregated into an index for a perspective and an entire BSC. The definition is general and implies no assumptions of the imple- mented utility functions and aggregation. Unlike, the general defi- nition the result section of this paper explicitly defines a specific set of utilization functions and the manner of aggregation. With re- gard to the formalization an implementation was made with co- operation with an orthopedics department. 2. Theory and methods The formalization of the Balanced Scorecard presented here is in part based on decision theory. Decision theory is the product of the joint efforts of economists, mathematicians, philosophers, social scientists, and statisticians toward making sense of how individu- als and groups make or should make decisions (Resnik, 1987). Decision theory provides a formal framework for making logical choices in the face of uncertainty. Given a set of alternatives, a set of consequences, and a correspondence between those sets, decision theory offers conceptually simple procedures for choice (Parmigiani & Inoue, 2009). It is thus usual to divide decision the- ory into two main branches: normative (or prescriptive) decision theory and descriptive decision theory (Resnik, 1987). The basic assumption is that decision functions are based on preferences. A function or representation of preferences is regarded to be a utility function or evaluation function if each alternative can be associ- ated with a real number (Resnik, 1987). 2.1. General formalization The functional criterion for evaluation of indicators for a clinical balanced scorecard needs an appropriate metric for an afterwards or ex post consideration. For each perspective j exist a set of nj dis- joint indicator xij , which can be combined to an indicator vector ~xj ¼ðx1j ; . . . ; xnjÞ T ; xij 2 R. 2.1.1. Utility functions An indicator vector of a perspective is an element of an nj- dimensional real number space. An indicator vector ~xj can be as- signed with a vector function U ! jð~xjÞ to an evaluation vector. The vector function U ! jð~xjÞ contains a number of nj one-dimensional functions UijðxijÞ that assigns each component to a single utility value. 2.1.2. Aggregation functions The resulting vector of U ! jð~xjÞ can be assigned to a single scalar Pj. This can be interpreted as aggregation function or index for a perspective Pj. Function Cj U ! jð~xjÞ � � incorporates the mathematical method of transferring the utility vector into a single real value. Pj ¼ Cj U ! jð~xjÞ � � : ð1Þ For a classical Balanced Scorecard with a financial, customer-, pro- cess- and innovation perspective let the number of perspective be m = 4. The aggregation function of the entire BSC for disjoint per- spective evaluations Pj with j 2 {1, . . . , m} is: B ¼ KðP1; . . . ; PmÞ¼ K C1 U ! 1ð~x1Þ � � ; . . . ; Cm U ! mð~xmÞ � �� � : ð2Þ The functions U, C und K incorporate preferences with weighting values. How this weighting is performed in a concrete situation is still undefined and left to the special definition. Fig. 1 show how the hierarchy level looks like. This hierarchy is general and abstract. Specific utilization and aggregation at this point are still left to indi- vidual definition. 2.1.3. Decision rule Coming to a decision rule it is useful to know before and there- fore ex ante how prospective changes can be evaluated. The under- lying assumption is that the decision maker is rational in a way that he wants to increase the entire score B of a BSC throughout different time period. Thus, each period t has one and only one en- tire evaluation Bt. If the decision rule from time period t to time period (t + 1) is to increase the score then the necessary condition is Bt 6 Bt+1. However, the decision maker does not only want to in- crease the score in successive time periods but also to maximize it. His sufficient condition is therefore to select one alternative that leads to a maximum value. Let Aj be the set of result from activities that are accomplishable. A rational decision maker would choose for a perspective Pj that alternative ~a which is the maximum: ð8~b : Cð~aÞ P Cð~bÞÞ^~a;~b 2 Aj: ð3Þ It is possible that there exist more than one indicator vector that satisfies the requirement. In this case the decision maker is indiffer- ent to these alternatives. These indifferent indicator vectors belong to one equivalence class that results by the homomorphism of the function. 3. Results The results are classified into utility functions, techniques to aggregation, an approach to fast computation and examples of an application of an expert system. 3.1. Special formalization The above general formalization establishes certain require- ments that must be satisfied before the exact scope of a utility function and aggregation can come into play. A special formaliza- tion precisely defines calculation principles that can be imple- 1948 H. Kunz, T. Schaaf / Expert Systems with Applications 38 (2011) 1947–1955 https://isiarticles.com/article/379 
work_fxfbp4hxrfcvvauhl4sgirwtmy	----	05_Miladinovic_Photoionization theories_final Kragujevac J. Sci. 38 (2016) 53-62. UDC 539.1 THEORETICAL AND EXPERT SYSTEM APPROACH TO PHOTOIONIZATION THEORIES Ivan D. Petrović, Violeta M. Petrović*, Tatjana B. Miladinović Department of Physics, Faculty of Science, Kragujevac University, Republic of Serbia *Corresponding author; E-mail: violeta.petrovickg@gmail.com (Received April 8, 2016) ABSTRACT. The influence of the ponderomotive and the Stark shifts on the tunneling transition rate was observed, for non-relativistic linearly polarized laser field for alkali atoms, with three different theoretical models, the Keldysh theory, the Perelomov, Popov, Terent’ev (PPT) theory, and the Ammosov, Delone, Krainov (ADK) theory. We showed that aforementioned shifts affect the transition rate differently for different approaches. Finally, we presented a simple expert system for analysis of photoionization theories. Keywords: ionization rate, expert system, tunneling ionization. INTRODUCTION Non-linear ionization of atoms and molecules in an intense laser field is a currently of interest problem in atomic physics. Theoretical explanation of ionization processes begun with Keldysh, who showed in his paper (KELDYSH, 1964) that the nature of multiphoton and tunneling ionization is essentially the same, and presented the two limiting cases of non-linear photoionization. These two mechanisms are distinguished by the value of what today is widely known as Keldysh parameter , where is unperturbed ionization potential for considered system, is laser frequency and is the field strength in . Atomic units were used throughout this paper unless otherwise indicated. If the multiphoton ionization takes place, while the tunneling ionization prevails when . After the appearance of the Keldysh work, its results were refined by Perelomov, Popov and Terent’ev in PERELOMOV et al. (1966). Next, Ammosov, Delone and Krainov, based on PERELOMOV et al. (1966), derived the ADK theory (AMMOSOV et al. 1986) for the case of the tunneling ionization for arbitrary complex atoms and atomic ions. The intent of this paper was to compare an influence of the ponderomotive potential and the Stark shift of the initial binding state of alkali atoms on the transition rate, as well as the influence of the spatial distribution of the laser beam shape. This was accomplished by analyzing three commonly used theoretical models for field ionization, Keldysh, PPT and ADK. 54 THEORETICAL FRAMEWORK The Keldysh theory provides a commonly accepted framework for a quantitative analysis of ionization processes. With the exponential accuracy on the incident laser field, the Keldysh transition rate (KELDYSH, 1964) has the form: , (1) where Z is the charge of the atomic residue and . Perelomov, Popov and Terent’ev (PERELOMOV et al. 1966) suggested a more accurate model for calculating the photoionization rate for an atom and atomic residual ions. This model takes into account the Coulomb interaction between outgoing photoelectron and ion. The PPT theory is valid for arbitrary values of the Keldysh parameter . It agrees quite well with the experimental results in both ionization regimes, multiphoton and tunneling (LIN et al. 2004). Based on the PPT theory, the ionization rate in linear laser field is given (VOLKOVA et al. 2006) by the formula: . (2) The most often used tunneling theory is the ADK theory, which is in excellent agreement with experiments in noble gases and small molecules. The ADK theory has extended the PPT theory for the tunneling ionization rate of arbitrary complex atoms and atomic ions. Transition rate formula for linearly polarized laser field in cases of zero momentum is: (3) In the case of non-zero initial momentum of the ejected photoelectrons, the ADK formula has the following form (RISTIĆ et al. 2009): , (4) where is the effective principal quantum number, and is the initial momentum of the ejected photoelectrons (LANDAU and LIFSHITZ, 1991; BAUER, 2006). In all of these theories the exponential dependence on field strength and the ionization potential is emphasized. However, at higher intensities the atomic level structure and the ionization process become increasingly influenced by the laser irradiation through two well- known effects, the ponderomotive potential and the Stark shifts. In process of laser induced ionization, a bound electron must acquire energy equal to its ionization potential increased by two additional terms, one which corresponds to the ponderomotive potential (DELONE and KRAINOV, 1998) and second to the Stark shift (DELONE and KRAINOV, 2000), where is the static polarizability of the atom (SCHWERDTFEGER, 2014). Accordingly, the field free ionization potential has the following form (VOLKOVA et al. 2011): , where denotes the shifted, effective ionization potential. Putting in Eq. 1, the Keldysh tunneling ionization rate is found to be: , (5) 55 where superscript denotes both the ponderomotive and the Stark shifts. Repeating the same procedure, and were obtained: , (6) . (7) Additionally, the laser beam shape is regarded as Lorentzian (SHCHATSININ IHAR, 2009; SHEALLY and HOFFNAGLE, 2006; PETROVIĆ and MILADINOVIĆ, 2014): . (8) where is the axial coordinate that is normal to the direction of the light ray, (ZHANG, 2010) and is called the laser beam waist, which represents the smallest spot size realized at . PETROVIĆ and MILADINOVIĆ (2014) have used Lorentzian and Gaussian beam shape and showed that both give satisfying results. Here we have chosen Lorentzian. Based on Eq. 8, the effective ionization potential can be written in the following form: . According to Eq. 12 and the inline equation for , Eqs. 5, 6, and 7 can be rewritten in the following form: , (9) , (10) And . (11) where L indicates the Lorentzian distribution. RESULT AND DISCUSSION In this paper were theoretically analyzed the tunneling transition rates, for , in the field of intensity , for linearly polarized Ti:sapphire laser, , ( ). The single ionized, , alkali atom of potassium, K is studied. The obtained results were compared for three different ionization models. Also, the influence of spatial distribution on the transition rate is discussed. We started from the transition rates, , and for the field-free ionization potential (see Eqs. 1, 2 and 4), for generally assumed laser beam shape, without any specifications. As a result, the following theoretical curves were obtained (see Fig. 1). 56 Figure 1. The ionization rates, , , vs laser field intensity, . As can readily be seen, the field-free transition rate curves exhibit a significantly different behavior for the Keldysh, PPT and ADK theory. Keldysh curve almost linearly increases, while the other two reach some maximal value and then decrease. It is also obvious that PPT curve has the higher values of the transition rate compared to the ADK curve, for the same field intensity. In order to perform a detailed analysis, the ponderomotive potential and the Stark shift were included in the equations for the transition rate (Eqs. 5, 6, 7). Here and in the following we used the notation: superscript denotes an included ponderomotive potential, while - ponderomotive potential and Stark shift. Fig. 2 demonstrates the dependence of the transition rates (with corrected ionization potential included) on the field intensity. Figure 2. a) Left, the ionization rates, , , vs laser field intensity, ; b) Right, the ionization rates, , , vs laser field intensity . All curves on Fig. 2 have the lower magnitude then on Fig. 1. The physical reason for this is the inclusion of the ponderomotive potential and the Stark effect which affect the ground state. In the intensity range under consideration, PPT curve is several orders of magnitude above the ADK, while Keldysh curve constantly increases in whole intensity range and so at the higher intensities have the higher values then PPT and ADK rates. It can be seen, by comparing left and right plot, that all curves have similar shape, but at different values of intensity. PPT and ADK curves have the maximum which is for the , i.e. shifted to the lower field intensity when compared to , i.e. . Also, the corresponding 57 maximal values for and are reduced compared to and which can be expected because the ionization energy is shifted to higher values. Based on everything mentioned, it follows that inclusion of the additional parameters into formulas for the transition rates leads to the significant changes in the physical picture. Next were plotted the curves for the Keldysh and the PPT transition rates, first without any corrections and then when all corrections are taken into account, respectively. Figure 3: a) Left, the ionization rates , , vs laser field intensity; b)Right, the ionization rates , vs laser field intensity. From Fig. 3, left plot, follows that all Keldysh curves, , , have the same shape. Inclusion of the ponderomotive potential and the Stark shift decreases only the transition rate at the same field intensity. One would expect that the K transition rate curve is significantly influenced by the Stark shift, but the graph shows that the decreasing is very low, which isn’t quite in accordance with theoretical prediction. Oppositely, right plot in Fig. 3 shows just the expected behavior of the PPT curve, as is significantly influenced by polarizability. Also, it is obvious that corrected , differ strongly from the uncorrected , and that with increasing laser field intensity the rates , approach zero. If the ponderomotive potential and the Stark shift of the energy of the ground state of an atom are both taken into account, the curves are shifted to the right, i.e. in the direction of lower field intensities. Laser interaction with atoms depends on several parameters related with laser sources. One of these parameters is a spatial distribution of a laser beam. So, it is of interest how exactly the distribution influences the transition rates. In Fig. 4 are shown the compared results for Keldysh rates and (see Eq. 1 and Eq. 9), and PPT rates and (see Eq. 2 and Eq. 10). Mark “L” in the superscript indicates the Lorentzian spatial distribution. This was accomplished for the corrections included. For the ADK theory the spatial dependence can be found in PETROVIĆ and MILADINOVIĆ (2014). The following figures were obtained. 58 Figure 4. a) Left, and vs laser field intensity ; b) Right, and vs laser field intensity . From Fig. 4 the influence of the spatial dependence is obvious. Both of Keldysh curves increase with increasing of the laser field intensity but this increase is much significant for . As a consequence has higher values at , compared to for the same field intensity. The main conclusion that follows from the comparison of the curves and is that the maximum for both transition rates is the same, but the curve reaches the maximum on the higher field intensity i.e. the maximum is shifted to the right. Both curves are asymmetric around the value of the field intensity on which and , respectively, have the maximum value, but this asymmetry is much more prominent for . Both curves then approach zero with further increasing of the laser field intensity. From aforementioned can be readily concluded that the transition rates are quite sensitive on the correction of the field free potential (potential without any corrections) and the laser beam shape, so the both effects must be taken into account. Many useful data and interesting relations and dependences could be obtained from observed transition rates (on which field intensity the transition rate has the maximal value, how the Keldysh parameter influences on the transition rate, etc.). In order to provide a more detailed outlook on obtained results, an expert system logic was implemented on the aforementioned analyses, which resulted in a development of an expert system. EXPERT SYSTEM In this section we give a short review about the expert system developed for analysis of photoionization theories with respect on the issues described above. An expert system (sometimes referred to as knowledge-based system) is a computer software that emulates the decision-making ability of a human expert (MANDAL et al. 2013). Beside the specific expert systems building tools, such as I2, CLIPS, PROLOG, LISP, an expert system can also be developed by writing programs using certain programming languages, such as Fortran, Pascal, C++ and Visual Basic (MUQUEEM, 2014). The expert system was built on Visual Basic, while the accuracy of ES reasoning was partially checked through ESBT I2+. For testing of ES the available theoretical and experimental results were used. We imitated the backward chaining logic where the system starts from a specific goal and attempts to satisfy the preconditions necessary for obtaining it (TILOTMA SHARMA et al. 2012). 59 Interactions between the users and the system are supported through a friendly graphical user interface running under Windows environment. The system starts by initializing the first interface screen and prepare the “questions” for the user who must input some necessary, initial data such as laser wavelength , chemical element observed, range of laser field intensity , the ion charge Z, and Keldysh parameter, . All the listed data must be entered. Fig. 5 (left plot) shows the initial form of GUI. Figure 5: a) Left, the initial form, b) Right, an error message After the user fills appropriate boxes, the ES then checks the data entered. If all data are valid, the ES continues to work. In an opposite case the “Error message” appears (right side on Fig. 5), so the user should re-enter a relevant data. For example, a valid value of the Keldysh parameter is less than as the observed ionization process is that of tunneling ionization. The button “Clear all” deletes all entered data. The initial form completely corresponds to the structure and has the function of the initial rule. Initial rule is first activated rule in an ES. In characteristic “if then” logic, the corresponding initial rule has the following form: RULE Initial IF Initial conditions AND Wavelength AND Element AND Ion charge AND Intensity range AND Keldysh parameter THEN Initial data ELSE Repeated input AND FORGET Wavelength AND FORGET Element AND FORGET Ion charge AND FORGET Intensity range AND FORGET Keldysh parameter AND CYCLE The “IF“ part of the initial rule contains premises bound with the logic operator “AND”. In order to activate the rule all premises must be true (LUCAS and VAN DER GAAG, 1991). It should be noted that we restricted ourselves only on above described issues. Otherwise, the ES can, based on the value of Keldysh parameter and intensity range conclude what ionization process is dominant and, based on that, perform appropriate analysis. The “Next” button loads the form “Transition rate” with displayed various options, Fig. 6 (on left side): 60 Figure 6. The form “Transition rate” Active buttons enable a calculation of transition rates by different theories, without or with corrections (see Section 2). Obtained results are then written in the table “Transition rates” that has a following structure: Table 1. The structure of table for writing data. Additionally, the form contains the button “Analysis” that calls a form Analysis shown on right side of Fig. 6, that offers the procedures necessary to determine on which laser field intensities the maximal values of the transition rates appear. “Back” option enables a return to the previous form; also the user can select some other options. Each button on the form activates a subroutine for the analysis of that particular transition rate. Obtained results are then written in appropriate columns of the “Transition rates” table. If the case when all of the transition rates are calculated, the following, completely filled table, is obtained: Table 2. A filled table, with fields that are given appropriate values during the work of the ES. 61 There is not the only given return from the ES. The user also can decide in which moment he wants an analysis to finish. The ending rule given below illustrates that. RULE End IF Initial data AND Transition rates OR Analysis THEN Analysis completed The rule will be activated if all of needed premises linked through the logic “AND” operator are satisfied and also if at least one premise linked through the logic “OR” operator is satisfied. CONCLUSION The ionization rates of alkali atoms by three theories, Keldysh, PPT and ADK for the case of non-relativistic linearly polarized laser field were studied. It was shown that the ponderomotive potential and the Stark shift influence the transition rates, even in the case of alkali atoms with relatively small polarizability. It was also shown that an expert system can be used for an analysis of the transition rate. The described system can be improved in several ways and this will be further pursued in forthcoming papers. Acknowledgements We are grateful to the Serbian Ministry of Education and Science for financial support through Project 171020. References [1] AMMOSOV, V.M., DELONE, N.B., KRAINOV, V.P., Tunnel ionization of complex atoms and atomic ions by an alternating electromagnetic field, Soviet Physics JETP 64 (6) (1986) 1191- 1194. [2] BAUER, D. Theory of intense laser-matter interaction, Max-Planck Inst., Heidelberg, Germany, (2006) 58, http://www.physik.uni-rostock.de/fileadmin/Physik/Bauer/tilmi.pdf [3] DELONE, N.B., KRAINOV, V.P., Multiphoton Processes in Atoms, 2nd edition, Springer (2000). [4] DELONE, N.B., KRAINOV, V.P., Tunneling and barrier-suppression ionization of atoms and ions in a laser radiation field, Physics Uspekhi 41 (5) (1998) 469-485. [5] KELDYSH, L.V., Ionization in the field of a strong electromagnetic wave, Soviet Physics JETP 20 (5) (1964) 1945-1957. [6] LANDAU, L.D., LIFSHITZ, E.M., Quantum Mechanics: Non-Relativistic Theory, 3rd ed. Pergamon, Oxford, (1991) 295-298., https://issuu.com/sonocontentodite/docs/landau-l.d.--- lifschitz-e.m.--vol.-1---mechanics-3 [7] LIN, S.H., VILLAEYS, A.A., FUJIMURA Y., Advances in Multi-photon Processes and Spectroscopy 16, Singapore: World Scientific Publishing Co. Pte. Ltd, (2004) 249 pp, 62 [8] LUCAS, P.J.F., VAN DER GAAG, L.C., Principles of Expert Systems, Centre for Mathematics and Computer Science, Amsterdam, Addison-Wesley (1991). [9] MANDAL S., CHATTERJEE S., NEOGI B., Diagnosis and troubleshooting of computer faults based on expert system and artificial intelligence, International Journal of Pure and Applied Mathematics 83 (5) (2013) 717-729, [10] MUQUEEM, S., Expert system application in library, Knowledge Librarian - An International Peer Reviewed Bilingual E-Journal of Library and Information Science, 1 (2) (2014) 168-175. [11] PERELOMOV, A.M., POPOV, V.S., TERENT’EV, M.V., Ionization of atoms in an alternating electric field, Soviet Physics JETP, 23 (5) (1966) 924-934. [12] PETROVIĆ, V.M., MILADINOVIĆ, T.B., Influence of the spatial and temporal distribution of an incident laser beam profile on the energy distribution of ionized photoelectrons, JETP 119 (4) (2014) 651-656. [13] RISTIĆ, V.M., MILADINOVIĆ, T.B., RADULOVIĆ, M.M., Calculating ionization transition rate for circularly polarized fields, including non-zero initial momentum, Acta Phys. Pol. A 116 (4) (2009) 504-506. [14] SHCHATSININ IHAR, Free clusters and free molecules in strong, shaped laser fields, Universität Berlin, PhD Thesis (2009) 51-71. [15] SCHWERDTFEGER, P., Table of experimental and calculated static dipole polarizabilities for the electronic ground states of the neutral elements (in atomic units) (2014), http://ctcp.massey.ac.nz/Tablepol2014.pdf [16] SHEALY, D.L., HOFFNAGLE, J.A., Laser beam shaping profiles and propagation. Applied Optics 45 (21) (2006) 5118-5131 DOI: 10.1364/AO.45.005118 [17] TILOTMA SHARMA, NAVNEET TIWARI, DEEPALI KELKAR, Study of difference between forward and backward reasoning, International Journal of Emerging Technology and Advanced Engineering 2 (10) (2012) 271-273. [18] VOLKOVA, E.A., GRIDCHIN, V.V., POPOV, A.M., TIKHONOVA, O.V., Tunneling ionization of a hydrogen atom in short and ultrashort laser pulses, JETP 102 (1) (2006) 40-52. [19] VOLKOVA, E.A., POPOV, A.M., TIKHONOVA, O.V., Ionization and stabilization of atoms in high- intensity, low-frequency laser field, JETP 113 (3) (2011) 394-406. [20] ZHANG, L., Intensity spatial profile analysis of a Gaussian laser beam at its waist using an optical fiber system, Chinese Physics Letters 27 (5) (2010) 054207-3 
work_fyjswziacfg3pe4tfkp35fqd3q	----	Distance difference and linear programming nonparallel plane classifier Distance difference and linear programming nonparallel plane classifier Qiaolin Ye a,⇑, Chunxia Zhao a, Haofeng Zhang a, Ning Ye b a School of Computer Science and Technology, Nanjing University of Science and Technology, Nanjing, China b School of Information Technology, Nanjing Forestry University, Nanjing, China a r t i c l e i n f o Keywords: GEPSVM LSTSVM TWSVM Standard eigenvalues Feature selection Input features Kernel functions a b s t r a c t We first propose Distance Difference GEPSVM (DGEPSVM), a binary classifier that obtains two nonparallel planes by solving two standard eigenvalue problems. Compared with GEPSVM, this algorithm does not need to care about the singularity occurring in GEPSVM, but with better classification correctness. This formulation is capable of dealing with XOR problems with different distribution for keeping the genuine geometrical interpretation of primal GEPSVM. Moreover, the proposed algorithm gives classification cor- rectness comparable to that of LSTSVM and TWSVM, but with lesser unknown parameters. Then, the reg- ularization techniques are incorporated to the TWSVM. With the help of the regularized formulation, a linear programming formation for TWSVM is proposed, called FETSVM, to improve TWSVM sparsity, thereby suppressing input features. This means FETSVM is capable of reducing the number of input fea- tures, for linear case. When a nonlinear classifier is used, this means few kernel functions determine the classifier. Lastly, this algorithm is compared on artificial and public datasets. To further illustrate the effectiveness of our proposed algorithms, we also apply these algorithms to USPS handwritten digits. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Eigenvalue based techniques are attractive for the classification of very large sparse datasets (Guarracino, Cifarelli, Seref, & Parda- los, 2007) such as generalized proximal SVM (GEPSVM for short) (Mangasarian & Wild, 2006). GEPSVM obtains each of the nonpar- allel planes by solving the eigenvector corresponding to a smallest eigenvalue of a generalized eigenvalue problem, such that each plane is as close as possible to the samples for its class and mean- time as far as possible from the samples for the other classes (Man- gasarian & Wild, 2006). The edges of two-class GEPSVM lie in its lower computational complexity and its better classification per- formance in terms of solving XOR problems with respect to stan- dard SVM that find one plane that separates the two classes. In Mangasarian and Wild (2006), Mangasarian et al. presented a sim- ple ‘‘cross planes’’ example that is a generalization of the XOR example, which indicated the effectiveness of GEPSVM over PSVM and SVM. Fig. 1 in Mangasarian and Wild (2006) demonstrates GEPSVM has classification correctness of 100% in XOR case. Re- cently, a lot of GEPSVM-based algorithms have been proposed. To improve the generalization of GEPSVM, Jayadeva et al. proposed Fuzzy GEPSVM (FGEPSVM) given its multi-category formulation. In 2007, Guarracino et al. (2007) introduced a new regularization technique to GEPSVM for reducing the time complexity of GEPSVM, but with two unknown parameters in linear case. These algorithms obtain two planes by solving generalized eigenvalue problems as GEPSVM does. However, for the symmetric matrices occurring in these algorithms such as H and M in the formulation (5) and (6), if both are semi-positive, an ill-defined operation will be obtained. Moreover, these algorithms weaken the genuine geo- metrical interpretation of the nonparallel plane classifier due to the adoption of regularization term that improves their generalization. Recently, a twin SVM algorithm (TWSVM for short), proposed by Jayadeva et al., was published in TPAMI (Jayadeva & Chandra, 2007). This algorithm, which is in the spirit of GEPSVM, obtains two planes by solving two smaller quadratic programming prob- lems (QPPs) than that of the standard SVM. Experimental results show the effectiveness of TWSVM over SVM and GEPSVM (Arun Kumar & Gopal, 2009; Jayadeva & Chandra, 2007). TWSVM takes O(1/4m3) operations which is 1/4 of standard SVM, whereas, GEPSVM takes O(1/4n3). Here, m is the number of training samples, n is the dimensionality and m � n (Arun Kumar & Gopal, 2009; Jay- adeva & Chandra, 2007). Obviously, GEPSVM is by far faster than TWSVM. To reduce the time complexity and keep the effectiveness of the twin SVM classifier, some scholars proposed its least squares version (LSTSVM for short) in 2009 (Arun Kumar & Gopal, 2009; Ghorai, Mukherjee, & Dutta, 2009). In fact, LSTSVM determines two nonparallel planes by solving two PSVM-type (Fung & Manga- sarian, 2001) problems. Compared with TWSVM, LSTSVM has les- ser computational time due to the fact that it only solves two systems of linear equations instead of two QPPs as for TWSVM. TWSVM and LSTSVM however, also lose the genuine geometrical interpretation of the nonparallel plane classifier. GEPSVM is 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.01.131 ⇑ Corresponding author. E-mail address: yeqiaolin65620868@163.com (Q. Ye). Expert Systems with Applications 38 (2011) 9425–9433 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.01.131 mailto:<xml_chg_old>yeqiaolin656208686@163.com</xml_chg_old><xml_chg_new>yeqiaolin65620868@163.com</xml_chg_new> http://dx.doi.org/10.1016/j.eswa.2011.01.131 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa proposed to solve the complex examples which are difficult classi- fication cases for typical linear classifiers just as XOR example does (Mangasarian & Wild, 2006). Each of the planes obtained by GEPSVM is as close as possible to the samples for its class and meantime as far as possible from the samples for the other classes (Mangasarian & Wild, 2006). However, TWSVM requires each of planes obtained to be as close as possible to the samples for its class and meantime at a distance of at least 1 from the samples for the other classes (Jayadeva & Chandra, 2007). LSTSVM requires each of the planes to be as close as possible to the samples for its class and meantime at a distance of 1 from the samples for the other classes. In intuition, when handling XOR examples of differ- ent distribution, TWSVM and LSTSVM may yield poor classification performance due to the difference from the optimization criterion of GEPSVM, although they can obtain good classification perfor- mance on UCI datasets due to the use of the loss function. More- over, another flaw of TWSVM and LSTSVM is that two penalty parameters are introduced to their objective functions instead of one regularization parameter as for GEPSVM. Undoubtedly, this will lead to the difficulty of parameter selection. In addition, when there are many noise variables, the 1-norm SVM (Zou, 2007; Zhou, Zhang, & Jiao, 2002) has advantages over the 2-norm SVM because the former is capable of generating sparse solutions that make the classifier easier to store and faster to compute. However, these GEPSVM-based algorithms, including GEPSVM, cannot generate very sparse solutions, even if we give their 1-norm formulations as in 1-norm SVM (Zou, 2007). This is so because the direction wi and threshold ri that determine the i th separating planes combines with the input samples. In this paper, we first propose a new but fast algorithm, termed as Difference GEPSVM (DGEPSVM). DGEPSVM need not consider the singularity occurring in GEPSVM due to the use of a similar for- mulation to the MMC (Jiang & Zhang, 2004). We show that the solution of DGEPSVM reduces to solving two simple eigenvalue problems. This property determines DGEPSVM is fast and at least comparable to GEPSVM. Moreover, DGEPSVM can deal with XOR examples with different distribution because it keeps the genuine geometrical interpretation of GEPSVM. Then, we further propose a feature selection algorithm for TWSVM, called FETSVM. This pro- posed algorithm can overcome such a flaw, that is, GEPSVM and other GEPSVM-based algorithms cannot generate the very sparse solutions. Lastly, the two algorithms are compared on artificial and UCI datasets. We also go onto illustrate their effectiveness for USPS handwritten digits application. Given four facts: (1) DGEPSVM need not care about the singu- larity occurring in GEPSVM and performs better in classification correctness than GEPSVM; (2) DGEPSVM surpasses TWSVM and LSTSVM in terms of solving XOR examples with different distribu- tion and gives comparable classification correctness on standard datasets; (3) DGEPSVM has lesser unknown parameters than TWSVM and LSTSVM; and (4) FETSVM performs faster than TWSVM and suppresses input features as well as giving compara- ble classification correctness. 2. Related work 2.1. Generalized Proximal Support Vector Machines (GEPSVM) (Mangasarian & Wild, 2006) Given m training points in n dimension input space Rn, denoted by the m1 � n matrix A belonging to class 1 and the m2 � n matrix B belonging to class-1, where m2 + m1 = m. The main purpose of GEPSVM is to find two nonparallel hyper- planes in n-dimension space, i.e., xT w1 � b1 ¼ 0; xT w2 � b2 ¼ 0 ð1Þ where (wi, bi) 2 (Rn � R) (i = 1, 2). This algorithm requires each plane to be as close as possible to the samples for its class and as far as possible from the samples for the other classes at the same time. Suppose (wi, bi) – 0, binary GEPSVM classifiers can be written as the optimization formulation as follows: ðGEPSVM1Þ Min ðw1;b1Þ – 0 ðAw1 �e1 b1ÞTðAw1 �e1 b1Þþd w1 b1 � �� �T w1 b1 � �� � ðBw1 �e2 b1Þ TðBw1 �e2 b1Þ ð2Þ ðGEPSVM2Þ Min ðw2;b2Þ – 0 ðBw2 �e2 b2ÞTðBw2 �e2 b2Þþd w2 b2 � �� �T w2 b2 � �� � ðAw2 �e1 b2ÞTðAw2 �e1 b2Þ ð3Þ where d is a regularization constant. The formulation (2) enables GEPSVM to obtain the plane, which is closest to the points for set +1 and furthest from set �1, and (3) enables GEPSVM to obtain the plane which is closest to the points for set �1 and furthest from the points for set +1. Let G ¼ ET E þ dI; H ¼ FT F L ¼ FT F þ dI; M ¼ ET E z1 ¼ w1 b1 � � ; z2 ¼ w2 b2 � � ð4Þ where E = [A � e1], F = [B � e2], E, F 2 Rn+1. The optimization problem (2) and (3) become: ðGEPSVM1Þ Min z1 –0 zT1 G z1 zT1 H z1 ð5Þ ðGEPSVM2Þ Min z2 –0 zT2 L z2 zT2 M z2 ð6Þ Using the well-known properties of Rayleigh quotient, we can obtain the solutions of (5) and (6) by solving the following two generalized eigenvalue problems: Gz1 ¼ k1Hz1; Lz2 ¼ k2 Mz2; zi – 0; i ¼ 1; 2: ð7Þ It can be seen from (2) and (3), Tikhonov regularization (Tikhonov & Arsen, 1977) is applied to each GEPSVM pair. This can improve the generalization of GEPSVM due to the fact that the regularization term is used for penalty. 2.2. TWIN Support Vector Machines (TWSVM) (Jayadeva & Chandra, 2007) In 2007, Jayadeva et al. proposed a stand-alone algorithm for binary classification, termed as Twin SVM (TWSVM) (Jayadeva & Chandra, 2007), which is in the spirit of GEPSVM. However, TWSVM has different formulation from GEPSVM. This algorithm obtains two nonparallel planes by solving two SVM-type problems. Experimental results show TWSVM outperforms GEPSVM and standard SVM, in terms of classification correctness. TWSVM can be written as follows: ðTWSVM1Þ Min 1 2ðAw1 � e1 b1Þ TðAw1 þ e1b1Þþ C1eT2n s:t: �ðBw1 � e2 b1Þþ n P e2; n P 0 ð8Þ ðTWSVM2Þ Min 1 2ðBw2 � e2b2Þ TðBw2 � e2b2Þþ C2eT1n s:t: ðAw2 � e1 b2Þþ n P e1; n P 0 ð9Þ where C1 and C2 are two penalty coefficients. From (8) and (9), we find that only constraints of the other class appear. The objective function does not sum up error over patterns of both the classes. These features show TWSVM is effective on skewed or unbalanced datasets. This may be a reason results in a better classification why 9426 Q. Ye et al. / Expert Systems with Applications 38 (2011) 9425–9433 https://isiarticles.com/article/25248 
work_g4xr6mfzkbcorko2azocjqkic4	----	NASA Technical Reports Server (NTRS) 
work_g62e2dobdngjhn2zvfq6rzqmey	----	An artificial bee colony approach for clustering An artificial bee colony approach for clustering Changsheng Zhang a,*, Dantong Ouyang b, Jiaxu Ning c a College of Information Science & Engineering, Northeastern University, Shenyang 110819, PR China b Key Laboratory of Symbol Computation and Knowledge Engineering of the Ministry of Education, Changchun 130012, PR China c Institute of Grassland Science Northeast Normal University, PR China a r t i c l e i n f o Keywords: Clustering Meta-heuristic algorithm Artificial bee colony K-means a b s t r a c t Clustering is a popular data analysis and data mining technique. In this paper, an artificial bee colony clustering algorithm is presented to optimally partition N objects into K clusters. The Deb’s rules are used to direct the search direction of each candidate. This algorithm has been tested on several well-known real datasets and compared with other popular heuristics algorithm in clustering, such as GA, SA, TS, ACO and the recently proposed K–NM–PSO algorithm. The computational simulations reveal very encouraging results in terms of the quality of solution and the processing time required. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction Clustering is an important problem that must often be solved as a part of more complicated tasks in pattern recognition, image analysis and other fields of science and engineering. Clustering procedures partition a set of objects into clusters such that objects in the same cluster are more similar to each other than objects in different clusters according to some predefined criteria. The exist- ing clustering algorithms can be simply classified into the follow- ing two categories: hierarchical clustering and partitional clustering (Sander, 2003.). Hierarchical clustering operates by par- titioning the patterns into successively fewer structures. Since it is not the subject of this study we will not mention it in detail. Partitional clustering procedures typically start with the patterns partitioning into a number of clusters and divide the patterns by increasing the number of partitions. The most popular class of partitional clustering methods is the center-based clustering algorithms. K-means has been used as a popular center-based clustering method due to its simplicity and efficiency, with linear time com- plexity. However, K-means has the shortcomings of depending on the initial state and converging to local minima (Selim & Ismail, 1984). In order to overcome these problems, many heuristic clus- tering algorithms have been introduced. A genetic algorithm based method to solve the clustering problem was proposed by Mualik and Bandyopadhyay (2002) and experimented on synthetic and real-life datasets to evaluate its performance. Krishna and Murty (1999) proposed a novel approach called genetic K-means algorithm for clustering analysis which defines a basic mutation operator specific to clustering called distance-based mutation. It has been proved that GKA converge to the best-known optimum through using the theory of finite Markov chain. A simulated annealing approach for solving the clustering problem is proposed by Selim and Al-Sultan (1991). The parameters of the algorithm were discussed in detail and it has been proved theoretically that a clustering problem’s global solution can be reached. Sung and Jin (2000) proposed a tabu search based heuristic for clustering. Two complementary functional procedures, called packing and releasing procedures were combined with the tabu search. Over the last decade, modeling the behavior of social insects, such as birds, ants, and bees for the purpose of search and optimi- zation has become an emerging area of swarm intelligence and suc- cessfully applied to clustering. An ant colony clustering algorithm for clustering is presented by Shelokar, Jayaraman, and Kulkarni (2004). The algorithm employs distributed agents who mimic the way real ants find a shortest path from their nest to food source and back. Its performance was compared with GA, tabu search, and SA. The particle swarm optimization which simulates bird flocking was used for clustering by Kao, Zahara, and Kao (2008). In order to improve its performance further, the PSO algorithm is hybridized with K-means and Nelder–Mead simplex search meth- od. Its performance is compared with GA (Murthy & Chowdhury, 1996) and KGA (Bandyopadhyay & Maulik, 2002) algorithm. Honey-bees are among the most closely studied social insets. Their foraging behavior, learning, memorizing and information sharing characteristics have recently been one of the most interest- ing research areas in swarm intelligence (Teodorovic et al., 2006). Recently, Karaboga and Basturk (2008) have described an artificial bee colony (ABC) algorithm based on the foraging behavior of honey-bees for numerical optimization problems. They have compared the performance of the ABC algorithm with those of other well-known modern heuristic algorithms such as genetic algorithm, differential evolutional algorithm and particle swarm 0957-4174/$ - see front matter � 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2009.11.003 * Corresponding author. Tel.: +86 0431 85166487. E-mail address: zcs820@yahoo.com.cn (D. Ouyang). Expert Systems with Applications 37 (2010) 4761–4767 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2009.11.003 mailto:zcs820@yahoo.com.cn http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa optimization algorithm for unconstrained optimization problems. In this work, ABC algorithm is extended for solving clustering prob- lems. The performance of the algorithm has been tested on a vari- ety of data sets provided from several real-life situations and compared with several other proposed clustering algorithms. This paper is organized as follows. In Section 2, we discussed the clus- tering analysis problems. The ABC algorithm and the ABC algo- rithm adapted for solving clustering problems are introduced in Section 3. Section 4 will present experimental studies that show that our method outperforms some other methods. Finally, Section 5 summarizes the contribution of this paper along with some fu- ture research directions. 2. The clustering problem Let O ¼fo1; o2; . . . ; ong be a set of n objects and let Xn�p be the profile data matrix, with n rows and p columns. Each ith objects is characterized by a real-value p-dimensional profile vector xiði ¼ 1; . . . ; nÞ, where each element xij corresponds to the jth real-value feature (j = 1, . . . , p) of the ith object (i = 1, . . . , n). Given X n�p , the goal of a partitional clustering algorithm is to determine a partition G ¼fC1; C2; . . . ; Ckg (i.e., Cg –U; 8g; Cg\ Ch ¼ U; 8g–h; [kg¼1Cg ¼ OÞ such that objects which belong to the same cluster are as similar to each other as possible, while objects which belong to different clusters are as dissimilar as possible. For this, a measure of adequacy for the partition must be defined. A popular function used to quantify the goodness of a partition is the total within-cluster variance or the total mean-square quanti- zation error (MSE) (Güngör & Ünler, 2007 ) which is defined as follows: Perf ðO; GÞ¼ Xn i¼1 Min koi � Clk 2jl ¼ 1; . . . k n o ð1Þ Where koi � Clk denotes the similarity between object oi and center Cl . The most used similarity metric in clustering procedure is Euclidean distance which is derived from the Minkowski metric. dðx; yÞ¼ Xm i¼1 jxi � yij r !1=r ) dðx; yÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXm i¼1 ðxi � yiÞ 2 q ð2Þ In this study we will also use Euclidean metric as a distance metric. The clustering problem is to find the partition G* that has optimal adequacy with respect to all other feasible solutions G ¼fG1; G2; . . . ; GNðn;kÞg (i.e., Gi–Gj; i – jÞ where Nðn; kÞ¼ 1 k! Xk g¼0 ð�1Þg k g � � ðk � gÞn It is the number of all feasible partitions. It has been shown that the clustering problem is NP-hard when the number of clusters exceeds three (Brucker, 1978). 3. Artificial bee colony based clustering 3.1. Honey bee modeling (Karaboga & Basturk, 2008) The minimal model of forage selection that leads to the emer- gence of social intelligence of honey bee swarms consists of three essential components: food sources, employed foragers and unemployed foragers, and two leading modes of the behavior, recruitment to a nectar source and abandonment of a source, are defined (Karaboga, 2005). A food source value depends on many factors, such as its proximity to the nest, richness or concentration of energy and the ease of extracting this energy. The employed foragers are associated with particular food sources, which they are currently exploiting or are ‘‘employed”. They carry with them information about these food sources and share this information with a certain probability. There are two types of unemployed for- agers, scouts and onlookers. Scouts search the environment sur- rounding the nest for new food sources, and onlookers wait in the nest and find a food source through the information shared by employed foragers. In ABC algorithm (Basturk & Karaboga, 2006; Karaboga & Bas- turk, 2008), the colony of artificial bees consists of three groups of bees: employed bees, onlookers and scouts. A food source repre- sents a possible solution to the problem to be optimized. The nec- tar amount of a food source corresponds to the quality of the solution represented by that food source. For every food source, there is only one employed bee. In other words, the number of em- ployed bees is equal to the number of food sources around the hive. The employed bee whose food source has been abandoned by the bees becomes a scout. As other social foragers, bees search for food sources in a way that maximizes the ration E/T where E is the energy obtained and T is the time spent for foraging. In the case of artificial bee swarms, E is proportional to the nectar amount of food sources discovered by bees. In a maximization problem, the goal is to find the maxi- mum of the objective function FðhÞ; h 2 RP . Assume that hi is the position of the ith food source; FðhiÞ represents the nectar amount of the food source located at hi and is proportional to the energy EðhiÞ. Let PðcÞ¼ fhiðcÞji ¼ 1; 2; . . . ; Sg (c: cycle, S: number of food sources being visited by bees) represent the population of food sources being visited by bees. As mentioned above, the preference of a food source by an on- looker depends on the nectar amount FðhÞ of that food source. As the nectar amount of the food source increases, the probability with the preferred source by an onlooker bee increases proportion- ally. Therefore, the probability with the food source located at hi will be chosen by a bee can be calculated as Pi ¼ FðhiÞPS k¼1 FðhkÞ ð3Þ After watching the dances of employed bees, an onlooker bee goes to the region of food source located at hi by this probability and determines a neighbor food source to take its nectar depending on some visual information, such as signs existing on the patches. In other words, the onlooker bee selects one of the food sources after making a comparison among the food sources around hi. The position of the selected neighbor food source can be calculated as hiðc þ 1Þ¼ hiðcÞ� /iðcÞ. /iðcÞ is a randomly produced step to find a food source with more nectar around hi . /ðcÞ is calculated by taking the difference of the same parts of hiðcÞ and hkðcÞ (k is a randomly produced index) food positions. If the nectar amount Fðhiðc þ 1ÞÞ at hiðc þ 1Þ is higher than that at hiðcÞ, then the bee goes to the hive and share her information with others and the position hiðcÞ of the food source is changed to be hiðc þ 1Þ, otherwise hiðcÞ is kept as it is. Every food source has only one employed bee. Therefore, the number of employed bees is equal to the number of food sources. If the position hi of the food source i cannot be improved through the predetermined number of trials ‘‘limit”, then that food source hi is abandoned by its employed bee and then the employed bee becomes a scout. The scout starts to search a new food source, and after finding a new source, the new position is accepted to be hi. Every bee colony has scouts that are the colony’s explorers. The explorers do not have any guidance while looking for food. They are primarily concerned with finding any kind of food source. As a result of such behavior, the scouts are characterized by low search costs and a low average in food source quality. Occasionally, the scouts can accidentally discover rich, entirely unknown food sources. In the case of artificial bees, the artificial scouts could have the fast discovery of the group of feasible solutions as a task. 4762 C. Zhang et al. / Expert Systems with Applications 37 (2010) 4761–4767 https://isiarticles.com/article/7366 
work_g65l3vvdijhoroo6zko6t3pm6m	----	LUND UNIVERSITY PO Box 117 221 00 Lund +46 46-222 00 00 An Architecture for Expert System Based Feedback Control Årzén, Karl-Erik 1988 Document Version: Publisher's PDF, also known as Version of record Link to publication Citation for published version (APA): Årzén, K-E. (1988). An Architecture for Expert System Based Feedback Control. (Technical Reports TFRT- 7399). Department of Automatic Control, Lund Institute of Technology (LTH). Total number of authors: 1 General rights Unless other specific re-use rights are stated the following general rights apply: Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal Read more about Creative commons licenses: https://creativecommons.org/licenses/ Take down policy If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim. https://portal.research.lu.se/portal/en/publications/an-architecture-for-expert-system-based-feedback-control(f7b7cc68-31fb-4e83-b338-d48dffd9cf01).html CODEN: IUTFD2/(TFRr-73 ee) I L-07 I (1es8) An Architecture for Expert System Based Feedback Control Karl-Erik Årzén Department of Automatic Control Lund Institute of Technology September 1988 Department of Automatic Control Lund fnstitute of Technology P.O. Box 118 5-221 00 Lund Sweden Documcnt na¡rc¡c Report Datc of itøuc September L988 Documcnt ÀIumbc¡ CoDEN: tUTFD2/(TFRT-73ee)/1-07l( 1e88) Author(s) Karl-Erik Årzén Supcrviso¡ Sponcoring org¡anísation STU Titlc dnd subtìtle An architecture for expert system based feedback control. Absttæt It is a recognized problem that many industrial control loops are badly tuned or run in manual mode. Two expert system approaches have been suggested for this problem. Fuzzy, rule-based control replace control algorithms by linguistic rules which model the operators manual control strategy. Knowledge-based control extends the range of conventional controllers by encoding general control knowledge and heuristics concerning tuning and adaptation in a supervisory expert system. An architecture for knowledge-based control is described where two concurrent processes are used for the knowledge-based system and the numerical algorithms. A modular, blackboard-based approach is used. This allows the decomposition of the problem into subtasks which are implemented as separate knowledge sourcee that can be rule-based with different inference strategies or procedural. The framework can be compared with a real-time operating system and has similar real-time primitives. The system has been implemented on a VAX tL/780 and used with good experiences. Key words Real-time expert systems, Feedback control Classífic¿tíon systcrm and./or indcx tcrms (íf any) S up plc ment ary b iblio graphic al ínfo tmat io n ISSN and kcy title ISBN Languagc English Numbc¡ oÍ pages 7 R.ccípìcnt's notes S ccuríty classìfrcat íon The report may bc o¡de¡ed from thc Department of Automatíc Cont¡ol or bortowcd úårough thc llnivercíty Librury 2, Box 7010, 3-227 O3 Lund, Swcdcn, Tclcx: 33248 lubbís lund. Paper to be presented at AI in real-time control, the IFAC Vüorkshop on 21-23 September, Wales. An architecture for expert systern based feedback control Kart-Erik Ä.rzén Department of Automatic Control Lund Institute of Technology Box 118 5-221 00 Lund, Swcden Abstract. It is a recognized problem that many industrial control loops are badly tuned or ru¡l in manual modc. Two cxpcrt systcm approachcs havc bccn suggcstccl ftrr tlris problem. Fuzzy, rule-based control replace control algorithms by linguistic rules rv¡ic[ model the operators m¿nual control strategy. Knowledge-based control extends the rangc of conventional controllers by encoding general control knowledge and heuristics concerning tuning and adaptation in a supervisory expert system. An architecture for knorvledge- based control is described where two concurrent processes are used for the knowledge-basecl system and the numerical algorithms. A modular, blackboard-based approach is used. Tiris allows the decomposition of the problem into subtasks which are implemented as separate knowledge sources that can be rule-based with different inference strategies or procedural. The framework can be compared with a real-time operating system and has similar real- time primitives. The system has beer- implemented on a VAX I1/780 and used rvith goocl experiences. Keyuords: Real-time expert systems, Feedback control 1. Introduction There is currently a significant interest in expert sys- tem techniques in the process control community, Applications of many different types have been pro- posed, implemented and a few also fielded. This pa- per considers the use ofexpert system, or knowledge- based system, techniques in the closed control loop. It is a recognized problem that many industrial con- trol loops are badly tuned or run in rnanual ¡node. This decreases the quality of the end product ancl thus increases cost. The manual control task also adds to thc already high cognitivc burclc¡r that pro- cess operators are exposed to in modern control sys- tems. The reasons for the poor control arc many. Onc could be that the control loop is badly tuned from the beginning. Another could be that the operating conditions have changed since the initialization of the controller. This could, c.g. bc duc to operation at dillerent operating points or time-varying dynamics. The conventional solution to the problern of poorly tuned control loops is to use adaptive controllcrs. Adaptive controllers, e.g., (,4.ström and Wittenmarrk, 1989), are currently beginning to be used in i¡rdustrial practice. There are, however, problems. Even though an explicit self-tuning regulator periodically updates the coefñcie¡rts of a process model thcre still a¡e rrÌany parameters that must be set explicitly. Examples are model orders and time scales. Such information can be diflìcult to provide and process operators typicall¡' lack the intuitive understanding that they have with conventional PID controllers- Two expert system approaches have been suggested for the described problem. Both i¡volve using the expert system as a part ol the feedback loop. In the u'ell-known fuzzy or rule-based approach, e.g., (Tong, 1984), the attempt is to model the manuai control strategy of the process operator. It is exprcssed as qualitative, Iinguistic rules lor horv to choose the con- trol signal in different situations, The rules replace conventional control algorithms. The intended appli- cations are control ofcomplcx proccsscs such as, c.g., cement kilns, for which either appropriate models do not exist or are inadequate. The sccond approach, from now on rcfcrrc<l f,o as knowledge-based control (Åström and Anton, 1g84; Ãström et al, 1986i Ârzén, 1987), instead uses ex- pert system techniques to extend the range of con- vcntional control ir.lgorithrns by crrcoding gcrrcral control knowlcdgc and heuristics rcgirrding tuning and adaptation in a supcrvisory expert system. Tlie kno*'ledge-based control approach is closer in spirit to co¡n'entional aclap[ivc control than fuzz¡' conlrol is. The approach is also motivaied by shortconrings of adaptive cont¡ollers. 'Ihe recursive identification algorithrns and the co¡r- trol algorithms of adaptive controlle¡s c¿r¡r l¡e sr:cn Supervlslon algorlthmsKnowledge Bas€d System algorlthme Conlrol algorllhms Process A prlorl informatlon Figure 1. A knowledge-based controller as the final algorithmic representation of an large amount of undcrlying thcoretical as v/ell as practi- cal control knowledgc. This reprcsent,ation is howevcr not enough. It has to be combined with heuristic logic that assures the controller performance under non-standard conditions. These conditions include switching between different operating modes, insuf- ficient process excitation, control signal saüuration etc, The concept safety jacket or saÍety nef (Isermann and Lachmann, 1985; Warvick, 1g88) has been estab- lished for this logic. The safety net part of a controller is often much larger than the actual algorithms and is designed mai¡ly from intuition, experience, ancl simulation. Safety nets tend to be very complex. Ex- perience has shown that design and testing is quite time consu:rring. The approach in knowledgc-based control is to use an expert system to represent the heuristic safety net. The controller consists of the combination of the expert system and a set of control algorithms, identification algorithms and supervision algorithms, as shown in Fig. 1. The topic of this paper is the organization and arch! tecture of a knowledge-based controller. I{nowledge- based control is a real-time expert system applica- lion and, as suclì, contains several difEcult problems such as non-monotoning reasoning, representation of time and temporal reasoning, reasoning under time constraints, responsiveness to asynchronous events etc. For an overview of these issues see Laífey et al (1988) or Chantler (198S). Some of these issues are still unsolved and will be probably never be com- pletely solved. In several cases, however, practical approaches exist that to some degree solve the prob- lems. There exist a widely spread mis-unclerstanding that, adding intelligent behaviour to a controller is simply a matter of generating a few rules and implementing them in an off-the-shelf expert system shell. This is far from the case. There is a strong interplay bctween the architecture of the expert system and the type of knowledge that ca¡r be naturally expressed in it. The majority of the expert system softrvare is st,ill intended for stand alone, off-line applications. Real- time capacities a¡e only available in ferv cases such as, e.9., G2 (Gensym, 1987), PICON (Moore el aI, 1985), and Muse (CCL, 1987). Section 2 describes the overall organization of an ex- pcrt system framework that has been developed and ,-/-\ Figure 2. Ove¡all implementation structure. implemented on a VAX lI1780. The frame-*'ork rvas developed for knowledge-based controi applications but is not restricted to it. The system has been uscd for design of intelligent tuning controllers (,4.r2én, 1987). In Section 3, the architecture of the expert system part of the controller is described t,ogether. Implementational issues are described in Section 4. Finally, Section 5 contains a discussio:r .rbout the sys- tem and a comparison *'ith other systems. 2. Overall architecture The knowledge-based controller consists of two major parts: the numerical algorithms and the knowledgc- based system. To assure that the execution of the nu- merical algorithms are not delayed by the knowledge- based system the parts are implemented as two com- municating concurrent VMS processes wlìere the nu- merical algorithms have the highest priority. The man-machine interface is implemented as a sepa- rate process. From tiris process, the user can interact directly with the knorvledge-based system and incli- rectly with the algorithms. The timer process is used t,o implement certain real-time interrupts described in the next section. The overall structure is shown in Fig 2. The knorvledge-baseci syste¡n and thc rnan-nrachi¡re interface are both writtcn in Lisp. Thc Lisp uscd is the UnLx dialect Franz Lisp, (Foderaro et aI, I?BB). The software package EUNICE, (Kashtan, 1g82), is used to create a Unix environment under Vi',{S. The reason for using Lisp is mainly its powerful syrnbolic processing capabilities. A drawback rvith Lisp in a real-time system is the garbage collection. The probiem can be avoided in t.rvo rvays. By using a Lisp system with i¡rcremental garbage coliection the garbage collection activily is spread uniformly in time and performed in the background. Thc seconcì approaclÌ is to write Lisp code that does not generate any garbage. This is rvhat is done in G2. Tlm. bo Process R.. u lt bor Timer System Basod Numerical algorithm Man-mach¡n€ communlcatio The nurnerical algorithms The numerical algorithm process is written in Pas- cal. It consists of a library of different algorithms such as PID algorithms, pole-placement algorithms, discrete filters, relays, recursive least-squares algo- ritlrrns, lcvcl crossing dctccl,ors, ctc. Thc algorilluns are uniformly coded and have well-defined interfaces. It is relatively straight forward to add new types of algorithms to the system. The process is connected to A/D and D/A converters. The algorithms can principally be divided into three groups: control algorithms, identification algorithms and monitoring algorithms. The cont,rol algorithms all compute a control signal based on command and measurement signals. Only one control algorithm can be running at a time. The identification and moni- toring algorithms all in some sense extract informa- tion from the numerical signal ff.ow. This informa- tion is sent to the knowledge-based system. The al- gorithms in these two groups can be viewed as filt,ers or feature extractors that send information to the knowledge-based system only when something signif- icant has happened. During steady-state operation, the knowledge-bascd system is not i¡volved and t,he system resembles a conventional controller. The sep- aration between the numerical algorithms and the knowledge-based system is favourable from the point of information flow. If a knowledge-based system was interfaced directly to a physical process or to an exist- ing control system, numerical information would have to be senú forth and back again at a high rate. The knowledge-based system also had to itself extract all useful symbolic i¡formation from the signals. This is a task that often is expressed in the form of numer- ical algorithms. Using expert system techniques for such tasks is often inefficient. Inter-pro cess cornnlunication The processes communicate by sending messages through mailboxes shown as rectangles in Fig. 2. The messages that are sent to the algorithms via Outbox are confìguration commands, parameter changes, and information requests from the knowledge-based sys- tem. The messages t,hat go to the knowledge-based system contains results obtained by an algorithm, alarrns that have been detected, answers to infor- mation requests, user cornmands, and timer inter- rupts. Messages to the knowledge-based system are normally sent to Inbox which is a standard first-in, first-out VMS ¡nailbox. Messagcs wilh prioritics in- dicating their importance are allowed, This is made possible by having an inte¡nal maill¡ox inside ùhe knowledge-based system process into which the mcs- sages in Inbox are inserted according to their pri- ority. Important messages such as alarms and timer interrupts have a high priority and are thus taken care of as fast as possible. Answerbox is used for re- sponses to information requests that, has been made by ihe knowledge-based system. Resultbox is used by the knowledge-based system to rcturn results to Lltc man-machine intcrfacc. The use of Lisp has interesting consequences for the communication between the knowledge-based system and thc ¡nan-¡n¿rchinc intcrfocc. Muilboxcs can bc seen as text files whcre a text line corrcsponcls to a message. In Lisp there is no syntactical difierence between dat,a objects and program code, i.e., Lisp functions. Lists are r¡sc<l to rcprcscnt both. This makes it possible to implement remote evaluation of Lisp functions. An arbitrary Lisp function can l¡e sent to the knowlcdgc-based systern from thc rna¡r- machine interface wherc it is cvaluatcd and the rcsull is returned in Resultbox, erronous Knowledge-based system architecture A standard ofÏ-the-shelf expert system framework, OPS 4, was used to implement the knorvledge-based part of the system in a first prototype (.Â.rzén, 1g86a, 1980b). OPS 4 (Forgy, 1979) is a simple rule-based, forward-chaining expert system framework. The rea- son for this choice was the data-driven nature of knowledge-based control. Data in the form of signif- icant events detected by the algorithms are sent to the knowledge-based system which should react and generate some response. The framework OPS 4 uses the incremental pattern-matching algorithm RETE (Forg¡ 1982) and is therefore also reasr¡nably fast. Another reason for the choice was simply that the system was alrailable to us and that rve rvanted to test the basic ideas rapidl¡'. Experiences of a prototype The first prototype was used to implement a relay- based PID auto-tuner (Åström and Hägglund, 1984). Ðxperiments with the prototype gave many results concerning both the feasibility of the approach and the demands on a expert system framex'ork for knowledge-based control. We were reassured in that the approach is feasible. The response times for the knowledge-based system were acceptable. The sam- pling rate for the numerical aigorithm process was 1 sccond. It took approximatively 2-3 sampling pe- riods from that a message tvas sent to the until a responding message was returned. A second positive result was a clean implementation of a relay auto- tuner that clearly benefited from the separation of logic and algorithms. The tin.re and ellort to make extensions to the controller were significantly smaller tha¡r for comparablc irnplcrncntaLions i¡r co¡rvc¡r{,i<.¡¡r¿rl languages. The negative experiences all concerncd OPS4. It be- camc clear that a sirnple forward-chaining rule sys- tem is not sufficient. An expert system framework must allow for a modular decomposition of both the rulebase and the database. For the database, this could be achieved by a frame system. The rulebase must be decomposable into rulc groups for the dilTer- ent subtasks. Furthcr it was found thaü one knowl- edgc represcntation technique is not enorrgh. Somc subtasks contai¡rs large sequential elements. 'Ihcse are mo¡e naturally represented procedurally tlran 3 Global Database (Blackboard) Schedu ler Knowledge sources TJ Figure 3. Knowledge based control st¡ucture with rules. It was also clear that backward chaining inferencing would be useful for some problems. Thc most, important drawback with OPS 4 was, however, that it is not designed for real-time operation. It has, e.9., no possibilities to havc timc-outs associatcd with database elements, no means for halting the rule exe- cution for a certain time, and no possibilities to check rules at given time intervals. A blackboard system Based on the experiences of the prototype, a real- time expert system framework has been developed. The reasoning model chosen as the basis for frame- work is the blackboard model, (Nii, 1986). A globat database, the blackboard, is available to dilïerent, co- operating knowledge sources. The database allows for frame structures for storing associated information. The knowledge sources can be thought ofas different actors, each ofwhich solves some subtask ofthe prob- lem. The knowledge sources also have thei¡ own local databases. Knowledge sources can be rule-based with either forward or backwa¡d chaining and procedural, The structure of the framework is shown in Fig. B. A knowledge source implements the domain knowl- edge for a certain task. It is ofien associated with one or more numerical algorithms. It could for example contain the hcrrristic logic surrounding an algorithm. The knowledge sources have primitives for adding, modifying, and deleting frames both gtobally and lo- cally. They also have primitives to halt their execu- tion for a certain time or until a certain dat¿base element is added to the blackboard. It is possible to have forward chaining rules that are tested with spe- cific time i¡terv¿ls and to associate validiiy intervals with database elements. The operation of the knowledge-based controller in- volves the actiyation of different knowledge sources both in sequence and in parailel. A typical case when knowledge sources are active in parallel is during the steady state control of the process. One knowl- edge source takes care of the actual control algorithm while other knowledge sources implernent diferent monitoring aspects. A separate rule-based module schedules the selcc- tion of knowledge sourccs at two different levels. The first level involves the sequential activation of dif- ferent knowledge sources. The second level involves the scheduling between different knowledge sources that are active simultaneously. This resembles the scheduling in an ordinary real-time, multi-tasking operating system where the knowledge sources are thc equivalents of concurrcnt proccsscs, Â knos,lcclgc source runs until it explicitly rcturns control to thc scheduler, e.g,, if it has to wait for some information or if it is finished. A simple extension rvhich allo.*,s interrupts among the knowledge sources is described in Årzén (i987). With this extension, priorities can be associated with knowledge sources. The scheduìer contains frames with inlormation of the state of the different knowledge sourccs and fra¡ncs rvhich co¡r_ tains information about the conditions on *,hich a knowledge source is waiting. Thc prinritivcs that involvcs waLil,ing ¿ ccrt¿¡in l,i¡nc are implemenied with the help of the timer process. A primitive that causes a knowledge source to .rvait a certain time gives rise to a mcssagc to thc timcr process. The message contains the desired wakeup time and a unique identifier for the rvaittime request. A high-priority message is returned to the scheduler when the waiting time has elapsed. This message causes the state of the waiting knowledge source to be changed to ready. Knowledge source combination The operation ofthe knowledge-based controller typ- ically consists of a sequence, with parallel parts, of knowledge source activations. Three diffe¡ent meth- ods for combiniag knowledge sources into scquences have been implemented. The most straightforward way is to use primitives that let knowledge sources acti'r'ate and deactivate each other. A knorvledge source has the possibility to wait until another knowledge source is finished. Procedural knowledge sources also have t,he possibil- ity to call other procedural knorvledge sources, and await and use their returned result. Anothcr alLcrnativc is to havc ¿l nunrL¡c¡ of ¡-rrc-sl,orcd sequc¡rces. One example of a sequence could be the i¡ritial tuning sequence. Othcr sequences could be used to ¡eturn to steady-state control when different alarm conditions have been de tected. Combination of knowledge sou¡ces into sequences is b¿sically a pro- cedural operation. It is therefore natural to express it with procedural knowledge sources. In order for this to be possible, rvait primitivcs that allows *'aiting for conjunctions and disjunctions of multiple events have been implemented. Tire last, and most complex mcthod is to dynanri- cally generate sequences. This is accom¡rlished by as- sociating goal states, i.e., post-conditions, and ini- tial states, i.e., pre-conditions, with each knowledge source. Each knowledge source can be vierved as an opcrator that transforms tìrc staie of the system front its initial statc to its goal statc. A sequence is recu¡- sivcly generated b¡. comparing the desircd goal and the current state u'ith the pre- and post-conditions of the operators. This formulation turns the problem into a planning problem. The scheduler generates a plan which then is executed. The possibility for diferent knowledge representa- tion techniques allows the user to choose the tech- nique most natural for each sub-problem. The v¿rious methods of combining knowledge sources give a rich and flexible structure. For instance, it is possible to have one knorvledge source that contains monitoring rulcs which are checked periodically. If something er- roneous is detected the rules can invoke other knorvl- edge sources that focuses on the problem. These knowledge sources could, e.g. be backward chainers that tries t,o verify some hypothesis concerning the error or proceclural knowledge sources that performs some procedural tests. Meta-knowledge sourccs with knowledge about, the applicability of othcr knowlcdge sources are also easy to implement. A knowledge source in a knowledge-based control application could be, e.g., contain knowledge about design of different controllers. Another knowledge source could contain knowledge about modelling ¿¡rcl ¡noclcl vnlidation. Othcr cxam¡rlcs cor¡lrl co¡rlni¡r knowledge of dillerent monitoring aspects of the con- troller. The possibility to refer to past signal values is important in a real-time environment. This is pos- sible through statistics knowledge sources that com- putes signal statistics over different time horizons. These knowledge sources are associated witîr numer- ical algorithms that collect the signal values. The described framework has been used for the de- sign of elaborate extensions of relay auto-tuning. This is described in ,A,rzén (1987). 4, Implementation The implementation of the expert system framework is built on the object-oriented system Flavors (Can- non, 1982) and the forward-chaining production sys- tem YAPS (Allen, 1983). The YAPS system is a pattern-matching system i¡ the same spirit as the OPS family with a similar optimized, incremental matching algorithm. The important difference is that YAPS is v¡ritten in Flavors and allows Flavor in- stances in its database. These Flavor instances can be instances of other YAPS systems. The YAPS system originally only allows arbitrarily nested list structures of containing numbers, atoms, and Flavor instances as database elements. The sys- tem has been modified to allow frame structures. The system has also been extended to allow for auto- matic explanations of how database elements have l¡een added to the system. The Scheduler is implement,ed as a flavor which in- herits a YAPS flavor. The scheduling strategy is rep- resented with rules. The different types of knowledge sources are implemented as diflerent flavors. Each in- dividual knowledge source is an instance of the cor- responding flavor. Each knorvledge source is repre- sented as a frame in the scheduler database. The frarne contains slots for the type of knorvledge source, Scheduler - YAPS system Figure 4. Implementation stnrcture e.g., forward or procedural, for the state of the knorvl- edge source, and for the actual flavor instance that implements the knowledge source. The implementation structure is illustrated in Fig, 4. The actual interface between the knorvledge sources and the schedrrler consists of a relat.ively small sct of messages for which the knowledge sou¡ce flavors should supply methods. This makes it easy t,o add new types of knowledge sources to the system. A slightly simplified example of a rule in the sched- uler is given below. (p schsdul€l 'rIf a knorledgo sourco is ready a¡d no othor knoçledge sourco is runaing thsn ru¡ this knosl_edge sourcs" (f ra¡e k¡o¡ledgo-source status activo 6tato t6ady instance -x) (' (fra¡oe krrorledge-source stato running)) (nodify 1 stat€ running) (<- -¡ 'run)) The forward chailing knowiedge sources are imple- mented as instances of a flavor that inherits a YAPS flavor. This givcs a st¡ucture where several YAPS sys- tcms residc as datab¿rsc clements insidc tirc Schcdulcr YAPS system. The backward chaining knowledge sources are based on an small expert systcm example in Winston and Horn, (1981), which has bcen extended and embed- ded in Flavors. Currently they can only ask qucs- tions to the operator when an answer cannot be au- tomatically deduced. A possible extension rvould be to, in that case, allow a backward chaining knowÌedge source to invoke a lorward or procedural knowledge source that returns the needed atìswer. The procedural knowledge sources cor.rsists of Lisp functions. In order to allow interrupts of these func- tions at arbitrary places the entire state of the Lisp computation must be saved. This is not possiblc rvith the Franz Lisp of the basic system. Inste¿Ld Lisp has been used to write a register nìacllirìe basecl intcr- preter for a procedural Lisp-like language that allorvs the computation to be suspended. ,._ìì I Knowledge ü:Þ Global Database Rules Scheduling rules 5. Summary Àrr gcrrcrul cx¡rcrL systcrrr frur¡rcwork fcrr rcul-tirrrc u¡_l- plications has becn prescntcd. It has becn dcvelopcd for knowledge-based control applications but is not restricted to it. The framework has real-time facilities. It is modu- larized into knowledge sources that can be compared with concurrent processes. This is similar to thc Muse system. In the current version, however, the knorvl- edge sources cannot interrupt each other. With a sim- ple extension this is possible. The knowledge sources have primitives to wait a certain time or for a cer- tain database element. These primitives are used to implement periodic rule testing in a way similar to G2 and Picon. Validity intervals can be used to indicate how long database elements remain valid. In contrast v¡ith G2 and Picon the validity interv¿ls are not propagated to inferred facts. History values of important signal values are maintained. It is not possible to store history va.lues of arbitrary frame attributes. The system allows for both rule-based and procedural representation which is very important, The flexible means of combining knowledge sources gives a rich structure. 1ìrere are many similarities between real-time op- erating systems and real-time knowledge-based sys- tems. Real-time operating systems for process control have evolved over a long period of time. This paper indicatcs a new system architccturc where real-time operating systems, databases, object-oriente(i pro- gramrning, and knowledge-based systems are com- bined. The exccution speed of the system is of the orcler of one forward chaining rule per second, The system is currently being ported to a Symbolics - IBM PC en- vironment where the knowledge-based system resides on the Symbolics and the numerical algorithms reside on the IBM PC. Preliminary results indicate a factor of 10 in increased speed. Other possible candidates for migration are systerns where powerful symbolic processing capacity is combined with conventional cornputing. One example of this is the ¿z-Explorer. ,4. cknorvle dge rncnts The author rvould like to thank Professor I{arl Johan Âström and Sven Erik Mattsson for many useful discussions. This work has been supported by the National Swedish Board for Technical Development (STU) under contract 85-3084. Refe¡ences ALLnu, E.M. (1983): ,,yApS: yet another production system," TR-1146, Department of Computer Science, University of Maryland. ÂnzÉN, K-8. (19S'6a): ,,Expert systems for process control,' in D. Sri¡am and R. Adey (trds.): proc. of Fj¡st Intetnational Conference on Applications of Artifrcial lntclligence in Engineering Practice, Springer Verlag, Berlin, pp. 1127-1138. .ÅnzÉx, K-8. (1986b): "Usc of expert svstems in closed loop feedback control," Proc. of American Control Conference, Seattle, Wr\. ÅnzÉx, K-8. (19S7): "Realization of cxpert system bascd feedback control," Ph.D. thesis CODEN: LUTFD2/ TFRT-1029, Department of Automatic Control, Lund Llul,iùutc of 'l'cchnology, Lund, Swcdcr¡. ÂsrRölt, K.J. and J..J. ANro¡ (1984): ,,Experr control,,, Proc. 9'th IFAC World Congress, Budapest, IIungar5,. Ås'rRöu, K.J. and 'I'. IIÃccLUND (1984): ,,Autorrratic tuning of simple regulators," Proc. IFAC g'th l!¡orld Co ngress, B udapest, Ilungary. Åsrnö¡'¡ K.J. and B. W¡TTnu¡¿lnx (fOSO): Adaptit,e Control, To appear, Addison-lVesley, Reading, lr,fA. ÅsrRölr, K.J., J.J. Ar.row and K.-E. ÅnzÉN (1986): "Expert control," Automatica, 22, 3, 277-286. CaNuoll, H.I. (1982); "Flavors: A non-hierarchical ap- proach to object-oriented programming,," unpublished paper. CCL (1987): "Muse product description,', Cambridge Consultants Liroited, 4 pages. Cu.lxrr,er, M.J. (1988): ,'Real-time aspects of experr systems in process control," Colloquium on Expert Systems jn P¡ocess Control, IEE, Savoy Place, Lonclon UK. FoouR,o.Ro, J.K., K.L. Sxr,ownR and K. Lay¡R (19g3): "The Franz Lisp Manual," ',JC Berkeley, California. FoRcY, C.L. (1979): "OPS4 IJser's manual," Technical report CMU-CS-79-132, Department of Computer Sci- ence, Carnegie-Melion University, FoRcY, C.L. (1982): "Rete: À fast algorithm for the many pattern/many object pattern match problem," Artifr.cial Intelligence, 75, l, 17-37. GnNsyl,t (1987): G2 User's manual, Gens¡.m Corp.. Cambridge, MA. IsERlaaNN, R. and K.H. Llcuuaun (1985): ,,pa¡ame- ter-adaptive control with configuration aids and sr.rper- vision functions," Äuf,omalica, 2L, 6, 625-638. K.a,snT,nN, D.L. (1982): "EUNICE: A system for porring UNIX programs to VAX/VMS," Artificial Intelligence Center, SRI Inte¡national, \,Ienlo Park, California. L,rrrnv, 1'.J., P.A. Cox, J.L. Scurvrrot, S.tr,I. K¡o. and J. Y READ (1988): "Real-time knowledge-based systems," AI Magazine, I, I, 27-45, MooRrì, R.L., L.R. IIrrwxrwsor,r, M.lì. i.¡;vtN;r¡rrì C.G. I(xrcxnRBocKÐR (1gS5): ,,Expert control,', proc. American Conttol Conf., Boston, \,f Ä, pp. ES5-887. Ntt, lI.P. (i986): "Blackboa¡d sy6tems: t he Ì¡lackl¡oarrl model of probiem solving and the evolution of black- board a¡chitectures," AI Magazine, 7, 2, 38-53. ToNc, R.M. (198a): "A retrospective vicw of ftzzy con- trol systems," Fuzzy Sets and ,Sysúerns, 14, lgg-210. WaRvtcr, K. (1988): Implementation of self-tuning controllers, Peter Peregrinus, London. 'Wnts,oN, P.H. and B.K.P. HoRN (1981): Iisp, Ad<ìi- son-\\Iesley, Reading, MA, 
work_g6ejnwnypvf3blhqqz2iyiswry	----	A direct interval extension of TOPSIS method Expert Systems with Applications 40 (2013) 4841–4847 Contents lists available at SciVerse ScienceDi rect Expert Systems with Applic ations j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A direct interval extension of TOPSIS method 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.02.022 ⇑ Corresponding author. Tel./fax: +48 34 3250 589. E-mail address: sevast@icis.pcz.pl (P. Sevastjanov). Ludmila Dymova, Pavel Sevastjanov ⇑, Anna Tikhonenko Institute of Comp. & Information Sci., Technical University of Czestochowa, Dabrowskiego 73, 42 201 Czestochowa, Poland a r t i c l e i n f o a b s t r a c t Keywords: TOPSIS Interval extension The TOPSIS method is a technique for order preference by similar ity to ideal solution. This technique cur- rently is one of the most popular methods for Multiple Criteria Decision Making (MCDM). The TOPSIS method was primary developed for dealing with only real-valued data. In many cases, it is hard to present precisely the exact ratings of alternatives with respect to local criteria and as a result these ratings are considered as intervals. There are some papers devoted to the interval extensions of TOPSIS method, but these extensions are based on different heuristic approaches to definition of positive and negative ideal solutions. These ideal solutions are presented by real values or intervals, which are not attainable in a decision matrix. Since this is in contradiction with basics of classical TOPSIS method, in this paper we propose a new direct approach to interval extension of TOPSIS method which is free of heuristic assumptions and limitations of known methods. Using numerical examples we show that ‘‘direct interval extension of TOPSIS method’’ may provide the final ranking of alternatives which is substantially different from the results obtained using known methods. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction The technique for order performanc e by similarity to ideal solu- tion (TOPSIS) (Lai, Liu, & Hwang, 1994 ) is one of known classical MCDM method. It was first developed by Hwang and Yoon (1981) for solving MCDM problems . The basic principle of the TOPSIS method is that the chosen alternative should have the shortest distance from the positive ideal solution and the farthest distance from the negative ideal solution. There exist a large amount of literature involving TOPSIS theory and applications. Is was shown by Garca-Cascal es and Lamata (2012) and Wang and Luo (2009) that ‘‘One of the problems attributable to TOPSIS is that it can cause the phenomeno n known as rank reversal. In this phenomeno n the alternative’s order of preference changes when an alternative is added to or removed from the decision problem. In some cases this may lead to what is called total rank reversal, where the order of preferences is totally inverted, that is to say, that the alternative considered the best, with the inclusion or re- moval of an alternative from the process, then becomes the worst. Such a phenomeno n in many cases may not be acceptable’’. Wang and Luo (2009) showed that rank reversal phenomeno n occurs not only in the TOPSIS method, but in many other decision making ap- proaches such as Analytic Hierarchy Process (AHP), the Borda–Ken- dall (BK) method for aggregating ordinal preferences, the simple additive weighting (SAW) method, and the cross-efficiency evalua- tion method in data envelopment analysis (DEA). Therefore, we can say that this problem is typical for known method of MCDM. In Garca-Cas cales and Lamata (2012), the authors proposed a new method for the solution of this problem in the framework of TOPSIS method. Nevertheless it was pointed out in Garca-Cascal es and La- mata (2012) that ‘‘the two methods ‘‘the classical’’ and ‘‘the new’’ do not have to give the same order. This is especially so in the case of evaluating alternatives which are very close’’. In other words, ‘‘the classical’’ and ‘‘the new’’ methods may provide different re- sults based on the same decision matrix. Hence, it is not obvious that ‘‘the new’’ method performs better than classical one if there is no need to add or remove an alternative from the decision problem. Therefore, hereinafter we shall consider only classical TOPSIS method and its interval extension. In classical MCDM methods, the ratings and weights of criteria are known precisely. A survey of these methods is presented in Hwang and Yoon (1981). In the classical TOPSIS method, the ratings of alternatives and the weights of criteria are presented by real values. Neverthel ess, sometimes it is difficult to determine precisely the real values of ratings of alternatives with respect to local crite- ria, and as a result, these ratings are presented by intervals. Jahansha hlo, Hosseinzade, and Izadikhah (2006) and Jahan- shahloo, Hosseinzade h Lotfi, and Davoodi (2009) extended the con- cept of TOPSIS method to develop a methodology for solving MCDM problem with interval data. The main limitatio n of this approach is that the ideal solutions are presente d by real values, not by http://dx.doi.org/10.1016/j.eswa.2013.02.022 mailto:sevast@icis.pcz.pl http://dx.doi.org/10.1016/j.eswa.2013.02.022 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 4842 L. Dymova et al. / Expert Systems with Applications 40 (2013) 4841–4847 intervals. The similar approach to determining ideal solutions is used in Yue (2011) and Sayadi, Heydari, and Shahanaghi (2009) this approach is used in context of the so-called VIKOR method which is based on the measure of ‘‘closenes s’’ to the ideal solutions, too. In this paper, we show that these extensions may lead to the wrong results especially in the case of intersect ion of some inter- vals representing the ratings of alternatives . In Chen (2011), Jahanshahloo et al. (2009), Jahanshahlo o, Khodabakhshi , Hosseinz adeh Lotfi, and Moazam i Goudarzi (2011), Tsaur (2011), Ye and Li (2009) and Yue (2011), the different definitions of interval positive and negative ideal solutions are pro- posed. They will be analysed in the next section, but their common limitation is that they are based on the heuristic assumptions (usu- ally without any analysis and clear justification) and provide inter- val ideal solutions that are not always attainable in the interval- valued decision matrix. Since this is in contradiction with basics of classical TOPSIS method, in this paper, we propose a new direct approach to inter- val extension of TOPSIS method which is free heuristic assump- tions and limitations of known methods. This approach makes it possible to obtain the positive and negative ideal solutions in the interval form such that (opposite to the known methods) these interval-valued solutions are always attainable on the initial interval-valued decision matrix. Since this approach is based on the interval comparison, a new simple, but well-justified method for interval comparison is develope d and presented in the special section. It is worth noting that the most general approach to the solu- tion of MCDM problems in the fuzzy setting is the presenta tion of all fuzzy values by correspond ing sets of a-cuts. There are no restrictions on the shape of membership functions of fuzzy values in this approach and the fuzzy TOPSIS method is reduced to the solution of MCDM problems using interval extended TOPSIS meth- od on the corresponding a-cuts (Wang & Elhag, 2006 ). Therefore, the development of a reliable interval extension of TOPSIS method may be considered as a first step in the solution of MCDM problems using the fuzzy TOPSIS method. The rest of the paper is set out as follows. In Section 2, we pres- ent the basics of TOPSIS method and analyse its known interval extensions. Section 3 presents the direct interval extension of TOP- SIS method and the method for interval comparison which is needed to develop this extension. The results obtained using the method proposed in Jahanshahlo et al. (2006) and Jahanshahlo o et al. (2009) are compare d with those obtained by the develope d new method. Section 4 concludes with some remarks. 2. The basics of TOPSIS method and known approaches to its interval extension The classical TOPSIS method is based on the idea that the best alternative should have the shortest distance from the positive ideal solution and the farthest distance from the negative ideal solution. It is assumed that if each local criterion is monotonically increasing or decreasing, then it is easy to define an ideal solution. The positive ideal solution is compose d of all the best achiev- able values of local criteria, while the negative ideal solution is composed of all the worst achievable values of local criteria. Suppose a MCDM problem is based on m alternatives A1,A2, . . . ,Am and n criteria C1, C2,. . .,Cn. Each alternative is evaluated with respect to the n criteria. All the ratings are assigned to alternatives and presented in the decision matrix D[xij]m�n, where xij is the rating of alternative Ai with respect to the criterion Cj. Let W = (w1,w2, . . . ,wn) be the vector of local criteria weights satisfying Pn j¼1wj ¼ 1. The TOPSIS method consists of the following steps: 1. Normaliz e the decision matrix: rij ¼ xijffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPm k¼1x 2 kj q ; i ¼ 1; . . . ; m; j ¼ 1; . . . ; n: ð1Þ Multipl y the columns of normalized decision matrix by the associ- ated weights: v ij ¼ wj � rij; i ¼ 1; . . . ; m; j ¼ 1; . . . ; n: ð2Þ 2. Determine the positive ideal and negative ideal solutions, respectivel y, as follows: Aþ ¼ vþ1 ; v þ 2 ; . . . ; v þ n � � ¼fðmax i v ijjj 2 K bÞðmin i v ijjj 2 K cÞg; ð3Þ A� ¼ v�1 ; v � 2 ; . . . ; v � n � � ¼fðmin i v ijjj 2 K bÞðmax i v ijjj 2 K cÞg; ð4Þ where Kb is the set of benefit criteria and Kc is the set of cost criteria. 3. Obtain the distances of the existing alternativ es from the posi- tive ideal and negative ideal solutions: two Euclidean distances for each alternatives are, respectivel y, calculated as follows: Sþi ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn j¼1 ðv ij � vþj Þ 2 vuut ; i ¼ 1; . . . ; m; S�i ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn j¼1 ðv ij � v�j Þ 2 vuut ; i ¼ 1; . . . ; m: ð5Þ 4. Calculate the relative closeness to the ideal alternativ es: RCi ¼ S�i Sþi þ S � i ; i ¼ 1; 2; . . . ; m; 0 6 RCi 6 1: ð6Þ 5. Rank the alternativ es accordin g to the relative closeness to the ideal alternatives: the bigger is the RCi, the better is the alterna- tive Ai. In Jahansha hlo et al. (2006) and Jahanshahlo o et al. (2009), an interval extension of classical TOPSIS method was proposed. This approach may be described as follows. Let ½xij� ¼ ½xLij; x U ij � be an interval value of jth criterion for ith alter- native (xLij and x U ij are the lower and upper bounds of interval, respectivel y), W = (w1,w2, . . . ,wn) be the weight vector satisfying Pn j¼1wj ¼ 1. Then D½½xLij; x U ij ��m�n is the interval-valued decision ma- trix. The method proposed in Jahanshahlo et al. (2006) and Jahan- shahloo et al. (2009) consists of the following steps: 1. Normaliz ing the decision matrix using the following expressions : rLij ¼ xLij Pm k¼1 x L kj � �2 þ xUkj � �2� �� �12 ; i¼1;. . .;m; j¼1;. . .;n; ð7Þ rUij ¼ xUij Pm k¼1 x L kj � �2 þ xUkj � �2� �� �12 ; i¼1;. . .;m; j¼1;. . .;n: ð8Þ 2. Taking into account the importance of criteria, the weighted normalized interval-valued decision matrix is obtained using the following expressions: v Lij ¼ wj � r L ij; v U ij ¼ wj � r U ij ; i ¼ 1; . . . ; m; j ¼ 1; . . . ; n: L. Dymova et al. / Expert Systems with Applications 40 (2013) 4841–4847 4843 3. The positive and negative ideal solutions are obtained as follows: Aþ ¼ vþ1 ; v þ 2 ; . . . ; v þ n � � ¼ max i v Uij jj 2 K b � � ; min i v Lijjj 2 K c � � ; ð9Þ A� ¼ v�1 ; v � 2 ; . . . ; v � n � � ¼ min i v Lijjj 2 K b � � ; max i v Uij jj 2 K c � � : ð10Þ 4. The separation of each alternative from the positive ideal solu- tion is calculated using the n-dimensional Euclidean distance: Sþi ¼ X j2Kb v Lij � v þ j � �2 þ X j2K c v Uij � v þ j � �2( )12 ; i ¼ 1; . . . ; m: ð11Þ Similarly, the separation from the negative ideal solution is calcu- lated as follows: S�i ¼ X j2Kb v Uij � v � j � �2 þ X j2K c v Lij � v � j � �2( )12 ; i ¼ 1; . . . ; m: ð12Þ 5. Calculate the relative closeness to the ideal alternatives : RCi ¼ S�i Sþi þ S � i ; i ¼ 1; 2; . . . ; m; 0 6 RCi 6 1: ð13Þ Table 1 Decision matrix. C1 C2 A1 [5, 7] [22, 32] A2 [0, 10] [25, 27] 6. Rank the alternatives according to the relative closeness to the ideal alternativ es: the bigger is the RCi, the better is the alterna- tive Ai. In Jahanshahloo et al. (2009) and Jahansha hloo et al. (2011), the interval extension of TOPSIS method based on interval-val ued ideal solution was proposed . This method is based on the assumption that the ideal and negative-ideal solutions are changed for each alternative. Nevertheles s, analysing this approach we can see that obtained interval ideal solutions are not always attainable in the interval decision matrix. In Ye and Li (2009), an interval TOPSIS method is used in the framework of group multi-attribute decision model. For the deci- sion maker k, the positive and negative interval-valued solutions are presented in Ye and Li (2009) (in our notation) as follows: Aþk ¼ mþkU1 ; m þkU 2 ; . . . ; m þkU n � � � ¼ max i mkLij ; max i mkUij � � ; A�k ¼ mþkU1 ; m þkU 2 ; . . . ; m þkU n � � � ¼ min i mkLij ; min i mkUij � � : There is no any justification of this approac h in Ye and Li (2009) and only what we can say about it is that in the case of only benefit cri- teria the upper bound of A+k may be obtained from (9) and the low- er bound of A�k may be obtained from (10). It is also important that interval-v alued ideal solution s obtained using the method proposed by Ye and Li (2009) are not always attainabl e in the interval -valued decision matrix. The similar problem may be found in Chen (2011). A more complicated approach to the definition of interval-val- ued ideal solutions was proposed by Tsaur (2011), where the author wrote ‘‘Theoretica lly, a for pivot value m _þ j ¼ ½m _þL j ; m _þU j � for criterion j in the positive ideal solution, we know that both of m _þL j and m _þU j might be obtained from different alternatives.’’. There- fore, the ideal solutions obtained using the method proposed by Tsaur (2011) are not always attainable in the interval-valued deci- sion matrix. In Yue (2011), the group decision making problem was solved using the modified interval extension of TOPSIS method. In this ap- proach, the positive and negative ideal solutions were presented by interval-valued matrices. For example, the negative ideal solution was presented as follows A� ¼ ð½m�Lij ; m �U ij �Þm�n , where m�Lij ¼ minkm kL ij ; m �U ij ¼ maxkm kU ij (k is a number of decision maker). It is easy to see that obtained A� is not always attainable in the interval decision matrices provided by decision makers. Summaris ing, we can say that the common limitatio n of known approach es to interval extension of TOPSIS method is that they (based on the different assumptions) provide interval-valued ideal solutions which are not always attainable in correspondi ng inter- val-value d decision matrices. This is in contradiction with basics of classical TOPSIS method and is a consequence of heuristic assumpti ons which are not usually justified enough. In the most of analysed approaches, the upper bound of positive interval-valued solution is calculated as in the expression (9) and the lower bound of negative interval solution is calculated as in (10). Hence, we can say that the approach developed in Jahanshahlo et al. (2006) and Jahansha hloo et al. (2009) providing real-valued ideal solutions attainable in the interval-valued decision matrix seems to be more justified than the other analysed here ap- proaches providing interval-val ued ideal solutions, which are not always attainable in the interval-valued decision matrix. Therefore, to compare a direct interval extension of TOPSIS method we pro- pose in this paper with other known approach es, it seems to be en- ough to compare the results obtained by our method with those obtained using the method develope d in Jahanshahlo et al. (2006) and Jahansha hloo et al. (2009) (see expressions (7)–(13)). 3. A new approach to the interval extension of TOPSIS method 3.1. The problem formulation Therefore, a more correct and straightforw ard approach to cal- culation of ideal solutions is representing them in the interval form using the expressions : Aþ ¼ vþL1 ; v þU 1 � � ; vþL2 ; v þU 2 � � ; . . . ; vþLn ; v þU n � �� � ¼ max i v Lij; v U ij h i jj 2 K b � � ; min i v Lij; v U ij h i jj 2 K c � � ; ð14Þ A� ¼ v�L1 ; v �U 1 � � ; v�L2 ; v �U 2 � � ; . . . ; v�Ln ; v �U n � �� � ¼ min i v Lij; v U ij h i jj 2 K b � � ; max i v Lij; v U ij h i jj 2 K c � � : ð15Þ As there are no any type reduction s (representation of interv als by real values) and additi onal assumpt ions concerned with expres- sions (14) and (15), we call our approach Direct Interval Extension of TOPSIS method. It is easy to see that expressions (14) and (15) provide the posi- tive and negative interval-valued ideal solutions which are always attainable in the corresponding interval-valued decision matrix. To perform the difference of proposed approach from known ones it is enough to compare it with the method developed in Jah- anshahlo et al. (2006) and Jahansha hloo et al. (2009) (see explana- tion at the end of previous section). 4844 L. Dymova et al. / Expert Systems with Applications 40 (2013) 4841–4847 Suppose we deal with the interval-valued decision matrix pre- sented in Table 1, where [x11] = [5, 7], [x12] = [22, 32], [x21] = [0,10] and [x22] = [25, 27] represent the ratings of alternatives A1 and A2 with respect to the benefit criteria C1 and C2. Since we deal with the only benefit criteria C1 and C2, then expression (9) in our case is reduced to Aþ ¼fmþ1 ; m þ 2 g¼ maxiðx U ijÞ. Using this expression, from the first column of Table 1 we get mþ1 ¼ 10 and from the second column m þ 1 ¼ 32. Therefore Aþ ¼fmþ1 ; m þ 2 g¼f10; 32g. Similarly , from (10) we get A ¼fm1; m2g¼ miniðx L ijÞ¼ f0; 22g. On the other hand, using any method for interval comparison (see Sevastjanov (2007), Wang & Kerre (2001) and subSection 3.2) we obtain that [x21] < [x11], [x22] < [x12] and for the positive and negative interval ideal solutions we get: ½Aþ�¼ f½m�þ1 ; ½m� þ 2 g¼ f½5; 7�; ½22; 32�g and ½A��¼ f½m��1 ; ½m� � 2 g¼f½0; 10�; ½25; 27�g. It is easy to see that fmþ1 ; m þ 2 g¼f10; 32g is not included in f½m�þ1 ; ½m� þ 2 g¼f½5; 7�; ½22; 32�g and fm � 1 ; m � 2 g¼f0; 22g is not included in f½m��1 ; ½m� � 2 g¼f½0; 10�; ½25; 27�g. Thus, we can say that in the cases when some intervals in the decision matrix intersect, the approach proposed in Jahanshahlo et al. (2006) and Jahansha hloo et al. (2009) may lead to wrong results. It is seen that our method provides the interval-valued ideal solutions which are strongly attainable in the considered decision matrix (see Table 1.). Since the other known methods analysed in previous section may produce interval-valued ideal solutions which are not attainable in the considered decision matrix (see Ta- ble 1.), they may produce the wrong results too. We use here the words ‘‘wrong results’’ to emphasize that only our approach based on the expressions (14), (15) guarantees that obtained interval-valued ideal solutions will be always attainable in the considered interval-valued decision matrix and this is in compliance with basics of classical TOPSIS method. As in (14) and (15) the minimal and maximal intervals must be chosen, the main difficulty in the implementati on of the above method is the problem of interval comparison. 3.2. The methods for interval comparison The problem of interval comparison is of perennial interest, be- cause of its direct relevance in practical modeling and optimization of real-world processes. To compare intervals, usually the quantitative indices are used (see reviews in Sevastjanov (2007) and Wang & Kerre (2001)). Wang, Yang, and Xu (2005) proposed a simple heuristic method which provides the degree of possibility that an interval is great- er/lesser than another one. For intervals B = [bL, bU], A = [aL, aU], the possibilities of B P A and A P B are defined in Wang et al. (2005) Wang, Yang, and Xu (2005) as follows: PðB P AÞ¼ maxf0; bU � aLg� maxf0; bL � aUg aU � aL þ bU � bL ; ð16Þ PðA P BÞ¼ maxf0; aU � bLg� maxf0; aL � bUg aU � aL þ bU � bL : ð17Þ The similar expression s were proposed earlier by Facchinett i, Ricci, and Muzzioli (1998) and by Xu and Da (2002). Xu and Chen (2008) showed that the expressions proposed in Facchine tti et al. (1998), Wang et al. (2005)<br/>and Xu and Da (2002) are equivalent ones. A separate group of methods is based on the so-called probabi- listic approach to the interval comparis on (see review in Sevastja- nov (2007)). The idea to use the probability interpretation of interval is not a novel one. Nevertheles s, only in Sevastjanov (2007) the complete consistent set of interval and fuzzy interval relations involving separated equality and inequalit y relations develope d in the framewor k of probability approach is presented. Neverthel ess, the results of interval comparison obtained using expressions (16) and (17) generally are similar to those obtained with the use of probabilistic approach to the interval comparison. The main limitations of described above methods is that they provide an extent to which an interval is greater/l esser than an- other one if they have a common area (the intersection and inclu- sion cases should be considered separately (Sevastjanov (2007))). If there are no intersections of compared intervals, the extent to which an interval is greater/l esser than another one is equal to 0 or 1 regardles s of the distance between intervals. For example, Let A = [1, 2], B = [3, 4] and C = [100, 200]. Then using described above approach es we obtain: P(C > A) = P(B > A) = 1, P(A > B) = 0. Thus, we can say that in the case of overlapping intervals the above methods provide the possibility (or probabili ty) that an interval is greater/lesser than another one and this possibility (or probabili ty) can be treated as the strength of inequality or (in some sense) as the distance between compared intervals. On the other hand, the above methods can not provide the mea- sure of intervals inequality (distance) when they have no a com- mon area. Of course, the Hamming distance dH ¼ 1 2 ðjaL � bLjþ jaU � bUjÞ: ð18Þ or Euclidean distance dE ¼ 1 2 ððaL � bLÞ2 þðaU � bUÞ2Þ 1 2 ð19Þ can be used as the distance between intervals, but these distances give no information about which interval is greater /lesser. It can be seen that they can not be used directly for interval comparis on especially when an interval is included into another one. Therefore, here we propose to use directly the operation of interval subtracti on (Moore, 1966 ) instead of Hamming and Euclidean distances. This method makes it possible to calculate the possibilit y (or probabili ty) that an interval is greater/l esser that another one when they have a common area and when they do not intersect . So for intervals A = [aL, aU] and B = [bL, bU], the result of subtrac- tion is the interval C = A � B = [cL, cU]; cL = aL � bU, cU = aU � bL. It is easy to see that in the case of overlappi ng intervals A and B, we al- ways obtain a negative left bound of interval C and a positive right bound. Therefore, to get a measure of distance between intervals which additional ly indicate which interval is greater/l esser, we propose here to use the following value: DA�B ¼ 1 2 ððaL � bUÞþðaU � bLÞÞ: ð20Þ It is easy to prove that for intervals with common center, DA�B is al- ways equal to 0. Really, expression (20) may be rewritten as follows: DA�B ¼ 1 2 ðaL þ aUÞ� 1 2 ðbU þ bLÞ � � : ð21Þ We can see that expression (21) represen ts the distance between the centers of compared intervals A and B. This is not a surprising result as Wang et al. (2005) noted that most of the proposed meth- ods for interval comparison are ‘‘totally based on the midpoint s of interv al numbers’’. It easy to see that the result of subtrac tion of interv als with common centers is an interval centered around 0. In the framewo rk of interval analysis, such interval is treated as the interv al 0. Table 2 Results of interval comparison. Method 1 2 3 4 5 6 7 P(A P B) 0 0.06 0.22 0.5 0.78 1 1 P(A 6 B) 1 0.94 0.78 0.5 0.22 0 0 dE 10.82 10 7.81 6 7.81 10 10.82 dH 9 8 6 6 6 8 9 DA�B �9 �8 �5 0 5 8 9 L. Dymova et al. / Expert Systems with Applications 40 (2013) 4841–4847 4845 More strictly, if a is a real value, then 0 can be defined as a � a. Similarly, if A is an interval, then interval zero may be defined as an interval A � A = [aL � aU, aU � aL] which is centered around 0. Therefore, the value of DA�B equal to 0 for A and B having a com- mon center may be treated as a real-valued representat ion of inter- val zero. The similar situation we have in statistics. Let A and B be sam- ples of measure ments with corresponding uniform probability dis- tributions such that they have a common mean (meanA = meanB), but different variances (rA > rB). Then using statistical methods it is impossibl e to prove that the sample B is greater than the sample A or that the sample A is greater than the sample B. Taking into account the above considerati on we can say that the interval comparison based on the assumption that intervals having a common center are equal ones seems to be justified and reasonable. Obviously, the comparison of intervals based on comparison of their centers seems to be too simple. Nevertheles s, as it is shown above, this approach is based directly on conventional operation of interval subtraction. Therefore it is not a heuristic one. Moreover this method coincides better with common sense than more com- plicated known approach es. Let us consider two intervals A = [3, 5] and B = [1, 4]. Since aU > bU and aL > bL, then according to the Moore (1966) and common sense we have A > B. Since A and B are not identical and have no a common center there is no chance for A and B to be equal ones. Finally, according to common sense in this case the possibility of A < B should be equal to 0. There is no chance for B to be greater than A as a whole, although some point belong- ing to B in the common area of A and B may be greater than the points of A in this area. Nevertheles s, in our case from (16) and (17) we get PðB P AÞ¼ 15 and PðA P BÞ¼ 45. Thus, we can see that the known approaches may provide counterintu itive results. In Table 1, we present the values of P(A P B), P(B P A) (see expressions (16), (17)), the Hammin g dH and Euclidean dH dis- tances (see expressions (18), (19)) between Ai and B, and DAi�B for intervals A1 = [4, 7], A2 = [5, 8], A3 = [8, 11], A4 = [13, 16], A5 = [18, 21], A6 = [21, 24], A7 = [22, 25] and B = [7, 22] placed as it is shown in Fig. 1. The numbers in the first row in Table 1 corre- spond to the numbers of intervals Ai, i = 1 to 7. (see Table 2). We can see that the values of DAi�B are negative when Ai 6 B and become positive for Ai P B. These estimates coincide (at least qual- itatively) with P(Ai P B) and P(B P Ai). So we can say that the sign of DAi�B indicates which interval is greater/lesser and the values of absðDAi�BÞ may be treated as the distances between intervals since these values are close the to the values of dE and dH in both cases: Fig. 1. Compared intervals. when intervals have a common area and when the there is no such an area. 3.3. The comparison of the direct interval extension of TOPSIS method with the known method Using DA�B, it is easy to obtain from (14), (15) the ideal interval solutions Aþ ¼f½vþL1 ; v þU 1 �; ½v þL 2 ; v þU 2 �; . . . ; ½v þL n ; v þU n �g; A� ¼f½v�L1 ; v �U 1 �; ½v �L 2 ; v �U 2 �; . . . ; ½v �L n ; v �U n �g: As in the framewo rk of our approac h the distance between interval s A and B is presented by the value of DA�B, there is no need to use Hamming or Euclidean distances for calculatio n of Sþi and S � i . Since DA�B is the subtraction of the midpoints of A and B, the values of Sþi and S � i may be calculated as follows: Sþi ¼ 1 2 X j2K B ððvþLj þ v þU j Þ�ðv L ij þ v U ijÞÞþ 1 2 X j2KC ððv Lij þ v U ijÞ �ðvþLj þ v þU j ÞÞ: ð22Þ S�i ¼ 1 2 X j2K B ððv Lij þ v U ijÞ�ðv �L j þ v �U j ÞÞþ 1 2 X j2KC ððv�Lj þ v �U j Þ �ðv Lij þ v U ijÞÞ: ð23Þ Finally , using expression (13) we obtain the relativ e closeness RCi to the ideal alternativ e. Let us consider some illustrative examples. Example 1. Suppose we deal with three alternatives Ai, i = 1 to 3 and four local criteria Cj, j = 1 to 4 presented by intervals in Table 3, where C1 and C2 are benefit criteria, C3 and C4 are cost criteria. Suppose W = (0.25, 0.25, 0.25, 0.25). To stress the advantag es of our method, in this example many intervals representi ng the val- ues of ratings intersect . Then using the known method for interval extension of TOPSIS method Jahanshahlo et al. (2006) and Jahanshahlo o et al. (2009) (expressions (7)–(13)) we obtain R1 = 0.5311, R2 = 0.6378, R3 = 0.3290 and therefore R2 > R1 > R3, whereas with the use of our method (expressions 7, 8, 22, 23 and 13) we get R1 = 0.7688, R2 = 0.7528, R3 = 0.0717 and therefore R1 > R2 > R3. We can see that there is a considerable difference between the final ranking obtained by the known method and using our method based on the direct extension of TOPSIS method. This can be ex- plained by the fact that the method proposed in Jahansha hlo et al. (2006) and Jahanshahlo o et al. (2009) has some limitatio ns Table 3 Decision matrix. C1 C2 C3 C4 A1 [6, 22] [10, 15] [16, 21] [18, 20] A2 [15, 18] [8, 11] [20, 30] [19, 28] A3 [9, 13] 12, 17] [42, 48] [40, 49] Table 4 Decision matrix. C1 C2 C3 C4 A1 [6, 22] [10, 15] [13, 19] [40, 48] A2 [3, 4] [17, 21] [20, 30] [22, 28] A3 [25, 28] [8, 10] [42, 48] [18, 20] 4846 L. Dymova et al. / Expert Systems with Applications 40 (2013) 4841–4847 concerned with the presentation of intervals by real values in the calculation of ideal solutions and using the Euclidean distance when intervals intersect. Example 2. In this example, there are no intersecting intervals in the columns of decision matrix (see Table 4). As in the previous example, C1 and C2 are benefit criteria, C3 and C4 are cost criteria, W = (0.25, 0.25, 0.25, 0.25). Then using the known method proposed in Jahanshahlo et al. (2006) and Jahanshahloo et al. (2009), in this example we get R1 = 0.4825, R2 = 0.4984, R3 = 0.5413 and therefore R3 > R2 > R1. With the use of our method we obtain R1 = 0.5671, R2 = 0.4675, R3 = 0.4488 and therefore R1 > R2 > R3. Hence, we can conclude that even in the case when there are no intersecting intervals in the col- umns of interval valued decision table, the known method pro- posed in Jahanshahlo et al. (2006) and Jahanshahloo et al. (2009) and our methods may provide very different final rankings of alter- natives. That may be explained by the fact that in the method pro- posed in Jahanshahlo et al. (2006) and Jahanshahlo o et al. (2009), the real-valued ideal solutions are used, whereas in our method they are presented by intervals attainable in the interval-valued decision table. Nevertheles s, when there are no intersecting intervals in the columns of interval-valued decision table, the final ratings ob- tained by the method proposed in Jahanshahlo et al. (2006) and Jahanshahlo o et al. (2009) may coincide with those obtained using our method. Consider the illustrative examples. Example 3. In this example we will use the decision table presented in Table 4, where C1 and C2 are benefit criteria, C3 and C4 are cost criteria, but W = (0.5, 0.1, 0.25, 0.15). Then using the method from Jahanshahlo et al. (2006) and Jahanshahlo o et al. (2009) we obtain R1 = 0.4812, R2 = 0.2232, R3 = 0.5798 and therefore R3 > R1 > R2. Using our method we get R1 = 0.6653, R2 = 0.2758, R3 = 0.6363 and therefore R3 > R1 > R2. Thus, in this case two considered methods provide coincided ratings. Example 4. Let us consider the decision matrix presente d in Table 5, which differs from Table 4 by only one element [x31] so that [x31] intersects with [x11]. There are no other intersection s in the columns of Table 5. Using, as in previous case W = (0.5, 0.1, 0.25, 0.15), and the method from Jahanshahlo et al. (2006) and Jahansha hloo et al. (2009) we get R1 = 0.4812, R2 = 0.25322, R3 = 0.5798 and therefore R3 > R1 > R2 as in the previous example, whereas with the use of our method we obtain R1 = 0.6653, R2 = 0.2758, R3 = 0.6363 and therefore R1 > R3 > R2. Table 5 Decision matrix. C1 C2 C3 C4 A1 [6, 22] [10, 15] [13, 19] [40, 48] A2 [3, 4] [17, 21] [20, 30] [22, 28] A3 [12, 28] [8, 10] [42, 48] [18, 20] Thus, we can see that even the change of only one element in a decision matrix which leads to the appearance of intersecting intervals in the correspondi ng column, may lead to the significant changes in the results obtained by our method, whereas the known method (Jahanshahlo et al. (2006) and Jahansha hloo et al. (2009)) does not provide different final ratings of compared alternatives. 4. Conclusion The critical analysis of known approach es to the interval exten- sion of TOPSIS method is presente d. It is shown that these exten- sions are based on different heuristic approaches to definition of positive and negative ideal solutions. These ideal solutions are pre- sented by real values or intervals, which are not attainable in a decision matrix. Since this is in contradictio n with basics of classical TOPSIS method, a new approach to the solution of MCDM problems with the use of TOPSIS method in the interval setting is proposed. This method called ‘‘direct interval extension of TOPSIS method’’ is free of heuristic limitations of known methods concerned with the def- inition of positive and negative ideal solutions and using the Euclidean distance when intervals in a decision matrix intersect. The main advantage of the proposed method is that (opposite to the known methods) it provides interval-valued positive and neg- ative ideal solutions which in compliance with the basics of classi- cal TOPSIS method are always attainable in the interval-valued decision matrix. It is shown that the use of known methods may lead to the wrong results as well as the use of the Euclidean distance when intervals representi ng the values of local criteria intersect. Using numerical examples, it is shown that the proposed ‘‘direct interval extension of TOPSIS method’’ may provide the final ranking of alternativ es which is substantially different from the results ob- tained using the known methods especially when some interval- valued ratings in the columns of decision table intersect . References Chen, T.-Y. (2011). Interval-valued fuzzy TOPSIS method with leniency reduction and a experimental analysis. Applied Soft Computing, 11, 4591–4606. Facchinetti, G., Ricci, R. G., & Muzzioli, S. (1998). Note on ranking fuzzy triangular numbers. International Journal of Intelligent Systems, 13, 613–622. Garca-Cascales, M. S., & Lamata, M. T. (2012). On rank reversal and TOPSIS method. Mathematical and Computer Modelling, 56, 123–132. Hwang, CL., & Yoon, K. (1981). Multiple Attribute Decision Making–Methods and Applications. Berlin Heidelberg: Springer. Jahanshahlo, G. R., Hosseinzade, L. F., & Izadikhah, M. (2006). An algorithmic method to extend TOPSIS for decision making problems with interval data. Applied Mathematics and Computation, 175 , 1375–1384. Jahanshahloo, G. R., Hosseinzadeh Lotfi, F., & Davoodi, A. R. (2009). Extension of TOPSIS for decision-making problems with interval data:Interval efficiency. Mathematical and Computer Modelling, 49, 1137–1142. Jahanshahloo, G. R., Khodabakhshi, M., Hosseinzadeh Lotfi, F., & Moazami Goudarzi, M. R. (2011). A cross-efficiency model based on super-efficiency for ranking units through the TOPSIS approach and its extension to the interval case. Mathematical and Computer Modelling, 53, 1946–1955. Lai, Y. J, Liu, T. Y., & Hwang, C. L. (1994). TOPSIS for MODM. European Journal of Operational Research, 76, 486–500. Moore, R. E. (1966). Interval analysis . Englewood Cliffs., N.J.: Prentice-Hall. Sayadi, M. K., Heydari, M., & Shahanaghi, K. (2009). Extension o VIKOR method for decision making problem with interval numbers. Applied Mathematical Modelling, 33, 2257–2262. Sevastjanov, P. (2007). Numerical methods for interval and fuzzy number comparison based on the probabilistic approach and Dempster–Shafer theory. Information Sciences, 177 , 4645–4661. Tsaur, R.-C. (2011). Decision risk analysis for an interval TOPSIS method. Applied Mathematics and Computation, 218 , 4295–4304. Wang, Y. M., & Elhag, T. M. S. (2006). Fuzzy TOPSIS method based on alpha level sets with an application to bridge risk assessment. Expert Systems with Applications, 31, 309–319. Wang, X., & Kerre, E. E. (2001). Reasonable properties for the ordering of fuzzy quantities (I) (II). Fuzzy Sets and Systems, 112 , 387–405. Wang, Y. M., & Luo, Y. (2009). On rank reversal in decision analysis. Mathematical and Computer Modelling, 49, 1221–1229. L. Dymova et al. / Expert Systems with Applications 40 (2013) 4841–4847 4847 Wang, Y. M., Yang, J. B., & Xu, D. L. (2005). A preference aggregation method through the estimation of utility intervals. Computers and Operations Research, 32, 2027–2049. Wang, Y. M., Yang, J. B., & Xu, D. L. (2005). A two-stage logarithmic goal programming method for generating weights from interval comparison matrices. Fuzzy Sets and Systems, 152 , 475–498. Xu, Z., & Da, Q. (2002). The uncertain OWA operator. International Journal of Intelligent Systems, 17, 569–575. Xu, Z., & Chen, J. (2008). Some models for deriving the priority weights from interval fuzzy preference relations. European Journal of Operational Research, 184 , 266–280. Ye, F., & Li, Y. N. (2009). Group multi-attribute decision model to partner selection in the formation of virtual enterprise under incomplete information. Expert Systems with Applications, 36, 9350–9357. Yue, Z. (2011). An extended TOPSIS for determining weights of decision makers with interval numbers. Knowledge-Based Systems, 24, 146–153. A direct interval extension of TOPSIS method 1 Introduction 2 The basics of TOPSIS method and known approaches to its interval extension 3 A new approach to the interval extension of TOPSIS method 3.1 The problem formulation 3.2 The methods for interval comparison 3.3 The comparison of the direct interval extension of TOPSIS method with the known method 4 Conclusion References 
work_g7vncqg54jbtlp6lcpqhvwtney	----	Jurnal Ilmu Komputer dan Informasi (Journal of Computer Science and Information). 7/2 (2014), 61-66 DOI: http://dx.doi.org/10.21609/jiki.v7i2.258 DIVERSITY-BASED ATTRIBUTE WEIGHTING FOR K-MODES CLUSTERING M Misbachul huda, Dian Rahma Latifa Hayun, and Annisaa Sri I. Informatics Engineering, information Technology Department Institut Teknologi Sepuluh Nopember Surabaya, East Java, Indonesia E-mail: annisaaindrawanti@gmail.com, misbachul@mhs.if.its.ac.id Abstract Categorical data is a kind of data that is used for computational in computer science. To obtain the information from categorical data input, it needs a clustering algorithm. There are so many clustering algorithms that are given by the researchers. One of the clustering algorithms for categorical data is k- modes. K-modes uses a simple matching approach. This simple matching approach uses similarity va- lues. In K-modes, the two similar objects have similarity value 1, and 0 if it is otherwise. Actually, in each attribute, there are some kinds of different attribute value and each kind of attribute value has different number. The similarity value 0 and 1 is not enough to represent the real semantic distance between a data object and a cluster. Thus in this paper, we generalize a k-modes algorithm for catego- rical data by adding the weight and diversity value of each attribute value to optimize categorical data clustering. Keywords: categorical data, diversity, K-modes, attribute weighting. Abstrak Data Kategorial merupakan suatu jenis data perhitungan di ilmu komputer .Untuk mendapatkan infor- masi dari input data kategorial diperlukan algoritma klastering. Ada berbagai jenis algoritma klas- tering yang dikembangkan peneliti terdahulu. Salah satunya adalah K-modes. K-modes menggunakan pendekatan simple matching. Pendekatan simple matching ini menggunakan nilai similarity. Pada K- modes, jika dua objek data mirip, maka akan diberi nilai. Jika dua objek data tidak mirip, maka diberi nilai 0. Pada kenyataannya, tiap atribut data terdiri dari beberapa jenis nilai atribut dan tiap jenis nilai atribut terdiri dari jumlah yang berbeda. Nilai similarity 0 dan 1 kurang merepresentasi jarak antara sebuah objek data dan klaster secara nyata. Oleh karena itu, pada paper ini, kami mengembangkan algoritma K-modes untuk data kategorial dengan penambahan bobot dan nilai diversity pada setiap atribut untuk mengoptimalkan klastering data kategorial. Kata Kunci: data kategorial, diversity, K-modes, pembobotan atribut. 1. Introduction Computer science is always related to the data. All computational processes need data not only small data but also big data. There are two data types, they are numeric data and categorical data. Arithmetic process can be given to numeric data but it cannot be given to categorical data. Catego- rical data is used in some systems or applications. For example, categorical data in intrusion detec- tion systems, population data and customer infor- mation in online shopping. Example of categorical data in intrusion detection system is IP address. IP address in each data must be different and it can- not be arithmetically compared. It will be in popu- lation data and customer information, too. Popu- lation data has such categorical data like gender, blood type and home address. Costumer informa- tion, in online shopping, has such categorical data. For example: the phone number and the used ba- nk. The categorical data clustering considers the similarity or dissimilarity between data. Similarity or dissimilarity can be considered by distance bet- ween the two data object. The shorter the distance between objects, the more similar the objects. Si- mple matching approaches can be used to calcu- late the distance. Example of simple matching ap- proach for categorical data is k-modes [7]. In in- trusion detection system, there is such new intru- sion that has not been known earlier. It needs a clustering algorithm to cluster and detect the new intrusion. Clustering is used to cluster population data and customer data to obtain the information needed, too. The study [7], k-means and k-modes are joi- ned together to cluster numeric and categorical data. K-means is a clustering algorithm for nume- ric data. K-means clusters the data type that can 61 62 Jurnal Ilmu Komputer dan Informasi (Journal of Computer Science and Information), Volume 7, Issue 2, June 2014 be arithmetically compared. Categorical data is a kind of data that cannot be arithmetically compar- ed each other. K-means cannot be used to cluster categorical data. To overcome the categorical da- ta, it uses k-modes algorithm. K-modes algorithm uses a simple matching approach to the process of matching dissimilarity clustering of data, replac- ing the means to modes. In k-means, to update a cluster centroid, it is used means formulae. On the other hand the k-modes, modes updating on the clustering process is determined by the frequency of occurrence of data so as to minimize the cost function. In [3], K-modes algorithm is supposed to ha- ve some deficiencies. In K-modes, the two similar objects have similarity value 1, and 0 if it is oth- erwise. According to [3], a simple matching app- roach 0 and 1 are not good enough to represent the real semantic distance between a data object and a cluster. Thus, in [3] the authors proposed a range between 0 and 1 which represents the wei- ght of similarity. In [8], focused on optimizing the k-modes algorithm for clustering categorical data with the new dissimilarity approach using the rou- gh set membership. According to the research, the shortcoming of the simple matching approach is having a weak intra-similarity. To solve the issue, the dissimilarity approach (frequency-based) bet- ween the two objects in [8] takes into account to the distribution of universal attribute values. The study [1] explains that sometimes the occurrence of low frequency data and high frequency data ha- ve the same overlap similarity value. With these problems, the study [1] approach takes into acco- unt to the similarity of the frequency distribution of the different attributes value to define the simi- larity between two categorical attribute values. In addition, the dissimilarity approach taking into account the frequency distribution of differ- ent attribute values can also be applied to the opti- mization of categorical data dissimilarity [1]. For example, there are two data, AAABB and ABC- DD. If using the k-modes algorithm [8], both dis- similarity values are equal to 1. As we can see, the diversity level between the first data and the se- cond data is different because the quantity of each letter is different. Thus, in this paper, the research focuses on the k-modes clustering with diversity- based attribute weighting and the research contri- bution is distance values diversity to optimize the categorical data dissimilarity. 2. Methods Categorical Data Categorical variables represent types of data whi- ch may be divided into groups. Analysis of cate- gorical data generally involves the use of data ta- bles. A two-way table presents categorical data by counting the number of observations that fall into each group for two variables, one divided into ro- ws and the other divided into columns. There is no intrinsic ordering to the catego- ries in categorical data. For example, gender is a categorical variable having two categories (male and female) and there is no intrinsic ordering to the categories. Hair color is also a categorical va- riable having a number of categories (blonde, bro- wn, brunette, red, etc.) and again, there is no agreed way to order these from highest to lowest. So that, computing the similarity and the dis- similarity between categorical data instances is not straightforward owing to the fact that there is no explicit notion of ordering between categorical values. To tackle this, there are some data-driven similarity measures that have been proposed for categorical data. In this section, we will describe the categorical data characteristics. But first, we will define the notation used in later explanation. Let’s consider for a categorical data set D containing N objects, defined over a set of l attribute where 𝐴𝐴𝑖𝑖 denotes the ith attribute. And let the ni be the number of values in each at- tribute𝐴𝐴𝑖𝑖. The next notation used are following: 𝑓𝑓𝑖𝑖(𝑥𝑥) = the frequency of value 𝑥𝑥 lies on attribute Ai in data set D. 𝑝𝑝𝑖𝑖(𝑥𝑥) = the probability of attribute 𝐴𝐴𝑖𝑖 to take the value 𝑥𝑥 in the data set D, defined as equation(1). 𝑝𝑝𝑖𝑖(𝑥𝑥) = 𝑓𝑓𝑖𝑖(𝑥𝑥) 𝑁𝑁 (1) 𝑝𝑝𝑖𝑖 2(𝑥𝑥) = another probability measure of attribute 𝐴𝐴𝑖𝑖 to take the value 𝑥𝑥 in the data set D, defined as equation(2). 𝑝𝑝𝑖𝑖 2(𝑥𝑥) = 𝑓𝑓𝑖𝑖 (𝑥𝑥)(𝑓𝑓𝑖𝑖(𝑥𝑥)−1) 𝑁𝑁(𝑁𝑁−1) (2) Size of the data, N. Most measure are typi- cally invariant to the size of the data. Number of attributes, l. In [1] experiment showed that the nu- mber of attributes does affect the performance of the outlier detection algorithm. The frequency of values taken by each attri- bute, ni. A data set may contain attributes that take some values and attributes that only take a few va- lues. A similarity measure may ignore another at- tribute while finding the more importance attri- bute. Distribution of 𝑓𝑓𝑖𝑖(𝑥𝑥). It refers to the distri- bution of frequency of values taken by attribute in the given data set. It is possible for a similarity measure to give more priority to frequently occu- rring attribute values or otherwise. M. Misbachul Huda, et a.l, Diversity-Based Attribute 63 Similarity and Dissimilarity Measure for Cate- gorical Data in K-Modes The classical theory of clustering in k-Modes, eit- her an element belongs to a cluster or it does not. The corresponding membership function is the characteristic function of the cluster that takes values 1 and 0. Suppose that there are 𝑥𝑥,𝑦𝑦 ∈ 𝐷𝐷, then the simple matching dissimilarity measure in k-Modes is defined following equation(3). 𝐷𝐷𝐷𝐷𝐷𝐷(𝑥𝑥, 𝑦𝑦) = � 1, 𝐷𝐷𝑓𝑓 𝑥𝑥 ≠ 𝑦𝑦 0, 𝑜𝑜𝑜𝑜ℎ𝑒𝑒𝑒𝑒𝑒𝑒𝐷𝐷𝐷𝐷𝑒𝑒 (3) With this concept, the distance between two objects computed often results in cluster with we- ak intra-similarity and disregards the similarity embedded in the categorical values [1]. Then, a new dissimilarity measure between the mode of a cluster and an object is introduced for the k-Modes Algorithm. To obtain such a clus- ter having the strong intra-similarity, the rough membership is used. It takes values between 0 and 1. Taking account of the frequency of mode components in the current cluster, Ng et al. [2] in- troduced a valuable dissimilarity measure into the k-Modes clustering algorithm. Let 𝑃𝑃 ⊆ 𝐴𝐴 and 𝑎𝑎 ∈ 𝑃𝑃, then the Ng dissimilarity measure 𝐷𝐷𝐷𝐷𝐷𝐷𝑃𝑃(𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖) between a categorical object 𝑥𝑥𝑖𝑖 and the mode of a cluster 𝑧𝑧𝑙𝑙 with respect to P is defined as the following equation(4). 𝐷𝐷𝐷𝐷𝐷𝐷𝑃𝑃(𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖) = �𝐷𝐷𝐷𝐷𝐷𝐷𝑎𝑎(𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖), 𝑎𝑎∈𝑃𝑃 where, 𝐷𝐷𝐷𝐷𝐷𝐷𝑎𝑎(𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖) = � 1, 𝐷𝐷𝑓𝑓 𝑓𝑓(𝑧𝑧𝑙𝑙, 𝑎𝑎) ≠ 𝑓𝑓(𝑥𝑥𝑖𝑖, 𝑎𝑎), 1 − 𝑚𝑚𝑎𝑎, 𝑜𝑜𝑜𝑜ℎ𝑒𝑒𝑒𝑒𝑒𝑒𝐷𝐷𝐷𝐷𝑒𝑒 where, 𝑚𝑚𝑎𝑎 = |{𝑥𝑥𝑖𝑖|𝑓𝑓(𝑥𝑥𝑖𝑖, 𝑎𝑎) = 𝑓𝑓(𝑧𝑧𝑙𝑙, 𝑎𝑎), 𝑥𝑥𝑖𝑖 ∈ 𝑐𝑐𝑙𝑙 }| |𝑐𝑐𝑙𝑙| (4) And |𝑐𝑐𝑙𝑙| is the number of objects in the lth cluster. For the k-Modes algorithm with Ng’s dis- similarity measure [2], the simple matching dissi- milarity measure is still used in the first iteration. So, Cao et al. in [3] proposed a new dissimilarity measure by using 𝑆𝑆𝐷𝐷𝑚𝑚𝑎𝑎(𝑥𝑥, 𝑦𝑦) defined in equation (5). 𝑆𝑆𝐷𝐷𝑚𝑚𝑎𝑎(𝑥𝑥, 𝑦𝑦) = 𝑓𝑓(𝑥𝑥, 𝑎𝑎) ≡ 𝑓𝑓(𝑦𝑦, 𝑎𝑎) ∑ 𝑓𝑓(𝑥𝑥, 𝑎𝑎) ≡ 𝑓𝑓(𝑧𝑧, 𝑎𝑎)𝑧𝑧∈𝐷𝐷 (5) From the equation(5) formula [3] introduced that the new dissimilarity measure is defined as the following equation(6). 𝑁𝑁𝐷𝐷𝐷𝐷𝐷𝐷𝑃𝑃(𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖) = ∑ 𝑁𝑁𝐷𝐷𝐷𝐷𝐷𝐷𝑎𝑎𝑎𝑎∈𝑃𝑃 (𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖) (6) where, 𝑁𝑁𝐷𝐷𝐷𝐷𝐷𝐷𝑎𝑎(𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖) = 1 − 𝑆𝑆𝐷𝐷𝑚𝑚𝑎𝑎(𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖) × 𝑚𝑚𝑎𝑎 As opposed to Ng’s dissimilarity measure, the similarity 𝑆𝑆𝐷𝐷𝑚𝑚𝑎𝑎(𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖) between object 𝑥𝑥𝑖𝑖and cluster 𝑧𝑧𝑙𝑙 is included in the proposed measure 𝑁𝑁𝐷𝐷𝐷𝐷𝐷𝐷𝑎𝑎(𝑧𝑧𝑙𝑙, 𝑥𝑥𝑖𝑖). In [3] introduced a weighted k- modes clustering algorithm, by considering that the similarity value between two objects is not al- ways 1, but it can be a value between 0 and 1. A pair of object 𝑥𝑥𝑖𝑖, 𝑥𝑥𝑗𝑗 is considered more similar th- an the second pair(𝑥𝑥𝑠𝑠, 𝑥𝑥𝑡𝑡), if and only if 𝑥𝑥𝑖𝑖and 𝑥𝑥𝑗𝑗 exhibit a less common attribute value match in the population. In other words, similarity among obje- cts could be decided by the un-commonality of their attribute value matches. Based on that, in [3] defined the “More Similar Attribute Set” of an at- tribute value 𝑎𝑎𝑗𝑗 (𝑟𝑟) as the following equation(7). 𝑀𝑀�𝑎𝑎𝑗𝑗 (𝑟𝑟)� = {𝑎𝑎𝑗𝑗 (𝑡𝑡)|𝑓𝑓�𝑎𝑎𝑗𝑗 (𝑡𝑡)|𝐷𝐷� ≤ 𝑓𝑓(𝑎𝑎𝑗𝑗 (𝑟𝑟)|𝐷𝐷)} (7) where 𝑓𝑓(𝑎𝑎𝑗𝑗 (𝑡𝑡)|𝐷𝐷) is the frequency count of attribute 𝑎𝑎𝑗𝑗 (𝑡𝑡) in the data set D. This is the set of attribute values with lower or equal frequencies of occu- rrence than that of 𝑎𝑎𝑗𝑗 (𝑟𝑟). Note that a value pair is more similar if it has lower frequency of occur- rence. The weighting function in [3] is defined as equation(8). 𝜔𝜔�𝑎𝑎𝑗𝑗 (𝑟𝑟)� = 1 − ∑ 𝑓𝑓(𝑎𝑎𝑗𝑗 (𝑡𝑡)|𝐷𝐷)(𝑓𝑓�𝑎𝑎𝑗𝑗 (𝑡𝑡) �𝐷𝐷�−1) 𝑛𝑛(𝑛𝑛−1)𝑎𝑎𝑗𝑗 (𝑡𝑡)∈𝑀𝑀(𝑎𝑎𝑗𝑗 (𝑟𝑟)) (8) where n is the frequency of objects in the data set D. If above function is used, less frequent values will make more contributions to the similarity va- lue. Diversity Index and Variable Uniqueness Although the above algorithms can effectively im- prove the accuracy of the clustering result of the k-modes algorithm, it is noted that the k-modes al- gorithm and its modified version cannot detect the diversity of the data in a certain attribute in data set. It is easily prove in simple matching similarity measure. Because in the value will be 1 if it is ide- ntical, and 0 if otherwise. In wk-modes, the basic 64 Jurnal Ilmu Komputer dan Informasi (Journal of Computer Science and Information), Volume 7, Issue 2, June 2014 TABLE 1 DATA SAMPLE Dataset/ object 𝑥𝑥1 𝑥𝑥2 𝑥𝑥3 𝑥𝑥4 𝑥𝑥5 Div 𝐷𝐷1 A A A A A 1 𝐷𝐷2 A A A A B 2 𝐷𝐷3 A A A B B 2 𝐷𝐷4 A A B B C 3 𝐷𝐷5 A A B C D 4 𝐷𝐷6 A B C C C 3 𝐷𝐷7 A B C D E 5 concept used is using individual Simpson diver- sity index (will be explained in the next part). By using this information, it is noted that this algo- rithm cannot detect the diversity of data in data set. Let us see Table 1 as the data sample of coun- ting the dissimilarity between two objects. Let the number of attribute in the given data set is only 1, and the number of data set is 7. The number of object in a dataset is 5. From the table we have the diversity value (Div) as variable uniqueness sho- wn in the table. The value of Div show us the number of dis- tinct value in a dataset. We assume that by using this Div value it will increase the dissimilarity between clusters based on the diversity of attribu- te value. From the Table 1, we will get 1 as the value of dissimilarity because each data is different. But, as we can see in the real in each data, there are some similar data in it. But, they have differ- rent diversity. For AAAAA, the diversity value is 1 because there’s just a single object data. But for ABCDE, the diversity value is 5 because actually there are 5 different data. Based on this diversity, we propose a new method to count the dissimi- larity measure to count the distance between two distinct objects. We use this diversity value to de- crease the inter-similar cluster value. Diversity index is formerly a concept in bio- logical area. It is a mathematical measure of spe- cies diversity in a community. Diversity indices provide more information about community com- position than simply species richness (i.e., the number of species present); they also take the re- lative abundances of different species into account [4]. Diversity indices provide important informati- on about rarity and commonness of species in a community. The ability to quantify diversity in th- is way is an important tool for biologists trying to understand community structure. One of the commonly used diversity index measures is Simpson Diversity Index (SDI). The basic concept of SDI represents the measurement of dissimilarity by frequency based. Simpson's Di- versity Index is a measure of diversity. In ecolo- gy, it is often used to quantify the biodiversity of a habitat. It takes into account the number of speci- es present, as well as the abundance of each spe- cies. Simpson's Index (D) measures the probabi- lity that two individuals randomly selected from a sample will belong to the same species (or some category other than species). The formula is defi- ned as below. 𝐷𝐷 = ∑𝑛𝑛(𝑛𝑛 − 1) 𝑁𝑁(𝑁𝑁 − 1) (9) where n is the total number of organisms of a par- ticular species, and N is the total number of orga- nisms of all species. Simpson's Index gives more weight to the more abundant species in a sample. The addition of rare species to a sample causes only small changes in the value of D [5]. In clus- tering theory, the property n can be defined as the number of object 𝑥𝑥𝑖𝑖 in a dataset D, and the pro- perty N can be defined as the number of all ob- jects in data set D. Diversity-based Attribute Weighting for K- Modes Clustering Taking into account of the intra and inter cluster information, we introduce the new dissimilarity measure in the equation(10). 𝐷𝐷𝐷𝐷𝐷𝐷𝑃𝑃�𝑥𝑥𝑖𝑖,𝑦𝑦𝑗𝑗� = 𝑆𝑆𝐷𝐷𝑆𝑆𝑖𝑖 × 𝐷𝐷𝐷𝐷𝐷𝐷 × 𝑛𝑛𝑖𝑖 𝑁𝑁 × 𝑆𝑆𝐷𝐷𝑆𝑆𝑗𝑗 × 𝐷𝐷𝐷𝐷𝐷𝐷 × 𝑛𝑛𝑗𝑗 𝑁𝑁 (10) where, 𝑆𝑆𝐷𝐷𝑆𝑆𝑖𝑖 is simpson diversity index in data set D for object 𝑥𝑥𝑖𝑖, 𝑆𝑆𝐷𝐷𝑆𝑆𝑗𝑗 is simpson diversity index in data set D for object 𝑦𝑦𝑗𝑗, 𝐷𝐷𝐷𝐷𝐷𝐷 is the number of category in data set D, 𝑛𝑛𝑖𝑖 is the frequency of 𝑥𝑥𝑖𝑖 in a cluster, 𝑛𝑛𝑗𝑗 the frequency of 𝑦𝑦𝑗𝑗 in a cluster, 𝑁𝑁 is the number of data in data set D. This formula is derived from the basic concept of probability count of object 𝑥𝑥𝑖𝑖 as defined below. 𝑝𝑝(𝑥𝑥𝑖𝑖) = 𝑁𝑁𝑥𝑥𝑖𝑖 𝑁𝑁 (11) where, 𝑝𝑝�𝑥𝑥𝑖𝑖,� is the probability of 𝑥𝑥𝑖𝑖, 𝑁𝑁𝑥𝑥𝐷𝐷 is the frequency of 𝑥𝑥𝑖𝑖, and 𝑁𝑁 is the number of data in data set D. In order to count the dissimilarity between two objects 𝑥𝑥𝑖𝑖,𝑦𝑦𝑗𝑗 , we modify the above formula as defined in equation(12). 𝐷𝐷𝐷𝐷𝐷𝐷𝑃𝑃�𝑥𝑥𝑖𝑖, 𝑦𝑦𝑗𝑗� = 𝑁𝑁𝑥𝑥𝑖𝑖 𝑁𝑁 × 𝑁𝑁𝑦𝑦𝑗𝑗 𝑁𝑁 (12) M. Misbachul Huda, et a.l, Diversity-Based Attribute 65 TABLE 2 SCALABILITY EXPERIMENTAL RESULT Pengujian Scalability Number of objects Diversity based Time(ms) Original Time (ms) 10 4 3 50 10 10 100 18 18 500 135 124 1000 737 585 5000 38010 18984 TABLE 3. DATA SET CHARACTERISTIC Dataset Class number Number of data Number of attributes Voting 2 435 16 Mushroom 2 8124 23 Soybean 4 47 36 TABLE 4 CLUSTERING EFFICIENCY RESULT Voting Mushroom Soybean Average Original K-Modes 0.8592 0.7381 0.8177 0.805 wk- Modes 0.8651 0.7905 0.897 0.850867 Zengyou k-modes 0.8734 0.7644 0.86 0.8326 Diversity k-modes 0.8861 0.7849 0.887 0.852667 Figure 1. Scalability Graphic between original k-modes and diversity-based k-modes where, 𝐷𝐷𝐷𝐷𝐷𝐷𝑃𝑃�𝑥𝑥𝑖𝑖, 𝑦𝑦𝑗𝑗�is the dissimilarity of 𝑥𝑥𝑖𝑖, 𝑦𝑦𝑗𝑗in attribute P, 𝑁𝑁𝑥𝑥𝑖𝑖is the number of object 𝑥𝑥𝑖𝑖in data set D, 𝑁𝑁𝑦𝑦𝑗𝑗is the number of object 𝑦𝑦𝑗𝑗in data set D and 𝑁𝑁is the number of data in data set D. To increase the accuracy of clustering pro- cess, we add the Simpson diversity index that has defined in the previous to the formula above. Our new proposing dissimilarity measure is defined as the following equation(13). 𝐷𝐷𝐷𝐷𝐷𝐷𝑃𝑃�𝑥𝑥𝑖𝑖,𝑦𝑦𝑗𝑗� = 𝑆𝑆𝐷𝐷𝑆𝑆𝑖𝑖 × 𝐷𝐷𝐷𝐷𝐷𝐷 × 𝑛𝑛𝑖𝑖 𝑁𝑁 × 𝑆𝑆𝐷𝐷𝑆𝑆𝑗𝑗 × 𝐷𝐷𝐷𝐷𝐷𝐷 × 𝑛𝑛𝑗𝑗 𝑁𝑁 (13) where, 𝑆𝑆𝐷𝐷𝑆𝑆𝑖𝑖is simpson diversity index in data set D for object 𝑥𝑥𝑖𝑖, 𝑆𝑆𝐷𝐷𝑆𝑆𝑗𝑗is simpson diversity index in data set D for object 𝑦𝑦𝑗𝑗, 𝐷𝐷𝐷𝐷𝐷𝐷is the number of category in data set D, 𝑁𝑁is the number of data in data set D. 3. Results and Analysis In this part, we will discuss about the experiment- tal result of scalability and efficiency of clustering algorithms. It employs three different algorithms. In first part of this section, we will explain the ex- periment environment and evaluation index. In se- cond part, we will explain the experimental result of scalability of original k-modes and our propos- ed method. And in the last part, we will show the experimental result of clustering efficiency of four different modified k-modes algorithm include our proposed method. Experiment Environment and Evaluation In- dex The experiment was done in core i3 (1.4 GHz), 4GB RAM, and windows 8.1x64 based computer. The implementation of the algorithm uses Java as the programming language. To evaluate the accu- racy of the algorithm, we use the following equa- tion(14). 𝑎𝑎𝑐𝑐𝑐𝑐 = ∑ 𝑎𝑎𝑖𝑖𝑘𝑘𝑖𝑖=1 𝑛𝑛 (14) where k is the number of known categories, 𝑎𝑎𝑖𝑖 is the number of object that lies in the right cluster in 𝐶𝐶𝑖𝑖(1 ≤ 𝐷𝐷 ≤ 𝑘𝑘). Scalability Evaluation To compare the scalability of original k-modes and our proposed method, we use synthetic data with the variant number of objects between 10 and 5000. This synthetic data have 23 different at- tributes. To decrease the random effect of k-mo- des algorithm, we did 100 times experiment for every scenario. For each experimental result shown in Table 2, is the average value of the 100 times experi- mental result. From these experiments, we get a scalability graphic shown in Figure 1. The experimental result of runtime from di- versity-based k-modes show that it needs longer time than the original k-modes to evaluate the gi- 66 Jurnal Ilmu Komputer dan Informasi (Journal of Computer Science and Information), Volume 7, Issue 2, June 2014 ven data set. It is caused by added step to compute the diversity of the data set. Clustering Efficiency Evaluation In this section, to compare the efficiency cluste- ring algorithms, we use four different modified k- modes algorithm include our proposed method. They are Simple-Matching k-modes, Zengyou’s k-modes, wk-Modes and our new proposed me- thod diversity-based k-modes. We use tree differ- rent data set from UCI Machine Learning [6]. Table 3 shows the characteristic of the data set. In experiments, the missing value is ignored. To decrease the random effect in k-modes, every experiment was conducted 100 times. Table 4 sh- ows the result of clustering efficiency evaluation between the four algorithms. Every value is the average of 100 times experiments result. 4. Conclusion The k-modes algorithm is widely used for clus- tering categorical data. Dissimilarity and simila- rity measure play crucial rules in this area. In this paper the limitation of the former algorithm on the usage of between cluster information is solved by using simpson diversity index to extend the value of the intra similarity index. The experimental re- sult shows that our proposed algorithm give better result. References [1] Shyam Boriah, Varun Chandola, Vipin Ku- mar. Similarity Measures for Categorical Data: A Comparative Evaluation. Depart- ment of Computer Science and Engineering University of Minnesota. 2008s [2] M.K. Ng, M.J. Li, Z.X. Huang, Z.Y. He, on the impact of dissimilarity measure in k-Mo- des clustering algorithm, IEEE Transactions on Pattern Analysis and Machine Intelli- gence 29 (3) (2007) 503–507. [3] Zengyou He, Xiaofei Xu, Shengchun Deng. Attribute value weighting in k-modes cluster- ing. Elsevier. 2011 [4] M. Beals, L. Gross, and S. Harrell, http://ww w.tiem.utk.edu/~gross/bioed/bealsmodules/s hannonDI.html, 2000, retrieved June 1, 2014 [5] Offwell Woodland & Wildlife Trust http:// www.countrysideinfo.co.uk/simpsons.htm., 2000, retrieved June 1, 2014 [6] UCI Machine Learning Repository http://ww w.ics.uci.edu/mlearn/MLRepository.html, 2009, retrieved June 1, 2014. [7] Huang, Z.J. Extension to the k-means algori- thm for clustering large data sets with catego- rical values. Data mining Knowledge Discov- ery, Vol. 2, No. 3, pp.283-304. 1998. [8] Fuyuan Cao, Jiye Liang, Deyu Li, Liang Bai, Chuangyin Dang. A dissimilarity measure for the k-Modes clustering algorithm. Elsevier. 2012. 
work_gavjto52cfb4hdzbfvxl4spuky	----	NASA Technical Reports Server (NTRS) 
work_gbi3ah5ymfdatpixrkkboizbqy	----	Northumbria Research Link Citation: Yang, Longzhi, Li, Jie, Chao, Fei, Hackney, Philip and Flanagan, Mark (2018) Job Shop Planning and Scheduling for Manufacturers with Manual Operations. Expert Systems. ISSN 1468- 0394 Published by: Wiley URL: https://onlinelibrary.wiley.com/journal/14680394 <https://onlinelibrary.wiley.com/journal/14680394> This version was downloaded from Northumbria Research Link: http://nrl.northumbria.ac.uk/34299/ Northumbria University has developed Northumbria Research Link (NRL) to enable users to access the University’s research output. Copyright © and moral rights for items on NRL are retained by the individual author(s) and/or other copyright owners. Single copies of full items can be reproduced, displayed or performed, and given to third parties in any format or medium for personal research or study, educational, or not-for-profit purposes without prior permission or charge, provided the authors, title and full bibliographic details are given, as well as a hyperlink and/or URL to the original metadata page. The content must not be changed in any way. Full items must not be sold commercially in any format or medium without formal permission of the copyright holder. The full policy is available online: http://nrl.northumbria.ac.uk/pol i cies.html This document may differ from the final, published version of the research and has been made available online in accordance with publisher policies. To read and/or cite from the published version of the research, please visit the publisher’s website (a subscription may be required.) http://nrl.northumbria.ac.uk/policies.html O R I G I N A L A R T I C L E Journal Section Job Shop Planning and Scheduling for Manufacturers with Manual Operations Longzhi Yang1∗ | Jie Li1 | Fei Chao2 | Phil Hackney1 | Mark Flanagan3 1Department of Computer and Information Sciences, Northumbria University, NE1 8ST, UK 2Cognitive Science Department, Xiamen University, Xiamen, China 3NHS Business Services Authority, Newcastle upon Tyne, NE15 8NY, UK Correspondence Longzhi Yang, Department of Computer and Information Sciences, Northumbria University, NE1 8ST, UK Email: longzhi.yang@northumbria.ac.uk Funding information National Natural Science Foundation of China, No. 91746103 Conflict of interest Conflicts of interest: none Job shop scheduling systems are widely employed to opti- mise the efficiency of machine utilisation in the manufac- turing industry, by searching the most cost-effective per- mutation of job operations based on the cost of each op- eration on each compatible machine and the relations be- tween job operations. Such systems are paralysed when the cost of operations are not predictable led by the involve- ment of complex manual operations. This paper proposes a new genetic algorithm-based job shop scheduling system by integrating a fuzzy learning and inference sub-system in an effort to address this limitation. In particular, the fuzzy sub-system adaptively estimates the completion time and thus cost of each manual task under different conditions based on a knowledge base which is initialised by domain experts and then constantly updated based on its built-in learning ability and adaptability. The manufacturer of Point of Sale and Point of Purchase products is taken in this pa- per as an example case for both theoretical discussion and experimental study. The experimental results demonstrate the promising of the proposed system in improving the ef- ficiency of manual manufacturing operations. K E Y W O R D S Job shop planning and scheduling, fuzzy learning and inference system, fuzzy rule interpolation, genetic algorithm, manual 1 2 Longzhi Yang et al. operation scheduling 1 | INTRODUCTION The manufacturing sector thrives on implementing efficiency initiatives through measuring and improving Key Perfor- mance Indicators (KPIs). Manufacturing heuristics that help senior management decision making are highly praised, particularly if they can deal with the inherent complexity of noisy shop floor data affecting planning scenarios. Man- ufacturing efficiency can be significantly improved by employing an intelligent job shop scheduling (JSS) system as shown by Operation Research (OR) and Artificial Intelligence (AI) solutions using combinatorial optimisations aiming to reduce the cost based on a given cost function, such as making span or economic cost (Çaliş and Bulkan (2015); Chaudhry and Khan (2016)). The success of intelligent JSS systems relies on an accurate estimation of cost, such as time of each individual operation on each discrete functional processing step at each operation/work centre, when the making span is taken as the cost function. Despite the machine processing time of each operation being readily estimable, based on the nature of the job and the characteristics of the machine, it is difficult to accurately estimate the completion time of whole series of complex manual operations, which often disables the JSS systems. For instance, the multiple complex manual operations of the bespoke print industry produce Point of Sale (POS) and Point of Purchase (POP) products, with particular emphasis on the complexity of 3D display products, and hard to estimate in advance especially for large seasonal shop promotions. An individual manufacturing job may require tens of operations, utilising several different machines coupled with manual tasks in different operation centres, and many of these jobs may need to be processed in parallel at any single time. Parallel lines of manual operations (or “lanes") are indeed common and expedient where possible; these “lanes" need to be planned alongside the job estimation for individual atomic operations such as the ‘time to glue’, ‘time to cut a shape’, ‘time to consolidate’. The scheduler plans the lanes to conform to the physical space available and sensible groupings of manual tasks. The existing system fails once the manual processes take precedent. The paralysed JSS fails the business in two ways and could lead to late jobs or additional ‘same day’ despatch costs incurred. The Sales Team is denied an accurate and realistic long-term schedule of operations and capacity management. Quotations are made available to customers simply based on the nature of the job rather than a reflec- tion of demand-supply relationship. The danger being to overcommit which exceeds the actual capacity and leads to additional problems in the scheduling having to content with unnecessary Work-In-Progress (WIP). Secondly, the ef- fective ‘management planning horizon’ is artificially shortened to a period as short as a week, instead of maintaining a strategic month or quarter perspective. This creates a cycle of “flood then famine” as a period of frantic hyper activity is replaced with insufficient work and quiet factory with a frustrated Sales Team. This common issue exists currently in most manufacturers where problems are seen in despatch but caused by planning, scheduling and estimating short comings. No despatch orientated solution exists as the answer lays in empowering the scheduler with stronger tools. This paper proposes a complete manual job scheduling solution to address these limitations using a genetic al- gorithm (GA) and an adaptive fuzzy inference system by further developing the seminal work reported in Yang et al. (2017b). The fuzzy inference sub-system accurately estimates the completion time of a manual operation under different situations. Through lack of viable data on manual operations and thus the exclusion of data-drive rule base generation such as Chen et al. (2018), the rule base of the fuzzy inference sub-system is initialised by expertise knowl- edge. The rule-base is then developed dynamically and adaptively upon its deployment whilst performing inferences. By taking the accurate completion time estimation as inputs, the GA calculates the optimal scheduling by considering not only meeting all the job deadlines, but also minimising the overall cost of all jobs. To demonstrate a ‘proof of Longzhi Yang et al. 3 principle’, a series of experiments were performed on data that simulate real world loading scenarios common in the manufacturing environment. The experimental results demonstrate the power of the proposed system in improving overall manufacturing efficiency. The rest of the paper is organised as follows. Section 2 briefs the theory underpinning of the proposed system. Section 3 overviews the proposed global scheduling system. Section 4 details the fuzzy inference and learning system in manual task completion time estimation. Section 5 presents the proposed manual job shop planning and scheduling system. Section 6 shows the effectiveness of the proposed approach through illustrative and simulated examples. Section 7 concludes the paper with future work highlighted. 2 | BACKGROUND The two integral components of the proposed system, that is the JSS system and the experience-based fuzzy inference system are reviewed in this section. 2.1 | Job Shop Scheduling JSS is a classical operation research problem or combinational optimisation problem, which has been extensively stud- ied by researchers in the fields of Operation Research and Artificial Intelligence (Pinedo (2015); Johnson and Mont- gomery (1974)). Briefly, a JSS problem involves a finite set of jobs to be processed on a finite set of machines. Each job comprises a set of operations that must be performed on one machine within a certain set of capable machines with various constraints and costs, in a given job-dependent order. Therefore, JSS is essentially a machine scheduling problem where jobs represent a series of activities and machines represent resources and each machine can process at most one job at a time. A typical objective of this process is to minimise the total completion time required for all jobs (i.e., the make span) or the economic cost, although other key performance indicators (KPIs) may individually or jointly be used in the objective function. As a classical research topic in the operation research community, common approaches developed by this commu- nity are linear programming and dynamic programming, whilst general search and optimization techniques developed by the community of artificial intelligence (AI) are utilised or adapted to JSS solutions, such as breadth-first search, Genetic algorithm (GA), Beam search (BS). A job shop scheduling problem is often represented and solved as a con- straint satisfaction problem (usually termed as SAT or CSP), which specifies both the mathematical representation of the problem and the corresponding solving solutions (Ghédira (2013)). A large number of solutions have been re- ported in the literature for CSPs, which revolve around a series of artificial intelligence technological advances that have occurred over the last three decades Pinedo (2015). Most of the JSS problems are NP-hard, with a small number of exceptions, that is the computational requirements for obtaining an optimal solution grow exponentially as the size of the problem increases. JSS solutions, no matter what form proposed by the operation research community or the artificial intelligence community, can be classified into three categories (Çaliş and Bulkan (2015); Chaudhry and Khan (2016); Miguel and Shen (2003); Gomes et al. (2005)): 1) the exact approaches based on hard search, 2) hard search with heuristics, and 3) inexact approaches based on soft search. The exact approaches or hard search algorithms can produce optimal solutions, but they are of exponentially computational time complexity. Typical hard search algorithms include branch and bound, integer linear programming and dynamic programming, amongst others. Various heuristics have been developed to speed up the hard search, with or without sacrificing the optimal solutions. By contrast, the inexact ap- 4 Longzhi Yang et al. proaches or soft search algorithms do not guarantee optimal solutions, but it generates near optimal solutions. In fact, soft search sacrifices the absolute optimal solutions for a reasonable computational effort and thus time span. Soft search approaches are developed mainly based on the advances in soft computing. Common soft search approaches include genetic algorithms, simulated annealing, ant colony optimisation and fuzzy logic, amongst others. These ap- proaches provide a full spectrum of compromise and balance between time and performance, which provides the users a great selection of options for problem solving. Genetic algorithms (GAs) computationally simulate the evolutionary process from nature to explore the optimal solutions from a large searching domain by repeating three basic operations, selection, crossover and mutation. As an adaptive heuristic search algorithm for solving both constrained and unconstrained optimisation problems, GA has been successfully and wildly applied to solve various JSS problems, such as Ak and Koc (2012); Wang et al. (2009). In particular, GA utilises a number of individuals (or organisms) that specially implement biologically inspired com- putational structures, most commonly chromosomes, to compose a population (Dawkins (1982)). Each individual of the population usually contains either the different job’s sequence number or allocated machines’ number for each job, which represents a valid scheduling solution. Based on the current generation of population, the genetic oper- ators, crossover and mutation, are employed to randomly generate the next generation of population, thus to form another valid scheduling solution. These operations are repeated until an optimal scheduling solution, which usually determined by minimising the cost such as the financial cost or the temporal cost of operations, is discovered. 2.2 | Fuzzy Inference and Interpolation Fuzzy sets and systems represent and reason on vague information that arises due to the lack of sharp bound- aries (Zadeh (1965)). The most widely applied fuzzy systems are fuzzy inference systems, such as the Mamdani inference (Mamdani (1977)) and the TSK inference (Takagi and Sugeno (1985)). In particular, these inference sys- tems are able to represent non-linear and high dimensional decision making problems as fuzzy rule bases. Rule bases are either translated from expert knowledge, or extracted from data sets (Mamdani (1977); Tan et al. (2016)). Given an input, a fuzzy inference system produces a system output by referring to those rules in the rule base whose an- tecedents overlap with the given input. However, a fuzzy inference system will fail if the given input is not covered by any rule in the rule base. Fuzzy rule interpolation (FRI) was proposed to address this limitation (Kóczy and Hirota (1993); Huang and Shen (2008); Yang and Shen (2013); Li et al. (2017)). FRI is essentially a fuzzy extension of piece-wise linear or polynomial interpolation or extrapolation (Yang et al. (2017b)). Accordingly, FRI enjoys the advantages of both fuzzy logic in terms of uncertainty management and piece- wise linear or polynomial interpolation or extrapolation by means of knowledge generalisation. Indeed, except being used as a supplementary to conventional fuzzy inference systems to work with sparse rule bases led by limited knowl- edge, FRI can also be used for system simplification for complex fuzzy models by omitting those rules that can be approximated by their neighbours. Given an input which does not overlap with any rule antecedent, the two closest neighbouring rules can be identified based on a given fuzzy distance metric, which is the aggregation (usually weighted average) of the distances between the antecedent items and their counterparts in the observation (Huang and Shen (2008)). Then, an indeterminate rule is generated such that its antecedents are as ’close’ (given a fuzzy distance metric) to the given input as possible. From this, a conclusion is derived by ensuring the shape distinguishability between the conclusion and the consequence of the intermediate rule is equal to that between the antecedents of the interpolated rule and intermediate rule. The technical details of the above approach can be found in Huang and Shen (2008), which are omitted here due to space limitation. FRI approaches have been further developed from different perspectives. For instance, adap- Longzhi Yang et al. 5 tive fuzzy interpolation was proposed to guarantee the interpolated results are consistent throughout the inference processes (Yang and Shen (2009, 2011); Cheng et al. (2016); Yang et al. (2017a)); rough-fuzzy rule interpolation was proposed for both representing the knowledge involving higher order uncertainty and facilitating rule interpolation with such knowledge (Chen et al. (2016)); dynamic version of fuzzy rule interpolation was proposed to add significant rules dynamically (Naik et al. (2017)); and experience-based fuzzy rule interpolation enables rule base involvement and revision whilst performing fuzzy inferences (Li et al. (2016)). The experience-based fuzzy rule interpolation is particularly useful in this work for manual task completion time estimation thanks to its ability in adaptively generating and revising the rule base when only limited training data and/or expert knowledge is available. In particular, this system firstly initialises the rule base with a very limited number of rules representing the limited incomplete knowledge. Then, based on the existing rule’s usage frequency information, historic performance information and distances between observations and existing rules, the two rules with the most important significances are selected for FRI rather than the two closest neighbouring rules as imple- mented in Huang and Shen (2008). Finally, the rule base is adaptively generated and revised, which is guided by the discrepancy between the interpolated result and the actual ground truth. 3 | OVERVIEW OF THE PROPOSED SYSTEM This work takes a typical manual operation centre, i.e., the collate and pack area in the print industry, as the example for theoretical development and experimentation. In particular, a collate and pack area deals with all the collating, packing and other hand finishing processes. In the print industry, this area is usually treated as a single cost centre in manufacturing Management Information System (MIS), although practically multiple production lines (usually named lanes in the industry) are often planned and applied by the area manager manually in the collate and pack area for the performance of multiple jobs in parallel in pursuit of running efficiency; this discrepancy basically disables the MIS. The proposed system herein accurately estimates the manual task completion time and automates the lane planning and scheduling. Note that this work focuses only on the collate and pack operations for simplicity, although the inclusion of tasks on machines is a scaled up version of this proposed system with more variables and constraints. The proposed system in this work is illustrated in Figure 1, which takes a set of manual jobs as inputs. The input jobs are annotated by the sales team which describes the nature of the job. Then the completion times of these jobs on different lane setups are estimated based on the job annotations, which are in turn fed forward to an optimisation algorithm, GA particularly in this work, for job planning and scheduling. After the planing/scheduling is implemented, the actual time used for each manual operation is compared with the estimated one, and the rule base is updated based on this feedback. The proposed system is able to learn whilst it performs and thus it will generally evolve and perform better along time. The two key components of the system, including the fuzzy inference sub-system and the GA subsystem, are detailed in the next two sections. 4 | MANUAL TASK COMPLETION TIME ESTIMATION The completion time of a manual task is estimated by the experience-based fuzzy interpolation system as introduced in Section 2.2, which is demonstrated in Figure 2. The system mainly comprises of three parts: a rule base, an FRI subsystem (particularly a transformation-based approach (T-based) used in this work) and a rule base revision mech- anism. In particular, the rule base is initialised based on limited expert knowledge, or more specifically based on the 6 Longzhi Yang et al. FIGURE 1 The overview of the proposed job planning and scheduling system expertise of the manager of the collate and pack area in this paper. The rule base is then constantly revised by the rule base revision mechanism whilst the system performs inferences by the transformation-based FRI. These three main components are detailed in the following subsections. Rule Base Revision Incoming Job Rule Base Updating information Decision T-Based FRI Job Review Neighbouring rules FeedbackInput FIGURE 2 Experience-based fuzzy interpolation system 4.1 | Rule Base Initialisation The lanes run in the collate and pack area are traditionally planned and scheduled by the area manager using their expertise knowledge. For simplicity and rule of thumb, the manager usually only implements three types of lanes: i) small lanes each implemented by one worker, ii) medium lanes each implemented by 4 workers, and iii) large lanes each implemented by 8 workers. Of course, other types of lanes may also be implemented for exceptional situations in order to achieve a deadline, but these are organised ad hoc without any serious planning and scheduling, and thus not considered in this work. The manager usually pursues both operational and managerial efficiency when setting up lanes. Note that in this work, it is assumed there is one shift per day, each lasting 8 hours. The expert knowledge is used in this work to initialise a fuzzy rule base for the estimation of the completion times of collate and pack tasks as discussed in the next subsection. According to the expert knowledge, the time for completion regarding a given manual task on a particular lane Longzhi Yang et al. 7 setup depends on a number of factors, typically including the object size, the task complexity, the number of parts, the total glue length, the number of applications of glue, the number of staples, and the number of folds; the feature values of each job are always provided by the sales team in the format of annotations. The unskilled nature of the assembly staff role implies a relatively short work life on this particular function. The worker may either move on to more skilled operations, be seasonal and only stay with the company in a matter of months, weeks, or even days, or threshold to an accepted level for their working life. The skill of the worker is therefore not a factor or parameter in the scheduling. This domain knowledge, which may vary from manufacturer to manufacturer, is used in this work to generate an initial rule base. The process of converting expert knowledge, usually expressed as linguistic rules, to fuzzy rules is beyond the scope of this paper, and a classical example of such process can be found in Mamdani (1977). Of course, if there are sufficient data available, data-driven approaches can also be used for rule base generation (Tan et al. (2016)), but this is usually not the case for most of the manufacturers due to the lack of data. As each piece of knowledge may be of different quality, a weight is also assigned to each rule indicating the quality of the rule. The variables used in the rule base and their domains are detailed in Table 1; and each rule is of one of the three following formats: R j 1 : IF x1 is A j 1 and x2 is A j 2 and · · · and x7 is A j 7 , THEN z1 is B j 1 (w j 1 ) R j 4 : IF x1 is A j 1 and x2 is A j 2 and · · · and x7 is A j 7 , THEN z4 is B j 4 (w j 4 ) R j 8 : IF x1 is A j 1 and x2 is A j 2 and · · · and x7 is A j 7 , THEN z8 is B j 8 (w j 8 ), (1) where A and B are fuzzy sets; w represents the weight of the rule expressing the confidence of the corresponding domain knowledge; zk represents the lane with k workers; and R j k represents the j th rule regarding a lane with k workers. TABLE 1 Variables and their domains Variable Meaning Domain x1 Object size {very small, small, medium, large, very large} x2 Task complexity {very easy, easy, medium, complex, very complex} x3 No. of parts A fuzzy integer number in a certin range x4 Total glue length A fuzzy real number in a certain range x5 No. of glue applications A fuzzy integer number in certain range x6 No. of staples A fuzzy integer number in a certain range x7 No. of folds A fuzzy integer number in a certain range z1 Time on a small lane A fuzzy real number in a certain range z4 Time on a medium lane A fuzzy real number in a certain range z8 Time on a large lane A fuzzy real number in a certain range 8 Longzhi Yang et al. Note that the rules above represent the rules of thumb in practice for manual lane planning and job allocations, but they usually do not lead to optimal operational efficiency. In order to enable the application of a JSS system for global optimisation of the manufacturing processes, an accurate estimation of the completion time for ad hoc lanes is of crucial importance. In other words, the times for completion need to be accurately estimated for all the combinations of inputs from the variable input domain with regard to various lanes of different sizes, which can be collectively represented as rules as: R j k : IF x1 is A j 1 and x2 is A j 2 and · · · and x7 is A j 7 THEN zk is B j k (w j k ), (2) where k = {1,2, · · · ,m}, m represents the largest number of workers on a single lane, and m takes 8 in this work. The complete rule base is impossible to be implemented based on the domain knowledge of the collate and pack area manager due to the complexity of the problem. Fortunately, this can be solved by adapting the recently proposed experience-based rule base generation and adaptation approach (Li et al. (2016)) as detailed in the next subsection. 4.2 | Fuzzy Rule Interpolation The initialised rule base is very sparse. Suppose now there is a new collate and pack task which can be described as x1 = A p 1 ,x2 = A p 2 , · · ·x7 = A p 7 as shown in the gray row in Table 2. This new input is not covered by any rule in the rule base. Then the neighbouring rules need to be identified to support the interpolation of completion times for the given new task based on different lane setups. Suppose that the closest neighbouring rules in the sparse rule base regarding the given input are identified and explicitly shown in Table 2, including: Rf1 : IF x1 is A f 1 and x2 is A f 2 and · · · and x7 is A f 7 THEN z1 = Bf1 (w f 1 ), R g 1 : IF x1 is A g 1 and x2 is A g 2 and · · · and x7 is A g 7 THEN z1 = B g 1 (w g 1 ), Rh4 : IF x1 is A h 1 and x2 is A h 2 and · · · and x7 is A h 7 THEN z4 = Bh4 (w h 4 ), R l4 : IF x1 is A l 1 and x2 is A l 2 and · · · and x7 is A l 7 THEN z4 = B l4 (w l 4). (3) TABLE 2 The inference process In order to estimate the time for completion by a small lane with one worker, the two ’closest’ rules Rf 1 and Rg 1 Longzhi Yang et al. 9 in the rule base are identified based on a given fuzzy distance metric. From this, the value of variable z1 for the given task can be estimated by means of fuzzy extrapolation using the approach briefed in Section 2.2. Similarly, the time for completion by a medium lane with four workers can be estimated using neighbouring rules Rh 4 and R l 4 through fuzzy interpolation. From this, the time lengths of completion with 2-worker and 3-worker lanes for the give task can be interpolated from these two extrapolated and interpolated rules by artificially taking the number of workers on the lane as a rule antecedent attribute. The time lengths of completion based on any other lane setups can be either interpolated or extrapolated in the same way. Note that some of the interpolated/extrapolated rules that have been proven accurate which does not present in the existing rule base will be added in the rule base. This means that the rule base will become denser and denser after the deployment of the proposed system to enable the system to adaptively learn from practice. The choosing of the closest neighbouring rules for interpolation/extrapolation can be different with the situation discussed above, and multiple options might be available. Suppose that the part of the rule base in the running example has been updated as illustrated in Table 3 (because interpolated/extrapolated rules Rp 3 and Rp 4 have been added in the rule base during the rule base revision which is discussed later in Section 4.3). Another new task presents which can be described as x1 = A∗1 ' A p 1 ,x2 = A ∗ 2 ' A p 2 , · · ·x7 = A ∗ 7 ' A p 7 as shown in the gray row in Table 3. In this case, it is clear that neighbouring rules Rh 4 and R l 4 can be used for interpolation and denote the interpolated results as Bp 4 ′. In the same time, by artificially taking the number of workers on lanes as an extra rule antecedent, the completion time can also be estimated from rules Rp 3 and Rp 5 and the result is denoted as Bp 4 ′′. In this case, the final result is a weighted aggregation of these two interpolated results in order to generate a global result, that is: B p 4 = λB p 4 ′ + (1−λ)B p 4 ′′ , (4) where λ is problem specific and 0.5 is used as a default value in this work. TABLE 3 The evolved inference process 4.3 | Rule Base Revision The rule base keeps being revised based on its performance whilst it performs fuzzy inferences after the system is deployed. This is mainly achieved using a feedback mechanism. Upon the completion of a collate and pack task, the real completion time is compared with the estimated completion time; and the comparison result is used to determine the quality of the interpolated rule which is expressed as the weight of the interpolated rule. In particular, the weight w of an interpolated/extrapolated rule is defined as: w = e− (t−t ′ )2 a , (5) 10 Longzhi Yang et al. where a(a > 0) is an adjustable parameter indicating the time error tolerance, and t and t′ represent the actual and estimated completion times, respectively. It is clear that the maximum value of weight is 1, which represents the time predicted by the system exactly matches the time spent in the real-world case. In the work, a = 30 is applied, which is provided by the experts. The interpolated rule may be used to replace an existing rule in the rule base, to extend the existing rule base or simply to be ignored. The framework of the rule base updating procedure is illustrated in Figure 3. An interpolated rule that has been proven accurate will be added into the rule base if there is not any similar rules included in the rule base. Suppose that R∗ is an interpolated rule, which is represented as “IF x1 is A∗1 and x2 is A ∗ 2 and · · · and x7 is A∗7 THEN zk = B∗k (w ∗ k )", and that rule R is a rule in the existing rule base which is represented as “IF x1 is A1 and x2 is A2 and · · · and x7 is A7 THEN zk = Bk (w)". The similarity degree of the these two rules S(R∗,R) can then be computed as: S(R∗,R) = 7∑ j=1 S(A∗ j ,Aj )+ S(B ∗,B) 7 + 1 , (6) where S(A∗ j ,Aj ) represents the similarity degree between the j th antecedents of the interpolated rule R∗ and the existing rule R , and S(B∗,B) indicates the degree of similarity between the two consequences. Different approaches have been developed for similarity calculation between fuzzy sets (Chen and Chen (2003)). For instance, if triangular fuzzy sets are employed, each fuzzy set A can be represented as A = {a1,a2,a3}, where (a1,a3) is the support of the fuzzy set and a2 is the core or normal point of the fuzzy set. In this case, the degree of similarity between two triangular fuzzy sets A∗ j and Aj can be calculated as: S(A∗j ,Aj ) = 1− 3∑ k=1 |a∗ jk − ajk | 3 . (7) Add R* Ignore R*Ignore R* Replace Ri by R* No Yes Rule in the rule base Interpolated rule No Yes No Yes Updated Rule Base FIGURE 3 The flowchart of rule base revision Given a similarity threshold δ, if the similarity degrees between the interpolated rule and all existing rules in the Longzhi Yang et al. 11 rule base are less than δ, a potential new rule for the rule base is identified. From this, if the weight of this newly interpolated rulew∗ is greater than a given weight threshold ϕ (i.e., w∗ > ϕ), this newly interpolated rule will be added into rule base; otherwise, the interpolated rule will be ignored due to its poor quality. If the similarity degree between the interpolated rule and one or more existing rules are greater than the given threshold δ, the set of pair-wisely similar rules will be determined. In this case, the system will compare the weights of all these rules, and only keeps the rule with the highest weight value in the identified set. This action ensures that only the most accurate rule within the set is kept in the rule base, such that the rule base is concise and also of a good generalisation ability. By following the rule base revision progress, the number of rules in the rule base and accordingly the system complexity can be controlled by adjusting the similarity threshold δ and weight threshold ϕ. 5 | MANUAL JOB SHOP PLANNING AND SCHEDULING Manufacturing efficiency can be significantly improved by employing the recent advances in artificial intelligence and this is true for the manufacturing of POS and POP which involves complex manual operations. A POS or POP is usually manufactured in multiple stages, such as printing, cutting, collating and packing, but, as stated in the Introduction section, only the manual collate and pack stage is considered in this paper. The objective of job planning and scheduling in this work is to minimize the make span, which not only leads to economic efficiency, but also provides informative information for the sales team such that they can take orders based on available manufacturing capacity. The task of an intelligent JSS system is then to find the best way of collating and packing with the shortest make span ensuring every job is completed within a given deadline, which requires the accurate estimation of completion time for manual tasks in different lane conditions based on the approach proposed in the last section. 5.1 | Problem Representation and Population Initialization Genetic algorithm (GA), as an adaptive heuristic search approach for solving both constrained and unconstrained optimisation problems, has been successfully and wildly utilised to solve JSS problems. In particular, the algorithm firstly initialises the population with random individuals, and then selects a number of individuals for reproduction to produce the next generation of individuals by employing the genetic operators, typically crossover and mutation. The algorithm repeats this process until a satisfactory solution is generated or a pre-defined maximum number of iterations has been reached. Assume that n incoming jobs are on the waiting list for manufacturing. A chromosome or individual, denoted as I , is designed to represent a potential ‘solution’ in this proposed system, as illustrated in Figure 4. An individual is formed by n genes, and each gene represents a job planning allocation. In particular, gene r represents that job Jr is allocated to lane kr which is a lane with from 1 worker to 8 workers. 5.2 | Population Initialisation The initial population Ð = {I1,I2, · · · ,I |Ð|} is formed by randomly assigning a job to an arbitrary valid lane. For a given job associated with a deadline, the required completion times for each of the 8 lanes with 1 to 8 workers can be accurately estimated by employing the proposed completion time estimation system as introduced in Section 4. A job can only be assigned to a valid lane on which the job can be completed within the required deadline. For instance, suppose that a collate and pack task requires to be done by the end of day 2 from the time of scheduling (i.e. two 12 Longzhi Yang et al. . . . 1 st Scheduled Job r th Scheduled Job n th Scheduled Job . . .Gene 1 Gene r Gene n FIGURE 4 Chromosome encoding shifts that is 16 hours), and that the estimated completion times on the 8 lanes from small to big are estimated as z1 = 18.3 hours, z2 = 16.56 hours, z3 = 14.53 hours, z4 = 12.65 hours, z5 = 10.74 hours, z6 = 8.87 hours, z7 = 6.95 hours, and z8 = 5.1 hours. It is clear based on the deadline that valid lanes are those with 3, 4, 5, 6, 7 or 8 workers. Small lanes with 1 or 2 workers are not valid as it will take over two days to complete the job on these lanes. 5.3 | Objective Function An objective function is used in the GA to determine the quality of individuals. The aim of the proposed system is to optimise the lane planning and job scheduling, by minimising the time cost of the collate and pack operations and ensuring all jobs can be done within the specified deadline. Therefore, the objective function in this work is defined as: min t, (t = t1 + · · · + tn),subject to : [tr , r ∈ [1, · · · ,n], tr ≤ dr , (8) where t is the required time to complete all the jobs in the waiting list, tr represents the estimated completion time for job Jr , and dr denotes the given deadline for corresponding job r . The individual with the smallest value of t represents the optimal solution in the population. 5.4 | Selection A number of individuals need to be selected to generate the next generation of population. The progress of the selection in this work is implemented by the fitness proportionate selection method, also known as the roulette wheel selection. Given the current population Ð, if fi is the fitness value of individual Ii in Ð, its probability of being selected to generate the next generation is: p(Ii ) = fi∑|Ð| j=1 fi , (9) where |Ð | is the total number of individuals in the population, or population size. In the proposed system, the fitness value fi of individual Ii is defined by adopting the linear-ranking algorithm ( Baker (1985); Li et al. (2018)), which is defined as: fi = 2−max + 2(max −1))(ri −1) |Ð | , (10) Longzhi Yang et al. 13 where max is a bias or selective pressure towards the fittest individuals in the population, ri is the ranking position of Ii in Ð. 5.5 | Reproduction For a selected number of individuals, two genetic operators, crossover and mutation, are applied to breed the next generation of individuals, as shown in Figure 5. In this work, a two-point crossover operation is adopted, which exchanges the genes between a pair of parent individuates in certain order to generate a pair of offspring (Anand and Panneerselvam (2016)). In particular, the crossover operation firstly randomly selects two crossover points, say p1 and p2 in parents Ii and Ij , and then copies the genes between the two selected crossover points from Ii and Ij to the same positions of offspring I ′ i and I ′ j . After that, the rest fields of I ′ i will be filled by Ij from the left side of the crossover point p2 in a cyclic order. If any job in a gene already exists in I ′i , this gene will be skipped. The second offspring I ′ j is generated by following the exact same operation. This crossover process is demonstrated in Figure 6, which takes 10 genes as an example. Select Genetic Operations Crossover Mutation . . . . . . Select OR fitn e s s Replace Valid ? Valid ? Valid ? Yes Yes Yes Replace Replace No No No Skip this individ ual Go to next iteration Skip this individ ual Go to next iteration FIGURE 5 Procedure of reproduction The second genetic operator is mutation, which is adopted in this work to maintain genetic diversity from one generation to the next. During the mutation operation, a gene from the parent is randomly selected and the planned lane in the gene is randomly mutated to another valid lane of a different size, thus to generate an offspring. The newly bred individuals and some of the best individuals in the current population Ð jointly form the next generation of the population (|Ð |). Note that only one crossover or mutation operation will be allowed to occur per generation and the pre-defined rates are used to control the percentage of operations. In addition, in order to guarantee the generated offspring represent a valid solution, a deadline constraint is always applied to make sure each job can be done within the required deadline. If a produced individual satisfies the applied deadline constraint, which means all scheduled jobs can be finished before the deadline, this individual will be kept for the next generation of population. Of course, if the number of valid individuals in the old generation is less than |Ð| which is usually the case during the first stage of iterations, the produced individual is also kept (despite not being a valid solution, because it increases genetic diversity in the gene pool). 14 Longzhi Yang et al. Parent 1 Parent 2 Off Spring 2 Off Spring 1 Crossover Point 1 Crossover Point 2 Start Filling Point FIGURE 6 The crossover operation 5.6 | Iteration and Termination The reproduction process as discussed above is repeated until a pre-defined maximum number of iterations is reached or the value of objective function of an individual is less than a pre-specified threshold. When the termination condi- tion is reached, the fittest individual in the current population is the optimal solution. Note that the proposed system may fail to generate a viable solution, if too many jobs need to be scheduled which are beyond the maximum manufac- turing capacity. In this case, the GA process will keep trying until the maximum number of iterations has been reached. This may be addressed by applying soft flexible scheduling techniques (Miguel and Shen (2003)), which searches a solution that minimises the cost of constraint violation. This extension is beyond the scope of the current project, which remains as a piece of future work. 6 | EXPERIMENTATION The two main components of the proposed system are validated and evaluated int this section using illustrative ex- amples and a simulated data set. 6.1 | Illustrative Examples on Rule Base Evolvement Three illustrative examples are taken in this section to demonstrate the working procedure of the proposed completion time estimation system. In these examples, the value ofλ in Equation 4 is pre-defined as 0.5, the threshold for similarity degree (δ) is set to 0.8, and the threshold (ϕ) for weight comparison is set to 0.8. For simplicity and to facilitate comprehensibility, only triangular membership functions are used in the example to represent fuzzy sets. 6.1.1 | Model Construction The model takes 7 inputs, as listed in Table 1, which jointly describe a given manual collate and pack task, and it predicts the time of completion for a particular type of lane of certain size. In particular, two inputs, including the object size and the task complexity, take values from fuzzy partitioned variable domain based on the domain knowledge as shown Longzhi Yang et al. 15 in Figure 7; and the other inputs are fuzzy numbers. These fuzzy numbers are usually provided by the in-house job estimating team. For instance, for a given task, the in-house job estimating team may estimate that the number of parts is between 5 and 8, but it is most likely to be 7. Then a triangular fuzzy number (4, 7, 8) will be used as the input for the attribute of the number of parts. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 VS S M L VL (a) Object size 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 VE E M C CV (b) Task complexity FIGURE 7 Fuzzy partition of antecedent attributes 6.1.2 | Scenario 1 Suppose that the initialised rule base only contains 5 fuzzy rules (R1 1 , R2 1 , R3 4 , R4 4 , and R5 8 ), which is extracted from the domain knowledge of the area manager, as shown in Table 4. There is a collate and pack task given as I = (x1 = (0.1,0.2,0.3),x2 = (0.1,0.2,0.3),x3 = (4,5,6),x4 = (2.0,2.2,2.8),x5 = (3,4,5),x6 = (2,4,5),x7 = (2,3,5)), which obviously does not overlap with any rule antecedents in the given rule base. In order to enable global planning and scheduling by an intelligent job shop scheduling system, the completion times of the task on various lanes are estimated first using FRI (and the transformation-based FRI is particularly used in this example). TABLE 4 The initialised rule base No. Antecedents Consequents x1 x2 x3 x4 x5 x6 x7 k=1 k=2 k=3 k=4 k=5 k=6 k=7 k=8 1 (0.05,0.15,0.25) (0.05,0.15,0.25) (1,2,3) (0.8,1.5,1.8) (1,2,3) (1,2,3) (1,2,3) 7.2(1) 2 (0.05,0.15,0.25) (0.05,0.15,0.25) (4,6,7) (3.0,3.5,4.0) (5,7,8) (4,6,8) (3,4,5) 12.0(1) 3 (0.2,0.35,0.5) (0.4,0.6,0.8) (7,9,10) (8.2,10.5,11.2) (6,7,8) (7,9,10) (5,6,7) 4.9(1) 4 (0.4,0.6,0.8) (0.7,1,1) (12,14,15) (18.5,20.8,22.5) (7,9,10) (13,14,16) (7,8,9) 8.9(1) 5 (0.7,1,1) (0.7,1,1) (17,18,20) (28.5,30.5,31.2) (10,11,12) (18,20,21) (9,10,11) 6.0(1) There are two rules available with the consequence being the completion time on a lane with 1 worker, two rules available with consequence being the completion time on a lane with 4 workers, and one rule available with consequence being the completion time on a lane with 8 workers in the initialised rule base. From these rules, the completion time required for the given task on a lane with 1 worker (i.e., z1 = 9.37) can be interpolated from neigh- bouring rules R1 1 and R2 1 ; and that on a lane with 4 workers (i.e., z4 = 2.24) can be extrapolated using neighbouring rules R3 4 and R4 4 . From this, the time of completion on the rest of lanes for the given task can either be interpolated or extrapolated based on the previously generated rules R∗ 1 and R∗ 4 by artificially taking the number of workers on the lane as an input attribute. The generated results are listed in Table 5. After feeding the results into the job shop scheduling system, assume that a globally optimised solution was generated which has planned the task on a lane with one worker, and such a lane was finally scheduled in the collate and pack area for the given task. Upon the completion of the task, the actual time taken for this task is t = 9.85. From this, the weight of the interpolated rule is calculated as w∗ 1 = 0.95 based on Equation 5. Therefore, the newly 16 Longzhi Yang et al. TABLE 5 Times of completion for scenario 1 Antecedents Consequents x1 x2 x3 x4 x5 x6 x7 k = 1 k = 2 k = 3 k = 4 k = 5 k = 6 k = 7 k = 8 (0.1,0.2,0.3) (0.1,0.2,0.3) (4,5,6) (2.0,2.2,2.8) (3,4,5) (2,4,5) (2,3,5) 9.37 4.69 3.12 2.24 1.87 1.56 1.34 1.17 interpolated rule can be represented as: R∗1 : IF x1 is (0.1,0.2,0.3) and x2 is (0.1,0.2,0.3) and x3 is (4,5,6) and x4 is (2.0,2.2,2.8) and x5 is (3,4,5) and x6 is (2,4,5) and x7 is (2,3,5) THEN z∗1 = 9.37 (0.95). (11) According to the rule revision procedure discussed in Section 4.3, once a new rule has been interpolated, the similarity degree between the newly interpolated rule and the existing rules in the rule base are computed to determine the actions for rule base revision. Given the newly interpolated rule as shown above in Equation 11, the similarity degrees between it and all existing rules in the rule base are calculated. Note that there are only two rules R1 1 and R2 1 whose consequences are the time of completion on a lane with one worker. The sub-results and the final results of the calculation based on Equation 6 are summarised in Table 6. As the similarity degrees are less than the given threshold δ = 0.8 and the weight of the rule (w∗ 1 = 0.95) is greater than the predefined threshold ϕ = 0.8, this interpolated rule R∗ 1 is added into rule base. TABLE 6 The details of similarity degree calculation i S(x∗ 1 ,x i 1 ) S(x∗ 2 ,x i 2 ) S(x∗ 3 ,x i 3 ) S(x∗ 4 ,x i 4 ) S(x∗ 5 ,x i 5 ) S(x∗ 6 ,x i 6 ) S(x∗ 7 ,x i 7 ) S(z∗ 1 ,z i 1 ) S(R∗ 4 ,R i 4 ) 1 0.75 0.75 0.4 0.59 0.5 0.54 0.6 0.77 0.61 2 0.75 0.75 0.88 0.65 0.6 0.6 0.82 0.78 0.73 6.1.3 | Scenario 2 The rule base keeps being revised whilst the system performs inferences. Suppose that the rule base has now been updated as shown in Table 7, and a new task, which is similar to the one discussed in Scenario 1, appears. Therefore, the time lengths of completion on different lanes need to be estimated. Of course, the completion time on lanes with 1 worker and 4 workers can be directly acquired from rules R2 1 and R2 4 . The time of completion on a lane with 2 workers can be interpolated either from R1 2 and R3 2 , or R2 1 and R2 4 by artificially taking the number of workers as a rule antecedent. In particular, result z2 2 ′ = 6.64 is interpolated using neighbouring rules R1 2 and R3 2 ; and result z2 2 ′′ = 6.99 is interpolated using neighbouring rules R2 1 and R2 4 by considering the number of workers as an input attribute. In this case, the final result is estimated as the average of the two interpolated results based on Equation 4, which is z2 2 = 6.82. The estimation of the completion time on other types of lanes can also be implemented in one of the ways discussed above, but the calculation details are omitted. Suppose that the lane with 2 workers has been selected by the job scheduling system, and the actual time taken for this manual task is t = 9.6. Based on Equation 5, the weight of the interpolated rule is determined as w∗ 2 = 0.79. Once the newly interpolated rule is constructed, the similarity between this interpolated rule R∗ 2 and the existing Longzhi Yang et al. 17 TABLE 7 The updated rule base No. Antecedents Consequents x1 x2 x3 x4 x5 x6 x7 k = 1 k = 2 k = 3 k = 4 k = 5 k = 6 k = 7 k = 8 1 (0.05,0.15,0.25) (0.05,0.15,0.25) (1,2,3) (0.8,1.5,1.8) (1,2,3) (1,2,3) (1,2,3) 7.2(1) 5.36(0.8) 1.7(0.8) 2 (0.1,0.2,0.3) (0.1,0.2,0.3) (4,5,6) (2.0,2.2,2.8) (3,4,5,) (2,4,5) (2,3,5) 9.37(0.95) 2.24(0.98) 3 (0.05,0.15,0.25) (0.05,0.15,0.25) (4,6,7) (3.0,3.5,4.0) (5,7,8) (4,6,8) (3,4,5) 12.0(1) 8.97(0.8) 2.9(0.8) 4 (0.2,0.35,0.5) (0.4,0.6,0.8) (7,9,10) (8.2,10.5,11.2) (6,7,8) (7,9,10) (5,6,7) 4.9(1) 5 (0.4,0.6,0.8) (0.7,1,1) (12,14,15) (18.5,20.8,22.5) (7,9,10) (13,14,16) (7,8,9) 8.9(1) 6 (0.7,1,1) (0.7,1,1) (17,18,20) (28.5,30.5,31.2) (10,11,12) (18,20,21) (9,10,11) 6.0(1) rules R1 2 and R3 2 are calculated as: S(R∗ 2 ,R1 2 ) = 0.59 and S(R∗ 2 ,R3 2 ) = 0.60, which are both less than the given threshold δ = 0.8. According to the rule base revision procedure as shown in Figure 3, the interpolated rule R∗ 2 therefore is ignored, and the rule base keeps unchanged. 6.1.4 | Scenario 3 Suppose that now another collate and pack task appears as I = (x1 = (0.5,0.7,0.9),x2 = (0.4,0.7,0.8),x3 = (10,11,12), x4 = (19.5,21.2,22.0),x5 = (7,8,9),x6 = (11,12,13),x7 = (7,8,9)). The times of completion on lanes with 1 worker, 2 workers and 4 workers for the given task can be either interpolated or extrapolated from the rule base shown in Table 7. In particular, the completion time on lanes with 1 worker and 2 workers can be extrapolated from neighbouring rules (R2 1 , R3 1 ) and (R1 2 , R3 2 ), respectively; and the completion time on a lane with 4 worker can be interpolated from neighbouring rules R4 4 and R5 4 . The time estimation based on other lane setup can then be interpolated or extrapolated from the previously interpolated and extrapolated rules by artificially taking the number of rules as an input attribute. The estimated results for all lane setups are listed in Table 8. TABLE 8 Times of completion for scenario 3 Antecedents Consequents x1 x2 x3 x4 x5 x6 x7 k = 1 k = 2 k = 3 k = 4 k = 5 k = 6 k = 7 k = 8 (0.5,0.7,0.9) (0.4, 0.7, 0.8) (10, 11, 12) (19.5, 21.2, 22.0) (7,8 ,9) (11,12,13) (7,8,9) 27.20 13.60 9.07 6.7 5.44 4.53 3.89 3.40 The job scheduling system finally selects the lane with 4 workers during the operation optimisation stage. After the job is completed, the weight for this interpolated rule is calculated as w∗ 4 = 0.78. The degrees of similarity be- tween this interpolated rule and the existing rules are computed as: S(R∗ 4 ,R1 4 ) = 0.26, S(R∗ 4 ,R2 4 ) = 0.40, S(R∗ 4 ,R3 4 ) = 0.46, S(R∗ 4 ,R4 4 ) = 0.73, S(R∗ 4 ,R5 4 ) = 0.82. It is clear that a similar rule R5 4 to the newly interpolated rule exists in the rule base as S(R∗ 4 ,R5 4 ) = 0.82. Then, the weight of the interpolated rule R∗ 4 is compared with that of the existing rule R5 4 , and w5 4 = 1 > w∗ 4 = 0.78. Consequently, the interpolated rule is discarded and the rule base remains as it was. 6.2 | Manual Assembly Job Planning and Scheduling Due to the commercial sensitivity of real-world performance data, two experiments based on a simulation data set are reported in this section to demonstrate the working of the proposed job scheduling system and to confirm the power of the proposed system in improving manufacturing efficiency. 6.2.1 | Simulation 1 - System Demonstration The data set contains 10 collating and packing jobs that need to be scheduled, which are with various complexities and different dispatch deadlines. The environment of the collate and pack area in this experiment is the same as the 18 Longzhi Yang et al. environment introduced in Section 6.1, which is able to accommodate 8 different sizes of lanes. There are maximally 13 workers available per day (8 working hours) to operate multiple jobs on multiple lanes at the same time. Consequently, based on these requirements, a large number of lane combinations are available for lane planning and job scheduling. For instance, lanes with 1, 4 and 8 workers are able to be operated at the same time, and the combination of lanes with 1, 3, 4 and 5 workers are also allowed amongst others. The aim of this demonstration is to use the proposed GA-based JSS system to find an optimal solution to schedule the 10 jobs and plan their lanes, thus to minimise the temporal cost of collating and packing and also to provide the preview of the available manufacturing capacity for the marketing and sales team. Also, for operational efficiency and as a common practice, all jobs are better to be finished within one day, such that the work shop can be cleared, cleaned and prepared for the jobs on the following day. This has been used as a hard constraint in the experiments. In this experiment, the parameters of GA are configured as follow: crossover rate = 0.85, mutation rate = 0.05, maximum iteration = 10,000 and the termination condition for an optimal solution is the time cost is not getting any better in 200 iterations. These parameters are empirically determined in this experiment. Assume that the completion times of each job on different lanes have been estimated by the proposed task completion time estimation system, which are listed in columns 2-9 and 12-19 in Table 9, and columns 10 and 20 indicate the dispatch deadline of the corresponding job in days. Based on the given deadlines of each job and the estimated completion time on each lane, the valid lanes to complete the individual jobs can be identified, as shown in Table 10. From this, the initialised population for the GA can be generated. Figure 8 shows two examples of randomly generated individuals. From the individual encoding method, the completion days for finishing all 10 jobs can be easily determined, which are 6 and 5 days, respectively, as also illustrated in the figure. Note that, during the population intialisation stage, the constraint of deadline is not considered for fast starting. Once the population has been initialised, the genetic operators of crossover and mutation, are employed to generate the next generation of population. 001 7 004 8 008 7 009 6 002 2 003 1 005 6 010 8 007 1 006 5 005 8 002 6 007 7 010 4 001 5 004 2 006 5 009 5 008 5 003 7 Individual 1 Individual 2 } Job ID Scheduled Lane Scheduled Sequence 1 st 10 th Day 1 Day 2 Day 3 Day 4 Day 5 Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 FIGURE 8 Two examples of randomly generated individuals for the initialised population In this work, a two points crossover operation is adopted. Suppose that two parent individuals have been selected for crossover operation, as shown in Figure 8, and two crossover points, 3 and 6, have been randomly selected. The process of the crossover operation is illustrated in Figure 9, which reproduced two new scheduling solutions, named as offspring 1 and offspring 2. These two offspring individuals represent two new lane plans and job schedules for all the 10 jobs, which are shown in Table 11. In this situation, it is clear that a better solution has been obtained by offspring 1 after the crossover operation, which reduced the total completion days to 3 days from 6 days, at the Longzhi Yang et al. 19 TABLE 9 Required completion times for the 10 simulated jobs Job Number Lane 1 Lane 2 Lane 3 Lane 4 Lane 5 Lane 6 Lane 7 Lane 8 Deadline (days) 1 12.80 12.43 10.28 8.66 7.05 5.55 4.57 2.99 2 2 8.97 7.11 5.93 5.78 4.14 3.15 1.98 0.99 2 3 7.70 6.72 5.74 4.75 3.70 2.79 1.81 0.82 4 4 9.25 7.98 7.02 5.91 4.63 3.68 2.57 1.45 2 5 16.21 14.75 12.92 10.35 10.47 7.66 6.18 4.07 2 6 14.98 13.48 11.77 10.17 7.98 6.96 5.37 3.76 3 7 7.74 6.76 5.77 4.78 3.54 2.81 1.83 0.84 4 8 19.02 17.36 15.43 13.36 11.23 9.35 7.44 5.38 4 9 16.64 15.77 13.30 10.80 10.77 7.79 6.46 4.35 2 10 13.00 11.65 9.13 7.70 6.94 5.83 4.41 2.96 2 TABLE 10 The smallest valid lane for each job Job 1 Job 2 Job 3 Job 4 Job 5 Job 6 Job 7 Job 8 Job 9 Job 10 Smallest valid lane 4 7 8 7 3 4 8 2 3 5 same time each job can be implemented in the specified deadline. However, offspring 2 leads to late completion for jobs 5 and 7, which consequently make an invalid solution. In this case, only offspring 1 will be used to replace often one of the worst individuals in the current population to produce the next generation of population. As an invalid solution, offspring 2 is ignored. The mutation operation in this work randomly mutates the scheduled lane number of a single job in a selected individual. In this demonstration, taking offspring 1 in Figure 9 as an example, the mutation point 9 has been randomly selected, and the scheduled lane size of job 6 in gene 9 is mutated to 8. As a result, the completion day of the new solution after the mutation operation was increased to 4 days from 3 days. Although it is not an optimal solution compared with current best individual, it is still a valid solution, as all jobs can be completed in time. Therefore, this offspring from mutation operation will still be used to replace one of the individuals usually with the worst fitness in the current population to generate a new generation of population. After a certain number of iterations, the GA is terminated; and the fittest individual in the population, as shown in Figure 10, indicates the optimal job schedule for the 10 current jobs. Of course, when a brand new job appears in the waiting list, the JSS system will be re-employed to update the lane plan and job scheduling. This will guarantee the planned lanes and scheduled jobs are always optimal and also the available capacity can be seen for a given deadline when the sales team commits to taking on a new job. 6.2.2 | Simulation 2 - System Evaluation Currently, many small and medium-sized enterprises (SMEs) and some large manufacturers still manually implement planning and scheduling ‘collation activities’ outside the core MIS functionality, usually utilising tools, such as spread- sheets embedded with crude heuristics. Due to jobs’ complexities, manual scheduling in this situation is often prac- tically implemented as a simple first in and first out queue (FIFO) or an Earliest Deadline First queue (EDF) driven by 20 Longzhi Yang et al. 001 7 004 8 008 7 009 6 002 2 003 1 005 6 010 8 007 1 006 5 005 8 002 6 007 7 010 4 001 5 004 2 006 5 009 5 008 5 003 7 Parent 1 Parent 2 009 6 002 2 003 1 010 4 001 5 004 2 006 5 008 5 005 8 007 7010 4 001 5 004 2 005 6 007 1 006 5 008 7009 6 002 2 003 1Offspring 1 Offspring 2 Crossover Point 1 Crossover Point 2 FIGURE 9 Crossover operation for two selected individuals 009 6 002 2 003 1 010 4 001 5 004 2 005 6 007 1 006 5 008 7 Day 1 Day 2 Day 3 FIGURE 10 The optimal solution dispatching deadlines. Briefly, FIFO is the simplest scheduling algorithm by which the jobs first taken by the manu- facturer are assumed to be manufactured first. EDF is a dynamic scheduling algorithm to place manufacturing jobs in a priority queue based on the deadlines. Whenever a new job occurs, the queue will be updated and the jobs with the closest deadlines will be manufactured first. Consequently, the optimal scheduling solution of all such operations for all the current jobs is hard to achieve, which not only restricts the efficiency of manufacturing process, but also often leads to unnoticed late jobs until the times come close to the dispatch deadlines. In order to evaluate how the proposed scheduling system can improve the efficiency of the POS/POP manufacturing process, a simulated data set of 50 collating and packing jobs associated with deadlines is utilised in this section. The required completion times of each job on each lane can be estimated by the proposed task completion time estimation system, which are shown Table 12. In order to estimate the proposed scheduling system in manufacturing efficiency, the proposed scheduling ap- proach is compared with the ones usually used by the collate and pack manager. Note that the collate and pack jobs are only planned and scheduled by the area manager manually on three types of lanes with 1, 4 or 8 workers; and the jobs are scheduled on these three lanes using two commonly used planning methods which are the first come first served (FIFO) method and the earliest deadline first (EDF) method. Different to these manual approaches, the proposed scheduling system is able to efficiently consider all 8 lanes thanks to the computational and search ability of the algorithms, which minimises the time cost for the scheduling. The scheduling results using the three different methods are listed in Table 13. In addition, for comparison purposes, the proposed JSS system is also applied for job scheduling based only on three types of lanes, small, medium and large, to match the situations of manual panning methods, and the results are also reported in Table 13. Longzhi Yang et al. 21 TABLE 11 Offspring 1 and Offspring 2 after the crossover operation Off Spring 1 Off Spring 2 Lane 1 Lane 2 Lane 3 Lane 4 Lane 5 Lane 6 Lane 7 Lane 8 Lane 1 Lane 2 Lane 3 Lane 4 Lane 5 Lane 6 Lane 7 Lane 8 Day 1 3 2 10 9 4 10 1 Day 2 4 1 5 3 2 9 Day 3 7 6 8 6 8 Day 4 5 Day 5 7 TABLE 12 The estimated time for completion on different sizes of lanes for the manual task data set Arriving Order Lane 1 Lane 2 Lane 3 Lane 4 Lane 5 Lane 6 Lane 7 Lane 8 Dispatch Deadline Arriving Order Lane 1 Lane 2 Lane 3 Lane 4 Lane 5 Lane 6 Lane 7 Lane 8 Dispatch Deadline 1 16.21 14.75 12.92 10.35 10.47 7.66 6.18 4.07 13 26 16.64 15.77 13.30 10.80 10.77 7.79 6.46 4.35 17 2 7.70 6.72 5.74 4.75 3.70 2.79 1.81 0.82 16 27 11.90 11.25 8.87 7.57 6.43 4.94 3.73 2.51 14 3 13.00 11.65 9.13 7.70 6.94 5.83 4.41 2.96 17 28 9.09 9.22 6.62 6.02 4.98 3.42 2.45 1.47 14 4 10.04 9.39 7.23 6.06 5.22 3.43 2.90 1.62 12 29 10.04 9.39 7.23 6.06 5.22 3.43 2.90 1.62 25 5 13.48 13.08 10.82 8.90 7.64 6.78 5.40 3.21 10 30 8.35 7.33 6.28 5.24 4.07 3.16 2.13 1.08 7 6 14.98 13.48 11.77 10.17 7.98 6.96 5.37 3.76 13 31 8.97 7.11 5.93 5.78 4.14 3.15 1.98 0.99 21 7 16.19 14.60 12.77 11.07 9.36 7.65 5.97 4.24 13 32 7.71 6.46 5.40 4.82 3.74 2.79 1.70 0.86 23 8 7.26 6.45 5.37 4.55 3.62 2.79 1.60 0.78 14 33 12.82 11.47 9.98 8.56 7.02 5.72 4.31 2.89 19 9 9.28 7.99 6.84 5.93 4.59 3.69 2.58 1.46 15 34 9.32 8.23 7.08 5.96 4.80 3.72 2.60 1.48 9 10 7.99 6.89 5.45 4.37 3.24 2.12 1.01 0.56 8 35 10.04 9.39 7.23 6.06 5.22 3.43 2.90 1.62 13 11 7.74 6.76 5.77 4.78 3.54 2.81 1.83 0.84 13 36 7.71 6.46 5.40 4.82 3.74 2.79 1.70 0.86 14 12 7.26 6.45 5.37 4.55 3.62 2.79 1.60 0.78 21 37 8.35 7.33 6.28 5.24 4.07 3.16 2.13 1.08 17 13 7.55 6.59 5.61 4.64 3.57 2.70 1.73 0.76 7 38 11.91 10.63 9.22 7.88 6.42 5.20 3.87 2.52 11 14 9.04 8.74 6.56 5.89 4.78 3.36 2.28 1.21 24 39 7.55 6.59 5.61 4.64 3.48 2.70 1.73 0.76 20 15 9.25 7.98 7.02 5.91 4.63 3.68 2.57 1.45 19 40 11.76 10.50 9.10 7.78 6.44 5.12 3.80 2.46 9 16 10.04 9.39 7.23 6.06 5.22 3.43 2.90 1.62 15 41 11.17 9.88 8.12 7.29 5.90 4.40 3.50 2.25 17 17 7.26 6.45 5.37 4.55 3.62 2.79 1.60 0.78 15 42 15.89 14.53 12.54 10.22 10.43 7.35 6.12 3.94 14 18 9.04 8.74 6.56 5.89 4.78 3.36 2.28 1.21 13 43 10.24 9.66 7.40 6.33 5.70 3.94 3.01 1.67 20 19 8.97 7.11 5.93 5.78 4.14 3.15 1.98 0.99 15 44 13.49 13.56 11.36 8.93 7.91 6.82 5.55 3.28 18 20 12.69 12.35 10.22 8.20 6.80 5.30 4.56 2.96 25 45 13.48 12.08 10.82 7.90 6.64 4.78 2.40 1.81 22 21 9.09 7.22 6.62 6.02 4.98 3.42 2.45 1.47 16 46 17.11 15.46 13.54 11.76 9.93 8.18 6.42 4.62 22 22 12.14 10.85 9.42 8.06 6.67 5.33 3.98 2.61 16 47 13.48 13.08 10.82 8.90 7.64 6.78 5.40 3.21 14 23 7.71 6.46 5.40 4.82 3.74 2.79 1.70 0.86 12 48 8.34 7.31 6.26 5.23 3.98 3.15 2.12 1.08 18 24 16.64 15.77 13.30 10.80 10.77 7.79 6.46 4.35 14 49 12.80 12.43 10.28 8.66 7.05 5.55 4.57 2.99 11 25 15.86 14.30 12.50 10.83 9.10 7.47 5.81 4.11 17 50 19.02 17.36 15.43 13.36 11.23 9.35 7.44 5.38 8 Table 13 clearly shows that all the 50 jobs can be efficiently scheduled within 14 days without any jobs delay or extra working hours by applying the proposed scheduling system, compared with FIFO method and EDF method that both require 17 days for completion. In particular, the FIFO method leads to multiple job delays, which can be very costly due to the breach of contracts or last minute premium shipping. Although the job schedule made by the EDF method does not cause any job delay, some of jobs require extra working hours to be finished on time, such as Jobs 50, 6 and 47, which consequently will increase the economic cost of the manufacturer. Interestingly, all 50 jobs can also be scheduled for completion using the proposed JSS system following the required dispatch deadlines without causing any extra hours, if only three packaging lanes are considered by the proposed system. However, in this case, the jobs progressing days will be increased to 20 days if this method is employed, which costs the business an additional 6 days. 22 Longzhi Yang et al. TABLE 13 The job scheduling made by different methods ("′" indicates the miss of the dispatch deadline, and "∗" represents the requirement of extra working hours) FIFO EDF The proposed System The proposed System with three fixed lanes Lane 1 Lane 4 Lane 8 Lane 1 Lane 4 Lane 8 Lane 1 Lane 2 Lane 3 Lane 4 Lane 5 Lane 6 Lane 7 Lane 8 Lane 1 Lane 4 Lane 8 Day 1 2 3 1 13 10 30 13 35 24 23 40 50 Day 2 5 4 6 50∗ 34 40 10 18 40 5 10 18 5 Day 3 8 9 7 49∗ 38 5 23 4 38 49 13 30 49 Day 4 10 11 12 23 4 1 34 30 1 4 38 Day 5 13 14 15 11 6∗ 7 11 6 7 11 34 1 Day 6 17 18 16 8 18 35 36 28 27 47 8 35 6 Day 7 19 21 20 28∗ 27 24 8 19 48 50 36 27 7 Day 8 23 22 24 36 47∗ 42 15 22 42 28 24 Day 9 25 27 26 17 9 16 2 17 21 46 17 48 42 Day 10 30′ 28 29 2 19 21 29 41 26 2 16 26 Day 11 32 31 33 22∗ 3 25 39 16 14 20 21 22 Day 12 36 34′ 35 37∗ 41 26 12 9 43 33 39 3 25 Day 13 39 37 38′ 48∗ 15 44 31 45 25 37 10 Day 14 40′ 41 42 39 43 33 32 37 3 44 9 13 Day 15 45 43 44 12 31 45 12 41 44 Day 16 47 48 46 32 14 46 32 15 33 Day 17 49′ 50′ 29 20 43 46 Day 18 45 31 Day 19 14 20 Day 20 29 7 | CONCLUSIONS This paper presented a manual job shopping planning and scheduling system using a GA algorithm and a dynamic fuzzy model. In particular, the fuzzy model, implemented by adapting the recently proposed experience-based fuzzy rule interpolation approach, estimates the time of completion for manual collate, assembly and pack tasks on different lanes with a variety of product configurations and sizes. This provides the necessary prerequisite in implementing the GA-based intelligent job shop scheduling systems which not only plans the lane arrangement but also schedules jobs based on the planned lanes. The system was validated and evaluated by illustrative examples and simulated experiments. The experimental results demonstrated the power of the proposed system in improving manufacturing productivity and efficiency. This is of high pragmatic and financial interest to the business as lower value work could be outsourced if it is found to cause delays to higher priority work. Despite of the success, the work can be further improved in different directions. Firstly, the proposed system was developed particularly for POS and POP manufacturers in this paper, but the underpinning artificial intelligence approaches are applicable to any industry with manual tasks involved in the manufacturing processes. Secondly, the proposed system only considers a simple situation that every worker takes one job per day. This constraint can be released to further improve the manufacturing efficiency; the implementation of this requires further investigation. Thirdly, it is worthwhile to investigate how the constraints can be relaxed when too many jobs beyond the maximum manufacturing capacity need to be scheduled. Longzhi Yang et al. 23 conflict of interest Conflicts of interest: none references Ak, B. and Koc, E. (2012) A guide for genetic algorithm based on parallel machine scheduling and flexible job-shop scheduling. Procedia - Social and Behavioral Sciences, 62, 817 – 823. World Conference on Business, Economics and Management (BEM-2012), May 4–6 2012, Antalya, Turkey. Anand, E. and Panneerselvam, R. (2016) A study of crossover operators for genetic algorithm and proposal of a new crossover operator to solve open shop scheduling problem. American Journal of Industrial and Business Management, 6, 774. Baker, J. E. (1985) Adaptive selection methods for genetic algorithms. In Proceedings of the 1st International Conference on Genetic Algorithms, 101–111. Hillsdale, NJ, USA: L. Erlbaum Associates Inc. Çaliş, B. and Bulkan, S. (2015) A research survey: review of ai solution strategies of job shop scheduling problem. Journal of Intelligent Manufacturing, 26, 961–973. Chaudhry, I. A. and Khan, A. A. (2016) A research survey: review of flexible job shop scheduling techniques. International Transactions in Operational Research, 23, 551–591. Chen, C., Parthaláin, N. M., Li, Y., Price, C., Quek, C. and Shen, Q. (2016) Rough-fuzzy rule interpolation. Information Sciences, 351, 1 – 17. Chen, S.-J. and Chen, S.-M. (2003) Fuzzy risk analysis based on similarity measures of generalized fuzzy numbers. IEEE Transactions on Fuzzy Systems, 11, 45–56. Chen, T., Shang, C., Su, P. and Shen, Q. (2018) Induction of accurate and interpretable fuzzy rules from preliminary crisp representation. Knowledge-Based Systems, 146, 152 – 166. URL: http://www.sciencedirect.com/science/article/pii/ S0950705118300546. Cheng, S.-H., Chen, S.-M. and Chen, C.-L. (2016) Adaptive fuzzy interpolation based on ranking values of polygonal fuzzy sets and similarity measures between polygonal fuzzy sets. Information Sciences, 342, 176 – 190. Dawkins, R. (1982) The Extended Phenotype. Oxford Oxford University Press. Ghédira, K. (2013) Constraint satisfaction problems: csp formalisms and techniques. John Wiley & Sons. Gomes, M., Barbosa-Póvoa, A. and Novais, A. (2005) Optimal scheduling for flexible job shop operation. International Journal of Production Research, 43, 2323–2353. Huang, Z. and Shen, Q. (2008) Fuzzy interpolation and extrapolation: A practical approach. Fuzzy Systems, IEEE Transactions on, 16, 13–28. Johnson, L. A. and Montgomery, D. C. (1974) Operations research in production planning, scheduling, and inventory control, vol. 6. Wiley New York. Kóczy, L. and Hirota, K. (1993) Approximate reasoning by linear rule interpolation and general approximation. International Journal of Approximate Reasoning, 9, 197–225. Li, J., Qu, Y., Shum, H. P. H. and Yang, L. (2017) TSK Inference with Sparse Rule Bases, 107–123. Cham: Springer International Publishing. Li, J., Shum, H. P. H., Fu, X., Sexton, G. and Yang, L. (2016) Experience-based rule base generation and adaptation for fuzzy interpolation. In 2016 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), 102–109. 24 Longzhi Yang et al. Li, J., Yang, L., Qu, Y. and Sexton, G. (2018) An extended takagi–sugeno–kang inference system (tsk+) with fuzzy interpolation and its rule base generation. Soft Computing, 22, 3155–3170. Mamdani, E. H. (1977) Application of fuzzy logic to approximate reasoning using linguistic synthesis. Computers, IEEE Trans- actions on, C-26, 1182–1191. Miguel, I. and Shen, Q. (2003) Fuzzy rrdfcsp and planning. Artificial Intelligence, 148, 11 – 52. Naik, N., Diao, R. and Shen, Q. (2017) Dynamic fuzzy rule interpolation and its application to intrusion detection. IEEE Trans- actions on Fuzzy Systems, PP, 1–1. Pinedo, M. (2015) Scheduling. Springer. Takagi, T. and Sugeno, M. (1985) Fuzzy identification of systems and its applications to modeling and control. IEEE Transactions on Systems, Man, and Cybernetics, SMC-15, 116–132. Tan, Y., Li, J., Wonders, M., Chao, F., Shum, H. P. H. and Yang, L. (2016) Towards sparse rule base generation for fuzzy rule interpolation. In 2016 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), 110–117. Wang, Y. M., Yin, H. L. and Wang, J. (2009) Genetic algorithm with new encoding scheme for job shop scheduling. The International Journal of Advanced Manufacturing Technology, 44, 977–984. Yang, L., Chao, F. and Shen, Q. (2017a) Generalized adaptive fuzzy rule interpolation. IEEE Transactions on Fuzzy Systems, 25, 839–853. Yang, L. and Shen, Q. (2009) Towards adaptive interpolative reasoning. In 2009 IEEE International Conference on Fuzzy Systems, 542–549. Yang, L. and Shen, Q. (2011) Adaptive fuzzy interpolation. Fuzzy Systems, IEEE Transactions on, 19, 1107–1126. Yang, L. and Shen, Q. (2013) Closed form fuzzy interpolation. Fuzzy Sets and Systems, 225, 1 – 22. Theme: Fuzzy Systems. Yang, L., Zuo, Z., Chao, F. and Qu, Y. (2017b) Fuzzy interpolation systems and applications. In Fuzzy Control Systems. InTech. Zadeh, L. (1965) Fuzzy sets. Information and Control, 8, 338 – 353. 
work_gda2m7oyejgx3jjbieakuk7mru	----	Creating the aspired intelligent assessment systems for teaching materials Expert Systems with Applications 38 (2011) 12168–12179 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Creating the aspired intelligent assessment systems for teaching materials Cheng-Hsiung Chen a, Gwo-Hshiung Tzeng b,c,⇑ a Department of Information Management, Kainan University, No. 1, Kainan Road, Shinshing Tsuen, Luchu Shiang, Taoyuan 338, Taiwan b Distinguished Chair Professor, Kainan University, No. 1, Kainan Road, Shinshing Tsuen, Luchu Shiang, Taoyuan 338, Taiwan c Institute of Management of Technology, National Chiao-Tung University, 1001, Ta-Hsueh Road, Hsin-Chu 300, Taiwan a r t i c l e i n f o a b s t r a c t Keywords: Intelligent assessment system Teaching materials MCDM (Multiple Criteria Decision Making) DEMATEL ANP (Analysis Network Process) VIKOR Core competence (CC) 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.03.050 ⇑ Corresponding author at: Distinguished Chair Pro 1, Kainan Road, Shinshing Tsuen, Luchu Shiang, Taoy 937 541 184. E-mail addresses: hsiung.chen@msa.hinet.net (C. edu.tw, ghtzeng@mail.knu.edu.tw, arthur@mail.knu.e The core objective of the Nine-Year Integrated Curriculum for primary schools is to ‘‘enable students to demonstrate their talents instead of just scoring high on exams’’ around the world. The determining fac- tor in education reform is to declare the competence indicators of necessary educational behavior in pri- mary and junior high school. In the reform process, for all domains, the enriched rate of competence indicators for educational materials and methods is very meaningful. Because educational materials and methods in different domains have their own style, we should evaluate these teaching materials sep- arately. Thus, in this research we propose a novel MCDM (Multiple Criteria Decision Making) framework for evaluating, comparing, and improving the effectiveness of competence indicators in the various pub- lications for teaching materials in primary school based on different viewpoints. The ANP (Analysis Net- work Process) weights are based on the DEMATEL technique with the MCDM method for resolving the problems of dependence and feedback among criteria. Then, a VIKOR technique with ANP weights is proposed for addressing and reducing the performance gaps for each criterion, thus hopefully improving, re-configuring and selecting the aspired Intelligent Assessment Systems (IAS) for teaching materials. An empirical study of Mandarin Chinese based on this system design of three publishers is illustrated to verify the effectiveness of the proposed method. The results can improve the efficiency and quality of the authored Mandarin Chinese teaching materials and may extend to other Learning Areas. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Every country has made the cultivation of human talent a priority in the 21st century. As other advanced countries propose education reform in different forms throughout the world, Taiwan also recognizes education as the bedrock of national development, implementing various education reforms such as pre-school edu- cation reform, grade 1–9 curriculum reform, the restructuring of secondary education, the enhancement of higher education, and lifelong learning projects (MOE, 2008). Therefore, the purpose of this research is to propose a novel technique and evaluation method that can improve, re-configure and select the aspired Intel- ligent Assessment Systems (IAS) for teaching materials to promote education levels. Since 1990, the four British educational reform contexts have been: (a) the right to national centralization reform; (b) the priority of compulsory education; (c) vocational education mainstreaming; and (d) the ability to pursue educational evaluation (DfE, 1991). ll rights reserved. fessor, Kainan University, No. uan 338, Taiwan. Tel.: +886 -H. Chen), ghtzeng@cc.nctu. du.tw (G.-H. Tzeng). The report of the Mayer Committee proposed to advise the Austra- lian Education Council (AEC) and the Ministers for Vocational Educa- tion, Employment and Training (MOVEET) on employment-related key competencies for post-compulsory education and training (Mayer, 1992). The Education Commission of Hong Kong proposed the following: ‘‘Learning is the key to one’s future, and education is the gateway to our society’s tomorrow’’. Education enables indi- viduals to develop their potential, construct knowledge and enhance the quality of their personal lives (EC, 2000). In general, no matter what style of education reform is used, the key competencies are the major concern for national education re- form at the beginning of the 21st century. The core objective of Nine-Year Integrated Curriculum for pri- mary schools in Taiwan is to ‘‘enable students to demonstrate their talents instead of just scoring high on exams’’. The determining factor in education reform is to declare the competence indicators for necessary educational behavior in primary school (MOE, 2002). In the reform process, for all domains, the enriched rate of compe- tence indicators for educational materials and methods is very meaningful. Because educational materials and methods in differ- ent domains have their own style, we should evaluate these teach- ing materials separately. Thus, in this research we propose a novel MCDM (Multiple Criteria Decision Making) framework based on the DEMATEL http://dx.doi.org/10.1016/j.eswa.2011.03.050 mailto:hsiung.chen@msa.hinet.net mailto:ghtzeng@cc.nctu.edu.tw mailto:ghtzeng@cc.nctu.edu.tw mailto:ghtzeng@mail.knu.edu.tw mailto:arthur@mail.knu.edu.tw http://dx.doi.org/10.1016/j.eswa.2011.03.050 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 12169 (Decision Making Trial and Evaluation Laboratory) technique for evaluating, comparing, and improving the effectiveness of compe- tence indicators in the various publications for teaching materials in primary school based on different viewpoints. The ANP (Analysis Network Process) weights are based on the DEMATEL technique with the novel MCDM method for resolving the problems of depen- dence and feedback among criteria. Then, a VIKOR technique with ANP weights is proposed for addressing and reducing the perfor- mance gaps for each criterion, thus hopefully improving, re-config- uring and selecting the aspired Intelligent Assessment Systems for teaching materials. An empirical study of Mandarin Chinese teach- ing materials in Grade 1 of primary school based on this system de- sign of three publishers is illustrated to verify the effectiveness of the proposed method. The results can improve the efficiency and quality of the authored Mandarin Chinese teaching materials and may extend to other Learning Areas. The remainder of this paper is organized as follows. In Section 2, the aspired intelligent assessment systems for teaching materials with MCDM are introduced. In Section 3, a novel MCDM method based on the DEMATEL technique is proposed. In Section 4, an empirical study for the aspired IAS for Mandarin Chinese teaching materials is presented to show the process that our proposed method entails, and discussions are conducted. Finally, in Section 5, concluding remarks are presented. 2. Intelligent assessment systems for teaching materials with MCDM method In recent decades, competence-based education has become the mega trend impacting the education reform strategies of the majority number of governments throughout the world. In the fol- lowing section, the literatures related to core competence (CC), the intertwined effects of an assessment system for teaching materials, will be reviewed as a foundation for the development of the theo- retical framework of this paper. We also give an example of Taiwan to explain the basic concepts of Educational Reform for teaching materials. 2.1. Educational reform of Taiwan (MOE, 2002) In keeping with the progress of the 21st century and the global trends of educational reform, Taiwan must engage in educational reform in order to foster national competitiveness and boost the overall quality of our citizen’s lives. The Ministry of Education (MOE) of Taiwan, therefore, has initi- ated curricular and instructional reforms in primary and junior high school education. These reforms have been based on the Table 1 Numbers for competence indicators for the mandarin Chinese curriculum for ea Criterion of CC Stag D1: Physical, mental, and spiritual mold C1: Self-understanding and exploration of potential 24 C2: Appreciation, representation, and creativity 20 C3: Career planning and lifelong learning 9 D2: Interpersonal and social relations C4: Expression, communication, and sharing 8 C5: Respect, care and teamwork 9 C6: Cultural learning and international understanding 6 C7: Planning, organizing and putting plans into practice 7 D3: The use of Life Science and Technology C8: Utilization of technology and information 4 D4: Logical thinking and reasoning C9: Active exploration and study 8 C10: Independent thinking and problem-solving 9 Sum 104 Sources: (MOE, 2002) Action Plan for Educational Reform approved by the Executive Yuan of Taiwan. Because the curriculum is not only the core of schooling but also the foundation on which teachers plan learning activities, the MOE places top priority on the development and implementation of the Grade 1–9 Curriculum. Curricular reforms are necessary and will be timely for the following reasons: (a) meeting national development needs; and (b) meeting public expectations. 2.2. Assessment system for mandarin chinese teaching materials Mandarin Chinese is one of the curricula included in the Lan- guage Arts. It is divided into 3 stages: Grades 1–3, Grades 4–6, and Grades 7–9. 2.2.1. Mandarin Chinese’s curriculum goal Based on the Curriculum Goals and CCs of the Grade 1–9 Curric- ulum, a more detailed description is given in sub Sections 4.1.2 and 4.1.3, with the Mandarin Chinese Curriculum Goal defined as fol- lows: (1) to utilize language to inspire individual potential and to cultivate learning; (2) to cultivate interest in writing and enhance ability to appreciate literature; (3) to equip with the self-learning ability for language learning and lay the foundation for lifelong learning; (4) to utilize language and words for expressions of emotion, experience-sharing and communications; (5) to utilize language expression to adapt to one’s environment and to demon- strate appropriate behavior; (6) to understand and recognize the Chinese, Taiwanese and foreign cultures and rituals through lan- guage learning; (7) to utilize the power of language to develop and implement plans effectively; (8) to combine the information of lan- guage and technology to enhance learning and to expand fields of study; (9) to cultivate an interest in language exploration and a pro- active attitude towards language-learning; (10) to utilize language for independent thinking and problem-solving. 2.2.2. Competence indicators of Mandarin Chinese’s curriculum of grade 1–9 curriculum Based on the requirements for children’s intellectual develop- ment, the numbers for the competence indicators of the Mandarin Chinese Curriculum for each stage are defined (see Table 1). In general, this includes listening, speaking, reading and writing of languages, and developing basic communication competencies. From the academic point of view, six terms were developed as fol- lows: (1) phonetic symbol applications; (2) listening skills; (3) the ability to speak; (4) literacy and writing ability; (5) reading skills; and (6) writing skills. ch stage. e1 Stage2 Stage3 Sum 19 18 61 17 17 54 13 10 32 9 12 29 9 13 31 8 6 20 7 6 20 11 8 23 7 9 24 8 7 24 108 106 318 Fig. 1. An analytical framework for aspired assessment system for teaching materials. s2 2 3 1 3 s1 s4 s3 s2 2 3 1 3 s1 s4 s3 Fig. 2. An example of the directed graph. 12170 C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 3. A novel MCDM method based on the DEMATEL technique with ANP The structure of the MCDM problem will be derived using the DEMATEL method. The priorities of every determinant are based on the structure derived by using the ANP. The VIKOR technique will be leveraged to calculate the compromise ranking of the alter- natives. Finally, the assessment system for Mandarin Chinese teaching materials will be derived. In summary, this evaluation framework consists of five main phases: (1) establishing determi- nants using the questionnaire survey method; (2) building the structure of network relation map (NRM) for the determinants by using DEMATEL; (3) calculating the priorities of every determinant using the ANP based on the structure of NRM derived by using DEMATEL in (2); (4) ranking the priorities of the assessment sys- tem for teaching materials using the VIKOR technique; and finally (5) establishing the assessment system for teaching materials and achieving the aspired levels (see Fig. 1). 3.1. DEMATEL Technique with ANP The DEMATEL technique was developed by the Battelle Geneva Institute: (1) to analyze complex ‘world problems’ dealing mainly with interactive man-model techniques; and (2) to evaluate qual- itative and factor-linked aspects of societal problems (Gabus & Fontela, 1972). The applicability of the method is widespread, with applications ranging from industrial planning and decision-making to urban planning and design, regional environmental assessment, the analysis of world problems, and so forth. It has also been suc- cessfully applied in many situations and fields, such as those of marketing strategies, control systems, safety problems, developing the competencies of global managers, group decision-making and so on (Chen, Lien, & Tzeng, 2010; Chiu, Chen, Tzeng, & Shyu, 2006; Huang, Shyu, & Tzeng, 2007; Lee, Tzeng, Hsu, & Huang, 2009; Li & Tzeng, 2009a, 2009b; Lin & Tzeng, 2009; Lin & Wu, 2008; Liou, Tzeng, & Chang, 2007; Ou Yang, Leu, & Tzeng, 2009; Ou Yang, Shieh, Leu, & Tzeng, 2008; Tzeng, Chen, Yu, & Shih, 2010; Wu & Lee, 2007). Furthermore, a hybrid model combining the two methods has been widely used in various fields – for example, e-learning evaluation (Tzeng, Chiang, & Li, 2007), teach- ing materials assessment (Chen and Tzeng, 2009), airline safety measurement (Liou et al., 2007), and innovation policy portfolios for Taiwan’s SIP Mall (Huang et al., 2007). Therefore, in this paper we use DEMATEL not only to detect complex relationships and build a NRM for the criteria but also to obtain the influence levels of each element over others; we then adopt these influence level values as the basis of the normalization supermatrix for determin- ing ANP weights to obtain the information about relative impor- tance. To apply the DEMATEL method smoothly, the authors refined the definitions based on the above authors and produced the essential definitions indicated below. The DEMATEL method is based upon graph theory, enabling us to plan and solve problems visually, so that we may divide multiple criteria into a relationship of cause and effect to better understand causal relationships. Direc- ted graphs (also called digraphs) are more useful than directionless graphs because digraphs will demonstrate the directed relation- ships of sub-systems. A digraph typically represents a communica- tion network or a domination relationship between individuals. Suppose a system contains a set of elements, S = {s1, s2, . . . , sn}, and particular pair-wise relationships are determined for modeling with respect to a mathematical relationship, MR. Next, portray the relationship MR as a direct-relation matrix that is indexed equally in both dimensions by elements from the set S. Then, extract the case for which the number 0 appears in the cell (i, j) if the entry is a positive integral that has the following meaning: the ordered pair (si, sj) is in the relationship MR; and has the kind of relation- ship regarding that element such that si causes element sj. The di- graph portrays a contextual relationship between the elements of the system, in which a numeral represents the strength of influ- ence (Fig. 2). The number between factors is the influence or influ- enced degree. For example, an arrow from s1 to s2 represents the fact that s1 influences s2 and that its influenced degree is two. The DEMATEL method can convert the relationship between the causes and effects of criteria into an intelligible structural model of the system (Chiu et al., 2006).. Definition 1. The pair-wise comparison scale may be designated as five levels, where the scores 0, 1, 2, . . . , 4 represent the range from ‘no influence’ to ‘very high influence’. Definition 2. The initial direct relation/influence matrix A is an n � n matrix obtained by pair-wise comparisons in terms of influ- ences and directions of influence between the determinants, in which aij is denoted as the degree to which the ith determinant affects the jth determinant. A ¼ a11 a12 � � � a1n a21 a22 � � � a2n .. . .. . . . . .. . an1 an2 � � � ann 2 66664 3 77775 Definition 3. The normalized direct relation/influence matrix N can be obtained through Eqs. (1) and (2), in which all principal diagonal elements are equal to zero. N ¼ zA ð1Þ Goal Criteria Subcriteria Control Criteria Goal Criteria Subcriteria Control Criteria C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 12171 where z ¼ 1 max max 16i6n Xn j¼1 aij; max 16j6n Xn i¼1 aij !, : ð2Þ In this case, N is called the normalized matrix. Because limq!1N q ¼ ½0�n�n A possible different network under each subcriterion of the control hierarchyA possible different network under each subcriterionof the control hierarchy Source: (Saaty, 1996) Fig. 3. The control hierarchy. Definition 4. Then, the total relationship matrix T can be obtained using Eq. (3), where I stands for the identity matrix. T ¼ N þ N 2 þ�� �þ N q ¼ NðI � NÞ�1 ð3Þ ½Explain� T ¼ N þ N 2 þ�� �þ N q ¼ NðI þ N þ N2 þ���þ N q�1ÞðI � NÞðI � NÞ�1 ¼ NðI � N qÞðI � NÞ�1 ¼ NðI � NÞ�1; when limq!1N q ¼ ½0�n�n where q ? 1 and T is a total influence-related matrix; N is a direct influence matrix and N = [xij]n�n;limq?1(N 2 + � � � + Nq) stands for a indirect influence matrix, 0 6 Pn j¼1xij < 1 and 0 6 Pn i¼1 xij < 1, and only one Pn j¼1 xij or Pn i¼1xij is equal to 1 for "i, j, but not all. So, limq?1 N q = [0]n�n. The (i, j) element tij of matrix T denotes the di- rect and indirect influences of factor i on factor j. Definition 5. The row and column sums are separately denoted as r and c within the total-relation matrix T through Eqs. (4) and (5). T ¼ ½tij� i; j 2f1; 2; . . . ; ng ð4Þ r ¼ ½ri�n�1 ¼ Xn j¼1 tij " # n�1 and c ¼ ½cj�n�1 ¼ Xn i¼1 tij " #0 1�n ð5Þ where the vectors r and c denote the sums of the rows and columns, respectively. Definition 6. Suppose that ri denotes the row sum of the ith row of matrix T. Then, ri is the sum of the influences of factor i on the other factors, both directly and indirectly. Suppose that cj denotes the column sum of the jth column of matrix T. Then, cj is the sum of the influences that factor j is receiving from the other factors. Fur- thermore, when i = j (i.e., the sum of the row sum and the column sum (ri + ci) represents the index indicating the strength of the influence, both dispatching and receiving), (ri � ci) is the degree of the central role that factor i plays in the problem. If (ri � ci) is positive, then factor i primarily is influencing the strength of the other factors; and if (ri � ci) is negative, then factor i primarily is receiving influence from other factors (Huang et al., 2007; Liou et al., 2007; Tamura, Nagata, & Akazawa, 2002). 3.2. The ANP method The ANP method, a multi-criteria theory of measurement devel- oped by Saaty (1996), provides a general framework for dealing with decisions without making assumptions about the indepen- dence of higher-level elements from lower-level elements and about the independence of the elements within a level as in a hier- archy. Compared with traditional MCDM methods (Saaty, 2005) – e.g. AHP (Analytic Hierarchy Process), TOPSIS (Technique for Order Preference by Similarity to an Ideal Solution), ELECTRE (ELimina- tion Et Choix Traduisant la REalité), etc. (Lee et al., 2007) – which usually assume independence between criteria, ANP, a new theory that extends AHP to deal with dependence in feedback and utilizes the supermatrix approach (Saaty, 2003), is a more reasonable tool for dealing with complex MCDM problems in the real world. In this section, the concepts of the ANP are summarized based on Saaty’s earlier works (Saaty, 1996, 1999, 2005). The ANP is a coupling of two parts. The first consists of a control hierarchy or network of criteria and subcriteria that control the interactions. The second is a network of influences among the ele- ments and clusters. The network varies from criterion to criterion, and a different supermatrix of limiting influence is computed for each control criterion. Finally, each of these supermatrices is weighted based on the priority of its control criterion, and the re- sults are synthesized through addition for all the control criteria (Saaty, 2004). A control hierarchy is a hierarchy of criteria and sub- criteria for which priorities are derived in the usual way with re- spect to the goal of the system being considered. The criteria are used to compare the components of a system, and the subcriteria are used to compare the elements. The criteria with respect to which influence is presented in individual super- matrices are called control criteria. Because all such influences ob- tained from the limits of the several supermatrices will be combined to obtain a measure of the priority of overall influence, the control criteria should be grouped in a structure to be used to derive priorities for them. These priorities will be used to weight the corresponding individual supermatrix limits and sum up. The analysis of priorities in a system can be thought of in terms of a control hierarchy with dependence among its bottom-level alter- natives arranged as a network as shown in Fig. 3. Dependence can occur within the components and between them. A control hierarchy at the top may be replaced by a control net- work with dependence among its components, which are collec- tions of elements whose functions derive from the synergy of their interaction and that hence have a higher-order function not found in any single element. The criteria in the control hierarchy that are used for comparing the components are usually the major parent criteria whose subcriteria used to compare the elements need to be more general than those of the elements because of the greater complexity of the components. A network connects the components of a decision system. According to size, there will be a system made up of subsystems, with each subsystem made up of components and each component made up of elements. The elements in each component interact or have an influence on some or all of the elements of another com- ponent with respect to a property governing the interactions of the entire system, such as energy, capital, or political influence. Fig. 4 demonstrates a typical network. Those components that no Source Component Source Component Source Component (Feedback loop) Source Component (Feedback loop) Intermediate Component (Transient State) Intermediate Component (Transient State) Sink Component (Absorbing State) Sink Component (Absorbing State) Intermediate Component (Recurrent State) Intermediate Component (Recurrent State) Outerdependence Innerdependence loop C1 C3 C4 C5 C2 Source Component Source Component Source Component (Feedback loop) Source Component (Feedback loop) Intermediate Component (Transient State) Intermediate Component (Transient State) Sink Component (Absorbing State) Sink Component (Absorbing State) Intermediate Component (Recurrent State) Intermediate Component (Recurrent State) Source Component Source Component Source Component (Feedback loop) Source Component (Feedback loop) Intermediate Component (Transient State) Intermediate Component (Transient State) Sink Component (Absorbing State) Sink Component (Absorbing State) Intermediate Component (Recurrent State) Intermediate Component (Recurrent State) Outerdependence Innerdependence loop C1 C3 C4 C5 C2 Source: (Satty, 1996) Fig. 4. Connections in a network. 12172 C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 arrow enters are known as source components, such as C1 and C2. Those from which no arrow leaves are known as sink components, such as C5. Those components which arrows both enter and exit leave are known as transient components such as C3 and C4. In addition, C3 and C4 form a cycle of two components because they feed back and forth into each other. C2 and C4 have loops that con- nect them to themselves and are inner-dependent. All other con- nections represent dependence between components, which are thus known to be outer-dependent. A component (dimension) of a decision network that was de- rived by the DEMATEL method in sub Section 3.1 will be denoted by Ch, h = 1, . . . , m; assume that it has nh elements (determinants), which we denote by eh1; eh2; . . . ; ehnh . The influence of a given set of elements (determinants) in a component (dimension) on any element in the decision system is represented by a ratio scale pri- ority vector derived from paired comparisons of the comparative importance of one criterion and another criterion with respect to the interests or preferences of the decision-makers. This relative importance value can be determined using a scale of 1–9 to repre- sent equal importance to extreme importance (Saaty, 1996). The influence of elements (determinants) in the network on other ele- ments (determinants) in that network can be represented using the following supermatrix: A typical entry Wij in the supermatrix is called a block of the supermatrix in the following form, where each column of Wij is a principal eigenvector of the influence of the elements (determi- nants) of the ith component of the network on an element (determi- nant) in the jth component. Some of its entries may be zero corresponding to those elements (determinants) that have no influence. W ij ¼ wi1 j1 wi1 j2 � � � wi1 jnj wi2 j2 wi2 j2 � � � wi2 jnj .. . .. . . . . .. . wini j1 wini j2 � � � wini jnj 2 666664 3 777775 After forming the supermatrix, the weighted supermatrix is derived by transforming all columns sum to unity exactly. This step is very much similar to the concept of the Markov chain in terms of ensur- ing that the sum of these probabilities of all states equals 1. Next, the weighted supermatrix is raised to limiting powers, such as limh?1W h to get the global priority vector or called weights (Huang, Tzeng, & Ong, 2005). 3.3. VIKOR The compromise ranking method (VIKOR) was proposed by Opricovic (Opricovic, 1998) as one applicable technique to imple- ment within MCDM. We assume that the alternatives are denoted as A1, A2, . . . , Ak, . . . , Am. The rating (performance score) of the jth criterion is denoted by fkj for alternative Ak, wj is the weight of the jth criterion, expressing the relative importance of the criteria, where j = 1, 2, . . . , n, and n is the number of criteria. The VIKOR method began with the following form of theLp � metric: Lpk ¼ Xn j¼1 wj f � j � fkj ��� ���� �h . f �j � f �j ��� ���� � #p( )1=p where 1 6 p 61, k = 1, 2, . . . , m; weight wj is derived using the ANP according to the NRM based on the DEMATEL method. The VIKOR method also uses Lp¼1k (as Sk) and L p¼1 k (as Qk) to formulate the rank- ing measure (Opricovic, 1998; Opricovic & Tzeng, 2002, 2004, 2007; Tzeng, Teng, Chen, & Opricovic, 2002, 2005). Sk ¼ Lp¼1k ¼ Pn j¼1 wj f �j � fkj ��� ���� �= f �j � f �j ��� ���� �h i Q k ¼ L p¼1 k ¼ maxj wj f � j � fkj ��� ���� �= f �j � f �j ��� ���� �jj ¼ 1; 2; . . . ; nn o The compromise solution minkL p k will be chosen because its value is closest to the ideal/aspired level. In addition, when pis small, group utility is emphasized (such as p = 1), and as pincreases to p = 1, the individual maximal regrets/gaps receive more importance, as shown by Yu (Freimer & Yu, 1976; Yu, 1973). Therefore, miniSi emphasizes the maximum group utility, whereas miniQi empha- sizes selecting the minimum of the maximum individual regrets. Based on the above concepts, the compromise ranking algorithm VI- KOR has the following steps. Step 1. Normalize the original rating matrix. In this step, we deter- mine the best f �j and the worst f � j values of all criterion functions, j = 1, 2, . . . , n. If we assume that the jth function represents a benefit, f �j ¼ maxifkj (or setting an aspired level) and f �j ¼ minifkj (or setting a tolerable level). Alter- natively, if we assume that the jth function represents a cost, f �j ¼ minifkj (or setting an aspired level) and f �j ¼ maxi fkj (or setting a tolerable level). Moreover, an ori- ginal rating matrix is transformed into a normalized weight-rating matrix with the following formula: rkj ¼ f �j � fkj ��� ���� �= f �j � f �j ��� ���� � Step 2. Compute the values Sk, Qk, when k = 1, 2, . . . , m, using the rela- tions (Ou Yang et al., 2008, Ou Yang, Leu, & Tzeng, 2009) C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 12173 Sk ¼ Xn j¼1 wjrkj and Q k ¼ maxjfrkjj ¼ 1; 2; . . . ; ng where Sk and Qk show the mean of group utility and maximal regret, respectively. Using the traditional VIKOR method, Qk is represented as maxj{wjrkjjj = 1, 2, . . . , n}, which implies that group utility is more important than maximal regret. Because Qk is only a part of Sk, Sk is unquestionably more than Qk. Therefore, Sk is emphasized more than Qk using the traditional VIKOR method. However, maximal re- gret is also very important in practice and is usually taken into ac- count to improve it. Therefore, to balance Sk and Qk, Qk = maxj{rkjjj = 1, 2, . . . , n} is used instead of the traditional VIKOR Qk. Step 3. Compute the index values Rk, i = 1, 2, . . . , m,using the relation Rk ¼ vðSk � S �Þ=ðS� � S�Þþð1 � vÞðQ k � Q �Þ=ðQ� � Q�Þ where S� ¼ minkSk; S � ¼ maxkSk; Q � ¼ mink Q k; Q � ¼ maxkQ k; (here, we can also set the best value to 0 and the worst value to 1) and 0 6 v 6 1, where v is introduced as a weight for the strategy of maximum group utility, whereas 1 � v is the weight of the individ- ual regret. Step 4. Rank the alternatives, sorting by the value of Sk, Qk, and Rk, for k = 1, 2, . . . , m, in decreasing order. Propose as a compro- mise the alternative (A(1)), which is ranked first by the measure min{Rkjk = 1, 2, . . . , m} if the following two condi- tions are satisfied: C1. Acceptable advantage: R(A(2)) � R(A(1)) P 1/(m � 1), where A(2) is the alternative with the second position on the ranking list by R; m is the number of alternatives. C2. Acceptable stability in decision making: Alternative A(1) must also be the best ranked by Sk or/and Qk, k = 1, 2, . . . , m. A set of compromise solutions is proposed if one of the condi- tions is not satisfied. The set of compromise solutions consists of the following: (1) alternatives A(1) and A(2) if only condition C2 is not satisfied; (2) alternatives A(1), A(2), . . . , A(M) if condition C1 is not satisfied. A(M) is determined by the relation R(A(M)) � R(A(1)) < 1/(m � 1) for maximum M (the positions of these alternatives are close). The compromise-ranking method (VIKOR method) determines the compromise solution; the obtained compromise solution is acceptable to the decision-makers because it provides the maxi- mum group utility of the majority (represented by minS) and a minimum of individual maximal regret for the opponent (repre- sented by minQ). The model uses the DEMATEL and ANP proce- dures in sub Sections 3.1 and 3.2 to obtain the weights of the criteria with dependence and feedback; it uses the VIKOR method to obtain the compromise solution. 4. An empirical study of the aspired assessment systems for mandarin chinese teaching materials In this section, an example being modified from a real case will be presented to demonstrate the effectiveness of the proposed novel MCDM framework using the DEMATEL technique. One empirical study example will be based on an example modified from three leading publishers. In this case study, Mandarin Chinese teaching materials (six books) for Grade 1 in primary school in Taiwan are selected; these materials were edited by those publishers. 4.1. Background description In the twenty-first century, major changes have taken place around the world in social, political, economic and cultural arenas. These changes are not only global but also national. In a drastic change, the most countries have become aware of the importance of education and culture. Education reform in these countries has been carried out to stimulate personal potential, to overtake the fine culture, and to promote social progress. After six decades of postwar development, Taiwan has transi- tioned from a traditional agricultural society into a modern indus- trial society. Political, economic, and cultural environments are faced with modernization, industrialization and the technological influences of structural adjustment and reconstruction. The impact of education reform is one of the most far-reaching and extensive implications of a major reform. It affects national self-positioning, slushing the social consciousness, establishing a new culture and developing national competitiveness for the new century. The effectiveness of economic development and of the develop- ment of democracy in Taiwan have aroused praise from more developed countries. It is a fact that the main contributing factor is the spread of education and enhancement of people’s quality of life. Taiwan is becoming an educational community. However, in the development of education over the years, many problems have emerged, and delays in solving the problems have made them even more complicated. In light of this, the Council on Education Reform was established in September, 1994. It was responsible for educational reform and the educational development of research and consideration, the Commission ‘‘Education overall reform Consultation Report’’ was presented in December of 1996. Taiwan should actively engage in reform. Issues that need to be addressed include the following: establishing a life-long learning society; enhancing equal opportunities in education; guiding examination biased towards intellectual culture; improving course materials and assessments; improve the multi-teacher education system; and improving the efficiency of the use of educational re- sources. The influence of education reform on the social and per- sonal is significant. We must ponder social change movements and carefully consider how the value of benchmarks of socio-cul- tural development is determined, when they occur in the context of education reform, rather than acting rashly. The Council has developed a comprehensive education reform proposal. It is divided into five dimensions as follows: (1) educa- tion deregulation, (2) good educational care for every student, (3) smoother avenues to higher education, (4) improved education quality, and (5) developing a life-long learning society. Speaking of the need to provide proper educational care for every student, it is clear that schools have not given adequate attention to students, especially disadvantaged students. This has been mainly due to the rigidity of primary school and junior high school education (compulsory education) in Taiwan, with a uni- form system and curriculum, coupled with a lack of long-term investment in resources, the unusual effect of teaching, and a sin- gle mode for entrance writing examinations at senior high schools and universities. Early on in the education of those disadvantaged students, they were not able to develop a good foundation for learning, and they were then exposed to large class size and a lack of timely and adequate care in general, so that their performance was different from that of the other students and were relatively vulnerable at their schools. 12174 C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 The Council proposed the following specific recommendations: (1) to reform the curriculum and instruction; (2) to reduce school size and class size; (3) to open autonomous schools; (4) to stimu- late schools to use the inner strength of their own power; (5) to as- sist each student with basic competence; (6) to establish a remedial teaching system; (7) to strengthen career counseling and provide multiple approaches; (8) to renew a system of student behavioral counseling; (9) to weaken the physical and psychologi- cal barriers to education; (10) to attach importance to aboriginal education; (11) to implement gender equality in education; and (12) to protect the quality of early childhood education. In the meantime, the Education Reform Committee, Taiwan, MOE proposed the curricular reforms for the Grades 1–9 (compul- sory education) Curricula. The current Curriculum Frameworks for Primary Schools and Junior High Schools were revised and promulgated in 1993 and 1994, respectively. Although the current Curriculum Frameworks have been gradually and properly implemented, the MOE believes that innovative thinking and practice in education are the prereq- uisites for success in the new century. Therefore, the MOE has launched plans for another curricular reform to build up consensus and integrate efforts at education reform to create a new and bet- ter environment for school education. The development of a new curriculum was divided into three stages. The duration and major tasks for each stage are shown in detail as follows: (1) Stage 1 (from April 1997 to September 1998): Establishing the Special Panel on the Development of Primary and Junior High Schools’ Cur- ricula; (2) Stage 2 (from October 1998 to November 1999): Estab- lishing the Panel on Researching and Formulating the Guidelines of Each Learning Area in the Grade 1–9 Curriculum, which covered the primary school and junior high school levels of education; and (3) Stage 3 (from December 1999 to August 2002): Establish- ing the Review Committee on the Revision and Formulation of Ele- mentary and Junior High School Curricula. 4.1.1. Core rationale The aim of education is to foster sound minds and character in students. Students should be taught democratic values, the Rule of Law, and humanitarian ideals; they should develop strong and healthy physiques, learn how to think for themselves and be crea- tive. Every government hopes that the school system will produce outstanding citizens with a sense of patriotism and the ability to adopt a global perspective. In essence, education is a learning pro- cess to help students explore their potential and develop their capacity to adapt and make the necessary efforts to improve their living environment. Given that, the following five basic aspects are emphasized and included in Grade 1–9 Curricula designed for the new century: developing humanitarian attitudes, enhancing the ability to integrate, cultivating democratic literacy, fostering both indigenous awareness and a global perspective, and building up the capacity for lifelong learning. The core components of each as- pect are as follows: (A) humanitarian attitude; (B) integration abil- ity; (C) democratic literacy; (D) native awareness and a global perspective; and (E) capacity for lifelong learning. 4.1.2. Curriculum goals The curricula for primary and junior high schools will adopt the following principles: (A) to involve all aspects of daily life that are related to students’ mental and physical development; (B) to encourage the development of individuality and the exploration of one’s potential; (C) to foster democratic literacy and respect for different cultures; and (D) to develop scientific understanding and competencies to meet the needs of modern life. The aim of national education is to teach students to obtain ba- sic knowledge and to develop the capacity for lifelong learning, thus cultivating able citizens who are mentally and physically healthy, vigorous and optimistic, gregarious and helpful to the community, intellectually curious and reflective, tolerant, and cre- ative, with a positive attitude and a global perspective. Schools will achieve such ideals through the promotion of educational learning activities that emphasize humanity, practicality, individuality, comprehensiveness, and modernity. Such activities include inter- actions between oneself and others, individuals and the commu- nity, and humans and nature. Regarding this aspect of national education, we must guide our students to achieve the following curriculum goals: (A) to enhance their self-understanding and ex- plore their individual potential; (B) to develop creativity and the ability to appreciate beauty and present their own talents; (C) to promote abilities related to career planning and lifelong learning; (D) to cultivate knowledge and skills related to expression, com- munication, and sharing; (E) to learn to respect others, care for the community, and facilitate teamwork; (F) to promote further cultural learning and international understanding; (G) to strength- en knowledge and skills related to planning, organizing, and imple- mentation; (H) to acquire the ability to utilize technology and information; (I) to encourage an attitude of active learning and studying; and (J) to develop abilities related to independent think- ing and problem-solving. 4.1.3. Core competencies To achieve the aforementioned goals, the curricular design of primary and junior high school education shall focus on the needs and experiences of students and aim to develop the core compe- tencies that a modern citizen should possess. Such CCs (core com- petencies) refer to the key competencies, which defined by the Mayer Committee are: (A) collect, analyze and organize informa- tion; (B) communicate ideas and information; (C) plan and orga- nize activities; (D) co-operate with others and help sustain the group’s ability to work; (E) use mathematical concepts and tech- nologies; (F) solve problems; (G) use technology (Mayer, 1992). Thus, the CCs for the Grade 1–9 curriculum reform in Taiwan may be categorized as follows (MOE, 2002): (A) self-understanding and exploration of potential; (B) appreciation, representation, and creativity; (C) career planning and lifelong learning; (D) expres- sion, communication, and sharing; (E) respect, care and teamwork; (F) cultural learning and international understanding; (G) plan- ning, organizing and putting plans into practice; (H) utilization of technology and information; (I) active exploration and study; and (J) independent thinking and problem-solving. With reference to curricular principles (see Section 4.1.2), the CCs may be divided into 4 dimensions: (1) the physical, mental, and spiritual mold (A-C); (2) interpersonal and social relations (D-G); (3) the use of Life Sciences and Technology (H); and (4) log- ical thinking and reasoning (I-J). 4.1.4. Learning areas To foster the CCs in citizens, the curricula for primary and junior high school education should emphasize three dimensions: indi- vidual development, community and culture, and natural environ- ment. Thus, the Grade 1–9 Curriculum encompasses seven major learning areas: (A) Language Arts, including Mandarin and English; (B) Health and Physical Education; (C) Social Studies, including his- tory and culture, geography, social institutions, morals and norms, etc., and the incorporation of the aforementioned subjects into one’s daily life; (D) Arts and Humanities, including music instruc- tion and instruction in the visual and performing arts; (E) Science and Technology, including learning about substances and energy, nature, the environment, ecological conservation, and information technology; (F) Mathematics, including acquiring the basic con- cepts of figures, shapes, and quantity; the ability to calculate and organize; and the ability to apply such knowledge and skills to dai- ly life; and (G) Integrative Activities, referring to activities that may Table 2 Rating the CCs relationships/influences for grades 1–3 mandarin curricula. D1 D2 D3 D4 C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 D1 C1 0.000 3.500 3.375 3.625 3.250 2.250 2.750 2.375 2.875 3.000 C2 3.250 0.000 2.875 3.000 2.625 2.875 2.750 2.250 2.500 2.375 C3 3.750 2.375 0.000 2.875 2.625 2.500 3.000 2.750 3.000 3.125 D2 C4 3.750 3.250 3.000 0.000 3.250 3.125 3.250 2.750 3.125 3.250 C5 2.375 3.000 2.125 3.500 0.000 3.500 3.375 2.375 2.750 3.000 C6 2.625 3.000 2.500 3.250 3.125 0.000 2.875 2.375 2.625 2.625 C7 2.625 2.625 3.375 3.375 3.375 2.875 0.000 2.750 3.000 3.125 D3 C8 2.375 2.500 2.875 3.125 2.375 2.500 3.000 0.000 3.250 3.250 D4 C9 3.000 3.000 3.375 3.250 3.000 2.875 3.500 3.125 0.000 3.750 C10 3.375 3.375 3.125 3.000 2.875 2.625 3.250 2.875 3.875 0.000 Remark: Surveyed based on the opinions of teacher(s) who are editing Mandarin Chinese teaching materials or have been teaching Mandarin Chinese teaching in primary school, Taiwan. Core Competencies (CCs) D1: Physical, mental, and spiritual mold D2: Interpersonal and social relations D3: The use of Life Science and Technology D4: Logical thinking and reasoning C1: Self-understanding and exploration of potential; C2: Appreciation, representation, and creativity; C3: Career planning and lifelong learning; C4: Expression, communication, and sharing; C5: Respect, care and teamwork; C6: Cultural learning and international understanding; C7: Planning, organizing and putting plans into practice; C8: Utilization of technology and information; C9: Active exploration and study; C10: Independent thinking and problem-solving. C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 12175 guide learners to practice, experience, and reflect upon the learning process as well as to testify to and apply what they have learned in real situations. Based on the analytical frame for expanding them, CCs were first selected as determinants. Then, the structure of the assessment system for the Mandarin Chinese teaching materials definition problem was established using DEMATEL. After that, the weight of each determinant for the decision structure would be decided by using the ANP. The determinants are CCs. The crite- ria were confirmed as Competence Indicators and as determinants for editing Mandarin Chinese teaching material. Meanwhile, the relationships between the determinants and the ANP derivations of the weights of each determinant would also be derived for the case study. The relationships between the determinants for the assessment system for Mandarin Chinese teaching materials were surveyed based on the opinions of teacher(s) who are editing Mandarin Chi- nese teaching materials or have been teaching Mandarin Chinese in primary school, Taiwan. The teachers are familiar with the assessment of Mandarin Chinese teaching materials. With the understanding of the determinants of the assessment system for Mandarin Chinese teaching materials, appropriate assessment strategies will also be proposed to shorten the gap be- tween the level of the selected current teaching material and the Table 3 Result of the DEMATEL analysis of CC Relationships/Influences for Grades 1–3 Mandarin C D1 D2 C1 C2 C3 C4 C5 D1 C1 1.1989 1.2839 1.2798 1.3786 1.2 C2 1.1962 1.0744 1.1645 1.2523 1.1 C3 1.2752 1.2146 1.1384 1.3175 1.2 D2 C4 1.3819 1.3446 1.3369 1.3402 1.3 C5 1.2329 1.2298 1.2038 1.3322 1.1 C6 1.1988 1.1890 1.1737 1.2813 1.1 C7 1.2867 1.2638 1.2847 1.3773 1.2 D3 C8 1.2067 1.1886 1.1998 1.2927 1.1 D4 C9 1.3678 1.3436 1.3544 1.4480 1.3 C10 1.3587 1.3351 1.3284 1.4204 1.3 Remark: Core Competencies (CCs) D1: Physical, mental, and spiritual mold D2: Interpe thinking and reasoning C1: Self-understanding and exploration of potential; C2: Appreci Respect, care and teamwork; C6: Cultural learning and international understanding; C7: P information; C9: Active exploration and study; C10: Independent thinking and problem- aspired levels of the determinants. Meanwhile, the proposed assessment system for Mandarin Chinese teaching materials will assist the publishers in surpassing the publisher leaders. Detailed procedures and results are illustrated below. 4.2. Decision problem network relation map structuring by DEMATEL The inter-relationships between the ten determinants will be deduced using the DEMATEL method introduced in sub Section 3.1. First, the direct relation/influence matrix A is introduced (see Ta- ble 2). After that, the direct relation/influence matrix A is normal- ized based on Eq. (1). Then, the total relationship matrix is deduced based on Eq. (3) (see Table 3). Finally, the strength of the influence for each determinant is deduced based on Eq. (5) (see Table 4) (see Fig. 6). The sequence for the strength of the influence of the determi- nants is as follows: (1) the use of Life Science and Technology (D3), which includes the utilization of technology and information (C8); (2) logical thinking and reasoning (D4), which is sequent as (a) inde- pendent thinking and problem-solving (C10) and (b) active explora- tion and study (C9); (3) physical, mental, and spiritual mold (D1), which is sequent as (a) career planning and lifelong learning (C3). (b) self-understanding and exploration of potential (C1) and (c) appreciation, representation, and creativity (C2); (4) interpersonal urricula. D3 D4 C6 C7 C8 C9 C10 721 1.1868 1.3055 1.1287 1.2787 1.3022 539 1.1083 1.2002 1.0342 1.1654 1.1802 173 1.1574 1.2736 1.1063 1.2450 1.2678 398 1.2754 1.3896 1.1999 1.3542 1.3787 317 1.1851 1.2818 1.0928 1.2346 1.2607 885 1.0379 1.2253 1.0562 1.1897 1.2083 810 1.2097 1.2242 1.1445 1.2880 1.3116 809 1.1305 1.2441 0.9939 1.2232 1.2419 390 1.2745 1.4038 1.2172 1.2641 1.4001 162 1.2490 1.3767 1.1927 1.3628 1.2651 rsonal and social relations D3: The use of Life Science and Technology D4: Logical ation, representation, and creativity; C3: Career planning and lifelong learning; C5: lanning, organizing and putting plans into practice; C8: Utilization of technology and solving. Table 4 ri + ci and ri � ci of the DEMATEL analysis of CC Relationships/Influences for Grades 1–3 Mandarin Curricula. CC ri + ci ri � ci D1: Physical, mental, and spiritual mold C1: Self-understanding and exploration of potential 25.3189 (5) �0.0886 (5) C2: Appreciation, representation, and creativity 23.9970 (8) �0.9377 (10) C3: Career planning and lifelong learning 24.6777 (6) �0.2515 (8) D2: Interpersonal and social relations C4: Expression, communication, and sharing 26.7817 (1) �0.0992 (6) C5: Respect, care and teamwork 24.6058 (7) �0.2351 (7) C6: Cultural learning and international understanding 23.5633 (9) �0.0659 (4) C7: Planning, organizing and putting plans into practice 25.5963 (4) �0.2534 (9) D3: The use of Life Science and Technology C8: Utilization of technology and information 23.0685 (10) 0.7357 (2) D4: Logical thinking and reasoning C9: Active exploration and study 26.0180 (3) 0.8069 (1) C10: Independent thinking and problem-solving 26.0219 (2) 0.3886 (3) Fig. 5. The analytic network based on the casual diagram of total relationships. 12176 C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 and social relations (D2), which are sequent as (a) respect, caring and teamwork (C5), (b) cultural learning and international understand- ing (C6), (c) expression, communication, and sharing (C4) and (d) planning, organizing and putting plans into practice (C7). 4.3. Calculating the weights of determinants by ANP Setting an appropriate assessment system as the goal, pairwise comparisons of the determinants were executed based on the total relationship matrix as deduced by DEMATEL. The inter-relation- ships between the goal and the dimensions of determinants of an assessment system are illustrated in Fig. 5 where the direction of the arrows indicates the direction of the influences. Finally, the to- tal relationship matrix serves as inputs for the ANP. By implement- ing the ANP, the limit super matrix W is calculated. Weights corresponding to each determinant (Table 5) are derived accord- ingly that will be used for the calculation of weighted averages and VIKOR scores. 4.4. Compromise ranking by VIKOR The VIKOR technique being introduced for compromise ranking was applied after the determinants’ weight calculations by ANP in sub Section 4.3. Meanwhile, the weighted averages for the Mandarin Chinese teaching materials (six books) for Grade 1 in primary school, which were edited by three leading publishers, are also calculated for comparison. In general, the calculation results (Table 6–8) demonstrate that both global and local reached the same conclusions: publisher A � publisher C � publisher B. 4.5. Discussions and implications Authoring teaching material is not an easy task. Meanwhile, there are no straightforward answers to the question of how teach- ing material should be authored to meet particular criteria. Not only criteria but also determinant, how teaching material should be authoring by considering factors being related to the authored curriculum. Also, from the perspective of MCDM, very little re- search has addressed the assessment system for teaching material, not to mention Mandarin Chinese teaching materials. In this research, a novel MCDM framework combining the DEMATEL tech- nique, ANP and VIKOR was proposed to address the above- mentioned problems; using this technique yielded satisfactory results. The novel MCDM model consisting of DEMATEL, ANP and VI- KOR was formed as follows: (1) overcome the issue of defining the assessment system for teaching material; (2) use innovative traditional MCDM approaches to resolve the problem of defining the assessment system for teaching material; (3) clarify the vague correlations between the determinants of teaching material; and (4) create the aspired priorities for the authored teaching material. For the assessment system for teaching materials, CCs were se- lected as determinants of the assessment, and categorized into 4 groups as the criteria of the assessment. These groups were as follows; D1: Physical, mental, and spiritual mold C1: Self-understanding and exploration of potential; C2: Appreciation, representation, and creativity; C3: Career planning and lifelong learning; D2: Interpersonal and social relations C4: Expression, communication, and sharing; C5: Respect, care and teamwork; C6: Cultural learning and international understanding; C7: Planning, organizing and putting plans into practice; D3: The use of Life Science and Technology C8: Utilization of technology and information; D4: Logical thinking and reasoning C9: Active exploration and study; C10: In- dependent thinking and problem-solving. Through application studies, we found the novel MCDM model to be applicable. DEMATEL establishes a reasonable assessment structure for dealing with criteria influence. Basically, the influence relationships of the results (see Fig. 6) were quite reasonable: Table 5 Weights of the determinants being derived by ANP. CC Local weights Globe weights (ANP) D1: Physical, mental, and spiritual mold 0.3014 C1: Self-understanding and exploration of potential 0.3377 0.1018 C2: Appreciation, representation, and creativity 0.3312 0.0998 C3: Career planning and lifelong learning 0.3312 0.0998 D2: Interpersonal and social relations 0.4054 C4: Expression, communication, and sharing 0.2656 0.1077 C5: Respect, care and teamwork 0.2455 0.0995 C6: Cultural learning and international understanding 0.2335 0.0947 C7: Planning, organizing and putting plans into practice 0.2554 0.1035 D3: The use of Life Science and Technology 0.0895 C8: Utilization of technology and information 1.0000 0.0895 D4: Logical thinking and reasoning 0.2037 C9: Active exploration and study 0.4959 0.1010 C10: Independent thinking and problem-solving 0.5041 0.1027 Table 6 Satisfaction of Grade 1 Curriculum. CC A Publisher B Publisher C Publisher 11 12 11 12 11 12 D1: Physical, mental, and spiritual mold C1: Self-understanding and exploration of potential 6.0000 5.7500 5.5000 5.2500 6.2500 6.2500 C2: Appreciation, representation, and creativity 6.5000 6.7500 5.5000 7.0000 5.7500 7.2500 C3: Career planning and lifelong learning 5.7500 4.7500 4.2500 4.5000 6.2500 4.7500 D2: Interpersonal and social relations C4: Expression, communication, and sharing 6.7500 7.5000 6.0000 6.5000 6.2500 7.2500 C5: Respect, care and teamwork 6.7500 6.2500 6.2500 6.5000 6.5000 6.5000 C6: Cultural learning and international understanding 5.2500 4.0000 4.2500 4.2500 4.0000 4.0000 C7: Planning, organizing and putting plans into practice 5.0000 5.0000 4.5000 4.7500 4.5000 4.2500 D3: The use of Life Science and Technology C8: Utilization of technology and information 3.7500 3.7500 4.0000 4.7500 4.5000 4.5000 D4: Logical thinking and reasoning C9: Active exploration and study 5.0000 5.2500 4.2500 4.2500 4.7500 4.5000 C10: Independent thinking and problem-solving 5.2500 4.7500 4.7500 4.5000 5.0000 6.0000 Remark: 11: the 1st Semester of Grade 1; 12: the 2nd Semester of Grade 1. Table 7 Aspired level for Grade 1 Curriculum. Criterion of CC Local weights Global weights (ANP) A Publisher B Publisher C Publisher 11 12 11 12 11 12 D1: Physical, mental, and spiritual mold 0.3014 Si 0.3917 0.4250 0.4914 0.4419 0.3916 0.3916 Qi 0.4250 0.5250 0.5750 0.5500 0.4250 0.5250 C1: Self-understanding and exploration of potential 0.3377 0.1018 0.4000 0.4250 0.4500 0.4750 0.3750 0.3750 C2: Appreciation, representation, and creativity 0.3312 0.0998 0.3500 0.3250 0.4500 0.3000 0.4250 0.2750 C3: Career planning and lifelong learning 0.3312 0.0998 0.4250 0.5250 0.5750 0.5500 0.3750 0.5250 D2: Interpersonal and social relations 0.4054 Si 0.4047 0.4263 0.4730 0.4472 0.4661 0.4459 Qi 0.5000 0.6000 0.5750 0.5750 0.6000 0.6000 C4: Expression, communication, and sharing 0.2656 0.1077 0.3250 0.2500 0.4000 0.3500 0.3750 0.2750 C5: Respect, care and teamwork 0.2455 0.0995 0.3250 0.3750 0.3750 0.3500 0.3500 0.3500 C6: Cultural learning and international understanding 0.2335 0.0947 0.4750 0.6000 0.5750 0.5750 0.6000 0.6000 C7: Planning, organizing and putting plans into practice 0.2554 0.1035 0.5000 0.5000 0.5500 0.5250 0.5500 0.5750 D3: The use of Life Science and Technology 0.0895 Si 0.6250 0.6250 0.6000 0.5250 0.5500 0.5500 Qi 0.6250 0.6250 0.6000 0.5250 0.5500 0.5500 C8: Utilization of technology and information 1.0000 0.0895 0.6250 0.6250 0.6000 0.5250 0.5500 0.5500 D4: Logical thinking and reasoning 0.2037 Si 0.4874 0.5002 0.5498 0.5624 0.5124 0.4744 Qi 0.5000 0.5250 0.5750 0.5750 0.5250 0.5500 C9: Active exploration and study 0.4959 0.1010 0.5000 0.4750 0.5750 0.5750 0.5250 0.5500 C10: Independent thinking and problem-solving 0.5041 0.1027 0.4750 0.5250 0.5250 0.5500 0.5000 0.4000 Integration 1.0000 Si 0.4374 0.4587 0.5056 0.4760 0.4606 0.4446 Qi 0.6250 0.6250 0.6000 0.5750 0.6000 0.6000 Remark: 11: the 1st Semester of Grade 1; 12: the 2nd Semester of Grade 1. C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 12177 (a) The aim of education is to foster students’ sound mind and char- acter – first, enhancing ‘‘D3: The use of Life Science and Technol- ogy’’ may strengthen ‘‘D4: Logical thinking and reasoning’’ for students; then, improving ‘‘D2: Interpersonal and social rela- tions’’ becomes the goal; finally, the curriculum goal gives stu- dents a ‘‘D1: Physical, mental, and spiritual mold’’. Table 8 VIKOR versus weighted average results. m A Publisher B Publisher C Publisher 11 12 11 12 11 12 D1 Physical, mental, and spiritual mold (Si) m = 1 0.0016(4) 0.3350(3) 1.0000(1) 0.5041(2) 0.0000(6) 0.0001(5) m = 0.5 0.0008(5) 0.5008(3) 1.0000(1) 0.6687(2) 0.0000(6) 0.3334(4) (Qi) m = 0 0.0000(5) 0.6667(3) 1.0000(1) 0.8333(2) 0.0000(5) 0.6667(3) D2 Interpersonal and social relations (Si) m = 1 0.0000(6) 0.3153(5) 1.0000(1) 0.6223(3) 0.8984(2) 0.6031(4) m = 0.5 0.0000(6) 0.6577(5) 0.8750(2) 0.6862(4) 0.9492(1) 0.8015(3) (Qi) m = 0 0.0000(6) 1.0000(1) 0.7500(4) 0.7500(4) 1.0000(1) 1.0000(1) D3The use of Life Science and Technology (Si) m = 1 m = 0.5 1.0000(1) 1.0000(1) 0.7500(3) 0.0000(6) 0.2500(4) 0.2500(4) (Qi) m = 0 D4 Logical thinking and reasoning (Si) m = 1 0.1478(5) 0.2933(4) 0.8568(2) 1.000(1) 0.4319(3) 0.0000(6) m = 0.5 0.0739(6) 0.3133(5) 0.9284(2) 1.0000(1) 0.3826(3) 0.3333(4) (Qi) m = 0 0.0000(6) 0.3333(4) 1.0000(1) 1.0000(1) 0.3333(4) 0.6667(3) Integration (Si) m = 1 0.0000(6) 0.3133(4) 1.0000(1) 0.5671(2) 0.3403(3) 0.1069(5) m = 0.5 0.5000(3) 0.6567(2) 0.7500(1) 0.2836(6) 0.4201(4) 0.3034(5) (Qi) m = 0 1.0000(1) 1.0000(1) 0.5000(3) 0.0000(6) 0.5000(3) 0.5000(3) Remark: 11: the 1st Semester of Grade 1; 12: the 2nd Semester of Grade 1. Fig. 6. The causal diagram of total relationship. 12178 C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 (b) For ‘‘D4: Logical thinking and reasoning’’, ‘‘C9: Active explo- ration and study’’ is better than ‘‘C10: Independent thinking and problem-solving’’ from the well-known Knowledge Management point of view. (c) For ‘‘D2: Interpersonal and social relations’’, starting from ‘‘C6: Cultural learning and international understanding’’ may shorten the spiritual distance of the individual; then, there will be improvement in ‘‘C4: Expression, communica- tion, and sharing’’; naturally, ‘‘C5: Respect, care and team- work’’ will be consolidated; and ‘‘C7: Planning, organizing and putting plans into practice’’ will be thoroughly considered. (d) For ‘‘D1: Physical, mental, and spiritual mold’’, ‘‘C2: Appreci- ation, representation, and creativity’’ are the benchmark of ’’C1: Self-understanding and exploration of potential’’; in particular, ’’C3: Career planning and lifelong learning’’ is bright and healthy. ANP provides a general framework for dealing with decisions without making assumptions about the independence of higher-le- vel elements from lower-level elements and about the indepen- dence of the elements within a level as in a typical hierarchy (Saaty, 2005). Thus, in defining the assessment system for teaching material illustrated in the paper, ANP is apparently a more reason- able tool for analysing the network structure with feedback. In- deed, the weighting sequence for the CCs (see Table 5) is seemingly the priority in the practice of CCs. It is contracted with the influence relationships mentioned above using the DEMATEL technique. For long-term concerns, who is responsible for assess- ing teaching materials is a serious concern. The VIKOR method uses an aggregating function R representing ‘‘closeness to the ideal’’. In comparison with the TOPSIS, which determines a solution with the shortest distance from the ideal solution and the farthest distance from the negative-ideal solution (Chen & Hwang, 1992; Tzeng, Shiau, & Teng, 1994), VIKOR can se- C.-H. Chen, G.-H. Tzeng / Expert Systems with Applications 38 (2011) 12168–12179 12179 lect the real ‘‘closest to the ideal’’ solution. On the other hand, a solution by TOPSIS is not always the closest to the ideal. A detailed comparison of TOPSIS and VIKOR has already been presented in the article by Opricovic and Tzeng (Opricovic & Tzeng, 2004, 2007). Furthermore, the Lp-metric of VIKOR represents the grouping le- vel in power p; in other words, p = 1 is the most emphasized group utility; p = 1 represents the most emphasized individuals. For example (see Table 8), regarding Book 11 from A Publisher, one can say that when m ¼ 1; RD1 ¼ 0:0016; RD2 ¼ 0:0000; RD3 ¼ 1:0000; RD4 ¼ 0:1478; RA11 ¼ 0:0000, and the sequences are, respectively, 4, 6, 1, 5, and 6. However, when m = 0, RD1 ¼ 0:0000; RD2 ¼ 0:0000; RD3 ¼ 1:0000; RD4 ¼ 0:0000; RA11 ¼ 1:0000, and the sequences are, respectively, 5, 6, 1, 6, and 1. In this case, Book 11 from A Publisher has good behavior on average, but D3 is the worst. Thus, when m = 0.5, the integration sequence is 3. This means that improving D3 will achieve higher performance. In contrast, Book 12 from B Pub- lisher may indicate the reverse conclusion. 5. Concluding remarks Based on effective concern, begin from Mandarin Chinese. This paper would have mainly advanced work in the field of the assess- ment system for teaching materials. First, a novel MCDM framework was proposed to define the determinants of teaching materials (not exclusively for Mandarin Chinese). Second, the traditional problem of the difficulty of defining the assessment system for teaching materials was resolved based on the novel MCDM approach being proposed. An important reason- ing conclusion was obtained: the more important determinant, the less influent determinant. The influent sequence of determi- nants is as follows: The use of Life Science and Technology (D3) � Logical thinking and reasoning (D4) � Interpersonal and so- cial relations (D2) � Physical, mental, and spiritual mold (D1) (see Fig. 5). However, the weighting sequence of determinants is as fol- lows: Interpersonal and social relations (D2, 0.4054) � Physical, mental, and spiritual mold (D1, 0.3014) � Logical thinking and rea- soning (D4, 0.2037) � The use of Life Science and Technology (D3, 0.0895) (see Table 5). Finally, the difficulty of the traditional MCDM approach in selecting ‘‘rotten apple(s)’’ was also resolved based on the concep- tual advance in achieving the aspired level of criteria. Acknowledgment Over 20 teachers provided feedback on the questionnaire. Although I do not know who are you, I sincerely appreciate your kindness. References Chen, C.H., & Tzeng, G.H. (2009a). Combined DEMATEL technique with a novel MCDM method for creating the aspired intelligent assessment systems for mandarin Chinese teaching materials. In The 10th Asia Pacific industrial engineering and management systems conference (APIEMS2009), Kitakyushu, Japan (pp. 2050–2061). Chen, C. H., & Tzeng, G.H. (2009b). Creating the aspired intelligent fuzzy assessment systems for mandarin Chinese teaching materials. In The 17th national conference on fuzzy theory and its application, Kaohsiung, Taiwan (pp. 787–792). Chen, Y. C., Lien, H. P., & Tzeng, G. H. (2010). Measures and evaluation for environment watershed plan using a novel hybrid MCDM model. Expert Systems with Applications, 37(2), 926–938. Chen, S. J., & Hwang, C. L. (1992). Fuzzy multiple attribute decision making: Methods and applications. Berlin: Springer. Chiu, Y. J., Chen, H. C., Tzeng, G. H., & Shyu, J. Z. (2006). Marketing strategy based on customer behavior for the LCD-TV. International Journal of Management and Decision Making, 7(2/3), 143–165. DfE (1991). Education and training for the 21st century. London: HMSO. Education Commission (2000). Reform proposal for the education system in Hong Kong. Hong Kong: Education Bureau. Freimer, M., & Yu, P. L. (1976). Some new results on compromise solutions for group decision problems. Management Science, 22(6), 688–693. Gabus, A., & Fontela, E. (1972). World Problems, an invitation to further thought within the framework of DEMATEL. Geneva, Switzerland: Battelle Geneva Research Centre. Huang, C. Y., Shyu, J. Z., & Tzeng, G. H. (2007). Reconfiguring the innovation policy portfolios for Taiwan’s SIP mall industry. Technovation, 27(12), 744–765. Huang, J. J., Tzeng, G. H., & Ong, C. S. (2005). Multidimensional data in multidimensional scaling using the analytic network process. Pattern Recognition Letters, 26(6), 755–767. Lee, W. S., Tzeng, G. H., Hsu, C. Y., & Huang, J. M. (2009). Combined MCDM techniques for exploring stock selection based on Gordon model. Expert Systems with Applications, 36(3/2), 6421–6430. Li, C. W., & Tzeng, G. H. (2009a). Identification of a threshold value for the DEMATEL method using the maximum mean de-entropy algorithm to find critical services provided by a semiconductor intellectual property mall. Expert Systems with Applications, 36(6), 9891–9898. Li, C. W., & Tzeng, G. H. (2009b). Identification of interrelationship of key customers’ needs based on structural model for services/capabilities provided by a semiconductor-intellectual property mall. Applied Mathematics and Computation, 215(6), 2001–2010. Lin, C. L., & Tzeng, G. H. (2009). A value-created system of science (technology) park by using DEMATEL. Expert Systems with Applications, 36(6), 9683–9697. Lin, C. J., & Wu, W. W. (2008). A causal analytical method for group decision-making under fuzzy environment. Expert Systems with Applications, 34(1), 205–213. Liou, J. J. H., Tzeng, G. H., & Chang, H. C. (2007). Airline safety measurement using a hybrid model. Air Transport Management, 13(4), 243–249. Mayer, E. (1992). Putting general education to work: The key competencies report. Canberra: Australian Government Publishing Service. Ministry of Education (2002). General guidelines of grade 1-9 curriculum of elementary and junior high school education. Taipei: Ministry of education. Ministry of Education (2008). Education in Taiwan 2008. Taipei: Ministry of Education. Opricovic, S. (1998). Multicriteria optimization of civil engineering systems. Belgrade: Faculty of Civil Engineering. Opricovic, S., & Tzeng, G. H. (2002). Multicriteria planning of post-earthquake sustainable reconstruction. Computer-Aided Civil and Infrastructure Engineering, 17(3), 211–220. Opricovic, S., & Tzeng, G. H. (2004). Compromise solution by MCDM methods: A comparative analysis of VIKOR and TOPSIS. European Journal of Operational Research, 156(2), 445–455. Opricovic, S., & Tzeng, G. H. (2007). Extended VIKOR method in comparison with outranking methods. European Journal of Operational Research, 178(2), 514–529. Ou Yang, Y. P., Leu, J. D., & Tzeng, G. H. (2009). A VIKOR-based multiple criteria decision method for improving information security risk. International Journal of Information Technology and Decision Making, 8(2), 1–21. Ou Yang, Y. P., Shieh, H. M., Leu, J. D., & Tzeng, G. H. (2008). A novel hybrid MCDM model combined with DEMATEL and ANP with applications. International Journal of Operations Research, 5(3), 1–9. Saaty, T. L. (1996). Decision making with dependence and feedback: The analytic network process. Pittsburgh: RWS Publication. Saaty, T.L. (1999). Fundamentals of the analytic network process. In Proceedings of international symposium on analytical hierarchy process, Kobe, Japan. Saaty, R.W. (2003). The analytic hierarchy process (AHP) for decision making and the analytic network process (ANP) for decision making with dependence and feedback. Pittsburgh, PA: Creative Decisions Foundation. Saaty, T. L. (2004). Fundamentals of the analytic network process – Dependence and feedback in decision-making with a single network. Journal of Systems Science and Systems Engineering, 13(2), 71–91. Saaty, T. L. (2005). Theory and applications of the analytic network process. Pittsburg, PA: RWS Publications. Tamura, M., Nagata, H., & Akazawa, K. (2002). Extraction and systems analysis of factors that prevent safety and security by structural models. In Proceedings of the 41st SICE annual conference, 3, Osaka, Japan (pp. 1752–1759). Tzeng, G. H., Chiang, C. H., & Li, C. W. (2007). Evaluating intertwined effects in e- learning programs: A novel hybrid MCDM model based on factor analysis and DEMATEL. Expert Systems with Applications, 32(4), 1028–1044. Tzeng, G. H., Chen, W. H., Yu, R., & Shih, M. L. (2010). Fuzzy Decision Maps–A Generalization of the DEMATEL Methods. Soft Computing, 14(11), 1141–1150. Tzeng, G. H., Lin, C. W., & Opricovic, S. (2005). Multi-criteria analysis of alternative- fuel buses for public transportation. Energy Policy, 33(1), 1373–1383. Tzeng, G. H., Shiau, T. A., & Teng, J. J. (1994). Multiobjective decision-making approach to energy supply mix decisions in Taiwan. Energy Sources, 16(3), 301–316. Tzeng, G. H., Teng, M. H., Chen, J. J., & Opricovic, S. (2002). Multicriteria selection for a restaurant location in Taipei. International Journal of Hospitality Management, 21(2), 171–187. Wu, W. W., & Lee, Y. T. (2007). Developing global managers’ competencies using the fuzzy DEMATEL method. Expert Systems with Applications, 32(2), 499–507. Yu, P. L. (1973). A class of solutions for group decision problems. Management Science, 19(8), 936–946. Creating the aspired intelligent assessment systems for teaching materials 1 Introduction 2 Intelligent assessment systems for teaching materials with MCDM method 2.1 Educational reform of Taiwan (MOE, 2002) 2.2 Assessment system for mandarin chinese teaching materials 2.2.1 Mandarin Chinese’s curriculum goal 2.2.2 Competence indicators of Mandarin Chinese’s curriculum of grade 1–9 curriculum 3 A novel MCDM method based on the DEMATEL technique with ANP 3.1 DEMATEL Technique with ANP 3.2 The ANP method 3.3 VIKOR 4 An empirical study of the aspired assessment systems for mandarin chinese teaching materials 4.1 Background description 4.1.1 Core rationale 4.1.2 Curriculum goals 4.1.3 Core competencies 4.1.4 Learning areas 4.2 Decision problem network relation map structuring by DEMATEL 4.3 Calculating the weights of determinants by ANP 4.4 Compromise ranking by VIKOR 4.5 Discussions and implications 5 Concluding remarks Acknowledgment References 
work_geahfhvb5rfztb2f4bdpjuhmyu	----	C:\Documents and Settings\Michael\MAJ Data\MJackson\MJOnWeb\NewSite\New Versions\PFrame10.prn.pdf PFrame10.doc 15/12/04 Page 1 Problem Frames and Software Engineering Michael Jackson, The Open University {jacksonma@acm.org} Abstract. A general account is given of the problem frames approach to the development of software-intensive systems, assuming that the reader is already familiar with its basic ideas. The approach is considered in the light of the long-standing aspiration of software developers to merit a place among practitioners of the established branches of engineering. Some of its principles are examined, and some comments offered on the range of its applicability. A view of the approach is suggested by an important account of engineering in the aeronautical industry: in particular, the problem classes captured by elementary problem frames are likened to those solved in established engineering branches by normal, rather than radical, design. The relative lack of specialisation in software development is identified as an important factor holding back the evolution of normal design practice in some areas. 1. INTRODUCTION It is widely held that software development should aspire to deserve recognition as an engineering discipline.1 This view motivated the two NATO Software Engineering conferences of 1968 and 1969, and is clearly stated in the report [25] of the first conference: “The phrase ‘software engineering’ was deliberately chosen as being provocative, in implying the need for software manufacture to be based on the types of theoretical foundations and practical disciplines, that are traditional in the established branches of engineering.” The report reflected what was clearly the consensus of the participants: software is, or should become, one more class of engineering product, to be set alongside such established engineering products as bridges, motor cars, chemical plants and aeroplanes. The consensus was unchanged one year later at the second conference in 1969, and has been widely accepted ever since. The participants assumed implicitly that the phrase software engineering was to be narrowly interpreted, to mean the engineering of the software itself—that is, of the structure and content of program texts—and that it was primarily concerned with the processes of software design, programming and testing, and with program execution. The alternative broader interpretation of the phrase, to mean the engineering of change in the world by devising and installing software-intensive systems, was not seriously considered. Surprisingly, the conference participants devoted little of their discussion to exploring and discussing the established practices in the engineering branches they hoped to emulate. (An exception was the talk by W D McIlroy at the first conference, in which he advocated the development and use of mass-produced software components, citing as possible examples trigonometric functions and input-output routines.) Instead they focused on the processes, products and challenges they recognised in their own practice of software development, without comparing them with those of other disciplines. A very interesting paper by Maibaum [24] discusses some aspects of the relationship between software engineering and the established branches, drawing heavily on Vincenti’s illuminating book What Engineers Know and How They Know It [31]. Vincenti writes chiefly about aeronautical engineering in the first half of the twentieth century when the field was 1 In this paper we will use the terms software development and software engineering interchangeably. PFrame10.doc 15/12/04 Page 2 being established by researchers and practitioners. He gives illuminating detail of five case studies including the problems of aerofoil design, a series of wind-tunnel experiments on propellers, the invention and development of flush riveting, the concept of a control volume as a theoretical tool in fluid dynamics, and the process by which the initially vague notion of ‘flying qualities’ became more exact and amenable to quantitative analysis and design. He also discusses the nature and anatomy of engineering knowledge, and the processes by which it evolves and increases over time. The present paper, too, draws on Vincenti’s book as a source of well-founded observations about engineering practice. Its purpose is to discuss some principles and aspects of the problem frames approach (which we will refer to as ‘PF’) to software development, and to relate them to the practice that Vincenti describes. The problem frames approach is not a development method. It is, rather, a perspective and a conceptual framework, embodying a certain way of looking at an important group of problem classes and of structuring the intellectual processes of developing good solutions. It is intended to be usable in several ways: constructively, to guide development; analytically, to help understanding of completed developments; and methodologically, to help understanding of existing and proposed approaches to software engineering. 2. THE MACHINE AND THE PROBLEM WORLD The PF view of software development takes as its starting point the alternative broader interpretation of the phrase software engineering, corresponding closely to Rogers’s definition [30] of engineering, quoted and amplified by Vincenti: “Engineering refers to the practice of organising the design and construction [and, I would add, the operation] of any artifice which transforms the physical world around us to meet some recognised need.” This definition succinctly captures the PF view. The development task is to design and construct an artifice. In PF we call this artifice the machine, constructed by building software that is then executed on a general-purpose computer, specialising the computer to serve a particular purpose. That purpose is to meet a recognised need, which we call the requirement. Satisfying the requirement involves transforming the physical world around us: in PF, the part of the world to be transformed, in which the requirement is located, is called the problem world. Using Polya’s term, we may say that the principal parts [27] of a software development problem are the machine, the problem world, and the requirement. Their relationships are shown in the generalised problem diagram in Figure 1: The machine interacts with the problem world at an interface of shared phenomena a. Typically, these phenomena are events and states, controlled either by the problem world or by the machine and shared at input-output ports or registers of the machine. (It is often convenient to raise the level of abstraction here—for example, by regarding an extremely reliable input-output device, such as an attached printer, as an integral part of the machine: the phenomena a would then include print events.) The machine and the problem world are both physical, conforming to their characterisation in Rogers’s definition; their interactions at a are unmediated and therefore also physical. The requirement is shown by a dashed oval, Figure 1. Generalised Problem Diagram Machine Problem World Requirement a b PFrame10.doc 15/12/04 Page 3 indicating its intangible quality. The requirement is not a tangible part of the problem: it is a predicate or condition on the problem world that the machine is required to bring about. The link between the requirement and the problem world is represented by a dashed line. This link denotes references by the requirement to physical phenomena b of the problem world. These are the phenomena that the customer for the system would observe to determine whether the requirement is satisfied. In general, the phenomena b, which we may call the requirement phenomena, are distinct—and sometimes disjoint—from the phenomena a. We may call the latter the specification phenomena: they are the phenomena in terms of which the external behaviour of the machine—that is, of the software—must eventually be specified. The generalised problem diagram of Figure 1 emphasises the physical phenomena of the requirement, the problem world and the specification: the machine and problem world, their interaction, and the requirement, are all to be understood in terms of physical phenomena rather than in terms of purely mathematical abstractions. That is not to say that PF eschews abstraction, but rather that it insists that abstractions must be firmly grounded in observable physical reality. For example, if we consider the striking of a key by a computer user to be a shared phenomenon at the specification interface, we are abstracting from the causal chain that runs from depression of the keytop to assigning the corresponding encoded value to a machine register. This causal chain may be very complex, involving interpretation of the key codes and execution of a debouncing algorithm; but we choose to regard the key depression and the assignment as a single shared event. Such abstractions are inevitable in any treatment of physical phenomena. What we insist on is that any abstraction can be unambiguously explained in terms of phenomena that we can observe in the physical world. Clearly this stipulation excludes certain problem classes. For example, PF makes no claim to offer guidance in the problems of factorising a large integer, finding the shortest path in a graph, or beating the world champion at chess or go. Nor can such problems be brought within the scope of PF by a gratuitous reification—for example, by stipulating a physical representation of the integer or graph or chessboard in a disk file. In some problems, such as the construction of a single-person computer game, the physical problem world may contain nothing other than the operator or user, and the problem world phenomena (other than the specification phenomena) are purely imaginary. The PF approach may then need at least some modification to accommodate the absence of an objective reality to which problem world descriptions can be referred. This kind of problem is discussed in [1]. However, it is important to recognise that problem worlds involving the behaviour of people interacting with the system are fully appropriate to PF: an interesting extension [2] of the approach allows the knowledge they acquire and use in their activities to be brought within the PF scope. 3. PROBLEM REDUCTION Because PF is concerned with engineering in the world, it demands a clear distinction among three descriptions: • the requirement R, describing the customer’s requirement—how the world is desired to be—in terms of the requirement phenomena b; • the specification S, describing, in terms of the specification phenomena a, the machine behaviour that must result from execution of the software to be developed; • the world properties W, describing the given properties of the problem world that hold independently of the machine’s behaviour and that govern the possible relationships between the specification and requirement phenomena. To show that the machine will indeed satisfy the requirement it is then necessary to show that: PFrame10.doc 15/12/04 Page 4 S,W ú R In other words, if a machine satisfying S is installed in a problem world satisfying W, then the requirement R will be satisfied. (A deeper and more detailed discussion of this entailment can be found in [10] and [12].) The gap between what is desired at the interface b and what can be directly monitored and controlled by the machine at interface a is bridged by the given properties of the problem world. Showing, formally or informally, that this relationship holds among the three descriptions may be called the basic problem concern. PF does not mandate a sequence of development tasks. But in concept we can imagine a development process that begins by capturing the customer’s requirement R, and proceeds, using the given properties W, to devise a machine behaviour specification S. We may call such a process full problem reduction: starting from a problem involving the phenomena a and b and the problem world properties, it derives a reduced problem in which the phenomena b and the problem world do not appear at all. A problem of engineering in the world has been reduced to the problem of building a machine with a specified external behaviour. When the problem world is more complex, and is structured as an interacting set of problem domains, the problem can be partially reduced in each of a series of steps, successively removing domains that are farthest from the machine. (Hall and Rapanotti have pointed out the connection between problem reduction and the weakest-environment calculus [21].) Eventually, of course, once this reduction process is complete, a programming process will create a program whose execution will satisfy S. Software development has traditionally dealt chiefly in what are, in this sense, reduced problems: the problem was seen solely as satisfying a given specification—often informal and sometimes unwritten—rather than including the process of deriving the specification. Early computers were connected very loosely and asynchronously, by data on punched card or paper or magnetic tape, to their problem worlds: it seemed natural to start at the point where the data was presented to the computer. Most software, apart from the compiler, the operating system kernel, and the software to manage input-output and system resources, consisted of batch programs performing mathematical computations or data processing tasks. Programs for mathematical computations were often written by people already fully familiar with the task to be performed, who saw no need for a written specification. Data processing tasks were often simply specified as reproducing, sometimes with little or no change, the behaviour of an existing manual or punch-card system. In general, the production, content and format of program specifications were not considered to be a significant part of the software development task. This view was memorably expressed by Dijkstra [7]: “The choice of functional specifications—and of the notation to write them down in —may be far from obvious, but their role is clear: it is to act as a logical ‘firewall’ between two different concerns. The one is the ‘pleasantness problem,’ ie the question of whether an engine meeting the specification is the engine we would like to have; the other one is the ‘correctness problem,’ ie the question of how to design an engine meeting the specification. ... the two problems are most effectively tackled by ... psychology and experimentation for the pleasantness problem and symbol manipulation for the correctness problem.” Dijkstra was concerned to maintain the mathematical purity of computer science, and rejected the notion of software engineering in any sense. He dubbed software engineering ‘the doomed discipline’, and claimed that its goal was self-contradictory because computers were purely symbol manipulation machines. The ‘pleasantness problem’—that is, obtaining the functional specification—was for others to address. Computer scientists and software developers should PFrame10.doc 15/12/04 Page 5 restrict their attention to the ‘correctness problem’, and should address it by formal mathematical techniques. This narrow view of the software development task has had an enduring effect. Although lip service is almost universally paid to the need to analyse and describe the problem world, the reality is that most descriptions—certainly most formal descriptions—constructed in software projects are descriptions of the machine rather than of the world. What may have begun as an effort to describe the world slips quickly into a focus on an object or database structure that is intended to represent it in the machine. The PF emphasis on the explicit investigation, description and use of the given problem world properties can be seen as an attempt to exert a countervailing intellectual influence. PF views the process of problem reduction as an integral and central part of the software engineering discipline. 4. REDUCED PROBLEMS IN ENGINEERING It may perhaps seem that in spite of Rogers’s definition the PF emphasis on the problem world is at odds with established engineering practice. At first sight, engineering practice seems to focus very directly on the product, or artifice, to be designed and constructed, leaving the world around it very much in the background. Automobile engineers appear to think much more about cars than about roads or drivers, so justifying an analogous emphasis in software engineering on the machine rather than the problem world. A distinction should be made here between what we may call local and ubiquitous problems. In a local problem the problem world is unique to a particular place, and the engineer must examine that problem world in all its particularity. Consider, for example, the building of a road or rail bridge over a navigable river. The bridge designer must obviously carry out an explicit and detailed investigation and analysis of the given properties of the particular problem world: the character of the ground on which the bridge supports will be erected; the river currents; the sizes and speeds of vessels navigating the river; the local winds; the density, nature and flow patterns of the planned traffic; and the seasonal variations in temperature and other atmospheric conditions. In a ubiquitous problem, by contrast, the engineering product is intended to be used anywhere that the operational environment exhibits certain broad characteristics that are essentially independent of the specific operational locality. Automobile engineering is concerned with wheeled vehicles that travel on terra firma; aeronautical engineering is concerned with vehicles that travel in the earth’s atmosphere. The problem world properties of interest are then location-independent properties, and do not demand a fresh explicit description of roadways or of the atmosphere for each new car or aircraft design. Rather than ad hoc products of the development project in hand, these descriptions are standard reference material. Vincenti writes of the indicative problem world properties to be considered by the aeronautical engineer: “Descriptive knowledge is knowledge of how things are. Descriptive data needed by [aeronautical] designers include physical constants (acceleration of gravity, for example) as well as properties of substances (failing strength of materials, coefficients of viscosity of fluids, etc.) and of physical processes (rate of chemical reactions and so forth). Occasionally they deal with operational conditions in the physical world (frequency and strength of atmospheric gusts for aircraft fatigue-loading calculations). As we have seen with flying qualities, they also encompass information on human beings (maximum forces exerted by pilots [on an aircraft’s manual controls]).” PFrame10.doc 15/12/04 Page 6 Certainly, some of this descriptive knowledge—strength of materials and rate of chemical reactions—is clearly about the substrate of the engineering artifice itself rather than its problem world or operating environment. Frequency and strength of gusts, however, and forces exerted by pilots, are certainly properties of the problem world. But they do not require to be established afresh on each occasion. In respect of these properties the engineers need only address the reduced problem, reduced in the light of standard knowledge of the properties of the problem world or environment. 5. NORMAL AND RADICAL DESIGN Vincenti draws on his own experience of aeronautical engineering and on the work of the philosopher Michael Polanyi [26] and the historian Edward Constant [6] to explain the central importance of normal—as opposed to radical—design for established engineers. Normal design is what allows them to succeed where they do and as often as they do. Of an engineer practising normal design he writes: “The engineer engaged in such design knows at the outset how the device in question works, what are its customary features, and that, if properly designed along such lines, it has a good likelihood of accomplishing the desired task.” “Every device possesses an operational principle, and, once the device has become an object of normal, everyday design, a normal configuration. Engineers doing normal design bring these concepts to their task usually without thinking about them.” In short, it is usually otiose, or even harmful, to rework from basic principles a problem whose requirements, design and solution have been firmly established by long and successful experience. Vincenti gives the example of automobile design: “Automobile designers of today usually (but not invariably) assume without much thinking about it that their vehicle should have four (as against possibly three) wheels and a front-mounted, liquid-cooled engine.” Vincenti lays some stress on the hierarchical structure of engineering products in the established branches. He distinguishes between devices and systems. Devices are “single, relatively compact entities”, while systems are “assemblies of devices brought together for a collective purpose.” The distinction is, of course, relative. The struts of an aircraft’s landing gear are a device—a part of the landing gear viewed as a system. The landing gear, viewed as a device, is a part of the aircraft regarded as a system. The aircraft itself can be viewed as a device forming a part of the airline system, or even of a national or global system of air transport. Although the distinction between devices and systems is relative in this sense, it is also closely associated with the distinction between normal and radical design. Normal design can be applied only to a device, to a “single, relatively compact entity”. Radical design is always necessary for a system, where there is no history of successful development of closely fitting precedents. Radical design may also be necessary when a new kind of device is required. Then the engineer’s ambitions must be dramatically lowered: “... how the device should be arranged or even how it works is largely unknown. The designer has never seen such a device before and has no presumption of success. The problem is to design something that will function well enough to warrant further development.” PFrame10.doc 15/12/04 Page 7 In effect, only normal design can realistically hope to produce a reliable product with no major defects or failures. Though less conspicuous than radical design, normal design makes up by far the bulk of day-to-day engineering enterprise. Unfortunately, this is not true of software engineering. 6. ELEMENTARY PROBLEM FRAMES Elementary problem frames can be regarded as defining problem classes in software engineering that are currently the object of normal design. Their machines, set firmly in the context of the problem worlds that provide their purpose, play the role that Vincenti’s well- understood devices play in the established engineering branches. Particular elementary problem frames, specialised from the universal problem diagram of Figure 1, capture classes of software development problems whose software solutions take the form of individual programs or, at most, very small systems of programs. For example: • In a WorkPieces problem a User edits a WorkPiece such as a text or graphic document. The requirement is that the edit commands issued by the User should effect appropriate corresponding changes in the WorkPiece. • In a Required Behaviour problem the Control Machine is required to impose a particular behaviour on a Controlled Domain. • In an Information Display problem the Information Machine is required to monitor the state and behaviour of a Real World and to display information about it on a Display. Problem frames do not aim to capture classes of problems of realistic size and complexity. Most of them capture simple subproblems that can appear only as ingredients or aspects of realistic problems. For example, the WorkPieces frame captures problems of simple document editing. The subsequent use of the document, and even the user’s ability to inspect different parts of the document while it is being edited, are ignored: the solution to a WorkPieces problem is not useful in isolation. Each problem frame constitutes what we might call a problem pattern. The very well- known work on object-oriented design patterns [9, 3] has some obvious similarities, and is clearly concerned with the identification of devices that, once identified, can become the objects of normal design. But design patterns are firmly located within the machine: all parts of each pattern are represented in programming terms. The work on analysis patterns [8] is an interesting treatment of some problem fragments in an object-oriented setting. A problem pattern stipulates: • a decomposition of the problem world into a particular set of physical problem domains interacting with each other and with the machine in a particular topology; • a characterisation of each problem domain according to its physical properties—causal, lexical or biddable—as they affect the problem (the Controlled Domain is causal, the WorkPiece is lexical, and the User is biddable); • a characterisation of each interface according to the types of the subsets of the phenomena—events, states, symbols—that are shared at the interface and to the assignment of control of each subset to one of the sharing domains; • the nature of the requirement and of its link—none, reference only, or constraining—to each problem domain. Problems that fit different elementary frames specialise the basic problem concern in different ways, giving rise to frame concerns of different forms. The specialisation arises PFrame10.doc 15/12/04 Page 8 because the given properties W of each problem domain d must be separately described, and these descriptions must be combined with each other and with the specification S and requirement R in a way that respects the connectivity of the frame diagram and the control of shared phenomena. The frame concern of a problem class can be regarded as loosely analogous to the operational principle of a device class in normal design—how the characteristic parts of the device fulfil their special function in combining to an overall operation which achieves the device’s purpose. Vincenti cites Sir George Cayley’s famous statement of 1809 [5] in which he characterised the principle of the fixed-wing aircraft: “to make a surface support a given weight by the application of power to the resistance of air.” The surface provides the lift because the applied power drives the surface against the resistance of the air. In addition to the frame concern each problem frame raises a number of specific concerns that must be addressed if the solution is to be acceptable. For example: • Initialisation. The machine necessarily begins operation in its initial state, and at that moment the problem world is also in one of some set of states. If the developer has not considered the problem world state that in fact holds initially in a particular case, the specified machine behaviour is unlikely to satisfy the requirement, at least for some initial period of operation and possibly for the whole lifetime of the system. • Breakage. A causal domain may be damaged if certain sequences of operations are performed on it. For example, some component of a Controlled Domain may be damaged if certain operation sequences are performed by the machine. The developer must identify such sequences, and ensure that the machine always avoids them. • Reliability. A causal problem domain typically satisfies its properties description with high—but never with perfect—reliability. If the reliability is too low in relation to the criticality of the system it is necessary to detect, or even to anticipate, failures and to prevent, mitigate or compensate their effects. • Information deficit. In an Information Display problem the information directly and currently available to the machine at the time it must produce each display output may be inadequate. It may have been available at an earlier time; or it may be information that must be accumulated from the beginning of execution; or it may be deducible only by elaborate calculation or estimation from the information available earlier. The developer must then find some effective way of overcoming this information deficit. In general, different problem frames raise different specific concerns. Information deficit does not arise in WorkPiece problems. Information Display problems do not raise reliability concerns. Breakage is not a concern in WorkPieces problems. To a significant extent the set of concerns that must be addressed, and the detail of each concern, depend on the characteristics of the frame’s problem domains. Breakage and reliability do not arise for lexical domains; information deficit does not arise for biddable domains. However, the importance of a particular concern can not be finally determined by considering abstract characteristics either of domains or of problem frames. It must be ultimately determined by the knowledge, experience and judgement embodied in a particular class of normal design task, applied to the particular problem world of the case in hand. The engineer practising normal design knows in advance what concerns will merit most of his attention. Many of the failures documented in P G Neumann’s Risks forum [29] can be attributed to the absence of established normal design practice for at least the failing part of the system. With hindsight the mistake or neglect that led to the failure is conspicuous, and we may often wonder why the developers did not see it. The reason is that they were doing radical design, PFrame10.doc 15/12/04 Page 9 not normal design, and their knowledge, experience and judgement were not specifically honed for the particular design task they were doing. In Vincenti’s words, they “had no presumption of success.” Depending on the circumstances, blame for the failure may be laid at the developers’ door for ignoring existing normal design practice; or at the customer’s door for stipulating requirements that went too far beyond what can be reliably delivered by current practice; or, more generally, at the software development community’s door for failing to evolve a normal design practice when enough knowledge and experience existed to do so. 7. INSTANTIATING ELEMENTARY FRAMES In an established branch such as aeronautical engineering [6], the task of normal design is “the improvement of the accepted tradition or its application under new or more stringent conditions”. Relatively small design changes give quantitative increments in cost, performance or economy, while retaining the customary configuration of parts and the accepted operating principle. Examples in software engineering are most readily found in the improvement of what we may call internal software components—components all of whose interactions are with other formally specified software components, within the machine. The classic handbook for a large variety of such software devices is Knuth’s magnum opus The Art of Computer Programming [17,18,19], where detailed discussions are given of the construction and performance of algorithms for sorting, tree traversal, garbage collection and many other purposes. Some examples can be cited closer to the problem world. Spell checkers, for instance, appear to have become the object of normal design, to judge by their greatly improved speed and reliability. In the design of solutions for problems fitting elementary problem frames the need for incremental quantitative improvement may be present, but is usually dominated by the need to instantiate known solutions for the detailed properties of specific problem domains. Instances of the WorkPieces frame, for example, will differ not only in response time and speed of saving the document being edited, but more importantly in the different lexical structures and operations of the WorkPieces they are required to handle. Similarly, instances of the Required Behaviour frame will differ in the structural and dynamic properties of the Controlled Domain. These differences provide the core subject matter for software development methods. For example, a model-based method such as Z [13] or VDM [15] is very appropriate to instantiating the common data structures that characterise a set of WorkPieces together with the repertoire of editing operations which can be applied to them. Similarly, problems fitting the Transformation frame, in which the task is to produce sequential output streams from sequential input streams, can often be appropriately solved by the JSP design method [14]: the sequential structures of the streams are described in regular expressions which are then combined to give the required program structure. 8. COMPOSITE ELEMENTARY PROBLEMS A problem fitting an elementary frame will sometimes be soluble by a simple machine, whose specification can be a single text. But sometimes a specific concern will demand a decomposition into two or more subproblems. For well-understood elementary problem frames this decomposition will be highly standardised. This standardised decomposition is another aspect of what Vincenti calls “a given operational principle and normal configuration”: it is understood not only how the device should work but also what arrangement of parts will best allow it to work in that way. PFrame10.doc 15/12/04 Page 10 Consider, for example, an Information Display problem in which the requirement is to maintain a Display showing information about a Real World. Figure 2 shows the problem frame diagram: The requirement is that the visible phenomena of the Display should correspond in a certain way to the state of the Real World. If the particular problem in hand presents a significant information deficit problem (as most Information Display problems do), the standard solution is to decompose the problem into two subproblems communicating by a Model domain, as shown in Figure 3: The Model domain is an invented lexical domain, often a database on disk. One subproblem constructs and maintains the Model, satisfying the requirement that the Model should correspond in a certain way to the Real World; the other subproblem maintains the Display, using the Model and satisfying the requirement that the display should correspond in a certain way to the Model. If we can express the requirements Display À Model and Model À Real World as relations, it seems clearly desirable that their relational composition should be equal to the original requirement Display À Real World. Introducing the Model domain is an unavoidable, and necessarily universal, practice in software development. Decomposing the problem into the two subproblems shown, however, is not standard practice: essentially it is a separation into two concurrent processes. They will subsequently need to be composed. In normal design of a composite device, the technique and mechanism of composition is standardised. Here, for example, a standard technique of mutual exclusion, applied to critical regions of appropriately chosen granularity, may be sufficient. 9. DECOMPOSING REALISTIC PROBLEMS If problems fitting elementary problem frames are devices in Vincenti’s terms, then realistic problems are usually systems—‘assemblies of devices brought together for a collective purpose’—and consequently the object of radical design. To repeat Vincenti’s words, “The Display À Real World Real World Display Info’n Machine Figure 2. Information Display Problem Diagram Figure 3. A Standard Decomposition Model À Real World Real World Model Builder Display À Model Display Model User Model Model PFrame10.doc 15/12/04 Page 11 designer has never seen such a device before and has no presumption of success. The problem is to design something that will function well enough to warrant further development.” The need for radical design springs from either or both of two sources. Some of the devices making up the system may be currently unfamiliar, themselves demanding radical design; and the particular combination of devices may be novel, posing unfamiliar problems of interaction. The PF approach aims to contain and minimise the radical aspects of development in any particular case by application of two interlocking principles. The first principle is that a realistic problem should be decomposed, so far as possible, into subproblems that fit known elementary problem frames. That is: all those parts or aspects of the system that can be properly regarded as familiar devices should be identified, and developed by the applicable techniques of normal design. What makes PF distinctive here is the granularity and compactness of the parts into which the problem is decomposed. Elementary problem frames capture closed systems, which, even if not realistically useful in isolation, are complete in themselves. Each subproblem has its own machine, its own problem world, and its own requirement, and has no external connections. This decomposition rule has important consequences for the granularity of the decomposition. Some commonly used development techniques, by contrast, decompose problems into open parts, usually of smaller granularity. This is true, for example, of decomposition into programming objects (which offer and invoke methods); of dataflow decomposition into sequential processes (which assume stream or database communication with other processes); and of requirements decomposition into use cases (each detailing one episode of user interaction within a larger, usually unspecified, context). Each of these techniques has its place; but from a PF point of view none is satisfactory as the chief form of problem decomposition. The second principle, flowing from the first, is that subproblem composition is a major development topic in its own right, demanding explicit and separate consideration. Because the devices identified in the decomposition are closed systems, and to a large extent standardised, their normal development pays—and should pay—no attention to the need to combine them with other devices into a realistic system. But eventually, of course, the decomposed subproblems must be recombined or composed. Composition brings its own concerns: the composition of the subproblem requirements must be shown to approximate closely enough to the original problem requirement; the subproblem machines must be configured for some kind of parallel execution, cooperating to whatever extent is necessary; conflicts between subproblems must be resolved. In the PF approach, the central principle to be applied here is that treatment of the composition concerns should be deferred until each subproblem has been identified and analysed to the point of obtaining a satisfactory machine behaviour specification and problem world description. Until this point, the identification and treatment of the decomposed subproblems should—in principle—take no account of the eventual need to make them work together. This principle stands in opposition to the most commonly used techniques, in which consideration of the composition concerns is integrated at the outset into the treatment of the subproblems. A seeming advantage of these traditional techniques is that composition can become a relatively trivial mechanical matter of matching actual to formal parameters or output stream to input stream definitions. In practice, however, the widespread emphasis on integration testing indicates that the common technique is hard to carry through successfully. An important advantage of the PF technique of deferring the composition concerns is that the subproblems are then seen in their purest forms, in which they correspond exactly to known PFrame10.doc 15/12/04 Page 12 devices, not yet complicated by their eventual interactions. Further, the composition concerns can be dealt with later as an explicit task of composing known devices into a system. If the deferred composition concerns then prove to demand substantial reworking of the subproblems to be composed, this is not a disadvantage of the separation: it is rather an indication that dealing simultaneously with the subproblems and their composition concerns would have been very difficult. 10. COMPOSITION CONCERNS Composition concerns arise when subproblems have parts of their problem worlds in common. Examples of composition concerns are: • Interference. Two subproblems interfere if one causes change in a problem domain and the other inspects the same domain: for example, the Model Builder machine changes the Model domain while the Model User machine inspects it. The Model User subproblem has a domain properties description of the Model, and it is necessary to ensure that the Model conforms to this description during each inspection. A suitable granularity must be chosen for mutual exclusion of atomic changes and inspections. • Scheduling. A scheduling concern arises when interference can not be dealt with by atomicity of changes and inspections. Suppose, for example, that in a library system one subproblem deals with membership and another deals with borrowing and returning books. In the books subproblem membership is treated as static, although in reality it is constantly changing as people join and leave the library. Questions then arise about book loans towards the end of a membership period. Can a member whose membership has only one week to run borrow a book for two weeks? What if a borrower resigns from membership before returning the book? • Conflict. The requirements of two subproblems that cause change in the same domain may sometimes conflict. For example, in a lift control problem the subproblem requirement of providing lift service may conflict with the subproblem of ensuring safe operation in the presence of mechanical and electrical faults: lift service may demand movement of the lift car to service a request, while safety demands that it be locked by the emergency brake at its current position in the shaft. Clearly two conflicting requirements can not both be satisfied: one must be given precedence. The interference concern can be regarded as an implementation matter. The properties description of the Model domain as seen by the Model User machine has already been given. To address the concern it is necessary only to ensure that this description holds at the relevant times. But the scheduling and conflict concerns, by contrast, raise requirements issues that were not visible when the subproblems were considered in isolation. 11. IMPLEMENTING COMPOSITION In the PF approach architecture can be seen as the composition and implementation of the subproblem machines in a realistic problem. Because a full-scale realistic problem is likely to be the object of radical design, there are several approaches that can be valuable in different cases. Two reported approaches aim to bring more realistic problem classes within the ambit of normal design. The first is reported in [11], where the machine of a subproblem is assumed to be specialised in a standard way that implements a particular scheduling and prioritisation scheme. The problem is to satisfy customer orders in a warehouse. Investigation of the requirement shows the potential for interference between orders for the same product, leading PFrame10.doc 15/12/04 Page 13 to duplicate allocation of the same stock. It also shows the need for a fair scheduling, ensuring that an order placed earlier will be allocated scarce stock in preference to a later competing order. The solution proposed is to use a specialised machine rather than the generic general- purpose computer usually assumed in an elementary problem frame. The specialised machine has built-in behavioural properties one-at-a-time and first-come-first-served, each capable of being instantiated for the problem in hand. In effect, aspects of the requirement are being delegated to the implementation. Some part of the requirement need not be stated in explicit detail because it is implicit in the choice of the specialised machine. The second of these approaches is reported in [28], where the idea of an architectural frame or AFrame is introduced as an elaboration of the problem frame diagram to accommodate an early architectural decision. One of their examples is a Transformation problem frame in which a standard decomposition into sequential processes is expected, and the architectural decision has been made to implement composition of the resulting machines by the pipe-and-filter architecture. Another example draws on the well-known MVC pattern [20]. Both of these approaches can be regarded as enlarging the set of devices that can be the object of normal design by defining specialised composite elementary problem frames. A different approach is reported in [22], where the task of composing conflicting requirements is addressed. In this approach an additional machine—which may be viewed as an architectural connector—is introduced between the conflicting machines and their common problem domains, and arbitrates in cases of conflict. In [22] the approach is examined in several different cases of real and potential conflict, and offers, at least in part, the possibility of application in the composition of many different realistic problems. 12. SPECIALISATION AND THE GROWTH OF KNOWLEDGE It is apparent from Vincenti’s account that an important part of the success of the established branches of engineering can be attributed to specialisation. Vincenti takes it entirely for granted that aeronautical engineers do not work on problems in other branches: they do not design bridges or chemical plants or even motor cars. The specialisation goes deeper, to much lower levels. For example, W F Durand and E P Lesley devoted their work from 1916 to 1926 on the design of propellers [31]. One whole chapter of Vincenti’s book is devoted to the development by the aircraft companies of flush riveting, in which the aerodynamic properties of wings and fuselages are improved by the use of rivets whose head do not protrude above the surface of the skin. Even a cursory inspection of the literature of an established branch of engineering reveals a similar level of specialisation. Whereas much software engineering literature (like the present paper, it might be said) tends to focus on general concerns or on methods or approaches proposed for universal or nearly universal application, the literature of established engineering tends rather to focus on highly specialised applications and problems. For example, one issue of The Journal of Structural Engineering [16] contains papers on such topics as “Performance Evaluation of Controlled Steel Frames under Multilevel Seismic Loads”, “Stress Concentration Factors of Doubler Plate Reinforced Tubular T Joints”, “Reliability Assessment of Highway Truss Sign Supports”, and “Wind Sensitivity of Recycled Plastic Soundwalls”. This kind of specialisation seems an essential foundation for the incremental growth of knowledge in the established branches of engineering. Specialisation is essential to the growth of knowledge because it allows experience to be accumulated effectively and systematically. The accumulation of experience and knowledge demands a conceptual structure within which it can be evaluated and stored, and from which it can be reliably and easily retrieved when it is relevant. Excessively narrow specialisation PFrame10.doc 15/12/04 Page 14 results, perhaps, in a smaller contribution to knowledge than would otherwise have been possible, but in a contribution nonetheless. Absence of specialisation often results in no contribution at all: the knowledge gained is too vaguely expressed, and is added to a largely unstructured corpus of software engineering knowledge in general rather than to a specific place or places in a detailed structure. The existence of a structured corpus of knowledge offers no guarantee that the knowledge will be used—that is, retrieved and applied when it should. The notorious failure of the Tacoma Narrows Bridge in 1940 was due to wind- induced oscillation of a kind that had been observed repeatedly in suspension bridges and recorded from as early as 1836 [23]. But the absence of such a knowledge structure virtually guarantees that it will not be used. ACKNOWLEDGEMENTS Tom Maibaum brought Vincenti’s book to my attention (he acknowledges his debt to William Newman, who brought it to his). Work with colleagues at the Open University, at the University of Newcastle, at the University of Leicester and at MIT has provided much enjoyment, illuminated many aspects of PF, and revealed many needs and opportunities for its further development and application. REFERENCES [1] Ian K Bray and Karl Cox; The Simulator; Another, Elementary Problem Frame? in Proceedings of the 9th International Workshop on Requirements Engineering: Foundation For Software Quality (REFSQ03), 2003. [2] John Brier, Lucia Rapanotti and Jon G Hall; Problem Frames for Socio-technical Systems: predictability and change; in Proceedings of the 1st International Workshop on Advances and Applications of Problem Frames, 2004. [3] Frank Buschmann, Regine Meunier, Hans Rohnert, Peter Sommerlad and Michael Stahl; Pattern-Oriented Software Architecture: A System of Patterns; John Wiley 1996. [4] J R Cameron; JSP & JSD: The Jackson Approach to Software Development; IEEE Computer Society Press, 2nd Edition 1989. [5] Sir George Cayley, Bart; On Aerial Navigation; Nicholson’s Journal, issues of November 1809, February 1810, March 1810. [6] E W Constant; The Origins of the Turbojet Revolution; John Hopkins University Press, Baltimore 1980. [7] E W Dijkstra; On the Cruelty of Really Teaching Computer Science; Communications of the ACM Volume 32 Number 12, December 1989. [8] Martin Fowler; Analysis Patterns: Reusable Object Models; Addison-Wesley 1996. [9] Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides; Design Patterns: Elements of Object-Oriented Software; Addison-Wesley, 1994. [10] Carl A Gunter, Elsa L Gunter, Michael Jackson and Pamela Zave; A Reference Model for Requirements and Specifications; Proceedings of ICRE 2000, Chicago Ill, USA; reprinted in IEEE Software Volume 17 Number 3, pages 37-43, May/June 2000. [11] Jon G Hall, Michael Jackson, Robin Laney, Bashar Nuseibeh and Lucia Rapanotti; Relating Software Requirements and Architectures Using Problem Frames; Proceedings of the 2002 International Conference on Requirements Engineering (RE’02), Essen, 2002. PFrame10.doc 15/12/04 Page 15 [12] Jon G Hall and Lucia Rapanotti; A Reference Model for Requirements Engineering; in Proceedings of the 11th Joint International Conference of Requirements Engineering (RE’03), 2003. [13] Ian Hayes (ed); Specification Case Studies; Prentice-Hall International, 1987. [14] M A Jackson; Constructive Methods of Program Design; in Proceedings of the 1st Conference of the European Cooperation in Informatics, pages 236-262; G Goos & J Hartmanis eds; Springer-Verlag LNCS 44, 1976; reprinted in [4] pages 46-62. [15] Cliff Jones; Systematic Software Development Using VDM; Prentice-Hall International, 2nd Edition 1990. [16] The Journal of Structural Engineering November 2002; Volume 128, Issue 11. [17] Donald E Knuth; The Art of Computer Programming; Volume 1: Fundamental Algorithms; Addison-Wesley, 1968. [18] Donald E Knuth; The Art of Computer Programming; Volume 2: Seminumerical Algorithms; Addison-Wesley, 1969. [19] Donald E Knuth; The Art of Computer Programming; Volume 3: Sorting and Searching; Addison-Wesley, 1972. [20] G E Krasner and S T Pope; A cookbook for using the Model-View-Controller user interface paradigm in Smalltalk-80; Journal of Object-Oriented Programming Volume 1 Number 3, pages 26-49, August/September 1988. [21] Luming Lai and J W Sanders; A Weakest-Environment Calculus for Communicating Processes; Technical Report PRG-TR-12-95, Oxford University Computing Laboratory, 1995. [22] Robin Laney, Leonor Barroca, Michael Jackson and Bashar Nuseibeh; Composing Requirements Using Problem Frames; in Proceedings of the 2004 International Conference on Requirements Engineering RE’04, IEEE CS Press, 2004. [23] Matthys Levy and Mario Salvadori; Why Buildings Fall Down: How Structures Fail; W W Norton, 1994. [24] Tom S E Maibaum; What we teach software engineers in the university: do we take engineering seriously? Proceedings of the 6th European conference held jointly with the 5th ACM SIGSOFT international symposium on Foundations of software engineering, Zurich, Switzerland, September 22-25, 1997. [25] Peter Naur and Brian Randell eds; Software Engineering: Report on a conference sponsored by the NATO SCIENCE COMMITTEE, Garmisch, Germany, 7th to 11th October 1968; NATO, January 1969. [26] Michael Polanyi; Personal Knowledge: Towards a Post-Critical Philosophy; Routledge and Kegan Paul, London, 1958. [27] G Polya; How To Solve It; Princeton University Press, 2nd Edition 1957. [28] Lucia Rapanotti, Jon G. Hall, Michael Jackson and Bashar Nuseibeh; Architecture- driven Problem Decomposition; in Proceedings of the 2004 International Conference on Requirements Engineering RE’04, Kyoto, IEEE CS Press, 2004. [29] http://www.catless.ncl.ac.uk/Risks [30] G F C Rogers; The Nature of Engineering: A Philosophy of Technology; Palgrave Macmillan, 1983. PFrame10.doc 15/12/04 Page 16 [31] Walter G Vincenti; What Engineers Know and How They Know It: Analytical Studies from Aeronautical History; The Johns Hopkins University Press, Baltimore, paperback edition, 1993. 
work_genlt3zrefdgplllgllgwnx3sm	----	Using situation calculus for e-business agents Using situation calculus for e-business agents Conan C. Albrecht, Douglas D. Dean, James V. Hansen* Marriott School of Management, Brigham Young University, Provo, UT 84602, USA Abstract As the Internet grows, it is becoming less feasible for customers and merchants to manually visit each web site, analyze the information there, and make sound business decisions regarding the trading of goods or services. To cope with this evolution, software agents can be designed that are capable of automating the more routine, tedious, and time-consuming tasks involved in current trading processes. At a higher level agents may also be able to negotiate and make autonomous decisions and commitments on behalf of their owners. This paper describes an agent implementation using the situation calculus, which offers a possibly unifying paradigm for dynamic agents. Interesting applications are currently being developed. Our contribution is a situation calculus agent system developed for e-business. Ongoing work is focused on implementing this system in an open marketplace environment. q 2003 Elsevier Science Ltd. All rights reserved. Keywords: Internet; E-business; Software agent 1. Introduction The Internet can be viewed as a large, distributed information resource, with connecting systems that are designed and implemented by many different organizations with various goals and agendas. Because of the breadth and flexibility of the Internet, e-business applications have grown rapidly in recent years, and this trend is expected to continue. Forrester Research and the Gartner Group estimate that e-business transactions will exceed $7 trillion by 2004. Thirty-seven percent of these transactions are estimated to be associated with electronic marketplaces, which provide resources to facilitate the interaction and exchange of commerce transactions among buyers, sellers, and other trading partners (Rahman & Bignall, 2001). Today’s e-business is increasingly characterized by the ability to create value through the gathering, synthesizing, and distribution of information. E-businesses progressively compete in real time rather than cycle time and operate through a constantly responsive dialogue with their customers and markets. Relationships among customers and markets are managed by technology through electronic channels. These technology-mediated channels are charac- terized by ongoing operations, which are subjected to measurement and tracking in new and innovative ways (Evans & Wurster, 2000). Such applications are being facilitated by e-business architectures that provide opportunity for distributed, service-based applications, where service providers may not know about each other at design time. Services include such components as credit card charging, final fulfillment of orders, or publication of product information. These services are often distributed and run on different computers in geographically dispersed locations. Generally, services are created and offered by separate businesses, which specialize in a particular type of service. Such entities can often achieve economies of scale and provide levels of service beyond what a firm may be able to provide internally. This environment offers new opportunities for online business advantage. Yet as the Internet grows in size and sophistication, it is becoming less feasible for customers and merchants to manually visit each web site, analyze the information there, and make sound business decisions regarding the trading of goods or services. It is inevitable that buyers may miss finding the best product or services in the vast milieu of information. In this new world, software agent technologies offer a robust paradigm for trading on the Internet and foreshadow a dramatically changed approach to conduct analysis, market research, and similar e-business functions. At one level, software agents can be designed to be capable of automating the more routine, tedious, and time-consum- ing tasks involved in current trading processes. At a higher level, agents can be designed to negotiate and make 0957-4174/03/$ - see front matter q 2003 Elsevier Science Ltd. All rights reserved. doi:10.1016/S0957-4174(02)00188-4 Expert Systems with Applications 24 (2003) 391–397 www.elsevier.com/locate/eswa * Corresponding author. Tel.: þ1-801-422-2308; fax: þ1-801-422-5933. E-mail address: james_hansen@byu.edu (J.V. Hansen). http://www.elsevier.com/locate/eswa autonomous decisions and commitments on behalf of their owners (Sandholm, 2001). 2. Agents and e-business Implementation of a conventional strongly coupled system for electronic data interchange (EDI) between trading partners requires that a business know the providers and consumers of the various goods and services being traded so that transactions can be exercised between appropriate parties. In today’s more open markets, deter- mining this information in advance is extremely difficult or even impossible. Consider an electronic system to support open trading, where orders are made available to any qualified bidder. Requiring the system designer to determine the sender and recipient of each transaction in advance is not realistic. On the contrary, open trading applications are only loosely coupled, since not all of the necessary information is available when the system is designed. Agents, however, can naturally support such appli- cations. Instead of specifying the individual entities to be interconnected and their interfaces with one another, an agent-based system need identify only the classes of entities in the system and their impact on the environment. Because the agent is designed to interact with the environment rather than with specific other agents, it can interact appropriately with any other agent that modifies the environment within the range of variation with which other agents are designed to handle (Wooldridge, 2002). Following Lesperance, Levesque, & Reiter (1999), we take an agent to be any active entity whose behavior is usefully described through mental notions such as knowledge, goals, abilities, commitments, and the like. At a low level, agents are currently being used to help suppliers reduce inventory carrying costs and marketing costs by offering special discounts unavailable to the general public. At a higher level agents are being programmed with rules or learning capability that have led to their being termed ‘intelligent’ agents. Intelligent agents are being used to help businesses determine which customers are most likely to be interested in purchasing the discounted items, to compare product and service information, to help users narrow the number of sites that contain desired information, and negotiating business transactions: prices, delivery schedules, quality, warrantees, etc. Firms are also using intelligent agent technology to continuously conduct environment scans. For example, an attorney that specializes in patent law can use pull technology to continuously monitor multiple sites and signal any developments such as new patent-related court cases or court decisions. (Greenstein & Vasarhelyi, 2002) Central to successful agent implementation is the issue of how to program agents to perform their functions. In one sense, agent programming can be viewed as a generalization of object-oriented programming; yet the notion of an agent is more complex than that of an object. Hence, it is important that tools for modeling and designing agents be based on sound theoretical foundations. Many of today’s agent paradigms are ad hoc or otherwise difficult to evaluate (Reiter, 2001). Although several powerful formalisms exist, finding the right formalism is a nontrivial challenge, as it must provide a level of expressiveness that serves the practical problems at hand in a tractable way (Lesperance, Levesque, Lin, & Scherl, 2000). Despite the range of choices, there is general agreement that formal methods do help in the long run in helping to develop a clearer understanding of problems and solutions (Kraus, 2001). This paper describes an approach and application grounded in theoretical and practical developments arising from recent research in cognitive robots and software agents. In particular the development of extensions to situation calculus opens new possibilities for agent programming (Boutelier, Reiter, Soutchanski, & Thrun, 2000). Situation calculus allows any defined set of complex action expressions to be viewed as the programming language for agents. That language can then be used to model the behavior of a set of agents and to actually implement them (Reiter, 2001). Relevant research in artificial intelligence aims at explaining and modeling intelligent behavior in terms of computational processes. The classical approach assumes intelligent behavior to be a result of correct reasoning on correct representations. In turn, reasoning is understood by means of formal logic. In research areas such as cognitive robotics and software agents, this approach to AI is applied to a crucial aspect of intelligent behavior, that of acting in a dynamic world. Situation calculus addresses this issue formally and directly (Lesperance et al., 2000). 3. High-level agent programming using situation calculus The construction of autonomous agents is paramount in artificial intelligence, with considerable research devoted to methods that will ease the burden of designing controllers for such agents. There are two principal ways in which the complexity of devising controllers can be constructed. The first is to provide languages, with which a programmer can specify a control program with relative ease, using high-level actions as primitives, and expressing the necessary operations in a natural way. The second is to simply specify goals and provide the agent with the ability to plan appropriate courses of action that achieve those goals. In this way the need for explicit programming is avoided (Wooldridge, 2002). The development of this type of computational agent has been the subject of considerable recent work. The core C. C. Albrecht et al. / Expert Systems with Applications 24 (2003) 391–397392 https://isiarticles.com/article/3668 
work_ghay4mkd3nczpcfwazq6dhtatu	----	A hybrid ant colony optimization for continuous domains A hybrid ant colony optimization for continuous domains Jing Xiao a,⇑, LiangPing Li b a School of Computer Science, South China Normal University, No. 55 Zhongshan West Road, Guangzhou 510631, China b Department of Computer Science, Sun Yat-sen University, No. 132 WaiHuan East Rd., Guangzhou Higher Education Mega Center, Guangzhou 510006, China a r t i c l e i n f o Keywords: Continuous optimization Ant colony optimization Continuous population-based incremental learning Differential evolution a b s t r a c t Research on optimization in continuous domains gains much of focus in swarm computation recently. A hybrid ant colony optimization approach which combines with the continuous population-based incre- mental learning and the differential evolution for continuous domains is proposed in this paper. It utilizes the ant population distribution and combines the continuous population-based incremental learning to dynamically generate the Gaussian probability density functions during evolution. To alleviate the less diversity problem in traditional population-based ant colony algorithms, differential evolution is employed to calculate Gaussian mean values for the next generation in the proposed method. Experimen- tal results on a large set of test functions show that the new approach is promising and performs better than most of the state-of-the-art ACO algorithms do in continuous domains. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Inspired by the ants’ foraging behavior, the first ant colony algo- rithm was proposed in (Dorigo, 1992). The core idea of an ant sys- tem is based on the fact that each ant deposits pheromone on the foraging path. In the early research, ant colony algorithms have been successfully applied for solving combinatorial optimization problems, including the traveling salesman problem (TSP) (Dorigo & Gambardella, 1997), the quadratic assignment problem (QAP) (Maniezzo, Colorni, & Dorigo, 1999) and the job shop scheduling problem (JSP) (Colorni, Dorigo, Maniezzo, & Trubian, 1994). The extension of the original ant colony algorithm to continuous domain can be accomplished by either discretizing the continuous domain into several regions or shifting from using a discrete probability distribution to using a continuous one such as a Gaussian probability density function (pdf). The first extension of ant colony algorithms to continuous domain is CACO (continu- ous ant colony optimization) (Bilchev & Parmee, 1995). CACO discretized the continuous neighborhood with nest structure. It initialized a nest on a given point of the search space and generates random vectors. The direction vector chosen was updated by the ants with better fitness values. In API (ant pachycondyla apicalis) (Monmarché, Venturini, & Slimane, 2000) each ant searched independently for a solution and the ants’ nest is periodically moved. CIAC (continuous interacting ant colony) (Dréo & Siarry, 2002) used some attractive points on the search space and direct communication between ants to improve the exploration. COAC (continuous orthogonal ant colony) (Hu, Zhang, & Li, 2008) also discretized the continuous search space and utilized the orthogo- nal design method to search the continuous domain completely. Another ensemble of ant colony optimization algorithms for continuous domain is mainly based on Gaussian probability den- sity function sampling. CACS (continuous ant colony system) (Pourtakdoust & Nobahari, 2004) employed a Gaussian probabil- ity density function whose mean and variance are adjusted dur- ing evolution to sample the continuous search space. MACACO (multivariate ant colony algorithm for continuous optimization) (Franca, Coelho, Von Zuben, & de Faissol Attux RR, 2008) opti- mized the search space with multivariate Gaussian pdf which is created by the information contained in the covariance matrix of the ant population. Speak of population, there are some work which fully utilize the information in previous ant population to generate the next one. PB-ACO (population-based ACO) (Guntsch & Middendorf, 2002) kept track of a certain number of best solu- tions found so far and used the archive to update the phero- mone. PB-ACO was applied to the TSP and QAP problems while ACOR (ant colony optimization in Rn) (Socha & Dorigo, 2008) ex- tended the population-based ACO with Gaussian pdf as phero- mone update. ACOR evolved several Gaussian pdfs in parallel and adopted the rank-based selection mechanism. ACOR also al- lowed exploiting the correlation between variables by coordinate rotation which is time-consuming. Inspired by PB-ACO and ACOR, our motivation for introducing a population based scheme is the dynamic generation of probability density functions in continuous domains. A hybrid ant colony opti- mization (HACO) incorporating with continuous population-based incremental learning (PBILc) and differential evolution (DE) (Sebag 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.02.151 ⇑ Corresponding author. Tel.: +86 20 85211352 401 E-mail address: xiaojing@scnu.edu.cn (J. Xiao). Expert Systems with Applications 38 (2011) 11072–11077 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.02.151 mailto:xiaojing@scnu.edu.cn http://dx.doi.org/10.1016/j.eswa.2011.02.151 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa & Ducoulombier, 1998) is proposed in this paper. HACO employs the multi-Gaussian pdfs sampling in parallel and learns the next generation’s mean and variance values dynamically by PBILc. The rank-based selection method in ACOR may suffer local optima when the solutions in the archive are very close to each other. To alleviate this, HACO involves a differential evolution operator with linear combination of the one best and two random individuals (Montes and Velazquez-Reyes, 2006) in the current population. The combination of differential evolution and ant colony has been applied to feature selection and power systems (Chiou, Chang, & Su, 2004; Khushaba, Al-Ai, Alsukker, & Al-Jumaily, 2008). The rest of this paper is organized as follows. Section 2 de- scribes HACO in detail. The pheromone representation and Gauss- ian pdf generation are elaborated in Section 2. Experimental results and analysis are presented in Section 3. Section 4 discusses how two different DE operators affect HACO. We conclude this paper and outline the future works in Section 5. 2. HACO: A hybrid ant colony optimization for continuous domains In this section, we introduce the HACO in detail. First, we describe the pheromone representation in HACO and the overall framework of the algorithm is given then. The solution construction and Gauss- ian pdf generation by two methods are introduced finally. 2.1. Pheromone representation In traditional ACO algorithms, ants use pheromone to commu- nicate with each other. In combinatorial optimization problem, ACO algorithms use a table to store pheromone information. At each construction step, an ant uses the table to form a discrete probability distribution. While in continuous optimization, differ- ent representation should be adopted. In this paper, we use the same idea to that proposed by Marco Dorigo in ACOR (Socha & Dorigo, 2008). HACO stores a number of solutions in a solution archive. Each solution contains the values of its n variables. The solution archive stores the solutions and also their values of the objective functions. As the pheromone information is not explicitly stored as in tra- ditional ACO, pheromone cannot be updated directly. In HACO, pheromone update is accomplished by updating the solution ar- chive. When a better solution is found, it will be added to the solu- tion archive with a worst solution being removed. With the best solutions ever found keeping in the solution archive, the ants can explore the domain effectively. 2.2. HACO framework The framework of HACO algorithm is described in Algorithm 1 and the notations of the algorithm are presented in Table 1. Algorithm 1. HACO Algorithm. set iteration t = 0; initialization_of_k_solutions(); while t < maxNumIterations do t = t + 1; for j = 1 to m/2 do for each dimension i do sampling the value of dimension i by the Gaussian kernel pdf by ranking; for j = m/2 + 1 to m do for each dimension i do sampling the value of dimension i by the Gaussian pdf by PBILc; for each ant do update_solution_archive(); generate_Gaussian_functions(); For an n-dimension problem, an ant constructs a solution in n steps. At each step i, an ant samples a value for variable xi. HACO keeps k solutions in the archive for pheromone update calculation. The ith variable of jth solution is denoted by sji . intializa- tion_of_k_solutions() initially generate k solutions from the scratch by uniform sampling. update_solution_archive() means that the pheromone update is accomplished by adding the set of newly generated better solutions according to the ranking or PBILc meth- od, then the same number of worst solutions in the archive is re- moved accordingly. In HACO generate_Gaussian_functions() performs two kinds of Gaussian generation. One of the ant groups applies the rank-based selection mechanism as in ACOR and the other applies the PBILc method which will be introduced later on. Then the two ant groups use the corresponding Gaussian func- tions to sample their values for each dimension. 2.3. Generate Gaussian functions by rank-based scheme For each dimension i, HACO generates a probability density function called Gaussian kernel pdf. First, each solution constructs an individual Gaussian pdf Ni(li, ri) for each dimension. Then the Gaussian kernel pdf is making up by the k separate Gaussian func- tions with different weights. These weights are associated with the fitness value of the objective function. Better solution will have higher weight. The weight wl of the solution s l is calculated by the following equation (Socha & Dorigo, 2008): wl ¼ 1 qk ffiffiffiffiffiffiffi 2p p e �ðl�1Þ 2 2q2 k2 ; ð1Þ which defines the weight to be a value of the Gaussian function with argument l, mean 1.0, and standard deviation qk, where q is a parameter to balance between diversification and intensification. When q is small, the best-ranked solutions are strongly preferred, and when it is large, the probability becomes uniform (Socha & Dor- igo, 2008). Sampling the Gaussian kernel pdf is done in two phases. First, choose one of the individual Gaussian function that compose the Gaussian kernel with probability pl given by Eq. (2). Then sample the chosen Gaussian function. pl ¼ wlPk r¼1 wr : ð2Þ We consider the chosen Gaussian function by (2) with li a standard solution for other ant solutions to explore. To establish the value of the standard deviation ri at step i, we calculate the average distance from li to all solutions in the archive, and multiply it by the param- eter n: Table 1 Notations in HACO. Symbol Description X = {x1, x2, . . . , xn} A set of n continuous variables in a certain problem k Size of the solution archive m Number of ants used in each iteration n Speed of convergence q Locality of the search process n Dimension of the problem a Learning rate N(l, r) Gaussian function with mean l and standard deviation r s Solution in the solution archive F Coefficient in differential evolution J. Xiao, L. Li / Expert Systems with Applications 38 (2011) 11072–11077 11073 https://isiarticles.com/article/7724 
work_gi7vf2tj2rcrfmlo3g4ws6qj5q	----	לימודים אקדמיים | הבינתחומי הרצליה ייתכן שאתה מנסה לגשת לאתר זה מדפדפן מאובטח בשרת. הפוך קבצי Script לזמינים וטען מחדש דף זה. הפעל מצב נגיש יותר בטל מצב נגיש יותר דלג על פקודות של רצועת הכלים דלג לתוכן ראשי בטל הנפשות הפעל הנפשות [{"Title":"עברית","eldUrl":{"Description":" He","Url":"https://www.idc.ac.il/he"},"eldPageNotFoundUrl":{"Description":" /he/Pages/pagenotfounderror.aspx","Url":"https://www.idc.ac.il/he/Pages/pagenotfounderror.aspx"},"eldFooterListFullUrl":{"Description":" /he/Lists/Footer/AllItems.aspx","Url":"https://www.idc.ac.il/he/Lists/Footer/AllItems.aspx"},"eldMapDirectsByCarLink":null,"eldMapDirectsPublicTransport":{"Description":" דרכי הגעה בתחבורה ציבורית","Url":"https://www.idc.ac.il/he/Pages/directions.aspx"},"eldMapDirectsNavDestination":{"Description":" נווטו אלינו ב-Waze","Url":"http://bit.ly/get2IDC"},"eldBrand":"\u003cimg alt=\"\" src=\"/he/PublishingImages/heb70.png\" style=\"BORDER: 0px solid; \"\u003e","eldBrandAffix":"\u003cimg alt=\"\" src=\"/he/PublishingImages/heb36.png\" style=\"BORDER: 0px solid; \"\u003e","eldEventsLobbyPageUrl":"https:// ","eldShareWithText":"שתפו ב:","eldAddToCalText":"סמנו ביומן","eldWhatsUpLobbyPageUrl":"/he/whatsup/Pages/main.aspx","eldSearchResultsPageUrl":"/he/Pages/search.aspx","eldDefaultImage":"http://"},{"Title":"English","eldUrl":{"Description":" En","Url":"https://www.idc.ac.il/en"},"eldPageNotFoundUrl":{"Description":" /en/Pages/pagenotfounderror.aspx","Url":"https://www.idc.ac.il/en/Pages/pagenotfounderror.aspx"},"eldFooterListFullUrl":{"Description":" /en/Lists/Footer","Url":"https://www.idc.ac.il/en/Lists/Footer"},"eldMapDirectsByCarLink":null,"eldMapDirectsPublicTransport":{"Description":" Arriving by public transport","Url":"https://www.idc.ac.il/en/Pages/directions.aspx"},"eldMapDirectsNavDestination":{"Description":" Navigate with Waze","Url":"http://bit.ly/get2IDC"},"eldBrand":"\u003cimg alt=\"\" src=\"/he/PublishingImages/eng70.png\" style=\"BORDER: 0px solid; \"\u003e","eldBrandAffix":"\u003cimg alt=\"\" src=\"/he/PublishingImages/eng36.png\" style=\"BORDER: 0px solid; \"\u003e","eldEventsLobbyPageUrl":"http://","eldShareWithText":"Share on","eldAddToCalText":"Add to Calendar","eldWhatsUpLobbyPageUrl":"/en/whatsup/Pages/main.aspx","eldSearchResultsPageUrl":"/en/Pages/search.aspx","eldDefaultImage":"http://"},{"Title":"chinese","eldUrl":{"Description":" 中文","Url":"https://www.idc.ac.il/chinese"},"eldPageNotFoundUrl":{"Description":" /en/Pages/pagenotfounderror.aspx","Url":"https://www.idc.ac.il/en/Pages/pagenotfounderror.aspx"},"eldFooterListFullUrl":{"Description":" /en/Lists/Footer","Url":"https://www.idc.ac.il/en/Lists/Footer"},"eldMapDirectsByCarLink":null,"eldMapDirectsPublicTransport":null,"eldMapDirectsNavDestination":{"Description":" Navigate by Waze","Url":"http://bit.ly/get2IDC"},"eldBrand":"\u003cimg alt=\"\" src=\"/SiteCollectionImages/brand_eng.png\" style=\"BORDER: 0px solid; \"\u003e","eldBrandAffix":"\u003cimg alt=\"\" src=\"/SiteCollectionImages/brandAffix_en.png\" style=\"BORDER: 0px solid; \"\u003e","eldEventsLobbyPageUrl":"https:// ","eldShareWithText":"https:// ","eldAddToCalText":"https:// ","eldWhatsUpLobbyPageUrl":"https:// ","eldSearchResultsPageUrl":"https:// ","eldDefaultImage":"https://"}] דלג לתפריט ראשי דלג לתוכן העמוד דלג למפת האתר מה תרצו לחפש? - בכדי להתחיל את החיפוש יש להקיש אנטר פתיחה/סגירה תפריט כפתור הניגודיות בחר שפה * * * המרכז הבינתחומי הרצליה עברית לימודים אקדמיים - הבינתחומי הרצליה   סגור טריילר סגור טריילר צפה בטריילר לעצירת הבאנר לניגון הבאנר ｜ ｜ סגור מודאל מפה ודרכי הגעה 
work_gif4gadfzngo3d44zddi5tyxym	----	Hydroelectric power plant management relying on neural networks and expert system integration Hydroelectric power plant management relying on neural networks and expert system integration J.M. Molina*, P. Isasi, A. Berlanga, A. Sanchis Departamento de InformaÂtica, Universidad Carlos III de Madrid. Avda. de la Universidad 30, 28911 LeganeÂs, Madrid, Spain Received 1 November 1998; accepted 1 January 2000 Abstract The use of Neural Networks (NN) is a novel approach that can help in taking decisions when integrated in a more general system, in particular with expert systems. In this paper, an architecture for the management of hydroelectric power plants is introduced. This relies on monitoring a large number of signals, representing the technical parameters of the real plant. The general architecture is composed of an Expert System and two NN modules: Acoustic Prediction (NNAP) and Predictive Maintenance (NNPM). The NNAP is based on Kohonen Learning Vector Quantization (LVQ) Networks in order to distinguish the sounds emitted by electricity-generating machine groups. The NNPM uses an ART-MAP to identify di�erent situations from the plant state variables, in order to prevent future malfunctions. In addition, a special process to generate a complete training set has been designed for the ART-MAP module. This process has been developed to deal with the absence of data about abnormal plant situations, and is based on neural nets trained with the backpropagation algorithm. 7 2000 Elsevier Science Ltd. All rights reserved. Keywords: Predictive maintenance; Neural networks; ART; LVQ; Power plants; Expert systems 1. Introduction The management systems of power plants continu- ously monitor di�erent features of several subsystems: energy distribution (located outside the power plant), energy generators, motors, turbines, etc. The features of each subsystem are related to a set of variables that de®ne the current situation in the subsystem. The evaluation of these variables gives some guidelines to human operators to detect abnormal situations in power plants. However, only a small set of variables can be observed and analyzed, to give useful infor- mation to the human observer. On the other hand, simple automatic monitoring systems are, in general, able to analyze all the input values and generate alarms. Simple monitoring systems warn when the numerical value of a variable is outside the range set by human experts. More sophisticated systems include the possibility of forecasting future events, in order to anticipate malfunction situations. The main advantages to be gained by using predictive modules are the optimization of productivity, operat- ing safety, and the protection of equipment against damage caused by malfunctions. These predictive sys- tems are especially useful in distributed control centers, since the general philosophy is to model the plant and predict future situations instead of merely taking pre- cautionary measures (Evans et al., 1994). The development of more complex management sys- tems, for predictive automatic surveillance, needs more advanced techniques. Some Arti®cial Intelligence (AI) techniques have been suggested in the literature for such systems (Raghavan and Simon, 1993; Loskiewicz- Buczak et al., 1993). These techniques give additional information for the identi®cation of potential future failures and their causes. They analyze the relations Engineering Applications of Arti®cial Intelligence 13 (2000) 357±369 0952-1976/00/$ - see front matter 7 2000 Elsevier Science Ltd. All rights reserved. PII: S0952-1976(00)00009-9 www.elsevier.com/locate/engappai * Corresponding author. Tel.: +34-91-624-9116; fax: +34-91-624- 9129. E-mail address: molina@ia.uc3m.es (J.M. Molina). between items of data coming from vital components of a power plant, thereby making it possible to per- form e�ective preventive maintenance. Knowledge-based systems (KBS) are able to provide valid ranges for variables, and to predict future abnor- mal situations from these values, using the expert knowledge of plant operators (Kim et al., 1994). In order to reduce the dependence of a ®nal KBS on the knowledge-acquisition phase, some machine learning techniques (Yoshikawa et al., 1995) have also been applied. In particular, the use of neural networks (NN) is a good approach to predictive control systems because of the highly non-linear relationships between the data and the abnormal situations (Isasi et al., 1996; Yon- ghong and van Cauwenberghe, 1995). While an expert system tries to mimic the response of a human operator, analyzing the same variables as the human, the neural networks overcome these limits and try to analyze non- linear relations between many di�erent signals. The man±machine communication is simple with an expert system, because the system could explain fail- ures and also indicate why some decisions have been taken. This communication is hard with an NN, because the output depends non-linearly on di�erent signals. So a complete integration of these two tech- niques could improve the global management system. Both the NN and the expert system information has been taken into account in the ®nal decision. The for- mer considers all the signals measured in the power plant, and makes a complex decision that considers the non-linear relations. The latter analyzes only a subset of the signals but can generate an explanation about the predicted events to support the human decision- making. In this work, a hybrid predictive system named MAPAIS (an abbreviation of the Spanish for an advanced system for predictive maintenance incorpor- ating audio and video) has been developed to forecast alarms in a hydroelectric power plant, and is included in a complete monitoring system named HYPERVI- SION. All the information that is processed to moni- tor the plant could be divided in three clusters: internal state variables, audio signals and video signals. The internal state variables are related to physical measures (temperature, pressure, water level, oil level, revolutions, etc.) taken in several devices (valves, pipes, rotors, switches, etc.). The audio signals are recorded in turbines in order to capture the sound of the rotors and the water in di�erent power situations, from normal to critical regimes. The video cameras record images from external power-plant components. All the recorded information is shown to human oper- ator through the HYPERVISION multimedia inter- face, and is used to evaluate the performance of MAPAIS. The predictive system (MAPAIS) relies on advanced arti®cial intelligence techniques, audio and video signal processing, expert knowledge systems and NN model- ing. Within the predictive modules of the system, the main goal of the NN maintenance module is to take advantage of the data recorded at every point in the hydroelectric plant. In this way, the predictive modules are acting as predictors of future incidents in the sense that they learn to detect situations of potential con- ¯ict. Inside the predictive system, two di�erent NN have been developed: the ®rst receives the values of all the internal state variables (NN Predictive Maintenance- NNPM), and the other takes the spectral values of di�erent acoustic signal measures in the turbines (NN Acoustic Prediction-NNAP). The NNPM developed here is an ART-MAP that predicts the future situation in the power plant on the basis of the internal state variables. The NNAP is an LVQ NN that classi®es audio spectrums into normal or abnormal acoustic noise. The use of NN using internal state variables when dealing with real situations (not simulations) adds an important problem: the acquisition of a complete set of data corresponding to abnormal situations is not possible, due to security and productivity restrictions in the plant. In this work, a special procedure for generating complete training sets has been developed. The non-linear relationships between the internal state variables have been obtained in a Multilayer Perceptron (MLP), trained using a backpropagation algorithm. The system is currently working in the hydroelectric plant named Villalcampo I (IBERDROLA), which is located in the Zamora province of Northwest Spain. This plant has also been used for implementing and further testing of the MAPAIS project. 2. System architecture In Fig. 1, a general view of the graphical supervisory system (HYPERVISION) is presented. HYPERVI- SION shows a schematic image of the electricity gener- ating machine group, constituted by three independent generators named group 1, group 2 and group 3, re- spectively. The plant operator is able to obtain infor- mation about any system parameter by descending a hierarchical hyperlink structure. The system could be decomposed into three levels (Fig. 2). . Acquisition: this level records values of the input variables (digital/analog values). . Prediction: the main part of the work, composed of several modules. J.M. Molina et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 357±369358 https://isiarticles.com/article/21704 
work_gihmwdoarrcofj5ql3ftyiqvze	----	MIJ2K Optimization using evolutionary multiobjective optimization algorithms This document is published in: Expert Systems with Applications, (2011), 38 (9), 10999–11010. DOI:http://dx.doi.org/10.1016/j.eswa.2011.02.143 © 2011 Elsevier Ltd. http://dx.doi.org/10.1016/j.eswa.2011.02.143 Ab tes pro com out Xip wit 1 Ke MIJ2K Optimization using evolutionary multiobjective optimization algorithms Alvaro Luis Bustamante ⇑, José M. Molina López, Miguel A. Patricio Univ. Carlos III de Madrid, Avda. Univ. Carlos III, 22, 28270 Colmenarejo, Madrid, Spain ⇑ Corresponding author. Tel.: +34 918561338. E-mail addresses: aluis@inf.uc3m.es (A.L. Bustamante), molina@ia.uc3m.es (J.M. Molina López), mpatrici@inf.uc3m.es (M.A. Patricio). jective a high s. We tr ession r ssed wi f the op ss the s stract: This paper deals with the multiob ted and highly accurate algorithm with blem including two competing objective peting objectives are quality and compr lined in this paper. Video will be compre h.org Foundation repository. The result o h different encoder parameters and discu Multi-objective, Optimization, Vi ferred to as bit rate). The best we could exp video quality with the smallest file size, bu conflict since better qualities inherently im ywords: deo, Encoder definition of video compression and its optimization. The opti-mization will be done using NSGA-II, a well- conver-gence speed developed for solving multiobjective problems. Video compression is defined as a y to find a set of optimal, so-called Pareto-optimal solutions, instead of a single optimal solution. The two atio maximization. The optimization will be achieved using a new patent pending codec, called MIJ2K, also th the MIJ2K codec applied to some classical vid-eos used for performance measurement, selected from the timization will be a set of near-optimal encoder parameters. We also present the convergence of NSGA-II uitability of MOEAs as opposed to classical search-based techniques in this field. . Introduction Nowadays, digital video is widely used for many purposes, ranging from mere entertainment, such as TV, video conferencing or video on demand, to more professional environments, such as remote video surveillance. This wide range of applications is possi- ble thanks to recent advances in digital video technology, like broadband connections to the Internet, computing capacity and the digital storage space of new devices. However, the most important part of a digital video system is the codec. The codec enables digital video compression (with the encoder) and/or decompression (with the decoder). This reduces the data necessary for representation. This is necessary because uncompressed digital video still exceeds common network band- widths for transmission and storage spaces for digital archiving. Compression usually employs lossy data compression. Lossy data compression reduces a file by permanently eliminating cer- tain information, especially redundant information. When the file is uncompressed, only a part of the original information is still there (although this may go unnoticed to the user, especially in vi- deo and sound compressions). It inherently implies a reduction of the quality in exchange for a reduction in the final amount of information. Systems like these introduce a complex trade-off between the quality and quantity of data needed to represent the video (also re- ect is to get the highest t these objectives are in ply a greater bit rate. Thus, when compressing digital video, the user usually estab- lishes encoder objectives or constraints, i.e. the maximum bit rate (normally used when there are bandwidth limits or storage space constraints (Jiang, 2006; Wang & Leou, 2003)), or video quality (rated on a 0 to 100 scale by some quality metric, etc., to ensure some quality of service (Ng, Leung, & Hui, 2005; Zhang, Zhu, & Ya-W, 2005)). The encoder should compress the digital video according to these objectives, but it is not usually easy to get a direct correlation between these high-level objectives and low-level encoder param- eters, because video encoder bit rate and video quality depend on several coding parameters, such as quantization parameter (Czuni, Csaszar, & Licsar, 2006), coding mode (kuang Chen, Vetro, & Sun, 1997), macroblock sizes (Tu, Yang, Shen, & Sun, 2003), or motion compensation algorithms (Hang, Chou, & Cheng, 1997). Each of these parameters has its own attributes or thresholds and may have different effects. We have developed MIJ2K, a new video co- dec, currently patent pending, and we need to optimize some parameters before its release. In this paper, we present an algorithm to dynamically optimize the two conflicting objectives discussed above (video quality/bit rate) as well as translate the objectives to the video encoder parameters. Such algorithms are known as multiobjective optimi- zation (MO) algorithms. MO optimization problems (Steuer, 1986; Sawaragi, Nakayama, & Tanino, 1985) are very common in many complex engineering situations, and can be found in many fields: product and process design, finance, aircraft design, the oil and gas industry, automobile design, or wherever optimal 1 decisions need to be taken in the presence of more than one, gen- erally conflicting objective, preventing the simultaneous optimiza- tion of each objective. Basically there are two major approaches for solving such MO problems. The first is to combine the individual objective functions into a single composite function and optimize this function only. This could be done with techniques such as the weighted sum method (Koski, 1988), utility theory (Thurston et al., 2006), etc. The second major approach is to determine an entire Pareto optimal solution set or a representative subset, i.e. a series of solu- tions that are non-dominated with respect each other. Pareto opti- mal solution sets are often preferred to single solutions because they can be practical when considering real-life problems and give the decision maker (DM) the option of evaluating the trade-offs between different solutions. The use of Multiobjective Evolutionary Algorithms (MOEAs) to output the Pareto front between these two objective functions (quality and bit rate) will satisfy all DM requirements, since the Pareto solution set will offer a wide range of solutions ranging from low quality, low bit rate to high quality, high bit rate, all of which are optimal in some sense. DMs could use the solution set for many applications with different constraints, such as real-time stream- ing, controlling the bit rate according to available bandwidth; high definition video, establishing high qualities; video storage with space constraints, etc. We will use the NSGA-II algorithm to obtain the Pareto front (Deb, Pratap, Agarwal, & Meyarivan, 2002). NSGA-II should maxi- mize two objective functions: the quality measured by the peak- to-signal noise ratio (PSNR), and the compression ratio (CR), which is a ratio between compressed video and original video sizes. To evaluate each set of parameters, we will test the video compressor with classical sequences used to evaluate encoders performance, like ‘Hall Monitor’ or ‘Akiyo’, selected from the well-known Xi- ph.org Foundation repository (Lora, 1994-2008), which is suitable for evaluating video compression codecs. Achieving compression and decompression to meet the fitness function for the quality and bit rate objectives is a tedious process due to the amount of data that has to be managed. Classical search- based techniques will take a long time, whereas MOEAs are well suited for such optimization problems. The paper is organized as follows. Section 2 gives a general description of how the MIJ2K works. Section 3 defines the multiob- jective problem. Section 4 specifies the conflicting objective func- tions and the decision variables for optimization. Finally, Section 5 presents the tests performed with the NSGA-II algorithm. 2. MIJ2K codec This section outlines the MIJ2K codec to give an understanding of how it works and the internal parameters needed to use evolu- tionary algorithms for optimization purposes. The basics of video compression rely on two major methods: intra-frame and inter- frame compression techniques (Shi & Sun, 2000). In intra-frame methods each video frame is an independent entity encoded with a still image compressor, usually JPEG (Pennebaker & Mitchell, 1993; Wallace, 1991) or JPEG2000 (Christopoulos, Skodras, & Ebrahimi, 2000; Rabbani & Joshi, 2002). This technique is extre-mely useful in real-time environments, like video surveillance, due to its low computing complexity. On the other hand, inter-frame techniques employ advanced coding methods, using algo-rithms like motion compensation. These techniques are more bandwidth-efficient than intra-frame methods at the expense of more complexity. In this paper, we optimize the internal parameters of the MIJ2K codec, which is based on the JPEG2000 image compression stan- dard, and a mixture of intra and inter-frame techniques. JPEG2000 is a wavelet-based (Adams & Ward, 2001) image compression stan- dard created by the Joint Photographic Experts Group committee in the year 2000, with the aim of superseding their original discrete cosine transform-based JPEG standard (dating from 1992). JPEG2000 offers a modest increase in compression performance compared with JPEG, but its main benefit is significant code- stream flexibility. The code stream obtained after compression of an image with JPEG2000 is scalable, meaning that it can be de- coded in a number of ways. For instance, by truncating the code stream at any point, we can get a representation of the image at a lower resolution or signal-to-noise ratio. By ordering the code stream in various ways, applications can achieve significant perfor- mance increases (Adams, 2001). Apart from the above features, the main benefit of using JPEG2000 for video streaming is that, unlike other video compres- sors including MPEG-4, compression could be done in real-time (Luis & Patricio, 2000) because it is an intra-frame codec. Thanks to this feature, JPEG2000 can be used in events that require real- time transmission, like video surveillance. As far as our proposal is concerned, however, the main advan- tage is that the image can, optionally, be partitioned into smaller independent non-overlapped rectangular blocks called tiles (ISO/ IEC, 2000). We will exploit this exceptional feature, provided by this compressor alone, to perform real-time inter-frame compres- sion using the proposed conditional tile replenishment method, which we optimize in this paper. We will employ a new block- based difference coding technique (Shi & Sun, 2000) for this task, having tiles assume the role of blocks. Tiles can be of any size, and the whole image can even be con- sidered as one single tile. Once the size has been chosen, though, all the tiles will be of the same size (except, optionally, tiles on the right and bottom borders). Dividing the image into tiles is advan- tageous in that the encoder/decoder will need less memory to en- code/decode the image. Also it can opt to encode/decode only selected tiles to achieve a partial coding/decoding of the image. It will provide full control of whatever area of the image is being compressed, decompressed, transmitted, etc. Fig. 1 shows an example of how the J2K code stream is struc- tured, and how an image is divided using tiles. The first marker present is the start of code stream (SOC). This is followed by a main header (MH), which includes the common parameters re- quired for image decoding. The tile-part header (TH) contains the necessary information for decoding each tile. It is followed by the corresponding tile-part bit-stream. Finally, the end of code-stream (EOC) marker denotes the termination of a J2K code stream. Notice that each region of the image occupies a definite region in the J2K code stream. Thanks to the header definition method (ISO/IEC, 2000), each region can also be accessed randomly. The inter-frame technique adopted in MIJ2K is a real-time spe- cific block-based difference coding, adapted to JPEG2000 streams. This technique is useful in the real-time transmission scheme be- cause it provides a low computational complexity. It also preserves the real-time latency provided by the native JPEG2000 intra-frame architecture. The MIJ2K architecture designed for this task is outlined in Fig. 2 and explained in more detail in the following sections. The general operating procedure is as follows. A common JPEG2000-based streaming system is basically di- vided into three steps. The first step is related to frame acquisition and compression. It is followed by the transmission of the resulting compressed frames. Finally, it ends with the reception and display of each frame. The acquisition and compression step in these systems is not complex. Each frame is acquired and compressed separately, and can be transmitted as soon as it is compressed. 2 Fig. 2. Overall functioning of the MIJ2K architecture, performing real-time selective tile compression and transmission. Fig. 3. ‘Akiyo’ sequence. Tiles detected as changing in frame 127. In this frame, a bandwidth saving of around 83% is achieved. Fig. 1. JPEG2000 code-stream structure. In the proposed MIJ2K streaming architecture, an extra process is inserted in the compression step. Instead of compressing each whole frame, it compresses and transmits only the areas that are different (changing tiles) from the previously transmitted frame. Compressing only changing tiles improves compression, transmis- sion, and decoding performance, since it hugely reduces the total amount of data for management. For example, Fig. 3 shows the tiles detected as changing in frame 127 of the ‘Akiyo’ sequence. They are marked with a green rectangle.1 Notice how only 17% of the tiles in this frame are de- tected as changing. When the MIJ2K method is applied to the whole ‘Akiyo’ sequence, it saves around 87% of bandwidth com- pared with a native JPEG2000 streaming system. Changing tiles are detected using a reference frame FRi that stores a representation of the last transmitted frame. Each new frame FSi to be transmitted is compared tile by tile with the refer- ence frame FRi in order to detect the tiles that differ with their counterparts. FRi is updated frame by frame with the tiles that are detected as ‘changing’ to create a real representation of what the client is viewing. The client receiving this modified stream (a JPEG2000 code stream containing just some of the tiles) will have to decode each 1 For interpretation of colour in Fig. 3, the reader is referred to the web version of this article. tile received and use it to update the displaying frame FDi . The cli- ent can easily locate each tile position since they are identified by a tile number. Knowing that all tiles (except image boundaries) are 3 of the same size, it is easy to calculate the position of the tile inside the full image. Some of the subsystems illustrated in Fig. 2 (necessary for understanding which parameters to optimize), and their design de- tails are explained in more detail in the following sections. 2.1. JPEG2000 encoder This module compresses the JPEG2000 still images. It basically takes and compresses the source frame FSi using tile partitioning. The encoder should be modified to assure that only the tiles spec- ified by the TINDEXi parameter, and not all the tiles of the frame are compressed. This is feasible since the tiles in JPEG2000 images can be compressed and decompressed separately. Ideally, then, only the changing tiles are compressed. This will save compression time, improving overall system operation. All frames will be com- pressed with the same quality, given by bppi. Fig. 4 is a diagram of this module. It shows all the required in- puts and outputs. These are described in more detail in the following: � Source frame FSi : this is the last image acquired by the digitizer board or digital camera, that is, the frame that is going to be transmitted. It should be formatted in some J2K encoder-under- standable format, for instance, a RAW 8bpp or 24bpp RGB image (depending on whether it is a gray-scale or color picture). � Quality bppi: this parameter is related to compression quantity, or target quality after each still image has been compressed. This parameter is usually expressed as the quantity of bits used to represent each pixel in the generated JPEG2000 code stream, that is, bits per pixel (bpp). � Tile size TSIZE: this parameter defines the size of the tiling per- formed in the J2K compressed image. Once the tile size has been set, all images will be compressed in separate squared regions JPEG2000 Encoder Source Frames Index of Changing Tiles: 3, 7, 15, … MAIN HEADER Tile #3 EOC SOC Tile #7 Tile #15 Parameters: Bits Per Pixel Tile Size (32x32) Resultant J2K Codestream Fig. 4. MIJ2K encoder operation. Fig. 5. Overlay areas with different tile sizes. Fro of the same size. Tile size is variable, but, theoretically, small tile sizes can achieve a better fitting to moving objects. This is illustrated in Fig. 5, where the soccer player is a moving object in the sequence, and little tiles are a better fit for the player. � Changing tiles index TINDEXi : this input is delivered by the motion measurement subsystem (this process is detailed later), as shown in Fig. 2. It indicates the index of the tiles that contain some movement, that is, the tiles that should be compressed. In this case, the JPEG2000 encoder should compress these tiles (regions) of the image only. This will improve the compression delay since the encoder does not have to compress the whole image. � J2K code stream FMIJ2Ki : this is the J2K code stream output after compression. It should contain only the changing tiles specified in the TINDEXi input. The tiles bit-stream should be between the main header and end of code-stream markers, as shown in Fig. 4. In this case, only tiles 3, 7 and 15 have been detected as changing. The FMIJ2K output will be passed directly to thei packetization subsystem, as illustrated in Fig. 2, which will transmit the frame over the network. This process is defined as in Eq. (1), where the J2K function rep- resents the encoder, and the inputs of the function are source frame FSi , index of tiles for compression T INDEX i , quality bppi and size of tiles TSIZE with the default value 32 � 32. FMIJ2Ki ¼ J2K F S i ; bppi; T INDEX i ; TSIZE � � ð1Þ Notice how the compression module introduces two variables for optimization: bppi and TSIZE. A priori we do not know what is the best value for the two variables, since we do not know how quality or tile size affect the motion compensation algorithm and the final video output. 2.2. Motion measurement This subsystem manages the motion measurement between two consecutive frames and detects the tile index of images that contain some movement. The complexity of the algorithm pro- posed for this task is low with a view to meeting the needs of real-time transmissions rather than aiming for the highly efficient motion detection performed by common inter-frame encoders. Such sophisticated techniques, as used in H.264/MPEG-4 AVC, are not usually feasible for critical real-time environments, such as video surveillance. This algorithm, as shown in Fig. 6, takes two frames, FSi and F R i (source and reference), for comparison. It also has to know the selected tile size TSIZE used in the compression step. m left to right 16 � 16, 32 � 32 and 64 � 64. 4 Motion Measurement Source Frames Tile Size (32x32) Index of Changing Tiles: 3, 7, 15, … Reference Frame Fig. 6. Motion measurement input/output. All the inputs and outputs of this system are described below: � Source frames FSi and FRi : FSi should be the same image as used in the J2K encoder in some affordable format where the pixel val- ues of the image can be directly manipulated. The other required image is the reference frame FRi . For comparison pur- poses, it should have the same format as the source frame FSi . � Tile size TSIZE: as compression is performed using the concept of tiling, motion should be measured in tile units. So, this subsys- tem must know the working tile size TSIZE. � Changing tiles index TINDEXi : as mentioned in the JPEG2000 enco- der section, this subsystem should provide the index of tiles containing some movement. The indexes could range from 0 to the total amount of tiles in any order and without any restric- tion. If no tiles with movement are detected (the current frame is almost equal to the last frame), this subsystem should some- how notify this circumstance and then send no tile for the current FSi frame (but the client side must be notified). TINDEXi ¼ Motion FSi ; FRi ; TSIZE � � ð2Þ So, this subsystem can be defined as in (2). In this case, we will specify what the ‘Motion’ function does in more detail, since this is the most important part of the adopted inter-frame technique. Fig. 7 is a detailed illustration of this function. We also describe all the tasks involved as follows: � Preprocessing: this task prepares the source frame FSi for motion measurement. The conversion performed is related to the extraction of image intensity information, that is, the Y-lumi- nance component of the YUV color space. The conversion spec- ified in (3) is applied for RGB images. This way we can work with one simple representation of the image, leading to a faster analysis of source frames. Sou Fram Tile Source Reference Tile Change Detector rce es Size (32x32) Index of Tiles Color Conversion Simple Blur Filter Preprocessing Reference UpdateReference Frame Fig. 7. Detailed MIJ2K motion measurement. Y ¼ 0:2999R þ 0:587G þ 0:114B ð3Þ Once the source frame has been correctly transformed to a gray-scale image, an optional simple blur filter is used to re- duce excessive detail and noise present in many surveillance cameras. We do not go into any more detail about the prepro- cessing algorithm and motion algorithm improvements, but we will try to determine whether the use of simple blur really does improve compression. So, the preprocessed frame output by this subsystem, Fi P, should be a gray-scale image conversion of Fi S passed through a simple blur filter, as described in (4). FPi ¼ Blur GrayScale FSi � �� � ð4Þ � Tile change detector: this module, illustrated in Fig. 7, should detect the tiles that contain movement for the purpose of selec- tive tile compression and transmission. The method used in this subsystem should be relatively simple in order to meet real- time requirements. This subsystem compares the whole preprocessed frame Fi P against the reference frame Fi R tile by tile. It detects how much movement there is in each tile to decide whether or not it should be transmitted. Fig. 8 shows how this subsystem works. It is also described in detail in the following: � Absolute difference ABSi[x]: each tile from both preprocessed and reference frames is passed through an absolute difference filter to detect absolute changes between two tiles. The compu- tational complexity of this operation is low, and it is useful for detecting objective differences between tiles. It generates a black-and-white image in which white pixels represent differ- ences, whereas black pixels signify no changes. This is described in (5), where FPi½x� and F R i½x� are tile x of frame i in the preprocessed P and reference R frames. On the other hand, ABSi[x] represents the absolute difference between the tiles. In this case, both FPi½x� and F R i½x� are two matrices of u � v containing all the tile’s pixel values. ABSi½x� ¼ FPi½x� � FRi½x� ��� ��� ð5Þ The tile image ABSi[x] output by the absolute difference process should be analyzed in order to detect how much movement there is, and thus decide whether or not it should be transmitted. Absolute Difference Tile Max. Pixel Value Mean Pixel Value Tile Thresholding Changing Tile (Yes/No) Ti le C ha ng e D et ec to r T hr ea d [ ] [ ] [ ] [ ] [ ] Fig. 8. MIJ2K tile change detector. 5 Changes should be measured somehow, and here we propose two efficient methods, which should work together. � Mean value MEANi[x]: this is the first measurement. It is the mean pixel value of the ABSi[x] tile. This process takes all the val- ues of the ABSi[x] tile, calculates the total and divides this by the number of tile pixels, as described in (6). MEANi½x� ¼ Pu�1 s¼0 Pv�1 t¼0 ABSi½x�ðs;tÞ u � v ð6Þ � Max value MAXi[x]: the second measurement is the maximum pixel value of the ABSi[x] tile, as described in (7). This is useful for detecting movement peaks in tiles, since they could be over- looked by the mean value when nearby pixels values are near to zero. Fig. 9. Pareto front solutions. 8s; t=s P 0 ^ s < u; t P 0 ^ t < v MAXi½x� ¼ ABSi½x�ðs;tÞ ():9ABSi½x�ðg;hÞ > ABSi½x�ðs;tÞ ð7Þ Operating in conjunction, MEANi[x] and MAXi[x] can detect all kinds of movements, ranging from small uniform variations (using the mean) to occasional big changes (using max thresh- old). Both metrics are easy to implement, providing a very low complexity method for detecting movement. � Thresholding: together MEANi[x] and MAXi[x] indicators tell us when the tile should be transmitted. In this way, there is said to be a big enough change in a tile to warrant transmission when both values are above some threshold. � Reference update: this subsystem is used to update the refer- ence frame FRi . The first reference frame, F R 0 , is an entirely black image, and it is updated frame by frame with the tiles that are different from the preprocessed frame FPi . Logically, the first frame F P0 will update all the tiles of the reference frame. The ref- erence frame is also updated directly from the blurred image output by the preprocessing subsystem. This will stop the refer- ence frame from having to be preprocessed each time and reduce system complexity. This subsystem introduces two new parameters for optimiza- tion, which are MEANi[x] and MAXi[x]. These variables determine motion detection algorithm sensitivity. A priori it is not feasible to set values for these parameters manually, and we will have evolutionary algorithms select the correct values. 3. Multiobjective optimization As discussed in the introduction, one way to solve MO problems is obtain the entire Pareto front. This is a common research field, where the real challenge is to obtain the Pareto front with the min- imum iteration length, because some objectives could be very costly to evaluate, and traditional search-based techniques are very time inefficient. In this way new evolutionary algorithms (EAs) (Back, Fogel, & Michalewicz, 1997) have been successfully extrapolated to multiobjective problems. EAs are well suited to MO optimization problems as they are fundamentally based on biological processes, which are inherently MO. In this case, we will apply Multiobjective Evolutionary Algo- rithms (MOEAs) (Deb et al., 2001; Coello, Lamont, & Veldhuizen, 2006) to obtain the Pareto front within the video compression field. Compression is done by reducing the redundant information, like spatial and temporal redundancies, present in video scenes to the minimum. This way we can output compressed videos with apparently no loss of information, where the final amount of infor- mation needed for representation is reduced. The problem in this case is that the loss of information (albeit redundant) amounts to a reduction in the quality or similarity of the compressed video compared with the original video. Fewer data necessarily imply lower quality, and, when trying to optimize the relation between conflicting objectives of compression and quality, we are faced with a MO problem. There is then not a single problem solution that satisfies both objectives, where we obtain the highest compression ratio with the highest quality. The solution in this case is to output the best quality to compression ratio across the full encoder operating range and let the DM select the solution that meets his or her requirements. The Multicriteria Decision Making (MCDM) (Ehrgott & Gandib- leux, 2002) literature describes two general approaches for solving user-preference mechanisms involving MO problems. In the first approach, the DM gives preferences first and the algorithm outputs a set of solutions of preferred regions of the Pareto front according to preferences (a priori methods). In the second approach, the algo- rithm provides almost all the Pareto front solutions, and the DM selects the interesting options (a posteriori methods) (Miettinen, 2001). A priori methods are preferred when there are many objective functions and the search space grows, since, in such cases, the computing resources can be focused on the preferred areas, return- ing results sooner. In this case, however, we are working with MO problems in the video compression field, where there are only two objective functions. Moreover, a wide encoder operating range must be covered, since DM constraints can easily vary depending on the use to which the encoder is put. So, a posteriori methods will be applied in this case to obtain the full Pareto front. As shown in Fig. 9 we expect to obtain the full Pareto front be- tween the conflicting objectives of quality, measured by PSNR, and CR. We can formally describe this problem as a set of objective functions to be jointly maximized (or minimized). This can be generally defined as a set of functions described in Eq. (8): maximize f n ð~vÞ where n ¼ 1; 2; . . . ; N ð8Þ In this case fnð~vÞ is an objective function, and the solution ~v is a vector of M decision variables, which is given by ~v ¼ðv 1; v 2; . . . ; v MÞ. The values of these decision variables should be between upper and lower bounds as defined by the problem. Once we have obtained solutions, we can compare them using the notion of dominance, that is, given a decision vector ~u ¼ðu1; u2; . . . ; ukÞ is said to dominate other ~v 2 ¼ðv 1; v 2; . . . ; v kÞ, if and only if 8i 2 f1; . . . ; kg; ui P v i ^ 9i0 2 f1; . . . ; kgjui0 > v i0 ð9Þ 6 In other words, the decision vector ~u is said to dominate ~v if and only if ~u is at least as good as ~v for all objectives and ~u is better than ~v for at least one objective. The Pareto optimal set is given when a decision vector is not dominated by any other vector in the search space. 4. MOEA approach to video encoder optimization In this section we discuss the decision variables that the enco- der uses, which we will optimize using NSGA-II, specifying the val- ues of the upper and lower bounds of each one, and how they affect the encoder. Moreover we detail the objective functions used to obtain the Pareto front. 4.1. Objective functions We will use two standard functions in the scope of video compression to evaluate the performance of each vector of solu- tions. The first function, f1, is the average PSNR (12), that repre- sents an objective quality metric, and the second function, f2, is the compression ratio (13), which describes a ratio of compressed to original video sizes. These functions are described below: MSEf ¼ 1 MN XM�1 i¼0 XN�1 j¼0 kIfði; jÞ� K fði; jÞk 2 ð10Þ PSNRf ¼ 20 � log10 255ffiffiffiffiffiffiffiffiffiffiffi MSEf p ! ð11Þ f1 ¼ P#Frames f¼1 PSNRf #Frames ð12Þ f2 ¼ CR ¼ UncompressedSize CompressedSize ð13Þ The mean square error (MSEf) is calculated pixel by pixel for each video frame f of M � N pixels between the original frame, repre- sented by If, and the reconstructed compressed frame, Kf. Then, the PSNRf is calculated for each video frame f over the MSEf value. Notice that MSE is calculated over the luminance component Y’ (brightness) of the YUV color space to compute the PSNR. These two objective functions should be maximized, since higher values of CR will mean lower compressed video sizes, and a greater number of dBs in the PSNR metric is equal to a lower MSE, leading to a greater similarity between original and com- pressed video frames. 4.2. Decision variables The encoder we intend to optimize is based on intra-frame en- coder techniques (Connor, Brainard, & Limb, 1972), but has been extrapolated to inter-frame coding (Brofferio & Rocca, 1977), basi- cally using a video scene motion detection system (MMS) with the aim of performing simple motion compensation. It is based on JPEG2000 (ISO/IEC, 2000) and its still images compression method. This essentially divides each frame into independent accessible square regions, also called tiles. Our encoder uses tiles for condi- tional replenishment depending on whether or not motion is detected in the same way as similar encoders do with empty macroblocks. Each frame or still image in the video sequence can be com- pressed using JPEG2000 to a specific quality, which will determine the final quality and size of the compressed video in combination with the motion detection system. All the decision variables or low level encoder parameters that we will try to optimize are summarized below, stating the range of values of each one, divided into functional units. 1. Tiling setup � TSIZE = {16 � 16, 24 � 24, 32 � 32} 2. Motion detection system � Smoothing = {Enabled, Disabled} � MEANi[x] = [0, 255] � MAXi[x] = [0, 255] 3. Still image quality control � Bits per pixel in 32 � 32 (bpp32) = [0.4, 7.5] � Bits per pixel in 24 � 24 (bpp24) = [0.7, 8.73] � Bits per pixel in 16 � 16 (bpp16) = [1.2, 11.0] Tiling setup represents the size of the squared regions of the still images once compressed. The width and height of each tile must be the same. In theory, lower tile sizes will perform better compression, since they make the motion compensation system more precise. On the other hand, they add an extra overhead, mak- ing the compressor less efficient. Motion detection parameters are the attributes that manage the motion measurement subsystem (MMS) in tile regions. Large val- ues of MEANi[x] and MAXi[x] will represent a high MMS sensitivity, and lower tile reusability, which will imply better quality but less CR. We think that, in the case of Smoothing, which is related to the simple blur filtering, improves the MMS by reducing the noise that is usually present in videos, but we really do not know exactly what effect it has and whether there is any real improvement. On the other hand, still image quality control only has one parameter. It represents the bits per pixel that the JPEG2000 com- pressor should use for each still frame. Again, higher values of this parameter will lead to better quality but less CR. Notice that, for each tile size, there is a specific operating range for this value. Observe how each of these low level encoder parameters actu- ally represents a MO problem. For this reason, we will try to obtain a set of parameter values that achieves a set of solutions near to the optimal Pareto front. 5. Experiments In this section we present all the tests performed to get the near-optimal Pareto solutions set of the encoder, using the NSGA-II MOEA. We detail both the purposes of the tests and the parameters used in each execution and discuss the results. For these experiments, we will use a posteriori methods, that is, output almost the entire Pareto front of the encoder, leaving it to the decision maker to choose the solution that meets his or her requirements a posteriori. The decision variables for optimization are the stated in Section 4.2. Notice that TSIZE and Smoothing are discrete variables. There- fore, we will perform additional tests to determine their best val- ues. We will use the concept of Pareto dominance to compare the solutions achieved by the NSGA-II algorithm. The NSGA-II algorithm will run a maximum of 200 generations with an initial population of 200. The population size has been set in order to ideally get an individual about every 0.15 dB, covering the 25 to 55 dB encoder operating range. This way the DM’s solu- tion selection will be more precise in terms of quality. Each test will be performed twice with the Hall Monitor Fig. 10(a) and Akiyo 10(b) test sequences. 5.1. Video test sequences The test video sequences used to evaluate the encoder are very commonly used to evaluate encoder performance. Selected se- quences are 10(a) and (b) from the Xiph.org repository. They are uncompressed CIF-format sequences with a length of 251 frames and a size of 352 � 288 pixels. 7 Fig. 10. Test video sequences. Fig. 11. Smoothing test [Hall Monitor]. 5.2. Smoothing test The first test performed is related to the MMS, and particularly to determining how the Smoothing parameter affects f1 and f2 objective functions when optimizing the MEANi[x], MAXi[x], and Bits per pixel decision variables. Smoothing is a low level encoder parameter that can be enabled or disabled only. It should improve the performance of the encoder when enabled, but we have no hard evidence of this. It should be the first test run since this parameter depends on the tile size parameter, but not viceversa. In this case then we can establish an arbitrary tile size, i.e. 32 � 32. The Bits per pixel range of values is set according to the tile size, as defined in the decision variables section and shown in Table 1. This test will generate two Pareto fronts that differ only as to whether the encoder uses Smoothing prefiltering. So, we could use the concept of Pareto dominance to determine whether Smoothing is really useful. Figs. 11 and 12 both show the Pareto solution set obtained for each test video with the Smoothing parameter enabled and dis- abled. It is clear that the Pareto solution sets with the Smoothing parameter enabled dominate the others with Smoothing disabled. This means that the Smoothing parameter really does improve the MMS and the final compression performance when enabled. So, for the best results, the Smoothing value should be enabled dur- ing the compression. In the following tests, Smoothing will always be enabled. 5.3. Tile size test In the second test we will try to determine the influence of the tile size on compression. We can determine whether there is a tile size that outperforms the others. A priori, lower tile sizes should perform better compression, improving the MMS, since smaller tile sizes usually fit any object in the image better. Also other video codecs use small macroblock sizes, i.e. H.264 (Wiegand, Sullivan, Bjntegaard, & Luthra, 2003) that works with 16 � 16, 8 � 8 and also 4 � 4. Table 1 Encoder parameters Test1. Test1 Run1 Run2 TSIZE 32 � 32 32 � 32 Smoothing Enabled Disabled MEANi[x] [0 255] [0 255] MAXi[x] [0 255] [0 255] bppi [0.4,7.5] [0.4,7.5] Fig. 12. Smoothing test [Akiyo]. 8 Note, however, that more data are necessary to represent smal- ler tile sizes in JPEG2000, i.e. observe how the lower bound of the Bits per Pixel low level encoder parameter grows as tile size de- creases. This indicates that there is an extra overhead on the inclu- sion of each tile. In this test we want to compare the Pareto fronts achieved by the MEANi[x], MAXi[x], and Bits per pixel decision variables with the different possible values for the Tile size parameter. We could determine the best value for this parameter comparing the Pareto solution sets. Table 2 describes the parameters used in this case for the three executions of this test, one for each tile size. Notice, in this case, that Smoothing is enabled in all tests, as a consequence of the pre- Table 2 Encoder parameters Test1. Test2 Run1 Run2 Run3 TSIZE 16 � 16 24 � 24 32 � 32 Smoothing Enabled Enabled Enabled MEANi[x] [0 255] [0 255] [0 255] MAXi[x] [0 255] [0 255] [0 255] bppi [1.2,11.0] [0.7,8.73] [0.4,7.5] Fig. 13. Tile sizes test [Hall Monitor]. vious result. This test will provide three Pareto fronts, indicating the best tile size to be used with the encoder. Fig. 13(a) and Fig. 14(a) show how all the Pareto fronts obtained are very similar for each tile size. But there is a slight performance improvement when using 32 � 32 tiles. Fig. 13(b) and Fig. 14(b) show a detailed area of the Pareto front, located at the Pareto front inflection point, clearly indicating how the 32 � 32 Pareto front dominates the other Pareto solution sets. From the comparison of these Pareto fronts generated with NSGA-II we can select the best value for the Tile Size low level en- coder parameter, which will be set at 32 � 32. We also have to prove how the overhead present with small tile sizes is not offset by the theoretical improvement in the MMS. 5.4. Results of optimization In the previous tests we have determined the best values for the discrete encoder parameters (TSIZE and Smoothing), and we now know that TSIZE = 32 � 32 and Smoothing = enabled improves the compression. The remaining parameters (MEANi[x], MAXi[x] and bppi) are continuous, and their values are determined for each vec- tor of solutions by the NSGA-II algorithm. Fig. 14. Tile sizes test [Akiyo]. 9 Fig. 15. Final Pareto solution set [Hall Monitor]. Fig. 16. NSGA-II converg The DM using the results of this optimization could select any solution from the Pareto front that meets his or her requirements (in terms of quality or compression ratio). The Pareto front con- tains a set of values of MEANi[x], MAXi[x] and bppi. The DM can also use the Pareto front to evaluate the trade-offs between all solu- tions. In this section we will report the Pareto solution sets for each video sequence. Figs. 15 and 17 represent the best Pareto solution sets for each compressed video. They illustrate how these solution sets cover a wide operating range in terms of quality and compres- sion ratio. Also we want to show the convergence speed of NSGA-II for such problems. Fig. 16(a)–(d) show the evolution of generations for the Hall Monitor test video, and how the solution approximated in the 20th generation is almost equal to the Pareto front solution set obtained with 200 generations, illustrated in Fig. 15. With the Akiyo test sequence, on the other hand, convergence speeds up, and, by the 5th generation (Fig. 18), we get a good Par- eto solution set, compared with the solution obtained in 200th generation, shown in Fig. 17. ence [Hall Monitor]. 10 Fig. 17. Final Pareto solution set [Akiyo]. Fig. 18. NSGA-II conv This demonstrates the suitability of using MOEAs for such prob- lems that have a large search space and are not suited for approx- imation using classical search-based techniques. In actual fact, they take about two minutes to evaluate f1 and f2 objective func- tions for each individual in this problem. 6. Conclusions In this paper we have worked with the multiobjective definition in the field of video compression. We have used NSGA-II to opti- mize the internal parameters of a new video codec called MIJ2K, but this procedure could be extrapolated to any video compressor with good results. Using MOEAs we can output an entire Pareto solution set covering the whole operating range of the codec for both quality and compression ratio. Also we show the suitability of using MOEAs in the field of video compression, since they achieve a fast convergence speed, as re- quired by such problems due to the cost of evaluating each objec- tive function. Even so, we intend to further research this problem. In this pa- per, we have optimized the low level encoder parameters in order ergence [Akiyo]. 11 to create a Pareto solution set for a specific test video. The solutions obtained can be used with different videos that share similar pro- files to the videos used here with the aim of reusing the optimization. In future work, we will research real-time optimization depend- ing on given constraints independently of the video that is being compressed. This is potentially very useful, for instance, when the video compressor is used for real-time streaming purposes and the network constraints are variable depending on load. The algorithm should then search new optimal solutions for the new given constraints. Acknowledgments This work was supported in part by Projects CICYT TIN2008- 06742-C02-02/TSI, CICYT TEC2008-06732-C02-02/TEC, SINPROB, CAM MADRINET S-0505/TIC/0255 and DPS2008-07029-C02-02. References Adams, M. 2001. The JPEG-2000 still image compression standard. ISO/IEC JTC 1/SC 29/WG 1, 2412. Adams, M., Ward, R. 2001. Wavelet transforms in the JPEG-2000 standard. In Proceedings of IEEE Pacific Rim Conference on Communications, Computers and Signal Processing, vol. 1, pp. 160–163. Back, T., Fogel, D. B., & Michalewicz, Z. (Eds.). (1997). Handbook of Evolutionary Computation. Bristol, UK, UK: IOP Publishing Ltd.. Brofferio, S., & Rocca, F. (1977). Interframe redundancy reduction of video signals generated by translating objects. IEEE Transactions on Communications [legacy, pre-1988], 25(4), 448–455. Christopoulos, C., Skodras, A., & Ebrahimi, T. (2000). The JPEG 2000 still image coding system: an overview. IEEE Transactions on Consumer Electronics, 46(4), 1103–1127. Coello, C. A. C., Lamont, G. B., & Veldhuizen, D. A. V. (2006). Evolutionary Algorithms for Solving Multi-Objective Problems (Genetic and Evolutionary Computation). Secaucus, NJ, USA: Springer-Verlag New York, Inc.. Connor, D., Brainard, R., & Limb, J. (1972). Intraframe coding for picture transmission. Proceedings of the IEEE, 60(7), 779–791. Czuni, L., Csaszar, G., & Licsar, A. (2006). Estimating the Optimal Quantization Parameter in h. 264. In ICPR ’06: Proceedings of the 18th International Conference on Pattern Recognition (pp. 330–333). Washington, DC, USA: IEEE Computer Society. Deb, K. (2001). Multi-Objective Optimization using Evolutionary Algorithms. Chichester: Wiley-Interscience Series in Systems and Optimization. John Wiley & Sons. Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002). A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE Transactions on Evolutionary Computation, 6, 182–197. Ehrgott, M., & Gandibleux, X. (2002). Multiple Criteria Optimization: State of the Art Annotated Bibliographic Surveys. Kluwer Academic Publishers.. Hang, H.-M., Chou, Y.-M., & Cheng, S.-C. (1997). Motion estimation for video coding standards. Journal of VLSI Signal Processings and Systems, 17(2–3), 113–136. ISO/IEC. 15444-1:2000 information technology jpeg2000 image coding system-part 1: core coding system. Technical report. Jiang, M. 2006. Adaptive rate control for advanced video coding. PhD thesis, Santa Clara, CA, USA. Adviser-Nam Ling. Koski, J. (1988). Multicriteria truss optimization. Engineering and in the Sciences. kuang Chen, Y., Vetro, A., Sun, H. and Kung, S.Y. 1997. Optimizing intra/inter coding mode decisions. In Proceedings of International Symposium on Multimedia Information Processing, pp. 561–568. Lora, M. 1994-2008. Xiph.org:: Test media. World Wide Web electronic publication. Luis, A., Patricio, M. Scalable Streaming of JPEG 2000 Live Video Using RTP over UDP. Miettinen, K. (2001). Some methods for nonlinear multi-objective optimization. Lecture Notes in Computer Science, 1–20. Ng, J. K.-Y., Leung, K. R., & Hui, C. K.-C. (2005). A qos-enabled transmission scheme for mpeg video streaming. Real-Time Systems, 30(3), 217–256. Pennebaker, W., & Mitchell, J. (1993). JPEG Still Image Data Compression Standard. Kluwer Academic Publishers.. Rabbani, M., & Joshi, R. (2002). An overview of the JPEG 2000 still image compression standard. Signal Processing: Image Communication, 17(1), 3–48. Sawaragi, Y., Nakayama, H., & Tanino, T. (1985). Theory of Multiobjective Optimization. Orlando: Academic Press. Shi, Y., Sun, H. 2000. Image and video compression for multimedia engineering: fundamentals, algorithms, and standards, CRC Pr I Llc. Steuer, R. E. (1986). Multiple Criteria Optimization: Theory Computation and Application. New York: John Wiley. pp. 546. Thurston, D. L. (2006). Multi-Attribute Utility Analysis of Conflicting Preferences. In Kemper E. Lewis et al. (Eds.), Decision Making in Engineering Design. New York, New York: ASME. Tu, Y.-K., Yang, J.-F., Shen, Y.-N., & Sun, M.-T. (2003). Fast Variable-Size Block Motion Estimation using Merging Procedure with an Adaptive Threshold. In ICME ’03: Proceedings of the 2003 International Conference on Multimedia and Expo (pp. 789–792). Washington, DC, USA: IEEE Computer Society. Wallace, G. 1991. The JPEG still picture compression standard. Wang, Y.-C., & Leou, J.-J. (2003). A Rate Control Scheme for h. 26l Video Transmission. In ICME ’03: Proceedings of the 2003 International Conference on Multimedia and Expo – Volume 3 (ICME ’03) (pp. 349–352). Washington, DC, USA: IEEE Computer Society. Wiegand, T., Sullivan, G., Bjntegaard, G., & Luthra, A. (2003). Overview of the H. 264/ AVC video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7), 560–576. Zhang, Q., Zhu, W., & Ya-W, Z. (2005). End-to-end qos for video delivery over wireless internet. Proceedings of the IEEE, 93(1), 123–134. 12 
work_givu6ffrhvdmlhy3rsho4ruonu	----	Corresponding author: Francesco Di Maio, Politecnico di Milano, Energy Department, Via La Masa 34, 20156, Milano, Italy. Tel: +39 02 23996372, Fax: +39 02 2399 8566. Email: francesco.dimaio@polimi.it A REGIONAL SENSITIVITY ANALYSIS-BASED EXPERT SYSTEM FOR SAFETY MARGINS CONTROL Francesco Di Maio1, Alessandro Bandini1, Michele Damato1, Enrico Zio1,2 1Energy Department, Politecnico di Milano, Milan, Italy 2Chair on System Science and Energetic Challenge, Fondation – EDF, Centrale Supélec, Paris, France Abstract Thermal-Hydraulic (TH) codes are used to simulate the response of nuclear safety systems under transient and accident conditions. The outcomes of the simulations are used to verify the safety margins required for safe operation and make decisions on how to maintain them. In this work, a novel Expert System (ES) based on Regional Sensitivity Analysis (RSA) is developed to guide a system undergoing an accident scenario towards the safest conditions in the optimal number of operation. The ES proceeds by firstly identifying the (uncertain) system controllable variables (i.e., control rods position, feed- water flow rate, void fraction inside the steam generator, etc.) that most affect the system response by RSA; then, the limit-state function is calibrated on a dataset of outcomes of TH code runs and the system failure boundary (i.e., the limit surface) is defined on the set of (uncertain) TH input variables. Application of the ES is firstly shown with respect to an analytical case study that artificially simulates the response of a NPP to an accident scenario and, then, to a practical case study concerning the response of the pressurizer of a Pressurized Water Reactor (PWR). Keywords: Nuclear System, Expert System, Best-Estimate Thermal-Hydraulic Code, Safety Margins, Limit-State Function, Regional Sensitivity Analysis, Pressurizer. 1. INTRODUCTION Safety remains a priority for Nuclear Power Plants (NPPs) design and operation, as the release of radioactive material can result in catastrophic consequences in terms of casualties, environmental pollution and financial losses [Hsieh et al., 2012]. For the safety of NPPs, accident-preventive design and operation, and effective consequence mitigation plans are developed [Ma et al., 2011]. In practice, Emergency Operating Procedures (EOPs) are defined to provide the technical basis for suitable operator response to Design Basis Accidents (DBAs) and, nowadays also, Beyond Design Basis Accidents (BDBAs) [IAEA, 2006]. Fault Detection and Diagnosis (FDD) methods have been developed for detecting different types of faults and supervising the plant behavior for accident prevention [Ma et al., 2011; Park et al., 2015; IAEA, 2000; IAEA, 2004; IAEA, 2005]. Upon localization and isolation, under certain abnormal conditions, manual actions are conducted by the plant operators for restoring mailto:francesco.dimaio@polimi.it 2 normal operating conditions [Hsieh et al., 2012; Park et al., 2015]. For this, Abnormal Operating Procedures (AOPs) are provided, where the sequence of actions to undertake are given for different Initiating Events (IEs). Two practical issues are: i. it is difficult to ensure sufficient coverage of all possible IEs through AOPs, as they are developed based on historical data and there may exist significant IEs that have not yet been experienced [Park et al., 2015]; ii. human errors may occur during abnormal situations and incorrect AOPs may, then, be followed [Hsieh et al., 2012]. In abnormal situations, time is critical, and the amount of information and data to be examined is large. Decision Support Systems (DSSs) can aid the operators and reduce the possibility of errors [Ma et al., 2011; Hsieh et al., 2012; Park et al., 2015]. Examples of DSSs are: the Alarm and Diagnosis-Integrated Operator Support (ADIOS) system that attempts to avoid that too many alarms influence the operators judgement in a wrong way [Kim et al., 2001]; the Hidden Markov Model (HMM) for recognizing accidents in NPPs, proposed in [Kwon et al., 1999]; the Fault Diagnosis Advisory System (FDAS), based on dynamic neural networks [Lee et al., 2007]; the Dynamic System Doctor (DSD), a system-independent interactive software for on- line state/parameter estimation in dynamic system [Aldemir et al., 2001]; the Analysis of Dynamic Accident Progression Trees (ADAPT) methodology [Hakobyan et al, 2008]; the unsupervised clustering technique for NPP components fault diagnosis, based on Haar Wavelet Transform (HWT) and Fuzzy C-Means (FCM) algorithm, in [Baraldi et al., 2013]; the hybrid approach for balancing false and missed alarms, based on Correlation Analysis (CA), Genetic Algorithms (GAs) and Sequential Probability Ratio Tests (SPRTs) in [Di Maio et al., 2013]; the non-parametric decision strategy in [Al-Dahidi et al., 2014] to detect whether NPP components are in abnormal conditions, using Prediction Intervals (PIs) and Auto-Associative Kernel Regression (AAKR) [Hines et al., 1998]. However, there is no guarantee of improvement in the operators’ performance and sometimes the result could be the increase of operator workload, with negative implications on performance [Yoshikawa 2005; Kim et al., 2007; Hsieh et al., 2012]. In this paper, we propose an Expert System (ES) [McBride et al., 1993; Liao, 2005; Ikram et al., 2015] for operator aid, based on the system response outcomes obtained from a Thermal- Hydraulic (TH) code and Regional Sensitivity Analysis [Wei et al., 2014]. The TH code 3 reproduces the system physical behavior according to a mathematical model m that receives an input vector n x  (comprised of n input variables) and generates the system response vector )(xmy  , containing (at least) one safety parameter y . The input variables set n  can be partitioned into two subsets: one with controllable variables ( q  ), i.e. the levers under control of the plant operator, which can be manipulated to increase plant safety (e.g., reactor control rods position, rate of feed-water flow through the plant primary loops, accumulator water temperature and pressure, repair times, etc.), and the other one with non-controllable variables ( qn  ) [Di Maio et al., 2016]. In practice, the inputs x are uncertain [Apostolakis, 1990; Helton, 1996; Oberkampf et al., 2004]. With reference to scenario FE , for which the response of interest Y must be lower than a threshold y γ (imposed by regulation), the limit-state function G is defined as [Bourinet et al., 2011]: y XYXG  )()( (1) The input space can be split into a failure domain }0)(:{  XGXF and a safe domain }0)(:{  XGXS . The system failure boundary (limit surface) }0)(:{  XGXF , which separates S from F , can be projected onto an input controllable subset ( q  ) [Di Maio et al., 2016], which contains the controllable variables q XXX ,,, 21  that can be varied to increase the system safety margin )( XYγy  [Zio et al., 2010]. While the system is operating in its nominal state ex (usually assumed equal to the expected values ][ XE of the input variables X ), the safety parameter )( ee xmy  falls well below the safety threshold y γ , providing a margin of safety with respect to uncertainties in the model [Zio et al., 2008]. In abnormal conditions, the safety margin may decrease and this needs to be kept under control in order to avoid diverging to catastrophic accidents. An additional difficulty in the problem is due to the fact that any generic controllable input variable j X , qj ,,2,1  , can be affected by epistemic and/or aleatory uncertainty. In this paper, this uncertainty is modeled by considering that when j X is set equal to jx̂ , the input variable actually ranges between jj xx ˆ and jj xx ˆ according to the modified Probability Density Function (PDF): 4 )ˆ()ˆ( )( )( * jjXjjX jX jX xxFxxF xt xf jj j j   , ]ˆ,ˆ[ jjjjj xxxxx  (2) where )( jX xt j is the truncated PDF of jX over range ]ˆ,ˆ[ jjjj xxxx  :             ),ˆ(,0 ˆ,ˆ),( )ˆ,(,0 )( jjj jjjjjjX jjj jX xxx xxxxxxf xxx xt jj (3) Coherently, the system (uncertain) initial state is: ][][][ 0,0,20,220,210,110,10 qqqq x,xxxx,xxxx,xxxX   (4) In this paper, we present the development of an Expert System (ES) for maximizing the safety margin by exploiting the results of a Regional Sensitivity Analysis (RSA) to guide the search for the range (or state of the system X ) closest to 0 X , where ])([ XG is the smallest. In detail, the ES proposed in this paper is based on the Revised Ratio Functions (RRFs) [Wei et al., 2014]) that measure the impact on the mean (Revised Mean Ratio Function, HM ) and variance (Revised Variance Ratio Function, HV ) of the model output distribution due to the reduction in the range of variability of an individual input. For our purposes, the effect on the variance is of particular interest and it is here exploited to find which input controllable variable allows obtaining the largest decrease in G (i.e., the largest increase in the expected safety margin) when the system is in its current abnormal state X . Assuming that the system is operating at X and that range ]ˆˆ[ jjjj xx,xx  , ,,,2,1 qj  is reduced to jx̂ , the expected value (  ) and variance (Var ) of G are, respectively:               11 11 ,1 * 1 ))(()ˆ|,,(ˆ| xx xx xx xx xx xx n j Xjqj qq qq dxxfxxxgxG       (5) 5                 11 11 ,1 *2 1 ))((ˆ|)ˆ|,,(ˆ| xx xx xx xx xx xx n j Xjjqj qq qq dxxfxGxxxgxGVar      (6) The original RSA technique defines j HV of input j X , qj ,,2,1  , as the ratio between the residual variance ]ˆ|[ jxGVar with respect to the residual mean ]ˆ|[ jxG , and the full variance  GVar [Wei et al., 2014]:  GVar xGVar HV j j ]ˆ|[  (7) where:             11 11 1 *2 1 ))(()][),,((][ xx xx xx xx xx xx n Xq qq qq dxxfGxxgGVar       (8)               11 11 1 * 1 ))((),,( xx xx xx xx xx xx n Xq qq qq dxxfxxgG       (9) As Eq. (7) shows, j HV indicates the actual reduction of the model output variance due to the restriction of the range of j X . The larger j HV is, the more function G varies (i.e., decreases or increases) for a perturbation of the other inputs X ( j ). Therefore, we define the revised * j HV as the ratio between the variance of G , computed over the range ]ˆˆ[ jjjj xx,xx  , qj ,,2,1  , and the reduced range ( x̂ ) of each of the other inputs ,X q,,2,1  , ,j and the full variance  GVar . It is clear that the larger *jHV is, the more function G varies and, thus, a larger increase in the system safety margins can be achieved through a perturbation of the j-th input. Through the approach described, we aim at providing the plant operator with an effective and unequivocal AOP to keep the safety margins and avoid system failure, giving clear indications of which controllable variables to modify and by how much, when an IE has occurred. The rest of the paper is organized as follows. Section 2 illustrates the proposed approach. Section 3 shows an application of the approach to an analytical case study that artificially 6 simulates the response of a NPP to an accident scenario. Section 4 contains the application of the approach to the pressurizer of a Pressurized Water Reactor (PWR). Conclusions are given in Section 5. 2. AN EXPERT SYSTEM FOR THE MAXIMIZATION OF SAFETY MARGINS DURING SYSTEM OPERATION For the sake of clarity, but without loss of generality, in this work the non-controllable input variables (i.e., nqq XXX ,,, 21   ) are considered to be fixed at some known values (i.e., nqq xxx ~,,~,~ 21   ) and )( XG is, consequently, a function of the controllable inputs only (i.e., q XXX ,,, 21  ). Given a current abnormal state of the system X , the most effective way to “escape” from it and reach safer conditions is to “perturb” the χ-th controllable variable X ,  q,1 , that is associated to the larger * j HV value, nj ,,2,1  , in order to attain the largest possible variation of )( XG (i.e., of the safety margin) by adding or subtracting a constant h to its nominal value j x̂ . When a perturbation of X does not result in a reduction of  )( XG (i.e., the expected safety margin value), it is considered that the system has reached normal conditions. Based on these knowledge rules, the ES algorithm performs the following four steps at each υ- th iteration of safety margins control, N,,1,0  (where 1N is the final number of iterations): 1. it computes: i. the realization )(  xgg  of G when the system is in its current nominal state ),,,( ,,2,1  qxxxx  ; ii. the expected value  )( XG over the range of input uncertainty, representing the system actual state       qqqq x,xxxx,xxxx,xxxX   ,,2,22,21,11,1  ; iii. * j HV of each input j X , qj ,,2,1  , over its own range   jjjj,j, xxxxX   ,, , with the initial ranges of all the other inputs X reduced to  ,,,2,1 ,,,,, qxxxx  ( q,,2,1  , j ). 7 2. it identifies the input variable X ,  q,1 , with the largest * j HV value (i.e.,   }{max: * , ,1 * ,  j qj HVHVX   that, once “perturbed”, generates the largest change in G ; 3. it computes two realizations  g and  g of function G in  x and  x , respectively, where: i. ),,,,,( ,,2,1  qxhxxxx    ; ii. ),,,,,( ,,2,1  qxhxxxx    ; iii. h is the distance between the current system nominal state x and the one that follows 1x (see below and Section 3 for further details). 4. it checks the following statements: i. if  gg ˆˆ   and ])([])([  XGXG   , then     xx 1 ; ii. otherwise, if  gg ˆˆ   and ])([])([  XGXG   , then     xx 1 . In other words, once the (χ-th) direction along which the largest change in G can be achieved is found (as in Step 3), Step 4 looks for the system nominal state that results in a decrease in G between  x and  x and it finally checks that ])([ XG is actually decreasing. The algorithm stops when neither one of the two above statements is verified because ])([])([  XGXG   and ])([])([  XGXG   , which means that moving the system from its nominal current state x to  x or to  x would not result in a decrease of ])([ XG . As it can be seen, the algorithm identifies a sequence of nominal states 1,,2,1 }{  Npp x  in S , obtained by shifting at each step υ the system nominal state x of a length h in the direction of the input variable ( X , ],1[ q ) that allows the largest reduction of  )( XGE (computed as in Step 1 Point iii). The choice of the step length h is a delicate task, because of two reasons: 1. the shorter h is, the more accurate is the final nominal state 1Nx , but the larger is the number of iterations N that the algorithm needs to perform to get to a final solution; 8 2. * ,jHV is computed at the υ-th step of the algorithm for each j-th input variable over the range ],[ , jjjj,j, xxxxRG   , whose length equals x2 ; it is, then, suggested that xh  because at step υ there is no information on the behavior of function g outside range j,RG . As ),,,( 21 q xxxg  is generally a complex function of q xxx ,,, 21  , the integrals in Eq. (7) are very complicated and cannot be solved analytically. Conversely, numerical techniques such as Monte Carlo integration [Zio, 2013], can allow approximating such integrals and computing ])([ XGE and ,jHV in a reasonable time at each υ-th iteration of the algorithm. In all generality, if we are trying to estimate the integral I of a function )( xvv  over an integration domain n  (i.e., xdxvI    )( ), where n x  , and a PDF )(xf X can be defined over  , then I is equivalent to [Zio, 2013]:          )( )( )( )( )( xf xv xdxf xf xv I X X X (10) where        )( )( xf xv is the expected value of the ratio )( )( xf xv computed over  with respect to a random variable distributed according to )(xf X . Eq. (10) is true for any PDF )(xf X on  , as long as 0)( xf X whenever 0)( xv . The value of        )( )( xf xv can, then, be estimated by: i) generating SN random samples )(c x (with SNc ,,2,1  ) of x according to )(xf X , ii) computing the ratio )( )( )( )( c X c xf xv for each sample and iii) finding the average SN I of these values:    S S N c c X c S N xf xv N I 1 )( )( )( )(1 (11) 9 According to the law of large numbers, as more and more samples are taken the Monte Carlo estimator SN I is guaranteed to converge to I , with a standard deviation that decreases as a function of the number of samples SN , by a rate of SN [Rubinstein, 1981; Kalos et al., 2008; Zio, 2013]. In this work, the pseudo-random samples )(c x are generated from the uniform distribution )(xu X , x , according to the algorithm originally proposed by von Neumann [Zio, 2013]. 3. ANALYTICAL CASE STUDY 3.1 Analytical Model Description Let us assume that the safety-significant response variable Y of the system is given by an analytical function m of two controllable inputs 1X and 2X as: 3 2 2 21 2 121 )1()2()3(8),(  XXXXXXmY (12) where  10,101 X and  10,102 X are random variables (such as, reactor control rods position, reactor coolant mass flow rate, steam generator feed-water mass flow rate, surge line mass flow rate, etc. [Cammi et al., 2006; Baraldi et al., 2013; Di Maio et al., 2015]), obeying two independent normal distributions )4,2(1N and )25.6,0(2N , respectively. The safety threshold limit to be applied to Y is 125 y γ and the resulting limit-state function of the system is [Bourinet et al., 2011; Zio et al., 2008]: 125)1()2()3(8),(),( 3 2 2 21 2 12121  XXXXγXXYXXG y (13) In Figures 1 and 2, the limit-state function G (continuous grid) and the system failure boundary F (solid line) are shown as functions of variables 1X and 2X . According to the definition of failure boundary given in Section 1, in this simplified case, F (Figures 1 and 2) can be easily found by the intersection between surface ),( 21 xxg (Figure 1) and plane 0g . In Figure 2, the safe ( S ) and failure ( F ) domains of the system are also shown. 10 Figure 1: limit-state surface G and failure boundary F of the analytical model m considered. Figure 2: failure boundary F of the analytical model m represented in the input space ]10,10[]10,10[  . 3.2 Assumptions As already discussed in Section 1, the uncertainty affecting input j X , 2,1j , with j X being fixed at some value  10,10ˆ  j x , is here modeled by considering j X to be ranging in the interval ]ˆ,ˆ[ jjjj xxxx  . For the sake of simplicity and without loss of generality, j x is independent of j x~ and it is computed as follows, for both 1X and 2X : 11 5.0   I jj j N ab x , 2,1j (14) where 10 j a and 10 j b are the lower and upper bounds of the domain ],[ jj ba of input j X , respectively, while 20IN is an arbitrary number of subintervals in which ],[ jj ba is partitioned. In a real case, the magnitude of j x would depend on the actual variability of j X and represents the confidence on how close j X actually is to its nominal value j x̂ (for instance, j X may describe the mass flow rate removed from the Steam Generator (SG) of a NPP by a set of safety valves during a “swell and shrink” accident [Di Maio et al., 2015], with j x̂ and j x equal to 25 [kg/s] and 5 [kg/s], respectively). Furthermore, the step length h is set equal to its maximum allowed value (Section 2) that is 0.5 in this case (for example, if j X represents the mass flow rate through the SG safety valves, h is the increase or decrease in j X resulting from regulating the valves operating position). 3.3 Results The methodological steps illustrated in Section 3.2 have been applied to five different initial nominal positions of the system )( 0  x , 5,,2,1  , in S . After applying on each )( 0  x the set of rules that constitutes the ES detailed in Section 2, five final nominal positions )( 1   N x have been obtained as reported in Table 1, where 1N is the final number of iterations needed to identify )( 1   N x . x 0,1 (ς) x 0,2 (ς) x Nς-1,1 (ς) x Nς-1,2 (ς) 1 -3.0 -2.0 0.0 -0.5 11 2 3.0 -2.0 0.0 -0.5 11 3 0.0 3.0 0.0 -0.5 15 4 8.0 6.0 9.0 10.0 12 5 0.0 6.0 9.0 10.0 28 N ς +1 Initial nominal states x 0 (ς) ς Final nominal states x Nς-1 (ς) 12 Table 1: initial ( )( 0  x ) and final ( )( 1   N x ) nominal states of the system in S , and number of performed iterations ( 1N ). From Table 1, one can read that )5.0,0.0( )3( 1 )2( 1 )1( 1 321   NNN xxx and ).0.10,0.9()5( 1 )4( 1 54   NN xx Indeed, depending on the initial nominal state of the system )( 0  x , )( 1   N x is either )5.0,0.0( P or )0.10,0.9(P , as )625.0,118.0( Q and P are two local minima of function g (see Figure 1). As an example, Figure 3 shows the optimal sequence 1,,2,1 )( }{    Npp x  of the system nominal states computed for 1 , where the system initial nominal position is )2,3()1( 0 x and the total number of iterations is 1111 N . To explain in more detail, Figure 4 illustrates each computed system state )( p x , 9,,2,1 p , coupled with its corresponding realization )( )1( pp xgg  of function G . As can be noticed, at each step the algorithm successfully identifies the direction along which function G varies the most between the 1X and 2X axes; then, it brings the system into a new optimal state, where p g is lower; finally, the algorithm stops when ])([ pXG cannot be reduced any further and the system is ultimately located in position )5.0,0.0( P . 13 Figure 3: sequence of system optimal nominal states 1,,2,1 )( }{    Npp x  for 1 . Figure 4: sequence of system optimal nominal states 1,,2,1 )( }{    Npp x  coupled with the corresponding realizations 1,,2,1 )( )}({     Nppp xgg  . For the sake of completeness, in Figure 5 the entire sequence of optimal nominal states of the system 1,21 )( }{   …,N,p=p x is illustrated for 5,,2,1  . Figure 5: sequences of system optimal nominal states 1,21 )( }{   …,N,p=p x computed for 5,,2,1  . 14 These results can be usefully compared with those obtained in [Di Maio et al., 2016] for the same analytical case study of Eq. (11). However, it should be honestly pointed out that these two works rest on two different rationales: i. in [Di Maio et al., 2016], the objective is to identify the farthest system state * x from the failure boundary F , regardless of the system working conditions when the accident scenario FE is triggered. Since * x is the farthest from F , it is also considered the optimal position for the system, i.e., the one where the final failure event is less likely to occur; ii. in this work, the initial abnormal state of the system 0 X is of paramount importance, as it is required to assess the feasibility of “escaping away” from failure starting from it. Indeed, the proposed ES evaluates where the system should go step-by-step in the input space depending on the behavior of function G , which is representative of the system safety margin. In [Di Maio et al., 2016], the only non-controllable variable y  (i.e., the system safety threshold) is assumed to be normally distributed on the range [−500,2500], with mean 500 y  and variance 2500 2  y  . Based on the randomness of variable y  , various safest operating conditions (i.e., states) ),( * 2 * 1 * xxx  are recommended for the system, depending on the operator risk attitude: i. risk-averse: )00.10,63.4( * x ; ii. risk-prone: )00.10,00.8( * x ; iii. mean: )95.9,91.5( * x ; iv. median: )00.10,25.4( * x . where the risk-prone and risk-averse conditions are defined as the 80-th and 20-th percentiles of a population of safest operating conditions computed for different values of y  , respectively. 15 In this case study, y  is supposed to be known and set equal to 125. As Table 1 and Figure 5 show, for 3,2,1 the final optimal nominal state )( 1   N x of the system (i.e., )5.0,0.0( P ) is far from all the solutions * x of [Di Maio et al., 2016]. This is due to the different criterion adopted in determining 1,21 )( }{   …,N,p=p x and ),( * 2 * 1 * xxx  : in [Di Maio et al., 2016] the (possibly unreachable) safest state * x is identified as the input variables setting for which the system is best sheltered against failures, whereas, in this work we provide the plant operator with instructions on how to change step by step the abnormal variables from the setting 0x to a normal state 1N x , which may (or may not) correspond to * x (that might not be realistic to reach for some accident scenarios FE ), depending on the actual behavior of function G . 4. PRESSURIZER CASE STUDY 4.1 Pressurizer Model Description The pressurizer of a PWR NPP, whose scheme is shown in Fig. 6, has been simulated through a Matlab SIMULINK model that is here taken as the TH model of the component [Baraldi et al., 2013]. The developed TH model is based on the application of the mass and energy conservation equations to the two regions of vapor and liquid. The exchanges between the two regions, due to evaporation of the liquid and condensation of the gas, have been taken into account [Kuridan et al., 1998, Todreas et al., 1990]. The system of non-linear differential equations describing the TH model is detailed in [Baraldi et al., 2010]. 16 Fig. 6: Scheme of a PWR NPP pressurizer. Let us assume the safety-relevant response Y of the system to be the pressure reached within the pressurizer when two controllable variables X1 and X2 (i.e., sprayers mass flow rate and heaters power, respectively) are the levels under control of the operators to counteract any developing surge line flow rate excursion scenario. To estimate g, a data set has been built in order to represent a realistic situation simulating 100 surge line transients, whose initial conditions have been taken to represent the realistic situation of a standard PWR pressurizer, (Table 2), whereas the surge mass flow rate excursion has been randomly sampled from the uniform distribution in the range [-10,10] kg/s. Initial condition Level 7.221 m Liquid temperature 342.1 °C Vapor temperature 342.3 °C Pressure 150.0 bar Table 2: Initial conditions of the pressurizer. 17 If 𝑋1 ∈ [0,7] kg/s is distributed like a N~(3.5, 2.4) and 𝑋2 ∈ [0,320] kW is distributed like a N~ (160,107.650), the resulting Y fitted to the available dataset is equal to Eq. (15). 𝑌 = 𝑚(𝑋1,𝑋2) = 160.6 − 11.2𝑋1 + 10 −2𝑋2 + 2.5 · 𝑋1 2 + 2.9 · 10−3𝑋1𝑋2 − 0.1 · 𝑋1 3 (15) The safety threshold γy is set equal to 155 bar and the resulting limit-state function G is given in Eq. (16). 𝐺(𝑋1,𝑋2) = 𝑌(𝑋1,𝑋2) − 𝛾𝑦 = 5.6 − 11.2𝑋1 + 10 −2𝑋2 + 2.5 · 𝑋1 2 + 2.9 · 10−3𝑋1𝑋2 − 0.1 · 𝑋1 3 (16) In Figures 7 and 8, the limit-state function G (continuous grid) and the system failure boundary 𝜕𝐹 (solid line) are shown as functions of variables X1 and X2. In Figure 8, the safe (S) and failure (F) domains are also shown. Figure 7: limit-state function G and failure boundary F of the pressurizer model m. 18 Figure 8: failure boundary F of the pressurizer model m represented in the input space ]320,0[]7,0[  . 4.2 Assumptions The uncertainty affecting the input Xj, j = 1,2, is here modelled considering Xj to be ranging in the interval [𝑥�̂� − ∆𝑥𝑗,𝑥�̂� + ∆𝑥𝑗]. The ∆𝑥𝑗 values to be used in what follows are listed in Table 3. j aj bj NI Δxj 1 0 7 18 0.39 2 0 320 18 17.8 Table 3: Values of Δxj. The step length h is set equal to its maximum allowed value (Section 2) for each one of the variables, i.e. h1 = 0.78 kg/s and h2 = 35.6 kW (with jX representing the mass flow rate through 19 the sprayers of the pressurizer and the power of its heaters respectively, j h is the increase or decrease in j X resulting from regulating the two different systems under operator’s control). 4.4 Results The methodological steps have been applied to three different initial nominal positions of the system �̅�0 (𝜍) , ς = 1, 2, 3, that belong to S and are close to 𝜕𝐹. The ES has been applied to these initial positions and the final system nominal states �̅�𝑁𝜍−1 (𝜍) have been obtained as shown in Table 4, where Nς+1 is the number of iterations needed to reach �̅�𝑁𝜍−1 (𝜍) . ς Initial nominal states �̅�0 (𝜍) Final nominal states �̅�𝑁𝜍−1 (𝜍) Nς+1 �̅�0,1 (𝜍) �̅�0,2 (𝜍) �̅�𝑁𝜍−1,1 (𝜍) �̅�𝑁𝜍−1,2 (𝜍) 1 0.79 0 3.89 0 5 2 0.79 35.6 3.89 0 6 3 1.56 320 3.89 0 13 Table 4: initial (�̅�0 (𝜍) ) and final (�̅�𝑁𝜍−1 (𝜍) ) states of the system in S, and number of performed iterations (Nς+1) At each step, the algorithm identifies the direction X1 or X2 along which the function G mostly decreases, leading the system into a new state where gp is lower. The algorithm automatically stops in the state in which the minimum value of gp is found. From Table 4, it can be seen that, irrespective of the initial nominal state, the ES leads the operator to the same final state �̅�𝑁𝜍−1 (𝜍) ≡ (3.89,0) ≡ 𝑃𝜗, which corresponds to a local minima of the limit-state function G. Without loss of generality, Table 5 shows the sequence of optimal nominal states for ς = 3, in which the final number of iterations is N3+1=13. p �̅�1 (3) �̅�2 (3) 1 1.56 320 2 2.33 320 3 3.11 320 4 3.89 320 20 5 3.89 284 6 3.89 249 7 3.89 213 8 3.89 178 9 3.89 142 10 3.89 107 11 3.89 71.1 12 3.89 35.6 13 3.89 0 Table 5: sequence of system optimal nominal states 1,,2,1 )( }{    Npp x  for 3 . In Figure 9, it is noteworthy comparing the pressure transient arising from the ES with the one deriving from the application of a PID control system, when the same in-surge accident scenario occurs (e.g., a total water mass entering in the pressurizer from the primary system of 1466 kg, with flow rate of 9 kg/s). Figure 9: Pressure transient in an in-surge accident scenario using the ES (dashed line) and the PID controller (thick solid line) to control the system. The failure boundary is represented by the horizontal thin line. In the first phase of the transient (up to 30 s), the PID system shows a constant pressure of 150 bar, while the ES produces a drop of 0.5 bar. The reason is that the safest position is 147.9 bar, 21 as it can be seen from the data at the end of the transient; thus, the ES works to reach this condition. At t = 30 s, the in-surge accident takes place and it is possible to notice a rise in pressure in both transients, steeper for the PID controller. But the main observation is that in the PID transient the pressure overcomes the failure boundary of 155 bar, whereas the ES manages to keep the pressure well below the failure boundary for the entire duration of the in-surge accident scenario. Lastly, as the accident scenario is recovered, the actions driven by the ES lead to a reduction in pressure, till it reaches its safest optimal working level of 147.9 bar. Finally, it is worth mentioning that the capability of the proposed ES in providing the plant operator with a live, effective and unequivocal AOP to keep the safety margins and avoid system failure only depends on the computational demand of the TH code utilized, and not on the ES algorithm itself: more complex simulation codes, that may become computationally intensive, may challenge the practical scalability of the proposed procedure and endanger the live guidance feature offered by the ES, despite that these could still be utilized for an off-line design of accurate, effective and unequivocal AOPs. 5. CONCLUSIONS In this work, an ES based on regional sensitivity indices (i.e., the Revised Variance Ratio Functions) has been proposed for improving the safety margin of a system when an IE occurs during operation. A RSA is performed to define the strategy for keeping under control the uncertainty affecting the controllable input variables. Such uncertainty has been modeled by centering the nominal value of the input variable on a reduced range of its domain. The ES successfully achieves the stepwise maximization of the system safety margin by acting on the model controllable inputs. The proposed approach has been proved effective on an analytical case study, which mimics the response of a NPP to an accident scenario, and on a case study of a pressurizer exposed to an accidental surge line transient. 22 REFERENCES [Aldemir et al., 2001] T. Aldemir, P. Wang, D.W. Miller. Parameter and State Estimation Using DSD, Transactions of the American Nuclear Society, volume 84; Jun 2001. [Apostolakis, 1990] G. E. Apostolakis. The Concept of Proability in Safety Assessments of Technological Systems. In Science, Vol. 250, pp. 1359-1364. December 1990. [Baraldi et al., 2010] P. Baraldi, A. Cammi, F. Mangili and E. Zio. An ensemble approach to sensor fault detection and signal reconstruction for nuclear system control. In Annals of Nuclear Energy, Vol. 37, pp. 778-790. Elsevier, April 2010. [Baraldi et al., 2013] P. Baraldi, F. Di Maio and E. Zio. Unsupervised Clustering for Fault Diagnosis in Nuclear Power Plant Components. In International Journal of Computational Intelligence Systems, pp. 764-777. Taylor and Francis, Ltd., 2013. [Bolado-Lavin et al., 2009] R. Bolado-Lavin, W. Castaings and S. Tarantola. Contribution to the Sample Mean Plot for Graphical and Numerical Sensitivity Analysis. In Reliability Engineering & System Safety, Vol. 93, pp. 1563-1573. Elsevier, 2009. [Bourinet et al., 2011] J. M. Bourinet, F. Deheeger and M. Lemaire. Assessing Small Failure Probabilities by Combined Subset Simulation and Support Vector Machines. In Structural Safety, Vol. 33, pp. 343-353. Elsevier, June 2011. [Buchanan et al., 1988] B. G. Buchanan and R.G. Smith. Fundamentals of Expert Systems. In Annual Review of Computer Science, Vol. 3, pp. 23-58. Annual Reviews Inc., June 1988. [Cammi et al., 2006] A. Cammi, L. Luzzi, A. A. Porta and M. E. Ricotti. Modelling and Control Strategy of the Italian LBE-XADS. In Progress in Nuclear Energy, Vol. 48, pp. 578- 589. Elsevier, 2006. [Di Maio et al., 2015] F. Di Maio, M. Vagnoli and E. Zio. Risk-Based Clustering for Near Misses Identification in Integrated Deterministic and Probabilistic Safety Analysis. In 23 Science and Technology of Nuclear Installations. Hindawi Publishing Corporation, February 2015. [Di Maio et al., 2016] F. Di Maio, A. Bandini, E. Zio, A. Alfonsi and C. Rabiti. An Approach Based on Support Vector Machines and a K-D Tree Search Algorithm for Identification of the Failure Domain and Safest Operating Conditions in Nuclear Systems. In Progress in Nuclear Energy, Volume 88, pp. 297-309, 2016. [Hakobyan et al., 2008] A. Hakobyan, T. Aldemir, R. Denning, S. Dunagan, D. Kunsman, B. Rutt, U. Catalyurek, "Dynamic generation of accident progression event trees," Nuclear Engineering and Design, 238, 3457-3467 (2008) [Helton et al., 1996] J.C. Helton and D. E. Burmaster. Guest Editorial: Treatment of Aleatory and Epistemic Uncertainty in Performance Assessments for Complex Systems. In Reliability Engineering & System Safety, Vol. 54, pp. 91-94. Elsevier, 1996. [Hines et al., 1998] J.W Hines, R.E. Uhrig, “Use of autoassociative neural networks for signal validation”, Journal of Intelligent and Robotic Systems, vol. 21, pp. 143–154, 1998. [Hsieh et al., 2012] M. H. Hsieh, S. L. Hwang, K. H. Liu, S. F. M. Liang and C. F. Chuang. A Decision Support System for Identifying Abnormal Operating Procedures in a Nuclear Power Plant. In Nuclear Engineering and Design, Vol. 249, pp. 413-418. Elsevier, April 2012. [Hurtado et al., 2004] J. E. Hurtado. Structural Reliability – Statistical Learning Perspectives. Lecture Notes in Applied and Computational Mechanics. Springer, 2004. [IAEA, 2000] VV. AA. IAEA Management of Aging of I&C Equipment in Nuclear Power Plant. IAEA-TECDOC-1147. International Atomic Energy Agency, 2000. [IAEA, 2004] VV. AA. Management Life Cycle and Aging at Nuclear Power Plants: Improved I&C Maintenane. IAEA-TECDOC-1402. International Atomic Energy Agency, 2004. 24 [IAEA, 2005] VV. AA. Effective Corrective Actions to Enhance Operational Safety of Nuclear Installations. IAEA-TECDOC-1458. Internation Atomic Energy Agency, July 2005. [IAEA, 2006] VV. AA. Development and Review of Plant Specific Emergency Operating Procedures. IAEA Safety Reports Series, No. 48. International Atomic Energy Agency, February 2006. [Ikram et al., 2015] A. Ikram and U. Qamar. Developing an Expert System Based on Association Rules and Predicate Logic for Earthquake Prediction. In Knowledge-Based Systems, Vol. 75, pp. 87-103. Elsevier, 2015. [Kalos et al., 2008] M. H. Kalos and P. A. Whitlock. Monte Carlo Methods. Wiley, 2008. [Kim et al., 2001] J. T. Kim, K. C. Kwong, I. K. Hwang, D. Y. Lee, W. M. Park, J. S. Kim and S. J. Lee. Development of Advanced I&C in Nuclear Power Plants: ADIOS and ASICS. In Nuclear Engineering and Design, Vol. 207, pp. 105-119. Elsevier, 2001. [Kim et al., 2007] J. H. Kim and P. H. Seong. The Effect of Information Types on Diagnostic Strategies in the Information Aid. In Reliability Engineering & System Safety, Vol. 92, pp. 171-186. Elsevier, February 2007. [Kwon et al., 1999] K. C. Kwon and J. H. Kim. Accident Identification in Nuclear Power Plants Using Hidden Markov Models. In Engineering Applications of Artificial Intelligence, Vol. 12, pp. 491-501. Elsevier, 1999. [Lee et al., 2007] S. J. Lee, K. Mo, P. H. Seong. Development of an Integrated Decision Support System to Aid the Cognitive Activities of Operators in Main Control Rooms of Nuclear Power Plants. In Proceedings of IEEE Symposyum on Computational Intelligence in Multicriteria Decision Making (MCDM), pp. 146-152. IEEE, 2007. [Liao, 2005] S. H. Lao. Expert System Methodologies and Applications – A Decade Review from 1995 to 2004. In Expert Systems with Applications, Vol. 28, pp. 93-103. Elsevier, 2005. 25 [Lucia et al., 2004] D. J. Lucia, P. S. Beran and W. A. Silva. Reduced-Order Modeling: New Approaches for Computational Physics. In Progress in Aerospace Sciences, Vol. 40, pp. 51-117. Elsevier, February 2004. [Ma et al., 2011] J. Ma and J. Jiang. Applications of Fault Detetion and Diagnosis Methods in Nucelar Power Plants: A Review. In Progress in Nuclear Energy, Vol. 53, pp. 255- 266. Elsevier, January 2011. [McBride et al., 1993] R. D. McBride and D. E. O’Leary. The Use of Mathematical Programming with Artificial Intelligence and Expert Systems. In European Journal of Operational Research, Vol. 70, pp. 1-15. Elsevier, October 1993. [Oberkampf et al., 2004] W. L. Oberkampf, J. C. Helton, C. A. Joslyn, S. F. Wojtkiewicz and S. Ferson. Challenge Problems: Uncertainty in System Response Given Uncertain Parameters. In Reliability Engineering & System Safety, Vol. 85, pp. 11-19. Elsevier, September 2004. [Park et al., 2015] J. Park and W. Jung. A Systematic Framework to Investigate the Coverge of Abnormal Operating Procedures in Nuclear Power Plants. In Reliability Engineering & System Safety, Vol. 138, pp. 21-30. Elsevier, February 2015. [Rubinstein, 1981] R. Y. Rubinstein. Simulation and the Monte Carlo Method. Wiley, 1981. [Tarantola et al., 2012] S. Tarantola, V. Kopustinskas, R. Bolado-Lavin, A. Kaliatka, E. Uspuras and M. Vaisnoras. Sensitivity Analysis Using Contribution to Sample Variance Plot: Application to a Water Hammer Model. In Reliability Engineering & System Safety, Vol. 99, pp. 62-73. Elsevier, March 2012. [Todreas and Kazimi, 1990] N. Todreas, M. S. Kazimi. Nuclear Systems 1, pp. 239- 287. Taylor & Francis, 1990. 26 [Wei et al., 2014] P. Wei, Z. Lu, W. Ruan and J. Song. Regional Sensitivity Analysis Using Revised Mean and Variance Ratio Functions. In Reliability Engineering & System Safety, Vol. 121, pp. 121-135. Elsevier, 2014. [Yoshikawa, 2005] H. Yoshikawa. Human-Machine Interaction in Nuclear Power Plants. In Nuclear Engineering and Technology, Vol. 79, pp. 151-158. Elsevier, 2005. [Zio et al., 2008] E. Zio and F. Di Maio. Bootstrap and Order Statistics for Quantifying Thermal-Hydraulic Code Uncertainties in the Estimation of Safety Margins. In Science and Technology of Nuclear Installations, Vol. 2008, Article ID 340164. Hindawi Publishing Corporation, 2008. [Zio et al., 2010] E. Zio, F. Di Maio and J. Tong. Safety Margins Confidence Estimation for a Passive Residual Heat Removal System. In Reliability Engineering & System Safety, Vol. 95, pp. 828-836. Elsevier, August 2010. [Zio, 2013] E. Zio. The Monte Carlo Simulation Method for System Reliability and Risk Analysis. Springer Series in Reliability Engineering. Springer, 2013. [F. Di Maio et al., 2013] F. Di Maio, P. Baraldi, E. Zio and R. Seraoui. Fault Detection in Nuclear Power Plants Components by a Combination of Statistical Methods. In IEEE Transactions on Reliability, Vol. 62, pp. 833-845. IEEE, October 2013. [S. Al-Dahidi] S. Al-Dahidi, P. Baraldi, F. Di Maio and E. Zio. A Novel Fault Detection System Taking Into Account Uncertainties in the Resconstructed Signals. In Annals of Nuclear Energy, Vol. 73, pp. 131-144. Elsevier, November 2014. 
work_gk74fystujhudgr5qfzyip62xq	----	Evolutionary algorithms based design of multivariable PID controller Expert Systems with Applications 36 (2009) 9159–9167 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Evolutionary algorithms based design of multivariable PID controller M. Willjuice Iruthayarajan *, S. Baskar Department of EEE, Thiagarajar College of Engineering, Madurai 625 015, Tamilnadu, India a r t i c l e i n f o Keywords: PID Control Evolutionary algorithm MIMO system On-line tuning Off-line tuning 0957-4174/$ - see front matter � 2008 Elsevier Ltd. A doi:10.1016/j.eswa.2008.12.033 * Corresponding author. Tel.: +91 94434 87093. E-mail address: willjuice@tce.edu (M.W. Iruthayar a b s t r a c t In this paper, performance comparison of evolutionary algorithms (EAs) such as real coded genetic algo- rithm (RGA), modified particle swarm optimization (MPSO), covariance matrix adaptation evolution strategy (CMAES) and differential evolution (DE) on optimal design of multivariable PID controller design is considered. Decoupled multivariable PI and PID controller structure for Binary distillation column plant described by Wood and Berry, having 2 inputs and 2 outputs is taken. EAs simulations are carried with minimization of IAE as objective using two types of stopping criteria, namely, maximum number of func- tional evaluations (Fevalmax) and Fevalmax along with tolerance of PID parameters and IAE. To compare the performances of various EAs, statistical measures like best, mean, standard deviation of results and average computation time, over 20 independent trials are considered. Results obtained by various EAs are compared with previously reported results using BLT and GA with multi-crossover approach. Results clearly indicate the better performance of CMAES and MPSO designed PI/PID controller on multivariable system. Simulations also reveal that all the four algorithms considered are suitable for off-line tuning of PID controller. However, only CMAES and MPSO algorithms are suitable for on-line tuning of PID due to their better consistency and minimum computation time. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction Proportional-integral-derivative (PID) control offers the sim- plest and yet most efficient solution to many real-world control problems. Three-term functionality of PID controller covers treat- ment of both transient and steady state responses. The popularity of PID control has grown tremendously, since the invention of PID control in 1910 and the Ziegler–Nichol’s straight forward tuning method in 1942. With the advances in digital technology, the sci- ence of automatic control now offers a wide spectrum of choices for control schemes such as adaptive control (Astrom & Witten- mark, 1995), neural network control (Fukuda & Shibata, 1992) and fuzzy logic control (Lee, 1990). However more than 90% of industrial controllers are still implemented based around PID con- trol algorithms, as no other controllers match the simplicity, clear functionality, applicability and ease of use offered by the PID con- trollers (Ang, Chang, & Li, 2005). Several approaches have been reported in literature for tuning the parameters of PID controllers. Ziegler–Nichols and Cohen– Coon are the most commonly used conventional methods for tuning PID controllers and neural network, fuzzy based approach, neuro-fuzzy approach and evolutionary computation techniques are the recent methods (Astrom & Hagglund, 1995). ll rights reserved. ajan). Many researches have already reported the optimal design of PID controller parameters using various EAs such as GA (Chen, Cheng, & Lee, 1995), MPSO (Gaing, 2004; Ghoshal, 2004; Mukher- jee & Ghoshal, 2007; Wang, Zhang, & Wang, 2006) and DE (Bingul, 2004) for SISO system. In general, EAs are robust search and optimization methodology, able to cope with ill-defined problem domain such as multimodality, discontinuity, time-variance, randomness and noise. GA approach for tuning of PID controllers for multi-input multi-output (MIMO) process is also reported (Chang, 2007; Zuo, 1995). In Chang (2007), decoupled multivariable PI controller tuning using GA with multi-parent crossover approach was presented. Simple three-parent differential crossover and uniform mutation operators have been employed. The better performance of three- parent crossover RGA over BLT and traditional two-parent cross- over based RGA was demonstrated in the paper. Recently, several modifications are carried out in crossover and mutation mechanisms of RGA such as SBX crossover, PCX crossover and non-uniform polynomial mutation to improve the perfor- mance of RGA. Self-adaptive simulated binary crossover (SBX) based RGA was successfully applied to various engineering optimi- zation problems (Deb, 2001). SBX crossover is self-adaptive in nat- ure which creates children solutions in proportion to the difference in parent solutions. The near parent solutions are monotonically more likely to be chosen as offspring than solutions distant from parents. mailto:willjuice@tce.edu http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 9160 M.W. Iruthayarajan, S. Baskar / Expert Systems with Applications 36 (2009) 9159–9167 Another EA, namely, covariance matrix adaptation evolution strategy (CMAES) with the ability of learning of correlations be- tween parameters and the use of the correlations to accelerate the convergence of the algorithm is recently proposed. Due to the learning process, the CMAES algorithm performs the search independent of the coordinate system, reliably adapts topologies of arbitrary functions, and significantly improves convergence rate especially on non-separable and/or badly scaled objective func- tions. CMAES algorithm has been successfully applied in varieties of engineering optimization problems (Baskar, Alphones, Sugan- than, Ngo, & Zheng, 2005).This algorithm outperforms all other similar classes of learning algorithms on the benchmark multi- modal functions (Kern et al., 2004). Covariance matrix adaptation evolution strategy algorithm and also recent modifications in other EAs were not applied for the tun- ing of PID controllers. Also, all the reported papers for EA based PID controller design, have considered one or two algorithms for the purpose of comparison. This paper focuses mainly on the performance evaluation of various EAs such as Self-adaptive RGA, MPSO, DE and CMAES on optimum design of multivariable PI and PID controllers for binary distillation column plant described by Wood and Berry (Chang, 2007). The essence of the paper lies in the determination of suit- able EA method for the tuning of PID controller for MIMO system. The remaining part of the paper is organized as follows. Section 2 introduces PID controller structure for SISO and MIMO systems. Section 3 describes the various EAs methods. Section 4 introduces the MIMO system considered for PID controller tuning. Section 5 presents the implementation of EA based multivariable PID con- troller design. Section 6 reveals the simulation results. Finally, con- clusions are given in Section 7. 2. PID controller structure A standard PID controller structure is also known as the ‘‘three- term” controller, whose transfer function is generally written in the ideal form in (1) or in the parallel form in (2) GðsÞ¼ K P 1 þ 1 T Is þ T Ds � � ð1Þ GðsÞ¼ K P þ K I s þ K Ds ð2Þ where KP is the proportional gain, TI is the integral time constant, TD is the derivative time constant, K I ¼ K P=T I is the integral gain and K D ¼ K P T D is the derivative gain. The ‘‘three-term” functionalities are highlighted below. � The proportional term – providing an overall control action pro- portional to the error signal through the all pass gain factor. � The integral term – reducing steady state errors through low fre- quency compensation by an integrator. � The derivative term – improving transient response through high frequency compensation by a differentiator. For optimum performance, KP, KI (or TI) and KD (or TD) are tuned by EAs by minimizing the performance measures such as IAE, ISE and ITAE. Yd + - Multivariable PID controllers K(s) E Fig. 1. A multivariable P 2.1. PID controller for MIMO system Consider a multivariable PID control structure as in Fig. 1, where, desired output vector: Yd = [yd1, yd2, . . ., ydn] T; Actual output vector: Y = [y1, y2, . . ., yn] T; Error vector: E = Yd � Y = [yd1 � y1, yd2 � y2, . . ., ydn � yn]T = [e1, e2, . . ., en] T; Control input vector: U = [u1, u2, . . ., un] T; n � n Multivariable processes: GðsÞ¼ g11ðsÞ � � � g1nðsÞ .. . . . . .. . gn1ðsÞ � � � gnnðsÞ 2 664 3 775 ð3Þ n � n Multivariable PID controller: KðsÞ¼ k11ðsÞ � � � k1nðsÞ .. . . . . .. . kn1ðsÞ � � � knnðsÞ 2 664 3 775 ð4Þ In this work, decoupled multivariable PID controller is consid- ered. So K(s) becomes KðsÞ¼ k1ðsÞ � � � 0 .. . . . . .. . 0 � � � knðsÞ 2 664 3 775 ð5Þ The form of ki(s) is either in (1) or (2). In this work, ‘‘parallel for- m” of PID controller in (2) is used and can be rewritten as kiðsÞ¼ kPi þ kIi s þ kDi s; ð6Þ For convenience, let hi ¼ ½kPi ; kIi ; kDi�, represents the gains vector of ith diagonal sub PID controller in K(s) .For multivariable PI con- troller, ki(s) in (6) can be rewritten as kiðsÞ¼ kPi þ kIi s ð7Þ hi ¼ ½kPi ; kIi� represents the gains vector of the ith diagonal sub PI controller in K(s). 2.2. Performance index In the design of PID controller, the performance criterion or objective function is first defined based on the desired specifica- tions such as time-domain specifications, frequency domain specifications and time-integral performance. The commonly used time-integral performance indexes are integral of the square error (ISE), integral of the absolute value of the error (IAE) and integral of the time-weighted absolute error (ITAE). Minimization of IAE as given in (8) is considered as the objective of this paper IAE ¼ Z 1 0 ðje1ðtÞjþ je2ðtÞjþ � � �þ jenðtÞjÞdt ð8Þ Multivariable Processes G(s) YU ID control system. M.W. Iruthayarajan, S. Baskar / Expert Systems with Applications 36 (2009) 9159–9167 9161 3. Evolutionary algorithms Evolutionary algorithms differ from the traditional optimization techniques in that EAs make use of a population of solutions, not a single point solution. EAs are inherently parallel, because they simultaneously evaluate many points in the parameter space (search space). Considering many points in the search space, EA has a reduced chance of converging to the local optimum and would be more likely to converge to the global optimum. An iter- ation of EA involves a competitive selection that weeds out poor solutions and offspring generation mechanism. Several evolution- ary search algorithms like GA, EP, MPSO, DE and CMAES were developed independently. These algorithms differ in selection, off- spring generation and replacement mechanisms. For solving global functional optimization problems, RGA, MPSO DE and CMAES algo- rithms are normally used and hence these algorithms are em- ployed in this paper. Over the years, several variants of these algorithms were proposed. Some of the recent variants of these algorithms are briefly explained in this section. 3.1. Real coded genetic algorithm with SBX crossover In general, a genetic algorithm has five components as follows: 1. A genetic representation of solutions to the problem. 2. A way to create an initial population of solutions. 3. An evaluation function rating solution in terms of its fitness. 4. Parent selection mechanism and genetic operators that alter the genetic composition of children during reproduction. 5. Values for the parameter of genetic algorithm. Real-number encoding is best used for function optimization problems. It has been widely confirmed that real-number encoding performs better than binary or gray encoding for constrained opti- mization. Owing to the adaptive capability, SBX crossover and polynomial mutation operators are employed. Tournament selec- tion is used as selection mechanism in order to avoid premature convergence. Simulated binary crossover (SBX) and polynomial mutation are briefly explained below. 3.1.1. Simulated binary crossover In SBX crossover (Deb, 2001), two children solutions are created from two parents as follows: Choose a random number ui, e [0, 1] and calculate bqi as given in (9) bqi ¼ ð2uiÞ 1 gcþ1; ui 6 0:5 1 2ð1�uiÞ � � 1 gcþ1 ; otherwise 8< : ð9Þ A spread factor bqi is defined as the ratio of the absolute differ- ence in offspring values to that of the parents. gc is the crossover index. Then compute the offspring xð1;tþ1Þi & x ð2;tþ1Þ i as xð1;tþ1Þi ¼ 0:5 ð1 þ bqiÞx ð1;tÞ i þð1 � bqiÞx ð2;tÞ i h i xð2;tþ1Þi ¼ 0:5 ð1 � bqiÞx ð1;tÞ i þð1 þ bqiÞx ð2;tÞ i h i : ð10Þ 3.1.2. Polynomial mutation Newly generated offspring undergoes polynomial mutation operation. Like in the SBX operator, the probability distribution can also be a polynomial function, instead of a normal distribution. The new offspring yð1;tþ1Þi is determined as follows: yð1;tþ1Þi ¼ x ð1;tþ1Þ i þðx U i � x L i Þdi: ð11Þ xUi and x L i are the upper and lower limit values.where the parameter di is calculated from the polynomial probability distribution PðdÞ¼ 0:5ðgm þ 1Þð1 �jdjÞ gm di ¼ ð2riÞ 1=ðgmþ1Þ � 1; if ri < 0:5 1 �½2ð1 � riÞ� 1=ðgmþ1Þ; if ri P 0:5 ( ð12Þ where gm is the mutation index. Newly generated individuals replace their parents and form the parents for the next generation and the above procedure is re- peated until a maximum number of function evaluations are completed. 3.2. Modified particle swarm optimization (MPSO) Particle swarm optimization provides a population-based search procedure in which individuals called particles, change their positions (states) with time. In a PSO system, particles fly around in a multidimensional search space. During flight, each particle ad- justs its position according to its own experience, and the experi- ence of neighboring particles, making use of the best position encountered by itself and its neighbors. The swarm direction of a particle is defined by the set of particles neighboring the particle and its history of experience. Let Xi and Vi represent ith particle position and its correspond- ing velocity in a search space. The best previous position of ith par- ticle is recorded and represented as Pbesti. The best particle among all the particles in the group is represented as Gbest. The updated velocity of individual i is given in (13) (Shi & Eberhart, 1998) V kþ1i ¼ x kV ki þ c1 rand1 Pbest k i � X k i � � þ c2 rand2 Gbest k � Xki � � ð13Þ where V ki Velocity of ith individual at iteration k xk Inertia weight at iteration k c1, c2 Acceleration factors rand1, rand2 Uniform random numbers between 0 and 1 Xki Position of ith individual at iteration k Pbestki Best position of ith individual at iteration k Gbestk Best position of the group until iteration k Each individual moves from the current position to the next one with the modified velocity as Xkþ1i ¼ X k i þ V kþ1 i ð14Þ The above procedure is repeated until a maximum number of func- tional evaluations have been reached. 3.3. Differential evolution Differential Evolution is a simple population-based, stochastic parallel search evolutionary algorithm for global optimization (Storm & Price, 1997). The initial population is chosen randomly and should cover the entire parameter space. At each generation, DE employs both mutation and crossover (recombination) to pro- duce one trial vector Ui,J+1 for each target vector Xi,J. Then, a selec- tion phase takes place, where each trial vector is compared with the corresponding target vector; the better one will enter the pop- ulation of the next generation. For each target vector Xi,J, a mutant vector is generated using Rand/Rand strategy as Ui;Jþ1 ¼ Xr3;J þ FðXr1;J � Xr2;JÞ ð15Þ 9162 M.W. Iruthayarajan, S. Baskar / Expert Systems with Applications 36 (2009) 9159–9167 where Ui;Jþ1 ith mutated individual for the next iteration X Population set F Mutation constant [0, 2] J Current iteration J+1 Next iteration Xr1,J. . . Xr3,J Randomly selected individuals from the population of the current iteration. To increase the diversity of the perturbed parameter vectors cross- over is performed after mutation V i;Jþ1 ¼ Xi;jð1 � CRÞþ Ui;Jþ1CR ð16Þ where, CR = crossover probability constant from interval [0, 1] The parents for the next iteration are selected as follows: Xi;Jþ1 ¼ V i;Jþ1 if fðV i;Jþ1Þ > fðXi;Jþ1Þ Xi;Jþ1 if fðXi;Jþ1Þ > fðV i;Jþ1Þ � ð17Þ where f(Vi,J+1) is the fitness function value of the ith individual of the population to which the mutation and crossover operators are ap- plied and f(Xi,J+1) is the fitness function value of the ith individual in the original population. The loss of the best individuals in the fol- lowing iteration is avoided by this selection mechanism, as the worst individuals are replaced by the best individuals. This process continues until the maximum function evaluation is reached. 3.4. Covariance matrix adaptation evolution strategy (CMAES) Covariance matrix adaptation evolution strategy is a class of continuous EA that generates new population members by sam- pling from a probability distribution that is constructed during the optimization process. One of the key concepts of this algorithm involves the learning of correlations between parameters and the use of the correlations to accelerate the convergence of the algo- rithm. The adaptation mechanism of CMAES consists of two parts, (1) the adaptation of the covariance matrix C and (2) the adapta- tion of the global step size. The covariance matrix C is adapted by the evolution path and difference vectors between the l best individuals in the current and previous generation. The detailed CMAES algorithm is presented in Kern et al. (2004). 3.4.1. CMAES algorithm Step 1: Generate an initial random solution. Step 2: The offspring at g + 1 generation xgþ1k are sampled from a Gaussian distribution using covariance matrix and global step size at generation g xðgþ1Þk ¼ zk; zk ¼ N hxi ðgÞ l ; r ðgÞ2 CðgÞ � � k ¼ 1; :::; k ð18Þ where hxiðgÞl ¼ Pl i¼1x ðgÞ i with l being the selected best individuals from the population. The parameters cc, ccov, cr and d required for further computa- tions are by default given in terms of the number of decision vari- ables (n) and l as follows: cc ¼ 4 n þ 4 ; cr ¼ 10 n þ 20 ; d ¼ max 1; 3l n þ 10 � � þ cr; ccov ¼ 1 l 2 ðn þ ffiffiffiffiffi 2Þ p 2 þ 1 � 1l � � min 1; 2l � 1 ðn þ 2Þ2 þ l ! ð19Þ The parameters cr and ccov control independently the adaptation time scales for the global step size and the covariance matrix. Note that if l � n, d is large and the change in r is negligible compared to that of C. The initial values are Pð0Þr ¼ P ð0Þ c ¼ 0 and Cð0Þ ¼ I. Step 3: The evolution path Pðgþ1Þc is computed as follows: Pðgþ1Þc ¼ð1 � ccÞ � P ðgÞ c þ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ccð2 � ccÞ p � ffiffiffiffi l p rðgÞ hxiðgþ1Þl �hxi g l � � ð20Þ Cðgþ1Þ ¼ ð1 � ccovÞ � CðgÞ þ ccov � 1 l Pðgþ1Þc P ðgþ1Þ c � �T þ 1 � 1 l � � 1 l Xl i¼1 1 rðgÞ2 xðgþ1Þi �hxi ðgÞ l � � xðgþ1Þi �hxi ðgÞ l � �T� ð21Þ The strategy parameter ccov 2 ½0; 1� determines the rate of change of the covariance matrix C. Step 4: Adaptation of global step size r(g+1) is based on a conju- gate evolution path Pðgþ1Þr Pðgþ1Þr ¼ð1 � crÞ � P ðgÞ r þ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi crð2 � crÞ p � BðgÞðDðgÞÞ�1ðBðgÞÞ�1 � ffiffiffiffi l p rðgÞ hxiðgþ2Þl �hxi g l � � ð22Þ the matrices B(g) and D(g) are obtained through a principal compo- nent analysis: CðgÞ ¼ BðgÞðDðgÞÞ2ðBðgÞÞT ð23Þ where the columns of B(g) are the normalized eigen vectors of C(g) and D(g) is the diagonal matrix of the square roots of the given eigen values of C(g). The global step size r(g+1) is determined by rðgþ1Þ ¼ rðgÞ exp cr d kPðgþ1Þr k EðkNð0; IÞkÞ ! � 1 ! ð24Þ Step 5: Repeat Steps 2–4 until the stopping criteria are satisfied. 4. MIMO system Most of the industrial processes belong to the category of MIMO system, which requires MIMO control techniques to improve per- formance, even though they are naturally more difficult to exploit than SISO system. Binary distillation column plant described by Wood and Berry (Chang, 2007; Wang, Zou, Lee, & Qiang, 1997) is considered. The transfer function of the above process is given in (25) GðsÞ¼ 12:8e�s 1þ16:7s �18:9e�3s 1þ21s 6:6e�7s 1þ10:9s �19:4e�3s 1þ14:4s " # ð25Þ The transfer function concerned with multivariable process has first order dynamics and significant time delays and it has a strong interaction between inputs and outputs. In this paper, multivari- able controller with PI and PID structures are used for optimizing the IAE performance for set point regulation using EAs. 5. EA implementation of multivariable PID controller System described by (25), has two inputs and two outputs. The decoupled multivariable PID controller K(s) for this system is given in (26) KðsÞ¼ k1ðsÞ 0 0 k2ðsÞ � ð26Þ In order to obtain the optimum performance, the parameters of K(s) i.e., ½h1; h2� ¼ ½kP1 ; kI1 ; kD1 ; kP2 ; kI2 ; kD2� are optimized by optimi- zation algorithms. The chromosome/particle representation is gi- ven in Fig. 2. First three elements are the parameters of k1(s) and next three elements are the parameters of k2(s). For multivariable PI control- ler, the structure of the chromosome is given in Fig. 3. First two 1P k 1I k 1D k 2P k 2I k 2D k Fig. 2. Chromosome/particle of multivariable PID controller. 1P k 1I k 2P k 2I k Fig. 3. Chromosome/particle of multivariable PI controller. Table 1 Parameter selection. Evolutionary algorithms Parameter RGA Pc = 0.8 Pm = 1/n gc = 5 gm = 20 MPSO C1 = 1 C2 = 1 Vmax = 0.1 x is linearly decreased from 0.9 to 0.2 over the iterations DE CR = 0.5 F = 0.8 CMAES Parameters are fixed automatically by the algorithm Table 2 Optimum parameters of multivariable PI controller – without tolerance. Method Optimum parameters of multivariable PI controller IAE kP1 kI1 kP2 kI2 BLT* 0.3750 0.0452 �0.0750 �0.0032 23.5568 RGA-multi-crossover* 0.9971 0.0031 �0.0141 �0.0071 10.5778 RGA-SBX crossover 0.8433 0.0026 �0.0127 �0.0069 10.4395 MPSO 0.8485 0.0026 �0.0132 �0.0069 10.4378 DE 0.8485 0.0026 �0.0132 �0.0069 10.4378 M.W. Iruthayarajan, S. Baskar / Expert Systems with Applications 36 (2009) 9159–9167 9163 elements are the parameters of k1(s) and the next two elements are the parameters of k2(s). All the elements of chromosomes/particles of population are randomly initialized within the search space specified by their lower and upper bounds of individual parameters as given in (Chang, 2007). The inequality conditions for the parameter ranges of PI and PID controllers are given in (27) and (28), respectively �1 6 K P1 6 1; �1 6 K I1 6 1; �1 6 K P2 6 1; �1 6 K I2 6 1 ð27Þ �1 6 K P1 6 1; �1 6 K I1 6 1; �1 6 K D1 6 1; �1 6 K P2 6 1; �1 6 K I2 6 1; �1 6 K D2 6 1: ð28Þ CMAES 0.8485 0.0026 �0.0132 �0.0069 10.4378 * Data taken from Chang (2007). Table 3 Optimum parameters of multivariable PI controller – with tolerance. Method Optimum parameters of multivariable PI controller IAE kP1 kI1 kP2 kI2 BLT* 0.3750 0.0452 �0.0750 �0.0032 23.5568 RGA-multi-crossover* 0.9971 0.0031 �0.0141 �0.0071 10.5778 RGA-SBX crossover 0.8622 0.0026 �0.0135 �0.0069 10.44 MPSO 0.8485 0.0026 �0.0132 �0.0069 10.4378 DE 0.8485 0.0026 �0.0132 �0.0069 10.4379 CMAES 0.8485 0.0026 �0.0132 �0.0069 10.4378 * Data taken from Chang (2007). Table 4 Statistical performance of EAs of multivariable PI controller – without tolerance. Method Best value Mean value Standard deviation Average computation time (s) RGA-SBX crossover 10.4395 10.9366 1.6638 137.8698 MPSO 10.4378 10.8157 1.6570 135.2763 DE 10.4378 10.4379 4.5272E-5 137.9376 CMAES 10.4378 10.4378 0 142.5264 Table 5 Statistical performance of EAs of multivariable PI controller – tolerance. Method Best value Mean value Standard deviation Average computation time (s) Average functional evaluations RGA-SBX crossover 10.44 10.4961 0.0810 134.8504 5840 MPSO 10.4378 10.5394 0.3410 110.3138 4805 DE 10.4379 11.2228 1.5740 107.3965 4676 CMAES 10.4378 10.4378 0 84.8255 3572 6. Simulation results In this work, two experiments namely design of multivariable PI and PID controller for binary distillation column plant described by Wood and Berry, using various EAs such as RGA, MPSO, DE and CMAES are conducted. For simulating Binary distillation column plant, MATLAB-SIMULINK software is employed. Simulations are carried out using Core 2 Duo Processor 2.2 GHz, 2GB RAM PC. IAE is determined for step response over 150 min time period. EAs sim- ulations are carried out using two types of stopping criteria namely, Fevalmax, Fevalmax with tolerance of design variables and tolerance of objective function. In this work, the Fevalmax is set at 6000 functional evaluations and tolerance is fixed as 10�5 for the last 20 generations. Owing to the randomness of the EAs, many trials with independent population initializations should be made to acquire a useful conclusion of the performance of the algorithm. Hence, best, mean, standard deviation of IAE measure and average computation in 20 independent trials of various EAs are reported and compared with the already reported results (Chang, 2007). 6.1. Parameter tuning Optimal parameter combinations for different EA methods are experimentally determined by conducting a series of experiments with different parameter settings before conducting actual runs to collect the results. The parameters actually used in the simulations are summarized in the Table 1. 6.2. Tuning of multivariable PI controller Best PI parameters and the corresponding IAE values for the 20 trials of multivariable PI controllers using different EAs with and without tolerance in stopping criteria are reported in Tables 2 and 3, respectively. For the purpose of comparison, already re- ported values obtained by conventional BLT method and GA with multi-crossover approach are directly taken from (Chang, 2007) and given in the Tables. 9164 M.W. Iruthayarajan, S. Baskar / Expert Systems with Applications 36 (2009) 9159–9167 Almost all the EAs are giving equal performance with respect to the best IAE. The statistical performances such as best, mean, stan- dard deviation of IAE and average computation time (ACT) of 20 trials using various EAs with and without considering tolerance are given in Tables 4 and 5, respectively. Without considering tol- erance in stopping criteria, CMAES and DE are the same with re- spect to mean value but CMAES is computationally slightly expensive due to complex mathematical manipulations. While 0 20 40 60 80 1 10 1 10 2 10 3 Generation IAE Fig. 4. Convergence characteristics of 0 50 100 150 200 0.7 0.8 0.9 1 Iteration K p1 1 0 50 100 150 200 -0.2 0 0.2 0.4 0.6 Iteration K p2 2 Fig. 5. Convergence characteristics considering tolerance, CMAES algorithm gives better performance in consistently achieving good results as compared to all other algorithms and also it requires less number of average functional evaluations (AFeval). MPSO algorithm gives better performance than DE and RGA. Fig. 4 shows the convergence characteristics of various EAs con- sidering tolerance. Higher value of IAE during the initial genera- tions/iterations indicates unstable PID control settings. For 00 120 140 160 180 200 / Iteration RGA-SBX MPSO DE CMAES EAs – multivariable PI controller. 0 50 100 150 200 -0.5 0 0.5 1 Iteration K i1 1 0 50 100 150 200 -0.15 -0.1 -0.05 0 0.05 Iteration K i2 2 of PID parameters using MPSO. 0 50 100 150 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Time in min O ut pu t re sp on se y 1 BLT RGA-MC PI-CMAES Fig. 6. Output response y1. 0 50 100 150 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Time in min O ut pu t re sp on se y 2 BLT RGA-MC PI-CMAES Fig. 7. Output response y2. M.W. Iruthayarajan, S. Baskar / Expert Systems with Applications 36 (2009) 9159–9167 9165 convenience, IAE for the unstable response is limited to 1000. Other than CMAES algorithm, the convergence characteristics of all algorithms are smooth. Due to the self-learning behavior of CMAES algorithm, convergence characteristics show large varia- tions during the initial search. Fig. 5 shows the convergence char- acteristics of multivariable PI parameters obtained by MPSO algorithm with tolerance in stopping criteria. Output responses y1 and y2 of the MIMO system with multivar- iable PI controller using best PI parameter obtained out of the 20 trials by using CMAES/MPSO algorithm is shown in Figs. 6 and 7, respectively. For comparison purpose, output responses using RGA with multi-crossover approach are also given in the same fig- ure. PI controller simulation results show the better performance of CMAES and MPSO algorithms as compared to the previously re- ported results obtained using GA with multi-crossover approach (Chang, 2007) and conventional BLT method. Time-responses spec- ifications for the designed PI controllers such as peak overshoot in%, rise time (min) and settling time for ±5% tolerance (min) for output responses y1 and y2 are summarized in Table 6. Peak over- shoot of system response with multivariable PI controller designed by CMAES/MPSO is approximately 50% lesser than GA with multi- crossover approach with a slight increase in rise time and settling time. 6.3. Tuning of multivariable PID controller Best PID parameters and the corresponding IAE values for the 20 trials of multivariable PID controllers using various EAs with and without considering tolerance in stopping criteria are reported in Tables 7 and 8, respectively. The CMAES and MPSO algorithms are almost giving equal performance with respect to the best IAE. The statistical performances such as best, mean, standard deviation of IAE, ACT and AFeval of 20 trials using various EAs with and with- out considering tolerance are given in Tables 9 and 10, respec- tively. CMAES, MPSO and RGA algorithms give better performance in consistently achieving good results as compared to DE. But AFeval required and ACT for RGA is larger than CMAES and MPSO. For clarity, transient portion (25 min) of output responses y1 and y2 of the MIMO system with multivariable PID controller using best PID parameter obtained using CMAES/MPSO algorithm is shown in Figs. 8 and 9, respectively. For comparison purpose, out- put responses using RGA with multi-crossover approach and mul- tivariable PI controller designed by CMAES are also given in the same figure. PID controller simulation results show the better per- formance of CMAES and MPSO algorithms as compared to the pre- Table 6 Time-response specifications of output responses – multivariable PI controller. Method y1 %Mp Rise time (min) Settling tim BLT 27.4365 4.9 32.2 RGA with multi-crossover 21.9866 2.5 7.1 CMAES 10.8390 2.8 7.3 Table 7 Optimum parameters of multivariable PID controller – without tolerance. Method Optimum PID parameters kP1 kI1 kD1 RGA-SBX crossover 1.0 0.0025 0.3892 PID-MPSO 1.0 0.0025 0.3872 DE 0.9945 0.0026 0.4021 CMAES 1.0 0.0025 0.3872 viously reported results obtained using GA with multi-crossover approach (Chang, 2007) and multivariable PI controller designed y2 e 5% (min) %Mp Rise time (min) Settling time 5% (min) 30.9972 9.2 88.2 17.8161 8.5 14.9 9.2414 8.9 10.6 IAE kP2 kI2 kD2 �0.0317 �0.0072 �0.0885 9.6859 �0.0332 �0.0073 �0.0909 9.6824 �0.0289 �0.0071 �0.0709 9.7108 �0.0332 �0.0073 �0.0909 9.6824 Table 8 Optimum parameters of multivariable PID controller – with tolerance. Method Optimum PID parameters IAE kP1 kI1 kD1 kP2 kI2 kD2 RGA-SBX crossover 1.0 0.0025 0.3860 �0.0358 �0.0073 �0.0991 9.6871 PID-MPSO 1.0 0.0025 0.3872 �0.0332 �0.0073 �0.0909 9.6824 DE 0.9825 0.0026 0.4239 �0.0563 �0.0082 �0.1563 9.9251 CMAES 1.0 0.0025 0.3872 �0.0332 �0.0073 �0.0909 9.6824 Table 9 Statistical performance of EAs of multivariable PID controller – without tolerance. Method Best value Mean value Standard deviation Average computation time RGA-SBX crossover 9.6859 10.2370 0.6968 136.6054 MPSO 9.6824 10.0619 0.4304 136.6904 DE 9.7479 9.8855 0.1135 138.9478 CMAES 9.6824 10.4451 0.7116 146.4358 Table 10 Statistical performance of EAs of multivariable PID controller – with tolerance. Method Best value Mean value Standard deviation Average computation time (s) Average functional evaluations RGA-SBX crossover 9.6871 10.0954 0.3128 139.4493 5943 MPSO 9.6824 10.1992 0.6073 133.0369 5705 DE 9.9251 46.4502 85.6846 71.8283 3075 CMAES 9.6824 10.3186 0.7618 138.6396 5677 0 5 10 15 20 25 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Time in min O ut pu t re sp on se y 2 RGA-MC PI-CMAES PID-CMAES Fig. 9. Transient portion of output response y2. 0 5 10 15 20 25 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Time in min O ut pu t re sp on se y 1 RGA-MC PI-CMAES PID-CMAES Fig. 8. Transient portion of output response y1. Table 11 Time-response specifications of output responses – multivariable PID controller. Method y1 y2 %Mp Rise time (min) Settling time 5% (min) %Mp Rise time (min) Settling time 5% (min) RGA-SBX 0.9602 2.8 12 11.9161 8.3 10.9 DE 2.1241 5.1 12.3 20.7696 8.1 11.7 CMAES/MPSO 0.8059 2.9 11.9 10.4336 8.4 10.6 9166 M.W. Iruthayarajan, S. Baskar / Expert Systems with Applications 36 (2009) 9159–9167 by CMAES. Time responses of y1 and y2 using best PID parameters using various EAs are summarized in Table 11. Results show the better performance of CMAES/MPSO in terms of very less peak overshoot and settling time of response y1 and y2 with slight in- crease in rise time. 7. Conclusions In this paper, performance evaluation of EAs such as RGA with SBX crossover, MPSO, DE and CMAES on the optimal design of mul- tivariable PI and PID controller for the binary distillation column plant is conducted. EAs simulations are carried out using two types of stopping criteria, namely, Fevalmax and Fevalmax along with tolerance of PID parameters and IAE. Multivariable PI/PID control- lers are designed by minimizing IAE and the results are compared with those of the already reported in literature namely, BLT and GA with multi-crossover approach. Simulation results are summarized as follows: (i) The better performance of evolutionary designed multivari- able PI controller over the already reported results. Also, multivariable PID controllers designed for the same system by various EAs are better than multivariable PI controller. (ii) Without tolerance, the best IAE obtained in 20 independent trials by all EAs is almost equal. This reveals that all EAs are equally applicable to off-line PID controller tuning. (iii) Considering the tolerance of PID parameters and IAE, CMAES and MPSO algorithms are more suitable for on-line tuning of PID controller due to their better consistency and minimum computation time. Also, MPSO is much more suitable for on- line tuning of PID controller due to the reduced computation time. M.W. Iruthayarajan, S. Baskar / Expert Systems with Applications 36 (2009) 9159–9167 9167 Acknowledgments The authors gratefully acknowledge the Management of the Thiagarajar College of Engineering, Madurai 625 015, Tamilnadu, India, for their continued support, encouragement, and the exten- sive facilities provided to carry out this research work. They also gratefully acknowledge the support of Dr. M. Chidambaram, Direc- tor, NIT, Trichy. References Ang, K. H., Chang, G., & Li, Yun (2005). PID control system analysis, design and technology. IEEE Transaction on Control System Technology, 13(4), 559–577. Astrom, K. J., & Wittenmark, B. (1995). Adaptive control (2nd ed.). Addison Wesley. Astrom, K. J., & Hagglund, T. (1995). PID controllers: theory, design, and tuning (2nd ed.). Instrument society of America. Baskar, S., Alphones, A., Suganthan, P. N., Ngo, N. Q., & Zheng, R. T. (2005). Design of optimal length low-dispersion FBG filter using covariance matrix adapted evolution. IEEE Photonics Technology Letters, 17(10), 2119–2121. Bingul, Z. (2004). A new PID tuning technique using differential evolution for unstable and integrating processes with time delay. ICONIP, Proceedings Lecture Notes in Computer Science, 3316, 254–260. Chang, W. D. (2007). A multi-crossover genetic approach to multivariable PID controllers tuning. Expert Systems with Applications, 33, 620–626. Chen, B. S., Cheng, Y. M., & Lee, C. H. (1995). A genetic approach to mixed H2/H1 Optimal PID control. IEEE Control Systems, 15(5), 51–60. Deb, K. (2001). Multiobjective optimization using evolutionary algorithms. Chichester, UK: Wiley. Fukuda, T., & Shibata, T. (1992). Theory and application of neural networks for industrial control systems. IEEE Transactions on Industrial Electronics, 39(6), 472–489. Gaing, Z. L. (2004). A particle swarm optimization approach for optimum design of PID controller in AVR system. IEEE Transactions on Energy Conversion, 19(2), 384–391. Ghoshal, S. P. (2004). Optimizations of PID gains by particle swarm optimizations in fuzzy based automatic generation control. Electric Power Systems Research, 72, 203–212. Kern, S., Müller, S. D., Hansen, N., Büche, D., Ocenasek, J., & Koumoutsakos, P. (2004). Learning probability distributions in continuous evolutionary algorithms – A comparative review. Natural Computation, 3(1), 77–112. Lee, C. C. (1990). Fuzzy logic in control systems: Fuzzy logic controller – Part I and II. IEEE Transactions on Systems Man and Cybernetics, 20(2), 404–435. Mukherjee, V., & Ghoshal, S. P. (2007). Intelligent particle swarm optimized fuzzy PID controller for AVR system. Electric Power Systems Research, 77(12), 1689–1698. Shi, Y., & Eberhart, R. C. (1998). A modified particle swarm optimizer. Proceedings of IEEE International Conference on Evolutionary Computation, 69–73. Anchorage, AK. Storm, R., & Price, K. (1997). Differential evolution – a simple and efficient heuristic for global optimization over continuous spaces. Journal of Global Optimization, 11, 341–359. Wang, J. S., Zhang, Y., & Wang, W. (2006). Optimal design of PI/PD controller for non-minimum phase system. Transactions of the Institute of Measurement and Control, 28(1), 27–35. Wang, Q. G., Zou, B., Lee, T. H., & Qiang, B. (1997). Auto-tuning of multivariable PID controllers from decentralized relay feedback. Automatica, 33(3), 319–330. Zuo, W. (1995). Multivariable adaptive control for a space station using genetic algorithms. IEE Proceedings – Control Theory and Applications, 142(2), 81–87. Evolutionary algorithms based design of multivariable PID controller Introduction PID controller structure PID controller for MIMO system Performance index Evolutionary algorithms Real coded genetic algorithm with SBX crossover Simulated binary crossover Polynomial mutation Modified particle swarm optimization (MPSO) Differential evolution Covariance matrix adaptation evolution strategy (CMAES) CMAES algorithm MIMO system EA implementation of multivariable PID controller Simulation results Parameter tuning Tuning of multivariable PI controller Tuning of multivariable PID controller Conclusions Acknowledgments References 
work_gkwile66brdu3haw7daczalhtm	----	Enriching Conflict Resolution Environments with the Provision of Context Information Davide Carneiro, Marco Gomes, Ângelo Costa, Paulo Novais, José Neves CCTC/Department of Informatics University of Minho, Braga, Portugal {dcarneiro, pjon, acosta, jneves}@di.uminho.pt, pg18373@alunos.uminho.pt Abstract It is a common affair to settle disputes out of courts nowadays, through negotiation, mediation or any other mean. This has also been implemented over telecommunication means under the so-called Online Dispute Resolution methods. However, this new technology-supported approach is impersonal and cold, leaving aside important issues such as the disputants’ body language, stress level or emotional response while being based on forms, e-mails or chat rooms. To overcome this shortcoming in this paper it is proposed the creation of intelligent environments for conflict resolution that can complement the existing tools with important knowledge about the context of interaction. This will allow decision- makers to take better framed decisions based not only on figures but also on important contextual information, similarly to what happens when parties communicate in the physical presence of each other. Keywords: Affective Computing, Role Playing Games, Emotions, Ambient Intelligence, Conflict Resolution Environments. 1 Introduction The increase in the transaction volume of global B2C e-Commerce led to a whole new way of doing commerce globally. Now, we talk of electronic contracting performed in part or wholly by means of electronic agents. However, disputes are still likely to arise in these transactions, namely because of late shipments or products of low quality. Courts, that were shaped after the industrial era and are still paper-based, are not ready for both the amount and the new characteristics of these disputes. The immediate consequence is an increase in the waiting queues of courts, rendering judicial systems slow and unresponsive. In a first attempt to address this problem, several alternatives to litigation in courts started to be adopted in the last decades – the so called Alternative Dispute Resolution (ADR) methods, including negotiation, mediation or arbitration (Brown and Marriot 1999). With the advent of the Information Society, these mailto:jneves%7d@di.uminho.pt 2 techniques started to be implemented in virtual environments as well, leading to what is today known as Online Dispute Resolution (ODR) (Katsch and Rifkin 2001). In its most basic form, Online Dispute Resolution simply implements already traditional methods over a communication mean, i.e., instead of negotiating in person, the disputant parties do it over a phone line or a similar mean. However, the latest research trends show that the role of technology in dispute resolution can be further enhanced, namely by using techniques from Artificial Intelligence (Lodder and Thiessen 2003). In such ODR systems, technology is used not only to put parties into contact but also to suggest solutions, plan strategies or compile useful information. However, this approach still has some flaws. Namely, when parties use an online tool, a significant amount of important information is lost. This information includes body language, context information or emotional state. All this would be taken into account by a judge or a jury in a litigation in court in a decision making process (Damásio 1994), even if in an unconscious way, but is lost when using an ODR tool. Its main usefulness is in determining how each aspect of the dispute resolution process affects each party (e.g. is the party happy with the proposed solution?, does the party fill threatened when talking to the other party?, is the party nervous when addressing a specific issue?). The lack of contextual information that would allow parties to take better decisions is thus due to the low richness of the media used to communicate, mostly text-based. In fact, it results difficult to convey emotions and other aspects of our rich communication processes using text only, something that we do intuitively and unconsciously when we are face-to-face. The development of ODR systems that are indeed able to perceive the state of the parties is thus of the utmost importance (Friedman and George 1997). The challenge addressed in this paper is thus the one of acquiring the necessary information to perceive this state. The use of soft computing techniques can be the answer to this challenge. Indeed, this field has been used recently to address a wide range of different real- world problems (Sedano et al. 2010; Corchado et al. 2010; Corchado et al. 2012; Zargari et al. 2012). Therefore, in this paper, a novel approach on ODR research is presented, merging different concepts: Ambient Intelligence (AmI), Role-Playing Games (RPG), Emotions, Conflict Handling Styles and Stress. Our objective is to build a knowledge environment that can be used by decision-makers to take more realistic and better framed decisions in a legal context. Ambient Intelligence is a recent technological paradigm in which traditional environments are empowered with the objective of providing useful context-aware services (Aarts and Grotenhuis 2011). This paradigm establishes the framework and the key ideas for a dynamic knowledge environment supporting the integration of the different concepts described next. Role-Playing Games are a type of game in which players “interpret” a character created in a given scenario (environment), allowing to create “social laboratories” (Nick 2003; Barreteau et al. 2003). To represent emotions the models proposed by William James (James 1884) and Carl Lange (Lange 1967) will be followed. This 3 theory suggests that emotions are the result of a response of human physiological external stimuli, i.e., every emotion is associated with a different physiological response (Cannon 1927). Conflict handling styles depict the way that each one of us reacts before a conflict scenario in terms of assertiveness and cooperativeness. Finally, stress has also a significant influence on our activities and, while a controlled level may even be positive in the generation of ideas and taking of decisions, an increased level of stress may jeopardize the whole process and damage future relationships. The approach presented in this paper develops around the idea of an intelligent environment for supporting conflict resolution, in which the evaluation of the context of interaction of the parties is an important step for efficiently achieving a mutually satisfactory outcome. Our vision is that parties can be in their own environments, using an ODR tool to solve a conflict, with the respective environments collecting important information and sharing it with the dispute resolution tool. To materialize it, this work brings together two projects: VirtualECare and UMCourt. 2 Bridging Ambient Intelligence and Online Dispute Resolution UMCourt and VirtualECare are two research projects maintained by the Intelligent Systems Lab, at the University of Minho. The VirtualECare project (Costa et al. 2008; Costa et al. 2009) focuses on the Ambient Intelligence paradigm and has as main objective to develop an agent-based environment able to monitor, interact and provide its customers with services to support daily living. Moreover, this project also focuses on the acquisition of data for supporting high-level decision making. In that sense, the system is able to read environmental and contextual information from its user’s environment, including environmental conditions, emotional state or physiological state. VirtualECare’s architecture (Figure 1) is a distributed one, bringing together two technologies: OSGi and Jade (Open Services Gateway Initiative, 2003; Bellifemine et al. 2007). 4 Fig. 1. A simplified view of the VirtualECare architecture. OSGi services are used to connect software agents and devices, hiding their singularities and allowing their functionalities to be accessed as standard services. Jade agents, on the other hand, are in charge of all the analysis and decisions making processes. VirtualECare also implements an OSGi-based fully functional simulation platform that allows for the creation and study of specific scenarios. All the data generated, either by an actual implementation of the system or by the simulation platform, is available to external applications by means of OSGi services. UMCourt, on the other hand, is a conflict resolution platform being developed under the TIARAC funded project that aims at the development of a multi-faceted agent-based ODR architecture, suited to be used in different legal domains (Andrade et al. 2010). The main objective is to empower the role of the parties in the dispute resolution process by providing meaningful and contextualized information, proposing solutions, strategies and guidance by means of mediation and negotiation algorithms. Similarly to VirtualECare, this architecture (Figure 2) is also based on OSGi and Jade. This allowed these two projects to come together, laying the path to a new research direction: Dispute Resolution Environments. A detailed description of the architecture and the implemented services is given in (Carneiro et al. 2010b; Carneiro et al. 2009). 5 Fig. 2. The organization of the agents that make up the UMCourt architecture. 3 Intelligent Environments Supporting Conflict Resolution Intelligent Environments can improve conflict resolution processes in many different ways. Specifically, different ways in which IEs can be used to provide different types of context information have been explored, to enrich the knowledge available for decision-making. Some of the approaches presented in this section are working implementations while others are simulations that allow us to improve the models used before their actual implementations. Two different high-level trends are presented in this paper. On the one hand two solutions that are already implemented and working as modules of the UMCourt conflict resolution platform will be detailed. These modules are described in subsections 3.1 and 3.2. The first classifies the conflict resolution style of the parties in real time by analyzing the proposals exchanges against boundary values defined in legal documents. The second assesses the level of stress of users based on their interaction patterns with the devices used as interfaces. Both are performed in a non-intrusive way and provide very important knowledge about the context of interaction to the mediator and to the conflict resolution platform itself. In the second trend the VirtualECare simulation platform is being used to generate knowledge that can be important for the decision-making processes. Simulation is paramount when it comes to create critical real-world scenarios where the margin of error must be minimal. The VirtualECare simulation tool is used to generate information about the parties, their context and their emotional state. The interest lies on the study of the effect of this information on the automated decision-making processes of UMCourt and assess their actual value for a human mediator. The tool allows simulating one instance of an intelligent environment setting, fully configurable in terms of devices, rooms, physical properties, user’s actions, internal and external atmospheric conditions, among others. This means that the tool allows simulating specific sensors such as temperature or humidity sensors but also vital sign ones. Given the scope of this work, focus will be on the simulation of the emotional state of the parties as an input for UMCourt. Our main aim is that either UMCourt or the mediator can take better decisions by encompassing knowledge about the emotional state of the parties, allowing to adapt and fine-tune strategies in real time. There is also interest in the possibility of creating user groups, based on role-playing games, in which each player can embody a certain character, with a given role and permissions (e.g. mediator, plaintiff, defendant, neutral). This is very important in a legal context since each person has different objectives (which have influence over parameters such as the conflict resolution style or the level of stress). Our objective in the long term is to replace these simulated sensors by real ones, in order to implement an actual 6 dispute resolution environment able to compile all this important information, similarly to what has already been done with the stress estimation and the conflict resolution style classification. In subsections 3.3 and 3.4 two components of the simulation platform are detailed that allow simulating important user’s parameters: the personality traits and the emotions. The simulation of these parameters is important to train and assess the behavior of the conflict resolution platform, allowing to understand how it would behave if actual information about the real users was being used. 3.1 Classification of the Conflict Resolution Style In alternative conflict resolution processes in which humans have a preponderant role, specifically in negotiation and mediation, the style of dealing with the conflict of each party will certainly influence the course of action and, consequently, the outcome. In this line of work focus is placed on developing methods for classifying the conflict resolution style of a person in a non-intrusive way, i.e., without using the traditional self-report instruments (such as questionnaires) with all its known disadvantages and inaccuracies. Kenneth Thomas and Ralph Kilmann formalized the way we respond to conflict situations into five different modes in terms of individual’s assertiveness and cooperativeness (Thomas and Kilmann 1974). The authors defined five different conflict handling styles: competing, accommodating, avoiding, collaborating and compromising. While the authors used a questionnaire that parties should fill in order to determine their conflict handling style, a different approach was followed. The interactions between the parties are analyzed in real time in order to estimate their style by analyzing the proposals exchanged while negotiating. This analysis is performed based on economical concepts, adapted to the legal domain. Specifically, to correctly analyze each proposal, we need to be aware of the BATNA and WATNA (respectively Best and Worst Alternative to a Negotiated Agreement) (De Vries et al. 20) which indicate the best and worst scenarios possible by law. The ZOPA – Zone of Potential Agreement is also taken into account (Raiffa 1982), as well as the the MLATNA – Most Likely Alternative to a Negotiated Agreement (describing the most likely scenario from a legal point of view) (Steenbergen 2005). 7 Fig. 3. The space that defines the personal conflict styles in function of the utility of the proposals and the values of the BATNA, BATNA and ZOPA. Generally, a negotiated process goes on with the parties exchanging successive proposals until they agree on a specific one or someone exits the process. The conflict handling style can be estimated by analyzing the proposals exchanged according to the intentions and objectives of the parties and also according to the space defined by the BATNA and WATNA of each party (Figure 3). In each round, each action of a party will contribute to the overall characterization of his conflict handling style. Thus, the conflict handling style is a continuous process, not defined by a single interaction but by a continuous analysis of the evolution of the interactions. In each round, two main paths are possible: either the party ignores the proposal or the party answers to the proposal. When a party ignores a proposal, he is exhibiting an Avoiding behavior given that he is satisfying neither his interests nor the ones of the other party. On the other hand, when the party makes a proposal or a counterproposal, he is cooperating on the process. However, in order to be more precise, the nature of the proposal must be analyzed. Our approach consists in analyzing each proposal in terms of its utility for each party. This analysis is subject to a set of principles. If the utility of the proposal is higher than the BATNA of the other party, he is showing a Competing style. In fact, the party is trying to maximize his own gain, probably in an unrealistically way, completely disregarding the other party. If the utility of the proposal is lower than the WATNA of the other party, the party is neglecting his own gain or even maximizing the gain of the other party. This may happen when a party is not aware of his chances, is facing a party with a relationship of power or just wants to end the process quickly. In such a scenario, the party is assumed to be exhibiting an Accommodating behavior. 8 When the utility of the proposal falls within the range of the ZOPA, it indicates that the party is being reasonable and trying to propose a settlement in which both parties will not win everything but will not lose everything either. In such a scenario, the conflict style is determined according to the distance to the average point of the ZOPA, as defined in equation 1. 𝛽 = � 𝑍𝑂𝑃𝐴𝑀𝐼𝑁 + 𝑍𝑂𝑃𝐴𝑀𝐴𝑋 2 � (1) Two points are defined that allow classifying the remaining conflict styles, as depicted in equations 2 and 3. 𝛼 = �𝑍𝑂𝑃𝐴𝑀𝐼𝑁 + 𝛽 − 𝑍𝑂𝑃𝐴𝑀𝐼𝑁 2 � = ( 𝑍𝑂𝑃𝐴𝑀𝐼𝑁 + 𝛽 2 ) (2) 𝛾 = �𝑍𝑂𝑃𝐴𝑀𝐴𝑋 − 𝑍𝑂𝑃𝐴𝑀𝐴𝑋 − 𝛽 2 � = ( 𝑍𝑂𝑃𝐴𝑀𝐴𝑋 + 𝛽 2 ) (3) When the utility of a proposal falls within the range defined by [𝛼, 𝛾], the proposing party is negotiating in an intermediary area of the ZOPA. This denotes that the party is trying to reach a compromise with potential losses in both sides. The party is thus evidencing a Compromising behavior. On the other hand, if the value of the utility is in the range defined by [𝑍𝑂𝑃𝐴𝑀𝐼𝑁, 𝛼[ ∪ ]𝛾, 𝑍𝑂𝑃𝐴𝑀𝐴𝑋], the party is proposing a solution closer to the limits of the ZOPA. This may indicate a party that is trying to work out a mutually agreeable solution, but trying to explore the weaknesses of the opposing party, trying to force him to accept it using some position of power. The conflict style of the party is this way defined as Collaborating. However, it is common that people do not make use of a single conflict style at a time. Moreover, people often change their conflict style according to the evolution of the process. In that sense, a more precise approach is proposed in which a main conflict style is inferred, together with a trend style. This indicates that a party shows a given style with a possible tendency towards another one. The following notation is used to denote a main conflict style with a trend to a secondary one: 𝑀𝑎𝑖𝑛→𝑠𝑒𝑐𝑜𝑛𝑑𝑎𝑟𝑦. Let 𝜑 be the value of the utility of a proposal. The following personal conflict handling styles are defined: 𝐶𝑜𝑙𝑙𝑎𝑏𝑜𝑟𝑎𝑡𝑖𝑛𝑔→𝐴𝑐𝑐𝑜𝑚𝑜𝑑𝑎𝑡𝑖𝑛𝑔 𝑖𝑓 𝜑 ∈ [𝑍𝑂𝑃𝐴𝑀𝐼𝑁, 𝑍𝑂𝑃𝐴𝑀𝐼𝑁+ 𝛼 2 [ 𝐶𝑜𝑙𝑙𝑎𝑏𝑜𝑟𝑎𝑡𝑖𝑛𝑔→𝐶𝑜𝑚𝑝𝑟𝑜𝑚𝑖𝑠𝑖𝑛𝑔 𝑖𝑓 𝜑 ∈ [ 𝑍𝑂𝑃𝐴𝑀𝐼𝑁+ 𝛼 2 , 𝛼[ 𝐶𝑜𝑚𝑝𝑟𝑜𝑚𝑖𝑠𝑖𝑛𝑔→𝐶𝑜𝑙𝑙𝑎𝑏𝑜𝑟𝑎𝑡𝑖𝑛𝑔−𝐴𝑐𝑐𝑜𝑚𝑜𝑑𝑎𝑡𝑖𝑛𝑔 𝑖𝑓 𝜑 ∈ [𝛼, 𝛽[ 𝐶𝑜𝑚𝑝𝑟𝑜𝑚𝑖𝑠𝑖𝑛𝑔→𝐶𝑜𝑙𝑙𝑎𝑏𝑜𝑟𝑎𝑡𝑖𝑛𝑔−𝐶𝑜𝑚𝑝𝑒𝑡𝑖𝑛𝑔 𝑖𝑓 𝜑 ∈ [𝛽, 𝛾[ 𝐶𝑜𝑙𝑙𝑎𝑏𝑜𝑟𝑎𝑡𝑖𝑛𝑔→𝐶𝑜𝑚𝑝𝑟𝑜𝑚𝑖𝑠𝑖𝑛𝑔 𝑖𝑓 𝜑 ∈ [𝛾, 𝑍𝑂𝑃𝐴𝑀𝐴𝑋+ 𝛾 2 [ 𝐶𝑜𝑙𝑙𝑎𝑏𝑜𝑟𝑎𝑡𝑖𝑛𝑔→𝐶𝑜𝑙𝑙𝑎𝑏𝑜𝑟𝑎𝑡𝑖𝑛𝑔−𝐶𝑜𝑚𝑝𝑒𝑡𝑖𝑛𝑔 𝑖𝑓 𝜑 ∈ [ 𝑍𝑂𝑃𝐴𝑀𝐴𝑋+ 𝛾 2 , 𝑍𝑂𝑃𝐴𝑀𝐴𝑋] 9 Performing this classification in each round of the process allows the mediator/negotiator to build a notion of the evolution of the conflict handling style of each party (Figure 4). With this knowledge, the neutral party will be able to determine the best moments to interfere in the process in an attempt to maximize the success of the outcome. 3.2 Assessment of the Stress Level Similarly to the conflict handling style, stress has a significant influence on the decisions taken during a negotiated process. In fact, stress has an influence (positive or negative) on virtually every decision we take. One of the first definitions of stress was proposed by Selye (1956). According to the author, stress can be seen as a non-specific response of the body to external demands. These demands (the load or stimulus that triggered a response) are denominated stressors while the internal body changes that they produce constitute the actual stress response. Fig. 4. The evolution of the conflict style of a party in 10 rounds. In a traditional conflict resolution setting in which all the intervenient parties gather in the physical presence of each other, evaluating the stress level is more or less natural for us: we do it by analyzing the body language and the effects of stress on the body and mind of our interlocutors, in an almost innate way. However, doing it over a virtual communication channel in which all this accessory information is lost results almost impossible. This constitutes a 10 significant obstacle to effective communication and jeopardizes the ability of the negotiator or mediator to take good decisions. In this line of work focus is placed on estimating the level of stress of the users in a non-intrusive and transparent way and providing this information to support decision making, ultimately contributing to more satisfactory and successful outcomes. Similar approaches can be found in the literature, despite in different fields. (Healey and Picard 2005) analyze the level of stress of the drivers of vehicles, although in an invasive way, whereas in (Vizer et al. 2009), detection of stress is performed using keystroke and linguistic features. In order to develop this model, an experiment was set up whose objective was to determine the effects of stress on behavioral, physical and cognitive parameters. In that sense, a game involving mental calculations and memorization was developed to be played on devices equipped with a basic set of sensors. The devices used to implement the experiment are depicted in Figure 5 and briefly described in Table 1. Table1. Brief description of the devices used to implement the experiment. Device Brief description Main features HP Touchsmart All-in-one PC touchscreen, web cam, large screen Samsung Galaxy Tab Tablet PC touchscreen, web cam, ac- celerometer, relatively large screen, mobile, Android OS HTC PDAs Smartphones touchscreen, camera, ac- celerometer, mobile, Android OS Sony FCB- EX780BP 25x Super HAD PAL Color Block Camera with External Sync 25x Optical Zoom, Image sta- bilizer, Day/Night Mode, Pri- vacy Zone Masking The participants of the experiment played this game in two different phases. In a first one, the participants would do it without any stressors, with all the time and quietness to think carefully on their decisions. In a second phase, the same participants played the same game but this time subject to stressors such as time constraints, vibrations on the device, annoying and unexpected sounds, among others. The main objective of the experiment was to determine, for each user, which of the studied parameters was affected by stress and in which measure. 11 Fig. 5. Devices used in the experiment to collect data about interaction and movement pat- terns. The parameters considered were as follows: • Touch pattern - the touch pattern represents the way in which a user touches the device and represents a variation of intensity over a period of time. This information is acquired from touchscreens with support for touch intensity. • Touch accuracy - a comparison between touches in active controls versus touches in passive areas (e.g. without controls, empty areas) in which there is no sense in touching. This information is acquired from touchscreens. • Touch intensity - the intensity of the touch represents the amount of force that the user is putting into the touch. It is analyzed in terms of the maximum, minimum and mean intensity of each touch. This information is acquired from touchscreens. • Touch duration - this represents the time span between the beginning and the end of the touch event. This data is acquired from devices with touchscreens. • Amount of movement - the amount of movement represents how and how much the user is moving inside the environment. An estimation of the amount of movement from the video camera is built. The image processing stack uses the principles established by Castillo et al. (2011) and uses image difference techniques to calculate the amount of movement between two consecutive frames (Fernández-Caballero et al., 2010). • Acceleration - the acceleration is measured from accelerometers in mobile devices. It is useful for building an estimation of how much the user is moving and how he is doing it (e.g. is the user having sudden movements?). Moreover, information from the accelerometer is used to support the estimation of the intensity of touch. With data about this parameters for each user and for each scenario (stress/no stress), significance tests were performed to determine, to which extent, each user was affected by stress in each parameter. This supported the development of personalized stress models based on the way users interact with the technological 12 devices and on the way they move. The whole experiment is described in detail in (Careiro et al. 2012). These stress models allowed us to develop a real-time solution for assessing the level of stress of users. It is non-intrusive since it is based solely on the analysis of the users’ interaction patterns. It is reliable since these interaction patterns are hard to fake, are innate and we exhibit them without planning them or thinking about them. It is also transparent: it can be used in conjunction with any application running on a handheld device and provide information about the stress level of the user to a server. The information compiled about stress includes not only the estimated level of stress but also the contribution of each of the parameters for the levels of stress (e.g. some parameters are more affected than others for given users, thus have different weights). Finally, it also provides a measure of the quality of information given in terms of the number and type of inputs available. In fact, there are moments in which there will be no data for a given input (e.g. in the last seconds the user did not touch the screen). The application can still provide a value of stress based on the other parameters (e.g. accelerometer, amount of movement), although with an expected smaller accuracy. The quality of information quantifies the accuracy of this estimation based on which parameters are available and on the results of the tests of significance (a parameter that has shown significant differences in the tests will provide more reliable information on stress than a parameter that has shown small differences). This transparent stress estimation layer allows for a mediator to have knowledge about the level of stress of the parties involved in the conflict resolution, allowing him to take better decisions during the process. These decisions may include a change in the strategy, pauses in the process or interrupting direct communication between the parties. With this approach, the mediator is no longer blindly and coldly analyzing proposals from a purely economic perspective. He does it with the support of all this context information, in a way that is more similar to what is done when the parties are together in a single physical location. This will allow for better decisions and, ultimately, for better possibilities of achieving satisfactory outcomes. 3.3 Simulation of User Traits One useful feature in simulation tools concerns the possibility to simulate different user-types. These types are created based on role-playing game techniques, i.e. each user-type will represent a different pre-defined role on the environment, which enables him to perform different actions or exhibit a given behaviour. Thus, the simulation platform allows defining the different actions that each role can perform in the conflict resolution environment (Figure 6). The use of role-playing game techniques permits the distinction of the different users with specific 13 characteristics. It also allows to assign different conflict styles to each user group, denoting the way that each one has to deal with the conflict. Fig. 6. A detail of the configuration of the vital signs of a given user. Vital signs can be configured to behave randomly according to a normal distribution with given parameters or they can be completely planned for the time of the simulation, allowing to plan specific scenarios. The tool also allows simulating vital signs. These are some of the most important factors in determining the emotional state of an individual. Moreover, vital signs can also be directly related with the stress level. However, the methods for acquiring this information may be invasive and even influence the results. In that sense, for the moment, these are being simulated. The configuration of these parameters is depicted in Figure 7. By creating different vital sign configurations, one can induce specific scenarios and see how the inference mechanisms create the associated emotional state of a party and, consequently, how the UMCourt platform adapts its strategies according to it. The simulation tool allows to independently configure the vital signs of each user. Two modes are possible: Random and Planned. In Random mode, the vital signs of the different users can be configured to develop randomly, according to configurable Gaussian functions. Alternatively, in the Planned mode, these vital signs can be completely planned. This means that it is possible to configure the exact vital signs of a user in each time instant of the simulation. This allows to, for example, induce a given physiological state for a given time instant and assess how the system reacts to it. 14 Fig. 7. A detail of the configuration of the users’ personality types. 3.4 Simulation of Emotions The simulation tool can also provide information regarding the emotional state of the users. In order to simulate the emotions of a user in a realistic fashion, a combined use of the user’s objectives (bridging with the work described in the previous subsection), the state of the environment and the type of personality is being made. For example, if the user has as objective to maximize his personal gain at all cost and the most likely outcome of the dispute resolution process is one in which his gains are low, the simulated emotion will be between sadness and disappointment. The time that this emotion lasts as well as its intensity is determined by the personality type. To model the personality of each user the OCEAN (Openness, Conscientiousness, Extroversion, Agreeableness and Negative emotionality) model is being used (McCrae et al. 2005). To define the emotions, a mutation of the OCC theory (Ortony et al. 1988) is being used, defined after the requirements of this specific field of application. Thus, at this moment, our emotional model considers the following basic emotions: happiness, sadness, fear, anger, disappointment and surprise. When defining this basic set of emotions, it was also considered the most suited emotions to express the opinion of the user about a suggestion or action of the system. Evidently, the same actions will trigger different emotions on different users, depending on the personality type. Moreover, the system is able to infer the emotional state of the dispute resolution environment as the average of the emotional state of all the participants in the dispute resolution process. A graphic similar to the one shown previously in figure 7 is computed using the average values of the emotions of all the users, that allows to understand the overall state of the environment. This is useful when the system must perform an action in the common scenarios in which there is a conflict of interests. In this sense, two approaches can be followed. On the one hand, the system can be configured to satisfy the emotions of a particular user or of a user with a given role. On the other hand, the system can be configured to maximize the satisfaction of all the users in the environment. In this case the mean of the preferences is maximized. This results interesting when we are training automated conflict resolution models to behave in ways that are similar to the ones conducted by human experts. 15 4 Conclusion and Future Work Current trends on ODR tools are not considering important information such as the body language or emotional states. This happens because current tools rely basically on web interfaces. In order to address this problem, a new approach is proposed, based on intelligent environments. It was argued that current tools must be complemented by environments that are able to acquire important context information about the parties and their surroundings. Using this information, ODR tools and mediators/negotiators will be able to determine to which extent a suggestion, an action or a given topic affects each party and, this way, adapt strategies in order to more efficiently achieve more mutually satisfactory outcomes. Given the complexity of such environments, a simulation platform was used to study and assess the development of the software modules, that are then used in the actual implementation of the UMCourt conflict resolution platform. This allows us to create specific scenarios, with specific user-types and determine how UMCourt and neutrals adapt strategies accordingly (case retrieval, solution proposal, mediation and negotiation conduction). In future work, more steps will be taken towards the actual implementation of all the modules in a real environment. This includes the acquisition of sensors that will gradually replace the sensors currently being simulated. A database of cases that store the emotional state before and after given actions taken by the ODR system is also being built. This allows us to use a nearest neighbor retrieval algorithm to examine past cases and predict, at each time and based on the conflict styles, how a given action may affect each party. This will allow the system to determine at which point a party moves from an avoiding conflict style to a compromising one (Carneiro et al. 2010a), for example. In fact, the determination of the conflict styles is a very important feature for a mediator. In future work machine learning techniques such as classification will be used in order to make the selection of cases and compare their effectiveness with the one of the retrieval algorithm. This, we believe, is the path to develop ODR tools that encompass very important context information that is being ignored by current research trends. Acknowledgments. This work is part-funded by ERDF - European Regional Development Fund through the COMPETE Programme (operational programme for competitiveness) and by National Funds through the FCT - Fundação para a Ciência e a Tecnologia (Portuguese Foundation for Science and Technology) within project FCOMP-01-0124-FEDER-028980 (PTDC/EEI-SII/1386/2012) and PEst-OE/ EEI/UI0752/2011. The work of Davide Carneiro is also supported by a doctoral grant by FCT (SFRH/BD/64890/2009). 16 References AARTS, E. AND GROTENHUIS, F. (2011) Ambient Intelligence 2.0: Towards Synergetic Prosperity. Journal of Ambient Intelligence and Smart Environments, 3(1):3–11. ANDRADE, F., NOVAIS, P., CARNEIRO, D., NEVES, J. (2010) Conflict Resolution in Virtual Locations. In Information Communication Technology Law, Protection and Access Rights: Global Approaches and Issues. Portela, I., Cunha, M. (Eds), IGI Global. BARRETEAU, O., LE PAGE, C. AND D'AQUINO, P. (2003) Role-Playing Games, Models and Negotiation Processes. In Journal of Artificial Societies and Social Simulation, vol. 6, no. 2. BELLIFEMINE, F.L., CAIRE, G., GREENWOOD, D. (2007) Developing Multi-Agent Systems with JADE, Wiley Series in Agent Technology, ISBN-13: 978-0470057476. BROWN, H., MARRIOTT, A. (1999) ADR Principles and Practice. Sweet and Maxwell. CANNON, W. (1927) The James-Lange Theory of Emotions: A Critical Examination and an Alternative. In The American Journal of Psychology, vol. 39, pp. 106-124. CARNEIRO, D., NOVAIS, P., ANDRADE, F., ZELEZNIKOW, J., NEVES, J. (2009) The Legal Precedent in Online Dispute Resolution. In Legal Knowledge and Information Systems, ed. Guido Governatori, Proceedings of the Jurix 2009, pp. 47-52. CARNEIRO, D., NOVAIS, P., ANDRADE, F., NEVES, J. (2010a) Using Mediation to Solve Disputes with Avoiding Parties. In Proceedings of the Fourth International Workshop on Juris-Informatics (Jurisin 2010), pp. 17-28. CARNEIRO, D., NOVAIS, P., COSTA, R., NEVES, J. (2010b) Developing Intelligent Environments with OSGi and JADE, The Third IFIP International Conference on Artificial Intelligence in Theory and Practice. CARNEIRO D., CARLOS CASTILLO J., NOVAIS P., FERNÁNDEZ-CABALLERO A., NEVES J. (2012) Multimodal Behavioural Analysis for Non-invasive Stress Detection, Expert Systems with Applications, Elsevier, http://dx.doi.org/10.1016/j.eswa.2012.05.065 CORCHADO, E., ARROYO, A., TRICIO V. (2010) Soft computing models to identify typical meteorological days. In: Logic Journal of the IGPL, Oxford University Press. COSTA, R., NOVAIS, P., LIMA, L., CARNEIRO, D., SAMICO, D., OLIVEIRA, J., MACHADO, J., NEVES. J. (2009) VirtualECare: Intelligent Assisted Living. In Electronic Healthcare, Dasun Weerasinghe (ed.), Springer-Verlag, pp. 138-144. CASTILLO, J.C., RIVAS-CASADO, A., FERNÁNDEZ-CABALLERO, A., LÓPEZ, M.T., MARTÍNEZ-TOMÁS, R. (2011) multisensory monitoring and interpretation framework based on the model-view-controller paradigm. In: Proceedings of the 4th International Workshop on the Interplay between Natural and Artificial Computation, vol 1, pp. 441–450. CORCHADO, E., SEDANO, J., CURIEL, L. AND VILLAR, J. R. (2012) Optimizing the operating conditions in a high precision industrial process using soft computing techniques. Expert Systems, 29: 276–299. doi: 10.1111/j.1468-0394.2011.00588.x. COSTA, R., CARNEIRO, D., NOVAIS, P., LIMA, L., MACHADO, J., MARQUES, A., NEVES, J. (2008) Ambient Assisted Living. In: Advances in Soft Computing, vol. 51, Springer- Verlag, pp. 86-94. http://dx.doi.org/10.1016/j.eswa.2012.05.065 17 DAMÁSIO, A. (1994) Descartes' Error: Emotion, Reason, and the Human Brain, Putnam Publishing, hardcover: ISBN 0-399-13894-3. DE VRIES B.R., LEENES R., ZELEZNIKOW J. (2005) Fundamentals of providing negotiation support online: the need for developping BATNAs, Proceedings of the Second International ODR Workshop, Tilburg, Wolf Legal Publishers, pp. 59-67. JAMES, W. (1884) What is an Emotion?. In: Mind, vol. 9, no. 34, pp. 188-205. FERNÁNDEZ-CABALLERO, A., CASTILLO, J.C., MARTÍNEZ-CANTOS, J., MARTÍNEZ-TOMÁS, R. (2010) Optical flow or image subtraction in human detection from infrared camera on mobile robot. Robotics and Autonomous Systems 58 (12), 1273– 1281. FRIEDMAN, GEORGE H. (1997) Alternative Dispute Resolution and Emerging Online Technologies: Challenges and Opportunities, 19 Hastings Comm. & Ent. L.J. 695. HEALEY, J., PICARD, R. (2005) Detecting Stress During Real-World Driving Tasks Using Physiological Sensors. IEEE Transactions on Intelligent Transportation Systems, 6(2):156–166. KATSCH E., RIFKIN J. (2001) Online dispute resolution – resolving conflicts in cyberspace. Jossey-Bass Wiley Company, San Francisco. LANGE, C. (1967) The emotions. New york: Harner Publishing co. LODDER, A., THIESSEN, E. (2003) The role of artificial intelligence in online dispute resolution. In Workshop on Online Dispute Resolution at the International Conference on Artificial Intelligence and Law, Edinburgh, UK. MCCRAE, R., COSTA, P., MARTIN, T. (2005) The neo-pi-3: A more readable revised neo personality inventory. In: Journal of Personality Assessment, 84 (3):261—270. NICK, Y. (2003) The Psychology of Massively Multi-User Online Role-Playing Games: Motivations, Emotional Investment, Relationships and Problematic Usage. In Avatars at Work and Play, Computer Supported Cooperative Work, vol. 34, pp. 187-207. ORTONY, A., CLORE, G., COLLINS, A. (1988) The cognitive structure of emotions. University Press. OPEN SERVICES GATEWAY INITIATIVE (2003) Osgi Service Platform, Release 3. IOS Press. RAIFFA, H. (1982) The art and science of negotiation: how to resolve conflicts and get the best out of bargaining, Cambridge, The Belknap Press of Harvard University Press. SEDANO J., CURIEL, L., CORCHADO, E., CAL E., VILLAR, J. (2010) A soft computing method for detecting lifetime building thermal insulation failures. In Integrated Computer-Aided Engineering, Vol. 17, Issue 2 , pp. 103-115, IOS Press. SELYE, H. (1956) The Stress of Life. McGraw-Hill. STEENBERGEN, W. (2005) Rationalizing Dispute Resolution: from best alternative to the most likely one, Proceedings of 3rd ODR Workshop, Brussels. THOMAS, K., KILMANN, R. (1974) Conflict and Conflict Management. Available at http://www.kilmann.com/conflict.html <last accessed July, 2012> 18 VIZER, L. M., ZHOU, L., SEARS, A. (2009) Automated stress detection using keystroke and linguistic features: An exploratory study. International Journal of Human-Computer Studies, 67(10):870–886. ZARGARI, S. A., SIABIL, S. Z., ALAVI, A. H. AND GANDOMI, A. H. (2012) A computational intelligence-based approach for short-term traffic flow prediction. Expert Systems, 29: 124–142. doi: 10.1111/j.1468-0394.2010.00567.x. 19 Biographies Davide Carneiro Davide Carneiro is a fellow researcher at the CCTC (Computer Science and Technology Center), at the Department of Informatics, University of Minho, Braga, Portugal. He is currently finishing his PhD, in which he worked under a joint Doctoral Programme in Computer Science that aggregates three top Portuguese Universities. He develops scientific research in the field of Artificial Intelligence, namely in Multi-agent Systems, Genetic Programming and Case- based Reasoning applied to the Legal field and Ambient Intelligence. He is the co- author of several publications in this field, including international journals, book chapters and conference proceedings. In 2008 he has been awarded the TLeIA08 - a National Award for Artificial Intelligence projects attributed by the Portuguese Artificial Intelligence Association, and in 2009 he has been awarded an Academic Merit Scholarship by the Portuguese Government. Marco Gomes Marco Gomes is a researcher at the CCTC (Computer Science and Technology Center), at the Department of Informatics, University of Minho, Braga, Portugal. He is currently enrolled in Doctoral Program on Informatics and develops scientific research in the field of Artificial Intelligence. He is the author of publications in conference proceedings and journals in the fields of Online Dispute Resolution and Ambient Intelligence. Ângelo Costa Ângelo Costa received the B.Sc. degree in Informatics Engineering, the M.Sc. in Informatics Engineering and Ph.D. in Biomedical Engineering in 2013 from the University of Minho. Currently he is a researcher at the CCTC (Computer Science and Technology Center) and the AAL4ALL project. He develops scientific research in the field of Artificial Intelligence, namely Cognitive Assistants and Multi-Agent Systems. Currently he is one of the main developers of cognitive assistants for mobile platforms. Paulo Novais Paulo Novais is an associate Professor of Computer Sciences at the Department of Informatics, in the University of Minho, Braga, Portugal and researcher at the CCTC (Computer Science and Technology Center) in which he is the coordinator of the Intelligent Systems Lab. From the same university he received a PhD in Computer Sciences in 2003 and his Habilitation in Computer Sciences in 2011. He develops scientific research in the field of Artificial Intelligence, namely Knowledge Representation and Reasoning, Machine Learning and Multi-Agent Systems, with applications to the areas of Law and Ambient Intelligence. He is the 20 co-author of over 120 book chapters, journal papers, conference and workshop papers and books. He is Vice-president of APPIA, the Portuguese Association for Artificial Intelligence. José Neves Jose Neves is Full Professor in Computer Science at the University of Minho, in Braga, Portugal, since 1983. Jose Neves is the Deputy Director of the Artificial Intelligence (AI) Group at the same university. He received his PhD in Computer Science from Heriot Watt University, Edinburgh, Scotland, in 1983. His current research interests are in the areas of knowledge representation and reasoning and evolutionary intelligence, aiming to construct dynamic virtual worlds of complex symbolic entities that compete against one another, in order to solve complex problems. 1 Introduction 2 Bridging Ambient Intelligence and Online Dispute Resolution 3 Intelligent Environments Supporting Conflict Resolution 3.1 Classification of the Conflict Resolution Style 3.2 Assessment of the Stress Level 3.3 Simulation of User Traits 3.4 Simulation of Emotions 4 Conclusion and Future Work 
work_glaaka4gpve2bihl6su7kfhsq4	----	Subtractive Clustering Fuzzy Expert System for Engineering Applications Procedia Computer Science 48 ( 2015 ) 77 – 83 Available online at www.sciencedirect.com 1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of scientific committee of International Conference on Computer, Communication and Convergence (ICCC 2015) doi: 10.1016/j.procs.2015.04.153 ScienceDirect International Conference on Intelligent Computing, Communication & Convergence (ICCC-2014) Conference Organized by Interscience Institute of Management and Technology, Bhubaneswar, Odisha, India Subtractive clustering Fuzzy Expert System for Engineering Applications U.Mohan Raoa,* , Y.R.Soodb, R.K.Jarialb aResearch Scholar Electrical Engineering Department, National Institute of Technology- Hamirpur,177005, India bDepartment of Electrical Engineering, National Institute of Technology- Hamirpur, 177005, India. Abstract Performance of a fuzzy expert system is related with how good the membership functions are normalized, tuned for a problem statement and correlation of antecedents and consequents. This paper helps in tuning and designing the membership functions that are best suited for the problem statement by integrating subtractive clustering method for fuzzy expert system design. Subtractive clustering algorithm is used to generate the tuned membership functions automatically in accordance to the domain knowledge. The proposed integrated design of clustering based fuzzy expert system acts in improving the accuracy and leads to a précised decision making environment. A practical example for ageing assessment of transformer insulation oil has been included to illustrate the method much effectively; the discussed example has been validated practically in the laboratory and is found that the proposed method is efficient in decision making © 2014 The Authors. Published by Elsevier B.V. Selection and peer-review under responsibility of scientific committee of Missouri University of Science and Technology. * Corresponding author. Tel.: +91-9882557759. E-mail address:mohan13.nith@gmail.com © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of scientific committee of International Conference on Computer, Communication and Convergence (ICCC 2015) (ICCC-2015) http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.04.153&domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.04.153&domain=pdf 78 U. Mohan Rao et al. / Procedia Computer Science 48 ( 2015 ) 77 – 83 Keywords:Clustering,Decession Making,Expert system, Fuzzy. 1. Introduction Decision making is a cautioned issue for thetechnocrats in various engineering domains which is being survived by the cutting edge advancements made to the methodologies that involved in thedecision making process. Choice of decision making is preferably done by using Multiple AttributeDecision Making (MADM) methods or Artificial Intelligent (AI)techniques depending on the problem constraints and available domain knowledge along with data related to the problem. MADM methods involve Simple additive weighting (SAW), Weighted Product Method (WPM),Analytical Hierarchy Process (AHP), TOPSIS, PROMETHEE [5]. Artificial Intelligent techniques include Artificial Neural Networks, Fuzzy systems, Evolutionary techniques, Data clustering algorithms; Knowledge based systems and Expert systems. The present paper is dealt with fuzzy expert system and subtractive clustering method which acts in proposing a novel alternative designing ideology for integrated fuzzy clustering expert system in detailed with practically validated example at the end of the day The accuracy of the decision making rely on the accuracy the decision maker involved with the problem statement because it is the choice of the decision maker to decide the weight age for the individual attributes for all the available alternatives in case of applying MADM methods. In case of AI techniques it is based on domain expert’s knowledge and framing of relation between IF parts (Antecedents) and THEN parts (Consequents). In order to bring down human errors and increase accuracy, tasks were shifted to machines which mimic the human thinking. At first expert systems were developed in 1970s and by 1080s these were not only confirmed to laboratories but introduced to the fields. Fuzzy Expert System (FES) has found increasingly interest in engineering applications for its behavior towards reduction of errors, it is to be noted that that FES performance is based in construction of rules and tuning of membership functions which is quite complex [3]. The basic process involved in designing a FES is as depicted in Fig. 1. Fig. 1. Design of Fuzzy Expert System So a step ahead this work presents a methodology for tuning the membership functions that deals in modification of the last step which is highlighted in the above figure. This tuning is done by integrating subtractive clustering algorithm as this algorithm has a remarkable feature of developing a fuzzy inference engine whose members ship functions were tuned according to the problem data, tuning of membership functions is done by identifying the number of linguistic variables as a cluster having a common center. An example of classifying the age group of power transformer insulation oil has been considered with New Age (NA), Medium Age (MA), and Old Age (OA) as linguistic variables.An expert system has been designed with the proposed method i.e. by integrating subtractive clustering method with existing method and found that this method is efficient by validating and comparing the obtained results with the practical results obtained in the laboratory. [4]. Observations on variations in shape of Identify Linguistic Variables Determine Fuzzy Sets Normalize Membership Functions Elicit Rule Base Develop Inference Engine Tune the System 79 U. Mohan Rao et al. / Procedia Computer Science 48 ( 2015 ) 77 – 83 membership functions were also compared and discussed to highlight the significance of the proposed method along with its performance in identifying the age of insulation oil. 2. Proposed Design of subtractive clustering FES The design of FES became much complex with tuning of system, which involves with tuning of membership functions. So an attempt is made to bring the problem from complex and high dimensional space to a designer friendly space by handling this particular tuning of membership function task to Subtractive clustering algorithm. After applying the subtractive clustering algorithm an inference engine with the tuned membership functions will beobtained and the same can be further employed for expert system design. The proposed design of FES is as shown For using subtractive clustering algorithm the following are required • Input file: A data sheet with a single column and number of rows as required, containing all the linguistic variables data in ASCII format. • Desired Output file: A data sheet with a single column and number of rows as required containing all the desired outputs for the corresponding input vectors present input file. It is to be noted that the number of rows for input and output files should be equal and the file is in ASCII format. • Radius value: An integer which decides the influence for the cluster in the space. However an accurate value of radius can be found by the performance parameters of the algorithm. At first an inference engine has been developed in Matlab environment which involves with identifying linguistic variables from the problem statement and developing membership functions in the universe of discourse, Evolution of all the antecedents and their implication to the associated consequents and develop an inference engine. The designing done so far is a usual method of designing a fuzzy inference engine, say IE-1 for instance, now instead of going for the tuning the inference engine(IE-1). Develop input file and desired output file as demonstrated above initiate both the files and confirm with an integer value as the radius value and declare the same value. Once the Fig. 2. Proposed Design for Fuzzy Expert System Inefficient Compare and check for certainty in Inference Engine Domain Expert User Interface End User History Explanation Generator Identify linguistic variables Determine fuzzy sets Normalize membership functions Elicit rule base Efficient Declare Radius Value Apply SCM Algorithm Evaluate Inference Engine Initiate I/P and O/P Files Desired Output file Input file 80 U. Mohan Rao et al. / Procedia Computer Science 48 ( 2015 ) 77 – 83 input and output files were initiated and radius value declared apply subtractive clustering algorithm the function for applying the said algorithm in Matlab environment is shown in Equation 1 below [C,S] = subclust([filein fileout],r); (1) Where, C = centre matrix, S = sigma values matrix, filein, fileout are input and output files, r = Radius After applying the algorithm function for input and output files, entire data is plotted in a multi-dimensional space and each point acts a data point to form a cluster. Cluster centres are calculated and all the data points having higher influence on any cluster center will fall within that cluster. The formulation of clusters is based on density calculation which is calculated by the Eqn 2 (2) Here ra is radius, xi, xj are data points. So a data point will have a high density value if there lie many number of data points in its surrounding. [2,6] The highest density point is chosen as first cluster center xc1 . In the next iteration the density measure of each data point xi is obtained from Eqn 3 below (3) This process continues until a sufficient number of clusters are attainted. The main tasks of the algorithm are as follows: • After calculation of density at each point first cluster center is identified as the point having highest density. • Eliminates all data points in the area around the first cluster center as laid by the radius value for obtaining the next cluster center. • The process continues until all the data points fall within the radii of a cluster center. It is a matter of fact that a center matrix C, Sigma values matrix S and an inference engine say IE-2 for instance will be developed automatically after the algorithm terminates or simply called as the products of the algorithm. Evolution of this inference engine (IE-2) can be done by having knowing about the relationship between the C and S matrices and cluster center values in accordance to the membership functions of IE-2. 3. Example of insulation Oil age assessment Problem statement: - To identify the ageing of power transformer insulation oil within three age groups viz New Age (NA), Medium Age (MA), and Old Age (OA). Step 1: To define linguistic variables: The first, and probably the most important, step in building any expert system is to specify the problem. Problem input and output variables and their ranges are to be determined. In the current problem of this work, three linguistic variables are New age, Medium age, and Old age.Step 2: To determine fuzzy sets: Fuzzy sets can have a variety of shapes. However, a triangle membership function is selected in random. Fig.3 and Fig.4. represents input and output membership functions. Input membership function is justified by extracting features from the UV/VIS (spectroscopy) response as (-8.2) to 0 for NA, (0.1) to 64 for MA, 64.1 and above for OA [1, 7]. Output membership function is classified as 0-3 for NA, 3.1-6 for MA and 6.1-9 for OA and sets for all linguistic variables accordingly. Step 3:Elicit fuzzy rules: All the rules that are to be accomplished with our system are to be developed with proper logical reasoning, in the current problem three rules based on If-Then; conditions were framed related to three 1 2 2 exp / 2 n ji i j a x x D r 1 2 2 exp / 2 n jii i i j a c x x D D D r 81 U. Mohan Rao et al. / Procedia Computer Science 48 ( 2015 ) 77 – 83 conditions in according to the age of oil.Step 4:Develop IF-1: Inference Engine-1 is developed using Matlab Fuzzy logic tool Fig.3. Input membership function of IE-1 Fig. 4. Output membership function of IE-2. Step 5: Files and radius: For input file is of A total of 150 P.I values has been generated, 50 values for each age group and organized as (150 X 1) matrix in ASCII format. And for output file 150 desired outputs were designated in the same order and ASCII format. And a radius value of 5 is given and SCM function is applied. An inference engine IE-2 is generated here and is evaluated for all the input vectors by using rule viewer in Matlab environment.The inference engine IE-2 obtained from subtractive clustering algorithm has its membership functions which are best suit and optimal for the insulation oil ageing problem statement. The input membership function obtained from algorithm is shown in Fig 5. The discrimination between the shapes of input membership functions of IE-1 and IE-2 is observed from Fig. 3. and Fig. 5. 82 U. Mohan Rao et al. / Procedia Computer Science 48 ( 2015 ) 77 – 83 Fig. 5. Input membership function of IE-2 Using the cluster centers and their range of influences a FIS is developed which will then be used to explore and understand oil ageing. In this case fuzzy inference engine is generated with three input membership functions which represent three clusters followed by three rules because only three cluster centers were identified by the algorithm. The parameters of the membership function obtained are [23.9 -4.19], where 23.9 represents the spread coefficient of the gaussian curve and -4.19 represents the center of the gaussian curve. Input membership function of cluster1 captures the position and influence of the first cluster for the input variable (C (1, 1) = - 4.19, S(1)=23.9). The devel oped syste m is teste d and the result s are discu ssed in Tabl e 1. Table 1.Performance comparison of proposed method with spectroscopy results. For testing the efficiency or decision making capability of the proposed methodology using the current problem statement 10 different insulation oil samples from ten different transformers have been considered for ageing analysis and are tested using Ultra violet visual infrared spectroscopy and from the absorbance curve obtained mean absorbance is calculated which is used in performance index (P.I.) calculation for a particular sample, and further these performance index values has been subjected to fuzzy rule evaluation of Inference engine (I.E-2) which is generated by the subtractive clustering algorithm and the subtractive clustering fuzzy expert system with Inference engine (I.E-2) has been simulated in the Matlab environment and found that the proposed developed system hasbeen tested for several samples and is found efficient in identifying the ageing of power transformer insulation oil. 4. Conclusion A methodology which is based on integrated clustering and fuzzy expert system has been proposed for effective and improved design of fuzzy expert system. The developed ideology helps in decision making and classification for any engineering indecisive statements. Design details along with implementation of the novel alternative method have been discussed. The proposed design methodology is explained step by step in detailed by considering an example of classification of power transformer oil ageing and its implementation is also discussed and found that the designed expert system is efficient in classifying the age group of the oil. This model helps in moving the decision making environment to a much précised environment. References 1. Hasmat Malik, Surinder Singh, Mantosh Kr, R.K. Jarial “UV/VIS Response Based Fuzzy Logic for Health Assessment of Transformer Oil., Elsevier, Procedia Engineering 30(2012), pp. 905-912. 2. Khaled Hammouda, Fakhreddine Karray “A comparative stydy of data clustering techniques” University of waterloo, Ontario-Canada. 3. Michael Negnevitsky “Artificial Intelligence-A guide to Intelligent systems”2nd Edition Addison Wesley, 2005, ISBN 0-321-20466-2. 4. Manual T90/T90+ “UV/VIS Spectrophotometer” TIFAC-CORE “NIT-HAMIRPUR”. 5. R. Venkata Rao “Decision Making in Manufacturing Environment Using Graph Theory and Fuzzy Multiple Attribute Decision Making Methods” Volume 2, Springer Series in Advanced Manufacturing 2013, Springer-Verlag, ISBN 978-1-4471-4375-8. S.no Ten Different Oil samples Spectroscopy Result Mean Absorbance Performance Index Proposed FES Result 1 S-1 New Oil -0.05774 -5.774 1(NA) 2 S-2 New Oil -0.063012 -6.3012 1(NA) 3 S-3 New Oil -0.035247 -3.5247 1(NA) 4 S-4 Medium Oil 0.430253 43.0253 2(MA) 5 S-5 Medium Oil 0.577060 57.7060 2(MA) 6 S-6 Medium Oil 0.379901 37.9901 2(MA) 7 S-7 Medium Oil 0.304751 30.4751 2(MA) 8 S-8 Old Oil 0.958037 95.8037 3(OA) 9 S-9 Old Oil 0.690279 69.0279 3(OA) 10 S-10 Old Oil 0.730380 73.0380 3(OA) 83 U. Mohan Rao et al. / Procedia Computer Science 48 ( 2015 ) 77 – 83 6. R.Naresh, Veena Sharma, Manisha vasistha, “An Integrated Neural fuzzy approach for fault diagnosis of transformers”, IEEE Trans. on power delivery Vol.23, No. 4; Oct-2008. 7. U.Mohan Rao, R.K.Jarial “ Fuzzy clustering approach to access the ageing of transformer insulation oil”, Annual IEEE-India Conference, Dec- 2013, pp.1-5. 
work_glbzbrgusjfitil7ejqyhr2csq	----	A Pro-performance appraisal system for the university Expert Systems with Applications 37 (2010) 2108–2116 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A Pro-performance appraisal system for the university Jui-Kuei Chen a,1, I-Shuo Chen b,* a Graduate Institute of Futures Studies, Tamkang University, 4F, No. 20, Lane 22, WenZhou Street, Taipei City 10648, Taiwan b Institute of Business & Management, National Chiao Tung University, 4F, No. 20, Lane 22, WenZhou Street, Taipei City 10648, Taiwan a r t i c l e i n f o a b s t r a c t Keywords: Fuzzy analytical network process (FANP) Decision-making trial and evaluation laboratory (DEMATEL) Original performance appraisal system (OPAS) Pro-performance appraisal system (PPAS) 0957-4174/$ - see front matter � 2009 Elsevier Ltd. A doi:10.1016/j.eswa.2009.07.063 * Corresponding author. Tel.: +886 911393602. E-mail addresses: chen3362@ms15.hinet.net (J.- com.tw (I-S. Chen). 1 Tel.: +886 912272961. Due to economic pressures and declining birth rates, universities in Taiwan are seeking ways to evaluate and improve operational performance to acquire a competitive advantage to attract more students. How- ever, current performance evaluation models have been criticized for two reasons. First, the measure- ment criteria currently used are not completely in accordance with the characteristics of different university types, research-intensive university, teaching-intensive university, and professional-intensive university. Second, the models assume independence of measured criteria. Nonetheless, in the real world, such measured criteria are seldom independent. To address these issues, we first reviewed the literature and interviewed Taiwanese higher education experts to integrate critical measurement criteria and develop an original performance appraisal system (OPAS). Next, we adapted a decision-making trial and evaluation laboratory (DEMATEL) method to present complex interdependent relationships and to construct a relation structure among measurement criteria for performance appraisal. A fuzzy analytic network process (FANP) was generated to address the dependence and feedback among each of the mea- surement criteria. Finally, we proposed a Pro-performance appraisal system (PPAS). This study offers a Pro-performance appraisal system (PPAS) to aid in future performance appraisals and improvements for all three university types. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction Higher education is the foundation for fostering high-tech tal- ent, the key factor in increasing national quality, and the main way to upgrade a nation’s competitive status (Fairweather, 2000; Meek, 2000). In recent years, the number of Taiwanese universities has increased to 157 in according to the Taiwanese Ministry of Education (Ministry of Education, 2006). However, the quality and operational performance among them has not increased pro- portionally (Taiwan Assessment and Evaluation Association, 2006). This has been a serious issue for the Taiwanese government and universities (Department of Higher Education, 2004). Visiting and standard procedure evaluations are the two meth- ods the Ministry of Education uses to evaluate universities. There are three types of Taiwanese universities: research-intensive uni- versities, teaching-intensive universities, and profession-intensive universities (Li, 2007). The current measurement criteria utilized are not completely in accordance with the characteristics of these three different university types. This is the main characteristic of the current system that some universities have argued has been ll rights reserved. K. Chen), ch655244@yahoo. unfair. In this study, we argue that since the measurement criteria utilized are assumed to be independent, each criterion may signif- icantly influence the operation performance appraisal results. However, the independence assumption is not consistent with the conditions in the real world. For the reasons above, we propose a more professional perfor- mance appraisal mechanism with more suitable measurement cri- teria for the three university types. In this study, critical measurement criteria were determined by summarizing the litera- ture and interviewing Taiwanese higher education experts. Then, a decision-making trial and evaluation laboratory (DEMATEL) meth- od was adapted to present complex interdependent relationships and to construct a relation structure among measurement criteria for performance appraisal. A fuzzy analytic network process (FANP) was constructed to solve the problem of dependence and feedback among each measurement criterion (Liou, Tzeng, & Chang, 2007). Here, we combined a DEMATEL, and a fuzzy ANP method to form a Pro-performance appraisal system (PPAS). 2. Pro-performance appraisal system (PPAS) Many studies offer insights on performance appraisal in higher education and some studies even develop evaluation models. However, there is inconsistency among different types of univer- sities and the relationships between measurement criteria are not http://dx.doi.org/10.1016/j.eswa.2009.07.063 mailto:chen3362@ms15.hinet.net mailto:ch655244@yahoo.com.tw mailto:ch655244@yahoo.com.tw http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa J.-K. Chen, I-S. Chen / Expert Systems with Applications 37 (2010) 2108–2116 2109 considered. Recent research suggests that the influential factors for operational performance for higher education vary across university types. After summarizing related literature and studies and performing in-depth interviews with higher education experts, we first developed an original performance appraisal system (OPAS). Then, we took a hybrid approach, combining DEMATEL, and Fuzzy ANP, which accounts for complex relation- ships using them to construct a Pro-performance appraisal system (PPAS). 3. Research methods To precisely quantify values with a complex measurement sys- tem is difficult. Nevertheless, such systems can be categorized into subsystems to make it easier to distinguish and evaluate each sub- group (Liou et al., 2007). Here, based on the original performance appraisal system (OPAS), DEMATEL is adapted to assess the inter- relations between each of the measurement criteria. Next, each cri- terion is weighted by conducting fuzzy ANP. Finally, we construct a Pro-performance appraisal system (PPAS) in accordance with above results. 3.1. The decision-making trial and evaluation laboratory (DEMATEL) It is difficult for a decision-maker to evaluate the single effect of a single factor while avoiding interference from the rest of the sys- tem because factors in a complex system may relate to each other directly or indirectly (Liou et al., 2007). In addition, an interdepen- dent system may result in passive positioning. For example, a sys- tem with a clear hierarchical structure may give rise to linear activity with no dependence or feedback, which may cause prob- lems that are distinct from those found in non-hierarchical systems (Tzeng, Chiang, & Li, 2007). The Battelle Geneva Institute created DEMATEL in order to solve difficult issues using interactive man-model techniques to measure qualitative and factor-linked aspects of societal problems (Gabus & Fontela, 1972). DEMATEL has been utilized in many additional contexts, such as industrial planning, deci- sion-making, regional environmental assessing, and even ana- lyzing world problems (Huang, Tzeng, & Ong, 2007). In each case, DEMATEL was used to confirm criteria interdependence and restrict the relationships that affect characteristics within an essential system and its developmental trends (Liou et al., 2007). The DEMATEL method is founded on graph theory. It allows decision-makers to analyze as well as solve visible problems. In doing so, decision-makers can separate multiple measurement cri- teria into cause and effect groups to identify causal relationships. In addition, directed graphs, called digraphs, are more useful than directionless graphs since they depict the directed relationships among subsystems. In other words, a digraph represents a commu- nication network or a domination relationship between entities and their groupings (Huang et al., 2007). The calculation steps of the DEMATEL are as follows (Liou et al., 2007; Yu & Tseng, 2006): Step 1: Calculate the initial average matrix by scores. Sampled experts are asked to point the direct effect based on their perception that each element i exerts on each other element j, as presented by aij, measured on a scale ranging from 0 to 4. No influence is repre- sented by 0, while a very high influence is represented by 4. Based on groups of direct matrices from samples of experts, we can generate an average matrix A in which each element is the mean of the corresponding elements in the experts’ direct matrices. Step 2: Calculate the initial influence matrix. After normalizing the average matrix A, the initial influ- ence matrix D, [dij]n�n, is calculated so that all principal diagonal elements equal zero. In accordance with D, the initial effect that an element exerts and/or acquires from each other element is given. The map depicts a contextual relationship among the elements within a complex system. Each matrix entry can be seen as its strength of influence. As a result, we can easily translate the relationship between the causes and effects of vari- ous measurement criteria into a comprehensive struc- tural model based on the influence degrees using DEMATEL. Step 3: Develop the full direct/indirect influence matrix. The indirect effects of problems decrease as the powers of D increase, e.g. D2, D3, ... , D1, which guarantees con- vergent solutions to the matrix inversion. Therefore, we can generate an infinite series of both direct and indirect effects. Let the (i, j) element of matrix A be pre- sented by aij, then the direct/indirect matrix can be ac- quired by following Eqs. (1)–(4). � D ¼ s A; s > 0 ð1Þ or ½dij�n�n ¼ s½aij�n�n; s > 0; i; j 2f1; 2; . . . ; ng; ð2Þ where 2 3 S ¼ Min 1 max1�i�n Pn j¼1 jaijj ; 1 max1�i�n Pn i¼1 jaijj 6664 7775 ð3Þ and lim m!1 Dm ¼ ½0�n�n where D ¼ ½dij�n�n; 0 � dij < 1: ð4Þ The total influence matrix T can be acquired by utilizing Eq. (5). Here, I is the identity matrix. T ¼ D þ D2 þ���þ Dm ¼ DðI � DÞ�1 when m !1: ð5Þ If the sum of rows and the sum of columns is repre- sented as vector r and c, respectively, in the total influ- ence matrix T, then T ¼ ½tij�; i; j ¼ 1; 2; . . . ; n; ð6Þ R ¼ ½ri�n�1 ¼ Xn j¼1 tij ! n�1 ; ð7Þ c ¼ ½cj� 0 1�n ¼ Xn i¼1 tij ! 1�n ; ð8Þ where the superscript apostrophe denotes transposition. If ri represents the sum of the ith row components of matrix T, then ri represents the sum of both direct and indirect effects of factor i on all other criteria. In addi- tion, if cj represents the sum of the jth column compo- nents of matrix T, then cj presents the sum of both direct and indirect effects that all other factors have on j. Moreover, note that j = i (ri + cj) demonstrates the degree to which factor i affects or is affected by j. Note that if (ri � cj) is positive, then factor i affects other fac- tors, and if it is negative, then factor i is affected by oth- ers (Liou et al., 2007; Tzeng et al., 2007). 2110 J.-K. Chen, I-S. Chen / Expert Systems with Applications 37 (2010) 2108–2116 Step 4: Set the threshold value and generate the impact relations map. Fig. 2. A fuzzy membership function for linguistic variable attributes. Table 1 Definition and membership function of fuzzy number. Fuzzy Linguistic variable Triangular fuzzy Finally, we must develop a threshold value. This value is generated by taking into account the sampled experts’ opinions in order to filter minor effects presented in ma- trix T elements. This step is needed to isolate the rela- tionship structure of the most relevant factors. In accordance with the matrix T, each factor tij provides information about how factor i affects j. In order to de- crease the complexity of the impact relations map, the decision-maker determines a threshold value for the influence degree of each factor. If the influence level of an element in matrix T is higher than the threshold va- lue, which we denote as p, then this element is included in the final impact relations map (IRM) (Liou et al., 2007). number number ~9 Extremely important/preferred (7, 9, 9) ~7 Very strongly important/ preferred (5, 7, 9) ~5 Strongly important/preferred (3, 5, 7) ~3 Moderately important/preferred (1, 3, 5) ~1 Equally important/preferred (1, 1, 3) 3.2. The fuzzy analytical network process (FANP) 3.2.1. Fuzzy set theory Fuzzy set theory was first developed in 1965 by Zadeh when he was attempting to solve fuzzy phenomenon problems, including problems with uncertain, incomplete, unspecific, or fuzzy situa- tions. Fuzzy set theory is more advantageous than traditional set theory when describing set concepts in human language. It allows us to address unspecific and fuzzy characteristics by using a mem- bership function that partitions a fuzzy set into subsets of mem- bers that ‘‘incompletely belong to” or ‘‘incompletely do not belong to” a given subset. 3.2.2. Fuzzy number We order the universe of discourse such that U is a collection of targets, where each target in the universe of discourse is called an element. A fuzzy number ~A is mapped onto U such that a random x ? U is appointed a real number, l~AðxÞ! ½0; 1�. If another element in U is greater than x, we call that element under A. The universe of real numbers R is a triangular fuzzy number (TFN), ~A, which means that for x 2 R; l~AðxÞ 2 ½0; 1�, and l~AðxÞ¼ ðx � LÞ=ðM � LÞ; L � x � M; ðU � xÞ=ðU � MÞ; M � x � U; 0 otherwise; 8>< >: Note that ~A ¼ðL; M; UÞ, where L and U represent fuzzy probabil- ity between the lower and upper boundaries, respectively, as shown in Fig. 1. Assume two fuzzy numbers ~A1 ¼ðL1; M1; U1Þ and ~A2 ¼ðL2; M2; U2Þ; then, (1) ~A1�~A2 ¼ðL1;M1;U1Þ�ðL2;M2;U2Þ¼ðL1 þL2;M1 þM2;U1 þU2Þ (2) ~A1 ~A2 ¼ðL1;M1;U1Þ ðL2;M2;U2Þ¼ðL1 L2;M1 M2;U1 U2Þ; Li > 0; Mi > 0; Ui > 0 (3) ~A1 �~A2 ¼ðL1;M1;U1Þ�ðL2;M2;U2Þ¼ðL1 �L2;M1 �M2;U1 �U2Þ (4) ~A1 ~A2 ¼ðL1; M1; U1Þ ðL2; M2; U2Þ¼ ðL1=U2; M1=M2; U1=L2Þ; Li > 0; Mi > 0; Ui > 0 ~A�11 ¼ðL1;M1;U1Þ �1 ¼ð1=U1;1=M1;1=L1Þ;Li > 0; Mi > 0; Ui > 0 Fig. 1. Triangular fuzzy number. 3.2.3. Fuzzy linguistic variable The fuzzy linguistic variable reflects different aspects of human language. Its value represents the range from natural to artificial language. When the values or meanings of a linguistic factor are being reflected, the resulting variable must also reflect appropriate modes of change for that linguistic factor. Moreover, variables describing a human word or sentence can be divided into numer- ous linguistic criteria, such as equally important, moderately important, strongly important, very strongly important, and extre- mely important, as shown in Fig. 2. Definitions and descriptions are shown in Table 1. For the purposes of this study, the 5-point scale (equally important, moderately important, strongly important, very strongly important and extremely important) is used. 3.2.4. Analytic network process (ANP) The purpose of the ANP approach is to solve problems involving interdependence and feedback between criteria or alternative solutions. ANP is the general form of the analytic hierarchy process (AHP), which has been used in multi-criteria decision-making (MCDM) in order to consider non-hierarchical structures. MCDM has been applied to numerous disciplines (Huang, Tzeng, & Ong, 2005). The beginning stage of an ANP uses pair-wise comparisons of the measured criteria to form a super matrix. The relative impor- tance-values of pair-wise comparisons can be categorized from 1 Fig. 3. The general form of the super matrix (Liou et al., 2007; Yu and Tseng, 2006). Table 2 Two simple cases. Number Case 1 Case 2 Structure type Matrix forming Table 4 The average initial direct-relation 10�10 matrix A. D1 D2 D3 D4 D5 D6 D7 D8 D9 D10 D1 0 3.17 1.02 0.24 0.11 0.43 0.11 0.30 2.14 0.10 D2 0.23 0 1.17 0.01 0.13 0.32 0.01 0.00 2.45 0.15 D3 3.33 3.21 0 1.12 1.86 3.37 0.23 0.14 1.07 0.40 D4 3.16 2.42 1.12 0 0.31 3.23 1.06 1.12 3.27 0.41 D5 0.13 0.08 0.03 1.03 0 2.18 1.22 1.19 3.36 0.27 D6 0.03 2.09 0.01 1.19 1.01 0 1.11 1.16 3.17 1.05 D7 2.21 0.12 2.33 3.05 3.42 3.18 0 1.10 3.16 3.31 D8 1.00 0.06 1.96 1.13 3.24 3.19 2.87 0 3.55 1.11 D9 1.23 0.23 0.05 1.06 2.04 1.31 0.38 0.21 0 0.20 D10 2.12 0.10 0.03 1.10 2.27 1.45 2.01 3.42 1.28 0 J.-K. Chen, I-S. Chen / Expert Systems with Applications 37 (2010) 2108–2116 2111 to 9 in order to represent pairs of equal importance (1) to extreme inequality in importance (9) (Saaty, 1980). Fig. 3 shows the general form of the super matrix where cm represents the mth cluster, emn represents the nth element in the mth cluster, and wij is the prin- cipal eigenvector measuring the influence of the jth cluster ele- ments on the ith cluster elements. In addition, if the jth cluster has no influence on the ith cluster, then wij = 0 (Yu & Tseng, 2006). The form of the super matrix depends on the variety of its struc- ture. In order to demonstrate how the structure is affected by the super matrix, Huang et al. (2005), Yu and Tseng (2006), and Liou et al. (2007) offer two simple cases that both involve three clusters to illustrate how to form a super matrix in accordance with differ- ent structures (see Table 2). Case 1 is much simpler than case 2, and based on each structure, the super matrices are given under each. Next, the weighted super matrix is generated by transforming all column sums to unity (Huang et al., 2005; Yu & Tseng, 2006). Then, we use the weighted super matrix to generate a limiting Table 3 An original 10-dimensional performance appraisal system (OPAS). System Measurement dimensions The original performance appraisal system (OPAS) Learning performance (D1) Life development (D2) Learning behavior (D3) Quality of teaching (D4) Research performance (D5) Professional skill performanc Organizational development External interactions (D8) School prestige (D9) Budget handling performance super matrix by using Eq. (9) to calculate global weights (Huang et al., 2005). lim k�1 wk ð9Þ Measurement criteria The enrollment rate of new students (C1) The graduation rate of current students (C2) The job acquiring rate of students (C3) The relations degree of students’ major and jobs (C4) Rate of continuing education (C5) Student innovative/creative ability (C6) The rate of borrowing books (C7) The rate of club participation (C8) The rate of course performance appraisal (C9) The rate of practical training course opening (C10) The rate of optional course opening (C11) Number of plans given by NSC (C12) The job promotion rate of faculty (C13) Number of articles published in international journals (C14) e (D6) Number of thesis winning of students (C15) Number of patents (C16) (D7) The performance of occupational refresher courses (C17) The supplemental budget of faculty research (C18) The budget of scholarships (C19) The budget of industry-university relations (C20) The budget of international relations (C21) Number of international conferences (C22) The satisfaction degree of students and faculty (C23) Employee turnover (C24) Score given by Ministry of Education (C25) (D10) The unit cost of each student (C26) The rate of tuition and schooling in school income (C27) The rate of government supplement in school overall income (C28) Table 5 Total influence T. D1 D2 D3 D4 D5 D6 D7 D8 D9 D10 D1 0.029 0.169 0.064 0.032 0.038 0.058 0.020 0.027 0.201 0.017 D2 0.033 0.020 0.059 0.017 0.031 0.043 0.011 0.010 0.203 0.015 D3 0.195 0.214 0.035 0.095 0.139 0.223 0.049 0.045 0.180 0.047 D4 0.202 0.184 0.092 0.059 0.097 0.236 0.094 0.092 0.286 0.058 D5 0.053 0.043 0.029 0.094 0.068 0.171 0.094 0.088 0.246 0.045 D6 0.052 0.128 0.033 0.100 0.114 0.080 0.090 0.089 0.243 0.078 D7 0.210 0.109 0.200 0.228 0.278 0.306 0.088 0.137 0.355 0.203 D8 0.093 0.077 0.133 0.141 0.260 0.283 0.200 0.069 0.336 0.110 D9 0.082 0.033 0.017 0.075 0.213 0.106 0.042 0.034 0.072 0.027 D10 0.158 0.063 0.053 0.121 0.201 0.186 0.156 0.205 0.220 0.052 Table 6 The sum of influences on measurement dimensions. Measurement dimensions ri + ci ri � ci D1 1.709 �0.503 D2 1.417 �0.661 D3 1.894 0.549 D4 2.362 0.441 D5 2.281 �0.419 D6 2.700 �0.688 D7 2.912 1.232 D8 2.494 0.901 D9 2.838 �1.615 D10 2.067 0.763 Fig. 4. The impact relations map of this study. 2112 J.-K. Chen, I-S. Chen / Expert Systems with Applications 37 (2010) 2108–2116 In this step, if the super matrix shows signs of cyclicity, then more than one limiting super matrix must exist. In this case, the Cesaro sum must be calculated to obtain the priority order of the multiple super matrices (Yu & Tseng, 2006). The Cesaro sum is calculated using Eq. (10) (Huang et al., 2005; Yu & Tseng, 2006) lim k�1 1 N � �XN k¼1 wk ð10Þ Table 7 The example of the local weight of criteria 23–25 under the effect of criteria 3. Measurement criteria C23 C24 C23 1.000 1.000 3.000 0.126 0 C24 4.583 6.708 7.937 1.000 1 C25 2.297 4.711 6.433 0.116 0 Eq. (10) calculates the average effect of a limiting super matrix. Otherwise, the super matrix would be raised to a large power to generate the priority weights (Liou et al., 2007; Yu & Tseng, 2006). 4. Empirical study of Pro-performance appraisal system (PPAS) Due to economic pressures and declining birth rates, Taiwanese universities are seeking ways to improve operational performance to acquire a competitive advantage and attract more students. In the previous section, we explained that performance appraisal cri- teria vary greatly from one university to another. In addition, crite- ria can affect each other. There are three university types in Taiwan, the research-intensive universities, the teaching-intensive universities, and the professional-intensive university, and the per- formance improvement and evaluation focuses are different for each university type. To overcome the challenges outlined and to meet the requirements above, we take the interrelationship be- tween criteria and the types of universities into account while developing a Pro-performance appraisal system (PPAS). 4.1. Forming an original performance appraisal system (OPAS) Due to the different measurement criteria involved, construct- ing a Pro-performance appraisal system (PPAS) for universities is complicated. The Pro-performance appraisal system (PPAS) must allow for interdependence and be in accord with real practice. In this study, we first categorized related measurement criteria from existing studies (Cameron, 1978; Lysons, Hatherly, & Mitchell, 1998). For a detailed summary of the measurement criteria, refer to Mei and Lee (2006). Twenty-five higher education experts were also consulted: ten from research-intensive universities, seven from professional-intensive universities, and eight from teaching- intensive universities. The works of the National Science Council (NSC) were consulted to form an original ten-dimensional perfor- mance appraisal system (OPAS) (Learning performance (D1), Life development (D2), Learning behavior (D3), Quality of teaching (D4), Research performance (D5), Professional skill performance (D6), Organizational development (D7), External interactions (D8), School prestige (D9), and Budget handling performance (D10)). Each category includes 2–3 measurement criteria (Table 3). A questionnaire was adapted and given to 66 experts from the three university types. Of the 66 questionnaires, 41 were used for this study: sixteen from the research-intensive universities (16/ C25 Local weight .149 0.218 0.155 0.212 0.435 0.075 .000 3.000 5.433 7.504 8.631 0.742 .133 0.184 1.000 1.000 3.000 0.182 Table 8 The un-weighted matrix. C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15 C16 C17 C18 C19 C20 C21 C22 C23 C24 C25 C26 C27 C28 C1 0 0 0 0 0 0 0 0 0.09 0.11 0.08 0 0 0 0 0 0.03 0.07 0.05 0 0 0 0 0 0 0 0 0 C2 0 0 0 0 0 0 0 0 0.03 0.02 0.02 0 0 0 0 0 0.02 0.006 0.03 0 0 0 0 0 0 0 0 0 C3 0 0 0 0 0 0.06 0.08 0.11 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 C4 0 0 0 0 0 0.04 0.03 0.06 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 C5 0 0 0 0 0 0.03 0.07 0.04 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 C6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.05 0.04 0.06 0 0 0 0 0 0 0 0 0 C7 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.02 0.04 0.03 0 0 0 0 0 0 0 0 0 C8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.01 0.01 0.02 0 0 0 0 0 0 0 0 0 C9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.07 0.09 0.07 0 0 0 0 0 0 0 0 0 C10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.06 0.07 0.08 0 0 0 0 0 0 0 0 0 C11 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.03 0.03 0.04 0 0 0 0 0 0 0 0 0 C12 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.11 0.11 0.1 0.17 0.2 0.16 0.24 0.31 0.27 0.21 0.16 0.15 C13 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.09 0.14 0.08 0.15 0.11 0.13 0.18 0.22 0.24 0.11 0.14 0.13 C14 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.13 0.12 0.19 0.18 0.23 0.24 0.58 0.47 0.49 0.24 0.19 0.16 C15 0 0 0 0 0 0.20 0.28 0.31 0.26 0.22 0.27 0 0 0 0 0 0.07 0.05 0.06 0.08 0.07 0.07 0 0 0 0 0 0 C16 0 0 0 0 0 0.67 0.54 0.48 0.41 0.36 0.36 0 0 0 0 0 0.10 0.11 0.09 0.13 0.08 0.16 0 0 0 0 0 0 C17 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.06 0.07 0.06 0 0 0 0 0 0 C18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.09 0.09 0.08 0 0 0 0 0 0 C19 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.04 0.06 0.03 0 0 0 0 0 0 C20 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.09 0.12 0.13 C21 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.1 0.13 0.11 C22 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.15 0.11 0.13 C23 0.05 0.16 0.07 0.08 0.21 0 0 0 0.02 0.04 0.02 0.17 0.16 0.16 0.19 0.17 0.01 0.004 0.005 0.01 0.02 0.01 0 0 0 0.01 0.01 0.03 C24 0.68 0.61 0.57 0.74 0.55 0 0 0 0.13 0.17 0.14 0.62 0.48 0.51 0.43 0.59 0.06 0.03 0.04 0.03 0.02 0.02 0 0 0 0.03 0.06 0.07 C25 0.27 0.23 0.36 0.18 0.24 0 0 0 0.06 0.08 0.11 0.21 0.36 0.33 0.38 0.24 0.04 0.06 0.05 0.06 0.05 0.04 0 0 0 0.06 0.08 0.09 C26 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.01 0.007 0.001 0 0 0 0 0 0 0 0 0 C27 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.03 0.01 0.001 0 0 0 0 0 0 0 0 0 C28 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.02 0.003 0.003 0 0 0 0 0 0 0 0 0 J.-K . C h en , I-S. C h en /E xp ert System s w ith A p p lica tio n s 3 7 (2 0 1 0 ) 2 1 0 8 – 2 1 1 6 2 1 1 3 Table 9 The weighted matrix. C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15 C16 C17 C18 C19 C20 C21 C22 C23 C24 C25 C26 C27 C28 C1 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 0.0024 C2 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 C3 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 0.0021 C4 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 0.0011 C5 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 0.0012 C6 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 C7 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 C8 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 C9 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 0.0077 C10 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 0.0071 C11 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 0.0034 C12 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 0.1659 C13 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 0.1303 C14 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 0.2439 C15 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 0.0762 C16 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 0.1371 C17 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 0.0069 C18 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 0.0095 C19 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 0.0047 C20 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 C21 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 0.0210 C22 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 0.0241 C23 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 0.0181 C24 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 0.0738 C25 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 0.0406 C26 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 C27 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 C28 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 2 1 1 4 J.-K . C h en , I-S. C h en /E xp ert System s w ith A p p lica tio n s 3 7 (2 0 1 0 ) 2 1 0 8 – 2 1 1 6 Fig. 5. The impact of direction map of measurement criteria. J.-K. Chen, I-S. Chen / Expert Systems with Applications 37 (2010) 2108–2116 2115 36), seven from teaching-intensive universities (7/11), and 18 from professional-intensive universities (18/19). The experts ranked the level of interrelationships among each measurement dimension on a scale from 0 to 4 (No influence, 0, to Very high influence, 4), and the importance between each measurement performance appraisal criterion on a 5-point scale described in Table 1. Table 10 Study result summary. System Measurement dimensions Meas The original performance appraisal system (OPAS) Learning performance (D1) The e The g Life development (D2) The j The r Rate Learning behavior (D3) Stude The r The r Quality of teaching (D4) The r The r The r Research performance (D5) Num The j Num (C14) Professional skill performance (D6) Num Num Organizational development (D7) The p (C17) The s The b External interactions (D8) The b The b Num School prestige (D9) The s Empl The s Budget handling performance (D10) The u The r The r incom 4.2. Evaluating the interrelationships of dimensions Here, the questionnaires from 41 higher educational experts were used to determine the level of relationship among each dimension. With the questionnaire data, we formed the average initial direct-relation 10�10 matrix A using pair-wise comparisons as shown in Table 4. Next, using Eqs. (1)–(3), we acquired the normalized direct- relation D from matrix A. After that, Eq. (5) was adapted to obtain total influence T (Table 5). Finally, we used Eqs. (7) and (8) to determine the total influence given to and received by each mea- surement dimension. The results are provided in Table 6. To avoid relationships too complex for our system, a threshold value under 0.20 was adopted after consulting with higher educa- tion experts. Fig. 4 shows the impact relations map (IRM). 4.3. Weighting the criteria and constructing a Pro-performance appraisal system (PPAS) After calculating the level of interrelationships among each measurement dimension, we utilized fuzzy ANP to acquire the weights of each measurement criterion. The importance of rela- tionships between measurement criteria is paralleled based on the impact relations map above. Such pair-wise comparisons are in accordance with the data in Table 5. In Table 7, we provide an example of the local weight calculated by the principle eigenvector of comparison, specifically of criteria 23–25 when under the effect of criterion 3. The complete result is an un-weighted super matrix, shown in Table 8. Next, we calculate the limiting power of the un-weighted matrix until it remains stable using Eq. (9). The result is shown urement criteria Global weights Ranking nrollment rate of new students (C1) 0.0024 C14 raduation rate of current students (C2) 0.0007 C12 ob acquiring rate of students (C3) 0.0021 C16 elations degree of students’ major and jobs (C4) 0.0011 C13 of continuing education (C5) 0.0012 C15 nt innovative/creative ability (C6) 0.0004 C24 ate of borrowing books (C7) 0.0003 C25 ate of club participation (C8) 0.0001 C22 ate of course performance appraisal (C9) 0.0077 C21 ate of practical training course opening (C10) 0.0071 C20 ate of optional course opening (C11) 0.0034 C23 ber of plans given by NSC (C12) 0.1659 C18 ob promotion rate of faculty (C13) 0.1303 C9 ber of articles published in international journals 0.2439 C10 ber of thesis winning of students (C15) 0.0762 C17 ber of patents (C16) 0.1371 C19 erformance of occupational refresher courses 0.0069 C11 upplemental budget of faculty research (C18) 0.0095 C1 udget of scholarships (C19) 0.0047 C3 udget of industry-university relations (C20) 0.0210 C5 udget of international relations (C21) 0.0210 C4 ber of international conferences (C22) 0.0241 C2 atisfaction degree of students and faculty (C23) 0.0181 C6 oyee turnover (C24) 0.0738 C7 atisfaction degree of industries (C25) 0.0406 C27 nit cost of each student (C26) 0.0001 C28 ate of tuition and schooling in school income (C27) 0.0002 C8 ate of government supplement in school overall e (C28) 0.0001 C26 Subsystem Appraisal Dimension Appraisal Criteria Core appraisal system Organizational development The employee turnover The percentage of promotion Academy performance Number of articles published in international journals Number of patents Number of winning student thesis External behavior Number of plans given by NSC The satisfaction degree of industries Financial support and budget planning Support appraisal system Fig. 6. A Pro-performance appraisal system (PPAS). 2116 J.-K. Chen, I-S. Chen / Expert Systems with Applications 37 (2010) 2108–2116 in Table 9. Last, using the above result, the impact-direction map is derived and is shown in Fig. 5. In Table 10, we summarize all of the above results. We propose a novel Pro-performance appraisal system (PPAS) in which the top seven most heavily weighted measurement criteria are extracted from the whole set of criteria. We argue that the operational per- formance appraisal and improvement considering only the highly weighted criteria will be better than that using all measurement criteria. Also, the content of organizational development (D7) is chosen due to its highest influential degree on the performance improvement conducting (as Fig. 6). In the PPAS we propose, a core appraisal system, containing three dimensions with seven mea- surement criteria, represents the key suitable items for all three university types to conduct operational performance appraisals and improvements with accuracy. This appraisal system supports financial and budget planning for all three university types to eval- uate appropriate financial decisions to support key initiatives and to control the budget. 5. Conclusions With increasing economic pressures and declining birth rates, Taiwanese universities are seeking ways to improve operational performance to acquire competitive advantages and to attract more students. Current performance appraisal models are inade- quate due to inconsistencies among different types of universities, inconsistencies in measurement criteria, and the treatment of interdependent criteria. To address these problems, we first re- viewed previous research and interviewed several higher educa- tion experts. Then, we combined a DEMATEL, and a fuzzy ANP to construct a Pro-performance appraisal system (PPAS), considering the interdependence and the relative weights of measurement cri- teria and the characteristics of three university types. The PPAS will be helpful in performing future performance appraisals and suggesting improvements for all three university types. References Cameron, K. S. (1978). Measuring organizational effectiveness in institutions of higher education. Administrative Science Quarterly, 23(4), 604–632. Department of Higher Education (2004). Operation book. Taipei: Taiwan Assessment and Evaluation Association. Fairweather, J. S. (2000). Diversification or homogenization: How markets and governments combine to shape American higher education. Higher Education Policy, 13, 79–98. Gabus, A., & Fontela, E. (1972). World problems, an invitation to further thought within the framework of DEMATEL. Geneva, Switzerland: Battelle Geneva Research Center. Huang, J. J., Tzeng, G. H., & Ong, C. S. (2005). Multidimensional data in multidimensional scaling using the analytic network process. Pattern Recognition Letters, 26(6), 755–767. Huang, J. J., Tzeng, G. H., & Ong, C. S. (2007). Marketing segmentation using support vector clustering. Expert Systems with Applications, 32(2), 313–317. Li, U. C. (2007). Enter university by 2.8 point each subject. Retrieved August 18, 2007, from http://news.msn.com.tw/print.aspx?id=210245. Liou, J. H., Tzeng, G. H., & Chang, H. C. (2007). Airline safety measurement using a hybrid model. Journal of Air Transport Management, 13(2007), 243–249. Lysons, A. F., Hatherly, D. J., & Mitchell, D. A. (1998). Comparison of measures of organizational effectiveness in U.K. higher education. The International Journal of Higher Education and Educational Planning, 36(1), 1–19. Meek, V. L. (2000). Diversity and marketisation of higher education: Incompatible concepts. Higher Education Policy, 13, 23–39. Mei, Y. F., & Lee, L. S. (2006). A study to develop performance indicators for colleges of technology in Taiwan. Journal of Education Research, 1(1), 25–67. Ministry of Education (2006). Department of Statistics. Retrieved November 23, 2008, from http://www.edu.tw/statistics/. National Science Council (2008). Fast connection. Retrieved November 23, 2008, from http://web1.nsc.gov.tw/mp.aspx. Saaty, J. T. (1980). The analytic hierarchy process. Columbus: McGraw-Hill. Taiwan Assessment and Evaluation Association (2006). Evaluation Bimonthly, 1, 48– 49. Tzeng, G. H., Chiang, C. H., & Li, C. W. (2007). Evaluating intertwined effects in e-learning programs: A novel hybrid MCDM model based on factor analysis and DEMATEL. Expert Systems with Applications, 32, 1028–1044. Yu, R., & Tseng, G. H. (2006). A soft computing method for multi-criteria decision making with dependence and feedback. Applied Mathematics and Computation, 180(2006), 63–75. http://news.msn.com.tw/print.aspx?id=210245 http://www.edu.tw/statistics/ http://web1.nsc.gov.tw/mp.aspx A Pro-performance appraisal system for the university Introduction Pro-performance appraisal system (PPAS) Research methods The decision-making trial and evaluation laboratory (DEMATEL) The fuzzy analytical network process (FANP) Fuzzy set theory Fuzzy number Fuzzy linguistic variable Analytic network process (ANP) Empirical study of Pro-performance appraisal system (PPAS) Forming an original performance appraisal system (OPAS) Evaluating the interrelationships of dimensions Weighting the criteria and constructing a Pro-performance appraisal system (PPAS) Conclusions References 
work_gllq5p3brbdkrgmr5mdj2jger4	----	Seasonal Fuzzy Time Series, Fuzzy C-means, Artificial Neural Network, Membership Degree, Air Pollution American Journal of Intelligent Sy stems 2013, 3(1): 13-19 DOI: 10.5923/j.ajis.20130301.02 A Novel Seasonal Fuzzy Time Series Method to the Forecasting of Air Pollution Data in Ankara Ozge Cagcag1, Ufuk Yolcu2, Erol Egrioglu1,*, Cagdas Hakan Aladag3 1Dep artment of Statistics, University of OndokuzM ay is, Samsun, 55139, Turkey 2Dep artment of Statistics, Giresun University, Giresun, 28000, Turkey 3Dep artment of Statistics, Hacettep eUniversity, Ankara, 06800, Turkey Abstract Fu zzy time series forecasting methods have been widely studied in recent years. This is because fuzzy time series forecasting methods are co mpatib le with fle xib le ca lculat ion techniques and they do not require constraints that exist in conventional time series approaches. Most of the real life time series e xh ibit periodical changes arising fro m seasonality. These variations are ca lled seasonal changes. Although, conventional time series approaches for the analysis of time series which have seasonal effect are abundant in literature, the number of fuzzy t ime series approaches is limited. In a lmost all of these studies, me mbership values are ignored in the analysis process. This affects forecasting performance of the approach negatively due to the loss of information as well as posing a situation that is incompatible with the basic features of fuzzy set theory. In this study, for the first time in literature, a new seasonal fuzzy time series approach which considers me mbe rship values in both identificat ion of fuzzy relat ions and defuzzification steps was proposed. In the proposed method, we used fuzzy C-means clustering method in fuzzification step and artificia l neural networks (ANN) in identification of fu zzy relation and defuzzification steps which consider me mbe rship values. The proposed method was applied to various seasonal fuzzy time series and obtained results were compared with some conventional and fuzzy time series approaches. In consequence of this evaluation, it was determined that forecasting performance of the proposed method is satisfactory. Keywords Seasonal Fuzzy T ime Se ries, Fuzzy C-means, Artific ia l Neura l Network, Me mbe rship Degree, Air Po llution 1. Introduction Nowadays, it is of vital impo rtance to make predict ions about the future in terms of planning and strategy formulat ion. This can be realized by accurate and realistic analysis of in formation and data that have e merged fro m past to present. This analysis can also be termed as time series analysis. Many different approaches have been proposed in literature for the analysis of time series. Each of these approaches has pros and cons. This leads to emergence of alternative methods which may enhance forecasting performance of the method namely fuzzy t ime series approaches which have a superior forecasting performance and which do not require hypothesis that are found in conventional approaches. On the other hand, due to the uncertainty that they contain, most of the time series encountered in real life should be considered as fuzzy time series. Fuzzy set theory proposed by Zadeh provides a basis for many studies as well as the fuzzy t ime series approaches [1]. The concept of fu zzy t ime series was first introduced by * Corresponding author: erole1977@y ahoo.com(Erol Egrioglu) Published online at http://journal.sapub.org/ajis Copyright © 2013 Scientific & Academic Publishing. All Rights Reserved Song and Chissom[2]. Fro m that day to this, fuzzy time series have been studied intensively and applied to many fie ld such as informat ion technologies, economy, finance, environment and hydrology. Since fuzzy time series approaches do not require constraints such as model assumption, the number of observations and normal distribution which e xist in conventional approaches and they are accordant with the use of fle xib le calcu lation methods, these approaches are becoming increasingly popular. Fu zzy time series approaches consist of three ma in steps as fuzzification, identification of fu zzy relat ions and defuzzificat ion. In literature, various approaches have been proposed for the improve ment of these steps. While some of these studies involve first order fuzzy time series forecasting models (such as[2-6]), others involve high order fuzzy time series forecasting models ([7-11]). A lthough, there are numerous approaches in literature fo r the analysis of fu zzy time series involving first and high degree fu zzy relat ions, few approaches have been proposed for the analysis of fuzzy time series involv ing seasonal fu zzy relat ions. However, most of the time series encountered in real life involve seasonal components. The use of seasonal fuzzy time series in the analysis of this type of fuzzy t ime series would be more rea listic and would provide superior forecas ting performance. In literature, the first approach for the analysis of seasonal 14 Ozge Cagcag et al. : A Novel Seasonal Fuzzy Time Series M ethod to the Forecasting of Air Pollution Data in Ankar a fuzzy t ime series was introduced by Song but this was not applied to any data[12]. In orde r to analy ze seasonal fuzzy time series, Egrioglu et al. proposed a hybrid fuzzy time series approach based on SARIMA and artificia l neural networks[13]. A lthough, the method proposed by Egrioglu et al. has some advantages, it uses universal set fragmentation in fuzzification step. These subjective judgments have negative impact on forecas ting performance of the method. In order to eliminate this proble m, Uslu et al. proposed an approach which does not require universal set frag mentation and which uses Fuzzy C-Means (FCM) in fuzzification step[14]. In all of these studies, only the fuzzy set having the highest me mbership value was considered in the analysis process and other fuzzy sets having lower me mbership values were ignored. This affects forecasting performance of the approach negatively due to the loss of informat ion as we ll as posing a situation that is incompatib le with the basic features of fuzzy set theory. In this study, we a imed to overcome the above mentioned factors which affect the forecasting performance of the method negatively in the analysis of a seasonal fu zzy time series. For this purpose, we proposed a new seasonal fuzzy time series fo recasting model wh ich considers me mbership values of each observations belonging to all fu zzy sets in both identification of fu zzy relat ion and defuzzificat ion steps. In the proposed model, we used FCM in fu zzification step and avoided subjective judgments and determined me mbe rship values with a systematic approach. We ut ilized ANN which considers all me mbe rship values in identification of fuzzy relat ion and prevented loss of informat ion and made use of fle xib le ca lculation ability of ANN. In defu zzificat ion step, artific ial neura l network which uses all me mbe rship values as input and real (crisp) values of time series as target was used for the first time in literature. The proposed method was applied to the amount of sulfur dio xide in Ankara and was co mpared with so me conventional time series approaches as well as fuzzy time series forecasting methods. In the second chapter, SARIMA models which were used in determining the model order, FCM wh ich was used in fuzzification step and ANN wh ich was used in determination of fu zzy relat ion and defuzzfication step will be introduced. Third chapter will deal with basic fu zzy t ime series concept and definitions. In the fourth chapter, proposed method and its algorith m will be g iven. In the fifth chapter, the proposed method will be applied to a rea l seasonal time series and obtained results will be presented with the other results obtained from other methods. In the last chapter, obtained results will be evaluated and discussed. 2. Review 2.1. SARIMA When a time series with 𝜇𝜇 mean, than the model is e xpressed in equation (1) t s t Dsds aBBZBBBB )()()()1()1)(()( Θ=−−−Φ θµφ (1) Model parameters can be given as follows; )1()( 1 p p BBB φφφ −−−=  (2) )1()( 1 q q BBB θθθ +++=  (3) )1()( 1 sP P s BBB Φ−−Φ−=Φ  (4) )1()( 1 sQ Q s BBB Θ++Θ+=Θ  (5) Detailed information on the model which is called seasonal autoregressive integrated moving average (SARIMA) and which is e xpressed as sQ)D,q)(P,d,SARIMA(p, can be obtained fro m Bo x and Jenkins[15]. 2.2. The Fuzz y C-Means (FCM) Clustering Technic al FCM clustering technical method is first introduced by Be zdek[16]. This is a most wide ly used fuzzy c lustering algorith m. FCM partit ions sets of n observation and each fuzzy c luster has a set center jv )1( ncc << . The me mbe rships of the observations are described by a fuzzy matrix µ with n rows and c columns in which n is the number of data objects and c is the number of clusters. ijµ , the ele ment in the 𝑖𝑖ith row and jth colu mn in µ , indicates the degree of association or me mbership function value of the ith object with the jth cluster. The characters of µ are as follows: [ ]0,1 1, 2, , ; 1, 2, ,ij i n j cµ ∈ ∀ = ∀ =  (6) 1 1 1, 2, , c ij j i nµ = = ∀ =∑  (7) 1 0 1, 2, , c ij j n j cµ = < < ∀ =∑  (8) The objective function of FCM algorith m is to minimize the equation (9) 1 1 ( , , ) c n ij ij j i J X V dβµ µ = = = ∑∑ (9) where, ij i jd x v= − (10) in wh ich, )1( >ββ is a scalar termed the weighting e xponent and controls the fuzziness of the resulting clusters and ijd is the Euclid ian distance from object ix to the cluster center jv . In this method, minimizing is done by an iterative a lgorith m. In each repetition the values of ijµ and jv are updated by the formulas given in equation (11) and equation (12). American Journal of Intelligent Sy stems 2013, 3(1): 13-19 15 1 1 n ij j j n ij i i x v β β µ µ = = = ∑ ∑ (11) ( ) 2 1 1 1 ij c ij ikk d d β µ − = =       ∑ (12) 2.3. Fee d For war d Ne ural Network Artificia l neural networks (ANN) can be defined as the mathe matica l a lgorith m that is inspired by the biological neural networks[17]. Art ific ial neural networks are much more d ifferent than biological ones in terms of their structure and ability[18]. Artific ia l neural networks compose of a mathe matica l mode l[19]. The learning capability of an artific ial neuron is achieved by ad justing the weights in accordance to the chosen learning algorith m. The basic architecture consists of three types of neuron layers: input, hidden, and output layers. In feed-forward networks, the signal flow is fro m input to output units, strictly in a feed-forward direct ion. Artific ial neura l network architectures are characterized by the follo wing attributes: Number of Layers: The a rtific ial neurons are arranged in an input layer, one or mo re h idden layers, and an output layer. Number of Neurons: The art ificia l neural network has to learn the features of the series for the analysis and forecasting of a fu zzy t ime series. As the number of neurons in the input and output layers are determined by the tra ining patterns, the number of neurons in the hidden layers can then be chosen arbitrarily (see Fig. 1). More a rtific ial neurons implies more we ighting matrices. Thus, fro m c lassical fie lds of application of artificia l neural networks (e.g., pattern recognition), the well-known proble m of over fitting must be considered. Fi gure 1. Archit ect ure of mult ilayer feed forward neural net work Activation Function: The proper selection of activation function that enables curvilinear matching between input and output units, significantly affect the performance of the network. Method of Training: The learning situations in neural networks may be classified into three d istinct sorts. These are supervised learning, unsupervised learning, and reinforce ment learn ing. In supervised learning, an input vector is presented at the inputs together with a set of desired responses, one for each node, at the output layer. The most wide ly used one is Back Propagation algorith m which updates weights based on the difference between ava ilab le data and the output of the network. Lea rning para meter which is used in back propagation algorithm and which can be taken fixedly or updated in the algorith m dyna mica lly, plays an important role in reaching optimal results. 3. Fuzzy Time Series The definition of fu zzy time series was firstly introduced by Song and Chissom[2]. Basic definitions of fuzzy time series not including constraints such as linear model and observation number can be given as follo ws; Definiti on 1 Fuzzy t ime series. Let ),2,1,0,)(( =ttY , a subset of real numbers, be the universe of discourse by which fuzzy sets )(tf j are defined. If )(tF is a collection of ),...(),( 21 tftf then )(tF is called a fuzzy t ime series defined on )(tY . Definiti on 2 First order seasonal fuzzy time series forecasting model. Let )(tF be a fu zzy t ime series. Assume there e xists seasonality in { })(tF , first order seasonal fuzzy t ime series forecasting model: )()( tFmtF →− (13) where m denotes the period. Definiti on 3 High order fu zzy t ime series forecasting model. Let )(tF be a fu zzy t ime series. If )(tF is caused by )2(),1( −− tFtF ,…, and )( ntF − , then this fuzzy logical relationship is represented by )()1(),2(),...,( tFtFtFntF →−−− (14) and it is called the nth order fuzzy t ime series forecasting model. Definiti on 4 First order b ivariate fu zzy t ime series forecasting model. Let F and G be two fu zzy time series. Suppose that ( 1) iF t A− = , kBtG =− )1( and jAtF =)( . A bivariate fuzzy logica l re lationship is defined as iA , jk AB → , where ki BA , are re ferred to as the left hand side and jA as the right hand side o f the b ivariate fu zzy logical re lationship. Therefore, first order bivariate fu zzy time series forecasting model is as follo ws: )()1(),1( tFtGtF →−− (15) 16 Ozge Cagcag et al. : A Novel Seasonal Fuzzy Time Series M ethod to the Forecasting of Air Pollution Data in Ankar a Definiti on 5 High order partia l bivariate fu zzy time series forecasting model. Let F and G be two fu zzy time series. If )(tF is caused by 1 1( ),.., ( ), ( ),k kF t m F t m F t m−− − − 1 1( ),.., ( ),lG t n G t n −− − )( lntG − , where im ),..,2,1( ki = and jn ),..,2,1( lj = are integers kmm <<≤ ...1 1 , lnn <<≤ ...1 1 then this FLR is represented by; )( )(),(),...,( ),(),(),...,( 11 11 tF ntGntGntG mtFmtFmtF ll kk →    −−− −−− − − (16) 4. Proposed Method Although, there are nume rous fuzzy t ime series approaches in literature, the number of approaches aiming at analyzing seasonal fuzzy time series which are frequently encountered in real life and wh ich inc lude seasonal components are limited. The first model proposed by Song involves one variable which belongs to only one period[12]. The approach proposed by Egrioglu et al. determines me mbe rship values of each observation belonging to fuzzy sets objectively[13]. A lthough, Uslu et a l. proposed a subjective judgment-free approach, she used set number representing the fuzzy set having the highest me mbership value of observations in identification of fu zzy re lations and defuzzificat ion[14]. Th is poses a situation which is incompatib le with fu zzy set theory as well as a ffecting the forecasting performance of the method negatively due to the loss of information. In this study, we proposed a new seasonal fuzzy time series forecasting model which does not require subjective judgments in all analysis processes and which uses SARIMA in determination of the model, FCM in fu zzification and ANN in defu zzificat ion steps. The advantages of the proposed model are as follows; • The proble m of determining the mode l o rder was eliminated by using SARIMA and delayed variab les in the model was determined systematically. • Subjective judgments were avoided by using FCM in fuzzification step and me mbership values which are compatible with the model we re determined by a systematic infrastructure. • Again, in the determination of fuzzy relat ions, the problem re lated to the number of input of ANN was eliminated by co-clustering of delayed variables data set and was limited by the set number. • Input and target values of ANN which a re used in the determination of fuzzy relat ion are not the set number but the me mbe rship values obtained from FCM. Thus, the approach becomes more realistic in e xposing fuzzy re lations in fuzzy time series. • In order to prevent loss of information, for the first time in literature me mbership values were used in fuzzification step. The algorithm of the proposed method in this study is given below; Algorithm Step 1 The model order is defined by SARIMA The time series concerned is analyzed by Bo x-Jen kins method after the model order is defined. Then we obtain residuals series )( ia . As an illustration let us suppose we have defined the model as SARIMA (1,1,0)(0,1,1)12 via Bo x-Jenkins method. This implies that tX will be a linear combination of the corresponding lagged variables. That is, ),,,,,( 1214131221 −−−−−−= ttttttt aXXXXXfX (17) Therefore, ),( lk representing the order of the model and the parameters kmm ,,1  and 1, , ln n are determined based on the inputs of the SARIMA model. Accordingly k and l are defined as 5 and 1 respectively. Then the model will be thlk ),( -order partia l b ivariate fu zzy time series forecasting model and the fuzzy relat ionship can be given as follow; )( )12(),14(),...,13( ),12(),2(),...,1( tF tGtFtF tFmtFtF →    −−− −−− (18) This imp lies ,14,13,12,2,1 54321 ===== mmmmm ,121 =n )(tF denotes the fuzzified time series tX and )(tG denotes the fuzzified residual series ta . Step 2 Data set of lagged variab les is created. Depending on the model order defined in previous step, for each time series wh ich should be included in the model tX , and residual series ta for each lagged variab les are lagged less than order of lagged variables and data set is created. In other words, when a model given in equation (18) is considered, lagged variables data set will include 111312111 ,,,,, −−−−− tttttt aXXXXX . Step 3 Data set of lagged variab les is clustered via FCM. The number of fuzzy set is determined with c where nc ≤≤2 and n is the nu mber of observation. Data set which covers the delays in times series is clustered via FCM clustering method. Thus, fuzzy set centers for each lagged variables constituting data set and me mbership values showing order of observations belonging to fuzzy sets for each observation are obtained. In this step, fuzzy sets are sorted according to set centers represented with , 1, 2, ,rv r c=  and , 1, 2, ,rL r c=  fu zzy sets are obtained. Step 4 Fu zzy re lations are determined via Feed Forward Artificia l Neura l Net works (ANN). The number of neurons in input and output layer of feed forwa rd artific ia l neural network used in determining fu zzy relations equals to number of fu zzy set )(c . The number of neurons in hidden layer is determined by trial and error. He re, American Journal of Intelligent Sy stems 2013, 3(1): 13-19 17 the point to take into consideration is that hidden layer unit number should be selected in a way that not losing generalization ab ility of feed forward a rtific ial neural network. The a rchitecture of feed forwa rd artificia l neural network having two hidden layers for a model including seven sets is presented in Figure 2. In Figure 2, ))(( tDS rL µ representsthe me mbership value of lagged data set belonging to thi fuzzy set at t time. Moreover, while me mbership value of observation of lagged data set belonging to c number fu zzy set at t time constitutes the inputs of ANN; me mbe rship value of observation of lagged data set belonging to c number fu zzy set at t time constitutes the outputs of ANN. In all layers of feed forwa rd artific ia l neural networks which is used in determining fu zzy re lation and whose architectural structure is e xe mp lified above, logistic activation function given in (19) equation is used. 1( ) (1 exp( ))f x x −= + − (19) Feed forwa rd artific ial neura l networks are trained according to Levenberg-Marquardt learning a lgorith m and optima l weights are obtained. Trained artificia l neural network lea rned the relation between consecutive time series observations and me mbership values of sets. Fi gure 2. Archit ect ure of feed forward art ificial neural net work for three set s Step 5 De fuzzification of forecasts. In order to obtain real forecasts of fuzzy time series at t time, me mbership values of observations belonging to fuzzy sets at t time depending on , 1, 2, ,rv r c=  fuzzy set center which was obtained from FCM method were determined and then these me mbership values were entered to feed fo rwa rd art ific ial neural networks as inputs and thus outputs of feed forward art ificia l neura l networks are created. These outputs represent forecast of observation at t time . A architecture of feed forward a rtific ial neural network for three sets is given Figure 3. Fi gure 3. A archit ect ure of feed forward art ificial neural net work in defuzzificat ion 5. Application The proposed method was applied to time series of “the amount of sulfur dio xide in Ankara p rovince between March 1994 and April 2006 (ANSO)”. The graph of ANSO time series is presented in Figure 4. In order to evaluate the performance of the proposed method, the last 10 observations were taken as test set and obtained results were co mpared with some conventional and alternative time series methods. In the applicat ion, in order to determine the order of fu zzy time series forecasting model, crisp time series is analyzed using Box-Jenkins method and optima l SA RIMA model is determined and residual time series ( )ta as well as tX time series are obtained. In this step, optimal model fo r ANSO t ime series was SARIMA (1,1,0)(0,1,1)12. As a linear function of tX , this mode l can be e xpressed as; 1 2 12 13 14 12( , , , , , )t t t t t ttX f X X X X X a− − − − − −= (20) Thus, the model will be (5,1)th order partia l high order fuzzy time series forecasting model where 5=k and 1=l . This model can be e xpressed as; )( )12(),14(),...,13( ),12(),2(),...,1( tF tGtFtF tFmtFtF →    −−− −−− (21) Fi gure 4. The t ime series dat a of t he amount of SO2 in Ankara After determining the model order o f partia l high order model, lagged variables data set for each lagged variable that should be included in the model is created. Lagged variables data set for (5,1)th order partia l model is created using 1 11 12 13 11, , , , ,t t t t t tX X X X X a− − − − − lagged variables. Here, it must be noted that lagged variables data set consists of one step leaded variable in partia l high order fuzzy time series forecasting model given in (20). Created data set is clustered via FCM. Clustering is applied to all lagged variable data sets together. In this step, data set is clustered by shifting the number of sets 5 to 15. Me mbership values of observations belonging to each fuzzy set are also determined via FCM method. The re lationship between these me mbership values, in other words, the nu mber of neurons in hidden layer of feed forwa rd artific ia l neuron network wh ich is used in determining fu zzy relat ion were shifted between 1 and 15. In the light of this informat ion, 1651511 =× different analyses were done and Root Mean Squared Error (RMSE) 0 50 100 150 18 Ozge Cagcag et al. : A Novel Seasonal Fuzzy Time Series M ethod to the Forecasting of Air Pollution Data in Ankar a and Mean Absolute Percentage Error (MAPE) were used as performance evaluation criteria . ( ) 1 1 ˆ T t t t RMSE x x T = = −∑ (22) 1 ˆ1 T t t t t x x MAPE T x= − = ∑ (23) Where ˆtx , and ˆtx , T represent crisp time series, defuzzified forecasts, and the number of forecasts, respectively. The algorith m of the proposed method is coded in Matlab version 7.9. In consequence of all analyses, the best forecasting performance was obtained in the case in which the nu mber of set is 14, the hidden layer unit number is 6 in the determination of fuzzy relat ion stage and the hidden layer unit number is 2 in the defu zzification stage. Results obtained fro m the proposed method and results of some other methods are summarized in Table 1. Table 1 c learly shows the superior performance of the proposed method in co mparison with conventional time series approaches as well as seasonal fuzzy t ime series approaches with respect to three criteria. Additionally, graph of the forecasts obtained fro m the proposed model with real values are given in Figure 5. When Table 1 and Figure 5 a re analyzed together, a ll the advantages as well as the superior forecasting performance of the proposed method can be seen easily. Table 1. Result of methods Data Set S ARI W ME [12] [13] [14] Th e 21 22.93 15.40 41.6 20.0 22.7 24.1713 27 22.35 16.11 27.5 30.0 22.7 24.1326 25 23.61 17.77 41.6 20.0 22.7 24.0786 28 28.81 25.12 41.6 30.0 22.7 24.1375 38 46.97 41.11 41.6 30.0 42.0 38.6718 45 54.62 46.12 46.7 50.0 42.0 39.4255 38 58.13 49.80 45.0 40.0 42.0 39.4255 36 46.99 44.24 46.7 30.0 42.0 39.4255 24 37.85 31.96 46.7 30.0 22.7 20,7401 22 24.76 18.39 27.5 20.0 22.7 26,5369 RMSE 9.62 7.11 12.7 4.56 3.66 3.32 MAP E 0.23 0.22 0.40 0.13 0.11 0.10 Fi gure 5. The graph of the result s obt ained from the proposed method and real t ime series 6. Discussion and Conclusions Diffe rent approaches have been proposed for the forecasting problemswh ich constitute an important ro le in future planning and strategy formulat ion. Fuzzy t ime series forecasting methods have attracted much attention in recent years. Although, numerous first and high order fuzzy time series forecasting models have been proposed in literature, these models are insuffic ient in the analysis of seasonal time series which are frequently encountered in real life. The approaches proposed in literature have specific outstanding features as well as some insufficiencies. The most significant disadvantage of these models is that they require subjective judgments and ignore me mbership values representing the degree of observations belonging to fuzzy sets in analysis process. This affects forecasting performance of these approaches negatively as well as posing a situation that is incompatib le with the basic features of fuzzy set theory. Seasonal fuzzy t ime series forecasting method proposed in this study eliminates this problem by considering the me mbe rship value in the determination of fuzzy relat ion and defuzzificat ion stages and presents fuzzy re lations mo re realistically. It is evident that partial high order seasonal fuzzy time series forecasting method which is proposed in this study and in which model order was determined via SARIMA and ANN was used in determining fu zzy re lations and defuzzification stages has some advantages and exhib its superior forecasting performance.It should be noted that these results are obtained for the para meter sets given above and ANSO t ime series e xa mined in the study. For instance, if the length of test set is shifted, the results can change or similarly if these parameter sets are used for other time series, the obtained results can change. Therefore, the obtained results are valid for only these parameter sets and this time series. In orde r to reach general results, a co mprehensive simu lation study has to be made. However, it is very hard to perform such a simulation study since there are many types of time series and many para meter co mb inations. REFERENCES [1] Zadeh, L. A., 1965, Fuzzy Sets, Inform and Control, 8, 338-353. [2] Son g Q., and Ch issom, B. S., 1993, Fuzzy time series and its models, Fuzzy Sets and Sy stems, 54, 269-277. [3] Son g Q., and Chissom, B. S., 1993, Forecasting enrollments with fuzzy time series- Part I, Fuzzy Sets and Sy stems, 54, 1-10. [4] Son g Q., and Chissom, B. S., 1994, Forecasting enrollments with fuzzy time series Part II, Fuzzy Sets and Sy stems, 62, 1-8. [5] Chen, S. M ., 1996, Forecasting enro llments based on fuzzy time-series, Fuzzy Sets and Sy stems, 81, 311-319. 20 25 30 35 40 45 1 2 3 4 5 6 7 8 9 10 data set the proposed method American Journal of Intelligent Sy stems 2013, 3(1): 13-19 19 [6] Yolcu, U., Egr io glu, E., Uslu, V. R., Basar an, M . A., and Aladag C. H., 2009, A New App roach for Determinin g the Len gth of Intervals for Fuzzy Time Ser ies, App lied Soft Comp uting, 9, 647-651. [7] Chen, S. M ., 2002, Forecasting enrollments based on high order fuzzy time series, Cy bernetics and Sy stems, 33, 1-16. [8] Aladag, C. H., Basaran, M . A., Egrio glu, E., Yolcu, U., and Uslu, V.R., 2009, Forecasting in high ord er fuzzy time series by using neural networks to define fuzzy relations, Exp ert Sy stems with App lications, 36, 4228-4231. [9] Egr io glu, E., Aladag, C. H., Yolcu, U., Uslu, V. R., and Basaran, M .A., 2009, A new app roach based on artificial neural networks for high order multivariate fuzzy time series, Exp ert Sy stems with App lications, 36, 10589-10594. [10] E. Egr io glu, V. R. Uslu, U. Yolcu, M . A. Basaran, and C. H. Aladag, A new app roach based on artificial neural networks for high order biv ariate fuzzy time series, J.M ehnen et al. (Eds.): App lications of Soft Comp uting, AISC 58, Sp ringer- Ver lag Ber lin Heid elber g, 265-273, 2009. [11] Egr io glu, E., Aladag, C. H., Yolcu, U., Uslu, V. R., and Basaran, M .A., 2010, Finding an op timal interval length in high ord er fuzzy time series, Exp ert Sy stems with App lications, 37, 5052-5055. [12] Son g, Q., 1999, Seasonal forecasting in fuzzy time series, Fuzzy Sets and Sy stems, 107(2), 235. [13] Egr io glu, E., Alad ag, C. H., Yolcu, U., Basaran, M .A., and Uslu, V. R., 2009, A new hy brid app roach based on SARIMA and p artial high order bivariate fuzzy time series forecasting model, Exp ert Sy stems with App lications, 36, 7424-7434. [14] V. R. Uslu, C. H. Alad ag, U. Yolcu, and E. Egr io glu, A n ew hy brid app roach for forecasting a season al fuzzy time series, 1st International Sy mp osium on Comp uting In Science & Engineer in g, Izmır –Turkey , 2010. [15] G. E. P. Box and G. M. Jenkins, Time series analy sis: Forecasting and control, San Francisco: CA: Holdan-Day , 1976. [16] J. C. Bezdek, Pattern recogn ition with fuzzy objective function algorithms, NY: Plenu m Press, 1981. [17] S. Gunay , E. Egrio glu E, and C. H. Alad ag, Introduction to single variable time series analy sis, Ankara, Turkey : Hacettep e University Press., 2007. [18] J. M . Zurada, Introduction of artificial neur al sy stems, St. Paul: West Publishin g, 1992. [19] Zhang, G., Patuwo, B. E., Hu, Y. M ., 1998, Forecasting with artificial neural n etworks: the state of the art, International Journal of Forecasting, 14, 35-62. 1. Introduction 2. Review 3. Fuzzy Time Series 4. Proposed Method 5. Application 6. Discussion and Conclusions 
work_gowycj3ytfhm5pn72bsssetr7m	----	NASA Technical Reports Server (NTRS) 
work_gqak356lgvhjrj626xo3zo5m7u	----	1 Citation: Nguyen, K.A. Stewart, R.A. Zhang, H. (2013) An autonomous and intelligent expert system for residential water end-use classification, Expert Systems with Applications, http://dx.doi.org/10.1016/j.eswa.2013.07.049. AN AUTONOMOUS AND INTELLIGENT EXPERT SYSTEM FOR RESIDENTIAL WATER END-USE CLASSIFICATION Abstract Intelligent metering technology combined with advanced numerical techniques enable a paradigm shift in the current level of water consumption information provision that is available to the customer and the water business. The aim of this study was to develop an autonomous and intelligent system for residential water end-use classification that could interface with customers and water business managers via a user-friendly web-based application. Water flow data collected directly from smart water meters includes both single (e.g., a shower event occurring alone) and combined (i.e., an event that comprises several overlapping single events) water end use events. The authors recently developed intelligent algorithms to solve the complex problem of autonomously categorising residential water consumption data into a registry of single and combined events using a hybrid combination of techniques including Hidden Markov Model (HMM), Dynamic Time Warping (DTW) algorithm, time-of-day probability functions, threshold values and various physical features. However, the issue still remained, which is the focus of this current paper, on how to integrate self-learning functionality into the visioned expert system, in order that it can learn from newly collected datasets from different cities, regions and countries, to that collected for the training data. Such versatility and adaptive capacity is essential to make the expert system widely applicable. Through applying alternate forms of HMM and DTW in association with a frequency analysis technique, a suitable self-learning methodology was formulated and tested on three independent households located in Melbourne, Australia with a prediction accuracy of between 80-90% for the major end-use categories. The three principle flow data processing modules (i.e. single and combined event recognition and self-learning function) were integrated into a prototype software application for performing autonomous water end-use analysis and its functionality is presented in the latter sections of this paper. The developed expert system has profound implications for government, water businesses and consumers, seeking to better manage precious urban water resources. 2 Key words: water end-use event, water micro-component, residential water flow trace disaggregation, hidden markov model, dynamic time warping algorithm, gradient vector filtering, adaptive analysis, adaptive function, water demand management 1. Introduction Following a long-standing drought for the second half of the last decade across most of Australia, most capital cities introduced a portfolio of water demand management strategies and constructed capital intensive rain-independent bulk supply sources to ensure the provision of a secure water supply (Willis et al., 2009a). Residential water consumption is often dependent on the water using fixtures or appliances within a dwelling, the household makeup, the regional location and a plethora of socio-demographic influences. A study of end-use water consumption aids water planners and consumers to identify where and when water is used in a household and hence, assists in driving proactive reductions in consumption (Loh and Coghlan, 2003; Stewart et al. 2010; Makki et al., 2011). However, the existing water end-use classification techniques require an extensive use of human resources to collect a combination of water use behaviours and appliance/fixture stock inventory data through a household audit followed by 2-3 hours of analyst time for each home (Stewart et al., 2011; Beal and Stewart, 2011). Presently, water end use or micro-component studies are restricted to the research domain, since it is not economically viable to complete citywide studies due to resource intensity of the flow data classification process. Intelligent and autonomous end use classification firmware is required along with bold large-scale roll-outs of high commercially available high resolution smart water meters in order to bring this level of water consumption information to the masses. Currently, an increasing number of smart water metering technologies have been introduced to the market. Such metering devices embrace two distinct elements: meters that use new technology to capture water use information and communication systems that can capture and transmit real-time water use information (Stewart et al., 2010). These forms of smart metering technology can provide total consumption data to the customer and utility at high levels of resolution; however, they fail to disaggregate this data into its end-use use categories. In the present study, an attempt to automate the domestic water end-use classification process and, thus, to enhance current practices in the urban water industry is required, and a robust hybrid model that employs HMM, DTW and event probability techniques is developed. The proposed system will allow individual consumers to log into their user-defined water 3 consumption web page to view their daily, weekly, and monthly consumption tables as well as charts on their water demand across major end-use categories (e.g., leaks, clothes washer, shower, irrigation). This system can rapidly alert customers of leak events so that they can immediately be addressed rather than waiting for the present slow feedback process from the traditional metering technology (e.g., the quarterly bill). The system will also benefit water businesses by rapidly providing water end-use reports of any desired property or suburb, thereby empowering them to develop more targeted conservation programs in water scarcity periods (e.g. Willis et al., 2011a; 2011b), improved water demand forecasting (e.g. Makki et al., 2011) and optimised pipe network modelling (e.g. Carragher et al., 2012; Beal and Stewart, 2013). Figure 1 summarises below the three key stages in the development of this system: • Stage 1: Develop a non-adaptive intelligent model that autonomously disaggregates collected water flow trace signatures that were collected from the intelligent water meters into a categorised registry of water end-use events (Nguyen et. al., 2013a, 2013b). • Stage 2: Equip the model with adaptive capabilities that enable it to interpret untrained water end-use signature traces, thereby allowing it to adapt to new situation context (e.g. different city to training dataset). • Stage 3: Develop an intelligent and user-friendly expert system and prototype firmware for use by consumers and businesses. [INSERT FIGURE 1] 2. Background 2.1. Existing water metering process and new paradigm Water consumption readings are usually recorded manually on a quarterly or half yearly basis. Under most situations, a whole year’s worth of water consumption data is described by only two to four data points in the water businesses billing system. Conventional water meters count each kilolitre of water as it passes through the meter and do not have the ability to record when (i.e., the time of day) and where the consumption takes place (e.g., washing machine, leaks) (Stewart et al., 2011). These systems produce limited and delayed water consumption information. The current water metering system does not typically provide real-time or continuous/frequent water consumption data, and in cases where it does, it does not provide a sufficient level of data resolution to allow water end-use event categorisation. While real-time 4 or near real-time water consumption data provisioning is now commercially viable with current smart metering technology, there is presently no firmware that can autonomously disaggregate this flow data into the ‘richer’ water end use categories of consumption. Until, such firmware is developed, powerful water end use information will be contained to expensive research studies (e.g. Beal and Stewart, 2011). 2.2. Intelligent system development using various pattern recognition techniques To overcome these limitations, intelligent metering technology is united with advanced pattern recognition techniques to enable a paradigm shift in the current level of water information provision available to the customer and water business. The aim of this project is to develop an autonomous and intelligent system for residential water end-use classification through the employment of various mathematical techniques, namely HMM, DTW and frequency analysis as presented below. A hidden Markov model (HMM) is a statistical Markov model in which the system being modelled is assumed to be a Markov process with unobserved (hidden) states. An HMM can be considered to be the simplest dynamic Bayesian network, which is one of the most popular techniques in the field of hand writing and speech recognition (Ephraim and Merhav, 2002). Principal theories and typical applications of this technique have been presented in Baum and Petrie (1966), Starner and Pentland (1995), Baum et al (1970), Cho et al (1995), Ghahramani and Jordan(1997), Chien and Wang (1997), Satish and Gururaj (2003) or Tapia (2004). In this study, HMM was utilised as the main classifier for water end use classification decision making. Another important mathematical tool is the Dynamic time warping (DTW) algorithm, which is a popular method for measuring the similarity between two time series of different lengths. In general, this task is performed by finding an optimal alignment between two series with certain restrictions. The sequences are extended or shortened in the time dimension to determine a measure of their similarity independent of certain non-linear variations in the time dimension (Myers and Rabiner, 1981). This technique has been widely applied in prototype selection (e.g. Nguyen et al., 2011), pattern recognition (e.g. Myers and Rabiner, 1981; Muller, 2007; Rabiner and Juang, 1993; Sakoe and Chiba, 1978; Manmatha and Srimal, 1999; and Marquez, 2001) or word image searching (Manmatha and Rath, 2002). DTW played an important role in this 5 study because it was utilised to the task of grouping similar unclassified events together to prepare for adaptive analysis. Probability analysis was also applied in this study. For the purpose of this study, frequency histogram data distributions were formulated from the training data to examine the likelihood of event characteristics occurring. A histogram comprises tabular frequencies, shown as adjacent rectangles, which are erected over discrete clusters (bins), with an area equal to the frequency of the observations in the interval. The height of a rectangle is also equal to the frequency density of the interval, i.e., the frequency divided by the width of the interval (Pearson, 1895). For example, suppose that we are given a vector that contains a volume of 10 events in litres, as follows: 𝒗 = [4.63, 4.57, 4.59, 4.63, 4.69, 4.75, 4.71, 4.79, 4.92, 4.76]. Then, the volume distribution using histograms with different numbers of clusters are presented in Figure 2. [INSERT FIGURE 2] In the present study, a distribution of all of the physical characteristics of a group of events (i.e., the volume, duration and flow rate of each water end use category) will be determined using the histogram method with 5 clusters, from which the representative value of each group can be obtained by selecting the entry that has the highest frequency. It should be noted that the selection of 5 clusters is based on the fact all events in each unclassified group after the grouping process will have approximate volume, flow rate and duration (i.e. the values of each feature do not spread over a wide range) regardless the end use category they actually belong to; therefore, the utilisation of 5 clusters is sufficient to determine the representative values. Given a group extracted from the tested home that contains 10 events whose volumes are presented above, the most typical volume representing this group is 4.6 L because it attains the highest frequency of 4 when using a histogram of 5 clusters. 3. Classification model development 3.1. Collected data for the study Data utilised for the development of the model is sourced from 252 residential households fitted with a smart meter and data logger and located in the urban south-east corner of the State of Queensland (SEQ), Australia, in both summer and winter for 2 years, 2010 and 2011. These households are consenting participants in the recently completed South-east Queensland 6 Residential End Use Study (SEQREUS) that was funded by the Queensland State Government (Beal and Stewart, 2011). A sample of properties is taken from the four key cities in this interconnected SEQ region, namely, Sunshine Coast Regional Council, Brisbane City Council, Ipswich City Council and Gold Coast City Council to use as a database for this study. The smart meters provided sufficient resolution (0.014L/pulse every five seconds) of water flow data to the household to complete a water end use or micro-component disaggregation process (i.e. each tap, shower, etc.). Participating households were also requested to participate in an appliance/fixture stock inventory audit and complete a questionnaire survey that was developed to assist in determining the socio-demographic characteristics and socioeconomic status of the households. All such data was required by the team on the SEQREUS in order to manually complete the water end use disaggregation process as well as for statistical analysis related to a number of objectives related to that study (e.g. Carragher et al. 2012; Beal et al., 2011a; Beal and Stewart, 2013). This studies budget was in excess of $US1,000,000 with a reasonable proportion of that budget assigned to water end use analysis process for a sample of 250 households, which is acceptable for a detailed research investigation but the disaggregation process needs to be automated for widespread application. Nonetheless, this extensive dataset of high resolution flow data and associated water end use event registry provided the training set for this study. 3.2. Stage 1: Non-adaptive classification model With the availability of data collected from SEQREUS, the building of an autonomous flow trace analysis system commenced (Figure 3). In Stage 1 of the study, a single event analysis module was developed to categorise all of the unclassified single events that occur in isolation. In this module, HMM, DTW and an event time-of-day probability function were applied to autonomously assign all of the single events to appropriate categories, with an average accuracy of 84.1% (Nguyen et al., 2013a)，which is slightly lower than that of combined event due to low recognition accuracy of bathtub and irrigation. Then, a combined event analysis (i.e., a group of concurrent single events) module, which remains one of the most complicated problems in the field of pattern matching, was developed. Several techniques were employed for splitting apart the various events in a combined water use event, including HMM, the Gradient Vector Filtering method and different probability functions that were extracted from the various physical features of the existing database (Nguyen et al., 2013b). The classification 7 outcomes have shown that approximately 88% of the combined events were accurately disaggregated into their end use components and then recognised. [INSERT FIGURE 3] 3.3. Stage 2: Adaptive classification model The classification model that was developed in Stage 1 was initially trialled in a different region (i.e. Melbourne) to that where the model was developed (i.e. different to the SEQ training data) to examine its versatility. Model accuracy dropped due to some fundamental causes, including, the presence of new end-use categories that have not been identified in SEQ (e.g., evaporative air conditioner) as well as some differences in water consumption behaviours for some of the end uses which may be due to a range of macro factors (i.e. different climatic conditions, government policy, etc.). To overcome this challenge, we needed to build some self-learning functionality into the model to make it more adaptive to different regions. Therefore, the objective of the research and focus of this paper was to integrate adaptive features into the model employing appropriate techniques. The establishment of this critical analysis module is articulated in the next section. 4. Adaptive classification model development 4.1. Overview of model architecture This function was developed to analyse all of the events that exhibit patterns which cannot be confidently recognised by the non-adaptive single and combined event modules. Figure 4 provides an overview of the overall analysis procedure for the adaptive model. [INSERT FIGURE 4] At the very first step for adaptive learning, the HMM threshold value, which is explained in the next section and is used to determine whether an event is classifiable by the existing non-adaptive analysis modules, will be applied when the model is operated in a new region. Classifiable events are initially analysed by these modules (Stage 2a), while all of the unclassifiable events are processed by a newly developed adaptive analysis unit (Stage 2b). At the end of this process, all of the unclassifiable events in stage 2b will be incorporated into the existing database to improve the HMM classifier. The advantages of the proposed technique in 8 comparison with other adaptive learning recognisers are the simple algorithms, the fast analysis time and the lower dependency on the existing database, which has been proven in a later section of this paper. A detailed technical development of this analysis module is presented in Figure 5. [INSERT FIGURE 5] The main objective of this analysis module is to address unclassified events that cannot be analysed by the existing non-adaptive model. The first required step is to group all of the events that are likely to belong to the same category together, using the DTW technique with various physical features that are extracted from each subjected event. The outcomes of this analysis step are several groups that contain similar unclassified events and a set of all of the events that cannot be assembled together. Grouped and ungrouped events are then analysed by HMM, DTW, the event time-of-day probability function and another set of physical parameters, which eventually results in all of the single events being classified and sometimes an additional set of unclassified events, which belong to a new end-use category. 4.2. New pattern identification using threshold values The threshold values that are applied in this analysis section were achieved through the training of the existing database that was collected in SEQ by using the HMM method. The determination of the threshold values for each end-use category can be explained as follows. Given that 𝑆𝑖 is a set of all single events that belong to category i, where 𝑖 = [1,2, … ,7] represents the shower, faucet, clotheswasher, dishwasher, toilet, bathtub and irrigation, respectively; 𝑆𝑖 is collected in SEQ and is used as a database to establish an HMM model to represent this category, which is denoted as 𝐻𝑀𝑀𝑖. 𝑇𝑉𝑖 is defined as the threshold value of category 𝑖 if 𝑇𝑉𝑖 is the minimum likelihood score that is achieved when using the 𝐻𝑀𝑀𝑖 to recognise 𝑆𝑖. As a result, when the model is operated in a different area, if one event is assigned to category 𝑖 by the existing single event analysis module but its likelihood is less than 𝑇𝑉𝑖, then it is considered to be an event with an unclassifiable pattern and will be set aside for further analysis. Following the comparison process against threshold values, a set that contains all of the unclassified events that require the application of a new analysis process is obtained. 9 4.3. Event grouping using DTW Given that 𝐀 = (𝐴1 , 𝐴2 , … , 𝐴𝑚 ) is a matrix that contains m unclassified events, and 𝐴𝑘 is the kth event of 𝐀. A DTW distance between 𝐴𝑘 and 𝐴𝑗, namely 𝐷𝑘,𝑗, is determined by using the DTW algorithm. By following the same process, a vector 𝐃 = (𝐷1,𝑗, … , 𝐷𝑖,𝑗, … 𝐷𝑚,𝑗) can be achieved to measure the similarity of all of the events to Event j, which is an arbitrary event that can be selected randomly from the subjected group. Nguyen et al. (2011) have found that, for 𝐴𝑘 and 𝐴𝑙, if � 𝐷𝑘,𝑗−𝐷𝑙,𝑗 𝐷𝑘,𝑗 � < 𝑒, where 𝑒 is defined as the threshold value in terms of similarity and j is an arbitrary event, then it is likely that events 𝐴𝑘 and 𝐴𝑙 have similar patterns. However, because of the fact that two completely different events could also have similar DTW distances to a reference event using the proposed method, additional parameters are required for the technique to be applicable to this specific study. In reality, one end-use event is described by four basic physical features, namely the volume (v), duration (t), maximum flow rate (qmax) and most frequent flow rate (qf) for all of the events in 𝑨. An aggregate DTW distance of Event 𝑘 to Event 𝑗, denoted as 𝐷𝐴𝑘,𝑗 , is determined as follows: 𝐷𝐴𝑘,𝑗 = 𝐷𝑘,𝑗 𝑣𝑘 𝑞𝑚𝑎𝑥𝑘 𝑡𝑘 𝑞𝑓𝑘 (1) Thus, if 𝑅𝐷 = � 𝐷𝐴𝑘,𝑗−𝐷𝐴𝑙,𝑗 𝐷𝐴𝑘,𝑗 � < 𝑒 (2) then events 𝐴𝑘 and 𝐴𝑙 will be assigned to the same group. At the end of this process, the first group (denoted as 𝑔1) that contains 𝑝1 similar events, whose relative differences in terms of DTW distances (i.e. 𝑅𝐷 in Equation 2) are less than the threshold value 𝑒, will be obtained. In the remaining (𝑚 − 𝑝1) events, the same process is then conducted to determine the second group, namely 𝑔2, which contains 𝑝2 similar events. This grouping process is repeated until no more groups can be achieved. The overall process will result in 𝑛 groups of events, denoted by 𝐺 = {𝑔1, 𝑔2 … 𝑔𝑛 }, which contain events that have similar patterns, and a set of all dissimilar events that cannot be gathered together. It can be seen from Equation (2) that the selection of the threshold value 𝑒 will affect the number of groups that are achieved after the grouping process, because all of the events that 10 have a relative difference of less than 𝑒 will be assembled. After considering the analysis time while the final classification accuracy is converged, 𝑒 = 0.1 has been adopted for this study. The next analysis step is to classify these grouped and ungrouped events. 4.4. Analysis of unclassified grouped events The classification of grouped events will be undertaken utilising the HMM technique in combination with other physical characteristics. An aggregate likelihood for each group to be classified into category 𝑖 can be determined by using the following proposed equation: 𝐿𝐿𝑖 = 𝑉𝑆𝑖 𝑇𝑆𝑖 𝑄𝑓𝑆𝑖 𝐻𝑀𝑀𝑆𝑖 (3) Where: • 𝑖 = [1, 2, … , 7] represents the shower, faucet, clotheswasher, dishwasher, toilet, bathtub and irrigation, respectively. • 𝑉𝑆, 𝑇𝑆 and 𝑄𝑓𝑆 are, respectively, the volume score, the duration score and the most frequent flow rate score, which are arbitrary numbers derived from the volume, duration and most frequent flow rate of the subjected group, and • 𝐻𝑀𝑀𝑆 is the representative HMM likelihood of the subjected group. from which an event can be classified to Category k if LLk is the maximum. 4.4.1. Determination of the required parameters The volume score, duration score, most frequent flow rate score and representative HMM score (i.e. 𝑉𝑆, 𝑇𝑆, 𝑄𝑓𝑆 and 𝐻𝑀𝑀𝑆 presented in Equation 3 are the four primary parameters that will be employed to aid the classification process. Given an unclassified group that contains 11 events (shown in Figure 6) and that was obtained after a grouping process while verifying the proposed technique against an independent home in Melbourne, the achievement of these features is described in the following key steps: [INSERT FIGURE 6] Step 1: Determine the volume, duration and most frequent flow rate of the classified events in all of the end-use categories that were achieved from the non-adaptive analysis, from which the 11 distribution range and distribution probability of the above-mentioned features of category 𝑖 can be obtained using the event probability histogram method. It should be noted at this step that a histogram with 10 clusters will be applied to shower, faucet, clotheswasher, bathtub and irrigation due to their widespread values, and 5 clusters for the other end use categories such as dishwasher and toilet because of their concentrated values. This step is undertaken only once for the recognition of all of the unclassified groups. The following example illustrates the process of obtaining these parameters for the dishwasher category based on all of the classified dishwasher events achieved through the existing non-adaptive single event analysis module (Table 1). It should be noted that the distribution probability for each feature is determined by using Equation 4. 𝐷𝑃𝑖 = 𝐹𝑖 𝑁𝑖 (4) Where: • 𝐷𝑃𝑖 is the distribution probability for category 𝑖, • 𝐹𝑖 is the frequency vector of category 𝑖 , and • 𝑁𝑖 is the number of events already classified into category 𝑖. [INSERT TABLE 1] Step 2: Determine the volume (𝑣), the most frequent flow rate (𝑞𝑓) and the duration (𝑡) of each event in the subjected group. From which, the representative volume, duration and most frequent flow rate for this group, denoted as 𝑉𝑟𝑒𝑝 , 𝐹𝑟𝑒𝑝 and 𝑇𝑟𝑒𝑝, can be obtained using the event probability histogram method with 5 clusters. As mentioned prior, the representative values are those that have the highest frequency. Steps 2 to 5 are illustrated in Table 2 over an example that is used to determine the aggregate likelihood of the above mentioned group to be assigned to dishwasher category. Step 3: Compare 𝑉𝑟𝑒𝑝 , 𝐹𝑟𝑒𝑝 and 𝑇𝑟𝑒𝑝 with the distribution of each end-use category to obtain the corresponding distribution probabilities, which are known as 𝑉𝑆, 𝑇𝑆 and 𝑄𝑓𝑆 and are presented in Equation 3. As presented in Table 2, the value of 𝑉𝑟𝑒𝑝 (4.64) falls into the range of 4.25-4.8 for the dishwasher (Table 1); therefore, the value for 𝑉𝑆4 is determined to be 42.8, which is the 12 corresponding distribution probability of this range (9/21). In the same way, the values of 𝑇𝑆4 and 𝑄𝑓𝑆4 are determined to be 26.5 and 66.7, respectively. However, if there is any category 𝑖 that has no classified single event achieved from the non-adaptive analysis process, which is very unlikely to happen, then the determination of the flow rate range and the corresponding distribution probability, as required in the previous step, cannot be undertaken for this category. As a result, the values of 𝑉𝑆, 𝑇𝑆 and 𝑄𝑓𝑆 will be considered to be 1, which means that the HMM likelihood of this group to be categorised as 𝑖 will not be magnified by any factor. In this case, the final recognition accuracy is only slightly affected because the verification process presented in section 5 has shown that the HMM method alone can explain most of the unclassified events correctly. Step 4: Determine the representative HMM likelihood for the subjected group, which is the HMM score that has the highest frequency when evaluated using a histogram method. Step 5: Determine the aggregate likelihood of the unclassified group. In this example, with the achieved values for 𝑉𝑆4, 𝑇𝑆4, 𝑄𝑓𝑆4 and 𝐻𝑀𝑀𝑆4, the overall likelihood of the group subjected to be classified as dishwasher is obtained through the utilisation of Equation 3. [INSERT TABLE 2] Following the same process from step 1 to 5, the likelihoods of this group to be assigned to other end uses such as shower, faucet, clotheswasher, dishwasher, toilet, bathtub and irrigation can be obtained. In this example, the subjected group was eventually assigned to the dishwasher category because the aggregate likelihood of this end use attains the maximum value. 4.4.2. Identification of a new end-use category In the context of this study, if the representative likelihood of the subjected group is 20 times less than the HMM threshold value of the category to which the group is assigned (i.e. 𝐻𝑀𝑀𝑆𝑖 < 0.05 𝑇𝑉𝑖), then this group is considered to be a new end-use category. This ratio of the threshold value was based on the analysis of several houses in Melbourne; however, for 13 identifying an appropriate value for the most accurate identification of a new end-use category, further research needs to be undertaken across a number of new regions to establish more rigorously founded criteria. If the prototype database expansion for a new end use category is proving difficult due to scarce examples of its use in households, then there is an opportunity to also supplement established prototypes with manually inputted ones that have been verified through the use of customer diaries or through fixture level sensors. The newly identified end use categories will then be incorporated into the existing resource so that the future system performance can be improved. 4.5. Analysis of ungrouped events The analysis of ungrouped events is conducted in a similar way to that for grouped events; however, modifications have been made to the formula for the establishment of the aggregate likelihood of one event, which is presented in Equation 5 below: 𝐿𝐿𝑢𝑖 = 𝑉𝑆𝑢𝑖 𝑇𝑆𝑢𝑖 𝑄𝑓𝑆𝑢𝑖 exp(𝑃𝑡) 𝐻𝑀𝑀𝑆𝑢𝑖 (5) Where: • 𝑉𝑆𝑢, 𝑇𝑆𝑢, 𝑄𝑓𝑆𝑢 and 𝐻𝑀𝑀𝑆𝑢 are the volume score, the duration score, the most frequent flow rate score and the representative HMM score of the subjected event. • 𝑃𝑡 is the probability index of event occurrence time of a day that was derived from the already classified events of the subjected household as explained in step 4 below. An in-depth assessment on the training database has found that most of the ungrouped events belong to the shower, faucet, abnormal toilet, bathtub and irrigation end-use categories, because they usually have highly variable patterns that cannot be gathered together. Therefore, the aggregate likelihood of an ungrouped event to be categorised as dishwasher and clotheswasher (i.e. 𝐿𝐿𝑢3 and 𝐿𝐿𝑢4) is zero. The achievement of 𝑉𝑆𝑢, 𝑇𝑆𝑢 and 𝑄𝑓𝑆𝑢 is similar to that for 𝑉𝑆, 𝑇𝑆 and 𝑄𝑓𝑆; however, 𝑉𝑟𝑒𝑝, 𝐹𝑟𝑒𝑝 and 𝑇𝑟𝑒𝑝 in this case are the volume, mode frequent flow and duration of the subjected unclassified event. To obtain 𝐿𝐿𝑢𝑖 for each ungrouped event, the following process is conducted: Step 1: Determine the distribution probability of all of the end-use categories based on all of the classified events, which also include the ones achieved from the analysis of the grouped events section (e.g., if there are 34 shower events obtained using the existing non-adaptive module and 20 events obtained from the grouped event analysis process, then the 14 determination of the range and distribution probability of the volume, duration and most frequent flow rate for this category is based on a total of 54 classified shower events). This process is also required once for the recognition of all of the ungrouped events. Again, histogram with 5 clusters will be applied to dishwasher and toilet, and 10 clusters to the other end use categories. Step 2: Determine the volume, the most frequent flow rate and the duration of the subjected event (i.e. 𝑉𝑟𝑒𝑝, 𝐹𝑟𝑒𝑝 and 𝑇𝑟𝑒𝑝) Step 3: Compare the volume, duration and mode flow rate of the subjected event with the distribution probability of each end-use category, to obtain the corresponding values for 𝑉𝑆𝑢, 𝑇𝑆𝑢 and 𝑄𝑓𝑆𝑢 Step 4: With the achievement of classified single events from all of the previous steps, a time of day probability index (𝑃𝑡) will be determined, which is shown in Figure 7 as an example for the tested home. It should be noted that the time of day probability index is suggested as an additional criterion for the classification process and is less critical than the other ‘classification’ inputs; therefore, it’s weighting contribution towards the overall likelihood score has been purposely kept limited. In this study, an exponential function is selected to limit the range of this magnifier factor to between 1 and 2.71, which corresponds to a time probability of 0% and 100%. For example, if an event occurred between 6-7 am, then its time-of-day (𝑃𝑡) probability to be classified as a shower is 0.0951 as presented in Figure 7 below. Step 5: Determine the HMM likelihood of the subjected event to be assigned to all of the end-use categories (i.e. 𝐻𝑀𝑀𝑆𝑢). Once all of the required parameters have been obtained, the aggregate likelihood of the unclassified data can be determined by using Equation 5. [INSERT FIGURE 7] 4.6. Adaptive learning procedure With the discovery of additional single water end use events from the above described adaptive analysis processes, the classification model can be enhanced through incorporating these new prototype event variants into the existing prototype database. This procedure will ensure that the prototype registry continuously grows and evolves to understand the new variants of water 15 end use flow signatures. The model updating process is described in a series of steps presented in Appendix 1 to determine a new HMM model )(λ . 5. Model calibration and verification The adaptive classification model has been verified against three random homes in a new Melbourne region using the proposed technique presented above. In this section, a comparison between the existing model developed for an SEQ application (i.e., a non-adaptive model) and the new model has also been undertaken by getting the SEQ formulated model to recognise three independent homes from the different city of Melbourne, Australia. It should be noted that homes 2 and 3 in this verification process have the presence of an evaporative air conditioner, which is a new end-use category that was not included in the existing prototype database since the region of SEQ has sufficient air humidity for running air conditioners. Detailed testing on these homes is displayed in Tables 3, both in terms of the number and respective volume of the events (denoted as ‘N’ and ‘V’, respectively). [INSERT TABLE 3] Applying the developed adaptive model to analyse the three Melbourne homes, the results were very promising, with an average recognition accuracy in terms of volume of at least 90% for the faucet, clotheswasher, dishwasher, toilet and irrigation (i.e., 90.1% 91.6%, 90.9%, 92.8% and 100%, respectively) and 81.5% for the shower. In terms of individual event recognition accuracy, it was lower (i.e. 81.5%, 89.7%, 89.1%, 91.4% and 86.2%) for the first five categories listed in Table 3. The classification of the bathtub end use category still remains a challenging problem, for which the accuracy indices are low both in terms of the volume (20.6%) and the specific events (40.5%) recognised. However, when considering the overall recognition accuracy for all end use events occurring within these three homes over the two week period, 85.7% of them were correctly classified using this autonomous recognition process. This is commendable given that these households were in a different region to the SEQ originated training dataset and new end use categories needed to be autonomously created by the adaptive model. Further testing also shows a considerable improvement of the new adaptive model compared to the existing model, which was built using SEQ training data (Figure 8). An increase in 16 accuracy has been experienced in most of the end-use categories, including shower, faucet, clothes washer, dishwasher and toilet. This enhancement can be explained by the fact that the original model for the SEQ region included the fixed boundary conditions for some of the physical characteristics that were derived from the SEQ database (e.g., the minimum shower volume is 7 litres) or applied the SEQ time-of-day probability information obtained from the SEQ training dataset. Therefore, the application of these features in Melbourne city has caused a minor reduction in categorisation accuracy. Moreover, when testing Home 2 and 3 using the SEQ-based model, most of the evaporative air conditioner events were misclassified as faucets or toilet, which resulted in a low accuracy in these end uses (i.e., when threshold values are utilised, the proposed model assigns evaporative air conditioner events into a “new end-use category” that does not affect the recognition accuracy of the other categories). [INSERT FIGURE 8] With a gradual expansion of the database through the adaptive learning process, the newly developed model can perform effectively in any new region without the requirement of a manual calibration if no end-use category exists. The new threshold values that are derived from the expanded database will identify the large majority of events that can be classified using the non-adaptive model, for which the effectiveness has been verified in Nguyen et al., (2013a, 2013b); the remaining variant events are left to the adaptive model. In case there is the presence of new end-use categories, new boundary values that determine whether a event belongs to an existing category or a new category should be re-identified and applied, as mentioned in section 4.4.2. 6. Residential water end use categorisation software application 6.1. Application outputs Through integrating and codifying the analytical processes contained in the single, combined and adaptive learning modules, a software application could be formulated that would be able to autonomously categorise remotely collected residential water consumption data received from smart meters into a repository of end use events. This software application offers different types of results presentations. Figure 9a shows the main interface, which provides important information to the customer, such as a summary of the classified volume of each end-use category during a specific period of time, which is supported by a detailed description of the start time, end time, volume, duration, maximum flow rate and most frequent flow rate of each 17 classified event. These results are directly achieved from the non-adaptive and adaptive analysis process presented in previous sections employing different mathematical techniques, such as HMM, DTW, Gradient Vector Filtering, time-of-day probability function or threshold values, etc. Each classified event is then plotted in a time series scale with different colours corresponding to different end-use categories. The application also allows all analysed results to be exported to an Excel file so that various statistical calculations and studies can be performed on the raw flow rate series of each individual water end use event for a particular household. Apart from the main interface, the analysis outcomes are also presented in terms of a pie chart showing the percentage of each contributing category, to give the user an instant overview of the end-use breakdown, and a bar chart that presents the average household water consumption in terms of litres per household per day (Figure 9b), which is achieved by taking the water consumed in each category divided by the total volume of water consumption in the analysis period. It should be noted that the current prototype software application and the outputs presented here are for the purposes of illustrating its functionality. Ultimately, the researchers seek to make the software embedded into the water businesses water consumption data collection repository, processing collected water use data from its entire customer fleet of meters autonomously and delivering that processed end-use information in a user-friendly form back to the customer via the web to their computer or phone. [INSERT FIGURE 9] 6.2. Optional manual adjustment functionality The present software prototype does not have the level of autonomous end use categorisation accuracy (>95%) considered necessary for commercial application. Therefore the present software prototype allows users to manually modify analysis decisions, such as changing, splitting or merging classified events. Any time that the user clicks on a classified event in the graphical figure, all of the physical characteristics of the event will be presented, with an option for manual modification (Figure 10a). The editing process is clearly demonstrated in Figure 10b, and once it has been finished, all of the edited events can be optionally updated into the existing database to improve classification accuracy in the future. This function was deemed a necessary inclusion in the present prototype software application, but ultimately this function 18 will be made redundant as more training data from a number of different regions enables the software to function with almost faultless accuracy. This is a key area of focus of the authors. 6.3.Daily end use diurnal demand functionality Another useful output of a water end use study is the daily end use diurnal demand graph (Figure 10c). This graph can be automatically created from the repository of classified end use events, and is highly beneficial to both consumers and water businesses seeking to better understand how residential water consumption is being used, at an end use level, across various significant days of the year (i.e. average weekday, average weekend day, peak day, average day peak month). This data is particularly useful for water infrastructure planning (e.g. water pipe network augmentation planning) as it informs network modelling engineers of the peak demand flow rates as well as the key end uses contributing to that peak demand (i.e. evening shower use combined with clothes washer contributes to morning peak). [INSERT FIGURE 10] 6.4. Benefits of the software application Future research aims to further improve the current software through creating a user-friendly presentation of the produced information, which can be interfaced by both the customer and water business professionals through a computer or smart phone accessible web-portal (Figure 11). For the customer, clever reports and diagrams on the following, as a minimum, will be designed: (a) daily water usage broken down on an end-use level for the past week and the average for the past month; (b) water end-use comparisons against a customer set budget, other households and best practice benchmarks; and (c) leak alerts and descriptions on likely leak types, with guidance on corrective actions. For the water business professional, automatically generated reports on the following will be created as a minimum: (d) water end-use averages for single or multiple properties from different suburbs (e.g., compare lower and higher socio-economic suburbs); (e) aggregated daily diurnal demand patterns and contributing end uses for specified days (i.e., peak day); and (f) water demand forecasting reports for selected regions based on just-in-time water end-use data provided. [INSERT FIGURE 11] 19 7. Conclusions, limitations and future directions The development of an autonomous and intelligent system for residential water end-use classification will be of significant benefit to both water consumers and utilities. It allows individual consumer to log into their user-defined water consumption program to view their daily, weekly, and monthly consumption tables, as well as charts on their water demand across major end use categories (e.g. leaks, clothes washer, shower, irrigation). It can also rapidly alert customers of leak events so that they can immediately be addressed rather than waiting for the present slow feedback process from the traditional metering technology (e.g. quarterly bill). This system will also help water businesses by rapidly providing water end-use reports of any desired property or suburb, thereby empowering them to develop more targeted conservation programs in water scarcity periods, improved water demand forecasting and optimised pipe network modelling. All of these opportunities can be realised by the proposed prototype expert system and associated software application integrating the single (Nguyen et al. 2013a), combined (Nguyen et al. 2013) and adaptive (current paper) analysis modules for categorising residential water flow data into end use event categories. The single event disaggregation model was comprehensively described in (Nguyen et al., 2013a), which employed HMM, DTW, event time-of-day probability function and other physical characteristics to assign an unclassified event into an appropriate water end use category. The formulation of a combined event analysis module was the logical second stage of research since a reasonable proportion of residential water consumption occurs simultaneously (Nguyen et al., 2013b). This modules utilises a hybrid combination of HMM, gradient vector filtering method, threshold values and various physical features to disaggregate combined events into several classified single events. The present analytical stage of this overall research project, which is the focus of this paper, had the goal to ensure that the model could adapt and self-learn variant water flow signature characteristics in different cities and regions without re-training or calibrating with that regions dataset. Through the application of HMM, DTW, threshold values and other physical features, the adaptive function has been successfully developed which allows the system to effectively analyse data from any new residential house in different regions. A verification process undertaken to assess the model capability displayed very promising outcomes with most of the 20 achieved recognition accuracies for all end use categories being approximately 90%. After this function was completed, a user-friendly automatic flow trace analysis application has been developed which integrates all available analysis modules into one comprehensive residential water end use event pattern recognition system. The only limitation with the adaptive module was its lower recognition accuracy for bathtub and irrigation events. While the present prototype software application is sufficient for conducting automated residential water end use analysis with an average of 80-90% recognition accuracy, the system would still require human input to achieve very high levels of recognition accuracy. Ultimately, accuracy in the order of 95-100% is required for commercially released software. Therefore, a future research program has been proposed by the researchers, which includes the following key tasks: 1) Apply genetic algorithms to explore the optimum states for the existing HMM classifier which may enhance accuracy and efficiency. 2) In addition to HMM and DTW, an Artificial Neural Network (ANN) with a back-propagation algorithm will be incorporated to analyse physical characteristics of each collected event (i.e. volume, duration and flow rate), which will likely have a significant impact on the recognition process accuracy. 3) Apart from the water end use event time-of-day likelihood functions that have been previously applied, other decision support parameters (i.e. social and demographic information) will be examined to improve accuracy. 4) Further train the analysis system using new water end use databases from other regions (i.e. Melbourne, Adelaide, United Kingdom) in order to improve its accuracy and sufficiently cater for different end use categories (e.g. evaporative air conditioners). References Baum, L. E. Petrie, T. 1966. Statistical inference for probabilistic functions of finite state Markov chains. The Annals of Mathematical Statistics 37 (6): 1554–1563. DOI:10.1214/aoms/1177699147. Baum, L. E. Petrie, T. Soules, G. Weiss, N. (1970). A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains. The Annals of Mathematical Statistics 41: 164. DOI:10.1214/aoms/1177697196. http://en.wikipedia.org/wiki/Digital_object_identifier http://dx.doi.org/10.1214%2Faoms%2F1177697196 21 Beal, C. and Stewart, R.A. (2011). South East Queensland residential end use study: final report. Technical Report No. 47 for Urban Water Security Research Alliance. Griffith University and Smart Water Research Centre, January 2012. Beal, C., Stewart, R.A., Huang, T.T., Rey, E. (2011a). SEQ residential end use study. Journal of the Australian Water Association 38 (1), 80-84. Beal, C.D. and Stewart, R.A. (2013) Identifying Residential Water End Uses Underpinning Peak Day and Peak Hour Demand. ASCE Journal of Water Resources Planning and Management Carragher, B.J., Stewart, R.A., Beal, C.D. (2012) Quantifying the influence of residential water appliance efficiency on average day diurnal demand patterns at an end use level: A precursor to optimised water service infrastructure planning. Resources Conservation and Recycling 62, 81-90. Chien, J.-T., Wang, H.-C. (1997). Telephone speech recognition based on Bayesian adaptation of hidden Markov models. Speech Communication, 22, 369-384. Cho, W., Lee, S.W., and Kim, J.H. (1995). Modelling and recognition of cursive words with HMM. Pattern Recognition, 28(12), 1941-1953. Ephraim, Y., Merhav, N. (2002). Hidden Markov processes. Information Theory, IEEE Transactions on, 48, 1518-1569. Ghahramani, Z., Jordan, M. I. (1997). Factorial Hidden Markov Models. Machine Learning 29 (2/3): 245–273. DOI:10.1023/A:1007425814087. Loh, M. and Coghlan, P. (2003). Domestic water use study in Perth, Western Australia 1998 to 2000. Water Corporation of Western Australia. Manmatha, R. and Srimal, N. (1999) Scale Space Technique for Word Segmentation in Handwritten Manuscripts. In: Proc. 2nd Int’l Conf. on Scale-Space Theories in Computer Vision, Corfu, Greece, September 26-27, 1999, pp. 22-33. Manmatha, R. and Rath, T. M. (2002): Word Image Matching Using Dynamic Time Warping. Multi-Media Indexing and Retrieval Group, Center for Intelligent Information Retrieval, University of Massachusetts. Technical Report Makki, A. Stewart, R.A. Panuwatwanich, K. and Beal, C. (2011) Revealing the determinants of shower water end use consumption: enabling better targeted urban water conservation strategies. Journal of Cleaner Production, DOI: 10.1016/j.jclepro.2011.08.007 22 Marquez, J.P. (2001) Pattern recognition: concepts, methods and applications. Springer, ISBN: 3-540-422978. Muller, M. (2007). Information Retrieval for Music and Motion, Chapter 4 . Springer, ISBN 978-3-540-74047-6. Myers, C. S., Rabiner, L. R. (1981). A comparative study of several dynamic time-warping algorithms for connected word recognition. The Bell System Technical Journal, 60, 1389-1409. Nguyen, K.A., Zhang, H., Stewart, R.A. (2011). Application of Dynamic Time Warping algorithm in prototype selection for the disaggregation of domestic water flow data into end use events. Proceeding of the 34th World Congress of the International Association for Hydro-Environment Engineering and Research, pp2137-2144, Brisbane, Australia, 26 June-1 July, 2011. Nguyen, K.A., Zhang, H., and Stewart, R.A. (2013a). Development of an intelligent model to categorise residential water end use events. Journal of Hydro-Environment Research, 10.1016/j.jher.2013.02.004 Nguyen, K.A., Zhang, H., and Stewart, R.A. (2013b). Intelligent pattern recognition model to automate the categorisation of residential water end-use events. Journal of Environment Modelling and Software, [under review]. Pearson, K. (1895). Contributions to the mathematical theory of revolution II: Skew variation in homogeneous material. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 186: 343–326. Rabiner, L., Juang, B. (1993). Fundamentals of speech recognition. Prentice-Hall, Inc., Chapter 4. Rabiner, L. R. 1990. A tutorial on hidden Markov models and selected applications in speech recognition. Readings in speech recognition. Morgan Kaufmann Publishers Inc. Sakoe, H., Chiba, S. (1978). Dynamic programming algorithm optimization for spoken word recognition, Acoustics, Speech and Signal Processing,19 IEEE Transactions on, vol. 26, no. 1, pp. 43{49, 1978. [Online]. Available: http://ieeexplore.ieee.org/xpls/abs all.jsp?arnumber=1163055 Satish, L., Gururaj, B. I. (2003). Use of hidden Markov models for partial discharge pattern classification. IEEE Transactions on Dielectrics and Electrical Insulation. Starner, T., Pentland, A. (1995). Real-Time American Sign Language Visual Recognition From Video Using Hidden Markov Models. Master's Thesis, MIT, Program in Media Arts. http://en.wikipedia.org/wiki/Special:BookSources/9783540740476 http://en.wikipedia.org/wiki/Special:BookSources/9783540740476 http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=212242 http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=212242 http://www.cc.gatech.edu/~thad/p/031_10_SL/real-time-asl-recognition-from%20video-using-hmm-ISCV95.pdf http://www.cc.gatech.edu/~thad/p/031_10_SL/real-time-asl-recognition-from%20video-using-hmm-ISCV95.pdf 23 Stewart, R.A., Willis, R.M., Giurco, D., Panuwatwanich, K., and Capati, B. (2010). Web-based knowledge management system: linking smart metering to the future of urban water planning. Australian Planner, 47(2), 66-74. Stewart, R.A., Willis, R.M., Panuwatwanich, K. and Sahin, O. (2011). Showering behavioural response to alarming visual display monitors: longitudinal mixed method study. Behaviour & Information Technology, DOI: 10.1080/0144929X.2011.577195, pp. 1-17. Tapia, E., Intille, S.S., Larson, K. (2004). Activity Recognition in the Home Using Simple and Ubiquitous Sensors. In: Ferscha, A., Mattern, F. (eds.) PERVASIVE 2004. LNCS,vol. 3001, pp. 158–175. Springer, Heidelberg Willis, R. M., Stewart, R.A., Panuwatwanich, K., Capati, B., Giurco, D. (2009a). Gold Coast domestic water end-use study. Water. Journal of Australian Water Association. 36(6), 9-85. Willis, R.M., Stewart, R.A., Panuwatwanich, K., Williams, P.R., Hollingsworth, A.L., (2011a). Quantifying the influence of environmental and water conservation attitudes on household end-use water consumption. Journal of Environmental Management. doi:10.1016/j.jenvman.2011.03.023 Willis, R.M. Stewart, R.A., Giurco, D., Talebpour, M.R., Mousavinejad, A. (2011b) End use water consumption in households: impact of socio-demographic factors and efficient devices. Journal of Cleaner Production, in-press, doi:10.1016/j.jclepro.2011.08.006 Appendices Appendix 1 Step1: Retrieve the initial state probability πi, state transition probability aij, and observation probability bj(ok) of the current model to use as starting probabilities. It should be noted that the subscript (i) here is used to indicate a state in HMM model training algorithm, not showing the end use category as in previous sections. Step 2: Based on the above values, determine the following parameters: • αt(i) : the probability of flow rate o1 through to ot and being in state i at time t ( iq t = ) given the HMM ( λ ) )|,...()( 21 λα iqoooPi ttt == (6) 24 • βt(i) : the probability of flow rate ot+1 through to oT, given the HMM ( λ ) and given that the model is currently in state i at time t ( iq t = ) ),|()( 21 λβ iqoooPi tTttt == ++  (7) • γt(i) : the probability of being in state i at time t given a water flow sequence ( O ) and HMM ( λ ) ∑ = = N j tt tt t jj ii i 1 )()( )()( )( βα βα γ (8) • ξt(i,j) : the probability of being in state i at time t, and in state j at time t+1, given a water flow sequence ( O ) and the HMM ( λ ). 1 ( , , | ) ( , ) ( | ) t t t P q i q j i j P λ ξ λ += == O O (9) where 1 1 1 1 ( | ) ( ) ( ) ( ), N N t kp p t t k p P k a b o pλ α β+ + = = = ∑∑O and 1 1 1( , , | ) ( ) ( ) ( )t t t ij j t tP q i q j i a b o jλ α β+ + += = =O Step 3: Calculate the following parameters for each water flow sequence ( O ) ∑ = T t t i 1 )(γ : expected number of times in state i for the water flow sequence ( O ) ∑ − = 1 1 )( T t t iγ : expected number of transition from state i for the water flow sequence ( O ) ∑ − = 1 1 ),( T t t jiξ : expected number of transition from state i to state j for the flow rate sequence ( O ) 25 Step 4: With the calculated values in step 3, the probabilities values of πi , aij and bj(ok) can be updated by performing Equations 10 to12: )( 1 ii γπ = (10) (11) (12) It should be noted that the above calculations of πi , aij and bj(ok) will be updated every time a new event is introduced to the existing HMM model for training. At the end of this process, a new HMM model ( λ ) will be achieved to cover both existing and new database. ∑ ∑ = = = = T t t T t t kj j j ob kt 1 such that 1 )( )( )( γ γ vo ∑ ∑ − = − == 1 1 1 1 )( ),( T t t T t t ij i ji a γ ξ 26 Figure captions Figure 1 Overview of proposed autonomous and intelligent water management system Figure 2 Example of frequency histogram with different number of clusters 27 Figure 3 Flowchart of the water end-use classification process Figure 4 Flowchart of adaptive model sequence 28 Figure 5 Adaptive model development Figure 6 Example of an unclassified group of events 29 Figure 7 Example of time of day probability for one particular home Figure 8a Adaptive and non-adaptive model comparison in terms of number of event 30 Figure 8b Adaptive and non-adaptive model comparison in terms of volume Figure 9a Software application main interface 31 Figure 9b Software application water end use pie and bar chart outputs Figure 10a Optional manual override reclassification of system classified event 32 Figure 10b Optional manual splitting of combined event into single event categories Figure 10c Software application output of water end use daily diurnal demand pattern 33 Figure 11 Proposed web interface application to customer and water utility 34 Table 1 Example of probability distribution for the dishwasher end use category Features Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Volume range (L) Frequency (No.) 2.62-3.17 4 3.17-3.71 1 3.71-4.25 6 4.25-4.8 9 4.8-5.3 1 Distribution probability (%) 19.1 4.8 28.5 42.8 4.8 Duration range (s) Frequency (No.) 80-93 2 93-106 4 106-119 1 119-132 6 132-145 9 Distribution probability (%) 9.5 19.1 4.8 28.5 42.8 Mode flow range (L/min) Frequency (No.) 2.3-2.7 14 2.7-3.1 1 3.1-3.5 1 3.5-3.9 1 3.9-4.3 4 Distribution probability (%) 66.7 4.7 4.7 4.7 19.2 35 Table 2 Determination of the aggregate likelihood of the grouped events to be classified to dishwasher category Event 1 2 3 4 5 6 7 8 9 10 11 Representative values 𝑣 (L) 4.63 4.57 4.59 4.63 4.69 4.75 4.71 4.79 4.92 4.76 5.18 4.64 Step 2 𝑡 (s) 125 125 125 125 130 130 130 135 135 135 140 126.5 𝑞𝑓 (L/min) 2.3 2.3 2.3 2.3 2.3 2.3 2.3 2.3 2.3 2.3 2.3 2.3 𝑉𝑠,4 (L) 42.8 26.5 66.7 Step 3 𝑇𝑠,4(𝑠) N/A 𝑄𝑓,4 (L/min) Step 4 𝐻𝑀𝑀4 (x10 -8) 7.4 54.5 48.9 16.2 75.8 19.9 6.5 2.1 4.0 3.3 13.2 9.5 Step 5 𝐿𝐿4 (x10 -4) 71.86 N/A 36 Table 3 End use event categorisation accuracy (%) using adaptive and non-adaptive models Adaptive model accuracy (%) Non-adaptive model accuracy (%) End use category Home 1 Home 2 Home 3 Average Home 1 Home2 Home 3 Average V N V N V N V N V N V N V N V N Shower 88.9 78.5 76.4 80.6 79.2 85.4 81.5 81.5 76.9 73.6 81.5 75.8 79.2 82.3 79.2 77.2 Faucet 97.0 93.4 77.3 79.3 95.2 96.3 90.1 89.7 93.6 90.4 68.9 78.3 86.9 84.3 83.1 84.3 Clotheswasher 96.7 90.1 86.3 81.8 91.8 95.5 91.6 89.1 85.2 83.1 82.6 81.8 84.2 85.5 84.0 83.4 Dishwasher 96.7 94.3 85.1 88.4 0 0 90.9 91.4 85.6 83.3 78.5 80.4 0 0 82.1 81.8 Toilet 91.4 87.8 89.8 86.5 97.1 84.4 92.8 86.2 78.6 75.3 70.2 74.6 72.1 75.4 73.6 75.1 Irrigation N/A N/A N/A N/A 100 100 100 100 N/A N/A N/A N/A 100 100 100 100 Bathtub 20.6 40.5 N/A N/A N/A N/A 20.6 40.5 20.6 40.5 N/A N/A N/A N/A 20.6 40.5 Note: Testing end use event categorisation accuracy by V (volume of end use) and N (number of end use events) correctly classified. 3.1. Collected data for the study 
work_gqmnbfeeevb27g4yjxdcqdqpmi	----	Expert Systems with Applications 00 (2012) 1–13 Expert Systems with Appli- cations Time-stamped Resampling for Robust Evolutionary Portfolio Optimization Sandra Garcı́aa,∗∗, David Quintanaa, Inés M. Galvána,∗, Pedro Isasia aComputer Science Department, Carlos III University of Madrid Avda. Universidad 30, 28911 Leganes, Spain http://www.evannai.inf.uc3m.es Abstract Traditional mean-variance financial portfolio optimization is based on two sets of parameters, estimates for the asset returns and the variance-covariance matrix. The allocations resulting from both traditional methods and heuristics are very dependent on these values. Given the unreliability of these forecasts, the expected risk and return for the portfolios in the efficient frontier often differ from the expected ones. In this work we present a resampling method based on time-stamping to control the problem. The approach, which is compatible with different evolutionary multiobjective algorithms, is tested with four different alternatives. We also introduce new metrics to assess the reliability of forecast efficient frontiers. Keywords: Financial Portfolio Optimization, Robust Portfolio, Multiobjective Evolutionary Algorithms 1. Introduction Asset allocation is one of the core topics in financial management. The search for the optimal choice of financial assets to be included in a portfolio has been the subject of research for a long time and it is one of the most active lines in finance. The academic literature on this subject is very large and mostly based on the work of Markowitz [12, 9]. Under this framework, the problem is introduced as a multiobjective optimization problem where the investor tries to figure out the weight that each of the investment alternatives should carry in the portfolio. The target of this investor would be both minimizing risk and maximizing return at the same time. The solution to the problem of optimizing for these two objectives in conflict defines a set of solutions called the efficient frontier. This Pareto front consists on portfolios that are neither better or worse than the rest. For each level of risk or return, there is no better alternative in terms of the other objective. This makes the election of one of them a choice to be made by investors according to the way they weight risk and rewards. The basic version of the problem can be solved using Quadratic Programming (QP). Unfortunately, this approach is built on a set of assumptions that are unlikely to hold in the real world. The quest for alternatives has driven attention to metaheuristics that might not suffer this limitation. This is the reason why the framework of evolutionary computation is getting traction on this area. ∗Corresponding author ∗∗Principal corresponding author Email addresses: sgrodrig@inf.uc3m.es (Sandra Garcı́a), dquintan@inf.uc3m.es (David Quintana), igalvan@inf.uc3m.es (Inés M. Galván), isasi@ia.uc3m.es (Pedro Isasi) 1 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 2 The solutions based on metaheuristics mostly fall into one this two categories: solutions that transform the multi- objective problem into a single-objective form, and those who deal with it using a mutiobjective approach. Among the first group, we would mention the work by Soleimani et al.[21]. These authors use a genetic algorithm and extend the mentioned classic mean-variance optimization model and consider constraints on transaction costs, round lots or cardinality. They tackle the multiobjective nature of the problem minimizing the risk objective while setting a range of minimum acceptable returns in the constraints. Chiranjeevi and Sastry [4] use another popular approach to transform the multiobjective problem into a version that can be handled by single-objective algorithms. Instead of keeping one of the objectives on the objective function and the other in the constraint, they consider the objectives in the fitness function weighting them. In this particular case, they manage five objectives that are obtained breaking down the basic two. Chang et al. [3] suggest a similar solution and optimize a function that weights risk and return using a risk aversion parameter. Zhu et al. [23] target a popular metric, the Sharpe Ratio, that combines both elements into a single expression. The development of Multiobjective Evolutionary Algorithms (MOEAs) has resulted into a number researchers exploring their performance in this area. Among these we could mention Skolpadungket et al. [20], who test a set of multi-objective algorithms (VEGA, SPEA2, NSGA-II...) on a constrained version of the two objective problem. More recently, Anagnostopoulos and Mamanis [1] compare the performance of different multiobjective algorithms, and Deb et al. [6] introduce a customized hybrid version of NSGA-II to tackle the problem. Finally, we will mention the work of Radziukyniene and Xilinskas [16], where authors compare FastPGA, MOCELL, AbYSS, and NSGA-II on both of the basic problems, and an extended version that considers the dividend yield as a third objective. Despite of the amount of research on portfolio optimization, there are still open issues. A key one is the robustness of results provided by algorithms. Among the most important factors that asset managers have to consider when evaluating the results provided by any of the above-mentioned methods, there is reliability. Very often, the expected efficient frontier lies far from the real one. This is due to the fact that the estimates for the expected risk and returns for the assets in the solution, and the portfolios derived from them, are very inaccurate. Understandably, this results on mistrust by some practitioners. The search for solutions for this problem has cleared the way for the field of robust portfolio optimization. It is in this area where we focus our contribution. We introduce a time-stamping method to control the population that enhances the reliability of the solutions provided by MOEAs. The process of optimizing the risk and return of a portfolio relies on two parameters: the estimates for the expected asset returns and the variance-covariance matrix. The values of these parameters are usually based on past data and they might be inaccurate due to, for instance, the presence of outliers. In this context, there are several potential ways to tackle the problem. The two most prevalent alternatives found in the literature rely on either putting an emphasis on having robust estimates for the parameters, or managing the optimization process itself. The first one usually tries to filter the estimates to control, for instance, the influence of extreme past events on their computation [14]. The authors focusing on the second alternative design approaches handle uncertainty in the parameters during the optimization process [15, 22]. The alternative suggested in this paper falls in the latter category. We will enhance the solutions of MOEAs testing the population for different values for the parameters, and selecting the portfolios that consistently offer a good performance. Optimizing for a single scenario, a set of expected asset returns and the use of a single variance-covariance matrix, bears the risk of getting solutions that might be extremely sensitive to deviations. This could potentially be a problem as it is almost certain that the estimates will not be accurate. We have to bear in mind that having perfect estimates for the expected returns for instance, implies that we can make perfect predictions for future prices, which is highly unlikely. For this reason we consider that assessing the candidate solutions in different likely scenarios and favoring those that consistently offer a good performance, might be a promising approach. The requirement of consistency is key. In order to achieve it, we introduce a time-stamping mechanism that, together with performance in terms of risk and return, will drive the evolution process to find robust and stable solutions. The approach introduced in this paper is related to alternatives based on resampling [19, 18]. The most compa- rable among the traditional approaches is the one described by Idzorek [10]. This author suggests using combining traditional QP with Monte Carlo simulation to derive a set of fronts that are merged into a single solution at a later stage. This idea is very interesting but, unfortunately, the approach suffers the shortcomings of QP, namely, the ability to deal with real-world constraints. This is the reason why we feel that adapting the idea to the framework of MOEAs is very promising, as they do not suffer from this limitation. There is a previous effort based of evolutionary algorithms along the lines of this work, but is based on a very simple resampling approach that optimizes for a different scenario 2 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 3 in each generation[8]. The time-stamping mechanism that we introduce in this work is based on the previous one extending the problem with a third implicit objective that favors solutions that are consistently reliable. As we will see in the experimental section, this results on significantly higher robustness. The proposed approach is compatible with a wide array of MOEAs. Given their different nature and behavior, we will test the approach on a set of popular algorithms. Specifically, the experimental section will consider NSGA- II, SPEA2, SMPSO and GDE3. NSGA-II [5] is one the most referenced algorithms in the field of multiobjective optimization. This one, together with SPEA2 [24] have been confirmed [20] to offer good performance in portfolio optimization. Apart from the mentioned two, we consider GDE3 [11], a differential evolution strategy, and SMPSO, [13] a multiobjective algorithm based on particle swarm optimization. In the context of multiobjective problems, an important issue is the metric used to evaluate the solutions. It is generally admitted that there is no ideal single metric that should be used to evaluate different objectives simultane- ously in every circumstance. The metrics that are most commonly used in this field, such as Hypervolume, Spread or SetCoverage are not appropriate indicators of stability in this context. For this reason, we define a set of metrics that capture different aspects of robustness in efficient frontiers. These, Estimation Error, Stability, Extreme Risk and Unrealized Returns draw on the basic principle that the expected risk and returns for the portfolios in the solution should be close to the observed ones ex-post. The rest of the paper is organized as follows. First, we make a formal introduction to the financial portfolio optimization problem. After, we describe in detail the evolutionary approach proposed in this work. This section includes a brief description of the MOEAs, the solution encoding and the fitness mechanism in order to find robust and stable portfolios. Next, the different metrics used in this work to evaluate robustness of solutions are described. That will be followed by the experimental results and, finally, there will be a section devoted to summary and conclusions. 2. Financial Portfolio Optimization Problem Financial portfolios can be defined as a collection of investments or assets held by an institution or a private individual. The Modern Portfolio Theory was originated in the article published by Harry M. Markowitz, in 1952 [12]. It explains how to use the diversification to optimize the Portfolio. In general, the portfolio optimization problem is the choice of an optimum set of assets to include in the portfolio and the distribution of investor’s wealth among them. Markowitz [9] assumed that solving the problem requires the simultaneous satisfaction of maximizing the expected portfolio return E(Rp) and minimizing the portfolio risk (variance) σ2p, that is, solving a multiobjective optimization problem with two output objective functions [21, 4, 20, 16]. The portfolio optimization problem can be formally defined as: • Minimize the risk (variance) of the portfolio: σ2p = Σ n i=1Σ n j=1wiw jσi j (1) • Maximize the return of the portfolio: E(Rp) = Σ n i=1wiµi (2) • Subject to: Σ n i=1wi = 1 (3) 0 ≤ wi ≤ 1; i = 1...n (4) where n is the number of available assets, µi the expected return of the asset i, σi j the covariance between asset i and j, and wi are the decision variables giving the composition of the portfolio. The constrains referenced in equations 3 and 4 require the full investment of funds and prevent the investor from shorting any asset, respectively. In a quantitative way, the risk is represented with the standard deviation σp. The solution to the problem should also consider some real world constraints [2] such as: 3 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 4 Figure 1. Efficient frontier • Cardinality constraint: it is possible to define the maximum Cmax and minimum Cmin number of assets in which it is possible to invest (wi , 0): Cmin ≤ Σ(wi , 0) ≤ Cmax (5) • Values limit constraint: each weight wi must have a value in the interval [limin f , limsup], where: 0.0 ≤ limin f ≤ wi ≤ limsup ≤ 1.0 (6) All of these equations are solved by a set of points that constitute the efficient frontier of the problem. The points will define a curve similar to fig.1, plotted in the risk-return space of all possible portfolios. The points of this curve represent portfolios which have the minimum amount of risk given a certain expected return (and viceversa). 3. Evolutionary Approach for Robust Portfolio Optimization As it was mentioned in the introduction, in this paper we tackle the problem of achieving robust or stable portfolios using MOEAs. In order to do that, we suggest replacing the traditional fitness function with a new one that extends it with a resampling mechanism and an implicit third objective to control the robustness of the solutions in the front. We will test the effectiveness of the approach on different MOEAs using the same chromosome structure and fitness evaluation procedure. In this section, we introduce briefly the MOEAs tested and provide details regarding the chromosome encoding and fitness function. 3.1. Tested Evolutionary Multi-objective Algorithms The following MOEAs will be tested: NSGA-II, GDE3, SMPSO, SPEA-II. All the algorithms are implemented in jMetal [7], a Java framework aimed at multiobjective optimization with metaheuristics. By reusing the base classes of jMetal, all the techniques share the same basic core components (solution encodings, operators, etc.). This ensures a fair comparison of the considered algorithms. NSGA-II, proposed by Deb et al. [5], is one of the most widely used multiobjective metaheuristics. It represents the new version of the NSGA algorithm developed by the same author. It is a generational genetic algorithm based on an auxiliary population derived from the original one applying the usual genetic operators (selection, crossover and mutation). Then, the two populations are merged and the individuals are sorted according to their rank. Inside each of these ranks, the crowding distance is used to sort the individuals from less to more crowded. A solution with a smaller value of this distance measure is, in some sense, more crowded by other solutions. Finally, the best solutions are selected to compose the new population that will be used to create a new population. GDE3 [11] is the third version of the Generalized Differential Evolution algorithm (GDE), an extension of Dif- ferential Evolution (DE) for global optimization with an arbitrary number of objectives and constraints. GDE3 starts with a population of random solutions. At each generation, an offspring population is created using the differential evolution operators; then, the current population for the next generation is updated using the solutions of both, the 4 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 5 offspring and the current population. Before creating the next generation, the size of the population is reduced using non-dominated sorting and a pruning technique aimed at diversity preservation, in a similar way as NSGA-II. How- ever, GDE3 modifies the crowding distance of NSGA-II in order to solve some of its drawbacks when dealing with problems having more than two objectives. SMPSO, introduced by Nebro et al. [13], is a multiobjective particle swarm optimization algorithm (PSO) char- acterized by the use of a strategy to limit the velocity of the particles. It is based in OMOPSO [17] but including the velocity constriction procedure. This mechanism is useful when the velocity becomes too large because it can produce new effective particles positions. SMPSO also relies on an external archive that stores the non-dominated solutions found during the search process. Polynomial mutation is used in the algorithm as a factor of turbulence. SPEA-II algorithm [24], developed by Zitzler et al., solves some weaknesses of a previous version by the same authors called SPEA. Among the improvements, we could mention a fitness functions that takes into account, for each individual, the number of individuals dominated by this one and the number of individuals which dominate it. This version also adds a density estimation for the population. This algorithm uses a core population and an archive. It assigns to each individual a fitness value that is the sum of its strength raw fitness plus a density estimation. In each generation, the non-dominated individuals of both the original population and the archive are used to update the archive; if the number of non-dominated individuals is greater than the population size, a truncation operator based on calculating the distances to the k-th nearest neighbor is used. All this procedure is known as Environmental Selection. Then, the algorithm applies the selection, crossover, and mutation operators to members of the archive in order to create a new population of offsprings which becomes the population for the next generation. 3.2. Solution encoding The encoding chosen will represent each portfolio as a vector of real numbers. Each of these numbers represents the percentage of investment per asset (also called weight: wi where i ranges from 1...n, and n is the number of investable assets). Here, each portfolio will be represented by a single element of the population. Every individual must meet the constraints specified by eqs. 3, 4 explained before. The sum of weights per individual must be 1, that is full investment is required. Also, the individuals must satisfy additional real-world constraints showed in eqs. 5 and 6. The satisfaction of these constraints is guaranteed by repairing algorithms. The individuals are repaired both after initializing the population (see alg. 1), and applying the genetic operators (see alg. 2). The repairing algorithms transform non-complying chromosomes into portfolio satisfying constraints. Whenever the number of invested assets do not belong to the interval [Cmin, Cmax], its number is adjusted to ensure compliance with the cardinality constraint. This is done adding or dropping assets until the requirement is met. In case the sum of weights per individual is not 1.0, the algorithm fine tunes the holdings adding or subtracting random amounts up to the required adjustment. These changes are forced meet the investment limits [limin f , limsup]. The details of this process are described in algorithms 1 and 2. Algorithm 1 Reparation after initialization Initialize population P as a set of vectors with real numbers xi = (xi1, ..., xin) ; xi j ∈ [limin f , limsup] for each individual xi of P do A random number ∈ [Cmin, Cmax] of values 0 is assigned to coordinates of vector xi while Σnj=1 xi j , 1 do if sum of the limin f of xi j , 0 is > 1 then select randomly one coordinate j of vector xi such that xi j , 0 and assign xi j = 0 end if if sum of the limsup of xi j , 0 is < 1 then select randomly one coordinate j of vector xi such that xi j = 0 and assign xi j = limsup end if if Σnj=1 xi j , 1 then select randomly one coordinate j of vector xi such that xi j , 0 add/subtract the quantity left to make the Σnj=1 xi j = 1 respecting the limits [limin f , limsup] end if end while end for Return(P) 5 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 6 Algorithm 2 Reparation after genetic operators for each individual xi of P do if there are less xi j = 0 than Cmin then set one random xi j = 0 to xi j = limin f end if if there are more xi j = 0 than Cmax then set the xi j , 0 with less value to xi j = 0 end if while Σnj=1 xi j , 1 do if sum of the limin f of xi j , 0 is > 1 then select randomly one coordinate j of vector xi such that xi j , 0 and assign xi j = 0 end if if sum of the limsup of xi j , 0 is < 1 then select randomly one coordinate j of vector xi such that xi j = 0 and assign xi j = limsup end if if Σnj=1 xi j , 1 then select one random xi j , 0 ∈ [limin f , limsup] add/subtract one random quantity left to make the Σnj=1 xi j = 1 respecting the limits [limin f , limsup] end if end while end for Return(P) 3.3. Fitness function. Time-stamped resampling In this paper we introduce a resampling strategy combined with an implicit age-based third objective to identify robust portfolios. In this subsection, the procedure to evaluate the fitness function is described in detail. The starting point for the evaluation of portfolios is the framework introduced in section 2, where part of the fitness of each individual is determined by evaluating two objective functions: the return E(Rp) (eq. 2) to be maximized and the risk σp (eq. 1) to be minimized. As it is apparent, E(Rp) and σp are very tightly related to the values of the expected asset returns and the variance-covariance matrix. One of the most important challenges that portfolio managers face when they operate within Markowitz’s framework is the dependence of the solutions on the estimates for these parameters. Given the challenges inherent to financial forecasting, it is normal that the mentioned parameters are not accurate. The mentioned difficulty is likely to result in a set of portfolios that could end up behaving in an unexpected way. Fig.2 shows a real example of one Pareto front where the solutions are evaluated using the forecast parameters (in red) vs. the real parameters (in green). The set of portfolios that were optimized for the traditionally considered most likely scenario (mean return for each asset over a period of time, and variance-covariance matrix computed using the same data) define the upper efficient frontier. However, once we calculate the risk and return for the same portfolios using the real observed parameters (actual assets returns instead of the mentioned averages), we realize that their profiles can be very different. As can see, the difference between the two fronts, that represent the same solution, might be potentially quite dramatic. A first step towards robust solutions would be resampling during the evolution process. This approach would require replacing the mentioned parameters in the fitness functions at every generation. Doing this, the evolution process would favor the individuals that show good performance in terms of risk and return over different scenarios, and would discard those overspecialized on specific values for the parameters, including the predicted ones. A key element of this process is the creation of these scenarios. We would need a set of solutions able to handle uncertainty, but there is no need for absolute generalization. The scenarios should be considered according to their likelihood. The approach that we use to generate likely scenarios is based on nonparametric bootstrap. As we mentioned before, the usual way to forecast the value of parameters for the model is averaging the returns and computing the associated variance-covariance matrix over a number of periods. Our algorithm has the same starting point, the definition of a time window. However, instead of using all the data to derive a single estimate for the parameters, data are resampled. The process selects a random set of time periods that has the same size as the original window (each period might be selected more that once). Then, we average the returns for those time periods and compute the variance-covariance matrix. Every time we do this, we generate new estimates for the parameters that are based on 6 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 7 Figure 2. Solution evaluated with forecast and real parameters past data. These estimates can subsequently be used to calculate the risk and returns of the portfolios. This process is described in 3. Algorithm 3 Resampling method S is the original sample set with a size Ns. S ′ is the new sample set with a size N′s. At the beginning, S ′ = ∅ and N′s = 0. while N′s , Ns do It selects one instance Xi, at random, from S . This instance is added to the new set. S ′ = S ′ + Xi. end while Return(S ′) Every generation, the algorithm evaluates all the individuals under a new resampled scenario. This is relevant because using a shared set of parameters makes the comparison for dominance purposes meaningful. This process alone would be enough to weed out overspecialized solutions over time. However, as it was described, it has a relevant flaw. The results would be very dependant on the last iterations of the evolution process. Every generation, we replace some solutions from the previous population with new ones and, chances the lasts scenarios generated might be very extreme. If that were the case, a percentage of the best solutions that might have fared well over time, might offer a poor performance. This could result on new solutions specialized on the last scenario replacing the old ones. If this happened during the first few generations, the specialized solutions would be replaced over time. However, in case we faced this in the final stages, we might end up obtaining poor results. We will prevent this adding to the described resampling an implicit third objective based on a time-stamping mechanism. This feature extends the resampled mean-variance model with a third objective, the age of the solution, to be maximized. Each individual will have a counter that will keep track of the number of generations it has remained in the population. The explicit maximization of this variable rewards good performance in previous generations and mitigates the risk described before. Once the evolution process is over, the population is evaluated once more using the forecast values for the parameters for the period. The final set of portfolios, provided as solutions, consists of the elements of the population that are not dominated in terms of risk and return. At this stage, all the information concerning the third objective is disregarded. Algorithm 4 shows the modifications carried out in the basic MOEAs to implement the fitness function described in this section. We expect that the sum of these changes in the traditional fitness function would discard those portfolios that, under normal circumstances, could potentially result in particularly bad performance, and prioritize those that offer consistently good solutions under likely scenarios. 7 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 8 Algorithm 4 Modifications on the basic MOEAs (Specific steps of MOEA) Reparation after initialization (alg. 1) for each generation do A new sample set (S ′) is generated by alg. 3 (Specific steps of MOEA) Reparation after genetic operators (alg. 2) for every existing individual p do do Calculate E(Rp) on S ′ Calculate σp on S ′ T p = T p + 1 end for (Specific steps of MOEA) end for (Specific steps of MOEA) MOEA returns a non-dominated solution’s set P Compute forecast parameters (S : original sample set) for each individual p in P do Discard time-stamping objective T p Calculate E(Rp) on S Calculate σp on S end for Get non-dominated individuals (P f ) from P Return (P f ) 4. Evaluation Metrics Solutions resulting from different runs of multiobjective algorithms must be compared using quantitative metrics. There is wide range of alternatives, some of them dependent on knowing the Pareto-optimal front. It is generally admitted that there is no single perfect metric that can be used to compare solutions in every circumstance. Actually, the choice of the right metric among the wide range available is itself, a multiobjective problem. The metrics most widely used in multiobjective context are Hypervolume and Spread [25]. These metrics could be useful when we look for good distributed and strong dominant fronts. However, they are not appropriate to capture the effect showed in fig. 2. In order to measure the robustness of the solutions, we introduce four metrics that evaluate different aspects related portfolio robustness. They are named: Estimation Error, Stability, Extreme Risk and Unrealized Returns. Next, we describe them in detail. • Estimation Error: It evaluates the average difference between the expected risk and return for every portfolio in the efficient frontier and the actual risk and return a posteriori once the real values of the parameters are observed. That is, the mean distance between the estimates for tn based on data from t0 to tn−1, and the actual values at tn. For this purposes, we consider the Mahalanobis distance, dM (x, y), which is defined as, dM (x, y) = √ (x − y)T Σ−1(x − y) (7) where x and y are the patterns to be compared, and Σ is the variance-covariance matrix. This metric is calculated measuring the average distance (dM ) between xp and xp ′ for all the portfolios in the solution. Here, xp represents, on one hand, the pair (E(Rnp),σ2np) for portfolio p and period tn calculated using forecast parameters. On the other hand xp ′ is defined by the pair (E(Rnp) ′,σ2np ′ ) where both return and risk for the same portfolio and moment in time are computed using the the real parameters (the ones observed for tn). Formally, it can be expressed as follows: EE = ΣNp=1[dM (xp, xp ′)]2 N (8) where N is the number of portfolios in the Pareto front. The smaller is the difference between the forecast and reality, the lower is the value of this metric and, therefore, the higher is the reliability of the original front. 8 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 9 • Stability: The Stability of a portfolio is measured averaging the Mahalanobis distances among the expected pair of risk/return in the efficient frontier and the expected risk/return in S different scenarios. Unlike the Estimation Error, which considers the difference between the expected scenario and the real scenario, this metric measures the average difference between the expected scenario and a wide range of feasible alternatives. The concern is the aggregate sensitivity to the distribution of potential scenarios, not to the one that happened to materialize. The metric is given by the equation: S T = ΣSi=1 ΣNp=1 [dM (xp,xpi )] 2 N S (9) where dM (xp, xpi) is the Mahalanobis distance between (E(Rnp),σ2np) and (E(Rnpi),σ 2 npi), the return and risk for portfolio p and period tn calculated using parameters for the scenario i. The set of scenarios is generated using nonparametric bootstrap technique described in Alg. 3. The relevance of this metric would subject to the size of the data set used to resample and the value of S . A larger set of scenarios is likely to result on a more accurate approximation to the potential distribution of parameters. As it was mentioned before the value of the metric is obtained averaging all the average distances. Therefore, high values of this metric would represent higher sensitivity to likely scenarios and lower reliability. • Extreme Risk: This metric evaluates the sensitivity of the solutions to worst-case scenarios. It is closely related to Stability, and it matches the same basic definition. The difference is that, instead of considering the average for all S scenarios, we will only take into account a small subset. Specifically, the computation would include only the w worst-case ones. For this purposes we define worse-case scenarios as the parameter combinations that result in the highest average Mahalanobis distance between the expected risk and returns, and the risk and returns for the same portfolios using the resampled parameters. The rationale for this indicator is providing an estimate for the expected average negative outcome of the realization of the worst scenario with probability w/S . The higher the metric, the higher the risk. • Unrealized Returns: This indicator provides information on the average potential return left on the table. Namely, it measures the average squared difference between the realized return and the maximum potential return for that risk level for all the portfolios in the solution. This is defined as: UR = ΣNp=1[Rnpe − Rnp] 2 N (10) where N is the number of portfolios in the solution, Rnp is the return of portfolio p computed using the ex-post parameters for time tn and Rnpe is the return of the portfolio, pe, with the most similar risk level in the effi- cient frontier derived using the ex-post parameters. High values on this metric would indicate large unrealized potential returns. 5. Experimental Validation In this section we test the aforementioned approach on a specific asset allocation problem. We try to identify the mixes of broad investment categories that provide the optimal balance between expected risk and return, while increasing robustness. As it was mentioned in the introduction, four different multiobjective meta-heuristics are considered on a constrained version of the problem. For each of them, NSGA-II, SPEAII, SMPSO and GDE3, we compare the performance of the standard version versus the robust equivalent. For experimental purposes, the cardinality constraints [Cmin, Cmax] and limits to the minimum and maximum weight that each asset can carry in the portfolio, [limin f , limsup], are set to [2,6] and [0.1,0.8] respectively. Subsection 5.1 describes the data sets that we used in this work. Subsection 5.2 gives details about the parameter settings for the experiments, multiobjective algorithms, and metrics. Finally, subsection 5.3 presents and analyzes the experimental results. 9 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 10 5.1. Data Sets For experimental analysis, we use a sample that consists of 240 monthly returns for eight broad financial indexes representing that many asset classes. The series of monthly returns cover the time period from January, 1990 to December, 2009 and the source for the data is Datastream. The list of indexes is provided in Table 1. Table 1. Data Sets Name Code Frank Russell 2000 Value FRUS2VA Frank Russell 2000 Growth FRUS2GR Frank Russell 1000 Value FRUS1VA Frank Russell 1000 Growth FRUS1GR S&P GSCI Commodity Total Return GSCITOT MSCI EAFE MSEAFEL BOFA ML CORP MSTR ($) MLCORPM BOFA ML US TRSY /AGCY MSTRAAA($) MLUSALM 5.2. Experimental Parameters The experimentation is based on a sliding window approach. Each window has a size n of 120 sets of returns for the eight indexes, that is 10-year worth of data. This means that the algorithm will rely on data from t0 to tn−1 to identify the best possible allocations for the period tn. The 10-year window will move one month at a time, 120 times. This will cover the date interval from 31/01/1990 to 31/12/2009. The algorithms will be run 20 times per window using the parameters described in table 2. This means that, for each window, we will obtain 20 solutions sets per algorithm. Table 2. Parameters. L = 8 (individual length). The termination condition is to compute 300 iterations. SPEA2 Population size 200 individuals Archive size 200 individuals Crossover SBX, pc = 0.9 Mutation Polynomial, pm = 1/L Selection of Parents Binary tournament NSGA-II Population size 200 individuals Crossover SBX, pc = 0.9 Mutation Polynomial, pm = 1/L Selection of Parents Binary tournament SMPSO Archive size 200 particles Swarm size 200 particles Mutation Polynomial, pm = 1/L GDE3 Population size 200 individuals Crossover DE crossover, CR = 0.9 Mutation DE mutation, F = 0.5 Selection of Parents DE selection As for the parameters required by the metrics, the number of scenarios S necessary to compute Stability and Extreme Risk, will be set to 500. The number of worst-case scenarios, w, considered for the second, will be 5. That means that comparison will be based on the 1% worst expected average outcomes. 5.3. Experimental Results This section shows the results from the experimental process described before. For every multiobjective algorithm tested, we compare the performance of the standard and the robust version of the algorithms over 20 runs using the metrics introduced in 4. In the tables, the first one is labeled as ”name-of-algorithm” and the second one as ”name-of- algorithm R+T” (R+T means resampling and third objective). Results are provided in terms of descriptive statistics 10 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 11 (average, median and variance) for metrics described in section 4. The percentage of improvement of the robust version over the basic one, labeled as % Av., is also reported. Table 3 shows the set of results for Estimation Error metric (EE). The advantage of using the robust approach over the standard one ranges from 27.46% to 54.93%. The lowest prediction errors are achieved using the robust version SMPSO, however the highest improvement takes place for SPEA2 which is the algorithm with highest initial EE value. We also note that the worse EE in the basic MOEA, the more it is enhanced by the robust approach. The results for the Stability metric are reported in table 4. The table shows how the proposed approach (R+T) improves significantly the stability of portfolios achieved by the standard MOEAs. Once again, the improvements are specially remarkable for the algorithms that, in their standard formulation, lead to less stable fronts. We observe that SPEA2 is the algorithm that provided the highest values for the metric in the basic version and the lowest values in its robust alternative. This makes the difference between using R+T mechanism and the standard algorithm specially large (67.53%). Conversely, GDE3 R+T shows the lowest improvement %Av (35.61%) and the worst final value for the metric, once we use the robust approach. Table 3. ”Estimation Error” Average Median Variance % Av. SPEA2 2,5198 1,7945 5,1421 SPEA2 R+T 1,1356 0,5538 2,4147 -0,5493 NSGAII 2,2823 1,7996 3,9817 NSGAII R+T 1,2401 0,6609 2,3748 -0,4566 SMPSO 1,5204 1,2878 1,7354 SMPSO R+T 1,0144 0,7403 0,9302 -0,3328 GDE3 1,4820 1,2580 1,5347 GDE3 R+T 1,0750 0,6721 1,2762 -0,2746 Table 4. ”Stability” Average Median Variance % Av. SPEA2 7,4057 7,0743 20,7329 SPEA2 R+T 2,4044 2,1024 2,2526 -0,6753 NSGAII 6,5508 6,1640 12,9661 NSGAII R+T 2,7433 2,5085 2,1520 -0,5812 SMPSO 5,2373 4,8981 7,2544 SMPSO R+T 3,2510 2,8582 3,6025 -0,3793 GDE3 5,1666 4,8507 6,8238 GDE3 R+T 3,3270 2,9755 3,8045 -0,3561 The results for the Unrealized Returns metric are reported in table 5. We observe that the solutions obtained by R+T using observed parameters are closer to the optimal efficient frontier than the solutions derived from the standard algorithms. The best results, once again, were obtained by SMPSO R+T. Results show improvements that range from 19.56% to 32.38%. GDE3 is the algorithm with the lowest % Av. despite not being the algorithm with the lowest value for the metric in the standard version. Finally, solutions evaluated under the less predictable parameters or scenarios (see table 6) show the same pattern of behavior already observed for EE and Stability metrics. It is remarkable that solutions obtained by R+T approach and evaluated under these worst scenarios are significantly closer to the observed parameters than the solutions of standard MOEAs evaluated under the forecast parameters. In this case, even though SPEA2 R+T offers the best median result, on average, it is beaten by the R+T versions of GDE3 and SMPSO, which provides both the smallest average value and the lowest variance for the metric. Table 5. ”Unrealized Returns” Average Median Variance % Av. SPEA2 3,5848 2,7989 8,5604 SPEA2 R+T 2,4240 1,8792 3,9853 -0,3238 NSGAII 3,4871 2,7326 7,8823 NSGAII R+T 2,5670 1,9975 4,1648 -0,2638 SMPSO 2,9805 2,4745 4,9961 SMPSO R+T 2,2514 1,7717 2,9235 -0,2446 GDE3 3,3518 2,9221 5,7049 GDE3 R+T 2,6962 2,1723 4,5393 -0,1956 Table 6. ”Extreme Risk” Average Median Variance % Av. SPEA2 3,5308 3,0614 4,9670 SPEA2 R+T 1,6622 1,0862 2,5638 -0,5292 NSGAII 3,2407 2,8451 3,9687 NSGAII R+T 1,8172 1,2318 2,5525 -0,4393 SMPSO 2,2485 2,0297 1,8884 SMPSO R+T 1,4671 1,2126 1,0257 -0,3475 GDE3 2,2091 1,9701 1,7397 GDE3 R+T 1,6425 1,2577 1,5775 -0,2565 The differences between the average values for the metrics for the basic and the suggested R+T versions of the algorithm, % Av., were tested for statistical significance. We used the protocol described in algorithm 5 and all the differences were significant at 1%. The best performing algorithm, in terms of robustness, seems to be SMPSO R+T. It offers the lowest values for three out of the four metrics considered (Estimation Error, Unrealized Returns and Extreme Risk). The best perform- ing one in terms of Stability is SPEA2 R+T. Among the standard MOEAS, GDE3 provides the highest robustness 11 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 12 Algorithm 5 Statical testing protocol if the values follow a Normal distribution (applying the Kolmogorov Smirnov test) then if the variances are homogeneity (the Levene test is used to check it) then A t-test is performed. else A Welch test is executed. end if else A Wilcoxon test is applied to compare the medians of the solutions end if followed by SMPSO. As we have have seen, the proposed R+T mechanism resulted in a systematic and significant increase in the reliability of solutions across the algorithms. The introduced R+T mechanism generates more robust solutions than any of the basic approaches of the MOEAs. The minimum decrease in average metric value was 19.56% evaluating Unrealized Returns for GDE3. The largest improvement, 67.53%, was achieved SPEA2 for Stability. Furthermore, we also observed that our approach helped to improve those algorithms which produced more unstable solutions. 6. Conclusions Portfolio optimization represents one of the main topics in financial research. The basic framework targets the combination of financial assets to be included in a portfolio in order to optimize the balance of risk and return. This process is usually based on two parameters: the estimates for the expected asset returns, and the variance-covariance matrix. Quadratic Programming can be used to solve the basic version of this problem. However, in the real world there are some constraints that cannot be easily tackled by traditional techniques. This is the reason why evolutionary multiobjective algorithms (MOEAs), that do not present these kind of limitations, are getting traction in the domain. One of the main problems that portfolio managers face is the uncertainty regarding the expected frontier derived from their forecasts for future returns. Very often, it lies far from the actual one, resulting in inaccurate forecast risk/return profiles for the portfolios. Our work is focused on this point, adding robustness to the results by preventing optimizing for a single expected scenario that may produce solutions that are hyper-specialized and might be extremely sensitive to likely deviations. We handle the uncertainty in these parameters during the optimization process testing the population for different values of parameters and selecting the portfolios that consistently offer a good performance. We also use a time-stamping mechanism that, in collaboration with performance in terms of risk and return, drives the evolution process to find robust and stable solutions. The assessment of the robustness of the efficient frontiers resulting from the process requires metrics that differ from the most traditional ones. For this reason, we introduce a set of four metrics ”Estimation Error”, ”Stability”, ”Extreme Risk” and ”Unrealized Returns”. The approach was tested through experimentation using four popular MOEAs (NSGA-II, SPEA2, SMPSO and GDE3). The results were compared in both the standard and the time-stamped resampled versions of the algorithms. They were tested on a sample of monthly returns for eight indexes representing different broad investment categories including stock, bonds, etc. The results show that the suggested approach enhances significantly the reliability of solutions for all algorithms. The time-stamped resampled versions improved the results across all the considered MOEAs. The improvements got up to 67.53% in ”Stability”, 54.93% in ”Estimation Error”, 52.92% in ”Extreme Risk” and 33.32% in ”Unrealized Returns”. These improvements were higher when standard MOEA provided worse results in terms of the metrics. We also observed that the R+T version of SMPSO tended to outperform the alternatives. Even though these results are good and promising in terms of robustness, there are several issues left open that could lead to future extensions of this work. Among them, the analysis of the performance of this robust approach on other MOEAs or the scalability of the results with different number of investment alternatives. 7. Acknowledgements The authors acknowledge financial support granted by the Spanish Ministry of Science under contract TIN2008- 06491-C04-03 (MSTAR), TIN2011-28336(MOVES) and Comunidad de Madrid (CCG10-UC3M/TIC-5029). 12 Garcia et al. / Expert Systems with Applications 00 (2012) 1–13 13 References [1] Anagnostopoulos, K., Mamanis, G., 2011. The mean-variance cardinality constrained portfolio optimization problem: An experimental evaluation of five multiobjective evolutionary algorithms. Expert Systems with Applications 38 (11), 14208–14217. [2] Barbosa, H. J., Lemonge, A. C., 2005. A genetic algorithm encoding for a class of cardinality constraints. In: Proceedings of the 2005 conference on Genetic and evolutionary computation. GECCO ’05. ACM, New York, NY, USA, pp. 1193–1200. [3] Chang, T.-J., Yang, S.-C., Chang, K.-J., 2009. Portfolio optimization problems in different risk measures using genetic algorithm. Expert Systems with Applications 36 (7), 10529 – 10537. [4] Chiranjeevi, C., Sastry, V. N., 2007. Multi objective portfolio optimization models and its solution using genetic algorithms. Computational Intelligence and Multimedia Applications, International Conference on 1, 453–457. [5] Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation 6 (2), 182–197. [6] Deb, K., Steuer, R. E., Tewari, R., Tewari, R., 2011. Bi-objective portfolio optimization using a customized hybrid NSGA-II procedure. In: EMO. pp. 358–373. [7] Durillo, J., Nebro, A., Alba, E., July 2010. The jmetal framework for multi-objective optimization: Design and architecture. In: CEC 2010. Vol. 5467 of Lecture Notes in Computer Science. Springer Berlin / Heidelberg, Barcelona, Spain, pp. 4138–4325. [8] Garcia, S., Quintana, D., Galvan, I. M., Isasi, P., 2011. Portfolio optimization using SPEA2 with resampling. In: Proceedings of the 12th international conference on Intelligent data engineering and automated learning. IDEAL11. Springer-Verlag, Berlin, Heidelberg, pp. 127–134. [9] H.M., M., 1959. Portfolio Selection: efficient diversification of investments. John Wiley & Son. [10] Idzorek, T. M., 2006. Developing robust asset allocations. Tech. rep. [11] Kukkonen, S., Lampinen, J., 2005. GDE3: The third evolution step of generalized differential evolution. In: IEEE Congress on Evolutionary Computation (CEC’2005). pp. 443 – 450. [12] Markowitz, H., 1952. Portfolio selection. The Journal of Finance 7 (1), 77–91. [13] Nebro, A., Durillo, J., Garcı́a-Nieto, J., Coello Coello, C., Luna, F., Alba, E., 2009. SMPSO: A new PSO-based metaheuristic for multi- objective optimization. In: 2009 IEEE Symposium on Computational Intelligence in Multicriteria Decision-Making (MCDM 2009). IEEE Press, pp. 66–73. [14] Perret-Gentil, C., Victoria-Feser, M.-P., Apr. 2005. Robust mean-variance portfolio selection. Fame research paper series, International Center for Financial Asset Management and Engineering. [15] Pflug, G., Wozabal, D., 2007. Ambiguity in portfolio selection. Quantitative Finance 7 (4), 435–442. [16] Radziukyniene, I., Xilinskas, A., 2008. Evolutionary methods for multi-objective portfolio optimization. In: Ao, S. I., Gelman, L., Hukins, D. W., Hunter, A., Korsunsky, A. M. (Eds.), Proceedings of the World Congress on Engineering 2008 Vol II, WCE ’08, July 2 - 4, 2008, London, U.K. Lecture Notes in Engineering and Computer Science. International Association of Engineers, Newswood Limited, pp. 1155– 1159. [17] Reyes, M., Coello, C. C., 2005. Improving PSO-based multi-objective optimization using crowding, mutation and �-dominance. In: Coello, C., Hernández, A., Zitler, E. (Eds.), Third International Conference on Evolutionary MultiCriterion Optimization, EMO 2005. Vol. 3410 of LNCS. Springer, pp. 509–519. [18] Ruppert, D., June 2006. Statistics and finance: An introduction. Journal of the American Statistical Association 101, 849–850. [19] Shiraishi, H., 2008. Resampling procedure to construct value at risk efficient portfolios for stationary returns of assets. Tech. rep. [20] Skolpadungket, P., Dahal, K., Harnpornchai, N., 2007. Portfolio optimization using multi-objective genetic algorithms. In: Proceeding of 2007 IEEE Congress on Evolutionary Computation. pp. 516–523. [21] Soleimani, H., Golmakani, H. R., Salimi, M. H., April 2009. Markowitz-based portfolio selection with minimum transaction lots, cardinality constraints and regarding sector capitalization using genetic algorithm. Expert Syst. Appl. 36, 5058–5063. [22] Tütüncü, R. H., Koenig, M., 2004. Robust asset allocation. Annals OR 132 (1-4), 157–187. [23] Zhu, H., Wang, Y., Wang, K., Chen, Y., 2011. Particle swarm optimization (PSO) for the constrained portfolio optimization problem. Expert Systems with Applications 38 (8), 10161 – 10169. [24] Zitzler, E., Laumanns, M., Thiele, L., 2001. SPEA2: Improving the strength pareto evolutionary algorithm. Tech. Rep. 103, Computer Engineering and Networks Laboratory (TIK), Swiss Federal Institute of Technology (ETH), Zurich, Switzerland. [25] Zitzler, E., Thiele, L., 1998. An evolutionary algorithm for multiobjective optimization: The strength pareto approach. Tech. Rep. 43, ss, Gloriastrasse 35, CH-8092 Zurich, Switzerland. URL citeseer.ist.psu.edu/article/zitzler99evolutionary.html 13 
work_gr5hjfqjdrfy5oniy6sunevkny	----	Associative learning on imbalanced environments: An empirical study L. Cleofas-Sáncheza, J.S. Sáncheza,, V. Garcı́ab, R.M. Valdovinosc aInstitute of New Imaging Technologies, Department of Computer Languages and Systems, Universitat Jaume I, Castelló de la Plana, Spain E-mail: cleofas@uji.es, sanchez@uji.es bDivisión Multidisciplinaria de Ciudad Universitaria, Universidad Autónoma de Ciudad Juárez, Ciudad Juárez, Chihuahua, Mexico E-mail: vicente.jimenez@uacj.mx cSchool of Engineering, Universidad Autónoma del Estado de México, Toluca, Mexico E-mail: li rmvr@hotmail.com Abstract Associative memories have emerged as a powerful computational neural network model for several pattern classification problems. Like most traditional classifiers, these models assume that the classes share similar prior probabilities. However, in many real-life applications the ratios of prior probabilities between classes are extremely skewed. Although the literature has provided numerous studies that examine the performance degradation of renowned classifiers on different imbal- anced scenarios, so far this effect has not been supported by a thorough empirical study in the context of associative memories. In this paper, we fix our attention on the applicability of the associative neural networks to the classification of imbal- anced data. The key questions here addressed are whether these models perform better, the same or worse than other popular classifiers, how the level of imbal- ance affects their performance, and whether distinct resampling strategies produce a different impact on the associative memories. In order to answer these ques- tions and gain further insight into the feasibility and efficiency of the associative memories, a large-scale experimental evaluation with 31 databases, seven classi- fication models and four resampling algorithms is carried out here, along with a non-parametric statistical test to discover any significant differences between each pair of classifiers. Keywords: Associative memory, Class imbalance, Resampling Preprint submitted to Expert Systems with Applications May 27, 2016 1. Introduction Associative memories have been an active research area in pattern recognition and neuroscience for over 50 years (Palm, 2013). Although typical applications of these connectionist models include image recognition and recovery, data analysis, control, inference and prediction (Danilo et al., 2015; Kareem and Jantan, 2011; Lou and Cui, 2007; Mu et al., 2006; Nazari et al., 2014; Štanclová and Zavoral, 2005), the associative memories have lately emerged as useful classifiers for a large variety of problems in data mining and computational intelligence (Aldape- Pérez et al., 2012, 2015; Sharma et al., 2008; Uriarte-Arcia et al., 2014). From a practical point of view, classification refers to the assignment of a fi- nite set of samples to predefined classes based on a number of observed variables or attributes. The effective design of classifiers is a complex task where the under- lying quality of data becomes critical to further achieve accurate classifications. Besides other factors, the performance of classifiers strongly depends on the ad- equacy of the data in terms of several intrinsic characteristics such as number of examples, relevance of the attributes, class distribution and completeness of data. In particular, a very common situation in real-life classification problems is to find data sets where the amount of examples available for one class is quite different from that of the other classes. This significant difference in the size of the classes is usually known as class imbalance, and it has been observed that most traditional classifiers perform poorly on the minority class because they assume an even class distribution and equal misclassification costs (Japkowicz and Stephen, 2002). Examples of applications with skewed class distributions can be found in many different areas ranging from bioinformatics and medicine to finance, network se- curity and text mining. For instance, in credit risk and bankruptcy prediction, the number of observations in the class of defaulters is much smaller than the num- ber of cases belonging to the class of non-defaulters (Kim et al., 2015). Gene expression microarray analysis for cancer classification shows significant differ- ences between the number of cancerous and healthy tissue samples (Blagus and Lusa, 2010). Web spam detection exhibits an asymmetric distribution between legitimate and spam pages (Fdez-Glez et al., 2015). Within this context, the major research line has traditionally been directed to the development of techniques to tackle the class imbalance at both algorithmic and data levels (He and Garcia, 2009; He and Ma, 2013; López et al., 2013). The methods at the algorithmic level modify the existing learning models for bias- ing the discrimination process towards the minority class; the data level solutions consist of a preprocessing step to resample the original data set, either by over- 2 sampling the minority class and/or under-sampling the majority class, until the classes are approximately equally represented. In general, the resampling strate- gies have been the most widely used because they are independent of the classifier and can be easily implemented for any problem (Cao et al., 2014; Garcı́a et al., 2012). Over the last years, however, research on this topic has also put the emphasis on studying the effect of imbalance together with other data complexity char- acteristics such as overlapping, small disjuncts and noisy data (He et al., 2015; López et al., 2013; Napierala et al., 2010; Prati et al., 2004; Stefanowski, 2013). Another critical subject that has attracted increasing interest in the scientific com- munity is how to assess the performance of a classification model in the presence of imbalanced data sets because most common metrics (e.g., accuracy and error rates) strongly depend on the class distribution and assume equal misclassification costs, which may lead to distorted conclusions (He and Garcia, 2009; Menardi and Torelli, 2014). In addition to the aforementioned questions, research has also placed its atten- tion on evaluating and comparing the performance of competing classifiers (Brown and Mues, 2012; Kwon and Sim, 2013; López et al., 2013; Liu et al., 2011; Seiffert et al., 2014), but many of these works do not take into consideration the impact of different levels of imbalance on the performance of each particular classification model, neither the implications of using this jointly with one resampling technique or another. While research in class imbalance has mainly concentrated on well-known classifiers such as support vector machines (Akbani et al., 2004; Hwang et al., 2011; Liu et al., 2011; Maldonado and López, 2014; Yu et al., 2015), kernel methods (Hong et al., 2007; Maratea et al., 2014), k-nearest neighbors (Dubey and Pudi, 2013), decision trees (Kang and Ramamohanarao, 2014) and multiple classifier systems (Dı́ez-Pastor et al., 2015; Galar et al., 2012; Krawczyk et al., 2014; Park and Ghosh, 2014), very few theoretical or empirical analyses have been done so far to thoroughly establish the performance of associative memories when learning from class imbalanced data. Therefore, the present study intends to extend the very preliminary existing works (Cleofas-Sánchez et al., 2013, 2014) by increasing the scope and detail at which a type of associative memory networks (the hybrid associative classifier with translation) performs in the framework of class imbalance. We believe that the extensive experimental analysis here carried out will allow to gain a deeper insight into the feasibility and efficiency of these associative memories and help researchers and practitioners to build effective as- sociative learning models and develop techniques based on associative learning 3 to handle the class imbalance problem. To sum up, the purpose of this paper is three-fold: • To explore the performance of the hybrid associative memory with transla- tion and compare this against other popular classification methods of differ- ent nature; • to investigate how the imbalance ratio affects the performance of these as- sociative memories; and • to analyze the impact of several resampling strategies on the performance of the associative neural network. The rest of this paper is organized as follows. Section 2 introduces the bases of the associative memory neural networks and in particular, of the hybrid associative classifier with translation. Section 3 briefly describes four resampling methods to deal with the imbalance problem, which will be further used in the experiments. The experimental set-up and databases are presented in Section 4, while the results and statistical tests are discussed in Section 5. Finally, Section 6 remarks the main conclusions and outlines some avenues for future research. 2. Associative memories The associative memory is an early type of artificial neural network that takes the form of a matrix M generated from a finite set of p previously known asso- ciations, which is called fundamental set {(xµ, yµ) | µ = 1, 2, . . . , p}, where xµ are the fundamental input patterns of dimension n and yµ are the the fundamental output patterns of dimension m. Then, xµj and y µ j denote the j-th component of an input pattern xµ and of an output pattern yµ, respectively. Functionality of the associative memories is accomplished in two phases: learn- ing and recall. The learning process consists of building a matrix M with a value for each association (xk, yk). In the recall phase, an output vector y is obtained from the associative memory; this vector is the most similar to the input vector x. These memories have the capability of storing huge amounts of data that allow to recover the input examples with low computational efforts. The associative memories can be of two types (Aldape-Pérez et al., 2012): heteroassociative (e.g., linear associator) and autoassociative (e.g., Hopfield net- work). The heteroassociative memories relates input patterns with output patterns 4 of distinct nature and formats (xµ ̸= yµ), while the autoassociative memories are a particular case where xµ = yµ and n = m. Some of the most widely-studied models of associative memories are lernma- trix (Steinbuch, 1961), the linear associator (Anderson, 1972; Kohonen, 1972), the Hopfield network (Hopfield, 1982), the bidirectional associative memory (Kosko, 1987), and the morphological associative memory (Ritter et al., 1998). The popu- larity of these models comes from their capability of storing huge amounts of data that allow to properly recover the most similar patterns given an input vector with low computational efforts, but they have some difficulties to discriminate between classes. 2.1. The hybrid associative classifier with translation With the aim of extending the applicability of classical associative memories to classification tasks, Santiago (2003) introduced the hybrid associative classifier with translation (HACT), which is based on the learning phase of the linear as- sociator and the recall phase of the lernmatrix. The learning phase of the linear associator comprises two steps: 1. For each association (xµ, yµ), compute the matrix (yµ)·(xµ)T , where (xµ)T is the transpose input vector. 2. Sum the p matrices to obtain the memory M = ∑p µ=1(y µ) · (xµ)T , being mi,j = ∑p µ=1(y µ i )(x µ j ) T its (i, j)-th component. On the other hand, the recall phase of the lernmatrix consists of finding the class of an input pattern xµ, that is, to find the components of the vector yµ as- sociated to xµ, whose i-th component is calculated according to the following expression: y µ i = { 1 if ∑n j=i mi,jx µ j = max p h=1 [∑n j=i mh,jx µ j ] 0 otherwise (1) Note that if xµ belongs to class c, this expression leads to an m-dimensional vector with all components equal to zero except for the c-th component whose value is equal to 1. In addition, the HACT model incorporates the translation of the axes to a new origin located at the centroid of the fundamental input patterns. Let A = {xµ |µ = 1, 2, . . . , p} be a set of n-dimensional fundamental input patterns that belong to 5 Algorithm 1 HACT – learning phase s ← 0 for all xµ ∈ A do s ← s + xµ end for x ← s/p {x is the new origin of the coordinates axes} for all xµ ∈ A do x̂µ ← xµ −x {Translate the input patterns to the new coordinates axes} c ← 1 {Assign an output vector of size m to patterns that belong to class c} while c ≤ m do if class(x̂µ) = c then yµc = 1 else yµc = 0 end if end while end for M ← 0 {Apply the learning phase of the linear associator} for all (x̂µ, yµ) do M ← M + (yµ) · (x̂µ)T end for m classes, then the learning phase of HACT to construct the matrix M consists in the steps of Algorithm 1. Once the matrix M has been generated, the classification of a new input pat- tern x consists of applying the recovery phase of lernmatrix according to expres- sion 1. 3. Resampling methods to handle class imbalance Much research has been conducted to tackle the class imbalance problem by applying different resampling strategies that change the class distribution of the data set (Yang et al., 2011). This can be done by either over-sampling the minor- ity class or under-sampling the majority class until both classes are approximately equally represented. Despite their advantages and popularity, several shortcom- ings remain associated with both these strategies because they artificially alter the prior class probabilities (Barandela et al., 2004): whilst under-sampling may re- sult in throwing away potentially useful information about the majority class and perturbing the a priori probability of the training set (Dal Pozzolo et al., 2015), 6 over-sampling worsens the computational burden of most classifiers, may increase the likelihood of overfitting and introduce noise that could result in a loss of per- formance. 3.1. Over-sampling The simplest method to increase the size of the minority class corresponds to random over-sampling (ROS), which balances the class distribution through the random replication of minority class examples. Although this method appears to be quite effective without adding new information into the data set, it may increase the likelihood of overfitting because it makes exact copies of the minority class examples. In order to avoid overfitting, Chawla et al. (2002) proposed the Synthetic Mi- nority Oversampling TEchnique (SMOTE) to up-size the minority class. Instead of merely replicating examples, this algorithm generates artificial examples of the minority class by interpolating existing instances that lie close together. It first finds the k positive nearest neighbors for each minority class example and then, the synthetic examples are generated in the direction of some or all of those near- est neighbors. SMOTE allows the classifier to build larger decision regions that contain nearby examples of the minority class. Depending upon the amount of over-sampling required, a certain number of examples from the k nearest neigh- bors are randomly chosen. 3.2. Under-sampling Random under-sampling (RUS) aims at balancing the data set through the ran- dom removal of examples that belong to the over-sized class. Despite important information can be lost when examples are discarded at random, it has empirically been shown to be one of the most effective resampling methods. Unlike the random approach, other under-sampling methods are based on a deterministic selection of the examples to be eliminated. For instance, Laurikkala (2001) introduced the Neighborhood CLeaning rule (NCL), which uses the Wil- son’s editing algorithm (Wilson, 1972) to remove only the majority class cases that border the minority class examples, assuming that they are noise. 4. Experimental set-up With the aim of evaluating the performance of the HACT model on imbal- anced domains, we have carried out a series of experiments on a large collection of imbalanced data sets using four resampling algorithms (ROS, SMOTE, RUS, 7 and NCL) to preprocess the original data sets and seven classifiers implemented with the default parameter settings given by the WEKA toolkit (Hall et al., 2009): the Bayes network (BNet), a multilayer perceptron network with backpropaga- tion (MLP), a normalized Gaussian radial basis function (RBF), a support vector machine with a linear kernel(SVM), a pruned decision tree (C4.5), the nearest neighbor with generalization (NNge), and the voted perceptron (VP). Note that we have chosen classifiers of different nature and complexity in order to draw meaningful conclusions, including four neural networks (BNet, MLP, RBF, VP), a linear classifier (SVM), a decision tree (C4.5), and an instance-based learner (NNge). 4.1. Data sets The experimental set-up has been designed following the common practice (Akbani et al., 2004; Barua et al., 2014; Dı́ez-Pastor et al., 2015; Galar et al., 2013; López et al., 2012, 2013; Yu et al., 2013, 2015), which consists of gener- ating skewed data sets with different levels of class imbalance. In particular, the scheme proposed in the literature is to transform multi-class data sets by com- bining several original classes to shape the majority and minority classes. For instance, Glass0123 vs 456 is an imbalanced version of Glass database where the classes 0, 1, 2 and 3 have been combined to form a unique minority class and the original classes 4, 5 and 6 have been joined to represent the majority class. Simi- larly, Glass6 has been created by leaving the original class 6 as the minority class and merging the rest of the classes into a single majority class. The results are based on 31 imbalanced data sets taken from the KEEL Data Set Repository (Alcalá-Fdez et al., 2011), whose main characteristics are summa- rized in Table 1. These data sets have been divided into two groups according to their imbalance ratio (IR), that is, the ratio of the majority class size to the minority class size. A first group comprises data sets with a low/moderate imbalance ratio (IR < 10), while the second category gathers the strongly imbalanced databases (IR ≥ 10). The five-fold cross-validation method has been adopted for the experimental design because it seems to be the best estimator of classification performance compared to other methods, such as bootstrap with a high computational cost or re-substitution with a biased behavior. Each original data set has randomly been divided into five stratified parts of of size n/5 (where n denotes the total number of samples in the data set); for each fold, four blocks have been pooled as the training set, and the remaining portion has been used as an independent test set. 8 Table 1: Data sets used in the experiments Data set #Samples #Attributes IR Data set #Samples #Attributes IR Wisconsin 683 9 1.86 Ecoli0347 vs 56 257 7 9.28 Yeast1 1484 8 2.46 Yeast05679 vs 4 528 8 9.35 Haberman 306 3 2.78 Ecoli067 vs 5 220 6 10.00 Vehicle1 846 18 2.90 Led7digit02456789 vs 1 443 7 10.97 Vehicle3 846 18 2.99 Ecoli01 vs 5 240 6 11.00 Glass0123 vs 456 214 9 3.20 Yeast1 vs 7 459 7 14.30 Ecoli2 336 7 5.46 Ecoli4 336 7 15.80 Ecoli1 336 7 3.36 Glass4 214 9 15.47 Segment0 2308 19 6.02 Yeast1458 vs 7 693 8 22.10 Glass6 214 9 6.38 Glass5 214 9 22.78 Yeast3 1484 8 8.10 Yeast2 vs 8 482 8 23.10 Ecoli3 336 7 8.60 Yeast4 1484 8 28.10 Ecoli034 vs 5 200 7 9.00 Yeast1289 vs 7 947 8 30.57 Yeast0256 vs 3789 1004 8 9.14 Ecoli0137 vs 26 281 7 39.14 Ecoli046 vs 5 203 6 9.15 Yeast6 1484 8 41.40 Ecoli0346 vs 5 205 7 9.25 The classification models have been applied to the sets preprocessed by the resampling strategies and also to each original training set, which can be deemed as a baseline for comparison. The results from classifying the test samples have been averaged across the five runs and then evaluated for significant differences using the non-parametric Wilcoxon signed-rank test (Demšar, 2006). 4.2. Evaluation criteria Although the accuracy (and its counterpart, the error rate) is the most com- monly used metric to evaluate the performance of classification systems, it has been demonstrated not to be appropriate when the prior class probabilities are very different because it does not consider misclassification costs, is strongly bi- ased to favor the majority class, and is very sensitive to class skews (Daskalaki et al., 2006; Japkowicz, 2013). Therefore, the effectiveness of classifiers in im- balanced domains has traditionally been evaluated by using other criteria, such as precision, area under the ROC curve, F -measure, and geometric mean of accura- cies. In this paper, the geometric mean of accuracies has been chosen because it represents a good trade-off between the rates measured separately on each class. Gmean = √ TPrate ·TNrate (2) where TPrate and TNrate denote the accuracy on the minority and majority classes, respectively. 9 5. Experimental results and discussion The purpose of these experiments is three-fold: (i) to evaluate the performance of the HACT model in comparison to that of other classifiers when the data set is imbalanced; (ii) to find out whether there exist differences in its behavior de- pending on the imbalance ratio; and (iii) to determine how the application of re- sampling strategies affects the effectiveness of the HACT classifier. At this point, we believe that the degree of difficulty of classification of imbalanced data sets depends on their class imbalance ratio. This hypothesis can be extended to the resampling techniques and to the different classifiers. This is the reason why the experiments and analysis have been performed according to two levels of imbal- ance. The Gmean results have been tested for statistically significant differences by means of non-parametric tests, which are generally preferred over the para- metric methods because the usual assumptions of independence, normality and homogeneity of variance are often violated due to the non-parametric nature of the problems (Demšar, 2006). Both pairwise and multiple comparisons have been used in this paper. First, we have computed the Friedman’s ranking of the clas- sifiers for each resampling strategy and each data set independently according to the Gmean results: as there are eight competing classifiers, the ranks for each data set go from 1 (best) to 8 (worst); in case of ties, average ranks are assigned. Then the average rank of each model across all data sets has been calculated. Afterwards, the Wilcoxon’s paired signed-rank test has been used to find out statistically significant differences between each pair of classifiers for two levels of significance (α = 0.10 and α = 0.05). This statistic ranks the differences in performances of two algorithms for each data set, ignoring the signs, and com- pares the ranks for the positive and the negative differences. 5.1. Results on databases with low/moderate imbalance The detailed results over each problem are given in the Appendix. Here the Friedman’s average ranks for the eight classification models have been plotted in Figure 1. For the original training sets, the MLP and NNge classifiers arise as the algorithms with the lowest rankings, that is, the highest performance in average; when using the under-sampling strategies to balance the data sets, it seems that the best classifiers depend on the particular algorithm used. With the over-sampling techniques, SVM yields the lowest rankings. Despite the HACT model has never obtained the lowest average ranks, it is worth noting that their rankings are not too far from those of the best classifiers. 10 Figure 1: Friedman’s average ranks for databases with low/moderate imbalance Tables 2–4 report the Wilcoxon’s test for the data sets with a low/moderate im- balance ratio. The upper diagonal half of each table corresponds to a significance level of α = 0.10 (10% or less chance), whereas the lower diagonal half is for α = 0.05. The symbol “•” indicates that the classifier in the row significantly out- performs the classifier in the column, and the symbol “◦” means that the classifier in the column performs significantly better than the classifier in the row. Table 2: Wilcoxon’s statistic for the original training sets with low/moderate imbalance (1) (2) (3) (4) (5) (6) (7) (8) (1) HACT – • • (2) BNet – o • • (3) MLP – • • • • (4) RBF ◦ – • • (5) SVM ◦ ◦ ◦ ◦ – ◦ ◦ • (6) C4.5 ◦ • – ◦ • (7) NNge • • – • (8) VP ◦ ◦ ◦ ◦ ◦ ◦ ◦ – Results in Table 2 show that the HACT model performs significantly better than the SVM and VP classifiers at both significance levels when the imbalanced training set has not been preprocessed. In this scenario, the performance of the MLP neural network is significantly better than that of the RBF, SVM, C4.5 and VP algorithms, while NNge outperforms SVM, C4.5 and VP. Although both MLP and NNge appears to be the best alternatives when data sets present a low or moderate imbalance ratio, it has to be noted that the differences between these 11 two classifiers and HACT are not statistically significant and their computational cost is high when the data set size is large. In Tables 3–4, one can observe that the behavior of HACT varies considerably depending on whether the data sets are under- or over-sampled: while the removal of majority class examples seems not to change its performance significantly, the over-sampling algorithms degrade the effectiveness of the HACT classifier, es- pecially with the SMOTE method (RBF, SVM and NNge perform significantly better than HACT for a significance level of α = 0.10). Table 3: Wilcoxon’s statistic for the under-sampled sets with low/moderate imbalance NCL RUS (1) (2) (3) (4) (5) (6) (7) (8) (1) (2) (3) (4) (5) (6) (7) (8) (1) HACT – • – ◦ • (2) BNet – ◦ • • – • • (3) MLP – • • • – • • • (4) RBF – • • • ◦ – ◦ • (5) SVM ◦ ◦ ◦ – ◦ ◦ • • • – • • (6) C4.5 ◦ • – ◦ • – • (7) NNge • • – • ◦ ◦ – (8) VP ◦ ◦ ◦ ◦ ◦ ◦ ◦ – ◦ ◦ ◦ ◦ ◦ ◦ ◦ – Table 4: Wilcoxon’s statistic for the over-sampled sets with low/moderate imbalance SMOTE ROS (1) (2) (3) (4) (5) (6) (7) (8) (1) (2) (3) (4) (5) (6) (7) (8) (1) HACT – ◦ ◦ ◦ – ◦ • (2) BNet – ◦ ◦ ◦ ◦ ◦ – ◦ ◦ ◦ • (3) MLP • – • • – • • (4) RBF • – • • – ◦ • • (5) SVM • • – • • • • • – • • • (6) C4.5 • – • ◦ ◦ ◦ – • (7) NNge • – • ◦ ◦ ◦ ◦ ◦ ◦ – (8) VP ◦ ◦ ◦ – ◦ – The three values in the cells of Table 5 show how many times the method of the column has been significantly-better/same/significantly-worse, with a significance level of α = 0.05, for the databases with a low/moderate imbalance rate. As can be observed, the MLP neural network performs the best with both the imbalanced sets and the under-sampled sets, while the SVM appears to outperform the other 12 classifiers when the data sets have been over-sized. The HACT model does not present significant differences with the best approaches when using the original training set and the sets under-sampled with the NCL algorithm and therefore, it can considered as one of the best alternatives. However, the use of SMOTE and ROS seems to produce some degradation on the performance of the HACT classifier. Table 5: Summary of the Wilcoxon’s statistic for the sets with low/moderate imbalance HACT Bnet MLP RBF SVM C4.5 NNge VP Original 2/5/0 2/5/0 4/3/0 2/4/1 1/0/6 2/3/2 3/4/0 0/0/7 NCL 1/6/0 2/5/0 3/4/0 2/5/0 1/1/5 2/3/2 3/4/0 0/0/7 RUS 1/5/1 1/6/0 3/4/0 1/4/2 4/3/0 1/5/0 1/4/2 0/0/7 SMOTE 0/6/1 0/2/5 1/6/0 2/5/0 3/4/0 1/6/0 2/5/0 0/4/3 ROS 1/5/1 1/3/3 3/4/0 3/3/1 6/1/0 1/4/3 0/1/6 0/6/1 5.2. Results on databases with high imbalance Figure 2 shows the Friedman’s average ranks for the eight classification mod- els under a high imbalance ratio. For the imbalanced sets, the lowest rankings correspond to the MLP, Bnet and HACT neural networks. Although there is not a best performing classifier with the resampling algorithms, it seems that the HACT model is one of the approaches with lower rankings independently of the tech- nique used. Figure 2: Friedman’s average ranks for strongly imbalanced databases 13 The Wilcoxon’s test reported in Table 6 reveals that the performance of HACT is significantly better than that of SVM, NNge and VP at both significance levels when the original training sets are strongly imbalanced. In this sense, its behavior seems to be similar to that of MLP since this classifier performs significantly better than RBF, SVM, C4.5, NNge and VP. Table 6: Wilcoxon’s statistic for the original training sets with high imbalance (1) (2) (3) (4) (5) (6) (7) (8) (1) HACT – • • • • (2) BNet – • • (3) MLP – • • • • • (4) RBF ◦ – • • (5) SVM ◦ ◦ ◦ ◦ – ◦ ◦ • (6) C4.5 ◦ – • (7) NNge ◦ ◦ • – • (8) VP ◦ ◦ ◦ ◦ ◦ ◦ ◦ – When the data sets with a high imbalance rate are under-sampled with the NCL algorithm, Table 7 shows that the HACT model achieves significantly better results than SVM and VP at a significance level of 0.05. The MLP neural network performs significantly better than BNet, RBF, SVM, NNge and VP, but we do not find statistically significant differences between the performance of MLP and that of HACT. In the case of random under-sampling, the SVM appears to be the best performing classifier since its Gmean results are significantly better than those of the remaining algorithms for α = 0.05. Nevertheless, the HACT neural network is still better than VP and is not significantly worse than the other classifiers. Table 7: Wilcoxon’s statistic for the under-sampled sets with high imbalance NCL RUS (1) (2) (3) (4) (5) (6) (7) (8) (1) (2) (3) (4) (5) (6) (7) (8) (1) HACT – • • • – • ◦ • (2) BNet – ◦ ◦ – ◦ • (3) MLP • – • • • • • – ◦ • (4) RBF ◦ – • • – ◦ • • (5) SVM ◦ ◦ ◦ – ◦ • • • • – • • • (6) C4.5 – • ◦ – • (7) NNge ◦ • – • ◦ ◦ – • (8) VP ◦ ◦ ◦ ◦ – ◦ ◦ ◦ ◦ ◦ ◦ – 14 On the other hand, when the training sets are over-sampled by means of the SMOTE method (see Table 8), it seems that no classifier performs significantly the best for any value of α. In this case, the BNet approach clearly corresponds to the worst model since its performance is significantly lower than that of any other classifier. With random over-sampling, the conclusions are similar to those drawn for the RUS algorithm, that is, the SVM seems to be the best performing method. However, in this scenario, the performance of HACT is statistically the same as that of the SVM and better than that of BNet and NNge. Table 8: Wilcoxon’s statistic for the over-sampled sets with high imbalance SMOTE ROS (1) (2) (3) (4) (5) (6) (7) (8) (1) (2) (3) (4) (5) (6) (7) (8) (1) HACT – • ◦ • – • • • (2) BNet ◦ – ◦ ◦ ◦ ◦ ◦ ◦ – ◦ ◦ ◦ ◦ ◦ (3) MLP • – • • – ◦ • (4) RBF • – • – ◦ • (5) SVM • – • • • • – • • (6) C4.5 • ◦ – • ◦ – • (7) NNge – ◦ ◦ ◦ ◦ ◦ – ◦ (8) VP • – • • – Table 9 summarizes how many times the method of the column has been significantly-better/same/significantly-worse, with a significance level of α = 0.05, for the strongly imbalanced databases. In this case, the HACT model is among the best performing methods, especially with the original training set and the over-sampled sets. With the RUS algorithm, the SVM is clearly the best al- ternative because it has been significantly better than all the remaining classifiers. On the other hand, VP and Bnet appear to be the worst models: the former with the under-sampling techniques and the latter with the over-sampling algorithms. Table 9: Summary of the Wilcoxon’s statistic for the sets with high imbalance HACT Bnet MLP RBF SVM C4.5 NNge VP Original 3/4/0 2/5/0 5/2/0 2/4/1 1/1/5 1/5/1 2/3/2 0/0/7 NCL 2/5/0 0/6/1 5/2/0 2/4/1 0/3/4 0/7/0 2/4/1 0/3/4 RUS 1/5/1 0/6/1 1/5/1 2/4/1 7/0/0 1/5/1 1/4/2 0/1/6 SMOTE 1/6/0 0/1/6 2/5/0 1/6/0 1/6/0 1/5/1 0/7/0 1/6/0 ROS 2/5/0 0/1/6 2/4/1 2/4/1 5/2/0 2/4/1 0/1/6 1/5/1 15 6. Conclusions and future work This paper intends to cover some questions on the application of the HACT neural network, which is a type of autoassociative memory, to problems charac- terized by skewed class distributions. Although earlier related works have given some insights into the effects of the class imbalance problem on this model, var- ious interesting issues have not been studied in depth. Taking this into account, the main contributions can be summarized as follows: (i) we have empirically analyzed the performance of the HACT model in comparison to that of several standard classifiers; (ii) we have discussed the behavior of these classifiers as a function of the imbalance ratio; and (iii) we have studied how the use of different resampling techniques may affect the performance of the associative memory. Experimental results on a large collection of imbalanced data sets using four resampling techniques and seven different classifiers allow to remark some im- portant findings that were not discussed in the few preliminary works: • When the imbalanced training set is not preprocessed by some resampling algorithm, the HACT model performs as well as the best classifier in most situations; • the superiority of HACT with respect to other classifiers is more significant in the case of strongly imbalanced data sets, both on the original (imbal- anced) training sets and on the balanced sets; • the resampling strategies affect differently the performance of HACT de- pending on the imbalance ratio. Thus, when the imbalance is low or mod- erate, the over-sampling methods degrade the effectiveness of the HACT model considerably. However, with strongly imbalanced data, both under- and over-sampling have given rise to very similar results. Despite the promising results in terms of the geometric mean of acuracies exhibited by the HACT model and supported by non-parametric statistical tests, there are still several interesting open research questions to work on. For instance, the present paper has focused on only the class imbalance problem, but a thor- ough and extensive analysis of the behavior of these neural networks when data sets are affected by multiple complexities deserves further future consideration. Another direction for extending this study is to investigate the applicability of the associative memories on imbalanced big data and data streams in a time-varying environment (concept change and concept drift), which are currently two hot top- ics in the literature related to expert and intelligent information systems. 16 Acknowledgment This work has partially been supported by the Mexican Science and Tech- nology Council (CONACYT-Mexico) through the Postdoctoral Fellowship Pro- gram [232167], the Mexican PRODEP [DSA/103.5/15/7004], the Spanish Min- istry of Economy [TIN2013-46522-P], and the Generalitat Valenciana [PROME- TEOII/2014/062]. Appendix This appendix provides five tables with the Gmean results for the experimen- tal analysis carried out in the present work. Table 10 contains the values for all the databases and classifiers achieved when using the imbalanced training sets, Table 11–14 shows the results with the sets preprocessed by NCL, random under- sampling, SMOTE, and random over-sampling. References Akbani, R., Kwek, S., Japkowicz, N., 2004. Applying support vector machines to imbalanced datasets. In: 15th European Conference on Machine Learning. Pisa, Italy, pp. 39–50. Alcalá-Fdez, J., Fernández, A., Luengo, J., Derrac, J., Garcı́a, S., Sánchez, L., Herrera, F., 2011. KEEL data-mining software tool: Data set repository, integration of algorithms and experimental analysis framework. Journal of Multiple-Valued Logic and Soft Computing 17, 255–287. Aldape-Pérez, M., Yañez Márquez, C., Camacho-Nieto, O., Argüelles-Cruz, A. J., 2012. An associative memory approach to medical decision support systems. Computer Methods and Programs in Biomedicine 106 (3), 287–307. Aldape-Pérez, M., Yañez Márquez, C., Camacho-Nieto, O., López-Yáñez, I., Argüelles-Cruz, A. J., 2015. Collaborative learning based on associative mod- els: Application to pattern classification in medical datasets. Computer in Hu- man Behavior 51 (Part B), 771–779. Anderson, J. A., 1972. A simple neural network generating an interactive memory. Mathematical Bioscience 14, 197–220. 17 Table 10: Geometric mean for the original data sets HACT BNet MLP RBF SVM C4.5 NNge VP wisconsin 97.70 97.38 95.60 96.35 96.98 94.84 96.81 91.56 yeast1 66.20 64.10 62.83 50.69 41.97 63.64 68.58 41.39 haberman 62.65 40.43 49.82 38.81 0.00 48.31 52.11 24,46 vehicle1 63.39 67.57 75.80 63.94 27.97 63.01 70.91 0.00 vehicle3 64.93 67.51 72.66 59.81 0.00 63.51 67.23 0.00 glass-0-1-2-3 vs 4-5-6 92.68 87.89 91.93 89.27 86.51 91.49 70.64 0.00 ecoli1 87.04 84.99 85.34 88.31 80.60 85.78 86.20 72.14 ecoli2 81.13 85.65 89.00 90.65 74.83 85.57 87.18 47.99 segment0 71.86 98.78 99.39 97.71 98.89 98.38 99.00 65.22 glass6 89.12 91.06 83.93 87.00 86.85 79.73 84.16 0.00 yeast3 76.38 84.50 85.07 86.51 67.02 84.99 87.45 24.89 ecoli3 80.54 83.91 76.16 58.37 0.00 68.65 64.31 0.00 ecoli-0-3-4 vs 5 77.46 83.20 88.18 91.42 83.43 79.03 87.94 39.86 yeast-0-2-5-6 vs 3-7-8-9 69.46 72.16 69.36 60.48 31.80 57.87 70.16 46.84 ecoli-0-4-6 vs 5 77.33 88.71 88.47 85.90 80.40 79.53 88.23 22.18 ecoli-0-3-4-6 vs 5 77.51 82.07 88.23 91.70 83.44 79.97 87.49 31.19 ecoli-0-3-4-7 vs 5-6 77.22 69.13 88.47 83.19 71.95 76.78 83.94 34.57 yeast-0-5-6-7-9 vs 4 74.07 41.53 68.70 28.10 0.00 62.54 62.18 20.00 ecoli-0-6-7 vs 5 78.28 80.42 85.73 86.39 67.08 73.60 88.77 0.00 led7digit-0-2-4-5-6-7-8-9 vs 1 79.58 87.67 88.92 81.65 90.17 87.34 56.32 82.62 ecoli-0-1 vs 5 74.47 86.20 89.03 89.03 83.67 79.70 82.72 38.73 yeast-1 vs 7 64.25 32.57 51.25 31.47 0.00 44.43 40.21 0.00 glass4 82.23 56.71 86.69 86.02 25.81 76.88 87.64 0.00 ecoli4 79.39 80.49 88.73 88.59 59.16 79.72 83.13 31.62 yeast-1-4-5-8 vs 7 59.24 0.00 18.22 0.00 0.00 0.00 0.00 0.00 glass5 87.23 91.33 89.00 82.84 0.00 89.23 70.36 0.00 yeast-2 vs 8 76.98 74.08 73.83 77.29 74.08 0.00 70.48 67.01 yeast4 71.36 53.13 54.09 0.00 0.00 44.35 52.21 13.49 yeast-1-2-8-9 vs 7 63.28 40.69 36.45 18.28 0.00 48.26 36.10 0.00 ecoli-0-1-3-7 vs 2-6 78.54 83.36 83.51 83.36 83.51 70.58 62.90 29.94 yeast6 72.33 82.45 69.39 0.00 0.00 75.25 65.09 0.00 18 Table 11: Geometric mean for the data sets preporcessed by NCL HACT BNet MLP RBF SVM C4.5 NNge VP wisconsin 97.58 97.99 96.43 96.68 96.98 94.07 97.29 91.65 yeast1 65.96 59.52 63.01 57.48 44.90 61.07 67.46 39.13 haberman 66.27 42.81 43.68 49.11 31.16 48.37 50.26 11.22 vehicle1 64.28 66.09 66.93 60.80 38.74 62.83 57.43 0.00 vehicle3 64.52 63.38 55.96 48.60 0.00 51.41 58.30 0.00 glass-0-1-2-3 vs 4-5-6 93.97 86.04 89.67 90.43 83.96 88.06 89.126 0.00 ecoli1 85.40 85.86 82.24 88.26 82.19 86.12 86.04 80.95 ecoli2 79.66 81.69 91.30 92.28 77.14 86.74 86.77 59.06 segment0 71.76 98.90 99.28 97.66 98.89 98.30 99.31 53.20 glass6 88.83 87.33 86.62 78.31 86.85 88.33 82.77 0.00 yeast3 7.41 79.81 85.14 85.24 67.00 82.37 87.52 36.79 ecoli3 76.21 62.93 74.46 76.15 0.00 66.83 79.02 16.91 ecoli-0-3-4 vs 5 77.46 85.87 88.18 88.69 83.43 79.94 91.94 44.34 yeast-0-2-5-6 vs 3-7-8-9 68.22 74.43 68.89 68.89 51.09 60.60 70.15 39.05 ecoli-0-4-6 vs 5 78.28 88.96 91.70 83.44 83.44 76.40 83.44 22.18 ecoli-0-3-4-6 vs 5 77.46 88.47 88.72 83.21 83.67 82.76 86.13 0.00 ecoli-0-3-4-7 vs 5-6 75.87 68.98 88.86 84.30 71.95 79.31 79.82 40.00 yeast-0-5-6-7-9 vs 4 71.20 23.70 58.35 57.05 0.00 53.70 59.22 0.00 ecoli-0-6-7 vs 5 77.06 80.42 85.51 80.22 63.25 77.07 83.46 0.00 led7digit-0-2-4-5-6-7-8-9 vs 1 77.22 87.67 91.12 89.04 90.03 83.72 90.51 87.44 ecoli-0-1 vs 5 75.98 86.20 89.03 88.82 83.67 82.90 86.01 38.64 yeast-1 vs 7 62.99 0.00 36.35 8.15 0.00 25.72 25.78 0.00 glass4 81.69 53.37 62.63 44.50 25.81 89.91 54.62 0.00 ecoli4 78.58 80.62 91.90 86.32 54.77 79.84 83.67 0.00 yeast-1-4-5-8 vs 7 56.27 0.00 0.00 0.00 0.00 0.00 0.00 40.00 glass5 69.56 0.00 100.00 44.72 0.00 83.26 44.72 40.00 yeast-2 vs 8 74.08 74.08 74.08 74.00 74.08 0.00 74.08 74.08 yeast4 68.88 0.00 37.10 19.53 0.00 14.14 19.93 14.14 yeast-1-2-8-9 vs 7 56.92 0.00 0.00 0.00 0.00 0.00 0.00 0.00 ecoli-0-1-3-7 vs 2-6 77.12 83.20 83.20 83.51 83.20 70.58 83.51 44.72 yeast6 70.74 79.67 62.98 73.55 0.00 71.59 56.04 0.00 19 Table 12: Geometric mean for the data sets preporcessed by RUS HACT BNet MLP RBF SVM C4.5 NNge VP wisconsin 95.88 97.48 96.20 96.45 97.36 95.23 95.96 89.27 yeast1 67.26 67.60 70.06 66.11 70.00 70.90 67.40 68.43 haberman 60.39 52.25 60.68 52.42 50.90 63.70 56.54 48.87 vehicle1 62.02 66.03 77.01 69.71 74.52 71.66 64.94 62.91 vehicle3 65.00 68.06 77.33 71.32 74.02 73.48 67.38 64.59 glass-0-1-2-3 vs 4-5-6 84.49 89.09 91.77 89.36 85.86 89.78 91.13 42.11 ecoli1 86.99 85.09 86.69 88.17 86.88 88.90 87.12 87.15 ecoli2 82.78 89.63 88.57 87.11 88.63 85.62 86.94 80.02 segment0 75.68 99.01 99.02 97.10 98.11 98.29 99.34 89.12 glass6 87.81 92.06 85.77 84.77 90.54 85.16 87.32 37.50 yeast3 85.30 90.53 90.48 88.71 89.39 91.57 92.14 86.37 ecoli3 88.07 87.15 82.76 83.86 86.36 85.21 80.92 81.04 ecoli-0-3-4 vs 5 88.24 93.30 87.19 85.63 88.54 83.67 82.20 81.85 yeast-0-2-5-6 vs 3-7-8-9 76.37 75.75 74.91 77.51 78.36 79.38 72.60 77.01 ecoli-0-4-6 vs 5 86.49 84.20 88.47 83.68 88.62 79.81 81.77 76.68 ecoli-0-3-4-6 vs 5 85.99 91.21 81.30 88.38 87.07 83.85 85.74 71.68 ecoli-0-3-4-7 vs 5-6 86.49 80.05 92.98 85.08 89.90 82.83 82.07 80.06 yeast-0-5-6-7-9 vs 4 79.60 78.87 75.40 73.75 78.38 72.80 73.00 78.37 ecoli-0-6-7 vs 5 85.25 85.91 87.46 79.37 86.24 84.85 86.24 69.61 led7digit-0-2-4-5-6-7-8-9 vs 1 88.54 88.45 85.83 85.41 89.47 89.24 73.80 85.96 ecoli-0-1 vs 5 89.67 88.78 85.71 86.55 88.34 85.49 82.68 72.37 yeast-1 vs 7 67.11 45.40 69.07 71.47 77.73 59.43 67.12 58.43 glass4 86.43 86.17 88.79 87.10 86.21 78.72 85.90 39.37 ecoli4 92.09 90.60 93.55 90.09 95.48 84.90 93.20 52.84 yeast-1-4-5-8 vs 7 54.00 0.00 53.36 57.36 59.90 53.14 57.30 41.94 glass5 56.57 90.52 91.59 96.27 90.79 88.66 84.84 18.49 yeast-2 vs 8 74.09 74.08 65.45 70.82 72.53 69.68 63.44 59.48 yeast4 81.39 81.13 77.50 79.96 81.70 83.97 78.58 76.96 yeast-1-2-8-9 vs 7 64.62 42.03 68.96 64.48 73.89 66.67 62.47 46.07 ecoli-0-1-3-7 vs 2-6 80.86 60.95 70.74 80.24 80.87 76.51 78.86 38.14 yeast6 87.81 87.40 83.03 88.67 88.69 82.10 80.03 86.58 20 Table 13: Geometric mean for the data sets preporcessed by SMOTE HACT BNet MLP RBF SVM C4.5 NNge VP wisconsin 97.70 97.38 95.60 96.35 96.98 94.84 96.81 91.56 yeast1 66.21 71.13 71.30 69.84 70.42 70.58 69.13 68.19 haberman 62.48 63.43 56.13 54.32 47.78 65.78 58.15 45.40 vehicle1 63.04 68.51 74.99 69.04 72.01 69.77 72.90 52.65 vehicle3 64.85 68.01 77.21 72.25 71.75 69.63 68.27 62.30 glass-0-1-2-3 vs 4-5-6 86.55 90.53 91.87 95.23 88.10 89.09 87.92 0.00 ecoli1 87.41 85.25 89.31 88.37 90.40 85.43 89.18 87.53 ecoli2 89.18 84.02 91.34 89.35 90.17 85.57 88.94 89.55 segment0 75.27 97.98 99.86 97.20 99.16 99.06 99.00 98.66 glass6 88.45 93.31 88.82 88.33 89.31 89.29 90.19 58.89 yeast3 86.85 46.94 90.71 89.16 88.80 91.88 92.83 89.59 ecoli3 87.47 78.85 85.16 88.81 89.56 77.40 85.73 87.64 ecoli-0-3-4 vs 5 89.07 87.93 88.18 90.12 89.32 89.59 88.19 91.33 yeast-0-2-5-6 vs 3-7-8-9 76.81 68.67 75.31 75.30 79.39 75.36 75.80 78.45 ecoli-0-4-6 vs 5 87.84 87.73 86.50 88.72 89.15 89.14 90.93 86.48 ecoli-0-3-4-6 vs 5 88.12 82.30 89.67 90.94 89.67 89.16 90.18 83.30 ecoli-0-3-4-7 vs 5-6 86.44 83.54 86.11 88.82 90.10 88.82 89.46 89.06 yeast-0-5-6-7-9 vs 4 81.06 70.67 74.60 75.80 78.68 77.41 74.62 77.30 ecoli-0-6-7 vs 5 83.49 81.47 87.46 84.62 86.49 87.46 88.43 85.38 led7digit-0-2-4-5-6-7-8-9 vs 1 88.83 84.41 88.05 83.72 88.10 90.28 56.18 88.44 ecoli-0-1 vs 5 89.43 88.82 88.23 85.61 89.86 82.32 85.42 90.67 yeast-1 vs 7 69.29 52.75 60.27 65.15 76.59 62.23 71.62 66.80 glass4 86.73 90.14 91.99 89.91 90.03 83.69 90.98 59.55 ecoli4 93.79 86.05 91.31 95.29 90.87 76.48 94.11 94.00 yeast-1-4-5-8 vs 7 64.59 17.98 51.54 56.97 64.46 33.45 55.34 62.41 glass5 49.73 89.00 94.17 83.26 60.57 99.03 70.54 50.18 yeast-2 vs 8 74.08 58.91 75.08 74.39 74.08 78.86 77.06 76.95 yeast4 82.88 33.54 80.95 81.68 81.98 76.14 79.18 82.56 yeast-1-2-8-9 vs 7 69.13 18.22 64.97 59.90 74.12 54.66 59.03 72.10 ecoli-0-1-3-7 vs 2-6 81.58 70.32 87.62 83.20 84.91 70.06 70.19 85.76 yeast6 87.34 33.74 83.55 87.51 87.86 78.99 84.18 88.26 21 Table 14: Geometric mean for the data sets preporcessed by ROS HACT BNet MLP RBF SVM C4.5 NNge VP wisconsin 95.88 97.80 96.19 96.46 97.28 95.34 96.58 92.09 yeast1 66.79 71.77 69.94 66.73 70.63 69.26 65.33 70.79 haberman 59.60 61.39 55.67 57.59 55.95 54.00 47.10 61.65 vehicle1 62.94 67.94 81.29 70.36 76.00 70.23 58.01 65.60 vehicle3 64.84 67.99 81.35 72.46 73.23 70.89 52.90 65.62 glass-0-1-2-3 vs 4-5-6 84.86 89.48 91.77 91.68 88.68 86.95 82.32 44.02 ecoli1 87.58 86.30 85.70 88.33 89.78 89.31 85.31 88.33 ecoli2 89.45 86.84 91.13 89.01 90.96 86.38 83.97 90.79 segment0 76.02 93.17 99.70 96.91 99.16 99.02 97.54 99.06 glass6 88.07 89.19 85.77 87.57 88.32 86.99 74.00 7.35 yeast3 88.03 89.59 91.32 89.70 88.76 88.96 89.59 88.91 ecoli3 84.12 82.25 86.12 86.45 85.95 78.33 74.33 85.77 ecoli-0-3-4 vs 5 87.46 88.94 87.43 88.44 89.32 92.73 85.14 88.03 yeast-0-2-5-6 vs 3-7-8-9 86.41 75.09 74.35 78.14 79.41 73.61 68.25 79.12 ecoli-0-4-6 vs 5 87.07 76.40 83.26 86.13 88.89 78.17 88.23 84.06 ecoli-0-3-4-6 vs 5 87.07 84.71 88.72 88.23 89.42 81.84 87.49 89.80 ecoli-0-3-4-7 vs 5-6 88.12 81.21 89.45 86.34 90.31 85.48 83.75 86.01 yeast-0-5-6-7-9 vs 4 81.19 66.06 72.85 77.09 78.68 63.91 53.70 77.22 ecoli-0-6-7 vs 5 84.30 85.73 90.80 87.94 86.49 85.29 89.21 86.69 led7digit-0-2-4-5-6-7-8-9 vs 1 88.54 88.31 87.48 87.03 87.97 86.31 89.01 87.96 ecoli-0-1 vs 5 89.39 83.29 88.02 91.35 89.00 76.39 82.72 90.24 yeast-1 vs 7 70.11 54.28 74.28 70.61 77.38 55.87 44.24 71.37 glass4 82.99 47.93 89.22 82.19 90.27 84.13 44.36 54.51 ecoli4 93.81 83.13 91.16 90.14 90.58 82.73 77.21 93.54 yeast-1-4-5-8 vs 7 65.65 41.71 33.80 55.48 64.61 46.83 25.75 64.53 glass5 64.19 44.72 89.23 70.19 93.69 94.17 44.72 48.39 yeast-2 vs 8 74.09 52.90 72.70 69.70 74.08 79.56 70.63 73.35 yeast4 82.47 57.13 74.85 75.88 80.45 63.42 36.74 82.45 yeast-1-2-8-9 vs 7 67.27 49.08 48.14 65.04 73.29 59.55 47.68 70.44 ecoli-0-1-3-7 vs 2-6 85.14 44.64 81.50 83.05 85.09 69.94 70.58 79.13 yeast6 87.90 74.56 79.62 86.38 87.73 78.31 50.62 86.44 22 Barandela, R., Valdovinos, R. M., Sánchez, J. S., Ferri, F. J., 2004. The imbal- anced training sample problem: Under or over sampling? In: Fred, A., Caelli, T. M., Duin, R. P. W., Campilho, A. C., de Ridder, D. (Eds.), Structural, Syntac- tic, and Statistical Pattern Recognition. Vol. 3138 of Lecture Notes in Computer Science. Springer, pp. 806–814. Barua, S., Islam, M., Yao, X., Murase, K., 2014. MWMOTE–Majority weighted minority oversampling technique for imbalanced data set learning. IEEE Trans. on Knowledge and Data Engineering 26 (2), 405–425. Blagus, R., Lusa, L., 2010. Class prediction for high dimensaional class- imbalanced data. BMC Bioinformatics 11 (523), 1–17. Brown, I., Mues, C., 2012. An experimental comparison of classification algo- rithms for imbalanced credit scoring data sets. Expert Systems with Applica- tions 39 (3), 3446–3453. Cao, P., Zhao, D., Zaiane, O., 2014. Hybrid probabilistic sampling with random subspace for imbalanced data learning. Intelligent Data Analysis 18 (6), 1089– 1108. Chawla, N. V., Bowyer, K. W., Hall, L. O., Kegelmeyer, W. P., 2002. SMOTE: Synthetic minority over-sampling technique. Journal of Artifial Intelligence Re- search 16 (1), 321–357. Cleofas-Sánchez, L., Camacho-Nieto, O., Sánchez, J. S., Valdovinos-Rosas, R. M., 2014. Equilibrating the recognition of the minority class in the imbalance context. Applied Mathematics and Information Sciences 8 (1), 27–36. Cleofas-Sánchez, L., Garcı́a, V., Martı́n-Félez, R., Valdovinos, R. M., Sánchez, J. S., Camacho-Nieto, O., 2013. Hybrid associative memories for imbal- anced data classification: An experimental study. In: Carrasco-Ochoa, J. A., Martı́nez-Trinidad, J. F., Rodrı́guez, J. S., di Baja, G. S. (Eds.), Pattern Recog- nition. Vol. 7914 of Lecture Notes in Computer Science. Springer, pp. 325–334. Dal Pozzolo, A., Caelen, O., Bontempi, G., 2015. When is undersampling ef- fective in unbalanced classification tasks? In: Appice, A., Rodrigues, P. P., Santos Costa, V., Soares, C., Gama, J., Jorge, A. (Eds.), Machine Learning and Knowledge Discovery in Databases. Vol. 9284 of Lecture Notes in Computer Science. Springer, pp. 200–215. 23 Danilo, R., Jarollahi, H., Gripon, V., Coussy, P., Conde-Canencia, L., Gross, W., 2015. Algorithm and implementation of an associative memory for oriented edge detection using improved clustered neural networks. In: IEEE Interna- tional Symposium on Circuits and Systems. Lisbon, Portugal, pp. 2501–2504. Daskalaki, S., Kopanas, I., Avouris, N., 2006. Evaluation of classifiers for an uneven class distribution problem. Applied Artificial Intelligence 20 (5), 381– 417. Demšar, J., 2006. Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research 7, 1–30. Dı́ez-Pastor, J. F., Rodrı́guez, J. J., Garcı́a-Osorio, C., Kuncheva, L. I., 2015. Random Balance: Ensembles of variable priors classifiers for imbalanced data. Knowledge-Based Systems 85, 96–111. Dubey, H., Pudi, V., 2013. Class based weighted k-nearest neighbor over imbal- ance dataset. In: Pei, J., Tseng, V., Cao, L., Motoda, H., Xu, G. (Eds.), Ad- vances in Knowledge Discovery and Data Mining. Vol. 7819 of Lecture Notes in Computer Science. Springer, pp. 305–316. Fdez-Glez, J., Ruano-Ordás, D., Fdez-Riverola, F., Méndez, J., Pavón, R., Laza, R., 2015. Analyzing the impact of unbalanced data on web spam classification. In: Omatu, S., Malluhi, Q. M., Gonzalez, S. R., Bocewicz, G., Bucciarelli, E., Giulioni, G., Iqba, F. (Eds.), 12th International Conference on Distributed Com- puting and Artificial Intelligence. Vol. 373 of Advances in Intelligent Systems and Computing. Springer, pp. 243–250. Galar, M., Fernández, A., Barrenechea, E., Bustince, H., Herrera, F., 2012. A review on ensembles for the class imbalance problem: Bagging-, boosting-, and hybrid-based approaches. IEEE Trans. on Systems, Man, and Cybernetics, Part C: Applications and Reviews 42 (4), 463–484. Galar, M., Fernández, A., Barrenechea, E., Herrera, F., 2013. EUSBoost: Enhanc- ing ensembles for highly imbalanced data-sets by evolutionary undersampling. Pattern Recognition 46 (12), 3460–3471. Garcı́a, V., Sánchez, J. S., Mollineda, R. A., 2012. On the effectiveness of pre- processing methods when dealing with different levels of class imbalance. Knowledge-Based Systems 25 (1), 13–21. 24 Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., Witten, I. H., 2009. The WEKA data mining software: an update. SIGKDD Explorations Newsletter 11 (1), 10–18. He, H., Garcia, E. A., 2009. Learning from imbalanced data. IEEE Trans. on Knowledge and Data Engineering 21, 1263–1284. He, H., Ma, Y., 2013. Imbalanced Learning: Foundations, Algorithms, and Appli- cations. Wiley – IEEE Press, Piscataway, NJ. He, M., Weir, J. D., Wu, T., Silva, A., Zhao, D.-Y., Qian, W., 2015. K Nearest Gaussian–A model fusion based framework for imbalanced classification with noisy dataset. Artificial Intelligence Research 4 (2), 126–135. Hong, X., Chen, S., Harris, C., 2007. A kernel-based two-class classifier for im- balanced data sets. IEEE Trans. on Neural Networks 18 (1), 28–41. Hopfield, J. J., 1982. Neural networks and physical systems with emergent collec- tive computational abilities. Proceedings of the National Academy of Sciences 79 (8), 2554–2558. Hwang, J. P., Park, S., Kim, E., 2011. A new weighted approach to imbalanced data classification problem via support vector machine with quadratic cost func- tion. Expert Systems with Applications 38 (7), 8580–8585. Japkowicz, N., 2013. Assessment metrics for imbalanced learning. In: Imbalanced Learning: Foundations, Algorithms, and Applications. Wiley – IEEE Press, pp. 187–210. Japkowicz, N., Stephen, S., 2002. The class imbalance problem: A systematic study. Intelligent Data Analysis 6 (5), 429–449. Kang, S., Ramamohanarao, K., 2014. A robust classifier for imbalanced datasets. In: Tseng, V., Ho, T., Zhou, Z.-H., Chen, A., Kao, H.-Y. (Eds.), Advances in Knowledge Discovery and Data Mining. Vol. 8443 of Lecture Notes in Com- puter Science. Springer, pp. 212–223. Kareem, E. I. A., Jantan, A., 2011. An intelligent traffic light monitor system using an adaptive associative memory. International Journal of Information Process- ing and Management 2 (2), 23–39. 25 Kim, M.-J., Kang, D.-K., Kim, H. B., 2015. Geometric mean based boosting al- gorithm with over-sampling to resolve data imbalance problem for bankruptcy prediction. Expert Systems with Applications 42 (3), 1074–1082. Kohonen, T., 1972. Correlation matrix memories. IEEE Trans. on Computers C– 21, 353–359. Kosko, B., 1987. Adaptive bidirectional associative memories. Applied Optics 26 (23), 4947–4960. Krawczyk, B., Woniak, M., Schaefer, G., 2014. Cost-sensitive decision tree en- sembles for effective imbalanced classification. Applied Soft Computing 14, Part C, 554–562. Kwon, O., Sim, J. M., 2013. Effects of data set features on the performances of classification algorithms. Expert Systems with Applications 40 (5), 1847–1857. Laurikkala, J., 2001. Improving identification of difficult small classes by balanc- ing class distribution. In: 8th Conference on Artificial Intelligence in Medicine. Cascais, Portugal, pp. 63–66. Liu, Y., Yu, X., Huang, J. X., An, A., 2011. Combining integrated sampling with SVM ensembles for learning from imbalanced datasets. Information Processing & Management 47 (4), 617–631. López, V., Fernández, A., Garcı́a, S., Palade, V., Herrera, F., 2013. An insight into classification with imbalanced data: Empirical results and current trends on using data intrinsic characteristics. Information Sciences 250, 113–141. López, V., Fernández, A., Moreno-Torres, J. G., Herrera, F., 2012. Analysis of pre- processing vs. cost-sensitive learning for imbalanced classification. open prob- lems on intrinsic data characteristics. Expert Systems with Applications 39 (7), 6585–6608. Lou, X., Cui, B., 2007. Robust asymptotic stability of uncertain fuzzy BAM neural networks with time-varying delays. Fuzzy Sets and Systems 158 (24), 2746– 2756. Maldonado, S., López, J., 2014. Imbalanced data classification using second-order cone programming support vector machines. Pattern Recognition 47 (5), 2070– 2079. 26 Maratea, A., Petrosino, A., Manzo, M., 2014. Adjusted F-measure and kernel scaling for imbalanced data learning. Information Sciences 257, 331–341. Menardi, G., Torelli, N., 2014. Training and assessing classification rules with imbalanced data. Data Mining and Knowledge Discovery 28 (1), 92–122. Mu, X., Artiklar, M., Watta, P., Hassoun, M., 2006. An RCE-based associative memory with application to human face recognition. Neural Processing Letters 23 (3), 257–271. Napierala, K., Stefanowski, J., Wilk, S., 2010. Learning from imbalanced data in presence of noisy and borderline examples. In: Szczuka, M., Kryszkiewicz, M., Ramanna, S., Jensen, R., Hu, Q. (Eds.), Rough Sets and Current Trends in Computing. Vol. 6086 of Lecture Notes in Computer Science. Springer, pp. 158–167. Nazari, S., Eftekhari-Moghadam, A.-M., Moin, M.-S., 2014. A novel image stenography scheme based on morphological associative memory and permuta- tion schema. Security and Communications Networks 8 (2), 110–121. Palm, G., 2013. Neural associative memories and sparse coding. Neural Networks 37, 165–171. Park, Y., Ghosh, J., 2014. Ensembles of α-trees for imbalanced classification prob- lems. IEEE Trans. on Knowledge and Data Engineering 26 (1), 131–143. Prati, R. C., Batista, G. E., Monard, M. C., 2004. Learning with class skews and small disjuncts. In: Bazzan, A. L., Labidi, S. (Eds.), Advances in Artificial Intelligence. Vol. 3171 of Lecture Notes in Computer Science. Springer, pp. 296–306. Ritter, G., Sussner, P., Diaz-de Leon, J. L., 1998. Morphological associative mem- ories. IEEE Trans. on Neural Networks 9 (2), 281–293. Santiago, R., 2003. Clasificador Hı́brido de Patrones basado en la Lernmatrix de Steinbuch y el Linear Associator de Anderson-Kohonen. Master’s thesis, In- stituto Politécnico Nacional, Centro de Investigación en Computación, Mexico D.F. (in Spanish). Seiffert, C., Khoshgoftaar, T. M., Van Hulse, J., Folleco, A., 2014. An empirical study of the classification performance of learners on imbalanced and noisy software quality data. Information Sciences 259, 571–595. 27 Sharma, N., Ray, A. K., Sharma, S., Shukla, K. K., Pradhan, S., Aggarwal, L. M., 2008. Segmentation and classification of medical images using texture- primitive features: Application of BAM-type artificial neural network. Journal of Medical Physics 33 (3), 119–3126. Stefanowski, J., 2013. Overlapping, rare examples and class decomposition in learning classifiers from imbalanced data. In: Ramanna, S., Jain, L. C., Howlett, R. J. (Eds.), Emerging Paradigms in Machine Learning. Vol. 13 of Smart Inno- vation, Systems and Technologies. Springer, pp. 277–306. Steinbuch, K., 1961. Die lernmatrix. Kybernetik 1 (1), 36–45. Uriarte-Arcia, A. V., López-Yáñez, I., Yáñez Márquez, C., 2014. One-hot vector hybrid associative classifier for medical data classification. PLoS ONE 9 (4), e95715. Štanclová, J., Zavoral, F., 2005. Hierarchical associative memories: The neural network for prediction in spatial maps. In: Roli, F., Vitulano, S. (Eds.), Im- age Analysis and Processing. Vol. 3617 of Lecture Notes in Computer Science. Springer, pp. 786–793. Wilson, D. L., 1972. Asymptotic properties of nearest neighbour rules using edited data. IEEE Trans. on Systems, Man and Cybernetics 2, 408–421. Yang, J., Yu, X., Xie, Z.-Q., Zhang, J.-P., 2011. A novel virtual sample generation method based on gaussian distribution. Knowledge-Based Systems 24 (6), 740– 748. Yu, H., Mu, C., Sun, C., Yang, W., Yang, X., Zuo, X., 2015. Support vector machine-based optimized decision threshold adjustment strategy for classifying imbalanced data. Knowledge-Based Systems 76, 67–78. Yu, H., Ni, J., Zhao, J., 2013. ACOSampling: An ant colony optimization-based undersampling method for classifying imbalanced DNA microarray data. Neu- rocomputing 101, 309–318. 28 
work_grruuptu7zbyrfv7w7nmxmz3fi	----	UbiPaPaGo: Context-aware path planning Expert Systems with Applications 38 (2011) 4150–4161 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a UbiPaPaGo: Context-aware path planning Chiung-Ying Wang a, Ren-Hung Hwang b,⇑, Chuan-Kang Ting b a Department of Information Management, TransWorld University, Taiwan, ROC b Department of Computer Science and Information Engineering, National Chung-Cheng University, Taiwan, ROC a r t i c l e i n f o Keywords: Ubiquitous network Human-centered computing Context-awareness Path planning Genetic algorithm (GA) Spatial conceptual map 0957-4174/$ - see front matter � 2010 Elsevier Ltd. A doi:10.1016/j.eswa.2010.09.077 ⇑ Corresponding author. Address: 168, University Taiwan, ROC. Tel.: +886 5 2720411x33112; fax: +886 E-mail addresses: wjy@cs.ccu.edu.tw (C.-Y. Wang) H. Hwang), ckting@cs.ccu.edu.tw (C.-K. Ting). a b s t r a c t The increased prevalence of digital devices with communication capability heralds the era of ubiquitous computing, as predicted by Mark Weiser. Ubiquitous computing aims to provide users with intelligent human-centric context-aware services at anytime anywhere. Optimal path planning in a ubiquitous net- work considers the needs of users and the surrounding context. This approach is very different from that applied by existing research on car navigation and mobile robots. This study proposes a context-aware path planning mechanism based on spatial conceptual map (SCM) and genetic algorithm (GA), referred to as UbiPaPaGo. The SCM model is adopted to represent the real map of the surrounding environment. The optimal path is planned using a GA, which is a robust metaheuristic algorithm. UbiPaPaGo attempts to automatically find the best path that satisfies the requirements of an individual user. A prototype of UbiPaPaGo is implemented to demonstrate its feasibility and scalability. Experimental results validate the effectiveness and the efficiency of UbiPaPaGo in finding the optimal path. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction consider the path distance, accessibility of WiFi Access Point (AP) The demand for intelligent navigation applications in path find- ing, such as PaPaGo (PAPAGO) and TomTom (TomTom Interna- tional), has risen significantly in recent years, helped by the rapid development of wireless and communication technologies. These systems provide mobile users, such as car drivers, with intelligent path planning, routing and navigation services, and direct users to their destination using electronic maps and Global Position System (GPS). However, these systems are computer-centric, and thus are unable to achieve the vision of Mark Weiser (Weiser, 1991) of a ubiquitous environment with human-centric applications. Con- text-awareness is a key to human-centric applications. A context- aware application extracts, interprets and uses context, and automatically adapts its functionality to the current context of an individual user. Consequently, mobile users can adopt context to plan the ‘‘best-fitting” path automatically. To help address the problem and develop the solution, a sce- nario is described for finding John’s optimal path that satisfies his requirements. John needs to drive to the Taipei Arena as quickly as possible to watch a show, because he is late. The automobile assistant connects to a streaming server to watch the live show re- motely before John arrives at Taipei Arena. In this case, John needs to find the shortest path to the Taipei Arena, while keeping connec- tivity with the streaming server. Therefore, the optimal path must ll rights reserved. Rd., Min-Hsiung Chia-Yi 621, 5 2720859. , rhhwang@cs.ccu.edu.tw (R.- or 3G Base Station (BS), accessibility of the streaming service and quality of service (QoS) of the required service. The problem con- sidered herein is formally defined as how to plan an optimal (shortest) path for an individual user based on his requirements and the context of his surrounding environment, such as user pref- erence, location, network connectivity, bandwidth, service avail- ability and quality of service (QoS). The objective is to find an optimal (shortest) path given constraints, such as higher band- width, available service and QoS guarantee. These constraints en- sure that the shortest path is also the one that best fits the user requirements. The context is very important for inferring the optimal path. The path planning framework relies on various contexts acquired from various sensing sources. The collected contexts are stored in an open and distributed context management server (Lu, Wang, & Hwang, 2009), called Ubiquitous Gate (U-gate), which is responsi- ble for context acquisition, context representation, context retrie- val, context protection and context inference. The path planning framework adopts an ontology as the underlying context model. The ontology model is implemented using RDF (Resource Descrip- tion Framework) ([RDF Primer] and [RDF Schema]) and OWL (Web Ontology Language) (OWL). The context model includes user con- texts, such as personal information and active applications, as well as environmental contexts, such as map information and network connectivity. The proposed solution, called UbiPaPaGo, is based on spatial conceptual map (SCM) (Kettani & Moulin, 1999) and genetic algo- rithm (GA) (Holland, 1975). A modified SCM is proposed to model the real world map and related environment context, and a path http://dx.doi.org/10.1016/j.eswa.2010.09.077 mailto:wjy@cs.ccu.edu.tw mailto:rhhwang@cs.ccu.edu.tw mailto:ckting@cs.ccu.edu.tw http://dx.doi.org/10.1016/j.eswa.2010.09.077 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 4151 planning map based on this model is then created. In addition, a GA is proposed to find the optimal path. Chromosomes are encoded based on the method proposed in Inagaki, Haseyama, and Kitajima (1999). The fitness function is carefully designed to fit the user requirements. Prototyping and simulations are performed to dem- onstrate the feasibility and the efficiency of the proposed UbiPaPaGo. The remainder of this paper is structured as follows. Section 2 describes related works on path planning and context-aware path prediction. Section 3 introduces the SCM model. Section 4 de- scribes the design of UbiPaPaGo based on GA. Section 5 provides implementation details of the prototyping system. Experimental results are discussed in Section 6. Conclusions are finally drawn in Section 7, along with our future research. 2. Related work Path planning problems have been studied extensively in the past, but mostly focusing on mobile robots (Bodhale, Afzulpurkar, & Thanh, 2009; Castilho & Trujilo, 2005; Lei, Wang, & Wu, 2006; Li, Zhang, Yin, Zhang, & Liu, 2006; Liu, Lu, Yang, & Chen, 2008; Tu & Yang, 2003), transportation systems (Li, Liu, et al., 2006) and game systems (Arikan, Chenney, & Forsyth, 2001; Wan, Chen, & Earnshaw, 2003). These proposed path planning algorithms emphasize features, such as collision avoidance and ease of track- ing. Although these solutions address information about roads, obstacles and maps, they do not adapt to user situations or context. Context-aware path finding (moving path) has recently been studied. Studies which focused on context-aware path prediction are most related to our work. These approaches adopt context, such as history data, user’s preference, location and map data, to reason the future moving path in advance. Path prediction ap- proaches are able to yield high prediction accuracy and efficient network resource management and pre-configuration, thus pro- viding high quality of service for users. Neural-network based pre- diction algorithm (Poon & Chan, 2000) predicts path from both global and user-level information. Profile-based next-cell predic- tion (PBNP) algorithm (Bharghavan & Jayanth, 1997), shadow clus- ter concept (Levine, Akyildiz, & Naghshineh, 1997), mobile motion prediction (MMP) algorithm (Liu & Maguire, 1996), per-user profile replication scheme (Jannink, Lam, Shivakumar, Widom, & Cox, 1997), hierarchic position-prediction (HPP) algorithm (Liu, Bahl, & Chlamtac, 1998), regular path recognition method (Erbas, Steuer, Kyamakya, Eggesieker, & Jobmann, 2001) and sectorized mobility prediction algorithm (Chellappa Doss, Jennings, & Shenoy, 2004) predict the further location from a user’s moving history. In addi- tion, HPP and mobility tracking algorithms (Zaidi & Mark, 2005) consider instantaneous RSSI measurements of surrounding cells. PBNP also considers location classification. Path prediction (Anag- nostopoulos, Anagnostopoulos, Hadjiefthymiades, Kalousis, & Kyriakakos, 2007) considers both spatial information, such as mov- ing profile and moving history, and temporal information, such as morning or afternoon. Moving history is one of the important con- texts. However, if the moving history is not available, then those approaches are not applicable. Hence, the mobility prediction architecture (Samaan & Karmouch, 2005) performs path prediction based on user context, such as user preferences, goals and spatial conceptual maps (SCM). Additionally, most approaches use high- level contexts in the prediction process. However, Sigg, Haseloff, and David (2006) use low-level context that contain additional information not available with high-level context to improve pre- diction accuracy. The prediction results of these approaches are useful for many context-aware applications. For instance, if the moving path is obtained, then the handoff process can be prepared such that seamless handoff can be better achieved. However, the high accuracy of preferable or habitual moving path may not be the optimal (best-fitting) path in terms of the user’s objective and requirement. Consequently, planning an optimal path (short- est path) based on various contexts that best fit the user require- ments is an essential issue. Genetic algorithm (GA) (Holland, 1975) is a search method for solving large-scale optimization problems. GA is inspired by bio- logical evolution and has proved to be a robust and powerful adap- tive search technique. Many researches use GA to solve path finding (Ji, Chen, & Kitti, 2004), path planning (Kanoh, 2007; Castillo & Trujillo, 2006; Kanoh & Hara, 2008) and route selection (Chakraborty, Maeda, & Chakraborty, 2005) problems. Numerical experiments indicate that GA can result in effective searching, shortest path, improved QoS, low system resource consumption and quick decision time. Therefore, the proposed UbiPaPaGo mech- anism adopts a GA-based approach to obtain the optimal path. 3. Design of UbiPaPaGo This section introduces the design goal of UbiPaPaGo. Section 3.2 presents the context model of UbiPaPaGo. Section 3.3 describes the architecture of UbiPaPaGo. Finally, the privacy policy of UbiPaPaGo is introduced in Section 3.4. 3.1. Design goal This study develops a human-centric, context-aware and intel- ligent path planning mechanism to satisfy individual user require- ments in a ubiquitous network. Accordingly, the following objectives are proposed: (1) The mechanism should be scalable and possible to implement in a distributed manner. (2) Based on various contexts acquired from the environment, the solution should be able to automatically plan an optimal path that best fits the user requirements, such as network connectivity, bandwidth and available services. UbiPaPaGo relies on various types of context stored in an open and distribute framework (Lu et al., 2009). This framework allows a digital device to join the ubiquitous environment and search for context-aware services based on standard protocols, such as Uni- versal Plug and Play (UPnP) (Universal Plug and Play), Hypertext Transfer Protocol (HTTP) (Fielding et al., 1999) and DNS. A service or application, such as UbiPaPaGo, can be easily deployed into the system through standard protocols and distributed context managers. 3.2. Context model Table 1 presents the context model of UbiPaPaGo. In a ubiqui- tous network, billions of sensors located in our physical environ- ment accumulate a huge amount of context. These contexts are fed into context model for context-aware path planning applica- tions. This study limits the description of context needed to sup- port path planning mechanism. The context model includes static and dynamic types describing the user, terminal and map. The user profile contains details of the user. The user preference allows a user to specify the required service and cost constraints. Meanwhile, the proposed system also considers user privacy. The interface privacy setting enables users to specify their privacy pol- icies, since user privacy is always one of the most important issues in developing context-aware applications. Users can hide sensitive information, such as schedule and location, or filter access to per- sonal context, using their privacy settings. The terminal profile includes ID, capability, software interface status and types of running application. The terminal capability describes the process speed, memory, screen size, resolution and Table 1 Context model of UbiPaPaGo. Static Dynamic User – User identity – User preference � Required service � Cost constrain � Habit – Privacy setting – Schedule – Location – Moving direction – User preferences Terminal – Terminal ID – Capability � Process speed � Memory � Screen size � Resolution � Supported interface types – Application type � VoIP, video conference � IPTV, mobile TV � Email, ftp � IM, web browsing – Interface status – Running application type Map – Block ID – Block distance – Accessible service � Conversational � Real-time streaming � Non-real-time streaming � Interactive � Background – QoS parameter of accessible service � RSSI � Required bandwidth � Delay � Jitter � Packet loss ratio – Accessible context server – Accessible AP/BS – QoS parameter of accessible AP/ BS � RSSI � Required bandwidth � Delay � Jitter � Packet loss ratio Server WLAN Service Providing Server Context Management Server Middleware (Web Service) C on text A cqu isition 3G GPS RFID Zigbee Hand-held Devices Client Application : UbiPaPaGo Sensor LAN WiMax Network Fig. 1. Architecture of UbiPaPaGo. 4152 C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 supported interface types. The terminal application types describe software applications installed in the terminal. Application types based on quality of service (QoS) parameters, such as response time, delay, jitters and bandwidth, can be categorized into four types, namely (i) conversational service, such as VoIP and video conferencing; (ii) real-time (RT) service, such as Internet Protocol Television (IPTV) and mobile TV; (iii) non-real-time (NRT) services, such as email and ftp; (iv) interactive services, such as instant mes- sage (IM) and web browsing. The map and environment profile includes block ID, block dis- tance, accessibility of AP/BS, QoS parameter of accessible AP/BS, accessible service, QoS parameter of accessibility service, and accessible context server (U-gate). These contexts are represented using an SCM model, which is described in detail in Section 4. The QoS parameters of a service are pre-defined based on service pro- file and stored in the context server. The QoS of AP/BS is dynami- cally updated to the context management server. 3.3. Architecture of UbiPaPaGo Fig. 1 shows the UbiPaPaGo architecture. The UbiPaPaGo servers are composed of a service providing server and a context manage- ment server. The service providing server provides the intelligent human-centric services. The context management server mainly acts as a semantic database server to manage context accumulated from the sensors in the environment. The context management server manages and refers to context by constructing an ontology tree based on OWL and RDF. The hand-held device includes a Ubi- PaPaGo client application to find the best-fitting path. The commu- nication model (Lu et al., 2009) among the device and the server is mainly based on the standard protocols, including Universal Plug and Play (UPnP), Simple Service Discovery Protocol (SSDP) (Simple Service Discovery Protocol), Simple Object Access Protocol (SOAP) (Simple Object Access Protocol), General Event Notification Proto- col (GENA) (General Event Notification Architecture) and Hyper- text Transfer Protocol (HTTP). 3.4. Privacy policy UbiPaPaGo relies on various contexts which are highly related to personal privacy. Protecting user privacy is always one of the most important issues in developing context-aware applications. Unfortunately, due to large amount of context information and de- vices in dynamic and heterogeneous environments, creating a total solution for managing security and privacy policies is known to be a very challenging open issue in ubiquitous environments. In cur- rent development, UbiPaPaGo tries to preserve privacy by using a privacy setting to hide sensitive information, such as schedule, location, moving direction and user preference in context manage- ment server, and filtering access to personal context based on user’s preference and privacy setting. Future versions of UbiPaPaGo will deploy advanced mechanisms or policies for preserving pri- vacy on the context management server. 4. UbiPaPaGo based on SCM model This section represents the trajectory map and environment context using a SCM model. The SCM is an abstraction of real map representation containing of a set of landmark objects (Oj), a set of medium objects (Wy) and the influence areas of those ob- jects. The landmark objects define those areas, such as buildings or user’s target destination. The medium objects (also called ‘‘ways”) define areas, such as streets, roads, trajectories and virtual connections between objects (Kettani & Moulin, 1999). A way ob- ject is further partitioned into continuous blocks, called Way Ele- mentary Areas (WEAs). To encode additional context information, a novel approach for partitioning a way object is proposed as follows. A WEA may be a landmark object (Oj) or a crossing point of two or more ways (Ket- tani & Moulin, 1999). The proposed solution requires additional contexts, including received signal strength indication (RSSI) of Fig. 2. A portion of Taipei City of Taiwan map. C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 4153 AP/BS and available bandwidth. Thus, besides partitioning a way object based on crossings, the proposed solution also considers characteristics of AP/BS, such as RSSI. Specifically, a way object is partitioned into several continuous blocks (WEAs) such that the RSSI of an AP/BS within each WEA is roughly at the same level. Fig. 2(a) illustrates a portion of the map of Taipei city in Taiwan, and the distribution of AP/BS (Wifly). To clearly show the distribu- tion and the context of AP/BS, the blue1 circle in Fig. 2(a) shows an enlarged portion of the map. In this study, a series of simulations were performed using the Qualnet network simulator (Qualnet) to implement the context of the AP/BS, as shown in Fig. 2(b). Fig. 3 shows its SCM representation. 1 For interpretation of color in Figs. 2, 3, 10–14, the reader is referred to the web version of this article. The Matrix of Orientation and Adjacency and Characteristic (MOAC) (Samaan & Karmouch, 2005) was adopted to represent the WEAs of the SCM. The MOAC contains characteristic informa- tion of each WEA and landmark object, and the relative informa- tion and displacement direction between two WEAs. Fig. 4 shows a portion of the MOAC of Fig. 3. The diagonal cell (x, x) represents the characteristics, i.e. context of WEA ax. The cell (x, y), where x – y denotes the characteristics of relative direction between ax and ay, which can be represented in eight-way directions, namely, east (E), west (W), south (S), north (N), northeast (NE), southeast (SE), southwest (SW) and northwest (NW). For instance, a19 is on the south (S) side of a18, and a3 is on the northwest side (NW) of a18. According to MOAC, the list of blocks can be inferred along a moving path, along with the direction of movement and all re- quired context information. Fig. 3. SCM representation of Taipei City of Taiwan map of Fig. 2. Fig. 4. A portion of MOAC of Fig. 3. 4154 C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 Fig. 6. Path planning map. C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 4155 The characteristic function Co( ) describes several characteris- tics of an object. Specifically, Co(Oj) and Co(ai) describe the context information of a landmark object and a WEA, respectively. More- over, since state information of APs is also considered in this study, each AP is associated with a characteristic function Co(APi). For in- stance, if O1 is a rapid metro station, then the characteristic func- tion denotes the characteristics in terms of name, location. Thus an example of Co(O1) could be {c1 = rapid transit station, c2 = a33}. Similarly, if WEA a132 is covered by two APs and provides a service, then the characteristic function of Co(a132) is denoted by {c1 = {AP1, AP3}, c2 = {streaming server}}. The characteristics of AP are described in terms of available RSSI (c1), bandwidth (c2), delay (c3), jitter (c4), and packet loss ratio (c5). An example of Co(AP1) is {c1 = �70 dbm, c2 = 11 Mbps, c3 = 30 ms, c4 = 10 ms, c5 = 5%}. 5. UbiPaPaGo based on GA This section describes our context-aware path planning mecha- nism based on GA. The GA procedure is shown in Fig. 5. Before applying GA, the SCM model obtained in Section 4 is first converted to path planning map, as shown in Fig. 6, where node ai denotes a crossing in Fig. 3. Restated, to reduce the length of chromosome and to increase the scalability of UbiPaPaGo, the path planning map considers only crossing nodes. Additionally, the path planning map needs to consider the source or destination node, if these are not at a crossing. However, the fitness function of a chromosome is still calculated according to all WEAs and their characteristic func- tions, as described later. The path planning map is an undirected graph G = (V, E), where V denotes a set of crossing, source and destination nodes, and E de- notes a set of edges (WEAs). The aim is to identify the shortest path for G. Each edge e e E includes blocks(ai). Corresponding to each edge, a weight xij denotes the fitness value calculated from the fit- ness function from node ai to node aj. Moreover, if the user’s re- quired service is assumed to be located at the edge, then this edge is called a constraint edge e0 e E. The found path ~S must con- tain this edge, so that the required service can be accessed. For in- stance, if the streaming services are located at block a60 and a46, respectively, then the links (28, 87) and (23, 81) are constraint edges that must be included in the path. Note that end-to-end Initialization Population Evaluate Fitness Termination Criterion Met? Solution Start End Selection/ Reproduction Crossover Mutation No Yes Fig. 5. The procedure of GA. QoS attributes are evaluated, e.g., the delay represents the end- to-end delay of a path. The UbiPaPaGo defines the optimal path as the shortest path subject to seven constraints, including received signal strength indication (RSSI), available bandwidth (BW), response time (RT), delivery latency (DL), jitter (J) and packet loss ratio (PL), available service (AS). The optimization problem is formulated as follows: minimize Dð~SÞ¼ X k2~S distðekÞ ð1Þ subject to 8 linkðekÞ 2~S : xðekÞ P hx; X ¼fRSSI; BWg; x 2 X ð2ÞX k2S yðekÞ6 hy; Y ¼fRT; DL; J; PLg; y 2 Y ð3Þ AS~S exists required service ð4Þ where Dð~SÞ denotes the total distance of moving path, dist(ek) de- notes the distance of edge ek, hx and hy are thresholds for RSSI, RT, etc. The constraints (2)–(4) ensure that the computed result indeed satisfies the user’s requirements. 5.1. Chromosome representation and population initialization Chromosome representation is a key to the design of GA for solving a problem. This study adopts the encoding method pro- posed in Wan et al. (2003), where all chromosomes have the same length, thus facilitating crossover operation. Fig. 7 gives an exam- ple of chromosome representation, where the upper sequence is the node id and the lower sequence is the chromosome. Here a chromosome is represented by a sequence of integer numbers; a gene denotes the node id through which it passes. For an edge (vi, vj) from node vi to node vj, node vj is put in the ith id of the chro- mosome. If the id is not along the path, then its corresponding gene is filled with a neighboring node randomly chosen from the set of its neighboring nodes. For instance, the chromosome in Fig. 7 de- notes a path from source node id 28 (block a28) to destination node id 113 (block a113), passing through a87, a83, a120 and a118. The initial population influences the convergence speed and the quality of the solution. Random and heuristic initializations are the most commonly used ways of generating candidate solutions. However, random initialization possibly causes a large computing overhead and loop path, while heuristic initialization (Arikan et al., 2001) has difficulty in finding a global optimal. Heuristic initializa- tion explores a small part of solution space, making it faster than random initialization. This study adopts the probability network approach (Poon & Chan, 2000) that combines the features of ran- dom and heuristic initialization to generate initial candidate paths in the gene pool. Starting with the source node, the next node is chosen from its directly connected neighbors. To speed up the con- vergence speed, each neighboring node is assigned a different selection probability. A neighboring node located in the direction id 18 20 23 25 28 31 76 81 83 87 90 113 118 120 124 127 137 143 gene 20 23 25 83 87 28 113 83 120 83 127 118 113 118 120 124 158 137 chromosome )2()1( (3) (4) (5) Fig. 7. Example of encoding path from a28 to a113. 4156 C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 toward the destination is assigned a higher probability. The se- lected next node then becomes the gene of the source node’s id. This procedure repeats until the path reaches the destination node. To avoid looping, a node is forbidden to visit twice in the same path during the next node selection procedure. Considering the example shown in Fig. 7, a87 has four neighboring nodes, a28, a83, a90 and a124 (cf. Fig. 6), among which a83 and a124 have higher probabilities to be chosen as the next node of a87, since they are located in the direction toward a113 and have not appeared in the path before. For nodes that are not in the path under consideration, their next nodes are randomly selected. For instance, a31, a87 and a127 can be chosen as the next node of a90. 5.2. Fitness function The fitness function measures the distance of the path that satis- fies all the quality constraints. The concept of penalty function is adopted to handle the constraints (Yeniay, 2005). In this study, the quality of a chromosome depends on properties of each link(ek) of the path, including received signal strength indication (RSSI), avail- able bandwidth (B), response time (RT), delivery latency (DL), jitter (J), packet loss (PL) and available service (AS). The fitness function is defined by F ¼ Dð~SÞ � ð1 þ aPÞ ð5Þ with P ¼ X k2~S pðekÞþ hASð~SÞ ð6Þ where pðekÞ¼ X x2fRSSI;BWg fxðekÞþ X y2fRT;DL;J;PLg gyðekÞ ð7Þ fxðekÞ¼ 1; if xðekÞ6 hx 0; otherwise � x 2fRSSI; BWg ð8Þ gyðekÞ¼ 1; if P k2S yðekÞ P hy 0; otherwise ( y 2fRT; DL; J; PLg ð9Þ id 18 20 23 25 28 31 76 81 83 parent 1 20 23 25 83 87 28 113 83 12 parent 2 76 18 81 23 25 90 113 76 87 mask 1 2 1 1 2 1 1 1 2 offspring 20 18 25 83 25 28 113 83 87 chro Fig. 8. An example of c hASð~SÞ¼ 1; if required service is not available along path 0; otherwise � ð10Þ 5.3. Selection The selection operation chooses a chromosome from the cur- rent population for reproduction. The UbiPaPaGo adopts tourna- ment selection (Goldberg & Deb, 1991) to select parents for producing offspring. In tournament selection, two individuals are chosen at random from the population, and the one with the higher fitness value is chosen as a parent for the crossover operation. The tournament process repeats until two parents are selected. 5.4. Crossover The crossover operation recombines two chromosomes (called parents) to produce a new chromosome (called offspring). This study proposes a modified uniform crossover by additionally con- sidering genes (nodes) that have not appeared (been visited) be- fore. The proposed crossover operation is carefully designed to avoid the path looping problem. The crossover starts with the gene at the id of the source node. If the two parental genes at this id are identical, then this gene is simply copied to the offspring. Other- wise, the gene that has not appeared before is chosen. A tie is bro- ken by random selection. After selecting the gene of the current id, the procedure moves to the id of the next node (i.e., the gene we just selected), and is repeated until the destination node is reached as the gene of the current id. Fig. 8 shows an example of the crossover operation. The path in parent 1 is a28 ? a87 ? a83 ? a120 ? a118 ? a113, and that of par- ent 2 is a28 ? a25 ? a23 ? a81 ? a76 ? a113. For the node id 87, the operation randomly selects among a83 and a90. If a90 is selected, then node id 90 is considered. The operation is then repeated until the destination node is reached. The new offspring a28 ? a25 ? a83 ? a87 ? a90 ? a127 ? a124 ? a120 ? a118 ? a113 is generated after the crossover operation. For the id’s in which the path does not pass through, e.g. 20, the genes at them are randomly selected from either parent, for example, a18 in node id 20. 87 90 113 118 120 124 127 137 143 0 83 127 118 113 118 120 124 158 137 90 87 76 81 137 120 90 143 127 2 1 1 1 1 1 1 1 2 90 127 118 113 118 120 124 158 127 mosome rossover operation. id 18 20 23 25 28 31 76 81 83 87 90 113 118 120 124 127 137 143 before 20 23 25 83 87 28 113 83 120 83 127 118 113 118 120 124 158 137 after 20 23 25 83 87 28 113 83 120 90 127 118 113 118 120 124 158 137 Fig. 9. An example of mutation operator. C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 4157 5.5. Mutation With a given probability, mutation introduces variation into the chromosome by adjusting partial chromosome to avoid local opti- ma. In this study, a random node (mutation node) that is neither a source nor a destination node along a path is chosen for mutation (Wan et al., 2003). The new gene value is randomly chosen from directly connected nodes of the mutation node. Fig. 9 shows an example of mutation operation. The old path is a28 ? a87 ? a83 ? a120 ? a118 ? a113. In the mutation operation, the node id a87 is randomly chosen as the mutation gene. The new path following the mutation operation is a28 ? a87 ? a90 ? a127 ? a124 ? a120 ? a118 ? a113. After crossover and mutation operations are completed, the off- springs are evaluated by the fitness function for another round of reproduction until a termination condition is satisfied. 6. Prototyping and implementation This section demonstrates the feasibility of the UbiPaPaGo by prototyping a real system in U-gate. Because U-gate is an open, effi- cient and distributed context management architecture and com- munication model based on standard protocols, UbiPaPaGo can be implemented in a distributed manner, enabling any service provider to deploy its service into the ubiquitous environment easily. For easy deployment, UbiPaPaGo was implemented in a small scope environ- ment. Therefore, the demonstration scenario is in our campus (Chung-Cheng University), as displayed in Fig. 10, and made up with user using EeePC with WiFi and RFID. In the scenario, the user uses UbiPaPaGo service in EeePC to connect UbiPaPaGo service provider (U-gate) in local area, as shown in Fig. 11(a) and (b), to plan a path from the college of engineering (O9) to the auditorium (O8) in the CCU campus. U-Gate collects the contexts of user’s current location, requirement, ID, schedule and SCM map to infer the path planning result, as shown in Fig. 11(c). Readers are referred to Lu et al. (2009) for the detailed implementation of U-gate. Fig. 10. The implementation 7. Experimental results and discussion This section evaluates the proposed UbiPaPaGo through com- puter simulation. The proposed UbiPaPaGo was implemented in a Java (JDK 6) program on a PC with a Core 2 Quad 2.4 GHZ CPU and 2G MB memory. Fig. 12 shows the road map of some part of Taipei city that was used in the experiments. In the experiment, the road map with 1481 nodes was first converted into a graph with 582 crossing nodes. The parameters are set as follows. The experiment tries five population sizes (N): 100, 200, 300, 400, and 500. Termination cri- terion is 500 generation. The crossover probability (Pc) was set to 1, and the mutation probability (Pm) was set to 0.05. The tournament size was set to 2. The alpha (a) value of the penalty function weight was empirically set as a = 2. The collected results were averaged over 100 runs. Notably, UbiPaPaGo uses probability network based initializa- tion and a modified crossover. To verify the effect of these two ele- ments, we compare the performance of UbiPaPaGo with that of GA, i.e. using random initialization and uniform crossover. The perfor- mance metric, mean best fitness (MBF), is defined by MBF ¼ sum of the best fitness value of each run number of runs ð11Þ 7.1. Optimal path evaluation Fig. 13(a)–(c) shows the best obtained path, the optimal path, and the shortest path for the indicated source and destination. The result of the proposed UbiPaPaGo in Fig. 13(a) coincides with the optimal path in Fig. 13(b) and the shortest path in Fig. 13(c). Table 2 summarizes the results of QoS constraint and parameters of the obtained path in Fig. 13(a)–(c). The QoS parameters of the resulting path means the minimum RSSI and bandwidth and max- imum delay, jitter and packet loss rate along the path. The best ob- tained path of UbiPaPaGo guarantees shorter path and QoS to scenario of UbiPaPaGo. Fig. 11. The demonstration of UbiPaPaGo. 4158 C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 satisfy user’s requirements, while the shortest path has shortest distance but does not satisfy user’s requirements. 7.2. Convergence behavior Fig. 14 compares the convergence behavior of the UbiPaPaGo and GA. The result is helpful for choosing the appropriate popula- tion size and termination condition. Moreover, it shows that a small population size reduces evaluation cost but results in prema- ture convergence. On the other hand, a large population size finds better solutions and yet requires longer computation time. We can also observe that UbiPaPaGo converges faster than GA. According to the result in Fig. 14, the best convergence curve of UbiPaPaGo was achieved with population size of 300 in 100 generations. Fig. 12. The portion of Taipei City. C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 4159 Therefore, unless otherwise stated, the following experiments of UbiPaPaGo were performed with population size 300 and 100 generations. The GA differs from UbiPaPaGo by adopting a random approach in the initialization and crossover operations, such that some chromosomes may contain loops. Hence, the UbiPaPaGo converges faster than the GA. The results indicate that the UbiPaPaGo improves the convergence speed by preventing looping in the initialization and crossover operations. Fig. 13. Path comparison (paths are Table 2 The QoS constraint and result of found path, optimal path and shortest path. Parameters Distance Fitness value Min RSSI Min b Range N/A N/A �40 to �100 2–54 Threshold N/A N/A P�60 P30 Best obtained path of UbiPaPaGo 124 124 �60 40 Optimal path 124 124 �60 40 Shortest path 123 327 �60 2a a The QoS parameter is not satisfied user’s requirement. 7.3. Optimization results comparison Table 3 compares the performance of the UbiPaPaGo and GA in terms of MBF, average computation time, probability of obtaining the best path. Note that the GA uses random initialization and uni- form crossover, while UbiPaPaGo uses probability network ap- proach initialization and modified uniform crossover. The MBF values indicate that UbiPaPaGo generally outperforms the GA in terms of fitness. The effectiveness of the UbiPaPaGo was measured under vari- ous numbers of generations. The UbiPaPaGo was found to have a shorter computation time and significantly higher success rate than the GA, because UbiPaPaGo is a probability network approach in which the next node is located forward of the destination with a higher probability. Additionally, UbiPaPaGo avoids looping. The probability of obtaining the optimal solution is higher than 93% with a population size of 300. This approach is also efficient, be- cause the population size is small and the evolution process is short. The result is very important and applicable for context- aware applications. Short computation time and high success prob- ability enable UbiPaPaGo to achieve satisfactory solution quality and real time service. Table 3 also shows that UbiPaPaGo has lower standard deviations than GA does. 7.4. Effect of alpha value Finally, Table 4 compares the fitness, distance, and penalty val- ues at different alpha values ranging from 0.5 to 2.0. Note that pop- ulation size 100 with 500 generations are adopted herein. The table shows that the increaseing alpha value causes decreasing penalty value, increasing average fitness, and increasing distance value. This is because a high alpha value intensifies penalty so that the evolutionary search tends to avoid penalty and to find a solution without penalty but with ordinary distance. Conversely, a low al- pha encourages finding a short route that may violate some constraints. denoted by blue dotted lines). andwidth Max delay Max jitter Max packet loss rate Available service 10–70 10–50 0.02–0.2 0.1 640 615 60.05 =1 10 10 0.03 1 10 10 0.03 1 70a 30a 0.2a 0a Fig. 14. Convergence curve of UbiPaPaGo and GA. Table 3 The QoS constrain and result of found path. Generation number Fitness Average computation time (s) Probability of obtaining the optimal path Best MBF Std UbiPaPaGo N = 1 124 132.07 6.26 2.67 0.05 N = 100 124 125.48 1.86 7.18 0.34 N = 200 124 124.74 1.8 9.49 0.73 N = 300 124 124.34 1.34 11.64 0.93 N = 400 124 124.13 0.48 13.73 0.95 N = 500 124 124.1 0.46 15.96 0.95 GA N = 1 124 135.73 5.8 15.47 0.01 N = 100 124 131.1 5.27 20.84 0.06 N = 200 124 129.9 4.14 23.24 0.25 N = 300 124 129.66 4.42 25.85 0.43 N = 400 124 129.2 2.9 28.01 0.43 N = 500 124 129.2 3.12 30.47 0.44 Table 4 Fitness, distance and penalty for alpha values 0.5, 1.0, 1.5, and 2.0. Alpha (a) Parameters Average fitness (F) Average distance (Dð~SÞ) Average penalty (P) 0.5 129.4 126.83 5.14 1 131.06 127.59 3.47 1.5 131.17 127.79 2.25 2.0 132.56 128.57 2.00 4160 C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 8. Conclusions and future work This study proposes an intelligent navigation application, Ubi- PaPaGo, which focuses on providing human-centric context-aware path planning mechanism in a ubiquitous network. Specifically, UbiPaPaGo considers requirements and contexts of users and envi- ronment when planning the best-fitting path. UbiPaPaGo adopts the SCM model to present map context and uses GA to find the optimal path. The map simplifies conversion to graph, thus reduc- ing the computational complexity and increasing scalability. Experimental results examine the effect of initialization and pro- posed crossover. Moreover, the results validate that the proposed UbiPaPaGo can efficiently achieve the shortest path that guaran- tees all QoS parameters. Some issues of the proposed UbiPaPaGo remain for further re- search. This study assumes that the destination is inferred from the user’s personal context, such as schedule. If the destination of the path is uncertain or unknown, predicting the destination from the related context becomes an a priori problem. Additionally, the proposed system only considers how to plan an optimal path for a user but does not consider issues, such as vertical handoff between AP and BS, and resource reservation along the path in advance to guarantee quality of service of the path. Finally, security and pri- vacy issues also need further investigation. Advanced framework or policies for preserving privacy will be deployed in the future. Acknowledgment The research was supported by the NSC97-2221-E-194-011- MY3 and NSC97-2221-E-194-012-MY3, National Science Council, ROC. References Anagnostopoulos, T., Anagnostopoulos, C., Hadjiefthymiades, S., Kalousis, A., & Kyriakakos, M. (2007). Path prediction through data mining. In IEEE international conference on pervasive services (ICPS), Istanbul, Turkey (pp. 128– 135). Arikan, O., Chenney, S., & Forsyth, D. A. (2001). Efficient multi-agent path planning. In Eurographic workshop on computer animation and simulation, Manchester, UK (pp. 151–162). Bharghavan, V., & Jayanth, M. (1997). Profile-based next-cell prediction in indoor wireless LAN. In IEEE Singapore international conference. Bodhale, D., Afzulpurkar, N., & Thanh, N. T. (2009). Path planning for a mobile robot in a dynamic environment. In IEEE international conference on robotics and biomimetics (pp. 2115–2120). Castilho, O., & Trujilo, L. (2005). Multiple objective optimization genetic algorithms for path planning in autonomous mobile robots. International Journal of Computers, Systems and Signals, 6(1), 48–63. Castillo, O., & Trujillo, L. (2006). Patricia Melin, multiple objective genetic algorithms for path-planning optimization in autonomous mobile robots. Soft Computing – A Fusion of Foundations, Methodologies and Applications, 11(3), 269–279. Chakraborty, B., Maeda, T., & Chakraborty, G. (2005). Multiobjective route selection for car navigation system using genetic algorithm. In IEEE mid-summer workshop on soft computing in industrial application (SMXia/05) (pp. 190–195). Chellappa Doss, R., Jennings, A., & Shenoy, N. (2004). Mobility prediction for seamless mobility in wireless networks. In 4th international network conference (INC) (pp. 435–443). Erbas, F., Steuer, J., Kyamakya, K., Eggesieker, D., & Jobmann, K. (2001). A regular path recognition method and prediction of user movements in wireless networks. In IEEE vehicular technology conference, Atlantic City (pp. 2672–2676). Fielding, R., Gettys, J., Mogul, J., Frystyk, H., Masinter, L., Leach, P., et al. (1999). Hypertext Transfer Protocol – HTTP/1.1, IETF RFC 2616, June 1999. General Event Notification Architecture (GENA). <http://www.tools.ietf.org/html/ draft-cohen-gena-p-base-01>. Goldberg, D. E., & Deb, K. (1991). A comparative analysis of selection schemes used in genetic algorithms. Foundation of genetic algorithms (pp. 69–93). San Mateo Calif: Morgan Kaufman. Holland, J. (1975). Adaptation in natural and artificial systems. The University of Michigan Press. Inagaki, J., Haseyama, M., & Kitajima, H. (1999). A genetic algorithm for determining multiple routes and its applications. In IEEE international symposium (ISCAS’99) (pp. 137–140). Jannink, J., Lam, D., Shivakumar, N., Widom, J., & Cox, D. (1997). Efficient and flexible location management techniques for wireless communication system. ACM/ Baltzer Wireless Networks, 3(5), 361–374. http://www.tools.ietf.org/html/draft-cohen-gena-p-base-01 http://www.tools.ietf.org/html/draft-cohen-gena-p-base-01 C.-Y. Wang et al. / Expert Systems with Applications 38 (2011) 4150–4161 4161 Ji, Z. W., Chen, A., & Kitti, S. (2004). Finding multi-objective paths in stochastic networks: A simulation-based genetic algorithm approach. In Congress on evolution computation (CEC 2004) (pp. 174–180). Los Alamitos: IEEE Press. Kanoh, H. (2007). Dynamic route planning for car navigation systems using virus genetic algorithms. International Journal of Knowledge-Based and Intelligent Engineering Systems, 11(1), 65–78. Kanoh, H., & Hara, K. (2008). Hybrid genetic algorithm for dynamic multi-objective route planning with predicted traffic in a real-world road network. In Genetic and evolutionary computation conference (GECCO 2008) (pp. 657–664). Kettani, D., & Moulin, B. (1999). A spatial model based on the notions of spatial conceptual map and of object’s influence areas. In Conference spatial information theory (COSIT) (pp. 401–416). Lei, L., Wang, H., & Wu, Q. (2006). Improved genetic algorithms based path planning of mobile robot under dynamic unknown environment. In IEEE international conference on mechatronics and automation (pp. 1728–1732). Levine, D., Akyildiz, I., & Naghshineh, M. (1997). A resource estimation and call admission algorithm for wireless multimedia networks using the shadow cluster concept. IEEE/ACM Transactions on Networking, 5(1), 1–12. Li, Q., Liu, G., Zhang, W., Zhao, C., Yin, Y., & Wang, Z. (2006). A specific genetic algorithm for optimum path planning in intelligent transportation system. In 6th international conference on ITS telecommunications (pp. 140–143). Li, Q., Zhang, W., Yin, Y., Zhang, W., & Liu, G. (2006). An improved genetic algorithm of optimum path planning for mobile robots. In The sixth international conference on intelligent systems design and applications (pp. 637–642). Liu, T., Bahl, P., & Chlamtac, I. (1998). Mobility modeling, location tracking, and trajectory prediction in wireless ATM networks. IEEE JAC, 16(6), 922–936. Liu, Z., Lu, Q., Yang, P., & Chen, L. (2008). Path planning for tractor-trailer mobile robot system based on equivalent size. In 7th world congress on intelligent control and automation (WCICA) (pp. 5744–5749). Liu, G., & Maguire, G. (1996). A class of mobile motion prediction algorithms for wireless mobile computing and communications. ACM/Baltzer MONET, 1(2), 113–121. Lu, J. H., Wang, C. Y., & Hwang, R. H. (2009). An open framework for distributed context management in ubiquitous environment. In International symposium on UbiCom frontiers – Innovative research, systems and technologies (Ufirst-09), Brisbane, Australia (pp. 88–93). OWL. W3C document. <http://www.w3.org/2004/OWL/>. PAPAGO. <http://www.papago.com.tw/>. Poon, W. T., & Chan, E. (2000). Traffic management in wireless ATM network using a hierarchical neural-network based predication algorithm. In 15th international conference on computers and their applications, New Orleans, United States (pp. 5–8). Qualnet, Scalable Network Technologies Inc. <http://www.scalable-networks.com>. RDF Primer. W3C document. <http://www.w3.org/TR/rdf-primer/>. RDF Schema. W3C document. <http://www.w3.org/TR/rdf-schema/>. Samaan, N., & Karmouch, A. (2005). A mobility prediction architecture based on contextual knowledge and spatial conceptual maps. IEEE Transactions on Mobile Computing, 4(6), 537–551. Sigg, S., Haseloff, S., & David, K. (2006). A novel approach to context prediction in ubicomp environments. In 17th Annual IEEE international symposium on personal, indoor and mobile radio communications (PIMRC’06) (pp. 1–5). Simple Object Access Protocol (SOAP) version 1.2. <http://www.w3.org/TR/soap/>. Simple Service Discovery Protocol (SSDP). <ftp://ftp.pwg.org/pub/pwg/ipp/ new_SSDP/draft-cai-ssdp-v1-03.txt>. TomTom International. <http://www.tomtom.com/>. Tu, J., & Yang, S. (2003). Genetic algorithm based path planning for a mobile robot. In IEEE international conference on robotics and automation, Taipei (pp. 1221– 1226). Universal Plug and Play (UPnP). <http://www.upnp.org/>. Wan, T. R., Chen, H., & Earnshaw, R. (2003). Real-time path planning for navigation in unknown environment. In Theory and practice of computer graphics (pp. 138– 145). Weiser, M. (1991). The computer for the twenty-first century. Scientific American, 94–104. Wifly. <http://www.wifly.com.tw/Wifly6/tw/WiflyNet/ServiceRange/HotPoint>. Yeniay, Ö. (2005). Penalty function methods for constrained optimization with genetic algorithms. Journal of Mathematical and Computation Application, 10(1), 45–56. Zaidi, Z. R., & Mark, B. L. (2005). Real-time mobility tracking algorithms for cellular networks based on kalman filtering. IEEE Transactions on Mobile Computing, 4(2), 195–208. http://www.w3.org/2004/OWL/ http://www.papago.com.tw/ http://www.scalable-networks.com http://www.w3.org/TR/rdf-primer/ http://www.w3.org/TR/rdf-schema/ http://www.w3.org/TR/soap/ http://www.pwg.org/pub/pwg/ipp/new_SSDP/draft-cai-ssdp-v1-03.txt http://www.pwg.org/pub/pwg/ipp/new_SSDP/draft-cai-ssdp-v1-03.txt http://www.tomtom.com/ http://www.upnp.org/ http://www.wifly.com.tw/Wifly6/tw/WiflyNet/ServiceRange/HotPoint UbiPaPaGo: Context-aware path planning Introduction Related work Design of UbiPaPaGo Design goal Context model Architecture of UbiPaPaGo Privacy policy UbiPaPaGo based on SCM model UbiPaPaGo based on GA Chromosome representation and population initialization Fitness function Selection Crossover Mutation Prototyping and implementation Experimental results and discussion Optimal path evaluation Convergence behavior Optimization results comparison Effect of alpha value Conclusions and future work Acknowledgment References 
work_gtlsaeip3jeprhm4mykzjjhvba	----	V tonF-mcfH-is UCRL- 102060 PREPRINT MAESTRO - A Model and Exnert System Tuning Resource for Operators D. L. Lager, H. R. Brand, W. J. Maurer F. E. Coffield and F. Chambers This paper was prepared for submittal to The International Conference on Accelerator f, Large Experimental Physics Control Systems Vancouver, British Columbia October 30 - November 3, 1989 UCRL—102060 DE90 006775 This is a preprint of a paper intended for publication in a journal or proceedings. Since changes may be made before publication, this preprint is made available with the understanding that it will not be cited or reproduced without the permission of the author. MASTER DISTRIBUTION OF THIS DOCUMENT IS UNLIMITED Disclaimer This document was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government or the University of California nor any of their employees makes any warranty, expressed or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or uscfulncssofany information, apparatus, product, or process disclosed, orrepresents that its use would not infringe privately owned rights. Reference herein to any specific commercial products, process, or service by trade name, trademark, manufacturer, or otherwise, does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United Suites Government or the University of California. The views and opinions of authors expressed herein do not necessarily stale or reflect those of the United States Government or the University of California, and shall not be used for advertising or product endorsement purposes. M A E S T R O — A Model And Expert S y s t e m T u n i n g R e s o u r c e for O p e r a t o r s Darrel L. LAGER, Hal R. BRAND, William J. MAURER, Fred COFFIELD, and Frank CHAMBERS Lawrence Livermore National Laboratory, Livermore, California, USA ABSTRACT We have developed MAESTRO, a Model And Expert System Tuning Resource for Operators. It provides a unified software environment for optimizing the performance of la.ge, complex machines, in particular the Advanced Test Accelerator and Experimental Test Accelerator at Lawrence Livermore National Laboratory. The system incorporates three approaches to tuning: • A mouse-based manual interface to select and control magnets and to view displays of machine performance. • An automation based on "cloning the operator" by implementing the strategies and reasoning used by the operator. • An automation based on a simulator model which, when accurately matched to the machine, allows downloading of optimal sets of parameters and permits diagnosing errors in the beamline. The latter two approaches are based on the Artificial Intelligence technique known as Expert Systems. * This work was performed jointly under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under contract W-7405-ENG-48, for the Strategic Defense Initiative Organization and the U. S. Army Strategic Defense Command in support of SDIO/SDC-ATC MIPR No. W43-GBL-D-5007. 1 I. INTRODUCTION There is a class of large, complicated machines prevalent in physics research experiments that require expert human control to obtain a satisfactory level of performance. Members of this class include the Advanced Test Accelerator (ATA) and Experimental Test Accelerator (ETA) particle-beam accelerators, the Stanford Linear Accelerator (SLAC), the Superconducting Super Collider (SSC), and laser chains for Laser Isotope Separation (LIS) and Inertial Confinement Fusion (ICF). Many such machines may have several levels of control, with operators responsible for overall global control decisions and closed- loop feedback controllers responsible for relatively small sections of the machine. Over a period of time the operators develop strategies based on rules-of-thumb (heuristics) for optimizing the performance of the machine. As the machines become more complicated and as the performance criteria become more stringent there is a need to provide the operators with an on-line system model to aid in tuning and in failure diagnosis. MAESTRO (Model And Expert System Tuning Resource for Operators) is a software environment that combines physics models of the system and operator heuristics. The MAESTRO acronym was chosen to emphasize the conductor metaphor for unifying and coordinating the activities of control, diagnostics, modelling, and post-run analysis. Traditionally these activities are separate; often different people within different groups in the organization are responsible for each of these aspects. The principal advantage of using one consistent environment is rapid turn-around. Being able to observe the simulation while the machine is running allows the operator to make much more informed control decisions. Also, discrepancies between the model and the machine are readily apparent so that physicists can quickly gain insight into the phenomena that are occurring. A. Intelligent Assistant MAESTRO functions as an intelligent assistant for tuning, much as its predecessor, ATES (Accelerator Tuning Expert System) [Lager]. Displays permit the operator to see where MAESTRO is in the tuning processes and allow him to interrupt it so he may do manual tuning. There is a machine interrogation and control interface (M1C1) for manually triggering data acquisition and for changing magnet settings by using a mouse. There are menu-selectable tools for characterizing the machine by sweeping focusing and steering i magnets through a range of settings and by graphing the resulting beam positions. Finally, MAESTRO provides a very flexible runtime and development environment for rapid prototyping, fast debugging, and online construction of diagnostic codes. B. Tuning Tuning is the art and science of optimizing the performance of a machine for a desired experiment The art is in knowing that a certain bump on a waveform indicates that a particular power supply is off line, or that having a certain waveshape at the end of the injector gives poorer performance through the accelerator but much better performance in the wiggler. \ s a science, tuning may be viewed as a multivariate nonlinear optimization of multiple cost functions from multiple measurements. Typical systems involve tens to hundreds of interacting control variables, five to ten cost functions, and three or four types of measurements. Some of these measurements are readily available for computer manipulation (for example, digitized oscilloscope traces) and others are not (for example, audio information from microphones placed near the machine). Systems are often under­ diagnosed so that a single measurement is affected by several control parameters. Multiple data sets must then be taken to decouple the effects of the individual controls. We have chosen the domain of tuning particle beam accelerators to implement the concepts embodied in MAESTRO as an example of a complex system that can benefit from this approach. II. TUNING APPROACHES MAESTRO blends three distinct tuning approaches to achieve better performance than possible ft om any one approach alone. A trade-off among these approaches can be made as the machine is modified or grows more complex, as the operators learn more of its idiosyncrasies, or as more is understood about its physics. The first approach is "cloning the operator." We encode as faithfully as possible the strategies and reasoning followed by the operator. This was the dominant approach taken for tuning ATA since it was not possible within the allotted time to develop an accurate model for the beamline. A disadvantage of using only this approach is that operators may not be able to develop successful strategies for the far more complex machines being designed, for example the SSC. 3 Tlie second approach is model-based tuning, where an on-lins simulator of the system is coupled with the system's control and diagnostics subsystems. This approach has the advantage that a complete set of tuning parameters can be determined using the model and then simply downloaded onto the real machine at a substantial time savings. The disadvantage is that it may not be economically feasible to develop a sufficiently accurate simulator. The third approach is to tune manually, but to provide the operator with more powerful tools for tuning the machine. This takes the form of displays derived from the raw data the operator is normally presented, and different interfaces making it easier to control the machine. For most applications the manual approach is needed because the operator will almost always have more knowledge of the machine than is economically feasible to encode into a computer program. He can take into account information that may not be readily accessible to the computer, such as the "sound" the machine makes when it is not quite running right. He may also be able to instantly diagnose a failure because for example he remembers what happened when that same situation arose on a machine he was tuning 20 years previous. The goal is to achieve a blend of these three approaches that minimi2es the time required to tune and maximizes the time available for performing physics experiments. Each of these approaches is discussed below. A. Cloning the Operator This was the dominant approach taken for tuning ATA since it was not possible within the allotted time to develop an accurate model for the beamline. The fundamental goal followed by the expert was to tune for a zero; that is, tune to put the beam on the center of the pipe. Given that goal, it was unnecessary to be concerned with rotation, tilt, and calibration-factor errors in the beamline components. He only needed to be qualitatively concerned with coupling errors between the vertical and horizontal steering magnets. We concentrated our efforts on tuning the transport section of ATA. The tuning process was to follow a global strategy by selectively applying local strategies involving two to four components in the section. The local strategies were implemented using the Monitored Decision Script (MDS) representation we developed for ATES [Lager]. The scrip', part of the MDS represented the step-by-step procedure followed by the operator, the decision part represented the choices made when there was anomalous behavior by the 4 machine, and the monitor part emulated the behavior on the pan of the operator when he would periodically "pop his head up" and check the overall status of the machine, especially those portions affecting the section being tuned. Our initial approach for implementing the local strategies was to decompose the problem into two simpler ones, namely to center the beam in X, and then center in Y. We had the capability to account for the relative rotations of the components and their various calibration factors. Unfortunately, when we attempted tuning ATA in February 1989, there was coupiing between the steering magnets that we could not accommodate, so we modified the local strategies to more closely follow the operator's procedure. There is now an initial phase where a pair of steerers are chosen, one nominally steering vertically, the other horizontally, and an experiment is performed to diagnose which most affects the X position and which affects ¥. These steerers are used to center the beam by iteratively halving the error in X, then in Y, using a uniplex adaptive optimization algorithm [King]. With this approach, we were able to successfully center the beam. The global tuning strategy is driven by the overriding constraint, "don't put the beam into the wall." As soon as the machine is powered up the operator checks the beginning of each major section to see if the beam is there. If not, he does coarse steering in the previous section to get the beam through it Once he has the beam to the tuning dump, he goes back to the beginning of the transport section and meticulously centers the beam while monitoring the current reaching the tuning dump. For the meticulous centering, he uses more sophisticated local strategies based on pairs of steering magnets, where the first member of the pair is used to position the beam at the center of the second member. The second is then used to remove any angular offset in the beam. We developed an inference engine [Chamiak] to implement the global strategy. The philosophy is to choose and execute an appropriate local strategy, based on the present state of the machine. The inference engine operates by repeatedly "matching," "selecting," and "executing" MDSs until the desired machine state has been achieved. Since each MDS has a pre-condition field defining the state of the machine when it is valid to be executed, and a post-condition field defining the state of the machine if it successfully executes, the "match" part of the cycle consists of scanning all the MDSs to find which ones match the present state of the machine. The "select" part of the cycle chooses one of the MDSs to execute if there is more than one match. Then the selected MDS is "executed." The actions of the MDS change the state of the machine, and the cycle is repeated. 5 B. Model-Based Tuning Model-based tuning promises to put more science into the art of tuning leading to a more rigorous understanding of the machines. The model can enable operators to "see" the effects of their tuning in the regions between sensors. Given sufficiently accurate models, it should be possible to determine a set of parameters that will change the present state ot the machine to a desired (i.e., tuned) state in a single step. With less accurate models, iteration will be necessary, but will typically converge rapidly. Experiments can be performed off-line using the model to determine, for example, the effects of adding new components to the beamline at far less expense than involving the real machine. MAESTRO directly couples the beamline model to the machine representation in order to make all the present magnet fields, locations, calibration factors, rotation angles, and tilt angles directly available to the model. A nonlinear least-squares algorithm performs least- squares fits to data sets of beam positions. Fit parameters include magnet field strengths and component displacements, rotations, and tilts. The art (and artificial intelligence) is in choosing the proper subset of parameters to be fit and in choosing between alternative fits when more than one explanation of the data is possible. Model-based tuning requires two distinct phases. The first is "commissioning" [Lee], where the simulator is matched with the real machine. This phase forces proper bookkeeping because it requires all the components to be properly characterized. Any rotations, tilts, and miscalibrations must be eliminated from the machine or incorporated into the model during this phase so that the simulator and real machiie produce the same beam trajectory. We have developed a variety of tools for commissioning. Steering sweeps measure the beam position as first the horizontal steerer is swept through a range of values, then the vertical steerer is swept. This produces the cross-shaped pattern in Fig. 1. Focus sweeps measure the beam position as the field in a focusing magnet is varied over a range. This produces the spiral-shaped pattern in Fig. 2. We have begun the commissioning on the smallest subset of ETA possible, one magnet followed immediately by a beam-bug in the injector section of the machine. Once these components have been characterized, we will commission the remaining injector magnets and then proceed down the accelerator, beginning with the first 10-cell set. Proceeding in this fashion insures that the smallest number of noncommissioned devices influence a measured value. A typical commissioning fit for a steering sweep is shown in Fig. 3 6 where the small squares are measured beam positions and the large squares are model predictions. The second or "operational" phase is actual accelerator tuning and diagnosis. With the simulation model and accelerator in relative agreement, tuning can be done by using the nonlinear parameter estimator to select the optimum tuning parameters for the injected beam after its launch conditions have been estimated. Obviously, a very large number of tuning parameters are available. Initially, an experienced operator will select a subset of these for tuning, and will iterate the process with different subsets. Eventually, the operator's knowledge and experience will be encoded into an expert system that will select the subset of tuning parameters and will decide how much iteration is necessary When the simulation model and the accelerator go out of relative agreement, the model can be used to find those regions within the accelerator where there is still agreement For a single component failure, there will be two regions of agreement roughly surrounding the failed component. From this information, a set of suspect components can be constructed. With the aid of an expert operator (or eventually, an expert system), a number of possible failures can be proposed. The nonlinear parameter estimator is then used to estimate both the magnitude of the proposed error and the improvement in the model's predictive ability, given the proposed error and the data currently available from the accelerator sensors. With this information, many proposed errors can be rejected because either the magnitude of the error is beyond reasonable limits, and/or the improvement in the model by the addition of the error is too small. In the ideal case, only one reasonable error remains. In other cases, one is normally left with only a few possibilities. Beam redirection experiments can then be performed to further isolate the accelerator-model discrepancy. Selection of these experiments will initially be done by experienced operators but will eventually be included within the expert system's capabilities. When the problem is finally found, then the accelerator can be fixed and/or the model can be updated with the (fit) error information, and tuning can proceed as before. C. Manual Interface The manual interface for MAESTRO is shown in the screen dump in Fig. 4. The interface consists of several windows that become visible on trc screen as necessary. The machine interrogation and control interface (MICI) window is the main one for interacting with the system. It consists of two panes, with the upper one for displaying the output from the simulator, showing the horizontal beam position vs location along the beamline as 7 a thick line and the vertical position as a thin line. The lower pane is a set of icons depicting the components in the beamline. The locations of the icons and their shape are derived from the information describing the beamline in the MAESTRO knowledge base. Components added to the beamline are automatically included in the MICI display, once the information has been added to the knowledge base. The operator controls the fields in the magnets by clicking a mouse button as a cursor is positioned over an icon. Clicking over a vertical steering magnet icon, for example, causes power supply control windows to become visible below the MICI window, as in Fig. 4. Clicking the mouse as the cursor is positioned in those windows causes the current in the appropriate power supply to be incremented or decremented. The scope display window is used to control and display data acquisition. Clicking the mouse over the appropriate label in the window causes data to be acquired and displayed as a set of three traces in the window, giving the time histories of the beam current, X position, and Y position. The actual location of the beam within the pipe is derived from the trace data, after accounting for sensor misalignment and is displayed in the circular "bulls-eye" displays in the upper part of the scope display window. Other windows display additional derived data. A position display window shows a value vs position down the beamline. A "bug-walk," for example, makes posiaon displays that show peak current, X position, and Y position at the locations of the beam-bug position monitors along the beamline. Similarly, a "magnet-walk" shows the fields at the centers of the magnets vs their location in the beamline. Three magnet walks appear in Fig. 4, showing the solenoidal fields (labeled PD-B), the horizontal steering fields (labeled PD- H), and the vertical steering fields (labeled PD-V) vs distance down beamline. There are also windows for displaying historical data. Clicking the mouse over an icon for a beam bug causes a "shot history" window to appear. By clicking the mouse over buttons on the window, the operator can view all the oscilloscope traces acquired for that bug for that day's run. I'f. FUTURE We are planning to extend the capabilities of the shot history mechanism to improve its flexibility for manipulating not only past shot data but also past machine configuration data. We want to have the capability to redo the signal processing with different control parameters, for example. We also want to make queries, such as: "during the last three X months what was the highest current magnitude measured when the machine had the long collimator installed?" The approach we sie taking is to develop an unstructured database based on the Artificial Intelligence representation scheme knows as a semantic network. We are also developing the capability to acquire image data from cameras that observe the beam as it strikes foils inserted into the beamline. We will apply image-understanding techniques to determine the positic-; and focus of the beam for making tuning decisions. IV. SUMMARY The MAESTRO software environment was developed to function as an intelligent assistant to an operator in tuning complex systems such as particle-beam accelerators. It incorporates three approaches to tuning. The "cloning the operator" approach uses an inference engine and the MDS representation to encode the strategies and reasoning followed by the operator. The model-based approach makes use of a beamline simulator and a nonlinear least-squares parameter estimator first to match the model with the machine and then to determine optimum tuning parameters after computing the beam launch conditions. The third approach lets the operator manually perform tuning and provides him with displays that easily let him determine the machine status. Finally, a history mechanism lets the operator view past data to compare the present tuns with ones previously obtained. ACKNOWLEDGEMENTS The authors gratefully acknowledge the contributions of Doyle Rogers and lohn Clark, the tuning experts. Their ability to make order out of chaos is truly amazing. REFERENCES Charniak, Eugene; Riesbeck, Christopher K.; and McDermott, Drew V., Artificial Intelligence Programming, Lawrence Erlbaum Associates, Hillsdale, New Jersey, 1980. King, Paul G. and Demming, Stanley N., "UNIPLEX: Single Factor Optimization of Response in the Presence of Error," Analytical Chemistry, 46, 1476-1481, (1974). 9 Lager, Darrel L.; Brand, Hal R.; Maurer, William J.; Searfus, Robert M.; and Hernandez, Jose E.; "An Expert System for Tuning Particle-Beam Accelerators," Proceedings of SPIE- The International Society for Optical Engineering, Vol. 1095, Applications of Artificial Intelligence VII, 28-30 March, 1989. Lee, Martin and Clearwater, Scott, "GOLD: Integration of Model-Based Control Systems with Artificial Intelligence and Workstations," Workshop on Model-Based Accelerator Controls, Brookhaven National Laboratory, Upton, New York, 17 August 1987. 10 Figure Captions Figure 1. Variation of X and Y beam position as the vertical and horizontal steering magnets are swept through a range of values. Figure 2. Variation of X and Y beam position as the solenoidal field is varied through a range of values. Figure 3. Comparison of fit between model and measured data for a steering sweep. The small squares indicate measured values, the large squares indicate model predictions. Figure. 4. Screen dump of the windows associated with the manual interface to MAESTRO. These windows include: (a) simulated beam position vs distance down beamline with theX position shown by the thin line and Y position by the thick line; (b) icons depicting the components according to their positions in thp beamline; (c) power supply control window; (d) scope display showing tta /, X, and Y waveforms from the selected beam bugs; and (e) field strength vs distance down the beamline for the solenoids, horizontal steerers, and vertical steerers. 11 ( . 1 2 1 , Figure 1. Figure 2. 12 r RESOLVE BEAMLINE SIMULATION PLOT 0 a E 0 a • m 0 a E I s a1 p • 0 • • 0 0 0 s 0 B 0 0 Q 0 Figure 3. k I-I 
work_gtr6ulhdsrgaxlb6rri4jrwvra	----	PII: 0898-1221(91)90112-H Computers Math. Applic. Vol. 21, No. 11/12, pp. 111-116, 1991 0097-4943/91 $3.00 + 0.00 Printed in Great Britain. All rights reserved Copyright~) 1991 Pergamon Press plc F U Z Z Y D E C I S I O N T A B L E S F O R E X P E R T S Y S T E M S ART LEW Department of Information and Computer Science University of Hawaii (Received July 1990) A b s t r a c t - - T h e use of fuzzy decision tables as a programming language for representing both the knowledge and the procedures in expert systems is discussed. Examples of their use for the generation of procedural code and for the generation of if-then rules are given. I. INTRODUCTION The main a~Ivantages of decision tables [1,2] are that: (a) they provide a systematic way to design algorithms; (b) they are amenable to certain forms of automated error-checking (such as for consistency [3]) and to formal verification (by correctness proofs [4]); and (c) a processor to compile and execute them is relatively easy to implement [5]. Decision tables (DTs) are compatible with most programming languages, including general- purpose languages such as Ada and C as well as special-purpose languages such as LISP [6] and Prolog; the processor described in [5] can be adapted, with varying degrees of difficulty, to translate DTs into any of these languages. We previously discussed the use of DTs as a general-purpose programming language [7]; here, we discuss their use for the implementation of rule-based expert systems. 2. F U Z Z Y A L G O R I T H M S In an earlier p a p e r [8], we d e m o n s t r a t e d t h a t an y ( n o n f u z z y ) a l g o r i t h m can b e i m p l e m e n t e d b y a D T . In principle, a n y f u z z y a l g o r i t h m [9,10], if it is t o b e realized (i.e., e x e c u t e d on an a c t u a l c o m p u t e r ) , can also be i m p l e m e n t e d using a D T , al b ei t possibly a n o n d e t e r m i n i s t i c one [11]. For e x a m p l e , a f u z z y version o f t h e D T in [7] for finding t h e s h o r t e s t p a t h in a g r a p h can be b ased on f u z z y o p t i m i z a t i o n a l g o r i t h m s [12]; some o t h e r ex am p l es o f f u z z y o p t i m i z a t i o n a l g o r i t h m s a p p e a r in [13-15]. 3. A P P L I C A T I O N O F D T S T O E X P E R T S Y S T E M S It is n a t u r a l t o use D T s to express the knowledge base o f an e x p e r t s y s t e m [16]. Each IF-THEN rule o f t h e f o r m IF P1 . o . AND Pn THEN CI AND Cm {with certainty factor of=n} can be r e p r e s e n t e d by a rule (or c o l u m n ) of a D T Ty p ~et by A ~ - ~ _ X 111 112 A. L~.w P1 : T Pn : T Cl :X am: X {cf = n} where the premises and consequents of the productions are the conditions and the actions of the DT. This idea is quite old; numerous rule-based expert systen~ incorporating decision-table concepts in their design and analysis have been reported (see, for example, [17219]). Much of this prior work concerned methods of detecting ambiguities, regarded as a problem in that other context; here, we regard ambiguities as a desirable way to express fuzziness. We distinguish between "procedural" DTs, in which the consequents are general actions (which may include "call" statements), and "nonprocedural" DTs in which the consequents are simple actions (i.e., value assignments). Generally speaking, we liken procedural DTs to forward-chaining production systems such as OPS5 [20], and nonprocedural DTs to backward-chaining inference systems such as Mycin [16]. Decision table processors can be designed to perform both forward and backward chaining. In this paper, we focus our attention on procedural DTs. 4. F U Z Z Y DTS DTs are suitable for expressing fuzziness in expert systems in a number of ways. The fuzzi- ness may be in multi-valueness of conditions, nondeterminism of rules, or a variety of forms of uncertainty in the entries of any of the four DT quadrants or with respect to any row or column. For example, in the above DT, Pn or Cm may be fuzzy expressions, their values may be fuzzy rather than truth-valued possibly with varying degrees of certainty, and rules themselves may have associated certainty factors and may or may not be mutually exclusive. These variations can all be accommodated within the framework of "ambiguous" DTs. Ambiguous procedural DTs can be implemented as nondeterministic algorithms. One problem in the handling of fuzziness for which there is no clear-cut solution relates to how fuzzy certainty factors should be defined and blended. Even in nonfuzzy contexts, different expert systems have adopted different conventions, none of which may be appropriate for a given application [21]. Use of DTs permits users to program their own blending formulas; unlike the case in most systems, many different formulas may be used in the same application and these formulas may be dynamically defined. 5. E X A M P L E S Examples of the application of fuzzy procedural DTs to the design of expert systems appear in [12], in which the nonfuzzy DTs used to solve the stock market problem as given in [2] were modified in what are rather minor ways. Even so, there are several errors in the fuzzy DTs of [12], such as the one given here as Table 1. Table 1. ~1 ~ 2 R3 R4 trading possible? : Y Y N stkavg (X1)\ } : XI+X2 XI.X2 bndavg (X2)/ XI/X2 call(stocks) : <=0.2 >0 call(bonds) : <=0.2 =0 call(account) : <=0.2 X process R4 : <=0.4 go again : <=0.2 X stop >0.4 X Fuzzy decision t a b l e s for e x p e r t systems Table 2. trading possible ? : s t k a v g ( X 1 ) \ ) bndavg (X2)/ R1 R2 R2 ~ R3 : R4 Y Y Y N X I + X 2 Xl.X2 Xl/X2 update averages : X X X call(stocks) : <=0.2 call(bonds) : <=0.2 call(account) : <=0.2 process R4 {goto~ : <=0.4 go again {repeat} : <=0.2 stop ~exit} : >0.4 X =0 >0 X X 113 (Here, + , . , and / are fuzzy max, min, and diff operators.) A corrected version is given in Table 2; note the reversal of the relations X1/X2 > 0 and Xl/X2 = 0. The adoption in these procedural tables of a left-to-right interpretation convention with implicit ELSE rules, as well as of conditional rules associated with "goto" statements (such as to rule R4), is, we believe, undesirable. Table 3 shows an equivalent table which utilizes different conventions in which the "logic" is expressed in the upper right quadrant. T a b l e 3. trading possible? : T T T T F stkavg+bndavg : <=0.2 >0.2 >0.2 >0.2 - stkavg.bndavg : - <=0.4 <=0.4 >0.4 - stkavg/bndavg : - >0 =0 - - update averages : X X X X - call(stocks) : X - X - - call(bonds) : X X - - - call(account) : X X X - - repeat : X X X - - exit : - - - X X One advantage of this format, in which the rules can be evaluated in any order, is t h a t optimal conversion algorithms (such as in [22]) may then be applied. (To simplify lexical processing, we used a slightly different "syntax" in our implementation, as illustrated below.) 6. I M P L E M E N T A T I O N We attribute the errors in Table 1 and the other fuzzy DTs of [12] to the unavailability of a programming system with which the tables could be tested. This provided us with the motivation to design such a programming system. Our system [which at this writing is not yet complete] is essentially a preprocessor which translates fuzzy DTs, expressed in a format similar to that used in Table 3, into a conventional (procedural) language in a manner like that described in [5]. The precise DT format we adopted is given in the Appendix 1; its translation into a procedural language (which can be executed using an ordinary compiler) is shown in Appendix 2. Using the slightly different syntax given in Appendix 3, the DT can be translated into the set of IF-THEN rules shown in Appendix 4, which in turn can be processed using a simple expert system such as in [23]. 7. C O N C L U S I O N Fuzzy DTs can be used to express the knowledge base of an expert system. DTs can also be used as the programming language in which expert systems are implemented (i.e., programmed, 114 A. L~.w in a c h o i c e o f e n v i r o n m e n t s , s u c h as L I S P a n d C ) . T h a t use o f D T s is a n a d v a n t a g e o u s w a y f o r h u m a n s t o i n t e r a c t w i t h c o m p u t e r s , e s p e c i a l l y as a m e a n s t o r e p r e s e n t p r o d u c t i o n rules, is d e m o n s t r a t e d b y t h e i r f r e q u e n t a d o p t i o n in t h e A I l i t e r a t u r e (e.g., see [13,20,24]). O f c o u r s e , s i n c e t a b l e s a r e c o m m o n l y u s e d for a v a r i e t y o f o t h e r k i n d s o f i n f o r m a t i o n , s u c h as r e l a t i o n a l d a t a b a s e s , d e s i g n o f a n i n t e g r a t e d s y s t e m e m p l o y i n g t a b u l a r r e p r e s e n t a t i o n s o f k n o w l e d g e a n d p r o c e d u r e s is w o r t h s e r i o u s c o n s i d e r a t i o n . A n o t h e r a d v a n t a g e o f t h e u s e o f f u z z y D T s t o i m p l e m e n t e x p e r t s y s t e m s is flexibility. D T s p e r m i t b o t h f o r w a r d a n d b a c k w a r d c h a i n i n g , a n d allow t h e i n c o r p o r a t i o n o f d i f f e r e n t w a y s o f h a n d l i n g f u z z i n e s s ( s u c h as b l e n d i n g c e r t a i n t y f a c t o r s ) w i t h i n t h e s a m e a p p l i c a t i o n . ( D e t a i l s o f a d e s i g n w h i c h is c o m p a r a b l e t o S y s t e m Z-11 [25] will b e d i s c u s s e d in a f o r t h c o m i n g p a p e r . ) W e c o n c l u d e t h a t t h e d e s i g n o f e x p e r t s y s t e m s u t i l i z i n g f u z z y D T s is w o r t h w h i l e a n d w a r r a n t s f u r t h e r r e s e a r c h a n d d e v e l o p m e n t , a n d we a r e so p r o c e e d i n g . O n e o p e n r e s e a r c h p r o b l e m is w h e t h e r b e t t e r e x p e r t s y s t e m s c a n b e i m p l e m e n t e d u s i n g f u z z y D T p r o c e s s o r s , i.e., b y i n c o r p o - r a t i n g f u z z i n e s s i n t h e p r o c e s s o r r a t h e r t h a n in t h e t a b l e s u p o n w h i c h t h e y o p e r a t e . REFEIIEN CES 1. CODASYL Task Group, A Modern Appraisal of Decision Tables, ACM, New York, (1982). 2. R.B. Hurley, Decision Tables in Software Engineering, Van Nostrand Reinhold, New York, (1983). 3. P.J.H. King, The interpretation of limited entry decision table format and relationships among conditions, Computer Journal 12,320-326 (1969). 4. A. Lew, Proof of correctness of decision table programs, Computer Journal 27,230-232 (1984). 5. A. Lew, A decision table processor, University o] Hawaii, technical report (1988). 6. B.M. Schwartz, LISP 1.5 decision tables interpreted for a small computer and proposed for parallel com- puters, S I G P L A N Notices 6 (8), 93-103 (1971). 7. A. Lew, Decision tables for general-purpose scientific programming, Software-Practice and Experience 13, 181-188 (1983). 8. A. Lew, On the emulation of flowcharts by decision tables, Commun. A C M 25, 895-905 (1982). 9. L. Zadeh, Fuzzy algorithms, Information and Control 12, 94-102 (1968). 10. L. Zadeh, The role of fuzzy logic in the management of uncertainty in expert systems, Fuzzy Sets and Systems 11, 199-227 (1983). 11. A. Lew, Decision table programmlng-some new perspectives, University o.f Hawaii, technical report (1983). 12. A. Kandel, Fuzzy Mathematical Techniques with Applications, Addison-Wesley, Reading, Mass., (1986). 13. R. Bellman and L.A. Zadeh, Decision-making in a fuzzy environment, Management Sci. 17 (4), B141-164 (1970). 14. R.L.P. Chang, Fuzzy decision tree algorithms, IEEE Trans. Sys. Man Cyb. 7, 28-35 (1977). 15. A.O. Esogbue, Dynamic programming, fuzzy sets, and the modeling of R & D management control systems, IEEE Trans. Sys. Man Cyb. 13, 18--40 (1983). 16. B. Buchanan and E. Shortliffe, Rule-Based Expert Systems, Addison-Wesley, Reading, Mass., (1984). 17. M. Trigoboff and C.A. Kulikowski, IRIS: A system for the propagation of inferences in a semantic net, Proc. 5th IJCAI, 274-280 (1977). 18. B.J. Cragun and H.J. Steudel, A decision-table-based processor for checking completeness and consistency in rule-based expert systems, Int. J. Man-Machine Studies 26,633-648 (1987). 19. R. Maes and J.E.M. VanDijk, On the role of ambiguity and incompleteness in the design of decision tables and rule-based systems, Computer Journal 31,481-489 (1988). 20. T. Cooper and N. Wogrin, Rule-Based Programming in OPSS, Morgan Kaufmarm, San Mates, Calif., (1988). 21. D. Kopso, L. Pipino and W. Rybolt, A comparison of the manipulation of certainty factors by individuals and expert system shells, J.MIS 5, 66-81 (1988). 22. A. Lew, Optimal conversion of extended-entry decision tables with general cost criteria, Commun. A CM 21,269--279 (1978). 23. B. Sawyer and D. Foster, Programming Expert Systems in Pascal, Wiley, New York, (1986). 24. M.A. Carrico, J.E. Girard and J.P. Jones, Building Expert Systems, McGraw-Hill, New York, (1989). 25. K.S. Leung and W. Lava, Fuzzy concepts in expert systems, Computer 21 (9), 43-53 (1988). Fuzzy decision tables for expert systems 115 A P P E N D I X A A F~zz~ Procedural DT. dtbegin! trading possible max(stkavg,bndavg) min(stkavg,bndavg) diff(stkavg,bndavg) update averages c a l l ( s t o c k s ) call(bonds) call(account) repeat! exit! dtend! :T T T T F !<=0.2 >0.2 >0.2 >0.2 - !- <=0.4 <=0.4 >0.4 - !- >0 =0 - - :X X X X - :X - X - - :X X - - - :X X X - - :X X X - - :- - - X X A P P E N D I X B Translation oJ the D T of Appendix A. while not(lambda<O) do begin if (trading possible) and (max(stkavg,bndavg)<=0.2) then begin update averages; call (stocks); call (bonds); call (account); lambda:=1; end else if (trading possible) and (max(stkavg,bndavg)>0.2) and (min(stkavg,bndavg)<=0.4) and (diff(stkavg,bndavg)>O) then begin update averages; call (bonds); call (account); lambda:=l; end else if (trading possible) and (max(stkavg,bndavg)>0.2) and (min(stkavg,bndavg)<=0.4) and (diff(stkavg,bndavg)=O) then begin update averages; call (stocks); call (account); lambda:=1; end else if (trading possible) and (max(stkavg,bndavg)>0.2) and (min(stkavg,bndavg)>0.4) then begin update averages; lambda:=-1; end else if not(trading possible) then begin lambda:=-1; end else begin lambda:=-2; end; end; 116 A. Lsw A P P E N D I X C A F~zz~ Exper~ S~/8~em (in ~ D T Formal). d t b e g i n ) t r a d i n g p o s s i b l e = m a x ( s t k a v g , b n d a v g ) min(etkavg,bndavg) diff(stkavg,bndavg) update averages c a l l ( s t o c k s ) c a l l ( b o n d s ) call(account) exit dtend! IT T T T F !<=0.2 >0;2 >0.2 >0.2 ! <=0.4 <=0.4 >0.4 ! >0 =0 !, m X X A P P E N D I X D R~le8 Asaociated wi~A the Expert S~mtern of A~pe.dix C. Rule 1.1: if trading possible=T and max(stkavg,bndavg)<=0.2 then update averages, and call(stocks), and call(bonds), and call(account). Rule 1.2: if trading possible=T and max(stkavg,bndavg)>0.2 min(stkavg,bndavg)<=0.4 diff(stkavg,bndavg)>O then update averages, and call(bonds), and call(account). Rule 1.3: if trading possible=T and max(stkavg,bndavg)>0.2 min(stkavg,bndavg)>=0.4 diff(stkavg,bndavg)=O then update averages, and call(stocks), and call(account). Rule 1.4: if trading possible=T and max(stkavg,bndavg)>0.2 min(stkavg,bndavg)>0.4 then update averages, and exit. Rule 1.5: if trading possible=F then and exit. and and and and and 
work_gueb22ja6jgt3pxyocpuq4udkq	----	PII: 0888-613X(88)90170-3 346 A b s t r a c t s assumes marginal independence between inputs. This demonstrates that the importance of updating formulas can outweigh that o f prior assumptions. Thus, when UISs are judged by their final accuracy after optimization, completely different results are obtained than when they are judged by whether or not their prior assumptions are perfectly satisfied. Structuring Causal Tree Models with Continuous Variables L e i X u Department o f Automation, Tsinghua University, Beijing, China J u d e a P e a r l Cognitive Systems Laboratory, Computer Science Department, UCLA, Los Angeles, California 90024-1600 This paper considers the problem o f invoking auxiliary, unobservable variables to facilitate the structuring o f causal tree models for a given set o f continuous variables. Paralleling Pearl's 1986 treatment o f bivalued variables, it is shown that if a collection of coupled variables are governed by a joint normal distribution and a tree-structured representation exists, then both the topology and all internal relationships of the tree can be uncovered by observing pairwise dependencies among the observed variables (i.e., the leaves o f the tree). Furthermore, the conditions for normally distributed variables are less restrictive than tho3e governing bivalued variables. The result extends the applications of causal tree models that were found useful in evidential reasoning tasks. Address correspondence to L. Xu. Implementing Evidential Reasoning in Expert Systems J o h n Y e n USC/Information Sciences Institute, 4676 Admiralty Way, Marina del Re),, California 90292 The Dempster-Shafer theory has been extended recently for its application to expert systems. However, implementing the extended D-S reasoning model in rule-based systems greatly complicates the task o f generating informative explanations. By implementing GERTIS, a prototype system for diagnosing rheumatoid arthritis, it is shown that two kinds of knowledge are essential for explanation generation: (1) taxonomic class relationships between hypotheses and (2) pointers to the rules that significantly contribute to belief in the hypothesis. As a result, the knowledge represented in GERTIS is richer and more complex than that of conventional rule-based systems. GERTIS not only demonstrates the feasibility o f rule-based evidential reasoning systems, but also suggests ways to generate better explanations and to explicitly represent various useful relationships among hypotheses and rules. Can Evidence Be Combined in the Dempster-Shafer Theory? J o h n Y e n USC/lnformation Sciences Institute, 4676 Admiralty Way, Marina del Rey, California 90292 
work_gv6cgj4ojjaqpj6znvialeabaq	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220980256 Params is empty 220980256 exception Params is empty 2021/04/06-03:05:21 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220980256 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:21 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_gwrgldbvi5amdc6a3gmpm33lzi	----	32331.141_149 BRIDGE DAMAGE ASSESSMENT THROUGH FUZZY PETRI NET BASED EXPERT SYSTEM By Weiling Chiang,1 Kevin F. R. Liu,2 and Jonathan Lee3 ABSTRACT: The reliability of the techniques adopted for damage assessment is important for bridge manage- ment systems. It is widely recognized that the use of expert systems for bridge damage assessment is a promising direction toward bridge management systems. However, several important issues need to be addressed, such as the management of uncertainty and imprecision, the efficiency of fuzzy rule based reasoning, and the need of an explanation facility to increase confidence about the assessment results. To address the issues arising from using expert systems, this paper is aimed at developing an expert system for assessing bridges based on an expert system shell, which is called the fuzzy Petri net based expert system (FPNES). Major features of FPNES include the ability to reason using uncertain and imprecise information, knowledge representation through the use of hierarchical fuzzy Petri nets, a reasoning mechanism based on fuzzy Petri nets, and an explanation of the reasoning process through the use of hierarchical fuzzy Petri nets. Therefore, this expert system for assessing bridges does not impose any restriction on the inference mechanism. Furthermore, this approach offers more informative results than other systems. An application to the damage assessment of the Da-Shi bridge in Taiwan is used as an illustrative example of FPNES. INTRODUCTION In recent years, many countries have been aware of bridge problems and have initiated the development of bridge man- agement systems to assist their decision-makers in establishing efficient repair and maintenance programs. Basically, bridge management systems consist of several modules to record the inventory data of bridges, store the inspection results, evaluate the damage states of bridges, propose maintenance and retrofit schemes, estimate the maintenance and retrofit cost, and al- locate the available funds appropriately. The reliability of the technique adopted for evaluating bridge states is playing an important role in bridge management systems. Damage as- sessment for a bridge is a process to evaluate the damage state of the bridge based on visual inspection and empirical tests. However, it is a difficult task due to a lack of complete un- derstanding of the mechanism of bridge deterioration. Bridge structures are too complex to analyze completely, and there- fore, numerical simulations require a host of simplified as- sumptions. Nevertheless, an experienced engineer who has closely studied these problems over the years could use his heuristic knowledge to achieve the task. The use of expert systems for damage assessment has been widely recognized as a promising solution for alleviating the lack of experts. An expert system is a computer program that emulates the problem-solving ability of human experts. Separating the knowledge base from the inference procedure makes it easy to extend and modify knowledge collected from all possible sources. Expert systems can perform task much more effi- ciently than experts, and can even be accessed through the Internet without time and space barriers to serve as a ‘‘knowl- edge server’’ (Eriksson 1996). Recently, researchers have be- gun to investigate the use of expert systems to assess damage assessment — for example, Yao (1980), Ishizuka et al. (1982), Ogawa et al. (1985a,b), Furuta et al. (1991, 1996), Ross et al. 1Prof., Dept. of Civ. Engrg., Nat. Central Univ., Chungli, Taiwan 32054. E-mail: chiang@cc.ncu.edu.tw 2PhD, Dept. of Civ. Engrg., Nat. Central Univ., Chungli, Taiwan 32054. E-mail: frliu@se01.csie.ncu.edu.tw 3Prof., Dept. of Comp. Sci. and Information Engrg., Nat. Central Univ., Chungli, Taiwan 32054. E-mail: yjlee@se01.csie.ncu.edu.tw Note. Editor: Sivand Lakmazaheri. Discussion open until September 1, 2000. To extend the closing date one month, a written request must be filed with the ASCE Manager of Journals. The manuscript for this paper was submitted for review and possible publication on September 28, 1998. This paper is part of the Journal of Computing in Civil En- gineering, Vol. 14, No. 2, April, 2000. qASCE, ISSN 0887-3801/00/ 0002-0141–0149/$8.00 1 $.50 per page. Paper No. 19364. (1990), Hadipriono and Ross (1991), Shiraishi et al. (1991), Krauthammer et al. (1992), Miyamoto et al. (1993), and Kush- ida et al. (1997). Several important issues need to be addressed in using expert systems for damage assessment of bridges: • The descriptions of heuristic damage assessment knowl- edge by experts usually take the form of natural language that contains intrinsic imprecision. For example, experts may treat the extent, degree, and seriousness of a defect as linguistic variables that have fuzzy values. Therefore, a mechanism that can cope with imprecise information in expert system is essential. • Uncertain and imprecise information involved in the dam- age assessment usually makes the problem harder. There- fore, a reasoning mechanism that can deal with uncertain and imprecise information is crucial for damage assess- ment. • To consider the efficiency of fuzzy rule based reasoning, it is important that fuzzy facts (input or inferred) find the matched fuzzy rules efficiently, rather than scanning through all the fuzzy rules. • In order to increase the confidence about the assessment results, an explanation facility that can describe how the conclusions are derived is instrumental for a computer- aided tool designed especially for damage assessment. To address the first two issues, the writers have proposed a possibilistic logic based approach (Lee et al. 1998) to handling uncertain and imprecise information. The writers have also proposed a fuzzy Petri net approach to addressing the last two issues. The writers have proposed a framework of integrated expert systems based on proposed fuzzy Petri nets, called fuzzy Petri net based expert systems (FPNES) (Lee et al. 1999). This paper is aimed at developing an expert system for assessing bridges based on FPNES. Unlike other researchers’ systems, our expert system does not impose any restriction on the inference mechanism, that is, the intended meaning is not required to be intact; meanwhile, the confidence level can be partially certain. Furthermore, it offers more informative re- sults because the explanation provided in the system and the confidence level of the conclusions can be used as a way of justification on whether to take the recommendations into ac- count or not. An application to the damage assessment of the Da-Shi bridge in Taiwan is used as an illustrative example of FPNES. The result through FPNES not only matches the ex- JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 141 perts’ judgments, but also provides more information than ex- perts do. FUZZY PETRI NET BASED EXPERT SYSTEM FPNES, a framework of integrated expert systems based on the proposed fuzzy Petri net, has been proposed by Lee et al. (1999) and is briefly described in this section. Major features of FPNES include the ability to reason using uncertain and fuzzy information, explanation of the reasoning process through the use of hierarchical fuzzy Petri nets, and a reason- ing mechanism based on fuzzy Petri nets. REASONING FOR UNCERTAIN AND FUZZY INFORMATION The distinction between imprecise and uncertain informa- tion can be best explained by the canonical form representation (i.e., a quadruple of attribute, object, value, and confidence) proposed by Dubois and Prade (1988). Imprecision implies the absence of a sharp boundary of the value component of the quadruple, whereas uncertainty is related to the confidence component of the quadruple, which is an indication of the reliability of the information. To represent uncertain imprecise information, a fuzzy prop- osition with a fuzzy valuation has been chosen by Lee et al. (1998), denoted as (r, t) (1) where r is fuzzy proposition of the form ‘‘X is F ’’ [i.e., X is a linguistic variable (Zadeh 1975, 1976) and F is a fuzzy set in a universe of discourse U]; and t is a fuzzy valuation. It should be noted that for every formula (r, t) (called a truth- qualified fuzzy proposition), we assume t is larger than or equal to t(rup) [i.e., t(rup) is the real fuzzy truth value derived from r and a possibility distribution p], which means mt(t) is the upper bound of the possibility that r is true to a degree t. The fuzzy set is to represent the intended meaning of imprecise information, while the fuzzy truth value serves as the repre- sentation of uncertainty for its capability to express the pos- sibility of the degree of truth. An inference rule for truth-qualified fuzzy propositions has been expressed as follows: (r ` r ` ? ? ? ` r ) → q, t1 2 n 1 r 9, t1 2 r 9, t (2)2 3 ??? r 9, tn n 11 q 9, tn 12 where ri, (i = 1 ; n), q, and q9 are fuzzy propositions andr 9i characterized by ‘‘Xi is Fi,’’ ‘‘Xi is and ‘‘Y is G,’’ ‘‘Y isF 9’’i G9,’’ respectively; tj ( j = 1 ; n 1 2) are fuzzy valuations for truth values and defined by Fi and are the subsetsm (t). F 9t j i of Ui; G and G9 are the subsets of V. There are three major steps for deriving q9 and of (2):tn 12 • The fuzzy rules and fuzzy facts with fuzzy truth-values are transformed into a set of uncertain classical proposi- tions with necessity and possibility measures. • The possibilistic entailment is performed on the set of uncertain classical propositions. • We reverse the process in the first step to synthesize all the classical sets obtained in the second step into a fuzzy set, and to compose necessity and possibility pairs to form a fuzzy truth-value. 142 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 Hierarchical Fuzzy Petri Nets To increase the efficiency of rule based reasoning, two is- sues are particularly relevant: the possibility of exploiting con- currency, and the use of smart control strategies. To achieve these goals, the writers have proposed the use of fuzzy Petri nets to model fuzzy rule based reasoning (Lee et al. 1998). Furthermore, the explanation of how to reach conclusions is expressed through the movements of tokens in fuzzy Petri nets. There are several rationales behind basing a computational paradigm for expert systems on Petri net theory. First, Petri nets achieve the structuring of knowledge within rule bases, which can express the relationships among rules and help ex- perts construct and modify rule bases. Second, Petri nets’ graphic nature provides the visualization of dynamic behavior of rule based reasoning. Third, Petri nets make it easier to design an efficient reasoning algorithm. Fourth, Petri nets’ an- alytic capability provides a basis for developing a knowledge verification technique. Last, the underlying relationship of con- currency among rules activation can be modeled by Petri nets, which is an important aspect where real-time performance is crucial. Fuzzy Petri Nets: Definition A fuzzy Petri net FPN is defined as a quintuple: FPN = (FP, UT, A, W, M ) (3)0 where FP = {( p1, F1), ( p2, F2), . . . , ( pm, Fm)} is a finite set of fuzzy places, where pi represents a fuzzy condition and Fi is a fuzzy subset of Ui to represent the fuzzy set of the con- dition; UT = {(t1, t1), (t2, t2), . . . , (tn, tn)} is a finite set of uncertain transitions, where tj represents the causal relationship of fuzzy conditions and tj is a fuzzy truth value to represent the uncertainty about the causal relationship of fuzzy condi- tions; A is a set of arcs; W is a weight function; and M0 = {M( p1), M(p2), . . . , M( pm)} is the initial marking, where M( pi) is the number of tokens in pi. Each token is associated with a pair of fuzzy sets ( , ti)F9i (called an uncertain fuzzy token). Fuzzy places with uncertain fuzzy tokens can be interpreted as uncertain fuzzy facts related to the fuzzy conditions modeled by the fuzzy places. Fuzzy Rule Based Reasoning and Fuzzy Petri Nets It is widely recognized that fuzzy Petri nets are a promising modeling mechanism for formulating fuzzy rule based reason- ing. The three key components in fuzzy rule based reasoning — fuzzy propositions, fuzzy rules, and fuzzy facts — can be formulated as places, transitions, and tokens, respectively. The mapping between fuzzy rule based reasoning and fuzzy Petri nets has been made below. • Fuzzy places: Fuzzy places correspond to fuzzy proposi- tions. The fuzzy sets, attached to the fuzzy places, rep- resent the values of fuzzy propositions. Fuzzy input and fuzzy output places of a truth-qualified transition are used to represent the antecedent and conclusions parts of a truth-qualified fuzzy rule, respectively. • Uncertain fuzzy tokens: An uncertain fuzzy token repre- sents a truth-qualified fuzzy fact. The fuzzy sets and fuzzy truth-values are attached to uncertain fuzzy tokens to rep- resent the values and our confidence level about the ob- served facts, respectively. • Uncertain transitions: Uncertain transitions are classified into four types: inference, aggregation, duplication, and aggregation-duplication transitions. The inference transi- tions represent the truth-qualified fuzzy rules; the aggre- FIG. 1. Modeling Fuzzy Rule Based Reasoning through Fuzzy Petri Nets (a) before Firing Inference Transition ; (b) after Fir-it 1 ing it 1 gation transitions are designed to aggregate the conclusion parts of rules that have the same linguistic variables; the duplication transitions are used to duplicate uncertain fuzzy tokens to avoid the conflict problem; and the ag- gregation-duplication transitions link the fuzzy proposi- tions with the same linguistic variables. These have been formally defined below. Type 1: Inference Transition (ti). An inference transition serves as modeling of a truth-qualified fuzzy rule. A truth- qualified fuzzy rule having multiple antecedents is represented as (r ` r ` ? ? ? ` r ) → q, t (4)1 2 n 1 where ri and q are of the forms of ‘‘Xi is Fi’’ and ‘‘Y is G,’’ respectively. In Fig. 1, after firing the inference transition it ,1 the tokens will be removed from the input places of a newit ,l token will be deposited into the output place of and theit ,1 fuzzy set and the fuzzy truth value attached to the new token are derived by three steps: transformation, inference, and com- position. Type 2: Aggregation Transition (ta). An aggregation transition is used to aggregate the conclusions of several truth- qualified fuzzy rules that have the same linguistic variable, and to link the antecedent of a truth-qualified fuzzy rule that also has the same linguistic variable. For example, there are m truth-qualified fuzzy rules having the same linguistic variable in the conclusions, denoted as (r → q , t ), (r → q , t ), ? ? ? , (r → q , t ) (5)1 11 1 2 12 2 m 1m m where q1i is ‘‘Y is G1i.’’ In Fig. 2, after firing the aggregation transition the tokens in input places of will be re-a at , tm 11 m 11 moved, a new token will be deposited into the output place of and the fuzzy set and the fuzzy truth value attached toat ,m 11 J FIG. 3. Modeling Duplication of Uncertain Fuzzy Token through Fuzzy Petri Nets (a) before Firing Duplication Transition ; (b) after Firingd dt t1 1 FIG. 2. Modeling Aggregation of Conclusions by Aggregation Transition (a) before Firing ; (b) after Firinga at tm11 m11 the new token are derived by three steps: transformation, ag- gregation, and composition. Type 3: Duplication Transition (td). The purpose of du- plication transitions is to avoid the conflict by duplicating the token. For example, there are l truth-qualified fuzzy rules hav- ing the same linguistic variable in the antecedents, denoted as (r → q , t ), (r → q , t ), ? ? ? , (r → q , t ) (6)11 1 1 12 2 2 1 1l l where r1i means ‘‘Xi is F1i.’’ They are linked by a duplication transition shown in Fig. 3. After firing the duplication transi- OURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 143 FIG. 4. Modeling Aggregation-Duplication of Uncertain Fuzzy Token through FPN (a) before Firing Aggregation-Duplication Transition ; (b) after Firingad adt t1 1 tion the tokens in the input place of will be removed,d dt , t1 1 new tokens will be added into the output places of and thedt ,1 fuzzy sets and the fuzzy truth values attached to the new to- kens are not changed. Type 4: Aggregation-Duplication Transition (tad). An aggregation-duplication transition is a combination of an ag- gregation transition and a duplication transition. It is used to link all fuzzy propositions that have the same linguistic vari- able. For example, there are m truth-qualified fuzzy rules hav- 144 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 ing a same linguistic variable in the conclusions and l truth- qualified fuzzy rules having the same linguistic variable in the antecedents, denoted as (r → q , t ), (r → q , t ), ? ? ? , (r → q , t ) (7a)1 11 1 2 12 2 m 1m m (q → s , t ), q( → s , t ), ? ? ? ,1(m11) 1 m 11 1(m12) 2 m 12 (q → s , t )1(m11) 1 m 11 (7b) where si means ‘‘Z is Hi.’’ They are linked by an aggregation- duplication transition shown in Fig. 4. After firing the aggre- gation-duplication transition the tokens in the input placesadt ,1 of will be removed, new tokens will be deposited into theadt 1 output places of and the fuzzy sets and the fuzzy truthadt ,1 values attached to the new tokens are derived by three steps: transformation, aggregation, and composition. Hierarchical Fuzzy Petri Nets To overcome the complexity arising from large sizes of rule bases and fuzzy Petri nets, two important features, modular- ized rule bases and hierarchical fuzzy Petri nets, have been adopted in FPNES. Modularization is used to partition rule bases into smaller parts for organizing rules. In a hierarchical fuzzy Petri net, each hierarchy contains a fuzzy Petri net that may or may not contain other hierarchies. The connections between hierarchies are achieved by defining importing and exporting fuzzy places (Fig. 5). That is, an exporting fuzzy place with respect to a hierarchy is defined as a fuzzy place that is connected to the hierarchy by an arc from the fuzzy place to the hierarchy; an importing fuzzy place with respect to a hierarchy is defined as a fuzzy place connected to the hierarchy by an arc from the hierarchy to the fuzzy place. In a graphical representation, a hierarchy is drawn as a double- lined square to connect the importing or exporting fuzzy places. A hierarchical fuzzy Petri net that contains a main hi- erarchy H0 and hierarchy H1 is illustrated in Fig. 5(a). The status of the fuzzy place P1 in Fig. 5(a) is shown in Fig. 5(b). In this figure, the fuzzy Petri net in the middle window is the main hierarchy at the top level of the hierarchical structure, FIG. 5. (a) Hierarchical Fuzzy Petri Net; (b) Place Status for Pl in Ho and the fuzzy Petri net in the bottom window is hierarchy H1 at the second level. In H0, fuzzy places P2 and P4 are the exporting fuzzy place with respect to hierarchy H1, and fuzzy place P5 is the importing fuzzy places with respect to hier- archy H1. In the hierarchy H1, fuzzy places P1 and P3 are the importing fuzzy places with respect to H0, and fuzzy place P5 is the exporting fuzzing place with respect to H0. When a token is inserted into the fuzzy place P2 in H0, it will be transited into hierarchy H1 and added to place P1 in hierarchy H1. Similarly, once a token enters into place P4 in H0, it will be sent into hierarchy H1 and reaches the fuzzy place P3 in H1. After firing transition t1, t2, and t3 in hierarchy H1, the token arrives the fuzzy place P5 in H1 and then enters the fuzzy place P5 in H0. There are two main benefits by having a hierarchical struc- ture in the system: (1) the notion of hierarchy makes easy the handling of complex systems through decomposition; and (2) a hierarchical Petri net facilitates the reusability, that is, each hierarchy can be considered as a reuse unit. Transforming Modularized Fuzzy Rule Bases into Hierarchical Fuzzy Petri Nets The explanation of how to reach conclusions is expressed through the movement of tokens in hierarchical fuzzy Petri nets. The first step toward this goal is to transform modular- ized fuzzy rule bases into hierarchical fuzzy Petri nets. To bridge the gap between fuzzy rule based expert systems and fuzzy Petri nets, it is important to have a mechanism to au- tomatically transform modularized fuzzy rule bases into hier- archical fuzzy Petri nets. In our earlier paper (Lee et al. 1999), two algorithms have been involved in the transformation. One is to transform modularized fuzzy rule bases into a hierarchical incidence matrix. The other is to transform the hierarchical incidence matrix into a hierarchical fuzzy Petri net. Reasoning Mechanism Based on Fuzzy Petri Nets To consider the efficiency of fuzzy rule based reasoning, it is crucial that fuzzy facts (input or inferred) find the matched J fuzzy rules efficiently, rather than scanning through all the fuzzy rules. Fuzzy Petri nets offer an opportunity to achieve this goal by using transitions and arcs to connect fuzzy rules as a net based structure. A data-driven reasoning algorithm has been developed by the writers; for details about the reasoning algorithm, see Lee et al. (1998). Overview of FPNES Tool FPNES has been implemented in Java with a client-server architecture, encompassing four main parts: fuzzy Petri net system (FPNS), user interface, transformation engine, and knowledge bases (Fig. 6). Java has been adopted as the pro- gramming language for the FPNES tool for its capability of running on multiple platforms through the Internet. FPNS is a modeling and analysis tool for fuzzy Petri nets, and serves as an inference engine, and explanation facility in FPNES. FPNS mainly contains the simulator and analyzer for fuzzy Petri nets. It provides the basic constructs for hierarchi- cal fuzzy Petri nets (e.g., hierarchies, fuzzy places, uncertain transitions, arcs, and uncertain fuzzy tokens). After judging the firing conditions, the simulator will compute the fuzzy sets and move tokens. The analyzer performs the tasks of analyzing the properties of fuzzy Petri nets, such as incidence matrix, reach- ability trees, and state equations. Users can edit modularized fuzzy Petri rule bases in the client site, including the assignments of linguistic variables, truth-qualified fuzzy rules, the relationship of modules, and modularized structures of input facts. When users finish edit- ing the modularized fuzzy rule bases and the corresponding facts, and decide to run them, the data is then sent to the transformation engine and transformed to a hierarchical fuzzy Petri net in FPNS. After FPNS processes the hierarchical fuzzy Petri net with the aid of the reasoning algorithm, it sends the results back to users. The results are presented in a hierarchical fashion to provide a flexible explanation facility. DAMAGE ASSESSMENT To develop our modularized rule bases for damage assess- ment of bridges, we define the scope of knowledge domain, FIG. 6. Overview of FPNES Tool OURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 145 146 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 FIG. 8. Part of Rule Base for Crack in I-Girder FIG. 7. Factors Related to Cracking in Concrete describe the process of linguistic assessment, and discuss how to construct the modularized rule bases. Knowledge Domain The type of defects that could occur in a bridge depends upon the construction materials, environmental conditions, and external loadings. Nowadays, prestressed concrete bridges are widely used in Taiwan. However, some of them become im- plicit threats to the public because they have served for a long period of time and lots of defects are developing. To start with, the scope of the system is limited to (1) the concrete bridges that have prestressed concrete I-girders and hammerhead piers; and (2) the defects that occur most commonly and influence the load-carrying capacity. Linguistic Assessment In order to keep a bridge functioning, each component serves several functions to meet demands such as traffic load- ing, earthquakes, erosion, and environmental corrosion. Based on the observed defects, an expert can identify what kind of function is insufficient, judge the effect, and determine the damage level. The inference procedure for experts to assess bridge damage is described as follows: 1. A team of inspectors visually investigates each compo- nent of a bridge and records the observed defects, such as cracking, delamination, spalling, honeycomb, efflores- cence, corrosion, and movement. 2. Based on the defect symptoms, such as locations and patterns, experienced bridge engineers can identify the possible causes of the defects. 3. The possible causes can induce what kinds of functions are eliminated due to the defects. 4. The damage level of each defect is evaluated according to the quantitative description of its symptoms. 5. The levels of functional insufficiency are inferred based on the damage levels of the defects. 6. The assessment of damage can be estimated by aggre- gating both the functional insufficiency and its levels. The expression of observed defects by inspectors is viewed as the input of the system. Defects affecting the load-carrying capacity of bridges include (1) concrete discontinuities caused by overloadings, such as shear or flexure cracks, delamina- tions, and spalls; (2) component displacements such as rota- tional movement and settlement; and (3) soil scours due to rushing flood. Various patterns and locations of the defects are used to determine the possible causes and effects on the func- tionality, and the magnitude of each defect relates to the level of damage. The level of damage, severity of defects, and intensity of confidence about descriptions are expressed linguistically and are considered as linguistic variables. The severity of damage in terms of linguistic variables is classified into seven fuzzy levels, such as (1) very severe [i.e., m(x) = x8]; (2) severe [i.e., m(x) = x4]; (3) fairly severe [i.e., m(x) = x2]; (4) fair [i.e., m(x) = x]; (5) fairly slight [i.e., m(x) = x ]; (6) slight [i.e., m(x) =1/2 ]; and (7) very slight [i.e., m(x) = x ], where x is a real1/4 1/8x number between zero and one, and denotes the degree of dam- age. The intensity of confidence about rules or facts is also considered in terms of linguistic variables with values such as (1) very true [i.e., m(t) = t2]; (2) true [i.e., m(t) = t]; (3) fairly true [i.e., m(t) = t ]; (4) fairly false [i.e., m(t) = (1 2 t) ];1/2 1/2 (5) false [i.e., m(t) = 1 2 t]; (6) very false [i.e., m(t) = (1 2 t)2]; and (7) unknown [i.e., m(t) = 1], where t is a real number between zero and one, and denotes the degree of truth. Modularized Rule Bases Crack, delamination, spalling, movement, and scour are five main defects in our system. Cracking is the most common defect and can be discovered in any kind of concrete structure. Its location and pattern are used to determine whether a crack affects the load-carrying capacity. The level of severity of a crack should not be determined directly without full consid- eration of the factors involved. the factors could include but are not necessarily limited to those shown in Fig. 7, and the related fuzzy rules of module ‘‘crack in I-girder’’ are presented in Fig. 8. Delamination is separation along a plane nearly par- allel to the surface of the concrete. Spalling is a depression resulting from the dislodgment surface of concrete. It is typi- cally found in old concrete structures and not in young con- crete. Before developing into a depression, spalling may be presented as a delamination below the concrete surface. Some- times it is as deep as the reinforcing steel. Vertical movement such as differential settlement can pro- duce serious distress in a bridge structure. Rotational move- ment of substructure can be considered to be the result of unsymmetrical settlement or lateral movements. Piers, abut- ments, and walls may undergo this type of movement. The degree of damage depends on the extent of vertical and rota- tional movement. The recommendation of strengthening or re- pairing is based on the possible cause. Scour of the substruc- ture support is one of the more frequent factors that may cause or lead to structural failure or foundation distress. Scour is the removal of the streambed, backfill, slopes, or other supporting material, by stream tidal action, dredging, propeller backwash, etc. The degree of damage depends on the extent of exposure of foundation and the extent of deterioration of foundation. FIG. 9. Fuzzy Rule Petri Net Transformed from Rule Base of Crack in I-Girder The whole rule base is subdivided into a lot of modules in order to overcome the complexity arising from too large sizes of the rule bases and to make the rules well organized. At present there are 34 modules in this system, in which 216 truth-qualified fuzzy rules and 224 recommendations are in- cluded. Seven modules related to the cracking in concrete, five modules dealing with the delamination in concrete, five mod- ules concerning the spalling in concrete, two modules about the movement of foundation, and two modules handling to the scour of support are constructed. Furthermore, two modules are used to evaluate the insufficiency of shear or flexure re- sistance, and 11 modules are used to assess the damage levels of components; 216 truth-qualified fuzzy rules and 224 rec- ommendations are accommodated in these modules. As was mentioned previously, fuzzy Petri nets’ graphical representa- tion can help us construct and modify fuzzy rule bases. Fig. 9 shows the fuzzy Petri nets after the transformation of the rule bases of a crack in an I-girder. EXAMPLE To demonstrate the use of FPNES, the Da-Shi bridge lo- cated in northern Taiwan was considered. It was rebuilt in 1960 as a simply supported, 12-spanned bridge, 550 m long and 7.8 m wide, across the Da-Han river. This bridge consists of 12 decks, 36 prestressed concrete I-girders, 120 diaphragms, 11 piers, and two abutments. In 1997, this bridge was inspected by the Center of Bridge Engineering Research at National Central University. Visual inspection revealed that many minor cracks accompanied with efflorescence spread over eight panels within deck 7. The PCI girders S9G1, S9G2, S9G3 in span 9 and S10G1, S10G2, S10G3 in span 10 had severe flexure, shear cracks, and some spalls. There were two diaphragms where several spalls were found. The foundations of piers P1 to P6 were exposed up to 5 m above the ground level, and some of them suffered severe spalling. Furthermore, pier 2 had rotational movement. After executing FPNES for damage assessment of the Da- Shi bridge, the hierarchical fuzzy Petri nets were constructed based on the modularized rule bases, and uncertain fuzzy to- J kens were transformed into these nets in order to fire transi- tions and perform the reasoning mechanism. The results of damage assessment using FPNES for the Da-Shi bridge were expressed in a hierarchical fashion to serve as an explanation mechanism to facilitate the retrieval of detailed information on damaged components from the top down to lower levels (Fig. 10). The recommendations embedded in rule bases were also provided on an if-needed basis. As a result, the damage to the Da-Shi bridge was evaluated to be severe with strong confi- dence (i.e., very true) since the overall damage to the super- structure was severe with common confidence (i.e., true) and the overall damage to the substructure was also severe with strong confidence (i.e., very true). The severity of superstruc- ture results from the overall damage to decks was fairly slight with fair confidence (i.e., fairly true), the overall damage to I- girders was severe with common confidence (i.e., true), and the overall damage to diaphragms was very slight with strong confidence (i.e., very true). Meanwhile, the severity of sub- structure as derived from the overall damage of piers was se- vere with strong confidence (i.e., very true). The FPNES re- sults not only match the experts’ judgments, but also provide more information than experts do, because the explanation provided in the system and the confidence level associated with the conclusions can be used to justify whether to take the recommendations into account. The following recommendations were made with strong confidence (i.e., very true). The observed defects or deficien- cies may affect the load-carrying capacity of the bridge. For the safety of the traveling public, traffic on the bridge should be limited. A second inspection and load capacity evaluation should be made. The second inspection is used to identify the compressive strength of concrete, the carbonation depth in concrete, the content of chloride in concrete, the corrosion of reinforcement, the crack depth in concrete, etc. The load ca- pacity evaluation is performed to ascertain the safety load ca- pacity for the current condition of the bridge. Based on the results of the load capacity evaluation, a plan can be made for rehabilitating, strengthening, or replacing the bridge. OURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 147 FIG. 10. Results of Damage Assessment for Da-Shi Bridge Using FPNES CONCLUSION This research has resulted in the development of a fuzzy Petri net based expert system (FPNES) for bridge damage as- sessment, which contains a reasoning mechanism to deal with uncertain and imprecise information, a fuzzy Petri nets ap- proach to modeling fuzzy rule based reasoning, and a tool supporting the damage assessment of bridges. Unlike other researchers’ systems, our expert system does not impose any restriction on the inference mechanism, that is, the intended meaning is not required to be intact; meanwhile, the confi- dence level can be partially certain. Furthermore, it offers more informative results because the explanation provided in the system and the confidence level of the conclusions can be used to justify whether to take the recommendations into account. An application to the damage assessment of the Da-Shi bridge in Taiwan is used as an illustrative example of FPNES. The FPNES results not only match the experts’ judgments, but also provide more information than experts do. Future research will consider issues related to damage as- sessments that remain to be addressed further: (1) how to en- hance FPNES to involve numerical data such as earthquake records, experimental results, and load testing information; and (2) how to integrate the numerical data and linguistic assess- ment into an overall evaluation. ACKNOWLEDGMENTS The writers would like to thank Sam Lin for his early work on this project. This research is partially sponsored by the National Science Council (Taiwan) under grants NSC 86-2213-E-008-008 and NSC 86- 2211-E-008-015. APPENDIX I. REFERENCES Dubois, D., and Prade, H. (1988). Possibility theory: An approach to computerized processing of uncertainty. Plenum, New York. Eriksson, H. (1996). ‘‘Expert systems as knowledge servers.’’ IEEE Expert, 13(3), 14–19. Furuta, H., He, J., and Watanabe, E. (1996). ‘‘A fuzzy expert system for 148 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 damage assessment using genetic algorithms and neural networks.’’ Microcomputers in Civ. Engrg., 11(1), 37–45. Furuta, H., Shiraishi, N., Umano, M., and Kawakami, K. (1991). ‘‘Knowl- edge-based expert system for damage assessment based on fuzzy rea- soning.’’ Comp. and Struct., 40(1), 137–142. Hadipriono, F. C., and Ross, T. J. (1991). ‘‘A rule-based fuzzy logic deduction technique for damage assessment of protective structure.’’ Fuzzy Sets and Sys., 44(3), 459–468. Ishizuka, M., Fu, K. S., and Yao, J. P. T. (1982). ‘‘Inference procedures under uncertainty for the problem-related method.’’ Information Sci., 28(3), 179–206. Krauthammer, T., Muralidharan, R., and Schimdt, W. (1992). ‘‘Combined symbolic-numeric explosion damage assessment for structures.’’ J. Comp. in Civ. Engrg., ASCE, 6(4), 417–434. Kushida, M., Miyamoto, A., and Kinoshita, K. (1997). ‘‘Development of concrete bridge rating prototype expert system with machine learning.’’ J. Comp. in Civ. Engrg., ASCE, 11(4), 238–247. Lee, J., Liu, K. F. R., and Chiang, W. (1998). ‘‘Fuzzy Petri nets for modeling rule-based reasoning.’’ Int. J. Artificial Intelligence Tools, 7(4), 463–485. Lee, J., Liu, K. F. R., and Chiang, W. (1999). ‘‘A fuzzy Petri nets based expert system for damage assessment of bridges.’’ IEEE Trans. on Sys., Man., and Cybernetics, Part B: Cybernetics, 29(3), 350–370. Miyamoto, A., Morikawa, H., Kushida, M., and Tokuyama, T. (1993). ‘‘A knowledge-based expert system application in structural safety as- sessment of concrete bridges.’’ Bridge management: Inspection, main- tenance, assessment and repair 2, J. E. Harding, G. E. R. Parke, and M. J. Ryall, eds., E & FN Spon, London, 96–109. Ogawa, H., Fu, K. S., and Yao, J. P. T. (1985a). ‘‘An inexact inference for damage assessment of existing structures.’’ Int. J. Man-Machine Studies, 22(2), 295–306. Ogawa, H., Fu, K. S., and Yao, J. P. T. (1985b). ‘‘SPERIL-II: An expert system for damage assessment of existing structures.’’ Approximate reasoning in expert systems, M. M. Gupta, A. Kadel, W. Bandler, and J. B. Kiszka, eds., Elsevier Science, Amsterdam, 731–744. Ross, T. J., Sorensen, C., Savage, J., and Carson, J. M. (1990). ‘‘DAPS: Expert system for damage assessment.’’ J. Comp. in Civ. Engrg., ASCE, 4(4), 327–348. Shiraishi, N., Furuta, H., Umano, M., and Kawakami, K. (1991). ‘‘An expert system for damage assessment of a reinforced concrete bridge deck.’’ Fuzzy Sets and Sys., 44(3), 449–457. Yao, J. P. T. (1980). ‘‘Damage assessment of existing structures.’’ J. Comp. in Civ. Engrg., ASCE 106(4), 785–799. Zadeh, L. A. (1975). ‘‘The concepts of a linguistic variable and its ap- plication to approximate reasoning (Part 1).’’ Information Sci., August, 199–249. Zadeh, L. A. (1976). ‘‘The concepts of a linguistic variable and its ap- plication to approximate reasoning (Part 2).’’ Information Sci., Septem- ber, 43 – 80. APPENDIX II. NOTATION The following symbols are used in this paper: A = set of arcs; F, G = fuzzy set; FP = finite set of fuzzy places; FPN = fuzzy Petri net; H = hierarchy; M0 = initial marking; M( pi) = number of tokens in pi; J p = fuzzy place; pi = fuzzy condition; q, r, s = fuzzy proposition; t = truth degree; tj = causal relationship of fuzzy conditions; ta = aggregaton transition; tad = aggregation-duplication transition; td = duplication transition; ti = inference transition; U, V = universes of discourse; UT = finite set of uncertain transitions; W = weight function; X, Y, Z = linguistic variable; m(x) = membership function; p = possibility distribution; and t = fuzzy truth value. OURNAL OF COMPUTING IN CIVIL ENGINEERING / APRIL 2000 / 149 
work_gwxxyoytejc6pm2frh5kmaj5gm	----	NASA Technical Reports Server (NTRS) 
work_gxb3jgybnreolk7frlfntepowa	----	PII: S0957-4174(03)00067-8 The collaborative filtering recommendation based on SOM cluster-indexing CBR Tae Hyup Roh a,*, Kyong Joo Oh b , Ingoo Han a a Graduate School of Management, Korea Advanced Institute of Science and Technology, 207-43 Cheongryangri-Dong, Dongdaemun-Gu, Seoul 130012, South Korea b Department of Business Administration, Hansung University, Seoul, South Korea Abstract Collaborative filtering (CF) recommendation is a knowledge sharing technology for distribution of opinions and facilitating contacts in network society between people with similar interests. The main concerns of the CF algorithm are about prediction accuracy, speed of response time, problem of data sparsity, and scalability. In general, the efforts of improving prediction algorithms and lessening response time are decoupled. We propose a three-step CF recommendation model, which is composed of profiling, inferring, and predicting steps while considering prediction accuracy and computing speed simultaneously. This model combines a CF algorithm with two machine learning processes, Self-Organizing Map (SOM) and Case Based Reasoning (CBR) by changing an unsupervized clustering problem into a supervized user preference reasoning problem, which is a novel approach for the CF recommendation field. This paper demonstrates the utility of the CF recommendation based on SOM cluster-indexing CBR with validation against control algorithms through an open dataset of user preference. q 2003 Elsevier Ltd. All rights reserved. Keywords: Collaborative filtering; Recommendation system; Self-organizing map; Case-based reasoning 1. Introduction The advent of the network world induced by the rapid development of the Internet and the accompanying adoption of the Web has promoted the chances to create greater business opportunities and to reach customers more easily. This 24 £ 7 on-line accessibility has resulted in the enlargement of choices, but customers are faced with information overload. Their own arduous efforts are required to retrieve information that matches their prefer- ences. It is an automated and sophisticated decision support system that is needed to suggest personalized information in a brief form without going through an annoying process. Collaborative filtering (CF) recommendation is a knowl- edge sharing technology for distribution of opinions and facilitating contacts in network society between people with similar interests. The CF recommendation is the process of multiple users sharing information on the preferences and actions of an affinity group tracked by a system which, then, tries to make useful recommendations to individual users based on the patterns it predicts (Herlocker, Konstan, Borchers, & Riedl, 1999; Kumar, Raghavan, Rajagopalan, & Tomkins, 1998). CF recommendation also provides a complementary tool for information retrieval systems that facilitates users’ navigation in a meaningful and personal- ized way. Most content retrieval methodologies use some type of a similarity score to match a query describing the content with key words, the individual titles or items, and then present the user with a ranked list of suggestions. However, conventional CF does not use any actual content (e.g. words, description, URL, etc.) of the items. It is rather based on preference ratings information to match users with similar interests together and to predict a user’s rating for an unseen item by examining his/her community’s rating for that item. The CF recommendation systems are built on the assumption that a good way to find interesting content is to find other people who have similar interests and then recommend items that those similar users like (Breese, Heckerman, & Kadie, 1998). Most research on recommendation systems can be divided into three categories: technical system development research, user behaviour and reaction research, and privacy issues. Our focus is on technical system development 0957-4174/03/$ - see front matter q 2003 Elsevier Ltd. All rights reserved. doi:10.1016/S0957-4174(03)00067-8 Expert Systems with Applications 25 (2003) 413–423 www.elsevier.com/locate/eswa * Corresponding author. Tel./fax: þ82-29583685. E-mail address: rohth@kgsm.kaist.ac.kr (T.H. Roh). http://www.elsevier.com/locate/eswa research, especially, the design and analysis of an algorithm for CF recommendation. As the number of users and items increases and the contents of each user’s preference to the items changes, typical CF recommendation needs exponen- tially growing computation time for finding an affinity group and predicting each user’s unknown preferences (Cho, Kim, & Kim, 2002; Claypool et al., 1999). We find the potential for improving prediction accuracy and efficiency simul- taneously by separating on- and off-line steps using recent clustering and reasoning machine learning techniques. This study presents a three-step CF model which is composed of SOM profiling, CBR inferring and CF predicting step. The SOM network has been studied as one of the most popular unsupervized neural network models for clustering and visualization in a number of real-world problems (Kohonen, Hynninen, Kangas, & Laaksonen, 1996). The CBR is well known as it benefits from the case-specific knowledge of past problems to find solutions to the new problems (Kim & Han, 2001). These two outperforming machine learning methods can be combined for CF and increase the accuracy and efficiency in the recommendation process. The rest of this paper is organized as follows. Section 2 provides a brief overview of CF models for several recommendation techniques and issues with an emphasis on algorithmic features shown by previous research. Details of the proposed CF model are provided in Section 3. Section 4 describes the dataset, evaluation metrics and experimental design. Experiments are run on an open dataset associated with the MovieLens preference rating dataset, six exper- imental protocols and two evaluation metrics for the various algorithms. Results are shown in Section 5 and conclusion are presented in Section 6. 2. Background 2.1. Collaborative filtering recommendation The fundamental function of CF is to predict the preferences of one user, referred to as the ‘active user’. The problem space can be formulated as a matrix of users versus items, with each cell representing a user’s rating on a specific item. Let I be the whole set of items, Ihð, IÞ be the subset that has been rated by the active user ðUaÞ; and Ir ¼ ðI > ICh Þ be the subset that has not been rated by the Ua: CF systems estimate Ua‘s preferences for items in Ir based on the overlap between his/her preference ratings for items in Ih and those of the other users. The key advantage of CF is that it does not consider the content of the items being recommended, but human determine the relevance, quality, and interest of an items in the information stream. As a result, filtering can be performed on items that are hard to analyze with computers, such as Multimedia Component, ideas, feelings, people, and so on. Rather than mapping users to items through ‘content attributes’ or ‘demographics’, CF treats each item and user individually. Accordingly, it becomes possible to discover new items of interest simply because other people like them. At the same time, CF’s dependence on human ratings can be a drawback. For a CF system to work well, several users must evaluate each item; even then, new items cannot be recommended until some users have taken the time to evaluate them. These limitations, often referred to as ‘data sparsity’ and ‘cold start problem’, cause trouble for users seeking obscure items (since nobody may have rated them) or advice on new items (since nobody has had a chance to evaluate them) (Good et al., 1999). CF related research starts from the Tapestry system, out of Xerox (Goldberg, Nichols, Oki, & Terry, 1992) which coined the term ‘collaborative filtering’ in the context of a system for filtering email using binary category flags. Tapestry was a full-featured filtering system for electronic documents—primarily electronic mail and Usenet postings. The GroupLens is a pioneering and ongoing effort in CF (Good et al., 1999; Herlocker et al., 1999; Konstan et al., 1997; Resnick, Iacovou, Sushak, Bergstrom, & Riedl, 1994; Schafer, Konstan, & Riedl, 2001). The GroupLens team initially implemented a neighbourhood-based CF system for rating Usenet articles. Several similar systems were developed around the same time as the GroupLens Usenet system, including the Ringo music recommender which used a number of measures of distance between users, including Pearson correlation, constrained Pearson correlation, vector cosine (Shardanand & Maes, 1995), and the Bellcore Video Recommender (Hill, Stead, Rosenstein, & Furnas, 1995). These three research systems used what have come to be called neighbourhood-based prediction algorithms. Due to their speed, flexibility, and understandability, neighbour- hood-based prediction algorithms are currently one of the most effective ways to compute predictions in CF. Breese et al. (1998) identify two major classes of CF prediction algorithms; memory-based CF and model-based CF. Memory-based algorithms operate over the entire user database to make predictions. The most common memory- based models are based on the notion of nearest neighbours, using a variety of distance measures. Model-based systems are based on a compact model inferred from the data. They compare a number of algorithms including Bayesian clustering, decision-tree modelling, and also show that neighbourhood-based CF performed better than Bayesian belief networks for non- binary domains. Bayesian network and correlation models are the best-performing if computational complexity is not taken into account. In this framework, our SOM cluster- indexing CBR CF predictor model would be considered model-based CF. More recently, a number of machine learning tech- niques and hybrid filtering techniques have been chal- lenged. Hybrid filtering models combine recommendations from multiple sources which include the content of the item or page, the ratings of users, content-based filtering, T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423414 and demographic information and so on. Balabanovı́c and Shoham (1997) apply ‘Selection agent’, which decides the recommendation algorithm between content-based filtering and CF. Pazzani (1999) shows the hybrid approach for recommendation that uses more of the available infor- mation and consequently has more precise recommen- dations. The strengths of the different approaches can be complementary. Basu, Hirsh, and Cohen (1998) present an inductive learning approach to recommendation that is able to use both ratings information and other forms of information about each item in predicting user prefer- ences. Delgado and Ishii (1999) suggest a weighted- majority rating approach and Pennock, Horvitz, Lawrence, and Giles (2000) suggest a hybrid memory and model based approach for personality diagnosis that computes the preference probability with the same personality grouping. However, these efforts of improving prediction algorithms are decoupled from computational complexity and response time issues. In this paper, we introduce a computational machine learning CF model which comprises off-line learning part and on-line preference predicting part. This model focuses on accuracy and efficiency simultaneously by lessening on- line computation complexity with SOM cluster-indexing CBR process. Since we adapt the dense user-item matrix using the reference data set induced by SOM clusters’ centroid value of each item, the correlation matrix is directly computed and then the active user’s preference to item is predicted. 2.2. Self-organizing map The SOM network, known as competitive learning or self-organization, has been shown as one of the most popular unsupervized competitive neural network learning models, for clustering and visualization in a number of real- world problems (Kohonen et al., 1996). It is capable of mapping high-dimensional similar input data into clusters close to each other. It has two-layer, fully connected networks with a weight matrix. Sometimes, SOM called ‘topology-preserving maps’, assumes a topological structure among the cluster units. A topological map is simply a mapping that preserves neighbourhood relations and performs a topology-preserving projection from the data space onto a regular two-dimensional grid. The resulting maps provide users an intuitive and familiar way of correlating and illustrating input data sets. Furthermore, SOM can be used for clustering, classification, and modelling. The versatile properties of SOM make it a valuable tool in data mining. Relating to the clustering capability of SOM, Mangiameli, Chen, and West (1996) demonstrate that it is a better clustering algorithm than hierarchical clustering with overlapped dispersion, irrele- vant variables, outliers or different sized populations. For that reason, SOM has been adapted as an analytical tool in various marketing domains including database marketing (Ha & Park, 1998), segmentation of on-line markets (Vellido, Lisboa, & Meehan, 1999) and automatic labelling of customer clusters (Yuan & Chang, 2001). In this suggested CF recommendation model, we focus on the clustering capability of SOM. Given a set of users’ preference patterns to items, X; the algorithm returns a prototype (a set of cluster’s centroid values) yi for each cluster i: The prototypes are sometimes called neurons. The number of clusters, M; is a parameter that must be provided a priori. In the algorithm, first each prototype yp is randomly initialized (line 4). In the main loop (lines 5 – 10), one randomly selects an element x [ X and determines the neuron yp that is nearest to x: In the inner loop (lines 8, 9), one considers all neurons y that are within a neighbourhood NðypÞ of yp; including yp; and updates them according to the formula in line 8. The effect of neuron updating is to move neuron y closer to pattern x: The degree by which y is moved towards x is controlled by the parameter g; which is called the learning rate. It has to be noted that gis dependent on the distance between y and yp; i.e. if neuron y [ NðypÞ has a smaller distance to yp than neuron y0 [ NðypÞ then y is moved towards x by a larger degree than neuron y0: After iterations through the repeat-loop, the learning rate g is reduced by a small amount, thus facilitating convergence of the algorithm. It can be expected that after a sufficient number of iterations the yi’s has moved into areas where many xj’s are concentrated. Hence each yi can be regarded as a cluster’s centroid value which is used for the next CBR process as the ‘cluster-indexed reference set’. The Pseudo code description is shown in Fig. 1. In spite of several excellent applications, SOM has some limitations that hinder its performance. The typical limitations and the settlements are due to the vulnerability of convergence along a number of cluster and weight initialization, network size, and stopping rule conditions (Kim & Han, 2001). To determine the number of cluster, we adapt the visualization techniques by use of principle component analysis (PCA). PCA was first introduced in 1901 by Pearson and Hotelling generalized it to random variables in 1933. The idea is to keep only the ‘principal’ eigenvectors (components). The number of eigenvectors to retain depends on the variances (eigenvalues) but is typically small. If v eigenvectors are retained, data are Fig. 1. Pseudo code description of self-organizing map. T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423 415 projected along the first n principal eigenvectors. In this study, PCA facilitates dimensionality reduction for off-line clustering of user and rapid online cluster assignment, and also users are projected onto the ‘eigen-plane’ in 2 or 3D scatter plot for visualization. 2.3. Case based reasoning CBR is a methodology for building an analogy process, one way of human reasoning, which is the inference that a certain resemblance implies further similarity. It makes direct use of past experiences or cases to solve a new problem by recognizing its similarity with a specific known problem and by applying to find a solution for the current situation (Chiu, 2002; Choy, Lee, & Lo, 2002). CBR applications can be broadly used for two main problem types: classification and synthesis tasks. The classification task is to match case against those in the case base to determine what type, or class, of case it is, and then the solution from the best matching case is reused. Synthesis task attempts to create a new solution by combining parts of previous solutions. CBR systems that perform synthesis tasks must make use of adaptation and are usually hybrid systems combining with other techniques. Its main advantages over other techniques are as follows: in the CBR system, most knowledge is acquired in the case base and so it reduces the knowledge acquisition effort. That is, it makes use of existing case database, so it requires less general knowledge which is very difficult to get. Second, it requires less maintenance effort. Since rule bases or models should consider many dependencies between rules and effects of changes of the rule base are hard to predict, it is difficult to maintain. However, case bases are easier to maintain, because cases are independent of each other, domain experts and novices understand cases quite easily and maintenance of the CBR system can be done by adding/deleting cases. CBR algorithms have been used for marketing decision making processes. Hui, Fong, and Jha (2001) present the hybrid CBR – ANN approach that integrates ANN with the CBR cycle to extract knowledge from service records for the web customer service. Choy et al. (2002) apply CBR to integrate customer relationship management (CRM) and supplier relationship management (SRM) for facilitating supply chain management of supplier selection. Chiu (2002) suggests a case based customer classification approach for direct marketing, which combines Genetic Algorithm and the CBR process. The traditional process involved in CBR can be represented by a schematic cycle as shown in Fig. 2. Aamodt and Plaza (1994), Bradley (1994) describe CBR as a cyclical process: representation, retrieval, reuse, revision, and retainment. Case retrieval searches the case base to select existing cases sharing significant features with the new case. Through the retrieval step, similar cases that are potentially useful to the current problem are retrieved from the case base. The computing of the degree of similarity between the input and the target case can usually be calculated using various similarity functions among which nearest-neighbour matching is one of the frequently used methods. Nearest- neighbour matching is a quite direct method that uses a numerical function to compute the degree of similarity. Usually, cases with higher degrees of similarity are retrieved. A typical numerical function is shown in the following formula (Kolodner, 1993). X n i¼1 Wi £ simðf I i ; f R i ÞX n i¼1 Wi where Wi is the weight of the ith feature, f I i is the value of the ith feature for the input case, f Ri is the value of the ith feature for the retrieved case, and sim ( ) is the similarity function for f Ii and f R i : In our suggested CF model, CBR process provides classification and synthesis with additional generalized knowledge derived from the users’ explicit preference patterns. Generalized knowledge can be acquired by the centroid values of clusters obtained using clustering techniques, which are added to the case base as representative cases and then used as a case indexing scheme in order to retrieve more relevant cases. The cluster-indexing approach assumes that there are some different subgroups (clusters) in each rated group. The centroid values of clusters are new artificial cases that extract the information from the whole case base and represent each clustered case. 3. SOM cluster-indexing CBR CF recommendation This study utilizes the outperforming SOM as a clustering tool and the strength of CBR as assistance to index and retrieve like-minded users. The point of this study is to support the usage of the SOM for finding clusters with centroid value of items and CBR for indexing and retrieval in the CF recommendation process. The centroid values of clusters are the values of weight vectors that are the interim results from SOM learning processes, and these are standardized due to difference of Fig. 2. CBR process as a schematic cycle comprising the five ‘Re’s. T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423416 rating scale. The standardized centroid values have the same representation scheme as raw users’ rating value in spite of being learned artificial cases. These values represent clustered users of the entire user-base, and they are used as an indexing tool for each user. These standardized centroid values are consistent if learning is implemented again with the same parameters, although the addition of new users modify the standardized centroid values just a little. The definite process of the cluster-indexing method is composed of three steps: profiling, inferring and predicting steps. In the profiling step operated in the back-office, PCA and preliminary SOM testing are performed to fit the stable cluster condition. Clusters are derived from the dense subset of user-item rating DB. All the training users are indexed by the SOM process in accordance with the similarity to the centroid values of each cluster. While the inferring step, CBR compares an active user with the centroid values. The most similar cluster is inferred from reference users indexed within the selected cluster. After the inferring step, preference prediction is performed on-line with correlation based CF between an active user and reference users of a selected cluster. Fig. 3 depicts the proposed model architecture. 3.1. Profiling step In the profiling step, PCA is used for visualizing users’ patterns and reducing the dimension of input items before SOM clustering. Clusters are then derived from the SOM process using reference user DB with dense preference rating data. The reference users are indexed by the cluster and comprise the cluster-indexing reference DB. Step 1. User clustering with the centroid values of clusters by the SOM. 1.1 Explore users’ distribution by PCA. 1.2 Determine number of clusters using PCA factors. 1.3 Initialize weight vectors of the SOM. 1.4 Find clusters and the standardized centroid values of clusters. 3.2. Inferring step When the item preference prediction of an active user is requested, an active user’s rating information is compared with the standardized centroid values of each cluster. Through indexing and retrieval part of the CBR process, the most similar cluster is determined and retrieved. Step 2. Active user indexing and retrieval with CBR. 2.1. Index an active user with the centroid values of clusters having minimum distance calculated by the k-Nearest Neighbour method, which is Min_D ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Xn m¼1 l va;m S 2 Cref;ml 2 vuut where m is the given item, ni;m is the active user’s rating value to item m; S is the standardizing factor for rating scale, and Cref;m represents the centroid values to item m of the fixed clusters. 2.2. Retrieve the neighbours that were indexed in the same cluster. 3.3. Predicting step The active user’s predicted preference value to the target item is calculated by a Pearson-correlation based filtering formula in the on-line prediction part. Step 3. Prediction of active user’s predicted preference value. 3.1. Calculate Pearson correlation between the active user and the most similar neighbours that were indexed in the same cluster wða; iÞ ¼ X j ðva;j 2 maÞ sa ðvi;j 2 miÞ si where wða; iÞ is the Pearson correlation coefficient to compute the weight for each user’s contribution that is indexed in the same cluster. 3.2. Compute the prediction Pa;t of active user ðUaÞ on target item It: Pa;t ¼ ma þ k X i–a wða; iÞðvi;t 2 miÞ where the sum is same cluster-indexing users in the reference DB, ni;m is the rating cast by the other user i on item t; and ma is Uas mean rating. The constant k in front of the sum is an appropriate normalization factor. The process of the cluster-indexing method is exempli- fied in Fig. 4 If an active user’s preference rating {5, _, 2, 2, …,?,…, 4} is given, at first, it is indexed as the most similar cluster C2 using the cluster-indexing DB, which is clustered by SOM in an off-line learning process. Next, this method retrieves the nearest neighbour users out of that Fig. 3. Model architecture of SOM cluster-indexing CBR CF recommendation. T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423 417 cluster-indexing user group (C2). Finally, the predictive value of the active user’s target item is calculated by use of the prediction formula. 4. Experiments Experiments are run on open datasets, six different experimental predictors, and two evaluation metrics for the various algorithms. Results are compared with base-line models and other comparative models in terms of the level of data sparsity and machine learning techniques. 4.1. Data: MovieLens dataset MovieLens data sets were collected by the GroupLens Research Project at the University of Minnesota (http://www.cs.umn.edu/Research/GroupLeans/data). The- Fig. 4. Computation example of SOM cluster-indexing CBR CF model. T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423418 http://www.cs.umn.edu/Research/GroupLeans/data historical dataset consists of 100,000 ratings from 943 users on 1682 movies with every user having at least 20 ratings and simple demographic information for the users (age, gender, occupation, zip code) is included. The ratings are on a numeric five-point scale with 1 and 2 representing negative ratings, 4 and 5 representing positive ratings, and 3 indicating ambivalence. We sample a reference user set that has enough rating information to discover similar user patterns. The number of users that are extracted as the reference sample is 251. Among 100,000 rating records, 6093 were generated from 251 users rating data to 33 items out of 10 movie genres. 4.2. Evaluation metrics Recommender systems researchers use several different measures for the quality of recommendations produced: statistical accuracy, decision-support metrics, coverage measures. We apply two metrics to our evaluation: normalized mean absolute error (NMAE) and the receiver operating characteristic (ROC) curve including the area under ROC curve, which are used by Goldberg, Roeder, Gupta, and Perkins (2001) and Good et al. (1999). NMAE. We look at the average absolute deviation of the predicted rating to the actual rating on items the users in the test set have actually voted on. That is, if the number of predicted ratings in the test set is in the active case, ma; then the average absolute deviation for an active user is MAE ¼ 1 ma Xma j¼1 lPa;j 2 va;jl Since our numerical rating scale gives ratings over the range [1, 5], we normalize to express errors as percentages of full scale: NMAE is NMAE ¼ MAE rmax 2 rmin ROC curve and area under ROC curve. This metric evaluates the performance of classification scheme in which there is one variable with two categories by which subjects are classified. ROC sensitivity is a signal proces- sing measure of the decision making power of a filtering system. Operationally, it is the area under the ROC—a curve that plots the sensitivity vs. 1—specificity of the test (Swets, 1988). Sensitivity refers to the probability of a randomly selected good item being accepted by the filter. Specificity is the probability of a randomly selected bad item being rejected by the filter. Points on the ROC curve represent trade-offs supported by the filter. The ROC sensitivity ranges from 0 to 1 where 1 is perfect and 0.5 is random. To test the difference of six protocols’ performance, we use one-way ANOVA with a post hoc test—Bonferroni procedure in the equal variance assumed for multiple comparisons statistics of MAE (Breese et al., 1998). In this study, the post hoc test procedure is used for investigating the differences between specific experimental protocols in conjunction with ANOVA comparing each protocol’s MAE difference. 4.3. Experimental setup At first, we build the dataset into three experimental sets for testing the effect of the available information level (the number of items on which active users have rated). In the first set, named Allbut1, we withhold one selected item for each user in the test set, and try to predict its value given all the other ratings the user has voted on. In the second and third set of experiments, we select five and 10 ratings from each test user as the observed ratings, and then attempt to predict those preference levels. We call these Given5, and Given10. The Allbut1 experiments measure the algorithms’ performance when given as much data as possible from each test user. The Given experiments look at users with less data available, and examine the performance of the algorithms when there is relatively little known about an active user. Secondly, we present metrics derived from empirical analysis of the proposed SOM cluster-indexing CBR CF model, hereafter referred to as SCP, compared to baseline models and comparative 3-step models. The list of experiment protocols is shown in Table 1. 4.3.1. SOM cluster-indexing CBR CF model On the belief that an affinity group can be clustered according to the distribution of their rating values, a couple of clustering and visualizing methods are performed to find the number of clusters. In this study, clustering techniques involve two distinct works: (1) the determination of the number of clusters present in the reference-base; and (2) the assignment of reference users to one cluster. The number of clusters, which is the number of nodes in the output layer, depends on the expected number of clusters, but there is currently no apparent practical or theoretical way of determining the optimal size of the output layer (Nour & Madey, 1996). There is a possible instability due to the randomness of clusters, so it requires a policy for initial cluster selection. To reduce this possibility, the SCP model contains the PCA process and preliminary SOM clustering. At first, each reference user’s distribution is Table 1 List of experiment protocols Protocol Description Proposed model SCP SOM cluster-indexing CBR CF predictor Comparative model Baseline model UAP By-user-average CF predictor IAP By-item-average CF predictor SPP Simple Pearson CF predictor 3-Step model SIP SOM cluster induction CF predictor SNP SOM cluster neural network CF predictor T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423 419 summarized by PCA using 6093 rating records in the 33 movie among 10 genres before SOM clustering. In this experiment, we chose three components by PCA, and data are projected onto the eigen-plane for pre-visualization shown as Figs. 5 and 6. In preliminary SOM clustering, the number of clusters is set as 2 – 10. When deciding the optimal number of clusters, the lowest cluster number is selected so that each cluster can have as many indexed reference users as possible. If a cluster has no or only a few users, this cluster does not have sufficient user cases to find a more similar user within indexed users. The total number of neurons in the output layer is decided as being four clusters according to the results of the preliminary SOM that was performed to find the number of clusters. Based on the SOM learning, each user in reference DB is indexed into a cluster. In addition to indexing, the centroid values of each item are deduced and its average values of each genre cluster suggest each cluster’s inferable genre preference characteristics shown in Table 2(a) and (b). It appears as if each cluster is grouped by rating level, and a quick glance seems to suggest that the average rating is going up, i.e. C4 (0.7810) is relatively more liberal than C1 (0.5400). However, a closer look shows that each cluster has different genre and movie preferences. For example, C1 is a comparatively negative group, but this group prefers tough genres such as horror, crime, and war Multimedia Component to soft ones such as drama, romance films. On the other hand, C3 rates higher than C1. This affinity group likes science-fiction and adventure rather than horror, which is the most preferred genre for C1. 4.3.2. Comparative models Previous research on CF algorithms has tended to compare the performance of algorithms exclusively within that research, so making comparisons of algorithm per- formance from paper to paper is difficult. We show three baseline predictors, by-user-average CF predictor (UAP), by-item-average CF predictor (IAP) and simple Pearson CF predictor (SPP), to provide benchmarks against which any predictor could be compared. The baseline algorithms are simple, efficient and return reasonable results. The UAP returns the average of the ratings the given user has already entered. The IAP gives the average rating for the given movie of all users that have voted for that movie. The SPP returns the Pearson correlation based neighbourhood prediction. To demonstrate the utility of the SCP model, we change the CBR process into neural network and induction technique as a classification method withholding the 1st SOM profiling step and the 3rd Pearson correlation-based prediction step. The SOM neural network CF Predictor (SNP) uses the well known back propagation neural network algorithm in the classification step and all decision coefficients are controlled to achieve the best prediction accuracy. The SOM induction CF Predictor (SIP) uses decision a tree technique known as induction, especially SEE 5.0 algorithm which is an upgraded version of Quinlan (1993) decision tree classifier C4.5. Fig. 6. Scatter plot of reference users projected onto the 3D ‘eigen-plane’ for visualization. Fig. 5. Scree curve of eigenvalue explained by components. The largest amount of eigenvalue is explained by the first component. The first three components can be the representative factors of the population dataset. Table 2a Genre average centroid values of SOM cluster Genre Children’s Drama Adventure Science-fiction Crime Thriller War Romance Comedy Horror Average Cluster C1 0.4845 0.4775 0.5432 0.5435 0.6473 0.5477 0.6090 0.4980 0.5540 0.6870 0.5400 C2 0.5595 0.6050 0.6756 0.6525 0.8367 0.6693 0.6570 0.5590 0.5540 0.5790 0.6490 C3 0.6910 0.5820 0.7218 0.7488 0.6340 0.6857 0.7090 0.5970 0.9313 0.5450 0.6860 C4 0.7180 0.6965 0.8336 0.8124 0.8210 0.7910 0.8820 0.6785 0.6897 0.7230 0.7810 Bold numbers mean the most preferred genre in each cluster, and italicized numbers are the worst. T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423420 5. Results To validate the effectiveness of SCP, prediction accuracy is compared with the comparative experimental algorithms in terms of NMAE and the area under the ROC curve. Table 3(a) and (b) tabulates the accuracy results of each protocol on the movie dataset. SPP reflects better users’ preference than simple average predictors (UAP, IAP) in the all experiment dataset, Allbut1, Given5 and Given10 at the 5% significance level among the baseline model, which are the same results as the ROC area metric. These results support that CF algorithm is more accurate than simple average predictors that have not considered the like-minded affinity group to predict active users’ preference. Comparing between UAP and IAP, when not enough rating data are available to apply a CF algorithm, user average value has more predictive power than item average value. SCP and SNP, which use a clustering-classification method, yield superior results to SPP, which is a type of memory-based model. However, the SIP model shows worse performance than all other experimental protocols. So model-based CF methods, especially machine learning technique-based CF models, have promising potential to improve, but their success depends on methodology suitability. Especially, the SCP model shows an out- performing result (0.1583 NMAE on Given 10 and an ROC area of 0.8461) compared with other comparative modelling by NMAE and ROC area metrics. It dominates UAP, IAP, SPP, SIP models at the 5% and 1% significance levels and yields better performance than SNP. According to our experiment, the SCP model can alleviate prediction error about 4%. From the viewpoint of data sparsity, Given10’s average NMAE (0.1790), which has less preference information, reports a higher error rate than Allbut1 (0.1564) and Given5 (0.1747). This implies that as explicit information, about preference rating to items, becomes sparse, the prediction accuracy decreases. In this study, we make a pre-filtered DB, named as reference DB which is composed of dense user-item rating data. Building a high-density preference DB can be one approach for alleviating the sparsity problem to achieve higher recommendation accuracy. Overall, among the ROC curves illustrated in Fig. 7, the SCP model is stably dominant to other clustering-classifi- Table 3a Performance results: Prediction accuracy and area under ROC curve of each protocol Protocol Metric NMAE ROC Area Allbut 1 Given5 Given10 UAP 0.1600 0.1863 0.1901 0.7127 IAP 0.1714 0.1930 0.1951 0.6071 SPP 0.1497 0.1719 0.1734 0.8006 SNP 0.1413 0.1557 0.1638 0.8166 SIP 0.1687 0.1892 0.1933 0.7430 SCP 0.1475 0.1524 0.1583 0.8461 Average 0.1564 0.1747 0.1790 Lower value of NMAE metric indicates better performance, and for ROC area, it is vice versa. Table 2b Centroid values of SOM cluster by item Genre Items Cluster C1 C2 C3 C4 Children’s 1 0.407 0.520 0.739 0.751 2 0.562 0.599 0.643 0.685 Drama 3 0.488 0.611 0.581 0.766 4 0.467 0.599 0.583 0.627 Adventure 5 0.431 0.666 0.675 0.759 6 0.690 0.790 0.841 0.966 7 0.543 0.637 0.741 0.828 8 0.690 0.752 0.754 0.913 9 0.362 0.533 0.598 0.702 Science-fiction 10 0.705 0.822 0.942 0.974 11 0.676 0.676 0.857 0.941 12 0.597 0.689 0.884 0.891 13 0.516 0.622 0.705 0.705 14 0.364 0.489 0.606 0.663 15 0.542 0.607 0.766 0.839 16 0.570 0.682 0.647 0.790 17 0.523 0.685 0.599 0.783 18 0.503 0.673 0.681 0.797 19 0.556 0.699 0.758 0.850 20 0.427 0.534 0.741 0.681 Crime 21 0.593 0.875 0.592 0.902 22 0.670 0.783 0.666 0.784 23 0.679 0.852 0.644 0.777 Thriller 24 0.466 0.628 0.733 0.835 25 0.467 0.560 0.588 0.615 26 0.710 0.820 0.736 0.923 War 27 0.609 0.657 0.709 0.882 Romance 28 0.569 0.642 0.645 0.779 29 0.427 0.476 0.549 0.578 Comedy 30 0.655 0.631 0.800 0.815 31 0.601 0.585 0.519 0.630 32 0.406 0.446 0.575 0.624 Horror 33 0.687 0.579 0.545 0.723 Table 3b Performance results: Statistical significance test—one-way ANOVA with the post hoc test IAP SPP SNP SIP SCP UAP 20.0202 0.0665** 0.1052*** 20.0131 0.1271*** IAP – 0.0867** 0.1253*** 0.0070 0.1472*** SPP – – 0.0387* 20.0797** 0.0605** SNP – – – 20.1183*** 0.0219 SIP – – – – 0.1402** Bonferroni procedure based on MAE: mean differences and signifi- cance level at ***1%, **5% and *10% level for the pair-wise comparison of performance between protocols. T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423 421 cation CF models and memory-based models. This implies that when the item recommendation criterion is changed, the SCP model can be flexibly applied. For example, even if the recommendable item criterion is changed from 5 to 4 stars in the movie recommendation, the SCP model does still works. 6. Conclusion In this paper, we propose an SCP model which applies two combing machine learning techniques, SOM and CBR to the consecutive CF prediction process as a new approach in the CF recommendation field. This study shows that cluster-indexing CBR is an effective user indexing and retrieval method for CF recommendation. Instead of using all user’s ratings for retrieval of nearest neighbour, SOM cluster-indexing CBR approach allevi- ates the on-line computational complexity by use of significant cluster-centroid values induced from SOM process. The SOM facilitates affinity user grouping and extraction of representative centroid values of each cluster’s items for assisting case indexing and retrieval of CBR. For SOM clustering of our study, most of the computational costs are involved in the training process and this process is done off-line in the profiling step. After that, the CBR process and CF prediction in the selected cluster are a compact representation of the raw ratings information, and thus the time and space complexities on making recommendations are quite low. The performance of our model yields superior results compared to memory-based CF techniques and other previous hybrid CF models. The NMAE values induced from our model indicate that predicted rating values will be within roughly 15% of the true rating values. So the items with predicted ratings well above the mean for a new user in many cases will correspond to desirable items for that user. These accuracies are comparable with those reported for a completely different data set (jokes); the algorithms in Goldberg et al. (2001) show NMAE from 0.187 to 0.237 in the 20 unit rating scale [210, þ10]. Herlocker et al. (1999) report MAE from 0.768 to 0.828. When these are normal- ized to the 4 unit rating scale [1, 5], they yield NMAE from 0.192 to 0.207 in the same MovieLens data set. According to Goldberg et al. (2001), if user ratings are distributed uniformly or normally, random predictions yield NMAE of 33 and 28%, respectively. Our model also yields superior performance when compared to other traditional memory- based CF algorithms and other NN and induction based CF prediction algorithms. This suggests that there is room for improved accuracy for all current CF algorithms. We are experimenting with a number of variations, such as k-means clustering and hybrid approaches with adaptive online weighting to further improve accuracy without altering online computation time. This study compares several computational approaches to CF recommendation, considering prediction accuracy and response speed simul- taneously. Especially, data mining techniques, such as SOM, NN and CBR have shown the potential for improving. The promising potential for CF systems can be investigated by integrating with product/customer-specific information profiling, implicit information analysis such as web-page navigation history and retrieval technology. In the future research, we will suggest hybrid recommendation algor- ithms and try to apply our model into a real-world personalized recommendation site. Acknowledgements This research was financially supported by Han Sung University in the year of 2003. References Aamodt, A., & Plaza, E. (1994). Case-based reasoning: Foundational issues, methodological variations, and system approaches. Artificial Intelligence Communications, 7(1), 39 – 59. Balabanovı́c, M., & Shoham, Y. (1997). Fab: Content-based, collaborative recommendation. Communications of the ACM, 40(3), 66 – 72. Basu, C., Hirsh, H., Cohen, W., (1998). Recommendation as classification: Using social and content base information in recommendation. Proceedings of the 1998 Workshop on Recommender Systems (pp.11 – 15) Bradley, P. S. (1994). Case-based reasoning: Business applications. Communication of the ACM, 37(3), 40 – 43. Breese, J.S., Heckerman, D., Kadie, C., (1998). Empirical analysis of predictive algorithms for collaborative filtering. Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence (UAI-98) (pp. 43 – 52) Chiu, C. (2002). A case-based customer classification approach for direct marketing. Expert Systems with Applications, 22(2), 163 – 168. Fig. 7. ROC curve comparison of CF prediction protocols. Upper line indicates more prediction accuracy. T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423422 Cho, Y. H., Kim, J. K., & Kim, S. H. (2002). A personalized recommender system based on web usage mining and decision tree induction. Expert Systems with Applications, 23(3), 329 – 342. Choy, K. L., Lee, W. B., & Lo, V. (2002). Development of a case based intelligent customer-upplier relationship management system. Expert Systems with Applications, 23(3), 281 – 297. Claypool, M., Gokhale, A., Miranda, T., Murnikov, P., Netes, D., Sartin, M., (1999). Combining content-based and collaborative filters in an online newspaper. ACM SIGIR’99 Workshop on Recommender Systems, Berkely, CA Delgado, P.J., Ishii, N., (1999). Memory-based weighted majority prediction for recommender systems. SIGIR Workshop on Recommen- der Systems Goldberg, D., Nichols, D., Oki, B. M., & Terry, D. (1992). Using collaborative filtering to weave an information tapestry. Communi- cations of the ACM, 35(12), 61 – 70. Goldberg, K., Roeder, R., Gupta, D., & Perkins, C. (2001). Eigentaste: A constant time collaborative filtering algorithm. Information Retrieval Journal, 4(2), 133 – 151. Good, N., Schafer, J.B., Konstan, J., Borchers, A., Sarwar, B., Herlocker, J., Riedl, J., (1999). Combining collaborative filtering with personal agents for better recommendations. Proceedings of the 1999 Conference of the American Association of Artificial Intelligence (AAAI-99) Ha, S. H., & Park, S. C. (1998). Application for data mining tool to hotel data mart on the Internet for database marketing. Expert Systems with Applications, 15(1), 1 – 31. Herlocker, J., Konstan, J., Borchers, A., Riedl, J., (1999). An algorithmic framework for performing collaborative filtering. Proceedings of the 1999 Conference on Research and Development in Information Retrieval Hill, W., Stead, L., Rosenstein, M., & Furnas, G. (1995). Recommending and evaluating choices in a virtual community of use. CHI 95, Denver, CO: ACM Press, pp. 194 – 201. Hui, S. C., Fong, A. C. M., & Jha, G. (2001). A web-based intelligent fault diagnosis system for customer service support. Engineering Appli- cations of Artificial Intelligence, 14(4), 537 – 548. Kim, K. S., & Han, I. G. (2001). The cluster-indexing method for case- based reasoning using self-organizing maps and learning vector quantization for bond rating cases. Expert Systems with Applications, 21(3), 147 – 156. Kohonen, T., Hynninen, J., Kangas, J., Laaksonen, J., (1996). SOM_PAK: The self-organizing map program package. Technical Report A31, Helsinki University of Technology, Laboratory of Computer and Information Science, FIN-02150 Kolodner, J. L. (1993). Case-Based Reasoning. Los Altos, CA: Morgan Kaufmann. Konstan, J. A., Miller, B. N., Maltz, D., Herlocker, J. L., Gordon, L. R., & Riedl, J. (1997). GroupLens: Applying collaborative filtering to Usenet news. Communications of the ACM, 40(3), 77 – 87. Kumar, R., Raghavan, P., Rajagopalan, S., Tomkins, A., (1998). Recommendation systems: A probabilistic analysis. Proceedings of the 39th Annual Symposium on Foundations of Computer Science Mangiameli, P., Chen, S. K., & West, D. (1996). A comparison of SOM neural network and hierarchical clustering methods. European Journal of Operation Research, 93, 402 – 417. Nour, A. N., & Madey, G. R. (1996). Heuristic and optimization approaches to extending the Kohonen self organizing algorithm. European Journal of Operation Research, 93, 428 – 448. Pazzani, M. J. (1999). A framework for collaborative, content-based and demographic filtering. Artificial Intelligence Review, 13(5/6), 393 – 408. Pennock, D.M., Horvitz, E., Lawrence, S., Giles, C.L., (2000). Collabora- tive filtering by personality diagnosis: A hybrid memory- and model- based approach. Proceedings of the 16th Conference on Uncertainty in Artificial Intelligence (UAI-2000) (pp. 473 – 480). Stanford, CA Quinlan, J. R. (1993). C4.5: Programs for Machine Learning. Morgan Kaufmann. Resnick, P., Iacovou, N., Sushak, M., Bergstrom, P., Riedl, J., (1994). GroupLens: An open architecture for collaborative filtering of netnews. Proceedings of the 1994 Computer Supported Collaborative Work Conference Schafer, J. B., Konstan, J. A., & Riedl, J. (2001). E-commerce recommendation applications. Data Mining and Knowledge Discovery, 5(1 – 2), 115 – 153. Shardanand, U., Maes, P., (1995). Social information filtering: Algorithms for automating ‘Word of Mouth’. Proceedings of ACM CHI ‘95 (pp. 210 – 217). Denver, CO Swets, J. A. (1988). Measuring the accuracy of diagnostic systems. Science, 240, 1285 – 1289. Vellido, A., Lisboa, P. J. G., & Meehan, K. (1999). Segmentation of the on- line market using neural networks. Expert Systems with Applications, 17(4), 303 – 314. Yuan, S., & Chang, W. (2001). Mixed-initiative synthesized learning approach for web-based CRM. Expert Systems with Applications, 20(2), 187 – 200. T.H. Roh et al. / Expert Systems with Applications 25 (2003) 413–423 423 The collaborative filtering recommendation based on SOM cluster-indexing CBR Introduction Background Collaborative filtering recommendation Self-organizing map Case based reasoning SOM cluster-indexing CBR CF recommendation Profiling step Inferring step Predicting step Experiments Data: MovieLens dataset Evaluation metrics Experimental setup Results Conclusion Acknowledgements References 
work_gxz4onlrcfhvvpeojp5vphrspm	----	A classifier fusion system for bearing fault diagnosis Expert Systems with Applications 40 (2013) 6788–6797 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A classifier fusion system for bearing fault diagnosis 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.06.033 ⇑ Corresponding author. Tel.: +1 (514) 396 8932; fax: +1 (514) 396 8595. E-mail addresses: luana.bezerra@gmail.com (L. Batista), bechirbadri@yahoo.fr (B. Badri), robert.sabourin@etsmtl.ca (R. Sabourin), marc.thomas@etsmtl.ca (M. Thomas). Luana Batista, Bechir Badri, Robert Sabourin ⇑, Marc Thomas École de Technologie Supérieure, 1100, rue Notre-Dame Ouest, Montreál, QC H3C 1K3, Canada a r t i c l e i n f o a b s t r a c t Keywords: Bearing fault diagnosis Vibration analysis Machine condition monitoring Support vector machines Iterative Boolean Combination ROC curves Classifier fusion In this paper, a new strategy based on the fusion of different Support Vector Machines (SVM) is proposed in order to reduce noise effect in bearing fault diagnosis systems. Each SVM classifier is designed to deal with a specific noise configuration and, when combined together – by means of the Iterative Boolean Combination (IBC) technique – they provide high robustness to different noise-to-signal ratio. In order to produce a high amount of vibration signals, considering different defect dimensions and noise levels, the BEAring Toolbox (BEAT) is employed in this work. The experiments indicate that the proposed strat- egy can significantly reduce the error rates, even in the presence of very noisy signals. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction Although the visual inspection of time- and frequency-domain features of measured signals is adequate for identifying machinery faults, there is a need for a reliable, fast and automated procedure of diagnosis (Samanta et al., 2004). Due to the increasing demands for greater product quality and variability, short product life-cy- cles, reduced cost, and global competition, automatic machine con- dition monitoring (MCM) has been gaining importance in the manufacturing industry (Liang et al., 2004). MCM systems allow for a significant reduction in the machinery maintenance costs, and, most importantly, the early detection of potential faults (Guo et al., 2005). Mass unbalance, rotor rub, shaft misalignment, gear failures and bearing defects are exemples of faults that may lead to the machine’s breakdown (Samanta et al., 2004). Besides the detection of the early occurence and seriousness of a fault, MCM systems may also be designed to identify the compo- nents that are deteriorating, and to estimate the time interval dur- ing which the monitored equipment can still operate before failure (Lazzerini and Volpi, 2011). These systems continuously measure and interpret signals (e.g., vibration, acoustic emission, infrared thermography, etc.), that provide useful information for identifying the presence of faulty symptoms. The focus of this work is in rotating machines, which usually operate by means of bearings. Since they are the place where the basic dynamic loads and forces are applied, bearings represent a critical component. A defective bearing causes malfunction and may even lead to catastrophic failure of the machinery (Tandon and Choudhury, 1999). Vibration analysis has been the most em- ployed methodology for detecting bearings defects (Thomas, 2011). Each time a rolling element passes over a defect, an impulse of vibration is generated. On the other hand, if the machine is oper- ating properly, vibration amplitude is small and constant (Alguin- digue et al., 1993). Another methodology successfully applied to this problem has been the acoustic emission (AE) (Elmaleeh and Saad, 2008; Tandon and Choudhury, 1999). Automatic bearing fault diagnosis can be viewed as a pattern recognition problem, and several systems have been designed using well-known classification techniques, such as Artificial Neu- ral Networks (ANNs) and Support Vector Machines (SVM). When these systems employ real vibration data obtained from bearings artificially damaged, they have to cope with a very limited amount of samples. Furthermore, with exception of a few works (Guo et al., 2005; Jack and Nandi, 2002) – which consider a validation set, be- sides the training and test sets –, the choice of the system’s param- eters, including the feature selection step, too often has been done by using the same datasets employed to train/test the classifiers. This may lead to biased classifiers that will hardly be able to gen- eralize on new data. Another important aspect that has been little investigated in the literature is the presence of noise, which dis- turbs the vibration signals, and how this affects the identification of bearing defects (Lazzerini and Volpi, 2011). In this paper, a classification system based on the fusion of dif- ferent SVMs is proposed to detect early defects on bearings in the presence of high noise levels. Each SVM classifier is designed to deal with a specific noise configuration and, when combined to- gether – by using the Iterative Boolean Combination (IBC) tech- nique (Khreich et al., 2010) – they provide high robustness to different noise-to-signal ratio. In order to produce a high amount of bearing vibration signals, considering different defect dimensions and noise levels, the BEAr- ing Toolbox (BEAT) is employed in this work. BEAT is dedicated to the simulation of the dynamic behaviour of rotating ball bearings http://crossmark.dyndns.org/dialog/?doi=10.1016/j.eswa.2013.06.033&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.06.033 mailto:luana.bezerra@gmail.com mailto:bechirbadri@yahoo.fr mailto:robert.sabourin@etsmtl.ca mailto:marc.thomas@etsmtl.ca http://dx.doi.org/10.1016/j.eswa.2013.06.033 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa L. Batista et al. / Expert Systems with Applications 40 (2013) 6788–6797 6789 in the presence of localized defects, and it was shown to provide realistic results, similar to those produced by a sensor during experimental measurements (Sassi et al., 2007). This paper is organized as follows. Section 2 presents the state- of-the-art in automatic bearing fault diagnosis. Section 3 describes the experimental methodology, including datasets, measures used to evaluate the system performance, and the IBC technique. Finally, the experiments are presented and discussed in Section 4. 2. The state-of-the-art in automatic bearing fault diagnosis Fig. 1 illustrates the general structure of a bearing. It is com- posed of six components: housing, outer race (OR), inner race (IR), rolling elements (RE) (i.e., rollers or balls), cage and shaft (Guo et al., 2005). As previously mentioned, the interaction of de- fects in rolling element bearings produces impulses of vibration. As these shocks excite the natural frequencies of the bearing ele- ments, the analysis of the vibration signal in the frequency-do- main, by means of the Fast Fourrier Transform (FFT), has been an effective method for predicting the health condition of bearings (Tandon and Choudhury, 1999). Each defective bearing component produces frequencies, which allow for localizing different defects occurring simultaneously. BPFO (Ball Pass Frequency on an Outer race defect), BPFI (Ball Pass Frequency on an Inner race defect), FTF (Fundamental Train Fre- quency) and BSF (Ball Spin Frequency) – as well as their harmonics, modulating frequencies, and envelopes – are examples of fre- quency-domain indicators, calculated from kinematic consider- ations – that is, the geometry of the bearing and its rotational speed (Sassi et al., 2007). It is worth noting that the shock amplitude is directly related to the defect dimension: the bigger the defect, the bigger the shock. Fig. 1. Typical roller bearing, showing different component parts. Adapted from Jack and Nandi (2002). Fig. 2. Example of a hypothetical defect located in the rolling element (a) and its corresp frequency). Adapted from Sassi et al. (2007). Fig. 2 presents an example of a defect located in the outer race and its corresponding vibration signal. Not only frequency- but also time-domain indicators have been widely employed as input features to train a bearing fault diagno- sis classifier. Time-domain indicators are adimensional, and allow for representing the vibration signal through a single scalar value. For instance, peak is the maximum amplitude value of the vibra- tion signal, RMS (Root Mean Square) represents the effective value (magnitude) of the vibration signal and Kurtosis describes the impulsive shape of the vibration signal. Table 1 presents the effec- tiveness (advantages and disadvantages) of some time-domain indicators in describing the presence (or absence) of faulty symp- toms (Kankar et al., 2011; Sassi et al., 2008; Tandon and Choudhu- ry, 1999). A bearing fault diagnosis system may be designed to provide different levels of information about the defect (s). The first and simpler issue investigated in the literature is the detection of the presence or absence of a defect (Jack and Nandi, 2002; Samanta et al., 2004). The second issue is the determination of the defect location, which may occur in different components of a bearing (Alguindigue et al., 1993; Bhavaraju et al., 2010). Often, the type of defect is considered along with the defect location. For instance, some authors consider the following classes: sandblasting of IR/OR, indentation on the roll, unbalanced cage (Lazzerini and Volpi, 2011; Volpi et al., 2010), crack on IR/OR, spall on IR/OR, spalls on rollers (Widodo et al., 2009), generalized fault of two balls (Alguin- digue et al., 1993), etc. Finally, the severity of a bearing defect is the last and perhaps the most difficult information to be predicted. Through this infor- mation, it may be possible to estimate the duration during which the equipment can still operate safely. In the literature, this issue has been partially investigated, by associating a different class to each defect dimension (Cococcioni et al., 2009a, 2009b; Widodo et al., 2009). Cococcioni et al. (2009a), for example, have employed three classes for describing the seriousness of an ‘‘indentation on the roll’’, namely, light (450lm), medium (1.1 mm) and high (1.29 mm). The drawback of this strategy is that other defect dimensions are not considered by the classifier. A more suitable solution would be the estimation of defect dimensions as a regres- sion problem. Table 2 presents a summary of different systems reported in the literature, with their respective employed classification tech- niques, types of signal, descriptors (features), types of defects and datasets. It is important to mention that the bearing defects may be categorized as distributed or local. Distributed defects are due to unavoidable manufacting imperfections, such as surface roughness, waviness, misaligned races and off-size rolling ele- ments (Sassi et al., 2007), whereas localized defects include cracks, onding shock impulses (b), where FTF is the Fundamental Train Frequency (or cage Table 1 Time-domain indicators. Indicator Advantage Disadvantage Peak May indicate the presence of a defect even at the initial stage The signal source is unknown; May create a false alarm Root Mean Square (RMS) Toward the end of the bearing life, the RMS level increases dramatically Low sensitivity to indicate a defect at the initial stage; The signal source is unknown Kurtosis Low sensitivity to the variations of load and speed; Well suited for detecting a defect at the initial stage When the defect is at an advanced stage, the Kurtosis value comes down to a value of an undamaged bearing; The signal source is unknown Crest Factor (CF) Impulse Factor (IF) Like Kurtosis, CF and IF are well suited for detecting a defect at the initial stage Same problems as Kurtosis Thikat (Sassi et al., 2008) May indicate the presence of a defect at any rotational speed Same problems as Kurtosis; No physical meaning; Needs the initial RMS value Talaf (Sassi et al., 2008) The talaf value constantly increases with the defect dimension; A slope change is an indication of impending failure; Indicates 4 levels of degradation The signal source is unknown; No physical meaning; Needs the initial RMS value Table 2 Survey of bearing fault diagnosis systems. (AE = Acoustic Emission, MLP = Multi-Layer Perceptron, SVM = Support Vector Machine, CHC = Convex Hull Classifier, PNN = Probabilistic Neural Network, RNN = Recirculation Neural Network, RBF = Radial Basis Function, GA = Genetic Algorithms, HMM = Hidden Markov Model, MFCC = Mel-Frequency Complex Cepstrum, SOM = Self-Organizing Maps, RVM = Relevance Vector Machine, QDC = Quadratic Discriminant Classifier, LDC = Linear Discriminant Classifier, PCA = Principal Component Analysis, ICA = Independent Component Analysis, EoC = Ensemble of Classifiers.) Refs. Classifiers Signals Features Defect classes Datasets Kankar et al. (2011) SVM MLP SOM real vibration signals, artificial defects 5 different speeds kurtosis, skewness, std (from wavelet coefficients) number of loaders, speed faultless bearing, IR fault, OR fault, RE fault, fault in all components 150 samples 10-fold cross-validation Bhavaraju et al. (2010) MLP SOM real vibration signals, artificial defects, 5 different speeds kurtosis, skewness, std (from wavelet coefficients), number of loaders, speed faultless bearing, IR fault, OR fault, RE fault, fault in all components 150 samples 50% training, 50% test Lazzerini and Volpi (2011) ensembles of MLPs real vibration signals, artificial defects, 10 different noise levels FFT parameters (performed forward feature selection) faultless bearing, indentation on IR, indentation on the roll, sandblasting of IR, unbalanced cage 12740 samples 70% training, 30% test (100 trials) Volpi et al. (2010) one-class CHC real vibration signals, artificial defects FFT parameters (performed forward feature selection) faultless bearing unbalanced cage, indentation on IR (450 lm), sandblasting of IR, indentation on the roll (450 lm, 1.1 mm and 1.29 mm) 12740 samples training with ‘‘fautless’’ class, test with all classes (30 trials) Widodo et al. (2009) RVM SVM real AE and vibration signals, artificial defects, considered only low-speeds (e.g., 20 and 80 rpm) statistical, time- and frequency-domain features selected with PCA/ICA faultless bearing, crack on IR (0.1 mm), spall on IR (0.6 mm), crack on OR (0.1 mm), spall on OR (0.7 mm), spalls on rollers (1 mm and 1.6 mm) 105 samples cross-validation Cococcioni et al. (2009b) LDC, QDC, MLP, RBF NN real vibration signals, artificial defects, 10 different noise levels FFT parameters (performed forward feature selection) faultless bearing, indentation on IR, indentation on the roll (450 lm, 1.1 mm and 1.29 mm), sandblasting of IR, unbalanced cage 12740 samples 70% training, 30% test (100 trials) Cococcioni et al. (2009a) LDC, QDC, MLP, EoC real vibration signals, artificial defects, 5 frequency ranges FFT parameters (performed forward feature selection) faultless bearing, indentation on IR, indentation on the roll (450 lm, 1.1 mm and 1.29 mm), sandblasting of IR, unbalanced cage 12740 samples 70% training, 30% test (10 trials) 6 7 9 0 L. B a tista et a l./E xp ert System s w ith A p p lica tio n s 4 0 (2 0 1 3 ) 6 7 8 8 – 6 7 9 7 Table 3 Survey of bearing fault diagnosis systems (continuation). Refs. Classifiers Signals Features Defect classes Datasets Sreejith et al. (2008) MLP real vibration signals, artificial defects time-domain features fautless bearing, RE fault, OR fault, IR fault 80 samples from CWRU bearing data center (Case Western Reserve University) 60% training, 40% test Teotrakool et al. (2008) SVM motor current signals, artificial defects, 4 different speeds RMS values from wavelet packet coefficients (feature selection with GA) faultless bearing vs. OR fault; faultless bearing vs. cage fault – Lei et al. (2008) improved fuzzy c-means real vibration signals from locomotive roller bearings time-domain features fautless bearing slight rub faults on OR, serious flaking faults on OR 150 samples for clustering Sugumaran et al. (2008) one-class & multi-class SVMs real vibration signals, artificial defects, 3 different speeds Kurtosis and statistical features (selected with a decision tree) faultless bearing, OR fault, IR fault, OR fault + IR fault – Sugumaran et al. (2007) SVM, proximal SVM real vibration signals, artificial defects, 3 different speeds Kurtosis and statistical features (selected with a decision tree) faultless bearing, OR fault, IR fault, OR fault + IR fault 600 samples 83% training, 17% test Abbasion et al. (2007) SVM real vibration signals, artificial defects Weibull negative log-likelihood function of time-domain signals faultless bearing, IR-drive fault, IR-fan fault, RE-drive fault, RE-fan fault, OR-drive fault, OR-fan fault 63 samples for test Rojas and Nandi (2006) SVM real vibration signals, speeds FFT parameters and statistical features faultless bearing, worn bearing, OR fault, IR fault, RE fault, cage fault 1920 samples 50% training, 50% test Guo et al. (2005) MLP, SVM real vibration signals, defects artificially introduced, 16 different speeds statistical, frequency- and time-domain features selected with GA fautless bearing, worn bearing, cage fault, IR fault, OR fault, RE fault 2880 samples 1/3 training, 1/3 test 1/3 validation, Table 4 Survey of Bearing Fault Diagnosis Systems (continuation). Refs. Classifiers Signals Features Defect classes Datasets Purushotham et al. (2005) HMMs (one per class) real vibration signals, artificial defects, considered multiple faults MFCC coefficients (wavelet analysis) 2 faults on IR + 1 fault on RE, 2 faults on OR + 1 fault on RE, one fault in each component training, test (4 different splits) Samanta et al. (2004) MLP, RBF NN, PNN real vibration signals, artificial defects statistical and time-domain features selected with GA fautless bearing vs. faulty bearing (OR fault) 288 samples 50% training, 50% test Samanta et al. (2003) MLP SVM real vibration signals, artificial defects statistical and time-domain features selected with GA fautless bearing vs. faulty bearing (OR fault) 288 samples 60% training, 40% test Samanta and Al-Balushi (2003) MLP real vibration signals, artificial defects statistical and time-domain features fautless bearing vs. faulty bearing (OR fault) 200 samples 60% training, 40% test Lou and Loparo (2004) neuro-fuzzy real vibration signals, artificial defects, 4 different load values std of wavelet coefficients fautless bearing, IR fault, RE fault 24 samples 50% training, 50% test Jack and Nandi (2002) SVM, MLP real vibration signals, artificial defects, 16 different speeds statistical and frequency-domain features selected with GA faultless (brand new bearing, worn bearing) vs. faulty (OR fault, IR fault, RE fault, cage fault) 2880 samples 1/3 training, 1/3 test, 1/3 validation Jack and Nandi (2001) SVM, MLP real vibration signals, artificial defects, 16 different speeds statistical and frequency-domain features faultless bearing, worn bearing, OR fault, IR fault, RE fault, cage fault 960 samples 1/3 training, 1/3 test, 1/3 validation Alguindigue et al. (1993) RNN, MLP real vibration signals, real and artificial defects high- and low- frequency features faultless bearing, fault on IR, generalized fault on IR, fault on OR, generalized fault on OR, artificial fault of a ball, generalized fault of two balls, generalized fault of all the components the test set contained samples from the training set L. B a tista et a l./E xp ert System s w ith A p p lica tio n s 4 0 (2 0 1 3 ) 6 7 8 8 – 6 7 9 7 6 7 9 1 Table 6 Classes of defects. OR IR RE class 0 0 0 0 class 1 1 0 0 class 2 0 1 0 class 3 0 0 1 class 4 1 1 0 class 5 1 0 1 class 6 0 1 1 class 7 1 1 1 Table 7 Data partitioning for each DB(nc) (1 6 nc 6 6). Positive class Negative class trn 3500 3500 vld 1750 1750 tst (per noise level) 1750 1750 Table 8 ROC AUC on validation data. System AUC S(nc=1) 1 S(nc=2) 0.9999 S(nc=3) 0.9999 S(nc=4) 0.9996 S(nc=5) 0.9992 S(nc=6) 0.9989 Fig. 3. DET curves of the selected systems S(nc), 1 6 nc 6 6, using their respective validation sets (vld). 6792 L. Batista et al. / Expert Systems with Applications 40 (2013) 6788–6797 pits and spalls on the rolling surfaces (Tandon and Choudhury, 1999). In Tables 2–4, only localized defects are considered. Some authors have worked with signals obtained from multiple rotational speeds. With exception of Widodo et al. (2009) and Sugumaran et al. (2007) – which developed a different system for each rotational speed –, the classifiers have been trained/tested with data corresponding to several speeds simultaneously (Guo et al., 2005; Jack and Nandi, 2002; Rojas and Nandi, 2006; Teotra- kool et al., 2008), and, sometimes, the rotational speed is employed as input-feature (Bhavaraju et al., 2010; Kankar et al., 2011). How- ever, these systems consider either non-rotating loads or no-load conditions, which means that the shock amplitudes are not af- fected if the rotational speed changes. So far, no work investigated the case where a same system has to deal with different speeds un- der a rotating load. Regarding non-rotating loads, few works have considered sig- nals obtained from multiple load conditions. While (Bhavaraju et al., 2010; Kankar et al., 2011) employed the number of loaders (which goes from 0 to 2) as input-feature, (Lou and Loparo, 2004) acquired vibration data from four load values (0, 1, 2 and 3 Horse Power (HP)). In both cases, the signals regarding the different load conditions were employed to train/test a same classifier. 3. Methodology The objective of this work is to detect the presence or absence of bearing defects by taking into account six levels of noise, i.e., sig- nal-to-noise ratio ranging from 40 to 5 db. Noise robustness is achieved through the incorporation of noisy data during the train- ing phase, along with the fusion of different SVMs, each one is de- signed to deal with a specific noise configuration. The BEAT simulator (Sassi et al., 2007) is employed to generate vibration signals coming from the operation of a ball bearing type SKF 1210 ETK9. The rotational speed is 1800 RPM, subjected to a non-rotating load of 3000 N. From the simulated data, the follow- ing time-domain indicators are calculated: RMS, peak, Kurtosis, crest factor, impulse factor and shape factor. As frequency-domain indicators, BPFO, BPFI, 2BSF, as well as their first two hamonics are calculated. It is worth noting that the frequency-domain indicators employed in this work are normalized with respect to the rota- tional speed. Regarding the time-domain indicators, they are inde- pendent of the rotational speed when the load is non-rotating. The rest of this section describes the datasets and the perfor- mance evaluation methods employed in the experiments, as well as the Iterative Boolean Combination technique. 3.1. Datasets Six noise configurations (nc = 1,2,3,4,5,6) are considered in this paper, as indicated in Table 5. For each noise configuration, there is a specific database, that is, DB(nc). Each sample in the databases is composed of a set of frequency and temporal indicators, plus the defect diameter, ddef, related to each bearing component, i.e., ddef(OR), ddef(IR) and ddef(RE). Eight classes of defects are defined in Ta- ble 6. The flag = 1 indicates that there is a defect in the correspond- ing component, while flag = 0 indicates the absence of defect. For Table 5 Noise configurations (nc). nc Training/validation Test 1 40 db 40, 30, 20, 15, 10, 5 db 2 40, 30 db 40, 30, 20, 15, 10, 5 db 3 40, 30, 20 db 40, 30, 20, 15, 10, 5 db 4 40, 30, 20, 15 db 40, 30, 20, 15, 10, 5 db 5 40, 30, 20, 15 10 db 40, 30, 20, 15, 10, 5 db 6 40, 30, 20, 15, 10, 5 db 40, 30, 20, 15, 10, 5 db instance, class 6 corresponds to two different defects occuring simultaneously: one in the outer race, and another in the ball. For the non-defective components, ddef goes from 0 mm to 0.016 mm. Regarding the defective components, ddef goes from 0.017 mm to 2.8 mm. Since the objective of this work is to indicate the presence or absence of a bearing defect, regardless its location, only two clas- ses are considered, i.e, faultless and faulty. The faultless class corresponds to the class 0 (see Table 6) and, in order to have two balanced classes, the faulty class contains subsets of samples from classes 1 to 7. Table 7 presents the way the samples are partitioned. Fig. 4. DET curves of the selected systems S(nc), 1 6 nc 6 6, using the test sets (tst). L. Batista et al. / Expert Systems with Applications 40 (2013) 6788–6797 6793 3.2. Performance evaluation methods The ROC (Receiving Operating Characteristics) curve – where the true positive rates (TPR) are plotted as function of the false po- sitive rates (FPR) – is a powerful tool for evaluating, comparing and combining pattern recognition systems (Khreich et al., 2010). Several interesting properties can be observed from ROC curves. First, the AUC (Area Under Curve) is equivalent to the probability that the classifier will rank a randomly chosen positive sample higher than a randomly chosen negative sample. This measure is useful to characterize the system performance through a single scalar value. In addition, the optimal threshold for a given class Table 9 Average EER ±r (%) on test data over 10 trials. tst S(nc=1) S(nc=2) S(nc=3) S(nc= 4) S(nc=5) S(nc=6) 40 db 0.02 ± 0.02 0.05 ± 0.05 0.10 ± 0.07 0.17 ± 0.10 0.38 ± 0.17 0.57 ± 0.27 30 db 1.40 ± 1.09 0.00 0.01 ± 0.01 0.09 ± 0.07 0.14 ± 0.10 0.32 ± 0.16 20 db 5.38 ± 0.81 7.51 ± 1.01 0.07 ± 0.06 0.06 ± 0.06 0.08 ± 0.08 0.10 ± 0.06 15 db 20.98 ± 7.4 34.12 ± 6.7 � 0.11 ± 0.03 0.16 ± 0.07 0.15 ± 0.06 10 db � � � � 0.27 ± 0.09 0.68 ± 0.26 5 db � � � � � 0.36 ± 0.08 10 −1 10 0 10 1 10 2 10 −1 10 0 10 1 10 2 FPR (%) FN R (% ) DET curves (vld) IBC MRDET Individual Systems 4 5 6 8 7 3 Fig. 5. DET curve obtained with IBC using a validation set containing all noise levels. The DET curves of the 6 individual systems and the Maximum Realizable DET curve (MRDET) are shown as well. Table 10 Operating points of IBC DET curve. Operating point FNR (%) FPR (%) Average (%) 1 100.00 0.00 50.00 2 0.89 0.00 0.45 3 0.65 0.06 0.36 4 0.48 0.24 0.36 5 0.42 0.42 0.42 6 0.24 1.13 0.69 7 0.18 2.68 1.43 8 0.12 6.25 3.19 9 0.00 15.65 7.83 10 0.00 100.00 50.00 Table 11 Decision thresholds associated to the EER operating point. Classifier Threshold c1 0.9919 c2 0.9816 c3 0.9916 c4 1.5587e�004 c5 0.0095 c6 0.0452 6794 L. Batista et al. / Expert Systems with Applications 40 (2013) 6788–6797 distribution lies on the ROC convex hull, which is defined as being the smallest convex set containing the points of the ROC curve. Fi- nally, by taking into account several operating points, the ROC curve allows for analyzing these systems under different classifica- tion costs (Fawcett, 2006). A similar way to evaluate systems is through a DET (Detection Error Trade-off) curve, in which the false negative rates (FNR) are plotted as function of the false positive rates, generally, on a logarithmic scale. In this work, ROC and DET curves are computed from the output probabilities provided by the classifiers. The validation set, vld, is used for this task. In order to test a given classifier, its correspond- ing ROC operating points (thresholds) are applied to the set, tst. Re- sults on test are shown as well in terms of equal error rate (EER), which is obtained when the threshold is set to have the false neg- ative rate approximately equal to the false positive rate. 3.3. Iterative Boolean Combination (IBC) Ensembles of classifiers (EoCs) have been used to reduce error rates of many challenging pattern recognition problems. The moti- vation of using EoCs stems from the fact that different classifiers usually make different errors on different samples. When the re- sponse of a set of C classifiers is averaged, the variance contribution in the bias-variance decomposition decreases by 1C, resulting in a smaller classification error (Tumer and Ghosh, 1996). It has been recently shown that the Iterative Boolean Combina- tion (IBC) (Khreich et al., 2010) is an efficient technique for com- bining systems in the ROC space. IBC iteratively combines the ROC curves produced by different classifiers using all Boolean func- tions (i.e., a _ b, :a _ b, a _:b, :(a _ b), a ^ b, : a ^ b, a ^:b, :(a ^ b), a � b, and a � b), and does not require prior assumption that the classifiers are statistically independent. At each iteration, IBC selects the combinations that improve the Maximum Realiz- able ROC (MRROC) curve – i.e., the convex hull obtained from all individual ROC curves – and recombines them with the original ROC curves until the MRROC ceases to improve. For more details on the IBC technique, please refer to Algorithms 1 to 3 in Khreich et al. (2010). 4. Simulation results and discussions Two main experiments are performed. In the first experiment, each database DB(nc)(1 6 nc 6 6) is employed in the generation of a baseline system S(nc). For each DB(nc): � trn is used to train n different classifiers ci, 1 6 i 6 n, by employ- ing different SVM parameters; � vld is used to validate each individual classifier ci, by means of ROC curves, and select that one with the highest AUC. The select classifier is called S(nc); � tst is used to test the performance of S(nc). In the second experiment, the IBC technique (Khreich et al., 2010) is used to combine the best classifier of each noise configuration. 4.1. Experiment 1 The goal of the first experiment was to obtain the best baseline system for each one of the noise configurations defined in Table 5. For each database DB(nc)(1 6 nc 6 6), several SVMs were trained using the grid search technique (Chang and Lin, 2001), so that 10 −1 10 0 10 1 10 2 10 −1 10 0 10 1 10 2 FPR (%) FN R (% ) DET curves (tst = 40db) IBC Individual systems 10 −1 10 0 10 1 10 2 10 −1 10 0 10 1 10 2 FPR (%) FN R (% ) DET curves (tst = 30db) IBC Individual systems 10 −1 10 0 10 1 10 2 10 −1 10 0 10 1 10 2 FPR (%) FN R (% ) DET curves (tst = 20db) IBC Individual systems 10 −1 10 0 10 1 10 2 10 −1 10 0 10 1 10 2 FPR (%) FN R (% ) DET curves (tst = 15db) IBC Individual systems 10 −1 10 0 10 1 10 2 10 −1 10 0 10 1 10 2 FPR (%) FN R (% ) DET curves (tst = 10db) IBC Individual systems 10 −1 10 0 10 1 10 2 10 −1 10 0 10 1 10 2 FPR (%) FN R (% ) DET curves (tst = 5db) IBC Individual systems Fig. 6. DET curve obtained with IBC using the test sets (tst). The DET curves of the 6 individual systems are shown as well. L. Batista et al. / Expert Systems with Applications 40 (2013) 6788–6797 6795 the SVM providing the highest AUC is selected. To train the SVMs with RBF kernel, the following values were employed: c = ({2�4,2�3,2�2,2�1,20} and C = {2�5,2�4,2�3,2�2,2�1,20,21,22,23, 24,25}. Since the obtained ROC curves reached AUC close to 1, as indi- cated in Table 8, DET curves on a log–log scale are presented in- stead (see Fig. 3 (a)). Note that the curve representing system Table 12 Average EER ±r (%) on test data over 10 trials. tst IBC technique Majority vote Single best (S(nc=6)) 40 db 0.04 ± 0.04 0.06 ± 0.06 0.57 ± 0.27 30 db 0.00 0.01 ± 0.02 0.32 ± 0.16 20 db 0.11 ± 0.06 0.06 ± 0.06 0.10 ± 0.06 15 db 0.11 ± 0.04 0.10 ± 0.04 0.15 ± 0.06 10 db 0.29 ± 0.09 � 0.68 ± 0.26 5 db 0.33 ± 0.07 � 0.36 ± 0.08 Table 13 Additional error rates ±r (%) obtained with IBC over 10 trials. tst FNR (%) FPR (%) Average (%) Expected FPR = 1% 40 db 0.02 ± 0.02 10.70 ± 7.66 5.36 30 db 0.01 ± 0.01 8.61 ± 5.09 4.31 20 db 0.13 ± 0.15 5.34 ± 3.92 2.73 15 db 0.16 ± 0.09 6.48 ± 3.50 3.32 10 db 0.18 ± 0.11 8.94 ± 2.78 4.56 5 db 0.09 ± 0.10 21.87 ± 8.68 10.98 Expected FPR = 0.1% 40 db 0.03 ± 0.03 0.37 ± 0.65 0.40 30 db 0.01 ± 0.02 0.15 ± 0.33 0.08 20 db 0.18 ± 0.13 0.01 ± 0.01 0.09 15 db 0.23 ± 0.10 0.09 ± 0.12 0.16 10 db 0.36 ± 0.13 1.35 ± 0.62 0.85 5 db 0.15 ± 0.08 5.74 ± 1.21 2.94 Expected FPR = 0.01% 40 db 0.05 ± 0.04 0.10 ± 0.27 0.07 30 db 0.03 ± 0.04 0.03 ± 0.07 0.03 20 db 0.50 ± 0.31 0.00 0.02 15 db 0.56 ± 0.36 0.02 ± 0.03 0.29 10 db 1.60 ± 1.14 0.14 ± 0.10 0.87 5 db 0.28 ± 0.13 0.95 ± 0.44 0.61 Expected FPR = 0.001% 40 db 0.09 ± 0.09 0.11 ± 0.26 0.10 30 db 0.07 ± 0.14 0.03 ± 0.07 0.05 20 db 0.63 ± 0.38 0.00 0.31 15 db 0.94 ± 0.66 0.01 ± 0.02 0.47 10 db 3.38 ± 3.20 0.06 ± 0.07 1.72 5 db 0.48 ± 0.29 0.51 ± 0.53 0.49 Expected FPR = 0.0001% 40 db 0.09 ± 0.10 0.11 ± 0.26 0.10 30 db 0.10 ± 0.18 0.03 ± 0.07 0.06 20 db 0.66 ± 0.37 0.00 0.33 15 db 0.96 ± 0.68 0.01 ± 0.02 0.48 10 db 3.68 ± 3.56 0.06 ± 0.07 1.87 5 db 0.52 ± 0.34 0.46 ± 0.55 0.49 6796 L. Batista et al. / Expert Systems with Applications 40 (2013) 6788–6797 S(nc=1) does not appear in the graphic because a complete separa- tion of both classes was obtained. Fig. 4 shows the DET curves obtained on test data (tst) using the validation operating points. Observe that DET curves plotted in a same graphic are the results of a same system on different test data. Therefore, these curves are useful in order to analyse the robustness of each system regarding individual noise levels. It is worth noting that system S(nc=1) provided a complete class separa- tion for 40 db (that’s why the corresponding DET curve does not appear in the graphic), and, in a similar way, S(nc=2) and S(nc=3) pro- vided a complete class separation for 40 db and 30 db. Table 9 presents the average EER – as well as the standard devi- ation, r – obtained for each noise level during test, over 10 trials. The symbol ‘�’ indicates that the system has a random (or worse than random) behaviour for a given test set. A similar situation was observed in the work ofLazzerini and Volpi (2011), where clas- sification accuracies of 50% or less were obtained for high levels of noise. As expected, the systems become more robust to higher noise levels as they are gradually incorporated to the training phase. 4.2. Experiment 2 In the second experiment, IBC was used to combine the best classifier of each noise configuration, found in the first experiment. For all classifiers, a same validation set containing all noise levels (i.e., 40, 30, 20, 15, 10 and 5 db) was employed. Since a high num- ber of combinations is performed, the number thresholds per curve was limited to 500 (in the previous experiment, all validation scores were employed as thresholds). Fig. 5 shows the DET curve obtained with IBC, along with the DET curves of the six systems employed during the combination process. Note that IBC improved the Maximum Realizable DET (MRDET) curve of the individual systems. The operating points falling on the IBC curve are presented on Table 10. Each point is the result of a Boolean combination of dif- ferent individual classifiers. For instance, the operating point 5, which gives the EER, corresponds to a boolean combination (BC) of all 6 classifiers (cj,1 6 j 6 6), that is, BC{EER} = (c1 ^ c2 ^ c3 ^ c4 ^ - c5 ^ c6), using the decision thresholds indicated in Table 11. It is worth noting that the AND rule emerges most of the time in the IBC curve of Fig. 5. In the ideal case, when the classifiers are conditionally independent, and their ROC/DET curves are proper and convex, the AND and OR combinations are proven to be opti- mal, providing a higher performance than the original ROC curves (Khreich et al., 2010). Indeed, the datasets employed to design the proposed system are independent, randomly generated by using the simulator BEAT. Fig. 6 shows the DET curves obtained on test data using the IBC points indicated in Fig. 5, and Table 12 presents the average EER (over 10 trials) obtained with IBC, Majority vote and with the best single classifier. The Majority vote rule reached very low EER with respect to 40, 30, 20 and 15 db noise levels. On the other hand, a random behaviour was observed for 10 and 5 db noisy data. The reason is due to the fact that the majority of the individual classi- fers presents a random behaviour for high levels of noise. Observe that IBC provided an improvement for almost all test datasets with respect to the single best classifier obtained in the previous experiment. Finally, Table 13 presents additional results of IBC on test data, when the threshold is set in order to reach FPR (%) = {1,0.1,0.010.001,0.0001}. These intermediate points are obtained by using interpolation (Scott et al., 1998). Note that the FPR de- creases at the expense of an FNR increasing. In practice, the trade-off between FPR and FNR can be adjusted by the operators according to the current error costs. 5. Conclusion In this paper, a new system based on the fusion of classifiers in the ROC space was proposed in order to detect the presence of absence of bearing defects in noisy environments. Noise robust- ness was achieved through the incorporation of noisy vibration signals (ranging from 40 to 5 db) during the training phase, along with the Iterative Boolean Combination (IBC) of different SVMs, each one designed to deal with a specific noise configuration. In order to generate enough vibration signals, considering as well different defect dimensions, the BEAring Toolbox (BEAT) was employed. Experiments performed using time- and frequency- domain indicators (i.e., RMS, peak, Kurtosis, crest factor, impulse fac- tor, shape factor, BPFO, BPFI, 2BSF, and hamonics) indicated that the proposed system can significantly reduce the error rates, even in the presence of high levels of noise. Future work consist of validating the proposed strategy with real vibration signals. L. Batista et al. / Expert Systems with Applications 40 (2013) 6788–6797 6797 References Abbasion, S., Rafsanjani, A., Farshidianfar, A., & Irani, N. (2007). Rolling element bearings multi-fault classification based on the wavelet denoising and support vector machine. Mechanical Systems and Signal Processing, 21(7), 2933–2945. Alguindigue, I., Loskiewicz-Buczak, A., & Uhrig, R. (1993). Monitoring and diagnosis of rolling element bearings using artificial neural networks. IEEE Transactions on Industrial Electronics, 40(2), 209–217. Bhavaraju, K., Kankar, P., Sharma, S., & Harsha, S. (2010). A comparative study on bearings faults classification by artificial neural networks and self-organizing maps using wavelets (vol. 2, no. 5, pp. 1001–1008). Case Western Reserve University, Bearing Data Center. <http://csegroups.case.edu/ bearingdatacenter/>. Chang, C., & Lin, C. (2001). LIBSVM: a library for Support Vector Machines. In <http://www.csie.ntu.edu.tw/�cjlin/libsvm>. Cococcioni, M., Lazzerini, B., & Volpi, S. (2009a). Automatic diagnosis of defects of rolling element bearings based on computational intelligence techniques. International Conference on Intelligent Systems Design and Applications, 970–975. Cococcioni, M., Lazzerini, B., & Volpi S. (2009b). Rolling element bearing fault classification using soft computing techniques. In IEEE international conference on systems, man and cybernetics, 2009 (pp. 4926–4931). Elmaleeh, M., & Saad, N. (2008). Acoustic emission techniques for early detection of bearing faults using LabVIEW, in: Fifth international symposium on mechatronics and its applications (pp. 1–5). Fawcett, T. (2006). An introduction to ROC analysis. Pattern Recognition Letters, 27, 861–874. ISSN 0167-8655. Guo, H., Jack, L., & Nandi, A. (2005). Feature generation using genetic programming with application to fault classification. IEEE Transactions on Systems, Man, and Cybernetics, Part B, 35(1), 89–99. Jack, L., & Nandi, A. (2001). Support vector machines for detection and characterization of rolling element bearing faults. Journal of Mechanical Engineering Science, 215(9), 1065–1074. Jack, L., & Nandi, A. (2002). Fault detection using support vector machines and artificial neural networks, augmented by genetic algorithms. Mechanical Systems and Signal Processing, 16(2–3), 373–390. Kankar, P., Sharma, S., & Harsha, S. (2011). Fault diagnosis of ball bearings using continuous wavelet transform. Applied Soft Computing, 11, 2300–2312. Khreich, W., Granger, E., Miri, A., & Sabourin, R. (2010). Iterative Boolean Combination of classifiers in the ROC space: An application to anomaly detection with HMMs. Pattern Recognition, 43, 2732–2752. ISSN 0031-320. Lazzerini, B., & Volpi, S. (2011). Classifier ensembles to improve the robustness to noise of bearing fault diagnosis. In Pattern Analysis and Applications (pp. 1–17). Lei, Y., He, Z., Zi, Y., & Hu, Q. (2008). Fault diagnosis of rotating machinery based on a new hybrid clustering algorithm. The International Journal of Advanced Manufacturing Technology, 35, 968–977. ISSN 0268-3768. Liang, S., Hecker, R., & Landers, R. (2004). Machining process monitoring and control: The state-of-the-art. Journal of Manufacturing Science and Engineering, 126(2), 297–310. Lou, X., & Loparo, K. (2004). Bearing fault diagnosis based on wavelet transform and fuzzy inference. Mechanical Systems and Signal Processing, 18(5), 1077–1095. Purushotham, V., Narayanan, S., & Prasad, S. A. (2005). Multi-fault diagnosis of rolling bearing elements using wavelet analysis and hidden Markov model based fault recognition. NDT & E International, 38(8), 654–664. Rojas, A., & Nandi, A. (2006). Practical scheme for fast detection and classification of rolling-element bearing faults using support vector machines. Mechanical Systems and Signal Processing, 20(7), 1523–1536. Samanta, B., & Al-Balushi, K. (2003). Artificial neural network based fault diagnostics of rolling element bearings using time-domain features. Mechanical Systems and Signal Processing, 17(2), 317–328. Samanta, B., Al-Balushi, K., & Al-Araimi, S. (2003). Artificial neural networks and support vector machines with genetic algorithm for bearing fault detection. Engineering Applications of Artificial Intelligence, 16(7-8), 657–665. Samanta, B., Al-Balushi, K., & Al-Araimi, S. (2004). Bearing fault detection using artificial neural networks and genetic algorithm. EURASIP Journal on Applied Signal Processing, 366–377. Sassi, S., Badri, B., & Thomas, M. (2007). A numerical model to predict damaged bearing vibrations. Journal of Vibration and Control, 13(11), 1603–1628. Sassi, S., Badri, B., & Thomas, M. (2008). Tracking surface degradation of ball bearings by means of new time domain scalar descriptors. International Journal of COMADEM, 11(3), 36–45. Scott, M., Niranjan, M., & Prager, R. (1998). Realisable classifiers: Improving operating performance on variable cost problems. Sreejith, B., Verma, A., & Srividya, A. (2008). Fault diagnosis of rolling element bearing using time-domain features and neural networks. In Third international conference on industrial and information systems (pp. 1–6). Sugumaran, V., Muralidharan, V., & Ramachandran, K. (2007). Feature selection using decision tree and classification through proximal support vector machine for fault diagnostics of roller bearing. Mechanical Systems and Signal Processing, 21(2), 930–942. Sugumaran, V., Sabareesh, G., & Ramachandran, K. (2008). Fault diagnostics of roller bearing using kernel based neighborhood score multi-class support vector machine. Expert Systems with Applications, 34(4), 3090–3098. Tandon, N., & Choudhury, A. (1999). A review of vibration and acoustic measurement methods for the detection of defects in rolling element bearings. Tribology International, 32(8), 469–480. Teotrakool, K., Devaney, M., & Eren, L. (2008). Bearing fault detection in adjustable speed drives via a support vector machine with feature selection using a genetic algorithm. In IEEE instrumentation and measurement technology conference (pp. 1129 –1133). Thomas, M. (2011). Fiabilité, maintenance prédictive et vibration des machines. 9782760533578. Presses de l’Université du Québec (D3357). Tumer, K., & Ghosh, J. (1996). Analysis of decision boundaries in linearly combined neural classifiers. Pattern Recognition, 29(2), 341–348. Volpi, S., Cococcioni, M., Lazzerini, B., & Stefanescu, D. (2010). Rolling element bearing diagnosis using convex hull. In International joint conference on neural networks (pp. 1–8). Widodo, A., Kim, E., Son, J., Yang, B., Tan, A., Gu, D., et al. (2009). Fault diagnosis of low speed bearing based on relevance vector machine and support vector machine. Expert Systems with Applications, 36(3, Part 2), 7252–7261. http://refhub.elsevier.com/S0957-4174(13)00428-4/h0005 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0005 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0005 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0010 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0010 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0010 http://csegroups.case.edu/bearingdatacenter/ http://csegroups.case.edu/bearingdatacenter/ http://www.csie.ntu.edu.tw/~cjlin/libsvm http://www.csie.ntu.edu.tw/~cjlin/libsvm http://refhub.elsevier.com/S0957-4174(13)00428-4/h0015 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0015 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0015 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0020 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0020 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0025 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0025 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0025 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0030 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0030 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0030 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0035 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0035 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0035 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0040 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0040 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0045 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0045 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0045 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0050 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0050 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0050 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0055 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0055 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0055 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0060 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0060 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0065 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0065 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0065 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0070 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0070 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0070 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0075 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0075 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0075 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0080 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0080 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0080 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0085 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0085 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0085 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0090 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0090 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0095 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0095 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0095 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0100 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0100 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0100 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0100 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0105 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0105 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0105 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0110 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0110 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0110 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0115 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0115 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0120 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0120 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0125 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0125 http://refhub.elsevier.com/S0957-4174(13)00428-4/h0125 A classifier fusion system for bearing fault diagnosis 1 Introduction 2 The state-of-the-art in automatic bearing fault diagnosis 3 Methodology 3.1 Datasets 3.2 Performance evaluation methods 3.3 Iterative Boolean Combination (IBC) 4 Simulation results and discussions 4.1 Experiment 1 4.2 Experiment 2 5 Conclusion References 
work_gyqaanb5hjhwxh4biznmhtgdie	----	CSDL | IEEE Computer Society 
work_h3tr2rthlzcvjd25zrvvtxerdq	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220982273 Params is empty 220982273 exception Params is empty 2021/04/06-03:05:24 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220982273 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:24 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_h6qrlbjhb5hdnacqgfcwkwzr4a	----	Introduction 2 Fuzzy Data Envelopment Analysis: A Discrete Approach Majid Zerafat Angiz L.1 Ali Emrouznejad†2 Adli Mustafa3 1 Department of Mathematics, Islamic Azad University, Firouzkooh, Iran 2 Aston Business School, Aston University, Birmingham, UK 3 School of Mathematical Sciences, Universiti Sains Malaysia, Penang, Malaysia Abstract Data envelopment analysis (DEA) as introduced by Charnes et al [3] is a linear programming technique that has widely been used to evaluate the relative efficiency of a set of homogenous decision making units (DMUs). In many real applications, the input-output variables cannot be precisely measured. This is particularly important in assessing efficiency of DMUs using DEA, since the efficiency score of inefficient DMUs are very sensitive to possible data errors. Hence, several approaches have been proposed to deal with imprecise data. Perhaps the most popular fuzzy DEA model is based on -cut. One drawback of the -cut approach is that it cannot include all information about uncertainty. This paper aims to introduce an alternative linear programming model that can include some uncertainty information from the intervals within the α-cut approach. We introduce the concept of “local α-level” to develop a multi- objective linear programming to measure the efficiency of DMUs under uncertainty. An example is given to illustrate the use of this method. Keywords: Fuzzy data envelopment analysis; interval data; local α-level, multi objective programming, decision making unit. 1. Introduction Data envelopment analysis (DEA) initially was proposed by Charnes et al. [3] is a non-parametric linear programming technique for measuring the relative efficiency of a set of homogeneous decision making units (DMUs), with the common set of inputs and outputs. Examples of DEA include the efficiency of hospitals in providing their services [17] measurement efficiency of health † Corresponding author: Ali Emrouznejad, Operations and Information Management Group, Aston Business School, Aston University, Birmingham, United Kingdom a.emrouznejad@aston.ac.uk file:///D:/HD/RH12-PB/Subm/-%20rj%20-%20Zarafat-FuzzyDEA-DiscreteApproach/-rj-2009-06-09-resubmit%20in%20FSS/a.emrouznejad@aston.ac.uk 3 centers [12] manufacturing efficiency [23, 24] productivity of OECD countries [5, 6 ,7]. For some computational calculation of DEA methods see Emrouznejad [8] and for a recent theoretical survey and full list of applications of DEA see Cook and Seiford [4] and Emrouznejad et al. [9]. Traditionally, all input/output values of DMUs are crisp data, hence, most of the previous studies dealt with precise data. Theoretically, DEA measures the efficiency of each DMU by finding the distance of the DMU to the best practice; therefore, the efficiency scores are very sensitive to the data. If there is an outlier, then the efficiency scores of many DMUs may change substantially. Therefore, a key to the success of the DEA is to measure all inputs outputs accurately. However, in real application of production process many complicated factors are involved that makes difficult to measure inputs and outputs precisely. This makes a case where we need to measure the efficiency of DMUs with inexact values or interval data. Several approaches have been developed to deal with fuzzy data in DEA. Sengupta [26] applied principle of fuzzy set theory to introduce fuzziness in the objective function and the right-hand side vector of the conventional DEA model [3]. Guo and Tanaka [13] used the ranking method and introduced a bi-level programming model. Lertworasirikul [20] developed the method in which first, inputs and outputs were defazified and then the model was solved using the α- cut approach. Their method is simple but the uncertainty in inputs and outputs is effectively ignored. Some other approaches based on α-cut can be found in [22, 16, 25] and the methods based on the interval efficiency are seen in [10, 14]. Lertworasirikula [19] considered each constraint in the DEA as a fuzzy event; hence he transferred fuzzy DEA model to possibility linear programming problem. Lertworasirikul et al. [21] further developed a fuzzy BCC model where the possibility and credibility approaches are provided and compared with an α -cut level based approach for solving the FDEA models. Using the possibility approach, they revealed the relationship between the primal and dual models of fuzzy BCC. Using the credibility approach they showed how the efficiency value for each DMU can be obtained as a representative of its possible range. A different approach based on possibility programming was developed in [1, 18]. Inuiguchi and Tanino [15] applied the extension principle to define fuzzy efficiency score using DEA. In this paper we develop an alternative method which is able to provide fuzzy efficiency measures for DMUs with fuzzy observations. We introduce the 4 concept of local α-level which can include more information from uncertain data into the model. Use of local α-level in the proposed DEA model enables us to capture as much information as possible from the uncertain DMU while in the standard α-cut some fuzzy characteristics of DMUs are ignored [27-29]. The rest of this paper is organized as follows. DEA and Fuzzy DEA are defined in Section 2. Using concept of local α-level, an alternative fuzzy DEA is proposed in Section 3. Further discussion is given in Section 4. This is followed by a numerical example and comparison with other models in Section 5. Conclusion is given in Section 6. 2. DEA and fuzzy DEA Data envelopment analysis (DEA) is a non-parametric technique for measuring the relative efficiency of the decision making units (DMUs) that have homogenous inputs and outputs. DEA applies linear programming techniques to the observed inputs /outputs of DMUs by constructing an efficient production frontier based on the best practices. Each DMU's efficiency is then measured relative to its distance to this frontier [4]. Consider a set of n DMUs, in which ( 1, 2,..., ) ij x i m and ( 1, 2,..., ) rj y r s are inputs and outputs of j DMU (j=1,2,…,n). The standard form of CCR model for assessing p DMU is written as: (1) 1 1 1 1 max . . 1 0 , ,              s r rp r m i ip i s m r rj i ij r i r i u y s t v x u y v x u v r i The above model can only be used for cases where the data are precisely measured. Fuzzy DEA is a powerful tool for evaluating the performance of DMUs with imprecise data (or interval data). Fuzzy input-output variables can be introduced to DEA in the following fuzzy linear programming model. 5 (2) 1 1 1 1 max . . 1 0 , , s r rp r m i ip i s m r rj i ij r i r i u y s t v x u y v x j u v r i               where “~”indicates the fuzziness. rj y and ij x are fuzzy inputs and fuzzy outputs, respectively. 0 is a non-Archimedean small positive number. Saati et al. [25] suggested a different CCR model for assessment of fuzzy data by transferring the standard CCR model to a possibilistic programming problem. Their basic idea is using α-cut approach to transform the fuzzy CCR model (3) into a crisp linear programming problem such as the standard DEA model. Their proposed approach assumes that the solution lies in the interval and the result for each DMU is an interval efficiency score rather than a crisp efficiency score. The main drawback in this approach is that their model can not retain the uncertainty information completely since it is based on simple α-cut approach. In other words, the fuzzy numbers are simply converted to intervals using the same membership numbers in the entire of interval. 3. An alternative fuzzy DEA model under uncertainty The proposed approach in this paper is based on alternative definition of the membership functions of the coefficients, i.e. input output variables. Assume ( )F  is the set of fuzzy number. We define “local α-level” as follows. Definition 1: The crisp set of elements that belong to the fuzzy set M with degree of at least h and less than 1h  , in which 1h h   , is called the “local α- level” set and it is represented as follows:          1| ( )h h hMM x X x (3) One advantage of using a local α-level is that the corresponding point on the membership function retained the whole of uncertainty in process of solving the problem. We define the membership function .  M N based on  M and  N [29]. 6 Definition 2: If , ( ) M N F are two fuzzy numbers that ( ), ( ) N M x y  are continuous membership functions, then  . . ( ) sup min ( ), ( )     M N M N z x y z x y (4) Obviously, if the bounded variable v is considered as a trapezoidal fuzzy number, the following corollary is obtained. Corollary 1: The product of bounded variable v in fuzzy number M is defined as  . ( ) ( ), . ,   v M Mz x z v x x X (5) In most DEA applications for the sake of computational efficiency and ease of data acquisition, trapezoidal or triangular membership functions are often used. Figure 1a: Triangular membership function Figure 1b: concept of local α-level Figure 1a shows a triangular “fuzzy number” and Figure 1b shows the local α-level for a fuzzy number which is corresponding with the figure 1a. As seen, each α-cut acts in a local determined domain. For instance, the corresponding domain of l is [ , ] [ , ]l l l lx y x y    . The range of each local α-level in the figure 1b has shown as the horizontal bold lines. Theorem 1. Let ),,( ~ ulm xxxA  be a symmetric triangular fuzzy number such that ),...,2,1( ni i  are local α-levels, thus, nn xxxx 221 ,...,,...,, make a partition on the interval of fuzzy number A ~ . Then, for each of the data point ],[ 21 n xxx  we have xxxxxx ii n i A m ii n i A    )()( 2 1 ~ 2 1 ~  (6) Proof. To prove this theorem, we rearrange (6) to 1 1 22 3 3 ll 1h  h h 1h  l l x y  l l x y  7 0))(( 2 1 ~   xxxxx i m ii n i A  (7) Suppose that ))(( 2 1 ~ xxxxxB i m ii n i A     Two cases are considered Case 1. m kk xxxx  1 In this case, B can be written in the following form. ))())((( )())((())())((( 2 1 ~ ~ 1 1 ~ xxxxx xxxxxxxxxxB i m n ni iiA i n ki i m iAii m i k i A           (8) Since A ~ is symmetric, therefore nlxx lnAlA 21)()( )1(2~~   . Thus, the right side of fuzzy number A ~ is divided to two areas as follows: ))())((( ))())((())())((( ))1(2())1(2())1(2( ~ ))1(2( 1 1 ))1(2())1(2( ~ 2 1 ~ xxxxx xxxxxxxxxx in m n ki ininA in m k i ininAi m n ni iiA               (9) Combining (8) and (9) we obtain ))]()(())())[((( ))]()(())())[((( )1(2())1(2( ~ )1(2()1(2( 1 1 ~ xxxxxxxxx xxxxxxxxxB in m ini n ki i m iA in m inii m i k i A            (10) In equation (10) we have )22)((0 ~ xxxB ii n ki A     8 Since xx i  for ki  , therefore, .0B Case 2. nxxxx kk m 2 1    Similarly, it can be proved that .0B ■ Hence, xxx ii n i Ax   )(min 2 1 ~ in this method is obtained the maximal degree of membership function which does not depend on number of partitions. Corollary 2. Suppose ,, xx i )(~ iA x satisfy the conditions of Theorem 1 then xxx ii n i Ax   )(min 2 1 ~ is unique. Using the concept of local α-level in model (2) the following fuzzy linear programming is proposed for assessing p DMU . (11)                                 2 1 1 1 1 min ( ) , min ( ) , max . . 1 0 , , , , ij rj ij x ijh i ij i ijh h rj y rjk r rj r rjk k s r rp r m i ip i s m r rj i ij r i l u ij ij ij l u rj rj rj r i Z x v x v x i j Z y u y u y r j Z u y s t v x u y v x j x x x i j y y y r j u v r i As seen in figure 1b, corresponding to each α, a set of subintervals is assigned to each fuzzy number. Theorem 2. The above model is an extension to proposed model of Kao and Liu [16], if we remove the two minimization distance functions from the objective of 9 the model, the largest value of optimal solution in each α-level is the same as those obtained in Kao and Liu [16]. Proof. Consider the following model: (12) 1 1 1 1 max . . 1 0 (1 ) (1 ) (1 ) (1 ) , 0 , s r rp r m i ip i s m r rj i ij r i m l m u rj rj rj rj rj m l m u ij ij ij ij ij r i u y s t v x u y v x y y y y y x x x x x u v r i                                 For α=0, consider n DMUs with m inputs and r outputs as follows: u l ij ij io ip l u rj rj ro rp x x j o x x y y j o y y       It is clear that for o DMU each dominated DMU (a DMU with higher level of inputs and lower level of outputs) will not be more efficient. So optimal value of mathematical programming (12) is equals to efficiency of p DMU . This is also valid for any arbitrary . In model (11), ijhx corresponds to the length of input value of xij located in the intersection of h  and sides of the corresponding triangular membership function. Similarly rjky corresponds to the length of output value of yij. Consider the following variable substitutions. ˆ ijx = i ijv x , ˆ rjy = , ,r rju y i r j 10 Hence, model (13) is concluded. (13)                                 1 1 1 1 ˆmin ( ) , ˆmin ( ) , ˆmax . . ˆ 1 ˆ ˆ 0 ˆ , ˆ , , , ij rj ij x ijh ij i ijh h rj y rjk rj r rjk k s p rp r m ip i s m rj ij r i l u i ij ij i ij l u r rj rj r rj r i Z x x v x i j Z y y u y r j Z y s t x y x j v x x v x i j u y y u y r j u v r i Assume that ˆ , ( , , ) ijh ij i ijhx x v x i j h    , so ( , , )ijh ijh ijhx x x i j h      and ˆ rjk rj r rjky y u y   , so ( , , )rjk rjk rjky y y r j k      . Applying these substitutions model (13) may be solved using the following multi-objective programming. (14)                                         1 1 1 1 min ( )( ) , min ( )( ) , ˆmax . . ˆ 1 ˆ ˆ 0 ˆ ( ) 0 , , ˆ -( ) 0 , , ij rj ij x ijh ijh ijh h rj y rjk rjk rjk k s p rp r m ip i s m rj ij r i ij i ijh ijh ijh rj r rjk rjk rjk l i ij Z x x x i j Z y y y r j Z y s t x y x j x v x x x i j h y u y y y r j k v x          ˆ , ˆ , , , u ij i ij l u r rj rj r rj r i x v x i j u y y u y r j u v r i 11 The model (14) is a multi-objective, hence, it can be solved by Archimedean goal programming model (15) as follows: (15)                                               1 1 1 1 1 1 1 1 min . . ( )( ) , ( )( ) , ˆ ˆ 1 ˆ ˆ 0 ˆ ( ij rj m n s n ij ij rj rj p p i j r j x ijh ijh ijh ij ij h y rjk rjk rjk rj rj k s rp p p r m ip i s m rj ij r i ij i ijh ijh w d w d w d s t x x x d t i j y y y d t r j y d t x y x j x v x x                    ) 0 , , ˆ -( ) 0 , , ˆ , ˆ , , , ijh rj r rjk rjk rjk l u i ij ij i ij l u r rj rj r rj r i x i j h y u y y y r j k v x x v x i j u y y u y r j u v r i In model (15), the w’s in the objective function are positive penalty weights and d’s measure the over-achievement and under-achievement from the target point t, i.e. ijt , rjt , pt . 4. Discussion In Figure 2a, consider the local α-level  1 . The membership values corresponding to interval 1 2 [ , ]n n are approximated by  1 . For instance, the membership value related to x is  1 instead of ( )x that is real membership value. Furthermore, assume that we include another local α-level  2 (See Figure 2b), it is seen that the membership function corresponding to x is now  2 12 instead of ( )x . It is clear that   1 2 ( )x and  2 is a better approximation than  1 for ( )x . According to Theorems 2, if the objective functions minimizing in (13) are deleted from the model, the optimal solution for inputs and outputs will be arisen at its endpoints of interval of fuzzy numbers. Furthermore, if the objective function maximizing in (13) is eliminated, Theorem 1 is adopted and its optimal solutions are fuzzy number. Figure 3 illustrates the above mentioned concept for evaluating P DMU . In this figure, the interior arrows represent the optimal solution when the objective function of maximizing is absent in (13) and the arrows located under fuzzy numbers construct the optimal solution (13) when only objective function of maximizing is present. Interaction between the objective functions of maximizing and minimizing in (13) cause the fuzzy optimal solution. In the methods based on α-level approach, all fuzzy numbers are dealt with the same level. Consider the example given in Kao and Liu [16] evaluating four DMUs with single input and single output. In total, 88 LPs should be solved using the α-cuts of the efficiency scores for eleven α values, they assumed that all fuzzy inputs and outputs are in the same α-level. In the proposed method different level of α’s are considered for inputs and outputs, hence we need to solve 11×11×88 LPs. Figure 2b: A presentation of local two α-levels Figure 2a: A presentation of local one α-level ij x~ ip x~ rp y~ 22  1 x 1 n 2 n 1 n 2 n  2 ( )x  1 x ( )x 13 5. An Application and comparison with other methods Consider 5 DMUs with two triangular fuzzy inputs and 2 triangular fuzzy outputs. Table (1) shows the data which are also used in Guo and Tanaka [13]. Table 1. Data for numerical example DMU Variable D1 D2 D3 D4 D5 I1 (4.0, 3.5, 4.5) (2.9, 2.9, 2.9) (4.9, 4.4, 5.4) (4.1, 3.4, 4.8) (6.5, 5.9, 7.1) I2 (2.1, 1.9, 2.3) (1.5, 1.4, 1.6) (2.6, 2.2, 3.0) (2.3, 2.2, 2.4) (4.1, 3.6, 4.6) O1 (2.6, 2.4, 2.8) (2.2, 2.2, 2.2) (3.2, 2.7, 3.7) (2.9, 2.5, 3..3) (5.1, 4.4, 5.8) O2 (4.1, 3.8, 4.4) (3.5, 3.3, 3.7) (5.1, 4.3, 5.9) (5.7, 5.5, 5.9) (7.4, 6.5, 8.3) Source: Guo and Tanaka (2001) Fuzzy efficiencies of DMUs using standard fuzzy DEA model (2) and with different α value solved by the method suggested in [24] is reported in Table (2). Table 2.the results of the model suggested in [24] DMU α D1 D2 D3 D4 D5 0 1.0 1.0 1.0 1.0 1.0 .5 0.995 1.0 1.0 1.0 1.0 .75 0.906 1.0 0.936 1.0 1.0 1 0.85 1.0 .86 1.0 1.0 Using the Kao and Liu’s approach, the cut  of the efficiency scores for four  values is presented in Table (3). Table 3. Table of efficiency using Kao & Liu fuzzy DEA model DMU α D1 D2 D3 D4 D5 0 (0.626,1.0) (0.835,1.0) (0.575,1.0) (0.855,1.0) (0.737,1.0) .5 (0.757,0.995) (0.988,1.0) (10,1.0) (10,1.0) (0.845,1.0) .75 (0.808,0.906) (1.0,1.0) (0.792,0.936) (1.0,1.0) (0.971,1.0) 1 0.85 1.0 .86 1.0 1.0 The results of the possibility approach introduced in [19] have given in Table (4). Table 4. Table of efficiency using Lertworasirikul’s possibilistic model [19] DMU α D1 D2 D3 D4 D5 0 1.107 1.238 1.276 1.52 1.296 0.25 1.032 1.173 1.149 1.386 1.226 0.5 0.963 1.112 1.035 1.258 1.159 0.75 0.904 1.055 0.932 1.131 1.095 1 0.855 1.000 0.861 1.000 1.000 14 Although the Lertworasirikul’s possibilistic model is not directly comparable with the new suggested method, the result of Table (4) is aggregated measure in a linear approach and an optimal cut  is obtained for each fuzzy numbesults to make it comparable. Fuzzy efficiencies of DMUs using the proposed model and with different α values are reported in Table (5). Table 5. Table of efficiency using proposed fuzzy DEA model α D1 D2 D3 D4 D5 0,0.25,0.5,0.6,0,7,0.75,1 0.915 1.0 0.948 1.0 0.991 0,0.25,0.5,0.7,0.75,1 0.909 1.0 0.945 1.0 0.991 0,0.5,0.75,1 0.903 1.0 0.941 1.0 0.84 0 1.0 1.0 1.0 1.0 1.0 The results shown in the first row of Table 4 are more accurate than the results suggested in [24], as seen in Theorem 2, Kao and Liu [16] model obtained the results at endpoints of the intervals only. Figure 4 shows the structure of the discrete triangular number concerned to input 1 of 1 DMU . As it can be seen we included 7 local α-levels for calculation of efficiency in this example.  h  0, 0.25, 0.5, 0.6, 0.7, 0.75 and 1. rj y~ Figur e 3: opti mal solut ions at end point inter vals 3.50 3.63 3.75 3.80 3.85 3.88 4 4.13 4.15 4.20 4.25 4.36 4.5 Figu re 4. Disc rete fuzz y num ber h x 11 0.25 0.5 0. 6 0.7 0.75 15 It should be noted that increasing the numbers of local α-levels will result more precise measure of efficiency. For example when measuring efficiency of DMU1 if α-levels are [0, 0.5, 0.75, 1] we obtain efficiency score of “0.909”, while if α-levels are [0, 0.25, 0.5, 0.7, 0.75, 1] we obtain efficiency score of 0.903 which is a more accurate estimation of efficiency. Comparing these results with Table (2) it is clear that our proposed approach is given a better estimation of the efficiency scores when data are in the form of interval values. One drawback of the proposed model is longer computational calculation; however this is not a major issue with development of high-speed computers. 6. Conclusion In evaluating DMUs in Fuzzy DEA there are four traditional approaches; the fuzzy ranking approach, the defuzzification approach, the tolerance approach and the α-cut based approach. Each of these methods has its advantages and drawbacks in the way they treat uncertain data in DEA models. Perhaps due to its simplicity, the α-cut based approach is frequently used by DEA scholars. As result of simplicity in this approach we will lose a lot of uncertainty information. This paper proposed an alternative fuzzy DEA technique for measuring efficiency of decision making units under fuzzy environment using local α-level concept and linear programming problem. The numerical example showed a better estimation of efficiency when using the proposed model. The final model presented is a multi-objective programming, its transformation to a linear programming and developing an efficient algorithm for large scale problems are subjects for future development. References [1] I. Alp, Possibilistic Data Envelopment Analysis, Mathematical and Computational Applications. 7 (2002) 1: 5-14. [2] B. Asady, A. Zendehnam, Ranking fuzzy numbers by distance minimization, Applied Mathematical Modeling. 31 (2007): 2589-98. [3] A. Charnes, W.W. Cooper, E. Rhodes, Measuring the efficiency of decision making units, European Journal of Operation Research. 2(1978): 429-44. [4] W. D. Cook, L. Seiford, Data envelopment analysis (DEA) – Thirty years on, European Journal of Operational Research, (2008) in press: doi:10.1016/j.ejor.2008.01.032 [5] A. Emrouznejad, An alternative DEA measure: A case of OCED countries, Applied Economic Letters. 10 (2003): 779-782. 16 [6] A. Emrouznejad, E. Thanassoulis, A mathematical model for dynamic efficiency using data envelopment analysis, Journal of Applied Mathematics and Computation. 160 (2005) 2: 363-378. [7] A. Emrouznejad, E. Thanassoulis, Measurement of productivity index with dynamic DEA, International Journal of Operational Research (2010), in press. [8] A. Emrouznejad, Measurement efficiency and productivity in SAS/OR, Journal of Computers and Operation Research. 32 (2005) 7: 1665-1683. [9] A. Emrouznejad, B. Parker, G. Tavares, Evaluation of research in efficiency and productivity: A survey and analysis of the first 30 years of scholarly literature in DEA, Socio-Economic Planning Sciences. 42 (2008) 3: 151-157. [10] T. Entani, Y. Maeda, H. Tanaka, Dual models of interval DEA and its extension to interval data, European Journal of Operational Research. 136 ( 2002) 1: 32-45. [11] R. E. Esteuer, Multiple criteria optimizing: theory, computation and application, John Wiley & Sons, Inc 1976. [12] K. Field, A. Emrouznejad, Measuring the Performance of Neonatal Care Units In Scotland, Journal of Medical Systems. 27 (2003) 4: 315-324. [13] P. Guo, H. Tanaka, Fuzzy DEA: A perceptual evaluation method." Fuzzy Sets and Systems. 119 (2001) 1: 149-160. [14] J. L. Hougaard, A simple approximation of productivity scores of fuzzy production plans, Fuzzy Sets and Systems. 152 (2005) 3: 455-465. [15] M. Inuiguchi and T. Tanino, Data Envelopment Analysis with Fuzzy Input-Output Data, Y. Haimes and R. E. Steuer: Research and Practice in Multiple Criteria Decision Making, 296-307, Springer (2000). [16] C. Kao, S. T. Liu, Fuzzy efficiency measures in Data Envelopment Analysis, Fuzzy Sets and Systems. 113 (2000): 427–437. [17] J. M. Kirigia, A. Emrouznejad, R. G. Vaz, H. Bastiene, J. Padayachy, A comparative assessment of performance and productivity of health centres in Seychelles, International Journal of Productivity & Performance Management. 57 (2008) 1. [18] T. Leon, V. Liern, J. L. Ruiz, I. Sirvent, A fuzzy mathematical programming approach to the assessment of efficiency with DEA models, Fuzzy Sets and Systems. 139 (2003): 2 407- 19. [19] S. Lertworasirikul, S. C. Fang, A. Jeffrey, J. A. Joines, H. L. W. Nuttle, Fuzzy Data Envelopment Analysis (DEA): a possibility approach, Fuzzy Sets and Systems. 139 (2003): 2 379-94. [20] S. Lertworasirikul, Fuzzy Data Envelopment Analysis for Supply Chain Modeling and Analysis, Dissertation Proposal in Industrial Engineering, North Carolina State University. 2001. [21] S. Lertworasirikul, S. C. Fang, H. L. W. Nuttle, J. A. Joines, Fuzzy BCC model for data envelopment analysis, Fuzzy Optimization and Decision Making 2 (2003): 337-358. 17 [22] Y. Meada, T. Entani, H. Tanaka, Fuzzy DEA with interval efficiency, Proceedings of 6th European Congress on Intelligent Techniques and Soft Computing. EUFIT '98. Aachen, Germany, Verlag Mainz. 2 (1998): 1067-71. [23] R. Mulwa, A. Emrouznejad, L. Muhammad, Economic efficiency of smallholder maize producers in western Kenya: A DEA meta-frontier analysis, International Journal of Operational Research. 6 (2008): 3 In Press. [24] R. Mulwa, A. Emrouznejad, F. M. Murithi, Impact of liberalization on efficiency and productivity of sugar industry: The case of Mumias sugar company in Kenya, Journal of Economic Studies. 2008 in Press. [25] S. Saati Mohtadi, S. A. Memariani, G. R. Jahanshahloo, Efficiency analysis and ranking of DMUs with fuzzy data, Fuzzy Optimization and Decision Making. 1 (2002) 255-267. [26] J. K. Sengupta, A fuzzy systems approach in Data Envelopment Analysis, Computers and Mathematics with Applications. 24 (1992) 8: 259-66. [27] M. Zerafat Angiz, A. Emrouznejad, A. Mustafa, A. Rashidi Komijan, Selecting the most preferable alternatives in a group decision making problem using DEA, Expert Systems with Applications. 36 (2009) 9599–9602. [28] M. Zerafat Angiz, A. Emrouznejad, A. Mustafa, A. Fuzzy assessment of performance of a decision making units using DEA: A non-radial approach, Expert Systems with Applications. 37 (2010) 5153–5157. [29] M. Zerafat Angiz, A. Emrouznejad, A. Mustafa, A. S. al-Eraqi (2010) Aggregating Preference Ranking with Fuzzy Data Envelopment Analysis, Knowledge-Based Systems, In Press. [30] H. J. Zimmermann Fuzzy set theory and its applications, Kluwer Academic Publishers. 2001. 
work_h7puv6wucjfeddwp6khk7y4mb4	----	sfunctions.eps Followee recommendation in Twitter using fuzzy link prediction Fernando M. Rodŕıguez†, Luis M. Torres‡, Sara E. Garza∗ School of Mechanical and Electrical Engineering (FIME) Universidad Autónoma de Nuevo León (UANL) San Nicolás de los Garza, NL, Mexico †fernando.aldape@gmail.com, ‡luis.torres.ciidit@gmail.com, ∗sara.garzavl@uanl.edu.mx Abstract In social networking sites, it is useful to receive recommendations about whom to contact or follow. These recommendations not only allow to es- tablish connections with people one might already know in real life, but also with people or users that have similar interests or are potentially interesting. We propose an approach that tackles contact (followee) recommendation in Twitter by means of fuzzy logic. This fuzzy approach handles recommenda- tion as a link prediction problem and uses three types of similarity between a pair of users: tweet similarity, followee id similarity, and followee tweet sim- ilarity. These similarities are calculated by extracting user profiles. These profiles are, in turn, obtained by considering Twitter as a heterogeneous in- formation network. To test our approach, we crawled a repository of 6,000 users and 2 million tweets, and we measured accuracy by comparing our re- sults with the actual followee lists of the users. These results, which are also compared against the results given by state-of-the-art methods, show a high accuracy. Other advantages of the fuzzy system include a self-explanatory capability and the ability to produce a non-binary friendship value. Keywords: fuzzy systems, recommender systems, Twitter, link predic- tion, expert systems, artificial intelligence 1 Introduction Social networking sites (SNS’s) play a key role in our daily lives. These sites not only enable users to publish and receive information they consider to be of interest, but also allow them to keep in contact with friends, or even to establish new relationships. Social networking sites, ultimately, serve as a platform for personal or social expression and organization — hence their importance for areas such as marketing, human-computer interaction, and research. 1 Due to the increasingly intensive use of SNS’s, it is common for users to be overloaded with information (Benito-Ruiz, 2009; Gantz and Reinsel, 2010), e.g. status updates to read or possible new friends to contact. For this reason, recommender systems have been adopted to filter content, activities, products, and social connections of actual interest for the user (Park et al., 2012). An important representative of SNS’s nowadays — with 190 million users, 60,000 daily comments, and ranking ten in Alexa’s Top Sites1 — is Twitter2. This site offers a microblogging service where users are able to publish short messages (140 characters maximum per post) called “tweets” and receive the posts of other users that they decide to follow, where follow- ing a user is similar to subscribing to an RSS news feed; the ones who follow a user are called the user’s followers and the ones followed by a user are called the user’s followees (in this work, we will use “contacts” and “friends” as synonyms of this term). Other distinctive features of Twitter include post forwarding, where the forwarded message is called retweet, user mentions in messages (where each user account starts with “@”), and a special form of message labeling, where the label is called hashtag and usually starts with “#”. Twitter, of course, is not exempt of the previously mentioned infor- mation overload and crowding. Recommendation approaches for this SNS so far include automatic suggestions of tweets, URL’s, hashtags, mentions, retweets, and followees (Kywe et al., 2012). We focus on this last recom- mendation item, since helpful followee recommendations aid users to receive actual content of interest, to get acquainted with people of interest, and to build stronger communities — either in Twitter or real life. Because the information in Twitter tends to be imprecise, incomplete, and could highly depend on the context where it is employed, we propose a fuzzy hybrid recommendation system. The rules of this system create rec- ommendations by predicting follower-followee relationships (which we shall refer to as “friendships”) between pairs of users given the similarity of their tweets, their followees, and the tweets of their followees. These similarities are derived from user profiles whose information is retrieved by considering Twitter as a heterogeneous information network. Our contributions thus include: - A fuzzy expert system that acts as a hybrid recommender system. 1Information available at: http://www.alexa.com/topsites. This list of Top Sites was retrieved in March 2014. 2Available at: http://twitter.com. 2 This fuzzy system predicts the degree of friendship between a pair of Twitter users and uses this degree as the basis for recommendation. - A formal conceptual model for extracting user profiles by considering Twitter as a heterogeneous network. - Evidence for the effectiveness of the approach in different contexts (Twitter, the Cit-HepTh citation network). - A comparison with state-of-the-art methods. The rest of this document is organized as follows: Section 2 reviews nec- essary background and Section 3 discusses related work; Section 4 describes our approach and Section 5 provides experiments and results; Section 6, finally, presents conclusions and future work. 2 Background This section introduces vocabulary and notation that will be used. It covers basic notions for recommender systems, graph theory, information retrieval, and fuzzy logic. 2.1 Recommender systems In the midst of overwhelming amounts of information, the aim of a recom- mender system is to suggest users those items that might be of interest, such as movies, records, books, and other users. Most recommender systems are personalized and build a profile to make suggestions based on the informa- tion and particular preferences of the target user (i.e., the recipient of the recommendations). Common approaches for this kind of recommendation include content-based and collaborative filtering systems; while the former find items that are similar to the ones highly rated by the user in the past, the latter find items that were highly rated by users that are similar to the target user (Ricci et al., 2011). In that sense, content-based systems are item-oriented and collaborative filtering systems are user-oriented. The usual output of a recommender system is a ranked list of items, where the items at the top — ideally — are the most likely to achieve high ratings by the target user. To build this list, several recommendation algo- rithms start off with a set of candidate items and gradually refine this set. When talking about friend or contact recommendations, the term “item” is usually replaced by “user”. Consequently, it is common to refer to candi- date users instead of candidate items and to recommended users instead of 3 recommended items; let us note that the latter are the ones included in the (final) recommendation list. Another important aspect of friend recommen- dation is that this task is sometimes stated as a link prediction problem, i.e. the problem of determining if a connection will appear between a pair of entities in a network (Getoor and Diehl, 2005). 2.2 Networks Networks depict related entities and are formally represented with graphs. A graph G = (V, E) consists of a set V = {v1, v2, . . . vn} of vertices and a set E of edges. While the former set represents the entities of the network, the latter stands for the relationships or connections among these entities; E ⊆ V ×V , typically3. If E is symmetric, i.e. the existence of an edge (u, v) implies the existence of (v, u) as well, then all edges run in both directions and the graph is said to be undirected; being this the case, an edge can be formally written as a 2-subset {u, v}. However, if the property of symmetry does not hold, the graph is directed and (u, v) ∈ E depicts an edge that goes out from u into v. In a directed graph, edges are commonly referred to as arcs and are drawn as arrows. When the vertices or edges have labels, the graph is said to be labeled, and an edge-labeled graph where the labels are numerical values is said to be weighted (it is otherwise known as an unweighted graph whose edges have indistinct values). A subgraph Gs = (Vs, Es) is a portion of a graph, such that Vs ⊆ V and Es = {{u, v} : u, v ∈ Vs}. The neighborhood of a vertex v (denoted by Γ(v)) is the set of vertices which share an edge with v, i.e. Γ(v) = {u : {u, v}∈ E} . For a directed graph, Γ(v) = Γ+(v)∪Γ−(v), where Γ+(v) is the out-neighborhood and Γ−(v) is the in-neighborhood. While the former consists of the vertices whose arcs go out from v, the latter includes vertices whose arcs go into v: Γ+(v) = {u : (v, u) ∈ E} , (1) Γ−(v) = {u : (u, v) ∈ E} . (2) In several contexts, such as scientific coauthorship, transportation, and recommendation (Newman, 2010), V contains vertices of different types (pa- pers and authors / buyers and products / users, resources, and tags). This can be formally represented with a k-partite graph, where V is composed of k disjoint subsets and the edges in E join pairs of vertices that belong to different subsets. When k = 2, the graph is said to be bipartite and thus 3Edges usually consist of vertex pairs, but three or more vertices can be simultaneously joined by a single edge. In this case, the edge is referred to as hyperedge and the graph is called hypergraph. 4 V = V1 ∪V2, (3) V1 ∩V2 = ∅, and (4) ∀e = {u, v} , e ∈ E ⇒ (u ∈ Vi)∧ (v ∈ Vj), i 6= j, (5) where Vi, Vj ∈{V1, V2}. Examples of bipartite graphs include author-paper, actor-movie, metabolite-reaction, and buyer-product networks. More complex situations can be modeled with heterogeneous information networks (Sun et al., 2012, 2009; Wang et al., 2012; Ji et al., 2011), which not only take into account different vertex types but also consider different edge types (directed, undirected, weighted, . . . ) and connections between arbitrary pairs of vertices (including the same type). Formally, for any heterogeneous graph V = V1 ∪ . . .∪Vn = n ⋃ i=1 Vi and (6) E ⊆{(u, v) : (u ∈ Vi)∧ (v ∈ Vj), i, j ∈{1, . . . n}} . (7) An example of a heterogeneous graph is the roll call data network pre- sented by Wang et al. (2012), which consists of two vertex types and three edge types; while the vertices either stand for legislators or bills, the edges either join legislator pairs (cosponsorship), bill pairs (semantic similarity), or legislator-bill pairs (this last edge type is directed and represents votes). 2.3 The Vector Space Model The vector space model is, so far, one of the central models for information retrieval (Liu, 2007). This model views each document as a bag of words (a representation where order is not important) and extracts a weight vector from this bag. Each vector’s length is equal to the vocabulary (i.e. unique words) of the whole document collection. The weight wi,j for a given word ki in a particular document dj is usually awarded by a text frequency–inverse document frequency or tf-idf score; this score indicates the importance of ki in dj: 5 wi,j = tfi,j × idfi idfi = log N ni tfi,j = fi,j M ∑ x=1 fx,j , (8) where fi,j is the number of times that ki appears in dj, M is the amount of unique words in dj, N is the amount of documents in the collection, and ni is the number of documents that ki appears in. A common metric for calculating similarity between document vectors is the cosine similarity (Baeza-Yates and Ribeiro-Neto, 1999), which is based on the dot product of the vectors: cosim(da, db) = da � db |da|× |db| , (9) where da and db are the documents, and da and db are the vectors. A similarity of 0 indicates that the documents have no common words and a similarity of 1 indicates that the documents are identical. 2.4 Fuzzy logic systems as linguistic processors Fuzzy systems encode knowledge, which is commonly expressed as linguistic information. These systems are based on fuzzy logic and were originally proposed by Zadeh (1965). A fuzzy system maps a set of given inputs to an output by means of an inference engine that uses a fuzzy rule base. To perform inferences with this rule base, the inputs have to be fuzzified, and the fuzzy result is defuzzified to obtain the corresponding output. This process is illustrated in Figure 3. The core of fuzzy system design is given by fuzzy sets, linguistic vari- ables, and membership functions. A fuzzy set assigns a membership value to every element, as opposed to a conventional set where elements are either present or absent; e.g., in the set Old = {(45, 0.1), (65, 0.5), . . . (90, 0.9)}, the element 90 has a membership value of 0.9. A membership value — whose range is between 0 and 1, inclusive — is assigned by a membership function. There are several forms of membership functions; however, the triangular and trapezoidal forms are commonly used: triangular(x, a, b, c) = max ( min ( (x − a) (b − a + ǫ) , (c − x) (c − b + ǫ) ) , 0 ) , (10) 6 trapezoidal(x, a, b, c, d) = max ( min ( min ( (x − a) (b − a + ǫ) , 1 ) , (d − x) (d − c + ǫ) ) , 0 ) , (11) where x is a given element of the fuzzy set and ǫ = 1 × 10−6 or another suitable constant. Parameters a, b, c, and d are additionally illustrated in Figures 1a and 1b. (a) Triangular (b) Trapezoidal Figure 1: Membership functions and parameters. A linguistic variable is a variable associated with a numeric variable x and does not take numbers as values, but instead takes linguistic values. For example: Volume = {“Low”, “High”, “Very high”, . . .} . Each linguistic value is related to a fuzzy set or membership function. A range of operation is defined for every linguistic variable and it is partitioned considering the linguistic values used. Usually, an odd number of linguistic partitions is used; three and five are the most common (see Figure 2). As we previously mentioned, a fuzzy system is an integration of the logic operations shown in Figure 3. Because we can use different membership functions, fuzzifiers, inference engines, and defuzzifiers, there is a variety of fuzzy systems. In this paper, as we will see later, we used a fuzzy system with a singleton fuzzifier, trapezoidal and triangular membership functions, a Mamdani max-min engine, and a centroid defuzzifier. Let us explain each of these. The singleton fuzzifier maps a value a into a fuzzy set A′ by setting membership µA′(x) to 1 at x = a and 0 at any other value: µA′(x) = { 1 if x = a 0 otherwise. (12) 7 Figure 2: (a) Five fuzzy sets or linguistic values: 1=“Very Low”, 2=“Low”, 3=“Medium”, 4=“High”, 5=“Very High”, (b) Three fuzzy sets or linguistic values: 1=“Low”, 2=“Medium”, 3=“High” Figure 3: Components of a fuzzy system. With respect to the inference engine and the fuzzy rule base, in first place, the fuzzy rule base is a matrix (which we shall denote by DB) where every row is a rule and every column is an integer that represents the lin- guistic value associated with that rule from antecedents to consequents. In the case of three variables and five linguistic values, a maximum of 53 = 125 rules or rows and 4 columns (three inputs and one output) build up the ma- trix. Second, considering a normalized interval for every linguistic variable, a type1(x, n) function can be defined as follows: type1(x, n) =            trapezoidal(x, 0, 0, 0.16, 0.33) if n = 1 triangular(x, 0.16, 0.33, 0.5) if n = 2 triangular(x, 0.33, 0.5, 0.66) if n = 3 triangular(x, 0.5, 0.66, 0.83) if n = 4 trapezoidal(x, 0.66, 0.83, 1, 1) if n = 5, (13) 8 where x is an input value and n corresponds to a specific membership func- tion of a linguistic variable. Considering a fuzzy system of three inputs, a max-min inference engine is defined as follows: I(r) = min(type1(x1, DB(r)), type1(x2, DB(r)), type1(x3, DB(r))) (14) where r is the fuzzy rule number, DB is the fuzzy rule base (as already stated), and I is the inference calculated. Finally, the output is defuzzified as follows: y = ∑R r=1(I(r)×ym(DB(r))) ∑R r=1 I(r) , (15) where ym is a precalculated center of mass of every membership function of the output (the predefined values are ym =[0, 0.25, 0.5, 0.75, 1]) and R is the number of rules. DB(r) establishes the correct value depending of the fuzzy rule. 3 Related Work As we will see throughout this section, related work is centered on followee recommendation in Twitter and fuzzy recommendation. Twitter, in general, has been a subject of study for some years now. For example, several works analyze the “Twittersphere” as a complex network, including topologic and geographic properties (Java et al., 2007); the impact of retweets, user in- fluence, and the longevity of trending topics (Kwak et al., 2010); and how users with similar interests are connected — a phenomenon known as ho- mophily (Huberman et al., 2008). Other works include outcome prediction for the American presidential election (Tumasjan et al., 2010), stock market prediction (Sakaki et al., 2010), rumor propagation (Liu and Chen, 2011), and epidemics detection (Culotta, 2010). Regarding followee recommendations in Twitter, Hannon et al. (2010, 2011) introduce the Twittommender, a real-time system where users can search for followees and receive followee recommendations as well. To gen- erate these recommendations, a user profile of tf-idf weights is extracted and presented as a query to the system. Nine profiling strategies are proposed; five of these use a single information source and four combine two or several sources — either of the same or different types (content vs. structure). The strategies consist of: (1) user tweets, (2) followee tweets, (3) follower tweets, (4) followee id’s, (5) follower id’s, (6) a combination of user, followee, and 9 follower tweets, (7) a combination of follower and followee id’s, (8) a combi- nation of user tweets and follower id’s, and (9) an ensemble of strategies 1-7 where users which frequently have a high position in recommendations are preferred over users whose frequencies or positions are low. These profiling strategies were evaluated by measuring precision (overlap between the lists of recommended users and the actual followee lists of the target users) and ranking effectiveness (if relevant recommendations were placed in high po- sitions of the list); a live trial where target users were asked if they would follow the recommended users was also carried out. Collaborative filtering strategies (4,5,7) gave, in general, better results than content-based strate- gies (1-3,6), although the latter tended to place relevant recommendations in higher positions. Similar results are presented by Armentano et al. (2011a,b), who propose two recommenders: purely based on content and purely based on structure. For the content-based recommender, a set of users is randomly selected from Twitter’s public timeline (a stream where the tweets of public accounts are added) and represented with term-vector profiles; cosine similarity between these users and the target user (whose profile is also a vector) is calculated and users exceeding a given threshold are added into the recommendation list. For the topology-based (structural) recommender, the followees of the target user’s co-followers (i.e. the group of users who share followees with the target user) are obtained and ranked according to a score that attempts to balance popularity and resemblance to user preferences; this score takes the product of a candidate recommendation’s proportion of repetitions (a higher score is awarded to users recommended several times), proportion of followers vs. followees, and proportion of mentions. Results were evalu- ated using discounted cumulative gain, precision, and mean reciprocal rank; even when no statistically significant differences were found between the two recommenders, the content-based approach showed to position relevant rec- ommendations at the top of the list. Other recommenders that make use of structural information are given in the works by Hsu et al. (2006) and Silva et al. (2010). The former describes a hybrid recommender that employs graph proximity and logistic regression to suggest friends in the LiveJournal weblog, and the latter uses complex network theory and genetic algorithms to recommend friends of friends in the Oro-Aro SNS. While several works treat recommendation as an information retrieval problem, Tsourougianni and Ampazis (2013) present a followee hybrid rec- ommender based on classification. The approach, which receives the target user (or target user and followee) tweets as a query, locates the target user’s friends-of-friends (FoF’s) and classifies their tweets as either positive or neg- 10 ative for recommendation; the users with the highest percentage of positives are added to the recommendation list. The features for the classifier (a Markovian classifier) consist of Orthogonal Sparse Bigrams, which are ex- tracted per tweet. As in other cases, the approach was evaluated using precision. While most of the results yielded a high precision, the authors imply that the presence of bots in Twitter could cause outlier results. Followee recommendations in Twitter are also tackled by Gavilanes et al. (2013) with a system that, unlike others, is built on top of human-generated recommendations. To show that these influence users’ behavior towards fol- lowing new people, the authors study a 24-week corpus created with manual recommendations from the #followfriday (or #ff ) trend, which consists of users recommending their followees other users to follow (celebrities, au- thorities in a certain topic, and the like); these recommendations are made on Fridays, hence the name of the hashtag. To rank the human-generated recommendations, several features are extracted and used with the Rota- tion Forest algorithm to train a binary classifier. The features are grouped into three categories: user, relation, and format. While the first category includes user popularity and level of activity in Twitter, the second con- siders level of communication between users and similarity (content-based and geographical), and the third includes recommendation repetition (i.e. how many times a user has been recommended) and context (e.g. day of the week). The results were evaluated using mean average precision, and the relation features were found to be the most useful; on the contrary, the format features were the least useful. Another recent approach concerns the Twilite system, which is intro- duced by Kim and Shim (2013); preceded by a similar model-based ap- proach named Twitobi (Kim and Shim, 2011), the Twilite system sits on Latent Dirichlet Allocation and matrix factorization, which help to recover the process of tweeting and following users. The resulting probabilistic gen- erative model (whose parameters are learned via expectation maximization) is used for followee and tweet recommendation. Still within the context of recommendation in social media are the sem- inal works by Chen et al. (2009, 2010) and Liu and Lee (2010). The first of these works (Chen et al., 2009) proposes and evaluates four algorithms that are applied on Beehive, an enterprise social networking site of IBM. One of these algorithms is purely content-based, as it creates tf-idf vectors for user profiles and generates the recommendation list based on cosine similarity; the second algorithm (which is hybrid) extends the former by increasing similarity if the users are linked. The third algorithm, purely structural, is based on FoF’s and mutual friendships, and the last algorithm (named 11 SONAR) is leveraged by organizational information, such as patents, the or- ganizational chart, and project wikis. It is important to note that, for every recommendation, an explanation is provided (this feature being also used in works by other authors). To assess the four algorithms, the authors con- ducted a survey where users were asked a variety of questions regarding the received recommendations; another form of evaluation consisted of a con- trolled field study where a recommender widget was introduced in Beehive and monitored for a sample of users. While all algorithms showed strengths and weaknesses, in general SONAR was considered as a fair alternative, since it takes advantage of more information. Posterior works by these authors focus on Twitter, specifically on URL recommendation with topic relevance and social voting (Chen et al., 2010). With respect to the work by Liu and Lee (2010), this work — which is applied on the Cyworld SNS — shows that collaborative filtering can be enhanced by combining nearest neighbors with friends for item recommendation. The approach by Li et al. (2011) suggests experts by using a fuzzy lin- guistic method and fuzzy text categorization to analyze task documents. Other items recommended using a fuzzy engine include electronic products (Cao and Li, 2007), music (Park et al., 2006), and academic resources inside universities (Porcel and Herrera-Viedma, 2010); fuzzy logic has been used, as well, for learning and constructing user profiles (Castellano et al., 2010; Xiang-Wei et al., 2009). More recently, a friend-to-friend recommendation system is proposed by Yigit et al. (2015), where a classification of data in four categories is made and a mathematical equation is used to generate a recommendation; this approach performs better than extended FoF, a graph-based system, and a conceptual fuzzy set based algorithm. Also, a fuzzy system is used to extract useful interaction behavior of Twitter users (Fu and Shen, 2014). Collective user behavior is analyzed to generate tweet reply or not reply categories by means of fuzzy association rules that consider activity time, number of friends/followers, and number of tweets as influen- tial factors. Other fuzzy systems have been used for sentiment analysis on social networks, for example, the analysis of opinions generated by Twitter and Facebook with the use of fuzzy propagation modeling (Trung and Jung, 2014). In the approach by Bollen et al. (2011), a self-organized fuzzy neural network is used to relate opinions generated in Twitter by mood tracking tools with changes in Dow Jones industrial average over time. 12 3.1 Discussion Reviewing the aforementioned related approaches, we can see that none of them does followee recommendation in Twitter using a fuzzy-based method (hence the uniqueness of the solution we propose). The state of the art, to the best of our knowledge, instead poses two kinds of approaches: followee recommendation in Twitter using various methods and fuzzy approaches for various tasks and contexts (other than followee prediction and recommen- dation in Twitter). Let us discuss related work of each kind. The followee recommendation task is not new and has been explored pre- viously; however, a number of the proposed methods have been developed over less complex (smaller, homogeneous, less noisy) contexts. Considering that Twitter is by no means a simple context, the few works that actually concentrate on this social network still have room for improvement — either talking about precision, efficiency, or robustness. For example, the works by Gavilanes et al. (2013) and Hannon et al. (2010) use a considerable number of variables (the greater the number of variables, the more time it takes to extract this information), while our approach proposes only three. On the aspect of precision, the work by Armentano et al. (2011b) reports results that could be further improved. On the aspect of robustness, Tsourougianni and Ampazis (2013) state that bots (noise) affect the performance of the ap- proach, and for this reason our approach is based on fuzzy logic. In summary, our approach attempts to improve — in any aspect — the current state of the art, or at least bring in new knowledge about followee recommendation in Twitter. With respect to fuzzy expert systems for various tasks and contexts, as we have seen, these approaches so far have not been devoted to the task of followee prediction and recommendation in Twitter; this is, therefore, an opportunity area for the field of expert systems. Tackling this opportunity area could bring novel perspectives on challenging problems and show that the solution proposed by expert systems keeps fitting into new contexts (no need of “reinventing the wheel”). As a result, it is possible to state that our contribution lies within two areas: fuzzy expert systems and followee recommendation in Twitter and similar contexts. For the fuzzy expert systems area, our contribution consists of a fuzzy system capable of predicting the degree of friendship between a pair of users of a sui generis (massive, dynamic, heterogeneous) social network; for the followee recommendation area, our contribution consists of a robust, efficient approach that combines several information types (text, structure) in a manner that has not been tried before. 13 4 Approach To recommend followees, we designed a fuzzy system that takes as input three similarity values between a target user ui and a candidate user uj: comment (tweet) similarity, followee similarity, and followee tweet similarity. The rules of the system produce the predicted degree of friendship (follower- followee relation) between the users, and this value is used to determine if uj is worth recommending. To calculate the similarities, we first extract user profiles. 4.1 Generation of user profiles As previously mentioned, a profile contains information that allows to char- acterize the user and measure similarity with respect to other users, thus enabling recommendations. We consider that a user profile can be con- structed with (a) the comments of the user, (b) the comments of the user’s followees, and (c) the id’s of the followees. This approach for generating user profiles includes both content and structure: while the former is given by comments that comprise topics and opinions, the latter is given by links that reveal how the user is connected. It also — from the recommender systems perspective — combines both content-based and collaborative fil- tering strategies by employing information from the user and information from contacts, respectively. In that sense, the approach is hybrid. Upon generating a user profile, it is important to take into consideration the amount of data per user; when this amount is not easily manageable, the profile might as well be generated with a random sample of comments and contacts. To the best of our knowledge, there is no standardized number of followees to include or the criteria to select these. 4.1.1 Formal conceptual framework for profile generation Let us formally depict the profile of a user ui with a triplet pi = (ci, ei, fi) of document vectors, where the vector ci represents the comments of ui, the vector ei represents the comments of ui’s followees, and the vector fi represents the id’s of ui’s followees; for later explanation, let us assume that a function fp(ui) produces pi. All vectors consist of tf-idf weights that are computed by conceiving Twitter as a heterogeneous information network and constructing documents from this network. 14 Let G = (V, E) represent a heterogeneous information network of users and comments. G is unweighted, vertex-labeled, and consists of two sub- networks: a follower-followee directed network Gf = (Vf, Ef) and a user- comment undirected bipartite network Gcf = (Vcf, Ecf). An example of this kind of network is illustrated by Figure 4. c6 c3 c4 c2 c7 c5 c1 u3 u1 u2 u5 u4 Figure 4: Example of a heterogeneous information network. For Gf, each vertex v ∈ Vf stands for a user and each arc (a, b) ∈ Ef represents a “who-follows-who” relationship where a follows b; these arcs only join pairs of vertices in Vf, such that Ef ⊂ Vf × Vf. With respect to Gcf, it represents a set of comments Vc generated by the set of users Vf (note that Vc∪Vf = Vcf = V ); an edge {a, c}∈ Ecf where a ∈ Vf and c ∈ Vc indicates that user a made comment c. From our example, we can see that Vf = {u1 . . . u5}, Vc = {c1 . . . c7}, (u1, u2) ∈ Ef, and {u1, c1}∈ Ecf. To retrieve followees and comments from users, let us introduce three kinds of neighborhoods on the vertices of G: the f-neighborhood, the com- ment neighborhood, and the extended comment neighborhood. The f-neighborhood of a vertex i, which we will denote by Γ+ f (i), is the out-neighborhood of i in Gf and represents the set of followees for a given user (see Eq. 1 in Section 2.2). In our example, Γ+ f (u2) = {u3, u5}. The comment neighborhood of a vertex i ∈ Vf, which we will denote by 15 Γc(i), is the neighborhood of i in the bipartite subgraph Gcf and represents the set of comments (tweets) for a given user: Γc(i) = {v : {v, i}∈ Ecf, i ∈ Vf} . In our example, Γc(u2) = {c3, c4}. The extended comment neighborhood of a vertex i ∈ Vf, which we will de- note by Γe(i), corresponds to the comment neighborhoods of the f-neighbors for i and represents the comments of the followees for a given user: Γe(i) = {v : (i, j) ∈ Ef, v ∈ Γc(j)} . Regarding our example, Γe(u2) = Γc(u3)∪Γc(u5) = {c5, c6}. It is also important to introduce a function L(i) that maps a vertex i to a label li, where the latter consists of a user id for v ∈ Vf or a comment (text) for v ∈ Vc. Taking the previous definitions into consideration, three types of doc- uments can be generated per user: the comment document, the extended comment document, and the followee document. These document types all emerge from the same procedure: the concatenation of labels that belong to a given neighborhood. Let us represent this procedure with a function Φ(N) = n j∈N L(j). In this case ci = Φ(Γc(i)), ei = Φ(Γe(i)), and fi = Φ(Γ + f (i)). For our exam- ple, let us assume that L(u2) = “@newton”, L(c3) = “shoulders of giants”, L(c4) = “action and reaction”, L(u3) = “@archimedes”, L(c6) = “eureka”, L(u5) = “@galileo”, and L(c5) = “it moves”. Here, cu2 = “shoulders of giants action and reaction”, fu2 = “@galileo @archimedes”, and eu2 = “eureka it moves”. By extending this concept, three corpora are constructed from G: a comment corpus C, an extended comment corpus E, and a followee corpus F. Each corpus consists of the union of its respective documents. Formally, X = ⋃ i∈Vf xi where X ∈{C, E, F} and xi ∈{ci, ei, fi}. Taking these considerations into account, the calculation of the tf-idf scores for the document vectors ci, ei, and fi is straightforward from Equation 8. 16 Given two profiles pi and pj, we create a similarity vector where each ele- ment corresponds to the cosine similarity (Eq. 9) between document vectors of the same type. Let us denote this vector with s(i,j) = [ c(i,j), e(i,j), f(i,j) ] , where c(i,j) = cosim(ci, cj), e(i,j) = cosim(ei, ej), and f(i,j) = cosim(fi, fj). For further explanation, let us assume that a function fsim(ui, uj) = s(i,j). For supervised learning approaches, the similarity vectors can actually be turned into a matrix S, where each row is an input vector. An additional vector yD of desired outputs is needed and serves for the training and test sets; each output o(i,j) ∈{0, 1} of this vector indicates whether the users are contacts or not: S =    c(1,2) e(1,2) f(1,2) ... ... ... c(n−1,n) e(n−1,n) f(n−1,n)    and yD = [ o(1,2) o(1,3) . . . o(n−1,n) ]T .(16) 4.2 Fuzzy system design As previously stated, our fuzzy system has three inputs and one output. The linguistic variables for the inputs are tweet similarity, followee tweet similarity, and followee similarity between a pair of users, while the output is the predicted degree of friendship between these users; this result could also be interpreted as the probability of a link (follower-followee relationship) creating between the users. A diagram of the fuzzy system is shown in Figure 5. Friendship prediction value Fuzzy system Followee id similarity Tweet similarity Followee tweet similarity Figure 5: Architecture of the fuzzy system for recommendation prediction. Each input variable has nine partitions, i.e. nine linguistic values (fuzzy sets): Extremely low, Very low, Considerably low, Low, Medium, High, Con- siderably high, Very high, and Extremely high. As we can see from Figure 6, the fuzzy sets at both extremes are trapezoidal functions (half-trapezoids) and the rest are triangular functions; these functions are not only common but also computationally cheap. These functions are described in Eq. 17: 17 type1(x, n) =                            trapezoidal(x, 0, 0, 0.1, 0.2) if n = 1 triangular(x, 0.1, 0.2, 0.3) if n = 2 triangular(x, 0.2, 0.3, 0.4) if n = 3 triangular(x, 0.3, 0.4, 0.5) if n = 4 triangular(x, 0.4, 0.5, 0.6) if n = 5 triangular(x, 0.5, 0.6, 0.7) if n = 6 triangular(x, 0.6, 0.7, 0.8) if n = 7 triangular(x, 0.7, 0.8, 0.9) if n = 8 trapezoidal(x, 0.8, 0.9, 1, 1) if n = 9 (17) 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 EL VL CL L M H CH VH EH Figure 6: Linguistic values used (represented as memberships functions). EL=“Extremely low”, VL=“Very low”, CL=“Considerably low”, L=“Low”, M=“Medium”, H=“High”, CH=“Considerably high”, VH=“Very high”, EH=“Extremely high”. To fuzzify the inputs, the singleton method (Eq. 12) is used, as this is the one usually employed and the input does not require further processing. With respect to the fuzzy rule base, there are 93 = 729 rules possible; how- ever, a smaller set was derived by manually analyzing 200 cases in which a pair of users was connected (had a follower-followee relationship). Regard- ing the inference engine, we use Mamdani’s max-min engine (Eq. 14). To defuzzify the output, the centroid (center of mass) method (Eq. 15) is used. 18 4.2.1 Recommendation algorithm To make recommendations for a target user, we select a set of candidate followees (a suitable choice is to use a friends-of-friends or FoF approach) and include in the recommendation list those candidates that exceed an acceptance threshold τ for the prediction value obtained by the fuzzy system; in our experiments, τ was set to 0.5. Additionally, only the top k elements of the list may be chosen (in this paper, we do not apply this filter). This procedure is shown in Algorithm 1. Algorithm 1 Recommendation algorithm Description: Receives a target user ui, a set C of candidate followees, an accep- tance threshold τ, and a list size k. Returns a recommendation list Lk. 1: function recommend(ui, C, τ, k) 2: pi ← fp(ui) 3: L ← () 4: for all uc ∈ C do 5: pc ← fp(uc) 6: s(i,c) ← fsim(pi, pc) 7: y(i,c) ← FS(s(i,c)) ⊲ Get prediction value from fuzzy system (FS). 8: if y(i,c) ≥ τ then 9: append(uc, L) 10: end if 11: end for 12: sort(L) ⊲ Sort L according to prediction values. 13: Lk ← L[1 : k] ⊲ Get the top k elements of L. 14: return Lk 15: end function 5 Experiments and Results To test our approach, we created 30 datasets with information crawled from Twitter and measured recommendation accuracy for each dataset, as it is conventionally done for evaluating recommender systems; we compared our results against a supervised variant of our approach, two state-of-the-art methods (previously described in Section 3) and a baseline. With the in- tent of evaluating our approach in other contexts, we also created a pair of datasets for the Arxiv High-Energy Physics Theory citation network (Cit- HepTh) and obtained accuracy for these datasets as well. Finally, to assess the performance of input variables (followee similarity, tweet similarity, and 19 followee tweet similarity), we calculated the correlation of each input vari- able with respect to the obtained output. 5.1 Experimental Setup Sample networks. Because the Twitter network is very large to han- dle, experiments were performed using smaller sample networks, where each sample network consisted of a heterogeneous network such as the one de- scribed in Section 4; each sample network had one target user, the followees of this target user, and a number of non-contacts (i.e., users who were nei- ther followees nor followers of the target user); the main idea was for the fuzzy system to be able to recover the target user’s followees as recommen- dations. As such, we evaluated the fuzzy system by calculating the accuracy (percentage of matches) between the resulting recommendation list and the actual followee list of the network’s target user. Since sample networks were grouped into datasets, we obtained the average accuracy per dataset. Twitter data. To generate the sample networks, a repository of users and comments was first crawled from Twitter4. We collected a seed set of 100 users, which resulted from searching “Monterrey” (a northeastern city of Mexico) in Twitter and selecting those users that complied with four criteria: having Monterrey as their location, writing in Spanish, having at least 30 followees, and having no more than 3,000 followers. The last two criteria intended to discard companies and celebrities such as singers, actors, and famous politicians; processing this kind of users is left as future work. To have a larger repository, the seed set was expanded by including those followees and followers that complied with the criteria previously mentioned; the expansion was breadth-first, such that the initial level consisted of seed users, the second level consisted of seed user contacts, the third level of contact contacts, and so on. The current repository contains 18,499 users5. From this repository, we randomly selected 6,000 users and approximately 2 million comments. These comments were filtered (stopwords were removed) and spelling correction was performed. Twitter datasets. With the data from Twitter, we constructed 3,000 heterogeneous sample networks, which were equally distributed to create 30 4All information was obtained using the Twitter API, available at https://dev. twitter.com/. 5This repository has also been used for other studies involving Twitter, such as opinion mining in Spanish (Rodŕıguez, 2013). 20 datasets with 100 networks each. While the number of users (size) of each network was variable, on average, there were 24.44 users per network. Cit-HepTh datasets. To assess our approach in a different context, we also created datasets for the Arxiv High-Energy Physics Theory (Cit-HepTh) citation network6, which is analogous to the Twitter heterogeneous network. In this case, we are dealing with reference recommendation, i.e. recommend- ing which paper to cite. In that sense, papers can be seen as equivalent to users, a paper b that is cited by a paper a can be seen as a’s followee, and the abstract of a is equivalent to the set of a’s tweets. For the case of Cit-HepTh, we generated two datasets, where each dataset consisted of 30 heterogeneous sample networks. Each network contained 100 papers on average. Comparative experiments. With the intent of having a wider perspec- tive on the effectiveness of the fuzzy approach, we performed a compari- son against two state-of-the-art methods: Twittommender (Hannon et al., 2010) and the co-followers structural strategy followed by Armentano et al. (2011b); for Twittommender, we used the ensemble strategy. We addi- tionally introduced a baseline, which consisted of a simple friends-of-friends (FoF) approach — i.e., recommending the followees of the target user’s fol- lowees (note that this would yield a high accuracy if the followee and FoF lists had a considerable overlap). Our supervised variant. Another point of comparison was given by the supervised variant of our approach, which was implemented to further as- sess the combination of our three input variables. This supervised variant consisted of a multilayer perceptron neural network (we ran experiments us- ing WEKA7 and its default parameters). We shall refer to this approach as “NN variant” or simply “NN”. 5.2 Results and Discussion A summary of our results is shown in Table 1 (note that this is the average of the average results per dataset), and a box plot for the Twitter datasets is depicted in Figure 7. As we can see from these results, our approaches (the fuzzy system and the neural network) are highly competitive with regard 6This network is available as part of Stanford’s Large Network Dataset Collection. URL: http://snap.stanford.edu/data/cit-HepTh.html. 7Available at: http://www.cs.waikato.ac.nz/ml/weka/. 21 to the rest of the methods; the neural network, in particular, clearly out- performs all competitors (the difference with respect to Twittommender in the Twitter datasets is, in fact, statistically significant with p = 1.16×10−6 using a two-sided paired t-test). This suggests that our three input variables act as reasonable link predictors and are thus feasible for recommendation. Moreover, let us note that almost all methods performed better on the Cit- HepTh datasets, except for the co-followers strategy, which seems to be better tailored for Twitter. This apparently suggests that Twitter’s nature is more “messy” (e.g. user’s comments) and probably more chaotic; it could be therefore convenient to employ the sui generis features of Twitter to try to improve accuracy. However, it would be desirable to have more results (more datasets) on the Cit-HepTh network to have a more confident point of view on this perspective. This is left as future work. Table 1: Average recommendation accuracy (30 Twitter datasets, two Cit-HepTh datasets). FS=Fuzzy system, NN=supervised variant, TM=Twittommender, CF=Co-followers, FoF=friends-of-friends (baseline). FS NN TM CF FoF Twitter 61.57% 71.75% 65.13% 29.99% 3.15% Cit-HepTh 83.25% 86.77% 86.86% 10.84% 8.74% Global 72.41% 79.26% 75.99% 20.41% 5.94% As we can also see from the results, the fuzzy system is a strong com- petitor both inside and outside Twitter. Although it ranks third, there is an evident advantage over the baseline and the co-follower strategy (needless to say, there is a statistically significant difference with these approaches, since p < 2.2 × 10−16); in that sense, our approach remains competitive. With respect to Twittommender, we can see that the fuzzy system’s accuracy lies behind for approximately 3%, but let us also note that the fuzzy system is only using three variables whereas Twittommender is using five (user tweets, followee id’s, followee tweets, follower tweets, and follower id’s), i.e. it needs to extract more information. Another disadvantage of Twittommender is that it is more of a “hit or miss” approach — at least with our Twitter results — and this can be appreciated in the box plot (Figure 7). On the contrary, the fuzzy system tended to give more stable results. With respect to the NN variant, the fuzzy system also stays behind. 22 In general, neural networks are more precise than fuzzy systems, but they are also black boxes and no explanation can be extracted from them; on the other hand, fuzzy systems explain the relationship between inputs and outputs by means of linguistic variables and rules. For example, consider the rule: IF v1 is Low and v2 is Medium High and v3 is Extremely High THEN y is Very High, where v1 is user tweet similarity, v2 is followee tweet similarity, v3 is followee id similarity, and y is the possibility of being contacts. In a linguistic manner, this rule explains that if there is a low user tweet similarity and there is a medium high followee tweet similarity and there is an extremely high followee id similarity, then there is an extremely high possibility to be contacts. We can choose any fuzzy rule and make the same interpretation. Fuzzy systems have the disadvantage of not being very accurate against other paradigms such as neural networks; however, because NN’s are black boxes, it is very difficult to establish an interpretation of the relationship between inputs and outputs. Moreover, a fuzzy system has a high interpretability because it generates rules that explain in an explicit and linguistic way the relationship between variables. 5.2.1 Individual variables As we stated previously, with a fuzzy approach it is possible to know how each input variable influences the output. To measure the individual influ- ence of input variables, we calculated the correlation coefficient between each input variable and the output. Our results for both Twitter and Cit-HepTh are displayed in Table 2. As we can see from this table, for both cases, the variable with the highest influence on the fuzzy system corresponds to fol- lowee tweets. This result allows us to gain insight on the factors that affect prediction, and thus make adjustments for future work. 5.2.2 Summary In summary, our results show that the fuzzy approach is suitable for link prediction in recommender systems, both in Twitter and in other contexts. The approach is highly competitive with other methods. In some cases, it clearly surpasses the accuracy obtained by its competitors; in other cases, it stays behind by a small percentage but excels in other aspects, such as 23 Table 2: Correlation between input variables and prediction output. Twitter Cit-HepTh User tweets 0.12 0.43 Followee id’s 0.42 0.32 Followee tweets 0.85 0.65 amount of information used and capability of explanation. This last goal is important for recommender systems in general (Bonhard and Sasse, 2006). 6 Conclusions and Future Work A new fuzzy system for followee recommendation in Twitter has been pro- posed in this paper. This system handles recommendation mainly as a link prediction problem and uses three linguistic variables that express similar- ity between users: tweet similarity, followee tweet similarity, and followee similarity. Every similarity contributes in a different way for the prediction. To calculate these similarities, a profile is built for each user. The profile consists of vectors that contain information from comments and followee relationships, and this information is extracted by considering the Twitter social networking site as a heterogeneous information network. From an objective point of view, the approach has both strengths and weaknesses. On one hand, as with any expert system, creating the rules requires knowledge and is an additional overhead of the approach; also, as we could see from the results, accuracy is still below other methods and could still be further improved. On the other hand, the system only requires three types of information (on the contrary of other methods which require more), is self-explanatory, and is capable of producing a degree instead of only a yes/no binary answer. The self-explanatory capability, in particular, allows the system to explain how variables interact to produce an output, as opposed to “black-box” methods, such as neural networks or support vector machines. With respect to the friendship degree, besides helping to construct the recommendation list for a particular user, it can also be used to carry out other tasks, such as discovering latent user communities or clusters. With respect to the impact of this work, our fuzzy approach brings in 24 a new perspective on the use and construction of expert systems for link prediction and contact recommendation in microblog-oriented social net- works. Our method, in particular, brings new knowledge on how to address recommendation in a sui generis, heterogeneous, massive, dynamic environ- ment, which is subject to inaccuracies (think, for example, about language inconsistencies) and incomplete information. One of our main findings, in particular, reveals that the combination of user tweet similarity, followee id similarity, and followee tweet similarity yield accurate followee recom- mendation results (these results are even better when the combination is used in a supervised approach); in that sense, the use of only three infor- mation types is promising — specially in the neural network case. Even when the fuzzy approach can be further improved, we have taken the first step towards a robust, efficient, accurate solution to the problem. Another theoretical contribution that it is important to highlight and that might be of interest is the conceptual model of Twitter as a heterogeneous network, as this model not only facilitates the understanding of Twitter’s underlying dynamics, but also serves as a basis for other data mining tasks, such as cyberbullying detection and interest mining. As future research directions, we propose several courses of action. One of these consists of modifying the existing rule base to incorporate more input variables (e.g. follower tweets, follower id’s, retweets, multimedia con- tent, etc.) with the intent of improving accuracy and exploiting Twitter’s unique features; of course, this implies an additional overhead and the trade- off between quality and efficiency must be assessed. Now, beyond the sole introduction of more variables, lies the issue of finding an optimal design for the fuzzy system (fuzzification and defuzzification methods, membership functions, partitions, etc.); in that sense, adaptive methods — such as ge- netic algorithms, PSO, and similar — could be employed to find this optimal design. Another important future research direction concerns testing and adapt- ing the fuzzy expert system for other contexts, both inside and outside Twit- ter. In that sense, it seems valuable to discover how general the system could be, and if it is possible to have a tailored version enhanced for Twitter (or a specific language, such as Chinese or Arab, whose alphabet is different) and a general version for heterogeneous networks analogous to Twitter (such as citation networks). Consequently, there is still plenty of room to create new “flavors” of the original fuzzy expert system. Related to this is the hybridization of the fuzzy system with a supervised technique to create a more powerful approach. A final future research direction is related to community discovery and 25 graph clustering. Because the fuzzy system produces a friendship degree for a pair of users, it is possible to view this degree as a network weighted edge and use this information to construct a friendship network (or at least parts of) and develop a local probabilistic graph clustering algorithm to find groups of similar users in Twitter (or other complex networks). On the contrary of the follow graph, which would be normally used for the clustering task in Twitter, this friendship network constructed from the fuzzy system could give new insights to community discovery in Twitter and social networks in general. 26 Fuzzy NN TM CF FoF 0 2 0 4 0 6 0 Approach A cc u ra cy ( % ) Figure 7: Comparison using the Twitter datasets. Each box shows the min- imum, first quartile, median, third quartile, and maximum for the accuracy obtained from the 30 datasets (one box per approach). NN= neural network (multilayer perceptron), TM= Twittommender, CF=co-followers strategy, FoF=friends-of-friends strategy. 27 References ARMENTANO, M. G., GODOY, D., and AMANDI, A. (2011a) Towards a followee recommender system for information seeking users in Twitter, in Proceedings of the Workshop on Semantic Adaptive Social Web (SASWeb 2011), CEUR Workshop Proceedings, volume 730, 27–38. ARMENTANO, M. G., GODOY, D. L., and AMANDI, A. A. (2011b) A topology-based approach for followees recommendation in Twitter, in Proceedings of the 9th Workshop on Intelligent Techniques for Web Per- sonalization and Recommender Systems, Barcelona, Spain. BAEZA-YATES, R. and RIBEIRO-NETO, B. (1999) Modern information retrieval. ACM Press, New York, NY, USA. BENITO-RUIZ, E. (2009) Handbook of research on Web, chapter Infoxica- tion 2.0, 60–79. IGI Global. doi: 10.4018/978-1-60566-190-2.ch004. BOLLEN, J., MAO, H., and ZENG, X. (2011) Twitter mood predicts the stock market, Journal of Computational Science, 2(1), 1–8. BONHARD, P. and SASSE, M. (2006) Knowing me, knowing you: Us- ing profiles and social networking to improve recommender systems, BT Technology Journal, 24(3), 84–98. CAO, Y. and LI, Y. (2007) An intelligent fuzzy-based recommendation sys- tem for consumer electronic products, Expert Systems with Applications, 33(1), 230–240. CASTELLANO, G., CASTIELLO, C., DELL’AGNELLO, D., FANELLI, A., MENCAR, C., and TORSELLO, M. (2010) Learning fuzzy user pro- files for resource recommendation, International Journal Of Uncertainty, Fuzziness, and Knowledge-based Systems, 18(4), 389–410. CHEN, J., GEYER, W., DUGAN, C., MULLER, M., and GUY, I. (2009) Make new friends, but keep the old: Recommending people on social networking sites, in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems: ACM, 201–210. CHEN, J., NAIRN, R., NELSON, L., BERNSTEIN, M., and CHI, E. (2010) Short and tweet: experiments on recommending content from information streams, in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems: ACM, 1185–1194. 28 CULOTTA, A. (2010) Towards detecting influenza epidemics by analyzing Twitter messages, in Proceedings of the 1st Workshop on Social Media Analytics: ACM, 115–122. FU, X. and SHEN, Y. (2014) Study of collective user behaviour in Twitter: a fuzzy approach, Neural Computing and Applications, 25(7-8), 1603–1614. GANTZ, J. and REINSEL, D. (2010) The digital universe decade: Are you ready? IDC Review. GAVILANES, R. G., O’HARE, N., AIELLO, L. M., and JAIMES, A. (2013) Follow my friends this Friday! An analysis of human-generated friendship recommendations, in Social Informatics: Springer, 46–59. GETOOR, L. and DIEHL, C. (2005) Link mining: a survey, ACM SIGKDD Explorations Newsletter, 7(2), 3–12. HANNON, J., BENNETT, M., and SMYTH, B. (2010) Recommend- ing Twitter users to follow using content and collaborative filtering ap- proaches, in Proceedings of the 4th ACM conference on Recommender Systems: ACM, 199–206. HANNON, J., MCCARTHY, K., and SMYTH, B. (2011) Finding useful users on Twitter: Twittomender the followee recommender, in Advances in Information Retrieval: Springer, 784–787. HSU, W. H., KING, A. L., PARADESI, M. S., PYDIMARRI, T., and WENINGER, T. (2006) Collaborative and structural recommendation of friends using weblog-based social network analysis, in Proceedings of the 2006 AAAI Spring Symposium: Computational Approaches to Ana- lyzing Weblogs: AAAI, 55–60. HUBERMAN, B., ROMERO, D. M., and WU, F. (2008) Social networks that matter: Twitter under the microscope, First Monday, 14(1). doi: http://dx.doi.org/10.5210/fm.v14i1.2317. JAVA, A., SONG, X., FININ, T., and TSENG, B. (2007) Why we twitter: understanding microblogging usage and communities, in Proceedings of the 9th WebKDD and 1st SNA-KDD 2007 workshop on Web mining and social network analysis: ACM, 56–65. JI, M., HAN, J., and DANILEVSKY, M. (2011) Ranking-based classifica- tion of heterogeneous information networks, in Proceedings of the 17th 29 ACM SIGKDD International Conference on Knowledge Discovery and Data mining: ACM, 1298–1306. KIM, Y. and SHIM, K. (2011) Twitobi: A recommendation system for Twitter using probabilistic modeling, in Proceedings of the IEEE 11th International Conference on Data Mining (ICDM): IEEE, 340–349. KIM, Y. and SHIM, K. (2013) TWILITE: A recommendation system for Twitter using a probabilistic model based on latent Dirichlet allocation, Information Systems, 42, 59–77. doi: http://dx.doi.org/10.1016/j.is.2013. 11.003. KWAK, H., LEE, C., PARK, H., and MOON, S. (2010) What is Twitter, a social network or a news media?, in Proceedings of the 19th International Conference on World Wide Web: ACM, 591–600. KYWE, S. M., LIM, E.-P., and ZHU, F. (2012) A survey of recommender systems in Twitter, in Social Informatics: Springer, 420–433. LI, M., LIU, L., and LI, C.-B. (2011) An approach to expert recommendation based on fuzzy linguistic method and fuzzy text classification in knowledge management systems, Expert Systems with Applications, 38(7), 8586– 8596. LIU, B. (2007) Web data mining: Exploring hyperlinks, content, and usage data. Springer, Chicago, IL, USA. LIU, D. and CHEN, X. (2011) Rumor propagation in online social net- works like Twitter–a simulation study, in Proceedings of the 3rd Inter- national Conference on Multimedia Information Networking and Security (MINES): IEEE, 278–282. LIU, F. and LEE, H. J. (2010) Use of social network information to enhance collaborative filtering performance, Expert systems with applications, 37 (7), 4772–4778. NEWMAN, M. (2010) Networks: An introduction. Oxford University Press, New York, USA. PARK, D. H., KIM, H. K., CHOI, I. Y., and KIM, J. K. (2012) A liter- ature review and classification of recommender systems research, Expert Systems with Applications, 39(11), 10059–10072. 30 PARK, H.-S., YOO, J.-O., and CHO, S.-B. (2006) A context-aware music recommendation system using fuzzy bayesian networks with utility theory, in Fuzzy systems and knowledge discovery: Springer, 970–979. PORCEL, C. and HERRERA-VIEDMA, E. (2010) Dealing with incomplete information in a fuzzy linguistic recommender system to disseminate in- formation in university digital libraries, Knowledge-Based Systems, 23(1), 32–39. RICCI, F., ROKACH, L., and SHAPIRA, B. (2011) Introduction to recom- mender systems handbook. In Ricci, F., Rokach, L., Shapira, B., and Kan- tor, P. B., editors, Recommender Systems Handbook, pages 1–35. Springer US. ISBN 978-0-387-85819-7. doi: 10.1007/978-0-387-85820-3 1. URL http://dx.doi.org/10.1007/978-0-387-85820-3_1. RODRÍGUEZ, F. M. (2013) Cuantificación del interés de un usuario en un tema mediante mineŕıa de texto y análisis de sentimiento. Master’s thesis, Universidad Autónoma de Nuevo León. SAKAKI, T., OKAZAKI, M., and MATSUO, Y. (2010) Earthquake shakes Twitter users: real-time event detection by social sensors, in Proceedings of the 19th International Conference on World wide web: ACM, 851–860. SILVA, N. B., TSANG, R., CAVALCANTI, G. D., and TSANG, J. (2010) A graph-based friend recommendation system using genetic algorithm, in Proceedings of the 2010 IEEE Congress on Evolutionary Computation (CEC): IEEE, 1–7. SUN, Y., YU, Y., and HAN, J. (2009) Ranking-based clustering of hetero- geneous information networks with star network schema, in Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Dis- covery and Data mining: ACM, 797–806. SUN, Y., HAN, J., AGGARWAL, C. C., and CHAWLA, N. V. (2012) When will it happen?– Relationship prediction in heterogeneous information net- works, in Proceedings of the 5th ACM International Conference on Web Search and Data mining: ACM, 663–672. TRUNG, D. N. and JUNG, J. J. (2014) Sentiment analysis based on fuzzy propagation in online social networks: a case study on tweetscope, Com- puter Science and Information Systems, 11(1), 215–228. 31 TSOUROUGIANNI, E. and AMPAZIS, N. (2013) Recommending who to follow on Twitter based on tweet contents and social connections, Social Networking, 2, 165–173. doi: http://dx.doi.org/10.4236/sn.2013.24016. TUMASJAN, A., SPRENGER, T., SANDNER, P., and WELPE, I. (2010) Predicting elections with Twitter: What 140 characters reveal about po- litical sentiment, in Proceedings of the 4th International AAAI Conference on Weblogs and Social Media, 178–185. WANG, J., VARSHNEY, K. R., and MOJSILOVIC, A. (2012) Legislative prediction via random walks over a heterogeneous graph, in Proceedings of the 2012 SIAM International Conference on Data Mining: SIAM, 1095– 1106. XIANG-WEI, M., YAN, C., LI-LI, Q., and TAO-YING, L. (2009) A new user profile model based on intuitionistic fuzzy set for personalized infor- mation analysis and sharing, in Proceedings of the International Confer- ence on Management Science and Engineering (ICMSE): IEEE, 64–69. YIGIT, M., BILGIN, B. E., and KARAHOCA, A. (2015) Extended topology based recommendation system for unidirectional social networks, Expert Systems with Applications, 42(7), 3653–3661. ZADEH, L. (1965) Fuzzy sets. Information and control, 8(3), 338–353. 32 
work_hc6qtyai7nco3l6j6os32vtjnm	----	Paper Title (use style: paper title) Construction of Powerful Online Search Expert System Based on Semantic Web Yasser A. Nada Chairman of Department of computer science, Faculty of Computers and Information Technology, Taif University - K S A, y_nada@yahoo.com Abstract— In this paper we intends to build an expert system based on semantic web for online search using XML, to help users to find the desired software, and read about its features and specifications. The expert system saves user's time and effort of web searching or buying software from available libraries. Building online search expert system is ideal for capturing support knowledge to produce interactive on-line systems that provide searching details, situation-specific advice exactly like setting a session with an expert. Any person can access this interactive system from his web browser and get some questions answer in addition to precise advice which was provided by an expert. The system can provide some troubleshooting diagnose, find the right products; … Etc. The proposed system further combines aspects of three research topics (Semantic Web, Expert System and XML). Semantic web Ontology will be considered as a set of directed graphs where each node represents an item and the edges denote a term which is related to another term. Organizations can now optimize their most valuable expert knowledge through powerful interactive Web-enabled knowledge automation expert system. Online sessions emulate a conversation with a human expert asking focused questions and producing customized recommendations and advice. Hence, the main powerful point of the proposed expert system is that the skills of any domain expert will be available to everyone. Keywords— Expert System, Semantic Web, Online Search, XML. I. INTRODUCTION The Semantic Web is an extension of the World Wide Web with new technologies and standards that enable interpretation and processing of data and useful information for extraction by a computer. The World Wide Web Consortium (W3C) recommends XML, XML Schema, RDF, RDF Schema and Web Ontology Language (OWL) as standards and tools for the implementation of the Semantic Web. Ontologies work as the main component in knowledge representation for the Semantic Web. It is a data model that represents a set of concepts and the relationships between those concepts within a domain. Building an ontology starting from scratch is not an easy task since it makes heavy demands on time in addition to expert knowledge related to the domain. Expert System (ES), also called a Knowledge Based System (KBS), is computer application programs that take the knowledge of one or more human experts in a field and computerize it so that it is readily available for use. It can also be integrated with textual database which can be used for explanation purposes of basic terms and operations to confirm and to reach conclusion in some situations [1]. A challenge for the Semantic Web is enabling information interoperability between related but heterogeneous ontology [2]. One of the most powerful attributes of expert systems is the ability to explain reasoning. Since the system remembers its logical chain of reasoning, a user may ask for an explanation of a recommendation and the system will display the factors it considered in providing a particular recommendation. This attribute enhances user confidence in the recommendation and acceptance of the expert system [3]. Each time we require certain software or product, we searched the internet twice and more, and maybe 10 times to find what we are searching for, that waste our efforts and time. Many times, we did not even found the software, and then we have to go and buy it from the CD's Shop. Also software fall in hundreds of categories, and sometimes we are very confused what to use and what is the exact features of each one of them. We search through the website of that software and find that all the data is crowded and we just can't differentiate between a software and another and we got confused from that. That paper intends to solve all of these problems, and serve people really by helping them in searching and downloading software. The system will be a web based expert system for searching and downloading software from all types and usability's. That expert system will offer classified arrangements for software and powerful searching method through the available software, also information (features, backwards, download instructions, installation instructions… etc), the expert system will be implemented as mentioned above as a web based system, we need to go through several steps to accomplish that goal. II. RELATED WORKS In [4] the authors presented the characteristics of disaster management is different in different disaster (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 12, 2013 181 | P a g e www.ijacsa.thesai.org mailto:y_nada@yahoo.com relief information, knowledge, standardization of operating procedures and the feasibility of the rescue program. Knowledge-sharing and case retrieval is to develop case-based intelligent decision support system facing the most important issue. Disaster rescue command for decision-making, the use of Web Ontology Language that state the information and knowledge of characteristics of the earthquake disaster rescue, a case-based reasoning and logic to describe the rescue planning business processes. In [5] presented the main ramification of the idea of ―data shouldn’t be dumb‖ on rules languages which is the language has to be simple and able to cover simple linkage situations. More complex applications of rules should be reserved for the intelligent applications that will populate the web. This isn’t to say that research and standardization on more powerful and flexible rule languages aren’t valuable or shouldn’t be done, it’s not just necessary or even desirable for the rule-based infrastructure of the semantic web. Keeping these things separate – the infrastructure from the residents of the web – lets us have a much simpler rule language as part of the web infrastructure. This paper [6] has proposed two extensions for the current generation of digital libraries. First, authors proposed to use ontology to represent scholarly information in digital libraries, thus making the libraries be able to share and exchange knowledge in the Semantic Web environment. Second, fuzzy theory is employed to process uncertain scholarly information as the forms of fuzzy ontology and fuzzy queries. A general architecture of digital libraries in the Semantic Web environment has been presented and an experimental system has also been developed to verify our ideas and techniques. In [7] proposed an ontology-based information retrieval model to improve effectiveness of information retrieval. The ontology embedded in the proposal model is a fuzzy taxonomy generated automatically from the documents. In [8] presented an approach to level one sensor fusion in the context of the Semantic Web has been presented using a Semantic Web Expert System Shell (SWEXSYS). Our approach allows for the delegation of sensor fusion tasks to agent-based systems like SWEXSYS. In our approach, we discussed the merit of having Semantic Web enabled data. By making data Semantic Web compliant, an agent may be capable of disambiguating the information; also, additional information that may be needed in the fusion process can be retrieved without human intervention from the Semantic Web by the fusion agent. III. KNOWLEDGE IN THE EXPERT SYSTEM Knowledge acquisition, knowledge representation, knowledge storage and knowledge application constitute the main content of expert system’s work progress [9], whose core is knowledge base and inference engine. The level of an expert system is basically determined by its knowledge base. Expert’s experience and knowledge are stored in the knowledge base, the more complete and true of the knowledge, the higher level of expert system is [10]. Thus whether it possesses plenty of knowledge is the key of expert system. There are representation techniques such as frames, rules, tagging, and semantic networks which have originated from theories of human information processing. Since knowledge is used to achieve intelligent behavior, the fundamental goal of knowledge representation is to represent knowledge in a manner as to facilitate inference (i.e. drawing conclusions) from knowledge [11, 12]. We got the following software categories that will be written inside the semantic network knowledge representing as follows: Internet browser Figure (1), Painting program, Downloading program, Recovery Program Figure ( 2), Compressors, File sharing Figure (3), Security, Video and audio players, E-books reader, Text Editors, Slides viewers, Images viewers, Cleaning programs. Figure 1: Knowledge Representation for Internet Browser (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 12, 2013 182 | P a g e www.ijacsa.thesai.org Ontologies in Semantic Web Figure 2: Knowledge Representation for Recovery Program Figure 3: Knowledge Representation for File Sharing Programs (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 12, 2013 183 | P a g e www.ijacsa.thesai.org IV. ONTOLOGIES IN SEMANTIC WEB Ontology is a data model, which can be used to describe a set of concepts and the relationships between those concepts within a domain. Ontology works as the main component in knowledge representation for the Semantic Web. Research groups in both America and Europe developed Ontology modeling languages as The DARPA Agent Markup Language (DAML) and Ontology Inference Layer (OIL.) The W3C Web Ontology Working Group has considered DAML+OIL as the starting point for the introduction of standardized and accepted ontology language for the Semantic Web as Web Ontology Language (OWL) [13]. OWL has three sublanguages: OWL Full, OWL DL and OWL Lite [14]. Existing Semantic Web ontology can be grouped into the following four major categories: metaontology, comprehensive, upper ontology, systematic domain specific ontology, and simple specialized ontology[15]. V. SYSTEM ARCHITECTURE The system further combines aspects of three research topics (XML, Semantic Web, and Expert System) to facilitate the finding of semantic web services for a client Figure (4). A. XML Extensible Markup Language (XML) is a set of rules for encoding documents in machine-readable form. It is defined in the XML 1.0 Specifications produced by the W3C, and several other related specifications, all gratis open standards. The design goals of XML emphasize simplicity, generality, and usability over the Internet. It is a textual data format with strong support via Unicode for the languages of the world. Although the design of XML focuses on documents, it is widely used for the representation of arbitrary data structures, for example in web services. Many application programming interfaces (APIs) have been developed that software developers use to process XML data, and several schema systems exist to aid in the definition of XML-based languages. Currently, web services interact by passing XML data, with data types specified using XML Schema. Simple Object Access Protocol [SOAP] can be used as the communication protocol [16], and the I/O signatures for web services are given by Web Services Description Language [WSDL] [17]. UDDI stands for Universal Description, Discovery and Integration [18] and provides the means to publish and discover web services through a UDDI registry. Today, XML is invading the world of computers and occupying most of its fields. It is widely spreading over the internet, networks, information systems, software and operating systems, DBMS, search tools, web development and services, communication protocols and other fields. As a result, XML data are floating within and between different applications and systems all over the internet and intranets. A core data representation format for semantic web is Resource Description Framework (RDF). RDF is a framework for representing information about resources in a graph form. It was primarily intended for representing metadata about WWW resources, such as the title, author, and modification date of a Web page, but it can be used for storing any other data. It is based on triples subject-predicate-object that form graph of data. All data in the semantic web use RDF as the primary representation language. The normative syntax for serializing RDF is XML in the RDF/XML form. Formal semantics of RDF is defined as well. B. Semantic Web Architecture The architecture of semantic web is illustrated in the Figure (5) below. The first layer, URI and Unicode, follows the important features of the existing WWW. Unicode is a standard of encoding international character sets and it allows that all human languages can be used (written and read) on the web using one standardized form. Uniform Resource Identifier (URI) is a string of a standardized form that allows to uniquely identifying resources (e.g., documents). A subset of URI is Uniform Resource Locator (URL), which contains access mechanism and a (network) location of a document. Another subset of URI is URN that allows identifying a resource without implying its location and means of dereferencing it - an example is urn. The usage of URI is important for a distributed internet system as it provides understandable identification of all resources. An international variant to URI is Internationalized Resource Identifier (IRI) that allows usage of Unicode characters in identifier and for which a mapping to Figure 4: System Architecture Semantic Web Ontology Expert System (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 12, 2013 184 | P a g e www.ijacsa.thesai.org http://en.wikipedia.org/wiki/Machine-readable http://en.wikipedia.org/wiki/W3C http://en.wikipedia.org/wiki/Gratis http://en.wikipedia.org/wiki/Open_standard http://en.wikipedia.org/wiki/Internet http://en.wikipedia.org/wiki/Unicode http://en.wikipedia.org/wiki/Data_structures http://en.wikipedia.org/wiki/Web_service http://en.wikipedia.org/wiki/Application_programming_interfaces http://en.wikipedia.org/wiki/XML_schema http://www.obitko.com/tutorials/ontologies-semantic-web/resource-description-framework.html http://www.obitko.com/tutorials/ontologies-semantic-web/resource-description-framework.html http://www.w3.org/TR/rdf-syntax-grammar/ http://www.w3.org/TR/rdf-mt/ http://www.obitko.com/tutorials/ontologies-semantic-web/semantic-web-architecture.html#semantic-web-layers http://unicode.org/ http://tools.ietf.org/html/rfc3986 http://tools.ietf.org/html/rfc3986 URI is defined. In the rest of this text, whenever URI is used, IRI can be used as well as a more general concept. RDFS have semantics defined and this semantics can be used for reasoning within ontology and knowledge bases described using these languages. To provide rules beyond the constructs available from these languages, rule languages are being standardized for the semantic web as well. Two standards are emerging - RIF and SWRL. For querying RDF data as well as RDFS ontology with knowledge bases, a Simple Protocol and RDF Query Language (SPARQL) are available. SPARQL is SQL-like language, but uses RDF triples and resources for both matching part of the query and for returning results of the query. Since RDFS are built on RDF, SPARQL can be used for querying ontology and knowledge bases directly as well. Note that SPARQL is not only query language; it is also a protocol for accessing RDF data. It is expected that all the semantics and rules will be executed at the layers below Proof and the result will be used to prove deductions. Formal proof together with trusted inputs for the proof will mean that the results can be trusted, which is shown in the top layer of the figure above. For reliable inputs, cryptography means are to be used, such as digital signatures for verification of the origin of the sources. On top of these layers, application with user interface can be built. C. Expert System An expert system is a computer program designed to simulate the problem-solving behavior of a human who is an expert in a narrow domain or discipline. An expert system is normally composed of a knowledge base (information, heuristics, etc.), inference engine (analyzes the knowledge base), and the end user interface (accepting inputs, generating outputs). The concepts for expert system development come from the subject domain of artificial intelligence (AI), and require a departure from conventional computing practices and programming techniques. One of the most powerful attributes of expert systems is the ability to explain reasoning. Since the system remembers its logical chain of reasoning, a user may ask for an explanation of a recommendation and the system will display the factors it considered in providing a particular recommendation. This attribute enhances user confidence in the recommendation and acceptance of the expert system. All expert systems are composed of several basic components: a user interface, a database, a knowledge base, and an inference mechanism. Moreover, expert system development usually proceeds through several phases including problem selection, knowledge acquisition, knowledge representation, programming, testing and evaluation. 1. Inference Engine The entire control and operation of the system are done by the inference engine; that is developed using C#; which handles the knowledge in format of XML to get the result from the XML file (Knowledge base) that stores the knowledge rules. Figure (6) shows the block diagram of inference engine. The main roles of the inference engine are summarized as: It applies the expert domain knowledge to what is known about the present situation to determine new information about the domain. The inference engine is the mechanism that connects the user inputs in the form of answers to the questions to the rules of knowledge base and further continues the session to come to conclusions. This process leads to the solution of the problem. The inference engine also identifies the rules of the knowledge base used to get decision from the system and also forms the decision tree. Figure 5: Semantic Web Architecture in Layers Figure 5: Semantic Web Architecture in Layers Figure 6: Inference Engine Block Diagram (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 12, 2013 185 | P a g e www.ijacsa.thesai.org http://www.w3.org/TR/rif-bld/ http://www.w3.org/Submission/SWRL/ http://www.obitko.com/tutorials/ontologies-semantic-web/rdf-query-language-sparql.html http://www.obitko.com/tutorials/ontologies-semantic-web/rdf-query-language-sparql.html http://www.obitko.com/tutorials/ontologies-semantic-web/semantic-web-architecture.html#semantic-web-layers In this paper the model we took the domain of software is an XML Semantic Web website, we implement the data knowledge based in the form of XML – RDF file to contain all the collected software which we will include in the website, the format of the RDF file as follows: The knowledge is represented in XML format, because it gives flexibility and being industry standard. Figure (7) shows the part of represented knowledge in XML. We define 4 attributes: - Category – Name – Link - Price VI. RESULTS AND TEST THE SYSTEM The system was evaluated with different users, including developers, and staff. The system has validated by experts in the domain of online search. Tests of the system were carried out by the developers to make sure the system would work correctly as well as the system is web based system. Figures (8, 9, 10, and 11) shows the snapshots of the developed system. Figure 7: Represented Knowledge in XML. Figure 8: Home Page Figure 9: Search Tool Page Figure 11: Search Results Page Figure 10: All Available Software Page (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 12, 2013 186 | P a g e www.ijacsa.thesai.org VII. CONCLUSIONS In this paper, an expert system based on semantic web for online search has been proposed. It helps users to find the desired software and read about its features and specifications. The expert system saves users’ time and effort consumed in web searching. Users can access and react with the proposed system easily using their web browsers. In addition, the proposed system produces interactive on-line objects that provide searching details and advices exactly like to setup a session with human expert. Furthermore, the proposed system combines three main research topics; Semantic Web, Expert System, and XML. Also, organizations can optimize their most valuable expert knowledge through the proposed powerful interactive Web-enabled knowledge automated expert system. Finally, online sessions which have been constructed by the proposed system emulate a human expert conversation, which receives questions to produce customized recommendations. REFERENCES [1] Khumukcham R., Shikhar Kr. S, "JESS Based Expert System Architecture for Diagnosis of Rice Plant Diseases: Design and Prototype Development", 2013 4th International Conference on Intelligent Systems, Modeling and Simulation, P 674-676, 2013. [2] Valerie Cross, Xueheng Hu., "Fuzzy Set and Semantic Similarity in Ontology Alignment", WCCI 2012 IEEE World Congress on Computational Intelligence, Brisbane, Australia June, 10-15, 2012. [3] Tanimoto, S. L. , " The Elements of Artificial Intelligence " , Computer Science Press. , 1990. [4] Wu Yun, Li Chan, "Semantic Web-based Seismic Disaster Management Expert System", Project 70601011 supported by National Natural Science Foundation of China, 2009. [5] Dean Allemang, "Rule-based Intelligence in the Semantic Web ", -or- "I'll settle for a web that's just not so dumb!", Proceedings of the Second International Conference on Rules and Rule Markup Languages for the Semantic Web (RuleML'06), 2006. [6] Q. Tho, H. Siu, and C. Tru , "Ontology-Based Fuzzy Retrieval for Digital Library", ICADL 2007, LNCS 4822, pp. 95–98. Springer-Verlag Berlin Heidelberg, 2007. [7] C. Been-chian, H. Chih-hung, J. Ming-yi, "Intelligent Information Retrieval Applying Automatic Constructed Fuzzy Ontology", Proceedings of the Sixth International Conference on Machine Learning and Cybernetics, Hong Kong, 19-22 August 2007. 1- 4244-0973-X/07, IEEE, 2007. [8] Omoju Thomas, David J. Russomanno, "Applying the Semantic Web Expert System Shell to Sensor Fusion using Dempster-Shafer Theory", 0-7803- 8808-9/05, IEEE,2005. [9] Feng Ding., "Neural Network Expert System", Beijing: Science Press, pp.1-10, 2006. [10] Wang Hua, Li Peng-bo. ―Fault Diagnosis Expert System of Inertial Nacigation System Based on CLIPS and .NET Platform [J]‖. Journal of Chinese Inertial Technology, Vol.14 No.6, pp.78-80, 2006. [11] George F. Luger and William, A. Stubblefield; " Artificial Intelligence and the Design of Expert Systems", The Benjamin/ Cummings publishing Co., Inc., 1989. [12] Keith D., "The Essence of Expert Systems", Prentice Hall, 2000. [13] G. Antoniou, F. V. Harmelen, "Handbook on Ontologies in Information Systems", Springer- Verlag , pp. 76-92, 2003. [14] L. Deborah McGuinness, F. V. Harmelen, ―OWL Web Ontology Language Overview‖, W3C Recommendation , Tech. Rep., 10, Feb, 2004. [15] Li Ding, et al., ―Using Ontologies in the Semantic Web: A Survey”, Department of Computer Science and Electrical Engineering , University of Maryland Baltimore County, Baltimore, Tech. Rep. TR CS-05-07, July, 2005. [16] http://www.w3.org/TR/soap/ [17] http://www.w3.org/TR/wsdl.html [18] http://www.uddi.org/ AUTHORS BIOGRAPHY DR. Yasser Ahmed Nada Was born in Ismailia, Egypt, in 1968. He received the BSc degree in pure Mathematics and Computer Sciences in 1989 and MSc degree for his work in computer science in 2003, all from the Faculty of Science, Suez Canal University, Egypt. In 2007, he received his Ph.D. in Computer Science from the Faculty of Science, Suez Canal University, Egypt. From September 2007 until now, he worked as Assistant Professor of computer science. Chair, Department of computer science, Faculty of Computers and Information Technology, Taif University, KSA. His research interests include Expert Systems, Artificial Intelligence, Object Oriented Programming, Computer Vision, and Genetic. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 12, 2013 187 | P a g e www.ijacsa.thesai.org http://www.w3.org/TR/soap/ http://www.w3.org/TR/wsdl.html http://www.w3.org/TR/wsdl.html http://www.uddi.org/ http://www.uddi.org/ 
work_hcjgrs4vozhbvoev4fpe5y2x4m	----	Investigation of publicﾒs perception towards rural sustainable development based on a two-level expert system Investigation of public’s perception towards rural sustainable development based on a two-level expert system Y.P. Cai a, G.H. Huang a,b,*, Z.F. Yang c, W. Sun a, B. Chen d a Environmental Systems Engineering Program, Faculty of Engineering, University of Regina, Regina, Saskatchewan, Canada S4S 0A2 b Chinese Research Academy of Environmental Science, Beijing Normal University, Beijing 100012-100875, China c State Key Laboratory of Water Environment Simulation, School of Environment, Beijing Normal University, Beijing 100875, China d Faculty of Engineering and Applied Science, Memorial University, St. John’s, NL, Canada A1B 3X5 a r t i c l e i n f o Keywords: Sustainable development Public’s perception Expert system Rural area China a b s t r a c t Sustainable development is of great significance in rural areas of China, which are under coupled pres- sures of poverty reduction, environmental protection and economic development. In these areas, agricul- ture is the primary sector in supporting their economies, where a number of the relevant production and processing practices cause many adverse impacts on environment and ecosystem. This poses many chal- lenges for decision makers in formulating sustainable development strategies without invoking potential controversies or objections from the public. In this study, a two-level expert system (TES) for facilitating the investigation of public’s perception towards sustainable development in rural areas was developed. Interactive relationships among objectives of socio-economic development and environmental protec- tion, as well as the related policy implications and public responses, were comprehensively examined and incorporated within TES. A two-level system was particularly adopted for facilitating participation of various stakeholders, including decision makers and farmers. The Yongxin County of the Shanjianghu Region in Jiangxi Province, China was selected to demonstrate the applicability of the developed TES. A series of questionnaire surveys were conducted both for acquiring knowledge about the interrelation- ships among sustainable development, agricultural production, poverty reduction and environmental protection. Public’s perception towards various system components of rural sustainable development were efficiently acquired and analyzed. The results indicated that sustainable development would be very significant on rural areas in China. Also, public awareness to the concept and profile of rural sustain- able development was still at a low level, especially for farmers. The developed TES can be not only used for acquiring knowledge of sustainable development, but also providing basic information for the formu- lation of relevant policies and strategies in rural areas of China. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction As the foundation for sustaining our daily life, agriculture is one of the most crucial sectors in supporting regional and national economies (Ambroise, Barnaud, Manchon, & Vedel, 1998; Doussan, Thannberger-gaillarde, & Thie0baut, 2000; Kim, Euan, & Luisa Isla, 2002; Landais, 1998; Lange, Hehl-Lange, & Brewer, 2008; Legg & Viatte, 2001; Marsh, 1997; Sartorius & Kirsten, 2007). Over the past decades, decision makers in many developing countries have been facing with the twin challenges of spurring agricultural produc- tions to feed a rapidly growing population and finding an economic engine to stimulate economic growth and reduce social poverty. In addition, they have had to confront the third urgent issue, i.e., the depleting natural resources and the degrading environmental qual- ity. Undoubtedly, human beings cannot afford to abandon goals of growth and poverty-alleviation for the sake of sustainability; un- less resources are managed in sustainable ways, current and future generations will find their welfare threatened (Harmancioglu, Fe- dra, & Barbaros, 2008; Huang, Huang, Hu, Maqsood, & Chakma, 2005; Murdiyarso, van Noordwijk, Puntodewo, Widayati, & Lusi- ana, 2008; Nader, Abi Salloum, & Karam, 2008). Thus, identify sus- tainable patterns for resources allocation, environmental protection and economic development in agricultural sector of rur- al areas is desired. Recently, many modes of sustainable development were exam- ined in China to mitigate intensive conflicts between economic development and environmental protection (Huang, Huang, Hu, Maqsood, & Chakma, 2005). Over the past two decades, the environment in China has been seriously impaired along with economic development. The accelerating economic development in this country has been resulting in a series of potential threats to the deteriorating ecosystem if there is lacking of effective 0957-4174/$ - see front matter � 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2008.11.032 * Corresponding author. Address: Faculty of Engineering, University of Regina, Regina, Saskatchewan, Canada S4S 0A2. E-mail address: huang@iseis.org (G.H. Huang). Expert Systems with Applications 36 (2009) 8910–8924 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a mailto:huang@iseis.org http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa management strategies for protecting the environment. Numerous events and studies have proved that the destruction of natural re- sources and the deterioration of environment could be disastrous to current and future socio-economic development. Particularly, sustainable development is critical in the efforts to achieve coordi- nation of social, economic, environmental goals in rural areas of China, which are more backward compared with prosperous cities in the country. However, in order to customize developing modes according to site-specific characteristics, many challenges are fac- ing by decision makers at multiple jurisdictions in rural areas, where up-to-date information and technologies, as well as suffi- cient economic resources are normally unavailable. Also, for pursu- ing aims of sustainable development in rural areas, various stakeholders (e.g., decision makers, residents, managers and third-party members) need to systematically consider many fac- tors and processes, such as targets of regional sustainability, con- sensus in local region for economic development and environmental protection, tradeoffs between eco-environmental preservation and socio-economic development. This poses a major obstacle for decision makers in formulating schemes of sustainable development without incurring potential controversies or objec- tions from the public. Therefore, it is desired to develop effective tools for aiding in acquiring and analyzing public’s perception to- wards local sustainable development and promoting the involve- ment of various stakeholders in rural areas of China, which can also be easily used by decision makers at multiple spatial scales. Previously, public participation/involvement was highly recom- mended as an effective way for solving public issues such as envi- ronmental protection and sustainable development, and to enhance acceptability of the relevant decision-making process (Cai, Xu, & Yang, 2002; Cai, Yang, & Xu, 2003; Cai, Huang, Nie, Li, & Tan, 2007; Cai, Huang, Yang, Lin, Bass, et al., 2008; Cai, Huang, Yang, Lin, & Tan, 2008; Cai, Huang, Yang, & Tan, 2008; Cai, Huang, Lin, Nie, & Tan, 2009; Tran, 2006). It is thus considered as a neces- sary process to formulate decision alternatives which may have di- rect or indirect impacts on the public. Over the past decade, there were many studies on public involvement/participation to help facilitate the management of natural resources, environmental quality, as well as the customization of sustainable development modes. For example, Haque, Kolba, Morton, and Quinn (2002) pro- posed a method for improving the degree of public participation in water resources management, water pollution control, as well as the related decision-making process. Dungumaro and Madulu (2003) made an attempt to incorporate various public participation measures within an integrated water resources management sys- tem for effectively identifying decision alternatives with high pub- lic acceptability in a watershed of Tanzania. Chopyak and Levesque (2002) examined the past, current, and future trends of the rela- tionship between science, technology, and society, indicating a shift to a raised public-participation level in relevant decision mak- ing across the world. In order to examine public’s perception to- ward potable and non-potable water reuse. Hartley (2006) proposed an interdisciplinary and integrative method and applied it to several communities in USA. Tran (2006) carried out a study to investigate public awareness on sustainable development in a small island, suggesting that public’s perception and involvement would be a significant factor for developing regional sustainable development schemes. Peterlin, Kontic, and Kross (2005) analyzed public’s perception toward environmental quality management in a coastal zone of Slovene through a number of statistics methods. In China, there were still quite a number of studies on public par- ticipation/involvement and its effects upon environmental man- agement, water resources management, ecological engineering, waste management and sustainable development. For instance, Tang, Wong, and Lau (2008) examined current situations/obstacles of public participation when conducting social impact assessment (SIA) and proposed a case study for public participation in Guang- dong Province, indicating that public’s perceptions toward results of SIA has received increasing attention from multi-level govern- ments across China. Dumreicher (2008) investigated the process of sustainable development in many villages of China through ana- lyzing public opinions on environmental quality and living space in rural areas. Li, Xiong, and Xu (2008) proposed a series of measures for: (a) improving public access to environmental information, (b) expanding environmental information disclosure among private enterprises, (c) increasing collection and disclosure of environmen- tal information, and (d) investigating and managing environmental information in rural areas. These studies either proposed various measures for promoting public participation in the process of conducting an individual pro- ject related to environmental management, ecological conserva- tion and sustainable development, or merely made a comprehensive review or investigation on the design, implementa- tion and evaluation of public participation. Moreover, a number of methods were proposed for acquiring public’s perception towards the management of land use, water resources, and environment, which are merely effective in a specific city and region. There is lacking of studies on public participation in rural areas for support- ing sustainable development, especially in developing countries. Also, the majority of the previous studies merely proposed meth- ods for acquiring public’s perception towards an individual devel- opment project or construction work at a problem-solving level, few of them suggested a universal tool for not only enhancing pub- lic participation but also improving effectiveness of decision mak- ing in environmental protection and socio-economic development. Furthermore, public participation in most of the studies was lim- ited to one specific group of people, such as local residents, deci- sion makers and enterprise managers. They could hardly handled public publication and acquired public’s perception in a broad and general context. There were few studies which could compre- hensively and dynamically integrate public participation into sus- tainable development schemes in rural areas. This would lead to a lack of effective and applicable measures for supporting sustain- able development especially in rural areas. Comparatively, expert system is considered as one of the effective tools in supporting decision making related to sustainable development at multiple scales. It can link a combination of decision-support tools (e.g., optimization modeling, and cost-benefit analysis) and various information components into a general decision-making process (Athanasiadis & Mitkas, 2007; Boyleand & Baetz, 1998; Carswell, Gardiner, Bertolotto, Rizzini, & Mandrak, 2008; Huang, Batez, & Patry, 1992, 1995a, 1995b; Lewis & Bardon, 1998; Tremblay, Hester, Mcleod, & Huot, 2004; Zeng, Zhang, Liu, & Yang, 2007; Zeng & Trauth, 2005; Yao, Liu, Fan, Yao, & Yang, 2006). With such an effective tool, both the knowledge of agricultural development and eco-environmental preservation, as well as public’s perception towards sustainable development can be interactively and dynam- ically acquired. Thus, a large quantity of data and information can be easily obtained without onerous efforts. Therefore, the objective of this study is to develop a two-level expert system (TES) for helping acquire public’s perception to- wards sustainable development in rural areas and apply it to the Sanjianghu Region of Jiangxi Province, China. Interactive question- naire design, visual display options and user interactions will be integrated into such a system, improving its applicability and robustness in dealing with real-world development problems in rural areas. Opinions from various stakeholders and decision mak- ers can thus be successfully identified for supporting the genera- tion of policies and strategies related to sustainable development, highly enhancing their public acceptability of the obtained plans. Practical measures and strategies can thus be properly customized in the study area. Relationships among socio-economy, eco-envi- Y.P. Cai et al. / Expert Systems with Applications 36 (2009) 8910–8924 8911 https://isiarticles.com/article/29346 
work_hecle6dizjhqjliitfldf33z6m	----	Microsoft Word - ESWA-semantic-similarity-final.doc Ontology-based semantic similarity: a new feature-based approach David Sánchez1, Montserrat Batet, David Isern, Aida Valls Intelligent Technologies for Advanced Knowledge Acquisition (ITAKA). Departament d’Enginyeria Informàtica i Matemàtiques, Universitat Rovira i Virgili, Avda. Països Catalans, 26. 43007 Tarragona (Spain) Abstract. Estimation of the semantic likeness between words is of great importance in many applications dealing with textual data such as natural language processing, knowledge acquisition and information retrieval. Semantic similarity measures exploit knowledge sources as the base to perform the estimations. In recent years, ontologies have grown in interest thanks to global initiatives such as the Semantic Web, offering an structured knowledge representation. Thanks to the possibilities that ontologies enable regarding semantic interpretation of terms many ontology-based similarity measures have been developed. According to the principle in which those measures base the similarity assessment and the way in which ontologies are exploited or complemented with other sources several families of measures can be identified. In this paper, we survey and classify most of the ontology-based approaches developed in order to evaluate their advantages and limitations and compare their expected performance both from theoretical and practical points of view. We also present a new ontology-based measure relying on the exploitation of taxonomical features. The evaluation and comparison of our approach’s results against those reported by related works under a common framework suggest that our measure provides a high accuracy without some of the limitations observed in other works. Keywords: semantic similarity, semantic relatedness, ontologies, feature-based similarity, WordNet 1 Corresponding author. Address: Departament d’Enginyeria Informàtica i Matemàtiques. Universitat Rovira i Virgili. Avda. Països Catalans, 26. 43007. Tarragona. Spain Tel.: +34 977 556563; Fax: +34 977 559710; E-mail: david.sanchez@urv.net. 1. Introduction With the enormous success of the Information Society and the World Wide Web, the amount of textual electronic information available has significantly increased. As a result, computer understanding of text has acquired great interest in the research community in order to enable a proper exploitation, management, classification or retrieval of textual data. One of the most basic problems when aiming to interpret textual data is the assessment of semantic likeness between terms because, as it has been demonstrated in psychological experiments (Goldstone, 1994), it acts as a fundamental organizing principle by which humans organize and classify objects. It is important to note that two different paradigms can be found in the literature. On one hand, semantic similarity states how taxonomically near two terms are, because they share some aspects of their meaning (e.g., dogs and cats are similar to the extend they are mammals). On the other hand, the more general concept of semantic relatedness does not necessary rely on a taxonomic relation (e.g., car and wheel or pencil and paper); other non taxonomic relationships (e.g., meronymy, antonymy, functionality, cause-effect, etc.) are also considered. Semantic similarity/relatedness computation has many direct and relevant applications. Some basic natural language processing tasks such as word sense disambiguation (Patwardhan, Banerjee, & Pedersen, 2003), synonym detection (Lin, 1998) or automatic spelling error detection and correction (Budanitsky & Hirst, 2001) rely on the assessment of words’ semantic resemblance. Direct applications can be found in the knowledge management field, such as thesauri generation (Curran, 2002), information extraction (M. Y. Chen, Chu, & Chen, 2010; P. Chen, Lin, & Chu, 2011; D. Sánchez & Isern, 2009; Stevenson & Greenwood, 2005) or ontology learning (D. Sánchez, 2010; David Sánchez & Antonio Moreno, 2008; D. Sánchez & A. Moreno, 2008), in which new terms related to already existing concepts, should be acquired from textual resources. The Semantic Web is an especially relevant application area, when dealing with automatic annotation of documents (Cimiano, Handschuh, & Staab, 2004; Chu, Chen, & Chen, 2009; D. Sánchez, Isern, & Millan, 2010) and text clustering (Song, Li, & Park, 2009). Despite its usefulness, robust measurement of semantic similarity/relatedness between textual terms remains a challenging task (Bollegala, Matsuo, & Ishizuka, 2007). Many works have been developed in the last years, especially with the increasing interest on the Semantic Web and the popularization of ontologies (Lanzenberger, et al., 2008). Proposed methods aim for automatically assessing a numerical score between a pair of terms according to the semantic evidence observed in one or several knowledge sources, which are used as semantic background. Ontologies have been of great interest for the semantic similarity research community as they offer a structured and unambiguous representation of knowledge in the form of conceptualizations interconnected by means of semantic pointers. These structures can be exploited in order to assess the degree of semantic proximity between terms. According to the principle in which the similarity/relatedness computation is based and the way in which the ontology is exploited and/or complemented with other sources (e.g., thesaurus, domain corpora, etc.), different families of methods can be identified. In this paper we survey and compare most of the ontology-based similarity/relatedness measures developed in recent years. We tried to collect as much relevant approaches as possible in order to offer an updated and detailed review and comparison of the expected performance of these measures both from theoretical and practical points of view. Concretely, for each family of functions, we identify their main advantages and limitations under the dimensions of expected accuracy, computational complexity, dependency on knowledge sources (type, size, structure- dependency and pre-processing) and parameter tuning. We also present a new feature-based method for semantic likeness assessment based on the exploitation of the taxonomical knowledge available in an ontology. By considering semantic evidence typically omitted by other feature-based approaches, this measure aims to rival state of the art ontology-based approaches in terms of accuracy, while retaining a low computational complexity. In order to compare all the semantic similarity/relatedness measures in a practical setting and evaluate them against our own approach in an objective manner, we collected the results reported by related works covered in the paper and tested our approach under the same conditions. Several widely used benchmarks have been considered in order to enable an objective comparison. The rest of the paper is organized as follows. Sections 2 surveys, reviews and compares related works in ontology-based semantic similarity/relatedness estimation, classifying them in different families. Section 3 presents the new method for measuring semantic likeness. Section 4 summarizes the evaluation results reported by related works for the analysed measures and tests the accuracy of our own approach under the same conditions. Section 5 discusses and compares the different measures according to the reported evaluation results. The final section contains the conclusions. 2. Ontology-based semantic similarity/relatedness Ontologies provide a formal specification of a shared conceptualization (Guarino, 1998). Being machine readable and constructed from the consensus of a community of users or domain experts, they represent a very reliable and structured knowledge source. Due to this reason, and thanks to initiatives such as the Semantic Web, which brought the creation of thousands of domain ontologies (Ding, et al., 2004), ontologies have been extensively exploited in knowledge-based systems (Valls, Gibert, Sánchez, & Batet, 2010) and, more precisely, to compute semantic likeness. A paradigmatic example is WordNet (Fellbaum, 1998), a domain-independent and general purpose thesaurus that describes and organizes more than 100,000 general English concepts, which are semantically structured in an ontological fashion. It contains words (nouns, verbs, adjectives and adverbs) that are linked to sets of cognitive synonyms (synsets), each expressing a distinct concept (i.e., a word sense). Synsets are linked by means of conceptual-semantic and lexical relations such as synonymy, hypernymy (is-a), six types of meronymy (part-of), antonymy, complementary and so on. The backbone of the network of words is the subsumption hierarchy which accounts more than an 80% of all the modelled semantic links, with a maximum depth of 16 nodes. The result is a network of meaningfully related words, where the graph model can be exploited to interpret the meaning of the concept. In fact, WordNet has been used as the background ontology in all related works. In this section, we survey and compare related works in ontology-based similarity assessment according to the following classification: 1. Edge-counting approaches. 2. Feature-based measures. 3. Measures based on Information Content. 2.1. Edge counting approaches Ontologies can be seen as a directed graph in which concepts are interrelated mainly by means of taxonomic (is-a) and, in some cases, non-taxonomic links. By mapping input terms to ontological concepts by means of their textual labels, a straightforward method to calculate their similarity is to compute the minimum Path Length connecting their corresponding ontological nodes via is-a links (Rada, Mili, Bichnell, & Blettner, 1989). The longer the path, the more semantically far the terms are. Let us define path(a,b)=l1,....lk as a set of links connecting the terms a and b in a taxonomy. Let |path(a,b)|=k be the length of this path. Then, considering all the possible paths from a to b, their semantic distance as defined by (Rada, et al., 1989) is (1). ( ) ( )bapatha,bdis iirad ,min ∀= (1) Several variations and improvements of this edge-counting approach have been proposed. On one hand, in addition to this absolute distance between terms, Wu and Palmer (Wu & Palmer, 1994) considered that the relative depth in the taxonomy of the concepts corresponding to the evaluated terms is an important dimension, because concept specializations become less distinct as long as they are recursively refined. So, equally distant pairs of concepts belonging to an upper level of a taxonomy should be considered less similar than those belonging to a lower lever. Wu and Palmer’s measure count the number of is-a links (N1 and N2) from each term to their Least Common Subsumer (LCS) (i.e., the most concrete taxonomical ancestor that subsumes both terms) and also the number of is-a links of the LCS to the root (N3) of the ontology (2). 321 3 & 2 2 ),( NNN N basim pw ×++ × (2) Based on the same principle, Leadcock and Chodorow (Leacock & Chodorow, 1998) also proposed a measure that considers both the number of nodes Np separating the ontological nodes corresponding to terms a and b, included themselves, and the depth D of the taxonomy in which they occur in a non-linear fashion (3). )2/log(),(& DNbasim pcl −= (3) Li et al., (Li, Bandar, & McLean, 2003) also proposed a similarity measure that combines the shortest path length and the depth of ontology information in a non-linear function (4). ( ) hh hh bapath li ee ee ebasim ii ββ ββ α − − − + − = ∀ ·),( ,min (4) , where h is the minimum depth of LCS in the hierarchy and α ≥0 and β> 0 are parameters scaling the contribution of shortest path length and depth, respectively. Based on a benchmark data, authors stated that the optimal parameters for the measure with respect to a concrete set of human judgements were: α =0.2; β =0.6. However, this is just an empirical finding for a specific setting. It lacks a theoretical basis and cannot be generalized. Al-Mubaid and Nguyen (H. Al-Mubaid & Nguyen, 2006) proposed a cluster-based measure that combines the minimum path length and the taxonomical depth. They define clusters for each of the branches in the hierarchy with respect the root node. They measure the common specificity of two terms by subtracting the depth of their LCS from the depth Dc of the cluster (5). )),((),( baLCSdepthDbaCSpec c −= (5) The common specificity is used to consider that lower level pairs of concept nodes are more similar than higher level pairs, as in Wu and Palmer approach. So, the proposed distance measure (sem) is defined as follows (6): ( ) ))()1,log((min),( kCSpecbapathbadis iisem +×−= ∀ βα (6) , where α>0 and β>0 are contribution factors of path length and common specify features and k is a constant. Authors use k=1 because with k≥1 they proved that the distance is positive. Moreover, in their experiments, they give an equal weight to the contribution of the two components (path length and common specify) by using α= β = 1. Both Li et al., And Al-Mubaid and Nguyen approaches are often considered in the literature (Petrakis, Varelas, Hliaoutakis, & Raftopoulou, 2006; Pirró, 2009) as “hybrid” approaches, as they combine several structural characteristics (such as path length, depth and local density) and assign weights to balance the contribution of each component to the final similarity value. Even though their accuracy for a concrete scenario (see evaluation section) is higher than more basic edge-counting measures, they depend on the empirical tuning of weights according to the ontology and input terms. Hirst and St-Onge (Hirst & St-Onge, 1998) extended the notion of taxonomical edge- counting by considering also non-taxonomic semantic links in the path (full_path). All types of relations found in WordNet together with rules that restrict possible semantic chains are considered, along with the intuition that the longer the path and the more changes in relation’s direction, the lower the likeness. The following path directions are considered: upward (such as hypernymy and meronymy), downward (such as hyponymy and holonymy) and horizontal (such as antonymy). The resulting formula is (7) ),(),(_),(& baturnskbapathfullCbasim sh ×−−= (7) , where C and k are constants (C = 8 and k = 1 are used by the authors), and turns(a, b) is the number of times the path’s direction changes. Due to the non-taxonomic nature of some of the relations considered during the assessment, Hirst and St-Onge’s measure captures a more general sense of relatedness than of taxonomical similarity, assessed by the approaches detailed above. The main advantage of the presented measures is their simplicity. They only rely on the graph model of an input ontology whose evaluation requires a low computational cost (in comparison to approaches dealing with text corpora, see Section 2.3). However, several limitations hamper their performance. First, edge-counting measures only consider the shortest path between concept pairs. When they are applied to wide and detailed ontologies such as WordNet that incorporate multiple taxonomical inheritance, the result is that several taxonomical paths are not taken into account. Other features also influencing the concept semantics, such as the number and distribution of common and non-common taxonomical ancestors are neither considered. As a result, by taking only the minimum path between concepts, many of the taxonomical knowledge explicitly modelled in the ontology is omitted. Another problem of path-based measures typically admitted (Bollegala, Matsuo, & Ishizuka, 2009; Wan & Angryk, 2007) is that they rely on the notion that all links in the taxonomy represent a uniform distance. In practice, the semantic distance among concept specializations/generalizations in an ontology would depend on the degree of granularity and taxonomic detail implemented by the knowledge engineer. 2.2. Feature-based measures Feature-based methods try to overcome the limitations of path-based measures regarding the fact that taxonomical links in an ontology do not necessary represent uniform distances. This is addressed by considering the degree of overlapping between sets of ontological features. As a result, they are more general and, potentially, they could be applied in cross ontology similarity estimation settings (i.e., when concept pairs belong to two different ontologies), a situation in which edge-counting methods cannot be directly applied (Petrakis, et al., 2006). So, on the contrary to edge-counting measures which, as stated above, are based on the notion of minimum path distance, feature-based approaches assess similarity between concepts as a function of their properties. This is based on the Tversky’s model of similarity (8), which, derived from the set theory, takes into account common and non common features of compared terms, subtracting the latter from the former ones. In fact, common features tend to increase similarity and non-common ones tend to diminish it (Tversky, 1977). Formally, let Ψ(a) and Ψ(b) be the features of terms a and b respectively, let Ψ(a) ∩ Ψ(b) be the intersection between those two sets of features, and Ψ (a)\ Ψ (b) the set obtained when eliminating the elements of Ψ(b) from the set of features of concept a, Ψ(a). Then, the similarity between a and b is proposed to be computed as a function of Ψ(a) ∩ Ψ(b), Ψ(a)\Ψ (b) and Ψ(b)\ Ψ (a) as. ))(\)(())(\)(())()((),( abFbaFbaFbasimtve ΨΨ⋅−ΨΨ⋅−Ψ∩Ψ⋅= γβα (8) , where F is a function that reflects the salience of a set of features, and α, β and γ are parameters that weight the contribution of each component. The definition of the set of features is crucial in this model. The existing approaches rely on information that is available in ontologies, in particular the set of synonyms (called synsets in Wordnet), definitions (i.e., glosses, containing textual descriptions of word senses) and different kinds of semantic relationships are considered. In Rodriguez and Egenhofer (Rodríguez & Egenhofer, 2003), the similarity is computed as the weighted sum of similarities between synsets, features (e.g., meronyms, attributes, etc.) and neighbour concepts (those linked via semantic pointers) of evaluated terms (9). ),(),(),(),(& baSvbaSubaSwbasim odsneighborhofeaturessynsetser ⋅+⋅+⋅= (9) , where w, u and v weight the contribution of each component, which depend on the characteristics of the ontology and S represents the overlapping between the different features, computed as: A\)),(1(B\),( ),( BbaAbaBA BA baS γγ −++∩ ∩ = (10) , where A, B are the terms evaluated for concepts corresponding to a and b, A\B is the set of terms in A but not in B and B\A the set of terms in B but not in A. Finally, γ(a, b) is computed as a function of the depth of a and b in the taxonomy as follows: ⎪ ⎪ ⎩ ⎪⎪ ⎨ ⎧ > + − ≤ + = )()(, )()( )( 1 )()(, )()( )( ),( bdepthadepth bdepthadepth adepth bdepthadepth bdepthadepth adepth baγ (11) In Petrakis et al., (Petrakis, et al., 2006) a feature-based function called X-similarity relies on the matching between synsets and a concept’s glosses extracted from WordNet (i.e., words extracted by parsing term definitions). They consider that two terms are similar if the synsets and glosses of their concepts and those of the concepts in their neighbourhood (following semantic relations) are lexically similar. The similarity function is expressed as follows: ⎪⎩ ⎪ ⎨ ⎧ = > =− 0),()},,(),,(max{ 0)(,1 ),( , baSifbaSbaS baSif basim synsetsglossesodsneighborho synsets SimilarityX (12) The similarity for the semantic neighbours Sneighborhoods is calculated as follows: ii ii i odsneighborho BA BA baS ∪ ∩ = ∈SR max),( (13) , where each different semantic relation type (i.e., is-a and part-of in WordNet) is computed separately and the maximum (considering all the synsets of all concepts up to the root of each hierarchy) is taken. Equivalently, the similarity for glosses Sglosses and synonyms Ssynsets are both computed as: BA BA baS ∪ ∩ =),( (14) , where A and B denote the set of synsets or glosses for the term a and b. Feature-based measures exploit more semantic knowledge than edge-counting approaches, evaluating both commonalities and differences of compared concepts. However, by relying on features like glosses or synsets (in addition to taxonomic and non-taxonomic relationships), those measures limit their applicability to ontologies in which this information is available. Only big ontologies/thesauri like WordNet include this kind of information. In fact, an investigation of the structure of existing ontologies via the Swoogle ontology search engine (Ding, et al., 2004) reveals that domain ontologies very occasionally model any semantic feature apart from taxonomical relationships. Another problem is their dependency on the weighting parameters that balance the contribution of each feature. In all cases, those parameters should be tuned according to the nature of the ontology and even to the evaluated terms. This hampers their applicability as a general purpose solution. Only the definition of Petrakis (Petrakis, et al., 2006) does not depend on weighting parameters, as the maximum similarity provided by each feature alone is taken. Even though this adapts the behaviour of the measure to the characteristics of the ontology, the contribution of other features is omitted if only the maximum value is taken at each time. 2.3. Measures based on Information Content Also acknowledging some of the limitations of edge-counting approaches, Resnik (Resnik, 1995) proposed to complement the taxonomical structure of an ontology with the information distribution of concepts evaluated in input corpora. He exploited the notion of Information Content (IC), by associating appearance probabilities to each concept in the taxonomy, computed from their occurrences in a given corpus. IC of a term a is computed according to the negative log of its probability of occurrence, p(a) (15). In this manner, infrequent words are considered more informative than common ones. )(log)( aPaIC −= (15) According to Resnik, semantic similarity depends on the amount of shared information between two terms, a dimension which is represented by their Least Common Subsumer (LCS) in an ontology. Two terms are maximally dissimilar if a LCS does not exist (i.e., in terms of edge-counting, it would be not possible to find a path connecting them). Otherwise, their similarity is computed as the IC of the LCS (16). )),((),( baLCSICbasimres = (16) One of the problems of Resnik’s metric is that any pair of terms having the same LCS results in exactly the same semantic similarity. Both Lin (Lin, 1998) and Jiang and Conrath (Jiang & Conrath, 1997) extended Resnik’s work by also considering the IC of each of the evaluated terms. Lin proposed that the similarity between two terms should be measured as the ratio between the amount of information needed to state their commonality and the information needed to fully describe them. As a corollary of this theorem, his measure considers, on one hand, commonality in the same manner as Resnik’s approach and, on the other hand, the IC of each concept alone (17). ))()(( ),(2 ),( bICaIC basim basim reslin + × = (17) The measure proposed by Jiang and Conrath is based on quantifying, in some way, the length of the taxonomical links as the difference between the IC of a concept and its subsumer. When comparing term pairs, they compute their distance by subtracting the sum of the IC of each term alone from the IC of their LCS (18). ),(2))()((),(& basimbICaICbadis rescj ×−+= (18) It is important to note that IC-based measures need, in order to behave properly, that the probability of appearance p monotonically increases as one moves up in the taxonomy (i.e., ∀ ci | cj is hypernym of ci => p(ci)≤ p(cj) ). This is achieved by computing p(a) as the probability of encountering any instance of a in the given corpus. In practice, each individual occurrence of any noun in the corpus is counted as an occurrence of each taxonomic class containing it (19) (Resnik, 1995). N wcount ap aWw ∑ ∈ = )( )( )( (19) , where W(a) is the set of nouns in the corpus whose senses are subsumed by a, and N is the total number of nouns in the corpus. As a result, an accurate computation of concept probabilities requires a proper disambiguation and annotation of each noun found in the corpus. This process is usually done manually to ensure the correctness of the tagging, hampering the scalability and applicability of this approach with large corpora. Moreover, if either the taxonomy or the corpus changes, re-computations are needed to be recursively executed for the affected concepts. So, it is necessary to perform a manual and time- consuming analysis of corpora and resulting probabilities would depend on the size and nature of input corpora. Moreover, the background taxonomy must be as complete as possible (i.e., it should include most of the specializations of each concept) in order to provide reliable results. Partial taxonomies with a limited scope may not be suitable for this purpose. All those aspects limit the scalability and applicability of those approaches (David Sánchez, Batet, Valls, & Gibert, 2009). Considering the limitations of IC-based approaches due to their dependency on corpora, some authors tried to intrinsically derive IC values from an ontology. These works rely on the assumption that the taxonomic structure of ontologies like WordNet is organized in a meaningful way, according to the principle of cognitive saliency (Blank, 2003). This states that humans specialise concepts when they need to differentiate them from already existing ones. So, concepts with many hyponyms (i.e., specializations) are more general and provide less information than the concepts at the leaves of the hierarchy. From the Information Theory point of view, they consider that abstract ontological concepts appear more probably in a corpus as they subsume many other ones. In this manner, the probability of appearance of a concept (i.e. the IC) is estimated as a function of the number of hyponyms and/or their relative depth in the taxonomy. Seco et al., (Seco, Veale, & Hayes, 2004) and Pirró and Seco (Pirró & Seco, 2008) base IC calculations on the number of hyponyms. Being hypo(a) the number of hyponyms of the concept a and max_nodes the number of hyponyms of the root node, they compute IC of a concept in the following way (20): )_log( )1)(log( 1)( nodesmax ahypo aICseco + −= (20) The denominator ensures that IC values are normalized in the range [0..1]. This approach only considers hyponyms of a given concept in the taxonomy; so, concepts with the same number of hyponyms but different degrees of generality appear to be equally similar. In order to tackle the problem, and in the same manner as for edge-counting measures, Zhou et al., (Zhou, Wang, & Gu, 2008) proposed to complement hyponym-based IC computation with the relative depth of each concept in the taxonomy. IC of a concept is computed as (21): ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ −+⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ + −= )_log( ))(log( )1( )_log( )1)(log( 1)( depthmax adepth k nodesmax ahypo kaIC zhou (21) In addition to hypo and max_nodes, which has the same meaning as eq. 20, depth(a) corresponds to the depth of the concept a in the taxonomy and max_depth is the maximum depth of the taxonomy. The factor k adjusts the weight of the two features involved in the IC assessment. They use k=0.5. Both ways of intrinsically computing IC have been applied directly on the similarity functions proposed by Resnik, Lin and Jiang and Conrath. Those approaches overcome most of the problems observed for corpus-based IC approaches (specifically, the need of corpus processing and their high data-sparseness) competing and even improving them in terms of accuracy (as it will be stated in the evaluation) when applied over WordNet. However, they require big, and fine grained taxonomies/ontologies with a detailed taxonomical structure in order to properly differentiate concept’s IC. For small or very specialized ontologies with a limited taxonomical depth and low branching factor, resulting IC values between concepts would become too homogenous to enable a proper differentiation. 3. A new feature-based measure exploiting taxonomical knowledge From the study of ontology-based semantic similarity measures presented in the previous section, the basic conclusions that we want to stress are the following: • Pure ontology-based measures, like the ones based on edge-counting, features and intrinsic IC computation are characterized by their simplicity and computational efficiency as they only exploit the semantic network provided by the ontology. Edge-counting measures are the simplest ones and, in consequence, their accuracies have been surpassed (as it will be stated in the evaluation section) by more complex approaches exploiting additional semantic evidence. Feature-based approaches, however, rely on features which are hardly found in domain ontologies, such as non taxonomic relationships, attributes, synonym sets or glosses. In consequence, their applicability and accuracy depend on the availability of this information. • Information Content approaches based on semantically-annotated textual data aim to improve pure ontology-based ones by capturing implicit semantic as a function of concept distribution in corpora. However, the association between the words found in a corpus and concepts which are needed to compute accurate concept appearance frequencies is not straightforward, requiring a process of manual word sense tagging for disambiguation. This hampers the applicability of these methods in practise, which are also affected by corpora availability and data sparseness. In this section we present a new measure that relies on taxonomical features extracted from an ontology. Being a feature-based method, our proposal follows a similar principle as proposed in the Tversky’s model (eq. 8), which considers that the similarity between two concepts can be computed as a function their common and differential features. Differently from previous feature-based approaches presented in section 2.2, we only rely on taxonomic information. This is due to the fact that available ontologies rarely model other kind of knowledge a part from taxonomical relationships (Ding, et al., 2004). As a matter of fact, as stated in section 2, taxonomical knowledge represent more than the 80% of the total amount of relationships modelled in WordNet. Moreover, on the contrary to other feature-based measures, as only one type of feature will be considered (i.e., taxonomic relationships) no tuning parameters will be used to weight the contribution of potentially scarce semantic features, overcoming one of the limitations observed in related works and improving the generality of our measure. 3.1. Evaluating concept dissimilarity Our measure considers as features the taxonomical categorization of concepts given by the ontology in order to evaluate the amount of dissimilarity (i.e. semantic distance) between concepts. Concretely, we consider that a term can be semantically distinguished from other ones by comparing the set of concepts that subsume it. Using the same notation introduced by Tversky in (Tversky, 1977), the following definitions formalize this idea. Definition 1. Let C be the set of concepts of an ontology, we define concept subsumption (≤) as a binary relation ≤ : C×C. Having two concepts ci and cj, ci ≤ cj is fulfilled if ci is a hierarchical specialization of cj or if ci=cj (i.e., they are the same concept even though they could be expressed by means of equivalent synonyms). The fact of including the concept itself in the subsumption relation assumes the notion of dominance as a reflexive relation (Partee, ter Meulen, & Wall, 1990). Definition 2. The set of taxonomical features describing the concept a is defined in terms of the relation ≤ as: }|{)( caCca ≤∈=φ (22) It is important to note that several immediate generalizations (i.e. categorizations) per concept may be available in ontologies modelling multiple inheritance, such in WordNet (Devitt & Vogel, 2004) or in detailed domain ontologies such as MeSH or SNOMED-CT (H. Al- Mubaid & Nguyen, 2006). So, in our case, the set of taxonomical features associated to the concept includes all the upper categories found when recursively going through all the upper taxonomical paths modelled in the ontology for that concept. Oppositely, ontology-based related works (Jiang & Conrath, 1997; Leacock & Chodorow, 1998; Lin, 1998; Rada, et al., 1989; Resnik, 1995; Wu & Palmer, 1994) deal with the case of multiple taxonomical generalizations per concept by taking the one that defines the maximum similarity with respect to another concept (e.g., taking the minimum path between both concepts). In our opinion, this simplification omits a large amount of explicitly available knowledge (i.e., other generalizations and their corresponding taxonomical paths). So, an important characteristic of our definition of taxonomical features (22) is that it does not introduce this limitation. Given two concepts which are semantically described according to their taxonomic features (those of eq. 22), we consider the degree of disjunction between their feature sets (non-common taxonomical features) as a function of their distance (or dissimilarity) whereas the degree of overlap (common ones) is proportional to their similarity. Considering the set of differential taxonomical features of a with respect to b as: )}()(|{)(-)()(\)( bcacCcbaba φφφφφφ ∉∧∈∈== , we formally define the dissimilarity between a and b as: Definition 3. The dissimilarity dis: C×C →ℜ between a and b is given by the cardinality of the set of differential features of a with respect to b and the set of differential features of b with respect to a: )(\)()(\)(),( abbabadis φφφφ += (23) In order to enable accurate comparisons between the dissimilarity computed for pairs of concepts corresponding to taxonomical branches with different degrees of taxonomical detail or multiple inheritance, the value given by (23) should be normalized taking into account the total size of the set of features. To include this normalizing factor, we divide the absolute dis value by the whole amount of features extracted for both terms. This value corresponds to the sum of cardinalities of differential and common taxonomical feature sets, that is )()()(\)()(\)( baabba φφφφφφ ∩++ . This will give a relative dissimilarity value in the [0..1] interval. As a final consideration, many authors argued that an information theoretic approach, in which semantic features are evaluated in a non linear fashion (H. Al-Mubaid & Nguyen, 2006; Leacock & Chodorow, 1998; Li, et al., 2003), approximates better the concept of similarity. Following a similar principle as intrinsic IC approaches (Seco, et al., 2004; Zhou, et al., 2008), in which the IC of the concept is computed as a non-linear function of the amount of semantic features, we introduce the logarithm in our calculation, obtaining the following formulation. Definition 4. The normalized dissimilarity between a and b according to their taxonomical features is calculated as: ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ ∩++ + += )()()(\)()(\)( )(\)()(\)( 1log),( 2 baabba abba badisnorm φφφφφφ φφφφ (24) In this expression, the logarithm is calculated by adding 1 to the ratio. In this manner, we avoid infinite values for equivalent terms (i.e., they are equal terms or exact synonyms corresponding to the same concept), which will lead to a zero numerator because there are not unique features differentiating them. Moreover, this expression maintains the value range of the results in the interval [0..1], being the boundary values: min_dis=log2(1+0)=0, for equivalent terms (i.e., ))()(|:()()( bxaxxiffba φφφφ ∈⇔∈∀= ). max_dis=log2(1+1)=1, for maximally different terms with disjoint feature sets resulting in equal numerator and denominator (i.e., ))()(|:()()( bxaxxiffba φφφφ ∉⇒∈∀≠ ). 3.2. Example To illustrate the behavior of our approach, let us consider the following portion of an ontology (Figure 1). At the lowest level we have four concepts: surfing, sailing, swimming and sunbathing. All of them are activities of leisure in the beach, but some of them are also sports related to wind and/or water. If only the minimal path connecting the concepts is considered, all of them are at the same distance because they are all brothers with respect to the concept beach leisure. So, this will be the answer given by the path length measures presented in section 2.1. With our proposal, which takes into account all the subsumers of the concepts, we are able to distinguish different levels of dissimilarity, which better captures the semantic distance that one would attach to those terms. <<Figure 1>> Figure 1. Ontology example. In our case, the set of features generated for the following four concepts are: { } { } { } { }activityleisurebeachsunbathingsunbathing urebeach_leisactivity,sport,water,swimming,swimming leisurebeachactivitysportwaterwindsailingsailing leisurebeachactivitysportwaterwindsurfingsurfing ,_,)( )( _,,,,,)( _,,,,,)( = = = = φ φ φ φ From these sets of features we can calculate the dissimilarity between any of those pairs of terms. Table 1 contains the results of applying eq. 24. As the measure is symmetric, we present only the lower triangular distance matrix. To give an example, the dissimilarity between sailing and sunbathing is calculated as follows: 78.0 214 14 1log),( 2 =⎟ ⎠ ⎞ ⎜ ⎝ ⎛ ++ + +=sunbathingsailingdisnorm , where the number of differential taxonomical features of sailing with respect to sunbathing is 4 (i.e., sailing, wind, water and sport) and the number of differential features of sunbathing with respect to sailing is 1 (i.e., sunbathing), and the set of common elements has a cardinality of 2 (i.e., beach_leisure and activity). Table 1. Dissimilarities calculated with disnorm for leaf concepts presented in Figure 1. Concepts Surfing Sailing Swimming Sunbathing Surfing 0 Sailing 0.36 0 Swimming 0.51 0.51 0 Sunbathing 0.78 0.78 0.65 0 For the ontology fragment in Figure 1, the concepts sailing and surfing are the most similar ones (least distance) as it was expected, because they are sports made in the water with the help of the wind, as well as, beach leisure activities (as it is explicitly stated in the ontology via taxonomical relationships). Swimming is more similar to sailing and surfing than sunbathing because it is also a water sport. In fact, sunbathing is only a relaxing activity in the beach not related with sports, so it is considered the semantically farthest to the rest of concepts. All those semantic evidences are properly captured by our approach by exploiting taxonomical knowledge (especially in the case of multiple inheritance) as it is presented in Table 1. 3.3. Properties In order to show the validity of the presented measure, we have studied the properties that a dissimilarity measure must fulfill. It is important to note that the fulfilment of those properties is a requirement if the measure is used in conjunction with some reasoning or data mining techniques (for example, similarity is a core element in case-based reasoning stages: case base building, case retrieval, and even case adaptation (O'Sullivan, Smyth, & Wilson, 2005)). In fact, the coherency of the results obtained in clustering algorithms relying on similarity measures may depend on the fulfilment of those properties (Everitt, Landau, & Leese, 2001). In this respect, several of the measures proposed in related works (especially weighted and feature- based ones (Li, et al., 2003; Rodríguez & Egenhofer, 2003; Tversky, 1977)), do not accomplish basic properties such as minimality, hampering their applicability in some data mining algorithms. Proposition 1. The function disnorm fulfills the properties of dissimilarity measures (Euzenat & Shvaiko, 2007): 0),(,, ≥Ο∈∀ badisba (positiveness) 0),(, =Ο∈∀ aadisa (minimality) ),(),(,, abdisbadisba =Ο∈∀ (symmetry) Proof. By adding 1 to the inner expression of the logarithm, values positively range from 0 to 1 so, positiveness is ensured. Equivalent terms will result in an empty set of unique features and in a disnorm value of log2(1)=0, accomplishing the second property that says that the minimum dissimilarity must be given by the comparison of one item with itself. Finally, as a and b are evaluated according to their feature sets (i.e., the order of the elements in the sets is irrelevant) and all the operations performed over those sets are commutative (i.e., difference and intersection), the measure accomplishes the symmetry property. □ 3.4. Dealing with polysemic terms When input terms are raw words extracted from text, some of them may be polysemous. For general ontologies such as WordNet, polysemic words correspond to several concepts (i.e., one per word sense) which can be found by mapping words to concept synsets. As a matter of fact, in WordNet 2, polysemic nouns correspond to an average of 2.77 concepts2. A proper disambiguation of input terms may solve the ambiguity, assigning input words to unique ontological concepts. However, as stated in the introduction, semantic similarity is promoted exactly for applications dealing with various levels of ambiguity because in texts we find words rather than concepts (Budanitsky & Hirst, 2006). In previous works, polysemic words have been tackled by retrieving all possible concepts corresponding to a term and then, computing individual similarities for each possible pair of 2 http://wordnet.princeton.edu/wordnet/man2.1/wnstats.7WN.html concepts and selecting, as the final result, the maximum similarity value obtained. The rationale for this criterion is that in order to evaluate the similarity between two non-disambiguated words (i.e., no context is available), human subjects would pay more attention to their similarities (i.e., most related senses) rather than their differences, as it has been demonstrated in psychological studies (Tversky, 1977). Therefore, we have taken the same approach to solve this problem, taking the minimum dissimilarity value obtained for all the possible combinations. Definition 5. The generalized dissimilarity measure which is able to deal with polysemic terms is defined as: )','(min),( ' ' badisbadis norm Bb Aadgeneralize ∈∀ ∈∀ = (25) , where A is the set of concepts (i.e., word senses) for the term a, and equally for term b. Proposition 2. The function disgeneralized fulfills the properties of dissimilarity measures: positiveness, minimality and symmetry. Proof. The proof is straightforward considering that disnorm fulfills the positiveness, minimality and symmetry as the minimum operator keeps the properties when applied to it. □ 4. Results As stated in (Bollegala, et al., 2009), an objective evaluation of the accuracy of a semantic similarity function is difficult because the notion of similarity is a subjective human judgement. In order to enable fair comparisons, several authors created evaluation benchmarks consisting of word pairs whose similarity were assessed by a set of humans. Rubenstein and Goodenough (Rubenstein & Goodenough, 1965) defined the first experiment in 1965 in which a group of 51 students, all native English speakers, assessed the similarity of 65 word pairs selected from ordinary English nouns on a scale from 0 (semantically unrelated) to 4 (highly synonymous). Miller and Charles (Miller & Charles, 1991) re-created the experiment in 1991 by taking a subset of 30 noun pairs which similarity was reassessed by 38 undergraduate students. The correlation obtained with respect to Rubenstein and Goodenough experiment was 0.97. Resnik (Resnik, 1995) replicated again the same experiment in 1995, in this case, requesting 10 computer science graduate students and post-doc researchers to assess similarity. The correlation with respect to Miller and Charles results was 0.96. Finally, Pirro (Pirró, 2009) replicated and compared the three above experiments in 2009, involving 101 human subjects, both English and non-English native speakers. He obtained an average correlation of 0.97 with respect to Rubenstein and Goodenough experiment, and 0.95 with respect to Miller and Charles experiment. It is interesting to see the high correlation obtained between the experiments even though being performed in a period of more than 40 years and with heterogeneous sets of people. This states that similarity between selected words is stable over the years, making them a reliable source for comparing measures. In fact, Rubenstein and Goodenough and Miller and Charles benchmarks have become de facto standard tests to evaluate and compare the accuracy of similarity measures. Authors quantify the accuracy of their measures by computing the correlation between the similarity ratings reported in these benchmarks against those obtained by means of the computerized assessments. If the two ratings are exactly the same which means that the similarity function perfectly mimics human judgements, correlation coefficient will be 1, whereas 0 means that automatic assessments are unrelated to human opinions. Spearman’s and Pearson’s correlations coefficients have been commonly used in the literature; both are equivalent if ratings sets are ordered (which is the case). They are also invariant to linear transformations which may be performed over results such as a change between distance and similarity by changing the sign of the value or normalizing values in a range. The use of these benchmarks and the correlation coefficient as a measure of evaluation enables an objective comparison between measures. In order to evaluate the accuracy of related works commented in section 2, we have taken the correlation values originally reported by related works for Rubenstein and Goodenough and Miller and Charles benchmarks (when available) and summarized them in Table 2. In the case in which a concrete measure depends on certain parameters (such as weights or corpora selection/processing) the best correlation value reported in authors’ experiments according to optimum parameter tuning was compiled. It is important to note that, even though some of them rely on different knowledge sources (such as tagged corpora), all measures use WordNet as background ontology. Concretely, WordNet 2 is the most common version used in related works. In cases in which original authors used an older version (WordNet 2 was released in July 2003), we took a more recent replication of the measure evaluation performed by another author in order to enable a fair comparison. At the end, we picked up correlation results reported by authors in papers published from 2004 to 2009. In order to evaluate and compare our approach against related works under the same conditions, we applied it to the two benchmarks also using WordNet 2 as ontology. The correlation values obtained in our case are shown in the last row of Table 2. Table 2. Correlation values for each measure. From left to right: authors, measure type, correlation for Miller and Charles benchmark, correlation for Rubenstein and Goodenough benchmark and reference in which those correlations where reported Measure Type M&C R&G Evaluated in Rada et al., (path length) Edge-counting 0.59 N/A (Petrakis, et al., 2006) Wu and Palmer Edge-counting 0.74 N/A (Petrakis, et al., 2006) Leacock and Chodorow Edge-counting 0.74 0.77 (Patwardhan & Pedersen, 2006) Li et al., Edge-counting 0.82 N/A (Petrakis, et al., 2006) Al-Mubaid and Nguyen (sem) Edge-counting N/A 0.815 (Hisham Al-Mubaid & Nguyen, 2009) Hirst and St-Onge Edge-counting 0.78 0.81 (Wan & Angryk, 2007) Rodriguez and Egenhofer Feature 0.71 N/A (Petrakis, et al., 2006) Tversky Feature 0.73 N/A (Petrakis, et al., 2006) Petrakis et al., (X-similarity) Feature 0.74 N/A (Petrakis, et al., 2006) Resnik IC (corpus) 0.72 0.72 (Patwardhan & Pedersen, 2006) Lin IC (corpus) 0.7 0.72 (Patwardhan & Pedersen, 2006) Jiang and Conrath IC (corpus) 0.73 0.75 (Patwardhan & Pedersen, 2006) Resnik (IC computed as Seco et al.,) IC (intrinsic) N/A 0.829 (Zhou, et al., 2008) Lin (IC computed as Seco et al.,) IC (intrinsic) N/A 0.845 (Zhou, et al., 2008) Jiang and Conrath (IC computed as Seco et al.,) IC (intrinsic) N/A 0.823 (Zhou, et al., 2008) Resnik (IC computed as Zhou et al.,) IC (intrinsic) N/A 0.842 (Zhou, et al., 2008) Lin (IC computed as Zhou et al.,) IC (intrinsic) N/A 0.866 (Zhou, et al., 2008) Jiang and Conrath (IC computed as Zhou et al.,) IC (intrinsic) N/A 0.858 (Zhou, et al., 2008) Our approach (eq. 25) Feature 0.83 0.857 - 5. Discussion From the results reported in Table 2, in the following, we will make a comparative analysis of the different measures covered in this paper. Together with their accuracy, their main advantages and drawbacks from the application point of view will be discussed. The basic path length measure (Rada, et al., 1989) presents the lowest accuracy (0.59) due to the fact that absolute lengths of the paths connecting two concepts may not accurately represent their specificity. This is the case of WordNet, since concepts higher in the hierarchy are more general than those lower in the hierarchy (Pirró, 2009). As a result, other edge-counting approaches also exploiting the relative depth of the taxonomy (Wu and Palmer (Wu & Palmer, 1994), Leadcock and Chodorow (Leacock & Chodorow, 1998)) offer a higher accuracy (0.74). It is remarkable the correlation values obtained by Li (Li, et al., 2003) and Al-Mubaid and Nguyen approaches (H. Al-Mubaid & Nguyen, 2006), which combine the length of the path with the depth of the concepts in a weighted and non-linear manner. However, they rely on empirical parameters whose values have been experimentally determined to optimize the accuracy for the evaluated benchmark, hampering their generality. Hirst and St-Onge (Hirst & St-Onge, 1998) presents a similar behaviour, also relying on tuning parameters but, in this case, using non-taxonomic relationships that consider a more general concept of relatedness. Feature-based methods try to overcome the limitations of path-based measures by considering different kinds of ontological features. The problem, which has been also noted for some edge-counting measures, is their dependence on the parameters introduced to weight the contribution of each feature (for Rodriguez and Egenhofer (Rodríguez & Egenhofer, 2003) and Tversky (Tversky, 1977) approaches). Correlation values are, however, very similar to those offered by edge-counting measures (0.71-0.74) in these benchmarks. This can be motivated by the fact that they rely on concept features, such as synsets, glosses or non-taxonomic relationships which have secondary importance in ontologies like WordNet in comparison with taxonomical knowledge. In fact, those kind of features are scarce in ontologies (Ding, et al., 2004), which causes those approaches are based on partially modelled knowledge. As a result, those measures, even being more complex, are not able to significantly outperform the state of the art of edge-counting measures. For IC-based measures, we observe that approaches relying on an intrinsic computation of IC (based on the number of concept hyponyms) clearly outperform approaches relying on corpora (0.72 vs. 0.84, in average). This is very convenient as corpora dependency seriously hampers the applicability of classic IC measures. The difference between both ways of computing IC is caused by two factors. Firstly, the data sparseness problem that appear when relying on tagged corpora (which would be necessary small due to manual tagging) to obtain accurate concept appearance frequencies. Secondly, the fact that WordNet’s taxonomy is detailed and fine-grained, which enables an accurate estimation of a term’s generality as a function of its number of hyponyms. With regards to the performance of each measure, Lin (Lin, 1998) tends to improve Resnik (Resnik, 1995) one when IC is computed intrinsically, as the former is able to differentiate terms with identical LCS but different taxonomical depths. With regards to the way in which intrinsic IC is computed, more complex approaches also exploiting relative depth and relying on weighting parameters (Zhou et al., (Zhou, et al., 2008)) offer the highest accuracy (0.86). Comparing the proposed measure with other ontology-based works, one can note that our approach’s accuracy surpasses the basic edge-counting approaches (0.83 vs. 0.74). In general, in complex and detailed ontologies like WordNet, where multiple taxonomical paths can be found connecting concept pairs (overlapping hierarchies), path-based measures waste explicitly available taxonomical knowledge as only the minimum path is considered as an indication of distance. Only the Li et al.,’s measure is able to achieve a very similar accuracy when the appropriate scaling parameters are empirically chosen. Feature-based approaches’ correlations are also surpassed (0.84 vs. 0.74), even though they are based on other non-taxonomical features and weighting parameters. This shows that taxonomical knowledge plays a more relevant role in stating term similarity than other more scarce features which are typically poorly considered in available ontologies. The same situation is repeated for corpus-based IC measures (0.84 vs. 0.73) showing that the exploitation of high quality taxonomical knowledge available in ontologies provides even more reliable semantic evidences than unstructured textual resources. This is coherent to what is observed for approaches computing IC in an intrinsic manner, which, conceptually, follow a similar principle as our approach. In their case, similarity is computed as a function on the number of hyponyms whereas in our case it is estimated as a function of overlapping and non- overlapping hypernyms. Moreover, Lin and Jiang and Conrath measures computing IC intrinsically follow the same principle as feature-based measures: similarity is proportional to feature overlapping (in their case, represented by the IC of the LCS) and inversely proportional to the differences (in their case, the IC of each individual term). So, if the IC is computed from taxonomical evidences (i.e., number of hyponyms) it is coherent that their correlation values are similar as those of our approach. The only case in which they surpass our measure’s correlation is when IC is computed as Zhou et al.’s (Zhou, et al., 2008), in which a weighting parameter is introduced to optimize the assessment. Summarizing, our measure is able to provide a high accuracy without any dependency on data availability, data pre-processing or tuning parameters for a concrete scenario. As it only relies on the most commonly available ontological feature, our measure ensures its generality as a domain-independent proposal. At the same time, it retains the low computation complexity and lack of constraints of edge-counting measures as it only requires retrieving, comparing and counting ontological subsumers. This ensures its scalability when it must be used in engineering or data mining applications, which may require dealing with large sets of terms (Armengol, 2009; Batet, Valls, & Gibert, 2008). Compared to other approaches based on taxonomical knowledge, the exploitation of the whole amount of unique and shared subsumers seems to give solid semantic evidences of semantic resemblance. First, the distinctive features implicitly include information about the different paths connecting the pair of terms. In the same manner, the depth of the Least Common Subsumers of those concepts is implicitly included in the set of shared subsumers (i.e., the deeper the LCS, the higher the amount of common features). Other features that have been identified in the literature, such as relative taxonomical densities and branching factors, are also implicitly considered, being all of them useful dimensions to assess semantic similarity. As any other ontology-based measure, the final accuracy will depend on the coverage, detail, completeness and coherency of taxonomical knowledge. Moreover, most of the improvements achieved by our approach are derived from the fact that similarity is estimated from the total set of subsumer concepts considering the different taxonomical hierarchies. If the input ontology offers little taxonomical detail or does not consider multiple inheritance, the accuracy improvements of our approach with respect to measures based on the minimum path are likely to be less noticeable. Fortunately, large and broad ontologies are being developed, like WordNet as a general purpose description of concepts, or the UMLS repository in the medical context. 6. Conclusions As it has been explained in the introduction, semantic similarity assessment is a crucial component embedded in many applications framed in the artificial intelligence research area. This paper provides an up-to-date survey of ontology-based semantic similarity measures that can be used to estimate the resemblance between terms. A new measure based on taxonomical features has been also presented and compared in the context of the survey. In this measure the set of features is built from the categorization (i.e., subsumers) of the concepts modelled in the ontology. In our case, we consider subsumers as labels that describe the meaning of the concept in different levels of generality. Differently to the other feature-based approaches, this measure only relies on taxonomic ontological knowledge (which is the most commonly available one), lacking of corpora- dependency or parameter-tuning. It is computational efficient as only taxonomical branches are explored and it fulfils the mathematical properties required in many applications for coherent similarity computation (Everitt, et al., 2001; O'Sullivan, et al., 2005). The paper has analysed, under a common framework, the pros and cons of both related works and our proposal, with the aim of giving some insights on their accuracy, applicability, dependencies and limitations. In addition, a complete comparison of all these measures in a practical setting is reported, using the two widely used benchmarks. The conclusions extracted from those analyses would help practitioners in selecting the measure that better fits with the requirements of a concrete application. In particular, the results reported by our measure for the two benchmarks suggest a promising accuracy, improving the correlations reported by most of other ontology-based approaches, while minimizing the constraints that may hamper its applicability both from the computational efficiency and resource-dependency points of view. For this reason, as future work, we want to study how the inclusion of this measure can improve the results in some concrete applications. In particular, we are studying semantically grounded data mining processes and statistical disclosure control methods, which may directly benefit from a more accurate similarity assessment. Acknowledgements This work has been partially supported by the Universitat Rovira i Virgili (a pre-doctoral grant of M. Batet, a post-doctoral grant of D. Isern and 2009AIRE-04) and the Spanish Ministry of Science and Innovation (DAMASK project, Data mining algorithms with semantic knowledge, TIN2009-11005) and the Spanish Government (PlanE, Spanish Economy and Employment Stimulation Plan). References Al-Mubaid, H., & Nguyen, H. A. (2006). A cluster-based approach for semantic similarity in the biomedical domain. In 28th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS 2006 (pp. 2713–2717). New York, USA: IEEE Computer Society. Al-Mubaid, H., & Nguyen, H. A. (2009). Measuring Semantic Similarity between Biomedical Concepts within Multiple Ontologies. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, 39, 389-398. Armengol, E. (2009). Using explanations for determining carcinogenecity in chemical compounds. Engineering Applications of Artificial Intelligence, 22, 10-17. Batet, M., Valls, A., & Gibert, K. (2008). Improving classical clustering with ontologies. In 4th World Conference of the IASC and 6th Conference of the Asian Regional Section of the IASC on Computational Statistics & Data Analysis, IASC 2008 (pp. 137-146). Yokohama, Japan: International Association for Statistical Computing. Blank, A. (2003). Words and Concepts in Time: Towards Diachronic Cognitive Onomasiology. In R. Eckardt, K. von Heusinger & C. Schwarze (Eds.), Words and Concepts in Time: towards Diachronic Cognitive Onomasiology (pp. 37-66). Berlin, Germany: Mouton de Gruyter. Bollegala, D., Matsuo, Y., & Ishizuka, M. (2007). Measuring Semantic Similarity between Words Using Web Search Engines. In C. Williamson & M. E. Zurko (Eds.), 16th international conference on World Wide Web, WWW 2007 (pp. 757-766). Banff, Alberta, Canada ACM. Bollegala, D., Matsuo, Y., & Ishizuka, M. (2009). A Relational Model of Semantic Similarity between Words using Automatically Extracted Lexical Pattern Clusters from the Web. In P. Koehn & R. Mihalcea (Eds.), Conference on Empirical Methods in Natural Language Processing, EMNLP 2009 (pp. 803–812). Singapore, Republic of Singapore: ACL and AFNLP. Budanitsky, A., & Hirst, G. (2001). Semantic distance in WordNet: An experimental, application-oriented evaluation of five measures. In Workshop on WordNet and Other Lexical Resources, Second meeting of the North American Chapter of the Association for Computational Linguistics (pp. 10- 15). Pittsburgh, USA. Budanitsky, A., & Hirst, G. (2006). Evaluating wordnet-based measures of semantic distance. Computational Linguistics, 32, 13-47. Cimiano, P., Handschuh, S., & Staab, S. (2004). Towards the self-annotating web. In 13th international conference on World Wide Web, WWW 2004 (pp. 462 - 471). New York, USA: ACM. Curran, J. R. (2002). Ensemble Methods for Automatic Thesaurus Extraction. In Conference on Empirical Methods in Natural Language Processing, EMNLP 2002 (pp. 222–229). Philadelphia, PA, USA: Association for Computational Linguistics. Chen, M. Y., Chu, H. C., & Chen, Y. M. (2010). Developing a semantic-enable information retrieval mechanism Expert Systems with Applications, 37, 322-340. Chen, P., Lin, S. J., & Chu, Y. C. (2011). Using Google latent semantic distance to extract the most relevant information Expert Systems with Applications, 38, 7349-7358. Chu, H. C., Chen, M. Y., & Chen, Y. M. (2009). A semantic-based approach to content abstraction and annotation for content management Expert Systems with Applications, 36, 2360-2376. Devitt, A., & Vogel, C. (2004). The topology of WordNet: Some Metrics. In P. Sojka, K. Pala, P. Smrz, C. Fellbaum & P. Vossen (Eds.), The Second Global Wordnet Conference, GWC 2004 (pp. 106- 111). Brno, Czech Republic: Masaryk University. Ding, L., Finin, T., Joshi, A., Pan, R., Cost, R. S., Peng, Y., Reddivari, P., Doshi, V., & Sachs, J. (2004). Swoogle: A Search and Metadata Engine for the Semantic Web. In thirteenth ACM international conference on Information and knowledge management, CIKM 2004 (pp. 652-659). Washington, D.C., USA: ACM Press. Euzenat, J., & Shvaiko, P. (2007). Ontology Matching: Springer Verlag. Everitt, B. S., Landau, S., & Leese, M. (2001). Cluster Analysis. London: Arnold. Fellbaum, C. (1998). WordNet: An Electronic Lexical Database. Cambridge, Massachusetts: MIT Press. Goldstone, R. L. (1994). Similarity, interactive activation, and mapping. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 3-28. Guarino, N. (1998). Formal Ontology in Information Systems. In N. Guarino (Ed.), 1st International Conference on Formal Ontology in Information Systems, FOIS 1998 (pp. 3-15). Trento, Italy: IOS Press. Hirst, G., & St-Onge, D. (1998). Lexical chains as representations of context for the detection and correction of malapropisms. In C. Fellbaum (Ed.), WordNet: An Electronic Lexical Database (pp. 305–332): MIT Press. Jiang, J. J., & Conrath, D. W. (1997). Semantic Similarity Based on Corpus Statistics and Lexical Taxonomy. In International Conference on Research in Computational Linguistics, ROCLING X (pp. 19-33). Taiwan. Lanzenberger, M., Sampson, J., Kargl, H., Wimmer, M., Conroy, C., O'Sullivan, D., Lewis, D., Brennan, R., Ramos Gargantilla, J. Á., Gómez-Pérez, A., Fürst, F., Trichet, F., Euzenat, J., Polleres, A., Scharffe, F., & Kotis, K. (2008). Making Ontologies Talk: Knowledge Interoperability in the Semantic Web. IEEE Intelligent Systems, 23, 72-85. Leacock, C., & Chodorow, M. (1998). Combining local context and WordNet similarity for word sense identification. In WordNet: An electronic lexical database (pp. 265-283): MIT Press. Li, Y., Bandar, Z., & McLean, D. (2003). An Approach for Measuring Semantic Similarity between Words Using Multiple Information Sources. IEEE Transactions on Knowledge and Data Engineering, 15, 871-882. Lin, D. (1998). An Information-Theoretic Definition of Similarity. In J. Shavlik (Ed.), Fifteenth International Conference on Machine Learning, ICML 1998 (pp. 296-304). Madison, Wisconsin, USA: Morgan Kaufmann. Miller, G. A., & Charles, W. G. (1991). Contextual correlates of semantic similarity. Language and Cognitive Processes, 6, 1-28. O'Sullivan, D., Smyth, B., & Wilson, D. C. (2005). Understanding case-based recommendation: A similarity knowledge perspective. International Journal on Artificial Intelligence Tools, 14, 215- 232. Partee, B., ter Meulen, A., & Wall, R. (1990). Mathematical Methods in Linguistics: Kluwer Academic Publishers. Patwardhan, S., Banerjee, S., & Pedersen, T. (2003). Using Measures of Semantic Relatedness for Word Sense Disambiguation. In A. F. Gelbukh (Ed.), 4th International Conference on Computational Linguistics and Intelligent Text Processing and Computational Linguistics, CICLing 2003 (Vol. 2588, pp. 241-257). Mexico City, Mexico: Springer Berlin / Heidelberg. Patwardhan, S., & Pedersen, T. (2006). Using WordNet-based Context Vectors to Estimate the Semantic Relatedness of Concepts. In EACL 2006 Workshop on Making Sense of Sense: Bringing Computational Linguistics and Psycholinguistics Together (pp. 1-8). Trento, Italy. Petrakis, E. G. M., Varelas, G., Hliaoutakis, A., & Raftopoulou, P. (2006). X-Similarity:Computing Semantic Similarity between Concepts from Different Ontologies. Journal of Digital Information Management, 4, 233-237. Pirró, G. (2009). A semantic similarity metric combining features and intrinsic information content. Data & Knowledge Engineering, 68, 1289-1308 Pirró, G., & Seco, N. (2008). Design, Implementation and Evaluation of a New Semantic Similarity Metric Combining Features and Intrinsic Information Content. In R. Meersman & Z. Tari (Eds.), OTM 2008 Confederated International Conferences CoopIS, DOA, GADA, IS, and ODBASE 2008 (Vol. 5332, pp. 1271-1288). Monterrey, Mexico: Springer Berlin / Heidelberg. Rada, R., Mili, H., Bichnell, E., & Blettner, M. (1989). Development and application of a metric on semantic nets. IEEE Transactions on Systems, Man, and Cybernetics, 9, 17-30. Resnik, P. (1995). Using Information Content to Evalutate Semantic Similarity in a Taxonomy. In C. S. Mellish (Ed.), 14th International Joint Conference on Artificial Intelligence, IJCAI 1995 (Vol. 1, pp. 448-453). Montreal, Quebec, Canada: Morgan Kaufmann Publishers Inc. . Rodríguez, M. A., & Egenhofer, M. J. (2003). Determining semantic similarity among entity classes from different ontologies. IEEE Transactions on Knowledge and Data Engineering, 15, 442–456. Rubenstein, H., & Goodenough, J. (1965). Contextual correlates of synonymy. Communications of the ACM, 8, 627-633. Sánchez, D. (2010). A methodology to learn ontological attributes from the Web. Data & Knowledge Engineering, 69, 573-597. Sánchez, D., Batet, M., Valls, A., & Gibert, K. (2009). Ontology-driven web-based semantic similarity. Journal of Intelligent Information Systems, 35, 383-413. Sánchez, D., & Isern, D. (2009). Automatic extraction of acronym definitions from the Web. Applied Intelligence, doi: 10.1007/s10489-009-0197-4 (in press). Sánchez, D., Isern, D., & Millan, M. (2010). Content Annotation for the Semantic Web: an Automatic Web-based Approach. Knowledge and Information Systems, doi:10.1007/s10115-010-0302-3, (in press). Sánchez, D., & Moreno, A. (2008). Learning non-taxonomic relationships from web documents for domain ontology construction. Data & Knowledge Engineering, 63, 600-623. Sánchez, D., & Moreno, A. (2008). Pattern-based automatic taxonomy learning from the Web. AI Communications, 21, 27-48. Seco, N., Veale, T., & Hayes, J. (2004). An Intrinsic Information Content Metric for Semantic Similarity in WordNet. In R. López de Mántaras & L. Saitta (Eds.), 16th Eureopean Conference on Artificial Intelligence, ECAI 2004, including Prestigious Applicants of Intelligent Systems, PAIS 2004 (pp. 1089-1090). Valencia, Spain: IOS Press. Song, W., Li, C. H., & Park, S. C. (2009). Genetic algorithm for text clustering using ontology and evaluating the validity of various semantic similarity measures. Expert Systems with Applications, 36, 9095-9104. Stevenson, M., & Greenwood, M. A. (2005). A semantic approach to IE pattern induction. In K. Knight (Ed.), 43rd Annual Meeting on Association for Computational Linguistics, COLING-ACL 2005 (pp. 379–386). Ann Arbor, Michigan, USA: Association for Computational Linguistics. Tversky, A. (1977). Features of Similarity. Psycological Review, 84, 327-352. Valls, A., Gibert, K., Sánchez, D., & Batet, M. (2010). Using ontologies for structuring organizational knowledge in Home Care assistance. International Journal of Medical Informatics, 79, 370-387. Wan, S., & Angryk, R. A. (2007). Measuring Semantic Similarity Using WordNet-based Context Vectors. In M. El-Hawary (Ed.), IEEE International Conference on Systems, Man and Cybernetics, SMC 2007 (pp. 908 - 913). Montreal, Quebec, Canada: IEEE Computer Society. Wu, Z., & Palmer, M. (1994). Verb semantics and lexical selection. In 32nd annual Meeting of the Association for Computational Linguistics (pp. 133 -138). Las Cruces, New Mexico: Association for Computational Linguistics. Zhou, Z., Wang, Y., & Gu, J. (2008). A New Model of Information Content for Semantic Similarity in WordNet. In S. S. Yau, C. Lee & Y.-C. Chung (Eds.), Second International Conference on Future Generation Communication and Networking Symposia, FGCNS 2008 (pp. 85-89). Sanya, Hainan Island, China: IEEE Computer Society. 
work_helpidgwavcrvfyywi2ro6i5sa	----	Decision support for foreign investment strategy under hybrid uncertainty Decision support for foreign investment strategy under hybrid uncertainty Po-yuan Chen a,⇑, Horng-jinh Chang b a Department of Financial and Tax Planning, Jinwen University of Science and Technology, No. 99, An-chung Rd., Sindian District, New Taipei City 23154, Taiwan, ROC b Graduate Institute of Management Sciences, Tamkang University, No. 151, Ying-chuan Rd., Tamsui District, New Taipei City 25137, Taiwan, ROC a r t i c l e i n f o Keywords: Decision support Foreign investment Hybrid uncertainty a b s t r a c t Empirical evidences show that Japan-based companies moved their major operations to the USA due to the currency appreciation of Japanese Yen in 1980s. However, the multinational firms moving their oper- ations abroad still face both the risk of foreign price and the risk of foreign exchange rate. According to the purchasing power parity (PPP) and interest rate parity (IRP), the foreign exchange rate has associa- tions with the relative price and the relative interest rate between two countries. Therefore, when the price and the interest rate evolve stochastically as proposed by many scholars, we can anticipate the ran- domness of the foreign exchange rate. By integrating these three sources of risks as a hybrid uncertainty, we propose a framework for corporate valuation and investment strategy. We analytically derive corpo- rate value for those multinationals in question and then numerically obtain the real option value by Monte Carlo simulations, based on which we investigate the optimal entry strategy. The sensitivity for corporate value, optimal entry time, and real option value are analyzed. The results suggest a decision support process for foreign investment timing strategy under the hybrid uncertainty. The managerial implication of this work is that the optimal investment strategy for the multinational firms with foreign operations depends on some market and risk factors, as well as correlations among them. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Global business community today faces risk exposures not only from volatile product prices but also from foreign exchange rates. Recently, many countries compete with their counterparts for currency depreciation, attempting to gain export advantage. As a consequence, the game of currency arises, making the foreign ex- change market more volatile and weakening the profits of multina- tional firms with foreign operations. For example, Japanese Yen reached its 15 yrs high up to $84 JPY/USD in 2010, causing many Japan-based firms abroad to suffer from losses denominated in domestic currency. Kim and Hur (2009) addressed the importance of foreign exchange management and proposed a model to inves- tigate the effects of foreign exchange on firm performance using genetic algorithm and VaR. Kogut and Chang (1996) found that in order to maintain the exports value of domestic operations un- der the appreciation of Japanese yen against the US dollar, Japanese companies switched their supporting export-oriented trades to establishing the foreign operations directly in the United States during the period of 1970s through 1980s. Studies of Rangan (1995) found that American firms tend to keep their ex factory dollar prices constant, allowing foreign currency prices to fluctuate with exchange rates. By contrast, Japanese and German firms were found to set foreign prices in a manner sensitive to market condi- tions. Some researchers also advocated that the volatility of foreign exchange rate is related to the trends towards foreign direct invest- ments. According to the purchasing power parity (PPP), the rela- tionship between the foreign exchange rate and the price has been tested by many researchers, for example by Perron and Vogelsang (1992). Dornbusch (1988) asserts that the movements in the nominal foreign exchange rate between two countries and the movements in their respective price indexes adjust through time to maintain the constant purchasing power. Moreover, the interest rate parity (IRP) suggests that exchange rate is also corre- lated with interest rates in two countries. We therefore consider the foreign exchange rate, the interest rate and the foreign price as mutually correlated stochastic processes. We aim to formulate a stochastic model to value the multinationals which move the operations to foreign countries. Regarding the corporate valuation approaches, Seng and Lai (2010) classifies business valuation methods into three categories: market-based, income-based (discounted cash flow, DCF) and as- set-based approaches. The asset-based model dates back to the work of Merton (1974), who proposed the stochastic asset value in a call option, whose exercise price is the contingent claim on debt. Based on his option framework, many related models have been developed by Black and Cox (1976), Landes and Loistl 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.10.007 ⇑ Corresponding author. Fax: +886 2 8212 2399. E-mail address: mikecpy@ms7.hinet.net (P.-y. Chen). Expert Systems with Applications 39 (2012) 4582–4589 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.10.007 mailto:mikecpy@ms7.hinet.net http://dx.doi.org/10.1016/j.eswa.2011.10.007 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa (1992), etc. More recent study on option-type investment strategy can be found in the work of Yuan (2009), who proposed a frame- work to investigate expansion strategy under price uncertainty from the aspect of an industry. Both the real option value and the investment timing strategy are obtained through simulation– optimization mechanism in this model. On the other hand, the works employing the DCF to valuate business can be found in the works of Rappaport (1986), and Copeland, Koller, and Murrin (1996). The main theme of these works lies in the discounted pres- ent value of all future cash inflows. Casey (2001) computed the discounted value from a series of uncertain future payments until the date of insolvency and then obtained the present value distribution. Inspired by previous research results, we propose a valuation framework by integrating the foreign exchange rate, the foreign price, and the interest rate as a hybrid uncertainty for those com- panies with foreign operations. Subsequently, we investigate the entry strategy for the foreign operations and obtain the corre- sponding real option value. In the work of Kim, Won, and Bae (2010), corporate value is de- rived from the present value of future dividend flows. Instead, the corporate value is derived in our work by discounting all future profits. These profit flows evolve as a stochastic process due to the assumption of randomness in interest rate, foreign exchange rate and foreign price, as illustrated in Figs. 1 and 2 based on the historical price quotes for the LCD product denominated in New Taiwan dollar (NTD) and on the quotes for the NTD exchange rate against the US dollar (USD) during the period of January 2000 through June 2009 acquired from the Taiwan Economic Journal (TEJ) database. The sensitivity of corporate value is analyzed to provide readers with more insights into the interactions among the following factors: the drift and the volatility of the stochastic foreign exchange rate and the foreign price, the price elasticity of demand, and the correlation between the exchange rate and the foreign price. This paper is organized as follows: Section 1 introduces the cor- responding valuation models, introduced by previous scholars. Section 2 presents the formulation of corporate value in the con- text of stochastic price, interest, and exchange rates under an in- verse demand function. The analytical solution for the proposed corporate value is then obtained. Based on derived corporate value, Monte Carlo simulations for real option value and optimal entry Fig. 1. The monthly price process denominated in US dollar for the LCD product. Fig. 2. The monthly foreign exchange rate process quoted as NT dollars against per US dollar. P.-y. Chen, H.-j. Chang / Expert Systems with Applications 39 (2012) 4582–4589 4583 https://isiarticles.com/article/12541 
work_hf2e4ztvcnd2nffwbuqu6amt7q	----	dp9625ll 1 GIS, Expert Systems and Interoperability Linda Lilburne1, George Benwell2 & Roz Buick3 1 Landcare Research P. O. Box 69 Lincoln New Zealand Ph +64 (3) 325 6700 Fax +64 (3) 325 2418 2 Department of Information Science Otago University P. O. Box 56 Dunedin New Zealand 3Trimble Navigation P. O. Box 8729 Christchurch New Zealand Email: lilburnel@landcare.cri.nz 1. Introduction How should geographic information systems be developed? There is a strong demand from users for enhanced functionality and power. Vendors can and do respond to these demands. But where will this lead? Will the result be one all-embracing and all- conquering program or geographic information system (GIS)? A GIS could grow to incorporate all statistical functions, all visualisation techniques, all data management functions etc. It is possible to perceive a scenario in which GIS is developed to ‘bloatware’ proportions. An alternative scenario is one in which a GIS is interfaced with other software systems. Embedding database bridges and other product-specific links, providing data import and export routines, and system calls are all ways of interfacing GIS with other systems. GIS vendors could opt to produce a ‘linkware’ GIS, interfaced to as many third party systems as possible. Given these two alternatives to GIS development, an interesting set of questions arises. How far do vendors go with enhancing their systems compared with interfacing with third party systems? Is there a balance? Or do GIS users just keep calling for ‘more’, regardless of the solution set? There is a balance. GIS is likely to be developed by being enhanced AND by being interfaced with third party software. In a way, this is a third developmental track leading to an increasingly functional GIS whose ability to interact with other systems is greatly improved. This interoperable GIS allows flexible combinations of system components while still providing a comprehensive range of spatial operations and analytical functions. Figure 1 depicts the three developmental tracks, leading to the ‘bloated’ GIS, the linked GIS, or the interoperable GIS in an environment in which systems can cooperate. 2 Of these three developmental tracks, this paper presents an example of what can be achieved with the interoperable GIS. Expert systems are introduced along with the client/server and object-oriented paradigms. By using these paradigms, a generic, spatial, rule-based toolbox called SES (spatial expert shell) has been created. SES is described using examples and contrasted with other documented expert system – GIS linkages. But first integration is modelled in three dimensions to highlight the need for improvements in how GISs can interact with other systems. 2. The integration cube Integration has been described by Fedra (1993), Goodchild (1992), and Nyerges (1992) from two perspectives: data and user interface. A more comprehensive model of integration can be described by adding a third perspective: interoperability. This relates to the ability of two systems, or processes, to cooperate with each other. Processes, which can exchange a variety of dynamically determined requests and information, have high interoperability. The interwoven nature of the requests allows the functions of each system to be highly integrated. More commonly however, integration is achieved by starting, executing and exiting a second process either after, or at some predefined point within, the first process. The flow of functionality between the systems is sequential. One process is followed by a second process, after which control may return to the first system. Interoperability is low. By using the three dimensions of user interface, data and interoperability, integration can be portrayed as a cube (Figure 2). This gives a more complete picture of a coupling between two systems. The data axis progresses from no exchange of data, transfer of data, to sharing data. The user interface axis extends from a position of no link between two interfaces, to a trigger (e.g. a menu option or button) between two interfaces, to two interfaces which have a similar look and feel, through to a single interface. The Bl oa t w ar e GI S Figure 1 Three developmental tracks 3 interoperability axis ranges from static, sequential integrations with limited functionality to dynamic, interleaved linkages with full functionality. The degree of integration is represented by its position in the cube. The front bottom left corner represents no integration; the back top right corner of the cube represents a fully integrated system. A sequentially integrated system can now be differentiated from a system in which the components of the integrated system can freely interact with each other. This model is defined in detail in Lilburne (1996). Lilburne also demonstrates how most integrations are positioned on the front plane of the cube, i.e. have low interoperability scores. 3. GIS and Expert Systems An expert system is a “computer program designed to model the problem-solving ability of a human expert” (Durkin, 1994 p. 7). It comprises a knowledge base (KB), an inference engine, and working memory. The expert’s domain knowledge is stored in the KB. The inference engine can process or reason with this knowledge to draw conclusions about a problem. The working memory contains the facts and deductions made in a session. Expert systems are used to identify, monitor, diagnose, predict, control, specify, design, configure, and plan (Jackson, 1990). An expert system can combine many types of knowledge including intuition, experience, qualitative beliefs, heuristics, empirical observations, and expert judgement. The advantages of integrating a GIS with an expert system have been recognised by a number of authors (Burrough, 1986; Fedra, 1995; Fischer, 1994; Fisher et al., 1988; Lein, 1992; Leung, 1993; Robinson et al., 1987; Smith and Yiang, 1991; Zhu and Healey, 1992). These authors saw the expert system as having the potential to add User Interface Integration 0 Figure 2: Integration cube 4 intelligence to GIS tasks, e.g. map design, generalisation, automated name placement, feature extraction, spatial query. Fischer (1994) notes the increasing use of AI techniques to represent meta-data. Burrough (1992) envisions how an expert system could be an adviser or tutor for using a GIS. Domain knowledge represented in an expert system together with spatial data can provide a decision support environment in which users are guided by the integrated system towards a recommendation. Table 1 contrasts the strengths and weaknesses typically observed in expert systems and GISs. This both shows strengths that are complementary and how some of the weaknesses of one system are matched by strengths in the other. Expert system GIS qualitative quantitative imprecise data precise data uses symbols uses geometric primitives eg point, line segment integrates knowledge integrates data handles incomplete data and knowledge does not easily handle incomplete data suited to unstructured problems suited to structured problems no spatial capability spatially capable handles incomplete data and knowledge does not easily handle incomplete data does not cope well with lots of data copes with large volumes of data explanation facility no explanation facility can represent knowledge is not designed to represent knowledge can manage knowledge can not easily manage knowledge has inference engine no inference or reasoning capability opportunistic algorithmic i.e. sequential no mapping/graphing capability variety of output maps/graphics can not efficiently do arithmetic operations can efficiently perform geometrical ops. Table 1 Comparison of some expert system and GIS characteristics The vision of those promoting expert system – GIS linkages was developed five or more years ago. To date, linkages have not fulfilled their vision. In part this is due to over- optimistic rhetoric about the capabilities of expert systems. It also relates to the level of interoperability between a GIS and an expert system. For example, one application of an expert system – GIS linkage is to assist with solving or understanding environmental problems which are often very complex. GIS allows the real world to be modelled in its spatial context. Knowledge of real world objects and processes can be represented in the expert system. Sometimes the best knowledge 5 available is in the form of heuristics. Combining spatial data with knowledge offers real opportunities in the management of our natural resources (Fedra, 1995). However, the real world is not a series of sequential processes; rather it is a complex tapestry of interactions between objects and processes. Hence the degree of interoperability between an expert system and a GIS will affect the ability of an integrated system to model the complexity of the real world. Recent technological advances offer new potential for closer interaction between systems, in particular the client/server and object-oriented technologies. 4. Technological paradigms Client/server technology refers to the software that allows a process to receive messages from another process. These messages request services of the receiving system (the server). The service might be to perform a specified action or to return some information to the requesting system (the client). Both processes remain in memory concurrently, avoiding the loss of performance that occurs when loading a system into memory every time one of its functions is required. There is no limit to the number of requests, nor are there any restrictions on the types of requests that can be made. Both systems must conform to a common client/server protocol. There are incompatible client/server protocols, e.g. DDE, OLE (PC), RPC (UNIX), APPC (IBM). Object-oriented (OO) technology is based on objects which have an identity, state and behaviour (Booch, 1994). ‘State’ refers to the data or values associated with the object at a particular point in time. ‘Behaviour’ is how an object acts and reacts. Key characteristics of OO technology that are important in a GIS – expert system link are abstraction, encapsulation, inheritance, and polymorphism. Abstraction is a simplified description of an object which encompasses all of its essential characteristics. Encapsulation, or information hiding, is the ability to hide implementation details from the user. Inheritance allows objects to inherit behaviour and state from parent objects. Polymorphism is the ability to redefine or override inherited behaviour from parent classes. Abstracting the essential qualities that are useful for a given domain and grouping those objects with similar qualities is a powerful way of simplifying the computer representation of a problem. These essential qualities characterise the state and the behaviour of objects. Behaviour is defined by abstracting the essential operations that an object can perform and which can be performed on it. A vector class in an expert system, defines the state and behaviour of GIS vector layers. The state of a vector object includes the name of the GIS layer, its description, default colour and its physical location. Behaviour is described in methods which are associated with the vector class. These methods are routines which encapsulate the GIS commands to draw, delete, create, modify and manipulate GIS layers. The commands are hidden from the user who does not need to know how GIS operations are implemented. Encapsulation also serves to hide the complexities of GIS data representation. Objects inherit state and behaviour from parent classes. For example, a roads object, representing a GIS road network layer, inherits the state and behaviour of its parent 6 vector class. Each object representing a GIS layer inherits a constructor method called create. This method encodes commands needed to select the required number of features from the layer, create a sub-object for each feature and transfer the GIS attributes to the sub-object. This allows the details of how to create objects that are sourced from a GIS to be associated with the appropriate object, rather than buried in a transformation routine written in a 3GL1. This facilitates maintenance of data exchange between the two systems. Multiple inheritance is very useful in a GIS – expert system combination, as it allows an object to inherit behaviour from application specific classes or objects as well as GIS related classes. For example, a road_segment object inherits state and behaviour from the GIS line class, the specific road_1 object(s) that the segment is part of and the roads vector layer object. Polymorphism is useful to override inherited behaviour. For example, objects can inherit a display method which instructs the GIS to draw the object appropriately. Polymorphism allows a specialised display method to override the inherited one during processing. 5. SES design SES (Spatial Expert Shell) integrates two commercial products: the GIS ARC/INFO (ESRI Inc., 1991), and the expert system shell, Smart Elements (Neuron Data Inc., 1994). ARC/INFO v7 includes some new commands (IAC2request, IACconnect, IACdisconnect, IACreturn) which create a framework for client/server communication with another process. Once a connection has been initialised, messages can be sent between the processes. Smart Elements is a combination of a hybrid frame, rule-based expert shell called Nexpert Object, and a GUI3 developers kit called Open Interface. It has an Application Programming Interface (API) which allows C routines to access Smart Elements functions. SES is developed on a Solaris SUN Workstation platform which both ARC/INFO and Smart Elements support. Both systems use Sun’s ONC-RPC client/server protocol. Smart Elements is the client and ARC/INFO is the server. A combination of C and ARC/INFO’s macro language AML is used to develop the client/server interface between ARC/INFO and Smart Elements. Essentially SES is a collection of spatial classes with predefined state and behaviour. The expert system shell is extended to include spatial classes. GIS elements (e.g. the display window, vector data, raster data) are modelled as classes under a top level class gisObject (Figure 3). Generalised classifications of spatial data are modelled as classes under the gisObjectType class. Appropriate spatial methods are associated with all of these classes. For example, polygon related state (area, perimeter) and behaviour (draw, adjacency, overlap) is defined in the slots and methods of the polygon class. An attribute of the state is stored in a ‘slot’. Methods describe the operations or behaviour of an 1 3GL = Third generation language 2 IAC = Inter-application communication 3 GUI = Graphical user interface 7 object. Vector and grid classes have slots and methods that define vector and raster layer behaviour, e.g. how layers should be drawn, how features can be selected, and in which colour they should be drawn. Figure 3 SES class diagram A spatial expert system application can be developed by creating objects belonging to the spatial classes that make up SES. For example, in an agricultural application, there might be GIS layers of paddocks, tracks, streams, buildings, soil types, and a DTM. Objects would be created for each of these layers, linked to either the vector or the grid class. The objects inherit slots from these parent classes, e.g. the paddocks vector object inherits the colour slot and its default value “white” from the vector class. The paddocks object’s default colour value could be redefined to “yellow”. Methods are also inherited. For example, the paddocks, tracks, streams, and buildings objects all inherit the selectFeature method from the vector class. This enables an application expert system 8 developer to transparently access a GIS function in which GIS features are selected from a map. For example, the DTM object inherits a raster specific draw method and default colour scheme from the parent grid class. The application developer creates additional slots defining the domain state of each vector layer object. In the agricultural example, the soils object might have inheritable slots which represent the GIS attributes of the soil layer, e.g. soil name, soil code, pH level, soil depth. The create method associated with the soils object will dynamically create sub-objects of the soils object. Each sub-object, soil_1, soil_2 etc., represents a single feature in the vector layer, e.g. a polygon. The sub-objects inherit domain slots, i.e. soil_name, soil_depth etc. which are populated by attribute values from the GIS. The sub-objects are linked to gisObjectType classes so that GIS state and behaviour can also be inherited. In Figure 4, sub-object soil_1 inherits GIS slots and methods from the polygon class and inheritable domain slots and methods from the soils object. The soils object inherits GIS slots and methods from the vector class. These are not inherited by its sub-objects. The spatial methods attached to spatial classes access C procedures which perform error checking, creation and management of the information request which is sent to ARC/INFO, and manipulation of the reply. All spatial operations are executed in ARC/INFO. No locational data is manipulated in the expert system. Methods are used to access spatial relationships between objects. This enables the spatial context of an object Figure 4 Soil example object diagram 9 to be referenced in heuristic knowledge. For example, a SES rule base could dynamically and transparently determine the truth of the following conditions: • if forest stand is adjacent to road • if the nearest firestation = “Albany” • if bus route is longer than 5km • if parcel is within the Christchurch District boundary • if paddock contains sandstone • if site is at least 200 m from the nearest water source • if habitat is above 1000m Each of the conditions requires the GIS to perform either an adjacency, nearest neighbour, route distance, contained within, overlay, buffer and/or a raster map algebra operation. These operations are encapsulated in generic spatial methods which are inherited by the domain objects (e.g. parcel, forest_stand) from spatial classes. In the first example, the forest_stand object sends its inherited getAdjacent method to the roads object. In addition to the GIS providing information about spatial data and relationships, the expert system can dynamically access the full presentational and analytical functionality of a GIS. This is possible through provision of a global gisExecute method in Smart Elements. This method takes a string argument. First the method determines and substitutes GIS names and locations for any spatial objects referred to in the string. The string, now a valid GIS command that runs a macro, is passed to the ARC/INFO process which then executes it. The expert system can request the GIS to display the results of an inference session on a map, or complex spatial analyses such as pattern analysis, multivariate analysis, location/allocation, hydrological or viewshed analysis can be requested. The results might then be interpreted by the expert system, upon completion of the request. Both raster objects and vector objects can be manipulated in a rule base. The Smart Elements network diagram in Figure 5 demonstrates how sample heuristics defining suitability of land for forestry can utilise combinations of raster and vector data. Forest suitability depends on the land’s aspect, elevation, soil type, landuse capability code, and distance from roads. The lri4 object refers to a soil polygon layer and the track object refers to a line layer. Aspect, elevation, nw, high, nw_high, sunny, pukaki, luc5_6e_6c, luc_lessthaneq_5, mid, road_at_least_200 and not_nw_high objects all represent raster maps. The road_at_least_200 object represents land that is at least 200m from a track. It is generated by executing its inherited create method with an argument string “> 200 track”. This string is stored in the criteria slot. Similarly the raster pukaki object is generated by executing the create method with an appropriate argument which selects, then rasterises all the polygons in the lri GIS layer with a soil type of Pukaki, i.e. “lri soil cn Pukaki”. The raster objects are weighted according to their relative importance. The raster objects and weights are passed to the GIS by the 4 lri = land resource inventory which is a national GIS layer of soil and vegetation data 5 luc = landuse capability code e.g. 6e 10 weight method which combines the GIS layers appropriately. Finally the suit_forest object is displayed by the GIS by sending an inherited draw method to itself. Figure 5 Network diagram modelling rule-based suitability of land for forestry 6. Discussion The key advantage of an approach with high interoperability is flexibility. In SES, the GIS linkages required by a knowledge-based application are determined at runtime. They are not hardwired into the application. Spatial information is only accessed when it is required by the inference engine. The alternative approach is to calculate all the spatial relationships that might be required, and preload them as facts into the knowledge base, e.g. (Bleecker et al., 1990; Loh and Rykiel, 1992). This approach is only suitable for a limited range of spatial information, e.g. an adjacency matrix, stream connectivity. The dynamic and generic nature of the linkage allows easy update of the domain knowledge base. The interwoven nature of the communication between the two processes allows the spatial and knowledge-based functions of the two systems to be effectively integrated. Figure 6 shows where SES is positioned in the integration cube. 11 There is very high interoperability between ARC/INFO and Smart Elements. Data is shared and there are two interfaces which can be designed to have a similar look and feel. This flexible linkage was achieved with minimal development effort as SES brings together the considerable power of the expert system shell Smart Elements and the GIS ARC/INFO. This was achieved by using a client/server approach. The strengths of each system are maintained. SES maintains the computational power of a GIS by performing all geometric and raster map operations in the GIS. Raster maps are mapped to objects in Smart Elements which can be managed by the rule base. The expert system does not manipulate individual cell values. Knowledge is represented in Smart Elements using a symbolic representation (objects). This is more intuitive to work with than a quantitative representation (Sharma et al., 1994). Knowledge can be easily modified. A knowledge base developer can modify knowledge on the fly without needing to alter code or the linkage mechanism. Explanation capabilities provided in Smart Elements are available to the integrated system, as is the ability to handle incomplete data and opportunistic reasoning. GIS database management facilities can be used by the integrated system to access and manage data. The full set of GIS presentational functions are accessible in SES. A limitation of SES is that while the intricacies of GIS operations are encapsulated in spatial classes, the use of spatial objects in rules is not completely transparent. For example, in a conditional statement: If paddock = high where high refers to the elevation, the syntax of the rule varies according to whether high is an aspatial attribute (i.e. “high”, “mod”or “low”), a raster map representation of cells with values over 1000m in an area of interest, or a fuzzy raster representation. The Interface Figure 6 SES in the integration cube 12 knowledge base developer must be aware of these differences and structure the rules accordingly. The design of SES is portable to other combinations of layer-based GISs and hybrid object/rule-based expert system shells that support client/server functionality. For example, Smart Elements could use the DDE protocol to make requests of an ARCVIEW (ESRI Inc., 1994) server by passing appropriate AVENUEJ commands. For SES to be portable between GISs, spatial services must be standardised and made accessible to client processes by vendors. Requests can then be made in a format that is independent of which GIS product is the server. There is a move in this direction by the Open GIS Consortium (1996) which is developing a specification for distributed geoprocessing. SES is based on a client/server relationship in which Smart Elements is the client and ARC/INFO the server. A reverse architecture is possible in which a client ARC/INFO process requests a Smart Elements server to run an inference session. For example, a rule base of diagnostic heuristics could be accessed from ARC/INFO. The user can be prompted for some information but all other information required by the rule base must be passed to Smart Elements before the rule(s) is processed. This is because a Smart Elements – ARC/INFO linkage is synchronous. In a synchronous linkage, the requesting process must wait for the server process to complete the request before it can continue processing. If in the diagnostic example, a GIS data value or relationship is required, this can not be dynamically accessed. The GIS is too busy waiting for its original request to finish, to service any requests made to it. The need to pass all potentially relevant facts to a rule base inhibits the flexibility of the system and is not very efficient. Spatial facts, especially distance related facts can quickly become quite extensive. An asynchronous link, in which processes do not have to wait for each other, would allow concurrent bi-directional requests. Knowledge accessed by a client needs to be modularised into discrete reasoning segments. This is usually possible with meta-data and classification knowledge. For example, an expert system server can inform a GIS client of the validity of a value, or an expert system can be requested to fire a rule base to classify a series of data values supplied by a GIS client. The expert system-as-server/GIS-as-client architecture is appropriate in a GIS controlled application where one or more modular rule bases are accessed to perform a specified classification, diagnosis, data validation, and/or recommendation. Meta-knowledge is not appropriate in this architecture. However, an expert system application is often a system in which many strands of knowledge are interwoven together. Control knowledge is integrated with domain knowledge, meta-data, and an intelligent interface. For example, a DSS6 expert system application might combine an intelligent interface including appropriate question windows, recommendations, and explanations, with both knowledge about the process of determining a solution and knowledge about the domain objects. The expert system- as-client/GIS-as-server architecture (i.e. SES) is more suitable for a system in which 6 Decision support system 13 multiple types of knowledge are integrated. GIS tasks can be modularised and thus made accessible to the expert system client which controls the integrated system. Other integration approaches are defined by Lilburne (1996). One approach taken by some authors is to develop an in-house GIS and/or expert system which is very demanding of resources, e.g. Davis (1991), Lam (1993). Loose and merged/embedded approaches where the systems are sequentially executed have low interoperability and are usually inflexible, but these approaches require minimal resources to implement a link. An enhanced approach where one of the systems is extended to incorporate functions normally performed by the other system results in a subset of the total functionality available to SES. For example, basic GIS display functions were incorporated into an expert system based on PROLOG (Crossland, 1990). A tight approach usually requires access to the source code of the systems being integrated. 7. Conclusion SES is a powerful, flexible, generic toolbox which can be used to represent knowledge from any domain in a spatial context. Use of the object-oriented and client/server paradigms have enabled a highly interoperable linkage, accessing the full range of functionality, to be established between two powerful systems. Minimal resources were required to achieve this. SES supports our belief that moving towards a truly interoperable GIS is essential in today’s interlinked world of distributed systems. An interoperable GIS allows effective use of external techniques and systems. In particular, an interoperable GIS expands the potential of combining knowledge with GIS. So, returning to the question of the balance between developing functionality versus interfaces, there is a need for vendors to follow the third developmental track and further improve the interoperability of GIS. 8. Acknowledgments This research was funded by the Foundation for Research, Science and Technology, New Zealand. Rhys Gibson’s cheerful assistance with Smart Elements is much appreciated, as are the comments by the reviewers: Paul Luckman, Megan Ogle- Mannering and Grant Hunter. 9. References Open GIS Consortium, Inc. 1996. Open GIS. WWW ref: http://www.opengis.org. Bleecker, M., Hutson, J. L., and Waltman, S. W. 1990. Mapping groundwater contamination potential using integrated simulation modeling and GIS. Proceedings of Application of geographic information systems, simulation models, and knowledge-based systems for landuse management, Blacksburg, VA. Booch, G. 1994. Object-oriented analysis and design, The Benjamin/Cummings Publishing Company Inc., Redwood City, California. 14 Burrough, P. A. 1986. Principles of geographical information systems for land resources assessment, Oxford University Press. Burrough, P. A. 1992. Development of intelligent geographical information system. International journal of geographical information systems, 6(1). Crossland, M. D. 1990. HyrdoLOGIC - a prototype GIS expert system for examining an AI application in a GIS environment. Proceedings of GIS/LIS '90. Davis, J. R., Cuddy, S. M., Laut, P., Goodspeed, M. J., and Whigham, P. A. 1991. Testing of soil moisture prediction model for army land managers. Journal of irrigation and drainage engineering, 117(4), p. 476-489. Durkin, J. 1994. Expert systems: design and development, Macmillan Publishing. Fedra, D. K. 1993. GIS and environmental modeling. In Environmental modeling with GIS, M. Goodchild, B. Parks, and L. Steyaert, eds., Oxford University Press. Fedra, K. 1995. Decision support for natural resources management: models, GIS and expert systems. AI Applications, 9(3), p. 3-19. Fischer, M. M. 1994. From conventional to knowledge-based geographic information systems. Computers, environment and urban systems, 18(4), p. 233-42. Fisher, P. F., Mackaness, W. A., Peacegood, G., and Wilkinson, G. G. 1988. Artificial intelligence and expert systems in geodata processing. Progress in physical geography, 12, p. 371-388. Goodchild, M., Haining, R., and Wise, S. 1992. Integrating GIS and spatial data analysis: problems and possibilities. International journal of geographical information systems, 6(5), p. 407-423. Jackson, P. 1990. Introduction to expert systems, Addison-Wesley Publishing Co. Lam, D. C. L., and Swayne, D. A. 1993. An expert system approach of integrating hydrological database, models and GIS: application of the RAISON System. Proceedings of HydroGIS 93: Application of geographical information systems in hydrology and water resources, Vienna, IAHS Publication 211, p. 23. Lein, J. K. 1992. Modeling environmental impact using an expert-geographic information system. Proceedings of GIS/LIS '92, San Jose, CA, USA, p. 436-44. Leung, Y. 1993. Towards the development of an intelligent spatial decision support system. In Geographic information systems, spatial modelling and policy evaluation, M. Fischer and P. Nijkamp, eds., Springer-Verlag. 15 Lilburne, L. 1996. The integration challenge. Proceedings of SIRC '96, Dunedin. Loh, D. K., and Rykiel, E. J. 1992. Integrated resource management systems: coupling expert systems with database management and GIS. Environmental management, 16(2), p. 167-177. Nyerges, T. L. 1992. Coupling GIS and spatial analytical models. Proceedings of 5th international symposium on spatial data handling, Charleston, South Carolina, p. 534-543. Robinson, V. B., Frank, A. U., and Karimi, H. A. 1987. Expert systems for GIS in resource management. AI Applications, 1(1). Sharma, J., Flewelling, D. M., and Egenhofer, M. J. 1994. A qualitative spatial reasoner. Proceedings of 6th international symposium on spatial data handling, p. 665-681. Smith, T. R., and Yiang, J. 1991. Knowledge-based approaches in GIS. In Geographical information systems: principles and applications, D. J. Maguire, M. F. Goodchild, and D. W. Rhind, eds., Longman Group UK Ltd., Marlow, UK. Zhu, X., and Healey, R. 1992. Towards intelligent spatial decision support: integrating geographical information systems and expert systems. Proceedings of GIS/LIS '92. 
work_hhxlyuwirvgnbmdyk4kpgbo35e	----	A New Distributed Expert System to Ontology Evaluation Procedia Computer Science 37 ( 2014 ) 48 – 55 1877-0509 © 2014 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the Program Chairs of EUSPN-2014 and ICTH 2014. doi: 10.1016/j.procs.2014.08.011 ScienceDirect Available online at www.sciencedirect.com The 5th International Conference on Emerging Ubiquitous Systems and Pervasive Networks (EUSPN-2014) A new distributed expert system to ontology evaluation Djellali Choukria,b aLANCI UQAM, Local W-5353 Case postale 8888, Succ. Centre-Ville Montréal (Québec) H2X 3Y7, Canada bLATECE UQAM, 201, PK 4470, Président Kennedy Montréal (Québec) H2X 3Y7, Canada Abstract This paper addresses the increasingly encountered challenge of ontology evaluation. The best known approaches to ontology evaluation focus on the used criteria and model domain. In the present study, we propose a new approach that uses a reasoning tool as distributed expert system based on symbolic structures of facts and rules. The semantic model found is used to verify the ontology consistency and unexpected relationships between the ontological artefacts. The evaluation is based on formal system defined in description logic SHOIN. The results show that our evaluation approach is independent of the conceptualization of domain model and considers the main features of ontology structure. Good experimental studies demonstrate the multidisciplinary applications of our approach. © 2014 The Authors. Published by Elsevier B.V. Selection and peer-review under responsibility of Elhadi M. Shakshuki. Keywords: Semantic Web, Ontology, logic, owl, consistency, subsumption, instantiation, DIG. 1. Introduction The ontology is proving to be the best solution for communication and knowledge sharing. It provides a deeper level of semantics with significant metadata and ontological commitments to share knowledge between the interaction partners. The underlying aim is to clarify the semantics of Web resources through metadata or annotations. Ontology evaluation remains a significant problem in Semantic Web. The literature contains many definitions of ontology evaluation; many of these contradict one another. djellali.choukri@courrier.uqam.ca © 2014 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the Program Chairs of EUSPN-2014 and ICTH 2014. http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2014.08.011&domain=pdf 49 Djellali Choukri / Procedia Computer Science 37 ( 2014 ) 48 – 55 However, the best known and most cited is (Gómez-Pérez, see [9]), which is also the definition we adopt in our paper « A technical judgment of the content of the ontology with respect to a frame of reference during every phase and between phases of their lifecycle ». Several ontology evaluation approaches have been proposed in the ontologies evaluation literature. The choice of an appropriate approach depends on the evaluation purpose; application and the criteria used to evaluate the ontology. These criteria focus on the characteristics of ontology and they are independent of the application domain. It is difficult to choose the best user who judges the ontology quality and the proposed criteria for the evaluation. Moreover, the choice of a suitable approach depends on the model domain. In order to overcome these drawbacks, we propose a new evaluation approach which is based on the property of soundness, completeness and decidability. The paper is organized as follows: In Section 2, we present the current state of the art in ontology evaluation, our research questions and the problematic of ontology evaluation. The conceptual architecture of our approach is given in Section 3. Before we conclude, we give in Section 4 a short evaluation with benchmarking model for our conceptual model. Then, a conclusion (Section 5) and future work (Section 6) end the paper. 2. State of the art, Problem and Research Questions There are several approaches that have been presented in the literature for ontology evaluation. The best known approaches fall into one of the following categories: - Gold Standard: following a benchmark comparison using the Gold Standard reference ontology. - Evaluation based on data: using a data-driven method to evaluate the degree of fit between the ontology and a Data Set describing the problem domain to which the ontology refers. Therefore, this approach requires traceability mechanisms to describe the concrete relations between the ontology entities and the Data Set. - Evaluation based on criterion: a qualitative approach following the ontology characteristics regardless of the application domain [10]. The most known criteria are grouped as shown in Table (1): Table 1. Summary of work on the criterion of ontology evaluation. Author Year Total criteria Objective Gruber 95 5 Design criteria Uschold &Grüninger 96 3 Design criteria Noy & Hafner 97 28 Design criteria Hovy 97 36 Compare linguistic ontologies Uschold 98 10 Identify the roles of ontology in applications - Evaluation based on tasks: accomplishing a given task using ontology and evaluating the results [1]. The different approaches proposed in the ontologies evaluation literature consult experts who use their opinions and experiences in the evaluation process. However, they are subject to extreme evaluations. Furthermore, the choice of a suitable approach depends on the used criteria and the model domain. In this paper, we propose a new approach that uses a reasoning tool as distributed expert system based on symbolic structures of facts and rules. The evaluation is based on formal system defined in description logic SHOIN. 50 Djellali Choukri / Procedia Computer Science 37 ( 2014 ) 48 – 55 3. The architecture of our evaluation system Consistency, subsumption and instantiation can accentuate the main features of the conceptual hierarchy as well as the ontological inference, i.e., decidable decisions about satisfiability, ontology hierarchy and completeness. Description logic is perfectly suited to this situation. It has a formal semantics based on logic and equipped by decidable decisions. In addition, the description logic differs from its predecessors, such as conceptual graphs, existential graphs, semantic networks and frames where it is governed by a formal semantics based on logic [6],[7]. . 4. Experimentation 4.1. Ontology The CRISP-DM-OWL1 ontology used in this project is integrated into a hybrid system DM, describing the artifacts and the basic rules to improve the intelligence level of the system. The ontology acts as a source of additional knowledge [4]. Figure (2) describes the CRISP-DM-OWL ontology with the artifacts involved in each section. The hierarchy of concepts is illustrated using the GraphViz tool [5]. In order to simplify the layout hierarchies of concepts and to allow a better understanding, we show only a few specializations. 1 http://www.elmanahel.ca/ontology/crisp-dm-owl.owl 2http://racer.sts.tuhh.de/ Fig. 1. The Descriptive Inference System. As shown in figure (1), the descriptive inference system used to verify the consistency, soundness and completeness of an ontology is based on the inference engine RacerPro2 (Renamed ABox and Concept Expression Reasoner) [8]. The ontology can be regarded as a T-Box/A-Box representation with a hierarchy of roles describing the domain in terms of classes (concepts) and properties (roles). Thus, we can consider the evaluation system as distributed expert system based on structures of facts and rules. It provides a symbolic reasoning used to verify the ontology completeness and soundness. The DIG protocol (XML standard) [3] is used to connect the applications to a semantic model based on ontology. The allocation of terminological knowledge base allows users to query the conceptual model. In this way, we can query the terminological knowledge base to ensure that all facts deduced from the ontology can be inferred from the Data Set. In other words, our main objective is to describe how the ontology realizes requirements and needs described in the requirements model, i.e., the specifications and ontological commitments. The inference system is divided into three main components: inference engine, terminology T-Box and description of the world A-Box. In each of these components, knowledges are declared in the form of rules and facts. The rules are related to the terminology reasoning operations (subsumption and instantiation). 51 Djellali Choukri / Procedia Computer Science 37 ( 2014 ) 48 – 55 Fig. 2. The different techniques of modeling. 52 Djellali Choukri / Procedia Computer Science 37 ( 2014 ) 48 – 55 4.2. Inference engine In order to ensure the ontology consistency, RacerPro provides several inference rules that deduce implicit knowledge from the explicitly represented knowledge. These inference rules are decidable and with low complexity. The inference engine transforms the ontology artefacts as descriptive rules to support formal reasoning. The rules are categorized and arranged according to the level of complexity. It should be noted that the ontology ),,,,,,( ARIHRHCO IRCC is considered in two parts: - The extensional part CRC , ou T-Box: representation and manipulation of concepts and roles in terminological level. - The intensional part IRI , ou A-Box: representation and manipulation of individuals in a factual level. In this symbolic definition, ontology is represented by a hierarchy of symbols CH ( RH ) connected by the subsumption relation . C : concepts are arranged in a schema hierarchy .CH CR : the set of relationships between concepts which are also arranged in a hierarchy .RH I : the instances that are interconnected by all instances of properties .IR A : the set of axioms used to express other relationships between concepts and to constrain their interpretations. RacerPro adopts Fixed Point Semantics in order to avoid the definitions of terminology, which contain terminological cycles (rewriting termination). We choose RacerPro reasoning as a tool for our approach because it offers more features and more graphics editors such as OilEd and RacerPorter (RACER Interactive Client Environment). It also offers a manipulation of symbolic reasoning for application developers. In addition, RacerPro offers several optimization techniques with proof tools: - RacerPro includes several optimization techniques to ensure good performance of search, in particular, the dependency-directed backtracking and DPLL-style semantic branching. - The verification of conceptual consistency. - The search for inconsistent concepts in terminology T-Box. - The determination of parents and children of a concept. - The verification of the consistency of the description of the world A-Box. - Testing the description of the world A-Box and the terminology T-Box. - Find the subsumed/subsuming ontological artefact in the terminology T-Box and description of the world A-Box. - Calculate the direct types of individuals [2], [7]. 4.3. Terminology T-Box This component contains terminological axioms that describe the relationship between concepts and roles. Thus, RacerPro define terminological axioms which have the following form: )()( 2121 rrBArrBA IRrrOBA 21 ,;, . The axioms of the first type are called inclusions, while axioms of the second type are called equalities. The semantic description of the terminological description T-Box is interpreted as a subset of a domain of interpretation. The subsumption is the terminological inference based on concept expressions in the terminological description T-Box [7]. 53 Djellali Choukri / Procedia Computer Science 37 ( 2014 ) 48 – 55 Several rules can be present in the terminology T-Box: - R1 : ).()( BAABBA - R2 : ).()( BABA - R3 : ).)(()( IBABA - R4 : ).)()(()( II BABABA - R5 : ).)(()( IBABA - R6 : ).()( AA I - ….. OBA , . For reasons of readability, we cannot present all the rules. Figure (3) illustrates the localhost connection with RacerPro server, which shows the consistency verification of terminological axioms T-Box. Fig. 3. The terminological reasoning T-Box. 4.4. Description of the world A-Box The second component of the terminological knowledge base is a description of the world A-Box. The ontological entities in this world can take the following form: OAooroA (),,(),( 211 , ),, 21 IRrIoo . The first type is called concept assertion and the second type is called the role assertion. In the description of the world A-Box, RacerPro declares all instances I that are interconnected with all relations IR . The inference engine uses the terminological instantiation reasoning to determine if an object o is an instance of a concept A (also known as the validity of the assertion )( oA ) [7]. Several instantiation rules may arise in the description of the world A-Box: - R1 : ).())(( oAABoB - R2 : ).()))((())(())(())()((( oCCBAandCBCAoBoA - R3 : ).()))(((),( 1221 oAoAralloor - R4 : ).())(()((),( 1221 oAoBrallBAoor - R5 : ).())()(),1(( BAoBoAni ii - ….., etc. OCBA ,, , ., 21 Ioo 54 Djellali Choukri / Procedia Computer Science 37 ( 2014 ) 48 – 55 Figure (4) shows the consistency treatment using the terminological instantiation reasoning in the description of the world A-Box. Fig. 4. The terminological instantiation reasoning of the world description A-Box. The inference rules mentioned above are very important because they are used to verify the ontology consistency, unexpected relationships between the ontological artefacts, terminological knowledge base unsatisfiability, determine subsumers and subsumed, avoid definitions of terminology that contain cycles and fixed points, etc. 5. Conclusion In this paper, we have introduced a new approach providing standardized evaluation for ontologies. Our approach is independent of the conceptualization of the domain model and considers the main features of the ontology structure and its population (concepts, instances, axiom, relationship, etc.). It has a formal semantics based on description logic and equipped by decidable decisions. In addition, our system offers several advantages: - Soundness, completeness and decidability. - Ability to transform a descriptive representation to a first order predicate logic representation. - Efficient for reasoning by classification. - A well-defined semantics. - The simpleness of modeling ontologies. - Duality expressiveness versus complexity. - The representativeness is described by two levels, i.e., the terminology description T-Box and world description A-Box. - Classification and instantiation are operations that are at the base of the terminological reasoning. 6. Future work The purpose of our next work is to propose a visualisation tool for representing the ontological knowledge. With the presentation of the ontology hierarchy we can basically understand the inheritance relationships between ontological artefacts. The visualization tool should at least provide an overview of the hierarchy or partial views, allowing the user to focus on a part of the ontology. These elements should be displayed so that you can discern the relevant information. 55 Djellali Choukri / Procedia Computer Science 37 ( 2014 ) 48 – 55 References 1. Brank, Janez, Marko Grobelnik et Dunja Mladenić. 2005. A survey of ontology evaluation techniques. 2. Haarslev, Volker, et Ralf Möller. Racer: A Core Inference Engine for the Semantic Web: EON, 2003. 3. S Bechhofer, R Moller, and P Crowther. The DIG description logic interface: DIG/1.1, 2003. 4. Shen Yanfen. A formal ontology for Data Mining: principles, design and evolution Thesis UQTR, 2007. 5. Gansner, E.R. Drawing graphs with Graphviz, Technical report, AT&T Bell Laboratories, Murray, 2009. 6. The description logic handbook: theory, implementation, and applications. Edited by. Franz Baader. Deborah L. McGuinness. Daniele Nardi. Peter F, 2003. 7. Napoli, Amedeo. An introduction to description logics. 1997. 8. Haarslev V, Hidde K, Mölr R, Wessel M. The RacerPro knowledge representation and reasoning system. Semantic Web 2011; 1-7. 9. Gómez-Pérez A, Fernández-López M, Chorcho O: Ontological Engineering, Springer-Verlag, London, UK, 2004. 10. Lozano-Tello, Adolfo, et Asunción Gómez-Pérez. «Ontometric: A method to choose the appropriate ontology». Journal of Database Management, 2004; 15: 1-18. 
work_hiqnrmu3offw7blky2jcvqsw5i	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy A novel anonymization algorithm: Privacy protection and knowledge preservation Weijia Yang a,*, Sanzheng Qiao b a Department of Computer Science, Shanghai Jiao Tong University, Shanghai 200030, China b Department of Computing and Software, McMaster University, Hamilton, Ont., Canada L8S 4K1 a r t i c l e i n f o Keywords: Data mining Privacy protection Data anonymization Knowledge preservation a b s t r a c t In data mining and knowledge discovery, there are two conflicting goals: privacy protection and knowl- edge preservation. On the one hand, we anonymize data to protect privacy; on the other hand, we allow miners to discover useful knowledge from anonymized data. In this paper, we present an anonymization method which provides both privacy protection and knowledge preservation. Unlike most anonymiza- tion methods, where data are generalized or permuted, our method anonymizes data by randomly break- ing links among attribute values in records. By data randomization, our method maintains statistical relations among data to preserve knowledge, whereas in most anonymization methods, knowledge is lost. Thus the data anonymized by our method maintains useful knowledge for statistical study. Further- more, we propose an enhanced algorithm for extra privacy protection to tackle the situation where the user’s prior knowledge of original data may cause privacy leakage. The privacy levels and the accuracy of knowledge preservation of our method, along with their relations to the parameters in the method are analyzed. Experiment results demonstrate that our method is effective on both privacy protection and knowledge preservation comparing with existing methods. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction Privacy protection is an important issue in data processing. Sup- pose a data set shown in Fig. 1 is used in a medical research article. Before the article is published, the data must be anonymized to protect privacy. For example, if Michael knows that Hannah is 32 working as a clerk in UK, then he can infer that Hannah has diabe- tes, even the data set is published without names. A group of attri- butes, such as fAge; Job; Countryg in the above example, which can be used to infer patients is commonly called quasi-identifier (Q-I). The attribute like disease is called sensitive attribute. In medical research, the associations such as ‘‘clerk ! hypertension” and ‘‘[50–60] ! diabetes” are often studied. This kind of associations compose non-sensitive knowledge, if individual names cannot be inferred from the associations. Thus, it is important that a data set is protected in such a way that the miner can hardly infer the sensitive data of an individual and, at the same time, discover non-sensitive knowledge. In recent years, many methods have been proposed to protect the data privacy in data mining (Agrawal & Aggarwal, 2001; Agra- wal & Srikant, 2000; Aggarwal & Yu, 2008; Clifton, Kantarcioglu, Vaidya, Lin, & Zhu, 2002; Du & Zhan, 2003; Lindell & Pinkas, 2000). Among these methods, data anonymization (Ciriani, Vim- ercati, Foresti, & Samarati, 2008; Machanavajjhala, Gehrke, Kifer, & Venkitasubramaniam, 2006; Sweeney, 2002) provides an effec- tive yet simple way of preventing the user from learning sensitive data. Referring to the k-anonymity method (Sweeney, 2002), any individual is indistinguishable from at least k � 1 other ones in the anonymized data set. Machanavajjhala et al. (2006), however, has pointed out that the user may guess the sensitive values with high confidence when the sensitive data is lack of diversity, and introduced the l-diversity method. Both k-anonymization and l- diversity have drawbacks. Most of the related works (Kifer & Gehrke, 2006; LeFevre, DeWitt, & Ramakrishnan, 2005, 2006; Li, Li, & Venkatasubramanian, 2007; Meyerson & Williams, 2004) gen- eralize the data values. Some attributes in the quasi-identifier may even be totally suppressed. Both the attribute generalization and suppression lead to loss of data details (Aggarwal, 2005). As a re- sult, the useful data knowledge such as the non-sensitive knowl- edge and data distributions can only be obtained in general forms, which may be of little value for the miner. Recently, the permutation method has been proposed by Xiao and Tao (2006) and Koudas et al. (2007). Instead of generalizing the values of attributes, it randomly permutes the sensitive values in record groups. In this way, the user can hardly match the sensi- tive values with the right individuals. However, it is also difficult to discover the data knowledge from the permuted data. In this paper, we propose a novel anonymization method. In- stead of generalizing or permuting attribute values as in most anonymization methods, our method randomly breaks the links 0957-4174/$ - see front matter � 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2009.05.097 * Corresponding author. Tel.: +86 21 5474 0000. E-mail addresses: weijia.yang@yahoo.com.cn (W. Yang), qiao@cas.mcmaster.ca (S. Qiao). Expert Systems with Applications 37 (2010) 756–766 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy between the value combinations of the quasi-identifier and the values of the sensitive data, while maintaining statistical relations between quasi-identifier data and sensitive data, thus preserving the non-sensitive knowledge for statistical study. It protects pri- vacy and, at the same time, allows the miner to discover non-sen- sitive knowledge with high confidence. Furthermore, an enhanced method is proposed for preventing the privacy leakage when the user has some prior knowledge of the original data associations. The paper is organized as follows: in Section 2, we introduce our privacy definition after presenting the preliminary knowledge of data anonymization. Our random anonymization (RA) algorithm is presented in Section 3. Then we analyze its level of privacy pro- tection in Section 4. The discussion of the accuracy of the knowl- edge preservation is given in Section 5. Afterwards, we present in Section 6 an enhanced RA algorithm to limit the extra privacy leak- age. Finally, we demonstrate our experiments in Section 7. 2. Data privacy Since our method provides a way of anonymizing data, it is necessary to first introduce some preliminary definitions of data anonymization in Section 2.1. The privacy is disclosed when the user is able to correctly link the individuals with their sensitive val- ues. In Section 2.2, we introduce our new definition of privacy by data association. 2.1. Preliminary In this section, we adopt several definitions commonly used in the k-anonymization and l-diversity methods from Sweeney (2002) and Machanavajjhala et al. (2006). First, we present the for- mal definition of quasi-identifier. Definition 1 (Quasi-Identifier (Q-I). Given a data set DðA1; A2; . . . ; AmÞ and an external table DE. For all records ri 2 D, if the value combination riðAj; . . . ; AkÞ; j; k 6 m;fAj; . . . ; Akg contains no identifiers, can be uniquely located in DE , we call the set of attributes fAj; . . . ; Akg a quasi-identifier. For example, in the data set in Fig. 1, the external table DE would consist of Name, Age, Job, and Country and a quasi-identifier would be fAge; Job; Countryg. The k-anonymization method uses the generalization tech- nique, formally defined by: Definition 2 (Generalization). Suppose that a domain M consists of disjoint partitions fPig; i ¼ 1 . . . n, and [Pi ¼ M. On a given value combination v, we call the generalization process as returning the only partition Pi containing v. By generalizing the quasi-identifier, each individual in a k- anonymous table is identical to at least k � 1 other ones with re- spect to the quasi-identifier: Definition 3 (k-Anonymous). Given a data set DðA1; A2; . . . ; AmÞ and its quasi-identifier QI. If for any subset C # QI and for any record ri 2 D, there exist at least k � 1 other records sharing the same values with ri on the attribute set C, then data set D is k- anonymous. When the data set is k-anonymous, we can group together the records with the same value combinations of the quasi-identifier: Definition 4 (Q-I Group). Given a data set D and its quasi- identifier QI. We define the Q-I group as the set of all the records with the same values on QI. Fig. 2 is a 2-anonymous data set generalized from Fig. 1, where the data set is partitioned into three groups. As shown in Fig. 2, the data of Age are grouped into wider intervals and Jobs are clustered. We denote by symbol ‘‘�” a wildcard in attributes. As a result, each record is indistinguishable from at least one other record by the quasi-identifier. However, k-anonymity has drawbacks. When an individual be- longs to the last group in Fig. 2, the user can infer that the individ- ual has hypertension with more than 66% confidence, since two out of three in the group have hypertension. Moreover, since quite a few data values have been generalized in the anonymous data set, non-sensitive associations such as ‘‘[50–60] ! diabetes” can- not be obtained. While the k-anonymization method only focuses on hiding the Q-I information of the individuals, the l-diversity model (Mach- anavajjhala et al., 2006) pays attention to the relations between the Q-I and sensitive data: Definition 5 (l-diversity). A Q-I group is said to satisfy l-diversity if it contains at least l ‘‘well-represented” values for the sensitive attribute. A data set is said to have l-diversity if every Q-I group of it satisfies l-diversity. In particular, for each Q-I group, if the entropy of the sensitive attribute is greater than ln l, then the data set is said to satisfy Fig. 2. A 2-anonymous data set by generalizing Fig. 1. Fig. 1. A sample of original confidential data. W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 757 Author's personal copy ‘‘entropy l-diversity”. When a user tries to guess the sensitivity of an individual from an anonymized data, the diversity quantifies the average number of the possible sensitive values. Fig. 3, also de- rived from Fig. 1, shows an anonymized data set satisfying more than (entropy) 2-diversity. From this data set, the user cannot infer the sensitive value of any individual with more than 50% confi- dence. However, Q-I values have been generalized further, thus the non-sensitive knowledge is hard to obtain. Moreover, the diversity of the sensitivity in each Q-I group is limited by the dis- tribution of sensitive data in the whole data set. To keep the data details, the permutation method permutes the sensitive values within each Q-I group. As shown in Fig. 4, the data set in Fig. 1 are permuted so that each group satisfies at least (en- tropy) 2-diversity. The permuted data set is actually divided into two views: the view of the quasi-identifier and the view of the sen- sitivity. During data mining, the user will join these two views for each record group. Although the data details are preserved, the non-sensitive associations can hardly be discovered from Fig. 4, be- cause their confidences decrease dramatically. 2.2. Measuring data privacy Most k-anonymization methods (Aggarwal, 2005; Kifer & Gehrke, 2006; LeFevre et al., 2005, LeFevre, DeWitt, & Ramakrish- nan, 2006; Meyerson & Williams, 2004; Sweeney, 2002) emphasize the protection of the quasi-identification information. The minimal size of the Q-I groups is used to quantify the privacy level of the anonymized data. The larger the minimal size is, the more difficult it is to identify an individual from a value combination of Q-I. Sim- ilarly, it is difficult to identify an individual from a value of sensi- tive data alone. However, if the user can establish a strong association between a value combination of Q-I and a value of sen- sitive data, then privacy can be compromised. Thus we propose that privacy is measured by the association between the value combinations of Q-I and the values of sensitive data. A probabilistic measurement of privacy or anonymity is given in Section 4. Also, the anti-monotone property of k-anonymity shows that if the qua- si-identifier is k-anonymous, any subset of it is also k-anonymous. The value combinations of the subset do not disclose more privacy than those of the quasi-identifier. Thus, unlike the k-anonymiza- tion methods, we do not group the value combinations of Q-I. How do we prevent the user from getting these important associ- ations? We will present our anonymization algorithm in the next section. 3. Anonymization algorithm To break the associations between the value combinations of Q- I and the values of sensitive data, it is not necessary to anonymize all the attributes in Q-I. Our main idea is to randomly replace part of the Q-I data for each record by using the distributions of the original values of an attribute in the Q-I. In this way, no new information is added to the anonymized data, the associations between Q-I and sensitive data are broken, and the original data distribution is preserved. Consequently, the associations in individ- ual records are broken, whereas the statistics of associations in the whole data set is preserved. We present the anonymization process in Algorithm 1. Algorithm 1. The RA Algorithm 1: Input : the original data set D, the Q-I attributes Q, and the probability distribution fp1; . . . ; pmg, where m ¼ jQj, the number of attributes in Q-I. 2: Output: the original data set D is overwritten by an anonymized one 3: begin 4: n :¼ jDj; 5: Dist :¼;; 6: for i: = 1 to m do 7: begin 8: Disti:= the distribution of the values of Q i; 9: end 10: for j: = 1 to n do 11: begin 12: Randomly select an attribute Q k in Q-I of the jth record with probability pk; 13: Randomly generate a new value for Q k based on Distk ; 14: Replace the value of Q k with the new value; 15: end 16: end For example, in the data set in Fig. 1, the Q-I attributes Q ¼fQ 1; Q 2; Q 3g¼fAge; Job; Countryg. The values of Q i and their corresponding distributions are:Fig. 4. An anonymous data set by permuting Fig. 1. Fig. 3. A 2-diversity data set derived from Fig. 1. 758 W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 Author's personal copy Values of Q 1 [20–30] [30–40] [40–50] [50–60] [60–70] Dist1 1/10 4/10 1/10 3/10 1/10 Values of Q 2 Doctor Clerk Trader Engineer Banker Dist2 1/10 4/10 2/10 1/10 2/10 Values of Q 3 USA Germany UK India Dist3 6/10 1/10 2/10 1/10 Setting p1 ¼ p2 ¼ p3 ¼ 1=3, for every record in Fig. 1, we equally likely select an attribute Q k from the three attributes in the Q-I. Then we randomly generate a value for the selected Q k based on Distk and replace the original value of Q k with the new value. Fig. 5 shows an anonymized data set produced by Algorithm 1. How do we choose pi in this algorithm? How does this algo- rithm protect privacy while preserving non-sensitive knowledge? We will address these issues in the following two sections. 4. Privacy analysis Since our RA algorithm, unlike the k-anonymity and l-diversity methods, anonymizes a data set by randomly breaking the associ- ations between its Q-I and sensitive data, we first define a statisti- cal measurement of anonymity in Section 4.1. Then we use this definition to analyze the anonymity of our method. Moreover, since we randomize the values of the selected attributes, for com- parison, we also use the measurement in Evfimievski, Gehrke, & Srikant (2003) to analyze the ‘‘privacy breaches” in our method in Section 4.2.1. In addition, in Section 4.2.2, we further discuss the privacy leakage when the user has some prior knowledge of the data distribution. 4.1. Probabilistic anonymity For a good anonymization, it should be very unlikely that the user can infer the original associations from the corresponding associations in an anonymized data set. Thus we propose the fol- lowing definition of anonymity. Definition 6 (Probabilistic anonymity). Suppose that a data set D is anonymized to D0. Let r be a record in D and r0 2 D0 be its anonymized form. Denote rðQIÞ as the value combination of the quasi-identifier in r. The probabilistic anonymity of data set D0 is defined by 1=PðrðQIÞjr0ðQIÞÞ, where PðrðQIÞjr0ðQIÞÞ is the probability that rðQIÞ (for all r 2 D) may be inferred given r0ðQIÞ. The probabilistic anonymity gives a measurement of how unli- kely the user can infer original associations. The greater the prob- abilistic anonymity, the less probable the user can guess the original data. Now that we have a measurement of anonymity, we will show how to determine pi in Algorithm 1 to maximize the anonymity of the data set produced by the algorithm. Proposition 7. Let Q i; i ¼ 1; . . . ; m be the ith Q-I attribute (category attribute) in a data set D and EntropyðQ iÞ be the value of the entropy of Q i. Then the probabilistic anonymity of D 0, the anonymized form of D, reaches the maximal value when each pi in Algorithm 1 is directly proportional to the value of eEntropyðQ iÞ. Proof 1. Let cij be the jth category value in Q i and FreqðcijÞ be the frequency of cij in Q i . We calculate PaðD 0Þ, the probabilistic ano- nymity of D0, by: ln PaðD0Þ ¼� Xm i¼1 XJi j¼1 ½pi FreqðcijÞ lnðpi FreqðcijÞÞ� ¼ Xm i¼1 XJi j¼1 pi FreqðcijÞ ln 1 pi � � þ Xm i¼1 XJi j¼1 pi FreqðcijÞ ln 1 FreqðcijÞ � � ; where Ji is the number of the category values in Q i . SincePJi j¼1 FreqðcijÞ¼ 1, we can also derive: ln PaðD0Þ ¼ Xm i¼1 pið� ln pi þ EntropyðQ iÞÞ ð1Þ Next, we use the method of Lagrange Multipliers (Chen, Jin, Zhu, & Ouyang, 1983) to find the maximal value of PaðD0Þ. We incorporate the constraint Pm i¼1 pi ¼ 1 into (1): Fðp1; . . . ; pm; lÞ¼ Xm i¼1 ½pið� ln pi þ EntropyðQ iÞÞ�þ l Xm i¼1 pi � 1 ! where l is an unknown scalar. Setting rp1;...;pm Fðp1; . . . ; pm; lÞ¼ 0, we have: @F @p1 ¼ EntropyðQ 1Þ� ln p1 � 1 þ l ¼ 0 . . . @F @pm ¼ EntropyðQ mÞ� ln pm � 1 þ l ¼ 0: 8>< >: Thus, for all i; j 2 ½1; m�, we have ln pi pj ¼ EntropyðQ iÞ� EntropyðQ jÞ; implying that pi pj ¼ eEntropyðQ iÞ eEntropyðQ jÞ : ð2Þ From (2), when pi ¼ e EntropyðQ iÞPm j¼1 e EntropyðQ jÞ ; PaðD0Þ reaches its maximal value. hFig. 5. An anonymized data set produced by Algorithm 1 from Fig. 1. W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 759 Author's personal copy The above proposition also shows a general method for calcu- lating the probabilistic anonymity. When pi ¼ 1m ; i ¼ 1; . . . ; m, we can have a quick estimation of the scaled PaðD0Þ by taking the geo- metric mean of the diversities of all Q-I attributes: ln PaðD0Þ¼ ln mþ Xm i¼1 1 m EntropyðQ iÞ � � ¼ ln m Ym i¼1 Diversityi !1 m 0 @ 1 A; ð3Þ where Diversityi ¼ eEntropyðQ iÞ is the entropy diversity of Q i (Mach- anavajjhala et al., 2006). For an arbitrary record, the probability that the user may guess its original Q-I values, given an anonymized data set, is 1=PaðD0Þ. This is also the user’s confidence in associating a sensitive value with an individual. From (3), PaðD0Þ is usually greater than the geometric mean of the diversities of all Q-I attri- butes. Similarly, PaðD0Þ is usually greater than the diversity of the sensitive attribute. Even when the user is sure that an individual is in the data set, the user’s maximal confidence in inferring the cor- responding sensitivity is 1=Diversitys, where Diversitys is the diver- sity of the sensitive attribute. In comparison, in the l-diversity method, the user’s confidence in guessing the data privacy is 1=l, which is usually greater than 1=Diversitys. Therefore, by disassociat- ing the relationships between the Q-I and sensitive values, the RA method provides a better protection of the privacy than the l-diver- sity method does. 4.2. Privacy breaches The measurement of probabilistic anonymity evaluates the gen- eral level of privacy protection. Although the privacy is protected on average, certain privacy breaches (Evfimievski et al., 2003) may still take place. 4.2.1. Common breaches Suppose q 2 D is one of the original value combinations of the quasi-identifier. And q0 2 D0 is its anonymized form. The privacy breaches can be represented by: there exists q 2 D and q0 2 D0 : Pðq � tjq0 � t0Þ� Pðo � tjq0 � t0Þ; for all o – q; where t 2 D is the original record and t0 2 D0 is its anonymized form. By the above definition, if the user is able to get the original q with high confidence once q0 appears in the anonymized data, then there are privacy breaches. Privacy breaches are quantified by the c- amplification (Evfimievski et al., 2003): Definition 8 (c-Amplification). Suppose a data set D is anonymized to D0. For some value v0 2 D0, if for all v 1; v 2 2 D; Pðv 1!v0Þ Pðv 2!v0Þ 6 c, then the anonymization is at most c-amplifying for v0. The anonymiza- tion is at most c-amplifying if it is at most c-amplifying for all v0 2 D0. By the definition, once the anonymized value v0 is revealed, when all candidate values are equally possible, i.e., they share a common probability, then the probability that the user may obtain its original value is at most c times the common probability. Thus, the closer the value of c approaches 1, the more difficult it is for the user to infer the original data. Note that c P 1. In the RA algorithm, each record is anonymized without refer- ring to its original Q-I data. It seems that we have c ¼ 1. But the original data distribution is used to generate the new values for the selected attributes. How does this affect c? In the following, we analyze the privacy breaches in our method. Suppose a data set D is anonymized to D0 by Algorithm 1. Let q0 ¼ fq01; q02; . . . ; q0mg be a value combination of the quasi-identifier in D0 where q0i is the value of the ith Q-I attribute Q i. We calculate the probability Pðq ! q0Þ for each possible value combination q 2 D: P q ¼f�; q02; q03; . . . ; q0mg! q0 � � ¼ p1 P Q 1 ¼ q01 � � P q ¼fq01;�; q03; . . . ; q0mg! q0 � � ¼ p2 P Q 2 ¼ q02 � � . . . P q ¼fq01; . . . ; q 0 m�1;�g! q 0 � � ¼ pm P Q m ¼ q0m � � 8>>>< >>>: If there exists a k such that pk PðQ k ¼ q0kÞ P pi PðQ i ¼ q 0 iÞ for all i ¼ 1; . . . ; m, then there is a privacy breach by Definition 8. Although the user can infer with high confidence that q0 is generated by anon- ymizing the kth attribute, he is unable to find the original value of the kth attribute, since each value is anonymized independently from the original one. Thus the user can hardly leverage the breaches in our method to identify the individuals or their sensitive data. However, things will be different when the user has some prior knowledge of the distributions of the Q-I data, especially the support of the value combinations of the quasi-identifier. This is probable when data set D contains all the individuals in the public data. In the next subsection, we analyze the data privacy that our method preserves in this situation. 4.2.2. Privacy breaches by prior knowledge Let t0 2 D0 be an anonymized record and q0 # t0 be its Q-I value combination, then t0 may be anonymized from one of the following two types of records in D: � The records which contain q0 both before and after the anonymization. � The records whose Q-I values have intersection with q0 of size m � 1 and are anonymized to q0. Definition 9 (Support). The support of a value combination q in D, denoted by suppDðqÞ, is the percentage of the records in D which contain all the attribute values in q. The support of a record t, denoted by suppDðtÞ, is the support of its value combination. A support gives the frequency of the occurrence of a value com- bination in D. Assuming pi ¼ 1=m; i ¼ 1; . . . ; m, we have the expec- tation of the support for the first type of records in D0: EðsuppD0ðft 0jq0 � t&q0 � t0gÞÞ ¼ suppDðq 0ÞPðq0is anonymized to itselfÞ ¼ suppDðq 0Þ 1 m Xm i¼1 P Q i ¼ q0i � � : ð4Þ For the second type, jq0 \ qj ¼ m � 1, assuming pi ¼ 1=m; i ¼ 1; . . . ; m, the expectation of the support is: EðsuppD0ðft 0jq0 � t0&q � tgÞÞ ¼ X q ðsuppDðqÞPðq is anonymized to q 0ÞÞ ¼ X q suppDðqÞ 1 m P Q wðq0;qÞ ¼ q0wðq0;qÞ � � � ; ð5Þ where wðq0; qÞ is the index of the Q-I attribute in which q0 and q differ. This time, by the prior knowledge, a breach for the user to guess the original value of the anonymized attribute may occur. Combining (4) and (5), we can derive the c-amplification for the breach: cq0 ¼ maxq suppDðq0Þ � Pm i¼1 PðQ i ¼ q 0 iÞ suppDðqÞ � PðQ wðq0;qÞ ¼ q0wðq0;qÞÞ ! If cq0 < 1, we set cq0 ¼ 1=cq0 . The value of cq0 varies with the distribu- tions of different data sets. It is desirable to have a constant mea- 760 W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 Author's personal copy surement. Thus, we derive an upper bound for the maximal proba- bility that the user may infer the original q from q0. Let maxSupp be the larger of max jq\q0j¼m�1 ðsuppDðqÞP Q wðq0;qÞ ¼ q 0 wðq0;qÞ � and suppDðq 0Þ Xm i¼1 P Q i ¼ q0i � � then the maximal probability is maxSuppP q suppDðqÞPðQ wðq0;qÞ ¼ q0wðq0;qÞÞ � þ suppDðq0Þ Pm i¼1 P Q i ¼ q0i � � ð6Þ The following proposition gives an upper bound for the maximal probability. Proposition 10. Suppose pi ¼ 1=m, for i ¼ 1; . . . ; m. Then the max- imal probability that the user may guess the original data is bounded above by max suppDðq0Þ miniðsuppDðq0 q0iÞÞ ; PðQ wðq0;qÞ ¼ q0wðq0;qÞÞPm i¼1 P Q i ¼ q0i � � !: Proof 2. We consider two cases based on the maxSupp in (6). In the first case when maxSupp ¼ suppDðq0Þ Pm i¼1 P Q i ¼ q0i � � , the probability that the user may guess the data is Pðq 2 Djq0 2 D0Þ ¼ suppDðq0Þ Pm i¼1P Q i ¼ q 0 i � � Pm i¼1 P Q i ¼ q0i � � suppDðq0Þþ P wðq0;qÞ¼i suppDðqÞ � � ¼ suppDðq0Þ Pm i¼1 PðQ i ¼ q 0 iÞPm i¼1 P Q i ¼ q0i � � � suppD q0 q0i � �� � 6 suppDðq0Þ mini suppDðq0 q0iÞ � � ; ð7Þ where q0 q0i is the value combination obtained by deleting q 0 i from q. If q ¼fqig, then suppDðq0 q0iÞ¼ 1. In the second case, maxSupp ¼ suppDðqÞPðQ wðq0;qÞ ¼ q0wðq0;qÞÞ for some q, we have Pðq 2 Djq0 2 D0Þ ¼ suppDðqÞPðQ wðq0;qÞ ¼ q0wðq0;qÞÞPm i¼1 P Q i ¼ q0i � � suppD q 0 q0i � �� � 6 PðQ wðq0;qÞ ¼ q0wðq0;qÞÞPm i¼1 P Q i ¼ q0i � � : ð8Þ Combining (7) and (8) completes the proof. h In summary, without any prior knowledge, the user is unlikely to associate sensitive data with individuals from anonymized data set. Even with the knowledge of the distribution of the quasi-iden- tifier, the user’s ability is limited by the upper bound given in Prop- osition 10. 5. Quasi-sensitive knowledge preservation Now that we have discussed the security of our anonymization method, in this section, we turn to the issue of knowledge preser- vation, which is important in data mining. Many k-anonymization methods (LeFevre et al., 2005, 2006; Meyerson & Williams, 2004) try to minimize the loss of data details while maximizing the level of privacy protection. Some recent algorithms (Fung, Wang, & Yu, 2007 ;Kifer & Gehrke, 2006) also try to retain the data utility for the purpose of data analysis and data mining. But few consider the issue of non-sensitive knowledge preservation. In Section 5.1, we first explain why we preserve the non-sensitive knowledge and also define its general form in the term ‘‘quasi-sensitive knowledge”. Then in Section 5.2, we discuss the accuracy of knowl- edge recovery. 5.1. Quasi-sensitive knowledge In most applications of data analysis and data mining, we are interested in finding out the data relationships among attributes, especially the associations between the Q-I and sensitive values. We call these associations the quasi-sensitive associations. Definition 11 (Quasi-sensitive associations). Let Q be the set of Q-I attributes in data set D and S be the sensitive attributes. Suppose attribute sets bQ # Q and bS # S. For any values q̂ 2 bQ and ŝ 2 bS, we call the associations q̂ ! ŝ the quasi-sensitive associations in D. All the quasi-sensitive associations make up the quasi-sensitive knowledge in D. For example, in the data set in Fig. 1, the two of its quasi-sensi- tive associations: � Clerk ! Hypertension (confidence: 75%) � [30–40], Clerk ! Hypertension (confidence: 67%) which are helpful in analyzing the causes of hypertension. We denote by jq̂j the number of values in q̂ and jQj the number of attributes in Q, and call jq̂j the length of the association q̂ ! ŝ. We can see that normally the closer jq̂j approaches jQj, the more sensitive the related association is. Especially when jq̂j ¼ jQj, the related association is most sensitive because the antecedent part of the association rule can infer a specific individual. To preserve the less sensitive associations, we try to make the discovery of the ‘‘shorter” associations more accurate than the ‘‘longer” ones. In the following section, we will show how to achieve this goal. 5.2. Accuracy evaluation of knowledge recovery We define the relative difference: RðqÞ¼ suppD0ðqÞ� suppDðqÞ suppDðqÞ 100%: The frequently used measurement relative bias is represented by BðqÞ¼ jEðRðqÞÞj. This provides a measurement of the difference be- tween the support in D0 and the support in D. Thus, from the defini- tion of support, the smaller BðqÞ is, the more accurate the knowledge discovery can be. In this section, we first derive the gen- eral form of the relative error in our method. Then, we analyze how the quasi-sensitive associations are treated with respect to their length. Recall that Algorithm 1 randomly selects Q-I attribute for anon- ymization. Intuitively, the more Q-I attributes q involves, the more likely that q will be changed. Consequently, it is more likely that the difference RðqÞ will be larger. The following proposition shows a relation between the relative bias BðqÞ and the length of q. First we introduce some notations. Again, let q be a value combination in the original Q-I data. We denote :q the set of value combina- tions: fq̂ such that jq̂j ¼ jqj & QIðq̂Þ¼ QIðqÞ & q̂ – qg. Suppose q ¼fq1; . . . ; qlg; l 6 m, we denote q qi; i ¼ 1; . . . ; l, the value com- bination fq1; . . . ; qi�1; qiþ1; . . . ; qlg, and q qx the value combination fq1; . . . ; ql; qxg. When jqj ¼ 1, we define suppDðq qÞ¼ 1. We also use liftðqi; q qiÞ¼ suppðqÞ suppðq qiÞ PðQ i¼qiÞ to measure the strength of the relationship between qi and q qi . A lift value greater than 1 indi- cates there is a positive association between qi and q qi , whereas a value less than 1 indicates there is a negative association (Giudici, 2003). Proposition 12. For a value combination q in D, the relative bias BðqÞ is positively proportional to P i 1 liftðqi;q qiÞ � 1 h i . W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 761 Author's personal copy Proof 3. From the definition of RðqÞ, we have EðRðqÞÞ¼ E suppD0ðqÞ� suppDðqÞ suppDðqÞ � � ¼ suppDðqÞPðq ! qÞþ suppDð:qÞPð:q ! qÞ� suppDðqÞ suppDðqÞ ¼ suppDð:qÞ suppDðqÞ Pð:q ! qÞ� Pðq !:qÞ: ð9Þ Then we get BðqÞ¼ jEðRðqÞÞj ¼ 1 m Xjqj i¼1 suppDðq qiÞ� suppDðqÞ suppDðqÞ PðQ i ¼ qiÞ� PðQ i – qiÞ � � ¼ 1 m Xjqj i¼1 1 liftðqi; q qiÞ � 1 � � ; ð10Þ which completes the proof. h The proposition shows that the more attributes q has, the larger the range of relative bias BðqÞ is. In particular, when liftðqi; q qiÞ P 0:5, then BðqÞ 6 jqj m , implying that the range of the expected relative difference between suppD0 and suppD in- creases linearly with the length jqj. As for the variance of the relative difference, we have: VarðRðqÞÞ¼ 1 m2 Xjqj i¼1 suppDðq qiÞ 2 suppDðqÞ 2 PðQ i ¼ qiÞ� PðQ i ¼ qiÞ 2 � ! : ð11Þ Let value combination ~q ¼ q qx; qx R q. If suppDðq qiÞ suppDðqÞ varies little from suppDð~q qiÞ suppDð~qÞ for i ¼ 1; . . . ; jqj, then VarðRð~qÞÞ P VarðRðqÞÞ. That means VarðRðqÞÞ grows almost linearly with the increase of jqj. In other words, the relative difference between suppD0ðqÞ and suppDðqÞ be- comes more uncertain as value combination gets longer. In particu- lar, in the case of single attribute Q-I, where m ¼ 1, from (10), we have zero bias: BðqÞ¼ 1 liftðq; q qÞ � 1 ¼ 0: In summary, the shorter the length of value combination is, the bet- ter the quasi-sensitive knowledge is preserved, and the more accu- rate the knowledge recovery is. Also, since short associations are usually less sensitive than long ones, the user is more likely to dis- cover the less sensitive associations than the sensitive ones. 6. Enhancement of privacy protection Although the RA algorithm prevents the user from associating the individuals with their sensitive data, the prior knowledge of the associations of the Q-I data may worsen privacy leakage. In Fig. 5, the probabilistic anonymity PaðD0Þ of the anonymized data set is about 11 by Proposition 7. Suppose the user knows in advance the association: ‘‘Clerk, USA ! [30–40]” (50%). Once the value combination {[20–30], Clerk, USA} occurs in the anonymized data, the user is able to guess with a confidence that the original Q- I value combination is {[30–40], Clerk, USA}, where a ¼ Pð\age" is anonymizedÞ 50% � 17% > 1 PaðD0Þ : Thus, the user’s inference about the data privacy is improved. To tackle this problem, we randomly anonymize more than one Q-I attribute. This will improve privacy protection. How- ever, knowledge recovery may become less accurate. In this section, we generalize Algorithm 1 by uniformly choosing kðk P 1Þ Q-I attributes for each record and replacing their val- ues by using their original distributions. We first present the algorithm, then analyze its privacy protection and knowledge preservation. Algorithm 2. The Enhanced RA Algorithm 1: Input : the original data set D, the Q-I attributes Q, and the number kðk 6 mÞ of the attributes to be anonymized 2: Output : the original data set D is overwritten by an anonymized one 3: begin 4: m :¼ jQj; 5: n :¼ jDj; 6: Dist :¼;; 7: for i :¼ 1 to m do 8: begin 9: Disti :¼ the distribution of the values of Q i; 10: end 11: for j :¼ 1 to n do 12: begin 13: Randomly select k attributes with equal probability in Q-I of the jth record; 14: for k :¼ 1 to k do 15: begin 16: attr :¼ kth selected Q-I attribute; 17: Randomly generate a new value for attr based on Distattr ; 18: Replace the value of attr with the new value; 19: end 20: end 21: end Apparently, the Algorithm 2 enhances privacy protection when k gets larger. What is the impact of k on knowledge preservation? In the following, we will show that both the bias BðqÞ and variance VarðRðqÞÞ increase as k gets larger. Since we choose the Q-I attributes uniformly at random, the probability that each combination of k Q-I attributes gets selected is 1 m k � �� . In the algorithm, a value combination cannot be anon- ymized to all the other forms. If two value combinations q and q0 differ in g values ðg 6 kÞ, then these two value combinations are convertible into each other, since all the g values may be anony- mized. Before the discussion of BðqÞ and VarðEðqÞÞ, in the following proposition, we investigate the impact of k on the probability that a value combination is anonymized to its convertible value combina- tion. We will show that both probabilities Pðq ! q0Þ and Pðq0 ! qÞ increase as k increases. Proposition 13. Suppose k is the number of the anonymized attributes in each record. Let q 2 D and q0 2 D0 be partial value combinations of Q- I, jqj ¼ jq0j 6 m, and they are convertible into each other. Then the probabilities Pðq ! q0Þ and Pðq0 ! qÞ increase directly as k. Proof 4. Suppose that q and q0 differ in g attributes, g 6 k. Let AttrSetsgþk be the set of all possible combinations of g þ k Q-I attri- butes where each combination AttrSetsgþki ; i ¼ 1; . . . ; jqj� g k � � , contains those g attributes and other k; k 6 jqj� g, arbitrary Q-I attributes of q. Let PðAttrSetsgþki ; q 0Þ be the probability that the values of the g þ k attributes in AttrSetsgþki are anonymized to the 762 W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 Author's personal copy corresponding values of q0. Then probability that q is anonymized to q0 is: Pðq ! q0Þ ¼ Xminðk�g;jqj�gÞ k¼0 m �jqj k � g � k � � m k � � XjAttrSetsgþkj i¼1 PðAttrSetsgþki ; q 0Þ 2 6664 3 7775 ¼ Xminðk�g;jqj�gÞ k¼0 Prk;gk where, for simplicity, Prk;gk denotes m �jqj k � g � k � � m k � � PjAttrSetsgþkj i¼1 PðAttrSetsgþki ; q 0Þ. As k is increased by 1, Prkþ1;gk Prk;gk ¼ m �jqj k þ 1 � g � k � � m k þ 1 � � m �jqj k � g � k � � m k � � ¼ ðk þ 1Þðm � k þ 1Þ ðk þ 1 � g � kÞðm � k þ 1 �jqjþ g þ kÞ : It then follows that 1 6 Prkþ1;gk Prk;gk 6 k þ 1; since g þ k 6 minðk; jqjÞ. This shows that the probability Pðq ! q0Þ increases directly as k. The proof for Pðq0 ! qÞ is similar. h Now let us look at Bðq0Þ and VarðRðq0ÞÞ. From (9), we get Bðq0Þ ¼ X fq2:q0g suppDðqÞ suppDðq0Þ Pðq ! q0Þ � � � Pðq0 ! qÞ and VarðRðq0ÞÞ ¼ X fq2:q0g suppDðqÞ suppDðq0Þ � �2 ðPðq ! q0Þ� Pðq ! q0Þ2Þ " þ Pðq0 !:q0Þ� Pðq0 !:q0Þ2 � # : As k increases, there are more value combinations q in D convertible to q0. Also, from Proposition 13, both Pðq ! q0Þ and Pðq0 ! qÞ vary directly as k. Therefore, the range of BðqÞ increases with k. As for the variance, Pðq ! q0Þ and Pðq0 ! qÞ are usually less than 1/2, then Pðq ! q0Þ� Pðq ! q0Þ2 and Pðq0 !:q0Þ� Pðq0 !:q0Þ2 increase with k. Thus, the variance also varies directly as k. In summary, by tuning the value of k, we can adjust the level of the average bias and the variance. Thus, we need to find a compro- mise between the protection of data privacy and the preservation of useful knowledge, as demonstrated in our experiments. 7. Experiments In this section, we present our experiment results to demon- strate the effectiveness of our methods. By common aggregate que- ries, we first compare our RA Algorithm 1 with the l-diversity and permutation methods on the SQL queries. Then we compare our method with the permutation method on the quasi-sensitive knowledge discovery. At the same time, we demonstrate the ef- fects of the parameter k in our enhanced RA Algorithm 2 and the association length on knowledge discovery. 7.1. Experiment setup The data set adopted in the experiments is the adult data set from the UCI machine learning repository (Hettich & Bay, 1999). We used the following nine original attributes to compose the qua- si-identifier: ‘‘education”, ‘‘race”, ‘‘sex”, ‘‘work-class”, ‘‘marital-sta- tus”, ‘‘age”, ‘‘relationship”, ‘‘native-country” and ‘‘salary”. The attribute ‘‘occupation” was retained as the sensitive attribute. The records with missing values were removed and the resulting data set contained 30,162 records. 7.2. Accuracy of SQL queries In the first part of the experiments, we generated three anony- mized data sets by respectively applying the entropy l-diversity, permutation, and our RA Algorithm 1 to the test data set. We implemented the entropy l-diversity method based on the k-anon- ymization algorithm in LeFevre et al. (2006). While the permuta- tion algorithm was implemented as in Xiao & Tao (2006). Then, we performed random queries on the three anonymized data sets and compared the average accuracy of their answers. The queries were in the forms similar to those in Xiao & Tao (2006): select count ð�Þ from tablename where Q 1 in ðR1Þ and . . . and Q m in ðRmÞ and S in ðRsÞ where Q i is the ith Q-I attribute, S the sensitive attribute, and Ri the query values for the ith attribute. 7.2.1. Record count estimation When a data set is anonymized by the l-diversity algorithm, we cannot run the queries directly since the data is generalized. In- stead, the record count of each query is obtained by the following estimate (Xiao & Tao, 2006): X p #pðRsÞ Ym i¼1 #pðRi \ ValuesðQ iÞÞ #pðValuesðQ iÞÞ ! ; ð12Þ where ValuesðQ iÞ is the set of the distinct values of attribute Q i and function #pðxÞ counts in the pth Q-I group the occurrence of the val- ues in the set x. By summing up the estimated record counts in all the Q-I groups, we get an estimate for the query. When a data set is anonymized by the permutation method, the answers to the queries are estimated by X p #pðRsÞ #pðrecords matching the Q -I conditionsÞ #pðrecords in the p th groupÞ � � : ð13Þ In this case, in the pth Q-I group, the records matching the condi- tions of Q-I in the where clause in a query can be directly retrieved as #p (records matching the Q-I conditions). The product in the above equation assumes all the associations between the value combinations of Q-I and sensitive data are equally possible. We call this kind of associations the uniform associations between the Q-I and sensitive values. When the data set is anonymized by the RA algorithm, we re- trieve the queries directly without estimation. 7.2.2. The results We conducted eight sets of experiments. In the jth set, 1 6 j 6 8, of experiments, we performed 1000 queries on the three anonymized data sets and each query was composed of the sensitive attribute and j attributes which were selected from the quasi-identifier uniformly at random. For each selected attribute, we chose a random number of categorical values from its domain to form its query values Ri. In each set of experiments, we W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 763 Author's personal copy calculated the average relative errors of the query results. The rel- ative error is defined by jAnsðquery; DÞ� Ansðquery; D0Þj Ansðquery; DÞ ; where Ansðquery; DÞ returns the answer to query on data set D. Since the numerator jAnsðquery; DÞ� Ansðquery; D0Þj compares the aggre- gated counts between D and D0, the above definition of relative error may exceed 1. The probabilistic anonymity of the test data set reached 34 by our RA anonymization method, meaning that the average probabil- ity that a user could correctly associate an individual with the cor- responding sensitive value was 1/34. While the l-diversity and permutation methods limited that probability to 1=l, which was further limited by the distribution of the sensitive values. In the test data set, the maximal value of l was about 10 (Entropy(occu- pation)� ln 10). Thus, we set l ¼ 10 for the l-diversity and permu- tation methods. As shown in Fig. 6a, when the queries become more sensitive, involving more attributes, the relative error in the permuted data varies less than those in the l-diverse data and the RA anonymized data. This is because the estimate (12) for the results of the queries on the l-diverse data assumes a uniform value distribution of each Q-I attribute. Thus, the more Q-I attributes are involved, the farther away their joint distribution is from the multivariate uniform dis- tribution. As for the RA method, from Proposition 12, the frequent and infrequent value combinations usually have higher values of BðqÞ than other value combinations. Since the proportions of the frequent and infrequent value combinations decrease when the number of attributes involved increases, the rise of the relative er- ror slows down as the number of attributes involved increases, as shown in Fig. 6a. For the permuted data set, the data links in the original data set are not completely uniform associations assumed by (13), causing the errors in the queries. Since the strength of those associations is not influenced by the number of the attributes involved in the queries, the relative error varies the least. We then decreased the value of diversity l to 8 and 5 and re- peated the comparison. This time, each query in the experiment consisted of the sensitive attribute and 5 random Q-I attributes. As shown in Fig. 6b, although the accuracy of the estimations in the l-diverse data and permuted data improved as l decreased, the estimation from the anonymized data set by our RA method consistently had the least amount of error. This is because the esti- mation from the RA anonymized data set assumes no relationships between the attributes. The RA algorithm tries to preserve the data relationships by perturbing a small portion of data set. 7.3. Quasi-sensitive knowledge preservation In this section, we investigate knowledge preservation in sev- eral anonymization techniques, the effect of the parameter k in the enhanced RA Algorithm 2 on knowledge discovery, and the im- pact of the association length on knowledge discovery. We first ran the Apriori algorithm (Agrawal & Srikant, 1994) on the original data set D to get the quasi-sensitive associations AssoD, which was used as the baseline result in the following experi- ments. Then we compared baseline results with the quasi-sensitive associations from each anonymized data set D0. In the comparison, we evaluated the accuracy of the knowledge discovery by calculat- ing the ‘‘relative percentage”: jAssoD \ AssoD0 j jAssoDj 100%: In the test data set, the highest frequency of the values of attribute ‘‘occupation” was about 13%. Thus, in the Apriori algorithm, we set the confidence threshold to 15% and the support threshold to 8% . Fig. 7 compares the effect of our RA Algorithm 1 with that of the permutation method with l ¼ 5; 8; 10 (We did not compare with l- diversity and k-anonymization since most data associations are suppressed there). More quasi-sensitive associations were recov- ered from our RA anonymized data than from the permuted data. Although the previous experiments of aggregate queries showed that the accuracy of the answers from the permuted data varied the least, in the current experiment, only a small part of the asso- ciations was recovered. This is caused by the assumption of the uniform associations between the Q-I and sensitive values. In the previous experiments, most queries consisted of independent attri- bute values. For example, in the first set of the previous experi- ments, the queries consisted of the sensitive attribute and one random Q-I attribute. We ran the Chi-Square test for each contin- gency table consisting of a pair of attributes in the queries. Most Fig. 6. (a) Comparison of the query accuracy among the data sets by different anonymization methods ðl ¼ 10Þ; (b) comparison of the query accuracy with different diversity values. Fig. 7. Comparison of the quasi-sensitive associations among different anonymi- zation methods. 764 W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 Author's personal copy results of v 2 jDj was less than 0.2. Thus, the relative errors from the queries consisting of associated attribute values contributed little to the average error. However, in the association discovery, it only discovered those highly associated attribute values. The more interesting the associations are, the farther away they are from the uniform relationship. While in our RA method, the data ran- domization had a relatively small impact on the data relationships. Therefore, it was capable of preserving more associations. To test the influence of k on the discovery of the quasi-sensitive knowledge, we implemented the enhanced RA Algorithm 2 with various numbers of Q-I attributes anonymized. We set the param- eter k to 1,3,5,7,9. As shown in Fig. 8a, with the increase of k, more artificial asso- ciations were generated and fewer original associations were dis- covered. This confirms our conclusions in Section 6. Since a quasi-sensitive association may cause more potential threats to the data privacy when it involves more attribute values, we further compared the discovery of the quasi-sensitive associa- tions with various association lengths. In the comparison, we set k ¼ 5. Fig. 8b shows that it is more accurate to recover the shorter associations than the longer ones. Most of the associations with 1, 2 or 3 attributes were discovered even when more than half of the attribute values were anonymized in each record. (Here, we de- note by length 1 the frequent items in the data set.) This means that the user is more likely to discover less sensitive associations. The above experiments demonstrate the following characteris- tics of our methods. The RA Algorithm 1 is capable of protecting data privacy with minor impact on the data distribution. The en- hanced RA Algorithm 2 can further decrease the risk of privacy leakage when some original data associations are disclosed. More- over, by tuning the value of k, it can make a balance between the useful knowledge preservation and privacy protection. 8. Conclusions In this paper, we introduce a novel data anonymization method by randomizing the data records. Different from the generalization methods such as the k-anonymization methods, we regard the data privacy as the links between the Q-I and sensitive values. By ran- domly replacing part of the values in each record while maintain- ing statistical relations in whole data set, the RA method not only achieves a higher level of the privacy protection but also preserves more non-sensitive knowledge than the other anonymization methods, as demonstrated by our experiments. Moreover, the use- ful associations which are less sensitive can be discovered more accurately than the sensitive ones. In particular, when association length is 1, our method delivers zero bias. Furthermore, the en- hanced RA method provides the extra privacy protection when the user has some prior knowledge of the Q-I data. Data anonymization is a popular direction in the research of pri- vacy preserving data mining. Integrating our method with other privacy preserving techniques, we may accomplish more tasks. This work serves as an initial step. Additional research will include more work on other types of data knowledge and the definition of data privacy. References Aggarwal, C. C. (2005). On k-anonymity and the curse of dimensionality. In Proceedings of the 31st international conference on very large data bases (pp. 901– 909). Trondheim, Norway. Agrawal, D., & Aggarwal, C. C. (2001). On the design and quantification of privacy preserving data mining algorithms. In Proceedings of the 20th ACM SIGMOD- SIGACT-SIGART symposium on principles of database systems (pp. 247–255). Santa Barbara, California. Agrawal, R., & Srikant, R. (1994). Fast algorithms for mining association rules in large databases. In Proceedings of the 20th international conference on very large data bases (pp. 487–499). Santiago, Chile. Agrawal, R., & Srikant R. (2000). Privacy-preserving data mining. In Proceedings of the 2000 ACM SIGMOD international conference on management of data (pp. 439– 450). Dallas, Texas. Aggarwal, C. C., & Yu, P. S. (2008). A general survey of privacy-preserving data mining models and algorithms. Privacy-preserving data mining (pp. 11–52). US: Springer. Chen, C., Jin, F., Zhu, X., & Ouyang, G. (1983). Mathematical analysis. Higher Education Press. Ciriani, V., Vimercati, S., Foresti, S., & Samarati, P. (2008). k-Anonymous data mining: A survey. Privacy-preserving data mining (pp. 105–136). US: Springer. Clifton, C., Kantarcioglu, M., Vaidya, J., Lin, X., & Zhu, M. Y. (2002). Tools for privacy preserving distributed data mining. ACM SIGKDD Explorations Newsletter, 4(2), 28–34. Du, W., & Zhan, Z. (2003). Using randomized response techniques for privacy- preserving data mining. In Proceedings of the ninth ACM SIGKDD international conference on knowledge discovery and data mining (pp. 505–510). Washington, DC. Evfimievski, A., Gehrke, J., & Srikant, R. (2003). Limiting privacy breaches in privacy preserving data mining. In Proceedings of the 22th ACM SIGMOD-SIGACT-SIGART symposium on principles of database systems (pp. 211–222). San Diego, California. Fung, B. C. M., Wang, K., & Yu, P. S. (2007). Anonymizing classification data for privacy preservation. IEEE Transactions on Knowledge and Data Engineering, 19(5), 711–725. Giudici, P. (2003). Applied data mining: Statistical methods for business and industry. John Wiley & Sons. Ltd. Hettich, S., & Bay, S. D. (1999). The UCI KDD archive. URL : http://kdd.ics.uci.edu. Kifer, D., & Gehrke, J. (2006). Injecting utility into anonymized datasets. In Proceedings of the 2006 ACM SIGMOD international conference on management of data (pp. 217–228). Chicago, Illinois. Koudas, N., Srivastava, D., Yu, T., & Zhang, Q. (2007). Aggregate query answering on anonymized tables. In Proceedings of the 23rd international conference on data engineering (pp. 116–125). Istanbul, Turkey. LeFevre, K., DeWitt, D. J., & Ramakrishnan, R. (2005). Incognito: Efficient full-domain k-anonymity. In Proceedings of the 2005 ACM SIGMOD international conference on management of data (pp. 49–60). Baltimore, Maryland. LeFevre, K., DeWitt, D. J., & Ramakrishnan, R. (2006). Mondrian multidimensional k- anonymity. In Proceedings of the 22nd international conference on data engineering (p. 25). Atlanta, Georgia. Li, N., Li, T., & Venkatasubramanian, S. (2007). t-Closeness: Privacy beyond k- anonymity and l-diversity. In Proceedings of the 23rd international conference on data engineering (pp. 106–115). Istanbul, Turkey. Lindell, Y., & Pinkas, B. (2000). Privacy preserving data mining. In Proceedings of the 20th annual international cryptology conference on advances in cryptology (pp. 36–54). Santa Barbara, California. Fig. 8. (a) Discovery of the quasi-sensitive associations with different values of k in the enhanced RA method; (b) discovery of the quasi-sensitive associations with different length ðk ¼ 5Þ. W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 765 Author's personal copy Machanavajjhala, A., Gehrke, J., Kifer, D., & Venkitasubramaniam, M. (2006). l- Diversity: Privacy beyond k-anonymity. In Proceedings of the 22th international conference on data engineering (p. 24). Los Alamitos, California. Meyerson, A., & Williams, R. (2004). On the complexity of optimal k-anonymity. In Proceedings of the 23rd ACM SIGMOD-SIGACT-SIGART symposium on principles of database systems (pp. 223–228). Paris, France. Sweeney, L. (2002). Achieving k-anonymity privacy protection using generalization and suppression. International Journal of Uncertainty, 10(5), 571–588. Xiao, X., & Tao, Y. (2006). Anatomy: Simple and effective privacy preservation. In Proceedings of the 32nd international conference on very large data bases (pp. 139– 150). Seoul, Korea. 766 W. Yang, S. Qiao / Expert Systems with Applications 37 (2010) 756–766 
work_hixbsaviinbf5aytigbyw2i3ja	----	Supplier selection using AHP methodology extended by D numbers Supplier selection using AHP methodology extended by D numbers Xinyang Deng a,1, Yong Hu b,1, Yong Deng a,c,⇑, Sankaran Mahadevan c a School of Computer and Information Science, Southwest University, Chongqing 400715, China b Institute of Business Intelligence and Knowledge Discovery, Guangdong University of Foreign Studies, Sun Yat-Sen University, Guangzhou 510006, China c Department of Civil and Environmental Engineering, School of Engineering, Vanderbilt University, Nashville, TN 37235, USA a r t i c l e i n f o Keywords: Supplier selection D numbers D-AHP Analytic hierarchy process Fuzzy preference relation a b s t r a c t Supplier selection is an important issue in supply chain management (SCM), and essentially is a multi-criteria decision-making problem. Supplier selection highly depends on experts’ assessments. In the process of that, it inevitably involves various types of uncertainty such as imprecision, fuzziness and incompleteness due to the inability of human being’s subjective judgment. However, the existing methods cannot adequately handle these types of uncertainties. In this paper, based on a new effective and feasible representation of uncertain information, called D numbers, a D-AHP method is proposed for the supplier selection problem, which extends the classical analytic hierarchy process (AHP) method. Within the proposed method, D numbers extended fuzzy preference relation has been involved to repre- sent the decision matrix of pairwise comparisons given by experts. An illustrative example is presented to demonstrate the effectiveness of the proposed method. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction Supply chain management (SCM) has emerged as an important element in improving the global competitiveness of companies in the highly competitive global economy (Othman & Ghani, 2008; Gunasekaran & Ngai, 2005). One of the key problems in SCM is sup- plier selection (Wu and Barnes, 2011; Huang and Keskar, 2007, i.e., find the best supplier among several alternatives according to var- ious criteria, such as cost, service, risk and others. This is a compli- cated multi-criteria decision-making problem (Chai, Liu, & Ngai, 2013). The identification of criteria and integration of experts’ assessments are two components of the supplier selection problem. With respect to the identification of criteria, selection of criteria and calculation of priority weights both should be considered. Lee, Kang, Hsu, and Hung (2009) investigated the green supplier selec- tion problem for high-tech industry, and identified six criteria, namely quality, technology capability, pollution control, environ- mental management, green production and green competency. In the automobile industry, Govindan, Kannan, and Haq (2010) devel- oped a framework to identify and rank the associated criteria, for instance asset specificity and supplier performance. Aksoy and Ozturk (2011) explored the problem of supplier selection in just-in-time (JIT) production environments. It is worth noting that environmental criteria have gradually begun to attract increasing attention in supplier selection. Humphreys, Wong, and Chan (2003) identified seven environmental categories, including envi- ronmental costs (pollutant effects), environmental costs (improve- ment), management competencies, ‘‘green’’ image, design for environment, environmental management systems, and environ- mental competencies. Generally speaking, the criteria for supplier selection highly de- pend on individual companies and industries. On the one hand, dif- ferent companies have different organizational structure, management strategy, enterprise culture and others. All of these influence the determination of supplier selection criteria. On the other hand, industry background causes huge difference and greatly impacts the selection of suppliers. Therefore, the identifica- tion of supplier selection criteria are on the basis of specific envi- ronments, and largely requires domain experts’ assessment and judgement. After identifying the criteria, a methodology is needed to inte- grate experts’ assessments in order to find the best supplier. At present, various methods have been applied to supplier selection, such as the analytic hierarchy process (AHP) (Chan, 2003; Wang, Chin, & Leung, 2009), fuzzy set theory (Ordoobadi, 2009; Labib, 2011), fuzzy extended AHP method (Chan & Kumar, 2007; Chan, Kumar, Tiwari, Lau, & Choy, 2008; Kilincci & Onal, 2011), TOPSIS method (Boran, Genc, Kurt, & Akay, 2009; Deng, 2006; Wang, Cheng, & Kun-Cheng, 2009; Zouggari & Benyoucef, 2012), Dempster–Shafer evidence theory (Deng & Chan, 2011) and others (Chen & Chao, 2012; Ferreira & Borenstein, 2012; Ng, 2008; Sana- yei, Mousavi, & Yazdankhah, 2010). Among these methods, AHP 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.07.018 ⇑ Corresponding author at: School of Computer and Information Science, South- west University, Chongqing 400715, China. E-mail addresses: ydeng@swu.edu.cn, prof.deng@hotmail.com (Y. Deng). 1 These authors contributed equally to this work. Expert Systems with Applications 41 (2014) 156–167 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2013.07.018&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.07.018 mailto:ydeng@swu.edu.cn mailto:prof.deng@hotmail.com http://dx.doi.org/10.1016/j.eswa.2013.07.018 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa (Ngai & Chan, 2005; Saaty, 1980) is popular and extensively used to deal with complex decision problem due to its simplicity in con- cept and convenience in operation. Even so, its inability to ade- quately handle the inherent uncertainty and imprecision in the data is also often criticized. The fuzzy extended AHP method (Buy- ukozkan, Cifci, & Guleryuz, 2011; Chan & Kumar, 2007; Chan et al., 2008; Sadiq & Tesfamariam, 2009; Wang, Luo, & Hua, 2007) pro- vides the ability to handle the fuzziness in the experts’ subjective judgment with the aid of fuzzy set theory. However, the inconsis- tency of the fuzzy evaluation comparison matrix is still an open is- sue (Leung & Cao, 2000; Wang & Chen, 2008). Some researchers have studied the concept of fuzzy preference relation (Chen & Chao, 2012; Garcia, del Moral, Martinez, & Herrera-Viedma, 2012; Herrera-Viedma, Herrera, Chiclana, & Luque, 2004; Herre- ra-Viedma, Alonso, Chiclana, & Herrera, 2007; Lee, 2012; Tanino, 1984; Wang & Fan, 2007; Xu, 2007) which directly enables the fuzz- iness of preference relation to be represented in the form of mem- bership function. Yet the method of fuzzy preference relation is still unable to handle situations when the information is incomplete. Supplier selection highly depends on large amounts of domain knowledge, where experts’ assessments play an important role. However, various uncertainties are present in domain experts’ sub- jective and qualitative judgment, such as imprecision, fuzziness, incompleteness and so on. Therefore it is necessary to develop a more effective method for supplier selection, which should be able to handle various types of uncertainties (Kara, 2011; Li & Zabin- skys, 2011; Wu, 2009). In order to effectively handle various uncertainties involved in the supplier selection problem, a new representation of uncer- tain information, called D numbers (Deng, 2012), is presented in this paper. The concept of D numbers extends the Dempster– Shafer evidence theory (Dempster, 1967; Shafer, 1976), and is more effective in representing uncertain information, as dis- cussed later. Secondly, a D numbers-based preference relation is proposed by extending the fuzzy preference relation approach. The proposed D numbers preference relation overcomes the defi- ciency of fuzzy preference relation. Finally, the AHP method is extended using the D numbers-based preference relation, leading to the proposed D-AHP method for decision-making under uncertain information. The rest of this paper is organized as follows. Section 2 gives a brief introduction to the classical AHP method. Section 3 develops the proposed D-AHP method for supplier selection. It contains the introduction of D numbers and fuzzy preference relation, the con- cept of D numbers preference relation, the calculation of priority weights and inconsistency degree in D numbers preference rela- tion, and the D-AHP method. After that, an illustrative example of supplier selection is given to show the effectiveness of the pro- posed D-AHP method in Section 4. Finally, Section 5 concludes the paper. 2. Analytic hierarchy process (AHP) AHP Saaty (1980) is a structured technique for handling com- plex decision problems. Both qualitative and quantitative factors are combined by using AHP in the decision-making process. Generally, the process of applying AHP can be divided into three steps. First, establish a hierarchical structure by recursively decom- posing the decision problem. Second, construct the pairwise com- parison matrix to indicate the relative importance of alternatives. A numerical rating including nine rank scales is suggested, as shown in Table 1. Third, calculate the priority weights of alternatives according to the pairwise comparison matrix by the following equation: Aw ¼ kmaxw; w ¼ðw1; w2; . . . ; wnÞ T ð1Þ where A is a n dimensional comparison matrix, kmax is the largest eigenvalue of A and w is the eigenvector corresponding to kmax. In AHP, a consistency index (C.I.) is defined to measure the inconsistency within the pairwise comparison matrix A. C:I: ¼ kmax � n n � 1 ð2Þ Accordingly, the consistency ratio C.R. is used to measure the degree of C.I. by the following equation: C:R: ¼ C:I: R:I: ð3Þ where R.I. is the random consistency index, its value is related to the dimension of the matrix, listed in Table 2. If C.R. < 0.10, the inconsistency degree of the comparison matrix A is consider acceptable and the eigenvector w is used as the weighting vector after normalization. Otherwise, the comparison matrix needs to be adjusted. 3. D-AHP for supplier selection 3.1. D numbers Information about the real world is affected by many sources of uncertainty. Many existing theories, such as probability theory, fuzzy set theory (Zadeh, 1965), Dempster–Shafer evidence theory (Dempster, 1967; Shafer, 1976), have been developed to model various types of uncertainty, and provide some desirable proper- ties. But these theories, however, still contain deficiencies that can- not be ignored. Take Dempster–Shafer evidence theory as an example. Due to its inherent advantages in the representation and handling of uncertain information, the Dempster–Shafer evi- dence theory is being studied for use in many fields, such as decision analysis, pattern recognition, risk assessment, supplier selection and others (Dempster, 1967; Deng, Shi, Zhu, & Liu, 2004; Deng, Su, Wang, & Li, 2010; Deng, Chan, Wu, & Wang, 2011; Deng, Sadiq, Jiang, & Tesfamariam, 2011; Deng, Jiang, & Sadiq, 2011; Deng, Liu, Hu, & Deng, 2013; Kang, Deng, Sadiq, & Mahadevan, 2012; Liu, Chan, Li, Zhang, & Deng, 2012; Sadiq, Klein- er, & Rajani, 2006; Sadiq, Najjaran, & Kleiner, 2006; Sadiq, Kleiner, & Rajani, 2007; Shafer, 1976; Zhang, Deng, Wei, & Deng, 2012; Wei, Deng, Zhang, Deng, & Mahadevan, 2013). In the mathematical framework of Dempster–Shafer theory, the basic probability assignment (BPA) defined on the frame of discernment is used to express the uncertainty in judgement. A problem domain indicated by a finite non-empty set X of mutually exclusive and exhaustive hypotheses is called a frame of discernment. Let 2X denote the power set of X, a BPA is a mapping m:2X ? [0, 1], satisfying mð;Þ¼ 0 and X A22X mðAÞ¼ 1 ð4Þ Table 1 Numerical rating in the AHP method. Scale Meaning 1 Equal importance 3 Moderate importance 5 Strong importance 7 Demonstrated importance 9 Extreme importance 2, 4, 6, 8 Intermediate values Table 2 Random consistency index R.I. n 1 2 3 4 5 6 7 8 9 10 R.I. 0 0 0.52 0.89 1.12 1.26 1.36 1.41 1.46 1.49 X. Deng et al. / Expert Systems with Applications 41 (2014) 156–167 157 https://isiarticles.com/article/46527 
work_hlhwknhymfg3jhmv2t3rh5wonu	----	Intelligent profitable customers segmentation system based on business intelligence tools Intelligent profitable customers segmentation system based on business intelligence tools Jang Hee Lee a,*, Sang Chan Park b,1 a School of Industrial Management, Korea University of Technology and Education, 307 Gajeon-ri, Byeong cheon-myun, Cheonan City, Choongnam Province 330-708, South Korea b Department of Industrial Engineering, Korea Advanced Institute of Science and Technology (KAIST), 373-1 Kusong-dong, Yusong-gu, Taejon 305-701, South Korea Abstract For the success of CRM, it is important to target the most profitable customers of a company. Many CRM researches have been performed to calculate customer profitability and develop a comprehensive model of it. Most of them, however, had some limitations and accordingly the customer segmentation based on the customer profitability model is still underutilized. This paper aims at providing an easy, efficient and more practical alternative approach based on the customer satisfaction survey for the profitable customers segmentation. We present a multi- agent-based system, called the survey-based profitable customers segmentation system that executes the customer satisfaction survey and conducts the mining of customer satisfaction survey, socio-demographic and accounting database through the integrated uses of business intelligence tools such as DEA (Data Envelopment Analysis), Self-Organizing Map (SOM) neural network and C4.5 for the profitable customers segmentation. A case study on a Motor company’s profitable customer segmentation is illustrated. q 2005 Elsevier Ltd. All rights reserved. Keywords: Customer relationship management; Customer profitability; Customer segmentation; Customer satisfaction survey 1. Introduction In today’s competitive business environment, the ability to identify profitable customers, build their long-term loyalty and steadily expand existing relationships is key competitive factors to a company. To meet these factors, companies across a wide range of industries have made Customer Relationship Management (CRM) one of the leading business strategies, integrating sales, marketing and service across multiple business units and customer contact points. CRM helps companies understand the value of custo- mers, target their most profitable customers, cultivate and maintain high-quality relationships that increase loyalty and profits. Precise evaluation of customer profitability and targeting the most profitable customers are crucial elements for the success of CRM. Many CRM researches have been performed to calculate customer profitability based on customer lifetime value and develop a comprehensive model of it. Most of them, however, had some limitations by not considering such as the change of profit contribution resulted from the customer defection (Berger & Nasr, 1998; Gupta & Lehmann, 2003). They need further extensions considering additional factors such as customer reactivation possibility, attracting/service cost and causes of customer defection. On the other hand, the customer segmentation based on their profitability to a company is still an underutilized approach. This study aims at providing an easy, efficient and more practical alternative approach based on the customer satisfaction survey for the profitable customers segmenta- tion instead of using a customer profitability model, which is an important tool for marketing and managing customer relationships by providing the information of overall satisfaction level, repurchase intentions, word-of-mouth intentions, etc. In our approach, we use intelligent tools such as Data Envelopment Analysis (DEA), Self-Organizing Map (SOM) 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.01.013 Expert Systems with Applications 29 (2005) 145–152 www.elsevier.com/locate/eswa * Corresponding author. Tel.: C82 41 560 1446; fax: C82 41 560 1439. E-mail addresses: janghlee@kut.ac.kr (J.H. Lee), sangpark@kaist.ac.kr (S.C. Park). 1 Tel.: C82 42 869 2960; fax: C82 42 869 5920. http://www.elsevier.com/locate/eswa neural network and C4.5 to segment profitable customers. DEA evaluates efficiency through the relation analysis between the company’s input costs for a customer (e.g. marketing cost, production cost, inventory cost, delivery cost, service cost and relationship management cost) and the output (e.g. his/her satisfaction level, repurchase intentions and word-of-mouth intentions in the customer satisfaction survey and his/her profit contribution to it). Through the successive mining of customer satisfaction survey and socio-demographic data by SOM and C4.5, we segment profitable customers among all the surveyed customers. We present a survey-based profitable customers seg- mentation system (SPCSS) that designs, executes (on-line, e-mail, etc.) the customer satisfaction survey for all customers in customer database of a company and conducts those mining works for the profitable customers segmenta- tion. SPCSS has an architecture based on intelligent agent technology and also the integration of those mining process into decision support system framework by means of applying that technology. This paper is organized as follows. Section 2 presents a review of literature in the profitable customer segmentation and customer satisfaction survey. In Section 3, we introduce our research methodology for profitable customer segmen- tation based on customer satisfaction survey and the basic structure of proposed system incorporated that methodology is presented. A case study on a Motor company’s profitable customer segmentation in South Korea is illustrated in Section 4 and the concluding remarks are presented in Section 5. 2. Profitable customer segmentation and customer satisfaction survey Traditional customer segmentation models were based on demographic, attitudinal, and psychographic attributes of a customer (Griffin, 2003). They gave too simple results and poor accuracy for today’s complicated business environ- ment. Recently, the customer segmentation based on customer transactional and behavioral data (e.g. purchases type, volume and history, call center complaints, claims, web activity data, etc.) collected by various information systems is commonly used. However, the customer segmentation based on his/her profitability to a company is still underutilized. Customer profitability is a customer-level measure that refers to the revenues less the costs which one particular customer generates over a given period of time and has been studied the name of Customer value, Customer Lifetime Value, LTV and Customer Equity. Many customer profit- ability researches focused on the future cash flow derived from the past profit contribution and did not considered the change of profit contribution resulted from the customer defection (Berger & Nasr, 1998; Gupta & Lehmann, 2003). Hwang, Jung, and Suh (2004) suggested a new customer profitability model considering past profit contribution, potential benefit indicated cross-selling and up-selling opportunity, and defection probability of a customer measured customer loyalty and segmented customers based on their model. However, they said that it had some limitations such as not considering the reactivation possibility of customers, attracting/servicing cost and causes of customer defection. It is difficult and complicated to develop an effective and exact customer profitability model and segment profitable customers based on that model. In this study, we provide an easy, efficient and more practical alternative approach through the customer satisfaction survey for the profitable customers segmentation instead of using that model. The typical customer satisfaction survey collects data on the causal context of satisfaction, i.e. antecedents (e.g. perceived performance of various product attributes/ser- vice) and consequences (e.g. overall satisfaction level, repurchase intentions and word-of-mouth intentions). According to the Satisfaction-Profit Chain principle (Anderson & Mittal, 2000), improving product and service attributes causes increased customer satisfaction, increased customer satisfaction leads to greater customer retention and improving customer retention greater profitability. Empirical Researches have shown that increasing overall satisfaction leads to greater repurchase intentions, as well as to actual repurchase behavior and companies with high customer satisfaction and retention can expect higher profits (Reichheld & Frederick, 1996). In this study, we use the customer’s overall satisfaction level, repurchase intentions, word-of-mouth intentions obtained from the customer satisfaction survey and his/her profit/loss to a company derived from the accounting database of it for the first step of profitable customers segmentation. 3. Profitable customers segmentation based on customer satisfaction survey We propose a survey-based profitable customers seg- mentation system (SPCSS) based on data mining and agent technology that designs, executes (on-line, e-mail, etc.) customer satisfaction survey and conducts predefined mining processes for the profitable customers segmentation. SPCSS has a multi-agent based architecture and the integration of predefined mining processes into decision support system framework (Fig. 1). There are three types of intelligent agents within the SPCSS architecture: Survey management (SM) agent with survey knowledge base that provides system co-ordination, facilitates (mined) knowledge communication, and takes the charge of design and execution of customer satisfaction survey, profitable customers segmentation (PCS) agent that segments profitable customers among all the surveyed J.H. Lee, S.C. Park / Expert Systems with Applications 29 (2005) 145–152146 https://isiarticles.com/article/671 
work_hmao7uaq3rftdo2plxoowkbtrm	----	PII: S0950-7051(99)00005-2 Applications of rule-base coverage measures to expert system evaluation V. Barrp Department of Computer Science, Hofstra University, Hempstead, NY 11550, USA Received 24 September 1998; received in revised form 5 January 1999; accepted 5 January 1999 Abstract Often a rule-based system is tested by checking its performance on a number of test cases with known solutions, modifying the system until it gives the correct results for all or a sufficiently high proportion of the test cases. This method cannot guarantee that the rule-base has been adequately or completely covered during the testing process. We introduce an approach to testing of rule-based systems, which uses coverage measures to guide and evaluate the testing process. In addition, the coverage measures can be used to assist rule-base pruning and identification of class dependencies, and serve as the foundation for a set of test data selection heuristics. We also introduce a complexity metric for rule-bases. q 1999 Elsevier Science B.V. All rights reserved. Keywords: Rule-base system; Casual-associational network; Logical path graph 1. Introduction Evaluation of a knowledge-based system is a multi- faceted problem, with numerous approaches and techniques. The results generated by the system must be evaluated, along with its features, the usability of the system, how easily it can be enhanced, and whether or not it has a posi- tive impact on the people who are using the system in place of an approach which is not computer based. The system’s performance must also be evaluated in light of its intended use [1]. If the expert system is meant to function as an intelligent assistant then it must satisfy the criterion of being a useful adjunct to the human problem solver. If the system is expected to emulate the reasoning of a human expert then a more rigorous evaluation of the system is needed. During the last 20 years there has been considerable development and use of knowledge-based systems for medi- cal decision support. In this period there has been heavy emphasis on functional analysis, addressing two primary questions: • Does the system give the results we expect on test cases? • Does the system improve the effectiveness of those who use it? The emphasis on functional analysis can lead to seemingly strong statistical statements about the correctness of a system, demonstrating that it gives the correct result, or the same result as a human expert, in a high percentage of test cases. However, functional testing does not guarantee that all parts of the system are actually tested. If a section of the rule-base is not exercised during the functional test then there is no information about that section of the system and whether it is correct or contains errors. Further, many performance problems for rule-bases result from unforeseen rule interactions [2]. A test suite of known cases may never trigger these interactions, though they should be identified and corrected before a system is put into actual use. The method we present enhances functional analysis of rule-based classification systems with a rule-base coverage assessment, overcoming limitations of common methods for rule-based expert systems evaluation. The underlying premise of this work is that an ideal testing method is one that guarantees that all possible reasoning paths through a rule-base have been exercised. As with procedural software, this is often an unreasonable and/or unattainable goal, possi- bly due to a lack of test data, to un-executable program paths, or to the size of the rule-base. Further, even if each possible path is exercised, we cannot realistically do so with each distinct set of test values that could cause its traversal. A reasonable goal is for the rule-base testing process to exercise every inference chain or provide information about the failure of the testing process to do so. Usually verification and validation (V & V) of rule-based systems involves a static structural analysis (verification) method to detect internal inconsistencies, followed by a dynamic, functional, validation in which system behavior on a set of test cases is compared with expected results. The Knowledge-Based Systems 12 (1999) 27–35 0950-7051/99/$ - see front matter q 1999 Elsevier Science B.V. All rights reserved. PII: S 0 9 5 0 - 7 0 5 1 ( 9 9 ) 0 0 0 0 5 - 2 pE-mail address: vbarr@magic.hofstr.edu (V. Barr) weakness of a strictly functional approach to validation is that the test data available may not adequately cover the rule-base, and, at best, limited information about coverage will be obtained. System performance statistics are usually presented as if they apply to the entire rule-base, rather than just to the tested sections. This can lead to false estimates of system performance in actual use. The system performance indicated by the comparison of actual and expected results is relevant only for the tested sections, while performance in the untested sections cannot be predicted. We must also consider completeness of the test set and age of the rule-base by the test data. Completeness of the test set refers to the degree to which the data represents all types of cases, which could be presented to the system under intended conditions of use. Coverage of the rule-base refers to how extensively possible combinations of inference rela- tions are exercised during test data evaluation. In the trivial case, with a correct rule-base and a complete test suite, the test data would completely cover the rule-base, all actual results would agree with expected results, and we could predict completely correct performance of the rule-base in actual use. In the more usual situation we may have errors and incompleteness in the rule-base, as well as inadequacies in the test data. If we only judge the system based on a comparison of actual and expected results, the rule-base could perform well on the test data, but actually contain errors which are not identified due to incompleteness of the test data. This could lead to a false prediction of correct performance on all cases, when in fact we cannot make any accurate prediction about performance of the rule-base in those areas for which there is an absence of test data. Our testing approach, as outlined in Fig. 1, allows clear identification of incompleteness in the test data and poten- tial errors in the rule-base through identification of sections of the rule-base that have not been exercised during func- tional test. This can indicate weaknesses in the test set and/ or sections of the rule-base that may not be necessary. An incomplete test set can be supplemented with additional cases chosen from the available population, guided by a series of heuristics and the coverage analysis information. Alternatively, if there is no test data which covers certain parts of the system, it is possible that those sections should be pruned from the rule-base or modified. Our approach carries out structural analysis of the rule- base using five rule-base coverage measures (RBCMs) which identify sections not exercised by the test data. This makes it possible to improve completeness of the test suite, thereby increasing the kinds of cases on which the rule-base will be tested and improving coverage of the rule-base. In addition to the coverage analysis, we employ a rule- base representation which facilitates application of the coverage measures; a set of heuristics for re-sampling the population of available test cases, based on coverage infor- mation, as shown in Fig. 2; strategies for rule-base pruning and identification of class-dependencies; a rule-base complexity metric. In another study [3] the utility of the aforementioned is illustrated extensively using rule-bases which were prototypes for the AI/RHEUM system [4] and the TRUBAC (Testing with RUle-BAse Coverage), a tool which implements the coverage analysis method. 2. Related work This work builds on both coverage-based testing methods for procedural software (see [5] for a review of methods and [6,7] for a data-flow approach to testing) and earlier work on rule-base analysis. Early approaches for rule-base analysis carried out only verification or validation. A number of systems, such as the ONCOCIN rule checker program (RCP) [8], CHECK [9,10], ESC (expert system checker) [11], and KB-Reducer [12,13] carry out only verification. Beyond their limitation to verification, these systems have additional weaknesses. RCP is limited to identification of static problems at the rule level, and cannot identify problems that result along longer reasoning chains. The CHECK system has better complexity than the RCP, but can be used only for systems developed using LES, the Lockheed expert systems development environment. ESC is very efficient if there are no conflicts or redundancies in V. Barr / Knowledge-Based Systems 12 (1999) 27–3528 Fig. 1. Rule-base evaluation with coverage analysis. Fig. 2. Evaluation with coverage analysis and data re-sampling. the rule-base, but can require exponential time if such problems exist. KB-Reducer is a verification tool, which operates on an implied network of rules. It has the advantage of checking a rule-base for inconsistency and redundancy over inference chains, and not just pairs of rules. A number of dynamic analysis tools have also been developed, such as TEIRESIAS [14], and SEEK2 [15]. TEIRESIAS aids in debugging and knowledge acquisition by allowing the alternation, deletion or addition of rules in order to fix errors in the rule-base that led to incorrect conclusions. However, this process requires that the system tester is sufficiently expert in the problem domain to identify errors in the reasoning the system used to reach a conclu- sion. SEEK2 is an automated rule-base refinement tool which tries out various rule refinements based on the system’s performance on known cases. However, the qual- ity of the refinements produced will be determined by the breadth of the test cases used. SEEK2 does not judge how well the test cases cover the range and domain of the system being evaluated. 2.1. Causal-associational network Graph-based methods for rule-base analysis involve the creation of a graph representation of the rules. An early example of such a representation (though not specifically used for V & V) is the causal-associational network (CASNET) for glaucoma diagnosis [16]. While our approach is similar to the approach used in CASNET, there are a number of differences, which stem from the ways the two methods use the graph structure. In CASNET the network serves as a direct representation of the knowl- edge, with interior nodes representing intermediate stages of disease progression. In our approach the graph is a direct representation of the knowledge base and intermediate nodes represent intermediate hypotheses in a logical sense, but may have no particular meaning relative to the problem domain unless the system designer built that into the rules. 2.2. KB-Reducer As mentioned earlier, KB-Reducer [12,13] uses the implied network of rules to do the rule-base analysis. The process of knowledge base reduction involves calculation of all possible logically independent and minimal sets of inputs under which the knowledge base will conclude each asser- tion (each class). In order for the reduction process to work, the rules of the knowledge base must form an acyclic network under the depends-on relation, defined in Ref. [12]. If the network is acyclic, then KB-Reducer proceeds to label each hypothesis H with the set of environments that lead to the assertion of H, where an environment is itself a set of findings. KB-Reducer carries out the labeling process on the rules in an order such that no rule is processed before any rules on which it depends. As each rule is processed, KB-Reducer updates the partial label for the hypotheses the rule asserts, and checks for redundancy and contradiction. 2.3. Completeness Verifier COVER (COmpleteness VERifier) [17–19] is another approach which combines both a functional and structural analysis of the rule-base. COVER carries out seven verifica- tion checks: redundancy, conflict, subsumption, unsatisfi- able conditions, dead-end rules, circularity and missing rules. The rules must either be written in or converted to a language based on first-order logic, and COVER must be given the set of final hypotheses (classes), as well as infor- mation about any semantic constraints. A primary difference between COVER and the work described here is in the granularity of the graph constructed. In COVER the nodes of the graph represent rules, and the edges represent dependencies or relations between rules, while in our representation each rule is itself represented by a small sub-graph, which allows us to carry out a more detailed analysis of which inference relations have and have not been exercised by the test data. 2.4. Pr/T nets The graph representation we are using is closer in some respects to the Pr/T net representation of a rule-base than it is to any of the other graph-based methods. In the majority of the graph-based methods a rule from the rule-base is equivalent to a node in the graph, while in our graph nodes correspond to individual findings or hypotheses, not to entire rules. Similarly, in the Pr/T net representation [20] the level of details is the findings and hypotheses, not the rules. However, only static analysis of the rule base is carried out using the Pr/T net representation. 2.5. Path Hunter and Path Tracer The VV & T approach based on the execution path model, incorporated in Path Hunter and Path Tracer [21,22], shares the fundamental premise upon which our work is based: functional validation may show that the system performs well on the test cases, but there may still be problems in portions of the rule base that were never exercised during testing. The goal of Path Hunter/Path Tracer is the selection of a set of test cases that exercise the structural components of the rule-base as exhaustively as possible. This involves firing all rules, and also firing every “causal sequence” of rules. The model used to identify all possible dynamic causal rule firing sequences is the rule execution path (equivalent to a sub-DAG in our representa- tion). Path Hunter uses structural path analysis to detect poten- tial interactions between rules in a rule-based and to identify problems within the rule-base, essentially using a path enumeration step. The complexity is controlled by precom- puting the logical completion for each subproblem and by V. Barr / Knowledge-Based Systems 12 (1999) 27–35 29 the use of equivalence classes of rules, formed by collecting redundant rules together into one class, which reduces the number of paths that Path Hunter must generate. (In our approach the step of explicit identification of redundant and ambiguous rules is unnecessary, as they will be identi- fied through the graph construction process. Therefore, while there may be redundant rules in the rule-base, there will be no duplication of those rules within our graph repre- sentation.) Path Tracer is a tool for structural rule-base testing, using the paths generated by Path Hunter, in conjunction with traces of dynamic rule firings, to determine how extensively the possible execution paths are covered by the test data. While there are a number of similarities in the premises which underlay our approach and Path Hunter/Path Tracer, there are also significant differences between the two approaches. First, the graph we construct models the rule- base at the level of findings and hypotheses, rather than at the rule level as is carried out in the Path Hunter representa- tion. While this may make the graph somewhat larger, it allows us to carry out both V & V with one representation. This is in contrast to COVER, which requires two represen- tations, the first-order logic translation of the rules and the dependency graph, in order to carry out verification alone. Second, the methods by which rule-base coverage are determined or measured are quite different. Path Tracer does its assessment of path coverage based on the number of causal dependencies observed in the trace file after the test data is run, using a number of strategies to map concrete paths observed at run time to the abstract paths generated by Path Hunter. In our approach, as the effect of concrete firings is indicated directly in the graph representation of the rule-base, we can determine rule-base coverage directly from the graph after the test data is run. We determine the extent of rule-base coverage by considering whether the state of the graph after the test data is run satisfies the four rule-base coverage measures. 2.6. Logical path graph model The logical path graph (LPG) model [23,24] is based on program control flow analysis. It attempts to apply cyclo- matic complexity and basis path testing to the rule-base environment to find a graphical representation of rule- bases which could then be used to determine rule-base complexity and determine a set of paths through the rule- base that, when executed, would adequately test the rules and their interactions. The logical path graph is a directed graph in which the nodes represent individual rules and the edges are deter- mined by logical paths through the rule base. The goal of Kiper’s work [23,24] is to use the LPG to determine a set of paths through the rule-base such that traversal of those paths during system testing represents an adequate test of the rules and their interactions. However, there are certain problems that arise in the use of logical paths. If there are multiple edges entering a node in the graph, they can be interpreted either as an AND or an OR relation. In order to avoid the possibility of OR edges the node must be replicated, in effect creating in the LPG the kind of redundancy which we usually try to remove from rule-bases. This results in the possibility of a single rule being represented by multiple nodes. If both AND and OR edges are to be allowed then the person evaluating the LPG must know what type each edge is. In addition, because of the possibility of multiple nodes representing a single rule, each node has to be labeled not just with the rule number but also with the condition set, the set of all conditions asserted by nodes on the path leading to the node. There are significant differences between the LPG and our approach. The graph structure we propose is based on findings and hypotheses and directly models the logical relations within rule antecedents, whereas the structure used for the LPG is built at the rule level. This difference in the graph construction allows us to avoid putting identi- fying information on edges or replicate rule representations, as is the case in the LPG. 3. Testing with rule-base coverage measures The first step in rule-base testing with coverage measures is to build a graph representation of the rule-base. Our method uses a directed acyclic graph (DAG) representation. We assume a generic propositional rule-base language [3] into which other rule-base languages can be translated. During construction of the DAG, pairwise redundant rules, pairwise simple contradictory rules and potential contradictions (ambiguities) are identified. After DAG construction is complete, static analysis (verification) of the rule-base reports dangling conditions (an antecedent component that is not defined as a finding and is not found as the consequent of another rule), useless conclu- sions, and cycles in the rule-base. At this point the rule- based could be modified to eliminate or correct any static problems. The static analysis phase is followed by dynamic analysis of the rule-base using test cases. As test cases are processed, one or more of several rule-base coverage measures (RBCMs) can be reviewed in order to determine the quality of the test data supplied thus far. Additional information about the rule-base and its testing can also be used by the system tester to guide the selection of future test data. The tester would start by providing sufficient test data to satisfy the simplest functional measure (conclude each class of the system) and proceed to the more difficult structural measures. Finally, if the user is not able to provide sufficient data to attain the desired degree of rule-base coverage (according to the selected criterion), the user can use the DAG representation to synthesize data, which can then be reviewed by an expert to determine if the data represents a valid case in the problem domain. V. Barr / Knowledge-Based Systems 12 (1999) 27–3530 This testing approach, described more fully later, has been implemented in the TRUBAC tool (Testing with RUle-BAse Coverage) [3,25]. 3.1. Rule-base representation In order to evaluate the rule-base coverage, the rule-base must be in a form which allows identification of sections which have and have not been covered by the test data. Our representation is based on the AND/OR graph implicit in the rule base [26]. The DAG has a source node, corresponding to working memory, and a sink node, corresponding to success in reaching one of the classes (diagnoses or goals) of the system. Interior nodes are sub-class nodes (SUBs), representing intermediate hypotheses, and operator nodes, representing the allowable operators AND, OR, and NOFM1 nodes. These operator nodes represent the fact that the conjunction and/or disjunction of multiple components of an antecedent must be true in order for the conclusion of a rule to be entered into working memory. There are edges from the source to each finding and from each class to the sink. The antecedent of each rule is represented by a subgraph, which connects findings and sub-class nodes to operators as indicated by the antecedent. Each antecedent- consequent connection represented by a rule is also repre- sented by an edge from the subgraph for the antecedent to the node for the consequent: For example, the representation of the two rules If P1 and P2 then R1 If R1 and R2 then Q is shown in Fig. 3. P1, P2 and R2 could be findings or sub- class nodes, while R1 and Q are either class or sub-class nodes. The complete graph of a rule-base is constructed by linking together the individual structures for successive rules. Using this framework we can easily represent rule- bases, including those in which the certainty factors are hard coded within the consequent definitions (rather than computed during the inference process) such as the rules in Fig. 4. 3.2. Rule-based coverage measures Using the DAG structure described earlier, we define execution paths in a rule-based system. Each reasoning chain through the rule-base corresponds to a sub-DAG, which includes all nodes and edges in the DAG correspond- ing to the rules fired during a particular chain of inferences. This include: source node; sink node; all nodes correspond- ing to the findings involved; the node for the concluded class; all nodes corresponding to antecedent components, operators, and rule consequents; and all edges involved in antecedent –consequent links formed by the rule connec- tions used in the reasoning chain. An individual rule firing involves all edges and nodes in the graph that corresponds to that rule. An execution path is, therefore, the sub-DAG that corresponds to all the rules fired along a particular reasoning chain executed due to a specific set of findings. Ideally, in testing a rule-base, we would like to provide sufficient test data to cause every possible execution path of the DAG to be traversed. This corresponds to firing all rules in every combination possible. As this is usually not reasonable in the testing environment, we propose the following hierarchy of rule-base coverage measures (RBCMs) in order to guide the selection of test data and give an objective measure of how well a test suite has covered the rule-base. Each-class :Satisfied if the test data causes traversal of one execution path to each class of the system. This is equivalent to providing, for each class, one test case that concludes that class. This is a very minimal coverage measure, which should be satisfied by all expert systems developers, regardless of their overall testing strategy. Each-hypoth: Satisfied if the test data causes traversal of execution paths such that each sub-class is reached, as well as each class. This shows that each sub-class is actu- ally reachable from the source and appears to be a rele- vant part of the system. The set of test data which satisfies this coverage measure is a superset of that which satisfies Each-class. Each-class-every-sub: There may be many execution paths that connect each sub-class to each class (and no execution paths for some sub-class to class combina- tions). This coverage measure is satisfied if, for each sub-class to class combination connected by some execu- tion path, at least one execution path, which includes the combination is executed. This coverage measure is stron- ger than Each-hypoth. Each-class-every-finding: There may be many execution paths that connect each finding to each class (and no execution paths for some finding-class combinations). This coverage measure is satisfied if, for each finding- class combination connected by some execution path, at least one execution path which includes the combination V. Barr / Knowledge-Based Systems 12 (1999) 27–35 31 Fig. 3. DAG representation of two related rules. Fig. 4. Rules with hard coded certainty factors. 1 NOFM nodes represent the construction “if N of the following M things are true, then…”, which is a feature of EXPERT [27] rule-bases. is executed. This coverage measure is stronger than Each- class but is incomparable to Each-hypoth as Each-hypoth can be satisfied without complete traversal of any execu- tion paths for some finding-class combinations. All-edges: This RBCM is satisfied if the data causes traversal of a set of execution paths such that every infer- ence relationship (every edge in the graph) is utilized along some inference path. While this will not guarantee that all rules will be used in every combination possible, it will guarantee that every rule is used in all possible ways along some execution path. In a rule-base with no NOFM nodes, the data that satisfies this RBCM will be a superset of the data necessary to satisfy Each-class-every-finding and Each-class-every-sub. In typical usage the tester will run a number of test cases and then query TRUBAC to determine whether each RBCM has been satisfied. If a coverage measure is not satisfied then TRUBAC will show what relationships remain to be covered in order to satisfy the RBCM. We assume the existence of an oracle that determines if the result given by the rule-base for a test case is correct 2 or not. In this work we do not consider the issue of specifica- tions and how we determine if an answer is correct or not. If an answer is wrong, then the execution path that led to its conclusion is suspect and must be studied for errors. If the answer is right then the path is only of interest to the extent that it helps satisfy the coverage measure(s) chosen by the tester. In a very complex rule-base, or one for which there is very little test data, the RBCMs help the tester determine ways in which the data is deficient in testing portions of the rule-base. The coverage measures provide the user with information that can lead to the acquisition or development of additional test cases, or will indicate errors in the rule- base. This approach can also address a difficulty often faced by those who test expert systems, which is a paucity of test data. An additional feature of this approach to rule-base testing can help the user gain insight into the correctness of the system even when little test data is available. In addition to evaluating the five coverage measures, the DAG framework can be used to generate test data, which would lead to traversal of any execution paths not exercised by the test data. This synthesized data can then be shown to one or more experts in order to determine if each synthe- sized test case, which is a reflection of the logic embodied within a section of rule-base (corresponding to an execution path), makes sense in the context of the problem domain which the rule-base was designed to handle. If the expert does not agree that the data represents a plausible case in the problem domain then the section of the rule-base repre- sented by that section of the graph must be reviewed for errors. 4. Applications of coverage analysis In addition to providing information about the testing process itself, the coverage analysis can be used to enhance testing and facilitate other kinds of rule-base analysis, as described below. 4.1. Heuristics for test data selection There may be a large redundant pool of available test data, from which a subset of cases must be selected for the test suite. Running all available cases can be infeasible due to the length of time it may require. A random selection of test cases may give statistical confirmation that the system works properly for the tested situations, but a poor selection of cases will lead to an incomplete test set which then leads to incomplete coverage of the rule-base. If a test set which we know is incomplete leads to complete cover- age of the rule-base then the rule-base is incomplete and is not capable of handling precisely those types of cases that are absent from the test suite. If an incomplete test set leads to incomplete coverage, we can use the coverage informa- tion as the foundation for a set of heuristics for test data selection from the available population, in order to construct a test set that will maximize rule-base coverage. To date we have restricted our work to classification systems in which each goal of the system is a class and we consider each intermediate hypothesis to be a sub- class. Assuming that we have a degree of meta-knowledge about the make-up of individual test cases (e.g. what facts or intermediate hypotheses are involved in each case), a set of heuristics for data selection is: 1. For each class, select a test case which concludes only that class. 2. For each class not yet tested, select a test case which will conclude it (and additional classes). At this point Each- class will be satisfied. 3. Select test cases which will conclude unused sub-classes. This satisfies Each-hypoth. 4. Select test cases which will cover sub-class to class rela- tions, and direct finding to class relations for findings which do not lead to an intermediate sub-class. This satisfies Each-class-every-sub. 5. Select test cases which will cover finding to class rela- tions for findings which represent alternative ways to conclude sub-classes. That is, while a sub-class to class relation may have been covered, there may be multiple ways to conclude the sub-class. This satisfies Each-class- every-finding. In [3,29] we show that the use of coverage information, V. Barr / Knowledge-Based Systems 12 (1999) 27–3532 2 For a given test case the experts may not agree on one answer, in which case the system will be considered correct if its answer is in the set of answers agreed on by the experts as possible for that case. There are a number of issues that are raised by the desire to have a “gold standard” [28] against which to measure the expert system’s answers. For example, if the system’s answer agrees with that of the expert, should we consider the system to be correct, even if we subsequently learn that both were wrong. along with meta-knowledge about the pool of available cases, can successfully be used to select the test cases in a way that contributes to a useful test of the rule-base. This results in greater assurance that the system has been tested and works correctly for the situations that we expect to encounter most often, as well as clearly identifying those parts of the system that still require more examination, test- ing, or refinement. 4.2. Class dependence Another aspect of rule-base analysis or evaluation is the identification of class dependencies, in which the rules that lead to the conclusion of one class are also highly involved in the conclusion of another class. If two classes, C1 and C2, are dependent, and we have a large number of test cases that will be classified as C1, it may be that some of those cases will also be classified as C2, although with differing certainty factors. Because of this, testing for one class may help achieve coverage for the other. Further, if C1 and C2 are dependent classes and we change rules for C1 then we should rerun test cases, which conclude C1, and rerun test cases, which conclude C2. This will verify that changing rules for C1 did not inadvertently affect the system’s ability to properly classify C2 as well. To determine if two classes have rules in common we look at sharing or overlap among the sets of sub-classes that can lead to a class. A high degree of overlap among sub-classes implies overlap among the rules and a degree of dependency between those classes. This information can be obtained immediately from data collected while the DAG representation is built. For each class a list of all the sub-classes which can help to conclude the class is formed, and then these lists are compared for pairs of classes. Table 1 shows the overlap figures for a small prototype of the AI/RHEUM system for rheumatology diagnosis [4]. These figures indicate that 57% of the sub- classes, which lead to RA (rheumatoid arthritis) lead to MCTD (mixed connective tissue disease), while 36% of the sub-classes, which can lead to MCTD also lead to RA. There is asymmetry in these figure’s results because the absolute number of sub-classes which can lead to the classes is different. These figures indicate that modifications made to the rules for RA would possibly also affect the perfor- mance of the system on cases that should be classified as MCTD, while there would be a lesser affect on the system’s classification of cases as RA if the rules for MCTD were modified. These results are consistent with those found in [30], which uses Monte-Carlo simulation-based techniques to carry out rule-base evaluation. However, in our approach we obtain this information as a by-product of DAG construction, without the overhead of rule-base execution. 4.3. Rule-base pruning Once a rule-base has been constructed, it is possible that not all the rules are necessary for the rule-base to perform correctly. For example, during incremental development some early rules may be supplanted by rules added later. If the system can be pruned by removing rules or compo- nents within rules, and the performance on test cases is not affected, then it is possible that there were unnecessary rules or that the test cases are not adequate to evaluate the entire rule-base [30]. Further, we expect that a smaller rule-base will run more efficiently. Coverage information can focus the pruning steps on sections of the rule-base which have not been executed by the test data. If a section of the rule-base is never executed during test suite execution then either there are unnecessary rules, which can be pruned, or the test suite is not sufficiently rich. In the latter case, additional test cases are necessary to cover the un-executed section of the rule- base. In general the developer and/or the expert will decide whether the proper approach is to prune or to add test cases. If the test suite is truly representative of cases that will be found in the application environment, and the rule-base performs correctly on the test set, then it is likely that the uncovered portion of the rule-base is in fact unnecessary and can be pruned. If the test set is complete and the rule-base performs incorrectly then the uncovered sections of the rule- base are candidates for rule refinement. The coverage measures provide information about why the rules were never used, based on findings and sub-classes which appear in rule antecedents but are not present in any test cases. Removal of these antecedent components from the rules may generalize them sufficiently that they will correctly handle some of the test cases and will be covered by an existing portion of the test suite. In experiments run on the RHEUM rule-base, the first phase of pruning was based on TRUBAC’s static analysis results. We were able to prune 5 rules out of 76, as well as eliminate 85 findings and 3 classes. This significantly reduced the overall size of the graph over which coverage was evaluated from 333 nodes to 241 nodes [3]. Two addi- tional pruning iterations, based on the coverage data and the assumption that the test set was complete, eliminated 20 rules, 6 findings, and 15 components of rule antecedents. This resulted in an overall 34% reduction in the number of rules and a 40% reduction in the graph size. V. Barr / Knowledge-Based Systems 12 (1999) 27–35 33 Table 1 Overlap of sub-classes for class pairs PM PSS SLE MCTD RA PM 0 0 22.22 22.22 PSS 0 16.67 16.67 16.67 SLE 0 9.09 0 0 MCTD 18.18 9.09 0 36.36 RA 28.57 14.29 0 57.14 4.4. A metric for rule-based systems The coverage measures can be very useful for the testing process. However, it would also be useful to have a way to predict or measure the complexity of the system and of the testing process before testing is begun. We would also like to be able to compare the complexity of different rule-bases, particularly if there are multiple rule-bases that handle problems within a common domain. This raises the issue of whether there is a reasonable analog to the control flow graph and, more importantly, a complexity metric for rule- bases which would reflect the relationship between rule- base structure and system complexity. The graph representation proposed here serves as a suita- ble foundation for such a complexity metric for rule-based systems. The graph imposes no particular execution order on the rules, and it represents all logical relations that are inherent within the rule-base. However, graph-based metrics such as McCabe’s cyclomatic complexity metric cannot adequately determine the number of execution paths in a rule-base. The actual number of execution paths is based on the logical relationships in the rule-base, using the following mechanism: • For each finding, we assume there is only one path to it. • For each OR or SUB node, consider the parent nodes. The number of paths to the OR or SUB node is the sum of the paths to the parent nodes. • For each AND node, compute the product of the number of paths that lead to each parent of the AND. • Class nodes are treated like OR or SUB nodes. The total number of paths through the rule-base is computed by adding up the number of paths to each class node. In Fig. 5 there are a total of five paths to G based on three paths to the SUB node and two paths to the OR node. This metric can be computed fairly easily by visiting the nodes in topological order and saving in each node the number of paths to that node. This execution path metric can serve a number of purposes in rule-base development and analysis. The total number of execution paths represents the maximum number of test cases needed for complete coverage of the rule-base according to the strongest rule-base coverage measure (All- edges). However, usually the actual number of data sets needed will be less than the number of execution paths, as often, particularly in diagnosis systems, one test set may cover a number of execution paths to different diagnoses. 5. Conclusions and future work This work shows that there are numerous uses of rule- base coverage data in the testing process. Rule-base perfor- mance evaluation can be misleading unless care is taken to identify problems with both the test data and the rule-base. Both the test data and the rule-base can be improved by using information about the extent to which the test data has covered the rule-base under test. This work can be extended in a number of directions. Quantitative performance prediction can be computed based on performance of the system on test cases, a measure of how well the test data covers the rule-base, and a measure of the degree to which the test set is representative of the population for which the system is intended. A second area of extension is for systems which have dynamic computa- tion of certainty factors, which requires modification of the rule-base coverage measures [3] as well as changes to the implementation and the data selection heuristics. It would also be useful to extend the approach to systems that are not acyclic and prepositional. This would greatly increase the practicality of this method for testing rule-bases in a variety of application areas. Another area which should be studied in the future is that of the relationship between rule-base complexity and the difficulty of carrying out the testing process. While intui- tively it may seem that a more complex rule-base should undergo more complex and stringent testing, it may in fact be the impact of an incorrect result that should determine the quality of the testing. For example, in a medical diagnosis system if an incorrect result could lead to not providing treatment to an ill patient then all steps possible should be taken to ensure that the system works correctly, no matter how complex or simple the rule-base is. Finally, it may be possible to extend this approach to analyze the Bayesian belief networks. The probabilistic relationships between nodes of the network, with informa- tion flow which is bi-directional along the arcs, and nodes which may be dependent in some contexts and independent in others, significantly complicates this task. References [1] P. Jackson, Introduction to Expert Systems, 2, Addison-Wesley, Reading, MA, 1990. [2] R. O’Keefe, D.E. O’Leary, Expert system verification and validation: a survey and tutorial, Artificial Intelligence Review 7 (1993) 3–42. [3] V. Barr, Applications of rule-base coverage measures to expert system evaluation, PhD thesis, Rutgers University, 1996. [4] L.C. Kingsland, The evaluation of medical expert systems: experi- ences with the AI/RHEUM knowledge-based consultant system in rheumatology, Proceedings of the Ninth Annual Symposium on Computer Applications in Medical Care, Washington DC, 1985 pp. 292–295. [5] W.R. Adrion, M.A. Branstad, J.C. Cherniavsky, Validation, verifica- tion, and testing of computer software, ACM Computing Surveys 14 (2) (1982) 159–192. V. Barr / Knowledge-Based Systems 12 (1999) 27–3534 Fig. 5. Graph of rule-base with OR and SUB. [6] P. Frankl, E. Weyuker, A data flow testing tool, Proceedings of IEEE Softfair II, San Francisco, December 1985. [7] S. Rapps, E. Weyuker, Selecting software test data using data flow information, IEEE Transactions on Software Engineering 11 (4) (1985) 367–375. [8] M. Suwa, S.C. Scott, E.H. Shortliffe, An approach to verifying completeness and consistency in rule-based expert system, AI Maga- zine 3 (4) (1982) 16–21. [9] T.A. Nguyen, W.A. Perkins, T.J. Laffey, D. Pecora, Checking an expert systems knowledge base for consistency and completeness, Proceedings of the Ninth IJCAI, Menlo Park, CA, 1985, pp. 374–378. [10] T.A. Nguyen, W.A. Perkins, T.J. Laffey, D. Pecora, Knowledge base verification, AI Magazine 8 (2) (1987) 69–75. [11] B.J. Cragun, H.J. Steudel, A decision-table-based processor for checking completeness and consistency in rule-based expert systems, International Journal of Man-Machine Studies 26 (1987) 633–648. [12] A Ginsberg, A new approach to checking knowledge bases for incon- sistency and redundancy, Proceedings of the Third Annual Expert Systems in Government Conference, Washington DC, 1987, pp. 102–111. [13] A. Ginsburg, Automatic Refinement of Expert System Knowledge Bases, Pitman, London, 1988. [14] R. Davis, Interactive transfer of expertise, in: B.G. Buchanan, E.H. Shortliffe (Eds.), Rule-Based Expert Systems, Addison-Wesley, Reading, MA, 1984, pp. 171. [15] A. Ginsberg, S. Weiss, P. Politakis, SEEK2: a generalized approach to automatic knowledge base refinement, Proceedings of IJCAI-85, 1985. [16] S.M. Weiss, C.A. Kulikowski, S. Amarel, A. Safir, A model-based method for computer-aided medical decision-making, Artificial Intel- ligence 11 (1978) 145–172. [17] A.D. Preece, Verification of rule-based expert systems in wide domains, Research and Development in Expert Systems VI, Proceed- ings of Expert Systems ’89, British Computer Society Specialist Group on Expert Systems, London, 1989, pp. 66–77. [18] A.D. Preece, R. Shinghal, Foundation and application of knowledge base verification, International Journal of Intelligent Systems 9 (1994) 683–701. [19] P.D. Grogono, A.D. Preece, R. Shinghal, C.Y. Suen, A review of expert systems evaluation techniques, Workshop on Validation and Verification of Knowledge-Based Systems, 11th National Conference on Artificial Intelligence, Washington DC, 1993 pp. 120–125. [20] D. Zhang, D. Nguyen, A technique for knowledge base verification, IEEE International Workshop on Tools for Artificial Intelligence, 1989, pp. 399–406. [21] C. Grossner, A.D. Preece, P.G. Chander, T, Radhakrishnan, C.Y. Suen, Exploring the structure of rule based systems, Proceedings of the 11th National Conference on Artificial Intelligence, Washington DC, 1993, pp. 704–709. [22] A.D. Preece, C. Grossner, P.G. Chander, T. Radhakrishnan, Structural validation of expert systems using a formal model, Workshop on Validation and Verification of Knowledge-Based Systems, 11th National Conference on Artificial Intelligence, Washington DC, 1993, pp. 19–26,. [23] U. Gupta, J. Kiper, B. Ly, A. Preece, Developing criteria for compar- ing specific V and V tools (from First Winter Workshop on Verifica- tion and Validation of Knowledge-Based Systems), AAAI-92 Workshop on Verification and Validation of Knowledge-Based Systems, San Jose, CA, 1992. [24] J. Kiper, Structural testing of rule-based expert systems, ACM Trans- action on Software Engineering and Methodology 1 (2) (1992) 168– 187. [25] V. Barr, TRUBAC: a tool for testing expert systems with rule-base coverage measures, Proceedings of the 13th Annual Pacific Northwest Software Quality Conference, Portland, OR, 1995. [26] P. Meseguer, Structural and performance metrics for rule-based expert systems, Proceedings of the European Workshop on the Veri- fication and Validation of Knowledge Based Systems, pp. 165–178, Cambridge, England, 1991. [27] S.M. Weiss, K.B. Kern, C.A. Kulikowski, M. Uschold, A guide to the EXPERT consultation system. Technical Report CBM-TR-94, Department of Computer Science, Laboratory for Computer Science Research, Rutgers University, 1987. [28] B.G. Buchanan, E.H. Shortliffe, Rule-Based Expert Systems The problem of evaluation, Addison-Wesley, Reading, MA, 1985. [29] V. Barr, Rule-base coverage analysis applied to test case selection, Annals of Software Engineering, 1997. [30] N. Indurkhya, Monte-carlo simulation-based evaluation and refine- ment of rule-based systems, Technical Report DCS-TR-277, Depart- ment of Computer Science, Rutgers University, 1991. V. Barr / Knowledge-Based Systems 12 (1999) 27–35 35 
work_ho4p6lmanrfsljzmvdcbb52o4q	----	Microsoft Word - 2011_02_10_Serratosa.docx 1 A Probabilistic Integrated Object Recognition and Tracking Framework Francesc Serratosa a , René Alquézar b & Nicolás Amézquita a a Departament d’Enginyeria Informàtica i Matemàtiques, Universitat Rovira i Virgili, Av. Països Catalans 26, 43007 Tarragona, Spain b Institut de Robòtica i Informàtica Industrial, CSIC-UPC, Llorens Artigas 4-6, 08028 Barcelona, Spain Abstract. This paper describes a probabilistic integrated object recognition and tracking framework called PIORT, together with two specific methods derived from it, which are evaluated experimentally in several test video sequences. The first step in the proposed framework is a static recognition module that provides class probabilities for each pixel of the image from a set of local features. These probabilities are updated dynamically and supplied to a tracking decision module capable of handling full and partial occlusions. The two specific methods presented use RGB colour features and differ in the classifier implemented: one is a Bayesian method based on maximum likelihood and the other one is based on a neural network. The experimental results obtained have shown that, on one hand, the neural net based approach performs similarly and sometimes better than the Bayesian approach when they are integrated within the tracking framework. And on the other hand, our PIORT methods have achieved better results when compared to other published tracking methods in video sequences taken with a moving camera and including full and partial occlusions of the tracked object. Keywords: Object tracking; object recognition; occlusion; performance evaluation; probabilistic methods; video sequences; dynamic environments 2 1. Introduction One of the most general and challenging problems a mobile robot has to confront is to identify, locate and track objects that are common in its environment. To this end, object models have to be defined or learned in conjunction with some associated recognition and tracking procedures. There are several issues that have to be considered while dealing with object locating and tracking which deserve some previous discussion. The first important issue is to determine the type of object model to learn, which usually depends on the application environment. For instance, in [1], the target was an aerial vehicle. And in [2,3,4,5] the targets were people. In [6], they used specific parameters of the object to be tracked. In [7], they track hands using textures. In [8], they developed and implemented a real system for simultaneous localization and mapping (SLAM) algorithm for mobile robots based on an extended Kalman filter. It was applied to indoor environments and used stereo vision based on two web-cam. The system diverges from ours in that we like to track objects captured from the mobile- robot cameras, instead of localize the position of our robot. While tracking the object, for instance people walking in the street, the system could try to recognize person the through a face recognition system. There is a lot of literature related to this field [9]. Moreover, other systems not only indentify subjects but detect the mood of these subjects or detect specific pathologies [10]. Face identification or recognition is not the scope of this paper. Finally, other field related to object tracking is automatic hand-gesture recognition [11]. In this kind of systems, hands have to be tracked and the trajectory (position, speed, acceleration) has to be analyzed to conclude the meaning of this movement. In our case, we want a mobile robot equipped with a camera to locate and track general objects (people, other robots, balls, wastepaper bins …) in both indoor and outdoor environments. A useful object model should be relatively simple and easy to acquire from the result of image processing steps. For instance, the result of a colour image segmentation process, consisting of a set of regions or spots, characterized by simple features related to colour, may be a good starting point to learn the model [7, 12]. Although structured models like attributed graphs or skeletons can be synthesized for each object from several segmented images [13, 14], we have decided to investigate a much simpler approach in which the object is just represented as an unstructured set of pixels. Other methods detect some characteristic points of the object to be tracked [15]. At a learning phase, the most repeatable object keypoints for the specific object are learned. Another interesting work is [16], in which the algorithm search for different region tracks. These methods have been proven to have a good performance when there is low variability of the features of the object. Nevertheless, with deformable objects, it is difficult to extract some representative points. One of the main drawbacks of structural methods is that the segmented images can be quite different from one frame to the other, and therefore it is difficult to match the structure in the current frame with the previous ones. In [17], the model was specially designed to segment and track objects from video sequences that suffer from abrupt changes. The starting point of our approach is to accept these differences between segmented images and use a more rudimentary model in which the basic element is not the spot or region of the segmented image but its pixels. An example of structural 3 method was reported in [14], where the object model was based on the skeleton of the object obtained in the segmented images. Since the skeletons resulting from two almost equal images can be very different, the applicability of such approach is limited. The tracking step was performed in [14] by an extension of the Kalman filter in which the skeleton and other geometrical features were considered. Other options has been [18,19] where the model specifically incorporated the relation between position and time. Finally, other methods are based on keeping the information of the silhouette of the object to be tracked. In [5], the method is based on learning a dynamic and statistical model of the silhouette of the object. In our case, we cannot use this system since we assume that the deformation of the object to be tracked is not predictable. A second significant issue is to provide the tracking procedure with the capacity of determining occlusions and re-emergencies of tracked objects, i.e. occlusion handling. Over recent years, much research has been developed to solve the problem of object tracking under occlusions [4], because, in real-world tracking, a target being partly or entirely covered by other objects for an uncertain period of time is common. Occlusions pose two main challenges to object tracking systems. The first challenge is how to determine the beginning and the end of an occlusion. The second challenge is how to predict the location of the target during and at the end of the occlusion. Determining occlusion status is very hard for the trackers where the only knowledge available on the target is its initial appearance. When some parts of an occluder are similar to those of the target, the occluder and the target are mistaken. Various approaches that analyze occlusion situations have been proposed. The most common one is based on background subtraction [4, 19, 20]. Although this method is reliable, yet it only works with a fixed camera and a known background, which is not our case in mobile robotics. In [4], they used several cameras, and tracking and occlusion of people is solved by a multi-view approach. In [20], they achieve real-time tracking with small images. Evidence is gathered from all of the cameras into a synergistic framework. Other approaches are based on examining the measurement error for each pixel [22, 23]. The pixels that their measurement error exceeds a certain value are considered to be occluded. These methods are not very appropriate in outdoor scenarios, where the variability of the pixel values between adjacent frames may be high. A mixture of distributions is used in [24] to model the observed value of each pixel, where the occluded pixels are characterised by having an abrupt difference with respect to a uniform distribution. Contextual information is exploited in [25, 27]. These methods have better performance in terms of analysing occlusion situations but tracking errors are observed to frequently occur and propagate away. In addition, in the case of using these approaches in a mobile robot application, there is a need of knowing a priori the robot surroundings. Determining the re-emergence of the target and recapture its position after it is completely occluded for some time is the other main challenge. Setting a similarity threshold is one method, yet the optimal threshold value is difficult to determine. This problem is circumvented in [22], where the image region that matches the best with the template over a prefixed duration is assumed to be the reappearing target. In [23], an observation model and a velocity motion model were defined. The observation model was based on an adaptive appearance model, and the velocity motion model was derived using a first-order linear predictor. Both approaches are defined in the framework of particle filter, with provisions for handling occlusion. 4 In the scenarios where the motion of the target is not smooth neither predictable most of the aforementioned methods would fail. Recently, new object tracking methods that are robust to occlusion have been reported with very promising results [27, 28]. The method reported in [27] relies on background subtraction (it works only for static cameras) and a k-NN classifier to segment foreground regions into multiple objects using on-line samples of object’s appearance local features taken before the occlusion. The method described in [28] relies on an adaptive template matching but it only handles partial occlusions and the matching process seems to be computationally costly. A third relevant issue, which generally is not so mentioned, is to integrate the recognition and tracking steps in a common framework that helps to exploit some feedback between them. To the best of our knowledge there are few existing works that combine recognition and tracking in an integrated framework [29, 30]. Object recognition and tracking are usually performed sequentially and without any feedback from the tracking to the recognition step [14]. These tasks often are treated separately and/or sequentially on intermediate representations obtained by the segmentation and grouping algorithms [31-33]. Sometimes, they are applied in a reverse order, with a first tracking module supplying inputs to the recognition module, as, for instance, in gesture recognition [34]. An integrated framework for tracking and recognising faces was presented in [30]. Conventional video-based face recognition systems are usually embodied with two independent components: the recognition and the tracking module. In contrast, an architecture was defined in [30] that tightly couples these two components within a single framework. The complex and nonlinear appearance manifold of each registered person was partitioned into a collection of sub-manifolds where each models the face appearances of the person in nearby poses. The sub-manifolds were approximated by a low-dimensional linear subspace computed by PCAs. Finally, Artificial Intelligence was applied to tracking objects in [35]. This paper describes thoroughly and in detail the current state of a probabilistic integrated object recognition and tracking (PIORT) methodology that we have developed in the latest years, as well as two particular methods derived from it. It also presents a collection of experimental results in test video sequences obtained by PIORT methods and alternative tracking methods. Previous stages in the development of PIORT, together with preliminary results, have been partially reported elsewhere [36- 39]. In the experimental evaluation carried out, PIORT methods have been compared to six state-of-the-art tracking methods of which we were able to get and apply their program codes to the test video sequences: - Template Match by Correlation (TMC) [40]; - Basic Meanshift (BM) [41]; - Histogram Ratio Shift (HRS) [42]; - Variance Ratio Feature Shift (VRFS) [42]; - Peak Difference Feature Shift (PDFS) [42]; and - Graph-Cut Based Tracker (GCBT) [43, 44]. 5 Their codes were downloaded from the VIVID tracking evaluation web site www.vividevaluation.ri.cmu.edu, which unfortunately seems not to be accessible anymore. We briefly summarise these methods next. In the TMC method [40], the features of the target object are represented by histograms. These histograms are regularised by an isotropic kernel which produce spatially smooth functions suitable for gradient-based optimisation. The metric used to compare these functions is based on the Bhattacharyya distance and the optimisation is performed by the mean-shift procedure. In [41] a general non-parametric framework is presented for the analysis of a multimodal feature space and to separate clusters. The mean-shift procedure (localisation of stationary points in the distributions) is used to obtain the clusters. Throughout this framework, a segmentation application is described. In [42] three different tracking methods are presented. They are based on the hypothesis that the best feature values to track an object are the ones that best discriminate between the object and the present background. Therefore, with several sample densities of the object and also of the background, the system computes the separability of both classes and obtains new features. The feature evaluation mechanism is embedded in a mean- shift tracking system that adaptively selects the top-ranked discriminative features for tracking. In the first method, Histogram Ratio Shift (HRS), the weights applied to each feature are dynamically updated depending on the histograms of the target and also of the background. In the second one, Variance Ratio Feature Shift (VRFS), the ratio between the variance of the target and the surrounding background is computed and considered for selecting the features. Finally, the Peak Difference Feature Shift (PDFS) softens the histogram of the features by a Gaussian kernel; moreover, it considers possible distracter objects near the target and dynamically changes the feature selection. And finally, in [43, 44], a method for direct detection and segmentation of foreground moving objects is presented called Graph-Cut Based Tracker (GCBT). The method first obtains several groups of pixels with similar motion and photometric features. The mean-shift procedure is used to validate the motion and bandwidth. And then, the system segments the objects based on a MAP framework. Our PIORT methodology is based on the iterative and adaptive processing of consecutive frames by a system that integrates recognition and tracking in a probabilistic framework. The system uses object recognition results provided by a classifier, e.g. a Bayesian classifier or a neural net, which are computed from colour features of image regions for each frame. The location of tracked objects is represented through probability images that are updated dynamically using both recognition and tracking results. The tracking procedure is capable of handling quite long occlusions. In particular, object occlusion is detected automatically and the prediction of the object’s apparent motion and size takes into account the cases of occlusion entering, full occlusion and occlusion exiting. In contrast with [27], our tracking method does not rely on background subtraction and a fixed camera and, to the contrary of [28], it can cope with complete occlusion and it does not involve any template to match and update. In our approach, the following assumptions are made: i) target objects may be distinguished from other objects and the background based on 6 colour features of their appearance, though these features may experiment slight variations during the image sequence; in fact, this is a requirement of the classifiers we currently use for static recognition and could be relaxed or changed if the classifier in this module were replaced or used a different set of object’s appearance features; ii) target object’s shape, apparent motion and apparent size can all vary smoothly between consecutive frames (non-rigid deformable objects are thus allowed); we think this is not a strong assumption if a typical video acquisition rate is used, as large changes in shape, motion and size are allowed for the whole sequence; iii) image sequences may be obtained either from a fixed or a slow moving camera (this is also quite realistic for most applications in practice); iv) target objects may be occluded during some frames, but their motion does not change abruptly during occlusion; this last assumption is certainly stronger and may fail in some cases, but it is caused by the need of predicting an approximate position of the object during occlusion based on its previous trajectory. The rest of the paper is organized as follows. A formal description of our probabilistic framework for object recognition and tracking is given in Section 2. As shown in Fig. 1, the system involves three modules: static recognition, dynamic recognition and tracking decision modules. The methods used for the static recognition module are specified in Section 3. The dynamic recognition module is explained in Section 4. The tracking decision module is described in detail in Section 5. The experimental results are presented in Section 6. Finally, conclusions and future work are discussed in Section 7. 7 2. A probabilistic framework for integrated object recognition and tracking Let us assume that we have a sequence of 2D color images I t (x,y) for t=1,…,L, and that there are a maximum of N objects of interest in the sequence of different types (associated with classes c=1,…,N, where N≥1), and that a special class c=N+1 is reserved for the background. Furthermore, let us assume that the initial position of each object is known and represented by N binary images, pc 0 (x,y), for c=1,…,N, where pc 0 (x,y)=1 means that the pixel (x,y) belongs to a region covered by an object of class c in the first image. If less than N objects are actually present, some of these images will be all-zero and they will not be processed further, so, without loss of generality, we consider in the sequel that N is the number of present objects to track. Hence, we would like to obtain N sequences of binary images Tc t (x,y), for c=1,…,N, that mark the pixels belonging to each object in each image; these images are the desired output of the whole process and can also be regarded as the output of a tracking process for each object. We can initialize these tracking images (for t=0) from the given initial positions of each object, this is ),(),( 00 yxpyxT cc = (1) In our approach, we divide the system in three modules. The first one performs object recognition in the current frame (static recognition) and stores the results in the form of probability images (one probability image per class), that represent for each pixel the probabilities of belonging to each one of the objects of interest or to the background, according only to the information in the current frame. This can be achieved by using a classifier that has been trained previously to classify image regions of the same objects using a different but similar sequence of images, where the objects have been segmented and labeled. Hence, we assume that the classifier is now able to produce a sequence of class probability images Qc t (x,y) for t=1,…,L and c=1,…,N+1, where the value Qc t (x,y) represents the estimated probability that the pixel (x,y) of the image I t (x,y) belongs to the class c, which has been computed taking into account a local feature vector (see Section 3). In general, the probability images Qc t (x,y) can be regarded as the output of a static recognition module defined by some function r on the current image: ( ) ),(),( yxIryxQ tt c = (2) In the second module (dynamic recognition), the results of the first module are used to update a second set of probability images, pc , with a meaning similar to that of Qc but now taking into account as well both the recognition and tracking results in the previous frames through a dynamic iterative rule. More precisely, we need to store and update N+1 probability images pc t (x,y), for c=1,…,N+1, where the value pc t (x,y) represents the probability that the pixel (x,y) in time t belongs to an object of class c (for c=1,…,N) or to the background (for c=N+1). In general, these dynamic probabilities should be computed as a certain function f of the same probabilities in the previous step, the class probabilities given by the classifier for the current step (which have been obtained from the actual measurements) and the tracking images resulting from the previous step: ( ) ),(),,(),,(),( 11 yxTyxQyxpfyxp tttt c −− = (3) 8 The update function f used in our system is described in Section 4, which incorporates some additional arguments coming from the tracking module to adapt its parameters. Finally, in the third module (tracking decision), tracking binary images are determined for each object from the current dynamic recognition probabilities, the previous tracking image of the same object and some other data, which contribute to provide a prediction of the object’s apparent motion in terms of translation and scale changes as well as to handle the problem of object occlusion. Formally, the tracking images Tc t (x,y) for the objects (1≤c≤N) can be calculated dynamically using the pixels probabilities p t (x,y) according to some decision function d: ( ) ),(),,(),( 1 yxTyxpdyxT t c tt c − = (4) in which some additional arguments and results may be required (see (12) and Section 5 for a detailed description of the tracking decision module). 3. Static recognition module In our PIORT (Probabilistic Integrated Object Recognition and Tracking) framework, the static recognition module is based on the use of a classifier that is trained from examples and provides posterior class probabilities for each pixel from a set of local features. The local features to be used may be chosen in many different ways. A possible approach consists of first segmenting the given input image I t (x,y) in homogeneous regions (or spots) and computing some features for each region that are afterwards shared by all its constituent pixels. Hence, the class probabilities Qc t (x,y) are actually computed by the classifier once for each spot in the segmented image and then replicated for all the pixels in the spot. For instance, RGB color averages can be extracted for each spot after color segmentation and used as feature vector v(x,y) for a classifier. In the next two subsections we present two specific classifiers that have been implemented and tested within the PIORT framework using this type of information. Figure 1: Block diagram of the dynamic object recognition and tracking process. 9 3.1 A simple Bayesian method based on maximum likelihood and background uniform conditional probability Let c be an identifier of a class (between 1 and N+1), let B denote the special class c=N+1 reserved for the background, let k be an identifier of an object (non-background) class between 1 and N, and let v represent the value of a feature vector. Bayes theorem establishes that the posterior class probabilities can be computed as ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )∑ = + == N k kPkvPBPBvP cPcvP vP cPcvP vcP 1 | | | | | (5) Our simple Bayesian method for static recognition is based on imposing the two following assumptions: a) equal priors: all classes, including B, will have the same prior probability, i.e. P(B)=1/(N+1) and P(K)=1/(N+1) for all k between 1 and N. b) a uniform conditional probability for the background class, i.e. P(v|B)=1/M, where M is the number of values (bins) in which the feature vector v is discretized. Note that the former assumption is that of a maximum likelihood classifier, whereas the latter assumes no knowledge about the background. After imposing these conditions, equation (5) turns into ( ) ( ) ( )∑ = + = N k kvP M cvP vcP 1 | 1 | | (6) and this gives the posterior class probabilities we assign to the static probability images, i.e. Qc t (x,y) = P(c | v(x,y)) for each pixel (x,y) and time t. It only remains to set a suitable M constant and to estimate the class conditional probabilities P(v | k) for all k between 1 and N (object classes). To this end, class histograms Hk are set up using the labeled training data and updated on-line afterwards using the tracking results in the test data. For constructing the histograms, let v(x,y) be the feature vector consisting of the original RGB values of a pixel (x,y) labeled as belonging to class k. We uniformly discretize each of the R, G and B channels in 16 levels, so that M =16×16×16= 4096. Let b be the bin in which v(x,y) is mapped by this discretization. To reduce discretization effects, a smoothing technique is applied when accumulating counts in the histogram as follows: ( ) ( ) ( ) ( ) bbbHbH bneighborsbHbH kk kk ofneighbor a is ' if1' : ' ))(#10( : += −+= (7) where the number of neighbors of b (using non-diagonal connectivity) varies from 3 to 6, depending on the position of b in the RGB space. Hence, the total count Ck of the histogram is increased by ten (instead of one) each time a pixel is counted and the 10 conditional probability is estimated as P(v | k) = Hk(b) / Ck where b is the bin corresponding to v. The above smoothing technique is also applied when updating the histogram from the tracking results; in that case the RGB value v(x,y) in the input image I t (x,y) of a pixel (x,y) is used to update the histogram Hk (and the associated count Ck) if and only if Tk t (x,y)=1. 3.2 A neural net based method In this method, a neural net classifier (a multilayer perceptron) is trained off-line from the labeled training data. The RGB color averages extracted for each spot after color segmentation are used as feature vector v(x,y) and supplied as input to the network in both training and test phases. To the contrary of the Bayesian method described previously, training data for the background class are also provided by selecting some representative background regions in the training image sequence, because the network needs to gather examples for all classes including the background. The network is not retrained on-line using the tracking results in the test phase (this is another difference with respect to the Bayesian method described). It’s well known that using a 1-of-c target coding scheme for the classes, the outputs of a network trained by minimizing a sum-of-squares error function approximate the posterior probabilities of class membership (here, Qc t (x,y) ), conditioned on the input feature vector [45]. Anyway, to guarantee a proper sum to unity of the posterior probabilities, the network outputs (which are always positive values between 0 and 1) are divided by their sum before assigning the posterior probabilities. 4. Dynamic recognition module Even though the static recognition module can be applied independently to each image in the sequence, this does not exploit the dynamic nature of the problem and the continuity and smoothness properties that are expected in the apparent motion of the objects through the sequence. Hence, a dynamic update of the pixel class probabilities pc t (x,y) is desired that takes into account these properties. To this end, not only the previous probabilities pc t-1 (x,y) and the results of the current static recognition Qc t (x,y) have to be combined but also the binary results of the tracking decision in the previous step Tc t-1 (x,y) have to be considered, since this permits to filter some possible misclassifications made by the static classifier. Typically, some background spots are erroneously classified as part of an object and this can be detected if these spots are situated far from the last known position of the object. Therefore, the update function f for the dynamic class probabilities can be defined as follows (for some adaptive parameters αc t , 0<αc t <1): ( ) ),()1(),(),( ),()1(),(),( ),( 1 1 11 11 ∑ + = −− −− −+ −+ = N k t k t k t k t k t k t c t c t c t c t ct c yxQyxpyxT yxQyxpyxT yxp αα αα (8) A tracking image for the background, which is required in the previous equation, can be defined simply as 11    ≤≤∀= =+ otherwise0 1:0),(if1 ),( 1 NccyxT yxT t ct N (9) and computed after the tracking images for the objects. The parameter αc t that weights the influence of the previous probabilities must be adapted depending on the apparent motion of the tracked object of class c. If this motion is very slow, αc t should reach a maximum αmax closer to 1, whereas if the motion is very fast, αc t should reach a minimum αmin closer to 0. In order to a set a proper value for αc t the areas (Ac t-1 and Ac t-2 ) and mass centers (Cc t-1 and Cc t-2 ) of the object in the two previous tracking images are used in the following way. Let π11 −− = tc t c Ar and π22 −− = tc t c Ar be the estimates of the object radius in the two previous frames obtained by imposing a circular area assumption. Let 21 −− −= t c t cc CCd be the estimated displacement of the object in the 2D image and let 21max −− += t c t cc rrd be the maximum displacement yielding some overlapping between the two former circles. If max cc dd ≥ we would like to set min αα =t c , whereas if 0= c d then the value of αc t should be set according to the change of the object apparent size: Let ( )2121 ,max −−−− −= t c t c t c t cc rrrrs be a scale change ratio. If 0= c s (unchanged object size) then we would like to set max αα =t c whereas in the extreme case 1= c s then we would set min αα =t c again. Combining linearly both criteria, displacement and scale change, we define the prior value ( ) ( ) ccc t c sdd minmax max minmaxmax ˆ αααααα −−−−= (10) which satisfies max ˆ αα ≤t c . Note that the value max ˆ αα =t c (maximum weight for the previous probabilities) is obtained when both 0= c d and 0= c s , what means that both the centers and the areas of the object are the same in the last two observations (no displacement and no scale change have occurred). Finally the parameter αc t is set as follows:      >∧< ≤∧< ≥ = min max min max min max min ˆifˆ ˆif if ααα ααα α α t ccc t c t ccc cc t c dd dd dd (11) The constants αmin and αmax were set to 0.1 and 0.6, respectively, in our experiments (see Section 6). 5. Tracking decision module As depicted in Figure 1, the tracking images Tc t (x,y) for the objects (1 ≤ c ≤ N) can be 12 calculated dynamically using the pixels probabilities p t (x,y) according to some decision function d. However, this function involves some additional arguments and results, as explained next. To give an initial estimate of the foreseen translation and rescaling of the object in the current step, the measurements of both the object mass center and area in the tracking images of the two previous steps are required. Hence, the areas Ac t-1 and Ac t-2 and the mass centers Cc t-1 and Cc t-2 , already used in the dynamic recognition module as we have seen, must also be supplied here. The application of the estimated transformation to the previous tracking image Tc t-1 (x,y) will serve to reduce the image area to explore using the class probabilities while filtering (blacking) the rest. This strategy alone permits to track visible objects reasonably well [26, 27] but it fails completely if the object becomes occluded for some frames [28]. In order to cope with occlusion, more information is needed in the decision function d. The key point is to distinguish between the a posteriori tracking image Tc t (x,y) and an a priori prediction ( )yxT t c ,ˆ , which could maintain some relevant information of the object before the occlusion such as area and movement. The object mass center Cc t and area Ac t needed for tracking should be measured either from Tc t (x,y) or ( )yxT t c ,ˆ depending on whether the object is visible or occluded. Hence, an occlusion flag Oc t has to be determined as an additional result. Moreover, the two previous flags Oc t-1 and Oc t-2 help to know whether the object is entering or exiting an occlusion. In addition, t c m r is a movement weighted average vector that represents the past trajectory direction of the tracked object, which is useful to solve some ambiguous cases that happen when the object crosses or exits an occlusion by another object with a similar appearance (same- class occlusion). Finally, it should be taken into account that the uncertainty in the prediction ( )yxT t c ,ˆ grows as the number of consecutive frames the object is occluded increases. In the original method described in [27], two constant parameters ε and δ were used to define an uncertainty region around each pixel transformation. Since we want to adjust the level of uncertainty based on the duration of the occlusion, these parameters have to be adaptive for each object, i.e. εc t and δc t . Summarizing, the decision function d involves the following arguments and results:           = −−−− −−−−− −− 1121 21211 11 ,,, ,,,,, ),,(ˆ),,(),,( ,,,,,),,(ˆ),,( t c t c t c t c t c t c t c t c t c t c t c t t c t c t c t c t c t c t c t c AA CCOOm yxTyxTyxp dACOmyxTyxT δε δε rr (12) This function is described in detail in the next subsections, which cover the different independent sub-modules of the tracking decision module. Figure 2 illustrates graphically some of the calculations that are explained in what follows. 5.1 A priori prediction of the tracked objects The first step is to give a priori estimates of the mass center and area of the object in time t. The mass center is predicted as follows: 13 otherwise2 ifˆ 21 121    − ¬∧ = −− −−− t c t c t c t c t ct c CC OOC C (13) When the object is exiting an occlusion, Cc t-2 is not reliable enough to be used together with Cc t-1 to predict the next movement; therefore, a conservative estimate is given, just the previous measured value. In the rest of cases (the object is visible, is occluded or is entering an occlusion), a constant rate prediction is used. Note, however, that when the object is occluded, the mass center is not measured on the a posteriori tracking image, but on the a priori one, as we will see later. It is interesting to notice that the above constant rate prediction can be proved to be equivalent to the one given by a linear Kalman filter for a particular setting of the filter parameters and equations. Let t c t c t c t c uBwAw ω++=+ 1 and t c t c t c wHd ν+= be respectively the state and measurement equations of a linear Kalman filter (KF) for predicting the mass center of object c. If we set A=I, B=I, H=I, dc t =Cc t as the measurement, uc t = Cc t - Cc t-1 as the input and R t =0 as the covariance matrix of the measurement noise νc t (which is assumed to be zero), then the a priori and a posteriori estimates of state wc t given by the KF are 2Cc t-1 - Cc t-2 and Cc t respectively. Figure 2: Geometrical illustration of the tracking process. Estimates of object’s area and mass center for step t are computed from previous values in t-1 and t-2. For each pixel in step t a rectangular region in step t-1 is determined which allows the assignment to the pixel of one of three labels: “certainly belonging to the object” (yellow diagonal-bar-shaded region), “uncertain” (blue brick-shaded region) and “certainly not belonging to the object” (the rest of the image). 14 The a priori estimate of the object area is calculated as follows: ( ) otherwise ifˆ 1 12221     ¬∧¬ = − −−−− t c t c t c t c t ct c A OOAA A (14) If the object has been visible in the two previous frames, a constant rate of a scale factor is used to predict the area. It can be proved that this prediction is equivalent to the one given by a (non-linear) extended Kalman filter for a particular setting of the filter parameters and equations. Let ( )t c t c t c t c uwfw ω,, 1 =+ and ( )t c t c t c whd ν, = be the state and measurement equations, respectively, of an extended Kalman filter (EKF) for predicting the area of object c. If we set ( ) t c t c t c t c t c t c uwuwf ωω +=,, , ( ) tctctctc wwh νν +=, , dc t =Ac t as the measurement, uc t = Ac t /Ac t-1 as the input and R t =0 as the covariance matrix of the measurement noise νc t (which is assumed again to be zero), then the a priori and a posteriori estimates of state wc t given by the EKF are (Ac t-1 ) 2 / Ac t-2 and Ac t respectively. In the rest of cases (the object is occluded or is entering or exiting an occlusion), the area is supposed to remain constant. From these predictions, a change of coordinates transformation can also be estimated that maps each pixel Pc t-1 = (xc t-1 ,yc t-1 ) of the object c in step t-1 (maybe occluded) into its foreseen position in step t: ( ) 111 ˆˆ ˆ −−− −+= t c t c t c t c t c t c AACPCP (15) Actually, we are interested in applying the transformation in the inverse way, i.e. to know which is the expected corresponding position in time t-1, ( )111 ˆ,ˆˆ −−− = tc t c t c yxP , of a given pixel Pc t = (xc t , yc t ) in t: ( ) 1 11 ˆ ˆ ˆ − −− − += t c t c t c t ct c t c AA CP CP (16) This is enough to compute the a priori tracking image ( )yxT t c ,ˆ in time t, either from the previous a posteriori or a priori tracking image, depending on the previous occlusion flag: ( ) ( ) ( ) otherwiseˆ,ˆˆ ifˆ,ˆ ,ˆ 111 1111    ¬ = −−− −−−− t c t c t c t c t c t c t ct c yxT OyxT yxT (17) where the values of 11 ˆ,ˆ −− t c t c yx are clipped whenever is necessary to keep them within the range of valid coordinates. 5.2 First computation of the tracking images To compute the a posteriori tracking image Tc t (x,y), the pixel class probabilities p t (x,y) are taking into account only in some image region that is determined from Tc t-1 (x,y) or 15 ( )yxT t c ,ˆ 1− (depending on Oc t-1 ) and the tolerance parameters εc t and δc t . Since the estimates of the translation and scale parameters in the coordinate transformation can be inaccurate, we define a rectangular region of possible positions for each pixel by specifying some tolerances in these estimates. To this end, we use the adaptive parameters εc t and δc t , which must be positive values to be set in accordance with our confidence in the translation and scale estimates respectively (the most confidence the smallest tolerance and vice versa), and which are adjusted according to the following rules: ( ) otherwise if,min 121 max    ∨+ = −−− ini t c t cincr t ct c OO ε εεε ε (18) ( ) otherwise if,min 121 max    ∨+ = −−− ini t c t cincr t ct c OO δ δδδ δ (19) where εini, δini are default values, εmax, δmax are the maximal allowed values and εincr, δincr are the respective increases for each successive step under occlusion. Note that the tolerances keep on growing when exiting an occlusion until the object has been visible in the two previous frames; this is needed to detect and track the object again. Let ( )111 , −−− = t Cc t Cc t c yxC and ( )t Cc t Cc t c yxC ˆ,ˆˆ = be respectively the previous mass center and the a priori estimate of the current mass center. The four vertices of the rectangular uncertainty region centered at 1ˆ −t c P are denoted (top-left) TLc t-1 = (xinfc t-1 , yinfc t-1 ), (top- right) TRc t-1 =(xsupc t-1 , yinfc t-1 ), (bottom-left) BLc t-1 =(xinfc t-1 , ysupc t-1 ) and (bottom-right) BRc t-1 = (xsupc t-1 , ysupc t-1 ), where: ( ) ( ) ( ) ( )        − −−− + ≥−−− + −−− + = − − − − − − − − otherwise ˆˆ ˆˆ 0ˆˆif ˆˆ ˆˆ 1 1 1 1 1 1 1 1 t c t c t c t c t cC t cC t c t cC t ct cC t cC t cC t c t cC t c t c t c t c t c t cC t cC t c t cC t ct cC t c AAA xxxx x xxxx AAA xxxx x xinf δ ε ε δ ε (20) ( ) ( ) ( ) ( )       + −+− + ≥−+− − −+− + = − − − − − − − − otherwise ˆˆ ˆˆ 0ˆˆif ˆˆ ˆˆ 1 1 1 1 1 1 1 1 t c t c t c t c t cC t cC t c t cC t ct cC t cC t cC t c t cC t c t c t c t c t c t cC t cC t c t cC t ct cC t c AAA xxxx x xxxx AAA xxxx x xsup δ ε ε δ ε (21) 16 ( ) ( ) ( ) ( )       − −−− + ≥−−− + −−− + = − − − − − − − − otherwise ˆˆ ˆˆ 0ˆˆif ˆˆ ˆˆ 1 1 1 1 1 1 1 1 t c t c t c t c t cC t cC t c t cC t ct cC t cC t cC t c t cC t c t c t c t c t c t cC t cC t c t cC t ct cC t c AAA yyyy y yyyy AAA yyyy y yinf δ ε ε δ ε (22) ( ) ( ) ( ) ( )        + −+− + ≥−+− − −+− + = − − − − − − − − otherwise ˆˆ ˆˆ 0ˆˆif ˆˆ ˆˆ 1 1 1 1 1 1 1 1 t c t c t c t c t cC t cC t c t cC t ct cC t cC t cC t c t cC t c t c t c t c t c t cC t cC t c t cC t ct cC t c AAA yyyy y yyyy AAA yyyy y ysup δ ε ε δ ε (23) The values of xinfc t-1 , yinfc t-1 , xsupc t-1 and ysupc t-1 are clipped whenever is necessary to keep them within the range of valid coordinates. In order to understand eqs. (20) to (23), they must be first compared with eq. (16), which gives the position of the rectangle center (jointly for x and y coordinates). Then, consider for instance eq. (20), since the other three are simply derived by symmetry; eq. (20) aims at setting the leftmost value of x in the uncertainty rectangle; to this end, some small proportion of the estimated center displacement in the x coordinate is subtracted in the numerator, and the scale ratio is either enlarged or shrunk in the denominator (by adding or subtracting respectively a small proportion of the new estimated area) depending on which of these two options yields the smallest (leftmost) x value; it’s easy to check that the last depends on the sign of the numerator. Now, each pixel Pc t = (xc t ,yc t ) is labeled, with respect to object c, as one of three labels (“certainly belonging to the object c”, “certainly not belonging to the object c” or “uncertain”) as follows. If Oc t-1 is false then: if all the pixels in the rectangular region delimited by TLc t-1 , TRc t-1 , BLc t-1 , BRc t-1 have a common value of 1 in Tc t-1 (x,y), it is assumed that Pc t is definitely inside and certainly belongs to object c; to the contrary, if they have a common value of 0 in Tc t-1 (x,y), it is assumed that Pc t is clearly outside and certainly does not belong to object c; otherwise, the rectangular region contains both 1 and 0 values, the pixel Pc t is initially labeled as “uncertain”. However, if Oc t-1 is true, Tc t-1 (x,y) will represent a totally or partially occluded object and we cannot rely on it, but on the predicted ( )yxT t c ,ˆ 1− , which is based on information previous to the occlusion. If all the pixels in the rectangular region delimited by TLc t-1 , TRc t-1 , BLc t-1 , BRc t-1 have a common value of 0 in ( )yxT t c ,ˆ 1− , it is assumed that Pc t does not belong to object c; otherwise (the rectangular region contains both 1 and 0 values or only 1 values), the pixel Pc t is labeled as “uncertain”. Only for the uncertain pixels (x,y) the dynamic probabilities p t (x,y) will be used. Recall that these probabilities will have been updated previously from the object recognition results in time t, Q t (x,y), also expressed as probabilities. More precisely, we propose the 17 following rule to compute the value of each pixel of the a posteriori tracking image for object c in time t:         += otherwise0 1 and 1between allfor maximum theis),(anduncertainisit objecttobelongscertainly),(if 1 ),( Nc yxpor cyx yxT t c t c (24) 5.3 Post-processing of the tracking images Sometimes, the tracking images Tc t (x,y) obtained by applying eq. (24) contain disconnected regions of 1-valued pixels, or, said in other words, more than one connected component t ci T , 1≤i≤I, I>1. This may be produced by a variety of causes, mainly segmentation or recognition errors, but also may be due to possible partial occlusions of the target object by an object of a different class. In addition, a particular problem that leads to object split occurs immediately after a same-class crossing or occlusion: when the target object has just finished crossing another object or region which is recognized to be in the same class (distracter), then the tracking method is misled to follow both the object and the distracter. It is very difficult to devise a general method that can always distinguish between erroneous components due to noise or distracters and correct object components, especially if separated components are allowed to cope with partial occlusions, but some useful heuristics based on properties such as size, movement or shape may be defined that work reasonably well in a majority of cases. In order to eliminate noisy regions and to circumvent the same-class crossing problem, while handling partial occlusions at the same time, we propose a post-processing step that removes from Tc t (x,y) some possible artifacts or distracters (setting some initially 1- valued pixels to zero). In fact, this step is only carried out if Tc t (x,y) contains more than one component. In such a case, we need to choose which components t ci T to keep (one or more) and which to discard. To this end, three heuristic filters are applied sequentially, whenever two or more components remain before the filter application. The first filter is aimed at deleting small noisy regions and is solely based on their size. Let t ci A be the area of the i-th connected component t ci T and let )(Area t c T be the total area covered by 1-valued pixels in Tc t (x,y). The i-th component is removed if the ratio )(Area t c t ci TA is below a given threshold κ , e.g. κ = 0.15. The second filter is aimed at deleting distracters, including those appearing after same- class occlusion, and is based on a comparison between the apparent movement of the remaining components and the previous recent trajectory of the tracked object represented by the movement vector 1−t c m r . Let t ci C be the mass center of the i-th connected component t ci T and define an associated movement vector 1− −= t c t ci t ci CCz r for each component. Then, 18 1 1 , θcos − − 〉〈 = t c t ci t c t ci ci mz mz rr rr (25) is a measure of the alignment between the vectors 1−t c m r and t ci z r , which is only reliable for our purposes if both vectors have a sufficiently large norm. Otherwise, the angle θci can be considered rather random, since may be affected a lot by adding small perturbations on the vectors. Consequently, abrupt trajectory changes (greater than 90 degrees) are penalized if we remove the i-th component t ci T when the condition λλ ≥∧≥∧< −10θcos t c t cici mz rr holds, where λ is another threshold, e.g. λ = 3. However, to guarantee that at least one component is kept, the remaining component for which t ci z r is the most collinear vector with respect to 1−t c m r , i.e. the component i such that         〉〈 = − t ci t c t ci z mz i r rr 1 , max arg (26) is never removed by this second filter. The third filter is also aimed at deleting distracters and is based on a comparison between the shapes of the components and that of the a priori prediction of the target object (represented by the 1-valued region in ( )yxT t c ,ˆ ). For each remaining component t ci T , the a priori prediction of the target object is moved from its original center t c Ĉ to the component center t ci C , thus resulting in a translated copy ( )yxT t ci ,ˆ , and the spatial overlap between both shapes is then measured as follows: ( ) ( )t ci t ci t ci t ci ci TT TT SO ˆ Area ˆ Area ∪ ∩ = (27) The components having a spatial overlap 243.0< ci SO (which is the overlap obtained between two circles of the same size when one of the centers is located in the border of the other circle) are deleted in this third filter, unless ci SO is the maximum spatial overlap of the remaining components. This exception guarantees the persistence of at least one component in the final tracking image. As a result of the post-processing, the pixels of all the components t ci T removed by any of the three filters are set to zero in the final tracking image Tc t (x,y). 5.4 Determination of occlusion and geometric measurements Once both Tc t (x,y) and ( )yxT t c ,ˆ have been determined, it is possible to detect the occurrence of an occlusion (i.e. to set the current occlusion flag) in the following way. Let Area(Tc t ) be the measured area of the 1-valued region in the final Tc t (x,y) and let Area( t c T̂ ) be the measured area of the 1-valued region in ( )yxT t c ,ˆ . Then, 19 ( ) ( ) ( ) ( ) otherwiseˆAreaArea ifˆAreaArea 2 1 1    < < = − rTT OrTT O t c t c t c t c t ct c (28) where 0 < r1 < r2 < 1 (for instance, r1=0.5, r2=0.75). Note that the condition for remaining in occlusion mode is harder than the condition for initiating an occlusion. This facilitates the recovery of the object track when exiting an occlusion or when a false occlusion has been detected. Next, the a posteriori estimates of the object mass center and area are selected between those of the a priori and a posteriori tracking images based on the value of the occlusion flag: ( ) ( ) otherwiseˆ if    ¬ = t c t c t ct c TMC OTMC C (29) where MC(Tc t ) is the measured mass center of the 1-valued region in the final Tc t (x,y) and MC( t c T̂ ) is the measured mass center of the 1-valued region in ( )yxT tc ,ˆ , and ( ) ( ) otherwiseˆArea ifArea    ¬ = t c t c t ct c T OT A (30) Finally, the movement weighted average vector t c m r is updated afterwards as follows: ( )( ) ( )   ≥−+ <−+ = − − βββ β 1if1 1if1 1 1 tmv ttmtv m t c t c t c t ct c rr rr r (31) where β is a positive parameter between 0 and 1, e.g. β=0.2, and t c v r is the current movement defined by 1− −= t c t c t c CCv r . Note that the second row in (31) is a typical moving average computation, while the first row denotes a simple average for the starting steps, and both give the same result for t=1/β. 6. Experimental results We were interested in testing both PIORT approaches in video sequences including object occlusions and taken with a moving camera. Nevertheless, we also performed a first set of validation experiments in video sequences taken with a still camera. In all tests we defined N=1 objects of interest to track. All images in the video sequences were segmented independently using the EDISON implementation of the mean-shift segmentation algorithm, code available at http://www.caip.rutgers.edu/riul/research/code.html. The local features extracted for each spot were the RGB colour averages. For object learning, spots selected through ROI (region-of-interest) windows in the training sequence were collected to train a two- layer perceptron using backpropagation and to build the target class histogram. When 20 using the neural net in the test phase, the class probabilities for all the spots in the test sequences were estimated from the net outputs. When using the histogram, the spot class probabilities were estimated according to equation (6). In both cases, the spot class probabilities were replicated for all the pixels in the same spot. For object tracking in the test sequences, ROI windows for the target object were only marked in the first image to initialise the tracking process. The recognition and tracking results for the test sequences of our PIORT approaches were stored in videos where each frame has a layout of 2 x 3 images with the following contents: the top left is the image segmented by EDISON; the top middle is the image of probabilities given by the static recognition module for the current frame; the top right is the a priori prediction of the tracking image; the bottom left is the image of dynamic probabilities; the bottom right is the original image with a graphic overlay that represents the boundaries of the a posteriori binary tracking image (the final result for the frame); and the bottom middle is an intermediate image labelled by the tracking module where yellow pixels correspond to pixels labelled as “certainly belonging to the object”, light blue pixels correspond to pixels initially labelled as “uncertain” but with a high dynamic probability, dark blue pixels correspond to pixels labelled as “uncertain” and with a low probability, dark grey pixels are pixels labelled as “certainly not belonging to the object” but with a high probability and the rest are black pixels with both a low probability and a “certainly not belonging to the object” label. For comparison purposes, tracking of the target objects in the test sequences was also carried out by applying the six following methods, which only need the ROI window mark in the first frame of the test sequence: Template Match by Correlation (TMC), which refers to normalized correlation template matching [30]; Basic Meanshift (BM) [31]; Histogram Ratio Shift (HRS) [32]; Variance Ratio Feature Shift (VRFS) [32]; Peak Difference Feature Shift (PDFS) [32]; and Graph-Cut Based Tracker (GCBT) [33, 30]. These methods have been commented briefly in Section 1. From the tracking results of all the tested methods, two evaluation metrics were computed for each frame: the spatial overlap and the centroid distance [46]. The spatial overlap SO(GTk,STk) between the ground truth GTk and the system track STk in a specific frame k is defined as the ratio ( ) ( ) ( ) kk kk kk STGT STGT , STGTSO Area Area ∪ ∩ = (32) and Dist(GTCk, STCk) refers to the Euclidean distance between the centroids of the ground truth (GTCk) and the system track (STCk) in frame k. Naturally, the larger the overlap and the smaller the distance, the better performance of the system track. Since the centroid distance can only be computed if both GTk and STk are non-null, a failure ratio was measured as the number of frames in which either GTk or STk was null (but not both) divided by the total number of frames. Finally, an accuracy measure was computed as the number of good matches divided by the total number of frames, where a good match is either a true negative or a true positive with a spatial overlap above a threshold of 0.243 (which is the overlap obtained between two circles of the same size when one of the centres is located in the border of the other circle). 21 6.1 Experimental results on video sequences taken with a still camera The first set of experiments comprised three test video sequences taken with a still camera that show indoor office scenes where the target to track is a blue ball moving on a table. A similar but different sequence was used for training a neural network to discriminate between blue balls and typical sample regions in the background and for constructing the class histogram of the blue ball (this training sequence is available at http://www-iri.upc.es/people/ralqueza/bluetraining.avi). In the first test sequence, http://www-iri.upc.es/people/ralqueza/S1S2.avi, two blue balls are moving on the table and one occludes temporally the other one during some frames. Two experiments were performed on this test sequence depending on the initialisation of the tracking. In test S1, the tracking was initialised at the right ball and in test S2, the tracking was initialised at the left ball. The static recognition module considers that both balls belong to the same class. In both tests, the temporal overlapping was correctly managed by our methods since the tracked ball is well relocated after exiting the occlusion. The corresponding videos displaying the results of PIORT methods (in the layout described above) are at http://www-iri.upc.es/people/ralqueza/S1_NN.mpg and S2_NN.mpg for the PIORT-Neural net method and at S1_Bayes.mpg and S2_Bayes.mpg for the PIORT-Bayesian method. Sequence S1: Blue balls crossed. Tracking initial right ball Sequence S2: Blue balls crossed. Tracking initial left ball In the second test sequence (test S3), http://www-iri.upc.es/people/ralqueza/S3.avi, the tracked blue ball is occluded twice by a box during 5 and 12 frames, respectively. Recognition and tracking results for the whole sequence using the PIORT-Neural Net and Bayesian methods are at http://www-iri.upc.es/people/ralqueza/S3_NN.mpg and S3_Bayes.mpg, respectively. The tracking of the blue ball is quite satisfactory since both occlusions are correctly detected and the ball is correctly relocated when exiting the occlusion. Sequence S3: Blue ball moving occluded by box 22 In the last test sequence of this group, http://www-iri.upc.es/people/ralqueza/S4.avi, there are again two blue balls and the target moving ball crosses twice, once in front of and once behind, the second ball, which does not move. As the recognition module classifies both balls in the same class, the same-class occlusion is not detected as an occlusion (the two balls are merged into a single blue object), but anyway the target ball is well tracked after the two crossings. The videos displaying the results of the PIORT- Neural Net and Bayesian methods for this sequence are at http://www- iri.upc.es/people/ralqueza/S4_NN.mpg and S4_Bayes.mpg, respectively. Sequence S4: Blue ball moving around another blue ball Table 1 presents the results (mean ± std. deviation) of the spatial overlap (SO) and centroid distance (CD) measures together with the failure ratio (FR) and accuracy (Acc) of each tracking method for the four tests S1 to S4, emphasizing the best values for each measure and test in bold. Our PIORT tracking methods worked fine in the four tests obtaining the best values of the four measures (except in the Accuracy measure for test S4, where the HRS method gave a slightly superior performance). All methods performed quite well in S1; only PDFS method performed comparably to PIORT approaches in S2; only PIORT methods worked in S3 while the rest failed; and only BM and HRS methods performed comparably to PIORT approaches in S4. 23 Table 1. Results of ball tracking on video sequences taken with a still camera. SO: Spatial Overlap; CD: Centroid Distance; FR: Failure Ratio; Acc: Accuracy Video Sequence Tracking method SO CD FR Acc S1 Blue balls crossed (Right ball) TMC 0.56 ± 0.10 5.07 ± 2.07 0 0.98 BM 0.60 ± 0.06 3.19 ± 1.21 0 1.00 HRS 0.46 ± 0.11 6.03 ± 2.05 0 1.00 VRFS 0.66 ± 0.07 1.15 ± 0.47 0 1.00 PDFS 0.63 ± 0.10 2.01 ± 0.94 0 1.00 GCBT 0.64 ± 0.18 13.20 ± 52.52 0.05 0.94 PIORT-Neural Net 0.84 ± 0.09 1.38 ± 1.39 0 1.00 PIORT-Bayesian 0.80 ± 0.07 0.75 ± 0.76 0 1.00 S2 Blue balls crossed (Left ball) TMC 0.22 ± 0.27 44.34 ± 52.24 0 0.41 BM 0.23 ± 0.29 42.51 ± 50.42 0 0.36 HRS 0.25 ± 0.31 44.93 ± 51.96 0 0.41 VRFS 0.28 ± 0.35 42.82 ± 52.62 0 0.41 PDFS 0.50 ± 0.30 36.27 ± 86.95 0.14 0.77 GCBT 0.20 ± 0.27 70.69 ± 68.80 0 0.36 PIORT-Neural Net 0.60 ± 0.23 3.94 ± 4.98 0 0.91 PIORT-Bayesian 0.46 ± 0.25 15.04 ± 52.64 0.05 0.73 S3 Blue ball moving occluded by box TMC 0.01 ± 0.04 173.40 ± 68.71 0.22 0 BM 0.01 ± 0.07 182.54 ± 68.14 0.22 0 HRS 0 187.85 ± 67.96 0.25 0 VRFS 0.02 ± 0.18 140.14 ± 93.44 0.20 0.17 PDFS 0.13 ± 0.41 131.07 ± 106.1 0.42 0.02 GCBT 0 237.02 ± 134.6 0.74 0.22 PIORT-Neural Net 0.81 ± 0.42 0.47 ± 0.38 0 1.00 PIORT-Bayesian 0.53 ± 0.37 8.39 ± 48.61 0.03 0.95 S4 Blue ball moving around still blue ball TMC 0.35 ± 0.22 13.10 ± 32.38 0.01 0.75 BM 0.56 ± 0.15 7.39 ± 29.05 0.01 0.93 HRS 0.60 ± 0.13 6.21 ± 29.16 0.01 0.96 VRFS 0.10 ± 0.62 74.68 ± 45.00 0.01 0.14 PDFS 0.13 ± 0.43 44.39 ± 36.14 0.01 0.17 GCBT 0.10 ± 0.53 201.60 ± 98.35 0.80 0.18 PIORT-Neural Net 0.74 ± 0.21 5.90 ± 29.33 0.01 0.94 PIORT-Bayesian 0.72 ± 0.20 5.58 ± 29.38 0.01 0.94 24 6.2 Experimental results on video sequences taken with a moving camera The second set of experiments comprised another three test video sequences where the target is a ball, but this time taken with a moving camera. The first of them (test S5) again shows an indoor office scene where a blue ball is moving on a table and is temporally occluded, while other blue objects appear in the scene. This test sequence can be downloaded at http://www-iri.upc.es/people/ralqueza/S5.avi. Sequence S5: Blue bouncing ball on table The other two test sequences in this group show outdoor scenes in which a Segway robot tries to follow an orange ball that is being kicked by a person. Both include multiple occlusions of the tracked orange ball and differ in the surface over which the ball runs, which is pavement in the case of test S6 and grass in test S7 (see http://www- iri.upc.es/people/ralqueza/S6.avi and S7.avi, respectively). A similar but different sequence was used for training a neural network to discriminate between orange balls and typical sample regions in the background and for constructing the class histogram of the orange ball (this training sequence is available at http://www- iri.upc.es/people/ralqueza/orangetraining.avi). Table 2. Results of ball tracking on video sequences taken with a mobile camera. 25 SO: Spatial Overlap; CD: Centroid Distance; FR: Failure Ratio; Acc: Accuracy The tracking results videos for the above test sequences are attainable at http://www- iri.upc.es/people/ralqueza/S5_NN.mpg, S5_Bayes.mpg, S6_NN.mpg, S6_Bayes.mpg, S7_NN.mpg and S7_Bayes.mpg. Sequence S6: Segway - Orange ball on pavement Video Sequence Tracking method SO CD FR Acc S5 Blue bouncing ball on table TMC 0.28 ± 0.48 74.65 ± 91.53 0.19 0.43 BM 0.23 ± 0.52 78.40 ± 90.33 0.19 0.37 HRS 0.16 ± 0.45 125.88 ± 11.80 0.43 0.30 VRFS 0.20 ± 0.38 96.72 ± 134.84 0.39 0.60 PDFS 0.28 ± 0.57 103.60 ± 36.77 0.41 0.59 GCBT 0.01 ± 0.29 188.79 ± 18.13 0.75 0.21 PIORT-Neural Net 0.60 ± 0.40 12.53 ± 59.38 0.05 0.95 PIORT-Bayesian 0.59 ± 0.39 12.46 ± 59.40 0.05 0.95 S6 Segway - Orange ball on pavement TMC 0.06 ± 0.40 146.35 ± 81.83 0.03 0.14 BM 0.09 ± 0.43 110.94 ± 76.70 0.03 0.19 HRS 0.09 ± 0.38 156.99 ± 103.80 0.41 0.21 VRFS 0.16 ± 0.68 70.46 ± 49.17 0.03 0.21 PDFS 0.14 ± 0.59 117.09 ± 81.43 0.03 0.21 GCBT 0.01 ± 0.34 233.56 ± 62.12 0.93 0.06 PIORT-Neural Net 0.72 ± 0.20 2.67 ± 19.21 0.01 0.98 PIORT-Bayesian 0.13 ± 0.73 202.14 ± 99.35 0.81 0.19 S7 Segway - Orange ball on grass TMC 0.02 ± 0.29 137.93 ± 84.53 0.04 0.04 BM 0.15 ± 0.27 125.13 ± 116.14 0.34 0.35 HRS 0.03 ± 0.33 190.63 ± 89.72 0.54 0.08 VRFS 0.59 ± 0.21 7.93 ± 38.85 0.02 0.95 PDFS 0.33 ± 0.50 121.46 ± 125.91 0.48 0.51 GCBT 0.01 ± 0.37 208.39 ± 83.88 0.79 0.04 PIORT-Neural Net 0.47 ± 0.23 17.02 ± 60.98 0.06 0.88 PIORT-Bayesian 0.25 ± 0.49 133.43 ± 126.22 0.53 0.42 26 Sequence S7: Segway - Orange ball on grass Table 2 presents the results (mean ± std. deviation) of the spatial overlap (SO) and centroid distance (CD) measures together with the failure ratio (FR) and accuracy (Acc) of each tracking method for the three tests S5 to S7, emphasizing the best values for each measure and test in bold. Our PIORT-Neural net method worked fine in the three tests obtaining the best values of spatial overlap and accuracy measures in tests S5 and S6 and yielding results a little bit under the performance of the VRFS method in test S7, in which the VRFS method gave the best values of the four measures. Our PIORT- Bayesian method worked well in test S5 but failed to track the orange ball correctly in tests S6 and S7. Only both PIORT methods performed well in S5; only PIORT-Neural net method worked in S6 while the rest failed; and only VRFS and PIORT-Neural net methods obtained satisfactory results in S7. Sequence S8: Pedestrian with red jacket The last set of experiments comprised another three test video sequences, taken with a moving camera in outdoor environments, where the targets are humans, more precisely, some part of their clothing. The first sequence in this group (test S8) is a long sequence taken on a street where the aim is to track a pedestrian wearing a red jacket (see http://www-iri.upc.es/people/ralqueza/S8.avi) and includes total and partial occlusions of the followed person by other walking people and objects on the street. In this case, a short sequence of the scene taken with a moving camera located in a different position (http://www-iri.upc.es/people/ralqueza/redpedestrian_training.avi) was used as training sequence. 27 The other two test sequences in this group, tests S9 and S10, show outdoor scenes in which humans riding Segway robots and wearing orange T-shirts are followed. In test S9 a single riding guy is followed, whereas in test S10, two men are riding two Segway robots simultaneously and crossing each other. These test sequences are at http://www- iri.upc.es/people/ralqueza/S9.avi and S10.avi and the training sequence associated with them is at http://www-iri.upc.es/people/ralqueza/T-shirt_training.avi. Sequence S9: Guy on Segway with orange T-shirt Sequence S10: Men on Segway with orange T-shirt The tracking results videos for the above test sequences are attainable at http://www- iri.upc.es/people/ralqueza/S8_NN.mpg, S8_Bayes.mpg, S9_NN.mpg, S9_Bayes.mpg, S10_NN.mpg and S10_Bayes.mpg. 28 Table 3 presents the results of the evaluation measures of each tracking method for the three tests S8 to S10, emphasizing the best values for each measure and test in bold. Both PIORT methods gave the best results, very similar between them, in tests S8 and S9, and PIORT-Neural net method performed clearly the best in test S10. Note that in the pedestrian sequence (S8), an occlusion by people carrying red bags distracted the attention of the PIORT tracking module and caused a momentarily impairment in performance, especially for the centroid distance measure, but the tracker was able to recover correctly the target after that occlusion. In this sequence S8, only the PDFS method performed comparably to PIORT approaches in terms of accuracy and centroid distance, although it achieved a rather lower spatial overlap. In test S9, the HRS, VRFS and PDFS methods obtained similar and reasonably well results, but not as good as those of PIORT methods. Finally, only the PIORT-Neural net method worked well in test S10, where the PIORT-Bayesian method performed poorly because it followed the other Segway-riding man after a crossing between both men. Table 3. Results of human tracking on video sequences taken with a mobile camera. SO: Spatial Overlap; CD: Centroid Distance; FR: Failure Ratio; Acc: Accuracy Video Sequence Tracking method SO CD FR Acc S8 Pedestrian with red jacket TMC 0.44 ± 0.31 25.25 ± 61.10 0.07 0.77 BM 0.24 ± 0.58 72.08 ± 64.33 0.07 0.34 HRS 0.35 ± 0.24 13.49 ± 38.27 0.02 0.64 VRFS 0.45 ± 0.32 34.27 ± 81.13 0.12 0.82 PDFS 0.50 ± 0.20 11.42 ± 45.11 0.03 0.95 GCBT 0.04 ± 0.32 194.7 ± 105.3 0.77 0.16 PIORT-Neural Net 0.79 ± 0.24 11.90 ± 50.87 0.04 0.96 PIORT-Bayesian 0.74 ± 0.24 11.15 ± 48.14 0.04 0.95 S9 Guy on Segway with orange T- shirt TMC 0.10 ± 0.53 130.3 ± 69.75 0.00 0.15 BM 0.22 ± 0.13 41.30 ± 58.70 0.01 0.40 HRS 0.53 ± 0.25 22.83 ± 58.43 0.05 0.86 VRFS 0.69 ± 0.25 27.69 ± 75.15 0.10 0.90 PDFS 0.56 ± 0.21 29.19 ± 74.65 0.10 0.90 GCBT 0.14 ± 0.22 101.6 ± 112.7 0.36 0.19 PIORT-Neural Net 0.73 ± 0.16 3.40 ± 14.78 0.00 0.97 PIORT-Bayesian 0.74 ± 0.13 3.70 ± 14.61 0.00 0.98 S10 Men on Segway with orange T- shirts TMC 0.06 ± 0.39 104.3 ± 83.15 0.03 0.10 BM 0.29 ± 0.28 42.10 ± 59.06 0.03 0.59 HRS 0.28 ± 0.30 38.72 ± 65.09 0.06 0.58 VRFS 0.38 ± 0.34 36.81 ± 64.53 0.06 0.61 PDFS 0.32 ± 0.36 91.14 ± 119.4 0.35 0.56 GCBT 0.04 ± 0.31 187.1 ± 103.1 0.72 0.08 PIORT-Neural Net 0.73 ± 0.18 8.37 ± 40.74 0.03 0.96 PIORT-Bayesian 0.16 ± 0.58 81.36 ± 62.93 0.03 0.22 29 7. Conclusions, discussion and future work In this paper we have described an updated version of the probabilistic integrated object recognition and tracking (PIORT) methodology that we have developed in the latest years, partially reported in [36-39], and presented a collection of experimental results in test video sequences, with the aim of comparing two particular approaches derived from PIORT, based on Bayesian and neural net methods, respectively, with some state-of- the-art tracking methods proposed by other authors. An improved method for object tracking, capable of dealing with rather long occlusions and same-class object crossing, has been proposed to be included within our probabilistic framework that integrates recognition and tracking of objects in image sequences. PIORT does not use any contour information but the results of an iterative and dynamic probabilistic approach for object recognition. These recognition results are represented at pixel level as probability images and are obtained through the use of a classifier (e.g. a neural network) from region-based features. The PIORT framework is divided in three parts: a static recognition module, where the classifier is applied to single-frame images, a dynamic recognition module that updates the object probabilities using previous recognition and tracking results, and a tracking decision module, where tracking binary images are determined for each object. This third module combines the recognition probabilities with a model that predicts the object’s apparent motion in terms of translation and scale changes, while coping with the problems of occlusion and re-emergence detection. Moreover, the tracking module can deal with object splitting, either due to partial occlusions or same-class object crossing, and, in most cases, is able to select and track only the target object after it crosses or is occluded by another object which is recognized as belonging to the same class, i.e. it is able to re-establish the identity of the target object. The experimental work reported in this paper has been focused on the case of single object tracking, just because the tracking methods we had available for the comparison only allowed single object tracking. However, as shown in [36], the PIORT system is capable of tracking multiple objects of different classes simultaneously and, as demonstrated in the experiments, it can be applied to video sequences acquired either by a fixed or a moving camera. The size, shape and movement of the target objects can vary softly along the sequence, but the appearance features used by the classifier (up to now, colour features) should remain rather stable for a successful tracking. It must be taken into account that the global performance of the system depends not only on the ability of the tracking method but also on the quality of the object recognition probabilities provided by the trained classifier. In this regard, false positive detections by the classifier can only be harmful for tracking when they are very close or “touching” the target, otherwise they are filtered by the second and third modules. Even in the first case, the tracking module is sometimes able to distinguish between the target and a false distracter, when the latter is different enough in terms of size, shape or motion trajectory. Concerning false negative errors by the classifier, they can be coped partially by the second module, especially when the apparent motion of the target is slow and hence the previous probabilities (adaptively) weight more than the current ones given by the classifier. Nevertheless, if the classifier 30 fails to detect the target object in just a few consecutive frames the tracker will assume a target occlusion and proceed in occlusion mode, which implies an assumption of constant motion directed by the previous trajectory and a growing uncertainty in the target position. In this case, object tracking can be sometimes recovered if the classifier redetects the target afterwards, depending basically on the real trajectory of the target and the gap duration. In this paper, we have presented two static recognition methods that can be embedded in the first module of PIORT, giving rise to two different instances of the methodology. Both methods are based on the use of a classifier that is trained from examples and provides posterior class probabilities for each pixel from a set of local features. The first classifier is based on a maximum likelihood Bayesian method in which the conditional probabilities for object classes are obtained from the information of the class histograms (for discretized RGB values) and a uniform conditional probability is assumed for the background. The second classifier is based on a neural net which is trained with the RGB colour averages extracted for each spot of the segmented images. Even though the characteristics of these two classifiers are quite different, the recognition and tracking results of PIORT using both approaches were excellent and very similar in five of the ten test sequences, which might mean that the good ability of PIORT to track the objects is mostly due to a smart cooperation of the three inner modules and is not very dependent on the specific method used for object recognition. However, in the remaining five test sequences, the tracking method based on a neural net classifier clearly outperformed the one based on a simple Bayesian classifier, which failed in three of these test sequences. Indeed, we observed that updating the histograms at each frame may cause severe drift errors when the tracker begins to fail, which result in a rapid breakdown of the Bayesian classifier performance in subsequent frames. Hence, depending on the particular application, it might be preferable not to update the histograms after training. The performance of both Bayesian and neural net classifiers also depends somewhat on the quality of the image segmentation process carried out previously. In the case of good segmentations, like the ones we obtained using EDISON for the test sequences, the probability images given by the classifiers are smooth (large areas with same values) and this eases the tracking, whereas in the case of over-segmentations, the probability images may be noisy due to an excess of spots and this may hinder a stable tracking. In the experimental comparison with other six methods proposed in the literature for object tracking, a PIORT method obtained the best results in nine of the ten test sequences and only a slightly inferior performance with respect to best method in the other one (VRFS). Except for the case of the first test sequence S1, where all methods worked fine, the six alternative methods tested mostly failed to track the target objects correctly in the test sequences, due to the difficult instances of occlusions and object crossings they contain. However, we are aware of the fact that the six alternative methods tested here are not model-based (i.e. they are not trained in advance) to the contrary of PIORT, and thus, it is little surprising that PIORT obtained the best results. The availability (for us) of their implementation was the main reason why we selected them, but we foresee to carry out future experimental comparisons of PIORT against some state-of-the-art model-based tracking methods like those by Cremers [5] and 31 Lepetit [15] (once we have available an implementation of these methods to run the experiments). Although further experimental work is needed, the new tracking module included in PIORT has demonstrated by now to be effective under several-frames occlusions produced by an object of a class different to that of the target object. If the occluding and the target objects are recognised as belonging to the same class, then the occlusion is not detected as such, both objects are merged temporarily, but despite this behaviour, the tracking method is able in most cases to recover and track the original target when the same-class object occlusion or crossing ends. However, as observed in some of the test sequences, still there are cases where the behaviour of the tracking decision module of PIORT should be improved, particularly in the step of object re-emergence after occlusion and when other objects of similar appearance are next to the target. The upgrade of this tracking module will be subject of future research. We think that PIORT approaches for object tracking are especially suitable in noisy environments where segmented images vary so much in successive frames that it is very hard to match the corresponding regions or contours of consecutive images. The empirical results presented are quite satisfactory, despite the numerous mistakes made by the static recognition module, which can be mostly ignored thanks to the integration with the proposed tracking decision module. A right criticism that can be raised against PIORT is that too many parameters need to be set. Apart from the parameters specific of the classifier in the first (static recognition) module, the dynamic recognition module uses two parameters, which are bounds on the linear adaptive weighting of previous and current probabilities, and the tracking decision module uses up to twelve parameters: six related to the uncertainty in the target position prediction, three for tracking image post-processing filters (one for each filter), two for occlusion mode determination and one more for a weighted average computation of the target movement vector. It is very difficult to get rid of these parameters in our approach, but the default values reported in the previous sections have been tuned carefully to yield a stable satisfactory behaviour of PIORT in all the test sequences. Of course, for new sequences, these default values may not be optimal and some further tuning might improve the performance. A sensitivity analysis for each one of the PIORT parameters would be extremely hard to do and assess, since the system response may depend a lot also on the specific features of the input sequences. By experience, we hypothesize and claim that, in general, small variations on the given default values do not affect importantly the obtained tracking results, but larger ones could do. As future work, we want to extend the experimental validation of PIORT by applying it to new and more difficult image sequences; in particular, sequences where multiple objects are tracked simultaneously in the scene. And, as commented before, new comparative studies against state-of-the-art model-based tracking methods (e.g. [5, 15]) would be very interesting to do whenever possible. For the two currently used approaches in the static recognition module, an obvious upgrade is to replace the RGB by the HSI colour space, since the latter seems to be more suited for matching or tracking objects, especially in natural environments with changing illumination. In addition, we are interested in implementing and testing new 32 classifiers in the static recognition module, which could exploit other features completely different to the basic colour features used up to now. For instance, an SVM classifier could be applied to a set of features formed by Gabor filter responses, provided that class probability values were estimated from margin values. Another possible extension would be to replace in the third module the simple rules used in the a-priori predictions of target centres and areas (equivalent to noiseless Kalman filters) by the whole Kalman filter formulation considering noise for both the dynamics and the observations. However, this replacement would increase even more the number of the system parameters and it is not clear that resulted in significant changes in the whole system behaviour. Acknowledgements This research was partially supported by Consolider Ingenio 2010, project CSD2007- 00018, by the CICYT project DPI 2007-61452 and by the Universitat Rovira i Virgili (URV) through a predoctoral research grant. References [1] H. Wang, J. Peng, L. Li, “Runway detection of an unmanned landing aerial vehicle based on vision”, IJPRAI 20(8), pp: 1225-1244, 2006. [2] H-D. Yang, S-W. Lee, S-W. Lee, “Multiple Human Detection and Tracking Based on Weighted Temporal Texture Features”, IJPRAI 20(3), pp: 377-392, 2006 [3] J-W. Hsieh, Y-S. Huang, “Multiple-person tracking system for content analysis”, IJPRAI 16(4), pp: 447-462 2002. [4] A. Villanueva, R. Cabeza, S. Porta, “Gaze Tracking System Model Based on Physical Parameters”, IJPRAI 21(5), pp: 855-878, 2007 [5] S. Kang, B-W. Hwang, S-W. Lee, “ Multiple People Tracking Based on Temporal Color Feature”, IJPRAI 17(6), pp: 931-949, 2003 [6] J. Ning, L. Zhang, D. Zhang, C. Wu, “Robust Object Tracking Using Joint Color-Texture Histogram”, IJPRAI 23(7), pp: 1245-1263, 2009. [7] A. Sanfeliu, F. Serratosa, R. Alquézar, “Second-order random graphs for modeling sets of attributed graphs and their application to object learning and recognition”, Int. Journal of Pattern Recognition and Artificial Intelligence 18 (2004) 375-396. [8] A. Chatterjee, O. Ray, A. Chatterjee, A- Rakshit, “Development of a real-life EKF based SLAM system for mobile robots employing vision sensing”, Expert Systems with Applications, (38), pp: 8266–8274, 2011. [9] Li, Stan Z.; Jain, Anil K. (Eds.) “Handbook of Face Recognition”, Springer, 2005. [10] K. Burçin, N.V. Vasif, “Down syndrome recognition using local binary patterns and statistical evaluation of the system”, Expert Systems with Applications, (38), pp: 8690–8695, 2011. [11] C.Y. Tsai, Y.H. Lee, “The parameters effect on performance in ANN for hand gesture recognition system”, Expert Systems with Applications, (38), pp: 7980–7983, 2011. [12] G. L. Foresti, “Object recognition and tracking for remote video surveillance”, IEEE Trans. on Circuits and Systems for Video Technology 9 (1999) 1045-1062. [13] Francesc Moreno-Noguer, Alberto Sanfeliu, Dimitris Samaras, “Dependent Multiple Cue Integration for Robust Tracking”, IEEE Trans. Pattern Anal. Mach. Intell. 30(4), pp: 670-685, 2008. [14] H. Lee, J. Kim, H. Ko, “Prediction Based Occluded Multitarget Tracking Using Spatio-Temporal Attention”, IJPRAI 20(6), pp: 925-938, 2006. [15] C-J. Chang, J-W Hsieh, Y-S. Chen, W-F. Hu, “Tracking Multiple Moving Objects using a Level-Set Method”, IJPRAI 18(2), 2004. [16] A. Senior, A. Hampapur, Y-L. Tian, L. Brown, S. Pankanti, R. Bolle, “Appearance models for occlusion handling”, Image and Vision Computing 24 (2006) 1233-1243. [17] H.T. Nguyen, A.W.M. Smeulders, “Fast occluded object tracking by a robust appearance filter”, IEEE Trans. Pattern Anal. Mach. Intell. 26 (2004) 1099-1104. [18] S.K. Zhou, R. Chellappa, B. Moghaddam, “Visual tracking and recognition using appearance-adaptive models in particle filters”, IEEE Trans. Image Process. 13 (2004) 1491-1506. 33 [19] A.D. Jepson, D.J. Fleet, T.F. EI-Maraghi, “Robust online appearance models for visual tracking,” IEEE Trans. Pattern Anal. Mach. Intell. 25 (2003) 1296-1311. [20] K. Ito, S. Sakane, “Robust view-based visual tracking with detection of occlusions”, in: Proc. Int. Conf. Robotics Automation, 2001, vol. 2, pp. 1207-1213. [21] K. Hariharakrishnan, D. Schonfeld, “Fast object tracking using adaptive block matching”, IEEE Trans. Multimedia 7 (2005) 853-859. [22] L. Zhu, J.Zhou, J. Song, “Tracking multiple objects through occlusion with online sampling and position”, Pattern Recognition 41 (2008) 2447-2460. [23] J. Pan, B. Hu, "Robust occlusion handling in object tracking", in: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR 2007), Minneapolis, Minnesota, June 2007. [24] Z. Tu, X. Chen, A.L. Yuille, S.C. Zhu, “Image parsing: unifying segmentation, detection, and recognition”, in: Proc. Ninth IEEE Int. Conf. on Computer Vision, 2003, pp 18- 25. [25] K-C. Lee, J. Ho, M-H. Yang, D. Kriegman, “Visual tracking and recognition using probabilistic appearance manifolds”, Computer Vision and Image Understanding 99 (2005) 303-331. [26] S.C. Zhu, A. Yuille, “Region competition: unifying snakes, region growing, and Bayes/MDL for multiband image segmentation”, IEEE Trans. on Pattern Analysis and Machine Intelligence 18 (1996) 884-900. [27] J. Malik, S. Belongie, T. Leung, J. Shi, “Contour and texture analysis for image segmentation”, Int. J. Computer Vision, 43 (2001) 7-27. [28] Z. Tu, S.C. Zhu, “Image segmentation by data driven Markov chain Monte Carlo”, IEEE Trans. on Pattern Analysis and Machine Intelligence 24 (2002) 657-673. [29] F-S. Chen, C-M. Fu, C-L. Huang, “Hand gesture recognition using a real-time tracking method and hidden Markov models”, Image and Vision Computing 21 (2003) 745-758. [30] L. Maddalena, A. Petrosino, A. Ferone, “Object Motion Detection and Tracking by an Artificial Intelligence Approach”, IJPRAI 22(5), 2008. [31] N. Amezquita Gomez, R. Alquézar, F. Serratosa, “Object recognition and tracking in video sequences: a new integrated methodology“, in: Proc. 11th Iberoamerican Congress on Pattern Recognition, CIARP 2006, LNCS 4225, pp. 481-490. [32] N. Amézquita Gómez, R. Alquézar, F. Serratosa, “A new method for object tracking based on regions instead of contours”, in: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR 2007), Minneapolis, Minnesota, June 2007. [33] N. Amézquita Gómez, R. Alquézar, F. Serratosa, “Dealing with occlusion in a probabilistic object tracking method”, in: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR 2008), Anchorage, Alaska, June 2008. [34] R. Alquézar, N. Amézquita Gómez, F. Serratosa, “Tracking deformable objects and dealing with same class object occlusion”, in: Proc. Fourth Int. Conf. on Computer Vision Theory and Applications (VISAPP 2009), Lisboa, Portugal. [35] D. Comaniciu, V. Ramesh, P. Meer, “Kernel-based object tracking”, IEEE Trans. on Pattern Analysis and Machine Intelligence 25 (2003) 564-577. [36] D. Comaniciu, P. Meer, “Mean shift: a robust approach toward feature space analysis”, IEEE Trans. on Pattern Analysis and Machine Intelligence 24 (2002) 603-619. [37] R. Collins, Y. Liu, “On-line selection of discriminative tracking features”, IEEE Trans. on Pattern Analysis and Machine Intelligence 27 (2005), 1631-1643. [38] A. Bugeau, P. Pérez, “Track and cut: simultaneous tracking and segmentation of multiple objects with graph cuts”, in: Proc. Third Int. Conf. on Computer Vision Theory and Applications (VISAPP 2008), Funchal, Madeira, Portugal. [39] Y. Boykov, O. Veksler, R. Zabih, “Fast approximate energy minimization via graph cuts”, IEEE Trans. on Pattern Analysis and Machine Intelligence 23 (2001) 1222-1239. [40] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford University Press, 1995. [41] F. Yin, D. Makris, S.A. Velastin, “Performance evaluation of object tracking algorithms”, in: Proc. 10th IEEE Int. Workshop on Performance Evaluation of Tracking and Surveillance (PETS’2007). [42] S.M. Khan & M. Shah, “Tracking Multiple Occluding People by Localizing on Multiple Scene Planes”, IEEE Trans. on Pattern Analysis and Machine Intelligence 31 (2009) 505-519. [43] D. Cremers, “Dynamical Statistical Shape Priors for Level Set Based Tracking”, IEEE Trans. on Pattern Analysis and Machine Intelligence 28 (2006) 1262-1273. [44] V. Lepetit & P. Fua, “Keypoint Recognition using Randomized Trees”, IEEE Trans. on Pattern Analysis and Machine Intelligence 28 (2006) 1465-1479. [45] M. de la Gorce, N. Paragios and David Fleet. “Model-Based Hand Tracking with Texture, Shading and Self- occlusions”, IEEE Conference in Computer Vision and Pattern Recognition, 2008. [46] Yan Huang, Irfan A. Essa: Tracking Multiple Objects through Occlusions. IEEE Conference in Computer Vision and Pattern Recognition 2005: 1051-1058. [47] T.Yang, S.Li, Q.Pan, J.Li, “Real-Time multiple objects tracking with occlusion handling in dynamic scenes”, ”, IEEE Conference in Computer Vision and Pattern Recognition, 2005, 970 - 975. 
work_hslnomdpfnfvbdezomxyw5n7gy	----	PII: 0888-613X(93)90008-2 Foreword: Special Issue on Fuzzy Expert Systems In this special issue on "Fuzzy Expert Systems," six papers cover a wide range of c o n c e r n s - - f r o m theory to applications including: (1) a rule base reorganization, (2) a linear interpolation, (3) a neuro-fuzzy approach to pairwise comparison, (4) properties of reduction, in transitive matrices, (5) a consistency checking procedure, and (6) a context dependency model. We present a brief review of these papers next. In "Rule Base Reorganization and Search with a Fuzzy Cluster Analysis," Tiirksen and Jiang proposes a fuzzy cluster analysis for the purpose of restructuring a knowledge base where rules are either S- or R-implication types. The proposed algorithm can be implemented either with a composi- tional or an analogical tolerance relation. It is shown that analogical tolerance relation is more efficient in a rule base restructuring and a rule firing schema, under the proposed search scheme. In "Approximate Reasoning by Linear Rule Interpolation and General Approximation," K6czy and Hirota investigate the problem of sparse fuzzy rule bases. They review various methods of analogical reasoning available in current literature. They discuss concepts of graduality, measurability, distance, and similarity in the fuzzy sense. The fundamental equation of rule interpolation is introduced with a-cuts and resolution principle. The method is extended to multi-dimensional spaces. In "Neuro-Fuzzy Approach to Data Analysis of Pairwise Comparisons," Ichihashi and Tiirksen propose an artificial neural network model with an iterative on-line learning scheme. Two quantification methods of pairwise comparisons are presented in order to derive the associated weights. The proposed approach can also be implemented even when we have access to incomplete pairwise comparisons. The proposed approach is compared with Guttman's method and Saaty's AHP. In " O n Reduction of Transitive Fuzzy Matrices and Its Applications," Di Nola, Kolodziejczjk, and Sessa show that the class of s-trarisitive (w-transitive) fuzzy matrices properly contain the class of m a x - m i n transi- tive fuzzy matrices. It is also shown that the basic properties of the reduction models remain valid also for s-transitive fuzzy matrices. Address correspondence to L B. Tiirksen, Department o f Industrial Engineering, University of Toronto, Toronto, Ontario M5S 1A4 Canada. Received March 12, 1992; accepted March 12, 1993. International Journal of Approximate Reasoning 1993; 9:165 166 © 1993 Elsevier Science Publishing Co., Inc. 655 Avenue of the Americas, New York, NY 10010 0888-613X/93/$6.00 165 166 I.B. Tfirksen et al. In "Consistency Checking for Fuzzy Expert Systems," Leung and So introduce an affinity measure based on a similarity measure and apply this measure for consistency checking in mixed fuzzy and non-fuzzy expert system environments. In "The Context Model: A n Integrating View of Vagueness and Uncer- tainty," Gebhardt and Kruse address two different types of partial igno- rance, i.e., vagueness and uncertainty. In this paper they restrict their presentation to Bayes Theory and D e m p s t e r - S h a f e r Theory. "Context Model" applications are illustrated for the spoiled sandwich effect, the three prisoners problem, and the unreliable alarm paradigm. We hope that our readers will find these topics of some interest to their ongoing theoretical and applied investigations. I. Burhan Tfirksen, Hideo Tanaka, and Junzo Watada 
work_htbmyqzu2zejlk55ix3okhxnue	----	NASA Technical Reports Server (NTRS) 
work_htfht4hrybdz3jty5t7367hreu	----	warwick.ac.uk/lib-publications Manuscript version: Author’s Accepted Manuscript The version presented in WRAP is the author’s accepted manuscript and may differ from the published version or Version of Record. Persistent WRAP URL: http://wrap.warwick.ac.uk/109440 How to cite: Please refer to published version for the most recent bibliographic citation information. If a published version is known of, the repository item page linked to above, will contain details on accessing it. Copyright and reuse: The Warwick Research Archive Portal (WRAP) makes this work by researchers of the University of Warwick available open access under the following conditions. © 2018, Elsevier. Licensed under the Creative Commons Attribution-NonCommercial- NoDerivatives 4.0 International http://creativecommons.org/licenses/by-nc-nd/4.0/. Publisher’s statement: Please refer to the repository item page, publisher’s statement section, for further information. For more information, please contact the WRAP Team at: wrap@warwick.ac.uk. http://go.warwick.ac.uk/lib-publications http://go.warwick.ac.uk/lib-publications http://wrap.warwick.ac.uk/109440 http://creativecommons.org/licenses/by-nc-nd/4.0/ mailto:wrap@warwick.ac.uk Hierarchical Viewpoint Discovery from Tweets Using Bayesian Modelling Lixing Zhua, Yulan Heb, Deyu Zhoua,∗ aSchool of Computer Science and Engineering, Southeast University, China bDepartment of Computer Science, University of Warwick, UK Abstract When users express their stances towards a topic in social media, they might elaborate their viewpoints or reasoning. Oftentimes, viewpoints expressed by different users exhibit a hierarchical structure. Therefore, detecting this kind of hierarchical viewpoints offers a better insight to understand the public opinion. In this paper, we propose a novel Bayesian model for hierarchical viewpoint discovery from tweets. Driven by the motivation that a viewpoint expressed in a tweet can be regarded as a path from the root to a leaf of a hierarchical viewpoint tree, the assignment of the relevant viewpoint topics is assumed to follow two nested Chinese restaurant processes. Moreover, opinions in text are often expressed in un-semantically decomposable multi-terms or phrases, such as ‘economic recession’. Hence, a hierarchical Pitman-Yor process is employed as a prior for modelling the generation of phrases with arbitrary length. Experimental results on two Twitter corpora demonstrate the effectiveness of the proposed Bayesian model for hierarchical viewpoint discovery. Keywords: Natural language processing, Opinion mining, Bayesian modelling 1. Introduction Stance classification aims to predict one’s stance in a two- sided debate or a controversial hot topic and has been inten- sively studied in recent years (Hasan and Ng, 2013; Elfardy et al., 2015). However, apart from detecting one’s stance, we are more interested in figuring out the reasons or key viewpoints why the person supports or opposes an issue of interest. More- over, viewpoints expressed by different users could be related and exhibit a hierarchical structure. Figure 1 illustrates an ex- ample hierarchical viewpoint tree, in which both User A and User B are supporters of Trump. However, the former just ex- pressed his support for Trump without mentioning any reasons behind, while the latter stated the reason that Trump is a charis- matic leader. We can also see that both User C and User D support Trump due to his economic policy, however with differ- ent reasons (‘higher employment rate’ vs. ‘trade protection’). Such a hierarchical viewpoint tree enables a better understand- ing of user opinions and allows a quick glimpse of reasons be- hind users’ stances. Mining hierarchical viewpoints from tweets is challenging for the reasons below: (1) In comparison with that the stance is either ‘Support’ or ‘Oppose’, the hierarchical structure of viewpoints is unknown a priori; (2) People tend to express their opinions in many different ways and even with infor- mal or ungrammatical language; (3) Opinion expressions often contain multi-word phrases, for example, ‘economic recession’ and ‘economic growth’. Simply decomposing phrases into uni- grams may lose their original semantic meaning. As such, sim- ∗Corresponding author. Fax.: 8602552090861. Email addresses: zhulixing@seu.edu.cn (Lixing Zhu), y.he@cantab.net (Yulan He), d.zhou@seu.edu.cn (Deyu Zhou) Trump running for president of the United States Support Oppose RacismPersonal charisma Vote for #Trump. User A Economic policy ... ... #DonaldTrump is completely disgusting for his racist. I am impressed with your patriotism and honesty, there needs to be more like you on capitol hill. User B Higher employ- ment rate User C User D Trade protec- tion Figure 1: An example of a hierarchical viewpoint tree on the topic “Trump run for election” from Twitter. ply applying a bag-of-words topic model might wrongly group them under the same topic due to the shared word ‘economic’. To tackle these challenges, in this paper, we propose a Bayesian model, called Hierarchical Opinion Phrase (HOP) model, for hierarchical viewpoint discovery from text. In such a model, the root node (level-1) contains the topic of inter- est (e.g., ‘Trump run for president’) and the level-2 topics in- dicate stance (e.g., either ‘Support’ or ‘Oppose’), while top- ics in the level-3 and below contain viewpoints under differ- ent stances. Assuming that viewpoints in each tweet are gen- erated from a path from the root to a leaf of a hierarchical viewpoint tree, the assignment of viewpoint topics can be re- garded as following two nested Chinese Restaurant Processes (nCRPs). Furthermore, a hierarchical Pitman-Yor process is employed as a prior to model the generation of phrases with October 8, 2018 arbitrary length. We have also explored various approaches for incorporating prior information such as sentiment lexicons and hashtags in order to improve the stance classification accu- racy. To the best of our knowledge, our work is the first attempt for hierarchical viewpoint detection. The proposed approach has been evaluated on two Twitter corpora. Experimental re- sults demonstrate the effectiveness of our approach in compar- ison with existing approaches for hierarchical topic detection or viewpoint discovery. Our source code is made available at https://github.com/somethingx86/HOP. 2. Related Work In this section, we give a brief review of four related lines of research: opinion mining based on topic models, hierarchical topic extraction, topical phrase models and deep learning for sentiment analysis. 2.1. Opinion Mining based on Topic models Topic models such as Latent Dirichlet Allocation (LDA) (Blei et al., 2003) have been proven effective for opin- ion mining. Lin and He (2009) proposed a Joint Sentiment Topic (JST) model that extends the standard LDA by adding a sentiment layer on top of the topic layer to allow the extrac- tion of positive and negative topics. A variant of JST called reverse-JST was studied in (Lin et al., 2010), where the gen- eration of sentiments and topics is reversed. Kawamae (2012) separated words into aspect words, sentiment words and other words. Aspect words were generated dependent on latent as- pects which were in turn sampled from sentiment-associated topics. Kim et al. (2013) modified the aforementioned model by using a recurrent Chinese Restaurant Process (rCRP) prior on the aspect variable. They mined a hierarchy of aspects from product reviews and associated each aspect with a positive or negative sentiment. In addition to online reviews, the LDA-based models have also been applied on debate forums and Twitter. Lim et al. (2014) proposed a Twitter opinion topic model which makes use of target-opinion pairs. Trabelsi et al. (2014) designed a joint topic viewpoint model by assuming that the distribu- tion over viewpoints is associated with latent topics. Thonet et al. (2016) treated nouns as topic words while adjectives, verbs, adverbs were treated as opinion words for topic-specific opin- ion discovery. Vilares and He (2017) focused on detecting per- spectives from political debates. They modelled topics and their associated perspectives as latent variables and generated words associated with topics or perspectives by following different routes. However, the aforementioned models are not able to generate a hierarchy of opinions. 2.2. Hierarchical Topic Extraction In general, there are two types of approaches for extract- ing topical hierarchies. The first one is based on probabilistic graphical models. For example, the Hierarchical LDA (HLDA) model was proposed for discovering the topical hierarchies from abstracts of scientific papers (Blei et al., 2010). In the model, each document is assumed to be attached to a path where each level is a topic. The path will induce a set of topics, and words will be generated in the same way as in LDA. The allo- cation of paths follows an nCRP prior. Kim et al. (2012) argued that topics should be distributed over the whole nodes of hier- archy. They achieved this by placing an rCRP prior on the tree. Jordan et al. (2015) proposed a nonparametric model called nested hierarchical Dirichlet process to allow shared groups among clusters. Their work extended nCRP by incorporating a hierarchical Dirichlet process. The second type of approaches to hierarchical topic extrac- tion is based on frequent pattern mining. An early work is fre- quent itemset-based hierarchical clustering model (Fung et al., 2003) which simply clustered documents according to their shared items. Wang et al. (2013) developed a phrase min- ing framework called CATHY (Constructing A Topical Hier- archY). It first builds a term co-occurrence network using a fre- quent pattern mining method which is commonly used in as- sociation rule mining. The initial network corresponds to the root topic. Then the network is clustered into subtopic net- works in a probabilistic way by assuming that each term co- occurrence was generated by a topic. The process is repeated until no subtopics can be found. 2.3. Topical Phrase Models In order to address the wide occurring of un-semantically de- composable phrases, various topical phrase models have been proposed in the literature. Wallach (2006) made the first attempt to extend LDA by incorporating hierarchical Dirichlet language model called Bigram Topic Model (BTM), where each word is generated from a distribution over vocabulary following a two- level hierarchical Dirichlet process. Wang et. al (2007) ex- tended BTM by adding a switch variable at each word position to decide whether to begin a new n-gram or to continue from the previously identified n-gram. El-Kishky et al. (2014) de- veloped a pipeline approach, called TopMine. It first extracts frequent phrases using a frequent pattern mining method. Then an LDA-based model is learned where words in the same phrase are generated from the same topic. He (2016) proposed a top- ical phrase model which extends TopMine by using the hierar- chical Pitman-Yor Process (HPYP) to model the generation of words in a phrase. 2.4. Deep Learning for Sentiment Analysis Recent years have seen a surge of interests of developing deep learning approaches for sentiment analysis. Many of them have been applied to sentiment classification on product re- views (Chen et al., 2016; Gui et al., 2017), news articles (Lai et al., 2015; Nguyen et al., 2017) as well as tweets (Ghiassi et al., 2013; Dong et al., 2014). Different neural network ar- chitectures have been explored including Convolutional Neural Network (CNN) models (dos Santos and Gatti, 2014; Severyn and Moschitti, 2015a,b), Long Short-Term Memory (LSTM) Networks (Tang et al., 2015; Wang et al., 2016) and models with attention mechanisms (Ren et al., 2016; Yang et al., 2016). However, these models rely on annotated datasets where each October 8, 2018 https://github.com/somethingx86/HOP γ α cd,3 cd,4 cd,L n[1,Nd] c1 cd,2λ .. . d[1,D] Gu k wd,n,1 θd k[1,] ... ... zd,n,1 zd,n,2 wd,n,2 zd,n,Ud,n wd,n,Ud,n u[1,Ud,n] G0 U ηβk k[1,] d[1,D] cd,1 cd,2 cd,3 cd,L γ α θd n[1,Nd] zd,n wd,n .. . (a) HLDA (b) HOP Figure 2: Graphical models of HLDA and the proposed Hierarchical Opinion Phrase (HOP) model. Boxes are plate notations representing replicates. Table 1: Notations Used in the Article. Symbol Description βk Topic k, which is a distribution over the vocabulary cd,l The lth level, whose value indexes a topic and follows CRP cd The path in the tree for tweet d, which follows nCRP θd Distribution over the levels for tweet d α Parameter for the Dirichlet distribution γ Concentration parameter for the CRPs η Parameter for the symmetric Dirichlet distribution λ Parameter for the Bernoulli distribution wd,n The token in tweet d, position n (HLDA) zd,n Level allocation for token wd,n (HLDA) wd,n,u The token in tweet d, phrase n, position u (HOP) zd,n,u Level allocation for token wd,n,u (HOP) G0 Base distribution for the HPYPs Gk Topic k, which is an HPYP Gku The PYP that generates the uth token in phrases of topic k document is either labeled by the sentiment class or annotated with more fine-grained opinion targets/words for training. As such, they cannot be used for hierarchical opinion discovery in the absence of annotated data. 2.5. Summary Our model is partly inspired by HLDA. While the root-level topics in HLDA mostly contain background words, our model is able to extract the key topic of interest from tweets at the root- level. Also, the level-2 topics in our model are constrained to be stance-related topics to allow for the incorporation of prior information as will be shown in the experiments section. Fur- thermore, we incorporate the hierarchical Pitman-Yor Process (HPYP) as the prior to deal with the generation of multi-word phrases and hence the proposed model can generate hierarchi- cal viewpoints with better interpretability. 3. Hierarchical Opinion Phrase (HOP) Model In this section, we propose the Hierarchical Opinion Phrase (HOP) model to learn a viewpoint hierarchy from text and also model the generation of phrases at the same time. Before pre- senting the details of HOP, we first describe the HLDA model. The notations used in our model and HLDA are summarized in Table 1. HLDA (Blei et al., 2010) as illustrated in Figure 2(a) as- sumes that each topic is tied to a node in a tree and the doc- uments are generated by first selecting a path in the tree, then choosing a topic at each level of the path and finally drawing words from the assigned topics. Notations c in Figure 2(a) are random variables indexing topics β. The per-tweet path cd = {cd,1, cd,2, . . . , cd,L} follows nCRP, whose valuing behavior functions as the affiliated tweet’s seating process, which will be described later in the introduction of nCRP. The arrows linking c indicate the constraint that the path’s lower-level node could only take the value of indices of the topics in the restaurant the upper level is pointing to, i.e., a tweet only counts those tweets whose upper level takes the same value in the nCRP of the gen- erative process. One problem of applying HLDA for viewpoint discovery is its unconstrained number of topics and the mixture of topics under different stances. The problem is aggravated when the corpus is noisy. We modify HLDA by placing a root topic that is shared by all documents to generate common words. Since the number of stances is fixed in our data (“Support” or “Op- pose”), a stance latent variable that has only two possible values is placed at level-2. The stance level will therefore hopefully separate two sets of opposing viewpoints. Since the number of level-2 topics is fixed at 2, it is now possible to incorporate side information into the model such as hashtags indicating the stances, which will be shown in the Experiments section. An- other problem with the original HLDA is that it operates on the bag-of-words assumption. As such, the topic results are less in- terpretable since many phrases are not semantically decompos- able. We therefore propose to generate phrases from the mod- ified HLDA model by incorporating the HPYP into the gener- ative process. HPYP has been previously explored for single- level topical phrase extraction in newswire stories and clinical October 8, 2018 documents (Lindsey et al., 2012; He, 2016). But it has never been explored for hierarchical topical phrase extraction. For comparison, the HOP model is illustrated in Figure 2(b) in which the root topic (level-1 topic) contains the topic of in- terest that is shared across all documents and the level-2 is a stance layer that has only two possible topics, either ‘Support’ or ‘Oppose’. Topics in level-3 and below capture viewpoints under different stances. The proposed approach assumes that phrases have been identified prior to model learning. Phrase extraction can be done by many different approaches. In this paper, we ex- tract phrases from data using an open source toolkit called Gensim.models.phrases1. It first discovers candidate phrases based on word collocation patterns, then transforms the phrases into distributed representations, and finally filters out irrelevant phrases (Mikolov et al., 2013). In the following subsections we will discuss the HOP model in more details. 3.1. Generative Process Suppose there is an L-level hierarchical opinion tree T as shown in Figure 1 where L is fixed, and each tweet contains L latent topics corresponding to a path from the root node to a leaf node. For example, for the tweet “I am impressed with your pa- triotism and honesty, there needs to be more like you on capitol hill.”, its hierarchical topics are [Trump run for election, sup- port, personal charisma]. The root node c1 of T is shared by all tweets in the collection. The second level of T is limited to have two topics corresponding to two stances, ‘Support’ or ‘Oppose’, which is assumed to follow a Bernoulli distribution parameterized by λ. The value of lower levels follows an nCRP prior, which can depict the unbound nature of viewpoints. We use cd = {cd,1, cd,2, . . . , cd,L} to denote the path assigned to the dth tweet. Therefore, the prior for each level can be expressed as cd,2 ∼ Bernoulli(λ), cd,l ∼ CRP(γ, cd,l−1, . . . , cd,2). nCRP. We briefly describe the nCRP (Blei et al., 2010) here. The nCRP is employed as the prior for the assignment of top- ics, which can organize the topics into a tree topology. As il- lustrated in Figure 3, tweets are analogous to customers and topics are analogous to tables. Assuming that the nth customer enters the root restaurant, he will choose an existing table βi proportional to the number of customers already sitting there or choose a new table proportional to the concentration parameter γ, that is p(occupied table i|other customers) = ni n − 1 + γ , p(new table i|other customers) = γ n − 1 + γ , where ni is the number of customers seated at table βi, n is the total number of customers including the present one, and γ is the concentration parameter normally set to 0.5, which indicates how likely the customer will sit in a new table. After choosing the table βi, the customer will proceed to choose a table in the lower-level restaurant which is pointed to by Table βi. This 1http://radimrehurek.com/gensim/models/phrases.html#id1 process will continue until the customer reaches depth L. As a result, a path composed of tables (or topics) from L different levels will be induced. This forms L hierarchical topics. HPYP. We detect phrases based on word collocations so that phrase detection is separated from topic inference. In HOP, phrases will be assigned with the same topic if they are assigned with the same level and their involved tweets happen to share the same table in that level. We use an HPYP to model the gen- eration of phrases assigned with the same topic. HPYP was first proposed in (Teh, 2006), which has been proven to be the best smoothing method for n-gram language models. In HPYP, the distribution of the first word is generated from PYP defined as G1 ∼ PYP(a0, b0, G0), where G0 is a uniform distribution over a fixed vocabulary, a0 is the discount parameter indicating the degenerating rate of to- generate power law distribution and b0 is the concentration pa- rameter controlling the amount of variability of G1 around G0. The distribution of the second word is generated using G1 as the base distribution, G2 ∼ PYP(a1, b1, G1). The process con- tinues until the end of the phrase. The generative process of drawing words from the prior can be simulated by the gener- alised CRP (Pitman et al., 2002). A restaurant corresponds to each Gu, where each table is served a dish whose value is cho- sen from the base distribution Gu−1. The first customer sits at the first table; the n + 1th customer chooses an occupied table in proportion to the number of customers already sitting there and takes the value of dish on the table, or chooses a new table proportional to a constant parameter and orders a dish from the base distribution. This process continues until the proxy cus- tomer sits in an existing table or there is no parent restaurant. As such, the probability of the uth word given the seating ar- rangement is p(w|Λu) = Cuw − au−1Tuw Cu· + bu−1 + au−1Tu· + bu−1 Cu· + bu−1 × p(w|Λu−1), where Cuw is the number of customers having dish w in the restaurant u, Tuw denotes the number of tables serving dish w in restaurant u, Cu· = ∑ w Cuw and Tu· = ∑ w Tuw. It is worth noting that the restaurant setup here is different from that in nCRP. In nCRP each tweet is a customer and each topic is a table. A tweet is assigned with L topics (or tables) from the root to the leaf. In HPYP for phrase generation, each word w is a customer and a restaurant u is the context of the word. For example, for an n-gram phrase, the context of the nth word is its preceding n − 1 words. Based on the above description, the generative process of HOP is given below. • For each topic k ∈ {1, 2, 3, . . . ,∞}: – Gk1 ∼ PYP(a0, b0, G0) – Gk2 ∼ PYP(a1, b1, G k 1) · · · – GkU ∼ PYP(aU−1, bU−1, G k U−1) • Set c1 to be the root restaurant • For each tweet d ∈ {1, 2, 3, . . . , D}: October 8, 2018 http://radimrehurek.com/gensim/models/phrases.html#id1 β 1 β 2 β 3 β 11 β 12 β 13 d3 d5 d9 d1 d2 d7 d8 d4 d6 d9 d5 d3 d1 d2 d7 d8 β 21 β 22 β 23 β 31 β 32 β 33 d4 d6 R 0 R 1 R 2 R 3 Figure 3: Illustration of the nested Chinese Restaurant Process (nCRP). Each circle represents a table which points to a unique restaurant denoted as a rectangle. Each customer (or tweet) will first choose a table in the upper level then follow the link pointed to by the table to reach a lower-level restaurant and choose another table there. – Select the Level-2 topic cd,2 ∼ Bernoulli(λ) – For level l ∈ {3, . . . , L}, select its corresponding topic cd,l ∼ CRP(γ, cd,l−1, . . . , cd,2) – Draw a distribution over levels θd ∼ Dirichlet(α) – For each phrase n ∈ {1, 2, . . . , Nd}: * For each word u ∈ {1, 2, . . . , Ud,n}: If it is the first word in a phrase (u = 1): · Assign a level zd,n,u|θd ∼ Discrete(θd ) · Draw a word wd,n,u|{zd,n,u, cd, G} ∼ Discrete(G cd [zd,n,u ] 1 ) Else: · Set zd,n,u = zd,n,u−1 · Draw a word wd,n,u|{zd,n,u, cd, G} ∼ Discrete(G cd [zd,n,u ] u ). Here, cd [zd,n,u] denotes the zd,n,uth component of vector cd , G0 is a uniform distribution over a fixed vocabulary W of V words, for ∀w ∈ W, G0(w) = 1 V . 2 Since phrases have already been identified prior to hierarchical opinion extraction, the bound- aries of phrases are observed and need not be sampled from data. 3.2. Inference and Parameter Estimation Given the observed variable w = {w1, w2, . . . , wD}, our goal is to infer the hidden variables c and z using the posterior distri- bution p(c, z|w,λ,γ,α, G0, Ω) where Ω = {a0, b0, . . . , aU, bU} denotes the hyper-parameters of HPYPs. Since exact infer- ence of the posterior distribution is intractable, Gibbs sam- pling (Griffiths and Steyvers, 2004) is employed to approximate the hidden variables, which sequentially samples each variable of interest using the conditional probability of that variable given the current values of all other variables and the data. With sufficient iterations, the sampling process will finally reach a status when all the samples can be seen as generated by a sta- tionary distribution. The variables used in the sampling process are: (1) zd,n,u, the level allocation of the uth word in the nth phrase of the dth tweet; (2) cd , the path of the dth tweet. The objective of the Gibbs sampler is to approximate p(c, z|w,λ,γ,α, G0, Ω). We omit Ω for clarity. In Gibbs sampling, we are focused 2Note that we abuse the notation here that we use n to denote the nth phrase. on the per-document posterior p(cd, zd|c−d, z−d, w,λ,γ,α, G0). For d ∈ {1, 2, . . . , D}, we first sample word-wise level alloca- tion p(zd,n,u|c, z−d,−n,−u, w,α, G0), where z−d,−n,−u is the vector of level allocations leaving out zd,n,u. Then we sample path p(cd|c−d, z, w,λ,γ, G0). 3.2.1. Level Allocation Sampling Given a path, we need to sample level allocations for the tweet d. According to Bayes’ rule and conditional indepen- dence, we obtain p(zd,n,u = l|c, z−d,−n,−u, w,α, G0) ∝ p(zd,n,u = l|zd,−n,−u,α)× p(wd,n|c, z, w−d,−n, G0). (1) The first term in Eq. 1 is a conditional probability marginal- izing out θd . Since the Dirichlet distribution p(θd|α) and the discrete distribution p(zd,n,u|θd ) form a Dirichlet-Multinomial conjugate, we have p(zd,n,u = l|zd,−n,−u,α) = Cld,−n,−u + α[l] C·d,−n,−u + ∑L l=1 α[l] . Here, Cld,−n,−u is the number of times level label l being assigned to some word tokens in tweet d leaving out zd,n,u, C·d,−n,−u =∑L l=1 C l d,−n,−u, L is the total depth of the hierarchy. The second term in Eq. 1 follows a zd,n,u-specified HPYP with wd,n as a new random variable given the particular status of that HPYP. We use the generalised CRP to perform sam- pling. Let Λcd [l] denote the current seating arrangement of topic cd [l], the second term is rewritten as p(wd,n|Λcd [l]). Thereafter, the random process can be simulated by the situation that the first word wd,n,1 enters the restaurant Λ cd [l] 1 as a customer. The probability of the next word w from Gcd [l]u can be calculated re- cursively as p(w|Λcd [l]u ) = Ccd [l]uw − au−1T cd [l] uw Ccd [l]u· + bu−1 + au−1T cd [l] u· + bu−1 Ccd [l]u· + bu−1 ×p(w|Λcd [l]u−1 ). Here, Ccd [l]uw denotes the number of customers eating dish w in the restaurant u owned by topic cd [l], T cd [l] uw denotes the number of tables serving dish w in restaurant u owned by topic cd [l], October 8, 2018 Ccd [l]u· = ∑ w C cd [l] uw and T cd [l] u· = ∑ w T cd [l] uw . The recursion ends when u = 1, that is p(w|Λcd [l]1 ) = Ccd [l]1w − a0T cd [l] 1w Ccd [l]1· + b0 + a0T cd [l] 1· + b0 Ccd [l]1· + b0 1 V . If wd,n,u is not the first word of a multi-term phrase, we just take zd,n,u = zd,n,u−1 and do not sample a new topic. 3.2.2. Path Sampling Given paths of other tweets and level allocations, we have to sample the path for tweet d. By applying the Bayes’ rule, we have p(cd|c−d, z, w,λ,γ, G0) ∝ p(cd|c−d,λ,γ) × p(wd|c, z, w−d, G0). (2) The first term in Eq. 2 is a prior over paths. It can be com- puted by first calculating a Bernoulli distribution then calculat- ing each level’s seating distribution in the corresponding restau- rant. The second term in Eq. 2 is the probability of a given tweet under a possible seating arrangement in the nCRP tree. It can be decomposed into probabilities of phrases/words occurred in the tweet, p(wd|c, z, w−d, G0) = p(wd,1|c, z, w−d, G0)× p(wd,2|c, z, w−d, wd,1, G0)× ·· · p(wd,Nd |c, z, w−d, wd,1, . . . , wd,Nd−1, G0). (3) Each term in Eq. 3 is a posterior distribution of phrase wd,n conditioned on other phrases allocated with the same topic. The distribution follows an HPYP, and thus can be sampled in the same way as described for the second term in Eq. 1. 3.2.3. Complete Sampling Procedure Algorithm 1 Gibbs sampling for HOP 1. Initialize the model by arbitrarily assigning a path to each tweet. Randomly assign a level number of the path to each word/phrase in the tweet. Initialize the HPYP configuration within each topic for each associated word/phrase. 2. For each tweet d ∈ {1, 2, . . . , D}: (a) Sample c(t+1)d using Eq. 2. (b) For each phrase n in tweet d, n ∈ {1, 2, . . . , Nd}: i. For each word u in phrase n, u ∈ {1, 2, . . . , Ud,n}: A. Sample z(t+1)d,n,u using Eq. 1. 3. Repeat step 2 until the global log-likelihood converges or a fixed number of iterations is reached. 4. Output the final sample {c, z}. Given the conditional distributions defined above, we are able to perform the full Gibbs sampling. Let {c(t), z(t)} denote the current state, the sampling process is described in Algo- rithm 1. 4. Experiments In this section, we first describe the datasets and baselines used in our experiments. We then evaluate HOP against the baselines quantitatively and qualitatively. 4.1. Experimental Setup Algorithm 2 Opinion tweets retrieval. 1. Define the seed patterns “[brexit|leaving the EU|staying in the EU]+ [would|can|might|won’t|can’t]” for Dataset I (Brexit), and “[Trump |Hillary|Donald Trump|Hillary Clinton]+[would|can|might|won’t|can’t]” for Dataset II (US General Election). These seed patterns are used for retriev- ing seed tweets such as “Top economic think tanks agree that brexit would harm economy. Vote Remain in the EU” or “Trump would start WW3! There is no one he hasn’t of- fended”. We denote the set of the seed patterns as P and the set of opinion tweets as T . 2. (a) Perform Part-of-Speech (POS) tagging on the seed tweets T . (b) Extract keywords which are tagged as the latter noun in the POS tag pattern “NN+MD+VB+NN”. (c) Enlarge the seed pattern set P by adding new rules based on the newly extracted keywords such as “keyword+[would|can|might|won’t|can’t]”. (d) Retrieve tweets based on the enlarged seed pattern set. Add the retrieved tweets to T . 3. Repeat step 2 until no more tweets are found. To evaluate HOP, two datasets are constructed. We devel- oped a crawler3 using the Twitter4j toolkit and Twitter Stream- ing API4. The crawler listens to the global timeline and fil- ters tweets with specific keywords. Dataset I contains 515, 113 tweets crawled using hashtags such as #EURef, #EU refer- endum and #Brexit, dated 14th-24th June 2016. Dataset II contains 1, 691, 294 tweets with hashtags #PresidentElection, #Election2016, #Hillary, #Trump, dated 1st-8th November 2016. To ensure tweets with opinions are kept, a rule-based method inspired by (Handler et al., 2016) is used and is de- scribed in Algorithm 2. The process runs iteratively until no more tweets are found. The statistics of the final datasets are shown in Table 2. It can be noticed that only 13.33% of tweets are kept in Dataset I and for Dataset II the proportion is 2.13%. Each tweet is pre-processed by removing common stop words. No stemming is used. Phrases are identified based on collocation patterns from data using Gensim5. We perform Gibbs sampling for a maximum of 10,000 iterations and output the intermediate results every 1,000 iterations. It usually takes 3https://github.com/somethingx86/EclipseTwitterStreamer 4https://developer.twitter.com/en/docs 5http://radimrehurek.com/gensim/models/phrases.html October 8, 2018 https://github.com/somethingx86/EclipseTwitterStreamer https://developer.twitter.com/en/docs http://radimrehurek.com/gensim/models/phrases.html Table 2: Statistics of the two datasets. Dataset Property Value I #tweets 68,672 #unigram tokens 1,219,484 vocabulary size 31,149 II #tweets 36,013 #unigram tokens 591,620 vocabulary size 21,595 around 1,000 iterations to reach a stationary status when the topic topology and tweet-topic associations no longer change, which can be visualized using a tree structure. 4.2. Methods for Comparison We compare our model with the following approaches which can generate topic hierarchies: CATHY (Constructing A Topical HierarchY) (Wang et al., 2013) builds a topical hierarchy where each topic is represented as a set of phrases. A term co-occurrence network is first con- structed using the Frequent Pattern (FP)-growth algorithm. The network edges are clustered according to their associated topics to obtain a hierarchy of topics. In our experiments, the hierar- chy depth is set to 3, the cluster number is set to 2 for level 2 and 4 for level 3. Other parameter settings follow (Wang et al., 2013). HLDA (Hierarchical LDA) (Blei et al., 2010) assumes that each document is generated by drawing an infinite L-level path ac- cording to the nCRP prior, and drawing a topic distribution over levels in the path according to a stick-breaking process. Words are drawn from the L topics which are associated with the restaurants along that path. HASM (Hierarchical Aspect-Sentiment Model) (Kim et al., 2013) produces hierarchical aspects in which each aspect is as- sociated with positive, negative or neutral polarities. The aspect hidden variable follows an rCRP. We use the default parame- ter settings in our experiments. Note that the prior sentiment knowledge from some common sentiment seed words is used to set the asymmetric Dirichlet prior for aspect-sentiment-word distributions in HASM. In our experiments, its default senti- ment lexicon is replaced by Sentiment140Lex6 (Mohammad et al., 2013) which was specifically constructed for Twitter sen- timent analysis. For HOP, the parameter settings are: λ = 0.5, γ = 0.5, α = [0.75, 0.15, 0.075, 0.025], au = 0.8, bu = 1 for HPYP, the maximum phrase length U = 3. We only keep topics which are associated with at least 1,000 tweets. 4.3. Topic Coherence Various measures have been proposed to evaluate the qual- ity of the topics discovered. Newman et al. (2010) found that the pointwise mutual information (PMI) of all word pairs in a 6http://www.saifmohammad.com/Lexicons/Sentiment140-Lexicon-v0.1. zip topic’s top ten words coincides well with human judgements. PMI is defined as follows: PMI(wi, w j) = log p(wi, w j) p(wi) p(w j) , (4) where p(wi, w j) is the co-occurrence likelihood of two words. It can be estimated by counting the co-occurrence of the word pair in sliding windows in an external large meta-document. Röder et. al (2015) studied the known coherence measures and pro- posed a new measure which was a combination of some exist- ing ones. Particularly, this metric first retrieves co-occurrence counts for the given words using a sliding window of size 110 in Wikipedia. For each top word a vector is built whose compo- nents are the normalized Point-wise Mutual Information (PMI) between the word and every other top words. The arithmetic mean of all vector pairs’ cosine similarity is treated as the co- herence measure of a given topic. I II Dataset 0.30 0.35 0.40 0.45 0.50 T o p ic C o h e re n c e 0.3796 0.4033 0.3979 0.4269 0.3946 0.4151 0.4207 0.4383 CATHY HLDA HASM HOP Figure 4: Topic coherence on two datasets. Following the measure proposed in (Röder et al., 2015), we report the average topic coherence scores computed on the top 10 words/phrases of each topic which is shown in Figure 4. It can be observed that CATHY scores the lowest compared to the other three methods on both datasets. HLDA gives better results compared to HASM. HOP outperforms all the other methods and the improvement over the second best performing model, HLDA, is more prominent on Dataset I. 4.4. Stance Classification Some previous studies used topic models to perform sentiment/stance classification and achieved comparable re- sults (Trabelsi and Zaıane, 2014; Thonet et al., 2016). Apart from HLDA and HASM, we select two more baselines which can output document-level stance labels: VODUM (Viewpoint and Opinion Discovery Unification Model) (Thonet et al., 2016) assumes nouns are topical words and adjectives, verbs and adverbs are opinion words. It uses dif- ferent generative routes to generate topical words and opinion words. Their model associates a viewpoint label (equivalent to a stance label here) with each document. October 8, 2018 http://www.saifmohammad.com/Lexicons/Sentiment140-Lexicon-v0.1.zip http://www.saifmohammad.com/Lexicons/Sentiment140-Lexicon-v0.1.zip sLDA (Supervised LDA) (Mcauliffe and Blei, 2008) modified LDA by adding an observed response variable to each docu- ment, which follows a Gaussian distribution whose mean is the averaged weighed sum of topic assignments of all the docu- ment’s tokens. To obtain the ground-truth for the evaluation of stance clas- sification, we hired three senior undergraduate students, who worked on NLP-related final-year projects, to manually anno- tate some randomly selected tweets from Dataset II. Tweets were discarded if there were disagreement among the annota- tors. In total, 80 tweets were discarded. In the end, we kept a to- tal of 1,000 tweets which consists of 748 positive and 252 neg- ative tweets. We compare HOP with baselines on this dataset for stance classification. For HOP, we use the level-2 topics for stance classification. For HLDA, there is no restriction on its level-2 topic number and we manually go through all the topics in level-2 to identify the likely ‘Support’ and ‘Oppose’ stances. Afterwards, each document’s stance can be identified accord- ingly. For VODUM, the number of viewpoints (stances) is set to 2 and all the hyperparameters take the default setting. We run each approach for 20 times and average the classification results over 20 such runs. HASM used the prior sentiment knowledge to set the asym- metric Dirichlet prior for aspect-sentiment-word distributions. In more details, it placed higher probabilities on some polarity seed words based on a given sentiment lexicon. The incorpora- tion of this kind of supervised information gave them an edge over the unsupervised alternatives. In order to compare our model with HASM fairly, we also experimented with a variant of HOP (call ‘pHOP’) by incorporating the prior information of stance and sentiment into the model. The prior information can be derived from either hashtags or existing sentiment lexi- cons. For hashtags, we use ‘#VoteTrump’ and ‘#NeverHillary’ as a proxy label for the ‘Support’ stance, and ‘#VoteHillary’ and ‘#NeverTrump’ as a proxy label for the ‘Oppose’ stance. In the experiments, the prior information was only utilized during model initialization, where the ‘Support’ and ‘Oppose’ tweets were assigned to the left and right path in Level-2, respectively. Furthermore, we also utilize the sentiment prior information of words which is obtained from Sentiment140Lex. We consider those with sentiment intensity larger than 1.0 as strong sen- timent words. We restrict those words with strong sentiment from Sentiment140Lex to be only generated from the Level-2 topics. During model initialization, if a tweet has a stance la- bel or a word token can be found in the sentiment lexicon, then the path or level allocation will be kept. Otherwise, a path or level allocation will be randomly initialized. There are about 10% tweets carrying proxy labels, and 1.15% words found in the sentiment lexicon in our data. For sLDA, proxy stance la- bels are treated as the observed response variables to train the model. Stance classification results are presented in Figure 5. HLDA, VODUM and HOP are all unsupervised approaches without the prior information. It can be observed that HLDA gives better results compared to VODUM but VODUM tends to be more stable. HOP outperforms both HLDA and VODUM. Although the best performance achieved is only about 58%, this VODUM HLDA HOP sLDA HASM pHOP Models 0.50 0.55 0.60 0.65 0.70 0.75 0.80 A c c u ra c y Figure 5: Stance classification accuracy. is still noticeable given that HOP is totally unsupervised. By incorporating the prior information into the model, we no- tice a significant boost in stance classification accuracy where pHOP achieves an accuracy of 71% and it outperforms sLDA by 2%. HASM only achieves an accuracy of 60% since it only utilizes the sentiment lexicon and cannot benefit from the proxy stance labels due to its unbounded aspect node arrangement for each sentence, which has to follow an rCRP. 4.5. Qualitative Results To evaluate the quality of opinion hierarchies discovered, we compare HOP with CATHY and HLDA. Figure 6 illustrates the opinion hierarchies generated by these three models respec- tively on Dataset I. It can be observed that CATHY can dis- tinguish two opposing stances at level-2, but it generated less coherent topics at level-3. Oftentimes, CATHY tends to group phrases sharing common constituent words into the same topic, e.g., Topic 2 and 4 under the right branch of the level-2 topic. HLDA generates more sensible topics at level-3. However, it is not able to distinguish two stances at level-2 since the num- ber of topics is unconstrained. Also, top topic words listed at level-3 for HLDA are dominated by unigrams and are less inter- pretable. This is not surprising since HLDA operates under the bag-of-phrases assumption. Phrases such as ‘immigration pol- icy’ and ‘immigration control’ are treated as two distinct tokens although they share the same constituent word ‘immigration’. Hence, running HLDA on data with phrases identified suffers from more severe data sparsity as the vocabulary size is signif- icantly enlarged. On the contrary, we can observe that HOP generates more distinct opinions7, especially at level-2. It can also be observed that HOP has more phrases appeared in its level-3 topics com- pared to HLDA. Overall, HOP offers better interpretability of the topics discovered. Moreover, when examining the hierar- chical topic assignment to each individual tweet, we notice that 7We manually add the topic labels in bold face for easy understanding of the results. October 8, 2018 CATHY HLDA vote leave EU Brexit remain referendum UK stay country Britain support against remain stay life house place live long today safe night immigration government public services economy Brexit campaign human rights western political sovereignty Brexit economy trade_deals markets economic_recession jobs workers_rights pound bank_England wages_rise HOP immigration EU control Brexit leave migration control_borders migrants immigrants immigration_control Brexit Labour leave Cameron EU tory Osborne prime_minister government party EU Brexit leave NHS money pay cost tax public_services spend race muslim Trump human_rights Obama president touch open_borders Hillary white EU vote Europe England France status_quo member_states German European_Union western_political unelected democracy freedom_movement stay parliament Scottish_independence laws sovereignty control general_election EU Brexit vote_leave UK Britain trade economy leaving jobs Europe Brexit_campaign lies remain facts read Jo_Cox_murder debate Boris_Johnson truth Farage leave EU vote Brexit stay referendum UK country remain happen EU Brexit UK leave leaving trade Britain remain economy immigration stay wait long place life man forever love longer friends EU control Cameron deal immigration government tory reform veto PM Brexit Britain economy market industry warns business world markets businesses cost funding pay spend paid housing increase extra higher immigrants stay house lol safe wanna gonna forever happy live hate mayor win gold para DA middle mayoral ref bath grave Brexit EU leave vote UK stay Britain remain life referendum leave_EU vote_leave UK_EU EU_referendum Brexit_UK vote_leave_EU stay stay_EU stay_leave stay_life vote_stay Boris_Johnson love_stay Jo_Cox leave_EU vote_leave vote_leave_EU EU_referendum remain_leave vote_leave_remain vote_EU_referendum vote_Brexit UK_vote_leave UK_leave_EU UK_EU leave_EU UK_leave_EU Britain_EU EU_trade leave_EU_trade Britain_leave_EU David_Cameron_EU Cameron_EU rights_leave_EU prime_minister Michael_Gove leave_EU post_Brexit EU_Brexit EU_change George_Osborne project_fear Michael_gove_EU Osborne_Brexit stay_EU stay_life stay_leave vote_stay stay_lane stay_place leave_EU_stay stay_forever stay_long UK_stay love_stay check_latest_updates stay_EU stay_house stay_home stay_safe stay_room stay_leave world_stay hope_stay Boris_Johnson Jo_Cox London_mayor Sadiq_Khan Ruth_Davison leave_Boris_Johnson Eddie_Izzard Bob_Geldof stay_EU Jo_Cox_murder Nigel_Farage vote_leave stay_EU EU_eurovision leave_EU stay_moment Amber_Rudd leave_racist social_media_stay vote_stay Brexit_UK Brexit EU_Brexit leaving_EU single_market EU_NHS EU_immigration Brexit_economy EU_economy control_immigration Figure 6: Opinion hierarchies discovered by CATHY, HLDA and HOP on Dataset I. in many cases, tweets are assigned with sensible hierarchical opinions. For example, “Briefing: Trade, investment and jobs will benefit if we Vote Leave” is assigned with [Support, Econ- omy]. It can be easily inferred from the topic result that the tweet is about the support of Brexit because of an economy- related reason. On Dataset II only the hierarchy generated by HOP is pre- sented in Figure 7 due to the space limit. Although the level-2 topics do not clearly represent two opposing stances since the HOP model is totally unsupervised, it can be inferred from the level-3 topics (‘Email scandal of Hillary’ and ‘Trump vowed to drain the swamp8’) that the left level-2 topic is about ‘Sup- porting Trump’. And similarly we can infer from the level-3 topics that the right level-2 topic is about ‘Supporting Hillary’. Also, under the ‘Drain the swamp’ topic, the 4th level gives more fine-grained topics on ‘Controlling illegal immigrants’ and ‘Trump will lead a unified Republican government’. 5. Conclusion In this paper, we have proposed an unsupervised Hierarchi- cal Opinion Phrase (HOP) model in which each document is 8Trump used this metaphor to describe his plan to fix problems in the federal government. Trump Hillary Donald_Trump president American vote Trump vote Hillary Hillary Trump Clinton Obama America email scandal drain swamp racist taxes Russia Hillary_Clinton FBI campaign emails Huma indicted spooky_story email FBI_investigation jail Trump America country great president repeal_obamacare change world drain_swamp won Trump taxes Obama tax release_tax_returns obamacare pay Clinton supreme_court law Trump Putin markets Russian economy crash Russia nuclear_war energy civil_war Trump win Hillary vote Clinton Trump women white racist black women_rights black_voters shit build_wall child_rape Donald_Trump reliable lead_unified_republican wind trusted Trump worst economy stock_market Mark_Cuban Trump lies Donald stop_lie cheat Trump deport hunger illegal_immigrants million_immigrants Figure 7: Opinion hierarchy discovered by HOP on Dataset II. associated with a path in a topic tree and with a document- specific distribution over different levels in the tree. Phrases are drawn from their associated topic-specific HPYP. Experi- mental results on two Twitter datasets show that our proposed HOP model is able to reveal hierarchical opinions from social media. It also shows that HOP significantly outperforms ex- isting approaches to hierarchical topic extraction in both topic October 8, 2018 coherence and stance classification. In our current work, all the paths in the generated hierarchical opinion tree have the same depth. In future work, we will explore modelling hierarchical opinion trees with varying depths of path. Acknowledgements We would like to thank the reviewers for their valuable com- ments and helpful suggestions. This work was funded by the National Natural Science Foundation of China (61528302, 61772132), the Natural Science Foundation of Jiangsu Province of China (BK20161430) and Innovate UK (grant no. 103652). References Blei, D. M., Griffiths, T. L., and Jordan, M. I. (2010). The nested chinese restaurant process and bayesian nonparametric inference of topic hierar- chies. Journal of the ACM (JACM), 57(2):7. Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent dirichlet allocation. Journal of machine Learning research, 3(Jan):993–1022. Chen, H., Sun, M., Tu, C., Lin, Y., and Liu, Z. (2016). Neural sentiment classification with user and product attention. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1650–1659. Dong, L., Wei, F., Tan, C., Tang, D., Zhou, M., and Xu, K. (2014). Adaptive re- cursive neural network for target-dependent twitter sentiment classification. In Proceedings of the 52nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 2: Short Papers), volume 2, pages 49–54. dos Santos, C. and Gatti, M. (2014). Deep convolutional neural networks for sentiment analysis of short texts. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 69–78. El-Kishky, A., Song, Y., Wang, C., Voss, C. R., and Han, J. (2014). Scalable topical phrase mining from text corpora. Proceedings of the VLDB Endow- ment, 8(3):305–316. Elfardy, H., Diab, M., and Callison-Burch, C. (2015). Ideological perspective detection using semantic features. Lexical and Computational Semantics (* SEM 2015), page 137. Fung, B. C., Wang, K., and Ester, M. (2003). Hierarchical document clustering using frequent itemsets. In Proceedings of the 2003 SIAM International Conference on Data Mining, pages 59–70. SIAM. Ghiassi, M., Skinner, J., and Zimbra, D. (2013). Twitter brand sentiment anal- ysis: A hybrid system using n-gram analysis and dynamic artificial neural network. Expert Systems with applications, 40(16):6266–6282. Griffiths, T. L. and Steyvers, M. (2004). Finding scientific topics. Proceedings of the National academy of Sciences, 101(suppl 1):5228–5235. Gui, L., Zhou, Y., Xu, R., He, Y., and Lu, Q. (2017). Learning representations from heterogeneous network for sentiment classification of product reviews. Knowledge-Based Systems, 124:34–45. Handler, A., Denny, M. J., Wallach, H., and OâĂŹConnor, B. (2016). Bag of what? simple noun phrase extraction for text analysis. NLP+ CSS 2016, 114. Hasan, K. S. and Ng, V. (2013). Frame semantics for stance classification. In CoNLL, pages 124–132. He, Y. (2016). Extracting topical phrases from clinical documents. In AAAI, pages 2957–2963. Kawamae, N. (2012). Hierarchical approach to sentiment analysis. In Semantic Computing (ICSC), 2012 IEEE Sixth International Conference on, pages 138–145. IEEE. Kim, J. H., Kim, D., Kim, S., and Oh, A. (2012). Modeling topic hierarchies with the recursive chinese restaurant process. In Proceedings of the 21st ACM international conference on Information and knowledge management, pages 783–792. ACM. Kim, S., Zhang, J., Chen, Z., Oh, A. H., and Liu, S. (2013). A hierarchical aspect-sentiment model for online reviews. In AAAI. Lai, S., Xu, L., Liu, K., and Zhao, J. (2015). Recurrent convolutional neural networks for text classification. In AAAI, volume 333, pages 2267–2273. Lim, K. W. and Buntine, W. (2014). Twitter opinion topic model: Extracting product opinions from tweets by leveraging hashtags and sentiment lexicon. In Proceedings of the 23rd ACM International Conference on Conference on Information and Knowledge Management, pages 1319–1328. ACM. Lin, C. and He, Y. (2009). Joint sentiment/topic model for sentiment analysis. In Proceedings of the 18th ACM conference on Information and knowledge management, pages 375–384. ACM. Lin, C., He, Y., and Everson, R. (2010). A comparative study of bayesian mod- els for unsupervised sentiment detection. In Proceedings of the fourteenth conference on computational natural language learning, pages 144–152. Association for Computational Linguistics. Lindsey, R. V., Headden III, W. P., and Stipicevic, M. J. (2012). A phrase-discovering topic model using hierarchical pitman-yor processes. In Proceedings of the Joint Conference on Empirical Methods in Natu- ral Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 214–222. Mcauliffe, J. D. and Blei, D. M. (2008). Supervised topic models. In Advances in neural information processing systems, pages 121–128. Mikolov, T., Sutskever, I., Chen, K., Corrado, G., and Dean, J. (2013). Dis- tributed representations of words and phrases and their compositionality. Advances in Neural Information Processing Systems, 26:3111–3119. Mohammad, S., Kiritchenko, S., and Zhu, X. (2013). Nrc-canada: Building the state-of-the-art in sentiment analysis of tweets. In Proceedings of the seventh international workshop on Semantic Evaluation Exercises (SemEval-2013), Atlanta, Georgia, USA. Newman, D., Lau, J. H., Grieser, K., and Baldwin, T. (2010). Automatic eval- uation of topic coherence. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 100–108. Association for Computational Linguistics. Nguyen, D., Vo, K., Pham, D., Nguyen, M., and Quan, T. (2017). A deep architecture for sentiment analysis of news articles. In International Con- ference on Computer Science, Applied Mathematics and Applications, pages 129–140. Springer. Paisley, J., Wang, C., Blei, D. M., and Jordan, M. I. (2015). Nested hierarchical dirichlet processes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(2):256–270. Pitman, J. et al. (2002). Combinatorial stochastic processes. Ren, Y., Zhang, Y., Zhang, M., and Ji, D. (2016). Context-sensitive twitter sentiment classification using neural network. In AAAI, pages 215–221. Röder, M., Both, A., and Hinneburg, A. (2015). Exploring the space of topic coherence measures. In Proceedings of the eighth ACM international con- ference on Web search and data mining, pages 399–408. ACM. Severyn, A. and Moschitti, A. (2015a). Twitter sentiment analysis with deep convolutional neural networks. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Re- trieval, pages 959–962. ACM. Severyn, A. and Moschitti, A. (2015b). Unitn: Training deep convolutional neural network for twitter sentiment classification. In Proceedings of the 9th international workshop on semantic evaluation (SemEval 2015), pages 464–469. Tang, D., Qin, B., and Liu, T. (2015). Document modeling with gated recurrent neural network for sentiment classification. In Proceedings of the 2015 con- ference on empirical methods in natural language processing, pages 1422– 1432. Teh, Y. W. (2006). A hierarchical bayesian language model based on pitman- yor processes. In Proceedings of the 21st International Conference on Com- putational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pages 985–992. Association for Computational Linguistics. Thonet, T., Cabanac, G., Boughanem, M., and Pinel-Sauvagnat, K. (2016). Vo- dum: a topic model unifying viewpoint, topic and opinion discovery. In European Conference on Information Retrieval, pages 533–545. Springer. Trabelsi, A. and Zaıane, O. R. (2014). Finding arguing expressions of diver- gent viewpoints in online debates. In Proceedings of the 5th Workshop on Language Analysis for Social Media (LASM)@ EACL, pages 35–43. Vilares, D. and He, Y. (2017). Detecting perspectives in political debates. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1573–1582. Wallach, H. M. (2006). Topic modeling: beyond bag-of-words. In Proceedings of the 23rd international conference on Machine learning, pages 977–984. October 8, 2018 ACM. Wang, C., Danilevsky, M., Desai, N., Zhang, Y., Nguyen, P., Taula, T., and Han, J. (2013). A phrase mining framework for recursive construction of a topical hierarchy. In Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 437–445. ACM. Wang, X., McCallum, A., and Wei, X. (2007). Topical n-grams: Phrase and topic discovery, with an application to information retrieval. In Data Mining, 2007. ICDM 2007. Seventh IEEE International Conference on, pages 697– 702. IEEE. Wang, Y., Huang, M., Zhao, L., et al. (2016). Attention-based lstm for aspect- level sentiment classification. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 606–615. Yang, Z., Yang, D., Dyer, C., He, X., Smola, A., and Hovy, E. (2016). Hier- archical attention networks for document classification. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1480– 1489. October 8, 2018 Introduction Related Work Opinion Mining based on Topic models Hierarchical Topic Extraction Topical Phrase Models Deep Learning for Sentiment Analysis Summary Hierarchical Opinion Phrase (HOP) Model Generative Process Inference and Parameter Estimation Level Allocation Sampling Path Sampling Complete Sampling Procedure Experiments Experimental Setup Methods for Comparison Topic Coherence Stance Classification Qualitative Results Conclusion 
work_htmuy3yn5jaqhbh64txayj7xj4	----	PDF hosted at the Radboud Repository of the Radboud University Nijmegen The following full text is a publisher's version. For additional information about this publication click this link. http://hdl.handle.net/2066/112344 Please be advised that this information was generated on 2021-04-06 and may be subject to change. http://hdl.handle.net/2066/112344 trends in analytical chemistry, vol. 9, no. Z,l990 58 References C. L. Forgy, Artificial Intelligence, 19 (1982) 17-37. F. Barachini, 8th International Workshop on Expert Systems & their Applications, Avignon General Conference Vol. 2, EC2, Gerfau, Paris, 1988, pp. 217-234. A. Gupta, 5th International Workshop on Expert Systems and their Applications, Avignon, EC2, Gerfau, Paris, 1985, pp. 26-57. M. R. Genesereth, Proceedings on the 3rd AAAZ Conference, Kaufman, Los Altos, CA, 1983, pp. 119-124. S. Russell, The Complete Guide to MRS, Report No. STAN- CS-851080, Stanford University, Stanford, June, 1985, 121 PP. 6 B. Silver, Studies in Computer Science and Artificial Intelli- gence, Vol. 1, North-Holland, Amsterdam, 1986. 7 J. Treur, Znformatie, 30 (12) (1988) 909-972. 8 R. Davis and B. G. Buchanan, in B. G. Buchanan and E. H. Shortliffe (Editors), Rule-Based Expert Systems, Addison Wesley, Massachussetts, 1984, pp. 507-530. 9 M. L. Wright, M. W. Green, G. Fiegl and P. F. Cross, An Ex- pert System for Real-Time Control, IEEE Software, IEEE Computer Society, March, 1986, pp. 16-24. M. De Winter, D. Grietens and Dr. M. Rijckaert are at the Chem- ical Engineering Department, KU Leuven, De Croylaan 46,303O Heverlee, Belgium. An expert system for the validation of high- performance liquid chromatographic methods L. M. C. Buydens and J. A. van Leeuwen Nijmegen, The Netherlands M. Mulholland Cambridge, U.K. B. G. M. Vandeginste and G. Kateman Nijmegen, The Netherlands A method validation expert system, developed as part of Es- prit project ESCA, is described. The proper validation of high-peeormance liquid chromatography methods is an is- sue of rapidly growing importance. This is due to the in- creasing demands of good laboratory practice. Since vali- dation involves a lot of statistical calculation and interpreta- tion with which most analysts are not familiar, it is often ne- glected in the method development process. An expert system that provides this knowledge and experience is therefore very useful. Introduction Method validation is gaining rapidly in importance due to the increasing demands of good laboratory practice. This is especially true in the area of phar- maceutical analysis, where regulatory bodies pose increasing demands on the analytical methods. The Esprit project ESCA aims to build demonstrator ex- pert systems for method development in high-per- formance liquid chromatography (HPLC) analysis of pharmaceutical compounds. The general ap- proach of the project is described in the paper on ESCA, elsewhere in this issue’. Such a system should include all necessary steps from the selection of initial HPLC conditions up to the validation of the method. This resulted in four expert systems, each having a specific task covering the whole development pro- cess. The expert system for the selection of initial conditions provides chromatographic conditions that yield acceptable retention times for all peaks in the chromatogram. When a chromatogram is obtained in this way it may be necessary to optimise the selec- tivity to obtain an optimal distribution of the peaks over the chromatogram. This is done by the expert system for the selectivity optimization. The expert system for the optimization of chromatographic and instrumental parameters aims to obtain an accepta- ble resolution and signal-to-noise ratio in the short- est possible time. The task of the method validation expert system is to validate the HPLC method that has been selected and optimised in the other expert systems of ESCA. All these domains have been de- scribed elsewhere’>*. This paper focuses on the ex- pert system for method validation. In the subsequent sections the scope and aproach in this expert system is described. Scope of the method validation expert system Method validation is a very broad concept. The full validation of a method comprises the testing of 0165-9936/90/$03.00. 0 Elsevier Science Publishers B .V. trends’ in analytical chemistry, vol. 9, no. 2,199O 59 different aspects of its performance such as accura- cy, precision, sensitivity, selectivity and limitations. The expert system described here concentrates on the precision testing of an HPLC method. Precision testing is the determination of the random error. Many methods will be applied under slightly differ- ent conditions from those of their development. The method may be used on other instruments, by other analysts or even in other laboratories. Therefore an estimation of the precision is a very important aspect in the method validation. According to Youden and Steiner3 one can distinguish a repeatability test and a reproducibility test. Repeatability testing involves testing the performance of a method when repeated under the same conditions, on the same instrument and by the same analyst. The reproducibility of a method is its precision under changing conditions and environments, e.g. on other instruments, by other analysts or in other laboratories. A reproduci- bility test hence involves an interlaboratory study which implies much effort and expense. The interla- boratory test of a method is doomed to failure when the method is not rugged for some slightly changing environmental and operational conditions that can be expected when the method is transferred to an- other site. Before organizing a large interlaboratory reproducibility test, the method can be intralabora- tory tested for its ruggedness with respect to ex- pected changes. The evaluation of the precision often involves the use of more or less complex statistics. Most analysts therefore consider it a tedious task to test the preci- sion of a developed method. An expert system that proposes the experiments that should be carried out, performs the appropriate statistical tests and inter- prets the results, and subsequently providing advice on the respecification of the method if the test fails, would be very useful. The expert system that has been developed in the ESCA project provides such advice on the repeatability and the ruggedness test- ing of HPLC methods435. In this article we focus on the ruggedness testing part of the expert system’. Ruggedness system The use of statistical experimental design has been proposed by Youden and Steiner3 to evaluate the ruggedness of a method for certain factors. More re- cently these designs have been applied in ractice for the ruggedness testing of HPLC methods ?9 - . The ruggedness expert system described here con- tains the necessary knowledge and experience for advising the user on the use and interpretation of ap- propriate statistics. The expert system basically con- sists of a separate module for each of the following four steps: l method description; l the selection of important factors; l the selection of the experimental design; l processing and interpreting the results. The method description module The user is asked to fully specify the method that will be tested. This includes information about sever- al aspects of the analytical procedure, all of which may influence the degree and extent of the testing procedure. This module contains knowledge on the normal range of the specifications. If the system ac- cepts the user’s method description, the user can safely rely on the result of a further consultation. Only when the method description is very incom- plete is some caution required. The conclusions may not be fully significant when too much information is missing. In all other cases the user can be confident about the results when the method description is ac- cepted by the system. The requested information is: l information about the analytical procedure of the HPLC method, such as sample preparation, col- umn, flow-rate, etc. l chromatographic results, such as retention times and minimal resolution between two peaks. l intended application area of the method. Differ- ent levels of validation are required, depending on whether the method is only used a few times or is applied for many samples in different laborato- ries. Also, before a method is submitted to regula- tory bodies, a thorough precision test should be carried out. l finally, knowledge about the availability of instru- ments is needed to build up the full validation pro- cedure. The method description module guides the user in providing all necessary information, without asking for superfluous or irrelevant data. It contains com- mon sense knowledge about chromatographic prac- tice and asks only information that applies to the spe- cific case. The consultation of this module is the basis for the further steps in the ruggedness system. The factor selection module In this step all factors that are likely to vary in the daily use of the method and that are suspected to in- fluence its performance are identified. Since HPLC is a rather complicated analytical technique, many parameters must be considered. The experienced analyst has however more or less strong ideas about the most important factors that are relevant to the behaviour of the system. For example, a method should be rugged for the temperature when it is clear 60 trends in analytical chemistry, vol. 9, no. 2; 1990 from the method description that no temperature control is provided. When the method is to be used over a long period of time and/or in other laborato- ries, then it is clear that the temperature of the envi- ronment will vary. When the method is not rugged for these variations this can cause alterations in its behaviour and wrong results may be obtained. The factor selection module covers the knowledge about the significant selection of these factors. This in- volves experience about the suspected variations of the selected factors in different environments”. Another feature of this module is that the user is allowed to add or delete factors that are chosen by the expert system. The levels of the selected factors can also be altered. This feature is provided to allow the user to introduce his/her knowledge about a spe- cific practical situation, For good laboratory practice these alterations are registered so that they can al- ways be traced afterwards. When the user agrees with the proposed set of factors, all information is stored and the system can continue. Module to select an experimental design The number of factors that may influence the method’s performance, when applied in practice, is very large. Even when only the most important fac- tors are selected a large number of experiments is still required. Experimental designs that allow as much information as possible to be extracted from a minimum number of experiments are necessary. When it can be assumed that there are no higher or- der interactions between the different factors, which is often a realistic assumption, then the fractional factorial designs are the most efficient for this pur- pose3. The designs that are implemented in the ex- pert system are: l full factorial designs; l half fractional factorial designs; l saturated factorial designs; and l reflected saturated factorial designs. The actual choice of a particular design depends on the number of factors and levels that are selected by the factor selection module. When the effects of only two or three factors are to be established at two levels then the full fractional designs are selected, otherwise the saturated fractional designs are pre- ferred. When the expert system has made a decision it presents the selected design together with the batch of experiments that are to be carried out. At this stage the user leaves the expert system to per- form the required experiments. Module for processing and interpretation of results When the experimental design has been selected an appropriate spreadsheet is created where the re- sults of the experiments can be entered. When all data have been collected the expert system can start processing the results. First the relevant chroma- tographic results from each experiment are calcu- lated. These include: concentrations, peak efficien- cy, resolutions etc. When these parameters are ob- tained the calculations can start to find the effects of the different factors”. The method is supposed to be repeatable before the ruggedness test is started. To check if this is valid over the whole experiment all standard errors are examined. Since the analyte con- centrations are the most important chromatographic results, the main effects of all factors on the calcu- lated concentrations are calculated first and com- pared with a prespecified tolerance level. If any fac- tor shows an unacceptably large effect on the con- centrations, the method fails on the ruggedness test. The main effects on all the other parameters, such as peak height and resolution, are then investigated in the same way. When the tolerance level for these pa- rameters is exceeded, the ruggedness test does not fail, but warnings are flagged to the user. If the rug- gedness test of a specified method fails, the user is advised to respecify the method and to repeat the ruggedness test. For obtaining meaningful advice on this issue, this expert system must be linked to the other expert systems that were developed within the ESCA project. The most straightforward is the inte- gration with the expert system for the optimization of instrumental and chromatographic parameters. This expert system can give advice on how to change the method in order to obtain improvements, e.g. in resolution. Integration of these two expert systems is proceeding. When the method passes the ruggedness test a ruggedness report is printed. This includes a set of system suitability criteria. These are the maximum and minimum values that are found for chroma- tographic parameters such as resolution and peak height. Whenever the method is applied, the analyst can then use these values to evaluate the chroma- tographic system regarding its suitability for the method. Implementation The method validation system is implemented in Goldworks, one of the selected expert system shells in the ESCA project. Since method validation is a very complex and broad area, of which only a part is implemented in the present system, it was necessary to build the expert system in a way that is easily ex- tendable. We chose a modular approach. The two main modules that constitute precision testing are the repeatability and the ruggedness system. These modules can be consulted separately. Each of these trend in analytical chemistry, vol. 9, no. 2,199O 61 / I h sd w di0mos3 rqeet repeat A A Repeatability Supervisor A V method factcf select diagnose description choice design Neeechess A A A A A Rhggedness I ?/ \1/ v V VV \/ Common Datastructure I I 1 Y explain ( Fig. 1. Implementation architecture of the precision testing expert system. * Screens > chromatograph questions > options I CRROMATOGRAPH Main questions and answers I Related answers Flow rate (ml/min) : 1.5 Number of solvents :2 PR : 2.5 Buffer cone (M) : Additives : Injection volume (ul) : 10 Temperature mode : CONTROLLED J I I Minimum solvent (%) : 25 Solvent 1 (%) : 75 Solvent 2 (%) : 25 Solvent 3 (%) : Solvent 4 (%) : Minimum additive (%) : Additive1 :o %( Additive2 :o %( Additive3 :o %( Additive4 :o %( i Additive5 :o %( 1 Temperature (deg. C) : 40 Fig. 2. Example of user interaction with the ruggedness expert system. 62 main modules are in turn built from different submo- dules, and each submodule can also be consulted in- dividually. A structure, called the supervisor, guides the user through the appropriate modules. It con- tains meta level or strategic knowledge on the next module to be consulted and therefore needs an over- view of the results and data that are used in all sepa- rate modules. All data and results of the modules are stored in common database, accessible by all mod- ules and the supervisor. The overall architecture is pictured in Fig. 1. This architecture allows the integration of mod- ules which contain very different types of knowl- edge. This is very useful since, as can already be seen from the ruggedness system modules, very different knowledge sources are necessary. Experiential as well as algorithmic knowledge sources must be com- bined in an integrated system to make it useful. This combination of modules containing different types of knowledge is a typical feature of second generation expert systems12. In the ESCA project the stand-alone expert sys- tems described elsewhere in this issue’ will be inte- grated. The architecture of the method validation expert system is flexible enough to allow an easy in- tegration with the other expert systems. Though user-friendliness was not the main con- cern in the project, attention was paid to developing an efficient and rather robust user interface. The questions are presented to the user in a window sys- tem. Explanations and additional help are provided where necessary. These can always be accessed by the user through special pop up windows. An exam- ple of the user interface is presented in Fig. 2. Conclusions The expert system for method precision testing, developed within the ESCA project, is a succesful stand-alone expert system. Validation of the system by means of 11 real test cases resulted in about 85% of success. Even for the cases where the conclusions were different from the real expert, the expert sys- tem choice was acceptable to the expert and in some cases even better than the real expert. A full descrip- tion of the validation and further evaluation of the system is published elsewhere13. The implementa- tion of the system allows future additions and inte- grations to be carried out with flexibility. Acknowledgement Part of this research is supported by the EEC as Esprit project P1570 Expert systems in chemical analysis (ESCA). trends in analytical chemistry, vol. 9, no. 2,’ 1990 References 1 J. A. van Leeuwen, L. M. C. Buydens, B. G. M. Vande- ginste and G. Kateman, Trends Anal. Chem., 9 (1990) 49. 2 D. Goulder. T. Blaffert, A. Blokland. L. M. C. Buvdens. A. 9 10 11 12 13 Chhabra, A. Cleland, N. Dunand, H. Hindriks, ‘G. Kate- man, J. A. van Leeuwen, D. L. Massart, M. Mulholland, G. Musch, P. Naish, A. Peeters, G. Postma, P. J. Schoenma- kers, M. DeSmet and G. B. M. Vandeginste, Chromatogra- phia, 26 (1988) 237-243. W. J. Youden and E. H. Steiner, Statistical manual of the AOAC, AOAC, Washington, DC, 1975. M. Mulholland, J. A. van Leeuwen and B. G. M. Vande- ginste, Anal. Chim. Acta, 223 (1989) 183-192. M. Mulholland, N. Dunand, A. Cleland, J. A. van Leeuwen and B. G. M. Vandeginste, J. Chromatogr., in press. M. Mulholland and H. Waterhouse, .I. Chromatogr., 395 (1987) 539-551. M. Mulholland and J. Waterhouse, Chromatographia, 25 (9) (1988) 769-774. M. Mulholland, P. J. Naish, D. R. Stout and J. Waterhouse, Chemometrics and Intelligent Laboratory Systems, 5 (1989) 262-270. A. Mulholland, Trends Anal. Chem., 7 (1988) 383. J. A. van Leeuwen, B. G. M. Vandeginste, G. Kateman, M. Mulholland and A. Cleland, submitted for publication. G. Box, W. Hunter and J. Hunter, Statistics for Experiments, an Introduction to Design, Data Analysis and Model Build- ing, Wiley, New York, 1978, pp. 291-453. B. Maitre, T. Laasri, F. Mondot, F. Charpillet and J. P. Ha- ton, Proceedings of the 9th International Workshop on Expert Systems and their Applications, Specialised Conference on Second Generation Expert Systems, Avignon, May 29-June 2,1989, EC2, Nauterre, 1989, pp. 237-251. M. Mulholland, J. A. van Leeuwen and L. M. C. Buydens, J. Chromatogr., in press. Drs. L. M. C. Buydens, .I. A. van Leeuwen and G. Kateman are at the Department of Analytical Chemistry, Catholic University of Nijmegen, Nijmegen, The Netherlands. Dr. M. Mulholland is at Philips Scientific, Cambridge, K. U. Dr. B. G. M. Vandeginste is at Unilever Research, Vlaardingen, The Netherlands. Computer Corner Contributions Contributions of between 400 and 900 words are wel- come in the following categories: hardware, software, chemical applications, mathematical tools and interfac- ing. Please send your papers either to: TrAC Computer Comer, D.L. Massart, Vrije Universiteit Brussel, Fakulteit der Geneeskunde en der Farmacie, Farmaceutische Scheikunde, Laarbeeklaan 103, B-1090 Brussels, Belgium. or TrAC Computer Comer, A.P. Wade, Department of Chemistry, University of British Columbia, 2036 Main Mall, Vancouver, B.C. Canada V6T lY6. 
work_hwqiyaijfbe53iyl4yuuyi3iwa	----	Role of Expert Systems in Identification and Overcoming of Dengue Fever (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 8, No. 10, 2017 82 | P a g e www.ijacsa.thesai.org Role of Expert Systems in Identification and Overcoming of Dengue Fever Dr. Nadeem Ahmed Department of Computer Science Superior University Lahore, Pakistan Muhammad Shoaib Department of Computer Science Sharif College of Engineering and Technology Lahore, Pakistan Adeed Ishaq Department of Computer Science & IT The University of Lahore Lahore, Pakistan Abdul Wahab Department of Computer Science Superior University Lahore, Pakistan Abstract—This paper presents a systematic literature review on expert systems which are used for identification and overcoming of Dengue fever. Dengue is a viral disease produced by Flavivirus. The expansion of Dengue fever is because of uncontrolled population and urbanization without suitable water administration. With the quick technological enhancement, we can identify and overcome Dengue fever by using expert systems. These expert systems require knowledge of the relevant problem and techniques to infer the result in order to make decisions. The literature review provides a comparison of techniques, methodologies, limitations and advantages of different Dengue expert systems. These expert systems facilitate both doctors and patients in Dengue detection. Multiple risk factors can be eliminated by the detection of Dengue fever through expert systems at early stages of Dengue. Furthermore, we find that enhancement of knowledge base improves accuracy of expert systems. Keywords—Expert system; rule based; Dengue; health care; disease; fever I. INTRODUCTION Dengue fever is one of the viral diseases around the world that is caused by blood feeding mosquito. Dengue fever happens due to a single stranded RNA Flavivirus, which is transmitted by mosquito bite [1], [2]. It is one of the most emerging diseases of the world [3]. There are about 20-80 million cases annually and majority of these are diagnosed as flu that causes about 24,000 deaths every year. Due to lack of proper treatment around 2.5% of the infection results to be fatal. The symptoms of this fever include joint pain, headache, muscle-pain and skin rashes [4]. Dengue ranks as the most important mosquito-borne viral disease in the world [5]. Countries which are most effected by Dengue are more than 110 in number [6]. Mainly this is due to bad sanitary conditions in underdeveloped countries. The experts provided a number of solutions regarding Dengue diagnosis but due to inexperienced doctors it is still misclassified [7]. Thus, throughout the world the experts are facing the problem of diagnosis of Dengue due to its misclassifications. The diagnosis process is quite time consuming which make things more complex. Much time is required for the identification of its correct classification. There is no problem in diagnosis if the number of patients is less but if they are huge in numbers and the country has limited resources then it becomes problematic to correctly diagnose this fever. The Dengue fever issue motivates computer experts to devise an automated and time efficient solution that correctly classify the Dengue. These automated systems for Dengue diagnosis can be deployed in faraway places as well as in epidemic situations to save precious lives. In computer science those systems that can solve a specific problem and can reason rationally are known as Expert Systems. Expert systems require knowledge of the regarding problem and techniques to infer the result in order to make decision [39]. Today there exists a number of expert systems for disease diagnosis as for bacterial infection MYCIN, for lung disease PUFF, for angiograms ANGY, for edema PIP, for meningitis diagnosis NEOMYCIN, for chest pain MED1, for bacterial infection GUIDON and for hematological disease HEME [8]- [10]. There are several types of expert systems based on knowledge base, i.e. rules, cases, hybrid approach (rules and cases) and hybrid approach with ANN like Medical Expert Solution (MES) system [25]. The MES uses a knowledge base which consists of two knowledge structures, named as symptom and diseases. The MES provides medical facilities to those users who don’t have access for it. Further, we will discourse research methodology, different mosquito borne diseases, related work, discussion, conclusion and future work. II. RESEARCH METHODOLOGY A. Rsearch Question RQ1: Does expert systems correctly identify Dengue fever? RQ2: Does knowledge base of expert system fulfill the requirements? RQ3: Does interfaces of expert systems are user friendly for all type of users? (IJACSA) - International Journal of Advanced Computer Science and Applications Vol. 8, No. 10, 2017 83 | P a g e www.ijacsa.thesai.org B. Search Strategy After observing and viewing all aspects of expert systems a research has been carried out on expert systems for Dengue. The field of computer science plays a major role in this perspective because it provides us the tools and techniques for gathering data from massive data stores/sources. 1) Identification of Search Terms: From research question extract major terms and check their synonyms for search related research papers. These terms verified in relevant papers. OR operator use for concatenation of synonyms words. AND operator use for concatenation of major terms. 2) Trial search: Trial search conducted by using the following search string in Science Direct, IEEEXplore, SpringerLink and ACM digital library. Table 1 contains complete information of digital libraries with search items and search conduction date. (“Dengue” OR “Fever” OR “Malaria" OR “Rule based” OR “Artificial Neural Network”) AND (“Identification” OR “Overcoming” OR “Expert system” OR “Task” OR “Factors” OR “Challenges” AND OR risks OR problems OR “issues”) TABLE I. TRAIL SEARCH RESULTS S. No. Digital library Search Conduction Date 1 IEEE Explore 144 01 January 2017 2 ACM Digital Library 118 10 January 2017 3 Google Scholar 101 15 January 2017 4 Science Direct 12 20 January 2017 5 Site Seer Digital Library 11 25 January 2017 6 Springer link 8 29 January 2017 C. Study Selection Criteria Some of the papers are relevant to our research and properly defining our research scope. These papers have been included here that helped in the research and those papers which are not relevant to our research, not properly defining our research scope and their language is different, are excluded. 1) Inclusive Criteria: After viewing different papers some papers are, giving the answer of our research questions, defining our search terms, their language is English and within our scope, these papers are included. 2) Exclusion Criteria: Some papers are not within a scope or not written in the English language which is prescribed in this paper, or not giving proper answers of our research questions, such type of papers is excluded. D. Data Extraction and Assessment of Study Quality As our inclusive criteria is defined, we have gathered the data which is relevant to our research, fulfilling our requirements and defining the scope of our research, also it will guide us for quality study. E. Electronic Data Stores We search related data from the following electronic data sources:  www.ieeexplore.ieee.org  www.acm.org  www.sciencedirect.com  www.citeseer.ist.psu.edu  www.scholar.google.com  www.springerlink.com Fig. 1. Research methodology. Fig. 1 shows complete research process and number of research paper included in our study. Duplicate publication and non-primary research studies exclude. III. DIFFERENT MOSQUITO BORNE DISEASES Here we describe the origin of different mosquito born disease. These diseases have some common symptoms with others. Some are viral while other are non-viral. Viral diseases are either due to mosquitos like Yellow fever, Malaria, Dengue and influenza [11], whereas non-viral are natural diseases such as heart diseases, diabetes and cancer. Unlike non-viral diseases, the viral diseases can spread in short spans of time and are difficult to handle. Fig. 2 expresses hierarchical structure of viral and non-viral diseases. (IJACSA) - International Journal of Advanced Computer Science and Applications Vol. 8, No. 10, 2017 84 | P a g e www.ijacsa.thesai.org Fig. 2. Viral and non-viral diseases. A. Diseases Due to Mosquitos These diseases are caused by viruses that are transmitted by mosquitos. Which includes Dengue, Yellow Fever and Malaria [12]. In past years’ research shows that there is leap in diseases caused by mosquitos. This motivates researchers and scientists to work to reduce these diseases [13]. Fig. 3 presents complete detail of mosquitos borne diseases. 1) Yellow Fever: It is one of the infectious diseases which is caused by Aedes mosquito that is infected by Flavivirus. This mosquito is usually found in sub-tropical or tropical regions [14]. Signs and Symptoms: Yellow Fever happens due to a virus cause by mosquito. The symptoms appear in 3 to 6 days after mosquito bite. There are three stages of this disease.  Stage 1 (infection): The infected person shows following symptoms in first 3 to 4 days, fever, loss of appetite, headache, joint pain, vomiting and jaundice. After day 4 these symptoms become brief.  Stage 2 (remission): In this stage, most of the symptoms are gone and people can be recovered but due to carelessness the situation can get even worst.  Stage 3 (intoxication): In organs such as liver, heart and kidney different problems may occur [15], [16]. 2) Malaria: Malaria is also a disease that happens due to virus by a female mosquito Plasmodium of genus Anopheles carry. If this disease is not treated in time it may lead to complications towards death. There are two types of malaria in general: simple and severe. Simple malaria is curable if treated within time while severe malaria may lead to death. Fig. 3. Mosquitos borne diseases. Signs and Symptoms:  Simple Malaria: This disease remains for 6 to 10 hours. Its symptoms are, cold stage that has sense of shivering, hot stage with headache and vomiting and sweating stage. Attacking schedule is with Tertain parasite, it occurs every second day & with Quartan parasite it occurs every third day. General symptoms are: Fever, Headache, Sweating, Vomiting, Body ache and Chills. The physical impact of malaria can be seen on the infected body that includes, getting sweaty, temperature rise up, mild jaundice, respiration increased and weakness in body.  Severe Malaria: During simple malaria if there arise a complication of infection in any part of body by failure of any organ it may create severe malaria. 3) Dengue: Aedes aegypti is certain type of mosquitos that causes of Dengue. It is a viral disease and the virus is single-positive strand RNA virus and commonly found in sub- tropical and tropical regions [17]. Signs and Symtoms: Major symptoms of Dengue are high-fever and combination of at least two which are, Pain in joints, Bleeding gums, Rashes on skin, Very low white-blood cells, Pain in bones and severe pain behind eyes. To get better a proper lookout is very necessary because after 3 to 7 days following things can happen, Red spots on body, Pale skin, Drowsiness, Pain in abdomen, Breathing in difficult. Dengue fever that stays for 2 to 7 days is normally called Dengue Hemorrhagic Fever (DHF). DHF warning appears when after 24 to 48 hours fever suddenly drops and at this time bleeding is started, due to this circulatory system is disturbed and it may lead to death [18]. World Health Organization (WHO) has worked on the Dengue risk factors that play the pivotal role in the disease diagnosis [19]. Similarly, resources have showed that Dengue is assorted disease, its symptoms are mixed with other diseases that make its detection and diagnosis more difficult [20]. IV. RELATED WORK A. Expert System using ANN In this paper, SOM (Self Organization Map) was used to distinguish between healthy and Dengue infected patients. The research consists of two maps, one that contain BIA (bio impedance analysis) factors and other contains disease symptoms. The research resulted these major BIA factors; Abdominal pain, Anorexia, bleeding and rashes that were present in Dengue patients. After BIA analysis, they found 100-unit best map size with error of 0.033 and quantization error of 1.313 [21]. Goal was to predict occurrence of Dengue by comparing weights. The learning rate varied from 0.5 to 0.9 [22]. Table 2 demonstrates advantages and disadvantages of expert systems by using artificial neural network. (IJACSA) - International Journal of Advanced Computer Science and Applications Vol. 8, No. 10, 2017 85 | P a g e www.ijacsa.thesai.org TABLE II. ADVANTAGES AND DISADVANTAGES OF EXPERT SYSTEM USING ANN Advantages Disadvantages Reprogramming is not needed Proper training is needed to operate Tasks are completed by adjusting weights Large data, high processing time Learning is by hit and trail Complexity is large if data is big Easy to use and learn by examples Lack in history maintenance Perform functions that a linear system can’t perform Highly dependent on data B. Emerge This is rule-based expert system which was developed to assist in emergency room. This system uses weighting factors determined by neural network. It is composed of input & output block and a hidden layer block that communicates input to output. Neural Network learns from example & predict an output based on this knowledge. It uses IF-THEN-UNLESS instead of IF-THEN that make more precise decision [29]. C. Expert System using Rule based Reasoning Here, Dengue was detected using Rule Based Classifier that are: Decision Tree, Naïve Bayes & Associative Classifiers. The research proved that result from more than one classifier is better than results by using one classifier. The result jumped up to 70% by using more than one classifier thus giving better result than single classifier [23]. In this paper, a Dengue diagnosis system was proposed that observe medical record procedures. It uses tree like model rather than DF and DHF. Out of the record of 41 infected 27 were classified DF, 9 were DHF-I and 5 were DHF-II [24]. Table 3 provides different advantages and disadvantages of expert systems by using rule based reason. TABLE III. ADVANTAGES AND DISADVANTAGES OF EXPERT SYSTEM USING RULE BASED REASON Advantages Disadvantages Representation of knowledge is natural Knowledge range is narrow Knowledge is kept separately Cost effective expert system Training is used to increase skills No creative solutions There is no overhead No proper rules Additional help is involved Can’t learn itself D. MYCIN MYCIN was one of the earliest diagnosis systems that used to diagnose infectious diseases caused by bacteria. First it describes the infection produced by bacteria then, it suggests antibiotic to give the infected person based on the decision [34]. E. PNEUMOCONIOSIS X-RAY Diagnosis Expert System This expert system was developed by Miriam Kubiska and Julie Herzner in 1992. The system has inference engine to examine the shadows on the x-ray. These shadows are used to determine the degree and type of PNEUMOCONIOSIS that is a lung disease. System includes the following:  The knowledge base that contain data of x-ray representation.  The interface that details the conclusions.  Knowledge acquisition mode that allow user to add or change information. F. XDIS XIDS had information of more than 250-300 internal diseases and syndromes that are frequently met in practice which help experts in diagnosis. The system has instructions for each disease, if disease is curable the system will prescribe medicine but if it is unable to identify the disease it will refer to a specialist. This process takes no more than 10 minutes [29]. G. Expert System using Integration of Rule base and ANN There is also a hybrid technique in which inference engine is updated and maintained whenever a new rule come. The purpose of new rule is to refining of time complexity and efficiency of symbolic rule. The authors explained about neural composition and how they can be shaped directly from results of experiments. Also, they explained the pattern based methodologies in comparison with Rule Based Cases [26]. The expert system was also available on mobile application. The data represented in form of decision tree & cases were stored in table in symbolic form. They maintained database for cases and rules which show that hybrid approach make system faster and always available due to web [27]. H. Role of Ultrasound in Dengue Detection In this paper, the authors perform a study using ultrasound rays to find that can be used to diagnose Dengue fever and also to be useful in predicting the severity of this disease. To test the study 128 people were taken as a subject that suffer Dengue disease. 40 people were identified to be positive case of Dengue, remaining underwent ultrasound again until those who were infected were differentiated from those who were not infected by Dengue virus. They found that after 5 to 7 days of ultrasound effect their gall bladder wall that got thick and had pleural effusion and ascites. They concluded that in epidemic of Dengue ultrasound can be used for diagnosis of Dengue fever [28]. Table 4 presents detail comparison of various expert systems and also express future guidelines for new researches. (IJACSA) - International Journal of Advanced Computer Science and Applications Vol. 8, No. 10, 2017 86 | P a g e www.ijacsa.thesai.org TABLE IV. COMPARISON OF DIFFERENT EXPERT SYSTEMS PROPOSED BY DIFFERENT RESEARCHERS Expert System Targeted Diseases Expert System Technique Expert System Methodology Future Guidelines Advantage Limitations Medical Expert Solution (MES) system [29] Tropical Diseases like Dengue Fever , Malaria 1. Inference engine 2. Knowledge Base MES consist of two main sections: 1. User-interface: This module used for Displaying information and perform user interaction with system 2. Expert system: It consists of inference engine and Knowledge Base. Knowledge Base consist of two elements, Diseases and their symptoms. MES use this data for diagnosis the diseases 1. Enhance Knowledge Base for getting accurate results. 1. Beneficial for those people who don’t have excess to medical facilities. 2. Provides first aid facility before going to medical consultant. 3. Reduce physicians work. 4. It can be used in epidemics and war environment. 1. Training required to use the system. Adaptive Neuro-Fuzzy Inference System (ANFIS) [30] Dengue , Dengue Fever , Malaria 1. Fuzzy inference system 2. Hybrid learning algorithm for ANIS 1. Optimization of ANIS structure 2. Optimization of hybrid learning training parameters 3. Evaluation of optimal models performance 1. Develop such kind of system which can be used by physicians in hospitals or clinics. 1. Diagnose risk in Dengue patients. 2. It has better performance as compare to noninvasive system which is developed for diagnosis risks in Dengue patients [32]. 3. Diagnostic accuracy achieved up to 86.13%. 4. More robust and reliable than Ibrahim et al. [32] 1. Not practically implemented yet. 2. No module for physicians. Adaptive- Network-Based Fuzzy Inference System (ANFIS)[31] Dengue , Dengue Fever 1. Fuzzy if-then rules or fuzzy conditional statements 2. fuzzy-rule- based systems 1. Architectures and learning algorithms 1.1 Architecture and Basic Learning Rule 1.2 Batch (Off-Line) Learning 1.3 Pattern (On-Line) Learning 2. Adaptive-network-based fuzzy Inference system 2.1 ANFIS architecture 2.2 Hybrid learning algorithm 2.3 fuzzy inference systems with simplified fuzzy if-then rules 1. The System proposed new fuzzy models for data classification and feature extraction. 2.Fixed structure of ANFIS. 1. Training issue regarding overlapping between adjacent membership functions and minimal Uncertainty. An Application of expert system (AExS) [33] Varicella (Chickenpox), Dengue Fever Hands-foot- and-mouth disease Epidemic parotitis (Mumps) 1.Forward chaining 2. Rule-base system 1. User-interface 2. Inference mechanism 3. Explanation facility 4. Knowledge acquisition facility 1. Construct AExs for large computer program. 2. Web portal design for easy excess of user anywhere. 1. It is important for diagnosis of viral infections. 2. Provide feedback in explanation component and inference engine. 3. Complete system with experimental results. 1. User can only select predefined symptoms. 2. Lack of self- inference ability. Web-based Patient Support System[35] Dengue, Diabetes 1. Artificial Intelligence in Medicine 2. Fuzzy logic 3. Neural Network 4.Centralized Databases and WWW 1. WEB-BASED MEDICAL DIAGNOSIS AND PREDICTION 1.1. User-interface 1.2. Predication modules 1.3. Diagnosis modules 1.4. Database 1. Developed distributed databases for improving quality of treatment and providing better diagnosis. 1. Artificial intelligence used to make consultancy more interactive. 2. Fuzzy logic used in uncertainty conditions. 3. ANN produce better results. 4. Provide Tele-health care. 1. Not deployed in working in environment. 2. Lack of experimental study. Rule based expert system[36] Dengue infections 1.Bioelectrical impedance analysis (BIA) 1. Development of a rule based expert system 1.1 Identifying the problem 1.2 Tree Diagram 1.3 Rule Based Expert System 1.4 Architecture of the Expert 1. Applying nonlinear modeling technique for improving classification 1. Perform Risk classification in DF and DHF. 1. Do not recognize complex patterns. (IJACSA) - International Journal of Advanced Computer Science and Applications Vol. 8, No. 10, 2017 87 | P a g e www.ijacsa.thesai.org System 1.5 Structure of the Patient Database accuracy like ANN. Predicting arboviral disease emergence using Bayesian networks [37] Dengue virus 1. Bayesian Belief Network (BBN) 1. Framework and process of BBN risk modelling and Mapping 2. Input data processing and classification 2.1 Climatic parameters 2.2 Road and railroad density 2.3 Seaports and airports 2.4 Human population density 2.5 Frequency of DENV introduction 2.6 Inclusion of endemicity risk node 3. Infection risk under current summer and winter Climates 4. Scenario modelling 5. Sensitivity analysis of risk node 1. Case study done on a large region and get better results for predicating Dengue virus risk. 2. By using this model other diseases like Chikungunya, Yellow fever, and Zika virus can be predicted. 1. Identify Dengue risk in different environments. 2. Dengue virus risk model developed which predict Dengue risk in particular condition. 1. Case study perform only for specific region. The Integrated Management of Health Care Strategies and Differential Diagnosis [38] Malaria, Typhoid fever, Cough 1. Combination of Integrated Management of Child illnesses (IMCI) and Health Information Systems (HIS). 1. Action and disease oriented differential Diagnosis of malaria and typhoid fever 2. Knowledge engineering 3. IMCI/HIS expert system interface architecture 4. System evaluation 1. Adaptive interfaces can be designed for urban and rural area users. 2. Provide support for distributed environments. 1. Used for diagnosis Malaria and Typhoid fever. 2. Used in rural area health care where maximum people are illiterate. 3. Provide user-interface for user interaction. 4. Reduce child mortality rate in rural areas. 1. It is stand- alone application which cannot be run in distributed environments. V. DISCUSSION AND CONCLUSION Dengue is a viral disease produce by Flavivirus. It has two forms: one is Dengue fever and second is Dengue hemorrhagic fever. DHF is more severe form of Dengue. It has become worldwide major problem in recent years. The expansion of Dengue fever is because of uncontrolled population and urbanization without suitable water administration. If infected female Aedes mosquitos bite human than Dengue virus transmitted in human blood. In computer science, systems that can solve a specific problem and can reason rationally are known as Expert Systems. These systems require knowledge of the relevant problem and techniques to infer the result in order to make decision. In recent years, different researchers have proposed different expert systems for identification, prescribe medicine and overcoming Dengue disease. These expert systems use different technique for getting better result like Bayesian Belief Network (BBN), Artificial Neural Networks (ANN), Fuzzy inference and Hybrid learning algorithm for ANIS. For implementing these techniques different researchers have proposed different methodologies for determining Dengue patient’s problems. The main common strategies in these methodologies are User-interface, Knowledge base engineering process and system evaluation process. User- interface works as bridge between system and user. During designing of user interfaces careful analysis of user tasks and context must be done [40]. We suggest that user interfaces should be well formed to convey knowledge according to user mental model so that a friendly communication may be done between user and system. Mostly knowledge Base consist of two elements, Diseases and their symptoms which can be used for determining disease and predict medical suggestions related to particular disease. We find during study that knowledge base has limited expressive power. Knowledge base should update itself according to environment as if it faces a new symptom which is not added in knowledge base than it should automatically be added in the knowledge base. Knowledge base should be supported for distributed environments. System evaluation used for checking the performance of expert system. A generic mechanism required for evaluation which can be followed by every researcher during designing of expert systems. Globally define certain set of rules to evaluate the expert system. VI. FUTURE WORK Furthermore, we can enhance Knowledge Base of expert systems for getting accurate results. Such kind of system should be developed which can be used by doctors and patients in web portal. Distributed databases should also be developed for improvement of quality of treatment and to provide better diagnosis. The expert system should be updated as it may be used for other diseases like Chikungunya, Yellow fever, and Zika. Adaptive interfaces should be designed for both literate and illiterate users which would help them to use these systems more easily and efficiently. Online assistance provided to semiliterate users during use of expert system can enhance their performance [42]. Ahmad, N. [41] introduced a new approach which use a virtual character for providing assistance to barely literate or deaf people. By using such kind of support, we can enhance expert system utility by offering additional services like automatic translation from text to sign language. (IJACSA) - International Journal of Advanced Computer Science and Applications Vol. 8, No. 10, 2017 88 | P a g e www.ijacsa.thesai.org Finally, autistic people are very sensitive and information technology is very useful for resolving their problem [43]. Therefore, expert system can be designed for autistic people because they also don’t know how to react in this situation. REFERENCES [1] Gubler, D. J. (1998). Dengue and Dengue hemorrhagic fever. Clinical microbiology reviews, 11(3), 480-496. [2] Gubler, D. J., Ooi, E. E., Vasudevan, S., & Farrar, J. (Eds.). (2014). Dengue and Dengue hemorrhagic fever. CABI.. [3] Fajardo-Dolci, G., Meljem-Moctezuma, J., Vicente-González, E., Venegas-Páez, F. V., Mazón-González, B., & Aguirre-Gas, H. G. (2011). [The Dengue fever in Mexico. Knowledge for improving the quality in health]. Revista Médica del Instituto Mexicano del Seguro Social, 50(6), 631-639. [4] Sulehri, M. A., Hussain, R., & Gill, N. I. (2012). Dengue fever its diagnosis, treatment, prevention and control. Gomal Journal of Medical Science, 6, 22-27. [5] World Health Organization. (2012). Global strategy for Dengue prevention and control 2012-2020. [6] Ranjit, S., & Kissoon, N. (2011). Dengue hemorrhagic fever and shock syndromes. Pediatric Critical Care Medicine, 12(1), 90-100.` [7] Salman, A., Lina, Y., & Simon, C. (2014). Computational Intelligence Method for Early Diagnosis Dengue Haemorrhagic Fever Using Fuzzy on Mobile Device. In EPJ Web of Conferences (Vol. 68, pp. 3-9). [8] Ashwinkumar, U. M., & Anandakumar, K. R. (2012, January). A Web- Based Patient Support System Using Artificial Intelligence to Improve Health Monitoring and Quality of Life. In 2012 Second International Conference on Advanced Computing & Communication Technologies (pp. 101-105). IEEE. [9] Kumar, K. A., Singh, Y., & Sanyal, S. (2009). Hybrid approach using case-based reasoning and rule-based reasoning for domain independent clinical decision support in ICU. Expert Systems with Applications, 36(1), 65-71. [10] Tomar, P. P., & Saxena, P. K. (2011). Architecture for medical diagnosis using rule-based technique. In First Int. Conf. on Interdisciplinary Research & Development, Thailand (Vol. 25, pp. 1- 25). [11] National Pesticide Information Center, Diseases Transmitted by Mosquitoes,NPIC. http://npic.orst.edu/pest/mosquito/diseases.html [12] Caraballo, H., & King, K. (2014). Emergency department management of mosquito-borne illness: malaria, Dengue, and West Nile virus. Emergency medicine practice, 16(5), 1. [13] World Health Organization. (2012). Global strategy for Dengue prevention and control 2012-2020. [14] Yellow Fever: Causes, Symptoms & Diagnosis - Healthline http://www.healthline.com/health/yellow-fever [15] Yellow fever: Nursing2016 - Wolters Kluwer Health vol. 35, no. 7, p. 75, 2005. [16] Ward, A. M., Gunaratne, J., & Garcia-Blanco, M. A. (2014). Identification of Dengue RNA binding proteins using RNA chromatography and quantitative mass spectrometry. Dengue: Methods and Protocols, 253-270. [17] Simmons, C. P., Farrar, J. J., van Vinh Chau, N., & Wills, B. (2012). Dengue. New England Journal of Medicine, 366(15), 1423-1432. [18] Hadinegoro, S. R. S. (2012). The revised WHO Dengue case classification: does the system need to be modified? Paediatrics and international child health, 32(sup1), 33-38. [19] Farrar, J. J., Hien, T. T., Horstick, O., Hung, N. T., Jaenisch, T., Junghanns, T., & Simmons, C. P. (2013). Dogma in classifying Dengue disease. The American journal of tropical medicine and hygiene, 89(2), 198-201. [20] Faisal, T., Ibrahim, F., & Taib, M. N. (2008, August). Analysis of significant factors for Dengue infection prognosis using the self- organizing map. In 2008 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (pp. 5140-5143). IEEE. [21] Husin, N. A., Mustapha, N., Sulaiman, M. N., & Yaakob, R. (2012, September). A hybrid model using genetic algorithm and neural network for predicting Dengue outbreak. In 2012 4th Conference on Data Mining and Optimization (DMO) (pp. 23-27). IEEE. [22] Bakar, A. A., Kefli, Z., Abdullah, S., & Sahani, M. (2011, July). Predictive models for Dengue outbreak using multiple rulebase classifiers. In Electrical Engineering and Informatics (ICEEI), 2011 International Conference on (pp. 1-6). IEEE. [23] Ibrahim, F., Taib, M. N., Sulaiman, S., & Abas, W. W. (2001). Dengue fever (DF) and Dengue haemorrhagic fever (DHF) symptoms analysis from an expert system perspective. In Multi Topic Conference, 2001. IEEE INMIC 2001. Technology for the 21st Century. Proceedings. IEEE International (pp. 212-215). IEEE. [24] Suwa, M., Scott, A. C., & Shortliffe, E. H. (1982). An approach to verifying completeness and consistency in a rule-based expert system. Ai Magazine, 3(4), 16. [25] Adekoya, A. F.@l,Akinwale,A. T.2 and Oke, O. E.3. A medical expert system for managing tropical diseases. Department of Computer Science, University of Agriculture, Abeokuta, Nigeria adekoyaaf@unaab.edu.ng1 www.journal.unaab.edu.ng/index.php/COLNAS/article/download/145/1 48 [26] Prentzas, J., & Hatzilygeroudis, I. (2005). Rule-based update methods for a hybrid rule base. Data & Knowledge Engineering, 55(2), 103-128. [27] Sulistiarini, E. B. (2013). Representing Knowledge Base into Database for WAP and Web-based Expert System. arXiv preprint arXiv:1312.3060. [28] Sai, P. V., Dev, B., & Krishnan, R. (2014). Role of ultrasound in Dengue fever. The British journal of radiology. [29] Adekoya, A. F., Akinwale, A. T., & Oke, O. E. (2008). A medical expert system for managing tropical diseases. College of Natural Sciences Proceedings, 74-86. [30] Faisal, T., Taib, M. N., & Ibrahim, F. (2012). Adaptive Neuro-Fuzzy Inference System for diagnosis risk in Dengue patients. Expert Systems with Applications, 39(4), 4483-4495. [31] Jang, J. S. (1993). ANFIS: adaptive-network-based fuzzy inference system. IEEE transactions on systems, man, and cybernetics, 23(3), 665- 685. [32] Faisal, T., Ibrahim, F., & Taib, M. N. (2010). A noninvasive intelligent approach for predicting the risk in Dengue patients. Expert Systems with Applications, 37(3), 2175-2181. [33] Chuang, L. Y. (2016). An Application of Expert System for Diagnosing Fever Caused by Viral Infection. Journal of Life Sciences and Technologies Vol, 4(1). [34] Shortliffe, E. (Ed.). (2012). Computer-based medical consultations: MYCIN (Vol. 2). Elsevier. [35] Ashwinkumar, U. M., & Anandakumar, K. R. (2012, January). A Web- Based Patient Support System Using Artificial Intelligence to Improve Health Monitoring and Quality of Life. In 2012 Second International Conference on Advanced Computing & Communication Technologies (pp. 101-105). IEEE. [36] Ibrahim, F., Mohamad, M. I., Makhtar, S. N., & Ibrahim, J. (2007). Classification of risk in Dengue fever and Dengue haemorrhagic fever using rule based expert system. In 3rd Kuala Lumpur International Conference on Biomedical Engineering 2006 (pp. 50-53). Springer Berlin Heidelberg. [37] Ho, S. H., Speldewinde, P., & Cook, A. (2017). Predicting arboviral disease emergence using Bayesian networks: a case study of Dengue virus in Western Australia. Epidemiology and Infection, 145(1), 54. [38] Adehor, A. B., & Burrell, P. R. (2008). The integrated management of health care strategies and differential diagnosis by expert system technology: a single-dimensional approach. World Academy of Science, Engineering and Technology, 44, 533-538. (IJACSA) - International Journal of Advanced Computer Science and Applications Vol. 8, No. 10, 2017 89 | P a g e www.ijacsa.thesai.org [39] McLeod, R., & Schell, G. P. (1998). Management information systems. Upper Saddle River, NJ: Prentice Hall [40] Iqbal, M. W., Ahmad, N., & Shahzad, S. K. (2017). Usability evaluation of adaptive features in smartphones. Procedia Computer Science, 112, 2185-2194. [41] Ahmad, N. (2014). People Centered HMI's for Deaf and Functionally Illiterate Users (Doctoral dissertation, Politecnico di Torino-Universität Potsdam, Germany). [42] Ahmad, N., Shoaib, U., & Prinetto, P. (2015). Usability of Online Assistance From Semiliterate Users’ Perspective. International Journal of Human-Computer Interaction, 31(1), 55-64. [43] Shoaib, M., Hussain, I., Mirza, H. T., & Tayyab, M. (2017, July). The role of information and innovative technology for rehabilitation of children with Autism: A Systematic Literature Review. In Computational Science and Its Applications (ICCSA), 2017 17th International Conference on IEEE (pp. 1-10). 
work_hxcodjpquneutiug2zkfy44nea	----	A regression tree approach using mathematical programming Expert Systems With Applications 78 (2017) 347–357 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa A regression tree approach using mathematical programming Lingjian Yang a , Songsong Liu a , b , Sophia Tsoka c , Lazaros G. Papageorgiou a , ∗ a Centre for Process Systems Engineerig, Department of Chemical Engineering, University College London, Torrington Place, London WC1E 7JE, UK b School of Management, Swansea University, Bay Campus, Fabian Way, Swansea SA1 8EN, UK c Department of Informatics, King’s College London, Strand, London WC2R 2LS, UK a r t i c l e i n f o Article history: Received 13 July 2016 Revised 4 February 2017 Accepted 5 February 2017 Available online 9 February 2017 Keywords: Regression analysis Surrogate model Regression tree Mathematical programming Optimisation a b s t r a c t Regression analysis is a machine learning approach that aims to accurately predict the value of contin- uous output variables from certain independent input variables, via automatic estimation of their latent relationship from data. Tree-based regression models are popular in literature due to their flexibility to model higher order non-linearity and great interpretability. Conventionally, regression tree models are trained in a two-stage procedure, i.e. recursive binary partitioning is employed to produce a tree struc- ture, followed by a pruning process of removing insignificant leaves, with the possibility of assigning multivariate functions to terminal leaves to improve generalisation. This work introduces a novel method- ology of node partitioning which, in a single optimisation model, simultaneously performs the two tasks of identifying the break-point of a binary split and assignment of multivariate functions to either leaf, thus leading to an efficient regression tree model. Using six real world benchmark problems, we demon- strate that the proposed method consistently outperforms a number of state-of-the-art regression tree models and methods based on other techniques, with an average improvement of 7–60% on the mean absolute errors (MAE) of the predictions. © 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license. ( http://creativecommons.org/licenses/by/4.0/ ) 1 r i r p t m e m Z & O & ( i v s l i l a a o n t t o m ( C t t p b s e h 0 . Introduction In machine learning, regression analysis seeks to estimate the elationships between output variables and a set of independent nput variables by automatically learning from a number of cu- ated samples ( Sen & Srivastava, 2012 ). The primary goal of ap- lying a regression analysis is usually to obtain precise predic- ion of the level of output variables for new samples. Examples of ethodologies for regression analysis in the literature include lin- ar regression ( Seber & Lee, 2012 ), automated learning of algebraic odels for optimisation (ALAMO) ( Cozad, Sahinidis, & Miller, 2014; hang & Sahinidis, 2013 ), support vector regression (SVR) ( Smola Schlkopf, 2004 ), multilayer perception (MLP) ( Hill, Marquez, ’Connor, & Remus, 1994 ), K-nearest neighbour (KNN) ( Korhonen Kangas, 1997 ), multivariate adaptive regression splines (MARS) Friedman, 1991 ), Kriging ( Kleijnen, 2015 ), and regression tree. Quite often, one would like to also gain some useful insights nto the underlying relationship between the input and output ariables, in which case the interpretability of a regression method ∗ Corresponding author. E-mail addresses: lingjian.yang.10@ucl.ac.uk (L. Yang), ongsong.liu@swansea.ac.uk (S. Liu), sophia.tsoka@kcl.ac.uk (S. Tsoka), .papageorgiou@ucl.ac.uk (L.G. Papageorgiou). c i i l t ttp://dx.doi.org/10.1016/j.eswa.2017.02.013 957-4174/© 2017 The Authors. Published by Elsevier Ltd. This is an open access article u s also of great interest. Regression tree is a type of the machine earning tools that can satisfy both good prediction accuracy nd easy interpretation, and therefore have received extensive ttention in the literature. Regression tree uses a tree-like graph r model and is built through an iterative process that splits each ode into child nodes by certain rules, unless it is a terminal node hat the samples fall into. A regression model is fitted to each erminal node to get the predicted values of the output variables f new samples. The Classification and Regression Tree (CART) is probably the ost well known decision tree learning algorithm in the literature Breiman, Friedman, Olshen, & Stone, 1984 ). Given a set of samples, ART identifies one input variable and one break-point, before par- itioning the samples into two child nodes. Starting from the en- ire set of available training samples (root node), recursive binary artition is performed for each node until no further split is possi- le or a certain terminating criteria is satisfied. At each node, best plit is identified by exhaustive search, i.e. all potential splits on ach input variable and each break-point are tested, and the one orresponding to the minimum deviations by respectively predict- ng two child nodes of samples with their mean output variables s selected. After the tree growing procedure, typically an overly arge tree is constructed, resulting in lack of model generalisation o unseen samples. A procedure of pruning is employed to remove nder the CC BY license. ( http://creativecommons.org/licenses/by/4.0/ ) http://dx.doi.org/10.1016/j.eswa.2017.02.013 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2017.02.013&domain=pdf http://creativecommons.org/licenses/by/4.0/ mailto:lingjian.yang.10@ucl.ac.uk mailto:songsong.liu@swansea.ac.uk mailto:sophia.tsoka@kcl.ac.uk mailto:l.papageorgiou@ucl.ac.uk http://dx.doi.org/10.1016/j.eswa.2017.02.013 http://creativecommons.org/licenses/by/4.0/ 348 L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 i d t s w Y e p a S a v t v ( L p t i t s s o u e M i ( 2 t t v c 1 o ( t r c a e c o a t a s a o a c w a s q s f i e t l sequentially the splits contributing insufficiently to training accu- racy. The tree is pruned from the maximal-sized tree all the way back to the root node, resulting in a sequence of candidate trees. Each candidate tree is tested on an independent validation sam- ple set and the one corresponding to the lowest prediction error is selected as the final tree ( Breiman, 20 01; Wu et al., 20 08 ). Alter- natively, the optimal tree structure can be identified via cross val- idation. After building a tree, an enquiry sample is firstly assigned into one of the terminal leaves (non-splitting leaf nodes) and then predicted with the mean output value of the samples belonging to the leaf node. Despite its simplicity, good interpretation and wide applications ( Antipov & Pokryshevskaya, 2012; Bayam, Liebowitz, & Agresti, 2005; Bel, Allard, Laurent, Cheddadi, & Bar-Hen, 2009; Li, Sun, & Wu, 2010; Molinaro, Dudoit, & van der Laan, 2004 ), the simple rule of predicting with mean values at the terminal leaves often means prediction performance is compromised ( Loh, 2011 ). The conditional inference tree (ctree) tackles the problem of re- cursive partitioning in a statistical framework ( Hothorn, Hornik, & Zeileis, 2006 ). For each node, the association between each inde- pendent input feature and the output variable is quantified, using permutation test and multiple testing correction. If the strongest association passes a statistical threshold, binary split is performed in that corresponding input variable; otherwise the current node is a terminal node. Ctree is shown to avoid the problem of build- ing biased tree towards input variables with many distinct levels of values while ensuring the similar prediction performance. Since almost all the tree-based learning models are constructed using recursive partitioning, an efficient yet essentially locally op- timal approach, the evtree implements an evolutionary algorithm for learning globally optimal classification and regression trees ( Grubinger, Zeileis, & Pfeiffer, 2014 ), and is considered an alterna- tive to the conventional methods by globally optimising the tree construction. Evtree searches a tree structure that takes into ac- count the accuracy and complexity, defined as the number of ter- minal leaves. Due to the exponentially growing size of the prob- lem, evolutionary methods are employed to identify a quality fea- sible solution. M5’, also knows as M5P, is considered as an improved version of CART ( Quinlan, 1992; Wang & Witten, 1997 ). The tree growing process is the same as that of the CART, while several modifica- tions have been introduced in tree pruning process. After the full size tree is produced, a multiple linear regression model is fitted for each node. A metric of model generalisation is defined in the original paper taking into account training error, the numbers of samples and model parameters. The constructed linear regression function for each node is then simplified by removing insignificant input variables using a greedy algorithm in order to achieve locally maximal model generalisation metric. Tree pruning starts from the bottom of the tree and is implemented for each non-leaf nodes. If the parent node offers higher model generalisation than the sum of the two child nodes, then the child nodes are pruned away. When predicting new samples, the value computed at the corresponding terminal node is adjusted by taking into account the other pre- dicted values at the intermediate nodes along the path from the terminal to the root node. The fitting of linear regression functions at leaf nodes improves the prediction accuracy of the regression tree learning model. M5’ is then further extended into Cubist ( RuleQuest, 2016 ), a commercially available rule-based regression model, which has re- ceived increasing popularity recently ( Kobayashi, Tsend-Ayush, & Tateishi, 2013; Minasny & McBratney, 2008; Moisen et al., 2006; Peng et al., 2015; Rossel & Webster, 2012 ). M5’ is employed to grow a tree first, which is then collapsed into a smaller set of if-then rules by removing and combining paths from the root to the ter- minal nodes. It is noted here that the if-then rules resulted from Cubist method can be overlapping, i.e. a sample can be assigned nto multiple rules, where all the predictions are averaged to pro- uce a final value. This ambiguity decreases the interpretability of he rule model. The Smoothed and Unsmoothed Piecewise-Polynomial Regres- ion Trees (SUPPORT) is another regression tree learning algorithm, hose foundation is based on statistics ( Chaudhuri, Huang, Loh, & ao, 1994 ). Given a set of samples, SUPPORT fits a multiple lin- ar regression function and computes the deviation of each sam- le. The samples with positive deviations and negative deviations re respectively assigned into two classes. For each input variable, UPPORT compares the distribution of the two classes of samples long this input variable by applying two-sample t test. The input ariable corresponding to the lowest P value is selected as split- ing node and the average of the two class mean on this splitting ariable is taken as break-point. The Generalised, Unbiased, Interaction Detection and Estimation GUIDE) adopts similar philosophy as the SUPPORT ( Loh, 2002; oh, He, & Man, 2015 ). Given a node, the same step of fitting sam- les with a linear regression model and separating samples into wo classes based on the sign of deviations is employed. For each nput variable, its numeric values are binned into a number of in- ervals before a chi-square test is used to determine its level of ignificance. The most significant input variable is used for binary plit. In terms of break-point determination, either a greedy search r median of the two class mean on this splitting variable can be sed. More other variants of the above regression tree models also xist in the literature, including SECRET ( Dobra & Gehrke, 2002 ), ART ( Elish, 20 09; Friedman, 20 02 ), SMOTI ( Malerbao, Espos- to, Ceci, & Appice, 2004 ), MAUVE ( Vens & Blockeel, 2006 ), BART Chipman, George, & McCulloch, 2010 ) and SERT ( Chen & Hong, 010 ), etc. In the above classic regression tree methodologies, the tradi- ional means of node splitting are dominated by either exhaus- ively searching the candidate split corresponding to the maximum ariance reduction by predicting of mean output values in two hild nodes ( Breiman et al., 1984; Quinlan, 1992; Wang & Witten, 997 ), or examining distribution of sample deviations from fitting ne linear regression function to all the samples in the parent node Chaudhuri et al., 1994; Loh, 2002 ). However, it is noticed that for hose algorithms where terminal leaf nodes are fitted with linear egression functions ( Quinlan, 1992; Wang & Witten, 1997 ), the hoice of splitting variable, break-point and regression coefficients re done sequentially, i.e. the splitting variable and break-point are stimated during tree growing procedure while regression coeffi- ients for each child node are computed at pruning step. A theoretically better node splitting strategy is to simultane- usly determine the splitting feature, the position of break-point nd the regression coefficients for each child node. In this case, he quality of a split can be directly calculated as the sum of devi- tions of all samples in either subset. A straightforward exhaustive earch algorithm for this problem can be: for each input variable nd each break-point, samples are separated into two subsets and ne multiple linear regression is fitted for each subset. After ex- mining all possible splits, the optimal split is chosen as the one orresponding to the minimum sum of deviations. The problem ith this approach is, however, that as the numbers of samples nd input variables grow, the quantity of multiple linear regres- ion functions need to be evaluated increases exponentially, re- uiring excessive computational time. For example, given a regres- ion problem of 500 samples and 10 input variables, we assume or each input variable, each sample takes a unique value. Then t requires construction of 9980 ( = 499 × 10 × 2) multiple lin- ar regression functions in order to find the optimal split for only he root node, which will only become worse as the tree grows arger. L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 349 g 2 n p f o u f p w f c m o d v p b o f 2 g m p i l o a s g a f n a s n a m n d o c 2 t c b g d d g w g i s t s n a f i m b i l β n r p a t 2 t t s b h a I In this work, we adopt a recently proposed mathematical pro- ramming optimisation model ( Yang, Liu, Tsoka, & Papageorgiou, 016 ), which solves the problem of splitting a node into two child odes to global optimality in affordable computational time. In our roposed framework, tree leaf nodes are fitted with polynomial unctions and recursive partition is permitted when the amount f reduction in deviation achieved by node splitting is above a ser-specific value, which is also the only tuning parameter in our ramework. Since the size of the tree is controlled via the tuning arameter, the pruning procedure is not implemented. The rest of the paper is structured as follows: In Section 2 , e describe the main features of the optimisation model adopted rom literature and introduces the framework of our proposed de- ision tree building process. In Section 3 , a number of bench- ark regression problems are employed to test the performance of ur proposed method. A comprehensive sensitivity analysis is con- ucted to evaluate how prediction accuracy varies with different alues of the tuning parameter. Later, prediction accuracy of our roposed method is compared against a number of decision tree ased algorithms and some other state-of-the-art regression meth- ds. Section 4 presents our main conclusions and discusses some uture directions. . Method In our previous work ( Yang et al., 2016 ), we have proposed a re- ression method based on piece-wise linear functions, named seg- ented regression. Segmented regression identifies multiple break- oints on a single independent variable and partitions the samples nto multiple regions, each one of which is fitted with a multiple inear regression function so as to minimise the absolute deviation f the samples. The core element of the segmented regression is mathematical programming optimisation model that, given one ingle input variable as splitting variable and the number of re- ions, simultaneously optimises the positions of the break-points nd the regression coefficients of one multiple linear regression unction for each region. In this work, we adopt this optimisation model to optimise bi- ary splitting of nodes. Given a node and a single input variable s splitting variable, the optimisation model is solved to find the ingle break-point and the regression coefficients for the two child odes. The model is solved when each input variable in turn serves s splitting variable once, and the input variable giving the mini- um absolute deviation is selected for splitting the current parent ode. Recursive node splitting terminates when the reduction in eviation drops below a user-specific threshold value. Below, the verview of the regression approach, and the detailed mathemati- al programming model for node partitioning are presented. .1. Regression tree approach As for other regression tree learning algorithms, recursive split- ing is used to grow the tree from root node until a split of node annot yield sufficient reduction in deviation. The pseudocode for uilding a tree is given below. Proposed regression tree algorithm Step 1. Fit a polynomial regression function of order 2 to root node minimising absolute deviation, recorded as ERROR root . Step 2. Start from the root node as the current node, and let E RROR current = E RROR root . Step 3. In each current root, for each input variable m, specify it as splitting variable ( m = m ∗ ) and solve the proposed Optimal Piece-wise Linear Regression Analysis model ( OPLRA ). The deviation is noted as ERROR split m . Step 4. Identify the best split corresponding to the minimum absolute deviation, noted as E RROR split = min m E RROR split m . Step 5. If E RROR current − E RROR split ≥ β × E RROR root , the current node is split; otherwise the current node is finished as a terminal node. Step 6. Apply step 3–5 to each remaining child node in turn. Given training samples, the first step of our proposed tree rowing strategy is to fit a polynomial regression function of or- er of 2 to the entire set of training samples minimising absolute eviation, which is noted as ERROR root . The used polynomial re- ression function can provide higher prediction accuracy. Note that hen the coefficient of the quadratic term is zero, the obtained re- ression model is a linear function. The absolute deviation is min- mised here, due to its simplicity and ease of optimisation. The ab- olute deviation of root node, multiplied by a scaling parameter β, aking value between 0 and 1 , is specified as the condition for node plitting. In other words, the current node is split into two child odes, only if the optimal split of the node results in reduction of bsolute deviation being greater than β × ERROR root . Then starting rom the root node as the current node, each feature m is specified n turn as splitting feature m ∗ once, while solving model OPLRA inimising the sum of absolute deviations of two child nodes. The est split of the current node is identified as the one correspond- ng to minimum absolute error. If the best split brings down abso- ute deviation from the current node ( ERROR current ) by more than × ERROR root , then the split takes place; otherwise the current ode is finalised as terminal leaf node. Note that the tuning pa- ameter β determines the size of the developed tree, and an ap- ropriate value of β can avoid the overfitting on the training data, nd achieve good prediction accuracy for testing. The flowchart of he whole procedure is illustrated in Fig. 1 . .2. Mathematical programming model for node partitioning For a given current node n and one feature m * for potential par- ition, the proposed mathematical programming model for the op- imal node split, OPLRA , is presented in this section. The indices, ets, parameters and variables associated with the model are listed elow. For better separation between the parameters and variables, ere lower case letters are for parameters, while upper case letters re for variables: ndices c child node of the current parent node n; c = l represents left child node,and c = r represents right child node m feature/independent input variable, m = 1 , 2 , . . . , M m ∗ the feature where sample partition takes place n the current parent node s samples in the data set, s = 1 , 2 , . . . , S Sets C n set of child nodes of the current parent node n S n set of samples in the current parent node n Parameters a sm numeric value of sample s on feature m y s real output value of sample s u a suitably large positive number � a suitably small positive number Continuous variables B c intercept of regression function in child node c D s absolute deviation between predicted output and real output for sample s P c s predicted output for sample s in child node c W 1 c m , W 2 c m regression coefficients for feature m in child node c X m ∗ break-point on partition feature m ∗ 350 L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 Fig. 1. Flowchart of the proposed regression tree approach. c w s s n c p P v Binary variables F c s 1 if sample s falls into child node c; 0 otherwise Binary variables F c s , taking value of either 0 or 1 , are introduced to model if sample s belongs to child node c or not. Modelling of which sample belongs to either child node is achieved with the following constraints: a sm ∗ ≤ X m ∗ − � + u (1 − F c s ) ∀ s ∈ S n , c = l, m ∗ (1) X m ∗ + � − u (1 − F c s ) ≤ a sm ∗ ∀ s ∈ S n , c = r, m ∗ (2) When sample s is assigned into left child node (i.e. F c s = 1 when c = l), Eq. (1) becomes A sm ∗ ≤ X m ∗ − � while Eq. (2) becomes re- dundant. On the other hand, when sample s is assigned into right hild node (i.e. F c s = 1 when c = r), Eq. (2) becomes A sm ∗ ≥ X m ∗ + � hile Eq. (1) is redundant. The insertion of � is to ensure strict eparation of the samples into two child nodes. The following con- traints restrict that each sample belongs to one and only one child ode: ∑ ∈ C n F c s = 1 ∀ s ∈ S n (3) For each child node c , polynomial functions of order 2 is em- loyed to predict the value of samples ( P c s ): c s = ∑ m a 2 sm W 2 c m + ∑ m a sm W 1 c m + B r ∀ s ∈ S n , c ∈ C n (4) For any sample s , its training error is equal to the absolute de- iation between the real output and the predicted output for the L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 351 c w D D i m a c t b m c B t p t 2 q a d l b f n m n d 3 h s t c r p o b 2 t f v t b n p o s c s m b n Table 1 Summary of benchmark data sets. Case study Number of samples Number of features Yacht hydrodynamics 308 6 Concrete strength 1030 8 Energy efficiency heating 768 8 Energy efficiency cooling 768 8 Airfoil 1503 5 White wine quality 4898 11 g t m f a m f A o i c p a T d m t h m o i s f a a w f 1 d i t i M c C ‘ K e l b M i b w t m i a 0 7 W hild node c where it belongs to (i.e. F c s = 1 ), and can be expressed ith the following two equations: s ≥ y s − P c s − u (1 − F c s ) ∀ s ∈ S n , c ∈ C n (5) s ≥ P c s − y s − u (1 − F c s ) ∀ s ∈ S n , c ∈ C n (6) The objective function is to minimise the sum of absolute train- ng errors of splitting the current node n into its child nodes: in ∑ s ∈ S n D s (7) The final OPLRA model consists of a linear objective function nd several linear constraints, and the presence of both binary and ontinuous variables define an MILP problem, which can be solved o global optimality by standard solution algorithms, for example ranch and bound. The optimisation model simultaneously opti- ises the break-point ( X m ∗ ), the allocation of samples into two hild nodes ( F c s ) and the regression coefficients ( W 1 c m , W 2 c m and c ) to achieve the least absolute deviation. Another advantage of his optimisation model is that there is no need to pre-process in- ut variable, i.e. input variables do not need to be binned into in- ervals for analysis. .3. Prediction for new samples After the regression tree is determined, prediction of new en- uiry samples can easily be performed. A new sample is firstly ssigned to one of the terminal leaf node, before yielding a pre- iction using the multivariate function derived for that particu- ar node. The predicted output value, if lies outside the interval ounded by the minimum and maximum of fitted output values or training samples in that particular node, is then adjusted to the earest bound. The proposed regression tree approach, referred to as Mathe- atical Programming Tree (MPTree) in this paper, is applied to a umber of real world benchmark data sets in the next section to emonstrate its applicability and efficiency. . Results and discussion In this section, we aim to comprehensively evaluate the be- aviour of the proposed MPTree using real world benchmark data ets. We first conduct a comprehensive sensitivity analysis for the uning parameter β in order to identify a robust value that gives onsistently good prediction accuracy. After that, prediction accu- acy comparison is performed to evaluate MPTree against certain opular regression tree learning algorithms in literature and some ther regression methodologies. A total number of 6 real world regression data sets have een downloaded from UCI machine learning repository ( Lichman, 013 ). The first regression problem Yacht Hydrodynamics predicts he residuary resistance of sailing yachts at the initial design stage rom 6 independent features describing the hull dimensions and elocity of the boat, including longitudinal position of the cen- re of buoyancy, prismatic coefficient, length-displacement ratio, eam-draught ratio, length-beam ratio and Froude number. The ext example, Concrete Strength ( Yeh, 1998 ), studies how com- ressive strength of different concrete are affected by attributes f the concretes. There are 1030 samples with 8 input attributes, uch as cement, blast furnace slag, fly ash, water, superplasticizer, oarse aggregate, fine aggregate and age. Energy Efficiency data ets ( Tsanas & Xifara, 2012 ) are obtained by running simulation odel. There are 768 samples, with each corresponding to one uilding shape, described by 8 features including relative compact- ess, surface area, wall area, root area, overall height, orientation, lazing area and glazing area distribution. The aims are to establish he relationship between either heating or cooling load require- ent of the building and the characteristics of these building. Air- oil data set concerns how the different frequencies, chord lengths, ngles of attack, free-stream velocities and suction side displace- ent thicknesses can predict the sound pressure level of an air- oil. The last case study, White Wine Quality ( Cortez, Cerdeira, lmeida, Matos, & Reis, 2009 ), aims to associate expert preference f white wine taste with 11 physicochemical features of the wines, ncluding fixed acidity, volatile acidity, citric acid, residual sugar, hlorides, free sulfur dioxide, total sulfur dioxide, density, pH, sul- hates and alcohol. The details of these data sets are provided s the supplementary material, and their sizes are summarised in able 1 . For each regression problem, we employ a 5-fold cross vali- ation to estimate the predictive accuracy of various regression ethods. Given a data set, 5-fold cross validation randomly splits he samples into 5 subsets of roughly equal size. One subset is old out as testing set, while the other 4 subsets of samples are erged to form training set. MPTree constructs a regression tree n the training set, whose prediction accuracy is estimated us- ng the holdout testing set. The process continues until each sub- et is hold out once as testing set. We conduct 10 rounds of 5- old cross validation by performing different random sample splits, nd the mean absolute errors (MAE) of the prediction are aver- ged over 50 testing sets as the final error. For each data set, e normalise each independent input variable with the following ormula so that the scaled input data take value between 0 and : A sm = A ′ sm −min s A ′ sm max s A ′ sm −min s A ′ sm ∀ s, m, where A ′ sm denotes the raw input ata. To assess the relative competitiveness of the proposed MPTree n terms of prediction accuracy, we compare the proposed MPTree o a number of popular regression methods in literature, includ- ng CART, M5’, Cubist, linear regression, SVR, MLP, Kriging, KNN, ARS, segmented regression ( Yang et al., 2016 ) and ALAMO. CART, tree, evtree and Cubist are implemented in R ( R Development ore Team, 2008 ) using the packages ‘rpart’, ‘party’, ‘evtree’ and Cubist’, respectively. M5’, linear regression, SVR, MLP, kriging and NN are implemented in WEKA machine learning software ( Hall t al., 2009 ). For KNN, the number of nearest neighbours is se- ected as 5, while for other methods their default settings have een retained. We use the MATLAB toolbox called ARESlab for ARS. ALAMO is reproduced using the General Algebraic Model- ng System (GAMS) ( GAMS Development Corporation, 2014 ), and asis function forms including polynomial of degrees up to 3, pair- ise multinomial terms of equal exponents up to 3, exponen- ial and logarithmic forms are provided for each data set. Seg- ented regression and the proposed MPTree are also implemented n GAMS. ALAMO, segmented regression and our proposed MPTree re solved using CPLEX MILP solver, with optimality gap set as . All computational runs were performed on a 64-bit Windows based machine with 3.20 GHz six-core Intel Xeon processor 3670 and 12.0 GB RAM. 352 L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 Fig. 2. Sensitivity analysis for β of all data sets. t t l A s s S c M r a f o a a v M fi 3.1. Sensitivity analysis for β In this section, we first perform a comprehensive sensitivity analysis on the single tuning parameter β in the proposed MPTree. Recall in the tree growing procedure, β controls termination of re- cursive node splitting. A node is split into two child nodes if the optimal split leads to reduction of absolute training deviation be- ing more than a threshold value, defined as the amount of abso- lute training deviation of a multiple linear regression analysis on the entire set of training samples ERROR root multiplied by the scal- ing parameter β. The tree grows larger as β decreases. Identifying a suitable value for β is a non-trivial problem as an excessively high value would terminate the node splitting prematurely with- out adequately describing the data, while a very small value can over-fit the unseen samples by constructing very large trees. In this work, we test a series of values, including 0.005, 0.01, 0.015, 0.025, 0.05 and 0.1 . The results of the sensitivity analysis are presented in Fig. 2 . According to Fig. 2 , we can clearly observe a phenomenon that as β is reduced from 0.1 to 0.015 , prediction error almost mono- onically drops. This improved prediction performance can be at- ributed to the fact that decreased β allows the tree to grow arger, and thus better describing the latent pattern in the data. s β further lowers down, MAE can even further decreases for ome examples, including Energy Efficiency Heating and Airfoil. For ome other examples, including Yacht Hydrodynamics, Concrete trength, Energy Efficiency Cooling and White Wine Quality, more omplicating trees do not predict unseen testing samples well, as AE worsens. It is well known that in data mining, parameter fine tuning is equired for a particular method to reach optimal performance for specific data set. Thus, it is our interest here to identify a value or β that corresponds to robust prediction accuracy for a range f different tested benchmark examples. In this study, β = 0.015 ppears to yield overall robust and accurate prediction as it usu- lly leads to lowest or second lowest MAE among all the tested alues. Higher values of β are shown to give significantly higher AE, while smaller value of β sometimes leads to noticeable over- tting, thus compromising the robustness of its performance. L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 353 Table 2 Prediction accuracy comparison across different regression methods, in terms of MAE. The proposed MPTree method is high- lighted in italic, and the best prediction accuracy of each data set is given in bold. Yacht Concrete Energy efficiency Energy efficiency Airfoil White wine hydrodynamics strength heating cooling quality Tree-based methods MPTree 0 .58 3 .85 0 .35 0 .80 0 .015 0 .52 CART 1 .61 7 .22 2 .00 2 .38 0 .035 0 .60 Ctree 0 .81 5 .99 0 .63 1 .40 0 .029 0 .58 Evtree 1 .05 6 .44 0 .56 1 .59 0 .032 0 .59 M5’ 0 .96 4 .72 0 .69 1 .21 0 .021 0 .56 Cubist 0 .60 4 .29 0 .35 0 .89 0 .017 0 .56 Non-tree-based methods Linear regression 7 .27 8 .31 2 .09 2 .27 0 .037 0 .59 SVR 6 .45 8 .21 2 .04 2 .19 0 .037 0 .58 MLP 0 .81 6 .23 0 .99 1 .92 0 .035 0 .62 Kriging 4 .32 6 .22 1 .79 2 .04 0 .030 0 .58 KNN 5 .30 7 .07 1 .94 2 .15 0 .026 0 .54 MARS 1 .01 4 .87 0 .80 1 .32 0 .035 0 .57 Segmented regression 0 .71 4 .87 0 .81 1 .28 0 .029 0 .55 ALAMO 0 .79 8 .04 2 .72 2 .76 0 .032 0 .64 Fig. 3. Prediction accuracy (MAE) comparison across different regression methods. For each benchmark example, the original MAE achieved by different methods in Table 2 are normalised between 0% and 100%, with 0% representing the lowest MAE and 100% representing the highest MAE. 3 r t a t e m w i o d b i a m h n i a 4 c m m m a m 7 .2. Performance comparison across different regression methods After identifying a value (i.e. 0.015 ) for the only user-specific pa- ameter, β, in the proposed MPTree, we now compare the predic- ion performance of the MPTree against a number of state-of-the- rt regression methods. To ensure unbiased comparison, β is set o 0.015 thorough all examples studied. For each of the benchmark xamples, we compare the MAE achieved by various competing ethods. The detailed prediction accuracies reported in Table 2 , in hich the performance of the proposed MPTree method is shown n italic, and the best prediction accuracy, i.e. the smallest MAE, f each data set is given in bold. The proposed has the best pre- iction accuracy in all data sets, compared to regression methods ased on a wide range of tree and non-tree methodologies. Only n the Energy Efficiency Heating data set, Cubist achieves the same ccuracy as MPTree. Overall, MPTree achieves 7–60% of improve- ent on MAE compared to each of other regression models. We ave also implemented MPTree with linear function at each child ode, the results of which still show great competitiveness as be- ng either the top or the second best method in all examples and chieves the overall best performance with MAE values of 0.60, .16, 0.36, 1.00, 0.014 and 0.55, respectively. The comparative results are summarised in Fig. 3 . This Radar hart is plotted to comprehensively visualise the prediction perfor- ance of different methods across all 6 data sets. For each bench- ark example studied, we normalise the MAE achieved by all ethods in Table 2 to scaled values between 0% and 100%, with 0% nd 100% respectively denoting the lowest and the highest MAE. To aintain the readability of the plot, prediction accuracies of only methods are plotted. It is clearly observed from Fig. 3 that the 354 L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 Table 3 Prediction accuracy comparison across different tree-based regression methods in terms of MSE. The proposed MPTree method is highlighted in italic, and the best prediction accuracy of each data set is given in bold. Yacht Concrete Energy efficiency Energy efficiency Airfoil White wine hydrodynamics strength heating cooling quality MPTree 2 .03 43 .88 0 .26 2 .65 0 .0 0 05 0 .59 CART 5 .41 86 .24 6 .85 9 .40 0 .0020 0 .58 Ctree 2 .79 63 .72 1 .33 4 .37 0 .0014 0 .55 Evtree 3 .02 69 .12 1 .00 4 .44 0 .0015 0 .56 M5’ 3 .08 40 .72 0 .95 3 .26 0 .0 0 08 0 .53 Cubist 1 .07 37 .77 0 .27 2 .76 0 .0 0 06 0 .51 Fig. 4. Constructed tree by CART on Energy Efficiency Heating example. Boxes represent the terminal leaf nodes with labels inside, while circles represent other nodes, where the symbol inside refers to the feature where the split takes place. The splitting rules are given on the corresponding paths. M c p m M r s r p f 3 t s p proposed MPTree forms the smallest area across all data sets, and performs better than other implemented tree-based learning algo- rithms, including ctree, evtree, M5’ and Cubist, and non-tree-based models, including MLP and Segmented regression. Overall, MPTree demonstrates clear advantage over the counterparts by managing the lowest MAE value for each and every tested benchmark exam- ple (including SVR, Kriging and KNN where results are not shown here). It is undoubtedly that the proposed MPTree, by optimising simultaneously the position of break-point and regression coeffi- cients per child node, representing significant improvement com- pared with other tree models in literature. In this work, MAE is adopted as the performance metric of regression models, which might not be suitable for all the data sets. Besides, other approaches might provide better fittings over another performance metric, e.g., mean squared error (MSE), root mean squared error (RMSE), Akaike Information Criterion, etc. When we compare the prediction accuracy in terms of MSE of all tree-based methodologies, Table 3 shows that the post-processed b SE values from the optimal solutions of MPTree are still very ompetitive with MSE values from other methodologies, even the roposed MPTree aims to minimise MAE. Although the perfor- ance of MPTree is not as dominant as it is considering MAE, PTree still ranks first on three data sets out of six, and is compa- able with Cubist, which performs the best for the other three data ets. These results demonstrate the impact of performance met- ics on the predication performance, and the consideration of other erformance metrics in MPTree would be an interesting direction or future research. .3. Comparison of actual constructed trees by different regression ree methods Last section has demonstrated that the novel MPTree regres- ion tree learning method offers superior prediction capacity. Com- ared to certain regression methods whose output models cannot e interpreted, for example kernel-based SVR and MLP, tree learn- L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 355 Fig. 5. Constructed tree by M5’ on Energy Efficiency Heating example. Boxes represent the terminal leaf nodes with labels inside, while circles represent other nodes, where the symbol inside refers to the feature where the split takes place. The splitting rules are given on the corresponding paths. i s m i p l b a u b s b s p s s c m ( i t t s n a n p ng algorithms are well-known for their easy interpretability. The equence of the derived rules can be simply visualised as tree, aking it easily understandable and possible to gain some insights nto the underlying mechanism of the studied system. The inter- retability of a constructed tree model decreases as the tree grows arger. In this section, attention is turned into comparing the num- er of terminal leaf nodes of the trees constructed by CART, M5’ nd MPTree. Taking Energy Efficiency Heating as an example and sing all the available samples as training set, the trees grown y CART, M5’ and MPTree are presented in Figs. 4 , 5 and 6 , re- pectively, in which the terminal leaf nodes are represented by oxes, and other nodes in the trees are represented by circles. The ymbol in each circle represents the feature where the split takes lace. According to Fig. 4 , CART has built a simple tree for the 768- ample example. On the top of the tree, CART splits the entire et of samples on feature m1 at break-point of 0.361 into two hild nodes, which are in turn further split on feature m7 and 1 , respectively. There are a total number of 7 terminal leaf nodes TN1–TN7) and the depth of the tree is equal to 4. From Fig. 5 , t is apparent that M5’ has constructed a much larger tree than he CART. The top part of the M5’ tree is almost identical as the ree built by CART, which is not surprising as the two algorithms hare great similarity during tree growing procedure and only sig- ificantly different from each other on pruning procedure. Over- ll, the tree grown by M5’ has a depth of 8 and 24 terminal leaf odes (TN1–TN24) , which is much harder to understand and inter- ret. Fig. 6 visualises the actual tree built by our proposed MPTree 356 L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 Fig. 6. Constructed tree by MPTree on Energy Efficiency Heating example. Boxes represent the terminal leaf nodes with labels inside, while circles represent other nodes, where the symbol inside refers to the feature where the split takes place. The splitting rules are given on the corresponding paths. Table 4 The number of terminal leaf nodes of the constructed trees by different regression tree learning methods. Yacht Concrete Energy efficiency Energy efficiency Airfoil White wine hydrodynamics strength heating cooling quality CART 5 13 7 4 18 7 M5’ 4 10 24 24 44 55 MPTree 5 14 7 12 14 6 M m M i i r c o O o t r c m s m t w c c A s I E 6 t U method. The size of the derived tree is similarly small as that of CART with 7 terminal leaf nodes (TN1–TN7) and a depth of 3, yet the two trees are quite different as the root nodes of the two trees are split on different features. MPTree, optimising the node split- ting, picks feature m3 as partition feature, in contrast to feature m1 selected by CART. Overall on the Energy Efficiency Heating exam- ple, CART and MPTree appear to build trees that are small in size, while M5’ outputs a significantly larger tree. The same analysis has been repeated on the other 5 benchmark data sets, and the results of which are available in Table 4 . The same observation can be made that for the other examples, CART and MPTree derive trees of similar numbers of terminal leaf nodes, while M5’ sometimes builds trees of comparable sizes as the other two (i.e. Yacht Hydrodynamics and Concrete Strength) but more of- ten outputs trees of several folds larger (i.e. Energy Efficiency Heat- ing, Energy Efficiency Cooling, Airfoil and White Wine Quality). 4. Concluding remarks Regression analysis is a data-driven computational tool that aims to predict continuous output variables from a set of indepen- dent input variables. In this work, we have proposed a novel re- gression tree learning algorithm, named MPTree. An optimisation model OPLRA recently published in literature has been adopted to optimise the binary node splitting. Given a specified splitting feature, OPLRA simultaneously determines the break-point position and the coefficients of the polynomial regression function in either child node so as to minimise residuals. An algorithm is introduced for recursive partitioning to grow the tree. A number of 6 real-world benchmark data sets have been used to demonstrate the applicability and efficiency of the proposed PTree. Popular regression learning algorithms have been imple- ented for comparison, including tree-based CART, ctree, evtree, 5’ and Cubist, and methods based on various other principles, ncluding MARS, MLP, kriging, segmented regression, etc. Cross val- dation experiment has been used to estimate the predictive accu- acy of different methods. The results clearly indicate that MPTree onsistently offers a much improved prediction accuracy than the ther competing methods for each of the benchmark data set. verall, we show that the proposed MPTree builds regression trees f better quality by optimising the node splitting. In the near future, we aim to explore a few aspects to refine he MPTree method. The existing regression tree learning algo- ithms, including the proposed MPTree, perform binary splits re- ursively to keep the tree growing. Splitting a parent node into ultiple child nodes, instead of two, is likely to better explore the tructure of the data set. Another potential avenue is to optimise ultiple levels of splitting simultaneously. Note that most of the ree building methods consider only splitting one node at a time, hile a look-ahead scheme that optimises also splitting of grand- hild nodes would lead to enhanced prediction performance of the onstructed tree. cknowledgement Funding from the UK Engineering and Physical Sciences Re- earch Council (to LY, SL and LGP through the EPSRC Centre for nnovative Manufacturing in Emergent Macromolecular Therapies, P/I033270/1), the UK Leverhulme Trust (to ST and LGP, RPG-2012- 86), the European Union (to ST, HEALTH-F2-2011-261366), and he Centre for Process Systems Engineering (CPSE) at Imperial and niversity College London (to LY) are gratefully acknowledged. L. Yang et al. / Expert Systems With Applications 78 (2017) 347–357 357 S f R A B B B B C C C C C D E F F G G H H H K K K L L L L L M M M M P Q R R R S S S T V W W Y Y Z upplementary material Supplementary material associated with this article can be ound, in the online version, at 10.1016/j.eswa.2017.02.013 . eferences ntipov, E. A., & Pokryshevskaya, E. B. (2012). Mass appraisal of residential apart- ments: An application of random forest for valuation and a CART-based ap- proach for model diagnostics. Expert Systems with Applications, 39 (2), 1772–1778. http://dx.doi.org/10.1016/j.eswa.2011.08.077 . ayam, E., Liebowitz, J., & Agresti, W. (2005). Older drivers and accidents: A meta analysis and data mining application on traffic accident data. Expert Systems with Applications, 29 (3), 598–629. http://dx.doi.org/10.1016/j.eswa.2005.04.025 . el, L., Allard, D., Laurent, J., Cheddadi, R., & Bar-Hen, A. (2009). CART algorithm for spatial data: Application to environmental and ecological data. Computa- tional Statistics & Data Analysis, 53 (8), 3082–3093. http://dx.doi.org/10.1016/j. csda.2008.09.012 . reiman, L. (2001). Statistical modeling: The two cultures. Statistical Science, 16 (3), 199–231 . reiman, L. , Friedman, J. H. , Olshen, R. A. , & Stone, C. J. (1984). Classification and regression trees . Taylor & Francis . haudhuri, P. L. , Huang, M. , Loh, W. , & Yao, R. (1994). Piecewise-polynomial regres- sion trees. Statistica Sinica, 4 , 143–167 . hen, A., & Hong, A. (2010). Sample-efficient regression trees (SERT) for semicon- ductor yield loss analysis. IEEE Transactions on Semiconductor Manufacturing, 23 (3), 358–369. doi: 10.1109/TSM.2010.204 896 8 . hipman, H. A., George, E. I., & McCulloch, R. E. (2010). BART: Bayesian additive regression trees. The Annals of Applied Statistics, 4 (1), 266–298. doi: 10.1214/ 09- AOAS285 . ortez, P. , Cerdeira, A. , Almeida, F. , Matos, T. , & Reis, J. (2009). Modeling wine pref- erences by data mining from physicochemical properties. Decision Support Sys- tems, 47 (4), 547–553 . ozad, A. , Sahinidis, N. V. , & Miller, D. C. (2014). Learning surrogate models for sim- ulation-based optimization. AIChE Journal, 60 (6), 2211–2227 . obra, A. , & Gehrke, J. (2002). SECRET: A scalable linear regression tree algorithm. In Proceedings of the eighth ACM sigkdd international conference on knowledge discovery and data mining (pp. 4 81–4 87). New York, NY, USA: ACM . lish, M. O. (2009). Improved estimation of software project effort using multiple additive regression trees. Expert Systems with Applications, 36 (7), 10774–10778. http://dx.doi.org/10.1016/j.eswa.2009.02.013 . riedman, J. (2002). Stochastic gradient boosting. Computational Statistics and Data Analysis, 38 (4), 367–378. doi: 10.1016/S0167- 9473(01)0 0 065- 2 . Cited By 637. riedman, J. H. (1991). Multivariate adaptive regression splines. The Annals of Statis- tics, 19 (1), 1–67 . AMS Development Corporation (2014). GAMS – A user’s guide . Washington, DC, USA. rubinger, T. , Zeileis, A. , & Pfeiffer, K. (2014). evtree: Evolutionary learning of glob- ally optimal classification and regression trees in R. Journal of Statistical Soft- ware, 61 (1), 1–29 . all, M. , Frank, E. , Holmes, G. , Pfahringer, B. , Reutemann, P. , & Witten, I. H. (2009). The WEKA data mining software: An update. SIGKDD Explorations, 11 (1), 10–18 . ill, T. , Marquez, L. , O’Connor, M. , & Remus, W. (1994). Artificial neural network models for forecasting and decision making. International Journal of Forecasting, 10 (1), 5–15 . othorn, T. , Hornik, K. , & Zeileis, A. (2006). Unbiased recursive partitioning: A con- ditional inference framework. Journal of Computational and Graphical Statistics, 15 (3), 651–674 . leijnen, J. P. C. (2015). Regression and Kriging metamodels with their experimental designs in simulation: Review . CentER Discussion Paper Series No. 2015-035. obayashi, T. , Tsend-Ayush, J. , & Tateishi, R. (2013). A new tree cover percentage map in eurasia at 500 m resolution using modis data. Remote Sensing, 6 (1), 209–232 . orhonen, K. T. , & Kangas, A. (1997). Application of nearest neighbour regression for generalizing sample tree information. Scandinavian Journal of Forest Research, 12 (1), 97–101 . i, H., Sun, J., & Wu, J. (2010). Predicting business failure using classification and re- gression tree: An empirical comparison with popular classical statistical meth- ods and top classification mining methods. Expert Systems with Applications, 37 (8), 5895–5904. http://dx.doi.org/10.1016/j.eswa.2010.02.016 . ichman, M. (2013). UCI machine learning repository . http://archive.ics.uci.edu/ml . oh, W. Y. (2002). Regression trees with unbiased variable selection and interaction detection. Statistica Sinica, 12 (2), 361–386 . oh, W. Y. (2011). Classification and regression trees. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 1 (1), 14–23 . oh, W. Y. , He, X. , & Man, M. (2015). A regression tree approach to identify- ing subgroups with differential treatment effects. Statistics in Medicine, 34 (11), 1818–1833 . alerbao, D. , Esposito, F. , Ceci, M. , & Appice, A. (2004). Top-down induction of model trees with regression and splitting nodes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26 (5), 612–625 . inasny, B., & McBratney, A. B. (2008). Regression rules as a tool for predicting soil properties from infrared reflectance spectroscopy. Chemometrics and Intelli- gent Laboratory Systems, 94 (1), 72–79. http://dx.doi.org/10.1016/j.chemolab.2008. 06.003 . oisen, G. G., Freeman, E. A., Blackard, J. A., Frescino, T. S., Zimmermann, N. E., & Jr. , T. C. E. (2006). Predicting tree species presence and basal area in utah: A comparison of stochastic gradient boosting, generalized additive models, and tree-based methods. Ecological Modelling, 199 (2), 176–187. http://dx.doi.org/10. 1016/j.ecolmodel.2006.05.021 . olinaro, A. M., Dudoit, S., & van der Laan, M. J. (2004). Tree-based multivariate re- gression and density estimation with right-censored data. Journal of Multivariate Analysis, 90 (1), 154–177. http://dx.doi.org/10.1016/j.jmva.20 04.02.0 03 . eng, Y., Xiong, X., Adhikari, K., Knadel, M., Grunwald, S., & Greve, M. H. (2015). Modeling soil organic carbon at regional scale by combining multi-spectral im- ages with laboratory spectra. PLoS ONE, 10 (11), 1–22. doi: 10.1371/journal.pone. 0142295 . uinlan, R. J. (1992). Learning with continuous classes. In 5th Australian joint con- ference on artificial intelligence (pp. 343–348). Singapore: World Scientific . Development Core Team (2008). R: A language and environment for statistical com- puting . R Foundation for Statistical Computing . Vienna, Austria. ossel, R. A. V., & Webster, R. (2012). Predicting soil properties from the australian soil visible near infrared spectroscopic database. European Journal of Soil Science, 63 (6), 848–860. doi: 10.1111/j.1365- 2389.2012.01495.x . uleQuest (2016). Data mining with cubist . https://www.rulequest.com/cubist- info. html . eber, G. , & Lee, A. (2012). Linear regression analysis. Wiley Series in Probability and Statistics . Wiley . en, A. , & Srivastava, M. (2012). Regression analysis: Theory, methods, and applications . Springer New York . mola, A. J. , & Schlkopf, B. (2004). A tutorial on support vector regression. Statistics and Computing, 14 (3), 199–222 . sanas, A. , & Xifara, A. (2012). Accurate quantitative estimation of energy perfor- mance of residential buildings using statistical machine learning tools. Energy and Buildings, 49 (0), 560–567 . ens, C. , & Blockeel, H. (2006). A simple regression based heuristic for learning model trees. Intelligent Data Analysis, 10 (3), 215–236 . ang, Y. , & Witten, I. H. (1997). Induction of model trees for predicting continu- ous classes. Poster papers of the 9th European conference on machine learning . Springer . u, X. , Kumar, V. , Quinlan, R. J. , Ghosh, J. , Yang, Q. , Motoda, H. , et al. (2008). Top 10 algorithms in data mining. Knowledge and Information Systems, 14 (1), 1–37 . ang, L. , Liu, S. , Tsoka, S. , & Papageorgiou, L. G. (2016). Mathematical programming for piecewise linear regression analysis. Expert Systems with Applications, 44 , 156–167 . eh, I.-C. (1998). Modeling of strength of high-performance concrete using artificial neural networks. Cement and Concrete Research, 28 (12), 1797–1808 . hang, Y. , & Sahinidis, N. V. (2013). Uncertainty quantification in CO 2 sequestration using surrogate models from polynomial chaos expansion. Industrial and Engi- neering Chemistry Research, 52 (9), 3121–3132 . http://dx.doi.org/10.1016/j.eswa.2017.02.013 http://dx.doi.org/10.1016/j.eswa.2011.08.077 http://dx.doi.org/10.1016/j.eswa.2005.04.025 http://dx.doi.org/10.1016/j.csda.2008.09.012 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0006 http://dx.doi.org/10.1109/TSM.2010.2048968 http://dx.doi.org/10.1214/09-AOAS285 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0011 http://dx.doi.org/10.1016/j.eswa.2009.02.013 http://dx.doi.org/10.1016/S0167-9473(01)00065-2 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0014 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0014 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0022 http://dx.doi.org/10.1016/j.eswa.2010.02.016 http://archive.ics.uci.edu/ml http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0026 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0026 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0028 http://dx.doi.org/10.1016/j.chemolab.2008.06.003 http://dx.doi.org/10.1016/j.ecolmodel.2006.05.021 http://dx.doi.org/10.1016/j.jmva.2004.02.003 http://dx.doi.org/10.1371/journal.pone.0142295 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0033 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0033 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0034 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0034 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0034 http://dx.doi.org/10.1111/j.1365-2389.2012.01495.x https://www.rulequest.com/cubist-info.html http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0037 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0037 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0037 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0037 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0038 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0038 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0038 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0038 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0039 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0039 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0039 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0039 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0040 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0040 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0040 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0040 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0041 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0041 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0041 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0041 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0042 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0042 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0042 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0042 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0043 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0043 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0043 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0043 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0043 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0043 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0043 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0043 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0044 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0044 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0044 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0044 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0044 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0044 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0045 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0045 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0046 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0046 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0046 http://refhub.elsevier.com/S0957-4174(17)30095-7/sbref0046 A regression tree approach using mathematical programming 1 Introduction 2 Method 2.1 Regression tree approach 2.2 Mathematical programming model for node partitioning 2.3 Prediction for new samples 3 Results and discussion 3.1 Sensitivity analysis for 3.2 Performance comparison across different regression methods 3.3 Comparison of actual constructed trees by different regression tree methods 4 Concluding remarks Acknowledgement Supplementary material References 
work_hxud3ocdqbgm7meeqxuk5h3hbq	----	A Practical Outlier Detection Approach for Mixed-Attribute Data Mohamed Bouguessa University of Quebec at Montreal Department of Computer Science Montreal, Qc, Canada bouguessa.mohamed@uqam.ca Abstract Outlier detection in mixed-attribute space is a challenging problem for which only few approaches have been proposed. However, such existing methods suf- fer from the fact that there is a lack of an automatic mechanism to formally discriminate between outliers and inliers. In fact, a common approach to outlier identification is to estimate an outlier score for each object and then provide a ranked list of points, expecting outliers to come first. A major problem of such an approach is where to stop reading the ranked list? How many points should be chosen as outliers? Other methods, instead of outlier ranking, implement var- ious strategies that depend on user-specified thresholds to discriminate outliers from inliers. Ad hoc threshold values are often used. With such an unprinci- pled approach it is impossible to be objective or consistent. To alleviate these problems, we propose a principled approach based on the bivariate beta mixture model to identify outliers in mixed-attribute data. The proposed approach is able to automatically discriminate outliers from inliers and it can be applied to both mixed-type attribute and single-type (numerical or categorical) attribute data without any feature transformation. Our experimental study demonstrates the suitability of the proposed approach in comparison to mainstream methods. Keywords: Data Mining, Outlier detection, Mixed-attribute data, Mixture model, Bivariate beta. Preprint submitted to Expert Systems with Applications February 4, 2016 1. Introduction1 Outlier detection is the practice of identifying data points which are consider-2 ably different from the remaining data (Aggarwal, 2013; Cao, Si, Zhang, and Jia,3 2010; Kriegel, Kroger, Schubert, and Zimek, 2011; Tan, Steinbach, and Kumar,4 2006). Outlier detection is also known as exception mining or deviation detec-5 tion because outlier points are exceptional in some sense or they have attribute6 values that deviate significantly from the expected or typical attribute values7 (Tan et al., 2006). Identifying outliers has practical applications in different do-8 mains such as intrusion and fraud detection, medical diagnosis, and many others9 (Fustes, Dafonte, Arcay, Manteiga, Smith, Vallenari, and Luri, 2013; Maervoet,10 Vens, Berghe, Blockeel, and Causmaecker, 2012; Alan and Catal, 2011). For11 example, in medical diagnosis, outliers may arise when the patient is afflicted12 with some disease, or suffers side-effects from a drug. Efficient detection of such13 outliers aids in identifying, preventing, and repairing the effects of malicious or14 faulty behavior (Penny and Jolliffe, 2011).15 Approaches to outlier detection can be categorised as supervised, semi-16 supervised, and unsupervised (Angiulli and Fassetti, 2014). In principle, super-17 vised, as well as semi-supervised learning methods, use labeled data to create18 a model which distinguishes outliers from inliers. On the other hand, unsuper-19 vised approaches do not require any labeled objects and detect outliers as points20 that are considerably dissimilar or inconsistent with respect to the remaining21 data using some quantified measures of outlierness (Aggarwal, 2013). To im-22 plement supervised and semi-supervised outlier detection methods, we should23 first label the training data (Wu and Wang, 2013). The problem here is that24 labeled data samples are more difficult, expensive and time consuming to obtain25 than unlabeled ones. This is why unsupervised approaches are more generally26 and widely used, since they do not require labeled information. In this paper27 we focus only on unsupervised outlier detection. For more surveys and details28 on outlier analysis, we refer the reader to Aggarwal (2013). In the following,29 we first describe some background information by providing a brief description30 2 of the key idea of some outlier detection approaches which are relevant to this31 work. Next, we discuss a number of elements that motivate this study and32 describe our contributions.33 1.1. Background Information34 Several unsupervised approaches have been proposed to identify outliers in35 numerical data. Such approaches can be broadly classified as statistical-based,36 distance-based, and density-based (Angiulli and Pizzuti, 2005). Statistical-37 based approaches attempt to fit the data set under investigation to a certain kind38 of distribution model (in general, the Gaussian model) (Yamanishi, Takeuchi,39 Williams, and Milne, 2000). Inliers occur in a high probability region of the40 model while outliers deviate strongly from the distribution. Distance-based41 approaches evaluate the outlierness of a point based on the distances to its k-42 nearest neighbors (kNN) (Angiulli and Pizzuti, 2005, 2002). Points with large43 kNN distance are defined as outliers. Finally, density-based approaches use the44 number of points within a specific local region of a data point in order to define45 local density (Breunig, Kriegel, Ng, and Sander, 2000). The local density val-46 ues could be then used to measure how isolated a point is with respect to the47 surrounding objects (Wu and Wang, 2013).48 The aforementioned approaches were specifically designed for numerical data.49 However, in several applications, attributes in real data sets are not numerical,50 but have categorical values. For categorical data sets, distance-based as well51 as density-based techniques must confront the problem of how to choose the52 measurement of distance or density (Wu and Wang, 2013). This poses a sig-53 nificant challenge in terms of generalizing algorithms for numerical data to the54 categorical domain (Aggarwal, 2013). To address this issue, a number of ap-55 proaches have been proposed to deal with categorical data (Koufakou, Secretan,56 and Georgiopoulos, 2011; He, Xu, Huang, and Deng, 2005). Some of these ap-57 proaches use the concept of frequent itemset mining to estimate an outlying58 score for each point. Inliers are those points which contain sets of items that59 co-occur frequently in the data sets, while outliers are likely to be the points60 3 Figure 1: Mixed-attribute data set with clustered objects and outliers. that contain rare itemsets (Koufakou et al., 2011).61 In many cases, categorical and numerical data are found in the same data62 set, as different attributes. This is referred to as mixed-attribute data (Ag-63 garwal, 2013). Outliers are those objects containing attribute values that are64 dissimilar to or inconsistent with the remaining objects in both the numerical65 and the categorical space (Koufakou and Georgiopoulos, 2010; Otey, Ghoting,66 and Parthasarathy, 2006). To illustrate, Fig. 1 shows a small data set composed67 of 18 objects with four numerical attributes (A1,A2,A3, and A4) and four cat-68 egorical attributes (A5,A6,A7, and A8). As can be seen from this figure, data69 objects O1,O2, . . . ,O15 are grouped into three clusters, while the remaining70 points, that is, O16,O17, and O18, are outliers which could not be located in71 any cluster. Note that in this figure each cluster is represented by a shade of72 gray and the unclustered background is white. Clusters thus exist in different73 subspaces spanned by different attributes. From Fig. 1, we can see that, in74 contrast to inliers (that is, the clustered objects), outliers contain dissimilar75 attribute values. In fact, compared to points that belong to clusters, outliers76 have non-correlated numerical attribute values along the numerical space and77 infrequent attribute values across the categorical space. On the other hand,78 4 from Fig. 1, we can see that objects grouped within clusters contain attribute79 values that are closely related along a specific subset of dimensions. For ex-80 ample, objects O1,O2,O3,O4, and O5, which form cluster 1, contain correlated81 attribute values along the numerical attributes A1,A2,A3, and a large number82 of common categorical attribute values along the categorical attribute A6.83 In practice, when faced with mixed-attribute data, it is common to discretize84 the numerical attributes and treat all the data as categorical so that categorical85 outlier detection algorithms can be applied to the entire data set. However, as86 suggested in Zhang and Jin (2010), discretizing numerical values into several87 bins could introduce noise or information losses. Improper discretizing thus88 would hamper the detection performance. To alleviate this problem, only few89 approaches (Koufakou and Georgiopoulos, 2010; Zhang and Jin, 2010; Otey,90 Ghoting, and Parthasarathy, 2006), have been proposed to handle outliers in91 the mixed-attribute space.92 The approach proposed in Otey et al. (2006) is based on the concept of93 frequent itemsets to deal with categorical attributes, and the covariance for94 continuous attributes. Specifically, the authors in Otey et al. (2006) assign to95 each point an outlier score inversely proportionate to its infrequent itemsets.96 They also maintain a covariance matrix for each itemset to compute anomaly97 scores in the continuous attribute space. A point is likely to be an outlier if it98 contains infrequent categorical sets, or if its continuous values differ from the99 covariance violation threshold. It is worth noting that the work proposed by100 Otey et al. (2006) has the merit of being the first outlier detection algorithm101 for mixed-attribute data.102 Koufakou and Georgiopoulos (2010) proposed an approach named ODMAD103 (Outlier Detection for Mixed Attribute Datasets). This algorithm calculates104 first, for each point in the categorical space, an outlier score which depends on105 the infrequent subsets contained in that point. Data points with score values less106 than a user-entered frequency threshold are isolated since they contain highly107 infrequent categorical values and may thus potentially correspond to outliers.108 This process results in a reduced data set based on which other outlier scores109 5 are calculated for the numerical space using the cosine similarity measure. As110 described in Koufakou and Georgiopoulos (2010), since minimum cosine simi-111 larity is 0 and maximum is 1, the data points with similarity close to 0 are more112 likely to be outliers. Experiments in Koufakou and Georgiopoulos (2010), show113 that ODMAD is fast and outperforms Otey’s approach.114 Zhang and Jin (2010) proposed a Pattern based Outlier Detection approach115 (POD). Patterns in Zhang and Jin (2010) are defined to describe the data objects116 as well as to capture interactions among different types of attributes. The more117 an object deviates from these patterns, the higher its outlier score. The authors118 in Zhang and Jin (2010) use logistic regression to learn patterns. These patterns119 are then used to estimate outlier scores for objects with mixed attribute. The120 top n points with the highest score values are declared as outliers. It is important121 to note that POD is not able to handle categorical values directly. To detect the122 target patterns, categorical attributes are first mapped into binary attributes.123 Then, these binary attributes are analyzed together with the original continuous124 attributes to detect outliers in the mixed-attribute space.125 1.2. Motivations and Contributions126 The area of outlier detection in mixed-attribute data offers several oppor-127 tunities for improvement. There are just very few approaches around in the128 literature so far, yet there are a number of problems still to solve. For instance,129 the output of POD (Zhang and Jin, 2010) is a ranked list of points that repre-130 sents the degree of outlierness of each point. The top n points in the list with131 the highest degree values are considered as outliers. This method encounters a132 major concern: at which level should this list be cut? Stated otherwise, starting133 from the first (ranked number one) object, how far should we go in that list? In134 general, no principled way is suggested on how many points should be selected135 from a ranked list. In some situations, the top n points are selected solely on the136 basis of specific knowledge of an application. Unfortunately, prior knowledge137 about the data under investigation is not always available.138 Since a ranked list has a particular disadvantage because there is no clear139 6 cut-off point of where to stop consulting the results, thresholding has turned140 out to be important in detecting outliers. For instance, ODMAD (Koufakou141 and Georgiopoulos, 2010) and the approach proposed by Otey et al. (2006)142 implement various strategies that depend on user-specified thresholds to detect143 outliers. In real situations, however, it is rarely possible for users to supply the144 threshold values accurately. Outlier detection accuracy can thus be seriously145 reduced if an incorrect threshold value is used. The experiments conducted in146 Koufakou and Georgiopoulos (2010) on the impact of using various threshold147 values on the outlier detection accuracy corroborate our claim. Finally, it is148 worth noting that ODMAD and Otey’s approach depend also on other input149 parameters such as the minimum support, the maximum length of itemset and150 the size of a window of categorical and numerical scores. Setting appropriate151 values of these parameters is not a straightforward task.152 To alleviate the aforementioned drawbacks of existing approaches for detect-153 ing outliers in the mixed-attribute space, we propose in this paper a principled154 approach which is able to automatically identify outliers. In our approach, we155 first estimate an outlying score, for each object, in the numerical space and156 another score in the categorical space. Next, we associate to each data point a157 two dimensional vector containing the estimated scores: one dimension contains158 the score estimated in the numerical space while the second one contains the159 outlying score calculated in the categorical space. We assume that, in both160 spaces, outliers are characterised by high score values. Finally, we propose a161 statistical framework, based on the bivariate beta mixture, in order to model162 the estimated outlier score vectors. The goal is to cluster the estimated vectors163 into several components such that data points associated to the component with164 the highest score values correspond to outliers.165 We have used the beta mixture mainly because it permits multiple modes166 and asymmetry and can thus approximate a wide variety of shapes (Dean and167 Nugent, 2013; Bouguila and Elguebaly, 2012; Bouguessa, 2012; Ji, Wu, Liu,168 Wang, and Coombes, 2005), while several other distributions are not able to do169 so. For example, the standard Gaussian distribution permits symmetric “bell”170 7 shape only. However, in many real life applications, the data under investigation171 is skewed with non-symmetric shapes. In this setting, as observed in Dean and172 Nugent (2013), and in Boutemedjet, Ziou, and Bouguila (2011), the standard173 Gaussian distribution may lead to inaccurate modeling (e.g. over estimation of174 the number of components in the mixture, increase of misclassification errors,175 etc.). In contrast to several distributions, the beta distribution is more flexible176 and powerful since it permits multiple symmetric and asymmetric modes, it177 may be skewed to the right, skewed to left or symmetric (Bouguila, Ziou, and178 Monga, 2006). This great shape flexibility of the beta distribution provides179 a better fitting of the outlier score vectors, which leads, in turn, to accurate180 detection of outliers. Our experimental results corroborate our claim.181 We summarize the significance of our work as follows:182 1. We view the task of identifying outliers from a mixture modeling perspec-183 tive, on which we devise a principled approach which is able to formally184 discriminate between outliers and inliers, while previous works provide185 only a ranked list of objects expecting outliers to come first.186 2. The proposed method automatically identifies outliers, while existing ap-187 proaches require human intervention in order to set a detection threshold188 or to manually define the number of outliers to be identified. Furthermore,189 our method is general, in the sense that it is not limited to mixed-attribute190 data and it can be applied to single-type attribute (numerical or categor-191 ical) data without any feature transformation.192 3. We conducted detailed experiments on several real data sets with mixed-193 attribute as well as with single-type attribute. The results suggest that194 the proposed approach achieves competitive results in comparison to main-195 stream outlier detection algorithms.196 The rest of this paper is organized as follows. Section 2 describes our ap-197 proach in detail. An empirical evaluation of the proposed method is given in198 Section 3. Finally, our conclusion is given in Section 4.199 8 Figure 2: Workflow of the proposed approach. 2. Proposed Approach200 We begin by fixing a proper notation to be used throughout the paper. Let201 D = {O1, . . . ,ON} be a set of N mixed-attribute data points. Each point con-202 tains An numerical attributes and Ac categorical attributes. The subspace of D203 that contains only numerical attributes is denoted Snum, while Scat refers to the204 subspace of D which contains only categorical attributes. In this paper, we rep-205 resent a data point Oi as Oi = [O n i ,O c i ], such that O n i = (o n i1, . . . ,o n il, . . . ,o n iAn )206 and Oci = (o c i1, . . . ,o c it, . . . ,o c iAc ), where onil designates the l th numerical attribute207 value and ocit corresponds to the t th categorical attribute value. In what follows,208 we will call onil a numerical 1D point and o c it a categorical 1D point.209 In our approach, we propose first to estimate, for each object Oi, an outlier210 score in the numerical space and another score in the categorical space. Then,211 we associate to each data point a two-dimensional outlier score vector ~Vi con-212 taining the two estimated scores. Finally, based on {~Vi}(i=1,...,N), we devise a213 probabilistic approach that uses the bivariate beta mixture model to automati-214 cally discriminate outliers from inliers in the full-dimensional space. Specifically,215 we first model {~Vi} as a mixture of m bivariate beta components. We then se-216 lect the component that corresponds to vectors with the highest score values.217 Data objects associated with the set of vectors that belong to the selected com-218 ponent correspond to outliers. Fig. 2 provides a simple visual illustration of the219 9 proposed approach. More details are given in the follows.220 2.1. Estimating Outlier Score in the Numerical Space221 It is widely accepted that outliers are data points that are considerably222 dissimilar from the remaining data (Aggarwal, 2013; Huang and Yang, 2011;223 Kriegel et al., 2011). In this setting, it is reasonable to assume that, in gen-224 eral, most of the attribute values of outlier objects projected along each of the225 dimensions in Snum tend to be far apart from the remaining attribute values226 (Tan et al., 2006). On the other hand, inliers have attribute values that tend to227 be closely related along several (or all) dimensions in Snum. Our assumption is228 based on the fact that inliers tend to form dense regions across several dimen-229 sions in the numerical space, while outliers are sparsely distributed. With this230 intuition in mind, we define the outlier score ON(Oni ) for an object Oi in the231 numerical attribute space as232 ON(Oni ) = An∑ l=1 log ( WN (o n il) + 1 ) (1) with233 WN (o n il) = k∑ j=1 [ dl ( onil,kNNj(o n il) )]2 (2) where, for a specific dimension l in Snum, kNNj(o n il) denotes the j th nearest (1D234 point) neighborhood of onil and dl denotes the distance between two numerical235 1D points. In our case, this distance simply corresponds to the absolute value236 of the difference between two numerical attribute values of a specific dimension.237 The outlier score defined in (1) is the sum, over all dimensions in the numer-238 ical space Snum, of the log of the weight WN (o n il). As described by (2), WN (o n il)239 computes the sum of the square of the distance between each 1D point onil and240 its k nearest neighborhoods in dimension l. Intuitively, a large value of WN (o n il)241 means that onil falls into a sparse region in which the k nearest neighborhood242 attribute values of onil are loosely related, while a small value indicates that o n il243 10 Figure 3: The estimated outlier scores in the numerical space for the data objects depicted by Fig. 1. belongs to a dense region in which the k nearest neighborhood of onil are closely244 related. Note that we have used the square power in (2) in order to favor the245 weight of the 1D points belonging to a sparse region.246 The weight WN (o n il) captures the degree of isolation of an attribute value247 with respect to its neighbors. The higher its weight, the more distant are its248 neighbors along dimension l of Snum. Accordingly, based on (2), we can surmise249 that objects that are sparsely distributed over Snum will receive high ON(Oni )250 values, while related points will receive low score values. This means that out-251 liers will be characterized by high score values in contrast to inliers. As an252 illustration, Fig. 3 shows the estimated outlier scores in the numerical space253 for the data objects depicted by Fig. 1. As can be seen from Fig. 3, outlier254 objects O16,O17, and O18 have high score values in comparison to inliers, that255 is, O1, . . . ,O15.256 It is important to note that we have used the logarithm function in (1)257 primarily to squeeze together the large values that characterize outliers and258 stretch out the smallest values, which correspond to inliers. This squeezing and259 stretching contributes to enhancing the contrast between largest and smallest260 values which helps in distinguishing outliers from the rest of the points. Finally,261 note that we have added 1 to WN (o n il) in equation (1) to avoid the null value in262 the calculation of the logarithm, since it is possible to have WN (o n il) = 0 in the263 11 likely case where an attribute has more than k duplicative values.264 It is clear that the calculation of the k nearest neighbors is, in general, an265 expensive task, especially when the number of data points N is very large. How-266 ever, since we are searching for the k nearest neighbors in the one-dimensional267 space, we can perform the task in an efficient way by sorting the values in each268 attribute and limiting the number of distance comparisons to a maximum of269 2k values. The computation of the kNN distance is sensitive to the value of270 k, which is a limitation common to all kNN based approaches. However, we271 believe the problem this limitation creates for our approach does not have a272 major impact. This is because, since we estimate the kNN distances in the273 one-dimensional space only, the choice of the value of k is not as critical as in a274 multi-dimensional case. As suggested in Bouguessa and Wang (2009), to gain a275 clear idea of the sparseness of the neighborhood of a 1D point, we suggest using276 k = √ N as a default value.277 2.2. Estimating Outlier Score in the Categorical Space278 Virtually, as suggested in previous studies (Koufakou et al., 2011; Koufakou279 and Georgiopoulos, 2010; He et al., 2005), outliers in the categorical space are280 those points that have infrequent attribute categorical values for all dimensions281 compared to normal points. This means that every categorical 1D point of282 outlier objects is infrequent across all dimensions of Scat, while inliers have283 several categorical 1D points which occur with higher frequency along several (or284 all) categorical attributes (Koufakou et al., 2011; Koufakou and Georgiopoulos,285 2010). Based on such a definition, the outlier score OC(Oci ) for an object Oi in286 the categorical attribute space is formulated as287 OC(Oci ) = Ac∑ t=1 log ( WC(o c it) ) (3) with288 WC(o c it) = f(o c it) (4) 12 where f(ocit) denotes the number of times o c it appears in a specific categorical289 dimension t of Scat.290 OC(Oci ) is defined as the sum, across all dimensions in the categorical space291 Scat, of the log of the weight WC(o c it), which, in turn, corresponds to the occur-292 rence frequency of ocit in the categorical attribute t. Here, it is clear that rare293 categorical attribute values projected along dimension t will receive low weight294 values, while larger WC(o c it) values indicate that o c it is shared by several objects295 within dimension t. Accordingly, based on (3), points that share common cate-296 gorical values across Scat will get large OC(Oci ) values, while data objects that297 have infrequent categorical values across Scat will receive low OC(Oci ) values. As298 a result, since outliers are those points whose attribute categorical values occur299 very rarely along each dimension in Scat (Koufakou et al., 2011), it is easy to300 see that small values of OC(Oci ) designate outliers and high scores correspond301 to inliers. Finally, note that, as with the numerical outlier score described by302 (1), we have used a logarithm function in (3) to enhance the contrast between303 larger and smaller weight values.304 In this paper, as mentioned in Section 1, we assume that outliers are char-305 acterized by large score values in contrast to inliers. However, as just discussed,306 large OC(Oci ) scores refer to inliers. To regularize such scores, we need to invert307 them. For this purpose, we simply take the difference between the observed308 score and the maximum possible estimated score OCmax. The inverted score is309 estimated as310 OCinv(Oci ) = OCmax −OC(O c i ) (5) It easy to show that this linear inversion doesn’t affect the ranking-stability of311 the inverted scores:312 OC(Oc1) ≤OC(O c 2) ⇐⇒ −OC(O c 1) ≥−OC(O c 2) ⇐⇒ OCmax −OC(Oc1) ≥OCmax −OC(O c 2) ⇐⇒ OCinv(Oc1) ≥OCinv(O c 2). 13 Figure 4: The estimated outlier scores in the categorical space for the data objects depicted by Fig. 1. Accordingly, based on such a linear inversion, outliers will receive large score313 values while inliers will receive the lowest score values. In the remainder of this314 paper, unless otherwise specified, we use only the inverted categorical outlier315 score values. As an illustration, Fig. 4 shows the estimated outlier scores in the316 categorical space for the data objects depicted by Fig. 1. As can be seen from317 Fig. 4, outlier objects O16,O17, and O18 have high score values in comparison318 to inliers, that is, O1, . . . ,O15.319 Finally, as the reader can notice, in our approach we treat numerical and320 categorical attributes independently in order to estimate outlier scores in the321 numerical and the categorical space. In other words, this means we assume the322 independence of both numerical and categorical attributes. Such an assump-323 tion is mainly based on the general definition of outliers, which relies on the fact324 that outlier objects contain attribute values that are dissimilar to or inconsistent325 with the remaining points. Stated otherwise, outliers may contain many atyp-326 ical attribute values across most (or all) attributes of the data in comparison327 to inliers. Accordingly, investigating individual attributes in order to localize328 attribute values that deviate significantly from the expected or typical attribute329 values is appropriate to effectively detect outliers in the whole space.330 14 2.3. Modeling Outlier Score Vectors331 Once the outlier scores are estimated in both the numerical and the categor-332 ical spaces, we now focus on how to automatically identify outliers in the mixed-333 attribute space. To this end, we associate to each object Oi a two-dimensional334 vector ~Vi such that the first element of this vector corresponds to the outlier335 score of Oi in the numerical space, while the second element represent the out-336 lier score of Oi in the categorical space. Then, based on the estimated vectors,337 we propose a probabilistic approach that uses the bivariate beta mixture model338 to automatically discriminate outliers from inliers in the full-dimensional space.339 The probabilistic model framework is described in the follows.340 2.3.1. The Bivariate Beta Mixture Model341 Since the beta distribution is defined on the interval [0,1], we should first,342 without loss of generality, normalize the estimated outlier score values between343 0 and 1. Let ~Vi = (Vi1,Vi2) T where Vi1 and Vi2 represent, respectively, the344 normalized outlier scores ON(Oni ) and OCinv(O c i ). Under a mixture of bivariate345 beta distribution,346 ~Vi ∼ M∑ m=1 λm Bm(~Vi|~xm,~ym) (6) where Bm(~Vi|~xm,~ym) is the mth bivariate beta distribution; M denotes the347 number of components in the mixture; ~x = {~x1, . . . ,~xM} and ~y = {~y1, . . . ,~yM}.348 ~xm and ~ym are the parameters of the m th component with ~xm = (xm1,xm2) T 349 and ~ym = (ym1,ym2) T . λ = {λ1, . . . ,λM} represents the mixing coefficients350 such that ∑M m=1 λm = 1 and λm > 0.351 The bivariate beta distribution can be obtained by cascading two beta vari-352 ables together, that is, each element in the two-dimensional vector ~Vi is a scalar353 beta variable. In other words, the bivariate beta is the product of two uni-354 variate beta densities. Accordingly, the probability density function of the mth355 15 bivariate beta component is expressed as356 Bm(~Vi|~xm,~ym) = 2∏ d=1 B(Vid|xmd,ymd) (7) B(Vid|xmd,ymd) is the probability density function of the univariate beta distri-357 bution which is given by358 B(Vid|xmd,ymd) = Γ(xmd + ymd) Γ(xmd)Γ(ymd) V xmd−1 id (1 −Vid) ymd−1 (8) where Γ(.) is the gamma function given by Γ(α) = ∫∞ 0 βα−1 exp(−β)dβ; β > 0.359 2.3.2. Maximum Likelihood Estimate360 A common approach for estimating the unknown parameters xmd and ymd,361 (m = 1, . . . ,M; d = 1, 2) is the maximum likelihood estimation technique. The362 likelihood function corresponding to the mth bivariate beta component Bm is363 defined as364 L ( Bm(~Vi|~xm,~ym) ) = ∏ ~Vi∈Bm Bm(~Vi|~xm,~ym) = ∏ ~Vi∈Bm 2∏ d=1 B(Vid|xmd,ymd) (9) The logarithm of the likelihood function is given by365 log [ L ( Bm(~Vi|~xm,~ym) )] = Nm∑ i=1 2∑ d=1 log [ B(Vid|xmd,ymd) ] (10) where Nm is the size of the m th component.366 We note that the parameters pair {xmd,ymd} is independent from all other pairs. The problem of estimating the parameters of the model can thus be reduced to the estimation of the parameters pair {xmd,ymd} independently over each dimension of the outlier score vectors belonging to component m. In this 16 setting, the value {x̂md, ŷmd} that maximizes the likelihood can be obtained by taking the derivative of the expectation of the log-likelihood function with respect to xmd and ymd and setting the gradient equal to zero as   ∂E ( log [ L(Bm(~Vi|~xm,~ym)) ]) ∂xmd ∂E ( log [ L(Bm(~Vi|~xm,~ym)) ]) ∂ymd   = 0 (11) where367 ∂E ( log [ L(Bm(~Vi|~xm,~ym)) ]) ∂xmd = Nm∑ i=1 [ ∂ ∂xmd log ( Γ(xmd + ymd) Γ(xmd)Γ(ymd) V xmd−1 id (1 −Vid) ymd−1 )] = Nm∑ i=1 [Γ′(xmd + ymd) Γ(xmd + ymd) − Γ′(xmd) Γ(xmd) + log(Vid) ] = Nm Γ′(xmd + ymd) Γ(xmd + ymd) −Nm Γ′(xmd) Γ(xmd) + Nm∑ i=1 log(Vid). (12) and368 ∂E ( log [ L(Bm(~Vi|~xm,~ym)) ]) ∂ymd = Nm∑ i=1 [ ∂ ∂ymd log ( Γ(xmd + ymd) Γ(xmd)Γ(ymd) V xmd−1 id (1 −Vid) ymd−1 )] = Nm∑ i=1 [Γ′(xmd + ymd) Γ(xmd + ymd) − Γ′(ymd) Γ(ymd) + log(1 −Vid) ] = Nm Γ′(xmd + ymd) Γ(xmd + ymd) −Nm Γ′(ymd) Γ(ymd) + Nm∑ i=1 log(1 −Vid). (13) Equations (11), (12) and (13) yield the following expression369   Nm [ ψ(xmd + ymd) −ψ(xmd) ] + ∑Nm i=1 log(Vid) Nm [ ψ(xmd + ymd) −ψ(ymd) ] + ∑Nm i=1 log(1 −Vid) ]   = 0 (14) 370 where ψ(.) is the digamma function given by ψ(α) = Γ′(α) Γ(α) .371 17 Since the digamma function is defined though an integration, a closed-372 form solution to (14) does not exist. So the parameters pair {xmd,ymd} can373 be estimated using the Newton-Raphson method (Ypma, 1995). Specifically,374 {xmd,ymd} are estimated iteratively:375   x (I+1) md y (I+1) md   =   x (I) md y (I) md  − [~hm]T [Hm]−1 (15) where I is the iteration index, hm and Hm are respectively the vector of the376 first derivatives and the matrix of the second derivatives of the log likelihood377 function of the mth component.378 The vector ~hm is defined as379 ~hm =   h1m h2m   =   ∂E ( log [ L(Bm( ~Vi| ~xm, ~ym)) ]) ∂xmd ∂E ( log [ L(Bm( ~Vi| ~xm, ~ym)) ]) ∂ymd   (16) and the matrix Hm is expressed as380 Hm =   ∂h1m ∂xmd ∂h1m ∂ymd ∂h2m ∂xmd ∂h2m ∂ymd  ,381 where382 ∂h1m ∂xm = Nm [ ψ′(xmd + ymd) −ψ′(xmd) ] , ∂h1m ∂ym = ∂h2m ∂xmd = Nm [ ψ′(xmd + ymd) ] , ∂h2m ∂ym = Nm [ ψ′(xmd + ymd) −ψ′(βmd) ] . (17) ψ′(.) is the trigamma function given by ψ′(α) = Γ′′(α) Γ(α) − [ Γ ′(α) Γ(α) ]2.383 The Newton-Raphson algorithm for the update of equation (15) converges,384 as our estimate of xmd and ymd change by less than a small positive value �385 with each successive iteration, to x̂md and ŷmd. Note that, we have used in386 18 our implementation the method of moments estimators of the beta distribution387 (Bain and Engelhardt, 2000) to define starting values for {x(0)md,y (0) md} in equation388 (15). In this technique, the expected mean of the distribution is equated to the389 sample mean and the expected variance to the sample variance. Specifically,390 the method of moments estimators are391 x̂ (0) md = µmd [ µmd(1 −µmd) σ2md − 1 ] , ŷ (0) md = (1 −µmd) [ µmd(1 −µmd) σ2md − 1 ] . (18) where µmd and σ 2 md denote respectively the sample mean and variance of the392 normalized outlier score vectors belonging to the mth component which are393 projected along dimension d.394 2.3.3. EM Algorithm for the Bivariate Beta Mixture395 Let P = {λ1, . . . ,λM,~x1, . . . ,~xM,~y1, . . . ,~yM} denote the set of parameters396 of the mixture and V = {~V1, . . . , ~VN} the set of the normalized outlier score397 vectors. The usual choice for obtaining the maximum likelihood of the distribu-398 tion parameters is the EM algorithm (Dempster, Laird, and Rubin, 1977). This399 algorithm is based on the interpretation of V as incomplete data. As mentioned400 in Figueiredo and Jain (2002), for finite mixture, the missing part is a set of N401 label vectors η = {~η1, . . . ,~ηN} associated with the N outlier score vectors, in-402 dicating to which component ~Vi belongs. Specifically, each ~ηi = (ηi1, . . . ,ηim) T 403 is a binary vector, where ηim = 1 if ~Vi belongs to component m and ηim = 0404 otherwise.405 The complete data is thus defined by the sets η and V. The likelihood of the406 complete data is then:407 L(V,η|P) = N∏ i=1 M∏ m=1 [ λm Bm(~Vi|~xm,~ym) ]ηim (19) 19 and the complete log likelihood is:408 log(L(V,η|P)) = N∑ i=1 M∑ m=1 ηim log [ λm Bm(~Vi|~xm,~ym) ] = N∑ i=1 M∑ m=1 ηim log [ λm 2∏ d=1 B(Vid|xmd,ymd) ] = N∑ i=1 M∑ m=1 ηim [ log(λm) + 2∑ d=1 log(B(Vid|xmd,ymd)) ] (20) The EM algorithm can now be used to estimate P. Specifically, the algo-409 rithm iterates between an Expectation step and an Maximization step in order410 to produce a sequence estimate {P̂}(I), (I = 0, 1, 2, . . . ), where I denotes the411 current iteration step, until the change in the value of the complete log-likelihood412 in (20) is negligible. Details of each step are given below.413 In the Expectation step: each latent variable ηim is replaced by its expecta-414 tion as follows415 η̂ (I) im = E[ηim|~Vi,P] = λ̂ (I) m Bm(~Vi|~̂xm,~̂ym)∑M j=1 λ̂ (I) j Bj(~Vi|~̂xj,~̂yj) (21) In the Maximization step: the mixing coefficients {λm} and the parameters416 {~x1, . . . ,~xM,~y1, . . . ,~yM} are calculated using the values of η̂im estimated in the417 Expectation step. Specifically, the mixing coefficients are calculated as418 λ̂(I+1)m = ∑N i=1 η̂ (I) im N , m = 1, . . . ,M (22) The parameters {~xm = (xm1,xm2)T}(m=1,...,M) and {~ym = (ym1,ym2)T}(m=1,...,M)419 are estimated using the Newton-Raphson algorithm, based on (15), as described420 in the previous subsection.421 Finally, note that, the EM algorithm requires the initial parameters of each422 component. Since EM is highly dependent on initialization, it will be help-423 ful to perform initialization by means of clustering algorithms (Figueiredo and424 Jain, 2002). For this purpose we implement the k-means algorithm in order to425 20 Algorithm 1: EM algorithm for the bivariate beta mixture Input : {~Vi}(i=1,...,N); M Output: P̂ = {λ̂1, . . . , λ̂M, ~̂x1, . . . , ~̂xM, ~̂y1, . . . , ~̂yM} begin1 // Initialization Apply the k-means algorithm to cluster the set {~Vi} into M components;2 Estimate the initial set of parameters of each component using (18);3 repeat4 // Expectation Estimate {η̂im}(i=1,...,N; m=1,...,M) using (21);5 // Maximization Estimate {λ̂m}(m=1,...,M) using (22);6 Estimate {x̂md, ŷmd}(m=1,...,M; d=1,2) using (15);7 until the change in (20) is negligible;8 Return P̂;9 end10 partition the set {~Vi}(i=1,...,N), into M components. Then, based on such par-426 tition, we estimate the initial parameters of each component using the method427 of moment estimator of the beta distribution (Bain and Engelhardt, 2000) and428 set them as initial parameters to the EM algorithm. The detailed algorithm is429 summarized in Algorithm 1.430 2.3.4. Estimating the Optimal Number of Components in the Mixture431 The use of mixture of the bivariate beta distribution allows us to give a432 flexible model to describe the outlier score vectors. To form such a model,433 we need to estimate M, the number of components, and the parameters for434 each component. Several model selection approaches have been proposed to435 estimate M (Bouguessa, Wang, and Sun, 2006). In this paper, we implemented436 a deterministic approach that uses the EM algorithm described in Algorithm437 1 in order to obtain a set of candidate models for the range value of M (from438 1 to M max, the maximum number of components in the mixture) which is439 assumed to contain the optimal M (Figueiredo and Jain, 2002). The number of440 components is then selected according to441 M̂ = argmin M { C ( P̂,M )} M=1,...,Mmax (23) 21 Algorithm 2: Estimating the number of components in the mixture Input : {~Vi}(i=1,...,N), M max Output: The optimal number of components M̂ begin1 for M = 1 to M max do2 if M==1 then3 Estimate {x̂d, ŷd}d=1,2 using (15);4 Estimate ICL−BIC(P̂,M) using (24);5 else6 Estimate the parameters of the mixture using Algorithm 1;7 Estimate ICL−BIC(P̂,M) using (24);8 end9 end10 Select M̂, such that M̂ = argmin M ICL−BIC(P̂,M); 11 end12 where C ( P̂,M ) is some model selection criterion. Ji et al. (2005) found that the Integrated Classification Likelihood-Bayesian Information Criterion (ICL-BIC) performs well in selecting the number of components in the beta mixture. ICL- BIC has been also used in Dean and Nugent (2013) to select the number of beta mixture components. Accordingly, we use in our method ICL-BIC to identify the optimal number of components. The ICL-BIC criterion is given by ICL − BIC(P̂,M) = −2 log(L(V, η̂|P̂)) + QM log(N) − 2 N∑ i=1 M∑ m=1 η̂im log(η̂im) (24) where QM denotes the number of parameters of the model with M components442 and log(L(V, η̂|P̂)) corresponds to logarithm of the likelihood at the maximum443 likelihood solution for the investigated mixture model. The number of compo-444 nents that minimize ICL − BIC(P̂,M) is considered to be the optimal value for445 M. The procedure for estimating the number of components in the mixture is446 summarized in Algorithm 2.447 2.3.5. Automatic Identification of Outlier448 Once the optimal number of components has been identified, we focus now449 on detecting the bivariate beta component that corresponds to outliers. To this450 22 end, we used the results of the EM algorithm in order to derive a classifica-451 tion decision about which outlier score vector ~Vi belongs to which component452 in the mixture. In fact, the EM algorithm yields the final estimated posterior453 probability η̂im, the value of which represents the posterior probability that ~Vi454 belongs to component m. We assign ~Vi to the component that corresponds to455 the maximum value of η̂im. We thus divide the set of outlier score vectors into456 several components. As discussed earlier, in our approach we assume that out-457 lier points are characterized by high score values. Therefore, we are interested458 by the bivariate beta component which contains vectors with the highest score459 values. To identify such a component, we first compute, for each component460 in the mixture, the average value of the numerical outlier scores and also the461 average value of the categorical outlier scores (that is, we compute the average462 of Vi1 and Vi2 per component). Then, we select the component with the largest463 average values as our target component. This simple strategy for determining464 which component to pick works well in practice since it fits our assumption,465 which is based on the fact that outlier points are characterized by large score466 values in both numerical and categorical space. Finally, once our target compo-467 nent is identified, we focus on the problem of detecting outlier objects. To this468 end, we identify the set of data objects that are associated with the outlier score469 vectors ~Vi that belong to the selected component. The identified objects are out-470 liers. The steps described in Algorithm 3 can be implemented to automatically471 identify outliers.472 Finally, it is worth noting that the proposed methodology could be also473 used to identify outlier objects in single-type (categorical or numerical) at-474 tribute data. In this particular case, we propose to associate to each object475 only one score (ON(Oni ) or OCinv(O c i ), depending on the attribute type of476 the data under investigation). Then, to automatically discriminate between477 outliers and inliers, we can model the estimated scores as a finite mixture dis-478 tribution using the univariate beta which is given by (8). Here, the problem479 is thus reduced from modeling a set of two-dimensional outlier score vectors480 {~Vi}(i=1,...,N) (in the case of mixed-attribute data) to modeling a list of scalar481 23 Algorithm 3: Automatic identification of outliers Input : A data set D Output: A set of outliers OUT begin1 Estimate {ON(Oni )}(i=1,...,N) using (1);2 Estimate {OCinv(Oci )}(i=1,...,N) using (3) and (5);3 Associate, to each object Oi in D, a vector ~Vi = (Vi1,Vi2)T where Vi1 and4 Vi2 represent, respectively, the normalized values of ON(Oni ) and OCinv(Oci ) in [0,1] ; Apply Algorithm 2 to cluster {~Vi}(i=1,...,N) into M bivariate beta5 components; Use the results of the EM algorithm to decide about the membership of the6 outlier score vectore ~Vi in each component; Select the bivariate beta component that contains vectors with the highest7 score values; Identify objects in D associated with the set of ~Vi that belong to the8 selected component and store them in OUT ; Return OUT ;9 end10 outlier score values ({ON(ONi )}(i=1,...,N) or {OCinv(O c i )}(i=1,...,N)). In this set-482 ting, the parameters of the univariate beta mixture model to be estimated are483 {λm,xm,ym}(m=1,...,M). These parameters and the optimal number of compo-484 nents in the mixture are estimated using the EM algorithm with the Newton-485 Raphson method and ICL-BIC as described in the above subsections. By doing486 so, we divide the outlier scores into several populations so that the larger scores487 can be identified and the associated objects are then declared as outliers.488 3. Experimental Evaluation489 In this section, we devise an empirical study to evaluate the suitability of the490 proposed approach. In the following, we first describe the technique that we have491 adopted to produce data for use in outlier detection and the performance metrics492 used in the evaluation. Next, we illustrate the effectiveness of our approach to493 identify outliers in mixed-attribute data. Finally, we devise further experiments494 to evaluate the performance of the proposed methodology in detecting outliers495 in single-type attribute data.496 24 3.1. Data Preparation and Metrics497 We draw the attention of the reader to the fact that, at the time of writing498 this paper, there is a shortage of standard benchmark data that can be used499 to evaluate outlier detection algorithms. Most of the publicly available labeled500 data are primarily designed for classification and machine learning applications.501 If the real data are unlabeled, then the evaluation of outlier detection accuracy502 must be done based on domain knowledge or with the help of a domain expert.503 However, this scenario is not practical for the purpose of evaluation since domain504 knowledge is not always available. All these factors make the evaluation of the505 proposed methodology a challenging task.506 In this paper, we saliently illustrate the performance of our approach in507 handling outliers using real data from the UCI Machine Learning Repository 1.508 Most of these data sets are labeled for classification purposes. Here, we have509 to be aware of the fact that these class labels are not the perfect ground truth510 in the sense that they do not correspond necessarily to potential outliers in511 the data. Keeping these issues in mind, we have adopted a principled way to512 produce real data for use in outlier detection.513 In our experiments, similar to the work in (Das and Schneider, 2007), we514 create simulated outlier objects by randomly selecting attribute values. Specif-515 ically, in the numerical attribute space, we first normalize the attribute values516 of each numerical attribute onto the interval [0, 1] and the then inject outlier517 points whose attribute values are randomly selected from [0, 1]. As a result of518 this process, all the points in our data sets have coordinates in the range [0, 1]519 and are either normal points or outliers. Note that the outliers are distributed520 at random throughout the entire space. On the other hand, to obtain outliers521 in the categorical space, we inject novel objects in the data set in such a way522 that, for each dimension t, the attribute value of the newly generated object is523 randomly selected from the whole set of distinct categorical values that form524 1http://archive.ics.uci.edu/ml/ 25 dimension t in the original data. Outliers in the mixed-attribute space are a525 random combination of the newly generated objects in both the numerical and526 the categorical spaces.527 For the purpose of evaluation, we used the following standard metrics: (1)528 Accuracy, which corresponds to the proportion of correctly partitioned objects,529 (2) True Positive Rate (TPR), measuring the proportion of outliers that are530 correctly identified as outliers, (3) False Positive Rate (FPR), corresponding to531 the proportion of inliers incorrectly classified as outliers, and (4) F-measure of532 the outliers class, corresponding to the harmonic mean between precision and533 recall of the outlier objects class.534 3.2. Experiments on Mixed-Attribute Data535 The goal of the experiments conducted in this section are to evaluate the536 suitability of the proposed approach in handling outliers in mixed-attribute data.537 We compare the performance of our approach to that of ODMAD (Koufakou538 and Georgiopoulos, 2010), the most recent approach for detecting outliers in the539 mixed-attribute space. Note that ODMAD requires a number of parameters to540 be set by the user. For fairness in comparison, several values were tried for the541 parameters of ODMAD, following the suggestions in its original paper, and we542 report results for the parameter settings that produced the best results. Note543 that the selection of the best result here refers to the best F-measure value,544 since this metric represents a good trade-off between TPR and FPR.545 We considered mixed-attribute data sets taken from the UCI Machine Learn-546 ing Repository. As mentioned in the previous subsection, to obtain data sets547 for use in outlier detection, we generated outlier objects by randomly flipping548 attribute values. We fixed the number of outliers injected in each set to 10%549 of the original data set size under investigation. Fig. 5 summarizes the main550 characteristics of the data sets used in our experiments. Note that some data551 sets (such as Credit Approval, Automobile and Cylinder Bands) originally con-552 tain a number of objects with missing attribute values. In our experiments, we553 simply ignore such objects.554 26 Figure 5: Mixed-attribute data sets characteristics. (a) Credit Approval (b) Heart (c) AutoUniv (au 6) Figure 6: Estimated density curve of the outlier score vectors that correspond to three mixed-attribute data sets. We used our approach to identify outliers in each of the mixed-attribute555 data sets considered in these experiments. To this end, we set M max to 5 and556 then, as discussed in Section 2, we selected the optimal number of components557 that minimize ICL-BIC. Here, the reader should be aware that the value of558 M max is not limited to 5 and the user can set any other value. Interestingly,559 we found that the estimated outlier score vectors in each of the ten data sets are560 well fitted by three bivariate beta components. For the purpose of illustration561 and in order to not encumber the paper, we show in Fig. 6 the estimated562 probability density function of the outlier score vectors, that corresponds to563 Credit Approval, Heart and AutoUniv (au 6) only. Data points associated with564 the bivariate beta component that contains the score vectors with the highest565 values correspond to outliers. Recall that the identification of the component566 containing the highest score values follows the procedure described in Section567 2.3.5.568 27 Figure 7: Performance results on mixed-attribute data sets. Fig. 7 compares the proposed method with ODMAD. Shaded regions in this569 figure correspond to the best values of the four evaluation metrics considered570 in the experiment. As can be seen from Fig. 7, our approach achieves the571 highest true positive rates and F-measure values across all the data sets under572 investigation and reports low false positive rates with high accuracy values. In573 fact, the proposed method achieves, on average, an accuracy of 96.53%, TPR574 and FPR of 92.45% and 3.01% respectively and finally an F-measure of 0.847, all575 pointing to fairly accurate results. On the other hand, the results provided by576 ODMAD are, on average, reasonable but less competitive than those achieved by577 our approach. As depicted by Fig. 7, ODMAD reports, on average, an accuracy578 of 94.75%, TPR and FPR of 72.22% and 2.93% respectively and finally an F-579 measure of 0.717. Overall, in term of Accuracy, TP rate and F-measure, the580 proposed method performs better than ODMAD while the FPR achieved by581 both approaches are comparable.582 From Fig 7, we observe that our proposed method reports an average 92.45%583 TP rate. This means that 7.55%, on average, of outliers were misclassified as584 inliers by our approach. This not necessarily an error, since data points have585 coordinates in the range [0, 1] and are either inliers or outliers. Outliers were586 randomly placed throughout the entire space. In this setting, it is probable that587 some of the outlier objects will have attribute values related to normal objects588 in the data set under investigation. Under these circumstances, it is possible589 that few outlier objects will have low outlier score values and consequently be590 28 considered as inliers.591 To summarize, the results presented in Fig. 7, suggest that the proposed592 method performs well on different data sets. Furthermore, in contrast to ODMAD593 which suffers from its dependency on several input parameters (detection thresh-594 old, minimum support, the maximum length of itemset and the size of a window595 of categorical and numerical scores), our approach is able to accurately iden-596 tify outliers in an automatic fashion. Such a notable feature of our approach597 illustrates its practical usability to effectively identify outliers in real-life appli-598 cations. Another advantage of our approach is that it is able to handle out-599 liers in single-type (numerical or categorical) attribute data without any feature600 transformation, while existing methods are not able to do so. The following601 two subsections investigate this point using real data sets characterized by only602 numerical or categorical attributes.603 3.3. Experiments on Numerical Data604 The experiments described in this section aim to illustrate the capability605 of the proposed methodology in detecting outlier objects in numerical data.606 As discussed at the end of Section 2, when the data contains only numerical607 attributes, we associate to each object the numerical score ON(Oni ) given by608 (1). Then, we model these scores as a mixture of univariate beta mixture.2609 The parameters of the model {λm,xm,ym}(m=1,...,M) and the optimal number610 of components in the mixture are estimated following the reasoning described611 in Section 2. This process results in grouping outlier scores into several beta612 components. Data objects associated with the beta component containing the613 highest score values are declared outliers.614 Fig. 8 summarizes the main characteristics of the UCI numerical data sets615 used in the experiments. Note that, as with the experiments on mixed-attribute616 data, we have adopted the same technique to produce outliers, that is, normal-617 izing the attribute values between 0 and 1 and then injecting outliers in the618 2To fit the beta distribution, the estimated outlier scores should be first normalized in [0,1]. 29 Figure 8: Numerical data sets characteristics. (a) Ecoli (b) Wine Quality - Red (c) Breast Cancer Figure 9: Estimated density curves of the numerical outlier scores that corre- spond to three numerical data sets. data by generating objects whose attribute values are randomly selected from619 the interval [0,1]. The number of outliers injected in each data set corresponds620 to 10% of the original data set size. For each numerical data set, we estimated621 ON(Oni ) for each object and then modelled these scores as a mixture of univari-622 ate beta distribution. To this end, we set M max to 5 and selected the optimal623 number of components that minimize ICL-BIC. We found that the number of624 components varies from two to three beta components. For the purpose of il-625 lustration, Fig. 9 shows the density curve of the numerical outlier scores that626 correspond to three UCI data sets: Ecoli, Wine Quality - Red and Wisconsin627 Diagnostic Breast Cancer. The last component in each plot depicted by Fig.628 9 represents the highest score values . Data points associated with the scores629 grouped in this component correspond to outliers.630 To demonstrate the effectiveness of our approach, we compared its perfor-631 mance to that of kNN weighed outlier algorithm (kNNW) (Angiulli and Pizzuti,632 2005, 2002). kNNW assigns a weight to each data point based on the sum of633 30 Figure 10: Performance results on numerical data sets. the distances separating that point from its k nearest neighbors in such a way634 that outliers are characterized by high weights while inliers receive low weight635 values. After ranking data points based on the estimated weights, the top n636 points are identified as outliers. The implementation of this algorithm, and637 many other outlier detection approaches, is available in the ELKI Data Min-638 ing Framework 3 (Achtert, Kriegel, Schubert, and Zimek, 2013). Note that we639 have chosen kNNW for its effectiveness. In fact, in our empirical investigation,640 we have evaluated several other mainstream outlier detection algorithms, such641 as COP (Kriegel, Kroger, Schubert, and Zimek, 2012), LDOF (Zhang, Hutter,642 and Jin, 2009), LOCI (Papadimitriou, Kitagawa, Gibbons, and Faloutsos, 2003)643 and LOF (Breunig et al., 2000), already implemented in ELKI. We found that644 kNNW was the algorithm which performs well.645 Fig. 10 illustrates the results of our approach and those of kNNW on the646 numerical data sets considered in the experiments. Shaded regions correspond647 to the best Accuracy, TPR, FPR and F-measure values. Recall that kNNW648 produces a ranked list of points expecting outliers to come first. Accordingly,649 to distinguish outliers from inliers, the user should specify the target number of650 outliers n. In this setting, and in order to compute the value of the four evalu-651 ation metrics used in the experiments (Accuracy, TPR, FPR and F-measure),652 we have simply set the value of n equal to the real number of outliers in the653 3http://elki.dbs.ifi.lmu.de 31 Figure 11: Categorical data sets characteristics. data set under investigation. Finally note that, as with ODMAD, we have tried654 multiple values of k for kNNW, and we only report the best results, that is,655 those which correspond to the highest F-measure value.656 As can be seen from Fig. 10, our approach achieves, on average, the highest657 Accuracy (97.60%), TPR (91.94%) and F-measure (0.884). On the other hand,658 kNNW reports the lowest average FPR (1.50%) while our approach achieves an659 average FPR of 1.88%. Overall, both competing algorithms show good perfor-660 mances. A significant advantage of our approach is that it is able to automati-661 cally discriminate outliers from inliers while with kNNW the user should specify662 how many points should be selected as outliers.663 3.4. Experiments on Categorical Data664 The aim of this section is to illustrate the suitability of the proposed ap-665 proach for handling outliers in data sets with categorical attributes only. To this666 end, we selected a number of categorical data from the UCI Machine Learning667 Repository. Recall that these data sets are principally labeled for classification668 purposes. Accordingly, as discussed in Section 3.1, to produce data for use in669 outlier detection, we inject novel data points in such a way that each attribute670 value of each newly inserted object is randomly selected from the set of distinct671 categorical values that initially form the corresponding attribute in the original672 data. As with our previous experiments, the number of outliers injected in each673 data set corresponds to 10% of the original data set size. The main characteris-674 tics of the categorical data sets used in the experiments are summarized in Fig.675 11.676 32 (a) Audiology (b) Vote (c) Lymphography Figure 12: Estimated density curves of the categorical outlier scores that corre- spond to three categorical data sets. To identify outliers in each of the categorical data sets considered in these677 experiments, we estimated first OCinv(Oci ) for each object. These scores are then678 normalized in [0,1] and modelled as a mixture of univariate beta distribution.679 To identify the optimal number of components in the mixture, we set M max to680 5 and selected the number of components that minimize ICL-BIC. Interestingly,681 as with the experiment on numeric data, we found that the optimal number of682 components varies from two to three. Fig. 12 illustrates the density curve of the683 outlier scores corresponding to three data sets: Audiology, Congressional Voting684 Records (Vote) and Lymphography. The last component in each plot depicted685 by Fig. 12 represents the highest score values . Data points associated with the686 scores grouped in this component correspond to outliers. The knowledgeable687 reader can also observe in this rendering, and also from the pervious illustration688 of the estimated density curves depicted by Fig. 9 and Fig. 6, the great shape689 flexibility of the beta distribution which leads to accurate partitioning of the690 outlier scores.691 Fig. 13 compares the effectiveness of our approach to that of a recent out-692 lier detection approach for categorical data named Information-Theory Based693 Single-Pass (ITB-SP) (Wu and Wang, 2013). It has been empirically illustrated694 that ITB-SP is an effective approach which outperforms several existing cat-695 egorical outlier detection algorithms. The implementation of this algorithm696 has been kindly provided by its authors. As the name implies, this approach697 harnesses information theory concepts to estimate an outlier score for each ob-698 33 Figure 13: Performance results on categorical data sets. ject. Specifically, the authors in Wu and Wang (2013) propose the concept of699 holoentropy as a new measure for outlier detection. As defined in Wu and Wang700 (2013), holoentropy is a combination between entropy and total correlation with701 attribute weighting, where the entropy measures the global disorder in the data702 and the total correlation measures the attributes relationship. Based on this703 concept, that is holoentropy, the authors formulate a function to estimate an704 outlier score for each object in such a way that outliers are characterized by705 high score values. The top n objects with the highest score values are declared706 as outliers. Note that, since ITB-SP requires the number of outliers in the data707 n to be specified by the user, and in order to compute Accuracy, TPR, FPR708 and F-measure, we have simply set the value of n equal to the real number of709 outliers in the data set under investigation.710 As can be seen from Fig. 13, the average performance results for our ap-711 proach and ITB-SP are quite similar except for the average TPR and FPR. Our712 method reports an average 94.84% of true positives while the average TPR of713 ITB-SP is 88.57%. This means that only 5.16%, on average, of outliers were714 misclassified as inliers by our approach while 11.43%, on average, of outliers were715 misclassified as inliers by ITB-SP. On the other hand, as illustrated by Fig. 13,716 we can see that ITB-SP achieves the lowest FPR, that is 3.40%, while the pro-717 posed method reports an average 4.48% of false positives. Overall, the results718 illustrated in Fig. 13 suggest that both approaches display good performance.719 Our approach has, however, the non-negligible advantage of automatically dis-720 criminating outliers from inliers while ITB-SP requires the number of outliers in721 34 the data to be specified by the user. As discussed earlier, in real applications for722 which no prior knowledge about the data is available, it is not always possible723 for the user to set accurately the value of this parameter.724 4. Conclusion725 In this paper, we have highlighted some limitations of existing outlier detec-726 tion approaches for mixed-attribute data, including their dependency on user727 parameters, such as the detection threshold and the target number of outliers to728 be identified, which are difficult to tune and their incapability of formally dis-729 criminating between outliers and inliers. To alleviate these problems, we have730 proposed a principled approach that performs outlier detection in an automatic731 fashion.732 In our approach, we first devised two functions in order to estimate, for each733 object, an outlier score in the numerical space and another score in the cate-734 gorical space. Outliers in both spaces are characterized by high score values.735 Next, we associate to each data point in the data set under investigation a two-736 dimensional vector such that the first element of this vector corresponds to the737 estimated outlier score in the numerical space, while the second element corre-738 sponds to the outlier score estimated in the categorical space. Then, we model739 these vectors as a mixture of bivariate beta. The bivariate beta component740 that corresponds to the highest score values represents outliers. The beta dis-741 tribution has been chosen due to its great shape flexibility which leads, in turn,742 to accurate fitting of the estimated outlier score vectors. We have described a743 statistical framework to illustrate how the bivariate beta mixture model can be744 used to identify outlier objects.745 Finally, we have devised a detailed empirical study to illustrate the suit-746 ability of our approach in detecting outliers using several UCI data sets with747 mixed-attributes. We have compared the performance of the proposed method748 to that of ODMAD, the most recent approach for detecting outliers in the mixed-749 attribute space. The results show that our approach achieves results that are,750 in most cases, better than those of ODMAD. Moreover, we have performed751 35 further experiments to demonstrate the capability of our methodology in han-752 dling outliers in single-type attribute data without any feature transformation.753 Tests and comparison with previous ranking approaches on several numerical754 and categorical UCI data sets show that the proposed methodology exhibits755 competitive results.756 Acknowledgements757 The author gratefully thanks Dr. Shu Wu for providing the implementation758 of the ITB-SP algorithm. The author also would like to thank the reviewers for759 their valuable comments and important suggestions. This work is supported by760 research grants from the Natural Sciences and Engineering Research Council of761 Canada (NSERC).762 References763 Achtert, E., Kriegel, H.-P., Schubert, E., & Zimek, A., 2013. Interactive data764 mining with 3D-parallel-coordinate-trees. In Proceedings of the ACM SIG-765 MOD international conference on Management of Data, 1009–1012.766 Aggarwal, C. C., 2013. Outlier Analysis. Springer.767 Alan, O., & Catal, C., 2011. Thresholds based outlier detection approach for768 mining class outliers: An empirical case study on software measurement769 datasets. Expert Systems with Applications, 38 (4), 3440–3445.770 Angiulli, F., & Fassetti, F., 2014. Exploiting domain knowledge to detect out-771 liers. Data Mining and Knowledge Discovery, 28 (2), 519–568.772 Angiulli, F., & Pizzuti, C., 2002. Fast outlier detection in high dimensional773 spaces. In Proceedings of the 6th European Conference on Principles of Data774 Mining and Knowledge Discovery, 15–26.775 Angiulli, F., & Pizzuti, C., 2005. Outlier mining in large high-dimensional data776 sets. IEEE Transactions on Knowledge and Data Engineering, 17 (2), 203–777 215.778 36 Bain, L. J., & Engelhardt, M., 2000. Introduction to Probability and Mathe-779 matical Statistics, 2nd Edition. Duxbury Press.780 Bezdek, J., 1981. Pattern Recognition with Fuzzy Objective Function Algo-781 rithms. Plenum.782 Bouguessa, M., 2012. Modeling outlier score distributions. In Proceedings of783 the 8th international conference on Advanced Data Mining and Applications,784 713–725.785 Bouguessa, M., & Wang, S., 2009. Mining projected clusters in high-dimensional786 spaces. IEEE Transactions on Knowledge and Data Engineering, 21 (4), 507–787 522.788 Bouguessa, M., Wang, S., & Sun, H., 2006. An objective approach to cluster789 validation. Pattern Recognition Letters, 27 (13), 1419–1430.790 Bouguila, N., & Elguebaly, T., 2012. A fully Bayesian model based on reversible791 jump MCMC and finite beta mixtures for clustering. Expert Systems with792 Applications, 39 (5), 5946–5959.793 Bouguila, N., Ziou, D., & Monga, E., 2006. Practical bayesian estimation of794 a finite beta mixture through gibbs sampling and its applications. Statistics795 and Computing, 16 (2), 215–225.796 Boutemedjet, S., Ziou, D., & Bouguila, N., 2011. Model-based subspace clus-797 tering of non-gaussian data. Neurocomputing, 73 (10-12), 1730–1739.798 Breunig, M. M., Kriegel, H.-P., Ng, R., & Sander, J., 2000. LOF: Identifying799 density-based local outliers. In Proceedings of the ACM SIGMOD interna-800 tional conference on Management of Data, 93–104.801 Cao, H., Si, G., Zhang, Y., & Jia, L., 2010. Enhancing effectiveness of density-802 based outlier mining scheme with density-similarity-neighbor-based outlier803 factor. Expert Systems with Applications, 37 (12), 8090–8101.804 37 Das, K., & Schneider, J., 2007. Detecting anomalous records in categorical805 datasets. In Proceedings of the 13th ACM SIGKDD international conference806 on Knowledge Discovery and Data Mining, 220–229.807 Dean, N., & Nugent, R., 2013. Clustering student skill set profiles in a unit808 hypercube using mixtures of multivariate betas. Advances in Data Analysis809 and Classification, 7 (3), 339–357.810 Dempster, A., Laird, N., & Rubin, D., 1977. Maximum likelihood from incom-811 plete data via the EM algorithm. Journal of Royal Statistical Society, (Series812 B), 39, 1–37.813 Figueiredo, M. A. T., & Jain, A. K., 2002. Unsupervised learning of finite mix-814 ture models. IEEE Transactions on Pattern Analysis and Machine Intelli-815 gence, 24 (3), 381–396.816 Fustes, D., Dafonte, C., Arcay, B., Manteiga, M., Smith, K., Vallenari, A., &817 Luri, X., 2013. SOM Ensemble for unsupervised outlier analysis. Application818 to outlier identification in the Gaia astronomical survey. Expert Systems with819 Applications, 40 (5), 1530–1541.820 He, Z., Xu, X., Huang, J., & Deng, S., 2005. FP-Outlier: Frequent pattern based821 outlier detection. Computer Science and Information System, 2 (1), 103–118.822 Huang, B., & Yang, P., 2011. Finding key knowledge attribute subspace of out-823 liers in high-dimensional dataset. Expert Systems with Applications, 38 (8),824 10147–10152.825 Ji, Y., Wu, C., Liu, P., Wang, J., & Coombes, K., 2005. Applications of beta-826 mixture models in bioinformatics. Bioinformatics, 21 (9), 2118–2122.827 Koufakou, A., & Georgiopoulos, M., 2010. A fast outlier detection strategy for828 distributed high-dimensional data sets with mixed attributes. Data Mining829 and Knowledge Discovery, 20 (2), 259–289.830 38 Koufakou, A., Secretan, J., & Georgiopoulos, M., 2011. Fast outlier detection in831 large high-dimensional categorical data. Knowledge and Information Systems,832 29 (3), 697–725.833 Kriegel, H.-P., Kroger, P., Schubert, E., & Zimek, A., 2011. Interpreting and834 unifying outlier scores. In Proceedings of the 11th SIAM international con-835 ference on Data Mining, 13–24.836 Kriegel, H. P., Kroger, P., Schubert, E., & Zimek, A., 2012. Outlier detection in837 arbitrarily oriented subspaces. In Proceedings of the 12th IEEE international838 conference on Data Mining, 379–388.839 Maervoet, J., Vens, C., Berghe, G. V., Blockeel, H., & Causmaecker, P. D., 2012.840 Outlier detection in relational data: A case study in geographical information841 systems. Expert Systems with Applications, 39 (5), 4718–4728.842 Otey, M. E., Ghoting, A., & Parthasarathy, S., 2006. Fast distributed outlier de-843 tection in mixed-attribute data sets. Data Mining and Knowledge Discovery,844 12 (2–3), 203–228.845 Papadimitriou, S., Kitagawa, H., Gibbons, P. B., & Faloutsos, C., 2003. LOCI:846 Fast outlier detection using the local correlation integral. In Proceedings of847 the 19th IEEE international conference on Data Engineering, 315–326.848 Penny, K., & Jolliffe, I., 2011. A comparison of multivariate outlier detection849 methods for clinical laboratory safety data. Journal of the Royal Statistical850 Society. Series D (The Statistician), 50 (3), 295–308.851 Tan, P.-N., Steinbach, M., & Kumar, V., 2006. Introduction to Data Mining.852 Addison Wesley.853 Wu, S., & Wang, S., 2013. Information-theoretic outlier detection for large-scale854 categorical data. IEEE Transactions on Knowledge and Data Engineering,855 25 (3), 589–602.856 39 Yamanishi, K., Takeuchi, J.-I., Williams, G., & Milne, P., 2000. On-line un-857 supervised outlier detection using finite mixtures with discounting learning858 algorithms. In Proceedings of the 6th ACM SIGKDD international confer-859 ence on Knowledge Discovery and Data Mining, 320–324.860 Ypma, T. J., 1995. Historical development of the Newton-Raphson method.861 SIAM Review, 37 (4), 531–551.862 Zhang, K., Hutter, M., & Jin, H., 2009. A new local distance-based outlier863 detection approach for scattered real-world data. In Proceedings of the 13th864 Pacific-Asia Conference on Advances in Knowledge Discovery and Data Min-865 ing, 813–822.866 Zhang, K., & Jin, H., 2010. An effective pattern based outlier detection ap-867 proach for mixed attribute data. In Proceedings of the 23rd Australasian Joint868 Conference on Advances in Artificial Intelligence, Lecture Notes in Artificial869 Intelligence, 6464, 122–131.870 40 Introduction Background Information Motivations and Contributions Proposed Approach Estimating Outlier Score in the Numerical Space Estimating Outlier Score in the Categorical Space Modeling Outlier Score Vectors The Bivariate Beta Mixture Model Maximum Likelihood Estimate EM Algorithm for the Bivariate Beta Mixture Estimating the Optimal Number of Components in the Mixture Automatic Identification of Outlier Experimental Evaluation Data Preparation and Metrics Experiments on Mixed-Attribute Data Experiments on Numerical Data Experiments on Categorical Data Conclusion 
work_hzl3hfqs4ndafeacepefr5vkiy	----	               City, University of London Institutional Repository Citation: Haibu, M., Margraf, C., Miranda, M. D. M. and Nielsen, J. P. (2016). Cash flow generalisations of non-life insurance expert systems estimating outstanding liabilities. Expert Systems with Applications, 45, pp. 400-409. doi: 10.1016/j.eswa.2015.09.021 This is the accepted version of the paper. This version of the publication may differ from the final published version. Permanent repository link: http://openaccess.city.ac.uk/12533/ Link to published version: http://dx.doi.org/10.1016/j.eswa.2015.09.021 Copyright and reuse: City Research Online aims to make research outputs of City, University of London available to a wider audience. Copyright and Moral Rights remain with the author(s) and/or copyright holders. URLs from City Research Online may be freely distributed and linked to. City Research Online: http://openaccess.city.ac.uk/ publications@city.ac.uk City Research Online http://openaccess.city.ac.uk/ mailto:publications@city.ac.uk Cash flow generalisations of non-life insurance expert systems estimating outstanding liabilities Munir Hiabu Munir.Hiabu.1@cass.city.ac.uk Carolin Margraf Carolin.Margraf.1@cass.city.ac.uk Maŕıa Dolores Mart́ınez Miranda mmiranda@ugr.es Jens Perch Nielsen Jens.Nielsen.1@city.ac.uk Cass Business School, City University London, U.K. September 17, 2015 Abstract For as long as anyone remembers non-life insurance companies have used the so called chain ladder method to reserve for outstanding liabilities. When historical payments of claims are used as observations then chain ladder can be understood as estimating a multiplicative model. In most non-life insur- ance companies a mixture of paid data and expert knowledge, incurred data, is used as observations instead of just payments. This paper considers re- cent statistical cash flow models for asset-liability hedging, capital allocation and other management decision tools, and develops two new such methods incorporating available incurred data expert knowledge into the outstanding liability cash flow model. These two new methods unbundle the incurred data to aggregates of estimates of the future cash flow. By a re-distribution to the right algorithm, the estimated future cash flow is incorporated in the overall estimation process and considered as data. A statistical validation technique is developed for these two new methods and they are compared to the other recent cash flow methods. The two methods show to have a very good per- formance on the real-life data set considered. Keywords: Stochastic Reserving; General Insurance; Chain Ladder; Claims In- flation; Incurred Data; Model Validation. 1 1 Introduction The non-life insurance business is an important part of the economy for most devel- oped countries with market revenues amounting to around five percent of GNP’s. The best estimate of outstanding liabilities - often called the reserve - is perhaps the single most important number on the balance sheet of most non-life insurance com- panies. Insufficient reserves is one common reason for non-life insurance companies to go broke. A mismanaged reserving process can also lead to spurious and volatile yearly results leading to uninformed management decisions. Finally, a smoother and more transparent reserving process leads to significant cost savings in almost any non-life insurance company. In this light it is perhaps surprising that statistical models for the most often used data set, the incurred run off triangle, are rarely considered in the literature. Incurred data is a mixture of historical payments of already settled claims and predicted severities of reported but not settled claims. The predicted severities are based on all available expert opinion in the company and are called case estimates. In that spirit, incurred data has added information about reported claims which are not available when only historical payments are considered. Actuaries often prefer working with chain ladder on the incurred triangle rather than on historical payments. But it is a bit unclear what this really means in terms of mathematical statistics. While incurred chain ladder probably makes good sense in a deterministic forecasting framework, the stochastic nature of the expert opinion in incurred data is not taken into account by this practice. There is a little literature acknowledging the added value of incurred data: the probably most famous Munich chain ladder approach by Quarg and Mack (2004), regression approaches by Halliwell (1997), Halliwell (2009), Venter (2008), and a paid-incurred chain reserving method by Posthuma et al. (2008), Merz and Wüthrich (2010), Happ et al. (2012) and Happ and Wüthrich (2013). These eight papers combine payment data and incurred data into one statistical model resulting in one reserve estimate. However, none of those papers take the common micro-structure of payment data and incurred data into account. Payment data and incurred data are both constructed via manipulations on the same underlying data; incurred data with the extra element of expert knowledge on expected future cash flows. To model this relationship one needs micro-structural assumptions about the underlying claim process. Pigeon et al. (2014) does exactly this. It is based on a recent trend to use so called granular data or micro data for reserving, see Antonio and Plat (2014) for one of the most interesting recent contributions in that area. See also Mart́ınez-Miranda et al. (2013a) for a continuous interpretation of the classical chain ladder methodology. While these approaches indeed seem to be favorable, 2 they are not well established yet and rely on data and granular information that is most often not available at hand. Double chain ladder, introduced in Mart́ınez-Miranda et al. (2012), builds on micro-structural assumptions but does not need granular data in the estimation procedure. It is based on the methods of Verrall et al. (2010) and Mart́ınez-Miranda et al. (2011) where the objective was to only rely on data that is already available in most reserving departments. It uses additional information of claim counts which is another triangle most often available in the data portfolio of a reserving department. The result is a full statistical model based on historical payments and counts capable of incorporating the information of the incurred data in a natural way. Agbeko et al. (2015) recently introduced a model reproducing the deterministic results of the earlier mentioned chain ladder method on incurred data by incorpo- rating the incurred data expert opinion into the well-defined full stochastic model of double chain ladder. One direct advantage of this approach is that the chain ladder model based on paid data can be validated against the chain ladder model based on incurred data. While the paid chain ladder and incurred chain ladder methods have been available for a long time as part of almost any non-life actuary’s tool kit, it has never before been possible to compare them against each other when only the typical aggregated data were available. Two other methods of incorporating expert knowledge of incurred data into these full cash flow models have been introduced in Mart́ınez-Miranda et al. (2013b) and Hiabu et al. (2016). The first of these two methods is extracting the inflation of the cost of a single claim from the incurred data and then incorporates that information in the double chain ladder model of Mart́ınez-Miranda et al. (2012). The second of these two methods suggests to incorporate a RBNS-preserving property. RBNS stands for Reported But Not yet Settled, and it can be estimated by the sum of all case estimates. This estimate is the best the claims department of an insurance company (with all the expert knowledge on the nature and severity of each claim available in such a department) is able to do. Hiabu et al. (2016) therefore produced a version of double chain ladder reproducing exactly the expert judgement of the RBNS reserves. All those mentioned double chain ladder extensions take advantage of the un- derlying structure of the incurred data, extract the relevant information from it and plug it into the original double chain ladder method. The advantage of this approach is that the simplicity and intuition of the simple chain ladder method is preserved and that the full statistical interpretation and stochastic cash flow formulation is inherited from double chain ladder. The two new stochastic cash flow methods developed in this paper build on the ideas and techniques of Hiabu et al. (2016). The first one treats the expert knowledge 3 of the incurred data as real data and incorporates it into the model, the second builds a second RBNS preserving cash flow model on top of this method. The approach is first to unbundle the incurred data to get back to the original aggregated RBNS numbers. These aggregated numbers are re-distributed according to the estimated delay such that the resulting algorithm takes both historical data and expert data into account in the final estimation. We therefore let the estimated future cash flow be incorporated in the overall estimation process by considering it as data. Note that both of these two new methods are cash flow models of the same nature as the models considered in Agbeko et al. (2015), and they can therefore be validated and compared to the models considered there. In the applied data example, this validation indicates, that the two new methods seem to take better advantage of the incurred expert data than previous methods did. Recent years have seen a growing interest in expert systems related to non- life insurance, see for example Belles-Sampera et al. (2014) and Abbasi and Guillén (2013), who consider ways of understanding risk in non-life insurance. Guelman and Guillén (2014) work with pricing of insurance claims and the customers sensitivity to that price; Guillen et al. (2012) and Kaishev et al. (2013) transfer knowledge from one business line to another to optimize cross-selling. Human judgement is important in all these insurance applications. When prices are set, there is a business intelligence department evaluating how much weight to put on the model at hand and how much weight to put on market prices as such. When risk is evaluated, human judgement calibrates the entering parameters. And when RBNS claims reserves are set, then there is an element of human judgement in the settlement of every single claim. It is also a human judgement when it is decided to use model based claims reserves for some subset of the claims, for example the smaller ones. We conclude this introduction by noting that Mart́ınez-Miranda et al. (2012) has two versions of double chain ladder; one version where the delay is not adjusted and another where it is adjusted. In this paper only the unadjusted version of double chain ladder is considered. One reason for the adjustment of the delay in Mart́ınez-Miranda et al. (2012) was to improve the performance of estimating the out-of-sample tail reserve. While this is a very important issue, it is beyond the scope of this paper to consider the out-of-sample tail reserve. The rest of the paper is structured as follows. Section 2 describes the data and the expert knowledge, introduces the notation and defines the model assump- tions. Section 3 discusses the outstanding loss liabilities point estimates. Section 4 describes four methods to estimate the parameters in the model: DCL, BDCL, PDCL, IDCL, EDCL and PEDCL. An application is considered in Section 5 and the validation of the six methods against each other is gone through in Section 6. Finally, Section 7 provides some concluding remarks. 4 2 Data and first moment assumptions This chapter introduces the data used in maybe every insurance reserving depart- ment to calculate their outstanding liabilities. Also the methods described in this paper rely on these data sets. They are often shortly called run-off triangles. These run-off triangles are the aggregated incurred counts (data), aggregated payments (data) and aggregated incurred payments (mixture of data and expert knowledge). All of those three objects have the same structural form, i.e., they live on the upper triangle I = {(i,j) : i = 1, . . . ,m,j = 0, . . . ,m− 1; i + j ≤ m}, m > 0, where m is the number of underwriting years observed. The parameter m also has another crucial role. If no tail-factors are considered, which will be assumed throughout this paper, then m−1 is the maximum delay, that is the time from the underwriting of the policy a claim is based on until its payment. This assumption is called being run-off, hence the name run-off triangles. Let’s first introduce the two data triangles. Aggregated incremental incurred counts: NI = {Nik : (i,k) ∈ I}, with Nik being the total number of claims of insurance incurred in year i which have been reported in year i + k, i.e. with k periods delay from year i. Aggregated incremental payments: XI = {Xij : (i,j) ∈I}, with Xij being the total payments from claims incurred in year i and paid with j periods delay from year i. A often confusing point is that the meaning of the second coordinate of the triangle I varies between the two different data. While in the counts triangle it represents the reporting delay, in the payments triangle it represents the development delay, that is reporting delay plus settlement delay. The definition of the incurred payments triangles is not that straight forward. To allow for a exact description, we first introduce micro structural variables of the claims process. We hereby follow the line of Mart́ınez-Miranda et al. (2012), since those variables will also play a key role in the underlying assumption of the double chain ladder model. By N paid ikl , we denote the number of the future payments originating from the Nik reported claims, which were finally paid with a delay of k+l, where l = 0, . . . ,m−1. Also, let X (h) ikl denote the individual settled payments which arise from N paid ikl , h = 1, . . . ,N paid ikl . Finally, we define Xikl = N paid ikl∑ h=0 X (h) ikl , (i,k) ∈I, l = 0, . . . ,m− 1, 5 i.e., those payments originating from underwriting year i, which are reported after a delay of k and paid with an overall delay of k + l. The aggregated incurred payments are then considered as unbiased estimators of∑m−1 l=0 Xikl. Technically, we model the expert knowledge as follows. Expert knowledge: Aggregated incurred payments: II = {Iik : (i,k) ∈I}, with Iik being Iik = k∑ s=0 m−1∑ l=0 E[Xisl| F(i+k)] − k−1∑ s=0 m−1∑ l=0 E[Xisl| F(i+k−1)], where Fh is an increasing filtration illustrating the expert knowledge at calen- dar period h. In this manuscript, we will only consider best estimates and therefore only need assumptions on the mean. We show that the classical CLM multiplicative structure holds under very general underlying dependencies on the mean. The first moment conditions of the DCL model are formulated below. For fixed i = 0, . . . ,m; k,l = 0, . . . ,m− 1, and h = 1, . . . ,Npaidikl , it holds that A1. The counts Nik are random variables with mean having a multiplicative parametriza- tion E[Nik] = αiβk, and identification ∑m−1 k=0 βk = 1. A2. The mean of the RBNS delay variables is E[N paid ikl |NI] = Nikπ̃l. A3. The mean of the individual payments size conditional on the number of pay- ments and the counts is given by E[X (h) ikl |N paid ikl ,NI] = µ̃lγi. Assumption A1 is the classical chain ladder assumption applied on the counts triangle, see also Mack (1991). The main point hereby is the multiplicativity be- tween underwriting year and reporting delay. Assumptions A2 and A3 are necessary to connect reporting delay, settlement delay and development delay - the main idea of DCL. See also Verrall et al. (2010), Mart́ınez-Miranda et al. (2011) and Mart́ınez- Miranda et al. (2012). Using A1 to A3, we have that E  N paid i,j−l∑ h=1 X (h) i,j−l,l|NI   = E  N paid i,j−l,l∑ h=1 E[X (h) i,j−l,l|NI,N paid i,j−l,l]|NI   = E[N paid i,j−l,lµ̃j−lγi|NI] = Ni,j−lπ̃lµ̃j−lγi. 6 Note that the observed aggregated payments can be written as Xij = j∑ l=0 Xi,j−l,l = j∑ l=0 N paid i,j−l,l∑ h=1 X (h) i,j−l,l. With the previous consideration, we derive E[Xij|NI] = γi k∑ l=0 Ni,j−lπ̃lµ̃j−l,l, and the unconditional mean is E[Xij] = αiγi j∑ l=0 βj−lµ̃j−l,lπ̃l. (1) Inspecting equation (1), we can reduce the amount of parameters by setting µ =∑m−1 l=0 π̃lµ̃l and πl = π̃lµ̃lµ −1, so that µπl = µ̃lπ̃l and therefore the unconditional mean of the payments becomes E[Xij] = αiγiµ j∑ l=0 βj−lπl. (2) Equation (2) is the key in deriving the outstanding loss liabilities. These are the values of (Xij) in the lower triangle. Consequently in the sequel, (α,β,π,γ,µ) = (α1, . . . ,αm,β0, . . . ,βm−1,π0, . . . ,πm−1,γ1, . . . ,γm,µ) are called the DCL parameters. In the next section, we will see that in a very natural way, we are able to distinguish between RBNS and IBNR claims. This is possible due to the separation of the development delay into the reporting delay β and the settlement delay π. 3 Forecast outstanding claims: the RBNS and IBNR reserves and predictive distributions In this section, we assume that the DCL parameters (α,β,π,γ,µ) are already derived and show how easily point forecasts of the RBNS and IBNR components of the reserve can be calculated. Note that when calculating the RBNS part, it is possible to replace the parameter (αi,βk) by the true value Nik, since the claims are already reported and thus Nik is observed. However, for the IBNR reserves, it is obviously necessary to use all DCL parameters, including the estimates of future numbers of 7 incurred claims αiβk. Using the notation of Verrall et al. (2010) and Mart́ınez-Miranda et al. (2011), we consider predictions over the triangle, J = J = {i = 2, . . . ,m; j = 0, . . . ,m − 1 with i + j ≥ m + 1}, illustrated in Figure 1. Figure 1: Index sets for aggregate claims data, assuming a maximum delay of m−1. We define the RBNS component as follows, where we consider two possibilities depending on whether the estimates of Nik are used or not. X̂ rbns(1) ij = j∑ l=i−m+j Ni,j−lπ̂lµ̂γ̂i, (i,j) ∈J . (3) and X̂ rbns(2) ij = j∑ l=i−m+j N̂i,j−lπ̂lµ̂γ̂i, (i,j) ∈J , (4) where N̂ij = α̂iβ̂j. In most cases, to shorten the notation, we will simply write X̂rbnsij for the RBNS estimates. However, whenever it is necessary, we will state which version is taken. The IBNR component always needs all DCL parameters: X̂ibnrij = i−m+j−1∑ l=0 N̂i,j−lπ̂lµ̂γ̂i, (i,j) ∈J . (5) The outstanding loss liabilities point estimates are then, X̂ij = X̂ rbns ij + X̂ ibnr ij (6) The outstanding liabilities per accident year are the row sums of IBNR and RBNS estimates. For a fixed i, we write J(i) = {j : (i,j) ∈ J}. Then the outstanding 8 liabilities per accident year i = 1, . . . ,m are R̂i = ∑ j∈J(i) X̂rbnsij + X̂ ibnr ij . (7) In the next section we describe several methods to derive the DCL parameters. 4 Estimation of the parameters in the Double Chain Ladder model To estimate the outstanding claims and thereby construct RBNS and IBNR reserves, we need to estimate the parameters involved in (2). In this section, we explore six different estimators. 4.1 The DCL method The DCL method is the original and maybe most simple method to derive the pa- rameters introduced in the previous section. It was introduced in Mart́ınez-Miranda et al. (2012). To estimate the DCL parameters in (2), assumptions on the pay- ments triangle XI are needed. DCL assumes the assumptions underlying the CLM method. B1 The payments Xij, with i = 1, . . . ,m, and j = 0, . . . ,m − 1, are random variables with mean having a multiplicative parametrization: E[Xij] = α̃iβ̃j, m−1∑ j=0 β̃j = 1. (8) Finally, merging (2) and (8), we conclude αiγiµ j∑ l=0 βj−lπl = α̃iβ̃k, and identify the parameters by αiµγi = α̃i, (9) j∑ l=0 βj−lπl = β̃j. (10) Again, many other micro-structure formulations might exist, thus the one spec- ified by (9) and (10) is only one of several possible. However, the above model can be considered as a detailed specification of the CLM. In Mart́ınez-Miranda et al. 9 (2013b) it is shown that if the RBNS component is calculated by (4), DCL com- pletely replicates the results of CLM. Now, the main idea to derive the DCL parameters is to estimate the chain ladder parameters (α̂, β̂) and (̂̃α, ̂̃β) (cf. A1, B1) by applying the classical chain ladder algorithm on the payments triangle XI and the counts triangle NI. Afterwards, the parameters left in (2) (this is (γ̂, µ̂, π̂)) can be calculated by simple algebra using (9) and (10). For illustration of the chain ladder algorithm, we assume an incremental triangle (Cij) (in our case this would be NI or XI), and that we want to estimate its chain ladder parameters (α̂, β̂). To apply the chain ladder algorithm, one has to transform the triangle (Cij) into a cumulative triangle (Dij): Dij = j∑ k=1 Cik. Then, the chain ladder algorithm can be applied on (Dij). It will produce estimates of development factors, λj, j = 1, 2, . . . ,m− 1, which can be described by λ̂j = ∑n−j+1 i=1 Dij∑n−j+1 i=1 Di,j−1 . These development factors can be converted into estimates of (α,β) using the fol- lowing identities which were derived in Verrall (1991). β̂0 = 1∏m−1 l=1 λ̂l β̂j = λ̂j − 1∏m−1 l=j λ̂l α̂i = m−i∑ j=0 Cij m−1∏ j=m−i+1 λ̂j Alternatively, analytical expressions for the estimators can also be derived directly (rather than using the chain ladder algorithm) and further details can be found in Kuang et al. (2009). Once the chain ladder parameters (α̂, β̂) and (̂̃α, ̂̃β) are derived, the settlement delay parameter π can be estimated just by solving the following linear system.   ̂̃ β0 ... ...̂̃ βm−1   =   β̂0 0 · · · 0 β̂1 β̂0 . . . 0 ... . . . . . . 0 β̂m−1 · · · β̂1 β̂0     π0 ... ... πm−1   . (11) 10 Let π̂ denote the solution of (11). Now we consider the estimation of the parameters involved in the means of indi- vidual payments. Of course, the model is technically over-parametrised since there are too many inflation parameters in (9). The simplest way to ensure identifiability is to set γ1 = 1, and then the estimate of µ, µ̂, can be obtained from µ̂ = ̂̃α1 α̂1 . (12) Using µ̂, the remaining estimates for γi, i = 2, . . . ,m, are directly derived from (9). The estimation procedure of double chain ladder is already programmed with the language R. We have used the R-package DCL (Mart́ınez-Miranda et al., 2013c) to derive Table 1, which shows the values of α̂, β̂, π̂ and γ̂ calculated from real data included in the DCL package. i k,l α̂(i) β̂(k) π̂(l) γ̂(i) 1 0 1078 0.7599 0.0592 1.0000 2 1 1890 0.2097 0.3098 1.1173 3 2 2066 0.0189 0.2032 1.4947 4 3 2353 0.0064 0.1996 1.7461 5 4 3015 0.0016 0.1388 2.1075 6 5 3727 0.0010 0.0440 2.0936 7 6 5057 0.0009 0.0227 2.2495 8 7 6483 0.0007 0.0095 2.1250 9 8 7727 0.0003 0.0018 1.9028 10 9 7134 0.0001 0.0029 2.0197 11 10 7319 0.0001 0.0002 2.0704 12 11 6152 0.0000 0.0026 2.2666 13 12 5242 0.0001 0.0019 2.3157 14 13 6150 0.0000 0.0032 2.4747 15 14 7028 0.0001 -0.0002 2.3829 16 15 6725 0.0000 0.0013 2.8391 17 16 5260 0.0000 -0.0004 3.1815 18 17 5869 0.0000 0.0000 4.1747 19 18 5953 0.0000 0.0000 6.7501 µ̂ = 2579 Table 1: DCL parameter estimates derived by the DCL method 11 4.2 The BDCL method The CLM and Bornhuetter-Ferguson (BF) methods are among the easiest claim reserving methods and, due to their simplicity, they are two of the most commonly used techniques in practice. Some recent papers on the BF method include Verrall (2004), Mack (2008), Schmidt and Zocher (2008), Alai et al. (2009) and Alai et al. (2010). The BF method introduced by Bornhuetter and Ferguson (1972) aims to ad- dress one of the well known weaknesses of CLM, which is the effect outliers can have on the estimates of outstanding claims. Especially the most recent underwriting years are the years with nearly no data and thus very sensitive to outliers. However, these recent underwriting years build the very major part of the outstanding claims. Hence, the CLM estimates of the outstanding liabilities might differ fatally from the true (unknown) values. Acknowledging this problem, the BF method incorporates prior knowledge from ex- perts and is therefore more robust than the CLM method, which relies completely on the data contained in the run-off triangle XI. In this section, we briefly summarize the Bornhutter-Ferguson double chain lad- der (BDCL) method introduced in Mart́ınez-Miranda et al. (2013b), which mimics BF in the framework of DCL. The BDCL method starts with identical steps as DCL but instead of using the estimate of the inflation parameters, γ and µ, from the tri- angle of paid claims XI, it deploys expert knowledge in the form of the incurred triangle II to adjust the estimation of the sensitive inflation parameter γ. This is done as follows. From assumptions A2,A3 and equation (9), we easily deduce that E[Iik] = αiµγiβk = α̃iβk. (13) Hence, the incurred triangle II has multiplicative mean and its underwriting year factor, α̃, is identical to the one of the payments triangle XI (cf. (8)). However, its estimation is less sensitive to outliers since it incorporates all incurred claims via expert knowledge. We conclude that we can replace the payments triangle by the incurred payments triangle when we calculate estimates of the inflation parameters, γ,µ, in (9). Note that the severity mean µ is going to remain the same since the first rows of XI and II are identical. Summarized, the BDCL-method can be carried out as follows. • Step 1: Parameter estimation. Estimate the DCL parameters (α,β,π,γ,µ) using the DCL method of Section 4.1 with the data in the triangles NI and XI and denote the parameter esti- mates by (α̂, β̂, π̂, γ̂, µ̂). 12 Repeat this estimation using the DCL method but replacing the triangle of paid claims XI by the triangle of incurred data II. Keep only the resulting estimated inflation parameters, denoted by γ̂BDCL. • Step 2: BF adjustment. Replace the inflation parameters γ̂ from the paid data by the estimate from the incurred triangle, γ̂BDCL. From Step 1 and Step 2, the final BDCL estimates of the DCL parameters are (α̂, β̂, π̂, γ̂BDCL, µ̂). ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 2 4 6 Underwriting period S e ve ri ty in fla tio n 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 ● DCL BDCL Figure 2: Plot of severity inflation estimates. DCL: γ̂i (green), BDCL: γ̂ BDCL i (blue). Using the R-package DCL (Mart́ınez-Miranda et al., 2013c), we derive Figure 2 which shows the severity inflation estimates derived by DCL and BDCL. BDCL, with the incorporated expert knowledge, seems to stabilize the severity inflation in the most recent underwriting years while keeping the values in the other years. The result is a more realistic estimate correcting the DCL parameter γ̂i exactly in its weakest point, that is in those years where the payments triangle XI has nearly no data. Again, those recent underwriting years contain the very major part of the outstanding liabilities. 4.3 The IDCL method This section gives a brief theoretical introduction to the IDCL method of Agbeko et al. (2015). In the BDCL definition, we incorporated an additional triangle of in- curred claims in order to produce a more stable estimate of the underwriting inflation parameter γi. The derived BDCL method becomes a variant of the Bornhuetter- Ferguson technique using prior knowledge contained in the incurred triangle. One natural question is whether the derived reserve estimate is the classical incurred 13 chain ladder. The answer is that this not the case; the IDCL method does not replicate the results obtained by applying the classical chain ladder method to the incurred triangle. Among practitioners, the incurred reserve seems to be more real- istic for many datasets compared to the classical paid chain ladder reserve. IDCL mimics the reserve estimate of chain ladder on the incurred triangles in the DCL framework. It is defined just by rescaling the underwriting inflation parameter es- timate from the DCL method. Specifically, we define a new scaled inflation factor estimate γ̂IDCL such that γ̂IDCLi = R̂∗i R̂i γ̂i, (14) where R∗i is the outstanding loss liabilities per accident year as predicted by applying the traditional CLM on incurred data, and (Ri,γ̂i) are the outstanding loss liabilities per accident year and the inflation parameter respectively, using the DCL method (cf. Section 4.1). The final IDCL estimates of the DCL parameters are then (α̂, β̂, π̂, γ̂IDCL, µ̂). With the new inflation parameter estimate γ̂IDCL, the outstanding liabilities derived by the DCL parameters completely replicate the CLM estimates on the incurred trian- gle. 4.4 The PDCL method This section gives a brief theoretical introduction to the PDCL method introduced in Hiabu et al. (2016). See also Nielsen (2016). In the last section, we have described a method which incorporates expert knowledge in form of the incurred triangle II. The values in II arise from case estimates for RBNS claims, developed in the case department of the insurance company, and claims which are already paid. Thus, if one subtracts these already paid claims (which are given via the payments trian- gle XI) from the incurred triangle, one can reconstruct the RBNS case estimates. However, as soon as this is done, it is obvious that these RBNS case estimates do not match with the RBNS estimates (3) and (4), using any DCL method (including BDCL). We conclude that the reserve department, using double chain ladder (and also chain ladder), calculates different RBNS estimates than those given by the case department. PDCL replaces the calculated RBNS estimates with the case estimates. The method, however, does not only preserve the RBNS case estimates by adding the IBNR esti- mates to conclude the reserve, it also takes the RBNS case estimates to correct the other DCL parameters. Therefore, also the total IBNR size will change. The first step is to construct a preliminary square (Sij), i = 1, . . . ,m, j = 14 0, . . . ,m− 1, which yields new estimators for the DCL parameters. The upper tri- angle of the square (i.e., (i,j) ∈ I) should have the same entries as the payments triangle (Xij). The lower triangle (i.e., (i,j) ∈J) should consist of preliminary esti- mates of the outstanding loss liabilities. The outstanding loss liabilites comprise an RBNS and an IBNR part (cf. (6)). However, only the IBNR part of these outstand- ing loss liabilities is estimated. For the RBNS component, the RBNS case estimates are taken. More precisely, one takes the DCL parameter estimates (α̂, β̂, π̂, γ̂, µ̂) and use these parameters to estimate the RBNS component (X̂rbnsij ) and IBNR compo- nent (X̂ibnrij ) using (4) and (5). As mentioned above, the RBNS estimate should be be equal to the RBNS case estimates, which can only be reconstructed per accident year. For i = 1, . . . ,m, they can be described as Xrbns.case.estimatei = m−i∑ j=0 Iij − m−i∑ j=0 Xij. (15) Hence, we define the RBNS preserving components X̂ rbns.pres ij = ∑m−i j=0 Iij − ∑m−i j=0 Xij∑ j∈J(i) X̂ rbns ij X̂rbnsij . (16) Note that ∑ j∈J(i) X̂ rbns.pres ij = X rbns.case.estimate i . Thus we define the preliminary square (Sij) as Sij =  Xij, if (i,j) ∈I,X̂rbns.presij + X̂ibnrij , if (i,j) ∈J . (17) One can easily see that payments square (Sij) has approximately multiplicative mean E[Sij] ≈ α̃iβ̃j. Therefore, we can use (Sij) to completely replace XI to estimate the DCL parameters (cf. (8)). Note that in the BDCL method we were only able to balance the estimator of the inflation parameter α̃i (cf. (13)). Again, while in the BDCL method, one uses the expert knowledge to only adjust the inflation parameters, here, one can take full advantage of the triangle II and also correct the delay parameters. Since (Sij) has a multiplicative structure, the CLM idea is used to estimate α̃i and β̃j. We define ̂̃αPDCLi = m−1∑ j=0 Sij, ̂̃ β PDCL j = ∑m i=1 Sij∑ (i,j)∈I∪J Sij . (18) 15 Exactly as in the previous sections, one can now apply (9) and (10) to derive the PDCL parameters (α̂i, β̂j, π̂ PDCL, γ̂PDCL ∗ , µ̂PDCL). Since this approach is still not RBNS preserving, γ̂PDCL ∗ is balanced by defining a new scaled inflation factor esti- mate γ̂PDCL such that γ̂PDCL = ∑m−i j=0 Iij − ∑m−i j=0 Xij X̂rbnsij , (19) where X̂rbnsij is calculated with the parameters (α̂, β̂, π̂ PDCL, γ̂PDCL ∗ , µ̂PDCL) using (4). 4.5 The EDCL method In this chapter we introduce a new method called expert double chain ladder (EDCL), indicating that expert knowledge in form of incurred data and RBNS’s are incorpo- rated into the system as pseudo data. The idea of the EDCL method is to replicate the basic steps of the previously introduced PDCL method (15)-(18), but without the adjustment of the severity inflation in (19). Instead those steps are iterated until convergence. The iteration forces a homogeneous solution which incorporates the incurred triangle, I, and the payment triangle, X, to one reserve. The discrep- ancy between estimated RBNS and RBNS provided by the case estimates can be explained by variation of the observations around their mean. Note that given the model assumptions, we are estimating the mean of the RBNS. The EDCL estimation can be described as follows. In the first step of the itera- tion, we start with the DCL parameter estimates (α̂, β̂, π̂, γ̂, µ̂), which we denote by (α̂, β̂, π̂EDCL,(0), γ̂EDCL,(0), µ̂EDCL,(0)). In the k−th step of the iteration, we take the EDCL parameter estimates we obtained in the (k−1)−th step (α̂, β̂, π̂EDCL,(k−1), γ̂EDCL,(k−1), µ̂EDCL,(k−1)) and use these parameters to estimate the RBNS component (X̂ rbns,(k−1) ij ), and IBNR com- ponent (X̂ ibnr,(k−1) ij ), using (4) and (5). Then, we calculate the k−th RBNS preserving components X̂ rbns.pres,(k) ij = ∑m−i j=0 Iij − ∑m−i j=0 Xij∑ j∈J(i) X̂ rbns,(k−1) ij X̂ rbns,(k−1) ij . Now, we are able to define the k−th preliminary square (S(k)ij ) as S (k) ij =  Xij, if (i,j) ∈I,X̂rbns.pres,(k)ij + X̂ibnr,(k−1)ij , if (i,j) ∈J . 16 Since (S (k) ij ) has an approximately multiplicative structure, we use the CLM idea to estimate α̃i and β̃j. We define ̂̃αEDCL,(k)i = m−1∑ j=0 S (k) ij , ̂̃ β EDCL,(k) j = ∑m i=1 S (k) ij∑ (i,j)∈I∪J S (k) ij . Exactly as in the previous sections, we can now apply (9) and (10) to derive the EDCL parameters (α̂, β̂, π̂EDCL,(k), γ̂EDCL,(k), µ̂EDCL,(k)). After iterating until convergence, we derive the final EDCL parameters denoted by (α̂, β̂, π̂EDCL, γ̂EDCL, µ̂EDCL). 4.6 The PEDCL method In this section, we introduce the RBNS-preserving expert double chain ladder (PEDCL). As the name suggests, this method builds on the PDCL method by using the EDCL parameters (α̂, β̂, π̂EDCL, γ̂EDCL, µ̂EDCL). As mentioned in the previous section, EDCL does not preserve the RBNS case estimates. The reason was mentioned in the beginning of the previous section and has a parallelism to the two different RBNS estimates in (3) and (4). If as in PDCL, one decides not to change the case estimates, one can thus replace the RBNS part by the RBNS derived from the case estimates. As this information is only available per underwriting year, it can be incorporated by changing the infla- tion parameter accordingly. More precisely, we define a new scaled inflation factor estimate as γ̂PEDCL = ∑m−i j=0 Iij − ∑m−i j=0 Xij X̂rbnsij , where X̂rbnsij is calculated with the parameters (α̂, β̂, π̂ EDCL, γ̂EDCL, µ̂EDCL) using (4). This new inflation parameter γ̂PEDCL is used to replace γ̂EDCL in the parameter set (α̂, β̂, π̂EDCL, γ̂EDCL, µ̂EDCL) when calculating only the RBNS part which pre- serves the case estimates. Therefore, in contrast to the previous methods, PEDCL possesses two inflation parameters. Firstly, γ̂PEDCL for the RBNS part, that is when calculating (4) and secondly γ̂EDCL for the IBNR part, that is when calculating (5). Note that since we are just using the EDCL parameters to calculate the IBNR part, the IBNR reserve of the PEDCL method is exactly the same as the IBNR reserve of the EDCL method. 5 Real Data Application In this section, we apply the methods to a data set obtained from a UK motor business. The data is also available via the DCL R package (Mart́ınez-Miranda 17 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 2 4 6 Underwriting period S e ve ri ty in fla tio n 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● DCL BDCL IDCL PDCL EDCL PEDCL Figure 3: Plot of severity inflation estimates. et al., 2013c). Figure 3 shows a plot of the six severity inflation parameters derived by DCL, BDCL, IDCL, PDCL, EDCL and PEDCL. There were some rather rough corrections going on for IDCL,PDCL and PEDCL in the first five years. These five years represent less than 0.1% of the total loss liabilities estimates (cf. Table 2-4) and they are not really important. We have therefore taken the first five years out for these three estimation methods. Otherwise, these unimportant first five years would have dominated the graph and perhaps have confused the reader. Figure 3 shows that IDCL,PDCL and PEDCL still are quite volatile with a tendency to have lower severity inflations than DCL, BDCL and EDCL. This is because IDCL,PDCL and PEDCL adjust to high or low RBNS’s in certain years, while in particular BDCL and EDCL take a more balanced point of view. From a modeling perspective the sever- ity inflations resulting from BDCL and EDCL seem more attractive, because they are more stable and therefore seem more realistic. This graphs illustrate very well why the PEDCL method uses the EDCL parameters while estimating the IBNR re- serve. These parameters seem more realistic than the RBNS-preserving parameters. However, the RBNS-preserving parameters might be more realistic when estimating the somewhat realised RBNS reserve. If one believes the RBNS estimates are of good quality or are of the best possible quality one can do with the data, then one should of course use a method preserving these RBNS estimates such as PDCL and PEDCL. However, for the IBNR’s it is another matter. There is no expert knowl- edge available for the IBNR estimates and therefore one should use the methodology with the most credible parameters. We believe that EDCL is the method with the most credible parameters and PEDCL is therefore - in our opinion - the optimal methodology if one wishes to preserve the RBNS estimates. If one is looking for a more balanced view, where the RBNS’s are allowed to impact the parameters, but 18 where the observed data also should play a role in some sort of validation of the RBNS as data, then one should use the EDCL method. We believe that the DCL method, the EDCL method and the PEDCL method are the three best methods to consider for practising actuaries. To visualise these arguments, Table 2 shows the outstanding loss liabilities per underwriting year for all six methods mentioned in chapter 4 as well as the classical chain ladder method. Furthermore, the splitted values of RBNS and IBNR per underwriting year are illustrated in Table 3 and Table 4. i CLM DCL BDCL IDCL PDCL EDCL PEDCL 1 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 2 0.0000 0.0005 0.0003 0.0005 0.0000 0.0000 0.0000 3 0.0000 0.0001 -0.0003 0.0001 0.0040 0.0001 0.0040 4 0.0000 0.0007 -0.0014 0.0007 -0.0095 0.0003 -0.0095 5 0.0173 0.0039 0.0151 0.0045 0.0365 0.0076 0.0365 6 0.0346 0.0309 0.0313 0.0122 0.0054 0.0135 0.0068 7 0.1381 0.1407 0.1408 0.0090 0.0014 0.0541 0.0040 8 0.2449 0.2485 0.2483 0.0927 0.0949 0.1102 0.0972 9 0.3522 0.3582 0.3563 0.0536 0.0591 0.1774 0.0643 10 0.3943 0.3818 0.3800 0.1507 0.1981 0.2148 0.2024 11 0.5524 0.5246 0.5206 0.1664 0.2515 0.3415 0.2595 12 0.6839 0.6309 0.6169 -0.1458 0.0015 0.3415 0.0300 13 1.0504 0.9764 0.9733 0.8648 1.2718 0.8896 1.2576 14 2.5361 2.5483 2.5164 2.0388 2.8302 2.3205 2.8153 15 5.7370 5.4483 5.2846 4.2095 6.2887 5.4225 6.2721 16 14.0889 15.4373 12.9824 6.8542 9.4207 11.9785 9.4735 17 21.0057 21.7407 17.0455 12.0924 13.4243 16.5608 13.5068 18 44.6877 44.4580 29.2840 23.0002 25.9634 30.6858 26.1851 19 98.9723 98.9722 41.8444 39.1522 42.1009 45.0575 42.8640 SUM 190.4957 191.9021 112.2385 88.5565 101.9427 114.3021 103.0696 Table 2: Outstanding loss liabilities per underwriting year in million 19 i CLM DCL BDCL IDCL PDCL EDCL PEDCL 1 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 2 0.0000 0.0005 0.0003 0.0005 0.0000 0.0000 0.0000 3 0.0000 0.0001 -0.0003 0.0001 0.0040 0.0001 0.0040 4 0.0000 0.0007 -0.0014 0.0007 -0.0095 0.0003 -0.0095 5 0.0173 0.0039 0.0151 0.0045 0.0365 0.0076 0.0365 6 0.0329 0.0291 0.0295 0.0115 0.0050 0.0118 0.0050 7 0.1355 0.1382 0.1382 0.0088 0.0014 0.0516 0.0014 8 0.2399 0.2436 0.2434 0.0909 0.0923 0.1053 0.0923 9 0.3455 0.3515 0.3496 0.0526 0.0576 0.1707 0.0576 10 0.3822 0.3697 0.3680 0.1459 0.1903 0.2027 0.1903 11 0.5339 0.5062 0.5023 0.1606 0.2411 0.3231 0.2411 12 0.6550 0.6020 0.5887 -0.1392 0.0014 0.4387 0.0014 13 1.0034 0.9294 0.9265 0.8231 1.2101 0.8420 1.2101 14 2.4415 2.4537 2.4229 1.9631 2.7197 2.2249 2.7197 15 5.5906 5.3020 5.1427 4.0964 6.1235 5.2739 6.1235 16 13.8419 15.1902 12.7746 6.7444 9.2492 11.7542 9.2492 17 20.5131 21.2482 16.6593 11.8184 13.0995 16.1534 13.0995 18 42.7693 42.5397 28.0204 22.0078 24.8281 29.3288 24.8281 19 74.0936 74.0942 31.3260 29.3108 31.4544 33.6479 31.4544 SUM 162.5956 164.0027 99.5059 77.1009 89.3045 100.5369 89.3045 Table 3: RBNS per underwriting year in million 20 i CLM DCL BDCL IDCL PDCL EDCL PEDCL 1-5 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 6 0.0018 0.0018 0.0018 0.0007 0.0004 0.0018 0.0018 7 0.0026 0.0026 0.0026 0.0002 0.0000 0.0026 0.0026 8 0.0049 0.0049 0.0049 0.0018 0.0026 0.0049 0.0049 9 0.0067 0.0067 0.0067 0.0010 0.0015 0.0067 0.0067 10 0.0121 0.0121 0.0120 0.0048 0.0078 0.0121 0.0121 11 0.0184 0.0184 0.0183 0.0058 0.0103 0.0184 0.0184 12 0.0289 0.0289 0.0282 -0.0067 0.0001 0.0285 0.0285 13 0.0471 0.0471 0.0469 0.0417 0.0617 0.0476 0.0476 14 0.0946 0.0946 0.0934 0.0757 0.1105 0.0957 0.0957 15 0.1464 0.1464 0.1419 0.1131 0.1652 0.1486 0.1486 16 0.2470 0.2471 0.2077 0.1097 0.1715 0.2243 0.2243 17 0.4926 0.4925 0.3862 0.2740 0.3248 0.4073 0.4073 18 1.9183 1.9183 1.2636 0.9924 1.1353 1.3570 1.3570 19 24.8787 24.8779 10.5185 9.8414 10.6465 11.4096 11.4096 SUM 27.9001 27.8993 99.5059 11.4556 12.6382 13.7651 13.7651 Table 4: IBNR per underwriting year in million 21 6 Model validation This section describes the validation process for the six methods DCL, BDCL, IDCL, PDCL, EDCL and IPDCL discussed in Section 4. The following validation was introduced in Agbeko et al. (2015). To our knowledge this is the only developed method which is able to validate reserving estimates based on incurred data with estimates based on paid data. That is to validate DCL against IDCL in our terminology. However, this validation methodology is sufficiently general to allow all these six procedures to be validated against each other. The validation process builds on back testing and is based on the fact that all introduced DCL methods provide reserve estimates by predicting into the same paid triangle, X. Therefore, all methods can be compared on the same scale so to speak. Below, we have omitted the most recent calendar year and the four most recent calendar years, respectively (in all three available triangles). Therefore, since our dataset consists of m = 19 years, there are 18 and 60 cells, respectively, to be compared with the true values. ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● CLM DCL BDCL IDCL PDCL EDCL PEDCL− 1 0 0 0 0 0 0 0 5 0 0 0 0 0 1 5 0 0 0 0 0 1 periods cut P re d ic tio n m in u s o b se rv e d ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● CLM DCL BDCL IDCL PDCL EDCL PEDCL− 1 0 0 0 0 0 0 0 5 0 0 0 0 0 1 5 0 0 0 0 0 4 periods cut P re d ic tio n m in u s o b se rv e d Figure 4: Box plot of the cell errors Figure 4 shows two box plots of the respectively 18 and 60 errors calculated by taking the difference between estimated and true values. One conclusion is that DCL and CLM seem to be inferior to the five methods that take advantage of expert knowledge. Between these five methods, it seems that BDCL and EDCL have similar performance with a slight advantage to the new EDCL method. Also PDCL and PEDCL have similar performance, which was to be expected because they only differ in the IBNR reserves while having identical RBNS reserves. The omnipresent IDCL method does not provide convincing performance in this data example. Earlier studies based on a number of data sets have shown that IDCL sometimes have very good performance, but equally often fails badly. Of the three methods available at the time, DCL, BDCL and IDCL, only the BDCL method had 22 a stable performance. Sometimes DCL was winning convincingly, other times IDCL was the winner. In the long run, however, the eternal runner-up was the BDCL method, and most of the time it did almost as well - but not quite - as the best of the two other methods. We do, however, find EDCL a more convincing method than BDCL. While they have some similarities and while both have stable underwriting year severities as we saw in the application, EDCL seems more theoretically correct in its way of exploring the expert knowledge and we believe it will be replacing BDCL in the long run. Based on long-term considerations, we think DCL, EDCL and PEDCL should be sufficient in the practical actuaries tool box. BDCL, IDCL and PDCL are - in our view - less convincing methods. CLM DCL BDCL IDCL PDCL EDCL PEDCL 1 periods cut S u m o f a b so lu te e rr o rs 0 e + 0 0 1 e + 0 7 2 e + 0 7 3 e + 0 7 4 e + 0 7 CLM DCL BDCL IDCL PDCL EDCL PEDCL 4 periods cut S u m o f a b so lu te e rr o rs 0 e + 0 0 1 e + 0 7 2 e + 0 7 3 e + 0 7 4 e + 0 7 Figure 5: Bar plot for the sum of absolute cell errors CLM DCL BDCL IDCL PDCL EDCL PEDCL 1 periods cut R e la tiv e e rr o r 0 .0 0 .1 0 .2 0 .3 0 .4 0 .5 CLM DCL BDCL IDCL PDCL EDCL PEDCL 4 periods cut R e la tiv e e rr o r 0 .0 0 .1 0 .2 0 .3 0 .4 0 .5 Figure 6: Bar plot for the relative errors 23 In the top panels of Figure 5, we have plotted the sum of the absolute cell errors (`1 error). That is Sum of absolute cell errors = ∑ (i,j)∈B |X̂ij −Xij|, B = {(i,j)| i = 2, . . . ,m− c; j = 0, . . . ,m− c− 1; i + j = m− c + 1, . . . ,m}, where c is the number of recent calendar years omitted for back testing (here: 1 and 4). The relative errors, that is Sum of absolute cell errors Sum of absolute true values = ∑ (i,j)∈B |X̂ij −Xij|∑ (i,j)∈B |Xij| , is shown in the bottom panels of Figure 5. The conclusion Figure 5 is similar to the conclusion of Figure 4. 7 Conclusions This paper has developed two new methods combining classical chain ladder method- ology with expert knowledge via the double chain ladder methodology. The new EDCL introduces RBNS’s as pseudo data and uses an iterative procedure to im- prove the originally estimated DCL parameters that did not take RBNS information into account. It shows to have very good performance. The new PEDCL method preserves RBNS estimates also after the reserve has been estimated. Validation is introduced for these two new methods and they are compared to the previous meth- ods DCL, BDCL, IDCL and PDCL. Our conclusion is that while DCL always will be some kind of benchmark in the actuaries tool box, then EDCL and PEDCL seem to have sufficient quality to replace BDCL, IDCL and PDCL. That also means that EDCL and PEDCL seem to have sufficient quality to replace incurred chain ladder as the actuaries preferred method of incorporating incurred data expert knowledge. This is important, because most reserves in the actuarial practise use incurred chain ladder as the basis of estimation. This incurred chain ladder might be manually manipulated according to expert knowledge and the values of paid chain ladder. However, it is the most common basis of estimation in actuarial practise. Only the future will be able to show whether EDCL and PEDCL or other similar innovations will be able to replace actuaries habit of using the - in our view outdated - incurred chain ladder approach. 24 References Abbasi, B. and Guillén, M. (2013). Bootstrap control charts in monitoring value at risk in insurance. Expert Systems with Applications, 40(15), 6125–6135. Agbeko, T., Hiabu, M., Mart́ınez-Miranda, M. D., Nielsen, J. P., Verrall, R. (2015). Validating the Double Chain Ladder Stochastic Claims Reserving Model. Vari- ance 8(2), 1–23. Alai, D., Merz, M. and Wüthrich, M. V. (2009). Mean square error of prediction in the Bornhuetter-Ferguson claims reserving method. Ann. Actuar. Sci. 4(1), 7–31. Alai, D., Merz, M. and Wüthrich, M. V. (2010). Prediction uncertainty in the Bornhuetter-Ferguson claims reserving method: revisited. Ann. Actuar. Sci. 5(1), 7–17. Antonio, K. and Plat, R. (2014). Micro-level stochastic loss reserving in general insurance. Scandinavian Actuarial Journal 2014(7), 649–669. Belles-Sampera, J., Guillén, M. and Santolino, M. (2014). Beyond value-at-risk in finance and insurance: GlueVaR distortion risk measures. Risk Analysis 34(1), 121–134. Bornhuetter, R. L. and Ferguson, R. E. (1972). The actuary and IBNR. Casualty Actuarial Society Proceedings, 181–195. Guelman, L. and Guillén, M. (2014). A causal inference approach to measure price elasticity in Automobile Insurance. Expert Systems with Applications 41(2), 387– 396. Guillén, M., Nielsen, J.P., Scheike, T.H. and Pérez-Maŕın, A.M. (2012). Time- varying effects in the analysis of customer loyalty: A case study in insurance. Expert Systems with Applications 39(3), 3351–3358. Halliwell, L.J. (1997). Cojoint prediction of paid and incurred losses. CAS Forum Summer. 241–379. Halliwell, L.J. (2009). Modeling paid and incurred losses together. CAS E-Forum Spring. Happ, S., Merz, M. and Wüthrich, M.V. (2012). Claims development result in the paid-incurred chain reserving method. Insurance: Mathematics and Economics, 51(1), 66–72. 25 Happ, S. and Wüthrich, M.V. (2013). Paid-incurred chain reserving method with dependence modeling. ASTIN Bulletin, 43(1), 1–20. Hiabu, M., Margraf, C., Mart́ınez-Miranda, M. D., Nielsen, J. P. (2016). The link between classical reserving and granular reserving through double chain ladder and its extensions. British Actuarial Journal. 21, 18–33. Kaishev, V.K., Nielsen, J.P., and Thuring, F. (2013). Optimal customer selection for cross-selling of financial services products. Expert Systems with Applications 40(5), 1748–1757. Kuang, D., Nielsen, B. and Nielsen, J. P. (2009). Chain Ladder as Maximum Like- lihood Revisited. Ann. Actuar. Sci. 4(1), 105–121. Mack, T. (1991) A simple parametric model for rating automobile insurance or estimating IBNR claims reserves. Astin Bull. 21(1), 93–109. Mack, T. (2008) The prediction error of Bornhuetter-Ferguson. Astin Bull. 38(1), 87–103. Mart́ınez-Miranda, M.D., Nielsen, B., Nielsen, J.P. and Verrall, R. (2011) Cash flow simulation for a model of outstanding liabilities based on claim amounts and claim numbers. Astin Bull., 41(1), 107–129. Mart́ınez-Miranda, M. D., Nielsen, J. P. and Verrall, R. (2012). Double Chain Ladder. Astin Bull. 42(1), 59–76. Mart́ınez-Miranda, M. D., Nielsen, J. P Sperlich, S. and Verrall, R. (2013a). Contin- uous Chain Ladder: Reformulating and generalising a classical insurance problem. Expert Systems with Applications 40, 5588–5603. Mart́ınez-Miranda, M. D., Nielsen, J. P and Verrall, R. (2013b). Double Chain Ladder and Bornhutter-Ferguson. N. Am. Actuar. J. 17(2), 101–113. Mart́ınez-Miranda, M. D., Nielsen, J. P and Verrall, R. (2013c). R-package “DCL”: Claims Reserving under the Double Chain Ladder Model (v. 0.1.0, 25 October 2013). http://cran.r-project.org/web/packages/DCL/index.html. Merz, M. and Wüthrich, M.V. (2010). Paid–incurred chain claims reserving method. Insurance: Mathematics and Economics, 46(3), 568–579. Nielsen, J. P. (2016). The link between classical reserving and granular reserving through double chain ladder and its extensions. Transcripts of the London discus- sion. British Actuarial Journal. 21, 1–17. 26 Pigeon, M., Antonio, K., Denuit, M. (2014). Individual loss reserving using paid– incurred data. Insurance: Mathematics and Economics. 58, 121–131. Posthuma, B., Cator, E.A., Veerkamp, W., Zwet, van E.W. (2008). Combined analysis of paid and incurred losses. CAS E-Forum Fall. 272–293. Quarg, G. and Mack, T. (2004). Munich chain ladder. Blätter der DGVFM 26(4), 597–630. Schmidt, K. D. and Zocher, M. (2008). The Bornhuetter-Ferguson Principle. Vari- ance 2(1), 85–110. Venter, G. G. (2008). Distribution and value of reserves using paid and incurred triangles. CAS E-Forum Fall. 348–375. Verrall, R. (1991). Chain ladder and Maximum Likelihood. Journal of the Institute of Actuaries 118, 489–499. Verrall, R. (2004). Stochastic models for the Bornhuetter-Ferguson technique. N. Am. Actuar. J. 8(3), 67–89. Verrall, R., Nielsen, J. P. and Jessen, A. (2010). Including Count Data in Claims Reserving. Astin Bull. 40(2), 871–887. 27 
work_i2b3r7sbnbgvdmmp377zahzrra	----	A scenario-based modeling method for controlling ECM performance Cristina Lópeza* and Alessio Ishizakab aUniversity Pablo of Olavide, Department of Business Administration and Marketing. Pedro R. Campomanes Building, Road Utrera, km. 1 – 41013 Seville, Spain bUniversity of Portsmouth, Portsmouth Business School. Richmond Building, Portland Street, Portsmouth PO1 3DE, United Kingdom *Corresponding author Cristina López. Telephone number: + 00 34 95 4977324. Fax number: + 34 95 4348353. Email address: clopvar@upo.es. Postal address: Ctra. Utrera, km. 1 – 41013 Seville, Spain ABSTRACT Firms are increasingly more integrating Enterprise Content Management (ECM) into their IT infrastructure in order to manage a growing volume of complex structured and unstructured data. These packages allow capturing, generating and delivering accurate information and contents in real time to support decision making in the whole business process. However, it is not easy to achieve the full benefits of an ECM package as this relies on a variety of indicators from the organizational and technical perspectives. This paper proposes a new expert system to monitor ECM performance. It is based on the innovative hybrid technique incorporating Fuzzy Cognitive Maps (FCM) and the Analytic Hierarchy Process (AHP). As these indicators are interrelated, experts’ knowledge is extracted, modeled, combined and processed in the new proposed expert system. This enables managers to precisely forecast the impact of changes in control indicators on system performance through the simulation of different scenarios over time. This approach has practical impacts as managers can make informed decisions based on the analysis of the expert system and thus prevent ECM malfunctions or misuses. Keywords: Enterprise Content Management; Decision Support Systems; System performance; FCM; AHP. 2 1. Introduction In our digital era, companies face the challenge of handling vast volumes of complex information and contents arising from their own worldwide daily activities. Therefore, executive managers require help from operational support systems to identify, generate and evaluate relevant information. This is a crucial step to get a more adequate acknowledgement and treatment of uncertainty in decision support endeavors. Hence, many firms have adopted diverse decision support tools in their computing infrastructure: Enterprise Information Systems (EIS) (Leidner & Elam, 1993), Expert Systems (ES) (Luconi, Malone & Morton, 1986), Decision Support Systems (DSS) (Sprague, 1980), and Group Decision Support Systems (GDSS) (DeSanctis & Gallupe, 1987). As technology evolves, new business analytic tools, such as Enterprise Content Management (ECM) systems, have also emerged to support the integrated wide management of all types of information and contents (Smith & McKeen, 2003; Tyrväinen, Päivärinta, Salminen, & Iivari, 2006). ECMs are computer-based information systems that employ web technologies for capturing, processing, storing, protecting, maintaining and delivering structured and unstructured contents related to business processes (AIIM, 2010; Jan Vom Brocke, Simons, & Cleven, 2011). These solutions have the capacity to efficiently handle the whole content’s lifecycle, as well as the possibility of integrating it with other sources of data. Likewise, they provide different means for analyzing the information over a wide range of managerial decisions. Thus, ECM can enhance the decision-support capabilities of adopter firms (Alalwan, 2013), as well as complying with legal requirements (Blair, 2004). 3 The research on ECM adoption has increased in the last years (Alalwan & Roland, 2012; Grahlmann, Helms, Hilhorst, Brinkkemper, & Van Amerongen, 2012; Tyrväinen et al., 2006). Recent studies have explored the key challenges in ECM adoption (Andersen, 2008; Hullavarad, O’Hare, & Roy, 2015; Vom Brocke, Simons, Herbst, Derungs, & Novotny, 2006). Researchers have also proposed a step-by-step development process to successfully implement ECM systems (Nordheim & Päivärinta, 2006; O’Callaghan & Smits, 2005). Once the ECM implementation is accomplished and the application is operative, the normal ECM activity can provide a wide range of benefits at an operational, tactical and strategic level (Alalwan, Thomas, & Weistroffer, 2014; Paivarinta & Munkvold, 2005; Smith & McKeen, 2003; Tyrväinen et al., 2006). Yet to achieve them, organizations have to control the performance of their enterprise systems until all bugs, misuse, etc., are corrected (López & Salmeron, 2014). Despite its significance, this issue has received little attention. The goal of this paper is to propose a scenario-based approach for controlling ECM performance. Our approach is based on a new knowledge acquisition technique for expert systems based on the hybrid method combining Fuzzy Cognitive Maps (FCM) and the Analytic Hierarchy Process (AHP). Many different industries have implemented expert systems (Wagner, 2017) . One of the most cited obstacles to a successful development of an expert system is the problem of acquiring specific domain knowledge and representing it (Holsapple, Raj, & Wagner, 2008). In our paper, we propose FCM, which are capable of modeling a target real-world dynamic system, which may not be well-defined (Özesmi & Özesmi, 2004). Therefore, the connections between them can be represented by fuzzy weights 4 on a linguistic scale (Kosko, 1986). Linguistic expressions are more intuitive than quantitative scales to represent knowledge. Nevertheless, the problem is how to transform this scale into a quantitative one. For this purpose, we use AHP (Saaty, 1977). Furthermore, FCM provide mechanics to study the evolution of a scenario at successive times. This soft computing technique also enables developing “what-if” analysis to investigate alternative scenarios. For all these reasons, FCM have been successfully applied in diverse areas such as marketing (Lee, Lee, Lee, & Lim, 2013), tourism (Kardaras, Karakostas, & Mamakou, 2013), agriculture (Papageorgiou, Markinos, & Gemptos, 2009), medicine (Lee, Yang, & Han, 2012), and energy (Espinosa-Paredes, Nuñez-Carrera, Laureano-Cruces, Vázquez-Rodríguez, & Espinosa-Martinez, 2008; Kyriakarakos, Dounis, Arvanitis, & Papadakis, 2012), among many others. In the present study, the new proposed expert system will enable the ECM performance to be forecasted by allowing relevant control indicators to interact with each other. That is, for example, how an increase in the numbers of business tasks supported by the ECM would affect system complexity. In this way, three scenarios related to technological, decision- making and organizational indicators were simulated. The results reveal that depending on managers’ actions, ECM performance can be improved or damaged to differing degrees. As a result of this analysis, practitioners will be able to apply adequate response actions aimed at maintaining a proper ECM performance. This research is organized into five sections. Section 2 presents a brief review of the ECM literature. Section 3 sets out the theoretical background of the new scenario-based method 5 proposed. Section 4 describes a real case study where the new approach is applied. Section 5 discusses contributions from both theoretical and practical perspectives. Finally, Section 6 offers the concluding remarks and the future research perspectives. 2. Enterprise Content Management Based on the convergence of the two previous mechanics - Document Management (DM) and Content Management (CM) - for managing unstructured information, ECM packages first appeared in the late 1990s (Boiko, 2001). This term was specifically coined in 2001 by the Association for Information and Image Management (AIIM), a global forum of information professionals (http://www.aiim.org/). Since then, it has been widely extended among key actors involved in managing organizations’ digital information assets. The ECM term is related to the integrated management of all information assets closely linked to the day-to-day conduct of business activity. These systems have been defined on many occasions from both the technical and the process perspectives. Smith and McKeen (2003) proposed one of the most referenced ECM definitions in the literature, which is: “the strategies, tools, processes and skills an organization needs to manage all its information assets (regardless of type) over their lifecycle”. This definition stresses the following aspects, which characterize ECM as regards other enterprise systems:  ECM can be considered not only as a set of isolated technologies (Jan Brocke, Seidel, & Simons, 2010), since it also implies people and processes (Blair, 2004). In fact, these solutions require the combination of strategies, processes, tools and even skills to 6 successfully manage the content of assets in the adopter company as a single solution. Over the past few years, ECM systems have rapidly evolved in line with business technology trends such as cloud computing (Alalwan & Roland, 2012; Hullavarad et al., 2015). This has enabled content to be made internally and externally accessible (i.e., between the adopter company and its suppliers and/or customers).  ECM packages are capable of supporting from well-structured data (i.e., reports, word processing documents, spreadsheets) to less-structured data (i.e., emails, webpages) and even non-informational assets (i.e., videos and music files) (Tyrväinen et al., 2006).  These enterprise-wide applications manage the entire information resource lifecycle by means of web technologies and a repository (on-site or in the cloud), which is accessible via internet through a central interface. That is, ECM systems allow the capturing, management, storage, preservation, and delivery of contents and information profoundly related to the organizational processes in the most efficient way (AIIM, 2010). ECM implementation can yield many benefits to adopter companies (Alalwan et al., 2014; Paivarinta & Munkvold, 2005; Salamntu & Seymour, 2015; Smith & McKeen, 2003; Tyrväinen et al., 2006). In order to achieve them, capabilities embedded within the system package have to fit the functionalities of the business processes in the implementing companies (Seddon & Calvert, 2010). Notwithstanding, processes are in constant change because they need to adapt to the environment’s requirements. ECM should be therefore 7 adjusted to business activities throughout its lifecycle. This encompasses the following main stages described hereafter (Alalwan & Roland, 2012). During the adoption stage, managers have to prepare a comprehensive project feasibility study. This should detect the effects of ECM adoption on the firm performance. Iverson and Burkart (2007) present a framework for evaluating these impacts. The feasibility study should explore a wide range of issues related to ECM implementation, such as project risks, legal issues, benefits, costs, barriers and challenges of change management required. It must stress the active user support in the change management required for the successful ECM adoption (Munkvold, Paivarinta, Hodne, & Stangeland, 2003). In particular, Van Rooij (2013) describes how legacy issues need to be taken into account in the ECM adoption. Allen (2007) explains how to develop a return on investment (ROI) model when investing in an ECM system. Based on the viability analysis, decision makers will or will not implement the ECM solution. An affirmative decision marks the beginning of the acquisition stage. This consists in looking for the ECM solution that best suits business requirements. Choosing the most suitable content management system from among the large number of options on the market is a complex and difficult process. The decision makers have to consider multiple functional and non-functional criteria of different levels of importance (Lin, Hsu, & Sheen, 2007) under the constraint of a limited budget, time and human resources. In order to support this task, several studies have proposed a framework to compare ECM tools under specific criteria (Escalona, Domínguez-Mayo, García-García, Sánchez, & Ponce, 2015; Vitari, Ravarini, & Rodhain, 2006). In the same line, Oztaysi (2014) builds a hybrid AHP-grey TOPSIS model to evaluate 8 several ECM alternatives. Moreover, Votsch (2001) provides a taxonomy to clearly identify different kinds of content management systems. The evolution stage covers all the activities needed to install the ECM package chosen in the IT infrastructure, to integrate the system within the existing information sources and to test it. Nordheim and Päivärinta (2006) describe in detail how they implement the ECM in a large Norwegian oil firm. Simons, Brocke, Lässer, & Herbst(2014) distinguish important lessons to bear in mind during ECM implementation, while Herbst, Simons, Brocke, & Derungs (2014) explore critical success factors for ECM readiness assessment. Vom Brocke et al. (2006) present an ECM-blueprinting framework to guide professionals in the system implantation and at the same time to re-design affected business processes. Hullavarad et al. (2015) identify the main challenges a company faces during this stage. They also provide a graphical analysis of how each implantation phase gradually improves the business efficiency. In addition, other studies describe critical factors for successfully implementing ECM systems (Haug, 2012; Horne & Hawamdeh, 2015). Although the ECM implementation can finish successfully, it does not guarantee the long- term performance of the ECM. Accordingly, during the evaluation stage, managers assess the system performance and its effect on business activity. In this regard, Scott (2011) explores how users perceived ECM performance and what factors lead to the system’s acceptance. Arshad et al. (2014; 2015) studied how ECM usage supports business processes and Alalwan et al. (2014) the decision making. Subsequently, managers compare evaluation results with expected benefits and impacts. They can thus plan response actions to correct misalignment, 9 underperformance and/or even system bugs. A previous work underlined the need to develop new ECM evaluation practices (Paivarinta & Munkvold, 2005). Nonetheless, little attention has been paid to addressing this issue. A recent work provides a specific ECM research agenda based on an extensive literature review (Alalwan & Roland, 2012). This encourages further research that determines how ECM performance should be assessed and what control indicators ought to be used to evaluate ECM performance from different perspectives. Hence, this study provides a novel technique to bridge this gap between the literature and the practice. The next section presents the method proposed. 3. Research method In this paper, we will use a hybrid method combining Fuzzy Cognitive Maps and the Analytic Hierarchy Process to capture the importance of elements. FCM comprise theoretical foundations of Cognitive Maps (CM). This method (also called mental maps) was introduced in psychology by Tolman (1948) to represent the mental representation of physical locations. Irrelevant or unimportant information is omitted from the mental map. Thus, CM can be very different from an actual place. The differences between the mental representation and the physical characteristics of a location may reveal what humans or animals consider important. Later, Axerold (1976) advanced the idea of CM for supporting decision making. In addition to the important points (called nodes), he added causal connections between them (called edges). CM have been thereafter useful in problem solving (Eden, 2004) when many decisional variables are causally interrelated (Kim & Lee, 1998) because this can help decision 10 makers to highlight and analyze hidden relationships that contribute the most to attaining relevant and significant solutions. CM are signed digraphs where only the direction of the change between two nodes is modeled. Fig. 1 presents a CM example. A positive edge (noted with a positive sign +) indicates that the causal node casually increases or decreases the effect node in the same direction. A negative edge means that the causal node increases or decreases the effect node in the opposite direction. FCM were introduced to model the intensity of the change when an event occurs to some degree (Kosko, 1986). This is a graph-based knowledge representation technique (Dickerson & Kosko, 1993) which models a static or dynamic system using causal dependencies between a set of n nodes 𝑉 = (𝑣1,𝑣2,…,𝑣𝑛). FCM with fuzzy weights have the ability to deal with uncertain and imprecise data. Fig. 2 includes an FCM example, where numbers in brackets represent negative connections. The entire relationships can be represented in an adjacency matrix with their sign and intensity (1). If there is no relationship, then the entry is empty. 𝑊 = ( … … … 𝑤1→𝑛 … … … … … … 𝑤𝑖→𝑗 … … … … … … … … 𝑤𝑛→𝑛) (1) where 𝑤𝑖→𝑗 is a directed edge which indicates the influence of the causal node (𝑣𝑖) on the effect node (𝑣𝑗). 11 Fig. 1 CM example Fig. 2 FCM example FCM can be based on interviews, a Delphi process, and group discussions, among others (Jetter & Kok, 2014; Kardaras et al., 2013; Özesmi & Özesmi, 2004). They can be easily modified or extended by adding new nodes or causal links, or changing the weight assigned to the causal link. The difficult part is to assign the weight to the causal links. Some researchers have used a linguistic evaluation (Mourhir, Rachidi, & Karim, 2015; Obiedat & Samarasinghe, 2016). However, the difficulty is only shifted because the question remains on how to translate a linguistic evaluation into a quantitative evaluation. In this paper, we propose a new method based on AHP (Ishizaka & Labib, 2011; Saaty, 1980) for this transformation. The n linguistic terms are pairwise compared in a square matrix A (2) on a 1-9 evaluation scale, where 1 indicates equal importance and 9 extreme importance. A = (2) where aij is the comparison between the linguistic term i and j The matrix is reciprocal aij = 1/aii with the diagonal being equal at the unity because the              1...... ....../1... ...... 1 1 21 112 n ijji ij n a aa aa aa 12 linguistic term is compared with itself. Therefore, only the upper part of the matrix is required from the decision maker (Ishizaka, 2012). If a matrix is sufficiently consistent, the weights are calculated as shown in formula (3) (Ishizaka & Lusti, 2006): AW=λmaxW (3) where A is the comparison matrix, λmax is the principal eigenvalue W is the vector of the weights. As A has a redundancy of information, the consistency of the judgments entered by the decision maker can be tested with the consistency ratio (CR): CR=CI/RI (4) where CI=(λmax−n)/(n−1) is the consistency index n is the dimension of the comparison matrix λmax is the principal eigenvalue RI is the ratio index. The ratio index (RI) is the average of the consistency index of 500 randomly filled matrices. Saaty (1977) considers that a consistency ratio exceeding 10% may indicate a set of judgments that is too inconsistent to be reliable and therefore recommends revising the evaluations. FCM can be used not only for a simple causal reasoning of the phenomena represented. They also predict ECM performance systems by means of simulations of scenarios. In doing so, FCM incorporate the concept of neurons in the sense that they can be “on” (+1) or “off” (- 1), but also states in-between and therefore fuzzy states. When a node changes its state, it 13 affects all other connected nodes. If the threshold level of the effect node is reached, it will also change state and by consequence may also change further nodes within the network. Nodes already activated may be even activated again due to a feedback loop. By consequence, the activation spreads in a non-linear manner until the system reaches its stability or shows chaotic behavior. The inference process begins by assigning an input value [0, 1] to each FCM node, which corresponds to the initial state vector: 𝑉𝑆𝑖 0 = (𝑣1 0 𝑣2 0 … 𝑣𝑛−1 0 𝑣𝑛 0) (5) where 𝑣𝑛 0 points out the value of the node at the instant 0. During the simulation process, inputs are computed through a finite number of interactions in chain according to the following formula (Stylios & Groumpos, 2004) 𝑣𝑖 𝑡+1 = 𝑓(𝑣𝑖 𝑡 +∑ 𝑣𝑗 𝑡 ∙𝜔𝑗→𝑖 𝑛 𝑖≠𝑗 ) (6) where 𝑣𝑖 𝑡+1 is the value of node 𝑣𝑖 at the instant 𝑡 +1, 𝜔𝑖→𝑗 is the fuzzy weight between nodes 𝑣𝑗 and 𝑣𝑖, and 𝑓(𝑥) is the transformation function. A new value is calculated with (6) for each node at each time step. The most commonly applied transformation functions are the sigmoid function, the hyperbolic tangent function, the step function and the threshold linear function (Bueno & Salmeron, 2009; Yaman & Polat, 2009). Researchers should assess them in order to select the one that is most suitable to the requirements of the study. 14 FCM inference finishes when the limit vector is reached. This happens when either 𝑉𝑡 = 𝑉𝑡+1 or 𝑉𝑡+1 −𝑉𝑡 ≤ 𝜀; where 𝜀 is a residua (Espinosa-Paredes et al., 2008). The inference process can also result in a limit cycle. This implies that the vector state continues changing around several fixed states. When a continuous function is used, chaotic behaviour is possible (Papageorgiou, 2011). This happens when the inference process finds different outputs for each time step and, therefore, the FCM does not attain stability. The next section presents an application validating the method proposed. 4. Case Study It is well known that companies must control the normal operation of their installed enterprise systems in order to achieve the expected benefits and this is also the case for an ECM adoption. Therefore, firms need an appropriate method to detect whether these solutions work as expected. Hence, the present study proposes using the hybrid FCM and AHP technique to monitor ECM performance. This approach is explained through a real case study. Unlike empirical studies which pursue generalized findings, case studies have proved to be an excellent vehicle to test the applicability of the technique in a concrete and real-life situation (Yin, 2013). 4.1 Company description The case study was carried out in a European company which is a pioneer in providing services for Business Process Outsourcing (BPO) and digital technology. It had over 18 years of experience and more than 70,000 employees in 2015 and generated a turnover of more than 15 4,800 million euros. Today, the case study company operates services in over 40 countries around the world. The case study was carried out in one of their platforms located in Spain. The platform activity mainly focuses on providing a set of services related to customer relationship management, data intelligence and a wide range of business solutions. This branch implemented an ECM solution several years ago. Since then, the computer-based system centralizes in a single electronic solution the management of information assets (including structured and unstructured) generated and used in their own business processes. The tool provides relevant, precise and up-to-date information to help users in decision making and business process development. For this purpose, the enterprise system and the business process have to be completely aligned (Seddon & Calvert, 2010). If this were not the case, ECM underperformance would not occur. To control if this is really happening, the participants of the case study shared with us their ECM experience from a technological and a business perspective. The following subsection describes their active participation in building the FCM model. 4.2 Building the FCM model Different methods can be used to build FCM (Kang, Lee, & Choi, 2004; Sangjae Lee & Ahn, 2009; Papageorgiou et al., 2009; Schneider, Shnaider, Kandel, & Chew, 1998). Often each expert builds his/her own FCM. They represent their knowledge of their particular expertise. As there may be differences between their FCM, it may be necessary to use other methodologies to reach a consensus between participants, such as Delphi or the augmented FCM method. 16 In the augmented FCM method, the adjacency matrix of each participant is added to obtain the final diagraph-based FCM model (Dickerson & Kosko, 1993). This approach does not need participants to change their former opinions to obtain a consensus as in the Delphi methodology (Salmeron, 2009). In addition, experts’ models are not constrained by a closed list of concepts in such a way as to ensure that the final FCM represent all the insights. For these reasons, we decided to use this approach to build the FCM. Fig. 3 summarizes the main activities during the FCM building process. In order to build the FCM, which helps managers to control the performance of their ECM solutions, we held individual face-to-face interviews with the participants during the months of September and October 2016. They then occupied the IT manager and the operation manager positions. Each one built his/her own FCM model during the interview itself. We therefore designed a questionnaire for the purpose of guiding participants during the FCM building process. The questionnaire was organized into four parts. In the first part, the participants answered general open questions about their company and the ECM implemented. The goal was to elicit a better understanding about the firm’s activities and the system in use. In the second part, the participants identified the most relevant indicators that control ECM performance in an effective way. To make this work easier, we prepared an initial list of indicators based on the literature (Alalwan et al., 2014; Brocke et al., 2010; López & Salmeron, 2014). The participants could check it and add new elements. They could also remove elements that according to their experience were irrelevant. 17 Fig. 3 FCM building process Table 1 summarizes the final FCM nodes and their descriptions. Each indicator (I) node represents one specific measure required to monitor or control the ECM performance. The ECM node corresponds to the system performance. The nodes I1-I5 are related to business tasks supported, the nodes I6-I11 are related to decision making and the nodes I12-I14 are related to the technological perspective. 18 Table 1 FCM nodes, in italics, are new nodes proposed by the participants Code Nodes Description I1 Business tasks supported Number of business tasks supported by the ECM system I2 Efficiency of business tasks supported Users can carry out their business task in a more efficient way I3 Automatizing business task Procedures are properly defined and therefore can be automated in the ECM system I4 Velocity of business task development Users develop more easily and therefore more quickly business tasks when supported by ECM I5 Compliance with standards, certifications and rules Adopter company complies with standards, certifications and rules supported by ECM I6 Decision-making speed Users make quicker decisions I7 Decision-making quality Outcomes of the decision are usually accurate (without errors) and precise I8 Relevant information available Users can access diverse sources of relevant information during business tasks development I9 Creativity and critical thinking ECM system encourages creativity and critical thinking of users I10 Problem identification speed ECM system helps users to shorten the time frame to identify key factors related to the problem I11 Alternatives evaluation Users can evaluate more alternatives I12 System complexity Interactions of the system with existing architectural layers as well as between components and subsystems. I13 System volatility System ability to keep operating despite continuous changes I14 System adaptability System ability to work in new environments ECM ECM performance Indicators describing ECM performance In the third part, we defined the five linguistic terms used in the research to reflect the influence or strength of the causality (𝑤𝑖→𝑗) between a pair of nodes < 𝑣𝑖,𝑣𝑗 >. These are depicted in Table 2. Yet, fuzzy weights need be translated into real numbers in the space [0, 19 1] to obtain the final digraph-based FCM model. With this in mind, the participants carried out a pairwise comparison between linguistic terms using a 9-point scale proposed in (Saaty, 1977). Fig 4 shows an extract of the questionnaire. Later, the comparisons were imported into Expert Choice (http://expertchoice.com/). This software enabled us to calculate real numbers with the AHP method (Section 3). Table 2 also shows the real number derived from it. Table 2 Linguistic variables and the associated quantitative value. Abbreviation Description Quantitative value VW Very Weak 0.025 W Weak 0.049 M Moderate 0.113 H High 0.243 VH Very high 0.571 Fig. 4 Extract of the questionnaire Circle one number per row below using the scale: 1 = Equal 3 = Moderate 5 = Strong 7 = Very strong 9 = Extreme 2, 4, 6, 8 are intermediate values Compare the relative performance of one linguistic term with all other linguistic terms to determine the strength of relationship between indicators of ECM performance. Very weak 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Weak Very weak 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Moderate Very weak 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Strong Very weak 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Very strong Weak 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Moderate Weak 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Strong Weak 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Very strong Moderate 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Strong Moderate 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Very strong Strong 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 Very strong http://expertchoice.com/ 20 Finally, in the fourth part, the participants drew their FCM model. To do so, they defined the causal relationships between nodes. This can be expressed in the form of IF-THEN rules, where the sender or influencing node follows a binary code (ON or OFF) and the receiver or influenced node increases (+) or decreases (-) by a level evaluated using linguistic terms. With this in mind, the participants assigned the 𝑤𝑖→𝑗 values and generated their adjacency matrices (𝑊𝑃𝑖). When a participant describes them, three points exist and must be kept in mind. In the first place, they should be aware that the 𝑤𝑖→𝑗 weight indicates how strongly the 𝑣𝑖 node influences the 𝑣𝑗 node. Secondly, the strength of the relationship is given by a fuzzy weight preceded by a positive or negative sign. This shows whether the connection is direct (+) or inverse (-). Lastly, they also define the causal relation between variables. That is, if the 𝑣𝑖 node is a cause of the 𝑣𝑗 node or vice versa. We generated the final adjacency matrix (𝑊𝐴𝑢𝑔) on the outputs achieved in the previous activities. To do this, we added the 𝑊𝑃𝑖 of each participant When more than one participant assigns the 𝑤𝑖→𝑗 value, then 𝑤𝑖→𝑗 𝐴𝑢𝑔 = ∑ 𝑤𝑖→𝑗 𝑃𝑘𝑛 𝑘=1 𝑛⁄ where k is the identifier of each participant and 𝑛 is the number of participants. The FCM model also incorporates connections indicated by one participant. These do not require any additional transformation. The 𝑊𝐴𝑢𝑔 obtained is the following: 21 𝑊𝐴𝑢𝑔 = ( .000 .000 .243 .000 .000 −.24 .178 .000 .000 .000 .000 .178 .243 .000 .178 .000 .000 .000 .000 .000 .000 .113 .000 .000 .000 .000 .000 .000 .000 .000 .000 .243 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .113 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .243 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .113 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .113 .000 .000 .000 .000 .000 −.11 .178 .000 .000 .000 .342 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .113 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .049 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .243 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 −.24 .000 .000 .000 .000 .000 .000 .113 .000 −.407 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 −.113 −.113 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .113 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 ) Finally, we drew a graph-based model in line with the adjacency matrix obtained. Fig.5 presents the FCM proposed. The model consists of the fourteen representative nodes of control indicators (I) as well as the representative node of system performance (ECM). It also shows the static connections existing between each of them. The numbers in brackets represent negative relationships. For example, 𝑤𝐼12→𝐸𝐶𝑀 = (.407) means that an increase of “System complexity” (I12) causes a moderate decrease of ECM performance. Fig. 5 FCM for controlling ECM performance 22 4.3 Simulating scenarios to control ECM performance ECM systems performance should be monitored from a technical and also an organizational perspective (Arshad et al., 2014; Bianco & Michelino, 2010). The FCM model incorporates indicators related to both perspectives. If we focus our analysis on the organizational perspective, we can observe nodes or indicators related to the business task supported and the decision making with ECM. If we focus our analysis on the technical perspective, we can supervise nodes related to the ECM technological issues. Accordingly, we defined the following “what-if” scenarios at the instant 0:  Scenario 1a activates indicators related to business task supported by ECM.  Scenario 1b activates indicators related to business task supported by ECM, with the exception of I1.  Scenario 2 activates indicators related to decision making using ECM.  Scenario 3 activates indicators related to the ECM technological perspective. Table 3 shows the inputs and outputs of the scenarios simulated. In this case, the fuzzy weights are within the range [-1, 1]. These values can only be transformed by the hyperbolic tangent function (7) (Feyzioglu, Buyukozkan, & Ersoy, 2007; Stylios & Groumpos, 2000). This was chosen for the simulation of the model proposed. tanh(𝑥) = 𝑒𝜆∙𝑥−𝑒−𝜆∙𝑥 𝑒𝜆∙𝑥+𝑒−𝜆∙𝑥 (7) The hyperbolic tangent function uses lambda (𝜆) as a constant for the function slope. Although 𝜆 = 5 has shown to get a good degree of normalization with the sigmoid function 23 (Bueno & Salmeron, 2009), it might tend to approximate outputs to extreme values, i.e., one and zero. On the contrary, for smaller values of 𝜆, the hyperbolic function approximates a linear function. Earlier studies researched the best 𝜆 (Mago, Mehta, Woolrych, & Papageorgiou, 2012; Miao & Liu, 2000). Based on their results, we used a 𝜆 value equal to 1 in the inference process. Table 3 presents the findings obtained by simulating Scenarios S1a, S1b, S2 and S3. They reached stability after 629, 536, 581 and 534 time steps respectively. The results express how an activated indicator may affect other indicators and the ECM’s performance. Table 3 Inputs and outputs indexed for scenarios. Scenarios S1a S1b S2 S3 Nodes Input Output Input Output Input Output Input Output I1 1 0.049 0 0.000 0 0.000 0 0.000 I2 1 0.617 1 0.588 0 0.000 0 0.000 I3 1 0.401 1 0.331 0 0.000 0 0.000 I4 1 0.049 1 0.053 0 0.000 0 0.000 I5 1 0.049 1 0.053 0 0.000 0 0.000 I6 0 -0.579 0 0.000 1 -0.209 0 -0.331 I7 0 0.571 0 0.546 1 0.617 0 0.000 I8 0 0.000 0 0.000 1 0.051 0 0.000 I9 0 0.000 0 0.000 1 0.051 0 0.000 I10 0 0.000 0 0.000 1 0.051 0 0.000 I11 0 0.000 0 0.000 1 0.364 0 0.000 I12 0 0.292 0 0.000 0 0.000 1 0.053 I13 0 0.487 0 0.000 0 0.000 1 0.259 I14 0 -0.516 0 0.000 0 0.000 1 -0.427 ECM 0 -0.751 0 0.534 0 0.489 0 -0.664 Interactions 629 536 581 534 25 The simulation of Scenario 1a shows how the changes of indicators related to business task supported may affect other indicators and the ECM performance. We see that the nodes are in the range of -0.752 to 0.572. This scenario produces remarkable influences on both the decision-making indicators group and the technological indicators group. As regards the first group, only “Decision-making speed” (I6) and “Decision-making quality” (I7) were altered in the inference process. Increases in indicators related to business task supported cause a high reduction (-0.579) in “Decision-making speed” (I6) with ECM. This is due to the fact that the negative impact of “Relevant information available” (I8) in “Decision-making speed” (I6) is higher than the positive impact of the “Problem-identification speed” (I10) effect. Indeed, if ECM users have to access diverse sources of relevant information during business tasks development, they will make slow decisions. By contrast, indicators related to business task supported have a slightly positive impact on “Decision-making quality” (0.572). In this case, the effect of “Relevant information available” (I8) is positive, the same as the “Alternatives evaluation” (I11). Therefore, if managers aim to make more accurate and precise decisions, they should increase the source of “Relevant information available” (I8) and improve “Alternatives evaluation with ECM” (I11), although these actions might slow down the decision- making processes. On the technical side, Scenario 1a leads to modifications throughout all its indicators. It exerts a positive influence on both the “System complexity” (I12) and the “System volatility” (I13) but with different intensities. “Systems volatility” receives a moderate impact (0.487), while “System complexity” grows at a low rate (0.292). “System adaptability” (I14) suffered a negative and moderate influence (-0.516). Moreover, the results reveal that Scenario 1a highly negatively affects the ECM performance (-0.751). It would be reasonable to believe that increases in indicators related to business task 26 supported causes a positive effect in system performance. However, as we can observe in the FCM model, the node number of “Business tasks supported” by the ECM system (I1) impacts positively on “System complexity” (I12) and “System volatility” (I13). Hence, we defined Scenario 1b. This scenario reveals the influence of the indicators related to business tasks supported when the number of “Business tasks supported” remains stable (I1). Contrary to Scenario 1a, this does not have any influence on any of the technological indicators. It does not slow down “Decision-making speed” with ECM (I6) as in Scenario1a. The influence on “Decision-making quality” (I7) is similar to Scenario 1a with (0.546). As before, Scenario 1b impacts moderately positively on ECM performance (.534). Therefore, the managers can carry out actions to improve indicators related to business tasks supported, although preserving “System complexity” (I12) and “System volatility” (I13). The simulation of Scenario 2 indicates how the changes of the indicators related to decision-making modify only the ECM performance. This affects moderately and positively the system performance (0.489). Nevertheless, Scenario 2 does not have any influence on the indicators related to business tasks supported and those related to technological issues. The simulation of Scenario 3 clarifies how the modification of the indicators related to technological issues may impact on other indicators and ECM performance. This reveals that the combined increase of “System complexity” (I12), “System volatility” (I13) and “System adaptability” (I14) causes a slightly moderate slowdown of the decision-making speed when using ECM (-0.331). As in Scenario 1a, Scenario 3 negatively affects ECM performance (-0.665). The comparison between the impacts of the four scenarios on the ECM performance highlights relevant issues. Fig. 6 represents the scenarios’ influence on the system 27 performance. We observe that these range from to -.752 to .534. Depending on the scenario simulated, ECM performance can be improved or damaged. Scenario 1b and Scenario 2 improved ECM performance, while Scenario 1a and Scenario 3 damaged it. It is interesting to note that Scenario 1a and Scenario 1b are very similar. The unique difference is that the node number of “Business tasks supported” remains stable (I1) in Scenario1b. Notwithstanding, there is a considerable disparity in their impacts on ECM performance. If we observe the static representation of the FCM, only I1 is directly connected to I12 and I13. This fact is reflected in the results of Scenario 1a. As mentioned before, the simulation of Scenario 3 impacts negatively on the ECM performance. In this way, the influence of node I1 would explain the negative effect of Scenario 1a on system performance. Therefore, if the users require adding new tasks supported by the systems, professionals should carefully control the system complexity and the system volatility as the ECM performance could be seriously affected. Fig. 6 Simulations effects on ECM performance. 5. Discussion ECM packages have proven to be an excellent enterprise-wide approach for managing all information assets. Nonetheless, as the above results show, system performance would 28 be altered in many circumstances. Hence, this study presents a novel expert system for controlling ECM performance since it is operative until its withdrawal. The findings highlight new advances in both the theoretical and the practical field. 5.1 Theoretical contribution Nowadays, numerous firms have already adopted ECM solutions in their IT infrastructures. In order to get a proper ECM performance, they have to continuously overcome challenges derived from the technology itself and the deployment of new work routines from both business process and decision-making perspectives. Laumer, Beimborn, Maier, & Weinert (2013) already characterized ECM packages on three levels. (1) The enterprise level makes possible the supply of required information for carrying out business-related procedures. Our model brings together indicators related to business tasks supported by system, which make it possible to control that level of ECM performance. (2) The system level represents technological artefacts on which ECM solutions run. Accordingly, ECM performance should be also controlled by technological indicators, and these are covered in the model. (3) The information level determines the data required by users. Other studies also considered the people dimension or the relationship between users and contents in their view of ECM system (Alalwan & Roland, 2012; Tyrväinen, Päivärinta, Salminen, & Iivari, 2006). This one is closely related to the information level and therefore our research proposes to combine them under the same level named decision making. Our case study enriches what we already know about ECM because it is the first time that an investigation identifies indicators used to control ECM performance from the above-mentioned levels. Nordheim & Päivärinta (2006) identified critical issues that had arisen during the ECM implementation stage, which cannot be extrapolated to ECM post- implementation stage. Brocke, Simons, Sonnenberg, Agostini, & Zardini, (2010) 29 presented a framework to measure the financial performance of business processes, while our article focuses on ECM performance. Wiltzius, Simons, Seidel, & Vom Brocke (2014) explored which acceptance factors can affect users’ perception of the usefulness and ease of use of ECM. Yet, a good level of ECM users’ acceptance does not necessarily imply a good system performance level. Hence, those factors cannot be applied to control ECM performance. Laumer, Maier, & Weitzel (2017) have recently provided indicators related to system quality, service quality and information quality for understanding users’ satisfaction and the manifestations of workarounds. Given that the dimensions under study are not the same, it is normal that most of the indicators do not match. This fact highlights the gap filled, so that there is a complementarity of our research with the existent investigations. Our research also provides a new advance in the ECM body of knowledge since the findings highlight how indicators of ECM performance are closely related. A change in one indicator can be without a direct effect on the system performance, but it may cause a cascading effect and thus impact on the package performance. With this in mind, we developed a specific expert system to predict these effects. This is based on a novel approach based on the hybridation of FCM and AHP. FCM is an artificial intelligence technique that incorporates advances from fuzzy logic and neural networks into CM approach (Ferreira, Ferreira, Fernandes, Meidutė-Kavaliauskienė, & Jalali, 2017). Table 4 summarizes the pros and cons of FCM-AHP with regard to previous modeling techniques (Baldi & Rosen-Zvi, 2005; Bañuls, López, Turoff, & Tejedor, 2017; Kitchenham, Pickard, Linkman, & Jones, 2003; López & Salmeron, 2014). 30 Table 4 Comparison of modeling methods in terms of pros and cons. Modelling techniques Pros Cons Systems Dynamics  Capable of modeling all possible relationships.  Capable of modeling directed graph with cycles.  The propagation does not follow an established pattern.  Cannot model uncertainty.  Cannot assume scarcity of information. Bayesian Networks  Capable of modeling all possible relationships.  Can incorporate uncertainty.  Cannot model directed graph with cycles.  Does not quantify the uncertainty in causal connections in a precise way.  The propagation follows an established pattern.  Cannot assume scarcity of information. Neural Networks  Capable of modeling all possible relationships.  Can incorporate uncertainty.  Assumes information is scarce.  Cannot model directed graph with cycles.  Does not quantify the uncertainty in causal connections in a precise way.  The propagation follows an established pattern. FCM  Capable of modeling all possible relationships.  Can model uncertainty.  Can model directed graph with cycles.  The propagation does not follow an established pattern.  Assumes information is scarce.  Does not quantify the uncertainty in causal connections in a precise way. FCM-AHP  Capable of modeling all possible relationships.  Can incorporate uncertainty.  Capable of modeling directed graph with cycles. 31  The propagation does not follow an established pattern.  Assumes information is scarce.  Capable of quantifying the uncertainty in causal connections in a precise way. FCM is not exempt of drawbacks. In fact, it lacks a clear instrument to transform linguistic evaluations of existing interactions into quantitative evaluations. This transformation is normally performed by using the centroid method, the max aggregation method or the mamdani inference mechanism (Mago et al., 2012), although these methods do not consider possible imprecisions in experts’ evaluations. Earlier studies bring to light that AHP can compute quantitative evaluations from the linguistic evaluation of the experts (Ishizaka & Nguyen, 2013; Meesariganda & Ishizaka, 2017). Additionally, this method provides a consistency measure, which strengthens the robustness of the final FCM model. 5.2 Practical contribution Expert systems have been long recognized to be of invaluable resources because if experts leave an organisation their experience and knowledge also goes (Dehghani & Akhavan, 2017). Our developed system is much more than an expert system. It incorporates expert knowledge but it also allows simulating scenarios that are too complex for a human mind to process. Results of simulations help practitioners to gain a better understanding on how to control and improve ECM performance. In this way, the application of our developed systems has provided important insights.  Increases in indicators related to business task supported may cause a high and negative effect on ECM performance, as well as “Decision-making speed” (I6) 32 and “System adaptability” (I14). Results highlight this effect is explained by the effect of “Business tasks supported” (I1) on “System complexity” (I12) and “System adaptability” (I13), which in turn may cause a damage in ECM performance. This consequence is very difficult to predict without the proposed system, since an increase of “Business tasks supported” (I1) directly prompts an improvement in ECM performance. Therefore, when a new business task is supported by the ECM, practitioners should monitor its effects on both system complexity and system adaptability.  Increases in indicators related to decision making may result in a better ECM performance. However, simulations reveal “Decision-making speed” (I6) can be damaged by increases in “Business tasks supported” (I1) and “System complexity” (I12). Practitioners should control them to avoid a slowdown in the final users’ decisions, and thus the possible manifestation of workarounds as explained in (Laumer, Maier, & Weitzel, 2017).  Increases in indicators related to technological issues may negatively impact on ECM performance. This happens because the “System complexity” (I12) and “System volatility” (I13) nodes impact negatively on system performance. During the post-implementation stage, IT staff identify, assess and carry out the modifications required for improving or maintaining system usability and performance. The maintenance operation can thus generate additional interactions of the system with architectural layers, as well as further components, artefacts and subsystems. Through a ripple effect, the modifications may affect other modules of the system, increasing system complexity. This might increasingly lead to costly maintenance and damages in the solution stability (Pereira, 2001). Concerning an uncontrollable volatility, it may even cause the system not to meet 33 user requirements and to need to be completely revised. Hence, increasing “System complexity” (I12) and “System volatility” (I13) should be avoided. 6. Conclusions and directions for future research The present study provides a coupled FCM-AHP method to control ECM performance during its post-implementation stage. The application of an FCM technique is proposed to represent a real-world dynamic system target. The findings provide a graphical description of the ECM performance and facilitate an increased understanding of this problem. In fact, the model enables the representation of the most relevant indicators related to ECM performance. These can be brought together into categories. The first includes the indicators related to business tasks supported by ECM. The second contains the indicators related to decision making using ECM. The third encompasses indicators related to ECM technological issues. The diagraph-based FCM shows the interaction between them. Furthermore, this illustrates their direct effects on the ECM performance. The usefulness of such a model is also explored through the FCM’s dynamic behavior. Moreover, implications for its use in ECM supervision decision making are described. FCM specifically predict ECM performance by allowing relevant indicators to interact with each other. Practitioners will thus be able to apply measures aimed at maintaining an adequate ECM exploitation. In this way, FCM simulations reveal how three categories of indicators affect ECM performance in different ways. While indicators related to decision making prompt better system performance, indicators related with technological issues cause underperformance. This is due to the fact that the negative effects of increased “System complexity” (I12) and “System volatility” (I13) outweigh the positive effects of “System adaptability” (I14) on ECM performance. This would be enhanced if managers took steps aimed at improving indicators related to decision making (I6-I11) 34 and “System adaptability” (I13). At the same time, managers should control an unavoidable rise in “System complexity” (I12) and “System volatility” (I13). The influence of the indicators related to business tasks supported is to a great extent conditioned by the effect of “Business tasks supported” (I1). If the number of business tasks is stable, advances in the other indicators may improve ECM performance. On the contrary, if the number of business tasks increases, the ECM performance will be negatively affected. This can be explained by the increased complexity and volatility of the system caused by I1. Hence, to avoid this negative chain effect on ECM performance, when new business tasks are automatized, managers ought to carry out actions to control complexity and system volatility. This study illustrates how a hybrid FCM-AHP method was applied to control ECM performance in a concrete case. It was able to quantify the uncertainty in causal connections in a precise way. Nonetheless, it would be relevant to validate the proposed model with Exploratory Factor Analysis (EFA) and/or Confirmatory Factor Analysis (CFA). This would help the development of a unique model generalizable to other cases. As the developed method is generic, flexible and easily adaptable, it can be adapted without difficulties to other studies. Concerning the artificial intelligence perspective of the developed expert system, the more data is available to represent the system, the better FCM becomes at reaching a more precise solution. However, the incompleteness and uncertainties in experts’ perceptions might hinder this and therefore a new method to overcome this issue should be developed. Furthermore, a widely recognized method that measures the accuracy of the FCM model does not in this case exist. Likewise, FCM enable what-if scenarios to be forecasted over time. Yet they lack a measure of time (e.g., when it will happen, tomorrow, in one year, etc.). These would be very relevant advances in the field. 35 Finally, it would be pertinent to study the importance of each indicator. The idea is to rank them according to the level of the importance perceived by the key actors (i.e., users, managers, analysts, and vendors). The findings from doing so would help decision makers to manage their ECM systems in a more effective way. References AIIM. (2010). What is enterprise content management (ECM)? Retrieved January 1, 2016, from http://www.aiim.org/ Alalwan, J. A. (2013). A taxonomy for decision support capabilities of enterprise content management systems. Journal of High Technology Management Research, 24(1), 10–17. 10.1016/j.hitech.2013.02.001 Alalwan, J. A., & Roland, W. H. (2012). Enterprise content management research: a comprehensive review. Journal of Enterprise Information Management, 25(5), 441–461. 10.1108/09564230910978511 Alalwan, J. A., Thomas, M. A., & Weistroffer, H. R. (2014). Decision support capabilities of enterprise content management systems: An empirical investigation. Decision Support Systems, 68, 39–48. 10.1016/j.dss.2014.09.002 Allen, D. (2007). Cost / Benefit analysis for implementing ECM, BPM systems. The Information Management Journal, (May-June), 34–41. Andersen, R. (2008). The rhetoric of Enterprise Content Management (ECM): confronting the assumptions driving ECM adoption and ransforming Technical Communication. Communication, Technical Communication Quarterly, 17, 61–87. Arshad, N. I., Bosua, R., & Milton, S. K. (2015). Towards a model to understand ECMS-use in supporting business processes. Procedia Computer Science, 72, 194–200. 10.1016/j.procs.2015.12.121 Arshad, N. I., Milton, S. K., & Bosua, R. (2014). Enterprise Content Management systems-use supports standardized business processes. International Journal of Engineering & Technology, 14(2), 81–87. Axelrod, R. (1976). Structure of decision :The cognitive maps of political elites. Princeton, New Jersey: Princeton University Press. Baldi, P., & Rosen-Zvi, M. (2005). On the relationship between deterministic and probabilistic directed Graphical models: from Bayesian networks to recursive neural networks. Neural Networks : The Official Journal of the International Neural Network Society, 18(8), 1080– 1086. 10.1016/j.neunet.2005.07.007 Bañuls, V. A., López, C., Turoff, M., & Tejedor, F. (2017). Predicting the impact of multiple risks on project performance: a scenario-based approach. Project Management Journal, 48(November), 1–20. Bianco, F., & Michelino, F. (2010). The role of content management systems in publishing firms. International Journal of Information Management, 30(2), 117–124. 10.1016/j.ijinfomgt.2009.11.001 Blair, B. T. (2004). An enterprise content management primer. Information Management Journal, 38(October), 64–66. Boiko, B. (2001). Understanding Content Management. Bulletin of the American Society for Information Science and Technology, 28(1), 8–13. 10.1002/bult.221 Brocke, J., Seidel, S., & Simons, A. (2010). Bridging the gap between Enterprise Content Management and creativity: a research framework. In Proceedings of the 43rd Hawaii International Conference on System Sciences (pp. 1–10). Honolulu, HI, USA. Brocke, J., Simons, A., Sonnenberg, C., Agostini, P. L., & Zardini, A. (2010). Value Assessment of Enterprise Content Management Systems: A Process-oriented Approach. In 36 Information Systems: People, Organizations, Institutions, and Technologies (1st ed., pp. 131–138). Berlin: Heidelberg: Physica-Verlag. 10.1007/978-3-7908-2148-2 Bueno, S., & Salmeron, J. L. (2009). Benchmarking main activation functions in fuzzy cognitive maps. Expert Systems with Applications, 36(3), 5221–5229. 10.1016/j.eswa.2008.06.072 Dehghani, M., & Akhavan, P. (2017). An experimental investigation of knowledge acquisition techniques. Journal of Management Development, 36(4), 493–514. 10.1108/JMD-07- 2016-0132 DeSanctis, G., & Gallupe, B. (1987). A foundation for the study of group decision support systems. Management Science, 33(5), 589–609. 10.1287/mnsc.33.5.589 Dickerson, J. A., & Kosko, B. (1993). Virtual worlds as fuzzy cognitive maps. In IEEE Virtual Reality Annual International Symposium (pp. 471–477). Seatle, WA, USA: IEEE. 10.1109/VRAIS.1993.380742 Eden, C. (2004). Analyzing cognitive maps to help structure issues or problems. European Journal of Operational Research, 159(3), 673–686. 10.1016/S0377-2217(03)00431-4 Escalona, M. J., Domínguez-Mayo, F. J., García-García, J. A., Sánchez, N., & Ponce, J. (2015). Evaluating enterprise content management tools in a real context. Journal of Software Engineering and Applications, 8(August), 431–453. 10.4236/jsea.2015.88042 Espinosa-Paredes, G., Nuñez-Carrera, A., Laureano-Cruces, A. L., Vázquez-Rodríguez, A., & Espinosa-Martinez, E. G. (2008). Emergency management for a nuclear power plant using fuzzy cognitive maps. Annals of Nuclear Energy, 35(12), 2387–2396. 10.1016/j.anucene.2008.07.007 Ferreira, F. A. F., Ferreira, J. J. M., Fernandes, C. I. M. A. S., Meidutė-Kavaliauskienė, I., & Jalali, M. S. (2017). Enhancing knowledge and strategic planning of bank customer loyalty using fuzzy cognitive maps. Technological and Economic Development of Economy, 4913(June), 1–17. 10.3846/20294913.2016.1213200 Feyzioglu, O., Buyukozkan, G., & Ersoy, M. S. (2007). Supply chain risk analysis with fuzzy cognitive maps. In IEEE International Conference on Industrial Engineering and Engineering Management (pp. 1447–1451). Singapore. 10.1109/IEEM.2007.4419432 Grahlmann, K. R., Helms, R. W., Hilhorst, C., Brinkkemper, S., & van Amerongen, S. (2012). Reviewing Enterprise Content Management: a functional framework. European Journal of Information Systems, 21(3), 268–286. 10.1057/ejis.2011.41 Haug, A. (2012). The implementation of enterprise content management systems in SMEs. Journal of Enterprise Information Management, 25(4), 349–372. 10.1108/17410391211245838 Herbst, A., Simons, A., Brocke, J. vom, & Derungs, R. (2014). Critical Success Factors in Enterprise Content Management: Toward a Framework for Readiness Assessment. In Enterprise Content Management in Information Systems Research (1st ed., pp. 109–124). Berlin: Heidelberg: Physica-Verlag. 10.1007/978-3-642-39715-8 Holsapple, C. W., Raj, V., & Wagner, W. P. (2008). An experimental investigation of the impact of domain complexity on knowledge acquisition (KA) methods. Expert Systems with Applications, 35(3), 1084–1094. 10.1016/j.eswa.2007.08.004 Horne, S. B., & Hawamdeh, S. (2015). Factors impacting the implementation of enterprise content management systems. Journal of Information & Knowledge Management, 14(1), 1–11. 10.1142/S0219649215500082 Hullavarad, S., O’Hare, R., & Roy, A. K. (2015). Enterprise Content Management solutions - Roadmap strategy and implementation challenges. International Journal of Information Management, 35(2), 260–265. 10.1016/j.ijinfomgt.2014.12.008 Ishizaka, A. (2012). Clusters and pivots for evaluating a large number of alternatives in AHP. Pesquisa Operacional, 32(1), 87–101. Ishizaka, A., & Labib, A. (2011). Review of the main developments in the analytic hierarchy process. Expert Systems with Applications, 38(11), 14336–14345. 10.1016/j.eswa.2011.04.143 Ishizaka, A., & Lusti, M. (2006). How to derive priorities in AHP: A comparative study. Central European Journal of Operations Research, 14(4), 387–400. 10.1007/s10100-006- 37 0012-9 Ishizaka, A., & Nguyen, N. H. (2013). Calibrated fuzzy AHP for current bank account selection. Expert Systems with Applications, 40(9), 3775–3783. 10.1016/j.eswa.2012.12.089 Iverson, J., & Burkart, P. (2007). Managing electronic documents and work flows. Enterprise Content Management at work in nonprofit organizations. Nonprofit Management & Leadership, 17(4), 18–20. 10.1002/nml Jetter, A. J., & Kok, K. (2014). Fuzzy Cognitive Maps for futures studies-A methodological assessment of concepts and methods. Futures, 61, 45–57. 10.1016/j.futures.2014.05.002 Kang, I., Lee, S., & Choi, J. (2004). Using fuzzy cognitive map for the relationship management in airline service. Expert Systems with Applications, 26(4), 545–555. 10.1016/j.eswa.2003.10.012 Kardaras, D. K., Karakostas, B., & Mamakou, X. J. (2013). Content presentation personalisation and media adaptation in tourism web sites using Fuzzy Delphi Method and Fuzzy Cognitive Maps. Expert Systems with Applications, 40(6), 2331–2342. 10.1016/j.eswa.2012.10.031 Kim, H. S., & Lee, K. C. (1998). Fuzzy implications of fuzzy cognitive map with emphasis on fuzzy causal relationship and fuzzy partially causal relationship. Fuzzy Sets and Systems, 97(3), 303–313. 10.1016/S0165-0114(96)00349-1 Kitchenham, B., Pickard, L. M., Linkman, S. G., & Jones, P. W. (2003). Modeling software bidding risks. IEEE Transactions on Software Engineering, 29(6), 542–554. 10.1109/TSE.2003.1205181 Kosko, B. (1986). Fuzzy Cognitive Maps. Int. Jornal of Man-Machine Studies, 1(April 1985), 65–75. 10.1016/S0020-7373(86)80040-2 Kyriakarakos, G., Dounis, A. I., Arvanitis, K. G., & Papadakis, G. (2012). A fuzzy cognitive maps-petri nets energy management system for autonomous polygeneration microgrids. Applied Soft Computing Journal, 12(12), 3785–3797. 10.1016/j.asoc.2012.01.024 Laumer, S., Beimborn, D., Maier, C., & Weinert, C. (2013). Enterprise Content Management. Business & Information Systems Engineering, 5(6), 449–452. 10.1007/s12599-013-0291-3 Laumer, S., Maier, C., & Weitzel, T. (2017). Information quality, user satisfaction, and the manifestation of workarounds: a qualitative and quantitative study of enterprise content management system users. European Journal of Information Systems, 26(4), 333–360. 10.1057/s41303-016-0029-7 Lee, K. C., Lee, H., Lee, N., & Lim, J. (2013). An agent-based fuzzy cognitive map approach to the strategic marketing planning for industrial firms. Industrial Marketing Management, 42(4), 552–563. 10.1016/j.indmarman.2013.03.007 Lee, S., & Ahn, H. (2009). Fuzzy cognitive map based on structural equation modeling for the design of controls in business-to-consumer e-commerce web-based systems. Expert Systems with Applications, 36(7), 10447–10460. 10.1016/j.eswa.2009.01.070 Lee, S., Yang, J., & Han, J. (2012). Development of a decision making system for selection of dental implant abutments based on the fuzzy cognitive map. Expert Systems with Applications, 39(14), 11564–11575. 10.1016/j.eswa.2012.04.032 Leidner, D. E., & Elam, J. J. (1993). Executive information systems: their impact on executive decision making. Journal of Management Information Systems, 10(3), 139–155. 10.1109/HICSS.1993.284314 Lin, H. Y., Hsu, P. Y., & Sheen, G. J. (2007). A fuzzy-based decision-making procedure for data warehouse system selection. Expert Systems with Applications, 32(3), 939–953. 10.1016/j.eswa.2006.01.031 Lopez, C., & Salmeron, J. L. (2014). Dynamic risks modelling in ERP maintenance projects with FCM. Information Sciences, 256, 25–45. 10.1016/j.ins.2012.05.026 López, C., & Salmeron, J. L. (2014). Modeling maintenance projects risk effects on ERP performance. Computer Standards & Interfaces, 36(3), 545–553. 10.1016/j.csi.2013.11.002 Luconi, F. L., Malone, T. W., & Morton, M. S. S. (1986). Expert systems: the next challenge for managers. Sloan Management Review, 27(4), 3–14. Mago, V. K., Mehta, R., Woolrych, R., & Papageorgiou, E. I. (2012). Supporting meningitis 38 diagnosis amongst infants and children through the use of fuzzy cognitive mapping. BMC Medical Informatics and Decision Making, 12(98), 1–12. 10.1186/1472-6947-12-98 Meesariganda, B. R., & Ishizaka, A. (2017). Mapping verbal AHP scale to numerical scale for cloud computing strategy selection. Applied Soft Computing Journal, 53, 111–118. 10.1016/j.asoc.2016.12.040 Miao, Y., & Liu, Z.-Q. (2000). On causal inference in fuzzy cognitive maps. IEEE Transactions on Fuzzy Systems, 8(1), 107–119. 10.1109/91.824780 Mourhir, A., Rachidi, T., & Karim, M. (2015). Employing fuzzy cognitive maps to support environmental policy development. In IEEE International Conference on Fuzzy Systems (FUZZ-IEEE) (pp. 1–8). Istambul, Turkey: IEEE. 10.1109/FUZZ-IEEE.2015.7337969 Munkvold, B. E., Paivarinta, T., Hodne, A. K., & Stangeland, E. (2003). Contemporary issues of Enterprise Content Management: the case of Statoil. In ECIS 2003 Proceedings (pp. 1– 23). Naples, Italy: AISeL. Retrieved from http://aisel.aisnet.org/ecis2003/11 Nordheim, S., & Päivärinta, T. (2006). Implementing enterprise content management: from evolution through strategy to contradictions out-of-the-box. European Journal of Information Systems, 15(6), 648–662. 10.1057/palgrave.ejis.3000647 O’Callaghan, R., & Smits, M. (2005). A strategy development process for enterprise content management. In Proceedings of the 13th european conference on information systems (ECIS) (pp. 1271–1282). Regensburg, Germany. Retrieved from http://aisel.aisnet.org/ecis2005%5Cnhttp://aisel.aisnet.org/ecis2005/148 Obiedat, M., & Samarasinghe, S. (2016). A novel semi-quantitative Fuzzy Cognitive Map model for complex systems for addressing challenging participatory real life problems. Applied Soft Computing, 48, 91–110. 10.1016/j.asoc.2016.06.001 Özesmi, U., & Özesmi, S. L. (2004). Ecological models based on people’s knowledge: A multi- step fuzzy cognitive mapping approach. Ecological Modelling, 176(1–2), 43–64. 10.1016/j.ecolmodel.2003.10.027 Oztaysi, B. (2014). A decision model for information technology selection using AHP integrated TOPSIS-Grey: The case of content management systems. Knowledge-Based Systems, 70, 44–54. 10.1016/j.knosys.2014.02.010 Paivarinta, T., & Munkvold, B. E. (2005). Enterprise Content Management: an integrated perspective on information management. In Proceedings of the 38th Annual Hawaii International Conference on System Sciences (Vol. 0, pp. 1–10). Big Island, HI, USA: IEEE. 10.1109/HICSS.2005.244 Papageorgiou, E. I. (2011). A new methodology for decisions in medical informatics using fuzzy cognitive maps based on fuzzy rule-extraction techniques. Applied Soft Computing Journal, 11(1), 500–513. 10.1016/j.asoc.2009.12.010 Papageorgiou, E. I., Markinos, A., & Gemptos, T. (2009). Application of fuzzy cognitive maps for cotton yield management in precision farming. Expert Systems with Applications, 36(10), 12399–12413. 10.1016/j.eswa.2009.04.046 Pereira, R. E. (2001). Computing ripple effect for software maintenance. Journal of Software Maintenance and Evolution, 13(4), 263–279. 10.1002/smr.233 Saaty, T. L. (1977). A scaling method for priorities in hierarchical structures. Journal of Mathematical Psychology, 15(3), 234–281. 10.1016/0022-2496(77)90033-5 Saaty, T. L. (1980). The analytic hierarchy process: planning, priority setting, resource allocation (Decision making series). New York, NY: McGraw-Hill International Book Co. Retrieved from http://www.amazon.com/The-Analytic-Hierarchy-Process- Allocation/dp/0070543712 Salamntu, L. T. P., & Seymour, L. (2015). Growth and maturation of ECM from the year 2001 to 2011. In Fifth International Conference on Digital Information Processing and Communications (pp. 31–37). Sierre/Siders, Switzerland: IEEE. 10.1109/ICDIPC.2015.7323002 Salmeron, J. L. (2009). Augmented fuzzy cognitive maps for modelling LMS critical success factors. Knowledge-Based Systems, 22(4), 275–278. 10.1016/j.knosys.2009.01.002 Schneider, M., Shnaider, E., Kandel, a, & Chew, G. (1998). Automatic construction of FCMs. Fuzzy Sets and Systems, 93(2), 161–172. 10.1016/S0165-0114(96)00218-7 39 Scott, J. E. (2011). User perceptions of an Enterprise Content Management system. In 44th Hawaii International Conference on System Sciences (pp. 1–9). Kauai, HI, USA: IEEE. 10.1109/HICSS.2011.473 Seddon, B. P. B., & Calvert, C. (2010). A multi-project model of key factors affecting organizational benefits from enterprise systems. MIS Quarterly, 34(2), 305–328. Simons, A., Brocke, J. vom, Lässer, S., & Herbst, A. (2014). Lessons Learned from Implementing Enterprise Content Management at the National Public Administration in Liechtenstein Alexander. In Enterprise Content Management in Information Systems Research (1st ed., pp. 199–216). Berlin: Heidelberg: Physica-Verlag. 10.1007/978-3-642- 39715-8 Smith, H. A., & McKeen, J. D. (2003). Developments in practice VIII: Enterprise Content Management. Communications of the Association for Information Systems, 11, 647–659. Retrieved from http://aisel.aisnet.org/cais/vol11/iss1/33 Sneed, H. M., & Brössler, P. (2003). Critical success factors in software maintenance - A case study. In International Conference on Software Maintenance (pp. 190–198). Amsterdam, The Netherlands. 10.1109/icsm.2003.1235421 Sprague, R. H. (1980). A Framework for the Development of Decision Support Systems for the Development of Decision Support Systems, 4(4), 1–26. Stylios, C. D., & Groumpos, P. P. (2000). Fuzzy Cognitive Maps in modeling supervisory control systems. Journal of Intelligent and Fuzzy Systems, 8(1), 83–98. 10.1016/S0960- 0779(98)00303-8 Stylios C.D., & Groumpos P.P. (2004). Modeling complex systems using Fuzzy Cognitive Maps. IEEE Transaction on Systems, Man, and Cybernetics - Part A: Systems and Humans, 34(1), 155–162. Tolman, E. C. (1948). Cognitive maps in rats and men. Psychological Review, 55(4), 198–208. Retrieved from http://psychclassics.yorku.ca/Tolman/Maps/maps.htm Tyrväinen, P., Päivärinta, T., Salminen, A., & Iivari, J. (2006). Characterizing the evolving research on enterprise content management. European Journal of Information Systems, 15, 627–634. 10.1057/palgrave.ejis.3000648 Van Rooij, J. C. G. M. (2013). Legacy Issues in the Implementation of Enterprise Content Management. International Journal of Information and Communication Technology Research, 3(3), 120–123. Vitari, C., Ravarini, A., & Rodhain, F. (2006). an Analysis Framework for the Evaluation of Content Management Systems. Communications of the Association for Information Systems, 18(2006), 782–804. Retrieved from http://search.ebscohost.com/login.aspx?direct=true&db=bth&AN=25591421&site=ehost- live Vom Brocke, J., Simons, A., & Cleven, A. (2011). Towards a business process-oriented approach to enterprise content management: The ECM-blueprinting framework. Information Systems and E-Business Management, 9(4), 475–496. 10.1007/s10257-009- 0124-6 Vom Brocke, J., Simons, A., Herbst, A., Derungs, R., & Novotny, S. (2006). The business drivers behind ECM initiatives: a process perspective. Business Process Management Journal, 17(6), 965–985. http://dx.doi.org/10.1108/09564230910978511 Votsch, V. (2001). A taxonomy for content management systems. The Seybold Report - Analyzing Publishing Technologies, 1(11), 13–19. Wagner, W. P. (2017). Trends in expert system development: A longitudinal content analysis of over thirty years of expert system case studies. Expert Systems with Applications, 76, 85– 96. 10.1016/j.eswa.2017.01.028 Wiltzius, L., Simons, A., Seidel, S., & Vom Brocke, J. (2014). Factors in the Acceptance of Enterprise Content Management Systems. In Enterprise Content Management in Information Systems Research (1st ed., pp. 37–61). Berlin: Heidelberg: Physica-Verlag. 10.1007/978-3-642-39715-8 Yaman, D., & Polat, S. (2009). A fuzzy cognitive map approach for effect-based operations: An illustrative case. Information Sciences, 179(4), 382–403. 10.1016/j.ins.2008.10.013 40 Yin, R. K. (2013). Case study research: design and methods (5th ed.). London, UK: SAGE Publications, In. 
work_i2jr2kurn5dxjmdc7jhztknzkm	----	Safe expert systems: simulating experts or building formal theories? The Knowledge Engineering Review, Vol. 5: 1,1990,1—4 Safe expert systems: simulating experts or building formal theories? J O H N F O X This issue of the Review includes two papers dealing with decision theory. In recent years computer based decision support systems (DSSs) have become increasingly used in fields like medicine and industrial process-control where decisions may have critical consequences. For example, outcomes may have consequences for safety and action but may be needed urgently though available information may be imperfect or incomplete and even the options open to the decision maker may be incompletely worked out. In parallel with the development of DSSs, and in particular the appearance of expert systems, concern has therefore been growing about the potential for catastrophic errors created by these systems and, worse, the potential for catastrophes whose causes cannot be established. Concern for the risks associated with expert systems is now so strong that it has spilled over into public discussion, such as a British television programme1 in which American and British practitioners and critics argued the dangers of using expert systems in medical, industrial, military and other applications, and there have been calls for restrictions on the deployment of unsupervised or autonomous systems in safety critical situations (The Boden Report, 1989). Decision making is pivotal to AI generally, not just expert systems. Systems for natural language understanding must decide between alternative interpretations of sentences; vision systems must resolve ambiguities in images. Problem solvers, planners and automated design involves deciding between alternative solutions, actions, or device components, taking into account the uncertainty arising from the unreliability and/or imprecision of available information. If significant progress in these areas leads to practical technologies then no doubt such technologies will all find their way into safety critical applications. The underlying problem, addressed in different ways by the paper on behavioural decision theory from Lehner and Adelman and on decision theory and planning from Haddawy and Rendell, is that research on sound decision procedures has not been given sufficient emphasis in either AI or knowledge engineering research. A great deal of theoretical work in developing formal foundations of AI, for example, has concentrated on qualitative reasoning and logical inference because these are seen as novel and promising foundations for building sophisticated automata. Consideration of the practical circumstances in which symbolic reasoning systems may be used has led to developments in defeasible reasoning, such as non-monotonic logic and truth-maintenance systems, that take account of the dynamic acquisition and use of unreliable information. Yet, in any practical application of AI, inference is only part of the story. Equally important is the decision-theoretic framework within which inference systems operate, which in the end determines the actions that should be taken by the decision maker. This must not only take into account whether inferences can be made soundly or not but also the cost utility of possible actions, how much information is needed to safety select them and what should happen when more information comes in and beliefs about decision outcomes are questioned. The paper by Haddawy and Rendell reviews a particular field of AI that has been largely concerned with qualitative reasoning - planning. While recognizing the technical sophistication that 1 Electric Avenue, BBC, 1989. https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0269888900005191 Downloaded from https://www.cambridge.org/core. Carnegie Mellon University, on 06 Apr 2021 at 02:05:35, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms https://doi.org/10.1017/S0269888900005191 https://www.cambridge.org/core JOHN FOX 2 the field has achieved, and the growing importance of formal logic, they argue that much more is needed and that classical theories of how decisions under uncertainty ought to be made has a significant contribution. The elements of classical decision theory are widely known but principled extensions to the theory to cover the new tasks addressed in AI, such as planning and design, time- dependent reasoning, meta-level inference and so forth, raise important challenges which range well beyond the authors' focus. A prominent idea in practical knowledge engineering and research has been that a (perhaps the) source of inspiration for system design is what human experts do. Taking this approach in designing DSSs we must elicit answers from experts to such questions as "what are the faults most likely to be seen on such and such a device?", "how likely is it that such and such a component will fail in such and such a way?". Lehner and Adelman draw attention to the large body of research into human judgement and decision making, carried out over several decades, which is not well known in the knowledge engineering community. Moreover they point out that this research has yielded a good knowledge of common biases and weaknesses in human judgement, particularly probabilistic judgement. Reasons for these weaknesses are not hard to find; our knowledge and use of quantiative parameters can be imprecise; our preferences may be inconsistent; our ability to recall and bear in mind all relevant parameters imperfect, and even experts are subject to information overloading and lapses of memory and attention. The accumulation of knowledge about human decision capabilities should be considered carefully before taking "expert knowledge" at face value. The authors provide a number of useful references for those wanting to follow up this subject. It seems clear that a good knowledge of both statistical decision theory and behavioural decision theory should be a part of the basic training of knowledge engineers and researchers in applied AI, and that there may be important gaps to be filled in both the curriculum and research agenda. However achieving a sound hybrid of knowledge-based and decision-theoretic techniques is not all that is needed if we are to build sophisticated DSSs. A hint about something else that is missing can be found in the psychological literature. The basic assumption of decision theory is that it deals with rational choice. The decision maker is a rational agent which (or who) attempts to maximize the expected value of its actions, in the light of well-calibrated estimates of the likelihood of events and the costs and benefits associated with the outcomes of alternative actions. As Lehner and Adelman discuss, many studies have indicated that human decision makers fall somewhere below this standard of rationality. It is easy to conclude, and often has been, that human decision making is irrational by comparison with formal statistical decision procedures. However the statistical criteria of rationality may be too restrictive. One can hardly deny the above observations about human performance but analysis of human decision making has tended to concentrate on its weaknesses rather than its strengths. As Shanteau (1987) remarks in an analysis of expert decision making " . . . my emphasis has been on investigating factors which lead to competence in experts, as opposed to the usual emphasis on incompetence". Shanteau identifies a number of positive characteristics of expert decision makers. Firstly, he observes that experts know a lot about their field of expertise. They know what is relevant to a specific decision, they know what to attend to in a busy environment, and they know when to make exceptions to general rules. Secondly, experts know a lot about what they know and they can make decisions about their decisions. They have good communication skills and abilities to articulate their decision processes. Furthermore they know which decisions to make, and which not to; they can adapt to changing task conditions, and they are able to find novel solutions to problems. Such observations raise additional issues for the designers of DSSs, whether these are statistical, knowledge based or hybrid systems. They fall into two main categories, performance and responsibility. Performance issues (1) The decision procedure used by the DSS must perform well (make or recommend good decisions) https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0269888900005191 Downloaded from https://www.cambridge.org/core. Carnegie Mellon University, on 06 Apr 2021 at 02:05:35, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms https://doi.org/10.1017/S0269888900005191 https://www.cambridge.org/core Safe expert systems: simulating experts or building formal theories'? 3 even in the face of degraded data. Robustness entails being able to assess the reliability of information sources and to seek alternatives where necessary, as well as merely cope with uncertainty. (2) Few practical situations involve just one class of decision (e.g. diagnosis); decision theory must surely address the problem of deciding what decision is required. (3) Many practical automata must face rapidly changing situations, not only in the information available but also in the problem that needs to be solved. Decision support systems must incorporate capabilities for altering their decision goals as circumstances develop. The central requirements for meeting these demands are that we have a sound yet flexible decision procedure. To date theoretical soundness has been the preserve of classical statistical decision procedures. However classical procedures are inflexible in the face of unanticipated events, and therefore may require critical monitoring and supervision. DSS assessment criteria must include the ability to be rationally flexible, including the ability to. a. recognize that a decision is needed b. identify the kind of decision it is c. establish a strategy for making it d. formulate the decision options e. revise any or all of the above in the light of new information A decision system should be capable of autonomously invoking and scheduling these processes as circumstances demand. Classical theory offers little guidance for developing the necessary techniques. Responsibility issues We must also achieve a high level of communication between human supervisors and/or auditors wishing to examine, and potentially to intervene in, any aspect of the decision process. (1) If decisions lead to errors it must be possible to establish the reasons for those errors. (2) Where it is practical and appropriate provision should be made for a skilled supervisor to exercise overriding control. In general a decision maker needs to be able to reflect on the decision procedure, to be able to examine the: f. decision options (what choices exist) g. data (the information available that is potentially relevant to a choice) h. assumptions (about viability of options, reliability of data etc.) i. conclusions (in light of data and knowledge of the setting) Reflective capabilities should extend to the decision process itself, including j . the goals of the decision (what is the decision supposed to achieve) k. the methods being pursued (what justifies the current strategy) 1. characteristics of specific procedures (applicability conditions, reliability, completeness etc.) A theory of rational decision making must asknowledge these requirements. Classical decision procedures may be optimal in the sense that they promise to maximise the expected benefits to the decision maker, but they must be viewed as unsatisfactory in other ways. It seems to me that the root cause is that even a clever combination of logical and probabilistic inference is not by itself an ade- quate basis for fully autonomous decision making (Fox, Clark, Glowinski and O'Neil, 1990). The papers in this issue rightly emphasize the importance of statistical decision theory and awareness of relevant research. An additional challenge for the future, however, may be to find ways of understanding how good human decision makers achieve what they achieve, to try to improve on https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0269888900005191 Downloaded from https://www.cambridge.org/core. Carnegie Mellon University, on 06 Apr 2021 at 02:05:35, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms https://doi.org/10.1017/S0269888900005191 https://www.cambridge.org/core JOHN FOX 4 these capacities using sound theories of logical and statistical reasoning, and to develop a theory of decision making which meets the need for reflection. References Boden, M (Chair), 1989. Benefits and risks of knowledge-based systems. Report of Council for Science and Society, Oxford: Oxford University Press. Fox, J, Clark, DA, Glowinski, AJ and O'Neil, M, 1990. "Using predicate logic to integrate qualitative reasoning and classical decision theory." IEEE Trans, on Systems, Man and Cybernetics (In press). Shanteau, J, 1987. "Psychological characteristics of expert decision makers" in J Mumpower (ed) Expert Judgement and Expert Systems, NATO ASI Series, vol. F35. https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0269888900005191 Downloaded from https://www.cambridge.org/core. Carnegie Mellon University, on 06 Apr 2021 at 02:05:35, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms https://doi.org/10.1017/S0269888900005191 https://www.cambridge.org/core 
work_i2ssxta7irelzjxod2rjouswra	----	r961.dvi Journal of Computing and Information Technology - CIT 15, 2007, 3, 227–235 doi:10.2498 /cit.1000961 227 An Expert System-based Site Security Officer Adesina Simon Sodiya1, Olusola Adeniran1 and Ronke Ikuomola2 1University of Agriculture, Abeokuta, Nigeria 2Federal College of Education, Akoka, Lagos, Nigeria A Site Security Officer (SSO) who is a network security staff that responds to alarms from an Intrusion Detection System (IDS), is always faced with the critical problem of low response time when the network becomes big. Even a skilled SSO is hard-pressed and less productive when collecting and analyzing IDS output manually as the frequency of intrusion increases. In this work, an Ex- pert System-based SSO (ExSSO) is designed to correct this problem. The design presents an architecture that en- codes associated expert rules for responding to different categories of intrusions into its rule-based component. The Intrusion Index (II), which determines the extent of intrusion, is calculated to classify intrusions into three categories namely low, high and very high. The inference engine component utilizes the encoded rules to interpret and respond to intrusions based on the Intrusion Index. Visual Basic 6.0 is used to implement the design because of its interactiveness and high ability to support database. Testing the new design with data from three different network environments, the result shows a system that can investigate and respond to an average of 57 intrusions per minute as against the maximum response time of 2 per three minutes in human-based SSO. Keywords: intrusion, IDS, network security, SSO, expert system, intrusion index 1. Introduction Systems and networks are subject to electronic intrusions. Computer intrusion has been a major concern in the computer security field for over two decades and it is certain that the problem is going to be on the increase. Since the prob- lem is everywhere, government agencies and commercial organizations are now implement- ing various intrusion detection systems that can monitor this computer security breaches. Intrusion detection systems (IDSs) are security tools that are used to detect traces of malicious activities which are targeted against networks and their resources (Toth and Kruegel, 2002). According to Sodiya et al. (2004), intrusion de- tection systems are systems that detect internal and external attack on computer systems and un- dertake some measures to eliminate them. An Intrusion Detection System (IDS) does not only helps the administrators to detect intrusions and limit damages, but also helps to identify the source of attacks, which sometimes acts as a deterrent, especially in case of insider attacks (Wang et al., 2006). Once intrusion detection systems have obtained information about the event, they analyzed it to find signs of intru- sion. Intrusion Response Systems (IRSs) take over after signs of an intrusion are identified and either record the attack or attempt to actively counter it. Although IRSs are tightly coupled with the ID systems themselves and also im- portant in defending against threats, not much research effort has been put into their study. Therefore, intrusion response, in most cases, remains a manual process, which has to be per- formed by the Site Security Officer (SSO). Every intrusion detection system needs to com- municate with the outside world. This may be done passively by notifying the SSO or actively by trying to hinder the intruder or striking back. It is often not possible to use active response since the IDSs generate too many false alarms and the response would hit innocent users or hosts. The only possible action is then to report what has happened to the SSO and perhaps pro- vide him with the means to investigate the event further. Toth and Kruegel (2002) mentioned that the current intrusion response systems can be di- vided into notification, manual response and 228 An Expert System-based Site Security Officer automatic response systems. Majority of IRSs operate as notification systems, which means that they simply display or forward output de- livered by the IDS (e.g. incident data) to the SSO. Usually, urgent notification is realized via e-mail or text message services over a mobile phone. Manual IRS allows the SSO to manu- ally launch countermeasures against a detected intrusion by choosing from a predetermined set of response mechanisms. With automated response, SSO gives the sys- tem power to take action upon detecting an in- trusion attempt or an intrusion. The techniques range from increasing response delays in TCP handshakes, to blocking IPs, to resetting con- nections, and removing routes. Ability of an IDS to automatically stop intrusions seems too good to be true, and, in fact, many experts see it as a sure way to shoot oneself in the foot. A typ- ical scenario might be a spoofed attack with the source address pointing at the DNS server. An automated system responding to an attack may block access to the DNS server, hence disabling the entire network (Dobruki, 2003). The two categories listed above are not proac- tive in countering an intrusion. Even when signs of an intrusion have been detected, coun- termeasures are not triggered automatically and defending the network remains the task of the SSO. This opens a time window of vulnerability between the time when the intrusion has been detected and the time when the first countermea- sure is launched. Also, it is sometimes difficult for the SSO to collect each day the output from the detection system for signs of intrusion and respond to them quickly in a case where the number and frequency of intrusion increases rapidly. This makes the function of an SSO important and complex in intrusion detection. This leads to the objective of this work, which is to remove the human involvement in IDS response and reporting by building an Expert system-based Site Security Officer (ExSSO). It is hoped that this will help the organization in adopting a kind of site security policy that will protect their database fast enough without any downtime delay. 2. Literature Review Many attempts have been made over the past ten years to improve on intrusion detection and response system. These efforts have only made ways into the effectiveness of intrusion detec- tion and have not eliminated the requirement for human expert intervention. A classification of response functions in other IDSs is given in Carver et al. (2000). The re- sponse function in a detection system can be categorized as a notification system, manual re- sponse system, or automatic response system. According to the authors, most systems today are notification systems. In Carver (2001), an Adaptive Agent-based In- trusion Response System (AAIRS) was pro- posed for intrusion response. This was the first response system implementing a notion of learning. In this work, the interface relies on human action to update its IDS confidence met- ric. An adaptive intrusion detection system is de- scribed in Ragsdale et al. (2000). This system is used together with AAIRS to provide both adaptive detection and response. The response system is relatively advanced. It keeps track of previous alarms and classifies attacks on the basis of whether they are a continuation of an existing incident or it is a new attack. Alarms from different IDS agents in the system have dif- ferent confidence metrics according to previous detection results. The confidence in a suspected incident and the nature of the incident affects the course of action taken. A study in Toth et al. (2002) proposed yet an- other promising model for automating intrusion response. The authors suggested a way of ap- proaching the problem of response to network intrusions by constructing dependency trees that model configuration of the network. Another significant work in the area of IRS in- cludes a thorough consideration of some intru- sion detection and response cost modeling as- pect by Lee et al. (2002). They provided a good introduction to modeling costs in intrusion de- tection and responses. Carver (2000) did a comprehensive and thor- ough survey to investigate 56 IDSs. From his findings, there were no deducted solutions for intrusion response. There were, however, some responses implemented in a variety of IDSs. Most of the IDSs were notification and manual response systems, which are not preferable so- lutions. There were, however, some automatic response systems as well, but these were rather An Expert System-based Site Security Officer 229 immature. There is this possibility of having a delay between an alert and human reaction when manual system responds to attack and also an automated system responding to an attack may likely block access to the DNS server, hence disabling the entire network. In his survey of ten different IDSs, Fisch asserts that most of the existing IDSs had the ability to generate daily or weekly reports of suspicion events or users (Fisch, 1996). The focus of his research work is not on the user interface for reporting and notifying the Site Security Offi- cer, but, instead, relies on automatic response. It probably had some notification capabilities, although these are not described in detail. From the literature, so far no work has fully imple- mented completely automated intrusions analy- sis and response. 3. Architecture for Expert System-based Site Security Officer (ExSSO) The main task of intrusion detection system (IDS) is to monitor the events occurring in the network and then analyze them for signs of in- trusion. Once an intrusion has been detected, IDS issues alert notifying the SSO, who then responds to the IDS alarm to make appropriate measure. The components of ExSSO is pre- sented in Figure 1. 3.1. Inference Engine Inference Engine (IE) is the central processing unit of an expert system. It uses knowledge base to draw conclusions for each situation. In- ference engine can be thought of as the system software which stimulates a situation and actu- ally makes the decision. ExSSO inference engine uses two major com- ponents in making decision. These are intrusion index and interpreter. Intrusion Index: It is used to measure the ex- tent or gravity of attack/intrusion so as to deter- mine the action to be taken when an intrusion is detected. The intrusion index is then calculated as fol- lows: SSO Intrusion Index (II) = n∑ i=1 Xi n∑ i=1 Xi max() , where: n = 3 (because there are three variables considered). The variables considered for the calculation of II are: a) Attack Category Score Confidentiality 1 Integrity 2 Availability 3 Figure 1. Architecture of an Expert system-based Site Security Officer (ExSSO). 230 An Expert System-based Site Security Officer b) Attack Implication: this has to do with the possible consequence of the attack. It is di- vided into: Low 1 High 2 Very High 3 c) Security Violation Level: this has to do with the depth of the security violation. It can also be seen as the extent of circumven- tion. Security Violation Level Score Low 1 High 2 Very High 3 X = variables score Xi max() = represents maximum score obtain- able for variable i The maximum value of SSO Intrusion is 1 (SSO Intrusion Index (II) <= 1). This is then divided into 3 ratings as follows: Intrusion Index is low when 0≤ II<0.3 Intrusion Index is high when 0.3≤ II<0.7 Intrusion Index is very high when 0.7≤ II<1 This means that if the calculated index is be- tween 0 and 0.3, then II is low, and so on. Interpreter: The inference engine contains in- terpreter that decides how to apply rules to infer new knowledge. The interpreter scans the con- dition parts of each rule until one is found that can be fired. Then the next cycle starts, and the interpreter tries to find another rule that can be fired. The execution ends if no rules are applicable. 3.2. Rule Base The major actions that are invoked when an in- trusion is found are classified into three: Deflect: Send a warning message to the sus- pected intruder and then store information of the suspected intrusion in the database. If the same kind of suspicion is found in that terminal again, the action will turn to prevent. Prevent: Prevent the user from the network re- sources. Preempt: Strikes offensively against the likeli- hood of a particular intrusion occurring later. The rule base consists of the following set of rules: R1: IF II is low THEN Deflect R2: IF II is high THEN Prevent R3: IF II is very high THEN Preempt Deflect Algorithm Attack flag = .F. Check S /* S represents the network system */ If Attack flag = .T. Scan database If Terminal = N /* N represents the terminal where the attack is coming from */ Check Sup dat If found() Prevent() else enable audit() Notify user /* by sending an e-mail warn- ing message */ Endif Endif Function audit() Scan audit database For Terminal = N Enable reporting-window /* Display all the activities occurring in the user’s terminal */ Add to Sup dat /* where Sup dat is the database keeping records of intrusion */ End Prevent Algorithm Attack flag = .F. Check S /* S represents the network system */ If Attack flag = .T. Scan database If Terminal = N /* N represents the terminal where the attack is coming from */ Prevention () /* Prevent the users from cer- tain resources */ Endif Endif Function Prevent() Disconnect N /* Disconnect terminal N from Network resources */ End Preempt Algorithm Attack flag = .F. An Expert System-based Site Security Officer 231 Check S /* S represents the network system */ If Attack flag = .T. Scan database If Terminal = N /* N represents the terminal where the attack is coming from */ Close terminal /* Disable and shut down the user’s terminal */ Endif Endif 3.3. ExSSO Organization ExSSO is not an independent IDS component and it is designed to process IDS output im- mediately. Since intrusion detection activities are real-time events, ExSSO collects data (IDS alarms) continuously from IDS and responds to these alarms. The Intrusion Index calculated determines the nature of the response (Deflect, Prevent and Preempt). Figure 2. A representation of ExSSO inference engine. 232 An Expert System-based Site Security Officer 4. Implementation Procedure 4.1. Programming Language Visual Basic 6.0 was used as the programming language for the implementation of the design. The design can run under the Window XP and uses the output (alarm) generated by the intru- sion detection system as its input. 4.2. Data Source The data source to our design is the output in form of alarm from an IDS. We got the data used for the implementation from a network organi- zation in Lagos State, Nigeria. The output of the IDS contains the following information: time & date of attack, source and destination IP address and port number, security violation level, attack implication, attack types, and attack categories. Figure 3a. Figure 3b. An Expert System-based Site Security Officer 233 Figure 3c. Figure 3d. Figure 3 (a – d): Interface design of an ExSSO. 4.3. Interface design The boxes on the ExSSO interface screen rep- resent the 20 systems on the network (Figure 3a). The flashing red signal on system 3 is an alert signal indicating that system 3 has been attacked by an intruder. The design gives detailed information about the intruder, the level of attack and the attack sys- tem. The ExSSO design output contains the following intrusion analysis : — Security Violation Level: is the attack level given the strength of suspicion (the attack level can fall within the range of low, high and very high). — Attack Implication: is the type of loss ex- perienced as a result of the attack which is graded as low, high and very high. — Attack Categories: are attacks under the three computer security properties which are: attack due to confidentiality, integrity and availability. 234 An Expert System-based Site Security Officer — Source IP address: the IP address of the system where the attack is coming from. — Destination IP address: the IP address of the attacked system. — Time: identification of time of operation. — Date: identification of date of operation. 5. Evaluation of ExSSO 5.1. Summary of the Intrusive Event In this work, the running and testing of the de- sign was carried out and the summary of the intrusive event is shown in the table below. Attack Level IncidentLevel N1 Incident Level N2 Incident Level N3 Very High 27 10 20 High 64 55 50 Low 33 18 30 Table 1. Summary of the intrusive event level. It was discovered that the intrusive event that occurred on the network during the experiment, categorized as very high, high and low has an incident level of 27, 64 and 33 on network 1, an incident level of 20, 50 and 30 on network 2 and an incident level of 10, 55 and 15 on network 3. Therefore, we can say that a high level attack has the highest incident level on the network. As a result, most of the users are prevented from certain network resources. 5.2. Summary of the Attack Categories Attack Categories Incident Level N1 Incident Level N2 Incident Level N3 Availability 14 10 16 Confidentiality 35 28 20 Integrity 75 45 64 Table 2. Attack categories. Attack categories under integrity violation al- lowing the attacker to change the system state of data residing on or passing through a system have the highest incident level of 75 on network 1, an incident level of 45 on network 2 and an incident level of 64 on network 3. The attacks that fall under these categories are determent to the state of the network system. 5.3. Summary of the Attack Types Attack Types IncidentLevel N1 Incident Level N2 Incident Level N3 Penetration 76 60 70 Scanning 34 21 20 Dos 14 2 10 Table 3. Attack types. It was discovered that most of the attacks launched on the network are penetration attacks, whereby intruders gain control of the system by exploit- ing a variety of software flaws. The penetration attacks have an incident level of 74 out of the 124 attacks on network 1, an incident level of 60 out of 83 on network 2 and an incident level of 70 out of 100 attacks on network 3. 5.4. Summary of the Response Time Network No. of Attacks Time taken for response (minutes) Response time per attack (seconds) 1 124 2.18 1.055 2 83 1.50 1.080 3 100 1.70 1.020 Average response time 1.050 Table 4. Average response time. Consequently, the number of intrusions that the systems can respond to in a minute is equal to 60/1.05 = 57.14, approximately 57 intrusions. 6. Conclusion and Future Work Since human beings are incapable of dealing with the speed and amount of information gen- erated by intrusion detection systems, this has been successfully carried out to eliminate the problems with human intervention in IDS. We designed an ExSSO algorithm for responding to attacks in the intrusion database. An Expert System-based Site Security Officer 235 The direct application of this problem in the net- work environment warrants the significance of this research effort. Experimental evaluations bring promises to this area. Our algorithms successfully respond to malicious attacks on the network. It might also be necessary to design an agent-based SSO that will move around the network and respond to intrusion alarm. References [1] C. A. CARVER, Intrusion Response Systems: A survey. Department of Computer Science, Texas A & M University, College Station, TX, (2000). [2] C. A. CARVER, U. W. POOCH, An Intrusion Response Taxonomy and Its Role in Automatic Intrusion Re- sponse. IEEE Systems, Man and Cybernetics Infor- mation Assurance and Security Workshop, (2000) pp. 129–135. [3] C. A. CARVER, Adaptive Agent-based Intrusion Re- sponse. Ph. D. Dissertation, (2001), Department of Computer Science, Texas A & M University, College Station, TX. [4] M. DOBRUCKI, Priorities in the deployment of net- work intrusion detection systems. Master Thesis, (2002), Helsinki University of Technology, Depart- ment of Computer Science and Engineering. [5] E. A. FISCH, A Taxonomy and Implementation of Automated Responses to Intrusive Behavior. Ph. D. Thesis, (1996), Texas A & M University, College Station, Texas. [6] W. LEE, W. FAN, M. MILLER, S. STOLFO, E. ZADOK, Toward Cost-sensitive Modeling for Intrusion. it Journal of Computer Security, Vol. 10, No. 1 – 2, (2002). [7] D. J. RAGSDALE, C. A. CARVER, J. W. HUMPHRIES, U. W. POOCH, Adaptation Techniques for Intrusion Detection and Intrusion Response Systems. IEEE International Conference on Systems, Man, and Cybernetics, (2000), 4, pp. 2344–2349. http://www.itoc.usma.edu/ragsdale/pubs/ adapt.pdf [8] A. S. SODIYA, H. O. D. LONGE, A. T. AKINWALE, A New Two-tiered Strategy to Intrusion Detection. Information Management and Computer Security, 12(1), (2004). [9] T. TOTH, C. KRUEGEL, Evaluating the Impact of Automated Intrusion Response Mechanisms. Proc. of the 18th Annual Computer Security Applications Conference, (2002). [10] Y. WANG, S. R. BEHERA, J. WONG, G. HELMER, V. HONAVAR, L. MILLER, R. LUTZ, M. SLAGELL, To- wards the Automatic Generation of Mobile Agents for Distributed Intrusion Detection System. The Journal of Systems and Software, (2006), Depart- ment of Computer Science, Iowa State University, United States. www.elsevier.com/locate/jss Received: September, 2006 Revised: February, 2007 Accepted: May, 2007 Contact addresses: Adesina Simon Sodiya University of Agriculture Abeokuta, Nigeria e-mail: sinaronke@yahoo.co.uk Olusola Adeniran University of Agriculture Abeokuta, Nigeria e-mail: ojadeniran@yahoo.com Ronke Ikuomola Federal College of Education Akoka, Lagos, Nigeria e-mail: deronikng@yahoo.com DR. A. S. SODIYA is a lecturer of computer science, University of Agri- culture, Abeokuta, Nigeria. His research interests are computer security, artificial intelligence, network management, and information systems. He has published his papers in both local and international journals. OLUSOLA ADENIRAN is currently a Senior Lecturer at the Department of Mathematics, University of Agriculture, Abeokuta, Nigeria. His area of specialization is Loop Theory – an area which has been found to have a lot of applications in coding and cryptology. He has been involved in many projects in computer science. RONKE IKUOMOLA is a lecturer of computer science at Federal College of Education, Akoka, Lagos, Nigeria. She has just been awarded M. Sc. degree in computer science. Her areas of interest are network security and data mining. << /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /All /Binding /Left /CalGrayProfile (Dot Gain 20%) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Warning /CompatibilityLevel 1.3 /CompressObjects /Tags /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJDFFile false /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.1000 /ColorConversionStrategy /LeaveColorUnchanged /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams false /MaxSubsetPct 100 /Optimize true /OPM 1 /ParseDSCComments true /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo true /PreserveFlatness true /PreserveHalftoneInfo false /PreserveOPIComments false /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Apply /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true ] /NeverEmbed [ true ] /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 150 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 300 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.50000 /EncodeColorImages true /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /ColorImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 150 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 300 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /GrayImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 1200 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 1200 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 >> /AllowPSXObjects false /CheckCompliance [ /None ] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName (http://www.color.org) /PDFXTrapped /Unknown /Description << /FRA <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> /JPN <FEFF3053306e8a2d5b9a306f30019ad889e350cf5ea6753b50cf3092542b308000200050004400460020658766f830924f5c62103059308b3068304d306b4f7f75283057307e30593002537052376642306e753b8cea3092670059279650306b4fdd306430533068304c3067304d307e305930023053306e8a2d5b9a30674f5c62103057305f00200050004400460020658766f8306f0020004100630072006f0062006100740020304a30883073002000520065006100640065007200200035002e003000204ee5964d30678868793a3067304d307e30593002> /DEU <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> /PTB <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> /DAN <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> /NLD <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> /ESP <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> /SUO <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> /ITA <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> /NOR <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> /SVE <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se these settings to create PDF documents with higher image resolution for improved printing quality. The PDF documents can be opened with Acrobat and Reader 5.0 and later.) /ENU (Use these settings to create PDF documents with higher image resolution for improved printing quality. The PDF documents can be opened with Acrobat and Reader 5.0 and later.) >> >> setdistillerparams << /HWResolution [2400 2400] /PageSize [666.142 926.929] >> setpagedevice 
work_i3q4vd2cqfgf5basystouk4g3a	----	Microsoft Word - $ASQ84be92bb200836fc-41787c73-1030c38ec1f-7eb0#0 R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 1 of 24 Measuring the understanding between two agents through concept similarity Adolfo Guzman-Arenasm Jesus M. Olivares-Ceja Centro de Investigacion en Computacion (CIC), Instituto Politecnico Nacional, Mexico {a.guzman, jesuso}@acm.org SUMMARY. Two agents previously unknown to each other cannot communicate by ex- changing concepts (nodes of their own ontology): they need to use a common communica- tion language. If they do not use a standard protocol, most likely they use a natural lan- guage. The ambiguities of it, and the different concepts the agents possess, give rise to im- perfect understanding among them: How closely concepts in ontology OA map1 to which of OB? Can we measure these mismatches? Given a concept from ontology OA, a method is provided to find the most similar con- cept in OB, and to measure the similarity between both concepts. The paper also gives an algorithm to gauge du(A, B), the degree of understanding that agent A has about the ontol- ogy of B. The procedures use word comparison, since no agent (except the Very Wise Crea- ture, VWC) can measure du directly. Examples are given. KEY WORDS: Ontology matching, natural language, concept similarity, degree of under- standing, imperfect knowledge. 1. Introduction and objectives Agents A and B having to communicate with each other in order to achieve their goals, do so in one of two ways: (a) (Easier) Through a private language or protocol and naming convention, which requires prior agreement among programmers’ team or a Standards Committee; (b) (More general) Through a common popular general-purpose language, most likely a natural language. This paper addresses concept communication in the (b) setting. Two points stand: (1) Such communication can not be fulfilled through direct exchange of concepts belonging to an ontology, since A and B do not share the same ontology (they do not know exactly the same, from the same point of view), and OA and OB are in different address spaces. (2) Unlike case (a), the communication language is very often ambiguous. Together, (1) and (2) give rise to imperfect understanding (confusion). Knowledge is stored in concepts, which are mapped by the talker into words of the communication language; perceived words are internalized as concepts by the listener. If the concepts exchanged are animals and plants, Latin is fine: Felix Leo represents the con- 1 OA and OB are the ontologies of agents A (the talker or sender) and B (the listener), in the rest of the paper. 1 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 2 of 24 cept lion-león-loin2 while Canabis Indica stands for the concept marijuana. Other examples of words or symbols with a unique (universal) meaning: 4, π, Abelian group, Mexico, (23° 22’ 57”N, 100° 30’W), Abraham Lincoln, Berlin Symphony Orchestra. There are also semi-universal (popular) conventions, such as standard naming for chemical compounds; the Library of Congress Catalog for books, or the USA Social Security Num- ber; they provide non-ambiguity for those who adhere. If two agents can select a non- ambiguous language (each of its words maps exactly to one concept) or convention to ex- change concepts, great: they fall in case (a). If not, they have to settle for an ambiguous lan- guage, such as English [10], falling in case (b). When A talks to B, can either discover what produces the confusion? Is it possible to measure it? How can I be sure you understand me? Can I measure how much you under- stand me? Can you measure it? These questions have intrigued sociologists; they are also relevant to agents “not previously known to each other”3 trying to “interact with free-will”,4 for which they have to exchange knowledge. The paper provides answers to: (i) What is the most similar concept cB∈OB to concept cA∈OA? How similar is cB to cA? (ii) how much does B know about OA? If two agents do not share a concept, at least partially, they can not com- municate it or about it. Thus, a measure of the amount of understanding is the number of concepts they share, and how well they share them.5 We will sharpen these measures for both the ambiguous and the non ambiguous communication language. 1.1 Related work Huhns [12] seeks to communicate several agents sharing a single ontology. The authors have constructed [9, 10] agents communicating with previously unknown agents, so that not much a priori agreement between them is possible.3 An ancestor of our sim (§3.1) matching mechanism is [4], based on the theory of anal- ogy. Most work on ontologies involve the construction of a single ontology [for instance, 15], even in those that do collaborative design [11]. Often, ontologies are built for man- machine interaction [13] and not for machine-machine interaction. Everett [3] tries to iden- tify conceptually similar documents, but use a single ontology. Gelbukh [5, 6] does the same using a topic hierarchy: a kind of ontology. Linguists [18] identify related words (se- mantic relatedness), not concepts, often by statistical comparisons. With respect to measuring how close is a concept to another, Levachkine [16, 17] makes simpler measurements between qualitative values (“words,” you can say) belonging to a hierarchy, to measure their confusion. More at §3.1.2.1 below. With respect to the communication language, we prefer [2] in decreasing order: 2 We represent concepts in Courier font or with a subscript c. A concept is language-independent: the concept cat is the same as the concepts gato-gata, gatto-gatta, chat-chatt, Katze, КОТ-КОШКА, ዊ�ዊ� (neko), meaning “a small domestic feline animal”. Concepts appear in English in this paper, for readers’ benefit. 3 It is much easier to design the interaction (hence, the exchange of concepts) between two agents (or pieces of software), when the same team designs both agents. In this sense, “they previously know each other:” each agent knows what the other expects, when (the calling sequences), and the proper vocabulary to use. 4 Not-free will or canned interactions are those that follow defined paths. For instance, the interaction between a program and a subroutine it calls, where the calling sequence (of arguments) is known to both. Free will usually requires goals, resources, and planning [19]. 5 Knowledge is also stored in the relations (or verbs, actions, processes) between objects (or nouns, subjects): §1.2. 2 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 3 of 24 OP 1. A language whose words (tokens, atoms) are formed by concepts [19]; 2. One with unambiguous words (a word represents only one concept). Examples: the Natural Numbers; the Proper Nouns; 3. One where each word has a small number of ambiguities, for instance, a natural lan- guage [18] (we use this option in this paper; see figure 1); 4. One where each word points to a foreign address space, thus representing “black boxes” that can only be compared with = and ≠. Example: a language consisting of tokens such as “toves,” “borogove,” “mome”, “outgrabe”, “fromelic,” “meratroping,” tokens that only admit equality as their comparator. These are the four choices to name a concept. Often, an ontology skips details by just keep- ing the names (and nothing else) of a concept, as when it says John [owns=wardrobe] and it does not say anything else about wardrobe. It is a shallow comcept. The book [1] gives other approaches and experimental results in semantic analysis. thing (thing, something, object, entity) { food { vegetable [contain=chlorophyll, structure=living] { onion [color=white] lemon [color=green, shape=circle_shape, taste=sour, size=3cm] } fruit { orange [color=orange_color, shape=circle_shape] } meat { pork [texture=soft, color=pink] beef [texture=medium, color=red] turkey [texture=medium, color=white] veal [texture=soft, color=red] } seed (seed, cereal) { rice [color=white, shape=oval_shape] wheat [color=yellow, shape=oval_shape] } } tool (tool, hardware) { wrench (wrench) electrical_tool (household, electrical_tool) { dishwasher [use=cleaning, structure=manmade, size=small, covering=paint] blender [use=blend, structure=manmade, covering=paint] drill mixer [use=mix, structure=manmade] } } furniture { bed table sofa (sofa, seat) bench } pet { cat [covering=hair, family=feline] dog [covering=hair, family=canine] canary [covering=feather] } relation { color { white yellow green purple orange_color } shape (shape, contour) { circle_shape (circle, oval) square_shape (square ) } taste { sweet sour } size { 3cm (3cm, 1-2in) } covering { hair feather paint (paint, painted) } structure { living (living, alive) manmade (handcrafted, manmade) } contain (contain, content) { chlorophyll } family { feline canine } use { cleaning blend mix } matter_states (matter_state) {liquid suspension gel solid gas } } } Figure 1. OP, the ontology of person P. Concepts such as sofa appear in Courier font, while the words w(sofa) = (sofa, seat) (Cf. §1.2) denoting the concept appear in Times Roman. These words do not appear in the listing if identical to the concept denoted. Thus, we see onion and not onion (onion). Facts (Cf. §1.2) appear as cat [covering=hair], meaning that, for cat, the value of relation covering is hair. OP contains 68 concepts. thing (thing, something) { 3 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 4 of 24 OS pharmacy (pharmacy, drugs, medicine, medicament) { self_service { personal_care { mouth_care feminine_hygiene burns (burns, solar_protector) repellent } diet (diet_substance) { diet_powders (diet_powders, diet_creams) diet_solutions diet_tablets } stomach { laxatives_suspension [is_a=suspension] laxatives_suppository [is_a=gel] laxatives_tablets [is_a=solid, shape=circle] antiparasites } } patent_pharmacy } general_merchandise { hardware { bathroom_kitchen electrical { socket multicontact } lightbulb manual screw } } cloth {ladies_stocking hosiery shoe } perishable { frozen_food meat { pork beef turkey veal } fruit_vegetable { garlic_onion { garlic white_onion (onion) [color=white] yellow_onion (onion) [color=yellow] purple_onion (onion) [color=purple] } bulk_fruit { bulk_citrics (citric) exotics annual_fruits season_fruits } packed_fruits {avocados packed_citrics (citric) } } cooked (cooked_food) seeds {rice cereal bean } bulk_vegetable {gourd potato lemon [color=green, shape=spheric_shape, size=3cm, taste=sour] } seafood (fish_seafood) { fish (fish) preparations (fish_preparations) } } groceries { oils (eatable_ oil) { aerosol_oil sunflower_oil corn_oil mixed_oil olive_oil } rice (seed_rice) {extra_rice integral_rice shushi_rice precooked_rice super_extra_rice (super_extra_rice) } sugar (sugar, sweetener) { aspartame (aspartame, diet_sugar) glass_sugar (sugar) black_sugar (sugar) refined saccharin standard } coffee {decaffeinated_grain grain sweetened_coffee (sweetened) normal_soluble sweetened_soluble decaffeinated_soluble special_soluble } hot_cereal { oats mixed granola } cold_cereals { cereal_bars(bar) fiber_bars granola_bars energetic_bars filled_bars cold_rice_cereal (rice) cold_oats_cereal (oat) cold_corn_cereal (corn) cold_wheat_cereal (wheat) cold_mixed_cereal (mixed_cereal) diet_cereal } chili{ancho_chili tree_chili guajillo_chili morita_chili pasilla_chili } dry_fruits(dry, fruits) {prune coconut jamaica tamarind raisin (raisin, dry_grape)} flour (flour) { rice_flour (rice) wheat_flour (wheat) corn_flour (corn) hotcakes_flour (hotcakes, flour) pastries_flour (pastries, flour) powder (powder) other_flour (barley, rye, flour) } } other { relation { color { white yellow green purple } shape (shape, contour) {circle (circle_shape) spheric_shape (sphere) square_shape (square ) } taste { sweet sour } size {3cm (3cm, 2in) } } matter_states (matter_state) { liquid suspension gel solid gas } } } Fig. 2. Superama ontology (150 concepts). To save space, words of a concept do not appear when identical to it 4 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 5 of 24 1.2 Definitions Named entity. An object, relation, property, action, process, idea or thing that has a name: a word (or word phrase) in a natural language. ♦6 A named entity is shared. Concept. The representation of a named entity. ♦7 Examples: peak-uttermost, to_- fly_in_air, angry-mad. So, concepts have names: those words (or word phrases, such as New York City) used to denote the named entity that the concept represents.2 Unfortunately, the names given by different people to concepts differ (synonymy) and, more unluckily, the same word is given to two concepts (examples: words peak; fly; mad). Thus, words are ambiguous, while concepts are not. A person or agent, when re- ceiving words from a speaker, has to solve their ambiguity in order to understand the speaker, by mapping the words to the “right” concept in his/her/its own ontology. This mapping is called disambiguation. There are also composite or complex concepts, such as “to ski in a gently slope under a fair breeze while holding in the left hand a can of beer.” These can be shared with other agents, too, but they do not possess a name: they have not been reified (Cf. §2.2). Concepts can enter also in relations (called restrictions in some Logics) with other con- cepts, as defined below. A concept represents three kinds of real-world entities: (a) sets, like animal (repre- sents all the animals); (b) individuals, like Abraham Lincoln (represents a particu- lar person), and (c) relations, such as buys. Relation. A relation of order k is a sequence of concepts (rel c1 c2 … ck)representing that relation rel holds between c1, c2, ..,ck. ♦ Notice that rel is also a con- cept (which may be shallow). Notation. For binary relations (rel c1 c2) we write in the ontology rel [c1=c2], for instance lemon [color=green] in figure 2. Notice the use of c {p q r …} to indi- cate that the subsets of c are p, q, r… More at [19]. Fact. A relation is a fact if it holds in reality. ♦ It agrees with the real world. It faithfully represents an aspect of reality. To discover facts is not easy: people make tedious ex- periments and observations to ascertain what “really happens in the real world” and what does not. Thinking or philosophizing is not enough. We are not interested in how to discover facts. Thus, we postulate that an agent “knows” or “believes” that its knowl- edge consists of only facts: truthful representations of reality. It is possible for some rel∈OA not to be a fact, but A does not know this. An agent does not lie to itself. Rela- tions not agreeing with reality are “lies.” Knowledge is the concrete internalization8 of facts among real-world entities ♦6 It is stored as facts and as concepts; it is measured (grosso-modo) in “number of concepts.” Ontology. It is a formal explicit specification of a shared conceptualization. [7] ♦ It is a taxonomy of the concepts I want to represent.9 See figure 1. 6 Symbol ♦ means: end of definition. Having a name in a shared language means that is known to many people. 7 We differ from Formal Concept Analysis (FCA) where a concept is any subset of attributes (or the corresponding set of objects) while we narrow concept to represent only an objects or thing that has a name in a natural language. 8 By “concrete internalization” we mean writing down (storing) the fact as a relation, for later use. 9 Each concept that I know and has a name is shared, since it was named by somebody else. More at §2.2. 5 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 6 of 24 Our definition. O = (C, R, root), an ontology is a structure formed by concepts and relations among concepts. Each concept that represents a set (such as the concept ani- mal) necessarily holds the relation subset with some other set, except the distin- guished concept root, which is subset of nobody, and which is superset of all other sets. Each concept that represents an individual must be in relation member of with some set-representing concept. ♦ Each concept and relation has as name (or has associ- ated) words from a natural language, since the ontology tries to represent a part of the real world. The required presence of relations subset and member of give an ontology a tree- like or taxonomy-like appearance; nevertheless, a set (apple) may be subset of more than one sets, such as fruit and food, and a person may be member of several sets. Word(s) associated with a concept: w(cA, OA) = the words associated in OA to concept cA. Example: (figure 1): w(tool, OP) = (tool, hardware). 2. Degree of know ledge about a concept; amount of know ledge of an agent How much does an agent know? In this section, we measure the amount of knowledge of an agent against the “total ontology” (an impractical abstraction). Later (§3.2), we meas- ure its knowledge relative to the knowledge of another agent. The idea is to know first how much an agent knows about a given concept, and then sum these amounts over all its known concepts. Total ontology. It is the ontology OK of an agent K that knows everything known to every agent. ♦ It is the union of all the ontologies of all the intelligent creatures, people and agents. Observations: 1. The total ontology is finite, and grows every day. 2. It is unique. Idea: think of a huge encyclopedia. 3. It contains only facts. To comply with 2 and 3 above, when trying to merge contradicting facts, K must dis- cover what “fact” is a lie. K could keep prevailing knowledge, but this can produce widespread knowledge that is nevertheless wrong, such as (Earth-World shape- form flat-plane). Truth is not discovered by majority voting. Since it is not the purpose of this article to unveil a method to find facts, we simply pos- tulate that the total ontology of K is formed with the help of a god or Very Wise Creature (VWC) that makes sure lies coming from some person or agent do not enter it. Thus, OK contains only facts,10 while an agent or person can still have parts of its/his/her ontology formed by lies. OK holds all known and “right” knowledge. The degree of knowledge (dk, §2.1) of an agent is measured against OK. 10 This total ontology grows every day, as new discoveries are performed. Also, notice that the ontology of a VWC must be larger or at least equal than the total ontology. Nevertheless, a VWC can not easily communicate its “ex- cess knowledge” to any person or agent A, because A will not understand some trios coming from VWC. Neverthe- less, VWC can (sequentially) teach this new knowledge to A, cf. §3.3.1. In most cultures’ cosmogony, there is a god that did just that. Such god is a VWC. In fact, we can define a VWC as an agent who has the total ontology. K=VWC ♦ 6 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 7 of 24 2.1 Degree of knowledge of an agent about a concept This degree is a measure of the (imperfect) grasp of a concept by an agent A. The idea is that the more relations that concept has in OA, the larger such degree of knowledge is. The degree of knowledge dkA(c) of A about a concept c is a number between 0 and 1, ob- tained by counting the number of arcs connecting to/from c in OA, adding the number of arcs labeled c in OA,11 and dividing into the similar calculation for c in the total on- tology OK. ♦ The closer dkA(c) is to 1, the less imperfect is A’s knowledge of c. 2 . 1 . 1 A mo u n t o f k n o w l e d g e o f a n a g e n t The idea is that every concept in OA contributes to the amount of knowledge of A. The amount of knowledge an agent has = Σ dkA(ci) over all ci∈OA. ♦ It is measured against the total ontology. It is approximated by the area under the histogram of con- cepts (see Clasitex in [8]). Similarly, the amount of knowledge of an agent in a disci- pline or area D⊂ OK is Σ dkA(ci) over all ci∈D. Definitions 2.1 “Degree of knowledge of an agent about a concept” and 2.1.1 “Amount of knowledge of an agent” are impractical since the total ontology is out of our reach. In their place, we shall define and compute, i.) (instead of 2.1) The degree of knowledge of an agent about a concept with respect to other agent’s knowledge of such concept, to be called in §3.1 the similarity value (sv). ii.) (instead of 2.1.1) The amount of knowledge of an agent with respect to another agent’s knowledge, to be called in §3.2 the degree of understanding (du) of A about B: how much A understands about what B knows. 2.2 Reification slightly increases knowledge Reification. The action of exposing the internal representation of a system in terms of pro- gramming entities that can be manipulated at runtime. The opposite process, absorption, consists of effecting the changes made to reified entries into the system, thus realizing the causal connection link [14]. ♦ To reify a (usually complex) concept or relation is to represent it by an atomic symbol (“to give it a name”), probably because we want to say more complex things about it; for instance, to affect it by other relations. Example: The complex concept “When Bill Clinton tries to convince a person about x, he explains and praises very much and with abundant examples the good things about x, but (without skipping any) explains lightly the bad things about x, thus giving the false appear- ance of an unbiased explainer”12 can be reified by giving it a name: clintonize (x). Then we can speak of clintonization, quasi-clintonization, and express clearly relations such as “clin- tonizing is immoral.” An ontology OA’=OA∪clintonize has one more concept than OA: the concept clintonize. A private concept, shared by few; not very useful for communications.13 In 11 This makes sense when c is a relation. 12 I believe this definition was introduced by President George Bush (Sr.) during a debate with then-candidate Bill Clinton. 13 An agent that reifies a concept could find this very useful for its own work (not for communication with others), much as an experienced locksmith invents a new tool by modifying other, in order to work faster: a private tool. 7 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 8 of 24 OQ our ontologies we will not cache (store) derived concepts [9], unless they are shared by many people (by having a name and a definition in a certain dictionary, say); that is, unless they are reified and given the same name by many people. thing (thing, something, entity) { eatable_thing (food, foodstuffs, groceries) { plant { vegetable [structure=living] {lettuce tomato onion potato (potatoes, potato, pome_de_terre) } fruit { strawberry banana } tree { lemon [color=green, shape=circle_shape] orange [color=orange_color] } cereal {rice[color=white] wheat[color=light_brown] }}} animal (animal, creature) [structure=living] { bird [covering=feather] { canary[size=small, covering=feather, color=yellow] } reptile[covering=scale] { lizard [size=medium, covering=scale, color=green] } mammal [covering=hair] {cat[covering=hair, size=medium, family=feline] dog [covering=hair, size=medium, family=canine]}} inanimate_thing (inanimate, object) { tool (tool,hardware) { manual (manual_tool) { wrench hammer } appliance (appliance, electrical) { drill [use=boring] electrical_saw (saw) [use=cutting] dishwasher [use=cleaning] blender [use=blend]}} furniture { bed table (table, surface) sofa (sofa, place) chair (chair, seat) } energy water } relation { color { white yellow purple green light_brown orange_color } shape (shape, contour) { circle_shape (circle, oval, round) square_shape (square ) } size { small medium } structure { living (living, alive) manmade (handcrafted, manmade) } covering { hair feather } family { feline canine } use { cleaning blend mix } | matter_state (matter_state) { liquid suspension gel solid gas plasma } } Figure 3. Ontology of person Q that buys in the supermarket of figure 2. It has 75 concepts. 3. Measuring the degree of understanding (comprehension) This section finds the concept cB∈OB most similar to cA∈OA, and how similar they are. From this, adding the similarities over all cB∈OB, we measure the degree of understanding of A with respect to OB. How well A knows the concepts in OB. 3.1 Finding the concept in your ontology most similar to one I have Consider two persons P and Q buying something specific at a real supermarket (www.superama.com.mx; its real ontology contains more than 500 concepts, we have cho- sen 150 of them for Ontology S, figure 2). Their ontologies are shown in figures 1 and 3. If 8 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 9 of 24 a person does not find exactly what he wants, he would like to know what is the most simi- lar item available. The sim algorithm (called “hallar (cA)” or COM in [19]) finds the most similar concept cB∈OB to concept cA∈OA. It also computes a similarity value sv∈[0, 1] expressing how similar was cB to cA. Agent A makes known concept cA to B by sending to B w(cA, OA) (words14 denoting cA), and also sending words w(pA, OA) denoting the father pA of cA. Four cases exist for cB = sim(cA).15 Case a) Node and father match. (Figures 4, 5) We look in OB for nodes pB and cB such that: (1) w(cB, OB)∩w(cA, OA)≠∅ (the intersection between the words associated to cB and those associated with cA is not empty); and (2) w(pB, OB)∩w(pA, OA)≠∅ (the intersection between words associated to pB and those associated to pA is not empty); and (3) pB is the father, grandfather or great-grandfather16 of cB. If such pB and cB are found, then cB is the answer and the algorithm returns sv = 1, too. Figure 4. Case a) Agent P (with ontology OP, figure 1) wants to find the most similar concept to cA = onion in OS (figure 2). Words (shown inside round parenthesis) from cA map to different concepts in B, but the most similar concept found (cB) is white_onion. We see that pA maps into pB, the grandfather of cB. sim does not compare concepts directly; it compares their words. 14 By §1, A can not send any node of OA to B. 15 Rigorously, sim is a function of two variables that returns two values, so it should be written (sv, cB) = sim (cA, OB). 16 If pB is found more than three levels up, the “semantic distance” is too high and sim says “no match.” B (ontology OS, supermarket) pA cA pB cB A (ontology OP) vegetable (vegetable) lemon (lemon) food (food) white_onion (onion)[COLOR=WHITE] onion (onion)[COLOR=WHITE] garlic_onion (garlic, onion) fruit_vegetable (fruit, vegetable) yellow_onion (onion)[COLOR=YELLOW] meat (meat) bulk_fruit (bulk fruit) (vegetable) (onion) 9 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 10 of 24 Figure 5. Case a) Father and son in OA find matches in OB. There are in OB three candidate nodes, the first that matches is the most similar concept, in this case white_onion. Case b) Father matches but node does not. Figures Error! Reference source not found. and Error! Reference source not found.. This case occurs when (2) of case (a) holds, but (1) and (3) do not. pB is found in OB but cB is not. In this case, sim is called recur- sively, and we try to compute pB' = sim(pA) to confirm that pB is the ancestor of cA, the concept of interest. (1) If the pB' found is thing, the root of OB, the algorithm returns not_found and concludes; sv = 0; (2) Otherwise, a certain child of pB, to be called cB', is searched in OB, such that: A. Most17 of the facts of cB' coincide with the corresponding facts of cA. Children of pB with just a few matching facts17 are rejected. If the candidate cB' analyzed has children, they are checked (using sim recursively) for a reasonable match17 with the children of cA. If a cB' is found with the desired properties, the algorithm re- ports success returning cB' as the concept in OB most similar to cA. Then sv = the fraction of facts of cB' coinciding with corresponding facts of cA. B. Otherwise cB' is sought among the sons of the father (in B) of pB; that is, among the brothers of pB; if necessary, among the sons of the sons of pB; that is, among the grandsons of pB. If found, the answer is cB'. sv = the sv returned by cB' multi- plied by 0.8 if cB' was found among the sons of pB,18 or by 0.82 = 0.64 if found among the grandsons of pB. C. If such cB' is not found, then the node nearest to cA is some son of pB, therefore sim returns the remark (son_of pB) and the algorithm concludes. sv = 0.5, an ar- bitrary but reasonable value. For example, in Figure 6, A sends words that corre- spond to the pair (cA = drill, pA = electrical), whereas B has the concept electrical but doesn't have drill nor any similar. In this case, the concept drillA is translated by B into (son_of electrical)B, which means “some electricalB I don't know” or “some electricalB I do not have in my on- tology. 17 We have found useful the threshold 0.5: more than half of the compared entities must coincide. 18 We have found that 0.8 allows for a fast decay as one moves up from father to grandfather and up. S P 10 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 11 of 24 Figure 6. Case b). Agent A (with ontology of person Q, figure 3) wants to find in ontology OS the concept most similar to cA = drill. Words from pA map to words from pB but cA has no equivalence in the ontology OS of B. Figure 7 shows the execution of sim for case (b)2(C). In this case concept drillA has no equivalent in B. Here son_of_electricalB is chosen from B as the most similar concept because parents electricalA and electricalB coincide and were confirmed. We assign sv = 0.5 Figure 7. Case b) The father of cA finds its corresponding pB, but cA does not find a matching cB Case c) The node matches but the father does not. This case occurs when (1) of case (a) holds but (2) and (3) do not. See figures 8 and 9. Concept cB is found but pB is not. We verify two conditions (A) and (B): (A) Most17 of the facts of cB should coincide (using sim) with those of cA ; and (B) Most of the children of cA should coincide (by sim) with most17 of the children of cB. cA PB A (ontology OQ) B (ontology OS supermarket) electrical (electrical) electrical_saw (saw) tool (tool, hardware) multicontact (multicontact) drill (drill) electrical (electrical) hardware (hardware) socket (socket) PA ? (electrical) (drill) Q S 11 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 12 of 24 (1) If the facts in (A) and the children in (B) coincide, the algorithm concludes with re- sponse cB, although it did not find the pB∈OB that corresponds to pA∈OA. Here sv = the fraction of facts and children of cB matching with corresponding entities of cA. (2) If even fewer properties and children are similar then response is (probably cB) and the algorithm finishes. Here sv is computed like in (1)B. (3) If neither properties nor children are similar, response is not_found and the algo- rithm finishes. sv = 0. Figure 8. Case c). Node matches but father does not. Agent A (with ontology OP of figure 1) wants to find a concept in B’s ontology OS corresponding to cA = lemon. Words from cA match with words from cB, but there is no equivalence for words from pA. Figure 9 shows an example of case (c)(2). In this case we use sim to seek in B the most similar concept to lemonA. Here concepts match but parents (vegetableA, bulk_vegetableB) do not (words are different for each parent), therefore similarity of the properties are used (calling recursively to sim). sv = 0.75 because parents do not coin- cide, and the answer is “probably lemon.” Case d) If neither cB nor pB are found, the algorithm concludes returning the response not_found. Figures 1 and 1. sv = 0. cA could not find a similar node in OB. The agents may have different ontologies (they know about different subjects) or they may not share a common cmmunication language. See figures 10 and 11. PB A (ontology OP) B (ontology OS supermarket) lemon (lemon)[COLOR=GREEN, SHAPE=SPHERIC_SHAPE, SIZE=3CM,TASTE=SOUR] bulk_vegetable (bulk_vegetable) perishable (perishable) PA vegetable (vegetable) lemon (lemon)[COLOR=GREEN, SHAPE=CIRCLE_SHAPE, TASTE=SOUR,SIZE=3CM] food (food) onion (onion)[COLOR=WHITE] ? CB CA (vegetable) (lemon) 12 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 13 of 24 Figure 9. Case c) Unable to find in B the father of cA , although cB corresponding to cA was found. Figure 10. Case d) Neither node nor father match. Agent A wants to find the most similar concept in B to cA = sofa. There are no words from cA that map into words of cB nor for pA and pB. Ontologies are those of fig- ures 2 (OS) and 3 (OP). Figure 11 shows the execution for case (d). Observe that A’s ontology OP fragment of interest is mainly about furniture while B’s ontology OS is mainly about groceries. There are some concepts in common, but not the involved concepts. sv = 0. Figure 11. Case d) Neither cA = bed nor pA = furniture find suitable matches in ontology S. cA A (ontology OP) B (ontology OS supermarket) pharmacy (pharmacy) thing (thing, something) perishable (perishable) PA cloth (cloth) ? ? sofa (sofa) bed (bed) (bed) (furniture) furniture (furniture) Self_service (self service) patent (patent) SP P S 13 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 14 of 24 Ph hepatitis type_A_hepatitis type_C_hepatitis type_V_hepatitis type_B_hepatitis hepatitis cirrosis John If cB is the concept most similar to cA, it is not necessarily true that cA is the concept most similar to cB. sim is not symmetric. Example: Refer to Figure 12. Levachkine [16] digs deeper into this. The function sim is only defined between a concept cA in OA and the most similar con- cept cB in OB; extensions sim’ amd sim’’ are in §3.1.2 below. 3 . 1 . 1 W h o r u n s sim? Who compares these two concepts, since they belong to different ontologies? That is, who runs sim? Either agent A or B can execute it, since sim compares words, not concepts. Nevertheless, when A runs sim, it needs the collaboration of B (and vice versa), which has to provide words to be used by sim (thus, by A). Also, even when A executes sim producing cB as result, A can not “have” or “see” cB: it is a pointer to the memory OB, a meaningless pointer for A, such as the tokens of point 4 of §1.1. What A can see of cB is: (1) the words which denote cB, as well as (the words for) the nodes related to cB; (2) corresponding words for the father, grandfather, sons... of cB (and words for their relations); (3) sv, indicating how similar that elusive cB is to its (very solid) cA. In fact, A still has cA as “the concept I have been thinking all along.” When A runs sim, B can see, of course, cB, but it can not “see” or “grasp” cA. The most of what B can see of cA is that “A wants to talk about some- thing of which the closest I have is cB”.19 B can sense from the words sent to it by A some differences between its solid cB and the elusive cA of A. More in §3.3. Figure 12. sim is not symmetric. Physician Ph knows four kinds of hepatitis, including the popular type_A_hepatitis and the rare type_V_hepatitis (viral), while John only knows hepatitis. The type_V_hepatitis of Ph (and all others) finds John’s hepatitis as “the most similar concept John has,” while John’s hepatitis best maps into Ph’s type_A_hepatitis. Ph knows more than John, so Ph can select a better target in his rich ontology for John’s vague concept. John can not make such selec- tion. 19 It will not help if B is more cooperative. For instance, dumping all its ontology OB into A’s memory will not help A, who will still see a tangled tree of meaningless pointers. Well, not totally meaningless—some understandable words are attached to each node. Yes: A can (patiently) understand (untangle) OB by comparing each of OB nodes with its own OA –that is, by using sim! 14 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 15 of 24 3 . 1 . 2 Ge n e r a l i z i n g s i m sim’ (cA, dA) for two concepts belonging to the same ontology, is defined as the 1/(1+length of the path going from cA to dA in the OA tree). ♦ sim’ (cA, dA) ∈ [0, 1]. sim’ is sym- metric. Example: see figure 14. 3.1.2.1 Relation to confusion In [16], the confusion conf(cA, dA) occurred by using cA instead of dA, is defined as the length of the descending20 path from cA to dA. ♦ This definition holds for hierarchies; it is here extended to ontologies. If we had defined sim’ (cA, dA) = (1/(1 + length of the descend- ing path going from cA to dA in the OA tree), we would have had sim’ (cA, dA) = 1/1+conf(cA, dA). We prefer, for ontologies, the definition of sim’ in §3.1.2, since it is symmetric, while conf is not. Example: for ontology D of figures 1 and 2, conf (liq- uid_food, food) = 0; the confusion when using liquid_food instead of food is 0, since liquid food is food. conf (food, liquid_food) = 1; when I want liquid food but I am given food, there is an error of 1 (a small error, you could think). For other concepts, we obtain the values in figure 13. Confusion and similarity for concepts x and y belonging to the same ontol- ogy OP of figure 1 conf (x, y); con- fusion in using x instead of y conf (y, x); confu- sion in using y instead of x sim’ (x,y) = sim’ (y,x); simi- larity between x and y x =plant, y =animal 1 1 1/3 x =plant, y =inanimate_thing 1 2 1/4 x =plant, y =cereal 1 0 1/2 x = fruit, y = tool 2 3 1/6 x = strawbery, y = furniture 2 4 1/7 x=thing, y=furniture 2 0 1/3 Figure 13. Examples of confusion and similarity (sim’) for two concepts of the same ontology OP, figure 1. conf is not symmetric; sim’ it is. For similarity between any two objects of different ontologies, we have: sim’’ (cA, dB) when dB is not the most similar concept in OB to cA, is found by making first s1 = sv returned by sim (cA) [this also finds cB, the object in OB most similar to cA]; then, find s2 = sim’ (dB, cB). Now, sim’’ (cA, dB) = s1s2. 3.2 Degree of understanding The value sv found in cB=sim(cA) in §3.1 can be thought of as the degree of under- standing that agent B has about concept cA. Each cA that forces B to answer sv=0 indicates that B has no idea (no concept) about this cA. We can add all these sv’s for every concept 20 Going towards more specialized concepts. “Using a person from Houston when I want to use a Texan person, confusion is 0; using a Texan person when a Houston person is needed causes confusion=1; using a US person causes confusion=2.” 15 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 16 of 24 cA∈OA and find the degree of understanding that agent B has about OA.21 It is as if A asks B, for each concept cA∈OA, «do you understand what is cA?» How much do you understand my cA? At the end, A has a good idea of the understanding of B (with respect to OA). The degree of understanding of B about OA, du(B, A) = {sum over all cA∈OA of sv re- turned by sim(cA)} / number of concepts in OA. ♦ [20] Similarly, we can measure the de- gree of understanding of B about some region of OA. du is not symmetric. In general, an agent understands some parts better than others. Notice that if OA is a large (approaching in size the total ontology of §1.2), then the degree of understanding of B with respect to OA approaches its amount of knowledge as given in §2.1.1. du(B, A)≤ 1; in the regions where B knows more than A, du= 1. Example: for person P of figure 1 and supermarket S (Fig. 2), the degree of understanding of P about PS is du(P,S)= 29.75 / 150 =0.2. See figures 14, 1 and 15. cS CP= sim (cS) sv cS CP= sim (cS) sv thing thing 1.0 purple purple 1.0 electrical electrical_tool 1.0 shape shape 1.0 socket son_of_electric al_tool 0.5 circle son_of_shape 0.5 multicon- tact son_of_electric al_tool 0.5 spheric_shape son_of_shape 0.5 meat meat 1.0 square_shape square_shape 1.0 pork pork 1.0 taste taste 1.0 beef beef 1.0 sweet sweet 1.0 turkey turkey 1.0 sour sour 1.0 veal veal 1.0 size size 1.0 white_onion onion 1.0 3cm 3cm 1.0 lemon probably lemon 0.75 matter_states matter_states 1.0 relation relation 1.0 liquid liquid 1.0 color color 1.0 suspension suspension 1.0 white white 1.0 gel gel 1.0 yellow yellow 1.0 solid solid 1.0 green green 1.0 gas gas 1.0 Figure 14. We are trying to assess du(P, S), the degree of understanding that person P (figure 1) has about S’s ontology (figure 2). Only rows where sv≠0 are shown. Columns 2 and 5 show the concept cA most similar to each cS∈OS; columns 3 and 6 show the corresponding similarity value sv. The degree of understanding du(P,S) is computed by adding over every cS∈OS the corresponding sv returned by cP=sim(cS) and dividing by 150 = number of concepts in OS. Thus, du(P, S)= (ΣSsv)/150= 29.75/150= 0.2. That is, P understands 20% of S. Results appear in Figure 15 21 B does not know how many concepts there are in OA, so it needs cooperation of A, for instance, when B asks A “give me the next concept in your ontology, please.” 16 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 17 of 24 Figure 15. Computing du(P,S) as detailed in figure 14 shows that P understands 20% of S. Since the size of S is larger than the size of P, probably du(S, P), the degree of under- standing of S about P, will be larger than du(P, S). “The more I know, the more I under- stand.” This is shown in figures 16 and 17. Figure 16. To compute du(S, P), the program adds all sv’s produced (see figure 18) by each answer cS=sim(cP) when every concept of P is visited. Thus, du(S, P) = 0.51. S understands 51% of P. S P S P 17 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 18 of 24 CP cS = sim (cp) sv CP cS = sim (cP) sv thing thing 1.0 orange_color son_of_relation 0.5 onion white_onion 1.0 shape shape 1.0 lemon probably lemon 0.7 5 circle_shape son_of_shape 0.5 meat meat 1.0 square_shape square_shape 1.0 pork pork 1.0 taste taste 1.0 beef beef 1.0 sweet sweet 1.0 turkey turkey 1.0 sour sour 1.0 veal veal 1.0 size size 1.0 tool hardware 1.0 3cm 3cm 1.0 wrench son_of_hardware 0.5 covering son_of_relation 0.5 electri- cal_tool electrical 1.0 structure son_of_relation 0.5 dish- washer son_of_electrical 0.5 contain son_of_relation 0.5 blender son_of_electrical 0.5 family son_of_relation 0.5 drill son_of_electrical 0.5 use son_of_relation 0.5 mixer son_of_electrical 0.5 matter_states matter_states 1.0 relation relation 1.0 liquid liquid 1.0 color color 1.0 suspension suspension 1.0 white white 1.0 gel gel 1.0 yellow yellow 1.0 solid solid 1.0 green green 1.0 gas gas 1.0 purple Purple 1.0 Figure 17. Now S is asked by P about each concept in OP. Each answer of S (second and fifth columns; only cs’s with sv≠0 are shown) produces a similarity value (third and last columns). Adding these numbers gives 34.75. Dividing into | P |=68, the degree of understanding of S with respect to P is found to be 0.51. We can also compute the “cross” knowledge among our two supermarket customers. Figures 18 and 19 show du(Q,P)=0.82. Figures 20 and 21 show that du(P, Q)=0.71. Thus, we can see that P knows more about Q than Q about P. Figure 18. The degree of understanding of Q about P is du(Q, P) = 55.5 / 68 = 0.82. Details in figure 20. P Q 18 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 19 of 24 CP CQ = sim (cP) sv CP CQ = sim (cP) sv thing thing 1.0 green green 1.0 food eatable_thing 1.0 purple purple 1.0 vegetable vegetable 1.0 orange_color orange_color 1.0 onion onion 1.0 shape shape 1.0 lemon probably lemon 0.5 circle_shape circle_shape 1.0 fruit fruit 1.0 square_shape square_shape 1.0 orange probably orange 0.5 taste son_of_relation 0.5 meat son_of_eatable_th ing 0.5 size size 1.0 seed cereal 1.0 3cm son_of_size 0.5 rice rice 1.0 covering covering 1.0 wheat wheat 1.0 hair hair 1.0 tool tool 1.0 feather feather 1.0 wrench wrench 1.0 paint son_of_covering 0.5 Electri- cal_tool appliance 1.0 structure structure 1.0 Dish- washer dishwasher 1.0 living living 1.0 blender blender 1.0 man_made man_made 1.0 drill drill 1.0 contain son_of_relation 0.5 mixer son_of_appliance 0.5 family family 1.0 furniture furniture 1.0 feline feline 1.0 bed bed 1.0 canine canine 1.0 table table 1.0 use use 1.0 sofa sofa 1.0 cleaning cleaning 1.0 bench son_of_furniture 0.5 blend blend 1.0 cat cat 1.0 mix mix 1.0 dog dog 1.0 matter_states matter_state 1.0 canary canary 1.0 liquid liquid 1.0 relation relation 1.0 suspension suspension 1.0 color color 1.0 gel gel 1.0 white white 1.0 solid solid 1.0 yellow yellow 1.0 gas gas 1.0 Figure 19. To compute du(Q, P), every cP∈P (first and fourth columns) is visited and the most similar cQ∈Q (second and fifth columns) is found, as well as its similarity value sv (third and sixth columns). Only rows where sv≠0 are shown. The sum of all these sv’s (55.5) divided by the number of concepts in P (68) gives du(Q, P) = 0.81 19 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 20 of 24 Figure 20. du(P, Q) = 54.1 / 75 = 0.71. P understands 72% of Q’s knowledge. Details in figure 21. 3.3 Can I understand the source of our disagreement? Can I mend it? §3.1.1 shows that A can not perceive or see cB directly. Given cA∈OA and its most similar concept cB∈OB, can A perceive in what way cB differs from its cA? After all, A knows from the value sv returned by sim(cA), how imperfect is the matching of cA to cB. The answer is yes, and it will be explained qualitatively. A can ask B about the facts involving cB. It will receive the answers in words. Then, A can proc- ess them (through sim, perhaps) to see how cA’s facts differ from those re- ceived. It can do the same with the father_of (cB), and with the sons_of (cB). And so on. Some words received will refer to concepts of which A is not sure (it has no knowledge of them, or there is ambiguity), so that more processing (Process P is called again) on these concepts is needed. Also, occasionally A will receive from B relations involving cB which A has as false for cA. 3 . 3 . 1 A n a g e n t l e a r n i n g f r o m a n o t h e r a g e n t A can collect all this information in a note N attached to cA: “N is what B knows about my concept cA, which differs from my knowledge”. It is like having A compute a belief: “B believes or knows N about cA, and N is not what I know about cA.” Or, perhaps, A can go ahead and decide to internalize N, that is, to “make it its knowledge”: to learn N about cA from OB. For this to happen, A needs to incorporate the relations in N in its OA,22 and to resolve the ambiguities and inconsistencies coming from N (some of N’s relations are known to A to be false; others make little sense to A). This has been solved for an agent teaching a person, but not yet for an agent teaching another agent. It can be done, perhaps, by using other knowledge services in the Web to referee disagreements between A and B and help A decide who is wrong about what (the “what” is already captured in N). We call this ontology merging: if A learns OB from B, its new OA will be its old OA merged (as in- formally described here) with OB. Cuevas’ Ph.D. Thesis [2] is along these lines. 22 This (to regard N as facts) is not difficult, once A decides to do it. Process P P Q 20 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 21 of 24 CP CQ = sim (cP) sv CP CQ = sim (cP) sv thing thing 1.0 purple purple 1.0 Eat- able_thing food 1.0 green green 1.0 plant son_of_food 0.5 light_brown son_of_color 0.5 vegetable vegetable 1.0 orange_color son_of_color 0.5 lettuce son_of_vegetable 0.5 shape shape 1.0 tomato son_of_vegetable 0.5 circle_shape circle_shape 1.0 onion onion 1.0 square_shape square_shape 1.0 potato son_of_vegetable 0.5 size size 1.0 lemon lemon 1.0 small son_of_size 0.5 cereal seed 1.0 medium son_of_size 0.5 rice rice 1.0 structure structure 1.0 wheat wheat 1.0 living living 1.0 canary probably canary 0.34 man_made man_made 1.0 cat probably cat 0.67 covering covering 1.0 dog probably dog 0.67 hair hair 1.0 tool tool 1.0 feather feather 1.0 manual son_of_tool 0.5 family family 1.0 appliance electrical_tool 1.0 feline feline 1.0 drill drill 1.0 canine canine 1.0 Electri- cal_saw son_of_electrical _tool 0.5 use use 1.0 dishwasher dishwasher 1.0 cleaning cleaning 1.0 blender blender 1.0 blend blend 1.0 furniture furniture 1.0 mix mix 1.0 bed bed 1.0 matter_state matter_states 1.0 table table 1.0 liquid liquid 1.0 sofa sofa 1.0 suspension suspension 1.0 chair sofa 1.0 gel gel 1.0 relation relation 1.0 solid solid 1.0 color color 1.0 gas gas 1.0 white white 1.0 plasma son_of_matter_s tates 0.5 yellow yellow 1.0 Fig. 21. To compute du(P, Q), concepts in Q (columns 2 y 5) most similar to each concept in P (columns 1 and 4), as well as the similarity value sv, are computed. Only rows with sv≠0 are shown. 3.4 Conclusions Methods are given to allow interaction and understanding between agents with different ontologies, so that there is no need to agree first on a standard set of concept definitions. Given a concept and associated words, a procedure for finding the most similar concept in another ontology is shown, with examples, as well as a measure of the degree of under- standing between two agents. It remains to test our methods with large, vastly different, or practical ontologies. Work reported is a step towards free will interactions4 among agents, instead of using canned interactions. Also, towards agents “strange to each other”3 that try to interact and 21 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 22 of 24 make sense of their utterances, opposing the current trend where only agents written by the same person or group, or following the same data exchange standards, can interact. Interaction through standards will still dominate the market for some time: it is easier to define and follow standards than to be “flexible, uncompromising and willing to try to un- derstand new concepts.” A standard ontology in a discipline is a good thing, although it feels rigid and archaic after a while.23 Nevertheless, knowledge can not be standardized, since each day more sprouts; standardization will always fall behind. It is preferable, I think, for me to be flexible and have general ways of trying to understand what you have to say (even new or unusual things), instead of forcing you to use a standard for concept- sharing with me (I already force you to use a shared communication language, perhaps a natural language, but that is unavoidable). Our work shows that a standard ontology for concept-sharing is not needed; it will be impossible in general, anyway. 3 . 4 . 1 S u g g e s t i o n s f o r f u r t h e r w o r k From text document to ontology. To help ontology merging (§3.3.1), write a converter that forms ontologies out of text documents. Notation to describe any complex concept. How do you describe complex concepts, such as “clintonize” of §2.2? Idea: express it as relations composed by relations. See [19] for a first try. Describe actions. Extend ontologies to make possible to describe a sequence of events [22]. Better notation for ontologies. � Tree notation (figure 1) is cumbersome, since only one “subset” relation is represented, and often a set S is partitioned in several partitions. Thus, a better notation could be: person {partition sex (=M : male_person) (=F : female_person) } {partition age (≤20 : young_person) (20< age ≤ 60 : adult_person) (>60 : old_person) } � Similarly, graphic notations make cumbersome to represent n-ary relations. � When char- acterizing the relations (as another branch of the ontology), you need to define types of par- titioning relations (sex, age…), or whether the partition is a “natural” one, like partition- ing vertebrate into fish, bird, reptile, batrachians and mammal. Agent interaction. Establish necessary or sufficient conditions for agent interaction that do not have a communication agreement, as mentioned in §1. 3.5 Acknowledgments Helpful discussions were held with Profs. Serguei Levachkine (CIC), Michael N. Huhns (U. of South Carolina), Alexander Gelbukh (CIC), and Hal Berghel (U. of Nevada Las Ve- gas). Work herein reported was partially supported by NSF-CONACYT Grant 32973-A and Project CGPI-IPN 2004442). Authors have SNI National Scientist Award (CONACYT) 23 Compare the UNESCO Catalog of Sciences (which is 30-years obsolete in Computer Science) with the ACM Computing Classification System, which is 2-years obsolete. 22 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 23 of 24 3.6 References 1. V. Alexandrov, S. Levachkine and A. Guzman. “Data Dynamical Structures for Image Treatment with Applications to Digital Cartography”. Book in preparation. 2. A. Cuevas-Rasgado. Automatic learning by an agent through ontology merging. Ph. D. Thesis (in preparation; in Spanish), CIC-IPN. 3. J. Everett, D. Bobrow, R. Stolle, R. Crouch, V. de Paiva, C Condoravdi, M van den Berg, and L Polyani. Making ontologies work for resolving redundancies across docu- ments. Comm. ACM 45 (2) (2002) 55-60. 4. K. Forbus, B. Falkenhainer and D. Gentner. The structure mapping engine: algorithms and examples. Artificial Intelligence 41 (1) (1989) 1-63. 5. A. Gelbukh, G. Sidorov and A. Guzman-Arenas. Use of a weighted document topic hierarchy for document classification. Text, Speech, Dialogue (Pilsen, Chech Republic, 1999)133-138. 6. A. Gelbukh, G. Sidorov, and A. Guzman-Arenas. Document comparison with a weighted topic hierarchy. DEXA-99, 10th International Conference on Database and Expert System applications, Workshop on Document Analysis and Understanding for Document Databases (Florence, Italy, 1999) 566-570. 7. T. Gruber. Toward Principles for the Design of Ontologies Used for Knowledge Shar- ing, in: Nicola Guarino and Roberto Poli (eds.), Formal Ontology in Conceptual Analy- sis and Knowledge Representation (Kluwer Academic Publishers 1993). 8. A. Guzman-Arenas. Finding the main themes in a Spanish document. Journal Expert Systems with Applications, 14 (1/2) (1998) 139-148. 9. A. Guzman-Arenas, J Olivares, A. Demetrio and C. Dominguez. Interaction of purpose- ful agents that use different ontologies. In: Lecture Notes in Artificial Intelligence 1793 (Springer 2000) 557-573. 10. A. Guzman-Arenas, C. Dominguez and J. Olivares. Reacting to unexpected events and communicating in spite of mixed ontologies In: Lecture Notes in Artificial Intelligence 2313 (Springer Heidelberg 2002) 377-386. 11. C. Holsapple and K. Joshi. A collaborative approach to ontology design. Comm. ACM 45 (2) (2002) 42-47. 12. M. N. Huhns, M. P. Singh and T. Ksiezyk. Global Information Management Via Local Autonomous Agents. In: M. N. Huhns and M. P. Singh (eds.): Readings in Agents (Morgan Kauffmann 1997). 13. H. Kim. (2002) Predicting how ontologies for the semantic web will evolve. Comm. ACM 45 (2) (2002) 48-54. 14. F. Kon, F. Costa, G. Blair and R. H. Campbell. The case for reflective middleware. Comm. ACM 45 (6) (2002) 33-38. 15. D. B. Lenat, R. V. Guha, K. Pittman, D. Pratt and M. Shepherd. Cyc: Toward Programs with Common Sense, Comm. of the ACM 33 (9) (1990) 30-49. 16. S. Levachkine and A. Guzman-Arenas. Hierarchies Measuring Qualitative Variables. In: Lecture Notes in Computer Science 2945 (Springer-Verlag 2004) 262-274. 17. S. Levachkine and A. Guzman-Arenas. Hierarchy as a new data type for qualitative variables. Submitted to Data and Knowledge Engineering. 23 of 24 Monday , April 04, 2005 Elsevier R ev ie w C op y Measuring comprehension between two agents 4-Apr-05 24 of 24 18. M. Montes-y-Gomez, A. Lopez-Lopez and A. Gelbukh. Information Retrieval with Conceptual Graph Matching. In: Lecture Notes in Computer Science 1873 (Springer- Verlag 2000) 312-321. 19. J. Olivares. An Interaction Model among Purposeful Agents, Mixed Ontologies and Unexpected Events. Ph. D. Thesis, CIC-IPN, 2002. (In Spanish) Available on line at http://www.jesusolivares.com/interaction/publica 20. Jesus M. Olivares-Ceja, Adolfo Guzman-Arenas. Concept similarity measures the un- derstanding between two agents. Accepted in NLDB 04. To appear in Lecture Notes in Computer Science (Springer 2004). 21. Jan Kupper, Horacio Saggion, et al. Multi-source in formation extraction and merging. Proc. IJCAI 03, 409-414 24 of 24 Monday , April 04, 2005 Elsevier 
work_i72n3bqvb5d5pesvagfykdppxq	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220972132 Params is empty 220972132 exception Params is empty 2021/04/06-03:05:12 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220972132 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:12 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_i6rnbhcflbchhbsvdprmugxlxy	----	Microsoft Word - Hierarchies 717 for dke 14sep04.doc S. Levachkine & A Guzman-Arenas 1 of 31 Hierarchy as a new data type for qualitative variables Serguei Levachkine, Adolfo Guzman-Arenas* Centro de Investigación en Computación, Instituto Politécnico Nacional “Lopez Mateos” Campus, Mexico City, MEXICO sergei@cic.ipn.mx, aguzman@ieee.org SUMMARY. Qualitative variables take symbolic values such as cat, orange, California, Africa. Often these values can be arranged in levels of deeper detail. For example, the vari- able place_of_birth takes as level-1 values Africa, Asia... as level-2 values Nigeria, Japan... as level-3 values California, Massachusetts... These values are organized in a hierarchy H, a mathematical construct among these values. Over H, the following are defined: (1) the function confusion resulting when using a symbolic value instead of another; (2) the close- ness to which object o fulfills predicate P; (3) a method which allows precision-controlled retrieval for relational databases whose objects have symbolic values. INDEX TERMS: hierarchy, ontology, approximate queries, confusion, knowledge represen- tation. 1. Introduction What is the capital of Germany? Berlin is the correct answer; Frankfurt is a close miss, Madrid a fair error, and sausage a gross error. What is closer to a cat, a dog or an orange? Can we measure these errors and similarities? Can we retrieve objects in a data base that are close to a desired item? Yes, by arranging these symbolic (that is, non-numeric) values in a hierarchy. For the sake of completeness, four different definitions of hierarchy are * Corresponding author. S. Levachkine & A Guzman-Arenas 2 of 31 given in §2. At least one of them is original. These definitions are important for under- standing what a hierarchy is. However, the confusion does not depend on particular defini- tion. 1. This arrangement allows the definition of confusion (§3.1) to measure the error when using one symbolic value associated with a node in a hierarchy in place of the (in- tended, correct) symbolic value associated with other node in the hierarchy. Variations of the definition are: a. When the values represent sets with some size, as population in {France, Italy, Spain, Sweden} we talk about percentage hierarchies (§3.1.1). b. When the values can be ordered, as temperature in {frigid, cold, warm, hot, burning} we talk about ordered hierarchies (§3.1.2). 2. Confusion is also defined (§3.3) for hierarchies whose nodes are associated with predi- cates (called variables in the paper) in addition to hierarchies whose nodes are associ- ated with values of a variable (item 1). 3. Confusion among values (item 1) and among variables (item 2) enables measurement of how close a given object o fulfils a given predicate P (§3.2.2), and we write Pε(o) for this measure. The main contributions are (1)-(3), pertaining to symbolic values. Of course, errors, dis- tances and approximate answers are well understood and developed for numerical values. The rest of the introduction discusses related work. Section 2 gives several definitions for hierarchies. Section 3.1 introduces confusion. Section 3.2 presents predicates on hierar- chies. Discussion of overall paper’s results is in Section 4 and conclusions form Section 5. S. Levachkine & A Guzman-Arenas 3 of 31 Related work. Artificial Intelligence, Natural Language and Knowledge Representation communities have been gauging the distance, proximity or “relatedness” between symbolic values. Relevant efforts: A. Hierarchies. The concept of a (generalization) hierarchy is not new. Hierarchies are used in data warehousing and data mining; see, for instance, the H-sets of Bhin [2]. A practical use of hierarchies in symbolic processing is Clasitex [9], which finds the themes of an article written in Spanish or English. It uses the concept tree, and a word (not in the tree) suggests the topic of one or more concepts in the tree. BiblioDigital© [4], a recent development, uses a large taxonomy (although not a hierarchy) to classify text documents; a (distributed) crawler in it retrieves “external” documents residing elsewhere in the Web. If a document is about (Cf. Clasitex) war, Iraq and President Bush, its URL will be stored in these three nodes in the concept tree. Hierarchies are simpler than ontologies, albeit very useful [13, 21]. The data modeling community, through the entity-relationship model, also organize items by their nature, properties and the relations among them. B. Natural Language. Linguists (see, for instance, Proceedings of CICLING 04, LNCS 2945, as referenced in [11]) have proposed many versions of semantic closeness, simi- larity, and other measures among words. Everett [6] identifies conceptually similar documents using a single ontology. Sidorov [8] does the same using a topic hierarchy: a kind of ontology. Montes y Gómez [19] builds trees of words, and by graph matching retrieves similar texts. Another common idea twisting around is to regard the represen- tation space with a “universal” measure of proximity of space’s elements and then an attempt to adapt it to different subject domains [16] [24]. Comments on this in §4. S. Levachkine & A Guzman-Arenas 4 of 31 WordNet [26] organizes information in logical groupings called synsets; each syn- set is a list of synonymous words or collocations (e.g., “fountain pen”, “take in”), and pointers that describe the relations between this synset and other synsets. A word or collocation may appear in more than one synset, and in more than one part of speech. The words in a synset are logically grouped such that they are interchangeable in some context. Nouns and verbs are organized into hierarchies based on the hy- pernymy/hyponymy relation between synsets. Two kinds of relations are represented by pointers: lexical and semantic. Lexical relations hold between word forms; semantic re- lations hold between word meanings. These relations include (but are not limited to), antonymy, entailment, and meronymy/holonymy. Additional pointers are used to indi- cate other relations. Budanitsky [3] compares five measures of similarity or semantic distance in Word- Net: Jiang and Conrath's measure (the best in the comparison: a spelling-corrector on real data); that of Hirst-St-Onge (seriously over-related), that of Resnik (seriously un- der-related [23]), and those of Lin [16] and of Leacock-Chodorow (in between). Note that all the measures except those of Hirst and St-Onge are similarity (not relatedness) measures considering only the hyponymy hierarchy of WordNet. The main problem (§4) with these approaches is that they use distances, thus obeying the symmetric prop- erty d(a,b) = d(b,a), while conf (§3.1) does not. C. Ontologies. At least three approaches appear when measuring similarity or relatedness of concepts (nodes in the ontology): S. Levachkine & A Guzman-Arenas 5 of 31 1. Syntactic approach. Methods that take into account only the organization of the tree or data structure of the ontology; for instance [19], those based on XML, or the “ontology merging” of Protégè [20]. 2. Standard ontology. Use of a common or agreed-upon ontology. Clearly, if dif- ferent people (or agents) use the same ontology, similarities among concepts will be consistently measured across users. CYC [7] was an early attempt to build the concept tree for common concepts. A common ontology is predicted in [11]; conceptually similar documents are identified in [6] by using a single ontology. In contrast, point (3) following shows use of different ontologies. 3. Measuring similarity across ontologies. LIA, a language for agent interaction [10, 13, 21], has an ontology comparator COM, that maps a concept from one ontology into the closest corresponding concept in another ontology. COM is used in sim of §3.4. By repeated use of sim, the degree of understanding du(B, OA) of agent B (with ontology OB) about ontology OA is found in [22]. Instead of using ontologies, this paper works on arbitrary hierarchies (§2). Why? Be- cause the problem-oriented interaction can be easier to maintain if the hierarchical structure is not a priori rigid as in the case of common hierarchies or ontologies. D. Pattern Classifiers. Our predicates with controlled precision or confusion (§3.2.1) are similar to Pattern Classifiers [18], but these classify objects according to the values of their properties, whereas hierarchies help to classify these values, when they are non- numeric. E. Distances and ultradistances. Traditionally [1, 25], the representation space is re- garded as a metric space with some “exotic” distance (e.g., ultrametric distance to S. Levachkine & A Guzman-Arenas 6 of 31 measure the “distances” between members of a hierarchy). Thus, §2.4 develops ul- trametric distances for hierarchies. However, often is not the case that such a distance meets the needs of the classification problem under consideration. Thus, we lean to- wards functions like conf (§3.1) that are not distances. 2. Theory This section continues with the focus on distances of item (E) of §1: we show how to build an ultradistance from a hierarchy (§2.4.1), how to build a hierarchy from an ultradistance (§2.4.2), whereas in Section 3 we move to a new approach that does not use distances. Element set E. A set whose elements are explicitly defined. ♦ 1 Example: {red, blue, white, black, pale}. Ordered set. An element set whose values are ordered by a < (“less than”) relation. ♦ Ex- ample: {very_cold, cold, warm, hot, very_hot}. Covering. K is a covering for set E if K is a set of subsets si ⊂ E, such that ∪ si = E. ♦ Every element of E is in some subset si ∈ K. If K is not a covering of E, we can make it so by adding a new sj to it, named “others”, that contains all other elements of E that do not belong to any of the previous si. Exclusive set. K is an exclusive set if si ∩ sj = ∅ , for every si, sj ∈ K. ♦ Its elements are mutually exclusive. If K is not an exclusive set, we can make it so by replacing every two overlapping si, sj ∈ K with three: si - sj, sj - si, and si ∩ sj. Partition. K is a partition of set E if it is both a covering for E and an exclusive set. ♦ 1 This symbol means: end of definition. S. Levachkine & A Guzman-Arenas 7 of 31 Symbolic value. A value that is not numerical, vector or quantitative. ♦ Example: red. Representation. A symbolic value v represents a set E, written v ∝ E, if v can be consid- ered a name or a depiction of E. ♦ v is associated with E. Example: strings ∝ {violin, viola, cello, guitar}. Qualitative variable. A single-valued variable that takes symbolic values. ♦ Its value can- not be a set,2 although such value may represent a set. father_of (v). In a tree, f is the father_of v if f is the node immediately following v in the path from v to the root. ♦ f is “the node from which v hangs.” We say that v is a son_of (f). ♦ Similarly, grand_father_of (v), brothers_of, aunt, ascendants, descendants... are defined, when they exist. ♦ The root is the only node that has no father. 2.1 Hierarchy Definition 1. A hierarchy H of an element set E is a tree whose root is E and if a node has sons then these form a partition of their father. ♦ Definition 2. For an element set E, a hierarchy H of E is a tree of nodes; each node n is either an element of E or a set of symbolic values vi, for i=1,…n, where vi ∝ Ei, and {E1, E2,…, En} is a partition of E. ♦ Example: for E = {chair table bed shirt loafer moccasin hammer paintbrush broom saw}, a hierarchy is (figure 1) H1 = {furniture ∝ {chair table bed} apparel ∝ {shirt shoe ∝ {loafer moccasin}} tool ∝ {hammer brush ∝ {paintbrush broom} saw}}. A hierarchy groups E into smaller sets of alike symbolic values. 2 Variable, attribute and property are used interchangeably. Some objects have an attribute (such as weight) while others do not: the weight of blue does not make sense, does not exist. A variable (color, height) describes an aspect of an object; its value (blue, 2 Kg) is such description (symbolic value) or measurement (numeric value). S. Levachkine & A Guzman-Arenas 8 of 31 merchandise furniture hammer tool chair table bed shirt shoe apparel brush saw loafer moccasin paintbrush broom Definition 1 emphasizes that the nodes of H are sets, while for definition 2 the nodes are symbolic values (such as furniture). To reconcile, we use the former definition v ∝ E, where a symbolic value v represents a set E. Some times, we add node “others” to certain level of a hierarchy, when we are not sure whether the nodes already present at that level will collectively exhaust their father. Thus, for instance, in Figure 1, we could add to the second level node “other_merchandise”, if we are not sure that furniture, apparel and tool comprise all the merchandise we are interested. A hierarchical variable is a single-valued qualitative variable whose values belong to a hierarchy. ♦ The data type of a hierarchical variable is hierarchy. Example: trades_in that takes values from H1, as trades_in = furniture, trades_in = broom. Fig. 1. Hierarchy H1 of articles for sale. 2.2 Partitions of a finite set Let E be a set of n elements. A partition P of E is a set of k subsets Ci of E such that (1) Ci∩Cj=∅ ; (2) ∪ iCi=E. ♦ Two elements x and y of E are equivalent in a partition P if they belong to the same class Ci; this is denoted by xPy. ♦ S. Levachkine & A Guzman-Arenas 9 of 31 Let P(E) be the set of all partitions of E; an order relation among the members of P(E), denoted by <, can be defined thus: for any two partitions P and P’, P<P’ iff xPy→xP’y. Partition P is said to be finer than P’; it has more classes than P’, i.e. k>k’. ♦ Example: let E={ a,b,c,d,e,f} . Then E is less fine (i.e. coarser) than {{ a} ,{ b,c,d} ,{ e,f}} which in turn is less fine than {{ a} ,{ b,c} ,{ d} ,{ e} ,{ f}} . A lattice structure for P(E) can be based on the order relation. For every pair of parti- tions P and P’ there is a least upper bound (l.u.b.) P∨ P’, and a greatest lower bound (g.l.b.) P∧ P’. ♦ Let us call Pk a partition of k classes where k is the level of Pk. A partition P’ is said to cover a partition P if and only if P’ results from combining two classes of P. ♦ Note that P’=bcd,a does not cover P=ab,c,d, because P’ cannot be obtained from the union of two classes of P, which would in fact give P’1=abc,d, P’2=abd,c and P’3=ab,cd, but not P. A chain in the lattice is a sequence of partitions in order, e.g. (P1, P2,...,Pj) where P1<P2 <...< Pj; the term is understood in the sense of an elementary chain in graph theory. ♦ If Pn is the finest partition, the height h(P) of P is the length of the chain joining P and Pn. ♦ If Pk is a partition of k classes, h(P)=n-k. It can be shown that all chains joining P and P’ have the same number of elements, equal to the difference h(P) – h(P’) between their heights, and (i) if P and Q each cover R, then P ∨ Q covers both P and Q; (ii) h(P) + h(P’) ≥ h(P ∨ P’) + h(P ∧ P’). S. Levachkine & A Guzman-Arenas 10 of 31 h1 h2 h3 h4 a b c d e f 2.3 Hierarchy, equivalent definitions Let E be a set of n elements, ℘ (E) the set of all subsets of E and P(E) the lattice of the partitions defined by the order relation P < Q. Let CH be a complete chain in the lattice, i.e. a chain linking the finest partition Pn, of n elements, to the coarsest partition P1=E. Now we can give two additional definitions of a hierarchy. Definition 3. A hierarchy is a set of partition classes constituting a complete chain, includ- ing in particular the set E itself and the n subsets formed by the elements of E. ♦ The passage from level k to level k-1 on CH corresponds to combining two classes. How- ever, several levels can be passed over. Let P and Q be two non-consecutive partitions on CH, so that the classes of Q are either those of P or combinations of two or more classes of P. This leads to another direct definition. Definition 4. A hierarchy is a subset H of ℘ (E) such that (1) E∈ H, (2) if x and y are ele- ments of E, then { x} ,{ y}∈ H, (3) if h and h’ are elements of H, then either h∩h’=∅ or h∩h’≠∅ , in which case either h⊂ h’ or h’⊂ h. ♦ Example: If E={ a,b,c,d,e,f} , then figure 2 represents the hierarchy formed by the subsets { a} ,{ b} ,{ c} ,{ d} ,{ e} ,{ f} with h1=E, h2={ a,b,c,d} , h3={ e,f} and h4={ a,b,c} Fig. 2. Representation of a hierarchy. 2.4 Ultrametrics and clustering A partial ordering of the elements of a hierarchy can be based on the inclusion relation and can be made a total ordering by the process of ascending a complete chain CH. In gen- eral, the same hierarchy can be defined by several different chains; thus in the example of S. Levachkine & A Guzman-Arenas 11 of 31 figure 2 we can use three chains CH1, CH2 and CH3, with their nodes numbered 0,1,2,3,4 as follows: CH1 a,b,c,d,e,f abc,d,e,f abc,d,ef abcd,ef abcdef CH2 a,b,c,d,e,f abc,d,e,f abcd,e,f abcd,ef abcdef CH3 a,b,c,d,e,f a,b,c,d,ef abc,d,ef abcd,ef abcdef 0 1 2 3 4 Two elements of E occur in the same subset at a given node of CH, this being a parti- tion of E. Given the chain, the node numbers characterize each pair of elements of E. We can now show how they can be used to define a special kind of distance. 2 . 4 . 1 U l t r a m e t r i c d i s t a n c e f r o m a h i e r a r c h y If i, j and k are three elements of a set E, the ultrametric distance δ is defined as a function of E×E in R+ as follows: • δ(i,i)=0, • δ(i,j)=δ(j,i), • δ(i,j) ≤ max[δ(i,k),δ(j,k)] ♦ (1). It is easy to prove the following theorem: every triangle is either isosceles or equilat- eral, with the base less than or equal to the equal sides. So we might define a distance between elements of E by means of a chain of partitions, and it will be clear that this is an ultrametric distance in the sense just defined. It will also be clear that infinity of ultrametric distances can be defined so as to be consistent with the order imposed by the chain CH, and we must remember that the same hierarchy can be specified by several different such chains. Conversely, an indexed hierarchy can be built (§2.4.2), given an ultrametric distance. S. Levachkine & A Guzman-Arenas 12 of 31 2 . 4 . 2 C o n s t r u c t i n g a h i e r a r c h y f r o m a n u l t r a m e t r i c d i s t a n c e An ultrametric space is a couple consisting of a set E, finite or otherwise, and an ul- trametric distance δ. In such a space, all triangles are isosceles or equilateral with the base less than or equal to the equal sides (see Theorem of §2.4.1). Conversely, if a generalized “distance” between any pair of elements of a set E is such that all triples give triangles hav- ing this property, this distance has the ultrametric property given by relation (1). Let B be a sphere with center i and radius r, j an interior point and k a point on the sur- face. By hypothesis δ(i,j)≤r, δ(i,k)=r. In the triangle (i,j,k) we must have δ(j,k)=r. This means that any internal point j is equidistant from all points on the surface, or that any in- ternal point can be regarded as the center of the sphere. Similarly, it can be shown that if two spheres have a point in common then one must be included in the other; this follows from taking the common point as center for each sphere. Let h denote the set of points of a sphere; all these points are equidistant from each other. If j is a point not belonging to h then, by δ(i,j)≤δ(j,k), it is equidistant from all points of h and this constant distance is greater than r. Thus h is an element of a hierarchy H de- duced from the ultrametric distance δ. We can use this result to define an algorithm for the construction of H from δ: Step 0. Set up the triangle table ∆ of ultrametric distances between the elements of E. Step 1. Find those elements hi,...,hk for which these distances are least (and equal, in the case of two elements only). Replace these in ∆ by hm. Recompute the distances between hm and the other elements hj: δ(hm,hm)=0, δ(hm,hj)= δ(hi,hj) Step 2. If ∆ has more than one column, repeat step 1, else end. S. Levachkine & A Guzman-Arenas 13 of 31 Example: Taking the example of figure 2, the successive tables (tables 1 to 5) are as fol- lows. Table 1 a b c d e f a 0 1 1 3 4 4 b 0 1 3 4 4 c 0 3 4 4 d 0 4 4 e 0 2 f 0 Table 2 h4 d e f h4 0 3 4 4 d 0 4 4 e 0 2 f 0 Table 3 h4 d h3 h4 0 3 4 d 0 4 h3 0 Table 4 h4 h3 h2 0 4 h3 0 Table 5 h1 h1 0 The hierarchy is the same as in Figure 2. A knowledge of ultrametric distances enables us to construct a hierarchy and therefore to form a sequence of partitions of decreasing fineness along a chain CH of the partition lattice, i.e. to perform a classification of variable fineness. What is now required is to find that ultrametric distance which meets the needs of the classification problem under con- S. Levachkine & A Guzman-Arenas 14 of 31 sideration. Since, in general, the data of a problem consist of distances in the ordinary sense, the requirement is to obtain an ultrametric distance from ordinary distance. 2 . 4 . 3 O b t a i n i n g a n u l t r a m e t r i c f r o m a m e t r i c Several algorithms that enable to us to derive an ultrametric which is close to the dis- tance in terms of which the problem data have been proposed by Lance and Williams [14] and others [15]. 2 . 4 . 4 D i s t a n c e s a n d l i n k a g e e f f e c t In this section we define the chain distance value that can lead to the linkage effect. Let E be a set of n elements and d(i,k) the distance between two elements. We define a path cjk as a sequence of elements (j,…,q,…,k) and Cjk as the set of paths cjk, and δ(cjk) = max q d(q,q+1). Thus δ is the length of the longest link in the path and we say that this defines the length of the path. Using the minimax criterion, in the set Cjk we find a path of shortest length, δ(j,k) say. This is an ultrametric distance, which we call the chain distance value. ♦ The use of a chain distance can lead to linkage effect, which appears in the nearest neighbor method of clustering (single linkage). 2 . 4 . 5 C l u s t e r i n g m e t h o d s The concepts of hierarchy, ultrametrics and clustering are closely linked. It is natural to place different objects into the same cluster according to a criterion of neighborhood. As this proceeds, the partitioning becomes progressively less fine and new objects h of a hierarchy H, are created [11]. Use of the nearest neighbor or single linkage criterion can sometimes be dangerous because of the linkage effect, and for this reason we S. Levachkine & A Guzman-Arenas 15 of 31 have to look for more satisfactory criteria which will take into account, among other things, the problem context under consideration. This is handled in next section. 2.5 Conclusions Distances and ultradistances help us to establish a partial or total order among elements x,y,z,u ∈ E, this is written (x,y) ≤ (z,u) meaning that x resembles y more than z resembles u. If the number of elements of E is large, the establishment of this order may be difficult, so that a practical manner to order them (at least partially) is to define a numerical function of similarity or dissimilarity (confusion) that can be computed in terms of the attributes of every element of E: the dissimilarity (confusion) λ(x,y) will be smaller the more closely x resembles y. Moreover, this function may not be a distance. That is developed in next Sec- tion. 3. Properties and functions on hierarchies I ask for furniture, and they bring me a chair. Is there an error? Now, I ask for a saw, and a tool is brought. Can we measure this error? Can we classify or organize these values? Yes, by using hierarchies (figure 1). In this section, we measure the error (called confusion) when one symbolic value is used instead of another (the intended or correct value). 3.1 Confusion in using r instead of s, for a hierarchy H If r, s ∈ H, then [12] the confusion in using r instead of s, written conf(r, s), is: • conf (r, r) = conf (r, any ascendant_of (r)) = 0. • conf (r, s) = 1 + conf (r, father_of(s)). ♦ 3 3 Without loss of generality we can define conf (r,s) = [1 + conf (r, father_of(s))] / height (H). In this case, 1 ≥ conf (r,s) ≥ 0 S. Levachkine & A Guzman-Arenas 16 of 31 LIVE BEING ANIMAL PLANT CITRIC pine MAMMAL orange snake lemon dog cat bird To measure conf, count the descending links from r (the replacing value) to s (the replaced or intended value). conf is not a distance, nor ultradistance. Example: conf(r, s) for the hi- erarchy H2 of figure 3 is given in Table 6. Fig. 3. A hierarchy H2 of living creatures. conf live being ani- mal plant mam mal bird snake citric pine cat dog lemon orange live being 0 1 1 2 2 2 2 2 3 3 3 3 ani- mal 0 0 1 1 1 1 2 2 2 2 3 3 plant 0 1 0 2 2 2 1 1 3 3 2 2 mam mal 0 0 1 0 1 1 2 2 1 1 3 3 bird 0 0 1 1 0 1 2 2 2 2 3 3 snake 0 0 1 1 1 0 2 2 2 2 3 3 citric 0 1 0 2 2 2 0 1 3 3 1 1 pine 0 1 0 2 2 2 1 0 3 3 2 2 cat 0 0 1 0 1 1 2 2 0 1 3 3 dog 0 0 1 0 1 1 2 2 1 0 3 3 lemon 0 1 0 2 2 2 0 1 3 3 0 1 orange 0 1 0 2 2 2 0 1 3 3 1 0 Table 6. Confusion in using row r instead of column s for the live beings of H2. s r S. Levachkine & A Guzman-Arenas 17 of 31 The confusion thus introduced catches the hierarchy semantics and resembles reality. For example, the error when using plant instead of live being is 0, since all plants are live beings. conf(plant, live_being) = 0: there is no error if they ask you for a live being and you give them a plant. Giving a live being when asked for a plant has error 1; conf (live_being, plant) = 1. The confusion among two brothers, such as orange and lemon, is 1. The confu- sion when using the father instead of the son is 1. Using a son instead of the father gives 0. conf is not a symmetric function. Using specific things instead of general things produces low errors, see the column animal. Using general things instead of specific things produces high errors, see the row live_being. Using any descendant of a node n instead of n produces no error: observe the 0’s in column plant. The lower triangular half has smaller errors than the upper triangular half of the table 4. 3 . 1 . 1 C o n f u s i o n c o n f b f o r h i e r a r c h i e s t h a t a r e f o r m e d b y b a g s Sometimes, the sizes of the sets that form the hierarchy matter. For example, for the hi- erarchy soldier ∝ {male_soldier, female_soldier}, our definition of conf would yield conf(soldier, male_soldier) = conf(soldier, female_soldier) = 1, when these numbers should be something like conf(soldier, male_soldier) = 0.02, conf(soldier, female_soldier) = 0.98, since approximately 98% of soldiers are male. A percentage hierarchy H of E is a hierarchy in which the number of elements of E in each set of H is known. ♦ The nodes of H are not sets, but bags: unordered collection where repetitions are allowed. 4 A loss of context appears if we use an ultrametric distance or any other symmetric function: these triangular parts would be equal. S. Levachkine & A Guzman-Arenas 18 of 31 For bags, the confusion in using r instead of s, confb(r, s) should take into account the rela- tive popularity of s in r, that is, confb(r, s) = 1 – relative proportion of s in r. This is made precise in the following definition. For percentage hierarchies, confb(r, s) = 1 – (|E∩r∩s| / |E∩r|). ♦ It is one minus the number of elements of E that are present in r∩s, divided by the number of elements in E present in r. The intersection ∩ preserves repeated elements. Example: For E = NorthAmerica(330M) ∝ {Canada(30M) ∝ {French_Canada(5M), Eng- lish_Canada(25M)}, USA(200M), Mexico(100M)}, where the population in millions is indicated, we show in table 7 the confusion confb. For instance, confb(Canada, NorthAme- rica) = 1 – (30M/30M) = 0; confb(NorthAmerica, Canada) = 1-(30M/330M)=1-.09 = 0.91. See table 7. confb N C U M Fr_C En_C N 0 0.91 0.39 0.7 0.99 0.93 C 0 0 1 1 0.63 0.17 U 0 1 0 1 1 1 M 0 1 1 0 1 1 Fr_C 0 0 1 1 0 1 En_C 0 0 1 1 1 0 Table 7. The confusion confb(r, s) using row r instead of column s is shown for bags NorthAmerica, Canada, USA and Mexico, represented as N, C, U, and M. 3 . 1 . 2 C o n f u s i o n c o n f L f o r h i e r a r c h i e s t h a t c o n t a i n l i s t s Sometimes, there is an order among the symbolic values that form a partition, as in H3 ={microscopic, tiny, small, medium, large, gigantic}. In this case, some nodes in the hier- archy are lists (ordered sets), not just sets. For these, the confusion among two brothers should not be 1 (as the ordinary definition of conf will say), but a number between 0 and 1 S. Levachkine & A Guzman-Arenas 19 of 31 related to the proximity of the two brothers. For a hierarchy composed of sets, some of which have an ordering relation, the confusion in using r instead of s, confL(r, s), is defined as follows: • confL(r, r) = confL(r, s) = 0, when s is any ascendant of r. • If r and s are brothers, confL(r, s) = 1 if the father is not an ordered set; else, confL(r, s) = the relative distance from r to s = the number of steps needed to jump from r to s in the ordering, divided by the cardinality-1 of the father. • confL(r, s) = 1 + confL(r, father_of(s)). ♦ Example: Refer to hierarchy H3. confL (micro- scopic, tiny)=1/5; confL(microscopic, small)=2/5; confL(microscopic, gigantic)=1. The rest of the paper will derive results for conf; those for confb and confL can be simi- larly derived. 3.2 The set of values that are equal to another, up to a given confusion A value u is equal to value v, within a given confusion ε, written u =ε v, iff conf(u, v) ≤ ε. ♦ v is the “correct” or intended value. It means that value u can be used instead of v, within error ε [12]. Example: Refer to figure 3. The set of values equal to CITRIC with confusion 0 is {CITRIC orange lemon}. The set of values equal to CITRIC with confu- sion 1 is {CITRIC orange lemon PLANT pine}. The set of values equal to cat with con- fusion 2 is {ANIMAL MAMMAL bird snake dog cat}. Notice that =ε is neither sym- metric nor transitive. S. Levachkine & A Guzman-Arenas 20 of 31 3 . 2 . 1 P r e d i c a t e s o n v a l u e s o f h i e r a r c h i c a l v a r i a b l e s An object can be characterized by several (variable, value) pairs, and some of the vari- ables may be hierarchical, such as (Sue (lives_in Canada) (trades_in furniture) (size small)). It is thus appropriate to define predicates that may hold for such objects. 5 In this section we extend the notion of predicate to “predicate that holds within a confu- sion”, by defining the set S of objects that satisfy predicate P within a given confusion ε. P holds for object o with confusion ε, or P holds for o within ε, iff • If P is formed by non-hierarchical variables, iff P is true for o. • For pr a hierarchical variable and P of the form (pr c), iff for value v of property pr in object o, v =ε c (if the value v can be used instead of c with confusion ε).6 • If P is of the form P1 ∨ P2, iff P1 holds for o within ε or P2 holds for o within ε. • If P is of the form P1 ∧ P2, iff P1 holds for o within ε and P2 holds for o within ε. • If P is of the form ¬ P1, iff P1 does not hold for o within ε. ♦ It is easy to see that if P1 holds for o within ε1 and P2 holds for o within ε2, then P1 ∨ P2 holds for o within min(ε1, ε2), whereas P1 ∧ P2 holds for o within max(ε1, ε2). Example (refer to hierarchies H1 and H2 above): Let the predicates be U = (trades_in furniture) ∨ (owns dog), V = (trades_in furniture) ∧ (owns dog), W = ¬ (trades_in apparel), X = ¬ (owns bird) 5 For variables that are not hierarchical, a match in value means conf = 0; a mismatch means conf = ∞. S. Levachkine & A Guzman-Arenas 21 of 31 and objects be (Carl (trades_in furniture) (owns snake)), (Don (trades_in apparel) (owns citric)), (Ed (trades_in shoe) (owns ANIMAL)), (Fred (trades_in table) (owns orange)), (Gal (trades_in tool) (owns PLANT)), (Hal (trades_in hammer) (owns dog)). Then we have the results of table 8. U holds within ε for: V holds within ε for: W holds within ε for: X holds within ε for: ε = 0 Carl, Fred, Hal (nobody) Carl, Fred, Gal, Hal (all) ε = 1 (all) Hal (nobody) Don, Fred, Gal ε = 2 (all) Carl, Ed, Hal (nobody) (nobody) Table 8. How the predicates U, V, W, X hold for several objects. 3 . 2 . 2 F u l f i l l m e n t o f a p r e d i c a t e b y a n o b j e c t In above example, how well an object such as Ed fulfils a predicate such as V? Looking at column V in table 8, V starts holding for o at ε=2. No smaller ε will do. Object o ε-fulfills predicate P at threshold ε, if ε is the smallest number for which P holds for o within ε. ♦ It is a non-negative integer defined between an object and a predicate. The closer is ε to 0, the “tighter” P holds for o. Compare with the membership function for fuzzy sets. 3 . 2 . 3 C o n f u s i o n b e t w e e n t wo o b j e c t s Similar to §3.1, for two objects (o (pr1 v1) (pr2 v2).. (prk vk)) and (o’ (pr1 v1’) (pr2 v2’)… (prl vl’)) with the same hierarchical variables, we can define the confusion when object o is 6 c is the intended or correct value, and it is assumed to be a value (a constant) in the hierarchy of pr. If c is a constant that does not belong to the hierarchy of pr, then (pr c) is always false. If c were not a constant, but another variable, so that P is of the form (pr var), for instance (size trades_in), [“I want the objects where size and trade_in have the same value”] the value is false unless pr and var take values on the same hierarchy. In this case, the value of var is the intended value. S. Levachkine & A Guzman-Arenas 22 of 31 used instead of o’ as the sum of the individual confusions conf(vi, vi’) between their respec- tive qualitative values. o’ is the intended or correct object. Confusion when object o is used instead of o’. CONF (o, o´) = Σi conf(vi, vi’). ♦ CONF (written in capital letters) is not symmetric, and is not a distance. 3.3 Confusion between variables (not values) that form a hierarchy What could be the error in “Carl is a relative of Mary,” if all we know is “Carl knows Mary”? And if what we know is “Carl is the grandfather of Mary”? That is, in addition to qualitative values forming a hierarchy, it is possible for the variables themselves to form a hierarchy. Example: hierarchy H4. handles ∝ {trades_in ∝ {buys sells rents} carries ∝ {transports_by_surface flies ships} maintains ∝ {oils fixes } } When considering H4 for (Carl (trades_in furniture) (owns snake)) of example § 3.2.1, he surely (handles furniture) with confusion 0, (rents furniture) with confusion 1, (rents merchandise) with confusion 1, and (rents table) with confusion 1+1=2. Thus, we can ex- tend the definition of predicate with confusion to have variables that are members of a hier- archy, by adding another bullet to the definition of §3.2.1, thus: • If P is of the form (var c), for var a variable member of a hierarchy, iff ∃ variable var2 for which (var2 c) holds for o within ε – conf (var2, var), where var2 also belongs to the hierarchy of var.♦ The confusion of the variables adds to the confusion of the values. S. Levachkine & A Guzman-Arenas 23 of 31 Example: For (Don (trades_in apparel) (owns citric)) of E2, predicate (handles apparel) holds with confusion 0; (trades_in apparel) holds with conf=0, (buys apparel) holds with conf=1, (buys shoe) holds with conf= 1+1=2, and (buys shoe) ∧ (owns lemon) with confu- sion = max(2, 1) = 2. Predicate (transports apparel) holds for Don with conf=1+1=2. 3.4 Confusion confdo for values in different ontologies It is possible that values v1 and v2 belong to hierarchies coming out of different element sets E1 and E2 expressed as hierarchies H1 and H2. If these hierarchies can be converted into ontologies (to be called O1 and O2), and thus map values v1 and v2 (v2 is the intended value) into objects (concepts) c1∈ O1 and c2∈ O2, we can use function sim of [13] to find confdo(v1, v2). Function sim(c1, O1, O2) finds sv, a number between 0 and 1 expressing the similarity between a concept c1 in O1 and its most similar concept c’2∈ O2. It also finds such c’2. Thus, to find confdo(v1, v2), execute these steps: 1. Find c’2=sim(c1, O1, O2), the concept c’2 most similar to c1, as well as sv. 2. Since c2 and c’2 belong to the same ontology O2, find l = length of the path going from c2 to c’2 in the OB tree. 3. confdo(v1, v2) = (1+l)/sv.♦ The steps are depicted in figure 4. See also [21]. S. Levachkine & A Guzman-Arenas 24 of 31 v1 c1 v2 c2 c'2 O1 (0) (0) (1) l (2) O2 Figure 4. Steps to find confdo(v1, v2) for values in different ontologies: (0) map v1, v2 into c1, c2; (1) find c’2 and sv; (2) find l; (3) (not shown) compute (1+l)/sv. 4. Discussion Starting the discussion, let us look again at the ordering and similarity of the elements of finite sets. A reflexive and transitive relation can be defined for the set F of all pairs of elements in E x E, which is called a partial order. With x,y,z,u ∈ E, this is written (x,y) ≤ (z,u) meaning that x resembles y more than z resembles u. Such a relation does not neces- sarily apply over the complete set F because they may be pairs of elements that are really not comparable. If it does apply over the complete set it becomes a total order. It is difficult in practice to set up such a partial order if the number of elements of E is large, and if it is possible it is difficult to make this order without running the risk of gener- ating contradictions. In fact, the only practical way to establish a partial order is to define a numerical function of similarity or dissimilarity (confusion) that can be computed in terms of the attributes of every element of E: the dissimilarity (confusion) λ(x,y) will be smaller the more closely x resembles y. The same partial order can be generated by any of an unlimited number such functions. Some dissimilarity functions, however, may not be distances (Cf. §3). Thus, by definition, a S. Levachkine & A Guzman-Arenas 25 of 31 distance d(x,y) satisfies: (a) d(x,x) = 0; (b) d(x,y) = d(y,x); (c) d(x,y) ≤ d(y,z) + d(x,z). Re- lation (a) can be satisfied by making min λ = 0, which is simply a change of origin, and for (b) it is sometimes necessary to make λ symmetrical. Relation (c) may not hold, however, although it can be made to hold by adding a sufficiently large constant λ0 to the values, whilst retaining λ(x,x) = 0. The corresponding partial ordering will not be changed. Thus the following general statement can be made: it is always possible for a finite set E to make simple modifications that will transform a dissimilarity into a distance measure without affecting the partial order. In early approaches [25], the measures of similarity between variables have been con- sidered using Kendall, Hamming, Russell & Rao, Jaccard, Kuzlinsky, Yule, and other dis- tances, and a contingency table. These result, however, insufficient for variables that take symbolic values. In context of the identification problem, distance measures for objects and concepts have been proposed too. The idea of classification carries with it in the implication that a descriptor or symbolic description can be defined for each class and will in some sense be representative of the class; possible descriptors are a symbolic description of the class; a feature represented as a relational structure. If X is the representation of an object, we can define a concept Ai as an entity which is such that an ordering of the couples (X,Ai) can be established for i = 1,...,k. A concept is not necessarily associated with one particular class; this may be so, e.g. in the case of descriptors, but in other cases the concepts may overlap, as for example for probabilities. The “distance” D(X,Ai) between an object and a concept can be used effectively as a measure of similarity or dissimilarity. In other words, the “ob- ject-concept” distance D is used as a characteristic function. The following functions have S. Levachkine & A Guzman-Arenas 26 of 31 been used: probability, fuzzy assignment, inertia and potential, nearest neighbors (q-NN), etc. The main problem with these approaches is that they often (always) omit the context of a classification problem under consideration. Let f(X;A) be a measure of similarity or dissimilarity between the representation of an object X and a concept A. Such a measure limited to a single A would be of little interest; if for example, A were the symbolic description of a class, there would clearly be at least two classes, A and ¬ A (“not-A”). However, it can happen that the concepts A do not coincide with the classes. Let {Ai} be the set of concepts under consideration: the assumption that these concepts can be grouped together into a set implies that the set operations of union (∪ ) and intersection (∩) can be applied. The Ai are not necessarily disjoint, i.e. several can apply simultaneously to a single object, although in the very special case that each identi- fies one class they are clearly disjoint. Let @ be the algebra generated by the Ai and opera- tions: ∩, ∪ , and ¬ ; the elements of @ form the interpretation space. If the concepts Ai are expressed by predicates, the operations are written as ∧ and ∨ respectively and we have Stone’s theorem: all distributive logic systems are homomorphic to a distributive lattice of subsets of a set. We may recall that if p,q and r are predicates then, ∧ is distributive with respect to ∨ . Given f(X;A) and f(X;B), the fundamental question is: what are the values of f(X;A∩B) and f(X;A∪ B)? This can be answered in terms of two laws or operations: (i) an additive law ⊕ , homomorphic with ∪ , and (ii) a multiplicative law ⊗ , homomorphic with ∩ and distributive over ⊕ , and the answer is [15, 25] (Cf. §3.2.1) f(X; A∪ B) = f(X;A) ⊕ f(X;B), S. Levachkine & A Guzman-Arenas 27 of 31 f(X;A∩B) = f(X;A) ⊗ f(X;B) (2). Union and intersection of ontologies and corresponding measures for them will be con- sidered in a forthcoming work. 4.1 Summing up We can emphasize in this summary the following key points: • The context of a problem under consideration may be lost when attempting to define a distance on hierarchies of symbolic values (to measure closeness between hierarchical elements) that holds its partial (total) order. • We show (§3.1) a way that takes into account the problem context, represents the in- trinsic data semantics and expresses the similarity (dissimilarity) function in terms of the data attributes. • Such approach permits to know the set of values that are equal to another up to a given confusion (§3.2), how close an object fulfils a predicate, and how close an object is to another (§3.2.3). • Confusions and similarity functions for values in different ontologies can be defined as well (§3.4). 4.2 Applications A) To queries, either to retrieve objects that hold for a predicate to a given threshold [5], or to measure the closeness of an object to a certain predicate (definition of §3.2.2). B) To handle partial knowledge. Even if we only know that Carl trades in furniture, we can productively use this value in precise searches (example of §3.2.1). S. Levachkine & A Guzman-Arenas 28 of 31 C) As an approximation to the manner in which people use gradation of symbolic values (ordered sets), to provide less than crisp, but useful, answers. D) As a measure of the confusion between two values (§3.1), such as in the answer “the capital of Germany is Frankfurt” (Cf. Introduction), versus “the capital of Germany is sausage.” E) To compare attributes (as opposed to values) that are similar, but not equal, such as my_neighbor and my_acquaintance (Cf. §3.3). F) As an alternative to fuzzy sets, using Pε(o) (§3.2.2) as the membership function of a set. G) As a supervised pattern classifier, by using CONF(o, o’) of §3.2.3 between two ob- jects. See also [18] for another approach to classifiers of qualitative data. H) As an alternative to data mining, where approximate answers are usually useful. I) As an accelerator of applications (G) and (H), if we store in cache memory the queries and results of previous tasks, and compare the (predicates of the) new queries to these cached data, before embarking in new searches [17]. 5. Conclusions and suggestions for further work 5.1 Conclusions The notions of hierarchy and hierarchical variable make possible to measure the confu- sion when a value is used instead of another. This creates a natural generalization for predi- cates and queries. The notions were introduced and developed for hierarchies formed by sets, but they can be extended to bags and lists, too. S. Levachkine & A Guzman-Arenas 29 of 31 The concepts and examples given here have practical applications, since they mimic the manner in which people process symbolic values. In a subsequent paper we will describe a mathematical apparatus and further properties of functions defined in §3. See also [13, 22] for ontologies instead of hierarchies. 5.2 Suggestions for further work 1. Extend definition 3.2 (the sets of values equal to another value) to fuzzy sets. 2. Extend definition 3.3 (confusion between variables, not values) to fuzzy sets. 3. Extend queries of 3.2.2 to queries (search) in text, with the help of Clasitex [9]. 4. Idem to queries in maps. 5. Idem to queries in semi-structured data. Acknowledgments The advice of the referees was very useful. Helpful discussions were held with Prof. Victor Alexandrov, SPIIRAS-Russia, Dr. Jesus Olivares and Prof. Gilberto Martinez, CIC- IPN. Work herein described was partially supported by NSF-CONACYT Grant 32973-A and Project CGPI-IPN 20010778. The authors have a SNI National Scientist Award. References 1. V. Alexandrov. Developed systems in science, technique, society and culture. (Saint Petersburg State Technical University, Russia, 2000). In Russian. 2. N.T. Bhin, A.M. Tjoa, and R. Wagner. Conceptual Multidimensional data model based on meta-cube. Lecture Notes in Computer Science 1909 (Springer-Verlag, 2000) 24-31 3. A. Budanitsky and G. Hirst. Semantic Distance in WordNet: An Experimental, Application-oriented Evaluation of Five Measures. In: Proc. North American Chapter of the Association for Computational Linguistics (NAACL-2000), Pittsburgh, PA. 4. V.P, de Gyves and A. Guzman-Arenas. A distributed digital text accessing and acquisition system. BiblioDigital.  SoftwarePro International. In: Lecture Notes in Computer Science 3061 (Springer-Verlag 2004) 274-283. S. Levachkine & A Guzman-Arenas 30 of 31 5. V.P. de Gyves. Precision-controlled retrieval of objects with symbolic values, in a relational database. B. Sc. Thesis in preparation, UPIICSA-Instituto Politecnico Nacional. In Spanish. 6. J. Everett, D Bobrow, et al. Making ontologies work for resolving redundancies across documents. Comm. ACM 45 (2) (2002) 55-60. 7. D. B. Lenat and R. V. Guha. Building Large Knowledge-Based Systems. (Addison- Wesley 1989). 8. A. Gelbukh, G. Sidorov, and A. Guzman-Arenas. Document comparison with a weighted topic hierarchy, In: Proc. DEXA-99, 10th International Conference on Database and Expert System applications, Workshop on Document Analysis and Understanding for Document Databases (Florence, 1999) 566-570. 9. A. Guzman. (1998) Finding the main themes in a Spanish document. Journal Expert Systems with Applications 14 (1/2) (1998) 139-148. 10. A. Guzman, J. Olivares, A. Demetrio and C. Dominguez. Interaction of Purposeful Agents that use Different Ontologies. In: Lecture Notes in Artificial Intelligence 1793 (Springer-Verlag 2000) 557-573. 11. A. Guzman and S. Levachkine. Graduated errors in approximate queries using hie- rarchies and ordered sets. In: Lecture Notes in Artificial Intelligence 2972 (Springer- Verlag 2004) 139-148. 12. S. Levachkine and A. Guzman-Arenas. Hierarchies Measuring Qualitative Variables. In: Lecture Notes in Computer Science 2945 (Springer-Verlag 2004) 262-274. 13. A. Guzman and J. Olivares. Finding the Most Similar Concepts in two Different Onto- logies. Lecture Notes in Artificial Intelligence 2972 (Springer-Verlag 2004) 129-138. 14. G.N. Lance and W.T. Williams. Mixed-Data Classificatory Programs I – Agglomerative Systems. Australian Computer Journal 1 (1) (1967) 15-20. 15. S. Levachkine. Qualitative data conversion and representation. MIEM (Moscow Institute of Electronics and Mathematics, 1994) Report (in Russian). 16. D. Lin. An Information-Theoretic Definition of Similarity. Proc. 15th Int. Conf. on Machine Learning (ICML 1998) Madison, Wisconsin. 17. G. Martinez-Luna. Incremental data mining. Ph. D. Thesis in preparation. Centro de Investigacion en Computacion, Instituto Politecnico Nacional. In Spanish. 18. F. Martinez-T and A. Guzman. The logical combinatorial approach to Pattern Recognition, an overview through selected works. Pattern Recognition 34 (2001) 741- 751. 19. M. Montes-y-Gómez, A. Lopez-Lopez, and A. Gelbukh. Information Retrieval with Conceptual Graph Matching. In: Lecture Notes in Computer Science 1873 (Springer- Verlag 2000) 312-321. 20. N. Noy, R.W. Fergerson and M.A. Musen. The knowledge model of Protégè-2000: combining interoperability and flexibility. Stanford Medical Informatics Technical Report, Stanford University, 2000 21. J. Olivares. A model of interaction among purposeful agents with mixed ontologies and unexpected events. Ph. D. Thesis. Centro de Investigacion en Computacion, Instituto Politecnico Nacional, 2002. In Spanish. Available on line at http://www.jesusolivares.com/interaction/publica S. Levachkine & A Guzman-Arenas 31 of 31 22. J. Olivares and A. Guzman-Arenas. Concept similarity measures the understanding between two agents. In Proceedings of NLDB 04. (2004) 23. P. Resnik. Disambiguating Noun Groupings with respect to WordNet Senses. In: Armstrong, S. et al. (eds.), Natural Language Processing Using Very Large Corpora. (Kluwer Academic Publishing, Dordrecht, 1995) 77-98. 24. P. Resnik. (1999) Semantic Similarity in a Taxonomy: An Information-Based Measure and its Application to Problems of Ambiguity in Natural Language. Journal of Artificial Intelligence Research 11 (1999) 95-130. 25. J-C Simon. Patterns and operators. The foundations of data representation. (McGraw- Hill, 1984 26. WordNet: A Lexical Database for the English Language http:// www.cogsci.princeton.edu/~wn/ 
work_i5reu7kixjhdzcgs5bjshjzdai	----	Document downloaded from: This paper must be cited as: The final publication is available at Copyright http://dx.doi.org/10.1016/j.eswa.2012.12.087 http://hdl.handle.net/10251/34934 Elsevier Navarro Llácer, M.; Heras Barberá, SM.; Botti Navarro, VJ.; Julian Inglada, VJ. (2013). Towards real-time agreements. Expert Systems with Applications. 40(10):3906-3917. doi:10.1016/j.eswa.2012.12.087. Towards Real-Time Agreements Mart́ı Navarro and Stella Heras and Vicente Botti and Vicente Julián Departamento de Sistemas Informáticos y Computación Universitat Politècnica de València Camino de Vera s/n 46022 Valencia (Spain) Abstract In this paper, we deal with the problem of real-time coordination with the more gen- eral approach of reaching real-time agreements in MAS. Concretely, this work proposes a real-time argumentation framework in an attempt to provide agents with the ability of engaging in argumentative dialogues and come with a solution for their underlying agree- ment process within a bounded period of time. The framework has been implemented and evaluated in the domain of a customer support application. Concretely, we consider a society of agents that act on behalf of a group of technicians that must solve problems in a Technology Management Centre (TMC) within a bounded time. This centre controls every process implicated in the provision of technological and customer support services to private or public organisations by means of a call centre. The contract signed between the TCM and the customer establishes penalties if the specified time is exceeded. Keywords: Agreement Technologies, Real-Time, Multi-Agent Systems 1. Motivation Many Multi-Agent Systems (MAS) operate in resource and time-constrained environ- ments. In such domains, scarce resources must be shared between di↵erent agents taking into account that these agents must perform their tasks before a deadline is met. In ad- dition, agents in open MAS are autonomous entities that can have their own knowledge resources, can play di↵erent roles and can have di↵erent objectives and preferences over values that they want to promote with their actions (e.g. an interested agent could want Email address: mnavarro@dsic.upv.es, sheras@dsic.upv.es (Mart́ı Navarro and Stella Heras and Vicente Botti and Vicente Julián) Preprint submitted to Expert Systems with Applications November 30, 2012 ���������� ������������������������������������ to promote its own wealth in spite of the fairness in a negotiation to allocate a scarce re- source). Also, they can be linked by di↵erent types of dependency relations in what we call an agent society (Heras et al., 2012). The high dynamism of the application domains of open MAS requires agents to have a way of harmonising the conflicts that come out when they have to collaborate or coordinate their activities. On top of the simpler ability to interact, agents need a mechanism to argue (persuade other agents to accept their points of view, negotiating the terms of a contract, etc.) and reach agreements (Sierra et al., 2011) to collaborate and coordinate their activities. Furthermore, if these agents operate in real-time environments, the appropriate agreement mechanism that they should use will have the added difficulty of allowing agents to reach agreements within a specified time. Therefore, in these environments there is a need for coordinating the agents that make use of the available resources and reach agreements in a real-time fashion. Methods for the development of real-time MAS have been proposed in the literature (V. Julián and V. Botti, 2004). Also, real-time coordination in MAS has been stud- ied from di↵erent perspectives, as robotics (O. Kahtib, 1986), traffic management (Choy et al., 2003) or route planning (Navarro et al., 2012). When the final objective of the coordination process is to reach an agreement, to coordinate agents by engaging in argu- mentation dialogues with their opponents in a discussion is a common approach (Kraus et al., 1998)(Parsons et al., 1998)(Amgoud and Prade, 2004). However, as the social context of agents determines the way in which agents can argue and reach agreements, this context should have a decisive influence in the computational representation of ar- guments, in the argument management process and in the way agents develop strategies to argue with other agents. To deal with this challenge, we have recently proposed a case-based argumentation framework that allows agents to argue in agent societies, tak- ing into account their roles, preferences over values and dependency relations. This work also proposes a reasoning process by which agents can automatically generate, select and evaluate arguments in an agent society and learn from argumentation experiences to be able to develop dialogue strategies that help them to reach their objectives more efficiently (Heras et al., 2013b). Nevertheless, our original reasoning process didn’t take real-time considerations into account to bound the time that agents can spend to reach agreements 2 and hence, cannot be applied in real-time scenarios. In this paper, we deal with the problem of real-time coordination with the more general approach of reaching real-time agreements in MAS. Therefore, our notion of a real-time agreement may refer to the fair allocation of a scarce resource, but also to the outcome of a negotiation process to make a deal in a sale operation, to the final label assigned in a classification problem, to the final state of the beliefs database of an agent in a beliefs revision problem, and many more. However, any of these processes must be temporal bounded. This is an important aspect to consider in any real-time argumentative process, where a community of agents are able to reach agreements through argumentation before a deadline is met. Any agreement reached after this deadline will be considered of a low quality, or even not valid if we are in a hard real-time environment. Therefore, it is necessary to control every step of the argumentative process in order to predict its duration in time (i.e. we must know the temporal cost of executing each step of the process). Taking into account this prediction, we can guarantee that the agents engaged in the argumentation dialogue will reach an agreement about the outcome of the agreement process (e.g. they will decide the final allocation for a resource, the label of an individual in a classification problem, etc.) before the specified deadline is met. The response provided by the agents does not need to be optimal, since in some cases the best response may require more processing time, but at least it will be provided on time. This work proposes a real-time argumentation framework in an attempt to provide agents with the ability of engaging in argumentative dialogues and come with a solution for their underlying agreement process within a bounded period of time. The structure of this paper is as follows: section 2 presents the methodology used in this work; section 3 shows the reasoning process that agents can follow to engage in in real-time argumentation processes; section 4 shows an application example of our proposal; related works are shown in section 5; finally, section 6 shows the conclusions of this work. 2. Background In open multi-agent argumentation systems the arguments that an agent generates to support its position can conflict with arguments of other agents and these conflicts are solved by means of argumentation dialogues between them. In (Heras et al., 2013a) we 3 presented a computational case-based argumentation framework that agents can use to reason about argumentation processes. Also, in (Navarro et al., 2011) new case-based reasoning (CBR) techniques, called Temporal Bounded CBR (TB-CBR), to adapt the CBR methodology for its use in real-time MAS were proposed. The combination of these works forms the background of the real-time argumentation framework proposed in this paper. Thus, this section briefly reviews these proposals. 2.1. Case-based Argumentation Framework As proposed in (Heras et al., 2013a), this section illustrates the main components of our original case-based argumentation framework by following a running example based on the evaluation scenario of section 4. This framework will be adapted in Section 3 to be used in real-time scenarios. Thus, let us suppose that we have a MAS that manages a call centre that provides customer support, where a set of agents representing technicians must reach agreements about the best solution to apply to di↵erent types of software or hardware problems reported to the centre. In addition, the company that runs the call centre has signed a Service Level Agreement (SLA) with the customer that has contracted the support service. Among other terms, this SLA specifies a maximum acceptable time within the customer must receive a solution for the problem that has reported to the call centre. Therefore, to solve each problem, a group of technicians engage in an argumen- tation dialogue proposing their individual solutions and justifying the reasons that they have to propose them as the best solution for the problem within a specific time. From our point of view, a problem can be characterised by a set of features that describe it. In our framework, agents can use two types of knowledge resources to generate, select and evaluate arguments: A case-base with domain-cases: that represent previous problems and their so- lutions. Agents can use this knowledge resource to generate their positions and arguments. The position of an agent represents the solution that this agent pro- poses. Also, agents increase their domain knowledge at the end of each real-time argumentation dialogue by adding new cases to their domain-cases case-base. A case-base with argument-cases: that store previous argumentation experiences 4 ����������� ������������������������ ���������������� ���������������������� ������������������ ��������� ���������������� ���������������� ��� ������� Figure 1: Structure of a Domain-Case in the Call Centre Scenario and their final outcome. Agents use this resource to select the best position and argument to put forward in a specific situation in view of how suitable a similar position or argument was in a similar real-time argumentation dialogue. Also, agents store the new argumentation knowledge gained in each real-time agreement process, improving the agents’ argumentation skills. Figure 1 shows the structure of a domain-case in a call centre domain. In this example a technician of the call centre, say operator 1 (O1), has to solve the problem of a HP computer with Windows XP that reboots at random. Here, O1 has found the domain- case DC1 that matches de description of the problem to solve (also including the extra feature ”Model”). With this case, O1 can build its position ”Reinstall OS” and promote the value ”efficiency”. The argument-case of Figure 2 represents an argument that O1 generated to justify its position when arguing with another operator O2. Argument-cases have three possible types of components: the problem description that characterises the state of the world when the argument was stored, with a domain context that consists of a set of premises and a social context that includes information about the proponent and the opponent of the argument and their group. Moreover, we also store the preferences (ValPref ) of each agent or group over the set of values pre-defined in the system and the dependency relation between the proponent and the opponent. We consider two types of dependency relations (Dignum and Weigand, 1995): Power, when an agent has to accept a request from other agent because of some pre-defined domination relationship between them; and Charity, when an agent is willing to answer a request from other agent without being obliged to 5 do so. In the argument-case of Figure 2 proponent and opponent play the operator role and have a charity dependency relation between them. Also, both operators belong to the same group, which provides software support and does not impose any specific value preference order over its members. In the solution part, the conclusion of the case, the value promoted, and the acceptability status of the argument at the end of the dialogue are stored. This status shows if the argument was deemed acceptable, unacceptable or unde- cided. Thus, the conclusion of the argument-case in Figure 2 shows the solution proposed by O1 when it was presented with the automatic reboot problem. The argument-case promotes the same value than the solution generated from DC1 (efficiency). In addition, the conclusion part includes information about the attacks that the argument received during the real-time argumentation process. This information is used in our framework to emphasize the persuasive power of an argument that was finally accepted when it received attacks and its proponent was able to rebut them. Finally, the justification part of an argument-case stores the information about the knowledge resources that were used to generate the argument represented by the argument-case. Thus, the justification part of the argument-case of Figure 2 includes the domain-case DC1. In addition, the justifica- tion of each argument-case has a dialogue-graph (or several) associated, which represents the dialogue where the argument was put forward (DG1 in Figure 2). In this way, the sequence of arguments that were put forward in a dialogue is represented, storing the complete conversation as a directed graph that links argument-cases. This graph is used in our framework to improve the efficiency of an argumentation dialogue, for instance, ending early a current dialogue that is very similar to a previous one that ended up in disagreement. In our proposal, arguments that agents interchange are tuples of the form: Arg = {φ, v, < S >}, where φ is the conclusion of the argument, v is the value that the agent wants to promote and < S > is a set of elements that justify the argument (the support set). This set consists of di↵erent elements, depending on the argument purpose. On the one hand, if the argument provides a justification for a proposed solution, the support set includes the set of premises that represent the context of the domain where the argument has been put forward (those premises that match the problem to solve and 6 ������������� ��������������������� ���������������� �������������������������������� ��������������� ������������� �������������������� ������� �������������� �������������� ������������������ ���������������� ��������� ���������������� ���������������� ��� �������������� ������ �������������� ��������� �������������������� ������� ������ �������������� ��������� �������������������� ��������� �������������������� ��������������� ���������� ����� ���������������� ������������� �������� ������ ������������ ����� ��������������� �������� ������������ ������� �������� ��������� ��������� ������������� ������� Figure 2: Structure of an Argument-Case in the Call Centre Scenario other extra premises that were assumed) and optionally, any knowledge resource used by the proponent to generate the argument (domain-cases) and select it (argument-cases). This type of argument is called a support argument. On the other hand, if the argument attacks the argument of an opponent, it is called an attack argument and its support set can include any of the allowed attacks of our framework. These are distinguishing premises or counter-examples, as proposed in (Bench-Capon and Sartor, 2003). A distinguishing premise is a premise that does not appear in the description of the problem to solve and has di↵erent values for two cases, or a premise that appears in the problem description and does not appear in one of the cases. A counter-example for a case is another case which problem description matches the current problem to solve and also subsumes the problem description of former case, but proposing a di↵erent solution. Therefore, the premise ”Model” would be a distinguishing premise between the domain- case of Figure 1 and another domain-case, say DC2, that has exactly the same problem description than DC1, but that stores a di↵erent model for the HP computer. Also, as- suming that DC2 proposes an alternative solution (e.g ”Check for Memory Errors”) that promotes the value accuracy, it could be used to generate a counter-example attack to DC1 and vice-versa. In our running example, if O1 generates the support argument SA1 = {”Reinstall”, Ef f iciency, < {W indows XP , HP , P avillion, OW Automatic Reboot}, {DC1}, -, -, -, -, - >} to justify its position in view of another di↵erent position generated by O2, the latter could attack this argument with an attack argument AA2 = {”Check for 7 Memory Errors”, Accuracy, < {W indows XP , HP , P avillion, OW Automatic Reboot}, -, -, -, -, -, {DC2} >}, which presents the counter-example DC2. However, as shown in Figure 2, O1 ’s support argument remained acceptable at the end of the dialogue (when the argument-case was retained), which means that O1 was able to rebut the attack or that O2 withdrawn it. 2.2. Temporal Bounded Case-based Reasoning Case-based Reasoning systems (CBR) provide agent-based systems with the ability to reason and learn from the experience of agents. To do it, these systems reuse or adapt past solutions to solve current similar problems. However, in real-time environments agents have a bounded time to reason and generate answers to the requests that they receive. Thus, CBR techniques must take into account temporal restrictions to observe real-time constraints. CBR systems are highly dependent on their application domain, and therefore, designing a general CBR model that might be suitable for any type of real-time domain is unattainable. In (Navarro et al., 2011) a new CBR methodology, called Temporal Bounded CBR (TB-CBR), is presented to provide some guidelines to adapt CBR for its use in real-time systems. Mainly, two factors have been controlled to temporal bound the reasoning process of the system that implements this approach: the case-base data structure and the temporal execution of the reasoning phases of the TB-CBR cycle (retrieve, reuse, revise and retain, as is common for CBR systems). The design decision about the data structure of the case-base and the di↵erent algo- rithms that implement each CBR phase are important factors for determining the execu- tion time of the CBR cycle. The number of cases in the case-base is another parameter that a↵ects the temporal cost of the retrieval and retain phases. Thus, a maximum num- ber of cases in the case-base must be pre-defined by the designer. Note that, usually, the temporal cost of the algorithms that implement these phases depends on this number. In any case, the retrieval and retention time can be reduced by using an indexing algorithm. These algorithms organize the case-base by selecting a specific feature (or set of features) from the cases, grouping together those cases that share the same values for these fea- tures. This reduces the cost of the search for similar cases (for retrieval or previous to the introduction of new cases in the case-base) to a specific set of cases with the same index 8 as the current case (Patterson et al., 2005) (Quinlin, 1993) (Wess et al., 1993). Moreover, the time dedicated to execute the di↵erent phases must be controlled. To do this, the worst case execution time (WCET) of each function executed in each phase must be known. The sum of all function times represents the cost of running a complete cycle of TB-CBR. By using the WCET of each function we are ensuring that the time to execute the TB-CBR cycle does not exceed this calculated time. Figure 3: Temporal Bounded CBR cycle. In (Navarro et al., 2011), we propose a modification of the classic CBR cycle in order to adapt it to be applied in real-time domains. Figure 3 shows a graphical representation of this approach. Firstly, we group the four reasoning phases that implement the cognitive task of the real-time agent into two stages defined as: the learning stage, which consists of the revise and retain phases and the deliberative stage, which includes the retrieve and reuse phases. Both phases will have their own execution time scheduled. Therefore, the designer can choose to either assign more time to the deliberative stage or keep more time for the learning stage (and thus, design agents that are more sensitive to updates). The timeM anager(tmax) function is in charge of completing this task. Using this function the designer must specify how the real-time agent acts in the environment. Regardless of the decision taken by the designer, the timeM anager function should allow sufficient time for the deliberative stage to ensure a minimal answer. These new CBR stages must be 9 designed as an anytime algorithm (Dean and Boddy, 1988), where the process is iterative and each iteration is time-bounded and may improve the final response. The anytime behavior of the TB-CBR is achieved through the use of two loop control sequences. The loop condition is built using the enoughT ime function, which determines if a new iteration is possible according to the total time that the TB-CBR has to complete each stage. In accordance with this, the operation of our time bounded CBR cycle is shown in Algorithm 1. Firstly, the main di↵erence that can be observed between the classic CBR cycle and the TB-CBR cycle is the starting phase. Our real-time application domain and the restricted size of the case-base (as explained in the following sections) gives rise to the need to keep the case-base as up to date as possible. Commonly, recent changes in the case-base will a↵ect the potential solution that the CBR cycle is able to provide for a current problem. Therefore, the TB-CBR cycle starts at the learning stage, checking if there are previous cases waiting to be revised and possibly stored in the case-base. In our model, the solutions provided at the end of the deliberative stage will be stored in a solution list (solutionQueue) while a feedback about their utility is received. When each new CBR cycle begins, this list is accessed and while there is enough time, the learning stage of those cases whose solution feedback has been recently received is executed. If the list is empty, this process is omitted. After this, the deliberative stage is executed. The deliberative stage is only launched if the real-time agent has a problem to solve. These problems are stored in a list (problemQueue) as they arise. Thus, the retrieval algorithm is used to search the case- base and retrieve a case that is similar to the current case (i.e. the one that characterizes the problem to be solved). The solutions are stored just after the end of the delibera- tive stage. Each time a similar case is found, it is sent to the reuse phase where it is transformed into a suitable solution for the current problem by using a reuse algorithm. Therefore, at the end of each iteration of the deliberative stage, the TB-CBR method is able to provide a solution for the problem at hand, although this solution can be improved in following iterations if the deliberative stage has enough time to perform them. According to the temporal analysis of each phase of the CBR cycle, the anytime behavior of the TB-CBR is achieved through the use of two loop control sequences. The 10 loop condition is built using the enoughT ime function, which determines if a new iteration is possible according to the total time that the TB-CBR has to complete each stage. Algorithm 1 TB-CBR algorithm Require: tmax, case-base 1: (tlearning,tdeliberative) − timeManager(tmax) 2: if SolutionQueue 6= ; then 3: while enoughTime(tnow,trevise,tretain,tlearning) and SolutionQueue 6= ; do 4: r − pop(SolutionQueue) 5: {adequate − f Revision(r)}trevise 6: if adequate then 7: {f Retention(r, case − base)}tretain 8: if ProblemQueue 6= ; then 9: problem − pop(ProblemQueue) 10: repeat 11: {cases − Push(f Retrieval(problem, case − base))}tretrieve 12: {solution − f Adaptation(cases)}treuse 13: bestSolution − bestSolution(solution,bestSolution) 14: until ¬enoughTime(tnow,tretrieve,treuse,tdeliberative) 15: solutionQueue − push(bestSolution) 16: Return bestSolution 3. Real-Time Argumentation This section presents the adaption to real-time environments of the framework in- troduced in Section 2.1, and thus the original model becomes a real-time case-based argumentation framework. The original process allows agents to use the knowledge re- sources presented before to generate, select and evaluate their positions and arguments in real-time and automatically learn from the experience. Now, to temporal bound the original reasoning process, we have used the TB-CBR approach (see Figure 4). Thus, the process can be divided into three phases: generation and selection of positions, where the agent generates its potential positions and select the best one to propose; evaluation of positions, where the agents evaluates the positions generated by other agents; and fi- nally, argument management, where agents can either defend their positions if they are attacked or else, attack other di↵erent positions proposed by their partners. To defend 11 their positions, agents have to evaluate the attack arguments received and generate and select the best support argument to propose. To attack di↵erent positions, agents have to ask the agent that they want to attack for providing them with a support argument for the position to attack. Then, agents can generate and select the best argument to attack the support argument provided. In this section we focus on the adaption of the original algorithms of the agents’ reasoning process to real-time scenarios. For a complete docu- mentation of the pseudocode of each algorithm we refer the reader to the work published in (Heras et al., 2013a). The following sections explain these phases from the perspective of their temporal dimension. Along them, we assume that a set of agents with di↵erent positions are arguing to reach a real-time agreement to solve a complex problem that could be described with a set of features. 3.1. Position Generation and Selection In the first step of our real-time reasoning process, an agent can generate its individual position to solve the problem at hand. To perform this process, the agent retrieves from the domain case-base those cases that match with the specification of the current problem and generates its solution (or a list of potential solutions) by reusing the solution(s) applied to the retrieved cases. Thus, the set of retrieved cases could provide di↵erent solutions for the same problem. Then, the agent can use its case-base of argument-cases to select the best position to propose. The first step for this selection is to order the positions in subsets, taking into ac- count the value promoted by each position. Thus, the agent will assign to each subset a Suitability Level (SL). Positions that promote the agent’s most preferred value will be labelled with suitability level 1, positions that promote the second most preferred value will be labelled with level 2 and so on. Then, positions will be ordered within each level by its Similarity Degree (SimD) with the problem to solve, computed by using a domain-dependent similarity measure. After that, the agent will use the argumentation knowledge stored in its argument- cases case-base. For each position generated, the agent creates an argument-case that represents this position. This is a necessary translation to add the current social context 12 ������������������ ��� � � ����������������� ���� �������������������� ������������� ��� � � � � ����������� ���������� ��� � � ��������������� ���� ��������� ���������� ���� ������������������� ���� � � � � � � � � � � � � � � � � � � � � � � � � � � � � � ��������������������� ������������������� ���������� ���������� ���������� �� ��� ������������������ ����������� ���������������������� ����������� ��������������������� ������������������� � � � � � � � � � �� � � � � � � � � �� � � � � � � � � � � � � � � � � �� ��� �������� ����������� ������� ������ �������� �������������� ���� �������� ����������� ������� ���������� ��� ��� ����� ���� ���� �� � � � � � � � � � � � � � � � � � � � � � � � � Figure 4: Real-Time Argumentation Process to the position and be able to make queries to the argument-cases case-base and retrieve argument-cases that represent similar arguments that justify or attack the position. Then, the agent compares the argument-case created for each position with its case-base of argument-cases and retrieves the sets of argument-cases that match it. In this way, the agent can assign to each position a Support Factor (SF) from the argumentation point of view and decide which argument-case (and thus, which position) is most suitable to propose in view of its past argumentation experience and its current social context. As criteria for making such decision, we consider the following parameters: • Persuasiveness Degree (P D): is a value that represents the expected persuasive power of an argument by checking how persuasive an argument-case with the same 13 problem description and conclusion was in the past. To compute this degree, the number of argument-cases that were deemed acceptable out of the total number of retrieved argument-cases with the same problem description and conclusion as the current argument is calculated. • Support Degree (SD): is a value that provides an estimation of the probabil- ity that the conclusion of the current argument was acceptable at the end of the dialogue. It is based on the number of argument-cases with the same problem de- scription and conclusion that where deemed acceptable out of the total number of argument-cases retrieved. • Risk Degree (RD): is a value that estimates the risk for an argument to be attacked in view of the attacks received for an argument(s) with the same problem description and conclusion in the past. It is based on the number of argument-cases that were attacked out of the total number of argument-cases with the same problem description and conclusion retrieved that were deemed acceptable. • Attack degree (AD): is a value that provides an estimation of the number of attacks received by a similar argument(s) in the past. To compute this degree, the set of argument-cases with the same problem description that were deemed acceptable is retrieved. Then, this set is separated into several subsets, one for each di↵erent conclusion that these argument-cases entail. The sets whose conclusion match with the conclusions of the arguments to assess are considered, while the other sets are discarded. Thus, we have a set of argument-cases for each di↵erent potential argument (and its associated conclusion) that we want to evaluate. For each argument-case in each set, the number of attacks received is computed (the number of distinguishing premises and counter-examples received). Then, for each set of argument-cases, the average number of attacks received is computed and the attack degree of each argument is calculated by a linear transformation. • Efficiency degree (ED): is a value that provides an estimation of the number of steps that it took to reach an agreement posing a similar argument in the past. It is based on the depth n from the node representing a similar argument-case retrieved 14 to the node representing the conclusion in the dialogue graphs associated with it. To compute this degree, the same process to create the subsets of argument-cases as in the above degree is performed. Then, for each argument-case in each subset, the number of dialogue steps from the node that represents this argument-case to the end of the dialogue is computed. Also, the average number of steps per subset is calculated. Finally, the efficiency degree of each argument is calculated by a linear transformation. • Explanatory Power (EP ): is a value that represents the number of pieces of information that each argument covers. It is based on the number of knowledge resources were used to generate each similar argument-case retrieved. To compute this number, the same process to create the subsets of argument-cases as in the above degrees is performed. Then, for each argument-case in each set, the number of knowledge resources in the justification part is computed (the number of domain- cases and argument-cases). For each set of argument-cases, the average number of knowledge resources used is computed and the explanatory power of each argument is calculated by a linear transformation. Finally, the suitability factor of a new argument-case and its associated position is computed by the formula: SF =((wP D ⇤ P D + wSD ⇤ SD + wRD ⇤ (1 − RD) + wAD ⇤ (1 − AD) + wED ⇤ ED + wEP ⇤ EP )) (1) where wi 2 [0, 1], P wi = 1 are weight values that allow the agent to give more or less importance to each decision criteria. Finally, positions are ordered from more to less suitability by following the equation: Suitability = wSimD ⇤ SimD + wSF ⇤ SF (2) where wi 2 [0, 1], P wi = 1 are weight values that allow the agent to give more or less importance to the similarity degree or the support factor. The most suitable position of suitability level 1 is selected as the one that the proponent agent is going to propose and defend first. However, each agent keeps the rest of positions to make alternative proposals if its original position is rebutted. 15 Algorithm 2 shows the pseudocode of the algorithm that implements the generation of positions, the generation of the associated argument-cases and the selection of positions. This pseudocode modifies the version presented in (Heras et al., 2013a) by adding the estimated temporal cost for each function. The temporal cost of executing this phase has been bounded to get the worst case execution time. This allows the agent to know how long it takes to extract a valid position. Thus, if the temporal restrictions of the application domain require the agent to propose a position in less time than it needs to generate al least one suitable position, the agent can reject the request and do not engage in the real-time agreement process. In this way, the agent does not waste computation time on tasks that it knows that it will not be able to complete on time. In the algorithm, the function generatePositions generates the k first positions by using the algorithm generatePositions. In the worst case, the function has to check the whole case-base of domain-cases, incurring a temporal cost O(n) (where n is the number the cases stored in the domain-cases case-base) for finding and extracting similar domain- cases. Then, for each domain-case extracted, the function reuses its solution to generate a position. Again, in the worst case (in the temporal sense), we can generate n positions. Therefore, the generatePositions function has an asymptotic temporal cost of O(n2). The generateArgumentCase is a function that generates for each position its associated argument-case. This generation only changes the way in which we represent positions and we assume for it a constant temporal cost that is negligible in terms of the total temporal cost of the algorithm 2. The retrieveSimilarityDegree is a function that retrieves the similarity degree of each position with regard to the problem to solve. This degree is stored with each position and we assume that the temporal cost for checking this information is constant and negligible in terms of the total temporal cost of the algorithm 2. The computeSF is a function that computes the support factor for each position by means of its associated argument-case. In the worst case, this function has to check the whole case-base of argument-cases twice: one to retrieve the target set of argument-cases for the persuasiveness degree, support degree and risk degree and another to retrieve the average number of attacks, steps and knowledge resources of each subset retrieved 16 to compute the attack degree, efficiency degree and explanatory power. Therefore, this process has an asymptotic temporal cost of O(m2), where m is the number the cases stored in the argument-cases case-base. The selectPosition is a domain-dependent function that sorts the set of positions from more to less suitable with respect to some domain-dependent criteria. Most comparison algorithms have a worst case temporal cost of O(q2), being q the number of positions generated by the algorithm. In our case, if we generate one position from each domain- case that is stored in the case-base, we have that q = n, where n is the size of the domain-cases case-base. Thus, we estimate this cost for selecting the best position to put forward as O(n2). Finally, the mostSuitable is a domain-dependent function that returns the most suit- able position to solve the problem (e.g. the first position from the sorted list computed by the selectPosition) function. Again, if we assume that the temporal cost for checking this information is constant, it can be negligible in terms of the total temporal cost of the algorithm 2. Algorithm 2 Position Generation and Selection Require: ProblemDescription, k, wSimD, wP D, wSD, wRD, wAD, wED, wEP //The description of the problem to solve, the maximum number of positions to generate, and the weights for each element of the similarity degree and the support factor 1: positions = ; 2: argumentCases = ; 3: SimD = ; 4: SF = ; 5: selectedPositions = ; 6: positions = generatePositions(ProblemDescription, k) //O(n2) 7: for [i = 1;i  lenght(positions);i + +] do 8: argumentCases[i] = generateArgumentCase(ProblemDescription, positions[i]) //O(1) 9: SimD[i] = retrieveSimilarityDegree(positions[i]) //O(1) 10: for [i = 1;i  lenght(argumentCases);i + +] do 11: SF[i] = computeSF(ProblemDescription, argumentCases[i], argumentCases, wP D, wSD, wRD, wAD, wED, wEP ) //O(m 2) 12: selectedPositions = selectPosition(positions, argumentCases, SD, SF) //O(n2) 13: Return mostSuitable(selectedPositions) //O(1) 17 Summarising, the asymptotic cost to execute the process to generate and select po- sitions in the worst case is O(n4 ⇤ m3) where n is the number the cases stored in the domain-cases case-base and m is the number of cases stored in argument-cases case-base. Therefore, to be able to make a prediction and bound the temporal cost of the algorithm 2, the maximum number of cases in the domain-cases and the argument-cases case-base must be known and fixed in advance. In this way, if the contents of the agent’s case-bases contain the necessary information to generate and select an appropriate position, the al- gorithm will be able to generate at least one position on time. In this sense, as shown in Figure 4 the position generation and selection phase is mandatory and the other two phases (i.e. position evaluation and argument management) are optional and executed while the agent has still time to perform its temporal bounded reasoning process. 3.2. Position Evaluation Once the process to generate and select positions has finished, agents of our framework can either receive attacks to their positions or decide to challenge the positions of other agents. Thus, agents are able to evaluate its position with regard to other positions. The first step to evaluate an agent’s position is to check if it is consistent with the positions of other agents. For the sake of simplicity, here we assume that a position is consistent with other position if they totally match (they are the same). Agents do not attack consistent positions, but they can still generate support arguments if their own positions are challenged. Another possibility arises when a proponent agent has been able to generate the position of another agent, but it has selected another position as more suitable to propose. In that case, the proponent would accept the opponent’s position if the latter has a power relation over the proponent and would try to attack the opponent’s position otherwise. Finally, if the opponent’s position is not in the set of positions generated by the proponent and the opponent does not have a power relation over the proponent, the proponent can try to generate an argument to attack the opponent’s position and otherwise, accept the opponent’s position. In our real-time argumentation scenario the number of agents engaged in the dialogue cannot be determined a priori. Therefore, the temporal cost of evaluating the positions of other agents cannot be bounded. Thus, this phase has been implemented as an anytime 18 process. The agent has a maximum time (deadline) to evaluate all positions generated by other agents. Once this deadline is exceeded, the agent stops the evaluation and proceeds to the next phase of its reasoning process. Next sections explain the type of arguments that agents can generate and how these arguments are selected and evaluated in a real-time agreement process. 3.3. Argument Management Agents generate arguments when they are asked to justify their positions (support arguments) or when they attack others’ positions or arguments (attack arguments). A support argument has a support set that consists of the set premises that describe the problem and of any of the knowledge resources used to generate the position to justify (domain-cases and argument-cases). Attack arguments are generated when the proponent of a position provides an argument to justify it and an opponent wants to challenge the position or else, when an opponent wants to attack the argument of a proponent. If the agent receives an attack, it must evaluate its current argument in view of the incoming argument that poses the attack. The attack arguments that an agent can generate depend on the elements of the support set of the attacked argument. On one hand, if the support set includes a set of premises, the agent can generate an attack argument with a distinguishing premise. On the other hand, if the support set includes a domain-case or an argument-case, the agent can check its case-bases to find counter-examples to generate the attack. As for the case of generating positions, agents can generate several attack arguments. Then, to select the best attack argument to put forward in a specific step of the real-time agreement process, the agent uses the information of its argument-cases case-base and selects such one that is expected to have higher persuasive power in view of the agent’s previous experiences. The process to generate and select arguments is very similar than the process to generate and select positions. In the worst case, this process has to check the whole domain-cases and argument-cases case-base, incurring the same temporal cost of algorithm 2, which is O(n4 ⇤ m3) where n is the number the cases stored in the domain-cases case- base and m is the number of cases stored in argument-cases case-base. Therefore, to be 19 able to make a prediction and bound the temporal cost of the argument generation and selection process, the maximum number of cases in the domain-cases and the argument- cases case-base must be also known and fixed. Finally, agents must evaluate their arguments in view of the arguments proposed by their opponents. Then, a proponent agent can decide if an opponent’s argument con- flicts with its argument and hence, its argument is deemed acceptable, non-acceptable or remains undecided (it cannot make a decision over it). This evaluation is performed by using the defeat relation between arguments defined in the original case-based argu- mentation framework (Heras et al., 2012), which specifies which attacks over arguments succeed. If the proponent argument defeats the opponent’s, the latter acceptability status will change to non-acceptable. Then, the opponent can try to generate a new argument to counter-attack. On the contrary, if the proponent’s argument cannot defeat the op- ponent’s, the proponent has to withdraw its last argument and, if there are any, send to the opponent an alternative argument or propose an alternative position from its list of generated positions. In this case, the proponent’s argument acceptability status would preliminary change to undecided. An argument that remained undefeated at the end of the real-time agreement process is deemed acceptable. The final acceptability status of arguments is decided by following a dialogue-game protocol (Jordán et al., 2011). At this external level, the number of agents participating in the argumentation process cannot be known in advance and hence, we cannot estimate the number of arguments that the agent has to evaluate and the underlying attacks that the agent can receive or generate. As in the case of the evaluation of positions, we use here an anytime process by which the agent has been assigned a deadline to perform all interactions with the rest of agents. When this deadline is finished, the argument management phase ends and the agent must provide its final solution. Therefore, in the worst case where the agent has not enough time to argue, it does nor participate in the argumentation dialogue. Otherwise, if the allowed time is large enough, it can make an intensive use of its domain and argumentation knowledge and engage in the argumentation process proposing and defending all positions that it is able to generate. 20 4. Integrating Real-time Agreement Agents in a real application In this section, we perform tests to evaluate the performance of the proposed real- time case-based argumentation framework. With this objective, the framework has been implemented in the domain of a customer support application. Concretely, we consider a society of agents that act in behalf of a group of technicians that must solve problems in a Technology Management Centre (TMC) (Heras et al., 2009b)(Heras et al., 2013a), which control every process implicated in the provision of technological and customer support services to private or public organisations by means of a call centre. Therefore, we set a society composed by call-centre technicians playing the role of operator. Operators form groups that must solve the problems that the call centre receives, commonly known as tickets in the call-centres jargon. The dependency relations in this society establish that technicians with the same role have a charity relation among them. In addition, the company that runs the call centre has signed a Service Level Agreement (SLA) with the customer that has contracted the support service. Among other terms, this SLA specifies a maximum acceptable time within the customer must receive a solution for the problem that has reported to the call centre. If the solution exceeds this time, the centre can receive an economic penalization in its contract with the customer, which is greater as time passes. In this application domain we assume that each technician has a helpdesk application to manage the big amount of information that processes the call centre. In addition, this helpdesk would implement an argumentation module to solve each ticket as proposed in our framework. Hence, we assume the complex case where a ticket must be solved by a group of agents representing technicians that argue to reach an agreement over the best solution to apply. Each agent has its own knowledge resources (acceded via his helpdesk) to generate a solution for the ticket. The argumentation module of each agent includes a Domain-CBR engine that makes queries to its domain-cases case-base and an Argumentation-CBR engine that makes queries to its argument-cases case-base. The system has been implemented by using the Real-Time Multi-Agent Platform jART (Navarro et al., 2004). This platform is implemented in RT-Java and allows to execute agents in a time-controlled way, ensuring the fulfillment of the agents tasks within a 21 specific time. The data-flow for the argumentation dialogue to solve each ticket is the following: 1. The system presents a group of technicians with a new ticket to solve. 2. An agent, called the initiator agent, opens a new argumentation dialogue. 3. Technicians enter in the dialogue. 4. If possible, each technician generates his own position by using the argumentation module and propose it. This module supports the argumentation framework pro- posed in this paper. In this sense, technicians propose their solutions before the maximum time specified in the SLA contracted by the customer is met. All techni- cians that are willing to participate in the argumentation process are aware of the positions proposed in each moment. 5. The technicians argue to reach an agreement over the most suitable solution by following a persuasion dialogue controlled by a dialogue game protocol (Jordán et al., 2011). The dialogue proceeds as a set of parallel dialogues between technicians. Therefore, one technician can only argue with another at the same time. In each parallel dialogue, two technicians take turns to challenge the positions of the other technician, to assert arguments to justify their positions or to attack the other technician’s positions and arguments. If a position is attacked but the opponent agent cannot defeat it, the position receives one vote. Otherwise, the proponent of the position must withdraw it from the dialogue. 6. The dialogue ends when an agreement is reached (i.e. only one position remains undefeated in the argumentation dialogue) or a deadline specified in the SLA is met. The best solution is proposed to the user and feedback is provided and registered by each technician helpdesk. In this way, agents update their case-bases with the information of the actual solution applied and the arguments generated during the argumentation dialogue. If there is no agreement reached, the best solution can be selected by using di↵erent policies, such as to select the most voted position or the most frequent. In this domain, we assume that the most efficient technicians are acknowledged and rewarded by the company. Therefore, each technician follows a persuasion dialogue with 22 their partners, trying to convince them to accept its solution as the best way to solve the ticket received, while observing the common objective of providing the best solution for a ticket on time. Note that with this approach, we are assuming that the TB-CBR process is working properly and as the systems evolves, agents learn correct solutions in their domain-cases case base and useful arguments in their argument-cases case-base, hence providing more accurate solutions. To evaluate the system, we have run several tests populating the call centre with two di↵erent types of agents: agents that implement our original case-based argumenta- tion framework, where no real-time considerations were taken into account in the agents’ reasoning process to manage positions and arguments (noRT agents); and agents that im- plement the real-time case-based argumentation framework proposed in this work, which follow a temporal bounded TB-CBR cycle to manage their positions and arguments (RT agents). Thus, the latter type of agents are able to provide answers within the specific time that the SLA requires. With these tests, we evaluate the efficiency of the system that implements the real-time case-based argumentation framework by comparing its per- formance with the one achieved by the original framework. For the tests, a real database of 200 tickets solved in the past is used as domain knowledge. Translating these tickets to domain-cases, we have obtained a tickets case- base with 48 cases. Despite the small size of this case-base, we have rather preferred to use actual data instead of a larger case-base with simulated data. The argument-cases case-bases of each agent are initially empty and populated with cases as the agents acquire argumentation experience in execution of the system. To diminish the influence of random noise, for each round in each test, all results report the average and confidence interval of 48 simulation runs at a confidence level of 95%, thus using a di↵erent ticket of the tickets case-base as the problem to solve in each run. The results report the mean of the sampling distribution. The experiments have been performed with a population of 7 agents representing operators (in (Heras, 2011, Chapter 6) we discuss that populations greater than this number are not appropriate to elicit useful results with the small size of our case-base). Each test has been repeated for di↵erent assignments of the domain and argumentation 23 knowledge that agents have. Concretely, the domain-cases of the case-bases of the agents were randomly populated and increased by 5 from 5 to 45 cases in each experimental round. Also, the argument-cases case-base of each agent were randomly populated with argument-cases and increased by 2 from 2 to 18 cases in each experimental round. These cases were obtained from a set of cases created by training the system performing several problem-solving executions. (Heras, 2011, Chapter 6) demonstrates that this is reasonable amount of argument-cases that are elicited from the possible argumentation experiences of 7 agents arguing with a maximum knowledge of 45 domain-cases in our experimental domain. Therefore, each experiment is repeated for each type of agents (noRT and RT agents) during 9 rounds. In the first round, agents have 5 domain-cases (dc) and 2 argument-cases (ac), in the second round agents have 10 domain-cases and 4 argument- cases and so on up to 45 domain-cases and 20 argument-cases. In each simulation, an agent is selected randomly as initiator of the discussion. This agent has the additional function of collecting data for analysis and controlling the dead- line specified in the SLA. However, from the argumentation perspective, its behaviour is exactly the same as the rest of agents and its positions and arguments do not have any preference over others (unless there is a dependency relation that states it). The initiator agent receives one problem to solve per run. Then, it contacts its partners (the agents of its group) to report them the problem to solve. If the agents do not reach an agreement after a maximum time1, the initiator chooses the most supported (the most voted) solu- tion as the final decision (or the most frequent in case of draw). If the draw persists, the initiator makes a random choice among the most frequent solutions. To check the solution accuracy (the average error in the solutions that the system provides), the solution agreed by the agents for each ticket requested is compared with its original solution, stored in the tickets case-base. In the first tests, the call centre is first managed by a group of RT agents and after that by a group of noRT agents and there is no time limitation to provide answers to the customers requests (no deadline specified in the SLA). Then, the percentage of problems 1If the SLA does not define a maximum time, the initiator allows agents to argue until they do not have more positions and arguments to interchange. 24 that the system is able to solve (provide a solution, regardless of its level of suitability), and the average solution error have been analysed. One can expect that having more domain and argumentation knowledge (more domain and argument-cases in the agents’ case-bases) will prevent less problems from being undecided and will increase the number of problems that were correctly solved (the mean error in the solution predicted decreases). As presented in Table 1, for both types of agents the system achieves the same performance results, which is the expected result since the core argumentation skills of noRT and RT agents are exactly the same (they only di↵er in the temporal management of their reasoning process). The percentage of problems solved by the system increases and the mean error decreases with the number of domain and argument-cases that agents have, up to the 100% of solved problems with no errors from 25 domain-cases and 10 argument- cases onwards. h h h h h h h h h noRT & RT Cases 5dc/2ac 10dc/4ac 15dc/6ac 20dc/8ac 25dc/10ac 30dc/12ac 35dc/14ac 40dc/16ac 45dc/18ac Solved Problems % 64.58% 87.50% 97.92% 97.87% 100% 100% 100% 100% 100% Mean Error % 39.58% 16.67% 6.25% 4.26% 0.00% 0.00% 0.00% 0.00% 0.00% Table 1: Percentage of problems that the system is able to solve and error average (with no deadline in the SLA) Also, Figure 5 shows that the average time that noRT and RT agents take to solve problems are very similar. At first, the mean resolution time increases as the contents of the case-bases do, since agents have more justification elements and argumentation skills to generate more positions to propose and to defend their positions from attacks, which gives rise to longer agreement processes. However, up to 25 domain-cases and 10 argument-cases onwards the case-bases of agents begin to include many redundant data and agents reach agreements more quickly (since they propose the same or very similar positions). Therefore, the average time of the agreement process decreases and tends to the minimum exchange of locutions to communicate data among agents. As shown in Figure 5, both noRT and RT agents need a minimum of 1200 milliseconds to provide a solution with the fewer possible amount of information in their case-bases (regardless of its quality) for the requests received by the system. Now, to test the performance of the system with a tight deadline specified in the SLA, we have repeated the above tests allowing agents having only 1100 milliseconds to provide solutions. After 25 Figure 5: Average Time of the Agreement Process in milliseconds (with no deadline in the SLA) this deadline, we consider that the problem has not been solved and thus, the agreement process has failed (counting as a new error). Table 2 compares the percentage of problems that the system is able to solve with noRT and RT agents and a strict deadline of 1100 milliseconds. In addition, Table 3 presents the error average achieved in these tests. The results obtained for both types of agents in the case that there is not deadline is also shown for comparative purposes in both tables. Results from Tables 2 show that RT agents are able to bound their reasoning cycle and to provide much more solutions on time than noRT agents, thus achieving much lower error percentages (see Table 3). However, although the percentage of solved problems increases and the average error in the solutions provided decreases with the amount of knowledge that agents have in their case-bases, they do not have enough time to solve all problems with no error, as they are with the appropriate amount of knowledge (up to 25 domain-cases and 10 argument-cases) if no deadline is specified. In fact, as presented in Table 3, the quality of the solutions provided by RT agents is worst than the one that they achieve operating without deadline, but still it is much better than the one achieved by noRT agents. Nevertheless, in many real-time application domains is preferable to have a good enough solution (even if it is not the optimum) than leave the problem unresolved. Therefore, RT agents prevent many problems to remain unresolved and the real-time process to end in disagreement. Finally, Figure 6 shows the average time that noRT and RT agents take to solve 26 h h h h h h h h h h hh Solved Problems % Cases 5dc/2ac 10dc/4ac 15dc/6ac 20dc/8ac 25dc/10ac 30dc/12ac 35dc/14ac 40dc/16ac 45dc/18ac noDeadline 64.58% 87.50% 97.92% 97.87% 100% 100% 100% 100% 100% noRT (1100 ms) 10.42% 20.83% 37.50% 37.50% 38.91% 45.83% 47.92% 52.08% 52.08% RT (1100 ms) 50.00% 70.83% 75.00% 76.60% 76.95% 77.08% 77.08% 81.25% 87.50% Table 2: Percentage of problems that the system is able to solve (deadline 1100 ms) h h h h h h h h h h hh Solved Problems % Cases 5dc/2ac 10dc/4ac 15dc/6ac 20dc/8ac 25dc/10ac 30dc/12ac 35dc/14ac 40dc/16ac 45dc/18ac noDeadline 39.58% 16.67% 6.25% 4.26% 0.00% 0.00% 0.00% 0.00% 0.00% noRT (1100 ms) 91.67% 79.17% 68.09% 68.09% 62.50% 54.17% 52.06% 47.92% 47.92% RT (1100 ms) 50.00% 31.25% 25.00% 23.40% 23.13% 22.92% 22.92% 18.75% 12.50% Table 3: Error average in the solutions provided by the system (deadline 1100 ms) problems with a deadline of 1100 milliseconds specified in the SLA. Opposite to RT agents, noRT agents are not able to bound their reasoning cycle to propose positions and generate attacks and counter-attacks within this deadline. Thus, the mean time of their agreement processes is almost the same than the one presented in Figure 5 (variations are due to di↵erent temporal costs incurred in communicating positions and arguments via the network). This does not meet the requirements of the SLA and gives rise to the bad results presented in Tables 2 and 3 for the percentage of solved problems and mean error. On the contrary, the mean time in the agreement processes of RT agents slightly variates around the specified deadline of 1100 milliseconds. Again, variations over this time are due to di↵erent temporal costs incurred in communicating positions and arguments via the network. Therefore, RT agents succeed in adjusting their reasoning cycles to the tight deadline they are allowed, and with a medium amount of knowledge (up to 25 domain- cases and 10 argument-cases) are able to solve more than the 77% of problems with less than the 23% of error. 5. Related Work Over the last years, the research on technologies to reach agreements in computing systems is experiencing an exponential growth. The term agreement technologies already appears as a topic in the main AI and MAS conferences (e.g. the International Conference of Autonomous Agents and Multi-agent Systems (AAMAS) and the International Joint Conference on Artificial Intelligence (IJCAI) and there are new conference tracks and workshops specialised on this area (e.g. Workshop on Agreement Technologies (WAT)). 27 Figure 6: Average Time of the Agreement Process in milliseconds (deadline 1100 ms) However, the transfer of these technologies to the industry is still a challenge, and the few commercial organisations which adopt them would be classified as early adopters (Luck and McBurney, 2008). In addition, if the adoption of agreement technologies is still at its early stages, the application of these technologies to real-time scenarios is even more novel. Focusing on argumentation as an agreement technology, the literature reports few contributions on real-time applications of argumentation frameworks. In fact, these contributions are not specifically focused on bounding the reasoning process that agents follow to reach real- time agreements, but consist of negotiation systems where agents stop the negotiation process before a specific time. This is the case of the the real-time negotiation model for reflective agents proposed in (Soh and Tsatsoulis, 2005), applied to resource allocation problems (concretely, to multi-sensor target tracking). The model is an case-based nego- tiation model that integrates a real-time BDI architecture for the agents with a temporal logic model. These features allow to establish when certain states of the negotiation pro- tocol have to be true and for how long, which makes real-time implementation feasible. However, in this framework cases are situated and need domain-specific adaption rules to devise strategies that are applicable to the current negotiation context from them. Opposite to our proposal, in this work persuasion is viewed as a negotiation protocol for information exchange between two agents. These agents try to reach a deal in which the initiator agent provides arguments to support a request to convince the responding agent 28 to share its sensing resources with it. Thus, this is a two-party agreement process where di↵erent positions and attacks among them are not considered. Also, the model assumes certain characteristics that can pose several drawbacks to its application to open MAS. On the one hand, the list of neighbours and their sensors must be known in advance. On the other hand, despite concurrency being admitted, the agents can only negotiate about one issue at the same time. Finally, the framework has strong assumptions about the honesty, cooperativeness and rationality of the agents that do not fit the reality of many real-time application domains. Opposite to Soh’s and our framework, most argumentation systems produce argu- ments by applying a set of inference rules. Rule-based systems require to elicit a explicit model of the domain. In open MAS the usual application domains are highly dynamic and the set of rules that model them is difficult to specify in advance. However, tracking the arguments that agents put forward in argumentation processes could be relatively simple. Therefore, these arguments can be stored as cases codified in a specific case representa- tion language that di↵erent agents are able to understand. This is easier than creating an explicit domain model, as it is possible to develop case-bases avoiding the knowledge- acquisition bottleneck. With these case-bases, agents are able to perform lazy learning processes over argumentation information. Another important problem with rule-based systems arises when the knowledge-base must be updated (e.g. adding new knowledge that can invalidate the validity of a rule). Updates imply to check the knowledge-base for conflicting or redundant rules. Case-based systems are easier to maintain than rule-based systems since, in the worst case, the addition of new cases can give rise to updates in some previous cases, but does a↵ect the correct operation of the system, although it can have an impact in its performance. Hence, a case-based representation of the domain knowledge of the system is more suitable for being applied in dynamic open MAS. From the first uses of argumentation in AI, arguments and cases are intertwined (Skalak and Rissland, 1992). Case-based argumentation particularly reported successful applications in American common law (Bench-Capon and Dunne, 2007). However, these models as- sumed human-computer interaction and cases were not thought to be only acceded by software agents. In MAS, the research in case-based argumentation has just a few pro- 29 posals (Heras et al., 2009a). These proposals are highly domain-specific (e.g. persuasion in negotiation (Sycara, 1990), sensor networks (Soh and Tsatsoulis, 2005) and classifica- tion (Ontañón and Plaza, 2007)) or centralise the argumentation functionality either in a mediator agent, which manages the dialogue between the agents of the system (Tolchin- sky et al., 2007), or in a specific module of the system itself (Karacapilidis and Papadias, 2001). The notion of real-time, it also appears in the literature on visualisation tools for collaborative argumentation (Kirschner, 2003, Chapters 6-8), which support human cog- nitive and discursive processes, and provide suitable representations, services and user interfaces. However, the term real-time is used here as a synonym of now and refers to produce and visualise arguments as the argumentation dialogue proceeds. The same meaning of real-time was used in the rIBIS real-time group hypertext system, which ex- tends the IBIS informal-logic argumentation framework (Rittel and Webber, 1973). This system allows a distributed set of users to simultaneously browse and edit multiple views of a hypertext network. Also, the work presented in (M. Capobianco et al., 2004) pro- poses the use of dialectical databases to comply with real-time issues when modeling agent interaction in a MAS. These databases are used to speed up the inference process in the ODeLP logic programming language by keeping track of all possible potential arguments and the defeat relation among them. Opposite to the interpretation of the real-time concept as doing things at the cur- rent time, in our approach argumentation is used to ensure that agents reach real-time agreements before a deadline is met. 6. Conclusions This work has presented a real-time argumentation framework that provides agents with the ability of engaging in argumentative dialogues and come with a solution for their underlying agreement process within a bounded period of time. In open multi-agent argumentation systems the arguments that an agent generates to support its position can conflict with arguments of other agents and these conflicts are solved by means of argumentation dialogues between them. In our proposal, a computational case-based argumentation framework that agents can use to reason about argumentation processes 30 and new case-based reasoning techniques, called Temporal Bounded CBR (TB-CBR), that adapt the CBR methodology for its use in real-time MAS have been combined to allow agents to cope with agreement processes in real-time scenarios. To evaluate the framework, it has been implemented in the domain of a customer support application. Concretely, we consider a society of agents that act in behalf of a group of technicians that must solve problems in a call centre, which control every process implicated in the provision of technological and customer support services to private or public organisations. Results shown that if the agents of the centre implement our real- time argumentation framework, they are able to temporal bound their reasoning processes and to provide suitable enough solutions within a specified deadline. However, in this paper we only cope with the temporal bounding of the argument management process to reach real-time agreements in MAS. As discussed in (G. Mahdi et al., 2010) ”there is a need of an integrated and comprehensive view of the real-time phenomenon when suggested for multi-agent systems”. Therefore, further work will ad- vance research in this area by taking into account real-time issues not only on the agents reasoning process, but also on the dialogue protocol that agents follow to interchange arguments and reach agreements. Acknowledgements This work is supported by the Spanish government grants [CONSOLIDER-INGENIO 2010 CSD2007-00022, and TIN2012-36586-C03-01] and by the GVA project [PROMETEO 2008/051]. References Amgoud, L., Prade, H., 2004. Reaching agreement through argumentation: A possibilistic approach, in: 9th International Conference on Principles of Knowledge Representation ans Reasonning, KR-04, AAAI Press. pp. 175–182. Bench-Capon, T., Dunne, P., 2007. Argumentation in artificial intelligence. Artificial Intelligence 171, 619–938. 31 Bench-Capon, T., Sartor, G., 2003. A model of legal reasoning with cases incorporating theories and values. Artificial Intelligence 150, 97–143. Choy, M., Srinivasan, D., Cheu, R., 2003. Cooperative, hybrid agent architecture for real-time traffic signal control. IEEE Transactions on Systems Man and Cybernetics Part A Systems and Humans 33, 597–607. Dean, T., Boddy, M., 1988. An analysis of time-dependent planning, pp. 49–54. Dignum, F., Weigand, H., 1995. Communication and Deontic Logic, in: Wieringa, R., Feenstra, R. (Eds.), Information Systems - Correctness and Reusability. Selected papers from the IS-CORE Workshop, World Scientific Publishing Co.. pp. 242–260. G. Mahdi, A. Gouäıch, F. Michel, 2010. Towards an integrated approach of real-time coordination for multi-agent systems, in: Proceedings of the 4th KES international conference on Agent and multi-agent systems: technologies and applications, Part I, Springer-Verlag. pp. 253–262. Heras, S., 2011. Case-Based Argumentation Framework for Agent Societies. Ph.D. thesis. Departamento de Sistemas Informáticos y Computación. Universitat Politècnica de València. http://hdl.handle.net/10251/12497. Heras, S., Botti, V., Julián, V., 2009a. Challenges for a CBR framework for argumentation in open MAS. Knowledge Engineering Review 24, 327–352. Heras, S., Botti, V., Julián, V., 2012. Argument-based agreements in agent societies. Neurocomputing 75, 156–162. Heras, S., Garćıa-Pardo, J.A., Ramos-Garijo, R., Palomares, A., Botti, V., Rebollo, M., Julián, V., 2009b. Multi-domain case-based module for customer support. Expert Systems with Applications 36, 6866–6873. Heras, S., Jordán, J., Botti, V., Julián, V., 2013a. Argue to agree: a case-based argumen- tation approach. International Journal of Approximate Reasoning 54, 82–108. Heras, S., Jordán, J., Botti, V., Julián, V., 2013b. Case-based strategies for argumentation dialogues in agent societies. Information Sciences 223, 1–30. 32 Jordán, J., Heras, S., Julián, V., 2011. A customer support application using argumenta- tion in multi-agent systems, in: 14th International Conference on Information Fusion (FUSION-11), pp. 772–778. Karacapilidis, N., Papadias, D., 2001. Computer supported argumentation and collabo- rative decision-making: the HERMES system. Information Systems 26, 259–277. Kirschner, Paul A. and Shum, S. (Ed.), 2003. Visualizing argumentation: software tools for collaborative and educational sense-making. Springer-Verlag. Kraus, S., Sycara, K., Evenchik, A., 1998. Reaching agreements through argumentation: A logical model and implementation. Artificial Intelligence 104, 1–69. Luck, M., McBurney, P., 2008. Computing as interaction: agent and agreement tech- nologies, in: IEEE International Conference on Distributed Human-Machine Systems, IEEE Press. M. Capobianco, C.I. Chesñevar, G.R. Simari, 2004. An Argument-Based Framework to Model an Agent’s Beliefs in a Dynamic Environment, in: 1sr International Workshop on Argumentation in Multi-Agent Systems (ArgMAS-04), pp. 163–178. Navarro, M., Heras, S., Julián, V., Botti, V., 2011. Incorporating temporal-bounded cbr techniques in real-time agents. Expert Syst. Appl. 38, 2783–2796. Navarro, M., Julián, V., Soler, J., Botti, V., 2004. jart: A real-time multi-agent platform with rt-java, in: 3rd IWPAAMS, pp. 73–82. Navarro, M., de Paz, J.F., Julián, V., Rodŕıguez, S., Bajo, J., Corchado, J., 2012. Tem- poral bounded reasoning in a dynamic case-based planning agent for industrial envi- ronments. Expert Systems With Applications 29, 7887–7894. O. Kahtib, 1986. Real-Time Obstacle Avoidance for Manipulators and Mobile Robots. The International Journal of Robotics Research 5, 90–98. Ontañón, S., Plaza, E., 2007. Learning and joint deliberation through argumentation in multi-agent systems, in: 7th International Conference on Agents and Multi-Agent Systems, AAMAS-07, ACM Press. 33 Parsons, S., Sierra, C., Jennings, N.R., 1998. Agents that reason and negotiate by arguing. Journal of Logic and Computation 8, 261–292. Patterson, D.W., Galushka, M., Rooney, N., 2005. Characterisation of a novel indexing technique for case-based reasoning. Artificial Intelligence Review 23, 359–393. Quinlin, J.R., 1993. C4.5 Programs for Machine Learning. Morgan Kau↵man. Rittel, H., Webber, M., 1973. Dilemmas in a general theory of planning. Policy Sciences 4, 155–169. Sierra, C., Botti, V., Ossowski, S., 2011. Agreement Computing. KI - Künstliche Intelli- genz DOI: 10.1007/s13218-010-0070-y. Skalak, D., Rissland, E., 1992. Arguments and cases: An inevitable intertwining. Artificial Intelligence and Law 1, 3–44. Soh, L.K., Tsatsoulis, C., 2005. A real-time negotiation model and a multi-agent sensor network implementation. Autonomous Agents and Multi-Agent Systems 11, 215–271. Sycara, K., 1990. Persuasive argumentation in negotiation. Theory and Decision 28, 203–242. Tolchinsky, P., Atkinson, K., McBurney, P., Modgil, S., Cortés, U., 2007. Agents de- liberating over action proposals using the ProCLAIM model, in: 5th International Central and Eastern European Conference on Multi-Agent Systems and Applications, CEEMAS-07, Springer. pp. 32–41. V. Julián, V. Botti, 2004. Developing real-time multi-agent systems. Integrated Computer-Aided Engineering 11, 135–149. Wess, S., Altho↵, K.D., Richter, M., 1993. Using k-d trees to improve the retrieval step in case-based reasoning, in: European Workshop, Topics in Case-based Reasoning, pp. 67–81. 34 
work_hmonw5eaebaelbnpnt6dmeypoy	----	Douglas Véras e Silva CD-CARS: CROSS-DOMAIN CONTEXT-AWARE RECOMMENDER SYSTEMS Universidade Federal de Pernambuco p o s g ra d u a ca o @ c i n. u fp e. b r www.cin.ufpe.br/~posgraduacao Recife 2016 Douglas Véras e Silva “CD-CARS: Cross-Domain Context-Aware Recommender Systems” A Ph.D. Thesis presented to the Centro de Informática of Universidade Federal de Pernambuco in partial fulfillment of the requirements for the degree of Philosophy Doctor in Ciência da Computação. Advisor: Prof. Dr. Carlos André Guimarães Ferraz Co-Advisor: Prof. Dr. Ricardo Bastos Cavalcante Prudêncio Recife 2016 Catalogação na fonte Bibliotecária Monick Raquel Silvestre da S. Portes, CRB4-1217 S586c Silva, Douglas Véras e CD-cars: cross domain context-aware recomender systems / Douglas Véras e Silva. – 2016. 240 f.: il., fig., tab. Orientador: Carlos André Guimarães Ferraz. Tese (Doutorado) – Universidade Federal de Pernambuco. CIn, Ciência da Computação, Recife, 2016. Inclui referências. 1. Inteligência artificial. 2. Sistemas de recomendação. 3. Filtragem colaborativa. I. Ferraz, Carlos André Guimarães (orientador). II. Título. 006.3 CDD (23. ed.) UFPE- MEI 2016-139 Douglas Véras e Silva CD-CARS: CROSS-DOMAIN CONTEXT-AWARE RECOMMENDER SYSTEMS Tese apresentada ao Programa de Pós- Graduação em Ciência da Computação da Universidade Federal de Pernambuco, como requisito parcial para obtenção do título de Doutor em Ciência da Computação. Aprovado em: 21/07/2016. ___________________________________ Prof. Carlos André Guimarães Ferraz Orientador do Trabalho de Tese BANCA EXAMINADORA _________________________________________________ Prof. Dra. Patricia Cabral de Azevedo Restelli Tedesco Centro de Informática/UFPE _________________________________________________ Prof. Dr. Kiev Santos da Gama Centro de Informática/UFPE _________________________________________________ Prof. Dr. Sérgio Ricardo de Melo Queiroz Centro de Informática/UFPE _________________________________________________ Prof. Dr. Byron Leite Dantas Bezerra Escola Politécnica/UPE _________________________________________________ Prof. Dr. Evandro de Barros Costa Instituto de Computação/UFAL I dedicate this thesis to my parents and fiancée. Acknowledgements First of all, I thank God for giving me health and strength necessary to conclude this work. Without Him the conclusion of this work would not be possible. I thank my parents, Maria Aparecida and Jailson Antônio, and brother, Matheus Véras, who always give me love and support in all my life. Also, I would like to thank my fiancée, Laura Regina, her brothers (Diego Dermeval and Amauri Junior) and her parents, Laura Maria and Amauri Campos, for their patience, comprehension, love and support. My sincere gratitude to professors Carlos Ferraz and Ricardo Prudêncio, respectively, my advisor and co-advisor, for their incentive, trust, support, knowledge transmitted and friendship. I also thank Alysson Bispo, Thiago Prota and Rafael Ferreira, my friends and colleagues at Universidade Federal de Pernambuco (UFPE) and Universidade Federal Rural de Pernambuco (UFRPE), for the incentive and help in the development of this thesis. I am very grateful to UFPE and UFRPE for the opportunity and support to develop my research in their facilities. Also, my sincere gratitude to Fundação de Amparo a Ciência e Tecnologia de Pernambuco (FACEPE) for the financial support to the development of this research. I thank to professors Byron Leite, Patricia Tedesco, Kiev Gama, Evandro Costa and Sérgio Queiroz for providing constructive reviews, corrections and suggestions in order to improve my thesis. Finally, I would like to thank all my family members and friends for the friendship and incentive as well as to all the people who contributed directly or indirectly to this research. “And if I have a prophet’s power, and have knowledge of all secret things; and if I have all faith, by which mountains may be moved from their place, but have not love, I am nothing.” (I Corinthians 13:2 - Holy Bible) Resumo Tradicionalmente, “sistemas de recomendação de domínio único” (SDRS) têm alcançado bons resultados na recomendação de itens relevantes para usuários, a fim de resolver o problema da sobrecarga de informação. Entretanto, “sistemas de recomendação de domínio cruzado” (CDRS) têm surgido visando melhorar os SDRS ao atingir alguns objetivos, tais como: “melhoria de precisão”, “melhor diversidade”, abordar os problemas de “novo usuário” e “novo item”, dentre outros. Ao invés de tratar cada domínio independentemente, CDRS usam conhecimento adquirido em um domínio fonte (e.g. livros) a fim de melhorar a recomendação em um domínio alvo (e.g. filmes). Assim como acontece na área de pesquisa sobre SDRS, a filtragem colaborativa (CF) é considerada a técnica mais popular e ampla- mente utilizada em CDRS, pois sua implementação para qualquer domínio é relativamente simples. Além disso, sua qualidade de recomendação é geralmente maior do que a dos algoritmos baseados em filtragem de conteúdo (CBF). De fato, a maioria dos “sistemas de recomendação de domínio cruzado” baseados em filtragem colaborativa (CD-CFRS) podem oferecer melhores recomendações em comparação a “sistemas de recomendação de domínio único” baseados em filtragem colaborativa (SD-CFRS), aumentando o nível de satisfação dos usuários e abordando problemas tais como: “início frio”, “esparsidade” e “diversidade”. Entretanto, os CD-CFRS podem não ser mais precisos do que os SD-CFRS. Por outro lado, “sistemas de recomendação sensíveis à contexto” (CARS) tratam de outro tópico relevante na área de pesquisa de sistemas de recomendação, também visando melhorar a qualidade das recomendações. Diferentes informações contextuais (e.g. localização, tempo, humor, etc.) podem ser utilizados a fim de prover recomendações que são mais adequadas e precisas para um usuário dependendo de seu contexto. Desta forma, nós acreditamos que a integração de técnicas desenvolvidas separadamente (de “domínio cruzado” e “sensíveis a contexto”) podem ser úteis em uma variedade de situações, nas quais as recomendações podem ser melhoradas a partir de informações obtidas em diferentes fontes além de refi- nadas considerando informações contextuais específicas. Nesta tese, nós definimos uma nova formulação do problema de recomendação, considerando tanto a disponibilidade de informações de diferentes domínios (fonte e alvo) quanto o uso de informações contextuais. Baseado nessa formulação, nós propomos a integração de abordagens de “domínio cruzado” e “sensíveis a contexto” para um novo sistema de recomendação (CD-CARS). Para avaliar o CD-CARS proposto, nós realizamos avaliações experimentais através de dois “conjuntos de dados” com três diferentes dimensões contextuais e três domínios distintos. Os resul- tados dessas avaliações mostraram que o uso de técnicas sensíveis a contexto pode ser considerado como uma boa abordagem a fim de melhorar a qualidade de recomendações de “domínio cruzado” em comparação às recomendações de CD-CFRS tradicionais. Palavras-Chave: Recomendação de Domínio Cruzado. Recomendação Sensível a Con- texto. Filtragem Colaborativa. Recomendação de Domínio Cruzado Sensível a Contexto. Abstract Traditionally, single-domain recommender systems (SDRS) have achieved good results in recommending relevant items for users in order to solve the information overload problem. However, cross-domain recommender systems (CDRS) have emerged aiming to enhance SDRS by achieving some goals such as accuracy improvement, diversity, addressing new user and new item problems, among others. Instead of treating each domain independently, CDRS use knowledge acquired in a source domain (e.g. books) to improve the recommendation in a target domain (e.g. movies). Likewise SDRS research, collaborative filtering (CF) is considered the most popular and widely adopted approach in CDRS, because its implementation for any domain is relatively simple. In addition, its quality of recommendation is usually higher than that of content-based filtering (CBF) algorithms. In fact, the majority of the cross-domain collaborative filtering RS (CD-CFRS) can give better recommendations in comparison to single-domain collaborative filtering recommender systems (SD-CFRS), leading to a higher users’ satisfaction and addressing cold-start, sparsity, and diversity problems. However, CD-CFRS may not necessarily be more accurate than SD-CFRS. On the other hand, context-aware recommender systems (CARS) deal with another relevant topic of research in the recommender systems area, aiming to improve the quality of recommendations too. Different contextual information (e.g., location, time, mood, etc.) can be leveraged in order to provide recommendations that are more suitable and accurate for a user depending on his/her context. In this way, we believe that the integration of techniques developed in isolation (cross-domain and context- aware) can be useful in a variety of situations, in which recommendations can be improved by information from different sources as well as they can be refined by considering specific contextual information. In this thesis, we define a novel formulation of the recommendation problem, considering both the availability of information from different domains (source and target) and the use of contextual information. Based on this formulation, we propose the integration of cross-domain and context-aware approaches for a novel recommender system (CD-CARS). To evaluate the proposed CD-CARS, we performed experimental evaluations through two real datasets with three different contextual dimensions and three distinct domains. The results of these evaluations have showed that the use of context-aware techniques can be considered as a good approach in order to improve the cross-domain recommendation quality in comparison to traditional CD-CFRS. Keywords: Cross-domain Recommendation. Context-Aware Recommendation. Collabo- rative Filtering Recommendation. Cross-Domain Context-Aware Recommendation. List of Figures Figure 1 – Cross-domain collaborative filtering recommendation (based on (CRE- MONESI; TRIPODI; TURRIN, 2011)(SANTOS et al., 2012)). . . . . . 25 Figure 2 – Context-aware collaborative filtering recommendation. . . . . . . . . . 26 Figure 3 – Cross-domain context-aware recommendation. . . . . . . . . . . . . . . 29 Figure 4 – “Domain” definitions according to attributes and types of recommended items (CANTADOR et al., 2015). . . . . . . . . . . . . . . . . . . . . . 39 Figure 5 – Cross-domain recommendation tasks (CANTADOR et al., 2015). . . . 40 Figure 6 – Possible scenarios of user and/or item overlap between the source and target domains (CREMONESI; TRIPODI; TURRIN, 2011). . . . . . . 43 Figure 7 – Cross-domain recommendation approaches taxonomy (CANTADOR et al., 2015). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 Figure 8 – Partitioning of data: (left) hold-out; (middle) leave-some-users-out; and (right) leaveall (CANTADOR et al., 2015). . . . . . . . . . . . . . . . . 46 Figure 9 – Paradigms for incorporating context in recommender systems (ADO- MAVICIUS; TUZHILIN, 2015). . . . . . . . . . . . . . . . . . . . . . . 55 Figure 10 – Merging user preferences approach (CANTADOR et al., 2015). . . . . . 57 Figure 11 – A contextual feature represented by dimensions, attributes and values. 70 Figure 12 – The pre-filtering cross-domain recommendation is made by filtering the target contextual user-rating tensor for a given context. . . . . . . . . . 76 Figure 13 – The cross-domain post-filtering recommendation is made over the aggre- gated user-rating matrices and then post-filtered according to contextual user preferences. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 Figure 14 – Category preferences tensor enhancement from association rules. . . . . 80 Figure 15 – The cross-domain modelling recommendation uses contextual informa- tion directly in the recommendation function as an explicit predictor of a user rating for an item. . . . . . . . . . . . . . . . . . . . . . . . . . . 81 Figure 16 – The cross-domain PreF algorithm can be used before the PostF algo- rithm in a possible combination. . . . . . . . . . . . . . . . . . . . . . . 83 Figure 17 – The cross-domain modelling algorithm can be used before the PostF algorithm in a possible combination. . . . . . . . . . . . . . . . . . . . 84 Figure 18 – Original (a) and enhanced (b) item-to-item connections. Solid circles represent items belonging to a single domain, whereas blank circles represent cross items that act as a bridge among different domains (CREMONESI; TRIPODI; TURRIN, 2011). . . . . . . . . . . . . . . . 89 Figure 19 – Example of a temporal dimension with its possible contextual attributes and values in a hierarchical view. . . . . . . . . . . . . . . . . . . . . . 94 Figure 20 – Process for gathering the location contextual information from the user information. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 Figure 21 – Example of a location dimension with its possible contextual attributes and values in a hierarchical view. . . . . . . . . . . . . . . . . . . . . . 98 Figure 22 – Example of a companion dimension with its possible contextual at- tributes and values in a hierarchical view. . . . . . . . . . . . . . . . . 101 Figure 23 – Data model class diagram focusing contextual aspects of the CD-CARS implementation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 Figure 24 – Data model class diagram focusing dataset aspects of the CD-CARS implementation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 Figure 25 – Class diagram illustrating entities used by the pre-filtering class. . . . . 114 Figure 26 – Example of the pre-filtering process considering the context of user- ratings and the recommendation context. . . . . . . . . . . . . . . . . . 115 Figure 27 – Example of selected categories in the post-filtering recommendation. . . 116 Figure 28 – A class diagram illustrating the main post-filtering entities. . . . . . . . 117 Figure 29 – Splitting training and test sets considering the target domain and context under test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 Figure 30 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: book, and target domain: television). . . . . . . . . . . . . . . . . . . . . . . . . . 133 Figure 31 – Overall prediction performance (MAE) boxplots for television domain in the temporal dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 Figure 32 – F-metric performance x top ‘N’ items for the television domain in the temporal dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 Figure 33 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal dimension (target domain: television, and source: book). . . . . . . . . . . . . . . . . . . . . . . . 136 Figure 34 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: book, and target domain: television). . . . . . . . . . . . . . . . . . . . . . . . . . 137 Figure 35 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: book, and target domain: television). . . . . . . . . . . . . . . . . . . . . . . . . . 138 Figure 36 – Overall prediction performance (MAE) boxplots for television domain in the location dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 Figure 37 – F-metric performance x top ‘N’ items for the television domain in the location dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 Figure 38 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the location dimension (target domain: television, and source: book). . . . . . . . . . . . . . . . . . . . . . . . 141 Figure 39 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: book, and target domain: television). . . . . . . . . . . . . . . . . . . . . . . 142 Figure 40 – Overall prediction performance (MAE) boxplots for television domain in the companion dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Figure 41 – F-metric performance x top ‘N’ items for the television domain in the companion dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 Figure 42 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the companion dimension (target domain: television, and source: book). . . . . . . . . . . . . . . . . . . . . . . . 145 Figure 43 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: book, and target domain: television). . . . . . . . . . . . . . . 147 Figure 44 – Overall prediction performance (MAE) boxplots for television domain in the temporal and location dimensions with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . 148 Figure 45 – F-metric performance x top ‘N’ items for the television domain in the temporal and location dimensions with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 Figure 46 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal and location dimensions (target domain: television, and source: book). . . . . . . . . . . . . . . 150 Figure 47 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: television, and target domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . 151 Figure 48 – Overall prediction performance (MAE) boxplots for book domain in the temporal dimension with different user overlap levels (source domain: television). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 Figure 49 – F-metric performance x top ‘N’ items for the book domain in the temporal dimension with different user overlap levels (source domain: television). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Figure 50 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal dimension (target domain: book, and source: television). . . . . . . . . . . . . . . . . . . . . . . . 154 Figure 51 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: television, and target domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . 156 Figure 52 – Overall prediction performance (MAE) boxplots for book domain in the location dimension with different user overlap levels (source domain: television). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 Figure 53 – F-metric performance x top ‘N’ items for the book domain in the location dimension with different user overlap levels (source domain: television). 158 Figure 54 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the location dimension (target domain: book, and source: television). . . . . . . . . . . . . . . . . . . . . . . . 159 Figure 55 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: television, and target domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . 160 Figure 56 – Overall prediction performance (MAE) boxplots for book domain in the companion dimension with different user overlap levels (source domain: television). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 Figure 57 – F-metric performance x top ‘N’ items for the book domain in the companion dimension with different user overlap levels (source domain: television). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 Figure 58 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the companion dimension (target domain: book, and source: television). . . . . . . . . . . . . . . . . . . . . . . . 163 Figure 59 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: television, and target domain: book). . . . . . . . . . . . . . . 164 Figure 60 – Overall prediction performance (MAE) boxplots for book domain in the temporal and location dimensions with different user overlap levels (source domain: television). . . . . . . . . . . . . . . . . . . . . . . . . 165 Figure 61 – F-metric performance x top ‘N’ items for the book domain in the temporal and location dimensions with different user overlap levels (source domain: television). . . . . . . . . . . . . . . . . . . . . . . . . 166 Figure 62 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal and location dimensions (target domain: book, and source: television). . . . . . . . . . . . . . . 167 Figure 63 – Predictive performance (MAE) for the algorithms by varying target domain (book and TV), contextual dimension and user overlap levels (dispersion diagram). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 Figure 64 – Predictive performance (RMSE) for the algorithms by varying target domain (book and TV), contextual dimension and user overlap levels (dispersion diagram). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 Figure 65 – Classification performance (F-metric with N=5) for the algorithms by varying target domain (book and TV), contextual dimension and user overlap levels (dispersion diagram). . . . . . . . . . . . . . . . . . . . . 172 Figure 66 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: book, and target domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Figure 67 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: book, and target domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Figure 68 – Overall prediction performance (MAE) boxplots for Music domain in the temporal dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 Figure 69 – F-metric performance x top ‘N’ items for the Music domain in the temporal dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 Figure 70 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal dimension (target domain: Music, and source: book). . . . . . . . . . . . . . . . . . . . . . . . . . 179 Figure 71 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: book, and target domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 Figure 72 – Overall prediction performance (MAE) boxplots for Music domain in the location dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181 Figure 73 – F-metric performance x top ‘N’ items for the Music domain in the location dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 Figure 74 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the location dimension (target domain: Music, and source: book). . . . . . . . . . . . . . . . . . . . . . . . . . 184 Figure 75 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: book, and target domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . 185 Figure 76 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: book, and target domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . 185 Figure 77 – Overall prediction performance (MAE) boxplots for Music domain in the companion dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 Figure 78 – F-metric performance x top ‘N’ items for the Music domain in the companion dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 Figure 79 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the companion dimension (target domain: Music, and source: book). . . . . . . . . . . . . . . . . . . . . . . . . . 188 Figure 80 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: book, and target domain: music). . . . . . . . . . . . . . . . . 189 Figure 81 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: book, and target domain: music). . . . . . . . . . . . . . . . . 189 Figure 82 – Overall prediction performance (MAE) boxplots for Music domain in the temporal and location dimensions with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 Figure 83 – F-metric performance x top ‘N’ items for the Music domain in the temporal and location dimensions with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 Figure 84 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal and location dimensions (target domain: Music, and source: book). . . . . . . . . . . . . . . . . 193 Figure 85 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: Music, and target domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . 195 Figure 86 – Overall prediction performance (MAE) boxplots for book domain in the temporal dimension with different user overlap levels (source domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 Figure 87 – F-metric performance x top ‘N’ items for the book domain in the temporal dimension with different user overlap levels (source domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 Figure 88 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal dimension (target domain: book, and source: Music). . . . . . . . . . . . . . . . . . . . . . . . . . 198 Figure 89 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: book, and target domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 Figure 90 – Overall prediction performance (MAE) boxplots for Music domain in the location dimension with different user overlap levels (source domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 Figure 91 – F-metric performance x top ‘N’ items for the book domain in the location dimension with different user overlap levels (source domain: Music). . . 201 Figure 92 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the location dimension (target domain: book, and source: Music). . . . . . . . . . . . . . . . . . . . . . . . . . 202 Figure 93 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: Music, and target domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . 203 Figure 94 – Overall prediction performance (MAE) boxplots for book domain in the companion dimension with different user overlap levels (source domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 Figure 95 – F-metric performance x top ‘N’ items for the book domain in the companion dimension with different user overlap levels (source domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 Figure 96 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the companion dimension (target domain: book, and source: Music). . . . . . . . . . . . . . . . . . . . . . . . . . 207 Figure 97 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: music, and target domain: book). . . . . . . . . . . . . . . . . 208 Figure 98 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: music, and target domain: book). . . . . . . . . . . . . . . . . 208 Figure 99 – Overall prediction performance (MAE) boxplots for book domain in the temporal and location dimensions with different user overlap levels (source domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . 209 Figure 100 –F-metric performance x top ‘N’ items for the book domain in the temporal and location dimensions with different user overlap levels (source domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . . 211 Figure 101 –Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal and location dimensions (target domain: book, and source: Music). . . . . . . . . . . . . . . . . 212 Figure 102 –Predictive performance (MAE) for the algorithms by varying target domain (book and music), contextual dimension and user overlap levels (dispersion diagram). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214 Figure 103 –Predictive performance (RMSE) for the algorithms by varying target domain (book and music), contextual dimension and user overlap levels (dispersion diagram). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 Figure 104 –Classification performance (F-metric with N=5) for the algorithms by varying target domain (book and music), contextual dimension and user overlap levels (dispersion diagram). . . . . . . . . . . . . . . . . . . . . 216 List of Tables Table 1 – Summary of techniques for representation of context (VIEIRA; TEDESCO; SALGADO, 2009). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 Table 2 – Cross-Domain CF-based RS using the Merging user preferences approach. 58 Table 3 – Classification of context-aware-based related works regarding cross- domain RS aspects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 Table 4 – Classification of context-aware-based related works with respect to CARS aspects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 Table 5 – Main limitations of context-aware-based related works in comparison to our proposed CD-CARS. . . . . . . . . . . . . . . . . . . . . . . . . . . 67 Table 6 – Classification accuracy of the companion extraction. . . . . . . . . . . . 102 Table 7 – Information gain of contextual attributes in different target domains for the book-television dataset. . . . . . . . . . . . . . . . . . . . . . . . . . 103 Table 8 – Information gain of contextual attributes in different target domains for the book-music dataset. . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 Table 9 – Cross-domain and single-domain “book-television dataset” properties with 100% of user overlap. . . . . . . . . . . . . . . . . . . . . . . . . . 105 Table 10 – “book-television dataset” properties with 50% of user overlap when “TV” is the target domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 Table 11 – “book-television dataset” properties with 10% of user overlap when “TV” is the target domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 Table 12 – “book-television dataset” properties with 50% of user overlap when “Book” is the target domain. . . . . . . . . . . . . . . . . . . . . . . . . 106 Table 13 – “book-television dataset” properties with 10% of user overlap when “Book” is the target domain. . . . . . . . . . . . . . . . . . . . . . . . . 107 Table 14 – Cross-domain and single-domain “book-music dataset” properties with 100% of user overlap. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Table 15 – “book-music dataset” properties with 50% of user overlap when “Music” is the target domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Table 16 – “book-music dataset” properties with 10% of user overlap when “Music” is the target domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 Table 17 – “book-music dataset” properties with 50% of user overlap when “Book” is the target domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 Table 18 – “book-music dataset” properties with 10% of user overlap when “Book” is the target domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 Table 19 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Temporal dimension (source domain: Book, and target domain: Television).132 Table 20 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Location dimension (source domain: Book, and target domain: Television).137 Table 21 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Companion dimension (source domain: Book, and target domain: Television). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 Table 22 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual value combina- tions from the temporal and location dimensions (source domain: Book, and target domain: Television). . . . . . . . . . . . . . . . . . . . . . . . 146 Table 23 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Temporal dimension (source domain: Television, and target domain: Book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 Table 24 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Location dimension (source domain: Television, and target domain: Book).155 Table 25 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Companion dimension (source domain: television, and target domain: book). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 Table 26 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual value com- binations from the temporal and location dimensions (source domain: Television, and target domain: Book). . . . . . . . . . . . . . . . . . . . 164 Table 27 – Overall predictive performance (MAE) of the proposed algorithms in comparison to the best baseline one by varying target domain (book and TV), contextual dimension and user overlap levels. . . . . . . . . . . . . 173 Table 28 – Overall classification performance (F-metric with N=5) of the proposed algorithms in comparison to the best baseline one by varying target domain (book and TV), contextual dimension and user overlap levels. . 174 Table 29 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Temporal dimension (source domain: Book, and target domain: Music). 175 Table 30 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Location dimension (source domain: Book, and target domain: Music). 180 Table 31 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Companion dimension (source domain: Book, and target domain: Music).184 Table 32 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual value combina- tions from the temporal and location dimensions (source domain: Book, and target domain: Music). . . . . . . . . . . . . . . . . . . . . . . . . . 188 Table 33 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Temporal dimension (source domain: Music, and target domain: Book). 194 Table 34 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Location dimension (source domain: Book, and target domain: Music). 198 Table 35 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Companion dimension (source domain: Music, and target domain: Book).203 Table 36 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual value combina- tions from the temporal and location dimensions (source domain: Music, and target domain: Book). . . . . . . . . . . . . . . . . . . . . . . . . . 207 Table 37 – Overall predictive performance (MAE) of the proposed algorithms in comparison to the best baseline one by varying target domain (book and music), contextual dimension and user overlap levels. . . . . . . . . . . . 213 Table 38 – Overall classification performance (F-metric with N=5) of the proposed algorithms in comparison to the best baseline one by varying target domain (book and music), contextual dimension and user overlap levels. 217 Contents 1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 1.1 Contextualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 1.3 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 1.4 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 1.5 Proposal Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 1.6 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 1.7 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2 BACKGROUND AND RELATED WORK . . . . . . . . . . . . . . . 32 2.1 Recommender Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.1.1 Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.1.2 User Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.1.3 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.2 Cross-Domain Recommender Systems . . . . . . . . . . . . . . . . . . 38 2.2.1 Definition of Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.2.2 Cross-Domain Recommendation Tasks . . . . . . . . . . . . . . . . . . . . 39 2.2.3 Cross-Domain Recommendation Goals . . . . . . . . . . . . . . . . . . . . 41 2.2.4 Cross-Domain Recommendation Scenarios . . . . . . . . . . . . . . . . . . 42 2.2.5 Cross-Domain Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 43 2.2.6 Cross-Domain Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 2.2.6.1 Evaluation Data Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 2.2.6.2 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 2.2.6.3 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 2.3 Context-Aware Recommender Systems . . . . . . . . . . . . . . . . . 47 2.3.1 Definition of Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 2.3.2 Modelling Contextual Information . . . . . . . . . . . . . . . . . . . . . . 48 2.3.3 Obtaining Contextual Information . . . . . . . . . . . . . . . . . . . . . . 51 2.3.4 Contextual Information Relevance . . . . . . . . . . . . . . . . . . . . . . 52 2.3.5 Context-Aware Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 54 2.3.6 CARS Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 2.4 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 2.4.1 Cross-Domain Recommendation based on Collaborative Filtering . . . . . . 56 2.4.2 Cross-Domain Recommendation based on Context-Awareness . . . . . . . . 61 2.5 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 3 CD-CARS PROPOSAL . . . . . . . . . . . . . . . . . . . . . . . . . 68 3.1 CD-CARS Problem Formalization . . . . . . . . . . . . . . . . . . . . 68 3.2 Modelling Contextual Information . . . . . . . . . . . . . . . . . . . . 69 3.2.1 Contextual Features Formalization . . . . . . . . . . . . . . . . . . . . . . 69 3.2.2 Obtaining and Selecting Relevant Contextual Information . . . . . . . . . . 72 3.3 CD-CARS Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 3.3.1 Proposed Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 3.3.1.1 Cross-Domain PreF Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 75 3.3.1.2 Cross-Domain PostF Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 76 3.3.1.3 Cross-Domain Modelling Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 80 3.3.1.4 Cross-Domain Hybrid Contextual Algorithms . . . . . . . . . . . . . . . . . . . . 82 3.3.2 Base Cross-Domain Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 83 3.3.2.1 Single-Domain as Cross-domain Algorithms . . . . . . . . . . . . . . . . . . . . . 84 3.3.2.1.1 Neighborhood-based Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 3.3.2.1.2 Matrix factorization algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 3.3.2.2 Cross-Domain Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 3.4 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 4 CD-CARS IMPLEMENTATION . . . . . . . . . . . . . . . . . . . . 92 4.1 Dataset Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 4.1.1 Obtaining Contextual Information . . . . . . . . . . . . . . . . . . . . . . 94 4.1.1.1 Temporal Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 4.1.1.2 Location Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 4.1.1.3 Companion Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 4.1.2 Selecting Relevant Contextual Attributes and Values . . . . . . . . . . . . 102 4.1.3 Cross-Domain Datasets Description . . . . . . . . . . . . . . . . . . . . . 104 4.1.3.1 Book-Television dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 4.1.3.2 Book-Music dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 4.2 Contextual Model Implementation . . . . . . . . . . . . . . . . . . . . 108 4.3 Proposed Algorithms Implementation . . . . . . . . . . . . . . . . . . 112 4.3.1 Pre-filtering Implementation . . . . . . . . . . . . . . . . . . . . . . . . . 112 4.3.2 Post-filtering Implementation . . . . . . . . . . . . . . . . . . . . . . . . . 115 4.4 Base Cross-domain Algorithm Implementation . . . . . . . . . . . . . 123 4.5 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 5 CD-CARS EVALUATION . . . . . . . . . . . . . . . . . . . . . . . . 127 5.1 Evaluation Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . 127 5.1.1 Settings of the Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 127 5.1.2 Predictive Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 5.1.3 Classification Performance . . . . . . . . . . . . . . . . . . . . . . . . . . 129 5.1.4 Sensitivity Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 5.1.5 Statistical Significance Analysis . . . . . . . . . . . . . . . . . . . . . . . 131 5.2 Evaluation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 5.2.1 Book-Television Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 5.2.1.1 Television as Target Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 5.2.1.1.1 Temporal Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 5.2.1.1.2 Location Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 5.2.1.1.3 Companion Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 5.2.1.1.4 Combining Contextual Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 5.2.1.2 Book as Target Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 5.2.1.2.1 Temporal Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 5.2.1.2.2 Location Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 5.2.1.2.3 Companion Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 5.2.1.2.4 Combining Contextual Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 5.2.1.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 5.2.2 Book-Music Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 5.2.2.1 Music as Target Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 5.2.2.1.1 Temporal Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 5.2.2.1.2 Location Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 5.2.2.1.3 Companion Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 5.2.2.1.4 Combining Contextual Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 5.2.2.2 Book as Target Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 5.2.2.2.1 Temporal Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 5.2.2.2.2 Location Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 5.2.2.2.3 Companion Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 5.2.2.2.4 Combining Contextual Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 5.2.2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210 5.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 5.3 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220 6 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222 6.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222 6.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 6.3 Lines for Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . 224 REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 23 1 Introduction In this chapter, we contextualize and motivate the problem addressed in this thesis, respectively, in Section 1.1 and Section 1.2. The problem statement is described in Section 1.3. In Section 1.4, we define the objectives of this thesis and in Section 1.5 we describe an overview of the proposal. At last, we highlight the expected contributions and describe the structure of this thesis (Section 1.7). 1.1 Contextualization In the last years, the growth of the Internet has increased the amount of information available for users. Consequently, the task of finding relevant information among a myriad of options became a problem. This problem is traditionally known as the information overload problem (RESNICK et al., 1994)(HILL et al., 1995)(SHARDANAND; MAES, 1995)(ADOMAVICIUS; TUZHILIN, 2005)(RICCI; ROKACH; SHAPIRA, 2011). Taking into account the variety of information provided by applications on the Internet, we can consider information as any item capable of being consumed and rated by a user (e.g. movies, books, music, and so on). Given that, the information overload problem makes the process of finding relevant items in an extensive amount of options difficult for a user. Fortunately, this problem has received notable attention of researchers in the Artificial Intelligence area. In this way, recommender systems (RSs) have been designed in order to solve this problem (ADOMAVICIUS; TUZHILIN, 2005)(RICCI; ROKACH; SHAPIRA, 2011). For example, a recommender system (RS) can be used for suggesting to users interesting movies to watch, books to read, music to listen to, etc. The suggestions (recommendations) are provided according the users’ profile, which could be inferred from the consumption log, for instance. Recently, a large number of Web sites and applications have adopted recommender systems to provide their users with more relevant items, such as Amazon1, Netflix2, Youtube3, Last.fm 4, BookCrossing5, Buscape6, among many others. However, most of these systems are developed to recommend items in a specific domain such as movies, videos, music, books, and so on. Thus, they are known as single-domain RS, since they just consider the user profile from a unique domain to recommend relevant items in the same domain. For example, a single-domain RS could recommend a movie based on the 1 Amazon e-commerce site, http://www.amazon.com 2 Netflix on-demand streaming media site, http://www.netflix.com 3 YouTube video sharing site, https://www.youtube.com 4 Last.fm on-line radio, http://www.last.fm 5 BookCrossing on-line book exchange site, http://bookcrossing.com 6 Buscape helps users to find best prices for products and services, http://www.buscape.com.br 24 Chapter 1. Introduction previous movies watched or recommend a book similar to other ones read by a user. It is important to mention that despite there are several definitions of “domain”, our notion of domain can be expressed as a “kind of item” (e.g. movies, music, books, news, among others) (FERNÁNDEZ-TOBÍAS et al., 2012). Although single-domain RSs have achieved a good quality in suggesting relevant items for users, some issues still are significant for the information overload problem (RICCI; ROKACH; SHAPIRA, 2011), such as: • cold-start: it is related to situations in which a RS is unable to generate recommen- dations due to an initial lack of user preferences; • sparsity: when the average number of ratings per user and item is low, which may negatively affect the quality of the recommendations, a sparsity problem might exist; • diversity: when very similar or redundant items are recommended, which may not satisfy the users; • accuracy: even when the issues mentioned above are resolved, the RS may still not be accurate, which means that incorrect predictions of ratings may be made as well as a list of recommended items may not satisfy a user; among others. In order to alleviate these problems, cross-domain recommender systems (WINOTO; TANG, 2008)(CREMONESI; TRIPODI; TURRIN, 2011) have arisen, aiming to improve the quality of single-domain recommendations (FERNÁNDEZ-TOBÍAS et al., 2012). Instead of treating each domain independently, cross-domain RSs use knowledge acquired in a source domain (e.g. books) to improve the recommendation of a target domain (e.g. movies). To illustrate this, consider a user with no information about his/her favorite movie genres. This lack of information can be inferred either from his/her favorite book genres or from his/her similarity with users across different domains, for example. One of the first studies on this emerging research topic was that presented in (WINOTO; TANG, 2008), which investigated if the consumption behaviors on related items from different domains could be useful to make recommendations in a target domain. As shown by Winoto e Tang (2008), joint recommendations of items from multiple domains may be less accurate, but more diverse than recommendations of items in a single domain. Since the first cross-domain recommender systems have arisen, several approaches have been proposed to deal with different goals (FERNÁNDEZ-TOBÍAS et al., 2012). For instance, knowledge-based recommender systems try to exploit knowledge about users and items besides the relationship between them in order to produce recommendations (TREWIN, 2000). In this way, knowledge-based recommender systems demand a great amount of knowledge about users and items (and their domains), which should be stored and organized in a way that enables inferring and reasoning. However, that knowledge 1.2. Motivation 25 acquisition is a very difficult process and a knowledge engineer is required to construct the knowledge-base which causes a bottle-neck for the knowledge-based recommender systems (AZAK, 2010). As described above, the knowledge-based recommender systems rely on an “ad-hoc” approach, which may be difficult to customize to new situations, once that they are usually designed for a specific domain (TREWIN, 2000). However, other approaches have been successfully adopted for cross-domain RS and, in general, require little domain knowledge (FERNÁNDEZ-TOBÍAS et al., 2012), since they are based on simple information obtained from user ratings. Fernández-Tobías et al. (2012) stated that domains can be explicitly or implicitly linked by means of content-based (CBF) or collaborative filtering (CF) characteristics asso- ciated with users and/or items, such as ratings, social tags, and latent factors. Cremonesi, Tripodi e Turrin (2011) surveyed and categorized cross-domain collaborative filtering recommender systems (CD-CFRS), which recommend items from target domain explor- ing the similarities between users considering ratings from source and target domains, as illustrated in Figure 1. Likewise single-domain RS research, collaborative filtering is considered the most popular and widely implemented approach in cross-domain RS, because its implementation and integration in existing domains is relatively easy as well as its quality is generally higher than other approaches (ADOMAVICIUS; TUZHILIN, 2005)(FERNÁNDEZ-TOBÍAS et al., 2012). Figure 1 – Cross-domain collaborative filtering recommendation (based on (CREMONESI; TRIPODI; TURRIN, 2011)(SANTOS et al., 2012)). 1.2 Motivation In fact, the majority of the CD-CFRS can give better recommendations in compar- ison to single-domain RS, leading to a higher user satisfaction and addressing cold-start, sparsity, and diversity problems, but not necessarily, they can be more accurate than single- domain collaborative filtering RS (WINOTO; TANG, 2008)(FERNÁNDEZ-TOBÍAS et al., 26 Chapter 1. Introduction 2012). Meanwhile, context-aware recommender systems (CARS) is another relevant topic of research in RS and has been used in order to enhance the quality of recommendations Adomavicius e Tuzhilin (2015), especially for providing accurate recommendations taking into account the user’s context. The context-aware approach uses different contextual information (e.g., loca- tion, time, mood, etc.) to improve the accuracy of recommendations (ADOMAVICIUS; TUZHILIN, 2015), as illustrated in Figure 2. In many applications, such as recommending a vacation package, recommending a TV program, among others, it may not be sufficient to consider only users and items, it is also important to incorporate contextual information into the recommendation process in order to recommend items to users under certain circumstances (ADOMAVICIUS; TUZHILIN, 2015). For example, using the temporal context, a travel recommender system could provide a recommendation in the winter that can be very different from the one in the summer. More specifically, on weekdays a user might prefer to watch news programs when he/she turns on his/her TV in the morning, or to watch soccer games at night, and on weekends to watch comedy movies. Figure 2 – Context-aware collaborative filtering recommendation. Researchers in the recommender systems area have recognized the importance of contextual information (ADOMAVICIUS; TUZHILIN, 2015). In addition, in the cross- domain RS field, Fernández-Tobías et al. (2012) and Cantador et al. (2015) highlighted that context can be treated as a bridge between different domains and just a few works 1.3. Problem Statement 27 have considered context-aware techniques in cross-domain recommender systems. The use of context-aware techniques in cross-domain RS is an interesting open research direction, since the majority of the works about cross-domain recommender systems only adopt CBF and CF approaches, both considering only users and considering items attributes, without taking into account any additional contextual information. Accurate prediction of user preferences undoubtedly depends upon the degree to which the recommender system has incorporated the relevant contextual information into a recommendation method (ADOMAVICIUS et al., 2005)(ADOMAVICIUS; TUZHILIN, 2015). In this way, using a context-aware approach in cross-domain recommender systems may be useful to make suggestions under difficult situations such as in the cold-start, sparsity and diversity problems while improving the accuracy of recommendations, in comparison to traditional CD-CFRS, by using context and knowledge from different domains. We believe that the integration of techniques developed in isolation to cross-domain and context-aware RS can be useful in a variety of situations, in which recommendations can be improved by information from different sources and can be refined by considering specific contextual information. 1.3 Problem Statement In this thesis, we address the cross-domain recommendation problem under the CF and context-awareness perspectives. Thus, besides considering user ratings, we also consider the context of these ratings in the recommendations, which implies one more dimension (context) in the user-rating matrices, or in this case, user-rating-context tensors (User x Item x Context). In our cross-domain context-aware recommendation problem, there exists a user-rating-context tensor for each domain (source and target). In both source and target tensors, there is no item in common (no item overlap), but the same set of contexts is observed (context overlap) and at least a user must have ratings in both tensors (user overlap). Therefore, our problem is to explore the user-rating-context tensors from source and target domains to improve recommendations in the target domain, i.e., to improve the estimation of unknown ratings for items in a target domain by exploiting the user- rating-context tensors from these domains. Based on the above problem, we state the hypothesis of this thesis: The application of context-aware techniques can improve the accuracy of cross-domain collaborative filtering recommendations. It is important to mention that accuracy, in that hypothesis, is one quality aspect 28 Chapter 1. Introduction of recommender systems. However, other aspects can also be addressed in this thesis such as alleviating the cold-start and sparsity issues, or improving the recommendation diversity or coverage, but they are not our main goals. 1.4 Objectives The main objective of this thesis is to improve the accuracy of cross-domain collaborative filtering recommendations by adding context-aware techniques. In order to achieve this, we aim to build a cross-domain context-aware recommender system (CD- CARS) through the proposal of novel algorithms for cross-domain recommendations. Specific objectives of this thesis are: • To allow that traditional CF-based algorithms can be used in combination with the proposed algorithms. • To allow that the proposed algorithms can be extended and configured. • To allow that cross-domain context-aware recommendations can be made either to more related domains (e.g. Book and Television) or less related ones (e.g. Book and Music) among themselves. We consider this relation among distinct domains according to the set of item genres of them. As more the domains have item genres in common the more related they are considered (e.g. Book and Television have several item genres in common such as “romance”, “educational”, “religion”, etc.). • To provide a solution for identifying the most relevant contextual features that can be used for a specific domain. • To build a method responsible for generating contextual user profiles on different domains. • To develop a solution for extracting relevant information (e.g. contextual information, user profiles in different domains, etc.) for the datasets used in the CD-CARS evaluation. 1.5 Proposal Overview As mentioned previously, this thesis aims to improve the quality of cross-domain collaborative filtering recommendations by adding context-aware techniques. In order to achieve this, it is necessary to understand current approaches available in the literature, especially, about two emerging research areas: context-aware recommender systems (CARS) and cross-domain recommender systems (CDRS). 1.5. Proposal Overview 29 After searching and exploring several approaches in these areas, we propose CD- CARS algorithms based on distinct context-aware paradigms (pre-filtering, post-filtering, and modelling) (ADOMAVICIUS; TUZHILIN, 2015) combined with a CD-CFRS aiming to improve its quality (accuracy of the recommendations). Thus, this CD-CFRS is transformed into a CD-CARS which recommends items from target domain exploring the similarities between users considering ratings, and also their contexts, from source and target domains, as illustrated in Figure 3. The CD-CFRS algorithms adopted in the proposal belong to two traditional CF-based categories: neighborhood-based and matrix factorization. Figure 3 – Cross-domain context-aware recommendation. To illustrate the CD-CARS, suppose that a user X, who enjoys to read romance books on weekdays and does not have any preference known about movies, is very similar to another user Y that also enjoys romance books on weekdays and likes to watch action movies on weekdays and comedy movies on weekends, so, a CD-CARS could prioritize movies enjoyed by the user Y on the top of the recommended item list for the user X in those particular contexts (comedy movies on weekends and action movies on weekdays), just by knowing the book preferences from user X without his/her movie preferences. Although many cross-domain RS have used contextual information in an ad- hoc way as part of their knowledge-based approaches (see Section 2.4.2) (BLANCO- FERNÁNDEZ et al., 2011)(MOE; AUNG, 2014b)(KAMINSKAS et al., 2014), to the 30 Chapter 1. Introduction best of our knowledge, there are not works in the literature addressing the cross-domain recommendation task by means of contextual features using systematic paradigms (pre- filtering, post-filtering, and modelling) (FERNÁNDEZ-TOBÍAS et al., 2012)(CANTADOR et al., 2015). Those knowledge-based cross-domain RSs may be difficult to customize to new situations (domains) and to compare their performances with other approaches. For instance, the knowledge-based framework proposed in (KAMINSKAS et al., 2014) is specific for the considered domains (points of interest and music). Our CD-CARS, in turn, relies on the use of systematic context-aware approaches (ADOMAVICIUS; TUZHILIN, 2015), which have been successfully adopted for single domain recommendation and in general require just a little knowledge about the domain (e.g user ratings) and can be customized to new domains in a simpler way. 1.6 Contributions The contributions of this thesis are multiple, mainly in the recommender systems area. The main ones are listed below: 1. The formalization of the cross-domain context-aware recommendation problem from the survey of two emergent research fields: cross-domain and context-aware RS; 2. Improving the quality of CD-CFRS, through the realization of a CD-CARS by the proposal of novel algorithms based on three distinct and systematic paradigms of context-aware recommendation, which were chosen rather than ad-hoc context-aware approaches; 3. Provision of real datasets for evaluating CD-CARS, taking into account different domains and contextual information; 4. Providing a CD-CARS that can be useful to recommend items from any domain (e.g. books, music, movies, etc.), which allows generating cross selling or bundle recommendations for items from multiple domains (e.g. the recommendation of a music accompanied of a movie to watch or a book to read). Through the findings of this thesis, we expect to contribute for the cross-domain RS area towards future research and challenges in cross-domain context-aware recommen- dations. 1.7 Thesis Outline This thesis is structured as follows: 1.7. Thesis Outline 31 • Chapter 1 introduces and states the cross-domain context-aware problem, describes the proposal of this thesis as well as its objectives and contributions. • Chapter 2 reviews the literature about recommender systems focusing on cross- domain RS and context-aware RS. Also in this chapter, we compare related work to our thesis. • Chapter 3 presents the proposed CD-CARS. For that, we describe the formalization of the CD-CARS problem, how the contextual information is modelled, the proposed recommendation algorithms, and cross-domain CF-based algorithms adopted in combination with the CD-CARS algorithms. • Chapter 4 describes particular details of an implementation of the CD-CARS pro- posal. • Chapter 5 presents the results of an experimental evaluation of the implemented CD-CARS as well as a discussion about the findings of this research. Besides, we describe details about the experiments’ settings and evaluation metrics. • Chapter 6 presents the conclusions, limitations and future works of this thesis. 32 2 Background and Related Work In this chapter, we provide an explanation of concepts related to this thesis. Initially, we describe the main concepts of recommender systems (Section 2.1) such as approaches, algorithms, user profiling, and evaluation metrics. Following, we introduce concepts of cross-domain recommender systems (Section 2.2) and their most common approaches. Finally, we describe the foundations of context-aware recommender systems (Section 2.3). These concepts are necessary as background for understanding the proposed CD- CARS and for positioning the proposal of this thesis in the state-of-the-art research. However, since the union of cross-domain and context-aware fields has not been deeply explored, we also describe and classify cross-domain RS and CD-CARS, both related to the proposal of this thesis (Section 2.4). 2.1 Recommender Systems In recent years, recommender systems (RSs) have been crucial for dealing with the information overload problem (ADOMAVICIUS; TUZHILIN, 2005). This issue is related to the explosive growth and variety of information available on the Web, which frequently has overwhelmed users with a large myriad of options. RSs are software tools and techniques providing suggestions of relevant items to users (RESNICK; VARIAN, 1997)(BURKE, 2002)(BURKE, 2007). The suggestions relate to various decision-making processes, such as what products to buy, what music to listen to, or what TV programs to watch. Therefore, recommender systems can help people to identify contents of their interest among a large set of options available. These systems became an important research area since the publication of landmark papers in the 1990’s, when the term “collaborative filtering” was coined (RESNICK; VARIAN, 1997). Since then, the number of research papers published has increased significantly in many application fields (books, documents, images, movies, music, shopping, TV programs, and others) (PARK et al., 2012), as well as the amount of commercial applications of recommender systems by large companies such as Amazon.com (LINDEN; SMITH; YORK, 2003), Google (DAS et al., 2007), Last.fm (EYKE, 2009), Netflix (BENNETT; LANNING, 2007), among others. For a better understanding of RSs, we describe some perspectives of them in the following subsections. 2.1. Recommender Systems 33 2.1.1 Strategies There are several RS strategies (or approaches) in the literature. A strategy can be seen as a type or category of RS, and it may vary according to the paradigm of its recommendation algorithm, i.e., how the recommendation is made. This variation of strategies leads to different classifications of RS. Below, we describe eight categories of RS, which are based on (BURKE, 2007), (RICCI; ROKACH; SHAPIRA, 2011) and (VÉRAS et al., 2015) classifications. 1. Non-personalized: Non-personalized recommender algorithms present any user a predefined list of items. Such algorithms usually serve as a baseline for more advanced personalized algorithms. For example, one non-personalized algorithm, called Top Popular (TopPop), recommends the top-N items (e.g. movies) with the highest popularity (largest number of ratings) (CREMONESI; GARZOTTO; TURRIN, 2012). 2. Content-based filtering (CBF): Content-based recommendation systems try to recom- mend similar items to those that a user has liked or consumed in the past (PAZZANI, 1999). Indeed, the basic process performed by a content-based recommender consists in matching up the attributes of a user profile, in which preferences and interests are stored, with the attributes of a item’s content, in order to recommend to that user new interesting items. For example, TV contents that are similar (based on their genres, actors, and so on) to those the user preferred in the past are recommended. Since such RSs tend to recommend items with the same characteristics as the ones that a user liked in the past, the recommended items typically lack novelty, meaning the RS proposes a limited variety of unexpected (but relevant) recommendations (ADOMAVICIUS; TUZHILIN, 2005). 3. Collaborative filtering (CF): Collaborative recommender systems ignore content and exploit collective preferences of the crowd, i.e., they generate recommendations using different users’ rating profiles and suggest items that other users with similar tastes liked in the past (PAZZANI, 1999). The degree to which two users’ preferences are considered similar is based on a similarity measure of their rating histories. This approach can be illustrated by the expression: “people who watched this TV program also watched...”. In addition, CF is considered the most popular and widely implemented approach in RS, because its implementation and integration in existing domains are relatively easy, and its quality is usually higher than that of CBF algorithms. A common criticism of CF recommenders is that they tend to be biased toward popularity, constraining the degree of diversity, so, they are not able to recommend unrated items (related to cold-start and sparsity issues) (ADOMAVI- CIUS; TUZHILIN, 2005). The “community-filtering” strategy can be considered a 34 Chapter 2. Background and Related Work specialization of the collaborative filtering one (KAMAHARA et al., 2005). The “community-filtering” strategy recommends items based on the preferences of the user’s friends (or friend of a friend)(KAMAHARA et al., 2005)(BOURKE; MC- CARTHY; SMYTH, 2011)(HAN et al., 2015). Evidence suggests that people tend to rely more on recommendations from their friends than on recommendations from similar but anonymous individuals (KAMAHARA et al., 2005). This observation, combined with the growing popularity of social networks, is generating a rising interest in community-based systems or, as or as they usually referred to, social recommender systems (KAMAHARA et al., 2005). 4. Data mining: Significantly, many researchers have used data mining techniques to improve the recommender systems performance (PARK et al., 2012). Data mining techniques are defined as extracting or mining knowledge from data. These techniques are used for the exploration and analysis of a large amount of data in order to discover meaningful patterns and rules. They can be used to lead decision making and to predict the effect of decisions. For example, TV programs can be classified into two classes: “watched” and “not watched” and a user profile is then a collection of attributes together with the number of times they occur in positive and negative examples. Hence, the RS computes prior probabilities that a TV program has to belong to one of those two classes and the conditional probability that a feature is present if a TV program is classified into either positive or negative class. It must be noted that features are, in this case, related to content (e.g. genre) or not (e.g. time of the day). 5. Context-awareness: Context-awareness aims to give to its applications advantages in the use of contextual information, such as the user’s location, to offer proactive services to the user without any explicit request (ABOWD et al., 1999). In this way, a personalization system based on context-awareness could adapt its functionality or behavior so that it reacts differently depending on the user’s context (location, friends, family, time, among others) and the resources available at that moment, in accordance with his/her personal preferences (RICCI; ROKACH; SHAPIRA, 2011). For example, a RS could recommend a TV program that suits the user and his current situation: staying at home or on the train, in the noon or the evening, being in front of his TV or smartphone. For each situation (context), the user’s preference may be different, so, some contextual patterns could be found and exploited by recommendation algorithms, e.g. knowing that a user particularly likes watching sports in the evening time, at home (MOON et al., 2009). 6. Semantic-based: Semantic Web is based on describing Web resources by semantic an- notations (meta-data), formalizing these annotations in an ontology and applying rea- soning processes aimed at discovering new knowledge (BERNERS-LEE; HENDLER, 2.1. Recommender Systems 35 2001). The synergy between recommender systems and Semantic Web has already been explored in many domains (including the TV domain), showing significant in- creases in the recommendation accuracy (BLANCO-FERNÁNDEZ; PAZOS-ARIAS, 2008). Instead of employing traditional syntactic approaches, Semantic-based strat- egy discovers semantic relationships between the users’ preferences and the items available in the domain ontology through semantic similarity metrics. For example, using the semantic approach, a RS could recommend a place to visit (e.g. offering a tourist package) according to the places showed in a movie or sports game that a user liked (BLANCO-FERNÁNDEZ et al., 2011). The approaches described above focused on making recommendations for individual users and do not consider the problem of group recommendation. The problem of group recommendation has also been investigated recently (QUEIROZ; CARVALHO, 2004)(RICCI; ROKACH; SHAPIRA, 2011). Various techniques have been proposed, targeting different types of recommendation items (e.g., movie, TV program, music) and different groups (e.g., family, friends, dynamic social groups). Most group recommendation techniques consider the preferences of individual users and propose various strategies to either combine the individual user profiles into a single group profile (a pseudo user) and make recommendations for that pseudo user (BRUSILOVSKY; KOBSA; NEJDL, 2007), or generate recommendation lists for individual group members and merge the lists for group recommendation(MARILLY et al., 2011). This kind of approach is usually adopted in the TV domain, because watching TV activity is, traditionally, performed by a group of people (e.g. a family). For example, a TV program could be recommended according to the average rating of a group of users (based on their individual preferences) (BRUSILOVSKY; KOBSA; NEJDL, 2007). 2.1.2 User Profiling The user profile, which usually is composed of preferences and personal character- istics, is one of the most important aspects of the recommendation process. Recommender systems necessarily make use of user profiles in order to recommend items related to those profiles. However, there are different approaches for creating a user profile. Each approach has its benefits and its limitations (UBERALL; MUTTUKRISHNAN, 2009). Therefore, this is an important perspective of any recommender system, which can be categorized into three categories as follows (VÉRAS et al., 2015): 1. Explicit Profiling: An explicit profile can be created by a user in the first time that he/she logs in the recommendation system. In this case, users set their preferences (interests) such as favorite TV shows or genre of TV shows, favorite actors of movies, favorite channels, ratings for movies (e.g. rating “four” for a TV show on a scale of 36 Chapter 2. Background and Related Work zero to five), among others. Furthermore, users can modify any information of their profiles at any moment through the system (UBERALL; MUTTUKRISHNAN, 2009). However, the explicit profiling approach could bother and tire users (REICHLING; WULF, 2009) in order to fulfill their profiles every time that they find something interesting on TV, for example. 2. Implicit Profiling: On other hand, an implicit profile can be created automatically by the recommender system. In this case, the RS logs and saves the viewing behaviour of a user (UBERALL; MUTTUKRISHNAN, 2009). Through the user’s log, like watched programs (watching time, watching duration, genre, etc.), his/her preferences (interests) are inferred (instead of explicitly set). This inference may result in user’s favorite TV shows (or genre of TV shows), favorite actors of movies, favorite channels, and even a rating for a movie (e.g. calculated using the ratio between the watching duration and the TV show duration), among others. Sometimes, this approach could have the problem of incorrectly expressing the user profile (HU; KOREN; VOLINSKY, 2008), because the user could be sleeping or doing something else while TV is on, for example. 3. Contextual Profiling: The “contextual profile” usually is generated by “Context- Aware Recommendation Systems”(ABBAR; BOUZEGHOUB; LOPEZ, 2009). In this approach, the user’s profile is created through the relationship between contexts and “common” user profiles (explicit or implicit profiling). Thus, the recommendation process is based on the contextual profile, which contains contextual information besides user preferences (MUKHERJEE et al., 2011). Users’ contextual information can be obtained explicitly or implicitly (most common), such as location, friends, family members, watching day/time, activity, and so on. Therefore, according to the manner that the contextual information is obtained - explicitly or implicitly, the same issues from these approaches are applied for the contextual profile. 2.1.3 Evaluation Evaluation of recommender systems is fundamental in assessing the quality of their recommendations. However, many different measures have been defined in the literature with the aim of making better choices in general or for a specific application area. Likewise the RS algorithms, we describe evaluation metrics in a top-level way, as follows (VÉRAS et al., 2015): 1. Qualitative measures: these measures are used when we want a model to minimize the number of errors. Hence, these metrics are usual in many direct applications of recommenders. Inside this category, some of these measures are more appropriate 2.1. Recommender Systems 37 for some kinds of recommenders, predictors or information retrieval tasks. Exam- ples of these measures are Accuracy (LEE; YANG, 2003), F-measure (LEKAKOS; GIAGLIS, 2004), Coverage (LEKAKOS; CARAVELAS, 2008), Diversity (CRE- MONESI; TURRIN, 2010), among others (RICCI; ROKACH; SHAPIRA, 2011). During the evaluation, items are commonly labeled as relevant or irrelevant for a user and a metric is adopted to measure the quality of the items once classified by the RSs. The recommendation problem is treated as a classification task. 2. Probabilistic (Predictive) measures: these measures are especially useful when we want an assessment of the reliability of the predictions returned by a RS, whether they have recommended a non-relevant item with high or low probability. The main examples of these measures are Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) (SHANI; GUNAWARDANA, 2011)(HERLOCKER et al., 2004). The recommendation problem, in this case, is usually treated as a regression task when the actual rate of an item is compared to a rate or score predicted by the RS. 3. Ranking (Classification) measures: these measures are very common in the RSs area because they are based on “how well” recommender systems rank the rec- ommended items. Thus, there are many examples of evaluation metrics in this category, such as Precision and Recall curves (ZHIWEN; XINGSHE, 2003), Nor- malized Discounted Cumulative Gain (NDCG) (BALTRUNAS; MAKCINSKAS; RICCI, 2010), Mean Average Precision (MAP) (HOPFGARTNER; JOSE, 2010), hit rate (HR) (O’SULLIVAN; SMYTH; WILSON, 2004), Fall-out (JOJIC; SHUKLA; BHOSAREKAR, 2011), Area under the ROC Curve (AUC) (ZHANG; ZHENG, 2005), Breese score (BREESE; HECKERMAN; KADIE, 1998), among others (RICCI; ROKACH; SHAPIRA, 2011). Differently from the previous category, these metrics assess the quality of a ranking of items returned by the RS instead of the average quality of the raw scores returned by the RS. The recommendation problem is treated in this case as a ranking task. 4. User satisfaction: in this category, there are empirical experiments with users in order to verify their satisfaction about the RS (BLANCO-FERNáNDEZ et al., 2008). Although this method is used in many RSs and collects personal feedback of users (LóPEZ-NORES et al., 2009), this kind of evaluation may have some problems, such as biases, lack of an objective measure for RS quality assessment, comparison between different systems, and so on (RICCI; ROKACH; SHAPIRA, 2011). 38 Chapter 2. Background and Related Work 2.2 Cross-Domain Recommender Systems Nowadays, the majority of recommender systems provide recommendations for items belonging to a single domain. For instance, Netflix recommends movies (BENNETT; LANNING, 2007), Last.fm recommends songs (EYKE, 2009), among others. These single- domain recommender systems have been successfully adopted by several websites, however, some of them such as Amazon1 and eBay2 usually maintain user preferences for items from multiple domains. In addition, it is common that users of social networks provide their preferences and interests for a variety of items from distinct domains (e.g. music, books, movies, etc.) (SHAPIRA; ROKACH; FREILIKHMAN, 2013). Leveraging all the user preferences available in several systems or domains may be useful for generating better recommendations, e.g., by alleviating the cold-start and sparsity problems in a target domain, or by providing recommendations for items from multiple domains. Thus, cross-domain recommender systems aim to generate or improve recommendations in a target domain (e.g. music, etc.) by exploiting knowledge from a source domain (e.g. books, etc.) (FERNÁNDEZ-TOBÍAS et al., 2012). The cross-domain approach is a challenging and emergent field of recommender sys- tems (CANTADOR et al., 2015). Although it has been addressed from distinct perspectives, there are several distinct definitions of the cross-domain recommendation task. According to surveys about cross-domain RS (FERNÁNDEZ-TOBÍAS et al., 2012)(CANTADOR et al., 2015), we describe some of its perspectives in the following subsections. 2.2.1 Definition of Domain In the literature, researchers have considered different definitions of “domain”. For instance, some of them have considered items like movies and books as belonging to distinct domains, while others have considered different item genres as different item domains (e.g. “action movies” and “comedy movies”). Cantador et al. (2015) define the “domain” concept regarding the attributes and types of recommended items. They consider that domain may be defined at four levels (see Figure 4): • (Item) Attribute level. Recommended items have the same type and the same attributes, but they differ in the value of a certain attribute. For instance, two movies of different genres (e.g. “action movies” and “comedy movies”) belong to distinct domains (CAO; LIU; YANG, 2010). • (Item) Type level. In this level, recommended items have similar types and have some attributes in common. For example, movies and TV programs belong to distinct 1 http://www.amazon.com 2 http://www.ebay.com 2.2. Cross-Domain Recommender Systems 39 domains, since they have some attributes in common (title, genre, etc.), but they also have different ones (e.g., airtime, channel, etc.) (HU et al., 2013)(LONI et al., 2014). • Item level. Recommended items have different types and attributes (or the majority of them). For instance, movies and books belong to distinct domains even with some attributes in common (title, release/publication year, etc.) (GAO et al., 2013)(ENRICH; BRAUNHOFER; RICCI, 2013). • System level. In this level, recommended items are from different systems, which are considered as distinct domains. For example, a user could rate a movie in MovieLens3 as well as in Netflix4 (PAN; XIANG; YANG, 2012)(PAN; YANG, 2013). Figure 4 – “Domain” definitions according to attributes and types of recommended items (CANTADOR et al., 2015). It is important to mention that the notion of domain adopted in this thesis is based on the Item level, considering that movies and books belong to different domains, for example. 2.2.2 Cross-Domain Recommendation Tasks In the literature about cross-domain RSs, the works usually aim to exploit knowledge from a source domain to generate better recommendations in a target domain. Despite 3 https://movielens.org/ 4 https://www.netflix.com 40 Chapter 2. Background and Related Work there is not a unified definition of cross-domain recommender systems, Cantador et al. (2015) identified three recommendation tasks of them: • Multi-domain recommendation. The task is to recommend items in both source and target domains by exploiting knowledge from both domains. In this case, a significant user overlap may be necessary (CARMAGNOLA; CENA, 2009). This recommendation task is becoming feasible since users maintain profiles in several interconnected social networks or websites (CARMAGNOLA; CENA; GENA, 2011). • Linked-domain recommendation. In this task, items are recommended only in the target domain by exploiting knowledge from the source and target domains. This recommendation task has been mainly explored to improve the recommendations in a target domain where there is a lack of user preferences either caused by cold-start or sparsity problems (LOW; AGARWAL; SMOLA, 2011). A minimal user overlap may be necessary to perform this task (VERAS et al., 2015), and some approaches aim to establish knowledge-based links between source and target domains (MORENO et al., 2012). • Cross-domain recommendation. This task aims to recommend items only in the target domain by exploiting knowledge only from the source domain. In this case, the task is to provide recommendations in a target domain where there is no information about the users in that domain. Therefore, there is not user overlap among domains, and approaches intend to establish knowledge-based links between domains (TIROSHI; KUFLIK, 2012) or to transfer knowledge from the source domain to the target domain (STEWART et al., 2009). Figure 5 illustrates the three cross-domain recommendation tasks identified by Cantador et al. (2015). In the figure, IS and IT are sets of items from source (DS) and target (DT) domains, respectively. US and UT are sets of users from source and target domain, respectively. Grey filled areas represent the target users and recommended items, and hatched areas represent the exploited data for generating recommendations. Figure 5 – Cross-domain recommendation tasks (CANTADOR et al., 2015). 2.2. Cross-Domain Recommender Systems 41 As mentioned before in Section 1.3, the problem of this thesis is to explore the user- rating-context tensors (also called “multidimensional matrices” or, informally, “cubes”) from source and target domains to improve recommendations in the target domain. According to this issue and the definitions of a cross-domain RS task mentioned in this section, we can consider that problem as: “how to improve the quality of the Linked-domain recommendation task?”. However, despite the task that we aim to perform is classified as a Linked-domain recommendation, we refer the recommender system proposed in this thesis as a Cross- domain recommender system. This referral is made as a matter of simplicity and based on the literature of cross-domain RS (CANTADOR et al., 2015), in which the majority of the papers that perform Linked-domain and Multi-domain recommendation tasks refers themselves as Cross-domain RS. 2.2.3 Cross-Domain Recommendation Goals Likewise the cross-domain recommendation tasks, the cross-domain recommenda- tion goals can vary. According to (CANTADOR et al., 2015), we present some of the most common goals addressed by cross-domain RSs: • Alleviating the cold-start problem. This issue may occur when a RS is unable to generate recommendations due to an initial lack of user preferences. A possible solution is to obtain the user preferences from other domain (source) in order to enrich the user preferences in the target domain (SHAPIRA; ROKACH; FREILIKHMAN, 2013). • Alleviating the new user problem. This issue may happen when a user begins using a RS that initially has no knowledge about his/her preferences. In this case, the RS cannot make recommendations. This issue may be alleviated by exploiting the user’s preferences from a different domain (source) (CREMONESI; TRIPODI; TURRIN, 2011)(WINOTO; TANG, 2008)(SANTOS et al., 2012). • Improving accuracy. Recommender systems may have to deal with a low average number of ratings per user or item, which may negatively affect the quality of the recommendations. Ratings obtained from other domain (source) could increase the rating density in the target domain, which may improve the recommendation quality (STEWART et al., 2009)(MORENO et al., 2012). • Increasing diversity. Recommender systems may provide similar or redundant items for users. Thus, their satisfaction may be compromised. In this case, the diversity of recommendations could be increased by considering item preferences from multiple domains (WINOTO; TANG, 2008). 42 Chapter 2. Background and Related Work Again, as mentioned in Section 1.3, the problem in this thesis is to improve the quality of cross-domain collaborative filtering recommender systems (CD-CFRS). This quality refers to accuracy improvement by addition of context-aware techniques while maintaining the advantages of CD-CFRS in relation to cold-start and sparsity issues. 2.2.4 Cross-Domain Recommendation Scenarios CD-CFRSs are based on the set of ratings provided by users about items of source and/or target domains. According to the overlap among users and/or items of both domains, Cremonesi, Tripodi e Turrin (2011) identified four different cross-domain scenarios: • No overlap. There is no overlap between users and items in the domains. In other words, each item belongs to only one domain, and each user only has preferences for items of one domain. In this case, traditional single-domain CF-based RS cannot make recommendations due to the lack of common data between the domains (ABEL et al., 2011)(SZOMSZOR et al., 2008). • User overlap. Some users have preferences for items of, at least, two domains (source and target), but each item belongs only to a single domain. For instance, this scenario may happen when a dataset has ratings of the same user for two domains (e.g. movies and books) (SAHEBI; BRUSILOVSKY, 2013)(CREMONESI; TRIPODI; TURRIN, 2011). • Item overlap. In this scenario, there are items belonging to distinct domains (source and target). Users can give different ratings for these items depending on their domains. For example, this scenario may occur when a user rates a TV program in multiple systems (e.g. Movielens and Netflix), which could be considered as domains by the RS (CREMONESI; TRIPODI; TURRIN, 2011). In this case, the domain can be classified according to the System level (BERKOVSKY; KUFLIK; RICCI, 2007), as described in Section 2.2.1). • User and item overlap. In this scenario, there is overlap between the users as well as between the items (BERKOVSKY; KUFLIK; RICCI, 2007)(TIROSHI; KUFLIK, 2012). Figure 6 illustrates possible scenarios of user or item overlap between the source and target domains. In the figure, IS and IT are sets of items from source (DS) and target (DT) domains, respectively. US and UT are sets of users from source and target domain, respectively. Grey filled areas represent the target users and recommended items, and hatched areas represent the user and/or item overlap. 2.2. Cross-Domain Recommender Systems 43 Figure 6 – Possible scenarios of user and/or item overlap between the source and target domains (CREMONESI; TRIPODI; TURRIN, 2011). As stated in the problem of this thesis (Section 1.3), it is necessary a User overlap among source and target domains whereas an Item overlap is not. In addition to the literature, we can say that, in our problem, it is also necessary a Contextual overlap among such domains, i.e., the same contexts observed in the source domains are observed in the target domain. 2.2.5 Cross-Domain Approaches As discussed earlier, the cross-domain recommendation has been addressed from various perspectives. This fact led to the development of a variety of recommendation approaches. In many cases, these approaches are difficult to compare, since each one may be based on a different algorithm and adopt different data models about user preferences. Cantador et al. (2015) analyzed some surveys about cross-domain RS (CRE- MONESI; TRIPODI; TURRIN, 2011)(FERNÁNDEZ-TOBÍAS et al., 2012), and uni- fied different categorizations from these surveys by proposing a two-level taxonomy of cross-domain RS approaches, focusing on the exploitation of knowledge in cross-domain recommendations. This taxonomy is presented below: • Aggregating knowledge. Knowledge from one or more source domains is aggregated to perform recommendations in a target domain. Three approaches are considered: 44 Chapter 2. Background and Related Work 1. Merging user preferences – user preferences of different forms and scales are aggregated in a single set. These preferences may be ratings, tags, “like/dislike” binary preferences, among others (BERKOVSKY; KUFLIK; RICCI, 2007)(SA- HEBI; BRUSILOVSKY, 2013). 2. Mediating user modeling data – user modeling data from several recommender systems are aggregated in a single model. For instance, user similarities and user neighborhoods (SHAPIRA; ROKACH; FREILIKHMAN, 2013)(STEWART et al., 2009). 3. Combining recommendations – recommendations or predictions of single-domain RS are aggregated in a single RS (ZHUANG et al., 2010)(GIVON; LAVRENKO, 2009). • Linking and transferring knowledge. In this approach, the knowledge is linked or transferred between domains (source and target). Three possible approaches are: 1. Linking domains – source and target domains are linked by a common knowl- edge, e.g., item attributes, association rules, semantic networks, among others (SHI; LARSON; HANJALIC, 2011)(CHUNG; SUNDARAM; SRINIVASAN, 2007)(AZAK, 2010). 2. Sharing latent features – the source and target domains are related using implicit latent features (ENRICH; BRAUNHOFER; RICCI, 2013)(PAN et al., 2010). 3. Transferring rating patterns – explicit or implicit rating patterns from source domains are exploited in the target domain (LI; YANG; XUE, 2009b)(GAO et al., 2013). Figure 7 illustrates the taxonomy about cross-domain approaches proposed in (CANTADOR et al., 2015). Figure 7 – Cross-domain recommendation approaches taxonomy (CANTADOR et al., 2015). 2.2. Cross-Domain Recommender Systems 45 2.2.6 Cross-Domain Evaluation Basically, two types of evaluation can be used to compare recommender systems in general (FREYNE; BERKOVSKY, 2013). Offline experiments evaluate a RS by analyzing past user preferences. They are typically the easiest evaluation type to perform since they do not require interaction with real users. Online experiments demands that a group of real users use the RS in a controlled environment and give feedback about their experience with it. Cantador et al. (2015) compared the corresponding evaluation methods based on the cross-domain recommendation goals (see Section 2.2.3) and verified that the most of the works about cross-domain RS adopt the offline experiments. In this way, we only describe the aspects of offline experiments, according to (CANTADOR et al., 2015), in the following subsections. 2.2.6.1 Evaluation Data Partitioning In the offline evaluation of traditional RSs, different data partitions can be adopted (e.g. Hold-out, Leave-some-users-out and Leave-all-users-out)(CANTADOR et al., 2015). Thus, the dataset is usually divided into three subsets of ratings: • Training profiles: contain the set of ratings from users for items that are used to train the algorithms under evaluation; • Test profiles: contain the set of users and their known ratings for items that are used as input by the trained algorithms under evaluation; and • Test ratings: contain the set of users and their hidden ratings for items that are used by the algorithms under evaluation for that they can estimate their actual (or known) ratings. Regarding the offline evaluation of cross-domain RS, the same partitions used for evaluation traditional RSs can be adopted. However, for evaluation cross-domain RS is necessary to consider the use of data from the source and target domains. Depending on that use and the cross-domain RS goal, a certain partition may be more suitable, as described in (CANTADOR et al., 2015) and mentioned below: • Hold-out (see Figure 8-left) can be used when the test profiles set is a subset of the training profiles set and contains ratings from source and target domains. The test profiles set is sampled and hidden from the original dataset, taking into account both domains, without partitioning the users. This kind of partitioning may be suitable to evaluate linked- and multi-domain RS with the accuracy goal (SAHEBI; BRUSILOVSKY, 2013)(PAN; YANG, 2013). 46 Chapter 2. Background and Related Work • Leave-some-users-out (see Figure 8-middle) can be adopted when there is not an intersection between the training profiles set and the test profiles set. Note that both sets contain ratings from source and target domains as well as the test profiles set. This partition type may be suitable to evaluate a cross-domain RS with the new user goal (ABEL et al., 2013)(LI; YANG; XUE, 2009b). • Leave-all-users-out (see Figure 8-right) can be adopted when there is not an intersec- tion between the training profiles set and the test profiles set, but in this case there is also not intersection between the training profiles set and the entire target domain data set. Besides, the test profiles and test ratings sets have only data from the target domain. Thus, this partition may be suitable to evaluate a cross-domain RS with the cold-start and new item goals (JAIN; KUMARAGURU; JOSHI, 2013)(GOGA et al., 2013). Figure 8 – Partitioning of data: (left) hold-out; (middle) leave-some-users-out; and (right) leaveall (CANTADOR et al., 2015). 2.2.6.2 Evaluation Metrics As described in Section 2.1.3, there are several metrics for evaluating recommender systems in general. All these metrics can be used in the cross-domain context depending on the cross-domain recommendation goals and tasks (CANTADOR et al., 2015). For instance, Probabilistic measures are preferred when the goal is to reduce the sparsity of the target domain; Ranking measures are adopted for testing user models, especially in cold- start situations; and Qualitative measures are best-suited for the top-N recommendation task. Finally, the majority of works about cross-domain recommendations adopts predic- tion metrics (CANTADOR et al., 2015). This is motivated by the fact that the addressed goal is to reduce sparsity and increase accuracy, and the algorithms designed for this are often based on error-metric optimization techniques, which are naturally evaluated using the category of predictive metrics. 2.3. Context-Aware Recommender Systems 47 2.2.6.3 Sensitivity Analysis The performance of a cross-domain recommender is mainly affected by three pa- rameters (CANTADOR et al., 2015): the overlap between the source and target domains (SHI; LARSON; HANJALIC, 2011)(CREMONESI; TRIPODI; TURRIN, 2011)(ZHAO et al., 2013)(ABEL et al., 2013), the density of the target domain data (CREMONESI; TRIPODI; TURRIN, 2011)(SHAPIRA; ROKACH; FREILIKHMAN, 2013)(CAO; LIU; YANG, 2010)(PAN et al., 2010), and the size of the target user’s profile (SAHEBI; BRUSILOVSKY, 2013)(BERKOVSKY; KUFLIK; RICCI, 2008)(SHI; LARSON; HAN- JALIC, 2011)(LI; YANG; XUE, 2009b). Thus, it is important to consider the sensitivity of the cross-domain algorithms regarding these three parameters. According to (CANTADOR et al., 2015), the majority of the works have assumed a full overlap of users between the source and target domains whereas only a few ones have been evaluated by varying the percentage level of user overlap, e.g., in the range 0%-50% (CREMONESI; TRIPODI; TURRIN, 2011) or in the range 0%-100% (ZHAO et al., 2013). 2.3 Context-Aware Recommender Systems As mentioned before, the context-aware approach uses different contextual infor- mation (e.g., location, time, mood, etc.) to improve the accuracy of recommendations (SETTEN; POKRAEV; KOOLWAAIJ, 2004)(ADOMAVICIUS; TUZHILIN, 2015). For some applications it may not be sufficient to consider only users and items.For example, using the temporal context, a travel recommender system could provide a recommendation in the winter that can be very different from the one in the summer. In another example, a user could prefer to watch news programs in the morning and to watch soccer games at night. Therefore, accurate prediction of user preferences might depend on the use of relevant contextual information by recommender systems (ADOMAVICIUS et al., 2005). In recent years, researchers and companies have developed context-aware recom- mender systems (CARS) and applied them in a variety of different domains such as movie (SHEPSTONE; TAN; JENSEN, 2014), restaurant (PESSEMIER; DOOMS; MARTENS, 2014), tourism (MAHMOOD; RICCI; VENTURINI, 2009), music (KAMINSKAS; RICCI, 2012)(BALTRUNAS et al., 2011), mobile information (CHURCH et al., 2007), news (LEE; PARK, 2007), among others. Likewise the cross-domain approach, CARS is a challenging and emergent field of recommender systems (ADOMAVICIUS; TUZHILIN, 2015). In this way, we describe some of its perspectives in the following subsections. 48 Chapter 2. Background and Related Work 2.3.1 Definition of Context The definition of “context” varies among different research areas, including Com- puter Science. Since context has been studied in multiple disciplines, there is not a standard definition of “context”. In Computer Science, one of the most known definitions is given by Dey, Abowd e Salber (2001). They refer to “context” as: (...) any information that can be used to characterize the situation of enti- ties (i.e., whether a person, place or object) that are considered relevant to the interaction between a user and application, including the user and the application themselves. (BAZIRE; BRÉZILLON, 2005) identified 150 different definitions of context from different fields and made the following observation: ... it is difficult to find a relevant definition satisfying in any discipline. Is context a frame for a given object? Is it the set of elements that have any influence on the object? Is it possible to define context a priori or just state the effects a posteriori? Is it something static or dynamic? Some approaches emerge now in Artificial Intelligence [...]. In Psychology, we generally study a person doing a task in a given situation. Which context is relevant for our study? The context of the person? The context of the task? The context of the interaction? The context of the situation? When does a context begin and where does it stop? What are the real relationships between context and cognition? In the recommender systems area, there is also not a standard definition of “context”. However, some authors (PALMISANO; TUZHILIN; GORGOGLIONE, 2008)(ADOMAVI- CIUS; TUZHILIN, 2015) have a similar point of view about “context” for recommender systems, which is the focus of this thesis. These authors consider “context” as dimensions (e.g. location, time, mood, etc.) and their attributes (e.g. country, city, year, day, sadness, happiness, etc.), which can be used to adapt the recommendations. Based on this definition, we model contextual information in our CD-CARS (Section 3.2). In the next section, we describe how the contextual information can be modelled. 2.3.2 Modelling Contextual Information Contextual models represent which contextual information is considered in a domain or application, and how this information affects the system’s behavior (VIEIRA; TEDESCO; SALGADO, 2009). In general, contextual models define the elements of a particular domain that are considered as context (e.g. location context, temporal context, 2.3. Context-Aware Recommender Systems 49 etc.). They structure entities of a domain and indicate features of these entities, which are managed by the system. However, only this definition of elements does not provide the notion of the context’s dynamic. Production rules are usually adopted for this purpose (VIEIRA; TEDESCO; SALGADO, 2009). Generic contextual models aim to describe the information that must be considered as the context in a generic way. These models provide a classification for an initial set of elements that compose the context in a certain domain (VIEIRA; TEDESCO; SALGADO, 2009). Different applications can reuse the modeled information by extending the model in order to deal with particularities of a particular application. Generic contextual models have been proposed in several areas such as pervasive systems (CHAARI et al., 2007), collaborative systems (VIEIRA; TEDESCO; SALGADO, 2005), data integration (SOUZA et al., 2008), and intelligent systems (GU; PUNG; ZHANG, 2005). In this direction, researchers have investigated the adoption of several techniques for representation of information and knowledge about context (STRANG; LINNHOFF- POPIEN, 2004)(BETTINI et al., 2010). Vieira, Tedesco e Salgado (2009) summarize some of these techniques, as adapted and described in Table 1. Each representation technique described in Table 1 has advantages and disadvan- tages. Thus, there is not a technique that is universally considered as suitable for a certain context-aware system, since different systems have different restrictions and capabilities (VIEIRA; TEDESCO; SALGADO, 2009). A hybrid approach, which combines two or more techniques, may also be adopted as a contextual model. For example, Henricksen e Indulska (2006) proposed a hybrid model that combines ontologies and a graph model based on Object-Role Modeling (ORM). (VIEIRA et al., 2008) proposed a hybrid model that combines ontologies and contextual graphs (BRÉZILLON, 2007) to represent the structure of the contextual information and context-aware behavior. With respect to CARSs, in general they deal with modelling and predicting user preferences by incorporating contextual information into the recommendation process. These preferences are usually modeled as user ratings for items under specific contexts. In this way, the user ratings can be accompanied by contextual information that may be modelled of different types, each type defining a certain aspect of context such as time, location, companion, mood, and so on (ADOMAVICIUS; TUZHILIN, 2015). For instance, by considering movie recommender system, its users and movies can be described according to the following attributes (ADOMAVICIUS; TUZHILIN, 2015): • Movie: id, title, length, release year, director, genre, among others. • User: id, name, address, age, gender, profession, and so on. 50 Chapter 2. Background and Related Work Table 1 – Summary of techniques for representation of context (VIEIRA; TEDESCO; SALGADO, 2009). Technique Advantages Disadvantages Brief Description “Key-Value” pair Simple structure, and easy to imple- ment and use. It does not consider hierarchy and is not suitable for applica- tions with complex structures. A linear search with exact matching of terms. Markup language It provides hierarchy moreover, a markup scheme that imple- ments the model it- self. It does not solve incompleteness and ambiguity. Also, it is not suitable for appli- cations with complex structures. A query language based on marking. Topic maps It facilitates the nav- igation between the contextual elements and the human read- ing. It is an immature technique with a low support of tools. Navigation for se- mantic networks. Ontologies It aggregates rules, concepts, and facts in a single model. Standards make the reuse and sharing easier. It allows semantic comprehen- sion between humans and machines It does not allow modelling the behav- ior of the context- aware system. Also, it is a recent technol- ogy with a low num- ber of tools. Inference engine and query languages based on OWL or frames. Graph models It facilitates the con- cepts specification and the definition of the context-aware system behavior. It does not allow to process the con- cepts: mapping for data structures. It can be translated for XML and makes XML processing. In addition, the contextual information may consist of the following three dimen- sions, which can also be defined according to attributes: • Location. The user’s location when his/her is watching a movie. It may be composed by the attributes: id, name, street, city, state, country, among others. • Temporal. The time when a movie is watched. It may be composed by the attributes: date, day of week5, day type6, month, year, etc. 5 “monday”, “tuesday”, “wednesday”, “thursday”, “friday”, “saturday”, “sunday” 6 “weekday” or “weekend” 2.3. Context-Aware Recommender Systems 51 • Companion. With who the user watches the movie. It may be composed by the attributes: companion type7, companion name8, and so on. Given that, a user may rate (or watch) a movie depending on “where” he/she will be, “when” he/she will watch and/or “whom” he/she will be with. Beyond these three contextual dimensions illustrated in the example above, Neto e Freitas (2007) identified another three basic dimensions, referred as “5W+1H”, which represents “who”, “what”, “where”, “when”, “why”, and “how”. In addition, each contextual dimension can have a complex structure and hierarchy of attributes and their corresponding values. Although this complexity may be modelled by different forms, traditional models adopt a hierarchical structure of contextual infor- mation represented as trees (ADOMAVICIUS et al., 2005)(PALMISANO; TUZHILIN; GORGOGLIONE, 2008). For instance, suppose two contextual dimensions: location and temporal. These dimensions could have the following hierarchies associated with them: • Location: Street → City → State → Country; • Time: Date → Day of Week → Month → Year. Besides the traditional trees for representing the hierarchical structure of contextual information, other ways have been adopted such as Online Analytical Processing (OLAP) (ADOMAVICIUS et al., 2005) and ontologies (KAMINSKAS et al., 2014). 2.3.3 Obtaining Contextual Information An important aspect of CARS is how to obtain contextual information. Adomavicius e Tuzhilin (2015) mention three of the most common methods: • Explicitly. The contextual information is obtained directly from users (LEE; KWON, 2014)(COLOMBO-MENDOZA et al., 2015). A CARS could have this information by asking direct questions about the users’ contexts. For example, a user could select one of the possible contexts provided by the CARS together with the item rating. • Implicitly. In this case, users are not aware about the contextual information gathering process by the CARS. This information can be implicitly obtained in several ways (OH et al., 2014)(PHAM; JUNG; VU, 2014). For instance, a CARS could detect the user location from his/her mobile device location. Another manner could be through temporal information that can be implicitly obtained from the ratings’ timestamps. Therefore, the CARS does not need to interact with users to obtain their contexts. 7 “alone”, “friends”, “girlfriend/boyfriend”, “family”, “co-workers”, etc. 8 “Joseph”, “Paul”, “Laura”, etc. In this case, the values could have an associated id. 52 Chapter 2. Background and Related Work • Inferring. In this method, the contextual information is also obtained implicitly, but the use of statistical or data mining methods is required since the context cannot be obtained in a direct way (SHEPSTONE; TAN; JENSEN, 2014)(WANG; LI; XU, 2015). For example, a CARS could infer the companion (context) of a user from his/her review about a TV program through text mining techniques or observing the kind of the TV program watched by comparing it by using statistical data (e.g. an adult watching a TV program for kids probably is accompanied with kids). Semantic interpretation can also be used for inferring contextual information(BOYTSOV et al., 2015). Recently, some works have used the term “situation” for representing a particular contextual information which is inferred by means of semantic interpre- tation(BOUNEFFOUF, 2013)(BOYTSOV et al., 2015). Usually, the “situation” is inferred from sensor data and characterizes situations in which a user interacts with the CARS(BOUNEFFOUF, 2013). For instance, consider a user associated to: a location defined by the coordinates from his phone’s GPS; the time from his phone’s watch; and the meeting with some person from his agenda. From this knowledge, the CARS could infer that the user is ”in a restaurant, with the general manager of a company, at midday, and it is a workday”. In this way, that inferred contextual information can be called “situation” represented by three contextual dimensions (temporal, location and companion). 2.3.4 Contextual Information Relevance Some contextual dimensions can be more relevant in a given application than some other types (ADOMAVICIUS; TUZHILIN, 2015). For example, the weather may be more relevant for recommending places to visit than for recommending movies to watch. There are several approaches to determine the relevance of a given dimension (or type) of contextual information (ADOMAVICIUS; TUZHILIN, 2015). In particular, this relevance can be verified either manually (e.g. by using domain knowledge of a expert for a given application domain) (BRÉZILLON, 2007) or automatically (e.g. by using several existing feature selection methods from machine learning, data mining, statistics, and so on.) (GUYON; ELISSEEFF, 2003)(LIU; MOTODA, 2012)(CHATTERJEE; HADI, 2015). In a same contextual dimension, there may exist contextual attributes more relevant than others, since a contextual dimension can be modelled as a hierarchical tree (e.g. country x city attributes). In addition, some parts of a contextual dimension may not be known or available. In this case, some authors classify the source of the contextual information according to the relevance of its acquisition and selection. The classification proposed by Adomavicius e Tuzhilin (2015) is divided into three categories: • Fully observable. The relevant contextual information to the application is known 2.3. Context-Aware Recommender Systems 53 explicitly as well as its structure and its values at the moment when recommendations are made (DOURISH, 2004). For example, a product recommender system may consider that only the Temporal, Purchasing Purpose, and Companion dimensions matter for it. In addition, the recommender system may know the entire structure (attributes and values) of all these three contextual dimensions. For instance, the “day type” attribute from the Temporal dimension can have three possible values: “weekday”, “weekend”, and “holiday”. • Partially observable. In this category, only some of the information about the contex- tual dimensions is known explicitly (PALMISANO; TUZHILIN; GORGOGLIONE, 2008). For example, the recommender system may consider all the contextual di- mensions, such as Temporal, Purchasing Purpose, and Companion, but not know all their structure (attributes and values). Note that there may exist different levels of “partial observability”. For example, a CARS could only have access to the Temporal dimension for a certain user, whereas for another user it knows all the other contextual dimensions (Purchasing Purpose, and Companion). • Unobservable. In this category, no information about contextual dimensions is explicitly available to the CARS, and it makes recommendations by considering only the inferred context in an implicit way. For example, a CARS could build a latent predictive model to estimate unknown ratings, where unobservable context is modeled using latent variables (KOREN, 2008). Another aspect of the contextual information relevance is whether and how its importance changes over the time. Adomavicius e Tuzhilin (2015) also classified the contextual dimension relevance into two categories: • Static. The relevant contextual dimensions and their structure remains the same (static) over the time (PALMISANO; TUZHILIN; GORGOGLIONE, 2008). For example, a product recommender system could have three contextual dimensions (Temporal, Purchasing Purpose, and Companion) and they could not change along the entire RS lifetime. In this case, for example, the structure (attributes and values) of the Purchasing Purpose dimension does also not change over the time. • Dynamic. In this category, contextual dimensions, attributes or values change in some way over the time (ANAND; MOBASHER, 2006). For example, a CARS (or a CARS designer) could identify that the Companion dimension is no longer relevant for the CARS and could remove it from the system. Besides, a CARS could change the structure of some of the contextual dimensions (e.g. by adding new attributes to the Purchasing Purpose dimension). 54 Chapter 2. Background and Related Work 2.3.5 Context-Aware Approaches According to (ADOMAVICIUS; TUZHILIN, 2015), there are three systematic paradigms (or approaches) found in the CARS literature: • Contextual pre-filtering. In this recommendation paradigm (illustrated in Figure 9), contextual information guides the data selection for that specific context. In other words, information about the current context is used for selecting the relevant set of data (i.e., user ratings) (ADOMAVICIUS et al., 2005)(VERAS et al., 2015). Then, ratings can be predicted using any traditional collaborative-filtering recommender system on the pre-filtered data. • Contextual post-filtering. In this recommendation paradigm (illustrated in Figure 9), contextual information is initially ignored and the ratings are predicted using any traditional collaborative-filtering recommender system on the entire data. Then, the resulting recommendations (or predictions) are adjusted (or filtered) depending on the contextual information of the users (PANNIELLO et al., 2009)(VERAS et al., 2015). • Contextual modeling. Unlike the pre-filtering and post-filtering paradigms, in this paradigm (illustrated in Figure 9) the contextual information is used directly in the recommendation or predictive process (neither before or after it). Although the pre- filtering and post-filtering paradigms can use traditional CF-based algorithms, the Modelling paradigm actually needs to make “multidimensional” recommendations by considering contextual information as another dimension, beyond the users and items. Several approaches can be used in this algorithm such as predictive models (e.g. decision tree, regression, probabilistic model, among others) (ANSARI; ESSEGAIER; KOHLI, 2000) (OKU et al., 2006), matrix (or tensor) factorization (KARATZOGLOU et al., 2010)(HIDASI; TIKK, 2012)(BALTRUNAS; LUDWIG; RICCI, 2011)(KIM; YOON, 2014), heuristic calculations(ADOMAVICIUS et al., 2005), among others. On the other hand, other ad-hoc approaches, which not necessarily need user- ratings, have also been found in CARS literature and could be used according to the paradigms described above. Véras et al. (2015) describe some of these ad-hoc approaches: • Contextual rules: in this category, there are all kinds of rules that allow recommender systems to sense and to react based on their context. In general, these rules follow the same approach of “event-condition-action” (ECA) rules (MOON et al., 2006), “Key-value” rules (SONG; MOUSTAFA; AFIFI, 2012), among others. 2.3. Context-Aware Recommender Systems 55 Figure 9 – Paradigms for incorporating context in recommender systems (ADOMAVICIUS; TUZHILIN, 2015). • Contextual Ontology: contextual ontologies are not algorithms, but they are crucial to other knowledge-based context-awareness techniques. Most studies that used contextual ontologies also adopted some semantic-based inference. Thus, the com- bination of semantic-based and context-awareness techniques is present in many studies (KAMINSKAS et al., 2014)(MOE; AUNG, 2014b). • Similarity-based: instead of using similarity metrics to compare user or items, in this approach, algorithms compare contexts in order to recommend items (ALHAMID et al., 2015)(VILDJIOUNAITE et al., 2009)(WANG; LI; XU, 2015). The context can be represented in several ways such as tags, key-value pairs, among others. • Supervised learning: in this approach, a set of labeled examples is produced, where each example is composed by features extracted from contextual attributes (e.g. time of the day, mood, etc.). The task of supervised learning is, given a training set, to learn a function that predicts the user preferences based on the contextual features. Examples of algorithms adopted in this approach include Support Vector Machines (VILDJIOUNAITE et al., 2009), case-based reasoning (VILDJIOUNAITE et al., 2009), reinforcement learning (MOON et al., 2009), among others. However, these ad-hoc approaches are difficult to reproduce in distinct domains, since they are usually designed for specific ones. Thus, the algorithms proposed in this thesis (described in Section 3.3.1) follow the systematic paradigms described in (ADOMAVICIUS; TUZHILIN, 2015). 56 Chapter 2. Background and Related Work 2.3.6 CARS Evaluation As described in Section 2.1.3, there are several metrics for evaluating recommender systems in general. All these metrics can be used for evaluating a CARS depending on its purpose (ADOMAVICIUS; TUZHILIN, 2015). However, the evaluation is one of the main research issues and directions for CARS (ADOMAVICIUS; TUZHILIN, 2015). Only a few works have deeply studied the performance evaluation of several CARS approaches and techniques, besides their benefits and limitations. One of these works is presented in (PANNIELLO; TUZHILIN; GORGOGLIONE, 2014), which performed a categorical evaluation and comparison of several contextual techniques under a variety of situations. For instance, the authors compared different recommendation tasks (e.g. to recommend all relevant items, to recommend only the top-n relevant items, etc.), different evaluation metrics (e.g. accuracy, diversity, etc.), and the granularity of the processed contextual information, as well as other evaluation perspectives. Another example is the work presented in (CAMPOS; DÍEZ; CANTADOR, 2014), which focused on exploring “time” as one of the most relevant and widely used contextual dimensions in many CARS. For example, the authors reviewed common evaluation practices and methodological issues related to the comparative evaluation of time-aware recommender systems. They also demonstrated that the choice of the assessment conditions impacts the classification (or ranking) performance of different recommendation strategies. For that, they proposed a methodological framework for a robust and fair evaluation process. The works mentioned above represent an important step in direction to a more reproducible and standardized evaluation methodology of CARS. Analogously to (PAN- NIELLO; TUZHILIN; GORGOGLIONE, 2014), we performed different evaluation tasks (prediction and classification), besides verifying the recommender system performance by using contextual information from distinct dimensions, as we describe in Section 5.1. 2.4 Related Works In this section, we present related works to this thesis, divided into two subsections. In Section 2.4.1, we present some cross-domain recommender systems based on collaborative filtering without considering context-aware techniques, whereas in Section 2.4.2, we describe related cross-domain recommender systems that use contextual information, highlighting their limitation in comparison to our proposed CD-CARS. 2.4.1 Cross-Domain Recommendation based on Collaborative Filtering As mentioned before, the cross-domain recommendation has been addressed from various perspectives. This fact led the development of a wide range of recommendation 2.4. Related Works 57 approaches, as categorized by Cantador et al. (2015) (see Section 2.2.5). In general, Merging user preferences from different domains is the most direct way to address the cross-domain recommendation problem and it is among the most widely used strategies for the cross-domain recommendation (CANTADOR et al., 2015). The massive use of this approach can be explained by the fact that it has been shown that enriching sparse user preference data in a certain domain by adding user preference data from other domains, can significantly improve the generated recommendations under cold- start and sparsity conditions (SHAPIRA; ROKACH; FREILIKHMAN, 2013)(SAHEBI; BRUSILOVSKY, 2013). Figure 10 illustrates the Merging user preferences approach, in which user rating matrices from source (DS) and target (DT) domains are merged, and traditional single-domain CF-based recommender systems can be used on the merged data to recommend items from target domain (IT). Figure 10 – Merging user preferences approach (CANTADOR et al., 2015). Given that and the fact that our proposed CD-CARS is based on the Merging user preferences approach, we describe some works related to it according to the cross-domain RS perspectives described in Section 2.2. Table 2 describes the classification of related papers for these perspectives. Berkovsky, Kuflik e Ricci (2007) focused on cross-domain mediation of user models in CF-based recommendations. In the cross-domain mediation, the user modeling data is imported from remote systems (source domains) exploiting the same CF recommendation technique as the target system (domain). Hence, both source and target domains represent the user models as a list of ratings provided by a user on both domains (user overlap). The CD-CFRS imports the complete set of nearest neighbors calculated by the remote system and uses these similarities in the target domain (cross-domain task). The prediction accuracy of that CD-CFRS was measured through the MAE metric (probabilistic) by using the EachMovie dataset (MCJONES, 1997). In this dataset, movies from different 58 Chapter 2. Background and Related Work Table 2 – Cross-Domain CF-based RS using the Merging user preferences approach. Paper Domain Level Task Goal Scenario Evaluation (BERKOVSKY; KUFLIK; RICCI, 2007) Item At- tribute Cross- Domain Accuracy User and Item Overlap Probabilistic (WINOTO; TANG, 2008) Item Linked- Domain Diversity and Accuracy User Overlap Probabilistic (NAKATSUJI et al., 2010) Item Cross- Domain Accuracy User Overlap Probabilistic (CREMONESI; TRIPODI; TUR- RIN, 2011) System Linked- Domain and Cross- Domain Cold- start and Accuracy User Overlap Ranking (SANTOS et al., 2012) Item Multi- Domain and Linked- Domain Cold- start and New user User Overlap Ranking (TIROSHI et al., 2013) Item Cross- Domain Accuracy User Overlap Ranking (SAHEBI; BRUSILOVSKY, 2013) Item Linked- Domain New User and Accu- racy User Overlap Probabilistic (SHAPIRA; ROKACH; FREI- LIKHMAN, 2013) Item Linked- Domain Cold- start and Accuracy User Overlap Probabilistic and Rank- ing (LONI et al., 2014) Item Cross- Domain Accuracy User Overlap Probabilistic Proposed CD- CARS Item Linked- Domain Cold- start, New user and Accuracy User Overlap Probabilistic and Rank- ing genres are considered as from distinct domains (item attribute domain level), and one item can belong to two or more genres (domains), i.e., there is an item overlap. Winoto e Tang (2008) investigated alternative benefits that cross-domain recom- mendations may have such as serendipity and diversity. For that, they applied a traditional single-domain CF-based algorithm for making cross-domain recommendations by consid- ering aggregated ratings from source and target domains. The unique change made for them in the algorithm was in the weight of the Pearson correlation, which was modified in order to find neighbors with higher number of co-rated items in the target domain. The CD-CFRS performance was measured through the MAE metric (probabilistic) by using 2.4. Related Works 59 a real collected dataset. This dataset was composed by several domains such as movies, books, games, among others (item domain level). On it, a user could rate items from all domains (user overlap). Instead of aggregating user preferences directly, several researches have focused on directed weighted graphs that link user preferences from multiple domains (NAKATSUJI et al., 2010)(CREMONESI; TRIPODI; TURRIN, 2011)(TIROSHI et al., 2013). Nakatsuji et al. (2010) created a domain-specific-user graph (DSUG) for each domain (source and target). In the DSUG, the nodes are users and sets weighted edges between user nodes according to the similarity of users computed in each domain. Also, the DSUGs from distinct domains are connected to create a cross-domain-user graph (CDUG). Thus, the cross-domain RS performs a Random Walk with Restarts (RWR)(LOVÁSZ et al., 1996) on the CDUG from the active user node, and extracts user nodes that are present in DSUGs from the target domain that do not include the node of the active user in the source domain. The authors evaluated the cross-domain RS by using a dataset from two different domains (movies and music) with user overlap, and verified that the accuracy (measured with the MAE metric) of their method is higher than one method that predicts user preference by merging the user ratings from all domains. Cremonesi, Tripodi e Turrin (2011) built a graph whose nodes are associated with items and whose edges reflect rating-based item similarities. In this case, the inter-domain connections are the edges between pairs of items in different domains. The authors also proposed to enhance inter-domain edges by discovering new edges and strengthening existing ones, through strategies based on the transitive closure. Through datasets from different systems (system domain level) and with the same item type (movies), they evaluated several CF-based (nearest neighborhood and latent factor techniques) algorithms using the built multi-domain graph. The authors estimated the accuracy of the algorithms in terms of F-metric (ranking) (CREMONESI; KOREN; TURRIN, 2010) by varying the user overlap level (sensitivity analysis) in the datasets. Santos et al. (2012) proposed an architecture for a recommender system in an inter- application environment and compared traditional (single-domain) and inter-applications recommendations through the Breese metric (Ranking). The proposed recommender system handles different profiles from various applications (with different user-rating scales and forms) by normalizing them. Besides, its recommendation module is based on traditional collaborative filtering techniques, which are adjusted for making cross-domain recommendations (Multi-domain and Linked-domain). The authors developed a web application in order to obtain real user preferences for generating three datasets with different item domains (movies, books, bands and singers) and user-rating scales/forms. In the experiments, 60 users evaluated at least 10 items in the analyzed item domains (user overlap) so that each user have, at least, 30 preferences in the system. In these 60 Chapter 2. Background and Related Work experiments, the authors evaluated several scenarios (combining distinct domains as target) in order to verify the quality of the proposed RS face to cold-start and new user issues. Tiroshi et al. (2013) merged data from source and target domains into a single bipartite user-item graph. From it, several statistical and graph-based features of users and items were extracted. These features were exploited by a machine learning algorithm that addressed the recommendation problem as a binary classification problem. Then, they applied a Random Forest classifier (LIAW; WIENER, 2002) in order to recommend items from the target domain based on the user preferences in the source domain, present in the unified bipartite user-item graph. The authors collected a dataset containing user preferences in multiple domains (book, movie and music) extracted from social network profiles (Facebook9, Last.fm10, LinkedIn11, etc.) with user overlap. They adopted Precision (ranking) as evaluation metric in order to verify the accuracy of their cross-domain RS. Sahebi e Brusilovsky (2013) examined the impact of the size of user profiles in the source and target domains on the quality of cross-domain recommendations, and showed that aggregating ratings from a dense source domain increases the accuracy of recommen- dations in the target domain under cold-start conditions. Basically, the authors applied the k-Nearest Neighbor (k-NN) algorithm (LAROSE, 2005) in the aggregated ratings in order to perform recommendation in the target domain (linked-domain recommendation). For evaluating the CD-CFRS, they adopted a dataset (SAHEBI; COHEN, 2011) with two item domains (book and movie) and user overlap. The accuracy of the system was measured with the RMSE metric (Probabilistic). Similar to (SAHEBI; BRUSILOVSKY, 2013), the work proposed in (SHAPIRA; ROKACH; FREILIKHMAN, 2013) showed significant accuracy improvement by using aggregation-based methods when the user preferences from the target domain are sparse. In this case, the authors used a dataset composed of unary Facebook “likes” as user preferences from several domains (movies, TV shows, and music). Two algorithms were adopted by them: the k-NN algorithm with Jaccard similarity (AMATRIAIN et al., 2011), since the user ratings are in unary form, and the “Facebook popularity” one, which simply lists the top-mentioned user preference items in the Facebook profiles. The CD-CFRS accuracy was mainly measured by Recall (ranking) and MAE metrics. Finally, Loni et al. (2014) proposed a CD-CFRS with factorization machines (RENDLE, 2012) capable of transferring knowledge from different auxiliary domains to a target domains to improve rating predictions in the target domain. The CD-CFRS encodes rating matrices from multiple domains as real-valued feature vectors. With these vectors, the factorization machine finds patterns between features from the source 9 http://www.facebook.com 10 http://www.last.fm 11 http://www.linkedin.com 2.4. Related Works 61 and target domains, and estimates preferences associated with the input vectors. The CD-CFRS accuracy is evaluated through MAE and RMSE metrics. Besides, the Amazon dataset (LESKOVEC; ADAMIC; HUBERMAN, 2007) is used for evaluation purposes. This dataset is composed by user ratings from three different domains (book, music, and television) with user overlap. It is important to notice that only four related papers are based on the “Linked- Domain” task (see Table 2) like our proposed CD-CARS. While (WINOTO; TANG, 2008), (SANTOS et al., 2012), (SHAPIRA; ROKACH; FREILIKHMAN, 2013) and (SAHEBI; BRUSILOVSKY, 2013) explored only traditional single-domain CF-based algorithms used for making cross-domain recommendations, (CREMONESI; TRIPODI; TURRIN, 2011) proposed a cross-domain CF-based algorithm and compared it with those traditional ones. The cross-domain CF-based algorithms presented in Table 2 differ from our proposed CD-CARS by the fact that they do not take into account any contextual information for making recommendations. Therefore, in the next section, we present some related works about cross-domain algorithms that use contextual information in order to improve the quality of their recommendations. 2.4.2 Cross-Domain Recommendation based on Context-Awareness This thesis focuses on investigating the use of contextual information to enhance cross-domain collaborative filtering recommendations. According to Fernández-Tobías et al. (2012), no previous work had addressed the cross-domain recommendation task by deploying of contextual features until then. Seminal works have been published more recently (BRAUNHOFER; KAMINSKAS; RICCI, 2013)(ZHANG; YUAN; YU, 2014)(MOE; AUNG et al., 2013)(MOE; AUNG, 2014b)(MOE; AUNG, 2014a)(KAMINSKAS et al., 2014) (TANG; WAN; ZHANG, 2014)(TEKIN; SCHAAR, 2015)(JI; SHEN, 2015), adopting various approaches to this issue, from semantic techniques to supervised learning, for instance. Taking into account the cross-domain RS and CARS aspects described in Section 2.2 and Section 2.3, respectively, we describe and categorize some related works that make use of context-awareness techniques for providing cross-domain recommendations. Table 3 presents a classification of these works regarding cross-domain RS aspects whereas Table 4 shows the categorization of them with respect to the CARS perspectives. TripFromTV+ (BLANCO-FERNÁNDEZ et al., 2011) selects personalized tourism resources (target domain) for Digital TV viewers, by inferring their particular preferences from the kind of TV programs (source domain) that they enjoyed and from their activity on social networking sites (user overlap). The user profile, contextual information, and item (resources) are modelled through ontologies, so, the recommendation is made with semantic reasoning methods. Specifically, relevant context information is associated with 62 Chapter 2. Background and Related Work Table 3 – Classification of context-aware-based related works regarding cross-domain RS aspects. Paper Domain Level Task Goal Scenario Evaluation (BLANCO- FERNÁNDEZ et al., 2010)(BLANCO- FERNÁNDEZ et al., 2011)(BLANCO- FERNÁNDEZ et al., 2011) Item Cross- Domain New Item and Diversity User Overlap User Satis- faction (BRAUNHOFER; KAMINSKAS; RICCI, 2013) Item Cross- Domain Diversity No Overlap User Satis- faction (YUAN et al., 2012)(ZHANG; YUAN; YU, 2014) Item Multi- Domain Diversity User Overlap Qualitative and Rank- ing (MOE; AUNG et al., 2013)(MOE; AUNG, 2014b)(MOE; AUNG, 2014a) Item Cross- Domain Accuracy No Overlap Ranking (KAMINSKAS et al., 2014) Item Cross- Domain Diversity No Overlap Ranking (TANG; WAN; ZHANG, 2014) Item At- tribute Cross- Domain Diversity No Overlap Ranking (TEKIN; SCHAAR, 2015) Item Multi- Domain Diversity No Overlap Qualitative (JI; SHEN, 2015) Item Linked- Domain Accuracy User Overlap Probabilistic Proposed CD-CARS Item Linked- Domain Cold- start, New user and Accuracy User Overlap Probabilistic and Rank- ing each tourism resource (e.g. opening times, dates, duration, location and ticket price) and matched by the recommendation strategy against the user’s partially observed and static context (e.g location, temporal, etc.). The authors performed a simple experiment with 95 users in order to measure their satisfactions in the use of the TripFromTV+. That work is difficult to compare with other cross-domain RSs since its evaluation is empirical and the TripFromTV+ adopts a knowledge-based method, which in general is domain-specific and requires a great knowledge about the domains and their interconnections. Braunhofer, Kaminskas e Ricci (2013) addressed the cross-domain recommendation task by developing a mobile application that selects music content (target domain) that fits a place of interest (source domain) visited by the user. For that, the application used 2.4. Related Works 63 Table 4 – Classification of context-aware-based related works with respect to CARS as- pects. Paper RepresentationObtention Relevance Approach (BLANCO- FERNÁNDEZ et al., 2010)(BLANCO- FERNÁNDEZ et al., 2011)(BLANCO- FERNÁNDEZ et al., 2011) Ontologies Implicitly Partially Observable (Static) Contextual Ontology (BRAUNHOFER; KAMINSKAS; RICCI, 2013) Key-Value Explicitly and Implic- itly Partially Ob- servable (Dy- namic) Similarity- based (YUAN et al., 2012)(ZHANG; YUAN; YU, 2014) Key-Value Explicitly Partially Ob- servable (Dy- namic) Modelling (MOE; AUNG et al., 2013)(MOE; AUNG, 2014b)(MOE; AUNG, 2014a) Ontologies Explicitly Fully Ob- servable (Static) Contextual Ontology (KAMINSKAS et al., 2014) Ontologies Explicitly Fully Ob- servable (Static) Contextual Ontology (TANG; WAN; ZHANG, 2014) Key-Value Inferred Fully Ob- servable (Dynamic) Supervised Learning (TEKIN; SCHAAR, 2015) Key-Value Implicitly Partially Observable (Static) Modelling (JI; SHEN, 2015) Key-Value Implicitly Partially Observable (Static) Modelling Proposed CD-CARS Key-Value Implicitly and In- ferred Partially Observable (Static) Pre-Filtering, Post-Filtering and Mod- elling the users’ location and emotional tags (contextual information) assigned to both music tracks and point-of-interests (POIs), and adopted similarity metrics (e.g cosine, Jaccard, etc.) for establishing a match between music tracks and POIs based on their emotional tags. These tags were given explicitly by users without any user overlap between the domains. Through a live user study with 10 users, the authors evaluated if the mobile application is capable of providing recommendation with a certain grade of diversity. That work is domain-specific and does not taking into account the users’ preferences in the cross-domain recommendation. In contrast, it recommends items from the target 64 Chapter 2. Background and Related Work domain (music) directly related to the source domain (POI) according to their contextual information. The user’s context is only used for identifying in which POI he/she is located. Therefore, the same recommendations may be made for different users located in that POI. Yuan et al. (2012) proposed a context-aware feature selection framework for cross media recommendation in a digital library. The recommended items in that digital library (DL) can be from different domains (e.g. book, movie and music) and have different tags defined by the users representing the contextual features such as emotions, location, and so on. Thus, the set of items, users and contexts is represented by a user-item-context tensor. The authors initially applied a tensor factorization method (TUCKER, 1966) in that tensor and then used a k-NN clustering algorithm to recommend the top-n items regardless the items’ domain (multi-domain recommendation task). At least, the authors performed experiments by using the Douban12 cross media dataset with user overlap in order to evaluate the quality of cross media recommendations by means of recall and diversity metrics. As we described, the goal of that work is to recommend items in several domains aiming to improve the diversity of the system. In this case, only items from a specific domain could be recommend without taking into account the current user’s context. Besides, the contextual features considered in that work are based on user tags, which can be too varied among different users and also are applied to items. In this way, the users’ contexts are not considered by that work. In (MOE; AUNG, 2014b), a cross-domain RS was developed to recommend cosmet- ics (target domain) related to skin care problems (source domain). The developed system represented the contextual information through ontologies. This contextual information was related to cosmetics such as Place Zone, Age Level, Cosmetics Brand, Season, and Price Range. The system was developed by using Taxonomic conversational case-based reasoning (Taxonomic CCBR) on ontological properties to manage personalization sys- tematically (GUPTA, 2001), Ford-Fulkerson algorithm (PARAMESWARAN; VENETIS; GARCIA-MOLINA, 2011) applied to build the bridge of the semantic concepts between source and target domains and a technique for gathering recommendations according to the users’ contexts (called TOPSIS) (JADIDI; FIROUZI; BAGLIERY, 2010). The accuracy of the developed system was measured by means of raking metrics such as Precision, Recall and F-measure in a simple dataset without user overlap and with information about cosmetics and skin care problems. Likewise (BLANCO-FERNÁNDEZ et al., 2011), the work presented in (MOE; AUNG, 2014b) relies on the extensive use of knowledge about two domains, in which their interconnections must be established a priori by the RS designer. Thus, their domain-specific approach may be difficult to be adjusted for other domains (e.g. book, movie, music, etc.). 12 http://www.douban.com 2.4. Related Works 65 By extending the work proposed in (BRAUNHOFER; KAMINSKAS; RICCI, 2013), Kaminskas et al. (2014) proposed a knowledge-based framework for semantic networks that link concepts from different domains. The framework propagates the node weights, in order to identify target concepts that are most related to the source concepts. Based on data from DBpedia13 without user overlap, the authors evaluated the framework for recommending music (target domain) related to places of interest (source domain) according to location and time as contextual information explicitly defined by the users. Similar to evaluated in (BRAUNHOFER; KAMINSKAS; RICCI, 2013), the authors evaluated the knowledge- based framework by means of a empirical experimentation with some users. Therefore, the same criticism that we mentioned for (BRAUNHOFER; KAMINSKAS; RICCI, 2013) can be applied to that work. Tang, Wan e Zhang (2014) defined a task of cross-language context-aware citation recommendation14, aiming to recommend English citations (target domain) for a given part of the text (context) where a citation is made (e.g. introduction, motivation, related work, etc.) in a Chinese paper (source domain). This task is very challenging because the contexts and citations are written in different languages and there is a language gap when matching them. To handle this problem, they adopted a method that uses machine translation (MT) to translate contexts and/or citations, and then the problem is reduced to the monolingual context-aware citation recommendation. With this reduced problem, they proposed a bilingual context-citation embedding algorithm (called BLSRec-I), which can learn a low-dimensional joint embedding space for both contexts and citations. They evaluated the proposed methods based on a real dataset that contains Chinese contexts and English citations. In this case, there is not the concept of user or item overlap, given that the paper is considered as a “user”. In this way, they adopted three ranking measures (Recall, MAP and Mean Reciprocal Rank) in order to evaluate the positions of the right citations in the ranking list for each given context. Therefore, that work is designed especially for the citation domain and intended for matching citations in the correct context, without taking into account users. Tekin e Schaar (2015) proposed a multimedia content aggregation framework, which gathers content generated by multiple sources in order to provide content on demand for its users. They proposed a content aggregation algorithm, called DIStributed COntent Matching (DISCOM), capable of learning which content to gather and performing a matching between it and users’ preferences, by exploiting similarities between user types. In that system, each user is represented together with its context, which is considered as the user’s type. Based on this user type (context), the content aggregation framework requests content from one of the multimedia sources (multi-domain recommendation). Thus, the context can be represented as user information such as age, gender, among 13 http://wiki.dbpedia.org 14 Despite being a cross-language task, we can classify it as a cross-domain one due to its approach. 66 Chapter 2. Background and Related Work others. In addition, it may also be represented by the device type that the user is using (e.g., computer, mobile phone, etc.). The authors adopted two datasets without user or item overlap for evaluation purposes: the Yahoo! Today Module (YTM) (LI et al., 2010) and a collected one with music items. Based on these datasets, they evaluated the diversity of the recommendations generated by the content aggregation framework. A limitation of that work is the fact that the contextualized recommendations are provided for user/device types, thus, they are not personalized to a single user. Ji e Shen (2015) proposed an improved group-aware CF-based algorithm15 which predicts a user rating using a weighted sum of similar ratings from multiple user subgroups. The algorithm is based on matrix factorization and CodeBook Transfer (CBT) (LI; YANG; XUE, 2009a). The user subgroups are defined according to contextual information available from their ratings. This contextual information can be divided into three categories: users’ contexts (age, gender, etc.), items’ contexts (genre, release date, etc.), and environments’ contexts from the user ratings (time, place, etc.). Experiments were done based on three datasets with distinct domains (book, movie and music) with user overlap. The accuracy of the proposed algorithm was evaluated through probabilistic measures (MAE and RMSE). As we can note, the same limitation that we mentioned for (TEKIN; SCHAAR, 2015) can be applied to that work. In summary, the majority of the cross-domain RS described and categorized above relies on ad-hoc approaches of CARS (Contextual Ontology, Similarity-based and Supervised Learning), which may be difficult to customize to new situations, once that they are usually designed for a specific domain and do not take into account context obtained from user ratings (ADOMAVICIUS; TUZHILIN, 2015). As it can be seen from Table 4, our proposed CD-CARS, in turn, relies on the use of systematic context-aware techniques (Pre-Filtering, Post-Filtering and Modelling). These techniques have been successfully adopted for single domain RS and, in general, require little domain knowledge, since they are based on context obtained from user ratings (ADOMAVICIUS; TUZHILIN, 2015). It is important to mention that some related works adopted the Modelling systematic context-aware technique (YUAN et al., 2012)(TEKIN; SCHAAR, 2015)(JI; SHEN, 2015), as it can be seen from Table 4. However, two of them (YUAN et al., 2012)(TEKIN; SCHAAR, 2015) perform different cross-domain tasks and have distinct cross-domain goals in comparison to our CD-CARS. In addition, the work proposed in (TEKIN; SCHAAR, 2015) is designed for making cross-domain recommendations with no overlap among users, in opposite to the proposed in our CD-CARS. The related work proposed in (JI; SHEN, 2015) is the most similar to the our proposed CD-CARS according to the classification in 15 The authors consider their work as a context-aware RS by determining that a group can be viewed as a user type (context). 2.5. Final Remarks 67 Table 3 and Table 4, but (JI; SHEN, 2015) differs from it once that they only propose a single systematic context-aware technique (Modelling) and it is also not based on the Merging user preferences approach from cross-domain CF-based algorithms, which allows that traditional CF-based algorithms can also be used. Finally, the work proposed in (JI; SHEN, 2015) is intended for making recommendations for group of users instead of recommending items to a singular user. Table 5 summarizes the main limitations of the related works mentioned above in comparison to our proposed CD-CARS. Table 5 – Main limitations of context-aware-based related works in comparison to our proposed CD-CARS. Paper Systematic approach Accuracy Goal Linked- Domain Task User Overlap Merging user pref- erences (BLANCO- FERNÁNDEZ et al., 2010)(BLANCO- FERNÁNDEZ et al., 2011)(BLANCO- FERNÁNDEZ et al., 2011) No No No Yes No (BRAUNHOFER; KAMINSKAS; RICCI, 2013) No No No No No (YUAN et al., 2012)(ZHANG; YUAN; YU, 2014) Yes No No Yes Yes (MOE; AUNG et al., 2013)(MOE; AUNG, 2014b)(MOE; AUNG, 2014a) No Yes No No No (KAMINSKAS et al., 2014) No No No No No (TANG; WAN; ZHANG, 2014) No No No No No (TEKIN; SCHAAR, 2015) Yes No No No Yes (JI; SHEN, 2015) Yes Yes Yes Yes No Proposed CD-CARS Yes Yes Yes Yes Yes 2.5 Final Remarks In this chapter, we presented the main concepts related to this thesis as well as its related works. The research about these concepts provides a background for the understanding of the proposed CD-CARS, described in the next chapter. 68 3 CD-CARS Proposal In this chapter, we describe the CD-CARS proposal. For that, we formalize the cross-domain context-aware recommendation problem (Section 3.1) and model the contextual information (Section 3.2). In Section 3.3, we describe the proposed CD-CARS algorithms whereas in Section 3.3.2 we present the cross-domain algorithms that can be adopted as base, in combination with the proposed CD-CARS. At last, in Section 3.4, we mention the final remarks of this chapter. 3.1 CD-CARS Problem Formalization As mentioned in Chapter 1, the majority of the proposed approaches to cross- domain recommendation deals with collaborative filtering (CF) (CREMONESI; TRIPODI; TURRIN, 2011)(FERNÁNDEZ-TOBÍAS et al., 2012). CF is more convenient for cross- domain recommendation due to the lack of homogeneous description of item content in different domains. It can rely only on user ratings of items, usually represented by user-rating matrices (User x Item). Also, most of the available cross-domain RS suggest items regardless of the contextual conditions, which can be important to predict the users’ preferences in a particular context. Despite the large number of existing cross-domain CF- based recommender systems (CD-CFRS), the synergy between them and context-aware techniques is still little explored (FERNÁNDEZ-TOBÍAS et al., 2012)(CANTADOR; CREMONESI, 2014). In this way, we address the cross-domain recommendation problem under the CF and context-awareness perspectives. For that, as defined by Adomavicius e Tuzhilin (2015), we consider the user ratings as a function of three dimensions: CR : User × Item×Context −→ Contextual Ratings Thus, user ratings can be stored in a multidimensional user-rating-context tensor for each item domain (e.g. books, movies, music, among others). Notice that the notion of domain adopted in this thesis is based on the “Item level” definition (described in Section 2.2.1) by considering movies and books belonging to different domains, for example. In order to formalize our cross-domain context-aware recommendation problem, we introduce the following definitions, considering a set of ‘n’ source domains (S1,S2, ...,Sn) and just one target domain (T). 3.2. Modelling Contextual Information 69 Definition 1. US1,US2, ...,USn,UT: sets of users for each domain; IS1,IS2, ...,ISn,IT: sets of items for each domain; CS1,CS2,CSn,CT: sets of contextual features for each domain; CRSi : USi ×ISi ×CSi (where i=1,2,...,n) and CRT : UT ×IT ×CT: contextual user-rating tensors (i.e., multidimensional matrices or cubes) for each domain; US,T = (US1 ∪US2 ∪ ...∪USn ) ∩UT 6= �: at least one user must have preferences for items in the target domain and, at least, a source domain (user overlap); IS,T = IS1 ∩ IS2 ∩ ...∩ ISn ∩ IT = �: there is no item overlap between domains; CS,T = CS1 ∩CS2 ∩ ...∩CSn ∩CT = CS1 ∪CS2 ∪ ...∪CSn ∪CT 6= �: the same set of possible contexts is observed for user ratings in all domains (contextual overlap). Hence, the problem to be solved in this thesis is to estimate unknown ratings for items in a target domain (IT) by exploiting the user-rating tensors from the source and target domains (CRSi where i = 1, 2, ...,n and CRT), assuming US,T , IS,T and CS,T . It is important to mention that the ratings from the contextual user-rating tensors can have different scales or forms in distinct domains. For example, ratings of music could be represented as a binary form such as “Like” or “Dislike” while the ratings of movies and books could be represented, respectively, by five-star or ten-star scales. Therefore, the recommendation algorithms have to deal with this issue. For instance, an algorithm could normalize the different scales from ratings among distinct domains (SANTOS et al., 2012). 3.2 Modelling Contextual Information In this section, we describe how the contextual features are formalized (Section 3.2.1) as well as the contextual information is obtained and selected considering its relevance (Section 3.2.2). 3.2.1 Contextual Features Formalization As mentioned in Section 2.3.2, contextual information can be of different “types”, each one defining a certain contextual dimension, such as time (e.g. “day of week”, “period of the day”, etc.), location (e.g. “at home”, “at work”, etc.), companion (e.g. “alone”, “with friends”, etc.), among others. Furthermore, each contextual dimension can have a hierarchical structure that can be represented as different attributes (e.g. Time: Date → 70 Chapter 3. CD-CARS Proposal DayOfWeek → TimeOfWeek, or Date → Month → Quarter → Year) (ADOMAVICIUS; TUZHILIN, 2015). According to the CD-CARS problem formalization described before, we modelled a set of contextual features (illustrated in Figure 11), for each domain (CS1,CS2, ...,CSn,CT), as a Cartesian product of k contextual dimensions: Cd = D1 × D2 × ... × Dk (where d=S1,S2, ...,Sn,T domains) (ADOMAVICIUS et al., 2005). Each dimension Dj (j = 1,2,...,k) can be represented by l contextual attributes (A1,A2, ...,Al). Each attribute Az (z = 1,2,...,l) has a set of m values (v1,v2,...,vm) representing a part of the contextual information. Moreover, “Unknown”, which represents a missing (or not observable) part of the contextual information, is a default value (v1) for any contextual attribute. Figure 11 – A contextual feature represented by dimensions, attributes and values. Thus, the contextual information can be represented as a tuple of w values from different contextual attributes and/or dimensions, i.e., a possible context (c’) of a set of contextual features can be denoted as c′ = (v1,v2, ...,vw), where each value vs (s=1,2,...,w) belongs to a different contextual attribute Az (z = 1,2,...,l) and/or dimension Dj (j = 1,2,...,k). Note that the order of these values in the tuple does not change the meaning of the represented context (c’). 3.2. Modelling Contextual Information 71 For instance, consider three contextual dimensions (k = 3): D1 = Temporal,D2 = Location,D3 = Companion. Each one can have different hierarchical representation through contextual attributes. Suppose that D1 has two (l = 2) attributes (A1 = Day,A2 = DayType), D2 has three (l = 3) attributes (A1 = City,A2 = State,A3 = Country) and D3 has one (l = 1) attribute (A1 = CompanionType). For each contextual attribute of those dimensions, there is a set of possible values such as: • Temporal dimension (D1): A1 = {v1 = Unknown,v2 = Sunday,v3 = Monday,v4 = Tuesday,v5 = Wednesday,v6 = Thursday,v7 = Friday,v8 = Saturday} with eight possible values (m = 8), A2 = {v1 = Unknown,v2 = Weekday,v3 = Weekend} with three possible values (m = 3); • Location dimension (D2): A1 = {v1 = Unknown,v2 = Aberdeen,...,v2839 = Zurich} with 2839 possible values (m = 2839), A2 = {v1 = Unknown,v2 = Alabama,...,v381 = Wisconsin} with 381 possible values (m = 381), A3 = {v1 = Unknown,v2 = Australia, ...v113 = Zambia} with 113 possible values (m = 113); • Companion dimension (D3): A1 = {v1 = Unknown,v2 = Alone,v3 = Accompanied, v4 = Family,v5 = Friends,v6 = Partner,v7 = Fellows} with seven possible values (m = 7), Given this example, a set of contextual features could be the combination of all possible values from different attributes (six) and dimensions (three). So, by multiplying the ‘m’ values of each attribute, in this case, it will be approximately twenty billion different contexts resulted from the Cartesian product. Notice that this contextual feature modelling does not guarantee the consistency of the information among different attributes. For example, a context c′ ={Sunday,Weekday,Recife,Alagoas,EUA,Unknown} would be valid according to our modelling, but inconsistent considering the real contextual information (e.g., a consistent context could be c′ ={Sunday, Weekend, Recife, Pernambuco, Brazil, Unknown}). In this way, we let the RS application that uses this modelling responsible for obtaining consistent contextual information. In fact, despite the huge amount of possible contexts in the example given above, a real dataset obtained by the RS application, in that example, would have approximately one hundred and sixty thousand possible contexts, which is the result of the multiplication of the ‘m’ values of the more discrete contextual attributes from different dimensions: Day (A1 from D1) with m = 8, City (A1 from D2) with m = 2839, and CompanionType (A1 from D3) with m = 7, instead of the twenty billion ones. It is important to mention that the proposed contextual feature modelling is based on the “Key-Value” model (referred in Section 2.3.2). In this case, the matching between 72 Chapter 3. CD-CARS Proposal the context of the recommendation, which is called “contextual criteria”, and the contextual information, represented by this model in the user-ratings, is made in a linear way. In other words, once the tuple of w contextual values is established, from different contextual attributes and/or dimensions, then a contextual criteria can be used as a query term (i.e., context of the recommendation), as described in Algorithm 1. A contextual criteria can also be represented as a tuple of w contextual values from the same contextual attributes and/or dimensions as the contextual information from ratings. Despite the “Unknown” value be always a possible one for each contextual attribute, it has distinct meanings in the contextual criteria and the contextual information from ratings. For the contextual criteria, “Unknown” (v1) can be viewed as a part of the context to be ignored (i.e., uninformed). In this case, this value means that any value of the contextual information from ratings is acceptable for that contextual attribute and dimension, including the “Unknown” one, which for the contextual information from ratings represents a missing (or not observable) part of the contextual information, as mentioned before. Therefore, the algorithm described above considers, for contextual matching purposes, only the values different from “Unknown” on the contextual criteria. This mechanism is sufficient for the proposed CD-CARS algorithms. Algorithm 1. Matching between the context of the recommendation and the contextual information from ratings. Input: C, I, n (where C is the contextual criteria array of contextual values, and I is the contextual information array of contextual values, considering both arrays with the same size n). Output: isMatched (a boolean value determining if there is a match between the context of the recommendation and the contextual information). 1: procedure contextualMatching(C, I, n) 2: for v=1 to n do 3: if C[v] 6= “Unknown′′ and C[v] 6= I[v] then return false 4: end if 5: end for 6: return true 7: end procedure end 3.2.2 Obtaining and Selecting Relevant Contextual Information In this thesis, we are not concerned about how to obtain contextual information. We let the RS application responsible for gathering this information and persisting it 3.2. Modelling Contextual Information 73 according to the proposed contextual model. As mentioned in Section 2.3.3, three methods are more often used to acquire contextual information: explicit, implicit, and inferred. So, depending on the data available in the datasets used by the RS application, some of these methods can be more useful than others. Although the source of contextual information is irrelevant for the proposed CD-CARS algorithms, the quality of the obtained contextual information remains relevant for them. On the other hand, after obtaining the contextual information, if there are many contextual attributes available for a contextual dimension in the contextual model, then selecting relevant contextual information is important for the quality and performance of the CD-CARS. Taking the example previously given, if the temporal dimension (D1) has two attributes: Day (A1 = {v1 = Unknown,v2 = Sunday,v3 = Monday,v4 = Tuesday,v5 = Wednesday,v6 = Thursday,v7 = Friday,v8 = Saturday}), and DayType A2 = {v1 = Unknown,v2 = Weekday,v3 = Weekend}, then CD-CARS could select just one of these attributes for recommendation. The same may occur among two distinct contextual dimensions, for example, also considering the location dimension (D2), CD- CARS could select just one of these dimensions for the recommendation. Some types of contextual information (e.g. temporal, location, companion, etc.) can be more relevant in a given domain (e.g. books, movies, music, etc.) than others. As mentioned in Section 2.3.4, the selection of a relevant contextual dimension or attribute can be made with a feature selection method from data mining (LIU; MOTODA, 1998). In the proposed CD-CARS, for each different target domain where the recommendation takes place, we can apply the information gain measure1 considering the user-rating as a class and each contextual attribute as a tested attribute of the information gain measure. For instance, the user-rating class could have five possible nominal values representing ratings of a five-star scale. Note that in that example we assume that all ratings from the source and target domains are in the same scale and form. However, as mentioned before, for different forms or scales of ratings among distinct domains an algorithm must normalize them(SANTOS et al., 2012). Besides, the information gain calculated for each contextual attribute may vary depending on the target domain in which the data mining method is applied. So, from the list of most relevant attributes generated by the information gain measure, the CD-CARS could select only the contextual attribute with higher information gain value. Then, it could execute performance experiments in the selected attribute and, progressively, select the next relevant attribute of a different contextual dimension if the performance difference is significative between the previously selected attribute and the next one. In the case of selecting contextual attributes in the same dimension, however, 1 InfoGain(Class,Attribute) = H(Class) - H(Class | Attribute), where H means “entropy”, defined in Information Theory. 74 Chapter 3. CD-CARS Proposal the CD-CARS could select only the top attribute with higher information gain value, since the subsequent attributes are nothing more than different representations of the top attribute selected. For instance, if the top contextual attribute was Day, followed by the DayType, then selecting only the Day attribute is sufficient, as it represents all values of DayType attribute in a more discrete way. 3.3 CD-CARS Algorithms The algorithms proposed in our work rely on the use of a base cross-domain recommender system, in which the predicted rating (R̂(u,i)) for a particular pair of user u and item i, belonging to the target domain item set (IT), can be formalized as: R̂(u,i) = CD(u,i,RS1,RS2, ...,RSn,RT ), i ∈ IT (3.1) In which, RSi (i=1,2,...,n domains) and RT are 2-dimensional user-rating matrices, respectively in the source and target domains. Notice that the base cross-domain RS does not take into account the contextual information. In addition, it is possible that different scales and forms of ratings from distinct domains have to be handled by the base cross-domain RS algorithm. As mentioned before, an algorithm could normalize the ratings among distinct domains (SANTOS et al., 2012). In this way, we assume that all ratings from the source and target domains are in the same scale and form. In our CD-CARS problem, we consider contextual user-rating tensors and we need a function (F) to make rating predictions of items (i) for users (u) in contexts (c) given the tensors from source (CRSi, where i=1,2,...,n domains) and target domains (CRT), as defined in Equation 3.2. ĈR(u,i,c) = F(u,i,c,CRS1,CRS2, ...,CRSn,CRT ), i ∈ IT (3.2) The function (F) can be implemented using any of the three proposed CD-CARS algorithms described in the following section. 3.3.1 Proposed Algorithms We designed the algorithms according to three different context-aware RS paradigms (ADOMAVICIUS; TUZHILIN, 2015): Pre-filtering (PreF), Post-filtering (PostF) and Modelling. These paradigms are usually adopted in single-domain RS, but we extended their directives for the cross-domain recommendation task by taking into account the contextual user-rating tensors from different domains. 3.3. CD-CARS Algorithms 75 3.3.1.1 Cross-Domain PreF Algorithm PreF algorithm initially uses contextual information to filter the contextual user- rating tensor from the target domain (CRT) in order to obtain a two-dimension (2D) user-rating matrix. On the other hand, the contextual user-rating tensors from the source domains (CRS1,CRS2, ...,CRSn) are collapsed into a two-dimension (2D) user- rating matrix, by aggregating ratings for the same user-item pair in different contexts, prioritizing the user-ratings of the context of the recommendation (c). Then, the base cross-domain algorithm is applied to these matrices to produce the predicted ratings (ĈR(u,i,c)). Figure 12 illustrates the PreF technique, which is formalized in three steps as follows. • Step (1) Define the 2D reduced matrix (context-filtered matrix) for the target domain: RcT (u,i) = CRT (u,i,c) (3.3) The context-filtered matrix only has ratings according to: R̂T (u,i) =   ˆCRT (u,i,c), if c = o not available, otherwise (3.4) where ‘o’ represents the rating’s context. • Step (2) Define the 2D aggregated matrices (prioritizing the user-ratings with the context of the recommendation) for the source domains: Rj(u,i) = CRj(u,i,c) (3.5) Where j = S1,S2, ...,Sn source domains. For each source domain ‘j’, the aggregated ratings are calculated as: R̂j(u,i) =   ˆCRj(u,i,c), if c = o∑ c∈Cj ˆCRj (u,i,c) |Cj| , otherwise (3.6) where ‘o’ represents the rating’s context. • Step (3): Apply the base cross-domain technique using the reduced matrices: ĈR(u,i,c) = CD(u,i,RS1,RS2, ...,RSn,R c T ), i ∈ IT (3.7) In the steps above, the matching between the user-rating context (c) and the context of the recommendation (o) is made according to Algorithm 1 (page 72). Besides, we assume that the user-ratings from the source and target domains are in the same scale and form, as mentioned before. 76 Chapter 3. CD-CARS Proposal Figure 12 – The pre-filtering cross-domain recommendation is made by filtering the target contextual user-rating tensor for a given context. Since PreF algorithm filters the contextual user-rating tensor from the target domain, it could have a few user-ratings to make recommendations in very specific contexts, thus, the recommendation process would be made using almost entirely only the user-ratings from the source domains. Due to this drawback, we also proposed other context-awareness techniques as described in the next sections. 3.3.1.2 Cross-Domain PostF Algorithm PostF algorithm initially produces a single unified user-rating matrix by aggregating ratings for the same user-item pair in different contexts, prioritizing the user-ratings of the context of the recommendation (c). The base cross-domain is then applied using as input the aggregated rating matrices. Finally, contextual information is used to filter the ratings produced by the cross-domain algorithm. This filtering is done by considering items contained in the set of preferred item categories (e.g. comedy, action, rock, etc.) by the user in a given context (e.g. considering only comedy movies on weekdays). Figure 13 illustrates this algorithm, which is formalized in the following steps: • Step (1) Define the 2D aggregated matrices (prioritizing the user-ratings with the context of the recommendation) for the source and target domains: Rj(u,i) = CRj(u,i,c) (3.8) 3.3. CD-CARS Algorithms 77 Where j = S1,S2, ...,Sn,T domains. For each domain ‘j’, the aggregated ratings are calculated as: R̂j(u,i) =   ˆCRj(u,i,c), if c = o∑ c∈Cj ˆCRj (u,i,c) |Cj| , otherwise (3.9) where ‘o’ represents the rating’s context. • Step (2): Apply the base cross-domain technique using the matrices from Step (1) and collect the predicted ratings: R̂(u,i) = CD(u,i,RS1,RS2, ...,RSn,RT ), i ∈ IT (3.10) • Step (3): Given a context of recommendation (o) and user u as input, a rating produced for an item (R̂(u,i)) is discarded if the number of “good” rated items of a given item category (gi) is less than a threshold value (θ) in that context. Otherwise, the rating predicted by the cross-domain algorithm is maintained: ĈR(u,i,c) =   R̂(u,i), if c = o and CP(u,c,gi) >= θnot available, otherwise (3.11) where the category preferences tensor (CP(u,c,g)) contains the number of “good” rated items for each item category g, from different domains, observed in a context c for a user u. The definition of a “good” rated item can be made according to the scale and form of the user-ratings from distinct domains. As mentioned before, we considered that all user-ratings are normalized among the different domains. In this way, a “good” rated item could have a rating of, at least, “four” in a five-star scale, for example. We let the responsibility of this definition for the CD-CARS implementation as well as the optimal θ value, which could be calculated considering the number of “good” rated items in general. For instance, if a user has fifty “good” rated items in general (regardless their categories), then the θ value could be set to 10% of that number (i.e., θ = 5), which would mean that only categories with at least five “good” rated items would be considered. An alternative way to define Equation (3.11) is: ĈR(u,i,c) =   R̂(u,i) × (1 + ω), if c = o and CP(u,c,gi) >= θR̂(u,i) × (1 −ω), otherwise (3.12) where ω is a factor to increase or decrease the predicted rating value (R̂(u,i)). The θ value still has the same meaning of Equation (3.11) by defining a threshold for determining which categories of items are relevant or not according to the minimal 78 Chapter 3. CD-CARS Proposal number of “good” rated items. Thus, relevant categories will have the predicted rating value increased whereas irrelevant categories will have the predicted rating value decreased. In this case, the ω could be an empirically defined value (e.g. “0.1” value would increase or decrease by 10% the predicted rating value). Also, it could be calculated, proportionally, according to the relevance of the item category preferred by the user in a given context (e.g. the higher is the number of “good” rated items the higher is the ω value). Therefore, the PostF algorithm could adjust the predicted rating instead of filtering it. Likewise the θ value, the ω value definition is responsibility of the CD-CARS implementation. Figure 13 – The cross-domain post-filtering recommendation is made over the aggregated user-rating matrices and then post-filtered according to contextual user pref- erences. It is important to mention that g could also be expressed as a set of attributes which characterize an item (e.g. user tags), instead of being expressed as item categories like item genres (e.g. comedy, action, rock, etc.), without losing generality. 3.3. CD-CARS Algorithms 79 Similar to PreF algorithm, in the PostF algorithm, we apply a base cross-domain algorithm. However, none of the input tensors (from source and target domains) are filtered by context. Instead, all contextual user-rating tensors are reduced to matrices (by aggregating user-ratings from different contexts) that serve as input for the base cross- domain algorithm. After applying the base algorithm, only the post-filtered ratings are taken into account. In this process, an important task is to build the category preferences tensor (CP(u,c,g)). We build the category preferences tensor from the contextual user-rating tensors of the source and target domains. Depending on the θ value, it is possible that a user only has category preferences in source domains. The same situation could happen in a scenario where a dataset contains just a few users with overlap. In these cases, the PostF algorithm would not be able to recommend items in the target domain. In order to alleviate this problem, some techniques can be used. For example, using association rule mining (HIPP; GÜNTZER; NAKHAEIZADEH, 2000) to discover usage patterns between different domains and contexts (e.g. we could infer that users who like to read romance books on weekdays also like to watch romance movies on weekdays). Thus, we propose the enhancement of the category preferences tensor by using association rules to infer other item categories preferred by the users according to the possible contexts. Figure 14 illustrates this idea, which is only required when a user receives a recommendation in the target domain and the category preferences tensor does not have information about his/her rated item categories in that domain. As it can be seen from Figure 14, the association rules input is generated from the category preferences tensor (CP(u,c,g)). Each entry of this input is extracted from a user (u) and represents a set of pairs, composed of an item category (g) and context (c) in which that user has the number of “good” rated items greater or equal than the theta value. With that input, we can use an algorithm for generating association rules such as, for example, the Apriori (AGRAWAL; IMIELIŃSKI; SWAMI, 1993). After applying that algorithm and obtaining the resulting association rules, we can select only the most relevant ones according to their confidence and support levels. Optimal values for these parameters can vary depending on the dataset used in the CD-CARS application. In addition, we are interested only in “cross-domain” rules, i.e., rules that relate item categories between a source domain and the target domain. We discard rules that relate item categories between two source domains (or for the same domain), since they do not make the PostF recommendation possible when a user only has item category preferences in source domains. Finally, these rules are used for enhancing the category preferences tensor, which will have inferred item categories and contexts beyond the original preferences. 80 Chapter 3. CD-CARS Proposal Figure 14 – Category preferences tensor enhancement from association rules. Note that source and target domains have different sets of categories (e.g. music x books). However, by using association rules in the category preferences tensor enhancement, we can make cross-domain PostF recommendations even to less related domains such as music and movies, making the CD-CARS domain-independent. We remember that we consider this relation among distinct domains according to the set of item genres of them. As more the domains have item genres in common the more related they are considered (e.g. Book and Television have several item genres in common such as “romance”, “educational”, “religion”, etc.). 3.3.1.3 Cross-Domain Modelling Algorithm Unlike the PreF and PostF algorithms, the Modelling algorithm does not need to use a base cross-domain algorithm. In fact, it makes “multidimensional” recommendations by considering contextual information beyond users and items without reducing the contextual user-rating tensors. In this thesis, we propose the extension of two single-domain context-aware Mod- elling approaches: heuristic calculations (ADOMAVICIUS et al., 2005) and matrix factor- 3.3. CD-CARS Algorithms 81 ization (BALTRUNAS; LUDWIG; RICCI, 2011) (as mentioned in Section 2.3.5). This extension goes beyond the inclusion of contextual information in a single-domain CF-based algorithm since we intend to perform cross-domain context-aware recommendations. Thus, we consider four dimensions (user, item, context, and domain) instead of three (user, item, and context) for the cross-domain context-aware recommendation. The heuristic calculation approach described in (ADOMAVICIUS et al., 2005) includes contextual information by using an n-dimensional distance metric instead of the user-user or item-item similarity metrics traditionally used in such techniques (RICCI; ROKACH; SHAPIRA, 2011) whereas the matrix factorization approach described in (BALTRUNAS; LUDWIG; RICCI, 2011) could be generalized to consider additional dimensions (e.g. item domain) for the representation of the data as a tensor of four dimensions (user, item, context, and domain). Figure 15 – The cross-domain modelling recommendation uses contextual information directly in the recommendation function as an explicit predictor of a user rating for an item. The Modelling algorithms proposed in this thesis (illustrated in Figure 15) are formalized in a single step as follows. • Step (1): Apply the extended version of the base cross-domain algorithm using the user-rating-context tensors: ĈR(u,i,c) = CD(u,i,c,CRS1,CRS2, ...,CRSn,CRT ), i ∈ IT (3.13) In this way, for the Modelling algorithm based on heuristic calculations, instead of using a similarity metric only for calculating the user-user and item-item distances, it could also include other dimensions as context and item domain. For example, if the similarity metric is the Euclidian distance, it could be defined as: dist[(u,i,c,d), (u′, i′,c′,d′)] = √ w1d 2 1(u,u′) + w2d22(i, i′) + w3d23(c,c′) + w4d24(d,d′) (3.14) 82 Chapter 3. CD-CARS Proposal where d1, d2, d3, and d4 are distance functions defined for dimensions User, Item, Context, and Domain, respectively, and w1, w2, w3, and w4 are the weights assigned for each of these dimensions. For instance, these weights can be set according to the relevance of the four dimensions or empirical values. As mentioned in Section 2.3.4, it is important to select properly the contextual information used in the recommendation dataset. In addition, depending on the way how contextual information is obtained, it could be more or less relevant. For example, the contextual relevance might be low in a system that a user rates an item without explicitly consider the context in that rating, i.e., the context is dissociated from the user-rating (ADOMAVICIUS; TUZHILIN, 2015). On the other hand, a system that collects the user contextual information together with his/her rating may be more reliable, obtaining a more relevant contextual information (ADOMAVICIUS; TUZHILIN, 2015). Therefore, the use of the Modelling algorithm based on heuristic calculations will depend on the relevance of the association between context and user-rating. To avoid this issue, we also propose to adapt some algorithms based on matrix factorization (or tensor factorization, to be more specific) that considers the contextual information, such as the proposed ones in (KARATZOGLOU et al., 2010), (HIDASI; TIKK, 2012), and (KIM; YOON, 2014). For these algorithms, the item domain (e.g. book, TV, music, etc.) could be considered as a contextual dimension. This adaptation is simple and maintains the original logic of those algorithms. Finally, it is important to notice that the Modelling algorithm does not consider the context of the recommendation, as opposed to the PreF and PostF algorithms. On the other hand, it only takes into account the context of the user-ratings. Thus, the Modelling algorithm can recommend items without knowing the context of the user at the moment of the recommendation. 3.3.1.4 Cross-Domain Hybrid Contextual Algorithms In the previous sections, we described three proposed CD-CARS algorithms. In this section, we discuss how these different algorithms can be combined. One possibility is to combine the PreF and PostF algorithms, as illustrated in Figure 16. Naturally, the PreF algorithm can be used before the PostF one, since the PreF only filters out the recommendation data before the base cross-domain algorithm is applied, whereas the PostF filters out only the outcome of the base cross-domain algorithm. Another possibility is to combine the Modelling and PostF algorithms, as illustrated in Figure 17. Again, the PostF algorithm can be used after the first proposed algorithm, which in this case is the Modelling one. Unlike the utilization of the Modelling algorithm alone, in which the context of the recommendation is not considered, in this combination 3.3. CD-CARS Algorithms 83 is necessary to take into account the context in the moment of the recommendation once that the PostF algorithm demands this information in order to filter irrelevant items recommended by the Modelling algorithm. Figure 16 – The cross-domain PreF algorithm can be used before the PostF algorithm in a possible combination. In the two hybrid algorithms mentioned above (PreF + PostF and Modelling + PostF), the combinations are made in a direct way, without requiring to adapt any proposed algorithm. Other combinations might demand adaptation of the proposed algorithms, so, this could be a future direction of our research. 3.3.2 Base Cross-Domain Algorithms In this thesis, we propose the adoption of single-domain and cross-domain algo- rithms. According to the taxonomy presented in Section 2.2.5, the adopted algorithms fit in the Aggregating knowledge category and, more specifically, in the Merging user preferences approach. Section 3.3.2.1 describes the single-domain CF-based algorithms whereas Section 3.3.2.2 describes the cross-domain CF-based algorithms. 84 Chapter 3. CD-CARS Proposal Figure 17 – The cross-domain modelling algorithm can be used before the PostF algorithm in a possible combination. 3.3.2.1 Single-Domain as Cross-domain Algorithms As stated by Cremonesi, Tripodi e Turrin (2011), if there is overlap among users and/or items, then standard single-domain CF algorithms can be used for generating cross-domain recommendations by merging user-rating matrices from different domains, considering that these matrices are normalized (see Section 2.2.5). Thus, these algorithms also can be used to make cross-domain recommendations considering our formalization of the CD-CARS problem, since we have an overlap of users among domains (see Section 3.1). Therefore, those CF-based algorithms can be used as a base cross-domain algorithm together with our proposed CD-CARS algorithms. In this way, we can apply single-domain collaborative filtering algorithms as cross-domain (CD) technique in equations 3.7 and 3.10, for PreF and PostF algorithms, respectively. Note that the Modelling algorithm does not require a base cross-domain algorithm, as mentioned before. The following sections describe some algorithms for two traditional classes of CF-based algorithms that can be applied as a base cross-domain algorithm. 3.3.2.1.1 Neighborhood-based Algorithms Neighborhood-based algorithms (RICCI; ROKACH; SHAPIRA, 2011) calculate the similarity between two users or items, producing a rating prediction, which is computed by averaging the ratings expressed by similar users or items, weighted with the respec- tive similarity values. We describe some of these algorithms below (RICCI; ROKACH; SHAPIRA, 2011): 3.3. CD-CARS Algorithms 85 • NNUserNgbr computes a neighborhood consisting of the nearest ‘n’ users to a given user. “Nearest” users are defined by a similarity metric. In other words, the recommendations are derived from a neighborhood of the ‘n’ most similar users. The optimal value of ‘n’ can be defined through experiments. • ThresholdUserNgbr computes a neighborhood through a similarity threshold and takes any users that are at least that similar to a given user. The threshold should be between −1 and 1. The higher is the threshold value, the more selective is the neighborhood. Again, the optimal threshold value can be estimated from experimentation. • GenericItemBasedCF is simpler than the user-based CF algorithms described above because there is no parameter to be adjusted (like as ‘n’ or threshold). This item- based CF algorithm compares series of preferences expressed by many users, for one item, rather than by one user for many items (user-based). Some similarity metrics used by user-based CF algorithms also can be used in order to compute the similarity between items. A crucial aspect of these algorithms is the similarity computation between items or users. Similarity in user-based and item-based CF algorithms can be computed by means of traditional similarity metrics, such as (RICCI; ROKACH; SHAPIRA, 2011): • Weighted Euclidean distance similarity computes the Euclidean distance (dist) between two such user or items points. The equation below denotes a generic calculation of this metric: dist[(u,i), (u′, i′)] = √ w1d 2 1(u,u′) + w2d22(i, i′) (3.15) where d1 and d2 are distance functions defined for two dimensions: User and Item, respectively, and w1 and w2 are the weights assigned for each of these dimensions. This similarity metric never returns a negative value, and the more similar two users are (i.e., the larger the similarity value between them is), the smaller is the distance between them. In addition, if we only need to calculate the distance between users, then we can consider i = i′. On the other hand, if we only need to calculate the distance between items, then we can consider u = u′. • Cosine similarity (cos) is a measure of similarity between two vectors of an inner product space that measures the cosine of the angle between them. In the context of item recommendation, this measure can be employed to compute user similarities by 86 Chapter 3. CD-CARS Proposal considering a user u as a vector xu ∈ R|I|, where xui = rui if user u has rated item i, and 0 otherwise. The similarity between two users u and v is then computed as cos(u,v) = ∑ i∈Iuv ruirvi√∑ i∈Iu r 2 ui ∑ j∈Iv r 2 vj (3.16) where Iuv denotes the items rated by both u and v users. The same idea can be used to obtain similarities between two items i and j, according to the equation below: cos(i,j) = ∑ u∈Uij ruiruj√∑ u∈Ui r 2 ui ∑ u∈Uj r 2 uj (3.17) Cosine similarity is particularly used in positive space, where the outcome is bounded in a [0,1] interval. • Pearson correlation (PC) similarity is a ratio of the covariance of two data sets to their standard deviations. Unlike the Cosine similarity, this metric considers the effects of mean and variance of the ratings made by users u and v. The equation below denotes the calculation of this metric: PC(u,v) = ∑ i∈Iuv (rui − r̄u)(rvi − r̄v)√∑ i∈Iuv (rui − r̄u)2 ∑ i∈Iuv (rvi − r̄v)2 (3.18) The same idea can be used to obtain similarities between two items i and j, according to the equation below: PC(i,j) = ∑ u∈Uij (rui − r̄i)(ruj − r̄j)√∑ u∈Uij (rui − r̄i)2 ∑ u∈Uij (ruj − r̄j)2 (3.19) While the sign of the outcome of this metric indicates whether the correlation is direct or inverse, its magnitude (ranging from 0 to 1) represents the strength of the correlation. More sophisticated similarity metrics can also be used such as the proposed ones in (DIDAY; BOCK, 2000), (BEZERRA; CARVALHO, 2004), (BEZERRA; CARVALHO, 2011), among others. The recommendation of these single-domain neighborhood-based algorithms, used for cross-domain purposes, is made as usual (CREMONESI; TRIPODI; TURRIN, 2011), except by the fact that only items from the target domain are recommended. 3.3.2.1.2 Matrix factorization algorithms Matrix factorization algorithms (RICCI; ROKACH; SHAPIRA, 2011) map users and items to a latent feature space, commonly with reduced dimensionality (f). User-item 3.3. CD-CARS Algorithms 87 interactions are modeled as inner products in that space. The latent space tries to explain ratings by characterizing both items and users on factors automatically inferred from user feedback. The major challenge of these algorithms is computing the mapping of each item and user to factor vectors. After the recommender system completes this mapping, it can easily estimate the rating a user will give to any item. Unlike neighborhood-based algorithms, matrix factorization algorithms do not use similarity metrics, but techniques for identifying latent semantic factors like singular value decomposition (SVD) (RICCI; ROKACH; SHAPIRA, 2011). For the SVD, consider that each item i is associated with a vector qi ∈ Rf and each user u is associated with a vector pu ∈ Rf. For a given item i, the elements of qi measure the extent to which the item possesses those factors, positive or negative. For a given user u, the elements of pu measure the extent of interest the user has in items that are high on the corresponding factors (again, these may be positive or negative). The resulting dot product2, qTi pu, captures the interaction between user u and item i, i.e., the overall interest of the user in characteristics of the item. The final rating is created by the rule above: r̂ui = µ + bi + bu + qTi pu (3.20) where µ is the overall average rating and the parameters bu and bi indicate the observed deviations of user u and item i, respectively, from the average µ. In order to learn the model parameters (bu, bi, pu and qi) the regularized squared error can be minimized: min b∗,q∗,p∗ ∑ (u,i)∈κ (rui −µ− bi − bu − qTi pu) 2 + λ4(b2i + b 2 u+ ‖ qi ‖ 2 + ‖ pu ‖2) (3.21) The constant λ4, which controls the extent of regularization, is usually determined by cross-validation. Minimization is typically performed by either stochastic gradient descent or alternating least squares. Alternating least squares techniques rotate between fixing the pu’s to solve for the gi’s and fixing the qi’s to solve for the pu’s. When one of these is taken as a constant, the optimization problem is quadratic and can be optimally solved (BELL; KOREN; VOLINSKY, 2007). An easy stochastic gradient descent optimization was popularized by Koren (2008). The algorithm loops through all ratings in the training data. For each given rating rui, a prediction (r̂ui) is made, and the associated prediction error eui = rui − r̂ui is computed. For a given training case rui, the parameters are modified to move in the opposite direction of the gradient, producing: 2 The dot product between two vectors x, y ∈ Rf is defined as xT y = ∑f k=1 xk · yk. 88 Chapter 3. CD-CARS Proposal • bu ← bu + γ · (eui −λ4 · bu) • bi ← bi + γ · (eui −λ4 · bi) • qi ← qi + γ · (eui ·pu −λ4 · qi) • pu ← pu + γ · (eui · qi −λ4 ·pu) As the single-domain neighborhood-based algorithms, single-domain matrix factor- ization algorithms can be used as the base of cross-domain recommendations, considering that a unique merged user-rating matrix from different domain contains information about users and items. The recommendation of these single-domain neighborhood-based algorithms, used for cross-domain purposes, is made as usual (CREMONESI; TRIPODI; TURRIN, 2011), except by the fact that only items from the target domain are recom- mended. It is important to mention that the single-domain matrix factorization algorithms are different from the tensor factorization described in the Modelling approach. Instead of considering the mapping between users and items, matrix factorization (or tensor factorization, to be more specific) algorithms can be generalized to consider the context for the representation of the data as a tensor (KIM; YOON, 2014). 3.3.2.2 Cross-Domain Algorithm In the previous section, we described the single-domain CF-based algorithms adopted as a base cross-domain algorithm. One of the CD-CARS’s advantages is to allow that the majority of traditional single-domain CF-based algorithms can be used in combination with the proposed CD-CARS algorithms. In this section, we describe an actual cross-domain algorithm adopted in this thesis. Once that this algorithm is originally intended to perform cross-domain recommendations, we can directly apply it as a base cross-domain algorithm in combination with our proposed CD-CARS algorithms. For that, we adopted a neighborhood-based (CF-based) cross-domain algorithm, due to its simplicity. The cross-domain neighborhood-based algorithm adopted is proposed by (CRE- MONESI; TRIPODI; TURRIN, 2011), NNUserNgbr-transClosure. It enhances the NN- UserNgbr algorithm (described in Section 3.3.2.1.1), which is intended for single-domain recommendations, by improving its user-to-user or item-to-item similarities calculations with a “transitive closure” method. This improvement is achieved by discovering indirect relations among elements (i.e., transitive closure discovers all n-steps similarity paths between any pair of users, extending their neighborhood), as illustrated in Figure 18. For instance, if there exist two direct links: user A = user B = 1 (e.g. full similarity by the 3.4. Final Remarks 89 Pearson metric) and user B = user C = 1, then the transitive closure allows to set user A = user C = 1. According to (CREMONESI; TRIPODI; TURRIN, 2011), this transitive closure procedure is described as follows. Given a binary relation S, where sij is equal to either 1 or 0, the algebraic transitive closure of S is the union of successive powers of the original matrix, i.e.: Strans = ⋃ n∈N Sn (3.22) where ⋃ is the union operator. Matrix S is represented by the weighted connections among a set of items. However, since this matrix does not represent a binary relation, Equation (3.22) has been adapted as follows. Figure 18 – Original (a) and enhanced (b) item-to-item connections. Solid circles represent items belonging to a single domain, whereas blank circles represent cross items that act as a bridge among different domains (CREMONESI; TRIPODI; TURRIN, 2011). The “union” operator, which is defined for binary relations, has been replaced by the “maximum” operator, Z = max(X,Y) where the maximum matrix Z between two similarity matrices X and Y has been defined so that zij = max(xij,yij). The maximum operator adds the similarities discovered for new links while maintaining the original values for existing connections (since original similarities are generally stronger than derived ones). (CREMONESI; TRIPODI; TURRIN, 2011) has limited the transitive closure to only two steps. Experiments showed that a transitive closure with more than two steps did not provide any sensible improvement in the recommendation accuracy while increasing computational requirements. Thus, the enhanced item-to-item similarity matrix was computed as: S∗ = max(S,S2) (3.23) Except by this user-to-user (or item-to-item) similarity calculation, the remaining logic of the NNUserNgbr-transClosure algorithm is the same as the NNUserNgbr one. 3.4 Final Remarks In this chapter, we described the CD-CARS proposal. For that, we formalized the cross-domain context-aware recommendation problem and modeled the contextual 90 Chapter 3. CD-CARS Proposal information. In addition, we described the proposed CD-CARS algorithms as well as the base cross-domain algorithms adopted in this thesis. Regarding the CD-CARS problem formalization, we remember that we assumed that all ratings from the source and target domains are in the same scale and form. As mentioned before, an algorithm could normalize the ratings among distinct domains (SANTOS et al., 2012). In the same way, the proposal of the CD-CARS algorithms considers that all ratings from different domains are normalized. As outlined in Section 3.2, the proposed contextual feature modelling is based on the “Key-Value” model (mentioned in Section 2.3.2), since it is simple and relatively easy to implement and use (VIEIRA; TEDESCO; SALGADO, 2009)(BETTINI et al., 2010). Besides, this kind of contextual model allows a quickly matching between the context of the recommendation and the context represented in the user-ratings, so, that model is sufficient for the proposed CD-CARS algorithms. One of the advantages of the proposed CD-CARS algorithms is the possibility of using traditional single-domain and cross-domain CF-based algorithms as a base algorithm, which is used in combination with the proposed ones. The adoption of other algorithms of different approaches (e.g. content-based filtering, semantic-based, and so on.) for serving as a base cross-domain algorithm may be studied in future research. A particular common aspect between the PreF and PostF algorithms relies on the aggregation of user-ratings from different contexts in order to reduce the contextual user-rating tensors into matrices. Despite the proposal of these algorithms is concerned about a theoretical scenario, in which a user can have multiple ratings for the same item in distinct contexts, this usually does not happen in real datasets, as it is the case of the adopted ones in this thesis. Thus, this important aspect of those algorithms could not be experimented. Taking into account the PreF and Modelling algorithms, we can say that they are capable of making domain-independent cross-domain recommendations, i.e., they do not use any information about the item attributes (unlike the PostF approach), so, any “kind of item” (domain) can be used in such approaches independently. Indeed, the pre-filtering approach only reduces the dataset to be used by any traditional CD-CFRS, which is capable of making domain-independent cross-domain recommendations (CREMONESI; TRIPODI; TURRIN, 2011)(FERNÁNDEZ-TOBÍAS et al., 2012). Also, a heuristic-based modelling approach only changes the user/item similarity metric used by any traditional CD-CFRS, so, the domain-independent cross-domain recommendation can be made as well. On the other hand, for making domain-independent cross-domain recommendations with the post-filtering approach, the used dataset must have the description of item 3.4. Final Remarks 91 categories/genres (or any other attribute that categorizes them). Hence, the post-filtering approach filters out (or adjusts), from the cross-domain recommended item list, the items of a category according to the category preferences of the users. These preferences are obtained according to the users’ contexts and can be generated by several approaches, from simple arithmetic calculations to more sophisticated ones as the solution described in this chapter. In the next chapter, we will present an implementation of the CD-CARS proposal and an experimental evaluation performed on real datasets. 92 4 CD-CARS Implementation This chapter describes particular details of an implementation of the CD-CARS proposal. For that, we investigate the problem outlined in Section 3.1 taking into account three contextual dimensions and three distinct domains from real datasets. Section 4.1 describes the properties of CD-CARS datasets with different contextual information and domains and how they were obtained. Also in this section, the process of selecting relevant contextual attributes and values is shown. Section 4.2 presents how the contextual model is implemented through the extension of a traditional framework available in the area of single-domain recommender systems. Section 4.3 describes two of the proposed algorithms (see Section 3.3.1) considering the implemented contextual model. Section 4.4 describes implementation details of two base cross-domain algorithms. Finally, Section 4.5 presents the final remarks of this chapter. 4.1 Dataset Acquisition One of the main difficulties for evaluating cross-domain recommender systems is the lack of publicly available data, representing the ratings of the same users on items classified in multiple domains (TANG et al., 2012). Although there are several datasets from different domains (television, music, books, etc.) separately, just a few of them have user-overlapped ratings, i.e., all users only have ratings in a single domain. However, our problem demands a cross-domain dataset with contextual information. In other words, this dataset must have at least a small number of users that have some ratings in the source and target domains (i.e. a cross-domain dataset with some level of user overlap) and some of these ratings must contain contextual information. In order to achieve that, we extracted two datasets based on the dataset from (LESKOVEC; ADAMIC; HUBERMAN, 2007), since it was not designed for evaluating cross-domain context-aware recommendations. This dataset contains product metadata and review information about different Amazon products (Books, music CDs, DVDs, and so on)1 and we implemented a method to extract only its relevant information for our problem. For instance, we removed duplicated ratings and irrelevant information (e.g. number of votes, Amazon sales rank, product reviews, etc.), besides the creation of methods for gathering contextual information of three contextual dimensions (see Section 4.1.1). One of the extracted datasets was used for evaluating the CD-CARS algorithms in two more related domains (Book and Television, named as “book-television dataset”) and another for evaluating it in two less related domains (Book and Music, named as 1 https://snap.stanford.edu/data/amazon-meta.html 4.1. Dataset Acquisition 93 “book-music dataset”). Both datasets contain a set of ratings, which are composed of: • User ID: a positive integer value; • Item ID: a positive integer value; • Rating value: a positive integer value, defined explicitly by the user on a five-star scale; and • Contextual information: an array of integer values that represent contextual values. Each index from the array represents a distinct contextual attribute (e.g. country) of a certain contextual dimension (e.g. Location), as described in Section 3.2.1. Unfortunately, the contextual information was not available directly in the original dataset, so, we had to obtain the contextual information implicitly from the ratings’ dates, users’ Web accounts (from their Amazon IDs), and by inferring the ratings’ reviews, as detailed in the Section 4.1.1. Finally, we discarded users that had less than twenty ratings and did not have ratings in both domains (book/television for one dataset, and book/music for the other one) from datasets, which means that only overlapped users were included (full overlap). From these datasets, we created reduced versions of them in order to evaluate the sensitivity of the CD-CARS algorithms for datasets with different levels of user overlap. Thus, beyond the full-overlapped “book-television” and “book-music” datasets, we have four variations of each one of them: two datasets with 10% of overlap (one for each domain as a target), and two with 50% of overlap (one for each domain as a target). We generated these reduced versions from the full-overlapped datasets by removing all ratings for items from the target domain of the users chosen randomly according to the overlap percentage. For example, the “book-television” dataset with 10% of user overlap and Television domain as the target has 10% of the users with ratings in both domains (source and target) and the remaining 90% of them have ratings only for items in the Book domain (source). So, this dataset can be used for evaluating the cross-domain recommendation in the Television domain as a target when just a few number of users has ratings in the target domain. Besides the set of ratings, the extracted datasets contain information about items such as item ID, title, domain, and categories. We extracted those categories from the original dataset for all items in all domains (book, television, and music). The categories of the Book and Television domains were mapped into a single set of 27 categories2 while 2 Based on Amazon’s Movie&TV categories: unknown, action&adventure, international, animation, anime, boxed sets, classics, comedy, documentary, drama, educational, health, religion, fantasy, LGBT, holiday&seasonal, horror, artistical, kids&family, war, musicals, mystery, romance, sci-fi, special, sports, westerns 94 Chapter 4. CD-CARS Implementation the Music domain had 19 categories3. Also, the extracted datasets contain information about users such as user ID, Amazon user ID, and address (obtained from the Amazon user ID, as detailed in Sec- tion 4.1.1.2). At last, they contain the user-rating reviews, which are used for obtaining companion contextual information (as detailed in Section 4.1.1.3). 4.1.1 Obtaining Contextual Information As mentioned in Section 2.3.3, three methods are more often used to gather contex- tual information: explicit, implicit, and inferred. According to the contextual information available in the extracted datasets, we considered three contextual dimensions in the CD- CARS implementation. For two of them (Temporal and Location), we implicitly obtained the contextual information from the user ratings, while for one of them (Companion), we inferred the contextual information from the user-rating reviews. 4.1.1.1 Temporal Dimension The contextual information in the Temporal dimension can be directly extracted from user-rating timestamps, which are present in the majority of the datasets containing user-ratings. In this way, the timestamps could be transformed into several contextual attributes with different values and hierarchical levels. For example, timestamps could represent the “period of the day” attribute (Dawn, Morning, Afternoon and Night) as well as “day type” (weekend or weekday), as illustrated in Figure 19. A discussion about the hierarchical levels and possible values of contextual attributes can be found in Section 3.2.2. Figure 19 – Example of a temporal dimension with its possible contextual attributes and values in a hierarchical view. 3 Based on Amazon’s CD music categories: unknown, jazz, rock, classic rock, international, classical, pop, blues, gospel, dance, new age, country, folk, vocal, alternative rock, hard rock, kids&family, rap, special 4.1. Dataset Acquisition 95 In this implementation, the real datasets used in the experiments only had date information of the user-ratings. For that reason, we could not extract contextual attributes related to the rating time (e.g. rating hour) or “period of the day”. So, only contextual attributes related to day or month could be extracted, such as “day type” or “period of the year”. It is important to mention that the contextual information extracted from user- rating date is not entirely reliable because a user can rate an item and consume it in distinct moments, which clearly generates an impact in the temporal context. For instance, a user could watch a movie on Saturday and rate it only on Sunday. This temporal gap could be more frequent for the “period of the day” attribute, since a user could watch a movie in the afternoon and, due to its duration, rate it only at night, when that movie ends. However, despite this risk, we believe that there are rating patterns according to the users’ contexts in the same way that there are consumption patterns from these users’ contexts. In other words, we can say that there are users that usually rate (instead of watching) comedy movies on Sundays, for example. Thus, we expect that the risk of gathering the temporal context in an implicit way is minimal, which can be observed in the evaluation of the proposed recommender system. 4.1.1.2 Location Dimension Usually, the contextual information in the Location dimension can be implicitly collected when the user is using some device with Internet access or with a Global Positioning System (GPS), for example. However, this information is not available in the real datasets adopted in this implementation. On the other hand, all user-ratings in the datasets have information about the user IDs (and the real Amazon user IDs), as mentioned before. From the actual Amazon user IDs, we created a web crawler responsible for extracting the address information in the profile web pages from the users’ accounts, which can be accessed at a Uniform Resource Locator (URL)4 containing the Amazon user IDs. The web crawler simply makes a GET request from the Hypertext Transfer Protocol (HTTP) and receives an HyperText Markup Language (HTML) page. This page is parsed by a regular expression based on a particular HTML tag in order to extract a string containing the user’s address information. However, this string is defined by the user and is not standardized. In this way, after obtaining the not-standardized address information from the user profile web page, we used a Representational State Transfer (REST) web service, called 4 http://www.amazon.com/gp/pdp/profile/AMAZONUSERID/ is an example of URL, where “AMA- ZONUSERID” represents the real Amazon user ID, omitted here for privacy issues. 96 Chapter 4. CD-CARS Implementation Google Maps Geocoding5, that provides an Application Programming Interface (API) to retrieve the standardized address information of the users. This service requires a string address as parameter and returns a JavaScript Object Notation (JSON) document with multi-level address information about that address, such as “country”, “locality” (repre- senting a “city”), “administrative_area_level_1” (representing a “state”, for example), “administrative_area_level_2” (representing a “county” or “district”, for example), and so on (“administrative_area_level_N”, with ‘N’ greater than two, representing more specific administrative areas). Listing 4.1 shows a response example for the requested string “recife”. Listing 4.1 – Example of a response (JSON document) from google maps geocoding api for the input “recife”. { " r e s u l t s " : [ { " a d d r e s s _ c o m p o n e n t s " : [ { " l o n g _ n a m e " : " R e c i f e " , " s h o r t _ n a m e " : " R e c i f e " , " t y p e s " : [ " l o c a l i t y " , " p o l i t i c a l " ] } , { " l o n g _ n a m e " : " R e c i f e " , " s h o r t _ n a m e " : " R e c i f e " , " t y p e s " : [ " a d m i n i s t r a t i v e _ a r e a _ l e v e l _ 2 " , " p o l i t i c a l " ] } , { " l o n g _ n a m e " : " P e r n a m b u c o " , " s h o r t _ n a m e " : " PE " , " t y p e s " : [ " a d m i n i s t r a t i v e _ a r e a _ l e v e l _ 1 " , " p o l i t i c a l " ] } , { " l o n g _ n a m e " : " B r a z i l " , " s h o r t _ n a m e " : " BR " , " t y p e s " : [ " c o u n t r y " , " p o l i t i c a l " ] } ] , ... } } The standardized address (composed by country, state, and city) is extracted from the JSON document (through a JSON parser) and persisted into a database, serving as an 5 A developer key is required to use the web service, as described at https://developers.google.com/maps/documentation/geocoding/intro 4.1. Dataset Acquisition 97 address catalog. This mechanism avoids future unnecessary requests to the service. Thus, for each address searched in the web service, we persisted the search string (raw address) and its corresponding standardized address. This information is also persisted together with the users’ information. Figure 20 illustrates the process for gathering the location contextual information, described above. Figure 20 – Process for gathering the location contextual information from the user infor- mation. Since we only obtained static address information (country, state, and city were obtained from users’ web profiles), we could not extract contextual attributes related to the abstract locations, as, “place” attribute (at home, at work, in a movie theater, etc.). Therefore, only contextual attributes related to geographical location could be extracted, as illustrated in Figure 21. It is important to mention that each user has the same location context for all his/her ratings once that his/her location context was extracted from his/her address defined in his/her static web profile. Despite the users’ geographical locations are extracted, there is no guarantee that they are actually in those locations where they are rating (or consuming) an item, given that their locations are extracted from their web profiles, which are defined in their initial registration in the system. However, we believe that the majority of the items are rated by the users in their registered location. Finally, it is important to say that not all users had the address information available at their web profile, thus, many users did not have any contextual information about their location. 98 Chapter 4. CD-CARS Implementation Figure 21 – Example of a location dimension with its possible contextual attributes and values in a hierarchical view. 4.1.1.3 Companion Dimension In opposite to the previous contextual dimensions, in which we implicitly obtained the contextual information, the contextual information of the Companion dimension was inferred from the user-rating reviews available in the real datasets. For that, we implemented a method based on (BAUMAN; TUZHILIN, 2014). (BAUMAN; TUZHILIN, 2014) proposed an unsupervised text mining algorithm for discovering relevant contextual information from the user-generated reviews. Initially, they observed that contextual information appears more likely in specific reviews (those that describe specific details of an item, such as a book or movie) than in generic reviews (that describes overall comments about an item). After they cluster the user reviews into two groups (specific and generic), they try to find key-words or topics describing the contextual information in those reviews. They state that these topics appear more frequently in the specific reviews than in the generic ones. In this way, they compare the frequencies of key-words (or topics) appearing in the specific and the generic reviews and then select these key-words that have high-frequency ratios, assuming that the selected key-words should contain most of the contextual information among the user reviews. Finally, they inspect the list of the selected key-words by manually identifying the relevant context-related topics. In this way, we applied that method to our dataset with some adaptations by considering different domains. This method is described in the following. In the first step of the method, we separated reviews into specific and generic reviews by using the measures proposed in the original method (BAUMAN; TUZHILIN, 2014): 4.1. Dataset Acquisition 99 • LogSentences: logarithm of the number of sentences in the review plus one6. • LogWords: logarithm of the number of words used in the review plus one. • VBDsum: logarithm of the number of verbs in the past tenses in the review plus one. • Vsum: logarithm of the number of verbs in the review plus one. • VRatio - the ratio of VBDsum and Vsum (V BDsum V sum ). With those measures, we used the classical K-means clustering method (JAIN, 2010) to separate all the reviews into the “specific” and “generic” clusters, as described in (BAUMAN; TUZHILIN, 2014). However, we applied the clustering separately for each domain (Book, Television, and Music). This was the first adaptation in the original method, which is intended for single domain reviews. As a result, the vast majority of the reviews (99.8%) were classified as “specific” for all domains in our dataset. This result might have occurred due to the nature of user-rating reviews available in the original dataset, in which they were analyzed by the dataset provider in order to maintain only relevant user-rating reviews. It is important to mention that the majority of user-ratings (76%) did not have reviews (only did the rating values), considering both datasets. Given that the great majority of the reviews were classified as “specific”, we simplified the word-based and LDA-based (BLEI; NG; JORDAN, 2003) methods proposed in (BAUMAN; TUZHILIN, 2014), since these methods rely on the separation of specific and generic reviews. The adapted word-based method is explained below: 1. For each review Ri, identify the set of nouns Ni appearing in it. 2. For each noun nk, determine its weighted frequencies ws(nk) corresponding to the specific (s) reviews, as follows ws(nk) = |Ri : Ri ∈ specific and nk ∈ Ni| |Ri : Ri ∈ specific| (4.1) 3. Filter out the words nk that have overall low frequency, i.e., w(nk) = |Ri : nk ∈ Ni| |Ri : Ri ∈ specific| < α, (4.2) where α is a threshold value for the application (e.g., α = 0.005) 6 The authors of the proposed method added one to avoid the problem of having empty reviews when logarithm becomes −∞ 100 Chapter 4. CD-CARS Implementation 4. For each remaining noun nk left after filtering in the previous step, find the set of senses synset(nk) using WordNet7 (MILLER, 1995). 5. Combine senses into groups gt having close meanings using WordNet taxonomy distance. Words with several distinct meanings can be represented in several distinct groups. 6. For each group gt determine its weighted frequencies ws(gt) through frequencies of its members as: ws(gt) = |Ri : Ri ∈ specific and gt ∩Ni 6= ∅| |Ri : Ri ∈ specific| (4.3) 7. Sort groups by its weighted frequencies ws(gt) in its descending order. The adapted LDA-based method is described below: 1. Build an LDA model on the set of the specific reviews. 2. Apply this LDA model to all the user-generated reviews in order to obtain the set of topics Ti for each review Ri with a probability higher than a certain threshold level. 3. For each topic tk from the generated LDA model, determine its weighted frequencies ws(tk) corresponding to the specific (s) reviews, as follows ws(tk) = |Ri : Ri ∈ specific and tk ∈ Ti| |Ri : Ri ∈ specific| (4.4) 4. Filter out the topics tk that have overall low frequency, i.e., w(tk) = |Ri : tk ∈ Ti| |Ri : Ri ∈ specific| < α, (4.5) where α is a threshold value for the application (e.g., α = 0.005) 5. Sort topics by its weighted frequencies ws(tk) in its descending order. Note, that we did not use the generic reviews in both adapted methods described above, unlike they are originally (BAUMAN; TUZHILIN, 2014). In addition, after generating the sorted lists of key-words (or topics), we manually selected in the list of topics for each item domain only the topics related to the Companion contextual dimension. In contrast, in the original method, there is no restriction about the contextual dimensions extracted from the key-words (or topics). 7 WordNet is a large lexical database of English. Words are grouped into sets of cognitive synonyms, each expressing a distinct concept. Function synset(word) returns a list of lemmas of this word that represent distinct concepts. 4.1. Dataset Acquisition 101 Figure 22 – Example of a companion dimension with its possible contextual attributes and values in a hierarchical view. In this way, we identified six contextual values (alone, accompanied8, family, friends, partner, and colleagues9) for only one contextual attribute of “companion”, from the word groups and topics selected. This contextual attribute is high-level and likewise other contextual dimensions, it could be expanded into more granular contextual attributes (illustrated in Figure 22), as, specific family members (father, mother, siblings, and so on.), or yet, the person itself (e.g. by pointing the companion’s name) that the user is with. However, we let the contextual attribute of the Companion dimension in high-level due to the source of contextual information (inferred from reviews), in which has a few number of reviews with particular description of companion (specific family members or companion’s name). Furthermore, contextual information of other contextual dimensions could be discovered (e.g. the task or purchase purpose). However, we did not consider other contextual dimensions due to the small percentage of user-rating reviews with their respective inferred contexts, in contrast to the Companion dimension, which had 85% of the total of user-rating reviews with the inferred contexts. This observation is reasonable because a few users usually mention their temporal contexts in the reviews and inferring the companion context of user-ratings seems to be easier than inferring the temporal context, for example. In order to evaluate the classification performance of the implemented method for the companion extraction, we adopted the same methodology described by the authors of that method in (BAUMAN; TUZHILIN, 2014). In this way, for each item domain we randomly selected 300 reviews from the entire set of user-reviews (i.e., book, television, 8 User-ratings are classified in this high-level contextual value only when a more particular value could not be inferred, such as “family”. 9 In opposition to the “friends” value, in this contextual value are considered only co-workers, classmates, etc. 102 Chapter 4. CD-CARS Implementation and music - 900 reviews in total), however, we let 50 reviews for each contextual value (i.e., alone, accompanied, family, friends, partner, and colleagues). Hence, we manually labeled these reviews according to their contextual values and measured the accuracy of the contextual classification by comparing the labeled reviews to the classified reviews. The accuracy was calculated considering the number of correct classifications in comparison to the total of tested reviews. Table 6 reports the results of this empirical evaluation considering the different domains and contextual values. Table 6 – Classification accuracy of the companion extraction. Target Domain Overall Accuracy Contextual Values Alone Accompanied Family Friends Partner Colleague Book 19.67% 94% 3% 8% 2% 8% 3% TV 17% 76% 9% 5% 4% 6% 2% Music 10.83% 52% 2% 4% 1% 5% 1% As it can be seen from table, the implemented method did not have a good performance in the companion extraction task. Book was the domain with better results in general, while the Music had the worst ones. This result may be associated with the length of the user reviews, which is greater in the Book domain in comparison to the other ones. In addition, the average implemented method achieved better results for the Alone contextual value than for other values. This result may have occurred due to the great presence of personal pronoun “I” in the user reviews, which can be considered as a “topic” by the implemented method. For that reason, that method infers that the user was “alone”. 4.1.2 Selecting Relevant Contextual Attributes and Values As mentioned in Section 2.3.4, there are several approaches to determine the relevance of a given type of contextual information (contextual dimension). If we consider that only relevant dimensions (Location, Temporal and companion) are present in the datasets, we have to determine the relevance of the contextual attributes of these dimensions. For that, each user-rating in the datasets has contextual information about all contextual dimensions and their attribute variations, as described below: • Temporal - in this dimension, we persisted two contextual attributes: day (with values: Sunday to Saturday) and day type (values: weekend and weekday); • Location - in this dimension, we persisted three contextual attributes: country, state, and city10; • Companion - in this dimension, we persisted one contextual attribute: companion type (with values: alone, accompanied, family, friends, partner, and colleagues). 10 112 countries, 380 states, and 2838 cities were found in the datasets used in this thesis 4.1. Dataset Acquisition 103 Given these contextual dimensions and their attributes, we applied a data mining method (as mentioned in Section 3.2.2) to select only the most relevant contextual attributes of each contextual dimension. For that, we adopted the InfoGainAttributeEval method from Weka (HALL et al., 2009), which evaluates the worth of an attribute by measuring the information gain for a “class”. The output of this method is a ranking of the attribute list which indicates the importance of the attributes in the task of classification. In our case, the task of classification serves to analyze the influence of the distinct contextual attributes in the user-rating value (class). In this way, we applied the Info- GainAttributeEval11 with the user-rating value as a class (five possible values: 1 to 5) and six attributes (day, day type, country, state, city, companion type) for the two datasets used in this thesis, considering all target domains separately. Table 7 and Table 8 report the information gain values12 of the contextual attributes in different target domains for these datasets. Table 7 – Information gain of contextual attributes in different target domains for the book-television dataset. Target Domain Temporal Dimension Location Dimension Companion Dimension Day Day Type Country State City Companion Type Book 2.6e-4 1.1e-5 2.3e-3 7.9e-3 2.8e-2 9.9e-5 TV 2.3e-4 8e-5 5e-3 1e-2 3.6e-2 4.6e-5 Table 8 – Information gain of contextual attributes in different target domains for the book-music dataset. Target Domain Temporal Dimension Location Dimension Companion Dimension Day Day Type Country State City Companion Type Book 3.9e-4 3.1e-5 2.2e-3 9e-3 3.2e-2 1.2e-4 Music 2.4e-4 2.6e-5 4.6e-3 8.3e-3 3.6e-2 2.5e-5 As these tables demonstrate, the information gain was similar in both datasets and their respective domains. For them, the “day” attribute was the most relevant in the Temporal dimension as well as the “city” attribute and the “companion type” attribute were the most worth in their respective contextual dimensions. In addition, the “city” attribute is the most relevant to them, followed by the “day” and “companion type” attributes, in that order. Therefore, we have chosen only these three attributes for the evaluation of the proposed CD-CARS. However, there is no guarantee that the quality of recommendation is better in the “city” attribute than in the others. The quality may depend on how good the recommendation algorithms explore the contextual information available. Thus, this analysis does not discard the necessity of experimental evaluations 11 The weka.attributeSelection.Ranker was the ranker used with the configuration: -T (generateRanking) -1.7976931348623157E308 (threshold) -N (startSet) -1 (numToSelect). 12 These values range between 0 and 1, where a higher value represents a more discriminating feature 104 Chapter 4. CD-CARS Implementation for measuring the quality of recommendations in different (or with less information gain) contexts. It is important to mention that this analysis considers each attribute independently, and so do not take into account any correlation between distinct contextual attributes. For that reason, we also made experimental evaluations combining the selected contextual attributes (see Chapter 5). Moreover, we selected just one contextual attribute for each contextual dimension, but, other relevant contextual attributes could be used in the Location dimension once not all user-ratings have information about their particular location (e.g. instead of “city”, a user can only have information about “country” or “state”). Therefore, a CD-CARS implementation could consider any contextual information available although, for evaluation purposes, we have selected only the contextual attributes with higher information gain per dimension in order to minimize the evaluation cost by considering several contextual attributes and their combinations. Another aspect of the contextual information selection refers to selecting the most relevant contextual values of the contextual attributes. In order to verified this aspect, we had generalized the Companion values into only two categories: “alone” or “not-alone”, which contained all the other values such as “accompanied”, “family”, “friends”, “partner”, and “colleagues”. However, we verified that the information gain was higher when the Companion values were more granular, so, we kept all the original companion values separately. However, we remember that the quality of the inferred contextual information is low, which may impact in the calculation of the information gain. 4.1.3 Cross-Domain Datasets Description In this section, we describe the properties of two extracted datasets. One of them for evaluating the CD-CARS in two more related domains (Book and Television, named as “book-television dataset”, described in Section 4.1.3.1) and another considering two less related domains (Book and Music, named as “book-music dataset”, described in Section 4.1.3.2). 4.1.3.1 Book-Television dataset Table 9 summarizes the properties of the “book-television dataset”, which can be split into two single-domain sub-datasets and three or more samples considering ratings from specific contextual dimensions. As it can be seen from the table, the Books domain has more ratings ('64% from total) than Television domain ('36% from total). Also, the table summarizes the cross-domain dataset properties according to three contextual dimensions (Temporal, Location and Companion). In the table, we can see that 100% of the ratings have information about Temporal dimension, while almost 45% and 20% of 4.1. Dataset Acquisition 105 them, respectively, have information about Location (city) and Companion dimensions. In Section 4.1.2, we described why these contextual dimensions were chosen. Table 9 – Cross-domain and single-domain “book-television dataset” properties with 100% of user overlap. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 15341 194615 1249949 81.47 6.42 Books (single-domain) 15341 165896 805102 52.48 4.85 Television (single-domain) 15341 28719 444847 28.99 15.48 Temporal context (both domains) 15341 194615 1249949 81.47 6.42 Location (city) context (both domains) 7405 118020 557018 75.22 4.72 Companion context (both domains) 13598 76295 251707 18.51 3.30 Location (city) AND Companion contexts (both domains) 6846 52032 131257 19.17 2.52 Note that, in Table 9, all ratings from the full cross-domain dataset have a Temporal context associated with13, since they have this information extracted from dates, as described in Section 4.1. For this reason, we omitted the properties of the combination between the Temporal and Location (or Companion) dimensions, since the number of ratings for these combinations is the same as the number of ratings considering only the Location (or Companion) dimension alone. For example, if we consider the dataset by selecting ratings only with Location (city) and Temporal contextual dimensions, then we will have a sample of the dataset with 557018 ratings, which is the same number of ratings from the Location (city) dimension alone. It is important to mention that Table 9 shows the properties of the “book-television dataset” with full overlap between users. However, samples of that dataset are used with other user overlap levels in order to perform a sensitivity evaluation (see Section 2.2.6.3). As mentioned in Section 4.1, we generated reduced versions from the full-overlapped datasets by removing all ratings for items from the target domain of the users chosen randomly according to the overlap percentage. In this way, Table 10 and Table 11 present, respectively, the “book-television dataset” properties by considering user overlap levels of 50% and 10% when Television is the target domain while Table 12 and Table 13 show the “book-television dataset” properties by considering the same user overlap levels when Book is the target domain. 4.1.3.2 Book-Music dataset Table 14 summarizes the properties of the “book-music dataset”. We can see on the table that the Books domain has more ratings ('72% from total) than Music domain ('28% from total). In addition, 100% of the ratings have information about Temporal 13 See the number of ratings in the 1st and 4th rows. 106 Chapter 4. CD-CARS Implementation Table 10 – “book-television dataset” properties with 50% of user overlap when “TV” is the target domain. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 15341 188402 1011324 65.92 5.37 Books (single-domain) 15341 165896 805102 52.48 4.85 Television (single-domain) 7671 22506 206222 26.88 9.16 Temporal context (both domains) 15341 188402 1011324 65.92 5.37 Location (city) context (both domains) 7405 113049 446297 60.27 3.95 Companion context (both domains) 13012 73353 206512 15.87 2.82 Location (city) AND Companion contexts (both domains) 6571 49216 106664 16.24 2.17 Table 11 – “book-television dataset” properties with 10% of user overlap when “TV” is the target domain. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 15341 178646 851680 55.57 4.38 Books (single-domain) 15341 165896 805102 52.48 4.85 Television (single-domain) 1534 12750 46578 30.36 3.65 Temporal context (both domains) 15341 178646 851680 55.57 4.38 Location (city) context (both domains) 7405 103862 529307 71.48 5.10 Companion context (both domains) 12546 68084 332271 26.49 4.88 Location (city) AND Companion contexts (both domains) 6363 44830 248592 39.06 5.55 Table 12 – “book-television dataset” properties with 50% of user overlap when “Book” is the target domain. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 15341 131456 819335 53.41 6.23 Books (single-domain) 7671 102737 374488 48.82 3.65 Television (single-domain) 15341 28719 444847 28.99 15.48 Temporal context (both domains) 15341 131456 819335 53.41 6.23 Location (city) context (both domains) 7405 90838 581614 78.54 6.40 Companion context (both domains) 12425 53684 361950 29.13 6.74 Location (city) AND Companion contexts (both domains) 6298 36102 281688 44.73 7.80 dimension, while almost 46% and 11% of them, respectively, have information about Location (city) and Companion dimensions. Note that, due to the same reason mentioned in Section 4.1.3.1, we omitted the properties of the combination between other contextual dimensions, remaining only the combination between the Location and Companion dimensions. 4.1. Dataset Acquisition 107 Table 13 – “book-television dataset” properties with 10% of user overlap when “Book” is the target domain. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 15341 62722 514965 33.57 8.21 Books (single-domain) 1534 34003 70118 45.71 2.06 Television (single-domain) 15341 28719 444847 28.99 15.48 Temporal context (both domains) 15341 62722 514965 33.57 8.21 Location (city) context (both domains) 7087 43859 244913 34.56 5.58 Companion context (both domains) 11420 24291 104353 9.14 4.30 Location (city) AND Companion contexts (both domains) 5788 17147 55803 9.64 3.25 Table 14 – Cross-domain and single-domain “book-music dataset” properties with 100% of user overlap. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 13189 219034 1031386 78.20 4.71 Books (single-domain) 13189 162449 742844 56.32 4.57 Music (single-domain) 13189 56585 288542 21.88 5.10 Temporal context (both domains) 13189 219034 1031386 78.20 4.71 Location (city) context (both domains) 6951 132830 478510 68.84 3.60 Companion context (both domains) 11519 75754 207010 17.97 2.73 Location (city) AND Companion contexts (both domains) 6412 53999 116100 18.11 2.15 Table 15 – “book-music dataset” properties with 50% of user overlap when “Music” is the target domain. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 13189 208427 897227 68.03 4.31 Books (single-domain) 13189 162449 742844 56.32 4.57 Music (single-domain) 6595 45978 154383 23.41 3.36 Temporal context (both domains) 13189 208427 897227 68.03 4.31 Location (city) context (both domains) 6921 120509 407003 58.81 3.38 Companion context (both domains) 11315 75233 197571 17.46 2.63 Location (city) AND Companion contexts (both domains) 6303 53474 110730 17.57 2.07 Regarding the sensitivity analysis, Table 15 and Table 16 present, respectively, the “book-music dataset” properties by considering user overlap levels of 50% and 10% when “Music” is the target domain while Table 17 and Table 18 show the “book-music dataset” properties by considering the same user overlap levels when Book is the target domain. 108 Chapter 4. CD-CARS Implementation Table 16 – “book-music dataset” properties with 10% of user overlap when “Music” is the target domain. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 13189 177368 770030 58.38 4.34 Books (single-domain) 13189 162449 742844 56.32 4.57 Music (single-domain) 1319 14919 27186 20.61 1.82 Temporal context (both domains) 13189 177368 770030 58.38 4.34 Location (city) context (both domains) 6897 102369 350664 50.84 3.43 Companion context (both domains) 11141 74144 190478 17.10 2.57 Location (city) AND Companion contexts (both domains) 6203 52438 106797 17.22 2.04 Table 17 – “book-music dataset” properties with 50% of user overlap when “Book” is the target domain. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 13189 154329 635947 48.22 4.12 Books (single-domain) 6595 97744 347405 52.68 3.56 Music (single-domain) 13189 56585 288542 21.88 5.10 Temporal context (both domains) 13189 154329 635947 48.22 4.12 Location (city) context (both domains) 6482 104249 317421 48.97 3.04 Companion context (both domains) 8209 51833 116360 14.17 2.25 Location (city) AND Companion contexts (both domains) 4578 36396 66499 14.53 1.83 Table 18 – “book-music dataset” properties with 10% of user overlap when “Book” is the target domain. Dataset Users Items Ratings Ratings perUser Item Cross-domain (both domains) 13189 87993 347805 26.37 3.95 Books (single-domain) 1319 31408 59263 44.93 1.89 Music (single-domain) 13189 56585 288542 21.88 5.10 Temporal context (both domains) 13189 87993 347805 26.37 3.95 Location (city) context (both domains) 5945 57465 172780 29.06 3.00 Companion context (both domains) 5454 15949 35785 6.56 2.24 Location (city) AND Companion contexts (both domains) 3071 10310 19971 6.50 1.94 4.2 Contextual Model Implementation In this section, we describe how the contextual model, outlined in Section 3.2, was implemented. To implement this contextual model, we extended the implementation of 4.2. Contextual Model Implementation 109 the Mahout14 framework (OWEN et al., 2011). Figure 23 and Figure 24 show two class diagrams that represent the contextual data model considering the extension/realization of three Mahout entities (RecommenderBuilder, AbstractDataModel and IDRescorer). Figure 23 – Data model class diagram focusing contextual aspects of the CD-CARS implementation. Figure 24 – Data model class diagram focusing dataset aspects of the CD-CARS imple- mentation. 14 Apache Mahout open source project is a machine learning library under Apache Software Foundation - http://mahout.apache.org/ 110 Chapter 4. CD-CARS Implementation In the following, we describe the main entities represented in those class diagrams: • RecommenderBuilder - this Mahout interface guides the implementation of classes responsible for building recommender algorithms in order to be evaluated based on a given realization of the AbstractDataModel Mahout abstract class. In turn, these recommender algorithms must implement the Recommender Mahout interface to recommend items for a user. • AbstractDataModel - this Mahout abstract class implements some basic methods defined by the DataModel Mahout interface, which represents a repository of infor- mation about users and their associated preferences for items (i.e., user-ratings). • ContextualRecommenderBuilder - this novel interface extends the Recommender- Builder Mahout interface in order to guide the building of recommender algorithms capable of taking advantage of context-awareness features. • PreFilteringContextualRecommenderBuilder - this class implements the Contextu- alRecommenderBuilder interface. It is responsible for building and preparing the pre-filtering recommender algorithm (see Section 3.3.1.1) for evaluation according to three parameters: target domain (restricting source domains declared in the ItemDo- mainRescorer class, showed in Figure 24), contextual data model (represented by the ContextualDataModel class), and the context of the recommendation (represented by the ContextualCriteria class). • PostFilteringContextualRecommenderBuilder - this class implements the Contextual- RecommenderBuilder interface and follows the same logic as the PreFilteringContex- tualRecommenderBuilder, but, considering the post-filtering recommender algorithm (see Section 3.3.1.2) instead of the pre-filtering one. • BaseCrossDomainRecommenderBuilder - this class implements the ContextualRec- ommenderBuilder interface. It is responsible for building and preparing the base cross-domain recommender algorithm (see Section 3.3.2) for evaluation according to two parameters: target domain (restricting source domains declared in the Item- DomainRescorer class, described later) and contextual data model (represented by the ContextualDataModel class). In contrast to the PreFilteringContextualRec- ommenderBuilder and the PostFilteringContextualRecommenderBuilder, this class does not take into account the context of the recommendation (represented by the ContextualCriteria class), since it does not use any contextual information to recommend items in the target domain. However, it is used by the PreFiltering- ContextualRecommenderBuilder and PostFilteringContextualRecommenderBuilder classes that bring the contextual data model, which in turn can be used by the 4.2. Contextual Model Implementation 111 BaseCrossDomainRecommenderBuilder only for evaluation purposes, for example, in specific contexts as a baseline (without interfering in the recommendation process). • ContextualDataModel - while the AbstractDataModel Mahout abstract class contains only user-ratings, this novel class contains, besides the user-ratings, contextual information about these user-ratings, thus, the ContextualDataModel extends the AbstractDataModel. It is important to mention that the implementation of the AbstractDataModel is worried about performance issues (OWEN et al., 2011) and, for that reason, represents all preferences of a user through a PreferenceArray object. This object contains a single user ID, an array of item IDs, and an array of preference values. In our implementation, we extended the PreferenceArray with a multidimensional array of contextual feature codes (ContextualPreferenceArray). • ContextualCriteria - this novel class encapsulates the context of the recommendation represented by a list of contextual values (according to their contextual dimensions and attributes), which in turn are implemented as enumeration objects that realize the AbstractContextualAttribute interface. The list of contextual values is composed of six enumeration objects, one for each implemented AbstractContextualAttribute. • IDRescorer - this Mahout interface allows the realization of classes that can filter out items from the recommended item list according to several attributes such as an item genre (e.g. action, comedy, etc.) or an item domain (e.g. book, music,etc.). The developer that creates a class to implement the IDRescorer is free to choose the appropriate attribute for his purposes. Therefore, this Mahout interface makes the cross-domain recommendation possible, as well as the post-filtering recommendation (see Section 3.3.1.2), since both item genre and item domain can be used as a filter. • ItemDomainRescorer - this class implements the IDRescorer Mahout interface in order to filter out from the recommend item list those items that belong to the set of items from the source domains. These domains are specified through ItemDomain enumeration objects. For that, the ItemDomainRescorer depends on the information from a dataset (e.g. we created an AmazonCrossDataset class, which is an implementation of the AbstractDataset class and contains only three domains: Book, Television and Music). • ItemCategoryRescorer - this class implements the IDRescorer Mahout interface in order to filter out from the recommend item list those items that belong to the set of categories specified by ItemCategory enumeration objects. Thus, this class is useful for the post filtering recommendation algorithm and also depends on the information from a dataset, which contains a set of categories retrieved from the set of items. • AbstractDataset - this abstract class encapsulates, besides the contextual data model, a set of meta-information about its users (UserDatasetInformation), items 112 Chapter 4. CD-CARS Implementation (ItemDatasetInformation), addresses (AddressDatasetInformation) and association rules (AprioriRuleItemCategoryDomain). • AprioriRuleItemCategoryDomain - this class generates and persists a set of association rules used by the post filtering recommendation (see Section 3.3.1.2). These general rules are generated from all user preferences in the contextual data model, and relate item categories (ItemCategory enumeration objects) between different domains, e.g. who likes action movies also likes rock music - {ACTION, TV} => {ROCK, MUSIC}. In addition, for each generated rule, the class maintain its confidence and support levels, besides all combinations of contexts (represented by a ContextualCriteria) from the instances that generated that rule. These combination of contexts are used a posteriori by the PostFilteringStrategyRecommendation class (described in Section 4.3.2) that generated more specific rules (e.g. {ACTION, TV, WEEKDAY} => {ROCK, MUSIC, WEEKEND}) by considering particular contexts. This decision avoids the generation of several rules with context that are not used by the Post-Filtering algorithm, once that these rules only are used when a user does not have any category preferences in a specific context. • ItemDatasetInformation - this class represents the set of ItemInformation objects. An ItemInformation object contains information about an item, such as ID, name/title, year released, category, domain, link, and so on. • UserDatasetInformation - this class represents the set of UserInformation objects. A UserInformation object contains information about a user, such as ID, Amazon user ID, raw address (a not-standardized string), and so on. • AddressDatasetInformation - this class represents the set of AddressInformation objects. An AddressInformation object contains information about a standardized address, such as raw address string, city, state, country, and so on. 4.3 Proposed Algorithms Implementation In this thesis, we implemented two of the proposed algorithms (Section 3.3): Pre- Filtering (PreF) and Post-Filtering (PostF). In the next subsections, we will describe some implementation particularities of these algorithms considering the contextual model entities mentioned in Section 4.2 and other Mahout’s entities. 4.3.1 Pre-filtering Implementation We implemented the PreF algorithm according to its proposal, described in Sec- tion 3.3.1.1. The PreF implementation filters out from the ContextualDataModel the target domain user-ratings whose contexts (ContextualPreferenceArray) are “contained” 4.3. Proposed Algorithms Implementation 113 in the context of the recommendation (defined by the ContextualCriteria). In other words, the user-rating context must be the same context as the context of the recommendation without considering unknown contextual attributes of the user-rating context. Besides, it is important to remember that the user-ratings are only filtered in the target domain (ItemDo- main), which is verified through the ItemDomainRescorer and the ItemDatasetInformation classes (described before). For illustrating the pre-filtering process, consider a context of the recommendation and the contexts of a set of user-ratings, both respectively represented by Contextual- Criteria and ContextualPreferenceArray entities in Figure 25. While the ContextualPref- erenceArray contains a multidimensional array representing different user-ratings (first index of the array) and its contextual values in a sequence representing different con- textual attributes (second index of the array), the ContextualCriteria contains a list of enumeration objects representing different contextual attributes, which have a contextual code for each contextual value. The same order of contextual attributes declared in the ContextualPreferenceArray is automatically used in the ContextualCriteria, since the ContextualFileAttributeSequence instance determines a unified order of the contextual attributes. As it can be seen in Figure 25, the user (userId = ‘1’) has ratings for five items (itemIds from ‘1’ to ‘5’), where each rating for an item has an array of contextual codes in a specific order (e.g. itemIds[0] = 1, ratings[0]=4.0, contextualPreferences[0] = {1,1,0,-1,-1,- 1}). Each contextual code represents a contextual value of a different contextual attribute in the sequence defined by the ContextualFileAttributeSequence instance15. In this way, considering that the ContextualCriteria from the figure can be expressed by the following contextual code sequence: {1 (“SUNDAY”), 1 (“WEEKEND”), -1 (“UNKNOWN”),-1 (“UNKNOWN”),-1 (“UNKNOWN”),2 (“FAMILY”)}, so, only one user-rating (contex- tualPreferences[4]) is discarded, whereas the others four ones (contextualPreferences[0] to contextualPreferences[3]) are maintained by the PreF implementation. This process is illustrated in Figure 26, in which a green square means “a matching” between the contextual attribute values from the user-rating contexts and the recommendation context, whereas a red square means “a non-matching”. Finally, a yellow square means that the contextual attribute was not considered for matching between the user-rating contexts and the recommendation context, since one (or both) of these contextual attributes is “unknown”. 15 The six codes are represented respectively by the DayContextualAttribute, DayTypeContextualAt- tribute, LocationCountryContextualAttribute, LocationStateContextualAttribute, LocationCityContextu- alAttribute, and CompanionContextualAttribute enumerations 114 Chapter 4. CD-CARS Implementation Figure 25 – Class diagram illustrating entities used by the pre-filtering class. 4.3. Proposed Algorithms Implementation 115 Figure 26 – Example of the pre-filtering process considering the context of user-ratings and the recommendation context. 4.3.2 Post-filtering Implementation In contrast to the PreF proposal, the PostF one (described in Section 3.3.1.2) allows that several strategies be implemented. Thus, we investigated some strategies to perform the PostF recommendation by varying its threshold value (θ). For instance, we set θ to 2/3 of the frequency of the most preferred category. So, suppose that a user has given good ratings (at least 4.0 in a scale from 1.0 to 5.0), in a given context, for thirty religion books, twenty-five educational books, twenty comedy books, nineteen romance books, and ten action books, as illustrated in Figure 27. By applying the threshold strategy, the minimal value of occurrences is twenty (2/3 of 30, which is the frequency of the most preferred category - religion) for an item category to be maintained on the recommendation list. So, only religion (30 occurrences), comedy (20 occurrences) and educational books (25 occurrences) are included in the resulting recommendation, i.e., books of other categories are ignored such as romance (19 occurrences) and action (10 occurrences) ones. The optimal value of θ can be set through experiments, as well as the minimal value to consider a good rating, which will vary depending on the rating value interval (e.g. in an interval of 1-10 we could consider 8.0 or more like a good rating). The higher the θ value, the less the number of categories included in the users’ preferred categories. 116 Chapter 4. CD-CARS Implementation Figure 27 – Example of selected categories in the post-filtering recommendation. Figure 28 shows the main entities of the post-filtering implementation. From this figure, we can see that PostFilteringStrategyRecommendation class uses two distinct databases of preferred categories: • UserCategoriesPrefsInContextsByDomain - this database contains all categories of rated items for each user in the dataset. However, only the categories of well-rated (or “good”) items were considered (at least 4.0 in a scale from 1.0 to 5.0). Besides, the number of occurrences of each category is associated with their observed contexts and domains (e.g. {RELIGION,BOOK,WEEKDAY} = 5, which means that a user rated five books as “good” in weekdays). Thus, this class contains one contextual preference tensor CP(u,c,g) (described in Section 3.3.1.2) for each item domain. • CategoryContextDomainRulesMap - when a user does not have any information about preferred categories in a given domain, general association rules are necessary so that the post-filtering algorithm can recommend items even in a domain that the user has not ratings (see Section 3.3.1.2). In this way, this database increments the set of rules initially generated by the AprioriRuleItemCategoryDomain class with contextual information. For instance, a {RELIGION,BOOK} => {RELIGION,TV} rule from the AprioriRuleItemCategoryDomain could be transformed into a {RE- LIGION,BOOK,WEEKDAY} => {RELIGION,TV,WEEKEND} rule. Each part 4.3. Proposed Algorithms Implementation 117 of a rule is represented by a RuleTuple class, which contains information about the context (ContextualCriteria), item category (ItemCategory) and item domain (ItemDomain). In this case, rules are composed by one precedent RuleTuple and one consequent RuleTuple. Later in this section, we will detail the implementation of the process of association rules generation. Figure 28 – A class diagram illustrating the main post-filtering entities. It is important to mention that these databases are updated periodically on demand, i.e., in execution time just when it is necessary. We made this design decision once that we are concerned about performance issues given that the generation of association rules is costly if we consider the multi-dimensional RuleTuple entities (item category, item domain and context). In order to generate these rules, we used the AprioriRuleItemCategoryDomain and CategoryContextDomainRulesMap implementations in a two-step process described in Algorithm 2 (1-step) and in Algorithm 3 (2-step). Algorithm 2. AprioriRuleItemCategoryDomain algorithm for association rules generation (1-step). Input: d (dataset), cl (minimal confidence level), sl (minimal support level), gr (minimal “good” rating threshold value) Output: rs (rules with category and domain information) 118 Chapter 4. CD-CARS Implementation 1: procedure generatePreferredCategoriesMapByUser(d,gr) 2: Create a uc data structure containing a list of users where each user (u) has a set of categories (cs) 3: for each u ∈ d do 4: Get the array (ur) of user ratings (r) from u 5: for each r ∈ ur do 6: v = value from r 7: if v ≥ gr then 8: i = item from r 9: ic = categories from i 10: acs = cs from uc for u 11: for each c ∈ ic do 12: if c * acs then 13: Add c in the acs 14: Add acs in uc for u 15: end if 16: end for 17: end if 18: end for 19: end for 20: return uc 21: end procedure 1: procedure generateAprioriCategoryDomainRules(uc, d, cl, sl) 2: for i = 0; i < size of the categories set from d; i = i + 1 do 3: for j = 0; j < size of the categories set from d; j = j + 1 do 4: di = domain from category i 5: dj = domain from category j 6: if di 6= dj then 7: Create a precedent tuple (pt) composed by the category i and di 8: Create a consequent tuple (ct) composed by the category j and dj 9: ptc = 0 10: ctc = 0 11: for each u ∈ uc do 12: acs = cs from uc for u 13: if category i ⊂ acs then 14: ptc = ptc + 1 15: if category j ⊂ acs then 16: ctc = ctc + 1 17: end if 4.3. Proposed Algorithms Implementation 119 18: end if 19: end for 20: Create a rule (r) composed by pt, ct, ptc and ctc. 21: Add r in rs 22: end if 23: end for 24: end for 25: for each r ∈ rs do 26: Get ptc and ctc from r 27: nu = number of users from uc 28: rcl = ctc/ptc 29: rsl = ctc/nu 30: if rcl < cl or rsl < sl then 31: Remove r from rs 32: end if 33: end for 34: return rs 35: end procedure end Algorithm 3. CategoryContextDomainRulesMap algorithm for association rules genera- tion (2-step). Input: rs, which is a set of rules from the 1-step algorithm; d (dataset); gr (minimal “good” rating threshold value); uc data structure containing a list of users where each user (u) has a set of categories (cs); a rule tuple condition (rtc) composed by an item category ict, an item domain idt and contextual criteria (cc); cl (minimal confidence level); and sl (minimal support level) Output: crs (set of rules with category, domain and contextual information) 1: procedure addContextualInformationInAprioriCategoryDomainRules(rs, d, gr, uc) 2: for each rule ∈ rs do 3: for each u ∈ d do 4: Get cs from uc for u 5: Get the precedent tuple (pt) from rule 6: Get the consequent tuple (ct) from rule 7: Get ic1 from pt 8: Get ic2 from ct 120 Chapter 4. CD-CARS Implementation 9: if ic1 ⊂ cs and ic2 ⊂ cs then 10: Get the array (ur) of user ratings (r) from u 11: Create an empty set of contexts for the rule precedent category (cpt) 12: Create an empty set of contexts for the rule consequent category (cct) 13: for each r ∈ ur do 14: v = value from r 15: if v ≥ gr then 16: i = item from r 17: ic = categories from i 18: if ic1 ⊂ ic then 19: urc = context from r 20: Add urc in cpt 21: else 22: if ic2 ⊂ ic then 23: Add urc in cct 24: end if 25: end if 26: end if 27: end for 28: if cpt 6= ∅ and cct 6= ∅ then 29: for i = 0; i < cpt size; i = i + 1 do 30: for j = 0; j < cct size; j = j + 1 do 31: Add context i from cpt in pt for the rule 32: Add context j from cct in ct for the rule 33: end for 34: end for 35: end if 36: end if 37: end for 38: end for 39: 40: end procedure 1: procedure getAprioriCategoryDomainContextRules(rs, rtc, cl, sl) 2: for each rule ∈ rs do 3: Get the precedent tuple (pt) from rule 4: Get the consequent tuple (ct) from rule 5: Get the item category ( ic) from pt 6: Get the item domain ( id) from pt 7: Get ict from rtc 4.3. Proposed Algorithms Implementation 121 8: Get idt from rtc 9: Get cc from rtc 10: if ic = ict and id = idt then 11: Get the total number contexts (nc) from pt 12: ptc = 0 13: ctc = 0 14: for i = 0; i < nc; i = i + 1 do 15: if contextualMatching(cc,context i from pt,size of the array cc) then //According to Algorithm 1 16: ptc = ptc + 1 17: if contextualMatching(cc,context i from ct,size of the array cc) then 18: ctc = ctc + 1 19: end if 20: end if 21: end for 22: rcl = ctc/ptc 23: rsl = ctc/nc 24: if rcl < cl and rsl < sl then 25: Get the item category ( icc) from ct 26: Get the item domain ( idc) from ct 27: Create a contextual rule (cr) containing the rtc as precedent, and a consequent rule tuple composed by icc, idc and cc 28: Add cr in crs 29: end if 30: end if 31: end for 32: return crs 33: end procedure end Both algorithms described above are based on Apriori algorithm (AGRAWAL; IMIELIŃSKI; SWAMI, 1993). They can be seen as simplified versions of it since we are interested only in a subset of rules. More precisely, once that the PostF algorithm only uses the rules base when a user does not have contextual preferences in the recommenda- tion domain, so, only rules that relate categories among different domains (source and target) are necessary. For example, we could obtain rules like “{RELIGION,BOOK} => {RELIGION,TV}” or “{ACTION,TV} => {ROCK,MUSIC}” from Algorithm 2. 122 Chapter 4. CD-CARS Implementation As it can be seen in this algorithm, first, we perform the GeneratePreferredCate- goriesMapByUser procedure considering only good ratings (we set the minGoodRatingValue to 4.0), and then we apply the GenerateAprioriCategoryDomainRules procedure, resulting in a 1-step rule base which contains rules among different domains considering only their item categories. For that, this procedure makes all possible combinations of precedent and consequent item categories of different domains (see lines 2 and 3 in the GenerateApriori- CategoryDomainRules procedure). Furthermore, there are two Apriori parameters (minConfidenceLevel and minSup- portLevel) that can be used to define a minimal threshold for considering rules or not. If the minimum confidence and support levels for mining the rules are high, then the algorithm may not obtain enough rules for the PostF recommendation. Again, these rules are only required when a user does not have any category preference in the context of the recommendation for the target domain. Therefore, for the datasets used in this implementation, we set the minConfi- denceLevel to 0.7 and the minSupportLevel to 0.01. Thus, the 1-step algorithm obtained 25 rules16 for the “book-television dataset” and 23 rules17 for the “book-music dataset”, both with full user overlap. If we take into account the number of categories in the three domains (27 for Book and Television, and 19 for Music), then we can say that almost a half of these categories have a rule with an associated category, inferred by the 1-step algorithm. This also means that the PostF algorithm will not be able to recommend items of the categories that do not have any association rule. We used the same values for minConfidenceLevel and minSupportLevel parameters in Algorithm 3. This 2-step algorithm has the set of rules generated by the 1-step algorithm as an input and generates a rule base containing rules among different domains considering their item categories and also their contexts. In this case, examples of inferred rules could be like “{ACTION,TV,WEEKEND_FAMILY} => {ROCK,MUSIC,WEEKEND_FAMILY}” or “{RELIGION,BOOK,WEEKDAY_CANADA} => {GOSPEL,MUSIC,WEEKDAY_ CANADA}”. Note that both precedent and consequent contexts of these rules are the same (according to the line 27 in the getAprioriCategoryDomainContextRules procedure). This rule “kind” (with the same precedent and consequent contexts) is sufficient for our purposes since the PostF algorithm only needs to infer a set of preferred categories in the target domain for a user with preferred categories in the source domain according to the 16 Examples of generated rules: {DRAMA,TV => FANTASY,BOOK} (confidence=0.75 and sup- port=0.5), {ARTISTIC,BOOK => DRAMA,TV} (confidence=0.70 and support=0.12), {WEST- ERNS,TV => DOCUMENTARY,BOOK} (confidence=0.70 and support=0.09), {HEALTH,TV => EDUCATIONAL,BOOK} (confidence=0.74 and support=0.03), among others. 17 Examples of generated rules: {POP,MUSIC => FANTASY,BOOK} (confidence=0.74 and sup- port=0.33), {CLASSICAL,MUSIC => DOCUMENTARY,BOOK} (confidence=0.73 and sup- port=0.14), {NEW AGE,MUSIC => EDUCATIONAL,BOOK} (confidence=0.72 and support=0.05), {KIDS,BOOK => KIDS,MUSIC} (confidence=0.78 and support=0.03), among others. 4.4. Base Cross-domain Algorithm Implementation 123 context of the recommendation. This context is used to filter the users’ preferred item categories from the source domain in order to obtain the inferred item categories in the target domain. Finally, it is important to remember that the PostF algorithm is applied after the base cross-domain recommendation by filtering recommended items out in the target domain according to the users’ item category preferences (inferred or not). 4.4 Base Cross-domain Algorithm Implementation As mentioned in Section 3.3.2, we apply collaborative filtering algorithms as base cross-domain algorithm. In this implementation, we adopted a neighborhood-based (user- based similarity) algorithm, due to its simplicity and based on a preliminary battery of experiments using other CF-based algorithms (e.g. item-based neighborhood and matrix factorization). This algorithm also has been used as a baseline for cross-domain recommendation purposes in (CREMONESI; TRIPODI; TURRIN, 2011), which proposed an enhanced version of that user-based algorithm, aiming to make cross-domain CF recommendations under user overlap conditions. The implementation of these algorithms is described in the following. Algorithm 4. Item rating estimation with the implementation of the NNUserNgbr algo- rithm. Input: ux (the user), un (the user neighborhood calculated by any traditional similarity metric), i (the item, from the target domain) Output: er (estimated rating) 1: procedure estimatePreferenceInNNUserNgbr(ux, un, i) 2: if un 6= ∅ then 3: p = 0.0 4: ts = 0.0 5: c = 0 6: for each u ∈ un do 7: if u 6= ux then 8: if u has preference for i then 9: Get the preference value (pv) of u for i 10: Get the similarity value (sv) between u and ux 11: p = p + (sv ×pv) 12: ts = ts + sv 13: c = c + 1 14: end if 15: end if 124 Chapter 4. CD-CARS Implementation 16: end for 17: if c > 0 and ts 6= 0.0 then 18: er = p/ts return er 19: end if 20: end if 21: end procedure end Algorithm 4 presents the item rating estimation with the implementation of the NNUserNgbr algorithm. It is important to notice that the item must be from the target domain, so, the cross-domain recommendation can be seen as a reduced version of a traditional single-domain CF-based recommendation. Besides, the user neighborhood can be calculated by any similarity metric described in Section 3.3.2.1.1. In addition, we implemented the enhanced version of the user-based algorithm proposed in (CREMONESI; TRIPODI; TURRIN, 2011) (NNUserNgbr-transClosure), as mentioned in Section 3.3.2.1.1. This algorithm differs from the NNUserNgbr only by the fact that the user’s neighborhood calculation is extended. Thus, the Algorithm 4 also represents the item estimation process of the NNUserNgbr-transClosure algorithm considering the calculation of the extended user neighborhood, which is presented in Algorithm 5. Algorithm 5. User neighborhood calculation for the NNUserNgbr-transClosure algorithm. Input: ux (user), un (the user neighborhood calculated by any traditional similarity metric), mn (neighborhood size limit) Output: unt 1: procedure userNeighborhoodInNNUserNgbr-transClosure(ux, un,mn) 2: if un 6= ∅ then 3: Create a data structure (tm) containing a list of users where each user u has a set of similar users with their respective similarity values (sv) 4: for each uA ∈ un do 5: if uA 6= ux then 6: Get the uA neighborhood (uAn) according to any traditional similarity metric 7: for each uB ∈ uAn do 8: if uB * un and uB 6= ux and uB 6= uA then 9: Get the similarity value (svAB) between uB and uA 10: if uB * tm then 4.4. Base Cross-domain Algorithm Implementation 125 11: Create a data structure of users (su) containing the uA associated with the svAB value 12: Add su in tm for uB 13: else 14: Get su in tm for uB 15: Add uA associated with the svAB value in su 16: end if 17: end if 18: end for 19: end if 20: end for 21: unt = un 22: Get the minimum similarity value (msv) from unt 23: Get the less similar user ( lsu) from unt 24: for each uA ∈ tm do 25: s = 0.0 26: c = 0 27: Get su from uA 28: for each uB ∈ su do 29: Get svAB from su for uB 30: Get the similarity value between uB and ux (svBU) from unt 31: s = s + (svAB ×svBU) 32: c = c + 1 33: end for 34: if c > 0 then 35: ns = s/c 36: Get number of similar users ∈ unt (nsu) 37: if mn > nsu then 38: Add in unt the uA associated with the respective ns value 39: if ns < msv then 40: msv = ns 41: lsu = uA 42: end if 43: else 44: if ns > msv then 45: Add in unt the uA associated with the respective ns value 46: Remove lsu from unt 47: Get the minimum similarity value (msv) from unt 48: Get the less similar user ( lsu) from unt 126 Chapter 4. CD-CARS Implementation 49: end if 50: end if 51: end if 52: end for 53: end if 54: return unt 55: end procedure end Note that the implemented user neighborhood calculation in Algorithm 5 can be considered as a two-step similarity path between two users, as described in Section 3.3.2.1.1. Besides, a similarity metric is used in this calculation (see line 6 in the userNeighborhoodIn- NNUserNgbr-transClosure procedure) once that the “transclosure” process only extends the similarities discovery among users. Finally, the maximum number of ’nearest neighbors’ is maintained by the “transclosure” process, which eliminates the smaller user similarities from the user neighborhood (see lines 46 to 48 from the userNeighborhoodInNNUserNgbr- transClosure procedure). 4.5 Final Remarks In this chapter, we presented particular details of an implementation of two proposed CD-CARS algorithms (PreF and PostF) as well as the implementation of two base cross-domain algorithms (NNUserNgbr and NNUserNgbr-transClosure). In addition, we showed the properties of two CD-CARS datasets (“book-television” and “book-music”) with different contextual information (Temporal, Location and Companion) and domains (Book, Television and Music) and how they were obtained. Also in this chapter, we described the process of selecting relevant contextual attributes and values through a data mining method (InfoGainAttributeEval from Weka tool (HALL et al., 2009)). Finally, we presented the implementation of the contextual model through the extension of the Mahout framework (OWEN et al., 2011). In the next chapter, we describe and discuss experimental evaluations of the implemented algorithms through the CD-CARS datasets. 127 5 CD-CARS Evaluation This chapter presents an experimental evaluation of two proposed CD-CARS algorithms in comparison to cross-domain CF-based ones. For that, Section 5.1 describes the evaluation methodology adopted and the algorithms’ settings. Section 5.2 describes the evaluation results for each dataset used in the experiments as well as a discussion about their findings. Finally, Section 5.3 presents the final remarks of this chapter. 5.1 Evaluation Methodology In this section, we describe the adopted methodology to evaluate the proposed algorithms as well as their configurations. Besides, we describe how the statistical significance of the results is verified. 5.1.1 Settings of the Algorithms Before evaluating the proposed CD-CARS algorithms, we performed a preliminary battery of experiments in the two datasets mentioned in Section 4.1.3 in order to adjust the settings of the base single-domain CF-based algorithm (NNUserNgbr) adopted in the CD-CARS evaluation. As mentioned in Section 3.3.2.1.1, that algorithm can also be used to perform cross-domain recommendations, thus, we intend to verify its performance for single-domain and cross-domain scenarios. In this way, we adjusted the NNUserNgbr settings according to several experiments performed in the Book domain for each dataset, i.e., performing a single-domain recom- mendation, once that Book is a common domain to the both datasets. We set the ‘n’ parameter of the NNUserNgbr algorithm to “475” and selected the Euclidian distance as the similarity metric for it. The same configuration and similarity metric were adopted for another base cross-domain CF-based algorithm (NNUserNgbr-transClosure) and for evaluation in other domains (Television and Music). As the proposed CD-CARS algorithms, PreF and PostF, can be performed in combination with the base NNUserNgbr and NN-UserNgbr-transClosure ones, so, the base algorithms were used with the same settings described before. In addition to these settings, we set the PostF threshold (θ) value to “2/3” of the frequency of the most preferred category, and only the categories of items that had good ratings (four or more in a five-star scale) were considered in computation of the frequency of the users’ preferred categories (see Section 3.3.1.2). In the PostF algorithm, we also set “0.7” and “0.01”, respectively, for the association rule confidence and support levels. This decision was also made by 128 Chapter 5. CD-CARS Evaluation considering preliminary experiments in the Book domain as target by observing the PostF performance in the two datasets. We evaluated the proposed algorithms in comparison to the baseline ones by using Probabilistic and Classification measures, as described in the following sections. 5.1.2 Predictive Performance We measured the predictive performance of the algorithms by using the Mean Average Error (MAE) and Root Mean Squared Error (RMSE) metrics (SHANI; GU- NAWARDANA, 2011). MAE is a measure of the deviation of recommendations from their actual user-rating values. For each ratings-prediction pair (pi,qi) this metric treats the absolute error between them. The MAE is computed by first summing these absolute errors of the ‘N’ corresponding ratings-prediction pairs and then computing the average. Formally, MAE = ∑N i=1 |pi − qi| N (5.1) Analogously, RMSE computes the square root of the average of the square of all of the error (punishing large errors), by means of this formula: RMSE = √ ∑N i=1(pi − qi)2 N (5.2) These metrics evaluate the performance of a RS by comparing the numerical recommendation scores against the actual user ratings for the user-item pairs in the test dataset. In this way, for a single dataset adopted in the CD-CARS evaluation, we split it into the training and test sets for each target domain (e.g. Music) and context under test (e.g. “on sunday with friends”). The training set is composed by 100% of ratings from source domain, 100% of ratings from the target domain in which their contexts are not under test and 90% of ratings from target domain in which their contexts are under test. The test set is composed by 10% of ratings from target domain in which their contexts are under test. Figure 29 illustrates the process of splitting training and test sets considering the target domain and context under test. It avoids the waste of ratings in the test set for those ones that are not used in the target domain and context under test. That process can be seen as Hold-out, according to the partitioning ways of evaluation data described in Section 2.2.6.1. Finally, for each target domain and context under test, we performed each evaluated algorithm five times in order to verify its standard deviation and apply statistical tests. 5.1. Evaluation Methodology 129 Figure 29 – Splitting training and test sets considering the target domain and context under test. 5.1.3 Classification Performance Regarding the classification performance of the algorithms, we adopted the F-metric proposed by Cremonesi, Tripodi e Turrin (2011), which is calculated according to the Precision and Recall values used for evaluating top-N recommendations and obtained through a testing methodology described in (CREMONESI; KOREN; TURRIN, 2010). Analogously to (CREMONESI; KOREN; TURRIN, 2010), we randomly extracted, approximately, 1.4% of the ratings from the original dataset in order to build a probe set, therefore, the training set was composed by 98.6% of the rating from the full dataset. On the other hand, the test set was composed exclusively by the 5-star ratings (the maximum rating value for that evaluation dataset) from the probe set, thus, the not 5-star ratings from the probe set were discarded. However, we adapted this methodology by considering the target domain and context under test in order to fulfill the training and probe sets in a similar way than is made in the probabilistic evaluation, as illustrated in Figure 29. This avoids the waste of ratings in the probe set for those ones that are not used in the target domain and context under test. Likewise the predictive performance evaluation, the dataset used in the classification performance is partitioned as Hold-out, according to the partitioning ways of evaluation data described in Section 2.2.6.1. After those steps, we trained the algorithms with the training set and for each rating in the test set, given by a user ‘u’ for an item ‘i’ from the target domain: • We predict the ratings for the item ‘i’ and for 100 additional items1 from the target 1 The original method empirically adopts 1000 additional items, but the authors let this number free to 130 Chapter 5. CD-CARS Evaluation domain randomly chosen from the ones unrated by the user ‘u’; and • In decreasing order, we sort the list of 101 items2 according to the predicted ratings. If the item ‘i’ appears in the top-N recommendation list, we have a “hit”. In this way, Precision, Recall and F-metric values, according to (CREMONESI; KOREN; TURRIN, 2010), are defined as: Recall(N) = #hits |test set| (5.3) Precision(N) = #hits N ∗ |test set| (5.4) F-metric(N) = 2 ∗Recall(N) ∗Precision(N) Recall(N) + Precision(N) (5.5) Likewise the probabilistic evaluation, for each target domain and context under test, we performed each evaluated algorithm five times. The execution of several trials is not specified by the methodology proposed in (CREMONESI; KOREN; TURRIN, 2010), but we believe that as more executions are made the more reliable should be the results. Finally, in the evaluation results of the algorithms for a particular target domain and user overlap level, we show their classification performances through the F-metric curves by varying the number of top ‘N’ items (from one to twenty3). Besides, given that most of the online recommender systems (e.g. Amazon4, IMDB5, etc.) usually recommend up to five items in their basic layout (CREMONESI; TRIPODI; TURRIN, 2011), we fixed the top ‘N’ value to “five” to verify the variation of the F-metric value across different user overlap levels (sensitivity evaluation). 5.1.4 Sensitivity Evaluation As mentioned in Section 2.2.6.3, the performance of a cross-domain RS can be affected by the density of the target domain data and the user overlap between source and target domains. In this way, we evaluated the quality of the cross-domain algorithms by varying the percentage of the user overlap (10%, 50%, and 100%). Section 4.1.3 describes the properties of the datasets adopted in the CD-CARS evaluation regarding the different overlap levels. be chosen depending on the dataset used. 2 In the original method, the authors adopted 1001 items. 3 We have chosen this maximum top ‘N’ value by observing the convergence in the F-metric curves of the algorithms. 4 http://www.amazon.com 5 http://www.imdb.com 5.2. Evaluation Results 131 In addition, we studied the impact of the density of the target domain data in comparison to the density of the source domain data. Thus, we evaluated the quality of the cross-domain algorithms by varying the target domain in both datasets. One of them for more related domains (Book and Television), where the Book domain has more data than the Television domain, and another for less related domains (Book and Music), also where the Book domain has more data than the Music domain. Therefore, we expect that enriching sparse user preference data in a certain domain by adding user preference data from other domain can significantly improve the quality of cross-domain recommendations (SAHEBI; BRUSILOVSKY, 2013) (FERNÁNDEZ-TOBÍAS et al., 2012). 5.1.5 Statistical Significance Analysis In order to verify the statistical significance of the evaluation results, we adopted the nonparametric Mann–Whitney U test (WINTER; DODOU, 2010), also called Mann–Whitney– Wilcoxon (MWW) or Wilcoxon rank-sum test. This test verifies the null hypothesis, which states that two samples are statistically the same, against an alternative hypothesis, which especially can determine if a particular population tends to have larger values than the other. In addition, unlike the t-test it does not require the assumption of normal distributions (WINTER; DODOU, 2010). In this way, we applied the statistical significance tests with a confidence level of 95% for all user overlap levels, contextual dimensions and target domains. These tests were applied with support from “R” software tool (R Core Team, 2015). For the tests of predictive performance, we verified if the errors of the baseline algorithms were greater than the errors of the proposed ones6. For the tests of classification performance, we verified if the F-metric values of the proposed algorithms were greater than the F-metric values of the baselines ones7, considering the F-metric values for N=5. In both cases, the applied Wilcoxon tests were not paired given that the samples were independent among the algorithms. 5.2 Evaluation Results According to the datasets and evaluation methodology described before, we present and discuss the results of the proposed CD-CARS algorithms in comparison to the baseline cross-domain CF-based algorithms. For that, we divided the experiments into two datasets. Section 5.2.1 shows the evaluation results about two related domains (Book and Television), whereas Section 5.2.2 presents the evaluation results about two less related domains (Book and Music). Finally, we discuss the evaluation results in Section 5.2.3. 6 wilcox.test(baseline,proposed_algorithm, paired=FALSE,alternative = “greater”) by using the “R” software tool. 7 wilcox.test(proposed_algorithm,baseline, paired=FALSE,alternative = “greater”) by using the “R” software tool. 132 Chapter 5. CD-CARS Evaluation 5.2.1 Book-Television Results As mentioned before, we evaluated the quality of the cross-domain algorithms by varying the target domain for each dataset in order to study the impact of the density of the target domain data in comparison to the density of the source domain data. Thus, the following sections present the results considering a different domain as a target. 5.2.1.1 Television as Target Domain According to the contextual dimensions present in the datasets, we describe the evaluation results for each contextual dimension in the following sections. In addition, we show the results for a combination of contextual dimensions in Section 5.2.1.1.4. 5.2.1.1.1 Temporal Dimension Table 19 reports the overall predictive performance of the recommender algorithms, considering all contextual values from the Temporal dimension and different user overlap levels for the Television domain as target. The rows 1 and 2 from the table show the NNUserNgbr predictive performance when it is applied in the single-domain and cross- domain recommendations. As it can be seen, the simple addition of user ratings from other domain (Book), by using the same algorithm for cross-domain recommendation, improved the recommendation performance in, approximately, 8–21% (MAE) and 2–10% (RMSE) depending on the user overlap level. Table 19 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Temporal dimension (source domain: Book, and target domain: Television). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.721 ± 0.024 1.020 ± 0.048 0.454 ± 0.008 0.759 ± 0.020 0.412 ± 0.006 0.734 ± 0.012 NNUserNgbr (cross-domain) 0.598 ± 0.022 0.922 ± 0.044 0.417 ± 0.007 0.742 ± 0.018 0.324 ± 0.005 0.661 ± 0.010 NNUserNgbr- transClosure 0.251 ± 0.014 0.548 ± 0.028 0.256 ± 0.005 0.556 ± 0.012 0.217 ± 0.002 0.531 ± 0.004 PreF with NNUserNgbr- transClosure 0.129 ± 0.020 0.382 ± 0.057 0.132 ± 0.006 0.374 ± 0.016 0.151 ± 0.003 0.413 ± 0.007 PostF with NNUserNgbr- transClosure 0.216 ± 0.012 0.486 ± 0.024 0.210 ± 0.003 0.469 ± 0.005 0.173 ± 0.004 0.445 ± 0.006 In addition, Table 19 presents the overall performance of the PreF and PostF algorithms, besides the base NNUserNgbr-transClosure algorithm, which outperformed 5.2. Evaluation Results 133 the NNUserNgbr algorithm (performed for cross-domain purposes) by achieving an im- provement that varied in, approximately, 33–58% (MAE) and 19–40% (RMSE) depending on the user overlap levels. As it can be seen from table, the PreF predictive performance was better than the NNUserNgbr-transClosure and the PostF algorithms in all user overlap levels. The im- provement achieved by the PreF algorithm in comparison to the NNUserNgbr-transClosure one varied in, approximately, 30–48% (MAE) and 22–32% (RMSE) depending on the user overlap level. The PostF predictive performance was also better than the NNUserNgbr- transClosure algorithm in all user overlap levels, however, its improvement was smaller than the achieved by the PreF algorithm. Figure 30 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. Figure 30 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: book, and target domain: television). The statistical significance tests verified that both the PreF and PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure baseline algorithm for all user overlap levels8. Figure 31a, Figure 31b and Figure 31c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. Regarding the classification performance, Figures 32a, 32b, and 32c present the results of the F-metric at different N values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Television domain as target, considering the Temporal dimension. As it can be seen, in all user overlap levels and top ‘N’ values, 8 p-value=0.005413 and W=100 for all tests, except in the comparison between the baseline and the PostF algorithms when the user overlap level was 10%, where the W=97 and the p-value=0.037 134 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 31 – Overall prediction performance (MAE) boxplots for television domain in the temporal dimension with different user overlap levels (source domain: book). the PostF classification performance was better or similar than the baseline algorithms. The PreF performance was better than the baseline ones for low top ‘N’ values when there were 10% and 50% of user overlap levels, and for any top ‘N’ value when there was 100% of user overlap. In addition, the PostF classification performance was better or similar than the PreF one. Figure 33 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for all user overlap levels9. Besides, the PreF F-metric values were greater than the 9 p-value=0.005413 and W=99 for all tests 5.2. Evaluation Results 135 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 32 – F-metric performance x top ‘N’ items for the television domain in the temporal dimension with different user overlap levels (source domain: book). 136 Chapter 5. CD-CARS Evaluation Figure 33 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal dimension (target domain: television, and source: book). NNUserNgbr-transClosure algorithm for 50% and 100% user overlap levels10. For 10% of user overlap, the NNUserNgbr-transClosure F-metric value was greater than the PreF algorithm11. 5.2.1.1.2 Location Dimension Table 20 reports the overall predictive performance of the recommender algorithms, considering all contextual values from the Location dimension and different user overlap levels for the Television domain as target. As it can be seen from the table, the addition of user ratings from other domain (Book), by using the same algorithm for cross-domain recommendation (corresponding to the two first rows of the table), improved the predictive performance in, approximately, 20–36% (MAE) and 8–20% (RMSE) depending on the user overlap level. Also, Table 20 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 21–58% (MAE) and 11–36% (RMSE) depending on the user overlap levels. As it can be seen from table, the PostF predictive performance was better than the NNUserNgbr-transClosure algorithm in all user overlap levels, with an improvement that varied in, approximately, 5–16% (MAE) and 4–14% (RMSE) depending on the user overlap level. In addition, if we consider the high standard deviation (std) of the PreF 10 p-value=0.005413 and W=99 for all tests 11 p-value=0.005413 and W=100 5.2. Evaluation Results 137 Table 20 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Location dimension (source domain: Book, and target domain: Television). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.721 ± 0.024 1.020 ± 0.048 0.454 ± 0.008 0.759 ± 0.020 0.412 ± 0.006 0.734 ± 0.012 NNUserNgbr (cross-domain) 0.573 ± 0.022 0.865 ± 0.044 0.363 ± 0.007 0.691 ± 0.018 0.261 ± 0.005 0.582 ± 0.010 NNUserNgbr- transClosure 0.240 ± 0.017 0.550 ± 0.039 0.247 ± 0.006 0.545 ± 0.013 0.206 ± 0.003 0.513 ± 0.006 PreF with NNUserNgbr- transClosure 0.305 ± 0.385 0.433 ± 0.541 0.212 ± 0.045 0.588 ± 0.123 0.242 ± 0.027 0.602 ± 0.064 PostF with NNUserNgbr- transClosure 0.200 ± 0.010 0.468 ± 0.028 0.233 ± 0.004 0.519 ± 0.011 0.194 ± 0.005 0.484 ± 0.011 algorithm showed in Table 20, then we can say that the PostF outperformed the PreF algorithm for all user overlap levels. Figure 34 (MAE) and Figure 35 (RMSE) illustrate the predictive performance of the proposed algorithms over different user overlap levels. Note that for this case we showed the figures for both predictive metrics, since we observed a difference in the PreF performance depending on the predictive metric used in the evaluation. Figure 34 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: book, and target domain: television). The statistical significance tests verified that the PostF predictive errors (MAE) 138 Chapter 5. CD-CARS Evaluation Figure 35 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: book, and target domain: television). were less than the NNUserNgbr-transClosure baseline algorithm for all user overlap levels12. On the other hand, the PreF predictive errors (MAE) were statistically similar to the NNUserNgbr-transClosure algorithm for 10% and 50% of user overlap levels13. For 100% of user overlap, the NNUserNgbr-transClosure predictive error was statistically less than the PreF one14. Figure 36a, Figure 36b and Figure 36c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. With respect to the classification performance, Figures 37a, 37b, and 37c present the results of the F-metric at different top ‘N’ values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap levels for the Television domain as target, considering the Location dimension. As it can be seen, in all user overlap levels considering ‘N’ up to five, the PostF classification performance was better than the baseline algorithms, whereas for 50% and 100% of user overlap, the PostF outperformed them for ‘N’ up to ten. On the other hand, the PreF classification performance was worse than all other algorithms in all user overlap levels and ‘N’ values. Figure 38 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for 12 For 10% of user overlap, W=100 and p-value=0.005413. For 50% of user overlap, W=95 and p-value=0.0001028. Finally, for 100% of user overlap, W=98 and p-value=0.02165 13 For 10% of user overlap, W=30 and p-value=0.09391. For 50% of user overlap, W=71 and p- value=0.0615 14 W=90 and p-value=0.0007523 5.2. Evaluation Results 139 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 36 – Overall prediction performance (MAE) boxplots for television domain in the location dimension with different user overlap levels (source domain: book). 50% and 100% of user overlap levels15, whereas their performances were statistically similar for 10% of user overlap16. Besides, the NNUserNgbr-transClosure F-metric values were greater than the PreF algorithm for all user overlap levels17. 15 p-value=0.005413 and W=99 for all tests 16 p-value=0.2 and W=71 17 p-value=0.005413 and W=99 for all tests 140 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 37 – F-metric performance x top ‘N’ items for the television domain in the location dimension with different user overlap levels (source domain: book). 5.2. Evaluation Results 141 Figure 38 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the location dimension (target domain: television, and source: book). 5.2.1.1.3 Companion Dimension Table 21 shows the overall predictive performance of the recommender algorithms, considering all contextual values from the Companion dimension and different user overlap levels for the Television domain as target. As it can be seen from the table, the addition of user ratings from other domain (Book), by using the same algorithm for cross-domain recommendation (corresponding to the two first rows of the table), improved the predictive performance in, approximately, 16–28% (MAE) and 10–14% (RMSE) depending on the user overlap level. Table 21 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Companion dimension (source domain: Book, and target domain: Television). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.721 ± 0.024 1.020 ± 0.048 0.454 ± 0.008 0.759 ± 0.020 0.412 ± 0.006 0.734 ± 0.012 NNUserNgbr (cross-domain) 0.583 ± 0.022 0.881 ± 0.044 0.380 ± 0.007 0.680 ± 0.018 0.295 ± 0.005 0.625 ± 0.010 NNUserNgbr- transClosure 0.249 ± 0.029 0.539 ± 0.048 0.279 ± 0.011 0.587 ± 0.024 0.246 ± 0.003 0.574 ± 0.006 PreF with NNUserNgbr- transClosure 0.931 ± 0.111 1.308 ± 0.158 0.858 ± 0.021 1.204 ± 0.026 0.842 ± 0.010 1.169 ± 0.012 PostF with NNUserNgbr- transClosure 0.232 ± 0.035 0.492 ± 0.065 0.256 ± 0.013 0.540 ± 0.025 0.221 ± 0.006 0.523 ± 0.010 142 Chapter 5. CD-CARS Evaluation Also, Table 21 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 16–57% (MAE) and 8–38% (RMSE) depending on the user overlap levels. Figure 39 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: book, and target domain: television). As it can be seen from table, the PostF predictive performance was better than the NNUserNgbr-transClosure algorithm in all user overlap levels, with an improvement that varied in, approximately, 6–10% (MAE) and 8–9% (RMSE) depending on the user overlap level. In addition, the predictive performance of the PreF algorithm was worse than all other algorithms in the Companion dimension considering all user overlap levels, as showed in Table 21. Figure 39 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. Except when the user overlap level was 10%, the statistical significance tests verified that the PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure algorithm for all other user overlap levels18. On the other hand, the NNUserNgbr- transClosure predictive errors (MAE) were statistically less than the PreF ones for all user overlap levels19. Figure 40a, Figure 40b and Figure 40c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. 18 For 10% of user overlap, W=65 and p-value=0.1399. For 50% of user overlap, W=89 and p- value=0.001045. Finally, for 100% of user overlap, W=100 and p-value=0.005413 19 W=89 and p-value=0.001045 for all tests 5.2. Evaluation Results 143 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 40 – Overall prediction performance (MAE) boxplots for television domain in the companion dimension with different user overlap levels (source domain: book). Figures 41a, 41b, and 41c present the results of the F-metric at different top ‘N’ values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Television domain as target, considering the Companion dimension. As it can be seen, the proposed algorithms only outperformed the baseline for 50% and 100% of user overlap with low values of top ‘N’. Figure 42 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the NNUserNgbr-transClosure F-metric values were greater than the PostF ones for 10% and 144 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 41 – F-metric performance x top ‘N’ items for the television domain in the com- panion dimension with different user overlap levels (source domain: book). 5.2. Evaluation Results 145 Figure 42 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the companion dimension (target domain: television, and source: book). 50% of user overlap levels20, whereas their performances were statistically similar for 100% of user overlap21. Besides, the NNUserNgbr-transClosure F-metric values were greater than the PreF algorithm for all user overlap levels22. 5.2.1.1.4 Combining Contextual Dimensions In the previous sections, we presented the evaluation results regarding the contextual dimensions separately. In this section, we report the results for a combination of two contextual dimensions considering the same evaluation metrics and methodology described before. As mentioned in Section 4.1.2, an important aspect of context-aware recommender systems is to determine the relevance of contextual dimensions, attributes (or even values) in order to select only the contextual features that actually matter for evaluation (or recommendation) purposes. We have seen in Section 4.1.2 that two of the contextual dimensions (Temporal and Location) provide a greater information gain than the Companion dimension, which is confirmed by the results presented in the previous sections. In this way, we evaluated the combination of those two contextual dimensions (Temporal and Location), aiming to verify its performance in comparison to their own performances evaluated separately. Table 22 reports the overall predictive performance of the recommender algorithms, considering all contextual value combinations from the Temporal and Location dimensions with different user overlap levels for the Television domain as target. As it was observed 20 p-value=0.005413 and W=99 for all tests 21 p-value=0.75 and W=60 22 p-value=0.003913 and W=100 for all tests 146 Chapter 5. CD-CARS Evaluation in the previous sections, the addition of user ratings from the Book domain also improved the predictive performance of the NNUserNgbr in, approximately, 20–37% (MAE) and 9–20% (RMSE) depending on the user overlap level (rows 1 and 2 from the table). Table 22 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual value combinations from the temporal and location dimensions (source domain: Book, and target domain: Television). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.721 ± 0.024 1.020 ± 0.048 0.454 ± 0.008 0.759 ± 0.020 0.412 ± 0.006 0.734 ± 0.012 NNUserNgbr (cross-domain) 0.571 ± 0.022 0.860 ± 0.044 0.360 ± 0.007 0.689 ± 0.018 0.259 ± 0.005 0.580 ± 0.010 NNUserNgbr- transClosure 0.224 ± 0.017 0.482 ± 0.039 0.250 ± 0.006 0.552 ± 0.013 0.207 ± 0.003 0.515 ± 0.006 PreF with NNUserNgbr- transClosure 0.396 ± 0.390 0.761 ± 0.560 0.720 ± 0.045 1.050 ± 0.123 0.333 ± 0.027 0.739 ± 0.064 PostF with NNUserNgbr- transClosure 0.226 ± 0.010 0.503 ± 0.028 0.190 ± 0.004 0.433 ± 0.011 0.161 ± 0.005 0.437 ± 0.011 Also, Table 22 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 20–60% (MAE) and 11–43% (RMSE) depending on the user overlap levels. As it can be seen from table, except when the user overlap level was 10%, the PostF predictive performance was better than the NNUserNgbr-transClosure algorithm in all other user overlap levels, with an improvement that varied in, approximately, 22–24% (MAE) and 15–21% (RMSE) depending on the user overlap level. Despite the NNUserNgbr- transClosure algorithm have outperformed the PostF when the user overlap level was 10%, they had a similar performance, separated only by their standard deviations. Figure 43 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. The statistical significance tests verified that the PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure algorithm for the 50% and 100% user overlap levels23. When the user overlap level was 10%, the applied tests could not determine any statistical difference between the NNUserNgbr-transClosure and PostF predictive 23 In both cases, with W=100 and p-value=0.005413 5.2. Evaluation Results 147 Figure 43 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: book, and target domain: television). errors (MAE)24. On the other hand, the applied tests verified that the NNUserNgbr- transClosure predictive errors (MAE) were less than the PreF ones for all user overlap levels25. Figure 44a, Figure 44b and Figure 44c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. Taking into account the classification performance, Figures 45a, 45b, and 45c present the results of the F-metric at different top ‘N’ values, respectively, in 10%, 50%, and 100% of user overlap level for the Television domain as target, considering the combination between the Temporal and Location dimensions. For all user overlap levels and top ‘N’ values, the PostF classification performance was better or similar than the baseline algorithms, while the PreF classification performance was worse than all other algorithms. Figure 46 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for all user overlap levels26. On the other hand, the applied tests also verified that NNUserNgbr-transClosure F-metric values were greater than the PreF ones for all user overlap levels27. 24 W=41 and p-value=0.5 25 For all cases, with W=100 and p-value=0.003914 26 p-value=0.005413 and W=97 for all tests 27 p-value=0.003968 and W=100 for all tests 148 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 44 – Overall prediction performance (MAE) boxplots for television domain in the temporal and location dimensions with different user overlap levels (source domain: book). 5.2.1.2 Book as Target Domain In the Section 5.2.1.1, we presented the results for the Television target domain, which had fewer ratings in the cross-domain dataset in comparison to Book source domain (as described in Section 4.1.3.1). In this section, we present the results when Book is the target domain and Television is the source domain. According to the contextual dimensions present in Section 4.1.3.1, we describe the evaluation results for each contextual dimension in the following sections. In addition, we show the results for a combination of contextual dimensions in Section 5.2.1.2.4. 5.2.1.2.1 Temporal Dimension Table 23 reports the overall predictive performance of the recommender algorithms, considering all contextual values from the Temporal dimension and different user overlap 5.2. Evaluation Results 149 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 45 – F-metric performance x top ‘N’ items for the television domain in the temporal and location dimensions with different user overlap levels (source domain: book). 150 Chapter 5. CD-CARS Evaluation Figure 46 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal and location dimensions (target domain: television, and source: book). Table 23 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Temporal dimension (source domain: Television, and target domain: Book). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.643 ± 0.022 0.963 ± 0.040 0.500 ± 0.008 0.794 ± 0.020 0.401 ± 0.006 0.719 ± 0.010 NNUserNgbr (cross-domain) 0.497 ± 0.020 0.836 ± 0.042 0.367 ± 0.007 0.674 ± 0.018 0.297 ± 0.005 0.620 ± 0.008 NNUserNgbr- transClosure 0.133 ± 0.007 0.361 ± 0.021 0.180 ± 0.002 0.437 ± 0.006 0.181 ± 0.002 0.459 ± 0.005 PreF with NNUserNgbr- transClosure 0.122 ± 0.009 0.340 ± 0.009 0.116 ± 0.002 0.340 ± 0.006 0.114 ± 0.001 0.330 ± 0.005 PostF with NNUserNgbr- transClosure 0.120 ± 0.008 0.319 ± 0.019 0.153 ± 0.003 0.375 ± 0.009 0.155 ± 0.003 0.399 ± 0.007 levels for the Book domain as target. The rows 1 and 2 from the table show the NNUser- Ngbr predictive performance when it is applied in the single-domain and cross-domain recommendations. As it can be seen, the simple addition of user ratings from other domain (Television), by using the same algorithm for cross-domain recommendation, improved the recommendation performance in, approximately, 22–26% (MAE) and 13–15% (RMSE) depending on the user overlap level. In addition, Table 23 presents the overall performance of the PreF and PostF algorithms, besides the base NNUserNgbr-transClosure algorithm, which outperformed the NNUserNgbr algorithm (performed for cross-domain purposes) by achieving an im- 5.2. Evaluation Results 151 provement that varied in, approximately, 38–73% (MAE) and 25–56% (RMSE) depending on the user overlap levels. As it can be seen from the table, the PreF predictive performance was better than the NNUserNgbr-transClosure for all user overlap levels, and better than the PostF algorithm for 50% and 100% of user overlap levels. The improvement achieved by the PreF algorithm in comparison to the NNUserNgbr-transClosure one varied in, approximately, 8–37% (MAE) and 5–28% (RMSE) depending on the user overlap level. The PostF predictive performance was better than the PreF algorithm for 10% of user overlap, and better than the NNUserNgbr-transClosure for all user overlap levels. Figure 47 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. Figure 47 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: television, and target domain: book). The statistical significance tests verified that both the PreF and PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure baseline algorithm for all user overlap levels28. Figure 48a, Figure 48b and Figure 48c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. Regarding the classification performance, Figures 49a, 49b, and 49c present the results of the F-metric at different N values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Book domain as target, considering the Temporal dimension. As it can be seen, in all user overlap levels and top ‘N’ values, the 28 p-value=0.005413 and W=100 for all tests, except when there was 10% of user overlap, where W=79 and p-value=0.0144 (PreF x NNUserNgbr-transClosure), and W=87 and p-value=0.001943 (PostF x NNUserNgbr-transClosure) 152 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 48 – Overall prediction performance (MAE) boxplots for book domain in the tem- poral dimension with different user overlap levels (source domain: television). PostF classification performance was better or similar than the NNUserNgbr-transClosure baseline algorithm, whereas the PreF was only better than that baseline for low top ‘N’ values with 50% and 100% of user overlap levels. In addition, the PostF classification performance was better than the PreF one for all user overlap levels and top ‘N’ values. Figure 50 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for 5.2. Evaluation Results 153 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 49 – F-metric performance x top ‘N’ items for the book domain in the temporal dimension with different user overlap levels (source domain: television). 154 Chapter 5. CD-CARS Evaluation Figure 50 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal dimension (target domain: book, and source: television). 50% and 100% of user overlap levels29, whereas their performances were statistically similar for 10% of user overlap30. The PreF F-metric value was greater than the NNUserNgbr- transClosure algorithm for 100% of user overlap31, whereas the opposite occurred when the user overlap levels were 10% and 50%32. 5.2.1.2.2 Location Dimension Table 24 reports the overall predictive performance of the recommender algorithms, considering all contextual values from the Location dimension and different user overlap levels for the Book domain as target. As it can be seen from the table, the addition of user ratings from other domain (Television), by using the same algorithm for cross-domain recommendation (corresponding to the two first rows of the table), improved the predictive performance in, approximately, 25–30% (MAE) and 13–18% (RMSE) depending on the user overlap level. Also, Table 24 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 38–73% (MAE) and 25–60% (RMSE) depending on the user overlap levels. As it can be seen from table, the PostF predictive performance was better or similar than the NNUserNgbr-transClosure algorithm in all user overlap levels, with an 29 p-value=0.003968 and W=100 for all tests 30 p-value=0.35 and W=60 31 W=77 and p-value=0.02163 32 p-value=0.003968 and W=100 for both tests 5.2. Evaluation Results 155 Table 24 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Location dimension (source domain: Television, and target domain: Book). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.643 ± 0.022 0.963 ± 0.040 0.500 ± 0.008 0.794 ± 0.020 0.401 ± 0.006 0.719 ± 0.010 NNUserNgbr (cross-domain) 0.470 ± 0.020 0.830 ± 0.042 0.349 ± 0.007 0.645 ± 0.018 0.297 ± 0.005 0.617 ± 0.008 NNUserNgbr- transClosure 0.125 ± 0.015 0.329 ± 0.034 0.177 ± 0.006 0.427 ± 0.013 0.182 ± 0.003 0.460 ± 0.006 PreF with NNUserNgbr- transClosure 0.184 ± 0.285 0.447 ± 0.503 0.121 ± 0.045 0.314 ± 0.087 0.179 ± 0.019 0.432 ± 0.040 PostF with NNUserNgbr- transClosure 0.127 ± 0.004 0.330 ± 0.012 0.173 ± 0.004 0.410 ± 0.011 0.175 ± 0.003 0.438 ± 0.009 improvement that varied in, approximately, 2–4% (MAE) and 3–4% (RMSE) depending on the user overlap level. In addition, we can see in Table 24 that the PostF outperformed the PreF algorithm for the majority of the user overlap levels (10% and 100%). As the results presented in Section 5.2.1.1.2 (source domain: Book, target domain: Television, and Location dimension), the PreF predictive performance had a high standard deviation. As mentioned in that section, this issue may be caused by the PreF feature of filtering ratings from the target domain for untested contexts, especially in the Location dimension, where there are several cities with a low number of ratings. Figure 51 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. The statistical significance tests verified that both the PreF and PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure baseline algorithm for 50% of user overlap level33. For 10% and 100% of user overlap, there was not a significant difference between the performance of the PostF algorithm and the NNUserNgbr-transClosure baseline algorithm34. The same occurred between the baseline and PreF algorithms for 100% of user overlap35, whereas the NNUserNgbr-transClosure predictive errors was less than the PreF one36 for 10% of user overlap. Figure 52a, Figure 52b and Figure 52c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. With respect to the classification performance, Figures 53a, 53b, and 53c present 33 W=95 and p-value=0.0001028 (PreF x NNUserNgbr-transClosure), while W=77 and p-value=0.02163 (PostF and NNUserNgbr-transClosure) 34 W=44 and p-value=0.6847 (10% of user overlap), while W=96 and p-value=0.06495 (100% of user 156 Chapter 5. CD-CARS Evaluation Figure 51 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: television, and target domain: book). the results of the F-metric at different top ‘N’ values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Book domain as target, considering the Location dimension. As it can be seen, for 10% and 50% of overlap levels and low top ‘N’ values, the PostF classification performance was better or similar than the NNUserNgbr- transClosure baseline algorithm, whereas for 100% of user overlap this can be seen for any top ‘N’ value. On the other hand, the PreF classification performance was worse than all other algorithms in all user overlap levels and ‘N’ values. Figure 54 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for 50% and 100% of user overlap levels37, whereas the opposite from this was observed for 10% of user overlap level38. The applied tests also verified that the NNUserNgbr-transClosure F-metric values were greater than the PreF ones for all user overlap levels39. 5.2.1.2.3 Companion Dimension Table 25 shows the overall predictive performance of the recommender algorithms, considering all contextual values from the Companion dimension and different user overlap levels for the Book domain as target. As it can be seen from the table, the addition of user overlap) 35 W=96 and p-value=0.06495 36 W=77 and p-value=0.02163 37 p-value=0.005814 and W=97 for all tests 38 p-value=0.005814 and W=97 39 p-value=0.003913 and W=100 for all tests 5.2. Evaluation Results 157 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 52 – Overall prediction performance (MAE) boxplots for book domain in the loca- tion dimension with different user overlap levels (source domain: television). ratings from other domain (Television), by using the same algorithm for cross-domain recommendation (corresponding to the two first rows of the table), improved the predictive performance in, approximately, 12% (MAE) and 11% (RMSE) for 10% of user overlap, and in, approximately, 11% (MAE) and 7% (RMSE) for 50% of user overlap. Also, Table 25 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 28–59% (MAE) and 17–38% (RMSE) depending on the user overlap level. 158 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 53 – F-metric performance x top ‘N’ items for the book domain in the location dimension with different user overlap levels (source domain: television). 5.2. Evaluation Results 159 Figure 54 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the location dimension (target domain: book, and source: television). Table 25 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Companion dimension (source domain: television, and target domain: book). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.643 ± 0.022 0.963 ± 0.040 0.500 ± 0.008 0.794 ± 0.020 0.401 ± 0.006 0.719 ± 0.010 NNUserNgbr (cross-domain) 0.565 ± 0.020 0.851 ± 0.042 0.442 ± 0.007 0.739 ± 0.018 0.409 ± 0.005 0.747 ± 0.008 NNUserNgbr- transClosure 0.229 ± 0.039 0.519 ± 0.087 0.265 ± 0.014 0.563 ± 0.031 0.293 ± 0.007 0.616 ± 0.014 PreF with NNUserNgbr- transClosure 0.611 ± 0.289 0.872 ± 0.346 0.757 ± 0.051 1.069 ± 0.059 0.789 ± 0.016 1.104 ± 0.022 PostF with NNUserNgbr- transClosure 0.241 ± 0.043 0.520 ± 0.079 0.264 ± 0.014 0.544 ± 0.039 0.277 ± 0.007 0.575 ± 0.010 As it can be seen from table, the PostF predictive performance was better than the NNUserNgbr-transClosure algorithm in, approximately, 0.4% (MAE) and 3% (RMSE) for 50% of user overlap, and in, approximately, 5% (MAE) and 6% (RMSE) for 100% of user overlap, however, the predictive performance of these algorithms were similar when the user overlap level was 10%. In addition, the predictive performance of the PreF algorithm was worse than all other algorithms in the Companion dimension, as showed in Table 25. Figure 55 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. 160 Chapter 5. CD-CARS Evaluation Figure 55 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: television, and target domain: book). The statistical significance tests verified that the PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure algorithm for 100% of user overlap40. For 10% and 50% of user overlap levels, the applied tests could not verify a statistical difference between the NNUserNgbr-transClosure and PostF predictive errors (MAE)41. The applied tests also verified that the NNUserNgbr-transClosure predictive errors (MAE) were less than the PreF ones for all user overlap levels42. Figure 56a, Figure 56b and Figure 56c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. Figures 57a, 57b, and 57c present the results of the F-metric at different top ‘N’ values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Book domain as target, considering the Companion dimension. As it can be seen, the proposed algorithms only outperformed the baseline for 100% of user overlap with low values of top ‘N’. Figure 58 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the NNUserNgbr-transClosure F-metric values were greater than the both proposed algorithms for all user overlap levels43. 40 W=99 and p-value=0.01083 41 For 10% of user overlap, W=60 and p-value=0.2406, while for 50%, W=55 and p-value=0.3697 42 W=99 and p-value=0.01083 for all tests 43 p-value=0.005814 and W=97 for all tests between the PostF and baseline algorithms, while p- value=0.003913 and W=100 for all tests between the PreF and baseline algorithms 5.2. Evaluation Results 161 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 56 – Overall prediction performance (MAE) boxplots for book domain in the com- panion dimension with different user overlap levels (source domain: television). 5.2.1.2.4 Combining Contextual Dimensions In the previous sections, we presented the evaluation results regarding the contextual dimensions separately. In this section, we report the results for a combination of two contextual dimensions considering the same evaluation metrics and methodology described before. We have seen in Section 4.1.2 that two of the contextual dimensions (Temporal and Location) provide a greater information gain than the Companion dimension, which is confirmed by the results presented in the previous sections. In this way, we evaluated 162 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 57 – F-metric performance x top ‘N’ items for the book domain in the companion dimension with different user overlap levels (source domain: television). 5.2. Evaluation Results 163 Figure 58 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the companion dimension (target domain: book, and source: television). the combination of those two contextual dimensions (Temporal and Location), aiming to verify its performance in comparison to the performance of them evaluated separately. Table 26 reports the overall predictive performance of the recommender algorithms, considering all contextual value combinations from the Temporal and Location dimensions with different user overlap levels for the Book domain as target. As it was observed in the previous sections, the addition of user ratings from the Television domain also improved the predictive performance of the NNUserNgbr in, approximately, 25–30% (MAE) and 13–18% (RMSE) depending on the user overlap level (rows 1 and 2 from the table). Also, Table 26 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 39–75% (MAE) and 25–60% (RMSE) depending on the user overlap levels. As it can be seen from table, except when the user overlap level was 10% (by considering only the MAE metric), the PostF predictive performance was better than the NNUserNgbr-transClosure algorithm in all other user overlap levels, with an improvement that varied in, approximately, 10–20% (MAE) and 8–17% (RMSE) depending on the user overlap level. Despite the NNUserNgbr-transClosure algorithm have outperformed the PostF when the user overlap level was 10%, they had a similar performance, separated only by their standard deviations. Figure 59 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. The statistical significance tests verified that the PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure algorithm for the 50% and 100% user overlap 164 Chapter 5. CD-CARS Evaluation Table 26 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual value combinations from the temporal and location dimensions (source domain: Television, and target domain: Book). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.643 ± 0.022 0.963 ± 0.040 0.500 ± 0.008 0.794 ± 0.020 0.401 ± 0.006 0.719 ± 0.010 NNUserNgbr (cross-domain) 0.470 ± 0.020 0.830 ± 0.042 0.349 ± 0.007 0.645 ± 0.018 0.297 ± 0.005 0.617 ± 0.008 NNUserNgbr- transClosure 0.117 ± 0.017 0.328 ± 0.039 0.176 ± 0.006 0.418 ± 0.013 0.180 ± 0.003 0.457 ± 0.006 PreF with NNUserNgbr- transClosure 0.344 ± 0.396 0.713 ± 0.545 0.200 ± 0.045 0.423 ± 0.126 0.331 ± 0.027 0.699 ± 0.064 PostF with NNUserNgbr- transClosure 0.121 ± 0.010 0.309 ± 0.028 0.158 ± 0.004 0.382 ± 0.011 0.143 ± 0.005 0.378 ± 0.011 Figure 59 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: televi- sion, and target domain: book). levels44. When the user overlap level was 10%, the NNUserNgbr-transClosure predictive error (MAE) was statistically similar to the PostF one45. On the other hand, the applied tests verified that the NNUserNgbr-transClosure predictive errors (MAE) were less than the PreF ones for all user overlap levels46. Figure 99a, Figure 99b and Figure 99c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. 44 In both cases, with W=99 and p-value=0.005413 45 W=41 and p-value=0.5 46 For all cases, with W=100 and p-value=0.003914 5.2. Evaluation Results 165 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 60 – Overall prediction performance (MAE) boxplots for book domain in the temporal and location dimensions with different user overlap levels (source domain: television). Taking into account the classification performance, Figures 61a, 61b, and 61c present the results of the F-metric at different top ‘N’ values, respectively, in 10%, 50%, and 100% of user overlap level for the Television domain as target, considering the Temporal and Location dimensions. For all user overlap levels and top ‘N’ values, the PreF classification performance was worse than the all other algorithms. This result also occurred when the evaluation was performed for the Television as target by combining the two contextual dimensions (see Section 5.2.1.1.4). On the other hand, the PostF classification performance was better or similar than the NNUserNgbr-transClosure baseline algorithm in all user overlap levels and top ‘N’ values. Figure 62 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for 166 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 61 – F-metric performance x top ‘N’ items for the book domain in the temporal and location dimensions with different user overlap levels (source domain: television). 5.2. Evaluation Results 167 Figure 62 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal and location dimensions (target domain: book, and source: television). 50% and 100% of user overlap levels47, whereas their performances were statistically similar for 10% of user overlap48. The applied tests also verified that NNUserNgbr-transClosure F-metric values were greater than the PreF ones for all user overlap levels49. 5.2.1.3 Summary In this section, we provide a summary of the results from the evaluation of the “book-television dataset”. Figure 63 shows a dispersion diagram illustrating the predictive performance (MAE) for the algorithms by varying target domain (Book and Television), contextual dimension and user overlap levels, whereas Figure 64 shows the same, but considering the RMSE metric. It is important to mention that these figures do not take into account the standard deviation and the statistical significance of the results. Table 27 presents the predictive performance (MAE) achieved by the PreF and PostF algorithms in comparison to the best baseline algorithm (NNUserNgbr-transClosure), by taking into account their statistical significance50 and different target domain, contextual dimension and user overlap levels. Regarding the classification performance, Figure 65 presents a dispersion diagram illustrating the F-metric performance (with N=5) for the algorithms by varying target domain, contextual dimension and user overlap levels. Once again, it is important to mention that we are not considering the standard deviation and the statistical significance of the results in that figure. 47 p-value=0.005814 and W=97 for all tests 48 p-value=0.5 and W=51 49 p-value=0.003968 and W=100 for all tests 50 In the table, “**” means that the result could not be considered statistically significant. 168 Chapter 5. CD-CARS Evaluation Table 28 shows the classification performance improvement (F-metric with N=5) obtained by the PreF and PostF algorithms in comparison to the best baseline algorithm (NNUserNgbr-transClosure), by taking into account their statistical significance51 and different target domain, contextual dimension and user overlap levels. As it can be seen, at least one proposed algorithm (PreF or PostF) achieved the best predictive performance among the algorithms (or it was similar to the best one) in all scenarios (with distinct target domains, contextual dimensions, and user overlap levels). By considering the classification metric, the PostF algorithm achieved the best performance among the algorithms (or it was similar to the best one) in the majority of the scenarios. By summarizing the main findings from the evaluation results described in this section, we can say that: • In all scenarios (with different target domains, contextual dimensions and user overlap levels), the addition of user ratings from an auxiliary (source) domain improved the predictive performance of the NNUserNgbr algorithm, which was not designed for making cross-domain recommendations. This fact can be also observed in almost all scenarios regarding the classification performance of that algorithm. Note that this occurred even when a source domain had less ratings than the target domain. • In all scenarios (with different target domains, contextual dimensions and user overlap levels), the NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one by considering their predictive performances. Regarding their classification performances this fact also occurred in almost all scenarios. • The proposed algorithms (PreF and PostF) had better predictive and classification performances in the Temporal dimension than others dimensions. This fact contrasts to the information gain calculated in Section 4.1.2. In this contextual dimension, the PostF outperformed the NNUserNgbr-transClosure algorithm in all scenarios (user overlap levels and target domains) by considering either their predictive or classification performances. The PreF outperformed the NNUserNgbr-transClosure algorithm in all scenarios (user overlap levels and target domains) by considering the predictive performance. With respect to the classification performance, the PreF outperformed the NNUserNgbr-transClosure algorithm for 100% of user overlap (regardless the target domain) and for 50% of user overlap when the Television was the target domain. • If we make a comparison between the proposed algorithms in the Temporal dimension considering different evaluation metrics (predictive or classification), we see distinct relations between the algorithms’ results. While the PostF algorithm outperforms 51 In the table, “**” means that the result could not be considered statistically significant. 5.2. Evaluation Results 169 the PreF one by considering the classification performance in almost all scenarios (user overlap levels and target domains), the opposite happens when we take the predictive performance into account. • The more is the user overlap level the better is the classification performance of the PostF algorithm, especially in the Temporal and Location dimensions (or their combinations). The same can be observed for the PreF algorithm, but only considering the Temporal dimension. Note that this fact did not seem to occur when we considered the predictive performance of the algorithms. • The predictive and classification performances of the PreF algorithm were more affected than the PostF ones by considering the quantity of the contextual information present in the user ratings (see Section 4.1.3). The more particular were the tested contexts the worse were the PreF performances (e.g. in the Location dimension and in the combination of Location and Temporal dimensions). In this way, the PreF performances had a high variation, depending on the contextual information present in the user ratings, whereas the PostF performances were more uniform and similar to the NNUserNgbr-transClosure algorithm. • Regarding the low quality of the contextual information obtained in the Companion dimension (see Section 4.1.1.3), we can see that both proposed algorithms had low predictive and classification performances in comparison to other dimensions. This could also have occurred due to the low quantity of contextual information present in the user ratings for that contextual dimension. In particular, the PostF algorithm achieved a good performance by considering only the predictive metrics. • With respect to the combination of contextual dimensions (Temporal and Location), we can see the PostF predictive and classification performances in that combination were close to their own performances using only the Temporal dimension as single source of contextual information, whereas the PreF predictive and classification performances were similar to their own performances using only the Location di- mension. The classification performances of both algorithms were reduced with the addition of contextual information from other contextual dimension. In particular, the predictive performance of the PostF algorithm was slightly improved depending on the user overlap level and target domain, whereas the predictive performance of the PreF algorithm was decreased in any case. 170 Chapter 5. CD-CARS Evaluation Figure 63 – Predictive performance (MAE) for the algorithms by varying target domain (book and TV), contextual dimension and user overlap levels (dispersion diagram). 5.2. Evaluation Results 171 Figure 64 – Predictive performance (RMSE) for the algorithms by varying target domain (book and TV), contextual dimension and user overlap levels (dispersion diagram). 172 Chapter 5. CD-CARS Evaluation Figure 65 – Classification performance (F-metric with N=5) for the algorithms by varying target domain (book and TV), contextual dimension and user overlap levels (dispersion diagram). 5.2. Evaluation Results 173 Table 27 – Overall predictive performance (MAE) of the proposed algorithms in comparison to the best baseline one by varying target domain (book and TV), contextual dimension and user overlap levels. Contextual dimension Target Domain User Overlap Level PreF Improve- ment PostF Improve- ment Temporal TV 10% 48.6% 13.9% Book 10% 8% 9.7% TV 50% 48.4% 18% Book 50% 35.7% 15% TV 100% 30.4% 20.3% Book 100% 37.4% 14.6% Location TV 10% -26.8%** 16.7%** Book 10% -46.7% -1.8%** TV 50% 14.2%** 5.7% Book 50% 31.9% 2.5% TV 100% -17.2% 5.9% Book 100% 1.7%** 4%** Companion TV 10% -273.7% 6.7%** Book 10% -166.8% -5.2%** TV 50% -207.9% 8.2% Book 50% -185.2% 0.4%** TV 100% -242.4% 10.2% Book 100% -169.4% 5.6% Temporal and Location TV 10% -76.8% -0.9%** Book 10% -194% -3.4%** TV 50% -188% 24% Book 50% -13.6% 10.2% TV 100% -60.9% 22.2% Book 100% -83.9% 20.6% 5.2.2 Book-Music Results As mentioned before, we evaluated the quality of the cross-domain algorithms by varying the target domain for each dataset in order to study the impact of the density of the target domain data in comparison to the density of the source domain data. Thus, the following sections present the results considering a different domain as a target (Music and Book). 5.2.2.1 Music as Target Domain According to the contextual dimensions present in the datasets, we describe the evaluation results for each contextual dimension in the following sections. In addition, we show the results for a combination of contextual dimensions in Section 5.2.2.1.4. 174 Chapter 5. CD-CARS Evaluation Table 28 – Overall classification performance (F-metric with N=5) of the proposed algo- rithms in comparison to the best baseline one by varying target domain (book and TV), contextual dimension and user overlap levels. Contextual dimension Target Domain User Overlap Level PreF Improve- ment PostF Improve- ment Temporal TV 10% -38% 22.4% Book 10% -113.4% 4.7%** TV 50% 35.4% 38% Book 50% -27.2% 16.7% TV 100% 45% 41.2% Book 100% 7.3% 26.7% Location TV 10% -435.2% 1.9%** Book 10% -329.4% -8.1% TV 50% -491.7% 30.4% Book 50% -496.7% 3.8% TV 100% -414% 32.7% Book 100% -467.1% 18.7% Companion TV 10% -148.4% -39.1% Book 10% -142.4% -26.8% TV 50% -39% -2.8% Book 50% -112.2% -36.9% TV 100% -6.2% -5.8%** Book 100% -60.3% -7.5% Temporal and Location TV 10% -457.2% 20% Book 10% -404.2% 0.05%** TV 50% -532.5% 35.7% Book 50% -482.4% 18.1% TV 100% -488.4% 41.6% Book 100% -456% 28.4% 5.2.2.1.1 Temporal Dimension Table 29 reports the overall predictive performance of the recommender algorithms, considering all contextual values from the Temporal dimension and different user overlap levels for the Music domain as target. The rows 1 and 2 from the table show the NNUserNgbr predictive performance when it is applied in the single-domain and cross- domain recommendations. As it can be seen, the simple addition of user ratings from other domain (Book), by using the same algorithm for cross-domain recommendation, improved the recommendation performance in, approximately, 5–7% (MAE) and 1–2% (RMSE) depending on the user overlap level. In addition, Table 29 presents the overall performance of the PreF and PostF algorithms, besides the base NNUserNgbr-transClosure algorithm, which outperformed the NNUserNgbr algorithm (performed for cross-domain purposes) by achieving an im- 5.2. Evaluation Results 175 Table 29 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Temporal dimension (source domain: Book, and target domain: Music). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.684 ± 0.026 0.972 ± 0.053 0.654 ± 0.016 0.970 ± 0.030 0.588 ± 0.010 0.903 ± 0.017 NNUserNgbr (cross-domain) 0.634 ± 0.022 0.949 ± 0.051 0.610 ± 0.013 0.943 ± 0.028 0.557 ± 0.008 0.891 ± 0.014 NNUserNgbr- transClosure 0.307 ± 0.010 0.631 ± 0.013 0.458 ± 0.013 0.793 ± 0.020 0.480 ± 0.006 0.816 ± 0.011 PreF with NNUserNgbr- transClosure 0.185 ± 0.054 0.616 ± 0.166 0.171 ± 0.005 0.476 ± 0.007 0.212 ± 0.007 0.550 ± 0.016 PostF with NNUserNgbr- transClosure 0.257 ± 0.031 0.519 ± 0.038 0.400 ± 0.004 0.707 ± 0.002 0.423 ± 0.005 0.725 ± 0.010 provement that varied in, approximately, 13–51% (MAE) and 8–33% (RMSE) depending on the user overlap levels. As it can be seen from table, the PreF predictive performance was better than the NNUserNgbr-transClosure and PostF algorithms in all user overlap levels, except when the user overlap level was 10% if we only consider the RMSE metric instead of the MAE one. The improvement achieved by the PreF algorithm in comparison to the NNUserNgbr- transClosure one varied in, approximately, 39–55% (MAE) and 2–40% (RMSE) depending on the user overlap level. The PostF predictive performance was also better than the NNUserNgbr-transClosure algorithm in all user overlap levels, however, its improvement was smaller than the achieved by the PreF algorithm. Figure 66 (MAE) and Figure 67 (RMSE) illustrate the predictive performance of the proposed algorithms over different user overlap levels. Note that for this case we showed the figures for both predictive metrics, since we observed a difference in the PreF performance depending on the predictive metric used in the evaluation. The statistical significance tests verified that both the PreF and PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure baseline algorithm for all user overlap levels52. Figure 68a, Figure 68b and Figure 68c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. Regarding the classification performance, Figures 69a, 69b, and 69c present the results of the F-metric at different N values (between one and twenty), respectively, in 52 p-value=0.005413 and W=97 for all tests. 176 Chapter 5. CD-CARS Evaluation Figure 66 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: book, and target domain: Music). Figure 67 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: book, and target domain: Music). 10%, 50%, and 100% of user overlap level for the Music domain as target, considering the Temporal dimension. As it can be seen, in all user overlap levels and top ‘N’ values, the PostF classification performance was better or similar than the baseline algorithms, whereas the PreF was only better than the baseline ones for 50% and 100% of user overlap levels with low top ‘N’ values. In addition, the PostF classification performance was better or similar than the PreF one for all user overlap levels and top ‘N’ values. 5.2. Evaluation Results 177 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 68 – Overall prediction performance (MAE) boxplots for Music domain in the temporal dimension with different user overlap levels (source domain: book). Figure 70 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for 50% and 100% user overlap levels53, whereas for 10% of user overlap the tests could not verify any statistical difference between their performances 54. Besides, the PreF F-metric value was greater than the NNUserNgbr-transClosure one for 100% of user overlap55, 53 p-value=0.005413 and W=99 in both cases 54 W=80 and p-value=0.1 55 p-value=0.005413 and W=99 178 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 69 – F-metric performance x top ‘N’ items for the Music domain in the temporal dimension with different user overlap levels (source domain: book). 5.2. Evaluation Results 179 Figure 70 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal dimension (target domain: Music, and source: book). whereas the opposite from this was observed for 10% and 50% of user overlap levels56. 5.2.2.1.2 Location Dimension Table 30 reports the overall predictive performance of the recommender algorithms, considering all contextual values from the Location dimension and different user overlap levels for the Music domain as target. As it can be seen from the table, the addition of user ratings from other domain (Book), by using the same algorithm for cross-domain recommendation (corresponding to the two first rows of the table), improved the predictive performance in, approximately, 6–7% (MAE) and 3–4% (RMSE) depending on the user overlap level (50% or 100%). Note that the NNUserNgbr predictive performance was better than its own performance, considering the cross-domain scenario, when the user overlap level was 10%. Also, Table 30 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 11–54% (MAE) and 6–32% (RMSE) depending on the user overlap levels. As it can be seen from table, the PostF predictive performance was better than the NNUserNgbr-transClosure algorithm in all user overlap levels, with an improvement that varied in, approximately, 3–16% (MAE) and 4–19% (RMSE) depending on the user overlap level. In addition, if we consider the high standard deviation of the PreF algorithm showed in Table 30, then we can say that the PostF outperformed the PreF algorithm for 56 p-value=0.006812 and W=97 for both tests 180 Chapter 5. CD-CARS Evaluation Table 30 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Location dimension (source domain: Book, and target domain: Music). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.684 ± 0.026 0.972 ± 0.053 0.654 ± 0.016 0.970 ± 0.030 0.588 ± 0.010 0.903 ± 0.017 NNUserNgbr (cross-domain) 0.710 ± 0.022 1.082 ± 0.051 0.602 ± 0.013 0.932 ± 0.028 0.552 ± 0.008 0.877 ± 0.014 NNUserNgbr- transClosure 0.326 ± 0.065 0.731 ± 0.084 0.468 ± 0.008 0.808 ± 0.013 0.487 ± 0.004 0.822 ± 0.008 PreF with NNUserNgbr- transClosure 0.246 ± 0.238 0.528 ± 0.321 0.284 ± 0.022 0.514 ± 0.020 0.208 ± 0.033 0.541 ± 0.130 PostF with NNUserNgbr- transClosure 0.273 ± 0.011 0.591 ± 0.027 0.437 ± 0.007 0.775 ± 0.003 0.473 ± 0.005 0.791 ± 0.014 Figure 71 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: book, and target domain: Music). 10% of user overlap level. For 50% and 100% of user overlap levels, the PreF algorithm had a low standard deviation and achieved the best predictive performance among the algorithms. Figure 71 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. The statistical significance tests verified that both the PreF and PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure baseline algorithm for 50% 5.2. Evaluation Results 181 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 72 – Overall prediction performance (MAE) boxplots for Music domain in the location dimension with different user overlap levels (source domain: book). and 100% of user overlap levels57. For 10% of user overlap, the applied tests could not verify a statistical difference between the performance of the proposed algorithms and NNUserNgbr-transClosure58. Figure 72a, Figure 72b and Figure 72c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. 57 For all tests, W=99 and p-value=0.005413 58 W=60 and p-value=0.35 (PreF x NNUserNgbr-transClosure), while W=80 and p-value=0.1 (PostF x NNUserNgbr-transClosure) 182 Chapter 5. CD-CARS Evaluation With respect to the classification performance, Figures 73a, 73b, and 73c present the results of the F-metric at different top ‘N’ values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Music domain as target, considering the Location dimension. As it can be seen, for 50% and 100% of user overlap the PostF outperformed the baseline algorithms for ‘N’ up to ten. On the other hand, the PreF classification performance was worse than all other algorithms in all user overlap levels and ‘N’ values. Figure 74 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for 50% and 100% of user overlap levels59, whereas the opposite from this was observed for 10% of user overlap60. On the other hand, the applied tests also verified that the NNUserNgbr-transClosure F-metric values were greater than the PreF ones for all user overlap levels61. 5.2.2.1.3 Companion Dimension Table 31 shows the overall predictive performance of the recommender algorithms, considering all contextual values from the Companion dimension and different user overlap levels for the Music domain as target. As it can be seen from the table, the addition of user ratings from other domain (Book), by using the same algorithm for cross-domain recommendation (corresponding to the two first rows of the table), improved the predictive performance in, approximately, 3–17% (MAE) and 6–16% (RMSE) depending on the user overlap level. Also, Table 31 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 9–68% (MAE) and 4–47% (RMSE) depending on the user overlap levels. As it can be seen from table, the PostF predictive performance was better than the NNUserNgbr-transClosure algorithm for 50% and 100% of user overlap levels if we do not consider the standard deviation of the algorithms for 50% of user overlap. Its improvement in comparison to the NNUserNgbr-transClosure performance varied in, approximately, 3–11% (MAE) and 6–11% (RMSE) depending on the user overlap level. In addition, the predictive performance of the PreF algorithm was worse than all other algorithms in the Companion dimension for 50% and 100% of user overlap levels, as showed in Table 31. Figure 75 (MAE) and Figure 76 (RMSE) illustrate the predictive 59 p-value=0.005413 and W=99 for all tests 60 p-value=0.006314 and W=97 61 p-value=0.003912 and W=100 for all tests 5.2. Evaluation Results 183 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 73 – F-metric performance x top ‘N’ items for the Music domain in the location dimension with different user overlap levels (source domain: book). 184 Chapter 5. CD-CARS Evaluation Figure 74 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the location dimension (target domain: Music, and source: book). Table 31 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Companion dimension (source domain: Book, and target domain: Music). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.684 ± 0.026 0.972 ± 0.053 0.654 ± 0.016 0.970 ± 0.030 0.588 ± 0.010 0.903 ± 0.017 NNUserNgbr (cross-domain) 0.660 ± 0.022 0.818 ± 0.051 0.537 ± 0.013 0.823 ± 0.028 0.535 ± 0.008 0.845 ± 0.014 NNUserNgbr- transClosure 0.205 ± 0.122 0.430 ± 0.188 0.479 ± 0.004 0.785 ± 0.014 0.486 ± 0.018 0.797 ± 0.022 PreF with NNUserNgbr- transClosure 0.438 ± 0.221 0.540 ± 0.283 0.666 ± 0.058 1.076 ± 0.024 0.711 ± 0.052 1.079 ± 0.090 PostF with NNUserNgbr- transClosure 0.381 ± 0.076 0.807 ± 0.279 0.465 ± 0.044 0.736 ± 0.102 0.431 ± 0.016 0.707 ± 0.016 performance of the proposed algorithms over different user overlap levels. Note that for this case we showed the figures for both predictive metrics, since we observed a difference in the PreF and PostF performances depending on the predictive metric used in the evaluation. The statistical significance tests verified that the PostF predictive error (MAE) was less than the NNUserNgbr-transClosure algorithm for 100% of user overlap level62, whereas none statistical difference between their performances was observed when the user 62 W=100 and p-value=0.003968 5.2. Evaluation Results 185 Figure 75 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: book, and target domain: Music). Figure 76 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: book, and target domain: Music). overlap levels were 10%63 and 50%64. The applied tests also verified that the NNUserNgbr- transClosure predictive errors (MAE) were less than the PreF ones for all user overlap levels65. Figure 77a, Figure 77b and Figure 77c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. 63 W=55 and p-value=0.5 64 W=30 and p-value=0.8 65 W=100 and p-value=0.003968 for all tests 186 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 77 – Overall prediction performance (MAE) boxplots for Music domain in the companion dimension with different user overlap levels (source domain: book). Figures 78a, 78b, and 78c present the results of the F-metric at different top ‘N’ values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Music domain as target, considering the Companion dimension. As it can be seen, the proposed algorithms only outperformed the baseline ones for 50% and 100% of user overlap with very low values of top ‘N’. Figure 79 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the NNUserNgbr-transClosure F- metric values were greater than the both proposed algorithms for all user overlap levels66. 66 p-value=0.005814 and W=97 for all tests between the PostF and baseline algorithms, while p- 5.2. Evaluation Results 187 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 78 – F-metric performance x top ‘N’ items for the Music domain in the companion dimension with different user overlap levels (source domain: book). value=0.003913 and W=100 for all tests between the PreF and baseline algorithms 188 Chapter 5. CD-CARS Evaluation Figure 79 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the companion dimension (target domain: Music, and source: book). 5.2.2.1.4 Combining Contextual Dimensions In the previous sections, we presented the evaluation results regarding the contextual dimensions separately. In this section, we report the results for a combination of two contextual dimensions considering the same evaluation metrics and methodology described before. Table 32 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual value combinations from the temporal and location dimensions (source domain: Book, and target domain: Music). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.684 ± 0.026 0.972 ± 0.053 0.654 ± 0.016 0.970 ± 0.030 0.588 ± 0.010 0.903 ± 0.017 NNUserNgbr (cross-domain) 0.708 ± 0.022 1.076 ± 0.051 0.597 ± 0.013 0.929 ± 0.028 0.548 ± 0.008 0.874 ± 0.014 NNUserNgbr- transClosure 0.302 ± 0.060 0.628 ± 0.078 0.474 ± 0.009 0.818 ± 0.014 0.489 ± 0.004 0.825 ± 0.008 PreF with NNUserNgbr- transClosure 0.302 ± 0.278 0.755 ± 0.426 0.485 ± 0.040 0.741 ± 0.036 0.265 ± 0.036 0.642 ± 0.156 PostF with NNUserNgbr- transClosure 0.304 ± 0.021 0.632 ± 0.036 0.338 ± 0.005 0.621 ± 0.002 0.376 ± 0.004 0.706 ± 0.010 Table 32 reports the overall predictive performance of the recommender algorithms, considering all contextual value combinations from the Temporal and Location dimensions 5.2. Evaluation Results 189 with different user overlap levels for the Music domain as target. As it was observed in the previous sections, the addition of user ratings from the Book domain also improved the predictive performance of the NNUserNgbr algorithm in, approximately, 6–8% (MAE) and 3–4% (RMSE) depending on the user overlap level (for 50% and 100% of user overlap). Figure 80 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: book, and target domain: music). Figure 81 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: book, and target domain: music). Furthermore, Table 32 presents the overall performance of the NNUserNgbr- transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUser-Ngbr one (performed for cross-domain purposes) by achieving 190 Chapter 5. CD-CARS Evaluation an improvement that varied in, approximately, 10–57% (MAE) and 5–41% (RMSE) de- pending on the user overlap levels. As it can be seen from table, except when the user overlap level was 10%, the PostF predictive performance was better than the NNUserNgbr- transClosure algorithm in all other user overlap levels, with an improvement that varied in, approximately, 23–28% (MAE) and 14–24% (RMSE) depending on the user overlap level. Despite the NNUserNgbr-transClosure algorithm have outperformed the PostF when the user overlap level was 10%, they had a similar performance, separated only by their standard deviations. Note that the PreF predictive performance had a high standard deviation, observed in Table 32, as occurred in the evaluation of the Location dimension alone. In addition, if we consider the high standard deviation of the PreF algorithm showed in Table 32, then we can say that the PostF outperformed the PreF algorithm for 10% and 50% of user overlap level. For 100% of user overlap, the PreF algorithm had a low standard deviation and achieved the best predictive performance among the algorithms. Figure 80 (MAE) and Figure 81 (RMSE) illustrate the predictive performance of the proposed algorithms over different user overlap levels. Note that for this case we showed the figures for both predictive metrics, since we observed a difference in the PreF and PostF performances depending on the predictive metric used in the evaluation. The statistical significance tests verified that the PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure algorithm for the 50% and 100% user overlap levels67. The applied tests also verified that the PreF predictive errors (MAE) were less than the NNUserNgbr-transClosure algorithm for 100% user overlap68, whereas the opposite from this was observed for 50% of user overlap69. When the user overlap level was 10%, the applied tests could not verify a statistical difference between the performance of the proposed algorithms and NNUserNgbr-transClosure70. Figure 82a, Figure 82b and Figure 82c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. Taking into account the classification performance, Figures 83a, 83b, and 83c present the results of the F-metric at different top ‘N’ values, respectively, in 10%, 50%, and 100% of user overlap level for the Music domain as target, considering the Temporal and Location dimensions. For 50% and 100% of user overlap levels, the PostF classification performance was better or similar than the baseline algorithms for all top ‘N’ values, while for 10% of user overlap the PostF classification performance was better or similar than them for very low top ‘N’ values (1 to 3). On other hand, the PreF classification 67 In both cases, with W=99 and p-value=0.005413 68 W=100 and p-value=0.003968 69 W=99 and p-value=0.005413 70 W=61 and p-value=0.5 (PostF x NNUserNgbr-transClosure), whereas W=71 and p-value=0.8 (PreF x NNUserNgbr-transClosure) 5.2. Evaluation Results 191 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 82 – Overall prediction performance (MAE) boxplots for Music domain in the temporal and location dimensions with different user overlap levels (source domain: book). performance was worse than all other algorithms. Figure 84 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for all user overlap levels71. On the other hand, the applied tests also verified that the NNUserNgbr-transClosure F-metric values were greater than the PreF ones for all user 71 For 10% of user overlap, with W=97 and the p-value=0.037, and p-value=0.005413 and W=97 for the other user overlap levels 192 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 83 – F-metric performance x top ‘N’ items for the Music domain in the temporal and location dimensions with different user overlap levels (source domain: book). 5.2. Evaluation Results 193 Figure 84 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal and location dimensions (target domain: Music, and source: book). overlap levels72. 5.2.2.2 Book as Target Domain In the Section 5.2.2.1, we presented the results for the Music target domain, which had fewer ratings in the cross-domain dataset in comparison to Book source domain (as described in Section 4.1.3.2). In this section, we present the results when Book is the target domain and Music is the source domain. According to the contextual dimensions present in Section 4.1.3.2, we describe the evaluation results for each contextual dimension in the following sections. In addition, we show the results for a combination of contextual dimensions in Section 5.2.2.2.4. 5.2.2.2.1 Temporal Dimension Table 33 reports the overall predictive performance of the recommender algorithms, considering all contextual values from the Temporal dimension and different user overlap levels for the Book domain as target. The rows 1 and 2 from the table show the NNUser- Ngbr predictive performance when it is applied in the single-domain and cross-domain recommendations. As it can be seen, likewise the results showed in Section 5.2.2.1.1 (source domain: Book, target domain: Music, and Temporal dimension), the simple addition of user ratings from other domain (Music), by using the same algorithm for cross-domain recommendation, improved the recommendation performance in, approximately, 9–13% (MAE) and 6–7% (RMSE) depending on the user overlap level. 72 p-value=0.003914 and W=99 for all tests 194 Chapter 5. CD-CARS Evaluation Table 33 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Temporal dimension (source domain: Music, and target domain: Book). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.671 ± 0.024 1.002 ± 0.048 0.505 ± 0.008 0.792 ± 0.020 0.419 ± 0.006 0.735 ± 0.012 NNUserNgbr (cross-domain) 0.610 ± 0.022 0.942 ± 0.044 0.443 ± 0.007 0.740 ± 0.018 0.363 ± 0.005 0.679 ± 0.010 NNUserNgbr- transClosure 0.130 ± 0.012 0.368 ± 0.034 0.197 ± 0.001 0.451 ± 0.005 0.202 ± 0.003 0.473 ± 0.006 PreF with NNUserNgbr- transClosure 0.109 ± 0.018 0.346 ± 0.063 0.087 ± 0.005 0.340 ± 0.006 0.090 ± 0.002 0.300 ± 0.005 PostF with NNUserNgbr- transClosure 0.114 ± 0.004 0.349 ± 0.011 0.176 ± 0.002 0.399 ± 0.004 0.175 ± 0.003 0.418 ± 0.007 In addition, Table 33 presents the overall performance of the PreF and PostF algorithms, besides the base NNUserNgbr-transClosure algorithm, which outperformed the NNUserNgbr algorithm (performed for cross-domain purposes) by achieving an im- provement that varied in, approximately, 44–78% (MAE) and 30–60% (RMSE) depending on the user overlap levels. As it can be seen from the table, the PreF predictive performance was better than the NNUserNgbr-transClosure for all user overlap levels. Besides, it was better than the PostF algorithm for all user overlap levels if we do not consider the standard deviation. The improvement achieved by the PreF algorithm in comparison to the NNUserNgbr- transClosure one varied in, approximately, 16–56% (MAE) and 6–36% (RMSE) depending on the user overlap level. The PostF predictive performance was also better than the NNUserNgbr-transClosure for all user overlap levels. Figure 85 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. The statistical significance tests verified that the PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure baseline algorithm for all user overlap levels73. Also, the tests verified that, except for 10% of user overlap74, the PreF predictive errors (MAE) were less than the NNUserNgbr-transClosure algorithm for all other user overlap levels75. Figure 86a, Figure 86b and Figure 86c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. 73 p-value=0.003968 and W=97 for all tests 74 W=79 and p-value=0.1 75 p-value=0.003968 and W=97 for all tests 5.2. Evaluation Results 195 Figure 85 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal dimension (source domain: Music, and target domain: book). Regarding the classification performance, Figures 87a, 87b, and 87c present the results of the F-metric at different N values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Book domain as target, considering the Temporal dimension. As it can be seen, the PostF and PreF classification performances were better or similar than the NNUserNgbr-transClosure baseline algorithm for 50% and 100% of user overlap levels (with any top ‘N’ value for the PostF algorithm and with very low top ‘N’ values for the PreF one). In addition, for 10% of user overlap, the PostF classification performance was better than the NNUserNgbr-transClosure baseline algorithm only for very low top ‘N’ values (1 and 2), whereas the PreF classification performance was worse than all algorithms. Finally, the PostF classification performance was better than the PreF one for all user overlap levels and top ‘N’ values. Figure 88 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for 50% and 100% of user overlap levels76, whereas their performances were statistically similar for 10% of user overlap77. The applied tests also verified that the NNUserNgbr-transClosure F-metric values were greater than the PreF ones for all user overlap levels78. 76 p-value=0.005413 and W=97 for all tests 77 p-value=0.35 and W=60 78 W=77 and p-value=0.02163 for 10% of user overlap. For 50% and 100%, p-value=0.005413 and W=97 196 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 86 – Overall prediction performance (MAE) boxplots for book domain in the temporal dimension with different user overlap levels (source domain: Music). 5.2.2.2.2 Location Dimension Table 34 reports the overall predictive performance of the recommender algorithms, considering all contextual values from the Location dimension and different user overlap levels for the Book domain as target. As it can be seen from the table, the addition of user ratings from other domain (Music), by using the same algorithm for cross-domain recommendation (corresponding to the two first rows of the table), improved the predictive performance in, approximately, 10–12% (MAE) and 6–10% (RMSE) depending on the user overlap level. 5.2. Evaluation Results 197 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 87 – F-metric performance x top ‘N’ items for the book domain in the temporal dimension with different user overlap levels (source domain: Music). 198 Chapter 5. CD-CARS Evaluation Figure 88 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal dimension (target domain: book, and source: Music). Table 34 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Location dimension (source domain: Book, and target domain: Music). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.671 ± 0.024 1.002 ± 0.048 0.505 ± 0.008 0.792 ± 0.020 0.419 ± 0.006 0.735 ± 0.012 NNUserNgbr (cross-domain) 0.587 ± 0.022 0.897 ± 0.044 0.450 ± 0.007 0.739 ± 0.018 0.369 ± 0.005 0.678 ± 0.010 NNUserNgbr- transClosure 0.120 ± 0.010 0.362 ± 0.036 0.200 ± 0.002 0.443 ± 0.006 0.199 ± 0.004 0.467 ± 0.007 PreF with NNUserNgbr- transClosure 0.122 ± 0.206 0.349 ± 0.480 0.149 ± 0.102 0.406 ± 0.276 0.096 ± 0.050 0.293 ± 0.151 PostF with NNUserNgbr- transClosure 0.109 ± 0.016 0.317 ± 0.036 0.193 ± 0.004 0.428 ± 0.009 0.193 ± 0.003 0.449 ± 0.010 Also, Table 34 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 45–79% (MAE) and 31–59% (RMSE) depending on the user overlap levels. As it can be seen from table, the PostF predictive performance was better than the NNUserNgbr-transClosure algorithm in all user overlap levels, with an improvement that varied in, approximately, 3–9% (MAE) and 3–12% (RMSE) depending on the user overlap level. In addition, we can see in Table 34 that the PostF outperformed the PreF algorithm 5.2. Evaluation Results 199 for the majority of the user overlap levels (10% and 50%) if we consider the high standard deviation of the PreF algorithm. Figure 89 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. Figure 89 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the location dimension (source domain: book, and target domain: Music). The statistical significance tests verified that, except when the user overlap level was 10%79, the PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure baseline algorithm for all other user overlap levels80. The applied tests also verified that the PreF predictive error (MAE) was less than the NNUserNgbr-transClosure one for 100% of user overlap81, whereas their performances were statistically similar when the user overlap level were 10%82 and 50%83. Figure 90a, Figure 90b and Figure 90c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. With respect to the classification performance, Figures 91a, 91b, and 91c present the results of the F-metric at different top ‘N’ values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Book domain as target, considering the Location dimension. As it can be seen, the PostF classification performance was better or similar than the NNUserNgbr-transClosure baseline algorithm for 100% of user overlap with any top ‘N’ value, whereas for 50% of user overlap it was better than the baseline algorithm only 79 W=76 and p-value=0.2 80 For all tests, W=22 and p-value=0.02778 81 W=97 and p-value=0.003968 82 W=40 and p-value=0.65 83 W=30 and p-value=0.8 200 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 90 – Overall prediction performance (MAE) boxplots for Music domain in the location dimension with different user overlap levels (source domain: book). for low top ‘N’ values (1 to 3). For 10% of user overlap the PostF was outperformed by the baseline algorithm. On the other hand, the PreF classification performance was worse than all other algorithms in all user overlap levels and ‘N’ values. Figure 92 shows the variation of the F-metric value in different user overlap levels 5.2. Evaluation Results 201 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 91 – F-metric performance x top ‘N’ items for the book domain in the location dimension with different user overlap levels (source domain: Music). 202 Chapter 5. CD-CARS Evaluation Figure 92 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the location dimension (target domain: book, and source: Music). by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric value was greater than the NNUserNgbr-transClosure baseline algorithm for 100% of user overlap84, whereas the opposite from this was observed for 10% and 50% of user overlap levels85. The applied tests also verified that the NNUserNgbr-transClosure F-metric values were greater than the PreF ones for all user overlap levels86. 5.2.2.2.3 Companion Dimension Table 35 shows the overall predictive performance of the recommender algorithms, considering all contextual values from the Companion dimension and different user overlap levels for the Book domain as target. As it can be seen from the table, the addition of user ratings from other domain (Music), by using the same algorithm for cross-domain recommendation (corresponding to the two first rows of the table), improved the predictive performance in, approximately, 2% (MAE) and 1% (RMSE) for 10% of user overlap. Also, Table 35 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 33–74% (MAE) and 23–52% (RMSE) depending on the user over- lap level. The PostF predictive performance was better than the NNUserNgbr-transClosure one for all user overlap levels with an improvement that varied in, approximately, 5–10% (MAE) and 5–11% (RMSE) depending on the user overlap level. 84 W=97 and p-value=0.037 85 W=97 and p-value=0.005413 for both cases 86 W=99 and p-value=0.003914 for all tests 5.2. Evaluation Results 203 Table 35 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual values from the Companion dimension (source domain: Music, and target domain: Book). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.671 ± 0.024 1.002 ± 0.048 0.505 ± 0.008 0.792 ± 0.020 0.419 ± 0.006 0.735 ± 0.012 NNUserNgbr (cross-domain) 0.653 ± 0.022 0.994 ± 0.044 0.513 ± 0.007 0.814 ± 0.018 0.461 ± 0.005 0.785 ± 0.010 NNUserNgbr- transClosure 0.168 ± 0.035 0.470 ± 0.085 0.290 ± 0.013 0.576 ± 0.017 0.306 ± 0.004 0.604 ± 0.136 PreF with NNUserNgbr- transClosure 0.984 ± 0.251 1.232 ± 0.154 0.693 ± 0.011 1.035 ± 0.030 0.721 ± 0.014 1.041 ± 0.016 PostF with NNUserNgbr- transClosure 0.151 ± 0.028 0.417 ± 0.119 0.264 ± 0.017 0.522 ± 0.032 0.289 ± 0.006 0.573 ± 0.019 In addition, the predictive performance of the PreF algorithm was worse than all other algorithms in the Companion dimension, as showed in Table 35. Figure 93 illustrates the predictive performance (MAE) of the proposed algorithms over different user overlap levels. Figure 93 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the companion dimension (source domain: Music, and target domain: book). The statistical significance tests verified that the PostF predictive error (MAE) was less than the NNUserNgbr-transClosure algorithm for 100% of user overlap87. For 87 W=99 and p-value=0.003968 204 Chapter 5. CD-CARS Evaluation 10% and 50% of user overlap, the NNUserNgbr-transClosure predictive errors (MAE) were statistically similar to the PostF ones88. On the other hand, the applied tests also verified that the NNUserNgbr-transClosure predictive errors (MAE) were less than the PreF ones for all user overlap levels89. Figure 94a, Figure 94b and Figure 94c show the boxplots with the prediction performance (MAE) of the algorithms, respectively, for the 10%, 50%, and 100% user overlap levels. Figures 95a, 95b, and 95c present the results of the F-metric at different top ‘N’ values (between one and twenty), respectively, in 10%, 50%, and 100% of user overlap level for the Book domain as target, considering the Companion dimension. As it can be seen, the proposed algorithms were outperformed by the NNUserNgbr-transClosure baseline algorithm for all user overlap levels and any values of top ‘N’. Figure 96 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the NNUserNgbr-transClosure F- metric values were greater than the both proposed algorithms for all user overlap levels90. 5.2.2.2.4 Combining Contextual Dimensions In the previous sections, we presented the evaluation results regarding the contextual dimensions separately. In this section, we report the results for a combination of two contextual dimensions considering the same evaluation metrics and methodology described before. Table 36 reports the overall predictive performance of the recommender algorithms, considering all contextual value combinations from the Temporal and Location dimensions with different user overlap levels for the Book domain as target. As it was observed in the previous sections, the addition of user ratings from the Music domain also improved the predictive performance of the NNUserNgbr in, approximately, 10–12% (MAE) and 6–10% (RMSE) depending on the user overlap level (rows 1 and 2 from the table). Also, Table 36 presents the overall performance of the NNUserNgbr-transClosure, PreF and PostF algorithms. The NNUserNgbr-transClosure algorithm outperformed the NNUserNgbr one (performed for cross-domain purposes) by achieving an improvement that varied in, approximately, 46–81% (MAE) and 31–59% (RMSE) depending on the user overlap levels. As it can be seen from table, the PostF predictive performance was better than the NNUserNgbr-transClosure algorithm in all user overlap levels, with an 88 For 10% of user overlap, W=60 and p-value=0.35, while for 50% of user overlap, W=71 and p-value=0.2 89 W=99 and p-value=0.003968 for all tests 90 p-value=0.005814 and W=97 for all tests between the PostF and baseline algorithms, while p- value=0.003913 and W=100 for all tests between the PreF and baseline algorithms 5.2. Evaluation Results 205 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 94 – Overall prediction performance (MAE) boxplots for book domain in the companion dimension with different user overlap levels (source domain: Music). improvement that varied in, approximately, 7–23% (MAE) and 8–18% (RMSE) depending on the user overlap level. In addition, if we do not consider the standard deviation of the PreF and PostF algorithms showed in Table 32, then we can say that the PreF predictive performance 206 Chapter 5. CD-CARS Evaluation (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 95 – F-metric performance x top ‘N’ items for the book domain in the companion dimension with different user overlap levels (source domain: Music). 5.2. Evaluation Results 207 Figure 96 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the companion dimension (target domain: book, and source: Music). Table 36 – Overall predictive performance (MAE/RMSE) with standard deviation (std) by varying the user overlap level for all contextual value combinations from the temporal and location dimensions (source domain: Music, and target domain: Book). Algorithm 10% overlap 50% overlap Full overlap MAE±std RMSE±std MAE±std RMSE±std MAE±std RMSE±std NNUserNgbr (single-domain) 0.671 ± 0.024 1.002 ± 0.048 0.505 ± 0.008 0.792 ± 0.020 0.419 ± 0.006 0.735 ± 0.012 NNUserNgbr (cross-domain) 0.587 ± 0.022 0.897 ± 0.044 0.450 ± 0.007 0.739 ± 0.018 0.369 ± 0.005 0.678 ± 0.010 NNUserNgbr- transClosure 0.111 ± 0.009 0.361 ± 0.034 0.199 ± 0.002 0.434 ± 0.005 0.197 ± 0.003 0.464 ± 0.006 PreF with NNUserNgbr- transClosure 0.180 ± 0.226 0.480 ± 0.465 0.209 ± 0.122 0.510 ± 0.296 0.140 ± 0.060 0.405 ± 0.161 PostF with NNUserNgbr- transClosure 0.103 ± 0.014 0.295 ± 0.030 0.175 ± 0.003 0.396 ± 0.007 0.150 ± 0.002 0.378 ± 0.006 (measured by the MAE metric) was better than the PostF one when the user overlap level was 100%. Figure 97 (MAE) and Figure 98 (RMSE) illustrate the predictive performance of the proposed algorithms over different user overlap levels. Note that for this case we showed the figures for both predictive metrics, since we observed a difference in the PreF performance depending on the predictive metric used in the evaluation. The statistical significance tests verified that the PostF predictive errors (MAE) were less than the NNUserNgbr-transClosure algorithm for the 50% and 100% user overlap 208 Chapter 5. CD-CARS Evaluation Figure 97 – Overall prediction error (MAE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: music, and target domain: book). Figure 98 – Overall prediction error (RMSE) for cross-domain algorithms by varying user overlap level in the temporal and location dimensions (source domain: music, and target domain: book). levels91. When the user overlap levels was 10%, the applied tests could not verify a statistical difference between the performance of the PostF and NNUserNgbr-transClosure92 algorithms. On the other hand, the applied tests verify that the NNUserNgbr-transClosure predictive errors (MAE) were less than the PreF ones for 10% and 50% of user overlap 91 In both cases, with W=100 and p-value=0.003968 92 W=51 and p-value=0.35 5.2. Evaluation Results 209 levels93, whereas the opposite from this was observed for 100% of user overlap94. (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 99 – Overall prediction performance (MAE) boxplots for book domain in the temporal and location dimensions with different user overlap levels (source domain: Music). Taking into account the classification performance, Figures 61a, 61b, and 61c present the results of the F-metric at different top ‘N’ values, respectively, in 10%, 50%, and 100% of user overlap level for the Music domain as target, considering the combination between the Temporal and Location dimensions. For all user overlap levels and top ‘N’ values, the PreF classification performance was worse than the all other algorithms. 93 For 10% of user overlap, W=99 and p-value=0.003968. For 50% of user overlap, W=97 and p-value=0.005413 94 W=99 and p-value=0.003968 210 Chapter 5. CD-CARS Evaluation On the other hand, the PostF classification performance was better or similar than the NNUserNgbr-transClosure baseline algorithm in the majority of the user overlap levels (50% and 100%) for any top ‘N’ values. Figure 101 shows the variation of the F-metric value in different user overlap levels by fixing the top ‘N’ value to five. The statistical significance tests verified that the PostF F-metric values were greater than the NNUserNgbr-transClosure baseline algorithm for 50% and 100% of user overlap levels95, whereas the opposite from this was observed for 10% of user overlap96. On the other hand, the applied tests also verified that the NNUserNgbr-transClosure F-metric values were greater than the PreF ones for all user overlap levels97. 5.2.2.3 Summary In this section, we provide a summary of the results from the evaluation of the “book-music dataset”. Figure 102 shows a dispersion diagram illustrating the predictive performance (MAE) for the algorithms by varying target domain (Book and Music), contextual dimension and user overlap levels, whereas Figure 103 shows the same, but considering the RMSE metric. It is important to mention that these figures do not take into account the standard deviation and the statistical significance of the results. Table 37 presents the predictive performance (MAE) achieved by the PreF and PostF algorithms in comparison to the best baseline algorithm (NNUserNgbr-transClosure), by taking into account their statistical significance98 and different target domain, contextual dimension and user overlap levels. Regarding the classification performance, Figure 104 presents a dispersion diagram illustrating the F-metric performance (with N=5) for the algorithms by varying target domain (Book and Music), contextual dimension and user overlap levels. Once again, it is important to mention that we are not considering the standard deviation and the statistical significance of the results in that figure. Table 38 shows the classification performance improvement (F-metric with N=5) obtained by the PreF and PostF algorithms in comparison to the best baseline algorithm (NNUserNgbr-transClosure), by taking into account their statistical significance99 and different target domain, contextual dimension and user overlap levels. As it can be seen, at least one proposed algorithm (PreF or PostF) achieved the best predictive performance among the algorithms (or it was similar to the best one) in almost all scenarios (with distinct target domains, contextual dimensions, and user overlap 95 p-value=0.005814 and W=97 for all tests 96 p-value=0.035 and W=75 97 p-value=0.003913 and W=99 for all tests 98 In the table, “**” means that the result could not be considered statistically significant. 99 In the table, “**” means that the result could not be considered statistically significant. 5.2. Evaluation Results 211 (a) 10% of user overlap. (b) 50% of user overlap. (c) 100% of user overlap. Figure 100 – F-metric performance x top ‘N’ items for the book domain in the temporal and location dimensions with different user overlap levels (source domain: Music). 212 Chapter 5. CD-CARS Evaluation Figure 101 – Overall classification performance (F-metric at 5) for the algorithms by varying user overlap level in the temporal and location dimensions (target domain: book, and source: Music). levels). By considering the classification metric, the PostF algorithm achieved the best performance among the algorithms (or it was similar to the best one) in the majority of the scenarios. Most of the findings mentioned in the summary of the evaluation results for the “book-television dataset” (see Section 5.2.1.3) can also be mentioned in this summary (“book-music dataset”). In this way, we only highlight the main differences found in this summary in comparison to those findings: • The addition of user ratings from an auxiliary (source) domain also improved the predictive performance of the NNUserNgbr algorithm, but in this dataset this fact has occurred in less scenarios than in the “book-television dataset”. This can also be observed for the classification performance of that algorithm. Likewise the results in the “book-television dataset”, that improvement occurred even when a source domain had less ratings than the target domain. • Likewise the results in the “book-television dataset”, the proposed algorithms (PreF and PostF) had better predictive and classification performances in the Temporal dimension than others dimensions. The same findings mentioned for the predictive and classification performances of the PostF algorithm can be observed in this summary, as well as the PreF predictive performance. On the other hand, the PreF algorithm outperformed the NNUserNgbr-transClosure one in less scenarios (user overlap levels and target domains) than in that dataset by considering the classification performance. The PreF algorithm outperformed the NNUserNgbr- transClosure one only when the Music was the target domain with 100% of user 5.2. Evaluation Results 213 overlap. • Likewise the results in the “book-television dataset” regarding the combination of con- textual dimensions (Temporal and Location), the PostF predictive and classification performances in that combination were close to their own performances using only the Temporal dimension as single source of contextual information, whereas the PreF predictive and classification performances were similar to their own performances using only the Location dimension. In addition, the predictive and classification performances of the PreF algorithm was also decreased with the addition of contex- tual information from other contextual dimension. Besides, the PostF predictive performance was also increased, however, its classification performance was increased with the addition of contextual information from other contextual dimension in opposite to the results from that dataset. Table 37 – Overall predictive performance (MAE) of the proposed algorithms in comparison to the best baseline one by varying target domain (book and music), contextual dimension and user overlap levels. Contextual dimension Target Domain User Overlap Level PreF Improve- ment PostF Improve- ment Temporal Music 10% 39.7% 16.2% Book 10% 16.1%** 12.4% Music 50% 62.6% 12.8% Book 50% 56% 10.8% Music 100% 55.8% 11.9% Book 100% 55.4% 13.3% Location Music 10% 24.5%** 16.1%** Book 10% -2.9%** 8.9%** Music 50% 39.2% 6.7% Book 50% 25.5%** 3.7% Music 100% 57.4% 2.9% Book 100% 52% 3.8% Companion Music 10% -113.5% -85.7%** Book 10% -484% 10.3%** Music 50% -39.1% 2.9%** Book 50% -139.3% 8.7% ** Music 100% -46.5% 11.2% Book 100% -136% 5.4% Temporal and Location Music 10% -0.2%** -0.9%** Book 10% -62.1% 7.3%** Music 50% -2.3% 28.6% Book 50% -4.9% 12.1% Music 100% 45.9% 23.1% Book 100% 29.3% 24% 214 Chapter 5. CD-CARS Evaluation Figure 102 – Predictive performance (MAE) for the algorithms by varying target domain (book and music), contextual dimension and user overlap levels (dispersion diagram). 5.2. Evaluation Results 215 Figure 103 – Predictive performance (RMSE) for the algorithms by varying target domain (book and music), contextual dimension and user overlap levels (dispersion diagram). 216 Chapter 5. CD-CARS Evaluation Figure 104 – Classification performance (F-metric with N=5) for the algorithms by varying target domain (book and music), contextual dimension and user overlap levels (dispersion diagram). 5.2. Evaluation Results 217 Table 38 – Overall classification performance (F-metric with N=5) of the proposed algo- rithms in comparison to the best baseline one by varying target domain (book and music), contextual dimension and user overlap levels. Contextual dimension Target Domain User Overlap Level PreF Improve- ment PostF Improve- ment Temporal Music 10% -86.6% 0.5%** Book 10% -121.6% -4.6%** Music 50% -22.4% 22.1% Book 50% -41.1% 6.9% Music 100% 13% 37.9% Book 100% -13.8% 13.9% Location Music 10% -157.4% -11.3% Book 10% -234% -15.9% Music 50% -311% 17.1% Book 50% -420.5% -3.4% Music 100% -223.9% 33.3% Book 100% -455% 6% Companion Music 10% -58.7% -23.3% Book 10% -126.4% -51.9% Music 50% -92% -54.1% Book 50% -197% -39.6% Music 100% -29.7% -35.2% Book 100% -162.2% -34.3% Temporal and Location Music 10% -168% 9.1% Book 10% -292.3% -7.1% Music 50% -339.3% 23.3% Book 50% -408% 12% Music 100% -270.7% 42.1% Book 100% -444% 17.2% 5.2.3 Discussion Given the evaluation results presented in the previous sections, we can say that the use of context-aware techniques has proven to be a good approach in order to improve the cross-domain recommendation quality in comparison to traditional cross-domain recom- mender systems based on collaborative filtering techniques, which do not take contextual information into account. This finding was observed in the presented experiments, in which we evaluated two CD-CARS algorithms performed in two different datasets (“Book- television” and “Book-music”) by varying their target domains (Television, Music and Book), contextual dimensions (Temporal, Location and Companion), and user overlap levels (10%, 50%, and 100%). As we could see in the experiments, Temporal was the contextual dimension in which the proposed algorithms had a better performance for all datasets, target domains 218 Chapter 5. CD-CARS Evaluation and user overlap levels. This may have happened due to the great amount of contextual information obtained in that contextual dimension (100% of the ratings had temporal information) in comparison to other ones (Location with, approximately, a half of ratings, and Companion with, approximately, 20% from the ratings, as described in Section 4.1.3). This fact contrasts to the information gain verified in Section 4.1.2, where the Location dimension with the City attribute had the greater value for all target domains. In this way, more studies and experiments may be made in the future in order to determine the best contextual dimensions, attributes and values before evaluating the proposed algorithms, especially in the combination of contextual dimensions. In addition, in these studies we could verify the quality of recommendation of the proposed algorithms by reducing the number of temporal information present in the user ratings (“contextual sensitivity”). In addition, the quality of the contextual information may also have influenced on the recommendation quality for the proposed algorithms. As we have seen in Section 4.1.1.3, the Companion dimension has a poor quality of contextual information. Especially for the PreF algorithm, which filters ratings out from the target domain that are not from the recommendation context, the recommendation quality was more impaired than the PostF one, which uses the same set of ratings of the baseline algorithm for prediction calculations, and only in the end of the recommendation process ignores predictions (instead of initial ratings). The combination of two contextual dimensions (Temporal and Location) in the recommendation process generated controversial results for the proposed algorithms. Inde- pendently of the dataset used, while the PreF had worse results in that combination than using only one contextual dimension (Temporal or Location), the PostF recommendation quality was improved for some situations, as it could be seen in the result summaries (Section 5.2.1.3 and Section 5.2.2.3). Again, this may have happened given the PreF feature, in which might be more susceptible to problems in a situation with just a few number of ratings, generated by the contextual specialization from the combination of two contextual dimensions. It is important to remember that all contextual information used in the CD-CARS was obtained implicitly or by inference (see Section 4.1.1). In this way, there is no assurance that the contextual information acquired reflects the actual contextual information of the ratings. For instance, a user could watch a movie on Saturday and rate it only on Sunday, when the rating timestamp was observed. Thus, the actual temporal information of that rating might have been compromised. However, even considering this issue, it was possible to verify that the proposed algorithms had a good performance in the Temporal dimension. Besides, the Location context is extracted from the users’ accounts through their original IDs (i.e. static and single location of the users are obtained from their website accounts). In this way, the contextual information is little exploited when a user receives 5.2. Evaluation Results 219 the recommendation of items for a location different from his/her location (e.g. a user that has all ratings in the United States and receives recommendations in Brazil), especially for the PreF algorithm. For it, the recommendation would be fully based on the user similarities from the source domain, since the user would not have any rating in the target domain (pre-filtered in a location that the user does not have any information). So, it would not be possible to calculate the similarity between the user and other users with ratings in the target domain. On the other hand, for the PostF algorithm, no recommendation would be possible without the association rules once that the user would not have any contextual preferences in that location. Taking into account the PostF and PreF recommendation performances, we could observe that they had distinct results depending on the evaluation metric adopted. For example, considering the Temporal dimension, the PostF classification performance (F- metric) was better than the PreF one, whereas the opposite was observed when they were evaluated by means of predictive error metrics (MAE and RMSE). In other contextual dimensions, independently of the dataset, we have seen that the PostF had a better recommendation quality than the PreF algorithm, which had worse results than the baseline algorithms. In fact, besides the PreF’s feature of filtering ratings in a preliminary way before its model training, it also differs from the PostF algorithm in relation to the use of information about item categories, once that the PostF uses a category preference tensor in its recommendation process, whereas the PreF does not. For alleviating this disparity, we could combine both algorithms in order to try having the best of their features in a single hybrid algorithm, for example (as described in Section 3.3.1.4). By varying the target domain in the two datasets used in the experiments, we studied the impact of the density of the target domain data in comparison to the density of the source domain data. In both cases, we have seen that the addition of ratings from a source domain improved the recommendation quality of the cross-domain based algorithms, independently of its amount of ratings in relation to the target domain. In addition, even for domains less related among themselves (Book and Music), we could see a improvement on the recommendation quality of the cross-domain based algorithms. Considering distinct user overlap levels (10%, 50% and 100%), we could see in the experiments that the proposed algorithms had a better recommendation quality as the user overlap level was higher. For the PreF algorithm, more user overlap may significate more ratings in filtered contexts, expanding the similarities among users in these contexts, whereas for the PostF, more user overlap level may expand the category preference tensor, with more contextual information about item category preferences of users. On the other hand, the baseline algorithms, especially the NNUserNgbr-transClosure, had a similar performance independently of the user overlap level. As mentioned in Section 5.1.1, the proposed algorithms used the baseline ones 220 Chapter 5. CD-CARS Evaluation with the same recommendation settings (e.g. n=475 for the CF-based algorithm as base). However, note that more tests can be done in order to achieve a “optimal” setting for each algorithm, especially in the PreF algorithm, which uses a contextual sub-dataset with smaller amount of data in relation to other algorithms. Besides, the PostF’s threshold can be adjusted to an “optimal” value for other datasets and its minimal rating value for an item to be considered “good” can also be changed depending on the datasets, as described in Section 5.1.1. For the two datasets used in the experiments, we considered this value as “four” in a five-star scale. However, it may be possible that for other datasets the ratings from the contextual user-rating tensors can have different scales or forms in distinct domains. For example, ratings of music could be represented as a binary form such as “Like” and “Dislike” while the ratings of movie and books could be represented, respectively, by five-star and ten-star scales. As mentioned before, the base recommendation algorithms have to deal with this issue. For instance, an algorithm could normalize the different scales from ratings among distinct domains (SANTOS et al., 2012). Due to the lack of publicly real datasets available with cross-domain and contextual information, a feasible alternative to our experiments without having to produce contextual synthetic data was extract contextual information implicitly or by inference for the three contextual dimensions used in the experiments. However, other contextual dimensions still can be extracted from the user reviews, as for example, the Task dimension. In addition, other contextual attributes of the same contextual dimensions can be used in order to verify the recommendation quality of the proposed algorithms (e.g. by using “countries” instead of “cities” in the Location dimension). Finally, other important aspect of recommender systems is the execution per- formance. Although we did not evaluate the proposed CD-CARS by considering this aspect, intuitively, we can say that the PreF algorithm may be capable of recommending items consuming less resources (e.g. time and memory consumption) than the baseline algorithms, once that the PreF uses these algorithms as base for a small set of ratings in comparison to them. On the other hand, the PostF algorithm may demand more resources, since it initially uses the baseline algorithms as base, and then, applies an addition step to them by using an additional data structure (category preference tensor). 5.3 Final Remarks In this chapter, we presented and discussed experimental evaluations of two proposed CD-CARS algorithms in comparison to cross-domain CF-based ones. The experimental evaluation was made by considering two distinct datasets (described in Section 4.1.3) with three different contextual dimensions (Temporal, Location and Companion), target domains (Television, Music and Book) and user overlap levels (10%, 50%, and 100%). 5.3. Final Remarks 221 The algorithms were evaluated regarding their predictive and classification performances, which were analyzed by means of statistical significance tests. Finally, the conclusions and future works of this thesis are described in the next chapter. 222 6 Conclusion In this thesis, we have found that context-aware techniques can be used in order to improve the accuracy of cross-domain recommendations. A traditional cross-domain CF-based algorithm provided better recommendations when used in combination with the implemented CD-CARS algorithms (Pre-Filtering and Post-Filtering). By considering contextual information from three dimensions (Temporal, Location and Companion), experimental evaluations conducted in two real datasets, one with two more related domains (Book and Television) and another with two less related domains (Book and Music), showed that generating predictions exploiting knowledge from a source domain improved predictive and classification performances in the target domain. For both datasets, we made experiments by swapping source and target domains and, regardless of these domains evaluated as source or target, the proposed algorithms achieved better results in comparison to the baseline ones, especially by using contextual information from the Temporal and Location dimensions. For the Companion contextual dimension, only one of the implemented algorithms had a good predictive performance, whereas its classification performance was not as good. As discussed in Chapter 5, the low quality and quantity of contextual information from that contextual dimension may have influenced on the negative classification performance, mainly for the Pre-Filtering algorithm, which also had a bad predictive performance by taking into account the Companion dimension. With respect to the user overlap level variation (10%, 50% and 100%), we conclude that it influenced the proposed algorithms that, in general, are more accurate as higher is the user overlap level. Finally, through a novel approach, we expect that the findings from this study con- tribute to the cross-domain RS area towards future research in cross-domain context-aware recommendations. In the following sections, we describe the contributions, limitations and future works of this thesis. 6.1 Contributions One of the contributions of this thesis is the novel study about the successful integration of two emergent and relevant approaches of the recommender system (RS) area: cross-domain and context-awareness. This integration can lead to further research in the area in order to improve the quality of recommender systems. While the proposed CD-CARS algorithms address, mainly, the accuracy aspect of RSs quality, the cross-domain collaborative filtering algorithms, which are adopted in combination with the proposed 6.1. Contributions 223 ones, may address other aspects such as cold-start and sparsity. Therefore, the proposed CD-CARS takes the best aspects of those RS approaches into account. Other contributions of this thesis are: • The formalization of the cross-domain context-aware recommendation problem from the survey of two relevant research fields: cross-domain and context-aware RS. For that, we considered user ratings as a function of three dimensions (ADOMAVICIUS; TUZHILIN, 2015): User, Item and Context. Thus, the user ratings can be stored in multidimensional user-rating-context tensors for each item domain (e.g. books, movies, music, among others). In addition, it is necessary that there are user and contextual overlap among distinct domains. At least, the proposed contextual feature modelling is based on the “Key-Value” model, since it is simple and relatively easy to implement and use (VIEIRA; TEDESCO; SALGADO, 2009)(BETTINI et al., 2010); • Proposal of novel CD-CARS algorithms based on three distinct and systematic paradigms of context-aware recommendation (Pre-Filtering, Post-Filtering and Modelling), which were chosen rather than ad-hoc context-aware approaches. One of the advantages of the proposed CD-CARS algorithms is the possibility of using traditional single-domain and cross-domain CF-based algorithms as a base algorithm, which is used in combination with the proposed ones. In addition, the proposed algorithms can be directly combined such as Pre-Filtering and Post-Filtering, or Modelling and Post-Filtering, generating hybrid versions of the proposed CD-CARS algorithms; • Providing systematic CD-CARS algorithms that can be useful to recommend items for several domains (e.g. books, music, movies, etc.) in a simple way, since little information about users and items is required. It allows generating cross selling or bundle recommendations for items from multiple domains (e.g. the recommendation of a music accompanied of a movie to watch or a book to read); • Provision of two real datasets for evaluating CD-CARS1, taking into account dif- ferent domains and contextual information. One of them for evaluating CD-CARS algorithms in two more related domains (Book and Television) and another con- sidering two less related domains (Book and Music). These datasets were adapted and extracted from (LESKOVEC; ADAMIC; HUBERMAN, 2007), which contains ratings (five-star scale), product metadata and review information about different Amazon products2. In addition to these data, we included contextual information 1 https://github.com/douglasveras/cd-cars-datasets 2 https://snap.stanford.edu/data/amazon-meta.html 224 Chapter 6. Conclusion regarding three contextual dimensions: Temporal, Location and Companion, respec- tively, inferred from the ratings’ dates, users’ static addresses (obtained from their account on Amazon), and users’ rating reviews. 6.2 Limitations The main limitations of this thesis are: • Absence of a mechanism in the proposed CD-CARS to handle ratings from the contextual user-rating tensors that have different scales or forms in distinct domains. For example, ratings of music could be represented as a binary form such as “Like” or “Dislike” while the ratings of movies and books may be represented, respectively, by five-star or ten-star scales. As mentioned in the CD-CARS proposal, the proposed CD-CARS algorithms let the responsibility of dealing with this issue with the cross- domain algorithms used as base. For instance, the base algorithm could normalize the different scales from ratings among distinct domains (SANTOS et al., 2012). Another solution is to normalize the ratings in these domains before the recommendation process. • A deeper experimentation in the combination of different contextual dimensions (Temporal, Location and Companion) could be made. As previously mentioned, we just performed experiments in the combination of Temporal and Location dimensions, since the quality of the Companion dimension is low. • Lack of realization of a concrete CD-CARS for evaluating the satisfaction of real users, and its capabilities in terms of execution time and memory required. 6.3 Lines for Further Work The CD-CARS proposed in this thesis allows further investigation in multiple research lines such as: 1. Improvement of the implemented algorithms and implementation of other proposed CD-CARS algorithms such as Modelling or the combination between Pre-Filtering and Post-Filtering, for example, as well as other state-of-the-art CD-CFRS algorithms, which could be based on different cross-domain approaches in order to expand the findings of this thesis (e.g. Linking and transferring knowledge instead of the Aggregating knowledge, adopted in this thesis). For instance, we could compare the proposed CD-CARS with algorithms from the Linking and transferring knowledge approach, such as: CodeBook Transfer (CBT)(LI; YANG; XUE, 2009a), (GAO et al., 2013), among others; 6.3. Lines for Further Work 225 2. Exploration of further contextual information in different domains: a) Other algorithms could be used, or proposed, to infer more precise contextual information from user reviews (e.g. use of supervised text mining techniques for a better inference quality of the Companion contextual dimension) (CHEN; CHEN, 2015)(DOMINGUES et al., 2014)(LAHLOU et al., 2013), which may lead to a better recommendation quality and information gain calculated; b) Creation of techniques for inference of other contextual dimensions (e.g. Task, Mood, etc.) from user reviews, independently on the item domain (e.g. music, books, games, etc.); c) Building mechanisms to explicitly collect contextual information from user ratings over multiple domains; d) Making the contextual modelling (adopted in the CD-CARS) more representa- tive in order to describe semantic relations between contexts and domains, for example. 3. CD-CARS evaluation: a) Building a concrete CD-CARS for evaluating the satisfaction of real users taking into account the use of real resources (e.g. time, memory, etc.); b) Development of a benchmark for a rapid and efficient evaluation of the CD- CARS; c) Improving the current evaluation methodology used in this thesis taking into account different evaluation metrics (e.g. Breese score, Normalized Discounted Cumulative Gain, etc.) and partitioning of training and test sets, for example; d) Investigating and providing data mining techniques in order to select the most relevant contextual dimensions, attributes and values (or their combination) before performing recommendation or evaluation, once that the verification of all possible situations is costly; e) Combining other domains (e.g. Music and Television) or contextual dimensions (e.g. Location and Companion), as well as evaluating the algorithm performances by considering different user overlap levels (e.g. 0%, 25% and 75%); f) To verify the impact of the amount of ratings with contextual information in distinct contextual dimensions (“contextual sensitivity”), by reducing the ratings from the Temporal dimension for that it have a similar number of ratings in comparison to the Location dimension, for example; g) In order to verify the use of association rules by the PostF algorithm, we can made an evaluation for specific situations where the use of these rules is required by it. For instance, when a user receives a recommendation in the 226 Chapter 6. Conclusion target domain and the category preferences tensor (used by the PostF) does not have information about his/her rated item categories in that domain yet. In this case, the association rules are used for enhancing the category preferences tensor, as described in Section 3.3.1.2. 4. Development of CD-CARS applications. Since the proposed algorithms are not domain-specific, a myriad of cross-domain (or cross-selling) applications can be developed for multiple domains (e.g. a recommender system on the TV that, beyond TV shows, also could recommend books, sites, or other relevant information to the TV programs) (FERRAZ; SILVA; SILVA, 2015). Besides, these applications could be developed to be accessed in a ubiquitous way, depending on the users’ contexts and their domains of interest (e.g. when a user watches TV in his/her bedroom, then the CD-CARS application could recommend movies, whereas when the user is in the living room with his/her friends, then the same application could recommend music for them). 227 References ABBAR, S.; BOUZEGHOUB, M.; LOPEZ, S. Context-aware recommender systems: A service-oriented approach. In: VLDB PersDB workshop. [S.l.: s.n.], 2009. p. 1–6. Cited on page 36. ABEL, F. et al. Analyzing cross-system user modeling on the social web. In: Web Engineering. [S.l.]: Springer, 2011. p. 28–43. Cited on page 42. ABEL, F. et al. Cross-system user modeling and personalization on the social web. User Modeling and User-Adapted Interaction, Springer, v. 23, n. 2-3, p. 169–209, 2013. Cited 2 many times on page 46 and 47. ABOWD, G. D. et al. Towards a better understanding of context and context-awareness. In: SPRINGER. Handheld and ubiquitous computing. [S.l.], 1999. p. 304–307. Cited on page 34. ADOMAVICIUS, G. et al. Incorporating contextual information in recommender systems using a multidimensional approach. ACM Transactions on Information Systems (TOIS), ACM, v. 23, n. 1, p. 103–145, 2005. Cited 7 many times on page 27, 47, 51, 54, 70, 80, and 81. ADOMAVICIUS, G.; TUZHILIN, A. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. Knowledge and Data Engineering, IEEE Transactions on, IEEE, v. 17, n. 6, p. 734–749, 2005. Cited 4 many times on page 23, 25, 32, and 33. ADOMAVICIUS, G.; TUZHILIN, A. Context-aware recommender systems. In: Recommender systems handbook (Second Edition). [S.l.]: Springer, 2015. p. 191–226. Cited 20 many times on page 9, 26, 27, 29, 30, 47, 48, 49, 51, 52, 53, 54, 55, 56, 66, 68, 70, 74, 82, and 223. AGRAWAL, R.; IMIELIŃSKI, T.; SWAMI, A. Mining association rules between sets of items in large databases. In: ACM. ACM SIGMOD Record. [S.l.], 1993. v. 22, n. 2, p. 207–216. Cited 2 many times on page 79 and 121. ALHAMID, M. et al. Recam: a collaborative context-aware framework for multimedia recommendations in an ambient intelligence environment. Multimedia Systems, Springer Berlin Heidelberg, online, p. 1–15, 2015. ISSN 0942-4962. Disponível em: <http://dx.doi.org/10.1007/s00530-015-0469-2>. Cited on page 55. AMATRIAIN, X. et al. Data mining methods for recommender systems. In: Recommender Systems Handbook. [S.l.]: Springer, 2011. p. 39–71. Cited on page 60. ANAND, S. S.; MOBASHER, B. Contextual recommendation. [S.l.]: Springer, 2006. Cited on page 53. ANSARI, A.; ESSEGAIER, S.; KOHLI, R. Internet recommendation systems. Journal of Marketing research, American Marketing Association, v. 37, n. 3, p. 363–375, 2000. Cited on page 54. http://dx.doi.org/10.1007/s00530-015-0469-2 228 References AZAK, M. CrosSing: A framework to develop knowledge-based recommenders in cross domains. Dissertação (Mestrado) — MIDDLE EAST TECHNICAL UNIVERSITY, 2010. Cited 2 many times on page 25 and 44. BALTRUNAS, L. et al. Incarmusic: Context-aware music recommendations in a car. In: SPRINGER. EC-Web. [S.l.], 2011. v. 11, p. 89–100. Cited on page 47. BALTRUNAS, L.; LUDWIG, B.; RICCI, F. Matrix factorization techniques for context aware recommendation. In: ACM. Proceedings of the fifth ACM conference on Recommender systems. [S.l.], 2011. p. 301–304. Cited 2 many times on page 54 and 81. BALTRUNAS, L.; MAKCINSKAS, T.; RICCI, F. Group recommendations with rank aggregation and collaborative filtering. In: Proceedings of the fourth ACM conference on Recommender systems. [S.l.: s.n.], 2010. p. 119–126. ISBN 9781605589060. Cited on page 37. BAUMAN, K.; TUZHILIN, A. Discovering contextual information from user reviews for recommendation purposes. In: CBRecSys. [S.l.: s.n.], 2014. p. 1–8. Cited 4 many times on page 98, 99, 100, and 101. BAZIRE, M.; BRÉZILLON, P. Understanding context before using it. In: Modeling and using context. [S.l.]: Springer, 2005. p. 29–40. Cited on page 48. BELL, R.; KOREN, Y.; VOLINSKY, C. Modeling relationships at multiple scales to improve accuracy of large recommender systems. In: ACM. Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining. [S.l.], 2007. p. 95–104. Cited on page 87. BENNETT, J.; LANNING, S. The netflix prize. In: Proceedings of KDD cup and workshop. [S.l.: s.n.], 2007. v. 2007, p. 35–36. Cited 2 many times on page 32 and 38. BERKOVSKY, S.; KUFLIK, T.; RICCI, F. Cross-domain mediation in collaborative filtering. In: User Modeling 2007. [S.l.]: Springer, 2007. p. 355–359. Cited 4 many times on page 42, 44, 57, and 58. BERKOVSKY, S.; KUFLIK, T.; RICCI, F. Mediation of user models for enhanced personalization in recommender systems. User Modeling and User-Adapted Interaction, Springer, v. 18, n. 3, p. 245–286, 2008. Cited on page 47. BERNERS-LEE, T.; HENDLER, J. The semantic web. a new form of web content that is meaningful to computers will unleash a revolution of new possibilities. Scientific American Magazine, v. 1, p. 34–43, maio 2001. Cited on page 35. BETTINI, C. et al. A survey of context modelling and reasoning techniques. Pervasive and Mobile Computing, Elsevier, v. 6, n. 2, p. 161–180, 2010. Cited 3 many times on page 49, 90, and 223. BEZERRA, B. L.; CARVALHO, F. d. A. de. A symbolic approach for content-based information filtering. Information Processing Letters, Elsevier, v. 92, n. 1, p. 45–52, 2004. Cited on page 86. BEZERRA, B. L. D.; CARVALHO, F. D. A. T. D. Symbolic data analysis tools for recommendation systems. Knowledge and Information Systems, Springer, v. 26, n. 3, p. 385–418, 2011. Cited on page 86. References 229 BLANCO-FERNÁNDEZ, Y. et al. Tripfromtv+: Exploiting social networks to arrange cut-price touristic packages. In: IEEE. IEEE International Conference on Consumer Electronics (ICCE). [S.l.], 2011. p. 223–224. Cited 3 many times on page 62, 63, and 67. BLANCO-FERNÁNDEZ, Y. et al. Exploiting digital tv users’ preferences in a tourism recommender system based on semantic reasoning. Consumer Electronics, IEEE Transactions on, IEEE, v. 56, n. 2, p. 904–912, 2010. Cited 3 many times on page 62, 63, and 67. BLANCO-FERNÁNDEZ, Y. et al. Tripfromtv+: targeting personalized tourism to interactive digital tv viewers by social networking and semantic reasoning. IEEE Transactions on Consumer Electronics, IEEE, v. 57, n. 2, p. 953–961, 2011. Cited 7 many times on page 29, 35, 61, 62, 63, 64, and 67. BLANCO-FERNÁNDEZ, Y.; PAZOS-ARIAS, J. An MHP framework to provide intelligent personalized recommendations about digital TV contents. Software: Practice and Experience, v. 38, n. October 2007, p. 925–960, 2008. Cited on page 35. BLANCO-FERNáNDEZ, Y. et al. Exploiting synergies between semantic reasoning and personalization strategies in intelligent recommender systems: A case study. Journal of Systems and Software, Elsevier Inc., v. 81, n. 12, p. 2371–2385, dez. 2008. ISSN 01641212. Cited on page 37. BLEI, D. M.; NG, A. Y.; JORDAN, M. I. Latent dirichlet allocation. the Journal of machine Learning research, JMLR. org, v. 3, p. 993–1022, 2003. Cited on page 99. BOUNEFFOUF, D. Situation-aware approach to improve context-based recommender system. arXiv preprint arXiv:1303.0481, 2013. Cited on page 52. BOURKE, S.; MCCARTHY, K.; SMYTH, B. Power to the people: exploring neighbourhood formations in social recommender system. In: Proceedings of the fifth ACM conference on Recommender systems. [S.l.: s.n.], 2011. p. 337–340. ISBN 9781450306836. Cited on page 34. BOYTSOV, A. et al. Situation awareness meets ontologies: A context spaces case study. In: SPRINGER. International and Interdisciplinary Conference on Modeling and Using Context. [S.l.], 2015. p. 3–17. Cited on page 52. BRAUNHOFER, M.; KAMINSKAS, M.; RICCI, F. Location-aware music recommendation. International Journal of Multimedia Information Retrieval, Springer, v. 2, n. 1, p. 31–44, 2013. Cited 5 many times on page 61, 62, 63, 65, and 67. BREESE, J. S.; HECKERMAN, D.; KADIE, C. Empirical analysis of predictive algorithms for collaborative filtering. In: MORGAN KAUFMANN PUBLISHERS INC. Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence. [S.l.], 1998. p. 43–52. Cited on page 37. BRÉZILLON, P. Context modeling: Task model and practice model. In: Modeling and Using Context. [S.l.]: Springer, 2007. p. 122–135. Cited 2 many times on page 49 and 52. BRUSILOVSKY, P.; KOBSA, A.; NEJDL, W. The adaptive web: methods and strategies of web personalization. [S.l.]: Springer, 2007. v. 4321. Cited on page 35. 230 References BURKE, R. Hybrid recommender systems: Survey and experiments. User modeling and user-adapted interaction, Springer, v. 12, n. 4, p. 331–370, 2002. Cited on page 32. BURKE, R. Hybrid web recommender systems. In: The adaptive web. [S.l.]: Springer, 2007. p. 377–408. Cited 2 many times on page 32 and 33. CAMPOS, P. G.; DÍEZ, F.; CANTADOR, I. Time-aware recommender systems: a comprehensive survey and analysis of existing evaluation protocols. User Modeling and User-Adapted Interaction, Springer, v. 24, n. 1-2, p. 67–119, 2014. Cited on page 56. CANTADOR, I.; CREMONESI, P. Tutorial on cross-domain recommender systems. In: ACM. Proceedings of the 8th ACM Conference on Recommender systems. [S.l.], 2014. p. 401–402. Cited on page 68. CANTADOR, I. et al. Cross-domain recommender systems. In: Recommender Systems Handbook. [S.l.]: Springer, 2015. p. 919–959. Cited 13 many times on page 9, 26, 30, 38, 39, 40, 41, 43, 44, 45, 46, 47, and 57. CAO, B.; LIU, N. N.; YANG, Q. Transfer learning for collective link prediction in multiple heterogenous domains. In: Proceedings of the 27th International Conference on Machine Learning (ICML-10). [S.l.: s.n.], 2010. p. 159–166. Cited 2 many times on page 38 and 47. CARMAGNOLA, F.; CENA, F. User identification for cross-system personalisation. Information Sciences, Elsevier, v. 179, n. 1, p. 16–32, 2009. Cited on page 40. CARMAGNOLA, F.; CENA, F.; GENA, C. User model interoperability: a survey. User Modeling and User-Adapted Interaction, Springer, v. 21, n. 3, p. 285–331, 2011. Cited on page 40. CHAARI, T. et al. A comprehensive approach to model and use context for adapting applications in pervasive environments. Journal of Systems and Software, Elsevier, v. 80, n. 12, p. 1973–1992, 2007. Cited on page 49. CHATTERJEE, S.; HADI, A. S. Regression analysis by example. [S.l.]: John Wiley & Sons, 2015. Cited on page 52. CHEN, G.; CHEN, L. Augmenting service recommender systems by incorporating contextual opinions from user reviews. User Modeling and User-Adapted Interaction, Springer, v. 25, n. 3, p. 295–329, 2015. Cited on page 225. CHUNG, R.; SUNDARAM, D.; SRINIVASAN, A. Integrated personal recommender systems. In: ACM. Proceedings of the ninth international conference on Electronic commerce. [S.l.], 2007. p. 65–74. Cited on page 44. CHURCH, K. et al. Mobile information access: A study of emerging search behavior on the mobile internet. ACM Transactions on the Web (TWEB), ACM, v. 1, n. 1, p. 4, 2007. Cited on page 47. COLOMBO-MENDOZA, L. O. et al. Recommetz: A context-aware knowledge-based mobile recommender system for movie showtimes. Expert Systems with Applications, Elsevier, v. 42, n. 3, p. 1202–1222, 2015. Cited on page 51. References 231 CREMONESI, P.; GARZOTTO, F.; TURRIN, R. Investigating the Persuasion Potential of Recommender Systems from a Quality Perspective. ACM Transactions on Interactive Intelligent Systems, v. 2, n. 2, p. 1–41, jun. 2012. ISSN 21606455. Cited on page 33. CREMONESI, P.; KOREN, Y.; TURRIN, R. Performance of recommender algorithms on top-n recommendation tasks. In: ACM. Proceedings of the fourth ACM conference on Recommender systems. [S.l.], 2010. p. 39–46. Cited 3 many times on page 59, 129, and 130. CREMONESI, P.; TRIPODI, A.; TURRIN, R. Cross-domain recommender systems. In: IEEE. IEEE 11th International Conference on Data Mining Workshops (ICDMW). [S.l.], 2011. p. 496–503. Cited 20 many times on page 9, 24, 25, 41, 42, 43, 47, 58, 59, 61, 68, 84, 86, 88, 89, 90, 123, 124, 129, and 130. CREMONESI, P.; TURRIN, R. Controlling Consistency in Top-N Recommender Systems. In: IEEE International Conference on Data Mining Workshops. [S.l.]: Ieee, 2010. p. 919–926. ISBN 978-1-4244-9244-2. Cited on page 37. DAS, A. S. et al. Google news personalization: scalable online collaborative filtering. In: ACM. Proceedings of the 16th international conference on World Wide Web. [S.l.], 2007. p. 271–280. Cited on page 32. DEY, A. K.; ABOWD, G. D.; SALBER, D. A conceptual framework and a toolkit for supporting the rapid prototyping of context-aware applications. Human-computer interaction, L. Erlbaum Associates Inc., v. 16, n. 2, p. 97–166, 2001. Cited on page 48. DIDAY, E.; BOCK, H.-H. Analysis of symbolic data: Exploratory methods for extracting statistical information from complex data. [S.l.]: Springer-Verlag, 2000. Cited on page 86. DOMINGUES, M. A. et al. Exploiting text mining techniques for contextual recommendations. In: IEEE. Web Intelligence (WI) and Intelligent Agent Technologies (IAT), 2014 IEEE/WIC/ACM International Joint Conferences on. [S.l.], 2014. v. 2, p. 210–217. Cited on page 225. DOURISH, P. What we talk about when we talk about context. Personal and ubiquitous computing, Springer, v. 8, n. 1, p. 19–30, 2004. Cited on page 53. ENRICH, M.; BRAUNHOFER, M.; RICCI, F. Cold-start management with cross-domain collaborative filtering and tags. In: E-Commerce and Web Technologies. [S.l.]: Springer, 2013. p. 101–112. Cited 2 many times on page 39 and 44. EYKE, J. W. Temporal Problems, with a Focus on Mood, in Music Recommendation Within Last. FM. Tese (Doutorado) — University of Sheffield, Department of Information Studies, 2009. Cited 2 many times on page 32 and 38. FERNÁNDEZ-TOBÍAS, I. et al. Cross-domain recommender systems: A survey of the state of the art. In: Spanish Conference on Information Retrieval. [S.l.: s.n.], 2012. Cited 10 many times on page 24, 25, 26, 30, 38, 43, 61, 68, 90, and 131. FERRAZ, C. A.; SILVA, D. V. e; SILVA, J. S. da. A collaborative tv-internet application model to enrich tv viewing experience in a pervasive way. In: IEEE. IEEE International Conference on Pervasive Computing and Communication Workshops (PerCom Workshops). [S.l.], 2015. p. 148–153. Cited on page 226. 232 References FREYNE, J.; BERKOVSKY, S. Evaluating recommender systems for supportive technologies. In: User Modeling and Adaptation for Daily Routines. [S.l.]: Springer, 2013. p. 195–217. Cited on page 45. GAO, S. et al. Cross-domain recommendation via cluster-level latent factor model. In: Machine Learning and Knowledge Discovery in Databases. [S.l.]: Springer, 2013. p. 161–176. Cited 3 many times on page 39, 44, and 224. GIVON, S.; LAVRENKO, V. Predicting social-tags for cold start book recommendations. In: ACM. Proceedings of the third ACM conference on Recommender systems. [S.l.], 2009. p. 333–336. Cited on page 44. GOGA, O. et al. Exploiting innocuous activity for correlating users across sites. In: INTERNATIONAL WORLD WIDE WEB CONFERENCES STEERING COMMITTEE. Proceedings of the 22nd international conference on World Wide Web. [S.l.], 2013. p. 447–458. Cited on page 46. GU, T.; PUNG, H. K.; ZHANG, D. Q. A service-oriented middleware for building context-aware services. Journal of Network and computer applications, Elsevier, v. 28, n. 1, p. 1–18, 2005. Cited on page 49. GUPTA, K. M. Taxonomic conversational case-based reasoning. In: Case-Based Reasoning Research and Development. [S.l.]: Springer, 2001. p. 219–233. Cited on page 64. GUYON, I.; ELISSEEFF, A. An introduction to variable and feature selection. The Journal of Machine Learning Research, JMLR. org, v. 3, p. 1157–1182, 2003. Cited on page 52. HALL, M. et al. The weka data mining software: an update. ACM SIGKDD explorations newsletter, ACM, v. 11, n. 1, p. 10–18, 2009. Cited 2 many times on page 103 and 126. HAN, X. et al. Alike people, alike interests? inferring interest similarity in online social networks. Decision Support Systems, Elsevier, v. 69, p. 92–106, 2015. Cited on page 34. HENRICKSEN, K.; INDULSKA, J. Developing context-aware pervasive computing applications: Models and approach. Pervasive and mobile computing, Elsevier, v. 2, n. 1, p. 37–64, 2006. Cited on page 49. HERLOCKER, J. L. et al. Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems (TOIS), ACM, v. 22, n. 1, p. 5–53, 2004. Cited on page 37. HIDASI, B.; TIKK, D. Fast als-based tensor factorization for context-aware recommendation from implicit feedback. In: Machine Learning and Knowledge Discovery in Databases. [S.l.]: Springer, 2012. p. 67–82. Cited 2 many times on page 54 and 82. HILL, W. et al. Recommending and evaluating choices in a virtual community of use. In: ACM PRESS/ADDISON-WESLEY PUBLISHING CO. Proceedings of the SIGCHI conference on Human factors in computing systems. [S.l.], 1995. p. 194–201. Cited on page 23. HIPP, J.; GÜNTZER, U.; NAKHAEIZADEH, G. Algorithms for association rule mining—a general survey and comparison. ACM sigkdd explorations newsletter, ACM, v. 2, n. 1, p. 58–64, 2000. Cited on page 79. References 233 HOPFGARTNER, F.; JOSE, J. Semantic user profiling techniques for personalised multimedia recommendation. Multimedia systems, v. 16, p. 255–274, 2010. Cited on page 37. HU, L. et al. Personalized recommendation via cross-domain triadic factorization. In: INTERNATIONAL WORLD WIDE WEB CONFERENCES STEERING COMMITTEE. Proceedings of the 22nd international conference on World Wide Web. [S.l.], 2013. p. 595–606. Cited on page 39. HU, Y.; KOREN, Y.; VOLINSKY, C. Collaborative filtering for implicit feedback datasets. In: IEEE. Eighth IEEE International Conference on Data Mining (ICDM). [S.l.], 2008. p. 263–272. Cited on page 36. JADIDI, O.; FIROUZI, F.; BAGLIERY, E. Topsis method for supplier selection problem. World Academy of Science, Engineering and Technology, Citeseer, v. 47, p. 956–958, 2010. Cited on page 64. JAIN, A. K. Data clustering: 50 years beyond k-means. Pattern recognition letters, Elsevier, v. 31, n. 8, p. 651–666, 2010. Cited on page 99. JAIN, P.; KUMARAGURU, P.; JOSHI, A. @ i seek’fb. me’: Identifying users across multiple online social networks. In: INTERNATIONAL WORLD WIDE WEB CONFERENCES STEERING COMMITTEE. Proceedings of the 22nd international conference on World Wide Web companion. [S.l.], 2013. p. 1259–1268. Cited on page 46. JI, K.; SHEN, H. Making recommendations from top-n user-item subgroups. Neurocomputing, Elsevier, v. 165, p. 228–237, 2015. Cited 5 many times on page 61, 62, 63, 66, and 67. JOJIC, O.; SHUKLA, M.; BHOSAREKAR, N. A probabilistic definition of item similarity. In: Proceedings of the fifth ACM conference on Recommender systems. New York, New York, USA: ACM Press, 2011. p. 229–236. ISBN 9781450306836. Cited on page 37. KAMAHARA, J. et al. A Community-Based Recommendation System to Reveal Unexpected Interests. In: 11th International Multimedia Modelling Conference. [S.l.]: IEEE, 2005. p. 433–438. ISBN 0-7695-2164-9. Cited on page 34. KAMINSKAS, M. et al. Knowledge-based identification of music suited for places of interest. Information Technology & Tourism, Springer, v. 14, n. 1, p. 73–95, 2014. Cited 9 many times on page 29, 30, 51, 55, 61, 62, 63, 65, and 67. KAMINSKAS, M.; RICCI, F. Contextual music information retrieval and recommendation: State of the art and challenges. Computer Science Review, Elsevier, v. 6, n. 2, p. 89–119, 2012. Cited on page 47. KARATZOGLOU, A. et al. Multiverse recommendation: n-dimensional tensor factorization for context-aware collaborative filtering. In: ACM. Proceedings of the fourth ACM conference on Recommender systems. [S.l.], 2010. p. 79–86. Cited 2 many times on page 54 and 82. KIM, S.; YOON, Y. Recommendation system for sharing economy based on multidimensional trust model. Multimedia Tools and Applications, Springer, p. 1–14, 2014. Cited 3 many times on page 54, 82, and 88. 234 References KOREN, Y. Factorization meets the neighborhood: a multifaceted collaborative filtering model. In: ACM. Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. [S.l.], 2008. p. 426–434. Cited 2 many times on page 53 and 87. LAHLOU, F. Z. et al. A text classification based method for context extraction from online reviews. In: IEEE. 8th International Conference on Intelligent Systems: Theories and Applications (SITA). [S.l.], 2013. p. 1–5. Cited on page 225. LAROSE, D. T. k-nearest neighbor algorithm. In: . Discovering Knowledge in Data. John Wiley and Sons, Inc., 2005. p. 90–106. ISBN 9780471687542. Disponível em: <http://dx.doi.org/10.1002/0471687545.ch5>. Cited on page 60. LEE, H.; KWON, J. Personalized tv contents recommender system using collaborative context tagging-based user’s preference prediction technique. International Journal of Multimedia & Ubiquitous Engineering, v. 9, n. 5, p. 231–240, 2014. Cited on page 51. LEE, H. J.; PARK, S. J. Moners: A news recommender for the mobile web. Expert Systems with Applications, Elsevier, v. 32, n. 1, p. 143–150, 2007. Cited on page 47. LEE, W.; YANG, T.-H. Personalizing information appliances: a multi-agent framework for TV programme recommendations. Expert Systems with Applications, v. 25, n. 3, p. 331–341, out. 2003. ISSN 09574174. Cited on page 37. LEKAKOS, G.; CARAVELAS, P. A hybrid approach for movie recommendation. Multimedia Tools and Applications, v. 36, n. December 2006, p. 55–70, 2008. Cited on page 37. LEKAKOS, G.; GIAGLIS, G. A Lifestyle‚ÄêBased Approach for Delivering Personalized Advertisements in Digital Interactive Television. Journal of Computer-Mediated Communication, v. 6, n. 1, p. 00–00, 2004. Cited on page 37. LESKOVEC, J.; ADAMIC, L. A.; HUBERMAN, B. A. The dynamics of viral marketing. ACM Transactions on the Web (TWEB), ACM, v. 1, n. 1, p. 5, 2007. Cited 3 many times on page 61, 92, and 223. LI, B.; YANG, Q.; XUE, X. Can movies and books collaborate? cross-domain collaborative filtering for sparsity reduction. In: IJCAI. [S.l.: s.n.], 2009. v. 9, p. 2052–2057. Cited 2 many times on page 66 and 224. LI, B.; YANG, Q.; XUE, X. Transfer learning for collaborative filtering via a rating-matrix generative model. In: ACM. Proceedings of the 26th Annual International Conference on Machine Learning. [S.l.], 2009. p. 617–624. Cited 3 many times on page 44, 46, and 47. LI, L. et al. A contextual-bandit approach to personalized news article recommendation. In: ACM. Proceedings of the 19th international conference on World wide web. [S.l.], 2010. p. 661–670. Cited on page 66. LIAW, A.; WIENER, M. Classification and regression by randomforest. R news, v. 2, n. 3, p. 18–22, 2002. Cited on page 60. LINDEN, G.; SMITH, B.; YORK, J. Amazon.com recommendations: Item-to-item collaborative filtering. Internet Computing, IEEE, IEEE, v. 7, n. 1, p. 76–80, 2003. Cited on page 32. http://dx.doi.org/10.1002/0471687545.ch5 References 235 LIU, H.; MOTODA, H. Feature Selection for Knowledge Discovery and Data Mining. [S.l.]: Springer Science & Business Media, 1998. Cited on page 73. LIU, H.; MOTODA, H. Feature selection for knowledge discovery and data mining. [S.l.]: Springer Science & Business Media, 2012. v. 454. Cited on page 52. LONI, B. et al. Cross-domain collaborative filtering with factorization machines. In: Advances in Information Retrieval. [S.l.]: Springer, 2014. p. 656–661. Cited 3 many times on page 39, 58, and 60. LóPEZ-NORES, M. et al. MiSPOT: dynamic product placement for digital TV through MPEG-4 processing and semantic reasoning. Knowledge and Information Systems, v. 22, n. 1, p. 101–128, mar. 2009. ISSN 0219-1377. Cited on page 37. LOVÁSZ, L. et al. Random walks on graphs: A survey. Combinatorics, Paul Erdos is Eighty, v. 2, p. 353–398, 1996. Cited on page 59. LOW, Y.; AGARWAL, D.; SMOLA, A. J. Multiple domain user personalization. In: ACM. Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining. [S.l.], 2011. p. 123–131. Cited on page 40. MAHMOOD, T.; RICCI, F.; VENTURINI, A. Improving recommendation effectiveness: Adapting a dialogue strategy in online travel planning. Information Technology & Tourism, Cognizant Communication Corporation, v. 11, n. 4, p. 285–302, 2009. Cited on page 47. MARILLY, E. et al. Community-based applications. Bell Labs Technical Journal, v. 15, n. 4, p. 93–109, mar. 2011. ISSN 10897089. Cited on page 35. MCJONES, P. Eachmovie collaborative filtering data set. DEC Systems Research Center, v. 249, 1997. Cited on page 57. MILLER, G. A. Wordnet: a lexical database for english. Communications of the ACM, ACM, v. 38, n. 11, p. 39–41, 1995. Cited on page 100. MOE, H. H.; AUNG, W. T. Building ontologies for cross-domain recommendation on facial skin problem and related cosmetics. International Journal of Information Technology and Computer Science (IJITCS), v. 6, n. 6, p. 33, 2014. Cited 4 many times on page 61, 62, 63, and 67. MOE, H. H.; AUNG, W. T. Context aware cross-domain based recommendation. In: International Conference on Advances in Engineering and Technology. [S.l.: s.n.], 2014. Cited 7 many times on page 29, 55, 61, 62, 63, 64, and 67. MOE, H. H.; AUNG, W. T. et al. Cross-domain recommendations for personalized semantic services. International Journal of Computer Applications Technology and Research, v. 2, n. 1, p. 72–76, 2013. Cited 4 many times on page 61, 62, 63, and 67. MOON, A. et al. Designing CAMUS based context-awareness for pervasive home environments. In: International Conference on Hybrid Information Technology. [S.l.: s.n.], 2006. v. 1, p. 666–672. ISBN 0769526748. Cited on page 54. MOON, A. et al. Two-step recommendation based personalization for future services. In: International Conference on Advanced Communication Technology. [S.l.: s.n.], 2009. v. 03, p. 2268–2272. ISBN 9788955191394. Cited 2 many times on page 34 and 55. 236 References MORENO, O. et al. Talmud: transfer learning for multiple domains. In: ACM. Proceedings of the 21st ACM international conference on Information and knowledge management. [S.l.], 2012. p. 425–434. Cited 2 many times on page 40 and 41. MUKHERJEE, D. et al. A context-aware recommendation system considering both user preferences and learned behavior. In: 7th International Conference on Information Technology in Asia. [S.l.: s.n.], 2011. p. 1–7. ISBN 9781612841304. Cited on page 36. NAKATSUJI, M. et al. Recommendations over domain specific user graphs. In: ECAI. [S.l.: s.n.], 2010. p. 607–612. Cited 2 many times on page 58 and 59. NETO, B.; FREITAS, R. de. Um processo de software e um modelo ontológico para apoio ao desenvolvimento de aplicações sensíveis a contexto. Tese (Doutorado) — Universidade de São Paulo, 2007. Cited on page 51. OH, S. et al. Comparison of techniques for time aware tv channel recommendation. In: IEEE. Soft Computing and Intelligent Systems (SCIS), 2014 Joint 7th International Conference on and Advanced Intelligent Systems (ISIS), 15th International Symposium on. [S.l.], 2014. p. 989–992. Cited on page 51. OKU, K. et al. Context-aware svm for context-dependent information recommendation. In: IEEE COMPUTER SOCIETY. Proceedings of the 7th international Conference on Mobile Data Management. [S.l.], 2006. p. 109. Cited on page 54. O’SULLIVAN, D.; SMYTH, B.; WILSON, D. Improving the quality of the personalized electronic program guide. User Modeling and User-Adapted Interaction, v. 14, p. 5–36, 2004. Cited on page 37. OWEN, S. et al. Mahout in action. [S.l.]: Manning, 2011. Cited 3 many times on page 109, 111, and 126. PALMISANO, C.; TUZHILIN, A.; GORGOGLIONE, M. Using context to improve predictive modeling of customers in personalization applications. Knowledge and Data Engineering, IEEE Transactions on, IEEE, v. 20, n. 11, p. 1535–1549, 2008. Cited 3 many times on page 48, 51, and 53. PAN, W. et al. Transfer learning in collaborative filtering for sparsity reduction. In: AAAI. [S.l.: s.n.], 2010. v. 10, p. 230–235. Cited 2 many times on page 44 and 47. PAN, W.; XIANG, E. W.; YANG, Q. Transfer learning in collaborative filtering with uncertain ratings. In: AAAI. [S.l.: s.n.], 2012. Cited on page 39. PAN, W.; YANG, Q. Transfer learning in heterogeneous collaborative filtering domains. Artificial intelligence, Elsevier, v. 197, p. 39–55, 2013. Cited 2 many times on page 39 and 45. PANNIELLO, U.; TUZHILIN, A.; GORGOGLIONE, M. Comparing context-aware recommender systems in terms of accuracy and diversity. User Modeling and User-Adapted Interaction, Springer, v. 24, n. 1-2, p. 35–65, 2014. Cited on page 56. PANNIELLO, U. et al. Experimental comparison of pre-vs. post-filtering approaches in context-aware recommender systems. In: ACM. Proceedings of the third ACM conference on Recommender systems. [S.l.], 2009. p. 265–268. Cited on page 54. References 237 PARAMESWARAN, A.; VENETIS, P.; GARCIA-MOLINA, H. Recommendation systems with complex constraints: A course recommendation perspective. ACM Transactions on Information Systems (TOIS), ACM, v. 29, n. 4, p. 20, 2011. Cited on page 64. PARK, D. H. et al. A literature review and classification of recommender systems research. Expert Systems with Applications, Elsevier, v. 39, n. 11, p. 10059–10072, 2012. Cited 2 many times on page 32 and 34. PAZZANI, M. J. A framework for collaborative, content-based and demographic filtering. Artificial Intelligence Review, Springer, v. 13, n. 5-6, p. 393–408, 1999. Cited on page 33. PESSEMIER, T. D.; DOOMS, S.; MARTENS, L. Context-aware recommendations through context and activity recognition in a mobile environment. Multimedia Tools and Applications, Springer, v. 72, n. 3, p. 2925–2948, 2014. Cited on page 47. PHAM, X. H.; JUNG, J. J.; VU, S.-B. P. L. A. Exploiting social contexts for movie recommendation. Malaysian Journal of Computer Science, v. 27, n. 1, p. 68–79, 2014. Cited on page 51. QUEIROZ, S. R. d. M.; CARVALHO, F. d. A. de. Making collaborative group recommendations based on modal symbolic data. In: SPRINGER. Brazilian Symposium on Artificial Intelligence. [S.l.], 2004. p. 307–316. Cited on page 35. R Core Team. R: A Language and Environment for Statistical Computing. Vienna, Austria, 2015. Disponível em: <http://www.R-project.org/>. Cited on page 131. REICHLING, T.; WULF, V. Expert recommender systems in practice: evaluating semi-automatic profile generation. In: ACM. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. [S.l.], 2009. p. 59–68. Cited on page 36. RENDLE, S. Factorization machines with libfm. ACM Transactions on Intelligent Systems and Technology (TIST), ACM, v. 3, n. 3, p. 57, 2012. Cited on page 60. RESNICK, P. et al. Grouplens: an open architecture for collaborative filtering of netnews. In: ACM. Proceedings of the 1994 ACM conference on Computer supported cooperative work. [S.l.], 1994. p. 175–186. Cited on page 23. RESNICK, P.; VARIAN, H. R. Recommender systems. Communications of the ACM, ACM, v. 40, n. 3, p. 56–58, 1997. Cited on page 32. RICCI, F.; ROKACH, L.; SHAPIRA, B. Introduction to recommender systems handbook. [S.l.]: Springer, 2011. Cited 11 many times on page 23, 24, 33, 34, 35, 37, 81, 84, 85, 86, and 87. SAHEBI, S.; BRUSILOVSKY, P. Cross-domain collaborative recommendation in a cold-start context: The impact of user profile size on the quality of recommendation. In: User Modeling, Adaptation, and Personalization. [S.l.]: Springer, 2013. p. 289–295. Cited 9 many times on page 42, 44, 45, 47, 57, 58, 60, 61, and 131. SAHEBI, S.; COHEN, W. W. Community-based recommendations: a solution to the cold start problem. In: Workshop on Recommender Systems and the Social Web, RSWEB. [S.l.: s.n.], 2011. Cited on page 60. http://www.R-project.org/ 238 References SANTOS, V. dos et al. A recommender system architecture for an inter-application environment. In: IEEE. 12th International Conference on Intelligent Systems Design and Applications (ISDA). [S.l.], 2012. p. 472–477. Cited 12 many times on page 9, 25, 41, 58, 59, 61, 69, 73, 74, 90, 220, and 224. SETTEN, M. V.; POKRAEV, S.; KOOLWAAIJ, J. Context-aware recommendations in the mobile tourist application compass. In: SPRINGER. Adaptive hypermedia and adaptive web-based systems. [S.l.], 2004. p. 235–244. Cited on page 47. SHANI, G.; GUNAWARDANA, A. Evaluating recommendation systems. In: Recommender systems handbook. [S.l.]: Springer, 2011. p. 257–297. Cited 2 many times on page 37 and 128. SHAPIRA, B.; ROKACH, L.; FREILIKHMAN, S. Facebook single and cross domain data for recommendation systems. User Modeling and User-Adapted Interaction, Springer, v. 23, n. 2-3, p. 211–247, 2013. Cited 8 many times on page 38, 41, 44, 47, 57, 58, 60, and 61. SHARDANAND, U.; MAES, P. Social information filtering: algorithms for automating “word of mouth”. In: ACM PRESS/ADDISON-WESLEY PUBLISHING CO. Proceedings of the SIGCHI conference on Human factors in computing systems. [S.l.], 1995. p. 210–217. Cited on page 23. SHEPSTONE, S.; TAN, Z.-H.; JENSEN, S. Using audio-derived affective offset to enhance tv recommendation. Multimedia, IEEE Transactions on, v. 16, n. 7, p. 1999–2010, Nov 2014. ISSN 1520-9210. Cited 2 many times on page 47 and 52. SHI, Y.; LARSON, M.; HANJALIC, A. Tags as bridges between domains: Improving recommendation with tag-induced cross-domain collaborative filtering. User Modeling, Adaption and Personalization, Springer, v. 6787, p. 305–316, 2011. Cited 2 many times on page 44 and 47. SONG, S.; MOUSTAFA, H.; AFIFI, H. Enriched IPTV services personalization. In: IEEE International Conference on Communications. [S.l.]: Ieee, 2012. p. 1911–1916. ISBN 978-1-4577-2053-6. Cited on page 54. SOUZA, D. et al. Towards a context ontology to enhance data integration processes. In: Proceedings of the 4th Workshop on Ontologies-based Techniques for Databases (in VLDB’08). [S.l.: s.n.], 2008. Cited on page 49. STEWART, A. et al. Cross-tagging for personalized open social networking. In: ACM. Proceedings of the 20th ACM conference on Hypertext and hypermedia. [S.l.], 2009. p. 271–278. Cited 3 many times on page 40, 41, and 44. STRANG, T.; LINNHOFF-POPIEN, C. A context modeling survey. In: Workshop Proceedings. [S.l.: s.n.], 2004. Cited on page 49. SZOMSZOR, M. et al. Semantic modelling of user interests based on cross-folksonomy analysis. [S.l.]: Springer, 2008. Cited on page 42. TANG, J. et al. Cross-domain collaboration recommendation. In: ACM. Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining. [S.l.], 2012. p. 1285–1293. Cited on page 92. References 239 TANG, X.; WAN, X.; ZHANG, X. Cross-language context-aware citation recommendation in scientific articles. In: ACM. Proceedings of the 37th international ACM SIGIR conference on Research & development in information retrieval. [S.l.], 2014. p. 817–826. Cited 5 many times on page 61, 62, 63, 65, and 67. TEKIN, C.; SCHAAR, M. van der. Contextual online learning for multimedia content aggregation. Multimedia, IEEE Transactions on, IEEE, v. 17, n. 4, p. 549–561, 2015. Cited 6 many times on page 61, 62, 63, 65, 66, and 67. TIROSHI, A. et al. Cross social networks interests predictions based ongraph features. In: ACM. Proceedings of the 7th ACM conference on Recommender systems. [S.l.], 2013. p. 319–322. Cited 3 many times on page 58, 59, and 60. TIROSHI, A.; KUFLIK, T. Domain ranking for cross domain collaborative filtering. In: User Modeling, Adaptation, and Personalization. [S.l.]: Springer, 2012. p. 328–333. Cited 2 many times on page 40 and 42. TREWIN, S. Knowledge-based recommender systems. Encyclopedia of library and information science, v. 69, n. Supplement 32, p. 180, 2000. Cited 2 many times on page 24 and 25. TUCKER, L. R. Some mathematical notes on three-mode factor analysis. Psychometrika, Springer, v. 31, n. 3, p. 279–311, 1966. Cited on page 64. UBERALL, C.; MUTTUKRISHNAN, R. Recommendation index for DVB content using service information. In: IEEE International Conference on Multimedia and Expo. [S.l.: s.n.], 2009. p. 1–4. Cited 2 many times on page 35 and 36. VÉRAS, D. et al. A literature review of recommender systems in the television domain. Expert Systems with Applications, Elsevier, v. 42, n. 22, p. 9046–9076, 2015. Cited 4 many times on page 33, 35, 36, and 54. VERAS, D. et al. Context-aware techniques for cross-domain recommender systems. In: IEEE. 2015 Brazilian Conference on Intelligent Systems (BRACIS). [S.l.], 2015. p. 282–287. Cited 2 many times on page 40 and 54. VIEIRA, V. et al. A context-oriented model for domain-independent context management. Revue d’intelligence artificielle, v. 22, n. 5, p. 609–627, 2008. Cited on page 49. VIEIRA, V.; TEDESCO, P.; SALGADO, A. C. Towards an ontology for context representation in groupware. In: Groupware: Design, Implementation, and Use. [S.l.]: Springer, 2005. p. 367–375. Cited on page 49. VIEIRA, V.; TEDESCO, P.; SALGADO, A. C. Modelos e processos para o desenvolvimento de sistemas sensíveis ao contexto. In: Jornadas de Atualização em Informática. [S.l.]: André Ponce de Leon F. de Carvalho, Tomasz Kowaltowski.(Org.), 2009. p. 381–431. Cited 6 many times on page 17, 48, 49, 50, 90, and 223. VILDJIOUNAITE, E. et al. Unobtrusive dynamic modelling of tv programme preferences in a finnish household. Multimedia systems, v. 15, p. 143–157, 2009. Cited on page 55. 240 References WANG, F.; LI, D.; XU, M. A location-aware tv show recommendation with localized sementaic analysis. Multimedia Systems, Springer Berlin Heidelberg, online, p. 1–8, 2015. ISSN 0942-4962. Disponível em: <http://dx.doi.org/10.1007/s00530-015-0451-z>. Cited 2 many times on page 52 and 55. WINOTO, P.; TANG, T. If you like the devil wears prada the book, will you also enjoy the devil wears prada the movie? a study of cross-domain recommendations. New Generation Computing, Springer, v. 26, n. 3, p. 209–225, 2008. Cited 5 many times on page 24, 25, 41, 58, and 61. WINTER, J. C. D.; DODOU, D. Five-point likert items: t test versus mann-whitney- wilcoxon. Practical Assessment, Research & Evaluation, Dr. Lawrence M. Rudner, v. 15, n. 11, p. 1–12, 2010. Cited on page 131. YUAN, Z. et al. Structural context-aware cross media recommendation. In: Advances in Multimedia Information Processing–PCM 2012. [S.l.]: Springer, 2012. p. 790–800. Cited 5 many times on page 62, 63, 64, 66, and 67. ZHANG, H.; ZHENG, S. Personalized TV program recommendation based on TV-anytime metadata. In: IEEE International Symposium on Consumer Electronics. [S.l.: s.n.], 2005. di, p. 242–246. ISBN 0780389204. Cited on page 37. ZHANG, J.; YUAN, Z.; YU, K. Cross media recommendation in digital library. In: The Emergence of Digital Libraries–Research and Practices. [S.l.]: Springer, 2014. p. 208–217. Cited 4 many times on page 61, 62, 63, and 67. ZHAO, L. et al. Active transfer learning for cross-system recommendation. In: CITESEER. AAAI. [S.l.], 2013. Cited on page 47. ZHIWEN, Y.; XINGSHE, Z. Design, implementation, and evaluation of an agent-based adaptive program personalization system. In: Fifth International Symposium on Multimedia Software Engineering. [S.l.: s.n.], 2003. p. 140–147. ISBN 0769520316. Cited on page 37. ZHUANG, F. et al. Cross-domain learning from multiple sources: a consensus regularization perspective. Knowledge and Data Engineering, IEEE Transactions on, IEEE, v. 22, n. 12, p. 1664–1678, 2010. Cited on page 44. http://dx.doi.org/10.1007/s00530-015-0451-z Dedication Acknowledgements Epigraph Resumo Abstract List of Figures List of Tables Contents Introduction Contextualization Motivation Problem Statement Objectives Proposal Overview Contributions Thesis Outline Background and Related Work Recommender Systems Strategies User Profiling Evaluation Cross-Domain Recommender Systems Definition of Domain Cross-Domain Recommendation Tasks Cross-Domain Recommendation Goals Cross-Domain Recommendation Scenarios Cross-Domain Approaches Cross-Domain Evaluation Evaluation Data Partitioning Evaluation Metrics Sensitivity Analysis Context-Aware Recommender Systems Definition of Context Modelling Contextual Information Obtaining Contextual Information Contextual Information Relevance Context-Aware Approaches CARS Evaluation Related Works Cross-Domain Recommendation based on Collaborative Filtering Cross-Domain Recommendation based on Context-Awareness Final Remarks CD-CARS Proposal CD-CARS Problem Formalization Modelling Contextual Information Contextual Features Formalization Obtaining and Selecting Relevant Contextual Information CD-CARS Algorithms Proposed Algorithms Cross-Domain PreF Algorithm Cross-Domain PostF Algorithm Cross-Domain Modelling Algorithm Cross-Domain Hybrid Contextual Algorithms Base Cross-Domain Algorithms Single-Domain as Cross-domain Algorithms Neighborhood-based Algorithms Matrix factorization algorithms Cross-Domain Algorithm Final Remarks CD-CARS Implementation Dataset Acquisition Obtaining Contextual Information Temporal Dimension Location Dimension Companion Dimension Selecting Relevant Contextual Attributes and Values Cross-Domain Datasets Description Book-Television dataset Book-Music dataset Contextual Model Implementation Proposed Algorithms Implementation Pre-filtering Implementation Post-filtering Implementation Base Cross-domain Algorithm Implementation Final Remarks CD-CARS Evaluation Evaluation Methodology Settings of the Algorithms Predictive Performance Classification Performance Sensitivity Evaluation Statistical Significance Analysis Evaluation Results Book-Television Results Television as Target Domain Temporal Dimension Location Dimension Companion Dimension Combining Contextual Dimensions Book as Target Domain Temporal Dimension Location Dimension Companion Dimension Combining Contextual Dimensions Summary Book-Music Results Music as Target Domain Temporal Dimension Location Dimension Companion Dimension Combining Contextual Dimensions Book as Target Domain Temporal Dimension Location Dimension Companion Dimension Combining Contextual Dimensions Summary Discussion Final Remarks Conclusion Contributions Limitations Lines for Further Work References 
work_i5tjkbben5e77ie5lwxz5vp5qe	----	LHT-02-2020-0038_proof 1..23 Intelligent libraries: a review on expert systems, artificial intelligence, and robot Asefeh Asemi Doctoral School of Business Informatics, Corvinus University of Budapest, Budapest, Hungary and University of Isfahan, Isfahan, Iran Andrea Ko Corvinus University of Budapest, Budapest, Hungary, and Mohsen Nowkarizi Department of Knowledge and Information Science, Ferdowsi University of Mashhad, Mashhad, Iran Abstract Purpose – This paper reviews literature on the application of intelligent systems in the libraries with a special issue on the ES/AI and Robot. Also, it introduces the potential of libraries to use intelligent systems, especially ES/AI and robots. Design/methodology/approach – Descriptive and content review methods are applied, and the researchers critically reviewed the articles related to library ESs and robots from Web of Science as a general database and Emerald as a specific database in library and information science from 2007–2017. Four scopes considered to classify the articles as technology, service, user and resource. It is found that published researches on the intelligent systems have contributed to many librarian purposes like library technical services like the organization of information resources, storage and retrieval of information resources, library public services as reference services, information desk and other purposes. Findings – A review of the previous studies shows that ESs are a useable intelligent system in library and information science that mimic librarian expert’s behaviors to support decision making and management. Also, it is shown that the current information systems have a high potential to be improved by integration with AI technologies. In this researches, librarian robots mostly designed for detection and replacing books on the shelf. Improving the technology of gripping, localizing and human-robot interaction are the main concern in recent librarian robot research. Our conclusion is that we need to develop research in the area of smart resources. Originality/value – This study has a new approach to the literature review in this area. We compared the published papers in the field of ES/AI and robot and library from two databases, general and specific. Keywords Library system, Intelligent systems, Artificial Intelligent (AI), Intelligent library, Smart library, Expert System (ES), Robot Paper type Literature review 1. Introduction In computer science, artificial intelligence (AI) is an important topic. In this context, the focus is on human behavior and how machines can imitate intelligent human behavior. AI involves A review: intelligent libraries © Asefeh Asemi, Andrea Ko and Mohsen Nowkarizi. Published by Emerald Publishing Limited. This article is published under the Creative Commons Attribution (CC BY 4.0) licence. Anyone may reproduce, distribute, translate and create derivative works of this article (for both commercial & non- commercial purposes), subject to full attribution to the original publication and authors. The full terms of this licence may be seen at: http://creativecommons.org/licences/by/4.0/legalcode. This research has been supported by the “Project no. NKFIH-869-4/2019 has been implemented with the support provided from the National Research, Development and Innovation Fund of Hungary, financed under the T�emater€uleti Kiv�al�os�agi Program 2019 funding scheme.” The current issue and full text archive of this journal is available on Emerald Insight at: https://www.emerald.com/insight/0737-8831.htm Received 23 February 2020 Revised 25 February 2020 Accepted 25 February 2020 Library Hi Tech Emerald Publishing Limited 0737-8831 DOI 10.1108/LHT-02-2020-0038 http://creativecommons.org/licences/by/4.0/legalcode https://doi.org/10.1108/LHT-02-2020-0038 amongst other expert systems (ESs), fuzzy logic, artificial neural network, evolutionary algorithms, case-based reasoning, image processing, natural language processing, speech recognition and robotics. These areas are not separate, and in many intelligent systems at the same time, two or more AI techniques contribute in problem to solving. AI techniques or tools have utilized in many areas such as business, management, medicine, military etc. Library and information science also have developed in using intelligent systems. “Library management and its activities apply repetitious and time-consuming activities. Hence, in order to increase efficiency and effectiveness, many libraries are moving toward automation of their activities” (Dwivedi et al., 2013). AI techniques give more accuracy to the automation of libraries. The ideas of the utilization of intelligent systems instead of the classic systems in libraries started in 1990. Intelligent systems are used in the library to provide knowledge- based services to users of the library system and end-users. These systems, as a complementary system to the main library system, can make intelligent decisions for the retrieval and use of information resources. The careful decision making of these systems is based on the knowledge created by users in the library system. These systems are an important competitor of human activities in the library. This could have implications for librarians. The presented papers at the 27th Annual Clinic on Library Applications of Data Processing dealt the capabilities relate to the library applications: descriptive cataloging, technical services and collection development, subject indexing, reference services, database searching and document delivery (Lancaster and Smith, 1990). A lot of research has already been made of the various uses of AI technologies in libraries. Even Hsieh and Hall (1989) examined the definition and history of AI and investigates the body of literature on AI in “Library Literature” and Lisa. In the same year, O’Neill and Morris (1989) looked at the challenge and implications of ESs technology for LIS. In 1998 under the direction of NCR, Inc. and its Future Mapping® process, experts worked together to map out the best scenario for the libraries of the future (Leslie, 1999). After several years, a fully automated 24/7 online librarian system designed to respond to routine and repeat inquiries from distance learners (Payne and Bradbury, 2002). It seems that AI’s applications have been considered in various aspects of library and information science. Based on the review of the major models that exist, the following factors are effective in using human intelligence in information systems: (1) Understanding the nature of the information needs and defining this need for the system, (2) Identifying information resources that are relevant to information needs, (3) Evaluation of existing information resources, evaluation of retrieved information, (4) Organizing existing information resources, organizing selected information from items retrieved, (5) Managing existing information resources, managing retrieved information, (6) Using existing information resources, using retrieved information, (7) Information and knowledge analysis, (8) Converting information to knowledge, (9) Dissemination and transfer of information and knowledge, (10) Interaction and exchange of information and knowledge. The above is a very useful list for using AI in scientific databases and library and information affairs. For example, recommender systems are very important in identifying information resources and selecting them. These systems can beveryhelpful inselecting the right resources LHT based on user behavior in using the retrieval system. “However, we probably cannot consider the system to be intelligent by human standards. The fact that we are transient organic beings that possess five senses and feel, as well as think. In short, computers lack: all that man is, all mere complexities, the fury and the mire of human veins.” (Bailey, 1991 and Bailey, 1992). Today, many years have passed since research into AI programming techniques and the use of WSs for providing technical services to libraries and information databases and providing public services to end-users. Now the dream of smart libraries has become a reality. Cao et al. (2018) have paid to the conceptualization of the smart library, and scientists and professionals have created systems that can be thought of and decided instead of the librarian. Cox et al. (2018) studied on the intelligent library. They paid to the thought leaders’ viewpoints on the impact of AI on the libraries. Even these systems can mimic the librarian’s behavior. The purpose of the study is to review the articles on intelligent libraries and the use of Artificial Intelligence (AI), Expert Systems (ESs) and robots in the libraries. The articles retrieved from Web ofScience (WoS)as a cumulative databaseand EmeraldInsight as a professionaldatabase in library and information science. The research questions are the following: (1) In WoS, which articles did authors write about ES/AI and robots’ application in the library? (2) In Emerald Insight, which articles did authors write about ES/AI and robots’ application in the library? 2. Artificial intelligence AI applies to different sciences. We can say in the library and information science, it more uses in scientific databases and library systems. Such as behavioral science, social sciences, psychology, management and library science and information science. It is related to some of the systems that apply different forms of intelligence such as learner systems, inferior systems, systems with natural language understanding or natural language interpretation, systems with visual scene perception and systems that perform other types of feat that require human types of intelligence (Bavakutty and Salih, 2006). In this branch of the science that involves machines, solutions are utilized to solve complex problems of human behavior. We can present computer-based algorithms based on human behavior and knowledge in using systems. “It is an interdisciplinary field making use of concepts from various fields like cybernetics, information theory, psychology, linguistics, logic, etc. it can use to simulate human behavior and for computer ailed instruction, ES, robots and for NLP. It can also use for Intelligent Retrieval from databases” (Bavakutty and Salih, 2006). In this way, computer software and the use of various computer-based products help in the operation of various types of libraries and their public services and the generation of output products. Automation implies the degree of mechanization where the routines and receptive jobs or operations are left to be performed by machines with little or no intervention by human beings. Lesser the degree of human intervention, greater the degree of automation; this does not mean that automation does away with human beings. On the contrary, human beings are relieved of routine chores, giving them more time for tasks, which require their intelligence. In view of the various features of a modern computer system, we find that it has been applied in several areas of library work. Book acquisitions, cataloging, serials control, and circulation, information retrieval and dissemination, interlibrary loan, cooperative acquisition and cataloging have been automated in the library (Lakshmikant and Vishnu, 2008). 3. Intelligent systems Intelligent systems (ISs) are defined as any formal or informal system that is able to obtain and process data, to interpret the data by applying technologies of artificial intelligence and A review: intelligent libraries business intelligence and to provide reasoned judgments based on that to decision makers as a basis for action (Sharda et al., 2017). ISs are computer-based systems that help in the task of subject indexing can be thought of as an ES (Lancaster, 1997). Lancaster has a clear statement relating to the scope of AI: “Computer programs have been developed, which exhibit human-like reasoning, which may be able to learn from their mistakes and which quickly and cleverly perform tasks normally done by scarce and expensive human experts.” AI has a wide application area. Figure 1 gives a good idea of this coverage. Technologies that are frequently used in intelligent systems: machine learning, case-based reasoning, genetic algorithms, fuzzy logic and natural language processing (NLP). NLP is another facility of an intelligent system that can use to retrieve information needs from different scientific databases. In the information retrieval process, the user can state his information requirement in natural language, making the searching more easily and fruitful this allows users to state complex retrieval languages (Bavakutty and Salih, 2006). Business intelligence (BI) as is the set of techniques and tools for the transformation of raw data into meaningful and useful information for business analysis/decision support purposes (Sharda et al, 2017). BI solutions include data access, storage, data analysis and visualization technologies to support better decision making. 4. Expert systems Expert systems (ESs) are computer-based systems that simulate human decision making. They can integrate with information systems to improve their accuracy and performance (Singh et al., 1996). Various librarian ES has been developed. Waters (1986) designed the National Agricultural Library’s microcomputer-based ES to help users obtain answers to simple reference questions. In general, they ask questions from the user and take the user’s answer as input, then explain the rationale for decision results. In general, these systems consist of two main elements: A knowledge base and inference engine. The knowledge base encompasses all the information needs that human/librarian experts are using to decide. This information is present in the knowledge base as facts and rules. ESs can make much better decisions than librarian decision makers because their knowledge base can involve the experiences of a team of the best experts. The manner of librarian experts to make decisions is emulated for the design rules of the knowledge base. The rules are consisting of two main phases: “if phase” and “then phase.” The “if phase” is consisting of conditions, and the “then phase” is consisting of results. ESs are distinguished from other computer systems with the application of reasoning through the inference engine. The inference engine simulates human decision makings based on the knowledge base and a rule base (Figure 2). The creation of an ES includes the extraction of the relevant knowledge from the expert human, and it is often nature heuristic. ESs use problem-solving in different areas such as Figure 1. AI Coverage (Lancaster, 1997) LHT medicine, business, computer science, law, defense, education, mathematics, engineering, geology, etc. (Bavakutty and Salih, 2006). Many of the library’s activities are specialized. For this reason, library software should be used to improve the library’s performance. The efforts have been made in this regard. Denning and Smith (1994) had a survey on the “Electronic Library Search Assistant.” Kruk and Krawczyk (2004) studied about intelligent resources search in virtual libraries. Devadason and Vespry (1996) studied about planning for the library staff. It encapsulates the expert knowledge of a library staff planner. They presented the LISPA (Library and Information center Staff Planning Advisor). By reviewing the research and literature, specialized library systems can have the following applications in the area of providing library and information services: (1) Knowledge-based indexing (Amin and Razmi, 2009); (2) Natural Language Processing and abstracting (Albayrak and Erensal, 2004); (3) Reference work (Amin and Razmi, 2009); (4) Cataloging (Weiss, 1994) and (Amin and Razmi, 2009); (5) Online information retrieval (Bellman and Zadeh, 1970), (Sacchanand and Jaroenpuntaruk, 2006) and (Bavakutty and Salih, 2006); (6) Using intelligent interfaces in online information storage and retrieval systems; (7) Information needs analysis and representation, including different services, such as classification, indexing and abstracting; (8) Reference services; (9) Development of collection; (10) Hypertext and hypermedia (Bavakutty and Salih, 2006). 5. Methodology Descriptive and content review methods are applied to the study. The researchers critically reviewed the articles related to ES/AI and robots in the library. According to this review, the application of ES/AI and robots classified as the following: Technology: The articles surveyed and evaluated the information management systems in the libraries belongs to this group. These articles relate to usability and implementation. They do not propose or propose an information system or model. Resource: These articles related to information resources. This category may include the selection, acquisition and use of information resources. User / End-user: Existing information and knowledge systems/models are usually working based on the opinion of experts/users and end-user behavior. Therefore, applying ES Knowledge Base If-then rules Ability to ask question get input and explain rationale for answer. Inference Engine User Interface Figure 2. ES elements A review: intelligent libraries technologies such as inference engine and fact/rule base will improve the performance and accuracy of considered systems. Service: The articles in this group have proposed an ES or related technology and methods that can be connected and included in ESs to present public or technical services. The public services present to end-users to fulfill their information needs and technical services present to the librarians or any professional user in library activities. 6. Findings The purpose of the study is to review the articles on intelligent libraries and the use of ES/AI and robots in the libraries. Based on the research questions, the findings presented in two sections. The first section is related to the review of the articles in WoS as a general database in different subjects. The second section is related to the review of the articles in Emerald as a professional database in the Library and Information Science. According to this review, the application of ES/AI and robots classified into four classes such as technology, resource, user/ end-user and service. 6.1 ES/ AI and robots’ application in the library (WoS) The topics of “expert system” and “library” were searched in the WoS database on 10th Oct 2017. We found 1,208 documents related to this topic. Then we have refined the topics through “Research Area,” “Document Type.” In the research area, we selected the area of “Information Science, Library Science.” We chose “article” for “document type” and excluded unrelated articles. Finally, found 14 articles as a result, which are shown in Table 1. The review of papers shows the fading of the ES/AI in recent studies. It is found that the majority (46%) of the paper worked on the experts’/users’ behavior. This is even though no research has been done on the use of intelligent resources using ESs between the years 2007–2017 on the WoS (Figure 3). However, the studies that are related to information systems have a closed relation with the knowledge and opinion of experts. Using ES technologies such as inference engine and fuzzy rule base may increase the accuracy of them. Therefore, the current information systems can be improved by integration with ES technologies. ESs use in intelligent libraries. In general, the information provided to users in a library leads to a change in the behavior of the user’s knowledge and creates learning. The intelligent library uses an appropriate protocol for the exchange of information. This protocol is unique, and it is designed to help, confirm or perform the terms of the agreement. The terms of the agreement include a series of guidelines that will be carried out automatically. These guidelines relate to information sources, services, and technology for distributing and exchanging information. For operating a smart library, resources and services must be available under the agreement. All users must use the digital signature and agree to the terms of the agreement. Smart libraries can exchange information based on the internet of Things (IoT). Recently the researchers tried to increase the ability of librarian robots by applying the new methods. We searched for the topic of “Librarian robot” using WoS on 10th Oct 2017. Then we limited the results to the duration of 2007–2017. We excluded unrelated articles and finally found 15 articles and proceeding papers as a result, which is shown in Table 2. In this table, we determine the research area related to applied methodologies of papers in the “Research area of publication source.” A summary of the applied method is explained in “Method,” and the main contribution of papers is mentioned in “contribution.” The most recent papers that are related to librarian robots are in the area of service (Figure 4). Improving the technology of gripping, localizing and human-robot interaction are the most discussed issues in librarian robots. Librarian robots can be used in large libraries. This robot reduces a lot of common and duplicate activities in different places of the library, LHT No. Author Contribution Application based on the reference Class 1 Asemi et al. (2012) Management Information System (MIS) Supportive tool for library operations and provides suitably detailed reports in an accurate, consistent and timely manner Technology User/end- user 2 Black (2011) Web Content Management System (CMS) Support a large distributed content model and shares the CMS trail method used, which directly included content provider feedback side-by-side with the technical experts Technology User/end- user 3 Chu et al. (2010) learning system Support context-aware ubiquitous learning Technology User/end- user 4 Chu et al. (2010) electronic library with supporting context-aware ubiquitous-learning Supporting learning activities conducted in real-world environments Technology User/end- user 5 Ding and Sølvberg (2007) Rule-based metadata interoperation Support querying across distributed digital libraries created in heterogeneous metadata schemas, without requiring the availability of a global schema Service User/end- user 6 Golub et al. (2014) Terminology registries (TRs) Provide the content of knowledge organization systems (KOS) available both for human and machine access Technology User/end- user 7 Huili and Bo (2017) Smart library, Library robot Making the robot more like a librarian, focus on key technologies to take the robot into the real library environment, and cultivate relevant technical talents Service User/end- user 8 Hwang et al. (2011) A grid-based knowledge acquisition approach and a Mind tool is proposed Help students organize and share knowledge for differentiating a set of learning targets based on what they have observed in the field Technology User/end- user 9 Iglesias (2013) Application Robots in Library Advancements in Library Automation Automating Reference, Storage, Technical Services, Circulation desketc. Service User/end- user 10 Ismail and Kareem (2011) Identifying how novice researchers search, locate, choose and use web resources Supporting information-seeking behavior of novice researchers by specific research tools Technology User/end- user 11 Kao and Wu (2012) personalized knowledge integration platform for digital libraries Providers users with personalized information and knowledge services Technology User/end- user 12 Mehtab Alam Ansari (2008) Online Public Access Catalogue (OPAC) Allowing a user to search online and retrieve records/catalogue and depending on the underlying library management software/ online reservation, circulation and so on Technology User/end- user 13 Mei et al. (2017) Intelligent Use of Library proposes a general framework to establish the dynamic movement primitives library (DMPL) for a mobile robot path planning in an unknown environment Service 14 Phillips (2017) AI in Library, Library Automation Development in robotics and AI, and the potential implications for library services. It explores the impacts of automation of human work, with a particular focus on recent advances in robotics and AI and how these may affect library services and library work in future Service Table 1. Articles related to ES/ AI and library (WoS) A review: intelligent libraries especially at the library’s repository. For example, this robot can be helpful in shelf-reading activity. There are some imaginations that the use of librarian robot creates a gap between information and people. Smart libraries and librarian robots are always faced with this challenge. But not a way out of using new technologies, because the development of information does not coincide with the development of expert human resources. In many libraries, librarian robots can be helpful in solving library problems. Only the small number of the studies are related to resources. It is shown that we need to develop our research in this area. The library should take special care of every aspect related to the man-machine interface: favoring systems standardization, avoiding the accumulation of different equipment, using a clear, brief and direct language, including images and sound, representing reality and reflecting the human mental patterns (De Prado, 2000). AI techniques such as genetic algorithms, artificial neural networks, ESs, and fuzzy logic or hybrid methods can improve librarian robots to reflect human mental patterns. 6.2 ES and robot’s application in the library using Emerald Insight [1] Table 3 shows the review of the papers in the field of ES/AI and robots, and library exported from the Emerald Insight as a specific database in the library and information science. Figure 5 shows the most recent papers exported from Emerald Insight, which are related to ES/AI and robot in the library are in the area of service. The finding is the same as the exported papers from the WoS database. The following trends show a line graph of the relative frequencies across the main category in the abstracts of the articles (Figure 6). The thematic interaction was observed in the main categories of the articles based on their keywords. Most common categories in the abstracts are digital, information, library, search, and user. Figure 7 shows a line graph of the relative frequencies across the main category of the keywords of the articles. The thematic interaction was observed in the main categories of the articles based on their keywords. Most common categories in the abstracts are digital, information, Internet, library, and systems. 7. Discussion The ES should be considered only when development is: “possible,” “appropriate,” and “justified” (Lancaster, 1997). This question must be answered before we initialed an ES project. Waters (1986) gives some good guidelines on when we should consider using ESs. An Resource 0% Service 19% Technology 35% User/end-user 46% Figure 3. The scope of the articles in the field of ES/AI and library (WoS) LHT No Author Contribution Application based on the reference Class 1 Comsa et al. (2014) Presenting some similar state-of- the-art developments, CAD models of two book manipulators and also an innovative design approach in designing library book handling gripper mechanisms Gripper prototype is manufactured using light-weight thermoplastic reinforced material for the mobile finger Technology Service 2 Du et al. (2011) Designing an embedded controller for the pneumatic manipulator of library robot Using PC/104 boards system and emphasizing parameter self- tuning fuzzy-PID algorithm of the controller Technology Service 3 Grigorescu et al. (2010) Propose a robust feature extraction for 3D reconstruction of segmented boundary objects Using means of including feedback control at image segmentation level for boundary feature extraction. The objective of feedback is to adjust segmentation parameters in order to cope with scene uncertainties, such as variable illumination conditions Robustly extracted 2D object features are provided as input to the 3D object reconstruction module of the FRIEND vision system Technology Service 4 Heyer et al. (2012) Propose a new approach for detecting and grasping the book reliably Combination of two algorithms for book detection and grasping and users stereo vision together with hand camera to achieve a high rate of success Resource Service 5 Huili and Bo (2017) Making the robot more like a librarian, focus on key technologies to take the robot into the real library environment, and cultivate relevant technical talents Based on extensive research literature and best practices of library robots, this paper states robot technology can effectively solve some problems in library management and service, and improve user satisfaction to a certain extent Technology Service User/end- user 6 Iglesias (2013) Advancements in Library Automation: Automating Reference, Storage, Technical Services, Circulation Desk, etc. Different Methods. In this book presented different articles about Automating Reference, Storage, Technical Services, Circulation Desk, etc. Technology Service User/end- user 7 Kim et al. (2009) Propose an information structured environment called u-RT to enable a librarian robot to arrange books on bookshelves using ambient intelligence The librarian robot consists of a manipulator to recognize and manipulate books, and a mobile platform to localize itself and navigate using ambient RFID tags embedded in a floor. The proposed u-RT space connects physical and virtual space using physical hyperlinks Resource Service (continued) Table 2. Articles related to library and robot (WoS) A review: intelligent libraries No Author Contribution Application based on the reference Class 8 Kim et al. (2008) Realize ambient intelligence in the ubiquitous robot technology space The ubiquitous space for the robotic library is introduced and an RFID technology-based approach for the librarian robot proposed Technology 9 Kim et al. (2013) Investigates whether assigning a caregiving role to a robot or to its human interacting has psychological effects on the quality of human-robot interaction (HRI) College students interacted with a social robot in a between-subjects experiment with two manipulated conditions: one where the robot played the role of an ophthalmologist and one where participants played the role of the ophthalmologist User/end- user Service 10 Lin et al. (2013) Incorporates the robotic assistance in investigating the book locating behaviors of child patrons, and develop a service robot for child patterns in library settings Describe the process of developing an assistant robot with locating resources in libraries. Consulting the stakeholders, including child patrons and librarians. Analyzing the needs and incorporating into the design of library robot User/end- user Service 11 Mei et al. (2017) Proposes a general framework to establish the dynamic movement primitives library (DMPL) for a mobile robot path planning in an unknown environment Mathematical model: before the library is building, the workspace of the mobile robot is divided into multiple sectors through a unique sampling technique. Then, using a joystick, a user operates the mobile robot moving from start to any sample point, simultaneously recording the states such as position, velocity and acceleration. The primitives will be extracted from the recorded state sequence, and the learned weights will be stored in the DMPL. In the second phase, the DMPL is used online to supply the path planning decision Service 12 Mikawa (2010) A practicum Track Using Librarian robot in a support program for contemporary educational needs Providing a training ground for creating new types of contents for the Internet age, where students of several specialized fields come together User/end- user Service 13 Modler et al. (2014) Presents an innovative design approach in grippers for library automation context The gripper CAD model and the experimental gripper prototype, developed using light-weight thermoplastic reinforced material for the mobile finger Resource Service 14 Modler et al. (2012) Proposes one CAD gripper model designed in solid works software. The CAD model for the gripper and FEM simulation is presented The parallel gripper prototype is still in the manufacturing process using light-weight glass-fiber reinforced material Resource Service Table 2. (continued) LHT ES has received a lot of attention from the research community in the 1980s. Unfortunately, much of the writing sensationalized the field expectations dramatically (Lakshmikant and Vishnu, 2008) fueled by public expectations began to over-promise misconceptions about what AI can and cannot do arise and they persist today. Many rushed into the field in search of quick answers and quick profits. Several Al researchers saw what was happing and feared a backlash. Once all the excitement wore off during 1988–90, things did begin to change some of the realities and limitations of the AI techniques became evident. An AI backlash has resulted in ascertaining to an extent, but fortunately, it has not been wide-scaled instead. The optimism remains with a better sense of realism than before, and both the benefits and limitations are better appreciated. In general, in the protocol of an intelligent library, clear the subject of the agreement, the digital signature, and the divisional platform must be clear. Smart library agreements can be applied in various fields, including the acquisition of information resources, presenting public services to the end-user, technical services, management, and many other library activities in many library activities. Advantages of intelligent libraries are the security of information, the cost of services, the speed of service to the user community, the pursuit of activities within the framework of the standard. Problems that may be encountered in intelligent libraries are human factor problems, the cost of implementing intelligent library agreements, and copyright issues. One of the key elements in the smart library is digital identity. Based on the intelligent library protocol, users can create or control their digital identities. This digital identity includes information, reputation and information resources used or required by the users. Smart library users in the context of the IoT can decide which information to be No Author Contribution Application based on the reference Class 15 Phillips (2017) Developments in robotics and AI, and the potential implications for library services. It explores the impacts of automation of human work, with a particular focus on recent advances in robotics and AI and how these may affect library services and library work in future An in-depth literature review, and the results of original research. The research consisted of a survey of the general population, including library users and workers, and a focus group with library workers only. Key themes explored include: general perceptions and experience of automation in libraries, potential acceptance levels of robots being used in libraries, and the predicted positive and negative outcomes Service Table 2. Technology 21% Service 47% Resource 14% User/end- user 18% Figure 4. The scope of the articles in the field of library and robot (WoS) A review: intelligent libraries No Author/s Keywords Application based on the reference Class 1 Ataman Kemal (2009) Archives management, library Information science, Records management, Information profession Library services present by the smart talking robot Xiaotu based on artificial intelligence Service 2 Bi et al. (2017) Internet of things, Planning and control, Robot, System interaction, Wireless sensors network (WSN) The application of intelligent agents in library services Service 3 Bilandzic and Foth (2013) Library as a place, Technology- enhanced learning, Library 2.0, Commons 2.0, Coworking, Urban informatics, library, User studies, Australia Learning Using library 2.0 tools in the library services and user learning User/End- user Service 4 Borgman (2000) Electronic publishing, library, Computer networks The application Web 2.0 tools for the electronic publishing Resource 6 Chen and Hsiang (2009) Generation and dissemination of Information, Digital storage, Academic Library, Open systems Web information seeking and retrieval in digital library contexts based on the artificially intelligent Resource Service 7 Chowdhury (1999) Internet, Information services, Information retrieval Digital library using context- awareness technology Resource Service 8 Connolly (2008) Artificial intelligence, Robotics, Pattern recognition The application mobile for library services Resource Service 9 Dimi�c et al. (2010) Cataloging Bibliographic standards Extensible Markup Language Information systems RFID integrated systems and libraries Technology 10 Dutton (2014) Open systems, Communication technologies, Surveillance, Internet of Things, Consumers, Social behavior Users and electronic libraries End-user 11 Frederick (2016) Data, library, Fourth Industrial Revolution Enterprise knowledge portals based on the industry 4.0 User 12 Gelfand (1998) Grey literature, Internet, Publishing Science : The application a roboti digital content: Breedbot Resource 13 Hahn (2012) Mobile computing, Library software, Augmented reality, Computer vision, Smartphones, Mobile applications, Library systems, Mobile communication systems, Citation Networked library services Service 14 Islam and Ikeda (2014) Knowledge management, Library services, Digital library, Online services, WWW Online digital reference Resource Service 15 Joint (2009) Communication technologies, University library, Worldwide web Sharing technology of the experiences in the library Technology 16 Kapoor et al. (2014) Adoption, TAM, RFID, Use Library Web site management Technology 17 Kesselman (2017) Library Education Technology Conferences Emerging Technologies Consumer electronics Virtual reference librarians (Chatbots) Service 18 Lam and Chan (2007) Archiving, Digital library Libraries, data and the fourth industrial revolution Recourse (continued) Table 3. Articles related to ES/ AI and Robot in the library (Emerald Insight) LHT No Author/s Keywords Application based on the reference Class 19 Liu (2011) Intelligent agents, Artificial intelligence, library, Information services, Digital library, Library systems Metadata and cataloging practices Service 20 McDonnell and Shiri (2011) Information retrieval, Search engines, Design, User interfaces, Worldwide web, Consumer behavior Information retrieval and user interface User/ End- user 21 Miglino et al. (2008) Robotics, Evolution, Education, Entertainment XML schema for UNIMARC and MARC 21 Service 22 Milella et al. (2008) Surveillance, Radio waves, Robotics, Environmental management, Workplace security Collaborative digital reference: An Ask a Librarian (UK) overview Service 23 Noh (2013) Context-aware computing, Next-generation digital library, Ubiquitous library, Context-awareness technology, Intelligent space, Sensor, library, Information systems What is trending in libraries from the Internet cybersphere–AI and other emerging technologies Resource Technology 24 Oyelude (2017) AI Emerging technologies The intelligent library: Thought leaders’ views on the likely impact of AI on academic libraries Technology 25 Park and Kim (2013) Long-term evolution, Technology acceptance, Perceived mobility, Perceived adaptively, System and service quality, Satisfaction Mobile communication systems, User satisfaction Libraries as coworking spaces: user motivations and social learning User/ End- user Service 26 Park et al. (2015) User studies, User satisfaction, Books Reading e-book devices Resource User/ End- user 27 Rudall (2006) Automation, Computers, Cybernetics Library automation Technology 28 Taha (2012) Networked library, Research information, Digital contents, Query processing, Academic library, Library networks AI and robotic hand-eye coordination in the library Resource 29 Vincze (2017) Chatbots Knowledge management in digital libraries Service 30 Wu et al. (2010) Digital library, Resources, Copyright law, Colleges, Students Digital libraries and resources Resource 31 Xiaobin and Jing (2009) Computer applications, Innovation, User interfaces, Communication, Digital library Teaching and exposing grey literature in the library User/ End- user 32 Yao et al. (2015) Artificial intelligence, Promotion, Participatory library service, Social networking, Talking robot, Virtual reference service Intelligent search agent in the library Service 33 Zimerman (2012) Digital natives , Search behaviour, Academic library, Millennials Information searches , Search engines, Searchers Requirements for information professionals in a digital environment User/ End- user Service Table 3. A review: intelligent libraries transmitted to them, and this opportunity provides an understanding of the user’s information behavior. Smart library agreements can facilitate the post-exchange process of information so that the librarian and user will no longer need to be involved. The exchange of information in intelligent libraries will help smart city innovation. In the following is stated the requirements for the development of an ES: (1) An expert of the problem available; (2) Experts have the time for the ES development project; (3) Experts can articulate their knowledge and methods; (4) The problem is not too complex, but knowledge intensive; (5) The problem is not poorly understood; (6) The problem requires cognitive skills only. Resource 21% Technology 13% User/ End User 29% Service 37% digital 1R el at iv e Fr eq ue nc ie s 0.000 0.001 0.002 0.003 0.004 0.005 2 3 4 5 6 7 8 9 10 information library search user digital information internet systemslibrary 1R el at iv e Fr eq ue nc ie s 2 3 4 5 6 7 8 9 10 0.000 0.005 0.010 0.015 0.020 Figure 5. The scope of the articles in field ES/AI and Robot in the library (Emerald Insight) Figure 6. Relative frequencies across the main categories in the abstracts of the articles and thematic interaction between them in the field of ES/ AI and robot and library (Emerald Insight) Figure 7. Relative frequencies across the main categories in the keywords of the articles and thematic interaction between them in the field of ES/ AI and robot and library (Emerald Insight) LHT Results are measurable and can be agreed upon by the experts (Lancaster, 1997). An obvious potential application of ES within libraries is for the selection of booksellers or other vendors of library materials carried to its logical conclusion. A system might be developed to select a vendor to automate ethical based on past performance in the supply of publications of a particular type such a capability would be especially valuable in the acquisition of material that is less routine-conference proceeding. A similar system, known as the Monographic Acquisitions consultant, was developed at Lowa State University in 1944. The system was designed to optimize the decision on which vendors are preferred. Types of monographs in the knowledge base of the system includes both descriptive and evaluative data on each supplier. Descriptive data deal with the type of publisher (foreign, university press, the publication of science materials) and relationship with the library (blanket order, approval plan, standing orders on the exchange list (Lakshmikant and Vishnu, 2008). Edelman (2006) studied about an intelligent design in the American Research Library Collections. Fourie (2003) investigated Current Awareness Services (CAS) in library acquisitions. Carneiro (2001) explained the role of intelligent resources in knowledge management. Switching Brains is a cloud-based Intelligent Resources Management (IRM) for the internet of Cognitive Things (IoCT) (Francisco and Arsenio, 2015). Other ESs, designed to help library users satisfy their own needs, have also included document- orders aid (Lakshmikant and Vishnu, 2008). Systems have also been designed within the library community to aid in the selection process, systems of this type have been discussed by Sowell (1989) and Meador and Cline (1992) (Lakshmikant and Vishnu, 2008). The term “referral system,” as used here, relates to systems that and are designed to refer library users to information sources likely to provide the answer to a particular question of the factuality of “information” type within the library community more work has been done on a system of this kind than on any other ES. Bailey (1992) studied about reference information system. He tries to help the user to select the appropriate printed and electronic references by ES. Other research has studied about the remote reference service. This study explores the possibilities afforded for collaborative reference work (Davenport et al., 1997). The application mostly such systems refer to users’ printed sources like Conventional reference books; but other types of sources, such as those in CD-ROM form, can also be included in the knowledge base. The objective of such systems is obvious: to guide library users with a reference suitable source when a librarian is not available to help them form reference, referral system cover knowledge as a whole in the coverage of a general reference library while others are restricted to the highly specialized domain (Mishra and Srivastava, 2008). One of the most important application of artificial intelligence in libraries is the use of recommender systems. Investigating user behavior in information retrieval and privatizing user services can effectively address his/her information needs in the least amount of time. The different techniques of recommender systems, the properties of these systems, and the evaluation’s methods of these systems, provide the best services and resources available to the user. Today, library systems include the hardware and software systems that are applied to different library process activities. As libraries present a huge content of printed materials, the automation of books handling becomes necessary (Comsa et al., 2014). The librarian robot contains manipulation that can identify books. It moves using RFID tags installed on the library shelves. “Jaume” is a librarian robot, which was developed in the Robotic Intelligence Lab of Universitat Jaume I. The new Bordeaux Public Library (June 1991). It offers a special video collection of more than 600 of the information resources. It uses a robot to give the users’ consultant service or reference service (Giannattasio and Bruckmann, 1992). UJI librarian robot searches and retrieves the requested books by users. The process starts when the user requests a book by its title or code, either through the Internet or by voice. The robot locates the book and gives it to the user. The book’s initial information is read by the robot’s vision system. This general application integrates several inter-disciplinary skills like path planning, visual perception or A review: intelligent libraries multisensory-based grasping, all linked together by reasoning capabilities (Prats et al., 2008; Prats et al., 2007; Ramos-Garijo et al., 2003). A librarian robot evaluates with the following criteria: (1) Reliable visual localization; (2) Robust and fault-tolerant force-guided extraction; (3) Performance adequate for books of different sizes and thicknesses; (4) Active book searching; (5) Combine navigation and active vision; (6) Fault-tolerant probabilistic strategy. In the context of robot librarians and AI has been investigated by limited number of studies. In the database of Emerald, only one article was found (Yao et al., 2015). They introduced a collaborative library service based on artificial intelligence. They developed an intelligent robot called Xiaotu (female). The task of this robot is to provide online reference services. Four factors are important in the success of this robot: artificial intelligence, self-learning, vivid logo and language, and modular architecture (Yao et al., 2011). Yuehu and Yanqing (2012) studied using the internet technology of objects. They have tried to look at smart sets along with the robot librarian. Then compare the smart library with other libraries. Kyrarini et al. (2017) presented a framework called “Skill Robot Library” (SRL). This framework has the authority to store key points of the route. In fact, this robot can store user’s behavior in information retrieval, and it will work based on this stored behavior. Behan and O’Keeffe (2009) designed as a mobile robotic assistant, called “LUCAS” for the University of Limerick. This assistant is a help system that supports users intellectually. Kim and Kohtaro (2009) tried to provide robots based on the structured data. This study introduces a conventional and intelligent environment for a librarian robot. This environment is based on RFID technology for these systems. In another study, reference services were investigated using the instant messaging (IM) smart robot. The Shanghai Jiaotong University Library is presented for example. This library provides the IM robot’s intelligent library service using BotPlatform (Yi et al., 2011). 8. Conclusion A review of the articles shows that we can use expert and intelligence systems in different library activities and information services. The main goal is to provide specialized services with the help of librarians and information resources specialists. Library services include technical and public services. Both categories use intelligent systems and ESs. These activities include the provision of information resources, the organization of information resources such as classification, indexing, and abstracting, the storage and retrieval of information from library systems, reference services, and circulation desk. We classified the scopes of the researches into four classes “technology,” “user,” “service,” and “resource.” A review of the articles shows that users’ information behavior is a very good way to design intelligent systems. The storing information in cloud and non-cloud spaces allow for the development of these systems. In big data and social networks where scientific information resources are exchanged, intelligent agents can play an important role. User profiles can be a good source for designing ES algorithms based on user knowledge. ESs are the most useable intelligent system in library and information science, which mimic librarian expert’s behaviors to support decision and management. However, individually using this technology is reduced in recent studies. Most information systems have a closed relation with the knowledge and opinion of experts. Using ES technologies such as inference engine and fuzzy LHT rule base may increase the accuracy of them. Therefore, the current information systems can be improved by integration with ES technologies. The librarian robot can reduce the usual and repetitive activities on library shelves. Almost the third of the articles in Emerald Insight in ES have related to the “user” scope, and in librarian robot (18% in WoS), most of the articles have related to the "service" in Emerald Insight and in WoS as well. One of our conclusion is that we need to further research in the area of smart resources. Note 1. https://www.emerald.com/insight References Albayrak, E. and Erensal, Y.C. (2004), “Using analytic hierarchy process (AHP) to improve human performance: an application of multiple criteria decision making problem”, Journal of Intelligent Manufacturing, Vol. 15 No. 4, pp. 491-503, doi: 10.1023/B:JIMS.0000034112.00652.4c. Amin, S.H. and Razmi, J. (2009), “An integrated fuzzy model for supplier management: a case study of ISP selection and evaluation”, Expert Systems with Applications, Vol. 36 No. 4, pp. 8639-8648, available at: http://isiarticles.com/bundles/Article/pre/pdf/19174.pdf. Asemi, A., Akbari, A., Kheradmandnia, M. and Farazi, A. (2012), “Investigating the level of desirability of information in management information systems in libraries in Iran”, The Electronic Library, Vol. 30 No. 6, pp. 833-843, doi: 10.1108/02640471211282136. Ataman, B.K. (2009), “Requirements for information professionals in a digital environment: some thoughts”, Program Electronic Library and Information Systems, Vol. 43 No. 2, pp. 215-228, doi: 10.1108/00330330910954415. Bailey, C.W. Jr. (1991), “Intelligent library systems: artificial intelligence technology and library automation systems”, in Hewitt, J.A. (Ed.) Advances in Library Automation and Networking, Vol. 4, JAI Press, Greenwich, CT, available at: http://www.digital-scholarship.org/cwb/intlibs. pdf (accessed 20 June 2018). Bailey, C.W. (1992), “The intelligent reference information system project. A merger of CD-ROM LAN and expert system technologies”, Information Technology and Libraries, Vol. 11 No. 3, p. 237, available at: https://www.learntechlib.org/p/146739/ (accessed 21 April 2019). Bavakutty, M. and Salih, M. (2006), Research on Library Computerisation, 1st ed., Ess Ess Publications. Behan, J. and O’Keeffe, D.T. (2009), “The development of an intelligent library assistant robot (Ireland)”, Journal of Korea Robotics Society, Vol. 4 No. 2, pp. 147-154. Bellman, R.E. and Zadeh, L.A. (1970), “Decision-making in a fuzzy environment”, Management Science, Vol. 17 No. 4, pp. B141-B164, doi: 10.1287/mnsc.17.4.B141. Bi, Z., Wang, G., Xu, L.D., Thompson, M., Mir, R., Nyikos, J., Mane, A., Witte, C. and Sidwell, C. (2017), “IoT-based system for communication and coordination of football robot team”, Internet Research, Vol. 27 No. 2, pp. 162-181, doi: 10.1108/IntR-02-2016-0056. Bilandzic, M. and Foth, M. (2013), “Libraries as coworking spaces: understanding user motivations and perceived barriers to social learning”, Library Hi Tech, Vol. 31 No. 2, pp. 254-273, doi: 10.1108/07378831311329040. Black, E.L. (2011), “Selecting a web content management system for an academic library website”, Information Technology and Libraries, Vol. 30 No. 4, pp. 185-189, doi: 10.6017/ital.v30i4.1869. Borgman, C.L. (2000), “Digital libraries and the continuum of scholarly communication”, Journal of Documentation, Vol. 56 No. 4, pp. 412-430, doi: 10.1108/EUM0000000007121. Cao, G., Liang, M. and Li, X. (2018), “How to make the library smart? The conceptualization of the smart library”, The Electronic Library, Vol. 36 No. 5, pp. 811-825, doi: 10.1108/EL-11- 2017-0248. A review: intelligent libraries https://www.emerald.com/insight https://doi.org/10.1023/B:JIMS.0000034112.00652.4c http://isiarticles.com/bundles/Article/pre/pdf/19174.pdf https://doi.org/10.1108/02640471211282136 https://doi.org/10.1108/00330330910954415 http://www.digital-scholarship.org/cwb/intlibs.pdf http://www.digital-scholarship.org/cwb/intlibs.pdf https://www.learntechlib.org/p/146739/ https://doi.org/10.1287/mnsc.17.4.B141 https://doi.org/10.1108/IntR-02-2016-0056 https://doi.org/10.1108/07378831311329040 https://doi.org/10.6017/ital.v30i4.1869 https://doi.org/10.1108/EUM0000000007121 https://doi.org/10.1108/EL-11-2017-0248 https://doi.org/10.1108/EL-11-2017-0248 Carneiro, A. (2001), “The role of intelligent resources in knowledge management”, Journal of Knowledge Management, Vol. 5 No. 4, pp. 358-367, doi: 10.1108/eum0000000006533. Chen, K. and Hsiang, J. (2009), “The unique approach to institutional repository, practice of National University”, The Electronic Library, Vol. 27 No. 2, pp. 204-221, doi: 10.1108/ 02640470910947566. Chowdhury, G.G. (1999), “The Internet and information retrieval research: a brief review”, Journal of Documentation, Vol. 55 No. 2, pp. 209-225, doi: 10.1108/EUM0000000007144. Chu, H.C., Hwang, G.J. and Tseng, J.C.R. (2010), “An innovative approach for developing and employing electronic libraries to support context-aware ubiquitous learning”, The Electronic Library, Vol. 28 No. 6, pp. 873-890, doi: 10.1108/02640471011093552. Comsa, A., Maniu, I., Modler, N., Lovasz, E.C. and Ciupe, V. (2014), “Automated book manipulator in libraries”, in Pisla, D., Bleuler, H., Rodic, A., Vaida, C. and Pisla, A. (Eds), New Trends in Medical and Service Robots. Mechanisms and Machine Science, Vol. 16, Springer, Cham, doi: 10.1007/978- 3-319-01592-7_6. Connolly, C. (2008), “Artificial intelligence and robotic hand-eye coordination”, Industrial Robot: International Journal, Vol. 35 No. 6, pp. 496-503, doi: 10.1108/01439910810909484. Cox, A.M., Pinfield, S. and Rutter, S. (2018), “The intelligent library: thought leaders’ views on the likely impact of artificial intelligence on academic libraries”, Library Hi Tech, Vol. 37 No. 3, pp. 418-435, doi: 10.1108/LHT-08-2018-0105. Davenport, E., Procter, R. and Goldenberg, A. (1997), “Distributed expertise: remote reference service on a metropolitan area network”, The Electronic Library, Vol. 15 No. 4, pp. 271-278, doi: 10.1108/ eb045567de. De Prado, R.L. (2000), “Do users dream of electronic libraries?”, The Electronic Library, Vol. 18 No. 3, pp. 202-209, doi: 10.1108/02640470010337508. Denning, R. and Smith, P.J. (1994), “Interface design concepts in the development of Elsa, an intelligent electronic library search assistant”, Information Technology and Libraries, Vol. 13 No. 2, pp. 133-147. Devadason, F.J. and Vespry, H.A. (1996), “LISPA (library and information center staff planning advisor): a microcomputer-based system”, Information Technology and Libraries, Vol. 15 No. 2, pp. 105-112. Dimi�c, B., Milosavljevi�c, B. and Surla, D. (2010), “XML Schema for UNIMARC and MARC 21”, The Electronic Library, Vol. 37 No. 3, pp. 418-435, doi: 10.1108/02640471011033611. Ding, H. and Sølvberg, I. (2007), “Rule-based metadata interoperation in heterogeneous digital libraries”, The Electronic Library, Vol. 25 No. 2, pp. 193-206, doi: 10.1108/02640470710741322. Du, M., Fang, J. and Wang, L. (2011), “A parameter self-tuning fuzzy-PID control system for pneumatic manipulator of library robot”, Paper presented at the International Conference on Electronics, Communications and Control (ICECC), 9-11 Sep, Ningbo, China, doi: 10.1109/ICECC.2011. 6067700. Dutton, W.H. (2014), “Putting things to work: social and policy challenges for the internet of things”, Info, Vol. 16 No. 3, pp. 1-21, doi: 10.1108/info-09-2013-0047. Dwivedi, Y.K., Kapoor, K.K., Williams, M.D. and Williams, J. (2013), “RFID systems in libraries: an empirical examination of factors affecting system use and user satisfaction”, International Journal of Information Management, Vol. 33 No. 2, pp. 367-377, doi: 10.1016/j.ijinfomgt.2012. 10.008. Edelman, H. (2006), “Intelligent design and the evolution of American research library collections”, Library Resources and Technical Services, Vol. 50 No. 4, pp. 234-238, doi: 10.5860/lrts.50n4.234. Fourie, I. (2003), “How can current awareness services (CAS) be used in the world of library acquisitions?”, Online Information Review, Vol. 27 No. 3, pp. 183-195, doi: 10.1108/ 14684520310481409. LHT https://doi.org/10.1108/eum0000000006533 https://doi.org/10.1108/02640470910947566 https://doi.org/10.1108/02640470910947566 https://doi.org/10.1108/EUM0000000007144 https://doi.org/10.1108/02640471011093552 https://doi.org/10.1007/978-3-319-01592-7_6 https://doi.org/10.1007/978-3-319-01592-7_6 https://doi.org/10.1108/01439910810909484 https://doi.org/10.1108/LHT-08-2018-0105 https://doi.org/10.1108/eb045567de https://doi.org/10.1108/eb045567de https://doi.org/10.1108/02640470010337508 https://doi.org/10.1108/02640471011033611 https://doi.org/10.1108/02640470710741322 https://doi.org/10.1109/ICECC.2011.6067700 https://doi.org/10.1109/ICECC.2011.6067700 https://doi.org/10.1108/info-09-2013-0047 https://doi.org/10.1016/j.ijinfomgt.2012.10.008 https://doi.org/10.1016/j.ijinfomgt.2012.10.008 https://doi.org/10.5860/lrts.50n4.234 https://doi.org/10.1108/14684520310481409 https://doi.org/10.1108/14684520310481409 Francisco, R. and Arsenio, A.M. (2015), “Switching brains: cloud-based intelligent resources management for the internet of cognitive things”, EAI Endorsed Transactions on Cognitive Communications, Vol. 1 No. 2, pp. 1-13, doi: 10.4108/cogcom.1.2.e4. Frederick, D. E. (2016), “Libraries, data and the fourth industrial revolution (Data Deluge Column)”, Library Hi Tech News, Vol. 33 No. 5, pp. 9-12, doi: 10.1108/LHTN-05-2016-0025. Gelfand, J. (1998), “Teaching and exposing grey literature: what the information profession needs to know-examples from the sciences”, Collection Building, Vol. 17 No. 4, pp. 159-166, doi: 10.1108/ 01604959810238301. Giannattasio, I. and Bruckmann, D. (1992), “An update of audiovisual and new technology in French libraries”, IFLA Journal-International Federation of Library Associations, Vol. 18 No. 3, pp. 252-257, doi: 10.1177/034003529201800313. Golub, K., Tudhope, D., Zeng, M.L. and Zumer, M. (2014), “Terminology registries for knowledge organization systems: functionality, use, and attributes”, Journal of the Association for Information Science and Technology, Vol. 65 No. 9, pp. 1901-1916, doi: 10. 1002/asi.23090. Grigorescu, S.M., Natarajan, S.K., Mronga, D. and Graser, A. (2010), “Robust feature extraction for 3D reconstruction of boundary segmented objects in a robotic library scenario”, Paper presented at the International Conference on Intelligent Robots and Systems (IROS), IEEE/RSJ, 18-22 Oct. Taipei, Taiwan. Hahn, J. (2012), “Mobile augmented reality applications for library services”, New Library World, Vol. 113 Nos 9/10, pp. 429-438, doi: 10.1108/03074801211273902. Heyer, S., Enjarini, B., Fragkopoulos, C. and Graeser, A. (2012), “Book detection and grasping in library scenario”, Paper presented at the ROBOTIK 2012; 7th German Conference on Robotics, 21-22 May, Munich, Germany. Hsieh, C.C. and Hall, W. (1989), “Survey of artificial-intelligence and expert systems in library and information-science literature”, Information Technology and Libraries, Vol. 8 No. 2, pp. 209-214, available at: https://www.learntechlib.org/p/170528/ (accessed 21 April 2019). Huili, F. and Bo, Sh. (2017), “A review on researches and practices of library robots at home and abroad”, Library Journal, Vol. 36 No. 6, pp. 88-94, available at: http://www.libraryjournal.com. cn/EN/Y2017/V36/I6/88 (accessed 21 April 2019). Hwang, G.J., Chu, H.C., Lin, Y.S. and Tsai, C.C. (2011), “A knowledge acquisition approach to developing Mindtools for organizing and sharing differentiating knowledge in a ubiquitous learning environment”, Computers and Education, Vol. 57 No. 1, pp. 1368-1377, doi: 10.1016/j. compedu.2010.12.013. Iglesias, E. (2013), Robots in Academic Libraries: Advancements in Library Automation, Central Connecticut State University, USA, IGI Global, doi: 10.4018/978-1-4666-3938-6. Islam, M.A. and Ikeda, M. (2014), “Convergence issues of knowledge management in digital libraries: steps towards state-of-the-art digital libraries”, Vine: The Journal of Information, Vol. 44 No. 1, pp. 140-159, doi: 10.1108/VINE-05-2013-0029. Ismail, M.A. and Kareem, S.A. (2011), “Identifying how novice researchers search, locate, choose and use web resources at the early stage of research”, Malaysian Journal of Library and Information Science, Vol. 16 No. 3, pp. 67-85. Joint, N. (2009), “The Web 2.0 challenge to libraries”, Library Review, Vol. 58 No. 3, pp. 167-175, doi: 10. 1108/00242530910942027. Kao, S.C. and Wu, C. (2012), “PIKIPDL A personalized information and knowledge integration platform for DL service”, Library Hi Tech, Vol. 30 No. 3, pp. 490-512, doi: 10.1108/ 07378831211266627. Kapoor, K., Dwivedi, Y., Piercy, N.C., Lal, B. and Weerakkody, V. (2014), “RFID integrated systems in libraries: extending TAM model for empirically examining the use”, Journal of Enterprise Information Management, Vol. 27 No. 6, pp. 731-758, doi: 10.1108/JEIM-10-2013-0079. A review: intelligent libraries https://doi.org/10.4108/cogcom.1.2.e4 https://doi.org/10.1108/LHTN-05-2016-0025 https://doi.org/10.1108/01604959810238301 https://doi.org/10.1108/01604959810238301 https://doi.org/10.1177/034003529201800313 https://doi.org/10.1002/asi.23090 https://doi.org/10.1002/asi.23090 https://doi.org/10.1108/03074801211273902 https://www.learntechlib.org/p/170528/ http://www.libraryjournal.com.cn/EN/Y2017/V36/I6/88 http://www.libraryjournal.com.cn/EN/Y2017/V36/I6/88 https://doi.org/10.1016/j.compedu.2010.12.013 https://doi.org/10.1016/j.compedu.2010.12.013 https://doi.org/10.4018/978-1-4666-3938-6 https://doi.org/10.1108/VINE-05-2013-0029 https://doi.org/10.1108/00242530910942027 https://doi.org/10.1108/00242530910942027 https://doi.org/10.1108/07378831211266627 https://doi.org/10.1108/07378831211266627 https://doi.org/10.1108/JEIM-10-2013-0079 Kesselman, M. (2017), “Conference report: 50th consumer electronics show”, Library Hi Tech News, Vol. 34 No. 3, pp. 1-8, doi: 10.1108/LHTN-03-2017-0014. Kim, B.K. and Kohtaro, O. (2009), “Information structured space and ambient intelligent systems for a librarian robot”, The Journal of Korea Robotics Society, Vol. 4, pp. 147-154, available at: http:// www.koreascience.or.kr/article/JAKO200917161877604. Kim, B.K., Ohara, K., Kitagaki, K., Ohba, K. and Sugawara, T. (2008), “Design of ubiquitous space for the robotic library system and its application”, IFAC Proceedings Volumes, Vol. 41 No. 2, pp. 8221-8225, doi: 10.3182/20080706-5-KR-1001.01391. Kim, B.K., Ohara, K., Kitagaki, K. and Ohba, K. (2009), “Design and control of librarian robot system in information structured environments”, Journal of Robotics and Mechatronics, Vol. 21 No. 4, pp. 507-514. Kim, K.J., Park, E. and Shyam Sundar, S. (2013), “Caregiving role in human–robot interaction: a study of the mediating effects of perceived benefit and social presence”, Computers in Human Behavior, Vol. 29 No. 4, pp. 1799-1806, doi: 10.1016/j.chb.2013.02.009. Kruk, S.R. and Krawczyk, H. (2004), “Intelligent resources search in virtual libraries”, in Intelligent Information Processing and Web Mining, Advances in Soft Computing, Vol 25, Springer, Berlin, Heidelberg, pp. 439-443, doi: 10.1007/978-3-540-39985-8_49. Kyrarini, M., Naeem, S., Wang, X. and Gr€aser, A. (2017), “Skill robot library: intelligent path planning framework for object manipulation”, 25th European Signal Processing Conference (EUSIPCO), Kos, 2017, pp. 2398-2402, doi: 10.23919/EUSIPCO.2017.8081640. Lakshmikant, M. and Vishnu, S. (2008), Automation and Networking of Libraries (Electronic Source): A Manual of Library Management Software and Applications of Computer Technology in Libraries, New Age International, India. Lam, K.T. and Chan, D.L.H. (2007), “Building an institutional repository: sharing experiences at the HKUST Library”, OCLC Systems and Services: International Digital Library Perspectives, Vol. 23 No. 3, pp. 310-323, doi: 10.1108/10650750710776440. Lancaster, F.W. and Smith, L.C. (1990), Artificial Intelligence and Expert Systems: Will They Change the Library?, 27th Clinic on Library Applications of Data Processing March 25-27, at the University of Illinois at Urbana- Champaign. From IDEALS, the Illinois Digital Environment for Access to Learning and Scholarship, available at: https://www.ideals.illinois.edu/handle/ 2142/1209. Lancaster, F.W. (1997), Artificial Intelligence and Expert System Technologies: Prospects. Libraries for the New Millennium: Implications for Managers, Library Association, London, pp. 19-38. Leslie, D. (1999), “Industry experts gather to map future of libraries”, The Electronic Library, Vol. 17 No. 3, pp. 149-153, doi: 10.1108/02640479910329734. Lin, W., Yueh, H.P., Wu, H.Y. and Fu, L.C. (2013), “Developing a service robot for a children’s library: a design-based research approach”, Journal of the Association for Information Science and Technology, Vol. 65 No. 2, pp. 290-301, doi: 10.1002/asi.22975. Liu, G. (2011), “The application of intelligent agents in libraries: a survey”, Program, Vol. 45 No. 1, pp. 78-97, doi: 10.1108/00330331111107411. McDonnell, M. and Shiri, A. (2011), “Social search: a taxonomy of, and a user-centred approach to, social web search”, Program, Vol. 45 No. 1, pp. 6-28, doi: 10.1108/00330331111107376. Meador, J.M. and Cline, L. (1992), “Displaying and utilizing selection tools in a user-friendly electronic environment”, Library Acquisitions: Practice & Theory, Vol. 16 No. 3, pp. 289-294. Mehtab Alam Ansari, A. (2008), “Awareness and use of OPACs in five Delhi libraries”, The Electronic Library, Vol. 26 No. 1, pp. 111-129, doi: 10.1108/02640470810851789. Mei, Zh., Chen, Y., Jiang, M., Wu, H. and Cheng, L. (2017), “Mobile robots path planning based on dynamic movement primitives library”, Control Conference (CCC), 2017 36th Chinese, 26-28 July, doi: 10.23919/ChiCC.2017.8028446. LHT https://doi.org/10.1108/LHTN-03-2017-0014 http://www.koreascience.or.kr/article/JAKO200917161877604 http://www.koreascience.or.kr/article/JAKO200917161877604 https://doi.org/10.3182/20080706-5-KR-1001.01391 https://doi.org/10.1016/j.chb.2013.02.009 https://doi.org/10.1007/978-3-540-39985-8_49 https://doi.org/10.23919/EUSIPCO.2017.8081640 https://doi.org/10.1108/10650750710776440 https://www.ideals.illinois.edu/handle/2142/1209 https://www.ideals.illinois.edu/handle/2142/1209 https://doi.org/10.1108/02640479910329734 https://doi.org/10.1002/asi.22975 https://doi.org/10.1108/00330331111107411 https://doi.org/10.1108/00330331111107376 https://doi.org/10.1108/02640470810851789 https://doi.org/10.23919/ChiCC.2017.8028446 Miglino, O., Gigliotta, O., Ponticorvo, M. and Nolfi, S. (2008), “Breedbot: an evolutionary robotics application in digital content”, The Electronic Library, Vol. 26 No. 3, pp. 363-373, doi: 10.1108/ 02640470810879509. Mikawa, M. (2010), A Practicum Track Using Librarian Robot in Support Program for Contemporary Educational Needs, University of Tsukuba, Japan, available at: http://www.iiis.org/CDs2010/ CD2010SCI/SCI_2010/PapersPdf/SA431IB.pdf (accessed 21 June 2018). Milella, A., Cicirelli, G. and Distante, A. (2008), “RFID-assisted mobile robot system for mapping and surveillance of indoor environments”, Industrial Robot, Vol. 35 No. 2, pp. 143-152, doi: 10.1108/ 01439910810854638. Mishra, L. and Srivastava, V. (2008), Automation and Networking of Libraries [Electronic Resource]: A Manual of Library Management Software and Applications of Computer Technology in Libraries, New Age International. Modler, N., Hufenbach, W., Comsa, A., Maniu, I., Zichner, M. and Friedrich, J. (2012), “Compliant structures in book handling applications”, Applied Mechanics and Materials, Vol. 162, pp. 543-548. Modler, N., Comsa, A., Maniu, I., Lovasz, E.C. and Ciupe, V. (2014), “Dedicated gripper for books handling in a library”, in Visa, I. (Eds), The 11th IFToMM International Symposium on Science of Mechanisms and Machines. Mechanisms and Machine Science, Vol. 18, Springer, Cham, doi: 10.1007/978-3-319-01845-4_51. Noh, Y. (2013), “A study on next-generation digital library using context-awareness technology”, Library Hi Tech, Vol. 31 No. 2, pp. 236-253, doi: 10.1108/07378831311329031. O’Neill, M. and Morris, A. (1989), “The contribution of library and information science to expert system development”, The Electronic Library, Vol. 7 No. 5, pp. 295-300, doi: 10.1108/eb044908. Oyelude, A.A. (2017), “What’s trending in libraries from the internet cybersphere–artificial intelligence and other emerging technologies”, Library Hi Tech News, Vol 34 No. 2, pp. 11-12, doi: 10.1108/ LHTN-02-2017-0008. Park, E. and Kim, K.J. (2013), “User acceptance of long-term evolution (LTE) services: an application of extended technology acceptance model”, Program, Vol. 47 No. 2, pp. 188-205, doi: 10.1108/ 00330331311313762. Park, E., Sung, J. and Cho, K. (2015), “Reading experiences influencing the acceptance of E-book devices”, The Electronic Library, Vol. 33 No. 1, pp. 120-135, doi: 10.1108/EL-05-2012-0045. Payne, G.F. and Bradbury, D. (2002), “An automated approach to online digital reference: the Open University Library OPAL Project”, Program-Electronic Library and Information Systems, Vol. 36 No. 1, pp. 5-12, doi: 10.1108/00330330210426076. Phillips, D. (2017), “Robots in the Library: gauging attitudes towards developments in robotics and ai, and the potential implications for library services”, Dissertation, University of London, doi: 10. 17613/M6535M. Prats, M., Sanz, P.J., Del Pobil, A.P., Mart�ınez, E. and Mar�ın, R. (2007), “Towards multipurpose autonomous manipulation with the UJI service robot”, Robotica, Vol. 25 No. 2, pp. 245-256, doi: 10.1017/S0263574706003304. Prats, M., Mart�ınez, E., Sanz, P.J. and Del Pobil, A.P. (2008), “The UJI librarian robot”, Intel Serv Robotics, Vol. 1 No. 4, pp. 321-335, doi: 10.1007/s11370-008-0028-1. Ramos-Garijo, R., Prats, M., Sanz, P.J. and Del Pobil, A.P. (2003), “An autonomous assistant robot for book manipulation in a library”, Paper presented at the IEEE International Conference on Systems, Man and Cybernetics, 8-8 October, USA: Washington DC, doi: 10.1109/ICSMC.2003.1244499. Rudall, B.H. (2006), “Contemporary systems and cybernetics”, Kybernetes, Vol. 35 Nos 1/2, pp. 209-216, doi: 10.1108/03684920610640326. Sacchanand, C. and Jaroenpuntaruk, V. (2006), “Development of a web-based self-training package for information retrieval using the distance education approach”, The Electronic Library, Vol. 24 No. 4, pp. 501-516, doi: 10.1108/02640470610689197. A review: intelligent libraries https://doi.org/10.1108/02640470810879509 https://doi.org/10.1108/02640470810879509 http://www.iiis.org/CDs2010/CD2010SCI/SCI_2010/PapersPdf/SA431IB.pdf http://www.iiis.org/CDs2010/CD2010SCI/SCI_2010/PapersPdf/SA431IB.pdf https://doi.org/10.1108/01439910810854638 https://doi.org/10.1108/01439910810854638 https://doi.org/10.1007/978-3-319-01845-4_51 https://doi.org/10.1108/07378831311329031 https://doi.org/10.1108/eb044908 https://doi.org/10.1108/LHTN-02-2017-0008 https://doi.org/10.1108/LHTN-02-2017-0008 https://doi.org/10.1108/00330331311313762 https://doi.org/10.1108/00330331311313762 https://doi.org/10.1108/EL-05-2012-0045 https://doi.org/10.1108/00330330210426076 https://doi.org/10.17613/M6535M https://doi.org/10.17613/M6535M https://doi.org/10.1017/S0263574706003304 https://doi.org/10.1007/s11370-008-0028-1 https://doi.org/10.1109/ICSMC.2003.1244499 https://doi.org/10.1108/03684920610640326 https://doi.org/10.1108/02640470610689197 Sharda, S., Delen, D., Turban, T. and King, K. (2017), Business Intelligence: A Managerial Approach, Global Edition, Pearson, available at: https://www.pearson.com/uk/educators/higher- education-educators/product/Sharda-Business-Intelligence-A-Managerial-Approach-Global- Edition-4th-Edition/9781292220543.html (accessed April 13, 2020). Singh, D.K., Singh, B.K. and Dubey, Y.P. (1996), “Expert systems and their application in library and information systems”, DESIDOC Bulletin of Information Technology, Vol. 16 No. 4, pp. 9-12, available at: http://publications.drdo.gov.in/ojs/index.php/djlit/article/download/ 3272/1727/ (accessed 21 April 2019). Sowell, S.L. (1989), “Expanding horizons in collection development with expert systems: development and testing of a demonstration prototype”, Special Libraries, Vol. 80 No. 4, pp. 45-50. Taha, A. (2012), “Networked library services in a research-intensive university”, The Electronic Library, Vol. 30 No. 6, pp. 844-856, doi: 10.1108/02640471211282145. Vincze, J. (2017), “Virtual reference librarians (Chatbots)”, Library Hi Tech News, Vol. 34 No. 4, pp. 5-8, doi: 10.1108/LHTN-03-2017-0016. Waters, S.T. (1986), “Answerman, the expert information specialist - an expert system for retrieval of information from library reference-books”, Information Technology and Libraries, Vol. 5 No. 3, pp. 204-212. Weiss, P.J. (1994), “The expert cataloging assistant project at the national-library-of-medicine”, Information Technology and Libraries, Vol. 13 No. 4, pp. 267-271. Wu, H.C., Chou, C., Ke, H.R. and Wang, M.H. (2010), “College students’ misunderstandings about copyright laws for digital library resources”, The Electronic Library, Vol. 28 No. 2, pp. 197-209, doi: 10.1108/02640471011033576. Xiaobin, L. and Jing, G. (2009), “Innovation community: constructing a new service mode for academic libraries”, The Electronic Library, Vol. 27 No. 2, pp. 258-270, doi: 10.1108/ 02640470910947601. Yao, F., Lei, J., Chengyu, Z. and Wu, Ch. (2011), “New attempt on real-time virtual reference service: the smart chat robot of tsinghua university library”, New Technology of Library and Information Service, Vol. 27 No. 4, pp. 77-81, available at: http://manu44.magtech.com.cn/Jwk_infotech_wk3/ EN/10.11925/infotech.1003-3513.2011.04.13. Yao, F., Zhang, Ch. and Chen, W. (2015), “Smart talking robot Xiaotu: participatory library service based on artificial intelligence”, Library Hi Tech, Vol. 33 No. 2, pp. 245-260, doi: 10.1108/LHT-02- 2015-0010. Yi, S., Bao, L. and Jianfeng, Q. (2011), “Design and implementation of library intelligent IM reference robot”, New Technology of Library and Information Service, Vol. 27 No. 5, pp. 88-92, available at: http://manu44.magtech.com.cn/Jwk_infotech_wk3/EN/10.11925/infotech.1003-3513.2011.05.14. Yuehu, W. and Yanqing, L. (2012), “An intelligent library based on the IOT and the robot technology”, Library Work and Study, Vol. 3 No. 5, pp. 250-276, available at: http://en.cnki.com.cn/Article_en/ CJFDTOTAL-TSGG201203006.htm. Zimerman, M. (2012), “Digital natives, searching behavior and the library”, New Library World, Vol. 113 Nos 3/4, pp. 174-201 doi: 10.1108/03074801211218552. Further reading Berube, L. (2004), “Collaborative digital reference: an ask a librarian (UK) overview”, Program, doi: 10. 1108/00330330410519189?fullSc51. Clyde, L.A. (2000), “A strategic planning approach to Web site management”, The Electronic Library, Vol. 18 No. 2, pp. 97-108, doi: 10.1108/02640470010325637. Detlor, B. and Arsenault, C. (2002), “Web information seeking and retrieval in digital library contexts: towards an intelligent agent solution”, Online Information Review, Vol. 26 No. 6, pp. 404-412, doi: 10.1108/14684520210452736. LHT https://www.pearson.com/uk/educators/higher-education-educators/product/Sharda-Business-Intelligence-A-Managerial-Approach-Global-Edition-4th-Edition/9781292220543.html https://www.pearson.com/uk/educators/higher-education-educators/product/Sharda-Business-Intelligence-A-Managerial-Approach-Global-Edition-4th-Edition/9781292220543.html https://www.pearson.com/uk/educators/higher-education-educators/product/Sharda-Business-Intelligence-A-Managerial-Approach-Global-Edition-4th-Edition/9781292220543.html http://publications.drdo.gov.in/ojs/index.php/djlit/article/download/3272/1727/ http://publications.drdo.gov.in/ojs/index.php/djlit/article/download/3272/1727/ https://doi.org/10.1108/02640471211282145 https://doi.org/10.1108/LHTN-03-2017-0016 https://doi.org/10.1108/02640471011033576 https://doi.org/10.1108/02640470910947601 https://doi.org/10.1108/02640470910947601 http://manu44.magtech.com.cn/Jwk_infotech_wk3/EN/10.11925/infotech.1003-3513.2011.04.13 http://manu44.magtech.com.cn/Jwk_infotech_wk3/EN/10.11925/infotech.1003-3513.2011.04.13 https://doi.org/10.1108/LHT-02-2015-0010 https://doi.org/10.1108/LHT-02-2015-0010 http://manu44.magtech.com.cn/Jwk_infotech_wk3/EN/10.11925/infotech.1003-3513.2011.05.14 http://en.cnki.com.cn/Article_en/CJFDTOTAL-TSGG201203006.htm http://en.cnki.com.cn/Article_en/CJFDTOTAL-TSGG201203006.htm https://doi.org/10.1108/03074801211218552 https://doi.org/10.1108/00330330410519189?fullSc=1 https://doi.org/10.1108/00330330410519189?fullSc=1 https://doi.org/10.1108/00330330410519189?fullSc=1 https://doi.org/10.1108/02640470010325637 https://doi.org/10.1108/14684520210452736 El-Sherbini, M. and Klim, G. (2004), “Metadata and cataloging practices”, The Electronic Library, Vol. 22 No. 3, pp. 238-248, doi: 10.1108/02640470410541633. MacDougall, J., Brittain, J.M. and Gann, R. (1996), “Health informatics: an overview”, Journal of Documentation, Vol. 52 No. 4, pp. 421-448, doi: 10.1108/eb026974. Mardikian, J. (1995), “Self-service charge systems: current technological applications and their implications for the future library”, Reference Services Review, Vol. 23 No. 4, pp. 19-38, doi: 10. 1108/eb049262. Morales, E.R. (1999), “Generis: the ec-jrc generalised software control system for industrial robots”, Industrial Robot: International Journal, Vol. 26 No. 1, pp. 26-32, doi: 10.1108/ 01439919910250197. Omekwu, C.O. (2006), “African culture and libraries: the information technology challenge”, The Electronic Library, Vol. 24 No. 2, pp. 243-264, doi: 10.1108/02640470610649218. Raitt, D.I. (1985), “Look—no paper! the library of tomorrow”, The Electronic Library, Vol. 3 No. 4, pp. 276-289, doi: 10.1108/eb044666. Reneker, M.H. and Buntzen, J.L. (2000), “Enterprise knowledge portals: two projects in the United States Department of the Navy”, The Electronic Library, Vol. 18 No. 6, pp. 392-403, doi: 10.1108/ EUM0000000005386. Tegenbos, J. and Nieuwenhuysen, P. (1997), “My kingdom for an agent? evaluation of autonomy, an intelligent search agent for the internet”, Online and CD-Rom Review, Vol. 21 No. 3, pp. 139-148, doi: 10.1108/eb024616. Corresponding author Asefeh Asemi and Mohsen Nowkarizi can be coneccted at: asemi.asefeh@uni-corvinus.hu For instructions on how to order reprints of this article, please visit our website: www.emeraldgrouppublishing.com/licensing/reprints.htm Or contact us for further details: permissions@emeraldinsight.com A review: intelligent libraries https://doi.org/10.1108/02640470410541633 https://doi.org/10.1108/eb026974 https://doi.org/10.1108/eb049262 https://doi.org/10.1108/eb049262 https://doi.org/10.1108/01439919910250197 https://doi.org/10.1108/01439919910250197 https://doi.org/10.1108/02640470610649218 https://doi.org/10.1108/eb044666 https://doi.org/10.1108/EUM0000000005386 https://doi.org/10.1108/EUM0000000005386 https://doi.org/10.1108/eb024616 mailto:asemi.asefeh@uni-corvinus.hu Intelligent libraries: a review on expert systems, artificial intelligence, and robot Introduction Artificial intelligence Intelligent systems Expert systems Methodology Findings ES/ AI and robots’ application in the library (WoS) ES and robot’s application in the library using Emerald Insight [1] Discussion Conclusion Note References Further reading 
work_ia4kyatfrndlniinjocs6tmqnm	----	[PDF] Real-time fault diagnosis using knowledge-based expert system | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1016/J.PSEP.2007.10.014 Corpus ID: 17900554Real-time fault diagnosis using knowledge-based expert system @article{Nan2008RealtimeFD, title={Real-time fault diagnosis using knowledge-based expert system}, author={C. Nan and F. Khan and M. T. Iqbal}, journal={Process Safety and Environmental Protection}, year={2008}, volume={86}, pages={55-71} } C. Nan, F. Khan, M. T. Iqbal Published 2008 Engineering Process Safety and Environmental Protection Abstract Abnormal operating conditions (faults) cost process industry billons of dollars per year and can be prevented if they are predicted and controlled in advance. Advanced software applications, based on the expert system, has the potential to assist engineers in monitoring, detecting, and diagnosing abnormal conditions and thus providing safe guards against these unexpected process conditions. Abnormal operating conditions (faults) could be modeled and predicted with high confidence using… Expand View via Publisher lue.ethz.ch Save to Library Create Alert Cite Launch Research Feed Share This Paper 79 CitationsHighly Influential Citations 1 Background Citations 23 Methods Citations 20 View All Figures and Tables from this paper figure 1 table 1 figure 2 table 2 figure 3 table 3 figure 4 figure 5 figure 6 table 7 figure 8 table 8 figure 9 table 9 figure 10 figure 11 figure 12 figure 13 figure 14 figure 15 figure 16 figure 17 figure 18 figure 19 figure 21 figure 23 figure 24 figure 25 figure 26 figure 27 View All 30 Figures & Tables 79 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency Fault diagnosis and cause analysis using fuzzy evidential reasoning approach and dynamic adaptive fuzzy Petri nets Hu-Chen Liu, Qing-Lian Lin, Ming-Lun Ren Engineering, Computer Science Comput. Ind. Eng. 2013 51 Save Alert Research Feed Risk‐based fault diagnosis and safety management for process systems Hui, Z. Bao, F. Khan, T. Iqbal, Yan-jun Chang Engineering 2010 18 PDF Save Alert Research Feed RDR-based knowledge based system to the failure detection in industrial cyber physical systems Dohyeong Kim, S. Han, Y. Lin, B. Kang, S. Lee Computer Science Knowl. Based Syst. 2018 13 Save Alert Research Feed A hybrid fault diagnosis method for mechanic-electronic-hydraulic control system based on simulated knowledge from virtual prototyping Deyu He, N. Hu, L. Hu Engineering 2016 View 1 excerpt, cites methods Save Alert Research Feed Design of a real-time fault diagnosis expert system for the EAST cryoplant Zhiwei Zhou, M. Zhuang, Xiaofei Lu, L. Hu, G. Xia Computer Science 2012 18 Save Alert Research Feed FTA-based fault diagnose expert system for hydraulic equipments Mengmeng Bian, J. Shi, S. Wang Engineering Proceedings of 2011 International Conference on Fluid Power and Mechatronics 2011 8 PDF View 1 excerpt, cites methods Save Alert Research Feed Knowledge acquisition development in failure diagnosis analysis as an interactive approach M. Yazdi, Hamzeh Soltanali Computer Science 2019 17 Save Alert Research Feed Process fault diagnosis with model- and knowledge-based approaches: Advances and opportunities Weijun Li, H. Li, Sai Gu, Tao Chen Computer Science 2020 1 Save Alert Research Feed Fault Detection and Isolation In Gas Turbine Engines Zakieh Sadough Engineering 2012 2 Highly Influenced PDF View 18 excerpts, cites methods Save Alert Research Feed A Novel Framework for Real-Time Fault Diagnosis Based on Dynamic Fault Tree Analysis R. Duan, Xiudong Ou Computer Science 2013 1 PDF View 2 excerpts, cites background Save Alert Research Feed ... 1 2 3 4 5 ... References SHOWING 1-10 OF 22 REFERENCES SORT BYRelevance Most Influenced Papers Recency Intelligent system for process supervision and fault diagnosis in dynamic physical systems C. Lo, Y. K. Wong, A. Rad Computer Science, Engineering IEEE Transactions on Industrial Electronics 2006 61 View 2 excerpts, references background Save Alert Research Feed Fuzzy rule-based expert system for power system fault diagnosis H. Monsef, A. Ranjbar, S. Jadid Computer Science 1997 62 View 2 excerpts, references background and methods Save Alert Research Feed Fault Detection and Isolation of Industrial Processes Using Optimized Fuzzy Models L. Mendonça, J. Sousa, J. M. S. D. Costa Computer Science The 14th IEEE International Conference on Fuzzy Systems, 2005. FUZZ '05. 2005 17 View 2 excerpts, references methods Save Alert Research Feed A neural network methodology for process fault diagnosis V. Venkatasubramanian, K. Chan Engineering 1989 362 View 1 excerpt, references background Save Alert Research Feed A fuzzy expert system for fault detection in statistical process control of industrial processes S. M. El-Shal, A. S. Morris Computer Science IEEE Trans. Syst. Man Cybern. Part C 2000 63 Highly Influential View 2 excerpts, references background Save Alert Research Feed Model-based fault diagnosis via parameter estimation using knowledge base and fuzzy logic approach L. Mohamed, A. S. Ibrahim Engineering 11th IEEE Mediterranean Electrotechnical Conference (IEEE Cat. No.02CH37379) 2002 11 View 2 excerpts, references background Save Alert Research Feed Fuzzy-logic based trend classification for fault diagnosis of chemical processes Sourabh Dash, R. Rengaswamy, V. Venkatasubramanian Engineering, Computer Science Comput. Chem. Eng. 2003 121 PDF View 3 excerpts, references background Save Alert Research Feed Model-based fault detection in information poor plants J. Howell Engineering, Computer Science Autom. 1994 27 Save Alert Research Feed Improved on-line process fault diagnosis through information fusion in multiple neural networks J. Zhang Engineering, Computer Science Comput. Chem. Eng. 2006 98 View 2 excerpts, references background Save Alert Research Feed Component-based multi-model approach for fault detection and diagnosis of a centrifugal pump A. Wolfram, D. Fussel, T. Brune, R. Isermann Engineering Proceedings of the 2001 American Control Conference. (Cat. No.01CH37148) 2001 44 View 2 excerpts, references background Save Alert Research Feed ... 1 2 3 ... Related Papers Abstract Figures and Tables 79 Citations 22 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_ic4cnnv7svfkboukympxr6ybyu	----	D a t a a n a l y ti c s e n h a n c e d c o m p o n e n t v ol a t ili t y m o d e l Yao, Y, Z h a i, J, C a o , Y, Di n g , X, Li u , J a n d L u o , Y h t t p : // d x. d oi. o r g / 1 0 . 1 0 1 6 /j. e s w a . 2 0 1 7 . 0 5 . 0 2 5 T i t l e D a t a a n a l y ti c s e n h a n c e d c o m p o n e n t v ol a t ili ty m o d e l A u t h o r s Yao, Y, Z h a i , J, C a o , Y, Di n g , X, Li u , J a n d L u o , Y Ty p e Ar ti cl e U R L T hi s v e r s i o n i s a v a il a b l e a t : h t t p : // u s ir. s a lf o r d . a c . u k /i d / e p r i n t / 4 2 3 1 4 / P u b l i s h e d D a t e 2 0 1 7 U S I R i s a d i gi t a l c oll e c t i o n of t h e r e s e a r c h o u t p u t of t h e U n iv e r s i t y of S a lf o r d . W h e r e c o p y r i g h t p e r m i t s , f ull t e x t m a t e r i a l h e l d i n t h e r e p o s i t o r y i s m a d e f r e e l y a v a il a b l e o n li n e a n d c a n b e r e a d , d o w n l o a d e d a n d c o p i e d fo r n o n- c o m m e r c i a l p r i v a t e s t u d y o r r e s e a r c h p u r p o s e s . P l e a s e c h e c k t h e m a n u s c r i p t fo r a n y f u r t h e r c o p y r i g h t r e s t r i c ti o n s . F o r m o r e i nf o r m a t i o n , i n cl u d i n g o u r p o li c y a n d s u b m i s s i o n p r o c e d u r e , p l e a s e c o n t a c t t h e R e p o s i t o r y Te a m a t : u s i r @ s a lf o r d . a c . u k . mailto:usir@salford.ac.uk 1 Data analytics enhanced component volatility model Author: Yuan Yao Institute of Management Science and Engineering, Business School, Henan University, 475004, Jinming District, Kaifeng, Henan Province, China prof.yuanyao@gmail.com Jia Zhai Salford Business School, University of Salford, 43 Crescent, Salford M5 4WT, UK j.zhai@salford.ac.uk jia.zhai1982@gmail.com Yi Cao Department of Business Transformation and Sustainable Enterprise, Surrey Business School, University of Surrey, Guildford, Surrey, GU2 7XH, United Kingdom Jason.caoyi@gmail.com Xuemei Ding School of Computing and Intelligent Systems, Ulster University, Magee campus, Northland Rd, Londonderry Northern Ireland, UK, BT48 7JL x.ding@ulster.ac.uk Faculty of Software, Fujian Normal University, Upper 3rd Rd, Cangshan, Fuzhou, Fujian Province, 350108, China xuemeid@fjnu.edu.cn mailto:prof.yuanyao@gmail.com mailto:j.zhai@salford.ac.uk mailto:jia.zhai1982@gmail.com mailto:Jason.caoyi@gmail.com mailto:x.ding@ulster.ac.uk mailto:xuemeid@fjnu.edu.cn 2 Junxiu Liu, School of Computing and Intelligent Systems, Ulster University, Magee campus, Northland Rd, Londonderry Northern Ireland, UK, BT48 7JL j.liu1@ulster.ac.uk Yuling Luo Guangxi Key Lab of Multi-Source Information Mining & Security, Faculty of Electronic Engineering, Guangxi Normal University, Diecai, Guilin, Guangxi, China, 541000 yuling0616@mailbox.gxnu.edu.cn Corresponding author: Jia Zhai Salford Business School, University of Salford, 43 Crescent, Salford M5 4WT, UK +44(0) 161 295 8147 j.zhai@salford.ac.uk jia.zhai1982@gmail.com mailto:j.zhai@salford.ac.uk mailto:jia.zhai1982@gmail.com 3 Data analytics enhanced component volatility model Abstract Volatility modelling and forecasting have attracted many attentions in both finance and computation areas. Recent advances in machine learning allow us to construct complex models on volatility forecasting. However, the machine learning algorithms have been used merely as additional tools to the existing econometrics models. The hybrid models that specifically capture the characteristics of the volatility data have not been developed yet. We propose a new hybrid model, which is constructed by a low-pass filter, the autoregressive neural network and an autoregressive model. The volatility data is decomposed by the low- pass filter into long and short term components, which are then modelled by the autoregressive neural network and an autoregressive model respectively. The total forecasting result is aggregated by the outputs of two models. The experimental evaluations using one-hour and one-day realized volatility across four major foreign exchanges showed that the proposed model significantly outperforms the component GARCH, EGARCH and neural network only models in all forecasting horizons. 1. Introduction Volatility is considered the “barometer for the vulnerability of financial markets and the economy” (Jiang, Ahmed, & Liu, 2016) (Poon & Granger, 2003) and is crucial for asset pricing, derivative valuation and risk management. Volatility modelling and forecasting is a much devoted area of research and have attracted many attentions of researches. A number of econometrics and machine learning models have been developed in past years. Conventional models include generalized autoregressive conditional heteroskedastic (GARCH) model, developed by (Tim, 1986) and (Bollerslev, 1990), was accepted as one of the most popular volatility models for years. At the same time, autoregressive fractionally integrated moving average (ARFIMA) model 4 has been proposed to capture the long memory property in realized volatility (Granger C. W., 1980) (Granger & Joyeux, 1980). In 1999, (Lee & Engle, 1999) introduced a component GARCH model, which decomposes volatility into two components: a permanent long run trend component and a transitory short-run one that is mean-reverted to the long-run trend (Harris, Stoja, & Yilmaz, 2015). After that, more empirical evidences suggested that two-component model can capture the volatility structure much better than the one-component models. For example, (Brandt & Jones, 2006) showed that daily volatility can be well characterized by two- component model structure with one highly persistent component and one strongly stationary component. A number of literatures have found that two-component volatility models perform better than the one- component models in explaining stock as well as exchange rate volatilities. Started from component GARCH model, introduced by (Lee & Engle, 1999), those two-component models usually consider the volatilities as a composition of a permanent long-run trend component and a transitory short-run component, both of which follow a mean-reverting process with a slow reversion speed in long-run trend and quick reversion speed in the short-run one. Additional to the traditional econometrics models, machine-learning algorithms have been widely used in financial modelling (Jia, Yi, Yuan, Xuemei, & Yuhua, 2017) (Zhai, Cao, Yao, Ding, & Li, 2017) (Yi, Yuhua, Sonya, Ammar, & Martin, 2015) in recent years. (Boyacioglu & Avci, 2010) proposed an adaptive neural fuzzy inference system (ANFIS) for predicting earning per shares on Istanbul stock market and concluded that their method performed well on monthly forecasting. This ANFIS model was also effectively applied in predicting closing price of Zagreb Stock Market index (Svalina, Galzina, Lujić, & Šimunović, 2013). Two hybrid models were proposed in (Hajizadeh, Seifi, Zarandi, & Turksen, 2012), where explanatory input variables were selected based on GARCH, EGARCH and GJR-GARCH models and were fed into neural network model for volatility forecasting. The experimental evaluations showed that the hybrid models outperforming traditional models effectively. A hybrid model of self-organized fuzzy neural network and ARIMA model has been proposed in (McDonald S. , Coleman, McGinnity, & Li, 2013) and applied on 5 financial markets. This hybrid model was thoroughly compared with other traditional forecasting models in (McDonald S. , Coleman, McGinnity, Li, & Belatreche, 2014). The results showed that hybrid model achieved better forecasting results in average. (Kristjanpoller, Fadic, & Minutolo, 2014) developed a hybrid model composed of neural network and GARCH model, where the volatility is modelled and forecasted by GARCH model at the first step and the output of the GARCH as well as the original volatility data were then fed into a neural network model for forecasting the volatility. It showed that the hybrid model significantly outperformed the traditional ARFIMA and GARCH model. The hybrid model was extended in (Kristjanpoller & Minutolo, 2016) for spot and future oil price and showed 30% increase on precision over previous models. Among the structures of most hybrid models, artificial neural network is widely used for modelling the non-linear part of the underlying variable. Artificial neural network is also widely used in time series forecasting in recent studies (Rubio, Elias, Cruz, & Pacheco, 2016) (Aljarah, Faris, Mirjalili, & Al- Madi, 2016) (Rubio, 2016) (Restrepo, Manotas, & Lozano, 2016 ) (Jesús, 2016) (Liu, et al., 2016). Those studies show that careful feature selection and certain neural network structure with tailor-made algorithms can reach a stable, convergent, and better accuracy when compared with the traditional neural network for the prediction of time-series in different areas, i.e., mechatronic processes, brain signal, and some other biomedical data. (Kristjanpoller, Fadic, & Minutolo, 2014) and (Kristjanpoller & Minutolo, 2016) further show that particularly designed hybrid model composed of the artificial neural network and traditional GARCH models achieves significantly better performance than the traditional models, i.e., GARCH, EGARCH, and ARFIMA in forecasting the volatility of foreign exchange and commodity future. The hybrid models are trained by carefully selected features that are related to the volatility. In addition, the work in (Sharda & Patil, Oct 1992) (Gorr, June 1994) (Zhang & Qi, 16 January 2005) show that autoregressive neural network (ARNN) outperforms other structure of artificial neural network in trend and seasonality forecasting of a non-linear time series, where the trend is defined as removing the short-term components from the raw time series and the seasonality is defined as removing the long-term components. (Zhang & Qi, 16 January 2005) and (Nelson, Hill, Remus, & O'Connor, 1999) show that prior data processing, removing either the 6 trend or the seasonality, can dramatically reduce forecasting errors and is critical to build an adequate ARNN based forecasting model. Inspired by previous work of two-component models in econometrics area and hybrid models in machine learning area, we propose an alternative, very simple-structured model to capturing and forecasting volatility across short and long forecasting horizons. Our measure of volatility is based on the realized volatility. We decompose the volatility to long-term and short-term components using a low-pass Hodrick-Prescott filter, and then model the long-term component using an autoregressive neural network and the short-term component as a stationary autoregressive process around the long-term component. The structure of a low- pass filter, a first order autoregressive process and an autoregressive neural network is simple but effective to capture the dynamics of two components of the realized volatility. Since neural network is one of the most popular data analytics algorithms, we name it as “data analytics enhanced component volatility model”. We evaluate the model’s out-of-sample forecasting performance based on the one-hour and one-day realized volatility constructed using EUR/USD, GBP/EUR, GBP/JPY and GBP/USD high frequency exchange rates over the period 27 September 2009 to 12 August 2015, which includes around 7.7 billion observations across 2145 days. As a benchmark, we compare the forecasting accuracy of the proposed model with those of the two-component GARCH model (Lee & Engle, 1999) and the traditional EGARCH model as well as the neural network only model. We evaluate the performances up to 500 time points, which are 50 days for one- hour volatility and roughly two years for one-day volatility. Consistent with the findings of (Lee & Engle, 1999) and (Brandt & Jones, 2006), our experimental evaluations also show that in almost all the cases, our proposed model provides a significant improvement in forecasting performance over Component GARCH and EGARCH as well as the neural network only model. The improvement is stable across short and long horizons. In particular, the forecasting accuracy of our proposed model is roughly consistent across all horizons while the performance of the Component GARCH and EGARCH model deteriorate significantly in 7 long horizon forecasting. Overall, the experimental evaluations show that our proposed model achieves a stable and much better forecasting performance than most of the traditional volatility models. The outline of the remainder of this paper is as follows. In Section 2, we present the structure and details of the proposed model. Section 3 describes the data used in the empirical analysis and the forecasting evaluation criteria. Section 4 presents the empirical results and Section 5 provides a summary, concluding remarks and some future research directions. 2. Methodology 2.1 Model structure In this paper, we follow this idea and the model format in (Lee & Engle, 1999), we assume the realized volatility follows a two-component process given by σt = 𝐿𝑡 + 𝑆𝑡 (1) 𝑆𝑡 = 𝛼𝑆𝑡−1 + εt (2) where 𝐿𝑡 and 𝑆𝑡 are the long and short term component of the realized volatility respectively and εt is the random error term with zero mean and constant variance. The short term component 𝑆𝑡 is an autoregressive, AR(1), process with the parameter 𝛼, which measures the speed of the short term revision to long term trend. This structure of the volatility is following the two-component characteristics in previous literatures, for example (Lee & Engle, 1999), (Alizadeh, Brandt, & Diebold, 2002), and (Brandt & Jones, 2006). However, the implementation of our two-component model is different from traditional Component GARCH model. We implement the two-component model given in equation (1) and (2) in three steps. Firstly, we decompose and extract the long-term component 𝐿𝑡 from the realized volatility using low-pass Hodrick-Prescott filter 8 (Hodrick & Prescott, 1997). After getting 𝐿𝑡, the short term component of the realized volatility can be obtained by 𝑆𝑡 = σt − 𝐿𝑡. In the second step, we train an artificial neural network (ANN) using the long-term component 𝐿𝑡 to obtain a forecasting model. A future value of 𝐿𝑡+𝑛 at time 𝑡 + 𝑛 can be forecasted by the trained ANN. In the last step, the AR(1) process model 𝑆𝑡 = 𝛼𝑆𝑡−1 + εt is estimated using the short term component 𝑆𝑡 , which is obtained from the first step. A future value of 𝑆𝑡+𝑛 at time 𝑡 + 𝑛 can be then forecasted by the estimated AR(1) model. Therefore, we can calculate the future value of the realized volatility at time 𝑡 + 𝑛 by 𝜎𝑡+𝑛 = 𝐿𝑡+𝑛 + 𝑆𝑡+𝑛. 2.2 Volatility decomposition In the first step, the long-term component 𝐿𝑡 is extracted from the realized volatility. To do this, the low-pass Hodrick-Prescott filter (Hodrick & Prescott, 1997) is applied to extract a low frequency non-linear component from a time-series. That low frequency component represents the trend of the long-term component of the realized volatility. Hodrick-Prescott filter (Hodrick-Prescott filter) is widely used in applied macroeconomics (Stock & Watson, 1999) (McElroy, 2008) (Stock & Watson, 2016) for removing the short-term cyclical component of a time series from raw data. Given the value of the smoothing parameter 𝜆, the long term trend component shall solve min 𝜏 (∑(𝑦𝑡 − 𝜏𝑡) 2 + 𝜆 ∑[(𝜏𝑡+1 − 𝜏𝑡) − (𝜏𝑡 − 𝜏𝑡−1)] 2 𝑇−1 𝑡=2 𝑇 𝑡=1 ) where the smoothing parameter 𝜆 penalizes the variations in the growth rate of the trend component. The larger the value 𝜆, the higher is the penalty. We follow the studies in (Baxter & King, 1999) (Ravn & Uhlig, 2002) (Harris, Stoja, & Yilmaz, 2015) to set the 𝜆 as the widely used empirical value of 100 multiplied by the squared frequency of the data, which for one hour realized volatility (assuming 360 trading days per year and 8 hours per day) is 𝜆 = 100 ∗ (360 ∗ 8)2 = 829440000. 9 2.3 Autoregressive Neural Network 2.3.1 Network structure In the second step of implementing our two-component model, we use an autoregressive neural network (ARNN), a special format of artificial neural network, to model and forecasting the long-term component 𝐿𝑡 of realized volatility. As the discussion in Section 1, the work in (Sharda & Patil, Oct 1992) (Gorr, June 1994) (Zhang & Qi, 16 January 2005) show that ARNN outperforms other structure of artificial neural network in trend and seasonality forecasting of a non-linear time series. (Zhang & Qi, 16 January 2005) and (Nelson, Hill, Remus, & O'Connor, 1999) further show that prior data processing, removing either the trend or the seasonality, can dramatically reduce forecasting errors and is critical to build an adequate ARNN based forecasting model. Therefore, in our paper, we follow the existing study and employ the ARNN to model and forecasting the trend, the long-term component 𝐿𝑡 of realized volatility. The reason of selecting ARNN is twofold. First, main strand finance literature usually assumes the volatility is composed of two autoregressive process, long term and short term. We follow the traditional assumption with augmented adjustment: ARNN, which models the lags as a nonlinear function. Secondly, compared with other artificial neural network, ARNN can achieve even better accuracy in “deseasonalized” financial time-series forecasting by a relatively simple structure, where appropriate lags instead of a number of additional features is crucial for the forecasting performance. ARNN is also proved to have advantages over recurrent feed-forward neural network and is less sensitive to the problem of long-term dependence (Mustafaraj, Lowry, & Chen, 2011). Compared with traditional feed-forward artificial neural network, the ARNN has interconnection from the lagged input to the output layer, which enhances its capability of forecasting more than one-step as a time- series predictor. We follow the traditional ARNN structure: 𝑌𝑡 = 𝛼0 + ∑𝛼𝑖𝑌𝑡−𝑖 𝑛 𝑖=1 + ∑Ψ(𝛾0𝑗 + ∑𝛾𝑖𝑗𝑌𝑡−𝑖 𝑛 𝑖=1 )𝛽𝑗 ℎ 𝑗=1 + 𝑡 where 𝑛 is the number of lags, ℎ is the number of hidden neurons, Ψ(.) is the activation function, 𝛾 is the weights between input and hidden neurons, 𝛽 is the weights between hidden and output neurons. The non- 10 linear part contains ℎ hidden neurons transforms the input variables, defined as the 𝑛-lagged long-term component 𝐿𝑡 of realized volatility, weighted by parameters 𝛾𝑖𝑗 plus a bias 𝛾0𝑗, via a non-linear activation function Ψ(.). If representing the number of lags and hidden neurons as 𝑖 and 𝑗, a hidden neuron can be denoted by Ψ(𝛾0𝑗 + ∑𝛾𝑖𝑗𝑌𝑡−𝑖 𝑛 𝑖=1 ) each of which is weighted by a parameter 𝛽𝑗 before it produces the output layer. Since we want to model and forecasting the long-term component of volatility, ARNN operates as a non-linear regression function at the end as 𝑌𝑡 = 𝐺(𝑌𝑡−1,𝑌𝑡−2,…,𝑌𝑡−𝑛) + 𝑡,𝐴𝑅𝑋𝑁𝑁 which maps the unknown relation, 𝐺(∙), between the input variable and the target function 𝑌𝑡 and the error term 𝑡,𝐴𝑅𝑋𝑁𝑁. For the purpose of this study, the hidden layer uses the hyperbolic tangent sigmoid transfer function, while the output layer uses a linear transfer function. The ARNN structure with only one hidden layer is considered since it operates as a non-linear regression function and can be trained to approximate most non-linear function arbitrarily well (Siegelmann, Horne, & Giles, 1997 ), (Andreou, Charalambous, & Martzoukos, 2008) (Mustafaraj, Lowry, & Chen, 2011). Following this structure, in this paper, to forecasting long-term component 𝐿𝑡 of the realized volatility, an autoregressive neural network (ARNN) with three layers structure is used: an input layer that includes lagged 𝐿𝑡 inputs to the network; a hidden layer with hyperbolic tangential activation function, and an output layers with a linear activation function. In previous studies, the estimation of the model parameters usually follows the back-propagation algorithm. It is shown in (Charalambous, June 1992) that the back-propagation algorithm is often unable to converge rapidly to the optimal solution. Therefore, we utilize the modified Levenberg-Marquardt (LM) algorithm, which is much more sophisticated and efficient in terms of time capacity and accuracy (Hagan & Menhaj, 1994). When training the ARNN, we divide the training dataset into three subsets, 80% of the data for training the ARNN, 10% for validation, and the last 10% for testing. During the training process, the errors on the training dataset 11 and validation dataset are monitored at the same time. When the validation error rises while the training error maintains, the ARNN begins to overfit the data. The weights and bias at the minimum of the validation error are saved as the trained ARNN 2.3.2 Network parameters Under this structure of ARNN, the number of lags 𝑛 and the number of hidden neurons ℎ are the crucial parameters for constructing the ARNN model. We follow the widely used configuration for the ARNN in (Siegelmann, Horne, & Giles, 1997 ) and (Mustafaraj, Lowry, & Chen, 2011), and investigate the performance using the number of lags from 2 to 5, which is 𝑛 = 2,3,4,5. We firstly use ℎ = 10 as the preliminary configuration for the number of hidden neurons and finds the appropriate lags. We use the long- term component 𝐿𝑡 of one hour realized EURUSD rate from 27 Sep 2009 to 06 Dec 2012 to train the ARNN model under 𝑛 = 2,3,4,5 and use 500 long-term component values of EURUSD rate from 07 Dec 2012 to 15 Jan 2013 to test the trained model. The forecasting error is defined as �̂�𝑡 − 𝐿𝑡, the difference between the original value 𝐿𝑡 and the forecasted one. The average errors across the forecasting horizons is listed in Table 1, where we can observe that the best average error is at 𝑛 = 4, although the errors do not show a significant difference among different configuration of lag numbers. Four lags also conform to the configuration suggested in (Siegelmann, Horne, & Giles, 1997 ) and (Mustafaraj, Lowry, & Chen, 2011). Therefore, in this paper, we use four lags autoregressive neural network (ARNN) following the suggested configuration and our empirical investigation. Table 1 This table contains the average forecasted error across the horizon from 07 Dec 2012 to 15 Jan 2013 under configurations of lag number from 2, 3, 4 and 5 𝑛 = 2 𝑛 = 3 𝑛 = 4 𝑛 = 5 Average error 1.22807E-05 1.22647E-05 1.21385E-05 1.2293E-05 The following Error! Reference source not found.Figure 1 shows an example of using ARNN model to forecasting the long-term component 𝐿𝑡 of one hour realized EURUSD rate from 07 Dec 2012 to 15 Jan 2013. The ARNN model was constructed using one hidden layer with 10 neurons and was trained using the 12 𝐿𝑡 of EURUSD rate from 27 Sep 2009 to 06 Dec 2012. The forecasting results �̂�𝑡 include 500 long-term component values of EURUSD rate from 07 Dec 2012 to 15 Jan 2013. The forecasting error, defined as �̂�𝑡 − 𝐿𝑡, between the original value 𝐿𝑡 and the forecasted one �̂�𝑡 are shown in the bottom sub-figure, from which we can observe that out-of-sample forecasting errors are around 10−5. Figure 1 This figure shows an example of out-of-sample forecasting long term component 𝑳𝒕 of EURUSD using ARNN. The top figure shows the long term component of EURUSD from 07 Dec 2012 to 05 Feb 2013 decomposed by Hodrick-Prescott filter. The middle figure shows the forecasted value �̂�𝒕 of the long term component of EURUSD from 07 Dec 2012 to 15 Jan 2013 using ARNN model. The bottom figure shows the differences �̂�𝒕 − 𝑳𝒕 between the original long term component of EURUSD and the forecasted one. 2.4 Autoregressive model In the third step, we estimate an autoregressive model for the short-term component 𝑆𝑡: 𝑆𝑡 = 𝛼𝑆𝑡−1 + εt. The n-step ahead forecasting of the short-term component 𝑆𝑡 is therefore given by �̂�𝑡+𝑛 = 𝛼 𝑛𝑆𝑡 + ∑ 𝛼 𝑖𝑛−1 𝑖 εt−i (3) We can obtain the forecasted the long-term component �̂�𝑡+𝑛 of the realized volatility through ARNN model. Therefore, the n-step ahead forecasting of the one hour realized volatility is given by Equation (1) as σ̂t+n = �̂�𝑡+𝑛 + �̂�𝑡+𝑛 . We name the proposed model as autoregressive neural network enhanced two-component model (shorted as: NNE2C) due to the implementation of the model structure. To estimate the autoregressive 13 model for the short-term component 𝑆𝑡, we utilize the method of moment through Yule–Walker equations, named for Udny Yule and Gilbert Walker (Yule, 1927) (Walker, 1931): 𝛾𝑚 = ∑ 𝛼𝑘𝛾𝑚−𝑘 𝑝 𝑘=1 + 𝜎𝜖 2𝛿𝑚,0 where 𝛾𝑚 is the autocovariance function of 𝑆𝑡, 𝜎𝜖 2 is the variance of the input noise process and the 𝛿𝑚,0 is the Kronecker delta function. For the case 𝑝 = 1, one lag autoregressive process AR(1), the 𝛼1 can be obtained by 𝛾1/𝛾0. 2.5 Model flow We summarize the NNE2C model workflow when applied in realized volatility forecasting. Different from other hybrid models, which usually use the neural network as the dominate one to model the non-linear part of the underline financial variable, i.e. volatility or foreign exchange and use the ARIMA (McDonald S. , Coleman, McGinnity, Li, & Belatreche, 2014) or GARCH model (Kristjanpoller, Fadic, & Minutolo, 2014) as the pre-processing part for modelling the linear part of the underline financial variable, the NNE2C model follows the traditional financial theory to consider the volatility as a combination of the long and short term while producing an enhanced mechanism through explicitly decomposing the two components and modelling them separately. The simple NNE2C model follows and enhances the financial theory by capturing the volatility structure explicitly and effectively. Algorithm 1 Workflow of proposed NNE2C model 1 Given the realized volatility σt of selected financial security, i.e. FX, we assume it is composed of long and short term components: σt = 𝐿𝑡 + 𝑆𝑡 2 We use Hodrick-Prescott filter as a low-pass filter to explicitly extract the long term component 𝐿𝑡. The smoothing parameter 𝜆 of Hodrick-Prescott filter is selected according to the empirical study: 𝜆 = 100 ∗ (360 ∗ 8)2 = 829440000. min 𝜏 (∑(σt − 𝜏𝑡) 2 + 𝜆 ∑[(𝜏𝑡+1 − 𝜏𝑡) − (𝜏𝑡 − 𝜏𝑡−1)] 2 𝑇−1 𝑡=2 𝑇 𝑡=1 ) 3 The short term component 𝑆𝑡 is then obtained by 𝑆𝑡 = σt − 𝐿𝑡 4 Modelling 𝐿𝑡 and 𝑆𝑡 simultaneously: 14 Using 𝐿𝑡 to train the ARNN with 4 lags and 10 hidden neurons, 𝑛 step ahead 𝐿𝑡+𝑛 is forecasted by the trained ARNN model; 𝐿𝑡 = 𝛼0 + ∑𝛼𝑖𝐿𝑡−𝑖 𝑛 𝑖=1 + ∑ Ψ(𝛾0𝑗 + ∑𝛾𝑖𝑗𝐿𝑡−𝑖 𝑛 𝑖=1 )𝛽𝑗 ℎ 𝑗=1 + 𝑡 Using 𝑆𝑡 to estimate the AR(1) model, 𝑛 step ahead 𝑆𝑡+𝑛 is forecasted by the estimated AR(1) model; 𝑆𝑡 = 𝛼𝑆𝑡−1 + εt 5 𝑛 step ahead realized volatility is obtained by σt+n = 𝐿𝑡+n + 𝑆𝑡+n 3. Data and forecasting evaluation 3.1 Data We use the neural network enhanced two-component model defined in Section 2 to forecasting the volatility of the EUR/USD, GBP/EUR, GBP/JPY and GBP/USD exchange rates. High frequency exchange rate data (tick data) were obtained from Oricode Inc for the period 27 September 2009 to 12 August 2015 and included around 7.7 billion observations across 2145 days. The unobserved true volatility, in principle, can be estimated arbitrarily accurately using a measure of realized volatility calculated through the intraday returns (Harris, Stoja, & Yilmaz, 2015). It is proved by (Torben, Tim, & Francis, 2009) that the sum of squared intraday returns converges to the unobserved true volatility with the intraday interval approaching to zero. In this study, we use realized volatility as the proxy of the unobserved true volatility. This is obtained by aggregating the intraday squared returns using the approach given by (Andersen & Bollerslev, 1998): �̂�𝑟𝑣,𝑡 2 = ∑ 𝑟𝑡,𝑛 2 𝑁 𝑛=1 where �̂�𝑟𝑣,𝑡 2 is the realized volatility for time 𝑡, and 𝑟𝑡,𝑛 2 is the squared log return on time 𝑡 for interval 𝑛 (𝑛 = 1,2,… ,𝑁). In this paper, we use one-hour and one-day realized volatilities, both of which are constructed from 10 millisecond log return, which is the highest frequency in our data. The one-hour realized volatility is obtained by aggregating 360,000 10-millisecond log returns 𝑟𝑡,𝑛 2 in each hour. The data from 27 September 2009 to 07 December 2012 (20000 observations) is used for the initial 15 estimation of the model, while the data from 08 December 2012 to 12 August 2015 (16681) is used for out- of-sample evaluation. The one-day realized volatilities is constructed by the sum of 3,600,000 log returns in each trading day. For the one-day realized volatilities, the data from 27 September 2009 to 07 December 2012 (1000 observations) is used for the initial estimation of the model, while the remained data (837 observations) is used for out-of-sample evaluation. Table 2 reports summary statistics for the one-hour and one-day realized volatilities of full observations of four exchange rates. Panel A reports the mean, standard deviation, skewness and excess kurtosis and Bera– Jarque statistic and Panel B reports the first six autocorrelation coefficients and the Ljung–Box Q statistic for autocorrelation of six lags for the realized volatilities. P-values are also reported in parentheses. In the Ljung– Box Q tests, the null hypothesis that the residuals of the returns are not autocorrelated is rejected in both one- hour and one-day realized volatilities. Therefore the two realized volatilities are all highly autocorrelated. Table 2 Summary statistics and autocorrelations mean Standard deviation Skewness Excess kurtosis Bera-Jarque Panel A: Summary statistics GBP/USD 1 Hr 1.6657E-06 9.1010E-06 1.0231E+02 1.1661E+04 2.0743E+11 GBP/JPY 1 Hr 3.6200E-06 7.6346E-06 4.5817E+01 3.4566E+03 1.8227E+10 GBP/EUR 1 Hr 1.8514E-06 3.1560E-06 2.5651E+01 1.1853E+03 2.1401E+09 EUR/USD 1 Hr 2.1207E-06 5.6940E-06 4.4465E+01 3.8006E+03 2.2055E+10 GBP/USD Daily 3.2698E-05 7.2746E-05 3.5359E+01 1.4143E+03 1.5292E+08 GBP/JPY Daily 7.0981E-05 6.8848E-05 7.5062E+00 1.0538E+02 8.2050E+05 GBP/EUR Daily 3.6060E-05 2.4165E-05 2.4964E+00 1.6728E+01 1.6342E+04 EUR/USD Daily 4.1188E-05 5.5114E-05 2.1497E+01 7.0608E+02 3.8019E+07 Panel B: Autocorrelations 1 2 3 4 5 6 Ljung–Box Q GBP/USD 1 Hr 0.6638 0.2696 0.0387 0.0210 0.0216 0.0225 1.8952E+04 (0.0000) GBP/JPY 1 Hr 0.2870 0.1879 0.1897 0.1569 0.1559 0.1502 1.4856E+04 (0.0000) GBP/EUR 1 Hr 0.2381 0.1711 0.1363 0.1149 0.0903 0.0784 5.8392E+03 (0.0000) EUR/USD 1 Hr 0.3683 0.2986 0.2442 0.2361 0.1952 0.1588 1.7166E+04 (0.0000) GBP/USD Daily 0.0569 0.0255 0.0214 0.0218 0.0753 0.1139 6.8927E+01 (2.7236E-07) GBP/JPY Daily 0.5392 0.2908 0.2817 0.2527 0.2763 0.3728 2.8913E+03 (0.0000) GBP/EUR Daily 0.4351 0.1860 0.1641 0.1646 0.3203 0.5706 3.3945E+03 (0.0000) EUR/USD Daily 0.1811 0.1007 0.0914 0.0894 0.1412 0.2041 6.3254E+02 (0.0000) 16 In the example in the following Figure 2(a-b), Hodrick-Prescott filter is applied to one hour realized volatility with smoothing parameter 𝜆 of 829440000 (blue curve in Figure 2(a)) of EURUSD exchange rate from 27 Sep 2009 to 07 Dec 2012 to extract the long term component 𝐿𝑡 (red curve in Figure 2(b)). Figure 2(c-d) shows an example of one-day realized volatility with long term component extracted by Hodrick-Prescott filter with smoothing parameter 𝜆 of 12960000 (assuming 360 trading days per year, the 𝜆 is calculated by 100 multiplied by the squared frequency of the data). (a) (b) (c) (d) Figure 2 (a) Blue curve in the figure: One-hour realized volatility of EURUSD exchange rate from 27 Sep 2009 to 07 Dec 2012; Red curve: the long term component 𝑳𝒕of the realized volatility extracted by low-pass Hodrick-Prescott filter; (b) Short term component of one-hour realized volatility of EURUSD exchange rate; (c) Blue curve in the figure: one-day realized volatility of EURUSD exchange rate from 27 Sep 2009 to 07 Dec 2012; Red curve: the long term component 𝑳𝒕of the realized volatility extracted by low- pass Hodrick-Prescott filter; (d) Short term component of one-day realized volatility of EURUSD exchange rate. 17 3.2 Forecasting evaluation The proposed neural network enhanced two-component model is used to calculate out-of-sample forecastings of the realized volatilities of up to 500 hours or days ahead across the evaluation period for the one-hour or one-day realized volatility respectively. As the benchmark, we select one-factor EGARCH and two-factor Component GARCH of (Lee & Engle, 1999). In addition, we also employ four lags autoregressive neural network applied directly on realized volatility as one of the benchmark models. We therefore estimate four models: a) four lags autoregressive neural network enhanced two-component model (NNE2C); b) one- component EGARCH model; c) two-component GARCH model; and d) four lags autoregressive neural network model (NNOnly). The four models are evaluated using one-hour and one-day realized volatilities respectively. For one-hour and one-day data, four models are initially estimated using the first 20,000 and 1,000 observations from 27 Sep 2009 to 07 Dec 2012 respectively and then the volatilities at different estimation periods are calculated. Following (Michael & Christopher, 2006), we consider forecasting horizons of 5, 20, 100, 200, 360, and 500 hours or days ahead for one-hour or one-day volatility respectively. We choose the root mean square error (RMSE) with respect to the true realized volatility as the measure of the forecasting performance: 𝑅𝑀𝑆𝐸 = [ 1 𝑇 ∑ (𝜎𝑡(𝜏𝑡,𝜏𝑡+𝑇) − �̂�𝑡(𝜏𝑖,𝑡,𝜏𝑡+𝑇)) 2 𝑇 𝑡=1 ] 1 2 (4) For the shorter evaluation horizons of 5, 20, 100, we calculate the RMSE over the forecasting horizon, i.e. (𝜏𝑡,𝜏𝑡+𝑇) = (1,5),(1,20) and (1,100). For the three longer horizons, we calculate the RMSE over 100 time points, i.e. (𝜏𝑡,𝜏𝑡+𝑇) = (100,200),(260,360) and (400,500). For the one-hour realized volatility, the RMSE is to evaluate the average performance over 10 days (10 trading hours in each day). For the one-day realized volatility, the RMSE is to evaluate the average performance over 3 month ahead. 18 4. Results 4.1 Model parameter In theory, three-layer autoregressive neural network can approximate most of the functions as long as a sufficient number of hidden neurons is provided. In this paper, we construct the NNE2C with 5, 10, and 15 neurons in the hidden layer and test the forecasting performance of the proposed model. The results are illustrated in Table 3 and Table 4 for one-hour and one-day realized volatility data respectively. For the comparison purpose, we also included the forecasting results with 10 hidden neurons in Table 3 and Table 4. From Table 3, it is very clear that for the NNE2C, the forecasting accuracies are roughly the same by using different number of hidden neurons. The forecasting accuracy does not increase with the number of hidden neuron increases. This result is consistent in all forecasting horizons across four currencies. The only case that the forecasting accuracy rises with the increase of the number of hidden neurons is highlighted in Table 3 as GBP/USD rate at (100,200) horizon. However, the accuracy increase is as subtle as around 10E-8. For all the cases in Table 4, the forecasting accuracy differences by different number of hidden neurons are also tiny. The only three cases that the forecasting accuracy increases (although tiny) with the rise of the number of hidden neurons are highlighted in Table 4: GBP/JPY rate at horizon (260,360), (400,500) and GBP/EUR at horizon (400,500). In addition, the forecasting accuracies of NNOnly model with different number of hidden neurons do not show significant differences in Table 3 and Table 4 as well. Our results also conform to the previous researches in (Kristjanpoller & Minutolo, 2016) and (Kristjanpoller, Fadic, & Minutolo, 2014). Therefore, based on our experimental evaluations in Table 3 and Table 4, we believe our proposed NNE2C can achieve the best forecasting performance under the structure of three-layer autoregressive neural network with 10 hidden neurons. Table 3 Forecasting performance of one-hour realized volatilities. This table reports the Root mean square error (RMSE) for the autoregressive neural network enhanced two-component model constructed by one hidden layer with 5 and 15 neurons (NNE2C) and autoregressive neural network model constructed by one hidden layer with 5, 10 and 15 neurons (NNOnly). 𝜏1 𝜏2 NNE2C (5 neurons) NNE2C (10 neurons) NNE2C (15 neurons) NNOnly (5 neurons) NNOnly (10 neurons) NNOnly (15 neurons) 1 5 EUR / USD 2.9838E-06 3.0167E-06 2.9828E-06 5.8882E-02 5.8289E-02 5.8018E-02 19 GBP / EUR 9.8432E-06 9.8434E-06 9.8435E-06 7.5266E-02 7.4758E-02 7.4355E-02 GBP / JPY 4.9453E-05 4.9452E-05 4.9452E-05 1.4364E-01 1.3907E-01 1.4022E-01 GBP / USD 1.1083E-05 1.1081E-05 1.1087E-05 5.5426E-02 5.2369E-02 5.4421E-02 1 20 EUR / USD 5.7274E-05 5.7239E-05 5.7276E-05 1.3165E-01 1.3015E-01 1.3121E-01 GBP / EUR 4.8869E-05 4.8870E-05 4.8870E-05 1.2097E-01 1.2032E-01 1.2054E-01 GBP / JPY 5.3459E-05 2.3459E-05 5.3459E-05 1.5310E-01 1.4879E-01 1.5035E-01 GBP / USD 2.6065E-05 2.6063E-05 2.6069E-05 7.4815E-02 7.2129E-02 7.3733E-02 1 100 EUR / USD 3.1612E-05 3.1565E-05 3.1614E-05 1.0184E-01 1.0078E-01 1.0153E-01 GBP / EUR 3.0537E-05 1.0537E-05 3.0538E-05 1.0201E-01 1.0147E-01 1.0153E-01 GBP / JPY 4.4369E-05 4.4369E-05 4.4369E-05 1.4828E-01 1.4397E-01 1.4558E-01 GBP / USD 2.6203E-05 2.6201E-05 2.6207E-05 7.4542E-02 7.1763E-02 7.3194E-02 100 200 EUR / USD 2.8230E-05 2.8183E-05 2.8229E-05 9.4983E-02 9.4030E-02 9.4606E-02 GBP / EUR 3.3874E-05 3.2874E-05 3.3874E-05 1.0056E-01 1.0006E-01 1.0004E-01 GBP / JPY 6.0882E-05 6.0882E-05 6.0882E-05 1.8717E-01 1.8326E-01 1.8483E-01 GBP / USD 2.3845E-05 2.3844E-05 2.3843E-05 7.8703E-02 7.6050E-02 7.7510E-02 260 360 EUR / USD 3.8015E-05 3.8012E-05 3.8012E-05 1.5028E-01 1.4931E-01 1.4995E-01 GBP / EUR 2.0364E-05 2.0363E-05 2.0364E-05 8.7187E-02 8.6600E-02 8.6657E-02 GBP / JPY 4.8720E-05 4.8719E-05 4.8720E-05 1.7955E-01 1.7590E-01 1.7743E-01 GBP / USD 1.0392E-04 1.0396E-04 1.0393E-04 5.5737E-01 5.5654E-01 5.5698E-01 400 500 EUR / USD 6.1426E-05 6.1437E-05 6.1424E-05 1.7433E-01 1.7332E-01 1.7402E-01 GBP / EUR 3.9610E-05 3.9608E-05 3.9608E-05 1.2262E-01 1.2199E-01 1.2209E-01 GBP / JPY 7.0257E-05 7.0253E-05 7.0258E-05 2.0427E-01 2.0026E-01 2.0216E-01 GBP / USD 4.7603E-05 4.7615E-05 4.7608E-05 1.9838E-01 1.9656E-01 1.9750E-01 Table 4 Forecasting performance of one-day realized volatilities. This table reports the Root mean square error (RMSE) for the autoregressive neural network enhanced two-component model constructed by one hidden layer with 5 and 15 neurons (NNE2C) and autoregressive neural network model constructed by one hidden layer with 5 and 15neurons (NNOnly) 𝜏1 𝜏2 NNE2C (5 neurons) NNE2C (10 neurons) NNE2C (15 neurons) NNOnly (5 neurons) NNOnly (10 neurons) NNOnly (15 neurons) 1 5 EUR / USD 1.2852E-04 1.2854E-04 1.2860E-04 4.4915E-01 4.5038E-01 4.5214E-01 GBP / EUR 2.3046E-04 2.3082E-04 2.3053E-04 5.0792E-01 5.0309E-01 5.0854E-01 GBP / JPY 2.7913E-04 2.7904E-04 2.7938E-04 7.6417E-01 7.6045E-01 7.6562E-01 GBP / USD 3.3468E-05 3.1831E-05 3.3231E-05 4.0386E-01 4.0349E-01 4.0417E-01 1 20 EUR / USD 1.7691E-04 1.7692E-04 1.7702E-04 5.7042E-01 5.7113E-01 5.7226E-01 GBP / EUR 1.3060E-04 1.3087E-04 1.3088E-04 4.4148E-01 4.3658E-01 4.4204E-01 GBP / JPY 2.2931E-04 2.2910E-04 2.3074E-04 7.8099E-01 7.7768E-01 7.8039E-01 GBP / USD 3.4261E-04 3.4074E-04 3.4227E-04 1.4302E+00 1.4300E+00 1.4300E+00 1 100 EUR / USD 1.8115E-04 1.8101E-04 1.8164E-04 5.6324E-01 5.6411E-01 5.6537E-01 GBP / EUR 1.9050E-04 1.9037E-04 1.9162E-04 5.8821E-01 5.8323E-01 5.8888E-01 GBP / JPY 2.9117E-04 2.9235E-04 2.9621E-04 9.1889E-01 9.1522E-01 9.1853E-01 GBP / USD 1.8631E-04 1.8424E-04 1.8571E-04 7.5809E-01 7.5764E-01 7.5780E-01 100 200 EUR / USD 1.7611E-04 1.7565E-04 1.7784E-04 5.5212E-01 5.5295E-01 5.5436E-01 GBP / EUR 1.7733E-04 1.7777E-04 1.7590E-04 5.6214E-01 5.5714E-01 5.6286E-01 GBP / JPY 3.0586E-04 3.0894E-04 3.0814E-04 9.5076E-01 9.4728E-01 9.5029E-01 GBP / USD 1.7191E-04 1.6691E-04 1.7008E-04 5.4682E-01 5.4626E-01 5.4663E-01 20 260 360 EUR / USD 1.3838E-04 1.3791E-04 1.4283E-04 4.2934E-01 4.3042E-01 4.3157E-01 GBP / EUR 1.6731E-04 1.6787E-04 1.6194E-04 5.2557E-01 5.2051E-01 5.2631E-01 GBP / JPY 2.0534E-04 2.0251E-04 1.7687E-04 6.4310E-01 6.3951E-01 6.4250E-01 GBP / USD 1.3880E-04 1.3370E-04 1.3637E-04 4.2614E-01 4.2548E-01 4.2597E-01 400 500 EUR / USD 1.0489E-04 1.0359E-04 1.1499E-04 3.2422E-01 3.2540E-01 3.2634E-01 GBP / EUR 1.4010E-04 1.3893E-04 1.3327E-04 4.3346E-01 4.2840E-01 4.3424E-01 GBP / JPY 1.6133E-04 1.4925E-04 1.0101E-04 5.0550E-01 5.0165E-01 5.0518E-01 GBP / USD 1.0201E-04 9.7793E-05 9.9039E-05 3.1800E-01 3.1736E-01 3.1824E-01 4.2 Experimental results Following the configurations, we employ our experiments using NNE2C with 4 lags and 10 neurons. Table 5 and Table 6 report the RMSE of the one-hour and one-day realized volatilities given by equation (4) for four models over six forecasting intervals for four currencies respectively. Overall, for all models in the experiments in Table 5 and Table 6, the RMSE measures fall at the first with the forecasting horizon and then rise for four currencies. This is due to the reason that initially the forecasting interval increases from five to 20 and then to 100 time points. After the horizon (𝝉𝒕,𝝉𝒕+𝑻) = (𝟏,𝟏𝟎𝟎), the forecasting interval is fixed at 100 time points, and therefore the forecasting error rises as the horizon increases. In one-hour realized volatility evaluations in Table 5, the NNE2C model shows the highest forecasting accuracy in 23 out of 24 cases (four currencies with each having 6 horizons) followed by CGARCH model. The only exception is the highlighted case in Table 5: CGARCH achieved RMSE of 3.2067E-05 at horizon (𝟏𝟎𝟎,𝟐𝟎𝟎) on GBP/EUR and was more accurate than the NNE2C, which had 3.2874E-05 RMSE at the same case. At horizon of (𝟏,𝟏𝟎𝟎), the performance of CGARCH model is lower but close to the performance of the NNE2C in all four currencies. In other horizons, the NNE2C is significantly more accurate than CGARCH model. Particularly, the decline of the forecasting accuracy of the CGARCH model as well as EGARCH and NNOnly model is obvious from horizon (𝝉𝒕,𝝉𝒕+𝑻) = (𝟏𝟎𝟎,𝟐𝟎𝟎) to (𝟐𝟔𝟎,𝟑𝟔𝟎) and then to (𝟒𝟎𝟎,𝟓𝟎𝟎). It shows that at longer forecasting horizon, the NNE2C has sufficiently stronger forecasting capability than the traditional models. Table 5 Forecasting performance of one-hour realized volatilities. This table reports the Root mean square error (RMSE) for the autoregressive neural network enhanced two-component model constructed by one hidden layer with 10 neurons (NNE2C 10 21 neurons), Component GARCH model, EGARCH model, and autoregressive neural network model (NNOnly), for the forecasting interval 𝝉𝟏 to 𝝉𝟐, where 𝝉𝟏 = 𝒎𝒂𝒙(𝟏,𝝉𝟐 − 𝟏𝟎𝟎), for the one-hour realized volatilities of four currencies. 𝜏1 𝜏2 NNE2C (10 neurons) CGARCH EGARCH NNOnly (10 neurons) 1 5 EUR / USD 3.0167E-06 5.1341E-04 8.1872E-04 5.8289E-02 GBP / EUR 9.8434E-06 1.7790E-04 5.8565E-04 7.4758E-02 GBP / JPY 4.9452E-05 1.9269E-04 1.0531E-03 1.3907E-01 GBP / USD 1.1081E-05 3.2305E-04 6.4933E-04 5.2369E-02 1 20 EUR / USD 5.7239E-05 1.6669E-04 1.7328E-04 1.3015E-01 GBP / EUR 4.8870E-05 1.7436E-04 1.3082E-04 1.2032E-01 GBP / JPY 2.3459E-05 2.9617E-05 3.5046E-04 1.4879E-01 GBP / USD 2.6063E-05 1.1403E-04 4.5078E-04 7.2129E-02 1 100 EUR / USD 3.1565E-05 2.2018E-05 2.4161E-04 1.0078E-01 GBP / EUR 1.0537E-05 1.9650E-05 2.1353E-04 1.0147E-01 GBP / JPY 4.4369E-05 6.9260E-05 4.0864E-04 1.4397E-01 GBP / USD 2.6201E-05 8.3107E-05 3.9228E-04 7.1763E-02 100 200 EUR / USD 2.8183E-05 5.1364E-05 3.0277E-04 9.4030E-02 GBP / EUR 3.2874E-05 3.2067E-05 2.5238E-04 1.0006E-01 GBP / JPY 6.0882E-05 1.7346E-04 4.3271E-04 1.8326E-01 GBP / USD 2.3844E-05 5.2547E-05 3.2808E-04 7.6050E-02 260 360 EUR / USD 3.8012E-05 2.2140E-04 3.0277E-04 1.4931E-01 GBP / EUR 2.0363E-05 1.2524E-04 2.6884E-04 8.6600E-02 GBP / JPY 4.8719E-05 1.0468E-04 1.2865E-03 1.7590E-01 GBP / USD 1.0396E-04 1.2240E-03 2.9355E-04 5.5654E-01 400 500 EUR / USD 6.1437E-05 5.3047E-04 1.6394E-04 1.7332E-01 GBP / EUR 3.9608E-05 1.7219E-04 2.3168E-04 1.2199E-01 GBP / JPY 7.0253E-05 4.0934E-04 8.0361E-04 2.0026E-01 GBP / USD 4.7615E-05 5.8900E-04 1.5979E-04 1.9656E-01 In Table 6, the NNE2C was significantly more accurate than other three models in all 24 cases. In the highlighted cases in Table 6, which include GBP/EUR rate at (1,5) and (1,20) horizons, and GBP/EUR rate at (400,500) horizon, CGARCH model achieved the forecasting accuracies lower but close to the performance of the NNE2C. In other cases, the NNE2C remarkably outperforms all other models. It is worth noting that the NNOnly model performs the worst in all cases while the NNE2C achieves the significantly accurate performance. Our results conform to that of (Zhang & Qi, 2005) and (Nelson, Hill, Remus, & O'Connor, 1999), which concluded that neural network is not able to model volatility directly but neural networks built with deseasonalized data could produce significantly more accurate forecasting than with non-deseasonalized data. 22 Table 6 Forecasting performance of one-day realized volatilities. This table reports the Root mean square error (RMSE) for the autoregressive neural network enhanced two-component model constructed by one hidden layer with 10 neurons (NNE2C 10 neurons), Component GARCH model, EGARCH model, and autoregressive neural network model constructed by one hidden layer with 10 neurons (NNOnly 10 neurons), for the forecasting interval 𝝉𝟏 to 𝝉𝟐, where 𝝉𝟏 = 𝒎𝒂𝒙(𝟏, 𝝉𝟐 − 𝟏𝟎𝟎), for the one-day realized volatilities of four currencies. 𝜏1 𝜏2 NNE2C (10 neurons) CGARCH EGARCH NNOnly (10 neurons) 1 5 EUR / USD 1.2854E-04 5.7251E-04 8.4922E-04 4.5038E-01 GBP / EUR 2.3082E-04 3.7841E-04 1.2353E-03 5.0309E-01 GBP / JPY 2.7904E-04 1.1783E-03 2.2118E-03 7.6045E-01 GBP / USD 3.1831E-05 3.9678E-04 1.2697E-03 4.0349E-01 1 20 EUR / USD 1.7692E-04 5.4301E-04 1.7308E-03 5.7113E-01 GBP / EUR 1.3087E-04 2.1771E-04 7.0773E-04 4.3658E-01 GBP / JPY 2.2910E-04 1.5067E-05 1.3846E-04 7.7768E-01 GBP / USD 3.4074E-04 3.1626E-03 5.9010E-04 1.4300E+00 1 100 EUR / USD 1.8101E-04 9.4802E-04 1.0034E-04 5.6411E-01 GBP / EUR 1.9037E-04 1.1753E-03 1.1850E-03 5.8323E-01 GBP / JPY 2.9235E-04 1.5569E-03 5.2037E-04 9.1522E-01 GBP / USD 1.8424E-04 1.8094E-03 1.4454E-03 7.5764E-01 100 200 EUR / USD 1.7565E-04 9.4161E-04 7.8610E-04 5.5295E-01 GBP / EUR 1.7777E-04 1.0306E-03 1.1111E-03 5.5714E-01 GBP / JPY 3.0894E-04 1.9363E-03 3.8808E-04 9.4728E-01 GBP / USD 1.6691E-04 1.7113E-03 1.7661E-03 5.4626E-01 260 360 EUR / USD 1.3791E-04 2.9667E-04 2.2160E-03 4.3042E-01 GBP / EUR 1.6787E-04 6.3274E-04 1.0282E-03 5.2051E-01 GBP / JPY 2.0251E-04 9.6631E-04 1.5438E-03 6.3951E-01 GBP / USD 1.3370E-04 6.2537E-04 8.5126E-04 4.2548E-01 400 500 EUR / USD 1.0359E-04 1.3025E-03 3.3809E-03 3.2540E-01 GBP / EUR 1.3893E-04 2.1427E-04 2.4499E-03 4.2840E-01 GBP / JPY 1.4925E-04 2.3339E-03 3.5560E-03 5.0165E-01 GBP / USD 9.7793E-05 4.3271E-04 3.0735E-04 3.1736E-01 5. Conclusions The fact that volatility comprises both a long-term trend component and a strongly oscillation short-term component has crucial implications for modelling and forecasting volatility over both short and long horizons. In this paper, we develop a simple but effective volatility-forecasting model. The model is based on a decomposition of intraday realized volatility into the long and short-term components using the low- pass Hodrick-Prescott filter. The three-layer autoregressive neural network with 10 hidden neurons models the long-term component and the short-term component is modelled as a simple autoregressive process. 23 Therefore, we name the proposed model as “neural network enhanced two-component volatility model”. The model was thoroughly evaluated using high frequency tick data of four currencies across six forecasting horizons. The out-of-sample forecasting results consistently and significantly outperform the Component GARCH and EGARCH models as well as the autoregressive neural network model, which is applied on modelling the volatility directly. The results reported in this paper are based on two simple structures: the low-pass Hodrick-Prescott filter uses a fixed smoothing parameter and the short-term component follows a first order autoregressive progress. Future work would be to consider using an optimized smoothing parameter that decomposes a stationary short-term component. A higher order ARMA process would provide a better fit for the decomposed short- term component and may bring improved out-of-sample forecasting performance. Moreover, a higher smoothing parameter brings more smoothed long-term component, which can be easier to forecasting by neural network with higher accuracy, and decomposes a more volatile short-term component, which may not follow a stationary process. Indeed, it would be a sufficiently further improvement if considering the forecasting model as an optimization framework to find an optimal parameter, which decomposes a stationary short-term process while maintaining a smooth long-term component, which can be forecasted by autoregressive neural network with high accuracy. Acknowledgement This research is partially supported by the Scientific Research Funds for the Returned Overseas Chinese Scholars from State Education Ministry, Nature and Science Funds from Fujian Province, China (Grant No. 2015J01236), and the Funds for Young Key Program of Education Department from Fujian Province, China ( Grant No. JZ160425 ) 24 6. References Alizadeh, S., Brandt, M. W., & Diebold, F. X. (2002). Range‐based estimation of stochastic volatility models. Journal of Finance, 1047-1091. Aljarah, I., Faris, H., Mirjalili, S., & Al-Madi, N. (2016). Training radial basis function networks using biogeography-based optimizer. Neural Computing and Applications, 1-25. Andersen, T., & Bollerslev, T. (1998). Answering the skeptics: Yes, standard volatility models do provide accurate forecasts. International economic review, 885-905. Andreou, P. C., Charalambous, C., & Martzoukos, S. H. (2008). Pricing and trading European options by combining artificial neural networks and parametric models with implied parameters. European Journal of Operational Research, 1415–1433. Baxter, M., & King, R. G. (1999). Measuring Business Cycles Approximate Band-Pass Filters for Economic Time Series. Review of Economics and Statistics, 575-593. Bollerslev, T. (1990). Modelling the coherence in short-run nominal exchange rates: a multivariate generalized ARCH model. The review of economics and statistics, 498-505. Boyacioglu, M. A., & Avci, D. (2010). An Adaptive Network-Based Fuzzy Inference System (ANFIS) for the prediction of stock market return: The case of the Istanbul Stock Exchange. Expert Systems with Applications, 7908–7912. Brandt, M., & Jones, C. (2006). Volatility forecasting with range-based EGARCH models. Journal of Business and Economic Statistics, 61–74. Charalambous, C. (June 1992). Conjugate gradient algorithm for efficient training of artificial neural networks. IEE Proceedings G - Circuits, Devices and Systems, 301 - 310. Gorr, W. L. (June 1994). Research prospective on neural network forecasting. International Journal of Forecasting, 10(1), 1-4. Granger, C. W. (1980). Long memory relationships and the aggregation of dynamic models. Journal of econometrics 14.2, 227-238. Granger, C. W., & Joyeux, R. (1980). An introduction to long‐memory time series models and fractional differencing. Journal of time series analysis, 15-29. Hagan, M. T., & Menhaj, M. B. (1994). Training feedforward networks with the Marquardt algorithm. IEEE transactions on Neural Networks, 989-993. Hajizadeh, E., Seifi, A., Zarandi, M. F., & Turksen, I. (2012). A hybrid modeling approach for forecasting the volatility of S&P 500 index return. Expert Systems with Applications, 431–436. 25 Harris, R. D., Stoja, E., & Yilmaz, F. (2015). A cyclical model of exchange rate volatility. Journal of Banking & Finance, 3055-3064. Hodrick, R. J., & Prescott, E. C. (1997). Postwar US business cycles: an empirical investigation. Journal of Money, credit, and Banking, 1-16. Jesús, R. J. (2016). A method with neural networks for the classification of fruits and vegetables. Soft Computing, 1–14. Jia, Z., Yi, C., Yuan, Y., Xuemei, D., & Yuhua, L. (2017). Coarse and fine identification of collusive clique in financial market. Expert Systems with Applications, 225–238. Jiang, Y., Ahmed, S., & Liu, X. (2016). Volatility forecasting in the Chinese commodity futures market with intraday data. Review of Quantitative Finance and Accounting, 1-51. Kristjanpoller, & Minutolo. (2016). Forecasting volatility of oil price using an artificial neural network- GARCH model. Expert Systems With Applications, 65, 233–241. Kristjanpoller, W., Fadic, A., & Minutolo, M. (2014). Volatility forecast using hybrid Neural Network models. Expert Systems with Applications, 2437–2442. Lee, G. G., & Engle, R. F. (1999). A Permanent and Transitory Component Model of Stock Return Volatility. In R. Engle, & H. White, Cointegration, Causality, and Forecasting: A Festschrift in Honor of Clive W. J. Granger (pp. 475-497). Oxford: Oxford: Oxford University Press. Liu, Q., Yin, J., Leung, V. C., Zhai, J.-H., Cai, Z., & Lin, J. (2016). Applying a new localized generalization error model to design neural networks trained with extreme learning machine. Neural Computing and Applications, 59–66. McDonald, S., Coleman, S., McGinnity, Li, Y., & Belatreche, A. (2014). A comparison of forecasting approaches for capital markets. 2104 IEEE Conference on Computational Intelligence for Financial Engineering & Economics (CIFEr). London: IEEE. McDonald, S., Coleman, S., McGinnity, T. M., & Li, Y. (2013). A hybrid forecasting approach using ARIMA models and self-organising fuzzy neural networks for capital markets. The 2013 International Joint Conference on Neural Networks (IJCNN). Dallas: IEEE. McElroy, T. (2008). Exact Formulas for the Hodrick-Prescott Filter. Econometrics Journal, 209–217. Mustafaraj, Lowry, & Chen. (2011). Prediction of room temperature and relative humidity by autoregressive linear and nonlinear neural network models for an open office. Energy and Buildings, 1452–1460. Nelson, M., Hill, T., Remus, W., & O'Connor, M. (1999). Time series forecasting using NNs: Should the data be deseasonalized first? Journal of Forecasting , 359–367. 26 Poon, S.-H., & Granger, C. W. (2003). Forecasting Volatility in Financial Markets: A Review. Journal of Economic Literature, 41(2), 478-539. Ravn, M. O., & Uhlig, H. (2002). On adjusting the Hodrick–Prescott filter for the frequency of observations. Review of Economics and Statistics, 371-375. Restrepo, A. R., Manotas, D. F., & Lozano, C. A. (2016 ). Self-generation of electricity, assessment and optimization under the new support schemes in colombia. IEEE Latin America Transactions, 1308 - 1314. Rubio. (2016). Least square neural network model of the crude oil blending process. Neural Networks, 88– 96. Rubio, J. d., Elias, I., Cruz, D. R., & Pacheco, J. (2016). Uniform stable radial basis function neural network for the prediction in two mechatronic processes. Neurocomputing, 122–130. Sharda, R., & Patil, R. B. (Oct 1992). Connectionist approach to time series prediction: an empirical test. Journal of Intelligent Manufacturing, 317–323. Siegelmann, H., Horne, B., & Giles, C. (1997 ). Computational capabilities of recurrent NARX neural networks. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 208 - 215. Stock, J. H., & Watson, M. W. (1999). Forecasting Inflation. Journal of Monetary Economics, 293–335. Stock, J., & Watson, M. (2016). Dynamic Factor Models, Factor-Augmented Vector Autoregressions, and Structural Vector Autoregressions in Macroeconomics. In J. B. Taylor, & H. Uhlig, Handbook of Macroeconomics (pp. Volume 2, 415–525). ELSEVIER. Svalina, I., Galzina, V., Lujić, R., & Šimunović, G. (2013). An adaptive network-based fuzzy inference system (ANFIS) for the forecasting: The case of close price indices. Expert Systems with Applications, 6055–6063. Tim, B. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of econometrics 31.3, 307-327. Torben, A., Tim, B., & Francis, D. (2009). Parametric and nonparametric measurements of volatility. In Y. Aït-Sahalia, & L. P. Hansen, Handbook of Financial Econometrics. Amsterdam: North-Holland. W, B. M., & S, J. C. (2006). Volatility forecasting with range-based EGARCH models. Journal of Business & Economic Statistics, 470-486. Walker, G. (1931). On Periodicity in Series of Related Terms. Proceedings of the Royal Society of London, 518–532. 27 Yi, C., Yuhua, L., Sonya, C., Ammar, B., & Martin, M. (2015). Adaptive Hidden Markov Model With Anomaly States for Price Manipulation Detection. IEEE Transactions on Neural Networks and Learning Systems, 318 - 330. Yule, G. U. (1927). On a Method of Investigating Periodicities in Disturbed Series, with Special Reference to Wolfer's Sunspot Numbers. Philosophical Transactions of the Royal Society, 267–298. Zhai, J., Cao, Y., Yao, Y., Ding, X., & Li, Y. (2017). Computational intelligent hybrid model for detecting disruptive trading activity. Decision Support Systems, 26–41. Zhang, G., & Qi, M. (16 January 2005). Neural network forecasting for seasonal and trend time series. European Journal of Operational Research, 501–514. Zhang, P., & Qi, M. (2005). Neural network forecasting for seasonal and trend time series. European Journal of Operational Research, 501–514. 
work_iccog4m2mjaotb6qyok52fxkki	----	1 Civil Engineering and Environmental Systems, Vol. 22, No. 4, 2005, pp. 205-215 A Prototype Expert System for Site Selection of a Sanitary Landfill K.W. Chau Department of Civil & Structural Engineering, The Hong Kong Polytechnic University, Hunghom, Kowloon, Hong Kong Abstract It is desirable to incorporate heuristic and empirical knowledge including hydrological and bio-geochemical considerations into the selection process of a potential landfill site. In this paper, a prototype expert system for the selection of a landfill site, with hybrid knowledge representation approach under object-oriented design environment in a blackboard architecture, is described. It incorporates an artificial neural network for training of partial hazardous scores and a fuzzy rule base for the representation of heuristic knowledge. The evaluation is based on the hazardous waste site ranking system recommended by the U.S. Environmental Protection Agency, adapted to Hong Kong conditions by incorporating the stipulation of some local regulations. It is shown to be a useful aid to assist novice engineers in the selection process of a potential landfill site during preliminary investigation. Keywords: Artificial neural network; Blackboard architecture; Expert system; Fuzzy inference; Hazardous waste; Sanitary landfill; Site selection Introduction A sanitary landfill is one of the most popular refuse disposal means in an urban environment. In order to alleviate the nuisances to public health or safety, extensive studies have to be carried out to select the most feasible location. This site selection process involves the study of a proliferation of factors including hydrological as well as bio-geochemical considerations such as composition of contained wastes, possible migration paths of the wastes, soil properties, planned use of the landfill, population and land use, etc. The evaluation process involves many decisions to be made by the designer based on heuristics. It is difficult for various teams composed of engineers and scientists to possess the broad expertise and knowledge in performing such site selections in a consistent manner. Specifically, a novice engineer may face many difficulties in the design process. It is desirable to encapsulate this knowledge into the decision making process. Previously, some algorithmic models were developed to deal with this problem, such as Leão et al. (2004), Gray et al. (2005), Shin et al. (2005), Tiruta-Barna et al. (2005), etc. In many instances, it is often difficult for anyone other than the model developer to use the model successfully. Moreover, heuristic knowledge is not necessarily expressed in an algorithmic manner. In fact, empirical rules, often being incomplete, cannot be easily placed in specific frameworks. With the advancement of artificial intelligence (AI), an expert system furnishes a solution to this decision making process through the incorporation of symbolic knowledge processing on the basis of pertinent heuristic rules. During the past decade, the potential of AI techniques for providing assistance in the solution of engineering problems has been recognized. Expert systems are considered suitable for solving problems that demand considerable expertise, judgment or rules of thumb. Areas of early applications of expert system technology include medical diagnosis, mineral This is the Pre-Published Version. 2 exploration and chemical spectroscopy. In recent years, expert systems have been applied to emulate domain problems in a variety of fields (Chau & Ng 1996, Ranga Rao & Sundaravadivelu 1999, Chau & Chen 2001, Lin & Albermani 2001, Chau & Anson 2002, Kao & Adeli 2002, Chau 2004, Chau & Albermani 2004, Yao et al. 2005). The blackboard architecture, appropriate in domains characterized by interaction between diverse knowledge sources, is one of the most popular systems in the implementation of expert systems in solving a wide range of tasks: control (Hayesroth 1985), speech recognition (Engelmore and Morgan 1988), dynamic rescheduling (Bharadwaj et al. 1994), industrial building design (Kumar 1995), interlaminar stress analysis of composite laminates (Adeli and Yu 1995), crankshaft design (Lander et al. 1996), damage assessment of steel bridge (Barai and Pandey 2000), control of a cryogenic cooling plant (Linkens et al. 2000), large space structures (Kao and Adeli 2002), etc. However, to the author’s knowledge, an expert system addressing site selection of a sanitary landfill has never been reported in the literature. The programming language designed for AI (LISP or PROLOG) entails tremendous programming effort. Nowadays, it is common in the literature that expert system shells, which can provide some knowledge representation methods and inference mechanisms, are used as tools to facilitate the development of expert systems. This paper delineates a prototype expert system for the selection of a landfill site, LANDFILL, which has been developed using an expert system shell VISUAL RULE STUDIO (Rule Machines Corporation 1998) under the Microsoft Visual Basic programming environment. The blackboard architecture with hybrid knowledge representation techniques including production rule system and object-oriented approach is adopted. An artificial neural network (ANN) for training of partial hazardous scores and a fuzzy rule base for the representation of heuristic knowledge are incorporated. The expert system developed is based on the uncontrolled hazardous waste site ranking system recommended by the US Environmental Protection Agency (1984), adapted to Hong Kong conditions by incorporating the stipulation of some local regulations, including Air Pollution Control Ordinance, Waste Disposal Ordinance, Waste Disposal (Chemical Waste) (General) Regulation, Water Pollution Control Ordinance, etc. Solution strategies and development techniques of the system are addressed and discussed. The innovation and original contribution of this manuscript is mainly on hybrid application of the latest AI technologies: expert system; ANN; and, fuzzy inference system, in this specific problem domain. Domain knowledge In general, landfills offer an economic and viable solution to the waste disposal problem. During the landfill operation process, engineering principles are applied to confine the refuse to the smallest practicable size, and to cover it with layers of soil cover at the end of daily operation or an interim cover at a higher frequency as appropriate. This operation requires systematically depositing, compacting, and covering the wastes in compliance with specifications such as placing 150 to 300 mm of soil over every 600 mm of compacted fill. In addition, the soil cover should have a minimum designated depth and the final cover should be grassed to prevent erosion. A common projected ultimate land use of a landfill is a park with recreational facilities that are not affected by long-term ground subsidence. Thorough investigation has to be undertaken to select the most appropriate location for a landfill site. In the site selection process, though there may exist much more factors, four main factors are considered, namely, potential migration routes of the wastes, waste characteristics, planned features of the landfill, and potential targets at risks (Noble 1976). 3 For potential migration routes, the study of contaminant movement away from the disposal site through air, surface water and ground water is focused on. The movement of wastes depends not only on physical, geotechnical, geographic and environmental characteristics of the site, but also on the waste characteristics. This feature can be quantified by its toxicity, persistence, corrosiveness, reactivity, flammability, radioactivity, solubility and volatility. The planned features of the facility will affect the degree of contamination. Potential targets subsume population centers, critical habitats and sensitive ecological systems in its vicinity. The knowledge can be represented so as to minimize or prevent a contaminant from entering into a potential migration route and arriving at a potential target at risk. The site selection process involves three types of factor, namely, (i) physical, geographic, climatic, soil, water quality and socioeconomic conditions; (ii) political considerations, guidelines, standards and regulations established by the pertinent authorities; and (iii) expert judgment and heuristics (US Environmental Protection Agency 1984). In this process, the expert plays a key role since an enormous expert effort is required in the synthesis and analysis of these components. The following tasks are also required, namely, designing the general scheme and a uniform ranking procedure, identifying constraining regulations, analyzing data, and selecting a specific landfill site together with size. The judgments and expertise employed by experts in solving the domain problem are then translated into a set of explicit rules. LANDFILL LANDFILL is a prototype expert system designed to provide surrogate consultation during preliminary hazardous waste site investigations. It provides a versatile framework for the interpretation, classification and diagnosis of environmental conditions at waste disposal sites. The objective of such a consultation is to obtain a site rating using the expert rules and the decision logic described in the system’s knowledge base. The ranking criteria are based on: relative risk or danger, taking into account the population at risk; the hazardous potential of the substances at a facility; the potential for contamination of drinking water supplies, for direct human contact, and for destruction of sensitive ecosystems; and other appropriate factors. The system is compiled and encrypted to create a run-only system. This run-only system is installed on a microcomputer for office use. The user can always overrule any options and recommendations provided by the system. However, a mechanism is built-in to ensure that the user’s overruled input is reasonable and consistent. In other words, it plays the role of a knowledgeable assistant only. Besides the usual components in a typical expert system, namely, knowledge base, inference mechanism, session context, user interface, knowledge acquisition and explanation modules, it also incorporates an ANN tool, fuzzy rule system, and a database. The schematic diagram of this prototype system is shown in Figure 1. Knowledge acquisition and representation Knowledge plays an important role in an expert system. The knowledge used has been acquired from written documents such as codes of practice, textbooks and design manuals and complemented by interviews with ten experts, from whom fairly consistent views were acquired within a six month period. In order to acquire knowledge, it is better to work with the expert in the context of solving particular problems, instead of directly posing questions about rules. Hybrid knowledge representation schemes, including object-oriented programming, procedural methods, and production rules are employed to express engineering 4 heuristics and standard design knowledge. This approach renders it possible to take advantage of the characteristics of each method and to tailor for each type of domain knowledge in the knowledge base. Object-oriented programming Figure 2 shows the details of the blackboard architecture, which are classified into knowledge modules and the blackboard. Knowledge modules corresponding to procedural expertise knowledge are divided into Decision Process and Process Control whilst objects in the blackboard are classified into Decision Stage and Decision Entities. The blackboard is partitioned into a number of hierarchical levels, corresponding to different stages of the decision process. Decision Stage only comprises a single object whereas there are several objects in the Decision Entities level. Data inside the Decision Stage are employed by the Process Control knowledge modules to determine the next possible action, or to check the validity of the function triggered by the user. Forward chaining inference mechanism is employed here to derive the next process. After a specific decision stage has been satisfied, the pertinent Decision Stage indicator will be assigned one of the preset values, which are numerical values from 0 to 9. Decision Process modules determine largely the scope of decision to be solved by the expert system. The attached procedural method is processed when the value of the attribute changes, either by assignment under another method or by the user. A mixed problem-solving strategy is used here. The user is required merely to supply the relevant data during each decision stage and the system will determine the order in which different decision knowledge modules are executed. Process Control modules ensure the proper and effective application of knowledge in Decision Process modules and undertake conflict resolution. They evaluate the current attribute values in Decision Stage of the blackboard, which provides the indicator to assist this decision making. The Main Decision Process class monitors the decision stage of all key tasks during the decision process and decides either to continue to next step or to prompt a warning message. All primary tasks in Process Control module are expressed on command buttons together with procedural methods attached. Process Control knowledge modules work closely with the user-interface module to produce user-friendly main menu displays. Moreover, the relevant entries and decision parameters under Decision Entities, the corresponding attribute values of Decision Stage are synchronized through the Process Control knowledge modules. Production rules Some heuristic knowledge is represented in the IF/THEN/ELSE production rules with confidence factors that can be assigned either automatically, or in response to the user’s request. These rules are a formal way of specifying how an expert reviews a condition, considers various possibilities, and recommends an action. The explicit expertise under the production rule format in the knowledge base are employed to rank the potential landfill locations to identify the more feasible sites for further detailed studies. The following is a typical example of the production rules. Rule to find HazardGroundwaterDepth: 1 of 8 IF GroundwaterRoute.DepthOfGroundwater >= 6 AND GroundwaterRoute.DepthOfGroundwater < 25 5 THEN HazardScore.HazardGroundwaterDepth:= risky CF 70 In the production rules, the confidence factor (CF) is employed as the determining factor to control the inference process for the evaluation of each parameter. The range is from 0 to 100, representing the degree of confidence with which the statement is known. A higher value represents higher degree of confidence and hence is better. The confidence factors are set by weighted opinions of various experts based on their heuristic and experience. Depending on the responses of the user, appropriate rules will be executed. The system accounts for the facts that describe the domain knowledge on preliminary landfill site selection. The production rule also incorporates the fuzzy description. For some continuously varying conditions, such as the total population or net seasonal precipitation, the user can specify them with fuzzy descriptions or with definite numerical values. The system can automatically transfer the numerical values into fuzzy descriptions with the fuzzy membership curve to calculate its relevant confidence of membership before searching the rule base. Figure 3 shows the fuzzy description of total population within 300 m of the site with different curves representing the definitions of “very low”, “low”, “medium”, “high” and “very high”, respectively. The membership functions are also set by weighted opinions of various experts. All landfill site selection parameters are categorized into four independent types of factors, namely, groundwater route characteristics, targets at risk, facility characteristics and waste characteristics. The groundwater route characteristics hazard score (Sgr) considers a number of parameters including the depth of the groundwater at the aquifer, net seasonal precipitation, soil permeability, physical state of the waste at the time of disposal, aquifer soil type, the depth to bedrock, geology, seismicity, faults, stratification, heterogeneity and isotropy. The target score (St), representing the level of risk to potential targets, is related to the potential use of the aquifer of concern, the vicinity of nearest production wells, and nearby population served by the groundwater. The executed assigned value about the planned containment characteristics of the proposed landfill facility is defined as the facility characteristics score (Sfc).The waste characteristics hazard score (Swc) considers the toxicity, persistence, corrosiveness, reactivity, ignitability, radioactivity, solubility and volatility of each potential waste substance, along with its projected disposed quantity. The substance with the highest sum of the assigned values among all projected waste substances is selected as the representative waste material. The user has to choose appropriate answers to all these questions in order to assign a hazard rating score to each parameter. The expert system then matches the selected answers with each rule, executes the appropriate ones, and computes the four partial scores from an ANN. The weighted sum of executed assigned hazard values represents independent variables. It ultimately derives the overall hazard rating score of the site (S), as follows: wcfctgr SwSwSwSwS 4321  (1) where w1, w2, w3, and w4 are the weights of the partial score respectively, with values assigned by expert opinion after iterated cycles of training and validation of the ANN. The overall site hazard rating score is normalized between 0 and 100, which is utilized as a measure for the relative ranking of the specific site. A lower value represents a safer potential site location. The scores can be employed as a yardstick for the relative safety of different 6 sites. In a preliminary study, locations with high scores might be eliminated as potential landfills, whereas sites with low scores can remain for further comprehensive investigation. Artificial neural network (ANN) An ANN is used as the learning mechanism to transfer engineering experience into knowledge in determining the hazard scores. The back-propagation learning algorithm is employed to train the network for extracting knowledge from training examples (Rumelhart et al. 1994). An ANN architecture with 36 inputs (one attribute and one confidence factor for each feature), 4 outputs (one for each partial score), and one hidden layer of 18 nodes is created for the problem. The learning control parameters, including learning rate = 1.0 and momentum factor = 0.5, are chosen to control learning process. The initial network weights are assumed with uniformly distributed random values from the interval -0.5 to 0.5. The S- shaped sigmoid curve, as shown in eq. (2), is used as a transfer function on each neuron to represent the input-output relation in the hidden layer and output layer whilst a linear function is employed for the input layer. xe xS   1 1 )( (2) The major advantage of employing eq. (2) is that it can normalize all the input data into the range between 0 and 1, which is more manageable in terms of data interpretation. The normalized root-mean-square error (NRMSE) between target and output results is computed to evaluate the training performance. Let N be the number of testing example, Tij, Oij be the target values and the computed value of the ith test example and jth output node respectively, and jT be the average target value of j th output node, then the definition of the above- mentioned statistical quantity is as follows:          N i j jij N i j ijij TT OT NRMSE 1 4 1 2 1 4 1 2 )( )( (3) Figure 4 shows the relationship between NRMSE and number of training cycles. It is found that the error rate is converged in about 30 cycles. Self-learning mechanism is accomplished by the use of ANN. System validation is performed through the validation process of the ANN and by comparison of the results with those by the experts. The knowledge base is dynamic and if more input-output data pairs are provided by other experts in different locations of the world, the generalization capability of the ANN will yield different output results. Inference engine The inference engine controls the strategies that determine how, from where, and in what order, a knowledge base draws its conclusions. It controls the selection of procedure methods and production rules from the knowledge base to derive a conclusion or decision context. All the decision steps can be seen explicitly on the main screen display. The validity of the user’s choice on the preferred sequence of decision processes is checked by Process Control knowledge modules, which act opportunistically upon being triggered. An event-driven inference processing mechanism is adopted so that the ensuing action of the system will 7 depend on the input made by the user. For example, the factors affecting the site characteristics are considered depending on the nature of ground in the vicinity. If rock has been selected, the user is prompted to enter whether or not there are any faults. If soil has been selected, the above question will not be asked. After the network is trained, this system is capable to diagnose new cases of site selection, even for the cases that some input values are unknown owing to unavailable or missing records. A hybrid reasoning strategy that combines forward and backward reasoning schemes is used in order to arrive at a reasonable conclusion with minimum information. Forward chaining inference is employed to infer the output values given the current input values, part of which may be unknown initially. The prototype arrives at a conclusion when no unknown input can alter the current decision significantly. However, if the system has not yet arrived at a certain conclusion with a defined threshold value of confidence factor, backward chaining reasoning is employed. The system automatically highlights the unknown input units that have a significant effect on the current most plausible conclusion, and prompt the user to enter their values. User interface The system offers a friendly user interface. Whilst input data entries are kept at minimum, they are provided by the user mostly through selection of appropriate values of parameters from the menus and answers to the queries made by the system. If the input data provided by the user is not within the specified range, it will be rejected and a warning message will be prompted. The system provides a contemporary multi-window graphics text display, which is valuable to novice engineers. Case example A case example is employed to demonstrate the application of the system. The overall landfill hazard rating of a potential rural site located at Tseng Kwan O is evaluated. User-friendly displays are used to interact with end users by prompting for values and showing the output data. The sample run commences with inputs on various groundwater route characteristics in the main menu. The depth of groundwater to aquifer is less than 6 m and the net seasonal precipitation is 2400 mm/year. The aquifer soil type is sandy soil with soil permeability 5 x 10-6 m/s. The depth to bedrock is more than 50 m. The soil properties are homogeneous in all directions without any stratification and there are no faults nearby. The waste is in a solid state and is unconsolidated. It will release unpleasant smell with pH value of 8. The waste is not inflammable and its temperature is below 25C. Its toxicity and radioactivity are low. Its total quantity is about 500 tons/m3. The landfill will have a non-permeable liner and a leachate collection system. The land use at the vicinity of the site is mainly public area and there are no potential hazardous installations within 1 km of the site. The total population within 300 m of the site is about 50 people. The distance to the nearest drinking water well is about 200 m. The system transforms these production rules into user-friendly interactive menus. Based on the responses of the user, the system searches the knowledge base and generates the overall landfill site rating score, which is shown in Figure 5. In general, a range of scores between 0 and 50 would indicate a site to be feasible for sanitary landfill. In this case, the score of 41 is of medium value and it is feasible as a site for sanitary landfill. The recommendation is that detailed investigations can be further undertaken, which was verified by independent but consistent assessments of experts. Conclusions 8 The integration of the heuristic and empirical knowledge into a decision support system is useful in the selection process of a potential landfill site, since a simple score can represent a diversity of complicated factors. A prototype expert system, which assists in making decisions on selection of an appropriate landfill site, was developed and implemented. It incorporates an ANN for training of partial hazardous scores and a fuzzy rule base for representation of the heuristic knowledge. It is shown that the hybrid application of these latest AI technologies is appropriate to act as storage for empirical knowledge so that advice on landfill site selection can be furnished in the preliminary design stage. The evaluation of the prototype system is based on the hazardous waste site ranking system recommended by the U.S. Environmental Protection Agency, adapted to Hong Kong conditions by incorporating the stipulation of some local regulations. The knowledge base is transparent and can easily be updated, which renders the expert system an ideal tool for incremental programming. Increase in efficiency, improvement, consistency of results and automated record keeping are among the advantages of such a system. Furthermore, the educational spin-off of an expert system in training novice engineers or in transferring knowledge is significant. References Adeli, H. and Yu, G. 1995. An integrated computing environment for solution of complex engineering problems using the object-oriented programming paradigm and a blackboard architecture. Computers & Structures, 54(2): 255-265. Barai, S.V. and Pandey, P.C. 2000. Integration of damage assessment paradigms of steel bridges on a blackboard architecture. Expert System with Applications, 19(3): 193-207. Bharadwaj, A., Vinze, A.S. and Sen, A. 1994. Blackboard architecture for reactive scheduling. Expert System with Applications, 7(1): 55-65. Chau, K.W. 2004. A prototype knowledge-based system on unsteady open channel flow in water resources management. Water International, 29(1): 54-60. Chau, K.W. and Albermani, F. 2004. An expert system on design of liquid-retaining structures with blackboard architecture. Expert Systems, 21(4): 183-191. Chau, K.W. and Anson, M. 2002. A knowledge-based system for construction site level facilities layout. Lecture Notes in Artificial Intelligence, 2358: 393-402. Berlin: Springer. Chau, K.W. and Chen, W. 2001. An example of expert system on numerical modelling system in coastal processes. Advances in Engineering Software, 32(9): 695-703. Chau, K.W. and Ng, V. 1996. A knowledge-based expert system for design of thrust blocks for water pipelines in Hong Kong. Journal of Water Supply Research and Technology - Aqua, 45(2): 96-99. Engelmore, R. and Morgan, T. 1988. Blackboard systems. Wokingham: Addison_Wesley. Gray, D., Pollard, S.J.T., Spence, L., Smith R. and Gronow, J.R. 2005. Spray irrigation of landfill leachate: estimating potential exposures to workers and bystanders using a modified air box model and generalised source term. Environmental Pollution, 133(3): 587-599. Hayesroth, B. 1985. A blackboard architecture for control. Artificial Intelligence, 26(3): 251- 321. Kao, W.M. and Adeli, H. 2002. Multitasking object-oriented blackboard model for design of large space structures. Engineering Intelligent Systems for Electrical Engineering and Communications, 10(1): 3-8. Kumar, B. 1995 Knowledge Processing for Structural Design, Topics in Engineering Volume 25, ed. Brebbia, C.A. and Connor, J.J. Southampton: Computational Mechanics. Lander, S.E., Corkill, D.D. and Staley, S.M. 1996. Designing integrated engineering 9 environments: blackboard-based integration of design and analysis tools. Concurrent Engineering Research and Applications, 4(1): 59-71. Leão, S., Bishop, I. and David Evans, D. 2004. Spatial-temporal model for demand and allocation of waste landfills in growing urban regions. Computers, Environment and Urban Systems, 28(4): 353-385. Lin, S. and Albermani, F. 2001. Lattice-dome design using a knowledge-based system approach. Computer-aided Civil and Infrastructure Engineering, 16(4): 268-286. Linkens, D.A., Abbod, M.F., Browne, A. and Cade, N. 2000. Intelligent control of a cryogenic cooling plant based on blackboard system architecture. ISA Transactions, 39 (3): 327-343. Noble, G. 1976. Sanitary Landfill Design Handbook : the science and art of site selection, investigation, and design. Westport, CT: Technomic Publishing Co. Ranga Rao, A.V. and Sundaravadivelu, R. 1999. A knowledge based expert system for design of berthing structures. Ocean Engineering, 26(7): 653-673. Rumelhart, D.E., Widrow, B. and Lehr, M.A. 1994. The basic ideas in neural networks. Communications of the ACM, 37(3): 87-92. Rule Machines Corporation 1998. Developer’s Guide for Visual Rule Studio. Indialantic: Rule Machines Corporation. Shin, H.C., Park, J.W., Kim, H.S. and Shin, E.S. 2005. Environmental and economic assessment of landfill gas electricity generation in Korea using LEAP model. Energy Policy, 33(10): 1261-1270. Tiruta-Barna, L., Rethy, Z. and Barna, R. 2005. Release dynamic process identification for a cement based material in various leaching conditions. Part II. Modelling the release dynamics for different leaching conditions. Journal of Environmental Management, 74(2): 127-139. US Environmental Protection Agency 1984. Uncontrolled Hazardous Waste Site Ranking System, A User Manual, HW-10. Washington D.C.: USEPA. Yao, L., Postlethwaite, I., Browne, W., Gu, D., Mar, M. and Lowes, S. 2005. Design, implementation and testing of an intelligent knowledge-based system for the supervisory control of a hot rolling mill. Journal of Process Control, 15(6): 615-628. 10 Figure 1. Schematic diagram of the prototype expert system Microsoft Windows Environment Visual Basic Programming Environment Knowledge Acquisition Module Expert System Shell (Visual Rule Studio) User Interfaces Explanation Module Knowledge Engineer Interview with Expert Knowledge from Literature User Training Input/ Output Data Pairs Artificial Neural Network Tool Microsoft Access Databases System Interfaces Defuzzification interface Knowledge Base Inference Engine Backward Chaining Forward Chaining Blackboard (Context) Knowledge Modules Decision Process Process Control Decision Entities Decision Stage Procedural Methods Fuzzification interface Requirements of Code of Practice 11 Legends: Figure 2. Details of the blackboard architecture Agenda triggered by user or situation Knowledge modules Blackboard partitioned into hierarchical levels Decision Stage 1st Decision Entities Level (n-1)th Decision Entities Level nth Decision Entities Level DPKM DPKM DPKM DPKM PCKM DPKM Inference Mechanism Decision context Decision Process Knowledge modules Process Control Knowledge modules DPKM PCKM Forward Chaining Backward Chaining 12 0 0.2 0.4 0.6 0.8 1 0 50 100 Total p op ulation within 300 m F u zz y m em b er sh ip f u n ct io n Very Low Low M edium High Very High Figure 3. Fuzzy description of total population within 300 m of the site with different curves 13 0 0.1 0.2 0.3 0.4 0.5 0 20 40 60 80 100 Number of training cycles N R M S E Figure 4. Relationship between NRMSE and number of training cycles 14 Figure 5. Screen displaying overall landfill site hazard rating score 
work_icczmocbzrcslf6mnqixythg2i	----	Assessing Organizational Readiness for Implementing ERP System Using Fuzzy Expert System Approach | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.4018/IJEIS.2017010105 Corpus ID: 4316499Assessing Organizational Readiness for Implementing ERP System Using Fuzzy Expert System Approach @article{Hajilari2017AssessingOR, title={Assessing Organizational Readiness for Implementing ERP System Using Fuzzy Expert System Approach}, author={Amir Beirami Hajilari and M. Ghadaksaz and Gholamreza Soltani Fasghandis}, journal={Int. J. Enterp. Inf. Syst.}, year={2017}, volume={13}, pages={67-85} } Amir Beirami Hajilari, M. Ghadaksaz, Gholamreza Soltani Fasghandis Published 2017 Engineering, Computer Science Int. J. Enterp. Inf. Syst. Efficient use of enterprise resource planning systems is the only way to achieve competitive advantage in many industries. However, many reports indicate high failure rate of ERP implementation projects and the lack of access to benefits and advantages that enterprises have expected from ERP deployment. Managers are concerned about being ready to deploy such a system. Despite its enormous cost, implementing enterprise resource planning systems fail in practice. Accordingly, this study aims at… Expand View via Publisher igi-global.com Save to Library Create Alert Cite Launch Research Feed Share This Paper 2 Citations View All Topics from this paper ERP Expert system Enterprise resource planning Failure rate Usability Software deployment 2 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency Application of FMEA to Study the Risk Perception of SMEs Throughout the ERP Adoption Life Cycle S. Bharathi, Kanchan Chandrayan Business, Computer Science Int. J. Enterp. Inf. Syst. 2017 Save Alert Research Feed Successful Strategies for Energy Sector Enterprise Resource Planning Projects Dora Kwei Arko Business 2019 1 Save Alert Research Feed References SHOWING 1-10 OF 51 REFERENCES SORT BYRelevance Most Influenced Papers Recency Developing a practical framework for ERP readiness assessment using fuzzy analytic network process J. Razmi, M. S. Sangari, R. Ghodsi Computer Science Adv. Eng. Softw. 2009 95 Save Alert Research Feed Developing a Practical Framework for ERP Project Implementation: A Proposed Research Design J. Sullivan, Mela Wyeth, Wade M. Chumney Computer Science CONFENIS 2006 4 PDF Save Alert Research Feed Identification of critical failure factors in the implementation of enterprise resource planning (ERP) system in Taiwan's industries Wen-Hsien Tsai, Shih-Wen Chien, P. Hsu, J. Leu Business 2005 55 Save Alert Research Feed Critical success factors for ERP implementation in SMEs M. Ahmad, Ruben Pinedo Cuenca Engineering 2013 101 View 2 excerpts, references background Save Alert Research Feed Identification and classification of ERP critical failure factors in Iranian industries Amin Amid, Morteza Moalagh, Ahad Zare Ravasan Business, Computer Science Inf. Syst. 2012 181 PDF View 1 excerpt, references background Save Alert Research Feed Why are enterprise resource planning systems indispensable to supply chain management? Y. Su, Chyan Yang Business, Computer Science Eur. J. Oper. Res. 2010 117 PDF Save Alert Research Feed Achievement assessment for enterprise resource planning (ERP) system implementations based on critical success factors (CSFs) Albert Y Sun, A. Yazdani, John D. Overend Computer Science 2005 243 PDF Save Alert Research Feed ERP success prediction: An artificial neural network approach S. Rouhani, Ahad Zare Ravasan Computer Science 2012 36 Save Alert Research Feed Identifying critical issues in enterprise resource planning (ERP) implementation Ike C. Ehie, M. Madsen Engineering, Computer Science Comput. Ind. 2005 467 PDF View 1 excerpt, references background Save Alert Research Feed Enterprise resource planning: A taxonomy of critical factors M. Al-Mashari, A. Al-Mudimigh, M. Zairi Computer Science Eur. J. Oper. Res. 2003 1,037 Save Alert Research Feed ... 1 2 3 4 5 ... Related Papers Abstract Topics 2 Citations 51 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_icie57r4e5gehclfikwxbl2x5q	----	Report 1 An example of expert system on numerical modelling system in coastal processes K.W. Chau and W. Chen Department of Civil & Structural Engineering, Hong Kong Polytechnic University, Hunghom, Kowloon, Hong Kong Abstract With the advent of artificial intelligence technology as well as the widespread popularity of desktop microcomputers in recent years, integration of this new technology with the traditional numerical modelling system becomes a current trend in order to solve various engineering problems. It renders a more intelligent and user-friendly system on the problem domain. In this paper, a knowledge-based expert system on numerical modelling system for coastal water processes is delineated. Expert system application, as a key branch of artificial intelligence technology, is integrated with traditional numerical modelling for simulating flow and water quality phenomenon in coastal waters. The knowledge bases are classified into five major types, namely, a variety of models, relations between various model parameters and real physical conditions, feasible options of model parameters, question base as an user-interface directing the user to depict the actual physical conditions, and the rules of inference deducing the feasible choice of model and its parameters. A hybrid expert system shell, Visual Rule Studio, is employed as an ActiveX Designer under Microsoft Visual Basic environment because it combines the advantages of both production rules and object-oriented programming technology. Both forward chaining and backward chaining are used collectively during the inference process, which is mainly driven by premises and conditions with the highest factors of confidence. The inference engine will drive the decision tree to explore the most probable option of numerical model and parameters matching the real problem specifications. It is shown that the application and integration of the knowledge- based expert system technology into numerical modelling for coastal processes can provide substantial assistance to novice users for selection of numerical model as well as parameters. Keywords Coastal water process, expert system, knowledge base, numerical modelling, object-oriented programming technique, production rule Introduction In order to mimic and predict accurately the changes of flow and water quality conditions in coastal water processes, numerical modelling techniques have often been used as the simulation tool. Over the past few decades, with the advancement of numerical modelling technique, enormous amount of mathematical models have been developed and applied to various hydraulic engineering and environmental engineering problems. However, extensive and detailed expertise knowledge is required before one can distinguish the special features and limitations of these individual numerical models and then select the appropriate model to apply in a particular circumstance. It is found that these existing models always lack in some user-friendly aid for their selection, which will be very useful to novice engineers or engineering students as the design aid or training tool. Nowadays, the development of numerical modelling system has been developed to such a state that it tends to encapsulate more and more recent and advanced computer technology. 1 This is the Pre-Published Version of the article published in Advances in Engineering Software, Vol. 32, No. 9, 2001, pp. 695-703. Abbott (1989) , Abbott (1991) and Cunge (1989) have classified different models into various stages of computer generation. The initial models, which are developed for very specialized scope of problem, can only be used by the developer and other special users after having undergone training of long duration. Gradually the models incorporate more user-friendly tools such as menu of parameters specification, automatic grid formation, pre-processing, post-processing, management of collected field data, etc. It is intended for use by a much wider range of general audience or users to solve a much wider scope of engineering problems in different nature. The current trend is that new models often incorporate recent artificial intelligence technology into the mathematical modelling to form a single system. Since the last decade, some researchers have started to integrate expert system technology with numerical hydraulic and environmental modelling (Blanpain and Chocat, 1994; Chau and Yang, 1992; Chau and Yang, 1993; Chau and Yang, 1994; Chau and Zhang, 1995; Knight and Petridis, 1992; Uzel et. al., 1988). These integrated systems for coastal water processes only dealt with the simplest one-dimensional modelling situation. However, even for that simplest case, the symbolic programming for the knowledge representation and selection procedure required enormous effort. The integration of expert system and two- dimensional or three-dimensional numerical modelling will be much more complicated. The basic requirement is that the system should be able to provide expert advice on selection of the most appropriate model as well as the entailed model parameters under that particular scenario. Since the numerical modelling programs have often been developed in some traditional programming languages such as Fortran, Pascal, C, etc., it is considered not cost- effective to re-write and replace these well-proven and validated programs whose development involved long hours of concerted effort. Instead, the program is often written in some embedded source codes and the integrated system has to be accompanied by a usage wizard, which provides assistance and guidance for use and direction of non-expert users. As such, the use of the recent expert system technology, through establishing the requisite knowledge bases for the problem domain as well as utilizing a proper reasoning inference engine, is found to have high potential in providing assistance for both selection and manipulation of numerical models. Up to date, research studies on the incorporation of expert system technology into manipulation of numerical modelling are, as a matter of fact, very scarce. The key objective of this study is thus to develop an integrated system for coastal water processes, which incorporates the recent expert system technology into traditional numerical models together with some pre-processing and post-processing tools. One of the major requirements of the integrated system for numerical simulation of coastal hydrodynamic and water quality processes is to serve as design aid and training tool for novice users. It will furnish expert advice on the selection of simulation methods as well as their pertinent model parameters. In additional to the usual components in a typical expert system, namely, knowledge bases encapsulating the domain knowledge, inference engine, context, user- interface, knowledge acquisition and explanation modules, it also incorporates executable numerical models and databases. The architecture of the integrated system is shown in Figure 1. Direct user-friendly medium of interaction should be set up between the user and the system so that the user can modify certain domain knowledge inside the knowledge base at his/her will. Output results obtained by executing various numerical models should be interpreted through the aid of post-processing graphic expedients. Based on the knowledge base, which acts as the repository of all the domain knowledge, comparison and evaluation of the 2 performance of different models and thus their relative advantages, applicability as well as limitations can be thoroughly explored. Aids on manipulation of numerical modelling should also be included. The system is verified and validated by application to some real prototype problems in Hong Kong coastal waters, which can, of course, be applicable to any other coastal zones in the world. The application of the expert system technology for selection of mathematical model is presented and delineated in this paper. Many modules with different functions will be integrated into the whole system and some of the tools are linked by employing add-in tools. It is found that it is difficult, if not impossible, to incorporate all these modules into traditional standalone expert system shells. Hence the expert system shell, Visual Rule Studio (VRS), which runs as an ActiveX Designer under the windows based programming language environment Microsoft Visual Basic 6.0 (VB), is employed here. VRS is a hybrid expert system shell, which combines the advantages of both production rules and object-oriented programming paradigm. (Rule Machines Corporation, 1998) VB, being an object-oriented programming language, includes all the usual control objects of the common interface under Windows 98 or 2000 environments such as command button, textbox, option button, check box, etc. The user-friendliness of the Windows-based VB interface is one of the main reasons in choosing this shell. Besides, VB has the capability to link or execute external programs written in other conventional programming languages such as Fortran 90, Pascal or C. More Internet techniques are also incorporated into the system so as to pave the way for its future Internet version. Knowledge acquisition The physics of flow and water quality problems in coastal water processes are often represented by mathematical models, which solve numerically the simulated and discretized partial differential equations. The procedure in undertaking numerical modelling is usually divided into a few steps. The physical law or equations representing the problems is first written out and then discretized. The proper boundary conditions, coefficients of turbulence and other requisite model parameters are then chosen. The appropriate problem-solving algorithm is then adopted to solve the discretized equations subjected to the above constraints. In the above steps, the number of parameters involved in the selection procedure is 40 or so. Hence, based on these steps, a full decision tree of model selection procedure can be composed. The domain knowledge entailed in the development of this expert system has been encoded mainly from literature review and interviews with expert numerical modellers. Experience from the experts are gleaned on a variety of aspects including the logical reasoning in making decision, relative importance of various parameters, individual comments on various mathematical models, their preference choice of models, etc. More discrepancies in the selection manners were found between well-experienced modellers and novice users than what is expected. This also reinforces the potentials of an expert system in assisting the selection of numerical model for coastal water processes. The combination of theoretical knowledge from literature and their own practical experience in numerical modelling render the model selection made by well-experienced modellers more reasonable. They can strike a better balance between accuracy and effectiveness in model simulation. The novice users lack in the crucial practical experience and hence attempt only to follow strictly what the literatures describe. In so doing, a good balance between accuracy 3 and effectiveness cannot be easily achieved. As such, it is very imperative to incorporate more heuristic experience of numerical modelling experts into the expert system. As a matter of fact, it is very difficult to glean experience of various expert modellers in the world. However, the expert system can be further developed and updated through frequent usage and feedback from the users as well as through validation of personal conclusions and experience in the previous literatures on numerical modelling for coastal water processes. In this prototype system, the involved parameters are classified into a total of seven types, namely, dimension and method, numerical scheme, grid formulation, boundary conditions, initial conditions, eddy turbulence, and miscellaneous, as detailed in Table 1. The classification method is designed so as to facilitate the query and reasoning processes in searching the knowledge bases. Besides, it also effects the knowledge compilation and representation with regard to numerical modelling for coastal water processes. The initiative to select a numerical model in solving a physical problem for coastal water processes usually comes from the user. In choosing the numerical model and its pertinent set of parameters, the key factors to be considered are the main purpose of the user in applying the numerical modelling, the real conditions of the physical problem, the previous experience of the user in numerical modelling and the capability of the user in the interpretation of model results. The first two factors, i.e. problem specification and physical conditions, have been considered in most of the previous expert system in numerical modelling. However, in this system, the experience of the user is also taken into account through soliciting his/her comments and deriving some conclusions from his/her responses. The fourth factor, i.e. the interpretation of the model results by the user, depends on the experience of the user and will become important in manipulation of numerical models, which is treated separately. Table 2 shows the classification of user specifications into these three aspects. The selection of the most feasible numerical model becomes a matching process to relate the model parameters listed in Table 1 with the physical and other conditions specified by the user in Table 2. Hence the use of expert system technology is mainly on the establishment of the connection between Table 1 and Table 2, through knowledge bases, rule base and an appropriate reasoning inference engine. Since any numerical model incorporates a specific set of model parameters, the selection of numerical model will inevitably also involves the selection of model parameters. The whole procedure of model selection is represented by a decision tree, which is gradually filled as the selection procedure proceeds forward until the solution to the problem is accomplished. At present, this prototype system incorporates only a few popular and commonly used numerical models. Nevertheless, the selection process based on Table 1 and Table 2 already includes all the succinct features and heuristic experience in model and parameter selection. Anyway, it demonstrates the skeleton of how the expert system technology can be integrated into numerical modelling for coastal water processes. As a future extension, more numerical models can be incorporated into the system and, in that situation, the selection process will be developed with mathematical basis. The rules and the knowledge bases are designed to be transparent such that it can be easily updated with new additional knowledge. That is in fact one of the purposes in adopting expert system approach. One of the key issues in determining the selection of numerical models is to strike a balance between two intrinsically conflicting components, namely, accuracy and effectiveness. Some general concepts on this aspect are also considered and included into the knowledge bases. For example, it is believed that the effectiveness of finite element methods is usually lower 4 than that of finite difference method. Another example is that larger spatial step or larger time step both contribute to higher effectiveness and lower accuracy. Through asking for the requirement of the user on the relative significance of accuracy and effectiveness, the system can search the rule base and recommend the most feasible selection. The knowledge-based expert system acts only as an assistant to guide the decision making process and does not automatically balance the accuracy and effectiveness. The relevant knowledge is encoded into the knowledge base with varying factor of confidence. The selection decision is designed to be driven by the pertinent rule with the highest factor of confidence. Nevertheless, the recent rates of advancement for both the speed and memory size of microcomputers are so fast that the accuracy consideration is expected to be attained relatively easily than the effectiveness aspect in the near future. The balance between accuracy and effectiveness may then be shifted. Flexibility should be allowed for modifications of the knowledge base to reflect that new balance. Knowledge bases The knowledge base of the expert system is divided into a few sub-bases, namely, relation base, selection base, question base, rule base and model base. The relation base contains the knowledge concerning all the factors related to each parameter, which can be represented in the form of relation tree. Figure 2 shows part of the relation tree for model parameters. Although, in general, the tree nodes are usually represented in one layer, they can also be represented in two layers if the related factor is related to some other factors as well. The selection base stores the information regarding the possible options for each parameter and can be represented under the structure of a selection tree. Figure 3 displays part of the selection tree for model parameters. Each node of the selection tree represents an individual parameter and the tree branches represent available options of that parameter. Hence the relation tree and the selection tree are designed to assist in the knowledge representation of the parameter relations and options respectively. During the query process, all the related factors are determined by searching the relation tree. In the decision making process, all the feasible options of each parameter are determined from the selection tree. In fact, most expertise knowledge can be represented in the form of a knowledge tree, which is a static tree with its structure defined before the model selection. Questions used to direct the user to input the problem specifications are grouped together into the question base. The questions are not static and the ensuing questions will depend on the answers made by the user on the previous questions. The relation base is involved to provide the required relevant parameters in order of priority. Figure 4 gives an example of the inference process of an individual parameter to be specified by the user. The queries together with the possible set of answers are derived. Once the user makes a response on the query, it is feedback to the system and the inference engine will drive the selection decision process forward until the final solution is attained. Figure 5 gives a typical example of the inference direction from a specification by the user to the inference by the system. The knowledge representation of the rule base is intrinsically different from those of the relation base or selection base. It encapsulates the whole set of inference production rules linking the specifications made by the user and the selection of parameters. The following example gives a typical production rule, which incorporates the fuzzy description: RULE to determine dimensions of model: 2 of 8 5 IF the water depth is deep AND the density stratification in the vertical direction is not significant THEN the 2-dimensional horizontal numerical model is selected with a confidence factor of 70 The IF statement of the above rule statement describes the premises or conditions that the water depth and the density stratification should be deep and not significantly respectively for the conclusion to be fired. The THEN statement of the rule gives the conclusion that the 2- dimensional horizontal numerical model is selected with a confidence factor 70. There are 300 or so rule statements with fuzzy relations in the rule base of the system. The confidence factor is employed as the determining factor to control the inference process and the selection of each parameter. The range of the factor is basically from 0 to 100 representing the degree of confidence with which the statement is known. The rule base is designed to provide the link between the specifications made by the user and the recommended parameter selection by matching the highest confidence factor. In the above example rule statement, the water depth and the vertical density difference are expressed as a fuzzy description. As an alternative method, the user can also enter their exact numerical values during the query process. The system can then transform the numerical values into the corresponding fuzzy description by a fuzzy member curve, which computes the pertinent confidence of membership prior to searching the rule base for conclusions. Figure 6 shows the fuzzy description of water depth, in which curves A, B and C depict the fuzzy definitions of “very deep”, “deep” and “shallow” respectively. The model base contains a number of popular and well-proven numerical models for simulating coastal water processes. Typical examples of case studies around the world employing numerical models are stored here and can be retrieved easily for study. As such, it effectively gives some references for numerical modelling with prototype examples. When the user completes the selection process of all modelling parameters in a particular case, the system will then attempt to match the models in the model base. If an existing model in the model base is found to have similar characteristics to the specified selection process by the user, the user will be prompted whether or not to adopt the selection based on the that existing model. Inference engine The confidence factor is the key determinant used by the inference engine to drive the selection of various parameters. A default value of 50 is set initially for the confidence factors of all parameter selection, which represents a half-and-half chance for it to proceed successfully. When the whole picture of specifications is gradually completed by the user through the query process, the inference process continues and the new confidence factors of various parameter selection are determined through the matching process. In general, a few selection options will have confidence factors greater than 50. In this case, the system will recommend the user to opt for the one with the largest value of confidence factor. However, since the questions asked by the system are dynamic and the descriptions specified by the user are varying, it is difficult to ensure that a solution exists in every case. If no rule can be matched with the user specifications or no higher confidence factor can be obtained, the selection procedure will not be triggered. The confidence factor of each attribute’s value is recorded and maintained as part of the state of the attribute in the session context. The complete state of an attribute consists of its actual 6 value as well as its confidence factor. Confidence factors determined as a result of the firing of a rule are based on a product space calculation. The confidence factor of an antecedent preceded by NOT is 100 minus the confidence of the non-negated antecedent. When an antecedent of a rule is evaluated, the confidence resulting from the evaluation of the antecedent is first determined by comparing the minimum confidence of all antecedents joined by AND with the maximum confidence of all antecedents joined by OR. A single confidence value of the antecedents is determined to be the lower of the above two confidence factors. The conclusion of the rule is given a final confidence that is calculated from the single confidence factor of the antecedents and the confidence specified in the knowledge base for the conclusion. The product of this multiplication is normalized by dividing with 100, and the result is then assigned as the final confidence in the session context. Thus, the inference engine is designed as a set of codes to search and match for a way, which can lead to increasing confidence factors of parameter selection. The user is also allowed to execute some existing models employing the default parameter values. Of course, the existing model may not be the most appropriate in this particular instance. If the actual physical data differs too much from the bathymetry of the existing model, a low confidence factor of the model selection will be acquired. If the user can provide the specifications in greater details, there is a higher possibility for the inference engine to match the parameter selection with higher confidence factor and the resulting suggestions become much trustworthy. The main purpose of the query process is thus to glean the purpose of the user and the physical conditions of the problem so as to increase the confidence of the model selection. Hence, if it is anticipated that any questions or answers will have very small contribution to the confidence of parameter selection, it may not enter into the interrogative dialogue interface. As mentioned earlier, each parameter is related to some other factors through the relation base. However, several physical conditions are usually involved in order to determine each of these related parameters. As a result, the system will need to ask an enormous amount of questions for the users to fill in prior to completely specify the particular case. The mechanism is designed such that the system will simulate the possible conditions and determine the possible largest change in the confidence of selection through matching with the rule base before the query process. Questions with higher potential to contribute to the confidence of selection will be asked with a higher priority. The concept of decision tree is employed as a pre-set route to drive forward the process of the inference engine. It is in stark contrast with the static knowledge tree mentioned earlier for the relation base and the selection base although it is also represented in tree form. The key differences are that decision tree is dynamic and the structure of the tree branches will change according to the specifications provided by the user. Here the set of parameters is expressed under the structure of a decision tree. The procedure of parameter selection then involves searching through the decision tree and filling in all the tree branches after querying the user to specify myriad conditions. Hence the size of the decision tree depends on the complexity of the problem. As the query process proceeds and specifications by the user to some conditions have been added, the inference engine searches the rule base and computes the new confidence factors for all branches of the decision tree. It is designed that the classification method will cover all the involving parameters as far as possible in numerical modelling of coastal water processes. According to the present classification method, it is inevitable that conflicts may exist between parameters when some parameters preclude the existence of another parameter. For example, if a 2-dimensional 7 model is selected, the vertical grid spacing will no longer be used. In order to eliminate these conflicts in the expert system, the rule base is designed carefully to filter the required parameters in the decision tree. The most improbable branches of selection, defined as those with confidence factor less than 20, are cut away from the decision tree. Thus the decision tree can also be considered as the guide in a tree skeleton format, directing the inference engine to query the user as well as to make decision based on the responses. This function is provided by the control object “TreeView” under the programming environment Visual Basic 6.0. It has many useful features of tree adding, tree cutting, tree searching, tree shape changing, and tree display on the user-interface. Two types of inferencing strategy, namely, forward chaining and backward chaining, are often used in an expert system to control the matching between the facts and the rules. In this system, forward chaining is employed to search from responses made to the query by the user and to modify the decision tree whilst backward chaining is used to find the requisite parameters and then to determine the query to the user. Figure 7 shows the cyclic relationship amongst the inference engine and the various knowledge bases. The inference engine first searches the decision tree to determine the parameter, which will have highest potential confidence to model selection. The relation base is then involved to find the problem specifications or physical conditions related to that particular parameter. The question base is then searched to prompt the query together with the possible answers to the user to glean the problem specifications. The rule base then enter the scene to match the parameter and the specification, to cut away the impossible tree branches of the decision tree, to compute the confidence factors for a variety of selection options, and to recommend the one with the highest confidence factor. The new decision tree as well as the newly selected parameter is recorded. Another cycle of inferencing process is then ready to run again, until the whole decision tree is filled up. Conclusions Through the successful development of this prototype system, it has been proved that the expert system technology can be integrated into the numerical modelling for simulating flow and water quality phenomenon in coastal waters in order to provide assistance on the selection of model and its pertinent parameters. The integration renders a more intelligent and user-friendly system in the problem domain. The use of the hybrid expert system shell, Visual Rule Studio, which runs together with Microsoft Visual Basic 6.0 is found to be very effective in producing the system under the popular Windows environment. It combines the advantages of both production rules and object-oriented programming paradigm. Of course, expertise knowledge on more extensive types of numerical models can be added to enrich further this prototype system. Nevertheless, the existing architecture of the expert system has shown that it is useful to achieve the above objective. Further research effort will be spent on the representation of the more difficult encoded dynamic knowledge, knowledge elicitation of expertise questionnaire through World Wide Web, verification and validation of effectiveness of artificial intelligence technology in aiding numerical modelling. Acknowledgements The work described in this paper was substantially supported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region (Project No. PolyU5084/97E). References: 8 1. Abbott, M.B. (1989), Review of recent developments in coastal modelling, in Falconer,R.A.,Goodwin,P.,and Matthew,R.G.S., Hydraulic and Environmental Modelling of Coastal, Estuarine and River Waters, Avebury Technical,Aldershot, pp.3-39 2. Abbott, M.B.,(1991), Hydroinformatics: information technology and the aquatic environment, Avebury Technical,Aldershot 3. Blanpain, O. and Chocat,B.,(1994), Introduction of expertise in a hydroinformatics system: Choice of hydraulic and hydrologic models, Hydroinformatics’94,Verwey, Minns, Babovic&Maksimovic(eds) 1994, Balkema, Rotterdam, ISBN 9054105127 4. Chau, K.W. and Yang, Wen-wu, (1992), A Knowledge-based Expert System for Unsteady Open Channel Flow, International Journal of Engineering Applications of Artificial Intelligence, Vol. 5, No. 5, pp. 425-430 5. Chau,K.W., and Yang Wen-wu,(1993), Development of an integrated expert system for fluvial hydrodynamics, Advances in Engineering Software, Vol. 17, pp. 65-72 6. Chau, K.W. and Yang, Wen-wu, (1994), Structuring and Evaluation of VP-Expert Based Knowledge Bases”, International Journal of Engineering Applications of Artificial Intelligence, Vol. 7, No. 4, pp. 447-454 7. Chau, K.W. and Zhang, X.N., (1995), An Expert System for Flow Routing in a River Network, Advances in Engineering Software, Vol. 22, No. 3, pp. 139-146 8. Cunge, J.,(1989), Review of recent developments in river modelling, in Falconer, R.A.,Goodwin, P.,and Matthew, R.G.S., Hydraulic and Environmental Modelling of Coastal, Estuarine and River Waters, Avebury Technical,Aldershot, pp. 393-404 9. Knight, B. and Petridis M.,(1992), Flowes : An Intelligent Computational Fluid Dynamics System, Engng Applic. Artif. Intell. Vol.5, No.1, pp. 51-58 10. Rule Machines Corporation, (1998), Visual Rule Studio – Developer’s Guide. 11. Uzel, A.R.,Edwards, R.J., and Button, B.L., (1988), A study into the feasibility of an intelligent knowledge based system (IKBS) in computational fliuid mechanics(CFM), Eng.Appli. of AI, Vol. 1,September 9 Table 1. The classification of parameters Type Dimension and Method Numerical Scheme Grid Formulation Boundary Conditions Initial Conditions Eddy Turbulence Miscellaneous Parameters Dimension Numerical method Co-ordinate system Numerical method in vertical direction Vertical co- ordinate system Error of scheme Stability Explicit or implicit Alternating direction algorithm Advection term Node number per element Shape of grid Horizontal grid spacing Vertical grid spacing Uniform or non-uniform grid Type of point setting Tidal constituents at open boundary Discharge at open boundary Water quality variables at open boundary Variation in open boundary Variation in close boundary High order variation in boundary Initial conditions of elevation Initial conditions of current Initial conditions of water quality Turbulence model Vertical eddy diffusion Horizontal eddy diffusion Vertical eddy viscosity Horizontal eddy viscosity Coliolis force Geometry Bathymetry Wind at surface Bottom drag coefficient Table 2. Classification of the user specifications Purpose of study Physical characteristics Previous experience of user Tidal propagation Spatial scale Grid formulation Water quality study Time scale Accuracy of scheme Salt water intrusion Water depth Effectiveness of scheme Storm surge Complexity of bathymetry Stability of scheme Hot water discharge Ratio of tidal elevation to depth Initial condition Sediment transport Longitude and latitude of location Interaction of boundary conditions if adjacent to each other Seasonal structure of current in coastal zone Existence of river Representation of the driving forces Water exchange between water bodies Vertical variation of density Boundary at surface and bottom Research on eddy turbulence Vertical variation of current Open boundary at river end and ocean end Waste water discharge Vertical mixing Turbulence model and vertical variation 10 User User interface Explanation facility Expert Inference mechanism External database External numerical models Knowledge base Context Knowledge acquisition facility Figure 1 Architecture of the expert system 11 Figure 2. Part of the relation tree of model parameters 12 Figure 3. Part of the selection tree of model parameters 13 Question 1 (any information about vertical density variation?) Answer 2 (large variation) Answer 4 (small variation) Answer 3 (moderate variation) Answer 1 (very large variation) Related factor 1 (vertical density variation) Question 2 (any information about vertical current variation?) Related factor 2 (vertical current variation) Related factor 3 (water depth) Question 3 (specify water depth ?) Parameter (dimension) Related factor 4 (geometry complexity) Related factor 5 (user’s concern in vertical structure) Question 5 (what is the user’s concern in the vertical structure?) Question 4 (any complicated geometry?) Answer 5 (negligible variation) Figure 4. An example of the inference direction from a parameter to questions and answers 14 Rule base Related factor 1 (vertical density variation) User’s answer (very large variation) Match 1 (if : vertical density variation =very large variation then: dimensions=2d horizontal, with confidence 70) Match 2 (if : vertical density variation = very large variation then: dimensions=1d , with confidence 10) Match 3 (if : vertical density variation = very large variation then: dimensions=3d layered, with confidence 40) Selection 1 (dimensions=1-d) confidence=10 Selection 2 (dimensions=2-d vertical) confidence=70 Selection 3 (dimensions=2-d horizontal) confidence=30 Selection 4 (dimensions=3-d layered) confidence=80 Selection 5 (dimensions=2d+1d ) confidence=50 Selection 6 (dimensions=3-d fully) confidence=85 Confidence calculation Selection item with highest confidence (dimensions=3-d fully) Figure 5. An example of the inference direction from the user’s specifications through the inference engine 15 Fuzzy Member (%) 0 20 40 60 80 100 10 50 100 150 200 250 300 350 400 450 500 A B C Depth (m) Figure 6. Fuzzy description of the water depth (A---very deep) (B---deep) (C---shallow) Formulate and modify decision tree Select parameter with highest priority order Find related factors from relation base Pick questions from question base User to enter conditions and specifications Match selection with high confidence in rule base Option to adopt default model in model base Query for user specification Inference engine Figure 7. The inter-relationships amongst the inference engine and various knowledge bases 16 Advances in Engineering Software, Vol. 32, No. 9, 2001, pp. 695-703 An example of expert system on numerical modelling system in coastal processes K.W. Chau and W. Chen Department of Civil & Structural Engineering, Hong Kong Polytechnic University, Hunghom, Kowloon, Hong Kong Abstract Keywords Introduction Knowledge acquisition Knowledge bases Inference engine Conclusions Acknowledgements Figure 1 Architecture of the expert system 
work_id5zxekuovexhaprgj7hcbt6tm	----	International Journal of Scientific Research in Computer Science, Engineering and Information Technology Copyright: © the author(s), publisher and licensee Technoscience Academy. This is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial License, which permits unrestricted non- commercial use, distribution, and reproduction in any medium, provided the original work is properly cited International Journal of Scientific Research in Computer Science, Engineering and Information Technology ISSN : 2456-3307 (www.ijsrcseit.com) doi : https://doi.org/10.32628/CSEIT206657 285 Developing an Expert System Application to Detect Childs’ Lung Disease Sulis Sandiwarno Department of Computer Science, Universitas Mercu Buana University, Indonesia sulis.sandiwarno@mercubuana.ac.id Article Info Volume 6, Issue 6 Page Number: 285-290 Publication Issue : November-December-2020 Article History Accepted : 24 Nov 2020 Published : 10 Dec 2020 ABSTRACT The development of information technology has supported many activities, especially in terms of health. Artificial Intelligence (AI) is the application of information technology that is currently developing well. Several previous studies have evaluated models from expert systems to diagnose lung disease in children using Naïve Bayes (NB) and Support Vector Machine (SVM). However, in conducting these evaluations they do not try to make an integrated application to facilitate evaluation. In this study we propose to build a system that integrates NB and SVM classifiers. Furthermore, in this study we used a sample of data from a clinic in Indonesia. The results of this study, we conclude that the existence of this system will make it easier to evaluate the lung disease experienced by children. Keywords: Artificial intelligence, NB, SVM, Lungs I. INTRODUCTION The application of information technology today is the focus of life that leads to the advancement of science education. The development of information technology is supported by the lives of everyone who wants to use this information technology to help solve existing problems [1–3]. Information technology development in the world of health has a very big role in everyday life in solving existing problems. Expert system is a computer program that contains knowledge from one or more human experts regarding a specific field [4]. The general form of an expert system is a program based on a set of rules that analyzes information (usually provided by the user of a system) about a specific class of problems as well as a mathematical analysis of the problem. In the standard system, there are two commonly used models, namely forward chaining rules and backward chaining rules [4]. Forward chaining is forward tracking that starts from a set of facts by looking for rules that match the existing assumptions / hypotheses to conclusions. Meanwhile, backward chaining is backward tracking that starts reasoning from the conclusion (goal), by looking for a set of hypotheses to the facts that support a set of these hypotheses. http://ijsrcseit.com/ http://ijsrcseit.com/ http://ijsrcseit.com/ http://ijsrcseit.com/ https://search.crossref.org/?q=10.32628/CSEIT206657&from_ui=yes https://search.crossref.org/?q=10.32628/CSEIT206657&from_ui=yes Volume 6, Issue 6, November-December-2020 | http://ijsrcseit.com Sulis Sandiwarno et al Int J Sci Res CSE & IT, November-December-2020; 6 (6) : 285-290 286 Previous research has conducted an evaluation of expert systems in the field of health, especially for the lungs in children by applying models from machine learning such as Naïve Bayes (NB) and Support Vector Machine (SVM) [5, 6]. Although previous research has implemented various types of classifiers such as NB and SVM, they have not considered creating a system that can make it easier to conduct research evaluations simultaneously. Therefore, in this study we propose to design a system that can perform simultaneous evaluation by utilizing NB and SVM classifiers in the detection of lung disease in children. II. RELATED WORK In this chapter we divide references based on our research topic, namely NB and SVM classifiers in detecting Childs’ lung disease. 2.1 Naïve Bayes (NB) to Detect Childs’ Lung Naïve Bayes (NB) is of the one most elegant machine learning technique that is practically used. NB is an efficiency, effectiveness, and iterability algorithm to classify the data [7–9]. Previous studies have used NB in evaluating lung symptoms in children [4] 2.2 Support Vector Machine (SVM) to Detect Childs’ Lung SVM is one of the successful methods to develop classification in supervised learning. In addition, Huajuan [10] illustrated SVM useful to apply data classification in data mining and machine learning, because SVM has successfully solved and predicted the problem in high dimensionality of data classification such as text categorization with excellent performances. Previous studies have used SVM in evaluating lung symptoms in children [5, 6]. III. PROPOSED MODEL The flow work of this research divided into six steps such as design of knowledge representation, prototyping system, validation expert, testing of program, results. 3.1 Design of knowledge representation a. Design of knowledge representation model that was employed on this system based on the production rule using pattern of IF – THEN. Each lungs’ symptom has determined the weight value (confidence factor) that was defined by the domain expert within range 0....1. This value represents the confidence value of each lungs’ symptom which causing particular diseases. b. The lungs’ disease diagnosis adopted a forward chaining inference. This system allows users to select the symptoms of the infected kids’ lung. The user could be selected the symptoms textual statement and sample image in the expert system. The system will process the choice of users and give the evaluation results of the infected kids’ lung. 3.2 System Prototyping The expert system, for diagnosing lung disease, consists of lung disease diagnosis, knowledge-based, inference engine and database. The expert system architecture can be shown in Figure 2. In research we adopted PHP program language with MySQL as database. Inference engine contains thought mechanism and system reasoning pattern used by an expert. This mechanism will analyze a symptom selected by the user for producing the conclusion of diagnosis result. This inference system used forward chaining method. http://www.ijsrcseit.com/ http://www.ijsrcseit.com/ Volume 6, Issue 6, November-December-2020 | http://ijsrcseit.com Sulis Sandiwarno et al Int J Sci Res CSE & IT, November-December-2020; 6 (6) : 285-290 287 Figure 1. The framework of Proposed Model 3.2.1 Naïve Bayes Classifier Given the test description of the document 𝑑 of an opinion represented by the vector < 𝑤1, 𝑤2, … . . , 𝑤𝑚 >, to classify the document d, MNB is defined as: 𝐶 𝑀𝑁𝐵(𝑑) = 𝑃(𝑐) ∏ 𝑃(𝑤𝑖 |𝑐) 𝑓𝑖𝑛 𝑖=1 (1) where, 𝑃(𝑐) is a prior probability that a document 𝑑 belongs to class 𝑐 , 𝑛 is a number of the features, 𝑃(𝑤𝑖 |𝑐) is the conditional probability that a word 𝑤𝑖 occurring in the class 𝑐 , 𝑤𝑖 is the word feature occurred in 𝑑, 𝑓𝑖 is the number of frequency count of a word 𝑤𝑖 in reporting 𝑑 , and 𝐶𝑀𝑁𝐵(𝑑) is the class label of 𝑑 predicted by the classifier [9]. 3.2.2 Support Vector Machine (SVM) SVM is a machine learning technique which is used in prediction, classification, and regression. The 𝑖𝑡ℎ opinion in SVM trained with all of the opinions in the 𝑖𝑡ℎ class with the positive labels, then all other opinions with negative and neutral labels. Given 𝑙 as training data (𝑥𝑖 , 𝑦𝑖 ), … , (𝑥𝑛 , 𝑦𝑛 ) , where 𝑥𝑖 ∈ 𝑅𝑙 and 𝑌𝑖 ∈ {1,2 … . , 𝑐) describe an opinion class of 𝑥𝑖. To classify the document 𝑥𝑖, SVM is defined as: 𝑚𝑖𝑛 𝑤𝑚 ∈ 𝐻, 𝑏 ∈ 𝑅 𝑘 , ξ ∈ 𝑅𝑙 𝑥 𝑘 1 2 ∑ 𝑤𝑚 𝑇 𝑤𝑚 + 𝑘 𝑚=1 𝐶 ∑ ∑ ξ𝑖,𝑡𝑡≠𝑦𝑖 𝑙 𝑖=1 (2) 𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 𝑤𝑦𝑖 𝑇 𝜑(𝑥𝑖 ) + 𝑏𝑦𝑖 ≥ 𝑤𝑦𝑖 𝑇 𝜑(𝑥𝑖 ) + 𝑏𝑡 + 2 − ξ𝑖,𝑡 (3) ξ𝑖,𝑡 ≥ 0 𝑖 = 1, … , 𝑙, 𝑡 ∈ {1, … , 𝑘}\𝑦𝑖 where, in training the opinions data 𝑋𝑖 is illustrated to highest dimensional space by the function 𝜑 and 𝐶 is presenting the penalty parameter. Minimizing 1 2 ∑ 𝑤𝑚 𝑇 𝑤𝑚 𝑘 𝑚=1 describes we shall like to maximize 2 ||𝑤𝑖|| the margins between three groups of opinions data. If the data training is not linear distinguishable, there is penalty term 𝐶 ∑ ∑ ξ𝑖,𝑡𝑡≠𝑦𝑖 𝑙 𝑖=1 that can be reduced the total number of training error. For the summary, the concept of SVM is finding for the balance between the rule term of 1 2 ∑ 𝑤𝑚 𝑇 𝑤𝑚 𝑘 𝑚=1 and training errors [11–13]. 3.3 Validation expert The expert system validation is the stage which the experts re-examine the design of a system that has been developed by the authors. Validation of the prototype expert system is done by two senior doctors of clinic at Indonesia. IV. RESULTS The results of the research we have done can be seen from the sub-chapters which will be explained in detail below. http://www.ijsrcseit.com/ Volume 6, Issue 6, November-December-2020 | http://ijsrcseit.com Sulis Sandiwarno et al Int J Sci Res CSE & IT, November-December-2020; 6 (6) : 285-290 288 4.1 Diagnosis Results Figure 2. The diagnosis results Figure 3. The NB diagnosis results 4.2 Classification Results This section will explain the results of the research that has been done in evaluating kids’ lung diseases. The results of data classification carried out by NB and SVM classifiers are then calculated using several techniques, such as: precision, recall, F1 and accuracy. TABLE I DATA CLASSIFICATION RESULTS FROM SVM From the results obtained from the two methods, it can be seen that NB is the best method of classifying the data in this study. The accuracy value obtained from NB is 80.32% while SVM is 79.5% with a difference of 8.2%. From the fold (#) value that has been done that the smallest result of NB is in the second iteration with a value of 77.49%, and the highest value is 82.82 in the 4th iteration. Meanwhile, the smallest value in SVM is in the 2nd iteration with a value of 78.77% and the highest value is 82.05% in the 4th iteration. The Recall value at NB is 76.66% and the value at SVM is 71.06%. with a difference in value of 5.6 Meanwhile, the value of fold (#) in NB displays the lowest result in the 7th iteration with a value of 75.49%, while in SVM is 66% in the second iteration. Fold (#) Accuracy (%) Recall (%) Precision (%) F1 (%) 1 80.05 66.41 71.90 69.05 2 77.49 66.00 72.79 69.23 3 78.97 71.43 72.41 71.92 4 82.82 72.03 79.23 75.46 5 78.72 66.67 77.04 71.48 6 78.21 67.36 71.85 69.53 7 80.77 79.26 69.48 74.05 8 78.72 66.91 71.54 69.14 9 77.95 79.03 62.03 69.50 10 81.28 75.52 73.97 74.74 Average 79.50 71.06 72.22 71.41 http://www.ijsrcseit.com/ Volume 6, Issue 6, November-December-2020 | http://ijsrcseit.com Sulis Sandiwarno et al Int J Sci Res CSE & IT, November-December-2020; 6 (6) : 285-290 289 TABLE 2 DATA CLASSIFICATION RESULTS FROM NB If seen in the 2nd iteration NB displays the results of 76.53%, then there is a difference of 10.53%, while in the 7th iteration SVM displays the results of 79.26% with a value of 3.77% difference from NB. In the precision section, NB displays the results of 81.96% and SVM 72.22% with a difference in value of 9.74%. Whereas in section F1, NB displays the data classification results of 74.7% and SVM displays the results of 71.41% with a difference in value of 3.29%. based on the results of data classification that has been done by both methods, the biggest difference in value occurs in precision. V. CONCLUSION In this paper, an empirical study was conducted to evaluate the application of expert system to detect the kids’ lung based on Naïve Bayes (NB) and Support Vector Machine (SVM). Based on Naïve Bayes is the best algorithm to classify the users’ opinions in this study, the accuracy reported is 80.32% and 79.50% of SVM classifier. To The findings of this study reveal that NB classifier outperformed than SVM classifier in evaluating the kids’ lung disease. VI. REFERENCES [1]. Sadikin M, Fanany MI, Basaruddin T (2016) A New Data Representation Based on Training Data Characteristics to Extract Drug Name Entity in Medical Text. Comput Intell Neurosci. https://doi.org/10.1155/2016/3483528 [2]. Sadikin M (2017) Mining relation extraction based on pattern learning approach. Indones J Electr Eng Comput Sci. https://doi.org/10.11591/ijeecs.v6.i1.pp50-57 [3]. Triana YS (2018) Monte Carlo Simulation for Modified Parametric of Sample Selection Models Through Fuzzy Approach. In: IOP Conference Series: Materials Science and Engineering [4]. Kurniawan R, Yanti N, Ahmad Nazri MZ, Zulvandri (2015) Expert systems for self-diagnosing of eye diseases using Naïve Bayes. In: Proceedings - 2014 International Conference on Advanced Informatics: Concept, Theory and Application, ICAICTA 2014 [5]. de Carvalho Filho AO, Silva AC, de Paiva AC, et al (2017) Lung-Nodule Classification Based on Computed Tomography Using Taxonomic Diversity Indexes and an SVM. J Signal Process Syst. https://doi.org/10.1007/s11265-016-1134-5 [6]. Naqi SM, Sharif M, Yasmin M (2018) Multistage segmentation model and SVM-ensemble for precise lung nodule detection. Int J Comput Assist Radiol Surg. https://doi.org/10.1007/s11548-018-1715-9 [7]. Wang S, Jiang L, Li C (2015) Adapting naive Bayes tree for text classification. Knowl Inf Syst 44:77–89. https://doi.org/10.1007/s10115-014-0746-y [8]. Balamurugan AA, Rajaram R, Pramala S, et al (2011) NB+: An improved Naïve Bayesian algorithm. Knowledge-Based Syst 24:563–569. https://doi.org/10.1016/j.knosys.2010.09.007 [9]. Jiang L, Zhang L, Yu L, Wang D (2019) Class- specific attribute weighted naive Bayes. Pattern Recognit 88:321–330. https://doi.org/10.1016/j.patcog.2018.11.032 [10]. Huang H, Wei X, Zhou Y (2018) Twin support vector machines: A survey. Neurocomputing 300:34–43. https://doi.org/10.1016/j.neucom.2018.01.093 [11]. Weston J, Watkins C (1999) Support Vector Machines for Multi-Class Pattern Recognition. Proc 7th Eur Symp Artif Neural Networks [12]. Hsu CW, Lin CJ (2002) A comparison of methods for multiclass support vector machines. IEEE Trans Neural Networks 13:415–425. https://doi.org/10.1109/72.991427 [13]. He X, Wang Z, Jin C, et al (2012) A simplified multi- class support vector machine with reduced dual Fold (#) Accuracy (%) Recall (%) Precision (%) F1 (%) 1 81.33 76.00 81.82 73.06 2 78.77 76.53 83.09 73.14 3 81.79 77.65 84.14 77.46 4 82.05 76.90 83.85 75.69 5 79.74 76.57 86.67 74.76 6 77.69 76.48 77.78 70.71 7 82.31 75.49 83.12 78.77 8 80.00 76.71 78.46 72.34 9 79.23 77.28 77.85 75.23 10 80.26 76.99 82.88 75.86 Average 80.32 76.66 81.96 74.70 http://www.ijsrcseit.com/ Volume 6, Issue 6, November-December-2020 | http://ijsrcseit.com Sulis Sandiwarno et al Int J Sci Res CSE & IT, November-December-2020; 6 (6) : 285-290 290 optimization. Pattern Recognit Lett 33:71–82. https://doi.org/10.1016/j.patrec.2011.09.035 Cite this article as : Sulis Sandiwarno, "Developing an Expert System Application to Detect Childs' Lung Disease", International Journal of Scientific Research in Computer Science, Engineering and Information Technology (IJSRCSEIT), ISSN : 2456-3307, Volume 6 Issue 6, pp. 285-290, November-December 2020. Available at doi : https://doi.org/10.32628/CSEIT206657 Journal URL : http://ijsrcseit.com/CSEIT206657 http://www.ijsrcseit.com/ https://search.crossref.org/?q=10.32628/CSEIT206657&from_ui=yes https://search.crossref.org/?q=10.32628/CSEIT206657&from_ui=yes https://search.crossref.org/?q=10.32628/CSEIT206657&from_ui=yes https://search.crossref.org/?q=10.32628/CSEIT206657&from_ui=yes http://ijsrcseit.com/CSEIT206657 http://ijsrcseit.com/CSEIT206657 
work_iefbp6no4fbardridq77a4czba	----	Expert systems: A cognitive science perspective Behavior Research Methods, Instruments, & Computers /989, 2/ (2), /95-204 SESSION VII TUTORIAL ON EXPERT SYSTEMS Doris Aaronson, Presider New York University Expert systems: A cognitive science perspective ROY LACHMAN University of Houston, Houston, Texas The theory and technology of knowledge-based systems are intrinsically interdisciplinary and are closely related to the formalisms of cognitive psychology. In this paper, strategies of incor- porating intelligence into a computer program are described along with a common architecture for expert systems, including choices of representation and inferential methods. The history of the field is traced from its origins in metamathematics and Newell and Simon's (1961) GENERAL PROBLEM SOLVER to the Stanford Heuristic Programming Project that produced DENDRAL and MYCIN (Buchanan & Shortliffe, 1984). MYCIN gave rise to EMYCIN and a shell technol- ogy that has radically reduced the development time and cost of expert systems. Methodology and concepts are illustrated by transactions with a shell developed for graduate education and a demonstration knowledge base for the diagnosis of senile dementia. Knowledge-based systems and conventional programs are compared with respect to formalisms employed, applications, program characteristics, procedures supplied by the development environment, consistency, certainty, flexibility, and programmer's viewpoint. The technology raises basic questions for cognitive psychology concerning knowledge and expertise. The purpose of this paper is to explore the properties of knowledge-based systems and to compare them to con- ventional software systems. Such an analytic comparison requires an account of the conceptual origins of expert systems, their development, basic architecture, and relationship to psychological theory. The reader is as- sumed to have some familiarity with programming in a high-level language and at least an elementary sense of basic data structures and other programming concepts. This appears to be the characteristic level of awareness of most research psychologists, whose major effort is naturally devoted to the factual content and methodology of their primary occupation-the research and teaching of psychology. My examination of expert systems will follow a top-down sequence of exposition. The relation- ship of expert systems to cognitive science will be dis- cussed first, followed by an examination of central issues, such as what it means to incorporate intelligence and knowledge into software systems, the architecture of in- telligent systems, and their control processes. The dis- cussion will focus on the MYCIN system, and on its Addresscorrespondence to the authorat the Department of Psychology, University of Houston, Houston, TX 77204-5341. predecessor and progeny, since those programs have produced major advances in theory and technology. The final sections contain examples of transactions with an ex- pert system used for demonstration and teaching, and an analysis and comparison of expert and conventional sys- tem functionality. The paper, in its entirety, is based on the premise that basic knowledge will be advanced and the effective use of expert systems in science will be facili- tated by a careful distinction between conventional computer-support systems and knowledge-based systems. EXPERT SYSTEMS AND COGNITIVE SCIENCE Expert systems are one type of knowledge-based sys- tem and as such unequivocally comprise a subfield of arti- ficial intelligence (AI), which itself is a division of com- puter science. However, the designers of expert systems rely on the factual content and methodologies of several autonomous disciplines. These are collectively called cog- nitive science, and include linguistics, cognitive psychol- ogy, and AI. These disciplines share in common the study of various kinds of operations on symbols and symbol sys- tems. Cognitive scientists tend to be interested in the na- ture and properties of intelligent systems, and seek, in 195 Copyright 1989 Psychonomic Society, Inc. 196 LACHMAN varying degrees, computational theories of the systems of interest to their discipline. The cognitive sciences differ, as do all disciplines, in their presuppositional structures; they have different conceptions of subject matter, the ap- propriateness of specific problems, the acceptability of solutions, and the concepts of proof. The different presup- positions concerning the nature of truth and the methods of proof (Lachman & Lachman, 1986) result in the selec- tive use of confirmation methodologies, such as the natural science model of experimentation, formal proofs, intu- itions of native speakers, and proof by demonstration of a running program. The differences are overshadowed, however, by the interdependence of the cognitive sciences in finding solutions for each discipline's sanctioned problems. In particular, several ofthe core problems of knowledge-based systems are indistinguishable from the central problems of cognitive psychology and may require the methodology of both disciplines for acceptable so- lutions. Finally, it should be emphasized that although conventional software can be explicated in the idiom of computer science, knowledge-based systems can only be understood in terms of the concepts and methods of several of the individual cognitive sciences. The standard archi- tectures of knowledge-based systems, discussed below, clearly show that dependence. INCORPORATING INTELLIGENCE IN A COMPUTER PROGRAM Traditional programs are rigidly organized procedures for performing a task or solving a problem interactively with the user. The procedures, developed in software- engineering environments, are coded into programs from specifications provided by domain experts, project direc- tors, or analysts, all of whom tend to have naive models of end-users. The experts that develop specifications for a program possess explicit knowledge of the problem area and base their software requirements on established problem-solving methods. Conventional programs are then constructed from those specifications. The program code represents algorithms, that is, step-by-step symbol- manipulating operations, including decision and branch points, that are infallible for obtaining some result or solv- ing a problem. In some problems, for which there are a limited number of potential solutions, the algorithm con- sists of an exhaustive search through all possibilities. However, many significant scientific and human problems cannot be solved by algorithms because the problem areas are ill defined or open-ended, or because an exhaustive search may lead to combinatorial explosion. In such a sit- uation, there are too many solution paths to search in a reasonable amount of time. AI-based expert systems, in contrast, rely on heuristic methods to limit the search space, or number of solution paths, for a problem. Heuristics are procedures based on the established facts of a field and the judgment, opinions, and intuitions of a domain expert. Expert systems ex- plicitly represent, in computer code, the knowledge that makes decisions possible. They can thus function with poorly specified problems just as the expert does; they are not confmed to the precisely specified problems that are the only type typically solved by conventional software. A major difference between expert systems and con- ventional programs is that the capabilities of conventional programs come largely from stored algorithms making up their procedures, whereas the intelligence in expert systems is derived primarily from the explicit knowledge stored in its knowledge base. The knowledge there may be coded as rules, frames, or semantic networks. However it is coded, the declarative knowledge representation is processed deductively by a family of algorithms variously called the inference engine or interpreter. Conventional programs, on the other hand, must be constructed in a thorough, precise, and exhaustively specified fashion or they will not run. They are rigid and absolutely intoler- ant of ambiguity; failure of the user to conform to a pro- gram's requirements explicitly, including hidden require- ments, leads a conventional program to abort or "hang up." Gaps in knowledge, therefore, are not tolerated. In contrast, knowledge-based software systems make no such demands, basically because intelligence is incorporated into the system in a segregated, declarative knowledge base. The knowledge, semantics, and intelligence of con- ventional programs are embedded in algorithms and dis- tributed throughout combinations of procedures. This makes them context-dependent, difficult to understand and to modify, and highly predictable. The intelligence in knowledge-based systems, on the other hand, is primar- ily in the declarative expression of rules, semantic nets, or frames. The knowledge is recorded as data, is context- independent, and is easy to understand and modify. Such systems are rather unpredictable, because part of their in- telligence is in the semantic routines that control the state of the system, and the consequences of that control can be very difficult to anticipate. The relative power of the two approaches for various kinds of scientific and practical problems is very differ- ent. Articles on psychology and computing, such as those dealing with computer-based decision aids, sometimes fail to distinguish between intelligent and conventional pro- grams. Indeed, at one level of analysis there is consider- able overlap between expert systems and standard pro- grams. Expert systems, depending on their design, may contain various structures and procedures that are also present in conventional programs, and the algorithms of standard programs do represent human knowledge. More- over, both types of systems run on coded formalisms that are Turing-complete (Minsky, 1967). A Turing-complete formalism coded into a procedure is one that can be ex- ecuted on a Turing machine, and hence on any instantia- tion of a Turing machine, including high-level computer languages. Thus, anything that can be coded in the for- malisms of expert systems can also be coded in conven- tional high-level computer languages, given enough programming skill, coding time, and patience (Haugeland, 1981). Nevertheless, software systems that contain raw expertise in a knowledge base are crucially different from those based solely on algorithmic methods. THE ARCmTECTURE OF EXPERT SYSTEMS A software system is one or more configurations of computer code (a particular program, a programming lan- guage, or a programming environment that consists of one or more languages and numerous utility programs). Var- ious representations of software systems differ in level of abstraction and amount of detail. Machine language is the most concrete level of software system and pseudo- code is the least. Pseudocode is a rough natural-language representation of the anticipated steps in a program. It tends to be very loose and tentative, and lacks detailed expressions of implementation. The level of description called the architecture is less abstract than pseudocode, and has greater detail. However, representation at the architecture level is more abstract and less detailed than any level of computer code. The architecture of a soft- ware system, therefore, represents the system at a level of abstraction intended to capture the functionality and principles of operation built into the system, but is un- cluttered by input/output and other details of computa- tional procedure. Architecture, presented in a block dia- gram, is a simplification and idealization of a system; the diagram is designed to show functional components and their interconnections, and to capture major properties of the system and render them comprehensible. The User Interface A common architecture for expert systems is presented in Figure 1. The natural-language interface represents software elements that control the system's output of queries and conclusions, and its use of input data from the user regarding problem states and parameter values. Typically, the screen display is the main output modality EXPERT SYSTEMS 197 and the keyboard is the input modality of choice. Design of the user interface is a very active area of psychologi- cal research (Carroll, 1987). Interface designs, with very few exceptions, are rendered in algorithmic programs. Effective heuristic design of the user interface, based on natural-language communication, requires cognitive science (linguistics, psycholinguistics, and perhaps even ethnomethodology). In expert systems, the user interface solicits facts from the user about the current problem, in the form of multiple-choice selections or open-ended questions with a limited set of acceptable responses. The output information presented to the user, although it often appears to have many features of generative natural language, is usually "canned." That is, the system de- veloper has previously constructed, in the vernacular of the knowledge domain, natural-language sentences that describe each possible problem solution and the reasons that a given solution was selected by the system. Although real progress has been made in computational linguistics, cognitive science and AI technology have a long way to go before a truly intelligent natural-language interface will be available (Winograd, 1982). The Knowledge Base and Knowledge Representation The components of the architecture in Figure 1 are ab- stractions. The knowledge base, for example, is a con- ceptual receptacle for real-world knowledge. Knowledge in the base is stored in a declarative form as data struc- tures that mayor may not contain procedural elements. During the construction of an expert system, knowledge, in the form of rules, frames, and semantic nets, can be entered in various ways, depending on the development environment. Knowledge is entered into the base either by coding in a high-level language such as LISP or by use of a shell, which is a particular type of expert-system software-development environment. There are many strate- gies for representing knowledge, and cognitive psychol- USER 1 NATURAL LANGUAGE INTERFACE J SUPPORT SOFTWARE PROGRAMS TO LOAD RULES, INITIATE QUIERIES. J INFERENCE ENGINE ALGORITHMS FOR INFERENCE AND CONTROL STRATEGIES e.g. search for hypotheses; recursively call rules: backward chaining: ... J KNOWLEDGE BASE o PRODUCTIONS (rules) o STRUCTURED OBJECTS (frames. semantic networks) o PROPOSITIONAL LOGIC DATABASE (STM) o INITIAL FACTS o INFERRED FACTS o SYSTEM STATUS Figure 1. A common architecture for knowledge-based systems. 198 LACHMAN ogists tend to be familiar with one or more of the ap- proaches, so we will not linger on the important topic of representation. Using Nilsson's (1980) classification, there are three major classes of representation: production sys- tems (also called rule-based systems), logic systems (e.g. , predicate calculus), and structured objects (e.g., seman- tic nets and frames). A fourth strategy is increasingly be- ing used; it consists of various hybrid combinations of representational formats. Each category of representation has subtypes, and the pure versions of the major types have been proven to be formally equivalent (Anderson, 1976; Minsky, 1967). Although the formal equivalence of representational systems is of considerable theoretical importance, it has limited consequences for the actual de- velopment of expert systems. During the historical develop- ment of knowledge-based systems, a strategy emerged that was quite successful in advancing AI science and technology. Declarative knowledge was conceptually and physically segregated, as much as possible, in the knowl- edge base, and procedural knowledge was identified with the inference engine. Knowledge representation in cur- rent expert systems tends to be declarative, but, as we have earlier noted, procedural knowledge sneaks in, in various guises (Buchanan & Shortliffe, 1984). Inference Engine and Database The inference engine represents a set of algorithms that produce inferences from declarative knowledge stored as rules (alias: production system), frames (alias: schema), semantic networks (alias: acyclic and directed graphs), or some combination of these representational formats. These algorithms control the "reasoning" process by combining the inferential schema implicit in rules, frames, and networks with characteristics of the current problem either obtained from the user or inferred from facts stored in the database. When an advisory session between ex- pert system and user is initiated, the inference engine de- termines what facts it needs to solve a problem, and either gets those facts or terminates the computer run. The al- gorithms in the inference engine also control the sequenc- ing and generation of inferences, add newly inferred facts to the database, and process confidence levels. Although the classic formalisms for generating inferences with productions and networks are Turing-complete and there- fore formally equivalent, the actual algorithms imple- mented differ from the classical form of a production sys- tem or graph. Modifications to the formalism are made for pragmatic design considerations. Inferential algorithms for each representational schema can be implemented in many different ways with different selections of features and, therefore, really represent distinguishable families of procedures associated with productions, frames, net- works, and propositional logic. Algorithms for making deductions from uncertain and incomplete information are often included in the infer- ence engine. Both knowledge-based and standard pro- grams work in a deterministic fashion when evaluating probabilities of events, but the user of a standard program usually supplies all the data, estimates missing data, or selects a rule for making that estimate. Expert systems, in contrast, are characteristically designed to handle miss- ing and uncertain data automatically. It is the forte of ex- pert systems to work with uncertain, incomplete, and fuzzy data. In productions (also known as conditionals or rule-based inferences), the simplest and most basic rule of inference is modus ponens. The schema is: If p then q; given p; infer q The conditionals are recorded in the knowledge base and the initial facts about the problem are obtained from the user and stored in the database. Algorithms of the infer- ence engine apply the facts as antecedents of the condi- tionals and derive new facts from the consequence which are then also stored in the database. The system searches through the database, queries the user, or does both in an effort to find new antecedents so that rules may be fired. Firing a rule means finding a pattern that matches p in the rule base so that q may be added to the database as a new pattern and therefore as a potential match to forthcoming occurrences of p. The process requires a starting point, a strategy for ordering the selection of rules, and a strategy for working either from facts to conclu- sions, from conclusions (or hypotheses) to supporting facts, or both ways. The former strategy is called "for- ward chaining" and the latter strategy "backward chain- ing"; both strategies originated in mathematical logic. The main menu of a shell that supports the develop- ment of simple expert systems is shown in Figure 2. Com- puter runs showing the output from the shell's algorithms that represent the forward- and backward-chaining strate- gies for an identical rule base called DEMENTIA.RB are presented in Figures 3 and 4, respectively. The computer transactions shown in the figures are for DEMENTIA rules, three of which are shown in Table 1. The forward- chaining algorithm, depending on its particular implemen- tation, elicits the initial problem parameters (i.e., the facts) from the user and proceeds to draw all possible conclu- sions, as is illustrated in Figure 3. It recursively calls the rules by searching the database for an antecedent and adding the consequences of fired rules to the database until all facts and rules are exhausted. The forward-chaining algorithm can be implemented in a conversational expert system or in a larger system with primarily conventional components. The forward-chaining strategy is followed when it is possible to assume that all of the required in- formation will be available at the start of a computer run. The backward-chaining algorithm can be designed to search the rule base recursively until it produces a set of terminal nodes or end-states such as the hypotheses in Figure 4. Backward-chaining and hybrid systems will tend to be conversational. Selection of a hypothesis is solicited from the user, and the necessary facts to prove or dis- card it are obtained interactively, as in Figure 4. The backward-chaining sequence establishes subgoals and treats them in a similar fashion as the original hypothe- EXPERT SYSTEMS 199 ARTIFICIAL INTELLIGENCE LAB (C1Copyright 1987 by Roy Lachman MAIN MENU Procedures Available From This Menu A Forward Chaining E.S. E Physical and Mental Tests B Backward Chaining E.S. F' Print I<ule Uase C Bays i .i n Expert. System G Menu #2 0 Enter New Rule Base H Return to DOS During Any Procedure: Press <Fl> for Main Menu, < H > Exit to DOS Figure 2. The main menu of an expert-system shell used for graduate training in the cognitive theory and technology of knowledge-based systems. sis. Subgoals are evaluated by testing the first rule and, if it fails to fire or to establish the subgoal, recursing on the remaining rules. Subgoals that are established as facts (i.e., are the consequence of a fired rule) are added to the database. The backward-chaining algorithm either es- tablishes the selected hypothesis, or if it cannot, it says so and terminates the computer run. Support Software Support software, in an expert-system development en- vironment, performs any of a number of utility and sup- port functions. These functions can deal with any aspect of the architecture. In a rule-based system, for example, procedures can be included that interactively construct rules, edit the rules, test them for syntax and internal con- sistency, or test the entire rule set for contradictions. Graphics capabilities, development programs, file loca- tion and reformatting, and the like are available for other components of the architecture. In fact, any procedures used in user-interface development environments or in general software engineering can be added to an expert- system shell. ORIGINS OF THE KNOWLEDGE-BASED SYSTEM ARCHITECTURE Although a number of intellectual achievements made the emergence of AI and cognitive psychology possible, the seminal work in metamathematics in the 1930s ulti- mately contributed to both. The joint origins of AI and cognitive psychology in mathematical logic is described from the cognitive perspective by Lachman and Lachman (1986), and from the perspective of AI by Newell and Simon (1976). A direct line of development can be traced from the publication of Godel's (1931/1965) incomplete- ness theorem and Turing's (1936) computability thesis to the physical symbol system hypothesis (Newell, 1980; Newell & Simon, 1976). Although the physical symbol system hypothesis was not explicitly expressed until the 1970s, it is implicit in the early landmark work of Newell and Simon described below. The production system for- malism, so important to contemporary psychological theory and AI, also originated in the metamathematics of the 1930s (Post, 1936). In the 1950s, Newell and Simon originated a research program, in the Lakatosian sense (Lakatos, 1970), that was to become a cornerstone of psychology's information- processing paradigm. The program had several goals. One was to understand and develop computer programs that could deal with ultracomplicated problems such as chess or theorem proving in mathematical logic. Another goal was the scientific explication of general problem-solving and decision-making methods that involved limited ration- ality, "satisficing," and selective search. The computer programs they developed in that connection, LOGIC THEORIST (Newell & Simon, 1956) and GENERAL PROBLEM SOLVER (GPS; Newell & Simon, 1961), are illustrative of both the main research emphasis and the kinds of theory that emerged from their laboratory. Their research program aimed at the elucidation of domain- independent general problem solving and the development of theories of human thought. Their psychological theory was instantiated in a series of running programs whose performance generally was comparable to that of human problem solvers (Newell & Simon, 1972). The AI objec- tive of the program was to achieve general intelligence and domain-independent problem solving in a heuristic machine. Students of Newell and Simon wrote theses and dissertations, and conducted postdoctoral research using, and in some cases extending, the approach enunciated in GPS. While preparing his dissertation under the tutelage of Simon, Feigenbaum worked within the prevailing paradigm and developed a simulation of verbal learning called EPAM (Feigenbaum, 1961). However, attempts to develop programs in that tradition as serious technol- ogy ran into insurmountable problems. 200 LACHMAN eZ-16-1988 ze:Zl:17 RULE BASE=DEMENTIA.RB FORWARD CHAINING INFERENCE ENGINE To control the screen output Press <C>, otherwise <ANY KEY>. Enter facts now - Enter * To stop. Follow each entry with <RETURN>. FACT? NOT SELF-CARE FACT? POOR-RECENT-MEMORY FACT? NOT COUNT FACT? MAINTAIN-CONVERSATION FACT? * STM (data base) contains: NOT SELF-CARE POOR-RECENT-MEMORY NOT COUNT MAINTAIN-CONVERSATION The system is trying to find a fact in STM and to match it to an applicable rule in the rule base. SYSTEM USING RULE 17 Rule 17 Deduces: SEVERE-DECLINE STM (data base) now contains: NOT SELF-CARE POOR-RECENT-MEMORY NOT COUNT MAINTAIN-CONVERSATION SEVERE-DECLINE SYSTEM USING RULE 3 Rule 3 Deduces: EARLY-SENILE-DEMENTIA-ALZHEIMER-TYPE-(SDAT) STM (data base) now contains: NOT SELF-CARE POOR-RECENT-MEMORY NOT COUNT MAINTAIN-CONVERSATION SEVERE-DECLINE EARLY-SENILE-DEMENTIA-ALZHEIMER-TYPE-(SDAT) NO MORE APPLICABLE RULES FINAL RESULT: EARLY-SENILE-DEMENTIA-ALZHEIMER-TYPE-(SDAT) eZ-16-1988 ze:zz:ee NORMAL TERMINATION Figure 3. Computer transactions for forward chaining on the DEMENTIA rule base. All user entries are made to the prompt "FACT?". The domain of expertise used to illnstrate these principles and methods is the differential diagnosis of senile dementia. Feigenbaum, in collaboration with the geneticist Leder- berg, first attempted to develop a comprehensive and domain-independent aid to scientific thinking. Their ef- fort was not fruitful until they focused the problem-solving heuristics of the project on a circumscribed domain, and abandoned their effort to construct a general problem- solving program. The resulting system of programs, called DENDRAL, was developed by Feigenbaum, Lederberg, Dejerassi, and others. DENDRAL is a forward-chaining system that describes the molecular structure of unknown organic compounds from mass spectrometer and nuclear magnetic response data. It contains productions for the data-driven components and a procedural representation for the molecular structure generator (Lindsay, Buchanan, Feigenbaum, & Lederberg, 1980). DENDRAL was a watershed in AI; it represented a shift from power-based to knowledge-based programming. The program became the forerunner of many subsequent expert-system projects, in particular MYCIN. Success in the construction of knowledge-based programs thus was achieved only when the AI design objectives were changed to the representa- tion of narrow, limited, and focused domains of exper- tise (Feigenbaum, Buchanan, & Lederberg, 1971). The major change of focus from general, domain- independent to domain-dependent problem solving, along with other lessons learned from DENDRAL, were incor- porated into the MYCIN family of programs of the Stanford Heuristic Programming Project (Buchanan & Shortliffe, 1984). Newell (1984) described MYCIN as the expert system "that launched the field." The develop- ment of MYCIN and related programs (see Figure 5) made explicit many of the issues that arise in the construc- tion of knowledge-based systems. Although considerable innovation in AI technique has been achieved and land- mark systems have been developed, numerous basic and applied problems in cognitive science that are related to the basic properties of knowledge and to characteristics of knowledge-based systems remain unsolved. The MYCIN projects, including their domains, epistemolog- ical properties, and development, provide what is likely the most informative perspective on expert systems. The continuing retrospective study of MYCIN (Buchanan & Shortliffe, 1984; Clancey, 1985) explains the basis of the rise to prominence of knowledge-based systems, the fac- tors that make them successful, and the scientific foun- dations upon which they are built. 0 2 -1 7 -1 9 8 8 1 4 :4 0 :0 8 V E R = S 14 R U L E -B A S E = D E M E N T 1 A .R B R U L E 17 D e d u c e s S E V E R E -D E C L IN E T ry T o e s ta b li s h o n e o f th e fo ll o w in g h y p o th e s e s fr o m th e r u le b a s e : 1 = N O R M A L -A G IN G 2 = B E N IG N -S E N E S C E N T -C O G N IT IV E -D E F IC IT S 3 = E A R L Y -S E N IL E -D E M E N T IA -A L Z H E IM E R -T Y P E -( S D A T ) 4 = M ID D L E -S E N IL E -D E M E N T IA -A L Z H E IM E R -T Y P E 5 = L A T E -S E N IL E -D E M E N T IA -A L Z H E IM E R -T Y P E 6 = M U L T I- IN F A R C T -D E M E N T IA 7 = P O T E N T IA L L Y -R E V E R S IB L E -D E M E N T IA E n te r h y p o th e s is N o . 3 T h e S y s te m is tr y in g to p ro v e g o a l E A R L Y -S E N IL E -D E M E N T IA - A L Z H E IM E R -T Y P E -( S D A T ) T h e S y s te m is tr y in g to p ro v e g o a l S E V E R E -D E C L IN E T h e S y s te m is tr y in g to p ro v e g o a l P O O R -R E C E N T -M E M O R Y A ll r e le v a n t r u le s h a v e f ir e d , c a n 't p ro v e P O O R -R E C E N T -M E M O R Y , w h ic h is n o t in S T M -D B n o r th e c o n s e q u e n t o f a R u le . S h o r t- te r m m e m o ry (D A T A B A S E ) no w c o n ta in s : P O O R -R E C E N T -M E M O R Y N O T C O U N T M A IN T A IN -C O N V E R S A T IO N S E V E R E -D E C L IN E T h e S y s te m is tr y in g to p ro v e g o a l N O T S E L F -C A R E A ll r e le v a n t r u le s h a v e f ir e d , c a n 't p ro v e N O T S E L F -C A R E , w h ic h is n o t in S T M -D B n o r th e c o n s e q u e n t o f a R u le . IS T H IS T R U E : N O T S E L F -C A R E Y = Y es N = N o D = D o n 't k n o w (N o t y e t o p e r a ti o n a l) W = W hy (S h o w c o n te n ts o f S T M d a ta b a s e ) R E S P O N S E : N U N A B L E T O P R O V E H Y P O T H E S IS : E A R L Y -S E N IL E -D E M E N T IA -A L Z H E IM E R -T Y P E - (S D A T ) IS T H IS T R U E : P O O R -R E C E N T -M E M O R Y Y = Y es N = N o D = D o n 't k n o w (N o t y e t o p e r a ti o n a l) W =W hy (S h o w c o n te n ts o f S T M d a ta b a s e ) R E S P O N S E : Y 0 2 -1 7 -1 9 8 8 1 4 :4 1 :0 4 N O R M A L T E R M IN A T IO N o o a T h e S y s te m is tr y in g to p ro v e g o a l N O T C O U N T A ll r e le v a n t r u le s h a v e f ir e d , c a n 't p ro v e N O T C O U N T , w h ic h is n o t in S T M -D B n o r th e c o n s e q u e n t o f a R u le . IS T H IS T R U E : N O T C O U N T Y = Y es N = N o D = D o n 't k n o w (N o t y e t o p e r a ti o n a l) W =W hy (S h o w c o n te n ts o f ST M d a ta b a s e ) R E S P O N S E : Y T h e S y s te m is tr y in g to p ro v e g o a l N O T S E L F -C A R E A ll r e le v a n t r u le s h a v e f ir e d , c a n 't p ro v e N O T S E L F -C A R E , w h ic h is n o t in S T M -D B n o r th e c o n s e q u e n t o f a R u le . IS T H IS T R U E : N O T S E L F -C A R E Y = Y es N = N o D = D o n 't k n o w (N o t y e t o p e r a ti o n a l) W =W hy (S h o w c o n te n ts o f S T M d a ta b a s e ) R E S P O N S E : Y F IN A L R E S U L T : E A R L Y -S E N IL E -D E M E N T IA -A L Z H E IM E R -T Y P E -( S D A T ) R U L E 3 D e d u c e s E A R L Y -S E N IL E -D E M E N T IA -A L Z H E IM E R -T Y P E -( S D A T ) S h o r t- te r m m e m o ry (D A T A B A S E ) no w c o n ta in s : P O O R -R E C E N T -M E M O R Y N O T C O U N T M A IN T A IN -C O N V E R S A T IO N S E V E R E -D E C L IN E N O T S E L F -C A R E E A R L Y -S E N IL E -D E M E N T IA -A L Z H E IM E R -T Y P E -( S D A T ) T h e S y s te m is tr y in g to p ro v e g o a l M A IN T A IN -C O N V E R S A T IO N A ll r e le v a n t r u le s h a v e f ir e d , c a n 't p ro v e M A IN T A IN -C O N V E R S A T IO N , w h ic h is n o t in S T M -D B n o r th e c o n s e q u e n t o f a R u le . IS T H IS T R U E : M A IN T A IN -C O N V E R S A T IO N Y = Y es N = N o D = D o n 't k n o w (N o t y e t o p e r a ti o n a l) W =W hy (S h o w c o n te n ts o f ST M d a ta b a s e ) 0 2 -1 7 -1 9 8 8 1 5 :0 4 :5 7 N O R M A L T E R M IN A T IO N m >< -e tT l ;; Q - l en -< en -l tT l ~ en R E S P O N S E : Y F ig u re 4. C om p u te r tr an sa ct lo n s fo r b ac k w ar d ch ai n in g on th e D E M E N T IA ru le b as e. A ll u se r en tr ie s ar e m ad e to th e p ro m p t " R E S P O N S E :" . E n tr ie s ar e li m it ed to th e n u m b er s 1 to 7 in d ic at in g th e ch oi ce o f h yp ot h es is , an d to th e le tt er s l< \'" fo r Y es , "N " fo r N o, an d " W " fo r W h y. t v o 202 LACHMAN DENDRAL (Organic Chemistry) DART (Computer RepairJ I Shell Technology I Knowledge Acquisition TEIJSIAS CENTAUR J. GRAVIDA WHEEZE CLOT (Pulmonary Disease) (Bleeding) MYCIN (infectious Disease) J. NEOMYCIN ONCOCIN (Cancer Chemotherapy) I ICAI: Explanation J~ I Question 'T'"' BAOBAB 1960s I 1970s T Figure S. Program developed in the Stanford Heuristic Programming Project (Buchanan & ShortlifJe, 1984). Rule 3 If Severe-Decline And Not Self-care Then Early-senile-dementia-Alzheimer-type (SDAT) Note-DEMENTIA.RB was constructed from expert opinions published in Pearce (1984), Shamoian (1984), and other sources. The rule base is used only for training in the cognitive theory and technology of rule- based systems. Table 1 Three Rules from the DEMENTIA Rule Base Diagnosing Levels of Alzheimer's Disease, Multi-Infarct Dementia, and Depressive Dementia fore developed to handle probabilistic components of rules and to combine separate elements of evidence for diag- nostic hypotheses (Buchanan & Shortliffe, 1975). The MYCIN system soon became the model for inves- tigating a number of fundamental problems in AI and for developing expert system tools. The family of AI research programs was part of the Stanford Heuristic Programming Project shown in Figure 5 (Buchanan & Shortliffe, 1984). Research was conducted and programs were developed for expert system programming environments, on-line knowledge acquisition, tutorial packages, and consulta- tion programs for a variety of domains, some of which are shown in Figure 5. The specific programs developed included inference, natural-language question answering Sudden-onset-cog-Decline Fluctuating-course Impaired-consciousness Vascular-disease Multi-infarct-dementia Poor-recent-memory Not count Maintain-Conversation Severe-Decline If And And And Then If And And Then Rule 6 Rule 32 The initial objective of the MYCIN project was the de- velopment of a computer-based consultation system for the diagnosis of infectious disease and for giving advice in the choice of antimicrobial therapy. In the case of acute infections, immediate treatment is necessary even before the results of all laboratory tests and cultures are avail- able. In addition, antimicrobial therapy is subject to con- siderable human error. Consequently, development of the MYCIN system had practical significance, in addition to the scientific importance of testing AI constructs. A modified production system approach with a backward- chaining strategy was selected for the program. In other words, the system works backward from the goal of a therapeutic regimen and tests various rule-based diagnoses against facts supplied by the user interactively. The suc- cess of a rule-based approach in the DENDRAL programs was one reason for selecting a production-system representation. Still more important was the discovery by the developers of MYCIN that rules, and lines of reason- ing based on chained rules, are easier to understand, cri- tique, and modify than are alternative representations. These characteristics made fast prototyping possible with rule-based systems, an important property of production- system-based representation. The separation of produc- tion rules from the inference algorithms made it possible to develop an explanation module. The advice offered in a consultation was explained by the system, at least in part, by tracing the chains of inferences, displaying a descrip- tion of the rules invoked, and indicating why the system requested certain pieces of information from the user. The developers discovered, early in the project, that the productions of MYCIN differed in a significant way from those of DENDRAL, and that many of the inferences produced in MYCIN were uncertain. A system was there- (Bonnet, 1980), intelligent tutoring (Clancey, 1986), computer-based knowledge acquisition (Davis, 1979), and the expert-system shell EMYCIN (van Melle, Shortliffe, & Buchanan, 1984). EMYCIN is of practical importance because it produced a technology that has beenduplicated andextended in numerous commercial packages. The shell was created by removing the medical productions from the rule base of MYCIN and adding other procedures to the rule base. This made it possible to formulate rules from a variety of domains, including medical, engineering, and commercial, and to build a working system at a consider- able reduction in time and cost. The shell developed in my laboratory and its output for DEMENTIA.RB shown in Figures 2, 3, and 4 were patterned on EMYCIN. Many additional scientific lessons were learned from MYCINand its progeny, some of whichare validated by the demonstration of the expert-levelperformance of the system. The MYCIN project demonstrated that signifi- cant modifications to a program can be accomplished, without any reprogramming, by changing or adding declarative statements. From the origin of AI as a dis- cipline, the capacity to improve performance without reprogramming has been viewed as a significant dimen- sion of machine intelligence (McCarthy, 1958). The abil- ity of the system to use the same knowledge in more than one way (e.g., for advising and for explaining) is also symptomatic of intelligence. An important lesson for cog- nitive psychologists is the repeated demonstration, through progeny of MYCIN, that a few hundred rules may suffice to represent the deep knowledge of various significant EXPERT SYSTEMS 203 domains. Perhaps the final lesson of the projectis the dis- covery of the limits of rule-based systems, of the kinds of problems for which theydo notwork well or at all (e.g., continuous monitoring tasks, such as monitoring of life support systems in intensive care medical facilities). COMPARING KNOWLEDGE-BASED SYSTEMS WITH CONVENTIONAL SOFfWARE The distinction between conventional and knowledge- based systems is both real and important. However, this distinction cannot be appreciated simply by reference to the basic formalisms that support either of the systems. There are several reasons that formal comparisons are problematic. First, experimental systems, such as MYCIN, explore variations in many aspects of a formalism, as well as in combinations of formalisms; this makes it unclear exactly what is being compared. Second, the previous description of procedural languages as being Turing- complete means that anything that can be computed with a production system can also be computed with a high- level procedural language (Haugeland, 1981). Thus, at the formal level of description, there is no distinction to be made between any of the Turing-complete computa- tional systems. In physics and elsewhere, vastly differ- ent ontological domains can be represented by an identical mathematical formalism (Lachman, 1960). Consequently, although categories suchas knowledge-based systems and conventional programs are identical at one level of representation, theyare vastly different at another. There- Table 2 Programming Environments Standard Programming System Knowledge-based System Applications Operates in well-defined Operates in poorly defined domains with well-defined domains with poorly structured problems problems Flexibility Literal, inflexible Certainty Algorithmic, guaranteed to be correct Consistency Requires absolute consistency Environment supplies Iteration, branching, subroutine calls, etc. Program A description of a set of calculations or formal operations Control flow Decision of process currently executing Viewpoint Calculations performed, procedures implemented Programmer's job Describe algorithms needed to solve problem Flexible Not guaranteed correct, but can reach performance level better than human expert Deals with some inconsistency Forward and backward reasoning, traces its decisions, recursive calling of rules, etc. A description of relationships between variables, objects, or a system and its components Select-execute loop, unity of data and control Knowledge or rules applied, expanded view of a program Select knowledge needed to solve problem 204 LACHMAN fore, both categories must ultimately be described, not only in formal terms, but in terms of their presupposi- tions and in their style of approach. Newell (1984) has observed that many kinds of conventional software are designated as expert systems, which "mongrelizes" the field of AI-based knowledge systems. It is conceptually worthwhile to distinguish between the two categories, and major differences have been described throughout this paper. A summary of the features of each approach is presented in Table 2. Several of the differences are, as previously described. only a matter of degree. However, the categories of Table 2 represent different perspectives on the concepts employed, the external systems modeled, the program, and the programmer. Taken together, the features of knowledge-based systems may alter human in- tellectual capacity to such a degree that both science and society are fundamentally changed. REFERENCES ANDERSON, J. R. (1976). Language, memory, and thought. Hillsdale, NJ: Erlbaum. BoNNET, A. (1980). Analysede textesau moyend'unegrammaire seman- tique et de schemas. Application a La comprehension de resumes medicauxen langagenature/. These d'etat, Universite Paris VI, Paris, France. BUCHANAN, B. G., & SHORTLIFFE, E. H. (1975). A model of inexact reasoning in medicine. Mathematical Biosciences, 23, 351-379. BUCHANAN, B. G., & SHORTLIFFE, E. H. (Eds.) (1984). Rule-based expert systems. Reading, MA: Addison-Wesley. CARROLL, J. M. (Ed.) (1987). Interfacing thought: Cognitive aspects of human computer interaction. Cambridge, MA: MIT Press. CLANCEY, W. J. (1985). Heuristicclassification. Anificial Intelligence, 27, 1-67. CLANCEY, W. J. (1986). From GUIDON to NEOMYCIN and HERACLES in twenty short lessons: ONR final report 1979-1985. AI Magazine, 7, 40-60. DAVIS, R. (1979). Interactive transferof expertise. Anificiallntelligence, 12, 121-157. FEIGENBAUM, E. A. (1961). The simulation ofverballeaming behavior. Proceedingsofthe WesternJointComputerConjerence, 19,121-132. FEIGENBAUM, E. A., BUCHANAN, B. G., & LEDERBERG, J. (1971). On generality and problemsolving: A case studyinvolving the DENDRAL program. In B. Meltzer & D. Michie (Eds.), Machine intelligence 6 (pp. 165-190). New York: Elsevier. GoDEL, K. (1965). Theundecidable: Basicpaperson undecidable propo- sitions. unsolvable problems and computable functions (M. Davis, Ed. & Trans.). Hewlitt, NY: Raven Press. (Original work published 1931) HAUGELAND, J. (1981). Semantic engines: An introduction to mind de- sign. In J. Haugeland (Ed.), Minddesign(pp. 1-34). Cambridge, MA: MIT Press. LACHMAN, R. (1960). The model in theory construction. Psychologi- cal Review, 67, 113-129. LACHMAN, R., & LACHMAN, J. L. (1986). Informationprocessing psy- chology: Origins and extensions. In R. E. Ingram (Ed.), Information processing approaches to psychopathology and clinical psychology (pp. 23-49). New York: Academic Press. LAKATOS, I. (1970). Falsification and the methodology of science research programs. In I. Lakatos & A. Musgrave, Criticism and the growth ofknowledge (pp. 91-195). Cambridge, England: Cambridge University Press. LINDSAY, R. K., BUCHANAN, B. G., FEIGENBAUM, E. A., & LEDER- BERG, J. (1980). Applications of artificial intelligence for organic chemistry: The DENDRAL project. New York: McGraw-Hill. MCCARTHY, J. (1958). Programswith commonsense. In M. L. Minsky (Ed.), Proceedingsofthe symposiumon the mechanization ofthought processes (pp. 403-409). Cambridge, MA: MIT Press. MINSKY, M. (1967). Computation: Finiteand infinite machines. Engle- wood Cliffs, NJ: Prentice-Hall. NEWELL A. (1980). Physical symbol systems. Cognitive Science, 4, 135-183. NEWELL, A. (1984). Forward. In B. G. Buchanan & E. H. Shortliffe (Eds.), Rule-based expertsystems(pp. xi-xvi). Reading, MA: Addison- Wesley. NEWELL, A., & SIMON, H. A. (1956). The logic theory machine: A complex information processing system. IRE Transactions on Infor- mation Processing, IT-2, 61-79. NEWELL, A., & SIMON, H. A. (1961). GPS, a program that simulates human thought. In H. Billing (Ed.), Lernende Automaten (pp. 109- 124). Munich: R. Oldenberg. NEWELL, A., & SIMON, H. A. (1972). Human problem solving. Englewood Cliffs, NJ: Prentice-Hall. NEWELL, A., & SIMON, H. A. (1976). Computer science as empirical inquiry: Symbols and search. Communications of the ACM, 19, 113-126. NILSSON, N. J. (1980). Principles of artificial intelligence. Palo Alto, CA: Tioga Press. PEARCE, J. M. S. (1984). Dementia: A clinical approach. Oxford, England: Blackwell. POST, E. L. (1936). Finitecombinatory processes-Formulation I. Jour- nal of Symbolic Logic, 1, 103-105. SHAMOIAN, C. A. (Ed.) (1984). Dementia in the elderly. Washington, DC: American Psychiatric Press. TURING, A. M. (1936). On computable numbers, with an application to the Entscheidungsproblem. Proceedings ofthe London Mathemat- ical Society (Series 2), 42, 230-265. VAN MELLE, W., SHORTLIFFE, E. H., & BUCHANAN, B. G. (1984). EMYCIN: A knowledge engineer's tool for constructing rule-based expert systems. In B. G. Buchanan& E. H. Shortliffe (Eds.), Rule- based expert systems (pp. 314--328). Reading, MA: Addison-Wesley. WINOGRAD, T. (1982). Languageas a cognitiveprocess. Reading, MA: Addison-Wesley. 
work_ifb2hwodtjhtbo6tms3tkioipy	----	Journal oJ Glaciology, r ol, 42, Ao, 141, 1996 Avalanche forecasting-an expert systeDl. approach J GR G S CHW EI ZER A KD P AU L M, B, F OH:\" Eidgellossisches i ns lillll fill' Sdmee - llnd Lawinenforschung, CH-7260 /rJleissflllhjoclz /D avos, Swi lzerland ABSTRACT, Al ala nc h e fo recas ting fo r a gi,'e n region JS still a diffi c ult tas k im 'o h-ing g r ea t res ponsibility , An y tool s as istin g th e ex p e rt in th e d ec isio n-m a king process a r c we lcome, H oweve r , a n effi cient a nd success ful to ol sho uld mee t th e n eed s of th e forecas te r, vVith thi s in mind , (wo m o d e ls, )ve re d evelop e d usin g a co mm e rcia ll y al'ail a bl e soft wa re: CYBER TEK -COGENSY S I CIf , a jud g m e nt processor fo r indu cti ve d ec ision-m a kin g - a prin c ip a ll y d a ta -b ased ex p ert sys te m, Us in g wea th er, sn ow a nd snow- cove r d a ta as inpu t p a r a m eters, th e m o d els eva lu a te fo r a region th e d eg ree of al'a la nch e h aza rd , th e as p ec t a nd altitud e o f th e mos t d a n gero us slopes, Th e o utput res ult is b ased o n th e snow- cO\'e r sta bility , Th e new model s w er e d eve loped a nd have bee n tes ted in th e Da l'os r eg io n (Swi ss Alps) fo r seve ra l yea r s , T o ra te th e m o del s, th eir o utpu t is c omp a red to th e a p os teri ori I'e rifi ed h aza rd, Th e fir st model is p urel y d ata- based, Comp a red to o th e r sta ti sti cal m o del s, th e diffe r e n ces a rc: m o r e input inform a ti o n a bou t th e sn o w cove r fro m sn o w profil es and Rutsc hbl oc k tes ts, th e spec ifi c m e th od to searc h for simil a r situ a ti o ns, th e co n cise o utput res ult a nd th e kn ow led ge b ase th a t includ es th e I'e rifi ed d eg ree o f ava la nche h aza rd, Th e perfo rm a n ce is a bout 60 %, Th e seco nd , m o r e-refin ed m o d e l, is both d a ta - a nd rul e- based, It tries to mod el th e d ec ision-m a kin g process of a prag m a ti c expe rt a nd has a perform a n ce o f a bout 70 0/0 , w hi ch is co mp a ra bl e to th e acc uracy of th e publi c ,\'a rnin g , 1. INTRODUCTION AI'a la nc h e for ecas tin g, in our co ntex t , m ean s th e da il y assess m e nt o f th e a l'a la nc h e h aza rd for a g iven region, i, e, fo recas tin g a t th e meso -sca le (M eClun g a nd Sc hae rer, 1993 ) , Th e res ultin g ava la n c h e wa rnin gs a nd reco mm en- d a ti o ns [o r th e publi c sh o uld d esc rib e th e a l'a la nche situ a tio n , i, e, gi,'e info rm a ti o n a bout th e pl ace, th e tim e a nd th e pro bability o f r e lease for a sp ec ifi c type of a l'a la nch e (sla b or slufT, la r ge o r sma ll , we t o r dry), Th e mos t co nl' e ni ent way to ha ndl e thi s so rt o f inform a ti on is to summ a ri ze it as a d eg r ee of al'a lanch e h aza rd, Si nce 1985, in Switze rl a nd , th e d eg ree of h aza rd has bee n d efin ed in d esce ndin g ord e r b y th e release pro b a bili ty , th e a rea l ex te n t of th e in st a biliti es a nd th e size of th e a l'a la n ch es (F ohn , 1985 ) , Th e sca le is ge n e r a ll y based on th e sn ow-cOl'e r sta bilit y, I t co pies th e d evelopm ent o r stepping o f th e mos t typi ca l al'a la neh e situ a ti ons a nd hence is no t lin ea r. I 'h e int ensit y o f a n ava la nche situ a ti o n in c reases stro ng ly fr om one d eg r ee to a no th er , m ay be el 'e n ex p onenti a ll y , Co nsequ entl y, th e fr equ ency of th e d eg rees of haza rd d ec reases acco rdin g ly with in cr easin g d egr ee of haza rd (Fi g, I ), An y ex p ert sys tem should p ro fit from thi s co n ce pt th a t was a d o pt ed in 1993 b y th e w o rkin g gro up o f th e Eu ro p ea n al'a la nche- wa rnin g se n 'ices , In thi s stud y, seve n d eg r ees o f ava la nc h e haza rd a re u sed acco rdin g to t h e stru cture d efin ed in 1985; d eta il s o f th e Swi ss haza rd sca le hal'e a lso bee n giv e n b y Nl cC lun g a nd Sc hae rer ( 1993 ) , Sin ce dr y-sla b ava la n ch es represe nt th e m os t impo r- ta nt threa t fo r ski ers a nd bac k- co untr y tra l'cll ers, \\'e 3 18 degree of haza rd Fig , 1, R ela tive .frequeJlc} oJ tlt e verified degree oJ haz ard ill Ihe D avos region : len winter seasolls including 1512d are considered, Lighl columns ( lift) Io r the old Swiss sel'fll -degree scale, dark [olul11ns ( right) and valuesJor the Ilew European Jive-degree scale , foc used o n th e hazard of dr y-sla b ava la nc h es, During sprin g tim e, ,,'e t-snow al'a la nc h es a re pa rtl y co n sid ered: th e d ail y in cr ease of th e haza rd du e to wa rmin g durin g th e d ay is no t ta ken into acco unt , LaC ha pcll e ( 198 0 ) d esc rib ed th e tec hni q u e fo r assessin g th e ava la nche h aza rd: wea th er , sn o w a nd snow- cove r d a ta o bse rved d a il y a nd meas ured a t se l'er a l loca ti ons r eprese nta ti,'e for a g ive n area a re eva lu a ted by hum a n ex p e rts usin g th eir kn o wl edge a nd lo ng -term ex peri ence combin ed with indi vidu al intuiti o n, Sin ce th en th e procedure has not c h a nged mu ch, Th e core is still form ed b y th e cl assica l process of sy no psis suppl e- Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Sclzweizer al/d FiJllIl : .4 volallclte forecasti ng - an e.\/)erl s),slell1 a/Jproaclt T : verif ica tion : sup po rting t ools: st ati sti cal determini stic expert systems Fig , 2 , The clo:,sical cOll vellliol1al melhod sll/J/)Iel/lenled willt differenl sllpjJorting lools 10 forecasl tlte allolanc/ze /1Ci::ard on l/ze regiollal scale, me nt ed n owa d ays by diffe re nt so rt s o f s u p po rtin g too ls (Fig,2 ) , H owel'e r , th e d e m a nd s are stead il y in c reasin g : m ore fi'Cq u e n t a nd m ore d e t a i led in for m a ti o n is d es ir ed, T o uri sm is sti ll d eve lo pin g a nd , as in m a ny o th e r mou n ta in o u s reg io ns, th e A lps ha l'C b ecome o nc of th e fa l'o urit e p layg ro und s in Euro pe, S ki lO ur ing is wid e r ra ng in g a n d mo re p o pul a r th a n it was prel' io usl y , Neve rth eless, th e num b er 0 [' a\'a la ne h e I'ic tim s has n o t acco rdin g ly in c reased (Fi g, 3 ), whi c h m ay be du e to th e be ll er edu ca ti o n a nd aware n ess of th e ski e rs, pres um a bl y a lso d u e lO th e stab ili zin g e ITec t of m o r e fr eq uent skiin g a ft e r eac h s noll'fa ll per io d o n po pul a r slo p es a nd ho pefu ll y du e to b e tter \\'a rni ng, In th e fo recast pr ocesses nO\I'adays, a lo t 0 [' electro ni c too ls a rc i Jl\'o lved: ac q UI S111 0 1l , tra nsfe r an d represe n t- atio n o f t h e da ta, da ta b ase, snow -co l'e r si m ul atio n , 1111 free terrain ~ traffic o buildings ~ 60 r-------------------------------~ ~ '5 :§ 40 Q) .J:. g 20 o '5 > o o 65 70 75 80 85 90 95 year Fig , 3, Amlanche falalilies ill Ihe SwiH : II/)s 1964 6510 J99-1- 95, The falalilies are appoillled 10 Iltree calegories describing T('herl' Ihe {wa/al/che accidenls hajJjJelled: in blli/dillgs, 011 COlIll/IUllicalion lilles (illcllldillg cOlllrolled ski runs) and in Ihefree lermin ( back (0111111],), The 30)'ear average is aboul 28 falalities, indicaled b)' l/ie Ihill broken line , T he solid bold lille shows 0 5 ) 'ear IIlOl'illg average , Percentage Illlmbers give Ihe /)I'o/)orliol1 of falalities ill Ihe free terrain ( 10 ) )eal' average) showing deor~y Ihe increasing number I!/jalaii lies in Ihe free lerrain, whereas lite lolal IIlImber Irend is slalional)'. n u m erica l wea th e r fo recas t , d ec ision -ma kin g too ls a nd ex pe rt syste m s, a nd info r ma ti o n di stributi o n , EI'e n so, th e d ata t ha t a re m eas ured a re n o t th e mos t re leva n t o n es, \Yh a t is rea ll y n eed ed is th e s tre n g t h (co mpress ive, te nsile a nd shear ) of indi vidu a l sn ow laye rs, the so -ca ll ed 10 11'- e n t ro p y d a l a ( L aC h ape ll e, 1980 ) o r cl ass I d a ta (M cC lun g a nd Sc hae rer , 1993 ) , Th e a l'a il a b le d a ta a rc jus t mo re o r less app ro pri a te to p a r a m e te ri ze th e re!el'a nt p rocesses , T h e int erpre ta ti o n o f th ese d ata is t h e m os t d e li ca te tas k a nd hence m a n y su ppo rtin g too ls h ave bee n d el 'C lo ped fo r hum a n expe rts, Because th e 3\'a lanc he h aza rd ca nn o t (ye t?) be ca lcu la ted full y in a st ric t se nse ( by a lgo rithm s) , thi s is a fi e ld fo r hu ma n ex p e rts a nd co rres po ndin g ly for ex pe rt sys te m s, 2. PRESENT APPROACHES Th e a p p roac h d esc rib ed b y L aC h a pe ll e ( 1980 ), th e c lassica l m e th od , sti ll fo rm s th e b as is of th e d ec isio n- ma kin g proced ure of mos t ava la n c h e-fo recast sen' ices , U p till now, n o n e o f th e supp o rtin g tools have b ee n re li a b le e n o ug h to s ubs titute fo r th e hum a n expe rt a nd will proba bl y n e l'Cr bc, But m ay th ey beco m e a n obj ec til 'e p a rtn er fo r " di sc uss io n "? A ge n e ra l O\'crl'iew 0 [' d iffe re nt me th ods has b ee n g il'en by F b hn a nd o th e rs ( 1977 ) , Buse r a nd o th e rs ( 1985 ) a nd m o re rece n l ly by l\l cC lun g a nd Sc hae rer ( 1993 ) , Ln th e fo ll ow in g we foc us o n fo recas tin g mod els a nd too ls, Statistical approache s Th e mos t po pu la r sta ti sti ca l m e th od s a re t he di sc rimin a nt ana lys is a nd t h e nea res t n e ig h bo urs (:\I cC lung a nd Sc hae rer, 1993 ) , Already, in th c 1970s th e firs t s tudi es usin g d iscr imin a nt a n a lys is h ad b ecn pe rfo rm ed to find th e re le l'a nt pa ra mete rs fo r <1 I'a la nche fo recas tin g (e,g , P e d a, 1970; J udso n a nd Eri c kso n , 1973; A n nst ro ng a nd o th ers, 197+: Bo is a nd o th e rs, 1975; Sa l way, 1976 ), Snow a n d weat he r d a ta a rc us u a ll y use d toge th e r w it h obse n 'a ti ons of al'a la nche ac ti l' it y, ~ew sno ll' dep th , tem pe ra lUre a nd win d spee d , to m e nti o n so m e, h a l'e pro l'e d to be i m po rta n t. Th e resu lts ha I'C co n fi rmed th e ex pe ri e nce o f th e al'a la nc he expe n s, But , as n one of t he pa r ame ters used is d irec tl y re la ted to t he p rocess 0 [' a l'a la nc he fo r ma ti o n, it h as not bee n p oss ib le to el'a lu a te t h e al'a la nc he h aza rd, Th e d a ta uscd , th e usu a l o bse n 'ed a nd m eas ured pa ra me te rs, a r e a ll ind ex va lu es, (Th e d ata used a r e t hose th a t a re a l'a il ab le a nd no t th ose th at a rc m os t rcle l'a nt to th e process o f ava la nche fo rm a ti o n ,) Th ey a re in s tru c til'e to a n ex pe rt a nd may g ilT th e co r rec t hints to th e key p rocesses, su c h as se tt lem e nt. H Ovl'e\'e r, th e res u lts o f'th ese sta ti stica l studi es hal'e im p roved th e und ersta ndin g a nd h a l'e hclJJ ed lO s tru c ture know le dge a nd fin a ll y to d el'C lo p rule-b ased sys te m s, Additi o n a ll y, as th e sta tis ti ca l m od els n eed lo ng- te rm da ta, ma n y I'a lu a bl e o bse n' a ti o n s hal'e b ec n init ia ted, T h e acc umu la ted d a ta base m ay n olV be used to im p ro l 'e th e me m o r y o f' th e expe rt. Op era ti o n a l sys tems based o n th e sta ti sti ca l a pproac h, a nd usin g a lo n g- term d a ta b ase, h ave bee n d eve lo ped in sCI'C ra l co untri es an d a rc n o \\' w id e ly used (Bu se r a nd o th ers, 1987; N ava rre a nd o th e rs, 1987: M cC lun g, 1994; 3 19 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ournal of Glaciolog)I M cC lu ng and Tweedy, 1994; }. I crind o l, 1995 ) bo th for loca l and for reg iona l ava la nche fo recas tin g . Except fo r rh e sys te m d eve loped b y M cClun g a nd Tweedy (1994) , whi ch is a co mbin at ion of the tw o stat isti ca l m et hod s, all o f'th em use th e nea rest-neig h bour meth od. It is ge nera ll y assumed th a t simil ar snow a nd weather co ndition s should lead to simil a r ava la nche situ a tion s, i.e. that obser\'ed a\'a la nches of's imil ar past da ys sho uld be rep rese ntat i\'e ofth c prese nt- day situ a ti o n. A geo me tri ca l di stan ce in th e input p a r a meter space is used in sea rch i ng for simil a r si tu a - tion s. Th e Euel idi an di stance betwee n th e ac tual-d ay data a nd the surroundin g pas t-da y data is in so me mod els ca lculat ed direc tl y in th e input para m e ter space (Bu se r and ot h ers . 1987 ); in ot her mod els, the input d ata a rc first t ra nsform ed in the space of th e prin cipa l co mpon ents that w as determin ed b y th e sta ti sti ca l a nal ys is o f the d a ta base (lV[ crindol , 1995 ) and then th e M a ha la nobi s di sta nce is used (M cClun g and Tw eed y, 1994 ) . In so me models, the input para mete rs a re weig hted accOl-din g to the ge n eral ex p erien ce o f the expert fo recas ter (Bu ser a nd o th ers, 1987 ) and th e we ig hts may eve n vary accordin g to th e ge nera l wea ther type (Bolognesi , 1993b ) . The output is ge nerall y bas ed on th e obse rved avalan ches of' th e ten or 30 neares t n eig hb o urin g d ays . Thi s info rm a tion has to be e\'a lu ated by th e for ecas ter. Th e simplest way to summariz e thi s so rt of inform a ti on is to just se para te betll'een avalanche days a nd non -avalanche da)'s (Obled a nd Good , 1980 ). In add iti o n to th e li st of th e 30 nea rest neig hb ours , M cC lun g and Tw eed y ( 1994 ) predi cted th e p roba bility o f a\'a la nching by using di sc rimin a nt ana lysis. For th e fo recastin g of a \'a lan ches at K oo te nay Pass (B.C., Canada ) th is pro ba bility is co m- bin ed , usin g Baysian statisti cs , with the ex pert 's op ini on , m a d e a pri ori , to tak e into account a ddition allo\\'-entropy d a ta (e.g . th e sn ow-cO\'er situ ation ) (W e ir a nd }'[CC lun g, 1994 ) . In oth er m od els. a number is g iven as o utput acco rding to th e class ifi ca tion o f o bse rv ed avala nch es used in th e country (Gu yo marc' h a nd M e rindol , 1994 ) o r a n a va lanc he in dex is ca lcul a ted (N avarre a nd o th ers, 1987 ). Th ese types of o utput are difIi cult to r e late to the ac tual h aza rd in a give n reg ion. H ence, it is difIicul t to assess th e r ea l quality o f th ese fore cas t models. The y ce rtainl y improve th e refle c ti ons of un ex pe ri enc ed forecaste rs a nd m ay influ ence ex p eri enced forecasters but rarely may the y be ca lled a dec isi\'e help in dete rminin g the d eg ree of h aza rd in a given region. DeterIIlinis tic approaches Th e a im of the d e te rmini stic a pproach is to simul a te th e ava la nch e rel ease. \\' hereas th e m os t rel eva nt snow -cover pro cesses may be mod ell ed (Bru n a nd ot hers, 1989, 1992 ) and so m e a ttempts ha \'e a lso been m a d e to model th e ava lanc he fo rm a ti o n (G ubl er a nd Bader, 1989 ), it is still a lmos t imp ossib le to simulate t he range of the num erou s avala nche-fo rm a tion processes o n a m ountain slope - not to sp eak of a lI'ho le a rea. A poss ibl e way out is to use d iffer ent me thod s, fo r exa mpl e, to d eve lo p a n ex p ert sys tem w hi ch a n a lyzes the simulated snow-cove r st r a t- ig raph y (Giraud, 199 1). Fohn a nd H aec hl er ( 1978 ) d e\'clo ped a determ ini sti c- sta ti sti ca l m od el whi ch re lates t h e snow acc umul a ti o n by snowfa ll , wind a nd se ttl e m ent to th e a\'a la nche ac tivit y. Th e mod el is a ppropri a te to d esc ribe a\'a lanc he situ at ions in pe riods of heavy snowfa ll. 32 0 E x pert systeIIls Exp er t sys tem s· represe nt the id ea of simul at ing th e d ec isio n-m a kin g process of a n expe rt . Mos t o f them a r e symboli c co mputin g sys t e m s, i.e. usin g rul es w hi ch we r e formulat ed ex pli citl y by hum a n expe rt s, e.g . }' IEPR A (G ira u d , 199 1) a nd A V ALOG (Bol og n es i, 1993a ) . The Fr e n c h sys tem l\lEPRA a nal yzes the sn ow -cove r stra tig r a ph y; th e sno w profi les a re simul ated b y th e snow- cover mod el C RO CUS (Brun a nd others , 1989 ) in p a r a ll e l w ith m e t eo r o logica l data prO\' id ed by SAF R A J (Dura nd and o thers, 1993 ) , a m ode l fo r o ptim a l int e rpo latio n of m eteo ro logical d a ta. AV A - LOG , assessin g th e ava la n ch e hazard slope-b y-slope, is a n ass ist in g too l fo r th e e ffi cient art ifi cia l release by ex plosives in the res tric t ed a rea ofa ski reso rt. R ece ntl y, a hybrid exp er t sys te m h as bepn d eve lop ed usin g a neura l n e twork a nd rules ext racted from th e data base with n e ura l netwo rk techniqu es (Schweize r a nd o th e rs, 1994a, b ) . Bolog nesi ( 1993b ) h as d e \' e loped ano th er h y brid sys tem ca ll ed NX-LOG b y co m bi ning the statis ti cal mod el NXD (Bu ser and ot h e r s, 1987 ) and th e rul e - based syste m AVALOG. A furth er rece nt d evelopm ent is a rul e-b ased ex pe rt syste m to interpre t data fr o m snoll' profi les with respect to snow stabilit y (M cClun g , 1995 ). 3. A NEW APPROACH WITH THE CYBERTEK. COGENSYS™ JUDGMENT PROCESSOR In 1989 , we bega n a n ew ap proac h with th e id ea of buildin g a sys tem for region a l a\'a la nche fo r ecas tin g simil a r to th e sta ti stic a l ones bu t wi th a d i fferen t method of sea rchin g for simi lar situations and with optimi zed inpu t a nd output pa ra me ters, ca ll ed DAVOS. \V e tri ed to inclu de so m e of the rel eva nt ph ys ica l processes, i.e. elab o ra ted in put parameters a nd to give as a res ult direc tly what th e ava la nc he forecaster wo uld like to have: th e degree of haza rd (Schweizer and o th ers, unpublished ) . In 199 1, we worked o u t a co mplet ely new app roac h , more process -oriented a nd part ly r ul e-based , wh ich tri ed to mod e l th e reaso nin g of th e ava lanc h e for ecas ter, ca ll ed l\ IOD UL. Both models a re ba sed on softw a re for indu ctiv e d ec ision -m a kin g : CYBE R TEK-COGENSYST~ I jud g - me nt pro cesso r (ve rsion 19 ) , which is prim aril y used in th e fin a n ce a nd in suran ce wor ld . The IIlethod: the judgIIlent proc e ssor The C YB E RTEK- COGENSYSD I jud g m ent processor is a co mm e rc ia ll y a\'a ilabl e software fo r indu cti\'e a uto - matic d ec ision-m a kin g . Sin ce we had no access to th e so urce cod e, we did not exact ly know what the sys tem does A traditi o na l ex pert syste m co nsists o Cfi ve eleme nts: th e dial og ue co mpo nent, the prob lem- so lvin g co mpon ent , th e kn ow led ge base, th e ex plan ato r y compo nent and th e co mpon ent for in c remental lea rnin g (D U DEN Informatik , 1988 ) . Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core SC/z11'ei::.er and FiJ/m : A1'OIall che Jorecastillg - an eljJerl 5,J'stelll ap/noarl! (h oweve r, it works ) . So , th e a lgo rithm ca nn o t be g iven in all detail, but w e sha ll tr y to o utlin e th e ge n era l id ea b e low . So far , th e co re of the sys tem m ay be consirlered a " black box" . The judg m ent pro cessor is based on th e fact th a t prag m atic ex p e rts d ec id e, u sin g th eir ex perience a nd in tui tion rather than ex plici t rul es . The more co mpl ex a problem, th e less stru ct ured is th e kn ow ledge . Experts are u s u a ll y a ble to decide correc tl y and fas t in a real situ atio n . H owe \'e r, th ey a r e usuall y h a rdl y a ble to ex pl a in their d ec isions co mpletely following exact rules. Th e e xpert 's approac h is to c ho ose the rcl e\'a nt d a ta (which m ay diner s ubs tantia ll y from o ne situ at ion lo a no th er) , to elas. if I' a nd to analyze th e d a ta a nd fin a ll y to reac h a co nclu sio n. Buildin g up a mode l im'o h- es the fo ll ow in g s teps: 1. A so -ca ll ed judgment jJ1'oblem co n sis ts of a li st of question s a nd th e logic r eq uired to a rriv e a t ajudgme nt - that is, to reac h a co nclu sion or m a ke a d ec isio n - based o n the a nswe rs to those ques tion s. By specifying th e questi o ns - to be a ns\\' ered by yes/ no, mu ltipl e c h oice or num er ica l responses - th e m e nt o r, th e expe rt buildin g up the sys tem , d efin es th e d a ta needed to reac h a spec ific d ec ision a nd th e c rit eria that a re used to ca tego ri ze or e \'a lu a te th e data . Th a t m e an s fo r num eri ca l ques ti o ns th a t th e poss ibl e answe rs mu s t be g rouped into so -ca lled logical rallges (up to fiv e ranges), so th a t th e sys tem ca n lea rn h o\\' th e respon se is norm a lly categori zed. Num e ri ca l qu es ti o ns ca n ta ke the form of calcu lation s in cludin g co nditi ons. 2. On ce a problem h as b ee n d e fin ed, th e mentor " teac hes" the jud g ment processo r by ent erin g exam - pl es (rea l or r ea li sti c d a ta ) and inte rpre tin g the situ at io ns rep rese nted b\' th ose exa mpl es. By obse n 'in g the relatio nship be t\\'ee n th e d a ta an d th e mento r' s d ec isions . th e jud g ment pro cessor build s a logica l m ode l th a t a ll ows it to e l11ul a tC' the m e nto r' s dec isio n s. Th e more co mpl ex th e pro bl em the m o re situ at io n s arc needed. H o we\'e r, as usua l in case-b ased reasonin g sys tem s, th e pe rfo rman ce of th e sys te m in creases fa s t a t th e beg innin g with increasing numb er of silU a ri o ns, reac hes a plateau an d fin a ll y ma y eH' n decrease ( Fi g. 4 ). 3 . Th e judgm e nt processo r calcu la tes the so-ca ll ed logiral im/Jortall l'e of eac h qu es tion based o n th e obse rnl ti o n of th e ment or\ de c ision. Th e log ica l importan ce is a m eas ure ofho\\' a parti c ul a r ques t io n co ntributes to th e logica l model as a \\' hole, base d o n h O\\' ma ny situati o ns within the kn o w led ge base wou ld b eco me indi stin gu ish- a ble if th a t qu es ti o n were to be re m o \·ed. Base d o n th e log ica l impo rt a n ce, gi\'e n as a number fi'o m 1 ... 100 , t h e qu es ti o ns are class ifi ed as so - ca ll ed II/ajor o r min or question s. Th e log ica l im po rt a n ce is co n tin uou sl y upd a ted, so th e sys tem can lea rn in c reme ntall y . After suffi c ie nt tra inin g of th e m o d e l by th e m e nto r, th e m od el perform s th e following s te ps to r eac h a co nclu sio n for a n e w situati o n e nte r ed : 1. If a new situ a ti on is enco unt e re d , th e sys te m tri es to g ive a proposi ti o n fo r the poss i bl e d ec isio n o n rh e basis of rhe past kn ow n situati o n s and o n what is learned a bo ut th e d ec isio n log ic; p a rti c ul ar ly, the class ifi cation into m aj o r a nd min o r qu es ti o n s b ased o n th e pre\'ious ly ca lcu la ted log ica l imp o rt a n ce is O.B IJ) IJ) Q) 0.6 r:::: h 0 -(.) ~ 0.4 0 (.) 0.2 500 1000 1500 number of situations in knowledge base F ig. -1. A {),pical exam/Jle oJ th e pelformanre oJ a case- based reasoning 5,ystem with increasing knowledge base Ol'e r time: th e DA l'OS4 model . use d. A new situ a tion is simil a r to a kn ow n p ast situation from th e knowled ge b ase, if a ce rtain numb e r of the a n swe rs (usua ll y 80 % ) to the m ajor qu es ti o ns arc \\' ithin th e sa me log ica l ra nge . That me ans th at primaril y o nl y the m aj or qu es tion s a r e co n sid ered fo r sea rc hin g for simil a l- s itu at ions . Two pa s t situati o ns th a t arc bo th s imi lar to a n ew s itu a ti o n ma y h ence co in cid e in diffe re nt qu es ti o n s, e .g. if a prob le m co nsists of fi\ 'e major qu es ti ons, four a n s\\'C rs ha \'C to b e in th e sa m e ran ge, and he nce fi\'e diffe re nt poss ibiliti es ex ist 10 r a s imi la r situ at io n , in add iti o n to th e case wh en a pre \'i o us situ at ion is found th a t co in c id es in a ll an swc rs to th e m aj or ques ti o ns. So simi la r situati o ns n ee d not co n se q- u e ntl y be nea r , in a geo m e tri ca l se nse in th e parameters space , to the new situ a ti o n. T I. Th e pro posed d ec is io n is d e ri\ 'C d fr o m the similar s i tu a ti o ns fo und u s i ng the so -call ed (lssertioll !el'el of th e difTerent sim il a r situ a ti o ns. All qu es ti o ns a nd th e ir log ica l imp o rt a n ce arc conside r ed to d e te rmin e th e a sse rt ion leve l. rt is a numb e r (o n a sca le of 1- 100 ) that ren ec ts h o\\' close ly the c urren t situation co mp a res to ex is tin g situ a ti o ns ill th e knowl edge ba se . Th e close r th e a sse rti on le\'e l is to 100, the more simil a r th is examp le is to prev ious ly encountered s itu a ti o ns. Th e less th e answers ag r ee. th e sm a ll e r is th e a sse rti o n le \"e l, i.e. for eac h ans\\'e r th at does not ag r ee, a ce rta in a m o unt is subtr acted ri'om 100 , depending o n th e number o f qu es ti o ns a nd th e log ica l imp o rt ance of th e qu es ti on; in the case of p e rfec t agree m e nt , a so -ca ll ed full match, th e asse rti o n le\'e l is h e n ce eq ual to 100. rn . Th e qua lit y o f the pro posed deci s ion, ba sed o n t h e s imilar situ a tio ns fo und , is d esc rib ed b y th e so -ca ll ed cO lifidenre lerel, a n ind ica to r of how cer tain th e syste m is that its inte rpretat ion is approp ri a te to the curre nt situ a ti o n: a n ex cl a mati o n mar k ( I ) for \ 'ery confid e nt , a p e ri od (.) fo r I'easo nab ly co nfid e nt or a qu es ti o n mark (?) for not co nfid e nt. A low !c\'cl of co nfid e n ce s u gges ts th a t th e re a r e fe \\' situ a ti o n s th a t th e sys tem co n s id e rs to b e logica ll y simil a r , o r th a t th ose s itu a t io ns that a r c simil ar have co nOi c ting int e r - pre ta ti o ns. Additiona ll y, th e simil a r situ at ions th a t a r e u sed to d e ri\ 'C th e deci sio n with th e acco rdin g asse ni o n le\'e l a r e a lso gi\'e n as a n ex p la n a ti o n . J f th e sys te m is not ab le to lind a dec isio n o n th e bas is o f th e 32 1 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ouma l ~! Glaciolog), prese nt kn ow led ge b ase it gives th e r es ult " not poss ible to m a ke a n int e rpre ta ti on", in t h e fo ll owing simpl y ca lled " no inr e rpre t a ti on" . In Fig ure 5 th e different s te ps to reae h a co nclu sion (d esc rib ed a b O\·e ) a re summ a ri zed in stro n g ly simplifi ed g ra phi ca l fo rm. In T a ble I a n exa mpl e o f th e sys tem output is gi, ·e n. Sin ce the sear ch fo r simil a r situ a ti ons fo rm s th e co re of th e m Clh od , it m ay be call ed , in the broa d es t se nse, a n ea res t-n e ig hb o ur met h od . H o ,,·e,·e r , th e m e tri c for sea rchin g fo r simil a r situ a ti o n s differs substa nti a ll y fr om th e co mmonl y used dista n ce m eas ure, e.g . th e Eu cl idi a n di sta n ce . Th e cat ego ri za ti o n of th e input d a ta, th e cl assifi ca tion int o maj or a nd min or qu es ti o n s a nd th e m e tric to sea rc h simi lar situ a ti o ns a re all no n -li n ea r. Brieny summ a ri zed , th e sys tem we ig h s a nd cl ass ifi es th e ca tegor- ized d a ta, sea rc hes fo r simil a r situ a ti ons stro ng ly using th e cl assifi ca ti o n a nd ca tegori za ti o n , d eri ves a res ult fr om th e simil a r situ at io ns, desc ribes th e qu a lity of th e r es ult a nd fi nall y li sts th e simil a r situ a ti o n s used for d e ri,·ing th e res ult 32 2 + and x: past situations with according output (+ o r x) 0 : new situation , output unknown similar situations (c) toget h er ,,· ith th e pert in ent s im il ar ity meas ur e. T he a dva ntage of th e m e th od is th e st ro ng co nce ntra ti o n on th e qu es ti ons th at a re co nsid ered im p o rta nt. Th e fac t th a t th e m a jorit y of th e a nswers to th e m aj or qu es ti o ns h ave to b e in th e sa me log ica l ra nge m a kes th e logica l imp o rta nce o[ a q u es ti on , co mp a ra ble to th e we ig ht used in a d iffe re nt sys tem , to a d ecisi ve [acLO r, in co ntras t to simil a r sys te ms. So differe nt ,·e rsio ns of a mod e l, as we will use, w i th th e sa m e qu es ti on s but with diffe r e nt logica l imp o rt a nce (ca lcul a ted by th e sys tem, no t g ive n arb itra ril y), leadin g to a differen t p a rtiti on of maj o r a nd m in or qu es ti o n s, ,\·ill find vc ry differ ent simil a r situ at io n s. The application: the avalanche hazard In o ur case th e judg m ent pro bl e m is th e avala n c h e h aza rd a nd th e qu es ti ons will b e ca lled input p a l-a m e te rs a nd a re, fo r exa m pie, th e 3 d s u m of n ew snow d e p t h or th e air temp e r at ure . Th e a n swe r s a re th e ,·a lu es o f th e input p a ra m e te r s in a rea l situ a ti o n , e .g . 15 c m o r - 5°C. A rea l situ a ti o n is he nce d esc rib e d by the se t o f in p ut Fig. 5. CrB E RTEK- COG E KS r ST. 1f J udgment P ro - cessor : the different ste/JS to reach a conclusion are given (strongly sil7l/JLified) Jor a problem with on0' three injJut parameters (X/, ){2, .1.3 ) and one output /Jarameter ("x" or "+") . (a) InjJu t /Jarwneter space . (b) Categorization, leading 10 a cube (x/ , \2, 1"3) of 125 identical boxes . (c) Reduction to the major /Jawmeters (x /, \"3), i.e. projection to this jJLane. Selecting simiLar situations (all situations in the shaded squares) based 0 11 the Jollowing similarilJ' conditio,, : similar situations are all past situations with either x/or X3 ill the same catego l)' as fhe lIew situalion . I II the sha ded sqllares are often several similar /Jas t situa tiolls; t/tese si tualions that may have different output differ ill the third (minor) parameter (X2) . Based on the similar siluations, referring to the Logical il7ljJortallce and the minor parameter . the system /Jroposes the result; ill the above case, the /Jroposed ou tput would be "+ ", with a confidence Level cif ")", i.e. not confident, since there are too mClll)' simiLar situations with different output. Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Schwei;:er alld F iJlzn: Al'ala l/ che jorecastillg - all eX/Hrt s),slem ap/Jroach T able I" Exa mple oj screell oUlpu l: A"lODeL model, sub-problem Rel ease proba bilit y n ew snow " !l J anuary , 1995 (s ee ledjor eX/Jlallaliol/ ojlerllls and Figure 8) " T he sub-/JToblem has six ill/Jul !)(/)"{//nelers; the S) 's lelll /Jroposed the oll l/JUI la rgejor th e release probabiLiry , bu t it is /l ol SlITe at aLL, so it /JII IS a question mark" As explanation,jolll" similar situalions are gIVell U se r ID : Pro bl e m cod e : SLF RPROB_.\ S 316 Situ a ti o n N o: " \ ro" DATA 01eslion i\ ew sn o \\" d e pth (5 , 15 , 3 0 , 50 cm ) 3 d sum o f n ew sn ow d e pth ( 10 , 30 , 60 , 100 cm ) D a t e : Pro bl e m n a m e: Appli ca n t cod e: Answer 81 cm 93 1] ] allll(//Y 1995 09: 14 Release probabiLi~)1 new SIlO W 1/0195 (j L/j jJelzded ) Ca lego l) ' #5 : > 50 c m #4: > 60 ::; 100c m I 2 3 Qu a lity o f n ew sn o w (v e r y loose" loose , s li g htl y co nso lid a ted , quite co n solid a ted , co n so lid a ted ) slight!)' consolida led #3 4 5 6 Sn ow te mpe r a ture (3 d ) (co ld , co ld --+ w a rm , wa rm --+ co ld , wa rm , w e t ) Pros p ec ti\ "e ch a n ges (s tro n g ly \\"ea ke nin g , sli g htl y wca k e nin g , with o ut a n y inOu en ce , sli g htl y co nso lid a tin g , stro ng ly co n so li- d a tin g ) I n c rease o f sn o w d e pth (3 d ) ( 10, 30 , 5 0 , 8 0 cm ) I:'-ITERPRETATION R e lease pro ba bilit y n e w sn ow: (ve r y lo w , low, m od e r a t e , la r ge , \"e r y la r ge ) EXPLANATIO:\T A t e nt a ti\ "e co nclu sio n fo r this exa mpl e is situ a ti o n :\T o : 170 raid slighl!)' weakening 77 cm la rge? R e lease p ro ba bili ty n ew sn ow: large [t is import a nt to n o te , h owev e r , th a t in th e a bo \ "e c a se n e,," sno " " d e pth is: in s tea d o f: L ess i m po rt a n t differ e n ces a re th a tin th is case: Th e A S SERTION LEVEL fo r t hi s t e nta ti I"C co nclu sio n is: 85 A te nt a ti ve co n c lu sio n fo r thi s exa mpl e is situ a ti o n N o : 4 1 NONE R e lease pro ba bilit y n ew sn ow: l'e!} Lcuge I t is imp o rta nt to n o te, ho ,,"e\"e r, th a t in th e a bo \ "e ca se qu a lit y o f n c \\' snow is: in s teael o r: L ess imp o rt a nt diffe r e n ces a re th a t in thi s ca se: i\O:\,E Th e ASSERTION LE\ "EL fo r thi s te nt a ti\"e co n c lu s io n is: 8 1 A te nt a ti\ "e co nclu sio n fo r this exa mpl e is situ a ti o n No : 86 Rel ease pro ba bilit y n e w sno \\' : hllge It is impo rta nt to n o te, ho ,,"e ve r , th a t in th e a b o \ "e case qu a lit y o f n c \\' sn ow is: in s tea d o f: L ess impo rt a nt diffe re n ces a re th a t in thi s case: Th e ASSERTION LEVEL [o r thi s t c ntati\ "C co nclu s io n is: 8 1 A t e n t a ti\ "C co n c l usio n fo r this exa mpl e is si tLl a ti o n N o : 29 1 R el ease pro ba bilit y n e w sn o w: It is imp o rt a nt to n o t e, h o weve r , th a t in th e a bove c a se n c,," snoll" d e pth is: in s tca d o f: L ess impo rt a nt diffe r e n ces a re th a t in thi s ca se: in crea se o f s n o ll" d epth (3 d ays ) is: in s t e ad o f: Th e A S SERTION LEVEL fo r thi s te nta ti I"C co nclu sio n is: 8 1 NONE modera le # 1 #2 #4": > 50 ::; 80 cm > 50 c m 2: 15 ::; 30 cm sli g h tl )" co nso lid a ted loosc sli g hll y co nso lid a ted quit e co nso lid a ted > 50 C I11 2: 5 :s; 15 c m > 5 0 ::; 8 0 cm 2: 30 < 50 cm p a r a m e te r va lu es (w eat h e r, sn o w a nd sn o w- cove r d a t a ) fo r th e g ive n d ay " Th e log ica l ra n ges a r e, fo r exa mpl e, in th e c a se of th e 3 d s um o f n ew s n o w d e pth 0 """ 10 , 10 """ 3 0 , 30"" " 6 0 , 6 0"" " 120 a nd 111 0 r e th a n 12 0 e l11" Fin a ll y , th e d ec isio n o r inte rpre ta tion is th e d eg ree o f haza rd a nd in m ost \"e rsi ons o f th e m o d e l DA \ 'OS (see below ) th e a ltitud e a nd th e as p ec t o f th e mos t d a nge ro u s slo pes" 32 3 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ournal 0/ Glaciolog), 4. INPUT, OUTPUT AND VERIFICATION \\'e h a y e c h ose n our input p a r a meters (call ed qu es ti ons in the jud g m e n t processo r ) fro m a d a ta se t wh ich is beli eved to be represe nt a ti\ 'C o f th e region co n sid e red (Da \·os : a bo ut 50 km 2) : seve n qu a ntities a re m eas ured in th e mo rnin g (a t th e e:-:p erim e nt a l pl ot of SFISAR below \\·eiss Ouhj oc h , 2540 m a.s .l. , o r a t th e littl e p ea k a bo\·e th e in stitute, th e so -ca ll ed [n stitutsgipfel, 2693 m a. s.l. , by th e a ut o m a ti c wea th er sta ti o n o f th e Swiss 1\1e teo rolog ica l Institut e ) , fo ur qu a ntiti es a re pros pec ti\ "C \·alu es for th e d ay co n sid cred a nd ten qu a nti ti es d esc ri be th e ac tu a l sta te o f th e sno\\' cO\·e r b ased on slo p e m eas urements perfo rm ed a bo ut every 10 d. Th ese prin cip a l d a ta a r e give n in T a bl e 2. A d esc ripti o n of th e a yal a n che h aza rd is associated \-,·ith eac h d a ta se t co nsistin g of th e a bo\"C wea th e r, snow a nd snow-cover d a ta. I t see m s m ost ap propri a te to choose as o utput o f a n e:-:p ert sys tem e:-:ac tl y th e stru cture th at is usua ll y used by forecas ters. So th e ass istin g tool ·'s peaks·' th e sa m e la ng uage as th e fo recas ter. The qu es tion thu s is: whi ch d eg ree of haza rd d esc ribes th e prese nt ava la nch e situ a ti o n a nd where is it loca ted ? Onl y th e hi g h es t ex istin g d egr ee o f h aza rd is gi\·e n ; th e loca ti on is give n b y th e slopes, d esc rib ed in term s of a ltitud e a nd aspec ts, th a t a re suppose d to be th e m os t d a ngero us in th e region. So th e gi\·e n d egree of h aza rd is not a t a ll a \ ·e raged o\·e r a ll as pec ts; it is d efinit ely w ro ng to d ea l with aYe rages in thi s co ntex t. Th e refo re, th e a \·a la n c h e h aza rd is fo rmul a ted first o f all as d eg r ee of haza rd (l .. . 7) . Seco ndly , t h e lowe r limit T ab le 2. Princi/Jal daLa lI sed ill the two different D il J ·OS and J10D UL models . D, D ata llsed in the DA J·OS model; AI , Data llsed ill the l\r/OD UL model 1. IVI eaSllrements D , M D , M D , M D , M D , M 1\1 ;,,1 11 . Prognostic data D , ~1 D , M ?If M Ill. Snow -cover data D D M M ~r M M :vr M ~I 324 new sn o w d epth to ta l sn o \\· d epth pen e tr a ti o n d epth wind p eed a nd wind direc ti on ai r te m pe ra tu re sn o w te m per a tu re n ew s n ow d ensit y progn ostic air temp e r a ture a t noo n progn os ti c ind ex of d a il y radi a ti o n progn os ti c mea n wind sp eed pro g n os ti c new sno w d e pth ind ex o f snow-cover s ta bility d epth o f criti ca l laye r res ult o f Rutsc hbl oc k tes t ty p e o f release (RE tes t) ty p e o f critica l laye r (RE test) to ta l sla b thi ckn ess (RE tes t) new sn ow sla b thi c kn ess (RB tes t ) ty p e o f profile (RE tes t ) sn o w d epth a t th e t est site d a t e o f Rutsc hbl oc k tes t of th e prim a ril y end a ngere d a ltitud es is giye n in s teps of u suall y 20 0111. ( > 1200, > 1600 , > 1800. > 2000 . > 2200, >2400, >2500 , >2600, >2800 ma.s.I. ). T hirdl y, th e ma in as pect is d esc rib ed as ei th e r o n e of th e mea n direct ions ( .\", XE . E , SE. S. ,\ ·W ) a nd a n acco rdi n g sec tor ( ±4S', ± 67', ± 90 ) or as extreme slopes o r all slopes . Th e \I·es terl y secto r is no t so fr equ entl y end a n ge r ed a nd if so, o th er as p ec ts a re a lso e nd a nge red , so it m ay be d esc rib ed b y th e o th er o n es. If th e haza rd is giYen, for exa mpl e, as 4, >2400 1110.5.1. , N E± 90°, thi s m ea ns hi gh h aza rd on slo p es with aspec t fr o m n orth wes t to so uth eas t a bo\·e 240 0 m a .s.1. Three exa mpl es a re give n in Fi g ure 6. N s 1, > 2500 m, N ± 45 ° N s 3, > 2000 m, NE ± 90 ° N s 5, > 1500 m, all Fig . 6. T hree e\'amples of hall' the regional avalanche haza rd is described. Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Sc/twei;:.er alld F6itll: A l'ClLa 11 che Jorecast ing - all e.\pert ~)Jstem ajJProac/z V erifica tion Th e "real" ava la nche ha za rd th at Vie use in the dat a base of the DA VOS mod e l to teach the sys te m is th e result of an "a pos te riori " criti ca l assess m e nt o f the h azard , th e so- ca ll ed ve rification (Fbhn a nd Schweizer , 1995 ) . Th e verification h as again th e same stru c ture as th e mod e l ou tput . Oth e rwi se, it is h ard '" poss ible to verify th e a\'a la nch e h aza rd. Se\'e r a l s tudi es 011 the \ 'e rifi ca tion o f th e a \'a la n c h e haza rd with the help of th e so -ca ll ed a \'a la nc h e - act ivi ty ind ex were no t sufTi c ie ntl v su ccessful U ud so n a nd King, 1985; Giraud a nd o th ers, 1987; Remund , 199 3 ). On e reaso n is that in th e case when no a \'a lan c h es are prese nt or o b se rv ed . th e a va la nch e haza rd is no r necessa rily \'e r v low o r even no n- ex is tent. H e nce, it is ob\' io u sly wrong to use th e o b se n 'ed a\'a lan c he ac ti\'it y as th e so le o utput pa ram e te r o f a n ass is ting tool fo r r eg iona l ava lanc he fo recas tin g. Th e ava la nc he \'C rifi ca ti o n IS not id e ntical to th e ,l\'alanche h aza rd desc rib e d in th e public a\'a la nch e warnin g . A s th e warning is prosp ec ti\'e a nd th e d ata ba se m ay b e in co mpl ete a t th e tim e of fo recas tin g. the a\'a lanc h e h aza rd might h a \' e been in fac t la rge r o r sm a ll e r. It is a hl'a ys eas ie r to assess th e a\'a la n c he haza rd in hind sig ht! Op era ti o nall y, th e \'e rifi ca ti o n has bee n d o ne some d ays la te r co n s id e rin g the obse n 'e d a\'a la nc he aC li\'it y ( natura ll y a nd a rtifi c ia ll y released ) , th e pas t weat her co nditi o ns, the a dditi o na l sno w- cover tes ts. th e bac k- co untry skiin g ac ti\'it y a nd se \ 'C ral o th e r , partly perso na l o b se n ·a ti ons. Snow-cover tes ts for m a n impor- ta nt part o f th e \'e rifi ca ti o n lI·o rk. Lik e th e rea l-tim e assess m en t itself, it is an ex p e rt task . W c est im a te th a t th e \'Crification d esc ribes th e ava la n c he situ a ti o n co rrec tl y in a bo ut 90'10 o f th e d ays a nd thu s is much more a cc ura te th a n th e forecas t. L' sin g the d eg ree of h aza rd fi 'o m th e ava lan c h e fore cas t as o utput parameter would o n ly be rea so nab le if the lI'ar nin g w e re a hl'ays co rr ec t. An e rro neo us i nt e rpreta ti o n s hou ld not be e n c losed in th e d ata base. Comparin g th e fo r ecas ted d eg ree of h azard to th e \·e rifi e ation. th e fo recast seems to be co rrec t in a bo ut 70% o f th e d ays . So it see m s ob\· io us that th c use of thi s typ e o f \ 'C rifi ca ti o n rep rese nts a substa ntia l impro\'C m e nt (or the d e vel o pm e nt o f' ex p ert sys tems in the fi e ld of reg io na l a \ 'a la nche for ec a s t i n g . Furtherm o re, \ 'C rifi ca t io n is a prereCJuisite [or c\'aluating th e qual ity of a n y assistin g too l for reg io n a l a \'a lanche for ecas tin g a nd , of co urse, a lso fo r impro \ 'in g th e lI'arning itself: 5. MODELS Us in g the CYBERTEK-COGENSYST\I jud g m e nt pro - cess in g sys te m , we d e\ 'C lo p e d tw o dif'rerent typ es of' mod e l: DA VOS and I\ fOD U L. Th e DA \ ' OS model uses 13 weather, sn ow and sno \\'- co\'e r param e te rs an d e \'aluales th e degre e o f' ha za rd , th e a ltilUd e a nd th e as p ec t. The m ode l is ex elu si\'e ly d a ta-b ased, wherea s th e ~IODU L m od e l is both d a ta - a nd ru le-b ased. It u ses 30 input p a ra m e te rs s tep ll'ise a nd the e \'a lu a ti o n o f the d eg ree of' h aza rd is th e res ult of 11 inter co nn ec te d jud g m e nt pro ble ms that a re fo rmu lated acco rdin g to th e re leva nt processes . Th e sys tem tr ies to m o d el th e d ec ision-m a kin g process o f a n ex pe rt a\'a la n c h e foreca ster. DA VOS IIlodel Gellf1'al Jeatllres : inpllt . OlltPllt and kll owLedge base Th e DAVOS m ode l uses th e input p a ra me te rs g i\ 'e n III Tab le 3. M os t of' th e va lues a r c calc ula ted fr o m nin e prin ci pa l va lu es (Tabl e 2 ) acco rd in g to o ur ex p e ri e n ce . Th e id ea was to ta ke int o accollnt ce rta in rele va nt pro cesses, e.g. n e w snoll' se ttl eme nt. D e ta il s h a \'e bee n g iven in Schw e ize r and oth er s ( unpub li shed ) . I n th e fo ll ow in g so m e o f th e ela bo ra ted p a r a m e ters are bri e fly d esc rib ed. The settLemellt-qllotienl p a ra m e te r co mpares th e in c rease in sn o ll' depth durin g th e las t 3 d wi t h th e s um of' th e nell' snow d e pth o f th e las t 3 d. Th e sma ll e r th e \' a lue th e be tt e r th e se ttl eme nt. H owe \ 'e r , the se tt le m e nt inelud es no t onl y th e new sn ow but a lso the o ld-sn ow se ttl e me nt. Th e co nso lidati o n o f th e surface laye r is d esc ri bed by th e /Jenelraliol1-qlloliml param e te r : th e p e n e t- r a ti on d ep th today di\·id ed b y th e pe ne tra ti o n d e pth yes te rda y . Th e h e at tra nsport int o th e snow cO\'e r is taken int o acco unt by a d eg ree -d ay p a r a m e te r: the slim of th e /Jo silive air tem/Jeralllres alllOOIl oJ Ih e lasl 2 d and the presenl dCI)' ( prospec ti\'ely ) al 200011/ a .s./ . ( the average a ltitud e o f th e reg io n co nsid e r e d ) . Th e snow tr a n s p o rt is includ ed by th e blowing -sl/ ow param ete r: th e s lim of a dditi o n a l wind- tr a nsported snow in leewa rd slo p es OH r th e la s t 3 d ( F b hn and Haec hl e r, 19 78 ) . Th e radialioll illde.\ IS an es tim a ti o n ( I , 2 o r 3 ) o f th e daily lo tal g loba l so lar radiation fo r th e prese nt day. I mean s below th e lo ng - term m ea n \'a lu c (or th e g i\'en d ay, 2 about a nd 3 a b o ve , res p ec ti\ 'C ly. Th e .mOW -CONr slabili£J' illde\ ( I to 5 ) is an es tim a le of th e s tate of th e sn ow CO \'e r co nsid e rin g th e sn ow pro (il es a nd Rutsc hb loc k tes ts th a t a re a\'ail a ble [or th e reg io n. The dep lh oJ Ih e crilical /c£J'er is a n es tim a ti o n fi 'o m th e snow profiles a nd th e Rutsc hbloc k tes ts. W e u s u a ll y di spose of th e sno w profile from th e stud y plot and at least of o ne t y pi ca l sno'" pro (ile with a Rutsc hb lock tes t ri'o l11 a slope in th e D a \'os a rea, th e la tt e r ofte n performed b y o urse h-es. Bes id es th e input param e te rs, w e h a\'e a lso c h ose n th e ran ges for each of' th e input p ara l11 e te rs acco rdin g to o ur T able 3. Inpul jJarameters al/d /ogi{{{/ rallgeJ ./or th e D.·( J 'OS model Inpul /Jaram elers Sum of ne\\' snow d ep th (3 d ) Penetra ti o n d e pth Tot a l snoll' d e p th (3 d befo re ) Sett le m en t quotient Pe netra tion CJ u o ti e n t SUI11 o f bl o win g s n o w (3 d ) Air te mpera ture T e mp e rature diITe re nce Sum o f th e positive te mp era tures at noo n a l2000I11 a .s.1. (3 d ) Ind ex o f radiati o n I nd ex of sno\l' -cove r sta b ility W ind direc ti o n D e pth of cr iti ca l lave r B ollnda ries / choices 10 / 30 /60 fl2 0 CI11 5/15 /30 /50 e l11 70 / 1 00fl50 /200 cm 0.01 /0. 36 /0 . 7/ 1.0 0.4/0.8 / 1.2/ 3 .0 2 / 5 / 1 Ofl5 Clll 15 / 8/-3/0 C 5 /0 /5 / 10 c: 0.0 1/ 3/6 / 10 C I , 2, 3 I, 2, 3,4,5 N W , NE , SE , S W , 00 20 / 40 /60 /90 c m 325 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ournal oJ Claciology ex p eri ence . As m e nti o ned above, eac h of th e in put d a ta is assoc iated with o n e of up to fi ve logica l ra nges. Aft er seve ra l yea rs o f d a ta acc umul a ti o n , we ar e fin a ll y ab le to ch ec k w heth e r th e c hose n ra n ges w c rc r easo na bl e o r n ot. T wo exa mpl es b ased on th e 9 yea r d a ta base a r e g ivc n in F ig ure 7. \ Vh e reas th e 3 d sum o f th e new sn ow d epth see ms to catego rize quit e we ll , co m p a red to th e ve rifi ed degree of haza rd , th e set ti e m e n t q uo ti en t sh ows n o sp ec ifi c trend. This is in acco rd 'vV i th th e st ud y o f P erl a ( 1970 ) but it is in co ntras t to th e o pinio n of expe ri en ced fo r ecaste rs. Th a t p roba bl y d oes no t mea n th at th e set tlement qu o ti ent is not imp o rt a nt a t a ll ; it mi g ht be releva nt but o nl y in ce rt ain situ a ti o ns, i. e. co mbin a ti o ns of input d a ta . Sta tisti ca l me thod s, in pa rti cul ar uni var- ia te, d o not tell th e w hole truth. Th e DA VO S m o d e l a lso d oes no t co nsid e r th e set ti em en t q uo tien t as im p o rra n t (T a bl e 4 ) . Th e output res ul t is th e ava la n ch e hazard d esc ri be d as deg r ee of h aza rd , a ltitud e a nd as pect of th e m ost d a n gero us slop es (see a bove ) . Th e kn ow led ge b ase of th e D A VO S mod el co n is ts of th e d a il y d a ta fr o m nin e winte rs ( I D ecemb er- 3D April ), i.e . 136 1 situ a tio n s. Durin g thi s tim e period , 22 si t u a tio ns we re pa irwise id e n ti ca l, i. e. eac h o f th e in p u t p ara m e ters of th e 2 d co nsid er ed belo ng to th e sa me logica l ra n ge. A bo u t 700 000000 situ a ti ons a r e theo reti call y bu t, of co urse, n ot ph ysica ll y p oss ibl e . DijJerent versions: D A I "OS] , D A VOS2, D A VOS3 1/DA VOS32 and DA VOS4 The o ri gin a l versio n of th e DA VO S model w as call ed DA VO S I. Th e ex p eri ence wi t h thi s ve rsion h as g ive n r ise to fu rt her ver sio n s. T he \'alu es of th e logica l imp o rt a n ce of th e o ri g in al D AVO S I ve rsio n (T a bl e 4 ) show cl ea rl y th at this ve rsion is hardl y a bl e to disc rimin a te well. T welve of th e 13 input pa ra m eters a re co n sid e red as m ajo r o n es. As a con seq uen ee, loo ki ng for si mil ar si tu a ti ons m ea ns th at te n of the prese n t-d ay inpu t par a m e ter va lu es h ave to be in th e same ra n ge as the past-d ay va lu es. T hi s r eprese nts a ve r y st ri ct co nditio n for th e simil a rit y a nd r es ul ts in a 7 r---~----------~---------------, 50 100 150 200 250 3 d sum of new snow de pth [c m] 7 ~------------~--------------------, "0 6 - - • - ~ .. + - - .. :- .. - - • - - - - ~ co: 5 -------. -:- 0 .•• -;... -;- --- - ~ '0 4 :.- . .. ... --. . ......... ~- <I> ~ 3 ____ "'!II!!'O ••• - - -. - Cl> .....--: {l 2 . "~=::::':....;.~ V 1 ~ __ ~~ __ ~~~~--+-----~-----4 o 0.5 1.5 s ettleme nt coefficient 2 2.5 Fig . 7. Comparison oJ the 3 d slim !if the new SIlO W depth (a bove) and the settlement quotient (below) with the deg ree of hazard to check whether the logicaL ranges chosen categorize the data appropriately . Average degree oJ hazard Jar each category is also shown . T he da ta base consists of 136/ situatiolls from nine winters. la rge numb er of unin terp reta bl e situ a ti o ns. Thi s fac t see m s d e finit ely to h e du e to th e d esired o utpu t res ult th at co n sis ts of three ind ep e nd ent a nd equi va lent co mp o n ents (d eg r ee of haza rd , a lti t ud e a nd as p ect ) . In th e case of different ind epend ent output res ults, th e CYBERT EK- CO GEN SY SH 1 judgment processo r offers the poss ibili ty or choos in g one of th em as th e d omin a nt ou tp ut r es u I t. H ence, we use d a seco nd ve rsion or the mod el D A V O S: DAV O S2. 'vVh ereas in the D A VO S I \'e rsion a ll three res ul ts are eq ua ll y importa nt, in th e D A VO S2 ve rsion T able 4. Values oJ lite logical imjJortance oJ the different versions of the D A VOS model. B old figllres indicate so -called maj or parameters InjJut parameters D AVOS l D AVOS2 DAVOS31 DAVOS32 D AVOS4 EHN < IDem ) EHN 2: ID el11 S um of new sn ow d ep t h (3 d ) 0 100 0 100 100 Penet rat ion d ept h 83 28 14 47 29 T o ta l snow d epth (3 d before) 83 65 79 60 65 Se ttl ement qu o ti ent 50 2 1 15 18 19 P e n e t ra ti on q uo ti en t 41 24 18 36 27 S um of b lowin g sn ow (3 d ) 66 33 35 41 32 A ir te m pe ra ture 66 23 22 18 15 T e mp era ture difference 24 15 7 17 13 S um o f the positi ve temperatu res a t n oo n a t 2000 m a.s. 1. (3 d ) 41 29 30 4 22 In dex or ra di a ti on 44 II 17 II 16 Ind ex of snow -cove r s tab ility 100 86 81 70 77 Wind di rec ti o n 33 26 26 18 2 1 D epth of criti ca l laye r 79 51 100 47 83 326 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Sclill'ei::er and Fiilzn : .. h'a/allclie forecasting all e.\/Jnt sjlstem ajJProacli th e first o ut p ut res ult , th e d eg ree of haza rd, is th e d o min a nt o ne. Thi s lea ds to different \'a lues of th e logica l impo rt a n ce a nd acco rdin gly to d ifferent int erpre ta ti o ns. ~los t re le\'a nt is th e sm a ll numb er (fo ur ) of maj or p a ra m e ters a nd he nce th e be tt er se lection pe rfo rm a nce: ha rdl y a n y unint erpre ta bl e situ a tio ns, In th e D A VO S2 \'e rsion , th e va lu es of th e log ica l impo rta nce see m to be close r to ge nera l ex peri ence th a n to th e DA VO S I \'e rsion ", he re, for exa mpl e, th e ne\\' snow d e pth has no impo r tance a t a ll, Th e reaso n is t he so n of ou tput resu lt a nd th e predo min a nce of situ at ions wi th no new snow , Fro m ex peri e n ce, it is o b\'io us th a t it is quite imp o rt a nt w h e th e r fo r a g i\'e n day th ere is n e w sn o \\' o r no t. H e n ce, wc tri ed to ta ke int o acco unt thi s fac t b y d e\'e lo pin g t\I 'O n ew ve rsio ns: o n c fo r th e mo re fi-cqu e nt situ ations with o ut sn o \\'fa ll a nd th e o th er fo r th e situ at io ns \I' ith sno \l'fa ll - th e DA \'O S3 l a nd D A \ ' O S32 \ 'e r sio ns, r es pec ti\'e ly, Th ese two \'e rsio ns co ncc ntra te o n th e d eg ree o f h aza rd lik e th e DA VO S2 \'e rsio n. Th e kn o wl ed ge base o f th e D A VO S3 1 I'C rsio n co nsists o f a ll d ays fr o m t he las t nin e w int e rs, w hen th e 3 d sum of ne\\' sn o w d e pth was less th a n 10 c m ; th e co mpl e m e nt a r y se t o f d ays w ith th e 3 d sum o f ne \\' sn o\l' d epth cC] u a l to or la rge r th a n 10 cm fo rm s th e kn o wled ge base o f th e DA VO S32 \ e rsio n. Th e d iffe re n ces in th e log ica l im po rt a nce in th e tw o \'e rsio ns a rc quite t y pi ca l (T a ble 4 ). Th e \'a lu es o f th e log ica l imp or ta n ce of th e DA \ 'O S3 1 (n o n e \l' sno\l' ) \'ers io n a re simil a r to th e log ica l imp o rt a n ce o f th e o ri g in a l DA VO S I \'e rs io n , w h e r eas th e log iea l impo rt a nce of th e DA \'O S32 ve r sio n (n ew sn o\l' ) a re simil ar to th e o nes o f th e D A \ 'O S2 \'e rsio n . Th e num be rs o f m aj o r pa ra m e te rs a rc fo ur a nd seve n , res p ee ti\ 'C ly, fo r th e D A VOS 3 1 a nd DA VOS 32 \'c rsio n s; so th ey sho uld di sc rimin a te q uit c we ll, Fin a ll y, th e o utput r es ult \l'as redu ced to the d eg r ee o f h aza rd : D A VO S4 . Th e D A \ ·OS 4· \'e rsio n is m os t a ppro pri a te fo r co mp a ri so n \I' ith s imil a r fo recas tin g m o d e ls a nd \I'C h o p ed th a t du e to th f' sin g le ty p e of' o utput th e DA \ ' O S 4- \'(' rsio n sh o uld di sc rimin a te b e tt e r th a n th e o th e r \'e rsio n s, In a ll th e diffe re nt I'C rsio ns th e input pa ra m ete rs d esc ribin g th e Sla te of th e sn ow cOl'e r pro \'ed to b e imp o rt a nt (T a ble 4 ) , MODU L n lOdel General featu res all d structure Th e ex p e ri ences w ith th e di ffe re nt \'e rs io ns of th e DA \ ' O S mod e l, d esc rib ed a b Ol'e, di rected u s to tr y a differe nt , m o r e d e te rministi c a pproac h, Ori g in a ll y, ~IT hoped th a t th e DA \ ' O S m o d e l. based o n th e jud g m e nt p rocesso r , \l' o uld be ab le to c h oose th e re le \'a nt p a ra me te rs fro m th e 13 input pa ra m e te rs a cco rdin g to th e situ a ti o n a nd ge n e rall y so m e h o \\' to recog ni ze th e hidd en stru c ture o f reaso nin g behind it. D esp it e th e sa ti sfac to l'\' res ults o f so m e of th e \'e rsio n s o f th e DA VO S m od el (sce sec ti o n b e low ) , it sec m s that thi s aim w as too a mbiti o u s; th e pro blem seems to b e too co mpl ex or th e m e th od no t good e n o u g h, So \I'e d ec id ed to " h e lp " th e sys te m b y s tru c turin g th e input d a ta. Th e d es ig n of th e :-IOD U L m o d e l is th e re fo re q uitc simil a r to th e \I'ay a prag m a ti c ex p e rt fo recas te r d ec id es ( Fi g, 8 ) . Firs t of a ll , it is d ec isi\'e ,,' hc th e r th e re is ne\\' sn ow o r no t. Eith er th e ex p e rt h as to assess th e n ew -sno \\' s tabilit y o r h e /sh e di rec tl y assesses , \\' ith o ut n e w sno\\', th e o ld- w eath er. snow and snow cover data Fig , 8, S tructllre qftli e MODeL model : 11 sub -problems alld their relatioll , Shaded bOles are 011£), considered in tlie ((oe (if lIel(, SIlOl('. sno \\' s ta bili ty w hi c h is o ft e n simil a r to th e stabilit y I cl befo re, exce pt if the re is, fo r exa mpl e, a large in c rease o f hea l tr a n s p o rt a nd /o r ra di a ti on , So h e / s h e s tru c tures t h e inpu l d a t a acco rdin g to th e dine- re nt ste p s in th e d ec isio n p rocess, I f bo th th e n e \\'-s n o w sta bilit y a nd th e old-sno \\' sta bilit y, in c ludin g b o th th e efIec t o f th e ,,'ea th e r as fo recas t fo r tod ay, h a \'c b ee n assesse d. th e t\l'O re lease p ro b a biliti es a rc co mbin ed. T a kin g int o acco unt th e effec t or th e te rra in a nd o f th e s kier as a tri gge r , th e d eg ree o f haza rd is fin a ll y d e te rmin ed, At th e m o m e nt, o nl y th e d eg ree o f ha za rd is g i\'e n ; th e a ltitud e a nd th e as pec t o f th e l1l os t d a nge ro us slo p es, as g i\'e n in m os t o f' th c \'e rsio n s o r LI lt' DA \ ' O S mod el, h ",'e no t )Tt b ee n impl eme nt ed. Sub -j1m blellls Eac h o r th e sub-pro bl e m s as , for exa mpl e, qllali{J' qf lIew SilO1£' o r stabili!J' DJ old SII01I' re prese nt s a jud g m e nt probl e m , as d esc rib ed a bo\'e, a nd is hence prin c ip a ll y stru c lured lik e t h e D A \ 'O S mo d e l. Th e d iffe re nt s ub-pro bl ems a r e just s m a ll e r th a n th e D A VO S model , i. e , co nsist o f o nl y three to e ig ht input p a r a m e te rs, O fte n. o nl y three o f th e input p a ra m e ters a rc co n sid e red as m aj o r pa ra m ete rs. Thi s is a g rea t ach 'a nt age, since a mu c h sm a ll e r kn o \\'led ge ba se is suffi c ie nt to ob ta in good int erprc ta - tio ns a nd th e SI's tem us u a ll\' lea rn s fas te r a nd bette r fr o m , , th e log ic b e hind th e d ec isio n process , S ub-pro blem s w ith o nl y a b o ut fi\ 'e input p a r a m e tcrs mos t o f t h e tim e m ay find a simil a r pas t situ a ti o n that is id e nti ca l to th e ne w situ a ti o n , b ased o n a kn oll' led ge base o f o nl y a bo ut 10 0 situ a ti o n s. i mjJ/icit I'll les It is e \'e n p oss ible no t o nl y to build up th e kn o \l'led ge base \I'ith r ea l situ at io n s but a lso to co n Slru c t rea li s ti c situ a ti o n s b y \'a ryin g th e m aj o r input p a ra m e ters in a reaso n a bl e se nse. Thi s is im poss ibl e in th e DA \ ' O S mod e l. S o, if th e ex pe rt fee ls sure a bo ut o n e o f th e sub- pro bl e m s o n th e innu e n ce o f o ne o f' th e input pa ra m e te rs, may b e in co mbin a ti o n \I' ith a no th e r o n e, he/shc m ay sys te m a ti ca ll y co nstru c t rea li sti c situ a ti ons a nd d ec id e sys te m a ti ca ll y , But thi s m ea ns no thi ng o th e r th a n in c ludin g a rul e, n o t ex pli c itl y but impli c itl y , An exa mple o r such a n impli c it d cc isio n rul e used in th e 327 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Journal oJ Glaciology T able 5. Gel/eral Tule to decide on the degree of lza::.ard in the sub-jJroblell1 fin a l m e rg in g; principally dep endent 011 the co mbin ed (na tura l) r e lease pro b a bilit y and the infl u e n ee of th e ski e r , but also de/mulml all lhe overa ll c ri ti ca l d ep th by lhe po telltial avalanche si::.e and volume and 0 1/ lhe d e pth of s ta bl e o ld sno w b)i the lerraill roughness Influence oJ skier L o w ?-.l od er ate Hi g h Combined release probabilil)' r 'e ly Low lvIoderate High low I 2 I 2 3 2 3 4 4 4 4 V a lid if o " e r a ll c ritica l d e pth H cril = 15 . .. 5 0 c m , else, if H"'il < 15cm.l d eg ree o f h azard lessor ifHcril ;:::50c m a nd com bin ed rel ease p ro b a bilit y hi g h , th e n d egree o f h aza rd = 6 or 7 a nd if d ep th o[ sta bl e o ld sn o w H bouncl > 6 0 c m , e lse, if H bound = 30 .. . 60 cm , th e n I d egree o f h aza rd less o r if H boLl ll cl < 30 cm , th e n 2 d eg rees o f h aza rd less exce pt i[ co mbin e d rel ease p ro b a bilit y = moderate, n o redu c ti o n of h aza rd , o r i[ co mbin ed re lease prob a bilit y co nsid e ra bl e, o nl y I d eg r ee of haza rd less . su b-pro bl e m Jinal merging is g i" e n in T ab le 5 . Thi s is, o [ co urse, r a th e r ex ha ustin g w o rk bUL th e a d va ntage is th a t on e is m o re fl exibl e in o n e's d ec isio n th a n in th e case w he re o n e uses a stri c t ex pli c it rul e. It is easy fo r exa mpl e to includ e no n-lin ea r rela ti o n s. Furth erm o r e, it is poss ibl e to co n s tru c t ex trem e but s till rea li sti c situ a ti o ns th a t a re usua ll y r a re but o f co urse v e r y imp o rta n t. S o o ne o f th e di sad va n tages o f prin ci p a ll y sta tisti c (o r d a ta )-b ased m o d els us in g rea l d a ta m ay be ovc rco m e. Fin a ll y, o n e a rri ves a t a knO\d ed ge b ase th a t is a mixture of rea l, hi storic si t u a ti o ns d ec id e d acco rdin g to th e " erifi ed h aza rd a t th ose tim es a nd rea li sti c si t u a ti o n s direc tl y d ec id ed acco rdin g to ge n e r a l kno wled ge a nd ex peri ence . Th e pro bl e m is to h ave th e a ppropri a te mi x ture. InpuL parameters Thirty input pa ra me te rs (T a bl e 6 ) a re u sed in II sub- probl e m s inte rconn ec ted p a rtl y b y rul es . Alread y, to o bta in so m e of th e d a ta, a use r Wilh ce rt a in skills a nd ex peri e n ce is required. Th e o utput res ult o f a s ub-pro bl e m is usu a ll y used as a n input p a ra me ter to a noth e r sub- prob le m th a t a pp ea rs la te r o n in th e d ec isio n-m a kin g process . ?-.1 a n y o f th e input-p a ra m eter va lu es a re ca lcul a ted usin g rul es th a t d epend th e m se lves o n th e input valu es . Th e overall critical depth [o r exa mple d epe nd s, a mon g o th e r things, o n th e 3d sum of blowing-sll ow dep Lh th a t is o nl y con sid e re d in ce rta in situ atio n s wh e n sn owdrift is lik ely. NI odifica tions Du e to th e m odular stru c ture, it is eas il y possibl e to m odify a n y o f th e s ub-probl e m s. Additi o n a ll y, th e rela ti ve ly sm a ll numb e r o f input p a r a m e te rs in eac h sub-pro bl e m e na bl es th e kn ow led ge base to a dapt qui ckl y to a n y m o difi ca tion , as [o r exa mpl e a ddin g a n ew input 32 8 T able 6. inp ut jJarameters used ill the Ai 0 D U L model . T he data are grouped according to the availability, i.e . how UIS} it is to gf't thp rlata A . Conventional data New snow d e pth S um o f n e w snow d ep th ( 3 d ) D ensi ty o f n e w snow Snow d e pth C ha nge o f sn ow d epth (3 d ) Coe fIi cie nt of se ttl eme nt (3 d ) Pene tra tion d e pth Coe fIi cie n t of penetra ti o n d e pth Snow te mp e r a ture M ean wind sp eed (3 d ) Sum of blowin g snow (3 d ) Air te mp e r a ture T emp e r a ture differen ce B. Prognostic data N ew snow d e pth in th e eve nin g T e mp e rature d eve lopment until noon M ea n wind sp eed until tom o rrow Progn os ti c ind ex o f ra dia ti o n fo r tod ay C. S/Jecial snow-cover daLa R es ult o f Rutsc hbl oc k tes t T ype o f r e lease (Rut schbloc k tes t ) T ype o f c riti ca l laye r (Rutsc hbl oc k tes t ) T o ta l sla b thi c kn ess (Rutsc hbl oc k tes t ) Ne w sn o w sla b thickn ess (Rutsc hbloc k tes t ) T ype o f ram pro fil e (Rutsc hbl oc k tes t) Age o f Rutsc hbl oc k tes t C ha n ge of sn ow d epth sin ce Rutschbl oc k tes t C riti ca l d e pth of ne\\' snow sla b C ri li cal d e pth o f old snow sla b O ve ra ll c ri ti ca l d e pth EfTcc ti ve c riti cal d epth for s ki e r tri gge rin g De pth o f s t a bl e o ld sn ow p a ra me te r. So th e impo rta nt s ub-p ro bl em influence oJ the skier is stea dil y improved a cco rdin g to th e r es ults o[ sp ec ifi c stud y on sla b- ava la n c h e release trigge r ed b y a ski er (Sc h weize r , 1993 ) . [n th e s u b-pro ble m sl1ow-jJ70Jile analysis, th e snow pro fil e with Rutsc hbl oc k tes t is ro ug hl y interpre ted , a n a im th a t would ac tu a ll y need a n ex pert sys tem itse lf. Ei g ht prin cipal v a lu es (T a ble 2 ) a re used ex clu sive ly for solving thi s sub-pro bl em. T oge th e r with th e sub-probl e m sLabili!), of old snow, it should s ubstitute th e mos t import a nt input p a r a m e ter index if snow-cover stability in th e DA VOS mod els. S o thi s sub-probl e m is a lso und er pe rman e nt improve m e nt. R ece ntl y, t)ipe of release and th e quali{y oJ the critical la)ier o f th e Rutse hbl oc k tes t we re in trodu c ed. Operalionaluse Th e mod e l has to be run int e ra ctive ly by a n ex p eri e nced use r. Th e mod e l sto ps if th e pro p osed d ecision in o ne o f th e sub-pro bl e m s d oes no t h ave a hi g h level o f co nfid e nce, a nd th e use r h as to co nfirm th e d ec isio n befor e th e m od el co ntinu es to run. In th e ex ampl e g ive n in T a bl e I , th e interpre ta tion of th e release /) To babilil)! oJ new snow b y th e Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Schweizer and F61111: Avalanche Jorecastillg~ all expert system ap/Jroac/z Table 7. Qj/OIiI)I requirements Jar determinillg th e pelJonna nce cif the D AVOS mo del Qualit)l D eviation : degree D evia tioll: alt iLude Deviation : aspect of hazard Good 0 ±400 m About ri g ht F a ir 0 ±400m Not co mpletely wro ng 0 \ '\Trong (a ny res ult ) Wron g (any r es ul t) ±l ± 400 m Abo ut ri g h t P oor ±l > ±400 m N ot compl ete ly wrong ± I > ±400 m Wro ng (any r es ult ) Wron g > ±l (any res ult ) (any r es ult ) sys te m IS large?, w hi c h m ea n s th a t th e sys te m is not co nfid ent a nd the int era ctive run wou ld stop. A s a n exp la nation, four simil ar situations are give n w ith th e in te rpr eta ti o n s large, lIel} large, large and moderate . Comp a rin g th e pr ese nt case w ith th e fir st simil ar s itu a ti on indi ca tes ra th er vel)' large th a n l{JIge as o utput for the prese nt situ at ion; th e seco nd a nd third simil ar situations indica te th e output is b e twee n large and vel)' large; th e fourth situ at ion is too far from th e prese nt case to be co nsid e red. So, based on simil ar situation s, the use r wo uld pres um ab ly c h a nge th e int e rpretation to ve l)! Large. Hmve\'e r, th e int erpre tation proposed by th e syste m , large , wou ld no t be wrong bu t vel~Y large see ms to b e more consistent w ith the pre se nt kn owledge base , Th e final output r es ult , th e d eg ree of haza rd , is w ell ex pl ain ed by th e output res ult s o f the diOere nt s ub- proble ms. If th e mod el proposes a differe nt d eg r ee of h aza rd than th e u se r h as ind e pendentl y es timated , the diffe re nc e us uall y b eco m es obvious by in spec tin g th e output res ults of th e sub-problem s. Du e to thi s featur e, th e use r exp e ri e nce s th e mod el not as a bla ck-b ox syste m (d es pite th e prin c ip a ll y unknown a lgorithm ) but as a r ea l supportin g too l to th e forecaster . Th e interac tive u se of th e mod el prov ed to b e \'e r y in st ructiv e. 6. RESULTS Thank s to ve rifi ca tion data, it is poss ible to r ate our m od els quit e obj ec tiv ely, co mparing th e model output da y-b y-d ay to the verifi ca ti o n. Durin g th e last 5 yea rs of opf'ra ti ona l use, th e kn ow ledge base has con ti n u ous ly in c reased. Sin ce th e pe rforman ce of th e mod els d e pend s s tro ngly on th e s ta te of th e kn owledge base, th e r es ults a re not homoge neo us. Thi s is es p ec ia ll y tru e for th e firs t wint e rs with vers io n s of th e DAVOS m od e l. For co n sistency b e tw ee n th e diITe ren t mod els and \'e rs io n s, w e w ill on ly prese nt in th e following th e performance res ults of three winters ( 199 1- 92, 1992 - 93 a nd 1993- 94 ) . DA VOS nlOde } To co mp a re the int e rpre tat io ns provided by th e sys te m with th e ve rifi ca tion , the require ments of quality (four classes: good, Jair, /J oor, wro ng ) (Tabl e 7) were defined . If th e verifi ed aspect is e.g . AE ± 45'" , the rating in th e following cases]v' ± 67>, N Il ' ± 90° and S ± 90° wou ld b e abo1lt rigId, not compleleL)1 wrong and wrong, res pecti\·ely. Cons id er in g the d eg r'ee of' hazard , the a ltitud e a nd th e aspect th e DA VOS I and DA VOS2 ve r sions hav e on average a performan cc of about 65 and 70 % good orJair (see T a bl e 7 for d efiniti ons ) int erpretations , res pec tivel y (Tabl e 8 ) . To be ab le to co mpare the res ults of the DAVOS 1 and DAVOS 2 versions to the res ults ofdilIer e n t sys te m s, it is more co n ve ni e nt to co nsid er on ly th e d egr ee of haz a rd. In that casc, in 52 and 54% of a ll situ atio n s, res p ec tively , for DA VOS I a nd DA VOS2 , the deg ree of haza rd "v as co rrec t co mp a r ed to th e ve rific a tion. 86 a nd 89% of a ll situation s, res p cc ti\'e h' , a re co rrec t or d eviate ± I d eg r ee of hazard fr o m th e ve rili ca tion. Table 8 PelJormante oJ th e D/l I 'OS 7 and DA rOS2 ve rsiolls considering all Ilzr ee Ollt/Jllt results: the degree oJ ha z ard, the altitude alld tile as/leet. ,\ Jean valll es ( proportions ) oJ tlte la st Ihr ee zt'in ters ( /99/- 92 to /993- 94 ) Jor tlte different qualities defin ed ill Table 5 are given QjJaliljl oJ res1Ilt DJ flGSJ DA I'OS2 Good 0.3 5 0.37 Fair 0.32 0.33 Poor 0.19 0.19 Wron g 0.07 0.04 No int erp retation ( n . i. ) 0.07 0.07 As th e o utput res ult no illter/Hetation is cons id e red as neith e r ri g ht nor wrong, th e performan cc ma y be g iven considerin g on ly th c inte rpreted situ a tion s. In that case , the proporti on of co rre c tl y int erpret ed situ ations (degre e of h azard ) is 55 a nd 58 %, r esp ec tivel y, for DAVOSl a nd DAVOS 2. Th e qualit y of these vers io n s ma y dilfer from one winte r to ano th e r b y abo ut 5%. An exa mpl e of th e performan ce durin g a whole winter is give n in Figure 9. The average p e rc e n t age of cor r ect in te rpreta tion s (co n sid er in g onl y the degre e o f' d a nger ) durin g the wint e r of 1993 - 94 for th e DAVOS 2 vers ion was 56% . 329 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Joumal oJ Glaciology 7 "E 6 <Il 5 N <Il ---j ~ verificati_on -- DAVOS2 J --: -- : -- .r: 4 0 Ql 3 Ql 0,2 Q) "0 1 31.12.93 30.01 .94 1.03.94 31.03.94 30.04.94 winter 1993-94 Fig . 9. The degr ee of haz ard projJosed b), the D A r -OS2 lIlode! comjJared 10 lite veriJied degree of haz ard Jor Ihe winler 1993-94 il1 the D avos area. HowevfT, as can be see n in Fi g ure 9, b es id es the average perform a n ce, it is th e perform a nce in criti ca l situ a tion s o f, [or exa mpl e, increas ing or d ec reasin g h aza rd a t the beg inning or end of a sn o wfa ll peri od that is d ec isi,-e. nfort un a tel y, it mu st b e ad mitt ed th at th e performance in these situ a tion s is fa ir to poo r. The performance of th e other ,·ersions of th e DA VOS mod el is sli g htl y beller than th at of th e DA VOSI a nd DA VOS2 y ersion s. This fo ll ows fr om th e co n ce pt: split of the kn ow led ge base (DAVOS 3 1/32 ) a nd only one output res ult (DA VOS4 ) . Th e p e rform a nce of the DA VOS 3 1 ve rsion is eve n ve r y goo d , abo ut 750/0 . H owe,·er, th e perfor m a n ce of the DAVOS 32 vers ion is rather bad a t a bo ut 42%, probabl y du e to th e sm a ll er and mo re co mpl ex knowled ge base. The co mbin ed performance is abo ut 6 1 %. Th e DA VOS4 ve rsion that only predicts th e d egree of haza rd is on average correc t in 63% of a ll situ a ti o n s. This result r eprese nts th e best p e rform ance of a ll th e different ,-ersions of th e DA VOS model. H owe,-e r, co nsid e ring th e perform a n ce degree -b y-cl egre e, the r es ult is so m ewh at disillu sioning (T able 9) . Th e ave rage of63% of co rrect interpretations is th e res ult o f 76, 55 , 47 a nd 59% o f co rrect interpreta ti o ns for th e d eg r ees of ha za rd J, 2, 3 a nd th e d egrees 4 to 7 co mbin ed, respectiyely . I t is clear that th e interm edi a te degrees of haza rd are th e most difTicu lt to forecas t. In th e case of low or very hi g h haza rd , th e data a re more often unambiguous . Th e ex tremes arc easier to d ec id e. H OII·e,·er, since the ex treme Table 9. Detailed pe/formanee oJ Ih e D A VOS4 version: jJrognostic degree oJ Ita z ard compared 10 verified degree oJ haz ard. D egrees 4- 7 ( 7 ne ver o CCllrred ) are condensed. All !line winters ( 1361 siluations ) are considered Degree of Verified Correctness ha;::ard % Prognostic 2 3 4 ... 7 I 464 122 22 2 76 2 8 1 280 137 15 55 3 8 80 102 26 47 4 .. . 7 0 2 7 13 59 Total 63 330 e, "( nts a t the upp er ma rg in of the sca le are rare , the correctness is n o t a ll tha t good for eith er of th ese d egrees of hazard . Th e exa mpl e sho\\·s quite clearly that probably a ll ge nera ll y statis ti cs -b ased model s usi ng a d ata base with real situatio n s, a rc not ab le to predict r are situ a ti ons co rrec tl y, sin ce these so rts of situ a ti o n a re too r are. The sys tems a lw ays te nd to foll ow the types of situati on th a t fo rm the majority of the data base. MODUL lDodel Generally, th e results of th e !\,fOD U L model arc better th a n th ose of the DA VO S model. Thi s foll ows from the deterministic concept, more input parameters, es pecia ll y o n th e snow cm-er, a nd much more kno wlcd ge in the fo rm of th e stru c ture (sub-prob lems) a nd of th e impli cit rul es. N ow , we a lso h ave three winters o f expe ri ence . Durin g thi s tim e, the mod el has been co ntinu o usly improved, e.g . the calc ul ation of cert a in input parameters has be en c h a nged. So the performance has become bet te r. As the m ode l run s interactively , th e expe rt may sli g htl y influ en ce the res ult during its op erati on a l use. Thus , the performance g i,·en be low may n o t b e quite co mp arab le to th e more rigorous ly determined p e rfo rm a nce of th e D A VOS model an d might be sli ghtl y too optim ist ic. Th e ave rage performance durin g the las t three winters was 73% co rrect interpre tations , i. e. th e propose d d egree of haza rd d id not d eviate fi-om the ver ifi ca tion. All d ays were inte rpreted , i. e. th e r es ult I/O illll'lprelalioll did not occ ur. D e, ·iat ions of more th an onc degree of hazard are rare, in less than 2% of a ll si tu at ions. An examp le of th e perfo rm a n ce durin g a whole winter sea son is giv en in Fi gure 10 . The model fo llows quite exact ly the " e rifi ed degree of haza rd a nd also in tim es of in creasing or d ec reas in g hazard. 7,-----0=====================,------, "E 6 gJ 5 <Il .r: 4 o Ql 3 Ql 0,2 Ql "01 o 1.12 .93 -~ verification -- MODUL 31 . 12.93 30.01.94 1.03.94 31 .03.94 30.04.94 winter 1993-94 Fig . 10 . Th e degree of hazard jJl'oposed b} the i\I/OD UL model comjJared 10 l/ie verijied degree oJ hazard Jor Ihe winler /993-94 ill Ihe Davos area . Experience s h o\\'s that the l\IODUL mod e l is more sens ili ,·e to single input parameters . A w ro n g input parameter or a wrong deci sion in o n e of the sub-prob lems may h a,·e sub st a nti a l conseque n ces in the cn d , i.e . sometim es eve n a ch ange in the degree of h azard o f one or tw o step s. This is espec ia ll y due to th e sm a ll e r numb er of inpul p a r a m e ters treated at o n ce in a sub-probl em, due a lso to the fac t th a t th e outpu t r esu lt ofa sub-probl em is o ft en used aga in as input in anot h er sub-probl e m , a nd partly due to the fact that the input d ata are st ri c tl y categorize d . Th e latter prob le m might be remm·ed by introdu cing fuzzy logic, i. e. defining blurred ca tegor ies . Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Scltwei;:er and F6'hn : Avalanche forecasling - all erjJert s),stem ajJ/Hoar/z COInparis on of different lDodels Beca use in th e las t \"inte rs a n ea r es t-n eig hbo urs a l 'a la nch e- for ecas tin g model was used b y the al'a la n c h e -w a rnin g sec tio n a t SF I SAR, it is poss ibl e to co mpa re o ur mod els to thi s type o f m o d e l, call ed NEX_yIOD (1'.I e ist e r , 1991 ) that was d evelo p ed from the NXD m od el (Bu se r a nd others, 1987 ) , Th e o utput res ult of thi s model is, b es id es th e obse n 'ed a\'a la n c h es of th e te n n ea res t days loo kin g bac k 10 yea rs, th e a I'erage of th e for ecas ted d eg ree o f h aza rd of th ese ten n ea r es t n eig hbouring d ays, in co ntras t to o ur m od els th a t use the I"C rifi ed d egr ee o f haza rd, Nevertheless, thi s type of o utput is th e o nl y o n c ava il ab le fo r co mp a ri so n \I'ith o ur I'erified d eg ree o f h aza rd , Th e p erce nta ge o f co rrec t il1l e rpreta ti ons of th e NEX_:\ IOD m o del durin g th e las t three winters ( 1991 -92 to 1993- 94 ) was 38- 47%, d e pe ndin g o n h ow th e average d eg ree o f h aza rd is ro und ed, H owel'er, sin ce this operationa l ne a r es t-n e ig h- bo urs sys te m n eed s slig htl y diffe re nt input pa r a m e te rs, thi s res ult does n o t m ea n th a t th e n ea res t-n eig hb o urs m e th od is not as good as o ur m e th od, It r e prese nts just a co mp a ri so n be twee n sys t e m s th a t gi\"es th e sa me o utput r esu lt : th e degree o f h aza rd, Fi g ure II is a co mpari so n o f differe nt fo r ecastin g m ode ls a nd th e effe c ti ve fore cast in g with th e I'e rifi ed d eg ree o f h aza rd for th e Davos area during th e last three w inters ( 199 1 92 to 1993-9'~ ) , Th e in creas in g p e rform a n ce of th e DA VOS I , DAVOS 2, DA VOS'~ a nd iVIODUL m od els fo ll o 'lNs fr o m th e concept : th e m o re num e rou s and m ore stru c tured th e input para m e te rs or th e less numerou s th e o utput parameters. th e b e tt e r th e perfor m a n ce. Th e tr a ns i ti o n to th e fiv e -d eg r ee haza rd sca le \I'o uld in c rease th e p e rfo rm a n ce o f a n y o f th e m o d e ls bl' proba bl y less th a n I %, sin ce silU a ti o ns with a hi g h d eg ree of h aza rd a re ra re. > u c Q) ::1 C' ~ -Q) > ''-:; I1l ~ 0 .8 1 0 -30 -2 -1 . 0 G'l1 02 0 3 0.6 0.4 0 ,2 0 J[-J~ -- ~ J h j R In. ~ ~ I NEX. M OD NEX. MOD DAVOS 1 D AVOS2 DAVOS4 MODUL BULLETIN rounded Fig , 11 , CO llljJ arisoll of the jJeljo1'lI7allce of the statistical forecas t model J'v,/!'X_ilIO D , of 01l1'fOlll diffe ren t forecast models DIII 'OSI. DAVOS2, DA l 'OS4 al/d MOD UL, and q/lhe /Jll blic lI '({millg BC LLET!. \' dl/ril/g the thlee l(lillters 199 1- 9210 1993-9-1, Th e relative freqlle /l C)1 of tlt e del'ialioll from tlte verified degree of ha::.ard ill th e DOl'OS area 1.\ gIVen, 7. CONCLUSIONS 1' ll e C\1BE I~ rl'EK-COGENSYST~ 1 ' I I '\. JUc g m e n t pro cesso r, foll ow in g th e id ea of indu c til'e d ec ision-mak ing, prOl'ed to be use ful so ftw a r e fo r d el 'e lo pin g spec ifi c a ppli ca ti o ns in th e fi eld 0 (' ava la n c h e-h aza rd assess m e nt. Us in g weather, sno\l' a nd snow-cOl'er d a ta as input parameters, th e d e,'eloped mod els el 'a lu a te th e a\'a lan c h e haza rd fo r a g i,'e n reg ion, Th e n e\l' f'ca tur es a re th e c h o ice o f ela bora te input parameters, es p ec ia ll y more sn ow -cOl'e r d ata, th e non- lin ea r ca tego ri za ti o n o f th e input data, lh e s p ec ifi c a lgo rithm for th e sea rc h fo r simil a r situ a ti ons a nd fin a ll y th e co ncise o utput r es ult. Th e ava la n c h e h aza rd is d escrib ed as d eg rce of h aza rd , a ltitude a nd as pect o f th e m os t e nd a nge red slopes, for th e first tim e acco rdin g to th e sca le used in the fo recas ts, Thi s so rt of o utput res ult is m os t e ffi c ie nt for th c purpose of a l'a la n c h e fo rccast ing; it is mu c h m o re ap prop ri a t c to th e prob lem th a n, for exa mpl e, th c o utput "al'a la n c h c / non- antl a nch e d ay" , Th e use o f o b- se n 'at io na l a l'a lanc h e d a ta a lo n e is in suffi cie n t fo r bot h fo r ecas tin g an d ve rifi ca ti o n, Th e g i I'en o u t pu t resu I t is p ossibl e due to th e effo rt of pe rm a n en tl y I'erifyin g t h e ava la nche haza rd, V erifi ca ti on is thc m os t st rikin g feat ure a nd co mpl e tes th e d a ta se t - al th e prese nt time nin e w inte rs o f wea th e r , snow a nd sn o w- cOl'er d a ta with a co rrespo ndin g l 'C' rifi ed de g ree o f h aza rd - pro babl y a u niq ue se ri es, Th e sn O\l'-CO I'e r d a t a p ro l'ed to be I'e r y import a n t. Actually, it is well knOlI'll th a t a l 'a lanc h e fo recas tin g d e p ends strong ly o n th e sta te o f th e s n o \l' cOl'e r. H o w e l 'e r , a p a rt fr o m th e Fre n c h i\IEPRA m o d e l, until now h a rdl y a n y of th e present m o d e ls halT t a ken in to acco unt this o b v io us fac l. fl l cC lun g a nd T\\' ee d y ( 1994 ) introdu ced th e sn o\l' (,01,(, 1' so m e h ow impli c itl y in th eir mod e l b y co mbinin g the es timat es o f' th e m o del a nd o f the expert. Of co urse, thi s so rt o f dat a is n o t eas il y ava il a bl e but it is a n illu sio n to expe ct th a t a supportin g too l without a n )' sn o\\' -col'(' r d a t a is as p owe rful as th e expe rt for ecas t e r, Th e present-day m e teo ro logy pl ays a n im portant ro le but m os t o f the tim e it is n o t th e d ec is il'C o n c, Th e int eraCl il'e u se of th e m o d e ls pro\'ed to b e a s ubs ta nti a l advantage a nd es pec ia ll y th e \IOD U L m o d e l is I'(' r y in stru Cl il'C , It is I'er)' a pprop ri a te fo r the tra inin g o f juni o r fo recas le rs \I'ith a ce rtain b as ic kn o\l'led ge , Th e m o d c l run b y a juni o r fo recas te r m ay ac hi el'e a bout th e sa me performan ce (a bout 70% ) o n averagc as a se ni o r ex p e rt fo rccaster u s in g th e cO lll 'e n ti o n a l m eth od s . Th e DA \ 'OS m od el, a dat a -b ased mode l. and the flIOD U L model, a co mbin ed d a ta - a nd rul e-b ased m o d e l, h al,(, ac hi cI'('d a pcrformancc of a b o ut 60% a nd 70- 75%, r es pec ti l'c ly, Th e r e ex ist no co mp a rab le o r similar r es ults, b ased on a lo n g -t e rm operat io n a l test, o f a n y diffe r e nt sys te m for th e forecasting of th e reg io n a l al'a la nc he h aza rd, H OII'CI 'cr, th e p e rform a nce of a sys te m ca nn o t be full y d esc ribed b y a n al'erage pcrce nt age o f co rrec t int e rprct- a ti o ns co mp a red to ve rifi ca ti on ; th e pe rfo rm a nce in c riti ca l situ a ti ons is d ec isil'e , In such situ a ti o ns, a ll prese nt m o d e ls a re still no t good e n o ug h, In el'a lu a tin g the pcrfo rm a n ce of o ur sys tems, lI'e hav e bee n quite ri go ro us. Of co urse, a d ev ia li o n of o n e deg r ee o f haza rd d oes no t m ea n th e sa m e in a ll situ a ti o ns; it d e pe nd s o n th e d eg r ee of' haza rd a nd o n the Li irec ti o n of th e d el 'ia ti on, H owe l Tr, in co ntra st t o th e pre fe rence of so m e a\'a la ncht: -fo rc cas lin g se n 'ices , we think th a t avalanche warning is onl y e ffi c ient and fa ir in th e lo n g term ifn o m a rg in o f se c ur it )' is in c lud ed in th e fo recas t. W e a dmit th a t, in a sp ecifi c criti ca l situ at io n , a warn in g that is, for exa mple , o nc d egr ee abol'c th e la tt e r o n I'e rifi ed h azard m ay help prevent accidents , But , if th e fo recas ted d eg r ee is usu a ll y LOO hi g h , th e lI'a rnin gs become in effi cie nt a nd th e warning sr n 'ice I,,-ill soo n lose its c redibilit y , S o, in co n c lu sio n, liT co nsid e r a c!el'ia ti o n o f o nc d egree e ith e r 33 J Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ournal oJ Glaciology up or down as a \\To ng d ec ision , ind ependentl y of th e d eg ree of haza rd . Th e nex t step in th e d e\'elopm e nt will be to appl y th e mod els in dilTerent region s to assess th eir p erforman ce. Addition a ll y, several of th e sub-p roblems will b e furth er improved and it is also planned to det erm ine th e a ltitud e and th e as p ec t of th e most d a n ge rous slo p es in the MODUL model. Th e corres ponding sub-probl e m s still h ave to be developed. Fin a ll y, th e hazard of wet-snow ava lanc h es in spring tim e will be tak en into acco unt more sp ec ifica lly . The MODUL mod e l co ntain s g rea t potenti a l [or future developm ents. ACKNOWLEDGEMENTS W e thank th e form er direc tor of o ur in stitut e, C. Ja cca rd , for hi s enco uragement at th e start of this stud y. U . Guggis- berg, hea d of I NFEXPERT th a t se ll s th e so ftwar e used , promot ed th e work with co nsid e ra bl e enthu siasm. Variou s memb ers a nd tra in ees from SFISA R assisted durin g fi eld wor k, data management and o peration al use. \ Ve a lso th ank C. Pili s, who co ll aborated in th e ea rli es t ph ase of th e proj ec t. \\'e are grateful to the two anonymo us re\·iew e rs and , in pa rticular , to th e scientifi c editor D. M.l\,I cC lun g; th eir criti ca l co mments helped to improve the p a per. REFERENCES Ar mstro ng, R. L. . E. R. LaC hapcl lc. i\f.J . BOI·is and J. D. h -cs. 1974. Daeio/)!lIflll oJ lI1elhodologr Jor emlualioll alld /Hediclioll of m'alallche IlG~ard ill Ih e SIIII JIIIIII 1II0ulllaill I/rell of soulhu.'"lem Colorado . Bo uld e r. CO, Un i\'C rsit )' of Co lo rado. Institute of Arctic and Alpine R esearch. INST. \ ,\R 14 -06 -7155 -3. ) Bo is, P .. C. Oblcd a nd W . Good . 1975. \! ulti\'a riat e data analysis a s a lOo l for d av -b y-d av avala nche fo recas l. IlIlemllliolllll Associal ioll 0/ H)'flrological Sciellces Publicalioll 114 (Symposium a t Grindcl\\'al d 197+ Sno\\' i\l ec hani cs ), 39 1 403 . Bo lognes i, R. 1993 a. Artificial intrlli gc ll cc a nd loc al anl lan e hc fo r ec a stin g: the sys te m A \ ·A LOG. III Proceedillgs . I lIl efl/llliolllll ElIIl'I'gellSJ' IIl/d E ngineerillg ConJerellCl' , . Irlillgloll. 1':1. Soc. of Computer Scie nce (S.C. S. ) , San Diego, CA , 1 13- 11 6. Bolog nes i. R. 1993b . :"IX-LOG: idees di recrrices . . \ 'eige; el :it'lI lallches 63, 22 2+. Brun. E .. E. \f art in , \' . Sim on, C. G endre and C. Co leo u. 1989. An e ne rgv a nd m ass model of sno\\' CO \'c r , uitable for ope ra ti ona l a\'a la nchc forecas tin g . J. Glaciol.. 35 ( 12 1 . 333- 342 . Brun . E" P. D avi d , \1. Sudu l a nd G. Brun o l. 1992. A numerical mode l lO simul a te sn ow-eOl'e r stral ig r ap h ) for operat io n a l a\'a la n che fo recast in g. J. Glaciol .. 38 ( 128 ). 13 22. Busc r , 0. , P. Fo hn, \\' . Good, H. Gublcr and B. Sa lm. 1985 . Diffe rent method s fo r th e a ssess mclll of ava la nc he dan ge r. Cold Reg . Sci. Tee/mol., 10 (3), 199-218. Bu se r , 0. , ~1. BLid er a nd \ \'. Good. 1987. Av a la nche fo recas t b y neares t neigh bour m e thod. II/ternllliol/ol Associolioll oJ I-/ )'drologicol Scieuces Publi({lliol/ 162 (Symposium a t D m'os 1986 - .J l'lIlal/che Forlllalioll . .llol'elllenl ol/d E;fJecls ), 557- 569. D U DE:"I I nformatik. 1988. EI/ C)'clojJaedia. ~I annheim , etc .. Dud ell\·e rl ag. Durand. Y .. E . Brun, L. ;\Ierind ol, G. Guyo marc·h. B. Lesa ffr e a nd E. ~ [ a rtin . 1993. A meteoro logica l estimation of rele\'a lll parameters for sno\\' models. AIIII . Glaciol .. 18, 65- 7 1. Fo hn , P. 1985. Das sc h\\'C izerische Law in e nbull et in - ei ne Illlerpre tatio n- shilfc fUr den BenLilzc r. Eidg. 1101. Se/lllee- ulld LIlIt'il1ell/orsch. '\I ill. 38. Fo hn , P. a nd P . H aec hl er. 1978. Prcv ision d e grosses a\'ala n eh es a u moye n d\1Il m od ele d etermini s tc-st a ti stiqu e. III Deurieme Rel/conlre IlIlernatiol/ole sur la N eige elles Avalanches . 12- 13 et /-Ianil 1978, Crenoble. Fral/ce. COIII/lles Relldus. Grenob le, Assoc ia ti on :"I a ti o na le pour I'Etud e d e la Nei ge et d es Ava la nches, 15 1 165. Fo hn , P. \\. B. a nd ]. ScJlII'cizer. 1995, \ 'e rifi cation of a\'ala n c he d anger wit h res pec t to ""a lan che fo recasting. III Sil'a rdi ere , F. , ed. The cOlllribulioll oj.lcielllijic resellrch 10 sllle!)' u:ilh SIlOIt' . ice olld amlallche . Atles de colloque. C/1II1110Ili.1 30 lIIai 3 juill 1995. Gre noble. Assoc iati on :"Iationale po ur J'Etud e de la :\' eigc er dcs ,\ \'a la nches AXE~A. 151 156. Fo hn. P .. W. Good. P. Bo is and C. Obled. 1977. E\'a luation and co mp a ri so n of stat isti ca l a nd cO I1\'C llli o n a l method s of fo recas ting ava lanche haza rd. J. Glaciol .. 19 f8 11, 375-387. G ira ud , G. 1991. ~IEPRA: m od clc ex pert d'aid e a la prh'ision du ri sque d 'a \·a lanc hes . III .. Icle.; du D'IIl/JOsilllll . . 1.\E,\ ·. I-CIS. /- I/t·.J/? , 4- 8 jn ill 19.9 1, Challlollir . Frallce. G re n oble, l\ ssoc ia ti o n :"I a ti onale po ur l' Etucle d e la N eige et d es i\\·ala n c hes . 248 254. Gi r a ud , G. , j. Lafe uill e and E. Pahaul. 1987. E\'aluation d e la Cjua li tc d e la p rc\'ision du ri sq u e d '<l\·a lanc h e. IlIlenwliolllll .'/ssocia lioll DJ H),drologiwl Sciellets Publica lion 162 (Symposium at D",'os 1986 . Il'alallche Formalioll . .If Ol'emelll alld EfJeclsl . 583- 591. Gub le r. H . a nd H. -P o Ba d e r. 1989 .. \ model of initi a l fa ilure in sla b- a \'a lanche rel ease . . ·11111. Claciol .. 13,90- 95. Gu),omarc·h. G. a nd L. \ICr indol. 1994·. Qu e f' lLll -il sa yoir s ur ASTRAL' . \ 'eiges el A mlal/ches 66, 2 1- 25. J ucl so n, A. a nd B.] . Eri c kso n. 1973. Prediclillg al'alallche illlensi!)' ji'Olll wea lher dalll : a sllll islical alla!) ,,-is. Fort Co llin s. CO. U.S. J) epa rtm e nt o r Ag ri c ulture. Fores t Service. Roc ky i\ [OLIIllain Fores t and R a n ge E xpe ri ment Sta tion. (R ese arch Pape r' R\!-11 2 .) J ud so n , A. and R . \1. Kin g. 1985 . . \ n ind ex of regiona l snow-pack sta bilit y ba sed on natural slab a\·a lan chcs. J. (;Ia [iol ., 31 ( 108 ), 67 - 73. LaC h apel le. E. R. 1980. Th e rundam e nt a l processes in co nve nti o n a l ava la nche fo recas tin g . J. Glaciol., 26 (94 ) , 75- 84. \ IcC lun g . D. \\. 1994. Comp ut er assisted a\'a lanche forecast in g . III Hip a !. K . \\'. a nd L. F ang, edJ . lochaslic alld slalislical melhods ill I~) 'drolog)' alld enl'irollllletllal fIIgilleerillg. Do nrecht , Ycriag: Kluw e r 1'01. -t , 301 7- 358 . lI/ cC lun g . D. \1. 1995. C o mput er ass istance in a\'alanc hc foreca stin g . 111 Proceedill.gs. IlIlmwliollal SIIOIl' Sciellce I I'orksho/!, 30 Oclober 3 . \ '01'ell/ber 199.f, SlIoll'bird, Sail Lake, Clah. 3 10 3 13 . i\lc C lun g, D. ~L and P . A. Scha crer. 1993. Th e {I/'{t!al/che halldb ook . Seat ti e, \\'A, Th e i\I o untain ecrs. ~[ cClun g, D. lIL and J. T\\' eed y. 199+. Nume ri ca l ,l\'alan che pred ic tion: Kootenay Pass, British Columbia , Canada. J. (;/II(iol. , 40 135 ), 350 358 . i\ I e is te r , R. 1991. Grundla ge n der sc h\\' eize ri sc hen La\\inen\\'arn un g. l ·ermes,<IIlIg. Phologrollllllelrie. /t 'lIllllrlechllik, 3 (91 ). 88-95. \ /Cr inclol. L. 1995. An a logou s mod el s for ,n-a la n c hc ri sk forecas tin g. III Brug nol. G. , ed. ll cles de r CI/ i/'e(sil! Ellro/leenlle Sill' les risque.; 1I11111reh lIeiges et IIl'alllllches . 14-25 sejJlelllbre 1992. Clw1I10lli\. Frallce. Grc n oblc, CE i\l:\GREF Editions, 293- 299. :"I a n ll'l'e, j. P., G. Gu)'omarc ' h and G. Giraud . 1987. en mocl e lc ~ tati~tiqllc pour la p r{y is ion loca le des a\·ala nch es. Illternational . J.rsocialioll oJ H l'drological S(iell(e.l Publiralioll 16 2 (Symposiu m at Da\'05 1986 - Avalallche Formatiol/ . .I/ovemenl alld EffeCIS ), 57 1- 580. Oblcd , C. a nd \\' . Good. 1980. Rece nt development s of ava la n c h e fo r ecas tin g b\· di sc rimin a nt a na lys is tec hniqu es: a methodol og ica l re\·icw a nd so me app li ca tion s lO th e Parse nn area (Da vos , Swit ze r- la nd ) . J. Glaciol .. 25 (92 ) . 3 15- 3+6. Pe r la, R. I. 1970. On con tri butory faclO rs in al'a lanche h aza rd eva lu a ti on. Can . Ceolech.]., 7 (4 :. +14- 4 19 . R e mund , J . 1993. \ 'c rifikati o n der reg io n a len La\\' in enge fa hre n p rog- n ose. I Dipl omarbeit. Eid ge n oss isc he T ec hni sc he Hoc hsc hul e, Zuri c h . Geog raphisc hes Inst it ul. ) Sa l",ay, A. A. 1976. Stati sti c a l es tima tio n a nd prediction 01' ll\'a la n c h e activ it y from meteo ro log ica l data for th e R oge rs Pass area of Briti sh Co lumbia. ( Ph.D. thesis. U ni\'e rsit y of Briti sh Columbia. ) Sc ir weize r , j. 1993. Th e influ cn ee of th e laye red c h aracter of sn ow cO\'C r o n the tri gge ring o f sla b a\·a lanc hcs .. 11111. Glaciol., 18, 19 3 198. Schweiz er, \\. , P. \1. B. F o hn,.J. Sc hwcizc r a nd A. C1tsch. 1994a. A h yb rid expert sys tem for 3\'a la n ch e fore cas tin g. III Schenler, \\'., B. Sc hmid , A.1\1. Tjoa a nd H . \\' enhn er, eds. IIIJormatioll alld COlllll1l1l1i((llioll.1 Il'cI/IIologies illlollrisll1 . :"Ie\\' Yor k, etc., Springc r-\' crlag, 148 153. Sc hll'e izer, \1 ., P.1I r. B. Fo hn a nd J. Schweizc r. 199 +b . Int egra ting neura l net\\'orks and rule based systems to build an a\'alane he forecas ting sys lem. III Hamza, r.f. H ., ed .. Irlijicial IlIlelligellce . EI/ml S)'Slems alld . \ 'eurolllll .\ eluwks. Proceedillgs I. ISTED IlIlemaliollal ConJerell ce . " (j JII!r 1.994. <Ii'rich . Su:il;:.erlalld. I. \ STE D Illl ern a ti o n a l Assoc ia tion o f Sc ien ce and T ec hn ology for De\'e lopmen t) Acta Press. I 4. Sc hw e ize r. J.. P. Fohn an d C. PILlSs. Unpub li shed. COGENSYS jud gement processo r (P a r a d ocs ) a ls Hil[smill el fUr di e L aw in e n wa r- nun g. \ \'c iss fluhjo ch jD a\'os, Eid ge nbss isc h es Insti tut fur Schn ee - und L a winenforsc hun g . ( I nr c rn er Bcri ch t 675, 1992. ) \\ 'e ir, P. and D. ~f. ~I cClung . 199+. Computer based a rtifi cia l int ell ige n ce as an a id w BC ~[oTH's aya la nche forecas te rs. :i l'lIlallche. \ fl"s, 43 ,2-4. ,vIS received 21 Sept ember 1994 alld acce/lled ill revisedJorm 29 November 1995 33 2 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Vol 42 Issue 141 page 318-332 - Avalanche forecasting - an expert system approach - Jürg Schweizer and Paul M.B. Föhn 
work_ifkk7qcmhbcghln4wiacd2doki	----	PII: 0888-613X(88)90140-5 334 Abstracts UCLA, Los Angeles, California 90024-1600 An inductive logic can be formulated in which the elements are not propositions or probability distributions, but information systems. The logic is complete for information systems with binary hypotheses, that is, it applies to all such systems. It is not complete for information systems with more than two hypotheses but applies to a subset of such systems. The logic is inductive in that conclusions are more informative than premises. Inferences using the formalism have a strong justification in terms of the expected value of the derived information system. Explanation of Probabilistic Inference for Decision Support Systems C h r i s t o p h e r E l s a e s s e r MITRE Washington Artificial _Intelligence Center and Department o f Engineering and Public Policy, Carnegie-Mellon University, Pittsburgh, Pennsylvania 15213 This paper reports work in progress on an explanation facility for Bayesian conditioning aimed at improving user acceptance of probability-based decision support systems. Design of the facility, which appears to be reasonably domain-independent, is based on an information processing model that accounts for both biased and normative behavior in reasoning about conditional evidence. Preliminary results indicate that the facility is both acceptable to naive users and effective in improving understanding of Bayesian conditioning. Automated Generation of Connectionist Expert Systems for Problems Involving Noise and Redundancy S t e p h e n I. G a l l a n t College o f Computer Science, Northeastern University, Boston, Massachusetts, 02115 When creating an expert system, the most difficult and expensive task is that of constructing a knowledge base. This is particularly true if the problem involves noisy data and redundant measurements. This paper shows how to modify the MACIE process for generating connectionist expert systems from training examples so that it can accommo- date noisy and redundant data. The basic idea is to dynamically generate appropriate training examples by constructing both a " d e e p " model and a noise model for the underlying problem. The use of winner-take-all groups of variables is also discussed. These techniques are illustrated with a small example that would be very difficult for standard expert system approaches. A Measure-Free Approach to Conditioning I. R. G o o d m a n Command and Control Department, Code 421, Naval Ocean Systems Center, San Diego, California 92152 
work_ifymfqj7jjbv5ldnhyofmazcy4	----	The role of idioms in sentiment analysis Expert Systems with Applications 42 (2015) 7375–7385 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a The role of idioms in sentiment analysis http://dx.doi.org/10.1016/j.eswa.2015.05.039 0957-4174/� 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). ⇑ Corresponding author. Tel.: +44 29 2087 0320; fax: +44 29 2087 4598. E-mail addresses: WilliamsL10@cardiff.ac.uk (L. Williams), BannisterCA@cardiff. ac.uk (C. Bannister), Arribas-AyllonM@cardiff.ac.uk (M. Arribas-Ayllon), PreeceAD@cardiff.ac.uk (A. Preece), i.spasic@cs.cardiff.ac.uk (I. Spasić). Lowri Williams a, Christian Bannister a, Michael Arribas-Ayllon b, Alun Preece a, Irena Spasić a,⇑ a School of Computer Science & Informatics, Cardiff University, 5 The Parade, Cardiff CF24 3AA, UK b School of Social Sciences, Cardiff University, King Edward VII Avenue, Cardiff CF10 3WT, UK a r t i c l e i n f o Article history: Available online 28 May 2015 Keywords: Emotion recognition Sentiment analysis Natural language processing User-generated content Tagging a b s t r a c t In this paper we investigate the role of idioms in automated approaches to sentiment analysis. To estimate the degree to which the inclusion of idioms as features may potentially improve the results of traditional sentiment analysis, we compared our results to two such methods. First, to support idioms as features we collected a set of 580 idioms that are relevant to sentiment analysis, i.e. the ones that can be mapped to an emotion. These mappings were then obtained using a web-based crowdsourcing approach. The quality of the crowdsourced information is demonstrated with high agreement among five independent annotators calculated using Krippendorff’s alpha coefficient (a = 0.662). Second, to evaluate the results of sentiment analysis, we assembled a corpus of sentences in which idioms are used in con- text. Each sentence was annotated with an emotion, which formed the basis for the gold standard used for the comparison against two baseline methods. The performance was evaluated in terms of three measures – precision, recall and F-measure. Overall, our approach achieved 64% and 61% for these three measures in two experiments improving the baseline results by 20 and 15 percent points respectively. F-measure was significantly improved over all three sentiment polarity classes: Positive, Negative and Other. Most notable improvement was recorded in classification of positive sentiments, where recall was improved by 45 percent points in both experiments without compromising the precision. The statistical significance of these improvements was confirmed by McNemar’s test. � 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http:// creativecommons.org/licenses/by/4.0/). 1. Introduction The proliferation of user-generated content (e.g. product reviews) on the Web 2.0 provides opportunities for many practical applications that require consumer opinion (e.g. market research) as an alternative or a supplement to more traditional qualitative research methods such as surveys, interviews and focus groups. However, the sheer scale of text data acquired from the Web poses challenges to qualitative analysis. Text mining has emerged as a potential solution to the problems of information overload associ- ated with reading vast amounts of text originating from diverse sources. In particular, sentiment analysis (or opinion mining) aims to automatically extract and classify sentiments (the subjective part of an opinion) and/or emotions (the projections or display of a feeling) expressed in text (Liu, 2010; Munezero, Montero, Sutinen, & Pajunen, 2014). Most research activities in this domain have focused on the problem of sentiment classification, which classifies an opinionated text segment (e.g. phrase, sentence or paragraph) in terms of its polarity: positive, negative or neutral (e.g. Aue & Gamon, 2005; Bethard, Yu, Thornton, Hatzivassiloglou, & Jurafsky, 2004; Breck, Choi, & Cardie, 2007). Features used to support sentiment analysis include terms, part of speech, syntactic dependencies and negation (Pang & Lee, 2008). Most commonly, opinionated words that carry subjective bias are used in a bag-of-words approach to classify opinions (e.g. Attardi & Simi, 2006). Opinionated words can be utilized from lexicons such as SentiWordNet (Esuli & Sebastiani, 2006), WordNet-Affect (Valitutti, Strapparava, & Stock, 2004) and NRC word–emotion association lexicon (Mohammad & Turney, 2010). Dynamic calcu- lation of word polarity (or semantic orientation) based on its statis- tical association with a set of positive and negative paradigm words is an alternative to predefined lexicons of opinionated words (Turney & Littman, 2003). Other features explored in senti- ment analysis include more complex linguistic models based on lexical substitution, n-grams and phrases (Dave, Lawrence, & Pennock, 2003). Using an n-gram graph based method to assign sentiment polarity to individual word senses, experiments implied that figurative language (i.e. the language which digresses from lit- eral meanings) not only conveys sentiment, but actually drives the polarity of a sentence (Rentoumi, Vouros, Karkaletsis, & Moser, 2012). Although the value of phrase-level features in sentiment http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.05.039&domain=pdf http://creativecommons.org/licenses/by/4.0/ http://creativecommons.org/licenses/by/4.0/ http://dx.doi.org/10.1016/j.eswa.2015.05.039 http://creativecommons.org/licenses/by/4.0/ mailto:WilliamsL10@cardiff.ac.uk mailto:BannisterCA@cardiff.ac.uk mailto:BannisterCA@cardiff.ac.uk mailto:Arribas-AyllonM@cardiff.ac.uk mailto:PreeceAD@cardiff.ac.uk mailto:i.spasic@cs.cardiff.ac.uk http://dx.doi.org/10.1016/j.eswa.2015.05.039 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 7376 L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 analysis has been acknowledged (Socher et al., 2013; Wilson, Wiebe, & Hoffmann, 2009), few approaches have extensively explored idioms as features of this kind (e.g. Thelwall, Buckley, & Paltoglou, 2012). Nonetheless, the error analysis of sentiment clas- sification results often reveals that the largest percentage of errors are neutral classifications when no opinionated words are present or when idioms are used to express sentiment (Balahur et al., 2010). Idioms are often defined as multi-word expressions, the mean- ing of which cannot be deduced from the literal meaning of con- stituent words, e.g. the idiom a fish out of water is used to refer to someone who feels uncomfortable in a particular situation. To distinguish idioms from related linguistic categories such as for- mulae, fixed phrases, collocations, clichés, sayings, proverbs and allusions, the following properties need to be considered (Nunberg, Sag, & Wasow, 1994): 1. Conventionality: Their meaning cannot be (entirely) predicted from the constituent words considered independently. 2. Inflexibility: Their syntax is restricted, i.e. idioms do not vary much in way they are composed. 3. Figuration: Idioms typically have figurative meaning stemming from metaphors, hyperboles and other types of figuration. 4. Proverbiality: Idioms usually describe a recurrent social situation. 5. Informality: Idioms are associated with less formal language such as colloquialism. 6. Affect: Idioms typically imply an affective stance toward some- thing rather than a neutral one. The last property emphasizes the importance of idioms in sen- timent analysis as it implies that an idiom itself may often be suf- ficient to determine the underlying sentiment. There are two requirements for idioms to be effectively utilized in sentiment analysis methods: (1) Idioms need to be recognized in text, and (2) the associated sentiment needs to be explicitly encoded. The inflexibility property (see property 2 above) makes the first requirement feasible. Lexico-syntactic patterns can be used to model idioms computationally and recognize their occurrences in text. A lot of the idioms are frozen phrases such as by and large, which can be recognized by simple string matching. Syntactic changes such as inflection (e.g. verb tense change) are often seen in idioms (Yusifova, 2013). Such linguistic phenomena can be mod- eled by regular expressions, e.g. spill[s|t|ed] the beans. More com- plex idioms have variables for open argument places (Jackendoff & Pinker, 2005) (e.g. put someone in one’s place), which can still be modeled by means of lexico-syntactic patterns (e.g. put NP in PRN’s place) and recognized in a linguistically pre-processed text. Less often, idioms are ‘‘syntactically productive’’, i.e. they can be changed syntactically without losing their figurative meaning, e.g. John laid down the law can be passivized to the law was laid down by John while retaining the original figurative interpretation that John enforced the rules (Gibbs & Nayak, 1989). Transformational grammars have been suggested as a framework to handle more complex syntactic changes such as nominalization (e.g. you blew some steam off vs. your blowing off some steam) (Fraser, 1970). In Polish, a highly inflected language, idioms were recognized using a cascade of regular expressions and their effect on senti- ment analysis results was evaluated on a corpus of product and service reviews, where idioms were found to occur rarely (Buczynski & Wawer, 2008). In English, despite the obvious need for regular expressions, idioms are usually recognized using a lexicon-based approach, which can only recognize those idioms that are syntactically unproductive or frozen. For example, (Shastri, Parvathy, Abhishek, J., & R., 2010) used a dictionary of idioms (e.g. at a snail’s pace) in order to recognize them in text and map them to their abstract meaning (e.g. slow), which is then utilized to infer the sentiment. In another lexicon-based approach (Beigman Klebanov, Burstein, & Madnani, 2013), idiom recognition was further limited to 46 noun–noun compounds (e.g. glass ceil- ing). The use of their sentiment profiles was found to improve the performance of sentiment classification on a corpus of test-takers essays. Our own study aims to go beyond a lexicon-based approach to recognition of English idioms and use regular expressions instead. The added overhead of handcrafting regular expressions allowed us to explore a much wider set of idioms (beyond the low-hanging fruits) as part of sentiment analysis. Assuming that idioms can be identified in text automatically, we need additional knowledge about the underlying sentiment in order to utilize them as features of sentiment analysis. While idioms have been extensively studied across many disciplines (e.g. linguistics, psychology, etc.), thus far there is no comprehen- sive knowledge base that systematically maps idioms to senti- ments. This is the main reason why idioms have been underrepresented as features used in sentiment analysis approaches with few exceptions (e.g. Xie and Wang (2014) describe a set of 8160 Chinese idioms). Due to the subjective nat- ure of the problem, multiple annotations are required in order to either determine the prevalent sentiment associated with an idiom or use a fuzzy logic approach to represent the sentiment with a degree of truthfulness and falsehood. To support this task, a web-based crowdsourcing approach can be used to efficiently col- lect a large amount of information relevant for sentiment analysis (Greenwood, Elwyn, Francis, Preece, & Spasić, 2013). We used crowdsourcing to systematically map 580 English idioms to 10 emotion categories, which represents the largest lexico-semantic resource of this kind to utilize in sentiment analysis. The purpose of this paper is to study the effect of idioms on sen- timent analysis. The study was designed as follows (see Fig. 1): (1) collect a set of idioms that can be mapped to sentiments, (2) map individual idioms to sentiments in order to support them as fea- tures for sentiment analysis, (3) assemble a corpus of sentences in which these idioms are used, (4) annotate sentences in the cor- pus with sentiments in order to create a gold standard for senti- ment analysis, (5) implement a sentiment analysis approach that incorporates idioms as features, and (6) compare the evaluation results against a traditional sentiment analysis approach using the gold standard created in (4). 2. Data collection 2.1. Idioms Idioms pose considerable difficulties for English language learn- ers. Failure to understand idioms in context significantly affects one’s understanding of language in a variety of personal and pro- fessional situations (Nippold & Martin, 1989). It is therefore not surprising that most syllabi for English as a second language pay special attention to studying idioms (Liu, 2003). As a result, there is an abundance of teaching material dedicated to the study of idioms. In this study, we relied upon an educational web site – Learn English Today (Learn English Today, 2013), which organizes idioms by themes, many of which can be mapped to emotions either directly (e.g. Happiness/Sadness) or indirectly (e.g. Success/Failure). We focused specifically on emotion-related idioms, as these are anticipated to have a substantial impact on sentiment analysis. We selected 16 out of a total of 60 available themes, listed in Table 1 together with a number of associated idioms. A total of 580 idioms were collected. Fig. 1. An overview of the study design. Table 1 Distribution of idioms across themes. Theme Total Theme Total Anger/Annoyance 45 Mistakes/Errors 5 Anxiety/Fear 14 Politeness 8 Arguments/Disagreements 37 Problems/Difficulties 57 Enthusiasm/Motivation 10 Safety/Danger 27 Feelings/Emotions 48 Sleep/Tiredness 11 Fun/Enjoyment 22 Success/Failure 84 Happiness/Sadness 21 Surprise/Disbelief 16 Madness/Insanity 11 Violence 6 Table 2 Distribution of sentences across themes. Theme Total Theme Total Anger/Annoyance 261 Mistakes/Errors 31 Anxiety/Fear 88 Politeness 42 Arguments/Disagreements 232 Problems/Difficulties 360 Enthusiasm/Motivation 41 Safety/Danger 176 Feelings/Emotions 280 Sleep/Tiredness 64 Fun/Enjoyment 107 Success/Failure 519 Happiness/Sadness 128 Surprise/Disbelief 92 Madness/Insanity 47 Violence 50 L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 7377 2.2. Corpus The British National Corpus (BNC) (BNC Consortium., 2014; Leech, 1993) is a large text corpus of both written and spoken English compiled from a variety of sources. As such, it has been a corpus of choice for a great many studies in computational linguis- tics and natural language processing (NLP), including those focused on idioms (e.g. Grant, 2005). We also used the BNC to assemble a corpus of idioms used in context. We collected 2521 sentences that contain an expression that can be matched to an idiom. In most cases, this expression will have a figurative meaning associated with an idiom, but in some cases it will convey a literal meaning. In this sense, some of the sentences will be false positives. From a lexico-syntactic perspective, most idioms can be modeled with local grammars, but it is more difficult to automate their recogni- tion from a semantic perspective. Consider, for example, the fol- lowing two sentences extracted for expression in the bag, the figurative meaning of which is virtually secured: The Welsh farmer’s son had the 1988 conditional jockeys’ title already in the bag. I looked in the bag, it was full of fish. It is necessary to include false positives in the corpus in order to evaluate how incorrectly recognized idioms may affect the results of sentiment analysis. The BNC was searched using its Simple Search function avail- able online. It can be used to search the BNC for a word of phrase and returns up to 50 random sentences for each query. The BNC was searched for content words found in idioms and the returned results were manually matched to an idiom. A maximum of 10 sen- tences was selected for each idiom used in either a figurative or lit- eral sense. Search results were non-empty for a total of 423 idioms from the original list of 580 idioms. The mean and median average number of sentences extracted for an idiom are both 6, with stan- dard deviation of 3.39. Table 2 summarizes the number of sen- tences collected for each theme associated with idioms. Fig. 2. Annotation pane. 7378 L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 3. Data annotation 3.1. Annotation scheme Taking a long-term view of our research on sentiment analysis, we want to address the limitations of state-of-the-art approaches by focusing on a full range of emotions rather than merely senti- ment polarity. However, there is no consensus among researchers about how to define, measure or conceptually organize the range of feelings and behaviors that correspond to an emotion. The main tension in the literature is whether emotions can be defined as dis- crete, universal categories of basic emotions or whether they are characterized by one or more dimensions. Categorical approaches are usually theory-driven accounts that suggest basic emotions are the functional expression of underlying biological and evolu- tionary processes (Damasio, 1999; Darwin, 1872; LeDoux, 1996). This view is supported by empirical findings of cross-cultural stud- ies based on recognition of facial expressions (Ekman, 1972). Ekman claimed that there are six basic emotions – anger, disgust, fear, happiness, sadness and surprise – though later studies identi- fied additional emotions (Ekman, 1992). Dimensional approaches are often data driven accounts of emotions incorporating aspects of valence, arousal or intensity derived from empirical studies of self-reported experience (Russell, 2003). There is considerable variation among dimensional models, many of which incorporate either two or three dimensions (Rubin & Talarico, 2009). Having considered the literature on emotion classification, tra- ditional taxonomies arising from psychological research are too complex for our present task. We therefore opted to base our anno- tation scheme on the Emotion Annotation and Representation Language (EARL) Version 0.4.0, an XML-based language for repre- senting and annotating emotions in technological contexts (The Association for the Advancement of Affective Computing, 2014). It has been designed for a wide range of applications, including cor- pus annotation and emotion recognition. It organizes 48 emotions into 5 positive and 5 negative categories (see Table 3). To facilitate the annotation task, we used the 10 top-level categories as they provide a manageable number of choices for a human annotator, which are also evenly distributed between positive and negative polarities. For annotation purposes, specific emotions were used as examples instead of formal definitions to explain each top-level category, e.g. Caring encompasses affection, empathy, friendliness and love. 3.2. Annotation process We implemented a bespoke web-based annotation platform. It provides a simple interface accessible via a web browser, which eliminates the installation overhead and minimizes the need for special training. The interface consists of two panes. One pane con- tains randomly selected text ready to be annotated. In our study, this was either an idiom together with its definition or a sentence that contains an idiom (see Fig. 1 for study design). The other pane contains four annotation choices: Positive, Negative, Neutral and Ambiguous. The selection of either Positive or Negative categories expands the menu to provide additional choices (see Fig. 2). The Table 3 Top-level emotion categories in EARL. Negative category Example Positive category Example Negative & forceful Annoyance Positive & lively Joy Negative & not in control Helplessness Caring Love Negative thoughts Doubt Positive thoughts Hope Negative & passive Sadness Quiet positive Relaxed Agitation Stress Reactive Politeness category Neutral was introduced in EARL to allow the absence of an emotion to be annotated explicitly. We introduced the addi- tional category Ambiguous to annotate cases of emotion that can- not be determined as either Positive or Negative without additional information such as context, tone of voice, or body language. A help button is available next to each annotation choice to pro- vide additional information if needed. The overall annotation scheme can also be viewed in a separate window. Each annotation can be scored on a 3-point scale to account for the annotator’s con- fidence in a particular choice: Low, Medium or High, with Medium being a default choice. Online accessibility provided us with the flexibility of choosing the physical location for annotation experiments. Users were tracked by their IP addresses to avoid duplication of annotations, not to identify individuals. No other personal information was col- lected. This was explained in the privacy policy. All annotation results were stored securely in a relational database. Group annotation sessions were conducted weekly, where new annotators were briefed about the study and their role as an anno- tator. All annotators were required to be of native or native-like English proficiency. All annotations were performed indepen- dently. The data were randomized individually for each annotator, so they were always annotated in a different order. No discussions about particular data items were allowed among the annotators during the group sessions. Table 4 Distribution of annotations in the gold standard. Annotation Total % Example Negative 1219 48.35 All right, do not jump down my throat Positive 677 26.85 I shall go the extra mile Other 625 24.79 Your mother used to sleep like a log L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 7379 3.3. Annotation results The data was annotated over a course of 13 weeks. A total of 18 annotators participated in the study with 44% being female. A total of 2900 annotations were collected for all 580 idioms described in Section 2.1 with 5 annotations per idiom. A total of 8610 annota- tions were collected for all sentences in the corpus described in Section 2.2 with at least 3 annotations per sentence. A total of 143 sentences had a maximum of 5 annotations. Overall, the mean and median average number of annotations per sentence were both 3 with a standard deviation of 0.60. The main goal of this study was to evaluate the role of idioms in sentiment analysis by comparing our method to existing approaches, hence our experiments needed to conform to their sentiment classification scheme. Most sentiment analysis approaches output sentiment polarity, i.e. classify text as being positive, negative or neutral. This also applies to SentiStrength (Thelwall, 2014; Thelwall, Buckley, Paltoglou, Cai, & Kappas, 2010) and Stanford CoreNLP’s sentiment annotator (Socher et al., 2013), two state-of-the-art methods chosen as the baseline in our experiments. Therefore, to create a gold standard that can allow comparison against the baseline, we projected categories specified in Table 3 onto Positive and Negative polarity and merged Neutral and Ambiguous categories into a single category called Other. Nonetheless, we will be able to use the original annotations to re-train the machine learning method described here to support emotion classification against the categories described in Section 3.1 as part of the future work. After projecting all annotations onto the simplified classifica- tion scheme (i.e. Positive, Negative and Other), we used Krippendorff’s alpha coefficient (Krippendorff, 1980) to measure the inter-annotator agreement. As a generalization of known relia- bility indices, it was chosen because it applies to: (1) any number of annotators, not just two, (2) any number of categories, and (3) incomplete or missing data (Krippendorff, 2004). Krippendorff’s alpha coefficient of 1 indicates perfect agreement, whereas 0 indi- cates chance agreement. Therefore, higher values indicate better agreement. The values for Krippendorff’s alpha coefficient were obtained using an online tool for calculating inter-annotator agreement (Geertzen, 2012). The expected disagreement on the idiom dataset was calculated to be De = 0.606, the observed disagreement was Do = 0.205, which resulted in a = 0.662. The corresponding values calculated for the corpus of sentences were as follows: Do = 0.414, De = 0.643 and a = 0.355. The relatively high agreement (a = 0.662) on idioms alone illustrates that idioms can be reliably mapped to sentiment polarity. Significantly lower agreement (a = 0.355) on sentences where idioms are used in context illus- trates the complexity of sentiment interpretation, where a range of different emotions may be conveyed in a single sentence (e.g. They were too embarrassed or didn’t want to make a mountain out of a molehill) or the underlying emotion may vary depending on the context (e.g. His jaw dropped). 3.4. Gold standard Annotated sentences were used to create a gold standard for sentiment analysis experiments. Prior to calculating the inter-annotator agreement, additional annotations from a new independent annotator were sought to resolve any disagreements. The annotation process was finalized once each sentence had the majority of at least 50% across the annotations with all previous ties broken. A total of 282 additional annotations were collected in three iterations. Subsequently, each sentence with an annotation agreed by the relative majority of at least 50% of the annotators was treated as the ground truth. Table 4 shows the distribution of ground truth annotations across the three categories together with an annotated example from each. A random subset of approximately 10% was selected for testing (500 sentences) and the rest was used for training (2021 sentences). 4. Sentiment analysis 4.1. Idiom recognition In order to incorporate idioms as features of sentiment analysis, we require the means of recognizing them in text. For this purpose, we modeled each idiom by a local grammar (Gross, 1997). For example, the following grammar: hidiomi ::= hVBi hPRP$i heart on hPRP$i sleeve hVBi ::= wear | wore | worn | wearing hPRP$i ::= my, your, his, her, its, our, their was used to successfully recognize the idiom wear one’s heart on one’s sleeve in the following sentence: Rather than hidiomi wear your heart on your sleeve h/idiomi, you keep it under your hat. Idiom recognition rules were implemented as expressions in Mixup (My Information eXtraction and Understanding Package), a simple pattern-matching language (Cohen, 2014). The pattern-matching rules were applied to the test dataset of 500 sen- tences in which a single annotator marked up all idiom occur- rences differentiating between the figurative and literal meaning, e.g. Phew, that was a hidiomi close shave h/idiomi. He has polished shoes, a hnonidiomi close shave h/nonidiomi, and too much pride for a free drink. The following performance was recorded for idiom recognition: P = 94.44%, R = 100% and F = 97.14%, where an idiom was consid- ered to be correctly recognized if the suggested text span matched exactly the one marked up by the annotator. 4.2. Negation As with other phrases, the polarity of idioms can be changed by negation. For example, the polarity of the idiom jump for joy is pos- itive, but when negated the overall polarity can be reversed as illustrated by the following sentence: I didn’t exactly jump for joy. It is, therefore, essential to consider negation when using idioms as features in sentiment analysis. We implemented pattern-matching rules to recognize explicitly negated idioms based on clues such as negative words (e.g. no, never, etc.), negative adverbs (e.g. hardly, barely, etc.) and negated verbs (e.g. doesn’t, isn’t, etc.). A total of 27 negated idioms were annotated in the test 7380 L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 dataset. The following performance was recorded for the recogni- tion of negated idioms: P = 86.21%, R = 92.59% and F = 89.29%. Given a small number of negated idioms in the test dataset, a larger corpus is required in order to better estimate the performance of the negation module. 4.3. Feature selection Once idioms are recognized in text, we require additional infor- mation about their sentiment polarity in order to utilize them as features of sentiment analysis. Note that the polarity information was obtained as part of data annotation (see Section 3) when each idiom was annotated independently 5 times. After projecting the original annotations onto the sentiment polarity scheme, a total of 5 annotations collected for each idiom were used to calculate their feature vectors. Each idiom was represented by a triple (Positive, Negative, Other), where each value represents the percent- age of annotations in the corresponding category. For example, the idiom wear one’s heart on one’s sleeve received one Positive annota- tion, zero Negative annotations and four Other annotations. It was, therefore, represented as the following triple: (20, 0, 80). Further, as we wanted to investigate the impact of idioms as features in sentiment analysis, we conducted two experiments in which we combined idioms with the results of two popular senti- ment analysis methods: (1) SentiStrength (Thelwall, 2014; Thelwall et al., 2010) and (2) Stanford CoreNLP’s sentiment anno- tator (Socher et al., 2013). In Experiment 1, we used SentiStrength as the baseline method and used its output as features in our own method and combined them with those based on idioms. SentiStrength is a rule-based system that assigns sentiment polarity to a sentence by aggregat- ing the polarity of individual words, e.g. Input: The party is over. Analysis: The party [1] is over [�1] . Output: result = 0, positive = 1, negative = �1 As illustrated in the given example, we used trinary classifica- tion output as features in our method and converted them into a 3-dimensional vector: (0, 1, �1). In our approach, the phrase party is over would be recognized as an idiom, which was annotated and subsequently mapped to the following triple: (0, 100, 0) denoting that all annotators considered it to be negative. We appended the two vectors to create a single feature vector for the given sen- tence as follows: ð 0; 1;�1 |fflfflfflffl{zfflfflfflffl} sentiment polarity 0; 100; 0 |fflfflfflfflffl{zfflfflfflfflffl} idiom polarity Þ In Experiment 2, we used a sentiment annotator distributed as part of the Stanford CoreNLP, a suite of core NLP tools. This method uses recursive neural networks to perform sentiment analysis at all levels of compositionality across the parse tree by classifying each sub-tree on a 5-point scale: very negative, negative, neutral, posi- tive and very positive (see Fig. 3 for an example). In addition to classification, it also provides probability distribution across the 5 classes, which we have used as features in our method by con- verting them into a 5-dimensional vector: (4, 27, 46, 20, 3). As before, the idiom party is over would be recognized and mapped to its polarity triple (0, 100, 0) and appended to create a single fea- ture vector for the given sentence as follows: ð4; 27; 46; 20; 3 |fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl} sentiment polarity 0; 100; 0 |fflfflfflfflffl{zfflfflfflfflffl} idiom polarity Þ If an idiom was recognized to be negated, then we reversed pos- itive and negative polarities in the idiom polarity vector based on an assumption that negation typically converts positive to negative polarity and vice versa. For instance, let us consider the following sentence in which the negatively charged idiom party is over is negated: The party is not over yet. In this particular example, the negation changes the overall polarity from negative to positive. We modeled this phenomenon associated with negation by simply reversing positive and negative polarities, e.g. the original idiom polarity triple (0, 100, 0) would be converted to (100, 0, 0). When multiple idioms are recognized, then the associated idiom polarity values are aggregated by summing up the polarity vectors previously taking into account the effects of negation. For instance, two idioms were recognized in the following sentence: He hidiomi stopped dead in his tracks h/idiomi, hidiomi rooted to the spot h/idiomi with horror. Their polarity triples (0, 60, 40) and (0, 40, 60) respectively are summed up to obtain (0, 100, 100), which illustrates how the neg- ative polarity is boosted from 60 and 40 to 100 in this particular case. In our corpus of 2521 sentences, 29 sentences contained more than one idiom. Finally, if no idiom was detected, then the idiom polarity values are set to zero: (0, 0, 0). 4.4. Sentiment classification Sentiment analysis can be viewed as a classification problem with three available classes: Positive, Negative and Other. A wide range of supervised learning approaches can be used for this pur- pose. We used Weka (Hall et al., 2009), a popular suite of machine learning software, to train a classifier and perform classification experiments. We based our choice of a machine learning method on the results of cross-validation experiments on the training data- set. We opted for a naïve Bayes classifier, more specifically a Bayesian network classifier, which outperformed other methods available in Weka. For example, a naïve Bayes classifier outper- formed support vector machines in terms of F-measure, proved more robust against the choice of features and provided more bal- anced classification performance across different classes. These findings agree with previous experiments when we successfully used a naïve Bayes approach to classify sentences by emotions they convey (Spasić, Burnap, Greenwood, & Arribas-Ayllon, 2012) and can be partially explained by the fact that a naïve Bayes classifier does not necessarily require a lot of training data to perform well (Domingos & Pazzani, 1997). 5. Evaluation To evaluate the impact of ignoring idioms in sentiment analysis, we conducted two experiments initially outlined in Section 4.3. In Experiment 1, we used SentiStrength (Thelwall, 2014; Thelwall et al., 2010) as the baseline method as well as a part of feature selection for our own method. In Experiment 2, we did the same with the Stanford CoreNLP’s sentiment annotator (Socher et al., 2013). The classification performance was evaluated in terms of three measures – precision (P), recall (R) and F-measure based on the numbers of true positives (TP), false positives (FP) and false negatives (FN). Tables 5 and 6 provide the comparison of these val- ues for the two methods considered. The overall performance rep- resents micro-averaged results across the three classes. Fig. 3. Sentiment analysis results from Stanford CoreNLP. Table 5 The evaluation results using SentiStrength as the baseline method. Class Method TP FP FN P R F Positive Baseline 40 53 98 43.01 28.99 34.63 Ours 102 63 36 61.82 73.91 67.33 Negative Baseline 111 66 127 62.71 46.64 53.49 Ours 170 54 68 75.89 71.43 73.59 Other Baseline 72 158 52 31.30 58.06 40.68 Ours 49 62 75 44.14 39.52 41.70 Overall Baseline 223 277 277 44.60 44.60 44.60 Ours 321 179 179 64.20 64.20 64.20 Table 7 Confusion matrices using SentiStrength as the baseline method. Actual Predicted P N O (a) Baseline P 40 36 62 N 31 111 96 O 22 30 72 (b) Our method P 102 18 18 N 24 170 44 O 39 36 49 Table 8 Confusion matrices using Stanford CoreNLP sentiment annotator as the baseline method. Actual Predicted P N O (a) Baseline P 41 74 23 N 25 170 43 O 19 86 19 (b) Our method L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 7381 With the exception of recall for Other class in Experiment 1 (a drop from 58.06% to 39.52%), our method demonstrates a consider- able improvement across all three measures. The overall improve- ment in Experiment 1 is 19.60 percent points (see Table 5) and 15 percent points in Experiment 2 (see Table 6). In terms of F-measure, there is an improvement across all three classes, but most notably in Positive classifications. This is mainly due to improving the recall considerably by 45 percent points in both experiments without compromising the precision. Confusion matrices given in Tables 7 and 8 show how classifica- tion outcomes are re-distributed across the three classes. Table 7a Table 6 The evaluation results using Stanford CoreNLP sentiment annotator as the baseline method. Class Method TP FP FN P R F Positive Baseline 41 44 97 48.24 29.71 36.77 Ours 104 81 34 56.22 75.36 64.40 Negative Baseline 170 160 68 51.52 71.43 59.86 Ours 181 82 57 68.82 76.05 72.26 Other Baseline 19 66 105 22.35 15.32 18.18 Ours 20 32 104 38.46 16.13 22.73 Overall Baseline 230 270 270 46.00 46.00 46.00 Ours 305 195 195 61.00 61.00 61.00 P 104 24 10 N 35 181 22 O 46 58 20 illustrates that SentiStrength is conservative in making both Positive and Negative predictions, thus less often misclassifying instances of Other class. Its classification outcomes were improved in all other cases by the use of idiom-based features (see Table 7b). Conversely, Stanford CoreNLP sentiment annotator proved to be more conservative in making Positive predictions compared to making Negative ones (see Table 8a), thus making fewer misclassi- fications when making Positive predictions. Nonetheless, its classi- fication outcomes were improved in all other cases by the use of idiom-based features (see Table 8b). Table 10 Distribution of idioms across the corpora. Corpus Idioms Unique idioms Ratio MR1 3250 359 51.26 MR2 464 161 62.53 MR3 75 44 63.56 HR 3904 294 17.05 CR 393 101 14.14 PR 462 113 36.09 TW 2788 309 36.54 7382 L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 Finally, in order to determine the statistical significance of the improvement over the baseline methods we performed the analy- sis of paired observations. We compared the sentiment classifica- tion results for each sentence before and after taking idioms in consideration by using continuity corrected McNemar’s test (Everitt, 1977) to check for statistically significant differences in error rates. Under the null hypothesis, the two methods compared should have the same error rate. McNemar’s test is based on the v2 test statistic and (approxi- mately) distributed as v2 with 1 degree of freedom. We used a vari- ant of McNemar’s test statistic that incorporates a correction for continuity to account for the fact that the statistic is discrete while the v2 distribution is continuous. The choice of this particular test was based on the following two facts: (1) McNemar’s test has been shown to have low type I error, in this case – the probability that it would incorrectly detect a difference when no difference exists. (2) Its statistical power is improved when compared with the com- monly used paired t-test (Dietterich, 1998). The specific v2(1) and p-values recorded for the data produced in Experiment 1, where SentiStrength was used as the baseline method, were v2(1) = 43.16 and p < 0.001. The values recorded for Experiment 2, where a sentiment annotator distributed as part of the Stanford CoreNLP was used as the baseline method, were v2(1) = 29.28 and p < 0.001. Therefore, in both cases the results of McNemar’s test confirmed that there was a statistically significant difference in error rates between the two methods. 6. Discussion We evaluated the impact of idioms as features of sentiment analysis by showing that they significantly improve classification results when such features are present. Their overall impact on sentiment classification can be estimated by combining this infor- mation with idiom distribution. For this purpose, we used corpora commonly used to evaluate performance of sentiment analysis (see Table 9). We used the idiom recognition module described in Section 4.1 to match a list of 580 idioms against the given cor- pora. Table 10 provides the number of matched idiom occurrences, the number of different idioms matched and the ratio of idiom occurrences against the corpus size in megabytes. These values illustrate that the distribution of idioms varies considerably across different genres. Idioms were most commonly found in the movie reviews (MR1–MR3). Focusing on full-text reviews, we observed that 6.02% of documents from corpus MR1 contained idioms, whereas this number in corpus MR2 rose to 18.95%, thus illustrating signif- icant impact idioms may have on document classification. If we compare full-text reviews from corpus MR2 to subjective snippets of such reviews from corpus MR3, we can conclude that subjective sentences are more likely to contain idioms than the rest of the text, again suggesting the value of idioms in sentence classification in terms of their sentiment polarity. Table 9 Description of the corpora. Identifier Document type Original source MR1 Movie reviews http://www.imdb.com/ MR2 Movie reviews http://www.rottentomatoes.com/ MR3 Movie reviews http://www.rottentomatoes.com/ HR Hotel reviews http://www.tripadvisor.co.uk/ CR Car reviews http://www.edmunds.com/ PR Product reviews http://www.cnet.com/ TW Tweets http://twitter.com/ As illustrated by corpora HR and CR, idioms were less com- monly found in hotel and car reviews. Contrary to previous find- ings that short conversations normally contain fewer idioms (Straessler, 1982), tweets proved to contain a relatively high pro- portion of idioms, which, at ratio of 36.54, is in line with that of product reviews. The number of unique idioms found was related to corpus size. Naturally, the highest number of different idioms was found in the largest corpora: MR1, HR and TW. More importantly, the variety of idioms used was found to be strongly correlated to the genre rather than the size. The use of idioms in each corpus followed the power law distribution with a small number of idioms used frequently and the rest used rarely, but the frequently used idioms differed across the genres. To explore the bias in idiom usage, we selected top 20 most frequently occurring idioms in each corpus and com- pared these sets using the Jaccard similarity coefficient (see Table 11), which is defined as the size of the intersection divided by the size of the union of the given sets. These similarities were used to cluster corpora based on idiom usage. Fig. 4 provides a den- drogram produced as a result of hierarchical clustering based on complete linkage and Euclidian distance, whose values are shown between the clusters. Table 12 illustrates the differences in most frequently used idioms across the corpora. Our attempt to generalize the evaluation results presented in Section 5 will be based on the following assumptions. Given a cor- pus of subjective sentences, let I refer to a subset of such sentences that contain idioms and let O refer to the remaining sentences. Further, let FI and FO denote the F-measure values achieved by the baseline sentiment classification method on these two subsets respectively. If we re-classify the sentiment on these subsets by combining the baseline method with idiom-based features as described in Section 4, then the value of FO will remain unchanged, whereas the F-measure on set I is expected to increase to FI + i, where i refers to the expected improvement. If p is the percentage of subjective sentences that contain idioms (i.e. p = |I|/|I [ O|), then we can use it to roughly estimate the overall value of the F-measure as the weighted average of the values achieved on the given subsets, i.e. F = p � (FI + i) + (1 � p) � FO. In that case, the overall improvement of the F-measure can be estimated as the value of the product p � i. In other words, in order to generalize the evaluation results presented in Section 5, the improvement of sentiment classification on sentences with idioms Coverage Size Source 50,000 reviews 63.40 MB (Maas et al., 2011) 2000 reviews 7.42 MB (Pang & Lee, 2004) 10,662 snippets 1.18 MB (Pang & Lee, 2005) 259,000 reviews 229.00 MB (Ganesan, Zhai, & Han, 2010) 42,230 reviews 27.80 MB (Ganesan et al., 2010) 300 products 12.80 MB (Ganesan, Zhai, & Viegas, 2012) 1,048,576 tweets 76.30 MB (Go, Bhayani, & Huang, 2009) http://www.imdb.com/ http://www.rottentomatoes.com/ http://www.rottentomatoes.com/ http://www.tripadvisor.co.uk/ http://www.edmunds.com/ http://www.cnet.com/ http://twitter.com/ Table 11 Similarity of idiom usage across the corpora. Corpus MR1 (%) MR2 (%) MR3 (%) HR (%) CR (%) PR (%) TW (%) MR1 100.00 33.33 25.00 11.11 8.11 8.11 17.65 MR2 33.33 100.00 17.65 14.29 2.56 2.56 8.11 MR3 25.00 17.65 100.00 2.56 2.56 2.56 5.26 HR 11.11 14.29 2.56 100.00 17.65 17.65 11.11 CR 8.11 2.56 2.56 17.65 100.00 33.33 11.11 PR 8.11 2.56 2.56 17.65 33.33 100.00 21.21 TW 17.65 8.11 5.26 11.11 11.11 21.21 100.00 Fig. 4. Clustering of corpora based on idiom usage. Table 12 Top five most frequently used idioms across the corpora. Corpus Idiom 1 Idiom 2 Idiom 3 Idiom 4 Idiom 5 MR1 fall flat save the day butterflies in stomach jaw drop guilty pleasure MR2 fall flat save the day guilty pleasure devil’s advocate jaw drop MR3 guilty pleasure fall flat jaw drop speak volumes close to home HR chill out mixed feelings go the extra mile lie in last resort CR come a long way never looked back take it easy happy camper leaps and bounds PR blockbuster without a hitch happy camper never looked back leaps and bounds TW chill out lie in take it easy bored to tears hit the sack L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 7383 should be scaled down by the percentage of such sentences. Based on our discussion (see Tables 10 and 11), the values of both p and i are expected to vary across different genres. For example, p was found to be over 0.70% and 0.27% on corpora MR3 and TW respec- tively (note that we only used a list of 580 idioms, so these num- bers are expected to be higher). On the other, the classification improvement i will vary depending on both the baseline method and the genre. Our sentiment classification experiments were con- ducted on a genre-neutral corpus of sentences that contain a wide spectrum of uniformly distributed idioms, and as such were used to estimate the expected improvement i of 15 to 20 percent points in an unbiased fashion. 7. Conclusions We have demonstrated the value of idioms as features of senti- ment analysis by showing that idiom-based features significantly improve sentiment classification results when such features are present. The overall performance in terms of precision, recall and F-measure was improved from 45% to 64% in one experiment, and from 46% to 61% in the other. In terms of F-measure, there is an improvement across all classes, but most notably in classifica- tions of positive sentiment. This is mainly attributed to consider- ably improved recall by 45 percent points in both experiments without compromising the precision. In order to generalize these findings, we combined them with information on idiom distribu- tion. For this purpose, we used corpora commonly used to evaluate performance of sentiment analysis. The improvement of sentiment classification on sentences with idioms should be scaled down by the percentage of such sentences. This number, however, varied across different corpora and requires further research. For exam- ple, it was found to be over 0.70% and 0.27% on corpora of movie reviews and tweets respectively. In addition to experimental results, this study provides resources that can support further research into sentiment analysis including. We created a comprehensive collection of 580 idioms annotated with sentiment polarity, which represents the largest lexico-semantic resource of this kind to utilize in sentiment analy- sis. In addition, we implemented a set of local grammars that can 7384 L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 be used to recognize occurrences of these idioms in text. Thus far, the use of idioms in sentiment analysis was based on ad hoc approaches focusing on a severely limited set of idioms that can be identified in text using dictionary lookup methods. Another obstacle in systematically investigating the role of idioms in senti- ment analysis is their relative rarity. As explained in previous dis- cussion, corpora commonly used for evaluation of sentiment analysis approaches are biased in their use of idioms, which pre- vents the findings on the role of idioms in sentiment analysis to be generalized. We assembled a corpus of 2521 sentences with a wide range of idioms used in context. Similarly to idioms them- selves, this corpus is also annotated with sentiment polarity, which can be used in systematic evaluation of sentiment analysis approaches that claim to use idioms as features. All resources are freely available for download from the following URL: http:// www.cs.cf.ac.uk/idioment. This is the first in our series of studies on the role of idioms in sentiment analysis. To our best knowledge, this is the first study of this kind in English. Given the positive findings of the initial study, we will be looking to scale up the approach described here. Its main limitation is a significant overhead involved in handcraft- ing lexico-semantic rules for recognition of idioms and their polar- ity. To address this bottleneck, we will attempt to automate two crucial steps: (1) encoding local grammars that enable idiom recognition in text, and (2) determining idiom polarity. The corpus of sentences containing idioms in combination with the set local grammars for idiom recognition can be used as the training set to learn their lexico-syntactic patterns, i.e. induce which idiom components are frozen and which ones vary and in which sense (e.g. inflection, substitution, etc.). These findings can be used to generate lexico-syntactic patterns that will automati- cally model idioms by means of local grammars. Such rules can be optimized in terms of precision and recall using our corpus annotated with idiom occurrences. Solving a problem of such high complexity would enable utilization of an arbitrary list of idioms without incurring additional implementation overhead. In order to utilize an arbitrary list of idioms in sentiment anal- yses, we also need their polarity. In this study, we used crowd- sourcing for this purpose. While such an approach proved to be efficient especially when using commercial platforms such as http://www.crowdflower.com/, we would like to fully automate acquisition of idiom polarity as part of scaling up our existing approach. We suggest repurposing existing idiom dictionaries aimed at English language learners. We will conduct sentiment analysis experiments against idiom definitions. The goal of this exercise is to determine whether sentiment analysis results over idiom descriptions (not idioms themselves) are correlated with manual annotations acquired in this study. If this is the case, then it will be possible to generate idiom-based features (i.e. their sen- timent polarities) automatically in order to utilize a much wider set of idioms in sentiment analysis. Combined together, both avenues of further research seek to fully automate the use of idioms in sentiment analysis and mini- mize the knowledge elicitation bottleneck associated with this task. Finally, taking a long-term view, we have initially annotated both idioms and sentences with a wide range of emotions rather than merely sentiment polarity. This will enable future studies to expand on our existing research to tackle the more complex prob- lem of emotion classification (Spasić et al., 2012). Acknowledgments Data cited herein have been extracted from the British National Corpus, distributed by Oxford University Computing Services on behalf of the BNC Consortium. All rights in the texts cited are reserved. We gratefully acknowledge the help of our colleagues and friends who supported this study by volunteering their time for data annotation. We would like to thank Prof. Mike Thelwall for kindly allowing us to use the Java version of SentiStrength. References Attardi, G., & Simi, M. (2006). Blog mining through opinionated words. In Text Retrieval Conference. Gaithersburg, Maryland, USA. Aue, A., & Gamon, M. (2005). Customizing sentiment classifiers to new domains: A case study. In Proceedings of the international conference on recent advances in natural language processing. Borovets, Bulgaria. Balahur, A., Steinberger, R., Kabadjov, M., Zavarella, V., Goot, E. v. d., Halkia, M., Pouliquen, B., & Belyaeva, J. (2010). Sentiment analysis in the news. In Proceedings of the international conference on language, resources and evaluation (pp. 2216–2220). Valletta, Malta. Beigman Klebanov, B., Burstein, J., & Madnani, N. (2013). Sentiment profiles of multiword expressions in test-taker essays: The case of noun–noun compounds. ACM Transactions on Speech and Language Processing, 10, 12. Bethard, S., Yu, H., Thornton, A., Hatzivassiloglou, V., & Jurafsky, D. (2004). Automatic extraction of opinion propositions and their holders. In Association for artificial intelligence spring symposium on exploring attitude and affect in text (p. 2224). Palo Alto, California, USA: The AAAI Press. BNC Consortium. (2014). The British National Corpus, version 3 (BNC XML Edition), URL: <http://www.natcorp.ox.ac.uk/>. Breck, E., Choi, Y., & Cardie, C. (2007). Identifying expressions of opinion in context. In Proceedings of the 20th international joint conference on artificial intelligence (pp. 2683–2688). Hyderabad, India. Buczynski, A., & Wawer, A. (2008). Shallow parsing in sentiment analysis of product reviews. In Proceedings of the LREC workshop on partial parsing: between chunking and deep parsing (pp. 14–18). Marrakech. Cohen, W. W. (2014). MinorThird: Methods for identifying names and ontological relations in text using heuristics for inducing regularities from data. URL: <http://www.minorthird.sourceforge.net>. Damasio, A. R. (1999). The feeling of what happens: Body and emotion in the making of consciousness. New York, USA: Harcourt Incorporated. Darwin, C. (1872). The expression of emotions in man and animals. London, UK: John Murray. Dave, K., Lawrence, S., & Pennock, D. M. (2003). Mining the peanut gallery: Opinion extraction and semantic classification of product reviews. In Proceedings of the 12th international world wide web conference. Budapest, Hungary. Dietterich, T. G. (1998). Approximate statistical tests for comparing supervised classification learning algorithms. Neural Computation, 10, 1895–1923. Domingos, P., & Pazzani, M. (1997). On the optimality of the simple Bayesian classifier under zero-one loss. Machine Learning, 29. Ekman, P. (1972). Universals and cultural differences in facial expressions of emotions. In J. Cole (Ed.). Nebraska Symposium on Motivation, 1971 (Vol. 19, pp. 207–283). Lincoln, Nebrasca, USA: University of Nebraska Press. Ekman, P. (1992). An argument for basic emotions. Cognition and Emotion, 6, 169–200. Esuli, A., & Sebastiani, F. (2006). SentiWordNet: A publicly available lexical resource for opinion mining. In proceedings of the international conference on language, resources and evaluation (pp. 417–422). Genoa, Italy. Everitt, B. S. (1977). The analysis of contingency tables. London: Chapman & Hall. Fraser, B. (1970). Idioms within a transformational grammar. Foundations of Language, 6, 22–42. Ganesan, K., Zhai, C., & Han, J. (2010). Opinosis: a graph-based approach to abstractive summarization of highly redundant opinions. In Proceedings of the 23rd international conference on computational linguistics (pp. 340–348). Beijing, China. Ganesan, K., Zhai, C., & Viegas, E. (2012). Micropinion generation: An unsupervised approach to generating ultra-concise summaries of opinions. In Proceedings of the 21st international conference on World Wide Web (pp. 869–878). Geertzen, J. (2012). Inter-rater agreement with multiple raters and variable, URL: <https://mlnl.net/jg/software/ira/>, Retrieved May 8, 2014. Gibbs, R. W., Jr., & Nayak, N. P. (1989). Psycholinguistic studies on the syntactic behavior of idioms. Cognitive Psychology, 21, 100–138. Go, A., Bhayani, R., & Huang, L. (2009). Twitter sentiment classification using distant supervision. Technical report, CS224N (pp. 1–12). Stanford Digital Library. Grant, L. E. (2005). Frequency of ‘core idioms’ in the British National Corpus (BNC). International Journal of Corpus Linguistics, 10, 429–451. Greenwood, M., Elwyn, G., Francis, N., Preece, A., & Spasić, I. (2013). Automatic extraction of personal experiences from patients’ blogs: A case study in chronic obstructive pulmonary disease. In Proceedings of the third international conference on social computing and its applications (pp. 377–382). Karlsruhe, Germany. Gross, M. (1997). The construction of local grammars. In E. Roche & Y. Schabes (Eds.), (pp. 329–354). The MIT Press. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., & Witten, I. H. (2009). The Weka data mining software: An update. ACM SIGKDD Explorations Newsletter, 11, 10–18. Jackendoff, R., & Pinker, S. (2005). The nature of the language faculty and its implications for evolution of language (Reply to Fitch, Hauser, and Chomsky). Cognition, 97, 211–225. http://www.cs.cf.ac.uk/idioment http://www.cs.cf.ac.uk/idioment http://www.crowdflower.com/ http://refhub.elsevier.com/S0957-4174(15)00375-9/h0020 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0020 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0020 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0025 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0025 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0025 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0025 http://www.natcorp.ox.ac.uk/ http://www.minorthird.sourceforge.net http://refhub.elsevier.com/S0957-4174(15)00375-9/h0050 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0050 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0055 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0055 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0065 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0065 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0070 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0070 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0075 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0075 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0075 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0080 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0080 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0090 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0095 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0095 https://mlnl.net/jg/software/ira/ http://refhub.elsevier.com/S0957-4174(15)00375-9/h0115 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0115 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0125 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0125 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0140 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0140 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0140 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0145 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0145 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0145 L. Williams et al. / Expert Systems with Applications 42 (2015) 7375–7385 7385 Krippendorff, K. H. (1980). Content analysis: An introduction to its methodology. Beverly Hills, CA: SAGE Publications Inc. Krippendorff, K. (2004). Reliability in content analysis: Some common misconceptions and recommendations. Human Communication Research, 30, 411–433. Learn English Today. (2013). URL: <http://www.learn-english-today.com/idioms/ idioms_proverbs.html>. LeDoux, J. E. (1996). The emotional brain. New York, USA: Simon & Schuster. Leech, G. (1993). 100 million words of English. English Today, 9, 9–15. Liu, D. (2003). The most frequently used spoken American English idioms: A corpus analysis and its implications. TESOL Quarterly, 37, 671–700. Liu, B. (2010). Sentiment analysis and subjectivity. In N. Indurkhya & F. J. Damerau (Eds.), Handbook of Natural Language Processing (2nd ed.. Boca: CRC Press, Taylor and Francis Group. Maas, A. L., Daly, R. E., Pham, P. T., Huang, D., Ng, A. Y., & Potts, C. (2011). Learning word vectors for sentiment analysis. In Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies (pp. 142–150). Portland, Oregon, USA: Association for Computational Linguistics. Mohammad, S. M., & Turney, P. D. (2010). Emotions evoked by common words and phrases: using mechanical Turk to create an emotion lexicon. In Proceedings of the NAACL HLT workshop on computational approaches to analysis and generation of emotion in text (pp. 26–34). Los Angeles, California, USA. Munezero, M. D., Montero, C. S., Sutinen, E., & Pajunen, J. (2014). Are they different? Affect, feeling, emotion, sentiment, and opinion detection in text. IEEE Transactions on Affective Computing, 5, 101–111. Nippold, M. A., & Martin, S. T. (1989). Idiom interpretation in isolation versus context: A developmental study with adolescents. Journal of Speech, Language and Hearing Research, 32, 59–66. Nunberg, G., Sag, I. A., & Wasow, T. (1994). Idioms. Language, 70, 491–538. Pang, B., & Lee, L. (2004). A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd annual meeting on association for computational linguistics. Pang, B., & Lee, L. (2005). Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the 43rd annual meeting on association for computational linguistics (pp. 115–124). Pang, B., & Lee, L. (2008). Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2, 1–135. Rentoumi, V., Vouros, G. A., Karkaletsis, V., & Moser, A. (2012). Investigating metaphorical language in sentiment analysis: A sense-to-sentiment perspective. ACM Transactions on Speech and Language Processing, 9, 6. Rubin, D. C., & Talarico, J. M. (2009). A comparison of dimensional models of emotion: Evidence from emotions, prototypical events, autobiographical memories, and words. Memory, 17, 802–808. Russell, J. (2003). Core affect and the psychological construction of emotion. Psychological Review, 110, 145–172. Shastri, L., Parvathy, A. G. K., Abhishek, Wesley, J., & Blakrishnan, R. (2010). Sentiment extraction: Integrating statistical parsing, semantic analysis, and common sense reasoning. In Proceedings of the twenty-second conference on innovative applications of artificial intelligence. Atlanta, Georgia, USA. Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C. D., Ng, A., et al. (2013). Recursive deep models for semantic compositionality over a Sentiment Treebank. In Proceedings of the conference on empirical methods in natural language processing (pp. 1631–1642). Seattle, Washington, USA: Association for Computational Linguistics. Spasić, I., Burnap, P., Greenwood, M., & Arribas-Ayllon, M. (2012). A naïve Bayes approach to topic classification in suicide notes. Biomedical Informatics Insights, 5, 87–97. Straessler, J. (1982). Idioms in English: A pragmatic analysis. John Benjamins Pub Co. The Association for the Advancement of Affective Computing. (2014). Emotion annotation and representation language (EARL), Version 0.4.0, 30 June 2006, URL: <http://emotion-research.net/projects/humaine/earl>. Thelwall, M. (2014). SentiStrength, URL: <http://sentistrength.wlv.ac.uk/>. Thelwall, M., Buckley, K., & Paltoglou, G. (2012). Sentiment strength detection for the social web. Journal of the American Society for Information Science and Technology, 63, 163–173. Thelwall, M., Buckley, K., Paltoglou, G., Cai, D., & Kappas, A. (2010). Sentiment strength detection in short informal text. Journal of the American Society for Information Science and Technology, 61, 2544–2558. Turney, P., & Littman, M. (2003). Measuring praise and criticism: Inference of semantic orientation from association. ACM Transactions on Information Systems, 21, 315–346. Valitutti, A., Strapparava, C., & Stock, O. (2004). Developing affective lexical resources. Psychology, 2, 61–83. Wilson, T., Wiebe, J., & Hoffmann, P. (2009). Recognizing contextual polarity: An exploration of features for phrase-level sentiment analysis. Computational Linguistics, 35, 399–433. Xie, S.-X., & Wang, T. (2014). Construction of unsupervised sentiment classifier on idioms resources. Journal of Central South University, 21, 1376–1384. Yusifova, G. I. (2013). Syntactic features of English idioms. International Journal of English Linguistics, 3, 133–138. http://refhub.elsevier.com/S0957-4174(15)00375-9/h0150 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0150 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0155 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0155 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0155 http://www.learn-english-today.com/idioms/idioms_proverbs.html http://www.learn-english-today.com/idioms/idioms_proverbs.html http://refhub.elsevier.com/S0957-4174(15)00375-9/h0165 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0170 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0175 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0175 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0180 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0180 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0180 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0185 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0185 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0185 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0185 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0185 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0195 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0195 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0195 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0200 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0200 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0200 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0205 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0220 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0220 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0225 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0225 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0225 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0230 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0230 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0230 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0235 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0235 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0245 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0245 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0245 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0245 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0245 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0250 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0250 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0250 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0255 http://emotion-research.net/projects/humaine/earl http://sentistrength.wlv.ac.uk/ http://refhub.elsevier.com/S0957-4174(15)00375-9/h0270 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0270 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0270 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0275 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0275 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0275 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0280 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0280 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0280 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0285 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0285 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0290 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0290 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0290 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0295 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0295 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0300 http://refhub.elsevier.com/S0957-4174(15)00375-9/h0300 The role of idioms in sentiment analysis 1 Introduction 2 Data collection 2.1 Idioms 2.2 Corpus 3 Data annotation 3.1 Annotation scheme 3.2 Annotation process 3.3 Annotation results 3.4 Gold standard 4 Sentiment analysis 4.1 Idiom recognition 4.2 Negation 4.3 Feature selection 4.4 Sentiment classification 5 Evaluation 6 Discussion 7 Conclusions Acknowledgments References 
work_ig7kgk7r6zf7nelmmag2y5dddu	----	Mining changes in customer behavior in retail marketing Mining changes in customer behavior in retail marketing Mu-Chen Chen a,*, Ai-Lun Chiu b , Hsu-Hwa Chang c a Department of Business Management, National Taipei University of Technology, No. 1, Section 3, Chung-Hsiao E. Road, Taipei 106, Taiwan, ROC b SAS, Taipei, Taiwan, ROC c Department of Business Administration, National Taipei College of Business, Taipei, Taiwan, ROC Abstract During the past decade, there have been a variety of significant developments in data mining techniques. Some of these developments are implemented in customized service to develop customer relationship. Customized service is actually crucial in retail markets. Marketing managers can develop long-term and pleasant relationships with customers if they can detect and predict changes in customer behavior. In the dynamic retail market, understanding changes in customer behavior can help managers to establish effective promotion campaigns. This study integrates customer behavioral variables, demographic variables, and transaction database to establish a method of mining changes in customer behavior. For mining change patterns, two extended measures of similarity and unexpectedness are designed to analyze the degree of resemblance between patterns at different time periods. The proposed approach for mining changes in customer behavior can assist managers in developing better marketing strategies. q 2005 Elsevier Ltd. All rights reserved. Keywords: Data mining; Association rules; Retailing; Customer behavior; Change patterns 1. Introduction Customized service is actually crucial in retail markets. Customized service has become a key issue in developing customer relationships. Marketing managers can develop long-term and pleasant relationships with customers if they can detect and predict changes in customer behavior. In the past, researchers generally applied statistical surveys to study customer behavior. Recently, however, data mining techniques have been adopted to predict customer behavior (Giudici & Passerone, 2002; Song, Kim, & Kim, 2001). Data mining techniques search through a database without any specific pre-determined hypothesis to obtain implicit, previously unknown, and potentially useful information including knowledge rules, constraints and regularities (Chen, Han, & Yu, 1996). Data mining is a stage in Knowledge Discovery in Databases (KDD), involving the application of specific algorithms for pattern extraction (Mitra, Pal, & Mitra, 2002). Various successful applications have been reported in areas such as marketing, finance and banking. Applications in these domains generally involve the collection and storage of large amounts of data. Data mining brings various techniques together to discover patterns (rules) and to construct models from databases. Currently businesses face the challenge of a constantly evolving market where customer needs are changing all the time. In such a situation, change mining can enable market analysts to better understand changes in customer needs and how those needs change. Change mining is more appropriate in dynamic business environments, and involves extensive human intervention (Song et al., 2001). Retail market managers must not only provide high- quality products and services, but also must react appro- priately to changes in customer needs. Data mining can be applied to identify useful customer behavior patterns from large amounts of customer and transaction data (Giudici & Passerone, 2002). As a result, the discovered information can be ascertained to support better decision-making in retail marketing. Data mining techniques have mostly been adopted to generate predictions and describe behaviors. Relatively little research has focused on mining changes in databases collected over time (Liu, Hsu, Han, & Xia, 2000). 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2004.12.033 Expert Systems with Applications 28 (2005) 773–781 www.elsevier.com/locate/eswa * Corresponding author. Tel.: C886 2 2771 2171x3417; fax: C886 2 2776 3964. E-mail address: bmcchen@ntut.edu.tw (M.-C. Chen). http://www.elsevier.com/locate/eswa In the dynamic retail market, understanding changes in customer behavior can help managers to establish effective promotion campaigns (Song et al., 2001). Liu et al. (2000) devised a method of change mining in the context of decision trees for predicting changes in customer behavior. Since decision tree is a classification-based approach, it cannot detect complete sets of changes (Song et al., 2001). Association rule extraction was widely used for analyzing the correlation between product items purchased by customers, and to support sales promotion and market segmentation (Changchien & Lu, 2001; Changchien, Lee, & Hsu, 2004). Song et al. (2001) employed an approach based on association rules to identify changes in customer behavior. Most previous customer behavior studies applied custo- mer demographic variables to analyze customer behavior (Song et al., 2001). However, valuable customer behavioral variables, such as recency, frequency, and monetary (RFM), can be used to differentiate customer contributions to a business (Miglautsch, 2000; Stone, 1995; Suh, Noh, & Suh, 1999; Tsai & Chiu, 2004). Researchers have observed that RFM is a widely used technique for customer behavioral analysis that can effectively investigate customer values and segment markets. Customer behavioral variables, RFM, respectively, measure the recency of customer purchasing behavior, the frequency of purchasing, and the average monetary expenditure on purchasing. RFM can be trans- formed using customer and transaction databases. This study attempts to integrate customer behavioral variables (RFM), demographic variables, and transaction database to establish a method of mining changes in customer behavior. Song et al. (2001) designed two measures of similarity and unexpectedness to analyze the degree of resemblance between patterns at different time periods. However, these two measures are limited to the analysis of patterns with a single attribute on the right-hand- side (consequent part) of an association rule. This study designs two modified measures of similarity and unexpect- edness to overcome the above limitations. In this study, the data required for analysis are integrated and transformed from customer, product, and transaction databases. The proposed approach for mining changes in customer behavior can assist managers in developing better marketing strategies. 2. Mining customer behavior changes In this study, customer behavior patterns are first identified using association rule mining. Following the association rules of customer behavior are discovered, the changes in customer behavior are identified by comparing two sets of association rules generated from two datasets of different periods. Based on previous studies, changes in customer behavior include emerging patterns, added patterns, perished patterns, and unexpected patterns (Dong & Li, 1999; Liu & Hsu, 1996; Liu, Hsu, Mun, & Lee, 1999; Padmanabhan & Tuzhilin, 1999; Song et al., 2001). The discovered change patterns can be further explained and assessed to provide a basis for formulating marketing strategies. Fig. 1 illustrates the flowchart of change mining for customer behavior. Further details of the change mining procedure are discussed below. 2.1. Data pre-processing Prior to analysis, data accuracy and consistency must be ensured to obtain truthful results. Generally, some useful variables can be hidden in a large quantity of raw data, and thus can be obtained through data integration and transformation. Customer behavioral variables (RFM) are hidden in customer and transaction databases, and can be extracted from data integration and transform- ation. Since the data required for analyzing association rules must be discrete, continuous variables are trans- formed to discrete variables, and a simple 3-4-5 rule is applied (Han & Kamber, 2001). In RFM, recency represents the interval between the most recent transaction time of individual customers and the evaluation time (Stone, 1995). Moreover, frequency rep- resents the number of purchases by individual customers during a specific period. Additionally, monetary represents the average expenditure of a customer during a specific period. Individual customers’ recency, frequency, and monetary are scored to calculate the value of the purchasing behavior of each customer. This study adopts the RFM scoring approach of Miglautsch (2000) to transform the customer behavioral variables. Data preprocessing and customer behavioral & demographic data transformation Mining customer behavior Mining change patterns: • Emerging • Added • Perished • Unexpected Supporting Marketing Customer Product Transaction Fig. 1. Flowchart of mining changes for customer behavior. M.-C. Chen et al. / Expert Systems with Applications 28 (2005) 773–781774 https://isiarticles.com/article/20882 
work_igv6r4iavzgn3mtitipb2naxam	----	Selection and fusion of facial features for face recognition Expert Systems with Applications 36 (2009) 7157–7169 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Selection and fusion of facial features for face recognition Xiaolong Fan, Brijesh Verma * School of Computing Sciences, Faculty of Business and Informatics, Central Queensland University, Rockhampton, QLD 4701, Australia a r t i c l e i n f o Keywords: Face recognition Neural networks Evolutionary algorithms Pattern recognition 0957-4174/$ - see front matter � 2008 Elsevier Ltd. A doi:10.1016/j.eswa.2008.08.052 * Corresponding author. E-mail addresses: x.fan@cqu.edu.au (X. Fan), b.verm a b s t r a c t This paper proposes and investigates a facial feature selection and fusion technique for improving the classification accuracy of face recognition systems. The proposed technique is novel in terms of feature selection and fusion processes. It incorporates neural networks and genetic algorithms for the selection and classification of facial features. The proposed technique is evaluated by using the separate facial region features and the combined features. The combined features outperform the separate facial region features in the experimental investigation. A comprehensive comparison with other existing face recog- nition techniques on FERET benchmark database is included in this paper. The proposed technique has produced 94% classification accuracy, which is a significant improvement and best classification accuracy among the published results in the literature. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction 1.1. Background Face recognition is one of the most remarkable capabilities of human beings. It develops over the early years of childhood, and is important for several aspects of our social life. Human beings can remember hundreds or even thousands of faces in their whole life and can easily identify a familiar face in different perspective variations, such as illumination variations, age variations, and pose variations. Face recognition together with other abilities, such as estimating the expression of people with whom we interact, has played an important role in the course of evolution. The problem of machine recognition of faces has been studied for more than 30 years. It has attracted research interest from sev- eral disciplines such as image processing, pattern recognition, computer vision, neural networks and computer graphics. Such interest has been motivated by the growth of Face Recognition Technology (FRT) used in the applications in many areas, including face identification in law enforcement and forensics, user authen- tication in building access or automatic teller machines, indexing of, and searching for, faces in video databases, intelligent computer user interfaces, etc. After the September 11, 2001, terrorist attacks, FRT has been gaining more interest due to its significant involve- ment in anti-terror activities. FRT numerously used in commercial and law enforcement applications poses a wide range of technical challenges and requires an equally wide range of techniques from different disciplines. ll rights reserved. a@cqu.edu.au (B. Verma). A general statement of the problem of machine recognition of faces can be described as follows: Given still or video images of a scene, identify or verify one or more persons in the scene using a stored database of faces. Available collateral information such as race, gender, age, facial expression or speech may be used in nar- rowing the search. The solution to the problem involves face detec- tion, feature extraction from face region, face verification or recognition. Face detection refers to the determination of the exact position and size of a human face from cluttered scenes. Feature extraction refers to obtaining the features that can be fed into a face classification system. Face recognition refers to comparing an input face against models of faces that are stored in a database of known faces and then indicating if a match is found. Face veri- fication refers to confirming or rejecting the claimed identity of the input face. Although human beings seem to recognize a face in cluttered scenes with relative ease, machine recognition is much more dif- ficult for a variety of reasons. Firstly, different faces may appear very similar, i.e. every face contains two eyes, two ears, one nose and one mouth, thereby necessitating an exacting discriminant task. Secondly, different views of the same face may appear quite different due to imaging constraints, such as changes in illumination and variability in facial expressions, and due to the presence of personal accessories, such as glasses, beards, and hats. Finally, when the face undergoes rotations out of the imaging plane, a large amount of detailed facial structure may be occluded. Therefore, until now in many implementations of face recognition algorithms, the face images are obtained in a constrained environment with controlled illumination, minimal occlusions of facial structures, uncluttered background, and so on. Face recognition in an unconstrained environment is still a quite challenging task. mailto:x.fan@cqu.edu.au mailto:b.verma@cqu.edu.au http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 7158 X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 1.2. Literature review In the last decade, face recognition has become one of most ac- tive research areas of pattern recognition. The most existing face recognition methods can be simply classified into three categories: holistic feature-based matching method, local feature-based matching method and hybrid matching method (Chellappa, Wilson, & Sirohey, 1995). In holistic feature-based matching method, the whole face region is used as raw input to the recogni- tion system, like Principal Component Analysis (PCA) projection method (Turk & Pentland, 1991), Fisher-face method (Belhumeur, Hespanha, & Kriegman, 1997) and Nearest Feature Line (NFL) method (Li & Lu, 1999). Recently, an Independent Gabor Features (IGF) method (Liu & Wechsler, 2003) and a kernel Associative Memory (kAM) models-based method (Zhang, Zhang, & Ge, 2004) were also applied to face recognition. In local feature-based matching method, the local features such as eyes, nose, and mouth are first extracted and then their locations and local statistics (geometric and/or appearance) are fed into a structural classifier. Geometrical features method (Brunelli & Poggio, 1992) and Elastic Bunch Graph Matching (EBGM) method (Wiskott, Fellous, & Malsburg, 1997) belong to this category. In hybrid matching method, both holistic and local features are used for the recognition. A feature combination scheme for face recognition by fusion of global and local features is presented in Fang, Tan, and Wang (2002). A fully automatic system for face rec- ognition in databases with only a small number of samples is pre- sented in Yan et al. (2004). Global and local texture features are extracted and used in the recognition. Genetic Algorithms (GAs) can be used to select an optimal fea- ture set for pattern classification problems. Some researchers have used GAs for face recognition. In Bala, Huang, Vafaie, DeJong, and Wechsler (1995), GA–ID3 (decision tree learning) method is proposed to find an optimal subset of discriminatory features for pattern classification. GAs were used to search the possible optimal subset of extracted features. ID3 was used to produce a decision tree based on a subset selected by GAs. The GA–ID3 method was experimented to recognize visual concepts in satellite and face images. The results showed a significant improvement in classification performance and a good reduction in feature set dimension. In Liu, Tang, Lu, and Ma (2004), a kernel scatter-difference-based discriminant analysis for face recognition is presented. In Sun and Yin (2005), a genetic algorithm was used to select features for 3D face recognition. The method presented in Sun and Yin (2005), tries to optimise features by capturing good features which can minimize the inner-class distance and maximize the intra-class distance. In Liu and Wechsler (2000), an evolutionary pursuit (EP) based on GAs has been applied to face recognition. The idea in EP is to search for face basis through the rotated axes defined in PCA space. The overall classification rate obtained by the existing techniques is unsatisfactory; there- fore there is a need for a better feature selection and fusion tech- nique which could improve the overall classification accuracy for face recognition. In this paper, a novel feature selection and fusion technique for face recognition is presented. GA for feature selection and Artificial Neural Network (ANN) for classification were incorporated into the proposed technique. The proposed technique has been tested on a separate feature set from each facial region and compared with the combined feature set. A large set of dataset from the FERET bench- mark database (Phillips, Wechsler, Huang, & Rauss, 1998) is used for testing. The main research questions are (1) How to select the most significant facial features and combine them to improve an overall classification rate of face recognition systems? (2) What is the best combination of these features to a specific classifier? The original contributions of the research presented in this paper are as follows: (1) Identification of local facial regions by using dis- tance threshold method based on center coordinate information of each facial region. The facial features are extracted from each facial region. (2) A Genetic Algorithms (GAs)-based approach for facial feature selection. The significant areas inside each facial region are located using this approach. (3) An Artificial Neural Network (ANN)-based approach for facial feature classification. The selected facial features from GA approach are passed to ANN for final clas- sification. The classification error is passed back to GA to calculate the fitness of each individual. (4) A combined technique for face recognition. The proposed approach is tested on the separate fea- ture set from each facial region and the combined feature set. The FERET benchmark database is adopted to evaluate and com- pare the proposed approach. A comprehensive comparison of the proposed technique with other existing face recognition ap- proaches has been conducted. 2. Proposed technique This section describes the proposed feature selection and fusion technique for face recognition. Section 2.1 provides an overview of the proposed methodology. Section 2.2 introduces the distance threshold method that is used to locate facial regions. The average grey level value features are discussed in Section 2.3. Section 2.4 describes PCA features. The details of incorporating GAs and neural networks for feature selection and classification are discussed in Section 2.5. 2.1. Overview The goal of the proposed technique is to select the most signif- icant facial features effectively and find the best combination of these features for the classifier. The proposed technique aims to lo- cate the significant areas in facial regions from which the signifi- cant features are extracted. Facial regions refer to the separate regions in the face that contain one local organ, such as left eye re- gion, right eye region, nose region and mouth region. These facial regions contain the most discriminant facial characteristics on hu- man faces. The facial regions are the basis for the local feature- based feature extraction techniques. Even on these discriminant fa- cial regions, some areas inside may be more important than the other areas in a recognition task. By locating the most significant areas on the facial regions, the proposed approach actually re- moves ‘‘noise” information caused by other non-significant areas of the facial region. It may also remove part of the variation infor- mation caused by changes in facial expression, head rotation and illumination. By concentrating on these significant areas, it allows us to extract the most significant facial features from them to rep- resent human faces. These features may improve the classification rate of face recognition systems. The first step in the proposed technique is to locate facial re- gions in the face images. The facial feature extraction technique is performed on these facial regions. After feature extraction, the features are selected, fused and classified. Through selection, the significant areas are located, and through classification, the input face image is recognised or verified. The block diagram of the proposed technique to conduct exper- iments using separate and combined features on FERET benchmark dataset is depicted in Fig. 2.1. The details are described in the fol- lowing subsections. 2.2. Locate facial regions We first locate facial regions on each face image and then we extract features. The experimental face images are extracted from the FERET database. The center coordinate information provided vahab Highlight vahab Highlight Locate Facial Regions Locate Significant Areas GA Small Face Dataset Feature Extraction Separate Feature Set Classification Recognized Face Feature Selection Large Face Dataset ANN Classification Error Combined Feature Set Fig. 2.1. Block diagram of the proposed technique. X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 7159 for each facial region, such as eye center coordinate, nose tip coor- dinate and mouth center coordinate, is used and the distance threshold method is applied to locate the local facial regions. The distance threshold method defines distance thresholds in vertical and horizontal directions for the local facial region. These thresholds decide the size of the facial region. With center coordi- nate information, the facial region is easy to locate. Based on the images in the experimental database, the distance thresholds are set as follows. The vertical distance threshold is set to 16 and the horizontal distance threshold is set to 30 for the eyes and nose re- gions. They are set to 12 and 60 separately for the mouth region. 2.3. Average grey level value feature After locating the facial regions, each facial region was equally divided into small-size rectangle areas. The average grey level va- lue features are extracted from these small rectangle areas. The average grey level value feature can be expressed as gi ¼ P pðx; yÞ w � h � v ð1Þ where gi is the average grey level value feature for the small rectan- gle area i, p(x, y) is grey level value of pixel p inside the rectangle area i, w is the width of the small rectangle area and h is the height. v is the maximum grey level value for the image, here 255 for the experimental database. After division, the average grey level value features are ex- tracted on these small rectangle areas from left to right, top to bot- tom. In the experiments, The size of the small rectangle area was chose to be 6 � 4 (w = 6, h = 4). Then for the left eye region (same as right eye region and nose region), the size of the extracted fea- ture set becomes 20. For the mouth region, the size of the extracted feature set increases to 30 due to the larger size of the mouth region. 2.4. PCA feature PCA projection method for face recognition, which is also called eigenface method, is a classical method for face recognition. The simple idea behind eigenface method is to capture the largest vari- ances among a set of face images, and then use this information to encode and compare face images. The advantage of eigenface method is reduction of dimensionality, while maximizing the scat- ter of all the projected samples. Let {X1, X2, . . . , XN} be as set of N sample images. It takes values in an n-dimensional image space and each image belongs to one of the c classes {x1, x2, . . . , xc}. A lin- ear transformation needs to be found to map the original n-dimen- sional image space into an m-dimensional feature space, where m < n. The new feature vector yk 2 Rm is defined by the following equation: yk ¼ W TXk; k ¼ 1; 2; . . . ; N ð2Þ where W 2 Rn�m is a matrix with orthonormal columns. W is chosen to maximize the determinant of the total scatter matrix S of the pro- jected samples. S ¼ XN k¼1 ðXk � lÞðXk � lÞ T ð3Þ W opt ¼ argmaxwjW TSWj ¼ ðw1; w2; . . . ; wmÞ ð4Þ where N is the number of sample images and l is the mean image of all the samples. {wi|i=1, 2, . . . , m} is the set of n-dimensional eigen- vectors of S corresponding to the m largest eigenvalues. In the experiments, the PCA projection method is applied to local facial re- gions instead of the whole face images to extract features. 2.5. GA–ANN technique The GA and ANN-based technique is used to identify the signif- icant areas in each facial region and perform fusion and selection of features for face recognition. In this research, GAs are used to find potential significant features which will generate a higher recogni- tion rate. The areas that contain these significant features are con- sidered to be the significant areas. The chromosomes represent the possible selection of the significant features. Binary encoding is used for the chromosomes, where 1 represents that the feature is selected and 0 represents that the feature is not selected. In one generation, each chromosome is multiplied by the input feature set to generate the input feature vector to ANN. The input feature vector F can be represented as F ¼ CP ð5Þ C ¼ðc1; c2; . . . ; clÞ; ci 2f0; 1g ð6Þ P ¼ L þ R þ N þ M ð7Þ where C is a single chromosome, ci is one gene in the chromosome. l is the length of the chromosome, which is the same as the size of the input feature set P. When testing on the separate feature set from each facial region, P represents the separate feature set. As mentioned in the last section, size of the left eye feature set L is 20, size of the right eye feature set R is 20, size of the nose feature set N is 20 and size of the mouth feature set M is 30. When combin- ing them together, the size of P is 90. Eq. (7) shows the combining feature set P. 7160 X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 The input feature vector F is fed to ANN for classification. An ANN with a single hidden layer is used in this technique. A resilient backpropagation algorithm is used to train the network. The test- ing classification error is used to calculate the fitness of the corre- sponding individual in GA. In the reproduction, the ‘fittest’ individual that achieves the best testing classification rate is re- tained in the next generation. The chromosomes in each genera- tion of GAs that achieve the best classification rate are recorded. The chromosomes indicate which feature is selected and which is not. After all generations, the total number of times that each fea- ture has been selected for the best classification rate is calculated. All the features are ranked according to the number of times they have been selected. The areas that contain the feature inside top n ranking are the top n significant areas. For the experiments, n is de- fined as 3. 3. Databases Three experimental databases were used in this research. All of them are extracted from the FERET benchmark database. The pre- liminary experimental database is a small subset of FERET database and consists of 13 classes (one class represents one distinct per- son). In each class, there are four face images. Three of them are randomly chosen for training and the remainder is chosen for test- ing. The total number of face images in the database is 52. The images are selected carefully in order to have minimum pose var- iation. Fig. 4.1 shows the example images from the preliminary database. Top 3 rows show training images and bottom row shows testing images. The advance databases consist of 50 classes, each class repre- sents one distinct person. In the original dataset (DB1) from our previous study, there are four face images in every class. Three of them are randomly selected for training and one for testing. The extended dataset (DB2) includes all the images from DB1 and more Fig. 4.1. Example images of the preliminary database. The top 3 rows show training images and the bottom row shows testing images. images. In DB2, each class has 4–12 images for training and one for testing. There are totally 376 images for training and 50 images for testing in DB2. 4. Experimental results This section describes the experimental results on small and large databases. The goal of the experiments is to evaluate the pro- posed technique and make comparison with the other existing techniques. All experimental databases are extracted from the FER- ET benchmark database. Section 4.1 presents the experiments which are based on the preliminary database. Section 4.2 describes the advance experiments which are based on larger databases. 4.1. Preliminary results The preliminary experiments were conducted using preliminary database as described in Section 3. Fig. 4.1 shows the example images from the preliminary database. Top 3 rows show training images and bottom row shows testing images. At first, the experiments on location of the significant areas using GA–ANN were conducted on each facial region separately. The features used in the experiments were average grey level va- lue features. The extracted average grey level value features from each facial region formed the input feature vector for that region. The size of the small rectangular area for feature extraction was chosen to be 6 � 4. The size of the extracted feature set L from left eye region was 20. The same size was for the extracted fea- ture set R from right eye region and the extracted feature set N from nose region. For mouth region, the size of the extracted fea- ture set M was 30. L, R, N, M can be expressed in the following equations L ¼ðl1; l2; l3; . . . ; l20Þ; li 2 ð0; 1Þ ð8Þ R ¼ðr1; r2; r3; . . . ; r20Þ; ri 2 ð0; 1Þ ð9Þ N ¼ðn1;n2; n3; . . . ; n20Þ; ni 2 ð0; 1Þ ð10Þ M ¼ðm1; m2; m3; . . . ; m30Þ; ami 2 ð0; 1Þ ð11Þ These extracted feature sets were then fed to GA–ANN separately for selection and classification. To make the experiments consistent, the parameters of GA–ANN were set exactly same for every set of experiments. The generation number was set to 50 and the popula- tion number was set to 15. The crossover rate was set to 0.9 and the mutation rate was set to 0.2. The hidden units of ANN were in- creased from 6 to 44 (each time increased 2 hidden units), and the selections that generated the best recognition rate were re- corded. The epoch for ANN was set to 3000. The best classification results for each facial region feature set are shown in Tables 4.1–4.4. The shadowed cells indicate the high- est testing classification rate. The results given in Tables 4.1–4.4 show that the eye region and the mouth region achieve better recognition rate than the nose region. For the nose region, the best recognition rate is just 76.92% when the hidden units are 34. For left eye region, the best recognition rate is 92.31% when hidden units are 14, 24 and 34. For the right eye region, the best recognition rate is 92.31% when the hidden units are 30 and 44. For mouth region, the best recognition rate is also 92.31% when hidden units are 10 and 36. When achieving the best recognition rate for each facial region, the corresponding feature selection and combinations that were obtained are shown in Table 4.5. The results given in Table 4.5, the left eye region has four different feature combinations, which contain the same feature l1, l8, l9, l20. The right eye region has two different feature combinations, which contain the same feature r2, r3, r4, r5, r7, r9, r10, r16, r19. The mouth region also has two different vahab Highlight vahab Highlight vahab Highlight vahab Highlight vahab Highlight Table 4.1 Best classification results for the left eye feature set Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate [%] RMS Error 1 0 0 0 0 0 0 1 1 0 0 0 0 1 0 1 0 1 0 1 100 92.31 0.04719 14 1 0 0 0 0 0 0 1 1 0 0 0 0 1 0 0 0 1 0 1 97.44 92.31 0.06064 1 1 1 0 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 100 84.62 0.03070 16 1 1 0 0 0 1 1 1 1 0 0 1 1 1 1 1 0 0 0 1 100 84.62 0.03960 24 1 0 0 0 1 1 0 1 1 0 0 1 1 0 1 1 0 0 1 1 100 92.31 0.03910 1 1 1 1 1 0 0 0 0 1 0 0 1 1 1 0 1 1 0 1 100 84.62 0.03758 30 1 1 1 1 1 0 0 0 0 1 0 0 1 0 0 0 1 0 0 1 100 84.62 0.03948 34 1 0 0 0 0 1 1 1 1 0 0 0 1 0 0 1 1 1 1 1 100 92.31 0.03698 Table 4.2 Best classification results for the right eye feature set Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate [%] RMS Error 14 0 0 0 1 1 0 1 0 0 1 0 0 1 0 1 0 0 1 0 1 100 84.62 0.04686 16 0 0 0 1 1 0 1 1 0 1 1 0 0 0 0 0 0 1 1 0 100 84.62 0.05209 1 0 0 1 0 1 1 1 0 1 0 0 1 0 0 1 0 1 1 1 100 84.62 0.04268 24 1 0 0 1 0 1 1 1 0 1 0 1 1 0 0 1 0 1 1 1 100 84.62 0.03669 30 0 1 1 1 1 1 1 0 1 1 1 0 0 0 1 1 0 0 1 1 100 92.31 0.03492 36 0 1 0 1 1 0 0 1 0 1 0 1 0 0 0 1 0 1 0 1 100 84.62 0.04036 Table 4.3 Best classification results for the nose feature set Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate [%] RMS Error 14 1 0 0 0 0 1 0 0 0 0 0 0 1 0 0 1 0 1 1 0 76.92 69.23 0.10628 1 1 1 1 1 0 0 0 0 1 1 0 1 0 0 0 0 0 1 1 92.31 61.54 0.08497 1 1 1 1 0 1 0 1 0 0 0 0 1 1 0 0 1 1 1 1 87.18 61.54 0.0787516 1 1 1 0 1 1 0 0 1 1 1 0 1 1 0 0 1 1 1 1 97.44 61.54 0.06720 0 1 0 0 0 1 0 1 0 0 1 0 1 0 1 0 1 0 0 0 87.18 61.54 0.09534 22 0 1 0 0 0 1 0 1 0 0 1 0 1 0 1 0 1 0 1 0 94.87 61.54 0.08828 32 1 1 0 0 0 0 0 0 1 1 1 0 1 0 0 0 0 0 1 1 89.74 69.23 0.07957 34 1 1 0 1 0 0 0 0 1 1 0 0 1 0 1 1 0 1 1 1 97.44 76.92 0.06574 X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 7161 feature combinations, which contain the same features m2, m3, m10, m12, m16, m20, m22, m29. The nose region just has one feature combination. All feature sets were combined together to feed to GA–ANN again for the experiments. The size of the input feature vector in- creased to 90. The parameters of GA–ANN were set exactly the same as those of the previous experiments. Table 4.6 lists the best classification results achieved. The recognition rate is improved to 100%. When the recognition rate is 100%, these selected features were added together to locate the most selected features. The Table 4.4 Best classification results for the mouth feature set Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate [%] RMS Error 10 011101000111101100011100100011 94.88 92.31 0.056066 14 011110000100100101110000100100 100 84.62 0.048529 18 100011011111011010011000110111 100 84.62 0.03591 111111100111110001111101000111 100 84.62 0.029281 111111100111111100111111100011 100 84.62 0.02807530 111111100111111100110011101111 100 84.62 0.02902 32 011011000000001001010010101111 100 84.62 0.038419 36 011010010101010101110100011010 100 92.31 0.03661 Table 4.5 Feature selections for achieving the best recognition rate Facial region Hidden units Feature selection Left eye region 14 l1, l8, l9, l14, l16, l18, l20 l1, l8, l9, l14, l18, l20 24 l1, l5, l6, l8, l9, l12, l13, l15, l16, l19, l20 34 l1, l6, l7, l8, l9, l13, l16, l17, l18, l19, l20 Right eye region 30 r2, r3, r4, r5, r6, r7, r9, r10, r11, r15, r16, r19, r20 44 r2, r3, r4, r5, r7, r9, r10, r14, r16, r19 Nose region 34 n1, n2, n4, n9, n10, n13, n15, n16, n18, n19, n20 Mouth region 10 m2, m3, m4, m6, m10, m11, m12, m13, m15, m16, m20, m21, m22, m25, m29, m30 36 m2, m3, m5, m8, m10, m12, m14, m16, m18, m19, m20, m22, m26, m27, m29 Table 4.6 Classification results for combined feature set Hidden units Training classification rate (%) Testing classification rate (%) RMS error 8 100 100 0.030384 24 100 100 0.008929 38 100 100 0.009836 44 100 100 0.01825 7162 X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 top 10 selected features are shown in Table 4.7. Most of these fea- tures are concentrated in the eye region and there is no feature coming from the nose region. Table 4.7 Top 10 most selected features Rank Features 1 l20 2 l11, r5 3 m21 4 r1 5 m28 6 r3, r4 7 r19 8 m27 9 l19 10 r6 4.2. Advance experiments The preliminary experiments achieved very good results. This indicates that the proposed technique is promising. Since the pre- liminary database is relatively small, the proposed technique needs to be investigated on much larger databases. Another two databases are set up to conduct the experiments, referred to as Databases 1 and 2. Same as the preliminary database, both Dat- abases 1 and 2 are extracted from the FERET database and consist of 50 classes each. In Database 1, there are 150 face images for training and 50 face images for testing. Database 2 includes all the images from Database 1 and increases the training set. In Database 2, there are totally 376 face images for training and 50 face images for testing. Section 4.2.1 presents the experimen- tal results from Database 1 and Section 4.2.2 presents the results from Database 2. 4.2.1. Database 1 results There are four face images per class in Database 1. Three of them are randomly selected for training and the left one is se- lected for testing. Example images from Database 1 could be found in Fig. 4.2. Two different sets of experiments were con- ducted on Database 1. In the first set of experiments, the average grey level value features were investigated. In the second set of experiments, the PCA features were investigated. Section 4.2.1.1 describes the experiments using average grey level value features. The experiments using PCA features are explained in Section 4.2.1.2. 4.2.1.1. Average grey level value features. For the experiments using average grey level value features, two different sizes of small rect- angular areas for feature extraction were investigated. In the experiments, the size of the small rectangular area was firstly set to 6 � 4 and then set to 10 � 4. Section Section 4.2.1.1.1 presents the results when the size of the small rectangular area is 6 � 4. 4.2.1.2. Small rectangular area size is 6 � 4. When the size of the small rectangular area for feature extraction is 6 � 4, the GA– ANN technique was firstly tested on each facial region feature set separately. During the experiments, the hidden units of ANN were increased from 8 to 64 (an increment of 4 hidden units each time), and the selections that generated the best recognition rate were recorded. To make the experiments consistent, the other parameters of GA–ANN were set exactly same for every experi- ment. The generation number was set to 40 and the population Fig. 4.2. Example images from Database 1. Top three rows show training images and bottom row shows testing images. X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 7163 number was set to 10. The crossover rate was set to 0.9 and the mutation rate was set to 0.2. The epoch for ANN was set to 10,000. Tables 4.8–4.11 list the best classification results achieved for each facial region feature set. In the above tables, the shadowed cells indicate the highest testing classification rate. The highest testing classification rate for the left eye feature set is 54% (hidden units 44, 56, 60), for the right eye feature set is 62% (hidden units 28), for the nose feature set is 38% (hidden units 52) and for the mouth feature set is 70% (hidden units 36). The results given in Tables 4.8– 4.11 show that the mouth region alone achieved the best classi- fication rate, while the nose region achieved the worst classifica- tion rate. The extracted average grey level value features from each facial region were combined together to form the input feature Table 4.8 Best classification results for the left eye feature set Classification Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testin [ 12 1 0 0 0 1 1 1 1 1 1 1 1 0 0 0 1 1 1 0 0 82 32 1 1 0 1 1 0 0 1 1 0 0 1 1 1 0 1 1 0 0 0 100 40 1 1 1 1 1 0 0 0 1 0 1 0 0 1 1 1 0 1 1 0 100 44 1 1 1 1 1 0 0 0 0 1 0 0 1 1 1 1 1 1 0 1 100 48 1 1 1 1 1 0 0 0 0 1 0 0 1 1 1 0 1 1 0 1 100 56 1 1 1 1 1 1 1 1 0 1 1 0 1 1 1 0 1 1 1 0 100 60 1 1 1 1 1 0 1 1 1 0 1 1 1 1 1 0 1 1 0 1 100 64 1 1 1 1 1 0 0 0 0 1 0 0 1 1 1 0 1 1 0 1 100 vector for GA–ANN. The size of the input feature vector increased to 90. The other parameters of GA–ANN were set exactly the same as those of the experiments using the separate facial feature set. The feature combination sequence was left eye, right eye, nose and mouth. Table 4.12 lists the results of the combined feature set that achieved above 80% testing classification rate on Database1. The results given in the table show that the best testing classification rate is 86% (hidden units 40, 56), and the best training classification rate is 100%. The shadowed cells indicate the best testing classifi- cation rate. The combined feature set outperformed the separate feature set from each facial region and improved the classification rate significantly. To investigate the effects of feature combination sequence on the recognition rate, the feature combination sequence was re- versed to mouth, nose, left eye and right eye to form a new input vector. Then the same experiments under the same parameters were conducted. Table 4.13 lists the best classification results of the combined feature set in the reverse order. The best recognition rate is still 86% (hidden units 64). When the recognition rate is 86% (hidden units 40, 56) for the combined feature set in the original order, the total selection times of each feature were calculated. By mapping the total selection times of each feature to its corresponding extraction area, Fig. 4.3 is generated. In Fig. 4.3, the shaded areas are the areas that contain the top selected features. These areas are considered to be the significant areas. There are totally 36 areas. Among these areas, there are 9 areas from the left eye region, 7 areas from right eye region, 5 areas from the nose region and 15 areas from the mouth region. The results in Table 4.13 show that the best recognition rate is still 86% when the feature combination sequence is reversed. When the recognition rate is 86% (hidden units 64), the total selec- tion times of each feature were calculated. Similarly, by mapping the total selection times of each feature to its corresponding extraction area, Fig. 4.4 is generated. In Fig. 4.4, the shaded areas are the areas that contain the top selected features. There are total 49 areas. Among these areas, there are 7 areas from left eye region, 12 areas from right eye region, 11 areas from nose region and 19 areas from mouth region. 4.2.1.3. PCA features. PCA features are extracted separately from each facial region, and then combined together to form the input feature vector for GA–ANN. After feature extraction, the sequence of feature combination is left eye, right eye, nose and mouth. Because we do not know how many eigenvectors should be suitable for encoding the face images, a different number of Rate g Rate %] RMS Error 48 0.090757 52 0.065463 48 0.055730 54 0.055398 50 0.055502 54 0.04531 54 0.045535 52 0.050977 Table 4.11 Best classification results for the mouth feature set Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate [%] RMS Error 32 101101110100011010100100100110 100 62 0.055177 36 111111111110111001011000110010 100 70 0.047729 40 111111001110011010000100100010 100 66 0.048634 44 001110110101100101100100101010 100 62 0.047705 48 001110110100011010100100011010 100 66 0.048532 111111110100011010100100100101 100 66 0.042634 52 111111110010011010100100111010 100 66 0.04244 56 110001001101101110011000110010 100 68 0.042558 001101001111101010101001010110 100 62 0.04191 001101001111101001010101010110 100 62 0.041762 001101010011101001010101010110 100 62 0.043019 60 001010101111101001010101010110 100 62 0.0413 64 101001110000010101101000111101 100 64 0.042305 Table 4.10 Best classification results for the nose feature set Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] TestingR ate[%] RMS Error 20 1 0 1 0 0 0 1 1 1 1 1 0 1 1 0 1 1 1 0 1 82 34 0.083747 32 1 1 1 1 1 1 1 0 1 0 0 1 0 1 0 1 0 1 1 1 94.67 32 0.074111 36 1 1 1 0 1 0 1 0 1 1 1 0 0 0 1 0 1 1 1 0 91.33 32 0.078527 44 1 0 0 1 0 1 1 0 0 0 0 1 0 0 1 0 0 1 0 1 87.33 34 0.087115 48 1 1 0 1 0 0 0 0 0 0 0 0 1 1 0 0 1 0 1 1 84.67 30 0.084655 52 1 1 1 1 1 1 1 1 0 1 1 0 1 1 1 0 1 1 0 1 98.67 38 0.06674 60 1 0 0 1 0 1 1 0 0 0 1 0 0 0 1 0 1 1 0 1 88 34 0.081849 64 1 0 1 1 1 0 0 0 0 1 0 0 1 1 1 0 1 1 0 1 97.33 32 0.075388 Table 4.9 Best classification results for the right eye feature set Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate[%] RMS Error 16 1 1 1 1 1 0 0 1 0 1 0 0 1 1 1 0 1 1 0 1 100 56 0.075269 24 1 1 1 1 1 0 0 0 0 1 1 1 0 0 0 1 0 1 0 1 98 56 0.067813 28 1 1 1 1 1 0 0 0 1 1 0 0 0 1 0 1 0 0 0 0 98.67 62 0.067738 36 1 1 1 1 1 0 0 0 0 1 0 0 1 0 0 1 0 0 1 0 99.33 60 0.061682 44 1 1 1 1 1 0 0 0 0 1 0 0 1 1 0 1 0 1 0 1 100 56 0.052249 52 1 1 0 1 1 0 1 1 0 1 0 0 0 0 0 1 0 0 0 1 99.33 60 0.058238 60 1 1 1 1 1 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 99.33 58 0.062595 64 1 1 1 1 1 0 0 1 0 1 0 0 0 0 0 1 0 0 0 1 99.33 58 0.057562 7164 X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 Table 4.13 Best classification results for combined feature set in the reverse order Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate[%] Testing Rate[%] RMS Error 28 10100101111111110100011011001110101 10101001000001001101100001000011110111 00111111111101101 100 80 0.038631 40 00111011010001101010010000101101100 01111101000000001001000100001101101000 00101110000011100 100 80 0.033954 52 00011010110001101010010010010101011 00011110000010110001100100110110001001 01101111000001101 100 84 0.023553 56 10011101011010010101011100001010011 10100111010011110110100100001101111110 00001000000010101 100 82 0.021568 60 00111101011010101010100011110101100 01011100101100110110111011110010010111 11011110111100001 100 80 0.017561 11011101011010100011110011101101110 00110011010011110111011011001101101000 00100110000011110 100 86 0.016062 64 11011101011010101011110011101101110 00110011010011110110111011001101101000 00100110000011110 100 86 0.015144 01100010100101010101011100001010011 10100100101100001001000100001101101000 00100001111100001 100 82 0.01995 68 10011101011010101010100011110101100 01011011010011110110111011110010010111 11011110000011110 100 82 0.015121 Table 4.12 Best classification results for combined feature set in original order Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate [%] RMS Error 1101000001001000001111001110110111 0001011011100100001011011010000111101 1111101101101110111 100 80 0.03366 36 1101000001001000001111001110110111 0001011011100101110100100101111000010 0000010010010001111 100 80 0.03500 40 0010111110100111110000111100111101 0010000101000110011111001101101001101 1100111111111101101 100 86 0.02964 48 1000101100001101001111001110110111 0001011011100101110100100101111000010 0000010010010001000 100 82 0.02819 56 1010111111011011101111001110110111 1110100100011010001010010101111100110 0111110111111101010 100 86 0.01939 60 0011110000110001110110100110111111 1100001101000101100011001001011010100 0001001010011011111 100 84 0.02054 X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 7165 eigenvectors was evaluated in the experiments. The experiments using 10 eigenvectors, 14 eigenvectors, 18 eigenvectors and 22 eigenvectors were conducted. The parameters of GA–ANN were set exactly same for every experiment. The generation number was set to 40 and the population number was set to 10. The cross- over rate was set to 0.9 and the mutation rate was set to 0.2. The epoch for ANN was set to 10000. The hidden units were increased from 8 to 68 (an increment of 4 hidden units each time). Table 4.15 Best classification results for 14 eigenvectors Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate [%] RMS Error 00001100101100110100101100110011100 001100111111100000100 100 68 0.059964 20 11110111101100110100101100110011100 001100111111100011011 100 68 0.055764 40 00001100101100110100101100110011100 001100111111100101001 100 74 0.038337 48 00001100101100110100101100110111100 001100111111100000100 100 76 0.034414 60 00001100101100110100101011110111100 001100111111101100001 100 78 0.026711 64 00001100101100110101011100010011100 001100111111100010010 100 78 0.027607 Table 4.14 Best classification results for 10 eigenvectors Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate [%] RMS Error 0000001011100100110000101101111010 111001 100 60 0.080433 12 1111110101001010001111001110001100 110011 98.67 60 0.076291 36 0000001011100100101100110100101100 110011 100 64 0.043994 1000011001111111000001000001011000 110010 100 62 0.035194 52 1000011001111111111001000001011000 110010 100 62 0.033143 56 0000001011100100101100110100101100 110011 100 66 0.033341 Fig. 4.3. Significant areas in facial regions (original order combination). Fig. 4.4. Significant areas in facial regions (reverse order combination). 7166 X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 Tables 4.14–4.16 present the best classification results for a dif- ferent number of eigenvectors. The results given in Table 4.14 show that the best testing classification rate for 10 eigenvectors is 66% when the hidden units are 56. The corresponding best train- ing classification rate is 100%. Table 4.15 shows that the best testing classification rate for 14 eigenvectors is 78% when the hidden units are 60 and 64. The cor- responding best training classification rate is 100%. Table 4.16 shows that the best testing classification rate for 18 eigenvectors is 80% when the hidden units are 60. The correspond- ing best training classification rate is 100%. 4.2.2. Database 2 results In Database 2, each class has 4–12 images for training and one for testing. Because the combined feature set (using average grey level value features) achieved much better results on Database 1, Table 4.17 Best classification results (Database 2) of hidden units 40 feature selection (Database 1) Hidden Units Training Classification Rate Testing Classification Rate 30 100% 88% 42 100% 94% 48 100% 88% 52 100% 90% 54 100% 92% 60 100% 90% 66 100% 88% 74 100% 94% 76 100% 90% 82 100% 88% 86 100% 94% Table 4.16 Best classification results for 18 eigenvectors Classification Rate Hidden Units Feature Selection (GA Chromosomes) Training Rate [%] Testing Rate [%] RMS Error 16 11011110101100101010101111110010000 1010001000110101000110110100000000000 100 70 0.064818 32 11111001101111101010100100111101111 0101110110110101000110010011101010100 100 72 0.040375 40 00001000111010101010111010111010000 0110100000101100000111100111011011100 100 74 0.036662 48 11101100011010101010100101001010011 0110100000101100000111100111011011100 100 68 0.030918 56 10010111100000111101011010111010000 0110100000101100000111100111011011100 100 74 0.026269 60 10101011101010101001001010111110000 0110100000110011011100111111111101101 100 80 0.024114 64 00010011100101010101011010111010000 0110100000101100000110110011010100001 100 74 0.027777 Table 4.18 Best classification results (Database 2) of hidden units 56 feature selection (Database 1) Hidden Units Training Classification Rate Testing Classification Rate 26 100% 88% 38 100% 92% 44 100% 92% 54 100% 90% 58 100% 94% 68 100% 90% 72 100% 92% 78 100% 88% 80 100% 88% 88 100% 90% X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 7167 so only the combined feature set (using average grey level value features) experiments were conducted on Database 2. To make the experiments faster, the best feature selection from the previous experiments on Database 1 was used directly to train and test the ANN. The epoch was increased to 15,000 because there are more face images in the database. More hidden units were also used in the experiments. When the size of the small rectangular area was 6 � 4, the hid- den units 40 and 56 feature selections achieved the best recogni- tion rate on Database 1. These two feature selections were directly used in the experiments on Database 2. The results based on the hidden units 40 feature selection are listed in Table 4.17. The results based on the hidden units 56 feature selection are pre- sented in Table 4.18. The results given in both tables, the highest recognition rate is improved to 94%. When the size of small rectangular area was 10 � 4, the hidden units 44 feature selection achieved the best recognition rate on Database 1. Table 4.19 shows the best classification results using hidden units 44 feature selection on Database 2. The results given Table 4.19 Best classification results of hidden units 44 feature selection Hidden Units Training Classification Rate Testing Classification Rate 28 99.20% 90% 30 99.73% 88% 34 100% 88% 36 99.47% 90% 38 100% 94% 42 100% 90% 50 99.73% 90% 52 100% 92% 54 100% 94% 56 100% 94% 62 100% 90% 64 100% 92% 66 100% 92% Table 5.2 Combined feature set results on DB1 (Database 1) Hidden Units Training Rate [%] Testing Rate [%] 40 100 86 44 100 82 48 100 82 56 100 86 60 100 84 64 100 84 68 100 82 Table 5.3 Combined feature set results on DB2 (Database 2) Hidden Units Training Rate [%] Testing Rate [%] 26 99.73 88 30 100 88 38 100 88 42 100 94 50 100 90 7168 X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 in Table 4.19 show that the highest recognition rate is also im- proved to 94%. 5. Comparative analysis The results obtained in this research are compared to the results of the other methods mentioned in a recent study (Zhang et al., 2004). The authors (Zhang et al., 2004) also extracted a dataset from the FERET database as their experimental database, which has 927 images corresponding to 119 persons. Three different methods are experimented on this dataset. These methods include kernel associative memory (kAM) method which is proposed in their study, PCA-nearest-neighbor method and a simple NN-based template matching method termed ARENA. In this study, we com- pared with the highest classification rate achieved in their (Zhang et al., 2004) study. They conducted two sets of experiments similar as in our research: the first one used 3 images per class for training and the second one used 4 images per class for training. Fig. 5.1 shows the comparison of the best recognition rates be- tween our DB1 (Database 1) experimental results and their first set results. Fig. 5.2 shows the comparison of the best recognition rates between our DB2 (Database 2) experimental results and their sec- ond set results. Both figures show that our approach achieves a better recognition rate. 6. Conclusions We have presented a feature selection and fusion technique for face recognition in this paper. The GA for feature selection and ANN for feature classification are incorporated in the proposed tech- nique. The technique performs fusion and selection of facial fea- tures for face recognition. The significant areas inside each facial region are located through the feature selection. The FERET benchmark database is adopted to evaluate and com- pare the proposed technique with the existing techniques. Three different databases were used in the experimental investigation and all of them were extracted from the FERET benchmark data- base. Database 2 is the largest database, containing 50 classes and 426 face images. The experiments are conducted on Cluster machine at Central Queensland University. The preliminary experiments were conducted simply to pre-test the proposed technique. The experiments investigated the separate facial region feature set and the combined feature set using the average grey level value features. The preliminary results were Table 5.1 Separtate facial region feature set results on DB1 Hidden Units Training Rate [% 44 100 56 100 60 100 Left Eye Region 64 100 20 97.33 28 98.67 36 99.33 Right Eye Region 52 99.33 20 82 44 87.33 52 98.67 Nose Region 56 96.67 36 100 40 100 48 100 Mouth Region 56 100 promising. The left eye feature set, right eye feature set and mouth feature set all achieved the highest recognition rate of 92.31%. The nose feature set just achieved the highest recognition rate 76.92% and was the worst performer. The combined feature set outper- formed the separate facial region feature set by improving the rec- ognition rate to 100%. On Database 1, many experiments were conducted to perform further investigation. The different size of the feature extraction area, the different feature extraction tech- nique and the different sequence for feature combination were considered in the experimental investigation. When average grey level value features were used and the size of the small rectangular area was 6 � 4, the left eye feature set achieved 54% recognition rate, right eye feature set achieved 62% recognition rate, nose fea- ture set achieved 38% recognition rate and mouth feature set achieved 70% recognition rate (see Table 5.1). The mouth feature set was the best performer and the nose feature set was the worst performer. The combined feature set outperformed the separate fa- cial region feature set by achieving 86% recognition rate. The com- bination sequence did not affect the recognition rate for combined feature set. For significant areas, the mouth region contributed the ] Testing Rate [%] 54 54 54 52 60 62 60 60 34 34 38 36 70 66 66 68 20% 30% 40% 50% 60% 70% 80% 90% 100% Left Eye Right Eye Nose Mouth Combined Feature Set C la ss if ic at io n R at e Training Rate Testing Rate Fig. 5.1. Classification rate comparison between different feature sets. 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% PCA ARENA kAM GA-ANN Approaches R ec og ni ti on R at e Fig. 5.2. Comparison with other approaches. (3 images per class for training set). 40.00% 45.00% 50.00% 55.00% 60.00% 65.00% 70.00% 75.00% 80.00% 85.00% 90.00% 95.00% 100.00% PCA ARENA kAM GA-ANN Approach R ec og ni ti on R at e Fig. 5.3. Comparison with other approaches. (4 or more images per class for training set). X. Fan, B. Verma / Expert Systems with Applications 36 (2009) 7157–7169 7169 most and the nose region contributed the least. When the size of the small rectangular area was increased to 10 � 4, the left eye fea- ture set achieved 54% recognition rate, the right eye feature set achieved 56% recognition rate, the nose feature set achieved 36% recognition rate and the mouth feature set achieved 64%. The mouth feature set was still the best performer and the nose feature set was still the worst performer. The combined feature set im- proved the recognition rate to 86% when compared to the separate facial region feature set (see Table 5.2). The combination sequence slightly affected the recognition rate. The original order combina- tion achieved of 84% recognition rate while the reverse order com- bination achieved slightly higher recognition rate 86%. For significant areas, the mouth region still contributed the most and the nose region contributed the least. The above results indicate the mouth region is the most important facial region. Combination of facial features from each facial region is much more useful in improving recognition rate compared to just one facial region fea- ture set. A different number of eigenvectors was used for PCA fea- tures experiments. The 18 eigenvectors achieved the highest recognition rate 80%. The average grey level value features (com- bined feature set) outperformed the PCA features by 6%. The experiments on Database 2 were conducted by just using the combined feature set of average grey level value features. The recognition rate was improved to 94% (see Table 5.3). The experimental results of the proposed approach were also com- pared with the results of the other three approaches based on FER- ET database: PCA, ARENA and kAM. Fig. 5.2 was based on Database 1 results and the proposed approach improved the recognition rate by 1.3% compared to kAM method, 36% compared to PCA method and 41% compared to ARENA method. Fig. 5.3 was based on Data- base 2 results and the proposed technique improved the recogni- tion rate by 2.4% compared to kAM method, 43.2% compared to PCA method and 48.3% compared to ARENA method. The proposed technique achieved the highest recognition rate among the exist- ing techniques based on FERET database. References Bala, J., Huang, J., Vafaie, H., DeJong, K., & Wechsler, H. (1995). Hybrid learning using genetic algorithms and decision trees for pattern classification. Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, 1, 719–724. Belhumeur, P. N., Hespanha, J. P., & Kriegman, D. J. (1997). Eigenfaces vs. fisherfaces: Recognition using class specific linear projection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19, 711–720. Brunelli, R., & Poggio, T. (1992). Face recognition through geometrical features. Proceedings of ECCV, 92, 792–800. Chellappa, R., Wilson, C. L., & Sirohey, S. (1995). Human and machine recognition of faces: A survey. Proceedings of the IEEE, 83, 705–740. Fang, Y., Tan, T., & Wang, Y. (2002). Fusion of global and local features for face verification. IEEE International Conference on Pattern Recognition, 2, 382–385. Li, S. Z., & Lu, J. (1999). Face recognition using the nearest feature line method. IEEE Transactions on Neural Networks, 10, 439–443. Liu, Q., Tang, X., Lu, H., & Ma, S. (2004). Kernel scatter-difference based discriminant analysis for face recognition. In International 17th conference on pattern recognition (Vol. 2, pp. 419–422). Liu, C., & Wechsler, H. (2000). Evolutionary pursuit and its application to face recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(6), 570–582. Liu, C., & Wechsler, H. (2003). Independent component analysis of Gabor features for face recognition. IEEE Transactions on Neural Networks, 14, 919–928. Phillips, P. J., Wechsler, H., Huang, J., & Rauss, P. (1998). The FERET database and evaluation procedure for face recognition algorithms. Image and Vision Computing, 16(5), 295–306. Sun, Y., & Yin, L. (2005). A genetic algorithm based feature selection approach for 3D face recognition. In Biometric consortium conference. USA. Turk, M., & Pentland, A. (1991). Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3, 71–86. Wiskott, L., Fellous, J. M., & Malsburg, C. (1997). Face recognition by elastic bunch graph matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19, 775–779. Yan, S., He, X., Hu, Y., Zhang, H., Li, M., & Cheng, Q. (2004). Bayesian shape localization for face recognition using global and local textures. IEEE Transactions on Circuits and Systems for Video Technology, 1(14), 102–113. Zhang, B., Zhang, H., & Ge, S. (2004). Face recognition by applying wavelet subband representation and kernel associative memory. IEEE Transactions on Neural Networks, 15, 166–177. Selection and fusion of facial features for face recognition Introduction Background Literature Reviewreview Proposed Techniquetechnique Overview Locate facial regions Average grey level value feature PCA feature GA-ANN GA-ANN technique Databases Experimental Resultsresults Preliminary Resultsresults Advance Experimentsexperiments Database1 ResultsDatabase 1 results Average Grey Level Value Featuresgrey level value features Small rectangular area size is 6 times 4 PCA Featuresfeatures Database 2 Resultsresults Comparative Analysisanalysis Conclusions References 
work_iiconkps45cr5gmnah3zudj2e4	----	Example-Dependent Cost-Sensitive Decision Trees Alejandro Correa Bahnsen, Djamila Aouada, Björn Ottersten Interdisciplinary Centre for Security, Reliability and Trust, University of Luxembourg Abstract Several real-world classification problems are example-dependent cost-sensi- tive in nature, where the costs due to misclassification vary between ex- amples. However, standard classification methods do not take these costs into account, and assume a constant cost of misclassification errors. In this paper, we formalize a new measure in order to define when a prob- lem is cost-insensitive, class-dependent cost-sensitive or example-dependent cost-sensitive. Afterwards, we propose an example-dependent cost-sensitive decision tree algorithm, by incorporating the different example-dependent costs into a new cost-based impurity measure and a new cost-based prun- ing criteria. Using three different databases, from three real-world applica- tions: credit card fraud detection, credit scoring and direct marketing, we evaluate the proposed method against state-of-the-art example-dependent cost-sensitive techniques, specifically, cost-proportionate sampling and Bayes minimum risk. The results show that the proposed algorithm is the best per- forming method for all databases. Furthermore, when compared against a standard decision tree, our method builds significantly smaller trees in only Email addresses: alejandro.correa@uni.lu (Alejandro Correa Bahnsen), djamila.aouada@uni.lu (Djamila Aouada), bjorn.ottersten@uni.lu (Björn Ottersten) Preprint submitted to Expert Systems with Applications October 13, 2014 a fifth of the time, while having a superior performance measured by cost savings. Keywords: Cost-sensitive learning, Cost-Sensitive Classifier, Credit scoring, Fraud detection, Direct marketing, Decision trees 1. Introduction Classification, in the context of machine learning, deals with the problem of predicting the class of a set of examples given their features. Traditionally, classification methods aim at minimizing the misclassification of examples, in which an example is misclassified if the predicted class is different from the true class. Such a traditional framework assumes that all misclassifi- cation errors carry the same cost. This is not the case in many real-world applications. Methods that use different misclassification costs are known as cost-sensitive classifiers. Typical cost-sensitive approaches assume a constant cost for each type of error, in the sense that, the cost depends on the class and is the same among examples (Elkan, 2001; Kim et al., 2012). Although, this class-dependent approach is not realistic in many real-world applications, for example in credit card fraud detection, failing to detect a fraudulent trans- action may have an economical impact from a few to thousands of Euros, depending on the particular transaction and card holder (Sahin et al., 2013). In churn modeling, a model is used for predicting which customers are more likely to abandon a service provider. In this context, failing to identify a profitable or unprofitable churner have a significant different financial im- pact (Glady et al., 2009). Similarly, in direct marketing, wrongly predicting that a customer will not accept an offer when in fact he will, has a different 2 impact than the other way around (Zadrozny et al., 2003). Also in credit scoring, where declining good customers has a non constant impact since not all customers generate the same profit (Verbraken et al., 2014). Lastly, in the case of intrusion detection, classifying a benign connection as malicious have a different cost than when a malicious connection is accepted (Ma et al., 2011). In order to deal with these specific types of cost-sensitive problems, called example-dependent cost-sensitive, some methods have been proposed re- cently. However, the literature on example-dependent cost-sensitive methods is limited, mostly because there is a lack of publicly available datasets that fit the problem (Aodha and Brostow, 2013). Standard solutions consist in mod- ifying the training set by re-weighting the examples proportionately to the misclassification costs (Elkan, 2001; Zadrozny et al., 2003). Recently, a direct cost approach has been proposed to make the classification decision based on the expected costs. This method is called Bayes minimum risk (BMR), and has been successfully applied to detect credit card fraud (Correa Bahnsen et al., 2013, 2014c). The method consists in quantifying tradeoffs between various decisions using probabilities and the costs that accompany such deci- sions. Nevertheless, these methods still use a cost-insensitive algorithm, and either by modifying the training set or the output probabilities convert it into a cost-sensitive classifier. Another way to introduce the cost-sensitivity into the algorithms is by modifying the methods. In particular this has been done for decision tree (Lomax and Vadera, 2013). However, the approaches that have been proposed only deal with the problem when the cost depends on the class and not on the example (Draper et al., 1994; Ting, 2002; Ling 3 et al., 2004; Li et al., 2005; Kretowski and Grześ, 2006; Vadera, 2010). In this paper we formalize a new measure in order to define when a prob- lem is cost-insensitive, class-dependent cost-sensitive or example-dependent cost-sensitive. Moreover, we go beyond the aforementioned state-of-the-art methods, and propose a decision tree algorithm that includes the example- dependent costs. Our approach is based first on a new example-dependent cost-sensitive impurity measure, and secondly on a new pruning improvement measure which also depends on the cost of each example. We evaluate the proposed example-dependent cost-sensitive decision tree using three different databases. In particular, a credit card fraud detection, a credit scoring and a direct marketing databases. The results show that the proposed method outperforms state-of-the-art example-dependent cost- sensitive methods. Furthermore, when compared against a standard decision tree, our method builds significantly smaller trees in only a fifth of the time. Furthermore, the source code used for the experiments is publicly available as part of the CostSensitiveClassification 1 library. The remainder of the paper is organized as follows. In Section 2, we ex- plain the background behind example-dependent cost-sensitive classification and we define a new formal definition of cost-sensitive classification problems. In Section 3, we make an extensive review of current decision tree methods, including by the different impurity measures, growth methods, and pruning techniques. In Section 4, we propose a new example-dependent cost-sensitive decision tree. The experimental setup and the different datasets are described 1https://github.com/albahnsen/CostSensitiveClassification 4 in Section 5. Subsequently, the proposed algorithm is evaluated on the dif- ferent datasets. Finally, conclusions of the paper are given in Section 7. 2. Cost-Sensitive Cost Characteristic and Evaluation Measure In this section we give the background behind example-dependent cost- sensitive classification. First we present the cost matrix, followed by a formal definition of cost-sensitive problems. Afterwards, we present an evaluation measure based on cost. Finally, we describe the most important state-of- the-art methods, namely: Cost-proportionate sampling and Bayes minimum risk. 2.1. Binary classification cost characteristic In classification problems with two classes yi ∈{0, 1}, the objective is to learn or predict to which class ci ∈{0, 1} a given example i belongs based on its k features Xi = [x 1 i ,x 2 i , ...,x k i ]. In this context, classification costs can be represented using a 2x2 cost matrix (Elkan, 2001), that introduces the costs associated with two types of correct classification, true positives (CTPi ), true negatives (CTNi ), and the two types of misclassification errors, false positives (CFPi ), false negatives (CFNi ), as defined below: Actual Positive Actual Negative yi = 1 yi = 0 Predicted Positive CT Pi CF Pi ci = 1 Predicted Negative CF Ni CT Ni ci = 0 Table 1: Classification cost matrix 5 A classification problem is said to be cost-insensitive if costs of both errors are equal. It is class-dependent cost-sensitive if the costs are different but constant. Finally we talk about an example-dependent cost-sensitive classification problem if the cost matrix is not constant for all the examples. However, the definition above is not general enough. There are many cases when the cost matrix is not constant and still the problem is cost- insensitive or class-dependent cost-sensitive. For example, if the costs of correct classification are zero, CTPi = CTNi = 0, and the costs of misclassifi- cation are CFPi = a0 ·zi and CFNi = a1 ·zi, where a0, a1, are constant and zi a random variable. This is an example of a cost matrix that is not constant. However, C∗FNi and C ∗ TPi are constant, i.e. C∗FNi = (a1 · zi)/(a0 · zi) = a1/a0 and C∗TPi = 0 ∀i. In this case the problem is cost-insensitive if a0 = a1, or class-dependent cost-sensitive if a0 6= a1, even given the fact that the cost matrix is not constant. Nevertheless, using only the simpler cost matrix is not enough to define when a problem is example-dependent cost-sensitive. To achieve this, we define the classification problem cost characteristic as bi = C ∗ FNi −C∗TPi, (1) and define its mean and standard deviation as µb and σb, respectively. Using µb and σb, we analyze different binary classification problems. In the case of a cost-insensitive classification problem, for every example i CFPi = CFNi and CTPi = CTNi , leading to bi = 1 ∀i or more generally µb = 1 and σb = 0. For class-dependent cost-sensitive problems, the costs are not equal but constants CFPi 6= CFNi or CTPi 6= CTNi , leading to bi 6= 1 ∀i, or µb 6= 1 and σb = 0. Lastly, in the case of example-dependent cost-sensitive 6 problems, the cost difference is non constant or σb 6= 0. In summary a binary classification problem is defined according to the following conditions: µb σb Type of classification problem 1 0 cost-insensitive 6= 1 0 class-dependent cost-sensitive 6= 0 example-dependent cost-sensitive 2.2. Example-dependent cost-Sensitive evaluation measures Common cost-insensitive evaluation measures such as misclassification rate or F −Score, assume the same cost for the different misclassification errors. Using these measures is not suitable for example-dependent cost- sensitive binary classification problems. Indeed, two classifiers with equal misclassification rate but different numbers of false positives and false neg- atives do not have the same impact on cost since CFPi 6= CFNi ; therefore there is a need for a measure that takes into account the actual costs Ci = [CTPi,CFPi,CFNi,CTNi ] of each example i, as introduced in the previous sec- tion. Let S be a set of N examples i, N = |S|, where each example is repre- sented by the augmented feature vector Xai = [Xi,Ci] and labelled using the class label yi ∈ {0, 1}. A classifier f which generates the predicted label ci for each element i is trained using the set S. Then the cost of using f on S 7 is calculated by Cost(f(S)) = N∑ i=1 ( yi(ciCTPi + (1 − ci)CFNi )+ (2) (1 −yi)(ciCFPi + (1 − ci)CTNi ) ) . (3) Moreover, by evaluating the cost of classifying all examples as the class with the lowest cost Costl(S) = min{Cost(f0(S)),Cost(f1(S))} where f0 refers to a classifier that predicts all the examples in S as belonging to the class c0, and similarly f1 predicts all the examples in S as belonging to the class c1, the cost improvement can be expressed as the cost savings as compared with Costl(S). Savings(f(S)) = Costl(S) −Cost(f(S)) Costl(S) . (4) 2.3. State-of-the-art example-dependent cost-sensitive methods As mentioned earlier, taking into account the different costs associated with each example, some methods have been proposed to make classifiers example-dependent cost-sensitive. These methods may be grouped in two categories. Methods based on changing the class distribution of the training data, which are known as cost-proportionate sampling methods; and direct cost methods (Wang, 2013). A standard method to introduce example-dependent costs into classifica- tion algorithms is to re-weight the training examples based on their costs, either by cost-proportionate rejection-sampling (Zadrozny et al., 2003), or over-sampling (Elkan, 2001). The rejection-sampling approach consists in selecting a random subset Sr by randomly selecting examples from S, and 8 accepting each example i with probability wi/ max 1,...,N {wi}, where wi is defined as the expected misclassification error of example i: wi = yi ·CFNi + (1 −yi) ·CFPi. (5) Lastly, the over-sampling method consists in creating a new set So, by mak- ing wi copies of each example i. However, cost-proportionate over-sampling increases the training since |So| >> |S|, and it also may result in over-fitting (Drummond and Holte, 2003). Furthermore, none of these methods uses the full cost matrix but only the misclassification costs. In a recent paper, we have proposed an example-dependent cost-sensitive Bayes minimum risk (BMR) for credit card fraud detection (Correa Bahnsen et al., 2014c). The BMR classifier is a decision model based on quantifying tradeoffs between various decisions using probabilities and the costs that accompany such decisions (Jayanta K. et al., 2006). This is done in a way that for each example the expected losses are minimized. In what follows, we consider the probability estimates pi as known, regardless of the algorithm used to calculate them. The risk that accompanies each decision is calculated. In the specific framework of binary classification, the risk of predicting the example i as negative is R(ci = 0|Xi) = CTNi (1 − p̂i) + CFNi · p̂i, and R(ci = 1|Xi) = CTPi · p̂i + CFPi (1 − p̂i), is the risk when predicting the example as positive, where p̂i is the estimated positive probability for example i. Subsequently, if R(ci = 0|Xi) ≤ R(ci = 1|Xi), then the example i is classified as negative. This means that the risk associated with the decision ci is lower than the risk associated with classifying it as positive. However, when using the output of a binary classifier as a basis for decision making, there is a need for a probability 9 that not only separates well between positive and negative examples, but that also assesses the real probability of the event (Cohen and Goldszmidt, 2004). 3. Decision trees Decision trees are one of the most widely used machine learning algo- rithms Maimon (2008). The technique is considered to be white box, in the sense that is easy to interpret, and has a very low computational cost, while maintaining a good performance as compared with more complex techniques Hastie et al. (2009). There are two types of decision tree depending on the objective of the model. They work either for classification or regression. In this section we focus on binary classification decision tree. 3.1. Construction of classification trees Classification trees is one of the most common types of decision tree, in which the objective is to find the Tree that best discriminates between classes. In general the decision tree represents a set of splitting rules orga- nized in levels in a flowchart structure. In the Tree, each rule is shown as a node, and it is represented as (Xj, lj), meaning that the set S is split in two sets Sl and Sr according to Xj and lj: Sl = {Xai |X a i ∈ S ∧x j i ≤ l j} and Sr = {Xai |X a i ∈ S ∧x j i > l j}, (6) where Xj is the jth feature represented in the vector Xj = [x j 1,x j 2, ...,x j N ] and lj is a value such that min(Xj) ≤ lj < max(Xj). The Tree is constructed by testing all possible lj for each Xj, and picking the rule (Xj, lj) that maximizes a specific splitting criteria. Then the training data is split according to the best rule, and for each new subset the procedure 10 is repeated, until one of the stopping criteria is met. Afterwards, taking into account the number of positive examples in each set S1 = {Xai |Xai ∈ S∧yi = 1}, the percentage of positives π1 = |S1|/|S| of each set is used to calculate the impurity of each leaf using either the entropy Ie(π1) = −π1 log π1 − (1 − π1) log(1 −π1) or the Gini Ig(π1) = 2π1(1 −π1) measures. Finally the gain of the splitting criteria using the rule (Xj, lj) is calculated as the impurity of S minus the weighted impurity of each leaf: Gain(Xj, lj) = I(π1) − |Sl| |S| I(πl1) − |Sr| |S| I(πr1), (7) where I(π1) can be either of the impurity measures Ie(π1) or Ig(π1). Subsequently, the gain of all possible splitting rules is calculated. The rule with maximal gain is selected (bestx,bestl) = arg max (Xj,lj ) Gain(Xj, lj), (8) and the set S is split into Sl and Sr according to that rule. Furthermore, the process is iteratively repeated for each subset until either there is no more possible splits or a stopping criteria is met. 3.2. Pruning of a classification tree After a decision tree has been fully grown, there is a risk for the algorithm to over fit the training data. In order to solve this, pruning techniques have been proposed in Breiman et al. (1984). The overall objective of pruning is to eliminate branches that are not contributing to the generalization accuracy of the tree Rokach and Maimon (2010). In general, pruning techniques start from a fully grown tree, and recur- sively check if by eliminating a node there is an improvement in the error or 11 misclassification rate � of the Tree. The most common pruning technique is cost-complexity pruning, initially proposed by Breiman Breiman et al. (1984). This method evaluates iteratively if the removal of a node improves the error rate � of a Tree in the set S, weighted by the difference of the number of nodes. PCcc = �(EB(Tree,node),S) − �(Tree,S) |Tree|− |EB(Tree,node)| , (9) where EB(Tree,node) is an auxiliary function that removes node from Tree and returns a new Tree. At each iteration, the current Tree is compared against all possible nodes. 4. Example-Dependent Cost-Sensitive Decision Trees Standard decision tree algorithms focus on inducing trees that maximize accuracy. However this is not optimal when the misclassification costs are un- equal Elkan (2001).This has led to many studies that develop algorithms that aim to introduce the cost-sensitivity into the algorithms Lomax and Vadera (2013). These studies have focused on introducing the class-dependent costs Draper et al. (1994); Ting (2002); Ling et al. (2004); Li et al. (2005); Kre- towski and Grześ (2006); Vadera (2010), which is not optimal for some appli- cations. For example in credit card fraud detection, it is true that false pos- itives have a different cost than false negatives, nevertheless, false negatives may vary significantly, which makes class-dependent cost-sensitive methods not suitable for this problem. In this section, we first propose a new method to introduce the costs into the decision tree induction stage, by creating new-cost based impurity mea- 12 sures. Afterwards, we propose a new pruning method based on minimizing the cost as pruning criteria. 4.1. Cost-sensitive impurity measures Standard impurity measures such as misclassification, entropy or Gini, take into account the distribution of classes of each leaf to evaluate the pre- dictive power of a splitting rule, leading to an impurity measure that is based on minimizing the misclassification rate. However, as has been previously shown Correa Bahnsen et al. (2013), minimizing misclassification does not lead to the same results than minimizing cost. Instead, we are interested in measuring how good is a splitting rule in terms of cost not only accuracy. For doing that, we propose a new example-dependent cost based impurity measure that takes into account the cost matrix of each example. We define a new cost-based impurity measure taking into account the costs when all the examples in a leaf are classified both as negative using f0 and positive using f1 Ic(S) = min { Cost(f0(S)),Cost(f1(S)) } . (10) The objective of this measure is to evaluate the lowest expected cost of a splitting rule. Following the same logic, the classification of each set is calculated as the prediction that leads to the lowest cost f(S) =   0 if Cost(f0(S)) ≤ Cost(f1(S)) 1 otherwise (11) Finally, using the cost-based impurity, the splitting criteria cost based gain of using the splitting rule (Xj, lj) is calculated with (7). 13 4.2. Cost-sensitive pruning Most of the literature in class-dependent cost-sensitive decision tree fo- cuses on using the misclassification costs during the construction of the al- gorithms Lomax and Vadera (2013). Only few algorithms such as AUCSplit Ferri et al. (2002) have included the costs both during and after the con- struction of the tree. However, this approach only used the class-dependent costs, and not the example-dependent costs. We propose a new example-dependent cost-based impurity measure, by replacing the error rate � in (9) with the cost of using the Tree on S i.e. by replacing with Cost(f(S)). PCc = Cost(f(S)) −Cost(f∗(S)) |Tree|− |EB(Tree,node)| , (12) where f∗ is the classifier of the tree without the selected node EB(Tree,node). Using the new pruning criteria, nodes of the tree that do not contribute to the minimization of the cost will be pruned, regardless of the impact of those nodes on the accuracy of the algorithm. This follows the same logic as in the proposed cost-based impurity measure, since minimizing the misclassification is different than minimizing the cost, and in several real-world applications the objectives align with the cost not with the misclassification error. 5. Experimental setup In this section we present the datasets used to evaluate the example- dependent cost-sensitive decision tree algorithm CSDT proposed in the Sec- tion 4. We used datasets from three different real world example-dependent cost-sensitive problems: Credit scoring, direct marketing and credit card 14 Database Set Observations %Positives Cost Credit Scoring Total 112,915 6.74 83,740,181 Training 45,264 6.75 33,360,130 Under-sampled 6,038 50.58 33,360,130 Rejection-sampled 5,271 43.81 29,009,564 Over-sampled 66,123 36.16 296,515,655 Validation 33,919 6.68 24,786,997 Testing 33,732 6.81 25,593,055 Direct Marketing Total 37,931 12.62 59,507 Training 15,346 12.55 24,304 Under-sampled 3,806 50.60 24,304 Rejection-sampled 1,644 52.43 20,621 Over-sampled 22,625 40.69 207,978 Validation 11,354 12.30 16,154 Testing 11,231 13.04 19,048 Credit Card Total 236,735 1.50 895,154 Fraud Detection Training 94,599 1.51 358,078 Under-sampled 2,828 50.42 358,078 Rejection-sampled 94,522 1.43 357,927 Over-sampled 189,115 1.46 716,006 Validation 70,910 1.53 274,910 Testing 71,226 1.45 262,167 Table 2: Summary of the datasets fraud detection. For each dataset we define a cost matrix, from which the algorithms are trained. Additionally, we perform an under-sampling, cost- proportionate rejection-sampling and cost-proportionate over-sampling pro- cedures. In Table 2, information about the different datasets is shown. 5.1. Credit scoring Credit scoring is a real-world problem in which the real costs due to mis- classification are not constant, but are example-dependent. The objective in credit scoring is to classify which potential customers are likely to default 15 Actual Positive Actual Negative yi = 1 yi = 0 Predicted Positive CT Pi = 0 CF Pi = ri + C a F P ci = 1 Predicted Negative CF Ni = Cli ·Lgd CT Ni = 0 ci = 0 Table 3: Credit scoring example-dependent cost matrix a contracted financial obligation based on the customer’s past financial ex- perience, and with that information decide whether to approve or decline a loan Anderson (2007). This tool has become a standard practice among financial institutions around the world in order to predict and control their loan portfolios. When constructing credit scores, it is a common practice to use standard cost-insensitive binary classification algorithms such as lo- gistic regression, neural networks, discriminant analysis, genetic programing, decision tree, among others Correa Bahnsen and Gonzalez Montoya (2011); Hand and Henley (1997); Ong et al. (2005); Yeh and Lien (2009). However, in practice, the cost associated with approving a bad customer is quite different from the cost associated with declining a good customer. Furthermore, the costs are not constant among customers. This is because loans have different credit line amounts, terms, and even interest rates. For this paper we follow the example-dependent cost-sensitive approach for credit scoring proposed in (Correa Bahnsen et al., 2014b). In Table 3, the credit scoring cost matrix is shown. First, the costs of a correct classification, CTPi and CTNi , are zero for all customers, i. Then, CFNi are the losses if the customer i defaults, which is calculated as the credit line Cli time the loss given default Lgd. The cost of a false positive per customer CFPi is 16 defined as the sum of two real financial costs ri and C a FP , where ri is the loss in profit by rejecting what would have been a good customer. The second term CaFP , is related to the assumption that the financial institution will not keep the money of the declined customer idle it will instead give a loan to an alternative customer (Nayak and Turvey, 1997). Since no further information is known about the alternative customer, it is assumed to have an average credit line Cl and an average profit r. Then, CaFP = −r ·π0 + Cl ·Lgd ·π1, in other words minus the profit of an average alternative customer plus the expected loss, taking into account that the alternative customer will pay his debt with a probability equal to the prior negative rate, and similarly will default with probability equal to the prior positive rate. We apply the previous framework to a publicly available credit scoring dataset. The dataset is the 2011 Kaggle competition Give Me Some Credit2, in which the objective is to identify those customers of personal loans that will experience financial distress in the next two years. The Kaggle Credit datasets contain information regarding the features, and more importantly about the income of each example, from which an estimated credit limit Cli can be calculated (see (Correa Bahnsen et al., 2014a)). The dataset contains 112,915 examples, each one with 10 features and the class label. The proportion of default or positive examples is 6.74%. Since no specific information regarding the datasets is provided, we assume that they belong to average European financial institution. This enabled us to find the different parameters needed to calculate the cost matrix. In particular we 2http://www.kaggle.com/c/GiveMeSomeCredit/ 17 Actual Positive Actual Negative yi = 1 yi = 0 Predicted Positive CT Pi = Ca CF Pi = Ca ci = 1 Predicted Negative CF Ni = Inti CT Ni = 0 ci = 0 Table 4: Direct marketing example-dependent cost matrix used the same parameters as in (Correa Bahnsen et al., 2014a), the interest rate intr to 4.79%, the cost of funds intcf to 2.94%, the term l to 24 months, and the loss given default Lgd to 75%. 5.2. Direct marketing In direct marketing the objective is to classify those customers who are more likely to have a certain response to a marketing campaign (Ngai et al., 2009). We used a direct marketing dataset (Moro et al., 2011) available on the UCI machine learning repository (Bache and Lichman, 2013). The dataset contains 45,000 clients of a Portuguese bank who were contacted by phone between March 2008 and October 2010 and received an offer to open a long-term deposit account with attractive interest rates. The dataset contains features such as age, job, marital status, education, average yearly balance and current loan status and the label indicating whether or not the client accepted the offer. This problem is example-dependent cost sensitive, since there are different costs of false positives and false negatives. Specifically, in direct marketing, false positives have the cost of contacting the client, and false negatives have the cost due to the loss of income by failing to contact a client that otherwise would have opened a long-term deposit. 18 We used the direct marketing example-dependent cost matrix proposed in (Correa Bahnsen et al., 2014c). The cost matrix is shown in Table 4, where Ca is the administrative cost of contacting the client, as is credit card fraud, and Inti is the expected income when a client opens a long-term deposit. This last term is defined as the long-term deposit amount times the interest rate spread. In order to estimate Inti, first the long-term deposit amount is assumed to be a 20% of the average yearly balance, and lastly, the interest rate spread is estimated to be 2.463333%, which is the average between 2008 and 2010 of the retail banking sector in Portugal as reported by the Portuguese central bank. Given that, the Inti is equal to (balance∗ 20%) ∗ 2.463333%. 5.3. Credit card fraud detection A credit card fraud detection algorithm, consisting on identifying those transactions with a high probability of being fraud, based on historical cus- tomers consumer and fraud patterns. Different detection systems that are based on machine learning techniques have been successfully used for this problem, in particular: neural networks (Maes et al., 2002), Bayesian learn- ing (Maes et al., 2002), hybrid models (Krivko, 2010), support vector ma- chines (Bhattacharyya et al., 2011) and random forest (Correa Bahnsen et al., 2013). Credit card fraud detection is by definition a cost sensitive problem, since the cost of failing to detect a fraud is significantly different from the one when a false alert is made (Elkan, 2001). We used the fraud detection example- dependent cost matrix proposed in (Correa Bahnsen et al., 2013). In Table 5, the cost matrix is presented. Where Amti is the amount of transaction i, and 19 Ca is the administrative cost of investigating a fraud alert. This cost matrix differentiates between the costs of the different outcomes of the classification algorithm, meaning that it differentiates between false positives and false negatives, and also the different costs of each example. For this paper we used a dataset provided by a large European card pro- cessing company. The dataset consists of fraudulent and legitimate trans- actions made with credit and debit cards between January 2012 and June 2013. The total dataset contains 120,000,000 individual transactions, each one with 27 attributes, including a fraud label indicating whenever a trans- action is identified as fraud. This label was created internally in the card processing company, and can be regarded as highly accurate. In the dataset only 40,000 transactions were labeled as fraud, leading to a fraud ratio of 0.025%. From the initial attributes, an additional 260 attributes are derived using the methodology proposed in (Bhattacharyya et al., 2011; Whitrow et al., 2008; Correa Bahnsen et al., 2013). The idea behind the derived attributes consists in using a transaction aggregation strategy in order to capture con- sumer spending behavior in the recent past. The derivation of the attributes consists in grouping the transactions made during the last given number of Actual Positive Actual Negative yi = 1 yi = 0 Predicted Positive CT Pi = Ca CF Pi = Ca ci = 1 Predicted Negative CF Ni = Amti CT Ni = 0 ci = 0 Table 5: Credit card fraud detection example-dependent cost matrix 20 hours, first by card or account number, then by transaction type, merchant group, country or other, followed by calculating the number of transactions or the total amount spent on those transactions. For the experiments, a smaller subset of transactions with a higher fraud ratio, corresponding to a specific group of transactions, is selected. This dataset contains 236,735 transactions and a fraud ratio of 1.50%. In this dataset, the total financial losses due to fraud are 895,154 Euros. This dataset was selected because it is the one where most frauds are being made. 6. Results In this section we present the experimental results. First, we evaluate the performance of the proposed CSDT algorithm and compare it against a classical decision tree (DT). We evaluate the different trees using them without pruning (notp), with error based pruning (errp), and also with the proposed cost-sensitive pruning technique (costp). The different algorithms are trained using the training (t), under-sampling (u), cost-proportionate rejection-sampling (r), and cost-proportionate over-sampling (o) datasets. Lastly, we compare our proposed method versus state-of-the-art example- dependent cost-sensitive techniques. 6.1. Results CSDT We evaluate a decision tree constructed using the Gini impurity measure, with and without the pruning defined in (9). We also apply the cost-based pruning procedure given in (12). Lastly, we compared against the proposed CSDT constructed using the cost-based impurity measure defined in (10), using the two pruning procedures. 21 DT notp DT errp DT costp CSDT notp CSDT errp CSDT costp 0 10 20 30 40 50 60 70 80 % S av in gs DT notp DT errp DT costp CSDT notp CSDT errp CSDT costp 0.0 0.1 0.2 0.3 0.4 0.5 F 1 S c o r e Fraud Detection Direct Marketing Credit Scoring Figure 1: Results of the DT and the CSDT . For both algorithms, the results are calculated with and without both types of pruning criteria. There is a clear difference between the savings of the DT and the CSDT algorithms. However, this difference is not observable on the F1Score results. Since the CSDT is focused on maximizing the savings not the accuracy or F1Score. There is a small increase in savings when using the DT with cost- sensitive pruning. Nevertheless, in the case of the CSDT algorithm, there is no change when using any pruning procedure, neither in savings or F1Score. In Figure 1, the results using the three databases are shown. In particular we first evaluate the impact of the algorithms when trained using the training set. There is a clear difference between the savings of the DT and the CSDT algorithms. However, that difference is not observable on the F1Score results. Since the CSDT is focused on maximizing the savings not the accuracy or F1Score. There is a small increase in savings when using the DT with cost- 22 Fraud Detection Direct Marketing Credit Scoring set Algorithm %Sav %Accu F1Score %Sav %Accu F1Score %Sav %Accu F1Score t DTnotp 31.76 98.76 0.4458 19.11 88.24 0.2976 18.95 93.42 0.3062 DTerrp 31.76 98.76 0.4458 19.70 88.28 0.3147 18.95 93.42 0.3062 DTcostp 35.89 98.71 0.4590 28.08 88.28 0.3503 27.59 93.41 0.3743 CSDTnotp 70.85 95.07 0.2529 69.00 85.51 0.2920 49.28 93.19 0.3669 CSDTerrp 70.85 95.07 0.2529 68.97 88.18 0.3193 48.85 93.19 0.3669 CSDTcostp 71.16 94.98 0.2522 69.10 81.75 0.2878 49.39 90.28 0.3684 u DTnotp 52.39 85.52 0.1502 49.80 70.80 0.3374 48.91 75.96 0.2983 DTerrp 52.39 85.52 0.1502 49.80 70.80 0.3374 48.91 75.96 0.2983 DTcostp 70.26 92.67 0.2333 53.20 74.51 0.3565 49.77 79.37 0.3286 CSDTnotp 12.46 69.34 0.0761 64.21 52.34 0.2830 30.68 93.19 0.2061 CSDTerrp 14.98 70.31 0.0741 66.21 60.06 0.2822 41.49 93.19 0.2564 CSDTcostp 15.01 70.31 0.0743 68.07 62.11 0.2649 44.89 78.08 0.2881 r DTnotp 34.39 98.70 0.4321 68.59 72.73 0.3135 48.97 87.07 0.3931 DTerrp 34.39 98.70 0.4321 68.79 73.39 0.3196 48.97 87.07 0.3931 DTcostp 38.99 98.64 0.4478 69.27 72.58 0.3274 50.48 82.69 0.3501 CSDTnotp 70.85 95.07 0.2529 66.87 57.91 0.2761 31.25 93.19 0.1940 CSDTerrp 70.85 95.07 0.2529 67.47 63.31 0.2581 40.69 93.19 0.2529 CSDTcostp 71.09 94.94 0.2515 68.08 62.60 0.2642 44.51 77.82 0.2869 o DTnotp 31.72 98.77 0.4495 60.30 79.46 0.3674 47.39 88.28 0.3994 DTerrp 31.72 98.77 0.4495 60.30 79.46 0.3674 47.39 88.28 0.3994 DTcostp 37.35 98.68 0.4575 69.44 70.07 0.3108 50.92 87.52 0.3977 CSDTnotp 70.84 95.06 0.2529 68.75 64.72 0.2935 41.37 93.19 0.2205 CSDTerrp 70.84 95.06 0.2529 68.75 64.72 0.2935 41.38 90.63 0.3457 CSDTcostp 71.09 94.94 0.2515 68.75 64.72 0.2935 41.65 78.26 0.2896 Table 6: Results on the three datasets of the cost-sensitive and standard decision tree, without pruning (notp), with error based pruning (errp), and with cost-sensitive pruning technique (costp). Estimated using the different training sets: training, under-sampling, cost-proportionate rejection-sampling and cost-proportionate over-sampling sensitive pruning. Nevertheless, in the case of the CSDT algorithm, there is no change when using any pruning procedure, neither in savings or F1Score. In addition, we also evaluate the algorithms on the different sets, under- 23 Training Under-Sampling Cost-Sensitive Rejection-Sampling Cost-Sensitive Over-Sampling 0 10 20 30 40 50 60 70 80 % S av in gs DT − errp DT − costp CSDT Figure 2: Average savings on the three datasets of the different cost-sensitive and standard decision tree, estimated using the different training sets: training, under-sampling, cost- proportionate rejection-sampling and cost-proportionate over-sampling. The best results are found when using the training set. When using the under-sampling set there is a decrease in savings of the algorithm. Lastly, in the case of the cost-proportionate sampling sets, there is a small increase in savings when using the CSDT algorithm. sampling, rejection-sampling and over-sampling. The results are shown in Table 6. Moreover, in Figure 2, the average results of the different algo- rithms measured by savings is shown. The best results are found when using the training set. When using the under-sampling set there is a decrease in savings of the CSDT algorithm. Lastly, in the case of the cost-proportionate sampling sets, there is a small increase in savings when using the CSDT al- gorithm. Finally, we also analyze the different models taking into account the com- plexity and the training time. In particular we evaluate the size of each Tree. In Table 7, and Figure 3, the results are shown. The CSDT algorithm creates significantly smaller trees, which leads to a lower training time. In particu- lar this is a result of using the non weighted gain, the CSDT only accepts splitting rules that contribute to the overall reduction of the cost, which is 24 Fraud Detection Direct Marketing Credit Scoring set Algorithm |Tree| Time |Tree| Time |Tree| Time t DTnotp 488 2.45 298 1.58 292 1.58 DTerrp 488 3.90 298 2.13 292 2.13 DTcostp 446 19.19 291 5.23 280 5.23 CSDTnotp 89 1.47 51 0.40 69 0.40 CSDTerrp 88 1.87 51 0.64 69 0.64 CSDTcostp 89 1.74 51 0.45 69 0.45 u DTnotp 308 1.10 198 1.00 167 1.00 DTerrp 308 1.43 198 1.14 167 1.14 DTcostp 153 2.59 190 1.34 142 1.34 CSDTnotp 14 0.20 23 0.17 42 0.17 CSDTerrp 14 0.23 23 0.19 42 0.19 CSDTcostp 14 0.24 23 0.18 42 0.18 r DTnotp 268 0.98 181 0.90 267 0.90 DTerrp 268 1.24 181 0.95 267 0.95 DTcostp 153 2.48 162 1.20 261 1.20 CSDTnotp 18 0.22 10 0.07 70 0.07 CSDTerrp 18 0.23 10 0.07 70 0.07 CSDTcostp 18 0.23 10 0.07 70 0.07 o DTnotp 425 2.30 340 1.80 277 1.80 DTerrp 425 3.98 340 2.65 277 2.65 DTcostp 364 10.15 288 5.99 273 5.99 CSDTnotp 37 1.58 51 0.38 70 0.38 CSDTerrp 37 1.90 51 0.45 70 0.45 CSDTcostp 37 1.98 51 0.42 70 0.42 Table 7: Training time and tree size of the different cost-sensitive and standard deci- sion tree, estimated using the different training sets: training, under-sampling, cost- proportionate rejection-sampling and cost-proportionate over-sampling, for the three databases. not the case if instead the weighted gain was used. Even that the DT with cost pruning, produce a good result measured by savings, it is the one that takes the longer to estimate. Since the algorithm first creates a big decision tree using the Gini impurity, and then attempt to find a smaller tree taking 25 Training Under Sampling CS Rejection Sampling CS Over Sampling 0 50 100 150 200 250 300 350 400 n u m be r of n od es (a) Tree size Training Under Sampling CS Rejection Sampling CS Over Sampling 0 2 4 6 8 10 m in u te s (b) Training time DT errp DT costp CSDT Figure 3: Average tree size (a) and training time (b), of the different cost-sensitive and standard decision tree, estimated using the different training sets: training, under- sampling, cost-proportionate rejection-sampling and cost-proportionate over-sampling, for the three databases. The CSDT algorithm create significantly smaller trees, which leads to a lower training time. into account the cost. Measured by training time, the CSDT is by all means faster to train than the DT algorithm, leading to an algorithm that not only gives better results measured by savings but also one that can be trained much quicker than the standard DT . 6.2. Comparison with state-of-the-art methods Additionally to the comparison of the CSDT and a DT , we also evaluate and compare our proposed method with the standard example-dependent cost-sensitive methods, namely, cost-proportionate rejection-sampling (Zadrozny et al., 2003), cost-proportionate over-sampling (Elkan, 2001) and Bayes minimum risk (BMR) (Correa Bahnsen et al., 2014c). 26 Fraud Detection Direct Marketing Credit Scoring set Algorithm %Sav %Accu F1Score %Sav %Accu F1Score %Sav %Accu F1Score t DT 31.76 98.76 0.4458 19.70 88.28 0.3147 18.95 93.42 0.3062 DT −BMR 60.45 65.05 0.2139 69.27 63.09 0.2416 33.25 79.06 0.2450 LR 0.92 99.75 0.1531 14.99 88.25 0.2462 3.28 93.47 0.0811 LR−BMR 45.52 66.42 0.1384 68.46 70.73 0.2470 29.55 80.86 0.2883 RF 33.42 76.33 0.2061 20.10 86.94 0.2671 15.38 93.57 0.2720 RF −BMR 64.14 62.85 0.2052 67.30 69.07 0.3262 47.89 81.54 0.2698 u DT 52.39 85.52 0.1502 49.80 70.80 0.3374 48.91 75.96 0.2983 LR 12.43 73.08 0.0241 49.60 73.32 0.3396 45.38 85.40 0.3618 RF 56.84 54.48 0.0359 41.64 67.14 0.3069 49.53 79.02 0.3215 r DT 34.39 98.70 0.4321 68.79 73.39 0.3196 48.97 87.07 0.3931 LR 30.77 76.02 0.1846 62.85 72.63 0.3313 48.80 84.69 0.3660 RF 38.12 75.03 0.2171 61.45 66.06 0.2949 47.34 81.47 0.3284 o DT 31.72 98.77 0.4495 60.30 79.46 0.3674 47.39 88.28 0.3994 LR 27.93 76.79 0.1776 55.63 81.65 0.3540 34.05 91.55 0.3923 RF 36.12 75.62 0.2129 22.80 85.52 0.2871 21.72 93.22 0.3301 Table 8: Results on the three datasets of the decision tree, logistic regression and random forest algorithms, estimated using the different training sets: training, under-sampling, cost-proportionate rejection-sampling and cost-proportionate over-sampling Using each database and each set, we estimate three different algorithms, in particular a decision tree (DT), a logistic regression (LR) and a random forest (RF ). The LR and RF algorithms were trained using the implemen- tations of Scikit-learn Pedregosa et al. (2011), respectively. We only used the BMR algorithm using the training set, as it has been previously shown that it is where the model gives the best results (Correa Bahnsen et al., 2013). The results are shown on Table 8. Measured by savings, it is observed that regardless of the algorithm used for estimating the positive probabilities, in all cases there is an increase in savings by using BMR . In general, for all datasets the best results are found when using a random forest algorithm 27 for estimating the positive probabilities. In the case of the direct marketing dataset, the results of the different algorithms are very similar. Nevertheless, in all cases the BMR produce higher savings. When analyzing the F1Score, it is observed that in general there is no increase in results when using the BMR. It is observed that the best models selected by savings are not the same as the best ones measured by F1Score. And the reason for that, is because the F1Score treat the false positives and the false negatives as equal, which as discussed before is not the case in example-dependent cost-sensitive problems. Finally, we compare the results of the standard algorithms, the algo- rithms trained using the cost-proportionate sampling sets, the BMR, and the CSDT . Results are shown in Figure 4. When comparing by savings, for all databases the best model is the CSDT , closely follow by the DT with cost-based pruning. It is interesting to see that both algorithms that the al- gorithms that incorporates the costs during construction, BMR and CSDT , gives the best results when trained using the training set. When measured by F1Score, there is not a clear trend regarding the different results. In the case of fraud detection the best model is the DT , however measure by savings that model performs poorly. In the case of direct marketing, by F1Score, DT with cost pruning performs the best, but that model is the second worst by savings. In the credit scoring dataset the best model is the same when measured by savings or F1Score. 28 Fraud Detection Direct Marketing Credit Scoring 0 10 20 30 40 50 60 70 80 % S av in gs Fraud Detection Direct Marketing Credit Scoring 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 F 1 S c o r e t-DT r-RF t-RF -BMR u-DT − costp t-CSDT Figure 4: Comparison of the different models on the three databases. Measured by savings, CSDT is the overall best method. However, by F1Score, there is not a clear trend regarding the different results. 7. Conclusions In this paper we define the classification problem cost characteristic, as a straight forward method to define when a problem is cost-insensitive, class- dependent cost-sensitive and example-dependent cost-sensitive. Moreover, we proposed an example-dependent cost-sensitive decision tree algorithm, by incorporating the different example-dependent costs into a new cost-based impurity measure and a new cost-based pruning criteria. We show the im- portance of using including the costs during the algorithm construction, both by using a classical decision tree and then the cost-based pruning procedure, 29 and by fully creating a decision tree taking into account the costs. Using three databases, from three real-world applications: credit card fraud detection, credit scoring and direct marketing, we show that the pro- posed algorithm significantly outperforms the standard decision tree when measured by savings, for all databases. In terms of complexity and train- ing time, our proposed algorithm creates significantly smaller trees, and avoiding overfit, in a fraction of the time. Furthermore, when compared against state-of-the-art example-dependent cost-sensitive techniques, specif- ically, cost-proportionate sampling and Bayes minimum risk, we found that in all cases our method provides the best performance measured by savings. Lastly, our results show the importance of using the real example-dependent financial costs associated with the real-world applications, since there are sig- nificant differences in the results when evaluating a model using a traditional cost-insensitive measures such as the accuracy or F1Score, than when us- ing the savings,leading to the conclusion of the importance of using the real practical financial costs of each context. References Anderson, R., 2007. The Credit Scoring Toolkit : Theory and Practice for Retail Credit Risk Management and Decision Automation. Oxford Univer- sity Press. Aodha, O. M., Brostow, G. J., Dec. 2013. Revisiting Example Dependent Cost-Sensitive Learning with Decision Trees. In: 2013 IEEE International Conference on Computer Vision. Washington, DC, USA, pp. 193–200. 30 Bache, K., Lichman, M., 2013. UCI Machine Learning Repository. School of Information and Computer Science, University of California. Bhattacharyya, S., Jha, S., Tharakunnel, K., Westland, J. C., Feb. 2011. Data mining for credit card fraud: A comparative study. Decision Support Systems 50 (3), 602–613. Breiman, L., Friedman, J., Stone, C. J., Olshen, R., 1984. Classification and Regression Trees. Chapman and Hall/CRC. Cohen, I., Goldszmidt, M., 2004. Properties and Benefits of Calibrated Clas- sifiers. In: Knowledge Discovery in Databases: PKDD 2004. Vol. 3202. Springer Berlin Heidelberg, pp. 125–136. Correa Bahnsen, A., Aouada, D., Ottersten, B., 2014a. Example-Dependent Cost-Sensitive Credit Scoring using Bayes Minimum Risk. In: Interna- tional Conference on Machine Learning and Applications. Correa Bahnsen, A., Aouada, D., Ottersten, B., 2014b. Example-Dependent Cost-Sensitive Logistic Regression for Credit Scoring. In: 2014 13th Inter- national Conference on Machine Learning and Applications. IEEE, Detroit, USA. Correa Bahnsen, A., Gonzalez Montoya, A., Dec. 2011. Evolutionary Algo- rithms for Selecting the Architecture of a MLP Neural Network: A Credit Scoring Case. In: IEEE 11th International Conference on Data Mining Workshops. Ieee, pp. 725–732. Correa Bahnsen, A., Stojanovic, A., Aouada, D., Ottersten, B., Dec. 2013. Cost Sensitive Credit Card Fraud Detection Using Bayes Minimum Risk. 31 In: 2013 12th International Conference on Machine Learning and Appli- cations. IEEE, Miami, USA, pp. 333–338. Correa Bahnsen, A., Stojanovic, A., Aouada, D., Ottersten, B., Apr. 2014c. Improving Credit Card Fraud Detection with Calibrated Probabilities. In: Proceedings of the fourteenth SIAM International Conference on Data Mining. No. January 2012. Philadelphia, PA, pp. 677 – 685. Draper, B., Brodley, C., Utgoff, P., 1994. Goal-directed classification using linear machine decision trees. IEEE Transactions on Pattern Analysis and Machine Intelligence 16 (9), 888–893. Drummond, C., Holte, R., 2003. C4.5, class imbalance, and cost sensitivity: why under-sampling beats over-sampling. In: Workshop on Learning from Imbalanced Datasets II, ICML. Washington, DC, USA. Elkan, C., 2001. The Foundations of Cost-Sensitive Learning. In: Seventeenth International Joint Conference on Artificial Intelligence. pp. 973–978. Ferri, C., Flach, P., Hernández-Orallo, J., 2002. Learning decision trees using the area under the ROC curve. In: ICML ’02 Proceedings of the Nineteenth International Conference on Machine Learning. San Francisco, CA, USA, pp. 139–146. Glady, N., Baesens, B., Croux, C., Aug. 2009. Modeling churn using customer lifetime value. European Journal of Operational Research 197 (1), 402–411. Hand, D. J., Henley, W. E., 1997. Statistical Classification Methods in Con- sumer Credit Scoring: A Review. Journal of the Royal Statstical Society. Series A (Statistics in Society) 160 (3), 523–541. 32 Hastie, T., Tibshirani, R., Friedman, J., 2009. The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd Edition. Springer, Stanford, CA. Jayanta K., G., Mohan, D., Tapas, S., Apr. 2006. Bayesian Inference and De- cision Theory. In: An Introduction to Bayesian Analysis. Vol. 13. Springer New York, pp. 26–63. Kim, J., Choi, K., Kim, G., Suh, Y., Mar. 2012. Classification cost: An empirical comparison among traditional classifier, Cost-Sensitive Classifier, and MetaCost. Expert Systems with Applications 39 (4), 4013–4019. Kretowski, M., Grześ, M., 2006. Evolutionary induction of cost-sensitive de- cision trees. In: Foundations of Intelligent Systems. Springer Berlin Hei- delberg, pp. 121–126. Krivko, M., Aug. 2010. A hybrid model for plastic card fraud detection sys- tems. Expert Systems with Applications 37 (8), 6070–6076. Li, J., Li, X., Yao, X., 2005. Cost-Sensitive Classification with Genetic Pro- gramming. In: 2005 IEEE Congress on Evolutionary Computation. Vol. 3. IEEE, pp. 2114–2121. Ling, C. X., Yang, Q., Wang, J., Zhang, S., 2004. Decision trees with mini- mal costs. In: Twenty-first international conference on Machine learning - ICML ’04. No. Icml. ACM Press, New York, New York, USA, p. 69. Lomax, S., Vadera, S., Feb. 2013. A survey of cost-sensitive decision tree induction algorithms. ACM Computing Surveys 45 (2), 1–35. 33 Ma, J., Saul, L. K., Savage, S., Voelker, G. M., Apr. 2011. Learning to detect malicious URLs. ACM Transactions on Intelligent Systems and Technology 2 (3), 1–24. Maes, S., Tuyls, K., Vanschoenwinkel, B., Manderick, B., 2002. Credit card fraud detection using Bayesian and neural networks. In: Proceedings of NF2002. Maimon, L. R. O., 2008. Data Mining with Decision Trees: Theory and Applications. World Scientific Publishing Company. Moro, S., Laureano, R., Cortez, P., 2011. Using data mining for bank direct marketing: An application of the crisp-dm methodology. In: European Simulation and Modelling Conference. No. Figure 1. Guimares, Portugal, pp. 117–121. Nayak, G. N., Turvey, C. G., 1997. Credit Risk Assessment and the Oppor- tunity Costs of Loan Misclassification. Canadian Journal of Agricultural Economics 45 (3), 285–299. Ngai, E., Xiu, L., Chau, D., Mar. 2009. Application of data mining techniques in customer relationship management: A literature review and classifica- tion. Expert Systems with Applications 36 (2), 2592–2602. Ong, C.-s., Huang, J.-j., Tzeng, G.-h., 2005. Building credit scoring models using genetic programming. Expert Systems With Applications 29, 41–47. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., 34 Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E., 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830. Rokach, L., Maimon, O., 2010. Decision Trees. In: Rokach, L., Maimon, O. (Eds.), Data Mining and Knowledge Discovery Handbook, 2nd Edition. Springer US, Ch. 9, pp. 149–174. Sahin, Y., Bulkan, S., Duman, E., Nov. 2013. A cost-sensitive decision tree approach for fraud detection. Expert Systems with Applications 40 (15), 5916–5923. Ting, K., 2002. An instance-weighting method to induce cost-sensitive trees. IEEE Transactions on Knowledge and Data Engineering 14 (3), 659–665. Vadera, S., 2010. CSNL: A cost-sensitive non-linear decision tree algorithm. ACM Transactions on Knowledge Discovery from Data 4 (2), 1–25. Verbraken, T., Bravo, C., Weber, R., Baesens, B., Oct. 2014. Development and application of consumer credit scoring models using profit-based clas- sification measures. European Journal of Operational Research 238 (2), 505–513. Wang, T., 2013. Efficient Techniques for Cost-Sensitive Learning with Mul- tiple Cost Considerations. Ph.D. thesis, University of Technology, Sydney. Whitrow, C., Hand, D. J., Juszczak, P., Weston, D. J., Adams, N. M., Jul. 2008. Transaction aggregation as a strategy for credit card fraud detection. Data Mining and Knowledge Discovery 18 (1), 30–55. 35 Yeh, I.-C., Lien, C.-h., Mar. 2009. The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients. Expert Systems with Applications 36 (2), 2473–2480. Zadrozny, B., Langford, J., Abe, N., 2003. Cost-sensitive learning by cost- proportionate example weighting. In: Third IEEE International Confer- ence on Data Mining. IEEE Comput. Soc, pp. 435–442. 36 
work_ijodnynjfzg6pajray343ungoq	----	A fuzzy real option approach for investment project valuation Expert Systems with Applications 38 (2011) 15296–15302 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A fuzzy real option approach for investment project valuation Shiu-Hwei Ho a,⇑, Shu-Hsien Liao b a Department of Business Administration, Technology and Science Institute of Northern Taiwan, No. 2, Xueyuan Road, Peitou, 112 Taipei, Taiwan, ROC b Graduate Institute of Management Sciences, Tamkang University, No. 151, Yingchuan Road, Tamsui, 25137 Taipei County, Taiwan, ROC a r t i c l e i n f o a b s t r a c t Keywords: Project valuation Real options Fuzzy numbers Flexibility Uncertainty 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.06.010 ⇑ Corresponding author. E-mail addresses: succ04.dba@msa.hinet.net (S edu.tw (S.-H. Liao). The main purpose of this paper is to propose a fuzzy approach for investment project valuation in uncertain environments from the aspect of real options. The traditional approaches to project valuation are based on discounted cash flows (DCF) analysis which provides measures like net present value (NPV) and internal rate of return (IRR). However, DCF-based approaches exhibit two major pitfalls. One is that DCF parameters such as cash flows cannot be estimated precisely in the uncertain decision making environments. The other one is that the values of managerial flexibilities in investment projects cannot be exactly revealed through DCF analysis. Both of them would entail improper results on strategic investment projects valuation. There- fore, this paper proposes a fuzzy binomial approach that can be used in project valuation under uncertainty. The proposed approach also reveals the value of flexibilities embedded in the project. Furthermore, this paper provides a method to compute the mean value of a project’s fuzzy expanded NPV that represents the entire value of project. Finally, we use the approach to practically evaluate a project. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction deficiencies of the traditional valuation methods through recogniz- DCF-based approaches to project valuation implicitly assume that a project will be undertaken immediately and operated con- tinuously until the end of its expected useful life, even though the future is uncertain. By treating projects as independent invest- ment opportunities, decisions are made to accept projects with positive computed NPVs. Traditional NPV techniques only focus on current predictable cash flows and ignore future managerial flexibilities, therefore, may undervalue the projects and mislead the decision makers. Furthermore, for high-risk investment pro- jects, the traditional NPV method may adopt higher discount rates to discount project cash flows for trade-off or compensation. How- ever, higher discount rates may result in the underestimation of project value and the rejection of a potential project. For instance, investments such as new drug development or crude oil exploita- tion may carry high risk, but may also bring higher returns. Since DCF-based approaches ignore the upside potentials of added value that could be brought to projects through managerial flexibilities and innovations, they usually underestimate the upside value of projects (Bowman & Moskowitz, 2001; Dixit & Pindyck, 1995; Luehrman, 1998; Trigeorgis, 1993; Yeo & Qiu, 2003). In partic- ular, as market conditions change in the future, investment project may include flexibilities by which project value can be raised. Such flexibilities are called real options or strategic options. The real options approach to projects valuation seeks to correct the ll rights reserved. .-H. Ho), Michael@mail.tku. ing that managerial flexibilities can bring significant values to pro- jects. According to real options theory, an investment is of higher value in a more uncertain or volatile market because of investment decision flexibilities. Real options approach, as a strategic decision making tool, bor- rows ideas from financial options because it explicitly accounts for future flexibility value. Real options analysis is based on the assumption that there is an underlying source of uncertainty, such as the price of a commodity or the outcome of a research project. Over time, the outcome of the underlying uncertainty is revealed, and managers can adjust their strategy accordingly. The objectives of this paper are to develop a fuzzy binomial approach to evaluate a project embedded with real options, to pro- pose a method suitable for computing the mean value of fuzzy NPV, and to explore the value of multiple options existing in projects. The paper is organized as follows. Section 2 provides a survey of real options analysis. We especially focus on pricing, applications and recent developments of real options analysis. Section 3 presents a fuzzy binomial approach to evaluate a project under vague situa- tions. This section also proposes a method to compute the mean va- lue of fuzzy NPV. Section 4 illustrates a project valuation based on our approach. In the example, the values of the real options are also assessed. Section 5 discusses the results and findings in the example. Finally, conclusions are drawn in Section 6. 2. Related works Based on real options theory, Chen, Zhang, and Lai (2009) pre- sented an approach to evaluate IT investments subject to multiple http://dx.doi.org/10.1016/j.eswa.2011.06.010 mailto:succ04.dba@msa.hinet.net mailto:Michael@mail.tku.edu.tw mailto:Michael@mail.tku.edu.tw http://dx.doi.org/10.1016/j.eswa.2011.06.010 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa S.-H. Ho, S.-H. Liao / Expert Systems with Applications 38 (2011) 15296–15302 15297 risks. By modeling public risks and private risks into a unified frame- work, they utilized the binomial model to evaluate an ERP develop- ment project. Wu, Ong, and Hsu (2008) argued that ERP may be best represented by a non-analytical, compound option model. However, most IT studies that employ the options theory only consider a single option, use an analytical model such as the Black–Scholes model (1973), and cannot deal with multi-option situations. Therefore, Wu et al. employed the binomial tree approach to implement an active ERP management which involves uncertainties over time. Hahn and Dyer (2008) proposed a recombining binomial lattice ap- proach for modeling real options and valuing managerial flexibility to address a common issue in many practical applications—underly- ing stochastic processes that are mean-reverting. The models were tested by implementing the lattice in binomial decision tree format and applying to a real application by solving for the value of an oil and gas switching option. Reyck, Degraeve, and Vandenborre (2008) proposed an alternative approach for valuing real options based on the certainty-equivalent version of the NPV formula, which eliminates the need to identify market-priced twin securities. More- over, Bowe and Lee (2004) also utilized the log-transformed bino- mial lattice approach to evaluate the Taiwan High-Speed Rail (THSR) project. In DCF, parameters such as cash flows and discount rates are difficult to estimate (Carlsson & Fuller, 2003). In particular, innova- tive investment projects may count on the subjective judgments of decision makers due to lack of past data for reference. These param- eters are essentially estimated under uncertainty. With respect to uncertainty, probability is one way to depict whereas possibility is another. Fuzzy set theory provides a basis for the theory of possibil- ity (Zadeh, 1999). Fuzzy logic may be viewed as an attempt at formalization of two remarkable human capabilities. One is the capability to converse, reason and make rational decisions in an environment of imprecision, uncertainty and incompleteness of information and the other one is to perform a wide variety of phys- ical and mental tasks without any measurements and computations (Zadeh, 2008). The outstanding feature of fuzzy logic is that in fuzzy logic everything is—or is allowed to be—a matter of degree. In the generalized theory of uncertainty, uncertainty is linked to informa- tion through the concept of granular structure—a concept that plays a key role in human interaction with the real world (Zadeh, 2005). Thus, these parameters can be characterized with possibilistic distributions instead of probabilistic distributions, and can be esti- mated by making use of fuzzy numbers. By modeling the stock price in each state as a fuzzy number, Muzzioli and Torricelli (2004) obtained a possibility distribution of the risk-neutral probability in a multi-period binomial model, then computed the option price with a weighted expected value interval, and thus determined a ‘‘most likely’’ option value within the interval. Muzzioli and Reynaerts (2008) also addressed that the key input of the multi-period binomial model is the volatility of the underlying asset, but it is an unobservable parameter. The volatility parameter can be estimated either from historical data (historical volatility) or implied from the price of European options (implied volatility). Providing a precise volatility estimate is difficult; therefore, they used a possibility distribution to model volatility uncertainty and to price an American option in a multi- period binomial model. Carlsson and Fuller (2003) mentioned that the imprecision in judging or estimating future cash flows is not stochastic in nature, and that the use of the probability theory leads to a misleading level of precision. Their study introduced a real option rule in a fuzzy setting in which the present values of expected cash flows and expected costs are estimated by trapezoi- dal fuzzy numbers. They determined the optimal exercise time with the help of possibilistic mean value and variance of fuzzy numbers. The proposed model that incorporates subjective judg- ments and statistical uncertainties may give investors a better understanding of the problem when making investment decisions. Carlsson, Fuller, Heikkila, and Majlender (2007) also developed a methodology for valuing options on R&D projects, in which future cash flows were estimated by trapezoidal fuzzy numbers. In partic- ular, they presented a fuzzy mixed integer programming model for the R&D optimal portfolio selection problem. In addition to the binomial model, the Black–Scholes model is another way to evaluate the option’s value. Owing to fluctuations in the financial market from time to time, some input parameters in the Black–Scholes formula cannot be expected to always be pre- cise. Wu (2004) applied the fuzzy set theory to the Black–Scholes for- mula. Under the assumptions of fuzzy interest rate, fuzzy volatility and fuzzy stock price, the European option price turns into a fuzzy number. This allows the financial analyst to pick a European option price with an acceptable degree of belief. Lee, Tzeng, and Wang (2005) adopted the fuzzy decision theory and Bayes’ rule as a basis for measuring fuzziness in the practice of option analysis. Their study also employed ‘‘Fuzzy Decision Space’’ that consisted of four dimensions—fuzzy state, fuzzy sample information, fuzzy action and evaluation function—to describe the decisions of investors. These dimensions were used to derive a fuzzy Black–Scholes option pricing model under fuzzy environments. Thiagarajah, Appadoo, and Thavaneswaran (2007) also addressed that most stochastic models involve uncertainty arising mainly from lack of knowledge or from inherent vagueness. Traditionally, these stochastic models are solved using probability theory and fuzzy set theory. In their study, using adaptive fuzzy numbers, they modeled the uncertainty of characteristics such as interest rate, volatility, and stock price. They also replaced fuzzy interest rate, fuzzy stock price and fuzzy volatil- ity with possibilistic mean values in the fuzzy Black–Scholes formula. Making a R&D portfolio decision is difficult, because the long lead times of R&D and the market and technology dynamics lead to unavailable or unreliable collected data for portfolio manage- ment. Wang and Hwang (2007) developed a fuzzy R&D portfolio selection model to hedge against the R&D uncertainty. Since tradi- tional project valuation methods often underestimated the risky project, a fuzzy compound-options model was used to evaluate the value of each R&D project. The R&D portfolio selection problem was formulated as a fuzzy zero-one integer programming model that could handle both uncertain and flexible parameters to deter- mine the optimal project portfolio. From the viewpoint of fuzzy random variables, Yoshida, Yasuda, Nakagami, and Kurano (2006) discussed, under uncertainty in financial engineering, an American put option model that was based on the Black–Scholes stochastic model. In their study, prob- ability is applied as the uncertainty such that something occurs or not with probability, and fuzziness is applied as the uncertainty such that the exact values cannot be specified because of a lack of knowledge regarding the present stock market. By introducing fuzzy logic to the log-normal stochastic processes for the financial market, they presented a model with uncertainty of both random- ness and fuzziness in output. The Garman–Kohlhagen (G–K) model is a closed-form solution of the European currency options pricing model based on the Black–Scholes model, but the input variables of the G–K model are usually regarded as real numbers. However, it is more suitable and realistic to price currency options with fuzzy numbers because these variables are only available with imprecise data or data related in a vague way. Therefore, Liu (2009) started from the fuzzy environments of currency options markets, introduced fuzzy tech- niques, and created a fuzzy currency options pricing model. By turning exchange rate, interest rates and volatility into triangular fuzzy numbers, the currency option price turns into a fuzzy num- ber. This allows financial investors to pick any currency option price with an acceptable degree of belief. Fig. 1. The single period binomial tree of underlying asset value. Fig. 2. The dynamics of option value. 15298 S.-H. Ho, S.-H. Liao / Expert Systems with Applications 38 (2011) 15296–15302 From a modeling perspective, real options valuation methods have tended to follow financial option pricing techniques. The Black–Scholes models are used to evaluate simple real option sce- narios such as delay decisions, research and development, licenses, patents, growth opportunities, and abandonment scenarios (Miller & Bertus, 2005). Despite its theoretical appeal, however, the prac- tical use of real option valuation techniques in industry has been limited by the complexity of these techniques, the resulting lack of intuition associated with the solution process, or the restrictive assumptions required for obtaining analytical solutions. On the other hand, Cox, Ross, and Rubinstein (1979) developed a binomial discrete-time option valuation technique that has gained similar popularity to evaluate real options due to its intuitive nature, ease of implementation, and wide applicability to variety of option attributes. In addition, analytical models such as the Black–Scholes formula focus on a single option and cannot deal with multi-option situations. Therefore, the binomial model is adopted as a basis to develop the fuzzy valuation approach in this study. 3. The valuation approach 3.1. Expanded net present value The NPV approach assumes a fixed scenario, in which a company starts and completes a project that then generates cash flows during some expected lifetime without any contingencies. The approach anticipates no contingency for delaying or abandoning the project if market conditions turn sour. However, the assumption about NPV does not fit the actual situation. In reality, if the market is unfa- vorable, the project could be postponed to undertake until market conditions turn better; or, the project may be abandoned during the operation to reduce losses; or, the project may be expanded or extended as market conditions turn around. The flexibilities of these investment decisions indicate that decision makers are capable of restricting loss risks and retaining the potential to raise profits infi- nitely. As a result, the valuation should include these flexibilities which are embedded as real options in investment projects. In considering option value, the traditional NPV can be expanded as: expanded NPV = static NPV + value of option from active management (Trigeorgis, 1993). The expanded NPV is also called strategic NPV. Static NPV is the NPV obtained using the tra- ditional discount method; it is also called passive NPV. 3.2. The fuzzy binomial valuation approach In this study, a fuzzy binomial valuation approach is proposed to evaluate investment projects that are embedded with real options. The value of the project is represented by its expanded NPV, which can be evaluated by the valuation approach. However, the parame- ters are estimated by fuzzy numbers when the expanded NPV is estimated; thus, the expanded NPV is called fuzzy expanded NPV (FENPV) in this study. The proposed valuation approach is based on Cox et al. (1979). Assuming there is a call option with the present value of underly- ing asset S0 and exercising price K, the value of the underlying asset has Pu probability to rise to uS0 or Pd probability to drop to dS0 in the next period. The factors u and d represent the jumping up and down factors of the underlying asset’s present value, respec- tively. A single period binomial tree of the underlying asset value is shown in Fig. 1. The option will be exercised at period t = 1 if the underlying value is higher than K, and forgone if the underlying value is lower than K. The dynamics of the option value is shown in Fig. 2. If the option is sold at price C0, then the pricing approach is generally based on the assumption of replicating portfolio and can thus be determined by the following expression C0 ¼ 1 ð1 þ rÞ ½PuC1u þ Pd C1d� ð1Þ in which r is risk-free interest rate, and Pu and Pd are risk-neutral probabilities, which are determined by the following formulas. Pu ¼ ð1 þ rÞ� d ðu � dÞ ð2Þ Pd ¼ u �ð1 þ rÞ ðu � dÞ ¼ 1 � Pu ð3Þ Therefore, the price or present value of the call option is the discounted result of the option values C1u and C1d with risk-neutral probabilities. Also, under the assumption of no arbitrage opportuni- ties, the condition 0 < d < 1 < (1 + r) < u must be satisfied. Further- more, the expected return of the underlying asset should be zero based on the no-arbitrage assumption: Pu uS0 1 þ r � S0 � � þ Pd dS0 1 þ r � S0 � � ¼ 0 ð4Þ That is uPu 1 þ r þ dPd 1 þ r ¼ 1 ð5Þ Thus, we have the following risk-neutral probabilities equations: Pu þ Pd ¼ 1 uPu 1þr þ dPd 1þr ¼ 1 ( ð6Þ From (1)–(3), we know that the main factors affecting the call option value are jumping factors u and d; it is not easy, however, to estimate their values in a precise manner due to the uncertainty of the underlying volatility. The cash flow models applied to many financial decision mak- ing problems often involve some degree of uncertainty. In the case of deficient data, most decision makers tend to rely on experts’ Fig. 3. A FENPV with right-skewed distribution. S.-H. Ho, S.-H. Liao / Expert Systems with Applications 38 (2011) 15296–15302 15299 knowledge of financial information when carrying out their finan- cial modeling activities. The nature of this knowledge often tends to be vague rather than random. Fuzzy theory, which is aimed at rationalizing the uncertainty caused by vagueness or imprecision, has provided a promising basis for manipulating such vague knowledge (Sheen, 2005). In the relevant application of financial decision making, there is an example of employing triangular fuzzy numbers to examine the profitability indexes (Chiu & Park, 1994). In strategic or innovative investment projects, information often tends to be vague rather than random. Therefore, this study considers possibilistic uncertainty rather than probabilistic uncer- tainty and employs fuzzy numbers instead of statistics to estimate the parameters. For lightening computation efforts, we utilize the triangular fuzzy numbers ~u ¼ u1; u2; u3½ � and ~d ¼ d1; d2; d3½ � to rep- resent the jumping factors of the underlying asset. Therefore, the risk-neutral probabilities equations can be rewritten as ~Pu � ~Pd ¼ ~1 ~u� ~Pu 1þr � ~d� ~Pd 1þr ¼ ~1 ( ð7Þ where fPu ¼ Pu1 ; Pu2 ; Pu3� � and ~Pd ¼ Pd1 ; Pd2 ; Pd3� �. Thus, we have Pu1 ; Pu2 ; Pu3 � � � Pd1 ; Pd2 ; Pd3 � � ¼ ½1; 1; 1� u1;u2;u3½ �� Pu1 ;Pu2 ;Pu3½ �u 1þr � d1;d2;d3½ �� Pd1 ;Pd2 ;Pd3½ � 1þr ¼ ½1; 1; 1� 8< : ð8Þ which are Pui þ Pdi ¼ 1 ui�Pui 1þr þ di�Pdi 1þr ¼ 1 ( for i ¼ 1; 2; 3 ð9Þ It can be solved by considering the following relationship Pui ¼ ð1 þ rÞ� di ui � di ð10Þ Pdi ¼ ui �ð1 þ rÞ ui � di ð11Þ Since the risk-free interest rate r and the exercising price K are usually known, they are crisp values, whereas, the option values C1u and C1d become fuzzy numbers as a result of the jumping factors being fuzzified. That is, ~C1u ¼ maxð~uS0 � K; 0Þ and ~C1d ¼ maxð~dS0 � K; 0Þ. The ranking of two triangular fuzzy num- bers ~A ¼ ½a1; a2; a3� and ~B ¼ ½b1; b2; b3� can be derived from maxð~A; ~BÞ¼ ½maxða1; b1Þ; maxða2; b2Þ; maxða3; b3Þ�. Thus, the pricing formula for the fuzzy call option is eC 0 ¼ 11 þ r eP d � eC 1d � eP u � ~C1u h i ð12Þ In practical application, the present value of the underlying asset is determined by the NPV of the investment project; the exercising price is the additional outlay to exercise the embedded option. Managerial flexibility to adopt future actions introduces an asymmetry or skewness in the probability distribution of the pro- ject NPV (Yeo & Qiu, 2003). In the absence of such managerial flex- ibility, the probability distribution of project NPV would be considerably symmetric. However, in the existence of managerial flexibility such as the exercising of options, enhanced upside potential is introduced and the resulting actual distribution is skewed to the right. In essence, identical results are obtained in the case of possibi- listic distribution which is adopted by this study to characterize the NPV of an investment project. In other words, the characteristic of right-skewed distribution also appears in the FENPV of an invest- ment project when the parameters (such as cash flows) are charac- terized with fuzzy numbers. Although many studies have proposed a variety of methods to compute the mean value (Carlsson & Fuller, 2001; Fuller & Majlender, 2003) and median value (Bodjanova, 2005) of fuzzy numbers, these works did not consider the right-skewed characteristic present in the FENPV. Therefore, this study proposes a new method to compute the mean value of the FENPV based on its right-skewed characteristic. This mean value can be used to represent the FENPV with a crisp value. Moreover, different FENPVs can be compared according to their mean values. Let ~C ¼ ½c1ðaÞ; c3ðaÞ� be a fuzzy number and k 2 ½0; 1�. Then, the mean value of ~C is defined as E ~C � � ¼ Z 1 0 ð1 � kÞc1ðaÞþ kc3ðaÞ½ �da ð13Þ The weighted index k is called the pessimistic-optimistic index in Yoshida et al. (2006), but the index is determined by a subjective decision in Yoshida et al. (2006). However, this study considers that the index can be determined objectively. Fig. 3 illustrates a case in which the FENPV is represented by a right-skewed triangular fuzzy number. The right-skewed characteristic of FENPV—meaning that the more skew to the right, the more optimistic the payoff of the project—provides a clue to determining k with k ¼ ARALþAR , where AL and AR are the left-part area and right-part area of the FENPV, respectively. Thus, when k is determined objectively and substi- tuted into Eq. (13), the mean value of the FENPV can be computed as follows EðFENPVÞ¼ ð1 � kÞc1 þ c2 þ kc3 2 ð14Þ 4. An illustrative example An enterprise must continually develop new products and introduce them into the market to create profit. Therefore, evaluat- ing projects of new product development is a crucial task that should be an ongoing effort of an enterprise. In this case, a local biotechnology company in Taiwan proposes a new product devel- opment project that needs evaluation. The project must go through two stages before the new product can be introduced into the mar- ket. Stage one of the project will require two years and an invest- ment of I1 = 40 (million NT$) toward product development. When this is done, the project will proceed to the second stage, which will require one year and an outlay of I2 = 80 (million NT$) to acquire the equipment and raw material for mass production. Experts estimate that the project will create cash inflows with a present value of 100 (million NT$). If we use the biannual risk-free interest rate r = 3% as the discounting rate and frame six months as one period, the NPV of the project can be calculated as follows: NPV ¼ 100 � 40 � 80 ð1 þ 0:03Þ4 ¼�11:08ðmillionÞ This negative NPV suggests that the project should be rejected. The above results are obtained under the assumption that cash inflows can be generated with certainty. However, this assumption is unrealistic. In fact, the cash inflows will vary with fluctuations in market conditions, such as the market demand of the new product. Fig. 4. Binomial tree of the project’s cash inflows. Fig. 5. The decision tree of the project with the option to defer. Fig. 6. The decision tree of the project with the option to abandon. 15300 S.-H. Ho, S.-H. Liao / Expert Systems with Applications 38 (2011) 15296–15302 According to experts’ estimation, the new product may have a rate of 20% � (1 ± 5%) fluctuation per year with regard to its market demand. Since the volatility is estimated under uncertainty, a trian- gular fuzzy number is employed to characterize the possibilistic uncertainty of the volatility. Based on the estimation, the triangular fuzzy number ~q ¼ ½ð1 � 0:05Þ� 0:2; 0:2; ð1 þ 0:05Þ� 0:2� ¼ ½0:19; 0:25em0:2; 0:21� is used to express the fuzzy volatility. From the fuz- zy volatility ~q, the fuzzy jumping factors ~u and ~d can be determined as ~u ¼ expð~q � ffiffiffi s p Þ and ~d ¼ 1=~u, where s is the chosen time interval expressed in the same unit as ~q and exp denotes the exponential function. In this case, the value of s is 0.5 because there are six months (0.5 y) in each period. As a result, we have ~u ¼ ½1:1438; 1:1519; 1:1601� and ~d ¼ ½0:8620; 0:8681; 0:8743�. The fuzzy risk- neutral probabilities are eP u ¼ ½0:5448; 0:5704; 0:5962� and eP d ¼ ½0:4038; 0:4296; 0:4552�, respectively. With the above conditions, a binomial tree of the project’s cash inflows can be established, as shown in Fig. 4. (For simplicity, the numbers in the binomial tree are represented to two digits after the decimal point.) Nevertheless, the project may have some decision flexibilities when the project is undertaken. For instance, when the market con- ditions are unfavorable, the project can be deferred one period to undertake or the project can abandon its second stage investment to prevent losses from mass production. Therefore, the project with deferring option and abandoning option will be evaluated in the following subsections, respectively. Moreover, the project with a sequential multiple options which is combined with deferring option and abandoning option will also be evaluated. 4.1. Option to defer First of all, considering the situation that decision maker defers one period to undertake the first stage investment and commits to undertake the second stage investment. In this case, the project’s total outlay that discounted to period one is calculated as follows: I defer ¼ 40 �ð1 þ 0:03Þþ 80 ð1 þ 0:03Þ3 ¼ 114:41 The decision tree is shown in Fig. 5, where V = 100,eI defer ¼ ½114:41; 114:41; 114:41� and ~0 ¼ ½0; 0; 0�. The root value in Fig. 5 is the FENPV of the project with deferring option and can be calculated as follows: FENPV ¼ eP u �½eC 1u � eP d � eC 1dh i=ð1 þ 0:03Þ¼ ½0; 0:43; 0:92� The mean value of the FENPV is 0.46 (million), and the value of the option to defer the first stage investment is 0.46 � (�11.08) = 11.54 (million NT$). 4.2. Option to abandon Furthermore, when the decision maker only possesses the option to abandon the second stage investment, this implies that the deci- sion maker has already completed the first stage investment without deferring. The decision tree is shown in Fig. 6, in which eI 2 ¼ ½80; 80; 80�. From the root value in Fig. 6, we can conclude that the FENPV of the project with option to abandon the second stage invest- ment is FENPV = [22.95, 30.37, 39.69] �~I1 = [�17.05, �9.64, �0.31], where eI 1 ¼ ½40; 40; 40�. In this case, the mean value of the FENPV is �8.68 (million), and thus, the value of the option to abandon the second stage investment is �8.68 � (�11.08) = 2.4 (million). 4.3. Sequential multiple options Finally, when the project involves these two options but with different expiration days, these two options form a sequential mul- tiple options. In the sequential multiple options, decision makers have the options not only to abandon the second stage investment but also to defer the first stage investment. Therefore, the decision in period one is maxðeC 1u �eI 1; ~0Þ and maxðeC 1d �eI 1; ~0Þ, where eC 1u and eC 1d are the project values in the up and down cases, respec- tively, during period one. Based on the values at period two, we can find that eC 1u ¼ ½33:81; 42:32; 52:45� and eC 1d ¼ ½12:92; 16:62; 21:11�. The FENPV of the project with sequential multiple options is FENPV = [0, 1.28, 7.21], its mean value is 3.60 (million), and the value of the sequential multiple options is 3.60 � (�11.08) = 14.68 (million) (See Fig. 7). 5. Discussion In Table 1, we summarize the evaluation results of the new product development project that embedded with three different real options, respectively. From the evaluation results, we can observe that if the project does not have any decision flexibility, the project’s NPV is �11.08 Fig. 7. The decision tree of the project with sequential multiple options. Table 1 A summary of the results (in million NT$). Type of option FENPV of the project Mean value of the FENPV Option value Option to defer [0, 0.43, 0.92] 0.46 11.54 Option to abandon [�17.05, �9.635, �0.31] �8.68 2.4 Multiple options [0, 1.28, 7.21] 3.60 14.68 S.-H. Ho, S.-H. Liao / Expert Systems with Applications 38 (2011) 15296–15302 15301 (million NT$) and the project should therefore be rejected. How- ever, when the project is embedded with some decision flexibili- ties, the decisions will be different. Confronting uncertain market conditions, the decision flexibilities, such as deferring investment in the first stage or abandoning investment in the second stage, have specific values. In this paper, we have verified the values of these flexibilities from the aspect of fuzzy real options. When the project involves the option to defer investment in the first stage, the mean value of the project’s FENPV is 0.46 (million NT$). The overall value of the project is positive, thus, the project become acceptable. Moreover, the value of the option to defer is 11.54 (million NT$). The option value stems from the flexibility that decision maker can defer investment in the first stage to avoid the downward losses at project initiation. Moreover, when the project includes the option to abandon the second stage investment, the mean value of the project’s FENPV is �8.68 (million NT$). Although this mean value is negative, it is still greater than the original NPV = �11.08 (million NT$). This reveals that the second stage option can still prevent losses when the mar- ket conditions are downward and can retain the upside potential of profit when the market conditions are upward. Therefore, this option to abandon the second stage investment has a value of 2.4 (million NT$)-lower than the value of option to defer. The reason is that the first stage investment has been completed without deferring, whatever the market conditions are. Thus, even though the market conditions are downward at the initiation of the pro- ject, the decision maker will only be able to prevent losses at the second stage. Due to the smaller extent of hedging, the second stage option has a lower option value than the first stage option. Lastly, when both options form a sequential multiple options, the mean value of the project’s FENPV is 3.60 (million NT$), which represents the total value of the project. Since this value is positive, the project is acceptable. The value of the sequential multiple options is 14.68 (million NT$). This option value is higher than the value of a single option. This result shows that the multiple options provide greater value than a single option because multiple options provide more flexibility. However, the value of multiple options does not equate directly to the addition of the values of both options. The value cannot be raised linearly because of the nonlinear operations in the valuation model and the trade-off between both options in the hedging process. 6. Conclusions Since the options-related commodities could not be priced objectively, they were not widely accepted in the market. Until 1973, Black–Scholes proposed a valuation model that allowed investors to price the options; investors no longer needed to rely on subjective judgments. Since then, transactions and innovations in options have continually developed. Options have become the most popular financial commodities and have satisfied the mar- ket’s needs for hedge and arbitrage. There is a similar problem in traditional capital budgeting. Because estimating the value of flexibilities is difficult, the tradi- tional capital budgeting methods cannot discover the value of man- agerial flexibilities or other alternatives within investment projects. As a result, a potential investment project will be undervalued and rejected. The rejection of a potential project may incur substantial losses. Nevertheless, once flexibilities are regarded as real options, they can be evaluated using option valuation models. The values of these flexibilities become measurable and the entire value of an investment project can be revealed. The binomial valuation technique has gained popularity in the valuation of real options due to its intuitive nature, ease of imple- mentation, and capability of dealing with multiple options. Further- more, in an uncertain economic decision making environment, information such as cash flows, interest rate, cost of capital, and so forth possess some vagueness but not randomness (Kahraman, Ruan, & Tolga, 2002). Consequently, this study has proposed the fuz- zy binomial valuation approach to evaluate investment projects with embedded real options in uncertain decision making environ ments. Finally, combining real options theory with other theories has been a significant development in recent years. Smit and Trigeorgis (2006), (2007) combined real options theory with game theory to serve as an analytical instrument for competitive strategies in an uncertain environment. They unify two major strands of economic theory—real options and games—into a single, coherent framework and demonstrate how these ideas can be applied to formulating cor- porate strategy. The integrated options-and-games perspective is particularly relevant for oligopolistic and innovative industries such as consumer electronics, telecommunications or pharmaceuticals. References Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of Political Economy, 81, 637–659. Bodjanova, S. (2005). Median value and median interval of a fuzzy number. Information Sciences, 172, 73–89. Bowe, M., & Lee, D. L. (2004). Project evaluation in the presence of multiple embedded real options: evidence from the Taiwan High-Speed Rail Project. Journal of Asian Economics, 15, 71–98. Bowman, E. H., & Moskowitz, G. T. (2001). Real options analysis and strategic decision making. Organization Science, 12, 772–777. Carlsson, C., & Fuller, R. (2001). On possibilistic mean value and variance of fuzzy numbers. Fuzzy Sets and Systems, 122, 315–326. Carlsson, C., & Fuller, R. (2003). A fuzzy approach to real option valuation. Fuzzy Sets and Systems, 139, 297–312. Carlsson, C., Fuller, R., Heikkila, M., & Majlender, P. (2007). A fuzzy approach to R&D project portfolio selection. International Journal of Approximate Reasoning, 44, 93–105. Chen, T., Zhang, J., & Lai, K. K. (2009). An integrated real options evaluating model for information technology projects under multiple risks. International Journal of Project Management, 27, 776–786. 15302 S.-H. Ho, S.-H. Liao / Expert Systems with Applications 38 (2011) 15296–15302 Chiu, C. Y., & Park, C. S. (1994). Fuzzy cash flow analysis using present worth criterion. The Engineering Economist, 39, 113–137. Cox, J., Ross, S., & Rubinstein, M. (1979). Option pricing: A simplified approach. Journal of Financial Economics, 7, 229–263. Dixit, A. K., & Pindyck, R. S. (1995). The option approach to capital investment. Harvard Business Review, 73, 105–115. Fuller, R., & Majlender, P. (2003). On weighted possibilistic mean and variance of fuzzy numbers. Fuzzy Sets and Systems, 136, 363–374. Hahn, W. J., & Dyer, J. S. (2008). Discrete time modeling of mean-reverting stochastic processes for real option valuation. European Journal of Operational Research, 184, 534–548. Kahraman, C., Ruan, D., & Tolga, E. (2002). Capital budgeting techniques using discounted fuzzy versus probabilistic cash flows. Information Sciences, 142, 57–76. Lee, C. F., Tzeng, G. H., & Wang, S. Y. (2005). A new application of fuzzy set theory to the Black–Scholes option pricing model. Expert Systems with Applications, 29, 330–342. Liu, F. Y. (2009). Pricing currency options based on fuzzy techniques. European Journal of Operational Research, 193, 530–540. Luehrman, T. A. (1998). Investment opportunities as real options: Getting started on the numbers. Harvard Business Review, 51–67. Miller, L., & Bertus, M. (2005). License valuation in the aerospace industry: A real options approach. Review of Financial Economics, 14, 225–239. Muzzioli, S., & Reynaerts, H. (2008). American option pricing with imprecise risk- neutral probabilities. International Journal of Approximate Reasoning, 49, 140–147. Muzzioli, S., & Torricelli, C. (2004). A multiperiod binomial model for pricing options in a vague world. Journal of Economic Dynamics & Control, 28, 861–887. Reyck, B. D., Degraeve, Z., & Vandenborre, R. (2008). Project options valuation with net present value and decision tree analysis. European Journal of Operational Research, 184, 341–355. Sheen, J. N. (2005). Fuzzy financial profitability analyses of demand side management alternatives from participant perspective. Information Sciences, 169, 329–364. Smit, H., & Trigeorgis, L. (2006). Real options and games: Competition, alliances and other applications of valuation and strategy. Review of Financial Economics, 15, 95–112. Smit, H., & Trigeorgis, L. (2007). Strategic options and games in analyzing dynamic technology investments. Long Range Planning, 40, 84–114. Thiagarajah, K., Appadoo, S. S., & Thavaneswaran, A. (2007). Option valuation model with adaptive fuzzy numbers. Computers and Mathematics with Applications, 53, 831–841. Trigeorgis, L. (1993). Real options and interactions with financial flexibility. Financial Management, 22, 202–224. Wang, J., & Hwang, W. L. (2007). A fuzzy set approach for R&D portfolio selection using a real options valuation model. Omega, 35, 247–257. Wu, H. C. (2004). Pricing European options based on the fuzzy pattern of Black– Scholes formula. Computers & Operations Research, 31, 1069–1081. Wu, L. C., Ong, C. S., & Hsu, Y. W. (2008). Active ERP implementation management: A Real Options perspective. Journal of Systems and Software, 81, 1039–1050. Yeo, K. T., & Qiu, F. (2003). The value of managerial flexibility – a real option approach to investment evaluation. International Journal of Project Management, 21, 243–250. Yoshida, Y., Yasuda, M., Nakagami, J., & Kurano, M. (2006). A new evaluation of mean value for fuzzy numbers and its application to American put option under uncertainty. Fuzzy Sets and Systems, 157, 2614–2626. Zadeh, L. A. (1999). Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets and Systems, 100(Supplement), 9–34. Zadeh, L. A. (2005). Toward a generalized theory of uncertainty (GTU) – an outline. Information Sciences, 172, 1–40. Zadeh, L. A. (2008). Is there a need for fuzzy logic? Information Sciences, 178, 2751–2779. A fuzzy real option approach for investment project valuation 1 Introduction 2 Related works 3 The valuation approach 3.1 Expanded net present value 3.2 The fuzzy binomial valuation approach 4 An illustrative example 4.1 Option to defer 4.2 Option to abandon 4.3 Sequential multiple options 5 Discussion 6 Conclusions References 
work_ijpwrqnasvcvxnxbbxgl2gza6m	----	Edinburgh Research Explorer Dynamic prediction of financial distress using Malmquist DEA Citation for published version: Li, Z, Crook, J & Andreeva, G 2017, 'Dynamic prediction of financial distress using Malmquist DEA', Expert Systems with Applications, vol. 80, pp. 94-106. https://doi.org/10.1016/j.eswa.2017.03.017 Digital Object Identifier (DOI): 10.1016/j.eswa.2017.03.017 Link: Link to publication record in Edinburgh Research Explorer Document Version: Peer reviewed version Published In: Expert Systems with Applications General rights Copyright for the publications made accessible via the Edinburgh Research Explorer is retained by the author(s) and / or other copyright owners and it is a condition of accessing these publications that users recognise and abide by the legal requirements associated with these rights. Take down policy The University of Edinburgh has made every reasonable effort to ensure that Edinburgh Research Explorer content complies with UK legislation. If you believe that the public display of this file breaches copyright please contact openaccess@ed.ac.uk providing details, and we will remove access to the work immediately and investigate your claim. Download date: 06. Apr. 2021 https://doi.org/10.1016/j.eswa.2017.03.017 https://doi.org/10.1016/j.eswa.2017.03.017 https://www.research.ed.ac.uk/portal/en/publications/dynamic-prediction-of-financial-distress-using-malmquist-dea(c2c245a9-3a0a-4208-977f-5e5a721e4945).html 1 Dynamic Prediction of Financial Distress Using Malmquist DEA Zhiyong Lia, b, *, Jonathan Crookb, Galina Andreevab a School of Finance, Southwestern University of Finance and Economics, Chengdu 611130, China b Credit Research Centre, University of Edinburgh Business School, Edinburgh EH8 9JS, UK Emails: zhiyong.li@ed.ac.uk (Z. Li), j.crook@ed.ac.uk (J. Crook), Galina.Andreeva@ed.ac.uk (G. Andreeva) *Corresponding author: Zhiyong Li Tel.: +86 185 0822 6040 Address: Room 521, Gezhi Building, School of Finance, SWUFE, 555 Liutai Avenue, Wenjiang, Chengdu 611130, Sichuan, China. 2 Abstract Creditors such as banks frequently use expert systems to support their decisions when issuing loans and credit assessment has been an important area of application of machine learning techniques for decades. In practice, banks are often required to provide the rationale behind their decisions in addition to being able to predict the performance of companies when assessing corporate applicants for loans. One solution is to use Data Envelopment Analysis (DEA) to evaluate multiple decision-making units (DMUs or companies) which are ranked according to the best practice in their industrial sector. A linear programming algorithm is employed to calculate corporate efficiency as a measure to distinguis h healthy companies from those in financial distress. This paper extends the cross-sectional DEA models to time-varying Malmquist DEA, since dynamic predictive models allow one to incorporate changes over time. This decision-support system can adjust the efficiency frontier intelligently over time and make robust predictions. Results based on a sample of 742 Chinese listed companies observed over 10 years suggest that Malmquist DEA offers insights into the competitive position of a company in addition to accurate financial distress predictions based on the DEA efficiency measures. Keywords: Malmquist DEA; Bankruptcy prediction; Financial distress; Efficiency; Dynamic model Highlights: - This paper is the first to apply dynamic time-varying efficiency scores in distress prediction. - Malmquist DEA is used to produce dynamic efficiency scores over several periods. - Various efficiency scores are compared in discrete time hazard models. - A comparison is made between generic and industry-specific models. 3 1. Introduction Decision-support systems are vital for creditors who need to distinguish between firms that will perform well and those that under-perform, and therefore may have difficulties in repaying their loans. Such systems use various techniques including traditional statistical models and expert systems or intelligent machine- learning algorithms to evaluate the creditworthiness of borrowers. Those predictive methods are often regarded as Early Warning Systems to give early signals of potential business bankruptcy or financial distress. Numerous applications of machine learning algorit hms include Neural Networks (NN) by López Iturriaga & Sanz (2015) , Genetic Algorithm (GA) by Gordini (2014), Support Vector Machines (SVM) by Yang, You, & Ji (2011), and Ensemble models that combine several statistical and intelligent classifiers (Fedorova, Gilenko, & Dovzhenko, 2013; Marques, Garcia, & Sanchez, 2012). These studies compared the predictive accuracy of differe nt algorithms and examined statistically significant predictors of bankruptcy or distress. It has been shown that machine learning algorithms outperform statistic methods in terms of classificat io n accuracy because the objective of machine learning is to minimise misclassified errors so that an optimal solution can be found. Yet predictive accuracy is not the only feature that lenders are interested in, since there is a need (and in many cases a regulatory requirement) to understand and explain the risk drivers or factors that affect the probability of financial distress. Many machine- learning techniques, e.g. neural networks cannot provide such an explanation in contrast to Data Envelopment Analysis (DEA), which is an optimisation algorithm based on linear programming. It allows one to find the efficiency frontier (or benchmark) so that relative efficiency of Decision Making Units (DMUs) can be measured by the distance to this frontier. Cielen, Peeters, & Vanhoof (2004) argued that DEA as a type of machine learning technique can provide insights into the value of a company’s efficiency for bankruptcy prediction. The idea was developed further by Xu & Wang (2009), Psillaki, Tsolas, & Margaritis (2010), Premachandra, Chen, & Watson (2011), Shetty, Pakkala, & Mallikarjunappa (2012) etc. These studies demonstrated that corporate efficiency measures can successfully distinguish between ‘good’ and ‘bad’ companies, while Min & Lee (2008) suggested using efficiency scores generated by DEA to predict bankruptcy directly as a practical approach. However, most studies on the significance of efficiency in estimating the probability of financ ia l distress used cross-sectional models that fail to capture temporal changes in efficiency, and yet internal and external conditions associated with company performance do change over time. So as Shumway (2001) argues, dynamic models, in contrast to cross-sectional or static models, are preferred in business failure prediction. Premachandra, Chen, & Watson (2011), Li, Crook, & Andreeva (2014), Wanke, 4 Barros, & Faria (2015) and Wanke, Azad, & Barros (2016) all strongly suggested using dynamic efficiency models to predict the risk of bankruptcy or financial distress. Yet so far, to the best of our knowledge, no study has conducted an analysis of efficiency scores in dynamic prediction models. This paper fills this gap and is the first to develop a dynamic model integrated with Malmquist DEA and hazard models in order to predict financial distress, which can be easily extended to other bankruptcy or business failure prediction models within the scope of credit assessment. We explore the question of whether changes in the efficiency of companies over time really affects the chance that they will suffer financial distress. Our paper adds to the literature in several important ways. First, we propose two stage DEA-based programming models as decision-support systems in identifying efficient (healthy) and ineffic ie nt (distressed) companies in advance. Second, it enhances the bankruptcy prediction literature by providing a dynamic model that offers insights into the competitive position of a business, in addition to accurate distress predictions. Third, we address the methodological limitations of existing studies Xu & Wang (2009), Yeh, Chi, & Hsu (2010), Cielen, Peeters, & Vanhoof (2004), Premachandra, Bhabra, & Sueyoshi (2009), Premachandra, Chen, & Watson (2011) which either assume constant returns to scale (CRS) or homogeneous production technology in their samples. Our estimation of corporate efficiency is more realistic given mixed industrial sectors. This paper builds on the work of Premachandra, Chen, & Watson (2011), Paradi, Asmild, & Simak (2004), Shetty, Pakkala, & Mallikarjunappa (2012) and Li, Crook, & Andreeva (2014), all of which employ cross-sectional analysis. Our analysis is based on a sample of 742 Chinese listed companies observed over 10 years, in total 5,490 company-years. Dynamic DEA efficiency scores calculated by the Malmquist Index are used to classify healthy and distressed companies and to predict the probability of distress as the firm progresses through its lifecycle. Observations are made about the predictive utility of several model specifications with different efficiency scores and assumptions that contribute to decision-support modelling for bankruptcy prediction. We find that computing the efficiency of a company relative to those across a broad range of industries gives more accurate predictions than computing the efficiency relative only to others in the same industry. We also find that if one does the latter, then comparing the efficiency of a company with the most effic ie nt companies at any time throughout the sample period is more accurate than using only technical or super efficiency. The rest of the paper is structured as follows. Section 2 reviews the literature on dynamic corporate credit risk models and dynamic DEA models, which is fundamental to our extensions of previous studies in both areas. Sections 3 and 4 introduce the methodology and the data used, respectively, with the descriptions of the sample and variables. The results of four comparative models that employ 5 different types of efficiency scores are presented and discussed in Section 5, and Section 6 concludes the paper. 2. Literature review 2.1 Dynamic credit risk models Altman (1968) introduced statistical methods (Discriminant Analysis or DA) to corporate bankruptcy prediction before the age of machine learning techniques. Statistical methods estimate coefficients of financial ratios in a parametric format and were more popular compared to machine learning techniques given the limited computing power decades ago. Nevertheless, Shumway (2001) claimed that half of the financial ratios that were found to be successful in cross-sectional models turned out to be unrelated to bankruptcy probability in later periods. As a result Shumway (2001) proposed a hazard model that has advantages over cross-sectional models. First, hazard models incorporate the effect of time on the risk of an event occurring. Second, hazard models can also incorporate Time-Varying Covariates (TVCs) relating to an individual firm and to macroeconomic factors, the latter representing systemic effects. Third, hazard models can make better predictions by utilising data observed over several time periods. Fourth, hazard models can handle censoring: where an event occurs but is not observed in the observation time window. All of these advantages imply that dynamic or hazard models should be preferred in credit risk modelling. Unlike static models, dynamic models imply a time varying hazard rate and thus are more appropriate to make predictions. An event (e.g. default, bankruptcy or financial distress) can happen any time during interval ],[ ttt  ( t is duration time) in Cox proportional hazard regression and that was applied by Bonfim (2009) to Portuguese firms over the period 1996 to 2002 to predict bankruptcy risk during different macroeconomic cycles. In corporate credit, the default event is usually defined to occur within a specific period of time, commonly one year (Carling, Jacobson, Linde, & Roszbach, 2007). Covariates are also observed only at given points of time when financial statements are disclosed. Therefore it is more appropriate to use a discrete time version rather than a continuous time hazard model. The discrete time hazard model (DHM) is equivalent to multi-period logistic regression in terms of computation but with an additiona l term 0 ( )h t as the baseline hazard function. Such a method was applied to predict corporate default risk by Shumway (2001), Carling, Jacobson, Linde, & Roszbach (2007), Nam, Kim, Park, & Lee (2008), Wilson & Altanlar (2014) etc. 6 2.2 Dynamic DEA models If DEA efficiency is one of the covariates in dynamic models, there is an obvious need to evaluate it in multiple periods correspondingly. Whilst conventional (i.e. static) DEA models are constructed for a single period, many researchers and practitioners are interested in how efficiency changes over time. Specifically, if a DMU can be observed at different points of time, its change in efficiency over the periods can be informative for predicting future financial distress. One possible approach is to solve DEA problems period by period separately and build a panel dataset consisting of these efficie nc y scores, as Bryan, Fernando, & Tripathy (2013) suggested. Yet it can be argued that methodologic a lly the scores in different periods are incomparable because DEA scores are based on the frontier formed by the peers in that period. That is, a relative efficiency of 0.5 in the second period may be no better than a relative efficiency of 0.3 in the first period since efficiency also depends on the frontier shift for the industry, for example a change in membership or in technology. A possible solution to this might be to still perform a static DEA analysis for each period separately but in the second stage using a standard regression method to estimate the change over time and then extend it to further periods. Emel, Oral, Reisman, & Yolalan (2003) and Min & Lee (2008) used this two-stage method to forecast DEA scores and hence bankruptcy. However, Cook & Seiford (2009) commented that this approach was unsatisfactory because it failed to capture the interaction of one period with another. Window DEA was introduced by Charnes, Clark, Cooper, & Golany (1984) to deal with the efficiency change in the sense of time series. The idea of Window DEA, similar to other window analyses, is to set up a fixed observation window, and to move it across the whole period. Finally the movements and stability of the results can be analysed across different panel subsets. However Cooper, Seiford, & Tone (2006) argued that a shortcoming of Window DEA was evident in the initial and last period where cases were less well evaluated. The Malmquist DEA model is particularly suitable in dealing with panel data. The original idea of the Malmquist Index (MI) was to compare the production technology of two economies, so it is a bilateral index. Let ( )f x be the production function of an economy, where x is a vector of inputs such as labour and capital. To calculate the MI between Economy a and Economy b of differe nt production functions, we can substitute a x in ( ) b f  and vice versa. So the MI is defined as ( ) ( ) MI ( ) ( ) / ( ) ( ) ( ) ( ) b a a a a a b a a b b b b b a b f f f f f f f f    x x x x x x x x . (1) Inspired by Caves, Christensen, & Diewert (1982) who introduced this index in productivity analysis, Färe, Grosskopf, Lindgren, & Roos (1992) and Färe, Grosskopf, Norris, & Zhang (1994) integrated 7 the MI into DEA and developed a DEA-based Malmquist productivity index. The Malmquist productivity index evaluates the total factor productivity change of a DMU between two periods where a and b in equation (1) each relate to a period. It is defined as the product of efficiency change (catch- up) and technological change (frontier-shift) where the catch-up effect describes how much closer a DMU gets to the most efficient production frontier, and the frontier-shift effect describes the technology improvement in the sample. The decomposed elements of the MI can determine how much of a relative efficiency increase from period t to 1t can be credited to individual effort and how much to industry innovation. The efficiency change reflects the extent to which a DMU improves or worsens its efficiency, while technological change reflects the change of the efficiency frontiers between two periods. Since the introduction of the MI, there have been various studies of productivit y change over time in different fields, for example in Italian manufacturing firms (Costa, 2012), Spanish government tax offices (Fuentes & Lillo-Bañuls, 2015), Taiwanese banks (Shyu & Chiang, 2012) and Korean universities (Sohn & Kim, 2012). These studies looked into the efficiency changes over time to derive managerial implications and strategic recommendations and did not consider distress prediction. Table 1 Summary of corporate credit prediction literature Literature Samp le Event M ethod Ty p e of p rediction Altman (1968) US M anufacturing firms Bankrup tcy DA Static Hua, Wang, Xu, Zhang, & Liang (2007) Chinese listed comp anies Distress SVM Static Sun, Jia, & Li (2011) Chinese listed comp anies Distress Ensemble models Static Yang, You, & Ji (2011) Polish firms Bankrup tcy SVM Static Cao (2012) Chinese listed comp anies Distress Choquet integral Static Fedorova, Gilenko, & Dovzhenko (2013) Russian manufacturing firms Bankrup tcy Ensemble models Static Gordini (2014) Italian manufacturing firms Bankrup tcy GA Static Lóp ez Iturriaga & Sanz (2015) US banks Bankrup tcy NN Static Bonfim (2009) Portuguese firms Default Prop ortional hazard model Dy namic Shumway (2001) US firms Bankrup tcy DHM Dy namic Chava & Jarrow (2004) US firms Bankrup tcy DHM Dy namic Carling, Jacobson, Linde, & Roszbach (2007) Swedish firms Default DHM Dy namic Nam, Kim, Park, & Lee (2008) Korean listed comp anies Bankrup tcy DHM Dy namic Wilson & Altanlar (2014) UK new firms Bankrup tcy DHM Dy namic Emel, Oral, Reisman, & Yolalan (2003) Turkish manufacturing firms Bankrup tcy DEA, DA Static Cielen, Peeters, & Vanhoof (2004) Belgian firms Bankrup tcy DEA Static Paradi, Asmild, & Simak (2004) US M anufacturing firms Bankrup tcy DEA Static M in & Lee (2008) Korean manufacturing firms Bankrup tcy DEA Static Xu & Wang (2009) Chinese listed comp anies Distress DEA+DA, Logit & SVM Static Psillaki, Tsolas, & M argaritis (2010) French manufacturing firms Bankrup tcy DEA+Logit Static Premachandra, Bhabra, & Suey oshi (2009) US firms Bankrup tcy DEA Static Yeh, Chi, & Hsu (2010) Taiwanese manufacturin g firms Bankrup tcy DEA+Rough Set Thoery & SVM Static Premachandra, Chen, & Watson (2011) US firms Bankrup tcy DEA Static Shetty , Pakkala, & M allikarjunap p a (2012) Indian firms Bankrup tcy DEA Static Bry an, Fernando, & Trip athy (2013) US firms Bankrup tcy DEA+DA Static Li, Crook, & Andreeva (2014) Chinese listed comp anies Distress DEA+Logit Static Paradi, Wilson, & Yang (2014) US Non-manufacturing firms Bankrup tcy DEA+DA Static Kingy ens, Paradi, & Tam (2016) US retail comp anies Bankrup tcy DEA Static Yang & Dimitrov (2017) US Non-manufacturing firms Bankrup tcy DEA+SVM Static 8 Studies such as Emel, Oral, Reisman, & Yolalan (2003), Paradi, Asmild, & Simak (2004), Cielen, Peeters, & Vanhoof (2004), have been amongst the first to explore accuracy of bankruptcy predictions using DEA efficiency (see Table 1). It is worth extending it to a panel analysis as Malmquist DEA models proved capable of analysing change in performance over time. Unfortunately, we have only seen applications restricted to cross-sectional analysis, even in the latest studies such as Kingyens, Paradi, & Tam (2016) and Yang & Dimitrov (2017). Having understood the advantages of Malmquist DEA in dealing with panel data over other methods, this paper applies Malmquist DEA scores to dynamic prediction of financial distress by taking the time dimension into account, which bridges the literature of DEA applications and dynamic credit risk modelling. The details of the methodology are presented next. 3. Methodology We compute the MI and dynamic efficiency scores first and then regress financial distress on the indicators of dynamic efficiency and other variables. Negative values are occasionally observed in financial data that can be used as inputs and outputs for DEA. Therefore, dealing with negative values becomes necessary. An appropriate model would be the input-oriented VRS Slack Based Model (SBM), given that only the outputs contain negative values (Cooper, Seiford, & Tone, 2006), as in our case. The Malmquist DEA model is based on these assumptions while other choices may not guarantee both translation and units invariance at the same time. 3.1 Malmquist DEA In order to build the required panel dataset, it is necessary to calculate the efficiency scores for each company in each year of observation. Using Malmquist DEA we assume multiple inputs and outputs when DMUs are repeatedly observed on a certain interval basis. Caves, Christensen, & Diewert (1982) defined a distance function ( )D  based on the Malmquis t productivity index (Malmquist, 1953) to calculate technical efficiency (TE). A company is efficient if ( , ) 1D x y  . Let 0 t x denote a vector of inputs and 0 t y denote a vector of outputs for DMU0, both at period t . The relative efficiency of DMU0 at period t , * 0 0 0 0 ( , ) t t t D x y , is calculated as the amount by which input 0 x can be reduced while producing the given output level 0 y compared to the most efficient company on the frontier. Similarly, 1 1 1 0 0 0 ( , ) t t t D    x y is its efficiency score at period 1t  . Thus, with multiple periods, 1 1 0 0 0 ( , ) t t t D   x y and 1 0 0 0 ( , ) t t t D  x y are actually efficiency scores using a set of 9 inputs and outputs in one period, 1t  and t respectively, compared with the frontier of the other period, t and 1t  respectively. Following the ideas of Farrell (1957) to decompose the total factor productivity into the efficie nc y change (EC) and the technology change (TC), Färe, Grosskopf, Lindgren, & Roos (1992) defined the input-oriented Malmquist productivity index (MI) to measure the productivity change of DMU0 between period t and 1t  as 1 1 1 1 1 1/ 20 0 0 0 0 0 0 1 0 0 0 0 0 0 1 1 1 1 1 1/ 20 0 0 0 0 0 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0 0 ( , ) ( , ) MI [ ] ( , ) ( , ) ( , ) ( , ) ( , ) [ ] ( , ) ( , ) ( , ) t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t D D D D D D D D D D                    x y x y x y x y x y x y x y x y x y x y . (2) The first part is the relative change of efficiency from period t to 1t  . Hence they defined 1 1 1 0 0 0 0 0 0 EC ( , ) ) ( , t t t t t t D D     x y x y (3) and 1 1 1/ 20 0 0 0 0 0 1 1 1 1 0 0 0 0 0 0 ( , ) ( , T ) [ ] ( , , C ) ( ) t t t t t t t t t t t t D D D D        x y x y x y x y . (4) Other than using a distance function ( )D  to calculate efficiency, under the nonparametr ic framework Färe, Grosskopf, Norris, & Zhang (1994) calculated the MI by an oriented radial DEA model, while other DEA models are also suitable (Cooper, Seiford, & Tone, 2006). Let ( 1, , ) t ij x i m and ( 1, , ) t rj y r q denote the inputs and outputs for DMUj ( 1, ,j n ) respectively at any given point of time t . The production possibility set of a VRS model is defined by Cooper, Seiford, & Tone (2006) as ( , ) ( , ) , 0 , 1, 0 n n t t t t j j j j j j X Y x y x x y y                 eλ λ , (5) where e is the row vector with all elements equal to one, n Rλ is the intensity vector. Let 0 0 0 ( , ) t t t  x y be the optimal solution to the programming problem (6): 0 0 0 1 0 0 0 1 ( , ) min 1- m . . 1 , m t t t i t i i t t t t s x s t X Y              x y x λ s 0 y λ eλ λ 0 s 0 , (6) 10 where  s is a vector of slacks, λ is a non-negative vector and 1 1 n j j    . The reciprocal efficiency +1 +1 0 0 0 ( , ) t t t  x y is the optimal solution of equation (7): 1 1 0 0 0 1 0 1 0 1 0 1 ( , ) min 1- m . . 1 , m t t t i t i i t t t t s x s t X Y                  x y x λ s 0 y λ eλ λ 0 s 0 . (7) By solving the linear programmes (6) and (7) four times for 0 0 0 ( , ) t t t  x y , 1 1 1 0 0 0 ( , ) t t t     x y , 1 1 0 0 0 ( , ) t t t    x y , and 1 0 0 0 ( , ) t t t x y  , we have the MI, to 1 0 t t M  , as 1 1 1 1 1 to 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 ( , ) ( , ) ( , ) ( , ) t t t t t t t t t t t t t t M              x y x y x y x y . (8) This study is interested in the relative efficiency of a DMU calculated by the interaction of periods. Multiplying 0 0 0 ( , ) t t t  x y by to 1 0 t t M  will give the relative efficiency at period 1t  compared to period t for each DMU. In the domain of Malmquist DEA, the reference set, namely period t mentioned above, to which the relative efficiency is compared, is of importance in our research. More specifically, the model defined on two consecutive periods t and 1t  can be seen using adjacent periods as the reference set, which was the original design in Färe, Grosskopf, Lindgren, & Roos (1992). Suppose there are five periods 1, , 5t  . By running Malmquist DEA with adjacent references, one can only get relative efficiency 1  , 2  compared to period 1, 3  compared to period 2 and so on. However, it is not intuit i ve for the other relative efficiency of period 3 compared to period 1 or period 4 compared to period 2 . Thus it is not possible to interpret the relative efficiency directly with adjacent moving references. A solution is to use a fixed reference set as suggested by Berg, Førsund, & Jansen (1992). Therefore, in this research, all relative efficiency is referred to the first period as the beginning of the observation. Thus it is not period 1t  compared to period t but period t ( 2t  ) compared to period 1. In this way, it is very likely that in later periods efficiency scores are larger than 1 as technology develops. It should be noted that apart from scores in period 1, all other scores larger than 1 do not necessarily imply being efficient in that period. 3.2 Discrete hazard model with DEA scores 11 Computing efficiency scores calculated by DEA in the first stage and inputting them into the second stage analysis where other classifying methods are involved is a popular approach. Examples can be found in Psillaki, Tsolas, & Margaritis (2010), Xu & Wang (2009), Yeh, Chi, & Hsu (2010), Li, Crook, & Andreeva (2014), etc. The advantage of doing this is that we can evaluate the marginal effects and the statistical significance of new variables conditional on other covariates such as financial ratios , which have been shown to have predictive power in detecting potential bankruptcy risk. Dyson et al. (2001) and Li, Crook, & Andreeva (2014) have argued that homogeneity of DMUs in terms of technology is important both in the DEA modelling step and in the regression analysis. Thus industry diversity becomes a critical issue in the process. Although the homogeneity requirement, as Li, Crook, & Andreeva (2014) commented, may increase the complexity of modelling, it is consistent with the findings in corporate credit risk modelling that attention should be given to the differe nces between industrial sectors (Bonfim, 2009; Chava & Jarrow, 2004). This research employs this two-stage modelling process and takes into account industrial differe nces separately in both stages, i.e. the sectors are separated in the Malmquist DEA programming, and in the discrete hazard model with the sectors being represented by dummy variables. After Malmquist efficiency scores of multiple periods are obtained in the way given in the last section, they enter the discrete time hazard model with sector dummies written as 1 0 0 1, , , 2 2 , 2 1 logit( ( )) ( ) S T e T r d s s i s t i t s h t h t D          β x β x , (9) where 1d  when a company suffers financial distress, 0 otherwise; 0 ( )h t is the baseline hazard function; , , 2 e i s t x is a vector of efficiency scores for sector s company i at time 2t  , 3, 4,..., i t T ; , 2 r i t x is a vector of financial ratios for company i at time 3, 4,..., i t T ; 1 s D  if company i is a member of sector s , 0 otherwise, 1,...,s S ; 0  is the coefficient of the baseline hazard to be estimated; 1,s β is a vector of parameters for efficiency scores for sector s to be estimated; 2 β is a vector of parameters for financial ratios to be estimated. MaxDEA Pro 6.1 is used to solve Malmquist DEA problems indicated in equations (6), (7) and (8). It can handle unbalanced panel datasets where some cases of late entry or early exit are censored. Combined with equation (9), the base model for financial distress prediction is proposed. 12 3.3 Global reference Further to fixed reference, Pastor & Lovell (2005) introduced the idea of global reference. In some cases where efficient frontiers of different periods cross each other, a global reference set represents the best practices in all periods. For example, in Figure 1, there are four DMUs lying on each of two frontiers ABCD and EFGH. The DMU to be evaluated, unit N, could be referred to frontier ABCD, frontier EFGH or the most efficient units ever over the observation period AFBGH. It is acceptable that when the observation window is long enough, all DMUs at the current period are under the cover of the best historical DMUs, possibly including themselves. Thus the relative efficiency in this circumstance can be treated as absolute efficiency, if the sample is very large. The scores to global reference would be less than or equal to 1. In practice, when the model is built, it is the historic data prior to the current moment that is used in model training and the historic global reference of the past that is available. Therefore, the efficiency calculated by the global reference as an option is embedded into the comparative models. Figure 1 An example of global reference 3.4 Super efficiency Cooper, Seiford, & Tone (2006) referred to the two intertemporal scores, 1 1 0 0 0 ( , ) t t t    x y and 1 0 0 0 ( , ) t t t   x y as the ‘exclusive schemes’. They explain that the exclusive scheme in solving intertempora l programming treats the DMU in the period to be evaluated as having been removed from the evaluator group of the other period. This is mathematically equivalent to what is known as ‘super efficiency’ in DEA. Super efficiency is used as a solution to the problem that common DEA models do not provide 13 an efficiency ranking for efficient units as their scores are all equal to 1 (Andersen & Petersen, 1993). The difference between a super efficiency model and standard models is that in super efficiency models the DMU to be evaluated is eliminated from the reference frontier, so its score can be greater than 1, as shown in Figure 2. Units A, B, C, D and E consist of the productivity possibility set. If unit E is to be evaluated, its efficiency score is 1 as it is on the frontier AECD of standard DEA models; in super models, the new frontier ABCD is employed. For another unit C, its new reference frontier is AED. In this way, though units E and C are both efficient (score = 1) in standard models, a difference between them can be observed by obtaining a new unbound score greater than 1. Figure 2 An example of super efficiency DEA as a frontier technique is arguably an outlier analysis. However, extreme outliers may change the local frontier sufficiently for other units referred to it to be incorrectly measured. In this circumstance, super efficiency can be used to identify outliers (Banker & Chang, 2006). Obviously, super efficiency scores offer more discriminant power between efficient units, which is particular ly useful in classifying good and bad companies in credit risk models. This can be found in the model of Premachandra, Chen, & Watson (2011) who employed super efficiency scores to predict corporate failure. Our paper considers super efficiency in a model for comparison with the base model. The Malmquist SBM DEA model with super efficiency is described by Tone (2002) where 0 = i i t i s x   as: 14 0 0 0 , 1 0 1 0 1 ( , ) min 1+ m . . (1+ ) ( 1, , ) 1 , m t t t i i m t i i j ij i t t s t x x i m Y                ξ λ x y y λ eλ λ 0 ξ 0 . (10) 3.5 Model specification This paper uses a two-stage analysis. In the first stage, DEA efficiency scores for each company at each period are calculated by DEA models defined in the previous sections. In the second stage, the proposed discrete hazard model incorporates efficiency scores as covariates in a panel dataset. Four models outlined in Figure 3 are to be compared with each other for the reasons given below. Figure 3 Model specification Model One, as introduced in Section 3.1 and 3.2, uses Malmquist DEA scores calculated by equations (6), (7) and (8) as covariates in the hazard model (equation (9)). The efficiency score is Technical Efficiency under the VRS assumption. Model One is the base model of this research and predicts the probability of financial distress in two years’ time given that the company has survived until the time of prediction. Model Two applies global reference as introduced in Section 3.3. In this model, the relative efficiency score of company i in sector s at period t is calculated with reference to the most 15 productive companies in all possible periods in the same sector. It is of interest to investigate the predictive power of efficiency scores compared in a cross-period scenario. Model Three follows the super efficiency setting in Section 3.4. Super efficiency scores provide more discrimination for efficient companies so one may expect to see some improvement in predictive accuracy. Model Four is the simplest method in terms of DEA calculation and regression, but heterogeneous technologies are combined. In the first stage when calculating the efficiency scores, all industries are pooled together so the efficient frontier may be pushed outward as more units are considered. In the second stage, the term 1, , , 2 1 S T e s s i s t s D    β x in equation (9) is replaced by a simpler form 1 , 2 T e i t β x without the sector dummies. This is in line with previous literature that pools heterogeneous samples for bankruptcy prediction. This approach is referred to as the ‘generic’ model. The predictive accuracy of models is compared on the out-of-time test sample by standard measures used in predictive modelling (Lessmann, Baesens, Seow, & Thomas, 2015): AUC (the area under Receiver Operating Curve), KS (the Kolmogorov-Smirnov statistic), Type I error (a distressed company that is wrongly classified as a non-distressed company) and Type II error (a non-distressed company that is wrongly classified as a distressed company). For the latter two measures the cut-off is set to the percentage of the distressed companies in the training set. 4. Data 4.1 Sample description The data in this research is taken from the two Chinese stock exchanges, which by 2014 were listing over 2,500 Chinese companies. The Chinese government impose ‘Special Treatment’ (ST) on listed companies in financial distress, so ST is chosen as the official indicator of distress (marked as 1d  ) in this research. Predicting financial distress of Chinese listed companies indicated by ST is consistent with many previous studies using various machine learning techniques, for example, Hua, Wang, Xu, Zhang, & Liang (2007), Sun, Jia, & Li (2011) and Cao (2012). Since the number of employees is only available in annual reports after the year 2000, the observation period is between 2001 and 2010. After the initial filtering, 2,027 individually listed companies over the period 2001 to 2010, a total of 12,431 firm years were left in the sample for analysis. Among them, there are 12,058 healthy firm years and 373 distressed firm years, giving a distress rate of approximately 3%. 16 As industry classification is essential to this research, the starting point was to consider all industr ies. Banking and insurance companies are excluded from the sample as their accounting conventions are different from those in other sectors. For some industries the ST numbers were very low, so we focus on three industries (a total of 742 individual companies) that accounted for nearly half of all distressed cases (49.87%). These are Raw Materials (sector code 1510), Industrial Equipment (sector code 2010) and Real Estate (sector code 4040). A sufficient number of ST companies is necessary for the follow ing two reasons. First, as the panel analysis covers ten years, the valid number of firm years falling in each period cannot be too small. Second, DEA models require that in each period the number of units is more than double the number of inputs and outputs (8 in our case) for good estimates (Dyson et al., 2001). In the end, 5,490 firm years in these three industries were left in the sample. Table 2 indicates that the average distress rate across all years is 3.37% (185/5490=3.37% ). The average number of observations for each company in ten years is 7.4 ( 5490 / 742 7.40 ). In the years 2002, 2003 and 2006 there are significantly more companies suffering financial distress than in other years. Figure 4 shows the distress rate of the three sample industries in each period of observation. In most years, there are more distressed Real Estate companies than in the other two sectors. In later years (2008 to 2010) the distress rate is considerably lower than those during 2002-2003 and 2006-2007. The whole sample is split into a training set covering 2001-2008, eight years, and the test set includes companies with covariates measured over 2009 to 2010 and a distress/non-distress indicator measured in 2011 and 2012. This provides an out-of-time sample and is consistent with Shumway (2001) and other studies using DHM (Nam, Kim, Park, & Lee, 2008; Wilson & Altanlar, 2014). Table 2 Distributions of samples in three industries over 2001-2010 Sector Raw Materials Industrial Equipment Real Estate Total N 328 277 137 742 year Distress Total Distress Total Distress Total Distress Total 0 1 0 1 0 1 0 1 2001 208 1 209 167 5 172 121 5 126 496 11 507 2002 217 8 225 176 6 182 115 9 124 508 23 531 2003 224 9 233 181 8 189 104 11 115 509 28 537 2004 234 6 240 193 7 200 101 5 106 528 18 546 2005 231 5 236 189 6 195 98 3 101 518 14 532 2006 237 7 244 189 13 202 86 14 100 512 34 546 2007 251 9 260 214 4 218 82 6 88 547 19 566 2008 273 5 278 221 5 226 82 2 84 576 12 588 2009 265 6 271 219 2 221 80 2 82 564 10 574 2010 254 10 264 214 5 219 79 1 80 547 16 563 Total 2394 66 2460 1963 61 2024 948 58 1006 5305 185 5490 17 Figure 4 Distress rates of the three sectors over 2001-2010 4.2 DEA inputs and outputs Variables for the MI are selected from physical or monetary items that are contained in standard annual reports. There are five inputs: number of employees, total liabilities, total costs, total assets, and share capital, and three outputs: total profits, total cash flow and total sales. The reason for keeping both total sales and total profits is that a large revenue does not necessarily imply a large profit. Having correlated variables in DEA does not lead to a problem because their weights can automatically adjust without a significant impact on the efficiency score (Dyson et al., 2001). On the contrary, Dyson et al. (2001) (p.249) argued ‘omission of a highly correlated variable can on occasion lead to significa nt changes in efficiencies’. This argument may also be applied to the inclusion of both the number of employees and total costs as the latter covers labour costs. Therefore, simplistically, companies make use of resources (measured by total assets and share capital), hire people (measured by the number of employees), pay for labour and raw materials (measured by total costs), turn them into products and services, sell them for revenue (measured by total sales) and aim for large earnings (measured by total profits) and positive cash inflow (measured by total cash flow). The descriptive statistics of the covariates are reported in aggregate because Malmquist DEA models are estimated on the whole dataset (both training and test samples). For convenience of presentation, only graphs of means over time are presented in Figure 5. 18 Figure 5 Descriptions of DEA variables over 2001-2010 Number of employees Debts (mCNY) Costs (mCNY) Assets (mCNY) Capitals (mCNY) Profits (mCNY) Cash (mCNY) Sales (mCNY) Generally, the size of listed companies (in terms of total assets) increased over the ten years under study. Their total debts, total costs and share capital had similar growth rates. It is the same for total 19 sales as for output. However, there were some changes that did not follow the trend. For example, the number of employees in sector 2010 (Industrial Equipment) nearly doubled in the three years 2006 - 2008. There was a noticeably large drop in profit for sector 1510 (Raw Materials) in years 2008 and 2009, which might be due to the influence of the global financial crisis. Additionally, there was a sharp net cash inflow in sector 4040 (Real Estate) in 2009. These large changes highlight the importance of running DEA peer comparison analyses separately for each industry so that the relative efficie nc y scores are not biased. 4.3 Duration time and variables The sample of this study consists of listed companies so following Shumway (2001) we choose the stock trading age to be the duration time in the hazard model, because companies met the same requirements to be listed on an exchange. The average trading age in the sample is 7.79 years. Following experimentation the baseline function of the duration time, ln( )t , proved to be a good fit in the models, as in Shumway (2001). The indicator of financial distress, ST, is a status indicator where a company can go to ST and recover from it. Here, only the first occurrence of ST is regarded as the event of distress and the information after that is ignored. All companies at the time of entering the observation window in 2001 are healthy companies (not in the status of ST). So the model predicts the probability of a company going into financial distress (ST) for the first time in the next two years, conditional on lagged values of covariates, given the duration time since the company was listed on the stock exchange. Six categories of financial ratios are considered in the regression model: profitability, liabilities and liquidity, capital and asset composition, cash flow, operation and growth rate. In the preliminar y analysis of group mean difference tests and collinearity, one ratio from each category is selected to represent that aspect of a company’s financial position. These six ratios are Return on Equity, Current Liabilities / Total Liabilities, Tangible Assets / Total Assets, Cash Flow from Operation per Share, Total Assets Turnover and Total Assets Growth. 5. Results 5.1 Dynamic DEA score 20 Table 3 Description of efficiency scores Sector Distress Stats Technical Efficiency Global Efficiency Super Efficiency Generic Efficiency Raw Materials No N 2394 2394 2394 2394 Mean 1.225 0.653 1.234 0.918 SD 0.772 0.167 0.788 0.315 Min 0.029 0.021 0.029 0.028 Max 8.638 1 8.638 5 Yes N 66 66 66 66 Mean 0.756 0.464 0.756 0.67 SD 0.487 0.207 0.487 0.313 Min 0.04 0.026 0.04 0.035 Max 3.458 0.929 3.458 2.234 Mean dif. btw groups F 17.484 53.593 21.914 30.198 p 0.000 0.000 0.000 0.000 Industrial Equipment No N 1963 1963 1963 1963 Mean 1.033 0.457 1.059 0.936 SD 0.53 0.18 0.603 0.405 Min 0.07 0.032 0.07 0.062 Max 7.378 1 7.378 6.178 Yes N 61 61 61 61 Mean 0.702 0.307 0.702 0.664 SD 0.329 0.193 0.329 0.293 Min 0.004 0.002 0.004 0.003 Max 1.93 1 1.93 1.821 Mean dif. btw groups F 4.970 13.121 25.262 28.014 p 0.026 0.000 0.000 0.000 Real Estate No N 948 948 948 948 Mean 0.972 0.702 1.017 0.929 SD 0.569 0.196 0.666 0.503 Min 0.059 0.049 0.059 0.059 Max 6.277 1 8.204 5.931 Yes N 58 58 58 58 Mean 0.66 0.518 0.667 0.628 SD 0.381 0.26 0.389 0.35 Min 0.024 0.021 0.024 0.023 Max 1.805 1 1.805 1.603 Mean dif. btw groups F 3.683 21.919 14.439 22.350 p 0.050 0.000 0.000 0.000 Mean dif. btw sectors F 4.598 104.354 3.192 0.187 21 p 0.010 0.000 0.041 0.829 Figure 6 Distributions of efficiency scores over 2001-2010 Technical Efficiency Global Efficiency Super Efficiency Generic Efficiency In order to see how efficiency and technology change over time, we plot graphs of mean efficiency scores across all periods in Figure 6. For convenience, graphs for three industrial sectors are drawn on one chart. Generally the efficiency and technology levels increased while, in the later years, there were some declines, presumably due to the influence of the financial crisis in 2008. 5.2 Regression results Six ratios, together with efficiency scores, are integrated in Models One to Four as showed in Figure 3. The results are presented in Table 4. The 2  tests indicate that all four models explain significa nt amounts of variation in the probability of distress. The coefficient for each type of efficiency has a negative sign, which indicates that the more efficient a company is, the less likely it is to go into financial distress. The value of the coefficient on each type of efficiency differs between the models because their mean values and distributions are different. For Models One to Three, when three 22 industries were treated separately, differences in the parameters are observed. All parameters are significant at the 5% level. The parameters of the financial ratios in Table 4 have the expected signs and all are statistically significant. Table 4 Model results Covariates Model One Model Two Model Three Model Four ln(duration) -0.042 -0.101 0.219 0.194 Technical Efficiency (1510) -2.646** Technical Efficiency (2010) -2.968** Technical Efficiency (4040) -3.056** Global Efficiency (1510) -4.762** Global Efficiency (2010) -7.630** Global Efficiency (4040) -3865** Super Efficiency (1510) -2.565** Super Efficiency (2010) -2.886** Super Efficiency (4040) -2.941** Generic Efficiency -5.485** Return on Equity -9.084** -9.386** -9.105** -8.656** Current Liabilities / Total Liabilities 6.776** 6.670** 6.809** 6.567** Tangible Assets / Total Assets -1.463** -1.553** -1.508** -1.555** Cash Flow from Operation per Share -0.971** -0.890** -0.964** -0.872** Total Assets Turnover -1.551** -1.136** -1.562** -0.546** Total Assets Growth -2.695** -2.595** -2.777** -2.185** Constant -4.721** -4.499** -4.703** -2.984** Log likelihood -453.36 -448.03 -451.59 -437.53 Number of observations 4017 4017 4017 4017 LR 2  380.4 391.04 383.94 412.06 Prob > 2  0 0 0 0 Pseudo R2 0.2955 0.3038 0.2934 0.3201 ** indicates the coefficients is significant at the 5% level of significance. 5.3 Predictive accuracy Type I error occurs when a distressed company is wrongly classified as a healthy company while Type II error occurs when a healthy company is wrongly classified as a distressed company. The results of Models One to Four in Table 5 show much larger Type I errors in the test set than those in the training set. However, the opposite is the case with Type II errors, which is attributed to the lower 23 distress rate in the later years, as the classifications are based on the cut-off which is measured by the percentage of distressed companies in the training set. The AUC and KS statistics measure relative rank ordering of predicted probabilities of distress for healthy and distressed companies, with higher values corresponding to better models (Lessmann, Baesens, Seow, & Thomas, 2015). Values in Table 5 indicate very similar accuracy of rank orderings between all the models. The overall predictive accuracy is around 95%, which is higher than what is found in the cross sectional logit model combined with DEA efficiency in Xu & Wang (2009) (overall accuracy 91%) and in Li, Crook, & Andreeva (2014) (overall accuracy 93%). Table 5 Predictive accuracy Training set Test set AUC KS Type I error Type II error AUC KS Type I error Type II error Model One 0.881 0.622 58.28% 2.28% 0.861 0.629 88.89% 1.45% Model Two 0.883 0.631 56.95% 2.22% 0.866 0.655 77.78% 1.27% Model Three 0.882 0.632 57.62% 2.25% 0.860 0.625 83.33% 1.36% Model Four 0.898 0.661 56.29% 2.20% 0.880 0.670 72.22% 1.18% Figures in bold indicate the best performance across all models while the next best is marked in italics. Model Four, which disregards industrial classification, seems to do consistently better than the other models. This indicates that by relaxing the assumption of homogeneity between DMUs of DEA and pooling all industries together before calculating DEA scores and probabilities of default in hazard models, practically one is more able to make more accurate predictions of future corporate distress than if we distinguish between industrial sectors. These results are similar to the findings in former studies in consumer credit of Banasik, Crook, & Thomas (1996) and Bijak & Thomas (2012), where segmentation did not produced the expected effect because of the sample size. Model Two is the next best model in terms of AUC and KS on the test set in Table 5. Model Two with global reference has a larger AUC (0.866) than those of Models Two and Four in the out-of-time predictions. The Super Efficiency model (Model Three) was slightly less accurate than other models. This suggests that greater discrimination between the most efficient companies is unnecessary for the prediction of financial distress. 6. Conclusion One of the aims of expert systems and machine- learning algorithms is to provide analytical support for business decisions based on intelligent data analysis. We present one way of providing such support which offers the benefits of insights into relative performance of companies and captures its change 24 over time. We contribute to previous research in the area of DEA in expert systems (Min & Lee (2008); Shetty, Pakkala, & Mallikarjunappa, (2012); Xu & Wang (2009)) by introducing the time-varying component; by comparing industry-specific models against the generic models; and by investigat ing the potential effect on predictive accuracy of different efficiency measure. Dynamic models have inherent advantages over static models in the context of event prediction because conditions and behaviours change over time, so predictions need to be adjusted by incorporating as much information as possible. In this paper, our financial distress hazard models are enhanced by dynamic DEA scores which provide insights into the efficiency of a company relative to others over time. In the domain of DEA, Malmquist DEA is the only one that catches temporal changes of DMUs so it allows efficiency to be compared in both cross-sectional and time series formats. Other DEA algorithms such as standard DEA, Network DEA or Window DEA do not quite fit the bankruptcy prediction paradigm. A Malmquist productivity index is defined as the product of efficiency change (catch-up) and technological change (frontier-shift) and is calculated by the standard DEA scores in two periods and two inter-temporal scores with reference to the efficienc y frontier of the other period. A weakness of Malmquist DEA is that it is computationally intensive and cannot handle very large datasets. Nevertheless, corporate loan portfolios are often relatively small in sample size as compared to retail credit portfolios, such as mortgages. No previous research has attempted a panel analysis of DEA efficiency in predicting the probabilit y of financial distress. This paper has bridged this gap by calculating dynamic relative efficiency scores using Malmquist DEA and incorporating them as covariates in hazard models. Our models therefore provide time-varying information about the probability of distress and use panel rather than cross- sectional estimators. We find all efficiency scores are negatively associated with the probability of financial distress. These results confirm the findings from previous literature that the more efficient a company is, the less likely it is to encounter financial difficulties. We have experimented with several types of efficiency measures that offer insights into relative performance of companies over time. This offers the possibility for lenders to understand risk drivers of a company’s financial distress and how they vary over time. Our findings imply that the highest predictive accuracy is achieved when pooling the industries together and using generic efficiency for prediction. This implies that lenders can achieve their goals of accurate predictions and interpretable results without the need for segmented models and component efficiency scores. Pooling all industries together rather than carrying out a DEA analysis for each industry separately to calculate DEA scores may be practically effective in detecting financial distress, since the best predicting model in our analysis is the Generic Efficiency one. However, the second best is Global Efficiency Malmquist DEA with only slightly less accurate predictions than the generic scores. The 25 Global efficiency takes all historic records into account and chooses the most efficient company years as the reference units. This implies that when the sample is sufficiently large, global efficiency can be seen as absolute efficiency which is generalised in all units and periods. Therefore, if one is concerned with the essential assumption of homogeneity of technology for DEA, the Global Efficiency model that keeps this assumption and at the same time produces accurate predictions is the one to choose. Although this paper only employs data from three sectors in the empirical analysis, it can obviously be extended to a large variety of industries as long as their production technologies are similar, otherwise they should be dealt with carefully. Our method provides a feasible solution to accommodate differences and commonalities in the business operations of companies. From the comparisons with cross-sectional peers and time series histories, managers of companies will be able to identify the weaknesses in their businesses and improve their performance to avoid financial difficulties. Another useful application of the current paper would be to extend the combination of Malmquist DEA and the discrete hazard model to a single sector such as financial institutions. The failure of banks was noticeable during the subprime crisis, and would have great impact on the real economy of all countries. Giving dynamic early warning of their failure by inclusion of the time dimension will be of interest in the scope of bankruptcy prediction. Finally, incorporation of DEA scores into machine - learning predictive algorithms can be another fruitful avenue of investigation. References Altman, E. I. (1968). Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. Journal of Finance, 23, 589-609. Andersen, P., & Petersen, N. C. (1993). A procedure for ranking efficient units in data envelopme nt analysis. Managemeng Science, 39, 1261-1264. Banasik, J. L., Crook, J. N., & Thomas, L. C. (1996). Does scoring a subpopulation make a difference. The International Review of Retail, Distribution and Consumer Research, 6, 180-195. Banker, R. D., & Chang, H. (2006). The super-efficiency procedure for outlier identification, not for ranking efficient units. European Journal of Operational Research, 175, 1311-1320. Berg, S. A., Førsund, F. R., & Jansen, E. S. (1992). Malmquist Indices of Productivity Growth during the Deregulation of Norwegian Banking, 1980-89. The Scandinavian Journal of Economics, 94, S211-S228. Bijak, K., & Thomas, L. C. (2012). Does segmentation always improve model performance in credit scoring? Expert Systems with Applications, 39, 2433-2442. Bonfim, D. (2009). Credit risk drivers: Evaluating the contribution of firm level information and of macroeconomic dynamics. Journal of Banking & Finance, 33, 281-299. Bryan, D., Fernando, G. D., & Tripathy, A. (2013). Bankruptcy risk, productivity and firm strategy. Review of Accounting and Finance, 12, 309-326. Cao, Y. (2012). MCELCCh-FDP: Financial distress prediction with classifier ensembles based on firm life cycle and Choquet integral. Expert Systems with Applications, 39, 7041-7049. 26 Carling, K., Jacobson, T., Linde, J., & Roszbach, K. (2007). Corporate credit risk modeling and the macroeconomy. Journal of Banking & Finance, 31, 845-868. Caves, D. W., Christensen, L. R., & Diewert, W. E. (1982). The Economic Theory of Index Numbers and the Measurement of Input, Output, and Productivity. Econometrica, 50, 1393-1414. Charnes, A., Clark, C. T., Cooper, W. W., & Golany, B. (1984). A developmental study of data envelopment analysis in measuring the efficiency of maintenance units in the U.S. air forces. Annals of Operations Research, 2, 95-112. Chava, S., & Jarrow, R. A. (2004). Bankruptcy Prediction with Industry Effects. Review of Finance, 8, 537-569. Cielen, A., Peeters, L., & Vanhoof, K. (2004). Bankruptcy prediction using a data envelopme nt analysis. European Journal of Operational Research, 154, 526-532. Cook, W. D., & Seiford, L. M. (2009). Data envelopment analysis (DEA) – Thirty years on. European Journal of Operational Research, 192, 1-17. Cooper, W. W., Seiford, L. M., & Tone, K. (2006). DATA ENVELOPMENT ANALYSIS A Comprehensive Text with Models, Applications, References and DEA-Solver Software (2nd ed.): Springer. Costa, R. (2012). Assessing Intellectual Capital efficiency and productivity: An application to the Italian yacht manufacturing sector. Expert Systems with Applications, 39, 7255-7261. Dyson, R. G., Allen, R., Camanho, A. S., Podinovski, V. V., Sarrico, C. S., & Shale, E. A. (2001). Pitfalls and protocols in DEA. European Journal of Operational Research, 132, 245-259. Emel, A. B., Oral, M., Reisman, A., & Yolalan, R. (2003). A credit scoring approach for the commercial banking sector. Socio-Economic Planning Sciences, 37, 103-123. Färe, R., Grosskopf, S., Lindgren, B., & Roos, P. (1992). Productivity changes in Swedish pharamacies 1980–1989: A non-parametric Malmquist approach. Journal of Productivity Analysis, 3, 85-101. Färe, R., Grosskopf, S., Norris, M., & Zhang, Z. (1994). Productivity Growth, Technical Progress, and Efficiency Change in Industrialized Countries. The American Economic Review, 84, 66-83. Farrell, M. J. (1957). The Measurement of Productive Efficiency. Journal of the Royal Statistical Society. Series A (General), 120, 253-290. Fedorova, E., Gilenko, E., & Dovzhenko, S. (2013). Bankruptcy prediction for Russian companies: Application of combined classifiers. Expert Systems with Applications, 40, 7285-7293. Fuentes, R., & Lillo-Bañuls, A. (2015). Smoothed bootstrap Malmquist index based on DEA model to compute productivity of tax offices. Expert Systems with Applications, 42, 2442-2450. Gordini, N. (2014). A genetic algorithm approach for SMEs bankruptcy prediction: Empirical evidence from Italy. Expert Systems with Applications, 41, 6433-6445. Hua, Z., Wang, Y., Xu, X., Zhang, B., & Liang, L. (2007). Predicting corporate financial distress based on integration of support vector machine and logistic regression. Expert Systems with Applications, 33, 434-440. Kingyens, A. T., Paradi, J. C., & Tam, F. (2016). Bankruptcy Prediction of Companies in the Retail- Apparel Industry Using Data Envelopment Analysis. In J. Aparicio, C. A. K. Lovell & J. T. Pastor (Eds.), Advances in Efficiency and Productivity (pp. 299-329). Cham: Springer Internatio na l Publishing. Lessmann, S., Baesens, B., Seow, H.-V., & Thomas, L. C. (2015). Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research. European Journal of Operational Research, 247, 124-136. Li, Z., Crook, J., & Andreeva, G. (2014). Chinese companies distress prediction: an application of data envelopment analysis. Journal of the Operational Research Society, 65, 466-479. López Iturriaga, F. J., & Sanz, I. P. (2015). Bankruptcy visualization and prediction using neural networks: A study of U.S. commercial banks. Expert Systems with Applications, 42, 2857-2869. Malmquist, S. (1953). Index numbers and indifference surfaces. Trabajos de Estatistica, 4, 209-242. 27 Marques, A. I., Garcia, V., & Sanchez, J. S. (2012). Exploring the behaviour of base classifiers in credit scoring ensembles. Expert Systems with Applications, 39, 10244-10250. Min, J. H., & Lee, Y.-C. (2008). A practical approach to credit scoring. Expert Systems with Applications, 35, 1762-1770. Nam, C. W., Kim, T. S., Park, N. J., & Lee, H. K. (2008). Bankruptcy prediction using a discrete- time duration model incorporating temporal and macroeconomic dependencies. Journal of Forecasting, 27, 493-506. Paradi, J., Asmild, M., & Simak, P. (2004). Using DEA and Worst Practice DEA in Credit Risk Evaluation. Journal of Productivity Analysis, 21, 153-165. Paradi, J. C., Wilson, D. A., & Yang, X. (2014). Data Envelopment Analysis of Corporate Failure for Non-Manufacturing Firms Using a Slacks-Based Measure. Journal of Service Science and Management, Vol.07No.04, 14. Pastor, J. T., & Lovell, C. A. K. (2005). A global Malmquist productivity index. Economics Letters, 88, 266-271. Premachandra, I. M., Bhabra, G. S., & Sueyoshi, T. (2009). DEA as a tool for bankruptcy assessment: A comparative study with logistic regression technique. European Journal of Operational Research, 193, 412-424. Premachandra, I. M., Chen, Y., & Watson, J. (2011). DEA as a tool for predicting corporate failure and success: A case of bankruptcy assessment. Omega, 39, 620-626. Psillaki, M., Tsolas, I. E., & Margaritis, D. (2010). Evaluation of credit risk based on firm performance. European Journal of Operational Research, 201, 873-881. Shetty, U., Pakkala, T. P. M., & Mallikarjunappa, T. (2012). A modified directional distance formulation of DEA to assess bankruptcy: An application to IT/ITES companies in India. Expert Systems with Applications, 39, 1988-1997. Shumway, T. (2001). Forecasting bankruptcy more accurately: A simple hazard model. Journal of Business, 74, 101-124. Shyu, J., & Chiang, T. (2012). Measuring the true managerial efficiency of bank branches in Taiwan: A three-stage DEA analysis. Expert Systems with Applications, 39, 11494-11502. Sohn, S. Y., & Kim, Y. (2012). DEA based multi-period evaluation system for research in academia. Expert Systems with Applications, 39, 8274-8278. Sun, J., Jia, M.-y., & Li, H. (2011). AdaBoost ensemble for financial distress prediction: An empirica l comparison with data from Chinese listed companies. Expert Systems with Applications, 38, 9305- 9312. Tone, K. (2002). A slacks-based measure of super-efficiency in data envelopment analysis. European Journal of Operational Research, 143, 32-41. Wanke, P., Azad, M. A. K., & Barros, C. P. (2016). Financial distress and the Malaysian dual baking system: A dynamic slacks approach. Journal of Banking & Finance, 66, 1-18. Wanke, P., Barros, C. P., & Faria, J. R. (2015). Financial distress drivers in Brazilian banks: A dynamic slacks approach. European Journal of Operational Research, 240, 258-268. Wilson, N., & Altanlar, A. (2014). Company failure prediction with limited information: newly incorporated companies. Journal of the Operational Research Society, 65, 252-264. Xu, X., & Wang, Y. (2009). Financial failure prediction using efficiency as a predictor. Expert Systems with Applications, 36, 366-373. Yang, X., & Dimitrov, S. (2017). Data envelopment analysis may obfuscate corporate financial data: using support vector machine and data envelopment analysis to predict corporate failure for nonmanufacturing firms. INFOR: Information Systems and Operational Research, 1-17. Yang, Z., You, W., & Ji, G. (2011). Using partial least squares and support vector machines for bankruptcy prediction. Expert Systems with Applications, 38, 8336-8342. Yeh, C.-C., Chi, D.-J., & Hsu, M.-F. (2010). A hybrid approach of DEA, rough set and support vector machines for business failure prediction. Expert Systems with Applications, 37, 1535-1541. 28 
work_ilj67xanejabtnq4n5brgqinhu	----	PII: 0196-8904(94)00075-B P e r g a m o n Energy Convers. Mgmt Vol. 36, No. 4, pp. 257-261, 1995 Copyright © 1995 Elsevier Science Ltd 0196-8904(94)00075-1 Printed in Great Britain. All rights reserved 0196-8904/95 $9.50 + 0.00 A N E X P E R T S Y S T E M A P P R O A C H T O T H E U N I T C O M M I T M E N T P R O B L E M D . P . K O T H A R P a n d A I J A Z A H M A D 2 1Centre for Energy Studies, Indian Institute of Technology, Delhi, Hauz Khas, New Delhi 110 016 and 2Electrical Engineering Department, Regional Engineering College, Srinagar, Kashmir, India (Received 26 March 1994; received for publication 8 December 1994) Abstract--An expert system plays a key role in the better conservation and management of electrical energy in a power station having a number of units that have to be committed for a given load. This paper presents a hybrid expert system dynamic programming approach to the unit commitment problem. Here, the scheduling output of the usual dynamic programming is enhanced by supplementing it with the rule based expert system. The proposed system limits the number of constraints and also checks the possible constraint violations in the generated schedule. The expert system communicates with the operator in a friendly manner, and hence, the various program parameters can be adjusted to have an optimal, operationally acceptable schedule. Unit commitment Dynamic programming Expert system 1. I N T R O D U C T I O N I n o r d e r to select the m o s t e c o n o m i c a l schedule o f units in a p o w e r system, different p r o g r a m a l g o r i t h m s h a v e been developed in the past. These algorithms are b a s e d o n m a t h e m a t i c a l p r o g r a m m i n g m e t h o d s , such as d y n a m i c p r o g r a m m i n g ( D P ) [ 1 , 7], L a g r a n g i a n relaxation [2], b r a n c h a n d b o u n d etc. Experience with unit c o m m i t m e n t ( U C ) using a D P technique has s h o w n that, in o r d e r to o b t a i n a g o o d schedule, o p e r a t o r experience should be included in setting u p the c o n t r o l p a r a m e t e r s . This results in tuning the heuristic d a t a a n d input p a r a m e t e r s to p r o d u c e a lower cost a n d o p e r a t i o n a l l y acceptable schedule. A h y b r i d Artificial N e u r a l N e t w o r k ( A N N ) - D P a p p r o a c h [3], h a s also been used. This p a p e r p r o p o s e s a h y b r i d expert system (ES) d y n a m i c p r o g r a m m i n g a p p r o a c h to unit c o m m i t m e n t . H e r e , the scheduling o u t p u t o f the usual D P is enhanced b y s u p p l e m e n t i n g it with the rule b a s e d ES. This ES c o m b i n e s the knowledge o f the U C p r o g r a m m e r a n d a set o f rules. T h e m a i n feature o f the ES, thus developed using T u r b o P r o l o g V2.0, includes m a k i n g possible the use o f the p r o g r a m b y a n o n - e x p e r t user t h r o u g h its ability to guide the user t h r o u g h the i m p o r t a n t decision m a k i n g processes in the p r o b l e m set u p period. T h e system guides the user in adjusting the m a r g i n a l constraints, thereby ensuring a feasible solution within e c o n o m i c a l time limits. T h e m e t h o d h a s been applied to a 10 unit system, wherein the schedule has been o b t a i n e d b y the ES to ensure a feasible solution. I t provides a flexible p r o g r a m m e structure t h a t is easily a d a p t a b l e to changes in the system o p e r a t i o n . 2. E X P E R T S Y S T E M D E V E L O P M E N T I n the past, the ES a p p r o a c h for U C has been applied b y M o k h t a r i e t al. [5]. H e r e , they h a v e developed a n ES to analyze a n d i m p r o v e the D P b a s e d U C . T h e ES tool selected is b a s e d on the inference engine o f the well k n o w n M y c i n ES a n d has been developed for the c o m b u s t i o n turbine cycling p r o b l e m . O u a n g a n d S h a h i d e p o u r [4] h a v e p r o p o s e d a U C ES which applies the E A S E + N E X P E R T shell to o b t a i n a c o m m i t m e n t schedule b y using heuristic rules. A user friendly interface a n d graphic utilities h a v e also been set up. ECM 36/*--D 257 258 KOTHARI and AHMAD: THE UNIT COMMITMENT PROBLEM In the work presented here, a rule based ES approach [6] for handling the UC problem is based on the application of an ES to supplement the DP technique so that a more appropriate solution is achieved. 2.1. Main features of the ES development The main features of the ES developed here are: (i) It combines the knowledge of the UC programmer with the rule base and, hence, can lead an inexperienced operator to a better unit schedule. (ii) The knowledge base (KB)[9, 10], which is the data or knowledge used to make decisions, consists of two parts, i.e. the working memory and the rule base. The KB, containing the knowledge of both schedulers and mathematical programmers, is represented by if/then rule statements. (iii) Each rule represents a piece of knowledge relevant to the problem and the rules can be added, removed and modified in the knowledge base conveniently. (iv) Production rules are close to natural language and, therefore, easy to understand. This enables the user to communicate with the ES in a friendly manner in order to adjust various programme parameters, so that an optimal and acceptable solution is approached. (v) It can give the steps that lead to the conclusions and user's queries about the main phase of the problem. The user can confirm or correct the conclusion by examining the explanations given by the inference engine. The ES developed functions in the following main capacities: (i) Preprocessor of the DP results. (ii) KB and reasoning. (iii) Postprocessor of the result. (iv) User consultant. 2.2. Preprocessor of the DP results There are many constraints in the UC problem which are too difficult to include in the DP algorithm because it increases the programm execution time exponentially. In order to adjust the input data properly, the operator must have a thorough knowledge of the UC programme. The constraints, like operational and complex, are accommodated in the ES as described below. 2.2.1. Inclusion of operational constraints. The operational constraints are included by the user who is guided by the preprocessor, and this is accomplished through an interactive question and answer session during which the user will enter data and answer questions concerning the constraints, such as spinning reserve requirements, unit minimum up and down time, must run status, unit maintenance schedules, unit minimum and maximum generating limits etc. The information regarding these constraints is requested by the system and has to be supplied by the user. Take the case of the constraint "must-run-status". Here, it is assumed that, during any given hour, one or more units in any group may be designated as "must run", i.e. those units for which there will never be any doubt about the need to commit them either because of outstanding efficiency or exceptional capacity. There may also be certain units which will be known, in advance, to be unavailable due to scheduled or forced outage. Such units are designated as "must out". All such information is supplied by the user. 2.2.2. Inclusion of complex constraints. The inclusion of complex constraints in the problem not only makes the problem difficult but increases the program execution time also. This is particularly true of constraints like "unit minimum up and down time", which are time dependent. Here, it is assumed that such a constraint is not violated frequently, so that it can be relaxed and handled external to the UC program itself by adjusting the input data which is fed to it by the preprocessor. Detections for constraint violations can be made by including rules in the knowledge. Also, constraints like fuel constrained units, crew constraints and pollution constraints are present in the actual system operation, and these can be generalized to form production rules that would be added to the knowledge base. Here, in the present study, although it has been tried to include as many constraints as possible, yet these constraints could not be included to avoid the complexity of the problem. KOTHARI and AHMAD: THE UNIT COMMITMENT PROBLEM 259 2.3. Knowledge base and reasoning All the knowledge in solving the U C problem is presented by either a property list or a production rule, i.e. facts and rules. The property list has a form o f object-attribute-value. F o r instance, if the m a x i m u m capacity o f unit-1 is 500 MW, this knowledge is stored in the knowledge base by a triple: (Unit-l, m a x i m u m capacity, 500). Similarly, if the value o f cost parameter A o f unit 10 is Rs. 2500, this knowledge is stored as: (Unit-10, cost A, 2500). Other facts and system constraints are also represented in the same manner. The value portion o f the triple can be update if required. A n o t h e r form o f knowledge is a production rule. A rule expresses the relationship between facts and has a form o f (if clause then clause). Such rule based representation allows the expert system to approach a problem in a way similar to a h u m a n expert. An example o f knowledge representation is given below: I f the unit start up cost is low, A n d the efficiency is high, A n d the unit cost parameter values are low, Then the unit has 'must-run-status'. Such a rule is represented in Turbo Prolog as follows: Hypothesis (must-run-status), condition (start-up-cost-low), condition (efficiency high), condition (cost-parameter-values-low). The inference engine examines the first predicate o f the rule. If the start up cost is low, the engine goes to check the second predicate, if the efficiency is high, then it goes to check the third predicate, and it the cost parameter values are low, then the engine concludes that the unit has must-run-status. The facts for all units, like m i n i m u m and m a x i m u m generating limits, cost curve parameters, start up times etc. and the load curve for the period are stored in the d a t a base. This corresponds to the prolog's static database. The preceding rule a b o u t must-run-status, for example, is a part o f the rule base/static database. 2.3.1. Production rules in the knowledge base. In order to have a fast solution, a production rule based and priority list based heuristics are implemented. The rules, here used, use the priority list as the starting point with additional heuristics leading to the optimal commitment decisions. The decisions in the ES developed here are made according to the following set o f rules: Rule 1: I f a particular unit which is on the top o f the priority list o f available uncommitted units but has a higher m i n i m u m up-time (say, 4 h) than o f the other units (say 1 h) and the spinning reserve requirement is insufficient only for say 1 h, then it is more economical to commit the second unit for 1 h rather than the first unit for at least the m i n i m u m up-time o f 4 h. Rule 2: If, at a n y stage, there is an outage associated with the normally committed units, then commit the next most economical unit. Rule 3: Assign the value o f m a x i m u m tolerable insecurity level (MTIL) to be 0.000445. If, due to some forced outage, this value o f M T I L is not satisfied, then commit the units to the next nearest higher value o f M T I L . Rule 4: I f the load is much less t h a n expected, then the " m u s t r u n " unit (if suggested by the user) should be one o f the committed units in such a case. Rule 5: I f a unit is committed, decommitted and committed again and the " o f f " status in between is close or equal to the m i n i m u m down time o f the unit, commit the unit continuously at the required output. Rule 6: At any instant, if some committed units o f smaller capacity can be replaced in an " o f f " state, then do such replacement, preserving the spinning reserve constraint. 260 KOTHARI and AHMAD: THE UNIT COMMITMENT PROBLEM Rule 7: All the c o m m i t t e d units must operate at least at their m i n i m u m o u t p u t conditions. Rule 8: I f the load d e m a n d is not met at any instant, first let the m o st efficient unit generate m o r e power until either its m a x i m u m o u t p u t is reached o r the load d e m a n d is fulfilled. Rule 9" At any stage, if the sum o f the m a x i m u m generating capacity o f all c o m m i t t e d units is greater t h a n the load demand, then d e c o m m i t only that unit which has the lowest efficiency. 2.4. Postprocessor T h e function o f the postprocessor is to guide the user in analyzing the schedule an d also m a k e suggestions to the user a b o u t adjustments to p r o g r a m p aram et ers to i m p ro v e the overall schedule cost. T h e parameters adjusted here include m i n i m u m up an d d o w n time, unit initial conditions a n d priority ordering. A n o t h e r function o f the ES postprocessor is to detect constraint violations, which is d o n e by including rules in the knowledge base. These rules can deduce logically a violation condition f r o m the schedule o u t p u t and, subsequently, advise the user an d r e c o m m e n d corrections for the problem. We have seen that m i n i m u m u p and d o w n times are particularly difficult to model and c a n n o t be i n c o r p o r a t e d directly in a D P routine. So, to detect such violations, a rule can be introduced in the knowledge base. F o r example, one o f such rules is: Check constraint (up-time-violated): condition (unit on), condition (unit off), condition (dur-less-than-1 h). This rule states that, if the unit is committed, then d e c o m m i t t e d an d the d u r a t i o n between o n an d off is less than 1 h, then the unit up time is violated. T h e condition " u p - t i m e - v i o l a t e d " will n o t be a fact in the database. T h e r e are a n u m b e r o f such rules which have been used in the p r o g r a m m e for checking constraint violations. 2.5. User interface A natural language permits the user to c o m m u n i c a t e with the ES during the consultation process using a h u m a n language. T h e main feature o f the user interface, here, is t h at it works as an e x p l a n a t o r y system which permits the user to answer a question with why. T h e system then responds with any i n f o r m a t i o n regarding the c o m m i t m e n t o f a particular unit at a particular time and so helps the user to u n d e r s t a n d the reasoning process. A n y question raised by the o p e r a t o r concerning the result is answered by the ES, an d i f the results seemed to be unsatisfactory, the K B was modified and the entire process r e p e a t e d until satisfactory g o o d results were obtained. 3. S A M P L E S Y S T E M The rule based ES developed here is applied to analyze an d improve the D P results o f the sample system o f Ref. [8]. F r o m various consultation sessions with the ES, it was seen t h at the system is quite helpful in improving the results. A n u m b e r o f U C runs were m a d e to test the capability and friendliness o f the ES. T h e p r o g r a m has been ru n o n an I B M - P C - A T . Th e consultation sessions provide almost all i n f o r m a t i o n a b o u t the various U C schedules given at different values o f M T I L . T h e system also gives an explanation a b o u t an y decisions t ak en by it. It was seen that, in o rd er to increase the effectiveness o f the system, a considerable database is required. As the rules included here in the K B have been obtained mainly f r o m a t h o r o u g h survey o f the literature an d are based o n heuristics, therefore, it is expected that, if the rules o b t ai n ed f r o m field experts are included here, the system will obviously pave the way for handling the practical U C problem. The sample system consists o f 10 thermal generating units with different generating capacities and start up times. T h e time span is 24 h. It was pointed out by the ES n o t to put unit 2 o ff fo r 1 h an d then restart it again. It suggested that the unit must be kept on line f r o m the f o u r t h to the sixth hour. F u r t h e r , it suggested t h at the value o f M T I L at which the earlier schedule was c o m p u t e d by D P can be increased so t h at the overall schedule will be m o r e economical. T h e final U C schedule is depicted in Table 1. KOTHARI and AHMAD: THE UNIT COMMITMENT PROBLEM 261 Table 1 Load Unit status Hour (MW) l 2 3 4 5 6 7 8 9 10 1 2000 0 1 0 2 1980 0 1 0 3 1940 0 1 0 4 1900 0 1 1 5 1840 0 0 0 6 1870 0 1 0 7 1820 0 0 0 8 1700 0 0 0 9 1510 0 0 0 I0 1410 0 0 0 11 1320 0 0 0 12 1260 0 0 0 13 1200 0 0 0 14 1160 0 0 0 15 1140 0 0 0 16 1160 0 0 0 17 1260 0 0 0 18 1380 0 0 1 19 1560 0 0 1 20 1700 0 0 1 21 1820 0 0 1 22 1900 0 1 1 23 1950 0 I 1 24 1990 0 1 1 I 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 I 1 1 1 1 1 1 1 I 0 1 0 0 1 0 0 1 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 1 0 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 When one compares the schedule obtained here to that obtained purely by D P [8], there is not much variation except the points discussed above, but the main feature is the capability o f the ES to approach the operator in a friendly manner. A better solution may be achieved if the knowledge o f a good number of field experts is included in the database. 4 . C O N C L U S I O N An expert system and dynamic programming hybrid has been presented for solving the unit commitment problem. The dynamic programming solution o f a sample system of 10 units was supplemented by the expert system, resulting in better management o f the units and, hence, better energy conservation. R E F E R E N C E S 1. P. G. Lowery, I E E E Trans. Power Apparatus Systems PAS-85, (1966). 2. Slobodan Ruzic and Nikola Rajakovic, I E E E Trans. Power Systems 6, No. 1 (1991). 3. Z. Ouang and S. M. Shahidepour, 1EEE Trans. Power Systems 6, No. 3 (1991). 4. Z. Ouang and S. M. Shahidepour, Electric Power System Res. 20, No. 3 (1990). 5. Sasan Mokhtari, Jagjit Singh and Bruce Wollenberg, I E E E Trans. Power Systems 3, No. 1 (1988). 6. S. K. Tong et al., I E E E Trans. Power Systems 6, No. 3 (1991). 7. C. K. Pang and H. C. Chen, I E E E Trans. Power Systems PAS-95, No. 4 (1976). 8. A. K. Ayub and A. D. Patton, I E E E Trans. Power Systems PAS-90, (1971). 9. T. Sakaguchi et al., I E E E Trans. Power Systems PAS-102, No. 2 (1983). 10. Tim Taylor et al., I E E E Trans. Power Systems 4, No. 1 (1989). 11. I. J. Nagrath and D. P. Kothari, Power System Engineering. McGraw-Hill, New Delhi (1994). 
work_im4f27vpcra25d6a3vezniatde	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 221023627 Params is empty 221023627 exception Params is empty 2021/04/06-03:06:14 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221023627 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:06:14 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_imdlzyqdajawjkevtvg6k3pdui	----	PII: 0888-613X(94)90009-4 Approximations Between Fuzzy Expert Systems and Neural Networks Yoichi Hayashi l b a r a k i University, Hitachi-shi, Ibaraki 316, J a p a n James J. Buckley University o f A l a b a m a at B i r m i n g h a m , A l a b a m a A B S T R A C T The fuzzy expert system we are concerned about in this paper is a rule-based fuzzy expert system using any method o f approximate reasoning to evaluate the rules when given new data. In this paper we argue that: (1) any continuous fuzzy expert system may be approximated by a neural net; and (2) any continuous neural net (feedforward, multilayered) may be approximated by a fuzzy expert system. We show how to train the neural net and how to write down the rules in the fuzzy expert system. K E Y W O R D S : N e u r a l networks, f u z z y expert systems, approximations 1. I N T R O D U C T I O N T h e fuzzy e x p e r t s y s t e m w e a r e c o n c e r n e d a b o u t in this p a p e r is a c o n t i n u o u s r u l e - b a s e d fuzzy e x p e r t s y s t e m using any m e t h o d o f approxi- m a t e r e a s o n i n g to e v a l u a t e t h e rules w h e n given new data. T h e r e will b e n o u n c e r t a i n t i e s in t h e d a t a o r t h e rules, a n d t h e r e is n o t h r e s h o l d i n g in the firing o f a rule. T h a t is, all rules fire given new data. T h e n e u r a l nets will b e c o n t i n u o u s ( n o t h r e s h o l d i n g within n e u r o n s ) f e e d f o r w a r d , multilay- ered, e m p l o y i n g any l e a r n i n g a l g o r i t h m [1]. I n the next section we b e g i n b y showing t h a t a t h r e e - l a y e r e d n e u r a l net c a n b e t r a i n e d t o a p p r o x i m a t e a given discrete, continuous, fuzzy e x p e r t system uniformly, to any d e g r e e o f accuracy, o v e r all inputs. T h e m o t i v a - Address correspondence to Dr. Yoichi Hayashi, Department of Computer & Information Sciences, lbaraki University, Hitachi-shi, Ibaraki 316, Japan. A Preliminary version of this paper was presented at the Second International Conference on Fuzzy Logic and Neural Networks (IIZUKA'92), Iizuka, Japan, July 17-22, 1992. Received November 17, 1992; accepted August 15, 1993. International Journal of Approximate Reasoning 1994; 10:63-73 © 1994 Elsevier Science Inc. 655 Avenue of the Americas, New York, NY 10010 0888-613X/94/$7.00 63 64 Yoichi Hayashi and James J. Buckley tion for this part of our research is to build fast parallel computation (a neural net) for discrete fuzzy expert systems. In [2] we discussed approxi- mating fuzzy expert systems by neural nets, but the discussion was based on an existence proof. That is, used results that state continuous function can be approximated uniformly on compact sets by neural networks. In this study we show how to train a three-layered neural net, having a sufficient number of neurons in the hidden layer, to approximate a given continuous, discrete, fuzzy expert system. In [3] the authors also train neural nets to approximate fuzzy rules in a fuzzy expert system. However, they train a neural net to approximate one or two explicit fuzzy rules. Our results are more general in that our continuous, discrete, fuzzy expert system has any number of rules and uses any method of approximate reasoning to infer its final conclusion. In the second part of the next section we explain how to construct a continuous, discrete, fuzzy expert system to approximate a given neural net uniformly, to any degree of accuracy, over all inputs. We are not suggest- ing replacing neural nets by fuzzy expert systems. This is an important theoretical result showing that fuzzy expert systems can be computationally equivalent to neural nets. In [2] we also discussed this approximation result, but there we presented more of an existence proof. We showed that given the neural net, we can build a discrete fuzzy expert system, in particular we showed that there exists a method of approximate reasoning, to approximate the net. In this paper we present a more general discussion on the construction of the discrete fuzzy expert system. We will use a bar over a symbol to represent a fuzzy set. So A, B, C . . . . are all fuzzy sets. Also, all our fuzzy sets will be subsets of the real numbers. If , 4 is a fuzzy set, then ,4(x) denotes its membership function evaluated at real number x. Similarly, B(x), C(x) are the membership functions for B, C, respectively. 2. APPROXIMATIONS We will first discuss how to approximate a fuzzy expert system with a neural net and then use a fuzzy expert system to approximate a neural net. 2.1. Fuzzy Expert System Consider a fuzzy expert system (abbreviated FES) having one block of rules d r : If X = . ~ and Y = Bj, then Z = Cp, (1) Fuzzy Expert Systems and Neural Networks 65 f o r 1 < r < n. F o r simplicity we have a s s u m e d t h at each rule has only two simple clauses in its a n t e c e d e n t . Also, we are using fuzzy sets directly in the rules instead o f having t h e i r equivalent linguistic values. In e q u a t i o n (1) we c o n n e c t e d the two clauses with an " a n d , " b u t o n e can use " a n d " o r " o r " in the rules. W e will assume that: (1) the interval [al, b 1] contains t h e s u p p o r t o f all the fuzzy sets t h a n can b e used fo r X ; (2) [a 2, b2] contains all the fuzzy sets that m a y b e used f o r Y; and (3) [a 3, b 3 ] has t h e s u p p o r t o f all the fuzzy sets that can b e identified with variable Z. T h e expert system will use s o m e m e t h o d o f a p p r o x i m a t e reaso n i n g to evaluate the rules w h e n given d a t a X = . ~ an d Y = B ' . O n e m e t h o d o f a p p r o x i m a t e reasoning involves: (1) choosing animplica_tion o p e r a t o r ; (2) picking a m e t h o d o f c o m p o s i n g the d a t a X = A ' , Y = B ' , with t h e infor- m a t i o n in a rule; and (3) deciding on how to c o m b i n e t h e results o f each rule into a final conclusion [4]. T h e r e are o t h e r m e t h o d s like first combin- ing all the rules into o n e fuzzy r e l a t i o n [4], b u t we n e e d n o t exactly specify any particular p r o c e d u r e of a p p r o x i m a t e reaso n i n g in this p ap er. W e now assume some m e t h o d o f a p p r o x i m a t e r e a s o n i n g has b e e n chosen, and we will d e n o t e this m e t h o d by ~ ¢ 2 . So, given the d a t a X = A and Y = B ' , t h e block o f rules .9~r, 1 < r < n, and s¢~2, the fuzzy ex p ert system p r o d u c e s its conclusion Z = C ' . Now assume we r u n this fuzzy e x p e r t system o n so m e test d at a X = ~zTk, Y - B~,, 1 _< k _< K. L e t the c o r r e s p o n d i n g conclusions b e Z = C~,, 1 < k _< K. In a c o m p u t e r o n e usually uses discrete versions o f t h e c o n t i n u o u s fuzzy sets A i , B j , C p , Z ' , B ' , C ' , etc. So let ~ b e a discretiza- tion o f the intervals [ai, bi], 1 < i _< 3. In this p a p e r we will c h o o s e t h e following discretization: (1) pick x i in [a~, b 1 ] as x o = a 1, x i = a 1 + i ( b ~ - a l ) / N 1 1 < i <_ N 1, f o r positive i n t e g e r N1; (2) c h o o s e Yi in [a2, b2] as Y o = a 2 , Yi = a 2 + i ( b 2 - a 2 ) / N 2 , 1 <_ i <_ N 2, f o r positive i n t eg er N2; and (3) let z i be in [a3, b 3] so that z 0 = a3, z i = a 3 + i ( b 3 - a 3 ) / N 3 , 1 < i < N 3, N 3 a positive integer. ~ consists o f x 0 . . . . . xN1, Y0 . . . . , YN2, Z o , ' " , ZN~" L e t M = N1 + N2. T h e n we input the n u m b e r s a ~ k ( x i ), 0 < i < N 1, fo r X = A ~ and the n u m b e r s B ~ ( y i ) . . . . O < i < N 2 an d Y = B k , ' into t h e fuzzy expert system and obtain the n u m b e r s ff~,(z~), 0 < i < N3, fo r Z = C'~,, l < _ k < K . W e now construct, and train, a neural n e t w o r k t h at will c o m p u t e (approximately) the same results as the fuzzy ex p ert system f o r t h e inputs ~.~k(x~), 0 < i < N ~ , B ~ ( y i ) , 0 < i < N 2, f o r 1 _< k _< K. T h a t is, the n eu ral net c o m p u t e s the same as the fuzzy e x p e r t system, with resp ect to .~, fo r the test d a t a X = -~k, Y = B~,, 1 < k < K. It is well-known t h a t n e u r a l nets are universal approximators. W h a t this m e a n s is that given a c o n t i n u o u s F : R d ~ R t h e r e is a n eu ral n e t t h at can u n i f o r m l y a p p r o x i m a t e F, to any d e g r e e o f accuracy, o n c o m p a c t subsets o f R a. See [5-17] f o r a survey o f this literature. F r o m t h e discussion above we 66 Yoichi Hayashi and James J. Buckley see t h a t a discrete fuzzy e x p e r t system F E S will b e a m a p p i n g f r o m [0, l] d into [0, 1 e, f o r d = M + 2, e = N 3 + 1. W e n o w a r g u e t h a t we w o u l d n o r m a l l y e x p e r t F E S to b e a c o n t i n u o u s m a p p i n g . I n [18] it is s h o w n t h a t Z a d e h ' s c o m p o s i t i o n a l rule o f i n f e r e n c e is a c o n t i n u o u s o p e r a t i o n o n discrete fuzzy sets, w h e n it is b a s e d o n a c o n t i n u - o u s t-norm• W e w o u l d t h e r e f o r e e x p e c t t h a t t h e m e t h o d o f a p p r o x i m a t e reasoning, u s e d within the fuzzy e x p e r t system is also continuous. This t h e n implies t h a t t h e F E S is a c o n t i n u o u s m a p p i n g f r o m [0, 1] d into [0, 1] e. I t can b e shown t h a t t h e r e is a n e u r a l n e t t h a t can a p p r o x i m a t e F E S , u n i f o r m l y to a n y d e g r e e o f accuracy, o n c o m p a c t [0, 1] d. H o w e v e r , this a r g u m e n t is only a n existence a r g u m e n t a n d it d o e s n o t tell y o u h o w to c o n s t r u c t a n d t r a i n t h e n e u r a l net. So, f o r t h e rest o f this subsection we will discuss h o w to o b t a i n a n e u r a l n e t t h a t will a p p r o x i m a t e t h e given FES. T h e n e u r a l n e t will h a v e M + 2 input n e u r o n s , o n e h i d d e n layer, a n d N 3 + 1 o u t p u t neurons• T h e r e a r e d i f f e r e n t m e t h o d s o f specifying a sufficient n u m b e r o f n e u r o n s in t h e h i d d e n layer ([19-21]), a n d we a s s u m e t h a t o n e such m e t h o d h a s b e e n c h o s e n so t h a t t h e h i d d e n l a y e r h a s a sufficient n u m b e r o f n e u r o n s to l e a r n the training set. L a b e l the i n p u t n e u r o n s 11, 1 2 , . . . , IM+2, a n d the o u t p u t n e u r o n s O 1 , . . . , ON3+l.._We n o w d e s c r i b e h o w to train the n e u r a l n e t w o r k . W e input A'k(x o) to ~'7 --t --r I 1 . . . . , A k ( x N) t o [N+l, B k ( y o ) t o I N + 2 , . . . , B k ( y N ) t o IM+ 2 a n d t h e l - - l - - 2 • • ! d e s i r e d o u t p u t is Ck~ 3 - z 0) f r o m O1 . . . . . Ck(z N ) f r o m ON~+V T h a t IS, the training set has K p a i r s o f i n p u t s - o u t p u t s with {A'k(xi)[0 < i < N 1} U --t _ _ _ _ {Ck(Zi)]O < i < N 3} the o u t p u t set. { B k ( y i ) [ O < i < N 2} the input set a n d - ' _ _ F i g u r e 1 shows h o w t h e n e u r a l n e t will a p p r o x i m a t e the fuzzy e x p e r t system f o r t h e special case o f N 1 = Ne = N 3 = 10. T h e n the n e u r a l n e t a n d t h e fuzzy e x p e r t s y s t e m c o m p u t e t h e s a m e o u t p u t , with r e s p e c t to ~ , f o r the test d a t a X ~7 - , = A k , Y = B k , 1 < k < K . N o w o n e w o n d e r s h o w t h e o u t p u t s , f r o m t h e two systems, c o m p a r e if we i n p u t n e w d a t a X = A'-; a n d Y = B ' w h e r e t h e s e fuzzy sets do n o t b e l o n g to the test d a t a set. T h e result d e p e n d s o n h o w we pick__ed t h e test d a t a set. S u p p o s e first t h a t we c h o s e A~ = A i, B'~ = Bj, C~ = Cp, the fuzzy sets in t h e rules. T h a t is, we d o n o t r u n the fuzzy e x p e r t system to c o m p u t e Z = C~, b u t set the i n p u t p a i r to b e the fuzzy sets in the a n t e c e d e n t o f a rule a n d the o u t p u t fuzzy set is the fuzzy set in t h e r u l e ' s c o n s e q u e n c e . T h e test set has all t h e fuzzy sets in t h e rules a n d n o o t h e r fuzzy sets. W e t h e n train t h e n e u r a l n e t only o n t h e k n o w l e d g e b a s e (rules) o f the fuzzy e x p e r t s y s t e m a n d the n e t will k n o w n o t h i n g a b o u t ~¢.~. So, the two systems can differ c o n s i d e r a b l y f o r new d a t a X = -,~ a n d Y = B ' . H o w e v e r , if the test d a t a set h a s a n u m b e r o f p a i r s {A'k, B~,}, w h e r e t h e s e fuzzy sets do n o t = Q , t h e n b e l o n g to s o m e r u l e ' s a n t e c e d e n t , a n d use J ~ ' to c o m p u t e Z - ' t h e n e t t r a i n e d o n this i n f o r m a t i o n will i n c o r p o r a t e s o m e o f ~¢~' into its Fuzzy Expert Systems and Neural Networks xox~ ' ~ j O u t p u t s 67 x3 xlo Yo Yl Y2 Y~o Inputs : V \ Figure 1. Neural net approximating fuzzy expert system. weights. Then the two systems can produce similar results for new data X = ~ and Y = B'. So, one should pick the test data so that the input pairs (A'k, B~,) broadly cover applications of the fuzzy expert system and then train the neural net on this information. Then the net will better approximate the FES on new data. In the appendix we present a more formal argument (mathematical) on why you should not use only the rules in the FES as the training set for the neural network. 2.2. Neural Net Consider a continuous neural net (NN) with m input neurons, any number of hidden layers, and n output neurons. Assume that all the input, and output, signals are bounded between zero and one. Therefore, NN is a continuous mapping from [0, 1] m into [0, 1] n. 68 Yoichi Hayashi and James J. Buckley Recently, t h e r e have b e e n a n u m b e r o f p a p e r s showing that certain types o f fuzzy systems are universal a p p r o x i m a t o r s ([22-28]). All o f these fuzzy systems r e s e m b l e a fuzzy c o n t r o l l e r in that t h ey have singleton (crisp) inputs and defuzzified (crisp) output. W e may use these results to obtain a generalized fuzzy system (multiple outputs) that can a p p r o x i m a t e N N uniformly, to any d e g r e e o f accuracy, o v e r [0, l] m. H o w e v e r , we shall n o t p u r s u e that idea in this p a p e r but instead we will b e i n t e r e s t e d in building a fuzzy expert system to a p p r o x i m a t e NN. This F E S will have o n e block o f rules, we will use a fuzzy relation to model the implication in each rule, use Z a d e h ' s compositional rule o f i n f e r e n c e to evaluate each rule given new data, and finally c o m b i n e the o u t p u t s f r o m all the rules into o n e final conclusion. Input, and output, f r o m the F E S will b e discrete versions o f c o n t i n u o u s fuzzy sets. T h e arguments that fuzzy systems are universal a p p r o x i m a t o r s are existence proofs, they do not show y o u how to build the approximating fuzzy system. O u r m e t h o d is m o r e constructive in that we show how to construct rules and we can specify the m e t h o d o f a p p r o x i m a t e reasoning to be used to evaluate the rules. W e first c h o o s e wj, 1 < j <_ J, uniformly spread a r o u n d [0, 1] m an d let N N ( w j ) = qj, 1 < j < J. W e are using functional n o t a t i o n w h e r e wj [0, l] m is input to the neural net and N N ( w j ) is its o u t p u t , a v e c t o r qj in [0, 1] n. T h e F E S will have o n e block o f rules . 9 ~ j : I f X = A j , t h e n Z = C j , 1 < j < J . (2) T h e interval that contains all the fuzzy sets f o r X ( Z ) is [1, m]([1, n]). T h e discretization o f these intervals is: (1) x 0 = 1, x I = 2 . . . . . xm i = m; an d (2) z 0 = 1, zl = 2 , . . . , z , _ ~ = n. W e define A i an d Ci with respect to this discretization as follows: (A) ~ . i ) = the i th c o m p o n e n t o f w i, 1 <_ i < m ; and_ (2) ~ ( i ) = the i th c o m p o n e n t o f qj, 1 _< i < n. T h e fuzzy sets ~ . an d C i have no physical m e a n i n g n o r do they r e p r e s e n t linguistic variables. T h e y are defined to match the input (wj)-output (q j ) pairs f r o m N N so that the F E S will be able to uniformly a p p r o x i m a t e the neural net. Next we n e e d to specify the m e t h o d o f a p p r o x i m a t e r e a s o n i n g used t o evaluate the rules given input u in [0, 1] m on X. T h a t is, d at a on X will be A , a fuzzy subset o f [0, m l, and the fuzzy exp ert system t h e n concludes Z - - t = C , a fuzzy subset o f [0, n]. H o w e v e r , discrete versions o f these fuzzy sets will be used so that the input data will be X = u in [0, 1] m, w h e r e ~zV(i) = the i th c o m p o n e n t o f u for 1 _< i < m, and the o u t p u t f r o m t h e F E S will be Z = u in [0, 1] ~, w h e r e ~,0) = the i th c o m p o n e n t o f u fo r 1 < i _< n. In functional n o t a t i o n F E S ( ~ , ) = u. T h e main r e q u i r e m e n t o f the m e t h o d o f a p p r o x i m a t e reaso n i n g is F E S ( w j ) = qj all j, o r F E S ( w ~ ) = N N ( w ~ ) all j. W h a t this m e a n s is if Fuzzy Expert Systems and Neural Networks 69 X = z~ = A~ (discrete version _~) in rule 3 i , t h e n the final conclusion from the expert system is Z = C ' = Cj (discrete version qi), for all rules. We first must g u a r a n t e e t h a t this will h a p p e n for each rule because the rules will all fire separately, and t h e n we will combine their results to get final o u t p u t Z = C ' . W e first combine the d a t a ( ~ and ~ ) in each rule ~ i into a discrete fuzzy relation Ri on [0, m] × [0, hi. Given input d a t a X = v in [0, 1] m each rule fires producing conclusion Z = u o R j, for some composition o p e r a t o r " o ." T h e n the FES combines the results v o Rj across all rules (1 < j < J ) into its final conclusion (output) Z = u in [0, 1] n. W e n e e d to choose the Rj and " o " so that w i o Rj = qj all j. T h e r e are a n u m b e r of different choices if the fuzzy sets are normalized (at least one c o m p o n e n t of wj is equal to one, all j). However, our (discrete) fuzzy sets are not necessarily normalized since wj = 0 (all c o m p o n e n t s zero) could be a choice for a wj. But, there are m e t h o d s of approximate reasoning ([2]) with the property ~) o Rj = qj all j, for any wj in [0, 1]". T h e n we have w~ o Rj equal to qj for each rule 3~j. All that is left to do is to specify how the FES combines the results into its final output. F o r any v in [0, 1] m, the FES averages the results wj o Rj, for those wi nearest to v in [0, 1] m. T h e n given e > 0, we choose the wj, 1 _< j _< J, uniformly spread a r o u n d [0,1] m, construct the F E S as described above, and obtain I N N ( v ) - FES(v)I<e for all v in [0,1] m. T h a t is, we may build an FES to approximate a neural net, uniformly to any degree of accuracy, over all inputs from [0, 1] m. 3. SUMMARY AND C O N C L U S I O N S In this paper we first argued that given any continuous, discrete, rule- based fuzzy expert system using any m e t h o d of approximate reasoning to evaluate the rules, we can build and train a neural net to uniformly approximate, to any degree of accuracy, the fuzzy expert system. We also argued that you should not train the neural net only on the rules in the fuzzy expert system. W e showed (in the Appendix) how to train the net to uniformly approximate the fuzzy expert system. A practical application of this result is to substitute fast parallel c o m p u t a t i o n (the neural net) for a fuzzy expert system. Next we argued that given any continuous, multilayered, feedforward neural net, one can construct a fuzzy expert system to uniformly approxi- mate, to any degree of accuracy, the neural net. W e showed how to construct the rules in the fuzzy expert system a n d discussed the properties n e e d e d in the m e t h o d of approximate reasoning to obtain the desired result. M a t h e m a t i c a l details may be f o u n d in [2]. 70 Yoichi Hayashi and J a m e s J. Buckley B o t h a p p r o x i m a t i o n r e s u l t s t o g e t h e r s a y t h a t c e r t a i n n e u r a l n e t s a n d f u z z y e x p e r t s y s t e m s a r e c o m p u t a t i o n a l l y e q u i v a l e n t . A P P E N D I X I n this a p p e n d i x w e first s h o w h o w t o t r a i n t h e n e u r a l n e t s o t h a t it c a n u n i f o r m l y a p p r o x i m a t e t h e F E S o v e r all p o s s i b l e i n p u t s . T h e n w e a r g u e t h a t o n e s h o u l d n o t t r a i n t h e n e t w o r k o n l y o n t h e r u l e s in t h e f u z z y e x p e r t system• L e t v ~ [0, 1] a b e a p o s s i b l e i n p u t t o t h e F E S o r t o t h e n e u r a l n e t ( N N ) . • 7 ) ~-5 - t T h a t is, v = ( A (Xo), A ( x 1) . . . . , B (YN2))" W e h a v e a s s u m e d t h a t t h e F E S is a c o n t i n u o u s m a p p i n g f r o m [0, 1] a i n t o [0, 1] e SO it is u n i f o r m l y c o n t i n u - o u s . W h a t this m e a n s is t h a t g i v e n s > 0 t h e r e is a 8 1 > 0 s o t h a t I F E S ( v I ) - F E S ( v 2 ) I < e / 2 if Iv t - v 2 1 < 6 l , v l , v z in [0,1] a. W e a r e u s i n g t h e f u n c t i o n a l n o t a t i o n o f v i i n p u t t o F E S p r o d u c i n g F E S ( v i) as ( d i s c r e t e ) o u t p u t , a v e c t o r in [0, 1] e. L e t N N b e a c o n t i n u o u s t h r e e - l a y e r e d , f e e d f o r w a r d , n e u r a l n e t w i t h d i n p u t n e u r o n s , e o u t p u t n e u r o n s , a n d a s u f f i c i e n t n u m b e r o f n e u r o n s in t h e h i d d e n l a y e r so t h a t it c a n l e a r n a d a t a s e t o f size L . S u p p o s e t h a t u t, 1 _< l _< L , is a s e t o f v e c t o r s s p r e a d a r o u n d [0, 1] a. L e t u ) = F E S ( u l) a v e c t o r in [0, 1] e, 1 _< l _< L . W e h a v e a s s u m e d t h a t N N c a n l e a r n ( u l, u)), 1 _< l _< L , w h i c h m e a n s t h a t t h e w e i g h t s c a n b e a d j u s t e d s o t h a t N N ( u t) = u ' t , l < l < L . N N is a c o n t i n u o u s m a p p i n g f r o m [0, 1] a i n t o [0, 1] e, all signals a r e in t h e i n t e r v a l [0, 1], s o it is u n i f o r m l y c o n t i n u o u s • T h e r e f o r e , t h e r e is a 82 > 0 s o t h a t [ N N ( v ~ ) - N N ( v 2 ) [ < s / 2 if Iv l - v2[ < 32, vl, v 2 in [0,1] a. A g a i n w e a r e u s i n g f u n c t i o n a l n o t a t i o n w i t h /"i i n p u t t o t h e n e u r a l n e t a n d N N ( v i) its o u t p u t v e c t o r in [0, l ] e . D e f i n e 6 t o b e t h e m i n i m u m o f 61 a n d t~ 2 . C h o o s e wj, 1 <_ j < J, u n i f o r m l y s p r e a d a r o u n d [0, 1] d w i t h t h e p r o p e r t y t h a t g i v e n a n y v in [0, 1] d t h e r e is a wj s o t h a t I v - wjl < 8. A s s u m e t h a t J < L . 1 L e t F E S ( w i) = qj, 1 < j <_ J. T h e l e a r n i n g d a t a f o r N N will b e (wj, qj), 1 < j < J. T h a t is, t r a i n t h e N N s o t h a t N N ( w j ) = qj all j . W e n o w a r g u e t h a t this N N will u n i f o r m l y a p p r o x i m a t e t h e F E S . L e t v b e a n y v e c t o r in [0, 1] d a n d c h o o s e wj a l s o in [0, 1] a s o t h a t Iv - wjl < 8. 1We k/lOW there is an NN (sufficient number of neurons in the hidden layer) that will approximate FES, uniformly over all inputs v in [0, 1] a, to any degree of accuracy e > 0. So this NN will have L large enough. Fuzzy Expert Systems and Neural Networks 71 T h e n [ F E S ( v ) - N N ( v ) I = f F E S ( v ) - F E S ( w j ) + N N ( w j ) - N N ( v)[ < [ F E S ( v ) - F E S ( w i ) l + I N N ( w j ) - N N ( v ) l < ~ / 2 + ~ / 2 = because: (1) F E S ( w j ) = N N ( w j ) ; (2) Iv - wjl < 6j; an d (3) Iv - wj[ < 82. So, this n e u r a l n e t will c o m p u t e , within e, t h e same as the F ES , across all possible inputs. Now suppose y o u train the N N only o n the rules (k n o w l ed g e base). L e t s r ~ [0, 1] d d e n o t e the discretization o f A i and Bj in t h e a n t e c e d e n t o f rule ~'r, 1 < r _< n. Next let t r in [0, 1 ] e be the discretization o f Cp, t h e fuzzy set in the conclusion o f rule ~q~r, 1 < r < n. T r a i n t h e N N o n the set (st, t~), 1 < r < n, so t h a t N N ( s ~ ) = t~ all r. Now pick any v in [0, 1] a. W e would not expect I F E S ( v ) - N N ( v ) I to be small because: (1) t h e s r, 1 _< r < n d o e s not necessarily f o r m a u n i f o r m span o f t h e input space [0, 1]a; and (2) it may h a p p e n , d e p e n d i n g o n ~ ¢ ~ , t h a t F E S ( s r) -~ t r fo r s o m e rules. T h a t is, if u is n o t n e a r any st, t h e n we would n o t ex p ect N N ( u ) to b e close to F E S ( v ) . Also, N N m a y differ f r o m F E S e v e n o n the training set. F o r these r e a s o n s we do n o t r e c o m m e n d training the neural n e t only o n t h e rules o f the fuzzy e x p e r t system. References 1. Rumelhart, D. D., Hinton, G. E., and Williams, R. J., Learning internal representations by error propagation, in Parallel Distributed Processing: Explo- rations in the Microstructures o f Cognition, Vol. 1, (D. E. Rumelhart and J. L. McClelland, eds.), MIT Press, Cambridge, MA, 675-695, 1986. 2. Buckley, J. J., Hayashi, Y., and Czogala, E., On the equivalence of neural nets and fuzzy expert systems, Fuzzy Sets and Systems 53, 129-134, 1993. 3. Keller, J. M., and Tahani, H., Implementation of conjunctive and disjunctive fuzzy logic rules with neural networks, Int. J. Approx. Reasoning 6, 221-240, 1992. 4. Dubois, D., and Prade, H., Fuzzy sets in approximate reasoning, part I: inference and possibility distributions, Fuzzy Sets and Systems 40, 143-202, 1991. 5. Blum, E. K., and Li, L. K., Approximation theory and feedforward networks, Neural Networks 4, 511-515, 1991. 6. Cardaliaguest, P., and Euvrard, G., Approximation of a function and its derivative with a neural network, Neural Networks 5, 207-220, 1992. 72 Yoichi Hayashi and James J. Buckley 7. Cotter, N. E., and Guillerm, T. J., The CMAC and a theorem of Kolmogorov, Neural Networks 5, 221-228, 1992. 8. Cybenko, G., Approximation by superpositions of a sigmoidal function, Math. o f Control, Signals, and Systems, 2, 303-314, 1989. 9. Geva, S., and Sitte, J., A constructive method for multivariate function approxi- mation by multilayer perceptrons, I E E E Trans. on Neural Networks, 3, 621-623, 1992. 10. Hornik, K., Approximation capabilities of multilayer feedforward networks, Neural Networks 4, 251-257, 1991. 11. Hornik, K., Stinchcombe, M., and White, H., Multilayer feedforward networks are universal approximators, Neural Networks 2, 359-366, 1989. 12. Ito, Y., Approximation of functions on compact sets by finite sums of a sigmoid function without scaling, Neural Networks 4, 817-826, 1991. 13. Ito, Y., Approximation of continuous functions on R a by linear combinations of shifted rotations of a sigmoid function with and without scaling, Neural Networks 5, 105-115, 1992. 14. Kreinovich, V. Y., Arbitrary nonlinearity is sufficient to represent all functions by neural networks: a theorem, Neural Networks 4, 381-383, 1991. 15. Kfirkov~, V., Kolmogorov's theorem is relevant, Neural Computation 3, 617-622, 1991. 16. Kfirkov~, V., Kolmogorov's theorem and multilayer neural networks, Neural Networks 5, 501-506, 1992. 17. Park, J., and Sandberg, I. W., Universal approximation using radial-basis-func- tion networks, Neural Computation 3, 246-257, 1991. 18. Full6r, R., and Zimmermann, H.-J., On Zadeh's compositional rule of infer- ence, Proc. Fourth IFSA World Congress, Vol. Artificial Intelligence, Brussels, 41-44, July 7-12, 1991. 19. Huang, S.-C., and Huang, Y.-F., Bounds on the number of hidden neurons in multilayer perceptrons, I E E E Trans. on Neural Networks 2, 47-55, 1991. 20. Mirchandani, G., On hidden nodes for neural nets, I E E E Trans. on Circuits and Systems 36, 661-664, 1989. 21. Sartori, M. A., and Antsaklis, P. J., A simple method to derive bounds on the size and to train multilayer neural networks, I E E E Trans. on Neural Networks 2, 467-471, 1991. 22. Buckley, J. J., Sugeno type controllers are universal controllers, Fuzzy Sets and Systems 53, 299-304, 1993. 23. Buckley, J. J., Universal fuzzy controllers, Automatica 28, 1245-1248, 1992. 24. Buckley, J. J., Controllable processes and the fuzzy controller, Fuzzy Sets and Systems 53, 27-32, 1993. Fuzzy Expert Systems and Neural Networks 73 25. Buckley, J. J., Applicability of the fuzzy controller, in Advances in Fuzzy Systems: Applications and Theory, (P.-Z. Wang and K. F. Loe, Eds.) World Scientific Press, Singapore, in press. 26. Kosko, B., Fuzzy systems as universal approximators, Proc. o f l E E E Inter. Conf. on Fuzzy Systems, San Diego, 1153-1162, March 8-12, 1992. 27. Nguyen, H. T., and Kreinovich, V., On approximation of controls by fuzzy systems. Unpublished manuscript. 28. Wang, L.-X., Fuzzy systems and universal controllers, Proc. of IEEE Inter. Conf. on Fuzzy Systems, San Diego, 1163-1170, March 8-12, 1992. 
work_iokziw56jfbmrkdzls4i3arm2e	----	An Expert System for Supporting Traditional Chinese Medicine Diagnosis and Treatment Procedia Technology 16 ( 2014 ) 1487 – 1492 Available online at www.sciencedirect.com ScienceDirect 2212-0173 © 2014 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the Organizing Committee of CENTERIS 2014. doi: 10.1016/j.protcy.2014.10.169 CENTERIS 2014 - Conference on ENTERprise Information Systems / ProjMAN 2014 - International Conference on Project MANagement / HCIST 2014 - International Conference on Health and Social Care Information Systems and Technologies An expert system for supporting Traditional Chinese Medicine diagnosis and treatment Paulo Silvaa,*, Pedro Gagob,c, José Carlos Bregieiro Ribeirob, Manuel F. Santosc, Filipe Portelac, António Abelhac, José Machadoc, Filipe Pintoa,c a – ESTG, Polytechnic Institute of Leiria, Leiria, Portugal b – ESTG, CIIC, Polytechnic Institute of Leiria, Leiria, Portugal c – Algoritmi Research Center. University of Minho. Guimarães. Portugal Abstract Portugal has recently become one the few European countries to fully acknowledge Traditional Chinese Medicine (TCM); this substantial paradigm shift calls for novel tools for TCM practitioners, students and patients alike. This paper describes an Expert System for supporting the TCM consultation process – both in terms of gathering and managing the patients’ personal and symptomatic data, and of obtaining accurate diagnoses and treatments under regulated and reviewed protocols. The proposed tool was designed and is being developed with the support of two TCM therapists, which act as experts and provide aid to the processes of building the knowledge base and the automatic diagnosis system. In terms of architecture, the current version of the framework includes a mobile client application for the Android platform, integrated with an online spreadsheet. A survey was conducted in order to assess some of the needs of the community of TCM practitioners, and allowed gathering information on their needs. © 2014 The Authors. Published by Elsevier Ltd. Peer-review under responsibility of the Organizing Committees of CENTERIS/ProjMAN/HCIST 2014 Keywords: traditional chinese medicine, expert system, mobile health. * Corresponding author. Tel.: +0-000-000-0000; fax: +0-000-000-0000. E-mail address: author@institute.xxx © 2014 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the Organizing Committee of CENTERIS 2014. 1488 Paulo Silva et al. / Procedia Technology 16 ( 2014 ) 1487 – 1492 1. Introduction Traditional Chinese Medicine (TCM) is the name commonly given to the set of practices of traditional medicine in use in China. Those practices were developed over thousands of years and are considered one of the oldest forms of Oriental medicine. TCM encompasses seven main methods of treatment: Tui-na; Acupuncture; Moxibustion; Cupping; Phytotherapy; Chinese food therapy; Physical practices. The main goal of this research is the creation of an expert system capable of helping therapists reach a diagnosis and establish a treatment plan. In order do to so, the expert system must guide the therapist in the data entry process by presenting the most relevant questions in an adequate sequence. Moreover, the expert system must use that data to find the most adequate diagnosis hypothesis and present them to the therapist. In addition, we are using a Google Spreadsheets to store all relevant information, including the knowledge necessary to reach a diagnosis and suggest a treatment plan. This option eases the creation of both desktop and mobile applications while also simplifying the fine tuning of the system. All TCM diagnosis and acupuncture treatments included in the expert system were collected from experts and are supported by B. Auteroche e P. Navailh book [1]. This paper focuses on presenting the TCM-SIS (Traditional Chinese Medicine – Support Information System) framework, while providing the background underlying this research and proving the pertinence of this study; also, we intend to introduce the technical and architectural approaches proposed for implementing the framework. 1.1. Traditional Chinese Medicine Traditional Chinese Medicine cannot be understood without a basic understanding of holistic medicine. Holistic medicine views the body as a whole, and tends to believe that “imbalances” in the body manifest as symptoms. Holistic medicine believes that, in the long term, it is not efficient to suppress symptoms (e.g., by means of aspirin for headaches), as these symptoms will keep returning until the body is restored into balance. Traditional Chinese medicine is based on the theory of holistic medicine, and this can be seen through remedies such as herbal medicine, acupuncture, massage (Tui na), exercise (qigong), and dietary therapy. TCM diagnosis aims to trace symptoms to patterns of an underlying disharmony, by measuring the pulse, inspecting the tongue, skin, and eyes, and examining the eating and sleeping habits of the patient among many other things [1]. The theoretical framework of TCM has a number of key components. The Yin-yang theory is the concept of two opposing – yet complementary – forces that shape the world and all life, and is central to TCM. In the TCM view, a vital energy or life force called qi circulates in the body through a system of pathways called meridians. Health is an ongoing process of maintaining balance and harmony in the circulation of qi. The TCM approach uses eight principles to analyse symptoms and categorize conditions: cold/heat, interior/exterior, excess/deficiency, and yin/yang (the chief principles). TCM also uses the theory of five elements – fire, earth, metal, water, and wood – to explain how the body works; these elements correspond to particular organs and tissues in the body [2,3]. 1.2. The Portuguese Reality The law bill n. º 71/2013 published on September 2013 [13] is a milestone of tbe Portuguese TCM history, as Portugal will be one of the few European countries to fully acknowledge TCM and its practitioners. This paradigm shift increases the relevance of software solutions as the one presented in this paper. With the purpose of getting to know some of the TCM therapists’ needs as related to the information technologies in their professional settings we conducted a survey during a 3-day TCM post-graduation session that had 220 therapists in attendance. The participation in the survey was completely voluntary and of those 220 therapists we had 55 answering the survey. The first question asked in the survey was “What would you value the most in a computer application tool developed with the purpose of assisting you in your professional TCM practice?”. The answers are detailed in Table 1. The second question asked in the survey was “Do you presently use any kind of information system to support your therapeutic practice”; we have obtained 34 negative answers and 21 positive answers. 1489 Paulo Silva et al. / Procedia Technology 16 ( 2014 ) 1487 – 1492 Table 1 – Question 1 What would you value the most in a computer application tool developed with the purpose of assisting you in your professional TCM practice?”. Least valued Most valued No answer 1 2 3 4 5 Total Automatic diagnostic support 1 8 5 14 9 18 55 Patient Clinical record 1 2 0 5 12 35 55 Remote access to clinical information 4 3 2 9 12 25 55 Information data security 1 0 1 2 8 43 55 Quick access to external entity information 4 2 1 9 17 22 55 Prescription 5 2 3 12 10 23 55 Human anatomy aid on acupuncture points 1 2 5 11 15 21 55 Billing 3 5 11 11 12 13 55 Total 20 24 28 73 95 200 The answers to these two simple questions reinforce the well known facts: that in all health related information systems, the data security is of paramount importance and that professionals in this area are eager to have Electronic Health Records available to them. Other than that, they show a degree of distrust on “Automatic diagnostic support” tools that is to be expected from people that still do not use any form of information system in their therapeutic practice (only 21 in 55 respondents use some kind of information system). Also, the survey contained one open-ended question: “Please specify the Information Systems you use”. Of the 21 therapists that use some kind of information system in their therapeutic practice, 17 stated that the information system they use is a simple word processor and/or a spreadsheet that they use to register their patients’ clinical records. 2. Methodology and Technical Approach This work focuses on developing an Expert System for assisting TCM therapists with the process of efficiently and assertively diagnosing and treating patients. The idea of having automated decision support tools supporting clinicians is very compelling and several such systems have been created. Some of the best known are MYCIN [6], CASNET [7] and CADUCEUS [8], just to name a few. Figure 1 - Architecture of the proposed framework. 1490 Paulo Silva et al. / Procedia Technology 16 ( 2014 ) 1487 – 1492 The Expert System’s knowledge base is being built with the aid of two TCM therapists; their expertise is the main support for the processes of constructing and validating the framework. The TCM experts will be utilizing the B. Auteroche e P. Navailh reference book [1] as the preferred source for clarifying any doubts or acquiring specific information. The knowledge base containing symptoms and diagnosis information, as defined by the experts and required by the Expert System’s algorithm to produce an outcome, will be stored in “the cloud”, in an online spreadsheet (namely, a Google Spreadsheet), which is accessible to the client application via a web API (Application Programming Interface) provided by the Google Spreadsheets platform. A mobile application for the Android platform was implemented for supporting the consultation process; this will allow the therapist to gather and upload patient data during the consultation itself, allowing for a less disruptive consultation process. The architecture proposed for the TCM-SIS platform is shown in Figure 1. Commonly, the TCM diagnosis process is performed as the result of a clinical consultation and examination, in which the TCM therapist examines the patient and indicates the most suitable treatment. During this consultation, various questions are posed (e.g., “do you have heart problems?”, “do you have hearing problems?”); the TCM-SIS client application provides support for selecting a valid choice among those defined by the experts. All the observations performed by the therapist can be added in the same way, resulting in a set of symptoms which will be the basis for the Expert System’s diagnosis process. The set of symptoms will then be uploaded to the remote online spreadsheet – i.e., to the Google Spreadsheet – at the client application user’s command which will, in turn, return the diagnosis yielded by the implemented Expert System and the suggested treatment. Utilizing Google Spreadsheets for storing and managing both symptomatic and diagnosis data brings several advantages. It should be noted that building and fine tuning an Expert System is an iterative process, which requires a significant amount of inputs from the experts in order to achieve a quality system; the fact that the knowledge base is stored in an online spreadsheet, accessible via a web API, allows for continuous adaptations and improvements with basis on the therapists’ feedback. Moreover, it is possible for the therapists to fine-tune the knowledge base themselves – in order to improve its outcome or to adapt it to their specific needs, sensibilities or experience – given that the online spreadsheet is manageable by any user with basic computer skills. Also, and given that the online spreadsheets used are part of Google’s application ecosystem, the proposed framework will benefit from all the conditions offered by this enterprise in terms of security and reliability. 2.1. The Expert System The Expert System works with basis on three sources of information stored in the online spreadsheet, which are accessible to the client programs (namely, to the mobile client application) via its web API, and allow managing patient, symptoms, diagnosis and treatment data: ● Patient Data – contains patient information (e.g., name, height, weight, birth date, address, etc.). ● Validation Data – contains the options available for the client application’s user (i.e., the therapist) to annotate the patient’s symptoms. Each column represents a symptom, and for each symptom a series of possible values are defined in accordance to the specifications provided by the experts. The client application’s interface is generated dynamically with basis on the parameters defined in this spreadsheet, so as to provide the therapist with the current symptoms set and the array of possible values for each symptom [Figure 2]. ● Diagnosis and treatment data – contains the diagnosis information, as defined by the experts, and encompasses the set of possible outcomes of the Expert System’s diagnosis process; the goal of the Expert Systems will thus be that of returning a list of suggested diagnoses, ordered by accuracy, as the result of the application of the Expert System’s algorithm to the symptoms provided as input. It also contains the set of acupuncture points associated with treating the pathology diagnosed by the Expert Systems. The Graphical User Interface of the mobile application is intended be simple, efficient, and follow the Springboard model [5], which is characterized by having a menu page which includes the application’s main functionalities; also, as was above mentioned, the symptoms’ input view is extremely customizable, as it is dynamically generated from the data in the Validation Data spreadsheet. Figure 2 showcases the information flow in the TCM-SIS framework. 1491 Paulo Silva et al. / Procedia Technology 16 ( 2014 ) 1487 – 1492 Figure 2 - Information Flow in the TCM-SIS framework. Step 1 depicts an example “Validation Data” spreadsheet, where a set of symptoms and a list of possible values for each symptom. Step 1 shows a view of the Android Client application in which a list of symptoms is filled automatically with basis on the data defined in the “Validation Data” spreadsheet. Step 3 shows the set of possible values for the “type of pain” symptom. Step 4 exemplifies the “Diagnosis and treatment data spreadsheet” filled out with sample patient information. 3. Related Work Existing frameworks in the TCM fields are mainly of educational nature and/or generic Electronic Health Records management systems; examples include ShenProfessional [9], AcuBase [10], AcuPartner [11]; moreover, they are commonly commercial tools, which are not publicly available for researchers to experiment with. The Chinese Acupuncture Expert System (CAES) [4] and the ACE - Acupuncture Expert [12] are, to the best of our knowledge, the tools that most closely resemble the proposed TCM-SIS framework. CAES is a very comprehensive information system, encompassing several modules (e.g., patient record, needle insertion position animation, acupuncture points usage statistic, etc.) in addition to the diagnosis and acupuncture prescription subsystems; ACE is a commercial information system, which focuses on automating the diagnosis process using an Expert System. The knowledge base is, in both cases, built with the aid of experts in the field. The main differences between these and TCM-SIS reside in the availability of a simple and effective means for therapists to manage the knowledge base (the online spreadsheet), the mobile nature of the proposed client application, and the inherently distributed architecture which allows its extension with additional modules and client applications. 1492 Paulo Silva et al. / Procedia Technology 16 ( 2014 ) 1487 – 1492 4. Conclusions and Future Work Portugal has recently become one the few European countries to fully acknowledge Traditional Chinese Medicine; this substantial paradigm shift calls for novel tools for TCM practitioners, students and patients alike. This paper describes a framework for supporting the TCM consultation process – both in terms of gathering and managing the patients’ personal and symptomatic data, and of obtaining accurate diagnoses and treatments under regulated and reviewed protocols. The proposed tool was designed and is being developed with the support of two TCM therapists, which act as experts and provide aid to the processes of building the symptoms knowledge base and the automatic diagnosis system. In terms of architecture, the current version of the framework includes a mobile client application for the Android platform, integrated with an online spreadsheet. A Google Spreadsheet is utilized as: a knowledge base of symptomatic diagnosis and treatment information; a backoffice for managing the knowledge base; a database of patient data; and a web service which allows for the client application to remotely access and modify the stored data. A survey was conducted in order to assess the needs of the community of TCM practitioners, and allowed gathering information on their needs. It also permitted reaching a better understanding of the new business perspectives set off by the recent advances in terms of legislation. The main hindrance posed to this research is related with the nonexistence of norms that regulate TCM practice (such as those that exist for conventional Medicine), which complicates the search for diagnoses and treatments associated with a specific disease; the experts must constantly resort to specific literature resources, and the Expert System must be subjected to constant validations and parameterization; this is the main reason for designing the TCM-SIS framework in a way that enables the constant update of the knowledge base, and the immediate mirroring of any changes made to the knowledge base to the client application. Future work involves defining the Expert System’s algorithm with basis on the information gathered from the experts, and integrating those advances into the final product. Also, we expect to be able to introduce additional modules into the framework (e.g. billing, fitotherapy prescription, etc.) in order to provide therapists with a all-in-one solution for their MTC needs. Acknowledgments The authors would like to thank the TCM experts Natália Gameiro and Fernando Soares for all their help and expertise and also to the private institution - Institute Van Nghi Portugal. This work has been supported by FCT – Fundação para a Ciencia e Tecnologia in the scope of the project: PEst- OE/EEI/UI0319/2014. References [1] B. Auteroche et P. Navailh. Diagnóstico de Medicina Tradicional Chinesa - São Paulo. 3rd ed. Editor Andrei, 1998. [2] Kaptchuk TJ. The Web That Has No Weaver: Understanding Chinese Medicine. 2nd ed. New York, NY: McGraw-Hill, 2000. [3] O'Brien KA, Xue CC. The theoretical framework of Chinese medicine. In: Leung PC, Xue CC. Cheng YC, editors. A Comprehensive Guide to Chinese Medicine. River Edge. NJ: World Scientific Publishing Co.; 2003. [4] Department of Computer Science and Engineering. The Chinese University of Hong Kong, Chinese Acupuncture Expert System (CAES) - A Useful Tool to Practice and Learn Medical Acupuncture. PubMed, 2011 [5] L. Lopes, Interface móvel para ambiente de urgência e emergência médica. Work Report. 2013 p. 7-9. [6] Buchanan B.G. and Shortliffe E.H. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Reading, MA: Addison-Wesley, 1984. [7] Kulikowski C.A., Weis S.M. Representation of expert knowledge for consultation: the CASNET and EXPERT projects. in: P. Szolovits, editors. Artificial Intelligence in Medicine, 1982, Westview Press, Boulder, 21-56. [8] Pople H.E., Evolution of an Expert System: from INTERNIST to CADUCEUS. In: Artificial Intelligence in Medicine, ed. by I. De Lotto and M. Stefanelli, 1985, p. 179-208. Elsevier Science Publisher, Amsterdam. [9] http://www.shenprofessional.com - accessed on 23/04/2014 [10] http://www.acup.com/acubase.html - accessed on 23/04/2014 [11] http://www.acupartner.com - accessed on 23/04/2014 [12] http://www.taomedic.com/html/acupuncture_online_store.html - accessed on 23/04/2014 [13] https://dre.pt/pdf1sdip/2013/09/16800/0543905442.pdf - accessed on 23/04/2014 [14] https://developers.google.com/google-apps/spreadsheets/ - accessed on 19/06/2014 
work_ip7jqc6emjbhvjli7jg34x32am	----	CRPIT.COM - One Stop Destination for Education & Job Skip to content Menu Home Result Recruitment Notification Admit Card Counselling Merit List Time Table BTEUP Results 2021 UP Polytechnic Diploma Semester Exam Result April 4, 2021 by CRPIT Admin BTEUP Diploma Results 2021: Board of Technical Education (BTE) Uttar Pradesh is all set to release the result of even/Odd semesters as such semester 1st, 2nd, 3rd, 4th, 5th and 6th in the second week of August 2021. The exams of even semester of BTU were conducted by the authorized officials in June 2021.  Students … Read more Categories Result RBSE 12th Rechecking Result 2021- Retotaling, Revaluation Arts, Science, Commerce Result April 3, 2021 by CRPIT Admin RBSE 12th Rechecking Result 2021 Rajasthan Senior Secondary Board Revaluation Result 2021 Arts, Science, Commerce Subjects, by Name, Roll Number, BSER Ajmer 12th Retotaling Result Subject Wise Marks at rajresults.nic.in or rajeduboard.rajasthan.gov.in RBSE 12th Rechecking Result 2021- The RSBE 12th students who submitted online applications for Arts/ Science/ Commerce Revaluation now must be looking for … Read more Categories Result Goa SSC Revaluation Result 2021- GBSHSE 10th Rechecking Result April 3, 2021 by CRPIT Admin Goa SSC Revaluation Result 2021- Check GBSHSE 10th Exam Rechecking Result by Name, Roll Number at gbshse.org, Goa 10th Board (SSC) Revaluation Result Available soon Goa Board of Secondary and Higher Secondary Education has made all the preparations to declare the Goa SSC Revaluation Result 2021. The result will be declared on the official website … Read more Categories Result TS SSC Revaluation Result 2021- Telangana 10th Class Reverification/Recounting Result April 3, 2021 by CRPIT Admin TS SSC Revaluation Result 2021 Released Check by Name, Roll Number, BSE Telangana Class 10th Reverification & Recounting Marks/Result, Telangana Secondary School Certificate (SSC) Examination Revaluation Result at www.bse.telangana.gov.in Board of Secondary Education, Telangana will release the Revaluation Result of the Class 10th (SSC) Exam.  After the announcement of the TS SSC Board Exam Result … Read more Categories Result Gulbarga University Result 2021 B.Com B.A B.Sc 2nd, 4th, 6th Sem Exam Result April 2, 2021 by CRPIT Admin Gulbarga University Result 2021: Students who have taken the examination of Even Sem UG courses conducted by Gulbarga University will be able to check their results online as the university will soon release the result of various UG and PG courses. After appearing in the Even Sem Gulbarga University exams, all students are eagerly waiting … Read more Categories Result OU Degree Result 2021 bA, B.Sc, B.com 2nd, 4th, 6th Sem Exams Results April 2, 2021 by CRPIT Admin OU Degree Result 2021 UG 2nd, 4th, 6th Sem Exam results Manabadi: Osmania University is a state university which is located in the city of Hyderabad, in the state of Telangana. OU Degree Results 2021 can be checked on the official website of the Osmania University @osmania.ac.in as well as from this page. OU Degree Result 2021 … Read more Categories Result DAVV Results 2021: BA, B.Sc, B.Com, BBA, BCA, MBA Odd/Even semester Results 2021 April 2, 2021 by CRPIT Admin DAVV Results 2021: Formerly known as Indore University, Devi Ahilya Vishwavidyalaya (DAVV) is a state public university which is located in the state of Madhya Pradesh. The students can now check their DAVV Results 2021 as the officials of the university have uploaded the results online. DAVV Results 2021 BA B.Sc B.Com Part 1, 2, … Read more Categories Result OU Revaluation Result 2021: Manabadi Check Osmania University Degree Supply UG/PG Semester Result Here April 2, 2021 by CRPIT Admin OU Degree Revaluation Result 2021: Osmania University gives its students the facility of revaluation in which those students can apply for OU Degree revaluation of the answer sheet who could not perform well in the examination. The university releases the notification for the OU Degree revaluation after releasing and publishing the result of the exam. … Read more Categories Result HBSE 12th Revaluation Result 2021- Arts, Science, Commerce Rechecking Result April 2, 2021 by CRPIT Admin HBSE 12th Revaluation Result 2021- Haryana 12th Arts, Science, Commerce Rechecking Result available soon at www.bseh.org.in, BSEH Haryana Senior Secondary Exam Rechecking Result by Name, Roll Number, Haryana 12th Board Revaluation Result 2021    Board of School Education, Haryana is ready to release the 12th Senior Secondary Revaluation Result. The result would be issued on … Read more Categories Result IIT JAM 2021- Exam Pattern, Syllabus, Application Form, Admit Card, Previous Papers, Result, Score Card April 2, 2021 by CRPIT Admin IIT JAM 2021 – Check Complete Exam Pattern, Syllabus, IIT Kharagpur JAM (Joint Admission Test for MSc) Application Form, Admit Card, Previous Years Papers, Result& Score Card at jam.iitkgp.ac.in Indian Institute of Technology (IIT Kharagpur) and Indian Institute of Science (IISc) organize the Joint Admission Test for MSc known as JAM in short form. Joint … Read more Categories News Post navigation Older posts Page1 Page2 Page3 Next → Load More Recent Posts BTEUP Results 2021 UP Polytechnic Diploma Semester Exam Result RBSE 12th Rechecking Result 2021- Retotaling, Revaluation Arts, Science, Commerce Result Goa SSC Revaluation Result 2021- GBSHSE 10th Rechecking Result TS SSC Revaluation Result 2021- Telangana 10th Class Reverification/Recounting Result Gulbarga University Result 2021 B.Com B.A B.Sc 2nd, 4th, 6th Sem Exam Result OU Degree Result 2021 bA, B.Sc, B.com 2nd, 4th, 6th Sem Exams Results DAVV Results 2021: BA, B.Sc, B.Com, BBA, BCA, MBA Odd/Even semester Results 2021 OU Revaluation Result 2021: Manabadi Check Osmania University Degree Supply UG/PG Semester Result Here Updates RRB NTPC Admit Card 2019 Sarkari Result Privacy Policy       Contact us       Disclaimer CRPIT.COM is not affiliated to any government website. Information provided here is for educational purpose only. Read more in Disclaimer. | All Right Reserved © 2021 | 
work_iqaoqole7nf33bhdqwmgmfeauq	----	Microsoft Word - MATHSYS-first.doc Fuzzy Dissimilarity Learning in Case-Based Reasoning Ning Xiong School of Innovation, Design and Engineering Mälardalen University P.O. Box 883, SE-72123 Västerås Sweden ning.xiong@mdh.se http://www.idt.mdh.se/personal/nxg01/ Abstract: - Case-based reasoning (CBR) attempts to solve new problems by using previous experiences. However traditional CBR systems are restricted by the similarity requirement, i.e., the availability of similar cases to new problems. This paper proposes a novel CBR approach that exploits dissimilarity information in problem solving. A fuzzy dissimilarity model consisting of fuzzy rules has been developed for assessing dissimilarity between cases. Further, it is indicated that the construction of fuzzy dissimilarity rules can be realized by learning from the case library. Empirical studies have demonstrated that fuzzy dissimilarity models can be built upon a small case library while still yielding competent performance of the CBR system. Key-Words: - case-based reasoning; similarity; dissimilarity; case library; fuzzy rles; fuzzy reasoning 1 Introduction Case-based reasoning (CBR) presents an important cognitive methodology in Artificial Intelligence, which aims to use previous experiences to solve new problems [1]. A fundamental principle for conventional CBR methods is the hypothesis that similar problems have similar solutions. Hence a CBR system usually first retrieves cases in the case base that are similar to a query problem and then refines the solutions of the retrieved cases to tackle the new situation at hand. However, the concept of similarity remains an unclear issue. The meaning of “similarity” may be subjective and vary from situation to situation. The metric of similarity is intrinsically defined on the space of problems yet there is no natural link between the problem space and the solution space to warrant that the most similar case is also the most relevant to the target problem. This implies that the basic CBR hypothesis that similar problems have similar solutions does not hold in all circumstances. Tackling situations when similar problems don‟t have similar solutions is a crucial challenge for CBR research [2]. This paper investigates a new CBR approach that relies on dissimilarity information for problem solving. We consider two cases to be dissimilar as long as they have distinct or “remote” solutions. The dissimilarity model will be established to enable identification of cases from the case library that are dissimilar to a query problem. A dissimilar case provides counter-evidence to some extent, suggesting the inappropriateness of using its solution for solving the query problem. It follows that the final decision from CBR will be the candidate solution that has accumulated the least amount of counter-evidence. Further the model of dissimilarity is represented as a set of fuzzy linguistic rules. We believe that fuzzy if-then rules present a powerful and flexible means to represent the rich domain knowledge for case evaluation. Fuzzy rule-based reasoning can be performed to predict whether and to which extent a case from the library is dissimilar to the problem in query. The construction of fuzzy dissimilarity rules can be realized by learning from the case library as a valuable resource. Our empirical studies have demonstrated that fuzzy dissimilarity models can be built upon a small case library while still yielding competent performance of the CBR system. The remaining of the paper is organized as follows: Section 2 surveys the related works. Section 3 outlines the new dissimilarity-based CBR approach proposed in the paper. The fuzzy dissimilarity model for case evaluation is presented in section 4. Then, in section 5, we discuss the issue of how to learn these fuzzy dissimilarity rules from the case library. In section 6, we illustrate experimental results for evaluation of the proposed method. Finally, section 7 gives the conclusion. Recent Advances in Systems Science and Mathematical Modelling ISBN: 978-1-61804-141-8 322 2 Related Works Similarity evaluation is a key part for conventional CBR systems. So far the main stream of the works for construction of similarity models has been focused on feature weighting [2, 3]. Features are assigned with different weights in accordance with their importance, and the global similarity metric is defined as a weighted sum of the local matching values in single attributes. Different approaches of interest have been proposed for identifying such weights automatically. Incremental learning attempts to modify feature weights according to success/failure feedback of retrieval results [4]. The probability of ranking principle was utilized in [5] for the assignment of weight values to features. Case-ranking information was utilized in [6, 7] for weight adaptation towards similarity degrees of retrieved cases consistent with a desired order. Accuracy improvement represents another way for adapting the set of weights as discussed in [8] and [9]. Nevertheless, no matter how the values of weights are derived, the capability of these similarity learning methods is inherently constrained by weighted combination of the local matching degrees. This limitation in the structure of similarity makes it hard to represent more general knowledge and criteria for case assessment. A new similarity model without feature weighting was proposed in [10] and [11] as an effort to seek more powerful representation of knowledge for case retrieval. The idea was to encode the information about feature importance into local compatibility measures such that feature weighting is no longer needed. Later, in [12], it was analyzed and demonstrated that the parameters of such compatibility measures can be learned from the case library in favor of coherent matching, i.e. to maximize the supportive evidence while minimize the amount of inconsistence derived from pairwise matching of cases from the case base. The integration of fuzzy systems with CBR methodology has been an interesting topic addressed by some other researchers. Yager [13] identified that there was a close connection between fuzzy system modeling and case based reasoning. In [14] the central notion of similarity in CBR was treated as a fuzzy relation and fuzzy operations were applied as a tool for building composite similarity measures. Dubois and Prade [15] formalized the fundamental hypothesis of CBR in the context of fuzzy rules. They established a formal framework in which case- based inference can be implemented as a special type of fuzzy set-based approximate reasoning. On the other hand, CBR can be used to assist fuzzy system modeling as well. In [16] and [17] it was demonstrated that CBR could be exploited as feature selection criterion for building complex process models and fuzzy systems. 3 Dissimilarity-Based CBR In this paper we propose a new CBR approach that relies on dissimilarity information, as is depicted in Fig. 1. It starts with comparison of the query problem with known cases in the case library. A properly defined dissimilarity function has to be employed at this stage. As the evaluated dissimilarity values reflect the strength of inappropriateness of solutions of the known cases to solve the new problem, they offer important information to be utilized in the next step of solution filtering to find out the least impossible choice as the final decision to solve the new problem. Case Library Dissimilarity Assessment ? Solution Filtering Dissimilarity degrees Solution Case Evaluation Query P Fig. 1. CBR based on dissimilarity information In solution filtering, we are concerned with estimating the degrees of impossibility of candidate solutions by using the case information from the case library. We assume solutions of cases to be represented by discrete and mutually exclusive labels in the context of this paper. We define the degrees of impossibility contributed by a single case Ci (from the case library) by       bCSolutionif bCSolutionifCQDsim bP i ii i )(0 )(),( )( (1) where b represents a candidate solution, and Dsim(Q, Ci) denotes the degree of dissimilarity between query problem Q and case Ci. It bears mentioning that the impossibility degrees in (1) indeed represent a degree of exclusion, which is supported by the observation of the dissimilar case Ci having solution b. On the other hand, we will have Pi(b)=0 if Ci has a solution different from b, whereas it merely means that no evidence against Recent Advances in Systems Science and Mathematical Modelling ISBN: 978-1-61804-141-8 323 solution b is derived from case Ci, rather than the support for b as the solution to query problem Q. Next we consider the overall impossibility degrees in light of the whole case library. For calculating the overall degree of impossibility Imposs(b) for solution b, we only need to focus on a subset of cases which have that solution. This is owing to the fact that all other cases in the case library contribute no information for the impossibility of solution b, as indicated in Eq. (1). The ordered weighted averaging (OWA) operators provide a class of aggregation operators lying between and and or aggregations. Herein we adopt the S-OWA-OR (OR-like) aggregating operators as the parameterized OWA functions to combine the degrees of impossibility given by the individual cases in the case subset. Let  bCSolutionS ib  )(i denote the set of indices of the cases having solution b, the overall impossibility value Imposs(b) is calculated as )()( 1 )1()(~)( bPbP S bPbImposs i Si Si i b i Si b b b      (2) 10  Finally we select the solution * b that has the lowest overall impossibility value as the final solution for query problem Q, i.e.,  )(minarg* bImPossb b  (3) 4 Fuzzy Dissimilarity Model This section introduces the structure of fuzzy rules that are used as representation of the dissimilarity model. Suppose there are n relevant features for problems in the underlying domain. A case Ci in the case library is described by an (n+1) tuple:   iiniii scccC ,,,, 21  where inii ccc ,,, 21  denote the feature values in this case and si is the corresponding solution. Likewise we use an n-tuple   n yyy ,,, 21  to represent a query problem Q with yj referring to the value of the jth feature in the problem. For comparing case Ci and query problem Q, we first need to calculate the values of differences ijjj cyx  on every feature j between them. Such feature differences are then employed as inputs for condition parts of the fuzzy rules, which collectively decide the dissimilarity value of case Ci with respect to the query problem. Assume that the fuzzy sets of feature difference xj (j=1 n) are represented by A(j,1), A(j,2), , A(j, q[j]) and q[j] is the number of linguistic terms for xj. By h() we denote an integer function mapping from {1, 2, ..., m (mn)} to {1, 2, ...., n} satisfying  ij, h(i)h(j). The fuzzy rules employed in this paper for assessing case dissimilarity are formulated as follows: VityDissimilarThen kmhAxandand khAxandkhAxIf mDk mh Dk h Dk h      )]),(([ )]),2(([)]),1(([ )( )( )2( )2( )1( )1(   (3) where D(i) {1, 2,  ,q[h(i)]} for i=1m, and V {1.0, 0}. Note that the conclusion of the rule in (3) is a singleton being either unity or zero, it can be regarded as a zero-order Sugeno fuzzy rule. It also bears noting that the premise structure defined above is very general, offering a large degree of flexibility in specification. If the premise of the above rule in (3) includes all input variables in it (e.g. m=n), we say that this rule has a complete structure, otherwise its structure is incomplete. Another important feature of the rules in form (3) is that a union of input fuzzy sets is allowed in their premises. Rules having incomplete structure or containing OR connections of input fuzzy sets can achieve larger coverage of input domain, leading to substantial reduction of the number of rules [18, 19]. Finally, with the availability of a set of fuzzy dissimilarity rules in the form of (3), the degree of dissimilarity between case Ci and query problem Q can be calculated as follows:       k ik k kik i CQt VCQt CQDsim ),,( ),( ),( (4) where Vk is the singleton conclusion for rule Rk, and tk denotes the firing strength of rule Rk when comparing case Ci and query problem Q. 5 Learning Fuzzy Dissimilarity Rules We now discuss how to generate fuzzy rules as formulated in the preceding section to build a dissimilarity model. Our aim is to elicit dissimilarity values between cases that can precisely reflect the level of distinction between their solutions. Supervised learning will be performed in this paper to acquire competent fuzzy rules for dissimilarity evaluation. In the following we will first explain how adequate training examples can be created for learning and then we shall outline a genetic algorithm (GA) for automatic generation of fuzzy rules to mimic the training examples. 5.1 Deriving Training Examples The training examples for fuzzy dissimilarity learning can be created by resorting to the case Recent Advances in Systems Science and Mathematical Modelling ISBN: 978-1-61804-141-8 324 library. Since case solutions there are available, it is straight forward to obtain the desired value of dissimilarity for a pair of cases by comparing their solutions. Hence the desired dissimilarity value between case Ci and Cj can be defined as: ),(),( jiji ssDistCCDsim  (5) where Dist represents distinction level, and si and sj are the solutions of cases Ci and Cj respectively. The criterion for judging distinction level between case solutions is usually domain dependent, thus we cannot further concretize equation (5) without considering problem context and specifics. Nevertheless, in this paper we assume case solutions are represented by discrete labels (e.g. classes), the distinction level between solutions can simply be expressed as follows: 1. If the solutions (labels) have no orders, the distinction level is a binary function as       ji ji ji ssif ssif ssDist 1 0 ),( (6) 2. If the solutions (labels) have ordinal values, the distinction level should reflect the relative distance in the order. Thus we have        ji ji ji ji ssif K sse ssif ssDist ),( 0 ),( (7) where K is the total number of labels and e(si, sj) denotes the number of labels between si and sj in the order. Pairs of problems Solution difference Fuzzy Rule Base Learning Algorithm Training samples Level of Distinction Value of Dissimilarity value Errors Fig. 2. Fuzzy learning from training samples The equations (5-7) enable us to acquire many training samples from pairs of cases in the case library. Since we can create a training example for every pair of cases, a much larger multitude of training samples than the number of cases will be created. Next, as shown in Fig. 2, the task of the learning algorithm is to identify optimal fuzzy rules together with associated membership functions such that the dissimilarity degrees assessed via fuzzy reasoning will comply with the distinction levels specified in the training samples. 5.2 Learning Rule Premises by Genetic Algorithms Learning the fuzzy rules formulated in (3) is solved by identifying suitable premises for different conclusions. For instance, we need to discover under what circumstances two cases in comparison should have a dissimilarity degree of unity. The issue as such is termed as premise learning. In this paper we apply the genetic algorithm (GA) introduced by Godenberg [20] to search for general premises of rules. The purpose is to take advantage of the strength of genetic search to find a set of suitable premise structures together with parameters of associated fuzzy set membership functions. In the following we shall state briefly about coding scheme, genetic operators and fitness function which present key points for the genetic learning of the fuzzy dissimilarity rules. Genetic Coding Scheme. The information concerning structure of rule premises can be considered as a set of discrete parameters, while the information about fuzzy set membership functions is described by a set of continuous parameters. Owing to the different natures between the information about rule structure and about membership functions, a hybrid string consisting of two substrings is used here as the coding scheme. The first substring is a binary code representing premise structure of the fuzzy knowledge base, and the second substring is an integer code corresponding to parameters of fuzzy sets used by the fuzzy rules. Usually membership functions of a feature difference as input are characterized by a set of parameters. Each of these parameters can further be mapped into an integer through discretization. The resulting integers are then combined to form an integer-vector depicting the fuzzy partition of that input variable. The integer code as one part of the hybrid string is formed by merging together integer- vectors for all inputs (feature differences) Regarding rule premises, it is easy to see from (3) that premise structure of general rules is decided by integer subsets D(i) (i=1, 2, m). This fact suggests that a binary code be a suitable scheme for encoding structure of premises, since an integer from {1, 2, , q[h(i)]} is either included in the subset D(i) or excluded from it. For feature difference xj which is included in the premise (i.e. h -1 (j)), q[j] binary bits need to be used to depict the subset D(h -1 (j)){1, 2,...., q[j]}, with bit "1" representing the presence of the corresponding Recent Advances in Systems Science and Mathematical Modelling ISBN: 978-1-61804-141-8 325 fuzzy set in the OR-connection and vice versa. If feature difference xj does not appear in the premise, i.e., h -1 (j)=, we use q[j] one-bits to describe the wildcard of „„don‟t care‟‟. For instance, the condition "if [x1=(small or large)] and [x3=medium] and [x4=(medium or large)]" can be coded by the binary group (101; 111; 010; 011). Further, the whole substring for the premise structure of the rule base is a merge of bit groups for all individual rule premises. Crossover. Owing to the distinct nature between the two substrings, it is preferable that the information in both substrings be mixed and exchanged separately. Here a three-point crossover is used. One breakpoint of this operation is fixed to be the splitting point between both substrings, and the other two breakpoints can be randomly selected within the two substrings respectively. At breakpoints the parent bits are alternatively passed on to the offspring. This means that offspring get bits from one of the parents until a breakpoint is encountered, at which they switch and take bits from the other parent. Mutation. Because of the distinct substrings used, different mutation schemes are needed. Since parameters of input membership functions are essentially continuous, a small mutation with high probability is more meaningful. Therefore it is so designed that each bit in the substring for membership functions undergoes a disturbance. The magnitude of this disturbance is determined by a Gaussian density function. For the binary substring representing the structure of rule premises, mutation is simply to inverse a bit, replace „1‟ with „0‟ and vice versa. Every bit in this substring undergoes a mutation with a quite low probability. Fitness Function. An individual (hybrid string), HS, in the population is evaluated according to its modeling accuracy with respect to the training examples. As many pairs of cases are included in the training data set, we have to consider the total sum of modeling errors for measuring the overall performance of the hybrid string. The total error function is given by    SIji jiji CCDsimssDistHSError ),( ),(),()( (8) where si and sj are the solutions of cases Ci and Cj respectively, and SI refers to the set of pairs of case indexes corresponding to pairs of cases included in the training data set. At last, the fitness of individual HS is defined with inverse relation to the mean modeling error as follows SI HSError HSFitness )( 1)(  (9) 6 Evaluation Results We have applied our proposed approach to the problems of classification and diagnosis. In this section we illustrate a case study made on the well- known benchmark problem of wine data classification. The wine data can be downloaded from the address ftp.ics.uci.edu/pub/machine- learning-databases. It consists of 178 samples with 13 continuous attributes from three classes. To test the feasibility of learning fuzzy dissimilarity rules from a small number of cases, we randomly selected 33% of the cases from the Wine data set as the case base used for learning and the remaining cases as test data containing query problems. The fuzzy rules learnt from the case base were then applied for dissimilarity assessment in case-based classification of the problems in the test data set. We performed such tests 10 times. Table 1 illustrates the classification accuracy on the test data in the 10 tests. It is interesting to observe that, despite the small case bases with about 66 instances in each, very good classification accuracy was still achieved by our CBR system employing the learned fuzzy dissimilarity models. Table 1: Classification accuracy on test data Numbers of tests Classification accuracy 1 97.48% 2 90.76% 3 94.12% 4 93.22% 5 92.44% 6 93.28% 7 91.53% 8 90.76% 9 90.76% 10 94.96% Average 92.93% In table 2, we compare our work with some other machine learning approaches in terms of classification accuracy (on test data) and the numbers of cases used for learning. The classification accuracy we obtained is rather close to the best result among the other works. In the other aspect, we employed a much lower number of cases for learning than any other work as indicated in the table. It demonstrates that our system can survive with learning from a small amount of examples. This is an attractive advantage distinguishing our Recent Advances in Systems Science and Mathematical Modelling ISBN: 978-1-61804-141-8 326 ftp://ftp.ics.uci.edu/pub/machine-learning-databases ftp://ftp.ics.uci.edu/pub/machine-learning-databases CBR system from other supervised learning systems for classification. Table 2: Comparison with other methods Learning methods Accuracy Number of cases for learning This paper 92.93% 59 ~ 60 C4.5 [21] 90.14% 160 ~ 161 Hu [22] 91.63% 160 ~ 161 MOP-1 [23] 96.01% 160 ~ 161 7 Conclusion The significance of this paper is of two folds. First, we proposed a new CBR approach that exploits dissimilarity information in problem solving. This new approach contributes to avoiding the similarity constraint with traditional CBR systems and thereby facilitating global utilization of more information from the case library. Secondly, we developed a new fuzzy dissimilarity model consisting of fuzzy rules for assessing dissimilarity between cases. We believe that fuzzy rules provide a powerful means to express rich domain knowledge for case evaluation in various situations. Further we have explained and demonstrated how competent fuzzy dissimilarity rules can be acquired from the case library by supervised learning. References: [1] R. L. D. Mantaras et al., Retrieval, reuse, revision and retention in case-based reasoning, The Knowledge Engineering Review, Vol. 20, 2005, pp. 215-240. [2] R. Kohavi, P. Langley, and Y. Yun, The utility of feature weighting in nearest neighbor algorithms, Proc. European Conf. Machine Learning, 1997. [3] D. Wettschereck, and D. Aha, Weighting features, Proc. 1st Int. Conf. on Case-based Reasoning, 1995, pp. 347-358. [4] A. Bonzano, P. Cunningham, and B. Smith, Using introspective learning to improve retrieval in CBR: A case study in air traffic control, Proc. 2nd Int. Conf. Case-based Reasoning, 1997, pp. 291-302. [5] N. Cercone, A. An, and C. Chan, Rule-induction and case-based reasoning: Hybrid architectures appear advantageous, IEEE Trans. Knowledge and Data Engineering, Vol. 11, 1999, pp. 166-174. [6] K. Branting, Acquiring customer preferences from return-set selections, Proc. 4th Int. Conf. Case-Based Reasoning, 2001, pp. 59-73. [7] L. Coyle, and P. Cunningham, Improving recommendation ranking by learning personal feature weights, Proc. 7th European Conference on Case-Based Reasoning, 2004, pp. 560-572. [8] J. Jarmulak, S. Craw, and R. Rowe, Genetic algorithms to optimize CBR retrieval, Proc. European Workshop on Case-Based Reasoning (EWCBR 2000), 2000, pp. 136-147. [9] H. Ahn, K. Kim, and I. Han, Global optimization of feature weights and the number of neighbors that combine in a case-based reasoning system, Expert Systems, Vol. 23, 2006, pp. 290-301. [10] N. Xiong, and P. Funk, Building similarity metrics reflecting utility in case-based reasoning, Intelligent & Fuzzy Systems, Vol. 17, 2006, pp. 407-416. [11] N. Xiong, and P. Funk, Combined feature selection and similarity modeling in case-based reasoning using hierarchical memetic algorithm, Proc. IEEE World Congress on Computational Intelligence, 2010, pp. 1537-1542. [12] N. Xiong, Towards coherent matching in case-based classification, Cybernetics and Systems, Vol. 42, 2011, pp. 198-214. [13] R. R. Yager, Case-based reasoning, fuzzy systems modeling and solution composition, Proc. 2nd Int. Conf. Case-Based Reasoning, Providence, RI, USA, 1997, pp. 633-643. [14] H.-D. Burkhard, and M. M. Richter, On the notion of similarity in case-based reasoning and fuzzy theory, S. K. Pal, T. S. Dillon and D. S. Yeung (Eds.), Soft Computing in Case-Based Reasoning, Springer, 2001, pp. 29-45. [15] D. Dubois, and H. Prade, Fuzzy set modeling in case-based reasoning, International Journal of Intelligent Systems, Vol. 13, 1998, pp. 345-373. [16] N. Xiong, A hybrid approach to input selection for complex processes, IEEE Transactions on Systems, Man, and Cybernetics, Part A: Systems and Humans, Vol. 32, No. 4, 2002, pp.532-536. [17] N. Xiong, and P. Funk, Construction of fuzzy knowledge bases incorporating feature selection, Soft Computing, Vol. 10, No. 9, 2006, pp. 796 – 804. [18] N. Xiong, and L. Litz, Learning premises of fuzzy rules for knowledge acquisition in classification problems, Knowledge and Information Systems, Vol. 4, 2002, pp. 96-111 [19] N. Xiong, and L. Litz, Reduction of fuzzy control rules by means of precise learning – method and case study, Fuzzy Sets and Systems, Vol. 132, 2002, pp. 217-231. [20] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning. New York: Addison-Wesley, 1989. [21] J. R. Quinlan, C4.5: Programs for machine learning, San Mateo: Morgan Kauffman, 1993. [22] Y. C. Hu, Finding useful fuzzy concepts for pattern classification using genetic algorithm, Information Sciences, Vol. 175, 2005, 1-19. [23] H. Ishibuchi, and Y. Nojima, Analysis of interpretability-accuracy tradeoff of fuzzy systems by multiobjective fuzzy genetic-based machine learning, International Journal of Approximate Reasoning, Vol. 44, 2007, pp. 4-31. Recent Advances in Systems Science and Mathematical Modelling ISBN: 978-1-61804-141-8 327 http://journals.cambridge.org/production/action/cjoGetFulltext?fulltextid=435264 http://journals.cambridge.org/production/action/cjoGetFulltext?fulltextid=435264 
work_iqxvluj6o5ca7jod3562ei7oza	----	Introduction Similarity measure models and algorithms for hierarchical cases Dianshuang Wu, Jie Lu, Guangquan Zhang Decision Systems & e-Service Intelligence (DeSI) Lab Centre for Quantum Computation & Intelligent Systems (QCIS) Faculty of Engineering and Information Technology, University of Technology, Sydney, P.O. Box 123, Broadway, NSW 2007, Australia Corresponding author: Jie Lu, Tel. : +61 02 95141838 E-mail: sd_wds@hotmail.com (D. Wu), jielu@it.uts.edu.au (J. Lu), zhangg@it.uts.edu.au (G. Zhang) Abstract Many business situations such as events, products and services, are often described in a hierarchical structure. When we use case-based reasoning (CBR) techniques to support business decision-making, we require a hierarchical-CBR technique which can effectively compare and measure similarity between two hierarchical cases. This study first defines hierarchical case trees (HC-trees) and discusses related features. It then develops a similarity evaluation model which takes into account all the information on nodes’ structures, concepts, weights, and values in order to comprehensively compare two hierarchical case trees. A similarity measure algorithm is proposed which includes a node concept correspondence degree computation algorithm and a maximum correspondence tree mapping construction algorithm, for HC-trees. We provide two illustrative examples to demonstrate the effectiveness of the proposed hierarchical case similarity evaluation model and algorithms, and possible applications in CBR systems. Keywords: Hierarchical similarity, Hierarchical cases, Tree similarity measuring, Case-based reasoning 1. Introduction Case-based reasoning (CBR) is the process of solving new problems based on the solutions for similar past problems (Aamodt & Plaza, 1994). CBR provides a powerful learning ability to use past experiences as a basis for dealing with new problems, and facilitates the knowledge acquisition process by reducing the time required to elicit solutions from experts. It is represented by a four-step (4Rs) cycle: retrieve, reuse, revise and retain (Aamodt & Plaza, 1994). In the first ‘R’ stage, when a new problem is input, CBR retrieves the most similar case from the case base. mailto:sd_wds@hotmail.com mailto:jielu@it.uts.edu.au mailto:zhangg@it.uts.edu.au Obviously, designing an effective case similarity evaluation method to identify the most similar cases is a key issue in CBR. Many models and algorithms have been developed to measure similarity between two cases, which are described in a set of attributes (Falkman, 2000). However, in practice, some cases can only be described by hierarchical tree structures. Therefore, we need to explore effective similarity measures for hierarchical cases in order to apply CBR systems. Fig. 1 shows an example of an avian flu case describing the infection situation of birds in an area at a given time (Zhang, Lu & Zhang 2009). Obviously, it is a hierarchical case and viewed as a tree structure. This tree case has seven nodes called “wild birds”, “farm poultry”,…, “water bird” and “no water bird”. The “water bird” node indicates that 40% of water birds were infected. From its edge, we can see 70% of farm poultry are water birds. Similarly, there are 60% of birds in the farm poultry area. A bird flu case Wild birds Farm poultry Long distance migratory Short distance migratory 0.4 0.6 0.60.4 Water bird No water bird 0.7 0.3 (0.3) (0.4) (0.4) (0.4) Fig.1. A hierarchical case: bird flu From this example, we can summarize the following features of tree cases: (1) every node is associated with a concept; (2) all concepts represented by nodes of a tree case, construct a hierarchical structure. Nodes at different depths represent concepts with different abstraction levels. The child nodes can be viewed as a refinement of the concept expressed by their parent node; (3) all leaves of a tree case are assigned values. Other nodes’ values can be assessed by aggregating their children’s; (4) every node is assigned a “weight” to represent its importance relative to its parent node. As different cases may arise from different sources at different times, the tree structures, nodes’ concepts, weights and values in different trees are probably not all the same. To evaluate the similarity between these tree-structured hierarchical cases, all the information should be considered. The research in this paper is related to work on tree similarity measure and structured case similarity measure. Tree structured data are used in many fields, such as e-business (Bhavsar, Boley, & Yang, 2004), bioinformatics (Tran, Nguyen, & Hoang, 2007), XML Schema matching (Jeong, Lee, Cho, & Lee, 2008), document classification and organization (Rahman & Chow, 2010) and case-based reasoning (Ricci & Senter, 1998). The similarity measure of tree structured data is essential for many applications. One widely used tree similarity measure is tree edit distance (Zhang, 1993; Kailing, Kriegel, Schonauer, & Seidl, 2004), in which edit operations including insertion, deletion and re-labeling with costs are defined, and the least cost of a sequence of edit operations needed to transform one tree to another is used as the similarity measure between the two trees (Bille, 2005). The main difference between various tree edit distance algorithms lies in the set of allowed edit operations and their related cost definitions (Yang, Kalnis, & Tung, 2005). In (Xue, Wang, Ghenniwa, & Shen, 2009), the conceptual similarity measure between labels was introduced in the cost of edit operations to compare concept trees of ontology. Another kind of tree similarity measure is based on a maximum common sub-tree (MCS) or sub-tree isomorphism between two trees (Akutsu & Halldorsson, 2000). This method uses the size of MCS between two trees, or metrics defined by MCS as the similarity measure. In (Torsello, Hidovic, & Pelillo, 2004), four novel distance measures for attributed trees based on the notion of a maximum similarity sub-tree isomorphism were proposed. In (Lin, Wang, McClean, & Liu, 2008), the number of all common embedded sub-trees between two trees was used as the measure of similarity. The methods mentioned above mostly deal with node-labeled trees. In (Bhavsar, Boley, & Yang, 2004), node-labeled, arc-labeled, arc-weighted trees were used as product/service descriptions to represent the hierarchical relationship between the attributes. To compare these trees, a recursive algorithm to perform a top-down traversal of trees and the bottom-up computation of similarity was designed. However, their trees had to conform to the same standard schema, i.e. the trees should have the same structure and use the same labels, though some sub-trees were allowed to be missing (Yang, Sarker, Bhavsar, & Boley, 2005). As trees for hierarchical cases in our research are different to previous ones, we need to develop a new similarity measure method for them. Structured case similarity measure in the literature is usually based on the maximal common sub-graph or sub-graph isomorphism (Burke, MacCarthy, Petrovic, & Qu, 2000; Sanders, Kettler, & Hendler, 1997). In (Ricci & Senter, 1998), the similarity measure on tree structured cases, taking into account both the tree structures and node labels’ semantics, was researched. A sub-tree isomorphism with the minimum semantic distance was constructed and the minimum semantic distance was used as the similarity measure. This research is closely related to ours. However, the positions of corresponding nodes are not restricted in their sub-tree isomorphism, and this is not suitable for our hierarchical cases, because nodes at different depths represent concepts at different abstraction levels. Also, nodes’ values are not involved in their similarity measure. In this paper, we present a comprehensive similarity evaluation model considering all the information on nodes’ structures, concepts, weights and values to compare tree structured hierarchical cases. To express the concept correspondence between nodes in different trees, the concept correspondence degree is defined. A maximum correspondence tree mapping based on nodes’ structures and concepts is constructed to identify the corresponding nodes between two trees. Based on the mapping, the values of corresponding nodes are compared. Finally, the similarity measure of trees is evaluated by aggregating both the conceptual and value similarities. This paper is organized as follows. In Section 2 we describe the features of hierarchical case trees by mathematical formulas. Section 3 presents a similarity evaluation model to compare any two hierarchical case trees. A set of algorithms to compute the similarity between hierarchical cases is provided in Section 4. Section 5 presents two examples to demonstrate the effectiveness of the proposed hierarchical case similarity evaluation model and algorithms, and possible applications in CBR systems. It also compares the proposed HC-tree similarity model with other approaches. Section 6 concludes the paper and discusses tasks for our further study. 2. Hierarchical case trees A tree is defined as a directed graph ),( EVT = where the underlying undirected graph has no cycles and there is a distinguished root node in V , denoted by )(Troot , so that for all nodes Vv ∈ , there is a path in T from )(Troot to node v (Valiente, 2002). In real applications, the definition can be extended to represent practical objects. To express the concepts, values and weights associated with the nodes of hierarchical cases and the hierarchical relationships between nodes, the original tree structure is enriched and a hierarchical case tree (HC-tree) is defined. Definition 2.1: HC-tree. An HC-tree is a structure ),,,,( RWAEVT = , in which V is a finite set of nodes, E is a binary relation on V where each pair Evu ∈),( represents the parent-child relationship between two nodes Vvu ∈, , A is a set of attributes assigned to each node in V , W is a function to assign each node a weight to represent its degree of importance to its siblings, thereby satisfying the sum of the weights of all the children of one node is 1, and R is a function to assign a value to every leaf node to describe the degree of its relevant attribute. Two features of the HC-tree should be highlighted. First, all nodes in the HC-tree represent concept meanings, which are obtained from the attributes. As a hierarchical structure, the concept of one node depends, not only on the attribute itself, but also its children’s. Therefore, nodes at different depths represent concepts at different abstraction levels, and nodes at higher layers represent more significant concepts than lower nodes. Secondly, every node in the HC-tree has a value. The leaves’ values are indicated by R , and the internal nodes’ values can be computed by aggregating their children’s. resident bird v1 v2 0.5 0.5 T1 A bird flu case v3 migratory bird (0.8) long distance migratory 0.50.5 (0.7)(0.3) v4 v5 short distance migratory long distance migratory short distance migratory 0.60.4 (0.2) u1 u2 0.4 0.6 T2 A bird flu case wild birds farm poultry water bird no water bird 0.7 0.3 (0.3) (0.3) (0.3) u3 u6 u7u4 u5 Fig.2. Two examples of HC-trees Two examples of HC-trees, both describing the situation of bird flu, are illustrated in Fig. 2. The labels beside the nodes represent their attributes. The number beside the edge is the weight of the child. The number under each leaf represents its value. In 1T , “A bird flu case” is described by two aspects, “migratory bird” and “resident bird”, with both taking the same weight. Similarly, “migratory bird” is described by two sub-aspects, “long distance migratory” and “short distance migratory”. From “long distance migratory”, we can see that 30% of birds were infected. As 1T and 2T are from different sources, their structures and nodes’ weights are different. Their attribute terms are also not identical. To evaluate the conceptual correspondence between attributes in different HC-trees, a conceptual similarity measure between attributes is introduced as in (Xue, Wang, Ghenniwa, & Shen, 2009). Definition 2.2: Attribute Conceptual Similarity Measure. An attribute conceptual similarity measure 21 , AA sc is a set of mappings from two attribute sets 1A , 2A used in different HC-trees to the set [0, 1], ]1,0[: 21, 21 →× AAsc AA , in which each mapping denotes the conceptual similarity between two attributes. For convenience, the sub-script 21 , AA is omitted so that there is no confusion. For 11 Aa ∈ and 22 Aa ∈ , we say 1a and 2a are similar if 0),( 21 >aasc , and the larger ),( 21 aasc is, the more similar the two attributes are. Conceptual similarity between two attributes can be given by domain experts or calculated based on linguistic analysis methods. As an example, we define the conceptual similarity between the attributes of 1T and 2T in Fig. 2 as follows: sc(migratory bird, wild birds)=0.7, sc(migratory bird, farm poultry)=0.1, sc(resident bird, wild birds)=0.6, sc(resident bird, farm poultry)=0.8, sc(resident bird, water bird)=0.4, sc(resident bird, no water bird)=0.4, sc(long distance migratory, water bird)=0.1, sc(long distance migratory, no water bird)=0.2, sc(short distance migratory, water bird)=0.2, sc(short distance migratory, no water bird)=0.2. 3. A similarity evaluation model for HC-trees A similarity evaluation model for HC-trees is proposed in this section. In the model, maximum correspondence tree mapping is constructed to identify the corresponding node pairs of two HC-trees based on nodes’ structures and concepts, and the conceptual similarity between two HC-trees is evaluated. Based on the mapping, the value similarity between two HC-trees is evaluated, and the final similarity measure between two HC-trees is assessed as a weighted sum of their conceptual and value similarities. 3.1 Maximum correspondence tree mapping To identify two corresponding nodes in different HC-trees, both their structures and concepts should be considered. There are two structural restrictions. First, as nodes at different depths represent concepts at different abstraction levels, it is reasonable to assume that the corresponding nodes in the mapping should be at the same depth. Therefore, the roots of two HC-trees should be in the mapping. Secondly, as the children nodes can be viewed as a refinement of the concept expressed by the parent node, two separate sub-trees in one tree should be mapped to two separate sub-trees in another. In addition to satisfying structural restrictions, it is important that the corresponding nodes have a high conceptual similarity degree. To express the concept correspondence between two nodes in two HC-trees respectively, the following definition is introduced: Definition 3.1: Node Concept Correspondence Degree. Let 1V and 2V be node sets of 1T and 2T respectively. A node concept correspondence degree cord is a set of mappings from 1V and 2V to the set [0, 1], ]1,0[: 21 →×VVcord , in which each mapping denotes the concept correspondence between two nodes of two HC-trees. cord is symmetric, i.e. for 1Vv ∈ and 2Vu ∈ we have ),( uvcord = ),( vucord . Let v and u be two nodes of 1T and 2T respectively. There are three cases: (1) both v and u are leaves, (2) v is a leaf and u is an internal node, (3) both v and u are internal nodes. In the first case, as nodes’ concepts are derived from the attributes assigned to them, the concept correspondence degree between v and u can be defined as the conceptual similarity of their attributes. In the other two cases, as the internal node’s concept is also affected by its children, the children’s concepts should be considered in the definition. Thus, they should be defined recursively. The definitions of cord for the three cases are presented respectively as follows: Definition 3.2: Concept Correspondence Degree between Two Leaves. Let v and u be two leaves of 1T and 2T respectively. The concept correspondence degree between v and u , ),( uvcord is defined as: ).,.(),( auavscuvcord = (3.1) where av. and au. represent attributes of v and u respectively. For example, 4v and 6u are two leaves of 1T and 2T respectively in Fig. 2. The attribute of 4v is “long distance migratory” and of 6u is “water bird”. The concept correspondence degree between 4v and 6u is defined by sc(long distance migratory, water bird), and ),( 64 uvcord =0.1. Definition 3.3: Concept Correspondence Degree between a Leaf and an Internal Node. Let v be a leaf of 1T , u be an internal node of 2T , and },...,,{)( 21 quuuuC = be u ’s children set. The concept correspondence degree between v and u , ),( uvcord is defined as: ∑ = ⋅⋅−+⋅= q i ii uvcordwauavscuvcord 1 2 ),()1().,.(),( aa (3.2) where a is the influence factor of the parent node and iw2 is the weight of iu . For example, 3v is a leaf node of 1T and 3u is an internal node of 2T in Fig. 2. The concept correspondence degree between 3v and 3u is computed by the formula )),(3.0),(7.0()1().,.(),( 73633333 uvcorduvcordauavscuvcord ⋅+⋅⋅−+⋅= aa . In the formula, ).,.( 33 auavsc is 0.8. ),( 63 uvcord and ),( 73 uvcord can be computed by Definition 3.2, and ),( 63 uvcord =0.4 and ),( 73 uvcord =0.4. If a is 0.5, we can achieve ),( 33 uvcord =0.6. Definition 3.4: Concept Correspondence Degree between Two Internal Nodes. Let v and u be two internal nodes of 1T and 2T respectively, and },...,,{)( 21 pvvvvC = and },...,,{)( 21 quuuuC = be v and u ’s children sets, respectively. Let ),(, EVG uv = denote the bipartite graph induced by v and u , which is constructed as follows: )()( uCvCV ∪= , )}(),(:),{( uCtvCstsE ∈∈= . The weights of edges are defined as ),(, tscordweight ts = . uvMWM , is the maximum weighted bipartite matching of uvG , . Then, the correspondence degree between v and u , ),( uvcord is defined as: ),()( 2 1 )1().,.(),( ,),( 21 ji MWMuv ji uvcordwwauavscuvcord uvji ∑ ∈ ⋅+⋅−+⋅= aa (3.3) where iw1 is the weight of iv in 1T and jw2 is the weight of ju in 2T . In Definition 3.4, the maximum weighted bipartite matching uvMWM , identifies the most correspondence node pairs amongst v and u ’s children. The contribution of their children can therefore be fully considered when evaluating their concept correspondence degree. For example, 2v and 3u are two internal nodes of 1T and 2T respectively in Fig. 2. To compute their concept correspondence degree, a bipartite graph 32 ,uv G is constructed as Fig. 3 (a), in which the numbers beside the edges represent their weights. The maximum weighted bipartite matching of 32 ,uv G is illustrated in Fig. 3 (b). Then, The concept correspondence degree between 2v and 3u is computed by the formula ),( 32 uvcord = ).,.( 32 auavsc⋅a + )),()2)7.05.0((),()2)3.05.0((()1( 6574 uvcorduvcord ⋅++⋅+⋅−a . In the formula, ).,.( 32 auavsc is 0.1. ),( 74 uvcord and ),( 65 uvcord are computed by Definition 3.2, and ),( 74 uvcord =0.2 and ),( 65 uvcord =0.2. Let a be 0.5, so that we achieve ),( 33 uvcord =0.15. u6 u7 v4 v5 0.1 0.2 0.2 0.2 u6 u7 v4 v5 0.2 0.2 (a) (b) Fig.3. A bipartite graph 32 ,uv G and its maximum weighted bipartite matching With the above definitions, the concept correspondence degree of any node pair between two HC-trees can be evaluated. The maximum correspondence tree mapping considering both the structural restrictions and nodes’ concept correspondence is defined as follows. Definition 3.5: Maximum Correspondence Tree Mapping. Let 1V and 2V be node sets of HC-trees 1T and 2T , respectively. A mapping 21 VVM ×⊆ is a maximum correspondence tree mapping if it satisfies the following conditions: 1. 2121 uuvv =⇔= for any pair ),( 11 uv , Muv ∈),( 22 2. MTrootTroot ∈))(),(( 21 3. Muparentvparent ∈))(),(( for all non-root nodes 1Vv ∈ and 2Vu ∈ with Muv ∈),( 4. 0),( >uvcord for all nodes 1Vv ∈ and 2Vu ∈ with Muv ∈),( 5. MMWM uv ⊂, for all nodes 1Vv ∈ and 2Vu ∈ with Muv ∈),( , where uvMWM , is the maximum weighted bipartite matching of bipartite graph uvG , constructed of v and u ’s children with edges weighted by their children’s concept correspondence degree. In the above Definition 3.5, the first condition ensures that the mapping is a one-to-one mapping. Conditions 2 and 3 ensure the mapping satisfies the structural restrictions. The last two conditions represent the conceptual restrictions. Condition 5 ensures that most correspondence node pairs are in the mapping. As an example, the maximum correspondence tree mapping between 1T and 2T in Fig. 2 is illustrated in Fig. 4, in which corresponding nodes are connected by dashes. The construction process of the mapping will be described in Section 5.1. resident bird v1 v2 0.5 0.5 T1 A bird flu case v3 migratory bird (0.8) long distance migratory 0.50.5 (0.7)(0.3) v4 v5 short distance migratory long distance migratory short distance migratory 0.60.4 (0.2) u1 u2 0.4 0.6 T2 A bird flu case wild birds farm poultry water bird no water bird 0.7 0.3 (0.3) (0.3) (0.3) u3 u6 u7u4 u5 Fig.4. Maximum correspondence tree mapping between 1T and 2T From the recursive definitions of node concept correspondence degree, it is obvious that ))(),(( 21 TrootTrootcord is computed by aggregating the cord of all corresponding node pairs, which reflects the conceptual similarity between two HC-trees. We can define the conceptual similarity between two HC-trees as follows: Definition 3.6: Conceptual Similarity between HC-trees. Let 1T and 2T be two HC-trees. The conceptual similarity between 1T and 2T , ),( 21 TTsct is defined as ),( 21 TTsct = ))(),(( 21 TrootTrootcord . Taking 1T and 2T in Fig. 2 as an example, their conceptual similarity, ),( 21 TTsct is computed as ),( 11 uvcord . 3.2 Value similarity between HC-trees Based on the maximum correspondence tree mapping M , the values of two HC-trees can be compared. The value similarity between two corresponding nodes in M is evaluated first. As only leaf nodes are assigned values in HC-trees initially, for any Muv ∈),( , there are two cases: (1) v is a leaf node, or none of v ’s children are in M , (2) some of v ’s children are in M . We provide the computation formulas of the value similarity between v and u , ),( uvsvM for the two cases respectively. For case 1, ),( uvsvM is computed as: ))(),((),( uvaluevvaluesuvsvM = (3.4) where )(vvalue denotes v ’s value and )(⋅s denotes a value similarity measure. If v is a leaf node, )(vvalue is assigned initially. Otherwise, it is computed by aggregating its children’s values. )(⋅s can be defined according to the specific applications. For example, let two attributes’ values be 1a and 2a , and their value range be r ; then their similarity measure can be defined as raaaas 2121 1),( −−= . In the example in Fig. 4, as values of nodes are all within [0, 1], the similarity between two values is calculated as one minus the distance between them. For 3v and 3u in Fig. 4, )( 3vvalue is assigned initially as 0.8, and )( 3uvalue can be computed as 0.3. The value similarity between 3v and 3u is then computed as 0.5. In case 2, let pvvv ,...,, 21 be v ’s children and quuu ,...,, 21 be u ’s children. ),( uvsvM is computed as: ),()( 2 1 ),( 21 ),( jiMji Muv M uvsvwwuvsv ji ⋅+= ∑ ∈ (3.5) where iw1 is the weight of iv in 1T and jw2 is the weight of ju in 2T . Take 2v and 2u in Fig. 4 as an example. Their value similarity is computed as ),( 22 uvsvM = ),()2)4.05.0(( 44 uvsvM⋅+ + ),()2)6.05.0(( 55 uvsvM⋅+ . In the formula, ),( 44 uvsvM and ),( 55 uvsvM can be calculated by Formula (3.4), and ),( 22 uvsvM =0.725. With Formula (3.4) and (3.5), the value similarity between any corresponding nodes in M can be computed. As the recursive characteristic of Formula (3.5), the value similarity between the roots of two HC-trees is computed by aggregating the value similarity of all corresponding node pairs, which represents the value similarity of the two HC-trees. Therefore, we define the value similarity between two HC-trees as follows. Definition 3.7: Value Similarity between HC-trees. Let 1T and 2T be two HC-trees, and M be their maximum correspondence tree mapping. The value similarity between 1T and 2T , ),( 21 TTsvt is defined as ),( 21 TTsvt = ))(),(( 21 TrootTrootsvM . Taking 1T and 2T in Fig. 4 as an example, their value similarity, ),( 21 TTsvt is computed as ),( 11 uvsvM . 3.3 Similarity measure of HC-trees Based on the conceptual similarity and value similarity of two HC-trees, the similarity measure of HC-trees is defined as follows. Definition 3.8: Similarity Measure of HC-trees. The similarity between 1T and 2T is defined as: ),(),(),( 21221121 TTsvtTTsctTTsim ⋅+⋅= aa (3.6) where 21 aa + =1. In this definition, both the concepts and values of two HC-trees are comprehensively considered. 1a and 2a are weights of the two parts, which can be defined according to the specific applications. 4. Similarity measurement algorithms for HC-trees Algorithms to compute the similarity between two HC-trees are presented in this section. The flowchart in Fig. 5 shows the entire process. Fig.5. Flowchart to compute the similarity between two HC-trees We can see from Fig.5 that ),( 21 TTsct is firstly computed by calling )),(),(( 21 BTrootTrootcord , where B is a node set list which is indexed by the nodes in 1T . All the maximum weighted bipartite matching solutions during computing ))(),(( 21 TrootTrootcord are recorded in B . The maximum correspondence tree mapping M is then constructed based on B by calling ),,( 21 TTBapConstructM . ),( 21 TTsvt is computed based on M . Finally, the similarity of 1T and 2T is returned by aggregating their conceptual and value similarities. The algorithm is illustrated as follows. Start Call )),(),(( 21 BTrootTrootcord to compute ),( 21 TTsct , record the maximum weighted bipartite matching solutions in B Initialization Call ),,( 21 TTBapConstructM to construct the maximum correspondence tree mapping M between 1T and 2T Compute ),( 21 TTsvt based on M ),(),(),( 21221121 TTsvtTTsctTTsim ⋅+⋅= aa End Algorithm 1. Similarity measure algorithm for HC-trees similarity( 1T , 2T ) input: two trees 1T and 2T output: similarity between 1T and 2T 1 for all Vv ∈ Φ←)(vB 2 )),(),(( 21 BTrootTrootcordsct ← 3 ←M ),,( 21 TTBapConstructM 4 ))(),(( 21 TrootTrootsvsvt M← 5 return svtsct ⋅+⋅ 21 aa The algorithm of concept correspondence degree computation function ),,( Buvcord is illustrated by algorithm 2 as follows. Algorithm 2. Node concept correspondence degree computation algorithm ),,( Buvcord input: two nodes v and u output: concept correspondence degree between v and u 1 if both v and u are leaves 2 return ).,.( auavsc 3 else if u is an internal node, and quuu ,...,, 21 be u ’s children, 4 return ∑ = ⋅⋅−+⋅ q i ii Buvcordwauavsc 1 2 ),,()1().,.( aa 5 else ←)(vC v ’s children pvvv ,...,, 21 6 ←)(uC u ’s children quuu ,...,, 21 7 for i=1 to p 8 for j=1 to q 9 ),,( Buvcordc jiij ← 10 ←m )),()(( cuCvCchingComputeMat ∪ 11 for each muv lk ∈),( , if 0>klc 12 }{)()( lkk uvBvB ∪← 13 return klmuv lk cwwauavsc lk∑ ∈ ⋅+⋅−+⋅ ),( 21 )2)(()1().,.( aa A recursive process follows from the definition of concept correspondence degree. The most important part in the procedure are lines 5-13, where both v and u are internal nodes. A bipartite graph is constructed, taking their children as nodes, and the correspondence degrees between their children as the weights of edges. Function )(⋅chingComputeMat (Jungnickel, 2008) returns the maximum weighted bipartite matching, which identifies most correspondence node pairs among v and u ’s children. The matches are recorded in B , which are local maximum correspondence matches. For one node in 1T , there may be more than one node matching it during the computation process. However, as proved in (Valiente, 2002), there is a unique maximum correspondence tree mapping 21 VVM ×⊆ so that BM ⊆ . Given B , the corresponding maximum correspondence tree mapping M can be reconstructed as follows: Set ))(( 1TrootM to )( 2Troot and, for all nodes 1Vv ∈ in pre-order, set )(vM to the unique node u with Buv ∈),( and Buparentvparent ∈))(),(( . The reconstruction procedure is illustrated by algorithm 3 (Valiente, 2002). Algorithm 3. Maximum correspondence tree mapping construction algorithm ),,( 21 TTBapConstructM input: node set list B , two HC-trees 1T and 2T output: maximum correspondence tree mapping M from 1T to 2T 1 )())(( 21 TrootTrootM ← 2 list )(_ 1TtraversalpreorderL ← 3 for all Lv ∈ http://find.lib.uts.edu.au/search.do?N=4294340440 http://find.lib.uts.edu.au/search.do?N=4294340440 4 if v is nonroot and Φ≠)(vB 5 for all )(vBu ∈ 6 if )())(( uparentvparentM == 7 uvM ←)( 8 break 9 return M 5. Two illustrative examples and comparison with other approaches The proposed HC-tree similarity model and algorithms are to be used in CBR systems, such as CBR-based warning systems (Zhang, Lu & Zhang 2009), CBR-based recommender systems (Lu et al 2010) and web mining systems (Wang, Lu & Zhang, 2007). To show the effectiveness of our model, two examples are provided in this section. In the first example, the process of computing the similarity between two HC-trees in Fig. 2 is presented to show the behavior of the proposed algorithms in Section 4. In the second example, our similarity model is used in the retrieve stage of a simple CBR system to demonstrate the effectiveness of the model. The proposed model is then compared with other tree similarity evaluation methods. 5.1 Similarity measure computation between two HC-trees The similarity between 1T and 2T in Fig. 2 is computed by the proposed similarity measurement algorithms as follows. First, the conceptual similarity between 1T and 2T , ),( 21 TTsct is computed by calling ),,( 11 Buvcord . Let the coefficient a be 0.5, ),( 21 TTsct is computed as 0.856. During the recursive computation process, many maximum weighted bipartite matching problems are resolved, and the solutions are recorded in B : }{)( 22 uvB = , }{)( 33 uvB = , },{)( 744 uuvB = , },{)( 655 uuvB = . Secondly, given B , the maximum correspondence tree mapping M between 1T and 2T is constructed by calling ),,( 21 TTBapConstructM : }{)( 11 uvM = , }{)( 22 uvM = , }{)( 33 uvM = , }{)( 44 uvM = , }{)( 55 uvM = . The mapping is illustrated in Fig. 4. Based on the mapping M , the value similarity between 1T and 2T , ),( 21 TTsvt is evaluated by computing ),( 11 uvsvM . The computation uses Formulas 3.4 and 3.5 to achieve ),( 21 TTsvt as 0.6. Finally, let the weights 1a and 2a be both 0.5; the final similarity measurement between 1T and 2T , ),( 21 TTsim is computed by ),(5.0),(5.0 2121 TTsvtTTsct ⋅+⋅ =0.73. 5.2 Similar cases retrieval The proposed similarity model is used to retrieve similar cases in a CBR system in the following example. R A B C D E F G 0.4 0.4 0.2 0.4 0.4 0.2 0.5 (0.6) (0.2) (0.4) (0.3) (0.3) T1 0.5 H (0.3) r a b c d e f 0.5 0.4 0.1 0.6 0.4 0.5 (0.7) (0.1) (0.2) (0.6) Ta 0.5 g (0.3) R A B C E F G 0.4 0.4 0.2 0.4 0.4 0.2 0.5 (0.1) (0.8) (0.9) (0.8) (0.4) T2 0.5 H (0.9) D R’ A’ B’ C’ D’ E’ F’ G’ 0.4 0.4 0.2 0.4 0.4 0.2 0.5 (0.6) (0.2) (0.4) (0.3) (0.3) T3 0.5 H’ (0.3) R A B C D E F G 0.3 0.2 0.5 0.1 0.3 0.6 0.5 (0.6) (0.2) (0.4) (0.3) (0.3) T4 0.5 H (0.3) R A JC D E F K 0.4 0.40.2 0.4 0.4 0.2 0.5 (0.6) (0.2) (0.4) (0.3) (0.3) T5 0.5 H (0.3) Fig.6. A new case aT and five existing cases in a case base As illustrated in Fig. 6, HC-tree aT represents a new problem to be resolved and 1T ,…, 5T represent five solved problems in a case base. The conceptual similarity between their attributes is defined as follows: sc(r,R)=0.7, sc(a,A)=0.9, sc(a,B)=0.6, sc(b,A)=0.5, sc(b,B)=0.8, sc(d,D)=1, sc(d,E)=0.5, sc(d,G)=0.4, sc(e,D)=0.5, sc(e,E)=0.9, sc(e,H)=0.4, sc(f,F)=1, sc(f,H)=0.6, sc(f,G)=0.7, sc(g,G)=0.9, sc(g,H)=0.7, sc(r,R’)=0.6, sc(a,A’)=0.7, sc(a,B’)=0.6, sc(b,A’)=0.5, sc(b,B’)=0.7, sc(d,D’)=0.9, sc(d,E’)=0.4, sc(d,G’)=0.3, sc(e,D’)=0.4, sc(e,E’)=0.8, sc(e,H’)=0.3, sc(f,F’)=0.7, sc(f,H’)=0.6, sc(f,G’)=0.6, sc(g,G’)=0.7, sc(g,H’)=0.6. To retrieve the most similar cases to aT , the similarities between aT and cases in the case base are evaluated using the similarity model proposed in this paper. Let the coefficients a , 1a and 2a be all 0.5 in the model, and the similarity between two values be calculated as one minus the distance between them. The results are illustrated in Table 1. As can be seen in Table 1, 1T is most similar to aT , so 1T is retrieved. Table 1 Similarity between aT and cases in the case base 1T 2T 3T 4T 5T ),( ia TTsct 0.703 0.703 0.600 0.623 0.548 ),( ia TTsvt 0.745 0.304 0.745 0.537 0.365 ),( ia TTsim 0.724 0.504 0.672 0.580 0.456 As seen from Fig. 6, 2T and 1T are the same except for their values. Therefore, the conceptual similarity ),( 1TTsct a and ),( 2TTsct a are equal. However, as 1T ’s values are much closer to aT ’s than 2T ’s, ),( 1TTsvt a is larger than ),( 2TTsvt a , which makes 1T more similar to aT than 2T . 3T and 1T are different in terms of attributes. The concepts of 1T ’s attributes are more similar to aT ’s than 3T ’s, which makes 1T more similar to aT than 3T . 4T and 1T have different attribute weights. The weights of nodes corresponding to aT in 4T are smaller than those in 1T , which makes 4T less similar to aT than 1T . The example shows that our similarity model takes into account all the information on nodes’ structures, concepts, weights and values and it can be used to retrieve the most similar cases effectively in CBR systems. 5.3 Comparison with other approaches From the above examples, it can be seen that the proposed similarity evaluation model for HC-trees has five features: (1) it considers nodes’ conceptual similarities; (2) it considers the hierarchical relations between concepts; (3) it compares corresponding nodes’ values; (4) it considers the influence of nodes’ weights; (5) it considers the semantics of nodes’ structures. We compare our method with other tree similarity evaluation methods for these five aspects. We take into account the methods of Ricci, & Senter’s (1998), Xue, Wang, Ghenniwa, & Shen’s (2009) and Bhavsar, Boley, & Yang’s (2004), as they can represent different types of methods, respectively. The comparison results are illustrated in Table 2, where “√” represents that the method has the related feature. Table 2 demonstrates that none of the earlier methods can compare the tree structured data as comprehensively as our method. However, these features are essential to evaluate the similarity between complex tree structured hierarchical cases. As different HC-trees usually have different structures and attribute terms, the corresponding nodes between two HC-trees must be identified by evaluating their conceptual similarity. As attributes in hierarchical cases construct a hierarchical structure, the hierarchical relations between concepts and the semantics of nodes’ structures must be considered. Nodes’ values and relevant weights are essential to describe the case, so they must be compared when comparing two cases. With the above five features, the HC-trees can be compared comprehensively and accurately, and the most similar cases can be retrieved. Therefore, the proposed HC-tree similarity evaluation model is extremely suitable for retrieval of similar cases in CBR systems. Table 2 Comparison between our proposed method and other methods Method Feature 1 Feature 2 Feature 3 Feature 4 Feature 5 Our method √ √ √ √ √ Ricci’s method √ √ Xue’s method √ √ Bhavsar’s method √ √ √ 6. Conclusion and future work This paper defines the hierarchical case trees (HC-trees) to represent hierarchical cases. A similarity evaluation model to compare HC-trees is proposed and the related algorithms are presented. The concept correspondence degree between nodes is defined in the model and the conceptual similarity between trees is evaluated; a maximum correspondence tree mapping based on nodes’ structures and concepts is proposed to identify the corresponding nodes of two trees; the value similarity between two trees is computed based on the mapping; the final similarity measure between two trees is evaluated by aggregating their conceptual and value similarities. Two illustrative examples show that our method is highly effective for use in CBR systems. Our future research includes (1) to define fuzzy-HC-trees and a fuzzy similarity evaluation model based on our previous study (Lu, Zhang, Ruan & Wu, 2007) and propose related algorithms in order to improve inference accuracy in CBR systems; (2) to develop software based on the proposed similarity evaluation model for integration into real CBR systems, such as our BizSeeker recommender system (Lu et al. 2010) helping measuring similarity between two business on their product trees and our CBR-based avian influenza risk early warning (Zhang, Lu & Zhang, 2009). Acknowledgements The work presented in this paper was supported by the Australian Research Council (ARC) under Discovery Project DP0880739 and the China Scholarship Council. References Aamodt, A., & Plaza, E. (1994). Case-based reasoning: foundational issues, methodological variations, and system approaches. AI Communication, 7(1), 39-59. Akutsu, T. & Halldorsson, M.M. (2000). On the approximation of largest common subtrees and largest common point sets. Theoretical Computer Science, 233, 33-50. Bhavsar, V.C. Boley, H. & Yang, L. (2004). A weighted-tree similarity algorithm for multi-agent systems in e-business environments. Computational Intelligence, 20(4), 584-602. Bille, P. (2005). A survey on tree edit distance and related problems. Theoretical Computer Science, 337(1-3), 217-239. Burke, E. MacCarthy, B. Petrovic, S. & Qu, R. (2000). Structured cases in case-based reasoning--re-using and adapting cases for time-tabling problems. Knowledge-Based Systems, 13(2-3), 159-165. Falkman, G. (2000). Similarity measures for structured representations: a definitional approach. In E. Blanzieri, & L. Portinale (Eds.), Advances in Case-Based Reasoning: 5th European Workshop, EWCBR 2000, Trento, Italy, September 6-9, 2000 Proceedings. Lecture Notes in Artificial Intelligence, 1898, 380-392. Springer-Verlag Berlin Heidelberg. Jeong, B. Lee, D. Cho, H. & Lee, J. (2008). A novel method for measuring semantic similarity for XML schema matching. Expert Systems with Applications, 34(3), 1651-1658. Jungnickel, D. (2008). Graphs, networks, and algorithms, Berlin: Springer, 419-430. Kailing, K. Kriegel, H.P. Schonauer, S. & Seidl, T. (2004). Efficient similarity search for hierarchical data in large databases. In E. Bertino et al. (Eds.), Advances in Database Technology -EDBT 2004: 9th International Conference on Extending Database Technology, Heraklion, Crete, Greece, March 14-18, 2004 Proceedings. Lecture Notes in Computer Science, 2992, 676-693. Springer-Verlag Berlin Heidelberg. Lin, Z. Wang, H. McClean, S. & Liu, C. (2008). All common embedded subtrees for measuring http://find.lib.uts.edu.au/search.do?N=4294340440 tree similarity. 2008 International Symposium on Computational Intelligence and Design, 1, 29-32. Lu, J., Zhang, G., Ruan, D. and Wu, F. (2007). Multi-objective group decision making: methods, software and applications with fuzzy set techniques, Imperial College Press, London. Lu J., Shambour, Q., Xu, Y., Lin, Q. and Zhang, G. (2010). A hybrid semantic recommendation system for personalized government-to-business e-services, Internet Research (Acceptance date: 30 January 2010). Rahman, M. & Chow, T.W. (2010). Content-based hierarchical document organization using multi-layer hybrid network and tree-structured features. Expert Systems with Applications, 37(4), 2874-2881. Ricci, F., & Senter, L. (1998). Structured cases, trees and efficient retrieval. In B. Smyth, & P. Cunningham (Eds.), Advances in Case-Based Reasoning: 4th European Workshop, EWCBR-98, Dublin, Ireland, September 23-25, 1998 Proceedings. Lecture Notes in Artificial Intelligence, 1488, 88-99. Springer-Verlag, Berlin Heidelberg New York. Sanders, K.E. Kettler, B.P. & Hendler, J.A. (1997). The case for graph-structured representations. In D.B. Leake, & E. Plaza (Eds.), Case-based Reasoning Research and Development: Second International Conference on Case-Based Reasoning, ICCBR-97 Providence, RI, USA, July 25–27, 1997 Proceedings. Lecture Notes in Artificial Intelligence, 1266, 245-254. Springer Berlin Heidelberg. Torsello, A. Hidovic, D. & Pelillo, M. (2004). Four metrics for efficiently comparing attributed trees. 17th International Conference on Pattern Recognition (ICPR'04), 2, 467-470. Tran, T. Nguyen, C.C. & Hoang, N.M. (2007). Management and analysis of DNA microarray data by using weighted trees. Journal of Global Optimization, 39(4), 623-645. Valiente, G. (2002). Algorithms on trees and graphs, New York: Springer, 16-22, 206-224. Wang, C., Lu, J. and Zhang, G. (2007), Mining key information of web pages: a method and its application, Expert Systems with Applications. Vol. 33, 425-433 Xue, Y. Wang, C. Ghenniwa, H.H. & Shen, W. (2009). A new tree similarity measuring method and its application to ontology comparison. Journal of Universal Computer Science, 15(9), 1766-1781. Yang, L. Sarker, B.K. Bhavsar, V.C. & Boley, H. (2005). A weighted-tree simplicity algorithm for similarity matching of partial product descriptions. Proceedings of ISCA 14th International Conference on Intelligent and Adaptive Systems and Software Engineering, 55-60. Yang, R. Kalnis, P. & Tung, A. (2005). Similarity evaluation on tree-structured data. Proceedings of the 2005 ACM SIGMOD international conference on Management of data, 754-765. Zhang, J., Lu, J. and Zhang, G. (2009), Case-based reasoning in avian influenza risk early warning, The Second Conference on Risk Analysis and Crisis Response (RACR), October, 2009, Beijing, China. Atlantis Press, scientific publishing Paris France, 246-251 Zhang, K. (1993). A new editing based distance between unordered labeled trees. In A. Apostolico et al. (Eds.), Combinatorial Pattern Matching: 4th Annual Symposium, CPM 93, Padova, Italy, June 2-4, 1993 Proceedings. Lecture Notes in Computer Science, 684, 254-265. Springer-Verlag London, UK. 
work_isdelkci4ravthy6kted5tfk7a	----	Hindawi Publishing Corporation Journal of Sensors Volume 2013, Article ID 254629, 11 pages http://dx.doi.org/10.1155/2013/254629 Research Article Supervised Expert System for Wearable MEMS Accelerometer-Based Fall Detector Gabriele Rescio, Alessandro Leone, and Pietro Siciliano Institute for Microelectronics and Microsystems, Italian National Research Council (CNR), Via Monteroni, c/o Campus Università del Salento, Palazzina A3, 73100 Lecce, Italy Correspondence should be addressed to Gabriele Rescio; gabriele.rescio@le.imm.cnr.it Received 8 February 2013; Revised 10 June 2013; Accepted 11 June 2013 Academic Editor: Andrea Cusano Copyright © 2013 Gabriele Rescio et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Falling is one of the main causes of trauma, disability, and death among older people. Inertial sensors-based devices are able to detect falls in controlled environments. Often this kind of solution presents poor performances in real conditions. The aim of this work is the development of a computationally low-cost algorithm for feature extraction and the implementation of a machine- learning scheme for people fall detection, by using a triaxial MEMS wearable wireless accelerometer. The proposed approach allows to generalize the detection of fall events in several practical conditions. It appears invariant to the age, weight, height of people, and to the relative positioning area (even in the upper part of the waist), overcoming the drawbacks of well-known threshold-based approaches in which several parameters need to be manually estimated according to the specific features of the end user. In order to limit the workload, the specific study on posture analysis has been avoided, and a polynomial kernel function is used while maintaining high performances in terms of specificity and sensitivity. The supervised clustering step is achieved by implementing an one-class support vector machine classifier in a stand-alone PC. 1. Introduction The problem of falls in the elderly has become a health care priority due to the related high social and economic costs [1]. In fact the European population aged 65 years or more, which may be in need of assistance is increasing. This trend asks care-holders institutions to employ more efficient and optimized methods in order to be able to grant the required service at lower costs. The consequences of falls in the elderly may lead to psychological trauma, physical injuries, hospi- talization, and even death in the worst scenario [2–5]. The main reason that pushed for the development of the presented system is to allow noncompletely self-sufficient people (e.g., older people) to live safely in their own houses as long as possible. This is important not only for aspects of health regarding assisted people, but also for the consequent social advantages. The European community issued and funded various projects and consortia. The mission focuses on sev- eral purposes, all addressed to older people, varying from the assistance in case of need, to the prevention of dangerous or unhealthy situations. The purpose of the work described in this paper is to focus on people fall detection. Many solutions have been proposed in the detection and prevention of falls, and some excellent review studies were presented [1, 6]. Basically, fall-detection solutions can be classified in three main classes: wearable devices, ambi- ent devices, and camera-based devices. The first approach requires that the elderly holds some kind of devices (e.g., an assistive cane) or wears sensors like accelerometers and/or gyroscopes to detect the motion of the body. In partic- ular, recent miniaturization and cost reduction of MEMS accelerometers and the availability of reliable wireless com- munication technologies enabled the realization of affordable wearable monitoring systems that can be worn by people performing their normal daily activities [7–10]. For these reasons, in the last few years, the use of portable devices in the health monitoring of chronic patients has increased con- siderably. However, these devices have some drawbacks: they are prone to be forgotten, worn in a wrong body position, or accidentally damaged. Regarding fall detection, with respect to vision or acoustic sensors, the accelerometer module has the advantage of not having to be set up and installed in all rooms of the “smart home,” as it is required for instance for 3D video trackers or acoustic scene analyzers. On the other hand, 2 Journal of Sensors ST-LIS3LV02DQ triaxial MEMS accelerometer SPI LP-FPGA UART ZigBee 2.0 wireless end device Figure 1: Picture and block diagram of the wireless accelerometer- based device. the camera-based approach is a nonintrusive solution, since the sensor is a camera installed on a wall and is able to detect falls, switching off when the monitored person is no longer alone [11]. On the other hand, camera-based monitoring solutions suffer when large occluding objects (i.e., furniture) obstruct camera’s viewing. Another significant drawback of any camera-based solution is the violation of monitored per- son’s privacy. In this paper a fall detector through a wearable triaxial MEMS accelerometer is presented. The proposed solution overcomes the limitation of well-known threshold-based approaches [12–17] for which the accelerometer-based device need, to be tuned in an appropriate way for each installation (i.e., the parameters setup may be different for different peo- ple). For these reasons, a machine-learning scheme [18] is used, showing high generalization capabilities in the fall detection discrimination process. The expert system uses robust features extracted taking into account important constraints and requirements of mobile solutions (workload). The extracted features are (quasi-)invariant both to specific characteristics of the mounting setup (device on chest, on waist, and on abdomen) and specific characteristics of the end users in terms of age, weight, height, and gender. 2. Materials and Methods 2.1. Triaxial Accelerometer Sensor System. The hardware used is a wearable device composed by commercial discrete cir- cuits, according to the design proposed in [19]. The pic- ture and the logical block diagram are shown in Figure 1. The system integrates an ST LIS3LV02DL tri-axial MEMS accelerometer with digital output, an FPGA for computing functionalities, and a ZigBee module for wireless communi- cation up to 30 m. The power consumption is about 190 mW in streaming mode and 9 mW in idle. The wearable device can operate in streaming (raw data are sent via ZigBee to an external computing platform for data analysis with a 10 Hz frequency) or in standalone modality, by running the threshold-based fall detection implementation on the on-board FPGA (the power consumption is limited since the ZigBee module is activated just when a fall event occurs). The LIS3LV02DL MEMS accelerometer is DC cou- pled, and it responds up to 0 Hz, with 16-bits resolution and a full scale in the range ±2𝑔. Data can be transmitted in hexadecimal format. The sensor measures both static and dynamic accelerations along the 3 axes and allows one to receive information on the 3D spatial relative position (com- pared to the Earth gravity vector) of the person who wears it. In stationary conditions, assuming a particular axis, the component of the acceleration (amplitude 𝐴, rif. (1)) is defined according to the value of the sine of the angle 𝛼 between the considered axis and the horizontal plane, which is perpendicular to the Earth gravity component (𝑔): 𝐴 = 𝑔 sin (𝛼) . (1) In this way, if the accelerometer relative orientation is known, the resulting data can be used to determinate the angle of the user posture respect to the vertical direction. 2.2. Simulated Falls and ADL Tasks. In order to analyse the waveforms along each axis in the presence of falls and other kind of events (Activities of Daily Living, ADLs), a data col- lection has been defined in controlled (simulated) conditions by involving 11 healthy male and 2 healthy female volunteers. The simulated falls were performed by using a crash mat (height 2 cm) and knee/elbow pad protectors, meeting safety and ethical requirements. The range of actors age was 39.3 ± 12.3 years, weighing 73.7 ± 13.4 kg, and a height of 1.76 ± 0.1 m. 450 actions were simulated in which 250 were falls compliant to the specifications proposed by Noury et al. [20]. The following falls have been simulated for study: (1) backward falls ending in the lying position, (2) backward falls with recovery, (3) forward falls ending in lying flat, (4) forward falls with recovery, (5) lateral falls. During the data collection the wearable device was placed with an elastic band in a different positioning area on the upper part of the torso (on the chest, on the waist, and on the abdomen). Some ADLs were simulated in order to evaluate the ability of fall-detection algorithms to discriminate falls from ADLs. The simulated ADL tasks belong to the following categories: (1) walking, (2) sitting down on a chair and then standing up, (3) lying down on a mat (height 33 cm) and then standing up, (4) lying down on a mat (height 2 cm) and then standing up, (5) kneeling on a mat (height 2 cm) and then standing up. Each actor performed more than 15 simulated ADLs for a total of 205. Journal of Sensors 3 Data acquisition Noise filtering Data preprocessing System calibration Figure 2: Logical framework overview. 2.3. Preprocessing and System Calibration. The first four com- putational steps (Figure 2) deal with data acquisition, data preprocessing, noise filtering, and system calibration. The acceleration data on three axes (𝐴 𝑥 , 𝐴 𝑦 , and 𝐴 𝑧 ) is read out from the device worn by a user during the data col- lection. Data was stored into a portable computer and con- verted into gravitational units to represent acceleration data in the range ±2𝑔, in order to make it possible to extract the angle 𝛼 (described previously) and for not having orders of magnitude too different in the features (during the training of the classifier used to detect the fall events). The samples coming from the device are filtered out by a low pass 8 order, 8 Hz cut-off FIR (Finite Impulse Response) filter to reduce the noise due to electronic components, environment, and human tremor. In order to correctly handle preprocessed data, a cali- bration procedure was accomplished by recovering the ini- tial conditions after the device mounting. During this step the correct placing of hardware is verified by checking if two acceleration axes are orthogonal to the Earth gravity 𝑔 (Figure 3): the acceleration values measured on the two orthogonal components must be close to zero. The calibration procedure is composed by the following steps: (i) the user wears the device in a standing position for 10 seconds; (ii) the calibration routine calculates the average of the acceleration on each axis over this period. These are the initial acceleration values 𝐴 𝑥0 , 𝐴 𝑦0 , and 𝐴 𝑧0 ; (iii) if 𝐴 𝑥0 , 𝐴 𝑦0 , and 𝐴 𝑧0 will be close to those expected, they will be considered as references in the fall detec- tor algorithm, otherwise a routine to compensate the sensed misplacement will be enabled. With respect to Figure 3, the acceleration along 𝑥-component will be close to −1 and almost zero for the others according to (1). The expected values of acceleration on three axes 𝐴 𝑥0 , 𝐴 𝑦0 , and 𝐴 𝑧0 are also reported in Table 1. In practice, it is difficult to locate the device exactly in the right position, and the measured acceleration values will differ slightly from those expected. The calibration procedure is concluded if the following conditions are satisfied at the x y z Figure 3: Mounting position and calibration. Table 1: Acceleration values along the three axes during the initial position of Figure 3. 𝐴 ref 𝑥0 𝐴 ref 𝑦0 𝐴 ref 𝑧0 Initial position −1 0 0 same time (and the values 𝐴 𝑥0 , 𝐴 𝑦0 , and 𝐴 𝑧0 will be recorded and used as references in the feature extraction phase): 󵄨 󵄨 󵄨 󵄨 󵄨 𝐴 𝑥0 − 𝐴 ref 𝑥0 󵄨 󵄨 󵄨 󵄨 󵄨 < 0.3, 󵄨 󵄨 󵄨 󵄨 󵄨 𝐴 𝑦0 − 𝐴 ref 𝑦0 󵄨 󵄨 󵄨 󵄨 󵄨 < 0.3, 󵄨 󵄨 󵄨 󵄨 󵄨 𝐴 𝑧0 − 𝐴 ref 𝑧0 󵄨 󵄨 󵄨 󵄨 󵄨 < 0.3. (2) Since the values (𝐴ref 𝑥0 , 𝐴ref 𝑦0 , and 𝐴ref 𝑧0 ) are (−1, 0, and 0); 󵄨 󵄨 󵄨 󵄨 𝐴 𝑥0 + 1 󵄨 󵄨 󵄨 󵄨 < 0.3, 󵄨 󵄨 󵄨 󵄨 󵄨 𝐴 𝑦0 󵄨 󵄨 󵄨 󵄨 󵄨 < 0.3, 󵄨 󵄨 󵄨 󵄨 𝐴 𝑧0 󵄨 󵄨 󵄨 󵄨 < 0.3. (3) Otherwise a routine to compensate the sensed misplacement is enabled, and the angles displacements of the sensor axes (𝛼𝐴 𝑥0 , 𝛼𝐴 𝑦0 , and 𝛼𝐴 𝑧0 ) are calculated using the following trigonometric equation: 𝛼𝐴 𝑥0 = arctan ( 𝐴 𝑥0 √(𝐴 𝑦0 ) 2 + (𝐴 𝑧0 ) 2 ) , 𝛼𝐴 𝑦0 = arctan ( 𝐴 𝑦0 √(𝐴 𝑥0 ) 2 + (𝐴 𝑧0 ) 2 ) , 𝛼𝐴 𝑧0 = arctan ( 𝐴 𝑧0 √(𝐴 𝑥0 ) 2 + (𝐴 𝑦0 ) 2 ) . (4) 4 Journal of Sensors 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 −1 −0.5 0 0.5 1 1.5 2 Time (s) Postfall phase Static acceleration before fall Impact Critical phase Static acceleration after the fall Peak due to the impact X-axis Y-axis Z-axis g (9 .8 1 m /s2 ) Figure 4: Typical acceleration waveform along each axis for a forward fall. These values are stored and will be used for the correction of the misplacement during the feature extraction phase, described below. 2.4. Feature Extraction. Robust features are extracted in the time domain in a 5 sec sliding window by considering both quick and relevant acceleration changing along each axis (due to the fall) and by the change in position registered after the fall. The aim is to produce robust features taking into account all the information for the distinction of falls from other events. It is also important that such features have a low dependence on both the position of the sensor (whether it is placed on the waist, on the chest, or on the abdomen) and the human body characteristics of the user. Moreover, the com- putational cost must be limited for integration on embedded computing solutions. The waveform in Figure 4 represents a typical accelera- tion signal of a forward fall on three axes and all phases are indicated according to the taxonomy proposed in [16]. For the features extraction process, both critical and postfall phases are of interest. In the former, the shock is measured due to the impact toward the ground, and a dynamic acceleration changing is registered. In the latter (the body is already lying on the floor) the static acceleration value records a great change due to the new position of the individual with respect to the calibration phase. The 5 sec sliding window considered for features extrac- tion is split in three parts as follow: (1) from 0 sec to 0.5 sec: the maximum and minimum values of the acceleration are found, and the ampli- tude dynamics Δ𝐴 is calculated. It is proportional to the shock strength; (2) from 0.5 sec to 2 sec: the system is in a transitory regime (samples are discarded); (3) from 2 sec to 5 sec: the static-averaged acceleration value 𝐴 is calculated to evaluate the new user position. Acceleration data is averaged to filter out tremors and the little movements of the end user. During the system calibration phase it is verified that the device is worn correctly (see previous section), and the posi- tion changing is calculated as the difference between the value of the 3D-static acceleration after the fall (in the third part of the sliding window) and the one stored in the calibra- tion phase. This difference, called Changing Position Offset (CPO), has a value in the range of 0 to 2, and it is proportional to the user displacement. In this way, for feature extraction, a study of posture was not made: only the relative varying posture analysis was considered causing a computational cost reduction and improving the robustness of the setup. The feature vector is made up of three parameters (one for each axis), coming from the modulation of the CPO and the dynamic acceleration peak, due to the impact of the fall. If the routine to compensate the device misplacement is enabled, for the CPO calculation, the axes angles of the sensor in the initial condition need to be taken into account (𝛼𝐴 𝑥0 , 𝛼𝐴 𝑦0 , and 𝛼𝐴 𝑧0 , stored during the calibration phase). Using (1), the CPO value for 𝑥-axis is obtained as CPO 𝑥 = 󵄨 󵄨 󵄨 󵄨 󵄨 sin (𝛼𝐴 𝑥 − 𝛼𝐴 𝑥0 ) 󵄨 󵄨 󵄨 󵄨 󵄨 , (5) where 𝛼𝐴 𝑥 is the 𝑥-axis angle of the sensor in the third part of the sliding window; it is calculated considering the static averaged acceleration 𝐴 𝑥 and the same trigonometric equation used in (4) as follows: 𝛼𝐴 𝑥 = arctan ( 𝐴 𝑥 √(𝐴 𝑦 ) 2 + (𝐴 𝑧 ) 2 ) . (6) Journal of Sensors 5 Let 𝐴 0 initial acceleration value measured after the calibration phase Let 𝛼𝐴 0 initial axis angle of the sensor measured after the calibration phase Δ𝐴 → shock dynamics due to the impact calculated in the first part of the sliding window 𝐴 → Static averaged acceleration in the third part of the sliding window If compensation device misplacement procedure is enabled 𝛼𝐴 is calculated → axis angle of the sensor in the third part of the sliding window CPO = 󵄨 󵄨 󵄨 󵄨 󵄨 sin (𝛼𝐴 − 𝛼𝐴 0 ) 󵄨 󵄨 󵄨 󵄨 󵄨 → Changing Position Offset else CPO = 󵄨 󵄨 󵄨 󵄨 󵄨 𝐴 − 𝐴 0 󵄨 󵄨 󵄨 󵄨 󵄨 → Changing Position Offset End if Let Δ𝐴 𝑥 , Δ𝐴 𝑦 , Δ𝐴 𝑧 , CPO 𝑥 , CPO 𝑦 , CPO 𝑧 , Δ𝐴 and CPO calculated for each axis [𝑋, 𝑌, 𝑍] = [Δ𝐴 𝑥 ⋅ CPO 𝑥 , Δ𝐴 𝑦 ⋅ CPO 𝑦 , Δ𝐴 𝑧 ⋅ CPO 𝑧 ] → feature vector input for SVM Pseudocode 1: Feature extraction pseudocode. Position A: right side lateral fall xy z Position B: backward fall x y z Position C: forward fall x y z Position D: sitting x y z Position E: walking x y z Position G: lying x y z Position F: kneeling x y z Figure 5: Example of different spatial configurations of the device during falls and ADLs. The same procedure must be done for the two other axes. Thanks to this routine, the change in the angle position of the user along the three axes is measured. In this way the information coming from the modulation of the CPO remains highly selective for fall detection recognition even if the device is placed incorrectly and the features remain nearly the same. In this way the system could work efficiently reducing the problems of device positioning at the expense of an increase in the computational costs. The process of feature extraction, described previous, is summarized in the pseudocode shown in Pseudocode 1. For the feature extraction, It makes sense to consider the acceleration signal on each axis singularly, because a fall event leads to a change in the value of the static acceleration in at least two of the three acceleration axes (due to the orientation change of sensing axes). This is evident comparing the positions A, B, and C in Figure 5 (postfall phase of the lateral, backward, and forward falls, resp., are depicted) with respect to the initial position (see Figure 3). On the other hand, when a sitting event (D in Figure 5) occurs, the change in static acceleration can be neglected with respect to that of a fall; it is possible to do the same considerations for kneeling and walking (E and F). Instead, the axes orientation changes for a lying event (position G), but the acceleration peak produced is slower and lower than a fall, and through the product between the value of the acceleration peak and CPO it is also possible to discriminate this ADL. Table 2 shows the nominal values of acceleration along the three axes 𝐴 𝑥 , 𝐴 𝑦 , and 𝐴 𝑧 and the value of CPO corresponding to the positions A–G of Figure 5. 6 Journal of Sensors 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 −1.5 −1 −0.5 0 0.5 1 1.5 Time (s) X-axis Y-axis Z-axis Peak due to the impact Impact on the chair Sitting Static acceleration after sitting Static acceleration before sitting g (9 .8 1 m /s2 ) Figure 6: Example acceleration waveforms for a sitting event. 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 −1.5 −1 −0.5 0 0.5 1 1.5 Time (s) X-axis Y-axis Z-axis Lying Static acceleration before lying Static acceleration after lying g (9 .8 1 m /s2 ) Figure 7: Example acceleration waveforms for a lying event. Table 2: CPO coefficients by varying spatial position. 𝐴 𝑥 𝐴 𝑦 𝐴 𝑧 CPO Initial condition −1 0 0 Position A 0 −1 0 (1, 1, 0) Position B 0 0 −1 (1, 0, 1) Position C 0 0 1 (1, 0, 1) Position D −1 0 0 (0, 0, 0) Position E −1 0 0 (0, 0, 0) Position F −1 0 0 (0, 0, 0) Position G 0 0 −1 (1, 0, 1) When the orientation of acceleration does not change, the values of CPO are (close to) zero, and the elements of the features also become (close to) zero. In Figures 6 and 7 the waveforms of sitting and lying events are reported to make a comparison with that of the fall (shown in Figure 4), when the device is worn on the abdomen. The difference among the features of the fall of Figure 4 and the sitting and lying events of Figures 6 and 7 is apparent in Figures 8, 9, and 10. It was actually verified that all the consideration made thus so far remains valid whether the device is placed on the Journal of Sensors 7 0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 Time (s) X-axis Y-axis Z-axis Y(t=i) X(t=i) Z(t=i) g (9 .8 1 m /s2 ) Figure 8: Features extracted of the forward fall in Figure 4. 0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 Time (s) X-axis Y-axis Z-axis Z(t=i) Y(t=i) X(t=i)g (9 .8 1 m /s2 ) Figure 9: Features extracted of the sitting event in Figure 6. waist or on the chest. In Figures 11 and 12 the features of a sequence of falls and daily events, when the device is worn on the waist and on the chest, are shown. So it is also evident that the features obtained discriminate the falls from ADLs when the device is placed in other area of the torso. It is important to highlight that the measurements of the actions as shown in Figures 11 and 12 have been carried out by simulating critical behaviours, that is, sitting with a strong impact on the chair; bending down, lying down, and standing up quickly; moderate backward fall. 2.5. Fall Detection Algorithm. Once features are extracted, the fall events are detected by a one-class support vector machine (OC-SVM). SVM is a robust classification tool (in the presence of outliers too) with a good generalization ability. Furthermore it is less computationally intensive than other algorithms like neural networks [21]. One class SVM divides all samples into an objective field and a nonobjective field and then nonlinearly maps those sample into a higher dimensional features space with some efficient operators, called kernel function. The target is to find 8 Journal of Sensors 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 Time (s) X-axis Y-axis Z-axis X(t=i) Y(t=i) Z(t=i)g (9 .8 1 m /s2 ) Figure 10: Features extracted of the lying event in Figure 7. 0 1 2 3 0 0.5 1 1.5 Time (min) X-axis Y-axis Z-axis Sitting Kneeling Backward fall Bending downLying Forward fall Standing up g (9 .8 1 m /s2 ) Figure 11: Features extracted of falls and ADLs simulated wearing the device on the waist. a sphere that contains most of the normal data such that the corresponding radius 𝑅 can be minimized as follows: min 𝑅2 + 𝐶 𝑛 ∑ 𝑖=1 𝜉 𝑖 , s.t. 󵄩󵄩󵄩 󵄩 𝑎 − V 𝑖 󵄩 󵄩 󵄩 󵄩 2 ≤ 𝑅 2 + 𝜉 𝑖 , 𝜉 𝑖 ≥ 0, (7) where 𝑎 is the centre of the sphere and V 𝑖 a positive sample set. The slack variables 𝜉 𝑖 allow some data points to lie outside the sphere and the parameter 𝐶 controls the trade-off between the volume of the sphere and the number of errors. The objective function is max 𝑛 ∑ 𝑖=1 𝛼 𝑖 ⟨V 𝑖 , V 𝑖 ⟩ − 𝑛 ∑ 𝑖,𝑗=1 𝛼 𝑖 𝛼 𝑗 ⟨V 𝑖 , V 𝑗 ⟩ , s.t. 0 ≤ 𝛼 𝑖 ≤ 𝐶, 𝑛 ∑ 𝑖=1 𝛼 𝑖 = 1. (8) The original data points are first mapped into a feature space, because some data are not spherically distributed in the input Journal of Sensors 9 0 1 2 2.3 0 0.5 1 1.5 Time (min) X-axis Y-axis Z-axis Lying Standing up Sitting Kneeling Bending down Backward fall Forward fall g (9 .8 1 m /s2 ) Figure 12: Features extracted of falls and ADLs simulated wearing the device on the chest. Table 3: Types of tested kernels. Kernel Function Linear 𝐾(V 𝑖 , V 𝑗 ) = 1 + V𝑇 𝑖 V 𝑗 Polynomial 𝐾(V 𝑖 , V 𝑗 ) = (1 + V𝑇 𝑖 V 𝑗 ) 𝑝 Gaussian Radial Basis Function 𝐾(V 𝑖 , V 𝑗 ) = exp (−𝛾 󵄩 󵄩 󵄩 󵄩 󵄩 V 𝑖 − V 𝑗 󵄩 󵄩 󵄩 󵄩 󵄩 2 ) Sigmoid 𝐾(V 𝑖 , V 𝑗 ) = tanh (𝑘V𝑇 𝑖 V 𝑗 − 𝛿) space. For this reason the kernel function 𝑘(⋅, ⋅) is used as follows: max 𝑛 ∑ 𝑖=1 𝛼 𝑖 𝑘 (V 𝑖 , V 𝑖 ) − 𝑛 ∑ 𝑖,𝑗=1 𝛼 𝑖 𝛼 𝑗 𝑘 (V 𝑖 , V 𝑗 ) . (9) The common types of kernel found in the literature are tested (in Table 3 the functions are shown). They were analyzed using the MATLAB tool STPRtool [22], changing the values of their parameters. In the literature the Gaussian Radial Basis Function (GRBF) is often used in fall detectors [23–25]. For the features extracted, GRBF and polynomial kernel functions give the best results in terms of performance, even the polynomial kernel shows a lower computational cost (relative number of support vectors and relative execution time are considered). The comparison with other kernel functions will be shown in the following section. The parameters of kernel were chosen to detect as many falls as possible, and a simple post-processing step has been implemented in order to reduce false negative events. 2.6. Postprocessing. In choosing the parameters of the tested kernel functions, the value of sensitivity was given priority. In this way some nonfall events are detected from SVM. To reduce this problem a simple filtering by voting was added: it detects an event such as a fall if the alarm coming from the SVM classifier remains high a given time 𝑡. With this approach the spikes or temporary anomalies are filtered too. The optimum time 𝑡 for postprocessing is obtained from a parametric analysis. The results are shown in the next para- graph. 3. Results This section will show a discussion about the measurements of the experimental data processed in MATLAB. To validate the implemented algorithm, the dataset previously described has been considered. Since the output of the fall detection step is binary, there are four possible results. (i) True Positive (TP): a fall happens, and the algorithm detects it. (ii) False Positive (FP): a fall does not occur, and the algorithm reveals a fall. (iii) True Negative (TN): a daily event is performed, and the algorithm does not detect it. (iv) False Negative (FN): a fall occurs, but the algorithm does not detect it. The performance of the system, with respect to both different kernel functions of SVM and their parameters, can be evaluated considering the following metrics: sensitivity = TP TP + FN , specificity = TN TN + FP . (10) 10 Journal of Sensors Table 4: Performance by varying the kernel functions. Sensitivity (%) Specificity (%) Relative support vectors Relative execution time Linear (𝐶 = 10) 98.6 52.3 0.51x 0.2x Linear (𝐶 = 100) 97.2 46.4 0.44x 0.2x Polynomial (𝐶 = 20, 𝑃 = 2) 97.5 81.25 0.94x 0.8x Polynomial (𝐶 = 2.8, 𝑃 = 3) 97.7 95.8 1x 1x GRBF (𝐶 = 2.6, 𝛾 = 2) 98.1 83.1 1.1x 1.3x GRBF (𝐶 = 2, 𝛾 = 3) 97.4 95.2 1.3x 1.5x Sigmoid (𝐶 = 1, 𝑘 = 1, 𝛿 = 2) 98.8 36 0.3x 4x Sigmoid (𝐶 = 100, 𝑘 = 5, 𝛿 = 2) 99.1 42 0.4x 4.1x For completeness the results obtained with postprocessing elaboration have been added. The algorithm is tested when the device is placed on the waist, abdomen, or chest. The one class SVM classifier has been trained by using about 40 falls and 50 ADLs belonging to a large dataset in which more than 250 falls and 200 daily events were performed. The remaining 210 falls and 150 ADLs have been used for testing. Table 4 shows the specificity and the sensitivity for the kernel functions already described. The complexity of the various kernel functions can be studied by considering the relative number of vectors required and also the relative execution time [26, 27]. The kernel with the most significant values for the parameters is considered. From Table 4 the polynomial kernel and GRBF kernel give better results. Their capacity to detect a fall is higher (more than 95% for sensitivity and specificity) than others. GRBF (with 𝜎 = 3 and 𝐶 = 2) and the polynomial (with 𝑃 = 3 and 𝐶 = 2.8) provide similar values of specificity and sensitivity, but the last one works faster, and its number of vector is slightly lower, thereby it was chosen for this work. Misclassifications are for falls presenting slow dynamics or falls with partial recovery. It is important to see the result of the linear function: its workload is very low (less support vectors number and very low execution time), but it gives many false positive, and its value of the specificity is low. The function of GRBF is the most common used in the literature for the detection of falls, due to its high ability to separate the classes, even in the presence of many outliers (at the expense of the convergence time). In this work it has been also possible to obtain a high accuracy with the polynomial function thanks to features extracted: they have a high degree of intrinsic separability. The implemented OC-SVM shows improvements in the specificity and sensitivity with respect to a threshold-based approach, and this can be verified in comparison with both the frameworks presented in [17, 28], choosing the algorithms where the same parameters of fall are used (the impact detection and posture monitoring). For the comparison of the fall detection algorithms were used the same hardware Table 5: Comparison of the proposed OC-SVM method and thresh- old-based algorithm. Sensitivity (%) Specificity (%) OC-SVM 97.7 94.8 Threshold based 89.2 85.7 Table 6: Performance after filtering by voting. Polynomial (𝐶 = 2.8, 𝑃 = 3) Sensitivity (%) Specificity (%) 𝑡 = 0.4 s 97.7 94.8 𝑡 = 0.8 s 96.2 97.2 𝑡 = 1.2 s 86.4 98.1 (shown in Figure 1), benchmark dataset and training/test sets described above. The values of thresholds were tuned according to the falls presented in the dataset used to train the OC-SVM scheme. The best results for threshold based algorithm were obtained with Bourke’s algorithm and are compared with the proposed OC-SVM in Table 5: the latter seems more efficient, demonstrating its higher capacity to generalize, detecting a bigger number of falls with low impact magnitude than threshold-based algorithm. However, the computational cost of thresholdbased is lower. Due to (a) the reduction in the number of the features, (b) the low computational cost of extracted features, and (c) the used kernel, the overall system workload implemented is compatible with an integration in embedded low-power solutions (DSP, FPGA, and microcontroller). A reduction of false positive (and so an improving of specificity value), for the polynomial chosen, has been measured with filtering by voting with a temporal window of prefixed dimension (0.8 sec); see Table 6. Finally it was veri- fied that the values of sensitivity and specificity remain almost the same even if the axes of the sensor are placed differently with respect to that shown in Figure 3 (the values of testing dataset have been changed simulating same misplacements of the device). 4. Conclusions The proposed supervised scheme overcomes the limitation of well-known threshold-based approaches in which a heuristic choice of the parameters is accomplished. High performance in controlled conditions (events simulated) in terms of sensitivity and specificity was obtained using only the 20% of dataset for training. Performance metrics of different kernels in one class SVM are compared: best results are obtained with polynomial function and Gaussian Radial Basis Function. The polynomial kernel is used in order to limit the compu- tational workload. A study of posture was not made, but only posture changing analysis was used, limiting the computa- tional cost. Through the calibration step, the approach allows to generalize the detection of fall events leading invariance to physical characteristic of the end users. Future work will be devoted to validate the solution in real conditions, test the Journal of Sensors 11 methodology with a large set of different MEMS accelerome- ters, and port the framework on embedded mobile solutions as FPGA or DSP. References [1] N. Noury, P. Rumeau, A. K. Bourke, G. ÓLaighin, and J. E. Lundy, “A proposal for the classification and evaluation of fall detectors,” IRBM, vol. 29, no. 6, pp. 340–349, 2008. [2] M. C. Chung, K. J. McKee, C. Austin et al., “Posttraumatic stress disorder in older people after a fall,” International Journal of Geriatric Psychiatry, vol. 24, no. 9, pp. 955–964, 2009. [3] S. Sadigh, A. Reimers, R. Andersson, and L. Laflamme, “Falls and fall-related injuries among the elderly: a survey of residen- tial-care facilities in a Swedish municipality,” Journal of Com- munity Health, vol. 29, no. 2, pp. 129–140, 2004. [4] A. Shumway-Cook, A. M. Ciol, J. Hoffman, J. B. Dudgeon, K. Yorkston, and L. Chan, “Falls in the medicare population: inci- dence, associated factors, and on health Care,” Physical Therapy, vol. 89, no. 4, pp. 324–332, 2009. [5] S. Elliott, J. Painter, and S. Hudson, “Living alone and fall risk factors in community-dwelling middle age and older adults,” Journal of Community Health, vol. 34, no. 4, pp. 301–310, 2009. [6] X. Yu, “Approaches and principles of fall detection for elderly and patient,” in Proceedings of the 10th IEEE International Con- ference on E-Health Networking, Applications and Service (HEALTHCOM ’08), pp. 42–47, Singapore, July 2008. [7] D. M. Karantonis, M. R. Narayanan, M. Mathie, N. H. Lovell, and B. G. Celler, “Implementation of a real-time human move- ment classifier using a triaxial accelerometer for ambulatory monitoring,” IEEE Transactions on Information Technology in Biomedicine, vol. 10, no. 1, pp. 156–167, 2006. [8] J. Y. Hwang, J. M. Kang, Y. W. Jang, and H. C. Kim, “Develop- ment of novel algorithm and real-time monitoring ambulatory system using Bluetooth module for fall detection in the elderly,” in Proceedings of the 26th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC ’04), pp. 2204–2207, San Francisco, Calif, USA, September 2004. [9] H. J. Luinge and P. H. Veltink, “Inclination measurement of human movement using a 3-D accelerometer with autocalibra- tion,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 12, no. 1, pp. 112–121, 2004. [10] G. Anania, A. Tognetti, N. Carbonaro et al., “Development of a novel algorithm for human fall detection using wearable sen- sors,” in Proceedings of the IEEE Sensors (SENSORS ’08), pp. 1336–1339, Lecce, Italy, October 2008. [11] A. Leone, G. Diraco, and P. Siciliano, “Detecting falls with 3D range camera in ambient assisted living applications: a prelimi- nary study,” Medical Engineering and Physics, vol. 33, no. 6, pp. 770–781, 2011. [12] J. Chen, K. Kwong, D. Chang, J. Luk, and R. Bajcsy, “Wearable sensors for reliable fall detection,” in Proceedings of the 27th Annual International Conference of the Engineering in Medicine and Biology Society (IEEE-EMBS ’05), pp. 3551–3554, Shanghai, China, September 2005. [13] M. Kangas, A. Konttila, P. Lindgren, I. Winblad, and T. Jämsä, “Comparison of low-complexity fall detection algorithms for body attached accelerometers,” Gait & Posture, vol. 28, no. 2, pp. 285–291, 2008. [14] A. K. Bourke, J. V. O’Brien, and G. M. Lyons, “Evaluation of a threshold-based tri-axial accelerometer fall detection algo- rithm,” Gait & Posture, vol. 26, no. 2, pp. 194–199, 2007. [15] A. K. Bourke, P. Van de Ven, A. E. Chaya, G. M. Olaighin, and J. Nelson, “Testing of a long-term fall detection system incorpo- rated into a custom vest for the elderly,” in Proceedings of the 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC ’08), pp. 2844–2847, Vancouver, Canada, August 2008. [16] A. K. Bourke, P. van de Ven, M. Gamble et al., “Evaluation of waist-mounted tri-axial accelerometer based fall-detection algorithms during scripted and continuous unscripted activi- ties,” Journal of Biomechanics, vol. 43, no. 15, pp. 3051–3057, 2010. [17] F. Bagalà, C. Becker, A. Cappello et al., “Evaluation of accel- erometer-based fall detection algorithms on real-world falls,” PLoS One, vol. 7, Article ID e37062, 2012. [18] S. -H. Liu and W. -C. Cheng, “Fall detection with the support vector machine during scripted and continuous unscripted activities,” Sensor, vol. 12, no. 9, pp. 12301–12316, 2012. [19] A. Leone, G. Diraco, C. Distante et al., “A multi-sensor approach for people fall detection in home environment,” in Proceedings of the Workshop on Multi-Camera and Multi-Modal Sensor Fu- sion Algorithms and Applications (M2SFA2 ’08), pp. 1–12, 2008. [20] N. Noury, A. Fleury, P. Rumeau et al., “Fall detection—princi- ples and methods,” in Proceedings of the 29th Annual Interna- tional Conference of IEEE-EMBS, Engineering in Medicine and Biology Society (EMBC ’07), pp. 1663–1666, Lyon, France, August 2007. [21] L. M. Manevitz and M. Yousef, “One-class SVMs for document classification,” Journal of Machine Learning Research, vol. 2, no. 1, pp. 139–154, 2001. [22] V. Franc and V. Hlavac, “Statistical Pattern Recognition Toolbox (STPRtool)”. [23] T. Zhang, J. Wang, L. Xu, and P. Liu, “Fall detection by wearable sensor and one-class SVM algorithm,” Lecture Notes in Control and Information Sciences, vol. 345, pp. 858–863, 2006. [24] J. Yin, Q. Yang, and J. J. Pan, “Sensor-based abnormal human- activity detection,” IEEE Transactions on Knowledge and Data Engineering, vol. 20, no. 8, pp. 1082–1090, 2008. [25] H. Cheng, H. Luo, and F. Zhao, “A fall detection algorithm based on pattern recognition and human posture analysis,” in Pro- ceedings of the IET International Conference on Communication Technology and Application (ICCTA ’11), vol. 2011, pp. 853–857, Beijing, China. [26] R. Sangeetha and B. Kalpana, “A comparative study and choice of an appropriate kernel for support vector machines,” in Infor- mation and Communication Technologies, V. V. Das and R. Vijaykumar, Eds., vol. 101 of Communications in Computer and Information Science, pp. 549–553, Springer, Heidelberg, Ger- many, 2010. [27] R. Sangeetha and B. Kalpana, “Performance evaluation of kernels in multiclass support vector machines,” International Journal of Soft Computing and Engineering, vol. 1, no. 5, pp. 138– 145, 2011. [28] M. Grassi, A. Lombardi, G. Rescio et al., “A hardware-software framework for high-reliability people fall detection,” in Proceed- ings of the IEEE Sensors (SENSORS ’08), pp. 1328–1331, Lecce, Italy, October 2009. International Journal of Aerospace Engineering Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Robotics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Active and Passive Electronic Components Control Science and Engineering Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Rotating Machinery Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Journal of Engineering Volume 2014 Submit your manuscripts at http://www.hindawi.com VLSI Design Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Shock and Vibration Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Civil Engineering Advances in Acoustics and Vibration Advances in Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Electrical and Computer Engineering Journal of Advances in OptoElectronics Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Sensors Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Modelling & Simulation in Engineering Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Chemical Engineering International Journal of Antennas and Propagation International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Navigation and Observation International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Distributed Sensor Networks International Journal of 
work_itak635pcncttnjessmuqdf3fi	----	Swine-Vet: A Web-based Expert System of Swine Disease Diagnosis Procedia Computer Science 63 ( 2015 ) 366 – 375 1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Program Chairs doi: 10.1016/j.procs.2015.08.355 ScienceDirect Available online at www.sciencedirect.com The 5th International Conference on Current and Future Trends of Information and Communication Technologies in Healthcare (ICTH 2015) Swine-Vet : a Web-based Expert System of Swine Disease Diagnosis Chutchada Nusaia,*, Sirisak Cheechangb, Somkid Chaiphechc, and Goragot Thanimkana aDepartment of Computer Science, Rajamangala University of Technology Srivijaya,Nakhon Si Thammarat, Thailand bDepartment of Veterinary Public Health, Rajamangala University of Technology Srivijaya,Nakhon Si Thammarat, Thailand cDepartment of Animal Science, Rajamangala University of Technology Srivijaya,Nakhon Si Thammarat, Thailand Abstract A web-based expert system of swine disease diagnosis was developed for swine farmers and animal husbandmen. Our expert system was divided into three steps. First step was disease screening. We established the novel model of knowledge representation for inference using swine's gender and age range which defined by the veterinarian. Second step was disease diagnosis using the symptoms. To make a diagnosis using symptoms which are accurate and efficient, we established the novel model of uncertain knowledge representation for inference using determination of significant weight of each symptom which defined by the veterinarian and using the certainty factor of occurred symptom, the value was specified by user. Third step was the disease diagnosis using swine necropsy lesion. We established the novel model of knowledge representation for inference using major lesion group which defined by the veterinarian for confirmation of morbidness. From the results of diagnosis by our expert system compared with veterinarian, we found that it could disease screening accurately for 97.50 %, could diagnose by symptom accurately for 92.48% and could diagnose by lesion accurately for 95.62%. And the results of evaluation of satisfaction with Likert-scale by the swine farmers and animal husbandmen were 4.7 and 4.5 respectively. © 2014 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the Program Chairs. Keywords: expert system; swine disease diagnosis; knowledge representation * Corresponding author. Tel.: +66890711348; fax: +6675773131. E-mail address: chutchadanu@gmail.com © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Program Chairs http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.08.355&domain=pdf 367 Chutchada Nusai et al. / Procedia Computer Science 63 ( 2015 ) 366 – 375 1. Introduction Currently in Thailand, there are issues with the disease in swine1 which swine; an economic animal that favourably treats. Therefore, we developed a web-based expert system in swine disease diagnosis for swine farmers and animal husbandmen. Our expert system covers all swine diseases (40 diseases) in Thai language. Previously, there was only one research in expert system of swine disease diagnosis in Thai language was expert system for pre- diagnosis of important swine gastrointestial diseases2. Our expert system is useful to swine farmers and animal husbandmen for the care of ill swine health which the disease diagnosis result, protection and therapy let them troubleshoot. And, also decrease the disease spread in the farm and neighbourhood and to decrease veterinary shortage problem. There are several researches on the expert system for the animal diseases diagnosis2-8. All studies have the same objective was to provide a diagnostic expert system is equivalent to the veterinarian. Therefore, there are researches on establishing the method or technique for developing the system that leads to an efficient expert system. Our expert system of swine disease diagnosis was divided into three steps. First step was disease screening, second step was disease diagnosis using the symptoms that user noticed and found by considering from all ten body systems and third step was disease diagnosis using animal necropsy lesion. First step and second step were suitable to swine farmer while the animal husbandman can use all three steps. First step, disease screening using swine's gender and age range, from the veterinarian's knowledge and experience found that each disease occurs in swine's different gender and age range. Some diseases occur in female only such as Mastitis metritis agalactia (MMA) and some diseases occur in some age range such as Swine pox occurs in sucking and nursing. Therefore, swine's gender and age range can be used for occurred disease screening. In this step, we established the novel model of knowledge representation for inference using swine's gender and age range which defined by the veterinarian for disease screening. Second step, the disease diagnosis using the symptoms, from the veterinarian's knowledge and experience found that various symptoms are important in identifying the disease is not the same and that symptom is important in identifying types of diseases that aren’t same as well. Furthermore, because animals cannot tell symptoms like people, so, users have to specify the symptoms from the observation which may cause specifying incomplete symptom and not confident in specifying occurred symptoms. To make a diagnosis using symptoms which are accurate and efficient, we established the novel model of uncertain knowledge representation for inference using determination of significant weight of each symptom which defined by the veterinarian. In each diagnosis, the user must also specify certainty factor of occurred symptom. Therefore, our expert system must have provided images and descriptions of the symptoms which caused the user to specify the certainty factor correctly. Third step, the disease diagnosis using swine necropsy lesion, the animal necropsy has major lesion group which is important and frequently found in such diseased animals. This could use in identifying the disease. Therefore, we established the novel model of knowledge representation for inference using major lesion group which defined by the veterinarian for confirmation of morbidness. Therefore, our model causes an accurate diagnosis in swine disease diagnosis expert system. The previous research on swine disease diagnosis had not applied swine's gender, a significant weight of each symptom, an occurred symptom certainty factor and major lesion group for the disease diagnosis. Such as expert system for pre-diagnosis of important swine gastrointestial diseases 2 used symptoms, severe disease level, age span and feces for diseases diagnosis. A swine's gender, age range, significant weight of each symptom, an occurred symptom certainty factor and major lesion group were applied in our research. 2. The swine disease knowledge acquisition The collection of veterinarian's knowledge and experience which passed validation and verification was corroborated by the veterinarian team, the Faculty of Veterinary Science, Rajamangala University of Technology Srivijaya. The swine disease knowledge includes 40 swine diseases frequently occurred in Thailand9-11. The disease occurs in males or females or both genders. The diseases occur in different 4 age range such as sucking, nursing, growing, and breeding. The example of swine gender and age range is shown in the Table 1. The symptoms occur in 10 body systems, which are digestive, endocrine, nervous, reproductive, respiratory, urinary, intergumentary, circulatory, muscular and skeletal system. The symptoms that caused various diseases consist of 86 symptoms, the symptoms images of 120, and a significant weight of each symptom indicated the disease. The example of the 368 Chutchada Nusai et al. / Procedia Computer Science 63 ( 2015 ) 366 – 375 significant weight of each symptom is shown in Table 2. The swine necropsy lesions of 183,the lesion images of 196, each disease consists of major lesion group which is important and frequently found for confirmation of morbidness and minor lesion group is less important and rarely found(the example is shown in section 5.3). In addition, the swine disease knowledge includes symptom order, disease prevention, diseased animal care and disease spread prevention. Table 1. The example of swine gender and age range Table 2. The significant weight of each symptom 3. Model of knowledge representation in knowledge base We used rule based knowledge representation in knowledge base for disease diagnosis. The disease diagnosis was divided into three steps. 3.1 First step: disease screening We established the novel model of knowledge representation using swine's gender and age range, which was defined by the veterinarian. Model of knowledge representation for disease screening The knowledge in the form of rule is expressed as equation (1). IF G AND (A1OR A2 OR … OR An) THEN H (1) Where, G is gender A1, A2, ,…, An is age range H is a disease of swine. If user input swine's gender and age range match with the rule, the rule was used for disease screening and the result showed the preliminarily possible diseases. 3.2 Second step: Disease diagnosis using symptom We established the novel model of uncertain knowledge representation using determination of significant weight of each symptom, which was defined by the veterinarian. In each diagnosis, the user must also specify certainty factor of occurred symptom. Model of uncertain knowledge representation for disease diagnosis using symptom The knowledge in the form of rule is expressed as equation (2). IF e1(ω1) AND e2(ω2) AND e3 (ω3) AND … AND en (ωn) THEN H (CF, Min) (2) Where, e1, e2, e3,…, en is the symptoms. ωi (i = 1,2,3,…,n) is a significant weight of symptom to disease diagnosis, 1 1 n i i H is a disease of swine. CF is a certainty factor of rule. Min is the acceptable minimum. CF and Min defined by the veterinarians. Disease Gender Age range -Classical swine fever Male,Female sucking, nursing, growing, breeding -Salmonellosis Male,Female nursing, growing, breeding -Mastitis metritis agalactia (MMA) Female breeding Disease Symptom Body system significant weight factor Classical swine fever cyanosis Circulatory system 0.2 high fever Circulatory system 0.1 watery yellowish diarrhea Digestive System 0.3 convulsion/seizures Nervous system 0.1 reddening of skin Intergumentary system 0.2 depress Nervous system 0.1 369 Chutchada Nusai et al. / Procedia Computer Science 63 ( 2015 ) 366 – 375 For the evidences, E = e1(ω1) AND e2(ω2) AND e3 (ω3) AND… AND en (ωn) The calculation of the certainty factor of E (CF(E)) uses equation (3) CF(E) = n i i 1 ieCF (3) Where, ieCF is the certainty factor of occurred symptom, the value was specified by the user. The range of certainty factor is 0 to 1. The calculation of the certainty factor of occurred disease (CF(H,E)) uses equation (4). CF(H,E) = CF(E) CF (4) If CF(H,E) Min , the rule was used for disease diagnosis and the result showed the possible disease. 3.3 Third step: Disease diagnosis using lesion We established the novel model of knowledge representation using determination of major lesion group and minor lesion group, which was defined by the veterinarian. Model of knowledge representation for disease diagnosis using lesion The knowledge in the form of rule is expressed as equation (5). IF Ls1 AND Ls2 AND Ls3 AND…AND Lsn OR(Lm1OR Lm2 OR Lm3 OR … OR Lmn)THEN H (5) Where, Ls1, Ls2, Ls3,… Lsn is lesion in major lesion group. Lm1,Lm2, Lm3, ….Lmn is lesion in minor lesion group. H is a disease of swine. If user inputted all lesions in major lesion group, the rule was used for confirmation of morbidness. 4. Inference process The swine disease diagnosis used a forward inference. The inference process was divided into three steps. First step was inference in disease screening, second step was inference in the diagnosis using symptom and third step was inference in the diagnosis using lesion are shown in fig.1 First step, the inference began from taking the input swine’s gender and age range. Result of disease screening was preliminarily possible disease. Second step, the inference began from taking the input symptoms and its certainty factor to calculate the certainty factor of the E (CF(E)), and then calculated the certainty factor of occurred disease (CF(H,E)), performed rule selection and made diagnosis respectively. Result of diagnosis was possible swine disease. Third step, the inference began from taking the input lesion and made diagnosis respectively. Result of diagnosis was confirmation of morbidness. 5. Application of knowledge representation model and inference We used the model of knowledge representation and inference in the process of swine disease diagnosis. 5.1 First Step : disease screening Example of disease screening is presented as follows. When a user inputs gender and age range would be: gender is female. age range is breeding. The preliminarily possible diseases consist of 31 diseases. 5.2 Second Step : the disease diagnosis using symptom Example of diagnosis is presented as follows. 370 Chutchada Nusai et al. / Procedia Computer Science 63 ( 2015 ) 366 – 375 Fig.1 Inference process Input symptoms(ei) and its certainty factor (CF(ei)) Calculate the certainty factor of the E (CF(E)) Does the CF(H,E) ≥ Min ? Make diagnosis using symptoms Start End Calculate the certainty factor of occurred disease (CF(H,E)) Possible swine disease Input lesions Make diagnosis using lesion Confirmation of morbidness Do the lesions match with the rule? First step Second step Y Y Y N N Input gender and age range Do the input data match with the rule? Preliminarily possible disease Third step N Y Do the symptoms match with the rule? N disease screening Selected rules 371 Chutchada Nusai et al. / Procedia Computer Science 63 ( 2015 ) 366 – 375 When a user inputs the symptoms and its certainty factor would be: cyanosis. And its certainty factor is CF(e1) = 0.8; high fever. And its certainty factor is CF(e2) = 1; watery, yellowish diarrhea. And its certainty factor is CF(e3) = 1; convulsion/seizures. And its certainty factor is CF(e4) = 1; reddening of skin. And its certainty factor is CF(e5) = 0.9; What disease did the swine have? From our expert system, the input symptoms match with thirteen rules. In this case, we present only three rules (with high CF(H,E) in descending order) as follows. R1: IF cyanosis (ω1 = 0.2) AND high fever (ω2 = 0.1) AND watery, yellowish diarrhea (ω3 = 0.3) AND convulsion/seizures (ω4 = 0.1) AND reddening of skin (ω5 = 0.2) AND depress (ω6 = 0.1) THEN Classical swine fever (1, 0.75) R2: IF cyanosis (ω1 = 0.13) AND high fever (ω2 = 0.14) AND watery, yellowish diarrhea (ω3 = 0.17) AND convulsion/seizures (ω4 = 0.16) AND restlessness (ω5 = 0.1) AND inappetence (ω6 = 0.13) AND hemorrhag (ω7 = 0.17) THEN Salmonellosis (1, 0.7) R3: IF cyanosis (ω1 = 0.15) AND high fever (ω2 = 0.1) AND ataxia (ω3 = 0.17) AND grinding of the teeth (ω4 = 0.18) AND septicemia (ω5 = 0.15) AND peddling (ω6 = 0.12) AND sudden death (ω7 = 0.13) THEN Streptococcal (1, 0.7) Then, calculate the certainty factor of E (CF(E)) . CF(E) = n i i 1 ieCF CF(E)R1 = 11eCF + 22eCF + 33eCF + 44eCF + 55eCF = (0.8×0.2)+ (1×0.1)+ (1×0.3)+(1×0.1)+ (0.9×0.2) = 0.84 CF(E)R2 = 11eCF + 22eCF + 33eCF + 44eCF = (0.8×0.13)+ (1×0.14)+(1×0.17)+(1×0.16) = 0.574 CF(E)R3 = 11eCF + 22eCF = (0.8×0.15) +(1×0.1) = 0.22 And calculate the certainty factor of occurred disease (CF(H,E)). CF(H,E) = CF(E) × CF CF(H,E)R1 = 0.84×1 = 0.84 (CFR1= 1) CF(H,E)R2 = 0.574×1 = 0.574 (CFR2= 1) CF(H,E)R3 = 0.22×1 = 0.22 (CFR3= 1) Does the CF(H,E) Min ? CF(H,E)R1 > MinR1 , ( MinR1 = 0.75) CF(H,E)R2 < MinR2 , ( MinR2 = 0.7) CF(H,E)R3 < MinR3 , (MinR3 = 0.7) So, the rule that CF(H,E) Min is R1 only which is selected rule. In this case, the rule R1 was used for disease diagnosis and the result showed the possible disease is Classical swine fever. 5.3 Third step : the disease diagnosis using lesion For example, we present diagnosis for morbidness confirmation of Classical swine fever. If user inputted all lesions in major lesion group, the rule was used for confirmation of morbidness. 372 Chutchada Nusai et al. / Procedia Computer Science 63 ( 2015 ) 366 – 375 When a user inputs the lesion would be: 1.Haemorrhages in the skin especially body, abdomimal, ears, nose and inner thigh 2.Enlarged/swelling in body lymph nodes, area pharyngeal, submendibular, thorax and abdominal 3.Atrophy of the thymus 4.Button ulcers in the mucosa of cecum and colon 5.Bronchopneumonia necrosis, lung congestion and necrosis 6. Congestive hear and congestive splenomegaly 7. Pinpoint haemorrhages on the kidneys 8.Haemorrhages in body organs especially intestinal mucosa, lymph node, kidney, urinary bladder, lung, larynx, epiglottis, gall bladder and heart 9. Pericarditis and adhesions of lung 10. Pale skin, Mucous membranes of the eyes are pale 11.Diffuse necrotic enterrocolitis 12.Ulcerative proctitis 13. Colonic dilatation which causes abdominal distension 14. Inflammation around the heart In this case, because the user inputted all lesions completely in major lesion group (No.1-9), the rule was used for morbidness confirmation of Classical swine fever. 6. Expert system design and development Our web-based expert system for swine disease diagnosis consisted of the disease diagnosis, searching and showing disease data, searching and showing hospital data, web board, knowledge base, inference engine, data management and database. The architecture of expert system is shown in fig.2. The web-based expert system application was developed by ColdFusion Markup Language(ColdFusion 8), manages database system by SQL server 2008 and tested software by Black-Box testing technique. 6.1 Database The database of swine disease diagnosis expert system consists of 16 tables; gender, age range, gender and age range relate with disease, disease, disease image, symptom group, symptom, symptom image, symptom relate with disease, lesion, lesion image, lesion relate with disease, hospital, question web board, answer web board, and admin. The relationship in the database is shown in Fig.3 Fig. 3. Shows the relationship in the database Fig.2 The architecture of expert system A D M IN PK id us ername pas s word firs tname las tname D IS E A S E PK disease_id name c aus e s yndrome treatment c ontrol t_ value c f_ value D IS E A S E _ IM G PK itemno F K 1 dis eas e_ id fileno filename fileextens ion filedes c ription images name fileattac hdatetime D IS E A S E _ L E S IO N S PK,FK1 disease_id PK,FK2 lesions_id flag_ c onfirm D IS E A S E _ S Y M P T O M PK,FK1 disease_id PK,FK2 symptom_id s ignific ant_ weight H O S P IT A L PK id name addres s dis tric t c ity provinc e zipc ode tel L E S IO N S PK lesions_id les ions _ name les ions _ des c ription L E S IO N S _ IM G PK itemno fileno filename fileextens ion filedes c ription images name fileattac hdatetime S Y M P T O M PK symptom_id F K 1 s ymptomgroup_ id s ymptom_ name s ymptom_ des c ription s ymptom_ des c ription_ c f S Y M P T O M _ IM G PK itemno F K 1 s ymptom_ id fileno filename fileextens ion filedes c ription images name fileattac hdatetime S Y M P T O M G R O U P PK symptomgroup_id s ymptomgroup_ name W E B B O A R D _ A N S PK,FK1 ask_id PK ans_id ans _ detail ans _ by ans _ datetime ans _ email W E B B O A R D _ A S K PK ask_id as k_ topic as k_ detail as k_ by as k_ datetime as k_ group as k_ email G E N D E R PK gender_id gender A G E R A N G E PK age_id age_ range D IS E A S E _ G E N D E R _ A G E PK,FK1 disease_id PK,FK2 age_id PK,FK3 gender_id Database Data Management Admin interface Admin Searching and showing hospital data User User Interface Disease diagnosis Searching and showing disease data Web board Inference engine Knowledge base Knowledge acquisition tool Veterinarian 373 Chutchada Nusai et al. / Procedia Computer Science 63 ( 2015 ) 366 – 375 6.2 Data Management The data management in the database of expert system is a role of admin. The admin interface for data management is shown as fig.4. Fig.4 The admin interface for data management 6.3 Disease diagnosis The disease diagnosis process was divided into three steps. 6.3.1 First step: disease screening The user input swine’s gender and age range. The user interface for input swine’s gender and age range is shown as fig.5. The result of disease screening, the system displayed the preliminarily possible disease which is shown as fig.6. Fig.5 The user interface for input gender and age range Fig.6 The result of disease screening 6.3.2 Second step: the disease diagnosis using symptom The user must specify the swine symptom and certainty factor of occurred symptom in 10 body systems. The user interface for input symptoms and certainty factor of occurred symptom is shown as fig.7. The user could view co-information namely, symptom image, detailed description of symptom and description of certainty factor which is shown as fig.8.The result of disease diagnosis using symptom, the system displayed the probability of morbidness. In this sample, the possible disease namely Classical swine fever of 0.84 (84% ) which is shown as fig.9. Fig.7 The user interface for input symptoms and certainty factor of occurred symptom Fig.8 shows co-information Fig.9 Result of disease diagnosis using symptom 374 Chutchada Nusai et al. / Procedia Computer Science 63 ( 2015 ) 366 – 375 6.3.3 Third step: the disease diagnosis using lesion The user must specify lesion. The user interface for input lesion is shown as fig.10. The user could view co- information namely, lesion images, detailed description of lesion which is shown as fig.11. The result of disease diagnosis using lesion, the system displayed result as confirmation of morbidness /no confirmation of morbidness. In this example, the system displayed result as morbidness confirmation of Classical Swine Fever which is shown in fig.12. Fig.10 The user interface for input lesion Fig.11 shows co-information Fig.12 Result of morbidness confirmation of Classical Swine Fever. 7. System efficiency 7.1 System accuracy We tested the diagnosis accuracy using comparison of obtained results between our expert system and veterinarian's diagnosis. 100% A C Accuracy (6) Where, C is the number of correct answer (expert system answer match with veterinarian’s answer) A is the number of total answer 120 samples were diagnosed by our expert system and 4 veterinarians (which were not veterinarian team collected knowledge). The result showed that our expert system gave the disease screening accuracy of 97.50 % (veterinarian1= 97.50 %,veterinarian2= 100 %, veterinarian3= 95 %, veterinarian4= 97.50%, mean= 97.50% ), the diagnosis by symptom accuracy of 92.48% (veterinarian1= 93.33 %,veterinarian2= 94.16 %, veterinarian3= 90.80 %, veterinarian4= 91.66 %, mean = 92.48% )and the diagnosis by lesion accuracy of 95.62%(veterinarian1= 97.50 %,veterinarian2= 97.50 %, veterinarian3= 92.50 %, veterinarian4= 95.00 %, mean = 95.62%). 7.2 System user’s satisfaction We evaluate satisfaction of the swine farmers and animal husbandmen with Likert-scale(5 points) in 4 aspects ; functional, usability, performance and reliability. The results of evaluation of satisfaction by 100 swine farmers and 20 animal husbandmen are 4.7 and 4.5 respectively. 375 Chutchada Nusai et al. / Procedia Computer Science 63 ( 2015 ) 366 – 375 8. Discussion and conclusion The accuracy of disease diagnosis depends on diagnosis method. Our expert system, in first step, the disease screening, we established the novel model of knowledge representation for inference causes an accurate disease screening which the accuracy depends on the veterinarian team expertise who defined swine's gender and age range that indicates the disease. Second step, the disease diagnosis using the symptoms, we established the novel model of uncertain knowledge representation for inference causes an accurate diagnosis which the accuracy depends on the veterinarian team expertise who defined a significant weight of each symptom that indicates the disease. Furthermore, another important part is the completeness of the occurred symptoms and symptom certainty factor which are specified by the user. From the experiment, the occurred symptoms are incomplete or uncertain, but we can diagnosis through the significant weight and the certainty factor of occurred symptom. In third step, the disease diagnosis using the lesion, we established the novel model of knowledge representation for inference causes an accurate diagnosis which the accuracy depends on the veterinarian team expertise who defined major lesion group for confirmation of morbidness. In experiment, the occurred lesions are incomplete, but we can confirm morbidness through major lesion group. The previous research on swine disease diagnosis had not applied swine's gender, a significant weight of each symptom, an occurred symptom certainty factor and major lesion group for the disease diagnosis. Such as expert system for pre-diagnosis of important swine gastrointestial diseases 2 used symptoms, severe disease level, age span and feces for diseases diagnosis; the accuracy of diagnosis was 75.4%. A swine's gender, age range, significant weight of each symptom, an occurred symptom certainty factor and major lesion group were applied in our research and the result of experiment showed that our system give the disease screening accuracy of 97.50%, the diagnosis by symptom accuracy of 92.48% and give the diagnosis by lesion accuracy of 95.62% that is higher than the previous research results. From the result of diagnosis, our system can disease diagnosis approximate to the disease diagnosis by the veterinarian. And the results of evaluation of satisfaction with Likert-scale by the swine farmers and animal husbandmen were 4.7 and 4.5 respectively. Currently, our expert system has been applied in the animal hospital, Faculty of Veterinary Science, Rajamangala University of Technology Srivijaya and Nakhon Si Thammarat and Songkhla province Livestock Offices. Farmers and animal husbandman can operate the system through the Internet http://www.swinevet.in.th/swineindex.cfm. Acknowledgment The authors would like to thanks for the veterinarian team expertise, Faculty of Veterinary Science, Rajamangala University of Technology Srivijaya. This research was supported by fund of National Research Council of Thailand(NRCT). References 1. Link of swine disease situation in 2015. www.oae.go.th/download/document_tendency/journalofecon2558.pdf 2. Farpailin Maneelertu-dom, Sujate Chaunchom, Neramit Sookmanee, Supaporn Thaipakdee. The development of expert system for pre- diagnosis of important swine gastrointestial diseases in Thailand. Kasetsart Journal 2012; 46:996-1008. 3. V.Carvalho,I.Nass,M.Mollo,V.Massafera. Prediction of the occurrence of lameness in dairy cow using a fuzzy-logic expert system. Agricultural Engineering International;2005. 4. Fu Zetian, Xu Feng, Zhou Yun,Zhang XiaoShuan. Pig-vet : a web-based expert system for pig disease diagnosis. Expert Systems with Applications 2005;29:93-103. 5. Daoliang Li, Zetian Fu, Yanqing Duan. Fish-Expert : a web-based expert system for fish disease diagnosis. Expert Systems with Applications 2002; 23:311-320. 6. Jutarat Tangkittiwat, Nalinpat Porrawatpreyakorn. A model for swine disease analysis using neural network. The Tenth Conference on Computing and Information Technology;2014. 7. Paolo Liberati, Paolo Zappavigza. Improving the automated monitoring of dairy cows by integrating various data acquisition systems. Computers and electronics in agriculture 2009; 68:.62-67. 8. Zhang Yu, Xiao Jian-hua, Fan Fu-xiang, Wang Hong-bin. The expert system of cow disease diagnosis basing on the uncertainty evidence illation. The international Conference on bioinformatics and biomedical engineering ;2010 9. Kijja Urairong. Diagnosis,treatment and control of swine diseases. Kasetsart University. 10. Sukhothai Thammathirat University. Animal Health Science. 11. Surapol chondamrongkul. Economic animal disease. Prince of Songkla Universtiy. 
work_ivgteuwjv5ft3efm7yur3kaso4	----	GA-ANFIS Expert System Prototype for Prediction of Dermatological Diseases Lejla BEGIC FAZLIC a,1, Korana AVDAGIC b, and Samir OMANOVIC a a University of Sarajevo – Faculty of Electrical Engineering b UCSF – School of Pharmacy Abstract. This paper presents novel GA-ANFIS expert system prototype for dermatological disease detection by using dermatological features and diagnoses collected in real conditions. Nine dermatological features are used as inputs to classifiers that are based on Adaptive Neuro- Fuzzy Inference Systems (ANFIS) for the first level of fuzzy model optimization. After that, they are used as inputs in Genetic Algorithm (GA) for the second level of fuzzy model optimization within GA-ANFIS system. GA-ANFIS system performs optimization in two steps. Modelling and validation of the novel GA-ANFIS system approach is performed in MATLAB environment by using validation set of data. Some conclusions concerning the impacts of features on the detection of dermatological diseases were obtained through analysis of the GA-ANFIS. We compared GA-ANFIS and ANFIS results. The results confirmed that the proposed GA-ANFIS model achieved accuracy rates which are higher than the ones we got by ANFIS model. Keywords. Adaptive Neuro-Fuzzy Inference System, Genetic Algorithm, Prediction, Dermatological Diseases Introduction When speaking about dermatologic diseases, diagnosis variety is a complex problem. The diseases observed in this group are Pityriasis Lichenoides, Pityriasis Alba, Lichen Striatus, Lichen Planus and Psoriasis. They all share clinical features of erythema and scaling with only slight differences [7] due to which is usually difficult to identify what type of disease the patient actually has. The fact is that some studies present really interesting research on how to easier identify the disease on the basis of the Artificial Intelligence (AI) techniques. In our previous study [7], we presented approach, based on ANFIS model, for the detection and recognition of different types of dermatologic diseases. We used five ANFIS classifiers in order to detect different types of dermatologic diseases. Each of the ANFIS classifiers was trained so to be more accurate for one, than for the other type of disease class. In the recent work [8] , researches presented application of a particular neuro-fuzzy system, named KERNEL, to the problem of diagnosis differentiating in erythemato-squamous diseases. The research study [1] proposed ANFIS model by combining neural network adaptive capabilities and fuzzy logic qualitative approach. In research paper [4] authors used ANFIS to train the mathematical models where GA was responsible to analyze database and possible correlations with consumption. 1 Tel: +387 33 956 122, lejla.begic@fds.ba Digital Healthcare Empowering Europeans R. Cornet et al. (Eds.) © 2015 European Federation for Medical Informatics (EFMI). This article is published online with Open Access by IOS Press and distributed under the terms of the Creative Commons Attribution Non-Commercial License. doi:10.3233/978-1-61499-512-8-622 622 Authors in research study [9] suggested hybrid GA-ANFIS model in which GA optimizes structure and the number of fuzzy IF-THEN rules by finding best value of sub-clustering method. Modern medicine is looking for a solution in a form of new technology that would help doctors in their work. Usually, biopsy is necessary for the correct and final diagnosis. Patients are evaluated by the physicians in two steps. First, they do clinical inspection of the degree of erythema, scaling and compiling historical data that configure 9 features. In the second step, skin samples are taken to histopatological examination with the use of microscope. A novel GA-ANFIS expert system prototype for prediction of dermatological diseases is presented in our study. The purpose of this study is to reach smaller prediction errors (in regard to ANFIS model) by developing new GA-ANFIS expert system. The effectiveness of the GA-ANFIS system approach is tested and verified by the use of validation data. Modeling and validation of the novel GA-ANFIS system approach is performed in MATLAB environment. The results are compared with ANFIS [7], and they show the effectiveness of the proposed approach. Better precisness leads to better applicable value as a help to a doctor in establishing clinical diagnosis. The paper is set up as follows; in introduction, we described problems when differentiating of dermatological disease diagnosis is concerned, and listed some research studies on different approaches. In section one; we explained the methods used in our research along with the description of ANFIS, genetic algorithm and our novel GA-ANFIS system approach covered in this work. We also described GA- ANFIS system design architecture and explained the steps taken in using GA-ANFIS Matlab user interface. Section two shows the numerical results of training and testing error compared with ANFIS results. Finally, the discussion and conclusion are summarized in section three. 1. Material and Methods Skin is the largest organ in the body so that protecting skin from diseases is important. Dermatology is a branch of medicine dealing with skin, hair, nail and its disease. Diseases observed in this group are Pityriasis Lichenoides, Pityriasis Alba, Lichen Striatus, Lichen Planus and Psoriasis. They all share the clinical features of erythema and scaling with very few differences [2]. Nine input attributes are: scaling, itching, family history, Koebner phenomenon, papules (follicular), head involvement, melanin incontinence, acanthosis and inflammatory mononuclear infiltrate. We used three different statuses for features (3- obviously present, 0-not present, 1-2 relative intermediate presented values). The total number in the base was 320 patients. That group was divided in two parts. In the model developing phase, we used 260 patients (i.e. data), who were used for training, testing and checking. 60 separated patients did not participate in the training and testing, but were applied on the finished model in the application phase (clinical validation). Model passed two validation types. In the developing phase, the method ANFIS itself (which is the core of GA-ANFIS system) divides 260 patients, being taken as a model, to 70% of trained, 15% for controlling overfitting-a structure (checking) and 15% for the model testing. Due to the fact that the model is an integral part of the GA-ANFIS system, we needed an additional validation of the system that goes to the application, which we achieved with the afoeremnetioned separated patients. In this approach is used triple ANFIS method [3] based on Sugeno models using fuzzy logic and fuzzy sets, each with three skin features. L. Begic Fazlic et al. / GA-ANFIS Expert System Prototype 623 Our Sugeno model has three inputs (skin features) and one output (type of diseases), as well as 27 IF-THEN knowledge base rules. The next method used in the research is genetic algorithm (GA) based on evolutionary paradigm: crossover, mutation, population, generation and fitness function. The main function of GA is to share four kinds of ANFIS parameter structure and then optimize fuzzy sets in knowledge rule base resulting in the best prediction of dermatological diseases. 2. Results The GA-ANFIS expert system prototype was developed in MATLAB software package. The process flow chart is shown in Fig.1. Figure 1. Process flow chart GA-ANFIS system algorithm, GP-Grid Partitioning, SC-Subtractive Clustering, TrChk -Training and Checking, CH-Chromosome, FS-Fuzzy structure, PV-Predicted Values, Res- Residuum The size of chromosome structure-fuzzy structure (Fig. 2) was generated by ANFIS computational complexity [5] (see Table 1). Chromosome length is defined by Eq.(1) Table 1. ANFIS computational complexity. For NumIn=3 (input variables) and NumMf=3 (number of membership functions) we have concrete chromosome structure (see Fig. 2). Figure 2. Chromosome structure GA-ANFIS system L. Begic Fazlic et al. / GA-ANFIS Expert System Prototype624 LChrom = L1 + L4 = 3*(NumMf*NumIn) +(NumIn+1)*NumMf NumIn (1) The Graphical User Interface (GUI) (Fig. 3) contains four parts. The first part (left upper) shows initialization and pre-processing dermatology data. The second part (left lower) is used for interactive training (learning) ANFIS models. The third part is aimed for initialization and parameterization of GA. The last part integrates all previously data and functions into integral GA-ANFIS operations with diseases score and errors predictions. We made validation of prototype through four experiments. In each experiment were 20 individuals for the GA initial random populations, 200 generation for the GA and fitness limit 0.015.The prediction accuracy for the training data was used as the fitness criterion in the evaluation function. Finally, Fig. 4 below shows the comparison of the predicted value versus target value of the test data. Fig. 5 shows one of hyper planes which represent part of multi-modal transformation (knowledge rules) with the possibility to map all disorder type to given input dermatological features. Figure 4. Real vs. predicted values. Figure 3. Screen shot of Graphical User Interface Figure 5. Surface of predicted disorders Table 2 shows the comparison results of the test success of GA-ANFIS for different optimization method versus one used in previous work [7]. Table 2. Test success of GA-ANFIS system vs. ANFIS in previous work [7] L. Begic Fazlic et al. / GA-ANFIS Expert System Prototype 625 3. Discussion, conclusion and future work The comparison results show that our expert system prototype has better performances than the previous classical approach that is based on single ANFIS model [7]. First of all, it is an approach based on cascade optimization, as we have presented in this work and it gives better results comparing to previous ANFIS approach. The results we have achieved by this novel approach are more useful because we got better prediction by knowing only clinical features. It is important, because patient is usually not medicated before obtaining histopathological results. The proposed GA- ANFIS expert system combines ANFIS and GA capabilities. The total classification accuracy of our GA-ANFIS model was about 98%. This improving of 98% referrs to our system that was optimized for ANFIS level. Originality is in the approach that uses global optimization of the genetic algorithm, so to find ANFIS structure whose parameters give the best preciseness, from the aspect of model accuracy. If it had worked without GA, finding of ANFIS optimal structure would have taken very long time, and we would not have objective results (because we did not do thorough research, as GA did). We have therefore concluded that the proposed model can be used in detecting classes of dermatological diseases by taking into consideration only clinical features. In our approach we use two-stage cascade optimization, where eight types of ANFIS were parallel processed within GA algorithm. The next phase of our research will be focused on fusion GA-ANFIS model with chromosome configuration composed of three parts: four ANFIS parameters, L1 and L4 parameters as in presented approach. References [1] E. Derya Ubeyli, I. Guler, Automatic detection of erythemato-squamous diseases using adaptive neuro fuzzy inference systems, Elsevier -Computers in Biology and Medicine 35 (2005) 421–433. [2] J.-S.R. Jang, ANFIS: Adaptive-network-based fuzzy inference system, IEEE Trans. Syst. Man Cybern. 23 (3), (1993) pp. 665–685. [3] J.-S.R. Jang, Self-learning fuzzy controllers based on temporal backpropagation, IEEE Trans. Neural Networks 3 (5), (1992) pp. 714–723. [4] K.Kampouropoulus, F.Andrade, A. Garcia, L. Romeral, A Combined Methodology of Adaptive Neuro- Fuzzy Inference System and Genetic Algorithm for Short-term Energy Forecasting, Advances in Electrical and Computer Engineering (2014), Volume 14, Issue 1, page(s): 9 – 14, ISSN: 1582-7445, e- ISSN: 1844-7600, DOI: 10.4316/AECE.2014.01002. [5] Pierro Bonissone, Adaptive Neural Fuzzy Inference System, Analysis and Application, http://homepages.rpi.edu/~bonisp/fuzzy-course/Papers-pdf/anfis.rpi04.pdf. [6] S.Olusanya Olatunji, H Arif, Identification of Erythemato Squamous Skin Diseases using Extreme Learning Machine and Artifical Neural Network, Ictact Journal on Soft Computing, 2013,Vol 4, Issue 01. [7] Z. Avdagic, L. B. Fazlic, H. Fatkic, Automatic detection of dermatological diseases using Adaptive Neuro Fuzzy Inference Systems, Acta Informatica Medica;2007, Vol. 15 Issue 2, p89. [8] G. Castellano, C. Castiello, A.M. Fanelli, C. Leone, Diagnosis of dermatological diseases by neuro fuzzy system, Conference: Proceedings of the 3rd Conference of the European Society for Fuzzy Logic and Technology, Zittau, Germany, September 10-12, 2003. [9] M.Zanganeh, S.J. Mousavi, A.E.-Shahidi, Genetic Algorithm based Fuzzy inference system based in prediction of wave parameter, Computational Intelligence, Theory and Application, 2006, Vol. 38, pp. 741-750. L. Begic Fazlic et al. / GA-ANFIS Expert System Prototype626 
work_iwt6segftjgepab2rmdx5ufxwi	----	Microsoft Word - IM_Corrigendum_RetrievingSinusoids_ESWA_bib.docx Corrigendum Corrigendum to “Retrieving sinusoids from nonuniformly sampled data using recursive formulations” Expert Systems with Applications 72 (2017) 245–257 Ivan Maric Rudjer Boskovic Institute, Bijenicka 54, 10000 Zagreb, Croatia Email: ivan.maric@irb.hr Phone: +385-1-4561191 _______________________________________________________________________________________________________ The author regrets to inform that six errors have been discovered in this article. The main error refers to the “log-likelihood” term in equations (4), (5), (34), which is shown as approximated by using the sum squared error (SSE) but in our simulations it was actually approximated by using the square root of the SSE and all the results presented in the paper are obtained in that way. Using the square root of the SSE in the log-likelihood term in the Eq. (5) has the same effect as when increasing its model-complexity penalty term by a factor of 2. A couple of other less significant errors are also corrected. The corrections are outlined below: • The first error is located at page 246, in the right side of Eq. (4), where the sum squared error should be replaced by its square root, and the Eq. (4) should be replaced by: ( ) ( )[ ] 5.0 1 2 ,ln 2 ln       ∑ −−= = K k MkMkM Pw K ,L ββ )) w . • The second error is the same type of error as the first one and is located at page 246, in the Eq. (5), which should be replaced by: ( )[ ] K K k MkMk M MCPw K M +       ∑ −= == 5.0 1 2 , ,...2,1,0 ln 2 minarg β )) . • The third error is located at page 248, in the Section 2.2, in the first item of the numbered list and refers to the text "(see Section 2.3)", which should be replaced by: "(see Section 2.4)". • The fourth error is located at page 248, in the Section 2.2, in the second item of the numbered list and refers to the text "(see Section 2.4)", which should be replaced by: "(see Section 2.5)". • The fifth error is the same type of error as the first one and is located at page 249, in the Eq. 34, which should be replaced by: ( ) 5.0 1 2 ln 2 0       ∑= = K k k w K EDC . • The sixth error is located at page 250, in the Table 1, in the first row below “Initialization” and refers to the text "E0=EDC(0) (34)" which should be replaced by the equation: " ∑= = K k k wE 1 2 0 ". These errors do not affect the correctness of the procedures and the results presented in the paper. The author would like to apologize to the Editor and the readers for these errors. _______________________________________________________________________________________________________ DOI of original article: 10.1016/j.eswa.2016.10.057 E-mail address: ivan.maric@irb.hr http://dx.doi.org/10.1016/j.eswa.2017.04.011 0 
work_ixa54wxsxzhdtm2sicqueh5nam	----	Insight SFI Research Centre for Data Analytics Search Insight Menu About Who we are What we do Our structure People Work With Us Senior leadership Principal Investigators Funded Investigators Research and Operations Research Application Domains Demonstrators Research Challenges Core Scientific Expertise Publications Projects European Funded Projects Business Masterclasses Business Team Public Engagement EPE Committee Citizen Science News Latest News Media Queries Newsletter Spotlight on Research Contact Search Search for: Close search Close Menu AboutShow sub menu Who we are What we do Our structure PeopleShow sub menu Work With Us Senior leadership Principal Investigators Funded Investigators Research and Operations ResearchShow sub menu Application Domains Demonstrators Research Challenges Core Scientific Expertise Publications Projects European Funded Projects BusinessShow sub menu Masterclasses Business Team Public EngagementShow sub menu EPE Committee Citizen Science NewsShow sub menu Latest News Media Queries Newsletter Spotlight on Research Contact Insight Impact Ep.15 Happy Poddy’s Day! Listen to the final episode of The Insight Podcast Read On Insight Careers Insight is hiring! Check out our jobs page Read On Work with us Insight Impact Professor Barry O’Sullivan wins the prestigious Nerode Prize Read On Insight Business Visit our business team and read industry case studies Insight Research Research informed by our vision: Enabling Citizens. Smarter Societies. Insight People Find an expert. Work with us. Public Engagement School visits, events, citizen science, public education and more Latest News ‘People believe VR is all about games!’ Lukasz Porwol talks to Silicon Repu... 24/03/2021 Insight partners with global leader Avaya on Next Gen AI 24/03/2021 Prof Alan Smeaton talks Deepfakes with Newstalk’s Pat Kenny 16/03/2021 Ep.15 Happy Poddy’s Day! Listen to the final episode of The Insight Podcast 16/03/2021 Empowering Citizens Smarter Societies Find out more Insight in Numbers 0 + Researchers € 0 + Million in Funding 0 + Industry Partners 0 Research Institutions Stay up to date with our Newsletter Receive all the latest News & Insights* Leave this field empty if you're human: Privacy Statement Copyright Statement Data Protection Notice 
work_iyfr52ipozd5lnyrow2embtrve	----	1875-0362/20 Send Orders for Reprints to reprints@benthamscience.net 57 DOI: 10.2174/1875036202013010057, 2020, 13, 57-73 The Open Bioinformatics Journal Content list available at: https://openbioinformaticsjournal.com RESEARCH ARTICLE An Expert System to Diagnose Spinal Disorders Seyed M.S. Dashti1,* and Seyedeh F. Dashti2 1Department of Computer Engineering, Kerman Branch, Islamic Azad University, Kerman, Iran 2Department of Advanced Research, Bushehr University of Medical Sciences, Bushehr, Iran Abstract: Objective: Until now, traditional invasive approaches have been the only means being leveraged to diagnose spinal disorders. Traditional manual diagnostics require a high workload, and diagnostic errors are likely to occur due to the prolonged work of physicians. In this research, we develop an expert system based on a hybrid inference algorithm and comprehensive integrated knowledge for assisting the experts in the fast and high-quality diagnosis of spinal disorders. Methods: First, for each spinal anomaly, the accurate and integrated knowledge was acquired from related experts and resources. Second, based on probability distributions and dependencies between symptoms of each anomaly, a unique numerical value known as certainty effect value was assigned to each symptom. Third, a new hybrid inference algorithm was designed to obtain excellent performance, which was an incorporation of the Backward Chaining Inference and Theory of Uncertainty. Results: The proposed expert system was evaluated in two different phases, real-world samples, and medical records evaluation. Evaluations show that in terms of real-world samples analysis, the system achieved excellent accuracy. Application of the system on the sample with anomalies revealed the degree of severity of disorders and the risk of development of abnormalities in unhealthy and healthy patients. In the case of medical records analysis, our expert system proved to have promising performance, which was very close to those of experts. Conclusion: Evaluations suggest that the proposed expert system provides promising performance, helping specialists to validate the accuracy and integrity of their diagnosis. It can also serve as an intelligent educational software for medical students to gain familiarity with spinal disorder diagnosis process, and related symptoms. Keywords: Spine, Expert system, Uncertainty factor, Spinal disorder, Knowledge engineering, Knowledge representation. Article History Received: October 10, 2019 Revised: March 21, 2020 Accepted: April 06, 2020 1. BACKGROUND The development and expansion of information technology have brought about new and essential functions in computer- based decision-making systems. Expert Systems (ES), as an ingredient of artificial intelligence, have a central role in these systems. Decisions in an ES are made via computers. In these systems, experts’ knowledge in a particular field of science is transferred to a computer, as a result of which experts’ way of thinking in that particular area can be emulated. More speci- * Address correspondence to this author at the Department of Computer Engineering, Kerman Branch, Islamic Azad University, Kerman, Iran, E-mail: dashti@iauk.ac.ir fically, the ES detects the logical models through which an expert makes a decision, making judgments similar to those of human beings [1, 2]. One of the specialized topics in physiotherapy and physical education is the detection of abnormalities. Spinal disorders are usually caused by some behaviors, bad habits, or some diseases. The spine, as one of the most vital parts of the human body, continually changes, and in some cases, it may experience anomalies due to its structure and the spinal arrangements [3]. One of the most severe anomalies involves a deformation of the spine and the upper body [4]. The side effects of these anomalies include scoliosis, flat back, round back (thoracic kyphosis), swayback (lumbar lordosis), and cervical lordosis [3]. Nevertheless, no https://openbioinformaticsjournal.com http://crossmark.crossref.org/dialog/?doi=10.2174/1875036202013010057&domain=pdf mailto:dashti@iauk.ac.ir mailto:reprints@benthamscience.net http://dx.doi.org/10.2174/1875036202013010057 58 The Open Bioinformatics Journal, 2020, Volume 13 Dashti and Dashti medical ES has been implemented or designed to detect and prevent spinal disorders. This study suggests an ES designed to simulate the knowledge of experts and helps patients detect and prevent their harmful postural habits that could, in some cases, lead to premature death. A medical ES is composed of some features that distinguish it from other conventional medical systems. One of the significant differences in these systems as far as obtaining results is concerned, is that they can emulate doctors’ reasoning and inference [5, 6]. A medical ES also helps to overcome some of the shortcomings that a human expert may display. The most critical shortcomings include outdated information, possible mistakes, geographical limitations, boredom, lack of quick or complete response time, unclear diagnostic processes, lack of flexibility, low accuracy, slow learning, and lack of varied expertise [5]. Patients at any age may develop spinal disorders. Spine and intervertebral discs are curved but have different structures. In a healthy person, typically, the spine has four curvatures, including cervical lordosis, thoracic kyphosis, lumbar lordosis, and sacral kyphosis [7]. These curves, in the bone and muscle structure of an average person, must be aligned at an acceptable angle. However, if they are deformed or experience abnormalities, they will cause anomalies and defects such as boredom, loss of movement functionality, respiratory restriction, reduced capacity of the heart, reduced blood circulation, and congenital disabilities. Furthermore, due to the alternations imposed on the person’s appearance, she/he may be psychologically vulnerable, while being exposed to the risk of increased injury and damage as a result of pressure engendered by abnormal curvature [8]. In the ES proposed in this study, five anomalies, namely scoliosis, flat back, round back, swayback (lumbar lordosis), and cervical lordosis, are examined. Two different theories form the basis of the proposed system, including backward chaining, the theory of uncertainty; each of them is described separately in the next chapters. The paper is organized as follows: In section 2, the rationale behind this research is discussed. In section 3, elements of inference are explained. Studies in the literature are reviewed in section 4. In section 5, the methodology used in knowledge engineering is discussed. Section 6 describes the system evaluation investigated in this study. Finally, the paper presents the concluding remarks in section 7. 2. RATIONALE Despite the development of medical science, industrial growth, and technology expansion in the modern world, there are still people threatened by new disorders, such as postural anomalies [9]. Spinal disorders which might be detected by the proposed expert system are described as follows: Scoliosis: A sideways curve in the spine is called “scoliosis” [10 - 12]. This anomaly is initially C- shaped but develops into an S-like shape in its advanced forms. Scoliosis can be detected by various methods [10, 11]. The disorder is occasionally very progressive. If the curve increases, due to the pressure it exerts on the lungs, it reduces lung capacity and respiratory functioning, as a result of which some patients experience a slow death [13, 14]. Flat back: Bounded curvature in the spine in the back or the waist, is called “flat back” [15]. In this anomaly, the upper body switches into a vertical position, and mobility in the spine decreases [15]. In the flat back, the cartilage and disc are damaged, and the defensive mechanism against imposed blows and spine resiliency both disappear [10]. Round back: Increased back curvature leads to a condition known as “round back” [16, 17]. Severe round back is of two types, namely the irreversible and the reversible [18, 19]. This anomaly disfigures the shape of the body, causing physiological effects and tightness in the chest. The tightness can disturb the activity of the cardio-pulmonary system [10]. Swayback: An increase in the back curvature is called “swayback”. In this anomaly, the curvature in the lumbar region increases [20]. The stomach area is pulled forward, usually accompanied by pain and fatigue in the back [21, 22]. Swayback leads to pressure on vertebral discs and supposedly engenders back pain, makes childbirth difficult in women, and causes erosion in the bones [10]. Cervical lordosis: Increased cervical curvature leads to a condition called “cervical lordosis.” As a result of this anomaly, the individual experiences a forward head posture, in which and the height of the neck appears to be shorter than its standard height [10]. Failure to detect and treat these anomalies may leave harmful impacts on physiological functions, as in the case of the kyphosis effect, which disturbs the respiratory system [23]. In addition to physical damage, this adverse condition may even bring about psychological and social effects; for instance, kyphosis can be associated with depression [24]. Postural anomalies, apart from hereditary factors, are caused by industrialization, lack of exercise, bad eating habits, and the use of non-standard equipment [25]. In everyday life, an individual may experience different postural positions, some of which can lead to some anomalies in the long run [26]. Additionally, individuals who use pieces of equipment that do not conform to ergonomic standards, and those who ignore anthropometric dimensions are more likely to develop structural and physiological disorders [27]. Recent longitudinal studies show that the most common form of scoliosis is late- onset idiopathic scoliosis [12] that leads to a physical disorder, fatigue, and back pain. If left untreated, this disorder can even contribute to the mortality rate of the public population, because strenuous physical activity causes cardiac arrest over time [13, 14]. There are two significant motivations behind designing and implementing the proposed ES. First, helping experts to diagnose spinal anomalies more accurately; Second, determining the potential risk of development of spinal disorders in healthy people. Even more, since the integrity of knowledge is ensured in the development of the proposed system, it could be easily used by medical students to gain an Diagnose Spinal Disorders The Open Bioinformatics Journal, 2020, Volume 13 59 understanding of the detection of spinal disorders and related symptoms. 3. BASIC ELEMENTS In this section. First, we discuss the technical architecture of expert systems. Next, fundamental theories and inference approaches used in the development of the proposed expert system are thoroughly explained to provide readers with the basics necessary to understand how the proposed approach works. 3.1. Rule-Based Expert System In order to find solutions, conventional computer problem- solving programs use well-structured algorithms, data structures, and crisp reasoning strategies. In the face of the severe problems with which expert systems are concerned, it may be more useful to employ heuristics: strategies that often lead to a correct solution but sometimes fail. Conventional expert systems based on rules use human expertise to solve real-world problems that would typically require human intelligence. Knowledge of experts is often expressed on a computer in the form of rules or data. Based on the problem condition, specific rules and information can be retrieved to solve problems. Rule-based expert systems have played a significant role in strategic goal setting, planning, development, scheduling, fault control, diagnosis, and so on in modern intelligent systems and their implementations [28]. Today's users can choose from hundreds of commercial software packages with friendly graphical user interfaces with the technological advances made in the last decade [29]. Conventional computer programs carry out tasks using a decision-making logic containing very little knowledge other than a basic algorithm to solve this particular problem. Basic knowledge is often incorporated as part of the programming code so that as knowledge changes, the program has to be updated. Small pieces of human knowledge are compiled from knowledge expert systems into a knowledge base; that is used to infer through a problem using appropriate knowledge. A significant benefit here is that using the same program without reprogramming attempts; a particular problem can be solved within the knowledge based domain. Besides, expert systems can explain the rationale process and tackle the level of confidence and ambiguity that traditional algorithms do not manage [30]. Some of the critical advantages of expert systems are as follows: Replication and maintenance of irreplaceable human experience; Ability to deliver a system which is more reliable than human experts in terms of consistency; Minimizing the need for human expert’s presence at several locations at the same time (especially in a dangerous environment that is hazardous to human health); Solutions can be developed much faster when compared to the training procedure of human experts; Figure 1 displays the essential components of the expert system. All relevant information, details, rules, and relations used by the experts are stored in the Knowledge Base. The knowledge base may incorporate many human experts ' expertise [28]. A rule is a conditional statement that links the premises to actions or results. Fig. (1). Architecture of a typical expert system. Knowledge Base Knowledge Acquisition Facility User Interface Explanation FacilityInference Engine Expert with Knowledge User 60 The Open Bioinformatics Journal, 2020, Volume 13 Dashti and Dashti Another method that is used to collect and store information in a knowledge base is the frame. It links an item or entity to a set of facts or values. Frame-based representation of knowledge is a well-suited method for object-oriented programming techniques. Also referred to as are expert systems that use frames to store knowledge in the knowledge base are usually referred to as frame-based expert systems. The inference engine aims to search for knowledge based information and relationships and provide answers, predictions and suggestions in the way a human expert might provide. The inference engine should find and compile the correct facts, definitions, and laws. Two types of inference approaches are widely used– backward chaining is the practice of beginning with hypotheses and moving backward to the facts that support them [31]. Forward chaining begins with the evidence and progresses to the conclusions [32]. The facility of explanation allows a user to understand how the expert system achieved those outcomes. The purpose of the knowledge acquisition facility is to provide an effective and secure means to collect and store all knowledge based components. Expert user interface software is very often used to model, upgrade, and use expert systems. The user interface's purpose is to make the use of an expert system simpler for designers, users, and administrators. 3.2. Inference in Rule-Based expert systems A rule-based expert system is made up of if-then rules a set of facts, and an interpreter controlling the execution of the rules, fed with the facts. Such if-then rule statements are used to formulate the conditional statements which form the full base of knowledge. A single if-then rule presumes the form 'if x is A then y is B' and if-part of rule 'x is A' is referred to as the antecedent or premise while the following part of rule ' y is B' is referred to as the consequence or conclusion for rule-based systems, there are two specific types of inference engines: forward chaining and backward chaining. The initial facts are first processed in a forward chaining system and continue to use the rules to draw new conclusions from those facts. The hypothesis (or solution/goal) we are trying to reach in a backward chaining system is processed first, and we continue to search for rules that would allow us to infer this hypothesis. New sub-goals are also set for testing as the processing progresses. Primarily, forward chaining systems are driven by data, whereas backward chaining systems are goal-driven. In circumstances where information is costly to obtain, forward chaining strategy is particularly appropriate. If there is not enough information about what the result of inference might be, or if there is some particular hypothesis to be validated, forward chaining systems may not be efficient. Backward chaining inference is useful in conditions where the amount of data is potentially very high and where it is of value to have some unique features of the system under consideration [33]. 3.3. Backward Chaining Inference In the case of backward chaining, the primary concern is to align the conclusion of a rule against some known goal. So the' then' (consequent) part of the rule is generally not expressed as an action to be taken but rather as a condition, which is valid if the antecedent part(s) is correct. The backward chaining inference is analogous to the validation testing of the hypothesis in human problem-solving. For instance, a health- care specialist might suspect a patient's problems, which he/she then tries to prove by checking for specific symptoms. This reasoning style is designed by a goal-driven quest in an Expert System and is called backward-chaining [34], [35]. It is a theoretical top-down model that starts with a goal or hypothesis and searches for rules to validate the hypothesis. It tries to balance the variables that lead to relevant data facts and shows that the inference moves backward from the intended goal to establish facts that would fulfill the goal [36]. The implementation of backward chaining in a rule-based expert system is as follows: (1) The system checks the memory to see if the target has been added. Another knowledge base may have already proven the goal, so this phase is required. The algorithm reviews its set of rules, and if the goal has not been proven before, it continues to look for one or more that contains the goal in its THEN portion. This type of rule is called the goal rule. (2) Then the method looks at whether the target rule premises are listed in the memory. If the premises are not specified, new goals or sub-goals are to be checked, and other rules can be used to support them. (3) The system continues in this recursive way until it discovers a premise not provided by any rule; it is called a ' primitive'. The algorithm asks the user for details about it when a primitive is identified. This knowledge is then used by the system to prove both the sub-goals and the original goal. 3.4. Certainty Factor Earlier rule-based expert systems assumed that all current information was either absolutely true or false. However, in the real world, there is often uncertainty associated with the set of rules and the data provided by the user [37]. Expert systems use the Certainty Factor (CF) to implement the theory of uncertainty, which produces a possible output value of certainty. Thus, knowledge engineers consider some degree of uncertainty generated by uncertain occurrences. This consideration is specifically essential in the diagnosis of some diseases with uncertain consequences [38]. The certainty value in this work is within the range of 0 to 100. In the proposed inference algorithm, Certainty Factor (CF)s are assigned both to the premises and antecedents. As the rule's premise is uncertain because of uncertain facts, and the conclusion is uncertain because of the rule's specification, the following formula is used to estimate the conclusion's certainty factor: (1) In some situations, there is more than one rule which supports a given conclusion. In this case, each of the rules can be fired and added to the CF. If a fact supports a conclusion and a rule fires for the favor of that conclusion, then the following equations will be used to evaluate the fact-related current CF: Diagnose Spinal Disorders The Open Bioinformatics Journal, 2020, Volume 13 61 (2) 4. LITERATURE REVIEW Until now, many different approaches have been introduced to analyze spine condition and diagnose anomalies and vertebrae defects. Detection of spinal anomalies is generally possible by non-invasive or invasive methods. Some invasive methods include X-ray images, fluoroscopic, CT, and MRI scans [39]. Non-invasive approaches fall into two categories. The first batch category includes the use of kyphometer, inclinometer, flexible ruler, spinal pantograph, electro-goniometer, spinal mouse. The second category, however, relies on non-contact methods, such as New York test and observation methods [40]. One of the oldest methods, which is considered to be a common practice among professionals, is Cobb's method. In addition to the risks of exposure to X-rays, time-consumption, and financial costs, this method is prone to errors caused by patients’ physical movement during the imaging procedure. This invasive process can also cause bone cancer in men, as well as breast cancer and abortion in women in some cases [41]. The approach proposed by a study [16] is a good example of non-invasive methods. In 2010, a method was introduced for detecting anomalies of the spine, drawing on the image process, and the markers installed on the spinous process [16]. The authors tested this method on forty male students of Birjand University. In this method, markers were installed on the naked body, and images of the individual's movement were recorded through the motion analyzer camera. Comparing their method with a flexible ruler, the researchers observed results that showed a Pearson correlation coefficient of 97% for kyphosis and 95% for lumbar lordosis [2]. This method involves some disadvantages, as it: Requires special equipment including a motion analyzer and markers; Requires the individual to be naked, which may not be the ideal option in all cultural contexts; Requires background and proper lighting; Lacks a specially designed software to be publically available; Involves a high markers installation error, especially when they are set up by non-experts; Has a non-inclusive algorithm, and it needs marker installation; an inclusive and comprehensive algorithm, nevertheless, could function without the use of the motion analyzer device and the markers. However, with the increasing power of computers and the availability of mobile devices to people, new opportunities were created in the field of health and medicine. First, it helped to increase the knowledge of people and communities about health-related issues and keys to a healthy life. Second, new multidisciplinary branches in the field of health and medicine, namely bioinformatics, health informatics emerged. These fields relied on artificial intelligence and mathematical techniques as their basis for computation and reasoning. Until now, many approaches have been introduced to address the challenging task of diagnosing orthopedic diseases, and more specifically, spine disorders and vertebral defects. Methods in this area are roughly divided into two classes: Methods based on machine learning and conventional rule-based approaches. 4.1. Methods Based on Machine Learning The first group of approaches can grasp complex, non- linear relationships in the existing data [42]. Another study [43] proposed a new approach based on random forests to diagnose Osteoarthritis and to provide an easy interpretation of results by using a continuous regression output. In another work [44], authors classified Osteoarthritis subjects using knee joints ' sodium MRIs, which were transformed into radiant 3D acquisitions before and after the fluid suppression technique was applied. Every patient was then defined using 12 main features extracted from target image regions. A probabilistic boosting tree classifier was introduced in order to achieve a reliable segmentation approach to 3D vertebra CT spine images [45]. Information on all vertebras shapes was exploited by using statistical shape modeling techniques that supported the segmentation process. A weighted neighbor distance approach was manipulated in a few studies [46, 47] along with a hierarchy combination of algorithms, which were first introduced in another study [48, 49]. This approach represented morphology and was applied for Osteoarthritis diagnosis in MRIs of articular cartilage scans. In practice, this program was an image classifier that extracted global features from a training set of images, and discriminately assigned weights to them. In a work, an SVM based approach for determining needle entry site was developed. Authors applied midline detection, and template matching approaches to ultrasound spine images; in order to find the best classification features. In this research, more than 1,000 images were analyzed, and the accuracy degree of 95% on the training set and 92% on the test set were respectively obtained. A variation of the SVM model, named Least Squares SVM (LS-SVM), was manipulated by a study [50] in order to distinguish among scoliosis curve types. These curve types were obtained from 3D trunk images by using optical digitizers. When obtained, the 3D images were divided into horizontal slices. Then, all the slices were broken down into patches, and extraction of geometric thoracic and lumbar descriptors was completed, which were later used as classification features after a thorough dimensionality reduction with Principal Component Analysis. In a research [51], a Computer-Aided Diagnosis system was developed to diagnose lumbar inter-vertebral discs degeneration anomaly 62 The Open Bioinformatics Journal, 2020, Volume 13 Dashti and Dashti automatically. The proposed method analyzed and extracted the features from both T1 and T2 weighted MR images in order to acquire different types of feature sets. A generative model was proposed to predict the progression and shape of idiopathic scoliosis affected spines, using 3D spine reconstruction of a set of X-Ray images [52]. The researchers modeled multiple probabilistic structures, a geometric space with adequate structures and transform operators where high-dimensional data are reduced in size while retaining high-dimensional properties. A model based on deep learning and clustering techniques was proposed to detect Adolescent Idiopathic Scoliosis (AIS) automatically [53]. In order to optimize the encoding-decoding of 3D spine model vectors, a neural network with an auto-encoder stack was trained. In two recent studies [54, 55], a neural network was used, which was trained with ultrasound images for automatic detection of optimal vertebra level and Percutaneous spinal needle injection plane. A neural network was designed and trained to partition ultrasound images into a multi-scale patch series recursively was designed [54]. In each iteration, Hadamard coefficients [56] were manipulated to transform standard wave-like ultrasound signatures into signatures of region-orientation correlations to identify distinguishing features of specific spinal patterns. Then, input images were classified for both epidural and facet joint injections as either belonging to or not belonging to the target plane. In a reported work [55], a real- time scanner system was introduced. This scanner was implemented by using a convolutional neural network and a finite state transducer. The convolutional neural network was trained using a transfer learning methodology, where an inter- domain transfer of information was manipulated as a precondition for predicting accurately. Thus, the convolutional neural network was able to detect and assign probabilities to three key positions along with the patient's vertebral scanning – sacrum, intervertebral gaps, and vertebral bones. In a more recent neural-based approach [57], two deep learning pipelines were employed; one for the cervical and the other for lumbar vertebra detection, using annotated clinical MRIs with information labels for each spinal vertebra as training data. Two CNNs were used for each pipeline for the identification of more general and specific vertebral features. Moreover, a component-based graphic model was built based on a layered graph to eliminate false positives and properly mark each vertebra. each layer of this graph shows the previously observed vertebrae and configurations of identified and marked vertebrae and measured the shortest path between them with the distance function based on mean and adjacent covariance matrices. Despite the ability of ML models, this capability is not the complete solution to any inquiry which a healthcare provider may make at the time of treatment. ML algorithms could overfit “predictions in false data correlations”, [[58]] and could identify predictors which “are not causes” despite their predictive power. 4.2. Rule-based Approaches For a long time, rule-based medical systems have been a reliable tool for health practitioners to validate their assumptions [59]. While providing a high degree of accuracy, these type of systems do not rely on any training data. The other advantage of a rule-based medical system, especially the medical rule-based expert system is transparency, being easier to understand the reasoning from which conclusions are derived. Until now, few rule-based works have been reported in the field of orthopedics. Authors believed the classification of cervical spine defects is the basis for surgical therapy and the recommendation for therapy [60]. They reported an automatic rule-based classifier. This classifier required evidence from patient records, converted into a standardized form to facilitate data exchange and multimodal interoperability. A computer-based expert system was introduced to discover ergonomic risks for work-related musculoskeletal disorders. The proposed expert system utilized a knowledge base to classify risk factors into two central knowledge based units, in addition to four secondary knowledge based units that were related to the work environment and organization factors. This work suggested that a knowledge base system may be useful to assess a full-body assessment accurately. In a recent study, an Electronic Health Records (EHR) retrieval system is developed [62]. In particular, a retrieval system was introduced as a model for EHRs to support the medical decision-making process in managing spine defects using the information extracted from textual data on patient records. The patient cases were categorized according to the classes of cervical spine defects, and the classification was based on the rules obtained from the relevant defect classification scheme. The classifier was applied to medical documents in a retrospective study. In another study [63], the authors presented a new 2-d segmentation, classification approach for structural changes in the hippocampal dendritic spines. An interactive rule-based module was leveraged for the classification of spines. Morphological features described essential attributes of the segmented spinal shapes; then, spines were categorized automatically into one of four classes: stubby, filopodia, mushroom, and spine head. A rule-based expert system for diagnosis of neck diseases was reported in a study [64]. The proposed expert system was based on forward chaining and decision trees inference; this system took user response in the form of YES/NO to complete the chain of inference. While lacking technical and scientific rigor, the authors claimed that their proposed approach is helpful in the diagnosis of neck pain diseases. Though many expert systems have been developed to help specialists with spine-related issues until now, no ES has been implemented for diagnosing and estimating the risk of development of spinal abnormalities. 5. KNOWLEDGE ENGINEERING METHOD The implementation of the proposed ES is displayed graphically in Fig. (2). The cycles in Fig. (2) reflect the processes of level 1, and the indices next to them describe the processes of level 2, which are directly related to level 1 processes. In the proposed methodology, each process is repeated for any number of times until the medical expert system is formed in high quality and performance. The methodology is explained in this section. Diagnose Spinal Disorders The Open Bioinformatics Journal, 2020, Volume 13 63 Fig. (2). Phases of knowledge engineering methodology to build the proposed expert system. 5.1. Knowledge Acquisition The necessary resources of knowledge acquisition for system design included the following items: (1) Four orthopaedists; (2) Two physiotherapists; (3) A physical education instructor; (4) 63 articles, 18 reports, and four research projects, including all related signs and symptoms. To design this system, the relevant knowledge for the detection and prevention of any deformities, as well as their symptoms and side effects, were identified by interviewing experts and using questionnaires. Fig. (3) explains the interview process thoroughly; moreover, a filled questionnaire to detect swayback abnormality is represented in Table 1) as a sample. Both of these are crucial parts of the knowledge acquisition process in the early stages. • Feasibility study • Stating the reasons and motives • Goals • Determining resourcesAssessment • Acquisition of knowledge • Inference algorithm design • Designing the pruning method Acquisition of knowledge and system design • Implementation of the inference engine • User interface implementation • Implementation of minor features in expert system Implementation of expert system • Functional testing • User-Friendly testing • Possible Bug Fixing • Expert system Evaluation System testing • System properties description Documentation • Fundamental changes and the growth of knowledge Support 64 The Open Bioinformatics Journal, 2020, Volume 13 Dashti and Dashti Fig. (3). An overview of interview process. Table 1. Sample of the filled questionnaire by an expert in the first review session. Type of Abnormality: Round Back Definition: Round Back is a spinal abnormality in which an excessive outward curve of the spine results in an abnormal rounding of the upper back. Typically, the thoracic spine should have a natural kyphosis between 20 to 45 degrees, while postural or structural abnormalities can result in a curve that is beyond this range. Patients of all ages might be diagnosed with Round Back. Nevertheless, this condition is common during adolescence, when the bones rapidly grow. Diagnosis: 1-Wall Test 2-Shortness of Hamstring Muscle test 3-checkerboard 4-plumb-line 5-Spinal mouse 6-x-rays 7-MRI scan Side Effects: 1- Intervertebral disk herniation 2- Pain in Trapezius muscle and Chest 3-Disk space narrowing 4-Pelvic obliquity 5-Muscle imbalance 6-Fatigue 10- Tear or wear of the lumbar (lower) spine 11-Loss of height 12-Mild to severe back pain 13-Difficulty standing straight upright 14-Limited physical functions 15-Digestive problems 16-Body image problems 17-Breathing problems Causes: 1-Degeneration of adjacent intervertebral discs 2-Overstretched or weak upper back muscles 3-Fractures 4-Osteoporosis 5-Scheuermann's disease 6- Imheritence and birth defects 7-Inheritance 8-Syndromes 9- Cancer and cancer treatments 10-Tuberculosis 11-Paget's disease 12- Spina bifida 13- Muscular dystrophy 14-Neurofibromatosis Next, the experts (orthopaedists) were asked to classify the symptoms based on their importance and rate of occurrence in patients with the spinal anomaly. After the classification of symptoms, experts assigned a unique certainty factor value to each symptom. This value determines the degree of certainty of observing a symptom related to an anomaly in unhealthy patients (Table 2). Finally, the certainty factors assigned by the experts were summed, and the average value was considered to be a single certainty factor. In the third column of Table 3, the final certainty factor can be seen. This factor was derived from the stages mentioned above. Similarly, after classifying the symptoms associated with each other, the certainty effect factor in the fourth column of Table 3 was obtained through maximizing the certainty factor of the symptoms via Formula 3. This factor denotes the CF of the rule premises and serves as a threshold to confirm the observance of symptom in a patient. Fig. (4) schematically illustrates certainty effect values related to flat back anomaly. Session 1 • Experts were asked to fill the questionaries. Highlights of interview session were recorded (a sample of filled questionna shown in table 1). Results • By reviewing the highlights, symptoms of each disorder w discovered. (a sample might be seen in the second column of Tab Session 2 • With regards to dependencies between symptoms, experts asked to assign a certainty factor as a probability value between and one, to each symptom. Results • Average of certainty factor values given by the experts to symptom was calculated and selected as that symptoms' cert factor (column three of Table 2 represents a sample of cert factor value.) Session 3 • In order to ensure no piece of information is missing, experts requested to review the complete list of symptoms related to abnormality. In case of any alterations, the calculations were m from the beginning. Results • The maximum certainty factor value of each symptom was select as the maximum certainty value (column three of table 3). Finally equation 1 was used to calculate the value of the certainty effect factor (column four of table 3). A graphical display of certainty eff values, related to flat back abnormality is shown in Fig 2. Diagnose Spinal Disorders The Open Bioinformatics Journal, 2020, Volume 13 65 Table 2. Sample of symptoms and related factors of flat back abnormality. # Symptoms Certainty from the Perspective of an Expert Probability Cumulative Probability Probability Amendment Symptoms Class 1 Flat thoracic spine (loss of natural curve in the upper back) 80% 0.355 0.355 0.644 A 2 Flat lumbar spine (loss of natural low back curve) 60% 0.266 0.622 0.733 A 3 Reduced flexibility of the spine 30% 0.133 0.755 0.866 B 4 Loss of intervertebral disk space 20% 0.088 0.848 0.911 C 5 Degenerative disk disease 20% 0.088 0.933 0.911 D 6 Compression fractures 10% 0.044 0.977 0.955 E 7 Digestive problems 5% 0.022 1 0.977 F Table 3. Sample of the certainty effect factor calculated using the certainty factor in Table 2 for the flat back anomaly. # Symptoms class Maximum Certainty from the Perspective of an Expert in Class Certainty Effect Uncertainty Effect Cumulative Certainty Effect 1 A 80% 0.484 0.515 0.484 2 B 30% 0.181 0.818 0.666 3 C 20% 0.121 0.878 0.787 4 D 20% 0.121 0.878 0.909 5 E 10% 0.060 0.939 0.969 6 F 5% 0.030 0.969 1 Fig. (4). A sample of certainty effect factor for classified symptoms of the flat back anomaly. (3) 5.2. The Proposed Algorithm of Inference Engine After knowledge acquisition, inference methods and decision-making processes were determined based on the existing knowledge. To emulate experts’ thinking and to avoid the redundant occurrence of the rules, a hybrid algorithm was formulated for inference to obtain an excellent performance in terms of completeness, optimality, time complexity, and space complexity. The algorithm was an incorporation of the BCI [32] and Uncertainty [37]. The most frequently used method of 66 The Open Bioinformatics Journal, 2020, Volume 13 Dashti and Dashti the uncertainty principle in the MESs relies on the certainty factor [38]. However, in the proposed hybrid algorithm, another factor, known as the “certainty effect factor” was used. Applying this factor in the BCI would make it possible for the system to function without storing users’ trail questions and answers. Moreover, in every stage of the decision-making process, one could provide the feedback of statistical results by knowing the previous step of questions and answers. Besides, one could prune the chain through a binary decision tree. Because ESs, especially medical diagnosis systems, are, in most cases, multi-agent, dynamic, inaccessible, uncertain, and continuous, the method proposed in this study drew on the certainty effect factor. This factor helped to simplify the system design by integrating symptoms into some classes and by removing redundant symptoms across the classes. This algorithm frames a “certainty memory” which functions similar to the human thinking process. The certainty memory sorts the rules in descending order according to the value of their symptoms’ certainty effect factor at each moment. The inference process begins when the user specifies the type of abnormality he/she wants to be diagnosed. Next, a question is asked, and the user responds to that question by entering a numerical value within a range of 0 and 100. If the input value is equal or higher than the present certainty effect value of the symptom in certainty memory, it is concluded that examined symptom is observed in the patient with a high degree of certainty. Then, the system will remove the original value of that symptom’s “certainty effect” value. Next, the user input value of that symptom will be replaced as the new “certainty effect factor” while the premise is satisfied. However, if the user response is any value smaller than the existing value, the predefined “certainty effect” value remains unchanged. That is due to the low chances of observing that symptom in the patient, and as a result, that premise will not be satisfied. Afterward, all the satisfied premises are matched with their antecedents, and goal-related rules are fired. As soon as the rules are fired, the certainty factor (CF) of each rule is calculated based on its premises. At each stage, when a new rule is fired, the average sum of certainty factor of rules will be calculated as “certainty degree” of the chain of inference. Equation 4 is used to calculate this factor, which represents the likelihood of existence or risk of developing an anomaly in a person. While delivering a good level accuracy, this approach made efficient reasoning possible. The proposed inference algorithm and approximate responses would require fewer rules than the conventional method of the decision tree and explicit responses. The flowchart of this algorithm can be seen in Fig. (5). Fig. (5). Flowchart of inference engine algorithm. Start Determining the certainty memory Displaying certainty memory to the user User’s response in working memory Update the knowledgebase and certainty memory Is there any rule left End Certainty effect percentage Exit Yes No Determine the goal Run goal selection rule Running next rule in knowledgebase Display certainty memory to user Change in remaining rules’ conditions Preparing knowledgebase and arraigning the rules in descending order according to certainty effect factor Diagnose Spinal Disorders The Open Bioinformatics Journal, 2020, Volume 13 67 (4) The advantages of this inference algorithm are as follows: In the proposed algorithm, the user’s questioning- answering process can be stopped at any point, while reaching an acceptable answer at any stage of the decision-making process, because the events are independent. There is no need to store the path because the certainty memory can compute the relevant statistics at any stage. The possibility to return to previous questions in this algorithm can be accomplished. 5.3. Implementation of Knowledge Base and User Interface After knowledge acquisition was completed, and the inference algorithm was developed, knowledge engineers implemented the knowledge base, using Java Drools Library. On the other side, system designers must realize that ordinary users are not software engineers or knowledge engineers who can work in a complex environment of an ES. There is no doubt that systems with an inconvenient interface cannot be functionally efficient in actual practice. In this study, after the knowledge base was deployed, the graphical interface was designed by JavaFX. It has communicational features such as text-to-speech tool. Fig. (6) illustrates the proposed ES during the questioning-answering process initiated by the user. During the evaluation, all the users could conveniently work with this system. The customized version of the ES is compatible with these operating systems: Windows XP, Windows 7, Windows 8, Windows 10, Windows Server, Windows Vista, Mac, Linux, and Open Solaris. Fig. (6). A sample view of the proposed expert system during question-answering process. 5.4. Diagnosis in Practice The developed expert system relies on the degree of certainty in a series of responses made by the user to determine the order of questions that have to be asked. After each response, user input is taken into account as a certainty amount, and the system state is updated. Then, the system provides the most suitable question to the user. The question answering process ends at the point where no related question is left in the working memory. A flow of questions and answers to diagnose scoliosis abnormality in a 31-year-old female patient is shown in Table 4. Note that the patient was diagnosed with scoliosis abnormality after a close check-up by experts. 68 The Open Bioinformatics Journal, 2020, Volume 13 Dashti and Dashti Table 4. Sample of question answering to diagnose scoliosis abnormality. # Questions User Input/Degree of Certainty 1 What is the patient’s sex? female 2 How old is the patient? 17 3 How tall is the patient? (in cm) 160 4 How much does the patient weight? 65 5 To calculate weekly physical activity, answer following questions: Average heart rate during activity: Average daily physical activity Average weekly physical activity 115 low low 6 Does the patient’s spine have a sideways curve? 89% 7 Does the patient have an “S”- or “C”-shaped spine? 97% 8 Does one shoulder looks higher than the other? 89% 9 Does one shoulder blade stick out more than the other? 90% 10 Does the patient usually carry heavy objects using only one hand? Or play tennis or throw spears? 0% 11 Are the hips uneven? 66% 12 Does the plumb line not pass within 1.7 cm of the posterosuperior corner of the S1 vertebral body ? 88% 13 Is one of the legs longer than the other? 97% 14 Are the abdominal muscles weak? 94% Result: Patient is diagnosed with scoliosis abnormality with certainty degree of 89% 6. RESULTS The developed expert system was evaluated in two separate phases. In the first phase, participants or real-world samples (patients diagnosed with spinal abnormalities and also healthy people) were used to examine the accuracy of the system. Whereas, in the second phase, four orthopaedists evaluated the system using medical records of patients suffering from spinal disorders; based on their knowledge and experience. Both phases are discussed as follows. 6.1. Phase 1: Evaluation Based on Real-world Samples Two random statistical samples (groups of participants) were selected to evaluate the accuracy of the proposed ES results. Four orthopaedists were asked to introduce the ES to the participants, to gain feedback about the functioning of the ES. The first sample included individuals with spinal anomalies, while the second class of samples comprised of people who had a healthy spine. The evaluation process at this stage included predicting the likelihood of existence or chance of development of a spinal disorder in an individual. In doing so, 400 individuals were selected to form the samples. One hundred twenty of them were already diagnosed with one of the spinal anomalies by the experts, while 280 of them seemed to be in reasonable condition. Each expert examined only 30 samples from the group diagnosed with spinal disorders and 70 samples from a healthy group of individuals. Every one of the experts ran the system as a total number of 30 times for unhealthy samples, and five times for each healthy sample in the second group. The reason behind repeating the question- answering process for the healthy samples was to determine the likelihood of the development of any of the five spinal disorders in that patient in the future. This could help the experts to prevent the development of spinal disorders in healthy patients by providing precautionary actions or giving the necessary tips to them. Finally, the system was run for the total number of 1520 executions, and results were gathered. Table 5 shows the statistical information about the participants with anomalies, and Table 6, represents the information related to healthy/normal individuals. X-, S and CV denote mean or average probability, standard deviation and coefficient of variation respectively. As Table 5 represents, evaluation results show a minimum average accuracy value of 0.923 for the correct detection of cervical lordosis in unhealthy patients. Also, the system was successful in true detection of kyphosis disorder with a maximum average accuracy value of 0.940 in samples group suffering from the spinal anomaly. Table 6 demonstrates the mean value of the likelihood of developing any of the five anomalies in healthy samples. Figs. (7 and 8) show the error bar for the detection of anomalies in both healthy and unhealthy groups of participants. Table 5. The results of the expert system in examining the sample with anomalies. # Anomaly Name X- S CV 1 Scoliosis 0.935 0.060 0.065 2 Flat back 0.952 0.085 0.089 3 Kyphosis. 0.946 0.037 0.039 4 Cervical lordosis 0.923 0.034 0.037 5 Swayback 0.930 0.022 0.024 Diagnose Spinal Disorders The Open Bioinformatics Journal, 2020, Volume 13 69 Fig. (7). The error bar, and phase1 evaluation results of the expert system for the samples with anomalies. Fig. (8). The error bar, and phase1 evaluation results of the expert system for the samples without anomaly. 70 The Open Bioinformatics Journal, 2020, Volume 13 Dashti and Dashti Table 6. The results of the expert system in examining the sample without anomalies. # Anomaly Name X- S CV 1 Scoliosis 0.071 0.045 0.634 2 Flat back 0.027 0.022 0.814 3 Kyphosis 0.042 0.022 0.523 4 Cervical lordosis 0.057 0.036 0.631 5 Swayback 0.069 0.045 0.652 6.2. Phase 2: Evaluations Based on Medical Records Four orthopaedists participated in evaluating the system. Experts were provided with medical records of 2000 patients diagnosed with spinal abnormalities in earlier examinations. The records comprised of 400 instances per each anomaly. All the records were segmented and placed under the related abnormality class. For each spinal disorder, 100 cases were selected by each expert. Experts were then asked to review the given medical records. Features of these cases were later fed to ES as inputs through question and answering process. Each orthopaedist ran the system 100 times for each anomaly, as a sum of 400 executions by four experts for each anomaly and total executions of 2000 for all the anomalies. Experts determined to what extent the results of the expert system are accurate in the real world applications, and how properly it can detect an abnormality. As Table 7 and Fig. (9) show, evaluation results from the four experts are very close, with minimum average accuracy value of 0.915 for true detection of cervical lordosis and swayback, and maximum average accuracy of 0.940 for true detection of kyphosis spinal disorder. Table 7. Evaluation results of the system by experts. # Anomaly Name Expert No.1 Expert No.2 Expert No.3 Expert No.4 X'- S CV 1 Scoliosis 0.853 0.948 0.957 0.961 0.925 0.050 0.054 2 Flat back 0.860 0.975 0.933 0.944 0.925 0.046 0.050 3 Kyphosis 0.923 0.911 0.963 0.975 0.940 0.029 0.031 4 Cervical lordosis 0.880 0.939 0.942 0.910 0.915 0.026 0.028 5 Swayback 0.891 0.916 0.911 0.954 0.915 0.025 0.027 Fig. (9). The error bar, and phase2 evaluation results of the system for each anomaly by experts. Diagnose Spinal Disorders The Open Bioinformatics Journal, 2020, Volume 13 71 6.3. Cut-off Values in the Evaluation Procedure When the four experts completed the two phases of evaluation, they were asked to give their feedback about the system performance. Based on their opinions, there are three different situations: first when a sample is correctly diagnosed, second when a sample needs closer examination, third when a given sample is healthy. Then, experts were asked to describe their degree of certainty about system’s accuracy in each of these three cases in terms of numerical values. Table 8 shows a complete list of cut-off values for the detection of spinal abnormalities, which were obtained based on the experts’ feedback. Threshold values were categorized into three classes: Table 8. Cut-off values estimated based on experts’ feed back. # Anomaly Name Expert No.1 Expert No.2 Expert No.3 Expert No.4 Cut-off- True Positive Detection 1 Scoliosis 0.725 0.740 0.770 0.805 0.760 2 Flat back 0.655 0.810 0.770 0.785 0.755 3 Kyphosis 0.790 0.780 0.815 0.835 0.785 4 Cervical lordosis 0.675 0.750 0.810 0.725 0.740 5 Swayback 0.680 0.745 0.735 0.820 0.745 True Negative Detection 1 Scoliosis 0.445 0.500 0.515 0.540 0.500 2 Flat back 0.445 0.530 0.455 0.510 0.485 3 Kyphosis 0.485 0.470 0.550 0.575 0.520 4 Cervical lordosis 0.395 0.495 0.510 0.460 0.465 5 Swayback 0.420 0.475 0.455 0.530 0.470 1-TPD or True Positive Detection: the minimum cut- off value showed that an examined sample is definitely suffering from any type of spinal disorders. 2-TND or True Negative Detection: the maximum cut- off value showed that a given sample is healthy. 3-RFI or Requirement of Further Examinations: any value lower than TPDs or higher than TNDs is regarded as a RFI, suggesting that a sample with diagnosis probability within this span of numbers requires close examinations. For each anomaly, the estimated average of values given by experts was considered as the final threshold value. 7. DISCUSSION Experts participating in the evaluation of the system unanimously agreed over the integrity of the knowledge and validity of assessments. However, the proposed work is still subject to a few limitations. First, while experts found the evaluations to be reliable, they still argued over the degree of correctness of the results. For example, once the expert system detected scoliosis abnormality with a minimum certainty of 76% in a sample, all the four orthopaedists were confident that the examined sample was suffering from scoliosis. Meanwhile, there are situations where the system cannot detect abnormalities with a high degree of accuracy. For instance, in the examination of patients, where scoliosis abnormality had been diagnosed with accuracy degrees between 50% and 75%, at least one out of four experts believed further physical examinations are required before reaching a final judgment. This condition is very similar to the real-world cases where an orthopaedist fails to make a definitive diagnosis and starts further investigations. At this stage, the health specialist may seek or ask for additional details and information. In cases where abnormalities were detected with certainty degree equal or below 50%, all experts agreed that the patient was healthy; nonetheless, that certainty value still reflected the risk of developing anomaly in the examined sample. Knowing that experts are prone to errors, primarily due to prolonged activity, the purpose of this research was to create an expert system in the form of intelligent software to help experts diagnose spinal anomalies more accurately. Evaluations show that the proposed approach was successful in detecting spinal anomalies and to assess the risk of anomaly development. CONCLUSION AND FUTURE STUDIES This study proposed a novel, non-invasive method for detecting and preventing spinal anomalies. The expert system designed and implemented in this research can be used as a reliable assistant by medical experts, specially orthopaedists. This expert system helps specialists to validate the accuracy and integrity of results of spinal disorders diagnosis. It may also serve as an intelligent educational software for medical students to learn the process of spinal disorder diagnosis, and discover the symptoms and variables related to each spinal anomaly. By suggesting an optimal algorithm for the inference engine to emulate experts’ inference, the proposed method could successfully reason based on an optimal number of rules in a flexible question-answering process. It could also bring about advantages in comparison with the malfunctioning of sole use of conditional probabilities. In the conducted evaluation, the system was able to successfully detect anomalies in an individual by using anthropometric dimensions, everyday life habits, anatomy, and physiological data. The extensive use of this expert system by specialists may help a more accurate diagnosis and prevent the spread of spinal anomalies, leading to a healthier life. The present research may be further improved by using fuzzy inference or neural networks. Besides, other factors of anthropometric dimensions 72 The Open Bioinformatics Journal, 2020, Volume 13 Dashti and Dashti of healthy participants, anomalies, and medical information by researchers may be incorporated to yield better results. LIST OF ABBREVIATIONS AIS = Adolescent Idiopathic Scoliosis BCI = Backward Chaining Inference CF = Certainty Factor CNN = Convolutional Neural Network CT = Computed Tomography EHR = Electronic Health Records ES = Expert System LS = Least Squares MES = Medical Expert System ML = Machine Learning MR = Magnetic Resonance MRI = Magnetic Resonance Imaging RFI = Requirement of Further Examinations SVM = Support Vector Machine TPD = True Positive Detection TND = True Negative Detection AUTHOR’S CONTRIBUTION Seyed Mohammad Sadegh Dashti and Seyedeh Fatemeh Dashti conceived the presented idea, accomplished the process of knowledge acquisition, developed the inference algorithm and encoded the knowledge to the knowledge-base, designed and developed the user interface, supervised carrying the experiments out and discussed the results. Seyed Mohammad Sadegh Dashti developed the theoretical formalism and verified the analytical methods. . Seyedeh Fatemeh Dashti performed the analytic calculations and the numerical simulations and wrote the whole manuscript with support from Seyed Mohammad Sadegh Dashti. CONSENT FOR PUBLICATION Not Applicable. FUNDING None. CONFLICTS OF INTERESTS The authors declare that there is no conflict of interest regarding the publication of this article. ACKNOWLEDGEMENTS Authors would like to greatly appreciate Dr. Azita Yazdani from the Department of Health Information Management, Tehran University of Medical Sciences, for providing the medical records without which this research could not be evaluated. REFERENCES Jibril I, Agajo J. Development of a Medical Expert System for[1] Hypertensive Patients Diagnosis: A Knowledge-Based Rules; Advances in Electrical and Telecommunication Engineering. AETE 2018. Chaudhuri SB, Rahman M. Design of a Medical Expert System (MES)[2] Based on Rough Set Theory for Detection of Cardiovascular Diseases. Progress in Advanced Computing and Intelligent Engineering. Advances in Intelligent Systems and Computing 2018. [http://dx.doi.org/10.1007/978-981-10-6872-0_30] Ghasemi G;The effect of eight weeks of NASM exercises on Sway[3] back of high school female students. The Scientific Journal of Rehabilitation Medicine 2018. Heidari Moghaddam R. Reviewing the role of major thalassemia major[4] in spinal abnormality development Reviewing the role of major thalassemia major in spinal abnormality development 2013. Tolouei A. Developing an expert system to detect blood cancer. J[5] Health Manag 2010. Jibril I Z, Agajo J, Ajao L A, Kolo J G, Inalegwu O C. Development[6] of a Medical Expert System for Hypertensive Patients Diagnosis: A Knowledge-Based Rules Advances in Electrical and Telecommunication Engineering 2018; 1 Elaine N. Marieb;Human Anatomy & Physiology. Pearson Education,[7] Inc 2018. Khoubi M. Corrective and therapeutic exercises for treatment of spinal[8] abnormalities. Physical Sciences publication 2015. Keshvari F, Mirdarikvandi M. Applicable Corrective Cxercises 2018.[9] Daneshmandi M, Sedaghati P. New corrective approaches to treatment[10] of Parkinson’s disease. Boshra publications 2017. Machida M. Neurological Research in Idiopathic Scoliosis.[11] Pathogenesis of Idiopathic Scoliosis 2018. [http://dx.doi.org/10.1007/978-4-431-56541-3_7] Mao SH, Qian BP, Shi B, Zhu ZZ, Qiu Y. Quantitative evaluation of[12] the relationship between COMP promoter methylation and the susceptibility and curve progression of adolescent idiopathic scoliosis. Eur Spine J 2018; 27(2): 272-7. [http://dx.doi.org/10.1007/s00586-017-5309-y] [PMID: 28951969] Akazawa T, Kotani T, Sakuma T, et al. Midlife changes of health-[13] related quality of life in adolescent idiopathic scoliosis patients who underwent spinal fusion during adolescence. Eur J Orthop Surg Traumatol 2018; 28(2): 177-81. [http://dx.doi.org/10.1007/s00590-017-2027-4] [PMID: 28798984] Nanjundappa S, Harshavardhana MD. Results of Bracing for Juvenile[14] Idiopathic Scoliosis. Spine Deform 2018. Brumitt J. Core Assessment and Training. Human Kinetics 2010.[15] Ilbeigi S, Mehrshad N, Afzalpour ME, Yousefi M. Diagnosing spinal[16] disorders by using markers mounted on spinous excrescences, sports medicine. 2010. Daneshmandi H, Pourhosseini H. Analyzing spinal disorders among[17] boys and girls Journal of motion 2004. Uei H, Tokuhashi Y, Maseda M, Nakahashi M, Nakayama E. Multiple[18] vertebral fractures associated with glucocorticoid-induced osteoporosis treated with teriparatide followed by kyphosis correction fusion: a case report. Osteoporos Int 2018; 29(5): 1211-5. [http://dx.doi.org/10.1007/s00198-018-4425-9] [PMID: 29476202] Kobets AJ, Komlos D, Houten JK. Congenital cervical kyphosis in an[19] infant with Ehlers-Danlos syndrome. Childs Nerv Syst 2018; 34(7): 1411-5. [http://dx.doi.org/10.1007/s00381-018-3750-9] [PMID: 29450629] Schule r Thomas C. Segmental Lumbar Lordosis: Manual Versus[20] Computer-Assisted Measurement Using Seven Different Techniques. J Spinal Disord Tech 2004. Cressey Eric. Strategies for Correcting Bad Posture – Part 4.[21] EricCressey.com 2014. Robertson PA, Armstrong WA, Woods DL, Rawlinson JJ. Lordosis[22] Re-Creation in TLIF and PLIF: A Cadaveric Study of the Influence of Surgical Bone Resection and Cage Angle. Spine 2018. [http://dx.doi.org/10.1097/BRS.0000000000002705] [PMID: 29697601] Sayari A, Farahani A, Ghanbarzadeh M. Review and Comparison of[23] Structural and Aerobic Corrective Exercises Affecting Pulmonary Function of Students with Kyphosis. Olympics Journal 2007. Ghafouri F. Relation of Kyphosis with Depression and Anxiety[24] Among Athletic and Non-Athlete Male Students in Tehran Universities Journal of Research in Sport Science 2007. MirBagheri R, Ghasemi B. Effect of eight weeks of massage therapy[25] on lumbar lordosis Second Conference on applicable researches in Sport Science 2018. Mohammadi Sh, Mokhtarinia HR. Investigating the effects of different[26] working postures on cognitive performance. Archives of Rehabilitation 2018. http://dx.doi.org/10.1007/978-981-10-6872-0_30 http://dx.doi.org/10.1007/978-4-431-56541-3_7 http://dx.doi.org/10.1007/s00586-017-5309-y http://www.ncbi.nlm.nih.gov/pubmed/28951969 http://dx.doi.org/10.1007/s00590-017-2027-4 http://www.ncbi.nlm.nih.gov/pubmed/28798984 http://dx.doi.org/10.1007/s00198-018-4425-9 http://www.ncbi.nlm.nih.gov/pubmed/29476202 http://dx.doi.org/10.1007/s00381-018-3750-9 http://www.ncbi.nlm.nih.gov/pubmed/29450629 http://dx.doi.org/10.1097/BRS.0000000000002705 http://www.ncbi.nlm.nih.gov/pubmed/29697601 Diagnose Spinal Disorders The Open Bioinformatics Journal, 2020, Volume 13 73 [http://dx.doi.org/10.21859/jrehab.18.4.1] Choubineh A, Moudi A. Human-Design and Ergonomics. Tehran:[27] PhisNet 2015. Smith S, Kandel A. Verification and validation of rule-based expert[28] systems. CRC Press 2018. [http://dx.doi.org/10.1201/9781315214238] Ignizio James P. Introduction to expert systems: the development and[29] implementation of rule-based expert systems. 1991. Giarratano JC. Gary Riley; Expert systems. PWS publishing co 1998.[30] Mohammadi Motlagh HA, Minaei Bidgoli B. Design and[31] implementation of a web-based fuzzy expert system for diagnosing depressive disorder. Appl Intell 2018. Al-Ajlan A. The comparison between forward and backward chaining.[32] Int J Mach Learn Comput 2015. [http://dx.doi.org/10.7763/IJMLC.2015.V5.492] Grosan Crina, Abraham Ajith. Rule-based expert systems In:[33] Intelligent Systems. 2011. [http://dx.doi.org/10.1007/978-3-642-21004-4_7] Caignya A, Coussementa K. A new hybrid classification algorithm for[34] customer churn prediction based on logistic regression and decision trees. Eur J Oper Res 2018. Jose A. A Prasad,Design and Development of a Rule Based Expert[35] System for AACR: A Study of the Application of Artificial Intelligence Techniques in Library and Information Field. Saarbrücken, Germany: VDM Verlag 2011. Sagheb-Tehrani M. Expert systems development: Some issues of[36] design process. Softw Eng Notes 2005; 30(2): 1-5. [http://dx.doi.org/10.1145/1050849.1050864] Shapiro EY. Logic programs with uncertainties: A tool for[37] implementing rule-based systems In: IJCAI. 1983. Baudryab G, Macharis C. Range-based Multi-Actor Multi-Criteria[38] Analysis: A combined method of Multi-Actor Multi-Criteria Analysis and Monte Carlo simulation to support participatory decision making under uncertainty. Eur J Oper Res 2018. Fon GT, Pitt MJ, Thies AC Jr. Thoracic kyphosis: range in normal[39] subjects. AJR Am J Roentgenol 1980; 134(5): 979-83. [http://dx.doi.org/10.2214/ajr.134.5.979] [PMID: 6768276] Mahmoudi F, Shahrjerdi S. Changes in Pain and Kyphosis Angle[40] Following a Corrective Exercise Program in Elderly Women: A Randomized Controlled Trial. JRUMS 2017. Kado DM, Prenovost K, Crandall C. Narrative review: hyperkyphosis[41] in older persons. Ann Intern Med 2007; 147(5): 330-8. [http://dx.doi.org/10.7326/0003-4819-147-5-200709040-00008] [PMID: 17785488] Chen JH, Asch SM. Machine learning and prediction in[42] medicine—beyond the peak of inflated expectations. N Engl J Med 2017. [http://dx.doi.org/10.1056/NEJMp1702071] Kotti M, Duffell LD, Faisal AA, McGregor AH. Detecting knee[43] osteoarthritis and its discriminating parameters using random forests. Med Eng Phys 2017; 43: 19-29. [http://dx.doi.org/10.1016/j.medengphy.2017.02.004] [PMID: 28242181] Madelin G, Poidevin F, Makrymallis A, Regatte RR. Classification of[44] sodium MRI data of cartilage using machine learning. Magn Reson Med 2015; 74(5): 1435-48. [http://dx.doi.org/10.1002/mrm.25515] [PMID: 25367844] Mirzaalian H, Wels M, Heimann T, Kelm BM, Suehling M. Fast and[45] robust 3D vertebra segmentation using statistical shape models In 2013 35th annual international conference of the IEEE engineering in medicine and biology society (EMBC). [http://dx.doi.org/10.1109/EMBC.2013.6610266] Ashinsky BG, Coletta CE, Bouhrara M, et al. Machine learning[46] classification of OARSI-scored human articular cartilage using magnetic resonance imaging. Osteoarthritis Cartilage 2015; 23(10): 1704-12. [http://dx.doi.org/10.1016/j.joca.2015.05.028] [PMID: 26067517] Ashinsky BG, Bouhrara M, Coletta CE, et al. Predicting early[47] symptomatic osteoarthritis in the human knee using machine learning classification of magnetic resonance images from the osteoarthritis initiative. J Orthop Res 2017; 35(10): 2243-50. [http://dx.doi.org/10.1002/jor.23519] [PMID: 28084653] Shamir L, Orlov N, Eckley DM, Macura T, Johnston J, Goldberg IG.[48] Wndchrm - an open source utility for biological image analysis. Source Code Biol Med 2008; 3: 13. [http://dx.doi.org/10.1186/1751-0473-3-13] [PMID: 18611266] Yu S, Tan KK, Sng BL, Li S, Sia AT. Lumbar ultrasound image[49] feature extraction and classification with support vector machine. Ultrasound Med Biol 2015; 41(10): 2677-89. [http://dx.doi.org/10.1016/j.ultrasmedbio.2015.05.015] [PMID: 26119460] Adankon MM, Dansereau J, Labelle H, Cheriet F. Non invasive[50] classification system of scoliosis curve types using least-squares support vector machines. Artif Intell Med 2012; 56(2): 99-107. [http://dx.doi.org/10.1016/j.artmed.2012.07.002] [PMID: 23017984] Oktay AB, Albayrak NB, Akgul YS. Computer aided diagnosis of[51] degenerative intervertebral disc diseases from lumbar MR images. Comput Med Imaging Graph 2014; 38(7): 613-9. [http://dx.doi.org/10.1016/j.compmedimag.2014.04.006] [PMID: 24972858] Kadoury S, Mandel W, Roy-Beaudry M, Nault ML, Parent S. 3-D[52] morphology prediction of progressive spinal deformities from probabilistic modeling of discriminant manifolds. IEEE Trans Med Imaging 2017; 36(5): 1194-204. [http://dx.doi.org/10.1109/TMI.2017.2657225] [PMID: 28129153] Thong W, Parent S, Wu J, Aubin CE, Labelle H, Kadoury S. Three-[53] dimensional morphology study of surgical adolescent idiopathic scoliosis patient from encoded geometric models. Eur Spine J 2016; 25(10): 3104-13. [http://dx.doi.org/10.1007/s00586-016-4426-3] [PMID: 26851954] Pesteie M, Abolmaesumi P, Ashab HA, et al. Real-time ultrasound[54] image classification for spine anesthesia using local directional Hadamard features. Int J CARS 2015; 10(6): 901-12. [http://dx.doi.org/10.1007/s11548-015-1202-5] [PMID: 26026697] Hetherington J, Lessoway V, Gunka V, Abolmaesumi P, Rohling R.[55] SLIDE: automatic spine level identification system using a deep convolutional neural network. Int J CARS 2017; 12(7): 1189-98. [http://dx.doi.org/10.1007/s11548-017-1575-8] [PMID: 28361323] Beauchamp KG. Applications of Walsh and related functions: with an[56] introduction to sequency theory. Academic press 1984. Forsberg D, Sjöblom E, Sunshine JL. Detection and labeling of[57] vertebrae in MR images using deep learning with clinical annotations as training data. J Digit Imaging 2017; 30(4): 406-12. [http://dx.doi.org/10.1007/s10278-017-9945-x] [PMID: 28083827] Obermeyer Z, Emanuel EJ. Predicting the future—big data, machine[58] learning, and clinical medicine. N Engl J Med 2016; 375(13): 1216-9. [http://dx.doi.org/10.1056/NEJMp1606181] [PMID: 27682033] Abu-Nasser B. Medical Expert Systems Survey; International Journal[59] of Engineering and Information Systems. IJEAIS 2017. Deng Y, Groll MJ, Denecke K. Rule-based Cervical Spine Defect[60] Classification Using Medical Narratives. Stud Health Technol Inform 2015; 216: 1038. [PMID: 26262337] Pavlovic-Veselinovic S, Hedge A, Veselinovic M. An ergonomic[61] expert system for risk assessment of work-related musculo-skeletal disorders. Int J Ind Ergon 2016. [http://dx.doi.org/10.1016/j.ergon.2015.11.008] Deng Y, Denecke K. Patient Records Retrieval System for Integrated[62] Care in Treatment of Cervical Spine Defect In VLDB Workshop on Data Management and Analytics for Medicine and Healthcare. Basu S, Plewczynski D, Saha S, et al. 2dSpAn: semiautomated 2-d[63] segmentation, classification and analysis of hippocampal dendritic spine plasticity. Bioinformatics 2016; 32(16): 2490-8. [http://dx.doi.org/10.1093/bioinformatics/btw172] [PMID: 27153678] Naser SS. ALmursheidi SH;A Knowledge Based System for Neck[64] Pain Diagnosis; World Wide Journal of Multidisciplinary Research and Development. WWJMRD 2016. © 2020 Dashti and Dashti. This is an open access article distributed under the terms of the Creative Commons Attribution 4.0 International Public License (CC-BY 4.0), a copy of which is available at: (https://creativecommons.org/licenses/by/4.0/legalcode). This license permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. http://dx.doi.org/10.21859/jrehab.18.4.1 http://dx.doi.org/10.1201/9781315214238 http://dx.doi.org/10.7763/IJMLC.2015.V5.492 http://dx.doi.org/10.1007/978-3-642-21004-4_7 http://dx.doi.org/10.1145/1050849.1050864 http://dx.doi.org/10.2214/ajr.134.5.979 http://www.ncbi.nlm.nih.gov/pubmed/6768276 http://dx.doi.org/10.7326/0003-4819-147-5-200709040-00008 http://www.ncbi.nlm.nih.gov/pubmed/17785488 http://dx.doi.org/10.1056/NEJMp1702071 http://dx.doi.org/10.1016/j.medengphy.2017.02.004 http://www.ncbi.nlm.nih.gov/pubmed/28242181 http://dx.doi.org/10.1002/mrm.25515 http://www.ncbi.nlm.nih.gov/pubmed/25367844 http://dx.doi.org/10.1109/EMBC.2013.6610266 http://dx.doi.org/10.1016/j.joca.2015.05.028 http://www.ncbi.nlm.nih.gov/pubmed/26067517 http://dx.doi.org/10.1002/jor.23519 http://www.ncbi.nlm.nih.gov/pubmed/28084653 http://dx.doi.org/10.1186/1751-0473-3-13 http://www.ncbi.nlm.nih.gov/pubmed/18611266 http://dx.doi.org/10.1016/j.ultrasmedbio.2015.05.015 http://www.ncbi.nlm.nih.gov/pubmed/26119460 http://dx.doi.org/10.1016/j.artmed.2012.07.002 http://www.ncbi.nlm.nih.gov/pubmed/23017984 http://dx.doi.org/10.1016/j.compmedimag.2014.04.006 http://www.ncbi.nlm.nih.gov/pubmed/24972858 http://dx.doi.org/10.1109/TMI.2017.2657225 http://www.ncbi.nlm.nih.gov/pubmed/28129153 http://dx.doi.org/10.1007/s00586-016-4426-3 http://www.ncbi.nlm.nih.gov/pubmed/26851954 http://dx.doi.org/10.1007/s11548-015-1202-5 http://www.ncbi.nlm.nih.gov/pubmed/26026697 http://dx.doi.org/10.1007/s11548-017-1575-8 http://www.ncbi.nlm.nih.gov/pubmed/28361323 http://dx.doi.org/10.1007/s10278-017-9945-x http://www.ncbi.nlm.nih.gov/pubmed/28083827 http://dx.doi.org/10.1056/NEJMp1606181 http://www.ncbi.nlm.nih.gov/pubmed/27682033 http://www.ncbi.nlm.nih.gov/pubmed/26262337 http://dx.doi.org/10.1016/j.ergon.2015.11.008 http://dx.doi.org/10.1093/bioinformatics/btw172 http://www.ncbi.nlm.nih.gov/pubmed/27153678 https://creativecommons.org/licenses/by/4.0/legalcode An Expert System to Diagnose Spinal Disorders [Objective:] Objective: Methods: Results: Conclusion: 1. BACKGROUND 2. RATIONALE 3. BASIC ELEMENTS 3.1. Rule-Based Expert System 3.2. Inference in Rule-Based expert systems 3.3. Backward Chaining Inference 3.4. Certainty Factor 4. LITERATURE REVIEW 4.1. Methods Based on Machine Learning 4.2. Rule-based Approaches 5. KNOWLEDGE ENGINEERING METHOD 5.1. Knowledge Acquisition 5.2. The Proposed Algorithm of Inference Engine 5.3. Implementation of Knowledge Base and User Interface 5.4. Diagnosis in Practice 6. RESULTS 6.1. Phase 1: Evaluation Based on Real-world Samples 6.2. Phase 2: Evaluations Based on Medical Records 6.3. Cut-off Values in the Evaluation Procedure 7. DISCUSSION CONCLUSION AND FUTURE STUDIES LIST OF ABBREVIATIONS AUTHOR’S CONTRIBUTION CONSENT FOR PUBLICATION FUNDING CONFLICTS OF INTERESTS ACKNOWLEDGEMENTS REFERENCES 
work_izxzr6jh5vbxtl5aqnkwkam7we	----	e-Participation: A Theoretical Perspective The University of Bradford Institutional Repository http://bradscholars.brad.ac.uk This work is made available online in accordance with publisher policies. Please refer to the repository record for this item and our Policy Document available from the repository home page for further information. To see the final version of this work please visit the publisher’s website. Access to the published online version may require a subscription. Link to publisher version: http://dx.doi.org/10.1016/j.eswa.2013.06.061 Citation: Irani Z and Kamal MM (2014) Intelligent Systems Research in the Construction Industry. Expert Systems with Applications. 41(4) Part 1: 934-950. Copyright statement: © 2014 Elsevier. Reproduced in accordance with the publisher's self- archiving policy. This manuscript version is made available under the CC-BY-NC-ND 4.0 license. http://dx.doi.org/10.1016/j.eswa.2013.06.061 http://creativecommons.org/licenses/by-nc-nd/4.0/ http://creativecommons.org/licenses/by-nc-nd/4.0/ 1 Profiling Two Decades of Intelligent Systems Research in the Construction Industry Zahir Irani 1 zahir.irani@brunel.ac.uk Brunel Business School Brunel University Uxbridge, Middlesex UB8 3PH, United Kingdom Telephone: 01895266054 Muhammad Mustafa Kamal muhammad.kamal@brunel.ac.uk Brunel Business School Brunel University Uxbridge, Middlesex UB8 3PH, United Kingdom Telephone: 01895267728 1 Corresponding Author mailto:zahir.irani@brunel.ac.uk mailto:muhammad.kamal@brunel.ac.uk 2 Profiling Two Decades of Intelligent Systems Research in the Construction Industry Abstract With the increasing complexity of problems in the construction industry, researchers are investigating computationally rigorous intelligent systems with the aim of seeking intelligent solutions. The purpose of this paper is therefore to analyse the research published on ‘intelligent systems in the construction industry’ over the past two decades. This is achieved to observe and understand the historical trends and current patterns in the use of different types of intelligent systems and to exhibit potential directions of further research. Thus, to trace the applications of intelligent systems to research in the construction industry, a profiling approach is employed to analyse 514 publications extracted from the Scopus database. The prime value and uniqueness of this paper lies in analysing and compiling the existing published material by examining variables (such as yearly publications, geographic location of each publication, etc.). This has been achieved by synthesising existing publications using fourteen keywords 2 ‘Intelligent Systems’, ‘Artificial Intelligence’, ‘Expert Systems’, ‘Fuzzy Systems’, ‘Genetic Algorithms’, ‘Knowledge-Based Systems’, ‘Neural Networks’, ‘Context Aware Applications’, ‘Embedded Systems’, ‘Human-Machine Interface’, ‘Sensing and Multiple Sensor Fusion’, ‘Ubiquitous and Physical Computing’, ‘Case-based Reasoning’ and ‘Construction Industry’. The prime contributions of this research are identified by associating (a) yearly publication and geographic location, (b) yearly publication and the type of intelligent systems employed/discussed, (c) geographic location and the type of research methods employed, and (d) geographic location and the types of intelligent systems employed. These contributions provide a comparison between the two decades and offer insights into the trends in using different intelligent systems types in the construction industry. The analysis presented in this paper has identified intelligent systems studies that have contributed to the development and accumulation of intellectual wealth to the intelligent systems area in the construction industry. This research has implications for researchers, journal editors, practitioners, universities and research institutions. Moreover, it is likely to form the basis and motivation for profiling other database resources and specific types of intelligent systems journals in this area. Keywords: Intelligent Systems, Artificial Intelligence, Construction Industry, Profiling, Scopus Database. Article Type: Profiling Paper 1. Introduction Global economic competition has prompted many organisations to explore potential opportunities for enhancing the delivery of their products or services (Park, Lee, Kwon, 2010). This trend has become apparent in the construction industry as well, with clients/partners demanding a better service and projects that meet their requirements more meticulously (Chen, Griffs, Chen, Chang, 2012; Chaphalkar & Patil, 2012). This inclination towards transformation of construction operations has challenged the industry to become more efficient, integrated and more attractive, both in the eyes of society and its prospective workforce (Bowden, Dorr, Thorpe, Anumba, 2006; Cheng, Tsai, Lai, 2009). In response to this challenge, several government, industry or research-led construction change initiatives have emerged in most developed countries (Courtney & Winch, 2002). In parallel with, and to serve these initiatives there has been a rigorous effort, within the research and academic sector, to investigate and implement existing and emerging intelligent solutions that facilitate the improvements required to develop the construction industry (Bowden et al., 2006). Globally, the construction industry is one of the main sectors that was estimated to reach an approximate US$5.5 trillion at the 2 The search was conducted using a combination of two keywords e.g. ‘Intelligent Systems AND Construction Industry’. This process was repeated with other keywords as well to extract relevant papers. This process is further explained in detail in the research methodology section. 3 end of 2007 (Harmon, 2003). The huge investment made in construction operations worldwide is representative of approximately 4.6% of gross domestic product expended at the national level (El- adaway, 2008). An industry of this size and magnitude, therefore, has across-the-board effects on the development and affluence of nations. The construction industry’s contribution to the nation’s economy is, however, inhibited by an increasing number of problems that unfold and often intensify as projects progress (Mahfouz & Kandil, 2012). In order to understand and address the complex problems in the construction industry many academics and practitioners have conceptually and empirically researched the intelligent systems area within different contexts. Some recent varied examples include:  Liu & Tsai (2012) presenting a fuzzy assessment approach for occupational hazards in the construction industry;  Jiang, Jang, Skibniewski (2012) exploring wireless technology for tracking construction materials;  Coelho & de Brito (2011) proposing a knowledge-based system for materials distribution;  Ooshaksaraie, Basri, Abu Bakar, Maulud (2012) developing an expert system in stormwater management planning for construction sites in Malaysia; and  Chen et al., (2012) developing an evolutionary fuzzy hybrid neural network to enhance project cash flow management in the construction industry. Research on applying intelligent systems (such as artificial intelligence techniques) to the management of construction industry projects started in the 1980s (Hua, 2008). These techniques were, in some instances, compared to traditional simulation and statistical regression approaches to evaluate enhancements in areas of labour productivity, litigation, forecasting demands, cost estimations, optimising construction site layout, cash flow prediction, and bidding in construction projects (Goh, 1996; Sonmez & Rowings, 1998; Seydel, 2003). More explicitly, in a review of the use of intelligent systems solutions (e.g. artificial neural networks) in the field of construction management, Boussabaine (1996) cited the benefits of artificial intelligence techniques over mathematical and statistical models in situations where the process to be modelled is complex and where traditional models lack the ability to learn by themselves, generate solutions and respond adequately to highly correlated, incomplete or previously unknown data. The scope and applicability of intelligent systems solutions clearly indicates that this area can and has addressed a multiplicity of organisational problems with the support of a large variety of techniques and methods – which can help solve complex decision-making queries and provide insights into them. 1.1 Research Aim Since the review conducted by Boussabaine in 1996 on artificial neural networks in the field of construction management and more recently by Hua in 2008 on the applications of quantitative analysis techniques in construction economics and construction management (for both traditional and artificial intelligence techniques), this paper attempts to broaden the scope of their reviews by further assessing the applicability of different types of intelligent systems in the construction industry. Explicitly in respect of Boussabaine’s and Hua’s conclusion for construction economics and construction management (where it only focused on artificial neural networks and quantitative analysis techniques), this research specifically aims to: “identify the historical trends and current patterns in the use of different types of intelligent system in the construction industry. These trends and patterns will support in anticipating the future propensities in the use of different intelligent systems in the construction industry.” 1.2 Research Objectives This research intends to assess the extant research published on intelligent systems in the construction industry by employing a profiling approach and attempting to highlight the most frequently used intelligent systems in the construction industry. From the empirical findings (using 14 4 keywords), initially 550 papers were identified from the Scopus database during the period 1990 to 2012. After assessing the 550 publications, 514 papers were finally considered relevant and taken forward for further investigation. Since 1990, a number of academic outlets including among others: Expert Systems with Applications, International Journal of Intelligent Systems, Construction Management and Economics, Computers and Operations Research, International Journal of Project Management, Journal of Construction Engineering and Management, Automation in Construction, etc., have been dedicated to publishing research on intelligent systems. These sources offer a true reflection of the intelligent systems area and have emerged as quality outlets for publishing research in this field. Contributors from across the world have made contributions to the intelligent systems area. Given the limited research in the area of intelligent systems in the construction industry (as evidenced by the empirical findings), the rationale for undertaking this research is to provide a better understanding of the types of intelligent systems employed in existing studies (including other variables). In this respect, a review of the relevant intelligent systems outlets would help to shed light on intelligent systems types adopted to depict the evolution of the domain in future. This will inform researchers and academics engaged in the area of the most widely employed intelligent systems types and editorial preferences of the journals selected as part of this research. Thus, the aim of this research is realised by means of the following objectives – i.e. to identify the:  number of publications in each year;  geographic location of each publication;  type of publication (i.e. research or technical paper, literature review, viewpoint);  type of research methods employed (i.e. experiment, case study, mixed method, analytical);  type of intelligent systems employed/discussed (i.e. artificial intelligence, expert systems, fuzzy systems, genetic algorithms, knowledge-based systems, neural networks, etc.);  type of journal;  context type (i.e. this variable specifically focuses on the type of construction sector e.g. road/highway construction, building construction, bridge construction, etc.); finally,  citation analysis for each paper (i.e. by accessing the citations from the Scopus database); The prime contribution of this research focuses on identifying the association between the following; these contributions, however, are derived as a result of the abovementioned objectives.  yearly publications and geographic location;  yearly publications and the type of intelligent systems employed/discussed;  geographic location and the type of research methods employed; and  geographic location and the type of intelligent systems employed/discussed. The above four prime contributions are presented based on a comparison between the last two decades i.e. between 1990 to 1999 and 2000 to 2012. This comparison will provide insights into the historical trends and current patterns in using different types of intelligent systems in the construction industry. In order to achieve these objectives, the authors contribute by conducting a systematic review of 514 intelligent systems papers. This type of profiling research is essential in order to establish an understanding of intelligent systems (and their different types) and the state-of-the-art growth in the theory and application of intelligent systems within the construction industry. The remainder of this paper is structured as follows. Section 2 highlights the issues in the construction industry and the significance of intelligent systems in the construction industry. Thereafter, Section 3 presents the research methodology highlighting the overall research conducted in this paper. Then, in Section 4, we analyse and discuss the different variables (i.e. objectives) in intelligent systems studies with Section 5 summarising the contributions of this research. Finally, conclusions, implications for theory and practice and limitations are given in Section 6. 2. Construction Industry and Intelligent Systems: A Literature Perspective The business world is replete with existing and emerging set of events that further complicate decisions and the decision-making process in organisations (Voordijk, Meijboom, de Haan, 2006; Al- 5 Zawahreh & Cox, 2009; Ning, Lam, Lam, 2011). For example, the construction industry is teeming with countless issues that make management exceptionally complex (Winch, 2010). The perils and uncertainties inherent in such a dynamic environment make the management of organisational resources even more crucial (Park et al., 2010). The latter argument is supported by Chen et al., (2012) who state that construction projects are becoming increasingly larger and more intricate in terms of physical size and cost; for this reason the risks and potential for losses require better management. According to Beavers et al., (2006), one of the foremost reasons for casualties in the construction industry is in deploying cranes or giant frames (i.e. derricks) while lifting operations are being carried out. This requires proper workforce training while using such machinery or perhaps developing intelligent solutions that can (if not completely) then partially replace the human resource (Munisamy, 2009; Al-Zawahreh & Cox, 2009). The construction industry is considered as one of the major high-risk industries globally, where accidents that include those that result from falling from a height being the most frequent (Guo, Li, Chan, Skitmore, 2011). For example, some recent events that received widespread exposure in the United States and Europe, respectively, include the ‘Big Dig’ ceiling collapse in Boston (2005) and the Cologne subway/historical archive collapse (2009), both of which involved casualties. In this regard, factors for success may also vary from project to project in the construction sector. Even though ‘human’ experts can often accomplish a reasonable project result, deficits almost always follow due to managers failing to take all relevant factors into consideration and/or lacking access to all relevant information (Cheng et al., 2009; Cheng, Tsai, Sudjono, 2012). The construction industry in the past has been plagued with similar problems and is categorised by (a) specific intricacy variables due to individual industry ambiguities and inter-dependencies, and (b) insufficiency of operations (Tah & Carr, 2000; Dubois & Gadde, 2002; Beavers, Moore, Rinehart, Schriver, 2006). Despite the availability of several intelligent systems, construction managers are still perplexed when faced with a new problem in decision-making, when they ought to establish which existing intelligent systems are most apposite given the nature of the system, the objectives for development, time constraints and computing capacity (Bolduc, Renaud Boctor, Paporte, 2008; Cui & Lu 2009; Bisaillon, Cordeau, Paporte, Pasin, 2011). Decision-making in the construction industry is more about making robust decisions rather than optimal decisions (Winch, 2010). Construction industry projects whether dealing with steel bridge construction (Cui & Lu, 2009), highway construction (He & Sun, 2010) or developing partnerships between client and contractors (Bresnen & Marshall, 2000) have been marked as risky with uncertain decision-making resulting in poor performance. The latter argument is supported by AbouRizk (2010), who reports that when construction projects become large and/or complex 3 , they become more difficult to manage by employing conventional techniques. According to Ruuska, Ahola, Artto, Locatelli, Mancini (2011), large and/or complex construction projects present exceptional challenges as a result of: (a) the dynamic network of organisations that combine resources, capabilities and knowledge of participating actors to achieve clients’ needs and (b) differing and often conflicting objectives and expectations not evident if the projects were carried out by individual firms. As such challenges are not effortlessly dealt with using conventional techniques; they may result in lack of performance and, in some situations, project failure (Ruuska et al., 2011). Thus, the use of computer simulation techniques becomes vital, as these offer effective and efficient tools essential for designing and analysing construction related processes, irrespective of complexity or project size. 2.1 Significance of Intelligent Systems in the Construction Industry The intelligent systems area has expanded remarkably over recent years; in terms of the range of techniques and number of applications where they often provide a competitive advantage when compared with other conventional approaches (Negnevitsky, 2005; Uraikul, Chan, Tontiwachwuthikul, 2007). Intelligent systems include a range of techniques (e.g. neural networks, fuzzy logic/systems, genetic algorithms and genetic programming, expert systems, case-based 3 In this paper, we denote ‘complexity’ of construction projects with an example i.e. the bundling of construction operations into fewer, but larger contracts signifies that more responsibility is transferred from major clients of the construction industry to contractors. The construction contractor in such a leading role is faced with the challenge of managing the interrelationships of the whole supply chain which make the projects complex (Hagan, Bower, Smith, 2012). 6 reasoning, etc.) that operate synergistically to improve strategic decision-making and provide flexible data/information processing abilities for dealing with real world situations (Li, 2007; Hines, Leeson, Martínez-Ramón, Pardo, Llobet, Iliescu, Yang, 2008). For example, the complexity of construction industry operations and the resulting problem-solving capacity required is recognised as a challenge. In this regard, intelligent systems have played a significant role in providing a standardised methodological approach to solving important and fairly complex problems, facilitate making consistent decisions by using appropriate artificial intelligence solutions, and obtaining consistent and reliable results over time (Byrd & Hauser, 1991; Arain & Pheng, 2006; Shi, 2012). Intelligent systems can thus exploit the tolerance for inaccuracy, indecision/uncertainties, estimated reasoning and fractional fact so as to realise compliance, vigour and cost-effective solutions (Hines et al., 2008). Intelligent systems have been defined in several ways, for instance according to:  Curry and Moutinho (1991), “intelligent systems are programs that stretch to represent the knowledge of an expert in a certain domain.”  Kumara, Joshi, Kashyap, Moodie, and Chang (1986), “an intelligent system is a tool which is able to conceive the special knowledge of a problem and by using intelligently this knowledge of this special domain, it can suggest alternative actions.”  Feigenbaum (1982), “intelligent system is an intelligent program that, in order to resolve a difficult problem necessitating a considerable experience for its solution uses a special knowledge and deduction procedures.” These definitions clearly indicate that intelligent systems are software programs that syndicate the knowledge of experts and attempt to resolve distinct problems by imitating the reasoning processes of experts (Matsatsinis & Siskos, 2003). In the context of the construction industry, Guo et al., (2011) recommended the use of artificial intelligent systems to enhance the safety levels i.e. using game technology-based safety training platforms in a virtual environment in order to reduce existing fatality rates. Behzadan, Aziz, Anumba, Kamat, (2008) report that the unplanned and vibrant nature of a construction site, with the dangers and hitches presented by on-site operations, demand the use of intelligent ways to support the on-site construction workforce. Context-aware information delivery offers the capacity to intelligently capture and infer the user context, and deliver data and services to the mobile-worker based on the user’s context (Behzadan et al., 2008). Similarly, in the context of tainting construction industry operations and projects as risky, Tah & Carr (2000) report that organisations will have to reconsider their tactics to the approaches in which risks are handled within their projects. In doing so, proposing the development of robust knowledge-based systems that are capable of enhancing risk analysis and management processes and, as a result, directing the construction industry to develop benchmarks that are maintainable and frequently enhanced. On the other hand, repetitive construction projects are a common phenomenon in the construction industry that require resources to perform similar operations in different units (Vanhoucke, 2006). Due to the frequent resource movement from one unit to another, an effective schedule is imperative to guarantee the continuous usage of resources in repetitive operations between units. In this regard, Long & Ohsato (2009) developed a genetic algorithm-based method for scheduling repetitive construction industry projects. According to anecdotal references, the use of intelligent systems (e.g. expert systems) will most likely be the most important application of intelligent solutions for construction over the next decade. In the years to come, there is good potential for increased use of self-directed robots controlled by expert systems. Such advanced-application robots would finish concrete and spray paint buildings (already being done in Japan), apply sprayed insulation to structural steel members, and even install structural steel. 3. Research Methodology Building a profile of the last two decades of intelligent systems publications necessitated that the authors systematically review a total of 514 papers to capture data on several variables (with the whole research design presented in Figure 1). In order to perform such research, the authors selected the Scopus database. The rationale for selecting the Scopus database was that it covers nearly 18,000 titles from over 5000 international publishers, including coverage of 16,500 peer-reviewed journals on 7 different areas. Therefore, it is possible to search for and locate a significant proportion of the published material on intelligent systems using the general and advanced search facility. The authors utilised both facilities (i.e. general and advanced search) within this research exercise. The reason for employing a ‘General Search’ approach was simply that it is easy to use and its characteristics facilitate the repetition of searches without any confusion; hence it is relatively straightforward to obtain consistent results in repetitive searches provided the same search criteria are applied (Dwivedi et al., 2008; Williams et al., 2009). The ‘Advanced Search’ was used to further identify specific intelligent systems papers primarily focusing on the construction industry. However, the search was not restricted to occurrences of any of the selected keywords appearing in the article title only but also the abstract, keywords and the whole paper was considered. Thus, the authors initially searched the Scopus database using the ‘intelligent systems’ keyword only. This search resulted in 4,437 publications from 1990 to 2012, specifically focusing on journal articles only. 8 First Search Results First Search Results Figure 1: Research Design Keyword: “Intelligent Systems” First Search Results Keywords: “Intelligent Systems” AND “Construction” 4,437 Papers = 179 Papers = Second Search Results 145 Papers 31 Papers 3 Papers Keywords: “Intelligent Systems” AND “Construction Industry” 3 Papers = Third Search Results Similar Papers Not enough to determine the trends/changes of intelligent systems research over the last two decades “Artificial Intelligence” AND “Construction Industry” 88 Papers (out of total 93) = Fourth Comprehensive Search Results “Expert Systems” AND “Construction Industry” 73 Papers (out of total 98) = “Fuzzy Systems” AND “Construction Industry” 26 Papers (out of total 26) = “Genetic Algorithm” AND “Construction Industry” 96 Papers (out of total 121) = “Knowledge-Based Systems” AND “Construction Industry” 74 Papers (out of total 115) = “Neural Networks” AND “Construction Industry” 94 Papers (out of total 116) = “Sensing or Multiple Sensor Fusion” AND “Construction Industry” 26 Papers (out of total 31) = “Context Aware” AND “Construction Industry” 3 Papers (out of total 3) = “Embedded Systems” AND “Construction Industry” 2 Papers (out of total 3) = “Ubiquitous or Physical Computing” AND “The construction industry” 7 Papers (out of total 23) = “Human-Machine Interface” AND “The construction industry” 2 Papers (out of total 2) = “Case-based Reasoning” AND “The construction industry” 20 Papers (out of total 41) = 511 Papers Overall Fourth Comprehensive Search Total + 3 Papers = 514 Papers Overall Total Number of Papers considered for further Analysis SEARCHING SCOPUS DATABASE Start of Profiling Research Presenting the Findings = Analysing & Presenting the Main Contributions Sections 4.1 to 4.7 Sections 5.1 to 5.4 Concluding the Research End of Profiling Research S te p 1 S te p 2 S te p 3 S te p 4 Step 5 Step 6 9 However, as the focus of this research is on the construction industry, the authors repeated this process by restricting the search using the ‘intelligent systems’ AND ‘construction’ keywords together. This exercise resulted in 179 publications from 1990 to 2012. After analysing these 179 papers, the authors identified seven relevant papers. On further analysing these seven papers, only three specifically focused on the construction industry. The remaining articles were either not narrated in English (31 papers) or not related to this research (145 papers). Although these discarded papers highlighted the keywords (i.e. ‘intelligent systems’ AND ‘construction’) they were not specifically from the construction industry, but used the phrase ‘construction’ as a general terminology. To avoid further avoid any confusion and extract the correct number of publications on intelligent systems, the authors decided to use intelligent systems and construction industry keywords together i.e. stated in the following manner ‘intelligent systems’ AND ‘construction industry’. This search resulted in three publications – similar to those that resulted after analysing 179 publications. This sample of three papers was certainly not realistic to determine changes in intelligent systems research over the last two decades. Therefore, the authors further explored the literature and identified different intelligent systems types i.e. Artificial Intelligence, Expert Systems, Fuzzy systems, Genetic Algorithms, Knowledge-Based Systems, Neural Networks, Context Aware Applications, Embedded Systems, Human-Machine Interface, Sensing & Multiple Sensor Fusion, and Ubiquitous & Physical Computing. The authors identified that these intelligent systems types are also central to the scope of the Expert Systems with Applications journal including its current special issue on ‘Intelligent Systems in Construction’. After applying the same search process on each intelligent system type (using ‘Construction Industry’ as the second keyword), the authors found the following number of publications for each. As mentioned below the exact number of papers for each keyword is highlighted in bold, whereas the total number of papers with each keyword is mentioned inside the brackets (as the actual number of papers considered for further investigation and analysis). This is because while conducting the search with a certain set of keywords (e.g. artificial intelligence and construction industry) there were some papers that were already extracted with another keyword (e.g. in this category are expert systems, genetic algorithms, knowledge-based systems, and case-based reasoning). Due to the latter, the number of types of intelligent systems differs from the actual number of papers extracted. This can be attributed to the fact that expert systems, genetic algorithms, knowledge-based systems, and case- based reasoning are, in essence, the core techniques used in developing artificial intelligent systems. This is further explained in the following Table 1 with each point: 10 Keywords Papers Extracted ‘Intelligent Systems’ AND ‘Construction Industry’  3 (out of a total of 7) papers. o Out of the 7 papers, only 3 papers either employed or discussed intelligent systems in the construction industry. ‘Artificial Intelligence’ AND ‘Construction Industry’  88 (out of a total of 93) papers. o From the 88 papers:  36 papers were on artificial intelligence,  10 papers on decision support systems,  9 papers on knowledge-based systems,  8 papers on neural networks,  8 papers on genetic algorithms,  6 papers on expert systems,  4 papers on case-based reasoning,  4 papers on different intelligent systems types (e.g. 2 papers on intelligent construction systems, and 1 each on intelligent scheduling systems and marketing intelligent systems),  2 papers on fuzzy systems, and  1 paper on sensor technology. ‘Expert Systems’ AND ‘Construction Industry’  73 (out of a total of 98) papers. o All the 73 papers either employed or discussed expert systems in the construction industry. ‘Fuzzy systems’ AND ‘Construction Industry’  26 (out of a total of 26) papers. o From these 26 papers:  21 were on fuzzy systems,  3 on artificial intelligence, and  2 on case-based reasoning. ‘Genetic Algorithms’ AND ‘Construction Industry’  96 (out of a total of 121) papers. o From the 96 papers:  91 were on genetic algorithms,  5 on artificial intelligence. ‘Knowledge-Based Systems’ AND ‘Construction Industry’  74 (out of a total of 115) papers. o From the 74 papers:  70 papers were on different types of knowledge-based systems,  1 paper on fuzzy systems, and  3 papers discussed case-based reasoning. ‘Neural Networks’ AND ‘Construction Industry’  94 (out of a total of 114) papers. o From the 94 papers:  89 discussed neural networks,  3 on artificial intelligence,  1 paper on case-based reasoning, and  1 on genetic algorithms. ‘Sensing’ OR ‘Multiple Sensor Fusion’ AND ‘Construction Industry’  26 (out of a total of 31) papers. o All 26 papers discussed sensing or sensor technology in the construction industry. ‘Context Aware’ AND ‘Construction Industry’  3 (out of a total of 3) papers. o All three papers discussed context-aware systems in the construction industry. ‘Embedded Systems’ AND ‘Construction Industry’  2 (out of a total of 3) papers. o Both the papers discussed embedded systems such as cyber physical systems and environmental sensors for telepresence. ‘Ubiquitous’ OR ‘Physical Computing’ AND ‘Construction Industry’  7 (out of a total of 23) papers. o All 7 papers were on ubiquitous computing in the construction industry. ‘Human-Machine Interface’ AND ‘Construction Industry’  2 (out of a total of 2) papers. o Both papers discussed human-machine interface related systems in the construction industry e.g. augmented reality computer-aided drawing and robotic welding systems. ‘Case-based Reasoning’ AND ‘Construction Industry’  20 (out of a total of 41) papers. o All 20 papers were on case-based reasoning. Table 1: Keywords Used and Number of Papers Extracted through Overall Search 11 All the above paper numbers and the type of intelligent systems employed/discussed are summarised in Figure 1 and Table 5, respectively. Papers on other intelligent system types such as multi-agent systems, ambient intelligence, intelligent robotic systems, and autonomous computing, etc., were also searched using the same process as above, but there were no papers returned from the search by the Scopus database. Moreover, our search activities were restricted to the number of journals extracted from the Scopus database only, but there are many well-known journals focusing on intelligent systems and artificial intelligence e.g.:  Artificial Intelligence: An International Journal,  International Journal of Neural Systems,  International Journal of Intelligent Systems,  Artificial Intelligence Review,  Journal of Intelligent Material Systems and Structures,  Journal of Intelligent Manufacturing,  IEEE Intelligent Systems,  Journal of Intelligent and Robotics Systems Theory and Applications,  Knowledge-based Systems,  Robotics and Autonomous Systems,  Applied Artificial Intelligence,  IEEE Intelligent Systems and their Applications,  Journal of Intelligent Systems,  Web Intelligence and Agent Systems,  Journal of Intelligent and Fuzzy Systems, and  Intelligent Automation and Soft Computing. Although these journals are also indexed in the Scopus database, they did not return any papers based on our 14 keywords. Figure 2: Total Number of Papers Published between 1990 and 2012 As presented in Figure 2, the largest number of publications were recorded for year 2011 (with C = 57, 11.19%), followed by year 2009 (with C = 53, 10.51%) and year 2010 (with C = 53, 10.31%). With fewer publications (i.e. below the 10 mark) are 1992 (with C = 8, 1.56%), 1993 (with C = 9, 1.75 %), 1994 (with C = 8, 1.56%), 1995 (with C = 8, 1.56%), 1998 (with C = 7, 1.36%), 2000 (with 15 22 8 9 8 8 12 3 7 12 5 19 24 18 22 24 36 24 38 54 53 57 36 0 10 20 30 40 50 60 Total Frequency of Papers from 1990 to 2012 12 C = 5, 0.97%) and 1997 with the least number of papers i.e. with C = 3, 0.58%. Figure 1 clearly highlights the trend of an increasing number of publications in the intelligent systems domain from 2000 onwards with 410 papers; whereas the 1990s decade generated rather small numbers of papers i.e. 104 out of the total 514. This trend evidently signifies that use of intelligent systems is gaining popularity in the research area of the construction industry, specifically from 1990 to 2011. However, there is a relatively sharp decrease in the number of papers in the year 2012 (with C = 36, 7.00%) compared to the previous year 2011, with 57 papers published. In 2011, the leading intelligent systems choice of researchers and practitioners was genetic algorithms with a total of 15 papers (published in different leading journals), followed by neural networks with a total of 11 papers (out of a total of 57). These two choices were followed by fuzzy systems (7 papers), knowledge-based systems and expert systems (with five papers each), artificial intelligence, case-based reasoning and sensor technology (3 papers each), decision support systems and intelligent systems (two papers each), and remaining one paper on context-aware computing. On the other hand, in 2012, the leading intelligent systems choice of researchers and practitioners was neural networks with a total of 8 papers, followed by genetic algorithms with a total of five papers (out of a total of 36). These two choices were followed by artificial intelligence and knowledge-based systems (four papers each), expert systems and fuzzy systems (three papers each), embedded systems, sensor technology and ubiquitous computing (with two papers each) with the remaining one paper on case-based reasoning. Although the number of papers decreased between 2011 to 2012, the types of intelligent systems employed/discussed in research studies during these two years were similar with a few exceptions (i.e. there are two papers on embedded systems and ubiquitous computing each in 2011, but none on these two types of intelligent system in 2012 and on the other hand, one paper on context-aware systems in 2012 and none on this type in 2011). The reasons for fewer papers published in 2012 may be attributable to the changes in the global economic situation. Most of the developed economies are still struggling to overcome the economic problems caused by the global financial crisis in 2007-2008 (Altman, 2008). While this recession resulted in the total collapse of many large financial organisations and a decline in stock markets around the world, the construction industry (specifically those companies involved in home construction) was also directly affected (Altman, 2008). According to the World Economic Situation and Prospects (WESP) 2012, although there were some dispersed signs of developments in 2011 following the aftermath of the global financial crisis, growth of manufacturing production and the construction industry operations specifically remained on the decline (owing to weaker external demand). After a noticeable slowdown in 2011, global economic development was anticipated to remain lukewarm in 2012, with most EU regions growing at a pace below potential (WESP, 2012; Langdon, 2012). Thus, it may be suggested that the weakening situation of the construction industry due to the global financial crisis, has resulted in researchers and academics paying less attention to research in this area during 2010-2011, which resulted in fewer publications in 2012 (owing to limited case study projects and a lack of primary data in 2010-2011). Moreover, it can also be argued that a number of papers may have been submitted sometime in late 2011 or early 2012 but not all papers appear or get published, resulting in lower number of papers in 2012 than previous years. In this context, the top ranking peer-reviewed journals such as ‘Journal of Construction Engineering and Management’, ‘Journal of Computing in Civil Engineering’, ‘Automation in Construction’, ‘Construction Management and Economics’, and ‘Expert Systems with Applications’ and other journals may have contributed to fewer papers in 2012 (i.e. taking a longer time to publish papers). Nevertheless, some anecdotal evidence suggests that regardless of adverse economic conditions, the global construction industry is predicted to reach US $8,929 billion in 2017 with a Compound Annual Growth Rate (CAGR) of 7.3% over the next five years. This increase in construction industry operations will be supported by the major global event organisers including among others: the 2014 FIFA World Cup, the 2016 Olympic Games in Brazil, the 2014 Winter Olympic Games, and 2018 FIFA World Cup in Russia (Lucintel, 2012). These global corporations are anticipated to provide stimulus to the construction industry over the forecasted period. Having seen the gradual increase in the number of papers over the years (i.e. from 1990 to 2011) and the increasing support to the construction industry by the global corporations, it can be argued that after the decrease in the number of papers in 2012, a relatively significant increase in the number of publications may be anticipated to follow from 2013 onwards, with researchers presenting more intelligent solutions to the ever-growing problems in the construction industry (Cheng et al., 2012; Chen et al., 2012). Moreover, the authors 13 also assert that those papers submitted sometime in 2012, but which were not published in 2012, may appear in 2013 and onwards. Analysis shown in subsequent sections also highlights the type of intelligent systems employed/discussed in different contexts of the construction industry e.g. use of different types of intelligent systems in road construction, bridge construction, building construction, construction site layout planning, case flow prediction, construction litigation, dispute resolution, etc. 4. Findings and Discussion The findings of this study are now presented under different subsections. Each of the seven subsections discusses the findings in relation to a particular variable. The variables are as follows: number of regions – geo-spatial coverage (Section 4.1), types of publication (Section 4.2), types of research methods employed (Section 4.3), types of intelligent systems employed/discussed (Section 4.4), types of journal (Section 4.5), publication context type (Section 4.6), and citation analysis (Section 4.7). 4.1 Number of Regions (Geo-Spatial Coverage) Table 2 highlights the number of publications from 48 different geographical regions across the globe between 1990 and 2012 on intelligent systems in the construction industry. From the total number of publications (i.e. 514) analysed, the largest number of contributions came from researchers from the USA (C = 120, 23.35%), followed by Taiwan (C = 56, 10.89%), the UK (C = 52, 10.12%) and Canada (C = 51, 9.92%). The results in Table 1 clearly indicate that the first seven regions (i.e. USA, Taiwan, UK, Canada, China, Korea and Hong Kong) lead on intelligent systems research in the construction industry. Although it is not the intention of this research to present the contributing co- authors’ geographic locations, most of the co-authors (i.e. 2 nd , 3 rd , 4 th and even 5 th ) are also from the top seven regions with the exception of a few from Egypt, India and Iran. Geo-Spatial Coverage Frequency of Publications Per cent Geo-Spatial Coverage Frequency of Publications Per cent USA 120 23.35% Poland 3 0.58% Taiwan 56 10.89% Switzerland 3 0.58% UK 52 10.12% Greece 2 0.39% Canada 51 9.92% Ireland 2 0.39% China 31 6.03% Qatar 2 0.39% Korea 29 5.64% Serbia 2 0.39% Hong Kong 25 4.86% Sri Lanka 2 0.39% Turkey 15 2.92% Chile 1 0.19% Iran 13 2.53% Croatia 1 0.19% Egypt 13 2.53% Cyprus 1 0.19% Australia 11 2.14% France 1 0.19% India 10 1.95% Indonesia 1 0.19% Singapore 8 1.56% Israel 1 0.19% Spain 7 1.36% Mexico 1 0.19% Saudi Arabia 7 1.36% Nigeria 1 0.19% Portugal 5 0.97% Norway 1 0.19% Italy 4 0.78% Oman 1 0.19% Malaysia 4 0.78% Philippines 1 0.19% Sweden 4 0.78% Romania 1 0.19% Netherlands 4 0.78% South Africa 1 0.19% Finland 3 0.58% Thailand 1 0.19% Japan 3 0.58% Tunisia 1 0.19% Jordan 3 0.58% Vietnam 1 0.19% Lithuania 3 0.58% Yugoslavia 1 0.19% Total 514 100.00% Table 2: Frequency of Publications in each Geographical Location between 1990 and 2012 14 4.2 Type of Publication In this section, the authors categorise their list of 514 papers in our data-set based on the type of publication. The authors used a fairly similar list of publication types to that employed by Dwivedi & Mustafee (2010). This list is also similar to those identified by the publisher – Emerald. The data presented in Table 3 illustrate that the vast majority of the publications are research papers (C = 458, 89.11%), followed by general review and literature review papers (with C = 16, 3.11%, respectively). The remaining types of publication with their frequencies are presented in Table 2. The large number of research papers clearly indicates the significance of the intelligent systems area in the construction industry and that most researchers are focused on developing and proposing intelligent solutions to the many problems within the construction industry e.g. construction site layout planning, case flow prediction, construction litigation, dispute resolution, etc. Type of Publication Total Frequency Per cent Research Paper 458 89.11% General Review 16 3.11% Literature Review 16 3.11% Conceptual Paper 14 2.72% Technical Note 8 1.55% Research Report 1 0.19% Short Paper 1 0.19% Total 514 100.00% Table 3: Classification of Publication Types between 1990 and 2012 4.3 Type of Research Methods Employed The findings suggest that although a total of 15 different research methods were recorded from our data analysis, the majority of studies employed analytical methods (C = 204, 39.70%), followed by case study (C = 89, 17.31%) and conceptual/descriptive/theoretical (C = 49, 9.53%) methods. With regard to the analytical method – it was denoted as a combination of five different methods i.e. statistics, computer programming, simulation, algorithm and mathematical modelling. The other categories with their associated counts and percentages are presented in Table 4. Type of Research Methods Employed Total Frequency Per cent Analytical 204 39.70% Case Study 89 17.31% Conceptual / Descriptive / Theoretical 49 9.53% Secondary Data Analysis 40 7.78% Experiment 38 7.39% Survey 29 5.64% Design Research 25 4.86% Interview 15 2.92% Mixed Method 11 2.14% Questionnaire 6 1.17% Primary & Secondary Data Analysis 3 0.58% Meta-Analysis 2 0.38% Focus Group Interviews 1 0.19% Field Trial 1 0.19% Observation 1 0.19% Total 514 100.00% Table 4: Research Methods Employed between 1990 and 2012 15 4.4 Type of Intelligent Systems Employed/Discussed The authors conducted a thorough analysis of the papers in order to identify the different types of intelligent systems employed (as highlighted in Table 5). This was to determine the most frequently employed intelligent systems types among the 514 papers used in this research. From the findings, it is clear that genetic algorithms (with C = 100, 19.45%), neural networks (with C = 97, 18.87%), knowledge-based systems (with C = 79, 15.37%) and expert systems (with C = 79, 15.37%) are the most frequently used among 514 papers assessed, followed by artificial intelligence (with C = 47, 9.14%), case-based reasoning (with C = 30, 5.84%), and sensor technology (with C = 27, 5.25%). Genetic algorithms have been widely studied, experimented and applied in many fields of engineering. Not only does the genetic algorithm provide an alternative method to problem-solving, it consistently outperforms other traditional methods in most problems. Many real-world construction problems (e.g. occupational hazards, materials distribution, increasing disputes that unfold and often escalate as construction projects progress, construction site layout planning, predicting the risks of contractor defaults, resource allocation and management in construction projects, and contractual claims) involve finding optimal parameters, which might prove difficult using traditional methods (e.g. developing knowledge management functionalities based on conventions of practice – such an approach is established using reactive problem-solving methods) but are ideal for genetic algorithms. However, due to its outstanding performance in optimisation, the genetic algorithm has been wrongly regarded as a function optimiser. Type of Intelligent System Employed/Discussed Total Frequency Per cent Genetic Algorithm 100 19.45% Neural Network 97 18.87% Knowledge-Based Systems 79 15.37% Expert Systems 79 15.37% Artificial Intelligence 47 9.14% Case-Based Reasoning 30 5.84% Sensor Technology 27 5.25% Fuzzy Systems 24 4.67% Decision Support Systems 10 1.94% Ubiquitous Computing 7 1.36% Intelligent Systems 7 1.36% Context-Aware Systems 3 0.58% Human-Machine Interface 2 0.39% Embedded Systems 2 0.39% Total 514 100.00 Table 5: Type of Intelligent System Employed/Discussed between 1990 and 2012 4.5 Type of Journal From the total of 514 papers selected, a list of 122 different types of journal was used by researchers, academics and practitioners to publish their research. To avoid lengthy tables in the paper, the authors excluded those journals that had a frequency of one and denoted with ‘…’ i.e. out of the 122 journals used, 91 journals had a frequency of one. According to the findings illustrated in Table 6, the Journal of Construction Engineering and Management was the most used by researchers (with C = 109, 21.21%), followed by the Journal of Computing in Civil Engineering (C = 58, 11.28%), Automation in Construction (with C = 57, 11.090%), Construction Management and Economics (C = 38, 7.39%) and Expert Systems with Applications (C = 27, 5.25%). The scope of these leading journals is broad, covering all stages of the construction lifecycle from initial planning and design, through construction of the facility, its operation and maintenance, to the eventual dismantling and recycling of buildings and engineering structures. Moreover, these journals are key resources for researchers, practitioners, and students on advances and innovation in construction engineering and management, scheduling, estimating, cost control, quality control, labour productivity, inspection, contract administration, construction management, computer applications, 16 and environmental concerns. The remaining journals and their frequencies and percentages are presented in Table 6. Type of Journal Extracted from Scopus Total Frequency Per cent Journal of Construction Engineering and Management 109 21.21% Journal of Computing in Civil Engineering 58 11.28% Automation in Construction 57 11.09% Construction Management and Economics 38 7.39% Expert Systems with Applications 27 5.25% Canadian Journal of Civil Engineering 15 2.92% Journal of Information Technology in Construction 13 2.53% International Journal of Project Management 12 2.33% Engineering, Construction and Architectural Management 9 1.75% Journal of Civil Engineering and Management 8 1.55% Journal of Management in Engineering 8 1.55% Computers & Structures 8 1.55% Building and Environment 7 1.36% Computer-Aided Civil and Infrastructure Engineering 7 1.36% KSCE Journal of Civil Engineering 7 1.36% Advanced Engineering Informatics 6 1.17% Advances in Engineering Software 4 0.78% Engineering Applications of Artificial Intelligence 3 0.58% Journal of Professional Issues in Engineering Education and Practice 3 0.58% Journal of Quality in Maintenance Engineering 3 0.58% Tsinghua Science and Technology 3 0.58% Computing in Civil Engineering 2 0.39% Energy and Buildings 2 0.39% Engineering Structures 2 0.39% Indian Journal of Engineering & Materials Sciences 2 0.39% Industrial Management & Data Systems 2 0.39% International Journal of Industrial and Systems Engineering 2 0.39% Journal of Engineering and Applied Sciences 2 0.39% Journal of Infrastructure Systems 2 0.39% Sensors and Actuators B: Chemical 2 0.39% WSEAS Transactions on Computers 1 0.19% ... ... ... Yugoslav Journal of Operations Research 1 0.19% Total 514 100.00 Table 6: Type of Journals and their Frequency between 1990 and 2012 4.6 Publication Context Type From the total of 514 papers selected, a list of 43 context types was used by researchers, academics and practitioners to publish their research. To avoid lengthy tables in the paper, the authors considered developing appropriate categories to sum the relevant context types under a suitable category. For example, the ‘Large Construction Project’ category includes papers on context types related to electric power construction projects, hydropower construction projects, rural construction, and tunnel construction. Similarly, the ‘Knowledge Management in Construction Project’ category includes papers on knowledge acquisition, evaluation, planning, and knowledge networks. The same process was repeated for all context types (out of a total of 43) that had more than three papers. Nevertheless, there are 15 papers that focus on ‘Highway Construction’ with the exception of one on road construction. The process did not occur by coincidence, it was initially planned and the authors evaluated the papers accordingly (especially when there were 514 publications to review). A similar categorisation of paper context type is conducted by Williams et al., (2009) and Dwivedi et al., (2008). According to the findings illustrated in Table 6, ‘Construction Project Operations’ was the most researched (with C = 92, 17.90%), followed by ‘Building Construction Projects’ (C = 55, 17 10.70%) and ‘Information Technology in Construction Projects’ (with C = 47, 9.15%). The remaining context types and their frequencies and percentages are presented in Table 7. Context Type Total Frequency Per cent Construction Project Operations 92 17.90% Building Construction Projects 55 10.70% IT in Construction Projects 47 9.14% Decision-Making in Construction Projects 29 5.64% Planning and Scheduling in Construction Projects 22 4.28% Computing in Construction Projects 18 3.50% Construction Project Management 18 3.50% Construction Site Investigation 18 3.50% Knowledge Management in Construction Projects 18 3.50% Highway Construction Projects 15 2.92% Construction Project Costs 14 2.72% Resource Allocation in Construction Projecta 14 2.72% Concrete Products in Construction 11 2.14% Construction Engineering Projects 11 2.14% Risk Assessment & Management in Construction Projects 11 2.14% Construction Litigation 9 1.75% Performance in Construction Projects 9 1.75% Safety Hazard Identification in Construction 8 1.55% Estimation in Construction Projects 8 1.55% Contractor-related Issues in Construction Projects 7 1.36% Dispute Resolution in Construction Project 7 1.36% Large Construction Projects 7 1.36% Forecasting in Construction Projects 5 0.92% Labour Deployment & Productivity in Construction Project 5 0.92% School & Housing Construction Projects 5 0.92% Cash Flow Predictions in Construction Projects 4 0.78% Claims in Construction Projects 4 0.78% Financial Crisis & Management in Construction 4 0.78% Recruitment in Construction Projects 4 0.78% Sustainable Construction 4 0.78% Construction Plant Maintenance 3 0.58% Environmental Impact Assessment in Construction 3 0.58% Heavy Construction Equipment 3 0.58% Information Management in Construction Projects 3 0.58% Stone Crushing and Cutting in Construction 3 0.58% Value Engineering in Construction 3 0.58% Bridge Construction Projects 2 0.39% Construction Management Research 2 0.39% Construction Supply Chain Management 2 0.39% Fibre Optic Technology in Construction 2 0.39% Procurement Selection System in Construction 2 0.39% Recycling Construction Waste 2 0.39% Welding in Construction Projects 1 0.19% Total 514 100.00% Table 7: Context Type and Frequency of Publications between 1990 and 2012 4.7 Citation Analysis Citation analysis was conducted to determine the research impact of the most influential research studies. With regard to recording the citation data pertaining to all 514 articles, the authors started extracting this information from Scopus as of 02 December, 2012 onwards. These data were successively updated until submission. Data gathered from the Scopus database on total citation count per article indicates that 76 articles received 20 or more citations (with one paper receiving 127 citations), 77 papers received between 10 and 19 citations, while 278 papers received between nine to one citation. The second largest portion of the data collected (i.e. 78 papers) received no citation. 18 Most of the articles with no citation were published between 2008 and 2011, with a few articles between 2000 and 2007. Citation frequencies for all the papers are presented in Table 7. In total, 43 studies with larger values of citation counts (with a maximum count of 127 and a minimum count of 30) from each year are listed in Table 8 which includes the study with the largest count by Chua (1999) with a citation count of 127 (as per Scopus database). As noted by Dwivedi et al., (2008, 2009), older papers are more likely to have larger numbers of citations, while newer papers are likely to have lower citation counts. This can be shown by the fact that papers possessing the largest number of citations were published in early volumes of the selected journals and very few of the papers from a relatively recent volume had a large citation count. This is not an exceptional case as a similar tread has been identified in previous studies, including the profiling of the Journal of Electronic Commerce Research (Dwivedi et al., 2008) and Information Systems Frontiers (Dwivedi et al., 2009). The high citation count (e.g. specifically for those papers with a citation count between 127 and 40) may reflect the interest generated in their respective topic by intelligent systems practitioners and academics. It may also be due to the high level of project problems in the construction industry reported in a number of studies (evaluated as part of this research) and which might have attracted academics to re- examine intelligent systems related topics and related methodological practices. Citation Analysis Scopus Database Citation Counts Number of Studies Scopus Database Citation Counts Number of Studies 127 1 27 3 92 1 26 2 82 1 25 6 70 2 24 5 68 1 23 7 66 1 22 5 54 1 21 1 51 1 20 3 49 1 19 4 48 2 18 8 47 1 17 6 46 3 16 4 45 4 15 8 43 2 14 8 42 1 13 7 41 1 12 9 40 2 11 15 39 1 10 8 38 1 9 11 37 2 8 13 36 3 7 16 35 1 6 21 34 2 5 35 33 1 4 31 32 2 3 41 31 2 2 61 30 2 1 49 28 1 0 78 Total Citation Count 514 Table 8: Frequency of Citation Counts between 1990 and 2012 5. A Comparison between 1990-1999 and 2000-2012: Highlighting the Main Contributions in this Research The findings presented in the previous section are general and based on straight counting of, for example, the number of papers published in each year, the geographical region of each publication, 19 etc. This section provides insights into the relationship between some of the variables analysed in the previous section. The variables that are associated in this section are: geo-spatial coverage and yearly publication (Section 5.1), geo-spatial coverage and type of research methods employed (Section 5.2), geo-spatial coverage and type of intelligent systems employed/discussed (Section 5.3), and the association between the yearly publications and type of intelligent systems employed/discussed (Section 5.4). 5.1 Association between Yearly Publications and Geographic Location According to Figures 3 and 4, the USA is in the lead in conducting research into intelligent systems with C = 54 (45% of a total of 120) papers in the 1990s and C = 66 (55% of a total of 120) papers from 2000 onwards until 2012, followed by Taiwan (with three [5.35% of a total of 56] papers in the 1990s and 53 [94.64% of a total of 56] papers from 2000 onwards), the UK (with C = 13 [25% of a total of 52] papers in the 1990s and 39 [75% of a total of 52] papers from 2000 onwards) and Canada (with 10 [21.27% of a total of 47] papers in the 1990s and 37 [78.72% of a total of 47] papers from 2000 onwards). However, comparing the findings of both decades, it is clearly evident that more research was conducted from 2000 (i.e. with a total of 410 papers) onwards on intelligent systems than during the 1990s (i.e. with a total of 104 papers). Moreover, from the number of publications extracted from the 1990s era, it can be argued that there was less attention given to intelligent systems research or a limited need to product innovative and intelligent solutions as compared to 2000 (and onwards). The latter argument may be considered as factual, particularly as global economic competition has compelled many organisations to explore and develop intelligent solutions to improve the delivery of their products or services. Figure 3: Yearly Publications for each Geographic Location from 1990 to 1999 0 2 4 6 8 10 12 14 16 18 20 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 Australia Canada China Croatia Finland Hong Kong Israel Italy Japan Korea Malaysia Mexico Singapore SriLanka Sweden Taiwan UK USA 20 Figure 4: Yearly Publications for each Geographic Location from 2000 to 2012 5.2 Association between Yearly Publications and the Type of Intelligent Systems Employed According to Figures 5 and 6, there has been a sharp increase in the number of publications on intelligent systems in the construction industry in the past decade. For example, from 1990 to 1999, there was a total of 104 papers published on types of intelligent systems, whereas, from 2000 onwards there were 410 papers published from the total of 514 papers. In the 1990s, expert systems seemed to gain the attention of researchers, with 50 papers generated on expert systems (48.175) out of total of 104 papers, followed by knowledge-based systems (with C = 18, 17.31%) and neural networks (with C = 11, 10.58%). On the other hand, from 2000 onwards, there are 96 publications that have either employed or discussed the significance of genetic algorithms (23.41% of 410 papers), followed by neural networks (C = 86, 20.97% of 410 papers) and knowledge-based systems (C = 61, 18.88% of 410 papers). The increasing trend towards the use of intelligent systems in the construction industry over the last decade clearly indicates that the construction industry is without doubt attempting to deal with its myriad problems. For example, forecasting of project completion dates, litigation, planning construction site layout, cost estimations, issues on building construction projects, evaluating construction projects, predicting the risks of contractor defaults, resource allocation and management in construction projects, sustainability, contractual claims, etc., all make management extremely complex. Thus, the need for appropriate intelligent solutions to solve these construction problems. Construction industry projects are highly information intensive and many factors (such as financial, social, human factors) need thorough attention. In this case, factors for success may vary from project to project. Even though ‘human’ experts can often accomplish a reasonable project result, deficits almost always follow due to managers failing to take all relevant factors into consideration and/or lacking access to all relevant information (Cheng et al., 2009; Cheng et al., 2012). Moreover, the increase in the use of intelligent systems can be attributed to the fact that today’s business environment is progressively transforming to a state of hyper-competitiveness. In this context, construction organisations need to continually explore innovative ways to re-orchestrate their products and services for their customers (Park et al., 2011; Sonmez, 2011; Chen et al., 2011; Mahfouz & Kandil, 2012). 0 10 20 30 40 50 60 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 Australia Canada Chile China Cyprus Egypt Finland France Greece Hong Kong India Indonesia Iran Ireland Italy Japan Jordan Korea Lithuania Malaysia Nigeria Norway Oman Philippines Poland Portugal Qatar Romania Saudi Arabia Serbia Singapore South Africa Spain Sri Lanka Sweden Switzerland Taiwan Thailand The Netherland Tunisia Turkey UK USA Vietnam Yugoslavia 21 Figure 5: Type of Intelligent System Employed/Discussed between 1990 to 1999 and 2000 to 2012 0 10 20 30 40 50 60 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 Artificial Intelligence Case-Based Reasoning Context-Aware Systems Decision Support Systems Embedded Systems Expert Systems Fuzzy Systems Genetic Algorithm Human-Machine Interface Intelligent Systems Knowledge-Based Systems Neural Network Sensor Technology Ubiquitous Computing 22 5.3 Association between Geographic Location and the Type of Research Methods Employed According to Figures 5 and 6, analytical research methods have been employed by most of the research studies conducted over the last two decades. In the 1990s there were 25 research studies overall (24.04% of a total of 104 papers) that employed analytical research methods (over 11 regions); with the USA leading with nine papers (i.e. 36% of 25), followed by Canada with six papers (i.e. 24% of 25 papers in total in 1990) and the remaining nine regions with one or two papers each. Following analytical research methods, 28 (26.92% of a total of 104 papers) research studies focused on conceptual/descriptive/theoretical research methods and 21 (20.19% of a total of 104 papers) focused on design research. On the other hand, from 2000 onwards there were 179 research studies (43.66% of a total of 410 papers) that employed analytical research methods (over 34 regions). However, from 2000 onwards, Taiwan had taken the lead in research studies employing analytical research methods (with C = 33 papers i.e. 18.43% of 179), followed by the USA (C = 24 papers i.e. 13.41 of 179), China (C = 19 papers i.e. 10.61 of 179) and Canada (C = 16 papers i.e. 8.94% of 179). Following analytical research methods, the case study method was in second place (with C = 85, 20.73% of 410 papers). These 85 papers using case study research methods are spread across 23 regions with the UK and USA in the lead with 14 papers each. In third place are experiments (with C = 33, 8.05% of 410), followed by papers on secondary data analysis (with C = 32, 7.80% of 410) and papers using survey research methods (with C = 25, 6.09% of 410). Figure 6: Research Methods Employed in each Geographic Location from 1990 to 1999 0 10 20 30 40 50 60 Analytical Case Study Conceptual/Descriptive/Theoretical Design Research Experiment Interview Mixed Method Primary & Secondary Data Analysis Secondary Data Analysis Survey 23 Figure 7: Research Methods Employed in each Geographic Location from 2000 to 2012 0 10 20 30 40 50 60 70 Analytical Case Study Conceptual/Descriptive/Theoretical Design Research Experiment Field Trials Focus Group Interview Interview Meta Analysis Mixed Method Observation Primary & Secondary Data Analysis Questionnaire Secondary Data Analysis Survey 24 5.4 Association between Geographic Location and the Type of Intelligent Systems Employed The findings presented in Table 9 are related to the type of intelligent systems employed/discussed in each research paper from 1990 to 2012. These findings illustrate that the USA is the leading research region in intelligent systems throughout this period. The total numbers of papers generated from this region was 120 with a primary focus on expert systems (with C = 25, 20.83% of 120), knowledge-based systems (with C = 23, 19.16% of 120), neural networks (with C = 20, 16.67% of 120) and genetic algorithms with (C = 17, 14.16% of 120). In second place is Taiwan which generated 56 papers, of which genetic algorithms were employed in 16 papers, followed by neural networks and fuzzy systems (with C = 11, 19.64% of 56 each, respectively). In third place is the UK with 52 papers. Papers generated from the UK mainly focused on knowledge-based systems (with C = 14, 26.92% of 52), followed by artificial intelligence (with C = 8, 15.38% of 52). One of the main findings from this analysis is that out of the top 12 regions (as illustrated in Table 9), 10 regions focused on genetic algorithms to generate intelligent solutions to the problems of the construction industry. Following genetic algorithms, in second place are neural networks which have been employed by all the top 10 regions (as illustrated in Table 9). Following the significance of genetic algorithms in the literature (e.g. Leu & Hung, 2002; Long & Ohsato, 2009; Huang & Sun, 2009; Li et al., 2010), neural networks have been the choice of many researchers due to their remarkable ability to derive meaning from complicated or imprecise data that can be used to extract patterns and detect trends that are too complex to be noticed by either humans or other computer techniques (Baalousha & Çelik, 2011; Cheng et al., 2012). A trained neural network can be thought of as an ‘expert’ in the category of information it has been given to analyse. This expert can then be used to provide projections given new situations of interest and answer ‘what if’ questions (Liao, 2004). Geographic Region Type of Intelligent System Employed Total Individual Count of Intelligent System Type for Each Region USA Expert System (25) 120 Knowledge-Based Systems (23) Neural Network (20) Genetic Algorithm (17) Sensor Technology (13) Artificial Intelligence (11) Intelligent System (2) Context-Aware System (2) Ubiquitous Computing (2) Case-based Reasoning (2) Fuzzy Systems (2) Decision Support System (1) Taiwan Genetic Algorithm (16) 56 Neural Networks (11) Fuzzy Systems (11) Artificial Intelligence (8) Case-based Reasoning (5) Knowledge-Based Systems (3) Intelligent System (1) Sensor Technology (1) UK Knowledge-Based Systems (14) 52 Artificial Intelligence (8) Neural Network (7) Expert System (7) Genetic Algorithm (6) Sensor Technology (3) Fuzzy Systems (1) Context-Aware System (1) Decision Support System (1) Fuzzy Reasoning Approach (1) 25 Artificial Intelligence (1) Case-based Reasoning (1) Intelligent Systems (1) Canada Expert Systems (18) 51 Genetic Algorithm (11) Neural Network (11) Knowledge-Based System (3) Artificial Intelligence (3) Case-based Reasoning (2) Intelligent Systems (1) Automated Data Collection Technology (1) Sensor Technology (1) China Genetic Algorithm (8) 31 Neural Network (7) Case-based Reasoning (4) Artificial Intelligence (3) Ubiquitous Computing (2) Sensor Technology (1) Fuzzy Systems (1) Intelligent System (1) Decision Support System (1) Expert System (1) Knowledge-Based Systems (1) Ubiquitous Computing (1) Korea Artificial Intelligence (8) 29 Case-Based Reasoning (6) Genetic Algorithm (5) Ubiquitous Computing (3) Expert System (3) Knowledge-Based Systems (3) Neural Network (1) Hong Kong Genetic Algorithm (8) 25 Knowledge-Based System (6) Neural Network (5) Decision Support System (2) Case-based Reasoning (1) Fuzzy Systems (1) Artificial Intelligence (1) Expert System (1) Turkey Neural Network (5) 15 Knowledge-Based Systems (3) Fuzzy Systems (2) Case-based Reasoning (2) Sensor Technology (1) Decision Support System (1) Expert System (1) Iran Genetic Algorithm (6) 13 Fuzzy Systems (3) Neural Network (1) Intelligent Systems (1) Decision Support System (1) Expert System (1) Egypt Genetic Algorithm (4) 13 Neural Network (3) Expert System (2) 26 Fuzzy Systems (1) Knowledge-Based System (1) Sensor Technology (1) Artificial Intelligence (1) Australia Knowledge-Based Systems (4) 11 Case-Based Reasoning (3) Expert System (2) Artificial Intelligence (1) Human-Machine Interface (1) India Genetic Algorithm (4) 10 Neural Network (3) Expert Systems (1) Artificial Intelligence (1) Fuzzy Systems (1) Singapore Artificial Intelligence (2) 8 Knowledge-Based Systems (2) Neural Network (2) Case-based Reasoning (1) Genetic Algorithm (1) Spain Neural Network (2) 7 Embedded Systems (1) Decision Support System (1) Artificial Intelligence (1) Fuzzy Systems (1) Sensor Technology (1) Saudi Arabia Neural Network (5) 7 Genetic Algorithm (2) Portugal Knowledge-Based Systems (2) 5 Genetic Algorithm (2) Human-Machine Interface (1) Italy Knowledge-Based Systems (1) 4 Expert System (1) Sensor Technology (1) Neural Network (1) Malaysia Expert System (3) 4 Decision Support Systems (1) Sweden Sensor Technology (2) 4 Knowledge-Based Systems (2) The Netherlands Sensor Technology (2) 4 Case-based Reasoning (1) Knowledge-Based System (1) Finland Knowledge-Based Systems (1) 3 Expert System (2) Japan Knowledge-Based Systems (2) 3 Expert System (1) Jordan Genetic Algorithm (3) 3 Lithuania Knowledge-Based System (1) 3 Expert System (1) Intelligent System (1) Poland Neural Network (2) 3 Expert System (1) Greece Decision Support System (1) 2 Knowledge-Based System (1) Ireland Genetic Algorithm (1) 2 Neural Network (1) 27 Qatar Genetic Algorithm (2) 2 Serbia Artificial Intelligence (1) 2 Case-based Reasoning (1) Sri Lanka Knowledge-Based Systems (1) 2 Neural Network (1) Switzerland Expert System (3) 3 Chile Fuzzy Inference System (1) 1 Croatia Expert System (1) 1 Cyprus Artificial Intelligence (1) 1 France Genetic Algorithm (1) 1 Indonesia Intelligent Systems (1) 1 Israel Expert System (1) 1 Mexico Expert System (1) 1 Nigeria Expert Systems (1) 1 Norway Knowledge-Based Systems (1) 1 Oman Knowledge-Based Systems (1) 1 Philippines Genetic Algorithm (1) 1 Romania Knowledge-Based Systems (1) 1 South Africa Sensor Technology (1) 1 Thailand Neural Network (1) 1 Tunisia Case-based Reasoning (1) 1 Vietnam Genetic Algorithm (1) 1 Yugoslavia Artificial Intelligence (1) 1 48 Regions 514 514 Publications Table 9: Association between Geographic Location and Intelligent Systems Employed/Discussed (1990 – 2012) 6. Conclusions This paper aimed to reveal the current state of intelligent systems research published in a number of key journals (extracted from the Scopus database), thereby focusing on the historical trends and current patterns in the use of different types of intelligent system in the construction industry. Using a systematic review of 514 papers, this paper has primarily analysed the type of intelligent systems employed in construction industry related research studies between 1990 and 2012 – thus fulfilling the aim of this research (as indicated in Section 1.1). Figure 2 and Table 5 clearly indicate the historical trends and current patterns in the number of papers published on intelligent systems and the type of intelligent systems employed/discussed in each paper. As per Table 5, the genetic algorithm takes the lead with 100 papers (19.45% out of total 100%), followed by neural networks with 97 (18.87% out of total 100%), and knowledge-based systems and expert systems both with 79 papers each (i.e. 15.37% each, out of total 100%). The extensive use of these four intelligent system types also indicates that in future research studies; academics, researchers and practitioners may focus on these intelligent system types to propose robust intelligent solutions to the ever increasing problems of the construction industry. The papers selected for this profiling study focused on 14 search phrases, ‘Intelligent Systems’, ‘Artificial Intelligence’, ‘Expert Systems’, ‘Fuzzy systems’, ‘Genetic Algorithms’, ‘Knowledge-Based Systems’, ‘Neural Networks’, ‘Context-Aware Systems’, ‘Embedded Systems’, ‘Human-Machine Interface’, ‘Sensing and Multiple Sensor Fusion’, ‘Ubiquitous & Physical Computing’, ‘Case-based Reasoning’ and ‘Construction Industry’. Within this context, employing an established methodology, the paper examined a number of dimensions in the 514 papers analysed: publications in each year; geographic location of the paper; type of publication; research methods employed; type and number of intelligent systems employed; citation analysis; type of journal; and the context in which each article was published. This study has analysed the largest number of papers compared to existing review articles on the theme of ‘Intelligent Systems in Construction’. The intention in conducting this investigation was to provide a useful and usable resource of information for future researchers. The findings of this study clearly indicate a significant increase in the number of papers published on the different types of intelligent system from 1990 to 2011, although there is a sharp decrease in the number of papers in 2012 (owing 28 to the global financial crisis in 2007-2008) with limited research conducted in earlier/later in 2011 and early 2012, resulting in fewer publications in 2012. However, increasing support to the construction industry given by the global corporations in the future is anticipated to encourage researchers and academics to publish more papers on different types of intelligent systems in the construction industry from 2013 onwards. Based on the research conducted and understanding acquired, the authors anticipate that this research may contribute in conducting future research studies. Two areas that may require further attention from academics and researchers in the future and further work may be carried out, e.g.:  According to Table 2, the USA region has taken the lead with the highest number of publications followed by Taiwan (with C = 56), UK (with C = 52) and Canada (with C = 51). From the total of 48 regions reported in this research, 36 regions produced less than 10 publications with 17 regions generating one publication each. Looking at the trends from 1990 to 2012, it is anticipated that more research will be conducted in the years to follow in the top four regions, leaving many regions behind with no substantial or quality research and few publications. Therefore, academics and researchers from these regions may need to explore more avenues for quality research both conceptually and empirically to generate more publications in the area of intelligent systems in the construction industry.  According to Table 5, four types of intelligent systems i.e. ubiquitous computing, context- aware systems, human-machine interface, and embedded systems are among the least employed/discussed intelligent systems in the construction industry amounting to 14 papers (out of a total of 514) only from 1990 to 2012. Nevertheless, a significant body of research exists for each of these specialised intelligent systems areas and a number of contributions have been made towards theory and practice (e.g. Marwedel, 2003; de Vos et al., 2008; Irani et al., 2010; Rezazadeh et al., 2011; Themistocleous et al., 2012). Table 4 results indicate that these four intelligent systems types are still underexplored areas in both the academic and practical fields of the construction industry. Further research in such directions would enrich the body of knowledge in the research areas of ubiquitous computing, context-aware systems, human-machine interface, and embedded systems with a specific focus on the construction industry. 6.1 Implications for Theory and Practice From the theoretical implications perspective, the authors assert that these findings may help new researchers in the intelligent systems field to identify more relevant journals to refer to and in which to publish their work. The review presented in this paper can also help new researchers to strengthen research into construction in intelligent systems by facilitating consideration of less used but useful alternative methodological perspectives. Whereas from the practical implications perspective, the findings in this study clearly indicate that genetic algorithms, neural networks, expert systems and knowledge-based systems appear to be preferred over other intelligent systems available such as sensor technology and case-based reasoning. The main rationale that can attributed to this is that there were more papers on genetic algorithms, neural networks, expert systems and knowledge- based systems compared to sensor technology and case-based reasoning or even context-aware systems and ubiquitous computing. We anticipate that intelligent systems and artificial intelligence researchers and practitioners will find this paper a useful source of information, especially if they wish to learn more about the various facets pertaining to the existing body of published intelligent systems research in different construction related journals. 6.2 Limitation We fully acknowledge that our study has a limitation, and readers and future academics and researchers should be aware of it and indeed interpret the material presented in this paper within the context of the limitation. Using the set of 14 keywords, our search activities were restricted to the number of journals extracted from the Scopus database only, but there are many well-known journals focusing on intelligent systems and artificial intelligence (such as Artificial Intelligence: An International Journal, International Journal of Neural Systems, International Journal of Intelligent 29 Systems, Artificial Intelligence Review, Journal of Intelligent Material Systems and Structures, Journal of Intelligent Manufacturing, IEEE Intelligent Systems, Journal of Intelligent and Robotics Systems Theory and Applications, Knowledge-based Systems, Robotics and Autonomous Systems, Applied Artificial Intelligence, IEEE Intelligent Systems and their Applications, Journal of Intelligent Systems, Web Intelligence and Agent Systems, Journal of Intelligent and Fuzzy Systems, and Intelligent Automation and Soft Computing) which, although indexed in the Scopus database, did not return any papers based on the keywords. This will clearly have limited our ability to identify all relevant papers, although further research is required to determine the extent of the influence of such factors. Even though we consider that this paper has analysed the largest number of papers compared to other existing review papers, we believe that more comprehensive research is required in order to reduce the impact of the existing limitation we have identified in order to provide a greater understanding of the field of intelligent systems research. Acknowledgement The authors gratefully acknowledge the reviewers for their constructive and detailed feedback/comments in the earlier draft of this paper. Their feedback/comments have helped us to improve this paper significantly. References Al-Zawahreh, A. A. S., & Cox, M. A. A. (2009). The practice of operations research in Jordanian industry: a field study. International Journal of Information and Operations Management Education, 3, 178-197. AbouRizk, S. (2010). Role of Simulation in Construction Engineering and Management. Journal of Construction Engineering and Management. 136, 1140-1153 Altman, R. C. (2008). The Great Crash – 2008: The Geopolitical Setback for the West, Accessed Online: http://www.foreignaffairs.com/articles/63714/roger-c-altman/the-great-crash-2008 Arain, F. M., & Pheng, L. S. (2006). A Knowledge-based System as a Decision Making Tool for Effective Management of Variations and Design Improvement: Leveraging on Information Technology Applications. Journal of Information Technology in Construction, 373-392. Baalousha, Y., & Çelik, T. (2011). An integrated web-based data warehouse and artificial neural networks system for unit price analysis with inflation adjustment, Journal of Civil Engineering and Management, 17(2), 157-167. Beavers, J. E., Moore, J. R., Rinehart, R., & Schriver, W. R. (2006). Crane-Related Fatalities in the Construction Industry. Journal of Construction Engineering and Management. 132, 901-910. Behzadan, A. H., Aziz, Z., Anumba, C. J., & Kamat, V. R. (2008). Ubiquitous location tracking for context-specific information delivery on construction sites. Automation in Construction, 17, 737- 748. Bisaillon, S., Cordeau, J. –F., Paporte, G., & Pasin, F. (2011). A large neighbourhood search heuristic for the aircraft and passenger recovery problem. 4OR: A Quarterly Journal of Operations Research, 9, 139-157. Bolduc, M. –C., Renaud, J., Boctor, F., & Paporte, G. (2008). A perturbation metaheuristic for the vehicle routing problem with private fleet and common carriers. Journal of Operational Research Society, 59, 776-787. Boussabaine, A. H. (1996). The use of artificial neural networks in construction management: a review. Construction Management and Economics, 14(5), 427-36. Bowden, S., Dorr, A., Thorpe, T., & Anumba, C. (2006). Mobile ICT support for construction process improvement. Automation in Construction, 15, 664-676. Bresnen, M. & Marshall, N. (2000). Building partnerships: case studies of client–contractor collaboration in the UK construction industry. Construction Management and Economics, 18, 819- 832. Byrd, T. A., & Hauser, R. D. (1991). Expert systems in production and operations management: research directions in assessing overall impact. International Journal of Production Research, 29, 2471-2482. 30 Chaphalkar, N. B., & Patil, S. K., (2012). Decision Support System for Dispute Resolution in Construction Contracts. KSCE Journal of Civil Engineering, 16(4), 499-504. Chen, S. –M., Griffs, F. H., Chen, P. –H., & Chang, L. –M. (2012). Simulation and analytical techniques for construction resource planning and scheduling. Automation in Construction, 21, 99- 113. Chen, Y. Q., Liu, J. Y., Li. B., & Lin, B. (2011). Project delivery system selection of construction projects in China. Expert Systems with Applications, 38, 5456-5462. Cheng, M. –Y., Tsai, H. –C., & Lai, Y. –Y. (2009). Construction management process reengineering performance measurements. Automation in Construction, 18, 183-193. Cheng, M. –Y., Wu, Y. –W., Wu, C. –F. (2010). Project success prediction using an evolutionary support vector machine inference model. Automation in Construction, 19, 302-307. Cheng, M. -Y., Tsai, H. –C., & Sudjono, E. (2012). Evolutionary fuzzy hybrid neural network for dynamic project success assessment in construction industry. Automation in Construction, 21, 46- 51. Coelho, A., & De Brito, J. (2011). Distribution of materials in construction and demolition waste in Portugal. Waste Management & Research, 29(8) 843–853 Courtney, R., & Winch, G., (2002). CIB Strategy for Re-engineering Construction, CIB/UMIST, Available online: http://cibworld.xs4all.nl/dl/priority_themes/Proposal.pdf Cui, Y., & Lu, Y. (2009). Heuristic algorithm for a cutting stock problem in the steel bridge construction. Computers and Operations Research, 15, 612-622. Curry, B., & Moutinho, L. (1991). Expert systems and marketing strategy: An application to site location decisions. Journal of Marketing Channels, 1(1), 23-37. de Vos, H., Haaker, T. I., Teerling, M., & Kleijnen, M. (2008). Consumer Value of Context Aware and Location Based Mobile Services. Proceedings of the 21 st Bled e-Conference e-Collaboration: Overcoming Boundaries through Multi-Channel Interaction, Slovenia, 50-62. Dwivedi, Y. K., Kiang, M., Lal, B., & Williams, M. D. (2008). Profiling research published in the Journal of Electronic Commerce Research. Journal of Electronic Commerce Research, 9, 77-91. Dwivedi, Y. K., Lal, B., Mustafee, N., & Williams, M.D. (2009). Profiling a decade of Information Systems Frontiers’ research. Information Systems Frontiers, 11(1), 87-102. Dwivedi, Y. K., & Mustafee, N. (2010). Profiling research published in the Journal of Enterprise Information Management. Journal of Enterprise Information Management, 23(1), 8-26. Dubois, A., & Gadde, L. –E. (2002). The construction industry as a loosely coupled system: implications for productivity and innovation. Construction Management and Economics, 20, 621- 631. El-adaway, I. H. (2008). Construction dispute mitigation through multiagent based simulation and risk management modeling. Ph.D. thesis, Dept. of Civil, Construction, and Environmental Engineering, Iowa State Univ., Ames, IA. Feigenbaum, E. A. (1982). The handbook of artificial intelligence. Pitan. Goh, B. H. (1996). Residential construction demand forecasting using economic indicators: a comparative study of artificial neural networks and multiple regression. Construction Management and Economics, 14(1), 25-34. Guo, H., Li, H., Chan, G., & Skitmore, M. (2011). Using Game Technologies to improve the safety of construction plant operations. Accident Analysis and Prevention, 1-10. Hagan, G., Bower, D., & Smith, N. (2012). Delivery of complex construction multi-projects in contractor-led procurement In: Smith, S.D (Ed) Proceedings of the 28 th Annual ARCOM Conference, 3-5 September 2012, Edinburgh, UK, Association of Researchers in Construction Management, 1121-1131. Harmon, K. J. (2003). Dispute review board and construction conflicts: Attitude and opinion of construction industry members. Journal of Management Engineering, 19(3), 121-125. He, Z. –Q., & Sun, X. –L. (2010). Discrete optimization models and methods for management systems of pavement maintenance and rehabilitation. Journal of Shanghai University, 14, 217-222. Hines, E., Leeson, M., Martínez-Ramón, M., Pardo, M., Llobet, E., Iliescu, D., & Yang, J. (2008). Intelligent Systems: Techniques and Applications. Shaker Publishing. Hua, G. B. (2008). The state of applications of quantitative analysis techniques to construction economics and management. Construction Management and Economics, 26(5), 485-497. 31 Huang, R, -Y., & Sun, K. –S. (2009). A GA optimization model for workgroup-based repetitive scheduling (WoRSM). Advances in Engineering Software, 40, 212–228 Irani, Z., Lee, H., Weerakkody, V., Kamal, M. M., et al., (2010). Ubiquitous Participation Platform for Policy Makings (UbiPOL) - A Research Note. International Journal of Electronic Government Research, 6(1), 78-106. Jiang, S., Jang, W.-S., & Skibniewski, M. J. (2012). Selection of Wireless Technology for Tracking Construction Material using a Fuzzy Decision Model. Journal of Civil Engineering and Management, 18(1), 43-59. Kumara, S., Joshi, S., Kashyap, R., Moodie, C., & Chang, T. (1986). Expert systems in industrial engineering. International Journal of Production Research, 24(5), 1107-1125. Langdon, D. (2012). World Construction 2012: Program, Cost, Consultancy, Accessed Online: http://www.davislangdon.com/upload/WorldConstructionReport_2012.pdf Leu, S. –S., & Hung, T, –H. (2002). A genetic algorithm-based optimal resource constrained scheduling simulation model. Construction Management and Economics, 20(2), 131-141. Li, S. (2007). Agentra: An Internet-based multi-agent intelligent system for strategic decision-making. Expert Systems with Applications, 33, 565-571. Li, S. H. A., Tserng, H. P., Yin, S. Y. L., & Hsu, C. –H. (2010). A production modelling with genetic algorithms for a stationary pre-cast supply chain. Expert Systems with Applications, 37, 8406- 8416. Liao, S-H. (2004). Expert system methodologies and applications – a decade review from 1995 to 2004, Expert Systems with Application, 28(1), 93-103. Liu, H, -T., & Tsai, Y. –L. (2012). A fuzzy risk assessment approach for occupational hazards in the construction industry. Safety Science, 50, 1067-1078. Long, L. D., & Ohsato, A. (2009). A genetic algorithm-based method for scheduling repetitive construction projects. Automation in Construction, 18, 499-511. Lucintel. (2012). Global Construction Industry 2012-2017: Trends, Profits and Forecast Analysis, Accessed Online: http://www.concreteconstruction.net/economic-conditions/global-construction- industry-outlook.aspx Mahfouz, T., & Kandil, A. (2012). Litigation Outcome Prediction of Differing Site Condition Disputes through Machine Learning Models, Journal of Computing in Civil Engineering, 26(3), 298-308. Marwedel. P. (2003) Embedded System Design. Springer. Matsatsinis, N. F., & Siskos, Y. (2003). Intelligent support systems for marketing decisions. Boston: Kluwer Academic Publishers. Munisamy, S. (2009). A Spreadsheet-Based Approach for Operations Research Teaching. International Education Studies, 2, 82-88. Negnevitsky, M. (2005). Artificial Intelligence: A Guide to Intelligent Systems. (2 nd Edi), Pearson Education Limited. Ning, X., Lam, K. –C., & Lam, M. C. –K. (2011). A decision-making systems for construction site layout planning. Automation in Construction, 20, 459-473. Ooshaksaraie, L., Basri, N. E. A., Abu Bakar, A., Maulud, K. N. A. (2012). RP3CA: An expert system applied in stormwater management plan for construction sites in Malaysia, Expert Systems with Applications, 39, 3692-3701. Park, J., Ki, D., Kim, K., Lee, S. J., Kim, D. H., & Oh, K. Y. (2011). Using decision tree to develop a soil ecological quality assessment system for planning sustainable construction. Expert Systems with Applications, 38(5), 5463-5470 Park, M., Lee, H. –S., & Kwon, S. (2010). Construction knowledge evaluation using expert index. Journal of Civil Engineering and Management, 16(3), 401-411. Rao, L., & Osei-Bryson, K-M. (2007). Towards Defining Dimensions of Knowledge Systems Quality, Expert Systems with Applications, 33, 368-378. Rezazadeh, I. M., Wang, X., Firoozabadi, M., & Golpayegani, M. R. H. (2011). Using affective human–machine interface to increase the operation performance in virtual construction crane training system: A novel approach. Automation in Construction, 20(3), 289-298. Ruuska, I., Ahola, T., Artto, K., Locatelli, G., & Mancini, M. (2011). A new governance approach for multi-firm projects: Lessons from Olkiluoto 3 and Flamanville 3 nuclear power plant projects. International Journal of Project Management, 29(6), 647-60. 32 Seydel, J. (2003). Evaluating and comparing bidding optimization effectiveness. Journal of Construction Engineering and Management, 129(3), 285-92. Sonmez, R. (2011). Range estimation of construction costs using neural networks with bootstrap prediction intervals. Expert Systems with Applications, 38, 9913-9917. Sonmez, R., & Rowings, J. E. (1998). Construction labour productivity modelling with neural networks. Journal of Construction Engineering and Management, 124(6), 498-504. Tah, J. H. M., & Carr, V. (2000). A proposal for construction project risk assessment using fuzzy logic. Construction Management and Economics. 18, 491-5 Themistocleous, M., Azab, N. A., Kamal, M. M., Ali, M., & Morabito, V. (2012). Location-Based Services for Public Policy Making: The Direct and Indirect Way to e-Participation. Information Systems Management, 29(4), 269-283. Uraikul, V., Chan, C. W., & Tontiwachwuthikul, P. (2007). Artificial intelligence for monitoring and supervisory control of process systems. Engineering Applications of Artificial Intelligence, 20, 115-131. Vanhoucke, M. (2006). Work continuity constraints in project scheduling, Journal of Construction Engineering and Management, 132(1), 14-25. Voordijk, H., Meijboom, B., & de Haan, J. (2006). Modularity in supply chains: a multiple case study in the construction industry. International Journal of Operations & Production Management, 26, 600-618. WESP – World Economic Situation and Prospects (2012), United Nations Publications, Accessed Online: http://www.un.org/en/development/desa/policy/wesp/wesp_archive/2012wespupdate.pdf William, M. D., Dwivedi, Y.K., Lal, B., & Schwarz, A. (2009). Contemporary trends and issues in IT adoption and diffusion research. Journal of Information Technology, 24, 1-10. Winch, G. M. (2010). Managing Construction Projects. Blackwell Publishing Ltd and 2002 Blackwell Science Ltd. 
work_j3yvpi6xnzaq3ckkwhl4yj4mb4	----	Methods of expert estimations concordance for integral quality estimation M. P. Kuznetsova,∗, V. V. Strijovb aMoscow Institute of Physics and Technology, Institutskiy lane 9, Dolgoprudny city, Moscow region, 141700, Russia bNational Research University Higher School of Economics, 20 Myasnitskaya Ulitsa, Moscow 101000, Russia Abstract The paper presents new methods of alternatives ranking using expert estima- tions and measured data. The methods use expert estimations of objects quality and criteria weights. This expert estimations are changed during the computa- tion. The expert estimation are supposed to be measured in linear and ordinal scales. Each object is described by the set of linear, ordinal or nominal crite- ria. The constructed object estimations must not contradict both the measured criteria and the expert estimations. The paper presents methods of expert esti- mations concordance. The expert can correct result of this concordance. Keywords: expert estimations, integral quality, object ranking, preference learning 1. Introduction To make decisions about managed objects, for example, about nature pro- tected areas or state regions, one must rank this objects according to an integral quality estimations, or, equivalently, construct a binary preference function over the set of objects. To construct the integral quality estimation several steps must be performed [1]. 1. A quality criterion must be chosen to compare the objects. 2. A set of objects must be selected according to the quality criterion. 3. The expert must select a set of features describing the objects. 4. A design matrix ”objects-features” must be fulfilled. 5. The expert estimations of the objects quality and the criteria weights must be collected. Further we suppose that multicollinearity of the criteria is not significant. ∗Corresponding author Email addresses: mikhail.kuznecov@phystech.edu (M. P. Kuznetsov), strijov@ccas.ru (V. V. Strijov) Preprint submitted to Expert Systems with Applications July 29, 2013 Inna Typewriter Kuznetsov M.P., Strijov V.V. Methods of expert estimations concordance for integral quality estimation // Expert Systems with Applications, 2014, 41(4-2) : 1988-1996. We consider various scales for the expert estimations [2]: linear, ordinal and nominal. Each scale defines a method of transformation to be applied: for instance, any monotonic transformation can be applied to the ordinal scale. Decision making and preference learning propose several methods to esti- mate the integral quality of objects [3, 4]. Unsupervised methods construct the estimation using the objects description and the quality criterion. This paper introduces the principal component analysis as an example of the unsupervised methods [5, 6]. According to this method, an integral quality estimation is a projection of the objects to a first component of the design matrix. Also Pareto slicing and metric method can be regarded as the unsupervised methods. The supervised methods use expert estimations of the objects quality or the criteria weights [7] besides the design matrix. This paper presents the linear regression method [1], where the target variable is a vector of the expert-given object estimations. We consider linear and ordinal scaled expert estimations. The paper considers a linear model for objects quality estimation [8]. The expert estimations of the criteria weights and objects quality are measured in the linear or ordinal scale. In general, model-computed object estimations doesn’t equal expert-given estimations. This means that expert estimations and model- computed estimations contradict each other [9, 10]. We propose the method of the expert estimations concordance. We consider three various cases corre- sponding to the different types of measurement scales. The first case considers linear scaled both expert estimations and measured data. We propose the estimations concordance method as follows. A set of the admissible expert estimations is a segment restricted by the maximum and minimum value of the estimation. The model uses a structure parameter to find the solution as a point of this segment. The second case considers expert estimations of objects and criteria weights to be ordinal scaled. Ordinal-scaled expert estimations define a convex polyhe- dral cone. The design matrix defines a linear mapping of this cone to object space. The mapped cone can intersect a cone defined by the expert estimations of the objects. In this case the expert estimations of criteria weights and ob- jects supposed to be concordant. In the converse case, we present a method of ordinal concordance. The method minimizes a distance between the vectors in the cones. The third case considers ordinal scaled criteria [11]. The proposed method of objects quality estimation is as follows. Each criterion corresponds to the convex polyhedral con in objects space. According to the linear model, an admissible set of the object estimations is a Minkowski sum of the corresponding cones [12, 13]. The computed estimation is a projection of the expert estimation to the admissible set of values. The data: Nature Protected Areas’ Annual Report. We use the report to illus- trate the proposed methods. Table 1 shows a part of the data. The problem is to estimate efficiency of each NPA using the measured data and the expert estimations. 2 Table 1: Nature Protected Areas report with the expert object qualities and criteria weights estimations O b je ct n u m b er N a ti o n a l p a rk n a m e E x p e r t e s ti m a ti o n s o f o b je c t q u a li ti e s N u m b er o f en v ir o n m en ta l ex p er ti se s N u m b er o f in d iv id u a l g ra n ts N u m b er o f P h D s N u m b er o f em p lo y ee s N u m b er o f o rg a n iz a ti o n s a t th e N P A te rr it o ry N u m b er o f jo u rn a l p a p er s N u m b er o f jo u rn a l a u th o rs N u m b er o f st u d en ts 1.00 0.85 0.78 0.70 0.69 0.58 0.48 0.29 x1 NPA 1 1.00 3 0 1 4.5 3 3 2.5 0 x2 NPA 2 0.83 1 7 1 8 2 8.5 5 40 x3 NPA 3 0.67 2 1 1.5 9 1.5 9.5 6.5 66 x4 NPA 4 0.63 1 3 3 5 4.5 18 11 7 x5 NPA 5 0.58 0 12 8 19 11 7.5 11 62 x6 NPA 6 0.50 0 0 2 5 2.5 2.5 3.5 11 x7 NPA 7 0.44 1 5 4 11 20 16.5 15.5 4 x8 NPA 8 0.38 0 0 3 7 1 4.5 2.5 0 x9 NPA 9 0.33 0 6 0 7 1.5 7 4.5 1 x10 NPA 10 0.17 0 0 1 4 2 2.5 3.5 14.5 3 2. The integral quality estimation problem Denote by X = {xij} m,n i,j=1 a design matrix ”objects-features”. The object description is a vector xi, the i th string of the matrix X. The object ranking problem is to find a binary relation ≺ defined on the set of object pairs, xi ≺ xk. To solve this problem we find a mapping f : X → R, where X is a set of object values. A set of values R of the mapping f has a natural binary relation, that is finding a mapping f is sufficient to solve object ranking problem. Denote an integral quality estimation of the object xi by yi. We consider a linear model, where the value of integral quality yi is a linear combination of the criteria, elements of the vector xi: yi = f(w, xi) = n∑ j=1 wjxij. (1) Denote by f a vector of the values of the function f over the set of objects, f = [f(w, x1), ..., f(w, xm)] T = Xw, (2) where w is a vector of criteria weights. Let each criteria be mapped to the scale [0, 1]: xij 7→ xij − min i xij max i xij − min i xij , i ∈ {1, ..., m}, j ∈ {1, ..., n}. This paper considers the case of the full rank of X, rank(X) = n, with m > n. 3. Unsupervised integral quality estimation We will consider the principal component analysis as an unsupervised method for integral quality estimation. This method finds the objects’ projections to the principal component coordinates such that the sum of squared distances between the objects and their projections to the first component is minimum. Consider the orthogonal matrix W from the linear combination ZT = XTW where columns z1, ..., zn of the matrix Z have the maximum sum of variances, n∑ j=1 σ2(Zj) → max, where σ2(Zj) = 1 m m∑ i=1 (zi − z)2, z = 1 m m∑ i=1 zi. Columns of the matrix W are the eigenvectors of the covariance matrix Σ = XTX. The matrix Σ = XTX can be found using singular values decomposition 4 of the matrix XTX. Since Σ = XTX = WΛWT, it follows that ΣW = WΛ is an eigenvectors system of the matrix ΣW. Therefore the vector of object estimations ŷPCA is the projection of the row- vectors of the matrix X to its first principal component, and w is the first row- vector of the matrix W. This vector corresponds to the maximum eigenvalue of the matrix Σ. 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 x6 y6 x7 y7 x8 y8 x9 y9 x10 y10 Number of PhDs N u m b e r o f e m p lo y e e s Figure 1: The principle component method illustration Figure 1 illustrates principal component analysis for objects quality estima- tion. The black points are the NPAs from the table 1 describing by the criteria ¡¡Number of PhDs¿¿ and ¡¡Number of Employees¿¿. The PCA estimations yi of the objects xi are the points projections to the first principal component indicated by the blue line. 4. Supervised integral quality estimation The supervised methods use the model (1), the design matrix X, the expert estimations of the objects y0 or of the criteria weights w0. 4.1. The linear-scaled expert estimations Weighted sum. Consider the linear-scaled expert estimations of the criteria weights w0. The vector of the object estimations is the linear combination, ŷ = Xw0. This is the simplest method of objects estimation. The main drawback is the lack of robustness of the result estimations. The small changes of the expert estimations w0 may cause enormous changes in the result estimation. 5 0 0.2 0.4 0.6 0.8 1 0.5 1 1.5 2 PCA estimation, ŷP C A E s t im a t io n , y 1 = X w 0 x1x2x3 x4 x5 x6 x7 x8 x9 x10 Figure 2: Comparison of the weighted sum method with the principal component method Fig. 2 shows a comparison of the weighted sum estimation y1 = Xw0 with the principal component estimation ŷPCA. Each point is an object, NPA. The x- axis shows the principal component estimation ŷPCA, the y-axis — the weighted sum estimation y1. Expert-statistical method. Consider the linear-scaled expert estimations y0. The method obtains the criteria weights ŵ as the argument of minimum of a distance between the expert estimations y0 and the computed estimations y ′ 0 = Xŵ, ŵ = arg min w∈Rn ∥Xw − y0∥2. The solution of this problem is given by the ordinal least squares method, ŵ = (XTX)−1XTy0. (3) The vector of the object estimations y′0 = Xŵ. This vector is contained in space of the columns of the matrix X and is the nearest vector to the y0. 5. The expert estimations concordance In this section we will consider both expert estimations of the objects y0 and of the criteria weights w0. 6 5.1. Linear concordance of the expert estimations Consider the computed object estimations y1 = Xw0 using the vector w0 and the computed criteria weights w1 = X +y0 using the vector y0. Here the pseudo-inverse linear operator X+ = (XTX)−1XT. In other words, the linear operator X maps the expert estimations w0 to the vector y1, and pseudo- inverse linear operator X+ maps the expert estimations y0 to the vector w1. In general case the computed and the expert-given estimations are different, y1 ̸= y0, w1 ̸= w0. Call the expert estimations y and w concordant if the following conditions hold: y = Xw, w = X+y. (4) Hereafter we find the expert estimations under the conditions of concordance (4). Denote by y′0 = XX +y0 the projection of the vector y0 to the space of the columns of the matrix X. α-concordance method of the expert estimations. To resolve the contradiction in expert estimations let us consider the estimations wα ∈ [w0, w1] and yα ∈ [y1, y′0]. (5) y0 y′0 X+ X y1 w1 w0 X subspace, dim n Objects space, dim mFeatures space, dim n wα yα Figure 3: The α-concordance method illustration The pair of vectors wα, yα for the given α is defined by the following condi- tions, wα = αw0 + (1 − α)X+y′0, yα = (1 − α)y ′ 0 + αXw0. Theorem 1. The vectors wα, yα are concordant (4). Proof. It is easily proved that Xwα = yα and X +yα = wα: Xwα = αXw0 + (1 − α)XX+y′0 = αXw0 + (1 − α)y ′ 0 = yα, X+yα = (1 − α)X+y′0 + αX +Xw0 = (1 − α)X+y′0 + αw0 = wα. 7 Fig. 3 illustrates the α− concordance method. The vectors y0 w0 are the expert estimations of the objects and of the criteria weights. y′0 is the nearest point to the y0 in the n-dimensional subspace of the columns of the matrix X. The pair of vectors yα, wα is the concordant pair of the expert estimations. The parameter α defines expert preferences to the expert estimations of the objects versus the expert estimations of the criteria weights. If α tends to zero the expert prefers the estimations of the objects; if α tends to zero the expert prefers the estimations of the criteria weights. One could allow expert to assign the parameter α according to his own preferences. Another way to define the parameter α is to compute it as the argument of minimum of the residuals sum Q, Q = ∥w0 − wα∥ n + ∥y′0 − yα∥ m → min α . (6) γ-concordance method of the expert estimations. The γ-concordance method refuses from the restrictions (5). It finds concordant estimations in the neigh- borhoods of the vectors w0, y ′ 0 as a solution of the optimization problem (6) with a regularization parameter γ2 ∈ [0, +∞): wγ = arg min w∈Rn (ε2 + γ2δ2), where ε2 = ∥w0 − wγ∥2 and δ2 = ∥y′0 − yγ∥2. The solution of this problem is the vector of criteria weights, wγ = (X TX + γ2In) −1(XTy′0 + γ 2w0), (7) and the concordant estimations of objects, yγ = Xwγ. The parameter γ 2 defines expert preferences to the expert estimations of the objects versus the expert estimations of the criteria weights, so as α. y′0 w0 Objects space, dim mFeatures space, dim n X+ X ε2 = ∥w0 − wα∥2 δ2 = ∥y′0 − yα∥2 wα yα Figure 4: The γ-concordance method illustration Fig. 4 illustrates the method of γ−concordance. The radiuses of the circum- circles of the points w0 and y ′ 0 equal ε and δ, respectively. 8 0 0.2 0.4 0.6 0.8 1 0.5 1 1.5 2 α, γ(α) O b je ct es ti m a ti o n s, ŷ y1 y1 y2 y2 y3 y3 y4 y4 y5 y5 y6 y6 y7 y7 y8 y8 y9 y9 y10 y10 α−concordance γ−concordance (a) The change of the object estimations 0 0.2 0.4 0.6 0.8 1 0.74 0.76 0.78 0.8 0.82 0.84 0.86 0.88 α, γ(α) E rr o r v a lu e, Q α−concordance γ−concordance (b) The change of the error function Figure 5: α- and γ-concordance illustration Fig. 5 compares α- and γ-concordance methods. The x-axis shows the values of the parameter α changing from 0 to 1, whereas parameter γ is the function of α, γ = α 1 − α , so γ changes from 0 to ∞. The left figure shows object estimations chang- ing. The blue lines indicate object estimations for α-concordance, the red lines indicate γ-concordance. The extreme cases correspond to the expert estima- tions y′0 and y1 = Xw0, respectively. The right figure shows changing of the error function (6). In the case of γ-concordance this function has a global min- imum corresponding to the optimal value of the parameter γ. Fig. 6 compares the expert-statistical method and the γ-concordance method for the optimal value of γ defined by minimal value of the error function (6). The x-axis shows the expert-statistical estimations ŷOLS. The y-axis shows the γ-concordance estimations ŷγ. 5.2. Ordinal concordance of the expert estimations Let the expert estimations y0, w0 be measured in ordinal scales. It means that the set of the estimations is linearly ordered. Consider a cone in Euclidean space corresponding to this set. Without loss of generality suppose that the cone is described by the set of vectors y ∈ Rm with the following restrictions on the components of y: y1 > y2 > ... > ym > 0; w1 > w2 > ... > wn > 0. The set of such vectors y is described by the system of linear inequalities, Jmy 6 0, 9 0.2 0.4 0.6 0.8 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 OLS estimation, ŷOLS γ -E s t im a t io n , ŷ γ x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 Figure 6: Comparison of the expert-statistical method with the γ-concordance method where Jm is the m × m matrix, Jm =   −1 1 0 · · · 0 0 0 −1 1 · · · 0 0 ... ... ... ... ... ... 0 0 0 · · · −1 1 0 0 0 · · · 0 −1   . In the case of a random order yi1 > yi2 > yim > 0, the matrix of the system will be constructed from the Jm by permutations of the corresponding columns. In the same way, the cone of vectors w corresponding to the expert estima- tions of the criteria weights is described by the system of linear inequalities with the n × n-matrix Jn. Therefore the expert estimations y0 and w0 correspond to the m×m and n× n matrices Jm and Jn, respectively. Correspondence of convex polyhedral cones to the expert estimations. Denote by Y0 and W0 the cones, corresponding to the expert estimations in the space of objects and in the space of features, respectively, Y0 = {y|Jmy 6 0}, W0 = {y|Jnw 6 0}. The linear operator X maps the cone W0 of the expert estimations of the criteria weights w0 to the computed cone XW0. The linear operator XX+ maps the 10 cone Y0 of the expert estimations of the objects y0 to the cone Y′0 = XX +Y0. The cone Y′0 consists of the vectors from the subspace of columns of the ma- trix X. Fig. 7 illustrated the defined cones. W0 X+Y′0 Y0 Y′0 XW0X+ X Objects space, dim mFeatures space, dim n Figure 7: Cones corresponding to the expert estimations The ordinal-scaled expert estimations w0 and y0 are called concordant if the cones XW0 and Y′0 have a non-empty intersection besides the vector 0. In this case we can found vectors ŷ ∈ Y′0 and ŵ ∈ W0 satisfying the concordance conditions (4). Properties of polyhedral cones. Now we introduce the following properties of the polyhedral cones corresponding to the expert estimations. Lemma 2. If two cones have vertices in the origin, their intersection is a polyhedral cone. Proof. A polyhedral cone with the vertex in the origin is described by the sys- tem of linear inequalities. Let the first cone be described by the system X1w > 0 and the second cone — by the system X2w > 0. The intersection of this cones is described by the system with the matrix ( X1 X2 ) . In other words, their inter- section is the polyhedral cone with a vertex in the origin. Lemma 3. The locus Xw is a cone if X is a linear mapping. Proof. For any vector w ∈ W a vector λw ∈ W. Therefore, if y ∈ Y, we get λy = λXw = X(λw) ∈ Y = XW. This completes the proof. It follows that if W is a polyhedral cone, the linear operator X maps it to the polyhedral cone XW. A corresponding pseudo-inverse operator X+ maps the cone Y to the cone X+Y. 11 Lemma 4. W0 ∩ X+Y′0 = {0} ⇐⇒ XW0 ∩ Y′0 = {0}. Proof. Let us note that the cones W0 and Y′0 = XX+Y0 have the same dimen- sion n since rank(X) = n. This means that operators X, X+ imply one-to-one correspondence from the features space to the space of the columns of the ma- trix X. Let the vector w ∈ W0 ∩ X+Y′0 w ̸= 0. Then the corresponding vector Xw ∈ XW0 as well as Xw ∈ XX+Y′0 = Y ′ 0. That is the cones XW0 Y′0 intersect at non-zero vector Xw. Let the vector y ∈ Y′0 ∩ XW0 y ̸= 0. Then the corresponding vector X+y ∈ X+Y′0 as well as X+y ∈ X+XW0 = W0. That is the cones W0 and X+Y′0 intersect at non-zero vector X+y. From Lemma 4 it follows that for any vector wp of the cone Wp = W0 ∩ X+Y′0 there exists a unique vector y′p ∈ Y′p = XW0 ∩ Y′0 such that the vectors wp, y′p satisfy the conditions of concordance 4. Moreover, from the definition of the cone Y′0 it follows that for any vector y′p ∈ Y′p there exists a unique vector yp ∈ Y0 such that y′p = XX+yp. This implies that a concordant pair ŵ, ŷ must satisfy following conditions,   Jnŵ 6 0, ŷ = XX+y, Jmy 6 0. Optimization problem for ordinal-scaled expert estimations concordance. Now we formulate an optimization problem for ordinal-scaled expert estimations con- cordance. We will find the nearest vectors ŵ and y1 in the cones W0 and Y0 as follows: (ŵ, y1) = arg min w∈W0,y∈Y0 ∥X+y − w∥, subject to y ∈ Y0, w ∈ W0, ∥X+y∥ = 1, ∥w∥ = 1, (8) where ∥ · ∥ is the Euclidean metrics in the space Rm. 12 W0 X+Y′0 ŵ X+y1 Features space Figure 8: The nearest vectors of the cones This means that the computed vector of criteria weights ŵ is a monotonic transformation of the vector w0. At the same time the objects estimation ŷ = XX+y1 is the nearest point to the expert estimation y0 from the subspace of the columns of the matrix X. This method illustrated with fig. 8. The problem of the nearest vectors can be solved by maximizing the rank correlation. That is, we will find the vectors ŵ ∈ W0 and y1 ∈ Y0 such that Kendall correlation between ŵ and y1 is maximum: (ŵ, y1) = arg max w∈W0,y∈Y0 ρ(X+y, w) : ∥X+y∥ = 1, ∥w∥ = 1. The algorithm of minimizing distance between vectors in cones. Rewrite the problem (8) as follows: minimize ∥X+y − w∥ subject to (X+y)T X+y = 1 and wT w = 1, Jnw 6 0 Jmy 6 0. To solve this problem we propose an iterative algorithm consequently finding approximations of vectors y(2k), w(2k+1) at every even and odd iteration. Define the vector w(0) = w0 at the iteration k = 0. Denote by a = y (2k) and b = w(2k+1) the solutions of two consequent optimization problems: 2k : 2k + 1 : minimize ∥X+a − w(2k)∥ minimize ∥X+y(2k+1) − b∥ subject to (X+a)T X+a = 1, subject to bT b = 1 = 1, Jma 6 0. Jnb 6 0. While solving the optimization problem, define the constants w(2k) = a(2k−1) and y(2k+1) = b(2k). Since the target function and inequality constraints are convex, the solution will be found for finite number of iterations. Methods of convex optimization 13 to solve this problem are provided in [14]. To solve the problem of the rank correlation maximization we use a genetic algorithm. In the case of a non-trivial intersection of the cones W0 and X+Y0 the solution of (8) is a vector ŵ from an intersection of this cones and a vector ŷ satisfying ŷ = XX+y1 = Xŵ. That is, vectors ŷ, ŵ satisfy the concordance conditions (4). If an intersection of the cones is trivial, the proposed algorithm find the nearest non-concordant vectors. As in the case of linear scales, we present a method of the expert estimations concordance using a structure parameter α, yα = (1 − α)ŷ + αXŵ. Here the vector yα and the corresponding vector wα = X +yα define the cones Yα and Wα, respectively. Furthermore, the intersection Yα ∩ XWα ̸= ∅. As in the case of linear-scaled expert estimation concordance, the parameter α defines expert preferences to the expert estimations of the objects versus the expert estimations of the criteria weights. Below we present a method of con- structing the linear-scaled estimations from the computed ordinal-scaled esti- mations. Stable estimations with respect to the design matrix disturbance. Consider the computed cone Yp = Yα ∩ Wα and the design matrix X. Disturb the elements of this matrix, X∆ = X + ∆, with a normal-distributed noise, ∆ = δI, δ ∼ N(0, σ2). An image of the linear mapping y = X∆w is also normally distributed. According to the hypothesis call a stable solution yp the central vector of the cone Yp under the condi- tion ∥yp∥ = 1. The so-called Chebyshev point yp is a center of an incircle of the cone Yp. Find the maximum distance from the target vector yp to the cone faces as follows: ŷp = arg max yp∈Yp {∥yp − b∥2 : b ∈ Rm \ Yp, ∥yp∥2 6 1}. (9) Fig. 9 illustrates the Chebyshev point yp of the cone Yp. Expert estimation concordance using isotonic regression. Consider the special case of the problem (8) such that the expert estimations of the objects y0 are linearly scaled and the expert estimations of the criteria weights are ordinal- scaled. In this case, the problem can be formulated in the terms of the well- known isotonic regression problem [15, 16]. 14 Yp ŷp Objects space Figure 9: An stable solution: Chebyshev point yp Let w̃ = X+y0. Find the monotonic sequence w1 6 ... 6 wn as the nearest to the vector w̃, ŵ = arg minw∈Rn n∑ i=1 (w̃i − wi)2, wn > ... > w1. To make the expert estimation concordant, rewrite this problem with the struc- ture parameter λ. If λ tends to zero the expert prefers the estimations of the objects; if λ tends to zero the expert prefers the estimations of the criteria weights. Find the vector ŵ such that: ŵ = arg min w∈Rn ( 1 2 n∑ i=1 (w̃i − wi)2 + λ n−1∑ i=1 (wi − wi+1)+ ) . To solve this problem we use an algorithm proposed at [15]. 6. Expert estimations concordance for the ordinal-scaled criteria This section considers the case of the ordinal-scaled criteria. Write the ma- trix X as the concatenation of its columns, X = [χ1, ..., χn]. In the case of the ordinal criteria, the geometric shapes corresponding to the columns of X are cones X1, ..., Xn. As described above, each cone is defined by the system of linear inequalities, Xj = {xj|Jmj xj 6 0}, j = 1, ..., m, with the m × m matrices Jmj . 15 Consider a linear model of the objects quality estimation. In the terms of ordinal scales it means that the admissible set X of the object estimations ŷ is a set consisting of the all possible sums of vectors, X = {x| x = x1 + ... + xn, x1 ∈ X1, ..., xn ∈ Xn}. For the following consideration let us recall definition of the Minkowski sum. The Minkowski sum of two subsets L1 and L2 of the linear space is a set L ′ consisting of all possible sums of vectors from L1 and L2. Call an admissible set for the linear model a Minkowski sum, X = X1 + ... + Xn. To estimate vector of object qualities we construct an admissible set as the Minkowski sum of the convex polyhedra X1, ..., Xn. To do this we use a method from [13]. The proposed method computes a matrix of a system of linear in- equalities describing the sum of the polyhedra. The description of this method is given below. Let two convex polyhedra X1 and X2 be described by the following system of inequalities: X1 = {x1|J1x1 6 b1}, X2 = {x2|J2x2 6 b2}. The Minkowski sum of the polyhedra is the vector x satisfying the following conditions:   x − x1 − x2 = 0, J1x1 6 b1, J2x2 6 b2. Transform the system replacing the variable x1 = x − x2:{ J1x − J2x2 6 b1, J2x2 6 b2. (10) The following lemma describes the Minkowski sum of two polyhedra. Lemma 5. x ∈ X if and only if it exists x2 satisfying (10). This means that to find a vector x one must solve the system of linear inequalities, Cx2 6 d, C = ( −J1 J2 ) , d = ( b1 − J1x b2 ) . To solve this system we use the following version of the Minkowski-Farkas lemma. Lemma 6. Let J and b be a matrix and a vector. The system of linear in- equalities Jx 6 b is solvable iff yb > 0 for any vector y satisfying the following conditions: y > 0, yJ = 0. 16 In our case write the Minkowski-Farkas lemma as following, ∃x2 : Cx2 6 d ⇔ ∀z : CT z = 0, z > 0 → (d, z) > 0. Let V be a fundamental system of solutions (FSS) for this case. Therefore V = ( V1 V2 ) , where Vi is a FSS, corresponding to the matrix Ji. It follows that the condi- tion (d, z) > 0 must be rewritten as VT1 (b1 − J1x) + V T 2 b2 > 0. Denote J = VT1 J1, b = V T 1 b1 + V T 2 b2, and obtain the parameters J, b of the system of inequalities describing the Minkowski sum X1 + X2. To find the non-negative FSS of the system with the matrix V we use a method proposed in [13]. The solution of the concordance problem is the point ŷ nearest to the expert estimation y0 such that ŷ ∈ X . Having constructed the set X , define the com- puted object estimations as the projection PX (y) ∈ X satisfying the following conditions: ŷ = PX (y) = arg min z∈X ∥y − z∥. (11) The projection is unique due to the convexity of the set X . Fig. 10 illustrates a projection of the vector y0 to the admissible set X . X y0 ŷ y1 y2 y3 Figure 10: Projection of the point y0 to the cone X Fig. 11 compares the method of ordinal expert data (9) with the method of ordinal criteria (11). The x-axis shows the Chebyshev point estimation ŷCheb. The y-axis shows the projections ŷCones to the admissible set X . 17 0.2 0.4 0.6 0.8 0.45 0.5 0.55 0.6 0.65 0.7 0.75 Estimation ŷCheb for ordinal experts E st im a ti o n ŷ C o n e s fo r o rd in a l fe a tu re s x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 Figure 11: Comparison of the method of the ordinal expert data with the method of ordinal criteria 7. Computational experiment Results for the real data. Table 2 shows estimations of the Nature Protected Areas obtained by the four proposed methods. The results are comparable. The specific method should be chosen according to some knowledge or assumptions about the data structure. Analysis of the algorithms accuracy. The results of four proposed algorithms are demonstrated in the table 3. As the quality criterion we propose a cor- relation coefficient between expert estimation of objects y0 and the computed estimations ŷ. We use Pearson correlation coefficient, r(y0, ŷ), to measure the Table 2: Integral quality estimations for the Nature Protected Areas Object number ŷOLS ŷγ ŷCheb ŷCones x1 1 2 1 1 x2 5 5 4 3 x3 6 3 2 2 x4 2 6 5 5 x5 3 1 6 6 x6 8 9 9 9 x7 4 4 3 4 x8 9 8 8 8 x9 7 7 7 7 x10 10 10 10 10 18 Table 3: Accuracy of the proposed algorithms ŷOLS ŷγ ŷCheb ŷCones Pearson, r 0.69 0.55 0.6 0.66 Kendall, τ 0.47 0.47 0.38 0.51 quality in the linear scales and we use Kendall correlation coefficient, τ(y0, ŷ), to measure quality in the ordinal scales. Note that the estimation ŷ was com- puted using Leave-One-Out method. Thus we can estimate a generalization ability of the proposed algorithms. The results show the object estimations ŷ computed by the four proposed methods. 1. ŷOLS — estimations computed by the expert-statistical method (3), 2. ŷγ — estimations computed by the γ-concordance method (7), 3. ŷCheb — estimations computed by the Chebyshev point finding (9), 4. ŷCones — estimations computed by ordinal criteria method (11). The results show that Pearson correlation has the maximum value for the OLS- estimation, whereas Kendall correlation is maximum for the ordinal criteria method. Analysis of the algorithms stability. To analyse stability of the proposed algo- rithms we disturb the elements of the matrix X. Consider the matrix X∆ = X+∆, where ∆(i, j) ∼ N(0, σ). We change the standard deviation σ of the dis- turbance from its minimum value σ = 0 to its maximum value σmax. Fig. (12) shows how changes quality criteria for the estimations ŷ = f(ŵ, X∆). The left figure shows the changing of Pearson correlation. For the non-disturbed matrix X the expert-statistical method gives the best result but this result is less stable. The most stable results are indicated with the green line (ordinal criteria) and the black line (ordinal expert data). Software implementation. The proposed methods were realized using MATLAB language. The open access code of algorithms and the computational experiment are located at [17]. The code consists of the two main modules. The module comparison.m performs pairwise algorithms comparison. The results of this module are illustrated at fig. (2), (11), (6). The module test_noise.m tests algo- rithms (9), (7), (11), (3) precision and stability. The results of this module are shown above in this section. The input data structures X, y0, w0 are the design matrix X and the expert estimations of the object qualities and of the criteria weights, respectively. This data correspond to the table 1. 19 0 0.1 0.2 0.3 0.4 0.5 0.2 0.3 0.4 0.5 0.6 Noise std, σ P e a r s o n r OLS Gamma Ordinal features Ordinal experts (a) Pearson correlation, r 0 0.1 0.2 0.3 0.4 0.5 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Noise std, σ K e n d a ll τ OLS Gamma Ordinal features Ordinal experts (b) Kendall correlation, τ Figure 12: Analysis of the algorithms stability for the disturbed matrix X 8. Conclusion The paper presents the methods of objects integral quality estimation based on expert estimations and measured data. Unsupervised and supervised meth- ods are considered. The paper proposes the methods of the linear and ordinal expert estimations concordance. The methods use a structure parameter defin- ing expert preferences to the expert estimations of the objects versus the expert estimations of the criteria weights. The paper presents the method of the stable object estimations construction and the method of the ordinal-scaled objects in- tegral quality estimation. The error analysis is performed. The integral quality estimations are constructed for the Nature Protected Areas’ annual reports. References [1] V. Strijov, G. Granic, et al, Integral indicator of ecological impact of the croatian thermal power plants, Energy 36 (2011) 4144–4149. [2] A. Khurshid, H. Sahai, Scales of measurements: An introduction and a selected bibliography, Quality and Quantity 27 (1993) 303–324. [3] J. Fuernkranz, E. Huellermeier, Preference learning, Springer, 2011. [4] V. F. Lopez, F. de la Prieta, M. Ogihara, D. D. Wong, A model for multi- label classification and ranking of learning objects, Expert Systems with Applications 39 (2012) 8878–8884. [5] I. T. Jolliffe, Principal Component Analysis, Springer, 2002. [6] D. Kim, B.-J. Yum, Collaborative filtering based on iterative principal component analysis, Expert Systems with Applications 28 (2005) 823–830. 20 [7] S. Jullien-Ramasso, G. Mauris, P. B. L. Valet, A decision support system for animated film selection based on a multi-criteria aggregation of referees ordinal preferences, Expert Systems with Applications 39 (2012) 4250– 4257. [8] S. R. Searle, Linear models, John Wiley and Sons, 1997. [9] V. I. Danilov, A. I. Sotskov, Social Choice Mechanisms, Springer-Verlang, 2002. [10] S. D. G., Mathematics and voting, Notices of the American Mathematical Society (4/2008). [11] H. M. Moshkovich, A. I. Mechitov, D. L. Olson, Rule induction in data mining: effect of ordinal scales, Expert Systems with Applications 22 (2002) 303–311. [12] E. Fogel, D. Halperin, Exact and efficient construction of minkowski sums of convex polyhedra with applications, Computer-Aided Design 39 (2007) 929–940. [13] M. Uhanov, Polygons sum construction algorithm, Bulletin of South Ural State University, Series ”Mathematics, Physics, Chemistry” 1 (2001) 39–44. [14] S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University Press, 2006. [15] R. Tibshirani, H. Hoefling, R. Tibshirani, Nearly-isotonic regression, Tech- nometrics 53 (2010) 54–61. [16] H. Hoefling, A path algorithm for the fused lasso signal approximator, Journal of Computational and Graphical Statistics 19 (2010) 984–1006. [17] M. P. Kuznetsov, V. V. Strijov, Algorithms of the integral quality estimation, 2013. URL: http://svn.code.sf.net/p/mlalgorithms/code/PreferenceLearning/. 21 
work_j4jqi43icreh5hh4xm22pbqwky	----	CCC 2018 Proceedings of the Creative Construction Conference (2018) Edited by: Miroslaw J. Skibniewski & Miklos Hajdu DOI 10.3311/CCC2018-137 Corresponding author: Author email: asadi@engr.psu.edu Creative Construction Conference 2018, CCC 2018, 30 June - 3 July 2018, Ljubljana, Slovenia Quantitative ways of measuring client’s preferences: a step toward creating an intelligent architectural design agent Mai Sherifa, Somayeh Asadia,*, Ebrahim Karanb aDepartment of Architectural Engineering, Pennsylvania State University, University Park, PA 16802 USA bDepartment of Applied Engineering, Safety, and Technology, Millersville University, Millersville, PA 17551, USA Abstract Technology-driven and digital modeling developments are undeniably changing design and construction professions across the architecture, engineering and construction (AEC) industry. Today, it is impossible to conceive of an architectural practice without using computer tools right from initial design conceptualization to the creation of construction drawings for a given project. The computer applications delivered new capabilities to automate the design process and offered various possibilities of assistant to the designer. However, they are not smart enough yet to understand clients or end users' (e.g. building occupants) needs or expectations and innovate a building design. Artificial Intelligence (AI) could be a solution to making architectural and building design process less time-demanding and more organized with minimal manual interference. The ultimate aim of this study is to make AI part of AEC’s digital journey, by making an intelligent agent able to both understand client needs and produce building designs. An intelligent design agent is defined as an autonomous entity that perceives client needs and makes design decisions towards achieving client satisfaction while balancing design and technical objectives. There are two challenges with respect to creating intelligent design agents; (1) finding quantitative ways of measuring client’s needs and preferences, and (2) providing a mathematical framework for making design decisions. This study focuses primarily on the first challenge and proposes a solution for creating measurable data about the client’s needs and preferences. First, an overview of technologies that support quantitative measurement of people's physical, emotional, and behavioral characteristics is presented. These technologies include eye tracking, facial expressions, electrocardiogram (ECG), electroencephalogram (EEG), electromyogram (EMG), and galvanic skin response (GSR). This overview will enable us to understand potential applications of psychological measurement in the AI domain. This will be followed by a hypothetical design experiment to test the hypothesis that we can determine a subject’s degree of satisfaction by recording and analyzing his or her eye-movements and facial expressions when presented with a collection of visual data, like window design options on a screen. The participants are provided with window design options and asked to evaluate each design based on their preference by giving it a score. While study subjects complete evaluation tasks for each design option, we record their interactions using eye-tracking and an automated facial expression tool. The objective is to find a relationship between subjects’ preferences (as the independent variables) and the emotion expressed and the time spent on each design (as the dependent variables). This will allow us to provide a first look at how subjects interact with different attributes. © 2018 The Authors. Published by Diamond Congress Ltd., Budapest University of Technology and Economics Peer-review under responsibility of the scientific committee of the Creative Construction Conference 2018. Keywords: Eye-tracking; Facial expressions; End-user sstisfaction, Building design 1059 1. Introduction The first stage of the design process is evaluating the clients’ needs and requirements but the final design does not necessarily meet users’ satisfaction and expectations. Although the integration of computers and various revolutionary programs massively aid in the building design process and help avoid multiple issues that may arise, the design process is not entirely computerized yet. The integration of building end users in early design stages has been addressed and highlighted by various researchers [1]; however, the automation of that process has not yet been studied in detail. Building occupants are people who spend most of their times indoor and do not necessarily have the knowledge for designing the building; yet, they may have preferences about the way it has been designed and built [2]. The computer applications delivered new capabilities to automate the design process and offered various possibilities of assistant to the designer. However, they are not smart enough yet to understand clients or end users' (e.g. building occupants) needs or expectations and innovate a building design. Self-learning systems, known as artificial intelligence (AI) are changing the way architecture is practiced, as they do our daily lives, whether or not we realize it. In recent years, AI has been used in lots of projects since it is now able to show emotion but its application in design is still less explored but gradually gaining impetus. One of challenges is a design should be visually eye-catching to consider it as a good design. To overcome this challenge and integrate user’s perception in the design process, the use of eye tracking technology as well as facial expression has been suggested. Eye tracking investigates user's perception of real-time reaction to design, while conventional methods (i.e., interviews, focus group, questionnaires, etc.) have generally failed since they depend on users' willingness and competency to express their feeling about that particular design. Recent studies found that user's needs is not solely about functionality, and they have transformed to more empirical perspectives, not only encompass usability, but also socio-cognitive and emotional aspects of user experience such as joy, surprise, sadness, positive feelings interacting with design, etc. [3]. To this end, the ultimate aim of this study is to make AI part of architecture, engineering and construction (AEC)’s digital journey, by making an intelligent agent able to both understand client needs and produce building designs. AI has the potential to influence the architectural design process at a series of different construction stages, from site plan to building operation and could be a solution to making architectural and building design process less time-demanding and more organized with minimal manual interference. In the context of architectural design, AI can be designed as the use of computers to simulate architect’s intelligence and perform activities normally considered to require architectural knowledge. Although there are various different definitions of AIs, a common view of an intelligent agent is a system that perceives its environment and takes actions that maximize its chances of success [4]. The architectural design environment is faced with a variety of design considerations and client’s expectations. A schematic of the intelligent agent model is provided in Fig.1. The perception of an intelligent architectural design agent is the capability to interpret client’s expectations in a manner that is similar to the way humans use their senses to relate to the world around them. Currently, there is no accurate and through definition of great design. However, if we can measure and quantify users’ emotional and biological data as they interact with buildings, it can be included in the design. The use of eye tracking technology to measure users’ satisfaction seems to be a suitable approach to achieve this objective. In particular, this study compares eye fixations as well as facial expressions to the scores provided by the participants in order to understand whether user’s satisfaction can be accurately predicted by the use of this technology. Fig. 1. Schematic of the intelligent architectural design agent CCC 2018 Proceedings DOI 10.3311/CCC2018-137 1060 2. Background Many studies have explored the automation of building design; however, the complete digitalization of the building design process has not yet been achieved. Although technology driven developments are shaping the fundamental cores of the AEC industry, there still lays a gap in achieving total client satisfaction. Several of the technologies used in the Human-Computer Interaction (HCI) domain can facilitate the ability of machines to process client’s needs and understand them. These technologies include eye tracking, facial expressions, electrocardiogram (ECG), electroencephalogram (EEG), electromyogram (EMG), and galvanic skin response (GSR). The use of eye tracking technology was previously explored in relevance to different applications, some of which examined the HCI as well as user’s satisfaction with web design. Sharma et al. (2014) highlighted the issues, scope, and limitations of the implementation of this technology to test a specific interface. They discussed the level of usability of eye tracking technology as a method HCI and summarized the different measurements of eye tracking techniques, such as fixations, saccades, pupil size, and blink rate. They found that eye tracking systems were proved to be resourceful in usability testing and an indication of satisfaction which make it possible for HCI aiding in decision making process [5]. Khalighy and Green (2015) developed a methodology for quantifying visual aesthetics in a product design by the use of eye-tracking technology. The obtained results were compared to declared preferences by the contributors. It was found that number, duration, and coordinated of eye fixations can certainly indicate accurate approximations to the product preference and satisfaction, solely based on aesthetics. This proves the plausibility to objectively evaluate designs before being taken into construction and manufacturing processes [6]. Ding et al. (2016) conducted a study to test the feasibility of using eye tracking technology to measure and assess the product design. T wo different tasks were designed for this study in order to analyze the captured attention and its relation to the user’s experience. Two factors were focused on throughout the tasks; first fixation and pupil dilations. It was concluded that time to first fixation, fixation time, and minimized pupil variations can be used to indicate end user preference and satisfaction [7]. The probable correlation between eye movements and end-user perception associated with modern web-based EUD (End-user development) environments were studied by Tzafilkou and Protogeros. The aim of this study was to assess whether end user perception and acceptance can be measured by eye behavior. Overall, the study confirmed the significant bond between fixation on eye position and end-user perception [8]. Furthermore, many studies highlighted the missing gaps of using eye tracking technologies to determine end-user satisfaction. The state of the art problems that were raised from previous researches of eye tracking technology are mentioned by Fengpei (2006). Some of the patterns that appeared in the usability studies were stated including a number of fixation points, time at fixation points, and average fixation time. Some of the concerning matters when dealing with eye tracking technologies that were highlighted are: eye movement data being lost, accurate extraction of eye data, and appropriate interpretation and explanation of the data collected. In the conclusion of this paper, it is advised to conduct future research answering to the previously mentioned issues of eye tracking technology [9]. In support of building end user incorporation in early design phases, the matter of whether an advanced level of user contribution can be used as an approach to develop sustainable energy technologies and buildings was raised by Ornetzeder and Rohracher (2006). Their study focused on user-led innovations managed by high levels of participation from different end-users. The results showed that the involvement of users in the design was extremely helpful and result in better-quality and more sustainable buildings [10]. Moreover, Ochoa and Capeluto (2008) conducted a study to examine the impact of integrating building intelligence to improve building energy performance and user comfort. They mainly highlighting the importance of the decisions made at early design stages rather than those added at the end. The result proved that truly intelligent buildings are a product of a design progression that integrates intelligence as well as end-user needs [11]. Though there are many efforts to participate building end-users in the design process, an efficient method of doing so has not yet been defined. A complete automation and computerization of this process at the conceptual phase of design would assist the achievement of ground-breaking means of efficient, sustainable, and intelligent building designs. A couple of studies discussed the use of eye-tracking and facial expression analysis as a method of evaluating the building end-user perception. Mohammadpour et al. (2015) concluded that higher visual attention was indeed assigned to parts that were more attractive to the [12]. CCC 2018 Proceedings DOI 10.3311/CCC2018-137 1061 Moreover, Amiri et al. (2015) highlighted the fact that the combination of facial expression analysis along with eye- tracking data enhances the accuracy and evaluation of user satisfaction [13]. To conclude from previous studies, the use of eye tracking technologies seems like a suitable way to introduce AI into the current building design process. When mastered, this could aid reach better user-oriented sustainable designs with less effort and time wasted in communicating clients’ needs, requirements, and rework. The addition of facial expression recognition analysis along with the data collected by eye tracking provides the possibility of addressing some of the reliability gaps in the use of the eye tracking technology on its own. The ultimate goal of this study is to create an intelligent automated design process. 3. Methodology The use of eye tracking technologies along with facial expression analysis provides the means to quantify the building’s end-user satisfaction level. The eye-tracking experiment was conducted in a quiet and soft light lab using Tobii X2-30 compact eye-tracker to record eye movement data. The technical specification of the eye-tracker includes gaze accuracy of 0.4 to 0.5-degree range, precision of 0.32 to 0.45 degrees, data rate 30 Hz, freedom of head movement 50 (width) 36 (height) × 90 (depth) cm, calibration 9 points (~15 seconds to complete), monocular and binocular tracking capability. To set up this experiment, the eye tracking software (iMotions 6.1: human behavior research, simplified software) was set on screen recording mode and was linked to the slideshow. A simplified experiment was designed consisting of various options of window designs with minor adjustments in a form of the slideshow as shown in Fig.2. The first slide consisted of directions that explained the process requirements. Each slide compared between two different window designs; both their front view and angled view were shown (refer to Fig. 2). The participant was asked to look at the designs for a set period of time (10 seconds), write down a score from 1-6 (1 being the worst, 6 being the best) for each design (e.g. both preferred and undesirable design options), and then click on the preferred design. According to the participant’s preferred design option, a different set of designs was shown on the next slide. The participant was to repeat the process until the slideshow ends. This is a within-subjects experiment in which every single participant was subjected to every single window design. The facial expression recognition software (i.e., iMotions – Emotient) analyses the pattern of wrinkles and crevices based on a visual database of facial expressions and shows seven basic emotion indices including anger, contempt, disgust, fear, joy, sadness, and surprise and two aggregated emotion indices; positive and negative. Positive emotions are defined as emotions that are positively correlated with the evaluation scores and negative emotions are the ones with negative correlation. The Emotient engine (formerly known as FACET and previously CERT) automates Paul Ekman’s FACS coding system (Ekman and Rosenberg 1997) and uses a neural network to analyse the pattern of wrinkles and crevices created by different facial expressions. Main outputs are: 7 basic emotions, 2 advanced emotions including confusion and frustration, 33 facial landmarks such as corners of the eyes, the tip of the nose, and mouth, and operational distance of 70 cm. The emotion index scores ranged from 0 to 5, higher numbers on the emotion index indicate higher levels of feeling (e.g., happiness). a b Fig. 2. (a) experiment slideshow; (b) instructions shown after a set period of time After 10 seconds CCC 2018 Proceedings DOI 10.3311/CCC2018-137 1062 To perform this experiment, first a pilot study was conducted to ensure accuracy. To start, the participant had to provide their name, age, and gender. Next, the eye tracker device was connected to the software and made sure to be working. Then, the calibration process was done to ensure the participant’s eye movements were accurately recorded. When the calibration process was successful, the slideshow started. Throughout the experiment, the tracker device collected data, and so did the camera for the facial expressions data. Twenty people of different ages and genders participated in this experiment as shown in Fig.2. Out of 20 participants, half of them were male and the other half was female. The average age of the participants was 24 with the standard deviation of 3.9. All participants had normal or corrected-to-normal vision. Since the data was recorded as a video, areas of interest (AOI) were defined before the experiment in order to analyse eye-movements. AOI(s) are relevant areas of the stimulus and are closely connected to the research question and the corresponding research design. For each AOI, several measures were calculated using iMotions: time to first fixation (time from the start of the design displayed until the participant fixated on the AOI for the first time), total fixation duration (duration of all fixations within an AOI), fixation count (number of times that a participant fixated on an AOI) and revisit (number of times that a participant returned to an AOI), and pupil diameter and dwell time (the sum time of fixation and saccade) for both views (i.e., front and side). More details on the eye movement measurements and experiment setup can be found in Yousefi et al. (2015). The measured eye-movement data was then compared to the scores given by the participants whilst conducting the experiment. The facial expression data was also recorded and compared to develop a trend. The correlations between subjective evaluation and eye- movement as well as the t-test analysis were conducted by SPSS 22.0 [14]. The setup of the slides presented to one of the participants based on the preferred designs along with the heat map is shown in Fig.3. Heat map shows the areas with a large amount of interest and the aggregated eye fixation of those areas which is shown in red. The absolute duration heat maps are created by combining data from all of the participants to provide a quick view of average existing viewing patterns. Heat maps clearly show which objects capture participants’ attention in an area. Although the absolute duration heat map demonstrates the average viewing pattern, the heat maps created per each participant for each image present individual visual attention. Fig.3 (a). Heat Map Fig.3 (b). A participant taking the experiment 4. Results and discussions Data on the target variable (i.e., evaluation score) is presented in Table 1. In the experiment, each participant was asked to complete 4 evaluation stages within about 1 minute. The participant was asked to evaluate two design options per stage and label a 6-point evaluation score to each design option, where 6 means the most satisfactory (or excellent) and 1 means the least (or poor). Then they would be guided to continue with the preferred design option (e.g., higher evaluation score) to the next evaluation stage. From Table 1, it can be observed that the average score for design options for all participants was 4.0 while preferred design options received 49% higher scores compared CCC 2018 Proceedings DOI 10.3311/CCC2018-137 1063 to undesirable design options. Another observation is that evaluation scores for preferred design options gradually increased when moving from one stage to another. Table 1. Average values of evaluation scores in the window design experiments Undesirable Design Preferred Design Total Stage 1 3.1 4.7 3.9 Stage 2 3.1 4.7 3.9 Stage 3 3.8 4.9 4.3 Stage 4 2.9 4.9 3.9 Overall Evaluation 3.2 4.8 4.0 The results of two-sample t-test associated with a 95% significant level for the window design experiment is shown in Table 2. The null hypothesis states “there is no difference in means response between eye movement measurements (i.e. fixation and time spent) or emotion indices (i.e., fixation, time spent, joy, positivity, sadness, and surprise) and preferred and undesirable design”. The significance (2-tailed) value for fixation time periods is 0.005 which means that there is a statistically significant difference between the mean number of fixation time for preferred and undesirable designs. Fixations are those times when participants’ eyes essentially stopped scanning about the design option. Since the fixation time average for the preferred design is greater than undesirable design, we can conclude that participants stare more at preferred design options compared to undesirable options. Although the difference in spent time between preferred and undesirable design options was not statistically significant (p>0.05), but participants spent more time on designs with higher evaluation scores. In addition, all the studied emotion indices including joy, sadness, positivity, and surprise are statistically significant at the 95% confidence interval (Table 2). Similar to a picture which is worth a thousand words, our faces can reveal lots of information. Facial expressions may have advanced as efficient methods to transmit feelings and intentions and make it possible to communicate intelligent emotions with a simple raised eyebrow or curl of the lip. Table 2. Quantitative analysis of t-test for different window design experiments Variables Design Evaluation Mean Score Std. deviation F Sig. (2-tailed) Fixation Preferred 7.18 4.03 0.672 0.005 Undesirable 5.48 3.55 Time Spent Preferred 2.12 1.28 0.008 0.078 Undesirable 1.72 1.57 Joy Preferred 1.04 0.38 0.64 <0.001 Undesirable 0.49 0.31 Positivity Preferred 1.07 0.40 0.45 <0.001 Undesirable 0.53 0.32 Sadness Preferred 0.48 0.19 31.8 <0.001 Undesirable 0.77 0.40 Surprise Preferred 0.47 0.19 32.7 0 Undesirable 0.76 0.38 The Pearson correlation between the evolution scores and the eye tracking and emotion indices is shown in Table 3. The correlation between positivity index and participant’s evaluation is higher in compare with other indices (correlation coefficient = 0.633) and statistically significant at 95% confidence level. This means that there is a very close relationship between positivity index and participant’s evaluation scores. In another words, participants showed relatively higher positive emotions for the designs (or components) they liked. The correlation between joy and participant’s evaluation scores is moderately high which indicate a close relationship between them. The negative significant correlation at 95% confidence level exists between both sadness and surprise with participant’s evaluation scores (respectively, -0.432 and -0.467). This means that low evaluation scores were associated with high surprise and participants expressed higher sadness emotions for the designs (or components) they disliked. Both fixation and time spent have a positive correlation with participant’s evaluation scores and are significant at the 95% confidence level. CCC 2018 Proceedings DOI 10.3311/CCC2018-137 1064 Table 3. Comparison of eye tracking and emotion indices with evolution scores Variables Pearson correlation Significance Fixation 0.372 <0.001 Time Spent 0.249 <0.001 Joy 0.578 <0.001 Positivity 0.633 <0.001 Sadness -0.432 <0.001 Surprise -0.467 <0.001 The correlation between eye movement measurements and emotion indices and evaluation score is illustrated in Fig. 4. Correlation plot is the visual presentation of a matrix of Pearson's rank correlation coefficients between each variable. The trend line showed in red indicates the direction between each variable and evaluation score and the black circles show the level of strength between them. A positive association exists between fixation, time spent, joy, and positive emotion with evaluation score while the sadness and surprise have a negative association. In addition, these figures illustrate a linear relationship which means that the points on the scatterplot closely resemble a straight line. A relationship is linear if one variable increases by approximately the same rate as the other variables changes by one unit. a b c d e f Figure 4. Correlation plot between eye movement measurements and emotion indices and evaluation score 5. Conclusion The ultimate aim of this study is to make AI part of AEC’s digital journey, by making an intelligent agent able to both understand client needs and produce building designs. As a first step toward achieving this goal, in this study a quantitative method was used to measure building end-user satisfaction at the early stages of design process. To achieve this objective, twenty people were asked to look at a set of two different window designs at each stage, evaluate both designs, and select a preferred one. Meanwhile, their eye movements were being tracked and recorded in addition to their facial expressions and reactions to the presented designs. A statistical analysis using SPSS software was performed to analyze the collected data. The results of two-sample t-test showed there is a statistically significant difference between the mean number of fixation time for preferred and undesirable designs at 95% confidence interval. This means that participants stare more at preferred design options compared to undesirable options. However, the difference in spent time between preferred and undesirable design options was not statistically CCC 2018 Proceedings DOI 10.3311/CCC2018-137 1065 significant (p>0.05), but it was found that participants spent more time on designs with higher evaluation scores. In addition, all the studied emotion indices including joy, sadness, positivity, and surprise are statistically significant at the 95% confidence interval. The correlation between eye movement measurements and emotion indices and evaluation score is illustrated in Fig. 4. Besides, a Pearson’s correlation study was performed which showed a positive association between fixation, time spent, joy, and positive emotion with evaluation score while the sadness and surprise have a negative association. The combination of the data collected by eye tracking and that from facial expression analysis provides us with a better understanding of how people evaluate designs and rate their level of satisfaction. It provides us with the first step in creating a smarter and digitalized design process. References [1] Negendahl, K. (2015). Building performance simulation in the early design stage: An introduction to integrated dynamic models. Automation in Construction, 54, 39-53. [2] McDonnell, J. (2009). "Collaborative negotiation in design: A study of design conversations between architect and building users." CoDesign, 5(1), 35-50. [3] Leitch, K. A., Duncan, S. E., O'Keefe, S., Rudd, R., & Gallagher, D. L. (2015). Characterizing consumer emotional respons e to sweeteners using an emotion terminology questionnaire and facial expression analysis. Food Research International, 76, 283-292. [4] L. Padgham and M. Winikoff, Developing Intelligent Agent Systems, John Wiley & Sons Ltd., England, 2004, pp.1 - 6. [5] Sharma, K., Jermann, P., & Dillenbourg, P. (2014). How Students Learn using MOOCs: An Eye-tracking Insight. Proceedings of the European MOOC Stakeholder Summit 2014, (2003), 147–154. [6] Khalighy, S., Green, G., Scheepers, C., & Whittet, C. (2015). Quantifying the qualities of aesthetics in product design u sing eye-tracking technology. International Journal of Industrial Ergonomics, 49, 31–43. https://doi.org/10.1016/j.ergon.2015.05.011. [7] Guo, F., Ding, Y., Liu, W., Liu, C., & Zhang, X. (2016). Can eye-tracking data be measured to assess product design?: Visual attention mechanism should be considered. International Journal of Industrial Ergonomics, 53, 229–235. https://doi.org/10.1016/j.ergon.2015.12.001. [8] Tzafilkou, K., & Protogeros, N. (2017). Diagnosing user perception and acceptance using eye tracking in web-based end-user development. Computers in Human Behavior, 72, 23–37. https://doi.org/10.1016/j.chb.2017.02.035. [9] Fengpei, H. (2006). The studies of eye tracking and usability test. 10.1109/CAIDCD.2006.329343. [10] Ornetzeder, M., & Rohracher, H. (2006). User-led innovations and participation processes: Lessons from sustainable energy technologies. Energy Policy, 34(2 SPEC. ISS.), 138–150. https://doi.org/10.1016/j.enpol.2004.08.037. [11] Ochoa, C. E., & Capeluto, I. G. (2008). Strategic decision-making for intelligent buildings: Comparative impact of passive design strategies and active features in a hot climate. Building and Environment, 43(11), 1829–1839. https://doi.org/10.1016/j.buildenv.2007.10.018. [12] Mohammadpour, A., Karan, E., Asadi, S., & Rothrock, L. (2015). Measuring end-user satisfaction in the design of building projects using eye-tracking technology. Congress on Computing in Civil Engineering, Proceedings, 2015–Janua(January), 564–571. https://doi.org/doi:10.1061/9780784479247.070. [13] Amiri, S. S., Masoudi, M., Asadi, S., & Karan, E. (2015). A Quantitative Way for Measuring the Building User Design Feedback and Evaluation, (Thompson 2012), 1813–1819. [14] Yousefi, M. V. (2015). Implementing Eye Tracking Technology in the Construction Process. 51st ASC Annual International Conference Proceedings, 752–759. CCC 2018 Proceedings DOI 10.3311/CCC2018-137 1066 https://doi.org/10.1016/j.ergon.2015.05.011 https://doi.org/10.1016/j.ergon.2015.12.001 https://doi.org/10.1016/j.chb.2017.02.035 https://doi.org/10.1109/CAIDCD.2006.329343 https://doi.org/10.1016/j.buildenv.2007.10.018 
work_j5q5tirszvanva633mw725nnu4	----	doi:10.1016/j.eswa.2005.01.003 Building credit scoring models using genetic programming Chorng-Shyong Ong a , Jih-Jeng Huang a , Gwo-Hshiung Tzeng b,c,* a Department of Information Management, National Taiwan University, Taipei, Taiwan b Institute of Management of Technology, National Chiao Tung University, Ta-Hsuch Rd, Hsunchu 300, Hsinchu 1001, Taiwan c College of Management, Kainan University, Taoyuan, Taiwan Abstract Credit scoring models have been widely studied in the areas of statistics, machine learning, and artificial intelligence (AI). Many novel approaches such as artificial neural networks (ANNs), rough sets, or decision trees have been proposed to increase the accuracy of credit scoring models. Since an improvement in accuracy of a fraction of a percent might translate into significant savings, a more sophisticated model should be proposed to significantly improving the accuracy of the credit scoring mode. In this paper, genetic programming (GP) is used to build credit scoring models. Two numerical examples will be employed here to compare the error rate to other credit scoring models including the ANN, decision trees, rough sets, and logistic regression. On the basis of the results, we can conclude that GP can provide better performance than other models. q 2005 Elsevier Ltd. All rights reserved. Keywords: Credit scoring; Artificial neural network (ANN); Decision trees; Genetic programming (GP); Rough sets 1. Introduction Credit scoring models have been widely used by financial institutions to determine if loan customers belong to either a good applicant group or a bad applicant group. The advantages of using credit scoring models can be described as the benefit from reducing the cost of credit analysis, enabling faster credit decision, insuring credit collections, and diminishing possible risk (Lee, Chiu, Lu, & Chen, 2002; West, 2000). Since an improvement in accuracy of a fraction of a percent might translate into significant savings (West, 2000), a more sophisticated model should be proposed to significantly improve the accuracy of the credit scoring model in this paper. In order to obtain a satisfied credit scoring model, numerous methods have been proposed. Roughly, these methods can be classified to parametric statistical methods (e.g. discriminant analysis and logistic regression), non-parametric statistical methods (e.g. k nearest neighbor and decision trees), and soft-computing 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.01.003 * Corresponding author. Address: Institute of Management of Techno- logy, National Chiao Tung University, Ta-Hsuch Rd, Hsunchu 300, Hsinchu 1001, Taiwan. Tel.: C886 3571212157505; fax: 886 35753926. E-mail address: ghtzeng@cc.nctu.edu.tw (G.-H. Tzeng). approaches (e.g. artificial neural network (ANN) and rough sets). Recently, ANNs are the most popular tool used for credit scoring and has been reported that its accuracy is superior to that of traditional statistical methods in dealing with credit scoring problems, especially in regards to non- linear patterns (Desai, Crook, & Overstreet, 1996, 1997; Mahlhotra & Malhotra, 2003; Jensen, 1992; Piramuthu, 1999). However, on the other hand, ANN has been criticized for its poor performance when incorporating irrelevant attributes or small data sets (Castillo, Marshall, Green, & Kordon, 2003; Feraud & Cleror, 2002; Nath, Rajagopalan, & Ryker, 1997). In order to build an effective discriminant function, two issues should be considered. First, the relationships among attributes and classes may be linear or non-linear. Second, the irrelevant attributes should be removed in order to increase the accuracy of the classification model. In this paper, GP is employed to automatically and heuristically determine the adequate discriminant functions and the valid attributes simultaneously. In addition, unlike ANNs which are only suited for large data sets, GP can perform well even in small data sets (Nath et al., 1997). In order to efficiently obtain the discriminant function, the data set is preprocessed by discretization. Two real-world Expert Systems with Applications 29 (2005) 41–47 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa C.-S. Ong et al. / Expert Systems with Applications 29 (2005) 41–4742 cases will be used below to compare the accuracy rate to other classification models including the logistic regression model, ANN, decision trees and rough sets. On the basis of the results, we can conclude that GP can provide better performance than other models. The rest of this paper is organized as follows. Section 2 describes the models for credit scoring. Discretization and genetic programming are proposed in Section 3. Two real- world examples are used to demonstrate the proposed method in Section 4. Discussions are presented in Section 5 and conclusions are in Section 6. 2. Credit scoring models In this section, we describe three popular models used in building credit scoring models. The first model is logistic regression, which is mostly used for classification problems in the area of statistics. The second model is ANN, which is known for its excellent ability of learning non-linear relationships in a system. The third model is rough sets, which is one kind of induction based algorithms, and has been widely used in classification problems since 1990s. 2.1. Logistic regression Logistic regression model is one of the most popular statistical tools for classification problems. Logistic regression model, unlike other statistical tools (e.g. discriminant analysis or ordinary linear regression), can suit various kinds of distribution functions such as Gamble, Poisson, normal, etc. (Press & Wilson, 1978) and is more suitable for the credit scoring problems. In additional, in order to increase its accuracy and flexibility several methods have been proposed to extend the traditional binary logistic regression model, including multinomial logistic regression model (Agresti, 1990; Aldrich & Nelson, 1984; DeMaris, 1992; Knoke & Burke, 1980; Liao, 1994) and logistic regression model for ordered categories (McCullagh, 1980). Therefore, the generalized logistic regression model is the general form of binary logistic regression model and multinomial logistic regression model. Let a p-dimensional explanatory variables x0Z(x1,x2,., xp) and Y be the response variable with categories 1,2,.,r. Then the multinomial logistic regression model be given by the equation logisticðpÞ Z ln PðY Z jjxÞ PðY Z kjxÞ � � Z x0bj; 0% j% r; j sk (1) where bj is a (pC1) vector of the regression coefficients for the jth variable. Let the last response level be the reference level and then the response probabilities p1,p2,.,pr can be calculated by the equations pr hPðY Z rjxÞ Z ex 0brP r lZ1 e x0bl Z ex 0br ex 0br C P rK1 lZ1 e x0bl Z 1 1 C P rK1 lZ1 e x0b (2) pj hPðY Z jjxÞ Z pr e x0bj ; 1% j% r K 1 (3) where l is a response level, and l Z lðbj; 1% j% r; j skÞ Z Xn iZ1 lnðPðY Z yijxiÞÞ; l 2½1; 2; .; r� (4) is the ln likelihood for the multinomial logistic regression model and {(yi,xi),1%i%n} denotes the sample of n objects. When the category is equal to two, the multinomial logistic regression model reduces to a binary logistic regression model. Although logistic regression model can perform well in many applications, when the relationships of the system are non-linear, the accuracy of logistic regression decreases and ANN has been proposed to deal with this problem. 2.2. Artificial neural network Artificial neural networks were developed to mimic the neurophysiology of the human brain to be a type of flexible non-linear regression, discriminant, and clustering models The architecture of ANN can usually be represented as a three-layer system, named input, hidden, and output layers. The input layer first processes the input features to the hidden layer. The hidden layer then calculates the adequate weights by using the transfer function such as hyperbolic tangent, softmax, or logistic function before sending to the output layer. Combining many computing neurons into a highly interconnected system, we can detect the complex non- linear relationship in the data. The simple three-layer perceptron, which is most used in credit scoring problems, can be depicted as shown in Fig. 1. Recently, ANN has been widely used in credit scoring problems, and it has been reported that its accuracy is superior to the traditional statistical methods such as discriminant analysis and logistic regression (Desai et al., 1996, 1997; Jensen, 1992; Mahlhotra & Malhotra, 2003; Piramuthu, 1999). However, as mentioned previously, ANN has been criticized for its poor performance when existing irrelevant attributes or small data sets. Although many methods have been proposed to deal with the problem of variable selection (Feraud & Cleror, 2002; Nath et al., 1997), it is time waste and makes the model more complicated. In addition, other scholars are criticized Fig. 1. Three-layer neural network. C.-S. Ong et al. / Expert Systems with Applications 29 (2005) 41–47 43 the limitations of its long training process in designing the optimal network’s topology in credit scoring problems (Chung & Gray, 1999; Craven & Shavlik, 1997). 2.3. Rough sets Rough sets, originally proposed by Pawlak (1982), is a mathematical tool used to deal with vagueness or uncer- tainty Compared to fuzzy sets, there are some advantages to rough set theory (Pawlak, Grzymala-Busse, Slowinski, & Ziarko, 1995). One main advantage is that rough sets do not need any pre-assumptions or preliminary information about the data, such as the grade of membership function in fuzzy sets (Grzymala-Busse, 1988). Recently, rough set theory and fuzzy set theory have been used to complement or incorporate (Chakrabarty, Biswas, & Nanda, 2000; Mordeson, 2001; Radzikowska & Kerre, 2002) each other rather than to compete (Dubois & Prade, 1991). More detailed discussion about the process of rough set theory can refer to Walczak and Massart (1999). The original concept of approximation space in rough sets can be described as follows. Given an approximation space apr Z ðU; AÞ where U is the universe which is a finite and non-empty set, and A is the set of attributes. Then based on the approximation space, we can define the lower and upper approximations of a set. Let X be a subset of U and the lower approximation of in A is apr � ðAÞ Z fxjx 2U; U=IndðAÞ3Xg (5) The upper approximation of X in A is �aprðAÞ Z fxjx 2U; U=IndðAÞh X sfg (6) where U=IndðAÞ Z fðxi; xjÞ2U$U; f ðxi; aÞ Z f ðxj; aÞ c a 2Ag (7) Eq. (5) represents the least composed set in A containing X, called the best upper approximation of X in A, and Eq. (6) represents the greatest composed set in A contained in X, called the best lower approximation. After constructing upper and lower approximations, the boundary can be represented as BNðAÞ Z �aprðAÞ K apr � ðAÞ (8) According to the approximation space, we can calculate reducts and decision rules. Given an information system IZ(U, A) then the reduct, RED(B), is a minimal set of attributes B4A such that rB(U)ZrA(U) where rBðUÞ Z P cardð � BXiÞ cardðUÞ (9) denotes the quality of approximation of U by B. Once the reducts have been derived, overlaying the reducts on the information system can induce the decision rules. A decision rule can be expressed as f0q, where f denotes the conjunction of elementary conditions, 0 denotes ‘indicates’, and q denotes the disjunction of elementary decisions. The advantage of the induction based approaches (e.g. rough sets and decision trees) is that it can provide the intelligible rules for decision-makers (DMs). These intelli- gible rules can help DMs to realize the contents of data sets. Although these induction methods have been well devel- oped and successfully used in credit scoring problems (Ahn, Cho, & Kim, 2000; Beynon & Peel, 2001; Dimitras, Slowinski, Susmaga, & Zopounidis, 1999), the main problem of induction based methods is the ability of forecasting. It is clear that if a newly entered object does not match any rule, it cannot be determined which class it belongs to. Next, we described the concepts of GP which is used here to build the credit scoring models in Section 3. 3. Genetic programming Genetic programming was proposed by Koza (1992) to automatically extract intelligible relationships in a system and has been used in many applications such as symbolic regression (Davidson, Savic, & Walters, 2003), and classification (Stefano, Cioppa, & Marcelli, 2002; Zhang & Bhattacharyya, 2004). The representation of GP can be viewed as a tree-based structure composed of the function set and terminal set. The function set is the operators, functions or statements such as arithmetic operators ({C,K,!,j}) or conditional statements (If.then.) which are available in the GP. The terminal set contains all inputs, constants and other zero-argument in the GP tree. For example to express xyC3/x, the GP tree can be represented as Fig. 2. Once we initialize a population of the GP tree, the following procedures are similar to genetic algorithms Fig. 3. The crossover operator of GP tree. Fig. 2. The representation of a GP tree. C.-S. Ong et al. / Expert Systems with Applications 29 (2005) 41–4744 (GAs) including defining the fitness function, genetic operators such as crossover, mutation and reproduction, and the termination criterion, etc. Next, we introduce three main operators, crossover, mutation and reproduction, to show the procedures of finding the (approximate) optimal generation. In GP, the crossover operator is used to swap the subtree from the parents to reproduce the children using mating Table 1 The discretization of the continuous attributes in Australian data set Value 1 2 3 4 5 6 A2 [*, 17.75) [17.75, 21.21) [21.21, 22.38) [22.38, 23.04) [23.04, 23.34) [2 24 A3 [*, 0.480) [0.480, 1.793) [1.793, 2.793) [2.793, 4.020) [4.020, 6.103) [6 A7 [*, 0.145) [0.145, 1.020) [1.020, 2.145) [2.145, *) A10 [*, 1) [1, *) A13 [*, 23) [23, 93) [93, 171) [171, 262) [262, *) A14 [*, 13) [13, *) selection policy rather than exchanging bit strings as in GAs. An example of a crossover in GP is shown in Fig. 3. Similar to GAs, GP uses the mutation operator in order to avoid falling into the local optimal solution. The mutation operator is used to randomly choose a node in a subtree and replace it with a new created subtree randomly. Finally, a new generation can be reproduced from two parents using the reproduction operator to represent a better solution. In order to determine the adequate discriminant function, the fitness function of GP can be described as ffi Z P N i absðoi K eiÞ N (10) where abs($) denotes the absolute operator, oi denotes the observed class and ei denotes the expected class. It should be highlighted that the function set and the terminal set should be varied enough to represent the relations among independent and response variables. More- over, in order to satisfy the principle of parsimony, the depth of the GP-tree should also be limited. In addition, in order to obtain the discriminant function efficiently and effectively, discretization of continuous attributes should be employed before GP. Many discretiza- tion algorithms such as Boolean reasoning algorithm, entropy algorithm and naı̈ve algorithm have been proposed to deal with this problem (Shan, Hamilton, Ziarko, & Cercone, 1996; Wu, 1996). In this paper, Boolean reasoning algorithm is employed to determine the adequate discrete values. Next, two empirical cases will be used follow to compare the proposed method and the other models. 4. Empirical analysis In this section, GP is compared to MLP, classification and regression tree (CART), C4.5, Rough sets, and logistic regression (LR) using two-real world data sets. The first data set includes Australian credit scoring data with 307 examples of credit worthy customers and 383 examples for credit unworthy customers. It contains 14 attributes, where six are continuous attributes and eight are categorical attributes. The second data set, called the German Credit Data Set, was provided by Prof. Hofmann in Hamburg. 7 8 9 10 11 3.34, .38) [24.38, 27.92) [27.92, 32.38) [32.38, 37.38) [37.38, 48.96) [48.96, *) .103, *) Table 2 The parameter settings of GP Parameter Value Population size 40 Fitness function Eq. (10) Function set {C,K,!,sin,cos,R,Z,%,and,or,not,if} Terminal set {Attributes,1,2,3,4,5,6,7,8,9,10,11,12,13,14} Maximum number of generation 1000 Selection Lexictour Crossover rate 0.9 Mutation rate 0.01 C.-S. Ong et al. / Expert Systems with Applications 29 (2005) 41–47 45 It includes customer credit scoring data with 20 features, such as age, gender, marital status, credit history records, job, account, loan purpose, other personal information, etc. There are 700 records judged to be credit worthy and 300 records judged to be credit unworthy. Both data sets are made public from the UCI Repository of Machine Learning Databases, and are mostly used to compare the performance of various classification models. The first step of the proposed method is to dissect the continuous attributes. For example of Australian data set, the results of the discretization can be shown as in Table 1. The discretization of the continuous attributes in German data set can be described as shown in Appendix A. Next, we set the GP parameters of Australian data set as shown in Table 2 and the parameters of German data set can also be shown in Appendix B. In order to build the discriminant function as flexible as possible, we incorporate the logic operators into the function set. On the other hand, due to the range of the discretization values is from 1 to 14, we incorporate the constants from 1 to 14 into the terminal set. Table 3 The comparison of the credit scoring models in Australian data set Australian data Sample 1 (%) Sample 2 (%) Sample 3 (% GP 0.1111 0.1280 0.1304 MLP 0.1352 0.1256 0.1352 CART 0.1497 0.1256 0.1449 C4.5 0.1594 0.1304 0.1400 Rough sets 0.1382 0.1729 0.1538 LR 0.1497 0.1449 0.1304 Table 4 The comparison of the credit scoring models in German data set German data Sample 1 (%) Sample 2 (%) Sample 3 (% GP 0.2166 0.2266 0.2200 MLP 0.2400 0.2382 0.2500 CART 0.2765 0.2617 0.2435 C4.5 0.2446 0.2500 0.2227 Rough sets 0.2533 0.2649 0.2631 LR 0.2400 0.2421 0.2500 Five sub-samples are used to compare the error rate of the credit scoring models. In addition, the holdout method is used for avoiding the problem of overfitting. The error rate of the test sets in both Australian and German data sets can be described as shown in Tables 3 and 4. On the basis of the results, we can conclude that the proposed method outperforms to other models in our empirical analysis. In addition, ANN and logistic regression also well perform in this study and can be other choices for the credit scoring model. Next, we provide the discussions based on our implementation. 5. Discussions Due to the huge growth rate of the credit industry, building an effective credit scoring model have been an important task for saving amount cost and efficient decision making. Although many novel approaches have been proposed, more issues should be considered for increasing the accuracy of the credit scoring model. First, the irrelevant variables will destroy the structure of the data and decreases the accuracy of the discriminant function. Second, the credit scoring model should determine the correct discriminant function (linear or non-linear) automatically. Third, the credit scoring model should be useful in both large and small data sets. For above reasons, GP is used to build the credit scoring models in this paper. On this basis of the simulated results, we can conclude that GP outperforms than other models. However, ANN and logistic regression can also provide the satisfied solutions and can be other alternatives. The accuracy of the induction based approaches (decision trees and rough set) is inferior in this study. It is clear that the decision rules are derived from ) Sample 4 (%) Sample 5 (%) Overall 0.1207 0.0966 0.1173 0.1062 0.1014 0.1207 0.1400 0.1497 0.1419 0.1014 0.1159 0.1294 0.1718 0.1777 0.1628 0.1304 0.1352 0.1381 ) Sample 4 (%) Sample 5 (%) Overall 0.2433 0.2266 0.2266 0.2433 0.2533 0.2449 0.3170 0.3721 0.2941 0.2926 0.3318 0.2683 0.2353 0.2551 0.2543 0.2479 0.2500 0.2460 Table B1 The parameter settings of GP Parameter Value Population size 40 Fitness function Eq. (10) Function set {C,K,!,R,Z,%,and,or,not,if} Terminal set {Attributes,1, 2, 3, 4, 5} Maximum number of generation 40 Selection Lexictour Crossover rate 0.9 Mutation rate 0.01 C.-S. Ong et al. / Expert Systems with Applications 29 (2005) 41–4746 the training set. However, if a newly entered object within the test set does not match any rule, it cannot be determined which class it belongs to. Compared to other models, we consider that GP is more suitable for the credit scoring problems for the following reasons. Unlike the traditional statistical methods need the assumptions of the data set and the attributes, GP is a non- parametric tool and suitable for any situations and data sets. Compared to ANNs, GP can determine the adequate discriminant function automatically rather than assigned the transfer function by decision-makers. In addition, GP can also select the important variable automatically. Finally, the discriminant function which is derived by GP can provide the better forecasting performance than the induction based algorithms. 6. Conclusions Building a credit scoring model involves the problems of variable selection and model identification. Although many approaches have been proposed, a flexible and accurate method is limited. In this paper, GP is employed to build the discriminant function for the credit scoring problems. On the basis of the empirical results, we can conclude that GP is more flexible and performs better accuracy in the credit scoring problems significantly. Appendix A The discretization of the continuous attributes in Germen data set using Boolean reasoning algorithms can be described as shown in Table A1. Appendix B The parameters of German data set can be shown as in Table B1. Table A1 The discretization of the continuous attributes in German data set Value 1 2 3 4 5 Checking [*, 1) [1, 2) [2, *) Duration [*, 12) [12, 23) [23, 32) [38, *) History [*, 2) [2, *) Amount [*, 714) [714, 1387) [1387, 2045) [2045, 3914) [3914, *) Saving [*, 2) [2, *) Employed [*, 2) [2, 3) [3, *) Installp [*, 4) [4, *) Resident [*, 2) [2, 4) [4, *) Age [*, 27) [27, 33) [33, *) Existcr [*, 2) [2, *) Job [*, 1) [1, *) References Agresti, A. (1990). Categorical data analysis. New York: Wiley. Ahn, B. S., Cho, S. S., & Kim, C. Y. (2000). The integrated methodology of rough set theory and artificial neural network for business failure prediction. Expert Systems with Applications, 18(2), 65–74. Aldrich, J. H., & Nelson, F. D. (1984). Linear probability, logit, and probit models. Beverly Hills, CA: Sage. Beynon, M. J., & Peel, M. J. (2001). Variable precision rough set theory and data discretisation an application to corporate failure prediction. OMEGA: the International Journal of Management Science, 29(6), 561–576. Castillo, F., Marshall, K., Green, J., & Kordon, A. (2003). A methodology for combining symbolic regression and design of experiments to improve empirical model building. Genetic and Evolutionary Compu- tation Conference , 1975–1985. Chakrabarty, K., Biswas, R., & Nanda, S. (2000). Fuzziness in rough sets. Fuzzy Sets and Systems, 110(2), 247–251. Chung, H. M., & Gray, P. (1999). Special section: Data mining. Journal of Management Information Systems, 16(1), 11–16. Craven, M. W., & Shavlik, J. W. (1997). Using neural networks for data mining. Future Generation Computer Systems, 13(2/3), 221–229. Davidson, J. W., Savic, D. A., & Walters, G. A. (2003). Symbolic and numerical regression: Experiments and applications. Information Sciences, 150(1/2), 95–117. DeMaris, A. (1992). Logit modeling. Beverly Hills, CA: Sage. Desai, V., Crook, J., & Overstreet, G. (1996). A comparison of neural networks and linear scoring models in credit union environment. European Journal of Operations Management, 95(1), 24–37. Desai, V., Crook, J., & Overstreet, G. (1997). Credit scoring models in the credit union environment using neural networks and genetic algorithms. IMA Journal of Mathematics Applied in Business and Industry, 8(4), 324–346. Dimitras, A. I., Slowinski, R., Susmaga, R., & Zopounidis, C. (1999). Business failure prediction using rough sets. European Journal of Operational Research, 144(2), 263–280. Dubois, D., & Prade, H. (1991). In Z. Pawlark (Ed.), Rough sets: Theoretical aspects of reasoning about data. Dordrecht, The Nether- lands: Kluwer. Feraud, R., & Cleror, F. (2002). A methodology to explain neural network classification. Neural Network, 15(2), 237–246. Grzymala-Busse, J. W. (1988). Knowledge acquisition under uncertain- ty—A rough set approach. Journal of intelligent and Robotic Systems, 1(1), 3–16. Jensen, H. L. (1992). Using neural networks for credit scoring. Managerial Finance, 18(1), 15–26. Knoke, D., & Burke, P. J. (1980). Log-linear models. Beverly Hills, CA: Sage. Koza, J. (1992). Genetic programming: On the programming of computers by means of natural selection. Cambridge, MA: MIT Press. C.-S. Ong et al. / Expert Systems with Applications 29 (2005) 41–47 47 Lee, T. S., Chiu, C. C., Lu, C. J., & Chen, I. F. (2002). Credit scoring using the hybrid neural discriminant technique. Expert Systems with Applications, 23(3), 245–254. Liao, T. F. (1994). Interpreting probability model: Logit, probit, and other generalized linear models. Beverly Hills, CA: Sage. Mahlhotra, R., & Malhotra, D. K. (2003). Evaluating consumer loans using neural networks. OMEGA: The International Journal of Management Science, 31(2), 83–96. McCullagh, P. (1980). Regression model for ordinal data. Journal of the Royal Statistical Society, Series B, 42(2), 109–142. Mordeson, J. N. (2001). Rough set theory applied to (fuzzy) ideal theory. Fuzzy Sets and Systems, 121(2), 315–324. Nath, R., Rajagopalan, B., & Ryker, R. (1997). Determining the saliency of input variables in neural network classifiers. Computers and Operations Researches, 24(8), 767–773. Pawlak, Z. (1982). Rough set. International Journal of Computer and Information Science, 11(5), 341–356. Pawlak, Z., Grzymala-Busse, J., Slowinski, R., & Ziarko, W. (1995). Rough sets. Communications of the ACM, 38(11), 88–95. Piramuthu, S. (1999). Financial credit-risk evaluation with neural and neurofuzzy systems. European Journal of Operational Research, 112(2), 310–321. Press, S. J., & Wilson, S. (1978). Choosing between logistic regression and discriminant analysis. Journal of the American Statistical Association, 73(4), 699–705. Radzikowska, A. M., & Kerre, E. E. (2002). A comparative study of fuzzy rough sets. Fuzzy Sets and Systems, 126(2), 137–155. Shan, N., Hamilton, H. J., Ziarko, W., & Cercone, N. (1996). Discretization of continuous valued attributes in attribute-value systems. Proceeding of the fourth International orkshop on Rough Sets, Fuzzy Sets, and MachineDiscovery, Tokyo, Japan , 74–81. Stefano, C. D., Cioppa, A. D., & Marcelli, A. (2002). Character preclassification based on genetic programming. Pattern Recognition Letters, 23(12), 1439–1448. Walczak, B., & Massart, D. L. (1999). Rough sets theory. Chemometrics and Intelligent Laboratory Systems, 47(1), 1–16. West, D. (2000). Neural network credit scoring models. Computers and Operations Research, 27(11/12), 1131–1152. Wu, X. D. (1996). A Bayesian discretizer for real-valued attributes. The Computer Journal, 39(8), 688–691. Zhang, Y., & Bhattacharyya, S. (2004). Genetic programming in classifying large-scale data: an ensemble method. Information Science, 163(1/3), 85–101. Building credit scoring models using genetic programming Introduction Credit scoring models Logistic regression Artificial neural network Rough sets Genetic programming Empirical analysis Discussions Conclusions Appendix A Appendix B References 
work_j6fancn7djfnfciiozvjebwri4	----	An expert system approach to quality control An expert system approach to quality control E.P. Paladini* Departmento de Engenharia de produção e Sistemas, Universidade Federal de Santa Catarina, CP 476, Trindade, 88010-970, Florianopolis, SC, Brazil Abstract This paper describes the basic actions proposed for quality inspection. These actions involve structuring a Decision Supporting Expert System (DSES) to help with those decisions related to the preliminary activities for inspection development —most of them relating to determining of the need or convenience of carrying out the inspection itself. Once the opportunity to carry it out is defined, the expert system helps the user to select the type of inspection to adopt from amongst: (1) automatic or sensorial inspection; (2) inspection by samples or complete; (3) acceptance or rectifying and, in the most relevant module, (4) inspection by attributes or by variables. The complementary documentation of the DSES contains the directions to operate it, the rules and qualifiers that make up the system, as well as the results achieved through its experimental implementation. q 2000 Elsevier Science Ltd. All rights reserved. Keywords: Expert system; Quality control; Total quality management 1. Introduction Having the current concept of quality in mind (and the context of Total Quality Management itself), the feasibility process of basic quality control actions is nowadays consid- ered to take place within a well-defined structure, known as quality evaluation system. Inspection, in turn, is the most important activity in the quality evaluation system of an industrial process. When correctly developed, the inspection makes possible to carry out a precise analysis of how the process operates and serves as a basis for a set of decisions that directly affect it, such as corrective and preventive actions which must be complied with in order to guarantee acceptable quality levels. Quality inspection has got a number of effective techni- ques with a wide range of applicability. Since there are several techniques available, there are many options for those who intend to devise an inspection process. Never- theless, this situation poses difficulties as to the correct use of the different techniques, since most of such techniques have their own utilization particularities and yield results valid only within certain contexts. The need of choosing the most adequate technique for each situation shows that it is necessary to organize the information related to quality inspection, or else the whole quality evaluation process could be seriously compromised. The present paper highlights this question, which is considered to be a relevant restriction to the perfect use of quality evaluation. A more efficient way of optimizing qual- ity evaluation development is proposed. We dedicate special attention to quality inspection by attributes, a much more difficult area when it comes to making the correct decision about quality evaluation of services, products and processes. The literature about quality inspection is plentiful and easily accessible. More often than not, classical texts on quality control deal with this subject by focusing on proce- dures of sampling for acceptance, whereas rectifying inspection is rather rarely analyzed. The development of the theoretical models of the Operating Characteristic (OC) Curve associated with basic concepts, such as Accep- table Quality Level (AQL), Lot Tolerance Percent Defec- tive (LTPD) and producer’s and consumer’s risk, is always presented as a means of motivating the analysis of sampling schemes or sampling plans and inducing the procedures which aim at structuring such plans and making feasible their performance analysis, mostly in terms of reliability and costs. Tables suggesting models of sampling plans are just as relevant and are present as well in almost all books of the area. The following authors can be cited as classical in this field: Besterfield (1990), Charboneau and Webster (1997), Juran (1999), and Montgomery (1998), to name a few. Texts of renowned authors, such as Dodge and Deming can be found in more specialized journals. A collection of articles by Dodge about inspection by attributes was made available by the American publication called Journal of Quality Technology in 1977 (see Dodge, 1977). As for the work of Deming, articles on this subject written by him can Expert Systems with Applications 18 (2000) 133–151PERGAMON Expert Systems with Applications 0957-4174/99/$ - see front matter q 2000 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 7 - 4 1 7 4 ( 9 9 ) 0 0 0 5 9 - 7 www.elsevier.com/locate/eswa * Tel.: 155-48-331-7000; fax: 155-48-331-7075. E-mail address: paladini@iaccess.com.br (E.P. Paladini). be found in the same journal published in 1985 (Papadakis, 1985). The types of inspection are manifold, each of them bear- ing their particularities, specific purposes and restricted adequacy to certain contexts. Such is the case of sampling inspection or complete inspection, as well as inspection by attributes or by variables, for instance. The choice of the correct type of inspection is extremely relevant for quality evaluation, since the adequacy of the inspection process to the context where it takes place has been considered in many situations and its importance has always been stressed. Let us consider the inspection of raw materials as an example. An incorrect development of this type of inspec- tion can give the suppliers the impression that quality is not important to their customers and of course the suppliers will not have to regard it as relevant. Such a situation, which can be seen in practice, is to be found in several texts. Horsnell (1988), for example, draws attention to this fact, by showing that the use of inspection by attributes when products enter the factory is of such usefulness that it transcends inspection proper, in addition to being an easily applicable technique. In a broader analysis, one can see that the quality of the process can be significantly (and positively) affected by the mere inclusion of inspection in the process (it has been considered many years ago—see, for instance, Whittle, 1964). Quality inspection is widely used nowadays. Nonethe- less, several problems have been encountered upon using inspection systems and such problems derive for the most part from an inadequate application of the prescribed meth- odology. This inadequacy results from decisions which ought to be made as regards the way the inspection is to be carried out. Such decisions are considered here to be intuitive. However, they do deserve detailed investigation. We under- stand that the lack of an adequate methodology for planning the decision-making process is a fundamental restraint to the use of quality inspection. This situation justifies the elaboration of a general Deci- sion-Making Support System. Due to its peculiarities, the system was organized under the format of an expert system, which benefits from the advantages of artificial intelligence (AI) in order to carry out an evaluation. It should be noted that since inspection is the most impor- tant part of Quality Evaluation, the expert system proposed here plays a fundamental role in the Total Quality Manage- ment area. It is easy to see that this paper describes a process of interaction between users and computers in industrial companies. The process becomes feasible through the development and application of an expert system related to the area of quality control. This paper describes the devel- oped expert system, its operating methodology and lists the results obtained after its application in the companies we have studied. The interaction between the human resources of the organizations and the computational resources was intense in the last few years. With the advent of AI this interaction has acquired new, rather specific characteristics. 2. Artificial intelligence applied to quality evaluation A brief review of the specific technical literature reveals that there are many cases of successful applications of Al techniques to the Quality Evaluation area. Such applications were helpful in the resolution of relevant problems. Below are some examples thereof: (a) Hosni and Elshennavy (1988) reported the develop- ment of a quality control system based on knowledge, suitable to specific procedures of inspection by variables and to the selection of control graphs, both by attributes and by variables. (b) Eyada (1990) developed an expert system aimed at auditing procedures of Quality Assurance involving both suppliers and products in the process. Some years earlier, in the same field, Gipe and Jasinski (1986) conducted an analysis of expert system adequacy to Quality Assurance problems, showing the feasibility of this application. Another interesting expert system was developed by Crawford and Eyada (1989) with the purpose of planning the allocation of resources for the Quality Assurance program. (c) There are several expert systems used to select control graphs (see Alexander & Jagannathan, 1986; Dagli, 1990; Dagli & Stacey, 1988). An expert system designed for Process Control at a general level was developed by Moore (1995). (d) Evans and Lindsay (1987) reported the development of an expert system for Statistic Quality Control, which not only selects control graphs, but also offers interpreta- tions of such graphs and provides conclusions about the control status of a process. (e) Brink and Mahalingam (1990) developed an expert system which evaluates quality at manufacturing level, so as to detect and correct defects occurring during the productive process. (f) Pfeifer (1989) reported the development of an expert system to detect defects during the productive process and described its successful application in Germany. (g) An expert system which makes use of pattern recog- nition in order to ‘visualize’ pieces under inspection was developed in 1989 in the United States (see Ntuen, Park & Kim, 1989). This system, called KIMS, was success- fully tested in a variety of experiments (one of them is reported by Ntuen, Park & Sohn, 1990), where the perfor- mance of the KIMS in image recognition tasks is assessed. (h) Fard and Sabuncuoglu (1990) developed an expert system which seeks to select sampling by attributes, determining which type of sampling is the most adequate for each case: simple, double or multiple. A project of a E.P. Paladini / Expert Systems with Applications 18 (2000) 133–151134 https://isiarticles.com/article/4495 
work_j6fm452usjdmdhgardrg2mgn2i	----	The exploration of customer satisfaction model from a comprehensive perspective The exploration of customer satisfaction model from a comprehensive perspective Wen-Bao Lin * National Formosa University, Department of Business Administration ADD:64, Wen-Hua Road, Huwei, Yunlin, Taiwan, ROC Abstract This study provides a model of customer satisfaction from a comprehensive perspective and tries to use the nonlinear fuzzy neutral network model to verify the assumptions of the study. Samples are taken from the information and tourism industries at a proportion of 2:1 based on the population in Taipei and Kaohsiung cities. A total of 207 questionnaires are returned. As the result of the empirical research shows, the interpersonal-based service encounter is better than the technology-based service encounter in functional quality, while the technology-based service encounter is better than the interpersonal-based service encounter in technical quality. The functional quality has a positive and significant effect on customer satisfaction; the service quality has a positive and significant effect on service value; the service value has a positive and significant effect on customer satisfaction. The service encounter has a positive and significant effect on relationship involvement and the relationship involvement has a positive and significant effect on customer satisfaction. � 2006 Elsevier Ltd. All rights reserved. Keywords: Customer satisfaction; Service encounter; Service value; Relationship involvement; Service quality 1. Introduction The industrial structure in Taiwan has transformed from traditional agricultural and manufacturing industries to the tertiary service industry that is booming and domi- nating the market with its high employment rate and pro- duction value. In the circumstances, it is very important for the service industry to improve their service quality. The core value provided by the service industry to consumers includes not only the uniqueness of tangible and intangible products, but also various factors involved in the process of service delivery to customers, such as physical facilities, company image, and quality of the service delivery. Emphasizing customers’ perception of service quality is important to enhance the extended perception and impres- sion of consumers on products, and to increase the added value of products. With the rise of consumer awareness, requirements of customers changing from mass production to customization and quality. Therefore, how to under- stand the requirements of customers and provide them with products and services they need, reduce customer costs and improve customer value, and continually trace the satisfac- tion of consumers are essential for an enterprise to march toward success. Much research on the cause of service quality and cus- tomer satisfaction has been published in recent years. This research focuses on the following dimensions: (1) discuss- ing the antecedent variable that affects customer satisfac- tion. Zeithaml and Bitner (2000), for example, found that customer satisfaction was affected by service quality, prod- uct quality, price, personal and situational factors. Oliver and DeSarbo (1988) found that service quality was the antecedent variable of customer satisfaction, while Ander- son, Fornell, and Lehmann (1994) found that customer sat- isfaction was dependent on service value; (2) presenting theoretical models to explain the importance and necessity of customer satisfaction. Anderson and Suillivan (1993) pointed out that the satisfaction or dissatisfaction of cus- tomers could be measured based on the gap between the 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.04.021 * Tel.: +886 5 6315775; fax: +886 5 6313842. E-mail addresses: e2667@ms23.hinet.net, wblin@sunws.nfu.edu.tw www.elsevier.com/locate/eswa Expert Systems with Applications 33 (2007) 110–121 Expert Systems with Applications mailto:e2667@ms23.hinet.net mailto:wblin@sunws.nfu.edu.tw expectations of customers before purchasing and the per- formance of the purchased products or services. Woodruff, Cadotte, and Jenkins (1983) raised the norms in models of consumer satisfaction. They found that consumers evalu- ated products of different brands based on ‘‘norms’’ and their satisfaction was determined by the conformity among these norms. Meuter, Ostrom, Roundtree, and Bitner (2000) introduced the Technology Infusion Matrix to ana- lyze how to enhance the provision of customized service by helping staffs or consumers use hi-tech tools and ensure the service failure compensation to improve customers satisfac- tion; (3) bringing up research dimensions and approaches for measuring customer satisfaction and service quality. For example, Perkins (1993) pointed out that industries were different in target market segmentation and, thus, measurement of customer satisfaction varied depending on the characteristics of individual industries. He raised three dimensions to be taken into consideration for empir- ical research on measurement of customer satisfaction: operation dimension (availability, delivery on time, price, credibility), service dimension (sales service, technical sup- port, product line), product dimension (technical value, reliability, design). Parasuraman, Zeithaml, and Berry (1985) referred to ‘‘encounter’’ and ‘‘feasibility’’ as key dimensions. After interviewing senior management and consumers of four service industries and conducting com- prehensive exploratory research, they raised the five-gap theory and brought up an important measurement approach. With this approach, excess of expected service (ES) over perceived service (PS) was represented by ES > PS, indicating that the overall service quality was not acceptable, while satisfactory service quality was repre- sented by ES = PS and excess of perceived service over expected service was represented by ES < PS, indicating an inclination toward ideal quality. Previous research on customer satisfaction had the fol- lowing characteristics: (1) emphasizing the discussion of performance. Oliver (1980), for example, found that satis- faction was a value consumers give to a deal. The higher the service level that consumers expect to have, the more difficult for the service provider to provide, the higher the customer satisfaction. According to the conformity theory of Mittal and Tsiros (1999), when the expectation of con- sumers was not in conformity with the actual performance of a product, they might change their perception of the product and incline to eliminate the gap between the expec- tation and performance, and thus improve the satisfaction; (2) emphasizing the discussion of cause and result of satis- faction. Cronin and Taylor (1992), for example, evaluated service quality based on performance and argued that ser- vice quality was the cause of customer satisfaction. Para- suraman, Zeithaml, and Berry (1988) researched the gap between service quality and customer satisfaction; (3) explaining the cause of customer satisfaction from a single perspective, such as the attribution theory of Bitner (1990) and the equity theory of Oliver and DeSarbo (1988). Kotler (2000) found that customer satisfaction was closely linked to customer value and the measurement of customer satis- faction should be carried out from the perspective of value. As some previous literature shows, the topics such as ‘‘ser- vice quality’’, ‘‘service value’’, and ‘‘customer satisfaction’’ were researched separately and only a limited number of literature focused on integration research in the relation- ship between the behavior intention of consumers and the service quality, service value, and customer satisfaction as a whole (Cronin, Brady, Tomas, & Hult, 2000). This study provides integrated research of these variables and focuses on their relationship as the first research motive; (4) most of the satisfaction models used in previous research empha- sized the perceptive (Kotler, 2000) and emotional perspec- tives (Price, Arnould, & Tierney, 1995) of consumers. However, what a provider of the service industry empha- sizes is not only the interaction with consumers to ensure their satisfaction, but also the method of changing the trade exchange to the value increment exchange (the rela- tionship continuum theory of Day, 2000). Therefore, this study proceeds to the research from the perspective of rela- tionship involvement as the second research motive. Besides, the multivariate statistic analysis approach or case study was commonly used for empirical research (Bol- ton & Drew, 1991; Cronin, Michael, Brady, Hightower, & Shemwell, 1997; Zeithaml, 1988) and the nonlinear research model (such as the nonlinear fuzzy neural network model) was not used often for theory verification. The non- linear fuzzy neural network model is maturely developed and has a wide range of applications. It is not only useful for prediction and assortment, but also for uncertain behavior systems. Therefore, this study uses the nonlinear fuzzy neural network model to verify the initial data of the research as the third research motive. Based on the above, the purpose of this study is listed as follows: (1) To provide a customer satisfaction model from a comprehensive perspective of relationship involve- ment, service encounter, service quality, and service value. (2) To verify collected data with the nonlinear fuzzy neu- ral network model and discuss its meaning for management. 2. Literature review Service encounters were classified as follows in the previ- ous research: (1) according to the level of service encounter with the staff or facilities, the service encounter was classi- fied into high service encounter (contacting staff more fre- quently or contacting customers widely) and lower service encounter (mainly contacting physical facilities). Industries belonging to the former include the hotel industry, insur- ance industry and tourism industry, while industries that belong to the latter include internet banks and couriers. However, many industries of high or middle service encoun- ters contact with customers and make deals with them W.-B. Lin / Expert Systems with Applications 33 (2007) 110–121 111 https://isiarticles.com/article/20075 
work_ja5wcuvf65hxrpyv62j6qsu2gy	----	DenGraph-HO: A Density-based Hierarchical Graph Clustering Algorithm Nico Schlitter, Tanja Falkowski and Jörg Lässig Abstract DenGraph-HO is an extension of the density-based graph clustering algorithm DenGraph. It is able to detect dense groups of nodes in a given graph and produces a hierarchy of clusters which can be efficiently computed. The generated hierarchy can be used to investigate the structure and the characteristics of social networks. Each hierarchy level provides a different level of detail and can be used as the basis for interactive visual social network analysis. After a short introduction of the original DenGraph algorithm we present DenGraph-HO and its top-down and bottom-up approaches. We describe the data structures and memory requirements and analyse the run time complexity. Finally, we apply the DenGraph-HO algorithm to real-world datasets obtained from the online music platform Last.fm and from the former U.S. company Enron. 1. Introduction In 2011, we proposed DenGraph-HO in order to fulfill the special needs of social network analysts (Schlitter, Falkowski & Lässig 2011). In most cases, the visual inspection of a network is the first step of the analytical process and helps to determine the basic graph characteristics and further actions. DenGraph-HO supports this early stage by providing a quick visual analysis of the network structure. It provides the ability of zooming into network clusterings and has proven its usefulness for our practical work. Our approach differs from traditional hierar- chical clustering methods in that DenGraph-HO is a non-partional clustering algorithm. We con- sider the fact that not all nodes are necessarily members of clusters. In addition, the proposed hierarchy is not strictly built up by the classic divisive or agglomerative approach that is known from literature. We generalize these methods and propose a top-down approach and a bottom-up approach by extending the hierarchy paradigms. The proposed hierarchy supports superordinate clusters that contain subclusters. Each level of the hierarchy represents a clustering that fulfills the original DenGraph paradigms, which will be presented in Section 2.1. The levels, respectively the clusterings, differ in the density that is required to form a cluster. While lower level clusterings aggregate nodes with a lower similarity, higher level clusterings require a higher similarity between nodes. The density-based cluster criteria are controlled by the parameters η and ε which are iteratively in- or decreased to obtain different levels of the hier- archy. An existing clustering is used to compute the clustering of the next level. The efficiency of our algorithm is based on this iterative sequence of cluster adaptations instead of a complete new clustering. The remainder of this article is organized as follows. Section 2 discusses related work and introduces the original DenGraph algorithm and its variations DenGraph-O and DenGraph-IO. Section 3 covers the proposed top-down and bottom-up approaches of DenGraph-HO as well as the used data structures, their memory require- ments and the algorithm’s run time complexity. Its usability is demonstrated in Section 4 by ap- plying DenGraph-HO to two real-world datasets. Finally, a conclusion and an outlook are given in Section 5. 2. Related and Previous Work Clustering is a data mining method which is used to detect hidden structures in huge data sets. Its purpose is to bundle of single objects into groups in such a way that objects of the same group are more similar to each other than to objects of other groups. K-means (MacQueen 1967) is a commonly used and well studied clustering algorithm for spatial data. The algorithm strictly groups data all points into clusters. The number of clusters has to be predefined and stays constant during the clustering process. Each data point belongs to exactly one cluster. This can been seen as one major drawback of the algorithm since it does no deal with the concept of outliers. Density-based approaches like DBSCAN (Ester, Kriegel, Sander & Xu 1996) require no previous knowledge of the number of clusters hidden in the data. Clusters are defined as regions that have a high density of data points and that are surrounded by regions of data points of lower density. Since data points may not belong to any cluster, DBSCAN is a non-partitioning clustering method, which provides the ability to handle outliers. The major drawback of DBSCAN is its inability to find nested or overlapping clusters. Therefore, the authors extended their original approach to overcome this problem. The algorithms OPTICS (Ankerst, Breunig, Kriegel & Sander 1999), HiCS (Achtert, Böhm, Kriegel, Kröger, Müller- Gorman & Zimek 2006), HiCO (Achtert, Böhm, Kröger & Zimek 2006) and DiSH (Achtert, Böhm, Kriegel, Kröger, Müller-Gorman & Zimek 2007) were designed to generate cluster hierarchies. Since density-based clustering approaches pro- vide significant advantages they have also been used for graph clustering purposes. DenShrink (Huang, Sun, Han & Feng 2011), a parameter- free algorithm for community detection, applies modularity optimization to reveal the embedded hierarchical structures with various densities in large-scale weighted undirected networks. DenGraph (Falkowski, Barth & Spiliopoulou 2007) and SCAN (Xu, Yuruk, Feng & Schweiger 2007) have been independently proposed by extending DBSCAN. Both methods operate on undirected graphs, use a local cluster criteria and provide the ability to detect communities in social networks. However, a comparison of both methods (Kriegel, Kröger, Ntoutsi & Zimek 2011) pointed out that they significantly differ in the applied similarity function. DenGraph oper- ates on weighted graphs and considers the weight of an edge between two nodes as similarity of these nodes. On the contrary, SCAN uses a sim- ilarity function that is based on the topology of the underlying graph. The similarity of two nodes correlates to the number of neighbors they share. Later, SCAN’s similarity function was also used for the divisive hierarchical clustering DHSCAN (Yuruk, Mete, Xu & Schweiger 2007) and the agglomerative hierarchical clustering AHSCAN (Yuruk, Mete, Xu & Schweiger 2009). Since 2007, the development of DenGraph continued as well. We proposed the extensions DenGraph-O that allows overlapping clusters and DenGraph-IO which tracks community evolu- tion over time (Falkowski, Barth & Spiliopoulou 2008). In the following, we briefly introduce the original DenGraph algorithm and its extensions. In section 3 we then propose the density-based hierarchical algorithm DenGraph-HO. 2.1. DenGraph Given a graph G = (V,E) consisting of a set of nodes V and a set of weighted, undirected edges E, the DenGraph algorithm produces a clustering ζ = {C1,...,Ck} where each clus- ter Ci (i = 1...k) consists of nodes VCi ⊆ V . Since DenGraph is a non-partitioning cluster- ing algorithm there can be noise nodes VN = {u ∈V | u /∈Ci} that are not part of any cluster Ci. The remaining non-noise nodes are members of only one cluster and either core nodes or border nodes. Definition The ε -neighborhood Nε (u) of a node u∈V is defined as set of nodes that are connected to u having a distance less than or equal to ε . Nε (u) = {v ∈V | ∃(u,v)∈ E ∧ dist(u,v)≤ ε}, where dist(u,v) is the distance between u and v. A node u ∈ V is considered as core node if it has an ε -neighborhood of at least η neighbor nodes (|Nε (u)|≥ η ). Nodes which are in the ε - neighborhood of a core node, but do not have an own ε -neighborhood are called border nodes. According to (Falkowski 2009), the actual cluster criterion is based on the concepts di- rectly density-reachable, density-reachable and density-connected which are defined below and illustrated in Figure 1. Definition Let u,v ∈V be two nodes. Node u is directly density-reachable from v within V with respect to ε and η if and only if v is a core node and u is in its ε -neighborhood, i.e. u ∈ Nε (v). Definition Let u,v ∈V be two nodes. Node u is density-reachable from v within V with respect to ε and η if there is a chain of nodes p1,..., pn such that p1 = v, pn = u and for each i = 2,...,n it holds that pi is directly density-reachable from pi−1 within V with respect to ε and η . Definition Let u,v ∈ V be two nodes. u is density-connected to v within V with respect to ε and η if and only if there is a node m ∈ V such that u is density-reachable from m and v is density-reachable from m. Definition Let G(V,E) be an undirected, weighted graph. A non-empty set C ⊆ V is Figure 1: Concepts of Connectivity (Falkowski 2009) denoted as Cluster with respect to ε and η if and only if: • For all u,v ∈ V , if u ∈ C and v is density reachable from u, then v ∈C. • For all u,v ∈C u is density-connected to v within V with respect to ε and η . The complete DenGraph procedure is de- scribed in Algorithm 1. It uses a stack in order to process the graph nodes. In a first step, all nodes V are marked as noise. Afterwards, each so far unprocessed node v is visited and checked if it has an ε -neighborhood. If the neighbor- hood contains at least η nodes (|N(v)|≥ η ) the node v is marked as core and a new cluster is founded. Each of v’s neighbors within the ε - neighborhood is marked as border, becomes a member of the new cluster and is pushed on the stack. After processing all neighbors, each node u from the stack is checked regarding having an ε -neighborhood and is marked correspondingly. If u became a core node, all of it’s neighbors are marked as border and pushed on the stack. This procedure is repeated until all nodes of the graph are processed. According to Algorithm 1 the time complexity of our procedure mainly depends on the number of nodes and the number of edges. Each node is visited once and each edge is processed up to two times - once from both end-nodes. Conse- quently, the overall run time complexity of the DenGraph algorithm is O(|V|+ |E|), where |V| is the number of nodes and |E| the number of edges. (Falkowski 2009) Figure 2 shows the clustering of an exemplary interaction graph. The graph was clustered by applying the original DenGraph. Core nodes are blue, border nodes are green and noise nodes are Algorithm 1: DenGraph input : Graph,Clustering ζ ,η ,ε output: Clustering ζ begin foreach (u ∈V|u.state = noise) do if (|Nε (u)|≥ η) then Cluster=CreateNewCluster(); Cluster.addNode(u); u.state=core; foreach n ∈ Nε (u) do Cluster.addNode(n); n.state=border; stack.push(n); repeat v=stack.pop(); if (|Nε (v)|≥ η) then v.state=core; foreach n ∈ Nε (v)|n.state 6= core do Cluster.addNode(n); n.state=border; stack.push(n); until stack is empty; return ζ ; Figure 2: DenGraph Clustering (Falkowski 2009) drawn in red color. Each cluster contains one or more core nodes which are connected with each other. Following the DenGraph paradigm, the total number of nodes per cluster is greater than η . Nodes that are not member of any cluster are considered as noise nodes. Figure 3: DenGraph-O Clustering (Falkowski 2009) 2.2. DenGraph-O Practical work with DenGraph in the field of Social Network Analysis revealed a minor draw- back: While in real world applications nodes - re- spectively human beings - might be part of more than one community, the original DenGraph al- gorithm does not allow for clusters to overlap. The affected nodes were exclusively assigned to only one cluster among the cluster candidates. This issue was addressed in (Falkowski et al. 2008). The extended version, called DenGraph-O, allows border nodes that are mem- bers of multiple clusters. As a consequence sev- eral clusters of a graph might be overlapped. An example for those overlapping clusters is illustrated in Figure 3. Cluster 1 overlaps with Cluster 2 and Cluster 5 while Cluster 3 overlaps with Cluster 5 and Cluster 4. 2.3. DenGraph-IO In 2007, Falkowski et al. proposed DenGraph-IO (Falkowski & Barth 2007, Falkowski et al. 2008) to analyse the dynamics of communities over time. The authors compared the changes between clusterings that were obtained in different points in time. For this, it is necessary to iteratively compute the graph clusterings that are subse- quently observed over time. A huge computa- tional effort would be necessary if the original DenGraph was used to process multiple con- secutive snapshots of social networks. However, as social structures often change slowly over time, the graphs Gt and Gt+1 differ just slightly. Therefore, a complete re-clustering, as the use of the original DenGraph algorithm would demand, is quite inefficient. The incremental clustering algorithm DenGraph-IO addresses this issue and updates an existing clustering based on the changes of the underlying graph. The graph changes either by adding/removing of nodes, adding/removing of edges or changing the weight of an existing edge. As a result of the graph updates, new clusters may appear or existing clusters may be removed, merged or split. Since the DenGraph-IO algorithm deals exclu- sively with the parts of the graph that changed from one point in time to the other, the compu- tational complexity is dramatically reduced and even huge networks can be processed in reason- able time. Our experiments using a real-world social network obtained from Last.fm show that handling 2,500 graph updates using DenGraph- IO is about 400 times faster than the re-clustering with DenGraph-O (Schlitter & Falkowski 2009, Falkowski 2009). 3. DenGraph-HO One challenge of the DenGraph algorithm is the choice of the parameters ε and η . Several heuris- tics were investigated (Falkowski 2009), how- ever, the ”right” parameter combination mainly depends on the aim of the analysis. If the analyst is for example interested in observing strongly connected nodes rather than in clusterings that show the overall structure, the parameters need to be chosen accordingly. DenGraph-HO addresses this issue by allowing a quick visual inspection of clusterings for a given set of parameter com- binations. The process of zooming into or out of the network can be interactively controlled by the analyst according to his needs. The proposed algorithm returns a hierarchical clustering that describes the structure of the underlying network. The resulting hierarchy pro- vides multiple views of the network structure in different levels of detail. Consequently, the cluster hierarchy is an ideal basis for an efficient (a) Hierarchical Clusterings (b) Hierarchy Figure 4: Visualization of an Hierarchical DenGraph-HO Clustering zooming implementation. Zooming-in is done by stepping up in the hierarchy. It provides a more detailed view of the current cluster by presenting its subclusters. A higher level of abstraction is reached by zooming-out, which is equivalent to merging similar clusters into superordinate clusters. Figure 4 shows an exemplary graph clustering and the corresponding hierarchy as a tree of clusters. For the sake of clarity we removed the graph edges. The root of the hierarchy tree represents the whole graph and children repre- sent subclusters of their superordinate cluster. Following this definition, the leaves of the tree correspond to the smallest clusters that have no subclusters. The proposed DenGraph-HO hierarchy is based on the concepts of the DenGraph algo- rithm. Each level of the tree (besides the root) represents a valid clustering that fulfills the DenGraph-O paradigms. The hierarchy can be built by repeatedly applying DenGraph-O while using specific parameter settings for each level of the tree. The choice of the parameters η and ε is limited by constraints in order to ensure that lower level clusters are subclusters of the superordinate cluster. Let us assume that the clustering ζl forms level l of the hierarchy and is computed by applying DenGraph-O with the parameters εl and ηl . Level l + 1 represents a clustering that is based on εl+1 and ηl+1 and guarantees a higher similarity of nodes in the cluster. According to the description above, ζl+1 has to contain subclusters of clusters that are element of ζl . In order to preserve this parent-child relationship we have to ensure the following constraints: 1) The parameter εl that is used to generate the clustering ζl has to be bigger than or equal to εl+1 which is used to compute the clustering ζl+1 : εl ≥ εl+1 . 2) The parameter ηl that is used to generate the clustering ζl has to be lower than or equal to ηl+1 which is used to compute the clustering ζl+1 : ηl ≤ ηl+1 . Increasing ε might lead to a transition of a node state from noise or border to core or from noise to border. By increasing ε a core node cannot lose its state. This explains why increasing ε might create a new cluster or expand an existing one and why it surely avoids cluster reductions or removals. The same argument holds for de- creasing η and shows why the demanded cluster- subcluster relation can be guaranteed by the given constraints. In the following, we discuss how the proposed cluster hierarchy can be efficiently generated based on a list of parameter settings that fulfill the discussed constraints. An obvious approach would be to perform multiple re-clusterings us- ing DenGraph-O until each parameter setting is processed. However, this is very inefficient and would be a huge computational effort. The proposed DenGraph-HO algorithm ad- dresses this issue and uses incremental parameter changes to generate the cluster hierarchy. Instead of computing a complete new clustering for each level, an existing clustering is used and adapted. In the following, we discuss how an existing clustering of level l can be used to compute the clusterings of level l + 1 and l −1. We propose a bottom-up and top-down approach and analyse their efficiency depending on the graph structure. The input for both approaches is a graph G = (V,E) and an existing cluster- ing ζl = { Cl1,...,C l k } that fulfills the DenGraph paradigms. 3.1. Top-down Approach: Cluster Reduction, Split or Removal The top-down approach performs a zoom-in op- eration and generates a new clustering for level l + 1 of the hierarchy. Clusters of level l might be reduced, split or removed. By decreasing ε and increasing η the state of nodes within a cluster might change. A former border node might become noise (Cluster Reduction). Former core nodes might get border state (possible Clus- ter Splitting) or noise state (Cluster Reduction, possible Cluster Splitting or Removal). Due to the DenGraph paradigms, it is guar- anteed that noise nodes can not reach border or core state by decreasing ε or increasing η . Thus, noise nodes will not change their state and do not need to be processed. Consequently, the top-down approach traverses just border and core nodes and performs a re-clustering for each existing cluster. For this purpose, we use the modified DenGraph-O procedure shown in Algo- rithm 3. Each cluster C of level l is re-clustered by applying the parameters of level l + 1. Algorithm 2 shows how the modified DenGraph-O algorithm is used to create a complete hierarchy. After generating a parameter list of (ε,η)-pairs an initial DenGraph-O clustering for level l=1 is iteratively adapted to generate the subsequent clusterings for the levels l+i. Algorithm 2: Top-down Approach input : Graph,lmax output: Clustering ζ PL=CreateParameterSettingList(lmax); ζ1=DenGraph(Graph,PL[1].η ,PL[1].ε ); for l = 2 to lmax do ζl =Top-down Step(ζl−1,PL[l].η ,PL[l].ε ); return ζ ; Obviously, the characteristic of the cluster hierarchy depends on the values specified in the parameter list. For a given social network analysis task, it is quite hard to find appropriate parameters, that lead to a useful hierarchy - especially if there is no prior knowledge about the given social network. We are approaching this problem by choosing parameter values which are equally distributed over the parameter search space. On this basis, parameter ranges that lead to useful hierarchal structures can be discovered. Later on, they can be explored in more detail by re-applying the clustering method with new parameter settings. Algorithm 3: Top-down Step input : Graph,ζ ,ε ,η output: Clustering ζ foreach (C ∈ ζ ) do foreach r ∈C do r.state=noise; foreach (u ∈C|u.state = noise) do if (|Nε (u)|≥ η) then Cluster=CreateNewCluster(); C.addCluster(Cluster); Cluster.addNode(u); u.state=core; foreach n ∈ Nε (u) do Cluster.addNode(n); n.state=border; stack.push(n); while stack is empty do v=stack.pop(); if (|Nε (v)|≥ η) then v.state=core; foreach n ∈ Nε (v)|n.state 6= core do Cluster.addNode(n); n.state=border; stack.push(n); return ζ ; 3.2. Bottom-up Approach: Cluster Creation, Absorption and Merging The bottom-up approach performs a zoom-out operation and generates a new clustering on level l −1 of the cluster hierarchy. Existing clusters of level l might grow through the absorption of nodes or through the merge with other clusters. Our approach deals with changes of η only, because changing ε as well would make it nec- essary to process all graph nodes. However, if we process all graph nodes our approach would not be more efficient than the original DenGraph algorithm. Algorithm 4 shows how we expand the exist- ing clusters by the iterative increase of η . Doing so, the state of a core node remains unchanged. A former noise node may become a core node (cluster creation) or a border node (absorption). A former border node could become core node (absorption). In case a former border node is member of multiple overlapping clusters, its tran- sition to core state leads to a merge of those clusters. Algorithm 4: Bottom-up Approach input : Graph,lmax,etamin,ε output: Clustering ζ η =ηmin; ζlmax =DenGraph(Graph,η ,ε ); for l = lmax −1 to 1 do η =η +1; ζl =Bottom-up Step(ζl+1,ηl ,ε ); return ζ ; Algorithm 5 describes the procedure that per- forms the bottom-up step by processing the changes of η . Since core nodes keep their state, there is no need to consider them in the bottom- up approach. Consequently, the proposed proce- dure traverses only noise and border nodes in order to determine their new state and to adapt the clustering accordingly. Following this argu- mentation, the procedure’s efficiency is based on the saved time that the original DenGraph-O would have spent for processing core nodes. First, the algorithm iterates over all existing clusters in order to expand them. For each cluster it traverses all border nodes and updates their state according to the parameters η and ε . In case a former border node becomes core, this new core node is pushed on the stack for further processing. After dealing with all border nodes, the procedure Expand checks if the new core nodes absorb their neighbors into the cluster. In case an absorbed node has less than η neigh- bors in its ε -neighborhood it gets border state, Algorithm 5: Bottom-up Step input : Graph,Clustering ζ ,ε ,η output: Clustering ζ foreach (C ∈ ζ ) do Cluster=CreateNewCluster(); Cluster.addSubCluster(C); foreach (u ∈C|u.state = border) do if (|Nε (u)|≥ η) then u.state=core; stack.push(u); Expand(stack,Cluster,ε ,η ); ζ =DenGraph(Graph,ζ ,ε ,η ); return ζ ; Algorithm 6: Expand input : Graph,Clustering,Cluster,ε ,η output: Clustering while Stack is not empty do u=stack.pop(); if (u is member of multiple clusters) then foreach (C ∈ ζ|u ∈C) do Cluster.addSubCluster(C); foreach (p ∈C|p.state = border) do if (|Nε (p)|≥ η) then p.state=core; stack.push(p); else p.state=border; foreach (n ∈ Nε (u)|n.state ∈{noise,border}) do Cluster.addNode(n); if (|Nε (n)|≥ η) then n.state=core; stack.push(n); else n.state=border; return Clustering; otherwise it becomes a core node. These newly discovered core nodes are pushed on the stack and the procedure is repeated until no further nodes are absorbed into the cluster. Due to the possibility of overlapping clusters, a new core node might have been a member of multiple clusters before. Consequently, its new core state leads to a merge of the affected clusters into a superordinate cluster. After dealing with all border nodes the ex- isting clusters are maximal expanded with re- spect to the changed η and the constant ε . The remaining noise nodes are processed to check whether their state has changed. In case a for- mer noise node becomes core, a new cluster is created. To search for new clusters we use the original DenGraph algorithm, which processes exclusively noise nodes. Please note, that during the handling of these noise nodes a newly created cluster will not merge with an existing one. If the new cluster were in ε -distance to another cluster, the nodes of the new cluster would have been already absorbed by this existing cluster during its expansion phase. 3.3. Data Structures In the last decade, the number of social net- works and their participating users has rapidly increased. Following this trend, DenGraph-HO was developed to process efficiently even huge graphs like facebook or twitter that have millions of nodes and edges. To achieve this challenging aim, the used data structures need to be opti- mized for the most frequent operations of the proposed clustering algorithm. In the following, we briefly describe the used data structures and their memory requirements. The DenGraph-HO algorithm is implemented using Java. All data structures including the graph G = (V,E), the nodes V and the edges E are modeled as single classes. The corresponding objects are instantiated during the graph built-up phase and each object is singularly materialized in memory even if there are several references to this object. According to Algorithm 2 and 4, the top-down approach as well as the bottom-up approach need a valid clustering as starting point which is gener- ated by the original DenGraph algorithm. During this initial application of DenGraph procedure, the algorithm traverses all noise nodes in order to update the node states based on the cardinality of their ε -neighborhood. For an efficient imple- mentation, a list structure is needed that contains references to the noise nodes, allows rapid node traversing and also provides efficient insert and remove operations for the current item. The class LinkedList provided by Java fulfills these require- ments. Traversing the list takes linear time and works in O(n), the insert and remove operations on the current item are handled in constant time O(1). Because it is basically a modified DenGraph which performs the top-down step, Algorithm 3 benefits from the list of noise nodes as well. We implemented a similar list of border nodes for each cluster because the bottom-up step, implemented in Algorithm 5, needs to traverse all border nodes. Therefore, traversing nodes is efficiently done for all procedures of the algo- rithm. In order to update the state of a node, the cardi- nality of a node’s ε -neighborhood must be deter- mined by counting edges that have weight values less than ε . Due to the fact that this operation is frequently used, we decided to implement a binning mechanism for each node which assigns, depending on the weight, the edges of each node to a specific bin. In case the parameter εl for each level l is specified in advance, the ranges of the binning mechanism can be set accordingly. This can be done for example during initializing the graph and takes linear time O(n) where n denotes the number of nodes in the graph. For example, choosing εl=1 = 0.7 and εl=2 = 0.25 to generate a hierarchy with two levels, the ranges for the binning mechanism should be set to the same values. By this, all edges with a weight less than 0.25 are put in bin 1, edges with a weight between 0.25 and 0.75 are stored in bin 2. The remaining edges are put into bin 3. According to Algorithm 2, for each non-noise node u the cardinality of N0.25(u), and N0.7(u) must be determined. Due to the binning mechanism, which counts the number of edges in each bin, this task can be performed in O(l) where l denotes the number of bins. 3.4. Run Time Complexity The run time complexity of the original DenGraph applied on a graph G(V,E) is O(|V|+ |E|), where |V| is the number of nodes and |E| the number of edges. (Falkowski 2009). By applying DenGraph k times with k different (ε,η)-pairs, we would be able to produce a hier- archy with k levels similar to the ones generated by our DenGraph-HO algorithm. However, this would be quite inefficient since for each iteration each node is traversed even if the node’s state definitely will not change. As described before, DenGraph-HO overcomes this problem and tra- verses only relevant nodes. In fact, for each level of the hierarchy, DenGraph-HO deals only with the subgraph G′(V ′,E′) where V ′ ⊂ V and E′⊂ E. In the worst case scenario, if the original graph G equals G′, DenGraph-HO has the same run time like applying the original DenGraph algorithm k times. However, our practical work has shown that depending on the parameters ε and η the subgraph G′ is up to 50% smaller than the original graph G. This leads to a significant run time reduction, which has a huge impact on practical work. However, since this improvement reduces the run time just by a constant factor, the complexity according to the O-notation does not change and is still O(|V|+|E|). 3.5. Memory Requirements In the following, the memory requirement of the data structures is briefly described. It depends on the number of nodes |V|, the number of edges |E| and the number of hierarchy levels lmax. Assuming a 32bit environment, each pointer to a single object requires 4 byte. Each edge contains references to the two nodes that are connected by this edge. In addition, the weight of this connection is stored. Conse- quently, each edge needs 2×4+4 = 12 byte. The required memory to store all edges is |E|×12 byte. For each node its identifier, the current state (noise, border or core), the cluster(s) to which the node belongs and a reference to the edge binning structure that holds information about the node’s edges are stored. Due to memory requirements of 4 + 4 +(4×lmax)+ 4 = 12 + 4×lmax byte per node, the whole set of nodes needs (12 + 4× lmax)×|V| byte. The graph structure stores a list of references to all edges, a list containing the border and core nodes and a list of all noise nodes. Since we use a LinkedList there is a need for a forward pointer, a backward pointer and the reference to the actual node. In total the requirement of these lists is 12×|E|+ 12×|V| byte. The binning structure for the edges contains lmax bins that are implemented as a LinkedList of edge references. Since each edge is listed in two binning structures, we calculate a total memory need of 2×|E|×12 byte for all binning structures of the graph. This already includes the for- and backward pointers of our list implemen- tation. The addition storage of the binning ranges takes |V|×lmax ×4 byte. Each cluster contains references to the nodes of this cluster. For each hierarchy level, we assume that each node is member of one cluster. As a result, the clustering demands for a total of |V|×lmax ×12 byte. During processing, additional memory for the stack structure is required. In the worst case scenario, this stack holds all nodes and is then limited to |V|×4 bytes. In total, we calculate a memory requirement of |E|×48 +|V|×(28 + 20×lmax) byte to store all data structures of the DenGraph-HO algorithm. Consequently, a 3-level-clustering of a graph that has 1200 nodes and 12600 edges requires about 710 KB. 4. Applications In the following, we apply the DenGraph-HO al- gorithm on different datasets and show its ability for explorative visual social network analyses. The presented algorithm was implemented as a plugin for the open-source data mining frame- work RapidMiner (Mierswa, Wurst, Klinken- berg, Scholz & Euler 2006) which was formerly known as YALE. For graph visualization we used the prefuse toolkit developed by Jeffrey Heer (Heer, Card & Landay 2005). The first case study demonstrates the algo- rithm’s usefulness and analyses the email com- munication of the former U.S. company Enron. This data was published by the Federal Energy Regulatory Commission during the investigation of the biggest bankruptcy case in US-history. For our second analysis, we use information about user’s music listening behavior provided by the online music platform Last.fm. 4.1. Enron Case Study The collapse of Enron, a U.S. company honored in six consecutive years by “Fortune” as “Amer- ica’s Most Innovative Company”, caused one of the biggest bankruptcy cases in US-history. To investigate the case, a data set of approximately 1.5 million e-mails sent or received by Enron employees was published by the Federal Energy Regulatory Commission. Because of the public interest and the rareness of available real world e-mail data many stud- ies has been carried out on the Enron e-mail dataset. The original dataset had some ma- jor integrity problems and was cleaned up by Melinda Gervasio at the independent, nonprofit research institute SRI. The revised dataset was used by Klimt and Yang when they presented their introduction of then Enron corpus (Klimt & Yang 2004). Bekkerman et al. use the dataset for experiments in e-mail foldering and classification (Bekkerman, McCallum & Huang 2004). Diesner et al. analysed the evolution of structure and communication behavior of the employees on different organizational levels (Diesner, Frantz & Carley 2005). Shetty and Adibi published a sub- set of the original data containing approximately 250,000 e-mails from/to 151 Enron employees which were sent during 1998 and 2002 (Shetty & Adibi 2004). In 2011, we used this dataset to analyse the evolution of communities over time (Falkowski et al. 2008, Falkowski 2009). For our experiments with the DenGraph-HO clustering algorithm, we use a subset of this dataset containing only messages sent from En- ron employees to Enron employees. Traditionally, clustering is based on the dis- tance between the objects to be clustered. On a graph of interactions, we model distance between two actors based on the number of their interac- tions. To deal with outliers we chose a value z so that about 1.5 percent of all edges have an edge weight larger than z. Afterwards, we ensure that the weight of these edges is bounded to z. Definition The distance between two actors u and v is defined as distance(u,v) = min{u → v,v → u,z}−1 z−1 , (1) where u→v is the number of messages sent from u to v, v → u the number of messages sent from v to u and z is a value specified for handling outliers. The distance function ranges in [0,1] and is symmetric. If only one reciprocated interaction exists between u and v, then their distance is one. For our experiments we used a graph con- sisting of 57 nodes and 81 edges that repre- sents the internal Enron communication between 2000/12/04 and 2000/12/10. We applied the pre- sented top-down approach and retrieved the clus- ter hierarchy as shown in Figure 6. For each hierarchy level the used parameters η and ε , the number of resulting clusters and their sizes are listed in Table 1. Figure 5 shows the graphical representation of the interaction graph and the computed clusters. Due to the fact that the four members of Cluster 2 did not communicate with the other employees there is no connection between Clus- ter 2 and the bigger Cluster 1. Within Cluster1 the three subclusters Cluster 3, Cluster 4 and Figure 6: Hierarchy of the Enron Clustering Table 1: Hierarchy of the Enron Clustering Level ε η Labels # Nodes 0 - - Graph 57 1 1 2 Cluster 1 53 Cluster 2 4 2 0.988 3 Cluster 3 15 Cluster 4 11 Cluster 5 5 3 0.975 4 Cluster 6 8 Cluster 7 5 4 0.963 5 No changes 5 0.95 6 Cluster 13 9 Cluster 5 appear. A reason might be, that the respective cluster members work together in a team or in a project. At level three, Cluster 3 splits into Cluster 6 and Cluster 7. Since the Figure 5: Hierarchical Clustering of the Enron Graph Table 2: Hierarchy of the Last.fm Clustering Level ε η Cluster Labels # Nodes 0 - - Entire Graph 1209 1 0.05 22 hip-hop 35 indie, rock, alternative, punk 631 death metal, metal, heavy metal 85 2 0.037 24 hip-hop 31 punk, rock, indie 86 punk, rock 33 indie, rock, alternative 316 death metal, metal, heavy metal 37 3 0.025 26 indie, indie rock, rock 120 parameter combination of η and ε at level four leads to no cluster changes, Cluster 13 is finally formed within Cluster 5 at level five. 4.2. Last.fm Case Study Last.fm is a music community with over 20 million active users based in more than 200 countries. After a user signs up, a media player plugin is installed and all tracks a user listens to are submitted to a database. Last.fm records - among others - all artists a user listens to and provides lists of the most frequently listened artists for each week over the lifetime of a user. Last.fm provides for each user a list of the most frequently listened artists and the number of times the artist was played. We obtained the user listening behavior of 1,209 users over a periode of 130 weeks (from March 2005 to May 2008) and use this information to build user profiles by extracting the genres of the most listened artists. For each artist, Last.fm provides the tags that are used by users to describe the artist. For example, the band “ABBA” has 41 tags. The most often used tags are “pop”, “disco”, “swedish” and “70s”. The singer “Amy Winehouse” has 34 tags, the most often used tags are “soul”, “jazz”, “female vocalists” and “british”. Much work has been done using the collabo- rative music tagging data from Last.fm. Chen et al. studied the automatic classification of music genres (Chen, Wright & Nejdl 2009). The au- thors demonstrate the benefits of a classification technique that uses the tags supplied by the users for accurate music genre classification. Konstas et al. created a collaborative recommendation system for Last.fm based on both the social an- notation and friendships between users (Konstas, Stathopoulos & Jose 2009). In 2009, we ap- Figure 7: Hierarchy of the Last.fm Clustering Figure 8: Hierarchical Clustering of the Last.fm Graph plied DenGraph-IO on the dataset to analyse the dynamics of music communities (Schlitter & Falkowski 2009, Falkowski 2009). Geleijnse et al. used the data provided by Last.fm to investi- gate whether the tagging of artists is consistent with the artist similarities found with collabo- rative filtering techniques (Geleijnse, Schedl & Knees 2007). Since the authors found the data both consistent and descriptive for creating a ground truth for artist tagging and artist simi- larity we are following their approach. We use the tags provided by Last.fm for each artist to describe the artist in a genre vector ~ai. Definition An artist ai is defined as a genre vector ~ai = (w1,w2,...,wk) of the k-most used genre tags wi ∈W Definition For each user u a user profile is defined as ~u = m ∑ i=1 ~ai ·ci, (2) while m is the number of artists listened to. The value ci is provided by Last.fm and describes how often the artist ai was listen to by user u. Based on the user profiles we calculated the pair- wise similarity of user music preferences and generated a graph in which the nodes represent the users and the edges encode the distance be- tween users based on the similarity of their music listening behavior (Schlitter & Falkowski 2009). Definition The similarity of music preferences between two users u and v is defined as sim(u,v) = ∑ m i=1 ui ·vi√ ∑ m i=1 u 2 i ·∑ m i=1 v 2 i (3) For our analysis we used a Last.fm graph consisting of 1,209 nodes and 12,612 edges. We applied DenGraph-HO and obtained the cluster hierarchy shown in Figure 7. The calculated cluster labels are based on the user profiles of the cluster members. Table 2 gives more details about the clustering of each hierarchy level. Shown are the parame- ters ε and η , the cluster label and the number of nodes for each cluster. The semantical plausibility of the parent-child- relationship between a superordinate cluster and its subclusters can be demonstrated using the (indie, rock, alternative, punk)-cluster. This clus- ter is formed at level l = 1 and splits in the subclusters (punk, rock, indie), (punk, rock) and (indie, rock, alternative) at level l = 2. While l increases, the size of the observable clusters decreases and their semantic meaning becomes more specific. While traversing the hierarchy tree from the leaves to the root, the number of nodes per cluster increases. Due to the parameter setting of ε and η the clusters grow either through merging of clusters or through absorption of nodes. These properties of DenGraph-HO and the efficient calculation of the cluster hierarchy are the basis for the proposed zooming purpose. Figure 8 shows the entire graph and the clus- ters found by DenGraph-HO. For the sake of clarity graph edges are not drawn. Since we limited the number of hierarchy levels in our example and due to the small number of nodes, the graph can easily be understood. However, graphs with millions of nodes and a hierarchy depth greater than ten ask for appropriate tools. As a result, the ability of zooming through the graph enriches our tool set and is an important step for studying the structure of huge graphs. 5. Conclusion and Outlook In this article we proposed the hierarchi- cal density-based graph clustering algorithm DenGraph-HO. The algorithm computes a hierar- chy of clusters where each level of the hierarchy fulfills the DenGraph paradigms. The advantage of our method is based on the iterative change of clustering parameters which allows an efficient extension of a given clustering. We demonstrated the algorithm’s practical use for explorative visual network analysis by ap- plying the algorithm both to a communication graph based on the e-mail correspondence within the former U.S. company Enron and to a social network obtained from the online music platform Last.fm. Applied to the Last.fm data our algorithm forms clusters by grouping users that have simi- lar music listening preferences. For each cluster we determined a cluster label that represents music genres and which is based on the mu- sic listening behavior of the cluster members. The resulting cluster hierarchy shows a plausible semantical relationship between superordinate clusters and subclusters. Furthermore, clusters that represent similar music genres are located closely both in the graphical representation and in the cluster hierarchy. Since DenGraph-HO has proven its usefulness for our practical work, our next step is to extend our algorithm towards a scalable version that is suitable for a high performance computing environment. We will look into MapReduce and grid computing technologies in order to paral- lelize the algorithm execution. Thus, we will be able to apply our algorithm to huge social networks like Twitter or Facebook and to retrieve the corresponding cluster hierarchy in acceptable time. In addition, we will investigate how to in- tegrate the incremental approach known from the DenGraph-IO algorithm. The result will be an hierarchical incremental density-based graph clustering algorithm which is able to track the evolution of cluster hierarchies over time. 6. Acknowledgment This study was supported by the members of the grid computing project distributedDataMining (http://www.distributedDataMining.org) which provided the necessary computational power for our graph clustering experiments. References Achtert, E., Böhm, C., Kriegel, H.-P., Kröger, P., Müller-Gorman, I. & Zimek, A. (2006), Finding hierarchies of subspace clusters, in J. Fürnkranz, T. Scheffer & M. Spiliopoulou, eds, ‘PKDD’, Vol. 4213 of Lecture Notes in Computer Science, Springer, pp. 446–453. Achtert, E., Böhm, C., Kriegel, H.-P., Kröger, P., Müller-Gorman, I. & Zimek, A. (2007), Detec- tion and visualization of subspace cluster hier- archies, in K. Ramamohanarao, P. R. Krishna, M. K. Mohania & E. Nantajeewarawat, eds, ‘DASFAA’, Vol. 4443 of Lecture Notes in Com- puter Science, Springer, pp. 152–163. Achtert, E., Böhm, C., Kröger, P. & Zimek, A. (2006), Mining hierarchies of correlation clusters, in ‘SS- DBM’, IEEE Computer Society, pp. 119–128. Ankerst, M., Breunig, M. M., Kriegel, H.-P. & Sander, J. (1999), ‘Optics: ordering points to identify the clustering structure’, SIGMOD Rec. 28(2), 49– 60. Bekkerman, R., McCallum, A. & Huang, G. (2004), ‘Automatic categorization of email into fold- ers: Benchmark experiments on enron and sri corpora’, Center for Intelligent Information Re- trieval, Technical Report IR 418. Chen, L., Wright, P. & Nejdl, W. (2009), Improving music genre classification using collaborative tagging data, in ‘Proceedings of the Second ACM International Conference on Web Search and Data Mining’, WSDM ’09, ACM, New York, NY, USA, pp. 84–93. Diesner, J., Frantz, T. L. & Carley, K. M. (2005), ‘Communication networks from the enron email corpus ”it’s always about the people. enron is no different”’, Comput. Math. Organ. Theory 11(3), 201–228. Ester, M., Kriegel, H.-P., Sander, J. & Xu, X. (1996), A density-based algorithm for discovering clusters in large spatial databases with noise, in ‘Proc. of 2nd International Conference on Knowledge Discovery and Data Mining (KDD-96)’, pp. 226– 231. Falkowski, T. (2009), Community Analysis in Dynamic Social Networks, Sierke Verlag, Göttingen. Falkowski, T. & Barth, A. (2007), ‘Density-based temporal graph clustering for subgroup detection in social networks’, Presented at Conference on Applications of Social Network Analysis. Falkowski, T., Barth, A. & Spiliopoulou, M. (2007), Dengraph: A density-based community detection algorithm, in ‘Proc. of the 2007 IEEE / WIC / ACM International Conference on Web Intel- ligence’, IEEE Computer Society, Washington, DC, USA, pp. 112–115. Falkowski, T., Barth, A. & Spiliopoulou, M. (2008), Studying community dynamics with an incre- mental graph mining algorithm, in ‘Proc. of the 14 th Americas Conference on Information Systems (AMCIS 2008)’, Toronto, Canada. Geleijnse, G., Schedl, M. & Knees, P. (2007), The quest for ground truth in musical artist tagging in the social web era, in S. Dixon, D. Bainbridge & R. Typke, eds, ‘ISMIR’, Austrian Computer Society, pp. 525–530. Heer, J., Card, S. K. & Landay, J. A. (2005), prefuse: a toolkit for interactive information visualization, in ‘Proceedings of the SIGCHI conference on Human factors in computing systems’, CHI ’05, ACM, New York, NY, USA, pp. 421–430. Huang, J., Sun, H., Han, J. & Feng, B. (2011), ‘Density-based shrinkage for revealing hierar- chical and overlapping community structure in networks’, Physica A Statistical Mechanics and its Applications 390, 2160–2171. Klimt, B. & Yang, Y. (2004), Introducing the enron corpus, in ‘First Conference on Email and Anti- Spam (CEAS)’, Mountain View, CA. Konstas, I., Stathopoulos, V. & Jose, J. M. (2009), On social networks and collaborative recommenda- tion, in ‘Proceedings of the 32nd international ACM SIGIR conference on Research and de- velopment in information retrieval’, SIGIR ’09, ACM, New York, NY, USA, pp. 195–202. Kriegel, H.-P., Kröger, P., Ntoutsi, I. & Zimek, A. (2011), Density based subspace clustering over dynamic data, in J. B. Cushing, J. C. French & S. Bowers, eds, ‘SSDBM’, Vol. 6809 of Lecture Notes in Computer Science, Springer, pp. 387– 404. MacQueen, J. B. (1967), Some methods for classifi- cation and analysis of multivariate observations, in L. M. L. Cam & J. Neyman, eds, ‘Proc. of the fifth Berkeley Symposium on Mathematical Statistics and Probability’, Vol. 1, University of California Press, pp. 281–297. Mierswa, I., Wurst, M., Klinkenberg, R., Scholz, M. & Euler, T. (2006), Yale: Rapid prototyping for complex data mining tasks, in ‘KDD ’06: Proceedings of the 12th ACM SIGKDD in- ternational conference on Knowledge discovery and data mining’, ACM, New York, NY, USA, pp. 935–940. Schlitter, N. & Falkowski, T. (2009), Mining the dynamics of music preferences from a social networking site, in ‘Proceedings of the 2009 International Conference on Advances in Social Network Analysis and Mining’, IEEE Computer Society, Washington, DC, USA, pp. 243–248. Schlitter, N., Falkowski, T. & Lässig, J. (2011), Dengraph-HO: Density-based hierarchical com- munity detection for explorative visual network analysis, in ‘Research and Development in In- telligent Systems XXVIII Incorporating Appli- cations and Innovations in Intelligent Systems XIX’, Springer, London, pp. 283–296. Shetty, J. & Adibi, J. (2004), ‘The enron email dataset database schema and brief statistical report’. URL: http://www.cs.cmu.edu/∼enron/ Xu, X., Yuruk, N., Feng, Z. & Schweiger, T. A. J. (2007), Scan: a structural clustering algorithm for networks, in P. Berkhin, R. Caruana & X. Wu, eds, ‘KDD’, ACM, pp. 824–833. Yuruk, N., Mete, M., Xu, X. & Schweiger, T. A. J. (2007), A divisive hierarchical structural clus- tering algorithm for networks, in ‘ICDM Work- shops’, IEEE Computer Society, pp. 441–448. Yuruk, N., Mete, M., Xu, X. & Schweiger, T. A. J. (2009), Ahscan: Agglomerative hierarchical structural clustering algorithm for networks., in N. Memon & R. Alhajj, eds, ‘ASONAM’, IEEE Computer Society, pp. 72–77. 7. The authors 7.1. Nico Schlitter Nico Schlitter is currently working as head of the bwLSDF project at the Steinbuch Centre for Computing, Karlsruhe Institute of Technol- ogy. Much of his current work is related to distributed storage solutions for the state of Baden-Wuerttemberg. He is also head of the multidisciplinary grid computing project www. distributeddatamining.org where he works in the fields of social network analysis and time series analysis. After receiving his degree in computer science from Chemnitz University of Technology in 2006, he has been working at University of Magdeburg in the field of RFID-based data analysis and at University of Applied Sciences Zittau/Görlitz in the area of simulation and opti- mization. 7.2. Tanja Falkowski Tanja Falkowski is currently working as head of international relations at University of Göttingen. In 2009, she received her Ph.D. in computer sci- ence for her research on the analysis of commu- nity dynamics in social networks. She developed algorithms and methods to efficiently analyse temporal dynamics of group structures in social networks. Tanja Falkowski studied information systems at Technical University Braunschweig and wrote her diploma thesis at Haas School of Business, University of California at Berkeley. 7.3. Jörg Lässig Jörg Lässig is a Full Professor at the Faculty of Electrical Engineering and Computer Sci- ence at the University of Applied Sciences Zit- tau/Görlitz. He holds degrees in computer sci- ence and computational physics and received a Ph.D. in computer science for his research on ef- ficient algorithms and models for the generation and control of competence networks at Chemnitz University of Technology. He has been working in various research and industrial projects and is currently focusing on sustainable IT tech- nologies. His research interests include efficient algorithms, bio-inspired methods, and green in- formation systems. http://www.cs.cmu.edu/~enron/ www.distributeddatamining.org www.distributeddatamining.org Introduction Related and Previous Work DenGraph DenGraph-O DenGraph-IO DenGraph-HO Top-down Approach: Cluster Reduction, Split or Removal Bottom-up Approach: Cluster Creation, Absorption and Merging Data Structures Run Time Complexity Memory Requirements Applications Enron Case Study Last.fm Case Study Conclusion and Outlook Acknowledgment References The authors Nico Schlitter Tanja Falkowski Jörg Lässig 
work_jamfcdcnvjbabcmoye5b2ppkly	----	ESP: An expert system for poisoning diagnosis and management | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.3109/17538157.2010.490624 Corpus ID: 5972869ESP: An expert system for poisoning diagnosis and management @article{BatistaNavarro2010ESPAE, title={ESP: An expert system for poisoning diagnosis and management}, author={R. Batista-Navarro and Diana A. Bandojo and M. A. J. K. Gatapia and Reggie Nicolo C Santos and A. Marcelo and L. Panganiban and P. Naval}, journal={Informatics for Health and Social Care}, year={2010}, volume={35}, pages={53 - 63} } R. Batista-Navarro, Diana A. Bandojo, +4 authors P. Naval Published 2010 Computer Science, Medicine Informatics for Health and Social Care We describe a clinical decision support system (CDSS) designed to provide timely information germane to poisoning. [...] Key Method The system is implemented as a rule-based expert system with two major components: the knowledge base and the inference engine. The knowledge base serves as the database which contains relevant poisoning information and rules that are used by the inference engine in making decisions.Expand View on Taylor & Francis ncbi.nlm.nih.gov Save to Library Create Alert Cite Launch Research Feed Share This Paper 8 CitationsBackground Citations 3 View All Topics from this paper Poisoning Expert Systems Knowledge Bases CNS disorder Rule (guideline) Decision Support Systems, Clinical Inference Signs and Symptoms Acquired Immunodeficiency Syndrome Knowledge acquisition Decision Making Engineering 8 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency A Laboratory Test Expert System for Clinical Diagnosis Support in Primary Health Care Rodrigo Fernandez-Millan, José-Amelio Medina-Merodio, R. B. Plata, Jose-Javier Martinez-Herraiz, J. Gutierrez-Martinez Computer Science 2015 11 PDF Save Alert Research Feed The Design of a Cloud-based Clinical Decision Support System Prototype: Management of Drugs Intoxications in Childhood Baya Naouel Barigou, F. Barigou, Chawki Benchehida, B. Atmani, Ghalem Belalem Computer Science Int. J. Heal. Inf. Syst. Informatics 2018 1 Save Alert Research Feed Machine Learning: A Means of Diagnosing and Prescribing Treatments for Tuberculosis Patients A. James, E. Avenir, Jonas D. Villota, Maryli F. Rosas Medicine 2015 1 PDF Save Alert Research Feed Decision Support for Intoxication Prediction Using Graph Convolutional Networks Hendrik Burwinkel, M. Keicher, +4 authors Seyed-Ahmad Ahmadi Computer Science MICCAI 2020 PDF View 1 excerpt, cites background Save Alert Research Feed Antidote Application: An Educational System for Treatment of Common Toxin Overdose J. Long, Yingyuan Zhang, V. Brusic, L. Chitkushev, Guanglan Zhang Computer Science BCB 2017 View 1 excerpt, cites background Save Alert Research Feed A Rule-Based Computer-Aided System for Managing Home Accidents in Childhood Baya Naouel Barigou, B. Atmani, F. Barigou Computer Science ISCRAM-med 2016 Save Alert Research Feed Antidote application: an educational system for treatment of common toxin overdose Jon Long, Yingyuan Zhang, L. Chitkushev, V. Brusic, Guanglan Zhang Computer Science, Medicine BCB 2014 2 Save Alert Research Feed Clinical Decision Support Systems: Effective Solution for Diagnosis, Treatment, and Management of Patients Affected by Poisoning Zahra Mahmoudvand, S. Shadnia, Sharareh ROSTAM NIAKAN, M. Ghazisaeedi Medicine Iranian journal of public health 2019 PDF View 2 excerpts, cites background Save Alert Research Feed References SHOWING 1-10 OF 22 REFERENCES SORT BYRelevance Most Influenced Papers Recency SETH: an expert system for the management on acute drug poisoning in adults. S. Darmoni, P. Massari, +4 authors J. Leroy Computer Science, Medicine Computer methods and programs in biomedicine 1994 19 Highly Influential View 1 excerpt, references background Save Alert Research Feed Case-based reasoning for medical decision support tasks: The Inreca approach K. Althoff, R. Bergmann, +6 authors S. Gurov Computer Science, Medicine Artif. Intell. Medicine 1998 63 Highly Influential PDF View 1 excerpt, references background Save Alert Research Feed ESTHER — expert system for the diagnostics of acute drug poisonings O. Larichev, A. Asanov, Y. Naryzhny, Sergey Strahov Medicine 2002 11 Highly Influential View 1 excerpt, references background Save Alert Research Feed Flexible support for trauma management through goal-directed reasoning and planning B. Webber, R. Rymon, J. Clarke Computer Science Artif. Intell. Medicine 1992 63 View 1 excerpt, references background Save Alert Research Feed Design of Expert System for Search Allergy and Selection of the Skin Tests using CLIPS S. Karagiannis, A. Dounis, T. Chalastras, P. Tiropanis, D. Papachristos Computer Science 2007 27 View 1 excerpt, references background Save Alert Research Feed Case Report: Paracetamol Poisoning in a 2-Year-Old Child – from International Overview to the Role of the Hong Kong Poison Information Centre Jys Sia, Y. Chan Engineering 2006 6 PDF View 1 excerpt, references methods Save Alert Research Feed Rete: a fast algorithm for the many pattern/many object pattern match problem C. Forgy Computer Science 1991 1,803 PDF View 1 excerpt, references methods Save Alert Research Feed The Organophosphate Pesticides J. Kang, V. H. Zettel, N. Ward Chemistry 1995 25 Save Alert Research Feed JClips Available: http://sourceforge.net/projects/jclips/ via the INTERNET 2008 A Tool for Building Expert Systems [Internet Available from: http://clipsrules.sourceforge.net/ 2009 ... 1 2 3 ... Related Papers Abstract Topics 8 Citations 22 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_jbj4htluffecrc677zqmr5efya	----	Article Reference OECD's "Better Life Index": can any country be well ranked? KASPARIAN, Jérôme, ROLLAND, Antoine Abstract We critically review the Better Life Index (BLI) recently introduced by the Organization for Economic Co-operation and Development (OECD). We discuss methodological issues in the definition of the criteria used to rank the countries, as well as in their aggregation method. Moreover, we explore the unique option offered by the BLI to apply one's own weight set to 11 criteria. Although 16 countries can be ranked first by choosing ad hoc weightings, only Canada, Australia and Sweden do so over a substantial fraction of the parameter space defined by all possible weight sets. Furthermore, most pairwise comparisons between countries are insensitive to the choice of the weights. Therefore, the BLI establishes a hierarchy among the evaluated countries, independent of the chosen set of weights. KASPARIAN, Jérôme, ROLLAND, Antoine. OECD's "Better Life Index": can any country be well ranked? Journal of Applied Statistics, 2012, vol. 39, no. 10, p. 2223-2230 DOI : 10.1080/02664763.2012.706265 Available at: http://archive-ouverte.unige.ch/unige:54900 Disclaimer: layout of this document may differ from the published version. 1 / 1 http://archive-ouverte.unige.ch/unige:54900 April 17, 2012 18:32 Journal of Applied Statistics OCDE_kasparian_rolland_20120410 Journal of Applied Statistics Vol. 00, No. 00, Month 200x, 1–8 RESEARCH ARTICLE OECD’s “Better life index”: can any country be well ranked? Jérôme Kasparian1 and Antoine Rolland2 1 Université de Genève, GAP-Biophotonics, Chemin de Pinchat 22, CH-1211 Geneva 4, Switzerland - jerome.kasparian@unige.ch 2 Laboratoire ERIC, Université Lumière Lyon 2, 69676 BRON cedex, France - antoine.rolland@univ-lyon2.fr (Received 00 Month 200x; in final form 00 Month 200x) We critically review the Better Life Index (BLI) recently introduced by the Organization for Economic Co-operation and Development (OECD). We discuss methodological issues in the definition of the criteria used to rank the countries, as well as in their aggregation method. Moreover, we explore the unique option offered by the BLI to apply one’s own weight set to 11 criteria. Although 16 countries can be ranked first by choosing ad-hoc weightings, only Canada, Australia and Sweden do so over a substantial fraction of the parameter space defined by all possible weight sets. Furthermore, most pairwise comparisons between countries are insensitive to the choice of the weights. Therefore, the BLI establishes a hierarchy among the evaluated countries, independently on the chosen set of weights. 1. Introduction In May 2011, the Organization for Economic Co-operation and Development (OECD) proposed a new well-being index named “Better Life Index” (BLI) [9]. Following a tradition of research in multi-crietria evaluation in the economic and social fields [8], this index aims at offering an alternative to the Gross Domes- tic Product (GDP) to compare countries, taking into account not only the global amount of their wealth, but also well-being indicators. The BLI evaluates the 34 member states of OECD on 11 criteria, like housing, income, education, education, governance, etc. Each criterion is evaluated on a scale ranging between 0 and 10. A global country score is obtained by a weighted mean of the criteria. As emphasized by OECD, the innovative aspect of the BLI is the possi- bility offered to anyone to choose her/his own weights (as integer values between 0 and 5) in order to represent her/his own preferences on well-being indicators: “The OECD is NOT deciding what makes for better lives. YOU decide for yourself.” [9] In this paper, we present a critical review of this statement. In Section 2, we discuss biases due to the way how the criteria are built, as has been done for the Shanghai index of universities ranking [1] or various other composite indicators [2]. In Section 3, we investigate whether and how an appropriate set of weights allows a given country to be ranked first and analyze the most frequent rankings in the space of all the possible weights. By doing so, we exhibit an implicit and quite rigid country ranking established by the BLI, merely independent on the chosen set of weights. ISSN: 0266-4763 print/ISSN 1360-0532 online c© 200x Taylor & Francis DOI: 10.1080/0266476YYxxxxxxxx http://www.informaworld.com April 17, 2012 18:32 Journal of Applied Statistics OCDE_kasparian_rolland_20120410 2 2. Critical discussion of the criteria It is well-known that an indicator is always a partial view of the reality: its choice is then a matter of personal preference, i.e., of political approach. Therefore, it is always criticizable. In this regard, the option offered by the BLI to arbitrarily chose the weightings and/or select specific criteria among the 11 proposed ones constitutes a progress. For this reason, we will not discuss here the choice of the eleven criteria included by OECD in the BLI. However, even within this framework, technical criticism should be made about the construction of the indicators and associated scores. 2.1 Completeness of the criteria and indicators definition The content of the criteria offered for inclusion into the BLI could be enhanced. For example, the “Environment” criterion includes a single indicator, namely the average number of PM10 (particulate matter of aerodynamic diameter above 10 µm) in cities above 100,000 inhabitants. Other aspects of the environmental quality, such as other air pollutants, either urban or background, biodiversity preservation, water quality, CO2 emissions and so on are not considered. Scoring the environmental criterion with reference to a unique indicator, instead of considering the complexity of the matter of interest, could yield results misleading to the non-expert users. 2.2 Scoring of the criteria OEDC has produced and/or compiled the wide set of data needed to evaluate each indicator for each member state. But these data are hard to compare directly since they have heterogeneous scales. Furthermore, some of these indicators have to be maximized, while others should be minimized: this is for example the case for income and atmospheric pollution, respectively. Best practices in the field of multicriteria decision-making use specific methods to fix trade-off values between criteria, taking into account scale differences (see for example [4, 5]). In the BLI, the score of each criterion is normalized by a ratio method, i.e., it is obtained by applying to the scores of each indicator a linear function that scales to 0 and 10, respectively, the worst and best country scores for the considered indicator. The scores of a country therefore do not constitute an absolute measurement of its performance. They are rather relative to that of the best and worst countries for this indicator. Consequently, a country can have a bad score on an indicator not because its performance is intrinsically bad, but because one or several other countries have better performances in the considered domain. If the actual performances are almost equivalent, slight differences will result in artificially large contrasts in the scores. This is the case for the indicator "Time devoted to leisure and personal care" of the "Work-life balance" criterion, where all countries have very similar performances. The normalization sets the score of Mexico to 0 because of a relatively modest difference of only 2 hours weekly with the best countries. This behavior prevents any clear interpretation of the performance of a country or of its temporal evolution, and limits the relevance of the BLI to comparisons between countries at a given time. As can be seen from this example, due to the use of relative scores, pairwise com- parisons between two countries depend not only of their respective performances on each indicator, but also on the performances of all other countries since the lat- ter can reduce (resp. increase) the dynamics of any criterion by compressing (resp. expanding) the corresponding ranking scale through the above-discussed normal- April 17, 2012 18:32 Journal of Applied Statistics OCDE_kasparian_rolland_20120410 3 ization. The relative ranking of two countries can even be influenced by the perfor- mances of a third one, as exemplified in Appendix 1. Finally, the scaling into the 0–10 range is applied to each indicator independently rather than on the composite criteria. When several indicators are aggregated in a criterion, the global score for this criterion is the mean of the individual scores of the indicators. Since this mean is not subsequently renormalized, the scores of multi-indicator criteria do not span over the full 0 – 10 range, and hence lose dynamics. Its effective weight in the BLI is reduced as compared to mono-indicator criteria, which use the full range of scores from 0 to 10. Although this effect can be compensated by adequately overweighting the multi-indicator criteria, the omission to specify this limitation may be misleading, especially to non-expert users. 2.3 Global score construction The global index score of each country is obtained by a weighted mean of the scores of all criteria. The originality of OECD’s BLI is to let the user choose its own weights. While this aggregation technique has the advantage of being easily under- standable by all users including non-experts, it is known to strongly constrain the possible rankings, especially because it allows countries with heterogeneous profiles to stay well-ranked. Although this discussion lies beyond the scope of our present paper, it would be of great interest to study alternative aggregation procedures [7] relying on the raw data, which are available directly from the OECD web site [10]. Such improved aggregation procedures with more desirable properties [3, 6] include the min operator, which favors equilibrated countries, the OWA operator for which the weights are associated to the values and not to the criteria, or the Choquet integral which can even model specific interactions between topics, as can e.g. be expected between health and air quality, and more generally living conditions. The use of integer weights over a short range (between 0 and 5) is also welcome as a simple approach, but limits the dynamics as well as the possibility of fine tuning of the weights. We shall examine the effect of the discretization of the weights in the next Section. 3. Which country is whatsoever the most pleasant place to live? In spite of the limitations discussed in Section 2, we shall discuss here the relevance of OECD’s slogan within the framework defined by the specification of the BLI and investigate to which extent the choice of the weights of each criterion influences the ranking of the countries. More precisely, we focus our study on two mains aspects, corresponding to complementary points of view. First, can the BLI be instrumented by e.g., a government to exhibit an order in which his/her country is ranked as well as possible? In other words, can one find a set of weights optimizing the ranking of a given country? Conversely, assuming that one has no preference among the proposed criteria, one could weight them randomly. What would be the influence on the ranking? We addressed this question by computing the probability for each country to be ranked first, or to be ranked better than a given other one for random weights. 3.1 Optimizing ranking of a predefined country Let us first observe that, among the 34 member states of OECD, 12 are Pareto- dominated by at least another one, i.e., they have lower scores on each of the 11 April 17, 2012 18:32 Journal of Applied Statistics OCDE_kasparian_rolland_20120410 4 criteria. Therefore, they cannot be ranked first, regardless of the the considered set of weights. This limits ad hoc manipulations of the BLI. To go beyond this simple observation, we maximized the score difference with the best of the remaining population (Table 1, second column). For a country A, this score difference is defined as the difference between the score of A, and (i) if A can be ranked first, the country ranked second when using the same set of weights, or (ii) if A cannot be ranked first, the country ranked first when using the same set of weights. In this optimization, which we performed using linear programming, the free parameters are the weights of the 11 criteria. We made the problem continuous by releasing the constraint of integer weights, allowing them to vary continuously between 0 and 1. To avoid convergence to "pathological" solutions (e.g., reducing all weights towards zero, in which case all countries would be ranked first ex-aequo), as well as to allow comparing the score differences between countries (See Col. 2 of Tables 1 and 2), the sum of the weights was normalized to 1. Optimizing the score difference allows a robust optimization since the optimizing observable is continuous. But its interest is mostly restricted to countries that can be ranked first or second. To gain information about the other countries, we minimized the ranking. As displayed in the third column of Table 1, 16 countries, i.e., almost half of the OECD, can be ranked first if adequate weight sets are chosen. However strong discrepancies are observed in the differences of scores with the best of all other countries. Two third of the OECD countries can be ranked in the first three positions. Conversely, the poor optimum ranking of several countries evidences that some countries are intrinsically better ranked in the BLI than others and illustrates the limitations to the possible manipulations. We then investigated the impact of discrete weights on this result, by comparing it with the systematic exploration of the 611 possible sets of integer weights between 0 and 5. The sum of weights of each set is then normalized to 1 in order to allow comparison with Table 1. Comparing the results displayed in Tables 1 and 2 shows that the discrete weights strongly constrain the ranking, and to a lesser extent the score differences. They prevent 6 countries from being ranked first. Among them, Japan and Korea can be only be ranked 12th and 13th at best with discrete weights. This discrepancy illustrates that imposing discrete weights substantially influence the freedom of the user to fine tune her/his preferences, or to manipulate the ranking. 3.2 Robustness analysis These results clearly show that, provided one accepts the criteria defining the BLI, the OECD member states do not all provide the same well-being, regardless the weights assigned to the BLI criteria. In other words, the BLI defines a hierarchy of the countries, which is only marginally affected by the choice of the weights. To get more insight into this hierarchy, we explored the [0;5]11 parameter space defined by the 11 weights, by performing Monte-Carlo calculations. Such approach is clas- sical for exploring parameter spaces too large to allow a systematic investigation. A textbook example of their use is the estimation of a sub-volume of a space, e.g. to estimate the volume of a sphere and estimate the value of π [11]. In our work, over 430 million sets of weights (i.e., of coordinates in the parameter space) were randomly chosen in the [0;5]11 parameter space. For each set, each weight was ran- domly chosen in the [0;5] interval and the corresponding BLI ranking was calculated for each country. Among the resulting sample of 430×106 rankings, we computed the probability of each country to be ranked first and collected its best ranking and relative score relative to other countries. These Monte-Carlo computed values April 17, 2012 18:32 Journal of Applied Statistics OCDE_kasparian_rolland_20120410 5 Country Best score relative to others Best possible ranking Probability of # 1 ranking Canada 0.97 1 0.47 Australia 1.25 1 0.39 Sweden 0.63 1 0.10 Iceland 1.05 1 1.1 × 10−2 New Zealand 0.40 1 7.0 × 10−3 Switzerland 0.81 1 7.0 × 10−3 United States 2.16 1 6.4 × 10−3 Denmark 0.83 1 5.4 × 10−3 Norway 0.44 1 4.1 × 10−3 Finland 0.63 1 ε Netherlands 0.14 1 ε Belgium 0.22 1 ε Ireland 0.11 1 ε Japan 0.40 1 ε Korea 4.5 × 10−3 1 ε United Kingdom 0.10 1 ε Austria -0.08 2 0 France -0.18 2 0 Estonia -0.09 2 0 Germany -0.06 2 0 Slovak Republic -0.27 3 0 Poland -0.28 3 0 Luxembourg -0.30 3 0 Israel -0.61 6 0 Spain -0.69 6 0 Chile -1.27 8 0 Slovenia -0.49 9 0 Hungary -0.58 9 0 Czech Republic -0.75 9 0 Greece -0.76 10 0 Portugal -0.97 10 0 Italy -0.77 11 0 Mexico -2.06 14 0 Turkey -2.34 15 0 Table 1. Optimized the score and ranking of each country, as determined by linear programing and considering continuously varying weights. The best score relative to others is the maximum achievable difference between the considered country and the best among the other ones. ε denotes probabilities below 10−4, where the Monte-Carlo algorithm may provide unreliable probabilities. provide approximations of the real ones. Owing to the extremely large number of trials, the 95% confidence interval are minimal, as the precision for each proportion is ±4.7 × 10−5 in the worst case. As shown in the last column of Table 1, only three countries are ranked first over a substantial fraction of the parameter space: Canada (47%), Australia (39%) and to a lesser extent Sweden (10%), for a total of 96%. 13 other countries can only be ranked first at the cost of the choice of a very specific set of weights, as evidenced by the small, or even infinitesimal volume of the parameter space in which their scores exceeds that of the other countries. This meta-view of the "neutral" approach of OECD confirms the hierarchy between the countries set by the BLI and contradicts to a large extent OECD’s statement that the BLI does not intrinsically bear a predetermined country ranking. April 17, 2012 18:32 Journal of Applied Statistics OCDE_kasparian_rolland_20120410 6 Country Best score relative to others Best possible ranking Probability of # 1 ranking Canada 0.95 1 0.52 Australia 0.84 1 0.34 Sweden 0.34 1 8.9 × 10−2 Iceland 0.35 1 1.4 × 10−3 New Zealand 0.044 1 8.1 × 10−6 Switzerland 0.54 1 1.0 × 10−2 United States 0.66 1 3.1 × 10−3 Denmark 0.80 1 2.8 × 10−2 Norway 0.26 1 8.7 × 10−3 Finland 0.14 1 9.3 × 10−6 Netherlands -0.014 2 0 Belgium -0.51 7 0 Ireland -0.070 2 0 Japan -1.2 12 0 Korea -1.3 13 0 United Kingdom -0.44 6 0 Austria -0.33 5 0 France -0.68 10 0 Estonia -2.2 22 0 Germany -0.65 10 0 Slovak Republic -1.4 17 0 Poland -1.7 18 0 Luxembourg -0.63 8 0 Israel -0.73 6 0 Spain -1.4 16 0 Chile -2.3 17 0 Slovenia -1.5 17 0 Hungary -2.4 26 0 Czech Republic -1.5 17 0 Greece -2.0 21 0 Portugal -2.6 24 0 Italy -1.5 17 0 Mexico -2.3 14 0 Turkey -3.9 28 0 Table 2. Systematic test of all possible set of integer scores between 0 and 5 for each country. The best score relative to others is the maximum achievable difference between the considered country and the best among the other ones. Countries are sorted in the same order as in Table 1 . Another approach in trying to rank the countries is to determine, for a given pair of them, the respective volumes of the weight space where one is preferred to the other, i.e. it obtains a better score. The result obtained for all pairs of countries using 100.000 random weight sets in a Monte-Carlo calculation, offering a 95 % confidence interval of ±0.003, is displayed in Figure 1. The most striking feature is that the pairwise comparisons depends little on the choice of the weight set. Over 528 non-trivial pairs, the preference of 262 ones (49.6%) cannot be inverted by the choice of the weight set, while for another 312 of them (59%), the preference in one direction is achieved over less than 0.1% of the parameter space. Conversely, only 33 pairs (6.3%) share the parameter space within 40–60% preference probability. This confirmation of an implicit underlying hierarchy of the countries is evidenced in Figure 1 by the large areas of the graph with intense red or green colors, as April 17, 2012 18:32 Journal of Applied Statistics OCDE_kasparian_rolland_20120410 REFERENCES 7 C o u n tr y A 100% 0% 50% Country B P ro b a b il it y f o r A t o b e b e tt e r ra n k e d t h a n B Canada Australia Sweden Iceland New Zealand Switzerland United States Denmark Norway Finland Netherlands Belgium Ireland Japan Korea United Kingdom Austria France Estonia Germany Slovak Republic Poland Luxembourg Israel Spain Chile Slovenia Hungary Czech Republic Greece Portugal Italy Mexico Turkey C a n a d a A u st ra li a S w e d e n Ic e la n d N e w Z e a la n d S w it ze rl a n d U n it e d S ta te s D e n m a rk N o rw a y F in la n d N e th e rl a n d s B e lg iu m Ir e la n d Ja p a n K o re a U n it e d K in g d o m A u st ri a F ra n c e E st o n ia G e rm a n y S lo v a k R e p u b li c P o la n d L u x e m b o u rg Is ra e l S p a in C h il e S lo v e n ia H u n g a ry C ze c h R e p u b li c G re e c e P o rt u g a l It a ly M e x ic o T u rk e y Figure 1. Preference probability for each pair of countries. On the axis, countries are sorted according to their probability within the parameter space to be ranked first, then according to their best ranking (See Table 1) opposed to light values. Furthermore, the hierarchy exhibited from the pairwise preference largely overlaps that issued from overall ranking, as evidenced by the segregation of preferences over the diagonal of Figure 1. 4. Conclusion As a conclusion, we critically reviewed the BLI recently introduced by the OECD. Methodological issues include the use of relative scores rather than absolute ones and aggregation of the several indicators used for each criterion. Some criteria could also be enriched by the addition of complementary indicators. Furthermore, the constraint for integer weights clearly constrains the rankings. We also explored the unique option offered by the BLI to the users to choose their own weight sets between 11 criteria. Although adequate weightings allow 16 countries to be ranked first, only Canada, Australia and Sweden do so over a substantial fraction of the parameter space defined by all possible weight sets. Furthermore, most pairwise comparisons between countries are either totally or highly insensitive to the choice of the weight set. Therefore, the choice of these weights does not affect much the country ranking. Rather, as a weighted sum model, the BLI defines a quasi-hierarchy among the evaluated countries. References [1] J.-C. Billaut, D. Bouyssou, and Ph. Vincke, Should you believe in the Shanghai ranking? An MCDM view, Scientometrics, 84 (2010), pp. 237–263 April 17, 2012 18:32 Journal of Applied Statistics OCDE_kasparian_rolland_20120410 8 REFERENCES [2] D. Bouyssou, T. Marchant, M. Pirlot, P. Perny, A. Tsoukias, and P. Vincke, Evaluation and Decision Models: A Critical Perspective, Springer, Heidelberg, 2001 [3] M. Grabisch, J.L. Marichal, R. Mesiar, and E. Pap: Aggregation functions, Cambridge, UK, Cambridge University Press, 2009 [4] E. Jacquet-Lagrèze and J. Siskos, Assessing a Set of Additive Utility Functions for Multicriteria Decision-Making, the UTA method, European Journal of Operational Research, 10 (1982), pp. 151-164 [5] R.L. Keeney and H. Raiffa, Decisions with Multiple Objectives: Preferences and Value Tradeoffs, J. Wiley, New York, 1976 [6] J.L. Marichal, Aggregation functions for decision making, Decision-Making Process - Concepts and Methods, ISTE/John Wiley, 2009, pp. 673-721. [7] P. Meyer and G. Ponthière, Eliciting Preferences on Multiattribute Societies with a Choquet Integral, Computational Economics, 37 (2011), pp. 133–168, 2011 [8] G. Munda, Social Multi-criteria Evaluation for a Sustainable Economy, Springer, Heidelberg, New York, 2008 [9] OECD, http://www.oecdbetterlifeindex.org/wpsystem/wp-content/uploads/2011/05/ YourBetterLifeIndex_ExecutiveSummary2.pdf [10] OECD, http://www.oecdbetterlifeindex.org/wpsystem/wp-content/uploads/2011/06/ BetterLifeIndex_Data_2011V6.xls [11] William H. Press, Saul A. Teukolsky, William T. Vetterling, and Brian P. Flannery, Numerical Recipes 3rd Edition: The Art of Scientific Computing, Cambridge University Press, 2007 Appendix A: Example 1 A sentence like “life is better in New Zealand than in Finland. But if South Africa and Canada are making efforts to reduce unemployment, life will be better in Fin- land than in New Zealand” seems dummy. But this counterintuitive behavior is induced by the way scores on each topic are calculated in the BLI, whatever the weight on each topic. Following example illustrates this possibility. Let us consider the scores of 4 fictious countries on two indicators, and the cor- responding normalized scores on the topics. In this example, we consider equal weights for both topics: Indicator data Scores Country topic 1 topic 2 topic 1 topic 2 global index Country A 100 100 10 10 10 Country B 55 40 5.5 4 4.75 Country C 45 60 4.5 6 5.25 Country D 0 0 0 0 0 Country C is ranked before country B with regard to the global index. Assume now that the situation has changed only on countries A and D as follows: Indicator data Scores Country topic 1 topic 2 topic 1 topic 2 global index Country A 100 200 10 10 10 Country B 55 40 4.375 2 3.18 Country C 45 60 3.125 3 3.06 Country D 0 20 0 0 0 Absolute performances of countries B and C have not changed, but country B is now better ranked than country C. 
work_jca34ttijfecfg2vq5xtbkygb4	----	PII: S0957-4174(96)00067-X Towards a Framework to Verify Knowledge Sharing Technology A S U N C I Ó N G Ó M E Z - P É R E Z Laboratorio de Inteligencia Artificial, Facultad de Informática, Universidad Politécnica de Madrid, Campus de Montegancedo sn., Boadilla del Monte, 28660 Madrid, Spain Abstract—Based on the empirical verification of bibliographic-data and other Ontolingua ontologies, this paper provides an initial framework for verifying Knowledge Sharing Technology (KST). Verification of KST refers to the engineering activity that guarantees the correctness of the definitions in an ontology, its associated software environments and documentation with respect to a frame of reference during each phase and between phases of its life cycle. Verification of the ontologies refers to building the correct ontology, and it verifies that (1) the architecture of the ontology is sound, (2) the lexicon and the syntax of the definitions are correct and (3) the content of the ontologies and their definitions are internally and metaphysically consistent, complete, concise, expandable and sensitive. Copyright © 1996 Elsevier Science Ltd 1. INTRODUCTION DURING RECENT YEARS, considerable progress has been made in developing the conceptual bases for building technology that allows the reuse and sharing of knowl- edge. As libraries of definitions, ontologies are essential for both building intelligent systems and enabling interoperation of agents. Evaluation of ontologies as well as evaluation of the documentation and software envi- ronments is critical to the integration of this technology in real applications. In this sense, judging an ontology to be used for knowledge representation is just as necessary as getting a thorough inspection before purchasing a second-hand car. Just as one does not buy a car without first taking it to a mechanic for a checkup to ensure that the car is mechanically sound, it is unwise to publish your ontology, or to implement a software application that relies on ontologies written by others, without first evaluating and assessing their definitions and axioms. A well-evaluated ontology will not guarantee the absence of problems, but it will make its use safer. The word "ontology" is a fashionable word in the knowledge engineering community. Although there exist many possible interpretations of this term, Guarino and Giaretta (1995) discuss different definitions and propose the following interpretation. An ontology is: (sense 1) a logical theory which gives an explicit, partial account of a conceptualization; (sense 2) a synonym of con- ceptualization. Although ontologies differ from knowledge bases (KBs) (Gómez-Pérez, 1994), both have a common foundational problem. Ontologies and KBs by nature are incomplete, as it is impossible to capture everything known about the real world in a finite structure. Since ontologies are developed incrementally by adding new definitions and modifying the old ones, one of the most important problems is to guarantee complete, consistent and concise definitions from the beginning, during each stage and between stages of its development process. The maintenance of ontologies would also require complete evaluation of the whole ontology if a definition is added, modified or removed. Therefore, the knowledge sharing community needs to draw up a complete framework (Gómez-Pérez et al., 1995) with terminology, definitions, criteria, methods and tools for Knowledge Sharing Technology (KST) evaluation. This evaluation includes: ontologies, soft- ware and documentation. Some works have been done: (a) Related to terminology and definitions of terms, the main terms are (Gómez-Pérez et al., 1995): "evaluation", "verification", "validation" and "assessment". In this paper, the differences between "evaluation" and "assessment" are also emphasized. • "Evaluation of KST" subsumes "verification" and "validation". "Evaluation" means to judge the ontologies, their associated software envi- ronments and documentation technically with respect to a frame of reference during each phase and between phases of their life cycle. Examples of frames of references are the real world, a set of requirements or a set of competency questions (Gruninger & Fox, 1994). • "Verification of KST" refers to the technical activity that guarantees the correctness of an ontology, its associated software environments and documentation with respect to a frame of reference during each phase and between phases of its life cycle. • "Validation of KST" guarantees that the ontolo- gies, the software environments and the documentation correspond to the systems that they are supposed to represent. • "Assessment of KST" refers to the usability and utility of the ontologies, software environments, and their documentation when they are reused by KB S or shared by software agents. This paper is focused on verification of KST, that is, verification of ontologies, verification of software and verification of documentation. The next sections cover these issues. (b) In relation to the criteria for evaluating knowl- edge sharing technology, a few examples of verification of class definitions, hierarchies and relations and functions appear in Gómez-Pérez (1995). (c) With regard to the methods, competency questions (Gruninger & Fox, 1994) are proposed as a methodology for evaluating ontologies in the domain of enterprise engineering. The compe- tency questions are the basis for a rigorous characterization of the problems that the ontology has to cover, and they specify the problem and what constitutes a good solution to the problem. (d) Concerning tools, Ontolingua (Gruber, 1993a) provides a parser for legal KIF (Genesereth & Fikes, 1993) sentences, analyses of whether definitions are well formed and generates a report on undefined concepts and intra-ontology dependencies. Since the evaluation of knowledge sharing technology is very immature and there is an absence of a deep core of previous ideas, this paper is focused exclusively on providing some criteria and examples that guide the verification of KST. The paper gathers the experience on the evaluation of the bibliographic-data (Gruber, 1994) and other Ontolingua (Gruber, 1993a) ontologies. It is organized as follows. Section 2 deals with how to verify ontologies. Section 3 shows how to verify software used to build ontologies. Finally, Section 4 shows verification of documentation. 2. VERIFICATION OF ONTOLOGIES Ontology verification refers to correct building of the ontology, that is, ensuring that its definitions1 correctly implement its requirements, its competence questions or perform correctly in the real world. Ontologies verifica- tion is orthogonal to the use of the definitions by any KBS or software agent. Ontology verification includes verification of: (1) Each individual definition and axiom. (2) Collection of definitions and axioms that are stated explicitly in the definitions of the ontology. (3) Definitions that are imported from other ontologies. (4) Axioms that can be inferred using other definitions and axioms. To verify an ontology we have to determine the correctness of definitions and axioms by figuring out what the ontology explicitly defines, does not define or defines incorrectly. We also have to look at the scope of definitions and axioms by figuring out what can be inferred, cannot be inferred or can be inferred incor- rectly. To guarantee that an ontology is well-verified, we have to judge its architecture, its lexicon and syntax and its content using criteria specified in Table 1. 2.1. Verification of the Architecture At this point, we look to see if the structure of an ontology has been developed following the principles of design of the environment in which the ontology is included. For example, ontologies built in the Ontolingua environment should satisfy the five design criteria given by Gruber (1993b). 2.2. Verification of the Lexis and Syntax The ontology definitions must be lexically and syntac- tically correct. The environment should provide a scanner to detect that the lexical structure of the expressions is correct, and a parser to detect that its syntactic structure is also correct. It is particularly important that the lexical and syntax analyzer compo- TABLE1 Levels and Criteria in the Verification of Ontologies Levels Criteria Verification of the architecture Verification of the lexicon and syntax Verification of the content soundness correctness consistency, completeness, conciseness, expandability and sensitiveness ' A definition is written in natural language (informal definition) and in a formal language (formal definition). nents of the software environment rigorously implement the definitions of the lexis and grammar rules for the portable language. Failure to do this will allow the writing of non-portable definitions. As an example we have the use of wrong keywords in formal definitions. 2.3. Verification of Content Verification of the content is concerned with the analysis of completeness, consistency, conciseness, expandability and sensitiveness of the definitions and axioms that are explicitly set out in the ontology, and with the analysis of those that can be inferred using other definitions and axioms. 2.3.1. Consistency. Consistency refers to whether it is possible to obtain contradictory conclusions from valid input data (Gómez-Pérez, 1996). With the goal of providing mechanisms that help to verify semantically the consistency of an ontology and its definitions, we assume that: • A definition Def is composed of an informal definition IDef and a formal definition FDef. • An informal definition IDef is a free text documentation written in English. • A formal definition FDef is a collection of sentences written in a formal language. r Def= ((Senti)...(Sentn)) Since the semantics of KIF (Knowledge Interchange Format)2 unambiguously determines the referent of any term and the truth or falsity of any sentence, we assume that formal definitions are written in this language. • Given a definition Def, the function InterpretationFDef (IDef) interprets the meaning of an informal definition IDef with respect to its formal definition FDef. This function maps the documentation string IDef into the truth values true or false. Interpretation FDejJDe¡):IDef= :> {true, false} • Defined(Def Ont) is a function that determines if the 2 The semantics of KIF is a correlation between the terms and sentences of the language and a conceptualization of the world. The semantic value of a term and the truth value of a sentence are defined using the notions of interpretation of constants and variable assign- ment. An interpretation is a function i that associates the constants of KIF with the elements of a conceptualization. A variable assignment is a function v that maps (1) individual variables V into objects in a universe of discourse O and (2) maps sequence variables W into finite sequences of objects. Given an interpretation and a variable assign- ment, the semantic value of every term in the language is a function siv from the set T of terms into the set O of objects in the universe of discourse. The truth value for sentences is defined as a function tw that maps sentences 5 into the truth values true ox false. definition Défis defined in the ontology Ont. Defined(DefOnt)= ( true if Def is defined in Ont \ false otherwise * Inferred(FSenl Def Ont) is a function that determines if the formal sentence F&„, is inferred using the definition Def and the ontology Ont. Inferred(FSeM Def Ont) = ( true ifFSent is inferred using Def and Ont [false otherwise An ontology Ont is semantically consistent S-Con- sistency{Ont) if, and only if, each definition Def in the ontology is semantically consistent. S- Consistencyi Ont)<¿> {Ç4Def)Defmed(Def Ont)A S-Consistency0n,(Def)) A given definition Def in the ontology Ont is semanti- cally consistent S-Consistency0„,(Def) if, and only if: (1) the individual definition is consistent and (2) no contradictory sentences may be inferred using other definitions and axioms. (VDef, Ont) S-Consistency0m(Def)<=> (S-Individual-ConsistencyDej(Def)A S-Inferred- Consistency0„,(Def) ) 2.3.1.1. Individual Consistency. A given definition Def is individually consistent S-Individual-ConsistencyDef (Def) if, and only if: (1) the definition Def is metaphys- ically consistent, that is, it is consistent with respect to the real world RW and (2) it is internally consistent. S-Individual-ConsistencyDeJ(Def)<=> S-ConsistencyRv^Def)/\S-ConsistencyDe^Def) To guarantee that the definition Def is metaphysically consistent S-ConsistencyRW(Def), we prove that its formal as well as its informal definitions are metaphysically consistent. S-ConsistencyRW(Def)< d> (S-ConsistencyRK(FDef)A S-ConsistencyRW(IDei)) A formal definition FDef is metaphysically consistent S- ConsistencyRW(FDef) if, and only if, there is no contradiction in the interpretation of the formal defini- tion with respect to the real world. The goal is to prove compliance of the world model (if it exists and is known) with the world modeled formally. So, S-Consis- tencyRW(FDef) maps a formal definition FDef into the truth values true or false. S-ConsistencyRW(FDef):FDej=>{true, false} Since a formal definition is a set of KIF sentences, the function S-ConsistencyRW(FDí.f) is equivalent to determin- ing the truth value of each KIF sentence Sent¡ in the formal definition. S-ConsistencyRW((Senti)...(Sentn)) = ( true &tiv(Sent¡)=true for all i in [ 1. .n] \ false otherwise An informal definition IDef is metaphysically consistent S-ConsistencyRW(IDef) if, and only if, there is no contra- diction in the interpretation of the informal definition with respect to the real world. The goal is to prove the compliance of the world with the world modeled informally. This function maps the documentation string IDef into the truth values true or false. S-ConsistencyRW(IDi;fy.IDef=*{true, false] We assure that the definition Def is internally consistent S-Consistency^Def), by proving that its formal as well as its informal definitions have the same meaning. 5- Consistency DeJ(Def)<^ (Interpretation FAIDef) - S-ConsistencyRW(FDe^) To prove the individual consistency of the definition MONTH-NAME in Example 1, we verify its internal and metaphysical consistency. As the terms used to name the months are the same in the formal and informal definitions, the definition of MONTH-NAME is inter- nally consistent. However, both, its formal and informal definitions are metaphysically inconsistent because the term "house" is not a month in the real world. If we replace the term "house" by the term "January" in the formal definition of MONTH-NAME, then the whole definition is internally inconsistent, the formal definition is metaphysically consistent, and the informal definition is metaphysically inconsistent. To solve the incon- sistencies, we replace the term "house" by "January" in the informal definition. However, if we were to replace the term "house" by the term "Enero" (this means January in Spanish), for those English speakers who are not Spanish speakers there is still a metaphysical inconsistency in the informal definition (something other than January is written in the informal definition). However, for those who are Spanish speakers, the formal definition and the informal definition are metaphysically consistent, but the whole definition is internally incon- sistent because the symbols that name the months are different. (Define-Class MONTH-NAME (?Month) "The months of the year are: House, February, March, April, May, June, July, August, September, October, November, December" :iff-def (Member ?Month (setof House February March April May June July August September Octo- ber November December))) Example 1. Internally consistent definition, but not metaphysically consistent 2.3.1.2. Inferred Consistency. For a definition to be inferentially and semantically consistent, it must be impossible to obtain contradictory conclusions using the meaning of all the definitions and axioms in the current logical theory. We guarantee the inferred consistency of a given definition Inferred-Consistency0nt(Def) by prov- ing that if A is the set of inferred sentences for a given definition Def, (1) each inferred formal sentence FSm is individually consistent with respect to the definition Def and that (2) the set A of inferred sentences is internally consistent. OfFSentF'SM€A) QiDefOnuOnt') (Inferred-Consistency0n,(Def)^ (Defined(DefOnt)A Inferred(FSemDefOnt')A S-Individual-FSen,-ConsistencyDe](Fse„t)A Inferred(F'SenlDefOnt')A S-A-Consistency(FSenlF'&„,)) To assure that an inferred formal sentence is individually consistent with respect to the definition S-Individual- Fsent-ConsistencyDe^FSenl), we prove that: (1) there are no contradictions between the interpretation of the formal definition FDef and the interpretation of the inferred formal sentence FSe„, with respect to the real world and (2) there are no contradictions between the interpretation of the informal definition I0ef regarding the formal definition FDef and the interpretation of the inferred formal sentence FSenl regarding the real world. (VDef,FSi.n,)S-Individual-FSenl-ConsistencyDe](FSeJ<^ (((S-ConsistencyRW(FDef)=S-ConsistencyRW(FSen,))A (Interpretation fDef(IDef)=S-ConsistencyRW(FSe„,))) We guarantee that a set A of inferred formal sentences is internally consistent S-A-Consistency(FSen, F'Se„,), by proving that there are no contradictions between the interpretation of any inferred formal sentence FSenl and the interpretation of any other inferred formal sentence F'senr (VFSenlF'SMeA)S-A-Consistency(FSenlF'Senl)** (S-ConsistencyRV¿FSenl)=S-ConsistencyRW(F'Sen,)) Taking definitions in Example 2, the definition of KEYWORD would seem to be individually consistent. Since KEYWORD is a subclass of BIBLIO-TEXT, we can infer the formal sentence (string ?keyword), which means that ?keyword is a string. So, there is a semantically inferred inconsistency between the mean- ings of the inferred formal sentence and the informal definition of KEYWORD. 2.3.2. Completeness. Incompleteness is a fundamental problem in ontologies. In fact, we cannot prove either the (Define-Class BIBLIO-TEXT (?String) "The general class of text objects" :def (String?String)) (Define-Class BIBLIO-NAME (?String) "A name of something in the bibliographic-data ontology" :def (Biblio-Text?String)) (Define-Class KEYWORD (?Keyword) "A keyword is a number used as an index" :def (Biblio-Name?Keyword)) Example 2. Inferred inconsistency. completeness of an ontology or the completeness of its definitions, but we can prove both the incompleteness of an individual definition to derive the incompleteness of the ontology and the incompleteness of an ontology if at least a definition is missed. So, an ontology is semanti- cally complete if, and only if: (1) All that is supposed to be in the ontology is explicitly set out in it, or can be inferred using other definitions and axioms. (2) Each definition is complete. Semantic completeness of a definition refers to the degree to which the definitions in a user-independent ontology cover the equivalent concepts in the real world. We determine the completeness of a definition by figuring out: (a) what information the definition defines or does not explicitly define about the world; and (b) for all the information that is not explicitly defined, but required, we check if it can be inferred using other axioms and definitions. If it can be inferred, the definition is complete. Otherwise, it is incomplete. Completeness of the definitions concerns complete- ness of their formal and informal definitions. An informal definition written in natural language is com- plete if it expresses the same knowledge that the formal definition provides. To determine whether a formal definition or collection of formal definitions is complete we need a frame of reference, certain criteria to measure the degree of completeness and some guidelines to perform it. If there are no requirements or competency questions to be used as a frame of reference, other sources of information such as: the real world, relevant experts in developing ontologies, relevant users, books, examples, other ontologies could be used. In this case, the incompleteness of an ontology can be established by failure of test of any of the following three properties. Scope, which specifies the variety of different types of applications that might reuse or share the definitions; exhaustiveness, which refers to the level of precision of the definitions; granularity, which denotes the level of detail reached in each individual definition, as well as in the ontology. In order to provide a mechanism to verify the completeness of an ontology, we assume that the world is conceptualized in terms of KIF objects, relations and functions. The following ordered set of activities might help you to find incomplete definitions in an ontology. Step 1: Check completeness of the class hierarchy in which the current definition is included. The goal is to determine whether the superclasses of a given class exactly and precisely delimit the subclasses/superclasses of the appropriate class in the real world. Errors appear when: (a) the superclasses or subclasses of a given class are imprecise, over-specified or when they include classes that are not appropriate in the real world and (b) information about subclasses that are mutually disjoint or exhaustive subclass partitions are missing in their superclasses. Assume the following ontolingua classes definitions3 Def. 1. (Define-Class DOCUMENT (?X) "A document is something created by author(s) that may be viewed, listened to, etc., by some audi- ence..." :def (And (Individual-Thing?X) (Has-One?X Title-Of) (Has-One?X Number-Of-Pages-Of)) :axiom-def (Subclass-Partition Document (Setof Book Thesis Miscellaneous-Publication))) Def. 2. (Define-Class BOOK (?X) "Pages in a bound cover. You can't judge it by its cover." :def (And (Document?X) (Has-Some?X Has-Author) (Has-One?X Title-Of))) Def. 3. (Define-Class THESIS (?X) "An official report on a bout of graduate work for which one receives a degree, published by the university. Never mind that some fields make a big deal about the difference between dissertations and theses. From the bibliographic perspective, they are both of the same family." :def (And (Document?X) (Has-One-Of-Type?X Organization-Of University))) Def. 4. (Define-Class MASTERS-THESIS (?X) "M.S. thesis document." :def(Thesis?X)) 3 These definitions might not correspond with definitions in the Bibliographic-Data ontology. Def. 5. (Define-Class DOCTORAL-THESIS (?X) "Ph.D. thesis document." :def(Thesis?X)) Def. 6. (Define Class MISCELLANEOUS-PUBLI- CATION (?X) "A miscellaneous category of documents that are infrequently found in bibliographic references" :def ((Document?X)) Def. (?X) 7. (Define-Class COMPUTER-PROGRAM 'The Has-Author is the programmer." :def (Miscellaneous-Publication ?X)) Def. 8. (Define-Class PICTURE (?X) :def (Miscellaneous-Publication ?X)) Their attached hierarchy is given in Fig. 1—which does not exactly correspond to the hierarchy of the bibliographic-data ontology (Gruber, 1994)—in which, bold words represent classes, italic words mean proper- ties attached to the class, plain lines between classes represent subclass-of relations between classes and dashed lines mean that the subclasses of a class are mutually disjoint. Notice that: • The classes Doctoral-Thesis and Master-Thesis are an exhaustive subclass partition of the class THE- SIS. The following Ontolingua sentence should be included in the definition of Thesis. :axiom-def (Exhaustive-Subclass-Partition Thesis (Setof Masters-Thesis Doctoral-Thesis))) • The classes Computer-Program and Picture should be mutually disjoint with respect to the class Miscellaneous-Publication. The following should be included in the definition of Miscellaneous-Publica- tion. :axiom-def (Subclass-Partition Miscellaneous-Publication (Setof Computer-Program Picture))) • The specialization of the class Book into different subfields, that is, Computer-Book, Chemistry-Book and so on is possible. Ontologies builders should decide before building ontologies the level of granularity and what to cut out. Step 2: Check the completeness of the domains and ranges of the functions and relations and that the domains of these functions and relations are defined in the class hierarchy of the ontology being ver- ified. The aim is to figure out whether the domain (or range) of each argument of each function or relation in the ontology exactly and precisely delimits the classes that are appropriate for that argument. Errors appear when the domains and ranges are imprecise, over- specified or completely wrong. For a given subgraph of the class hierarchy of the current ontology, we find errors in the domain and range of its functions and relations when we fill in their tables of domains and ranges. These tables allow us to compare the old and new domains and ranges of the functions and relations in a hierarchy. In them, column 1 gathers the names of all the functions and relations whose domains are in the hierarchy. Columns 2 and 4 represent their original domains and ranges (as they are defined in the ontology you are verifying). Finally, columns 3 and 5 are the new domains and ranges of the functions and relations if they have to be modified (as you think they Document Title-Of Number-Of-Pages-Of - T i Thesis MiscellaneousJPublication Orgamzation-Of k a I I Picture Doctoral-Thesis Computer-Program Master-Thesis FIGURE 1. A classes/subclasses hierarchy and their properties. Book Has-Author Title-Of TABLE 2 Domains and Ranges for the Functions of the Hierarchy in Fig. 1 Definition Title-Of Number-Of-Pages-Of Organization-Of Conference-Of University-Of Original domain New domain Original range Document Document Book Proceeding Document Document Title Natural Organization Thesis University New range Conference TABLE 3 Domains and Ranges for the Relations of the Hierarchy in Fig. 1 Definition Original domain New domain Original range New range Has-Author Document — Author should be defined). Assume the following Ontolingua functions and relations definitions. Tables 2 and 3 summarize the domains and ranges of some functions and relations that have as a domain some classes in the hierarchy of the Fig. 1. Def. 9. (Define-Relation Has-Author (?Doc? Author) "The creator(s) of a document. Not necessarily the author of a work published in the document, but often so. The author is a real agent, not a name of an agent." :def (And (Document?Doc) (Author?Author))) Def. 10. (Define-Function Title-Of (?Doc):—?Title "The title of a document. Not necessarily the title of a work published in the document." :def (And (Document?Doc) (TitleTTitle))) Def. 11. (Define-Function Conference-Of (?Proc):-+ ?Conference 'The conference associated with a proceedings." :def (And (Proceedings ?Proc))) Def. 12. (Define-Function Number-Of-Pages-Of (?Doc):—?N "Number of pages contained in a document. Not the page numbers of an article." :def (And (Document?Doc) (Natural?N))) Def. 13. (Define-Function Organization-Of (?Doc):—^Organization "The institution that publishes a document, like a University or trade association." :def (And (Book?Doc) (Organization?Organization))) Def. 14. (Define-Function University-Of (?Doc):-+?University 'The University that publishes a thesis." :def (And (Document ?Doc) (Organization ?University))) Taking these tables and hierarchy, we can say that: (a) The domain and range of the functions Title-Of (the title of a document) and Number-Of-Pages-Of (num- ber of pages of a document) are well-defined. (b) The domain and range of the relation Has-Author (the author of a document) is well-defined. (c) The domain of the function University-Of (the university of a thesis) is over-specified. (d) The domain of the function Organization-Of (the organization that publishes a document) is impre- cise—any document has an institution that publishes it. So, definitions 11,13 and 14 are modified as follow: Def. 11. (Define-Function Conference-Of (?Proc):—• ?Conference 'The conference associated with a proceedings." :def (And (Proceedings ?Proc) (Organization ^Conference))) Def. 13. (Define-Function Organization-Of (?Doc):—• ?Organization "The institution that publishes a document, like a university or trade association." :def (And (Document ?Doc) (Organization ?Organization))) Def. 14. (Define-Function University-Of (?Doc):—?University 'The university that publishes a thesis." :def (And (Thesis ?Doc) (Organization ?University))) Step 3: Check the completeness of the classes. The goal is to know if the class gathers as much information as possible. So, the class should be defined by a predicate defined by necessary and sufficient conditions, and the set of properties attached to a given class represent the set of properties that the class owns in the real world. In this case, errors appear when: (a) There are missed properties in the definition of a class. We discover missed properties by checking that all the functions and relations that have the class as a domain are included as properties in the definition of the class. For example, going through columns 2 and 3 in Tables 1 and 2, we detect some potential properties (Title-Of, Number-Of-Pages- Of, Has-Author, Organization-Of) of the class DOCUMENT by selecting those functions and relations whose domain is DOCUMENT. Since the class DOCUMENT (see definition 1) only owns the properties Title-Of and Number-Of- Pages-Of, we can say that the definition is incomplete. To make the definition complete, we introduce the missed properties (Organization-Of and Has-Author) in the class DOCUMENT to guarantee that the class as well as its subclasses can have these properties defined. The following Ontolingua sentences should be included in the formal definition of DOCUMENT: (Has-Some ?X Has-Author) (Has-One ?X Institution-Of) (b) There are errors in the cardinality of any property. We detect errors in the cardinality by comparing that the cardinality of the properties in the world modeled formally is that which it is supposed to have in the real world. Check also that the minimum cardinality of a property is inferior to the maximum cardinality. For example, in the class THESIS we constrain the values of the inherited properties Has-Author (a thesis only has one author) when we use the following Onto- lingua sentence: (Has-One ?X Doc.Author) (c) Different classes have the same formal definition. We find equal formal definitions by checking classes that: (1) are classified under the same superclasses, (2) own the same set of properties and (3) the properties have the same cardinality. Checking definitions 4 and 5, we find that there are no semantic differences between the classes MASTER-THESIS and DOCTORAL-THESIS because they do not have any property that differentiates them. We solve the problem by defining a new function Degree-Of in the domain of THESIS and in the range of DEGREE. We differentiate between the two classes by including the Ontolingua sentence (=Degree-Of Ph.D.) in the formal definition of DOCTORAL-THESIS and (=Degree-Of M.S.) in the formal definition of MASTER-THESIS. The definitions of the func- tion Degree-Of and the definition of the class DEGREE are given in Example 3. (Define-Function DEGREE-OF (?Thesis) :—»?Degree "The degree of a thesis work." :def (and (Thesis?Thesis) (Degree ?Degree))) (Define-Class DEGREE (?Degree) "The degree of a study." :axiom-def (Subclass-Partition Degree (SetofB.S. M.S. Ph.D.))) Example 3. Definitions of a new function and class (d) The class does not include properties that it cannot have in the real world. The goal is to find out which properties the class cannot have in the real world and we include this information in the definition of the class. In particular they should be included if they may be inherited from its superclasses. Looking at Fig. 1, we know that a PICTURE is a subclass of the class DOCUMENT, and that all documents can have pages. The Ontolingua sentence (Cannot-Have ?X Number- Of-Pages-Of) — or the equivalent in your language — would forbid the definition of the property Number-Of-Pages-Of in the instances of PICTURE. 2.3.3. Conciseness. Conciseness refers to whether all the information gathered in the ontology is useful and precise. Conciseness does not imply absence of redun- dancies. Sometimes, some degree of controlled redundancy can be useful in definitions. A priori, it is difficult to prognosticate the conciseness of an ontology or set of ontologies because they provide as many abstract definitions as possible for a given domain. An ontology is concise if: (a) It does not store any unnecessary or useless definition. (b) Explicit redundancies do not exist between definitions. For example, if a class is extension- ally-defined by enumerating a set of objects, and these objects are defined as instances in the ontology, the ontology is redundant. Example 4 shows a case of explicit redundancy. (Define-Class MONTH-NAME (?Month) 'The months of the year, specified as an extensionally-defined (i.e. enumerated) set of objects, in English. Instances of this class of months are not symbols, they are months that may be denoted by object constants." :iff-def (Member?Month (setof January february march april may june july august September October november december))) (Define-Instance JANUARY (Month-Name)) Example 4. Explicit redundancy (c) Redundancies cannot be inferred using axioms attached to other definitions. Examples of inferred redundancies are: • A property that can be inherited from a superclass is defined explicitly in any of its subclasses. For example, the sentence (Has-One ?X Title-Of) in the class THESIS, would make it redundant because we can get this property from DOCUMENT by using inheritance. So, it must be removed in BOOK. • A subclass-of relation could be inferred using other definitions. Given the definitions in Exam- ple 5 we could infer from the definition of EXACT-RANGE—the EXACT-RANGE is the class whose instances are exactly those that appear in the last item of any tuple in the relation—that the class AGENT-NAME is a subclass of BIBLIO-NAME. Since AGENT- NAME is the EXACT-RANGE of the AGENT.NAME function, and a range of AGENT.NAME is BIBLIO-NAME, and since the EXACT-RANGE of a binary relation is a subclass-of any of ranges, then it follows that AGENT-NAME is a subclass of BIBLIO- NAME. Consequently, the definition of AGENT.NAME is concise and the inclusion of the constraint in the definition makes it redun- dant. (Define-Class AGENT-NAME (?Name) "A string that is the name of some agent." :def (Biblio-Name?Name) :axiom-def (Exact-Range Agent.Name Agent-Name)) (Define-Function AGENT.NAME (?Agent) :—>?Name "Function from an agent to the name by which it goes." :def (and (Agent?Agent) (Biblio-Name?Name))) (Define-Class BIBLIO-NAME (?String) "A name of something in the bibliographic- data ontology." :def (Biblio-Text?String)) Example 5. An implicit redundancy is inferred (d) A definition is itself redundant. Given the defini- tions in Examples 5 and 6, we can say that the definition of ORGANIZATION.NAME is redun- dant. It is explicitly said in Example 6 that the function ORGANIZATION.NAME is a special- ization of the function AGENT.NAME. If the domain and range of the ORGANIZATION- .NAME function are specializations of the domain and range of the AGENT.NAME function, then all the tuples in the ORGANIZATION.NAME func- tion are specializations of those in AGENT.NAME. We have to delete the sentence (Agent.Name ?Organization ?Name) in ORGANI- ZATION.NAME to make it non-redundant. (Define-Function ORGANIZATION.NAME ( ?Organization) :—» ?Name "The name by which organizations go by. One name per place." :def (and (Organization?Organization) (Biblio-Name?Name) (Agent.Name?Organization?Name))) (Define-Class ORGANIZATION (?X) "An organization is a corporate orsimilar institution, distinguished from persons and other agents." :def (Agent?X)) Example 6. The function ORGANIZATION.NAME is itself redundant 2.3.4. Expandability and Sensitiveness. Expandability refers to the effort required in adding new definitions to an ontology, as well as the effort needed to add new information to a definition, without altering the set of well-defined properties that are already guaranteed after the ontologies verification process. 2.3.5. Sensitiveness. Sensitiveness relates to how small changes in a given definition alter the set of well-defined properties that are already guaranteed. After including or modifying a definition, this criterion must guarantee that: (1) The architecture of the ontology and the architecture of its definitions are still sound. (2) The definitions are lexically and syntactically cor- rect. (3) The ontology and its definitions of conciseness, consistency and completeness are tightly connected. 3. VERIFICATION OF SOFTWARE Software verification refers to building the software right, which means that the software that builds, reuses and shares definitions and axioms correctly and com- pletely implements its requirements. Software engineering methodologies, techniques and tools provide the appropriate framework to verify KST software in each stage and between stages of its life cycle. 4. VERIFICATION OF DOCUMENTATION Documentation verification refers to building the docu- ments correctly. It seeks to guarantee that all the required documents have been written, that nothing has been overlooked in any document and that the documents evolve in step with definitions and software environ- ments in each phase and between phases of the life cycle. Verification of the documentation includes: the natural language string in each definition, general information about the ontology, basic ontological commitments, a summary of definitions, cases studies, definitions taken from other ontologies and also documentation about the software that the environment provides, installation manual, reference manual, release notes, frequently asked questions and tutorials. Special attention is required if WWW documents are indexed automatically using a program. In this case, mistakes in the indexes of the natural language doc- umentation appear easily due to the creative and flexible use of the language. From the information retrieval point of view, four categories of words can be found in the indexed text. (a) Correctly indexed words represent words in the free text documentation that are properly indexed with a word in the ontology vocabulary. (b) Correctly non-indexed words represent words in the free text documentation that are not indexed with a word in the ontology. (c) Incorrectly indexed words include words that have been wrongly indexed with the ontology vocabu- lary. Errors in the indexes are classified in the following categories: • An index semantic blunder arises in natural language documentation when the meaning of the word in the documentation string is not the same as the meaning of the term pointed in the ontology vocabulary. • A context error appears when there are no semantic errors in the pointer, but the word in the documentation string is not used in the ontology theory context. • Miscellaneous mistakes cover loops in indexes and problems in polymorphical definitions. While the former deal with indexes from words in the natural language documentation of a definition to the definition itself, polymorphical errors deal with several and different definitions of the same word in different ontologies. Multiple definitions create ambiguity in the selection of the indexes. (d) Incorrectly non-indexed words concern words used in the free text documentation in the ontology theory context that are not indexed with the ontology vocabulary because it is spelt differently. A study performed on Ontolingua ontologies reveals that the majority of the errors can be easily avoided if the ontology writer writes the words to be indexed using certain conventions (i.e. using uppercase for all the words, and/or using hyphenated strings of words). Assuming that the natural language documentation has been written following these conventions, the following heuristics will provide new semantic, context and morphological capabilities in the program that automat- ically generates the indexes: (1) Pluralization of hyphenated and non-hyphenated words in the lexicon. (2) Detection of situations in which a word is followed by unusual punctuation marks. (3) Automatic generation of hyphenated words. (4) Prevention of pointers to words that are out of the scope of the current ontology and ontologies that are included in the current ontologies. (5) If a polymorphic word and a name of an ontology appear together in a sentence, the polymorphical word should point to the definition in that ontology. (6) If a polymorphic definition is made in an ontology, any index of the word in the ontology should point to its definition, unless the name of any other ontology appears in the sentence. (7) Given a word, any index to that word from its natural language documentation must be prevented. 5. CONCLUSIONS Based on the empirical verification of ontolingua ontolo- gies, a novel approach to verify KST has been illustrated. The main contributions are: (a) We create a framework to verify KST. This framework includes terminology, definitions, cri- teria and examples to carry out the verification. (b) We split the verification process in three proc- esses: verification of ontologies, verification of software for building, reusing and sharing defini- tions and verification of documentation. The most important is verification of the ontologies. Soft- ware engineering provides the framework to verify KST software and documentation. (c) Verification of the ontologies includes verifying that: the architecture of the ontologies and defini- tions are sound, the lexicon and syntax are correct and the content of the definitions is consistent, complete, concise, expandable and sensitive. Regarding the content, we provide: • A formal definition of internal, metaphysical and inferred consistency, and examples that show how to deal with these new concepts. • An informal definition of completeness, and stereotype of errors that make relations, func- tions and classes incomplete. • An informal definition of conciseness and kinds of errors that make ontologies redundant. • We define expandability and sensitiveness of an ontology, and we identify which kind of ver- ification has to be performed when definitions are added or modified in an ontology. Finally, we remark that conciseness, consistency and completeness are tightly connected. An ontology can be complete and not be concise if the formal sentence written in a formal definition can be inferred using other definitions. However, if the sentences are not explicitly written and they cannot be inferred, the ontology could be concise or not, but it is not complete. Actually, we are developing a tool called ONE-T (ONtologies Evaluation Tool) for Ontolingua ontologies. The tool detects mistakes and omisions in ontologies and corrects them automatically whenever it is possible. So, it increases the performance and quality of the evaluation process. Acknowledgements—The research work has been performed at the Knowledge Systems Laboratory in Stanford University, Stanford, CA, U.S.A. It has been sponsored by grant number PF94-8821929 of the Ministerio de Educación y Ciencia in Spain. Thanks to Richard Fikes for his comments on the paper and advice during the work and to Mike Uschold for his comments. Genesereth, M. R. & Fikes, R. E. (1993). Knowledge interchange format. Version 3.0. Reference manual. Report Logic-92-1, Com- puter Science Department, Stanford University. Gómez-Pérez, A. (1994) From knowledge based systems to knowledge sharing technology: Evaluation and assessment. Technical Report KSL 94-73, Knowledge Systems Laboratory, Stanford University. Gómez-Pérez, A. (1995). Some ideas and examples to evaluate ontologies. In Proceedings of the Eleventh IEEE JConfereéce on Artificial Intelligence for Applications, pp. 299-305^ New York: IEEE Computer Society Press. Gómez-Pérez, A. (1996). Guidelines to verify compfeieness and consistency in ontologies. The Third World Congress hn Expert Systems. Gómez-Pérez, A., Juristo, N. & Pazos, J. (1995). Evaluation and assessment of knowledge sharing technology. In Towards very large knowledge bases: Knowledge building and knowledge sharing, pp. 289-296. IOS Press. Gruber, T. (1993a). A translation approach to portable ontology specifications. Knowledge Acquisition, 5, 199-220. Gruber, T. (1993b). Toward principles for the design of ontologies used for knowledge sharing. Technical Report KSL 93-04, Knowledge Systems Laboratory, Stanford University. Gruber, T. (1994). Bibliographic-Data ontology. Available at hpp. stanford.edu, as http://www-ksl.stanford.edu/knowledge-sharing /ontologies/html/bibliographic-data/index.html Gruninger, M. & Fox, M. S. (1994). The role of competency questions in enterprise engineering. IFIP WG5.7 workshop on benchmarking, theory and practice. Trondheim, Norway. Guarino, N. & Giaretta, P. (1995). Ontologies and knowledge bases: Towards a terminological clarification. In Towards very large knowledge bases: Knowledge building and knowledge sharing, pp. 25-32. IOS Press. http://stanford.edu http://www-ksl.stanford.edu/knowledge-sharing 
work_jcxx2b72dzgzdlmln4dydjqxly	----	Context Aware Environments : from Specification to Implementation • Authors : Patrick Reignier, Oliver Brdiczka, Dominique Vaufreydaz, James L Crowley, Jerôme Maisonnasse • Keywords : Context recognition, Petri nets, hidden Markov models, automatic cameraman, interaction group detection • Authors Address : Inria Rhône-Alpes, 655 avenue de l’Europe, 38334 Saint Ismier Cedex, France • Contact : Patrick Reignier • Contact’s email : Patrick.Reignier@inrialpes.fr • Contact’s phone number : +33 476 615 411 • Contact’s fax number : +33 476 615 210 Context Aware Environments : from Specification to Implementation Patrick Reignier, Oliver Brdiczka, Dominique Vaufreydaz James L Crowley, Jerôme Maisonnasse Inria Rhône-Alpes, 655 avenue de l’Europe, 38334 Saint Ismier Cedex, France {reignier,brdiczka,vaufreydaz,crowley,maisonnasse}@inrialpes.fr Abstract This article deals with the problem of implementing a context model for a smart environ- ment. This problem has already been addressed several times using many different data or problem driven methods. In order to separate modeling phase from implementation, we first represent the context model by a network of situations. Then, different implementations can be automatically generated from this context model depending on user needs and underlying perceptual components. Two different implementations are proposed in this paper: a deter- ministic one based on Petri nets and a probabilistic one based on hidden Markov models. Both implementations are illustrated and applied to real world problems. Keywords : Context recognition, Petri nets, hidden Markov models, automatic cameraman, interaction group detection 1 Introduction Pervasive and ubiquitous computing [Weiser(1991)] seek to integrate computation into everyday environments. People are enabled to move around and interact with computers more and more naturally. One of the goals of these computerized spaces is to enable devices to sense changes in the environment and to automatically adapt and act based on these changes. A main focus is laid on sensing and responding to human activity. Human activity does not strictly follow plans but is very situation dependent [Suchman(1987)]. Computerized spaces and their devices need hence to use this situational information, i.e. con- 1 text [Dey(2001)] , to respond correctly to human activity. In order to become context-aware, computer systems must maintain a model describing the environment, its occupants and their activities. Unfortunately, there is no generic algorithm capable of robustly recognizing situations from perceptual events coming from sensors. Various approaches, based on logic, Bayesian rule, fuzzy logic etc., have been explored and evaluated. However, their performance is very ”problem1 and environment dependent”. In order to be able to use several approaches inside the same application, we have decided to clearly separate the specification of context (scenario) and the implementation of the program that recognizes it. The transformation between a specification and its implementation must be as automatic as possible. In this article, we first review basic concepts for modelling context. The central notion of situ- ation [Dey(2001)] is defined and context is represented by a network of situations [Crowley(2002)]. Situations within this network are in temporal relationships (described by Allen’s temporal opera- tors [Allen(1983)]). This is the formal specification of the context. We then propose two different approaches for implementing it : a deterministic implementation transforming the situation net- work into a synchronized Petri net and a probabilistic implementation transforming the situation network into a hidden Markov model. Both implementations have been applied to real world prob- lems. An automatic cameraman system (section 3.3) has been implemented using synchronized Petri nets and a detector for interaction group configurations (section 4.3) has been implemented using hidden Markov models. Both example implementations have been tested with success. 2 Representing Context by Situation Models The notion of context is not new and has been explored in different areas like linguistics, natural language processing and knowledge representation. Dey defines context as ”any information that can be used to characterize the situation of an entity” [Dey(2001)]. An entity can be a person, place or object considered relevant to user and application. Context-aware applications need this contextual information to deliver the correct service to the correct user, at the correct place and time, and in the correct format for the environment [Weiser(1991)]. The structure and represen- tation of this information must be determined before being exploited by a specific application. 1The problem can be to recognize a situation in a well defined scenario (e.g. ”presentation” or ”brainstorming” in a meeting scenario) or to recognize the scenario within a set of possible scenarios (e.g. is it a meeting, party or candlelight dinner?) 2 Context and activity are separable. The context describes features of the environment within which the activity takes place [Dourish(2004)]. Loke states that situation and activity are, however, not interchangeable, and activity can be considered as a type of contextual information which can be used to characterize a situation [Loke(2005)]. Following Dey’s context definition, situation is a central notion describing context. We consider context to be a network of situations [Crowley(2002)]. Dey defines situation as ”description of the states of relevant entities” [Dey(2001)]. Situation is thus a temporal state within context. Allen’s temporal operators [Allen(1983)] can be used to describe relationships between situations (Fig. 1). Figure 1: Context of a ”presentation with questions”. The conceptual events characterizing the situations are not detailed. There are different concepts used to characterize a situation. As described above, activity is such a concept. Activity models like in [Muehlenbrock(2004)] identify the different states of entities. We extend this concept to roles and relations between entities [Crowley(2002)]. An entity is observed to play a role if it passes a role acceptance test on its properties. A relation is defined as a predicate function on several entities playing roles. Changes in activity, role or relation are signaled by events. Behavior within the environment can be described by a script. A script corresponds to a sequence of situations in the situation network reflecting (human) behavior in the environment. Scripts are not necessarily linear. Situations and the underlying concepts (activity, roles or relations) are the context meta-model. This meta-model can be used by the software developer to specify the situations and the associated reactions (the model). This model is only a specification. It must be transformed into a working system listening to perception events, interpreting them and verifying that this flow of events is 3 Figure 2: The ”leaving an object” context specification compatible with the model specification. Let us consider for instance a “leaving object” context. An alarm must be triggered when someone enters a room, leaves an object and run away. It can be specified by the model figure 2. The perception events can be provided by a video tracker. The Entering a room event is triggered when a moving target goes through a detection zone in front of the door. The Leaving object event is detected by a target split. The run event is triggered by a moving target with a speed higher than a specified threshold. Based on this context specification, the next step must be to write a program checking that the run event is coming after the leaves an object event. Unfortunately, there is no generic algorithm capable of robustly recognizing situations from perceptual events coming from sensors. Various approaches have been proposed but they are very problem dependant. We argue that instead of trying to find a “universal solution”, we must have an automatic way to transform a context specification into a set of context recognition modules based on various approaches. It is then easier to try several approaches on a particular application. The developer can also modify the context specification without having to recode the various implementations. In the following sections, we present a deterministic and a probabilistic implementation of situation models. The deterministic implementation is based on Petri nets, while the probabilistic 4 implementation is based on hidden Markov models. 3 Deterministic Implementation: Petri Nets We begin this section with an informal review of Petri nets. We then describe how Petri nets are used to implement a situation network representing a context model and how to evaluate a script. 3.1 Petri Nets A Petri net is a graphical mathematical tool used to model dynamical systems with discrete inputs, outputs and states. Such models have first been defined by C. M. Petri [Petri(1962)]. A Petri net is an oriented graph, composed of arcs and two categories of nodes (places and transitions). Places are the state variables of the system containing zero or a positive number of marks. Transitions model the system evolution and are connected from places to places. A transition is valid if all the ”from” places have at least one mark. When a valid transition is fired: one mark is removed from every ”from” place and one mark is added to every ”to” place. Only one transition can be fired at a time. A more formal definition of Petri nets is given in [Murata(1989)]. Finite-state machines like situation models can be equivalently represented by a subclass of Petri nets [Murata(1989)]. Several extensions of Petri nets have been proposed. One of them is the synchronized Petri net. A synchronized Petri net is a Petri net with events associated to each transition. A transition can now be fired if it is valid and the corresponding event has been triggered (Fig. 3). 3.2 Implementation and Script Evaluation using Petri Nets A context model is defined by a network of situations. The connections between the situations are temporal constraints based on Allen’s temporal operators [Allen(1983)]. Events indicate state changes of the concepts (activity, roles, or relations) describing situations. A situation change is triggered by these events. For instance, consider two situations S1 and S2 such that S2 meets S1 (Fig. 4). To trigger a change from the current situation S1 to situation S2, we need to observe at least two events: 1. a first event invalidating the situation S1 (¬V alidS1), followed by 2. a second event validating situation S2 (V alidS2) . 5 Figure 3: Synchronized Petri net with places S1, S2, S3 and transition T1 triggered by event E. Figure 4: A small context : S2 meets S1 6 This example can be directly represented as a synchronized Petri net (Fig. 5). We associate a situation to every place. The situation is active if there is at least one mark in the corresponding place: we are currently in situation S1. Figure 5: Synchronized Petri net implementation of the context In a ”standard” Petri net, the transition between S1 and S2 is automatically fired as soon as S1 is active. We need to control this transition based on the perceptual events coming from the environment. This transition control is managed using the transition event of the synchronized Petri net. We generate this transition event T only when Situation S1 is no more valid and Situation S2 becomes valid (as seen in the previous paragraph): T = ¬ValidS1 ∧ ValidS2 More generally, the event function T of a transition is the conjunction of a function associated to the place before the transition and a function associated to the place after the transition. We will call those two functions respectively P reT and P ostT . We have: T = P reT ∧ P ostT P reT corresponds to what we are currently observing in the environment. P ostT indicates what the system should expect to see next based on the contextual model. This short example showed how to transform the ”meet” operator into a corresponding synchro- nized Petri net. This transformation can be done for all the other Allen operators as summarized in Table 1. Note that Petri nets are very well adapted for implementing situation models containing par- allelism. As shown in Table 1, all Allen temporal operators can be implemented by Petri nets. However, Petri nets are not suitable for applications with erroneous perceptions or uncertain 7 Petri net Transition function { T 1 = ¬V alidS1 T 2 = V alidS2    P ostT 1 = V alidS1 T 2 = V alidS2 T 3 = ¬V alidS1 P reT 4 = ¬V alidS2 8 < : P ostT 1 = V alidS1 ∧ V alidS2 P reT 2 = ¬V alidS1 P reT 3 = ¬V alidS2  P ostT 1 = V alidS1 ∧ V alidS2 P reT 2 = ¬V alidS1 ∧ ¬V alidS2 8 >>< >>: P ostT 1 = V alidS2 T 2 = V alidS1 T 3 = ¬V alidS1 P reT 4 = ¬V alidS2    P ostT 1 = V alidS1 P ostT 2 = V alidS2 P reT 3 = ¬V alidS1 ∧ ¬V alidS2 Table 1: Synchronized Petri nets for the Allen operators: the ”Unspec” place stands for an unspecified situation, which means that we do not model what might happen. 8 perception expectations. 3.3 Example : Automatic Cameraman The automatic cameraman [FAME(2004)] is an audio-video recording system that automatically records a lecture. The lecture room is equipped with multiple cameras and microphones. The system is context-aware selecting at every time, based on the current situation, the appropriate camera to provide images (Fig. 6). Figure 6: Different camera images recorded by the automatic cameraman system. The available actions are Start recording, Filming the whole room, Filming the lecturer, Film- ing the audience and Filming the slides. The situation list is: • Init → start recording and filming the whole room. • The lecturer speaks → filming the lecturer. • Someone in the audience asks a question → filming the audience. • There is a new slide → filming the new slide. • Some one enters the room → filming the whole room. 9 The lecture is an alternation of ”lecturer speaking” and ”audience asking a question”. ”New slide” and ”Someone entering the room” can happen in parallel. The situation network is given in Fig. 7, the corresponding Petri net is given in Fig. 8. Figure 7: Situation network of the automatic cameraman system. Code generation is done by automatically transforming the synchronized Petri net into a cor- responding program in JESS [Jess]. JESS is an expert system programming environment (facts database with forward chaining rules). The input of the generated program are facts based on events describing state changes of the concepts (roles and relations here)2. The output are the current situation(s) and the associated action(s). We are going to illustrate this code generation process on a small subset of the previous Petri net (Fig. 9). The generated rules can be decomposed into three groups: 1. The Petri algorithm. 2. The events calculation (the synchronization). 3. The action generation. 2Activity, roles, and relations are reconstructed from data provided by perceptual components (video tracker, speech activity detectors, etc.) 10 Figure 8: Petri net of the automatic cameraman system. Figure 9: The lecturer speaks and then the audience asks a question 11 Petri algorithm: These rules implement the calculation of the marks: if there is at least one mark in place S1 and the associated event is triggered, we remove a mark in S1 and add a new one in S2. Events calculation: The transition event describes that S1 is no longer valid and that S2 has become true. A situation is described here by a set of roles and relations. A situation becomes true if its roles and relations are validated. A situation becomes invalid when there is a change in the assignment of entities to roles or a change in the relations between those entities. As explained in subsection 3.2, the transition event describes that S1 is no longer valid and that S2 has become true. A situation is described by a set of roles and relations. A situation becomes true if its roles and relations are validated. A situation becomes invalid when there is a change in the assignment of entities to roles or a change in the relations between those entities. • Situation S1: – Roles : lecturer, speaker. – Relations : lecturer is same as speaker. • Situation S2: – Roles : audience, speaker. – Relations : audience is same as speaker. 12 Action generation: These rules associate (if needed) an action to every situation. This approach recorded a four day seminar on ”Language Technology” and ”Language, Cog- nition, and Evolution” [Seminar(2004)], which was held on the premises of the FORUM2004 [FORUM(2004)] in Barcelona. The automatic cameraman has also been used to record and broadcast real lectures in the amphitheatre at INRIA Rhône-Alpes (example images can be seen in Fig. 6). 3.4 Experimental results Table 2 shows the confusion matrix for two hours of recording. A ground truth dataset has been produced by manually annotating the situations on the video. The three situations are : • Slides : a new slide is projected on the screen 13 • Speaker : the speaker is talking or answering a question • Audience : someone in the audience is asking a question The lines of the matrix contain the detection results, while the columns contain the expected response. Slides Speaker Audience Slides 0.96 0.04 0.0 Speaker 0.0 0.89 0.11 Audience 0.37 0.05 0.58 Table 2: Confusion matrix We have obtained a recognition rate of 93.7%. As the context determination is deterministic, this confusion matrix mainly shows the robustness of the underlying perception algorithms (the speech activity and the ”new slide” detector). In the last line of the matrix, we can see that there is a big confusion between ”audience asking a question” and ”new slide”. Indeed, when someone starts answering a question, at the same time, he is often seeking for slides in his presentation. This example illustrates that in some case, deterministic approach can be advantageously replaced by more fuzzy ones. We have also made an informal survey after broadcasting the lectures in the amphitheatre at INRIA. Remote spectators could access two streams : a first stream provided by a fix camera and a stream provided by our automatic cameraman. They all prefered the automatic cameraman stream, allowing them to better understand the presentation. 4 Probabilistic Implementation: Hidden Markov Models A probabilistic implementation of the situation model integrates uncertainty values into the model. These uncertainty values can both refer to confidence values for events and to a less rigid repre- sentation of situation and situation transitions. The situation model is a finite-state machine. The natural choice for a probabilistic imple- mentation is then a probabilistic finite-state machine [Vidal(2005)a]. A probabilistic finite-state machine is a probabilistic automaton (PFA) defined over a finite alphabet ∑ . A language is a subset of ∑∗. A PFA defines a stochastic language, which is a probability distribution over ∑∗. The distribution must verify: ∑ x∈P∗ P robability(x) = 1 14 A formal definition of PFA can be found in [Vidal(2005)a]. 4.1 Hidden Markov Models A hidden Markov model (HMM) [Rabiner(1989)] is a stochastic process where the evolution is managed by states. The series of states constitute a Markov chain which is not directly observable. Such a chain is said to be ”hidden”. Each state of the model generates an observation. Only the observations are visible. A more formal definition of an HMM is given in [Rabiner(1989)]. The two following propositions hold [Vidal(2005)b]: Proposition 1: Given a PFA A with m transitions and P robability(epsilon) = 0, there exists an HMM M with at most m states such that the stochastic language DM of M is equal to the stochastic language DA of A. Proposition 2: Given an HMM M with n states, there exists a PFA A with at most n states such that the stochastic language DA of A is equal to the stochastic language DM of M . As we are only interested in PFA without epsilon transitions, i.e. PFA the transitions of which are triggered by events, language-equivalent HMMs can be used to implement PFA. 4.2 Implementation and Script Evaluation using Hidden Markov Mod- els The situations of a context model can be implemented by the states of a HMM. Events indicating state changes of the concepts (activity, roles, or relations) generate the observations for the HMM. A state (situation) is characterized by a particular probability distribution of these observations. The activation of a new situation (state) is thus determined by the transition probability from the current state to this new state as well as by the probability of the given observations in this new state. The connections in the situation network are represented by non-zero transition probability values. The observation probability distributions for the situations as well as the transition probabilities between the situations need to be specified (or learned) when implementing a situation model. We are interested in three basic problems: 1. Given a sequence of observations (based on events) and a situation model implemented by a HMM, how to choose the corresponding state sequence (situation sequence)? This includes the determination of the (most likely) current situation and the determination of likely following situations. 15 2. Given a sequence of observations (based on events) and a situation model implemented by a HMM, how to compute the probability of the observation sequence, given the model? This corresponds to the likelihood of the situation model (based on the given events). 3. How to adjust the HMM model parameters? This corresponds to adjusting probability distributions of situations based on given event data. [Rabiner(1989)] gives several solutions to these problems. The Viterbi algorithm is used to determine the most probable state sequence, given a HMM and an observation sequence (Problem 1). The probability of a HMM, given an observation sequence, can be computed using the Forward- Backward algorithm (Problem 2). The expectation-maximization (EM) Baum-Welch algorithm adjusts the HMM model parameters, given observation sequences (Problem 3). Note that the HMM implementation of a situation model is particularly suitable for applications that deal with erroneous perceptions as well as situations that are characterized by a particular frequency of events. A HMM implementation is, however, less suitable for representing parallelism (not all Allen operators can thus be represented by a classical HMM). 4.3 Example : Detection of Interaction Groups This example addresses the problem of detecting changing interaction group configurations in a smart environment. During a meeting, participants can form one big group working on the same task, or they can split into subgroups doing independent tasks in parallel. Our objective is to determine the current small group configuration, i.e. who is interacting with whom and thus which interaction groups are formed. As we focus on verbal interaction, one group has a minimum size of two individuals (assuming that isolated individuals do not speak). The speech of each meeting participant is recorded using a lapel microphone. An automatic speech detector [FAME(2004)] parses this multi-channel audio input and detects which participant stops and starts speaking. We admit the use of lapel microphones in order to minimize correlation errors of speech activity of different participants, i.e. speech of participant A is detected as speech of participant B. The proposed approach is based on a HMM implementation of the context model. The obser- vations of the HMM are a discretization of speech activity events sent by the automatic speech detector. Fig. 10 shows the situation network for a meeting with 4 participants. Each possible interaction group configuration is represented by one situation. These situations are transformed to the states of a HMM (Fig. 11). 16 Figure 10: Situation model for a meeting of 4 participants A, B, C, D. Figure 11: States of the HMM implementation of the situation model for a meeting of 4 participants A, B, C, D. 17 The probability distributions of the different states are specified based on conversational hy- potheses. These conversational hypotheses assume that speech within an interaction group is more regulated than speech between distinct interaction groups. The transition probabilities between the states are set to a very low level in order to stabilize the detection of state changes assuming hence that group changes occur in reasonable delays. To detect different group configurations, we apply the Viterbi algorithm (solution to Problem 1 in section 4.2) to the flow of arriving observations. Figure 12: Example of a configuration of 2 groups of 2 participants 4.4 Experimental results To evaluate, we recorded three small group meetings with 4 participants (Fig. 12). Using the HMM detector, we obtained a total recognition rate for the small group configurations of 84.8 % [Brdiczka(2005)]. Table 3 shows the confusion matrix for the 3 experiments. This matrix indicates for each group configuration the correct and wrong detections. The lines of the matrix contain the detection results, while the columns contain the expected response. The figure 13 gives an overview of the detection of group configurations over time. The lines of the chart correspond to different group configurations. The continuous line indicates the correct group configuration expected as detection result. The results are encouraging and tend to validate the conversational hypotheses to distinguish 18 (ABCD) (AB)(CD) (AC)(BD) (AD)(CB) (ABCD) 0.88 0.03 0.06 0.03 (AB)(CD) 0.08 0.87 0.05 0.00 (AC)(BD) 0.22 0.01 0.77 0.00 (AD)(CB) 0.04 0.04 0.08 0.84 Table 3: Confusion matrix Figure 13: The upper left figure corresponds to the group configuration detection for the exper- iment 1 (duration: 9 min. 22 sec.). The upper right figure corresponds to the experiment 2 (duration: 15 min. 16 sec.). The third figure corresponds to the experiment 3 (duration: 16min. 19sec.) 0 on the axis corresponds to the group (ABCD), 1 is for the two groups (AB) and (CD), 2 for (AC)(BC) and 3 for (AD)(BC). interaction groups. The Viterbi algorithm executed on long observation sequences is quite robust to wrong detections of the speech activity detector. However, a minimum number of correct speech activity detections is necessary, as the method relies on the information of who speaks at which moment. The use of lapel microphones made it possible to limit wrong detections as these microphones are attached to a particular person (and thus should only detect his/her speech). 5 Conclusion This article proposed two different implementations for the situation model representing context: a deterministic one based on Petri nets and a probabilistic one based on hidden Markov models. Both approaches have been applied to real world problems with success: an automatic camera- man system (Petri nets) and an interaction group detector (HMMs) have been implemented. 19 Each implementation is well adapted for particular applications. Petri nets can implement all Allen temporal operators, in particular those describing parallelism. However, Petri nets can not model erroneous perceptions and uncertain situations (uncertain expectations of perceptions for a situation and uncertain situation transitions). Hidden Markov models are less rich in modelling temporal constraints (in particular parallelism), but HMMs permit to model erroneous input and uncertain situations. Thus the choice of the implementation depends on the application that is envisaged. References [Allen(1983)] J., Maintaining Knowledge about Temporal Intervals, Communications of the ACM, 26 (11):832-843, 1983. [Brdiczka(2005)] Brdiczka, O., Maisonnasse, J., and Reignier, P., Automatic detection of inter- action groups. Proceedings of the International Conference on Multimodal Interfaces, Trento, Italy, October 2005. [Crowley(2002)] Crowley, J.L., Coutaz, J., Rey, G. and Reignier, P., Perceptual Components for Context Aware Computing, Proceedings of the International Conference on Ubiquitous Computing, Göteborg 2002. [Dey(2001)] Dey, A.K., Understanding and Using Context, Personal and Ubiquitous Computing 5:4-7, 2001. [Dourish(2004)] Dourish, P., What we talk about when we talk about context, Personal and Ubiquitous Computing 8:19- 30, 2004. [FAME(2004)] FAME: Facilitating Agent for Multi-Cultural Exchange (WP4), European Com- mission project IST- 2000-28323 October 2001. [FORUM(2004)] FORUM2004: Universal Forum of Cultures, Barcelona, July 2004, http://www.barcelona2004.org. [Jess] Jess: the rule engine for java, http://herzberg.ca.sandia.gov/jess/. [Loke(2005)] Loke, S.W., Representing and reasoning with situations for context-aware pervasive computing: a logic programming perspective, The Knowledge Engineering Review, 19(3):213- 233, 2005. 20 [Muehlenbrock(2004)] Muehlenbrock, M.; Brdiczka, O.; Snowdon, D., Meunier, J., Learning to Detect User Activity and Availability from a Variety of Sensor Data, Proc. of PerCom, 2004. [Murata(1989)] Murata, T., Petri Nets: Properties, Analysis and Applications, Proc. of IEEE 77(4):541-580, 1989. [Petri(1962)] Petri, C. A., Kommunikation mit Automaten, Ph.D. thesis, Institut fuer Instru- mentelle Mathematik, Bonn, 1962. [Rabiner(1989)] Rabiner, L., A tutorial on Hidden Markov Models and selected applications in speech recognition, Proc. of IEEE 77(2):257-286, 1989. [Seminar(2004)] Consorci Universitat Internacional Menéndez Pelayo de Barcelona, ”Tecnologies de la llengua: darrers avenços” and ”Llenguatge, cognició i evolució”, Barcelona, July 2004, http://www.cuimpb.es. [Suchman(1987)] Suchman, L., Plans and Situated Actions: The Problem of Human-Machine Communication, Cambridge University Press, 1987. [Vidal(2005)a] Vidal, E., Thollard, F. de la Higuera, C. Casacuberta, and Carrasco, R. C., Prob- abilistic Finite- State Machines Part I. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005. [Vidal(2005)b] Vidal, E., Thollard, F., de la Higuera, C., Casacuberta, and Carrasco, R. C., Prob- abilistic Finite- State Machines Part II. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005. [Weiser(1991)] Weiser, M., The Computer for the Twenty-First Century, Scientific American Pub- lishers, 1991. 21 
work_jd2wgu7rtje5fdfbbvwb4plsem	----	Online joint palmprint and palmvein verification 1 2 3 4 5 7 89 10 11 12 13 14 15 16 17 1 8 Q1 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 Expert Systems with Applications xxx (2010) xxx–xxx ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Online joint palmprint and palmvein verification David Zhang ⇑, Zhenhua Guo, Guangming Lu, Lei Zhang, Yahui Liu, Wangmeng Zuo Biometrics Research Centre, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong a r t i c l e i n f o 19 20 21 22 23 24 25 26 Keywords: Biometric authentication Palmprint Palmvein Fusion Liveness detection Texture coding Matched filter 27 28 29 30 31 32 33 0957-4174/$ - see front matter � 2010 Published by doi:10.1016/j.eswa.2010.08.052 ⇑ Corresponding author. E-mail address: csdzhang@comp.polyu.edu.hk (D. Please cite this article in press as: Zhang, D., et j.eswa.2010.08.052 a b s t r a c t As a unique and reliable biometric characteristic, palmprint verification has achieved a great success. However, palmprint alone may not be able to meet the increasing demand of highly accurate and robust biometric systems. Recently, palmvein, which refers to the palm feature under near-infrared spectrum, has been attracting much research interest. Since palmprint and palmvein can be captured simulta- neously by using specially designed devices, the joint use of palmprint and palmvein features can effec- tively increase the accuracy, robustness and anti-spoof capability of palm based biometric techniques. This paper presents an online personal verification system by fusing palmprint and palmvein informa- tion. Considering that the palmvein image quality can vary much, a dynamic fusion scheme which is adaptive to image quality is developed. To increase the anti-spoof capability of the system, a liveness detection method based on the image property is proposed. A comprehensive database of palmprint– palmvein images was established to verify the proposed system, and the experimental results demon- strated that since palmprint and palmvein contain complementary information, much higher accuracy could be achieved by fusing them than using only one of them. In addition, the whole verification proce- dure can be completed in 1.2 s, which implies that the system can work in real time. � 2010 Published by Elsevier Ltd. 34 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 1. Introduction Biometric characteristics, including fingerprint, facial features, iris, voice, signature, and palmprint, etc. (Jain, Bolle, & Pankanti, 1999) are now widely used in security applications. Each biometric characteristic has its own advantages and limitations. Among var- ious biometric techniques, palmprint recognition is getting popular in personal authentication because it provides robust features from a large palm area and the palmprint image can be captured with a cost-effective device. In general, a typical palmprint acquisition de- vice operates under visible light and can acquire three kinds of fea- tures: principal lines (usually the three dominant lines on the palm), wrinkles (weaker and more irregular lines) and ridges (pat- terns of raised skin similar to fingerprint patterns). A resolution of about 100 dpi (dots per inch) (Han, Cheng, Lin, & Fan, 2003; Zhang, Kong, You, & Wong, 2003) can be used to acquire principal lines and wrinkles while a higher resolution, usually 500 dpi, is required to acquire ridge features. However, such a high resolution will in- crease significantly the computational cost to extract ridge features because of the large image size of palm, and hence prevents the system from being implemented in real time. Therefore, most of the palmprint base systems capture low resolution palmprint images using CCD (charge-coupled device) cameras and many algorithms have been proposed for feature extraction and match- 83 84 85 Elsevier Ltd. Zhang). al. Online joint palmprint and p ing (Connie, Jin, Ong, & Ling, 2005; Han et al., 2003; Hennings-Yeo- mans, Kumar, & Savvides, 2007; Hu, Feng, & Zhou, 2007; Kong & Zhang, 2004; Kumar & Zhang, 2005; Ribaric & Fratric, 2005; Su, 2009a, 2009b; Wu, Zhang, & Wang, 2003; Zhang et al., 2003). Although palmprint recognition has achieved a great success, it has some intrinsic weaknesses. For example, some people may have similar palm lines, especially principal lines (Zhang et al., 2003); also it is not so difficult to forge a fake palmprint (Kong, Zhang, & Kamel, 2009). These problems can be addressed by using multi-biometric systems, such as fusing facial trait and palmprint trait (Yao, Jing, & Wong, 2007) or fusing iris and palmprint traits (Wu, Zhang, Wang, & Qi, 2007). However, such systems are clumsy as they involve two separate sensors to sense two traits. One way to improve the discriminativeness and anti-spoofing capability of palmprint systems is to use more features from the palm, such as the veins of the palm. The veins of the palm mainly refer to the inner vessel structures beneath the skin and the palm- vein images can be collected using both far infrared (FIR) and near- infrared (NIR) light (Zharov et al., 2004). Obviously, palmvein is much harder to fake than palmprint. There is a long history of using NIR and FIR to collect vein biometrics (Cross & Smith, 1995; Kono, Ueki, & Umemura, 2002; Lin & Fan, 2005; Macgregor & Welfold, 1991; Socolinsky, Wolff, Neuheisel, & Eveland, 2001; Wang, Leedham, & Cho, 2008; Wu & Ye, 2009), while recently palmvein system has also been proposed (Watanabe, Endoh, Shio- hara, & Sasaki, 2005). Intuitively, since both palmprint and palm- vein are from the palm, it is possible to establish a convenient almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://dx.doi.org/10.1016/j.eswa.2010.08.052 mailto:csdzhang@comp.polyu.edu.hk http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 Fig. 1. The prototype system. 2 D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 multi-biometric system to acquire and use the two traits jointly, and the complementary information provided by the two traits will make the system more accurate in personal identification and more robust to spoof-attack. A number of studies have been conducted to fusing palmprint and palmvein information for personal recognition. Wang, Yau, Suwandy, and Sung (2008) developed such a system and obtained good results but their system uses two separate cameras and re- quires a time-consuming registration procedure, which makes it difficult to use in real-time. Hao, Sun, and Tan (2007) evaluated various image level fusion schemes of palmprint and palmvein images. However, they evaluated the method using a very small database (only 84 images), making it hard to draw strong conclu- sions. Toh et al. (2006) captured the palm image under IR (infrared) lighting and then extracted the palmvein and palmprint features separately; however, since there was only one IR light source, some valuable palmprint information was lost. In this paper, we design and construct a new device that can ac- quire palmprint and palmvein images simultaneously and in real- time. The device involves one CCD camera and two light sources (one NIR light source for palmvein and one visible blue light source for palmprint). The light sources switch quickly and the two images can then be acquired within 1 s. More details of the device are provided in Section 2. With the captured palmprint and palm- vein images, the personal verification framework has four main parts. (1) First, it does liveness detection by analyzing the bright- ness and texture of the acquired image. (2) Second it performs tex- ture coding to extract palmprint features. (3) Third, it uses matched filters to extract the palmvein features. (4) At last, a fu- sion matching score is computed through dynamic weight sum by considering the palmvein image quality. The rest of the paper is organized as follows. Section 2 describes the joint palmprint and palmvein system structure. Section 3 intro- duces the verification framework of the system. Section 4 reports experimental results on the established comprehensive database. Section 5 makes the conclusion and suggests some future work directions. 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 2. System description The developed data acquisition device is made up of two light sources, a light controller (which is used to switch the lights on or off), a grey level CCD camera, lens, and an A/D (analogue-to-dig- ital) converter connecting the CCD and the computer. The CCD is fixed at the bottom of the device. We use both visible white light and NIR light as the illumination sources. The palm lines can be clearly acquired under the visible light, while it is possible to ac- quire the vein structures beneath the palm skin but not palm lines or wrinkles under the NIR light. The uniformly distributed 880 nm LEDs (light emitting diode) are used for the NIR illumination. The light within range 700–1000 nm can penetrate human skin to a depth of 1–3 mm and it has been shown that 880–930 nm NIR light can provide a relatively good contrast of subcutaneous veins (Zha- rov et al., 2004). In order for a wider spectrum range for more com- plementary information, the uniformly distributed Blue LEDs (peaking at 470 nm) is used as the visible light. In order to reduce the cost of imaging system, a standard CCTV (Closed Circuit Television) camera, instead of a near-infrared sensi- tive camera, is used in our system. Its sensitivity in the NIR range is not as strong as that of NIR sensitive camera but has a much lower price. On the other hand, since this camera is also used to capture the palmprint image under the visible light illumination, no NIR fil- ter is used with the camera to cut out visible light. As a result, the quality of NIR palmvein images captured by our device is not as good as those by NIR cameras and/or with an NIR filter. However, Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 as we will see in this paper, the fusion of palmvein can still contrib- ute much in improving the performance of palmprint recognition. Fig. 1 shows the prototype of our system. The images can be captured in two different sizes, 352 * 288 and 704 * 576. Users are asked to put their palms on the platform with several pegs serving as control points for the placement of the hand. The com- puter then collects one palmprint image and one palmvein image under two different lighting conditions. The switching between two types of light is very fast and allows us to capture the two images in a very short time (<1 s). As shown in Fig. 2, the transla- tion or rotation is very small between the two images, so registra- tion between the two images can be omitted. Before feature extraction, it is necessary to extract from the original images a specific portion to work with. This is known as extraction of region of interest (ROI). The ROI extraction has two important advantages. First, it serves as a pre-processing to remove the translation and rotation of palmprint/palmvein images introduced in the data collection process. Second, ROI extraction extracts the most informative area in the images. It reduces a lot the data amount without losing much useful information. This will speed up the following feature extraction and matching processes. In this study, we set up ROI coordinates on palmprint image using the algorithm proposed in Zhang et al. (2003) and then use the coordinates to crop the ROI from palmprint and palmvein images. Fig. 3 shows some samples of ROI. 3. Joint palmprint and palmvein verification The flowchart of proposed online joint palmprint and palmvein verification system is illustrated in Fig. 4. It has four main stages. First the ROI is extracted; then a liveness detection algorithm based on image brightness and texture is applied; if the input images pass the liveness detection, palmprint features will be ex- tracted by texture coding and palmvein features will be extracted by matched filters; finally, the score level fusion is applied through dynamic weighted sum for decision making. The details of ROI extraction can be found in Zhang et al. (2003). In the following we discuss the processing of other stages. 3.1. Liveness detection There are various liveness detection methods in biometric sys- tems. For example, perspiration detection in a sequence of finger- print images (Parthasaradhi, Derakhshani, Hornak, & Schuckers, almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 220220 221 222 223 224 225 226 228228 229 230 231 232 233 Fig. 2. Sample images of (a) palmprint and (b) palmvein from the same palm. Fig. 3. ROI sample images. The top row shows the palmprint ROIs and the second row shows the associated palmvein ROIsQ2 . D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx 3 ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 2005); using additional hardware to acquire life signs; utilizing inherent liveness features such as facial thermograms (Schukers, 2002). However, these methods can be time-consuming, require addition hardware and costly. In our system, the low-cost NIR LED is used for illumination. It has been proved that 700– 1000 nm NIR light could penetrate human skin 1–3 mm inside, and blood will absorb more NIR energy than the surrounding tis- sues (e.g. fat or melanin), so the vein structure is darker than other areas in the palmvein image (Zharov et al., 2004). However, since the skin of some people, especially female, is relatively thicker (Lee & Hwang, 2002), their palmvein structures cannot be clearly captured (e.g. Fig. 3f). On the other hand, the fake palm made by some materials can also lead to dark lines under NIR illumination by using our system, e.g. Fig. 5a. Therefore, it will be difficult to ap- ply liveness detection by detecting only the existence of dark lines in the palmvein image. As human skin has special reflectance and absorbance proper- ties under the NIR spectral, the features associated with these properties can be extracted from the image brightness and texture for telling true palm from fake palms. Fig. 5 shows the palmvein images of several fake palms we made from different materials. After observing these images and palmvein images from true palms, we found that the image brightness and gray level co-occur- rence matrix (GLCM) (Haralick, Shanmugam, & Dinstein, 1973) Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 entropy could provide enough discriminant information to distin- guish them. Thus, we propose a liveness detection algorithm by analyzing palmvein image brightness and texture features. The brightness feature is defined as the average of the intensity over the image: B ¼ 1 M�N XM x¼1 XN y¼1 fðx; yÞ; ð1Þ where f(x, y) represents the gray value at pixel (x, y), and M and N represent the numbers of rows and columns in the image. On the other hand, the GLCM is a widely used texture operator in image processing and pattern recognition. For a given angle h and distance d, a GLCM is defined as: ph;dði;jÞ ¼ #f½ðx1;y1Þ;ðx1 þDx;y1 þDyÞ�2 Sjfðx1;y1Þ¼ i&fðx1 þDx;y1 þDyÞ¼ jg #S ; ð2Þ where Dx = d*cos h and Dy = d*sin h. S is the set of pixels in the im- age and ‘‘#” means the number of the elements in a set. (i, j) is the coordinate in the GLCM. With GLCM, several statistics could be computed, such as entro- py, contrast, correlation, energy, and homogeneity. Among them, almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 234 235 236 237 239239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 264264 265 266 267 268 269 270 Palmprint and palmvein image inputs ROI extraction No Live sample? Yes PalmVein Feature Extraction and Matching Palmprint Feature Extraction and Matching Dynamic Score Level Fusion End Fig. 4. Flowchart of online joint palmprint and palmvein verification. 4 D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 entropy is a popular feature to represent the uniformity of image texture. The more uniform the texture distributes, the bigger the entropy is. The GLMC entropy is computed as E ¼ XL i¼1 XL j¼1 pði; jÞ�ð� ln pði; jÞÞ; ð3Þ where L is the level of quantization. Fig. 5. NIR palmvein images of fake palms made from (a) cardboard; (b) foam; Table 1 The brightness and GLCM entropy of true and fake palm samples. 3b 3d 3f 5a 5b B 105.9 104.6 107.1 110.2 108.0 E 5.5 5.1 5.1 6.6 7.5 Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 Table 1 shows the brightness and GLCM entropy of the palm- vein images in Figs. 3 and 5. We can see the brightness and GLCM entropy values of fake palms and true palms are very different. Therefore, with some training samples, a classifier can be learned to tell the true palm from fake palms. In Section 4, we will establish a training dataset and learn the classifier. A testing dataset is also built to test the liveness detection method. 3.2. Palmprint feature extraction and matching In general there are three kinds of palmprint feature extraction algorithms, subspace learning (Connie et al., 2005; Hu et al., 2007; Ribaric & Fratric, 2005; Wu et al., 2003), line detection (Han et al., 2003) and texture-based coding (Kong & Zhang, 2004; Zhang et al., 2003). Among them, orientation texture-based coding (Kong & Zhang, 2004) is preferred for online system as it could achieve high accuracy. It is also fast for matching and can be easily implemented in real-time. The orientation of palm lines is stable and can serve as distinc- tive features for personal identification. To extract the orientation features, six Gabor filters along different orientations (hi = jp/6, where j = {0, 1, 2, 3, 4, 5}) are applied to the palmprint image. Here the real part of the Gabor filter is used and it is defined as: wðx; y; x; hÞ¼ xffiffiffiffiffiffiffiffiffiffi 2pj p e� x2 8j2 ð4x02þy02Þ eixx 0 � e� j2 2 � � ; ð4Þ where x0 = (x � x0)cos h + (y � y0)sin h, y0 = �(x � x0)sin h + (y � y0)cos h, (x0, y0) is the center of the function; x is the radial frequency in radians per unit length and h is the orientation of the Gabor functions in radians. j is defined by j ¼ ffiffiffiffiffiffiffiffiffiffiffiffi 2 ln 2 p 2dþ1 2d�1 � � , where d is the half- amplitude bandwidth of the frequency response. To reduce the influ- ence of illumination, the direct current is removed from the filter. (c) glove; (d) plaster; (e) plastic; (f) plasticine; (g) print paper; and (f) wax. 5c 5d 5e 5f 5g 5h 115.5 126.7 113.9 71.1 127.6 114.4 7.2 6.3 6.5 6.8 6.6 6.5 almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 287287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 304304 305 306 307 308 309 310 311 312 313 315315 316 317 318 319 320 321 322 324324 325 326 327 328 329 330 331 332 333 334 335 337337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 368368 369 370 371 Fig. 6. Orientation feature map of palmprint. (a) Original palmprint image; and (b) extracted feature map (different gray levels represent different orientation). D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx 5 ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 By regarding palm lines as the negative lines, the orientation corresponding to the minimal Gabor filtering response (i.e. the negative response but with the highest magnitude) is taken as the feature for this pixel (Kong & Zhang, 2004). Because the con- tour of Gabor filters is similar to the cross-section profile of palm lines, the higher the magnitude of the response, the more likely there is a line. Since six filters are used to detect the orientation of each pixel, the detected orientation {0, p/6, p/3, p/2, 2p/3, 5p/ 6} can then be coded by using three bits {000, 001, 011, 111, 110, 100} (Kong & Zhang, 2004). Fig. 6 shows an example of the ex- tracted orientation feature map, where different gray levels repre- sent different orientation. Based on the extracted 3-bit orientation feature maps, the Ham- ming distance between two maps can be calculated as follows: DðP; QÞ¼ PM y¼0 PN x¼0 P3 i¼1ðP b i ðx; yÞ� Q b i ðx; yÞÞ 3M�N ; ð5Þ where P and Q are two feature maps, Pbi Q b i � � is the ithbit plane of P(Q) and � is bitwise exclusive OR. To further reduce the influence of imperfect ROI extraction, we translate one of the two feature maps vertically and horizontally from �4 to 4 when matching with another feature map. The minimal distance obtained by translated matching is regarded as the final distance. 3.3. Palmvein feature extraction and matching It is observed that the cross-sections of palmveins are similar to Gaussian functions. Fig. 7 shows some examples. Based on this observation, the matched filters (Hoover, Kouznetsova, & Gold- baum, 2000; Zhang, Li, You, & Bhattacharya, 2007), which are widely used in retinal vessel extraction, can be a good technique to extract these palmveins. The matched filters are Gaussian- shaped filters along angle h: grhðx; yÞ¼�exp � x02 r2 � � � m; for jx0j6 3r; jy0j6 L=2; ð6Þ where x0 = xcos h + ysin h, y0 = �xsin h + ycos h, r is the standard devi- ation of Gaussian, m is the mean value of the filter, and L is the length of the filter in y direction which is set empirically. In order to suppress the background pixels, the filter is designed as a zero- sum. For one r, four different angle filters (hj = jp/4, where j = {0, 1, 2, 3}) are applied for each pixel, and the maximal response among these four directions is kept as the final response for the given scale1: 372 373 374 375 1 Different from the CompCode in Section 3.2 where the minimal response is used, here, the shape of used matched filters is identical to the cross-section of vein, thus the maximal response is kept. Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 RrF ¼ max R r hj ðx; yÞ � � ; j ¼f0; 1; 2; 3g; Rrhjðx; yÞ¼ g r hj ðx; yÞ�fðx; yÞ; ð7Þ where f(x, y) is the original image and * denotes the convolution operation. As shown in Bao, Zhang, and Wu (2005), Zhang et al. (2007), the multi-scale product of filtering responses is a good way to enhance the edge structures and suppress noise. The product of matched fil- ter responses at two scales r1 and r2 is defined as: Pðx; yÞ¼ Rr1F ðx; yÞ � R r2 F ðx; yÞ: ð8Þ The scale parameters r1 and r2 are set empirically. After computing the multi-scale production, we binarize it by using a threshold which is empirically set based on a training dataset. The vein pixel, whose scale production response is greater than the threshold, is represented by ‘‘1”, while the background pixel is represented by ‘‘0”. At last, some post-processing operations are performed to re- move some small regions. Fig. 8 illustrates the whole procedure of palmvein extraction. The extracted palmvein maps are binary images, and the dis- tance between two palmvein maps is computed as: DðP; QÞ¼ 1 � PM y¼0 PN x¼0ðP bðx; yÞ&Q bðx; yÞÞPM y¼0 PN x¼0ðP bðx; yÞjQ bðx; yÞÞ ; ð9Þ where P and Q are two palmvein feature maps, ‘‘& ” is bitwise AND operator and ‘‘j” is bitwise OR operator. The dissimilarity measure- ment in (9) is different from the Hamming distance used in Zhang et al. (2007). This is because most of pixels in the palmvein map are non-palmvein pixels. For example, in our database the average ratio of non-palm-vein pixels is about 86%. Such an uneven distribution of palm-vein and non-palm-vein pixels makes the Hamming dis- tance less the discriminative (Daugman, 2003). Similar to that in palmprint feature map matching, we translate one of the palmvein feature maps vertically and horizontally from �4 to 4 and match it with another palmvein feature map. The min- imal distance obtained by translated matching is regarded as the final distance. 3.4. Palmprint and palmvein fusion The information presented by multiple traits can be fused at various levels: image level, feature level, matching score level or decision level (Ross, Nadakumar, & Jain, 2006). Although image and feature level fusion can integrate the information provided by different biometric, the required registration procedure is too time-consuming (Wang et al., 2008). As to matching score fusion and decision level fusion, it has been found (Ross et al., 2006) that the former usually works better than the later because match scores contain more information about the input pattern and it is easy to access and fuse the scores generated by different matchers. For these reasons, matching score level fusion is the most com- monly used approach in multimodal biometric systems. In this work, we test sum and weighted sum on palmprint and palmvein matching score fusion: FDSum ¼ DPalmprint þ DPalmvein; ð10Þ FDWeight sum ¼ W Palmprint DPalmprint þ W PalmveinDPalmvein; ð11Þ where DPalmprint and DPalmvein are the palmprint and palmvein matching scores obtained by Eqs. (5) and (9), respectively. WPalmprint and WPalmvein are the weights for palmprint and palmvein feature in the fusion. Considering that not all palmvein images have clear vein struc- tures (referring to Fig. 3), it is intuitive that good quality palmvein images should have higher weight in the fusion than those poor almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 376 377 378 379 380 382382 383 384 385 386 387 388 389 390 391 392 393 394 396396 397 399399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 (a) (b) (c) 0 2 4 6 8 10 12 90 95 100 105 110 115 Position G ra y V al ue vein (c) vein (b) vein (a) (d) Fig. 7. (a)–(c) Some palmvein images and (d) the cross-section of vein structures. 6 D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 quality images. Here a dynamic weighted sum fusion scheme by incorporating the palmvein image quality is proposed. We define an objective criterion (Daugman, 2003) to evaluate the palmvein image quality: d0 ¼ jlvein � lnon-veinjffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi r2vein þ r 2 non-vein � � =2 q ; ð12Þ where lvein and lnon-vein are the average intensity values of vein pix- els and non-vein pixels extracted by in Section 3.3, and rvein and rnon-vein are the standard deviation of vein and non-vein pixels, respectively. For a clear palmvein image, the boundary between vein and non-vein structures is clear, so a higher d0 will be obtained. While for an unclear palmvein image, the boundary is not clear and the d0 will be smaller. For example, the d0 values of Fig. 3b, d and f are 1.63, 1.11 and 0.16, respectively. It shows that these images could be well classified by the proposed criterion. By incorporating the palmvein image quality into consideration, a dynamic weighted sum scheme is proposed as: FDWeight sum ¼ W Palmprint DPalmprint þ W PalmveinDPalmvein ¼ 1 � �d0 2 ! DPalmprint þ �d0 2 DPalmvein; ð13Þ �d0 ¼ d0 � d0min d0max � d 0 min ; ð14Þ where d0min and d 0 max are the minimal and maximal value of d 0 com- puted by a training dataset. As shown in (13), if quality of palmvein image is very good, the weights of both palmprint and palmvein Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 will be close to 0.5; if quality of palmvein image is very poor, the system will depend on palmprint solely. The experiments in Section 4.4 will validate the effectiveness of our dynamic fusion scheme. 4. Experimental results 4.1. Database establishment By using the developed joint palmprint and palmvein device, we established a database which includes palmprint and palmvein images from 500 different palms. The subjects were mainly volun- teers from the Shenzhen Graduate School of Harbin Institute of Technology and The Hong Kong Polytechnic University. In this database, 396 palms are from male and age ranges from 20 to 60. The images were captured by two separate sessions. The aver- age time interval between two sessions was 9 days. On each ses- sion, the subject was asked to provide six samples from his/her palms. For each shot, the device collected one palmprint and one palmvein image simultaneously (less than 1 s). Finally, the data- base contains 6000 palmprint and palmvein images of resolution 352 * 288. 4.2. Experimental results of liveness detection We established a fake palm database, which includes 489 images from the fake palms made by eight different materials: cardboard, foam, glove, plaster, plastic, plasticine, print paper, and wax. This data set is randomly partitioned into two equal parts, a training set and a test set. The same division strategy is almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 427 428 429 430 Fig. 8. Palmvein feature extraction. (a) Original palmvein images; (b) matched filter response with scale r1; (c) matched filter response with scale r2; (d) response product of the two scales; (e) binary image after thresholding; and (f) final vein map after post-processing. D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx 7 ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 applied to the true palm database. The distribution of brightness (B, refer to Eq. (1)) and entropy (E, refer to Eq. (3)) of the training Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 and test sets are plotted in Fig. 9. Because the skin reflectance and absorbance are different from those eight materials, we can almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 431 432 433 434 435 436 438438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 8 D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 see there is a clear boundary between them. In most of cases, dif- ferent materials are clustered into specific regions as they have specific properties under NIR illumination. Four thresholds (i.e. the boundaries of the rectangle in Fig. 9) are computed based on the training samples: Bmin ¼ minðBðTSÞÞ� rBðTSÞ; Bmax ¼ maxðBðTSÞÞþ rBðTSÞ; Emin ¼ minðEðTSÞÞ� rEðTSÞ; Emax ¼ maxðEðTSÞÞþ rEðTSÞ; ð15Þ where TS represents the whole training set, rB(TS) and rE(TS) are the standard deviation values of B and E in the training set, respectively. With these four thresholds, all of the training samples could be correctly classified as shown in Fig. 9a. For the test set, only one genuine palmvein is wrongly rejected as shown in Fig. 9b. In prac- tice, we can use more strict thresholds to reject the imposter palms. In case the user is wrongly rejected, he/she may put his/ her palm one more time on the system for verification. Currently, the fake palm database is a relatively small, and setting up a large fake palm database and investigate more advance feature extrac- tion methods will be our future focus. 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 4 4.5 5 5.5 6 6.5 7 7.5 8 60 70 80 90 100 110 120 130 140 Entropy B rig ht Cardboard Foam Glove Plaster Plastic Plasticine Print Paper Wax True (a) 4 4.5 5 5.5 6 6.5 7 7.5 8 60 70 80 90 100 110 120 130 140 Cardboard Foam Glove Plaster Plastic Plasticine Print Paper Wax True Entropy B rig ht Brightness and GLCM entropy distribution of test set. (b) Fig. 9. Brightness and GLCM entropy distribution of fake and true palm under NIR illumination. The rectangle is the boundary learned from the training set. (a) Distribution of the training set. (b) Distribution of the test set. Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 4.3. Experimental results on palmprint verification To obtain the verification accuracy for palmprint, each palmprint image is matched with all the other palmprint images. A match is counted as genuine matching if the two palmprint images are from the same palm; otherwise, the match is counted as impostor matching. The total number of palmprint matches is 6000 � 5999/2 = 17,997,000, and among them there are 33,000 (12 * 11/2 * 500) genuine matching, others are impostor matching. Equal error rate (EER), a point when false accept rate (FAR) is equal to false reject rate (FRR), is used to evaluate the performance. The distance distribution of genuine and impostor is shown in Fig. 10 and the receiver operating characteristic (ROC) curve is showed in Fig. 11. Using palmprint, we can get FRR = 2.10% when FAR = 5.6e�6%. The EER is about 0.0352%. The accuracy on palmprint is comparable to those of state-of-the-art (EER: 0.024%, multi-scale feature extraction) (Zuo, Yue, Wang, & Zhang, 2008) on the public palmprint database (PolyU Palmprint Database, 2006) collected under white illumination. 4.4. Experimental results on palmvein verification Using the same matching scheme as in Section 4.3, the distance distribution of genuine and impostor of palmvein images is illus- trated in Fig. 12. The ROC curve is displayed in Fig. 13. For compar- ison, the curve by using Hamming distance as in Zhang et al. (2007) is also plotted. From Fig. 13, we can see the proposed dissimilarity measure- ment could improve the verification accuracy significantly over the widely used Hamming distance. The EER of palmvein verifica- tion is about 0.3091%, which is not as accurate as palmprint verifi- cation but better than the result reported in Zhang et al. (2007) (98.8% GAR when FAR = 5.5%). This is largely due to the relatively low quality of palmvein images. As discussed in Section 2, in order to capture palmprint and palmvein image simultaneously, by using our developed device the palmvein image quality is not as good as that of palmprint. To better evaluate the effect of image quality on the verification accuracy, we partition the palmvein images into three equal sets by the proposed criterion d0: good quality images, average quality images and poor quality images. The ROC curves for three sets are plotted in Fig. 14. Good quality palmvein image set has much bet- ter results than poor quality image set. The EER values for good, average and poor quality image sets are is 0.0898%, 0.1214% and 0.3199%, respectively. Fig. 10. Matching distance distribution of palmprint. almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 492 493 494 495 496 497 498 499 500 501 503503 504 505 506 507 508 509 510 Fig. 11. ROC curve of palmprint. Fig. 12. Matching distance distribution of palmvein. Fig. 13. ROC curve of palmvein. Fig. 14. ROC curves of palmvein on different image quality. Fig. 15. ROC curves for different fusion schemes. D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx 9 ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 4.5. Experimental results by palmprint and palmvein fusion The distance distributions of palmprint and palmvein are differ- ent, which can be seen in Figs. 10 and 12. For example, the impos- tor scores of palmprint concentrate on 0.45 with standard deviation being 0.015, and that of palmvein concentrates on 0.89 with standard deviation being 0.023. Thus normalizing the match- ing scores before fusion is necessary. As the impostor distribution looks like a Gaussian shape, the z-normalization (Ross et al., 2006) is used here: DN ¼ D � limpostor rimpostor : ð16Þ The ROC curves of palmprint and palmvein fusion are shown in Fig. 15. The EER values of sum (Eq. (11)) and weighted sum (Eq. (12)) are 0.0212% and 0.0158%, respectively. Because palmvein contains complementary information to palmprint, the fusion of them could improve the system accuracy significantly. Fig. 16 shows an example, the two palms have similar palmprint patterns and they will be falsely accepted by using only palmprint images as almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 Table 2 Execution time. Time (ms) Image acquisition <1000 ROI extraction 63 Liveness detection 0.67 Palmprint feature extraction 36 Palmvein feature extraction 21 Palmprint matching 0.10 Palmvein matching 0.19 Fig. 16. An example pair of palms with similar palmprint which may be recognized wrongly by palmprint, but different palmvein features could separate them well. (a)–(b) from one palm; (c)–(d) from another palm. 10 D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 inputs. However, their palmvein patterns are very different. Thus, by combining palmprint with palmvein, they could be separated easily. Since the proposed weighted sum fusion incorporates palm- vein image quality, better accuracy could be achieved as shown in Fig. 15. Compared with the sum rule, the weighted sum could re- duce the EER up to 25%. 4.6. Speed The system is implemented using Visual C++6.0 on a Windows XP, T6400 CPU (2.13 GHz) and 2 GB Ram PC. The execution time for each step is listed in Table 2. The total execution time of verifica- tion is less than 1.2 s, which is fast enough for real-time applica- tion. As the speed of matching is fast, it can be easily extended to identification system. For example, for 1-to-1000 identification, the total time cost is only 1.4 s which could meet the real-time application. 5. Conclusion In this paper, we designed and developed an online palmprint verification system by fusing palmvein information. To improve anti-spoofing ability of the system, a liveness detection algorithm Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 based on the analysis of brightness and texture of image is pro- posed, and the experiment results show the effectiveness of the proposed method. Because the palmprint and palmvein informa- tion is very different, the improvement of simple sum is significant better than either information alone. Further improvement could be achieved by the proposed fusion scheme. Through experiments on a large database, our system shows that it can verify a person in 1.2 s with EER only about 0.0158%. In summary, we conclude that fusion of palmprint and palm- vein is a good method to get an accurate and robust personal ver- ification system. For further improvement of the system we will focus on three directions: (1) to investigate image and feature level fusion schemes; (2) to improve the palmvein image quality; (3) to collect more fake palm samples. Acknowledgements The work is partially supported by the GRF fund from the HKSAR Government (PolyU 5351/08E), the central fund from Hong Kong Polytechnic University, Key Laboratory of Network Oriented Intelligent Computation (Shenzhen), Science Foundation of Shenz- hen City (CXQ2008019), and the Natural Science Foundation of China (NSFC) (Nos. 60620160097 and 60803090). References Bao, P., Zhang, L., & Wu, X. L. (2005). Canny edge detection enhancement by scale multiplication. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27, 1485–1490. Connie, T., Jin, A., Ong, M., & Ling, D. (2005). An automated palmprint recognition system. Image and Vision Computing, 23, 501–515. Cross, J. M., & Smith, C. L. (1995). Thermographic imaging of subcutaneous vascular network of the back of the hand for biometric identification. In Proceedings of IEEE 29th international Carnahan conference on security technology (pp. 20–35). Daugman, J. (2003). The importance of being random: statistical principles of iris recognition. Pattern Recognition, 36, 279–291. Han, C., Cheng, H., Lin, C., & Fan, K. (2003). Personal authentication using palm-print features. Pattern Recognition, 36, 371–381. Hao, Y., Sun, Z., & Tan, T. (2007). Comparative studies on multispectral palm image fusion for biometrics. In Asian conference on computer vision (pp. 12–21). Haralick, R. M., Shanmugam, K., & Dinstein, I. (1973). Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics SMC, 3, 610–621. Hennings-Yeomans, P. H., Kumar, B. V. K., & Savvides, M. (2007). Palmprint classification using multiple advanced correlation filters and palm-specific segmentation. IEEE Transactions on Information Forensics and Security, 2, 613–622. Hoover, A., Kouznetsova, V., & Goldbaum, M. (2000). Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Transactions on Medical Imaging, 19, 203–210. Hu, D., Feng, G., & Zhou, Z. (2007). Two-dimensional locality preserving projections (2DLPP) with its application to palmprint recognition. Pattern Recognition, 40, 339–342. Jain, A., Bolle, R., & Pankanti, S. (Eds.). (1999). Biometrics: Personal identification in network society. Boston: Kluwer Academic. Kong, A., & Zhang, D. (2004). Competitive coding scheme for palmprint verification. In International conference on pattern recognition (pp. 520–523). Kong, A., Zhang, D., & Kamel, M. (2009). A survey of palmprint recognition. Pattern Recognition, 42, 1408–1418. Kono, M., Ueki, H., & Umemura, S.-I. (2002). Near-infrared finger vein patterns for personal identification. Applied Optics, 41, 7429–7436. Kumar, A., & Zhang, D. (2005). Personal authentication using multiple palmprint representation. Pattern Recognition, 38, 1695–1704. Lee, Y., & Hwang, K. (2002). Skin thickness of Korean adults. Surgical and Radiologic Anatomy, 24, 183–189. Lin, C. L., & Fan, K. C. (2005). Biometric verification using thermal images of palm- dorsa vein patterns. IEEE Transactions on Circuits and Systems for Video Technology, 14, 58–65. Macgregor, P., & Welfold, R. (1991). Veincheck: Imaging for security and personnel identification. Advanced Imaging, 6, 52–56. Parthasaradhi, S. T. V., Derakhshani, R., Hornak, L. A., & Schuckers, S. A. C. (2005). Time-series detection of perspiration as a liveness test in fingerprint devices. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, 35, 335–343. PolyU Palmprint Database (2006). <http://www.comp.polyu.edu.hk/�biometrics>. Ribaric, S., & Fratric, I. (2005). A biometric identification system based on eigenpalm and eigenfinger features. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27, 1698–1709. almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://www.comp.polyu.edu.hk/~biometrics http://www.comp.polyu.edu.hk/~biometrics http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 D. Zhang et al. / Expert Systems with Applications xxx (2010) xxx–xxx 11 ESWA 5101 No. of Pages 11, Model 5G 3 September 2010 Ross, A., Nadakumar, A. K., & Jain, A. K. (2006). Handbook of multibiometrics. Springer. Schukers, S. A. C. (2002). Spoofing and anti-spoofing measures. Information security technical report 7 (pp. 56–62). Socolinsky, D., Wolff, L., Neuheisel, J., & Eveland, C. (2001). Illumination invariant face recognition using thermal infrared imagery. In Proceedings of the IEEE computer society conference on computer vision and pattern recognition (pp. 527– 534). Su, C. (2009a). Palm extraction and identification. Expert Systems with Applications, 36, 1082–1091. Su, C. (2009b). Palm-print recognition by matrix discriminator. Expert Systems with Applications, 36, 10259–10265. Toh, K. A., Eng, H. L., Choo, Y. S., Cha, Y. L., Yau, W. Y., & Low, K. S. (2006). Identity verification through palm vein and crease texture. In International conference on biometrics (pp. 546–553). Wang, L., Leedham, G., & Cho, D. S.-Y. (2008). Minutiae feature analysis for infrared hand vein pattern biometrics. Pattern Recognition, 41, 920–929. Wang, J., Yau, W., Suwandy, A., & Sung, E. (2008). Person recognition by fusing palmprint and palm vein images based on ‘‘Laplacianpalm” representation. Pattern Recognition, 41, 1531–1544. Watanabe, M., Endoh, T., Shiohara, M., & Sasaki, S. (2005). Palm vein authentication technology and its applications. In The biometric consortium conference. Please cite this article in press as: Zhang, D., et al. Online joint palmprint and p j.eswa.2010.08.052 Available at <http://www.fujitsu.com/downloads/GLOBAL/labs/papers/ palmvein.pdf>. Wu, J., & Ye, S. (2009). Driver identification using finger-vein patterns with Radon transform and neural network. Expert Systems with Applications, 36, 5793–5799. Wu, X., Zhang, D., & Wang, K. (2003). Fisherpalm based palmprint recognition. Pattern Recognition Letter, 24, 2829–2838. Wu, X. Q., Zhang, D., Wang, K. Q., & Qi, N. (2007). Fusion of palmprint and iris for personal authentication. Advanced Data Mining and Applications, 466–475. Yao, Y. F., Jing, X. Y., & Wong, H. S. (2007). Face and palmprint feature level fusion for single sample biometrics recognition. Neurocomputing, 70, 1582–1586. Zhang, Y. B., Li, Q., You, J., & Bhattacharya, P. (2007). Palm vein extraction and matching for personal authentication. In 9th International conference VISUAL (pp. 154–164). Zhang, D., Kong, W., You, J., & Wong, M. (2003). Online palmprint identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25, 1041–1050. Zharov, V. P., Ferguson, S., Eidt, J. F., Howard, P. C., Fink, L. M., & Waner, M. (2004). Infrared imaging of subcutaneous veins. Lasers in Surgery and Medicine, 34, 56–61. Zuo, W., Yue, F., Wang, K., & Zhang, D. (2008). Multiscale competitive code for efficient palmprint recognition. In International conference on pattern recognition (pp. 1–4). almvein verification. Expert Systems with Applications (2010), doi:10.1016/ http://www.fujitsu.com/downloads/GLOBAL/labs/papers/palmvein.pdf http://www.fujitsu.com/downloads/GLOBAL/labs/papers/palmvein.pdf http://dx.doi.org/10.1016/j.eswa.2010.08.052 http://dx.doi.org/10.1016/j.eswa.2010.08.052 Online joint palmprint and palmvein verification Introduction System description Joint palmprint and palmvein verification Liveness detection Palmprint feature extraction and matching Palmvein feature extraction and matching Palmprint and palmvein fusion Experimental results Database establishment Experimental results of liveness detection Experimental results on palmprint verification Experimental results on palmvein verification Experimental results by palmprint and palmvein fusion Speed Conclusion Acknowledgements References 
work_jfopvflxkna3jp2tovexw7imbi	----	Author Guidelines for 8 1 Signal transmission model for the substations grounding grid Xianghui Xiao 1, 2 , Minfang Peng 1 , Jaime S. Cardoso2, Lei Wang 1 , Junfeng Guo 3 (1. Institute of Elec and Information Engineering in Hunan University, Changsha 410082, Peoples R China. 2. INESC Porto and Faculty of Engineering, University of Porto, Porto 4099-002, Portuguese Republic. 3.Wuhan National Laboratory for Optoelectronics of Huazhong University of Science and Technology, Wuhan 430074, Peoples R China) ABSTRACT The signal of the wireless sensor network in Grounding grid, owing to energy loss, network congestion, path constraints and other factors, is easy to delay even partially losing. In order to ensure that the signal can be transmitted effectively in grounding grids for the substation, this paper presents a method based on traffic model of back-off balanced multiple sensor network cooperation model. As we all know, cognitive radio (CR) technology is adopted in multi-channel wireless networks to provide enough channels for data transmission. The MAC protocols should enable the secondary users to maintain the accurate channel state information to identify and utilize the leftover frequency spectrum in a way that constrains the level of interference to the primary users. We proposed a novel cooperation spectrum sensing scheme in which the secondary users adopt backoff-based sensing policy based on the traffic model of the primary users to maximum the throughput of the network. To obtain the full accurate information of the spectrum is a difficult task so that we propose the backoff sensing as a sub-optimal strategy. Since the secondary users sense only a subset of the channels in our proposed scheme, less time is spent to get the channel state information as more time is saved for the data transmission. And while dealing the signal data, I combine the intensity transfer method instead of the priority method. This can effectively reduce the network congestion, to ensure that the main information can be transfer well. It is also very useful to signal transmission for the Multi-sensor in Grounding Grids for Substations. Index Terms— Cooperative Sensing, Backoff, Qos Ⅰ. INTRODUCTION As the grounding grids for the substation that the author is studying, is similar to the existing wireless sensor network, so in the establishment of the model of signal transmission, we no longer emphasizes grounding grids. In fact, the model is also practical for other wireless network. Allocating a fixed spectrum band to each wireless service has been proved to be inefficient since more than 90% of the spectrum is unused in most of the time while spectrum scarcity becomes a serious problem. As a result, Cognitive radio (CR) is developed to solve the problems resulting from limited spectrum resources and low utilization of the spectrum. CR network (CRN) allows the unlicensed users dynamically access the licensed spectrum to enable communication or improve service quality while no transmission of the primary users processing. Although the basic idea of the CR technology is simple, it imposes new challenges to the design of the MAC protocol. One of the most difficult, but important, design problems is how the secondary users effectively detect the existence of the primary users. Different sensing policies are proposed in [1]-[4] as possible solutions to this problem. Particularly, the authors in [1] suggest the secondary users shall know the state of each channel of the primary network which is modeled as an ON-OFF source alternating between state ON (active) and state OFF (inactive), however, when there are many channels to sense, the reporting phase will take a large portion of the slot which cannot be ignored. Therefore, the efficiency can be quite low due to the fact that the data transmission occurs only in the negotiating phase. In [2], the author use the statistical channel allocation to predict future spectrum usage and decide which channel to sense and access, but unfortunately, the complexity of this scheme increases quickly with the number of the licensed channels. The authors of [3] adopt stopping rule for the spectrum sensing as hardware restrictions are taken into consideration. However, the sensing priority is not discussed problem in this work. In [4], dissemination scheme for channel state information is studied with sensing priorities considered. The backoff-based sensing strategy proposed in this paper is actually a sensing priority adjustment scheme based on the traffic model of the primary users. Meanwhile, using the intensity transfer method, we transfer the data directly. The length of the traffic is modeled with exponential distribution, which is simple but effective, especially for the on-line traffic. The rest of the paper is organized as follows: Section Ⅱ introduces the traffic model of the primary network and presents the proposed cooperative sensing scheme. This is followed by Section Ⅲ which provides the performance of the proposed scheme with some numerical results obtained 2 from the simulation for the Grounding Grids in some Substation. Finally, the paper is concluded with section Ⅳ. Ⅱ. PROPOSED CHANNEL SENSING SCHEME A. System Model Overview We consider a licensed spectrum band divided into n data channels. Each licensed channel is time-slotted such that the primary users communicate with each other in a synchronous manner. Meanwhile, a number of ad-hoc network users, which are synchronized with the primary users, opportunistically access the licensed spectrum without imposing interference to the primary users. In this paper, we mainly focus on the scenarios where all secondary users utilize the licensed channels used by the same set of primary users. This implies that the licensed channel availability information sensed by each secondary user is consistent among all secondary users. The spectrum band is divided into N data channels index by i with i=1, 2… N and one control channel. Spectrum band of the control channel is pre-assigned and is disturbed from the primary users. In our proposed cognitive radio-based multi-channel MAC protocols, each secondary user is equipped with two transceivers. The first transceiver is devoted to operating over the dedicated control channel. The secondary users use their control transceivers to obtain the sensing results of the un-used licensed channels from other secondary users, and to negotiate with the other secondary users through the contention-based algorithms, such as IEEE 802.11 distributed coordination function (DCF) and Carrier Sense Multiple Access (CSMA) protocols. The second transceiver consists of a SDR module such that it can tune to any one of the n licensed channels to sense for spare spectrum, receive/transmit the secondary users’ packets. For convenience, we call the first transceiver the control transceiver and the second transceiver the SDR transceiver, respectively, in the rest of this paper. Each time slot is divided into sensing phase and negotiating phase. Sensing phase is used for channel sensing and information reporting while negotiating phase for data transmission and resources allocation among secondary users. Four types of packets are introduced into our scheme for exchanging control messages: 1) C-RTS/C-CTS: contention and spectrum reservation during the negotiating phase. 2) T-RTS/T-CTS: notify the other secondary users the completion of the transmission. 3) D-RTS/D-CTS: negotiate with the receiver node and initial the transmission. 4) RP: notify the other secondary users the sensing results. B. Traffic Model Traditionally, we assume that the primary user transmission request arrival is a Poisson stream, while the service time for each primary user is exponentially distributed. Thus, applying M/M/1 queuing model, we can model each channel as a Markov chain as shown in Fig. 2, where the variable in the circle represents the number of the primary user waiting for spectrum resources. The transition probability, denoted by ji q , , of the Markov chain can be written as:                    2, 1, 1, |Pr , jitO jitOt jitOt itSjttSq ji   (1) Thus, we are able to derive the probability transition matrix for the Markov chain  tS j , denoted by Q , as follows:                       02120 01 01 ,    ji qQ (2) Since the channel will be busy when   1tS i , we can simplify this Markov model to an ON/OFF channel usage model alternating between state ON (active) and state OFF (inactive). An ON/OFF channel usage model specifies a time slot in which the primary user signals is or isn’t occupying a channel. The secondary users can utilize the OFF time slot to transmit their own packets. Suppose that each channel changes its state independently. Let i  be the probability that i-th channel transits from state ON to state OFF and i  be the probability that i-th channel transits from state OFF to state ON, where Ni 1 . Then, the channel state can be characterized by a two-state Markov chain as shown in Fig. 1. The probability transition matrix is derived as follows:            i i ji qQ   1 1 , (3) Let  nP denotes the probability that the transmission on i- th channel lasts for n time slots, thus, we are able to derive that  nP is geometrically distributed:      11 1   nnP i n i  (4) Based on Eq. (4), the mathematic expectation, denoted by   nPE , of the traffic length can be derived as:      in nnPnPE  1 1     (5) 3 Fig. 1. two-state Markov chain C. Intensity transfer method for the data The so-called intensity transfer method is as below. When the control system is accurate input, the exact value put the linguistic variable value acquired from the former conditional statements to the next linguistic variable value. In most cases, fuzzy controller is a double input and single output controller. Wherein, an input for the deviation, another input for the rate of change of the deviation. Set input deviation X1, the change rate of the deviation X2, output Y, and their linguistic variables respectively Ai1, Bi2 and Yi said. Therefore, X1= {A11，A21，…，Aj1} X2= {B12，B22，…，Bn2} Y= {Y1，Y2，…，Yr，…，Ym} In addition, If X1=A11 and X2=B12 then Y=Y1 If X1=A11 and X2=B22 then Y=Y1 …… If X1=A11 and X2=Bn2 then Y=Yr if X1=A21 and X2=B12 then Y=Y2 …… If X1=Aj1 and X2=Bn2 then Y=Ym Detailed steps for the Intensity Transfer Method are as follows: 1) For a previous linguistic variable intensity Set input x1=a1, x2=a2; A1, A2 is an accurate value. According to the first rule, A1, A2 corresponding to language variables B11, B12 in the parameter X1, X2, then membership grades respectively are )()( 2111 1211 aaw BA   Analog， )()( 2112 2211 aaw BA   …… )()( 211 211 aaw nBAn   Can infer )()( 2121 1221 aaw BA   …… )()( 21 21 aaw nj BAjn   2) Seek the result of reasoning Due to the intensity transfer method is putting on the former role to transfer to a later. Then put the exact value to the membership grade of the next fuzzy quantity before. Therefore, for the first conditional, the reasoning results are 1111 * 11 )( YywY  Analogy, from the second conditional to the j×k conditional, the reasoning results are 1112 * 12 )( YywY  …… rrnn YywY )(1 * 1  2221 * 21 )( YywY  …… mmjnjn YywY )( *  According to the general result of reasoning, we can get the element that the later correspond to y1，y2，…,ym 3) Get the output accurate value b from the general results of reasoning Generally speaking we use the method with the centre of gravity to get the exact output value B, i.e. 4    dyyY ydyyY b )( )( * * 4) Take * 11Y as the membership grade for Y1, * 12Y as the membership grade for Y2, * 1nY as the membership grade for Yr; Meanwhile, take * 22Y as the membership grade for Y2, * jnY as the membership grade for Ym, At this time, when Y1, Y2,..., Ym is a monotone function, B can also be expressed as:         j p n q pq j p t n q pq Y yY b 1 1 * 1 1 * In equation above, t=0, 1, 2… m. After contrasting to the Mamdani fuzzy control method, fuzzy Smith control method and strength transfer method, we find the Mamdani fuzzy control method has two shortcomings whose adaptive ability is poor and the control precision is not high. Going through a two degree unstable object, dynamic performance and stability of the Mamdani fuzzy control method is weak [5]. Combining the Fuzzy control method and the Smith pre-estimation control method together, we can solve the time delay, time-varying, nonlinear complex system problems, but it is very complex to achieve [6]. And the variable domain of Intensity- transferring method is simple, intuitive, easy to understand, but also has a high control accuracy, According to the above analysis, we chose the variable domain of Intensity- transferring method in this method. D. Backoff-based Channel Sensing Policy The basic idea of our backoff-based channel sensing policy is that the secondary user will postpone the next sensing when the channel sensing result is turned out to be busy. Using Eq.4 to get the numerical results, we can further obtain that the probability decreases exponentially as the time slot increases. Therefore, it makes sense that the optimal performance is expected when adopting exponential backoff sensing policy. For the rest of this section, we will discuss details of our backoff-based channel sensing policy. Fig. 2. Frame Construction of our proposed scheme As Fig. 2 shows, for the i-th channel in time slot indexed by t with t =1, 2,… , T, (T +1), (T +2), …, the state of the i-th channel, denoted by )(ts i , corresponds to a binary random variable, with 0 and 1 representing the idle and the active states, respectively. A Backoff Time Counter (BTC) is used to record the backoff slots remained for each channel. From the view of the secondary users, the Let )(ta i denotes the channel state of the i-th channel kept in the channel state table by the secondary users, )(ta i may have the one of the four following values and it is updated whenever the secondary user senses the i-th channel or receives sensing results of the i-th channel from other secondary user: ·0: Unsensed: Channel that are not sensed and occupied by the other secondary users. ·1: Busy: Channel is occupied by primary users or other secondary users. ·2: Available: Channel is available according to the sensing results or the information in the report packets received from other secondary users. ·3 Backoff-Active: Backoff-based channel sensing policies active on this channel ( 0)( tb i ). Let )(td j denotes the j-th secondary user state, )(td j may have one of the three following values: 0: The secondary user reserves enough channel resources for its data transmission, 1: The secondary user needs more resources for data transmission. 2: The secondary user has no data to transmit. From the view point of the secondary users, the licensed channels are split into four subsets 30 ~ NN according to 5 the )(ta i values in any given t-th time slot, )3...0}()(|{  jjtaiN ij . Note that no common elements exist between any two of these sets and 2 N has no elements at the start of a slot since no channel has been sensed. The sensing phase of each slot is divided into n mini-slots, where n is the number of elements in 0 N . Following the settings in IEEE 802.11a [7], we assume each mini-slot last for MS T =9μs, which is long enough to determine whether a channel is busy or not, the sensing process is error free. If we denote S T , RP T , NP T as the time duration of the time slot, the reporting phase, and the negotiating phase, respectively, then we obtain: NPMSNPRPS TnTTTT  (2) Transmission only happens during the negotiating phase, let IDLE N denote the number of channels sensed to be idle, then we have the reward function in the given time slot: )()( MSSIDLENPIDLE nTTNTNtF  (3) We try our best to maximum F(t) by reducing n remarkably while IDLE N is comparable with the number of available channels in the network. At each mini-slot, a secondary user with 0)( td j randomly selects a channel in 1 N to sense while the other secondary users waiting for the channel report. Consider the j-th secondary user select i-th channel to sense, the SDR transceiver is tuned to the dedicated channel to get the channel information. Therefore, )(ts i is known by the j-th secondary user. If 1)( ts i , the secondary user updates )(ta i to 3 and sets a backoff period for this channel. The channel will be adding to 3 N and won’t be sensed until it expires. The length of the backoff window is set as 0 W slots, where the integer 0 W denotes the minimum window size. After waiting for this amount of backoff time, update )(ta i to 0 and sense the channel again. Every time the sensing result turns out to be busy, the backoff window for the channel will be multiplied by the backoff factor r. We predefined a constant value K represents the maximum times that the backoff-based sensing policy can be executed. After K times unsuccessful channel sensing, the backoff window will remain unchanged until the channel sensed to be idle. If )(ts i =0, the j-th secondary user realize the i-th channel is idle and will make the decision whether to keep it for transmission or not. If )(td j =1, the j-th secondary user will reserve i-th channel for data transmission. Therefore, the j-th node update )(ta i to 1 and )(td j base on the channel resources which has been reserved and spectrum required for transmission. )(ta i will be updated to 2 when )(td j =2, since the j-th node have no data to transmit. After a secondary user has reserved enough channel resources for its data transmission, transmission will be initialized by sending D-RTS packet to the receiver. After successfully receive a D-CTS packet since sending the last D-RTS, it gets the permission to transmit data packets in the coming next time slot. Since primary users may access channels at any time, the channels which are reserved by the secondary users still need to be periodically sensed to avoid colliding with the primary users. However, it is unnecessary to sense these channels with cooperation from all the secondary users of the network. Cooperation will only be operated between the sender and receiver which access these channels for transmission. Therefore, the frames have a different construction as shown in Fig.1. When primary user become active on the channels kept by the secondary user or the secondary user have finished their transmission, the secondary user will free the spectrum resources by sending T-RTS/T-CTS, which carries the index of the channels kept. The other secondary users, however, mark these channels with )(ta i =0 at the start of the next time slot which indicates that channels are available for sense. At the end of each mini-slot, )(ta i will be passed out to the other secondary users by sending a RP packet. The other secondary users, however, update their channel information table upon receiving the RP packet. Ⅲ. PERFORMANCE EVALUATION A. Simulation Model To the grounding grid for an 110KV substation, we make a simulated test. The substation grounding conductor consists of flat steel structure, the cross section is 40 6mm mm and the depth is 0.9m . The substantially shape can be shown in figure 3. 传感器 接地网 A D E F CB A, ··· , F 为所分区域 Fig. 3. Schematic diagram of real grounding grid 6 The 6 group of sensors are embedded in the A, B, C, D, E, F Figure 3 shows for the first. Assuming sensors and grounding conductor are in the same horizontal position, then the grounding conductor and the sensor are in the same soil environment. It can be identified that all the drift current comes from the grounding conductor will go through the electrode of the sensor entirely. So the sensor can transmit all the information of the grounding conductors. It is worth to point out, the A, B, C grid region is variable, and the final results should be decided by Tabu search. And the grounding grids for the substation is too small to transmit much information. The paper presents the date comes frome the tower. We compare the performance of our proposed scheme with the negotiation-based sensing policy proposed in [1]. The duration of traffic subject is to exponential distribution with an average length of 500 time slots. The spectrum utilization is denoted by γ, while the number of the data channels is denoted by n. We investigate the throughput results of the backoff-based sensing policy with different γ’s and n’s. Secondary users are distributed over a square with a side length of 100m. The secondary users’ movement and sensing errors are not considered in order to concentrate on the spectrum sensing. The length of time slot and mini-slot are set as 0.54ms and 9μs, respectively. Each secondary user generates packets of 40 bytes by Poisson process at rate 100 packets per second. The physical transmission rate is 40Mbps for the spectrum of data channels. The backoff factor(r) set as 2 and maximum backoff times set as 5. Total simulation time is 1 hour. B. Simulation Results We investigate the throughput performance differences between our proposed scheme and the comparison scheme. We obtain the numerical results for the saturation throughput achieved by two different channel-sensing policies under different situations. Fig. 4 and Fig. 5 plot the average throughput of the network against the channel utilization (γ) of the primary users and the average throughput of the secondary users against the number of channels (x), respectively. First, we investigate the impact of the channel utilization of the primary users on the performance of our proposed MAC protocols. The number (x) of licensed channels is set to 40. From Fig. 4, we observe that the average throughput of the network achieved by the backoff-based sensing policy is larger than that achieved by the negotiation-based sensing policy proposed in [1] for the same γ and x. Note that under the backoff-based sensing policy the average throughput of γ= 0.5 is the close to that achieved by the comparison scheme. This is because based on the backoff-based sensing policy, some time slots of the available channels are left to be unsensed, in other words, wasted, and the wasted transmission time cannot be ignored when the utilization is high. On the other hand, the performance become closer when γ is small, which makes sense since backoff-based sensing policy is only executed for a few channels. Fig. 4. the average throughput of the network achieved Fig. 5. The average throughput of the system according to the number of the channels Fig. 5 plots the impact of different number of channels on the average throughput of the system under both the backoff- based sensing policy and the comparison scheme, when the channel utilization is set as 0.2. Note that the data rate for a single channel remains the same during the simulation; the increase in the number of channels means a wider spectrum of the primary network. The average throughput of the network decreases remarkably against the number of the channels because more time is spent to get the message about the channel usage. Given γ and x, the saturation throughput achieved by the backoff-based sensing policy is higher than that by the comparison policy. Note that the length of the time slot is typically very short, while the number of channels of is typically very large. Take GSM cellular network for an example, the time-slot length is 577 μs and the number of traffic channels is 172. The throughput of the comparison scheme will decrease rapidly as the number of channels increases which is expected since a large portion of a time slot is used for getting the channel state map of the network. Our proposed scheme provides a possible solution to this problem. 7 Ⅳ. CONCLUSION We proposed and analyzed the opportunistic multi- channel MAC protocols for the cognitive radio-based wireless networks. Specifically, we provide a solution for selecting a channel for sensing and exchanging sensing results between the secondary users. It is hard for a secondary user to perfectly know the channel environment of the primary network; therefore, backoff-based channel sensing policy is studied as a possible solution. Instead of getting the accurate information of all the channels, we proposed backoff-based sensing policy as a sub-optimal solution. Applying the negative-exponential traffic model, we develop analytical models to evaluate the performance of our proposed scheme and the comparison scheme under different combination of channel utilization and number of channels for the saturation network case. And combining the intensity transfer method can effectively reduce the incidence of the network congestion. Because we only discuss the dissemination of information, the grounding wireless sensor networks are of not too much exposition [8]. Our analyses provide the guidelines to support the QoS requirements over cognitive radio based wireless networks. REFERENCES [1]. Hang Su, Xi Zhang, “Cross-Layer Based Opportunistic MAC Protocols for QoS Provisionings Over Cognitive Radio Wireless Networks”, IEEE Journal on Selected Areas in Communications Vol. 26, No. 1, January 2008. [2]. A. Hsu, D. Wei, and C. Kuo, “A cognitive MAC protocol using statistical channel allocation for wireless ad- hoc networks,” Proc. IEEE WCNC, March 11-15, 2007. [3]. Juncheng Jia, Qian Zhang, Xuemin Shen, HC-MAC: “A Hardware-Constrained Cognitive MAC for Efficient Spectrum Management” IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 26, NO. 1, JANUARY 2008. [4]. Jeong Ae Han, Wha Soak Jean, “Efficient Cooperative Channel Sensing in Cognitive Radio Ad Hoc Networks”, IEEE Journal on Selected Areas in Communications, Vol. 26, No. 1, January 2008 [5]. Li Hongxing. Adaptive fuzzy controllers based on variable universe [J]. Science in China(Series E: Technological Sciences). 1999.01: 11-20. [6]. Yin Haidong, Liu Feng, Xin Mingying. The Optimized Disign of the Fuzzy Controller (Ⅰ)——The predigested disquisition of rules of fuzzy control[J]. Journal of Northeast Agricultural University (English Edition), 2002.02: 153-157. [7]. IEEE Standard 802.11 - 1999; Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications; November 1999. [8]. Morteza Jalalat-evakilkandi, bolreza Mirzaei. A New Hierarchical-Clustering Combination Scheme Based on Scatter Matrices and Nearest Neighbor Criterions [J]. 2010 5th International Symposium on Telecommunications (IST’2010): 904-908. Date of script：2012-03-19。 Author introduction：Xinghui Xiao (1983- ), male, was born in shaoyang, China. PHD, lecturer. The main research directions are circuit testing and diagnosis, electromagnetic field theory and application, smart grid and so on; E-mail: xiaoen3@126.com Minfang Peng(1964-), female. She acquired her doctor’s degree in 2006. Professor, supervisor of a Ph.D. student. The main research directions are circuit testing and diagnosis, electromagnetic field theory and application, smart grid and so on; http://acad.cnki.net/kns55/oldNavi/Bridge.aspx?LinkType=BaseLink&DBCode=cjfq&TableName=cjfqbaseinfo&Field=BaseID&Value=JEXG http://acad.cnki.net/kns55/oldNavi/Bridge.aspx?LinkType=BaseLink&DBCode=cjfq&TableName=cjfqbaseinfo&Field=BaseID&Value=JEXG http://acad.cnki.net/kns55/popup/knetsearchNew.aspx?sdb=CJFQ&sfield=%e4%bd%9c%e8%80%85&skey=YIN+Hai++dong&scode= http://acad.cnki.net/kns55/popup/knetsearchNew.aspx?sdb=CJFQ&sfield=%e4%bd%9c%e8%80%85&skey=LIU+Feng&scode= http://acad.cnki.net/kns55/popup/knetsearchNew.aspx?sdb=CJFQ&sfield=%e4%bd%9c%e8%80%85&skey=XIN+Ming++ying++(Northeast+Agricultural+University&scode= http://acad.cnki.net/kns55/detail/detail.aspx?QueryID=4&CurRec=6&DbCode=CJFQ&dbname=CJFD9902&filename=DBYN200202012 http://acad.cnki.net/kns55/detail/detail.aspx?QueryID=4&CurRec=6&DbCode=CJFQ&dbname=CJFD9902&filename=DBYN200202012 http://acad.cnki.net/kns55/detail/detail.aspx?QueryID=4&CurRec=6&DbCode=CJFQ&dbname=CJFD9902&filename=DBYN200202012 http://acad.cnki.net/kns55/oldNavi/Bridge.aspx?LinkType=BaseLink&DBCode=cjfq&TableName=cjfqbaseinfo&Field=BaseID&Value=DBYN http://acad.cnki.net/kns55/oldNavi/Bridge.aspx?LinkType=BaseLink&DBCode=cjfq&TableName=cjfqbaseinfo&Field=BaseID&Value=DBYN http://acad.cnki.net/kns55/oldNavi/Bridge.aspx?LinkType=IssueLink&DBCode=cjfq&TableName=cjfqyearinfo&ShowField=cname&Field=BaseID*year*issue&Value=DBYN*2002*02 
work_jgbhrqh7u5ardlyufhz757f2wy	----	EM-REPS180016 119..138 Improving administrative decisions through expert systems: empirical analysis Marwa Gaber Ahmed Fahim Business Administration Department, Modern Academy for Computer Science and Management Technology in Maadi, Cairo, Egypt Abstract Purpose – The purpose of this paper is to introduce and identify the basic components, tasks and application areas of expert systems (ESs) as a decision support system that has been increasingly used in the business world lately and explore its potential for improving the effectiveness of administrative decisions in the public sector. Empirically, the paper explains the role of ESs in fostering decision-making processes at the Ministry of Investment and International Cooperation (MIIC) in Egypt. Design/methodology/approach – The design of this research is descriptive in the theoretical section and quantitative in the empirical one. Theoretically, the study adopted both the analytical approach and systems approach to demonstrate main concepts and relationships, while it conducted an empirical study to investigate the correlations in practice. Findings – The research concluded that the usage of ESs is deemed to be on the top of the technical solutions that might help public organizations develop their management quality and maintain competitive strength. In addition, the results proved that ESs contribute to administrative decisions at MIIC. Practical implications – The paper provides profitable findings and recommendations which can be applied by Egyptian public executives, in an attempt to ensure high quality and successful decisions using modern technology. Originality/value – This study has valuable implications for theory and practice together, as it offers numerous contributions to literature in the area of concern. Keywords Expert systems, Decision support systems, Managerial decision making, Public sector reform, Ministry of investment and international cooperation in Egypt Paper type Research paper 1. Introduction For further development of organizations in the present highly dynamic environment, their analysis is not just anticipated, but also required (Bolfikova et al., 2010, p. 155). Certainly, decision-making is becoming the remaining basis of excellence and competitive advantage that can guarantee superior returns for organizations (Harvey and TIS, 2007, p. 3). Successful organizations outperform their competitors in at least three ways; they make better decisions, they take decisions faster, and they implement decisions more. While most would agree with this, much less is known about what makes good or bad decisions (Dillon © Marwa Gaber Ahmed Fahim. Published in Review of Economics and Political Science. Published by Emerald Publishing Limited. This article is published under the Creative Commons Attribution (CC BY 4.0) licence. Anyone may reproduce, distribute, translate and create derivative works of this article (for both commercial and non-commercial purposes), subject to full attribution to the original publication and authors. The full terms of this licence may be seen at http://creativecommons.org/ licences/by/4.0/legalcode Expert systems and administrative decisions 119 Received 4 October 2018 Accepted 4 October 2018 Review of Economics and Political Science Vol. 3 No. 3/4, 2018 pp. 119-138 EmeraldPublishingLimited 2631-3561 DOI 10.1108/REPS-10-2018-011 The current issue and full text archive of this journal is available on Emerald Insight at: www.emeraldinsight.com/2631-3561.htm http://dx.doi.org/10.1108/REPS-10-2018-011 et al., 2010, p. 229). On the other hand, computers have been remarkable means used in managerial decision-making for decades. However, the hottest computerized decision aid during the 1980s was expert system (ES); that is the rising need for expert knowledge in decision-making within management can be most efficiently satisfied with the use of ESs, which would enable skills, experience and intuition to be used in real time and for an indefinite number of problematic situations (Dasic et al., 2011, p. 30) Hence, many organizations – private and public – have used ESs to assist their managers make better decisions. Indeed, understanding the technology and its implications and limitations is a central and critical aspect of any technology-related reform (Ahmad and Munir, 2016, p. 2). In general, the major difficulty in designing and implementing ESs is still the lack of knowledge and techniques on how to develop and apply them properly (Jayaraman and Srivastava, 1996, p. 27). Additionally, another vital challenge is to capture and encode the needed expertise/ wisdom from experts (Artificial intelligence and expert systems: knowledge-based systems, 2018). This problem seems to be more obvious in public sectors, putting into consideration the cognitive limits within the conservative organizational culture of these institutions. In sum, despite the progress and widespread usage of ESs, there is an alleged uncertainty about how people will react to this advanced technology, whether those using them as decision support systems (DSSs) or the others affected by the decisions reached, especially in public organizations, which makes ES applicability in governments a little bit questionable. For that reason, much of this discourse in literature is associated with business. Otherwise, past research in ESs has been minimal owing to the relative scarcity of information that exists on ES application within an organizational setting (Jayaraman and Srivastava, 1996, p. 27). Liao (2005) considered ES applications development as a problem- oriented domain. Consequently, until very recently it has been noticed the lack of a strong theoretical basis for viewing ESs – mostly DSSs – from a managerial perspective, not a technical one (Arnott and Pervan, 2008; Shim et al., 2002), in spite of the significance of this topic to management commonly and managerial decision-making particularly. The publications in this topic were scattered over many specializations, like artificial intelligence (AI), computer science, control engineering, logic, operations research and decision-making. Therefore, the ability to obtain new understanding using different social science methodologies should be the driving power of the future ES work (Liao, 2005). ES research needs to be based on more contemporary behavioral decision theory imported from management and other related fields to provide a stronger theoretical foundation for projects. In addition, this solid theoretical basis needs to be based on an increased number of interpretive case studies to illuminate areas of contemporary practice (Arnott and Pervan, 2008, pp. 667-669). So that, the researcher here believes this is a meaningful area that entails more attention and study, with specific reference to Egypt. Over and above, it is worth mentioning the distinct methodology adopted herein (open system approach) to address the concept of ES and the whole subject as well; is that the present research differentiates between ES outputs (recommended decisions) and actual administrative decisions, which is – this separation – not always found in literature (Jaradat et al., 2009). Thence, outputs are regarded here as a sub-indicator of the independent variable (which forms a system itself), not the dependent one. In conclusion, this work is designed to discover an obvious, concrete, and helpful means to know how to employ ESs in managerial decision-making. Thus, this study contributes to the existing body of knowledge by displaying an overview of ESs and the different decision areas in this regard. Besides, it is of interest to academics investigating the applications of these systems in public organizations. Moreover, the paper provides a way of exploring some of the changes and challenges that maybe encountered in practice. Hopefully, it could REPS 3,3/4 120 be a modest step towards conducting future inclusive research to appraise the technical and organizational renovations adopted by the Egyptian public sector, in its quest to achieve the sustainable development strategy (Egypt’s Vision 2030) with regard to maximizing the efficiency of governmental institutions. 2. Problem statement This research examines the beneficial aspects and the applicability of using ESs in governments, with the sake of bolstering the effectiveness of public executives’ decisions. The study provides a framework for introducing this modern decision instrument, and then it identifies the elements influencing the success of decision-making in the organizational practices, and analyzes the generic types of decisions that often require expertise to see when and if they are applied in public settings. Furthermore, this article looks at the bulk of technical reforms implemented in Egypt during the past years and evaluates the progress in this arena, through assessing the main characteristics of ESs and the organizational context that may affect the quality of the decision-making process in the case of the Ministry of Investment and International Cooperation (MIIC) to point out the limitation of the current procedures and to propose some practical solutions. Therefore, the study investigates a major research question which is: RQ: How can public sectors benefit from using ESs in improving the quality of administrative decisions and to what extent do ESs contribute to the managerial decision-making at MIIC? To answer RQ, the paper intends to detect the answers of the following sub-questions: Q1. What is the meaning of ES and what are its various tasks and benefits to management? Q2. What are the main classifications and stages of administrative decisions? Q3. How do ESs influence the decision-making processes in public organizations? Q4. How could ESs be applied for enhancing the quality of decisions at MIIC, considering the organizational factors? Subsequently, the research tests its master variables as illustrated in Figure (1): Figure 1. Research conceptual model Administrative Decisions (Dependent Variable) Organizational Factors (Moderating Variable) Expert Systems (ESs) (Independent Variable) Inputs Processes Outputs Feedback Source: Prepared by the researcher Expert systems and administrative decisions 121 3. Literature review: concepts and relationships description 3.1 Expert systems The field of AI is concerned about ways of developing systems that display elements of intelligent behavior. Those systems are designed to simulate human capabilities of sensing and thinking (Expert systems and applied artificial intelligence (2018)). ESs are a class of computer applications developed by researchers in AI. In essence, they are computer programs made up of a series of rules that analyze information about a nominated class of problems, as well as providing analysis of these problems, and depending upon their design they suggest a course of action to implement solutions or corrections (Alasgarova and Muradkhanli, 2008, p. 297). In that way, ES is a problem-solving package that imitates a human expert in a certain field, so it combines computer equipment, software, and specialized information to mimic expert human reasoning and advice. Typically, there are three key types of information transfer in ES which differentiate it from other DSSs (Arnott and Pervan, 2008). ES requires relevant information from the user about the problem domain. It also provides a recommendation based on the data given by the user, and if a non- expert requests information it offers a justification for the suggested decision (Jayaraman and Srivastava, 1996, pp. 27-28). As a result, ES differs from traditional DSS in that its knowledge base is way complex because of less reliance on the end-user to interpret and evaluate findings [Artificial intelligence and expert systems: knowledge-based systems, (2018)], although the current study believes that the two terms are used interchangeably in somehow. Early ESs performed fundamentally medical diagnoses and geological prospecting. With the mounting number of successful applications of ESs in those areas, AI specialists started to think that this technology might also be used to assist management in taking and implementing decisions. Research in this sphere has typically focused on how information technology can uphold the efficiency with which a user makes a decision and the effectiveness of that decision (Shim et al., 2002, p. 2). Nowadays, ESs are applied widely in commercial and industrial settings (Sahin et al., 2012). They have been used in manufacturing sectors, production scheduling, mining operations, agricultural activities, medical services, financial transactions, sales, control, and security departments. Moreover, they have been utilized to perform plenty of functions; the most popular among them are; consulting and recommending certain behaviors, designing and putting specifications, diagnosing and verifying objects, interpreting and confirming reliability, planning and creating actions, predicting future events, monitoring and evaluating signals, teaching and propagation of knowledge (Jayaraman and Srivastava, 1996, p. 29-30), classifying and identifying objects, and eventually generating options and solutions. In general, ESs have four main components; knowledge base, inference engine, knowledge acquisition module/justifier and user interface. The knowledge base contains the domain specific knowledge derived from the expert’s experience. One of the most commonly used ways to represent knowledge is as rules. The inference engine is a computer program that offers a methodology for reasoning and formulating conclusions (Expert systems and applied artificial intelligence (2018)). The knowledge acquisition module enables experts to save their knowledge in the knowledge base to draw and deduce new knowledge from the existing one via a machine learning process. Therefore, the justifier presents the whole chain of the decision-making process and the rules used to reach a conclusion. Finally, the interface for input/output is designed to communicate and interact with the user, environment and other systems such as databases (Alasgarova and Muradkhanli, 2008, p. 298). REPS 3,3/4 122 In this respect, the major players in developing ES are the knowledge engineer and the domain expert who act as builders. Once the system is completed, it is subject to consultation by end-users [Artificial intelligence and expert systems: knowledge-based systems, (2018)]. Actually, there are five essential stages for building ESs as follows (Chau, 1991): (1) Identification: in which the domain expert and knowledge engineer agree on the system’s purposes, specify the needed computer facilities, conduct a feasibility study and set a time frame for the project. (2) Conceptualization: the expert and knowledge engineer here determine the system design, basic concepts, data relationships and information-flow traits that describe the problem solving process. (3) Formulation: where the expert and engineer develop a knowledge base using the key concepts and relationships. The engineer must choose a programing language, and with the help of the expert, they represent these key relationships within the language framework. (4) Implementation: here the engineer makes this formalized knowledge compatible with information-flow. The resulting set of rules and associated control structure identify a prototype program capable of being executed then revised. (5) Testing and Maintenance: it involves judging the prototype program and refining/ modifying it to conform to the standards of excellence set by the expert. Besides, the knowledge database must be frequently updated to maintain timeliness and relevance. In fact, the usage of ESs guarantees numerous advantages. Indeed, ES is not a substitute for a knowledge worker’s overall performance of the problem-solving task, but it can dramatically minimize the amount of work the individual must do to fix a problem. Here are some possible organizational benefits of ESs [Sheikhtaheri et al., 2014, p. 110; Expert systems and applied artificial intelligence (2018)]: � ESs can complete their part of the task much faster than an expert, and so raising the speed of response. � The error rate of successful ESs is often much lower than human error for the same task, which enhances the efficiency and quality of work. � ESs can make consistent and reliable recommendations, and thus improving decisions by non-experts and shortening the whole process. � ESs can capture rare knowledge and maximize the use of scarce expertise, and thereby elevating profits while reducing costs. � When used as training vehicles, ESs result in a rapid learning curve for novices. Nevertheless, no technology can offer perfect solution. Large ESs are costly and require significant development time and resources. Therefore, there are various kinds of obstacles which would hinder organizations from developing successful ESs. Technical difficulties are always a problem, whereas non-technical barriers as well might pose an equally serious challenge. For instance, [Sheikhtaheri et al., 2014, p. 110; Chau, 1991; Artificial intelligence and expert systems: knowledge-based systems, (2018)]: � Problems with knowledge acquisition: maybe the employment of required experts is highly expensive, and sometimes they are not available, capable, or cooperative. In addition, knowledge transfer is often subject to biases and mistakes. Expert systems and administrative decisions 123 � Problems with knowledge engineers: help from knowledge engineers is costly and hard to obtain. In addition, knowledge engineers sometimes do not have much expertise in building business-related ESs. � Problems with management: sometimes management is not supportive enough or does not know how to deal with the organizational change arising from the operation of ESs. � Cognitive limits of users: in some cases, end-users do not prefer and trust the system or trust it too much. Hence, the term ES has been controversial. On the one hand, it creates high expectations and has been used as a buzzword for funding and a flag to wave for all types of projects. On the other hand, many people have criticized its feasibility (Liang, 1987, p. 14). In sum, successful ESs should have at least four indispensable features; selection of an appropriate problem sphere, realistic expectations, unhesitating commitment of top management, and comprehensive testing of the product (Jayaraman and Srivastava, 1996, p. 28). Eventually, it is worth mentioning that a primary idea in ESs technology is that problem solving could be generally accomplished through applying specific knowledge rather than specific techniques. This reflects the belief that human experts do not process their knowledge differently, but they do possess diverse knowledge. According to this philosophy, when ES does not produce the desired results, work must begin to expand the knowledge base, not to redesign procedures (Alasgarova and Muradkhanli, 2008, p. 299). 3.2 Administrative decisions In organizations, decisions are the signals of action and the portents to accomplishment or failure. Failure, in turn, marks the need for new decisions. Therefore, decisions and decision- making processes are best viewed as one major determinant of performance (Bozeman and Pandey, 2003, p. 2). Many researchers and practitioners believe that any organization faced with bad or negative decisions will be unproductive and continue to fail. In contrast, creative and innovative decision-making is of pivotal importance to the growth and success of all institutions (Ejimabo, 2015, p. 1). Undoubtedly, in a modern world, producing surplus of information and making administrative decisions have become more complicated. It is mainly about the fact that even the best and most efficient intelligence of decision-makers can be totally insufficient. Otherwise, it is necessary to exploit the array of decision-making tools which draw on such disciplines as psychology, sociology, economics, law, political science, computer science and others (Raczkowski, 2016, p. 33). For that reason, managerial decision-making is a notion that is seriously considered critical in the operations of the organization. It has gained the attention of scholars all over the world and it has almost several definitions. For example, some argued that decision- making is a decisive and deliberative social action concerned with selecting what to do in face of a problem, while others regarded it as a commitment to action or a concrete and discrete phenomenon driven by rationality. Thereby, administrative decision could be defined as a choice has been made from among two or more alternative objects or courses of action, given because of the advantages and disadvantages of supporting information about each (Ejimabo, 2015, p. 2). In that way, decision-making as a theory focuses exclusively on choice and the capability of decision-maker/leader to select the best option from the many available alternatives (Glaholt et al., 2010, p. 1). In other words, decision-making is deemed to be an integral part of every organization. Indeed, administrative decisions are made in organizations on a daily basis. Additionally, organizations are making decisions at all managerial levels. In this regard, there are three REPS 3,3/4 124 levels of management decision-making: strategic long-term management, tactical middle- range management and operational short-term management (Harvey and TIS, 2007, p. 3). Furthermore, there are three types of decision structure: unstructured decisions (related to the long-term strategy of the firm), semi-structured (some decision procedures can be pre- determined, but not enough to lead up to a certain decision) and structured decisions (the procedures can be specified in advance) (Management information systems and decision- making: an overview, 2018). On the other hand, much of the decision literature differentiates between technical and political aspects of decisions. (Bozeman and Pandey, 2003, pp. 5-6) stated that political decisions require more external players/actors and involve higher levels of conflict and a tendency to concentrate on ends rather than means. Similarly, they had expectations for technical decisions that they entail higher levels of economic rationality, technique and modeling are likely to be more paramount, and participants’ roles are mitigated by their technical status and specialization. According to Daft (1989), a rational/technical approach is idealistic under conditions of high goal consensus and high technical knowledge. On the contrary, situations of low goal consensus and less technical knowledge are better suited for political/non-technical decision operations. Perhaps, it is worth to think of decision content as a combination of technical and political content, with the pure technical content as a mooring at one end and pure political content at the other. Actually, managerial decision-making is inherently complicated. It is well known that effective administrative decisions are a result of a systematic process with obviously defined elements handled in a distinct sequence of steps. (Kryssanov et al. (2018), p. 4) presented a framework for the decision-making process. It is usually initiated by an external inquirer, and begins from gathering information to identifying a problem/task. Then, associated information is recognized and relevant/local reality models are chosen. Next, the models are evolved and possible alternatives/solutions are assessed. At the end, the best solution is selected for the final decision. After the decision is made, the new reality is considered and the process could be repeated under other circumstances. In this respect, managing implementation requires that the decision should be well communicated and the expected consequences are reflected in performance metrics. This guarantees that goals will be accomplished. Trial and error maybe allowed as tactical experiments within acceptable risk parameters, but duplicating past mistakes should be inexcusable. The outcome of previous decisions should be captured as a part of the corporate memory of the organization to ensure that lessons are learned (Harvey and TIS, 2007, p. 8). Generally speaking, the level of managerial decision-making can be promoted through providing proper decision support. The nature of this support depends on a number of variables, including the organizational context of the decision (public vs private sector), the quality of available information, the willingness of the decision maker to take advantage of such support (Dillon et al., 2010, p. 229), and of course the availability of resources, particularly technical ones. According to Nagy et al. (2011, p. 14), there are three indicators of good decisions; consistency, integration, and transparency. Whereas Bozeman and Pandey (2003, pp. 8-13) examined some familiar decision criteria for evaluation such as cost- effectiveness (doing more with less), fairness (sharing the pain), technical feasibility and usefulness (meeting the goal). This is in addition to the amount of time/timeframe required for the decision (the cost of taking so long to make decisions), its stability (permanence), participants (internal and external), the information quality (including accuracy and objectivity (intrinsic dimension), relevance and timeliness (contextual dimension), Expert systems and administrative decisions 125 interpretability and ease of understanding (representational dimension) and access (accessibility dimension)) (Dillon et al., 2010, p. 230). Finally, a crucial challenge here is to develop an appraisal system to guide the process of making and selecting administrative decisions, by consolidating the up-to-date technology available at an early phase of the decision-making process (Alasgarova and Muradkhanli, 2008, p. 297). The decision traits like speed, capacity, quality, desired output. . ., should be harmonized with the distinctive equipment attributes to arrive at best solutions. The next part of this research will discuss this point in some detail. 3.3 Expert systems and managerial decision-making in public organizations The debate on public performance has had major impact on public reforms in the late decades. Since the 1970s, a number of reforms have been adopted by governments; most of them aiming to reduce the size and expenses of bureaucracies, to improve accountability and transparency, to enhance efficiency and democracy, and to provide citizens greater access to services and better quality of life (Rauta, 2014, p. 58). Consequently, public management authors have regarded decision-making as a central focus for public performance reform. Managerial decision-making can be illustrated as a proposition considered by decision makers in the setting of the organization and its strategic status. Alternatives, risks, opportunities and potential outcomes are investigated, and thereby a decision is taken. Hertz (2013) highlighted what the decision-making challenges are in terms of faith, trust, and information excess; is that the decision maker sometimes has to take thousands of decisions a day, and doing this in an intelligent manner needs concentration and time (Raczkowski, 2016, p. 28). In addition, the decision-making process is mostly subject to human errors, as leaders have diverse personalities, prejudices, self-interest biases and varied attitudes towards risk (Harvey and TIS, 2007, p. 5). Hence, there is an urgency to think about how to make administrative decisions more properly and consistently. Some argued that many factors contribute in the complicated decision-making process, such as the development of new technologies, reducing access to financial resources, and limited capacity of decision makers (Ahmad and Munir, 2016, p. 3). In this respect, the complicacy of decisions justifies the utilization of ESs. ES as an advanced technology designed to provide information and uphold decision-making for various business functions, can ensure that decisions are evidence-based and that different probabilities are taken into account. Moreover, the use of ESs is a considerably cheaper, more accessible and rational way of solving problems in this arena (Dasic et al., 2011, p. 27). Actually, ESs are designed commonly for problems in which there is no single correct solution that can be encoded in a conventional algorithm. One would not use ES to find out shortest paths through graphs or sort data, as there are easier ways to perform these tasks. Simple systems use simple true/false logic and probability theory to assess data, but more sophisticated systems are capable of doing at least some evaluation considering real-world uncertainties, using methods as fuzzy logic (Sahin et al., 2012). Such sophistication is difficult to develop and yet comparatively imperfect in practice (Alasgarova and Muradkhanli, 2008, p. 297). In the health-care domain, for instance, ESs have shown many advantages, such as gaining and using rare expertise, more consistent medical decisions and shorter decision-making processes. Besides, ESs have assisted in making efficient and environmentally-sound agricultural decisions. They also have served as a training tool and in natural resources conservation, controlling, planning, marketing, and financial analysis (Baig et al., 2005, p. 208). Nevertheless, they have faced some challenges including problems with knowledge acquisition and data entry, wrong recommendations and responsibility, appraisal and maintenance of the system’s performance, aside from the REPS 3,3/4 126 limited scope of such systems and the need for integrating them into the routine work flow (Sheikhtaheri et al., 2014, p. 110). In spite of this, there are still a lot of direct benefits of applying ESs in organizations. ESs are capable of handling enormously-intricate tasks and functions, as well as an extremely-rich knowledge database content and structure. As such, they lessen production downtime and lift outputs and productivity. In addition, disseminating expert knowledge and advice, especially to remote and far locations using digital communications, enhances quality and reliability. As a result, the ongoing usage of ESs may be less expensive and more prosperous than assigning a human expert in specific situations (Expert systems, 2018). Furthermore, the use of ESs, particularly in strategic management, boosts the pace, efficiency, and consistency of decisions based on them. With the employment of ESs, there would be changes reflect in management hierarchy, where the authority at middle and top senior management levels is transferred to lower levels, as even convoluted problems are successfully-resolved within the scope of lower operations management with the help of ESs. Thus, extra time is left to middle and top managers which they use to solve other problems. Therefore, the functioning of the entire management system could be promoted to a higher level (Dasic et al., 2011, pp. 28-29). Otherwise, the organizational impact of ESs is deemed substantial. It may range from the establishment of additional departments (re-structuring), the modification in the communication system, in the degree of centralization/decentralization, in requirements and duties because of replacement-based ESs, and in the power of certain individuals and groups, to the improvement of the decision-making process and organizational effectiveness and efficiency of the whole entity (O’leary and Turban, 1987, p. 12). One relevant question remains here which is; “how will people react to the implementation of such advanced systems?” In fact, losing managerial control and offending employees are on the top of concerns that might be likely to surface. So that, despite the progress and wide-spread application of ESs, there is an alleged suspicion about how people will respond to this modern technology, both individuals using it as a decision aid and those affected by its decisions, especially in public sector organizations, taking into account the cognitive limits within the contrastive organizational culture of these institutions. This emphasizes the possibility that knowledge transfer maybe subject to some kind of interception or biases on the part of experts, or the potential for lack of trust or over-trust on the part of end-users. In this regard, O’leary and Turban (1987, pp. 12-13) stated that the influence of ESs on managerial decision-making depends on some variables such as the industry/sector in which the organization belongs to, the organizational climate, percentage of people affected by ESs, frequency of using ESs, number of operating ESs, suitability of the tasks, software and computing environment. Concerning the sectoral impact, it is emphatic that private and public sectors generally have diverse environments. The private sector is more noticeably associated with market forces, while the public sector is typically shaped by political considerations. One is about business and the other is about government. One tends to be decentralized and the other centralized. These different contexts imply unlike decision content (Dillon et al., 2010, p. 231). It needs to be realized that decisions made by public decision makers are indeed social decisions of social choice theory. In addition, it needs to be well recognized that decisions in the public sector are very often made on the basis of conflict between accepted and important values, thus reaching a decision-making compromise here is not easy (Raczkowski, 2016, p. 31). Whereas in public management like in commercial one, the decision-maker must learn how to accept chaos existing in modern global economy. This means decisions taken today maybe totally different tomorrow, both much better or worse (Hertz, 2013). A famous author Expert systems and administrative decisions 127 compared public and private sector decision-making using the metrics of analysis and bargaining, and concluded that private sector managers are more supportive of analysis- based decisions, while public executives are more supportive of bargaining-based decisions. Analyzing the two sectors separately suggests that public sector decision-making is typically more open, transparent and structured, while private sector decision-making is generally more ad-hoc and occurs more proactively. Contextual influences and constraints play significant roles in structuring public sector decisions. Because of the restricted nature of decisions made in this environment, human behavioral aspects have much less impact (Dillon et al., 2010, p. 233-234). Thereby, in a structured domain exactly as in public sectors, where qualitative reasoning is crucial to problem solving and expertise is quite available, developing an ES to bolster repetitive decisions, in particular, maybe convenient (Liang, 1987, p. 7). Here, ESs can reduce the workload of a public officer and allow him to take care of certain severe cases. Moreover, with the decreased impact of human factor, administrative decisions based on ESs are standardized and objective to a great extent. For that reason, the current research intends to continually examine the applicability of ESs in public organizations, with specific reference to Egypt. As the reviewed literature has shown that ESs are not completely unified and as such scholars tended to focus on several determinants to analyze (e.g. components, tasks, stages) depending on the theoretical perspective. Then, as long as this paper adopted the systems approach in management (open system) and claimed that ES is a system in the first place, so it could be best identified through its four basic dimensions: inputs, processes, outputs and feedback. Besides, the quality of administrative decisions is measured here through their major indicators such as consistency, integration, transparency, timeliness. . ., which were extracted from previous research (Nagy et al., 2011; Bozeman and Pandey, 2003). On the other hand, the present study believes that the impact of ESs on the performance of administrative decisions often relies on the various organizational factors such as leadership support, organizational culture and resources availability, as it was confirmed before by some researchers. Eventually, for successful implementation ES decision must be regarded as a good advice which can, but does not have to be fully approved by the decision maker (Dasic et al., 2011, p. 29). This explains the methodology adopted in this research and the necessity of distinguishing between ES outputs (recommended decisions) and the actual administrative decisions taken by leaders. In the next part, the study will test the effect of ESs on the efficiency of managerial decision-making at a critical Egyptian public agency, which is the MIIC; is that the role of technology here is fundamental because of the significance, accuracy, and promptness of the output desired to offer quality services to the investors in the Egyptian market. 4. Application: empirical study discussion 4.1 Methodology: research design and techniques The design of the study is descriptive in the theoretical part and quantitative in the empirical one. Theoretically, the research jointly adopted the analytical approach and the systems approach to demonstrate main concepts and to determine the relationship between variables as well, while it used an applied study to investigate correlations in practice. Hence, in addition to providing a concise overview of relevant literature, a field survey was conducted during the first two weeks of May 2018 to capture the executives’ attitudes towards developing and applying ESs, and their impact on the quality and effectiveness of administrative decisions at the MIIC. This entity was opted as the case study here because it is considered one of the crucial Egyptian economic ministries that have launched many substantial reforms, particularly REPS 3,3/4 128 technical ones, during the past years to enhance work processes and provide access to improved public services to all citizens (MIIC official website, 2018). Furthermore, after conducting a number of interviews with some prominent governmental executives inside and outside MIIC, it was noticed that this type of systems (ES), which is not widespread in the Egyptian public sector, is found in some way in this case although it is not yet complete/ perfect, as they are still working on some planned developments to allow for the full utilization of its capabilities in the near future. Therefore, it is necessary to explore the current situation and to evaluate the efficiency of existing ESs and how their effectiveness could be boosted to promote the quality of administrative decisions at the ministry. This might be regarded as a kind of process evaluation which takes place while implementing a project to discover the aspects that need to be changed during delivery. Its goal is to find ways of improving the program while it is realizing. For the purpose of gathering the required primary data, an Arabic-language structured questionnaire was adopted, as Arabic is the official language in Egypt (it was initially written in English and then translated). The questionnaire was formulated on the basis of literature review (Nagy et al., 2011; Bolfikova et al., 2010; Faisal and Hanzal, 2009). It encompasses 52 items other than demographic data and consists of three sections: determinants of ESs (inputs-processes-outputs-feedback) (independent variable), organizational factors (moderating variable), and administrative decisions criteria (dependent variable) (Appendix). Here, it is worth mentioning that the study utilized the questionnaire to collect data using a five-point Likert scale as the measurement tool, ranging from 1 = strongly disagree to 5 = strongly agree. Finally, statistical package for social survey (SPSS-V.18) and Analysis of MOment Structures were both the tools for compiling and processing data in this research. Multiple statistical tools were also used for data analysis, which are descriptive analysis, Pearson correlation coefficient, simple linear regression and structural equation modeling. Moreover, Cronbach’s alpha test was applied to assess the reliability of the research measures, and it was noticed that all coefficients are above 0.50, so there is evidence that the variables seem to be stable, consistent, reliable, and valid. 4.2 Demographic and professional data of respondents The whole population was around (120) people who perform as managers at MIIC (headquarter/main premises in Cairo). A manageable and convenient sample size of (52) people was examined. This sample was randomly selected. Knowing that 60 people (50 per cent of the population) were sampled and 52 only responded, thus the response rate was 87 per cent. Empirical results indicate that the majority of 73.1 per cent of the sample are males while 26.9 per cent are females, 86.6 per cent their ages are less than 50 years and 84.6 per cent of the sample are post graduates, whereas 15.4 per cent have only a BSc degree. In addition, the majority of 78.8 per cent of respondents are department managers while 21.2 per cent are considered upper level management, and eventually 53.9 per cent have spent at least 10 years working for MIIC. 4.3 Building indicators of the research variables Statistical technique was used to combine each group of related questions in one indicator. It is important to mention that seven indicators were already created: inputs, processes, outputs, feedback, organizational factors, administrative decisions and ESs, as shown in Table I which presents the descriptive statistics of them. The first six indicators were composed by using equal weights method, via adding the scores of the questions which are Expert systems and administrative decisions 129 related to each indicator and then dividing the sum by the number of those questions. However, the last indicator (ESs) was calculated through the following formula (0.4 � inputs þ 0.3 � processes þ 0.2 � outputs þ 0.1 � feedback) to take into consideration the hierarchy of these sub-items. From Table I, it is obvious that the values of the mean for all indicators are around 3 and 4 (in Likert scale), which clarifies that the respondents tended to be neutral or agreed to the questions that measured those indicators in common. It is clear also that the highest mean value is for the one of processes (4.00), which means that the presence of this indicator was good at MIIC. On the other hand, the least mean value is the one of administrative decisions (3.46), which means the level of this indicator was average. Table II indicates the relation between the whole indicator of ESs and its sub-variables using Pearson correlation coefficient. As shown in this table, there is a significant (p-value is less than 0.05) positive relationship at significance level a = 0.05 (with confidence level 95 per cent) between ESs whole indicator and each sub-variable. Knowing that the highly correlated indicator is inputs (0.838) and the least is feedback (0.488), which is logical in light of their hierarchy and reflects their influence on ESs at MIIC. In addition, this confirms the fact that feedback is oftentimes neglected in reality. 4.4 Testing the research hypotheses To accomplish the empirical goals, the research set out the following hypotheses: 4.4.1 First hypothesis H01. “There is no significant relationship at significance level a = 0.05 between ESs and the quality of administrative decisions at MIIC.” H01.1. “There is no significant relationship at significance level a = 0.05 between inputs and the quality of administrative decisions.” H01.2. “There is no significant relationship at significance level a = 0.05 between processes and the quality of administrative decisions.” Table I. Descriptive statistics of the research indicators Indicator Minimum Maximum Mean SD Inputs 3 5 3.94 0.59 Processes 3.2 5 4 0.42 Outputs 2.79 4.36 3.63 0.34 Feedback 2.33 4.67 3.92 0.44 Organizational factors 3 5 3.94 0.55 Administrative decisions 2.69 4.56 3.46 0.59 ESs 3.35 4.67 3.89 0.34 Table II. Pearson coefficient of ESs sub-variables Sub-variable Inputs Processes Outputs Feedback ESs Pearson coefficient 0.838 0.569 0.588 0.488 p-value 0 0 0 0 REPS 3,3/4 130 H01.3. “There is no significant relationship at significance level a = 0.05 between outputs and the quality of administrative decisions.” H01.4. “There is no significant relationship at significance level a = 0.05 between feedback and the quality of administrative decisions.” To prove whether the previous hypotheses are acceptable or not, simple linear regression was used to test the total effect of ESs, along with the influence of each component separately on the quality of administrative decisions, as presented in Table III. Note that the hypothesis will be denied if the significance of the model is less than 0.05, and vice versa. Table III indicates that: � For the first model, when inputs indicator is the independent variable: It is obvious that there is a significant positive relationship between inputs and administrative decisions at confidence level 95 per cent, and this appears from the value of beta. From adjusted R-squared, it is noticed that inputs have the ability to explain about 35.7 per cent from the variation in administrative decisions at MIIC. � For the second model, when processes indicator is the independent variable: It is clear that there is a significant positive relationship between processes and administrative decisions, and that processes have the ability to explain about 8.3 per cent only from the variation in decisions. � For the third model, when outputs indicator is the independent variable: It is obvious that there is a significant positive relationship between outputs and administrative decisions, and that outputs have the ability to explain about 35.2 per cent from the variation in decisions. � For the fourth model, when feedback is the independent variable: It is clear that there is no significant relationship between feedback and administrative decisions cause the significance of this model is greater than 0.05, which means that feedback does not affect the quality of decisions at all. This is compatible with the result that feedback is the least-correlated sub-indicator with ESs and it is usually ignored. � For the fifth model, when the independent variable is the overall ESs indicator: It is evident that there is a significant positive relationship between ESs and the quality of administrative decisions at confidence level 95 per cent, and that ESs have the ability to explain about 43.2 per cent from the variation in decisions. From the previous results, the study can conclude that there is a significant relationship at significance level a = 0.05 between ESs and the quality of administrative decisions at MIIC; is that it was proven that inputs, processes, and outputs of ESs have significant positive Table III. Simple linear regression models of the dependent variable on the different independent variables Simple regression model Dependent variable Independent variable Beta Significance of the model Adjusted R-squared First Administrative decisions Inputs 0.605 0 0.357 Second Processes 0.449 0.023 0.083 Third Outputs 1.053 0 0.352 Fourth Feedback 0.203 0.324 0.02 Fifth ESs 1.382 0 0.432 Expert systems and administrative decisions 131 effects on administrative decisions at MIIC, whereas feedback has insignificant effect on the quality of these decisions, which means that the first hypothesis as a whole and its first three sub-hypotheses as well are all rejected, while merely the fourth sub-hypothesis is accepted. 4.4.2 Second hypothesis H02. “There is no significant relationship at significance level a = 0.05 between ESs and the quality of administrative decisions at MIIC, putting into consideration the organizational factors.” To show if the previous hypothesis is acceptable or not, a structural equations model, illustrated in Table IV, was used to test whether the moderating variable has significant impact on the relationship between the independent and dependent variables or not. Knowing that the moderating variable will have significant influence on this relation if there is a significant effect of ESs on organizational factors, as well as a significant effect of organizational factors on administrative decisions. In that way, Table IV clarifies that: � The organizational factors have significant effect on the relationship between ESs and the quality of administrative decisions at MIIC at confidence level 95 per cent, as their p-values are less than 0.05 in both ways. � The direct effect of ESs on administrative decisions = 0.663, while the indirect effect through organizational factors = 1.164 � 0.611 = 0.711, which means that when taking into account the organizational factors, the effect of ESs on decisions increases from 0.663 to 0.711 or the relation becomes stronger. From the previous results, the research can conclude that there is a significant relationship at significance level a = 0.05 between ESs and the quality of administrative decisions at MIIC, taking into consideration the organizational factors, so that H02 is also rejected. 5. Conclusion: concluding remarks ESs are gaining widespread admission in the business world today. Thousands of cases have proven that ESs are invaluable decision-making tools. However, managers still need to learn how to overcome obstacles of successful implementation. Therefore, it seems that there is much to be learned about the potential use of ESs in the realm of public sectors. This research has productive implications for both theory and practice; first it offers multiple contributions to literature, as it expands on the existing knowledge by formulating a conceptual framework identifies the desired role of ESs in improving managerial decision- making, specifically in public organizations, and second the paper provides a practical directory to executives and policy makers, particularly in the Egyptian government, to help them enhance their choices using modern technology, aside from providing an educational material for MIIC staff and a catalyst to support the application of ESs there. Nevertheless, the empirical study here has few limitations. All results are based on information taken from Table IV. Results of structural equations model Dependent variable Independent variable Estimate S.E. C.R. p-value Organizational factors / ESs 1.164 0.218 5.345 0 Administrative decisions / Organizational factors 0.611 0.111 5.513 0 Administrative decisions / ESs 0.663 0.215 3.08 0.002 REPS 3,3/4 132 respondents. This is in addition to time limitations, i.e. findings represent the Egyptian economy within a period of great challenges and transformations in Egypt. Through the theoretical and applied parts of this research together, it has reached the next results and recommendations: 5.1 Results and findings analysis In general, this research claims that ESs do not aim to fully replace human experts in problem solving activities; however, they can serve as an extremely-useful expert advisor for numerous management tasks and issues. In that way, ESs are earnest assistants to public decision-makers and not substitutes for them. In this respect, the main results here are: � The study confirms that ES is one of the most commercially fruitful branches of AI. ESs are more recommended when the problem field is structured, the decision is repetitive, and the knowledge involves qualitative reasoning. Besides, there are several functional categories suitable for this technology which may include interpretation, prediction, diagnosis, design, planning, monitoring and control. � The paper highlights that managerial decision-making is one of the most challenging, dynamic, and ongoing tracks in every organization. Nowadays, there is a great need to modify the problem solving approaches among organizational leaders while accommodating technology and globalization. Hence, ESs can provide great potentials for automatization and modernization of this act. � Findings indicate that because of the structured, complicated, and restricted nature which is an obvious trait of problems within public management, it is possible to resolve them with the help of ESs. For that reason, properly designed and implemented ESs could be a powerful addition to the ever-expanding tool box of public executives. � Practical implications clarify that the presence of most of the indicators of ESs, administrative decisions, and organizational factors at MIIC is above average, which could be normal and expected in light of the fact that the ministry is still working on developing its systems to allow for the full utilization of their capabilities in future decision-making actions. This means there is yet an urgent need to support and strengthen these concepts in the work context of MIIC. Otherwise, it was noticed that individuals who had the opportunity to use ESs did not feel highly threatened by this decision instrument; instead, they thought it was a time-saver. � The results of this analysis provide evidence that existing ESs have contributed to the decision-making process at MIIC (by 43.2 per cent), as demonstrated by the positive relationship between the independent and dependent variables; is that it was proven that inputs, processes and outputs are positively related to the quality of administrative decisions, while only feedback does not affect decisions at all. This refers to the quite-weak interdependence relation between ESs dimensions that might impact decisions, in addition to the relatively poor compatibility between ESs applications, as an advanced decision aid technology, and the practices and activities of decision-making at the ministry. Knowing that this could also reflect in somehow the moderating variable influence in this relation. In this regard, it was emphatic that the relationship between ESs and administrative decisions is moderated by organizational factors, which means when considering the organizational factors at MIIC; the effect of ESs on managerial decision-making becomes greater. Expert systems and administrative decisions 133 5.2 Recommendations and future research In light of what was aforesaid, the research has made these recommendations: � Problem solving is better achieved through applying specific knowledge rather than certain techniques. Accordingly, when ES does not produce the desired results, then work should begin to enrich the knowledge base, not to reprogram the procedures. � When recruiting, employers must consider the decision-making experience of nominees. At the same time, nominees ought to be fully aware of the decision- making context within which the organization operates. � Public managers need to clearly understand their essential part and substantial role in the ES development project. They should provide financial, informational, and psychological support to the project team. Moreover, hiring consultants to help implement the plan is recommended if management is inexpert in handling organizational changes, along with providing adequate technical training to assist in overcoming barriers. � The empirical results suggest that there must be more genuine efforts to strengthen, integrate, and achieve consistency between ESs dimensions and applications on one hand, and between them and the managerial decision-making at MIIC on the other hand, which for sure will impact positively the quality and effectiveness of administrative decisions. In this respect, giving greater attention to the feedback that was evidenced it did not affect decisions completely, and at the same time, it was the least correlated indicator with ESs at the ministry, and linking this feedback together with inputs in a continuous cycle, or in other words taking advantage of feedback reviews in addressing current problems and promoting the level of inputs and thereby the other dimensions, all of this would undoubtedly boost the efficiency of ESs and thus the quality of the overall decision-making process at MIIC. Also, to foster the effect of the organizational factors that positively moderated the relationship between ESs and administrative decisions at MIIC, the researcher here recommends to place greater emphasis on these elements consecutively; the importance of leadership willingness and commitment to feeding and updating the ministry’s knowledge bases with the latest data, which was proven that it had no impact among different organizational factors there. As well as, the need for allocating sufficient funds and recourses that had the least influence among organizational factors, and then the need for high-level technical assistance from inside and outside the ministry, in addition to encouraging creativity, transparency and empowerment in the existing work climate. � Concerning future research, an obvious sequel to this study is to conduct longitudinal research to reveal over time if the ES feedback reviews reported at one point are associated with better inputs and outputs at a later point. This opens the way to replicate the framework of this research again at MIIC, but after a reasonable period which may give them the opportunity and time to complete the development of ESs and eliminate current limitations, to allow for fair evaluation of the efficiency of operations and satisfaction of service recipients. In addition, it is necessary to find and assess more applications of ESs in different service and manufacturing areas within both private and public sectors, and to test the various organizational antecedents and implications of using such advanced technology in these institutions, especially with its high requirements. Besides, further studies should also consider other moderating variables, such as organizational structure, system software, and computing environment. Finally, future work examining international rapprochement is urgently requested, particularly in close cultural contexts. REPS 3,3/4 134 References Ahmad, A. and Munir, Z.B. (2016), “Improving quality of healthcare through the selection of right technology: an expert system approach for a public sector hospital”, paper presented at the 2nd International Conference on Quality Engineering and Management, 13-15 July, Portugal, available at: www.researchgate.net/publication/306015346_Improving_Quality_of_Healthcare_ through_the_Selection_of_Right_Technology_An_Expert_System_Approach_for_a_Public_ Sector_Hospital (accessed 10 April 2018). Alasgarova, A. and Muradkhanli, L. (2008), “Expert system for decision-making problem in economics”, International Journal “Information Technologies and Knowledge, Vol. 2, pp. 297-299. Arnott, D. and Pervan, G. (2008), “Eight key issues for the decision support systems discipline”, Decision Support Systems, Vol. 44 No. 3, pp. 657-672. Artificial intelligence and expert systems: knowledge-based systems ( (2018), ) “Artificial intelligence and expert systems: knowledge-based systems”, available at: http://vfu.bg/en/e-Learning/ Artificial-Intelligence–AI_and_ES_Nowledge_base_systems.pdf (accessed 28 February 2018). Baig, F., Nawaz, N. and Saif, u-r. (2005), “Expert systems for decision making in agriculture sector”, Journal of Agriculture and Social Sciences, Vol. 1 No. 2, pp. 208-211. Bolfikova, E., Hrehova, D. and Frenova, J. (2010), “Manager’s decision-making in organizations – empirical analysis of bureaucratic vs. learning approach”, Journal of Economics and Business (ZB RAD EKON FAK RIJE), Vol. 28 No. 1, pp. 135-163. Bozeman, B. and Pandey, S.K. (2003), “Public management decision-making: technical vs. political decisions”, paper prepared for the National Public Management Research Conference, 9-11 October, Georgetown University, Washington, DC, D.C., available at: https://pdfs. semanticscholar.org/78ac/8ffa6479e0771a6b5ffddbdf6e209988bc4c.pdf (accessed 5 April 2018). Chau, P.Y.K. (1991), “Better decision making through expert systems for management”, SAM Advanced Management Journal, Vol. 56 No. 4. Daft, R.L. (1989), Organization Theory and Design, West Publishing Company, New York, NY. Dasic, M., Trajkovic, S. and Tesanovic, B. (2011), “The necessity of using expert systems in strategic decision making”, International Journal of Economics and Law, Vol. 1 No. 1, pp. 27-35. Dillon, S., Buchanan, J. and Corner, J. (2010), “Comparing public and private sector decision making: problem structuring and information quality issues”, proceedings of the 45th Annual Conference of the ORSNZ, November, pp. 229-237, available at: http://citeseerx.ist.psu.edu/viewdoc/ download?doi=10.1.1.455.9291&rep=rep1&type=pdf (accessed 1 April 2018). Ejimabo, N.O. (2015), “The influence of decision making in organizational leadership and management activities”, Journal of Entrepreneurship and Organization Management, Vol. 4 No. 2, pp. 1-13. Expert systems and applied artificial intelligence ( (2018), ) “Expert systems and applied artificial intelligence”, available at: www.umsl.edu/�joshik/msis480/chapt11.htm (accessed 27 January 2018). Expert systems ( (2018), ) “Expert systems”, available at: www.referenceforbusiness.com/encyclopedia/ Ent-Fac/Expert-Systems.html (accessed 7 February 2018). Faisal, G. and Hanzal, Q. (2009), “The effect of the elements of management information systems in the effectiveness of taking administrative decisions”, Tikrit Journal for Administrative and Economics Sciences, Vol. 5 No. 13, pp. 80-98. Glaholt, M.G., Wu, M.C. and Reingold, E.M. (2010), “Evidence for top-down control of eye movements during visual decision making”, Journal of Vision, (15, May), Vol. 10 No. 5, pp. 1-10. Harvey, Jasmin and Technical Information Service (TIS) (2007), “Effective decision making”, Topic Gateway Series, Vol. 40, The Chartered Institute of Management Accountants (CIMA), London, December. Hertz, N. (2013), Eyes Wide Open: How to Make Smart Decisions in a Confusing World, HarperCollins, UK. Expert systems and administrative decisions 135 http://www.researchgate.net/publication/306015346_Improving_Quality_of_Healthcare_through_the_Selection_of_Right_Technology_An_Expert_System_Approach_for_a_Public_Sector_Hospital http://www.researchgate.net/publication/306015346_Improving_Quality_of_Healthcare_through_the_Selection_of_Right_Technology_An_Expert_System_Approach_for_a_Public_Sector_Hospital http://www.researchgate.net/publication/306015346_Improving_Quality_of_Healthcare_through_the_Selection_of_Right_Technology_An_Expert_System_Approach_for_a_Public_Sector_Hospital http://vfu.bg/en/e-Learning/Artificial-Intelligence&hx2013;AI_and_ES_Nowledge_base_systems.pdf http://vfu.bg/en/e-Learning/Artificial-Intelligence&hx2013;AI_and_ES_Nowledge_base_systems.pdf https://pdfs.semanticscholar.org/78ac/8ffa6479e0771a6b5ffddbdf6e209988bc4c.pdf https://pdfs.semanticscholar.org/78ac/8ffa6479e0771a6b5ffddbdf6e209988bc4c.pdf http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.455.9291&rep=rep1&type=pdf http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.455.9291&rep=rep1&type=pdf http://www.umsl.edu/&hx223C;joshik/msis480/chapt11.htm http://www.referenceforbusiness.com/encyclopedia/Ent-Fac/Expert-Systems.html http://www.referenceforbusiness.com/encyclopedia/Ent-Fac/Expert-Systems.html Jaradat, A., Alajlouni, M. and Almashaqba, Z. (2009), “The role of management information systems in managerial decision making quality: a study in the housing bank for trade and finance”, Tishreen University Journal for Research and Scientific Studies – Economic and Legal Sciences Series, Vol. 31 No. 1, pp. 73-93. Jayaraman, V. and Srivastava, R. (1996), “Expert systems in production and operations management: current applications and future prospects”, International Journal of Operations and Production Management, Vol. 16 No. 12, pp. 27-44. Kryssanov, V.V. Abramov, V.A. Fukuda, Y. and Konishi, K. (2018), “A decision-making support system based on know-how”, available at: https://arxiv.org/ftp/cs/papers/0606/0606010.pdf (accessed 3 March 2018). Liang, T.P. (1987), “Expert systems as decision aids: issues and strategies”, working paper no. 1378, Bureau of Economic and Business Research (BEBR), College of Commerce and Business Administration, University of Illinois, July. Liao, S.H. (2005), “Expert system methodologies and applications – a decade review from 1995 to 2004”, Expert Systems with Applications, Vol. 28 No. 1, pp. 93-103. Management information systems and decision-making: an overview ( (2018), ) “Management information systems and decision-making: an overview”, available at: https://onlinecampus.bu. edu/bbcswebdav/pid-843933-dt-content-rid-2221759_1/courses/13sprgmetad715_ol/module_03a/ metad715_m03l02t02_managementinfosystems.html (accessed 1 February 2018). MIIC official website ( (2018), ), available at: www.miic.gov.eg/english/pages/default.aspx (accessed 15 April 2018). Nagy, O. El-Deik, M. and El-Qet, A. (2011), “The impact of management information systems on the quality of administrative decisions at jawwal company – palestine”, available at: https://eco. najah.edu/sites/eco.najah.edu/files/%20 20%ىلع20%ةيرادإلا20%تامولعملا20%مظن20%ريثأت ةيرادإلا20%تارارقلا20%ةدوج .pdf (accessed 16 March 2018). O’leary, D.E. and Turban, E. (1987), “The organizational impact of expert systems”, Human Systems Management, Vol. 7 No. 1, pp. 11-19. Raczkowski, K. (2016), Public Management: Theory and Practice, Springer International Publishing, Switzerland. Rauta, E. (2014), “A decision-making model for public management: the existence of a policy framework for performance in Romania”, International Review of Social Research (IRSR), (February), Vol. 4 No. 1, pp. 57-74. Sahin, S., Tolun, M.R. and Hassanpour, R. (2012), “Hybrid expert systems: a survey of current approaches and applications”, Expert Systems with Applications, (March), Vol. 39 No. 4, pp. 4609-4617. Sheikhtaheri, A., Sadoughi, F. and Dehaghi, Z.H. (2014), “Developing and using expert systems and neural networks in medicine: a review on benefits and challenges”, Journal of Medical Systems, (September), Vol. 38 No. 9, pp. 110. Shim, J.P., Warkentin, M., Courtney, J.F., Power, D.J., Sharda, R. and Carlsson, C. (2002), “Past, present, and future of decision support technology”, Decision Support Systems, Vol. 931, pp. 1-16. Appendix. Questionnaire statements A. ESs at MIIC Inputs: 1. ES utilizes smart/modern technology. 2. ES requires qualified staff. 3. Adequate financial resources are needed. REPS 3,3/4 136 https://arxiv.org/ftp/cs/papers/0606/0606010.pdf https://onlinecampus.bu.edu/bbcswebdav/pid-843933-dt-content-rid-2221759_1/courses/13sprgmetad715_ol/module_03a/metad715_m03l02t02_managementinfosystems.html https://onlinecampus.bu.edu/bbcswebdav/pid-843933-dt-content-rid-2221759_1/courses/13sprgmetad715_ol/module_03a/metad715_m03l02t02_managementinfosystems.html https://onlinecampus.bu.edu/bbcswebdav/pid-843933-dt-content-rid-2221759_1/courses/13sprgmetad715_ol/module_03a/metad715_m03l02t02_managementinfosystems.html http://www.miic.gov.eg/english/pages/default.aspx https://eco.najah.edu/sites/eco.najah.edu/files/&hx0025;20 https://eco.najah.edu/sites/eco.najah.edu/files/&hx0025;20 4. Vital/various data is stored. 5. Scientific principles/practical experience are stored. 6. ES is fed with updated information. Processes: 7. ES is used frequently. 8. ES performs numerous tasks. 9. Target is set up in line with ES purposes. 10. ES receives sufficient information to determine each case. 11. ES analyzes data and applies rules to address problems. Outputs: 12. ES provides clear/accurate professional recommendations. 13. ES provides practical recommendations. 14. ES provides timely recommendations. 15. Recommendations look for proven solutions. 16. Recommendations look for novel solutions. 17. Recommendations suggest best solutions. 18. Recommendations justify chosen alternatives. 19. Recommendations set implementation plans. 20. Recommendations are usually applied. 21. ES outcomes support executives’ decisions. 22. Outcomes support different stages of decision process. 23. Outcomes support a wide variety of decisions. 24. Outcomes affect decision structure. 25. Outcomes impact a broad range of citizens. Feedback: 26. ES is simple/easily-used. 27. ES is flexible/responsive. 28. ES is comprehensive/integrated. 29. ES is reliable/trusted. 30. ES is efficient/effective. 31. Outcomes could be improved by developing the technological structure/knowledge base. B. Organizational characteristics 32. MIIC assigns technical experts to design/implement/modify ES. 33. MIIC has a separate IT support unit. 34. MIIC has sufficient funds for developing ES and training employees. 35. Administrative leaders facilitate the implementation of ES and the enrichment of databases. 36. MIIC has an open/decentralized organizational culture. C. Administrative decisions at MIIC 37. Decisions are in line with public policies objectives of the state. 38. Decisions are in line with environmental changes. 39. Decisions are timely. 40. Decisions are announced. 41. Employees are consulted in relevant decisions. 42. Decisions provide appropriate solutions for problems. Expert systems and administrative decisions 137 43. Decisions deal with different aspects of problems. 44. Reasons behind decisions are often obvious. 45. Available alternatives are usually clear. 46. General/collective interest is given priority. 47. Decisions are consistent. 48. Decisions are complementary. 49. Decisions are instantly executed. 50. Decisions are easily followed up. 51. Decisions are periodically assessed. 52. ES is utilized to promote decision-making within MIIC. Corresponding author Marwa Gaber Ahmed Fahim can be contacted at: marwa_ecma@yahoo.com For instructions on how to order reprints of this article, please visit our website: www.emeraldgrouppublishing.com/licensing/reprints.htm Or contact us for further details: permissions@emeraldinsight.com REPS 3,3/4 138 mailto:marwa_ecma@yahoo.com Improving administrative decisions through expert systems: empirical analysis 1. Introduction 2. Problem statement 3. Literature review: concepts and relationships description 3.1 Expert systems 3.2 Administrative decisions 3.3 Expert systems and managerial decision-making in public organizations 4. Application: empirical study discussion 4.1 Methodology: research design and techniques 4.2 Demographic and professional data of respondents 4.3 Building indicators of the research variables 4.4 Testing the research hypotheses Undefined namespace prefix xmlXPathCompOpEval: parameter error xmlXPathEval: evaluation failed Undefined namespace prefix xmlXPathCompOpEval: parameter error xmlXPathEval: evaluation failed 5. Conclusion: concluding remarks 5.1 Results and findings analysis 5.2 Recommendations and future research References A. ESs at MIIC 
work_jgpukl2a5fg3xn2gvz37oonyoi	----	doi:10.1016/j.eswa.2006.09.014 www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 551–563 Expert Systems with Applications An expert system for dynamic re-coordination of distributed workflows q William L. Kuechler Jr. a,*, Vijay K. Vaishnavi b a Department of Accounting and Information Systems, University of Nevada at Reno, United States b Department of Computer Information Systems, Georgia State University, United States Abstract A persistent problem in the use of automated workflow management systems for inter-organizational workflows has been the need for manual redefinition of coordination points in the process models when either organization changes its processes. Coordinating commu- nications between production-chain organizations are usually based on the notification of completion of tasks by one party which con- stitute pre-conditions for activity in another organization; autonomous task changes disrupt coordination. This paper describes a workflow re-coordination model and corresponding expert support system based on workflow goals which are more stable than the low-level machine, role and technique dependent activities which implement them. Following development of the model, we describe a set of three elementary disruption cases which span a large number of common workflow changes. Using these cases we demonstrate that common coordination disruption situations can be totally or partially repaired by use of an expert coordination subsystem. � 2006 Elsevier Ltd. All rights reserved. Keywords: Workflow management system; Coordination; Goal-based process modeling 1. Introduction Workflow management is the study and analysis of busi- ness processes with the explicit intent of interfacing, merg- ing, expediting and/or redesigning those processes, when they are considered in their full context. Thus, workflow management differs from traditional industrial engineering process studies primarily in scope. Workflow management intends to cross divisional and traditional single-output- process boundaries in the search for global (organizational) process effectiveness. Workflow management includes anal- ysis, description, design, and augmentation of workflows (Georgakopoulos, Sheth, & Hornick, 1994). As the concep- tion of the enterprise becomes global and virtual (Morita, 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.09.014 q This work is partially supported by NSF Research Grants, IIS- 9810901 and IIS-9811248, and a research grant to the second author from the Robinson College of Business, Georgia State University. * Corresponding author. Tel.: +1 775 784 6910. E-mail addresses: kuechler@unr.edu (W.L. Kuechler Jr.), vvaishna@ gsu.edu (V.K. Vaishnavi). Mukaigaito, & Hayami, 1996; O’Leary, Kuokka, & Plant, 1997) the challenges of semantic ambiguity and dynamic task change and substitution are introduced into inter- organizational workflow management by the necessary autonomy of cooperating organizations (Vaishnavi & Kuechler, 2005). Most existing WFMS are based on ‘manufacturing models’ of work where precise specification of activities and their execution times are (presumed) critical. Yet many work processes, especially knowledge (office) work, neces- sarily incorporate slack time to handle indeterminacy and frequently specify desired results intentionally, that is the process goal(s) are significant, but the details of execution are not (Davenport, Jarvenpaa, & Beers, 1996; Hewitt, 1991). As organizations become more virtual and an increasing number of non-strategic functions are subcon- tracted, traditional production workflows take on the same character of being more readily specified by intention than by concrete task. Fig. 1 shows a macro view of the business environment. Processes in any organization (X, Y or Z) are dependent on mailto:kuechler@unr.edu mailto:vvaishna@ gsu.edu mailto:vvaishna@ gsu.edu Organization X Organization Y Organization Z processes process communications Client Process X Process Request to Y Process Reply from Y Server Process Y Process initiated by X Reply to X (1) Process Reply from Y Reply to X (2) Business Process Communication Client/ Server Process Interaction triggers Fig. 1. Cooperation through Process interaction (adapted from Morita et al., 1996). 552 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 information from concurrent processes in their trading partners. The client–server portion of the diagram illus- trates this dependency. Though the process interactions diagrammed in Fig. 1 may be manual or automated, our research is concerned with situations where the processes are controlled by automated workflow management systems (WFMS), and the process communications are between interoperating WFMS. Each WFMS operates from a model of a work process. A trigger (Joosten, 1994) in a WFMS signals the satisfaction of a set of precon- ditions that cause the enactment engine to perform (or schedule or assign resources to) a pending activity in a sequential process. Recent work by the process modeling and workflow management communities defines the coordination of tasks across semantically heterogeneous environments as a sig- nificant problem area (Andersson, Bider, Johannesson, & Perjons, 2005; Buhler & Vidal, 2005; Khomyakov & Bider, Verify education Check FBI for criminal record Check for in-state criminal record preliminary_ clearance_ complete Duration Activity (original) 1 day (phone) 3 days (fax) 2 weeks Durati 1 day (pho 3 weeks 2 monthsTrigger Fig. 2. WFMS trigger de 2001). We directly address some of the semantic issues in task interpretation that have been identified by these authors and others, for example (Agostini & De Michelis, 2000; Klein, Dellarocas, & Bernstein, 2000; van der Aalst & Jablonski, 2000), as not adequately addressed by prior WFMS and process research. The problem of coordination disruption through auton- omous activity changes in work processes is illustrated in Fig. 2. The figure shows a highly simplified definition of a security clearance process (Vaishnavi, Joosten, & Kuechler, 1996) in which a trigger, preliminary_clearance_complete in a process is described in terms of the completion of activity: check for in state criminal record. The trigger results in train- ing for a new hire in parallel with the conclusion of the secu- rity clearance and having the training begin prior to a complete clearance saves significant time. When the process is redefined, as would likely be the case if the process were shifted to a subcontracting site, the trigger no longer exists Verify education Check interpol for international criminal record Military intelligence investigation for national criminal activity on Activity (redefined) ne) Lost coordination ! The activity defining the trigger has been eliminated. ? Trigger fined as process-state. W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 553 and training is delayed until coordination can be reestab- lished between the contractor and subcontractor. Traditional work process modeling methods such as SADT and trigger modeling (Joosten, 1994) assume a deterministic, fixed element model (Lei & Singh, 1997; Yu, 1995). The system is closed, that is, the activities that make up the various processes are predefined. As indicated above, this assumption is almost certain to be violated when autonomous WFMS seek to interoperate. Autono- mous groups, especially in outsourcing situations, must be free to change process details. Thus we seek some more robust method of specifying trigger conditions than match- ing to fixed patterns of activities. Our prior research in this area (Vaishnavi & Kuechler, 2005; Kuechler, Vaishnavi, & Kuechler, 2001; Vaishnavi et al., 1996) has addressed the issues at a conceptual level and has developed formal models of the contractor–sub- contractor relationship based on the workflow manage- ment consortium (WFMC) workflow enactment model. The research reported in this paper builds on that research to articulate a computable goal-based coordination model to implement the formal models and develops the comput- able model into a working prototype. In the next section of the paper (Section 2) we survey the work of other researchers to introduce flexibility into pro- cess definitions. Section 3 develops the goal-based work- flow coordination model and Section 4 presents the architecture of the corresponding expert system. Section 5 discusses the expert system implementation, introduces three exemplars of coordination disruption as test cases and describes the system response to these cases. In a con- cluding section we discuss some of the ways in which the system and model can be augmented, and our on-going research with the technique. 2. Search for flexibility in workflow and process modeling In our research on coordinating inter-organizational WFMS, we have surveyed literature from multiple fields for approaches to similar problems and alternative app- roaches to coordinating complex systems. Coordination problems are ubiquitous, and in all fields, we find the recognition of the inadequacy of fixed representations for modeling real-world work situations. Cao and Sanderson (1995) find fixed representations of process inadequate even for the constrained field of robotic workcells. Traditional deterministic planning representa- tions are simply not robust enough under real-world condi- tions. Their approach is to expand traditional Petri net models of process by allowing the transitions between states to be expressed as fuzzy variables, interpreted by production rules. While suggesting interesting techniques, their research addresses a sufficiently different problem that their solution is not immediately applicable to workflows. Specifically, their domain requires fixed sequencing of activities, and many workflows, as noted in the literature, are dynamically resequenced in response to changes in the environment. The HFBP model (Morita et al., 1996) of business work- flows is specifically intended to describe work processes between interoperating workflows in autonomous organi- zations. Drawing from research in fault tolerant software systems, they model processes in terms of an attribute grammar. While successful in addressing certain types of error recovery in interoperating WFMS (due to inherent capabilities of attribute grammars) they note that the more general problem of dynamic change in process definitions for interoperating workflows is not sufficiently handled by the attribute grammar approach alone. Much research in office information systems (OIS) describes the complexity of actual work environments, and is acutely aware of the inadequacy of fixed process models for support of real-world work situations (Boland, Maheshwari, Te’eni, Schwartz, & Tenkasi, 1992; Gerson & Star, 1986; Hewitt, 1986, 1991). The following five models come from that area: (i) The partially shared view scheme (Lee & Malone, 1990) is a technique for enhancing the utility of com- puter supported collaborative work systems used between autonomous groups with differing world views and correspondingly different functional requi- rements. The scheme has been implemented in the Object Lens collaborative work support system, and attaches a type hierarchy to each work object the sys- tem supports. The type hierarchy is assumed at least partially shared (especially at its higher levels) by all users of the system. By determining correspondences with a known hierarchy, an object otherwise unknown to a user group can be at least partially recognized and usefully processed. Our workflow model adopts this technique by analogy: the know- ledge in our hierarchy is different, but our partial matching to a hierarchical structure to make infer- ences about an unknown entity is very similar. (ii) The AMS formalism (Ang & Hong, 1994) is both a model and a collaborative work support system implementation. The formalism can be character- ized as ‘‘loosely grammatical’’ in that the activity sequences that make up a workflow are sequentially expanded from a root ‘‘abstract activity’’ by a series of productions. Stressing flexibility, however, it does not attempt to formalize the grammar (and obtain the benefits of following a well researched formalism) in the manner of HFBP. The allowed productions by which non-atomic activities can be expanded are (1) an elementary activity, (2) an activity network, and (3) a memory organization packet for activities (MOPA). The MOPA draws from the work of Schank and Abelson (1977, 1995) and others in natu- ral language understanding and is responsible for much of the flexibility of the system. A greatly simpli- fied description of a MOPA is as an abstract script 554 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 that represents an activity sequence in general terms, and is instantiated to actual activities by the run-time context. Though more accommodating to actual work situations than the Petri net models it aug- ments, AMS is essentially a single system model, and has no mechanisms for preserving coordination between interoperating systems which make use of its flexibility. (iii) The Promanand system (Karbe, Ramsperger, & Weiss, 1990) is an office support system using a ‘‘cir- culating folder’’ metaphor. It is unique in that it is the most determined attempt to date to support a deter- ministic view of office work through the exhaustive enumeration of exceptions conditions for a given work environment. The lack of success of the system is viewed by Ellis and Wainer (1994) and others as support for the position that any non-adaptive sys- tem will ultimately prove inadequate to real-world work environments. Further support comes from Hewitt (1991) who has characterized office systems as open systems which cannot be analyzed a priori. (iv) Mahling and King (1999) point out that nearly all of the problems of inflexible computer assisted work- flows were first encountered by researchers in office information systems (OIS) in the 1980s. They recog- nize that workflow models based on goals are more flexible and robust than task descriptions, and their PolyFlow system dynamically expands goal-based process definitions into activity plans for the extremely demanding office work environment. Their approach coordinates changed activity sequences for different agents in a distributed workflow through automated support of human–human negotiation. While acceptable in the single non-repeating project office environment they envision, a wholly automated re-coordination approach seems desirable for repeti- tive, production workflows. Table 1 Flexible workflow models Model Authors Term Primary flexibility en Fuzzy Petri nets Cao and Sanderson (1995) Short Fuzzy transitions to HSBP Morita et al. (1996) Short Attribute grammar d PSV Lee and Malone (1990) Short Documents carry typ AMS Ang and Hong (1994) Both Generalized scripts, w Prominand Karbe et al. (1990) Long Exhaustive enumerat PolyFlow Mahling and King (1999) Long Goal-described proce support for exception AgentWork Muller et al. (2004) Long Rule-based reschedul temporal) are detecte Multiagent workflow enactment Buhler and Vidal (2005) Long Workflows are initiat definition language. T sequences Case-handling van der Aalst et al. (2005) Both Activities are enabled (v) Similar in many respects to PolyFlow is the case-han- dling approach of van der Aalst, Weske, and Grun- bauer (2005). Directed also at office workflows, the distinguishing feature of case-handling, is that it is data driven rather than process driven; activities are enabled based on data availability rather than activ- ity completion preconditions. While allowing for very flexible sequencing of activities, case-handling systems require continuous manual attention and extreme visibility of activity and context data as is unlikely to occur in a contractor–subcontractor pro- duction workflow. Buhler and Vidal (2005) propose that adaptive workflow enactment will, based on historic and functional consider- ations, ultimately be enacted through agent-based systems. In this vision a form of high-level business process descrip- tion language will be used to compose ensembles of agents into an ‘‘initial social order’’ for enactment of the work- flow. As these software agents negotiate with each other and continuously search for web-based alternative services, workflows will be dynamically responsive to environmental changes throughout their enactment. AgentWork (Muller, Greiner, & Rahm, 2004) is a WFMS that seeks to adapt to workflow exception condi- tions by specifying high-level logical rules for approaches for rescheduling activities. The system evolved in a medical environment where timely treatment is vital and so this sys- tem focuses on the detection of temporal exceptions and even projects time overruns before they occur. When excep- tions are detected, a rule-driven scheduling system resched- ules activities and may introduce substitute activities with different execution times or treatment applicability. We note that while the workflow enactment agents of Buhler and Vidal (2005) as well as the rules of the Agent- Work system (Muller et al., 2004) implicitly use goals as the higher level ordering principle for adaptive workflows, hancement augment Petri net models escription of process e/behavior hierarchy allowing partial interpretation hich are situationally instantiated ion (attempted) of all exception conditions for given environment sses dynamically expanded to executable task sequences; negotiation s ing or substitution of activities when workflow exceptions (especially d ed by composing ensembles of software agents with a business process he agents then dynamically direct the enactment of workflow activity based on data availability rather than activity completion preconditions W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 555 neither system explicitly mentions goals. However, actions of the software agents are guided by goals in their search for alternative activity sequences and the rules that guide AgentWork have of necessity been a priori derived from the higher level goals of the processes they enact. Table 1 summarizes the adaptive/flexible work process models discussed in this section. Soffer (2005) distinguishes between two forms of workflow flexibility: long term adapt- ability sometimes referred to as agility in processes and short term flexibility which is the ability to recover from small, short term changes in a way of working. We have categorized the models in Table 1 according to this criteria for comparison to our model, which is specifically aimed at dynamic recovery from small workflow changes (short term). While all the research overviewed in this section is moti- vated by the understanding that actual work processes are highly situational and require adaptive models, none except HSBP considers the distinct, but closely related problems inherent in modeling multiple, tightly-coupled autonomous production WFMS. HSBP outlines a grammatical model for single systems, but specifically reserves the modeling of interacting systems for later research. In the next section we demonstrate a reconceptualization of workflow coordi- nation that allows both process flexibility and tight cou- pling between trading partner WFMS. 3. Goal-based coordination model framework Enhancing conventional activity-based process models with goal (intentional) information has been suggested as a basis for introducing greater flexibility into automated process enactment since at least the early 1980s (Mahling & King, 1999). In analyses of workflow meta-models (Lei & Singh, 1997), it is suggested that no other technique can provide the directive information required to dynami- cally re-sequence tasks in response to error conditions. However few models or WFMS have been developed that actually operationalize and use goal information. Our workflow coordination model captures goal infor- mation about processes in a formal structure and uses the information straightforwardly to reason about changes to process activity sequences. We have named our model after subpr satifie verify education criminal record check (state) tr satifies_using satifies_using Fig. 3. WFMS trigger reconceptualized as a semantic its most distinguishing feature: a hierarchical overlay of process intentions (HOPI). Unlike agent-based approaches which require dedicated workflow systems (cf. Mahling & King, 1999), HOPI works as an overlay to augment most existing workflow management systems. Several HOPI attributes are unique in workflow and process modeling: first, the goal hierarchy is used not for planning, replanning or evaluating a process but rather for recognition of similarity between processes; second, coordinating communications between cooperating work- flow actors are elevated to the status of first order activities rather than pre or post-condition methods within an activity object. As such they are not bound to any spe- cific activity, but like all activities within HOPI, are linked instead to the goal hierarchy. Indeed, unique to HOPI, coordinating communications are assigned goals indepen- dently of activities so that they can be rescheduled indepen- dently of activities. Fig. 3 shows a high-level object model of the same process modeled in Fig. 2. Note that the trigger, preliminary_clearance_complete is an object at the same semantic level as the other activities in the process. Subject to semantic constraints derived from process goals in addi- tion to timing and other hard constraints, it can be resched- uled and replanned to preserve coordination even under process change or activity substitution. The primary data structure of the HOPI model is a work definition (WD) that specifies a hierarchy of intentions (goals) for the activities of a work process. The WD is an overlay to the activity sequence of the workflow and thus is generally applicable to any process representation and to the many WFMS currently in use. Fig. 4 illustrates the definition of the WD used by the model. The root goal is the overall purpose of the work, and the name by which the work is commonly known in the organization. The layer of subgoals is generated during process design or specification, and represents a process of stepwise decomposition of the root goal into its compo- nents (Simon, 1977). When the subgoals are well enough defined to be implemented, they are linked to the general- ized functionality by which they will be enacted. Incorpo- rating the concept of functions into the WD more closely models the manner in which workers conceive their ocess: security_clearance s_using criminal record check (federal) igger: preliminary _clearance_complete satifies_using ally defined, independent communications event. HOPI (Any conventional process representation) Problem Space (Intentional) Task Space Root goal (R) Sub-goals (S) Generalized functions (F) Activities (A) Trigger (T) Solution Space Legend: Goal-Sub-goal edge (GE) Goal-Function edge (FE) Function-Activity Edge (AE) Predecessor-Successor links Semantics: Sub-goal Contributes-to goal Function Operationalizes sub_goal Activity Enacts function Activity precedes/follows Activity Fig. 4. A hierarchical structure of intentions as an overlay of process representations. 556 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 activities than use of goals alone (Schank & Abelson, 1995). The terminal (leaf) nodes of the WD are the actual work activities containing predecessor/successor infor- mation. Fig. 5 shows the work definition instantiated for one of our three exemplar test cases: a simple activity substitution in the manufacture of a woman’s suit, consisting of a vest and a skirt. Production of the vest is decomposed into sew- ing the vests and storing the vests until final packaging. The function that implements the storage subgoal is load&ship- ViaTruck, which is concretely realized by the activity vest- sToWH3. The interprocess trigger, a communication to a trading partner to order cloth for other manufacturing pro- cesses (not shown) is linked to the completion of vest- sToWH3. The linkage of a coordinating communication to a specific predefined activity state is common in WFMS and renders systems brittle under change. Following work redefinition (K5a ! K5b in Fig. 5), the concrete activity Fig. 5. Original and altered wo has been changed to VestsTo9thStWH, resulting in coordi- nation disruption and the failure to order the cloth in a timely manner. The dynamic workflow re-coordination model depends on similarity inferences between an original process defini- tion and an altered definition, both modeled with HOPI. Fig. 6 illustrates the similarity concepts used to interpret the intentionally specified work definitions. The WD is a tree structure in which each element is a node. The path from the root goal to any activity node is termed the inten- tional context (IC) of the node, since tracing the path yields the semantics of the node. When ICs of activity nodes in different WDs are compared beginning at the root, the low- est identical node of the ICs is called the lowest common point (LCP) of the ICs. For example, if nodes 1 through 4 of WD’s A and B in Fig. 6 are identical, but node 5 differs in A and B then node 4 is the LCP of the ICs. The depth of the LCP is a measure of IC similarity. Activity nodes also rk definitions for test case. Fig. 6. Similarity inferences based on the work definition. W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 557 have a functional context (FC), which is loosely defined as the set of activities adjacent to a given activity node. Func- tional context similarity between activity nodes i and x in two WDs is measured by an index ranging from 0 (no sim- ilarity) to 1 (identical). The index is computed by adding a weight to the index for each activity node position relative to i and x for which the nodes are identically defined. Weights diminish geometrically with distance from i and x, as shown in Fig. 6. When information on the expected duration of the activities is available, a further measure of functional context similarity is functional context dura- tion, the ratio of the times from the start of the process to the trigger activity and from the trigger activity to the end of the process in the original and the altered process expressed as a percentage. Using the example of Fig. 5 to illustrate the use of the similarity measures, note that both activity hvests to WH3i in process K5a and activity hvests to 9th StWHi in process K5b have identical intentional contexts; the path down the intentional tree from the root goal hmakeSuiti is the same in both cases. Further, the activities that precede and follow activity hvests to WH3i in process K5a are iden- tical to those that precede and follow activity hvests to 9th StWHi in process K5b and those activities in both pro- cesses implement the same function hload&shipViaTrucki. Thus the functional contexts are also identical. Together these measures strongly indicate that the change to the workflow activity is nominal and the trigger may be attached to the new activity with no loss of information. This is precisely the determination of the expert system as shown later in Fig. 13. The similarity heuristics are based on Barsalou’s (1983) work on ad hoc categories and the work of Schank and Abelson (1977, 1995) on the relation of scripts and goals to understanding. The key concept is that when activities and their goals are structured in a conceptual hierarchy as in our WDs, each higher level node serves as a cognitively organizing principle for the nodes beneath it, enabling inten- sional classification and recognition of lower level nodes. In the state-oriented view of processes (Khomyakov & Bider, 2001) a business process is a trajectory in a multidi- mensional state space where each change of state (motion in the state space) takes the process closer to one or more goal states. Under this view, two processes are considered similar if (1) there is an isomorphic mapping from one state space to another (2) the process spaces have similar goal states (3) the processes have the same valid movements in the state space toward the goal (Andersson et al., 2005). Although one is derived from a cognitive model of process similarity perception and the other from a state-space for- malism, HOPI and the Khomyakov–Bider process views are mutually reinforcing; our re-coordination model oper- ationalizes all three Khomyakov–Bider similarity metrics. Coordinating through goals rather than specific activi- ties is closely analogous to the software engineering proce- dures by which subroutines are made loosely coupled (less dependent) in order that the system as a whole should be more robust (less fragile). Prior to discussion of the operation of these principles within the expert system, a brief discussion of the usage sce- nario envisioned for the model will provide a high-level context (see Fig. 7). Two business entities, a contractor who requests work and a subcontractor who provides the work are involved in each use of the model. Both entities manage their processes with compatible WFMS and at some initial point share a common process definition for the subcontracted work. At a later time the subcontractor unilaterally changes the definition of the work being requested by the contractor; however the contractor requires notice of some intermediate point in the process (a trigger) to enable scheduling efficiencies. The coordina- tion model attempts to determine the appropriate point Fig. 7. A usage model for the coordination subsystem. 558 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 in the altered definition to schedule the trigger, eliminating the user communication and reprogramming that would otherwise be required. This usage model corresponds to a reference process- based multi-enterprise process (Schuster, Georgakopoulos, Cichocki, & Baker, 2000) in which cooperating partners construct relatively long-lived virtual enterprises. Also referred to as outsourced workflows (Schulz & Orlowska, 2004), such inter-organizational processes are normally tightly coupled through work state dependencies which limit flexibility. However, the HOPI subsystem enables autonomous process redefinition, an adaptiveness normally found only in service-based workflows while retaining close, state-based coupling. 4. An architectural view of the coordination subsystem The subsystem is divided into two major components, a process architecture and a knowledge architecture. This divi- sion is typical of knowledge engineering, where a Process and an Expertise analysis model are defined for every sys- tem (Chandrasekaran, 1986; Giarrantano & Riley, 2005; Tansley & Hayball, 1993). The knowledge architecture, the intelligence used in maintaining coordination, is an articulation of the HOPI intentional description of the workflow rationale as described in Section 3. The process architecture, which illustrates the relationship of the major functional modules required for dynamically maintaining coordination between loosely coupled (autonomous) workflows, is shown in Fig. 8. The basic structure is similar to well known architectures for analogous tasks: plan- ning, understanding from plans, and constraint-based scheduling. 4.1. Process overview Moving from left to right in Fig. 8: the original work- flow definition (activities, their ordering, etc. as described in Section 3) and a modification of that original definition are available to the model. The comparison module deter- mines differences and similarities between the definitions using the concepts of functional context and intentional con- text as set out in the previous section. Note that compari- son is restricted to those portions of a Work Definition that contain triggers. The model does not concern itself with changes in any portions of a process that do not constitute conceptual or functional contexts for triggers. After comparison, difference data is passed to the inter- pretation module. The interpretation module attempts to ‘make sense’ of the differences, and reconcile them to known elements in the original process definition. If this activity is successful, as judged by evaluation heuristics, the ‘recognized’ activity set is passed to the scheduler. That module attempts to reschedule coordinating triggers in a manner that satisfies both ‘hard’ constraints, such as fixed times for events, and fixed predecessor/successor relations and the semantics attached to the coordinating triggers. The scheduler applies domain specific and domain inde- pendent heuristics to the scheduling of coordinating com- munications. This is required, for example, when changes Difference data orand 2 Recognition/ interpretation 1 comparison Modified workflow definition Original workflow definition Reconciled workflow definition ERROR No acceptable definition found Eval Data flow Control flow Principal function Knowledge Process 3 scheduling Activity similarity rankings, original and modified WD’s Fig. 8. Knowledge and process architecture for the dynamic WFMS coordination subsystem. W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 559 to the workflow definition have resulted in the elaboration of a single task in the original definition into multiple activ- ities in the changed definition. Heuristics may be used to determine a point within the elaborated activity at which to communicate state completion to the modified workflow. A process and a modification of it can always be recon- ciled at some level of abstraction. In general, very rough confidence factors can be assigned to a reconciliation (trig- ger rescheduling) based on measures of intentional and functional context. These factors along with the scheduling actions taken by the rescheduling algorithm are posted to a log file by the system for operator review. 5. Coordination system implementation The coordination subsystem prototype has been imple- mented in CLIPS, the widely used NASA developed production system language (Riley, 2005). The HOPI workflow description hierarchy (described in Section 3) Table 2 CLIPS data structures and instantiations for a HOPI work definition HOPI node structure definition P (deftemplate goalNode ( (slot name) (slot parent) (slot workProcess) ) (deftemplate activity ( (slot name) (slot parent) (slot scheduledDuration) (slot predecessor) (slot successor) (slot workProcess) ) has been modeled as a tree structure of deftemplates nodes linked by predecessor/successor pointers. The architecture modules, comparison, recognition/interpretation and sched- uling (described in Section 4) have been modeled by functions within different CLIPS modules (namespaces). We have chosen to use the CLIPS 6.23 Windows inter- face, even though it is primarily text-based since at this stage of the project only the developers are using the system. Each case – a set of CLIPS deffacts representing a HOPI model of two workflows, original and altered – is stored in a text file maintained with any text editor. In keeping with the usage model of Fig. 7, in the WD’s the original workflow definition is termed internal (to the contracting company) and the altered definition in use at the sub- contracting site is termed external. A partial listing of HOPI data structures definitions and data instances is shown in Table 2. Pseudo-code for one of the heuristic rules that reschedule triggers within the altered workflow definition is shown in Fig. 9. artial HOPI instantiation deffacts case1ExtGt ‘‘goal tree for external process, case 1X’’ (goalNode (name XverifyReferences) (parent xVerifyPersonalInfo) (workProcess XsecurityClearance) (goalNode. . . deffacts case1Acts ‘‘activity sequence for case 1X’’ (activity (name XphoneFormerEmployers) (parent Xvalidate2) (scheduledDuration 48) (predecessor XfaxSchool) (successor XfaxToFBI) (workProcess XsecurityClearance) (activity. . . Fig. 10. Simple activity substitution (equivalent function, same level). Fig. 9. A trigger rescheduling heuristic. 560 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 6. System run with test cases 6.1. Coordination disruption exemplars (test cases) Any workflow activity change1 of arbitrary complexity can be resolved into a repeated, sequential application of the elementary change patterns diagramed below in Figs. 10–12: (1) simple activity substitution involving a termino- logical change (identical function, different terminology), (2) multiple activities are subsumed in a new, more encom- passing process step, (3) a single process step is articulated into multiple activities. The cases we describe are patterns as defined in Andersson et al. (2005) that provide the isomor- phic mappings from an original process to an altered process. 6.2. System performance under test cases Running the prototype with test cases is straightfor- ward. The coordination subsystem code (HOPI.clp) is 1 Process redesign (substantial, deliberate activity resequencing) is not discussed in this paper nor addressed by the coordination subsystem in its current form. loaded into CLIPS using the (load) command. When run, the system asks for the name of a case file of facts defining an original and altered workflow as described above. Out- put from a run of the system with the elementary activity substitution example described in Section 1 (manufacture of a women’s suit) is shown in Fig. 13 (cf. discussion of context similarity and intentional similarity in Section 3). If attached to a working WFMS, the subsystem could transmit the new, reconciled workflow – the altered work- flow with the trigger activity from the original workflow rescheduled within it – directly to the subcontractor’s workflow enactment engine as described in the usage model of Fig. 7. Alternatively, the reconciled workflow could be transmitted to supervisory personnel for review prior to transmission to the workflow enactment engine. 7. Summary and future work The functionality of the coordination subsystem in its present form is useful in many real-world situations since many alterations to work processes conform to our elemen- tary change types. More complex changes typically evolve from the execution of multiple elementary changes over Fig. 11. New activity subsumes original activity (goal equivalent; level shifted up in the abstraction hierarchy). Fig. 12. Articulation of original activity (goal equivalent; level shifted down in the abstraction hierarchy). W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 561 extended periods of time. Encouraged by the performance of the coordination model and prototype to date we have projects either planned or in process to extend this work in two directions. First, we wish to validate and extend the heuristics used in the similarity computations and rescheduling rules of the model. We have already performed a pilot study to validate the instruments that we will use in a full concurrent verbal protocol exploration of expert process coordination prob- lem solving. Analysis of the protocols should yield greater understanding of how experts reason when modifying coordinated inter-organizational processes. This experi- ment will validate the cognitive model on which HOPI is based. A second research direction seeks to make HOPI and its associated re-coordination system more practical by par- tially automating the construction of the goal hierarchy through the machine analysis of textual documentation on a process. Most work processes in large organizations are accompanied by process manuals, design documents, memoranda to personnel involved in the process and other textual material. Using many of the same natural language processing (NLP) techniques that are currently used to monitor large bodies of e-mail and other textual mate- rial and for conceptual clustering of www inquiry results, we hope to construct tools for extracting process goals and functions from textual material ancillary to a work process. Fig. 13. Re-coordination system output. 562 W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 The same NLP techniques can be used to automatically construct a lexicon (process ontology) that describes the process, its activities, its performers and their relationships. Using a lexicon, distributed artificial intelligence inference techniques (Vincenzo & Sessa, 2001) can be used to signif- icantly extend the reasoning capabilities of the coor- dination subsystem. For example, a change from one functionally equivalent process step to another involving only a change of tool could be identified with much greater confidence using a lexicon. Inferences as to the similarity of new goals to existing goal/subgoal chains would be likewise enhanced. An additional project which we hope to launch as soon as practical is to embed the coordination subsystem in an actual working inter-departmental workflow system pro- posed for the university of the first author. The information gained from real-world experience with the system would be impossible to gather any other way, and we look for- ward to the challenges this project holds. References Agostini, A., & De Michelis, G. (2000). In W. v. d. Aalst, J. Desel, & A. Oberweis (Eds.), Improving flexibility of workflow management sys- tems.. Business process management: Models, techniques and empirical studies (1806 of LNCS, pp. 218–234). Berlin: Springer Verlag. Andersson, B., Bider, I., Johannesson, P., & Perjons, E. (2005). Towards a formal definition of goal-oriented business processes patterns. Business Process Management Journal, 11(6), 1–14. Available fromhttp:// www.ibissoft.se/publications/publications.htm. Ang, J., & Hong, J. (1994). AMS formalism: an approach to office modeling and OIS development. DataBase, 25(4), 25–38. Barsalou, L. W. (1983). Ad hoc categories. Memory and Cognition, 11, 211–217. Boland, R., Maheshwari, A., Te’eni, D., Schwartz, D., & Tenkasi, R. (1992). Sharing perspectives in distributed decision making. In Proceedings of CSCW ‘92, November, 1992 (pp. 306–313). Buhler, P., & Vidal, J. (2005). Towards adaptive workflow enactment using multiagent systems. Information Technology and Management, 6, 61–87. Cao, T., & Sanderson, A. C. (1995). Task sequence planning using fuzzy Petri nets. IEEE Transactions on Systems, Man and Cybernetics, 25(5), 755–768. Chandrasekaran, B. (1986). Generic tasks in knowledge based reasoning: high level building blocks for expert system design. IEEE Expert, 3, 23–30. Davenport, T., Jarvenpaa, S., & Beers, M. (1996). Improving knowledge work processes. Sloan Management Review(Summer), 53–65. Ellis, C., & Wainer, J. (1994). Goal-based models of collaboration. Collaborative Computing, 1, 61–86. Georgakopoulos, D., Sheth, A., & Hornick, M. (1994). An overview of workflow management: from process modeling to workflow automa- tion infrastructure. International Journal of Distributed and Parallel Databases, 3(2). Gerson, E., & Star, S. (1986). Analyzing due process in the workplace. ACM Transactions on Office Information Systems, 4(3), 257–270. Giarrantano, J., & Riley, G. (2005). Expert systems: principles and programming. Boston, MA: Thomson. Hewitt, C. (1986). Offices are open systems. ACM Transactions on Office Information Systems, 4(3), 271–287. Hewitt, C. (1991). Open information systems semantics for distributed artificial intelligence. Artificial Intelligence., 47(79), 79–106. Joosten, S. (1994). Trigger modeling for workflow analysis. Proceedings of CON ’94 (Workflow Mgt. Div.). Vienna: R. Oldenburg. Karbe, B., Ramsperger, N., & Weiss, P. (1990). Support of cooperative work by electronic circulation folders. In Proceedings of the 1990 ACM COIS ‘90 (pp. 109–117). Khomyakov, M., & Bider, I. (2001). Achieving workflow flexibility through taming the chaos. Journal of Conceptual Modeling, 21, 1–14. Klein, M., Dellarocas, C., & Bernstein, A. (Eds.). (2000). Adaptive workflow systems [Special issue]. Journal of Computer Supported Cooperative Work, 9, 3–4. Kuechler, W., Vaishnavi, V., & Kuechler, D. (2001). Supporting optimi- zation of business to business e-commerce relationships. Journal of Decision Support Systems, (31), 363–377. Lee, J., & Malone, T. (1990). Partially shared views: a scheme for communicating among groups that use different type hierarchies. ACM Transactions on Information Systems, 8(1), 1–26. http://www.ibissoft.se/publications/publications.htm http://www.ibissoft.se/publications/publications.htm W.L. Kuechler Jr., V.K. Vaishnavi / Expert Systems with Applications 34 (2008) 551–563 563 Lei, Y., & Singh, M. (1997). A comparison of workflow metamodels. In Stephen W. Liddle (Ed.), Proceedings of the ER’97 workshop on behavioral models and design transformations: Issues and opportunities in conceptual modeling. Available from http://osm7.cs.byu.edu/er97/ workshop4/ls.html. Mahling, D., & King, R. (1999). A goal-based workflow system for multiagent task coordination. Journal of Organizational Computing and Electronic Commerce, 9(1), 57–82. Morita, M., Mukaigaito, T., & Hayami, H., (1996). HFBP: a hierarchal and functional business process model for describing distributed autonomous business environments. In Proceedings of the NSF workshop on workflow and process automation in information systems, Athens, GA, UGA, 1996 (pp. 91–99). Muller, R., Greiner, U., & Rahm, E. (2004). AgentWork: a workflow system supporting rule-based workflow adaptation. Data and Knowl- edge Engineering, 51, 223–256. O’Leary, D., Kuokka, D., & Plant, R. (1997). Artificial intelligence and virtual organizations. Communications of the ACM, 40(1), 52–59. Riley, G. (2005). Clips reference manual: Volume 1 – Basic programming guide, June, 2005. Available from http://www.ghg.net/clips/clips.html. Schank, R., & Abelson, R. (1977). Scripts, plans, goals and understanding: an inquiry into human knowledge. Hillsdale, NJ: Lawrence Erlbaum. Schank, R., & Abelson, R. (1995). Knowledge and memory: the real story. Hillsdale, NJ: Lawrence Erlbaum. Schulz, K., & Orlowska, M. (2004). Facilitating cross-organizational workflows with a workflow view approach. Data and Knowledge Engineering, 51, 109–147. Schuster, H., Georgakopoulos, D., Cichocki, A., & Baker, D. (2000). Modeling and Composing Service-Based and Reference Process-Based Multi-enterprise Processes. In B. Wangler & L. Bergman (Eds.). Proceedings of advanced information systems engineering: 12th interna- tional conference, CAiSE 2000, Stockholm, Sweden, June 2000 (LNCS 1789/2000). Berlin: Springer Verlag. Simon, H. (1977). The new science of management decision (Revised Edition). Englewood Cliffs, NJ: Prentice Hall. Soffer, P. (2005). On the notion of flexibility in business processes. In Proceedings of the 6th workshop on business process modeling, devel- opment and support, Porto, Portugal. Available from http://lam- swww.epfl.ch/conference/bpmds05/program/soffer_4.pdf. Tansley, D., & Hayball, E. (1993). Knowledge-based systems analysis and design: a KADS developers handbook. New York: Prentice Hall. Vaishnavi, V., Joosten, S., & Kuechler, W. (1996). Representing workflow management systems with smart objects. In Proceedings of NSF workshop on workflow and process automation in information systems, Athens GA. Vaishnavi, V., & Kuechler, W. (2005). Universal enterprise integration: challenges of and approaches to web-enabled virtual organizations. Information Technology and Management, 6(1), 5–16. van der Aalst, W., & Jablonski, S. (2000). Dealing with workflow change: identification of issues and solutions. International Journal of Computer Systems, Science and Engineering, 15(5), 267–276. van der Aalst, W., Weske, M., & Grunbauer, D. (2005). Case handling: a new paradigm for business process support. Data and Knowledge Engineering, 53, 129–162. Vincenzo, L., & Sessa, S. (Eds.). (2001). Soft computing agents: new trends for designing autonomous systems. Heidelberg [Germany], New York: Physica-Verlag. Yu, E. (1995). Models for supporting the redesign of organizational work. In Proceedings of the Conference on Organizational Computing Systems (pp. 225–236). Milpitas, CA: ACM Press. http://osm7.cs.byu.edu/er97/workshop4/ls.html http://osm7.cs.byu.edu/er97/workshop4/ls.html http://www.ghg.net/clips/clips.html http://lamswww.epfl.ch/conference/bpmds05/program/soffer_4.pdf http://lamswww.epfl.ch/conference/bpmds05/program/soffer_4.pdf An expert system for dynamic re-coordination of distributed workflows Introduction Search for flexibility in workflow and process modeling Goal-based coordination model framework An architectural view of the coordination subsystem Process overview Coordination system implementation System run with test cases Coordination disruption exemplars (test cases) System performance under test cases Summary and future work References 
work_jgz56hlogncihjdwscuvfwh4vq	----	Untitled May 21, 2010 21:20 chaos˙v2 Fuzzy weather forecast in forecasting pollution concentrations D. Domańska Department of Modelling and Computer Graphics, Institute of Informatics, University of Silesia, Bȩdzińska 39 Sosnowiec, 41-200, Poland ddomanska@poczta.onet.pl M. Wojtylak Institute of Meteorology and Water Management (IMGW), Bratków 10 Katowice, 40-045, Poland monitoring.katowice@imgw.pl In this paper we want to analyse fuzzy weather forecasts, which are computed in our system and used to forecast pollution concentrations. The system works on real data: weather forecasts, meteorological situations and pollution concentrations. We compare defuzzification of the fuzzy weather forecast with weather forecast from Institute of Meteorology and Water Management. This comprehensive analysis allows us to investigate the effectiveness of forecasting pollution concentrations, putting the dependence between particular attributes describing the weather forecast in order and proving the legitimacy of the applicable fuzzy numbers in air pollution forecasting. Model is created for data, which is measured and forecast in Poland. By reason of this data our model is tested in real sets of data and effects are received in active system. Keywords: Fuzzy system models, Fuzzy numbers, Fuzzy matrix, Fuzzy weather forecast, Air pollution forecasting 1. Introduction Weather forecasting is increasingly perfected in subsequent centuries. In recent years many prediction ap- proaches, such as statistical [1], fuzzy [2; 3], neural networks [4; 5], neuro-fuzzy predictor [6] have emerged. Using numerical short-term weather prediction, research into the forecasting of air pollution concentrations began [7; 8]. This task is very difficult because apart from the information about meteorological conditions, the emission of air pollution depends first of all on the immission. Fuzzy logic [9; 10; 11] is an established methodology that is widely used in model systems in which variables are continuous, imprecise, or ambigu- ous. Use of this method is known in many mathematical forecasting models. It is usually used when the information transferred to the model is imprecise or incomplete [12; 13]. Many everyday phenomena of an ambiguous, continuous and imprecise nature may be effectively described using this theory. The main problem is with the knowledge. We do not have precise knowledge about the weather in the future. We only have numerical forecasting, i.e. conditions which may announce many similar meteorological situations. The result of a working Air Pollution Forecasting Model (APFM) is a forecast of air pollution concen- tration, among others PM10 for the next day. It is a specially chosen pollution because PM10 has a huge influence on human life [14; 15]. In each stage we use meteorological data with a mathematical apparatus [16; 17]. In particular in 1 May 21, 2010 21:20 chaos˙v2 2 APFM we use the weather forecasts derived from the Consortium for Small Scale Modelling (COSMO) model based on the Local-Model (LM) of Deutscher Wetter Dienst (DWD). Objects appearing in the paper are vectors and matrices. In vector space Rd for vectors we use (1) as the distance between objects. ddk(x, y) = ( d ∑ i=1 |xi − yi|k ) 1 k , x, y ∈ Rd, k > 0 . (1) For k > 1, k ∈ N the function (1) is metric. The distance between matrix objects is composition of vector objects. For terms distance between matrices we use (2). d n×m k1k2 (A, B) = dnk1 ([d m k2 (ai·, bi·)], 0n), for A = [aij ], B = [bij ], A, B ∈ Rn×m , (2) where 0n is a zero vector and ai·, bi· are i-th rows in matrices A and B. In the first instance we introduce a term, time horizon set T , in which the forecast will be calculated. T = {t = i · ∆t : i = 0, . . . , nT }, ∆t > 0, where ∆t means a time step (usually ∆t = 1 hour). We will identify the term from set T with 0 hour UTC. We assume that for each term from T we have values of df parameters of a numerical weather forecast (e.g. temperature, sea level pressure, wind direction and speed, cloud cover — high, medium, low). For the term weather forecast we will understand a matrix F ∈ R(nT +1)×df . Moreover, we assume that we possess the data from days from many years in every term t ∈ T . Every term t describes the state of the atmosphere with the aid of the ds parameters measured near the surface (e.g. temperature, wind direction) and the value of concentrations whose size we are forecasting. The set of meteorological data for each subsequent term t ∈ T defines the meteorological situation. The meteorological situation will be represented by a matrix S ∈ R(nT +1)×ds . The aerosanitary situation is the number of sequences of concentrations in t ∈ T terms, so it is a time series belonging to RnT +1. In order for the model to function properly it is essential to have all the historical data. Let us denote the set of weather forecasts as W F , the set of meteorological situations as M S, the set of pollution concentrations as AS. In the first stage, because of the huge data range, we start from min − max normalisation for every weather forecast in every column separately. Let us define: f ∗ ∈ W F — a chosen weather forecast for which we are calculating the forecast of pollution concentrations. k = k1 = k2 — first parameter used to control APFM system, it decides about dispersion between elements from set W F . Θ — a set of real numbers representing distances between every normalised weather forecast from W F and f ∗ so Θ = {ω1, . . . , ωq}, q = |W F | − 1, where ωi = d (nT +1)×df k (fi, f ∗) and fi ∈ W F . |A| means cardinality of set A. Fractional distance — distance (1) for k1, k2 ∈ (0, 1). In the next step we define second parameter ε. Parameter ε decides about the cardinality of similar elements from set W F . ε is decided in (3). ∀i=1,...,q ωi < ε, (3) where ωi ∈ Θ for i ∈ {1, . . . q}, i is the number of an element. The result of the first stage is set ε–W F (f ∗). In the second stage, in connection with results from the first stage, we create subset ε–M SF ⊂ M S. Every weather forecast is related with meteorological situation with date. Therefore for ε–M SF we consider pairs (f, s), f ∈ W F, s ∈ M S. Then, we set parameters describing the meteorological situations and the time horizon. After review of the chosen meteorological situations, we get a sequence of values: ∀t∈T ∀i=1,...,ds (ξ (1) t,i , . . . , ξ (m) t,i ), where m = |ε–M SF | . (4) We modify this sequence into a fuzzy number using a special form of the fuzzy number given by (5). For each attribute i and in each hour t ∈ T we have individual fuzzy number. This fuzzy number is approximate May 21, 2010 21:20 chaos˙v2 3 to the Gaussian function. The fuzzy number (5) was chosen based on our own calculations and based on paper [18]. µi,t(x) =          exp( −(x−m1i,t ) 2 2·σ2 1i,t ) if x ≤ m1i,t , 1 if x ∈ (m1i,t , m2i,t ) , exp( −(x−m2i,t ) 2 2·σ2 2i,t ) if x ≥ m2i,t , (5) where m1i,t ≤ m2i,t , σ1i,t > 0, σ2i,t > 0 for m1i,t , m2i,t , σ1i,t , σ2i,t ∈ R, for each attribute i in each hour t ∈ T . An individual fuzzy weather forecast consists of a time series (5). We receive fuzzy weather forecast (6) — equivalent to the real weather. φ∗ = [µi,t] for 1 6 i 6 ds, t ∈ T (6) where i is a number of attribute and t is an hour. φ∗ is a function for which we determine membership matrix composed from fuzzy numbers. φ∗it : R ds → [0, 1]. In φ∗ we can take property values. We receive φ∗(s) = [µit(sit)] which is membership matrix, s ∈ R(nT +1)×ds . A time series is a sequence of the regularly sampled quantities of an observed system. Fuzzy weather forecast needs to be able to clearly and precisely define the quality of a weather forecast and assign meteorological situation relative to a fuzzy weather forecast [13]. The assignment of a coefficient quality of a weather forecast is the point of entrance for the exact determination of the individual influence of an attribute on forecasting pollution concentrations in the future. In the third stage we review all meteorological situation s ∈ M S ⊂ R(nT +1)×ds . Then, for every meteorological situation s we calculate φ∗(s) and number ̺(s) = |φ∗(s)| using formula (7). |φ∗(s)| = dnT +1 k ([dds k (φ∗·t(s·t), 1ds )]t∈T , 0nT +1), s ∈ M S, k ∈ [0, 1] , (7) Let us fix η ∈ [0, 1] and determine a set η–M S ⊂ M S that |η–M S| = r, where ̺(s) < η for s ∈ η–M S. For subset η–M S we consider pairs (s, p), p ∈ AS. Then we fix the weight of the meteorological situations using following formula w(s) = 1 − ̺(s), s ∈ η–M S. In the fourth stage we choose r time series from set AS, where r = |η–M S| > 1. Afterwards, for every chosen time series we get a function p(j) : T → R+0 , j = 1, . . . , r with weight w(j) ∈ R, representing pollution concentrations. For each t ∈ T we create a sequence (p(1)(t), . . . , p(r)(t)). Then we take these sequences and we carry out an aggregation process to obtain one time series. We have used for example method αβ-aggregation (9). This and another the methods are described in details in paper [17]. We base these methods on the well-known methods: (8). ∀t∈T ua,t = ∑r i=1 w (i)p(i)(t) ∑r i=1 w (i) , ∀t∈T um,t = max i=1,...,r {p(i)(t)} . (8) where a means average aggregation, m means maximum aggregation. Let us denote for each t ∈ T the following time series ua,t and um,t as a time series received from methods (8) and ur,t as a time series received from the actual researched data. For l ≈ nT4 we determine two numbers based on knowledge about actual aerosanitary situation. We forecast pollution concentrations having partial knowledge, that is real number l of pollution concentrations. We execute some calculations on time series and we get parameters α, β from the second function (9). When we determine the optimal value of the parameters α, β for (9) we receive formulas (10), (11). h(α, β) = l ∑ t=0 (αua,t + βum,t − ur,t)2 . (9) where α = ∑l t=0 ua,tur,t ∑l t=0 u 2 m,t − ∑l t=0 um,tur,t ∑l t=0 ua,tum,t ∑l t=0 u 2 m,t ∑l t=0 u 2 a,t − ( ∑l t=0 um,tua,t) 2 , (10) May 21, 2010 21:20 chaos˙v2 4 (a) (b) Fig. 1. Fuzzy weather forecast for wind speed attribute on a) 9 January 2006 b) 10 January 2006. (a) (b) Fig. 2. Fuzzy weather forecast for temperature attribute on a) 9 January 2006 b) 10 January 2006. if ∑l t=0 u 2 m,t ∑l t=0 u 2 a,t − ( ∑l t=0 um,tua,t) 2 6= 0. β = ∑l t=0 um,tur,t ∑l t=0 u 2 a,t − ∑l t=0 um,tua,t ∑l t=0 ua,tur,t ∑l t=0 u 2 m,t ∑l t=0 u 2 a,t − ( ∑l t=0 um,tua,t) 2 , (11) if ∑l t=0 u 2 m,t ∑l t=0 u 2 a,t − ( ∑l t=0 um,tua,t) 2 6= 0. From (9) we receive parameters α, β given by (10) and (11). Then using method (9) and having knowledge about collateral information i.e. first ten hours real pollution concentrations in the day we being forecast, we can calculate for each t ∈ T the final time series uf,t. 2. Characteristics of fuzzy weather forecast A fuzzy weather forecast φ∗ is determined for each attribute i individually and is evenly distributed on each hour t ∈ T . It is valued on the basis of data similarity and proper weights of classification. We researched the behaviour of the fuzzy weather forecasts using different sets of forecast data. This is necessary because we have weather forecasts from a short period of time (only six years). Therefore, continuous work in a COSMO LM model weather forecast [19], [20] is not heterogeneous for finding the period of a weather forecast which is the best estimate of real meteorological situations. In Figs 1, 2, 3 fuzzy weather forecasts are shown along with real meteorological situations. The fuzziness is a good measure with which to mark the quality of a weather forecast both its elements and the whole weather forecast because fuzziness characterises the scattering of real data around the prognosis. 3. Features to estimate the quality of a fuzzy weather forecast The first feature is research volume for all attributes i in each hour t ∈ T . We receive the first number F i,t F i,t = ∞ ∫ −∞ µi,t(x)dx = √ 2Π 2 (σ1i,t + σ2i,t ) + m2i,t − m1i,t , (12) where i ∈ {1, . . . , ds}, t ∈ {0, . . . , nT }, σ1i,t , σ2i,t , m1i,t , m2i,t ∈ R(nT +1)×ds . Let us define fi(t) = F i,t for each t ∈ T . In Figs 4, 5, 6 we see the functions fi(t) for all t ∈ T and for the chosen attributes i are shown using real meteorological situations: wind speed, temperature and humidity. May 21, 2010 21:20 chaos˙v2 5 (a) (b) Fig. 3. Fuzzy weather forecast for humidity attribute on a) 9 January 2006 b) 10 January 2006. (a) (b) Fig. 4. fi for wind speed attribute on a) 9 January 2006 b) 10 January 2006. (a) (b) Fig. 5. fi for temperature on a) 9 January 2006 b) 10 January 2006. In Fig. 7 fuzzy weather forecast is shown for all attributes. The second feature is researching the quality of a fuzzy weather forecast by comparing it to a real meteorological situation. In this way we keep an attribute characterised by a grade of membership for each hour. In Figs 8, 9 grades of membership for wind speed, temperature and humidity are shown. In Figs 1, 2, 3 the fuzzy weather forecasts are clearly and explicitly shown and it can be seen that the fuzzy weather forecast has a little fuzziness when the grade of membership is large. By analysing other examples, a fairly significant dependence between small fuzziness and membership can be observed, a reverse correlation between them, because if the fuzziness is greater then the checkness of the pollution concentrations forecasting is minor. (a) (b) Fig. 6. fi for humidity attribute on a) 9 January 2006 b) 10 January 2006. May 21, 2010 21:20 chaos˙v2 6 (a) (b) Fig. 7. fi for all attributes on a) 9 January 2006 b) 10 January 2006. (a) (b) Fig. 8. The grade of membership µi,t for wind speed attribute and T = 72 on a) 9 January 2006 b) 10 January 2006. (a) (b) Fig. 9. The grade of membership µi,t for a) temperature attribute b) humidity attribute and T = 72 on 9 January 2006. 4. Verifiability of the weather forecast Taking into account that the key to the pollution concentration forecasting is reliable information coming from the weather forecast we need to separate the forecasts which we think that are fulfiled from the forecasts that have not fulfil. To make this we will use fuzzy weather forecast, weather forecasts and meteorological situations. Let us introduce following defuzzyfication method for µi,t(x): φi,t = ∞ ∫ −∞ µi,t(x)· x dx ∞ ∫ −∞ µi,t(x) dx = 1 2 (m22i,t − m 2 1i,t ) + σ22i,t − σ 2 1i,t + √ 2π 2 (m1i,t σ1i,t + m2i,t σ2i,t ) F i,t (13) Fuzzy weather forecast fi,t is defuzzyficated according to equation (13), we obtain a real value for every hour t ∈ T and for every attribute i. The verification of the weather forecast is based on similarity of the data: (1) value of the attributes describing the weather forecast to the values of the meteorological situations attributes (2) φi,t to the value of the attributes describing the weather forecast (3) φi,t to the value of the meteorological situations attributes May 21, 2010 21:20 chaos˙v2 REFERENCES 7 Table 1. Results of the tests for the temperature attribute. Date Mfi,si Mφi,t,fi Mφi,t,si 1 January 2006 0.363 0.562 0.082 5 January 2006 0.525 0.424 0.125 Table 2. Results of the tests for the wind speed at- tribute. Date Mfi,si Mφi,t,fi Mφi,t,si 1 January 2006 0.069 0.076 0.018 5 January 2006 0.133 0.172 0.046 10 January 2006 0.053 0.050 0.066 We use Mean Absolute Error [22] to estimate the verifiability of the weather forecast. For each attribute i the Mean Absolute Error between fuzzy weather forecast and weather forecasts is given by: Mφi,t,fi = 1 nT + 1 nT ∑ t=0 |φi,t − fi| (14) where f ∈ W F and |x| is the absolute value of x. Analogous is with the differences between meteorological situations and fuzzy weather forecast. We denote the Mean Absolute Error in this case as Mφi,t,si , s ∈ M S and for the differences between meteoro- logical situations and weather forecasts we use the notation Mfi,si , s ∈ M S, f ∈ W F for the Mean Absolute Error. In Tabs. 1, 2 we have the obtained average differences between the weather forecast attributes and real situations, weather forecast and φi,t, real situations and φi,t. At first we can see in Tab 1 that for greater fuzzines differences between weather forecasts and me- teorological situations, between fuzzy weather forecast and weather forecast are greater than between meteorological situations and fuzzy weather forecast. At second we see possibility to improve the weather forecast. Analogous is with the wind speed attribute in Tab 2, however small fuziness give us minor possi- bility to improve the weather forecast. 5. Conclusion Computations were performed for weather forecasts in 2003-2007, meteorological situations in 1997-2007 and pollution concentrations in 1998-2007 with ∆t = 1. We have df = 28 attributes describing weather forecasts, while the number of meteorological situations was equal to ds = 9. Attributes describing meteo- rological situations were chosen based on investigations by [21]. The effect of the suggested method for the prediction of a weather forecast was introduced for data from COSMO LM model, but the same method can be used for different weather forecasts based on numerical models. The condition which has to be met is to have real meteorological data. In the paper we have introduced the analysis of the fuzzy weather forecast, with specification of the method which allows us to improve the calculated weather forecast. In Figs. 1, 2, 3 is shown fuzzy weather forecast and in Figs 8, 9 particular grades of membership. The performed experiments have shown that if the fuzzines is smaller than the membership to the real situations is bigger. In Tabs 1, 2 we see that for φi,t and real situation we have the smallest difference between obtained attributes, what allows us to improve the quality of the weather forecast. References [1] G. Rigatosa, Q. Zhangb, Fuzzy model validation using the local statistical approach, Fuzzy Sets and Systems 160(7) (2009) 882-904. May 21, 2010 21:20 chaos˙v2 8 REFERENCES [2] C. Lee, A. Liu and W. Chen, Pattern Discovery of Fuzzy Time Series for Financial Prediction, IEEE Trans. Knowl. Data Eng 18(5) (2006) 613-625. [3] H. Kunhuang, Heuristic Models of Fuzzy Time Series for Forecasting, Fuzzy Sets and Systems 123(3) (2001) 369-386. [4] W. Jiang and P. Wang, Research on Interval Prediction of Nonlinear Chaotic Time Series Based on New Neural Networks, in: Proc. 6th World Congress Intell. Control and Automation, 2006, pp. 2835-2839. [5] Ajit Kumar Gautam, A.B. Chelani, V.K. Jain, S. Devotta, A new scheme to predict chaotic time series of air pollutant concentrations using artificial neural network and nearest neighbor searching, Atmospheric Environ. 42(18) (2008) 4409-4417. [6] M.J.L. Aznarte, J. Manuel Beńıtez Sánchez, D. Nieto Lugilde, C. de Linares Fernández, C. Dı́az de la Guardia, and F. Alba Sánchez, Forecasting Airborne Pollen Concentration Time Series with Neural and Neuro-Fuzzy Models, Expert Systems with Applications 32(4) (2007) 1218-1225. [7] L. Ośródka, E. Krajny, M. Wojtylak, The use of numerical weather forecast for air pollution forecast- ing in an urban industrial agglomeration, in: Proc. 4th Annual Meeting EMS and 5th EC on Applied Climatology, 2004, EMS Vol. 1. [8] L. Ośródka, M. Wojtylak, E. Krajny, K. Rorbek, Improvement of the high pollution concentrations fore- casting methods in urban-industrial agglomerations with the help of numerical models of meteorological forecasts, Wiad. IMGW 27(1) (2004) 105-116. [in Polish] [9] L.A.Zadeh, Fuzzy Sets, Inf. and Control 8 (1965) 338-353. [10] G. Klir, T. Folger, Fuzzy Sets, Uncertainty and Information, Prentice Hall PTR, Englewood Cliffs, NJ, 1988. [11] H.J. Zimmerman, Fuzzy Set Theory and its Applications, second ed., Kluwer, Dordrecht, 1991. [12] W. Silvert, Ecological impact classification with fuzzy sets, Ecol. Model. 96 (1997) 1-10. [13] B.K.Hansen, D.Riordan, Weather Prediction Using Case-Based Reasoning and Fuzzy Set Theory, in: Proc. Workshop on Soft Computing in Case-Based Reasoning, International Conf. on Case-Based Reasoning, Canada, 2001, pp.175-178. [14] M. Kowalska, L. Hubicki, E.J. Zejda, L. Ośródka, E. Krajny, M. Wojtylak, Effect of ambient air pollution on daily mortality in Katowice Conurbation, Poland, Polish J. Environ. Stud. (2007) 227-232. [15] M. Kowalska, J.E. Zejda, L. Ośródka, K. Klejnowski, E. Krajny, M. Wojtylak, Relationship between ambient air pollution and daily mortality in the Urban Area of Katowice-comparison on two periods 1994-1995 and 2001-2002, in: Proc. Public Conf. ”Particles and Health-State of the Research and Policy Implications”, (9), 2007. [16] D. Domańska, M. Wojtylak, Development data serving to forecasting pollution concentrations, Deci- sion Support Systems, Katowice, 2009, pp. 351-359. [in Polish] [17] D. Domańska, M. Wojtylak, Selection criteria of forecast pollution concentrations using collateral informations, Computer Methods and Systems, Kraków, 2009, pp.213-218. [18] L. Ośródka, M. Wojtylak, E. Krajny, R. Dunal, K. Klejnowski, Application Data Mining for forecasting of high-level air pollution in urban-industrial area in southern Poland, in: Proc. of the 10th Int. Conf. on Harmonisation within Atmospheric Dispers. Modelling for Regulatory Purposes, 2005, pp. 664-668. [19] G. Brunet, The first hundred years of numerical weather prediction, Proceedings of the 19th Int. Symposium on High Performance Computing Systems and Applications, Canada (2005). P. C. Banacos and D. M. Schultz The use of moisture flux convergence in forecasting convective initiation: historical and operational perspectives, Weather and Forecasting, 20, 351-366 (2005). [20] W. Stawiany, P. Caban, B. Cimander, E. Cisowska, I. Ho lda, E. Krajny, A. Krucza la, E. Ostrowska, L. Ośródka, K. Rorbek, S. Socha, J. Świȩch-Skiba, M. Wojtylak, Monograph of automatic measurement systems of the air quality in Katowice agglometaion (1993-1999), Biuletyn Regionalnego Monitoringu Środowiska: Wojewoda Śla̧ski (2000) 1-324. [in Polish] [21] Rob J. Hyndman and Anne B. Koehler, Another look at measures of forecast accuracy, International Journal of Forecasting (2006). 22(4), 679-688. 
work_jjzc4fo46fhb3jzjht3pgs26ni	----	Redalyc.A Novel Web-based Human Advisor Fuzzy Expert System Journal of Applied Research and Technology ISSN: 1665-6423 jart@aleph.cinstrum.unam.mx Centro de Ciencias Aplicadas y Desarrollo Tecnológico México Rafe, Vahid; Hassani Goodarzi, Mahdi A Novel Web-based Human Advisor Fuzzy Expert System Journal of Applied Research and Technology, vol. 11, núm. 1, febrero, 2013, pp. 161-168 Centro de Ciencias Aplicadas y Desarrollo Tecnológico Distrito Federal, México Available in: http://www.redalyc.org/articulo.oa?id=47426212015 How to cite Complete issue More information about this article Journal's homepage in redalyc.org Scientific Information System Network of Scientific Journals from Latin America, the Caribbean, Spain and Portugal Non-profit academic project, developed under the open access initiative http://www.redalyc.org/revista.oa?id=474 http://www.redalyc.org/articulo.oa?id=47426212015 http://www.redalyc.org/comocitar.oa?id=47426212015 http://www.redalyc.org/fasciculo.oa?id=474&numero=26212 http://www.redalyc.org/articulo.oa?id=47426212015 http://www.redalyc.org/revista.oa?id=474 http://www.redalyc.org 161Journal of Applied Research and Technology A Novel Web-based Human Advisor Fuzzy Expert System Vahid Rafe*1, Mahdi Hassani Goodarzi2 1,2 Department of Computer Engineering, Faculty of Engineering, Arak University, Arak ,38156-8-8349, Iran. Department of Computer Engineering, Islamic Azad University- South Tehran Branch, Iran. *v-rafe@araku.ac. ABSTRACT The applications of the Internet-based technologies and the concepts of fuzzy expert systems (FES) have created new methods for sharing and distributing knowledge. However, there has been a general lack of investigation in the area of web-based fuzzy expert systems. In this paper, the issues associated with the design, development, and use of web-based applications from a standpoint of the benefits and challenges of development and utilization are investigated. The original theory and concepts in conventional FES are reviewed and a knowledge engineering framework for developing such systems is revised. For a human advisor to have a satisfying performance, expertise is a must. In addition, some of advisory rules are subject to change because of domain knowledge update. The human requests may have linguistic or crisp forms and a conventional expert system (ES) is not able to overcome the fuzziness in the problem nature. In this research, a Web-based fuzzy expert system for Common Human Advisor (FES-CHA) is developed and implemented to be used as a student advisor at the department's web portal. The system is implemented by using Microsoft Visual Studio .NET 2010, MVC and Microsoft SQL Server 2012. Keywords: fuzzy expert systems, web-application, common human advisor, total average. 1. Introduction Knowledge-based and decision making systems are the branches of artificial intelligence which are based on imitating the human demeanor in finding the pattern of solutions to problems. In the real world, if definite and straightforward solution cannot be found, human expertise is needed. Experts often follow a trial-and-error approach for problem solving. Since there is no specific solution for this kind of problems, defining a certain computer method for achieving the solution is difficult. Therefore, expert systems are used to reach this goal. In these systems, the program consists of a set of rules. The knowledge in an expert human brain is also a set of if-then rules. M. H. Goodarzi [1,2,3] proposed the fuzzy application in student evaluating system, portfolio advisor system and educational advisor system. Fuzzy concepts can convert multiple crisp inputs to specific linguistic variables and use fuzzy rules to infer. In [4] the fuzzy-based advisor for elections and the creation of political communities was proposed. In [5], a web-based fuzzy expert system is used to help inexperienced Indian farmers in the use of pesticide for their farms. The initial version of this software was introduced in 1995 in a single-user form. In forums, usually a user starts a discussion and expresses his/her opinions and approaches to a particular problem. [6] Proposes a model for creating a fuzzy-based expert forum that intelligently responds to questions asked by users. Finding the right broker at the right time is another issue that requires expertise. This may be the reason for which inexperienced investors loose in stock markets. In [7] a stock expert system model is proposed. The goal of this system is to make a good suggestion based on information about goods and market in order to reduce the loss and increase the benefit. Educational consulting system tries to mimic the behavior of the staff addressing the educational consulting issues. In [8] a fuzzy expert system for intelligent tutoring systems with a cognitive mapping is proposed. Human cognition has become one of the most attractive areas of research and application in artificial intelligence in which human susceptibility is emulated. In [9] a new fuzzy method for hotel selection is introduced as a hotel advisory system.. In [10] the student A Novel Web‐based Human Advisor Fuzzy Expert System, Vahid Rafe / 161‐168 Vol. 11, February 2013 162  achievements and education system performance in a developing country is proposed. The current paper, includes five major sections: in the next section some of related works are reviewed. The third section describes the fuzzy rule-based and decision making systems and introduces the proposed model. In section 4 the proposed system is discussed in details. Section 5 includes a sample of the advisor system implemented in a university. Finally, a conclusion is provided. 2. Literature review In a real voting the total of both, positive and negative votes for the candidates are collected. A major problem for the voters is when they have to select their deputies from a large list of candidate. The problem is more serious in cases where the candidates are unknown to the voters. However, the creation of political societies interested in addressing political issues is a hurdle to overcome. In [4] an advisor system for elections and creation of political communities based on fuzzy logic is proposed. In this approach the recommendation engine works with a modified fuzzy C-means algorithm and the Sammon mapping technique used for visualization of recommendations. Each year in India, many farms are destroyed due to pests attack and insufficient experience of the farmers. In 2001, the loss was about 6.3 billion dollars. Soybean pest expert system (SOYPEST) [5] is a fuzzy expert system that asks fuzzy questions in order to generate a web-based response for the user. SOYPEST is created by using JESS and gradually became more accurate by receiving feedback from the users and the experts. Mutual information interchange and the creation ng of forums on the web are important issues which captivate many researchers. One of the best known content management systems (CMS) tools for this purpose is Vbulletin. There are reasons that support the possibility of receiving irrelevant answers, no answers at all, different confusing answers from several other users and unclear answers. These drawbacks may be considered as the Achilles heel of such systems. In [6], linguistic expressions are categorized and then n-gram algorithm is used to edit and convert the sentences to a proper format. This system supports 15 languages and by default the questions are multiple choice questions. The strong point of this system is its gradual improving knowledge base, the extended number and the expanded fields of topics; however, no considerable effort is done to find the best answer and the problem is solved through partial simulation of the human brain. 3. Fuzzy decision making system Most humans face difficulties related to life rules and regulations during their problem solving process. These rules and regulations are changing every now and then therefore, an expert is needed to memorize these rules in order to be able to help humans in their issues. The state of each person regarding the rules and regulations may differ from that of other people. Human state (HS) is a member of a fuzzy set with a degree of membership equal to HS . The First Step: determining and fuzzificating the inputs to the system by using fuzzy rules. Following are some examples of the fuzzy sets of the system on hand: Fuzzy set for Law in judgment system. Fuzzy set for passed courses in university. Fuzzy set for marks of selected courses in university. Fuzzy set for the rank and grade of student in the entrance exam And so on… 3.1 One sample for fuzzification of the crisp variable of TA in university system Total grade point average (GPA) of students can be categorized into these groups: A, B, C, D and E. This categorization can be expressed through linguistic terms as Excellent, Good, Middle, Weak, Very Weak. The Second Step: determining the degree of membership of linguistic terms including 3 following phases: A Novel Web‐based Human Advisor Fuzzy Expert System, Vahid Rafe / 161‐168 Journal of Applied Research and Technology 163 Phase 1. For each term, the value nearest to the numeric equivalent of the linguistic term which has the maximum degree of membership is selected. Here, the highest value for the linguistic term “Excellent” is 20 and 0 has the highest value in “Very Weak” fuzzy set. Phase 2.For each term, the value (or values) which has (have) the membership degree of 0 is (are) determined. Phase 3. point with 1 are connected to points with 0 by lines to form a gaussian (exponential) membership function. In cases in which there is more than one point with 1 , a gaussian membership function is obtained. In this model, membership functions can be gaussian, Z- shaped and S-shaped. The membership function of GPA variable is shown in Figure 1. Figure 1. Student's GPA membership function For example: to fuzzify TA=3 into a linguistic variable, first of all, we write the formula for Z- shaped and gaussian membership function.                               xc cxb ac cx bxa ac ax ax cbaxZ 1 21 2 0 ),,,( 2 2 (1)         2 2)(5.0 expFGaussian  cx M (2) Then the result of equations for TA =3 are calculated giving the following equations: 125.0)4/3(2 21)4,2,0,3( )( 2           TAx ac ax Z  1353.0)2exp( 2 )73(5.0 exp 2 2 )(        TAx TA = 3 with 125.0)( TAx exists in "very weak" fuzzy collection and with 1353.0)( TAx is a member of "weak fuzzy" collection. With competitive method the TA is changed to weak linguistic variable. 3.2 Rule extraction The core of the system is very flexible and can be applied in many advisory environments by substituting the knowledge-base of the system with an appropriate medical, judicial or sport, etc. advisory knowledge-base. For example, after an interactive negotiation with an advisor lecturer at a university, fuzzy rules can be elicited and used in the ES. The previous section illustrates how GPA can be fuzzified. The following rules are the result of the negotiations: If The GPA is Moderate And The number of semesters for which the student is registered is small And The courses that are not passed can be taken. And passed in one semester And The student has not given any pledges And The student has not received any disciplinary notices And The student has not reached the maximum time period for his/her studies And The student has 5 marks between 10-12 And Some other student has had conditions similar to this Student A Novel Web‐based Human Advisor Fuzzy Expert System, Vahid Rafe / 161‐168 Vol. 11, February 2013 164  Then To a large degree, it is possible that the student is allowed to continue his/her studies in the university by giving an official pledge of achieving a GPA over 14 in the next semester. Else According to the status of the student, the fuzzy system is not able to provide an answer. A human expert's opinion is needed. 3.3 Publishing the system on the web The rapid increase of information on the Internet is currently a key issue when one is looking for relevant information. The development of the World Wide Web and applying multimedia tools along accessibility of web sites from any place in the world makes feasible the design ofuser interface compatible with the web. Many expert systems in different fields of expertise are developed (EXSYS CORVID, SOYPEST, etc.) however, few are applied. Since linguistic terms and fuzzy sets are used, the process for inference should be done on the client rather than the server to reduce the server’s busy time. This procedure can be executed in browser by script languages like JavaScript, Java, VB Script, XML, AJAX and Applet. 3.4 One application of the proposed system in a university portal By implementing the proposed fuzzy advisor system in the university portal, before enrollment of the next semester, the students are checked and those who should be excluded are determined and prevented from registration for the next semester. This advisor system addresses his/her issue according to rules and regulations. In section 6 some questions and answers, which were provided by the advisor and the student, are shown. 4. A look inside the system The proposed system is analyzed and designed by UML methodology and documents are generated with rational rose case tool. The software is built on 5-tier layers such that when one of the layers is reconfigured or rebuild, other layers don't change. The framework of system is shown in figure 2. Figure 2. System framework. a) Initially, the user selects the type of advisory service and enters crisp data in web application layer via a web browser. b) The input data related to the system is controlled for GPA, number of official notices received, educational level in university, criminal records (if any), type of illness in medical system (if any), etc. These are executed in Business Facade and business rules layer. c) A request for fuzzification the Crisp variables and rules generation is submitted to knowledge-base of the system by data access layer with ADO.NET. Then the linguistic variables are generated just by view select, stored procedure & user define function execution. This section makes a database abstraction and prevents SQL-injection. d) Linguistic variables are sent to inference engine and processed with mamdani model [11]. This section is accomplished by one stored procedure in database, named UstpInference. e) Fuzzy answers are defuzzified and crisp output values are generated. This step is accomplished by one-user defined function in database, named UdfDefuzzifier. Finally, the advisor system extracts answers to be shown to the user. Data object model of system is shown in Appendix 1. Client side (JavaScript & JQuery) Server side (ASP.NET & MVC & C#) Business layer (business façade & business rules) .NET Framework classes Common (enumerations & data objects & data sets) Data access layer (ADO.NET) Database & databaseoObjects (Stored procedures, user define functions, views) A Novel Web‐based Human Advisor Fuzzy Expert System, Vahid Rafe / 161‐168 Journal of Applied Research and Technology 165 Figure 3. Implementation architecture. 4.1 Implementation Environment The system is implemented in the 3 following layers: Web Application layer: This layer includes Web Forms, Web User Controls, Web Component (Infragestics Grid Control) and Model View Control (MVC). Business facade, business rules layer: Includes controls and business methods and attributes in C# classes:  Fuzzy_decision.cs,  Fuzzy_ruleInference.cs,  Fuzzy_set.cs,  Fuzzy_linguisticVariables.cs. Data access layer: The class Cls_DataAccess.cs supports ADO.NET 3 and connectionless performance. First of all, users login to the portal site and select the advisor system link and select their problems category. System asks the questions related to the problems. Collecting the answers provided by the user, the CHA translates the user inputs to linguistic variables, makes the fuzzy rules and generate the fuzzy answer. Then with Segono model the fuzzy answer are defuzzified to crisp output and is reported to user. The implementation architecture of the system is shown in Figure 3. In implementation architecture diagram 3 sources can submit requests to the web server user, knowledge engineer and web service). Then the web server controls the requests and sends the appropriate interface for this request. After entering user information, the fuzzy question generator creates the fuzzy questions for the user. User enters the inputs and submits the data to server. When the user inputs are received by the web server, with rule generator, the fuzzifier and knowledge base, the If- part for the rules is generated. Then the inference engine refers to the system Knowledge base and if a match is found for the pattern within a fuzzy statement by using the Mamdani model, the fuzzy answers for theThen-part of the rules are generated. In case that the input data does not match any patterns in the rule base, an appropriate message stating that the system cannot find the answer is displayed. If the fuzzy answer is found, the system transfers that to the deffuzzifier and finally the crisp answer is reported to the user in the desired format including MS Excel, HTML, XML, histogram and report file. This system can connect to another database for refining the output and generates the additional information. A Novel Web‐based Human Advisor Fuzzy Expert System, Vahid Rafe / 161‐168 Vol. 11, February 2013 166  4.2 Advantages of Fuzzy Expert systems The major advantage of these systems is that knowledge gradually turns into wisdom and can be used as a decision making tool in critical situations which replaces the conventional FAQ. Some other features are:  More accessibility: Many experiments can be done. Simply an expert system is a mass production of experiments.  Cost reduction: The cost of gaining experience by the user is decreased considerably.  Risk reduction: The expert system can work in environments dangerous, harmful or unpleasant for human.  Eternity: Obviously, these systems don’t die.  Multiple experts: An expert system can be the result of knowledge elicitation from several experts.  More reliability: These systems don’t get tired or sick, they do not go on a strike and they do not conspire against their managers. On the contrary, these are often done by human experts.  Explanation capability: An expert system can explain the way in which the results are obtained. On the contrary, due to many reasons (fatigue, unwillingness, etc.) human experts are not able to provide such explanations all the time.  Quick response: Expert systems respond quickly.  Responsibility in any condition: In critical conditions and/or emergencies an expert may be unable to make the right decision due to stress or other factors while an expert system’s decision making is not affected by these events.  Experience base: An expert system can provide access to a massive amount of experience.  User training: An expert system can act like an intelligent tutor, i.e., problems are presented to the system and the way of reasoning can be obtained.  Ease of knowledge transmission: one of the most important advantages of expert systems is its convenience to move the knowledge from the system to somewhere else on the globe. 5. One experimental result of common advisor system in university A student is going to be dismissed from the university and is going to lose a bachelor degree. The advisor asks a few questions to provide an answer. Advisor: How many semesters have gotten a GPA under 12? Student: 2 Advisor: How many undertakings have you been given? Student: 0 Advisor: How many disciplinary notices have you received from the university? Student: 1 Advisor: How many semesters have you passed successfully? Student: 11 Advisor: How many grades under 10 have you gotten? Student: 23 Advisor: How many course units have you passed out of 144? Student: 95 Advisor: What is your GPA? Student: 11.16 The crisp student's answers, fuzzy values and the linguistic values of each question are shown in table 1. Question Answer Fuzzy variable/(fv) )(x Linguistic Variable 1 2 0.69 High 2 0 0 Very Low 3 1 0.32 Low 4 11 0.89 High 5 23 0.96 Very High 6 95 0.63 Middle 7 11.16 0.31 Low Table 1. Fuzzification of the experimental input crisp variables A Novel Web‐based Human Advisor Fuzzy Expert System, Vahid Rafe / 161‐168 Journal of Applied Research and Technology 167 Inference phase: IF fv1 is high And fv2 is very low And fv3 is low And fv4 is high And fv5 is very high And fv6 is middle And fv7 is low THEN You are dismissed with a probability of 33 percent. ELSE Not in knowledge base. 6. Conclusion This paper glanced at the definitions and introductory concepts of fuzzy logic and fuzzy decision making and some implemented examples of such systems were presented. Finally, a web- based student consulting expert system was proposed and its capability in enhancing the consulting process has been shown. Acknowledgements This project supporting by the Islamic Azad University, South Tehran Branch References [1] M. Hassani Goodarzi, "Evaluating Students' learning progress by using Fuzzy Expert Systems", Conf. Rec. IEEE 5th Int. Conf. International Conference on Natural Computation (ICNC'09) and the 6th International Conference on Fuzzy Systems and Knowledge Discovery (FSKD'09), Tianjin, China , pp. 248-252, Aug 2009. [2] M. Hassani Goodarzi, " A web-based Implementation of a Portfolio Advisor System based on Fuzzy Expert Systems ", in Conf. Rec. IEEE 6th Int. Conf. Information & Communication Technology and System ICTS, Surabaya, Indonesia, pp. 15-22, Sep 2010. [3] M. Hassani Goodarzi, Vahid Rafe " Educational Advisor System Implemented by Web-based Fuzzy Expert Systems", in Journal of Software Engineerin and Application(JSEA), Volume. 5 No. 2, July 2012. [4] L. Teran, " A Fuzzy-Based Advisor for Elections and the Creation of Political Communities", in IEEE Journal 978-0-9564263-8/3, pp. 180-185, 2011. [5] H. S. Saini, R.Kamal, A. N. Sharma, "Web Based Fuzzy Expert System For Integrated Pest Management in Soybean", International Journal of Information Technology, Vol. 8, No. 1, pp. 55-74, August 2002. [6] Y. Min Huang, J. Nan Chen, Y. Hung Kuo, Y. -Lin Jeng, "An intelligent human-expert forum system based on fuzzy information retrieval technique", Elsevier, Expert Systems with Applications Volume 34, pp. 446– 458, (2008). [7] P. E. Merloti, "A Fuzzy Expert System as a Stock Trading Advisor", www.merlotti.com/EngHome/Computing/fes.pdf. [8] M.H. Fazel Zarandi,M. Khademian, B. Minaei-Bidgoli, "A Fuzzy Expert System Architecture for Intelligent Tutoring Systems: A Cognitive Mapping Approach", SciRes, Journal of Intelligent Learning Systems and Applications, 2012, 4, pp.29-40 doi:10.4236/jilsa.2012.41003 Published Online February 2012. [9] E.W.T Ngai, F.K.T Wat, "Design and development of a fuzzy expert system for hotel selection", Elsevier Volume 31, Issue 4, pp. 275-286, May 2003. [10] J. H. Marshall, Ung Chinna, Ung Ngo Hok, "Student achievement and education system performance in a developing country ", Springer, Educ Asse Eval Acc, DOI 10.1007/s11092-012-9142-x, February 2012. [11] W. Siler, J. J. Buckley. Fuzzy Expert Systems and Fuzzy Reasoning, JOHN WILEY & SONS, INC(ISBN 0- 471-38859-9,2005). A Novel Web‐based Human Advisor Fuzzy Expert System, Vahid Rafe / 161‐168 Vol. 11, February 2013 168  Appendix 1: System Object Model 
work_jheoydk3s5butm6zfif4wf6wnm	----	PII: 0004-3702(95)00009-7 ELSEVI.ER Artificial Intelligence 83 (1996) l-58 Artificial Intelligence Measures of uncertainty in expert systems Peter Walley * Department of Mathematics, The Universiry of Western Australia, Nedlands, WA 6907, Australia Received April 1991; revised January 1995 Abstract This paper compares four measures that have been advocated as models for uncertainty in expert systems. The measures are additive probabilities (used in the Bayesian theory), coherent lower (or upper) previsions, belief functions (used in the Dempster-Shafer theory) and possibil- ity measures (fuzzy logic). Special emphasis is given to the theory of coherent lower previsions, in which upper and lower probabilities, expectations and conditional probabilities are constructed from initia I assessments through a technique of natural extension. Mathematically, all the measures can be regarded as types of coherent lower or upper previsions, and this perspective gives some insight into the properties of belief functions and possibility measures. The measures are eval- uated according to six criteria: clarity of interpretation; ability to model partial information and imprecise assessments, especially judgements expressed in natural language; rules for combining and updating uncertainty, and their justification; consistency of models and inferences; feasibility of assessment; and feasibility of computations. Each of the four measures seems to be useful in special kinds of problems, but only lower and upper previsions appear to be sufficiently general to model t.he most common types of uncertainty. Keywords: Inference; Decision; Prevision; Bayesian theory; Dempster-Shafer theory; Belief functions; Possibility theory; Lower probability; Upper probability; Imprecise probabilities; Conditional probability; lndependerlce 1. Introduction My aim in this paper is to compare and evaluate mathematical measures of uncer- tainty that can be used in expert systems. The measures that I will consider are Bayesian probabilities, coherent lower previsions, belief functions and possibility measures. Spe- cial emplhasis is given to the theory of coherent lower previsions as this approach has * E-mail: peter@maths.uwa.edu.au. 0004-3702/96/$15.00 @ 1996 Elsevier Science B.V. All rights reserved SSDf 0004-3702(95)00009-7 2 P Walley/ArtQicial intelligence 83 (1996) I-58 received less attention than the others in the AI literature, although it is (in my view) the one that is most widely applicable. On a mathematical level, the other measures can be regarded as special types of lower or upper previsions. They can also be given the behavioural interpretation of lower or upper previsions. This point of view leads to some illuminating comparisons between the theories. For example, the four theories differ greatly in the calculus they use for defining, updating and combining measures of uncertainty, especially the rules they use to define conditional probabilities and expectations and how they model judgements of indepen- dence. The rules used in the theory of lower previsions, which are based on a general procedure called natural &ension, can be applied to belief functions and possibility measures, and thus they can be compared with the rules used in Dempster-Shafer theory and possibility theory. The paper is not intended to be a survey of the four theories from a neutral position and it is certainly not intended to be a wide-ranging survey of mathematical models for uncertainty. There is a substantial and quickly growing literature on numerical measures of uncertainty in expert systems. References [46,47,88] are good introductions to the relationships between probability measures, belief functions and possibility measures. Readers can learn more about the subject by consulting the surveys in [ 8,3 1,79,86,89] and the proceedings of the annual workshops on Uncertainty in Artificial Intelligence. I shall not discuss non-numerical methods of reasoning with uncertainty such as the theory of endorsements [ 131, default reasoning and nonmonotonic logics [ 59,63,67,68], com- parative probability [ 27,104] and modal logics [59,104]. Surveys of the non-numerical approaches can be found in [47,79,86]. I take it for granted that most practical reasoning involves uncertainty, partial ignorance and incomplete or conflicting information, and that it is often useful to formally model the uncertainty. This raises the questions (a) what is the best way to model uncertainty? (b) how should we assess, combine and update measures of uncertainty? and (c) how should we use the measures to make inferences and decisions? These questions are relevant to all kinds of reasoning with uncertainty, not only in expert systems. However expert systems are an especially good testing-ground for theories of uncertainty because they aim to formalise and automate as much as possible of the reasoning process. As an expert system is concerned only with a narrow domain of application, it is possible to formulate special assessment strategies, models and patterns of reasoning which are appropriate to that application. Many of the relevant uncertainties can be assessed by domain experts and these assessments can be encoded in the system. A user may be required to supply some further assessments but the expert system should be able to guide him through this process. The measures of uncertainty that are combined in an expert system may come from various sources. Some may be “objective” measures, based on relative frequencies or on well established statistical models. (In medical diagnosis, for example, there may be data concerning the frequency of a disease in a population, or the statistical association between a symptom and a disease.) Other assessments of uncertainty may be supplied by the domain expert (e.g., concerning the irrelevance of specific observations in diagnosis, or the association between a symptom and a disease when there is little statistical data). Indeed several experts may have contributed in building the system. Further assessments I! Walley/Art@cial Intelligence 83 (1996) l-58 3 may be made by the user of the system (e.g. about the uncertainties in the symptoms exhibited by a patient). All these measures of uncertainty need to be combined by the expert system to make inferences and decisions (e.g. to make a diagnosis for this patient). What, then, is the meaning of the combined measures of uncertainty? Whose uncer- tainties do they measure? It is the user of the system who will act on the conclusions of the system, provided he is satisfied that its uncertainty assessments are acceptable to him. One can regard the expert system as a consultant that supplies various models and assessments, elicits others from the user, combines all the judgements, and finally informs the user “if you accept all these judgements then you should draw these con- clusions”. So the expert system constructs a single model for uncertainty which the user then considers adopting as a model for his own uncertainty and as a basis for action. The expert system should be able to justify its assessments of uncertainty and its rea- soning procedures, when requested by the user, to make its model and conclusions more convincing. It should also be able to modify some of its assumptions and assessments if requeste:d by the user. In the end, the uncertainty measures on which conclusions are based mus,t be acceptable to the user. 2. Criteria for evaluating measures of uncertainty In the rest of this paper I will compare measures of uncertainty according to the following broad criteria. (a) Intl?);pretution. The measure should have a clear interpretation that is sufficiently definite to be used to guide assessment, to understand the conclusions of the system and use them as a basis for action, and to support the rules for combining and updating measures. (b) Imprecision. The measure should be able to model partial or complete ignorance, limited or conflicting information, and imprecise assessments of uncertainty. (c) Culculus. There should be rules for combining measures of uncertainty, updating them after receiving new information, and using them to calculate other uncer- tainties, to draw conclusions and to make decisions. Some justification must be given for the rules. Special attention will be given to the rules for computing conditional probabilities and expectations from unconditional probabilities. (d) Cortsistency. There should be methods for checking the consistency of all uncer- tainty assessments and default assumptions used by the system, and the rules of the calculus should ensure that the conclusions are consistent with these assess- ments. In the Bayesian theory and the theory of lower previsions, the intuitive notion of consistency is formalised in mathematical principles of coherence. (e) Assessment. It should be practicable for a user of the system to make (and feel comfortable with) all the uncertainty assessments that are needed as input. The system should give some guidance on how to make the assessments. It should be able to handle judgements of various types, including expressions of uncertainty in natural language such as “if A then probably B”, and to combine qualitative judgements with quantitative assessments of uncertainty. P Walley/Artijicial Intelligence 83 (1996) 1-58 (f) Computation. It should be computationally feasible for the system to derive inferences and conclusions from the assessments. Readers may wish to add other desiderata to this list. It does seem to me that it is essential for a theory of uncertainty to attempt to satisfy all six criteria. We might distinguish the first four criteria, which are theoretical, from the last two, which are practical. The first four criteria are “theoretical” in the sense that one would expect an adequate theory of uncertainty to show that they can be satisfied, irrespective of the specific application. The last two criteria are “practical” in the sense that they will be satisfied in some applications but not in others, depending on the type of model involved, the number of assessments needed, practical constraints of time and computing power, and the abilities of the user. The two practical criteria are obviously necessary if an expert system is to be im- plemented in practice. The first criterion, interpretation, seems essential in order to give meaning to, and to justify, the conclusions of the system. The second, imprecision, is necessary because partial ignorance and conflicting information are common in practice. A calculus is needed in order to derive conclusions from the uncertainty assessments, so the third criterion is needed. The fourth criterion, consistency, is needed to avoid erroneous and irrational conclusions. See [99] for further justification of these criteria, especially (a), (b), (d) and (e). There is a striking divergence between workers in probability and philosophy, on the one hand, and those in expert systems, artificial intelligence, computer science and engineering, concerning their attitude to these criteria. Naturally enough, the second group has emphasised practical criteria, especially computation (f), and has mostly recognised the need for imprecise measures of uncertainty (b) and natural-language judgements (e). They have given much less attention to issues of interpretation (a), consistency (d) and justification of the calculus (c) . The literature in probability and philosophy pays more attention to these theoretical issues, but less to the practical issues of assessment and computation. There is surprisingly little attention, in all the literature, to the problem of assessment. Much of the work in expert systems is typified by MYCIN, perhaps the best-known and most influential expert system, which was designed to help physicians diagnose bacterial infections of the blood [ 6,47,80,81]. In MYCIN uncertainty is measured in terms of certainty factors. MYCIN does well on the practical criteria (e) and (f), largely because of the modularity of its inference system (dependencies amongst variables are ignored) and the use of simple rules to combine certainty factors, but it does badly on the theoretical criteria (a), (c) and (d), largely because of the lack of a clear interpretation of certainty factors and a lack of justification for the rules of combination. (An earlier version of this paper [ 1001 contained a detailed comparison of certainty factors with other measures of uncertainty but this has been omitted on the advice of the referees.) Although all six criteria seem essential, I regard the first criterion-the need for a clear interpretation-as the most fundamental, because an interpretation is needed to support the rules of the calculus, to formulate principles of coherence (consistency), to guide assessment, and to understand conclusions. Criterion (a) is therefore a prerequisite for criteria (c), (d) and (e). (This point is well made in [ 331.) The importance of P: Walley/Art@ial Intelligence 83 (1996) 1-58 5 interpretation tends to be underrated by workers in expert systems, perhaps because it is unclear to them how an interpretation of uncertainty measures could be used to derive anrd justify rules for combining and updating the measures. Readers may find it illuminating to compare the theories of de Finetti [28,29] or Walley [99], in which a behavioural interpretation of linear previsions or lower previsions is used to support principles of coherence and hence to derive the entire calculus, with the literature on certainty factors and fuzzy logic, in which the rules of the calculus seem quite arbitrary. In the following sections I will outline de Finetti’s Bayesian theory and the theory of coherent lower previsions, which emphasise interpretation and consistency, followed by the Dempster-Shafer theory of belief functions, which emphasises a simple rule of combination, and finally consider the possibility measures used in fuzzy logic, which emphasises judgements in natural language. Although the four theories differ greatly in their interpretation of uncertainty measures, their methods of assessment and their calculus, all the measures can be regarded, from a mathematical point of view, as special types of coherent lower or upper previsions, and the four theories will be compared from this perspective. 3. Bayesian probabilities The most highly developed and best understood theory of uncertainty is the Bayesian theory. The book [56] is a good introduction to the theory and [28,29] are wor- thy of careful study. In the field of expert systems, [47] is a good introduction and [50,60,63,90] are important references. Bayesian probability models have been used in quite different ways in the expert systems PROSPECTOR [20,21], GLADYS [91] and MUNIN [ 43,47,50]. More recent applications include HUGIN [ 21, PATHFINDER [ 371, BAIES [ 141, and [38]. The discussion here will be brief, as the theory and its strengths and weaknesses are quite well known. The main aim is to summarise the Bayesian approach in a form that will illuminate the comparison with the other approaches. In the theory uncertainty is measured by unconditional probabilities P(A) or by conditional probabilities P (A 1 B), numbers between zero and one which are interpreted as fair betting rates. That is, the person whose uncertainties are being modelled is taken to be willing to bet on or against event A at rates arbitrarily close to P(A), or at rates arbitrarily close to P(A 1 B) on condition that the bet is called off unless event B occurs. The key assumption here is that a person should always have the same marginal rate for betting on or against an event. This assumption is called the Buyesian dogma of precision. Bayesians have given several arguments to support the assumption, but none of the arguments is at all compelling; see 199, Chapter 51 for a thorough discussion. Apart from countable additivity, all the familiar properties of probabilities can be derived from the behavioural interpretation and the dogma of precision, together with the coherence principle that there should be no finite combination of acceptable bets that is certain to produce a net loss. These assumptions imply that the probability function P is a normalised, nonnegative, finitely-additive set function. Further coherence principles imply that updated probabilities after observing an event B should agree with conditional 6 R Walley/Art@cial Intelligence 83 (1996) 1-58 probabilities P(A 1 B), and that these should be related to unconditional probabilities through Bayed rule P( A n B) = P( A 1 B) P( B) . If the unconditional probabilities P( A n B) and P(B) have been assessed and P(B) is nonzero then the conditional probability P(A 1 B) is uniquely determined through Bayes’ rule. Thus Bayes’ rule can be used as a rule for computing conditional probabilities from unconditional probabilities. It can also be used as a rule for computing P( A fl B) from P( A 1 B) and P(B) , and indeed it is frequently used for that purpose. If an unconditional probability measure P is specified, the prevision (or expectation) of a random variable X, denoted by P(X), can be computed from P(X) = J XdP. When the possibility space fi is finite, P(X) = CoEn P(w)X(w). Again these should be regarded as coherence relationships. They can be used to compute the value of P(X) from assessments of probabilities, but they could also be used to provide information about probabilities from assessments of previsions, as in [ 281. In general, assessments of previsions determine upper and lower probabilities rather than precise values. If coherent probabilities are specified for all events then, for every random variable X, there is a unique value of P(X) that is coherent with the specified probabilities. This means that, for Bayesians, no information is lost by modelling uncertainty in terms of probabilities; previsions are uniquely determined by probabilities. This result does not carry over to imprecise probabilities: upper and lower previsions are not determined, in general, by upper and lower probabilities. The Bayesian theory of probability is closely related to a theory of decision making. Suppose that the utility resulting from each feasible action a can be measured by a precise number U(a, w) which depends on the unknown state w. Define the random variable X, by X,(w) = V( a, w) . A Bayesian would compute the prevision P( X,), the expected utility of action a, for each feasible action, and attempt to choose an action to maximise expected utility. The Bayesian theory is applied in practice by selecting some events or variables whose (conditional) probabilities or previsions can be precisely assessed, adding any judgements of independence or exchangeability, and then applying the rules of the theory to calculate other (conditional) probabilities or previsions. For example, if {A,, AZ,. . . , Ak} is an exhaustive set of mutually exclusive hypotheses and B is an observable event, one might make precise assessments of the prior probabilities P(Ai) and likelihoods P( B 1 Ai) for 1 6 i < k, and use these to calculate the predictive prob- ability P(B) = cff, P( B 1 Ai) P( Ai). After event B is observed, provided P(B) is not zero, one might update uncertainty about the hypotheses by calculating their posterior probabilities using Bayes’ rule, P(Ai 1 B) = P(B I Ai)P(Ai)/P(B). (There is nothing in the Bayesian theory that forces one to calculate P(Ai I B) in this way-it might be easier to assess P (Ai I B) directly than to assess the quantities P( Ai) and P( B I Ai) .) Uncertain conclusions will typically be expressed in terms of posterior probabilities that are conditional on all the available information. When k = 2, Bayes’ rule can be written more conveniently in the form p(B) = PA(B), where p = P(Al)/P(Az) is the prior odds on Al, p(B) = P(AI I B)/P(Az I B) is the posterior odds on AI, and A(B) = P(B 1 Al)/P(B I AZ) is the likelihood ratio generated by B. That is, posterior odds = prior odds x likelihood ratio. l? Walley/Artijicial Intelligence 83 (1996) l-58 I In principle, the Bayesian approach can be applied in any problem involving uncer- tainty. In practice, it can be difficult to make the many precise assessments of probabil- ities that are needed to determine a complete probability model, and to check that the assessments are coherent and that they determine a unique probability model. Unless sufficiently many assessments are made, the probabilities of interest will not be precisely determined-we obtain upper and lower probabilities [60,62]. But if a larger number of assessments are made, so the probabilities of interest are overdetermined, typically the assessments will be incoherent. There are also computational difficulties in verifying whether a given set of probability assessments and independence judgements is coherent, which is equivalent to checking whether a system of linear and quadratic equations has a solution. To alleviate the difficulties of assessment and computation, the early expert system PROSPECTOR incorporated the simplifying assumption that separate pieces of evidence are probabilistically independent conditional on the hypotheses of interest, and used simple rules (similar to the max/min rules of fuzzy logic) to combine pieces of evidence. However these rules are inconsistent with the Bayesian calculus and they can produce incoherent probabilities. More recently, attention has been directed to special types of models, notably the “belief nel works”, “causal networks” or “directed acyclic graphs” studied in [ 50,63,90]. These models involve judgements of conditional independence, based on an expert’s understanding of the causal relations between variables, which can be represented graph- ically by directed trees and which are reasonable in many practical problems. For these models, the effort of assessment and computation is greatly reduced: assessments of conditional1 probabilities are needed only for the links in the tree, and the effect of new information on probabilities can be propagated locally. Belief networks can also be elaborated into “influence diagrams” by adding information about possible actions and utilities, and thereby used to make decisions. Assuming that precise probabilities can be assessed for all events, the rules of the probability calculus are uncontroversial. It is the assumption of precision that is unaccept- able. When there is little information concerning a possible event A it is inappropriate to assess any precise probability P(A). (Suppose, for example, that I produce an urn containing coloured balls. Without any further information about the balls, how would you assess a precise probability that the first ball drawn from the urn will be red? See [ 1011 for discussion of this example.) Similarly the Bayesian approach cannot deal with imprecise, qualitative or natural-language judgements such as “if A then probably B” [ 1041. Conclusion The Bayesian theory does very well on criteria (a), (c) and (d), but poorly on (b) and (e) . Bayesian probabilities have a simple behavioural interpretation. The rules of the probability calculus can be justified through this interpretation, and the rules guarantee consistency (coherence). Computations are feasible for some important types of models, notably for singly-connected belief networks. The theory is highly developed, especially for dealing with judgements of conditional independence [ 631, and useful in many practical problems. 8 I? Walley/ArttjIcial Intelligence 83 (1996) l-58 The fundamental difficulties with the Bayesian theory concern the dogma of precision. Because they demand precise probability models, Bayesians cannot adequately model ignorance, partial information, assessments of uncertainty in natural language, or conflict between expert opinions. There is a compelling argument for allowing imprecise assess- ments of uncertainty (see [99] for detailed discussion), and this has been accepted in much of the expert systems literature. In the rest of this paper I consider measures of uncertainty which do admit imprecision. 4. Coherent lower previsions This section summarises the theory of coherent lower previsions. The theory is de- veloped in detail in [ 991, using mathematical concepts suggested in [ 28,87,1 lo]. For related work in expert systems see [ 25,30,35,60,62,66,94,113], but note that these pa- pers (except the last one) differ in some important respects from the approach outlined here; they emphasise upper and lower probabilities rather than previsions, and they adopt a different interpretation which leads to a different concept of independence. Fuller discussion of many of the ideas in this section can be found in [99]. Assessment First consider how an expert system might, ideally, elicit assessments of uncertainty from the domain expert and the user. The system should give some guidance on how to make the assessments. It might suggest what probabilities could be assessed to constrain the probabilities of interest, or what kinds of independence judgements and probability models may be reasonable. But the user should not be forced to accept any of these judgements or to make assessments of any particular type. For example, the system may suggest default assumptions such as conditional independence, but if these are unacceptable to the user then the system should be able to operate without them, typically obtaining weaker conclusions. Generally the system should be able to work with whatever combination of judgements and expressions of uncertainty the user is able to make, including precise or imprecise assessments of unconditional or conditional probability, judgements in natural language such as “A is more probable than B”, “if A then probably B” or “A is very likely”, judgements of (conditional) independence, and various other kinds of judgements. The system would check whether these judgements are mutually consistent, combine them to construct an overall probability model and to draw conclusions about the questions of interest, and report the model and conclusions to the user. If the conclusions were indeterminate, the user might try to make further assessments in order to make the probability model more precise, but he may not always be able to do so. The assessment process involves a sequence of judgements. After each judgement, the expert system could compute and display summaries of the current probability model, and analyse the model to suggest what kinds of judgements should be considered next in order to reduce the indeterminacy in inferences or decisions. In the light of the current model, the user may choose to reconsider and modify earlier judgements, make further P. Walley/Arhjicial Intelligence 83 (1996) 1-58 9 assessments, update the model to take account of new information, refine or reformulate the possibility space, or terminate the process. Each of these steps modifies the current model in a simple way: see [99, Section 4.31 for the mathematical details. For example, the user may decide to retract some of the assumptions or judgements (a kind of nonmonotonic reasoning) because the current model is incoherent or has unacceptable implications, he may recognise that his previous possibility space is not exhaustive and decide tc’ consider other possibilities, or he may use imprecise probabilistic information about one possibility space to provide information about a second possibility space that is related. to the first space through a multivalued mapping. The key step in this process is the construction of an overall probability model from an arbitrary combination of uncertainty judgements. This can be carried out by the expert system, without further input from a user, through a mathematical procedure called natural extension. Interpretation Before we can give a formal definition of natural extension we must characterise the kind of probability model that we aim to construct. In fact there are several types of models for imprecise probabilities that are more or less equivalent, subject to appropriate consistency requirements [99, Section 3.81. (Lower and upper probabilities are not an adequate model in general, for reasons to be explained later.) The simplest model is a lower prevision I’, which is a real-valued function defined on the set C of all gambles. A gamble is a bounded mapping from the possibility space of interest, 0 (a set whose elements represent possible states of affairs or “possible worlds”), to the real numbers, and is interpreted as an uncertain reward in units of utility. The interpretation of the quantity p(X) is that you are disposed to pay any price less than e(X) for the gamble (uncertain reward) X. Loosely, we may call p(X) a supremum buying price for X: it is the supremum of prices which the model asserts that you are willing to pay for X. A conjugate upper prevision P is defined by P(X) = -p( -X). The interpretation is that you are disposed to sell the gamble X for any price greater than p(X). (The theory could be presented equivalently in terms of 7, but here we concentrate on P.) The model does not say anything about whether you will buy or sell X if the price lies between p(X) and H(X); either course of action may be reasonable. Similarly, a conditional lower prevision e( X 1 B) is interpreted as a supremum of buying prices for X that you would be willing to pay if you learned only that event B has occurred. Buying and selling gambles are somewhat artificial activities and they are introduced here merely to give a simple interpretation for p(X) and P(X). These quantities also have implications in more practical decision problems (outlined near the end of Section 4), and they could be interpreted in terms of their implications for other types of decisions. This interpretation of upper and lower previsions is epistemic and behavioural, but not necessarily subjective. In some problems, Z’(X) and F(X) can be given a logical interpretation, as marginal buying and selling prices that are uniquely determined by the available evidence; an expert system that is concerned with a sufficiently narrow domain 10 I! Walley/Artifcial Intelligence 83 (1996) 1-58 and that does not require any subjective input from users might encode a system of inductive logic [ 1011. The interpretation is epistemic in the sense that upper and lower previsions (like all the other measures considered in this paper) reflect a particular state of information and will usually change when more information is obtained or when information is reassessed. The quantities p(X) and P(X) need not be maximally precise. They may be merely lower and upper bounds for quantities that could be specified more precisely, in the same way that we could give lower and upper bounds for a person’s lifetime, e.g. 400 and 300 B.C. in the case of Aristotle. The values p(X) and P(X) could be generated by merely qualitative judgements, as in the football example below. This should make it clear that, contrary to common objections, the specification of P(X) and P(X) does not demand “twice as much precision” as a Bayesian model. Note also that, while it may be possible to sharpen the assessments of F’(X) and p(X), there is no reason to suppose that a precise assessment p(X) = P(X) could be made, any more than Aristotle’s lifetime could be represented by a precise point in time. Imprecise (upper and lower) previsions may be needed to model incompleteness or conflict in the available information. In particular there is no justification for a Buyesiun sensitivity analysis interpreta- tion of Z’(X) and F(X), which regards them as lower and upper bounds for some underlying linear prevision P(X) that is not known precisely. Upper and lower prob- abilities are similarly interpreted as upper and lower bounds for an unknown, precise probability value. This interpretation seems to have been taken for granted in most pre- vious publications on upper and lower probability, including those in the AI literature [ 25,30,35,36,41,60,62,66,94]. Upper and lower probabilities with this interpretation are sometimes called “probability bounds ” “probability intervals” or “generalised probabil- , ities”. (The term “upper and lower probability” also may be misleading if it suggests upper and lower bounds for a precise probability; this seems to be why Shafer [ 711 preferred the new term “belief function”.) The Bayesian sensitivity analysis interpretation is both misleading and unnecessary. It is misleading because, in most problems, no useful meaning can be given to the “underlying linear prevision”. It is unnecessary because upper and lower previsions can be given a direct behavioural interpretation, in terms of buying and selling prices for gambles or in terms of their implications in other decision problems, and the behavioural interpretation is sufficient to justify the axioms and calculus of the theory. The distinc- tion is important because the behavioural interpretation and sensitivity analysis lead to different methods for modelling independence and other structural judgements [ 991. One of the important contributions of the Dempster-Shafer theory is that it has emphasised that belief functions should not be given a sensitivity analysis interpretation [ 751. Every belief function can be represented as a lower envelope of a set of probability measures. This is merely a mathematical representation, however; it is misleading and unnecessary to regard a belief function as a lower bound for some unknown probability measure. In the same way, every coherent lower prevision can be represented as a lower envelope of a set of linear previsions, but this is no reason to regard the lower prevision as a model for partial information about an unknown linear prevision. In the theory of coherent lower previsions [ 991, all the axioms and rules are justified purely in terms of the behavioural interpretation, without appealing to a sensitivity R Walley/Artijicial Intelligence 83 (1996) l-58 11 analysis interpretation. (This includes the axioms for conditional previsions and the rules for conditioning or updating.) So the theory of coherent lower previsions does not rely in any way on a sensitivity analysis interpretation, and in this respect it does not differ from the Dempster-Shafer theory or possibility theory. Of course there are some examples where E does represent partial information about an unknown Bayesian probability measure and a sensitivity analysis interpretation is justified. The behavioural interpretation still applies in these cases and they can be modellecl by coherent lower previsions. But the behavioural interpretation is much more general. It can be applied, for example, to belief functions and possibility measures, for which the sensitivity analysis interpretation is usually inappropriate. This enables us, in Sections 5 and 6, to view multivalued mappings and inexact judgements in natural language as sources of coherent lower and upper previsions. A behavioural interpretation of these kinds of models need not exclude the interpretations they are normally given as “measures of evidential-support” or “degrees of possibility”, although it does seem somewhat more definite and useful. Coherence Coherence of the lower prevision e can be characterised by three axioms: (Pl) p(X) 2 inf{X(w): w E 0) for all X E L, (P2) p( AX) = Ap(X) for all X E 13, A > 0, (P3) p(X+ Y) 2 p(X) +p(Y) for all X,Y E L. There are various equivalent characterisations of coherence in [ 991. Wilson and Moral [ 1131 have expressed the coherence axioms in the form of a logic, with an associated proof theory and semantics, and this may appeal to readers who are familiar with classical logic. Coherence implies the inequalities (for all gambles X and Y) P(X) +P(Y) <p(x+Y) <P(X) +p(Y) <P(x+Y) <P(X) +B(Y). The three axioms are consistency requirements which can be justified through the behavioural interpretation of L. Axiom (Pl) asserts a willingness to pay any price less than the infimum possible value of X for the gamble X. (Recall that X is interpreted as an uncertain reward in units of utility.) This is reasonable because such a transaction is certain to increase utility. Axiom (P2) says, in effect, that a gamble 2 is acceptable if and only if AZ is acceptable for any positive constant A. Thus the acceptability of a gamble does not depend on the unit of utility in which its rewards are expressed. To justify the superlinearity axiom (P3), consider two acceptable gambles which involve paying up to p(X) for X and up to p(Y) for Y. The combined gamble is equivalent to paying up to P(X) +P( Y) for X+ Y. Hence E( X+ Y), the supremum buying price for X + Y, should be at least E(X) + E( Y). Axioms (P2) and (P3) are compelling only if gambles are expressed in terms of a linear utility scale. They would not be reasonable if gambles were expressed in monetary units, as the acceptability of a gamble might then depend on whether its units were dollars or thousands of dollars and the combination of two acceptable monetary transactions might not be acceptable. 12 P: Walley/Arti&ial Intelligence 83 (1996) 1-58 In general, conditional previsions will be specified as well as unconditional ones. Co- herence axioms that apply to general specifications of conditional previsions are given in [ 991. In the simplest case, where unconditional lower previsions &’ and conditional lower previsions e(. 1 Bi) are specified for each event Bi in a finite partition {Bt , B2,. . . , Bk}, coherence can be characterised in terms of axioms such as (Cl) P(X) 2 min{P(XI Bi),P(X 1 &),...,P(X 1 &)}, (C2) P(X) 6 m={P(X I h),F(X I &I,. . . ,p(X I Bd}, (C3) l’(Bi(X-P(XI Bi)))=Ofori=1,2,...,k. The Bayesian theory of de Finetti [28,29] is presented in terms of linear previsions, which are simply coherent lower previsions that satisfy the precision condition p(X) = p(X) for all gambles X. Linear previsions are maximally precise. At the other extreme are the vucuous previsions l’(X) = inf{X( w): w E 0) and F(X) = sup{X( w): w E a}, which maximise the degree of imprecision P(X) - p(X) and model complete ignorance about a. In general, the degree of imprecision in previsions can reflect both the amount of information on which they are based and the degree of conflict between different types of information (e.g. between the assessments of several experts, or between prior infor- mation and statistical data). In turn, greater imprecision in previsions leads to greater indeterminacy in conclusions (we may be unable to say which of two hypotheses is more probable) and greater indecision (we may be unable to say which of two actions is better). For example, if several experts assess precise previsions Pt (X) , . . . , P,,(X) then it is natural to define p(X) = min{Pi( X) : 1 < i 6 n} and F(X) = max{ Pi( X) : 1 6 i 6 n}. Then p is the most precise model that is acceptable to every expert (it represents the behavioural dispositions that are common to all the experts, a kind of “consensus”), and the degree of imprecision P(X) - p(X) measures the extent of conflict or disagree- ment amongst the experts concerning X [ 1031. Other ways of aggregating information from several sources are studied in [97; 99, Sections 4.3, 5.3 and 5.4; 1021. The combination of assessments from different sources is an important problem in expert systems. Any coherent lower prevision E can be written as the lower envelope of a closed convex set M of linear previsions, i.e. Z’(X) = min{P(X): P E M}, or as the lower envelope of the set of extreme points ext M, i.e. E(X) = min{ P (X) : P E ext M}. Often the simplest way to characterise the lower prevision E is by specifying the probability mass function (or probability density function) of each linear prevision in ext M. Now we return to the problem of natural extension. Suppose that a user makes finitely many judgements of uncertainty. The problem for the expert system is to combine these with any other judgements in the system (including those supplied by domain experts) to compute a coherent lower prevision P. This can be done in three steps: ( 1) translate the judgements into behavioural terms, hence into constraints on E; (2) check that all the constraints are mutually consistent; (3) compute the minimal coherent lower prevision e that satisfies the constraints. This procedure is illustrated by the following simple example. P Walley/ArtQicial Intelligence 83 (1996) 1-58 13 Example 1 (Football game). Consider a football game whose possible outcomes are win ( W), draw (D) or loss (L) for the home team. To express his uncertainty about the outcome, the user makes the judgements: (a) prclbably not W, (b) W is more probable than D, (c) D is more probable than L. To construct the natural extension of these judgements we first translate them into behavioural terms. We take (a) to mean that the user is willing to give up one unit of utility if W occurs, provided he gains one unit if not W. We will use de Finetti’s notation, in which the gamble that pays one unit if event W occurs and zero otherwise (the indicator function of W) is denoted also by W. Then (a) yields the constraint p( W) < i, or equivalently Z’( D + L) 3 i. According to (b), the subject is willing to pay one unit if D occurs, provided he receives one unit if W occurs. This yields the constraint Z’( W - D) 3 0. Similarly (c) yields Z’( D - L) 2 0. The natural extension of the three judgements can be computed by finding the set M of all probability mass functions (w, d, I) that are consistent with the judgements, by solving the system of linear inequalities: d + I 2 i, w 2 d, d 2 1, w 2 0, d 2 0, 1 > 0, w + d + I = 1. Because this system has solutions, the three judgements are consistent. The extreme points of M are the three probability mass functions (3,:) $), (i, 4 , 0) , (i, $, a). The lower prevision of any gamble X can then be calculated as the minimum expected value of X under these three mass functions. For example, EJW)=min{t,&,#}={, F(W)=max{4,i,$}=f. Many other kinds of qualitative or quantitative judgements could be added to the three we have considered, for example, (d) if rtot D then W is very likely, (e) W is between 1 and 2 times as probable as D, ( f) I am willing to bet on L at odds of 4 to 1, (g) W has precise probability 0.4. Some orher natural-language judgements are considered in Section 6 and further exam- ples are given in [ 991. In more complicated problems, one might also make judgements of conditional probabilities, conditional independence, positive or negative dependence, permutabi‘lity or exchangeability, intervals of measures, upper and lower density func- tions or distribution functions or quantiles. All these judgements can be regarded as constraints on E. They can be combined (in principle) by natural extension, although the linear programming problem may become intractable when many assessments are made, and then it may be necessary to restrict attention to special types of assessment or model. Note especially that ordinary-language judgements such as (a), (b) and (c) can be combined with numerical assessments; the two types of judgement occur together in many applications. How would Bayesians deal with the football example? They would require a user to make more precise assessments in order to determine a single probability measure that is consistent with judgements (a), (b) and (c). But he may be unable to go beyond these qualitative judgements, except by choosing arbitrary numbers which would not 14 P. Walley/Arti$cial Intelligence 83 (1996) 1-58 reflect his state of uncertainty about the game, because he lacks either information about the game or expertise in assessing probabilities. One popular method for selecting a unique probability measure from M is by max- imising entropy. This gives the mass function ($, 3, 3) in the football example. But there are alternative methods, such as assigning a second-order probability distribution on M, which yield different answers, and any choice of a single probability measure seems arbitrary. Note that maximum entropy does not distinguish the partial information provided by the three judgements from complete ignorance: the same probability model would be selected if we had no information at all about the game. No precise probability measure can reflect the imprecision of the three judgements. For discussion of maximum entropy, see [ 4232,991. Upper and lower probability By applying the behavioural interpretation of lower and upper previsions, the lower (or upper) probability of event A can be interpreted as specifying acceptable rates for betting on (or against) A. Consider a choice of whether to bet on or against A at betting rate x, meaning odds of x to 1 - x on A. You will bet OIE A if x is less than P(A), you will bet against A if x is greater than F(A), and your choice is not determined (it may be reasonable to choose either way) if x is between &A) and p(A). Upper and lower probabilities have the basic properties E(8) = p(8) = 0, where 0 denotes the empty set, P( 0) = p( 0) = 1, p(A) = 1 - E( AC), where AC denotes the complement of A (hence upper probabilities are determined by lower probabilities, and vice versa), 0 < l’(A) 6 p(A) < 1, and P(A)+P(B) <P(AUB) <P(A)+P(B) <P(AUB) <B(A)+B(B) when A and B are disjoint. These properties are necessary but not sufficient for coherence. The further property of 2-monotonicity, that p( A U B) +F’( A JIB) 2 c(A) +F’( B) for all events A and B, is sufficient for coherence but not necessary. Some of the mathematical theory of coherent lower probabilities simplifies when the lower probabilities are 2-monotone (also known as “Choquet capacities of order 2”); see especially the simple formulas for conditional probabilities and expectations at the end of this section, and the theory in [40,96,102]. One way of constructing a lower prevision R is to assess upper and lower probabilities p(A) and l’(A) for all events A and to construct _P_ by natural extension. Not all coherent lower previsions can be constructed in this way, because many coherent lower previsions can generate the same upper and lower probabilities. In the football example, suppose that the initial assessments are upper probabilities i, i, 4 and lower probabilities f , a , 0 for the events F D, L respectively. (These are the upper and lower probabilities generated by the three judgements considered earlier.) Their natural extension to a lower prevision &, calculated as in the previous example, is the lower envelope of the five probability mass functions (i,:, i), (i, i,O), (4, a, f), (i, i, i) and (5, i, 3). This is different from the lower prevision Et constructed earlier, e.g. if (X(W),X(D),X(L)) = (l,-1,O) then P,(X) =0 but&(X) =-i, although e, and & generate the same upper and lower probabilities for all events. l? Walley/Art$cial Intelligence 83 (1996) 1-58 15 Thus lower previsions (defined on all gambles) contain more information than lower probabilities (defined on all events), and information may be lost when uncertainty is modelled solely in terms of upper and lower probabilities. That is why we take lower prevision to be the fundamental model, rather than lower probability. Lower (or upper) previsions are more expressive than lower (or upper) probabilities. The extra information in lower previsions is often crucial in defining conditional upper and lower probabilities. Without it the conditional probabilities may be excessively imprecise. That can be seen in the football example. Taking B = {K D}, the first (lower prevision) model gives p(W 1 B) = $ and F(W 1 B) = 3, whereas the second (lower probability) model gives less precise values E( W 1 B) = $ and F( W 1 B) = 3. (In both cases the conditional probabilities were calculated by conditioning the extreme points and finding upper and lower envelopes. This agrees with natural extension.) In general, conditional upper and lower probabilities are not uniquely determined by unconditional upper and lower probabilities through the coherence axioms. But con- ditional upper and lower probabilities and previsions are uniquely determined by un- conditional lower previsions, through axiom (C3), provided the conditioning event has nonzero lower probability. Another reason for regarding lower previsions as fundamental is that they can often be assessed more easily than lower probabilities. If we are primarily concerned with a real-valued quantity whose value is uncertain, such as a person’s age, it may be easier to assess upper and lower previsions for the quantity rather than upper and lower probabilities for the events that the quantity belongs to particular sets. Lower previsions are also needed when we start by assessing upper and lower probabilities, then observe an event and condition by natural extension; again information may be lost if only the updated upper and lower probabilities are reported [41]. Thus upper and lower probabilities are often inadequate models for uncertainty. This is a defect of all theories of upper and lower probability, including the theory of belief functions and, on the interpretation adopted here, possibility theory. A more general theory, dealing with upper and lower previsions or an equally general alternative, is needed. In particular, upper and lower probabilities are inadequate for modelling natural- language mdgements of uncertainty. Natural extension Now consider the general procedure of natural extension. Suppose that a user makes finitely many judgements which can be translated into constraints on conditional lower previsions, say p(Xi I Bi) > pi for 1 < i < k, where ,zi are specified real numbers. (The case of unconditional previsions is included by taking Bi = R, and the case of precise judgements P( X I B) = p by taking p(X 1 B) > ,x and J’( -X I B) 2 -p.) Then the natural extension to any conditional lower prevision P( X 1 B) can be computed by solving a linear program, using the formula k ,UU: B(X-p) 2 CAiBi(Xi-pi) for some A; 20 (1) i=I 16 t! Walley/Art@cial Intelligence 83 (1996) l-58 where Y > Z denotes Y(w) > Z(w) for all w E 0, B and Bi stand for indicator functions, and p and pi denote constant gambles. This definition of natural extension can be justified through the behavioural interpretation of lower previsions: if you are willing to pay up to ,ui for Xi conditional on Bi then, by combining positive mul- tiples of such gambles, I can induce you to pay up to p(X ( B) for X conditional on B. The resulting value of p(X 1 B) is finite provided the initial judgements are consistent in the sense that they “avoid sure loss”, a much weaker requirement than coherence [ 991. Natural extension summarises all the inferences that can be derived from the initial judgements through the rules of coherence. In fact, the natural extensions e are the minimal lower previsions that satisfy the initial constraints and are coherent. There may be other coherent lower previsions p’ that satisfy the initial constraints, but they must dominate the natural extensions in the sense that E”( X 1 B) 2 I’( X 1 B) for all gambles X and all events B, and therefore they incorporate additional information that is not implied by the initial set of judgements. For example, Dempster’s rule of conditioning often disagrees with natural extension because it implicitly involves judgements of conditional independence that are not implied by the initial belief function and which may not be consistent with the initial belief function. Natural extension is a very general method of inference. Indeed the following impor- tant constructions can be regarded as special types of natural extension: construction of (upper and lower) expectations from probabilities; construction of conditional (upper and lower) probabilities from unconditional ones; the “Fundamental Theorem of Proba- bility” of de Finetti [ 281; construction of inner and outer measures; construction of joint probabilities from marginal and conditional ones; and construction of probability models from qualitative judgements. The initial constraints are quite general and this allows a user to express his uncertainty in whatever forms are most convenient, e.g. through qualitative judgements such as the natural-language expressions listed in Section 6, or through a combination of qualitative and quantitative judgements. The results of natural extension are also very general; it can be used to construct a lower prevision p( X 1 B) for any conceivable gamble X and event B, hence to construct upper and lower proba- bilities and preferences between gambles. Inferences can be made by computing lower previsions of important variables conditional on all available information, and decisions by computing lower previsions of differences between utility functions, both of which are linear programming problems. Alternatively the natural extension can be computed, as in the football example, by solving a set of linear inequalities to obtain the extreme probability mass functions that are consistent with the initial constraints, and then forming their lower envelope. This is the dual linear programming problem. In the special case where all the initial probability judgements are precise and no independence judgements are made, it is equivalent to what is known as “probabilistic logic” [60,62]. This alternative procedure breaks down when independence constraints are included because these are nonlinear. On the behavioural interpretation of lower previsions, a judgement that two events are independent means that the upper and lower probabilities of one event would not change if you learned whether or not the other event occurred. This is quite different from judging that the events are independent under some Bayesian P Walley/Artijicial Intelligence 83 (1996) 1-58 17 probability measure that satisfies the other constraints, which would be the appropriate definition of independence under a Bayesian sensitivity analysis interpretation [ 1051. C0nside.r the simplest case where upper and lower probabilities are assessed for two events A and B that are judged to be independent, say P(A) = cx, p(A) = 6, p(B) = p and p(B) = j?. Then the behavioural interpretation of independence imposes the 12 constraints P(A) 2 g, F’( A 1 B) 2 ct, l’(A 1 BC) 2 cy, . . ., P( BC 1 A) 2 1 - fi, Z’( BC ) AC) 3 1 - p, and the natural extension of the judgements can be computed by natural exlension of these constraints, using E!q. (1). Bayesian sensitivity analysis would model the judgements in a different way, by finding the: set of joint probability measures P which satisfy the constraints g < P(A) < 5, j? 6 P(B) < fi and P(A n B) = P(A) P( B). (The independence constraint is nonlinear and this makes the computations difficult in more complicated examples.) Upper and lower previsions are then defined as upper and lower envelopes of the set of solutions. The two approaches do produce different numerical answers; see the example of two unreliable witnesses in Section 5. All previous work in AI, e.g. [ 25,30,35], seems to have taken the sensitivity analysis definition of independence for granted, without considering the behavioural definition. A comparative study of the two approaches is needed, with regard to both interpretation and computational methods. Using the behavioural definition of independence, the computation of natural extension is more complicated when many judgements of conditional independence, or other structural ‘constraints on lower previsions, are made. Computations can be carried out, in general, by solving a finite sequence of linear programs with progressively stronger constraints. At each stage the independence and structural constraints are applied to the conditional lower previsions that were produced by natural extension in the previous stage, giving a stronger set of constraints for the next application of natural extension. It is not yet clear whether this is feasible for complex problems with moderately large numbers of conditional independence judgements, or for the types of belief networks that have been studied by Bayesians. There has been a substantial amount of work in recent years concerning the propaga- tion of upper and lower probabilities [ 25,30,94] or convex sets of Bayesian probability measures [ 10,941 using the sensitivity analysis definition of independence. Natural extension can be applied to a completely arbitrary collection of assessments, but the linear programming computations will become impracticable when the number of assessments is sufficiently large, especially when many independence judgements are made. (Of course Bayesian computations also become intractable in unstructured problems.:1 When this happens, several approaches might be considered. First, we could use simpler formulas, involving only local computations, to give lower bounds for E’( X 1 B) and upper bounds for Ti( X 1 B). Some examples of such formulas are given in the following subsections (see also [ 66,941) . This approach produces conclusions that are always valid but less precise than those produced by natural extension. Alternatively we might try to approximate p(X I B) without necessarily finding a lower bound. Some approximation methods are suggested in [60] for the case in which ,a11 the probability assessments are precise. This approach is less appealing than the first as it may produce invalid conclusions. 18 P Walley/Art$cial intelligence 83 (1996) l-58 Finally, the approach that seems most likely to be useful in practical problems is to develop special types of imprecise probability models, such as the 2-monotone lower probabilities or models defined in terms of upper and lower density functions or mass functions, for which natural extensions can be computed explicitly, without linear pro- gramming. Examples are given in the following subsections. The expert system INFERNO [66], designed to diagnose faults on oil rigs, uses upper and lower probabilities to measure uncertainty but differs in two important ways from the approach suggested here. First, INFERNO works by propagating simple con- straints on upper and lower probabilities. This has the advantage of simplifying and localising computations. However, because the propagation method and constraints are much weaker than those given by natural extension, the conclusions of the system, while valid, may often be too weak to be useful. Second, the system provides no way of updating probabilities by conditioning on new evidence. Any new information can only be regarded as specifying further constraints on a fixed, unconditional probability measure. Calculus All the rules of the theory follow from the principles of coherence and natural extension, and they can therefore be justified through the behavioural interpretation of lower previsions. An important example is the generalised Bayes rule (GBR) : when p is a coherent lower prevision defined on a sufficiently large set of gambles and p(B) > 0, the conditional lower prevision !‘(X 1 B) is the unique solution n of the equation &‘( B( X - x) ) = 0. This equation can sometimes be solved explicitly, and otherwise by simple iterative algorithms. (For example, take xa to be any estimate of x, and define asequence ofestimates by x,,+t =x,+2f’(B(X--x,))/(p(B) +P(B)). Then the sequence converges to the solution x, and the error is bounded by CCP where a = (p(B) - I’(B)) /(P( B) + f’(B)) .) Thus conditional previsions are uniquely determined by unconditional ones, provided the lower previsions involved in the GBR have been specified and J’(B) > 0. The GBR can be used to update the initial prevision p after learning that event B has occurred. When p is a linear prevision, the GBR reduces to P (X 1 B) = P (BX) /P(B) , a version of the usual Bayes’ rule. As another example of natural extension, suppose {Al, AZ,. . . , Ak} is an exhaustive set of mutually exclusive hypotheses, B is an observable event, and upper and lower probabilities P(Ai),I’(Ai),P(B 1 Ai),P(B 1 Ai) are assessed for 1 < i < k. (This generalises the Bayesian model considered in Section 3.) Assume that the assessments satisfy the coherence conditions 0 < p( B 1 Ai) < P( B 1 Ai) 6 1, 0 < p( Ai) < P(A;) < 1, and F(Ai) +Cj#eP(Aj) < 1 < P(Ai) +C,P(Aj) for every 1 < i < k. We might wish to calculate the natural extensions of these assessments to predictive probabilities J’(B) and P(B), and to posterior probabilities P(Ai I B) and P( Ai I B). To do so, it is convenient to define the extremal probability measures P which satisfy l’(A,i) < P( A,/) < B(A,i) for 1 Q j < k. Any ordering of the hypotheses, denoted by A,,,&,..., A,,, determines an extremal P as follows. Set P(A,) equal to p( A,j) if j < r, equal to p( A,j) if j > r, and intermediate between e( A,,,) and F( A,,, ) if j = r, where the values of r and P( A,,,) are determined by $=i P(A,,) = 1. P Walley/Art@cial Intelligence 83 (1996) l-58 19 Using [99, Theorem 6.7.21, the natural extension can be written as P(B) = C;=l P(13 I Aj)Pl (Aj)T a weighted mean of the values P( B 1 A,i), where Pt is the extremal probability measure determined b r- ordering the hypotheses to have decreasing values of F’( B 1 Al). Similarly fr( B) = cj=, P (B 1 Aj) P2( Aj), where P2 is defined by ordering the hypotheses to have increasing P( B I Aj). (These and the following for- mulas simplify when Z’( A,!) = P( Al) for all j, as then PI (Aj) = P2( Aj) = e( Aj) .) In general J?(B) is bounded by C;=t P(B I Aj)P(Aj) < P(B) 6 & P(B I A,/)P(Aj), but these bounds may be far from sharp. Using a result from [99, 8.5.41, the posterior lower and upper probabilities after observing B are PC4 I B) = P(B I Ai>P3(Ai) P(B I Ai>& + Cj#P(B 1 AjJP3CA.i) ’ (2) where 91 is defined by setting nl = i and ordering the other hypotheses to have increasing p( B 1 A,i), and F( Ai 1 B) = - P(B I Ai)fi(Ai) f’(B I A,)P4(Ai) + Cj#iP(B I Aj)fi(Aj) ’ where P4 is defined by setting nk = i and ordering the other hypotheses to have decreasing P( B 1 Al). In this problem the computations are quite simple. If needed, simpler bounds could be obtained by substituting either p( AJ) or P( Aj) for P3 (Aj) and Pd(Aj) in Eqs. (2) and (3). A special case of these results was used in [25]. More general models involving belief networks are studied in [ 941. Other results, allowing general prior upper and lower previsions concerning the hypotheses Aj, are in [ 96,991. These formulas simplify in the case where there are only two hypotheses, Al and A?. Let p = 7’( Al > /I’( AZ) and p = &( Al ) /P( AZ) denote the ptior upper and lower odds on AI, and let x(B) = F(B-1 Al)/l’(B I AZ) and A(B) = P(B I Al)/itr(B I AZ) be the upper and lower likelihood ratios generated by B. The posterior upper and lower probabilities of the hypotheses are determined by the posterior upper and lower odds on A 1, which are given by the multiplicative formulas F(B) = F’(A, 1 B) _- f’(A2 1 B) =pA(B)’ These generalise the Bayesian formula: Conditional probabilities P(AI I B) p(B) = = P(A2 I B) = phW. posterior odds = prior odds x likelihood ratio. Suppose that coherent upper and lower probabilities, P and p, are specified for all subsets of a, and that we wish to construct conditional upper and lower probabilities B(. ) B) and F’(. ( B), e.g. to update beliefs after observing the event B. This problem is of interest because the other theories examined in this paper (Bayesian, belief func- tions and possibility theory) attempt to define conditional probabilities and expectations in terms of unconditional probabilities. (This seems to be a hangover from the classical 20 P. Walley/Art@cial Intelligence 83 (1996) 1-58 theory of probability.) It is important to recognise, however, that this is a special case of a much more general problem. In general we will make whatever probability assess- ments we can and use natural extension to construct other probabilities and previsions; it may often be easier and more informative to assess some conditional probabilities and previsions directly than to assess all unconditional probabilities (e.g. consider the problem in the previous subsection). So the problem considered here, in which uncondi- tional upper and lower probabilities are assessed for all events but no other assessments are made, should be regarded as atypical. Let B be any subset of 0 such that f’(B) > 0, and let M denote the set of all probability measures P such that P(C) > p(C) for all C C 0. Then lower probabilities conditional on B can be computed by applying the general formula for natural extension, giving P(A 1 B) =SUp { PU: B(A-,u) > kA,i(Ai-P(Ai)) i=l for some n 2 0, Ai G 0, hi 2 0 1 =inf{P(A fl B)/P(B): P E M}. (4) Provided fi is finite, P(A 1 B) can be computed by linear programming techniques. (Much simpler formulas can be used in the special case where E is 2-monotone, discussed below.) A more general formula, which applies whenever P(B) > 0, is &(A ) B) = inf{P(A fl B)/P(B): P E M, P(B) > 0). The conditional upper probabilities are defined by F(A 1 B) = sup{P(A n B)/P(B): P E M, P(B) > 0). The conditional probabilities p( e 1 B) and P(- 1 B) defined by these formulas are always coherent lower and upper probabilities. Thus conditioning by natural extension preserves coherence. Moreover, &a 1 B) and B( - 1 B) are always coherent with the unconditional probabilities e and P. We will see that the families of 2-monotone lower probabilities, belief functions and possibility measures are also closed under conditioning by natural extension. Provided p(B) > 0, the conditional probabilities satisfy the following inequalities: P(A n B) <P(A I B) < min J’(AnB) F(AnB) P(AnB)+P(ACnB> F’(B) ’ P(B) ’ P(AnB) >P(A I B) 2 max P(AnB) B(AnB) P(AnB)+lJACf7B) Z’(B) ’ B(B) > ’ These hold whenever p and P are coherent, but the lower bound for P(A I B) is actually achieved whenever p has the stronger property of 2-monotonicity. In that case the natural extensions are R Walley/Artifcial Intelligence 83 (1996) I-58 21 &PI 1 B) = Pt.4 n B) P(AnB) +P(ACnB)’ P(P, 1 B) = _ F(AnB) P(AnB) +P(ACnB)’ (5) provided Z’(B) > 0 [9,15,23,96]. (The same formulas apply when F(B) > 0 and E(B) = 0, provided the denominators are nonzero; set P(A 1 B) = 1 if the first denominator is zero, and H(A 1 B) = 0 if the second denominator is zero. This case is uninteresting because &A ] B) = 0 and p(A 1 B) = 1 for every A that satisfies p( A n B) > 0 and F( AC n B) > 0.) Numerical examples of this rule will be discussed in Sections 5 and 6. It has been generalised in [96,102,108] to characterise the posterior probabilities generated by a statistical likelihood function and prior lower probabilities that are Z!-monotone. When p is 2-monotone, the conditional lower probabilities P(. 1 B) are 2-monotone as well as coherent. Thus conditioning by natural extension preserves 2-monotonicity. (See [96] for a proof.) Expectations Again suppose that coherent upper and lower probabilities are specified for all sub- sets of 0. We wish to construct upper and lower previsions or “expectations”, F(X) and P’(X) , for gambles (bounded random variables) X. For example, in order to make decisions we would need to compute upper and lower previsions of differences between utility functions. But again we must point out that the case considered here, in which unconditional probabilities are assessed for all events but no other previsions are di- rectly assessed, is atypical; especially as some judgements, such as the natural-language judgements in the football example, cannot be modelled adequately in terms of upper and lower probabilities. Again the upper and lower previsions can be computed from the general formula for natural extension ( 1)) giving p(X)=SUp ,UU: X-p>k&(Ai--P(Ai)) forsomen>O, AiG 0, Ai>O i=l = inf (6) P(X)=inf /LU: X-/L<eAi(Ai-P(Ai)) forsomen>O, AiLa, Ai>O i=l = sup {I XdP: P E M . > (7) Again these formulas simplify in the special case where the lower probability E is 2-monotone. Define F;x and &, the upper and lower distribution functions of X, by 22 R Walley/Art@cial Intelligence 83 (1996) 1-58 Fx(x) = P({,: X(w) < x)1 and &(x> = p( {w: X(o) < x}). Provided c is 2-monotone, the natural extensions can be written as Choquet integrals [96] M 02 P(X) = s xdFx(x), P(X) = s x@,(x). (8) -co --oo Decision making Upper and lower previsions are used to make decisions in the following way. Suppose we need to choose an action from a finite set of possible actions {al, a2,. . . , ak}, where the utility U( a, w) of action a depends on the unknown w. (We assume that utilities are specified precisely; otherwise the decision problem is more complicated.) Define a gamble Xj by Xj( w) = U(aj, W) for each j = 1,2,. . . , k. TO compare two actions ai and a,j we compute the upper and lower previsions P( Xi - Xj) and p( Xi - X,j) based on all available information. Then action ai is preferred to aj if &(Xi - Xj) > 0, a,j is preferred to ai if P(Xi - X,j) < 0, and if neither condition holds there is insufficient information to determine a preference. Say that action ai is maximal if there is no other action aj that is preferred to ai. The non-maximal actions can be eliminated, and it is reasonable to choose any of the maximal actions. For more details and applications to real decision problems see [99, Sections 3.9 and 5.6; 101; 1031. This method produces only a partial preference ordering in general, and there may be more than one maximal action. Methods for generating complete preferences are discussed in [57,85,99]. All such methods seem somewhat arbitrary. It should be ac- knowledged that, when probability judgements are imprecise, there may be more than one reasonable course of action. Imprecise conclusions The inferences produced by natural extension are often imprecise. That can be seen in simple examples, e.g. if A and B are logically independent events and an expert assesses that each has precise probability i but makes no judgement about their degree of depen- dence, then the natural extension to upper and lower probabilities for their intersection is p( A n B) = i, E( A n B) = 0. When natural extension is used to compute conditional upper and lower probabilities from unconditional ones, the conditional probabilities may be highly imprecise, especially when the initial probabilities are 2-monotone. (Two ex- amples are discussed later in the paper.) In such cases, other methods of conditioning such as Dempster’s rule often produce conditional probabilities that are more precise. Advocates of the alternative methods have suggested that natural extension tends to produce excessive imprecision. Wilson and Moral [ 1131 have given an interesting example of this, also discussed in [ 1201. Suppose that an expert makes two assessments of conditional lower probabilities: E( B 1 A) = 1 and Z’(C 1 B) 2 0.999, where A, B and C are logically independent events. What do these judgements imply about p(C 1 A)? P WaNey/Arttj?cial Intelligence 83 (1996) 1-58 23 It may seem that we can make a fairly strong inference in this case, because the second judgement is “close to” P(C 1 B) = 1 which, together with P(B 1 A) = 1, would yield the inference E( C 1 A) = 1 by natural extension. (To see that, apply the general formula for natural extension ( 1), and use A n Cc C (A rl BC) u (B n Cc) to show that A (C - 1) 2 A (B - 1) + B (C - 1) .) However, natural extension of the expert’s judgements produces only the trivial inferences p( C 1 A) = 0 and P( C 1 A) = 1. These inferences may indeed seem “excessively imprecise”, but I believe that they are reasonable because nothing more can be derived from the two explicit judgements without making further assumptions. Suppose that an alternative method of inference produced lthe conclusion p(C 1 A) 2 S from the two judgements, where 6 is a specified positive number. The expert might then make some further assessments about the events, Suppose he makes the further judgements that P(A) > 0.001 and P(A rl C) = 0. These judgements are perfectly consistent with the two initial judgements, since all four judgements are consistent with a Bayesian probability measure P that satisfies P(B) = 1, P(AflB) =P(A) =O.OOl,P(BnC) =P(C) =0.999andP(ArlC)=O.Howeverthe new judgements are not consistent with the inference p(C ( A) > S because they imply F(C I A) = 0. This shows that the inference goes beyond the information contained in the two initial judgements; it relies on extra (implicit) assumptions which may be inconsistent with the expert’s other beliefs. This argument can be generalised as follows. Let Dt and V2 denote two sets of judgements and let Et denote a set of inferences produced from Dt by applying the rules of the calculus. Then we require the following consistency principle: if the overall set of judgements (Dt U 2)~) is consistent then (Et U VT) should be consistent. The force of this principle depends on the technical meaning that is given to “consis- tency”. In the theory of lower previsions “consistency” is identified with “avoiding sure loss”, the rules of natural extension do satisfy the consistency principle, and they are the strongest rules which do so. That is, the inferences given by natural extension are the most precise inferences possible, if the consistency principle is to be satisfied. It seems, therefore, that natural extension produces exactly the inferences that are implied by the explicit judgements and assumptions. Inferences may be excessively imprecise, not because natural extension is the wrong method of inference, but rather because the judgements and assumptions are excessively imprecise. Indeed the compu- tation of natural extensions will often reveal indeterminacy in conclusions or decisions that compels us to make our judgements more precise. For example, the problem that unconditional upper and lower probabilities tend to generate very imprecise conditional probabilities can sometimes be resolved by assessing upper and lower previsions, which are more informative than probabilities and therefore generate more precise inferences. One way to sharpen inferences in many problems is to add judgements of condi- tional independence or dependence. Consider again the judgements P(B I A) = 1 and p( C I B) 2 0.999. The natural extensions must satisfy the coherence conditions P(CIA)>P(BnCIA)>R(CIAnB)P(BIA), hence p( C I A) 2 p(C ) A n B) in this case. This is useful only if we can relate p(C I An B) to P’(C I B). One way to do so is to judge that A and C are conditionafly independent given B, which gives p(C I A n B) = p(C 1 B) 2 0.999 and hence 24 F! Walley/Art$cial Intelligence 83 (1996) l-58 I’(C 1 A) 3 0.999, a very strong conclusion. The same inference is produced by the weaker judgement that A and C are nonnegatively correlated conditional on B, so that P(C 1 A n B) > p(C 1 B). Either condition suffices to rule out the possibility that A and C are essentially incompatible events, and this must be ruled out before any nontrivial inference can be obtained. Another way to sharpen inferences in this problem is to make a further numerical assessment of &(A 1 B). Coherence requires that p(C 1 A) Zp(C IAnB) aP(AnC 1 B)/(fJAnC 1 B) +F(AnCC 1 B)). Here Hence P( C I A) 2 1 - O.OOl/p( A I B), and any assessment of f’(A I B) greater than 0.001 will produce a nontrivial lower bound for &( C 1 A). For example, the judgement that Z’( A 1 B) 2 0.01 gives the strong conclusion E( C I A) > 0.9. It appears that, in moderately complex expert systems, judgements or assumptions of conditional independence are needed to reduce the effort of assessment and produce useful conclusions, Many expert systems use expert knowledge about causal relationships to build “belief networks” based on assumptions of conditional independence. In other systems independence constraints are taken as default assumptions; if the expert or user supplies no information about the relationship between two events or variables then they are assumed to be independent. Such default assumptions may sometimes be needed to produce useful conclusions, but it is important that they always be made as explicit as possible (they may be inconsistent with an expert’s other beliefs), that users be encouraged to consider whether they are reasonable in a particular application, and that they can be easily retracted if the overall set of judgements becomes inconsistent. Ideally, inferences should be computed both with and without default assumptions so that a user can compare their effects. A logic of default assumptions that is compatible with the theory of lower previsions is outlined in [113]. Conclusion The theory of coherent lower previsions is a general theory of reasoning in the presence of uncertainty and partial ignorance. Lower previsions are more general and more expressive than lower and upper probabilities. They have a simple behavioural interpretation as supremum buying prices for gambles and they should not be interpreted, in general, as lower bounds for an unknown Bayesian prevision. The theory certainly satisfies criteria (a)-(d) of Section 2. Lower previsions have a clear behavioural interpretation which supports the principles of coherence and natural extension. All the rules of the theory can be derived from these principles, and they can be used to check consistency of the initial assessments and to ensure consistency of assessments with conclusions. The imprecision of lower previsions can be used to model P. Walley/Artifcial Intelligence 83 (1996) 1-58 25 a lack of information, conflict between several types of information or between expert opinions, or the vagueness of probability judgements in natural language. Coherent models can be produced by combining qualitative (natural-language) judgements with precise or imprecise numerical assessments. Because lower previsions have a behavioural interpretation, it is quite easy to understand the practical meaning of conclusions that are expressed in terms of them, and to use them in making decisions. The task of assessment can be handled, in principle, by allowing the user to make whatever judgements he finds most comprehensible and natural from a wide variety of admissible judgements. In particular he can express his uncertainties in ordinary language; some ways of doing so are discussed in Section 6. Other important sources of coherent lower previsions include multivalued mappings (Section 5)) partial information about precise probabilities, combination of expert opinions [ 1031 and various models based on statistical data [99,101,102]. It seems that, in complex problems, assumptions of independence or conditional independence will be needed to reduce the effort of assessment and produce useful conclusions. Further work is needed to compare several ways of modelling independence judgements, to study how independence can be used as a default assumption in expert systems, and to determine its effect on the precision of coaclu,sions. The genera1 method for making inferences and decisions in the theory is natural extension. In general, the computation of natural extension can be reduced to a linear programming problem, or (in the case of independence judgements) to a finite sequence of linear programs. Again, these problems may be intractable in moderately large ex- pert systems when many assessments are made, and then (as in the Bayesian theory) special types of models are required. Again further work is needed to find computation- ally efficient methods for computing natural extensions, especially when independence judgements are involved, to develop tractable types of models (e.g. using 2-monotonicity or upper and lower densities), and to develop efficient methods for propagating lower previsions in belief networks. 5. Belief functions The th’eory of belief functions was initiated by Dempster in a series of papers in the 1960s and developed by Shafer [ 711. Its relevance to expert systems is discussed in [ 34,741]. For more recent developments see [64,75] and the ensuing discussion [ 19,65,76,84,92,107,112]. Smets has developed an interesting variant of the theory called “the transferable belief model” [ 82-851. Applications of belief functions in expert systems include OASES [5] and its shell [4], MacEvidence [ 391, PSEIKI [44,47], PULCINELLA [69] and [49,58]. A belicffunction p is a real-valued function, defined on all subsets of a possibility space 0, which can be written in the form P(A) = CBCA m(B) for all subsets A, where m is a probability mass function on subsets of 0, i.e. m(0) = 0, m(B) 2 0 for all B C 0, and ~ecnm(B) = 1. Here C denotes set inclusion, not necessarily strict. (In Smets’ theory m@) may be positive.) The conjugate upper probabilities are defined by B(A)= 1 -f'(AC)=~,,,gm(B). 26 P Walley/Art$cial Intelligence 83 (1996) l-58 The mass function m is called the probability assignment for E. It is determined by p through the Mobius inversion formula m(B) = C,,,( -1) lB-Alp( A). Any lower probability function p determines a function m through-this formula, and p is a belief function if and only if m is a probability mass function. One can think of m(B) as a fluid probability mass that is free to move to any element of B. Bayesian probability measures are a special type of belief function for which m(B) = 0 unless B is a singleton set. The vacuous lower probability is a belief function, defined by m( 0) = 1. Interpretation The natural extension of a belief function E to a lower prevision is defined for all gambles X by p(X) = xBcam( B) inf{X( w): w E B}. It is easily verified that p satisfies the coherence axioms (Pl)-(P3) in Section 4, and it follows that every belief function is a coherent lower probability function. So belief functions can be given the behavioural interpretation of lower probabilities: P(A) is a supremum of acceptable rates for betting on event A. Various other interpretations of belief functions have been discussed in [36,63- 65,72,75,76,85,104] ; see [ 831 for a survey. Shafer prefers to interpret belief functions by drawing an analogy with a “canonical example” of a “randomly coded message” [ 72,751. In Shafer’s canonical example, the belief function is generated from an under- lying precise probability measure through a multivalued mapping (defined below). The underlying probability measure does appear to have a behavioural interpretation in terms of rational betting rates, and the belief function inherits this behavioural interpretation through the multivalued mapping [99, p. 1821. So it seems to me that a behavioural interpretation of belief functions is not only compatible with Shafer’s interpretation, but also a necessary consequence of his interpretation. Indeed, I regard Shafer’s semantics and the other interpretations of belief functions as possible ways of elaborating the behavioural interpretation. On [ 99, pp. 20 and 611, I call the behavioural interpretation “minimal” because it is compatible with a wide variety of elaborations. It requires that belief functions have certain implications for betting and other decisions, but it does not exclude other semantics which relate belief functions to the evidence on which they are based. Shafer [75] describes the random coding example as a “metaphor” which may pro- vide some guidance in constructing belief functions. The behavioural interpretation is concerned with how belief functions are used in making decisions. The two types of interpretation are compatible and I think that both are needed. (I prefer to call Shafer’s canonical example an “assessment strategy” rather than an “interpretation” of belief functions, but that is not to deny its utility.) In practice we need to construct belief functions from evidence, but we also need to use belief functions to make decisions and this seems to require some kind of behavioural interpretation [ 331. Without one, the practical meaning of inferences that are expressed in terms of belief functions is somewhat unclear. See [ 1071 for an interesting comparison of the two interpretations. In [ 721 Shafer does accept the behavioural interpretation of belief functions, although he argues that other aspects of their meaning are more important. In [75] he argues vehemently against a Bayesian sensitivity analysis interpretation of belief functions, I? Walley/Art@cial Intelligence 83 (1996) l-58 27 but that argument is irrelevant to the present discussion as I also reject the sensitivity analysis interpretation. The theory of coherent lower previsions and the rules of natural extension rely only on a behavioural interpretation. Of course I am not claiming that the whole Dempster-Shafer theory is compatible with a behavioural interpretation of belief functions. Much of the theory is based on Dempster’s rule for combining belief functions, and in many problems Dempster’s rule produces inferences that are unacceptable under a behavioural interpretation. Some presentations of the theory [75,76] give the impression that Dempster’s rule follows naturally from the “random coding” or “multivalued mapping” semantics for belief functions, together with an inocuous assumption of unconditional independence. In fact, even under Shafer’s own semantics, the justification of Dempster’s rule relies on stronger assumptions of conditional independence which seem to be unreasonable in many applications. (These assumptions are discussed later in this section.) Nor do the other interpretations of belief functions provide a convincing justification for Dempster’s rule. Shafer’s semantics and the concept of a multivalued mapping are compatible with a behavioural interpretation, but the indiscriminate use of Dempster’s rule is not. Assessment Belief functions are often generated by a multivalued mapping [ 151, which is a mapping A. from points of an underlying space P = ($1,. . . , &} to subsets of 0. For simplicity I assume that the sets A( @I ) , . . . , A( en) are distinct. It may be possible to assess a Bayesian probability measure P on !P, using either frequency information or subjective judgement, and this induces a belief function on 0 through m( A(q$)) = p($i). Example Z! (An unreliable witness). In this simplest example, an unreliable witness claims that he observed an event C. Either (@I ) he did observe C, so A(& ) = C, or (& ) he observed nothing, so A( &) = R. Suppose that, after hearing his report, we judge the witness to have credibility a, so P( $1) = (Y and P (I&) = 1 -a. This generates the probability assignment m(C) = (Y and m( 0) = 1 - (Y, and the corresponding belief function has P(C) = cy and Ti( C) = 1. Thus there is some evidence in favour of C (we would bet on C at any odds better than 1 - (Y to (Y), but no evidence against C (we are not prepared to bet against C at any odds). The imprecision of this belief function simply reflects the absence of information about the “base rate” frequency of C, P(C 1 &). A Bayesian would need to make a precise assessment of P(C 1 $2) and then compute P(C) = (Y + (1 - a)P(C 1 92). Depending on the practical context, there may or may not be information on which to assess P( C 1 t,&t). In general, we may only be able to assess lower probabilities c:<C 1142) andP(fit) =LX, and then weobtainP(C) =g+(l-(y)p(C 11,bz) and P(C) = 1 by natural extension. Note that Bayesians require two precise assessments, of a and I’( C 1 CCI;! ) , the multivalued mapping requires one precise assessment (a), but we can reach useful conclusions without making any precise assessments. 28 R Walley/Artijicial Intelligence 83 (1996) l-58 Belief functions can often be assessed by using the model of a multivalued mapping and assessing underlying Bayesian probabilities. (Although it is somewhat ironic that precise assessments should be needed in a theory of imprecise probabilities! A more general model, which does not assume that precise probabilities can be specified on the underlying space P’, is given in [99, Section 4.3.53.) In other problems we may be able to assess the probability assignment m more directly. For example there may be frequency information about the outcomes Bt , . . . , B, of previous trials, where the outcome is recorded as a subset Bi of 0 rather than a single element. We might then take the past relative frequency of B as our assessment of m(B) [ 19,104]. Some other assessment strategies have been suggested in [ 71,73,75]. Example 3 (Football example). It is important to recognise, however, that there are many types of information which cannot be modelled by belief functions. One simple example is the football example of Section 4, where three judgements were expressed in ordinary language. Let p be the lower probability function constructed by natural extension of the three judgements, which is the lower envelope of the three probability mass functions ($,i,i), (i,i,O), (i,i,i) onR={wD,L}.BytheMGbiusinversion formula, m(W=P(Q -P(WUD) -P(WUL) -P(DUL) +p(W) +p(D) +p(L) &$;_;+;+~+o=-~. As m( L?) is negative, m cannot be a probability mass function so P is not a belief function. It appears that probability judgements in natural language cannot be modelled, in general, by belief functions. Note also that, even when such judgements do yield a belief function, this may be less informative than the appropriate lower prevision. (As explained in Section 4, lower probabilities are less expressive than lower previsions.) Even when a belief function p is generated by a multivalued mapping, we may be able to make further assessments to sharpen p, and the resulting lower probability may not be a belief function. There are other examples in [48,64,96,99] of models and assessment strategies which produce coherent lower probabilities that are not belief functions. (Consider, for example, two tosses of a fair coin with unknown correlation between the outcomes.) There seems to be no good reason to restrict attention to belief functions, rather than coherent lower probabilities or coherent lower previsions. Nor is it clear why unconditional belief functions are taken to be the fundamental measures of uncertainty. Direct assessments of conditional lower and upper probabilities are often needed to measure the uncertainty associated with the “if-then” rules that are prevalent in expert systems [ 641. Dempster’s rule of combination Dempster’s rule is extensively used in the theory to combine and update belief func- tions. Let ml and m2 be probability assignments based on separate (“independent”) P Walley/Artijicial Intelligence 83 (1996) 1-58 29 bodies of evidence. A combined probability assignment m is defined by m(C) =p-’ C ml (A)m2( B) for all non-empty sets C, ArlB=C where P= c ml (A)MB) (9) is a normalizing constant, provided p is nonzero. It appears to be computationally feasible Ito use Dempster’s rule to combine belief functions that are defined on cer- tain kinds of tree structures [ 77,78,114]. Other computational results are discussed in [34,61,79,111,112]. Example 4 (Two unreliabEe witnesses). As a simple example of Dempster’s rule, sup- pose that there are two unreliable witnesses of the type considered earlier. The first witness, who has credibility cyt, reports that he observed event Cr. The second wit- ness, with credibility a~, reports C2. So ml and m2 are defined by ml (Cl) = al, ml(O) q : 1 -at, mz(C2) = ff2, m?(a) = 1 - (~2. Suppose that Ct and C2 are logi- cally indlependent events. Then p = 1 in Dempster’s rule and the combined probability assignment [71] is m(C1 n C2) = cqcq, m(CI) = cut(l - Ly2), m(C2) = (1 - CWI)CQ, m(Q)=(l-cut)(l-a2). When should beliefs be combined by Dempster’s rule? The rule appears to give reasonable answers in some problems, but it produces unsatisfactory and seemingly er- roneous tconclusions in other problems (an example is given below). Its applicability seems to rely, in general, on some rather delicate judgements of conditional indepen- dence. In. the example, the rule is applicable provided that beliefs, based on the report of one witness, about whether her report is correct would be unchanged by further information which specified both the report of the other witness and whether his report was correct. This is a type of conditional independence, conditional on the reports of both wit- nesses. It is quite different from the type of unconditional independence that is suggested by Shafer [ 72,75,76] as a justification for Dempster’s rule. (A similar distinction is made in [ 951 .:I In the example, each witness sometimes reports correctly and sometimes in- correctly, and the two witnesses may be unconditionally independent in the sense that learning whether one witness reported correctly (without knowing what the report was) would not change our belief that the other witness reported correctly. Unconditional independence is a simple judgement and often a reasonable one, for instance when the witnesses are known to have no interaction. But it is not sufficient to justify the use of Dempster’s rule, because the belief functions involved in Dempster’s rule are based on (or “conditional on”) the specific reports of the two witnesses. The credibility of witness i (q) must be a posterior probability, conditional on the report of witness i, in order to determine beliefs based on all the available evidence. (Shafer [ 72, p. 51, equates this posterior probability with the prior probability that witness i will report correctly, but Levi [55] has shown that this is unjustified.) 30 P: Walley/Art@cial Intelligence 83 (1996) 1-58 Dempster’s rule therefore relies on an assumption of conditional independence and this is harder to justify. If the two witnesses gave detailed reports which agreed in most details and were otherwise compatible, for example, then learning that one report is correct would certainly give us more confidence that the other report is correct and conditional independence would fail. It would fail also if the two reports were inconsistent, even though unconditional independence might be reasonable in these two cases. I am not suggesting that Dempster’s rule of combination is always inappropriate in such examples, merely that the implicit judgements of conditional independence should be made explicit and carefully considered. But conditional independence seems less reasonable in most other, more complex examples than in the case of the two witnesses, because learning one body of evidence and which mechanism or “code” produced it will typically provide information about the true state w and hence change beliefs about which mechanism produced the other body of evidence. That is especially clear in cases of conditioning, where one body of evidence actually restricts w to a subset of the initial possibility space LL Dempster’s rule is therefore unreliable unless careful attention is given to the implicit assumptions of conditional independence. For a general mathematical formulation of these assumptions and for further discussion, see [99, Section 5.131. The conditions given in [99] are similar to, but slightly different from, those given by Voorbraak [ 95, p. 1881. Voorbraak reaches a similar conclusion about the limited applicability of Dempster’s rule. Example 5 (Independent witnesses). Even in the simple example of the two witnesses, there are other ways of modelling independence. One way is to make the assessments p( Ct ) = at, f3 C2) = q, and judge that events Ct and C2 are independent (conditional on the testimony of both witnesses), meaning that upper and lower probabilities for one event would not change if we learned whether or not the other event occurred. Coherent lower previsions can then be computed by natural extension of the judgements P(CI) = P(C1 I c2> = P(Cl I cg = ~1 and p(C2) =p(C2 1 Cl) =J’(C2 1 Cl”> = LYE. Natural extension produces a lower probability model that is not a belief function (it is not even 2-monotone), and which is more precise than the belief function produced by Dempster’s rule. For example, let A denote the event that Ct and C:! have the same truth value (i.e. both occur or neither occurs). Then Dempster’s rule gives &,(A) = LYICX~, whereas natural extension gives F’,(A) = alcy2/( cq + (~2 - qq) which is strictly larger than &(A) provided 0 < (~1 < 1 and 0 < LUZ < 1. (Both rules give F(A) = 1.) This is one type of application in which natural extension produces inferences that are more precise than Dempster’s rule. (Compare with conditioning.) A third approach, suggested by Bayesian sensitivity analysis, is to look for all Bayesian probability measures P which satisfy P( Cl ) > CYI and P(Q) > a~, and under which Ct and C2 are independent [ 10.51. By forming the lower envelope of all such measures, we obtain a lower probability model that is more precise than the two previous models and again is not a belief function. For this model P,(A) =min{crt,a:!,Qta2 + (1 -nt)(l -a2)} P: Walley/Arti&ial Intelligence 83 (1996) 1-58 31 and &(.4) > &(A) > Z&(A), provided 0 < q < 1 and 0 < a2 < 1. If each witness has credibility $, for example, we obtain Pe (A) = 6, &(A) = 3 and Z’o (A) = $. The three models do agree on some probabilities, in particular p(Ct II C2) = qcq and p( Cl n C2) = 1 for each model. (A fourth model, based on possibility theory, will be given in Section 6. This gives &‘(CI II C2) = min{q, a~}, which seems less reasonable than the other solutions.) The differences between the four models may be important because judgements of independence are common in expert systems. It is not clear which model is most appropriate for practical applications, but see [99, Chapter 9;105] for some comparisons. Dempster’s rule of conditioning Suppose that the second probability assignment in Dempster’s rule is defined by m2( B) I= 1, representing knowledge that event B has occurred. Dempster’s rule of combination then reduces to Dempster’s rule of conditioning, which can be written most simply in terms of upper probabilities as PO (A 1 B) = P( A fl B) /F( B), defined whenever p(B) > 0, with &( A 1 B) = 1 --FD(A~ ) B). This rule is used in the theory of belief functions to update beliefs after receiving new information. Compare this rule with the rule of natural extension given in Section 4. Because every belief function E is 2-monotone, the conditional probabilities defined by natural extension are given by the simple formulas &(A I B) = P(A n B) P(AnB) +P(ACnB)’ PE(A / B) = _ P(AnB) P(AnB)+I’(ATB) (10) provided the denominators are nonzero. Several authors [23,41,93] have shown that if p is a bselief function then so is &(a 1 B). (This is true whenever F(B) > 0; see [ 4 I] .) Thus conditioning a belief function by natural extension produces another belief function. The conditional probabilities defined by Dempster’s rule are always at least as precise as those defined by natural extension [ 151, in the sense that &(A 1 B) < &,(A 1 B) < FD(A 1 B) < BdA I B). (11) Both rules yield the same conditional probabilities when the initial probability measure is precise (and then both agree with Bayes’ rule), but, in other cases, Dempster’s rule typically yields conditional probabilities that are more precise than natural extension. (See [ 1!>,23,48] for interesting comparisons of the two rules.) Natural extension pro- duces conditional upper and lower probabilities that are coherent with the unconditional upper and lower probabilities. In some cases the conditional probabilities defined by Dempster’s rule may also be coherent with the unconditional probabilities, even when Dempster’s rule and natural extension produce different answers. For instance, Demp- ster’s rule seems to produce reasonable inferences when it is used to update possibility 32 P: Walley/Art@cial intelligence 83 (1996) l-58 measures (see the examples in Section 6). But there are examples, such as the following, in which Dempster’s rule produces inferences that seem to be seriously wrong. Example 6 (e-contamination model). To compare the two rules of conditioning, con- sider the following statistical model. A random variable X takes values in the sample space X = {1,2,3,... , N}. To be specific we will take N = 104. (In fact Dempster’s rule will produce the same type of inconsistency for any value of N greater than 1, but the degree of inconsistency increases with N.) With probability 0.99, X is generated by the uniform probability distribution on X. With probability 0.01, X is generated by another, completely unknown, probability distribution on X. Thus almost all observa- tions follow a uniform distribution but one observation in a hundred is expected to be a “gross error”, in the sense that it is generated by a completely unknown (but possi- bly drastically different) mechanism. This sampling model is called an &-contamination neighbourhood of the uniform distribution (here E = 0.01) [ 401. Let B, denote the event that X = x, let A denote the event that X is generated by the uniform distribution, and define y = 1 if A occurs and y = 0 otherwise. There is uncertainty about both the value of X and whether A occurs, so we take the possibility spacetobefi={(x,y): x=1,2 ,..., N; y=O,l}. The uncertainty about D can be modelled by a belief function whose probability assignment m is defined by m(A fl B,) = 0.99/N for x = 1,2,. . . , N, m(AC) = 0.01, and m(C) = 0 for all other sets C. This model is imprecise because we do not know how to distribute the probability mass 0.01 amongst the alternatives to a uniform distribution. Before observing X, the probability of A is precisely 0.99. Now suppose we make a single observation X = x. How should we update our uncertainty about A? Dempster’s rule of conditioning produces Fo(A I B,) = F(An B,) m(A n B,) 0.99/N F(&) = m(A fl B,) + m(AC) = 0.99/N + 0.01 99 =- 99 + N < 0.01 when N = 104. Similarly PO ( AC 1 B,) = N/( 99 + N), hence &(A 1 B,) = 1 -Po(AC 1 B,) =PD(A 1 B,). Thus Po( A I B,) = &(A 1 B,) < 0.01 for every possible value of x. After observing x, the updated probability of A is precise and smaller than 0.01, whatever the value of x. Initially we are very confident that X will be generated by the uniform distribution. But if we use Dempster’s rule to update our beliefs then we will become very confident, whatever value of X we observe, that X was not generated by the uniform distribution! Intuitively there is a strong inconsistency between the initial and updated probabilities. Indeed an observer who knew that Dempster’s rule would be used to update probabilities could exploit the inconsistency to make a sure gain, by initially betting against event A and, after x is observed, betting on A at a more favourable rate. The initial and updated probabilities violate the coherence axioms (Cl) and (C2) of Section 4 and they “incur sure loss” in the mathematical sense of [ 991. P. Walley/Artifcial Intelligence 83 (1996) 1-58 33 In simple terms, Dempster’s rule produces the wrong answer because it treats the probability mass m(AC) = 0.01, which is spread over all possible values of x, as if it were entirely focused on the x that is actually observed. Indeed Dempster’s rule produces the inferences that a Bayesian would obtain from the precise probability assessments P( A n B, 11 = 0.99/N and P( AC II B,) = 0.01, which are incoherent when asserted for every possible x. I mentioned earlier that Dempster’s rule is applicable when a particular type of conditional independence holds. In this problem the required condition is that, for each x, learning that X = x would not change the relative likelihood of A n 8, and AC. Of course this condition is unreasonable. There appear to be many problems in which Dempster’s rule of conditioning or com- bination produces incoherent or unacceptable inferences; other examples can be found in [99, Section 5.131 and in [36,48,64,65,98]. Compare Dempster’s rule of conditioning with the method of natural extension, which always produces coherent inferences. The conditional upper and lower probabilities gen- erated by natural extension can be easily computed as follows: F&A 1 B,) =_ F(AnB,) 0.99/N P(AnB,) +P(ACnB,) = 0.99/N+O = ” &(A I B,) = P(A i-7 Bx) 0.99/N P(A n B,) + P(Ac n B,) = 0.99/N + 0.01 =Z<OOl 99+N * ’ In this problem, these are the unique updated upper and lower probabilities that are coherent with the initial belief function. The two rules of conditioning produce the same updated lower probability for A, but they produce very different upper probabilities. The updated probabilities given by Dempster’s rule are precise, whereas those given by natural extension are highly im- precise. The reason is that, for each possible i, the initial model is consistent with the hypothesis Hi that, whenever AC occurs, X takes the value i for certain. If we knew H* to be true we would obtain precise probability P( A 1 B,) = 99/(99 + N), but if we knew /4 to be true (where i $ x) we would obtain P( A 1 B,) = 1. The range of lower to upper updated probabilities must cover both values. The initial probability of A is precisely 0.99 under all the hypotheses Hi, but the updated probabilities are very different under different hypotheses and this produces imprecision in the updated upper and lower probabilities. Despite this explanation, some readers may feel that observation of x should have absolutely no effect on uncertainty about A and that the updated probabilities should be P( A 1 B,) = 0.99 (precisely) for every possible x to agree with the initial prob- ability P (,4) = 0.99. However these updated probabilities are not coherent with the initial belief function. If you take the updated probabilities to be precisely 0.99 then you must modify the initial probability model to make it precise. Indeed, we can use natural extension to compute the unconditional probabilities that are generated by the assessments P(A I B,) = 0.99 and P(A rl B,) = 0.99/N for all x, and we obtain P(AC fl Bx) = 0.01/N (precisely) for all x. This is tantamount to assuming that the probability mass m(AC) = 0.01 can be distributed uniformly over X, i.e. that X is generated by a uniform distribution when AC occurs, as well as when A occurs! If 34 I? Walley/Artificial Intelligence 83 (1996) I-58 there is really no information about how X is generated when AC occurs then the initial probability model should be imprecise, and this inevitably leads to imprecision in the updated probabilities. Expectations It is assumed in the Dempster-Shafer theory that uncertainties are modelled in terms of upper and lower probabilities. As discussed in Section 4, upper and lower probabilities may not be sufficient. In many problems upper and lower previsions or conditional prob- abilities are needed and information is lost if these are defined in terms of unconditional probabilities. Assuming that beliefs are specified in terms of a belief function c with probability assignment m, lower and upper previsions of a gamble X can be computed by natural extension, through the formulas cc P(X) = s xdFx(x) = cm(B) inf{X(w): w E B}, -cc BCfJ cc B(X) = s xdF,(x) = cm(B) sup{X(w): w E B}, (12) --oo EC&f where Fx and & are the upper and lower distribution functions of X under 41 (see Eq. (8) ). Preferences between actions can be constructed by computing upper and lower previsions of differences between utility functions, as outlined in Section 4. Other ways of using belief functions in decision making are described in [ 85,921. Conclusion Belief functions are a special type of coherent lower probability. They can model various types of partial ignorance and limited or conflicting evidence. Because they can be represented in terms of a probability mass function m, belief functions appear to be mathematically and computationally simpler, in some ways, than the general class of coherent lower previsions. Computationally efficient methods have been developed for combining and propagating belief functions. Belief functions can be assessed through multivalued mappings or in other ways, and the assessment strategies suggested by the theory are useful in many applications. (Shafer’s emphasis on constructing belief functions from simple evaluations of evidence is especially valuable.) However, many other assessment strategies produce coherent lower probabilities that are not belief functions. Some important types of uncertainty, e.g. judgements of probability in ordinary language, cannot be adequately modelled by belief functions. In fact belief functions are much less expressive than coherent lower previsions. The theory gives no justification for restricting attention to belief functions rather than coherent lower probabilities, or to lower probabilities rather than lower previsions. P Walley/ArtijEcial Intelligence 83 (1996) I-58 35 The calculus of belief functions relies heavily on Dempster’s rule of combination. Dempster’s rule may be useful in some problems where it is supported by explicit judgements of conditional independence but it is unreliable in general and it can produce inferences that are intuitively inconsistent and formally incoherent. More attention needs to be given to the exact conditions under which Dempster’s rule is applicable. Apart from the fundamental role it gives to Dempster’s rule, the Dempster-Shafer theory appears to be broadly compatible with the theory of coherent lower previsions. Belief functions that are constructed through a multivalued mapping have a behavioural interpretation and can be regarded as coherent lower probabilities, and they could be combined and updated through the rules of natural extension rather than by Dempster’s rule. 6. Possibility measures The the.ory of fuzzy sets was introduced by Zadeh [ 1161 for the purpose of mod- elling the imprecision and ambiguity of ordinary language. Uncertainty is often mea- sured in this theory by possibility measures, which are described in [ 16,17,46,115,118]. Concerning the use of possibility measures and fuzzy logic in expert systems, see [ 17,120,1:2 1 ] . Specific applications to expert systems include DIABETO [ 71, CADIAG- 2 [ I 1, SP.HINX [ 261, TAIGER [ 241, SPII-2 [ 5 11, PULCINELLA [ 53,691. Zadeh [ 1211 argues that “classical probability is insufficiently expressive to cope with the multiplicity of kinds of uncertainty which one encounters in AI and, more particularly, in expert systems”. In particular, Bayesian probabilities are inadequate for dealing wj th vague (inexactly defined) events or predicates, such as “young” or “tall”, or with natural-language expressions of probability, such as “probably”. These two types of vaguenfess occur together in the expression “it is likely that Mary is young”. Many of the production rules in expert systems such as MYCIN are fuzzy in these ways. (See [ 22,120] for many examples.) If expert systems are to be widely used, it does seem necessary to provide mathematical models for vague expressions in natural language. Dejnitions Consider the expression “Mary is young”. According to Zadeh, this provides some informatia’n about Mary’s precise age X, and the information can be modelled by a possibility distribution function TX, defined on the set D of possible ages. The number VQ (w) 1ie:s between zero and one and is read as “the degree to which it is possible that Mary has precise age o, given that she is young”. It is assumed that sup{~x( w) : w E 0) = 1, t:hat is, the function 7~ is normalised to have supremum value one. (Zadeh [ 1181 did not require this normalisation but it has been assumed by most later authors.) Zadeh identifies the function 7~ with the membership function for the fuzzy set “young” on 0. Thus a possibility distribution function is a type of membership function, and this enables th,e basic theory of fuzzy sets to be carried over to possibility distributions. A possibility measure T is defined for all subsets A of L? by v(A) = sup{rx (w) : w E A}, with V( 8) = 0. It follows that +7r has the properties (for all subsets A and B) : 0 < 36 P Walley/Art@cial Intelligence 83 (1996) l-58 n(A) < 1, ~(0) = 1, max{rr(A),7r(AC)} = 1, and ?T(A UB) = max{r(A),T(B)}. In general rr( A n B) < min{lr( A), T(B)} but there is equality in the important special case where A and B are “non-interactive” events [ 1171. Non-interaction appears to correspond to independence in probability theory. A possibility distribution function is analogous to a probability mass function or density function, and a possibility measure is analogous to a probability measure. Interpretation Thus Zadeh translates natural-language expressions into the mathematical formalism of possibility measures. But how should we interpret these measures? Zadeh clearly wishes to distinguish degrees of possibility from degrees of probability: “A basic as- sumption which underlies our approach to approximate reasoning is that the imprecision which is intrinsic in natural languages is, in the main, possibilistic rather than probabilis- tic in nature. . . . possibility relates to our perception of the degree of feasibility or ease of attainment, whereas probability is associated with the degree of belief, likelihood, frequency, or proportion.” [ 1191 It is widely recognised that possibility is distinct from probability and this distinction is central to all theories of probability, e.g. [ 281. Degrees of probability can be interpreted as betting rates or as relative frequencies, but it is less clear that it is meaningful to talk of “degrees of possibility”. Zadeh [ 1181 refers to expressions such as “slightly possible” and “quite possible” but admits that these usually indicate degrees of probability rather than degrees of possibility. Zadeh’s explication of degrees of possibility [ 118,119], and the example he gives to illustrate it, takes them to be measures of how easy or feasible it is to perform an action. This is somewhat different from the way that degrees of possibility are used in fuzzy logic, where they seem to measure how plausible it is that a proposition is true. Given the information “Mary is young”, for example, ~~(20) might measure how plausible it is that Mary is aged 20, but it is difficult to understand it as a measure of “how easy” or “how feasible” it is for Mary to be aged 20. The general usage is consistent with the following interpretation of degrees of possibility as upper probabilities, which are sometimes called “degrees of plausibility”. Various other interpretations of degrees of possibility have been suggested [ 16,171, but I do not know of any that justifies the usual min-max rules of combination. It is essential to have a clear interpretation of “degree of possibility” for the reasons mentioned earlier. If we do not understand the meaning of the numbers Q(W) then it will be difficult to assess them in practical problems, to understand conclusions that are expressed in terms of them, and to justify the calculus that is used to manipulate them. Interpretation as upper probability In fact, as suggested by Giles [ 32,331, possibility measures can be given a behavioural interpretation as upper probabilities. That is, we regard a possibility measure 7r as an upper probability measure, H(A) = T(A) = sup{7rx(w): w E A}, and interpret T(A) as a marginal acceptable betting rate for betting against A. The corresponding lower I? Walley/Artifcial Intelligence 83 (1996) l-58 37 probabilities are defined by P(A) = 1 - sup{~(w): w E AC} = 1 - r(AC). The lower probabilities are sometimes called necessity measures [ 17,461. They have the properties (for all subsets A, B of 0): F(A U B) = max{P(A) ,B( B)}, P(A II B) = min{P(A),P(B)}, P(A) 6 F(A), min{P(A),P(AC)} = 0, and either F’(A) = 0 or F(A) = 1. Suppose, for example, that A is the event that Mary’s age is at least 30 years. Given the information “Mary is young”, we might assess r(A) = 0.4 and V( AC) = 1, so that p(A) = 0.4 and P(A) = 0. The behavioural interpretation is that we would bet against A at odds of 3 to 2 against, but we would not bet on A at any finite odds. (As noted in Section 4, a more general behavioural interpretation could be given in terms of the implications of possibility measures for decision making. Decisions would be made, as in Section 4, by computing upper and lower previsions of differences between utility functions, using Eqs. ( 13) and ( 14) below.) All possibility measures are coherent upper probabilities. One way to see that is to construct their natural extensions to lower and upper previsions. If the possibility measure has possibility distribution IQ, its natural extensions are (for any gamble Z defined on a) 1 p(Z) = J inf{Z(w): TX(O) > u}du 0 sup Z =supz - J sup{~xW: Z(w) < y)dy infZ (13) 1 F(Z) = J sup{Z(o): Q(W) 3 u}du 0 SUD z =infZ + J SUP{~X(W): Z(w) > y)dy (14) infZ where inf Z and sup Z denote the infimum and supremum values of Z(o) over all w E 0. The second versions of each formula follow from Eqs. (8). See [ 102, Lemma I] for a derivation of these formulas. From the first expression for ZJ Z), it is easy to verify that the lower prevision p satisfies axioms (Pl)-(P3) of Section 4 and is therefore coherent. Then verify that its restriction:5 to upper and lower probabilities are defined by P(A) = sup{rx( w): w E A} and P(A) = 1 - sup{~(w): w E AC}, hence these upper and lower probabilities are coherent. (Alternatively, verify that p is 2-monotone and use the result that all 2-monotone lower probabilities are coherent.) There are great advantages in adopting this interpretation of possibility measures, especially as the theory of coherent upper and lower previsions could then be used to guide the assessment of possibility distributions and to derive rules for combining them. But note that possibility measures are a very special type of coherent upper 38 I? Walley/Artijicial Intelligence 83 (1996) I-58 probability. Their characteristic property P( AUB) = max{P( A), p( B)} is not necessary for coherence, e.g. if A and B are logically independent events and you assess F(A) = p(B) = i, the natural extension is P(A U B) = 1, and any assessment of P(A U B) between $ and 1 would be coherent with the initial assessments, whereas p( A U B) = i is necessary for P to be a possibility measure. For finite 0, a possibility measure r is just a consonant plausibility function in the sense of Shafer [ 711, and the corresponding lower probability p(A) = 1 - 7r( AC) is a special type of belief function, characterised by the property that the sets B for which m(B) > 0 form a nested sequence [46]. Indeed it seems more appropriate to call z-(A) a degree of plausibility rather than a degree of possibility. Event A is more or less plausible according to the amount of evidence pointing against A, e.g. if there is no evidence against A then A is fully plausible, T(A) = 1, and there is no reason to bet against A at any odds. We could also interpret F’(A) = 1 - r( AC) as a “degree of potential surprise” that we would experience if A failed to occur [54,70], or as a “degree of provability” of A [ 121. The uncertainty measures proposed by Shackle [ 701 and Cohen [ 121 appear to be mathematically equivalent to possibility measures and thus they are special types of coherent lower probabilities. Imprecision Possibility measures can be used to model some types of imprecise or partial infor- mation. For example, complete ignorance about a variable w can be modelled by the possibility distribution Q(W) = 1 for all w in 0, which corresponds to the vucuous upper and lower probabilities. The degree of imprecision concerning an event A can be measured in general by p(A) - I’( A) = T(A) + T( AC) - 1 = min{rr(A), r( AC)}. Some natural-language judgements, such as “Mary is young” and the variants discussed in the following subsections, can be modelled by possibility measures. However, first-order possibility measures do not seem to be sufficiently flexible to model many common types of uncertainty. The football example and other examples involving natural-language judgements of uncertainty cannot be adequately modelled by belief functions, and certainly not by first-order possibility measures which correspond to a special type of belief function. (Second-order possibility measures, considered later, can be used to model natural-language judgements of uncertainty, but they are considerably more complicated than first-order measures.) Most examples of multival- ued mappings and belief functions, such as the e-contamination model in Section 5, involve coherent upper probabilities that are not possibility measures. Nor can (first- order) possibility measures model precise probability assessments. Bayesian probability measures are a special type of belief function or coherent lower probability, but not a special type of possibility measure. The upper and lower probabilities defined by a non-degenerate possibility distribution are always imprecise and usually very imprecise, i.e. P(A) - l’(A) is large for many events A. Example 7 (ModeDing vague predicates). Zadeh [ 1211 argues that Bayesian probabil- ities are unable to model judgements like “Mary is young” or “it is likely that Mary is young”. He shows how these judgements can be modelled by possibility distributions, t? Walley/Art@cial Intelligence 83 (1996) l-58 39 although he does not give numerical assessments of the required distributions. I want to indicate how these judgements can be modelled using the theory of coherent lower previsions and the extent to which this is consistent with fuzzy reasoning. Other models for the same judgements are proposed in [ 11,181. Consid’er the proposition “Mary is young”. This provides some (but not much) infor- mation about Mary’s age. We could model our uncertainty about Mary’s age, given the information that she is young, by a coherent lower prevision. There are many possible ways of constructing a lower prevision. One way, which seems suitable for relating terms such as “young”, “old”, “tall” and “rich” to an appropriate numerical scale, is through assessments of upper and lower distribution functions [ 991. Let X denote Mary’s precise age. The upper and lower distribution functions of X are defined by F(w) = P(X < w) and F(w) = p(X 6 o) for all w > 0. Given only that Mary is young, it is entirely plausible that X < w for any specific w, so it is natural to take F(w) = 1 for all w. (This is a vacuous assessment, so F could be ignored altogether.) But “Mary is young” does provide some evidence that she is under 30, and strong evidence that she is under 40. On this basis, one might assess F(w) = 0 for 0 < w < 15, F(w) = (w - 15)/25 for 15 < o < 40, F(w) = 1 for w 3 40. (Alternatively, one might make a few qualitative judgements such as “probably X < 25” and “very probably X < 30” which can be translated into constraints on E. Note that the information provided by the assertion “Mary is young” is highly sensitive to the context in which it is made-is she “young to be walking” or “young to be retiring”?-- and it is debatable whether a useful model can be given that is independent of context:. We must assume, at least, that Mary is human!) A coherent lower prevision is then constructed by natural extension of these judge- ments. For any set A of possible ages (positive real numbers), we obtain upper and lower probabilities p(A) = sup{ 1 - E(w) : w E A}, p(A) = inf{& w) : w E AC}. This probability model is highly imprecise, as one would expect. Here the upper probability P is actually a possibility measure, with possibility dis- tribution Function 7~ (o) = 1 -F(w) = P(X > w), so 7rx(w> = 1 if 0 < w 6 15, Q(W) = (40-w)/25 if 15 < w < 40, and TX(W) = 0 if w > 40. In this case rrx(w), the degree to which it is possible that Mary has precise age w, can be identified with the upper probability that Mary’s age exceeds w, and the analysis based on natural extension of the lower distribution function F agrees with the analysis based on the possibility distribution rx; both generate the same possibility measure i? In general, when both the upper and lower distribution functions are non-vacuous, the upper probability produced by natural extension will not be a possibility measure. Conditional possibilities Suppose that unconditional upper probabilities are defined through a possibility dis- tribution :r by P(A) = sup{7r(o): w E A} and we wish to construct conditional upper probabilities P(. 1 B). Assume that r(w) > 0 for some w in B, so that F(‘(B) > 0. (Otherwise there is no useful information in 7r from which to construct P(. 1 B) .) The conditional probabilities can be constructed by natural extension, using Eq. (5). 40 l? Walley/Art$cial Intelligence 83 (1996) 1-58 Because the corresponding lower probabilities are 2-monotone, F(A 1 B)=- F(A n B) P(AnB) +p(ACnB) sup{~( w) : w E A r-7 B} =sup{a(w): w cAnB}+l -sup{~(w): we AUP} = sup{~( w 1 B) : o E A} where the conditional possibility distribution IT(. 1 B) is defined by (15) dw) T(W ( B) = T(W) + 1 - max{7r(w),/3}’ if w E B and n-(w) > 0, (16) 0, if w E BC or T(W) = 0, and p =P(BC). It can be verified that sup{~( w 1 B): w E L?} = 1 so that T(. 1 B) is a possibility distribution. This shows that if the unconditional upper probability P is a possibility measure then so is the conditional upper probability p(. 1 B) defined by natural exten- sion, provided p(B) > 0. Thus conditioning a possibility measure by natural extension produces another possibility measure. Example 8 (Examples of conditioning). Consider the model for the judgement “Mary is young”, characterised by the possibility distribution T(W) = 1, if 0 < o < 15, z-(w) = (40 - w) /25 if 15 < w < 40, T(W) = 0 if w 2 40. Suppose we learn the additional information that Mary’s age belongs to a specified set of real numbers, B. We can update upper and lower probabilities concerning Mary’s age simply by updating the possibility distribution T to T(. 1 B). First let B be the event that Mary is no older than 30 years. Using Eq. (16) we find that T(W / B) = r(w) if 0 < w < 30, T(W 1 B) = 0 if w > 30. Here T is updated simply by truncating it at age 30; the plausibility of ages below 30 does not change. As a second example, if B is the event that Mary’s age is not between 20 and 30 years, we find that TT( w 1 B) = n-(w) if 0 < w < 20, n-( w I B) = 0 if 20 < w < 30 or w 3 40, and T( w 1 B) = (40 - w) /(45 - w) if 30 < o < 40. Here the plausibility of ages below 20 does not change, but the new information makes ages between 30 and 40 more plausible than before. Finally, notice that if B does not contain the entire interval (0,151 then the updated probabilities will be vacuous, in the sense that a(o I B) = 1 if w E B n (0.40) and T(W I B) = 0 otherwise. If we learn, for example, that Mary is at least 5 years old then T( w I B) = 1 if 5 < w < 40 and T(W 1 B) = 0 otherwise; all ages between 5 and 40 become fully plausible. More generally, if BC contains a state w that is fully plausible (i.e. T(W) = 1) then F(BC) = 1 and hence the probabilities conditional on B are vacuous, in that rr(w 1 B) = 1 if w E B and T(W) > 0, while T(W I B) = 0 otherwise. In this last case, where P(B) = 0, the conditional probabilities defined by natural extension are essentially vacuous. This is generally the case when natural extension P. Walley/Arti@ial Intelligence 83 (1996) l-58 41 is used to condition on an event with lower probability zero. It can be understood through the consistency principle stated near the end of Section 4. In the example let A denote “Mary’s age is less than 30” and let B denote “Mary’s age is at least 5”. Then P(B) = 0, and it would be consistent with the initial possibility measure to make a further judgement that P(A rl B) = 0 precisely, which would produce P(A 1 B) = 0. To be consistent with this we need E(A 1 B) = 0. The same argument applies when A is any proper subset of the interval [ 5,40). So natural extension cannot produce any nontrivial inferences when the conditioning event B has P(B) = 0. (Unfortunately, for possibility measures there are many such events B. In fact, for every event B, either E(B) = 0 or E( BC) = 0.) The problem is that the initial model is too imprecise to determine conditional probabilities. One way to avoid this is to specify a more precise model, in particular one in which P(B) > 0. Of course the corresponding upper probability may no longer be a possibility measure. Dempster’s rule of conditioning Another option is to use a different rule of conditioning when p(B) = 0. When applied to a possibility measure ?r for which F(B) > 0, Dempster’s rule of conditioning gives the conditional upper probabilities FD(A 1 B) = sup{?s)(w 1 B): w E A}, where the conditional possibility distribution fl(. I B) is defined by ,&u 1 B) = if o E B, if o E BC, (17) and k = sup{~-( w): w E B} = p(B). The conditional upper probability FD(. I B) is a possibility measure. Writing rrE(. 1 B) and FE(. I B) for the conditional possibility distribution and conditional upper probability defined by natural extension, we have #(w I 13) < T~(OJ I B) for all w, and PD(A I B) < F&A I B) for all sets A. Thus the conditional probabilities determined by Dempster’s rule are always at least as precise as those defined by natural extension. Consider the three examples of conditioning the information about Mary’s age. In the first example, where B is the interval (0,301, rD (. I B) = c#(. I B) and the two rules produce the same solution. In the second example, where B = [ 20,30]‘, T’ (w 1 B) agrees with ,rrE( w 1 B) except when 30 < o < 40, in which case &‘( w I B) = T(W) is smaller than 7rE ( w I B) . The biggest difference between the two rules occurs in the third example, where B = [ 5, cm): wD (w 1 B) agrees with the initial degree of possibility n-(w) if 5 < w < 40, whereas rE (w I B) = 1. In this case conditioning by Dempster’s rule preserves the information in the initial possibility distribution but natural extension does not, and Dempster’s rule is arguably more reasonable. (In fact &‘(w I B) = T(W) for all w in B in all three examples, because k = p(B) = 1 in each case.) Dempster’s rule for conditioning possibility distributions (17) can also be derived within possibility theory. The information that event B has occurred is represented by a possibility distribution rf which is simply the indicator function of B. This is combined with the initial possibility distribution r using the minimum rule, producing the conditional possibility distribution ~(w I B) 0: min{r(w), n’(w)}, so ~(0 I B) 42 P. Walley/Arti$cial Intelligence 83 (1996) I-58 c( z-(w) if w E B and V(W 1 B) = 0 if w E BC. The proportionality constant is determined by the fact that r(. 1 B) is a possibility distribution, and the solution agrees with Dempster’s rule ( 17). A different rule for conditioning possibility distributions is proposed in [ 171. Example 9 (Probabilistic quali$cation). Now consider the qualified judgement “it is likely that Mary is young”. Zadeh [ 1211 treats this as information about an underlying probability density function for Mary’s age, whereas I regard it simply as information about Mary’s age. Let B denote the event that Mary is young, according to the criteria used by the speaker. The conditional probabilities P(A 1 B) and Z’( A 1 B) can be identified with the (unconditional) probabilities based on the information “Mary is young” which were defined previously. The judgement that B is likely is translated as p(B) 2 i. The natural extensions of these assessments are P(A) =iP(A 1 B) = iinf{F(w): w E AC} = i - $sup{r(w): OJ E AC}, P(A) = i + iP(A 1 B) = 1 - i inf{F(w): w E A} = i + $ sup{rr(w): w E A}, for any set A of possible ages. Again the upper probability is a possibility measure, with possibility distribution function 7~( w) = 1 - iF( w). The effect of the qualification “it is likely that” is simply to reduce the lower distribution function I; by a factor of $. (Equivalently, replace r by i + ir.) Although it produces a possibility distribution for Mary’s age X, this analysis is much simpler than that of Zadeh [ 1211, who requires a possibility distribution to be defined on the space of all probability density functions for X. More generally, suppose that B(. I B) and p(. I B) are upper and lower probabilities based on information B, and this information is qualified in some way that can be translated as J’(B) 2 /?, e.g. by asserting that B is “very probable” or by using one of the other natural-language expressions listed later in this section. Then upper and lower probabilities based on the qualified information can be computed by e(A) = pl’( A I B) and p(A) = pP( A 1 B) + 1 - p. This is a simple method of discounting information. To see that such an analysis will not always produce a possibility measure, consider the upper probabilities generated by natural extension from the two judgements “it is likely that Mary is young” and “it is likely that Mary is older than 10 years”. Let A and B denote the events “Mary is younger than 10 years” and “Mary is older than 40 years”. Then ‘ir( A) = P(B) = i but P( A U B) = 1 > max{P( A), p( B)}, so H is not a possibility measure. Cheeseman [ 1 l] suggests several Bayesian models for the uncertainty about Mary’s age, given the information “Mary is probably young”. He would first model the uncer- tainty given “Mary is young” by specifying a probability density function for Mary’s age, and similarly for the uncertainty given “Mary is not young”. To model “proba- bly” he would construct a second-order probability density on the unit interval. This is analogous to Zadeh’s second-order possibility distribution for “likely”. Thus Cheeseman’s analysis requires second-order assessments of similar complexity to Zadeh’s. However Cheeseman claims that “For the accuracy appropriate to this type P Walley/Artifcial Intelligence 83 (1996) l-58 43 of linguistic information, it is sufficient to extract a single estimate (probability)” as an approximation to the second-order density, and he chooses the precise probability 0.9 (the mean of his second-order density) to represent “probably”. So Cheeseman, given only the information “probably A”, would be willing to bet on A at odds of 9 to 1 on! Both the second-order density and the precise value 0.9 have strong implications concerning betting rates and other decisions, and they are therefore inadequate models for an imprecise term like “probably”. Cheeseman’s overall model for the uncertainty about Mary’s age is a precise probability density function, which is similarly inadequate to model the vagueness of the information on which it is based. Vague probability judgements (second-order possibility measures) Now consider the translation of vague probability judgements such as “event A is likely” or “A is very likely” into possibility distributions. Zadeh [ 117,121] treats such a judgement as information about the precise probability (p) of event A, and models it through a possibility distribution function 7rt, defined on the probability interval [0, 11. For instance, he models “it is likely that A” by a possibility distribution function rP(p) that is zero when 0 < p < i and increases nonlinearly on i < p 6 1 to a maximum 7rP( I) = 1. He models “it is very likely that A” by squaring the function rt,. Watson, Weiss and Donnell [ 1091 model “pretty likely” by a possibility distribution function that is zero on the interval 0 < p 6 0.55 and roughly constant on 0.65 < p < 0.9. More generally, a judgement such as “it is likely that Mary is young” would be translated by assigning a degree of possibility m(f) to every conceivable probability density function f for Mary’s age. It is strange that the theory of possibility measures, apparently designed to model the imprecisiamn in ordinary-language judgements of uncertainty, needs to refer to underlying probabilities that are precise. The number r,.,(p) measures the degree of possibility that the true probability of A is precisely p (given the judgement “it is likely that A”), i.e. it measures second-order uncertainty about a first-order probability. I have already remarked that it is unclear what is meant by “degree of possibility”, but there is a more basic difficulty here: it is not even clear what is meant by “the true probability of A”, and why this should be assumed to have the properties of a Bayesian probability measure. The most natural interpretation is that the “true probability of A” is the subjective probability assigned to A by the speaker, who provides only partial information about it when he asserts “it is likely that A”. But there is no reason to suppose that the speaker has made (or could make) a precise assessment of this probability-the vagueness of his assertion suggests just the opposite. (This issue is discussed at length in [99] .) Similarly, there will rarely be a “true probability” in any objective sense. If we do not understand what is meant by “the true probability of A”, how can we hope to assess the degree of possibility that p is the true probability? Compare Zadeh’s translation of “it is likely that A” with the behavioural translation of “probably A” that was suggested in Section 4. (Zadeh [ 1171 regards “likely” and “probable” as “more or less synonymous”.) The behavioural translation is that the speaker is willing to accept an even-money bet on A, which is modelled by P(A) 2 i. I! Walley/Art@cial Intelligence 83 (1996) I-58 45 regarded as “non-interactive” and a joint distribution formed by taking minima. (This is questionable because the first judgement effectively restricts the range of possible values for w - d. But this kind of “interaction” between judgements is very common, and it is not clear how possibility theorists would combine the possibility distributions if not by the minimum rule.) Thus, assuming w + d + 1 = 1 so that (w, d, Z) is a probability distribution, ~(w,d,Z)~min{~l(w),~~(W-d),~~(d-Z)} cx max{O, min{ 1 - 2w, w - d, d - I}}. It is necessary to renormalise the right-hand expression to have maximum value one, and this yields the joint possibility distribution 7r( w, d, Z) = 9 max{O, min{ 1 - 2w, w - d, d - I}}. The maximum possibility is achieved by w = 8, d = f, I= $. Compare this with the lower prevision constructed in Section 4, which is the lower envelope of the set M of all probability distributions (w, d, 2) that are consistent with the three _iudgements, i.e. satisfy w < 1, w 2 d, d b 1. The joint possibility rTT( w, d, 1) is positive when (w, d, 1) lies in the interior of M, it increases with distance from the boundary of M, and it is zero otherwise. Marginal possibility distributions can be computed from r by maximisation. For example, after some calculations the marginal possibility distribution for 2 is found to be rrL(I) =max{7r(w,d,Z): 0 < w < 1, 0 < d < 1, w+d = 1-Z) =9max{O,min{$, i- 1}}, which is unimodal with mode m = f . Hence, using Eqs. ( 19), this model generates upper and lower probabilities 1 P(L) =m+ s rL(y)dy= $ =0.2 ,?1 78, p(L) =m - I rTTL(y) dy = ; =O.ll 0 1. These probabilities are more precise than the values P(L) = 5 and p(L) = 0 generated by the model in Section 4; here P(L) -c(L) = & r~(y) dy = i, compared with 3 for the earlier model. Calculus As seen in the preceding example, the rules for combining and manipulating possi- bility distributions involve operations of forming maxima and minima. These rules are derived from the basic rules of fuzzy sets proposed by Zadeh [ 1161. Without a definite interpretation of possibility measures the rules appear quite arbitrary. But if possibility measures are interpreted as coherent upper probabilities, as suggested earlier, the rules 46 P: Walley/Artificial Intelligence 83 (1996) 1-58 need to be evaluated according to whether they preserve coherence, in the sense that, after applying the rules, the overall upper probability model is coherent. In some spe- cial cases, such as (b) below, the rules agree with natural extension and then they do preserve coherence. This gives some justification for these rules. But other rules, such as (c), disagreee with natural extension in general. It needs to be investigated whether such rules can produce upper probabilities that are incoherent or incur sure loss. I am currently studying these questions and the results will be reported elsewhere. The most important rules are as follows. (a) Given a joint possibility distribution 7rx,r for two variables X and Y, a marginal possibility distribution for X is defined by Q(X) = sup{7~x,r(x, y): y E L$} where 0, = {y: (x, y) E 0) and D is the joint possibility space. This rule is essentially built into the definition of a possibility measure. It does preserve coherence of the corresponding upper probabilities. (b) Given two marginal possibility distributions TX and rr~~y, where X and Y are logi- cally independent variables, a joint possibility distribution for X and Y is defined by ?TX,J( x, y) = min{rx(x), 7ry( y)}. This implies, in terms of the correspond- ing upper probabilities, that Px,r( A x B) = min{Fx( A), &( B)} for product sets A x B, and this rule agrees with natural extension of the marginal upper probabilities to a joint upper probability. However the two rules can disagree for sets that are not products. (c) Given two possibility distributions ~1 and ~2, concerning the same unknown w but based on “non-interactive” bodies of evidence, a combined possibility distribution is defined by 7~( w) c( min{rt (w), 772( to)}, where the normalising factor is chosen so that n has supremum value 1. This rule was used to combine judgements in the football example. Rule (b) can be regarded as a special case of (c) with w = (x, y). Another special case is the rule of conditioning discussed earlier, where ~2 is the indicator function of the conditioning event, which agrees with Dempster’s rule of conditioning ( 17). The more general rule (c) appears to have a similar role in combining possibility distributions to Dempster’s rule for combining belief functions. It is used in expert systems to combine information from different sources. However the rule disagrees with natural extension in general (although one special case of agreement was noted in (b)), and it need not produce coherent inferences. The next example shows that rule (c) and Dempster’s rule of combination can produce quite different answers. Example 11 (Two unreliable witnesses). Consider the example in Section 5 of two unreliable witnesses. The first witness has credibility at and reports that event Ct occurred. This can be modelled by a marginal possibility distribution 7~1 that assigns ~1 (Cl) = 1 and ~1 (C-f) = 1 - “1. This generates the upper and lower probabilities assumed in Section 5, PI (Cl ) = 1 and et (Ct ) = cyt . Similarly the information provided by the second witness is modelled by a possibility distribution ~2 (C2) = 1 and 7r2( Cg) = 1 -cQ. If we combine the two marginal possibility distributions by the minimum rule, we obtain degrees of possibility 1, 1 - LYZ, 1 - LYI and 1 - max{Lyt , a~} for the elementary P. Walley/Artijicial Inrelligence 83 (1996) l-58 47 events CI flC2, CI nC;, CfflC2 and CynCi. This generates upper and lower probabilities p( Cl n C;!) = 1 and ~(CI fl C2) = min{cul, CQ}. The lower probability is quite different from the value alay2 which is given by all three models in Section 5, based on different ways of formalising the judgement that the two reports are independent. There is also a discrepancy between the values of F(C,” fl Ci), which are 1 - max{crl , a~} for this model and ( 1 - CYI ) ( 1 - a~> for the three earlier models. In this example the combined possibility distribution seems less satisfactory than the three earlier models. When crl = cy:! = $, for example, the earlier models give p( Cl n Cz) = i whereas the rule for combining possibility measures gives p( Cl n CT) = 4. The latter value seems unreasonable; one would expect p( Cl n C2) to be somewhat less than the marginal lower probabilities a and E( CZ), which are each i, unless there is evidence that the two reports are correlated in a particular way. In fact, for any joint possibility measure, the corresponding lower probability must satisfy p(Cl n C;?) = min{P(CI),P(C2)} since this is a general property of necessity measures. But this property seems inappropriate when the two sources of information are judged to be independent. So it does not seem possible to adequately model independence judgements using possibility measures. Computation The computation of inferences and decisions from possibility distributions requires, in general, the solution of a nonlinear programming problem. (Compare with the com- putation of natural extension suggested in Section 4, which involves only linear pro- gramming.) For example, decisions are made from second-order possibility distributions by computing “fuzzy expectations”, i.e. degrees of possibility for all possible expected values x of a random variable X, by maximising n-(P) over all probability distributions P which satisfy P(X) = x. Because rr( P) will generally be a nonlinear function of the probability masses, this entails a nonlinear programming problem-in fact, a separate problem for each possible value of x! (See [ 109,117,121] for more details.) Similarly the computation of a marginal possibility distribution involves, in general, many non- linear programs. So computations will often be difficult, despite the apparent simplicity of the calculus. The computations are generally easier for first-order possibility distri- butions than for second-order distributions because the optimisations are generally over lower-diml:nsional spaces. Assessment How di:fficult is it to make the assessments required for a fuzzy analysis? The aim is to allow users of the system (or domain experts) to express their uncertainty in natural language, in whatever forms are most convenient. This is valuable, as it makes the task OF assessing uncertainty relatively simple for the system user. The difficulties are passed on to the experts on fuzzy logic, who must translate the natural-language judgements into possibility distributions. For example, they must translate “it is likely that A” into a function rr defined on the probability interval [0, l] . More complex 48 I? Walley/Art@cial Intelligence 83 (1996) 1-58 judgements of uncertainty concerning 0 must be translated into a function 71 defined on the space of all probability distributions on 0. A great deal of input is needed to specify these functions, and there seems to be a great deal of arbitrariness in selecting them. Again it is generally easier to assess a first-order possibility distribution than a second-order distribution because the former is usually defined on a lower-dimensional space. There seems to be little guidance in the literature on possibility theory about how to select the required functions. (But see [ 17, p. 191, for some suggestions.) Specific assessments of possibility distributions are sometimes given to illustrate the methodology, but these appear quite arbitrary and no justification is offered. For example, Zadeh [ 1171 specifies two quite different possibility distributions to model the judgement “likely” in successive examples: his Example 1.1 has 40.8) = 0.9 while Example 1.2 has 7r(O.8) = 0.5. Another translation, shown graphically in his Fig. 1, seems to be different again. No reasons are given to support any of these assessments, and it is hard to see how they could be supported until the meaning of the numbers m(p) is clarified. In practice one might use standard translations for common terms such as “likely”, “very likely” or “more likely than”. That is, all instances of the term “very likely” would be translated into the same possibility distribution. The choice of this standard distribution may still be somewhat arbitrary but no further input would be needed. The danger of using standard translations is that different people seem to use expres- sions like “very likely” in different ways and usage varies with context [ 3,106]. Ideally, one would like to use a different translation for each person and context. That would be impracticable in most cases, as it would require too much input from the person. However one could require the standard translation of “very likely” to encompass the meanings intended by most speakers in most contexts. We can illustrate the idea as follows, using the behavioural interpretation of upper and lower probabilities. Standard translations into lower probabilities Consider the expression “probably A”. Most people who use this expression would be willing to bet on A at even stakes. Hence they would accept the behavioural translation p(A) > 0.5. We could, in principle, determine a lower probability E(A) for person i by finding the least favourable odds at which he is willing to bet on A. Provided &(A) 2 0.5 for all persons i, the behavioural translation l’(A) 2 0.5 is acceptable to all, and we may say that it encompasses the meanings of “probably A” for all the persons. It could then be used as a standard (“cautious” or “default”) translation of “probably A” that is acceptable in the great majority of problems. In any particular problem we would try to (a) check with the person that he accepts the behavioural translation of his judgement, and (b) encourage him to make this more precise, e.g. if he is willing to bet on A at odds of 3 to 2 on then he will accept the more precise judgement P(A) > 0.6. If he was unwilling to do this, or if the checks could not be carried out, then the standard translation would be used. The standard translations should be based on empirical studies of the meanings of terms such as “probably”. It seems, for example, that most people do not apply the term P. Walley/Art@ial Intelligence 83 (1996) 1-58 49 “probable” to events with very high probability, so that one might include the constraint P(A) 6 0.9, as well as l’(A) 2 0.5, in the standard translation of “A is probable”. In the light of empirical studies such as [ 3,45,106], I suggest the following translations of common judgements in natural language. In all these expressions I take “probable” to be synonymous with “likely”. I identify negative expressions such as “A is improbable” with positive expressions such as “AC is probable”. The latter is translated into &( AC) > 0.5, which is equivalent to P(A) < 0.5. This translation is somewhat cautious, as there is evidence that most people would accept a stronger translation P(A) < 0.4. Many of the other translations could be strengthened considerably in suitable contexts. l A is extremely probable + l’(A) 2 0.98. l A has very high probability --) l’(A) 2 0.9. l A is highly probable -+ I’(A) > 0.85. l A is very probable + P(A) 2 0.75. l A has a very good chance + P(A) 2 0.65. l A is quite probable + P(A) 2 0.6. l A is probable [likely] + l’(A) > 0.5. l A is improbable [unlikely] --+ F(A) < 0.5. l A is somewhat unlikely [quite improbable] + P(A) < 0.4. l A is very unlikely --f P(A) < 0.25. l A has little chance --+ P(A) < 0.2. l A is highly improbable + P(A) < 0.15. l A has very low probability + p(A) < 0.1. l A is extremely unlikely -+ P(A) < 0.02. l A has a good chance -+ Z’(A) 2 0.4, P(A) < 0.85. l The probability of A is about (Y --+ Z’(A) 2 a - 0.1, P(A) 6 a + 0.1. l A is more probable than B + F’(A - B) 2 0. There is some degree of arbitrariness in choosing a single number (0.85) to translate an expression such as “highly probable”, but this is much less than the arbitrariness in choosjng a possibility distribution function r, i.e. a degree of possibility r(p) for every value of p between 0 and 1. Similar translations, from imprecise probabilities to lingui:stic expressions, could be used to make a system’s conclusions and reasoning more comprehensible to a user. Possibility theory and fuzzy logic have made a substantial contribution to our un- derstanding of uncertainty by drawing attention to the important problem of modelling natural-language judgements and suggesting possibility measures as suitable models. Provided second-order measures are allowed, possibility measures can mode1 a wide variety of uncertainty judgements, including imprecise judgements in natural language. Bayesian probabilities are inadequate to mode1 vague predicates and vague probability judgements, as in “Mary is probably young”, but it appears that upper and lower previ- sions may be adequate, especially as possibility measures can be regarded as a special type of coherent upper probability. SO f? Walley/Art$cial Intelligence 83 (1996) 1-58 It is important to distinguish between first- and second-order possibility measures. First-order possibility measures, which are used to model the uncertainty generated by vague predicates like “young” and “tall”, can be interpreted as coherent upper probabili- ties. They are mathematically simpler than the other types of coherent upper probabilities as they can be defined in terms of possibility distributions, functions whose domain is LJ rather than the power set of a. This may simplify computations and assessment, e.g. possibility measures can sometimes be assessed through upper or lower distribution functions. First-order possibility measures are likely to be useful models in many expert systems where information is elicited in terms of vague predicates. The main defect of first-order possibility measures is that they cannot model many of the common types of uncertainty. In particular they cannot model natural-language judgements of uncertainty or precise probability assessments. Possibility measures are a very special type of upper probability (in fact they correspond to a special type of belief function), and upper probability is itself inadequate in many problems; upper and lower previsions are needed in general. Second-order possibility measures, defined on the set of all probability distributions, are much more expressive than first-order measures. They can model both precise and imprecise judgements of uncertainty. But they are much more complicated than first- order measures and they have some serious defects. Second-order models are difficult to interpret and assess and they seem overly complicated to model qualitative judgements. Theoretical papers on fuzzy logic tend to ignore the practical problem of assessing second-order functions, each defined on a space of probability distributions. Compu- tations, e.g. of marginal possibility distributions or fuzzy expected values, generally require nonlinear programming. The rules for combining possibility measures are simple (they involve operations of maximising and minimising), but they do not appear to have any compelling jus- tification. If possibility measures are interpreted as coherent upper probabilities then the rules can be compared with the rules of natural extension and we can investi- gate whether they preserve coherence. They appear to do so in some cases but not in general. 7. Comparison and evaluation Most practical reasoning involves uncertainty. In expert systems, as in many other fields, it is frequently necessary to measure the uncertainty. What measure should we use? Four measures have been considered in this paper. Here is a summary of the extent to which they satisfy the criteria proposed in Section 2. Interpretation, calculus and consistency These criteria are satisfied only by the theories of Bayesian probability and coherent lower previsions, which start with a simple behavioural interpretation and use this to jus- tify principles of coherence and to derive rules for combining and updating probabilities. There is a general method, called natural extension, for computing new previsions from R Walky/Art~cial Intelligence 83 (1996) 1-58 51 an arbitrary set of judgements. Natural extension can be used to make inferences and decisions. There are general methods for checking consistency of the initial judgements, and the rules of natural extension ensure that conclusions will be consistent with the judgements. Possibility theory and the Dempster-Shafer theory do not do so well on these cri- teria. These theories propose mathematical properties to characterise their uncertainty measures and simple rules for combining measures, but they fail to give any compelling justification for their properties and rules. Belief functions can be interpreted in terms of multivalued mappings but this supports Dempster’s rule of combination only when some strong assumptions of conditional independence are added. Both theories lack methods for checking the consistency of models and conclusions, and their rules can produce conclusions that are intuitively inconsistent with the initial model. Imprecision Bayesian probabilities cannot adequately model ignorance, imprecise or qualitative judgements of uncertainty, or vague predicates in natural language. The other measures can do so to some extent, but belief functions and first-order possibility measures are not sufficiently general to model common types of imprecise judgements. Assessment Insufficient guidance is given in these theories (especially possibility theory) about how to make assessments of uncertainty, although the theories of belief functions and lower previsions do take this problem seriously. All the theories seem to need judgements of independence or non-interaction, in multivariate problems, to reduce the number of assessments. Lower previsions and second-order possibility measures can model a wide variety of uncertainty judgements, including qualitative judgements in ordinary language, although assessments of second-order possibility distributions seem complicated and arbitrary. Assessment is onerous for Bayesians because they require precise assessments and a complete probability model; the other theories are more flexible. Computation For all the measures, computational feasibility will depend on the type and complexity of the model and the number of assessments. For lower previsions, the computation of inferences and decisions by natural extension involves linear programming. More work is needed to develop computationally efficient methods and tractable models, particularly based on conditional independence. Bayesian probabilities, belief functions and first-order possibility measures, as special types of lower or upper previsions, may be computationally simpler in some cases. (They are tractable in singly-connected belief networks, for example.) Computational methods are most highly developed for Bayesian models. !$econd-order possibility measures are less tractable than the other measures- computations involve nonlinear programming. 52 P. Walley/Artifcial Intelligence 83 (1996) l-58 8. Conclusion So what measures of uncertainty should be used in expert systems? I believe that Bayesian probabilities, upper and lower probabilities, belief functions and first-order possibility measures can all be useful in different types of problems, for instance when the information is in the form of extensive statistical data, bounds on probabilities, multivalued mappings or natural-language judgements respectively. All these measures can be useful in special types of problems, but none of them is adequate as a general model of uncertainty. For example, none of them can adequately model the three qualitative judgements in the football example. Bayesian probabilities are not sufficiently genera1 because, in many problems, information is scarce and judge- ments are imprecise. Belief functions and possibility measures are not sufficiently genera1 because many examples involve lower probabilities that are not even 2-monotone. Up- per and lower probabilities are not sufficiently general because they do not uniquely determine upper and lower previsions and conditional probabilities; upper and lower previsions produce greater precision in inferences and decisions. I suggest that upper and lower previsions, which include the other measures as special cases, are sufficiently general to model the most important types of uncertainty. This raises the question: to what extent is the theory of coherent lower previsions compatible with the other theories? It is compatible with the Bayesian theory as the two theories have a similar behavioural interpretation and the rules of the Bayesian theory are special types of natural extension. So the theory of lower previsions can be regarded as a generalisation of the Bayesian theory; the two theories agree in the special case where all probability models are precise. The theory of Bayesian sensitivity analysis or “probability intervals” is also, to a large extent, compatible with the theory of lower previsions. Possibility theory and Dempster-Shafer theory appear to be less compatible with the theory of lower previsions. However I believe that the theory of lower previsions can incorporate some of the most useful features of these theories. In particular, one of the main contributions of the two theories has been to suggest some flexible and powerful methods for modelling particular types of partial information, notably through multivalued mappings and natural-language judgements. These can be used as methods for assessing lower previsions. The theories differ in their interpretation of uncertainty measures. The interpretation of belief functions and possibility measures is unsettled and controversial, but it seems to me that at least some of the interpretations that have been proposed for these mea- sures are consistent with the behavioural interpretation of lower and upper previsions. Two advantages of the behavioural interpretation are that it relates uncertainty measures to decisions and thereby explains how they can be used, and it leads to coherence principles which can be used to check consistency of models with conclusions. (Both features are lacking in Dempster-Shafer and possibility theory.) The behavioural inter- pretation is sufficiently genera1 to encompass multivalued mappings, inexact judgements and lower envelopes of Bayesian probability measures as sources of lower previsions, yet specific enough to support the principles of coherence and the rules of natural extension. I? Walley/Art@cial Intelligence 83 (1996) 1-58 53 The theories have quite different rules for combining and updating uncertainties, and in this respect they do appear to be incompatible. Dempster’s rule of combination and the minimum rule for combining possibility distributions can, in some problems, produce upper and lower probability models that are incoherent, and in these cases the rules are not compatible with the theory of lower previsions. These rules are controversial. They can lead to intuitive inconsistencies as well as mathematical incoherence and they seem to be applicable only in a limited range of problems. If these rules were given a more restricted role in the Dempster-Shafer theory and in possibility theory then the three theories would be considerably more compatible. Natural extension is an alternative to Dempster’s rule and the minimum rule, and it is worth investigating other rules that preserve coherence but produce stronger conclusions than natural extension. This may be a way of developing the calculus of belief functions and possibility measures. The behavioural interpretation and principles of coherence can impose some much needed discipline on the theories of belief functions and possibility measures, without which these theories can produce inconsistencies. Of course the theory of lower previsions is itself undeveloped in some important repects. Further investigation is needed into the practical problems of modelling, as- sessment and computation. Particular problems are to determine how best to model independence judgements, and how to compute natural extensions and propagate lower previsions efficiently. It is also important to compare the four approaches in some real- istically complex expert systems. I hope that this paper will persuade some people to consider these problems. Acknowledgements I am grateful to Dr. Hitoshi Furuta for inviting me to lecture on this subject in Osaka in April 1990 and for suggesting this paper. Many people commented on an earlier version of the paper [ 1001 which was circulated in April 1991. In particular I must thank Terrence Fine, Frank Hampel and Philippe Smets for stimulating discussions, and Smets, Nit Wilson and George Klir for sending me copies of relevant papers. Nit Wilson contributed some exceptionally detailed and insightful comments on an earlier version and the final version has benefited greatly from his suggestions. References 1 1 1 K.-P. Adlasnig and G. Kolarz, CADIAG-2: computer-assisted medical diagnosis using fuzzy subsets, in: M M. Gupta and E. Sanchez, eds., Approximate Reasoning in Decision Analysis (North-Holland, Amsterdam, 1982) 219-247. 12 1 S.K. Andersen, K.G. Olesen, F.V. Jensen and E Jensen, HUGIN-a shell for building Bayesian belief universes for expert systems, in: Proceedings IJCAI-89, Detroit, MI (Morgan Kaufman% San Mateo, CA, 1989) 1080-1085. 13 1 R. Beyth-Marom, How probable is probable? Numerical translations of verbal probability expressions, J. Forecasting 1 (1982) 257-269. 54 P Walley/Artijicial Intelligence 83 (1996) I-58 14 1 G. Biswas and T.S. Anand, Using the Dempster-Shafer scheme in a mixed-initiative expert system shell, in: L.N. Kanal, T.S. Levitt and J.F. Lemmer, eds., Uncertainty in Artificial Intelligence 3 (North- Holland, Amsterdam, 1989) 223-239. 15 I G. Biswas, M. Oliff and R. Abramczyk, OASES (Operations Analysis Expert System): an application in fiberglass manufacturing, Int. J. Expert Sysf. Res. Appl. 1 (1988) 193-216. 16 I B.C. Buchanan and E.H. Shortliffe, eds., Rule-Eased Expert Systems (Addison-Wesley, Reading, MA, 1984). 17 I J.-C. Buisson, H. Farreny and H. Prade, The development of a medical expert system and the treatment of imprecision in the framework of possibility theory, InjI Sci. 37 (1985) 21 l-226. 18 I R. Buxton, Modelling uncertainty in expert systems, Int. J. Man-Mach. Stud. 31 ( 1989) 415-476. 19 I L.M. De Campos, M.T. Lamata and S. Moral, The concept of conditional fuzzy measure, fnr. J. Intell. Syst. 5 ( 1990) 237-246. I IO I J.E. Cano, S. Moral and J.F. Verdegay, Propagation of convex sets of probabilities in directed acyclic networks, in: Proceedings Fourth IPMU Conference (1992) 289-292. I 1 I I F Cheeseman, Probabilistic versus fuzzy reasoning, in: L.N. Kanal and J.F. Lemmer, eds., Uncertainty in Artificial bzfelligence (North-Holland, Amsterdam, 1986) 85-102. I I2 I L.J. Cohen, The Probable and fhe Provable (Clarendon Press, Oxford, 1977). I I3 I P.R. Cohen, Heuristic Reasoning about Uncertainty: An Art@cial Intelligence Approach (Pitman, London, 1985). 1 I4 I R.G. Cowell, BAIES-a probabilistic expert system shell with qualitative and quantitative learning, in: J.M. Bemardo, J.O. Berger, AI? Dawid and A.F.M. Smith, eds., Bayesian Statistics 4 (Clarendon Press, Oxford, 1992) 595-600. [ IS I A.P. Dempster, Upper and lower probabilities induced by a multivalued mapping, Ann. Math. Stat. 38 (1967) 325-339. [ I6 I D. Dubois and H. Prade, An introduction to possibilistic and fuzzy logic& in: P Smets, A. Mamdani, D. Dubois and H. Prade, eds., Non-Standard Logics for Aufomated Reasoning (Academic Press, London, 1988) 287-326. [ I7 I D. Dubois and H. Prade, Possibility Theory (Plenum Press, New York, 1988). 1 I8 I D. Dubois and H. Prade, Modelling uncertain and vague knowledge in possibility and evidence theories, in: R.D. Shachter, T.S. Levitt, L.N. Kanal and J.F. Lemmer, eds., Uncertainty in Arti’cial Intelligence 4 (North-Holland, Amsterdam, 1990) 303-318. I I9 I D. Dubois and H. Prade, Evidence, knowledge, and belief functions, Int. J. Approx. Reasoning 6 ( 1992) 295-J 19. [ 20 I R.O. Duda, PE. Hart and N.J. Nilsson, Subjective Bayesian methods for rule-based inference systems, in: Proceedings 1976 National Computer Conference, AFIPS 45 (1976) 1075-1082; also in: B.W. Webber and N.J. Nilsson, eds., Readings in Artificial Infelligence (Morgan Kaufmann, Los Altos, CA, 1981) 192-199. 12 I I R.O. Duda and R. Reboh, AI and decision making: the PROSPECTOR experience, in: W. Reitman, ed., Artificial fntelligence Applicationsfor Business (Ablex, Norwood, NJ, 1984) I 11-147. I22 I A. Dutta, Reasoning with imprecise knowledge in expert systems, Inf Sci. 37 ( 1985) 3-24. I23 1 R. Fagin and J.Y. Halpern, A new approach to updating beliefs, in: PP. Bonissone, M. Henrion, L.N. Kanal and J.F. Lcmmer, eds., Uncertainty in Arttficial Intelligence 6 (North-Holland, Amsterdam, 1991) 347-374. 125 I 24 I H. Farreny, H. Pnde and E. Wyss, Approximate reasoning in a rule-based expert system using possibility theory: a case study, in: H.J. Kugler, ed., Information Processing ‘86 (North-Holland, Amsterdam, 1986) 407-413. K.W. Fertig and J.S. Breese, Interval influence diagrams, in: M. Henrion, R.D. Shachter, L.N. Kanal and J.F. Lemmer, eds., Uncertainty in Artificial Intelligence 5 (North-Holland, Amsterdam, 1990) 149-161. M. Fieschi, M. Joubert, D. Fieschi, G. Botti and M. Roux, A program for expert diagnosis and therapeutic decision, Med. I$ 8 ( 1983) 127-135. T.L. Fine, Theories of Probability (Academic Press, New York, 1973). B. de Finetti, Theory of Probability Vol. I (Wiley, London, 1974). B. de Finetti, Theory of Probability Vol. 2 (Wiley, London, 1975). 1% I 27 128 I 29 I? Walley/Art$cial Intelligence 83 (1996) I-58 55 I 30 ] L.C. van der Gaag, Computing probability intervals under independency constraints, in: l?P Bonissone, M. Hencion, L.N. Kanal and J.F. L.emmer, eds., Uncertainty in Art@cial Intelligence 6 (North-Holland, Amsterdam, 199 1) 457-466. I 3 I I W.A. Gale, ed., Artificial Intelligence and Statistics (Addison-Wesley, Reading, MA, 1986). 1321 R. Gales, Foundations for a theory of possibility, in: M.M. Gupta and E. Sanchez, eds., Fuzzy 133 [34 I35 1% 137 138 I39 I42 I E.T. Jaynes, Pupers on Probability, Sfutistics and Statistical Physics (Reidel, Dordrecht, 1983). I43 ) F.V. Jensen, SK. Andersen, U. Kjaerulff and S. Andreassen, MUNIN-on the case for probabilities in medical expert systems, Lecture Notes in Medical Informutics 33 (Springer, Berlin, 1987). 1441 A.C. Kak, K.M. Andress, C. Lopez-Abadia, MS. Carol1 and J.R. Lewis, Hierarchical evidence accumulation in the PSEIKI system, in: M. Henrion, R.D. Shachter, L.N. Kanal and J.F. Lemmer, eds., Uncertainty in Artijcial Intelligence 5 (North-Holland, Amsterdam, 1990). 145 I S. Kellt, Words of estimated probability, Stud. Intell. 8 ( 1964) 49-65. I46 ] G.J. Klir and T.A. Folger, Fuzzy Sets, Uncertainty, und Information (Prentice-Hall, Englewood Cliffs, NJ, 1!)88). 147 I F? Krause and D. Clark, Representing Uncertain Knowledge (Kluwer, Dordrecht, 1993). I48 I H.E. Kybuxg Jr, Bayesian and non-Bayesian evidential updating, Art& Intell. 31 (1987) 271-293. 149 I K.B. Laskey, M.S. Cohen and A.W. Martin, Representing and eliciting knowledge about uncertain evidence and its implications, IEEE Trans. Sysf. Man Cybern. 19 ( 1989) 536-557. I 50 I S.L. Lauritzen and D.J. Spiegelhalter, Local computations with probabilities on graphical structures and their application to expert systems (with discussion), J. Roy. Stat. Sot. Ser. B 50 (1988) 157-224. 15 I I J. Lebailly, R. Martin-Clouaire and H. Prade, Use of fuzzy logic in rule-based systems in petroleum geology, in: E. Sanchez and L.A. Zadeh, eds., Approximate Reasoning in Intelligent Systems, Decision und Control (Pergamon Press, Oxford, 1987) 125-144. I&m&on and Decision Processes (North-Holland, Amsterdam, 1982) 183-195. R. Giles, Semantics for fuzzy reasoning, Int. .I. Man-Mach. Stud. 17 (1982) 401-415. J. Gordon and E.H. Shortliffe, A method for managing evidential reasoning in a hierarchical hypothesis space. Artif: Intell. 26 (1985) 323-357. I B.N. Grosof, An inequality paradigm for probabilistic knowledge, in: L.N. Kanal and J.F. Lemmer, eds., Uncertainty in Artificial Intelligence (North-Holland, Amsterdam, 1986) 259-275. J.Y. Halpem and R. Fagin, Two views of belief: belief as generalized probability and belief as evidence, Artif: Intell. 54 (1992) 275-317. D. Heckerman, E. Horvitz and B. Nathwani, Toward normative expert systems 1: the PATHFINDER project, Methods Inky Med. 31 (1992) 90-105. M. Henrion, J.S. Breese and E.J. Horvitz, Decision analysis and expert systems, AI Muguzine 12 (1991) 64-91. Y.-T. Hsia and PP. Shenoy, An evidential language for expert systems, in: 2. Ras, ed., Methodologies for Intelligenf Systems 4 (North-Holland, New York, 1989) 9-16. I 40 1 l?J. Huber, Robust Stafisfics (Wiley, New York, 198 1) 141 ) J.-Y. Jaffmy, Bayesian updating and belief functions, IEEE Trans. Syst. Man Cybern. 22 ( 1992) 1144- 1152. I 52 I J.F. Hemmer, Generalised Bayesian updating of incompletely specified distributions, Large Scale Syst. 5 (1983) 51-68. I53 I K.S. Leung, W.S.F. Wong and W. Lam, Applications of a novel fuzzy expert system shell, Expert Syst. 6 (1939) 2-10. I54 I 1. Levi, Potential surprise: its role in inference and decision making, in: L.J. Cohen and M. Hesse, eds., Applications of Inductive Logic (Oxford University Press, Oxford, 1980) l-27. I55 I 1. Levi, Consonance, dissonance and evidentiary mechanisms, in: P Gbdenfors, B. Hansson and N.E. Sahlin, eds., Evidentiury Value, Library of Theoria 15 (Gleerups, Lund, 1983) 27-43. I 56 I D.V. Lindley, Making Decisions (Wiley, London, 197 1). I57 I R.P. Loui, Interval-based decisions for reasoning systems, in: L.N. Kanal and J.F. Lemmer, eds., L’ncertuinty in Artijicial Intelligence (North-Holland, Amsterdam, 1986) 459-472. 158 I J.D. Lowrance, Automated argument construction, J. Stat. Planning Inference 20 (1988) 369-387. I59 I R.C. Moore, Semantical considerations on nonmonotonic logic, Art$ Intell. 25 ( 1985) 75-94. I60 I N.J. Nilsson, Probabilistic logic, Artif Intell. 28 (1986) 71-87. 56 P. Walley/Ar~ifcial Inrelligence 83 (1996) l-58 I61 I P. Orponen, Dempster’s rule is #P-complete, A@ Intell. 44 ( 1990) 245-253. I62 I G. Paass, Probabilistic logic, in: P Smets, A. Mamdani, D. Dubois and H. Prade, eds., Non-Standard Logic.7 fiw Aufumated Reasoning (Academic Press, London, 1988) 213-25 1. I63 I J. Pearl, Probabilistic Reasoning in Intelligent Systems (Morgan Kaufmann, San Mateo, CA, 1988). I64 I J. Pearl, Reasoning with belief functions: an analysis of compatibility, Int. J. Approx. Reasoning 4 (1990) 363-389. I65 I J. Pearl, Rejoinder to comments on “Reasoning with belief functions: an analysis of compatibility”, Int. J. Approx. Reasoning 6 (1992) 425-443. 1661 J.R. Quinlan, Inferno: a cautious approach to uncertain inference, Cornput. J. 26 ( 1983) 255-269. I67 I R. Reiter, A logic for default reasoning, Artif Intell. 13 ( 1980) 81-132. 168 1 R. Reiter, Nonmonotonic reasoning, Ann. Rev. Comput. Sci. 2 (1987) 147-186. I69 I A. Saffiotti and E. Umkehrer, PULCINELLA: a general tool for propagating uncertainty in valuation networks, in: B.D. D’Ambrosio, P. Smets and PP. Bonissone, eds., Proceedings Seventh Inrernarional Cwtference on Uncertainfy in AI, Los Angeles, CA (Morgan Kaufmann, San Mateo, CA, 1991) 323-331. I70 I G.L.S. Shackle, Decision, Order; and 7ime in Human Affairs (Cambridge University Press, Cambridge, 1969). I 7 I I G. Shafer, A Mathematical Theory of Evidence (Princeton Universtiy Press, Princeton, NJ, 1976). I72 I G. Shafer, Constructive probability, Synrhese 48 ( 1981) L-60. I73 I G. Shafer, Belief functions and parametric models (with discussion), J. Roy. Star. Sot. Ser. E 44 ( 1982) 322-352. I74 I G. Shafer, Probability judgement in artificial intelligence and expert systems, Star. Sci. 2 ( 1987) 3-16. I75 I G. Shafer, Perspectives on the theory and practice of belief functions, Int. J. Approx. Reasoning 4 ( 1990) 323-362. I76 I G. Shafer, Rejoinders to comments on “Perspectives on the theory and practice of belief functions”, Int. .I. Approx. Reasoning 6 ( 1992) 445-480. I77 I G. Shafer and R. Logan, Implementing Dempster’s rule for hierarchical evidence, A@ Intell. 33 (1987) 271-298. I78 1 P. Shenoy and G. Shafer, Propagating belief functions with local computations, IEEE Expert 1 (3) ( 1986) 43-52. I79 I F.K.J. Sheridan, A survey of techniques for inference under uncertainty, Artif: Intell. Rev. 5 ( 1991) 89-l 19. I80 I E.H. Shortliffe, Computer-Based Medical Consultations: MYCIN (Elsevier, New York, 1976). 181 182 I 183 I 84 E.H. Shortliffe and B.G. Buchanan, A model of inexact reasoning in medicine, Math. Biosci. 23 ( 1975) 35 l-379. P. Smets, Belief functions, in: P Smets, A. Mamdani, D. Dubois and H. Prade, eds., Non-Standard Logics for Automated Reasoning (Academic Press, London, 1988) 253-286. P. Smets, The transferable belief model and other interpretations of Dempster-Shafer’s model, in: PP. Bonissone, M. Henrion, L.N. Kanal and J.F. Lemmer, eds., Uncertainty in Arrifcial Intelligence 6 (North-Holland, Amsterdam, 1991) 375-383. P Smets, Resolving misunderstandings about belief functions, Int. J. Approx. Reasoning 6 ( 1992) 321-344. I85 1 P. Smets and R. Kennes, The transferable belief model, Artif: Intell. 66 (1994) 191-234. I86 I P. Smets, A. Mamdani, D. Dubois and H. Prade, eds., Nun-Standard Logics for Automated Reasoning (Academic Press, London, 1988). I 87 ] C.A.B. Smith, Consistency in statistical inference and decision (with discussion), J. Roy. Stat. Sot. Ser. B 23 (1961) l-37. [ 88 I M. Smithson, Ignorance and Uncertainty (Springer, Berlin, 1989). 1891 D.J. Spiegelhalter, A statistical view of uncertainty in expert systems, in: W.A. Gale, ed., Arrifcial Intelligence and Staristics (Addison-Wesley, Reading, MA, 1986). ( 901 D.J. Spiegelhalter, AI? Dawid, S.L. Lauritzen and R.G. Cowell, Bayesian analysis in expert systems (with discussion), Stat. Sci. 8 (1993) 219-283. [ 9 I I D.J. Spiegelhalter and R.P. Knill-Jones, Statistical and knowledge-based approaches to clinical decision- support systems, with an application in gastroenterology (with discussion), J. Roy. Stat. Sot. Ser. A 147 ( 1984) 35-76. P Walley/Arti$cial Intelligence 83 (1996) l-58 57 [ 92 I T.M. Strat, Decision analysis using belief functions, ht. J. Approx. Reasoning 4 ( 1990) 391-417. 1931 C. Smdberg and C. Wagner, Generalized finite differences and Bayesian conditioning of Choquet capacities, Unpublished manuscript (1991). I94 I B. Tessem, Interval probability propagation, Int. J. Approx. Reasoning 7 ( 1992) 95-120. I 95 I F. Vaorbraak, On the justification of Dempster’s rule of combination, Artif Intell. 48 ( 199 1) 17 l- 197. I 96 I F? Walley, Coherent lower (and upper) probabilities, Statistics Research Report, University of Warwick, Coventry (1981). I97 I P. Walley, The elicitation and aggregation of beliefs, Statistics Research Report, University of Warwick, Coventry (1982). I98 I F? Walley, Belief function representations of statistical evidence, Ann. Stat. 15 (1987) 1439-1465. I99 1 P. Walley, Statistical Reasoning with Imprecise Probabilities (Chapman and Hall, London, 1991). I 1001 P. Walley, Measures of uncertainty in expert systems, Statistics Research Report, Department of I101 I102 [ 103 Mathematics, University of Western Australia, Perth ( 199 1) . I? W;rlley, Inferences from multinomial data: learning about a bag of marbles (with discussion), J. Roy. Stat. Sot. Ser. B 58 (1996) 3-57. P. Walley, Statistical inferences based on a second-order possibility distribution (with discussion), J. Ant. Srat. Assoc. 91 ( 1996). I! Walley and FM. Campello de Souza, Uncertainty and indeterminacy in assessing the economic viabil!ity of energy options: a case study of solar heating systems in Brazil, Energy Syst. Policy 14 ( 19913) 28 I-304. I 104 I P. Walley and T.L. Fine, Varieties of modal (classificatory) and comparative probability, Synthese 41 ( 197’9) 321-374. I 105 I P. Walley and T.L. Fine, Towards a frequentist theory of upper and lower probability, Ann. Stat. 10 (1982) 741-761. I 1061 T.S. ‘Wallsten, D.V. Budescu, A. Rapoport, R. Zwick and B. Forsyth, Measuring the vague meanings of ptobability terms, J. Exper. Psych. General 115 (1986) 348-365. I 107 1 L.A. Wasserman, Comments on Shafer’s “Perspectives on the theory and practice of belief functions”, Int. J. Approx. Reasoning 6 (1992) 367-375. I 108 I L.A. Wasserman and J. Kadane, Bayes’ theorem for Choquet capacities, Ann. Stat. 18 (1990) 132% 1339 I 109 I S.R. Watson, J.J. Weiss and M.L. Donnell, Fuzzy decision analysis, IEEE Trans. Svst. Man Cvbern. 9 1110 1 III I112 I113 (1979) 1-9. P.M. Williams, Indeterminate probabilities, in: M. Przelecki, K. Szaniawski and R. Wojcicki, eds., Formal Methods in the Methodology of Empirical Sciences (Reidel, Dordrecht, 1976) 229-246. N. Wilson, A Monte-Carlo algorithm for Dempster-Shafer belief, in: B.D. D’Ambrosio, P Smets and PP Elonissone, eds., Proceedings Seventh International Conference on Uncertainty in AI, Los Angeles, CA (Morgan Kaufmann, San Mateo, CA, 1991) 414-417. N. Wilson, The combination of belief: when and how fast?, Int. J. Approx. Reasoning 6 (1992) 377-,388. N. Wilson and S. Moral, A logical view of probability, in: A. Cohn, ed., Proceedings Eleventh European Conference on Artificial Intelligence (Wiley, London, 1994) 386-390. I 1 14 I H. XII, An efficient implementation of belief function propagation, in: B.D. D’Ambrosio, E! Smets and PP. Eionissone, eds., Proceedings Seventh International Conference on Uncertainty in AI, Los Angeles, CA (Morgan Kaufmann, San Mateo, CA, 1991) 425-432. I 1 IS I R.R. Yager, An introduction to applications of possibility theory (with discussion), Human Syst. Manage. 3 (1983) 246-269. I I 16 I L.A. Zadeh, Fuzzy sets, I@ Control 8 ( 1965) 338-353. I I 17 I L.A. Zadeh, The concept of a linguistic variable and its application to approximate reasoning (Part 3), I@ Sci. 9 (1976) 43-80. I I 18 I L.A. Zadeh, Fuzzy sets as a basis for a theory of possibility, Fuzzy Sets Syst. 1 (1978) 3-28. I II91 L.A. Zadeh, A theory of approximate reasoning, in: J. Hayes, D. Michie and L.I. Mikulich, eds., Machine Intelligence 9 (Halstead Press, New York, 1979) 149-194. I 1201 L.A. Zadeh, The role of fuzzy logic in the management of uncertainty in expert systems, Fuzzy Sets Syst. 11 (1983) 199-227. 58 P: Walley/Artijcial Intelligence 83 (1996) I-58 1 1211 L.A. Zadeh, Is probability theory sufficient for dealing with uncertainty in AI: a negative view, in: L.N. Kanal and J.E Jxmmer, eds., Uncertainty in Art$cial Intelligence (North-Holland, Amsterdam, 1986) 103-l 16. 
work_jjzvyc6pn5h57haekram3wec2i	----	An integrated and intelligent DSS for manufacturing systems An integrated and intelligent DSS for manufacturing systems Dursun Delen a,*, David B. Pratt b a Department of Management Science and Information Systems, William S. Spears School of Business, Oklahoma State University, 700 N. Greenwood Avenue, Tulsa, OK 74106, USA b School of Industrial Engineering and Management, College of Engineering, Architecture and Technology, Oklahoma State University, Stillwater, OK, USA Abstract There is no room for error in making managerial decisions in today’s global business environment marked by mergers, acquisitions, and increasing economic instability. In order to succeed in this ever so demanding, fast pace business environment, managers need integrated and intelligent information systems capable of supporting them throughout the decision-making lifecycle, which starts with structuring a problem from a given set of symptoms and ends with providing the information needed to make the decision. In this manuscript, we report on a collaborative research effort whose aim has been to fill this need by developing novel concepts and to demonstrate the viability of these concepts within an advanced modeling environment. q 2005 Elsevier Ltd. All rights reserved. Keywords: Intelligent decision support; Structured problem; Automated tool selection; Manufacturing systems modeling; Base model; Expert systems 1. Introduction Today’s highly competitive, fast pace business environ- ment makes it an absolute requirement on behalf of the managers to continuously make the best decisions in the shortest possible time. The notion of ‘learning from mistakes’ has left its place to ‘one strike and you’re out’ reality. That is, there is no room for error in making managerial decisions in this global environment marked by mergers, acquisitions, and ever-increasing economic instability. Success (or mere survival) depends on quickly aligning the organizational resources towards meeting (and exceeding) the actual (and perceived) needs and wants of the customer. In order to succeed in such an unforgiving environment, managers need integrated ‘intelligent’ decision support systems (DSS) that are capable of using a wide variety of models along with data and information resources available to them at various internal and external repositories. In this paper, we report on the up-to-now results of an advanced modeling research project. 1 The main goal of this project has been to develop novel concepts and necessary technologies to make decision support systems (DSS) readily accessible to practitioners and decision makers. To the best of our knowledge, to this date, this research is the only one to attempt to conceptualize, design and develop a comprehensive approach to intelligent and integrated DSS for manufacturing managers. Most other studies (similar to the one reported herein) limit themselves to a specific manufacturing system problem (e.g. scheduling, product design, layout, procurement, maintenance, etc.) or to a specific modeling and analysis tool (simulation, queueing, Petri nets, etc.), and therefore lack in addressing the real needs of the decision makers in a feasible manner. The main contribution of this research (to the body-of- knowledge in general and to the manufacturing systems modeling in specific) is that it advances the state-of-the-art in model-based decision support systems by addressing the most commonly pronounced shortcomings of the tra- ditional methodologies including lack of model reusability (Mize, Bhuskute, Pratt, & Kamath, 1992; Pratt, Farrington, Basnet, Bhuskute, Kamath, & Mize, 1999), lack of Expert Systems with Applications 30 (2006) 325–336 www.elsevier.com/locate/eswa 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.07.017 * Corresponding author. Tel.: C1 918 594 8283; fax: C1 918 594 8281. E-mail address: delen@okstate.edu (D. Delen). 1 This research project has been receiving funding from NSF, OCAST and AT&T Foundation since 1992. http://www.elsevier.com/locate/eswa integration of models to real world (Dalal, Kamath, Kolarik, & Sivaraman, 2004; Delen, Dalal, & Benjamin, 2005), and lack of model utility/accessibility (Beynon, Rasmequan, & Russ, 2004; Delen, Pratt, & Kamath, 1996). First, most decision models have been viewed as single purpose, throwaway efforts. They are built from scratch to address a particular problem or question, and then are often discarded with no thought given to additional use. This single-use, throwaway mentality of decision modeling is very expensive, time consuming and wasteful. Second, the models are isolated from the environment that they represent, portraying a static depiction of the system, specific to the problem being analyzed, lacking the much needed integration with the organizational data/infor- mation sources, which are mostly dynamic in nature. Third, lack of usability/accessibility of decision models by non-modeling specialists has limited their widespread usage and value. In order to address these shortcomings we have developed several novel concepts (e.g. base model concept, separation concept, tool independent model representation concept, knowledge-based expert advisors) along with a prototype software implementation where the viability of these concepts are measured as part of the advanced modeling environment. Fig. 1 illustrates our vision of the decision-making life cycle for manufacturing systems managers. As shown, decision makers are constantly exposed to problems and opportunities (as is the case in many business environments nowadays). In order to solve these problems and/or take advantage of the opportunities, they have to make decisions, which may result in changing the manufacturing system. In doing so, they have to be effective (making the correct decisions) and they have to be efficient (making them in a timely fashion). Therefore, manufacturing managers need all the help they can get to make the best possible decisions in the shortest possible time. This is where the need for a new generation of information systems emerges: DSS that are capable of supporting decision makers throughout the decision-making life cycle starting from problem definition and going all the way to results interpretation. In this study, we designed and developed an integrated software environment (and its prototype implementation) for an intelligent decision support system for manufacturing systems (IDSS–MS), capable of providing help to managers throughout the decision making life-cycle, which includes (1) structuring a problem from a given set of symptoms, (2) once the problems is structured, determining the best analysis tool (i.e. model type) to address the problem, (3) automatically generating the executable models specific to the structured problem, (4) conducting the analysis, and (5) providing the results back to the decision maker in an easily understandable format. The remainder of the paper is organized as follows. Section 2 gives a brief review of the related literature and places this research in its respective place. Section 3 introduces the architecture of the proposed solution and explains the innovative concepts underlying the architec- ture. Section 4 explains the software implementation and development of the knowledge bases. Section 5 summarizes the validation process and the usability issues. The paper ends with Section 6 where the concluding remarks and further research directions are identified. BASE MODEL Intelligent Manufacturing DSS Problems Opportunities Input Interaction Output Changes Construct/Update Characteristics Mfg. System Model Builder Decision Maker Fig. 1. Manufacturing systems decision making life-cycle. D. Delen, D.B. Pratt / Expert Systems with Applications 30 (2006) 325–336326 https://isiarticles.com/article/3590 
work_jldtmks7svht3leto6at5f65jq	----	Archive - California Agriculture Skip to Content Menu California Agriculture Share Print Site Map Give Facebook Twitter Reddit Pinterest Tumblr Delicious LinkedIn StumbleUpon Enter Search Terms Search Home Archive Advanced search About About Contact us Associate editors Subscribe Submit articles Share Print Site Map Enter Search Terms Give x Facebook Twitter Reddit Pinterest Tumblr Delicious LinkedIn StumbleUpon University of California, Agriculture and Natural Resources University of California California Agriculture Home Archive Advanced search About About Contact us Associate editors Subscribe Submit articles University of California California Agriculture Archive University of California Division of Agriculture and Natural Resources Site Information Get PDF Reader sbnum=9680 | pagenum=52363 © 2021 Regents of the University of California Nondiscrimination Statement Accessibility Webmaster Email: wsuckow@ucanr.edu 
work_jlfwf3dlofbqhni3odkzekgvzy	----	PDF hosted at the Radboud Repository of the Radboud University Nijmegen The following full text is a publisher's version. For additional information about this publication click this link. http://hdl.handle.net/2066/112320 Please be advised that this information was generated on 2021-04-06 and may be subject to change. http://hdl.handle.net/2066/112320 Analytrca C/zumcu Acta, 272 (1993) 41-51 Elsevler Science Pubhshers B V . Amsterdam 41 Expert systems in chromatography. Results of the ESCA project L Buydens Department of Analyttcal Chemistry, Kathoheke Unrverxte~t N&]megen, Toemoorveld, 6525 ED Ntpegen (Netherlands) P Schoenmakers Phrlps Research, P 0 Box 80000,5600 JA Emdhoven (Netherlands) F Marls and H Hmdnks Organon Intematwnal BC: Analytrcal R & D Laboratones, P 0 Box 20,534O BH Oss (Netherlands) (Recewed 9th July 1992) Abstract The final results of the ESCA project (Expert Systems for Chemical Analysis) are presented This IS one of the major projects m the field of expert systems for chromatography Expert systems have been developed that cover the unportant areas of method development m LC In the last part of the project attention was concentrated on two Lssues, the study of mtegratlon posslblhtles of the drfferent stand-alone systems and the Important aspect of vabdatlon and evaluation of the developed expert systems The mtegratlon studies and the results of the vahdatlon and evaluation are discussed Keywords Expert systems, Llquld chromatography, ESCA project The results of automation efforts by manufac- turers of chromatographlc instruments have led to an increased apphcablhty of chromatographlc mstruments for routme analysis The bottleneck of analysis IS situated mainly m the development of an optnnum method and m the interpretation of the results These processes usually require a lot of expertise and experience to solve the prob- lems that arise for each particular case Also, the quality control stage becomes an mcreasmgly nn- portant aspect to be automated As a result of automation, the numbers of analyses and results have grown so much that automatic quahty mom- tormg 1s necessary In view of the mcreasmg demands of good laboratory and management practice (GLP and GMP), this aspect will become even more unportant The mcorporatlon of ex- pertlse and expenence m instruments 1s therefore the next step to be taken Expert systems are software programs m which human expertise 1s unplemented Therefore, they seem to be the right approach for further au- tomation of instruments In other areas of chem- Istry they have already been demonstrated to be useful [l-3] In chromatography, a large amount of research has been carned out m a Jomt re- Correspondence to L Buydens, Department of Analytical Chemistry, Kathoheke Umversltelt NlJmegen, Toemooweld, 6525 ED Nqmegen (Netherlands) 0003-2670/93/$06 00 0 1993 - Elsevler Science Publishers B V All nghts reserved 42 L Buydens et al /Anal Chm Acta 272 (1993) 41-51 selection of application domaln t t ’ knowledge acquisition selection of tools knowledge implementation- validation Fig 1 Dlfferent steps III the development process of expert systems search proJect on the apphcablhty of expert sys- tems m chemical analysis (ESCA) [41 Partners from Industry and umversltles have cooperated to study the posslblhtles of the expert system ap- proach m LC In the first stage the development of LC methods for pharmaceutical compounds was selected as an apphcatlon area Within this area expert systems were developed that are rep- resentative of the whole area of method develop- ment selection of lmtlal method parameters, op- tlmlzatlon of selectlvlty and mstrumentatlon and finally vahdatlon of the method obtained As part of the project all aspects of the expert system building process have been investigated Aspects of integration and cooperation of expert systems have also been covered The project started m May 1987 and officially finished m May 1990 Durmg this period mterme- dlate results and fmdmgs have been commum- cated by means of presentations and so far about 25 papers have been published m mternatlonal Journals and numerous lectures and posters have been presented In this paper an overview of the most nnportant results 1s presented EXPERT SYSTEM DEVELOPMENT PROCESS The process of the development of expert sys- tems consists of several stages and has many different aspects It requires close cooperation between workers who have the necessary apphca- tlon knowledge and expertise and those response- ble for the lmplementatlon of the expert systems (“knowledge engmeers”) The different tasks are shown schematically m Fig 1 It shows clearly the sequence that has to be followed, the interaction between tasks and the loops that can occur The aim of the project was to study and demonstrate the apphcatlon of expert systems m chromatography Therefore, it was felt that a single apphcatlon was too lnmted to demonstrate the objective For that reason, a number of (rela- tively) small domams were selected based on cn- terra of usefulness, difficulty and vanety These domains were derived from LC method develop- ment as shown m Fig 2 This process resulted originally m four expert systems method selection and retention optimization 1 selectivity optimaization .l. system optimization 1 method validation Fig 2 Stages m the chromatographlc method development L Buydens et al /Anal Chm Acta 272 (1993) 41-51 43 Knowledge acqulsltlon 1s the process of ex- tracting knowledge as complete as possible from the chromatography expert This knowledge should then be unplemented m a choosen tool by the knowledge engmeer, which m this instance was either a chemometrlclan or a software spe- cialist The latter reqmres more time to become familiar with the domain knowledge, but may reahse better systems m terms of completeness, consistency and appearance Chemometnaans, on the other hand, can acquire more chemical knowledge m a shorter period of time, but the quality of the resulting software can be inferior To implement expert systems m a computer, a software tool 1s required This tool can be a standard computer language such as Pascal or C However, the lmplementatlon of expert systems with classical languages requires considerable software engineering expenence and a large ef- fort m terms of manpower In recent years dedl- cated tools for developing expert systems have become available These tools are often referred to as “expert system shells” The available tools range from simple to very sophlstlcated One of the purposes of ESCA was also to evaluate the sultablhty of these tools Therefore, it was neces- sary to select some suitable tools to implement the apphcatlons This selection was based on the lmplementa- tlon of a small test knowledge base m different tools of varying size (large, mid-sized and small, see Table 1) The test knowledge base contained essential features of the final knowledge and was obtained from earlier work by De Smet et al [5] TABLE I List of tools evaluated Size Name Orlgm Runmng on Small Delfi 2 Medmm Goldworks a KESa Mylog Nexpert object a Large Sl KEE Knowledge Craft Netherlands PC USA PC USA PC France PC USA PC USA WorkstatIon USA WorkstatIon USA WorkstatIon TABLE 2 Summary of development environment of the expert systems Domam Shell Expertise centre a Know- ledge engl- neermg centre a Method selectlon and KES VUB VUB retention optimization Organon Optlmlzatlon systems Selectivity optlmlzatlon KES WB WB Phlhps NL System optlmlzatlon Nexpert Phdlps NL Phlhps object + Hamburg Pascal Method vahdatlon systems Repeatablhty system Gold- works Umcam UK KUN Ruggedness system Gold- works Umcam UK KUN a WB = Free Umverslty Brussels, Belgmm, KUN = Cathohc Umverslty Nljmegen, Netherlands, Phlhps NL = Research Lab, Emdhoven, Netherlands, Phdlps Hamburg = Research lab, Hamburg, Germany, Organon = AnalytIcal R&D Labs, Oss, Netherlands, Umcam UK = Umcam, Cambndge, UK In general, it was concluded that large tools such as KEE or Knowledge Craft were too comph- cated They also require large workstations or microcomputers to run on The small tools were clearly not adequate, e g , limited lmplementatlon posslblhtles, limited or no access to external databases and poor quality of the end user and knowledge engineering interface It was finally decided to use a selected group of mid-sized tools, which had the additional advantage of run- ning on normal PCs [6] Table 2 summarizes the eventual tools from which the systems were built Expert systems can only be expected to be useful m practice if they are reliable Conven- tional software products can be tested thoroughly through a number of standard procedures How- ever, for testing expert systems no standard pro- cedures exist With expert systems both the soft- ware and the knowledge base must be reliable and correct The fact that the knowledge base often contams a lot of heurlstlc knowledge poses specific problems to the testing phase [7,8] Con- sidering the increasing demands of GLP, this part m the development process cannot be overestl- d These were the tools selected 44 L Buydens et al /Anal Chun Acta 272 (1993) 41-51 mated In view of this, considerable effort was devoted to the vahdatlon and evaluation of expert systems during the last part of the ESCA project EXPERT SYSTEMS OF ESCA trast to DASH Compounds that are subjected to a purity check usually contam less then 5% of unknown lmpuntles Optmuzatlon 1s then usually not required LABEL was added to be able to study the integration of method selection systems with optlmlzatlon expert systems The expert systems that were developed m the ESCA project can be dlvlded mto two categories stand-alone systems and integrated systems A complete overview 1s given m Table 3 In this section the stand-alone systems are considered LIT 1s a small expert system that helps to select all important parameters of a literature method and that checks whether a hterature method can be treated by SLOPES Method selection always starts with the choice of the chromatographlc mode, be It GC or LC In LC a further refinement can be made by choos- mg, for example, the normal- or reversed-phase mode, and further a C, or a cyano phase In this way a decision tree can be built When the experunental results are not satls- factory, all three expert systems have an exten- sion by which adaptations of the method are suggested m order to obtain an acceptable reten- tion range of the compounds DASH (Drug Analysis System m HPLC) 1s the system that assists m the selection of LC starting condltlons for the purity check of pharmaceutical compounds Because of the complextty of the relationship between the structure (input) and suitable percentage of modifier (output), this sys- tem was developed only for heterocychc basic compounds As most of these compounds are new chemical entitles, there 1s no literature avall- able on LC analyses for these compounds [9l The next step m method development 1s selec- tivity optlmlzatlon Thus step typically mvolves the optlmlzatlon of the mobile phase composition m order to obtain an optimum dlstrlbutlon of the peaks over the chromatogram LABEL 1s an expert system that was devel- oped by one of the partners (VUB) before ESCA started It selects a method for the LC of drugs m pharmaceutical formulations (label clann analy- sis) [5] It was included m the project because It covers he sltuatlon that one sample must be analysed Ior different compounds This 1s m con- SLOPES (SeLectlWy OPtlmlzatlon Expert System) 1s an expert system which typically ad- dresses one of the nnportant aspects of selectlvlty optmuzatlon Imtlally attention was focused on the selection of an appropriate optlmlzatlon cn- terlon In the past, many optnnlzatlon crlterla have been put forward However, it should be recogmzed that a smgle criterion 1s not always the best one m all situations [lo] SLOPES will help the chromatographer to select the most approprr- ate criterion, which will then be used to Judge the quahty of the chromatogram [ill This selection of a criterion depends, for example, on the se- lected expernnental design and on the ObJectwe of the optumzatlon (e g , best spreadmg of peaks TABLE 3 OvervIew of the ESCA expert systems Method develoument Stand-alone Integrated stage 1 Imtlal method selection and retention optumzatlon 2 Selectivity optimization 3 System optmuzation 4 Vahdatlon a Names are explamed m the text expert systems a DASH, LABEL, LIT SLOPES SOS REPS RES expert systems (INT) a INTI DASH LABEL LIT + SLOPES INT II INT III SOS + SOS + REPS RES L Buydens et al /Anal Chun Acta 272 (1993) 41-51 45 or muumum analysis tune) Once the optnnum flow-rate It predicts also the reqmred analysis selectlvlty has been obtamed the mobde phase tune and the crltlcal resolution A result of a composltlon and stationary phase are kept con- consultation and the experunental venficatlon 1s stant for the next step, system optmuzatlon shown m Fig 3 SOS (System Optnmzatlon expert System [12]) The final step m method development 1s the can be used here to select a column with the vahdatlon of the method This means that the shortest analysis tune from a column set grven by quality of the results should be guaranteed to a the user In addltron, the user should also provide certain extent The nnportance of validation 1s a set of avadable detector cells and a hst of stdl increasing m view of mcreasmg GLP de- allowed time constants Fmally, some hnuts mands The level of vahdatlon depends mamly on should be given, such as the required mmmmm the intended use of the method A higher level of resolution between a relevant pair of peaks, maxi- vahdatlon IS required if, for example, the method mum pressure drop and maxunum flow-rate 1s to be used m a large number of laboratones Wlthm these constraints SOS recommends the over a long period of tnne Methods to be used column, Instrument parameters and optimum for regulatory analysis need the highest level of a Time (min) Fig 3 Optlmrzatlon mth the SOS expert system (a) Chromatogram before optlmlzatlon, (b) chromatogram obtamed with the condltlons as advtsed by SOS, by which an analysis tnne and a resolution of 6 2 mm and 16, respectwely, were predrcted 46 L Buydens et al /Anal Chm Acta 272 (1993) 41-51 vahdatlon Vahdatlon of a method requires the testing of its speclficlty, precision, accuracy and hnutatlons These so-called performance charac- teristics may be tested by vahdatlon procedures Many of these procedures make use of complex mathematics and rely on statistical designs, such as the Plackett-Burman design (e g , for precl- slon testing) Expert systems can be of great help here m guiding the (mexpenenced) user m the set-up of such advanced designs, m the calcula- tion and m the interpretation of the results In the ESCA project, precision testing was chosen as the most challenging Item m method vahdatlon to demonstrate the apphcablhty of expert systems Precision testing mvolves, m fact, three separate parts repeatability, reproduclblhty and rugged- ness In the repeatability test the analysis 1s re- peated a number of times under Identical condl- tlons Thrs 1s m contrast to the reproduclblhty test, where different conditions, e g , different instruments m different laboratories, are applied Finally, m a ruggedness test the effect of small changes m the operating condltlons, 1 e , temper- ature, flow-rate and mobile phase composltlon, on precision 1s tested Repeatability and rugged- ness testing were the SubJect of different expert systems REPS (Repeatablhty testing System [13,14]) is an expert system m which Goldworks software is combined with the Lotus l-2-3 spreadsheet package The expert system 1s used to select test procedures for repeatability Based on the usage reqmrements, an experimental design IS set up The spreadsheet can run the algorithms and cal- culates the variances for peak areas and heights and retention times The expert system 1s able to interpret the results and to perform a diagnosis based on how the above parameters vary to- gether An example of a rule can illustrate how a diagnosis proposal IS reached For instance, if the variance of retention time and the variance m peak areas are large, and the variance of peak heights 1s small, then it 1s concluded that the problem of repeatablhty 1s unpreclslon of the flow-rate RES (Ruggedness Expert System [15,161) IS a modular expert system m which the Goldworks software 1s combined with the procedural lan- guage C It 1s intended to assist m the proper set-up of a complete ruggedness test This m- volves heurlstlcal (experience-based) knowledge to select the proper factors (and appropriate iev- els) to which the method should be rugged Sta- tlstlcal knowledge 1s necessary to choose a proper design based on the selected factors and the intended usage of the method The experimental results are interpreted If applicable system suit- ability criteria are provided, factors that cause problems are Identified INTEGRATION STUDIES [17,18] The stand-alone expert systems described above all tackle a specific sub-problem of the method development process These systems are implemented m different shells and run on dlffer- ent hardware Ideally, the chromatographer should be able to consult the system that 1s needed in a specific situation as part of a complete method development expert system Also, from the vlewpomt of the knowledge engineers it was seen as a challengmg task to Integrate stand-alone systems of different orlgms Because of the three knowledge engineer centres involved m the pro- ject, It was decided to study three partial mtegra- tlons (see Figs 4-6) There are two important aspects to this First, the analytical experts had to reahze that to pro- duce meaningful integrations It was necessary to fill knowledge gaps between the different stand- alone systems, so that additional knowledge ac- qulsltlon was inevitable This resulted m conad- erable extensions of the exlstmg systems and m the addition of new expert systems, such as LA- BEL and LIT Second, the knowledge engineers had the dlfflcult task of hnkmg sub-systems of different origins mto an acceptable architecture ZNT Z [17] The structure of the architecture of INT I IS given in Fig 4 In this scheme the supervisor 1s the essential part, having the strate- gic knowledge to route the end user to the dlffer- ent expert systems INT I is a typical example for which relatively much addltlonal knowledge was necessary for the integration and Integration was necessary m order to obtain a sultable system As L Buydens et al /Anal Chm Acta 272 (1993) 41-51 Fig 4 Structure of the Integrated system 1 (INT I) an extension, it IS felt that the integration with the SOS system 1s necessary because using SOS strongly influences the selection of the optlmlza- tion criterion ZNTZZ [19] Figure 5 shows the structure of the second integrated system Two of the five subsys- tems (REPS and SOS) are the orlgmal stand-alone systems whereas the other modules were built to add flexlblhty to the system This architecture allows the user to consult the system m three different situations It can be used to assess the repeatability of a new method or to check the repeatability of a previously validated method Also, the posslblhty of usmg the system as a trouble-shooting tool turned out to be a valuable feature ZNT ZZZ Another posslblhty to link stand-alone expert systems 1s shown m Fig 6 In the rugged- ness expert system (RES) the system optimization system (SOS) 1s incorporated as an extra module The scheduler has knowledge on when to activate which module The different modules take care of the different tasks m the ruggedness test and the SOS module has been added to provide solu- tions for problems that have been detected by the diagnosis module SOS can be used m a number of situations Pnmanly, SOS can help to improve 41 a method when the resolution has fallen below a critical level during the ruggedness testing SOS can then propose new condltlons based on the requirement for higher resolution Both systems had to be adapted slightly and/or extended m order to make a sensible integration These studies show that integration often re- sults m complex structures, 1 e , they are less user friendly This endorsed m fact the orlgmal dea- slon to build limited stand-alone systems On the other hand, it was shown that integration 1s use- ful m situations where the chromatographer often has to switch between systems Slmllar conclu- sions have also been reported elsewhere [20] VALIDATION AND EVALUATION OF THE ESCA EX- PERT SYSTEMS Considerable attention was paid to the testing of the expert systems [21,22] It 1s important to note that systems were tested with special empha- SIS on their performance rather than on appear- ance aspects such as a nice user interface The latter should, however, be of sufficient quality to make an understandable system Two mam stages have been dlstmgulshed, the vahdatlon and the evaluation stage The vahdatlon process involved checking the software and testing the knowledge base by the responsible expert The procedure that was fol- lowed involved the selection of a number of test cases by the expert The expert solved the test cases manually, while the expert system was also Requirements Editor subsystem Method Editor subsystem System Optimization subsystem Repeatability Testing subsystem Perform User Action subsystem Fig 5 Structure of the Integrated system 2 (INT II) I Common Datastructure Fig 6 Structure of the integrated system 3 (INT III) consulted The test cases were selected wlthm the scope of the systems and to make use of as much of the knowledge base as possible Whenever differences between the expert systems solution and that of the expert were seen, the cause of the discrepancy was identified This led to the addl- tlon of mlssmg knowledge or to the correction of L Buydens et al /Anal Chrm Acta 272 (1993) 41-51 exlstmg knowledge To decide on the proper per- formance of the systems, a set of pass/fall cnte- na were defined by the knowledge engineer and the expert prior to testing The systems were improved untd agreement was reached between the expert and the expert system After vahdatlon the systems were subjected to the second stage, the external evaluation The evaluation phase consisted of testing the expert system in practical situations, to evaluate the system’s performance m dally practice Gen- erally, these tests were performed by external evaluators, 1 e , chromatographers not involved m budding the system Unbiased problem cases were put to the expert system All inputs and outputs of the systems were registered and, whenever possible and appropriate, verified with expen- ments A list of performance criteria were ldentl- fled by the knowledge engineers and the experts for each system These criteria took mto account aspects of the man/machine interface, the con- sistency of the system and its hmltatlons A hst of criteria IS given m Table 4 The evaluations were carried out by different persons, ranging from experts m method develop- ment to students with little or no experience Summarlzmg the evaluations of the three mte- grated systems the followmg conclusions can be drawn The user friendliness expressed among others m a good user interface, clear screen text and easy help functions was Judged to be good m TABLE 4 Example of evaluation criteria Man-machine interface (user Interface) Choice of phrases Explanation Consistency testing System limits Operation (mouse, keyboard, file Input, etc 1 Usabdlty/ease of use Accuracy (correct answer, quality of advice) Reproducibility (repeatability, same mput same output) Robustness of software (does the system lock up or fall over) Ruggedness (small changes m input small changes m output, similar cases, similar answers) Conflict (two rules with the same input give a different output) Missing rules (input leads to no realistic output) Are essential parts missing? Are there examples of strange answers m extreme cases, e g , incomplete input, nonsense output? Techmcal content do the systems do a useful Job7 L Buydem et al /Anal Chm Acta 272 (1993) 41-51 49 INT III and INT III Some of these aspects are closely related to the quality of the shell KEiS (INT I) belongs to the older generation of shells m which the above features can be improved Case study pH optlmrzatwn With respect to the knowledge, a great variety exists between the systems The knowledge col- lected m INT I IS most complex and generally of heuristic nature Although thrs system was re- stricted to basic pharmaceutrcal compounds, there 1s still a lot of chemical and analytical knowledge to add The knowledge m the other systems 1s better defined and proved to be complete m a broad field of apphcatlons The evaluators found the item “factor choice m the ruggedness module very flexible On the other hand, they asked for more flemblhty m the experunental design In the followmg example an mterestmg apph- cation of expert systems 1s shown m which algo- rithmic-based knowledge 1s combined with heuristic knowledge The complex@ of some steps m the method development process til be demonstrated INT I deals wth method selection and selec- tlvlty optunlzatlon In this system three modules are present for the method selection and subse- quent retention optmuzation After an experi- ment it has to be decided whether the selectivity has to be optmuzed The expert system ade- quately helps to select a method for the selectlv- ity optimization, ~12, sequential or simultaneous approach, and which parameters have preferably 1400 1 a 1200 j 400- 7 1400- 1200- 1000- 600- 600- 400 3 0 “_ 5 10 1s 20 Tlma (mtn 1 Fig 7 Optmuzatlon with SLOPES (a) Chromatogram m one of the expenments, (b) result obtamed after optlmuatlon wth SLOPES 50 L Buydens et al /Anal Chm Acta 272 (1993) 41-51 to be optlmlzed, VU, percentage of modifier, mixture design, temperature, pH The next step 1s to carry out the optnnlzatlon It was chosen to implement only the software tools to carry out pH optlmlzatlon m a simultaneous approach The pH optlmlzatlon was selected because this 1s a relatively new area For other types of optlmlza- tlon one can use commercially available software tools Before the experiments for the pH optlmlza- tlon can be carried out, the parameter space has to be defined, an expernnental design has to be selected and a criterion has to be chosen for the calculation of the optimum In these three mod- ules heuristic knowledge for the selection of pa- rameter space for aads and for bases, chemomet- nc knowledge for the selection of an appropriate design and algorlthmlc knowledge for the calcula- tion of the optimum 1s used In Fig 7, two chromatograms are shown, one before and one after pH optlmlzatlon The opti- mlzatlon was done by means of SLOPES The mam hmltatlon of the system IS to describe accu- rately the retention behavlour of each solute over the parameter space selected (retention surface) It 1s well known that the relationship between retention and pH 1s an S-shaped curve The cal- culation of the retention surface through the measuring points cannot be done by a quadratic function However, to keep the number of meas- urements small it was decided to fit a quadratic function through the data points and to study the pH variation over only a small pH range of 3 units By this means a reasonably accurate pH optlmlzatlon could be achieved Main results of the evaluation Expert systems can provide very powerful as- sistance during method development because of the heuristic knowledge (expertise and expen- ence of a specialist) that 1s implemented m these systems However, during the evaluation phase of all the expert systems it became clear that the attitude towards expert systems 1s strongly de- pendent on the expertise level of the evaluator The accesslblhty of the specialist’s expertise was clearly appreciated by inexperienced users The mtroductlon of the expert systems resulted for those users m a considerable amount of time saving m the development process Experienced users could appreciate the quality of advice given by the systems They are more mterested, how- ever, m comparing the expertise m the systems with then own experience When the strategy implemented m the system did not agree with then own expertise, it resulted m dlssatlsfactlon wrth the expert system, because their own expen- ence, probably better adapted to their specific atuatlon, IS not considered by the system This 1s especially the case when the knowledge domam of the expert system 1s strongly susceptible to mdlvldual opmlons This gives rrse to a second aspect, that at present expert systems are not flexible enough (Minor) changes that could result m a better performing expert system for a partlc- ular situation may be lmposslble to make by the user Only the knowledge engineer who 1s aware of the structure of the knowledge base 1s able to make such changes without unexpected conse- quences A third conclusion of the evaluation 1s that the mam attention should be devoted to the mtegra- tlon of expert systems with laboratory mstru- ments When this 1s reahzed, expert systems can be used to obtain rapidly accurate advice while the user remains free to choose his or her own approach Concluszons The ESCA project can be considered as a pioneer project for the apphcatlon of expert sys- tem technology m analytical chemistry Method development m LC was selected as the apphca- tlon expertise area The expertise that has been considered m the project covers the important areas of method development m LC It was only possible to cover this large area because many recogmzed experts participated m the project It would be almost lmposslble to fmd a single expert to cover all the different aspects of method development Different expert systems resulted from the pr@ Ject Most of these are still m the research phase, but a few have been further developed for com- mercialization (system optimization system, ruggedness system) From the results of vahda- L Buydens et al /Anal Chrm Acta 272 (1993) 41-51 tlon and evaluation it can be concluded that expert systems are potentially very useful for method development m LC The benefits of the systems are that method development can be done more consistently and more efficiently and that better optimized and validated methods are produced Even when the systems were not yet complete this conclusion became clear Expert systems still have to fmd their way mto the chromatographlc laboratory Users will have to accept computer programs that assist during tasks such as method development This requires, as stated above, expert systems that are flexible and easy to mtegrate It should be possible to add new knowledge or to adapt the system according to changes m the apphcatlon environment Re- search on this subject remains necessary This research proJect was partly funded by the EEC, as ESPRIT project P1570 ESCA T Blaf- fert, A Cleland, T Hamolr, G Kateman, J A van Leeuwen, D L M Massart, M Mulholland, H PlnJns, B G M Vandengmste, N Walker and all other temporary participants are acknowl- edged for their contributions which made this final report possible REFERENCES B A Hohne and T H Pierce (Eds 1, Expert System Apph- catlons m Chemistry (ACS Symposmm Series, No 4081, Amencan Chemical Society, Washmghton, DC, 1989 J W Dolan, L R Snyder and MA Quarry, Chro- matographla, 24 (1987) 261 S S Wdhams, Trends Anal Chem ,9 (1990) 63 D Goulder, T Blaffert, A Blokland, L Buydens, A Chhabra, A Cleland, N Dunand, H Hmdnks, G Kate- man, H van Leeuwen, D Massart, M Mulholland, G 10 11 12 13 14 15 51 Musch, P Nalsh, A Peeters, G Postma, P Schoenmakers, M de Smet, B Vandegmste and J Vmk, Chro- matographla, 26 (1988) 237 M De Smet, A Peeters, L Buydens and D L Massart, J Chromatogr ,457 (1988) 25 J A van Leeeuwen, B G M Vandegmste, G J Postma and G Kateman, Chemom Intell Lab Syst , 6 (1989) 239 P Pohtakls and S M Weiss, ArtIf Intel], 22 (1984) 23 R Wehrens, L Buydens and G Kateman, Chemom In- tell Lab Syst , 12 (1991) 57 H Hmdnks, F Mans, J Vmk, A Peeters, M De Smet, D L Massart and L Buydens, J Chromatogr ,485 (1989) 255 P J Schoenmakers, The Optlmlzatlon of Chromatographlc Selectlvlty, a Guide to Method Development, Elsevler, Amsterdam, 1986 A Peeters, L Buydens, D L Massart and P J Schoen- makers, Chromatographla, 26 (1988) 101 P J Schoenmakers, N Dunand, A Cleland, G Musch and T Blaffert, Chromatographla, 26 (1988) 37 M Mulholland, J A van Leeuwen and B G M Vandegm- ste, Anal Chum Acta, 223 (1989) 183 M Mullholland, N Dunand, A Cleland, J van Leeuwen and B Vandegmste, J Chromatogr , 485 (1989) 283 J A van Leeuwen, L M C Buydens, B G M Vandegmste, G Kateman, P J Schoenmakers and M Mulholland, Chemom Intell Lab Syst , 10 (1991) 337 16 J A van Ixeuwen, L M C Buydens, B G M Vandegmste, G Kateman, P J Schoenmakers and M Mulholland, Chemom Intell Lab Syst , 11 (1991) 37 17 18 19 20 21 22 P Contl, H Puyns, N Vandendnesche, M Desmet, T Hamolr, F Mans, H Hmdnks, P Schoenmakers and D L Massart, Chemom Intel1 Lab Syst , 11 (1991) 27 L Buydens, J Van Leeuwen M Mulholland, B Vandegm- ste and G Kateman, Trends Anal Chem , 9 (1990) 58 M Mulholland, N Walker, F Mans, H Hmdrlks, L Buydens and T Blaffert, J Chromatogr , 550 (1991) 257 F A Settle and M A Pleva, Chemom Intell Lab Syst , 11 (1991) 13 J A van Leeuwen, L Buydens, B G M Vandegmste, G Kateman, A Cleland, M Mulholland, C Jansen, F A Mans, PH Hoogkamer and J H M van den Berg, Chemom Intel1 Lab Syst , 11 (1991) 161 F Mans, H Hmdrlks, J Vmk, A Peeters, N Vandendrle- sche and D L Massart, J Chromatogr ,506 (1990) 211 
work_jl47oly4sjbc5igyllnr5wmgli	----	liu-relative-2016.pdf Relative Preference-based Recommender Systems By Shaowu Liu BIT (Hons) Submitted in fulfilment of the requirements for the degree of Doctor of Philosophy Deakin University June, 2016 sfol Retracted Stamp sfol Retracted Stamp To my unborn baby... iv Table of Contents Table of Contents v List of Tables viii List of Figures ix Acknowledgements x Publication List xi Abstract xiii 1 Introduction 1 1.1 Background and Motivation . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.3 Overview of the Proposed Methodology . . . . . . . . . . . . . . . . . 5 1.4 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2 Literature Review 9 2.1 Notation and Problem Formulation . . . . . . . . . . . . . . . . . . . 9 2.2 Recommendation Techniques . . . . . . . . . . . . . . . . . . . . . . . 12 2.2.1 Content-based Recommender Systems . . . . . . . . . . . . . 12 2.2.2 Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . 14 2.2.3 Context-Aware Recommendation . . . . . . . . . . . . . . . . 16 2.2.4 Graph-based Recommendation . . . . . . . . . . . . . . . . . . 17 2.2.5 Trust-based Recommendation . . . . . . . . . . . . . . . . . . 18 2.3 Relative Preference-based Recommender Systems . . . . . . . . . . . 18 2.3.1 Notation and Problem Statement . . . . . . . . . . . . . . . . 20 2.3.2 Memory-based Models . . . . . . . . . . . . . . . . . . . . . . 24 v 2.3.3 Model-based Models . . . . . . . . . . . . . . . . . . . . . . . 25 2.4 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.4.1 Accuracy Metrics . . . . . . . . . . . . . . . . . . . . . . . . . 29 2.4.2 Diversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.4.3 Coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.4.4 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.5 Metrics for Relative Preference-based Models . . . . . . . . . . 31 2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 3 Ordinal Random Fields for Recommender Systems 34 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 3.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.2.1 Matrix Factorization . . . . . . . . . . . . . . . . . . . . . . . 39 3.2.2 Ordinal Matrix Factorization . . . . . . . . . . . . . . . . . . 40 3.2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 3.3 Ordinal Random Fields . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.3.1 Markov Random Fields . . . . . . . . . . . . . . . . . . . . . . 43 3.3.2 ORF: Unifying MRF and OMF . . . . . . . . . . . . . . . . . 45 3.3.3 Feature Design . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.3.4 Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . 47 3.3.5 Preference Prediction . . . . . . . . . . . . . . . . . . . . . . . 48 3.4 Experiment and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 49 3.4.1 Experimental Settings . . . . . . . . . . . . . . . . . . . . . . 49 3.4.2 Result and Analysis . . . . . . . . . . . . . . . . . . . . . . . . 52 3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 4 Preferecen Relation-based Recommender System 58 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 4.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 4.2.1 Preference Relation . . . . . . . . . . . . . . . . . . . . . . . . 62 4.2.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . 63 4.2.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 4.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.3.1 Preference Relation-based Matrix Factorization . . . . . . . . 67 4.3.2 Markov Random Fields . . . . . . . . . . . . . . . . . . . . . . 70 4.3.3 Conditional Random Fields . . . . . . . . . . . . . . . . . . . 73 4.3.4 PrefCRF: Unifying PrefNMF and CRF . . . . . . . . . . . . . 76 4.4 Experiment and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 82 4.4.1 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . 85 vi 4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 5 Learning from Heterogeneous Data Sources for Improved Top-N Recommendation 97 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 5.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 5.2.1 Heterogeneous Sources . . . . . . . . . . . . . . . . . . . . . . 99 5.2.2 Preference Relation . . . . . . . . . . . . . . . . . . . . . . . . 100 5.2.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 5.3 Preference Relation-based Conditional Random Fields . . . . . . . . . 104 5.3.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . 104 5.3.2 Preference Relation-based Matrix Factorization . . . . . . . . 105 5.3.3 Conditional Random Fields . . . . . . . . . . . . . . . . . . . 106 5.3.4 Ordinal Logistic Regression . . . . . . . . . . . . . . . . . . . 109 5.3.5 PrefCRF: Unifying PrefNMF and CRF . . . . . . . . . . . . . 110 5.4 Experiment and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 115 5.4.1 Experimental Settings . . . . . . . . . . . . . . . . . . . . . . 115 5.4.2 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . 118 5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 6 Conclusion and Future Work 127 6.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 Bibliography 130 vii List of Tables 2.1 Notations used in rating-based RecSys . . . . . . . . . . . . . . . . . 11 2.2 Capabilities of existing methods . . . . . . . . . . . . . . . . . . . . . 19 2.3 Notations used in Relative Preference-based RecSys . . . . . . . . . . 23 3.1 Summary of Major Notations (Chapter 3) . . . . . . . . . . . . . . . 38 3.2 Performance of ORF model . . . . . . . . . . . . . . . . . . . . . . . 53 3.3 Paired t-test for the ORF and the OMF. . . . . . . . . . . . . . . . . 53 4.1 Summary of Major Notations (Chapter 4) . . . . . . . . . . . . . . . 65 4.2 Capabilities of PR-based methods . . . . . . . . . . . . . . . . . . . . 67 4.3 Results on MovieLens-1M dataset . . . . . . . . . . . . . . . . . . . . 86 4.4 Results on EachMovie dataset . . . . . . . . . . . . . . . . . . . . . . 87 4.5 Paired t-test for PrefMRF and PrefNMF. . . . . . . . . . . . . . . . . 93 5.1 Performance of PrefCRF on MovieLens-1M dataset . . . . . . . . . . 120 5.2 Results over ten runs on Amazon dataset. . . . . . . . . . . . . . . . 121 5.3 Results over ten runs on EachMovie dataset. . . . . . . . . . . . . . . 122 5.4 NDCG@10 on MovieLens-20M dataset. . . . . . . . . . . . . . . . . . 123 viii List of Figures 2.1 An overview of Recommender System . . . . . . . . . . . . . . . . . . 13 3.1 Example of undirected graphs . . . . . . . . . . . . . . . . . . . . . . 44 3.2 Impact of Minimum Correlation Threshold . . . . . . . . . . . . . . . 54 3.3 Impact of Minimum Correlation Threshold . . . . . . . . . . . . . . . 55 3.4 Impact of Regularization Coefficient . . . . . . . . . . . . . . . . . . . 56 4.1 Example of MRF graphs . . . . . . . . . . . . . . . . . . . . . . . . . 72 4.2 Example of CRF graphs . . . . . . . . . . . . . . . . . . . . . . . . . 74 4.3 Performance of different position T on MovieLens-1M dataset (Sparse). 89 4.4 Performance of different position T on MovieLens-1M dataset (Dense). 90 4.5 Performance of different position T on EachMovie dataset (Sparse). . 91 4.6 Performance of different position T on EachMovie dataset (Dense). . 92 4.7 Impact of Sparsity Levels. . . . . . . . . . . . . . . . . . . . . . . . . 94 4.8 Impact of Parameters (MovieLens-1M) . . . . . . . . . . . . . . . . . 95 5.1 Flow from user preferences to PR . . . . . . . . . . . . . . . . . . . . 101 5.2 Undirected graphs for users u and v . . . . . . . . . . . . . . . . . . . 108 5.3 Varying sparsity on MovieLens-1M dataset. . . . . . . . . . . . . . . 121 5.4 Varying position T on EachMovie dataset. . . . . . . . . . . . . . . . 123 5.5 Varying biases on MovieLens-20M dataset. . . . . . . . . . . . . . . . 124 5.6 Varying parameters on MovieLens-1M. . . . . . . . . . . . . . . . . . 125 ix Acknowledgements I would like to thank Dr. Gang Li, my supervisor, for his many suggestions and constant support during this research. I am also thankful to my associate supervisors Prof. Gleb Beliakov and A/Prof. Ping Xiong for useful suggestions and friendly encouragement. I would like to thank my research mates and friends: Dr. Jia Rong, Dr. Huy Quan Vu, Dr. Truyen Tran, Dr. Veelasha Moonsamy, Dr. Yongli Ren, Dr Tianqing Zhu, and so many more. I am also thankful to the support staff, in particular Ms. Lauren Browne, Ms. Judy Chow for their administrative help. Of course, I am grateful to the patience and love from my parents, who gave me birth at the first place and raised me up. I am also thankful for my parents in law for their support throughout my research studies. Last but not the least, I would like express my sincere gratitude to my beautiful wife, Manchun Li, for her patience and love. Without her everlasting encouragement and understanding, this work would never have come into existence. Melbourne, Australia Shaowu Liu June 2016 x Publication List Refereed Journal Articles 1. Shaowu Liu, Gang Li, Truyen Tran, and Yuan Jiang. Preference Relation- based Markov Random Fields for Recommender Systems. Machine Learning. (ERA 2010 ranking A*) 2. Veelasha Moonsamy, Jia Rong, Shaowu Liu, Mining Permission Patterns for Contrasting Clean and Malicious Android Applications. Future Generation Computer Systems, 2014, 36: 122–132. (ERA 2010 ranking A) 3. Shaowu Liu, Rob Law, Jia Rong, Gang Li, and John Hall. Analyzing Changes in Hotel Customers’ Expectations by Trip Mode. International Journal of Hos- pitality Management, 2013, 34: 359–371. (ERA 2010 ranking A) 4. Gleb Beliakov, Gang Li, Shaowu Liu, Parallel bucket sorting on Graphics Processing Units based on convex optimization. Optimization, 2015, 64(4): 1033-1055. (Impact Factor: 0.936) 5. Huy Quan Vu, Gang Li, Nadezda S. Sukhorukova, Gleb Beliakov, Shaowu Liu, Carole Philippe, Helene Amiel, Adrien Ugon. K-complex Detection using a Hybrid-Synergic Machine Learning method. IEEE Transactions on Sys- tems, Man, and Cybernetics Part C: Application and Reviews (IEEE TSMCC), 2012, 42(6): 1478–1490. (Impact Factor: 2.171) xi xii 6. Huy Quan Vu, Shaowu Liu, Xinghua Yang, Zhi Li, Yongli Ren. Identify- ing Microphone from Noisy Recordings by Using Representative Instance One Class-Classification Approach. Journal of Networks (JNW), 2012, 7(6): 908- 917. (ERA 2010 ranking A) Refereed Conference Papers 1. Shaowu Liu, Gang Li, Truyen Tran, Yuan Jiang. Preference Relation-based Markov Random Fields. In Proceedings of the 7th Asian Conference on Machine Learning (ACML 2015), pp. 157-172, Journal of Machine Learning Research: workshop and conference proceedings, 2015. 2. Shaowu Liu, Truyen Tran, Gang Li, Yuan Jiang. Ordinal Random Fields for Recommender Systems. In Proceedings of the 6th Asian Conference on Machine Learning (ACML 2014), pp. 283-298, Journal of Machine Learning Research: workshop and conference proceedings, 2014. 3. Veelasha Moonsamy, Jia Rong, Shaowu Liu, Gang Li, Lynn Batten. Contrast- ing Permission Patterns between Clean and Malicious Android Applications. In Proceedings of the 9th International Conference on Security and Privacy in Communication Networks (SecureComm 2013). 4. Huy Quan Vu, Shaowu Liu, Zhi Li, Gang Li. Microphone Identification using One Class-Classification Approach. In Proceedings of the 2nd Workshop on Applications and Techniques in Information Security (ATIS 2011). Book Chapter 1. Gleb Beliakov and Shaowu Liu. Parallel Monotone Spline Interpolation and Approximation on GPUs. Designing Scientific Applications on GPUs, CRC Press, Taylor and Francis Group, 2013, 295–310. Abstract Recommender systems aim to recommend users with some of their potentially inter- esting items by exploiting various information, especially the absolute ratings. Nev- ertheless, recent literature has suggested that rating-based systems are less reliable comparing to those based on relative preferences, i.e., “which one is better?” instead of “what do you think of this one?” However, a problem of these emerging relative preference-based models is that they consider either the second order interactions, such as similarities between users, or the higher order interactions, such as latent factors. This limitation reduces the performance of relative preference-based systems as the two types of interactions are complementary. On the other hand, due to the change of input format, existing relative preference-based systems do not consider side information such as user profiles and item content, which can be helpful to further improve the performance. Furthermore, the potential of relative preference-based systems to merge heterogeneous data sets was not identified in literature, which can help alleviate the cold-start problem of having limited information for new users or items. In this thesis, we tackle these three issues. We propose a novel model to exploit the ordinal properties possessed by ratings, where both the second and higher order interactions are considered. In this model, ratings are no longer considered as num- bers, but a sequence of ordinal labels. The proposed model used Markov Random Fields to combine two types of interactions. Another type of relative preference is Preference Relation (PR), i.e., compar- isons of items. For PR-based systems, we proposed a modified version of Markov xiii xiv Random Fields which accepts PR instead of ordinal preferences, by converting PR into user-wise preferences, and then into ordinal distributions through ordinal logis- tic regression. This process produces the first PR-based recommender system that captures both types of interactions. For incorporating side information, we extended the Markov Random Fields to Conditional Random Fields, in which the users profiles and item content are considered by designing new features. Despite of improving existing PR-based systems, we also identified a great po- tential of such systems to merge heterogeneous data sets. Specifically, data sets in different format, such as 5-star ratings, binary ratings, page views, and mouse clicks can all be converted into PR format and used by PR-based systems. This observation makes it possible to alleviate the cold-start problem by generating a much denser data set, which could not be done for rating-based systems. To evaluate the performance of proposed models, we conducted experiments on different public data sets against the state-of-the-art relative preference-based models measured by different metrics. The results presented in the experiment sections of each chapter show statistically significant improvement over existing models. The main contributions of this research are proposing the first relative preference-based models that can capture both types of interactions, and using PR-based models to alleviate the cold-start problem. Keywords: Recommender Systems, Preference Relation, Collaborative Filtering, Relative Preference, Ordinal Preference Chapter 1 Introduction 1.1 Background and Motivation Recommender Systems (RecSys) aim to suggest items (books, movies, tourism attrac- tions, etc.) that are potentially to be liked by the user. To identify the appropriate items, RecSys use various sources of information, such as historical ratings given by users [38] or content of items [5]. RecSys were originally designed for users with in- sufficient personal experience or with limited knowledge on the items. However, with the rapid expansion of Web 2.0 and e-commerce, overwhelming number of items are offered, and now every user can be benefited from RecSys [66]. Over the last decade there have been rapid advances in RecSys, from both academia and industry [7,20,37,44,49,76]. One of the most important events in RecSys was the one million Netflix Prize [7] launched in 2006, which sought for RecSys that out- perform Netflix company’s own RecSys. The dataset released in this competition contains historical ratings on movies given by individuals. The Netflix dataset, to- gether with other datasets released by companies such as Amazon and Yahoo! have become the popular benchmark datasets in this field. Due to the extensive use of 1 2 these datasets, which contain ratings, most RecSys to date are designed to exploit ratings [40,41,69]. However, user feedbacks are not always expressed in form of absolute ratings, and it is often expensive to collect such explicit feedbacks. Furthermore, studies [12,42] have reported that absolute ratings may not be completely trustworthy. For example, the rating 4 out of 5 may in general indicate high quality, but it can mean just OK for critics. In fact, users’ quantitative judgment can be affected by a number of irrelevant factors such as the mood when rating, and in psychology this is called misattribution of memory [71]. While users are not good at making consistent quantitative judgment, the relative preferences such as ordinal preferences [42,46,79,82] and preference relations (PR) [12, 18,19] have been considered as more consistent form of feedbacks across like-minded users. For example, by measuring the relative order between items, the PR is usually less variant to irrelevant factors: a user in bad mood may give lower ratings to all items but the relative orderings between items remain the same. Being a more reliable type of user preferences, PR is also easier to collect comparing to ratings as it can be inferred from implicit feedbacks. For example, the PR between two items can be inferred by comparing their ratings, page views, played counts, mouse clicks, etc. This property is important as not all users are willing to rate their preferences, where collecting feedbacks implicitly delivers a more user-friendly recommender system. In addition, as the ultimate goal of RecSys, obtaining the ranking of items by itself is to obtain the relative preferences, a more natural input than absolute ratings [42,81]. Despite of its potential, the newly emerged relative preference-based RecSys pro- vides less features comparing to the well-established rating-based RecSys. Meanwhile, 3 relative preference-based RecSys provides an alternative view of user preferences, thus can be used to resolve issues of rating-based RecSys. Currently, relative preference- based RecSys still faces the following unresolved issues: • Different Structures in User Preferences: Existing recommendation techniques can be largely divided into two forms: memory-based [65,69] and model-based [38]. Memory-based approaches focus on capturing the second-order interactions be- tween similar users [65] or items [69]. This type of information is called Local Structure (LS) of user preferences. On the other hand, model-based approaches focus on discovering the weaker but higher-order interactions among all users and items. This type of information is called Global Structure (GS) of user preferences. Previous studies have suggested that these two types of struc- ture are complementary since they address different aspects of the preferences [38,46,79]. However, there is yet no relative preference-based RecSys that can capture both LS and GS. • Side Information of User Preferences: While existing recommendation tech- niques focus on exploiting user preferences, side information such as item con- tent and user attributes [79] are also shown to be useful in improving recom- mendation quality. However, due to the change of input format, there is yet no relative preference-based RecSys can incorporate side information. • User Preferences from Heterogeneous Sources: Last decade has seen a growing trend towards creating and managing more profiles in Online Social Networks. User are now providing feedbacks on different platforms in different formats, such as 5-star ratings, thumbs up/down, as well as implicitly as mouse clicks. 4 These rich, but heterogeneous, user preferences provide an opportunity of allevi- ate the cold-start problem [72]. However, existing recommendation techniques usually assume the user preferences are in the same format, and therefore are unable to exploit these heterogeneous user preferences. The first two issues have constrained the potentials of relative preference-based RecSys, while the third issue is faced by all existing recommendation techniques. This thesis aims to address these issues to make relative preference-based ReSys more effective and applicable. 1.2 Research Objectives The objective of this work is to overcome the aforementioned weaknesses of existing relative preference-based RecSys, as well as resolving the heterogeneous data sources issue of traditional RecSys. More specifically, the research objectives of this work are: • Learning Local and Global Structures: Capturing both the local and global structures of user preferences have been done in rating-based recommendation techniques [38]. However, existing approaches are not directly applicable to relative preference-based RecSys as the format of input has changed. Recent advances in Markov Random Fields-based RecSys [79] have made it possible to capture both structures in a principled way by utilizing the flexibility of graphical models. This thesis will investigate how the two structures can be compiled into a single model in a probabilistic manner. • Incorporating Side Information: How to incorporate side information such as 5 item content and user attributes is a problem for relative preference-based Rec- Sys. In fact, no existing relative preference-based RecSys has attempted this task as these models are designed particularly for user preferences and have no flexibility to incorporate side information in a proper way. On the other hand, Conditional Random Fields [79], as an the extended version of Markov Ran- dom Fields, can easily incorporate side information in a probabilistic manner. However, it remains unknown how Conditional Random Fields can accept rel- ative preferences as input. This thesis will investigate how to design a relative preference-based Conditional Random Fields model. • Learning from Heterogeneous Data Sources: How to unify user preferences in different formats has been a problem for traditional rating-based RecSys. The main difficulty is that there is no suitable method to convert user preferences among formats without introducing noises. Furthermore, some conversions are impractical, such as converting mouse clicks to 5-star ratings. Fortunately, the relative preference provides a unified interface for all kinds of user preferences, where both mouse clicks and 5-star ratings can be converted into pairwise preference relations. In this thesis, we will investigate how to learn from het- erogeneous data sources using relative preference-based RecSys. 1.3 Overview of the Proposed Methodology Firstly, to address the problem of learning both local and global structures, we propose two Markov Random Fields-based models to capture and unify both the LS and GS information. Specifically, the proposed model employs Markov Random Fields 6 (MRF) to investigate the LS information while the Ordinal Matrix Factorization (OMF) captures the GS information. In this way, we take advantages of both the representational power of the MRF and the ease of modeling ordinal preferences by the OMF. Experimental result on public datasets demonstrates that the proposed model can capture both types of interactions, resulting in improved recommendation accuracy. On the other hand, when the input format is pairwise preference relation, the Preference Relation-based Markov Random Fields model is proposed to deal with the input format of pairwise comparisons of items. Secondly, to address the problem of incorporating side information, we extended the proposed Markov Random Fields-based models to Conditional Random Fields- based models, in which the side information are modeled as global observations of the graphical models. We performed experiments on public datasets and demonstrate that side information has been properly incorporated, and significantly improved recommendation performance has been achieved and validated by statistical tests. Finally, to address the problem unifying information from heterogeneous data sources, we employed the several models to convert and exploit user preferences of different formats. Specifically, all types of user preferences are converted into the unified preference relation format and modeled by our proposed models. Experiment results on public datasets demonstrate that our solutions to unifying data from het- erogeneous sources have successfully minimized the noises information introduced, resulting improved recommendation quality, especially in cold-start cases where each data source provides a limited amount of data. 7 1.4 Thesis Outline This section presents the overall organization of this thesis. As the objective of this thesis is to address the problems of relative preference-based RecSys, the content of each chapter is organized as follows: • Chapter 2 presents a comprehensive survey on recommender systems in gen- eral with a focus on relative preference-based RecSys. Specifically, the relevant concepts, assumptions, and emerging research issues in this area will be dis- cussed. Efforts have been made to identify current and future issues of relative preference-based RecSys. • Chapter 3 focuses on resolving the issue of learning from both the local and global structures in ordinal user preferences. This chapter specifically inves- tigates the scenario of using ordinal type of preference as input. An Ordinal Random Fields (ORF) method is proposed to capture and unify both types of structures in a principled way. Experiments on multiple public datasets are con- ducted to show that the proposed method effectively improves the performance of recommendation by utilizing both types of structures. • Chapter 4 proposes a novel Preference Relation-based Markov Random Fields model to address the issue of learning from both the local and global structures in preference relations. This chapter also proposes a Preference Relation-based Conditional Random Fields model, which incorporates side information of users and items. The proposed model does not rely on ratings but pairwise compar- isons of items, thus offers better reliability and can be applied to a wider range of applications. To validate its performance, we conducted experiments on several 8 public datasets together with side information, and performance improvements have been confirmed statistically. • Chapter 5 addresses the issue of learning from multiple heterogeneous data sources. This chapter identifies and formalizes the heterogeneous data sources problem, and proposes the preference relation-based method to unify heteroge- neous data. With consideration of multiple data sources, the proposed method can reduce the effect of cold-start problem where each data source provides limited amount of data. With the help of the proposed method, implicit user preferences such as page views and mouse clicks can be easily exploited to alle- viate the cold-start problem. • Chapter 6 summarizes the contributions of this thesis, as well as discusses some possible extensions and directions of future research. To maintain readability, some essential concepts, definitions, and motivations are recounted in each chapter to make it self-contained. For basic concepts of recom- mender systems, readers may refer to the recommender system handbook [66]. Chapter 2 Literature Review This chapter is devoted to provides an extensive literature review on recommender systems by racing the trends and directions of current research. We chronologically review contributions along each research direction regarding recommender systems with a focus on relative preference-based RecSys. Specifically, Section 2.1 introduces the basic notations and related concepts of recommender systems. Section 2.2 reviews popular recommendation techniques along with the latest developments. Section 2.3 focuses on relative preference-based RecSys, and will introduce the recent develop- ments on this emerging topic. Section 2.4 presents evaluation metrics that are used to evaluate recommendation performance. 2.1 Notation and Problem Formulation RecSys use historical data to predict future interest in items by users. Two objects are involved in RecSys: items and users. Let U = {u1,u2, ...,um} denote a set of m users, and T = {t1, t2, ..., tn} denote a set of n items, such as books, movies, etc. The interest of user u ∈ U in item t ∈ T is encoded as the preference ru,t ∈ R, where 9 10 R captures the known preferences for all users U. A typical form of preferences is the ratings (e.g. 1 − 5 stars), though many other forms exist, such as like/dislike, clicked/not clicked, etc. Definition 1 (Recommender System). Given item collection T and known preferences R of all users U, RecSys aims to identify the item t̂ ∈ T that maximizes the preference rua,t of the active user ua ∈ U [1]: t̂ = arg max t∈T (rua,t) (2.1.1) This definition often implies that individual items are suggested to the individual users, however, real-world applications may require suggesting a set of items and/or to a group of users. To handle such cases, Definition 1 needs to be extended, how- ever, it remains a challenging task to making recommendations to groups of users or recommending a set of items. For ease of reference, notations used by rating-based RecSys are summarized in Table 2.1. Over the last decade, the development on RecSys has been carried out along two research lines: Recommendation Techniques and Evaluation Metrics. Works on the recommendation techniques focus on how to generate recommendations based on various information sources, ranging from the item content, the known preferences, to more recent sources such as the context [2] and the social trust [28]. After the recommendations have been generated, the next task is to evaluate the quality of recommendations using evaluation metrics. Evaluating common machine learning tasks such as classification are in general less difficult as the ground truth is available to assess the predictions. Accuracy metrics such as mean absolute error (MAE) are often employed to assess the performance of machine learning tasks as well as the RecSys. However, it becomes tricky in RecSys where the ground truth 11 Table 2.1: Notations used in rating-based RecSys Notations Mathematical Meanings U set of all users U = {u1,u2, ...,un} T set of all items T = {t1, t2, ..., tm} G a user group, G ⊆ U K an item package, K ⊆ T R available preferences data of all users ru,t the preference of user u on item t R(ux) the set of items rated by user ux S(tx) the set of users rated item tx Txy the set of items co-rated by user ux and uy rG,K the group G’s preference on package K Ir(G,K) the inter-relevance between group G and package K Ih(K) the aggregation of inherence properties of package K |·| the cardinality of the set ·̂ the prediction, e.g. t̂ is the predicted item t ·̄ the arithmetic mean 12 (user’s satisfaction) may not be well represented by the preferences data such as ratings. For example, a user rated 5-star for the movie Titanic but he/she may not want to watch Titanic Extended Version, but a RecSys focuses on ratings may still consider Titanic Extended Version as a 5-star recommendation. For this reason, research of RecSys has gone beyond the accuracy metrics to many novel metrics such as diversity [11], novelty [25], etc. Figure 2.1 provides an overview of recommendation techniques and evaluation metrics in RecSys. 2.2 Recommendation Techniques Recommendation techniques aims to identify the right item for the user, where two fundamental approaches are Content-based methods [62] and Collaborative Filtering methods [65]. Conventionally, Content-based methods generate recommendations by exploiting regularities in the item content, while Collaborative Filtering methods generate recommendations based on available preferences data of users. More recent approaches are exploiting extra information such as the context [2] and the social trust [28]. In this section, we briefly review these recommendation techniques. 2.2.1 Content-based Recommender Systems Content-based methods generate recommendations for the active user ua based on the contents of related items, where other users’ information is not utilized. The basic idea is to identify the unrated items that are similar to the active user’s highly rated items. 13 R ec om m en de r S ys te m R ec om m en da tio n T ec hn iq ue s E va lu at io n M et ric s C on te nt -b as ed C ol la bo ra tiv e F ilt er in g C on te xt -A w ar e G ra ph -b as ed T ru st -b as ed M em or y- ba se d M od el -b as e S up er vi se d le ar ni ng U ns up er vi se d le ar ni ng M at rix F ac ot riz at io n M as sa e t a l., 2 00 4 G uy e t a l., 2 00 9 et c.S al to n, 1 98 9 et c. N ak am ur a et a l., 1 99 8 R es ni ck e t a l., 1 99 4 et c. M iy ah ar a et a l., 2 00 0 et c. H of m an n et a l., 1 99 9 S ar w ar e t a l., 2 00 2 et c. K or en , 2 00 8 et c. A do m av ic iu s et a l., 2 01 1 K or en , 2 00 9 et c. Z ho u et a l., 2 00 7 et c. A cc ur ac y M et ric s S ar w ar e t a l., 2 00 1 Z ho u et a l., 2 00 7 et c. D iv er si ty G e et a l., 2 01 0 Z ho u et a l., 2 00 8 et c. N ov el ty Lü e t a l., 2 01 2 et c. C ov er ag e G e et a l., 2 01 0 et c. S ta bi lit y A do m av ic iu s et a l., 20 12 et c. Le ge nd S ub je ct M et ho d R ep re se nt at iv e W or ks B ey on d A cc ur ac y F ig u re 2 .1 : A n o v er v ie w o f R ec o m m en d er S y st em 14 The prediction of active user’s preference rua,t on unrated item t is calculated based on known preferences of items similar to t. The similarity between two items is measured by comparing the content of the items. For example, two movies can be compared in terms of the actors, directors, genres, etc [47]. For text-based items, the features can be represented by keywords using term frequency/inverse document frequency (TF-IDF) [67]. Given the features, the similarity can be calculated using standard metrics such as the cosine distance. Despite of the simplicity, content-based methods have three limitations. Firstly, it can be difficult to define features or extract content from some types of items, such as audio, videos, and pictures. Secondly, the user will always be recommended with items that are highly similar to the items he/she liked, which leads to the lacking of diversity [11]. Finally, it is difficult to identify items for new users or users with few ratings, and this is referred to as the cold-start problem [72]. 2.2.2 Collaborative Filtering Collaborative Filtering (CF) looks for items highly rated by users similar to the active user. CF methods can be classified into two classes: memory-based methods and model-based methods. Memory-based Methods In memory-based methods [59,65], the preference pre- dictions are based on the entire collection of known preferences. The idea is that similar users should rate the same movie similarly. The preference rua,t of unrated item t for active user ua is calculated based on the preference ruj,t from every user uj ∈ U who is similar to the active user ua. The similarity between two users is defined by comparing their known preferences, and two 15 popular measures are Pearson Correlation Coefficient (PCC) [65] and Vector Space Similarity (VS) [1]: Pearson Correlation Coefficient (PCC) sim(ux,uy) = ∑ ti∈Txy(rxi − r̄x)(ryi − r̄y)√∑ ti∈Txy(rxi − r̄x) 2 ∑ ti∈Txy(ryi − r̄y) 2 (2.2.1) Vector Space Similarity (VS) sim(ux,uy) = cos(ux,uy) = ∑ ti∈Txy rxiryi√∑ ti∈Txy r 2 xi ∑ ti∈Txy r 2 yi (2.2.2) where Txy = {ti ∈ T|rxi �= Ø,ryi �= Ø} denotes the set of items co-rated by both ux and uy. Model-based Methods In contrast to memory-based methods, model-based meth- ods [57,68] will construct a model from the known preferences, and make future recommendations based on the model. Model-based methods often take more training time than memory-based methods, however, they are more efficient in generating recommendations. According to the type of model, model-based methods can be further divided into three classes: supervised learning-based [57], unsupervised learning-based [31,70], and matrix factorization-based [9,38]. Similar to content-based methods, CF also suffers from cold-start problem. In addition, new items rated by a small number of users will have a low chance to be recommended. However, CF has been one of the most popular recommendation techniques for to its efficiency and high quality recommendations. 16 2.2.3 Context-Aware Recommendation Both content-based methods and CF focus on the preferences, however, users’ inter- ests could also be affected by the context. For example, whether a user would like a movie not only depends on the user’s taste, but also the context such as when, where, and with whom. One of the first considered context is temporal information. In 2001, Zimdars et al. [86] treated CF as a uni-variate time series problem, where a user’s next preference is predicted based on the previous preference. However, temporal information did not attract much attention until the successful of timeSVD++ method [39] in Netflix Progress Prize competition. The timeSVD++ method predicts preference of active user ua on item i at time t as: r̂ai(t) = μ + bi(t) + ba(t) + q T i ⎛ ⎝pa(t) + |R(ua)|12 ∑ j∈R(ua) yj ⎞ ⎠ , (2.2.3) where μ denotes the overall average preference, bi(t) is the item’s bias at time t, ba(t) is the user’s bias at time t, R(ua) is the set of items rated by user ua, qi and yj are item-factor vectors, and pa(t) is the user-factor vector at time t, pa(t), qi, and yj are in a joint latent factor space as used in matrix factorization techniques [38]. In this formulation, temporal information is modeled by the time-based bias, and this makes it superior to other competitors. Recently, it has been recognized that temporal is not the only important context and various kinds of context can be exploited to improve the recommendation quality, and this kind of RecSys is referred to as Context-Aware Recommender Systems [2]. 17 2.2.4 Graph-based Recommendation Graph-based methods consider the recommendation task as a link prediction problem of bipartite graph [85]. On bitpartite graph, users U and items T are represented by two sets of nodes, and each pair of < user,item > can be connected with an edge. These edges represent users’ interests in items, and making recommendations is the same as connecting the missing edges. A representative work is the Network-Based Inference (NBI) proposed by Zhou et al. [85] which generates recommendations based on the resource-allocation process. To make predictions for user ua, NBI first initializes the network as: AC(ua, ti) = ⎧⎨ ⎩1 if ti ∈ R(ua) 0 otherwise (2.2.4) where R(ua) denotes the set of items rated by user ua, and AC(ua, ti) is the Allocated Resource (AC) to node of item ti that represents user’s interests. In this initialization, 1 is assigned to every item ti if rated by user ua, and 0 otherwise. After this initial- ization for all users, the allocated resources will be redistributed among all items in the following two steps and the item with the most AC at the final stage will be recommended to the active user ua. Spreading Step In spreading step, all initially allocated resources will flow from all items T to all users U. The resources flow to each user ux is calculated as: AC(ux) = |T |∑ i=1 cxiAC(ux, ti) |S(ti)| (2.2.5) where S(ti) denotes the set of users who rated item ti, AC(ux, ti) denotes the initial resources allocated to item ti by user ux, and cxi is 1 if ux rated item ti and 0 otherwise. 18 Redistribution Step In this step, the resources will flow back to items T from users U. The final resources allocated to each item ti is calculated as: ÂC(ti) = |U|∑ x=1 cxiAC(ui) |Tx| (2.2.6) The item with most resources ÂC(ti) allocated will be recommended to the active user ua. 2.2.5 Trust-based Recommendation Similarity in typical recommendation methods is often defined by standard metrics such as cosine. However, instead of finding recommendations from similar users, it is also reasonable to find recommendations from familiar users [28]. Intuitively, an item liked by the user’s good friend has the potential to be liked by the user. Recent developments in social networks have further revealed the social trust rela- tionships among users, and Massa and Avesani [53] termed this kind of recommender systems as Trust-Aware Recommendation. Empirical results from Guy’s work [28] in- dicated that familiarity-based methods can be superior to similarity-based methods. Despite of the performance comparison, the key advantage of trust-aware methods is that it provides a promising approach to cold-start problem [27,28]. 2.3 Relative Preference-based Recommender Sys- tems 19 User preferences can be modeled in three types: pointwise, pairwise, and listwise. Though RecSys is not limited to pointwise absolute ratings, the recommendation task is usually considered as a rating prediction problem [38, 40, 69, 78]. Recently, a considerable literature [12, 19, 45, 64, 74] has grown up around the theme of rela- tive preferences, especially the pairwise PR. Meanwhile, recommendation task is also shifting from rating prediction to item ranking [56,74,83] in which the ranking itself is also relative preferences. The use of relative preferences has been widely studied in the field of Information Retrieval for learning to rank tasks [21,22,35]. Recently, PR-based [12,19,45,64] and listwise-based [74] RecSys have been proposed. Among them, the PR-based approach is the most popular, which can be further categorized as memory-based methods [12] that capture local structure and model-based methods [19,45,64] that capture global structure. We summarize the capabilities of the existing methods in Table 2.2. Table 2.2: Capabilities of existing methods Method Input Output LS GS Pointwise Memory-based Ratings Ratings � Pointwise Model-based Ratings Ratings � Pointwise Hybrid Ratings Ratings � � Pairwise Memory-based Preference Relations Item Rankings � Pairwise Model-based Preference Relations Item Rankings � 20 2.3.1 Notation and Problem Statement Preference Relation A preference relation (PR) encodes user preferences in form of pairwise ordering between items. This representation is a useful alternative to absolute ratings for three reasons. Firstly, PR is more consistent across like-minded users [12,19] as it is invariant to many irrelevant factors, such as mood. Secondly, PR is a more natural and direct input for Top-N recommendation, as both the input and the output are relative preferences. Finally, and perhaps most importantly, PR can be obtained implicitly rather than asking the users explicitly. For example, the PR over two Web pages can be inferred by the stayed time, and consequently applies to the displayed items. This property is important as not all users are willing to rate their preferences, where collecting feedbacks implicitly delivers a more user-friendly RecSys. In addition, PR- based RecSys provides an opportunity to utilize the vast amount of implicit data that have already been collected over the years, such as activity logs. With these potential benefits, we shall take a closer look at the PR, and investigate how they can be utilized in RecSys. We formally define the PR as follows. Let U = {u}n and I = {i}m denote the set of n users and m items, respectively. The preference of a user u ∈ U between items i and j is encoded as πuij, which indicates the strength of user u’s PR for the ordered item pair (i,j). A higher value of πuij indicates a stronger preference on the first item over the second item. 21 Definition 2 (Preference Relation). The preference relation is defined as πuij = ⎧⎪⎪⎪⎪⎪⎪⎨ ⎪⎪⎪⎪⎪⎪⎩ (2 3 ,1] if i � j (u prefers i over j) [1 3 , 2 3 ] if i � j (i and j are equally preferable to u) [0, 1 3 ) if i ≺ j (u prefers j over i) (2.3.1) where πuij ∈ [0,1] and πuij = 1 − πuji. This definition is similar to [19], however, we allocate an interval for each pref- erence category, i.e., preferred, equally preferred, and less preferred. Indeed, each preference category can be further break down into more intervals. Similar to [12], the PR can be converted into user-wise preferences over items. Definition 3 (User-wise Preference). The user-wise preference is defined as pui = ∑ j∈Iu[[πuij > 2 3 ]] − ∑ j∈Iu[[πuij < 1 3 ]] |Πui| (2.3.2) where [[·]] gives 1 for true and 0 for false, and Πui is the set of user u’s PR related to item i. The user-wise preference pui falls in the interval [−1,1], where −1 and 1 indicate that item i is the least or the most preferred item for user u, respectively. The user- wise preference measures the relative position of an item for a particular user, which is different from absolute ratings. Preference relation has been widely studied in the field of Information Retrieval [14, 21,22,35]. Nevertheless, PR-based RecSys have only emerged recently [12,19,45,64]. Problem Statement Generally, the task of PR-based RecSys is to take PR as input and output Top-N recommendations. Specifically, let πuij ∈ Π encode the PR of each user u ∈ U. 22 Each πuij is defined over an ordered item pair (i, j), denoting i ≺ j, i � j, or i � j as described in Eq. 2.3.1. The goal is to estimate the value of each unknown πuij ∈ Πunknown, such that π̂uij approximates πuij. This can be considered as an optimization task performs directly on the PR: π̂uij = arg min π̂uij∈[0,1] (πuij − π̂uij)2 (2.3.3) However, it can be easier to estimate the π̂uij by the difference between the two user-wise preferences pui and puj, i.e., π̂uij = φ(p̂ui−p̂uj), where φ(·) is a function that bounds the value into [0,1] and ensures φ(0) = 0.5. For example, the inverse-logit function φ(x) = e x 1+ex can be used when user-wise preferences involve large values. Therefore, the objective of this paper is to solve the following optimization problem: (p̂ui, p̂uj) = arg min p̂ui,p̂uj (πuij − φ(p̂ui − p̂uj))2 (2.3.4) which optimizes the user-wise preferences directly, and Top-N recommendations can be obtained by simply sorting the estimated user-wise preferences. Let us consider an instance space X = {xi} (e.g. items) and a finite set of labels (e.g. ratings) Y = {yi|i = 1,2, ...,k}. One task of preference learning is to find a label ranking for any instance, e.g., determine the most likely rating for an item. The other task is to find an object ranking, e.g., to determine the ranking of items. For ease of reference, notations used in Relative Preference-based RecSys are summarized in Table 2.3. The letters u, v, a, b represent users, and the letters i, j, k, l represent items. 23 Table 2.3: Notations used in Relative Preference-based RecSys Notations Mathematical Meanings U the set of users I the set of items Π the set of preference relations pui the user-wise preference of user u on item i G an undirected graph encodes relations of user-wise preferences V the set of vertices each represents a user-wise preference E the set of edges each connects two vertices fuv the correlation feature between users u and v fij the correlation feature between items i and j wuv the weight associated to the user-user correlation feature fuv wij the weight associated to the item-item correlation feature fij Q(pui | u,i) the ordinal distribution o the side information, e.g., user attributes and item content 24 2.3.2 Memory-based Models A memory-based model is proposed in [12] to take preference relations as input and compute similarities between users. The proposed model has the following three steps: Collecting User Profiles When preference relations are employed, four values are possible for user preferences: • i � j indicates item i is preferred over item j • i ≺ j indicates item j is preferred over item i • i ≈ j indicates item i and j are equally preferable Each value correspond to one question answered by a user. However, there are too many possible questions which cannot be asked. Therefore, a decision must be made to decide which subset of questions to ask. Computing Similarities Between Users Let Iu be the set of preference relations of user u, and fu1,u2(i,j) indicating whether two users u1 and u2 agree on their preference on the two items i and j. Given an item pair (i,j), the function fu1,u2(i,j) gives the value 1 if the two users have the same preference, and 0 otherwise. The similarity measure between two users u1 and u2 is then defined as: cos�(u1,u2) = ∑ (i,j)∈I1∩I2 fu1,u2(i,j)√∑ (i,j)∈I1 fu1,u1(i,j) · √∑ (i,j)∈I2 fu2,u2(i,j) = ∑ (i,j)∈I1∩I2 fu1,u2(i,j)√ |I1| · |I2| (2.3.5) 25 where the numerator represents the number preferences that both users agreed, and the denominator normalizes the result. Making Recommendations To make recommendations, the preference relations are first converted into user-wise preferences. Denote: • c+u,i: the number of preference relations that i is preferred • c=u,i: the number of preference relations that i is equally preferred to others • c−u,i: the number of preference relations that i is less preferred Then the user-wise preference of item i by user u is defined as: pui = c−u,i − c+u,i c−u,i + c + u,i + c = u,i (2.3.6) With the user-wise preferences computed, the preference over an unknown item j can be predicted by: puj = ∑ v∈Nu sim(u,v) · pvj∑ v∈Nu sim(u,v) (2.3.7) where Nu is the set of users that have similar profiles to user u. 2.3.3 Model-based Models Ordinal Matrix Factorization The ordinal nature of preferences has been overlooked in RecSys literature, until recently Ordinal Matrix Factorization (OMF) [32,42,61,82] has emerged to explore the ordinal properties of ratings. 26 In general, OMF aims to generate an ordinal distribution Q(rui|u,i) over all pos- sible rating values for each user/item pair. Predicting the rating for user u on item i is then equivalent to identifying the rating with the greatest mass in the ordinal distribution Q(rui|u,i). While traditional RecSys approaches make only a point esti- mate, the OMF produces a full distribution and each prediction is associated with a probability as a confidence measure. Typical OMF approaches assume the existence of a latent utility xui that captures how much the user u is interested in the item i. The latent utility xui can be defined in different ways [32, 42, 61, 82], but under the same framework of Random Utility Models [55] xui = μui + �ui (2.3.8) where μui is an internal score represents the interaction between the user u and the item i. The �ui is the random noise normally assumed to follow the logistic distribution in practice [42]. The latent utility xui is then generated from a logistic distribution centred at μui with the scale parameter sui proportional to the standard deviation xui ∼ Logi(μui,sui) (2.3.9) In collaborative filtering, the user-item interaction is often captured by MF tech- niques, thereby the internal score μui can be substituted with the MF term bui +p T u qi xui = bui + p T u qi + �ui (2.3.10) where pu and qi are, respectively, the latent feature vectors of the user u and the item i. Modelling the latent utility with MF reflects the name OMF. Despite how the latent utility is modelled, an ordinal assumption is required to convert the numerical utility into ordinal values. A common approach is the ordinal 27 logistic regression originally described by McCullagh [54], which assumes that the rating is chosen based on the interval to which the utility belongs rui = l if xui ∈ (θl−1,θl] for l < L and rui = L if xui > θL−1 (2.3.11) where L is the number of ordinal levels and θl are the threshold values of interest. Other assumptions [51] are also possible but McCullagh’s model is by far the most popular. The probability of receiving a rating l is therefore Q(rui = l|u,i) = ∫ θl θl−1 P(xui|θ) = F(θl) − F(θl−1) (2.3.12) where F(θl) is the cumulative logistic distribution evaluated at θl F(xui ≤ l|θl) = 1 1 + exp(−θuil−μui sui ) (2.3.13) where the thresholds θl can be parameterised to depend on user or item. This paper employs the user-specific thresholds parameterisation described in [42]. Therefore a set of thresholds {θul}Ll=1 is defined for each user u to replace the thresholds θuil in Eq. 2.3.13. Given the learned ordinal distribution Q(rui|u,i), not only the ratings can be predicted but also the confidence for each prediction. Preference Relation-based Matrix Factorization Matrix Factorization (MF) [41] is a popular approach to RecSys that has mainly been applied to absolute ratings. Recently, the PrefNMF [19] model was proposed to adopt PR input for MF models. The PrefNMF model discovers the latent factor space shared between users and items, where the latent factors describe both the taste of users and the characteristics of items. The attractiveness of an item to a user is then measured by the inner product of their latent feature vectors. 28 Formally, each user u is associated with a latent feature vector uu ∈ Rk and each item i is associated with a latent feature vector vi ∈ Rk, where k is the dimension of the latent factor space. The attractiveness of items i and j to the user u are u�u vi and u�u vj, respectively. When u � u vi > u � u vj the item i is said to be more preferable to the user u than the item j, i.e., i � j. The strength of this preference relation πuij can be estimated by u�u (vi − vj), and the inverse-logit function is applied to ensure π̂uij ∈ [0,1]: π̂uij = eu � u (vi−vj) 1 + eu � u (vi−vj) (2.3.14) The latent feature vectors uu and vi are learned by minimizing regularized squared error with respect to the set of all known preference relations Π: min uu,vi∈Rk ∑ πuij∈Π∧(i<j) (πuij − π̂uij)2 + λ(‖uu‖2 + ‖vi‖2) (2.3.15) where λ is the regularization coefficient. The optimization can be done with Stochastic Gradient Descent for the favor of speed on sparse data, or with Alternating Least Squares for the favor of parallelization on dense data. 2.4 Evaluation Metrics Classic problems such as classification often have some agreed evaluation metrics. However, recommendation techniques are evaluated in many different ways depending on the form of recommendation as well as the goal of recommendation. For example when predicting a user’s rating on a movie, accuracy metrics are often used to measure how close the predicted rating is to the true rating. On the other hand, if the RecSys predicts a ranking of items for a user, then other metrics will be required to measure 29 the correctness, diversity, etc. In this section, we describe some commonly used evaluation metrics for both rating and ranking based RecSys. 2.4.1 Accuracy Metrics To measure the recommendation quality, various accuracy metrics can be used. Two popular metrics are Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), which measure how close the prediction is to the ground truth Let ∑ a|R(ua)| be the number of unrated items by user ua, and r̂i be the predicted rating of item ti, the definition of MAE and RMSE are as follows: MAE = ∑ a,i |R̂x,i − Ra,i|∑ a|R(ua)| (2.4.1) RMSE = √∑ a,i |R̂a,i − Ra,i|∑ a|R(ua)| (2.4.2) MAE and RMSE are the commonly used metric in literature [68, 69] as well as in various competitions [7]. However, the prediction accuracy can also be measured in terms of correlations between the predicted and the ground truth. Different cor- relation measures exist and a popular one is Pearson Correlation Coefficient (PCC) defined as: PCC = ∑ a(R̂a − R̂)(Ra − R)√∑ a(R̂a − R̂)2 · √∑ a(Ra − R) (2.4.3) Other accuracy metrics are also developed, such as Accuracy/Precision [29], and Area Under ROC Curve (AUC) [85]. 30 2.4.2 Diversity Traditionally, the evaluation of RecSys is mainly based on accuracy metrics such as RMSE. However, the accuracy metrics can not evaluate the some properties of the items other than the preferences, such as Serendipity [25], Diversity [11], etc. One diversity metric is Personalization, in which the uniqueness of each user’s recommendation list is measured. Personalization refers the inter-user diversity [84] Personalization = 2 m(m − 1) ∑ x�=y ( 1 − |Lk(ux) ∩ Lk(uy)| |Lk(ux)| ) , (2.4.4) where m is the number of users, and (1 − |Lk(ux)∩Lk(uy)||Lk(ux)| ) is the Hamming distance between recommendation lists Lk(ux) and Lk(uy). 2.4.3 Coverage Coverage refers to the percentage of items of all items a RecSys can recommend. This metric is based on the observation that some items may not have the chance to be recommended to any user, which reduces the coverage of the system. Let N be the number of top places to be considered, Ld be the number of distinct items in all top-N recommendation lists, and L be the number of distinct items in all recommendation lists. The N-dependent coverage is defined as [25]: Coverage(N) = Ld/L (2.4.5) A low coverage means the RecSys can only make recommendations on a small number of distinct items, in other words, it always recommends the popular items. It can be shown that RecSys with high coverage implies higher diversity [48]. 31 2.4.4 Stability Stability measures consistency of recommendations for the same user [3]. The rec- ommendations generated by a stable RecSys should be similar after some new pref- erences are added. For example, the first recommendation of an unstable RecSys predicts movie A as 5-star and movie B as 1-star. Then the user watched movie A and rated it as 5-star. With this new preference added to the preferences data, the unstable RecSys then generates the second recommendation that predicts movie B as 5-star. The 5-star movie B which was 1-star, may lead to user confusion and lower the trust of the RecSys. The Stability property has been studied in detail in [4]. With various evaluation metrics available, the choice highly depends on the goal of the RecSys. In general, accuracy metrics such as MAE and RMSE are standard metrics for benchmark, and other metrics such as diversity and novelty are used to fulfill some additional requirements. Unlike other machine learning tasks which have agreed metrics, the evaluation of RecSys has gained great research interests and new metrics are keeping emerged. 2.4.5 Metrics for Relative Preference-based Models Traditional recommender systems aim to optimize RMSE or MAE which emphasizes on absolute ratings. However, the ultimate goal of recommender systems is usually to obtain the ranking of items [42], where good performance on RMSE or MAE may not be translated into good ranking results [42]. Therefore, we employ two evaluation metrics: Normalized Cumulative Discounted Gain@T (NDCG@T) [34] which is popular in academia, and Mean Average Precision@T (MAP@T) [13] which 32 is popular in contests 1. Among them, the NDCG@T metric is defined as NDCG@T = 1 K(T) T∑ t=1 2rt − 1 log2 (t + 1) (2.4.6) where rt is the relevance judgment of the item at position t, and K(T) is the normal- ization constant. The MAP@T metric is defined as MAP@T = 1 |Utest| ∑ u∈Utest T∑ t=1 Pu(t) min(mu, t) (2.4.7) where mu is the number relevant items to user u, and Pu(t) is user u’s precision at position t. Both metrics are normalized to [0,1], and a higher value indicates better performance. These metrics, together with other ranking-based metrics, require a set of relevant items to be defined in the test set such that the predicted rankings can be evaluated against. The relevant items can be defined in different ways. In this paper, we follow the same selection criteria used in the related work [12, 38] to consider items with the highest ratings as relevant. 2.5 Summary Numerous recommendation techniques have been developed over the last decade, ranging from the basic ones of content-based methods to the recent ones of context- based methods. These recommendation techniques performed well in real-world ap- plications such as Amazon, MovieLens, and Netflix. However, due to the extensive use of these rating-based datasets, existing models are specifically designed for the ratings format, whereas vast amount of implicit feedback such as log files has been 1KDD Cup 2012 and Facebook Recruiting Competition 33 stored but not utilized. The new emerged relative preference-based models provide a solution to make use of such implicit feedback, but lacks of modeling abilities com- paring to the well-established rating-based models. In the remaining chapters of this thesis, we identify and tackle weaknesses of relative preference-based models to make them more effective and applicable. Chapter 3 Ordinal Random Fields for Recommender Systems 3.1 Introduction Recommender Systems (RecSys) aim to suggest items that are potentially of interest to users, where the items can be virtually anything such as movies and attractions for travel. To identify the appropriate items, RecSys use various sources of infor- mation including item content [5] and user preferences [41]. By far, Collaborative Filtering [41, 69] is one of the most popular RecSys techniques, which exploits user preferences especially the numerical preferences. However, numerical preferences are often difficult to collect as users may find it easier to tell which item is preferable to others, rather than expressing the precise degree of liking. Furthermore, researchers argued that numerical preferences may not be completely trustworthy [12, 42]. For example, the internal scales of users can be different, where the rating 4 out of 5 generally indicates high quality, but it is possible to be just fine for critical users. While users are not good at making consistent quantitative judgment, ordinal preferences are considered to be more consistent across 34 35 like-minded users [18]. Ordinal preferences is an alternative view of user preferences, in which the relative orders between items are measured. To adopt ordinal preferences, substantial research efforts have been made over the past five years [42, 61, 73, 82]. While most data collections are still dominated by numerical preferences, the shift from numerical to ordinal is a slow process. Instead of going solely ordinal preferences in a sudden, most existing ordinal approaches begin with exploiting the ordinal properties possessed by numerical preferences. Among them, Ordinal Matrix Factorization (OMF) has been suggested as an effective method in recent developments [32,42,61,82]. In contrast to the numerical approaches, OMF makes weaker assumptions as the user preferences are no longer required to be interpreted as numbers, instead, only the ordering of items matters. Despite of its effectiveness in modeling ordinal properties, OMF is incapable of exploiting the local structure described as follows. Typical collaborative filtering methods discover two types of information: the neighborhoods and the latent factors, which we refer to as the local and the global structures of the preferences: Local Structure The local structure (LS) refers to the second-order interactions between similar users or items. This type of information is often used by neighborhood-based collaborative filtering, in which the predictions are made by looking at the neighborhood of users [65] or items [69]. Though the majority of preferences will be ignored in making predictions, LS-based approaches are effective when the users/items correlations are highly localized. Global Structure The global structure (GS) refers to the weaker but higher-order interactions among all users and items. This type of information is often used by 36 latent factor models such as SVD [41] and LDA [52], which aim at discovering the latent factor spaces in the preferences. GS-based approaches are often competitive in terms of accuracy as well as computational efficiency. Exiting literature has suggested that the LS and the GS are complementary since they address different aspects of the preferences [38,79]. In 2008, a unified framework has been proposed by Koren [38] to capture both structures, but only for numerical preferences. To the best of our knowledge, there is yet no method for the OMF to capture both the LS and the GS. Recent advances in Probabilistic Graphical Models, especially the Markov Random Fields (MRF), have provided methods of building RecSys capable of exploiting both the LS and the GS [79]. However, there has been little attempt to address the ordinal preferences issue due to the complication of modeling ordinal preferences with the MRF. This chapter aims to develop a unified model in which the OMF and the MRF are seamlessly combined to take advantages of both the representational power of the MRF and the ease of modeling ordinal preferences by the OMF. The proposed Ordinal Random Fields (ORF) model is not designed for a particular OMF but can incorporate any OMF model that produces ordinal distributions such as those in [32, 42,61,82]. While this work primarily focuses on exploiting the LS, the representational power of the ORF is by no mean limited to this. For example, the MRF employed in ORF can be extended to Conditional Random Fields (CRF) [43,78] to fuse auxiliary information such as the item content [5] and social relations [50]. These information has been shown helpful in making better recommendations [6,50], and becomes even more valuable when the preferences data are highly sparse. Besides the extensibility, 37 the ORF inherits other advantages of the probabilistic graphical models as well, such as supporting missing data by its nature, and disciplined learning and inferences techniques. The remaining part of this chapter is organized as follows. Section 3.2 reviews the basic concepts of the Matrix Factorization and the OMF which form the basis of this work. Section 3.3 is devoted to the proposed ORF model. In Section 3.4, experimental results of the proposed ORF model are presented. Finally, Section 3.5 concludes this chapter by summarizing the main contributions and future works. 3.2 Preliminaries RecSys usually predict users’ future interest in items. Let U and I, denote the set of all users and the set of all items, respectively. The interest of the user u ∈ U in the item i ∈ I is encoded as the preference rui ∈ R, where the rating matrix R contains all known preferences. Definition 4 (Recommender System). RecSys aims to identify the item î ∈ I that maximizes the interest of the target user u ∈ U [1] î = arg max i∈I (rui) (3.2.1) In the rest of this section, we briefly review two RecSys approaches: Matrix Fac- torization and Ordinal Matrix Factorization that form a basis of this work For ease of reference, notations used throughout this chapter are summarized in Table 3.1, and the term preference and rating will be used interchangeably. 38 Table 3.1: Summary of Major Notations (Chapter 3) Notations Mathematical Meanings U the set of all users I the set of all items R the set of known preferences G an undirected graph which encodes relations of preferences V the set of vertices each represents a preference E the set of edges each connects two vertices ru the set of all preferences by user u fij the correlation feature between items i and j wij the weight associated to the correlation feature fij L the number of rating levels, and the ratings are integers from 1 to L 39 3.2.1 Matrix Factorization Matrix Factorization (MF) [41] is a popular and accurate approach to RecSys. This approach discovers the latent factor spaces shared between users and items, where the latent factors can be used to describe both the taste of users and the characteristics of items. The attractiveness of an item to a user is then measured by the inner product of their latent feature vectors. Formally, each user u is associated with a latent feature vector pu ∈ Rk and each item i is associated with a latent feature vector qi ∈ Rk, where k is the number of factors. The aim of MF is then to estimate r̂ui = bui + p T u qi such that r̂ui � rui. The bias term bui = μ + bu + bi takes the biases into consideration, where μ is the overall average rating, bu is the user bias, and bi is the item bias. The latent feature vectors are learned by minimizing regularized squared error with respect to all known preferences min pu,qi∈Rk ∑ (u,i)∈R (rui − bui − pTu qi)2 + λ(‖pu‖ 2 + ‖qi‖2) (3.2.2) where λ is the regularization coefficient. The optimization can be done with Stochastic Gradient Descent for the favor of speed on sparse data, or with Alternating Least Squares for the favor of parallelization on dense data. Comparing to neighbor-based approaches [69], MF-based approaches [38,40] have shown advantages in terms of accuracy and computational efficiency. Nevertheless, all of these approaches treat the preferences as numerical and are incapable of exploiting ordinal preferences. 40 3.2.2 Ordinal Matrix Factorization The ordinal nature of preferences has been overlooked in RecSys literature, until recently Ordinal Matrix Factorization (OMF) [32,42,61,82] has emerged to explore the ordinal properties of ratings. In general, OMF aims to generate an ordinal distribution Q(rui|u,i) over all pos- sible rating values for each user/item pair. Predicting the rating for user u on item i is then equivalent to identifying the rating with the greatest mass in the ordinal distribution Q(rui|u,i). While traditional RecSys approaches make only a point esti- mate, the OMF produces a full distribution and each prediction is associated with a probability as a confidence measure. Typical OMF approaches assume the existence of a latent utility xui that captures how much the user u is interested in the item i. The latent utility xui can be defined in different ways [32, 42, 61, 82], but under the same framework of Random Utility Models [55] xui = μui + �ui (3.2.3) where μui is an internal score represents the interaction between the user u and the item i. The �ui is the random noise normally assumed to follow the logistic distribution in practice [42]. The latent utility xui is then generated from a logistic distribution centered at μui with the scale parameter sui proportional to the standard deviation xui ∼ Logi(μui,sui) (3.2.4) In collaborative filtering, the user-item interaction is often captured by MF tech- niques, thereby the internal score μui can be substituted with the MF term bui +p T u qi xui = bui + p T u qi + �ui (3.2.5) 41 where pu and qi are, respectively, the latent feature vectors of the user u and the item i. Modelling the latent utility with MF reflects the name OMF. Despite how the latent utility is modeled, an ordinal assumption is required to convert the numerical utility into ordinal values. A common approach is the ordinal logistic regression originally described by McCullagh [54], which assumes that the rating is chosen based on the interval to which the utility belongs rui = l if xui ∈ (θl−1,θl] for l < L and rui = L if xui > θL−1 (3.2.6) where L is the number of ordinal levels and θl are the threshold values of interest. Other assumptions [51] are also possible but McCullagh’s model is by far the most popular. The probability of receiving a rating l is therefore Q(rui = l|u,i) = ∫ θl θl−1 P(xui|θ) = F(θl) − F(θl−1) (3.2.7) where F(θl) is the cumulative logistic distribution evaluated at θl F(xui ≤ l|θl) = 1 1 + exp(−θuil−μui sui ) (3.2.8) where the thresholds θl can be parameterized to depend on user or item. This paper employs the user-specific thresholds parameterization described in [42]. Therefore a set of thresholds {θul}Ll=1 is defined for each user u to replace the thresholds θuil in Eq. 3.2.8. Given the learned ordinal distribution Q(rui|u,i), not only the ratings can be predicted but also the confidence for each prediction. 3.2.3 Summary Matrix Factorization has been one of the most popular RecSys approaches, which primarily focuses on numerical preferences such as ratings. Nevertheless, the nature 42 of user preferences is often ordinal, and the importance of modeling ordinal properties has been recognized in recent works on OMF [32,42,61,82]. Although the OMF en- ables the modeling of ordinal properties, the employment of MF makes it only focuses on the higher-order interactions (the GS) regardless of the localized interactions (the LS), whereas both information are valuable [38, 79]. Furthermore, the OMF by its nature cannot model auxiliary information such as content [5] directly. The powerful representation of Markov Random Fields (MRF) offers an oppor- tunity to take advantages from all of these information, and have been developed in recent works [78, 79]. Nevertheless, exploiting the ordinal properties is not an easy task for MRF [79], therefore the strengths of the OMF and the MRF are nicely complementary. This observation leads to a naturally extension of unifying these two approaches, and motivates the present work. 3.3 Ordinal Random Fields In this section, we propose the Ordinal Random Fields (ORF) to model the ordinal properties and capture both the LS and the GS. Here we exploit the LS of the item- item correlations only, while the user-user correlations can be modeled in a similar manner. The rest of this section introduces the concept of the Markov Random Fields followed a detailed discussion of the ORF including its feature design, parameter estimation, and predictions. 43 3.3.1 Markov Random Fields Markov Random Fields (MRF) [16, 78] models a set of random variables having Markov property with respect to an undirected graph G. The undirected graph G consists a set of vertices V connected by a set of edges E without orientation, where two vertices are neighborhood of each other when connected. Each vertex in V encodes a random variable, and the Markov property implies that a variable is conditionally independent of other variables given its neighborhoods. In this work, we use MRF to model user preferences and their relations respect to a set of undirected graphs. Specifically for each user u, there is a graph Gu with a set of vertices Vu and a set of edges Eu. Each vertex in Vu represents a preference rui of user u on item i, and each edge in Eu captures a relation between two preferences by the same user. As we consider only the item-item correlations in this work, two preferences are connected by an edge if and only if they are given by the same user. Fig. 3.1 shows an example of two graphs for users u and v. Note that vertices of different graphs are not connected directly, however, the edges between the same pair of items are associated to the same item-item correlation. For example, the edge between rui and ruj and the edge between rvi and rvj are associated to the same item-item correlation between items i and j (see the green dashed line in Fig. 3.1). Formally, let I(u) be the set of all items rated by user u and ru = {rui|i ∈ I(u)} be the joint set of all preferences (the variables) related to user u, then the MRF defines a distribution P(ru) over the graph Gu: P(ru) = 1 Zu Ψ(ru) (3.3.1) 44 ruj ruk rvj rvk rui rvi i-j correlation j-k correlati on i-k correlati on Figure 3.1: Example of undirected graphs for users u and v Ψ(ru) = ∏ (ui,uj)∈Eu ψij(rui,ruj) (3.3.2) where Zu is the normalization term that ensures ∑ ru P(ru) = 1, and ψ(·) is a positive function known as potential. The potential ψij(rui,ruj) captures the correlation between items i and j ψij(rui,ruj) = exp{wijfij(rui,ruj)} (3.3.3) where fij(·) is the feature function and wij is the corresponding weight. The cor- relation features capture the LS, while the weights realize the importance of each correlation feature. In ORF, the weights also control the relative importance between the LS and the GS. With the weights estimated from data, the unknown preference rui can be predicted as r̂ui = arg max rui P(rui|ru) (3.3.4) where P(rui|ru) serves as the confidence measure of the prediction. 45 3.3.2 ORF: Unifying MRF and OMF The standard MRF approach captures the LS by modeling item-item correlations under the framework of probabilistic graphical models. However, it employs the log-linear modeling as shown in Eq. 3.3.3, and therefore does not enable a simple treatment of ordinal preferences. OMF, on the other hand, can nicely model the ordinal preferences in a probabilistic way but is weak in capturing the LS. The com- plementary between these two techniques calls for the unified ORF model to take all of the advantages. Essentially, the proposed ORF model promotes the agreement between the GS discovered by the OMF and the LS discovered by the MRF. More specifically, the ORF model combines the item-item correlations (Eq. 3.3.3) and the point-wise ordinal distribution Q(rui|u,i) obtained from the OMF (Eq. 3.2.7) P(ru) ∝ Ψu(ru) ∏ rui∈ru Q(rui|u,i) (3.3.5) where Ψu(ru) is the potential function capturing the interaction among items, and ru is the set of preferences from user u. The potential function Ψu(ru) can be further factorized into pairwise potentials based on Eq. 3.3.3 and Eq. 3.3.2: Ψu(ru) = exp ⎛ ⎝ ∑ rui,ruj∈ru wijfij(rui,ruj) ⎞ ⎠ (3.3.6) where fij(·) is the correlation feature between items i and j to be defined shortly in Section 3.3.3, and wij is the corresponding weight controls the relative importance of each correlation feature (LS) to the ordinal distribution (GS). Put all together, the 46 joint distribution P(ru) for each user u can be modelled as P(ru) ∝ exp ⎛ ⎝ ∑ rui,ruj∈ru wijfij(rui,ruj) ⎞ ⎠ ∏ rui∈ru Q(rui|u,i) (3.3.7) where there is a graph for each user but the weights are optimised by all users. In fact, the user-user correlations can also be captured as P(R) ∝ ∏ i Ψi(ri) ∏ u Ψu(ru) ∏ u,i Q(rui|u,i) (3.3.8) but we limit our discussion to item-item correlations in this paper. 3.3.3 Feature Design A feature is essentially a function f of n > 1 arguments that maps the (n-dimensional) input onto the unit interval f : Rn → [0,1], where the input can be ratings or auxiliary information such as content [78]. The item-item correlation is captured by the following feature f(rui,ruj) = g(|(rui − r̄i) − (ruj − r̄j)|) (3.3.9) where g(α) = 1−α/(L−1) does normalization, and r̄i and r̄j are the average ratings for items i and j, respectively. This correlation feature captures the intuition that correlated items should receive similar ratings by the same user after offsetting the goodness of each item. Though this work focuses on the item-item correlations, the feature for user-user correlations can be designed in a similar manner: f(rui,rvi) = g(|(rui − r̄u) − (rvi − r̄v)|) (3.3.10) where r̄u and r̄v are the global average ratings for users u and v respectively. 47 Although the user and item bias have been modeled by the underlying OMF, the ORF itself can also model the bias with identity features for item i and for user u fi(rui, i) = g(|rui − r̄i|), fu(rui,u) = g(|rui − r̄u|) (3.3.11) Indeed, auxiliary information such as content [5] and social relations [50] can also be modeled by designing corresponding features. That being said, the ORF is a generic framework with great extensibility to integrate multiple sub-components such as neighborhood, content, and ordinal ratings. Nevertheless, this work focuses on the item-item correlation features only. Since one correlation feature exists for each possible pair of co-rated items, the number of correlation features can be large, and this makes the estimation slow to converge and less robust. Therefore we only keep the correlation features if strong correlation exists between two items i and j. Specifically, the strong correlation features are extracted based on the Pearson correlation and a user-specified minimum correlation threshold. 3.3.4 Parameter Estimation In general, MRF models cannot be determined by standard maximum likelihood approaches, instead, approximation techniques such as Markov Chain Monte Carlo (MCMC) [26] and Pseudo-likelihood [8] are often used in practice. The pseudo- likelihood leads to exact computation of the loss function and its gradient with respect to parameters, and thus faster. The MCMC-based methods may, on the other hand, lead to better estimation given enough time. As the experiments involve different settings and large number of features, this study employs the pseudo-likelihood tech- nique to perform efficient parameter estimation by maximizing the regularized sum 48 of log local likelihoods L(w) = ∑ rui∈R log P(rui|ru\rui) − 1 2σ2 ∑ u∈U wTu wu (3.3.12) where σ is the regularization coefficient, and wu is the subset of weights related to user u. The local likelihood is defined as P(rui|ru\rui) = 1 Zui Q(rui|u,i) exp ⎛ ⎝ ∑ ruj∈ru\rui wijfij(rui,ruj) ⎞ ⎠ (3.3.13) where Zui is the normalization term. Zui = L∑ rui=1 Q(rui|u,i) exp ⎛ ⎝ ∑ ruj∈ru\rui wijfij(rui,ruj) ⎞ ⎠ (3.3.14) To optimize the parameters, we use the stochastic gradient ascent procedure that updates the parameters by passing through the set of ratings of each user: wu ← wu + η∇L(wu) (3.3.15) where η is the learning rate. More specifically, for each rui and its neighbor ruj in the set of ratings ru by user u, update the weight wij using the gradient of the log pseudo-likelihood ∂logL ∂wij = fij(rui,ruj) − L∑ rui=1 P(rui|ru\rui)fij(rui,ruj) (3.3.16) 3.3.5 Preference Prediction The prediction of rating rui is straightforward, which can be done by identifying the rating with the greatest mass in local likelihood: r̂ui = arg max rui P(rui|ru) (3.3.17) 49 where the local likelihood is given by Eq. 3.3.13. Prediction made in this approach identifies the most likely rating from discrete values 1 to L, and the local likelihood serves as a confidence measure. For predictions of scalar values, the expectation can be used instead: r̂ui = L∑ rui=1 ruiP(rui|ru) (3.3.18) Finally, Alg. 1 summarizes the learning and prediction procedures for the ORF. 3.4 Experiment and Analysis To study the performance of the proposed ORF model, comparisons were made with the following representative algorithms: a) K-Nearest Neighbors (K-NN) [65, 69], which represents the methods exploiting the LS; b) OMF [42], which exploits the GS and ordinal properties; c) and finally the ORF model, which takes ordinal prop- erties into account and exploits both the LS and the GS. Details of the experimental settings and results are presented in this section. 3.4.1 Experimental Settings Datasets Experiments were conducted on two public movie rating datasets: the MovieLens-100K and the MovieLens-1M 1 datasets. The MovieLens-1M dataset contains roughly 1 million ratings by 6040 users on 3900 movies. The MovieLens- 100K dataset contains 100K ratings by 943 users on 1682 movies. Ratings are on the 1 − 5 scale. To perform a reliable evaluation, we keep only users who rated at least 30 1http://grouplens.org/datasets/movielens 50 Algorithm 1 Ordinal Random Fields Algorithm Input: User preferences R; the ordinal distribution Q from Eq. 3.2.7. Step 1: Generate strong correlation features: fstrong ← {fij|Pearson(i,j) ≥ minCorr} Step 2: Initialize the weights: ∀wij ∈ w, wij ← N(0,0.01); Step 3: Repeat for each u ∈ U do for each rui,ruj ∈ ru, i �= j do if fij ∈ fstrong then Compute correlation feature fij according to Eq. 3.3.9 Compute normalization term Zui according to Eq. 3.3.14 Compute local likelihood according to Eq. 3.3.13 Compute the gradient for weight wij according to Eq. 3.3.16 Update wij with the gradient wij ← wij + η∇L(wij) end if end for end for Until stopping criteria met Predictions: * Predict most likely rating with confidence measure using Eq. 3.3.18. * Predict expectation using Eq. 3.3.17 51 movies, and each dataset is shuffled and split into the disjoint training set, validation set and test set. For each user, 5 ratings are kept in the validation set for tuning the hyper-parameters, 10 ratings are reserved for testing, and the rest for training. Evaluation Metric The Mean Absolute Error (MAE) and the Root Mean Square Error (RMSE) are used as the evaluation metric MAE = 1 |Rtest| ∑ (u,i)∈Rtest |r̂ui − rui|, RMSE = √∑ (u,i)∈Rtest(r̂ui − rui)2 |Rtest| (3.4.1) where Rtest is the test set kept aside until all parameters have been tuned. A smaller MAE or RMSE value indicates better performance. Although both metrics are used, we consider the MAE metric to be more suitable for ordinal preferences. The reason is that it makes more scenes to consider being off by 4 is just twice as bad as being off by 2 when the preferences are ordinal. The RMSE metric, on the other hand, can be skewed to methods that are optimized for numerical preferences. Parameters To perform a fair comparison, we fix the number of latent factors to the typical value of 50 for all algorithms, and all weights are randomly initialized from N(0,0.01). The number of neighbors for K-NN algorithms is set to 10 and 30. The minimum correlation threshold for the ORF is set to reasonable values considering both the prediction performance and computational efficiency. We will also report the effect of varying the minimum correlation threshold. 52 3.4.2 Result and Analysis We first compare the performance of the proposed ORF model with related algo- rithms: user-based K-NN, item-based K-NN and OMF, where the OMF is the tar- geted baseline. Then the impact of parameters is investigated for the ORF model, in particular the regularization coefficient and the minimum correlation threshold. Comparison with Other Methods The comparison results in terms of prediction accuracy are shown in Table 3.2. The global average is used only as a benchmark, which uses the average rating as the predictions. The following observations can be made based on the results. Firstly, the K-NN methods, especially the item-based K-NN, perform quite well. As the K-NN methods exploit only the LS, this result indicates the effectiveness of the LS. However, the ignorance of the GS makes the K-NN methods not less generalized and thus highly susceptible to noisy data. Secondly, the OMF fits the data quite well when predicting the most likely rat- ings for the MAE metric. However, it exploits only the GS and therefore further improvements are possible by incorporating the LS information. Finally, the ORF has made further improvements upon the OMF by unifying the modeling of both the LS and the GS, as well as ordinal properties. Note that the performance of the ORF relies on the ordinal distributions generated by the underlying OMF, which can be implemented in different ways [32, 42, 61, 82]. In present work, the improvements over the OMF are soley based on incorporating the LS information. To confirm the improvements, a paired t-test (two-tailed) with a confidence level of 53 Table 3.2: For both the OMF and the ORF, the expectation values (Eq. 3.3.18) are used for RMSE and the most likely values (Eq. 3.3.17) are used for MAE. MovieLens-100K MovieLens-1M Method RMSE MAE RMSE MAE Global Ave. 1.1186 0.9430 1.1123 0.9401 UserKNN, K=10 0.9687 0.7584 0.9350 0.7328 ItemKNN, K=10 0.9372 0.7305 0.9032 0.7041 UserKNN, K=30 0.9463 0.7413 0.9149 0.7173 ItemKNN, K=30 0.9315 0.7295 0.8987 0.70478 OMF 0.9525 0.7226 0.9144 0.6918 ORF, minCorr=0.4 0.9475 0.7185 0.9117 0.6887 ORF, minCorr=0.3 0.9448 0.7148 0.9093 0.6870 Table 3.3: Paired t-test for the ORF and the OMF. t-test statistics Methods df t p-value ORF vs. OMF on MAE 9 6.0163 0.0002 ORF vs. OMF on RMSE 9 4.8586 0.0009 54 95% has been applied to the ORF and the OMF. Results shown in Table 3.3 confirm that the performance of methods with and without capturing the LS is statistically significant. Impact of Minimum Correlation Threshold As mentioned in Section 3.3.3, the ORF model requires a minimum correlation thresh- old to control the number of correlation features. The reason is that the number of correlation features can be very large, which makes the model less robust and slow to converge. Specifically, when this threshold goes to minimum (e.g. −1 for Pearson cor- relation), the potential number of correlation features can be as large as n2/2 where n is the number of items. On the other hand, the number of correlation features goes to zero when the threshold goes to maximum, and the ORF reduces to the OMF. (a) Number of Correlation Features (b) Coverage of Correlation Features Figure 3.2: Impact of Minimum Correlation Threshold on Number of Correlation Features 55 Fig. 3.2(a) shows the number of correlation features for different minimum corre- lation thresholds. Given that the MovieLens-100K dataset contains less items com- paring the MovieLens-1M dataset, there are even more correlation features remained in the MovieLens-100K dataset when the threshold becomes larger. This observation implies that the items in the MovieLens-100K dataset are more correlated with each other. We also show the coverage of correlation features for both datasets, and the MovieLens-100K has consistently higher coverage of correlation features. (a) MAE (b) RMSE Figure 3.3: Impact of Minimum Correlation Threshold Having these statistics result, we further examine the impact of the minimum cor- relation threshold on prediction accuracy, as plotted in Fig. 3.3. It can be observed that the prediction accuracy improves as the minimum correlation threshold becomes smaller. However, we notice that the performance on the smaller MovieLens-100K dataset is not as stable as that on the MovieLens-1M dataset, where the curve of the MovieLens-1M dataset is smoother and shows better monotonicity. One explanation 56 is that the MovieLens-100K dataset may not have enough data to make robust esti- mation for large number of weights. However, given adequate data and time, the best prediction performance can be achieved by including all correlation features, i.e., the minimum correlation threshold is set to minimum. Impact of Regularization Coefficient While the number of correlation features can be large, the model might be prone to over-fitting. Therefore we investigate the impact of varying the the regularization coefficient. Fig. 3.4 shows the performance of the ORF under different regularization (a) MAE (b) RMSE Figure 3.4: Impact of Regularization Coefficient settings. We observe that by varying the regularization coefficient the prediction performance was not affected too much. One possible explanation is that the ordinal distribution employed in the ORF is generated by the underlying OMF with its own regularization mechanism, whereas the regularization term in the ORF controls only 57 the weights for the second-order item-item correlation features. In other words, the ORF model by itself is unlikely to over-fit the data given that the underlying OMF model has been properly regularized. 3.5 Summary In this work we presented the ORF model that takes advantages of both the rep- resentational power of the MRF and the ease of modeling ordinal preferences by the OMF. While the standard OMF approaches exploit only the GS, the ORF model is able to capture the LS as well. In addition, the ORF model defines a uniformed inter- face for different OMF approaches with various internal implementations. Last but not least, the ORF model is a generic framework that can be extended to incorporate additional information by designing more features. A future extension could take the user-user correlations into account as we modeled only the item-item correlations in this work. Incorporating the user-user correlations may further improve the prediction performance. Another future work is to take auxiliary information into consideration by replacing the MRF with the Conditional Random Fields [43]. Fusing auxiliary information such as the item content and social relations could improve the prediction performance especially when the data is highly sparse. Chapter 4 Preferecen Relation-based Recommender System 4.1 Introduction RecSys aim to recommend users with some of their potentially interesting items, which can be virtually anything ranging from movies to tourism attractions. To identify the appropriate items, RecSys attempts to exploit user preferences [41] and various side information including content [5,6], temporal dynamics [40], and social relations [50]. By far, Collaborative Filtering [41] is one of the most popular RecSys techniques, which exploits user preferences, especially in form of explicit absolute ratings. Nevertheless, relying on solely absolute ratings is prone to the cold-start problem [72] where few ratings are known for cold users or items. To alleviate the cold-start problem, additional information, which is usually heterogeneous [6] and implicit [64] must be acquired and exploited. Recently, a considerable literature [12, 19, 45, 64, 74] has grown up around the theme of relative preferences. The underlying motivation is that relative preferences are often easier to collect and more reliable as a measure of user preferences. For 58 59 example, it can be easier for users to tell which item is preferable than expressing the precise degree of liking. Furthermore, studies [12,42] have reported that absolute ratings may not be completely trustworthy. For example, rating 4 out of 5 may in general indicate high quality, but it can mean just OK for critics. In fact, users’ quantitative judgment can be affected by irrelevant factors such as the mood when rating, and this is called misattribution of memory [71]. While users are not good at making consistent quantitative judgment, the pref- erence relation (PR), as a kind of relative preference, has been considered as more consistent across like-minded users [12,18,19]. By measuring the relative order be- tween items, the PR is usually less variant to irrelevant factors. For example, a user in bad mood may give lower ratings to all items but the relative orderings between items remain the same. Being a more reliable type of user preferences, PR is also eas- ier to collect comparing to ratings as it can be inferred from implicit feedbacks. For example, the PR between two items can be inferred by comparing their ratings, page views, played counts, mouse clicks, etc. This property is important as not all users are willing to rate their preferences, where collecting feedbacks implicitly delivers a more user-friendly recommender system. In addition, as the ultimate goal of RecSys, obtaining the ranking of items by itself is to obtain the relative preferences, a more natural input than absolute ratings [42,81]. While the PR captures the user preferences in the pairwise form, most existing works [42,46] take the pointwise approach to exploiting ordinal properties possessed by absolute ratings. To accept the PR as input and output item rankings, pair- wise approaches to RecSys have recently emerged in two forms: memory-based [12] and model-based [19, 45, 64]. These studies have shown the feasibility of PR-based 60 methods, and demonstrated competitive performance comparing to their underly- ing models, such as memory-based K-Nearest Neighbor (KNN) [12] and model-based Matrix Factorization (MF) [19]. However, the limitations of these underlying models have constrained the poten- tials of their PR extensions. More specifically, both KNN and MF based methods can only capture one type of information at a time, while both the local and the global information are essential in achieving good performance [38,46,79]. We refer these two types of information as the local and the global structures of the preferences: Local Structure The local structure (LS) refers to the second-order interactions between similar users [65] or items [69]. This type of information is often used by neighborhood-based collaborative filtering, in which the predictions are made by looking at the neighborhood of users [65] or items [69]. LS-based approaches ignore the majority of preferences in making predictions, but are effective when the users or items correlations are highly localized. Global Structure The global structure (GS) refers to the weaker but higher-order interactions among all users and items [41]. This type of information is often used by latent factor models such as Matrix Factorization [41], which aim to discover the latent factor space in the preferences. GS-based approaches are often competitive in terms of accuracy and computational efficiency [41]. Previous studies have suggested that these two structures are complementary since they address different aspects of the preferences [38, 46, 79]. However, to the best of our knowledge, there is yet no PR-based method that can capture both LS and GS. Another problem of existing PR-based methods is that side information such as item content and user attributes can’t be easily incorporated, which is critical in 61 cold-start cases. All the above reasonings lead to the desired model with the following properties: 1) Accept PR as input; 2) Capture both LS and GS; 3) Side information can be easily incorporated; 4) Output item rankings. Recent advances in Markov Random Fields-based RecSys [16,46,77,79] have made it possible to achieve the above objectives. MRF-based RecSys was first developed in [79] to capture both LS and GS. Later on, it has been extended in [46] to exploit ordinal properties possessed by absolute ratings. Nevertheless, all of these attempts rely on absolute ratings. This work aims to push the MRF-based RecSys one step further by fitting it into the PR framework, namely the Preference Relation-based Markov Random Fields (PrefMRF) and the Preference Relation-based Conditional Random Fields (PrefCRF) when side information is incorporated. The remaining part of this paper is organized as follows. Section 4.2 introduces the concepts of PR-based RecSys and formalizes the problem, followed by a review of related work. Section 4.3 is devoted to the proposed PrefMRF and PrefCRF models. Benchmark results on Top-N recommendation are presented in Section 4.4. Finally, Section 4.5 concludes this work by summarizing the main contributions and envisaging future works. 4.2 Preliminaries RecSys aim at predicting users’ future interest in items, and the recommendation task can be considered as a preference learning problem, which aims to construct a predictive preference model from observed preference information [58]. Existing preference learning methods are based on different learning to rank approaches [23]. 62 Among them, the pointwise approach is the choice of most RecSys [38, 69], which exploit absolute ratings, though pairwise approach that exploits PR has been largely overlooked until recently. The rest of this section describes the basic concepts and formalizes the PR-based RecSys followed by a review of related work. 4.2.1 Preference Relation A preference relation (PR) encodes user preferences in form of pairwise ordering between items. This representation is a useful alternative to absolute ratings for three reasons. Firstly, PR is more consistent across like-minded users [12,19] as it is invariant to many irrelevant factors, such as mood. Secondly, PR is a more natural and direct input for Top-N recommendation, as both the input and the output are relative preferences. Finally, and perhaps most importantly, PR can be obtained implicitly rather than asking the users explicitly. For example, the PR over two Web pages can be inferred by the stayed time, and consequently applies to the displayed items. This property is important as not all users are willing to rate their preferences, where collecting feedbacks implicitly delivers a more user-friendly RecSys. In addition, PR- based RecSys provides an opportunity to utilize the vast amount of implicit data that have already been collected over the years, such as activity logs. With these potential benefits, we shall take a closer look at the PR, and investigate how they can be utilized in RecSys. We formally define the PR as follows. Let U = {u}n and I = {i}m denote the set of n users and m items, respectively. The preference of a user u ∈ U between items i and j is encoded as πuij, which indicates the strength of user u’s PR for the ordered 63 item pair (i,j). A higher value of πuij indicates a stronger preference on the first item over the second item. Definition 5 (Preference Relation). The preference relation is defined as πuij = ⎧⎪⎪⎪⎪⎪⎪⎨ ⎪⎪⎪⎪⎪⎪⎩ (2 3 ,1] if i � j (u prefers i over j) [1 3 , 2 3 ] if i � j (i and j are equally preferable to u) [0, 1 3 ) if i ≺ j (u prefers j over i) (4.2.1) where πuij ∈ [0,1] and πuij = 1 − πuji. This definition is similar to [19], however, we allocate an interval for each pref- erence category, i.e., preferred, equally preferred, and less preferred. Indeed, each preference category can be further break down into more intervals. Similar to [12], the PR can be converted into user-wise preferences over items. Definition 6 (User-wise Preference). The user-wise preference is defined as pui = ∑ j∈Iu[[πuij > 2 3 ]] − ∑ j∈Iu[[πuij < 1 3 ]] |Πui| (4.2.2) where [[·]] gives 1 for true and 0 for false, and Πui is the set of user u’s PR related to item i. The user-wise preference pui falls in the interval [−1,1], where −1 and 1 indicate that item i is the least or the most preferred item for user u, respectively. The user- wise preference measures the relative position of an item for a particular user, which is different from absolute ratings. 4.2.2 Problem Statement Generally, the task of PR-based RecSys is to take PR as input and output Top-N recommendations. Specifically, let πuij ∈ Π encode the PR of each user u ∈ U. 64 Each πuij is defined over an ordered item pair (i, j), denoting i ≺ j, i � j, or i � j as described in Eq. 4.2.1. The goal is to estimate the value of each unknown πuij ∈ Πunknown, such that π̂uij approximates πuij. This can be considered as an optimization task performs directly on the PR: π̂uij = arg min π̂uij∈[0,1] (πuij − π̂uij)2 (4.2.3) However, it can be easier to estimate the π̂uij by the difference between the two user-wise preferences pui and puj, i.e., π̂uij = φ(p̂ui−p̂uj), where φ(·) is a function that bounds the value into [0,1] and ensures φ(0) = 0.5. For example, the inverse-logit function φ(x) = e x 1+ex can be used when user-wise preferences involve large values. Therefore, the objective of this paper is to solve the following optimization problem: (p̂ui, p̂uj) = arg min p̂ui,p̂uj (πuij − φ(p̂ui − p̂uj))2 (4.2.4) which optimizes the user-wise preferences directly, and Top-N recommendations can be obtained by simply sorting the estimated user-wise preferences. For ease of reference, notations used throughout this chapter are summarized in Table 4.1. The letters u, v, a, b represent users, and the letters i, j, k, l represent items. 4.2.3 Related Work User preferences can be modeled in three types: pointwise, pairwise, and listwise. Though RecSys is not limited to the pointwise absolute ratings, the recommendation task is usually considered as a rating prediction problem. Recently, a considerable literature [12,17,19,24,36,45,64,74] has grown up around the theme of relative pref- erences, especially the pairwise PR. Meanwhile, recommendation task is also shifting 65 Table 4.1: Summary of Major Notations (Chapter 4) Notations Mathematical Meanings U the set of users I the set of items Π the set of preference relations pui the user-wise preference of user u on item i G an undirected graph encodes relations of user-wise preferences V the set of vertices each represents a user-wise preference E the set of edges each connects two vertices fuv the correlation feature between users u and v fij the correlation feature between items i and j wuv the weight associated to the user-user correlation feature fuv wij the weight associated to the item-item correlation feature fij Q(pui | u,i) the ordinal distribution produced by PrefNMF o the side information, e.g., user attributes and item content 66 from rating prediction to item ranking [56,74,83], in which the ranking itself is also relative preferences. Preference relation has been widely studied in the field of The use of relative preferences has been widely studied in the field of Information Retrieval [14,21,22,35,64] for learning to rank tasks. Recently, PR-based [12,19,45, 64] and listwise-based [74] RecSys have been proposed. Among them, the PR-based approach is the most popular, which can be further categorized as memory-based methods [12] that capture local structure and model-based methods [19, 45, 64] that capture global structure. To the best of our knowledge, there is yet no PR-based method that can capture both LS and GS. Advances in Markov Random Fields (MRF) and its extension Conditional Random Fields (CRF) have made it possible to utilize both LS and GS by taking advantages of MRF’s powerful representation capability. Nevertheless, exploiting the PR is not an easy task for MRF and CRF [46,79]. This observation leads to a natural exten- sion of unifying the MRF models with the PR-based models, to complement their strengths. We summarize the capabilities of the existing and our proposed PrefMRF and PrefCRF models in Table 4.2. 4.3 Methodology In this section, we propose the Preference Relation-based Markov Random Fields (PrefMRF) to model the PR and capture both LS and GS. When side information is taken into consideration, this model extends to Preference Relation-based Conditional Random Fields (PrefCRF). In this work, we exploit LS in terms of the item-item correlations as well as the user-user correlations. The rest of this section introduces 67 Table 4.2: Capabilities of PR-based methods Method Input Output LS GS Side Information Pointwise Memory-based Ratings Ratings � Pointwise Model-based Ratings Ratings � Pointwise Hybrid Ratings Ratings � � Pairwise Memory-based Preference Relations Item Rankings � Pairwise Model-based Preference Relations Item Rankings � PrefMRF Preference Relations Item Rankings � � PrefCRF Preference Relations Item Rankings � � � the concept of the Preference Relation-based Matrix Factorization(PrefNMF) [19] that will be our underlying model, and then followed a detailed discussion of the PrefMRF and PrefCRF on issues including feature design, parameter estimation, and predictions. 4.3.1 Preference Relation-based Matrix Factorization Matrix Factorization (MF) [41] is a popular approach to RecSys that has mainly been applied to absolute ratings. Recently, the PrefNMF [19] model was proposed to adopt PR input for MF models. Like MF models, the PrefNMF model discovers the latent factor space shared between users and items, where the latent factors describe both the taste of users and the characteristics of items. The attractiveness of an item to a user is then measured by the inner product of their latent feature vectors. 68 Point Estimation The PrefNMF model described in [19] provides a point estimation to each preference, i.e., a single value. Formally, each user u is associated with a latent feature vector uu ∈ Rk and each item i is associated with a latent feature vector vi ∈ Rk, where k is the dimension of the latent factor space. The attractiveness of items i and j to the user u are u�u vi and u � u vj, respectively. When u � u vi > u � u vj the item i is said to be more preferable to the user u than the item j, i.e., i � j. The strength of this preference relation πuij can be estimated by u � u (vi − vj), and the inverse-logit function is applied to ensure π̂uij ∈ [0,1]: π̂uij = eu � u (vi−vj) 1 + eu � u (vi−vj) (4.3.1) The latent feature vectors uu and vi are learned by minimizing regularized squared error with respect to the set of all known preference relations Π: min uu,vi∈Rk ∑ πuij∈Π∧(i<j) (πuij − π̂uij)2 + λ(‖uu‖2 + ‖vi‖2) (4.3.2) where λ is the regularization coefficient. The optimization can be done with Stochastic Gradient Descent for the favor of speed on sparse data, or with Alternating Least Squares for the favor of parallelization on dense data. Distribution Estimation The original PrefNMF [19] computes the attractiveness of an item to a user by the product of their latent feature vectors which results a scalar value, where the likeli- hoods of other possible values remain unknown. However, in order to be combined with MRF models, we wish to have the distributions over a set of possible values. 69 Therefore the Random Utility Models [55] and the Ordinal Logistic Regression [54] are applied to perform the conversion. Random Utility Models [55] assume the existence of a latent utility xui = μui +�ui that captures how much the user u is interested in the item i, where μui captures the interest and �ui is the random noise that follows the logistic distribution as in [42]. The latent utility xui is then generated from a logistic distribution centered at μui with the scale parameter sui proportional to the standard deviation: xui ∼ Logi(μui,sui) (4.3.3) Recall that PrefNMF computes attractiveness of item to user by the product of their latent feature vectors, thereby the internal score μui can be substituted with the term u�u vi: xui = u � u vi + �ui (4.3.4) where uu and vi are, respectively, the latent feature vectors of the user u and the item i. The Ordinal Logistic Regression [54] is then used to convert the user-wise prefer- ences pui into ordinal values, which assumes that the preference pui is chosen based on the interval to which the latent utility belongs: pui = l if xui ∈ (θl−1,θl] for l < L and pui = L if xui > θL−1 (4.3.5) where L is the number of ordinal levels and θl are the threshold values of interest. The probability of receiving a preference l is therefore Q(pui = l | u,i) = ∫ θl θl−1 P(xui | θ) dθ = F(θl) − F(θl−1) (4.3.6) 70 where F(θl) is the cumulative logistic distribution evaluated at θl with standard de- viation sui: F(xui ≤ l | θl) = 1 1 + exp(−θuil−μui sui ) (4.3.7) The thresholds θl can be parameterized to depend on user or item. This paper employs the user-specific thresholds parameterization described in [42]. Therefore a set of thresholds {θul}Ll=1 is defined for each user u to replace the thresholds θuil in Eq. 4.3.7, and is learned from data. Given the learned ordinal distribution Q(pui | u,i), not only the preferences can be predicted but also the confidence for each prediction. The ordinal distribution Q(pui | u,i) captures the GS information in a probabilistic form, and will be incorporated into MRF and CRF in Section 4.3.4. Note that the user preference is quantized into ordinal values in this process. 4.3.2 Markov Random Fields Markov Random Fields (MRF) [16,78] model a set of random variables having Markov property with respect to an undirected graph G. The undirected graph G consists a set of vertices V connected by a set of edges E without orientation, where two vertices are neighborhood of each other when connected. Each vertex in V encodes a random variable, and the Markov property implies that a variable is conditionally independent of others given its neighborhoods. In this work, we use MRF to model user-wise preference and their interactions respect to a set of undirected graphs. Specifically for each user u, there is a graph Gu with a set of vertices Vu and a set of edges Eu. Each vertex in Vu represents a preference pui of user u on the item i. Note that the term preference is used instead 71 of rating because in the new model the preference is not interpolated as absolute ratings but user-wise ordering of items. Each edge in Eu captures a relation between two preferences by the same user. Two preferences are connected by an edge if they are given by the same user or on the same item, corresponding to the item-item and user-user correlations, respec- tively. Modeling these correlations is actually capturing the LS information in the preferences. However, it is not easy to model two types of correlations at the same time as it will result a large graph. Instead, we model the item-item and user-user correlations separately, and merge their predictions. Fig. 4.1 shows an example of four graphs for user u, user v, item i, and item j. Note that vertices of different graphs are not connected directly, however, the weights are estimated across graphs when the edges correspond to the same correlation. For example, the edge between pui and puj and the edge between pvi and pvj are associated to the same item-item correlation ψij between items i and j. Formally, let Iu be the set of all items evaluated by the user u and Ui be the set of all users rated the item i. Then denote pu = {pui | i ∈ Iu} be the joint set of all preferences (the variables) expressed by user u, and pi = {pui | u ∈ Ui} be the joint set of all preferences (the variables) rated on item i. Under this setting, the MRF defines two distributions P(pu) and P(pi) over the graphs Gu and Gi, respectively: P(pu) = 1 Zu Ψu(pu) , P(pi) = 1 Zi Ψi(pi) (4.3.8) Ψu(pu) = ∏ (ui,uj)∈Eu ψij(pui,puj) , Ψi(pi) = ∏ (ui,vi)∈Ei ψuv(pui,pvi) (4.3.9) where Zu and Zi are the normalization terms that ensure ∑ pu P(pu) = 1 and ∑ pi P(pi) = 1. The term ψ(·) is a positive function known as potential. 72 pui puk puj pul ψij pvi pvk pvj pvl ψij ψik ψik puj pvj paj pbj ψua ψuv pui pvi pai pbi ψua ψuv Figure 4.1: Example of MRF graphs. u, v, a, and b are users. i, j, l, and k are items. The potentials ψij(pui,puj) and ψuv(pui,pvi) capture the correlation between items i and j and correlation between users u and v, respectively: ψij(pui,puj) = exp{wijfij(pui,puj)} (4.3.10) ψuv(pui,pvi) = exp{wuvfuv(pui,pvi)} (4.3.11) where fij(·) and fuv(·) are the feature functions to be designed shortly in Section 4.3.4, and wij and wuv are the corresponding weights. These correlation features capture the LS information, while the weights realize the importance of each correlation feature. With the weights estimated from data, the unknown preference pui can be predicted using item-item correlations: p̂ui = arg max pui∈[−1,1] P(pui | pu) (4.3.12) 73 or using user-user correlations: p̂ui = arg max pui∈[−1,1] P(pui | pi) (4.3.13) where the confidences of the predictions can be measured by P(pui | pu) and P(pui | pi). 4.3.3 Conditional Random Fields Despite of the user preferences, various side information including content [5, 6], temporal dynamics [40], and social relations [50] are also important in making quality recommendations. While there exist methods to incorporate side information, there is yet no PR-based method that can achieve this. One advantage of Markov Random Fields is its extensibility, thus side informa- tion can be easily incorporated by extending the MRF to Conditional Random Fields (CRF). In MRF, the item-item and user-user correlations are modeled in a set of graphs, where each graph has a set of vertices representing the preferences. To incor- porate side information, the MRF is extended to CRF by conditioning each vertex on a set of global observations o, i.e., the side information in our context. Specif- ically, each user u is associated with a set of attributes {ou} such as gender and occupation. Similarly, each item i is associated with a set of attributes {oi} such as genres of movie. This side information is encoded as the set of global observations o = {{ou},{oi}}. The graphs for item-item and user-user correlations conditioned on global observations are illustrated in Fig. 4.2. Using the same settings as MRF in Section 4.3.2, the CRF models the conditional 74 p p p i k l j u p p p i k l j v p p p u v b a i p p p u v b a j Figure 4.2: Example of CRF graphs. u , v , a , and b are users. i , j , l , and k are items. 75 distributions P(pu | o) and P(pi | o) over the graphs Gu and Gi, respectively: P(pu | o) = 1 Zu(o) Ψu(pu,o) , P(pi | o) = 1 Zi(o) Ψi(pi,o) (4.3.14) Ψu(pu,o) = ∏ (ui)∈Vu ψui(pui,o) ∏ (ui,uj)∈Eu ψij(pui,puj) (4.3.15) Ψi(pi,o) = ∏ (ui)∈Vi ψui(pui,o) ∏ (ui,vi)∈Ei ψuv(pui,pvi) (4.3.16) where Zu(o) and Zi(o) are the normalization terms ensure ∑ pu P(pu | o) = 1 and∑ pi P(pi | o) = 1. The term ψ(·) is a positive function known as potential. The potential ψui(·) captures the global observations associated to the user u and the item i: ψui(pui,o) = exp{w�u fu(pui,oi) + w�i fi(pui,ou))} (4.3.17) The potentials ψij(·) and ψuv(·) capture the item-item and user-user correlations, respectively: ψij(pui,puj) = exp{wijfij(pui,puj)} (4.3.18) ψuv(pui,pvi) = exp{wuvfuv(pui,pui)} (4.3.19) where fu, fi, fij, and fuv are the features to be designed shortly in Section 4.3.4, and wu, wi, wij, and wuv are the corresponding weights realize the importance of each feature. With the weights estimated from data, the unknown preference pui can be predicted using item-item correlations: p̂ui = arg max pui∈[−1,1] P(pui | pu,o) (4.3.20) or using user-user correlations: p̂ui = arg max pui∈[−1,1] P(pui | pi,o) (4.3.21) where P(pui | pu,o) and P(pui | pi,o) give the confidence of the predictions. 76 4.3.4 PrefCRF: Unifying PrefNMF and CRF The CRF model captures the LS information by modeling item-item and user-user correlations under the framework of probabilistic graphical models. However, it em- ploys the log-linear modeling as shown in Eq. 4.3.18 and Eq. 4.3.18, and therefore does not enable a simple treatment of PR. The PrefNMF model, on the other hand, can nicely model the PR but is weak in capturing the LS and side information. The complementary between these two techniques calls for the unified PrefCRF model to take all of the advantages. Essentially, the proposed PrefCRF model promotes the agreement between the GS discovered by the PrefNMF, the LS discovered by the MRF, and the side information discovered by the CRF. More specifically, the PrefCRF model combines the item-item and user-user correlations (Eq. 4.3.15 and Eq. 4.3.16) and the ordinal distributions Q(pui | u,i) over user-wise preferences obtained from Eq. 4.3.6: P(pu | o) ∝ Ψu(pu,o) ∏ pui∈pu Q(pui | u,i) (4.3.22) P(pi | o) ∝ Ψi(pi,o) ∏ pui∈pi Q(pui | u,i) (4.3.23) where Ψu is the potential function capturing the interaction among items evaluated by user u, and Ψi is the potential function capturing the interaction among users rated item i. Put all together, the joint distribution P(pu) for each user u can be modeled as: P(pu) ∝ exp ⎛ ⎝ ∑ pui,puj∈pu wijfij(pui,puj) + ∑ (ui)∈Vu ψui(pui,o) ⎞ ⎠ ∏ pui∈pu Q(pui | u,i) (4.3.24) 77 and the joint distribution P(pi) for each item i can be modeled as: P(pi) ∝ exp ⎛ ⎝ ∑ pui,pvi∈pi wuvfuv(pui,pvi) + ∑ (ui)∈Vi ψui(pui,o) ⎞ ⎠ ∏ pui∈pi Q(pui | u,i) (4.3.25) where there is a graph for each user or item but the weights are optimized by all users or all items. Feature Design A feature is essentially a function f of n > 1 arguments that maps the n-dimensional input into the unit interval f : Rn → [0,1]. We design the following kinds of features: Correlation Features The item-item correlation is captured by the feature: fij(pui,puj) = g(|(pui − p̄i) − (puj − p̄j)|) (4.3.26) and the user-user correlation is captured by the feature fuv(pui,pvi) = g(|(pui − p̄u) − (pvi − p̄v)|) (4.3.27) where g(α) normalizes feature values and α plays the role of deviation. The terms p̄i, p̄j, p̄i, and p̄j are the item or user averages. The item-item correlation feature captures the intuition that correlated items should be ranked similarly by the same user after offsetting the goodness of each item. Similarly, the user- user correlation feature captures the intuition that correlated users should rate the same item similarly. 78 Attribute Features Each user u and item i has a set of attributes ou and oi, respec- tively. These attributes are mapped to preferences by the following features: fi(pui) = oug(|(pui − p̄i)|) fu(pui) = oig(|(pui − p̄u)|) (4.3.28) where fi models which users like the item i and fu models which classes of items the user u likes. Since one correlation feature exists for each possible pair of co-rated items, the number of correlation features can be large which makes the estimation slow to con- verge and less robust. Therefore we only keep the correlation features if strong item- item correlation or user-user correlation exists. Specifically, the strong correlation features fstrong are extracted based on the Pearson Correlation and a user-specified minimum correlation threshold. Note that the correlation is calculated based on the user-wise preferences generated from PR thus the rule of using PR as input is not violated. Parameter Estimation In general, MRF-based models cannot be determined by standard maximum likeli- hood approaches, instead, approximation techniques are often used in practice such as Pseudo-likelihood [8] and Contrastive Divergence (CD) [30]. The Pseudo-likelihood leads to exact computation of the loss function and its gradient with respect to pa- rameters, and thus faster. The CD-based methods may, on the other hand, lead to better estimation given enough time. As the experiments involve different settings and large number of features, this study employs the Pseudo-likelihood technique to perform efficient parameter estimation by maximizing the regularized sum of log local 79 likelihoods: logL(w) = ∑ pui∈Π log P(pui | pu,o) − 1 2σ2 w�w (4.3.29) where w are the weights and 1/2σ2 controls the regularization. To make the notation uncluttered, we write pu instead of explicitly as pu\pui. In this section we describe the parameter estimation of item-item correlations, where the user-user correlations can be estimated in the same way by replacing items with users. The local likelihood in Eq. 4.3.29 is defined as: P(pui | pu,o) = 1 Zui(o) Q(pui | u,i)ψui(pui,o) ∏ puj∈pu ψij(pui,puj) (4.3.30) where Zui(o) is the normalization term: Zui = lmax∑ pui=lmin Q(pui | u,i)ψui(pui,o) ∏ puj∈pu ψij(pui,puj) (4.3.31) where lmin is the first and lmax is the last interval, i.e., 1 and 3 in our settings. To optimize the parameters, we use the stochastic gradient ascent procedure that updates the parameters by passing through the set of ratings of each user: w ← w + η∇logL(w) (4.3.32) where η is the learning rate. More specifically, for each pui we update the attribute weights wo = {wu,wi} and correlation weight wij for each neighbor puj ∈ pu using the gradient of the log pseudo-likelihood ∂logL ∂wo = fo(pui,o) − lmax∑ pui=lmin P(pui | pu,o)fo(pui,o) − wi σ2 (4.3.33) ∂logL ∂wij = fij(pui,puj) − ∑lmax pui=lmin P(pui | pu,o)fij(pui,puj) − wijσ2 (4.3.34) 80 Item Recommendation The ultimate goal of RecSys is often to rank the items and recommend the Top-N items to the user. To obtain the item rankings, PrefMRF estimates distributions over user-wise preferences which can be converted into point estimate: The PrefCRF produces distributions over the user-wise preferences, which can be converted into point estimate: Most Likely Preference The preference can be determined by selecting the pref- erence with the greatest mass in local likelihood: p̂ui = arg max pui P(pui | pu,o) (4.3.35) where the local likelihood is given by Eq. 4.3.30. The local likelihood serves as a confidence measure. Smoothed Expectation When the prediction is not strict to discrete values, the expectation can be used instead: p̂ui = lmax∑ pui=lmin puiP(pui | pu,o) (4.3.36) where l refers to the intervals of user-wise preferences: from least to most preferred. Note that l is limited to the simplest case of 3 intervals in our settings, but more intervals are possible. The predictions by item-item correlation and user-user correlations can be merged by taking the mean value, and then items can be sorted and ranked accordingly. Finally, Alg. 2 summarizes the learning and prediction procedures for the PrefCRF. 81 Algorithm 2 PrefCRF Algorithm Input: Explict or implicit preferences. Step 1: Infer PR from preferences. Step 2: Predict user-wise preferences p̂ui using Eq. 4.2.2. Step 3: Predict distribution for each p̂ui using Eq. 4.3.6. Step 4: Repeat for each u ∈ U do for each pui ∈ pu do Compute normalization term Zui using Eq. 4.3.31 Compute local likelihood using Eq. 4.3.30 Compute attribute feature fi and fu using Eq. 4.3.28 Compute gradients for attribute features fo using Eq. 4.3.33 Update wo with the gradient using Eq. 4.3.32 for each puj ∈ pu, i �= j ∧ fij ∈ fstrong do Get correlation feature fij and fuv using Eq. 4.3.26 and Eq. 4.3.27 Get gradient for correlation feature fij using Eq. 4.3.34 Update wij with the gradient using Eq. 4.3.32 end for end for end for Until stopping criteria met Predictions: * Predict user-wise preferences using Eq. 4.3.36 or Eq. 4.3.35. * Select Top-N items according to estimated preferences. 82 Computational Complexity We perform the computational complexity analysis on the PrefMRF and its under- lying PrefNMF algorithms. Given n users and m items each with du and di prefer- ences, respectively. Let us temporarily ignore the user-specified latent factors. Then the complexity of both PrefNMF and PrefMRF is O(nd2u). However, in practice few item co-rated by the same user are strong neighbors of each other due to the correla- tion threshold defined in Section 4.3.4. As a result, the computation time of PrefMRF tends to be O(nduc) where c is a factor of correlation threshold. 4.4 Experiment and Analysis Datasets Ideally, the experiments should be conducted on datasets that contain user preferences in two forms: PR and absolute ratings. Unfortunately no such a dataset is publicly available at the moment, therefore we choose to compile the rating-based datasets into the form of PR. We use the same conversion method as in [19] by comparing the ratings of each ordered pair of items co-rated by the same user. For example, 1 is assigned to the PR πuij if pui > puj; 0 is assigned if pui < puj, and 0.5 is assigned if pui = puj. Experiments were conducted on two datasets: the MovieLens-1M 1 and the Each- Movie 2 datasets. The MovieLens-1M dataset contains more than 1 million ratings 1http://grouplens.org/datasets/movielens 2http://grouplens.org/datasets/eachmovie 83 by 6,040 users on 3,900 movies. The EachMovie dataset contains 2.8 million ratings by 72,916 users on 1,628 movies. The minimum rating is 1 and we cap the maximum at 5 for both datasets. The impact of side information is studied on the MovieLens- 1M dataset which provides gender, age, and occupation information about users and genres of movies. For a reliable and fair comparison, each dataset is split into train and test sets, and the following settings are aligned to related work [83]. As the sparsity levels differ between the MovieLens-1M and the EachMovie datasets, different number of ratings are reserved for training and the rest for testing. Specifically, for each user in the MovieLens-1M we randomly select N = 30, 40, 50, 60 ratings for training, and put the rest for testing. Some users do not have enough ratings thus were excluded from experiments. The EachMovie has less items but much more users comparing to MovieLens-1M, therefore it is safe to remove some less active users and we set N = 70, 80, 90, 100 to investigate the performance on dense dataset. Evaluation Metrics Traditional recommender systems aim to optimize RMSE or MAE which emphasizes on absolute ratings. However, the ultimate goal of recommender systems is usually to obtain the ranking of items [42], where good performance on RMSE or MAE may not be translated into good ranking results [42]. Therefore, we employ two evaluation metrics: Normalized Cumulative Discounted Gain@T (NDCG@T) [34] which is popular in academia, and Mean Average Precision@T (MAP@T) [13] which 84 is popular in contests 3. Among them, the NDCG@T metric is defined as NDCG@T = 1 K(T) T∑ t=1 2rt − 1 log2 (t + 1) (4.4.1) where rt is the relevance judgment of the item at position t, and K(T) is the normal- ization constant. The MAP@T metric is defined as MAP@T = 1 |Utest| ∑ u∈Utest T∑ t=1 Pu(t) min(mu, t) (4.4.2) where mu is the number relevant items to user u, and Pu(t) is user u’s precision at position t. Both metrics are normalized to [0,1], and a higher value indicates better performance. These metrics, together with other ranking-based metrics, require a set of relevant items to be defined in the test set such that the predicted rankings can be evaluated against. The relevant items can be defined in different ways. For example, for each user we can consider only 5-star items in the testing set as relevant items, or those items above each user’s average as relevant items. In this paper, we follow the same selection criteria used in the related work [12,38] to consider items with the highest ratings as relevant. Parameter Setting For a fair comparison, we fix the number of latent factors to 50 for all algorithms, the same as in related work [15]. The number of neighbors for KNN algorithms is set to 50. We vary the minimum correlation threshold to examine the performances with different number of features. Different values of regularization coefficient are also tested. 3KDD Cup 2012 and Facebook Recruiting Competition 85 4.4.1 Results and Analysis We first compare the performance of the proposed PrefMRF and PrefCRF models with four related models: KNN, NMF, PrefKNN, and PrefNMF, where the PrefNMF is the targeted model, and then investigate the impact of parameter settings. Comparison on Top-N Recommendation Comparison of these algorithms is conducted by measuring the NDCG and the MAP metrics on Top-N recommendation tasks. Each experiment is repeated ten times with different random seeds and we report the mean results with standard deviations on MovieLens-1M dataset in Table 4.3 and EachMovie dataset in Table 4.4. Note that only the MovieLens-1M dataset has side information which is used by the PrefCRF model. The PrefMRF as well as other models are based on only preferences data. We also report the NDCG and MAP values by varying the position T (i.e., how many items to recommend) in Fig. 4.3 and Fig. 4.4 for MovieLens-1M dataset and in Fig. 4.5 and Fig. 4.6 for MovieLens-1M dataset. The following observations can be made based on the results. Firstly, the KNN and the PrefKNN methods didn’t perform well on MovieLens- 1M comparing with Matrix Factorization based methods. One possible reason is that predictions are made based only on the neighbors, and as a result too much information has been ignored especially when the dataset is large. However, the performance of KNN -based methods has improved on the EachMovie dataset as we reserved more ratings for training, i.e., better neighbors can be found for prediction. Secondly, PrefNMF outperforms NMF on MovieLens-1M dataset which is con- sistent to the results reported in [19]. However, PreNMF does not perform well 86 Table 4.3: Results over ten runs on MovieLens-1M dataset. Given 30 Algorithm NDCG@5 NDCG@10 MAP@5 MAP@10 UserKNN 0.3969 ± 0.0020 0.4081 ± 0.0029 0.2793 ± 0.0021 0.2744 ± 0.0025 NMF 0.5232 ± 0.0057 0.5195 ± 0.0040 0.3866 ± 0.0055 0.3549 ± 0.0037 PrefKNN 0.3910 ± 0.0044 0.4048 ± 0.0038 0.2745 ± 0.0043 0.2720 ± 0.0037 PrefNMF 0.5729 ± 0.0049 0.5680 ± 0.0041 0.4387 ± 0.0046 0.3992 ± 0.0033 PrefMRF 0.6020 ± 0.0050 0.5934 ± 0.0039 0.4721 ± 0.0050 0.4244 ± 0.0036 PrefCRF 0.6316 ± 0.0076 0.5966 ± 0.0028 0.6254 ± 0.0073 0.4245 ± 0.0028 Given 40 Algorithm NDCG@5 NDCG@10 MAP@5 MAP@10 UserKNN 0.4108 ± 0.0040 0.4252 ± 0.0036 0.2936 ± 0.0036 0.2877 ± 0.0034 NMF 0.5323 ± 0.0050 0.5291 ± 0.0034 0.3976 ± 0.0045 0.3631 ± 0.0035 PrefKNN 0.4122 ± 0.0024 0.4283 ± 0.0024 0.2944 ± 0.0023 0.2904 ± 0.0023 PrefNMF 0.5773 ± 0.0037 0.5732 ± 0.0028 0.4437 ± 0.0041 0.4019 ± 0.0032 PrefMRF 0.6215 ± 0.0029 0.6140 ± 0.0023 0.4844 ± 0.0025 0.4420 ± 0.0020 PrefCRF 0.6435 ± 0.0064 0.6092 ± 0.0023 0.6420 ± 0.0062 0.4392 ± 0.0021 Given 50 Algorithm NDCG@5 NDCG@10 MAP@5 MAP@10 UserKNN 0.4273 ± 0.0040 0.4424 ± 0.0027 0.3078 ± 0.0038 0.3015 ± 0.0026 NMF 0.5360 ± 0.0041 0.5326 ± 0.0036 0.4010 ± 0.0040 0.3669 ± 0.0025 PrefKNN 0.4326 ± 0.0027 0.4483 ± 0.0030 0.3125 ± 0.0024 0.3070 ± 0.0022 PrefNMF 0.5761 ± 0.0067 0.5745 ± 0.0035 0.4424 ± 0.0064 0.4019 ± 0.0033 PrefMRF 0.6248 ± 0.0053 0.6172 ± 0.0032 0.4896 ± 0.0053 0.4460 ± 0.0027 PrefCRF 0.6648 ± 0.0055 0.6158 ± 0.0018 0.6580 ± 0.0059 0.4471 ± 0.0024 Given 60 Algorithm NDCG@5 NDCG@10 MAP@5 MAP@10 UserKNN 0.4480 ± 0.0044 0.4622 ± 0.0035 0.3266 ± 0.0036 0.3163 ± 0.0027 NMF 0.5462 ± 0.0068 0.5409 ± 0.0063 0.4109 ± 0.0069 0.3734 ± 0.0055 PrefKNN 0.4526 ± 0.0062 0.4689 ± 0.0039 0.3301 ± 0.0051 0.3223 ± 0.0033 PrefNMF 0.5756 ± 0.0062 0.5733 ± 0.0048 0.4409 ± 0.0059 0.4007 ± 0.0037 PrefMRF 0.6422 ± 0.0037 0.6301 ± 0.0037 0.5112 ± 0.0035 0.4600 ± 0.0026 PrefCRF 0.6772 ± 0.0074 0.6242 ± 0.0018 0.6715 ± 0.0072 0.4536 ± 0.0016 87 Table 4.4: Results over ten runs on EachMovie dataset without side information. Given 70 Algorithm NDCG@5 NDCG@10 MAP@5 MAP@10 UserKNN 0.7088 ± 0.0020 0.7115 ± 0.0015 0.6012 ± 0.0027 0.5767 ± 0.0017 NMF 0.7581 ± 0.0022 0.7577 ± 0.0017 0.6524 ± 0.0026 0.6225 ± 0.0020 PrefKNN 0.7260 ± 0.0022 0.7307 ± 0.0018 0.6197 ± 0.0020 0.5990 ± 0.0016 PrefNMF 0.7408 ± 0.0033 0.7348 ± 0.0039 0.6330 ± 0.0035 0.5800 ± 0.0038 PrefMRF 0.8317 ± 0.0032 0.8245 ± 0.0029 0.7512 ± 0.0039 0.6921 ± 0.0034 Given 80 Algorithm NDCG@5 NDCG@10 MAP@5 MAP@10 UserKNN 0.7146 ± 0.0018 0.7168 ± 0.0017 0.6070 ± 0.0021 0.5825 ± 0.0019 NMF 0.7636 ± 0.0021 0.7638 ± 0.0018 0.6583 ± 0.0025 0.6286 ± 0.0018 PrefKNN 0.7337 ± 0.0028 0.7377 ± 0.0018 0.6271 ± 0.0029 0.6057 ± 0.0021 PrefNMF 0.7422 ± 0.0036 0.7319 ± 0.0040 0.6329 ± 0.0039 0.5774 ± 0.0033 PrefMRF 0.8364 ± 0.0036 0.8232 ± 0.0030 0.7553 ± 0.0038 0.6991 ± 0.0032 Given 90 Algorithm NDCG@5 NDCG@10 MAP@5 MAP@10 UserKNN 0.7191 ± 0.0022 0.7279 ± 0.0028 0.6120 ± 0.0021 0.5933 ± 0.0013 NMF 0.7712 ± 0.0039 0.7692 ± 0.0033 0.6663 ± 0.0043 0.6431 ± 0.0034 PrefKNN 0.7418 ± 0.0028 0.7421 ± 0.0015 0.6357 ± 0.0030 0.6192 ± 0.0020 PrefNMF 0.7456 ± 0.0031 0.7358 ± 0.0038 0.6357 ± 0.0040 0.5819 ± 0.0036 PrefMRF 0.8394 ± 0.0035 0.8249 ± 0.0032 0.7474 ± 0.0037 0.7046 ± 0.0032 Given 100 Algorithm NDCG@5 NDCG@10 MAP@5 MAP@10 UserKNN 0.7279 ± 0.0028 0.7277 ± 0.0015 0.6238 ± 0.0032 0.5973 ± 0.0021 NMF 0.7741 ± 0.0030 0.7717 ± 0.0028 0.6719 ± 0.0034 0.6411 ± 0.0030 PrefKNN 0.7505 ± 0.0019 0.7511 ± 0.0012 0.6478 ± 0.0020 0.6231 ± 0.0014 PrefNMF 0.7391 ± 0.0033 0.7298 ± 0.0034 0.6318 ± 0.0039 0.5761 ± 0.0039 PrefMRF 0.8418 ± 0.0031 0.8277 ± 0.0030 0.7546 ± 0.0038 0.7063 ± 0.0036 88 on EachMovie where its performance is only slightly better than user-based KNN. The reason behind could be the EachMovie is much denser than the MovieLens-1M dataset, which makes the number of PR huge and difficult to tune optimal parame- ters. Besides, we observe that PrefNMF in general only achieves a slight improvement with more training data and even drops a bit with Given 60. Similarly for the Each- Movie dataset. With these observations, it appears that for a given number of users, the PrefNMF can be trained reasonably well with fewer data. Finally, the proposed PrefMRF and PrefCRF have made further improvement on both datasets upon the PrefNMF through capturing both LS and GS, as well as exploiting side information. From Fig. 4.3 and Fig. 4.4 we can see that the algorithms stabilized around position 10 and PrefMRF and PrefCRF consistently deliver better performance than others. It should be noted that the performance of PrefMRF and PrefCRF rely on their underlying model that captures the GS. In other words, the performance may vary when the PrefNMF is replaced with other alternative methods such as [45]. To confirm the improvements, a paired t-test (two-tailed) with a significance level of 95% has been applied to the best PrefMRF and the second best PrefNMF. Re- sults shown in Table 4.5 confirm that the performance of models with and without capturing the LS is statistically significant. Performance on Various Data Sparsity Levels To thoroughly examine the performance of these algorithms, we compare their per- formances under different settings of training set sizes: from Given 30 to Given 60 on MovieLens-1M dataset, and from Given 70 to Given 100 on EachMovie dataset. 89 (a) NDCG@T (Given 30) (b) MAP@T (Given 30) (c) NDCG@T (Given 40) (d) MAP@T (Given 40) Figure 4.3: Performance of different position T on MovieLens-1M dataset (Sparse). 90 (a) NDCG@T (Given 50) (b) MAP@T (Given 50) (c) NDCG@T (Given 60) (d) MAP@T (Given 60) Figure 4.4: Performance of different position T on MovieLens-1M dataset (Dense). 91 (a) NDCG@T (Given 70) (b) MAP@T (Given 70) (c) NDCG@T (Given 80) (d) MAP@T (Given 80) Figure 4.5: Performance of different position T on EachMovie dataset (Sparse). 92 (a) NDCG@T (Given 90) (b) MAP@T (Given 90) (c) NDCG@T (Given 100) (d) MAP@T (Given 100) Figure 4.6: Performance of different position T on EachMovie dataset (Dense). 93 Table 4.5: Paired t-test for PrefMRF and PrefNMF. Settings t-test statistics Dataset Sparsity Metric df t p-value MovieLens Given 60 NDCG@10 9 16.6218 < 0.00001 MovieLens Given 60 MAP@10 9 23.5517 < 0.00001 EachMovie Given 100 NDCG@10 9 72.4189 < 0.00001 EachMovie Given 100 MAP@10 9 72.1346 < 0.00001 Results are plotted in Fig. 4.7. It can be observed that in general more training data result in better performance. However, PrefNMF does not gain much benefit from more data and even perform slightly worse in Given 60. The PrefMRF on the other hand consistently gains performance from more data as the LS information can be better captured. Impact of Minimum Correlation Threshold As described in Section 4.3.4, a minimum correlation threshold is required to control the number of features in the PrefMRF model. By default, each pair of co-rated items has a feature which results in a large number of features. However, many of these features are useless if the item-item correlation are weak. To make the model more robust and with faster convergence, a minimum correlation threshold is applied to remove weak features. Specifically, the feature is removed if two items has a correlation measured by Pearson correlation less than the threshold. Results are 94 (a) Mean NDCG by training set sizes on MovieLens-1M dataset. (b) Mean MAP by training set sizes on MovieLens-1M dataset. (c) Mean NDCG by training set sizes on Each- Movie dataset. (d) Mean MAP by training set sizes on Each- Movie dataset. Figure 4.7: Impact of Sparsity Levels. 95 plotted in Fig. 4.8(a). It can be observed that a smaller correlation threshold delivers better performance, however, the number of features will also increase. To balance the performance and computation time, it is wise to select a moderate level of threshold depending on the dataset. (a) NDCG@10 by Threshold (b) NDCG@10 by Regularization Figure 4.8: Impact of Parameters (MovieLens-1M) Impact of Regularization Coefficient As the number of features in PrefMRF can be large, the model might be prone to over-fitting. Therefore, we investigate the impact of regularization settings as plotted in Fig. 4.8(b). We observe that the performance is better when a small regularization penalty applies. In other words, the PrefMRF can generalize reasonable well without too much regularization. This can be explained as the weights of item-item correlations 96 are not user-specific but shared by all users, thus they cannot over-fit every user perfectly. 4.5 Summary In this chapter we presented the PrefMRF model, which takes advantages of both the representational power of the MRF and the ease of modeling preference relations by the PrefNMF. To the best of our knowledge, there was no PR-based method that can capture both LS and GS, until the PrefMRF model is proposed in this work. In addition, side information can be easily incorporated by extending the PrefMRF model to the PrefCRF model. Experiment results on public datasets demonstrate that both types of interactions have been properly captured by PrefMRF, and significantly improved Top-N recommendation performance has been achieved. In addition, the PrefMRF model provides a generic interface for unifying various PR-based methods other than the PrefNMF used in this paper. In other words, any PR-based method that captures the GS should be able to take advantages from PrefMRF to capture LS as well. For future work, we would like to work on several directions. First, the compu- tation efficiency of PR-based matrix factorization needs to be improved given that the number of preference relations is usually much larger than absolute ratings. This is feasible as each user has his/her own set of preference relations, thus the learning process can be parallelized. Secondly, it would be interesting to see how PR-based methods perform on real implicit preferences dataset, such as page views and mouse clicks. Chapter 5 Learning from Heterogeneous Data Sources for Improved Top-N Recommendation 5.1 Introduction RecSys aim to recommend users with some of their potentially interesting items, which can be virtually anything ranging from movies to tourism attractions. To identify the appropriate items, RecSys attempts to exploit user preferences [41] and various side information [6]. However, the cold-start problem [72] raises when little information is known for cold users or items, e.g., a newly registered user. To alleviate the cold-start problem, additional information, which is usually heterogeneous, must be acquired. The last decade has seen a growing trend towards creating and managing more profiles in Online Social Networks (OSN), such as Facebook, LinkedIn, and Netflix etc. In light of this trend, it becomes possible to alleviate the cold-start problem by learning user preferences from heterogeneous data sources, e.g., a cold user of Netflix may have been used Facebook for a while. Nevertheless, user preferences from heterogeneous sources are heterogeneous whereas existing recommendation techniques 97 98 require the user preferences to be homogeneous. For example, user preferences are expressed as 5-star ratings for Netflix 1, 6-star ratings for EachMovie 2, and binary ratings for Facebook. Sometimes the user preferences may not even be expressed as ratings but as implicit feedbacks such as page views and clicks. Moreover, user preferences collected from heterogeneous sources may have different biases, as the user preferences not only reflect the quality of the items but also the quality of service providers. When the user preferences are heterogeneous and with biases, existing recommendation techniques cannot be directly applied. To the best of our knowledge, no previous work has considered the heterogeneous sources problem, in which the user preferences are heterogeneous with biases. In this work, we identify that the preference relations (PR), which measures the relative ordering between items, could be a key to the heterogeneous sources problem. With the assistance of PR, user preferences from heterogeneous sources can be fused seamlessly. For example, user preferences expressed as 5-star ratings, binary ratings, and page views can not be directly fused in general. However, all those user preferences can be deduced into the PR format by performing pairwise comparison on items. Once the user preferences are represented in PR, a direct merge can be performed. In fact, converting user preferences into PR not only provides a method to merge heterogeneous data but also reduces the biases that come with heterogeneity, i.e., the relative ordering of items is resistant to biases. In addition to collecting more user preferences, another method to alleviate the cold-start problem is to utilize the heterogeneous side information, such as attributes like age or occupation of users and items. 1http://www.netflixprize.com 2http://grouplens.org/datasets/eachmovie 99 This chapter aims to propose the Preference Relation-based Conditional Random Fields (PrefCRF) model to learn from heterogeneous data sources as well as the het- erogeneous side information The remaining part of this paper is organized as follows. Section 5.2 introduces the basic concepts of learning from heterogeneous sources and preference relations, followed by a review of related work. Section 5.3 is devoted to the proposed PrefCRF model. Benchmark results on Top-N recommendation are presented in Section 5.4. Finally, Section 5.5 concludes this chapter by summarizing the main contributions and envisaging future works. 5.2 Preliminaries This section briefly summarizes necessary background related to the heterogeneous sources problem and the preference relations that form the basis of our solution. 5.2.1 Heterogeneous Sources User preferences are usually assumed to come from a single homogeneous source. This assumption is becoming invalid given the rapid development of OSN, in which users maintain multiple profiles and the form of preferences diverges. We define two sources as heterogeneous if their preferences are 1) in different forms, e.g., ratings and clicks; 2) in different scales, e.g., 5-star scale and 6-star scale; 3) or biased differently due to factors irrelevant to the items’ quality, e.g., quality of the service providers. Based on this definition, not only the physically separated 100 sources are heterogeneous but a source changed significantly is also considered het- erogeneous to itself. In general, user preferences from heterogeneous sources cannot be merged directly as they may be in different forms. Even if their forms are the same, the scales could be different, where a force casting may change the meaning of preferences. In case that the scales are the same, biases are still introduced by the sources which make the recommendations inaccurate. 5.2.2 Preference Relation Preference relation (PR) encodes user preferences in the form of relative ordering between items, which is a useful alternative representation to absolute ratings as suggested in recent works [12, 19]. In fact, existing preferences such as ratings or other types of preferences can be easily represented as PR and then merged into a single dataset as shown in Fig. 5.1. . This property is particularly useful for the cold-start problem but has been overlooked in literature. We formally define the PR as follows. Let U = {u}n and I = {i}m denote the set of n users and m items, respectively. The PR of a user u ∈ U between items i and j is encoded as πuij, which indicates the strength of user u’s preference relation for the ordered item pair (i,j). A higher value of πuij indicates a stronger preference to the first item over the second item. 101 U se r Pr ef er en ce s Im pl ic it Pr ef er en ce s So ur ce 1 So ur ce 2 So ur ce n PR PR Pu rc ha se s Pl ay ed c ou nt C lic ks Pa ge v ie w s ... Ex pl ic it Pr ef er en ce s So ur ce 1 So ur ce 2 So ur ce n Th um bs u p/ do w n 5- st ar ra tin gs 6- st ar ra tin gs ... 5- st ar ra tin gs PR PR PR PR PR F ig u re 5 .1 : F lo w fr o m u se r p re fe re n ce s to P R 102 Definition 7 (Preference Relation). The preference relation is defined as πuij = ⎧⎪⎪⎪⎪⎪⎪⎨ ⎪⎪⎪⎪⎪⎪⎩ (2 3 ,1] if i � j (u prefers i over j) [1 3 , 2 3 ] if i � j (i and j are equally preferable) [0, 1 3 ) if i ≺ j (u prefers j over i) (5.2.1) where πuij ∈ [0,1] and πuij = 1 − πuji. An interval is allocated for each preference category, i.e., preferred, equally pre- ferred, and less preferred. Indeed, each preference category can be further break down into more intervals, though here in this paper we consider the minimal case of 3 intervals. Similar to [12], the PR can be converted into user-wise preferences over items which encode the ranking of items evaluated by a particular user. Definition 8 (User-wise Preference). The user-wise preference is defined as pui = ∑ j∈Iu\i[[πuij > 2 3 ]] − ∑ j∈Iu\i[[πuij < 1 3 ]] |Πui| (5.2.2) where Iu is the set of items related to user u, [[·]] gives 1 for true and 0 for false, and Πui is the set of user u’s PR related to item i. The user-wise preference pui falls in the interval [−1,1], where 1 and −1 indicate that item i is the most and the least preferred item for u, respectively. 5.2.3 Related Work Preference Relation (PR) has been widely studied in the field of Information Re- trieval [33]. Nevertheless, PR-based RecSys have only emerged recently [12,19,45,64]. 103 User preferences are usually categorized into three classes: pointwise, pairwise, and listwise. The pointwise preferences correspond to the absolute ratings widely used in RecSys, the pairwise preferences correspond to the PR presented in this work, and the listwise preferences is indeed the output of Top-N RecSys. Though RecSys is not limited to absolute ratings, the recommendation task is usually considered as a rating prediction problem. Recently, a considerable literature [12,19,45,64,74] has grown up around the theme of relative preferences. Meanwhile, recommendation task is also shifting from rating prediction to item ranking [74,83], in which the ranking itself is also relative preferences. Recently, pairwise preference relation-based [12, 19, 45, 64] and listwise-based [74] RecSys have been proposed. Among them, the pairwise approach is the most popular, which can be further cat- egorized as memory-based methods [12] and model-based methods [19, 45, 64]. The pairwise PR-based RecSys has the potential to unifying heterogeneous user prefer- ences, and this property, to the best of our knowledge, is first recognized in the present paper. Though the PR-based RecSys has the potential to alleviate the cold- start problem, it does not utilize the side information. Recent advances in PR-based RecSys [12, 19, 45, 64] and Conditional Random Fields (CRF) [78] have made it possible to unify the heterogeneous preferences into the unified PR format as well as modeling side information. This observation leads to a natural extension presented in this work to unify the CRF-based method with the PR-based methods, to complement their strengths. There exist two similar research topics that should be distinguished from our work. The first one is the cross-domain RecSys [60] which considers heterogeneous items, i.e., mixing movies and books, while our work considers the heterogeneous preferences 104 associated to the same type of items. The other topic is the mixture of experts [75], in which the non-personalized expert opinions from different sources are weighted into the personalized recommendations, while our work considers heterogeneous prefer- ences of the same user distributed across various sources. 5.3 Preference Relation-based Conditional Random Fields In this section, we propose the Preference Relation-based Conditional Random Fields (PrefCRF) to model both the heterogeneous preferences and the side informa- tion. The rest of this section defines the PR-based RecSys problem, and introduces the concept of the PrefNMF [19] that forms our underlying model, followed by a detailed description of the PrefCRF and discussion on issues such as feature design, parameter estimation, and predictions. 5.3.1 Problem Statement Generally, the task of PR-based RecSys is to take PR as input and output Top-N recommendations. Specifically, let πuij ∈ Π encode the PR of each user u ∈ U, and each πuij is defined over an ordered item pair (i, j), denoting i ≺ j, i � j, or i � j. The main task towards Top-N recommendations is to estimate the value of each unknown πuij ∈ Πunknown, such that π̂uij approximates πuij. This can be considered 105 as an optimization task that performs directly on the PR π̂uij = arg min π̂uij∈[0,1] (πuij − π̂uij)2 (5.3.1) However, it can be easier to estimate the π̂uij by the difference between two user- wise preferences pui and puj, i.e., π̂uij = φ(p̂ui − p̂uj), where φ(·) is a function that bounds the value into [0,1] and ensures φ(0) = 0.5. For example, the inverse-logit function φ(x) = e x 1+ex can be used when user-wise preferences involve large values. The objective of this paper is then to solve the following optimization problem (p̂ui, p̂uj) = arg min p̂ui,p̂uj (πuij − φ(p̂ui − p̂uj))2 (5.3.2) which optimizes the user-wise preferences directly, and Top-N recommendations can be obtained by simply sorting the estimated user-wise preferences. 5.3.2 Preference Relation-based Matrix Factorization Matrix Factorization (MF) [41] is a popular RecSys approach that has mainly been applied to absolute ratings. Recently, the PrefNMF [19] model was proposed to accommodate PR input for MF models. Like traditional MF models, the PrefNMF model discovers the latent factor space shared between users and items, where the latent factors describe both the taste of users and the characteristics of items. The attractiveness of an item to a user is then measured by the inner product of their latent feature vectors. Formally, each user u is associated with a latent feature vector uu ∈ Rk, and each item i is associated with a latent feature vector vi ∈ Rk, where k is the dimension of the latent factor space. The attractiveness of items i and j to user u are u�u vi and u�u vj, respectively. When u � u vi > u � u vj, the item i is said to be more preferable 106 to the user u than item j, i.e., i � j. The strength of this preference relation πuij can be estimated by u�u (vi − vj), and the inverse-logit function is applied to ensure π̂uij ∈ [0,1]: π̂uij = eu � u (vi−vj) 1 + eu � u (vi−vj) (5.3.3) The latent feature vectors uu and vi are learned by minimizing regularized squared error with respect to the set of all known preference relations Π: min uu,vi∈Rk ∑ πuij∈Π∧(i<j) (πuij − π̂uij)2 + λ(‖uu‖2 + ‖vi‖2) (5.3.4) where λ is the regularization coefficient. The optimization can be done with Stochastic Gradient Descent for the favor of speed on sparse data, or with Alternating Least Squares in favor of parallelization on dense data. 5.3.3 Conditional Random Fields Conditional Random Fields (CRF) [78] model a set of random variables having Markov property with respect to an undirected graph G, and each random variable can be conditioned on a set of global observations o. The undirected graph G consists of a set of vertexes V connected by a set of edges E without orientation, where two vertexes are neighboring to each other when connected. Each vertex in V encodes a random variable, and the Markov property implies that a variable is conditionally independent of others given its neighbors. In this work, we use CRF to model interactions among user-wise preferences conditioned on side information with respect to a set of undirected graphs. Specifically for each user u, there is a graph Gu with a set of vertexes Vu and a set of edges Eu. Each vertex in Vu represents a user-wise preference pui of user u on the item i. Each 107 edge in Eu captures a relation between two preferences by the same user. Specifically, two preferences are connected by an edge if they are given by the same user. Fig. 5.2 shows an example of two graphs for users u and v. Note that vertexes of different graphs are not directly connected, however, the edges between the same pair of items are associated to the same item-item correlation. For example, the edge between pui and puj and the edge between pvi and pvj are associated with the same item-item correlation ψij between items i and j. Each vertex is conditioned on a set of global observations o, which is the side information in our context. Specifically, each user u is associated with a set of L attributes {ou}L such as age, gender and occupation. Similarly, each item i is associated with a set of M attributes {oi}M such as genres for movie. Those side information is encoded as the set of global observations o = {{ou}L,{oi}M}. Formally, let pu = {pui | i ∈ Iu} be the joint set of preferences expressed by user u, then we are interested in modeling the conditional distribution P(pu | o) over the graph Gu. P(pu | o) = 1 Zu Ψu(pu,o) (5.3.5) Ψu(pu,o) = ∏ (ui)∈Vu ψui(pui,o) ∏ (ui,uj)∈Eu ψij(pui,puj) (5.3.6) where Zu(o) is the normalization term that ensures ∑ pu P(pu | o) = 1, and ψ(·) is a positive function known as potential. The potential ψui(·) captures the global observations associated to the user u and the item i, and the potential ψij(·) captures the correlations between two preferences pui and puj ψui(pui,o) = exp{w�u fu(pui,oi) + w�i fi(pui,ou))} (5.3.7) 108 pui puk puj pul i item attributes k item attributes l item attributes j item attributes u user attributes ψijψuj ψuj (a) Graph of user u pvi pvk pvj pvl i item attributes k item attributes l item attributes j item attributes v user attributes ψijψvj ψvj (b) Graph of user v Figure 5.2: Undirected graphs for users u and v 109 ψij(pui,puj) = exp{wijfij(pui,puj)} (5.3.8) where fu, fi, and fij are the features to be designed shortly in Section 5.3.5, and wu, wi, and wij are the corresponding weights realizing the importance of each feature. With the weights estimated from data, the unknown preference pui can be predicted as p̂ui = arg max pui∈[−1,1] P(pui | pu,o) (5.3.9) where P(pui | pu,o) measures the prediction confidence. 5.3.4 Ordinal Logistic Regression The original PrefNMF [19] computes the attractiveness of an item to a user as the product of their latent feature vectors which results a scalar value. Instead of point estimation, we wish to have the distributions over ordinal values. Therefore the Random Utility Models [55] and the Ordinal Logistic Regression [54] are utilized for the conversion. Random Utility Models [55] assume the existence of a latent utility xui = μui +�ui that captures how much the user u is interested in the item i, where μui captures the interest and �ui is the random noise, and here assumed to follow the logistic distribution [42]. The Ordinal Logistic Regression [54] is then used to convert the user-wise prefer- ences pui into ordinal values, which assumes that the preference pui is chosen based on the interval to which the latent utility belongs: pui = l if xui ∈ (θl−1,θl] and pui = L if xui > θL−1 (5.3.10) where L is the number of ordinal levels and θl are the threshold values of interest. 110 The probability of receiving a preference l is therefore: Q(pui = l | u,i) = ∫ θl θl−1 P(xui | θ) dθ = F(θl) − F(θl−1) (5.3.11) where F(θl) is the cumulative logistic distribution evaluated at θl with standard de- viation sui. F(xui ≤ l | θl) = 1 1 + exp(−θuil−μui sui ) (5.3.12) The thresholds θl can be user-specific or item-specific, and this work uses the user- specific parametrization same as in [42]. Then the thresholds θuil in Eq. 5.3.12 are replaced with a set of user-specific thresholds {θul}Ll=1 for each user u. These thresh- olds are then estimated from data for each user. 5.3.5 PrefCRF: Unifying PrefNMF and CRF The CRF provides a principled way of capturing both the side information and in- teractions among preferences. However, it employs the log-linear modeling as shown in Eq. 5.3.6, and therefore does not enable a simple treatment of PR. The PrefNMF, on the other hand, accepts PR but is weak in utilizing side information. The com- plementary between these two techniques calls for an unified PrefCRF model to take all the advantages. Unification Essentially, the proposed PrefCRF model captures the side information and promotes the agreement between the PrefNMF and the CRF. Specifically, the PrefCRF model combines the item-item correlations (Eq. 5.3.8) and the ordinal distributions Q(pui | 111 u,i) over user-wise preferences obtained from Eq. 5.3.11. P(pu | o) ∝ Ψu(pu,o) ∏ pui∈pu Q(pui | u,i) (5.3.13) where Ψu is the potential function capturing the side information and interaction among preferences related to user u. Though there is a separated graph for each user, the weights are optimized across all graphs. Feature Design A feature is essentially a function f of n > 1 arguments that maps the n-dimensional input into the unit interval f : Rn → [0,1]. We design the following kinds of features: Correlation Features The item-item correlation is captured by the feature fij(pui,puj) = g(|(pui − p̄i) − (puj − p̄j)|) (5.3.14) where g(α) normalizes feature values and α plays the role of deviation, and p̄i and p̄j are the average user-wise preference for items i and j, respectively. This correlation feature captures the intuition that correlated items should be ranked similarly by the same user after offsetting the goodness of each item. Attribute Features Each user u and item i has a set of attributes ou and oi, re- spectively. These attributes are mapped to preferences by the following features fi(pui) = oug(|(pui − p̄i)|) fu(pui) = oig(|(pui − p̄u)|) (5.3.15) where fi models which users like the item i and fu models which classes of items the user u likes. 112 Since one correlation feature exists for each pair of co-rated items, the number of correlation features can be large, and makes the estimation slow to converge and less robust. Therefore, we only keep strong correlation features fstrong extracted based on the Pearson correlation between items using a user-specified minimum correlation threshold. The correlations are computed using user-wise preferences generated from PR. Parameter Estimation In general, CRF models cannot be determined by standard maximum likelihood esti- mations, instead, approximation techniques are used in practice. This study employs the pseudo-likelihood [8] to estimate parameters by maximizing the regularized sum of log local likelihoods: logL(w) = ∑ pui∈Π log P(pui | pu,o) − 1 2σ2 w�w (5.3.16) where w are the weights and 1/2σ2 controls the regularization. To make the notation uncluttered, we write pu instead of explicitly as pu\pui. The local likelihood in Eq. 5.3.16 is defined as: P(pui|pu,o) = 1 Zui Q(pui|u,i)ψui(pui,o) ∏ puj∈pu ψij(pui,puj) (5.3.17) where Zui(o) is the normalization term: Zui = lmax∑ pui=lmin Q(pui | u,i)ψui(pui,o) ∏ puj∈pu ψij(pui,puj) (5.3.18) where lmin is the first and lmax is the last interval, i.e., 1 and 3 in our settings. 113 To optimize the parameters, we use the stochastic gradient ascent procedure that updates the parameters by passing through the set of ratings of each user: w ← w + η∇logL(w) (5.3.19) where η is the learning rate. More specifically, for each pui we update the attribute weights wo = {wu,wi} and correlation weight wij for each neighbor puj ∈ pu using the gradient of the log pseudo-likelihood ∂logL ∂wo = fo(pui,o) − lmax∑ pui=lmin P(pui | pu,o)fo(pui,o) − wi σ2 (5.3.20) ∂logL ∂wij = fij(pui,puj) − ∑lmax pui=lmin P(pui | pu,o)fij(pui,puj) − wijσ2 (5.3.21) Item Recommendation The ultimate goal of RecSys is often to rank the items and recommend the Top-N items to the user. As the input of PrefCRF is the preference relations, the final output will be item rankings instead of ratings. The PrefCRF produces distributions over the user-wise preferences, which can be converted into point estimates by computing the expectation p̂ui = lmax∑ pui=lmin puiP(pui | pu,o) (5.3.22) where l refers to the intervals of user-wise preferences: from the least to the most preferred. Given the predicted user-wise preferences, the items can be sorted and ranked accordingly. Alg. 3 summarizes the procedures of PrefCRF. 114 Algorithm 3 PrefCRF Algorithm Input: Heterogeneous preferences from sources. Step 1: Infer and merge PR. Step 2: Predict each user-wise preferences p̂ui using Eq. 5.3.3 and Eq. 5.2.2. Step 3: Predict distributions of p̂ui with Eq. 5.3.11. Step 4: Repeat for each u ∈ U do for each pui ∈ pu do Get local likelihood using Eq. 5.3.17 Get attribute feature fv using Eq. 5.3.15 Get attribute feature gradients using Eq. 5.3.20 Update wo with gradients using Eq. 5.3.19 for each puj ∈ pu, i �= j ∧ fij ∈ fstrong do Get correlation feature fij using Eq. 5.3.14 Get corr. feature gradient using Eq. 5.3.21 Update wij with gradient using Eq. 5.3.19 end for end for end for Until stopping criteria met Predictions: * Predict user-wise preferences using Eq. 5.3.22. * Sort and select the Top-N items. 115 Computational Complexity We perform the computational complexity analysis on the PrefCRF and its underly- ing PrefNMF algorithms. Given n users and m items each with du and di preferences, respectively. Let us temporarily ignore the user-specified latent factors. Then the complexity of both PrefNMF and PrefCRF is O(nd2u). However, in practice few item co-rated by the same user are strong neighbors of each other due to the correlation threshold defined in Section 5.3.5. As a result, the computation time of PrefCRF tends to be O(nduc) where c is a factor of correlation threshold. 5.4 Experiment and Analysis To study the performance of the proposed PrefCRF model, comparisons were done with the following representative algorithms: KNN [65], NMF [41], PrefKNN [12], and PrefNMF [19]. We implemented PrefCRF as well as the compared algorithms according to their original papers. 5.4.1 Experimental Settings Datasets and Experiment Design Experiments are conducted on four public datasets: MovieLens-1M 3, Amazon Movie Reviews 4, EachMovie 5, and MovieLens-20M 3. These datasets or their subsets are transformed to simulate four scenarios of heterogeneous data: 3http://grouplens.org/datasets/movielens 4http://snap.stanford.edu/data/web-Movies.html 5http://grouplens.org/datasets/eachmovie 116 Side Information The impact of side information is studied on the MovieLens-1M dataset which provides gender, age, and occupation information about users and genres of movies. The dataset contains more than 1 million ratings by 6,040 users on 3,900 movies. For a reliable comparison, the dataset is split into training and test sets with different sparsities, similar to related work [80,83]. Specifically, for each user we randomly select N = 30, 40, 50 and 60 ratings for training, and put the rest for testing. To ensure that each user has at least 10 movies for testing, users with less than 40, 50, 60 or 70 ratings are removed. Different Forms Amazon Movie Reviews dataset contains two forms of preferences: textual reviews and 5-star ratings. We extracted a dense subset by randomly selecting 5141 items with at least 60 reviews for each, and 2000 users with at least 60 movies reviews for each, and this results in 271K ratings. For each user, 50 random reviews are selected for training, and the rest are put aside for testing. The training set is further split into half ratings and half textual reviews. Rating-based models are trained on the ratings only, where PR-based models utilize textual reviews as well. Different Scales EachMovie dataset contains ratings in 6-star scale that can be easily converted into binary scale, i.e., ratings 1 − 3 and 4 − 6 are mapped to 0 and 1 respectively. We extract a subset by randomly selecting 3000 users who have rated at least 70 items as a dense dataset is required for splitting. The resultant dataset contains 120K ratings on 1495 items. For each user we randomly select 60 ratings for training and leave the rest for testing, and half of the ratings in the training set are mapped into binary scale. Rating-based models are trained on the 6-star ratings while PR-based models will exploit the 117 binary ratings as well. Different Biases We study the impact of biases by adding biases into a stable dataset with minimal existing biases. To prepare such dataset we extract a stable subset from the latest MovieLens-20M released on April-2015. Specif- ically, 258K ratings by 2020 users on 4408 movies released between 2010 and 2015 are extracted, where each user has rated at least 60 ratings. Biases are then introduced by adding a different Laplace noise sampled from Laplace(0,b) to each user and item. For PR-based methods, the same conversion method as in [19] is used to converted ratings into PR. For example, 1, 0 and 0.5 are assigned to the preference relation πuij when pui > puj, pui < puj, and pui = puj, respectively. Evaluation Metrics Traditional recommender systems aim to optimize RMSE or MAE which emphasizes on absolute ratings. However, the ultimate goal of recommender systems is usually to obtain the ranking of items [42], where good performance on RMSE or MAE may not be translated into good ranking results [42]. Therefore, we employ two evaluation metrics Normalized Cumulative Discounted Gain@T (NDCG@T) [34] that is popular in academia, and Mean Average Precision@T (MAP@T) [13] that is common in contests. The NDCG@T metric is defined as NDCG@T = 1 K(T) T∑ t=1 2rt − 1 log2 (t + 1) (5.4.1) 118 where rt is the relevance judgment of the item at position t, and K(T) is the normal- ization constant. The MAP@T metric is defined as MAP@T = 1 |Utest| ∑ u∈Utest T∑ t=1 Pu(t) min(mu, t) (5.4.2) where mu is the number of relevant items to user u, and Pu(t) is user u’s precision at position t. Both metrics are normalized to [0,1] with higher value indicates better performance. These metrics, together with other ranking-based metrics, require a set of relevant items to be defined in the test set such that the predicted rankings can be evaluated against. In this paper, we follow the same heuristics as in the related work [12,63] and consider items with the highest ratings as relevant. Parameter Setting For a fair comparison, we fix the number of latent factors to 50 for all algorithms, which is the same setting as used in [63].The number of neighbors for KNN algorithms is set to 50. We vary the minimum correlation threshold for the PrefCRF to examine the performance with different number of features. Different values of regularization coefficient are also tested. 5.4.2 Results and Analysis Algorithms are compared on four heterogeneous scenarios: side information, differ- ent forms, different scales and different biases. The impact of sparsity levels and parameters is studied on the MovieLens-1M dataset, while these settings for other experiments are fixed. Each experiment is repeated ten times with different random 119 seeds and we report the mean results with standard deviations. For each experiment, we also performed a paired t-test (two-tailed) with a significance level of 95% on the best and the second best results, and all p-values are less than 1 × 10−5. Fusing Side Information Table 5.1 shows the NDCG and MAP metrics on Top-N recommendation tasks by compared algorithms. It can be observed that the proposed PrefCRF, which captures the side information, consistently outperforms others. To confirm the improvement, we plot the results in Fig. 5.3(b) by varying the position T . The figure shows that PrefCRF not only outperforms others but has a strong emphasize on top items, i.e., T < 5. The impact of sparsity is investigated by plotting the results against sparsity levels as in Fig. 5.3(a). We can observe that the performance of PrefCRF increases linearly given more training data, while its underlying PrefNMF model is less extensible to denser dataset. One possible reason is that incorporating side information has extended the modeling capability of the model, resulting better utilization of more data. Fusing Preferences in Different Forms In this experiment, we first converted textual reviews into negative (−1), neutral (0), and positive (1) values using the NLTK library [10], and then converted them into PR. We study how these additional information can assist PR-based methods, and results over ten runs are shown in Table 5.2. Surprisingly, the performance of all PR-based methods except PrefCRF have decreased by incorporating textual reviews. 120 T a b le 5 .1 : R es u lt s o v er te n ru n s o n M o vi eL en s -1 M d a ta se t. G iv en 3 0 G iv en 4 0 A lg o ri th m N D C G @ 1 N D C G @ 1 0 M A P @ 1 M A P @ 1 0 N D C G @ 1 N D C G @ 1 0 M A P @ 1 M A P @ 1 0 U se rK N N 0 .4 3 0 6 ± 0 .0 0 1 1 0 .4 0 8 1 ± 0 .0 0 2 9 0 .3 5 3 9 ± 0 .0 0 7 1 0 .2 7 4 4 ± 0 .0 0 2 5 0 .3 6 9 5 ± 0 .0 0 4 8 0 .4 2 5 2 ± 0 .0 0 3 6 0 .3 6 6 3 ± 0 .0 0 4 7 0 .2 8 7 7 ± 0 .0 0 3 4 N M F 0 .5 2 7 4 ± 0 .0 0 8 4 0 .5 1 9 5 ± 0 .0 0 4 0 0 .5 2 2 5 ± 0 .0 0 8 1 0 .3 5 4 9 ± 0 .0 0 3 7 0 .5 4 2 4 ± 0 .0 0 6 7 0 .5 2 9 1 ± 0 .0 0 3 4 0 .5 3 7 7 ± 0 .0 0 6 6 0 .3 6 3 1 ± 0 .0 0 3 5 P re fK N N 0 .3 4 6 2 ± 0 .0 0 7 3 0 .4 0 4 8 ± 0 .0 0 3 8 0 .3 4 3 0 ± 0 .0 0 7 2 0 .2 7 2 0 ± 0 .0 0 3 7 0 .3 6 5 1 ± 0 .0 0 6 5 0 .4 2 8 3 ± 0 .0 0 2 4 0 .3 6 2 0 ± 0 .0 0 6 3 0 .2 9 0 4 ± 0 .0 0 2 3 P re fN M F 0 .5 7 7 8 ± 0 .0 1 1 2 0 .5 6 8 0 ± 0 .0 0 4 1 0 .5 7 2 4 ± 0 .0 1 0 9 0 .3 9 9 2 ± 0 .0 0 3 3 0 .5 8 8 3 ± 0 .0 0 7 3 0 .5 7 3 2 ± 0 .0 0 2 8 0 .5 8 3 2 ± 0 .0 0 7 3 0 .4 0 1 9 ± 0 .0 0 3 2 P re fC R F 0 .6 2 0 6 ± 0 .0 0 7 6 0 .5 8 5 6 ± 0 .0 0 2 8 0 .6 1 5 0 ± 0 .0 0 7 3 0 .4 1 9 5 ± 0 .0 0 2 8 0 .6 3 9 5 ± 0 .0 0 6 4 0 .5 9 9 0 ± 0 .0 0 2 3 0 .6 3 4 0 ± 0 .0 0 6 2 0 .4 2 9 4 ± 0 .0 0 2 1 G iv en 5 0 G iv en 6 0 A lg o ri th m N D C G @ 1 N D C G @ 1 0 M A P @ 1 M A P @ 1 0 N D C G @ 1 N D C G @ 1 0 M A P @ 1 M A P @ 1 0 U se rK N N 0 .3 8 3 1 ± 0 .0 0 6 3 0 .4 4 2 4 ± 0 .0 0 2 7 0 .3 8 0 3 ± 0 .0 0 6 0 0 .3 0 1 5 ± 0 .0 0 2 6 0 .4 0 3 5 ± 0 .0 0 9 0 0 .4 6 2 2 ± 0 .0 0 3 5 0 .4 0 0 2 ± 0 .0 0 8 5 0 .3 1 6 3 ± 0 .0 0 2 7 N M F 0 .5 4 3 0 ± 0 .0 0 8 3 0 .5 3 2 6 ± 0 .0 0 3 6 0 .5 3 9 0 ± 0 .0 0 8 2 0 .3 6 6 9 ± 0 .0 0 2 5 0 .5 5 4 7 ± 0 .0 1 0 9 0 .5 4 0 9 ± 0 .0 0 6 3 0 .5 5 0 4 ± 0 .0 1 1 3 0 .3 7 3 4 ± 0 .0 0 5 5 P re fK N N 0 .3 8 3 1 ± 0 .0 0 9 2 0 .4 4 8 3 ± 0 .0 0 3 0 0 .3 8 0 3 ± 0 .0 0 8 9 0 .3 0 7 0 ± 0 .0 0 2 2 0 .3 9 7 9 ± 0 .0 0 7 5 0 .4 6 8 9 ± 0 .0 0 3 9 0 .3 9 4 8 ± 0 .0 0 6 9 0 .3 2 2 3 ± 0 .0 0 3 3 P re fN M F 0 .5 8 7 3 ± 0 .0 0 9 6 0 .5 7 4 5 ± 0 .0 0 3 5 0 .5 8 3 0 ± 0 .0 0 9 8 0 .4 0 1 9 ± 0 .0 0 3 3 0 .5 8 5 4 ± 0 .0 1 4 5 0 .5 7 3 3 ± 0 .0 0 4 8 0 .5 8 0 8 ± 0 .0 1 4 2 0 .4 0 0 7 ± 0 .0 0 3 7 P re fC R F 0 .6 5 4 8 ± 0 .0 0 5 5 0 .6 0 6 8 ± 0 .0 0 1 8 0 .6 4 9 9 ± 0 .0 0 5 9 0 .4 3 7 2 ± 0 .0 0 2 4 0 .6 6 7 7 ± 0 .0 0 7 4 0 .6 1 3 9 ± 0 .0 0 1 8 0 .6 6 2 5 ± 0 .0 0 7 2 0 .4 4 3 6 ± 0 .0 0 1 6 121 (a) NDCG VS. Sparsity (b) NDCG@T (Given 60) Figure 5.3: Varying sparsity on MovieLens-1M dataset. We suspect that this is due to the misclassification errors introduced by sentiment classification on text. Indeed, converting preferences in different forms into PR can be accomplished in different ways, and may require domain knowledge fall beyond the scope of this paper. However, in the next subsection we will see that an accurate conversion can actually improve the performance. Table 5.2: Results over ten runs on Amazon dataset. Ratings Ratings + Textual Reviews Algorithm NDCG@10 MAP@10 NDCG@10 MAP@10 UserKNN 0.6244 ± 0.0040 0.4599 ± 0.0035 0.6244 ± 0.0037 0.4599 ± 0.0025 NMF 0.8073 ± 0.0040 0.6689 ± 0.0038 0.8073 ± 0.0041 0.6689 ± 0.0000 PrefKNN 0.6410 ± 0.0038 0.4690 ± 0.0029 0.5765 ± 0.0039 0.4083 ± 0.0029 PrefNMF 0.7495 ± 0.0040 0.5924 ± 0.0031 0.7377 ± 0.0030 0.5806 ± 0.0031 PrefCRF 0.8223 ± 0.0033 0.6813 ± 0.0027 0.8259 ± 0.0035 0.6890 ± 0.0026 122 Fusing Preferences in Different Scales In this experiment preferences in different scales are fused into PR to boost the performance of PR-based methods. The binary scale ratings are similar to the pos- itive/negative textual reviews, however without incorrect values introduced by text classification. Table 5.3: Results over ten runs on EachMovie dataset. 6-star Ratings 6-star Ratings + Binary Ratings Algorithm NDCG@10 MAP@10 NDCG@10 MAP@10 UserKNN 0.4374 ± 0.0047 0.3418 ± 0.0029 0.4374 ± 0.0047 0.3418 ± 0.0029 NMF 0.5211 ± 0.0078 0.3710 ± 0.0034 0.5211 ± 0.0078 0.3710 ± 0.0034 PrefKNN 0.4908 ± 0.0070 0.3793 ± 0.0031 0.5074 ± 0.0061 0.3938 ± 0.0044 PrefNMF 0.5233 ± 0.0061 0.3820 ± 0.0033 0.5454 ± 0.0060 0.3881 ± 0.0032 PrefCRF 0.5439 ± 0.0056 0.4006 ± 0.0045 0.5506 ± 0.0053 0.4038 ± 0.0043 Table 5.3 shows the performance of each method and the impact of position T is illustrated in Fig. 5.4. From the table, we can observe that the performance of all PR-based methods has increased by incorporating additional binary ratings, while the performance of rating-based methods remains the same. Fusing Preferences with Different Biases In this experiment we investigate the impact of different biases, particularly the user- wise and item-wise biases, which are sampled from Laplace(0,b) for each user and each item. Results for unbiased and biased datasets are reported in Table 5.4. For user-wise biases, we can see that the performance of rating-based methods 123 (a) NDCG@T (b) MAP@T Figure 5.4: Varying position T on EachMovie dataset. Table 5.4: NDCG@10 on MovieLens-20M dataset. Algorithm Bias = None User-bias = Laplace(0,2) Item-bias = Laplace(0,2) UserKNN 0.4465 ± 0.0033 0.3729 ± 0.0033 0.2914 ± 0.0017 NMF 0.4982 ± 0.0034 0.4566 ± 0.0032 0.3074 ± 0.0019 PrefKNN 0.4683 ± 0.0027 0.4683 ± 0.0027 0.3157 ± 0.0021 PrefNMF 0.4950 ± 0.0035 0.4950 ± 0.0035 0.3137 ± 0.0017 PrefCRF 0.5288 ± 0.0037 0.5288 ± 0.0037 0.3729 ± 0.0023 124 has decreased while PR-based methods are unaffected by such biases. This is further illustrated in Fig. 5.5(a). For item-wise biases, the performance of all methods has decreased, however, to different extent. Fig. 5.5(b) further shows the better resistance of PrefCRF to item-wise biases. (a) NDCG@10 (b) NDCG@10 Figure 5.5: Varying biases on MovieLens-20M dataset. Impact of Regularization and Correlation Threshold The proposed PrefCRF method has two user specified parameters: the regularization coefficient and a minimum correlation threshold that controls the number of corre- lation features. We examine the impact of these parameters on the MovieLens-1M dataset and the results are plotted in Fig. 5.6. For the regularization, we can see from Fig. 5.6(a) that the performance gets better when a small regularization penalty applies. In other words, PrefCRF can generalize reasonable well without too much regularization. One reason is that the model has already been regularized by its underlying PrefNMF model. Another reason is that 125 the weights of item-item correlations are not user-specific but shared by all users, thus they cannot over-fit every user perfectly. For the correlation threshold, Fig. 5.6(b) shows that a smaller threshold results better performance by including more correlation features, however, at the cost of more training time and more training data. (a) NDCG@10 (b) NDCG@10 Figure 5.6: Varying parameters on MovieLens-1M. 5.5 Summary In this chapter we proposed the PrefCRF model, which takes advantages of both the representational power of the CRF and the ease of modeling PR by the PrefNMF. To the best of our knowledge, this is the first study on unifying heterogeneous user preferences and heterogeneous side information. Experiment results on four public datasets demonstrate that various heterogeneous data have been properly handled by 126 PrefCRF, and significantly improved Top-N recommendation performance has been achieved. For future work, the computation efficiency of PR-based methods can be further improved given that the number of PR is usually much larger than ratings. Paral- lelization is feasible as as each user has a separated set of PR that can be processed simultaneously. Chapter 6 Conclusion and Future Work The research presented in this thesis consists of two parts: the first part focuses on the local and global structure and the side information issues and while the second part focuses on the heterogeneous data source issue. Several new research problems have been identified together with solutions. This chapter summarizes the research results and the main contributions of this thesis. 6.1 Contributions Theoretical and experimental results have led to the conclusion and main contribution of this thesis: • The Ordinal Random Fields (ORF) model is proposed to capture both the local and global structures of ordinal user preferences. Through this novel method, both structures are properly captured, resulting improved recommen- dation quality. ORF is one of first attempts on modeling both structures for ordinal user preferences. 127 128 • The Preference Relation-based Markov Random Fields (PrefMRF) model is pro- posed to capture both the local and global structures of Preference Relations. Due to the flexibility of previous preference relation-based models, only one type of structure can be modeled at a time. Through our novel method, both structures can be modeled, resulting significantly improved recommendation performance. PrefMRF is the first preference relation-based model that captures both types of structures. • The Preference Relation-based Conditional Random Fields (PrefCRF) model is proposed to incorporate side information such as item content and user profiles. PrefCRF is the first preference relation-based model that can incorporate side information in a principled way. • The preference relation-based models proposed in this thesis is applied to resolve the heterogeneous data sources problem. This is the first time this problem is spotted and we provide effective solutions to unify heterogeneous data. By exploiting multiple heterogeneous data sources help to alleviate the cold-start problem in which limited data is provided by each data source. 6.2 Future Work Although the proposed methods have addressed the aforementioned three issues to a certain extend, there remains several problems that need to be addressed in the future. We summarize these follows: • Parallelization of Preference Relation-based Models: The number of preference 129 relations are usually much larger than the number of traditional ratings. As a re- sult, the modeling process is often slower than rating-based methods. However, we observe that the preference relations of each user are kind of independent from other users’. Therefore, it it possible to parallelize the modeling process for each user to achieve faster modeling process. • Identifying Key Preference Relations: While there are so many possible prefer- ence relations, not all of them are important in making recommendations. It remains unknown how to measure the importance of each preference relation to the recommendation quality. If this information can be obtained in future work, then the number of preference relations to be used in modeling can be significantly reduced while the recommendation quality is preserved. Bibliography [1] G. Adomavicius and A. Tuzhilin. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Trans- actions on Knowledge and Data Engineering, 17(6):734–749, 2005. [2] G. Adomavicius and A. Tuzhilin. Context-aware recommender systems. In Rec- ommender systems handbook, pages 217–253. Springer, 2011. [3] G. Adomavicius and J. Zhang. On the stability of recommendation algorithms. In Proceedings of the fourth ACM conference on Recommender systems, pages 47–54. ACM, 2010. [4] G. Adomavicius and J. Zhang. Stability of recommendation algorithms. ACM Transactions on Information Systems (TOIS), 30(4):23, 2012. [5] M. Balabanović and Y. Shoham. Fab: content-based, collaborative recommen- dation. Commun. ACM, 40(3):66–72, 1997. [6] J. Basilico and T. Hofmann. Unifying collaborative and content-based filtering. In ICML’04, pages 65–72. ACM, 2004. [7] J. Bennett and S. Lanning. The netflix prize. In Proceedings of KDD cup and workshop, volume 2007, page 35, 2007. [8] J. Besag. Spatial interaction and the statistical analysis of lattice systems. Jour- nal of the Royal Statistical Society, Series B, 36(2):192–236, 1974. 130 131 [9] D. Billsus and M. J. Pazzani. Learning collaborative information filters. In ICML, volume 98, pages 46–54, 1998. [10] S. Bird, E. Klein, and E. Loper. Natural language processing with Python. O’Reilly Media, Inc., 2009. [11] K. Bradley and B. Smyth. Improving recommendation diversity. In Proceedings of the Twelfth National Conference in Artificial Intelligence and Cognitive Science (AICS-01), pages 75–84. Citeseer, 2001. [12] A. Brun, A. Hamad, O. Buffet, and A. Boyer. Towards preference relations in recommender systems. In Preference Learning (PL 2010) ECML/PKDD, 2010. [13] O. Chapelle, D. Metlzer, Y. Zhang, and P. Grinspan. Expected reciprocal rank for graded relevance. In CIKM’09, pages 621–630. ACM, 2009. [14] W. W. Cohen, R. E. Schapire, and Y. Singer. Learning to order things. J. Artif. Int. Res., 10(1):243–270, May 1999. [15] P. Cremonesi, Y. Koren, and R. Turrin. Performance of recommender algorithms on top-n recommendation tasks. In RecSys’10, pages 39–46. ACM, 2010. [16] A. Defazio and T. Caetano. A graphical model formulation of collaborative filter- ing neighbourhood methods with fast maximum entropy training. In ICML’12, pages 265–272, 2012. [17] M. S. Desarkar and S. Sarkar. Rating prediction using preference relations based matrix factorization. In UMAP Workshops, 2012. [18] M. S. Desarkar, S. Sarkar, and P. Mitra. Aggregating preference graphs for collaborative rating prediction. In RecSys’10, pages 21–28. ACM, 2010. 132 [19] M. S. Desarkar, R. Saxena, and S. Sarkar. Preference relation based matrix factorization for recommender systems. In UMAP’12, pages 63–75. Springer, 2012. [20] D. Fleder and K. Hosanagar. Blockbuster culture’s next rise or fall: The impact of recommender systems on sales diversity. Management science, 55(5):697–712, 2009. [21] Y. Freund, R. Iyer, R. E. Schapire, and Y. Singer. An efficient boosting algorithm for combining preferences. Journal of machine learning research, 4:933–969, 2003. [22] J. Fürnkranz and E. Hüllermeier. Pairwise preference learning and ranking. In Machine Learning: ECML’03, pages 145–156. Springer, 2003. [23] J. Fürnkranz and E. Hüllermeier. Preference learning. Springer, 2010. [24] Z. Gantner, L. Drumond, C. Freudenthaler, S. Rendle, and L. Schmidt-Thieme. Learning attribute-to-feature mappings for cold-start recommendations. In 2010 IEEE International Conference on Data Mining, pages 176–185. IEEE, 2010. [25] M. Ge, C. Delgado-Battenfeld, and D. Jannach. Beyond accuracy: evaluating recommender systems by coverage and serendipity. In Proceedings of the fourth ACM conference on Recommender systems, pages 257–260. ACM, 2010. [26] P. J. Green. Reversible jump markov chain monte carlo computation and bayesian model determination. Biometrika, 82(4):711–732, 1995. [27] I. Guy, I. Ronen, and A. Raviv. Personalized activity streams: sifting through the river of news. In Proceedings of the fifth ACM conference on Recommender systems, pages 181–188. ACM, 2011. [28] I. Guy, N. Zwerdling, D. Carmel, I. Ronen, E. Uziel, S. Yogev, and S. Ofek- Koifman. Personalized recommendation of social software items based on social 133 relations. In Proceedings of the third ACM conference on Recommender systems, pages 53–60. ACM, 2009. [29] J. Han, M. Kamber, and J. Pei. Data mining: concepts and techniques. Morgan kaufmann, 2006. [30] G. Hinton. Training products of experts by minimizing contrastive divergence. Neural Computation, 14(8):1771–1800, 2002. [31] T. Hofmann and J. Puzicha. Latent class models for collaborative filtering. In IJCAI, volume 99, pages 688–693, 1999. [32] N. Houlsby, J. M. Hernandez-lobato, and Z. Ghahramani. Cold-start active learning with robust ordinal matrix factorization. In ICML’14, pages 766–774, 2014. [33] E. Hüllermeier, J. Fürnkranz, W. Cheng, and K. Brinker. Label ranking by learning pairwise preferences. Artificial Intelligence, 172(16):1897–1916, 2008. [34] K. Järvelin and J. Kekäläinen. Cumulated gain-based evaluation of ir techniques. ACM TOIS, 20(4):422–446, 2002. [35] R. Jin, L. Si, and C. Zhai. Preference-based graphic models for collaborative filtering. In UAI’02, pages 329–336. Morgan Kaufmann Publishers Inc., 2002. [36] N. Jones, A. Brun, and A. Boyer. Comparisons instead of ratings: Towards more stable preferences. In Web Intelligence and Intelligent Agent Technology (WI-IAT), 2011 IEEE/WIC/ACM International Conference on, volume 1, pages 451–456. IEEE, 2011. [37] B. P. Knijnenburg, M. C. Willemsen, Z. Gantner, H. Soncu, and C. Newell. Explaining the user experience of recommender systems. User Modeling and User-Adapted Interaction, 22(4-5):441–504, 2012. 134 [38] Y. Koren. Factorization meets the neighborhood: a multifaceted collaborative filtering model. In KDD’08, pages 426–434. ACM, 2008. [39] Y. Koren. Collaborative filtering with temporal dynamics. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 447–456. ACM, 2009. [40] Y. Koren. Collaborative filtering with temporal dynamics. Commun. ACM, 53(4):89–97, 2010. [41] Y. Koren, R. Bell, and C. Volinsky. Matrix factorization techniques for recom- mender systems. IEEE Computer, 42(8):30–37, 2009. [42] Y. Koren and J. Sill. Ordrec: an ordinal model for predicting personalized item rating distributions. In RecSys’11, pages 117–124. ACM, 2011. [43] J. D. Lafferty, A. McCallum, and F. Pereira. Conditional random fields: prob- abilistic models for segmenting and labeling sequence data. In ICML’01, pages 282–289, 2001. [44] X. Li, G. Xu, E. Chen, and Y. Zong. Learning recency based comparative choice towards point-of-interest recommendation. Expert Systems with Applications, 42(9):4274–4283, 2015. [45] N. N. Liu, M. Zhao, and Q. Yang. Probabilistic latent preference analysis for collaborative filtering. In CIKM’09, pages 759–766. ACM, 2009. [46] S. Liu, T. Tran, G. Li, and Y. Jiang. Ordinal random fields for recommender systems. In ACML’14, pages 283–298. JMLR: workshop and conference proceed- ings, 2014. [47] P. Lops, M. de Gemmis, and G. Semeraro. Content-based recommender systems: State of the art and trends. In Recommender Systems Handbook, pages 73–105. Springer, 2011. 135 [48] L. Lü and W. Liu. Information filtering via preferential diffusion. Physical Review E, 83(6):066119, 2011. [49] L. Lü, M. Medo, C. H. Yeung, Y.-C. Zhang, Z.-K. Zhang, and T. Zhou. Recom- mender systems. Physics Reports, 519(1):1–49, 2012. [50] H. Ma, D. Zhou, C. Liu, M. R. Lyu, and I. King. Recommender systems with social regularization. In WSDM’11, pages 287–296. ACM, 2011. [51] R. D. Mare. Social background and school continuation decisions. Journal of the American Statistical Association, 75(370):295–305, 1980. [52] B. M. Marlin. Modeling user rating profiles for collaborative filtering. In NIPS’03. MIT Press, 2003. [53] P. Massa and P. Avesani. Trust-aware collaborative filtering for recommender systems. In On the Move to Meaningful Internet Systems 2004: CoopIS, DOA, and ODBASE, pages 492–508. Springer, 2004. [54] P. McCullagh. Regression models for ordinal data. Journal of the Royal Statistical Society, Series B, 42(2):109–142, 1980. [55] D. McFadden. Econometric models for probabilistic choice among products. Journal of Business, 53(3):S13–S29, 1980. [56] S. M. McNee, J. Riedl, and J. A. Konstan. Being accurate is not enough: how accuracy metrics have hurt recommender systems. In CHI EA’06, pages 1097– 1101. ACM, 2006. [57] K. Miyahara and M. J. Pazzani. Collaborative filtering with the simple bayesian classifier. In PRICAI 2000 Topics in Artificial Intelligence, pages 679–689. Springer, 2000. 136 [58] M. Mohri, A. Rostamizadeh, and A. Talwalkar. Foundations of machine learning. MIT press, 2012. [59] A. Nakamura and N. Abe. Collaborative filtering using weighted majority pre- diction algorithms. In ICML, volume 98, pages 395–403, 1998. [60] S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345–1359, 2010. [61] U. Paquet, B. Thomson, and O. Winther. A hierarchical model for ordinal matrix factorization. Statistics and Computing, 22(4):945–957, 2012. [62] M. J. Pazzani and D. Billsus. Content-based recommendation systems. In The adaptive web, pages 325–341. Springer, 2007. [63] Y. Ren, G. Li, and W. Zhou. Learning rating patterns for top-n recommenda- tions. In ASONAM’12, pages 472–479. IEEE Computer Society, 2012. [64] S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme. Bpr: Bayesian personalized ranking from implicit feedback. In Proceedings of the twenty-fifth conference on uncertainty in artificial intelligence, pages 452–461. AUAI Press, 2009. [65] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. Riedl. An open ar- chitecture for collaborative filtering of netnews. In CSCW’94, pages 175–186. ACM, 1994. [66] F. Ricci, L. Rokach, and B. Shapira. Introduction to recommender systems hand- book. Springer, 2011. [67] G. Salton. Automatic Text Processing: The Transformation, Analysis, and Re- trieval of. Addison-Wesley, 1989. 137 [68] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl. Application of dimension- ality reduction in recommender system-a case study. Technical report, DTIC Document, 2000. [69] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl. Item-based collaborative fil- tering recommendation algorithms. In WWW’10, pages 285–295. ACM, 2001. [70] B. M. Sarwar, G. Karypis, J. Konstan, and J. Riedl. Recommender systems for large-scale e-commerce: Scalable neighborhood formation using clustering. In Proceedings of the fifth international conference on computer and information technology, volume 1. Citeseer, 2002. [71] D. L. Schacter and C. S. Dodson. Misattribution, false recognition and the sins of memory. Philosophical Transactions of the Royal Society B: Biological Sciences, 356(1413):1385–1393, 2001. [72] A. I. Schein, A. Popescul, L. H. Ungar, and D. M. Pennock. Methods and metrics for cold-start recommendations. In SIGIR’02, pages 253–260. ACM, 2002. [73] A. Sharma and B. Yan. Pairwise learning in recommendation: experiments with community recommendation on linkedin. In RecSys’13, pages 193–200. ACM, 2013. [74] Y. Shi, M. Larson, and A. Hanjalic. List-wise learning to rank with matrix factorization for collaborative filtering. In RecSys’10, pages 269–272. ACM, 2010. [75] X. Su, R. Greiner, T. M. Khoshgoftaar, and X. Zhu. Hybrid collaborative filtering algorithms using a mixture of experts. In WI’07, pages 645–649. IEEE, 2007. [76] G. Takács, I. Pilászy, B. Németh, and D. Tikk. Scalable collaborative filtering approaches for large recommender systems. The Journal of Machine Learning Research, 10:623–656, 2009. 138 [77] T. Tran, D. Phung, and S. Venkatesh. Collaborative filtering via sparse markov random fields. arXiv preprint arXiv:1602.02842, 2016. [78] T. Tran, D. Q. Phung, and S. Venkatesh. Preference networks: Probabilistic models for recommendation systems. In AusDM’07, pages 195–202. ACS, 2007. [79] T. Tran, D. Q. Phung, and S. Venkatesh. Ordinal boltzmann machines for collaborative filtering. In UAI’09, pages 548–556. AUAI Press, 2009. [80] T. Tran, D. Q. Phung, and S. Venkatesh. Probabilistic models over ordered partitions with applications in document ranking and collaborative filtering. In SDM’11, pages 426–437. SIAM, 2011. [81] T. Tran, D. Q. Phung, and S. Venkatesh. Learning from ordered sets and appli- cations in collaborative ranking. In ACML’12, pages 427–442. JMLR: workshop and conference proceedings, 2012. [82] T. Tran, D. Q. Phung, and S. Venkatesh. A sequential decision approach to ordinal preferences in recommender systems. In AAAI’12, 2012. [83] M. Weimer, A. Karatzoglou, Q. V. Le, and A. Smola. Maximum margin matrix factorization for collaborative ranking. In NIPS’07, pages 1593–1600, 2007. [84] T. Zhou, L.-L. Jiang, R.-Q. Su, and Y.-C. Zhang. Effect of initial configuration on network-based recommendation. EPL (Europhysics Letters), 81(5):58004, 2008. [85] T. Zhou, J. Ren, M. Medo, and Y.-C. Zhang. Bipartite network projection and personal recommendation. Physical Review E, 76(4):046115, 2007. [86] A. Zimdars, D. M. Chickering, and C. Meek. Using temporal data for making recommendations. In Proceedings of the Seventeenth conference on Uncertainty in artificial intelligence, pages 580–588. Morgan Kaufmann Publishers Inc., 2001. 
work_jmwy2qts4bh4tn2bnpbqflcqou	----	Risk assessment model selection in construction industry Expert Systems with Applications 38 (2011) 9105–9111 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Risk assessment model selection in construction industry AmirReza KarimiAzari a, Neda Mousavi b,⇑, S. Farid Mousavi c, SeyedBagher Hosseini a a Iran University of Science and Technology, Tehran, Iran b Science & Research Branch, Islamic Azad University, Tehran, Iran c Tarbiyat Modares University, Tehran, Iran a r t i c l e i n f o Keywords: Risk management Risk assessment model Construction industry Fuzzy TOPSIS 0957-4174/$ - see front matter � 2010 Elsevier Ltd. A doi:10.1016/j.eswa.2010.12.110 ⇑ Corresponding author. E-mail addresses: nmousavi2930@gmail.com (N. com (S.F. Mousavi). a b s t r a c t Construction industry faces a lot of inherent uncertainties and issues. As this industry is plagued by risk, risk management is an important part of the decision-making process of these companies. Risk assessment is the critical procedure of risk management. Despite many scholars and practitioners recognizing the risk assessment models in projects, insufficient attention has been paid by researchers to select the suitable risk assessment model. In general, many factors affect this problem which adheres to uncertain and imprecise data and usually several people are involved in the selection process. Using the fuzzy TOPSIS method, this study provides a rational and systematic process for developing the best model under each of the selection criteria. Decision criteria are obtained from the nominal group technique (NGT). The proposed method can discriminate successfully and clearly among risk assessment methods. The proposed approach is demonstrated using a real case involving an Iranian construction corporation. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction The research on projects has expanded during the last dec- ades (Naaranoja, Haapalainen, & Lonka, 2007). A project is an organization of people dedicated to the deployment of a set of resources for a specific purpose or objective (Steiner, 1969). Pro- ject management is defined as planning, directing, and control- ling resources to achieve specific goals and objectives of the project (Fan, Lin, & Sheu, 2008). Managers need to ensure deliv- ery of projects to cost, schedule and performance requirement. To achieve this involves identifying and managing the risks to the project at all project stages from the initial assessment of strategic options through the procurement, fabrication, construc- tion and commissioning stage (Tah & Carr, 2001). The less ‘‘pre- dictable’’ nature of projects makes them riskier than day to day business activities (Elkington & Smallman, 2002). Risk is a possi- ble undesirable and unplanned event that could result in the project not meeting one or more of its objectives (Teneyuca, 2001). As the underlying concept of risk management is to man- age risks effectively, risk management is a critical part of project management (Lyons & Skitmore, 2004). Construction industries, face a lot of inherent uncertainties and issues like company’s fluctuating profit margin, competitive bid- ding process, weather change, productivity on site, the political sit- uation in a country, inflation, contractual rights, market ll rights reserved. Mousavi), mousavifd@gmail. competition, etc. Thus the construction industry, more than others, has been plagued by risk (Carr & Tah, 2001) and there is no con- struction project with risk free (Lam, Wang, Lee, & Tsang, 2007). With the rapid advancement in the construction industry, an in- creased number of uncertainties are bound to occur (Thevendran & Mawdesley, 2004). It is essential that the construction companies conquer these risks and uncertainties in order to assess the effect of these sources in order to decide which of the projects is more risky, plan for the potential sources of risk in each project and man- age each source during construction (Zayed, Amer, & Pan, 2008). Therefore it is paramount for construction companies to be sensi- tive to the issue of embracing and managing uncertainty and risk discussed above. Project related risk management has attracted steady stream of interest in the academic literature (Bannerman, 2008). One of the major steps in project risk management is to identify and assess the potential risks (El-Sayegh, 2008). Despite many scholars and practitioners recognizing the risk identification methods and assessment models in projects insufficient attention has been paid by researchers to select a suitable risk assessment model. This pa- per attempts to address this limitation and the gap in the current literature and provide a framework for determining optimal risk assessment model. In Section 2, some relevant literature is described. In Section 3, the problem of the risk assessment model selection is intro- duced. Section 4 concentrates on the proposed model. A real case study is presented to illustrate the application of the pro- posed method in Section 5. In the final section some conclusions are drawn. http://dx.doi.org/10.1016/j.eswa.2010.12.110 mailto:nmousavi2930@gmail.com mailto:mousavifd@gmail. com mailto:mousavifd@gmail. com http://dx.doi.org/10.1016/j.eswa.2010.12.110 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 9106 A. KarimiAzari et al. / Expert Systems with Applications 38 (2011) 9105–9111 2. Literature review 2.1. Risk The concept of risk became popular in economics during the 1920s. Since then, it has been successfully used in theories of deci- sion making in economics, finance, and the decision science (Ngai & Wat, 2005). Risk has different meaning to different people; that is, the concept of risk varies according to viewpoint, attitudes and experience. Engineers, designers and contractors view risk from the technological perspective; lenders and developers tend to view it from the economic and financial side (Baloi & Price, 2003). The traditional view of risk is negative, representing loss, haz- ard, harm and adverse consequences. But some current risk guide- lines and standards include the possibility of upside risk or opportunity, i.e. uncertainties that could have a beneficial effect on achieving objectives (Hillson, 2002). Project risk is defined by Project Management Body of Knowledge (PMBOK) published by the Project Management Institute (PMI) as an uncertain event or condition that, if it occurs, has a positive or a time, cost, span or quality, which implies an uncertainty about identified events and conditions. PMBOK describes risk through the notion of uncer- tainty; however, these two phenomena are not synonymous (Perminova, Gustafsson, & Wikstrom, 2008). According to the Olsson (2007) and Hillson (2004) attempts to link risk with uncer- tainty based on the distinction between aleatory and epistemic uncertainty in the following couplet: � Risk is measurable uncertainty. � Uncertainty is immeasurable risk. This implies that, when measurable, an uncertainty is to be con- sidered a risk. PMBOK’s definition of risk and uncertainty is the considered definition through the entire paper because this defini- tion implies that risk is quantifiable and lends itself to assessment. 2.2. Risk management If a risk is not identified it cannot be controlled, transferred or otherwise managed (Bajaj, 1997) and trying to eliminate all risks in projects is impossible. Thus, there is need for a formal risk man- agement process to manage all types of risks. The project success usually depends on the combination of all risks, response strategies used to mitigate risks and a company’s ability to manage those (Dikmen, Birgonul, & Han, 2007). Hence, the underlying concept of risk management is to manage risks effectively (Thevendran & Mawdesley, 2004). Risk management can lead to a range of project and organizational benefits including: (Bannerman, 2008) � Identification of favorable alternative courses of action. � Increased confidence in achieving project objectives. � Improved chances of success. � Reduced surprises. � More precise estimates (through reduced uncertainty). � Reduced duplication of effort (through team awareness of risk control actions). PMBOK included risk management as one of the nine focuses in project management and described it as the process concerned with conducting risk management planning, identification, analy- sis, responses, and monitoring and control on a project (Zou, Zhang, & Wang, 2007). Risk management in construction is a tedious task as the objective functions tend to change during the project life cycle, and the scenarios are numerous due to sensitivity of projects to uncontrollable risks stemming from the changes in the macro-environment, existence of high number of parties involved in the project value chain, and one-off nature of the construction process (Dikmen, Birgonul, Anac, Tah, & Aouad, 2008). Project risk management is an integrated process which in- cludes activities to identify project uncertainty, estimate their im- pact, analyze their interactions, control them in the execution stage, and even provide feedback to the maintenance of collective knowledge asset (Williams, 1995). Risk management based on con- sensus in the literatures, used the following three-step approach (Zayed et al., 2008): � Risk identification. � Risk assessment. � Risk mitigation. The first step in risk management is risk identification. Before risks can be managed, they must be identified. Identification sur- faces risks before they become problems and adversely affect a project. It refers to the evidences from previous experience or sim- ilar cases which would apply to the current project, in order to avoid or ameliorate the probability of compromising the project’s success. Construction risks can be categorized in a number of ways based on the source of risk, impact of risk or by project phase (Klemetti, 2006). In the most reference one, project risks are di- vided into two groups, according to their source, into internal and external. Internal risks are initiated inside the project while external risks originate due to the project environment (El-Sayegh, 2008). In risk identification step all internal and external risks must be identified. After the establishment of a list of risk events that had actually occurred in the process of project performance, these risks must be assessed. The primary objective of risk assessment is to estimate risk by identifying the undesired event, the likelihood of occurrence of the unwanted event, and the consequence of such event. Risk assessment involves measures, either conducted quantitatively or qualitatively, to produce the estimation of the significance level of the individual risk factors to the project, so as to produce the estimation of the risk of the potential factors to project success. However, this step results will become the input to the determina- tion of the optimum decision. With a better quantification measur- ing result, the managers can recognize which risks are more important and then deploy more resources on it to eliminate or mitigate the expected consequences. The identification and assessment of project risk are the critical procedures for projecting success, and they usually become the essential factors in the decision-making process (Williams, 1995). Most authors refer to the processes which include risk identifica- tion and assessment, as the stage called ‘‘risk analysis’’. Risk anal- ysis can provide insight to the specific sources of project risk and enable management to devise targeted remedial action. Several methods have been proposed and utilized thorough re- search by a lot of scholars to help contractors and subcontractors to evaluate and select the best projects in order to decide which pro- jects are more risky. And so these models help to plan for the po- tential sources of risk in each project and manage each source during construction. Currently project management teams have more options from which to choose. Risk assessment methods have ranged from simple classical methods to fuzzy approach mathematical models. Many construc- tion project risk assessment techniques currently used are compar- atively mature tools (Zeng, An, & Smith, 2007). Monte Carlo Simulation (White, 1995), Sensitivity Analysis (White, 1995), Critical path method (Kaufmann & Gupta, 1988), Fault tree analysis (Terano, Asai, & Sugeno, 1992), Event tree analysis (Huang, Chen, & Wang, 2001), Failure mode, effects and A. KarimiAzari et al. / Expert Systems with Applications 38 (2011) 9105–9111 9107 criticality analysis (Bowles & Pelaez, 1995) are the classical quan- titative methods, used in construction industry for risk assessment. These methods only use data that are quantitative so, for effective application of these sophisticated quantitative techniques high quality data are a prerequisite (Zeng et al., 2007). Only on a few projects and contracts are risk considered in a consistent and log- ical manner; much assessment is too subjective (Mills, 2001). So, some other models suggested, involve both quantitative and qual- itative ones. Fuzzy risk assessment methods have also been deployed with some scholars too. Mustafa and Al-Bahar (1991) investigated the subject of risk assessment and developed a scheme of classifying the various sources of risk in construction projects. They applied Analytical Hierarchical Process (AHP) in assessing the riskiness of a real-life constructing project (Mohammad & Al-Bahar, 1991). Sadiq and Husain (2005) developed a three-stage hierarchical structure aggregative risk model for grouping of risk items. For this grouping, an analytical hierarchy process was used for assessment. Another hierarchical risk breakdown structure is described to represent a formal model for qualitative risk assessment by Carr and Tah (2001). In their paper, using fuzzy approximation and composition, the relationships between risk sources and the consequences on project performance measures were identified and quantified. Cho, Choi, and Kim (2002) proposes another methodology for incor- porating uncertainties using fuzzy concepts into conventional risk assessment frameworks in construction industry. Choi, Cho, and Seo (2004) presents fuzzy risk assessment methodology for under- ground construction projects. A formalized procedure and associ- ated tools were developed to assess and manage the risks involved in underground construction. The suggested risk assess- ment procedure is composed of four steps of identifying, analyzing, evaluating, and managing the risks inherent in construction pro- jects. Zeng et al. (2007) developed a methodology to deal with risks associated with the construction projects in the complicated situa- tions. This model can handle with the expert knowledge, engineer- ing judgment and the historical data for risk assessment and in this model the risk can be evaluated directly using linguistic terms which are employed in risk assessment. Zayed et al. (2008) intro- duces a model, based on AHP to help practitioners to assess Chinese highway risk projects and prioritize them. This methodology quan- tifies the qualitative effect of subjective factors of risk. These methods differ in a variety of ways and they have their own advantages and disadvantages. So an ideal risk assessment method which would suit all organizations does not exist, as each of the organizations and projects possesses its own unique charac- teristics (Lichtenstein, 1996), so, an organization and project man- agement team need to select the most appropriate methodology on its specific. This problem labeled as the risk assessment model selection. 3. Risk assessment model According to Lichtenstein (1996) selecting which model is suit- able for the organization or project is affected by many factors. The cost of employing the technique, the level of external party’s approval, Organizational structure, Agreement, Adaptabil- ity, Complexity, Completeness, Level of risk, Organizational size, Organizational security philosophy, Consistency, Usability, Feasi- bility, Validity, Credibility and Automation are factors to be consid- ered in the selection of a risk assessment method (Lichtenstein, 1996). Owning to several quantitative and qualitative factors, some of which them may be in conflict with the others, the risk assessment technique selection is complicated. Risk assessment model involves two other key modeling as- pects: First, construction is described as a collaborative teamwork process where parties with different interests, functions, and objectives, share a common goal, which is successful completion of a project. Thus in this problem solving it is vital to involve sev- eral people from different parties. A second important consideration of the risk assessment model selection is that much knowledge in the real world is imprecise than precise (Olcer & Odabasi, 2005), thus the preference informa- tion provided to model selection may be imprecise or incomplete. As a result, multiple factors, which are either quantitative or qualitative and may be in conflict with each other, impact the risk assessment model selection problem and problem arises in group setting with incomplete, vague and uncertain information. In line with the multidimensional characteristics of the risk assessment model selection, the problem is a kind of multi-criteria decision-making (MCDM) problem, which requires MCDM meth- ods for an effective problem solving. MCDM refers to screening, prioritizing, ranking or selecting a set of alternatives under usually independent, incommensurate or conflicting attributes (Hwang & Yoon, 1981). It can rank differ- ent methods when they are compared in terms of their overall performance. Over the years, a variety of MCDM theories and techniques have been proposed by different behavioral scientists, operational researchers and decision theorists. The methods differ in many areas of theoretical background, type of questions asked and the type of results given (Hobbs & Meier, 1994). The availability and selection of such a method depends on the structure of the model and the information that can be collected. For brevity, a short introduction to TOPSIS is provided and the reader is referred to Saremi, Mousavi, and Sanayei (2009) and Shih, Syur, and Lee (2007) for a more in-depth treatment. The Technique for Order Preference by Similarity to Ideal Solu- tion (TOPSIS), proposed by Hwang and Yoon (1981), referring to the positive and negative ideal solutions as the ideal and anti-ideal solutions is a widely used MCDM method. It based on the concept that the chosen alternative should have the shortest distance from the positive ideal solution (PIS) and the farthest from the negative- ideal solution (NIS) for solving a multiple criteria decision-making problem. According to the simulation comparison of Zanakis, Solomon, Wishart, and Dublish (1998), this technique has the few- est rank reversals among the eight methods of MCDM (Shih, 2008). Chen (2000) extended TOPSIS to fuzzy environments; this ex- tended version used fuzzy linguistic value (represented by fuzzy number) as a substitute for the directly given crisp value in grade assessment. This modified TOPSIS is a practical method and fits hu- man thinking under actual environment (Wang, Cheng, & huang, 2009). The difference between TOPSIS and fuzzy TOPSIS chiefly lies in rating approaches. The merit of fuzzy TOPSIS is using fuzzy num- bers instead of precise numbers (Chen & Tsao, 2008). Fuzzy TOPSIS is flexible and efficient method that is easily under- stood by practitioners and researchers. Fuzzy TOPSIS method is extended in this paper for selecting a proper risk assessment model. 4. Proposed method A systematic approach to extend the TOPSIS is proposed to solve the risk assessment model selection problem under a fuzzy envi- ronment in this section. Assume that there is a committee of k decision makers K ¼ð1; 2; . . . ; kÞ who are responsible for assessing the project risks. Once the set of possible models are selected, the committee determines the best of these models. Model selection first requires Table 1 Linguistic variables for the ratings. Linguistic variables Fuzzy triangular Very low/very poor (0, 0, 1) Low/poor (0, 1, 3) Medium low/medium poor (1, 3, 5) Medium/fair (3, 5, 7) Medium high/medium good (5, 7, 9) High/good (7, 9, 10) Very high/very good (9, 10, 10) 9108 A. KarimiAzari et al. / Expert Systems with Applications 38 (2011) 9105–9111 identification of decision attributes (criteria). Various techniques exist in order to reach a consensus among the experts (Bryson, Mobolurin, & Joseph, 1997). Nominal Group Technique (NGT), Delphi (Van De Ven & Delbecq, 1974), Focus Groups and Brainstorming (Stewart & Shamdasani, 1990) are formal and more useful group management techniques. When comparing the NGT with other group processes the NGT has a number of advantages over other group processes (Potter, Gordon, & Hamer, 2004) thus the NGT technique is suggested to obtain decision criteria/factors. Selected criteria can be classified into two types: benefit factors (C1) and cost ones (C2). After this, members in the risk assessment group are required to provide their judgment on the basis of their knowledge and exper- tise for each model. Decision makers make decisions on the basis of their knowledge of the facts and personal experience. Their judgments and prefer- ences are often vague, inexact, imprecise and uncertain by nature which makes it difficult to estimate their preference with an exact numerical value since crisp data are inadequate to model real-life situations. Decision makers describe their preference with words or sentences in a natural or artificial language. In these circum- stances values are not numbers but linguistic terms. A linguistic variable is a variable whose values (namely linguistic values) have the form of phrases or sentences in a natural language (Von Altrock, 1996). Linguistic values can be represented by fuzzy num- bers. A fuzzy number is a convex fuzzy set, characterized by a given interval of real numbers, each with a grade of membership be- tween 0 and 1 (Wang & Elhag, 2006). In the following, some basic definitions of fuzzy sets theory will be reviewed briefly from Cheng and Lin (2002), Kaufmann and Gupta (1985), and Raj and Kumar (1999). A real fuzzy number A is described as a fuzzy subset of the real line R with member function fA that represents uncertainty. A membership function is defined from universe of discourse to [0, 1] (see Fig. 1). A triangular fuzzy number can be defined as a triplet (a, b, c); the membership function of the fuzzy number A is defined as: fA ¼ 0; x 6 a; x�a b�a ; a 6 x 6 b; c�x c�b ; b 6 x 6 c; 0; x 6 c: 8>>>< >>>: ð1Þ This representation is useful for arithmetic operation on fuzzy numbers. With this notation, the arithmetic operations on fuzzy numbers are defined as follows: ða1; b1; c1ÞðþÞða2; b2; c2Þ¼ ða1 þ a2; b1 þ b2; c1 þ c3Þ; ð2Þ ða1; b1; c1Þð�Þða2; b2; c2Þ¼ ða1 � c2; b1 � b2; c1 � a2Þ; ð3Þ ða1; b1; c1Þð�Þða2; b2; c2Þ¼ ða1 � a2; b1 � b2; c1 � c2Þ; ð4Þ ða1; b1; c1Þð�Þða2; b2; c2Þ¼ ða1 � c2; b1 � b2; c1 � a2Þ; ð5Þ ða1; b1; c1Þ �1 ¼ 1 c1 ; 1 b1 ; 1 a1 � � ; ð6Þ k �ða1; b1; c1Þ¼ ðka1; kb1 þ kc1Þ: ð7Þ Fig. 1. A triangular fuzzy number. According to the vertex method stated by Chen (2000), the dis- tance between fuzzy numbers (a1, b1, c1) and (a2, b2, c2) is calculated as: dðA1; A2Þ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 3 ða1 � a2Þ 2 þðb1 � b2Þ 2 þðc1 � c2Þ 2 h ir : ð8Þ Obviously, in the risk assessment model selection problem the models and the factor set Cj = (j = 1, 2, . . . , j) are finite, so it is very convenient to denote the rating of models on factors by fij. Then a decision problem can be concisely expressed as the fol- lowing decision matrix: Dk ¼ ½f kij �m�n; f kij ¼ðaij; bij; cijÞ is a linguistic variable, indicating the rating of each ith method with respect to each jth factor respect to kth DM. These linguistic variables can be described by triangular fuzzy numbers as shown in Table 1. In the decision making process, different attributes have differ- ent importance. Suppose wi = (i = 1, 2, . . . , m) is the relative weight of factor Ci, where wi P 0 and Pi¼n i¼1 wi ¼ 1. Denote a weight vector by w = (w1, w2, . . . , w6) T. Establishing the relative importance of fac- tors can be obtained by either directly assigning or indirectly using pair-wise comparisons (Cook, 1992). In many real-life cases, a deci- sion maker cannot generally specify exact attribute weights but can provide value ranges (Xu & Chen, 2007) thus it is suggested in this paper that linguistic variables are used for assigning the pri- ority weights of factors. The linguistic variable schemes in the rat- ing set and weighting set, shown in Tables 1 and 2, respectively, are used in this study to evaluate the ratings of strategies with respect to different factors and the importance of the factors. If there is consensus among DMs with respect to rating and importance of factors suppose k = 1 in proposed procedure. The procedure of the TOPSIS method consists of the following steps: 4.1. Establish a normalized matrix The decision matrix must first be normalized so that the ele- ments will be unit-free. The structure of the normalized matrix for the kth decision maker can be expressed as follows: Table 2 Linguistic variables for the importance weight of each criterion. Linguistic variables Fuzzy triangular Very low (0, 0, 0.1) Low (0, 0.1, 0.3) Medium low (0.1, 0.3, 0.5) Medium (0.3, 0.5, 0.7) Medium high (0.5, 0.7, 0.9) High (0.7, 0.9, 1) Very high (0.9, 1, 1) A. KarimiAzari et al. / Expert Systems with Applications 38 (2011) 9105–9111 9109 Rk ¼ ½rkij�m�n; k ¼ 1; 2; . . . ; K; i ¼ 1; 2; . . . ; m; j ¼ 1; 2; . . . ; n; ð9Þ where the rkij is the normalized value of f k ij ¼ðaij; bij; cijÞ which be cal- culated by the following relations: � If jth criterion is a benefit one: rkij ¼ aij c�j ; bij c�j ; cij c�j ! ; ð10Þ where c�j ¼ max cij. � And if jth criterion is a cost one: rkij ¼ a�j cij ; a�j bij ; a�j aij � � ; ð11Þ Table 3 Importance weight of criteria from three decision-makers. DM1 DM2 DM3 C1 MH M M C2 VH H H C3 L VL VL C4 H MH VH where a�j ¼ min aij . 4.2. Construct the weighted normalized decision matrix The columns of the normalized decision matrix for kth decision maker by the associated priority weights of factors are multiplied to construct the weighted normalized decision matrix as follows: V k ¼ ½v kij�m�n; k ¼ 1; 2; . . . ; K; i ¼ 1; 2; . . . ; m; j ¼ 1; 2; . . . ; n; ð12Þ where v kij ¼ r k ijð�Þw k j and v k ij ¼ðv ij1; v ij2; v ij3Þ is a triangular fuzzy number. 4.3. Calculate the separation measure from the ideal and the negative ideal solutions for each decision maker The positive ideal solution indicates the most preferable alter- native, and the negative ideal solution indicates the least prefera- ble alternative. So determine the fuzzy positive ideal solution (FPIS, A+) and fuzzy negative ideal solution (FNIS, A�) as follows (Chen, Lin, & Huang, 2006): Aþ ¼ ðvþ1 ; v þ 2 ; . . . ; v þ n Þ; A � ¼ ðv�1 ; v � 2 ; . . . ; v � n Þ; ð13Þ where vþj ¼ maxiðv ij3Þ and v � j ¼ miniðv ij1Þ; i ¼ 1; 2; . . . ; m; j ¼ 1; 2; . . . ; n. For kth, the distance of each alternative from A+ and A� can be currently calculated as: dkþi ¼ Xn j¼1 dðv kij; v þ j Þ; i ¼ 1; 2; . . . ; m; j ¼ 1; 2; . . . ; n ð14Þ and dk�i ¼ Xn j¼1 dðv kij; v � j Þ; i ¼ 1; 2; . . . ; m; j ¼ 1; 2; . . . ; n; ð15Þ where d(⁄, ⁄) represented the distance measurement between two fuzzy numbers. 4.4. Calculate the overall separation measure from the ideal and the negative ideal solutions To derive group preferences provided by multiple decision mak- ers and combine the group synthesis and prioritization stages into a single integrated stage, the geometric mean with the modified TOPSIS approach is employed. The overall separation measure calculated as: �dþi ¼ YK k¼1 dkþi !1 k ; i ¼ 1; 2; . . . ; m ð16Þ and �d�i ¼ YK k¼1 dk�i !1 k ; i ¼ 1; 2; . . . ; m: ð17Þ 4.5. Calculate the relative closeness to the ideal solution The relative closeness of the alternative Aj with respect to ideal solution A+ is defined as: Ci ¼ �d�i �dþi þ �d�i ; ð18Þ where Ci range belong to the closed interval [0, 1] and i ¼ 1; 2; . . . ; m. 4.6. Rank the alternatives A set of alternatives can now be preference ranked according to the descending order of Ci. The one with the maximum value of Ci is the best. 5. Numerical example To illustrate the group based fuzzy TOPSIS approach introduced above; risk assessment model selection problem faced by XYZ is presented. The case company XYZ is one of the Iranian construc- tion companies. Recently FNP Co. has taken a huge project in road construction. Risk management team is formed to manage risks in the project. Three experts with high qualification regarding project risks are se- lected to form a group. Risk assessment model selection is one of the fundamental tasks of the team. The proposed method com- monly taken decides among a three possible models. After the NGT technique is employed, the group identifies four criteria, for the risk assessment model selection, as follows: � Implementation cost. � External party’s approval. � Complexity. � Usability. The group, respectively, compares the four criteria and evalu- ates their degree of satisfaction with every model. The compared and evaluated grades are shown in Table 3 (see Tables 1 and 2 for the linguistic value and degree of importance). The ratings of the three consultants by the decision makers against the various criteria are shown in Tables 4. The linguistic evaluations are shown in Tables 3 and 4 are converted into trian- gular fuzzy numbers. After establishing a normalized matrix, the weighted normalized fuzzy decision matrix is calculated. To save space the other matrixes are omitted and only the separation from the ideal and the negative ideal solutions for each DM is shown in Table 5. Next, to derive group priorities, the group’s aggre- gated separation distances are generated by its geometric mean. Table 4 Ratings of the 3 consultants by the DMs under the various criteria. C1 C2 C3 C4 DM1 A1 M L ML ML A2 H VL VH L A3 MH VH M VH DM2 A1 ML ML M L A2 VH VL VH VL A3 M VH ML VH DM3 A1 ML M ML L A2 H L VH VL A3 M VH M VH Table 5 The distance measurement. DM1 DM2 DM3 d�1 d þ 1 d�1 d þ 1 d�1 d þ 1 A1 2.33 1.1 2.11 0.9 2.02 1.1 A2 2.78 0.4 2.58 0.1 2.58 0.3 A3 1.18 2.1 1.17 1.7 1 1.9 Table 6 The final closeness coefficient of each model. Model Overall A1 0.441 A2 0.321 A3 0.545 9110 A. KarimiAzari et al. / Expert Systems with Applications 38 (2011) 9105–9111 Table 6 represents the result. At last according to the closeness of the 3 consultant the A3 is the best. 6. Conclusion Despite its importance to the success of project management, risk management is rarely approached with the same rigor as other project management processes such as project scope and schedul- ing. A process of risk management has involved risk identification, risk assessment and risk mitigation. The identification and assess- ment of project risk are the critical procedures for projecting suc- cess. Many construction project risk assessment techniques are currently used in the construction industry but insufficient atten- tion has been paid by researchers to a select suitable risk assess- ment model. To address this decision problem, in this paper a group based fuzzy TOPSIS approach is developed with an effective algorithm to improve the quality and effectiveness of decision making. TOPSIS provides good evaluations and it appears to be more appropriate than other MCDM methods. Construction project would require interaction between dissim- ilar, yet contractually integrated parties, owners, designers, con- tractors, sub-contractors, suppliers, manufacturers, and others. As a result, construction is described as a collaborative teamwork pro- cess where parties with different interests, functions, and objec- tives, share a common goal, which is successful completion of a project. In line with a group decision environment of the problem, the approach provides a simple and effective mechanism to make comparative and absolute judgments in a conventional manner. The proposed method can discriminate successfully and clearly among risk assessment methods. Further research can apply this method to other decision situa- tions in construction industry, like project selection, performance selection, vendor selection, etc. References Bajaj, J. (1997). Analysis of contractors approaches to risk identification in New South Wales, Australia. Construction Management and Economics, 15, 363–369. Baloi, D., & Price, A. D. F. (2003). Modelling global risk factors affecting construction cost performance. International Journal of Project Management, 21, 261–269. Bannerman, P. L. (2008). Risk and risk management in software projects: A reassessment. The Journal of Systems and Software, 81(12), 2118–2133. Bowles, J. B., & Pelaez, C. E. (1995). Fuzzy logic prioritization in a system failure mode, effects and criticality analysis. Reliability Engineering and System Safety, 50, 203–213. Bryson, N., Mobolurin, A., & Joseph, A. (1997). Generating consensus fuzzy cognitive maps. Intelligent Information Systems, 8(10), 231–235. Carr, V., & Tah, J. H. M. (2001). A fuzzy approach to construction project risk assessment and analysis: Construction project risk management system. Advanced in Engineering Software, 32, 847–857. Chen, C. T. (2000). Extensions of the TOPSIS for group decision-making under fuzzy environment. Fuzzy Sets and Systems, 114, 1–9. Chen, C.-T., Lin, C.-T., & Huang, S.-F. (2006). A fuzzy approach for supplier evaluation and selection in supply chain management. International Journal of Production Economics, 102, 289–301. Chen, T.-Y., & Tsao, C.-Y. (2008). The interval-valued fuzzy TOPSIS method and experimental analysis. Fuzzy Sets and Systems, 159, 1410–1428. Cheng, C. H., & Lin, Y. (2002). Evaluating the best main battle tank using fuzzy decision theory with linguistic criteria evaluation. European Journal of Operational Research, 142(1), 174–186. Cho, H. N., Choi, H. H., & Kim, Y. B. (2002). A risk assessment methodology for incorporating uncertainties using fuzzy concepts. Reliability Engineering & System Safety, 78, 173–183. Choi, H., Cho, H., & Seo, J. W. (2004). Risk assessment methodology for underground construction projects. Journal of Construction Engineering and Management, 130(2), 258–272. Cook, R. L. (1992). Expert systems in purchasing applications and development. International Journal of Purchasing and Management, 18, 20–27. Dikmen, I., Birgonul, M. t., Anac, C., Tah, J. H. M., & Aouad, G. (2008). Learning from risks: A tool for post-project risk assessment. Automation in Construction, 18, 42–50. Dikmen, I., Birgonul, M. T., & Han, S. (2007). Using fuzzy risk assessment to rate cost overrun risk in international construction projects. International Journal of Project Management, 25, 494–505. Elkington, P., & Smallman, C. (2002). Managing project risks: A case study from the utilities sector. International Journal of Project Management, 20, 49–57. El-Sayegh, S. M. (2008). Risk assessment and allocation in the UAE construction industry. International Journal of Project Management, 26, 431–438. Fan, M., Lin, N., & Sheu, C. (2008). Choosing a project risk-handling strategy: An analytical model. International Journal of Project Management, 112, 700–713. Hillson, D. (2002). Extending the risk process to manage opportunities. International Journal of Project Management, 20, 235–240. Hillson, D. (2004). Effective opportunity management for projects – exploiting positive risk. New York: Marcel Dekker. Hobbs, B. F., & Meier, P. M. (1994). Multi criteria methods for resource planning: An experimental comparison. IEEE Transactions on Power Systems, 9(4), 1811–1817. Huang, D., Chen, T., & Wang, M.-J. (2001). A fuzzy set approach for event tree analysis. Fuzzy Sets and Systems, 118, 153–165. Hwang, C. L., & Yoon, N. (1981). Multiple attributes decision making methods and application. Berlin: Springer-Verlag. Kaufmann, A., & Gupta, M. M. (1985). Introduction to fuzzy arithmetic: Theory and applications. New York: Van Nostrand Reinhold. Kaufmann, A., & Gupta, M. M. (1988). Fuzzy mathematical models in engineering and management science. Amsterdam: North-Holland. Klemetti, A. (2006). Risk management in construction project networks. Report 2006/ 2. Finland: Laboratory of Industrial Management, Helsinki University of Technology. Lam, K. C., Wang, D., Lee, p. T. K., & Tsang, Y. T. (2007). Modelling risk allocation decision in construction contracts. International Journal of Project Management, 25, 485–493. Lichtenstein, Sh. (1996). Factors in the selection of a risk assessment method. Information Management and Computer Security, 4, 20–25. Lyons, T., & Skitmore, M. (2004). Project management in the Queensland engineering construction industry: A survey. International Journal of Project Management, 22, 51–61. Mills, A. (2001). A systematic approach to risk management for construction. Structural Survey, 19(5), 245–252. Mohammad, A. M., & Al-Bahar, J. F. (1991). Project risk assessment using the analytic hierarchy process. IEEE Transactions on Engineering Management, 38(1), 46–52. Mustafa, M. A., & Al-Bahar, J. F. (1991). Project risk assessment using the analytic hierarchy process. IEEE Transactions on Engineering Management, 38(1), 46–52. Naaranoja, M., Haapalainen, P., & Lonka, H. (2007). Strategic management tools in projects case construction project. International Journal of Project Management, 25, 659–665. Ngai, E. W. T., & Wat, F. K. T. (2005). Fuzzy decision support system for risk analysis in e-commerce development. Decision Support Systems, 40, 235–255. Olcer, A. I., & Odabasi, A. Y. (2005). A new fuzzy multiple attributive group decision making methodology and its application to population/maneuvering system selection problem. European Journal of Operational Research, 166, 93–114. A. KarimiAzari et al. / Expert Systems with Applications 38 (2011) 9105–9111 9111 Olsson, R. (2007). In search of opportunity management: Is the risk management process enough? International Journal of Project Management, 25, 745–752. Perminova, O., Gustafsson, M., & Wikstrom, K. (2008). Defining uncertainty in projects – A new perspective. International Journal of Project Management, 26, 73–79. Potter, M., Gordon, S., & Hamer, P. (2004). The nominal group technique: A useful consensus methodology in physiotherapy research. NZ Journal of Physiotherapy, 32(3), 126–130. Project Management Institute (2000). A guide to the project management book of knowledge (PMBOK). Newtown Square, PA: Project Management Institute. Project Management Institute (2004). A guide to the project management book of knowledge (PMBOK) (3rd ed.). Newtown Square, PA: Project Management Institute. Raj, P. A., & Kumar, D. N. (1999). Ranking alternatives with fuzzy weights using maximizing set and minimizing set. Fuzzy Sets and Systems, 105, 365–375. Sadiq, R., & Husain, T. (2005). A fuzzy-based methodology for an aggregate environmental risk assessment: A case study of drilling waste. Environmental Modelling & Software, 20, 33–46. Saremi, M., Mousavi, S. F., & Sanayei, A. (2009). TQM consultant selection in SMEs with TOPSIS under fuzzy environment. Expert Systems with Applications, 2742–2749. Shih, H.-S. (2008). Incremental analysis for MCDM with an application to group TOPSIS. European Journal of Operational Research, 186(2), 720–734. Shih, H. S., Syur, H. J., & Lee, E. S. (2007). An extension of TOPSIS for group decision making. Mathematical and Computer Modeling, 45, 801–813. Steiner, G. A. (1969). Top management planning. New York: Macmillan. Stewart, D. W., & Shamdasani, P. N. (1990). Focus groups: Theory and practice. London: Sage. Tah, J. H. M., & Carr, V. (2001). Towards a framework for project risk knowledge management in the construction supply chain. Advanced in Engineering Software, 32, 835–846. Teneyuca, D. (2001). Organizational leader’s use of risk management for information technology. Information Security Technical Report, 6(3), 54–59. Terano, T., Asai, K., & Sugeno, M. (1992). Fuzzy systems theory and its applications. San Diego, CA: Academic Press. Thevendran, V., & Mawdesley, M. J. (2004). Perception of human risk factors in construction projects: An exploratory study. International Journal of Project Management, 22, 131–137. Van De Ven, A. H., & Delbecq, A. L. (1974). The effectiveness of nominal, delphi, and interacting group decision making process. Academy of Management Journal, 17(4), 605–621. Von Altrock, C. (1996). Fuzzy logic and neurofuzzy applications in business and finance. New Jersey: Prentice-Hall. Wang, J.-W., Cheng, Ch.-H., & huang, K.-C. (2009). Fuzzy hierarchical TOPSIS for supplier selection. Applied Soft Computing, 9(1), 377–386. Wang, Y. M., & Elhag, T. M. S. (2006). Fuzzy TOPSIS method based on alpha level sets with an application to bridge risk assessment. Expert Systems with Applications, 31, 309–319. White, D. (1995). Application of systems thinking to risk management: A review of the literature. Management Decision, 3(10), 35–45. Williams, T. (1995). A classified bibliography of recent research relating to project risk management. European Journal of Operational Research, 85, 18–38. Xu, Z., & Chen, J. (2007). An interactive method for fuzzy multiple attribute group decision making. Information Sciences, 177, 248–263. Zanakis, S. H., Solomon, A., Wishart, N., & Dublish, S. (1998). Multi-attribute decision making: A simulation comparison of selection methods. European Journal of Operational Research, 107, 507–529. Zayed, T., Amer, M., & Pan, J. (2008). Assessing risk and uncertainty inherent in Chinese highway projects using AHP. International Journal of Project Management, 26, 408–419. Zeng, J., An, M., & Smith, N. J. (2007). Application of a fuzzy based decision making methodology to construction project risk assessment. International Journal of Project Management, 25, 589–600. Zou, P. X. W., Zhang, G., & Wang, J. (2007). Understanding the key risks in construction projects in China. International Journal of Project Management, 25, 601–614. Risk assessment model selection in construction industry Introduction Literature review Risk Risk management Risk assessment model Proposed method Establish a normalized matrix Construct the weighted normalized decision matrix Calculate the separation measure from the ideal and the negative ideal solutions for each decision maker Calculate the overall separation measure from the ideal and the negative ideal solutions Calculate the relative closeness to the ideal solution Rank the alternatives Numerical example Conclusion References 
work_jqqvf7vkqfaztm3eqhkr7qvrzi	----	Sentimental Causal Rule Discovery from Twitter Rahim Dehkharghani, Hanefi Mercan, Arsalan Javeed and Ycel Saygn Dept. of Computer Science and Engineering, Sabancı University, Istanbul, Turkey {rdehkharghani, hanefimercan, ajaveed, ysaygin}@sabanciuniv.edu Abstract—Sentiment analysis refers to the task of extracting the sentiment of people from textual data. This analysis can be employed to reflect the public’s ideas about an issue stated in natural language text. On the other hand, causal rules are associations between different concepts in a context that one (or several concepts) cause(s) the other(s). In this paper, we propose a new concept ”sentimental causal rules” which incorporate causal relationships among different aspects extracted from textual data and the sentiment towards those aspects. We also describe techniques to extract sentimental causal rules and show their effectiveness on Twitter as the data resource. As a case study, we investigated about the Kurdish issue in Turkey which is very important topic in our region causing heated debates that the governments need to follow. The experiments on Twitter data on the chosen case study suggest that sentimental causal rules are an effective way to summarize important aspects in textual data and to suggest causal relationships among those aspects which may lead to better policy making. Index Terms-Sentiment analysis, data mining, causal rules, sentimental causal rules, machine learning, Twitter; I. INTRODUCTION Textual information entered by users in various media may contain their opinion which indicate the sentiment of the authors towards a specific issue. Opinions can be about issues such as a product, a service, or a political issue. To extract the opinions of people from text, one method is to manually analyse each piece of text which would certainly be time consuming. A more efficient method employs a mechanism that could automatically process the textual data using text mining techniques. In this paper we propose the concept of ”Sentimental Causal Rules” where causal rules based on various aspects extracted from textual data sources are associated with sentiment. We also propose sentimental causal rule discovery techniques which combine sentiment analysis and causal rule discovery. A more detailed and formal explanation of this concept is presented in sections III nd IV. Sentiment analysis also possesses broad applicability in political issues which can be sensitive to a nation’s welfare. One such political issue in Turkey, for example, is the conflict between Kurdish parties and the government of Turkey. This crucial issue has been discussed by many users in Twitter mostly in Turkish and somewhat in English. As looking at and processing those tweets manually is considerably time consuming, therefore the problem in this paper is tackled in a semi-automatic methodology to demonstrate the effectiveness of sentimental causal rules to process the tweets. This paper proposes a four-fold methodology for senti- mental causal rule discovery. The goal of the first step is to extract the aspect keywords from the tweets. The second step is extracting the causal association rules based on the previously extracted keywords. The goal of the third step is to facilitate automatic extraction of sentiment from textual data which results in automatically labelling each tweet as positive, negative, or neutral (objective). Finally the forth step assigns polarity values (percentages) to each aspect keyword and also each causal rule separately. Because of the short length of the tweets, usually the overall polarity of the tweet can be considered as the polarity of the aspects seen in the tweet. Our overall method for sentimental causal rule discovery method is presented in figure 1. Fig. 1. The overall flowchart of our system for sentimental causal rule discovery The majority of tweets about the Kurdish issue in Turkey is in Turkish but in this paper, we worked only on English tweets. For this purpose, the relatively small number of tweets downloaded -5000- reflects the homogeneity of tweets, i.e. Turkish tweets. We analysed the tweets by data mining and sentiment analysis techniques. First, we extracted keywords from all tweets which were later used in causal rule extrac- tion. In other words, causal rules were extracted based on the keywords already obtained from the tweets. Then, we did sentiment analysis on tweets labelling them as positive, negative, or neutral. After extracting the aspect keywords, they helped to, then extract the aspect polarities, which means, for example, how much an aspect such as Syria or a causal rule such as student, Kurds ⇒ attack has gained positive or negative sentiment of users on Twitter. With only pure sentiment analysis on tweet level, the results would be the percentage of positive, negative, or neutral tweets about our case study, which itself does not tell much. However, we sentimental causal rules we were able to see why and in which aspects hold positive or negative opinions by Twitter users. II. RELATED WORK Investigating the attitudes of Twitter users on different issues is a new research area which has attracted many researchers. This section briefly summarizes the most relevant research on sentiment analysis in Twitter and also causal rule inference. Designing online tools for sentiment analysis in Twitter is an interesting method of investigating the Twitter users’ ideas. Sentiment140 [1] is one such free tool. This tool allows its users to discover the sentiment of a brand, product, or topic on Twitter. It classifies tweets about a topic (based on input keywords) into positive or negative, taking the advantage of machine learning methods such as Support Vector Machines. This tool uses features such as unigrams and bigrams, ex- tracted from a tweet message. The contribution of this work can be the use of emoticons for distant supervised learning. The smiling emoticon :), for example, in a tweet indicates that it contains positive sentiment and :( indicates that the tweet contains negative sentiment. The designers propose this method to avoid the time spent for manually labelling the training data as positive or negative. In other words, all tweets in training data should have emoticons but for the test data, they are not mandatory. Agarwal et. al. [2] also examine sentiment analysis on Twitter data. The authors experimented with three types of models: the unigram model, a feature based model, and a tree kernel based model. They assumed the unigram model as baseline. They investigated two kinds of models: tree kernel and feature based models and demonstrated that both these models outperform the unigram baseline. Twitter can be employed to build a corpus for sentiment analysis and opinion mining purposes. An example of research for this purpose is by Pak and Paroubek [3]. In this work, the authors automatically collected a large number of tweets as a corpus. Then, they performed statistical linguistic analysis of the collected corpus. Thereafter, they classified to the tweets into positive, negative, or neutral. Their approach is based on the distribution of word frequencies in the corpus. TreeTagger was used for English to tag all the posts in the corpus. Pak and Paroubek took also advantage of the presence of n-grams as a binary feature in training the classifier. A slightly different work has been accomplished by Kouloumpis et. al [4]. This work used a method for building a data model with Twitter hashtags. The features extracted from tweets include n-grams, POS tag of words, and us- ing resources such as MPQA (Multi-Perspective-Question- Answering) subjectivity lexicon. After analysing the impor- tance of all features, the authors conclude that POS features are less useful than others such as presence of intensifiers and positive/negative/neutral emoticons and abbreviations. To our knowledge, very few works have been accomplished on sentiment analysis of Turkish political issues. One example is [5] which investigated the Turkish political news in media. In this work, unigrams and bigrams together with polar Turk- ish terms are used as classification features, which in turn are used to train a classifier to classify unlabelled documents. The authors used four different classifiers: Naive Bayes, Maximum Entropy, SVM and the character based N-Gram Language Model, and compared their efficiency with each other. They concluded that Maximum Entropy and N-Gram Language Model are more efficient than the SVM and the Nave Bayes classifiers. The classification accuracy in different cases are from 65% to 77%. Apart from sentiment analysis research, several works have studied causal association rule mining. Some literature is concerned with discovering itemset patterns for the market basket analysis [6] [7]. In these papers, the authors searched for the nature of the relationships between the itemsets rather than merely extracting the statistical relationships between the itemsets, by using different algorithms such as CCC and CCU. [6] and [7] experimented on several data types and compared their results with each other. It was shown that finding causal models is applicable for large datasets and computationally faster. These works brought a new perspective to causal rule mining area. A hybrid approach (the combination of constraint-based search and Bayesian methods) for discovering causal relations from sparse data, has been proposed by Dash and Druzdzel [8]. The algorithm used in this approach searches the space of equivalence classes of models (essential graphs) using heuristic techniques. Each essential graph is converted into a directed acyclic graph and scored using a Bayesian scoring metric. Two variants of the algorithm are developed and tested using data from randomly generated networks of sizes from 15 to 45 nodes with data sizes ranging from 250 to 2000 records. Both variations have been compared with each other; it was observed that in each experimental step, both variations of greedy search outperformed the previous (experimental) steps. Mush research on text mining focuses on causal relations. They use hand-coded patterns to extract causation information from the text. One example has been done by Roxana Girju and Dan Moldovan [9]. The authors focused only on the causal relations and proposed a method for automatic detection of causation patterns. They accomplished a semi-automatic validation of ambiguous lexico-syntactic patterns referring to the causation. First, they discovered lexico-syntactic patterns that could express the causal relations; and then, the authors validated and ranked the ambiguous patterns acquired based on the semantic constraints on nouns and verbs. The method automatically discovers the applicable lexico-syntactic patterns referring to the causation and removes the ambiguity of the causal relationships obtained from the patterns in text. This work brought a considerable improvement in time and user work, compared to other previous attempts [10], [11]. To our knowledge the current work is the first one investi- gating Twitter user views using sentimental causal rules. Also we believe the case study used in this paper, Kurdish political issue in Turkey, is being processed for the first time through sentiment analysis and causal rule mining techniques. III. SENTIMENTAL CAUSAL RULES Causal rules, are association rules in which, one side of the association implies the other side based on the inductive reasoning mechanism. These rules show not only the depen- dency, but causality between different concepts (variables) in a context. For example in such a rule, X, Y → Z, instead of saying that three variables, X, Y and Z are dependent, it is said that X and Y cause Z. A real example in marketing can be soya, mushroom → ketchup meaning that for people buying soya and mushroom causes buying also ketchup. But it does not imply that buying ketchup causes also buying soya and mushroom. This causality can also happen in political domain. For example peace, Kurds → live means that the result of making peace with Kurds in Turkey will let them live (peacefully). In this work, before extracting the causal rules, important aspect keywords (variables stated above) have been extracted (explained in section IV-A) from tweets to be used in causal rule extraction step. The results of aspect keyword extraction step is a set of keywords (kw1, kw2, ...kwn) as the important aspects of the political Kurdish issue in Turkey. We presented each tweet as a vector of aspect keywords (ak1, ak2, ...akn), while n is the number of total aspect keywords which is 80. The value of aki for a tweet is 1 if the keyword kwi has been seen in that tweet, and 0 otherwise. After feeding these vectors to our causal rule extractor (section IV-B), causal rules are extracted in the form of aki, akj → akk which indicates that existence of terms aki and akj in a tweet cause the existence of akk. The real example related to the subject of this paper can be the previously stated rule (peace, Kurds → live). Similar to each tweet or each aspect keyword in a tweet, causal rules can be polar. Extracting this polarity gives extra information as the attitude of the Twitter users about dif- ferent aspects discussed in the tweets. For this reason, we investigated the polarity of each rule and extracted it via statistical methods; although other methods may work for it. This approach is explained in detail in subsection IV-D. The rule,deal, Iran ⇒ Kurds ,for example, has been mostly seen in negative tweets, therefore, the label neg is assigned to it. Apart from rule polarities, aspect polarities have been also investigated. The results of this investigation are presented in section V-B. The method for computing these polarities is again based on the occurrences of aspects in positive, negative or objective tweets. Based on our statistics, no general relation could be found between aspects and rule polarities. In other words, the aspects in different sides of a causal rule could have different polarities. This fact suggests that causal rules do not necessarily preserve the polarity of the aspects in two sides of the rule. A more detailed and step by step explanation of the sentimental causal rule discovery is given in the following section. IV. SENTIMENTAL CAUSAL RULE EXTRACTION In this section, the details of sentimental causal rule extrac- tion from textual data is explained. First, Causal rules need to be extracted (step two) based on the aspect keywords that have already been extracted in step one. Apart from causal rules, sentiment analysis is accomplished on tweets which labels all (3000) tweets as positive, negative or objective (step three). The extracted causal rules take the advantage of sentimental analysis techniques, which in turn need extracting the polarity features from the tweets. The results obtained from the sentiment analysis step as the pos/neg/obj percentage of tweets, each rule has been extracted from, are used as criteria to classify the rules in three classes (step 4). Each step in this four-fold methodology is a subsection of this section. A. Aspect Keywords Extraction That a sentence can express an opinion which always contains an extractable target renders it useful in sentiment analysis. These targets often offer essential information or key- words relating to the content of the opinion. The term Aspect in this work is assigned to those terms that are usually noun and carry a high importance in the above mentioned issue, such as Syria. Each aspect is an important feature related to the context, from which it has been extracted. In our research, we extracted the aspect keywords from the tweet set using a third party commercial service (www.alchemyapi.com), a cloud based text mining platform. The extraction process is accomplished by feeding the content of the subject tweet set to our Java based software whose keyword extraction feature has been built around the AlchemyApi’s SDK. For aspect keywords extraction in NLP, several techniques exist; the designers of AlchemyApi chose to take advantage of machine learning techniques and also the frequent nouns and phrases technique. The nouns can be identified using a POS tagger. In frequency based techniques, the most frequent words are specified and marked as aspects; Then the unrelated ones are removed after human evaluation. Although this method is simple, it is nonetheless quite effective. No further information is available regarding the keyword extraction method of this tool. After analysis of the tweet set, a group of 80 different keywords was reported; this group, in turn was used by our software utility to generate vectors for finding causal association rules using causal inference. B. Causally Associative Rule Extraction Association rule mining finds all the rules existing in the database that satisfy minimum support and confidence constraints. In the current study, instead of mere associations, we consider the problem of determining causal relationships between variables (aspect keywords) as they infer each other. In many previous works on association rule mining, only pure association rules as the dependency between variables have been investigated; those works, however, did not extract the causalities from the extracted rules, such as variable X causes Y in X → Y . Since this causality information can be quite useful in our case (Information extraction from political comments), we decided to use this technique. In previous research on causal rule discovery, Bayesian analysis was the major method to extract the rules. In this method, causal Bayesian network is used as a network in which, nodes represent the variables (attributes) and the arcs represent probabilistic dependence. The predecessors of a node (variable) is supposed to cause its associated nodes. Constraint-based methods, on the other hand, can be an alternative to Bayesian methods. The PC and FCI algorithms [12] -two example algorithms of this method- use observa- tional data -which has been already seen- to constrain the possible causal relationships between variables. Cooper [13] has described an algorithm called Local Causal Discovery (LCD) that is a special case of the PC and FCI algorithms. The LCD (Local Causal Discovery) algorithm (Fig. IV-B) is a polynomial constraint based algorithm which uses the test of variable dependence and independence. The crucial part of this technique is Markov Condition [12]: Definition (Markov Condition): let A be a node in a causal Bayesian network, and let B be any node that is not a descendant of A in the causal network. Then the Markov condition holds if A and B are independent, conditioned on the parents of A. Simply, any node in Bayesian network is conditionally independent of its non-descendants, given its parents. Input : A vector of terms as the tweet, D and CD stand for Dependence and Conditional Dependence respectively Output : list of possible causal relationships for all aspect keywords aki 6= akj if not D(aki, akj) for all aspect keywords akj not in (aki, akj) if D(aki, akk) and D(akj, akk) and CD(aki, akj|akk) output ”aki and akj cause akk” Fig. 2. LCD algorithm (Local Causal Discovery) for our case, adapted from [7] Rule (CCU causality): Assuming that Markov Condition holds, the existence of causal relations can be investigated in tweets. Assume there are three variables (aspect keywords): X, Y and Z. Using the Chi-Square test [14], dependency between these variables can be figured out. For example, suppose that X and Y are dependent. Then it can be concluded that one is caused by the other (directly or indirectly) i.e, one is descendant of the other, or possibly the third variable cause both of them: X → Y , Y → X or X ← Z → Y If Z is dependent on both X and Y , then these three variables should be causally related. There are several possibilities for the structure of ordering these variables. Nonetheless, these conditions: X and Z are dependent, Y and Z are dependent but Y and Z are independent; however, they become depen- dent when conditioned on X, decrease the possibilities into one: X → Z ← Y ; which means that X and Y cause Z. This case has been used in our rule extraction method. Apart from methodology, some assumptions on dataset (tweets) have been considered: DiscreteV ariables : Every variable has binary value; the word exists in a tweet or not. CausalFaithfulness : If two word are causally related to each other, they have to be dependent. Reverse may not be true in general. NoSelectionBias : The dataset is normally distributed. Because CausalRules concept has been already explained in detail in other papers such as [13], and it is only used in our work, we do not explain it more than this in this paper. C. Sentiment Analysis of Tweets There are three types of approaches to sentiment analysis: machine learning based, lexicon based and linguistic analysis based approaches [15]. In this work, we took advantage of first two methods. We did not analyse tweets linguistically. For example the the improvement in accuracy gained by con- sidering the negation compared to the case without covering it, was very low (less than 1%), therefore we did not cover the negation problem. In supervised training based approaches, labelled reviews (tweets in this work) are used for training a classifier to distinguish between positive, negative, or neutral reviews, considering the extracted features. Then, given a sample review in testing phase, the same features are extracted and compared to the learned models of positive, negative, or neutral reviews. The features used in this phase are based on the the average polarities and term frequencies of words in each tweet message, as summarized in Table I. The terms used in this table such as polarity ( numerical value indicating the positivity, negativity or neutrality of a review) and PMI (Pairwise Mutual Information) are explained more in IV-C2. Some of these features are well-known and used in the literature ([16], [17]) as well as our previous works [18] [19] [20] in a slightly different way, whereas, in this paper, the overall methodology for sentiment analysis is unique in its combined methodology. The first seventy features, F1 to F35 and F36 to F70, are computed based on the presence of a set of positive and a set of negative seed words built by Hu and Liu [21] (called SeedWordSet in this paper, including 2005 positive, and 4783 negative words) in different positions of tweets; while features F71 to F76 are computed using word polarities obtained from the SentiWordNet (SW) and the SenticNet (SN) -two known polarity resources in sentiment analysis- and PMI values of positive/negative seed words. Finally feature F77 is simply the occurrence of question mark (’?’) in a tweet. A more precise explanation of features is presented in the following subsections. 1) Features Based on the Occurrences of seed words in different positions of the tweets (F1-F70): In this set of features, the number of positive and negative seed words in a tweet has been counted. In other words, for each tweet, a feature vector with 70 values has been built. The number 70 is chosen because no tweet in our set has more than 35 words (2 * 35 = 70). The value of each feature can be zero or one. The value 1 for ith feature indicates that the ith word in tweet is positive, if i <= 35 or it is negative, if i > 35. The positivity TABLE I FEATURES EXTRACTED FROM EACH TWEET. SW AND SN STAND FOR SENTIWORDNET AND SENTICNET RESPECTIVELY Feature type Feature name features based on polar terms F1-F35: positive terms in each position in different positions F36-F70: negative terms in each position F71: avg. polarity of neg. terms in SW features based on F72: Avg. polarity of pos. terms in SW polarity resources F73: Avg. polarity of neg. terms in SN F74: avg. polarity of pos. terms in SN F75: avg. PMI value of pos. terms F76: avg. PMI value of neg. terms F77: occurrence of ”?” in tweet or negativity of a word or a set of words has been specified according to SeedWordSet. 2) Features Based on polarity resources (F71-F76): In SentiWordNet, three numerical values are assigned to each connotation of a word (or an expression): positivity, negativity or objectivity [22]. The summation of these three scores must equal to one which is presented in 1: Pos.Score(w) + Neg.Score(w) + Obj.Score(w) = 1 (1) where w stands for a given word; and the three scores stand for its positivity, negativity or objectivity scores, respectively. Furthermore, we define the polarity of a word w as: Pol(w) = Pos.Score(w)−Neg.Score(w) (2) There are several senses and consequently polarity values (each value for each sense) for each word in SentiWordNet; to compute the polarity of a word in a context, we did not do word sense disambiguation because it is a different problem that not yet been completely solved. We employed the average polarity of all senses of a word in SentiWordNet as its overall polarity. Consequently, the average polarity of all words in a tweet as a review, r, denoted by AP(r) is computed as in (3). AP(r) = 1 |r| ∑ wi∈r Pol(wi) (3) where |r| is the number of words in review (tweet) r and Pol(wi) is the polarity of the word wi as defined above. Features F71 and F72 are based on the average polarity concept (AP): the average polarity of only the negative and only the positive words in a tweet, respectively. A word w is decided as positive if Pol(w) > 0, or otherwise decided as negative. Those words with Pol(w) = 0 are not considered in our work. The features F73 and F74 are computed in similar way to F71 and F72 using the SenticNet. However, in SenticNet, unlike SentiWordNet, there is only one numerical value between −1 and 1 assigned to a word or phrase as its overall polarity; therefore, we did not encounter issues such as word sense disambiguation like what we had in SentiWordNet. In SenticNet a word or a group of words is supposed to be negative if its overall polarity score is less than 0 or otherwise supposed positive. The words with polarity value of zero are not used n this work. The last pair of features, F75 and F76 are computed using the PMI -Pairwise Mutual Information- of seed words in SeedWordSet. The formula used in computing the PMI value for each seed word is 4. PMI(t1, t2) = P(t1, t2) P(t1)∗P(t2) (4) t1 and t2 are the terms that the PMI value is being computed for them. Here, t1 is any of the seed words in SeedWordSet and t2 is any of the pure positive words such as excellent or pure negative words such as awful. P(ti) is the probability of seeing ti in a tweet and consequently P(t1, t2) is the prob- ability of seeing both t1 and t2 in a tweet. These probabilities have been computed using the tweets in training set. We computed the PMI value of each positive seed word with very positive words and also the PMI value of each negative seed word with very negative words presented in Table II. The value assigned to each seed word in the SeedWordSet is the average PMI value of that word with all negative or positive words in Table II. The computed values are used in the last pair of features F75 and F76. These features are the average PMI values of seed words in the SeedWordSet that have been seen in a tweets of training set. TABLE II LIST OF VERY POSITIVE AND VERY NEGATIVE WORDS USED TO COMPUTE PMI VALUE OF SEED WORDS IN SeedWordSet pos. words excellent, fantastic, good, outstanding, extraordinary neg. words awful, unpleasant, bad, sad, disgusting Finally feature F77 is simply the occurrence of the question mark in a tweet. Usually tweets with question mark have more of a tendency to be neutral compared to those without question mark. Although this feature does not use any resource, it is placed among the second group of features to prevent adding a new feature group. D. Assigning Polarity Values to Causal Rules Based on the statistical processing of the tweets, we as- signed three polarity values to each causal rule, which indicate how much a given rule has been extracted from positive, negative or objective tweets. As mentioned earlier, usually the overall (tweet level) polarity of the tweet can be considered as the polarity of the aspect(s) discussed in it. These scores are presented in results section (Table VI). Consequently we assign obj/pos/neg label to each rule if it satisfies one of the constraints below: • else if obj.score > (neg.score + pos.score) or neg.score == pos.score then the rule is considered as objective. • else if (pos.score > neg.score) and (pos.score > obj.score) then the rule is considered as positive. • else if (neg.score > pos.score) and (neg.score > obj.score) then the rule is considered as negative. The three scores mentioned above stand for the positivity, negativity and objectivity scores of each rule based on the percentage of tweets it has been extracted. The proposed constraints for labelling is a suggested algorithm, while other methods such as strict threshold could be used; we chose the proposed one because it is more flexible than the strict threshold and also could generate more real results. In strict threshold method, similar to confidence and support values in association rule extraction, a threshold is set and those rules with higher scores than the threshold are used for some goal. In most cases, rules are subjective and carry polar informa- tion, therefore the necessary condition for a rule to be objective is harder to satisfy (obj.score > (neg.score + pos.score)) than the conditions to be positive or negative. A rule is labelled as objective if its objectivity score is higher than the sum of both positivity and negativity scores or its positivity and negativity values are equal without regarding the objectivity score. But to be positive (or negative), having the maximum score as positive (or negative) is enough. Those rules that do not fit in any of the cases above, cannot be labelled in our method. V. EXPERIMENTAL EVALUATION In this section, the evaluation of proposed methodology explained in section IV-C has been presented including the dataset, results and discussion on results. A. Data Sets Twitter allows its users to post real time messages of 140 characters in length. The messages are called as tweets. Due to quick (time) and short nature of tweets, Twitter users often make spelling mistakes, use special characters to express special meanings and use emoticons to express feelings. For this paper, tweets have been collected on three topics Kurdish and Turkey, PKK (Partiya Karker Kurdistan which means Kurdistan Worker’s Party) and, Kurdistan and Turkey from a third party commercial service i.e. www.tweetarchivist.com. The data has been collected by cap- turing real-time tweets on a topic and archiving them. Neither kind of restriction such as language or location was imposed during the tweet collection process. The archive was made through collecting tweets in first three months of 2013 which resulted in around 10000 tweets in multiple languages. For this study, we sorted them and filtered only English tweets which resulted in around 5000. The filtered tweets were further analysed to remove scrap, irrelevant and duplicated tweets. The training dataset consisted about 2000 tweets. For the training data set each tweet was labelled manually as positive, negative or neutral. After training the classifier using the training set, we extracted another 1000 cleaned and unlabelled tweets and labelled them using the trained classifier. Preprocessing of Tweets: Due to the nature of tweets, they need to be preprocessed to treat missing, incomplete or noisy data. To do so, tweets were passed through a Java utility created specifically for this goal. The preprocessing involves the following steps: • Removal of URLs: some tweets contain URLs for hyper- linking different webpages. For sentiment analysis task, they do not contribute; therefore they are removed from tweets. • Removal of retweets: retweeting is the process of for- warding someone else’s tweet and posting to contact list of forwarding person. Retweets are duplicates of original tweets; so, such tweets have to be identified and removed from the dataset. • Treatment of special characters: special characters e.g. []()/!, have to be removed in the preprocessing step because they often are concatenated with words such as /begin ning/ which make the word of no use because of unavailability in the dictionary and failure to assign polarity to such words. • Treatment of emoticons: tweets contain emoticons to express users’ feelings, e.g. :- ) for happiness and <3 for love. Emoticons play a vital role in decoding the exact sense of tweets. For this purpose, a dictionary of emoti- cons has been created by collecting 80 emoticons and their emotional state. Each occurrence of emoticon in a tweet has been replaced with its corresponding emotional state; e. g. <3 substituted with love. An alternative for this, is using classification features as the presence f each emoticon in tweets. • Treatment of acronyms: a similar approach has been followed for treatment of acronyms. For acronym treat- ment, a dictionary of 500 acronyms had been created. The acronyms have been collected through the Internet resource (www.noslang.com). Each tweet has been pro- cessed to replace any acronym with its corresponding meaning e.g. lol replaced with laughing out loud or dah replaced with dumb as hell. The sentiment analysis in tweets is a difficult task because 1) the length of tweets is short and 2) people usually do not express their ideas clearly in tweets about an issue; they rather use an informal language. B. Results The results obtained from our methodology can be catego- rized in different groups. All the results have been extracted from about 3000 English tweets. The first group is the accuracy of our sentiment analysis system. This accuracy is different for binary and ternary classifications, specially when different classifiers are used (Table III). These accuracies are computed using 5-fold cross validation on training data in Weka 3.6. Although the accuracy of the classification is less than ideal, however, it may worth to save the time spent on manually labelling the test tweets by automatic classification. We ap- plied ternary classification on test data by Neural Networks (multilayer perceptron) classifier to predict the label of test set tweets. The overall percentage of positive, negative or neutral tweets about the Kurdish political issue in Turkey is the next set of results which reflects the overall attitude of public. These percentages can be seen in Table IV. It seems that many users think negative rather than positive about the mentioned issue. TABLE III ACCURACY OF BINARY AND TERNARY CLASSIFICATION OF TWEETS BY DIFFERENT CLASSIFIERS: SVM, LOGISTIC AND NEURAL NETWORKS (%) Classes SVM Logistic NN Binary 72 70 76 Ternary 67 64 68 TABLE IV OVERALL PERCENTAGE OF POSITIVE, NEGATIVE OR NEUTRAL TWEETS ABOUT THE KURDISH POLITICAL ISSUE IN TURKEY (%) Positive Negative Neutral 32 48 20 The third group of results is the keywords extracted from those tweets using an online keyword extractor (ex- plained in IV-A). These keywords with the percentage of positive/negative/objective tweets including them are shown in Table V. These percentages indicate the percentage of pos/neg/obj tweets including each aspect keyword. Finally, the association rules extracted from labelled tweets can bee seen in Table VI. In this table, support and confidence thresholds are taken as 0.004 and 0.2 respectively. The associ- ation rules revealed through causal inference provided useful insights into the overall summary and general sentiment of people about different aspects, discussed in tweet set. Based on rule labelling algorithm stated above, most of the rules presented in results section (Table VI) are labelled as negative because the negativity percentage of them are higher than the other scores. This negativity may be because no complete solution has been found for the Kurdish political issue in Turkey and it still remains as a negative issue in people’s opinion. Several hypotheses can be extracted from those rules. Although some of these hypotheses may not be completely true, they can provide useful information for politicians. • The rule rebel, Kurds ⇒ withdraw may owe to the fact that the Kurdish rebel will cause them withdrawing from being under control of Turkish government. Based on our rule classification algorithm, this rule is classified as objective because pos.score = neg.score. • association rules such as deal, Kurds ⇒ Syria, Syria, fight ⇒ peace, and strategic, war ⇒ Syria encircle the group of tweets discussing about the impact of peace deal on the security situation of Syria. users comments and several news-tweets on Twitter argue that after evacuating Turkey, the rebel groups are moving to Syria and will operate there. This action will bring momentum to their strategic terror activities from now on for Syrian region. These rules mostly contain the negative keywords such as fight, attack and Syria, which shows the negative attitude of Twitter users about this issue. Also two of these rules are classified a negative and one as positive. • The rule corridor, war ⇒ Syria has a means of the creation of Kurdish corridor in the region can be eased by fuelling war in Syria. This one is also classified as negative. Syria is an strategic region in Kurdish issue because Kurds are the largest ethnic minority in Syria. • The rule Islam, Kurds → attack suggests that some people blame Muslim Kurds to be the reason of potential war in Turkey. The high negativity percentage (100%) of this rule shows the pure negative attitude of Twitter about this this subject. On the other hand, the keyword Islam has also higher negative score than positive which approves the mentioned assumption. • The rule student, kurd ⇒ attack implies that some Kurds force the students to participate in war against the Turkey. This issue had been mentioned in Turkish TV channels already. Not surprisingly this rule is fully negative. • The rule PKK, deal ⇒ peace suggests that the peace will come to the area if a debate process (deal) starts between two sides of this struggle (Turkey and the PKK). The higher negative percentage of this rule indicates the negative attitude of people to make peace with the PKK. • The rule PKK, Kurdistan ⇒ deal indicated the fact that if PKK and Kurdish people can have their own Kurdistan (country), the peace deal will happen, how- ever, this assumption is approached negatively by the users, based on the obtained pos/neg/obj percentages. • A few association rules such as deal, Iran ⇒ Kurds and Iraq, Iran ⇒ Kurds indicate that the Kurdish issue in Turkey has dependency on the political involvement and internal situation of Iraq and Iran. BAsed on the negativity scores of both rules, this point is also considered as negative. TABLE V A PART OF EXTRACTED ASPECTS AND THE PERCENTAGE OF POS./NEG/OBJ. TWEETS INCLUDING THEM aspect name pos% obj% neg% PKK 29 55 16 Turkey 33 46 21 Kurdish 31 47 22 rebel 58 28 14 Iran 17 70 13 Syria 17 61 22 Iraq 36 46 18 leader 54 24 22 peace 54 28 18 corridor 21 64 15 war 19 41 40 Islam 12 62 26 Process 52 32 16 Kurds 26 53 21 Withdraw 54 33 13 Fight 35 42 23 Strategic 21 48 31 Attack 6 88 6 student 20 66 14 Deal 15 72 13 VI. CONCLUSION AND FUTURE WORK In this paper, we proposed the new concept of sentimental causal rules which consists of both sentiment analysis and TABLE VI ASSOCIATION RULES WITH THE PERCENTAGE OF POS./NEG/OBJ. SCORES EXTRACTED BASED ON THE ASPECTS Causal rule pos obj neg label rebel, Kurds ⇒ withdraw 50 0 50 obj Syria, fight ⇒ peace 33 50 17 pos deal, Kurds ⇒ Syria 6 5 89 neg strategic, war ⇒ Syria 29 29 42 neg corridor, war ⇒ Syria 29 29 42 neg Islam, Kurds ⇒ attack 0 0 100 neg deal, Iran ⇒ Kurds 6 6 88 neg Iraq, Iran ⇒ Kurds 4 10 86 neg PKK, deal ⇒ peace 22 7 71 neg PKK, Kurdistan ⇒ deal 15 13 72 neg student, Kurds ⇒ attack 0 0 100 neg causal rule concepts. As the case study, we investigated the challenging problems between Turkish government and Kurdish parties. We investigated the ideas of Twitter users about this issue. A four-phase approach is proposed in this work: 1) aspect keywords extraction, 2) causal rules extraction based on aspect keywords already obtained from the tweets, 3) sentiment analysis on tweets and 4) assigning the polarity values to each aspect and also causal association rules. The achieved results, specially those related to the opinion of users about different aspects of the Kurdish issue in Turkey seem fairly well and approximate the current situation of Turkey. This paper reflects the ideas of Twitter users (mostly non- Turkish) about the above mentioned issue; Turkish people usually write their tweets in Turkish. These kind of works can help politicians know the ideas of people about different political issues. We will investigate Turkish tweets in this area in our future work. We found the sentiment analysis of tweets a difficult task mostly because of the short length of tweets, many junk words and inevitable syntax errors in them. REFERENCES [1] A. Go, R. Bhayani, and L. Huang, “Twitter sentiment classification using distant supervision,” CS224N Project Report, Stanford, pp. 1–12, 2009. [2] A. Agarwal, B. Xie, I. Vovsha, O. Rambow, and R. Passonneau, “Sentiment analysis of twitter data,” in Proceedings of the Workshop on Languages in Social Media. Association for Computational Linguistics, 2011, pp. 30–38. [3] A. Pak and P. Paroubek, “Twitter as a corpus for sentiment analysis and opinion mining.” in LREC, 2010. [4] E. Kouloumpis, T. Wilson, and J. Moore, “Twitter sentiment analysis: The good the bad and the omg!” in ICWSM, 2011. [5] M. Kaya, G. Fidan, and I. H. Toroslu, “Sentiment analysis of turkish political news,” in Proceedings of the The 2012 IEEE/WIC/ACM Inter- national Joint Conferences on Web Intelligence and Intelligent Agent Technology-Volume 01. IEEE Computer Society, 2012, pp. 174–180. [6] L. Cavique, “A scalable algorithm for the market basket analysis,” Journal of Retailing and Consumer Services, vol. 14, no. 6, pp. 400–407, 2007. [7] C. Silverstein, S. Brin, R. Motwani, and J. Ullman, “Scalable techniques for mining causal structures,” Data Mining and Knowledge Discovery, vol. 4, no. 2-3, pp. 163–192, 2000. [8] D. Dash and M. J. Druzdzel, “A hybrid anytime algorithm for the construction of causal models from sparse data,” in Proceedings of the Fifteenth conference on Uncertainty in artificial intelligence. Morgan Kaufmann Publishers Inc., 1999, pp. 142–149. [9] R. Girju and D. I. Moldovan, “Text mining for causal relations.” in FLAIRS Conference, 2002, pp. 360–364. [10] D. Garcia, “Coatis, an nlp system to locate expressions of actions connected by causality links,” in Knowledge Acquisition, Modeling and Management. Springer, 1997, pp. 347–352. [11] C. S. Khoo, S. Chan, and Y. Niu, “Extracting causal knowledge from a medical database using graphical patterns,” in Proceedings of the 38th Annual Meeting on Association for Computational Linguistics. Association for Computational Linguistics, 2000, pp. 336–343. [12] P. Spirtes, C. Glymour, and R. Scheines, Causation, prediction, and search. The MIT Press, 2000, vol. 81. [13] G. F. Cooper, “A simple constraint-based algorithm for efficiently mining observational databases for causal relationships,” Data Mining and Knowledge Discovery, vol. 1, no. 2, pp. 203–224, 1997. [14] W.-C. Hou, “Quality of association rules by chi-squared test.” 2009. [15] E. Haddi, X. Liu, and Y. Shi, “The role of text pre-processing in sentiment analysis,” Procedia Computer Science, vol. 17, pp. 26–32, 2013. [16] B. Ohana and B. Tierney, “Sentiment classification of reviews using SentiWordNet,” in 9th. IT & T Conference, 2009, p. 13. [17] A. Hamouda and A. Rohaim, “Reviews classification using sentiwordnet lexicon,” in Journal on Computer Science and Information Technology (OJCSIT), Vol. (2), No.(1), 2011. [18] R. Dehkharghani, B. Yanikoglu, D. Tapucu, and Y. Saygin, “Adapta- tion and use of subjectivity lexicons for domain dependent sentiment classification,” in Data Mining Workshops (ICDMW), 2012 IEEE 12th International Conference on. IEEE, 2012, pp. 669–673. [19] G. Gezici, R. Dehkharghani, B. Yanikoglu, D. Tapucu, and Y. Saygin, “Su-sentilab: A classification system for sentiment analysis in twitter.” [20] R. Dehkharghani and C. Yilmaz, “Automatically identifying a software products quality attributes through sentiment analysis of tweets.” [21] B. Liu, M. Hu, and J. Cheng, “Opinion observer: analyzing and comparing opinions on the web,” in WWW ’05: Proceedings of the 14th international conference on World Wide Web. New York, NY, USA: ACM, 2005, pp. 342–351. [22] S. Baccianella, A. Esuli, and F. Sebastiani, “Sentiwordnet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining,” in Proc. of LREC, 2010. 
work_jsnmrvhjlfgxfkyxshfjgxeayu	----	Document downloaded from: This paper must be cited as: The final publication is available at Copyright http://dx.doi.org/10.1016/j.eswa.2012.01.140 http://hdl.handle.net/10251/34812 Elsevier Domenech, J.; De La Ossa Perez, BA.; Sahuquillo Borrás, J.; Gil Salinas, JA.; Pont Sanjuan, A. (2012). A taxonomy of web prediction algorithms. Expert Systems with Applications. 39(9):8496-8502. doi:10.1016/j.eswa.2012.01.140. A taxonomy of web prediction algorithms Josep Domenech, Bernardo de la Ossa, Julio Sahuquillo, Jose A. Gil and Ana Pont Universitat Politecnica de Valencia. Camı́ de Vera, s/n. 46022 Valencia (Spain) jdomenech@upvnet.upv.es; berospe@doctor.upv.es; {jsahuqui,jagil,apont}@disca.upv.es November 10, 2011 Abstract Web prefetching techniques are an attractive solution to reduce the user-perceived latency. These techniques are driven by a prediction en- gine or algorithm that guesses following actions of web users. A large amount of prediction algorithms has been proposed since the first pre- fetching approach was published, although it is only over the last two or three years when they have begun to be successfully implemented in com- mercial products. These algorithms can be implemented in any element of the web architecture and can use a wide variety of information as input. This affects their structure, data system, computational resources and ac- curacy. The knowledge of the input information and the understanding of how it can be handled to make predictions can help to improve the design of current prediction engines, and consequently prefetching techniques. This paper analyzes fifty of the most relevant algorithms proposed along 15 years of prefetching research and proposes a taxonomy where the algorithms are classified according to the input data they use. For each group, the main advantages and shortcomings are highlighted. 1 Introduction Web prediction algorithms pursue to discover patterns that permit them to foresee subsequent web user demands. Applications of these predictions include recommendation systems [11], e-commerce [1], content personalization [34], and reduction of user-perceived latencies by means of web prefetching [36] and cache prevalidation [17]. Since the goals of these applications are completely different, prediction algorithms have different requirements when used for each of them. For instance, prediction algorithms for web prefetching might work at object- level granularity because prefetching one object might reduce the user-perceived latency. However, recommendation systems focus on page-level granularity since users expect that the system recommends a page, not a single object composing a page. This paper focuses on prediction algorithms applied to web prefetching. Web prefetching is a technique aimed at reducing user-perceived latencies by process- ing (usually downloading) user requests before they actually demand them. For 1 example, if a user is likely to visit page B after visiting page A, the browser can download and cache page B just after requesting page A, i.e., while the user is reading page A. In this case, when the user actually requests page B, this is displayed without network latency since it is already in cache. Accurate predictions are required because, on misprediction, prefetched objects waste client, server and/or network resources, and this could degrade the overall sys- tem performance. For this reason, the prediction algorithm is the core of web prefetching. A wide range of prediction algorithms for web prefetching have been pro- posed in the literature over the last 15 years. Although early proposals focused on exploiting the object popularity to determine the objects that are more likely to be requested in the near future, the most explored option is to predict subsequent requests from past sequences of requests. However, over the last years, other data sources beyond the sequence of accesses, such as the site link structure, the HTTP headers or the page contents, are being included in the prediction algorithms. Understanding and identifying advantages and shortcomings of prediction algorithms is a hard task even for an experienced researcher in this field because of the large amount of proposals that work according to different rules. This means that selecting the best approach or developing a new one for a given environment represents an arduous task. Few attempts classifying prediction algorithms have been published and they mainly focus on how sequences of accesses are analyzed. Rabinovich and Spatscheck [41] classify different approaches depending on how predictions are made. They differentiate three categories of Markov models depending on how many previous accesses are used. Other category includes algorithms based on the popularity that web objects exhibit, and another one is dedicated to those algorithms exploiting the structure of the web. The last classification published, presented by Davison [15], divides the prediction algorithms into five categories according to how the sequence of accesses is analyzed. Although it is mentioned that more objects can be predicted if extra information sources are used (e.g., web page content), this is not analyzed in depth. Unfortunately, the complexity of prediction algorithms published in the liter- ature from Davison’s classification has noticeably increased. As a consequence, many of these proposals do not fit well in any of the previously identified cate- gories. Therefore, a new effort is required to order and clarify them to provide a sound understanding of their advantages, progresses and real possibilities of use. In this paper, we propose a prediction algorithm taxonomy in which the algorithms are classified by the type of information they use to make predictions. The algorithms are organized in four main categories depending on whether they take as input the object characteristics, the sequence of requests, the HTTP headers and/or the content of web pages. This classification is proposed because the input of the predictors determine what kind of accesses can be predicted and how complex the analysis is. For instance, a recently added link can be predicted by an algorithm using as input the sequence of requests only if at least one user has already followed that link. The main contribution of this paper is to present a comprehensive review of web prediction algorithms grouped by a novel taxonomy, thus easing the understanding of how predictors work and the implications of selecting a given 2 prediction algorithm when designing a new prefetching system. In other words, this paper is aimed at examining the pros and cons of each group of algorithms and identifying current and future trends in web prediction engines. The remainder of the paper is organized as follows. Section 2 describes the elements that compose the web prefetching architecture, highlighting the func- tion of the prediction algorithm. Section 3 classifies web prediction algorithms according to the proposed taxonomy. Finally, Section 4 draws some concluding remarks. 2 Web prefetching architecture Web prefetching systems are implemented by extending the generic web archi- tecture (with clients, servers and proxies) with two extra elements: the predic- tion and the prefetching engines. The function of the prediction engine is to guess future client actions, e.g., requests. Its implementation corresponds to a prediction algorithm whose output is a hint list. In most proposals, this hint list includes references to the predicted objects, although some few approaches suggest other type of actions [12]. These predictions (or hints) are usually made basing on previous experience about user accesses and preferences. These hints are provided to a prefetching engine, which is aimed at preprocessing (e.g., downloading) those objects predicted by the prediction engine. By preprocess- ing the requests in advance, the user-perceived latency is reduced when these objects are requested by users. The amount and scope of the information that can be used by the prediction engine depends on the element of the Web architecture (the client, the proxy or the server) at which the prediction engine is located. When the prediction engine is located at the client, the algorithm can only use access patterns, con- tents seen and preferences coming from one single user. Consequently, those predictions are only useful for this user. If the prediction engine is located at a proxy server, it can take advantage of the multi-user and multi-server informa- tion gathered at this element to perform the predictions. When the predictor is located at the server side (origin server or replica in the context of Content Dis- tribution Networks (CDN)), it makes predictions based on multi-user accesses and preferences to the same web site. This option has been the most explored in the research literature because of the accuracy of the predictions made and its potential use in real scenarios. Finally, in more complex proposals, the pre- dictions can be performed by several elements in collaboration. This benefits the accuracy and coverage of the predictions. The prefetching engine is independent from the predictor and can be located in any element of the web architecture that receives the hint list. However, the common trend is to locate it at the client side as it is currently implemented in Mozilla-based browsers. The prefetched objects are stored in a cache until they are demanded or evicted. Therefore, only those objects that can be stored at the web client must be predicted or prefetched. This is not a drawback because, although a web page can be the result of a dynamic response, many of the objects composing that page are usually static and consequently cacheable. Furthermore, a dynamically generated response can be cached if its headers properly mark the response. As all the cacheable objects can be prefetched (precached), prefetching techniques 3 Figure 1: Web prediction algorithm taxonomy cover, among other technologies, dynamic web server and application program- ming, and browser application programming like AJAX, Java, Flash, and other Rich Internet Applications. The performance of prediction algorithms is usually measured by means of metrics that quantify both the efficiency and the efficacy of the proposal. Although some proposals use some specific indexes, the general trend when evaluating prediction algorithms is to trade-off Recall and Precision, as main performance metrics. Recall is defined as the ratio of requested objects that were previously predicted. As the recall quantifies the weight of the predicted objects over the amount of objects requested by the user, it can give an idea about the usefulness or coverage of the predictions. Precision is defined as the ratio of good predictions to the number of predictions, so it is related to the quantity of wasted resources. 3 Prediction engine taxonomy This section presents the proposed taxonomy, which breaks down the prediction algorithms into four categories according to the information used to make the predictions. Earlier proposals simply consider isolated object popularity, while some oth- ers analyze the sequence of user requests to discover user navigation patterns. To discover the structure of web pages and the relation between requests, more re- cent algorithms take into account the HTTP headers or examine page contents. Content-based algorithms may involve the analysis of the list of hyperlinks, the text surrounding hyperlinks or the labels with meta-information included by the web designer. This taxonomy, represented in Figure 1, summarizes the prediction algo- rithms ranging from the earliest works in 1995 to the most recent ones found in 2010. The algorithms were designed considering the characteristics of their contemporary web sites and internet technologies. In this sense, this survey represents a review that covers more than 15 years of research following the evolution of the Web, from the genesis when static pages had few embedded ob- jects and hyperlinks to the current web, with dynamic and personalized pages. 4 3.1 Object characteristics Most of the algorithms falling in this category were proposed early in the Web. In those years, more than one decade ago, the Web structure was much simpler than nowadays. As a consequence, and in the absence of other approaches, the most intuitive and simplest algorithms were proposed. These proposals focus on analyzing object characteristics, like temporal lo- cality and frequency of accesses (i.e., popularity), which were the first aspects studied in the Web. Algorithms falling in this category usually follow the long- term approach, where clients are subscribed to proxies storing specific objects, which are in charge of keeping updated copies of the objects. The opposite is referred to as the short-term approach, which aims to prefetch those objects that are likely to be referenced in the near future, given user’s recent accesses. The work by Bestavros et al. [5] provides an overview of the performance benefits of propagating information from servers to proxies that are closer to their clients. No algorithm is described but only some ideas on which the al- gorithm should be based. The suggested algorithm takes as input the rate of transferred bytes among clients and proxies and the cumulative distribution of accesses to the objects as well. This idea was deeply worked in other work by the same author [6], which explores the performance gains of propagating information from servers to proxies. That is, the information is replicated at proxies which are closer to the final users. The level of propagation is based on the popularity of the objects and on the expected reduction in traffic. This approach exploits the properties of temporal and geographical locality of ref- erences, which, unlike the previous proposals, require a large community (e.g., thousands) of users. The Top-10 approach proposed by Markatos and Chronaki [33], also pub- lished in the last century, is one of the first attempts to exploit the web doc- uments’ popularity for prefetching purposes. Additional information obtained from the client side is also used to adapt prefetching to different client profiles. In this approach, the web server publishes a list of the most popular objects and this information is complemented with client access preferences to predict future user accesses. This approach has been included in this category since the main idea of the proposal is the use of the object popularity to predict future accesses, although client profile information is also used to control and adapt prefetching. In [26], Jian et al. propose an algorithm based on prefetching objects exhibit- ing a given characteristic. In this work, three main characteristics are explored and quantified: object lifetime, probability of being accessed, and user request arrival rate. The algorithm prefetches those objects where the estimated val- ues for a given characteristic exceed an specific threshold. The proposal works among clients and, either proxies or servers. The server collects information characterizing the objects and makes it available to the prefetching engine as well. Wu et al. [47] propose an algorithm aimed at improving a given performance metric, e.g., hit ratio, bandwidth consumption, or a ratio of both (H/B). De- pending on the target metric, the algorithm prefetches objects showing some particular characteristic (e.g., popularity, longer update intervals or a mix of both). The best performance is achieved by the algorithm which maximizes the H/B metric. The approach assumes that web servers push fresh copies into web 5 caches (assuming unlimited cache size) or proxy servers whenever such copies are updated. In [46], Venkataramani et al. propose a long-term approach where proxy caches are subscribed to those popular objects that are updated with low fre- quency. This approach is aimed at being used in the last level caches of a CDN hierarchical structure while the typical prefetching (i.e., short-term approach) should remain in the lower level caches. In [30], Lau and Ng propose a prefet- ching algorithm that analyzes the browser’s cache according to the long-term approach. The algorithm predicts and prefetches Web documents and their linked documents as well. Two main characteristics are analyzed for each ob- ject: its access rate and the last time it was accessed. As in [26], those objects whose characteristic exceeds a given threshold are prefetched. The algorithm uses the detection theory to determine the threshold value. In addition, users can take an active part in the prefetching engine by using a graphic user interface where they can modify any computed prefetching schedule. The work by Rangarajan et al. [42] can be characterized as a short-term approach. This paper presents a technique that dynamically groups users based on their frequency of access to each object. In this way, the server structure can be reorganized according to the needs of each group of users. To avoid overloading the network, the proposal predicts in a cluster-based fashion instead of predicting for each single user; that is, whenever a host connects to the server or a proxy, the prefetching strategy returns the URIs predicted for the cluster to which the host belongs. Unlike the previous proposals, in the work by Duchamp [21], servers manage not only statistics about the objects they store but also about objects in other servers. All this information is then summarized and a probability of being used is estimated. The prediction algorithm takes information both from the server and client sides. Once the prediction is received, the client side can interact to decide whether or not to prefetch the hinted hyperlinks according to some local parameters. Algorithms in this category made sense when they were proposed (i.e., early in the Web), since the Web was simple and static regarding the object’s point of view as well as its lifetime. Nowadays, this approach fits better to CDNs, which control both the origin and replica contents. 3.2 Sequence of requests Most of the web prediction algorithms proposed in the literature lie in the anal- ysis of the paths followed by each client, that is, in the sequence of requests. By comparing the user’s recent accesses to the paths previously recorded, algo- rithms are able to predict future accesses. Algorithms falling into this category require to distinguish which client is issuing the request in order to keep track of the right sequence of accesses. Al- though solving this issue is trivial when the prediction engine is implemented in the web browser, it becomes more complex when it is implemented in the proxy (keeping track of the clients IP could be enough) or in the server (frequently requiring the use of cookies). The element of the web architecture (i.e., client, proxy or server) in which the prediction engine is implemented is also a performance limiting factor for prediction engines of this type [19, 29]. This is because predictors falling in this 6 category cannot predict accesses to those resources that have not been requested before to the element where the predictor is placed (first-seen resources). As expected, the ratio of first-seen resources is low when predictors are located at the server but high when located at the client [19]. However, this limitation can be overcome if additional sources of information are provided to the prediction engine. Therefore, algorithms in this category have been divided into pure and hybrid, depending on whether the predictor considers only the sequence of requests or also takes into account additional information to refine predictions. 3.2.1 Pure predictors There is a large number of proposals specifically addressed to predict subse- quent accesses from the sequence of requests that a web server receives. Some variations of Markov models are often used to predict and hint resources that other element of the web architecture (i.e., the proxy or the clients) will prefetch [36, 13, 4, 38]. The dependency graph (DG) proposed by Padmanabhan and Mogul [36] is usually taken as baseline for performance comparisons. The graph has a node for every object that has ever been accessed. There is an arc from node A to B if and only if at some point in time a client accessed B within w accesses after A, where w is the lookahead window size. The weight of the arc is the ratio of the number of accesses to B within a window after A to the number of accesses to A itself. Some other algorithms [43, 31] use the predictions to preprocess resources at the server side. Schechter et al. [43] use session paths to make predictions for generating some dynamic content in advance. In the context of a distributed web server cluster, the proposal of Lee et al. [31] combines a prediction algorithm with a workload distribution algorithm to prefetch objects from disk in each web server. Kim et al. [28] propose a predictor that takes into account the sequence of requests from the client perspective. This sequence is used to predict the most likely link in the current page that the user will follow. Analyzing the sequence of requests at the client side has as main advantage that no accesses are hidden due to the effects of caching. However, since it considers single user information, it can only predict previously followed links, but not accesses to new content. Focusing on the proxy perspective, Fan et al. [22] employ the well-known prediction by partial matching (PPM) algorithm in such a way that it aggregates all users data to make predictions for each single user in a client-proxy context. The joint benefits of caching and prefetching are also analyzed by Teng et al. [45] in a proxy-server context. Its prediction algorithm is based on rules, each one with a confidence level, in which the next requested object depends on the sequence of objects recently requested. In contrast, in the proposal of Huang et al. [25], the prediction engine tries to make a prediction from the same user’s previous accesses. Only if it is not possible, the algorithm uses the aggregate information of other users. Since the range of accesses that are recorded is high, an important issue when the prediction engine is located at the proxy is to reduce the resource consump- tion. To deal with this concern, Yang et al. [49] propose a data structure to record web sessions in which it is possible to apply any data mining algorithm. The proposed algorithm works in two steps: first, user sessions are clustered; second, transition probabilities between pages in the same cluster are computed. 7 In a similar way, the algorithm proposed by Pallis et al. [37] identifies clusters of pages whose accesses are correlated as its first step to make predictions. Other approaches take into account the sequence of requests from several points of view to complement the sequences perceived in one element. In this context, requests that hit in a proxy cache are hidden for the origin server. This fact distorts the sequence of requests at the server and prevents the server from making a prediction for that request. Chen et al. [9] deal with this problem by proposing an algorithm that coordinates servers and proxies to make predictions. Another coordinated approach is taken by Pons [39] in the context of a web application, but considering coordination between proxies and clients. In this proposal, the proxy builds a generic prediction model that each client receives and adapts according to its own patterns. Some other prediction algorithms use the sequence of accesses as the only input, but they are not proposed to work in a specific element of the architecture. To improve PPM prediction accuracy, Ban et al. [3] propose to include extra data in the model to describe a node prediction capability. Chen and Zhang [8] propose a memory-efficient version of a PPM algorithm whose prediction tree branches have different height depending on the root object popularity. A similar algorithm, using Markov models of different orders, is proposed by Dongshan and Junyi [20] to deal with the scalability of the prediction engine. The model presented by Gunduz and Ozsu [24] compare session similarities to make predictions. Sessions are modeled taking into account the sequence of user requests and the time spent in each page. The work by Bonino et al. [7] describes an initial set of predictors (repre- sented as finite-state machines) that evolve according to genetic operators. One of these predictors is selected every time a request is received in order to make a prediction. The graph representing user sessions is different from those usually employed in Markov models, since requests are represented by edges, and nodes represent predictions instead of objects. 3.2.2 Hybrid predictors Predictors falling into this subcategory are those that complement the sequence of requests with another source of information, e.g., the HTTP headers or page contents. The main reasons to take into account extra sources are usually either to increase the prediction accuracy or to control the computational complexity of the algorithms. Three of the prediction algorithms analyzed by Zukerman et al. [50] are based on Markov models that employ the sequence of requests at the server as the main source. One of them, the Linked Space-Time Markov model, com- bines the sequence of accesses with the referrer information found in the HTTP headers. In a later work [2], they combine all the models to build a single predictor. The Double Dependency Graph (DDG) prediction algorithm (see Section 3.3) proposed by Domenech et al. [18] employs the sequence of user accesses to build a request dependency graph similar to the DG algorithm proposed in [36]. This predictor also requires to analyze response headers to distinguish container from embedded objects. The prediction engine proposed by Georgakis and Li in [23] (described in Section 3.4) also takes as input the sequence of requests received by the server. 8 However, it is more interested in the sequence of content read than in the path followed by the user. Nanopoulos et al. [35] propose an algorithm that builds association rules from the sequence of accesses, allowing only those page transitions found in the link structure of the web site. In a similar way, the proposal by Mabroukeh and Ezeife [32] builds a Markov model to make the predictions. However, in this case, the prediction tree is pruned by using the concept of semantic distance. To compute this distance, a domain ontology, which is provided during the design of the web site, is required. 3.3 HTTP headers This category includes those prediction algorithms that use the HTTP headers either as the only input or in combination with other inputs. Web client requests are issued by sending HTTP requests, which contain several headers providing information about request details, besides the URI. One of the most used request headers is the “Referer”. Although this request header is optional in the standard protocol, the most popular web browsers always provide it in their requests. A prediction algorithm can discover first- order dependences by only observing (Requested Object URI, Referrer URI) pairs. Together with the requested object, the web server returns some HTTP response headers. By looking at the response headers, a prediction algorithm can take some features of the requested objects as input. Response headers used in prediction algorithms include “Content-Type” and “Content-Length” headers. Zukerman et al. [50] compare several Markov-based prediction models that take into account different factors: the sequence of documents requests, the web site structure inferred by using the “Referer” request header, and a combina- tion of both. Their results show that the model that combines all these factors in the Markov model yields the best predictive power, in terms of time and space consumption. The same authors propose, in a later work [2], the max- Hybrid prediction algorithm, which uses as input the sequence of accesses with timestamp, the “Referer” request header, and the “Content-Length” response header. The DDG algorithm [18], classified and introduced in the previous section, employs the “Content-Type” response header to distinguish between container and embedded objects. Container objects are those that the user demand ex- plicitly, while secondary objects are those embedded in the page and requested automatically by the web browser. This allows DDG to differentiate the depen- dences of objects of the same page and objects of different pages. Thus, DDG represents the structure of the web site more faithfully than DG. The Referrer Graph (RG) prediction algorithm proposed by Ossa et al. [16] is based on DDG. Both algorithms use the “Content-Type” response header to classify objects as primary (container) and secondary (embedded) objects. However, unlike DDG, which uses the temporal sequence of accesses, RG uses the URI and the “Referrer” request header provided in each individual object request to build the arcs in the graph. In this way, RG builds arcs only between objects that are really linked with hyper references in the web site, and not 9 between objects that have been requested sequentially by a web browser. This allows RG to represent the structure of a web site more faithfully than DDG. HTTP headers can be considered accurate since they are provided explicitly by the web client and server. In other words, there is no derived guessing that might be erroneously interpreted as happens in the sequence of requets category. As it is unnecessary to access the full object content, these algorithms allow shorter processing time and are less intrusive than those in the content category. The headers are provided continually, and this allows the algorithms to actively capture user accesses. 3.4 Content Prediction algorithms that use the information extracted from the content of the web pages as main input are included in this category. These prediction engines use the current web page to make predictions of future accesses, usually without requiring history information. The algorithms in this category are specially designed to deal with objects that are newly created or never visited before, and with dynamic sites where a URL could have time-dependent contents. Despite the wide variety of proposals falling into this category, we have identified three trends when using page content as main input information. 3.4.1 Using hyperlinks The first approach proposed only used the references to objects found when parsing the current web page, without any further analysis. Chinen and Ya- maguchi [10] designed and implemented a prefetching proxy server called the WWW Collector (Wcol) by using this idea. This proxy prefetches and stores referenced pages, and also shares prefetched objects with its other clients. Cohen and Kaplan describe in [12] three simple prefetching techniques (pre- resolving host-names, pre-connecting and pre-warming) that can be used with non-prefetchable URLs. Predictions are performed by extracting URLs from anchor links of previously visited pages. The study considers two classes of requests. The first class includes pages linked in the results page of a search engine or web portal, whereas the second class contains all requests that were not preceded by a request to the server in the last 60 seconds. More recently, in the context of CDNs, Sidiropoulos et al. [44] identify clus- ters of “correlated” web pages in a site (web site communities), and make these communities the basic outsourcing unit of content to be prefetched. The ap- proach is based only on the hyperlink information, since they assume that two pages are “similar” or “correlated” if there exists a link between them. The algorithm by Nanopoulos et al. [35], described in Section 3.2, also falls into this category because it complements the prediction model obtained with the sequence of requests with the hyperlink graph of the web site. This graph is applied to the prediction model to allow only those page transitions that are included in it. 3.4.2 Semantic approaches Algorithms in this subcategory tend to capture the user surfing interest from semantic characteristics of visited pages. 10 A keyword-based semantic prefetching approach is proposed by Xu and Ibrahim [48]. This algorithm lies in the fact that client surfing is often guided by some keywords in the text that surrounds hyperlink definitions (hrefs) in web pages. Therefore, this approach predicts future requests based on semantic preferences of past retrieved objects. To do so, the algorithm ranks the links of the last visited page according to the similarity of the keywords found in their “anchor text” to those in the previously accessed links. More recently, Georgakis and Li [23] propose an algorithm that also considers the anchored text around each of the outbound hyperlinks. To assign a weight to each of these links, the algorithm keeps counters of the frequencies of appearance of bigrams for visited and non-visited links. Davison [14] proposes an algorithm that predicts the user’s next web page request by conducting an analysis of the text of his/her recently requested web pages. The algorithm compares the text in and around the hypertext anchors of the current page to the text contained in previously visited pages. 3.4.3 Labeled approaches This subcategory includes all those proposals where the web designer is in- volved in the prediction process by including extra information that will be later used to predict future accesses. The common advantage of these specific and designer-dependent proposals is that they exhibit interesting results for the environments where they are tested, that is, those for which the proposals have been developed. This is also their main disadvantage, because they are not generic solutions that could be easily adapted to general and real scenarios. Khan and Tao identify in [27] four dominant patterns in hyperlink structure (Chain, Tree, Complete Graph, and Tree with Complete Core). Their pro- posal consists on labeling links with this kind of information to develop smarter prefetching techniques when the structure of webspace is also brought into con- sideration. Pons [40] proposes to conduct web prefetching by using semantic links. That is, semantic information explicitly embedded during the web design and associ- ated with each web page link. A semantic link is different from a hyperlink, since the semantic link represents a pointer with a type or meaning directed from the predecessor web page to the successor web page. Examples of semantic link types are sequential, similar-to, cause-effective, implication, etc. The proposal of Mabroukeh [32] has been previously classified into the group of algorithms using the sequence of request as input because this is the main information used for prediction. Nevertheless, it is actually a hybrid algorithm that also employs semantic information to prune the Markov model according to a computed semantic distance. This proposal assumes that the semantic information is available in the form of domain ontology, provided during the design of the web site. 4 Conclusions Since the inception of web prefetching, the proposals of prediction algorithms have adapted to the increasing complexity of the web content. Thus, the com- plexity of the predictors has increased, too. In this paper, we have summarized 11 Table 1: Surveyed papers by category and year CATEGORY 1995-2000 2001-2005 2006-2010 Obj. characteristics [6, 5, 33, 21] [26, 46, 30, 42] [47] Seq. requests pure [4, 36, 43, 13, 22, 38] [8, 20, 7, 24, 28, 49, 9, 25, 39, 45] [31, 3, 37] hybrid [2, 50] [35] [23, 32, 18] HTTP headers [2, 50] [16, 18] Content [10] [12, 14, 27, 35, 48] [23, 40, 44, 32] and categorized more than 15 years of research in prediction algorithms ap- plied to web prefetching techniques. To this end, a new taxonomy based on the input information used by the prediction algorithms has been proposed and successfully applied. Through the proposed taxonomy, the pros and cons of each category of algorithms have been discussed. Table 1 summarizes the classification of the surveyed papers according to the taxonomy proposed in this work and the year of publication. This table permits us to identify trends in algorithm design. As one can observe, dur- ing the first period of five years, most of the proposed algorithms dealt only with object characteristics or sequence of requests to make predictions. In the next five-year period, the predominant choice was to use only the sequence of accesses as input of the prediction algorithms, although the number of content- based proposals also increased. Finally, during the last five years, researchers have moved towards more sophisticated inputs. Algorithms considering several input types as well as predictors that analyze web page contents or require the collaboration of the web page designer are gaining relative importance in the current trend. Furthermore, it has also been observed the increasing concern for the computational costs of the algorithms. Despite of this, there are several gaps in the design space of prediction algo- rithms that remain unexplored. Few papers, i.e., [21, 9, 39], analyze several web architecture points of view simultaneously, and none of them combine different input categories. Besides, the use of HTTP headers has not been intensively explored and no algorithm combining header and content analysis has been proposed. As for future work, we plan to classify other aspects of web prefetching systems: proposals of prefetching engines according to its location in the web architecture; and proposals of web prefetching architectures according to how prediction and prefetching engines are coordinated, thus providing a complete study about web prefetching techniques from its beginnings to current days. Acknowledgments This work has been partially supported by Spanish Ministry of Science and Innovation under grant TIN2009-08201, Generalitat Valenciana under grant GV/2011/002 and Universitat Politecnica de Valencia under grant PAID-06- 12 10/2424. References [1] Silvana Vanesa Aciar, Christian Serarols-Tarres, Marcelo Royo-Vela, and Josep Lluis de la Rosa i Esteva. Increasing effectiveness in e-commerce: rec- ommendations applying intelligent agents. International Journal of Busi- ness and Systems Research, 1(1):81–97, 2007. [2] David W. Albrecht, Ingrid Zukerman, and Ann E. Nicholson. Pre-sending documents on the WWW: A comparative study. In Proceedings of the Six- teenth International Joint Conference on Artificial Intelligence, Stockholm, Sweden, 1999. [3] Zhijie Ban, Zhimin Gu, and Yu Jin. A ppm prediction model based on stochastic gradient descent for web prefetching. In Proceedings of the IEEE 22nd International Conference on Advanced Information Networking and Applications (AINA), pages 166–173, Okinawa, Japan, 2008. [4] Azer Bestavros. Using speculation to reduce server load and service time on the WWW. In Proceedings of the 4th ACM International Conference on Information and Knowledge Management, Baltimore, USA, 1995. [5] Azer Bestavros. Speculative data dissemination and service to reduce server load, network traffic and service time in distributed information systems. In Proceedings of the Twelfth International Conference on Data Engineering, pages 180–187, New Orleans, USA, 1996. [6] Azer Bestavros and C. Cunha. Server-initiated document dissemination for the WWW. IEEE Data Engineering Bulletin, 1996. [7] Dario Bonino, Fulvio Corno, and Giovanni Squillero. A real-time evolu- tionary algorithm for web prediction. In Proceedings of the IEEE/WIC International Conference on Web Intelligence, Halifax, Canada, 2003. [8] Xin Chen and Xiaodong Zhang. Popularity-based PPM: An effective web prefetching technique for high accuracy and low storage. In Proceedings of the International Conference on Parallel Processing, Vancouver, Canada, 2002. [9] Xin Chen and Xiaodong Zhang. Coordinated data prefetching for web contents. Computer Communications, 28:1947–1958, 2005. [10] K. Chinen and S. Yamaguchi. An interactive prefetching proxy server for improvement of www latency. Proceedings of Seventh Annual Conference of the Internet Society (INET’97), 06/1997 1997. [11] Wei Chu and Seung-Taek Park. Personalized recommendation on dynamic content using predictive bilinear models. In Proceedings of the 18th Inter- national Conference on World Wide Web, pages 691–700, Madrid, Spain, 2009. 13 [12] Edith Cohen and Haim Kaplan. Prefetching the means for document transfer: a new approach for reducing web latency. Computer Networks, 39(4):437–455, 2002. [13] Edith Cohen, Balachander Krishnamurthy, and Jennifer Rexford. Efficient algorithms for predicting requests to web servers. In Proceedings of the IEEE INFOCOM ’99 Conference, New York, USA, 1999. [14] Brian D. Davison. Predicting web actions from HTML content. In Proceed- ings of the 13th ACM Conference on Hypertext and Hypermedia, College Park, USA, 2002. [15] Brian D. Davison. Learning web request patterns. In Web Dynamics - Adapting to Change in Content, Size, Topology and Use, pages 435–460. Springer, 2004. [16] Bernardo de la Ossa, Ana Pont, Julio Sahuquillo, and J. A. Gil. Referrer graph: a low-cost web prediction algorithm. ACM, 2010. [17] Josep Domenech, Jose A. Gil, Julio Sahuquillo, and Ana Pont. Speculative validation of web objects for further reducing the user-perceived latency. In Proceedings of the 9th International IFIP TC 6 Networking Conference, Chennai, India, 2010. [18] Josep Domenech, Jose A. Gil, Julio Sahuquillo, and Ana Pont. Using current web page structure to improve prefetching performance. Computer Networks, 54(9):1404 – 1417, 2010. [19] Josep Domènech, Julio Sahuquillo, José A. Gil, and Ana Pont. The impact of the web prefetching architecture on the limits of reducing user’s perceived latency. In Proceedings of the 2006 IEEE / WIC / ACM International Conference on Web Intelligence, Hong Kong, China, 2006. [20] Xing Dongshan and Shen Junyi. A new Markov model for web access prediction. Computing in Science and Engineering, 4(6):34–39, 2002. [21] Dan Duchamp. Prefetching hyperlinks. In Proceedings of the 2nd USENIX Symposium on Internet Technologies and Systems, Boulder, USA, 1999. [22] Li Fan, Pei Cao, Wei Lin, and Quinn Jacobson. Web prefetching between low-bandwidth clients and proxies: Potential and performance. In Proceed- ings of the ACM SIGMETRICS Conference on Measurement and Modeling Of Computer Systems, pages 178–187, Atlanta, USA, 1999. [23] A Georgakis and H Li. User behavior modeling and content based specu- lative web page prefetching. Data & Knowledge Engineering, 59:770 – 788, 12/2006 2006. [24] Sule Gunduz and M. Tamer Zsu. A web page prediction model based on click-stream tree representation of user behavior. In KDD ’03: Proceed- ings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535–540, New York, USA, 2003. ACM Press. 14 [25] Yin-Fu Huang and Jhao-Min Hsu. Mining web logs to improve hit ratios of prefetching and caching. In Proceedings of the 2005 IEEE / WIC / ACM International Conference on Web Intelligence, Compiegne, France, 2005. [26] Yingyin Jiang, Min-You Wu, and Wei Shu. Web prefetching: Costs, bene- fits and performance. In Proceedings of the 7th International Workshop on Web Content Caching and Content Distribution, Boulder, USA, 2002. [27] Javed I. Khan and Qingping Tao. Exploiting webspace organization for ac- celerating web prefetching. In Proceedings of the IEEE/WIC International Conference on Web Intelligence, Halifax, Canada, 2003. [28] Yuna Kim and Jong Kim. Web prefetching using display-based predic- tion. In Proceedings of the IEEE/WIC International Conference on Web Intelligence, Halifax, Canada, 2003. [29] Thomas M. Kroeger, Darrell D.E. Long, and Jeffrey C. Mogul. Exploring the bounds of web latency reduction from caching and prefetching. In Proceedings of the 1st USENIX Symposium on Internet Technologies and Systems, Monterey, USA, 1997. [30] Kelvin Lau and Yiu-Kai Ng. A client-based web prefetching manage- ment system based on detection theory. In Proceedings of the Web Con- tent Caching and Distribution: 9th International Workshop (WCW 2004), pages 129–143, Beijing, China, 2004. [31] Heung Ki Lee, Gopinath Vageesan, Ki Hwan Yum, and Eun Jung Kim. A proactive request distribution (prord) using web log mining in a cluster- based web server. In Proceedings of the International Conference on Parallel Processing (ICPP’06), Columbus, USA, 2006. [32] Nizar R. Mabroukeh and Christie I. Ezeife. Semantic-rich markov models for web prefetching. In Proceedings of the IEEE International Conference on Data Mining Workshops, pages 465 – 470, Miami, FL, USA, 2009. IEEE, IEEE. [33] Evangelos Markatos and Catherine Chronaki. A top-10 approach to pre- fetching on the web. In Proceedings of the INET’ 98, Geneva, Switzerland, 1998. [34] Bamshad Mobasher. The Adaptive Web, volume LNCS 4321, chapter Data Mining for Web Personalization, pages 90–135. Springer, 2007. [35] Alexandros Nanopoulos, Dimitrios Katsaros, and Yannis Manolopoulos. A data mining algorithm for generalized web prefetching. IEEE Trans. Knowl. Data Eng., 15(5):1155–1169, 2003. [36] Venkata N. Padmanabhan and Jeffrey C. Mogul. Using predictive pre- fetching to improve World Wide Web latency. Computer Communication Review, 26(3):22–36, 1996. [37] G Pallis, A Vakali, and J Pokorny. A clustering-based prefetching scheme on a web cache environment. Computers & Electrical Engineering, 34:309 – 323, 07/2008 2008. 15 [38] Themistoklis Palpanas and Alberto Mendelzon. Web prefetching using par- tial match prediction. In Proceedings of the 4th International Web Caching Workshop, San Diego, USA, 1999. [39] A Pons. Improving the performance of client web object retrieval. Journal of Systems and Software, 74:303 – 311, 02/2005 2005. [40] A Pons. Semantic prefetching objects of slower web site pages. Journal of Systems and Software, 79:1715 – 1724, 12/2006 2006. [41] Michael Rabinovich and Oliver Spatscheck. Web Caching and Replication. Addison-Wesley, 2002. [42] S.K. Rangarajan, V.V. Phoha, K.S. Balagani, R.R. Selmic, and S.S. Iyen- gar. Adaptive neural network clustering of web users. Computer, 37:34 – 40, 04/2004 2004. [43] Stuart Schechter, Murali Krishnan, and Michael D. Smith. Using path profiles to predict http requests. In Proceedings of the 7th International World Wide Web Conference, Brisbane, Australia, 1998. [44] Antonis Sidiropoulos, George Pallis, Dimitrios Katsaros, Konstantinos Sta- mos, Athena Vakali, and Yannis Manolopoulos. Prefetching in content distribution networks via web communities identification and outsourcing. World Wide Web, 11:39 – 70, 3/2008 2008. [45] Wei-Guang Teng, Cheng-Yue Chang, and Ming-Syan Chen. Integrating web caching and web prefetching in client-side proxies. IEEE Transactions on Parallel and Distributed Systems, 16(5):444–455, 2005. [46] Arun Venkataramani, Praveen Yalagandula, Ravindranath Kokku, Sadia Sharif, and Mike Dahlin. The potential costs and benefits of long-term prefetching for content distribution. Computer Communications, 25:367– 375, 2002. [47] Bin Wu and Ajay D. Kshemkalyani. Objective-optimal algorithms for long- term web prefetching. IEEE Transactions on Computers, 55(1):2–17, 2006. [48] Cheng-Zhong Xu and T.I. Ibrahim. A keyword-based semantic prefetching approach in internet news services. IEEE Transactions on Knowledge and Data Engineering, 16:601 – 611, 05/2004 2004. [49] Qiang Yang, Joshua Zhexue Huang, and Michael Ng. A data cube model for prediction-based web prefetching. Journal of Intelligent Information Systems, 20(1):11–30, 2003. [50] Ingrid Zukerman, David W. Albrecht, and Ann E. Nicholson. Predict- ing users’ requests on the WWW. In Proceedings of the 7th International Conference on User Modeling, Banff, Canada, 1999. 16 
work_jt4hkqjjjvf6rll3zjt3z4qbwi	----	Context-aware systems: A literature review and classification Expert Systems with Applications 36 (2009) 8509–8522 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Context-aware systems: A literature review and classification Jong-yi Hong *, Eui-ho Suh, Sung-Jin Kim POSMIS Lab, Industrial Management Engineering Building, Pohang University of Science Technology, (790-784) San 31, Hyoja-dong, Nam-gu, Pohang, Kyungbuk, South Korea a r t i c l e i n f o Keywords: Context-aware systems Classification framework Literature reviews 0957-4174/$ - see front matter Crown Copyright � 2 doi:10.1016/j.eswa.2008.10.071 * Corresponding author. Tel.: +82 54 279 5920; fax E-mail addresses: xman@postech.ac.kr (J.-y. Hong Suh), sciv@postech.ac.kr (S.-J. Kim). a b s t r a c t Nowadays, numerous journals and conferences have published articles related to context-aware systems, indicating many researchers’ interest. Therefore, the goal of this paper is to review the works that were published in journals, suggest a new classification framework of context-aware systems, and explore each feature of classification framework. This paper is based on a literature review of context-aware sys- tems from 2000 to 2007 using a keyword index and article title search. The classification framework is developed based on the architecture of context-aware systems, which consists of the following five lay- ers: concept and research layer, network layer, middleware layer, application layer and user infrastruc- ture layer. The articles are categorized based on the classification framework. This paper allows researchers to extract several lessons learned that are important for the implementation of context- aware systems. Crown Copyright � 2008 Published by Elsevier Ltd. All rights reserved. 1. Introduction Emerging ubiquitous or pervasive computing technologies offer ‘anytime, anywhere, anyone’ computing by decoupling users from devices (Dey, 2001; Hill et al., 2004; Kwon, Choi, & Park, 2005; Kwon, Yoo, & Suh, 2005; Schilit, Adams, & Want, 1994). To provide adequate service for the users, applications and services should be aware of their contexts and automatically adapt to their changing contexts-known as context-awareness (Bolchini, Schreiber, & Tan- ca, 2007; Dey, 2001; Zhu, Mutka, & Ni, 2005). Context is very important, since it provides information about the present status of people, places, things and devices in the environment (Korpipää, Mäntyjärvi, Kela, Keränen, & Malm, 2003; Kwon, 2004). Context is any information that can be used to characterize the situation of an entity. An entity is a person, place, or object that is considered rel- evant to the interaction between a user and an application, includ- ing location, time, activities, and the preferences of each entity (Dey, 2001). Context-awareness means that one is able to use con- text information. A system is context-aware if it can extract, inter- pret and use context information and adapt its functionality to the current context of use (Byun & Cheverst, 2004). The term context- aware computing is commonly understood by those working in context-aware, where it is felt that context is a key in their efforts to disperse and transparently weave computer technology into our lives. One goal of context-aware systems is to acquire and utilize information on the context of a device in order to provide services that are appropriate to the particular people, place, time, event, etc. 008 Published by Elsevier Ltd. All r : +82 54 279 2870. ), ehsuh@postech.ac.kr (E.-h. These systems aim to provide context-aware access to information, communication and computation. In late 1980s, there was a period of beginning activity on con- text-aware computing. A few of context-aware computing has met the interest. However, the activity seems to be increasing dra- matically. Nowadays, to overcome new challenges and require- ments found in context-aware systems, many researchers have made efforts to design and implement network, user infrastructure and middleware which effectively provide users with context- aware services. Numerous articles of journals and conferences have published research related to context-aware systems. In other words, many people are interested in context-aware systems. Therefore, we feel that this is a good time for a review analysis, since it has been over a year since many papers were published. Currently, it is difficult to compare articles, because the available research is published in quite different journals. Accordingly, the main objectives of this review are: � To classify and summarize research relevant for context-aware systems. � To provide a conceptual framework for the integration and clas- sification of articles. � To derive suggestions for context-aware systems researchers based on the literature review. Chen and Kotz (2000) surveyed the literature related with con- text-aware computing in mobile computing. They defined the terms context and context awareness, listed the context-aware applications that have been built, discussed approaches to sense and model the context, and looked into supporting infrastructures, security and privacy issues. However, in 2000, articles published in ights reserved. mailto:xman@postech.ac.kr mailto:ehsuh@postech.ac.kr mailto:sciv@postech.ac.kr http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Potentially related articles: 1419 Articles retrieved for further review: 423 Articles included in this paper: 237 Exclusion of articles based on the selection criteria: 186 Exclusion of articles based on titles/abstracts/full texts: 996 Potentially related articles: 1419 Articles retrieved for further review: Articles included in this paper: Exclusion of articles based on the selection criteria: Exclusion of articles based on the selection criteria: Exclusion of articles based on titles/abstracts/full texts: Exclusion of articles based on titles/abstracts/full texts: Fig. 1. The procedures to select the articles. 8510 J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 journal were not enough to analyze context-aware systems. Fur- thermore, a classification standard does not involve all fields of context-aware systems. Therefore, a new survey research on con- text-aware systems is needed. Baldauf, Dustdar, and Rosenberg (2007) suggested the abstract layer architecture for context-aware systems and introduced various existent middleware and server- based approaches to ease the development of context-aware appli- cations, where each layer is explained based on various research. However, suggestions for further study, literature review related with each layer, introducing framework of many research and pre- senting current research issues are not broad. The remainder of this paper is structured as follows. In Section 2, a methodology to extract the literature is illustrated. Section 3 presents general characteristics of the literature research. Section 4 illustrates the proposed classification framework. Section 5 de- scribes each feature of classification framework (concept and re- search, network, user infrastructure, middleware, application and service). Section 6 contains discussions and suggestions. Finally, Section 7 concludes the paper with brief concluding remarks. 2. Procedures A total of 237 articles from 2000 to 2007 were obtained and re- viewed. Articles were found via computerized search of the topic areas. The search was narrowed using the terms context-aware. A detail illustration of methodology for extracting articles is followed. 2.1. The selection criteria The selection criteria were used to select and accept the con- text-aware articles. If the papers did not meet the selection criteria, then they were excluded. The three criteria are described as follows: � The literature was based on a search in the descriptor for ‘con- text-aware’. Various online journal databases were selected to search context-aware literatures. Context-aware articles are found in comprehensive subjects such as computer science, engineering, etc. Subjects of online database which are summa- rized according to fields of context-aware systems are shown in Table 1. � This paper surveys the articles published from 2000 to 2007. The reason for selecting this time period is that many journals and conferences have published researches related to context-aware computing since 2000. Table 1 The online databases and subjects. Online database Subjects IEEE Xplore Biomedical communication and information communication Networks/systems Computational intelligence Computer engineering Computer graphics Control systems Distributed computing/real-time systems Socio-political aspects of technology Software design development Wiley InterScience Computer science engineering � This paper covers only journal articles. Other publication forms (The conference proceedings, unpublished working papers, mas- ter’s and doctoral dissertations, newspapers and books, etc.) were not included. Furthermore, LNCS (Lecture Notes of Com- puter Science) is excluded, because of numerous articles having contents of computer science. The reason for choosing the jour- nal is that both practitioners and academicians use journals fre- quently to obtain knowledge and spread their study findings. Thus, journal articles indicate the highest level of research. 2.2. Data sources and procedures to extract articles The papers were selected according to the procedures shown in Fig. 1. First of all, the articles were searched using six online dat- abases. The total number of articles is 1419. The number of articles by each online database is as follows: Science Direct (288), IEEE Xplore (100), Springer Link Online Libraries (43), ACM Digital Li- brary (200), Wiley InterScience (94), Ingenta Journals (421), and EBSCO (Electronic Journal Service) (273). Next, 996 articles were excluded because they did not have the word of context-aware systems in the titles and abstracts. Next, the articles were carefully reviewed to select those that considered context-aware as the core part. Two hundred and thirty seven articles remained because 186 articles did not meet the second selection criteria. Based on these procedures, a total of 237 articles met all the selection criteria. Online database Subjects Science Direct Business, management and accounting Computer science Decision science Engineering Psychology Social science EBSCO (Electronic Journal Service) Computer science Library science Mathematics ACM Digital Library Computer science Engineering Social sciences Ingenta Journals Computer science Education Engineering Springer Link Online Libraries Computer science Engineering Fig. 2. Classification of articles by publication year. Table 3 Classification of articles based on the journal. Journal articles Number of articles IEEE Pervasive Computing 23 Personal and Ubiquitous Computing 10 IEEE Internet Computing 6 Wireless Personal Communications 5 IEEE Intelligent Systems 5 Mobile Networks and Applications 5 IEEE Transactions on Software Engineering 4 The Others 139 Expert Systems with Applications 10 Computer Communications 6 Journal of Systems and Software 6 Pervasive and Mobile Computing 5 World Wide Web 5 IEEE Wireless Communications 5 Interacting with Computers 4 Total 237 J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 8511 3. The general characteristics of the articles 3.1. Classification of articles by publication year The number of articles by publication year is depicted in Fig. 2. Numerous context-aware articles have grown considerably since 2000. The number of articles in 2007 have becomes 7 times more than the number of articles in 2000. The number of articles from 2001 to 2004 had been increased continuously, and the number of articles from 2004 to 2006 has been almost the same. It is abso- lute that the concern about context-aware systems was increased and will be continued. 3.2. Classification of articles by online database The article by online database is categorized in Table 2. There are a total of 181 articles from online databases. We use 7 online databases to search the articles. In Table 2, IEEE Xplore is the high- est percentage of articles (74 articles, 31%), because it offers arti- cles of many journals (IEEE Pervasive Computing, IEEE Internet Computing, IEEE Wireless Communications and IEEE Transactions on Software Engineering) which have subject relevant to context- aware systems. Science Direct stores journal articles of various study fields, therefore this database is the second highest of articles (60 articles, 25.6%). Other online databases are Springer Link On- line Libraries (57 articles, 23.9%), Ingenta Journals (18 articles, 8%), ACM Digital Library (16 articles, 7%), Wiley InterScience (9 articles, 4%) and EBSCO (Electronic Journal Service) (3 articles, 1%). When the same article is found repetitively among different online database, the article number of database having more num- ber of articles is increased. 3.3. Classification of articles by journal The articles by journal articles are categorized in Table 3. There were 65 journals that published context-aware articles. Most of Table 2 Classification of articles by publication year. Online database Number of articles IEEE Xplore 74 Science Direct 60 Springer Link Online Libraries 57 Ingenta Journals 18 ACM Digital Library 16 Wiley InterScience 9 EBSCO (Electronic Journal Service) 3 Total 237 them were related to the computer science, electronic engineering and information management. Table 3 specifies journals that pub- lished four or more context-aware articles and the others that pub- lished only one article or no more than 3 related with context- aware systems are omitted. Pervasive computing and ubiquitous computing is closely connected with context-aware computing. Therefore, journals which focus on pervasive computing or ubiqui- tous computing have published articles relevant to context-aware systems. As Table 3 shows, IEEE Pervasive Computing published the most articles on context-aware systems (23 articles, 10%). Ex- pert Systems with Applications and Personal and Ubiquitous Com- puting published 10 articles (4%). Distribution of articles according to the journal shows that various journals have published articles relevant to context-aware systems. 4. Classification framework 4.1. Abstract architecture of context-aware systems The abstract architecture of context-aware systems is drawn based on the literature that explores the context-aware prototype, systems, and application to offer classification criteria for dividing the literature appropriately. To make the abstract architecture, SO- CAM, ACAI, NAMA, PeCAN, X-CAF, CyberDesk, WebPADS, CAPIAs, Hycon service framework, Culliver’s Genie, Intelligent Agent framework, context-aware agent architecture, reference frame- work for multi-target user interfaces, etc. are reviewed in detail. As the results of literature review related with context-aware architecture, we present general abstract layer architecture of con- text-aware systems. As Fig. 3 shows, this architecture consists of four layers: (1) network layer involves a network supporting con- text-aware systems and sensor collecting low-level of context information; (2) middleware layer manages processes and stores context information; (3) based below layers, application layer pro- vides users with appropriate service; and (4) to offer suitable inter- Fig. 3. Abstract layer architecture of context-aware systems. 8512 J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 face to users, interface of context-aware systems is managed in user infrastructure layer. 4.2. Classification framework of context-aware systems The classification framework is developed for classifying the lit- eratures related with context-aware systems, based on the abstract architecture of context-aware systems. The classification frame- work developed here consists of the following five layers: concept and research layer, network layer, middleware layer, application layer and user intrastate layer. Without the concept and research layer, layers of classification framework and layers of context- aware systems architecture are the same. As Fig. 4 shows, each layer has detailed categories for dividing the literatures in layers. Fig. 4 depicts a graphical classification framework for context- aware systems. Context-aware applications and services require technology from the foundation of wireless network, user infra- structure, and middleware. Moreover, application and technology related with network, middleware and user infrastructure are based on concept and research activities. Concept and research layer involves overview, algorithm, development guideline, frame- work, context data management, evaluation and privacy and secu- rity categories. Network layer consists of protocol, sensing, network requirement and network implementation. Middleware layer is classified as agent-based middleware, metadata based mid- dleware, tuple space based middleware, OSGI based middleware, reflective middleware and sensor selection middleware. User infra- structure layer is divided into interface and usability categories. Application and service layer consists of smart space, tour guide, information systems, communication systems, m-commerce and web service. Concept & Research Framework Evaluation Cont Dat Mg Overview Algorithm Network Infrastructur Protocol Network imp leme Middleware Lay Agent based Metadata based Ada Obje ba Tup le Sp ace based User Infrastructure User Interface Application Laye Smart Sp ace Tour Guide IS Commu Sys Fig. 4. Classification framework 5. Result of article classification The articles by subjects are categorized in Tables 4–9. Table 4 shows the number of articles, percentage of subject, and percent- age of all subjects, while Tables 5–9 represent all references of the context-aware systems articles. The context-aware concept and research has the highest per- centage of context-aware articles (87 articles, 37%). In context- aware concept and research, the category can be divided into 7 subjects. Most of the articles are related to ‘‘development guide- line”, ‘‘context data management” (13 articles) and ‘‘algorithm” (28 articles). The ‘‘algorithm” is divided into 3 categories, namely ‘‘agent” (3 articles), ‘‘context reasoning” (14 articles) and ‘‘service recommendation” (10 articles). There are many articles related to guidelines, context data management and algorithm to develop the context-aware application and systems. Other topics are ‘‘over- view” (8 articles), ‘‘framework” (10 articles), ‘‘privacy and security” (12 articles), and ‘‘evaluation” (3 articles). The second highest per- centage of context-aware articles is related to context-aware appli- cation (63 articles, 26%), which has six categories. The majority of this section is related to the ‘‘m-commerce” (19 articles)”. The ‘‘smart spaces” (11 articles) are the second portion of context- aware application and are composed of ‘‘home” (3 articles), ‘‘hospi- tal” (7 articles) and ‘‘class room” (1 article). Other topics are ‘‘tour guide” (6 articles) and ‘‘communication systems” (10 articles), ‘‘information systems” (7 articles) and ‘‘Web Service” (9 articles). Network infrastructure has the third highest percentage of con- text-aware articles (41 articles, 17%). The articles are mostly re- lated to the ‘‘internet protocol” (11 articles), ‘‘handoff management” (7 articles), ‘‘Sensing” (13 articles), ‘‘network requirement” (4 articles) and ‘‘network implementation” (5 arti- Develop - ment Guideline ext a t Privacy & Security e Layer Sensing ntation er Network requirement Reflective based OSGI based ptive ctive sed Layer Usability r nication tems M-commerce Web Service of context-aware systems. Table 4 Classification of articles by subject. Classification criteria Number of articles Percentage of subject (%) Percentage of all subjects (%) 1. Concept and research 87 100 36.6 1.1. Algorithm 28 32.2 11.8 1.1.1. Agent 4 4.6 1.7 1.1.2. Context reasoning 14 16.1 5.9 1.1.3. Service recommendation 10 11.5 4.2 1.2. Development guideline 13 14.9 5.5 1.3. Framework 10 11.5 4.2 1.4. Context data management 13 14.5 5.5 1.5. Evaluation 3 3.4 1.3 1.6. Privacy and security 12 13.8 5.0 1.7. Overview 8 9.2 3.4 2. Network infrastructure 41 100 17.2 2.1. Protocol 11 26.8 6 2.2. Handoff management 7 17.1 2.9 2.3. Sensing 13 31.7 5.5 2.4. Network requirement 4 9.8 1.7 2.5. Network implementation 6 14.6 2.8 3. Middleware 31 100 13.0 3.1. Agent-based middleware 11 35.5 4.6 3.2. Reflective middleware 5 16.1 2.1 3.3. Metadata based middleware 2 6.5 0.8 3.4. Tuple space based Middleware 3 9.7 1.3 3.5. Adaptive and objective based middleware 6 19.4 2.5 3.6. OSGI based middleware 4 12.9 1.7 4. User infrastructure 16 100 6.7 4.1. Interface 13 81.3 5.5 4.2. Usability 3 18.8 1.3 5. Application 63 100 26.5 5.1. Smart space 11 17.5 4.6 5.1.1. Home 3 4.8 1.3 5.1.2. Hospital 7 11.1 2.9 5.1.3. Class room 1 1.6 0.4 5.2. Tour guide 6 9.5 2.5 5.3. Information systems 7 12.7 3.0 5.4. Communication systems 10 15.9 4.2 5.5. m-Commerce 19 30.2 8.0 5.6. Web service 9 14.3 3.8 Total 237 100 J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 8513 cles). The articles related to the middleware have 31 articles (13%). The category of this middleware part can be also divided into six subjects. The articles are related to ‘‘agent-based middleware” (11 articles), ‘‘reflective middleware” (5 articles), and ‘‘adaptive and objective based middleware” (6 articles). The fewest of con- text-aware articles is related to user infrastructure (16 articles, 7%). The user infrastructure can be divided into two parts, ‘‘inter- face” (13 articles) and ‘‘usability” (3 articles). In result, the concept and research has the highest percentage of context-aware articles, in which the most explored subject is the algorithm. This represents that the context-aware studies are in developing stages and much research have been conducted in this area. The fewest of context-aware articles is related to user infrastructure. Fig. 5 shows the number of distribution of articles by the year and classification framework. Articles of concept and research layer and application and service layer have been increased con- tinuously by increasing the total number of articles from 2000 to 2007. Almost Articles which have issues of privacy and security are published in 2005 and 2007. It means that the concern about privacy and security has increased recently. Context-aware sys- tems can provide users with better service based on analyzing physical context and personal context such as user schedule, user preferences and so on. However, grasping user context causes risk of security and privacy relevant to personal informa- tion. Articles which focus on context reasoning are distributed almost from 2002 to 2004, because many algorithms for finding user context such as Bayesian network algorithm, decision tree, case based reasoning and rule based reasoning has been already developed. Approximately 65% of articles involved in application layer were published from 2005 to 2007. In accordance with developing fields of the other layer, context-aware computing application and service have been implemented recently. How- ever, articles in user infrastructure, network infrastructure and middleware layer are decreased after the peak publishing point of each layer. For example, articles in middleware layer were in- creased until 2004, and then decreased until. This phenomenon was caused by recent research that covers several category of framework from concept to application. In other words, the ten- dency of recent research focuses not only on basic concept but also implantation or case-study. Therefore, articles of application category in our classification framework have been increased continuously. 5.1. Concept and research layer This represents the references according to the detailed catego- ries in concept and research layer in Table 5. ‘‘Context-aware con- cept and research” involves theories and foundation to construct context-aware systems. Algorithm is an essential part in develop- ing the context-aware systems and is divided into three sections such as agent, context reasoning, and service recommendation. The algorithm category has research related to an AI methodology that provides the foundation for a technology of intelligent sys- tems. Here, it is applied as three indexes within the concept and re- search layer: (1) designing and modeling algorithm, message bus encoded by XML, role and communication methods of agents, (2) increasing accuracy and efficiency of algorithm for extracting high-level context from low-level context, and (3) developing algo- rithm or method to recommend appropriate service according to the context of user. Context data management represents the framework to manage the context and context information. Devel- opment guideline describes the guideline related to the develop- ment of many kinds of context-aware applications and services. The framework of context-aware systems illustrates acquiring, dis- covering, interpreting and accessing various contexts to build con- text-aware services. The framework of context-aware systems is similar to an architectural sketch of a house. It gives users a general idea of what the systems will look like and how it will implement systems. The framework shows the general capabilities of the sys- tems, a user interfaces system functions, system (data) flow, sys- tem management, etc. The evaluation is focused on the evaluation method of the system that is developed in each paper. The overview deals with the general definition, characteristics and history of context-aware. Ontology is generally defined as a ‘representation of a shared conceptualization of a particular domain’, and further defined by ‘axioms that offer additional constraints and meaning as well as rules and heuristics that can derive additional useful information’. Ontology language has been developed in the Semantic Web and has converged to a new W3C standard called Web Ontology Lan- guage (OWL; Smith, Welty, & McGuinness, 2004). Semantic web metadata standards such as Resource Description Framework (RDF) and OWL provide standard representation for context data Table 5 References of concept and research layer. Classification criteria References Concept and research Algorithm Agent Arcos and Plaza (2002), Hattori, Cho, Ohsuga, Isshiki, and Honiden (2004), Kao and Yuan (2005), Krause, Smailagic, and Siewiorek (2006) Context reasoning Anagnostopoulos, Ntarladimas, and Hadjiefthymiades (2007), Brunato and Battiti (2005), Castro and Munz (2000), Chetan, Ranganathan, and Campbell (2005), Drakatos, Pissinou, Douligeris, and Makki (2007), Ladd, Bekris, Rudys, Kavraki, and Wallach (2005), Niemegeers and Heemstra De Groot (2005), Ranganathan, Al-Muhtadi, and Campbell (2004), Sajal, Nirmalya, and Abhishek (2006), Samaan and Karmouch (2005), Satoh (2003), Smailagic and Kogan (2002), Weber, Sorkine, Lipman, and Gotsman (2007), Xiong and Hassan (2006) Service recommendation Byun and Cheverst (2004), Hatala, Wakkary, and Kalantari (2005), Kocaballı and KoçyiğitGranular (2007), Kouadri and Younas (2007), Kwon and Kim (2004), Meeuwissen, Reinold, and LiemInferring (2007), Pertselakis, Ferles, Tsiolis, and Stafylopatis (2005), Soldatos, Stamatis, Azodolmolky, Pandis, and Polymenakos (2007), Syukur and Loke (2007), Yuan and Peng (2004) Development guideline Anhalt et al. (2001), Augustin et al. (2006), Avrahami, Gergle, Hudson, and Kiesler (2007), Bolchini et al. (2007), Casas, Cuartielles, Marco, Gracia, and Falco (2007), Henricksena and Indulska (2006), Hong and Landay (2001), Julien, Payton, and Roman (2004), Kwon (2004), Kwon, Choi, and Kim (2007), Qing, Károly, Christian, Paulo, and Bernhard (2006), Qiu, Chang, Lin, and Shi (2007), van Sinderen, van Halteren, Wegdam, Meeuwissen, and Eertink (2006) Framework Dix et al. (2000), Doulkeridis, Loutas, and Vazirgiannis (2006), Johnson (2007), Jones, Grandhi, Terveen, and Whittaker (2004), Kao and Yuan (2004), Panagiotakis and Alonistioti (2006), Rogers and Muller (2006), Roussaki, Strimpakou, and Anagnostou (2007), van Kranenburg, Bargh, Iacob, and Peddemors (2006), Vuković, Kotsovinos, and Robinson (2007) Context data management Abecker, Bernardi, Hinkelmann, Kühn, and Sintek (2000), Chaari, Ejigu, Laforest F., and Scuturici (2007), Demiris, Vlahakis, Makri, Papaioannou, and Ioannidis (2005), Korpipää et al. (2003), Liao, He, and Tang (2004), Reignier, Brdiczka, Vaufreydaz, Crowley, and Maisonnasse (2007), Roussaki, Strimpakou, and Pils (2007), Satoh (2007), Serrano and Serrat (2006), Simons and Wirtz (2007), Smith, Ma, and Ryan (2006), Qi and Venkatasubramanian (2007), Zimmermann, Specht, and Lorenz (2005) Evaluation Griswold et al. (2004), Morla and Davies (2004), O’Grady et al. (2005) Privacy and security Bhatti, Bertino, and Ghafoor (2005), Bhatti, Samuel, Eltabakh, Amjad, and Ghafoor (2007), Falkovych and Nack (2006), Hill et al. (2004), Jiang and Landay (2002), Jorns, Jung, and Quirchmayr (2007), Jutla, Bodorik, and Zhang (2006), Minami and Kotz (2005), Tentori, Favela, and Rodriguez (2006), Vassilis, Loukas, Dimitris, and Stavros (2006), Yee, Korba, Lin, and Shih (2006), Zhang, Qi, Zhao, and Hou (2007) Overview Brown and Randel (2004), Chalmers (2004), Dey (2001), Intille (2004), Lieberman and Selker (2000), Liechti (2000), Loke (2006), Prekop and Burnett (2003) Table 6 References of network layer. Classification criteria References Network infrastructure Internet protocol Benedetto and Nardis (2006), Boukerche, Pazzi, and Araujo (2006), Chen and Mohapatra (2005), Friday et al. (2003), Giaffreda, de Carvalho, Melia, Giaffreda, and de Carvalho (2007), Kellerer, Wagner, Balke, and Schulzrinne (2004), Lin, Tsai, and Lai (2005), Singh and Acharya (2005), Tyan and Mahmoud (2005), Yau and Karim (2003), Zhu et al. (2005) Handoff management Balasubramaniam and Indulska (2004), Bellavista, Corradi, and Foschini (2007), Marias, Prigouris, Papazafeiropoulos, Hadjiefthymiades, and Merakos (2004), Mcnair and Zhu (2004), Qing et al. (2006), Rocha and Endler (2006), Yang, Wu, Yang, and Liu (2007) Sensing Brilingaitė and Jensen (2007), Gellersen, Schmidt, and Beigl (2002), Gellersen, Kortuem, Schmidt, and Beigl (2004), Goulev, Stead, Mamdani, and Evans (2004), Gross, Egla, and Marquardt (2006), Jonker, Persa, Caarls, de Jong, and Lagendijk (2003), Knight et al. (2007), Matsushita (2001), Michahelles and Samulowitz (2004), Pfeifer (2005), Schmidt and Laerhoven (2001), Xu, Fu, Lee, and Winter (2007), Zhong and Gilbert (2005) Network requirement Dixit (2002), Harter, Hopper, Steggles, Ward, and Webster (2002), Kanter (2002), Yang et al. (2007) Network implementation Aguiar, Sarma, Bijwaard, Marchetti, and Pacyna (2007), Beigl, Gellersen, and Schmidt (2001), Cano, Ferrandez-Bell, and Manzoni (2005), Gonzalez-Castano, Garcia-Reinoso, Gil-Castineira, Costa-Montenegro, and Pousada-Carballo (2005), Riva, Nadeem, Borcea, and Iftode (2007), Wang et al. (2004) Table 7 References of middleware layer. Classification criteria References Middleware Agent-based middleware Alahuhta, Löthman, Helaakoski, Koskela, and Röning (2007), Bellavista, Corradi, Montanari, and Stefanelli (2006), Bellavista, Corradi, and Stefanelli (2002), Chen et al. (2004), Julien and Roman (2006), Khedr and Karmouch (2005), Khedr and Karmouch (2004), Kumar et al. (2003), Riekki et al. (2003), Román et al. (2002), Soldatos, Dimakis, Stamatis, and Polymenakos (2007) Reflective middleware Capra, Emmerich, and Mascolo C. (2003), Chan and Chuang (2003), Chan, Chuang, Cao, and Leong (2004), Qin, Shi, and Suo (2007), Sacramento et al. (2004) Metadata based middleware Bellavista, Corradi, Montanari, and Stefanelli (2003a), Bellavista, Corradi, Montanari, and Stefanelli (2003b) Tuple space based middleware Cabri, Leonardi, Mamei, and Zambonelli (2003), Curino et al. (2005), Murphy and Picco (2004) Adaptive and objective based middleware Alex, Kumar, and Shirazi (2008), Trumler, Bagci, Petzold, and Ungerer (2005), Yau and Karim (2004a), Yau and Karim (2004b), Yau et al. (2006), Wu et al. (2007) OSGI based middleware Choi, Shin, and Shin (2005), Gu, Pung, and Zhang (2005), Gu, Pung, and Zhang (2004), Yu, Zhou, Yu, Zhang, and Chin (2006) 8514 J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 in context-aware systems. In other words, ontology will play a cru- cial role in enabling the processing and sharing of information and knowledge on the middleware. The reasons are as follows: � Providing a shared and common understanding of a domain that can be communicated across people and application systems (Chen & Finin, 2003). Table 8 References of application and service layer. Classification criteria References Applications and services Smart space Home Baek, Lee, Lim, and Huh (2005), Intille (2002), Schulzrinne, Wu, Sidiroglou, and Berger (2003) Hospital Agarwal, Joshi, Finin, Yesha, and Ganous (2007), Bottazzi, Corradi, and Montanari (2006), Favela, Rodríguez, Preciado, and González (2004), Favela et al. (2007), Kjeldskov and Skov (2007), Muñoz, Rodriguez, Favela, Martinez-Garcia, and González (2003), Rodríguez, Favela, Martínez, and Muñoz (2004) Class room Sung et al. (2005) Tour guide Bellotti, Berta, De Gloria, and Margarone (2005), Cano, Manzoni, and Toh (2006), Cheverst, Mitchell, Davies, and Smith (2000), Cheverst, Smith, Mitchell, Friday, and Davies (2001), O’Hare and O’Grady (2003), Pashtan, Heusser, and Scheuermann (2004) Information systems Celentano and Gaggi (2006), Kwon (2006a), Kwon, Yoo, et al. (2005), Kwon, Yoo, and Suh (2006), Norrie et al. (2007), Signer, Grossniklaus, and Norrie (2007), Wilson, Doyle, Weakliam, Bertolotto, and Lynch (2007) Communication systems Fogarty, Lai, and Christensen (2004), Goularte, Pimentel, and Moreira (2006), Raento, Oulasvirta, Petit, and Toivonen (2005), Ranganathan, Campbell, Ravi, and Mahajan (2002), Schmidt, Takaluoma, and Mantyjarvi (2000), Schilit, Hilbert, and Trevor (2002), Sumi and Mase (2000), Sumi and Nishida (2001), Udugama, Kuladinithi, Gorg, Pittmann, and Tionardi (2007), Yuan and Chen (2007) M-commerce Anagnostopoulos, Tsounis, and Hadjiefthymiades (2007), Becerra-Fernandez, Cousins, and Weber (2007), Bouvin, Christensen, Frank, and Hansen (2003), Broens, Halteren, Sinderen, and Wac (2007), Chakraborty et al. (2007), Genco, Sorce, Reina, and Santoro (2006), Kwon (2003), Kwon and Sadeh (2004), Kwon, Shin, and Kim (2006), Maiden, Omo, Seyff, Grunbacher, and Mitteregger (2007), Mandato, Kovacs, Hohl, and Amir-Alikhani (2002), Riva and Toivonen (2007), Roussos, Marsh, and Maglavera (2005), Skov and Høegh (2006), Stylianos, Nikia, Kia, and Christos (2007), Syukur and Loke (2006),Wac, Halteren, Bults, and Broens (2007), Wohltorf, Cissée, and Rieger (2005), Yu et al. (2006) Web service Blake, Kahan, and Nowlan (2007), Debaty, Goddi, and Vorbau (2005), Ceri, Daniel, Facca, and Matera (2007), Gandon and Sadeh (2004), Kanter (2003), Kwon (2006b), Kwon, Choi, et al. (2005), Kwon, Yoo, et al. (2005), Lee (2007), Pashtan, Kollipara, and Pearce (2003) Table 9 References of user infrastructure layer. Classification criteria References User infrastructure Interface Alexander and Matthias (2006), Bell, Feiner, and Höllerer (2002), Calvary et al. (2003), Hatala and Wakkary (2005), Hong, Dickson, Chiu, Shen, and Kafeza (2007), Korhonen et al. (2007), Kurvinen, Lähteenmäki, Salovaara, and Lopez (2007), Lieberman and Chu (2007), Lum and Lau (2002), Mäntyjärvi and Seppänen (2003), Rehman, Stajano, and Coulouris (2007), Selker (2004), Smailagic and Siewiorek (2002) Usability Barnard, Yi, Jacko, and Sears (2005), Burrell and Gay (2002), Kaasinen (2003) Fig. 5. Distribution of articles number by year and classification framework. J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 8515 � Facilitating knowledge sharing and reuse in an open and dynamic distributed systems (Kwon, Choi, et al., 2005; Kwon, Yoo, et al., 2005; Gruber, 1993). � Deriving fresh knowledge and facts based on reasoning contex- tual data and information by using inference engines. � Allowing devices and agents not expressly designed to work together to interoperate, achieving ‘‘serendipitous interoperabil- ity” (Wang, Dong, Chin, Hettiarachchi, & Zhang, 2004). Although some research only focuses on ontology, the number is very few. Ontology is generally used for reasoning and express- ing context data. Therefore, they do not need to have a separate category for ontology. 5.2. Network layer As shown in Table 6, we categorize the network infrastructure layer into internet protocol, handoff management, sensing, net- work requirements and network implementation. Context-aware computing needs to dynamically adapt to changes in this context and connect entities based on network. Therefore, many researches are conducted to offer appropriate network for providing context- aware computing. Internet protocol category involves presenting mechanisms and design of session initiation protocol (SIP), mobile IPv6, an integral part of constructing self-configuring mobile ad hoc networks and object discovery. Handoff management category has article about the process of transferring an ongoing call or data session from one channel connected to the core network to an- other. Article topic of handoff management category is treating heterogeneous network for seamless communication environ- ments. The sensing category is responsible for capturing, abstract- ing and acquiring context information of the situations in the environment. When implementing context-aware systems, the ba- sic idea is to use the sensor data and predict the current situation. Therefore, the sensing category consists of sensing algorithm, sens- ing technology and wearable sensing mechanism. Network requirements category means collecting requirements for the adaptation method which can support seamless computing infra- structures, framework for integration for user services out to the 8516 J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 mobile devices and heterogeneous network infrastructures. Net- work implementation category involves realizing the network which can provide context-aware computing environments. Although many researches related with context-aware computing illustrate network layer, the research which are specially focused on network are only involved in this category. A part of research related with constructing context-aware systems involves an illus- tration of sensing. However, these researches do not belong to the network category, because network sensing is not their core part. 5.3. Middleware layer Table 7 indicates the references according to the detailed cate- gories in middleware layer. The context-aware systems can be made up of several basic parts. Almost all the middleware parts gather context information, process it and derive meaningful ac- tions from it. Context-aware systems must have middleware sup- port for providing customized services. The middleware allows agents to acquire contextual information easily, reason about it using different logics and then adapt themselves to changing con- texts. Therefore, the middleware facilitates the development of context-aware agents. There are many types of middleware. This paper divides the currently available middleware for context- aware systems into seven categories. The essential part of agent- based middleware is mobile agents. Mobile agents are regarded as a highly potential technology for many application areas. They emerge as a middleware technology suitable to develop context- aware services, and have been developed to implement the needed active infrastructure and the MA-based middleware design and implementation. Reflective middleware possesses the unique abil- ity to model itself through self-representation, such that the manipulation of its behavior may be changed. A middleware sys- tems with self-representation is causally connected if changes made to the self-representation directly affect the implementation of the middleware. The opposite is true if changes to the middle- ware implementation change the self-representation. Metadata provides information on users, devices resources, and the preferred binding strategies. Metadata based middleware facilities perform runtime binding management actions based on metadata and con- text and location visibility. Tuple spaces have been considered for use in sensor networks and as a blackboard with a mechanism that permits to retrieve partially known information. This is a shared repository of information, which can be accessed via well-defined primitives. Adaptive middleware is a middleware for context- aware applications in smart-home setups. In this scheme, the mid- dleware matches the Quality of Context (QoC) requirements with the QoC available with the sensors. The application’s QoC require- ments are mapped to a utility function using the QoC attributes of the sensors available. OSGi-based middleware reliably manages context-aware services to support context acquisition, discovery, and reasoning. The middleware using agents for managing complexity in the delivery of context-aware systems have recently emerged. Auton- omous, proactive, rational and social agents are engaged in oppor- tunistic collaboration in the solution of appropriate service, because context-aware systems are comprised of many dynami- cally interacting components. Numerous incarnations of agents are found from the reactive agent that responds in a stimulus re- sponse manner, the agent that minimizes the need for a complex representational context model, to the agent that support reason- ing in a collaborative context. 5.4. Application and service layer Table 8 indicates the references according to detailed categories in applications and services layer. There are many types of context- aware application, which provide the user with smart environment such as home, hospital, class room, etc. When a user wants to go to unfamiliar places such as museum, they play an important role as a tour guide. In addition, context-aware applications include infor- mation systems, especially decision support systems, communica- tion systems as social community, m-commerce, and web service. Based on the analysis of the literature review related with con- text-aware application and service, characteristics of context- aware computing application and service are extracted as follows: � Compute device to operate independently of human control. To provide the user with services autonomously, context-aware applications are customizable by using user context and prefer- ence. Automatic execution of service is conducted when in a cer- tain context. � Context-aware applications act in anticipation of future goals or problems. Context-aware applications not only handle current task, situation and action but also, anticipate future behavior, moving point and problem of user. � In context-aware application, various devices are interconnected among each other and they recognize each other on a certain distance, because proximal devices distance between user and device or between devices are important factors to offer appro- priate service to user. � Context-aware applications provide a rich set of capabilities and services to the nomad as she moves from places to place in a transparent and convenient form. 5.5. User infrastructure layer Table 9 shows the references according to the detailed catego- ries in user infrastructure layer. It is characteristic for handheld de- vices and their users that they are continuously moving in several simultaneous fuzzy contexts. The usefulness of a given context- aware systems is most likely to suffer from the small screen and a dynamic context of users. Mobility is at its core the very essence of context-awareness. The dynamic environment sets special requirements for usability and acceptance of context-aware sys- tems. Therefore, research of user interface (UI) and usability of handheld device are carried out. In UI category, a content adapta- tion system for overcoming small interface such as PDA and cellu- lar phone, user modeling and human–computer interaction for considering the emergence of ubiquitous and mobile computing environments are presented. In UI research, it is suggested that the contents were confusing when a screen changed due to a loca- tion change. Usability category involves investigating the user needs based on user interviews, field evaluations with users, and expert evaluations of context-aware services. Appearing context- aware devices and services introduces a new level of complexity when it comes to usability evaluation. Usability evaluation meth- ods have been measured against desktop environments. Some researchers investigated general guidelines for evaluating con- text-aware systems, while others have evaluated novel interfaces and ways of interacting. They suggested a hybrid evaluation meth- od comprised of cognitive walkthrough and heuristic evaluation for UI supporting mobility of context-aware systems. 6. Discussions and suggestions Context-aware systems are still developing in order to improve. Many researchers have been concerned about context-aware sys- tems and context-awareness, and as shown in Fig. 2, the concern has increased from 2000 to 2007. Moreover, reports or white paper of many project have been published increasingly. However, con- J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 8517 text-aware systems are not fully implemented in real life. As shown in Table 4, much research has focused on the concept and research layer. The scope of applications or services in most articles is limited to small regions; laboratory, school, hospital, smart room and so on. In addition, strategic alternatives or business models for gaining the revenue by using context-aware systems are very few. Furthermore, technologies related to context-aware systems are merely standardized. The architecture, method of context model- ing, inference algorithm, network implementation and devices of users in each project are different. Furthermore, middleware, applications and services make use of different level of contexts and adapt the way they behave according to the current context. Therefore, according to the level and type of contexts along with the goal of context-aware systems, the context modeling method- ology, inference algorithm, agent structure, and interaction meth- od of agent are changed. Although the interaction between agents in the same middleware and cooperation between compo- nents of the same architecture are investigated, a standard for interaction, cooperation and operation in the different context- aware application has not been studied. Finally, we found that con- text sensing, context managing and context-aware services and applications are included in ubiquitous computing environment. Context awareness is a key factor for new applications in the area of ubiquitous computing. The immediate future of context-aware application and service may increasingly lie in its everyday life; as researchers are begin- ning to realize the impossibility of developing systems and appli- cation only in laboratory environments. Therefore, possible future research agenda may include answers to the following questions. 6.1. How to extract and use the cognitive context in context-aware application? Most of the context-aware systems focus on the external con- text, called physical context. External context means context data collected by physical sensors. It involves context data of the phys- ical environment, location data, distance, function on to other ob- jects, temperature, sound, air pressure, time, lighting levels surrounding users, and so on. The external context is important, and very useful for context-aware systems, because context-aware systems provide recommended services for a person based on ana- lyzing the external data. However, to provide personalized services according to user preferences, task and emotional state of user, the cognitive domains, such as information retrieval, decision making, situation monitoring, and so on, are needed. However, a few authors have addressed utilizing the cognitive elements of a user’s context. Several researchers have proposed models to capture the internal elements of context. This model differs from many of the previous approaches, because it focuses on extracting a user’s cog- nitive activities, rather then extracting the user’s movement based on a physical environment. Cognitive context information is key information to satisfy user by providing personalized context- aware computing services. In the early stage of research, only the physical context information was collected to be aware of user’s context. Recently, some of the literatures which focus on cognitive context have been introduced. However, it is insufficient to estab- lish context-aware systems that reflect cognitive context. To pro- vide suitable and personalized service for the user, more studies need to be conducted in this area. 6.2. What are design patterns of context-aware systems? Design patterns are defined as recurring solutions to design problems you see over (Berczuk 1994). They focus on the reuse of recurring architectural design themes, and have a recurring de- sign problem that is generated in a specific domain and suggest a solution (Buschmann, Meunier, Rohnert, Sommerlad, & Stal, 1996). Design patterns from another format sharing knowledge such as development guidelines are distinguished by the following reasons. � To help designers address high-level problems as well as low- level ones by showing the hierarchical structure between design patterns. � To capture the essence of recurring problems and their solutions in a compact form. � To have many examples of actual designs, alternatives to apply the solution, and some of the tradeoffs in applying the solution. Design patterns have been proposed in many domains as a for- mat for capturing and sharing design knowledge between practi- tioners. They can also assist designers in developing context- aware systems. Design patterns of context-aware systems will pro- vide an effective way for sharing solutions to design problems and reuse prior design knowledge. Also, the designer of the context- aware systems will avoid dangerous design problem and easily solve the difficult recurring problems. Some effort has aimed at unearthing common design patterns for context-aware systems. However, suggested design patterns are not specific, nevertheless, general and very few research has been conducted. Therefore, re- search on design pattern of context-aware systems should be con- ducted, because there are many potential benefits in developing a design pattern. 6.3. Which is the best inferring algorithm to extract user context and provide service to user? Various algorithms used in context-aware systems are classified into two parts. First, algorithm is utilized to infer high-level con- text of user. According to levels of abstraction, context is divided into low-level context and high-level context. Low-level context is raw data collected directly from physical sensors, while high-le- vel context is inferred from low-level context. This part involves the algorithm of context reasoner to extract high-level context and inferring algorithm to extract correct position of user, near ob- ject, and environments. Second, algorithm is recommending suit- able service. According to the activity, location and preference of user, recommended services of context-aware systems are differ- ent. Although location is critical to determine service, user profile (such as sex, age, style, preference, and so on) also has influence on recommending service. Therefore, for increasing user satisfac- tion by recommending service that customer wants to receive, an inferring algorithm based on context and user profile is used. In the two parts, Bayesian networks for handling causal relationships between various events, probabilistic logic, fuzzy logic, decision tree, neural network and support vector machine (SVM) are ap- plied. Moreover, many researches focus on improving accuracy and efficiency of these algorithms. However, the research for com- paring these algorithms according to the context data type, amount of context data, research scenario and components of envi- ronments has not been carried out. Since finding the suitable ser- vice quickly and correctly from on low-level context is very important in limited computing environments, research related with the best inferring algorithm for extracting user context and providing the service with users should be conducted. 6.4. How to deal with concurrently enormous data, information and knowledge having different format to serve suitable service to users? The raw data of low-level context are usually gathered from dif- ferent physical sensors. Data type, formats and abstraction level 8518 J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 from different physical sensors are different. Devices and physical sensors of context-aware systems use various scale and unit, and low-level context has different elements. Context-aware systems store data, information and knowledge that have different relation- ship, format and abstraction level in the context base. Furthermore, context-aware systems collect context history storing sensor data over time to offer proactive service. Context history stores huge amount of data on location, temperature, lighting level, task, uti- lized devices, state of devices, selected services and so on. To quickly provide suitable service to users, context-aware systems should manage variety, diversity and numerous amount of context. However, previous research suggested only a concept to control this problem. Therefore, a methodology and real implication for treating variety, diversity and numerous mount of context are needed. 6.5. How to extract the best solution when the context of users is conflicted? Conflicts occur on context-aware computing environments when many users use the several physical sensors and service sources. Thus, the conflicts of context-aware systems are classified into sensing conflict, service resource conflict and user preference conflict. � Sensing conflict: Not matching results from several physical data sources. If the coordinates of GPS and spotting of a camera are different, then sensing conflict is generated. � Service resource conflict: Selecting only several users among many users that want to be provided with the same service, due to limited service resource. If context-aware systems offer services without possessing all the necessary service resources, then service resource conflict occurs. � User preference conflict: Providing only several users with per- sonalized service, because the preference of users is not the same, although the same context of users is identified. If family members like different TV programs at the same time, then pref- erence difference between family members leads to user prefer- ence conflict. The problem of conflicts has been approached by using informa- tion fusion, time stamps and fuzzy algorithm. However, it is still not solved perfectly, because it is difficult to provide personalized services under the limited resource. Therefore, a study on method- ology and actual guideline for harmonizing all conflicts are required. 6.6. How to reflect the preference of users for satisfying user needs? The preference of each user is different according to user con- text and profile. Predicting the preferences of users and providing the personalized services based on users’ preferences have carried out by various research. Especially, predicting the preferences of users in e-commerce users for providing the personalized services is one of the important issues. Recommendation systems providing personalized services have been carried out by many researchers. However, the research considering users’ preferences on Context- aware computing is a relatively insufficient in the context-aware systems research field (Byun & Cheverst, 2004). There were some limitations in previous research for providing the personalized ser- vices on context-aware systems. First, the users have to input their preferences directly. Second, they did not provide automated ser- vices. Finally, it is difficult to provide new user with the personal- ized services. Thus, it is difficult to provide the users with the automatic personalized services due to these limitations. Hence, proposing the context-aware systems to provide the personalized services based on context history in context-aware computing is needed. 6.7. How to save users information in context-aware systems? As Table 5 shows, the research has been conducted in the area of security and privacy. The context-aware systems that do not consider user privacy and security in systems that deal with per- sonal data will not be publicly accepted. Therefore, general and several approaches to implement security and privacy in context- aware systems have been studied. Although privacy and security are an important aspect of context-aware systems, applicable con- text-aware systems utilizing these approaches have not been investigated yet, due to their complexity. In other words, most re- searches have only focused the concept for protecting privacy and security. The more research related with privacy and security is needed for context-aware systems to become a reality. A user utilizing context-aware systems uses many available computers embedded in everyday artifacts, like PDA, navigation, and so on. Each user uses many personal computing devices, and at the same time, the same publicly available device is used by many users. Furthermore, context-aware systems store and handle sensitive and personal data. Common to context-aware systems is the lack of appropriate security and privacy mechanisms that can effectively ensure a secure user authentication. Therefore, the need for authentication arises as the information and data stored in con- text base is highly personal and should be protected. Hence, more research on context-aware systems that can be protected from unauthorized access is needed. 6.8. How to evaluate performance of context-aware systems? It is essential to improve managing and planning context-aware systems based on performance evaluation, because ‘‘what you measure is what you get”. Various models such as Balanced Score Card (BSC), IS Success Model, Control Objectives for Information and Related Technology (COBIT), Information Economics, and so on, have been applied to suggest measures that can evaluate the performance of information systems by considering diverse per- spectives. Due to the unique characteristics of context-aware sys- tems such as portability, mobility, proactiveness, nomadicity, and so on, the current technology for evaluating information systems (IS) does not provide techniques to formally define, verify, imple- ment, and analyze them. Therefore, research on the model and per- formance measure reflecting the characteristics of context-aware systems should be conducted. 7. Conclusions The review was organized according to the framework devel- oped with the purpose of providing a comprehensive overview of research on context-aware systems. The review introduced the context-aware systems concept, network infrastructure, middle- ware, application, user infrastructure and presented an exhaustive list of each layer of context-aware systems. In this paper, we re- viewed the literatures for the concept and applications, and exam- ined them using dimensions related to ongoing and emerging issues in context-aware systems. This paper is based on a literature review of context-aware systems from 2000 to 2007 using a key- word index and article title search. Overall, we found that the activity related with context-aware computing seems to increase dramatically and can be classified into designing and implement- ing network, user infrastructure on handheld device, middleware which are enough to provides users with context-aware services, concept and research being the base of context-ware application, J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 8519 and application and service. It provides sufficient literature for the researchers on the use of context-aware systems, and we hope that it will motivate the researchers and practitioners to examine con- text-aware systems issues and its applications. The framework of context-aware systems offers a general development guideline, fundamental component and relationships within component of context-aware systems. Furthermore, on the basis of this frame- work and the literature review, selected opportunities for future research were discussed. It is concluded that different social science methodologies, such as psychology, cognitive science, and human behavior could imple- ment context-aware computing. Integration of qualitative, quanti- tative and scientific methods and integration of context-aware theory and concept studies may broaden our horizon on this sub- ject. Finally, the ability to continually change and obtain new understanding is the power of context-aware computing. Although considerable attention was given to the framework de- sign and classification of the literature review, some limitations still exist. First, some relevant articles might have been overlooked. Much the literature has been found by the title, keyword or abstract only. Although the title in most cases describes the content quite well this is not always so. In order to conduct a comprehensive lit- erature review, the topical focus was kept relatively narrow on con- text-aware systems. Moreover, this paper makes a brief literature review of context-aware systems from 2000 to 2007 in order to ex- plore how context-aware systems have developed in this period. White paper, practical reporter and many articles related with con- text-aware systems were not reviewed, because seven online dat- abases were searched in this paper. Period and searching limitation may not satisfy the need of readers looking for a review on context-aware systems. Finally, the correlation between the framework and real projects of context-aware systems are not illus- trated. The review did not involve the role and interaction methods of logical components in architectures of SOCAM, CASS, CoBrA, GUIDE, Hydrogen, Gaia, Context Toolkit, STU21 project and so on, since only articles of journals were extracted and analyzed. References Abecker, A., Bernardi, A., Hinkelmann, K., Kühn, O., & Sintek, M. (2000). Context- aware, proactive delivery of task-specific information: The KnowMore project. Information Systems Frontiers, 2(3), 253–276. Agarwal, S., Joshi, A., Finin, T., Yesha, Y., & Ganous, T. (2007). A pervasive computing system for the operating room of the future. Mobile Networks and Applications, 12(2–3), 215–228. Aguiar, R. L., Sarma, A., Bijwaard, D., Marchetti, L., & Pacyna, P. (2007). Pervasiveness in a competitive multi-operator environment: The daidalos project. Communications Magazine, IEEE, 45(10), 22–26. Alahuhta, P., Löthman, H., Helaakoski, H., Koskela, A., & Röning, J. (2007). Experiences in developing mobile applications using the apricot agent platform. Personal and Ubiquitous Computing, 11(1), 1–10. Alex, H., Kumar, M., & Shirazi, M. (2008). MidFusion: An adaptive middleware for information fusion in sensor network applications. Information Fusion, 9(3), 332–343. Alexander, Z., & Matthias, J. (2006). Implementing adaptive mobile GI services based on ontologies: Examples from pedestrian navigation support. Computers, Environment and Urban Systems, 30(6), 784–798. Anagnostopoulos, C. B., Ntarladimas, Y., & Hadjiefthymiades, S. (2007). Situational computing: An innovative architecture with imprecise reasoning. Journal of Systems and Software, 80(12), 1993–2014. Anagnostopoulos, C. B., Tsounis, A., & Hadjiefthymiades, S. (2007). Context awareness in mobile computing environments. Wireless Personal Communications, 42(3), 445–464. Anhalt, J., Smailagic, A., Siewiorek, D. P., Gemperle, F., Salber, D., Weber, S., et al. (2001). Toward context-aware computing: Experiences and lessons. IEEE Intelligent Systems, 16(3), 38–46. Arcos, J. L., & Plaza, E. (2002). Context aware personal information a gents. International Journal of Cooperative Information Systems, 11(3–4), 245–264. Augustin, I., Yamin, A. C., da Silva, L. C., Real, R. A., Frainer, G., & Geyer, C. F. R. (2006). ISAMadapt: Abstractions and tools for designing general-purpose pervasive applications. Software: Practice and Experience, 36(11–12), 1231–1256. Avrahami, D., Gergle, D., Hudson, S. E., & Kiesler, S. (2007). Improving the match between callers and receivers: A study on the effect of contextual information on cell phone interruptions. Behaviour and Information Technology, 26(3), 247–259. Baek, S. H., Lee, H. J., Lim, S. Y., & Huh, J. D. (2005). Managing mechanism for service compatibility and interaction issues in context-aware ubiquitous home. IEEE Transactions on Consumer Electronics, 51(2), 524–528. Balasubramaniam, S., & Indulska, J. (2004). Vertical handover supporting pervasive computing in future wireless networks. Computer Communications, 27(8), 708–719. Baldauf, M., Dustdar, S., & Rosenberg, F. (2007). A survey on context-aware systems. International Journal of Ad Hoc and Ubiquitous Computing, 2(4), 263–277. Barnard, L., Yi, J. S., Jacko, J. A., & Sears, A. (2005). An empirical comparison of use-in- motion evaluation scenarios for mobile computing devices. International Journal of Human–Computer Studies, 62(4), 487–520. Becerra-Fernandez, I., Cousins, K. C., & Weber, R. O. (2007). Nomadic context-aware knowledge management systems: Applications, challenges and research problems. International Journal of Mobile Learning and Organisation, 1(2), 103–121. Beigl, M., Gellersen, H. W., & Schmidt, A. (2001). Media cups: Experience with design and use of computer-augmented everyday artefacts. Computer Networks, 35(4), 401–409. Bell, B., Feiner, S., & Höllerer, T. (2002). Visualization viewpoints. IEEE Computer Graphics and Applications, 23(2), 6–9. Bellavista, P., Corradi, A., & Foschini, L. (2007). Context-aware handoff middleware for transparent service continuity in wireless networks. Pervasive and Mobile Computing, 3(4), 439–466. Bellavista, P., Corradi, A., Montanari, R., & Stefanelli, C. (2003a). Context-aware middleware for resource management in the wireless internet. IEEE Transactions on Software Engineering, 29(12), 1086–1099. 2003. Bellavista, P., Corradi, A., Montanari, R., & Stefanelli, C. (2003b). Dynamic binding in mobile applications: A middleware approach. IEEE Internet Computing, 7(2), 34–42. 2003. Bellavista, P., Corradi, A., Montanari, R., & Stefanelli, C. (2006). A mobile computing middleware for location- and context-aware internet data services. ACM Transactions on Internet Technology, 6(4). Bellavista, P., Corradi, A., & Stefanelli, C. (2002). The ubiquitous provisioning of internet services to portable devices. IEEE Pervasive Computing, 1(3), 81–87. Bellotti, F., Berta, R., De Gloria, A., & Margarone, M. (2005). Implementing tour guides for travelers. Human Factors and Ergonomics in Manufacturing, 15(4), 461–476. Benedetto, M. D., & Nardis, L. D. (2006). Tuning UWB signals by pulse shaping: Towards context-aware wireless networks. Signal Processing, 86(9), 2172–2184. Berczuk, S. (1994). Finding solutions through pattern languages. IEEE Computer, 27(12), 75–76. Bhatti, R., Bertino, E., & Ghafoor, R. (2005). A trust-based context-aware access control model for web-services. Distributed and Parallel Databases, 18(1), 83–105. Bhatti, R., Samuel, A., Eltabakh, M. Y., Amjad, H., & Ghafoor, A. (2007). Engineering a policy-based system for federated healthcare databases. Knowledge and Data Engineering, IEEE, 19(9), 1288–1304. Blake, M., Kahan, D., & Nowlan, M. (2007). Context-aware agents for user-oriented web services discovery and execution. Distributed and Parallel Databases, 21(1), 39–58. Bolchini, C., Schreiber, F. A., & Tanca, L. (2007). A methodology for a very small data base design. Information Systems, 32(1), 61–82. Bottazzi, D., Corradi, A., & Montanari, R. (2006). Context-aware middleware solutions for anytime and anywhere emergency assistance to elderly people. Communications Magazine, IEEE, 44(4), 82–90. Boukerche, A., Pazzi, R. W. N., & Araujo, R. V. (2006). Fault-tolerant wireless sensor network routing protocols for the supervision of context-aware physical environments. Journal of Parallel and Distributed Computing, 66(4), 586–599. Bouvin, N. O., Christensen, B., Frank, K. G., & Hansen, A. (2003). HyCon: A framework for context-aware mobile hypermedia. New Review of Hypermedia and Multimedia, 9, 59–88. Brilingaitė, A., & Jensen, C. S. (2007). Enabling routes of road network constrained movements as mobile service context. GeoInformatica, 11(1), 55–102. Broens, T., Halteren, A. V., Sinderen, M. V., & Wac, K. (2007). Towards an application framework for context-aware m-health applications. International Journal of Internet Protocol Technology, 2(2), 109–116. Brown, B., & Randel, R. (2004). Building a context sensitive telephone: Some hopes and pitfalls for context sensitive computing. Computer Supported Cooperative Work, 13(3–4), 329–345. Brunato, M., & Battiti, R. (2005). Statistical learning theory for location fingerprinting in wireless LANs. Computer Networks, 47(6), 825–845. Burrell, J., & Gay, G. K. (2002). E-graffiti: Evaluating real-world use of a context- aware system. Interacting with Computers, 14(4), 301–312. Buschmann, F., Meunier, R., Rohnert, H., Sommerlad, P., & Stal, M. (1996). Pattern- oriented software architecture – A system of patterns. Wiley & Sons Ltd.. Byun, H. E., & Cheverst, K. (2004). Utilizing context history to provide dynamic adaptations. Applied Artificial Intelligence, 18(6), 533–548. Cabri, G., Leonardi, L., Mamei, M., & Zambonelli, F. (2003). Location-dependent services for mobile users. IEEE Transactions on Systems, 33(6), 667–681. Calvary, G., Coutaz, J., Thevenin, D., Limbourg, Q., Bouillon, L., & Vanderdonckt, J. (2003). A unifying reference framework for multi-target user interfaces. Interacting with Computers, 15(3), 289–308. 8520 J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 Cano, J. C., Ferrandez-Bell, D., & Manzoni, P. (2005). Evaluating bluetooth performance as the support for context-aware applications. Telecommunication Systems, 28(3–4), 333–347. Cano, J. C., Manzoni, P., & Toh, C. K. (2006). UbiqMuseum: A bluetooth and java based context-aware system for ubiquitous computing. Wireless Personal Communications, 38(2). Capra, L., Emmerich, W., & Mascolo, C. (2003). CARISMA: Context-aware reflective middleware system for mobile applications. IEEE Transactions on Software Engineering, 29(10), 929–945. Casas, R., Cuartielles, D., Marco, A., Gracia, H. J., & Falco, J. L. (2007). Hidden issues in deploying an indoor location system. IEEE Pervasive Computing, 6(2), 62–69. Castro, P., & Munz, R. (2000). Managing context data for smart spaces. IEEE Personal Communications, 7(5), 44–46. Celentano, A., & Gaggi, O. (2006). Context-aware design of adaptable multimodal documents. Multimedia Tools and Applications, 29(1). Ceri, S., Daniel, F., Facca, M. F., & Matera, M. (2007). Model-driven engineering of active context-awareness. World Wide Web, 10(4), 387–413. Chaari, T., Ejigu, D., Laforest, F., & Scuturici, V. M. (2007). A comprehensive approach to model and use context for adapting applications in pervasive environments. Journal of Systems and Software, 80(12), 1973–1992. Chakraborty, D., Dasgupta, K., Mittal, S., Misra, A., Gupta, A., Newmark, E., et al. (2007). Businessfinder: Harnessing presence to enable live yellow for small, medium and micro mobile businesses. Communications Magazine, IEEE, 45(1), 144–151. Chalmers, M. (2004). A historical view of context, a historical view of context. Computer Supported Cooperative Work (CSCW), 13(3), 223–247. Chan, A. T. S., & Chuang, S. N. (2003). MobiPADS: A reflective middleware for context-aware mobile computing. IEEE Transactions on Software Engineering, 29(12), 1072–1085. Chan, A. T. S., Chuang, S. N., Cao, J., & Leong, H. V. (2004). An event-driven middleware for mobile context awareness. The Computer Journal, 47(3), 278–288. Chen, H., & Finin, T. (2003). An ontology for a context aware pervasive computing environment. In Proceedings of IJCAI workshop on ontologies and distributed systems, IJCAI, <www.cs.vu.nl/~heiner/IJCAI-03/Papers/Chen.pdf>. Chen, G., & Kotz, D. (2000). A survey of context-aware mobile computing research. Dartmouth College Technical Report TR2000-381. Chen, H., Finin, T., Joshi, A., Kagal, L., Perich, F., & Chakraborty, D. (2004). Intelligent agents meet the semantic web in smart spaces. IEEE Internet Computing, 8(6), 69–79. Chen, H., & Mohapatra, P. (2005). A context-aware HTML/XML document transmission process for mobile wireless clients. World Wide Web, 8(4), 439–461. Chetan, S., Ranganathan, A., & Campbell, R. (2005). Towards fault tolerant pervasive computing. IEEE Technology and Society Magazine, 24(1), 38–44. Cheverst, K., Mitchell, K., Davies, N., & Smith, G. (2000). Exploiting context to support social awareness and social navigation. ACM SIGGROUP Bulletin, 21(3), 43–48. Cheverst, K., Smith, G., Mitchell, K., Friday, A., & Davies, N. (2001). The role of shared context in supporting cooperation between city visitors. Computers & Graphics, 25(4), 555–562. Choi, J. H., Shin, D. K., & Shin, D. I. (2005). Research and implementation of the context-aware middleware for controlling home appliances. IEEE Transactions on Consumer Electronics, 51(1), 301–306. Curino, C., Giani, M., Giorgetta, M., Giusti, A., Murphy, A. L., & Picco, G. P. (2005). Mobile data collection in sensor networks: The TinyLime middleware. Pervasive and Mobile Computing, 1(4), 46–469. Debaty, P., Goddi, P., & Vorbau, A. (2005). Integrating the physical world with the web to enable context-enhanced mobile services. Mobile Networks and Applications, 10(4), 385–394. Demiris, A. M., Vlahakis, V., Makri, A., Papaioannou, M., & Ioannidis, N. (2005). intGuide: A platform for context-aware services featuring augmented-reality, based on the outcome of european research projects. Signal Processing: Image Communication, 20(9–10), 927–946. Dey, A. K. (2001). Understanding and using context. Personal and Ubiquitous Computing, 5(1), 4–7. Dix, A., Rodden, T., Davies, N., Trevor, J., Friday, A., & Palfreyman, K. (2000). Exploiting space and location as a design framework for interactive mobile systems. ACM Transactions on Computer–Human Interaction (TOCHI), 7(3), 285–321. Dixit, S. (2002). Wireless IP and its challenges for the heterogeneous environment. Wireless Personal Communications, 22(2), 261–273. Doulkeridis, C., Loutas, N., & Vazirgiannis, M. (2006). A system architecture for context-aware service discovery. Electronic Notes in Theoretical Computer Science, 146(1), 101–116. Drakatos, S., Pissinou, N., Douligeris, C., & Makki, K. (2007). A context-aware cache structure for mobile computing environments. Journal of Systems and Software, 80(7), 1102–1119. Falkovych, K., & Nack, F. (2006). Context aware guidance for multimedia authoring: Harmonizing domain and discourse knowledge. Multimedia Systems, 11(3). Favela, J., Rodríguez, M., Preciado, A., & González, V. M. (2004). Integrating context- aware public displays into a mobile hospital information system. IEEE Transactions on Information Technology in Biomedicine, 8(3), 279–286. Favela, J., Tentori, M., Castro, L. A., Gonzalez, V. M., Moran, E. V., & Martínez-García, A. I. (2007). Activity recognition for context-aware hospital applications: Issues and opportunities for the deployment of pervasive networks. Mobile Networks and Applications, 12(2–3), 155–171. Fogarty, J., Lai, J., & Christensen, J. (2004). Presence versus availability: The design and evaluation of a context-aware communication client. International Journal of Human–Computer Studies, 61(3), 299–317. Friday, A., Wu, M., Finney, J., Schmid, S., Cheverst, K., & Davies, N. (2003). Network layer access control for context-aware IPv6 applications. Wireless Networks, 9(4), 299–309. Gandon, F. L., & Sadeh, N. M. (2004). Semantic web technologies to reconcile privacy and context awareness. Web Semantics: Science, Services and Agents on the World Wide Web, 1(3), 241–260. Gellersen, H., Kortuem, G., Schmidt, A., & Beigl, M. (2004). Physical prototyping with smart-its. IEEE Pervasive Computing, 3(3), 74–82. Gellersen, H. W., Schmidt, A., & Beigl, M. (2002). Multi-sensor context-awareness in mobile devices and smart artefacts. Mobile Networks and Applications, 7(3), 341–351. Genco, A., Sorce, S., Reina, G., & Santoro, G. (2006). An agent-based service network for personal mobile devices. IEEE Pervasive Computing, 5(2), 54–61. Giaffreda, F., de Carvalho, R., Melia, T., Giaffreda, F., & de Carvalho, T. (2007). Media Network enablers for enhanced mobility management. BT Technology Journal, 25(2), 95–105. Gonzalez-Castano, F. J., Garcia-Reinoso, J., Gil-Castineira, F., Costa-Montenegro, E., & Pousada-Carballo, J. M. (2005). Bluetooth-assisted context-awareness in educational data networks. Computers Education, 45(1), 105–121. Goularte, R., Pimentel, G. C. M., & Moreira, E. D. S. (2006). Context-aware support in structured documents for interactive-TV. Multimedia Systems, 11(4). Goulev, P., Stead, L., Mamdani, E., & Evans, C. (2004). Computer aided emotional fashion. Computers & Graphics, 28(5), 657–666. Griswold, W. G., Shanahan, P., Brown, S. W., Boyer, R., Ratto, M., Shapiro, R. B., et al. (2004). ActiveCampus: Experiments in community-oriented ubiquitous computing. IEEE Computer, 37(10), 73–81. Gross, T., Egla, T., & Marquardt, N. (2006). Sensation: A service-oriented platform for developing sensor-based infrastructures. International Journal of Internet Protocol Technology, 1(3), 159–167. Gruber, T. (1993). A translation approach to portable ontology specifications. Knowledge Acquisition, 5(2), 199–220. Gu, T., Pung, H. K., & Zhang, D. Q. (2004). Toward an OSGi-based infrastructure for context-aware applications. IEEE Pervasive Computing, 3(4), 66–74. Gu, T., Pung, H. K., & Zhang, D. Q. (2005). A service-oriented middleware for building context-aware services. Journal of Network and Computer Applications, 28(1), 1–18. Harter, A., Hopper, A., Steggles, P., Ward, A., & Webster, P. (2002). The anatomy of a context-aware application. Wireless Networks, 8(2–3), 187–197. Hatala, M., & Wakkary, R. (2005). Ontology-based user modeling in an augmented audio reality system for museums. User Modeling and User-Adapted Interaction, 15(3–4), 339–380. Hatala, M., Wakkary, R., & Kalantari, L. (2005). Rules and ontologies in support of real-time ubiquitous application. Web Semantics: Science, Services and Agents on the World Wide Web, 3(1), 5–22. Hattori, M., Cho, K., Ohsuga, A., Isshiki, M., & Honiden, S. (2004). Context-aware agent platform in ubiquitous environments and its verification tests. Systems and Computers in Japan, 35(7), 13–23. Henricksena, K., & Indulska, J. (2006). Developing context-aware pervasive computing applications: Models and approach. Pervasive and Mobile Computing, 2(1), 37–64. Hill, R., Al-Muhtadi, J., Campbell, R., Kapadia, A., Naldurg, P., & Ranganathan, A. (2004). A middleware architecture for securing ubiquitous computing cyber infrastructures. IEEE Distributed Systems Online, 5(9). Hong, D., Dickson, K. W., Chiu, V. Y., Shen, S. C. C., & Kafeza, E. (2007). Ubiquitous enterprise service adaptations based on contextual user behavior. Information Systems Frontiers, 9(4), 343–358. Hong, J. I., & Landay, J. A. (2001). An infrastructure approach to context-aware computing. Human–Computer Interaction, 16, 287–303. Intille, S. S. (2002). Designing a home of the future. IEEE Pervasive Computing, 1(2), 76–82. Intille, S. S. (2004). A new research challenge: Persuasive technology to motivate healthy aging. IEEE Transactions on Information Technology in Biomedicine, 8(3), 235–237. Jiang, X., & Landay, J. A. (2002). Modeling privacy control in context-aware systems using decentralized information spaces. IEEE Pervasive Computing, 1(3), 59–63. Johnson, S. (2007). A framework for mobile context-aware applications. BT Technology Journal, 25(2), 106–111. Jones, Q., Grandhi, S. A., Terveen, L., & Whittaker, S. (2004). People-to-people-to- geographical-places: The P3 framework for location-based community systems. Computer Supported Cooperative Work (CSCW)(3), 249–282. Jonker, P., Persa, S., Caarls, J., de Jong, F., & Lagendijk, I. (2003). Philosophies and technologies for ambient aware devices in wearable computing grids. Computer Communications, 26(11), 1145–1158. Jorns, O., Jung, O., & Quirchmayr, G. (2007). Transaction pseudonyms in mobile environments. Journal in Computer Virology, 3(2), 185–194. Julien, C., Payton, J., & Roman, G. C. (2004). Reasoning about context-awareness in the presence of mobility. Electronic Notes in Theoretical Computer Science, 97, 259–276. Julien, C., & Roman, G. C. (2006). EgoSpaces: Facilitating rapid development of context-aware mobile applications. IEEE Transactions on Software Engineering, 32(5), 281–298. Jutla, D. N., Bodorik, P., & Zhang, Y. (2006). PeCAN: An architecture for users’ privacy-aware electronic commerce contexts on the semantic web. Information Systems, 31(4–5), 295–320. http://www.cs.vu.nl/~heiner/IJCAI-03/Papers/Chen.pdf J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 8521 Kaasinen, E. (2003). User needs for location-aware mobile services. Personal and Ubiquitous Computing, 7(1), 70–79. Kanter, T. G. (2002). HotTown, enabling context-aware and extensible mobile interactive spaces. IEEE Wireless Communications, 9(5), 18–27. Kanter, T. G. (2003). Attaching context-aware services to moving locations. IEEE Internet Computing, 7(2), 43–51. Kao, T. H., & Yuan, S. M. (2004). Designing an XML-based context-aware transformation framework for mobile execution environments using CC/and XSLT. Computer Standards Interfaces, 26(5), 377–399. Kao, T. H., & Yuan, S. M. (2005). Automatic adaptation of mobile applications to different user devices using modular mobile agents. Software Practice and Experience, 35(14), 1349–1391. Kellerer, W., Wagner, M., Balke, W. T., & Schulzrinne, H. (2004). Preference-based session management for IP-based mobile multimedia signaling. European Transactions on Telecommunications, 15(4), 415–427. Khedr, M., & Karmouch, A. (2004). Negotiating context information in context- aware systems. IEEE Intelligent Systems, 19(6), 21–29. Khedr, M., & Karmouch, A. (2005). ACAI: Agent-based context-aware infrastructure for spontaneous applications. Journal of Network and Computer Applications, 28(1), 19–44. Kjeldskov, J., & Skov, M. B. (2007). Exploring context-awareness for ubiquitous computing in the healthcare domain. Personal and Ubiquitous Computing, 11(7), 549–562. Knight, J. F., Bristow, H. W., Anastopoulou, S., Baber, C., Schwirtz, A., & Arvanitis, T. N. (2007). Uses of accelerometer data collected from a wearable system. Personal and Ubiquitous Computing, 11(2), 117–132. Kocaballı, A. B., & KoçyiğitGranular, A. (2007). The best match algorithm for context- aware computing systems. Journal of Systems and Software, 80(12), 2015–2024. Korhonen, J., Ojala, T., Ristola, A., Kesti, M., Kilpelänaho, V., Koskinen, M., et al. (2007). Mobile fair diary: Hybrid interface for taking, browsing and sharing context-aware notes. Personal and Ubiquitous Computing, 11(7), 577–589. Korpipää, P., Mäntyjärvi, J., Kela, J., Keränen, H., & Malm, E. J. (2003). Managing context information in mobile devices. IEEE Pervasive Computing, 2(3), 42–51. Kouadri, S. M., & Younas, M. (2007). Context-oriented and transaction-based service provisioning. International Journal of Web and Grid Services, 3(2), 194–218. Krause, A., Smailagic, A., & Siewiorek, D. P. (2006). Context-aware mobile computing: Learning context-dependent personal preferences from a wearable sensor array. IEEE Transactions on Mobile Computing, 5(2), 113–127. Kumar, M., Shirazi, B. A., Das, S. K., Sung, B. Y., Levine, D., & Singhal, M. (2003). PICO: A middleware framework for pervasive computing. IEEE Pervasive Computing, 2(3), 72–79. Kurvinen, E., Lähteenmäki, M., Salovaara, A., & Lopez, F. (2007). Are you alive? Sensor data as a resource for social interaction. Knowledge, Technology, and Policy, 20(1), 39–49. Kwon, O. B. (2003). I know what you need to buy: Context-aware multimedia-based recommendation system. Expert Systems with Applications, 25(3), 387–400. Kwon, O. B. (2004). Modeling and generating context-aware agent-based applications with amended colored petri nets. Expert Systems with Applications, 27(4), 609–621. Kwon, O. B. (2006a). Multi-agent system approach to context-aware coordinated web services under general market mechanism. Decision Support Systems, 41(2), 380–399. Kwon, O. B. (2006b). The potential roles of context-aware computing technology in optimization-based intelligent decision-making. Expert Systems with Applications, 31(3), 629–642. Kwon, O., Choi, K., & Kim, M. (2007). User acceptance of context-aware services: Self-efficacy, user innovativeness and perceived sensitivity on contextual pressure. Behaviour Information Technology, 26(6), 483–498. Kwon, O. B., Choi, S. C., & Park, G. R. (2005). NAMA: A context-aware multi-agent based web service approach to proactive need identification for personalized reminder systems. Expert Systems with Applications, 29(1), 17–32. Kwon, O. B., & Kim, M. Y. (2004). MyMessage: Case-based reasoning and multicriteria decision making techniques for intelligent context-aware message filtering. Expert Systems with Applications, 27(3), 467–480. Kwon, O. B., & Sadeh, N. M. (2004). Applying case-based reasoning and multi-agent intelligent system to context-aware comparative shopping. Decision Support Systems, 37(2), 199–213. Kwon, O. B., Shin, J. M., & Kim, S. W. (2006). Context-aware multi-agent approach to pervasive negotiation support systems. Expert Systems with Applications, 31(2), 275–285. Kwon, O. B., Yoo, K. D., & Suh, E. H. (2005). UbiDSS: A proactive intelligent decision support system as an expert system deploying ubiquitous computing technologies. Expert Systems with Applications, 28(1), 149–161. Kwon, O. B., Yoo, K. D., & Suh, E. H. (2006). ubiES: Applying ubiquitous computing technologies to an expert system for context-aware proactive services. Electronic Commerce Research and Applications, 5(3), 209–219. Ladd, A. M., Bekris, K. E., Rudys, A., Kavraki, L. E., & Wallach, D. S. (2005). Robotics- based location sensing using wireless ethernet. Wireless Networks, 11(1–2), 189–204. Lee, W. P. (2007). Deploying personalized mobile services in an agent-based environment. Expert Systems with Applications, 31(4), 1194–1207. Liao, S. S., He, J. W., & Tang, T. H. (2004). A framework for context information management. Journal of Information Science, 30(6), 528–539. Lieberman, H., & Chu, A. (2007). An interface for mutual disambiguation of recognition errors in a multimodal navigational assistant. Multimedia Systems, 12(4–5), 393–402. Lieberman, H., & Selker, T. (2000). Out of context: Computer systems that adapt to learn from, context. IBM Systems Journal, 39(3–4), 617–632. Liechti, O. (2000). Awareness and the WWW: An overview. ACM SIGGROUP Bulletin, 21(3), 3–12. Lin, P., Tsai, C. Y., & Lai, Y. C. (2005). A SIP-based mobility management platform for WLAN location-aware broadcasting and multicasting application. Wireless Communications and Mobile Computing, 5, 647–663. Loke, S. W. (2006). Context-aware artifacts: Two development approaches. IEEE Pervasive Computing, 5(2), 48–53. Lum, W. Y., & Lau, F. C. M. (2002). A context-aware decision engine for content adaptation. IEEE Pervasive Computing, 1(3), 41–49. Maiden, N., Omo, O., Seyff, N., Grunbacher, P., & Mitteregger, K. (2007). Determining stakeholder needs in the workplace: How mobile technologies can help. Software, IEEE, 24(2), 46–52. Mandato, D., Kovacs, E., Hohl, F., & Amir-Alikhani, H. (2002). CAMP: A context- aware mobile portal. IEEE Communications Magazine, 40(1), 90–97. Mäntyjärvi, J., & Seppänen, T. (2003). Adapting applications in handheld devices using fuzzy context information. Interacting with Computers, 15(4), 521–538. Marias, G. F., Prigouris, N., Papazafeiropoulos, G., Hadjiefthymiades, S., & Merakos, L. (2004). Brokering positioning data from heterogeneous infrastructures. Wireless Personal Communications, 30(2), 233–245. Matsushita, S. (2001). A headset-based minimized wearable computer. IEEE Intelligent Systems, 16(3), 28–32. Mcnair, J., & Zhu, F. (2004). Vertical handoffs in fourth-generation multinetwork environments. IEEE Wireless Communications, 11(3), 8–15. Meeuwissen, E., Reinold, P., & LiemInferring, C. (2007). Inffering and predicting context of mobile users. Bell Labs Technical Journal, 12(2), 79–86. Michahelles, F., & Samulowitz, M. (2004). Smart CAPs for smart ITs – context detection for mobile users. Personal and Ubiquitous Computing, 6(4), 269–275. Minami, K., & Kotz, D. (2005). Secure context-sensitive authorization. Pervasive and Mobile Computing, 1(1), 123–156. Morla, R., & Davies, N. (2004). Evaluating a location-based application: A hybrid test and simulation environment. IEEE Pervasive Computing, 3(3), 48–56. Muñoz, M. A., Rodriguez, M., Favela, J., Martinez-Garcia, A. I., & González, V. M. (2003). Context-aware mobile communication in hospitals. IEEE Computer Society, 36(9), 38–46. Murphy, A. L., & Picco, G. P. (2004). Using coordination middleware for location- aware computing: A lime case study. Lecture Note in Computer Science, 263–278. Niemegeers, I. G., & Heemstra De Groot, S. M. (2005). FEDNETS: Context-aware ad- hoc network federations. Wireless Personal Communications, 33(3–4), 305–318. Norrie, M. C., Signer, B., Grossniklaus, M., Belotti, R., Decurtins, C., & Weibel, N. (2007). Context-aware platform for mobile data management. Wireless Networks, 13(6), 855–870. O’Grady, M. J., O’Hare, G. M. P., & Sas, C. (2005). Mobile agents for mobile tourists: A user evaluation of Gulliver’s genie. Interacting with Computers, 17(4), 343–366. O’Hare, G. M. P., & O’Grady, M. J. (2003). Gulliver’s genie: A multi-agent system for ubiquitous and intelligent content delivery. Computer Communications, 26(11), 1177–1187. Panagiotakis, S., & Alonistioti, A. (2006). Context-aware composition of mobile services. IT Professional, 8(4), 38–43. Pashtan, A., Heusser, A., & Scheuermann, P. (2004). Personal service areas for mobile web applications. IEEE Internet Computing, 8(6), 34–39. Pashtan, A., Kollipara, S., & Pearce, M. (2003). Adapting content for wireless web services. IEEE Internet Computing, 7(5), 79–85. Pertselakis, M., Ferles, C., Tsiolis, K., & Stafylopatis, A. (2005). Wireless distributed implementation of a fuzzy neural classification system. International Journal of Artificial Intelligence Tools, 14(4), 661–682. Pfeifer, T. (2005). Redundant positioning architecture. Computer Communications, 28(13), 1575–1585. Prekop, P., & Burnett, M. (2003). Activities, context and ubiquitous computing. Computer Communications, 26(11), 1168–1176. Qi, H., & Venkatasubramanian, N. (2007). Timeliness-accuracy balanced collection of dynamic context data. Parallel and Distributed Systems, IEEE, 18(2), 158–171. Qin, W., Shi, Y., & Suo, Y. (2007). Ontology-based context-aware middleware for smart spaces. Tsinghua Science & Technology, 12(6), 707–713. Qing, W., Károly, F., Christian, P., Paulo, M., & Bernhard, P. (2006). Context-aware handover using active network technology. Computer Networks, 50(15), 2855–2872. Qiu, L., Chang, L., Lin, F., & Shi, Z. (2007). Context optimization of AI planning for semantic web services composition. Service Oriented Computing and Applications, 1(2), 117–128. Raento, M., Oulasvirta, A., Petit, O., & Toivonen, H. (2005). Context phone: A prototyping platform for context-aware mobile applications. IEEE Pervasive Computing, 4(2), 51–59. Ranganathan, A., Al-Muhtadi, J., & Campbell, R. H. (2004). Reasoning about uncertain contexts in pervasive computing environments. IEEE Pervasive Computing, 3(2), 62–70. Ranganathan, A., Campbell, R. H., Ravi, A., & Mahajan, A. (2002). Conchat: A context- aware chat program. IEEE Pervasive Computing, 1(3), 51–57. Rehman, K., Stajano, F., & Coulouris, G. (2007). An architecture for interactive context-aware applications. IEEE Pervasive Computing, 6(1), 73–80. Reignier, P., Brdiczka, O., Vaufreydaz, D., Crowley, J. L., & Maisonnasse, J. (2007). Context-aware environments: From specification to implementation. Expert Systems, 24(5), 305–320. 8522 J.-y. Hong et al. / Expert Systems with Applications 36 (2009) 8509–8522 Riekki, J., Huhtinen, J., Ala-Siuru, P., Alahuhta, P., Kaartinen, J., & Röning, J. (2003). Genie of the net, an agent platform for managing services on behalf of the user. Computer Communications, 26(11), 1188–1198. Riva, O., Nadeem, T., Borcea, C., & Iftode, L. (2007). Context-aware migratory services in ad hoc networks. Mobile Computing, IEEE, 6(12), 1313–1328. Riva, O., & Toivonen, S. (2007). The DYNAMOS approach to support context-aware service provisioning in mobile environments. Journal of Systems and Software, 80(12), 1956–1972. Rocha, R. C. A., & Endler, M. (2006). Middleware: Context management in heterogeneous, evolving ubiquitous environments. Distributed Systems Online, IEEE, 7(4), 1. Rodríguez, M. D., Favela, J., Martínez, E. A., & Muñoz, M. A. (2004). Location-aware access to hospital information and services. IEEE Transactions on Information Technology in Biomedicine, 8(4), 448–455. Rogers, Y., & Muller, H. (2006). A framework for designing sensor-based interactions to promote exploration and reflection in play. International Journal of Human- Computer Studies, 64(1), 1–14. Román, M., Hess, C., Cerqueira, R., Ranganathan, A., Campbell, R. H., & Nahrstedt, K. (2002). A middleware infrastructure for active spaces. IEEE Pervasive Computing, 1(4), 74–83. Roussaki, L. G., Strimpakou, M., & Anagnostou, M. E. (2007). Designing next generation middleware for context-aware ubiquitous and pervasive computing. International Journal of Ad Hoc and Ubiquitous Computing, 2(3), 197–206. 2007. Roussaki, I., Strimpakou, M., & Pils, C. (2007). Distributed context retrieval and consistency control in pervasive computing. Journal of Network and Systems Management, 15(1), 57–74. Roussos, G., Marsh, A. J., & Maglavera, S. (2005). Enabling pervasive computing with smart phones. IEEE Pervasive Computing, 4(2), 20–27. Sacramento, V., Endler, M., Rubinsztejn, H. K., Lima, L. S., Gonçalves, K., Nascimento, F. N., et al. (2004). MoCA: A middleware for developing collaborative applications for mobile users. IEEE Distributed Systems Online, 5(10), 1–14. Sajal, K. D., Nirmalya, R., & Abhishek, R. (2006). Context-aware resource management in multi-inhabitant smart homes: A framework based on Nash H-learning. Pervasive and Mobile Computing, 2(4), 372–404. Samaan, N., & Karmouch, A. (2005). A mobility prediction architecture based on contextual knowledge and spatial conceptual maps. IEEE Transactions on Mobile Computing, 4(6), 537–551. Satoh, I. (2003). SpatialAgents: Integrating user mobility and program mobility in ubiquitous computing environments. Wireless Communications and Mobile Computing, 3(4), 411–423. Satoh, I. (2007). A location model for smart environments. Pervasive and Mobile Computing, 3(2), 158–179. Schilit, B. N, Adams, N. I, & Want, R. (1994). In Context-aware computing applications (pp. 85–90). IEEE Computer Society. Schilit, B. N., Hilbert, D. M., & Trevor, J. (2002). Context-aware communication. IEEE Wireless Communications, 9(5), 46–54. Schmidt, A., & Laerhoven, K. V. (2001). How to build smart appliances? IEEE Personal Communications, 8(4), 66–71. Schmidt, A., Takaluoma, A., & Mantyjarvi, J. (2000). Context-aware telephony over WAP. Personal Technologies, 4(4), 225–229. Schulzrinne, H., Wu, X., Sidiroglou, S., & Berger, S. (2003). Ubiquitous computing in home networks. IEEE Communications Magazine, 41(11), 128–135. Selker, T. (2004). Visual attentive interfaces. BT Technology Journal, 22(4), 146–150. Serrano, J. M., & Serrat, J. (2006). Information modeling and handling for context- aware multimedia services. IEEE Wireless Communications, 13(5), 104–111. Signer, B., Grossniklaus, M., & Norrie, M. C. (2007). Interactive paper as a mobile client for a multi-channel web information system. World Wide Web, 10(4), 529–556. Simons, C., & Wirtz, G. (2007). Modeling context in mobile distributed systems with the UML. Journal of Visual Languages & Computing, 18(4), 420–439. Singh, A., & Acharya, A. (2005). Multiplayer networked gaming with the session initiation protocol. Computer Networks, 49(1), 38–51. Skov, B., & Høegh, T. (2006). Supporting information access in a hospital ward by a context-aware mobile electronic patient record. Personal and Ubiquitous Computing, 10(4). Smailagic, A., & Kogan, D. (2002). Location sensing and privacy in a context-aware computing environment. IEEE Wireless Communications, 9(5), 10–17. Smailagic, A., & Siewiorek, D. (2002). Application design: Wearable and context- aware computers. IEEE Pervasive Computing, 1(4), 20–29. Smith, M. K., Welty, C., & McGuinness, D. L. (2004). OWL Web Ontology Language Reference. World Wide Web Consortium (W3C) Recommendation, <www.w3.org/TR/owl-ref>. Smith, D., Ma, L., & Ryan, N. (2006). Acoustic environment as an indicator of social and physical context. Personal and Ubiquitous Computing, 10(4), 241–254. Soldatos, J., Dimakis, N., Stamatis, K., & Polymenakos, L. (2007). A breadboard architecture for pervasive context-aware services in smart spaces: Middleware components and prototype applications. Personal and Ubiquitous Computing, 11(3), 193–212. Soldatos, J., Stamatis, K., Azodolmolky, S., Pandis, I., & Polymenakos, L. (2007). Semantic web technologies for ubiquitous computing resource management in smart spaces. International Journal of Web Engineering and Technology, 3(4), 353–373. Stylianos, D., Niki, P., Kia, M., & Christos, D. (2007). A context-aware cache structure for mobile computing environments. Journal of Systems and Software, 80(7), 1102–1119. Sumi, Y., & Mase, K. (2000). Supporting awareness of shared interests and experiences in community. ACM SIGGROUP Bulletin, 21(3), 35–42. Sumi, K., & Nishida, T. (2001). Telme: A personalized, context-aware communication support system. IEEE Intelligent Systems, 16(3), 21–27. Sung, M., Gips, J., Eagle, N., Madan, A., Caneel, R., Devaul, R., et al. (2005). Mobile-IT Education (MIT. EDU): m-Learning applications for classroom settings. Journal of Computer Assisted Learning, 21(3), 229–237. Syukur, E., & Loke, S. W. (2006). Implementing context-aware regulation of context- aware mobile services in pervasive computing environments. International Journal of Web and Grid Services, 2(3), 260–305. Syukur, E., & Loke, S. W. (2007). Policy based control of context aware pervasive services. Journal of Ubiquitous Computing and Intelligence, 1(1), 110–131. Tentori, M., Favela, J., & Rodriguez, M. D. (2006). Privacy-aware autonomous agents for pervasive healthcare. Intelligent Systems, IEEE, 21(6), 55–62. Trumler, W., Bagci, F., Petzold, J., & Ungerer, T. (2005). AMUN-autonomic middleware for ubiquitous environments applied to the smart doorplate project. Advanced Engineering Informatics, 19(3), 243–252. Tyan, J., & Mahmoud, Q. H. (2005). A comprehensive service discovery solution for mobile ad hoc networks. Mobile Networks and Applications, 10(4), 423–434. Udugama, A., Kuladinithi, K., Gorg, C., Pittmann, F., & Tionardi, L. (2007). NetCAPE: Enabling seamless IMS service delivery across heterogeneous mobile networks. Communications Magazine, IEEE, 45(7), 84–91. van Kranenburg, H., Bargh, M. S., Iacob, S., & Peddemors, A. (2006). A context management framework for supporting context-aware distributed applications. Communications Magazine, IEEE, 44(8), 67–74. van Sinderen, M. J., van Halteren, A. T., Wegdam, M., Meeuwissen, H. B., & Eertink, E. H. (2006). Supporting context-aware mobile applications: An infrastructure approach. Communications Magazine, IEEE, 44(9), 96–104. Vassilis, K., Loukas, H., Dimitris, K., & Stavros, K. (2006). A dynamic context-aware access control architecture for e-services. Computers & Security, 25(7), 507–521. Vuković, M., Kotsovinos, E., & Robinson, P. (2007). An architecture for rapid, on-demand service composition. Service Oriented Computing and Applications, 1(4), 197–212. Wac, K., Halteren, A. V., Bults, R., & Broens, T. (2007). Context-aware QoS provisioning in an m-health service platform. International Journal of Internet Protocol Technology, 2(2), 102–108. Wang, X., Dong, J. S., Chin, C. Y., Hettiarachchi, S. R., & Zhang, D. (2004). Semantic space: An infrastructure for smart spaces. IEEE Pervasive Computing, 3(2), 32–39. Weber, O., Sorkine, O., Lipman, Y., & Gotsman, C. (2007). Context-aware skeletal shape deformation. Computer Graphics Forum, 26(3), 265–274. Wilson, D., Doyle, J., Weakliam, J., Bertolotto, M., & Lynch, D. (2007). Personalised maps in multimodal mobile GIS. International Journal of Web Engineering and Technology, 3(2), 196–216. Wohltorf, J., Cissée, R., & Rieger, A. (2005). BerlinTainment: An agent-based context- aware entertainment planning system. IEEE Communications Magazine, 43(6), 102–109. Wu, Z., Wu, Q., Cheng, H., Pan, G., Zhao, M., & Sun, J. (2007). ScudWare: A semantic and adaptive middleware platform for smart vehicle space. Intelligent Transportation Systems, IEEE, 8(1), 121–132. Xiong, L., & Hassan, A. K. (2006). Location awareness through trajectory prediction. Computers, Environment and Urban Systems, 30(6), 741–756. Xu, Y., Fu, T. Y., Lee, W. C., & Winter, J. (2007). Processing k nearest neighbor queries in location-aware sensor networks. Signal Processing, 87(12), 2861–2881. Yang, K., Wu, Y., Yang, Y., & Liu, E. (2007). Policy-based service-driven dynamic planning of heterogeneous wireless networks. International Journal of Mobile Network Design and Innovation, 2(1), 67–77. Yau, S. S., & Karim, F. (2003). An energy-efficient object discovery protocol for context-sensitive middleware for ubiquitous computing. IEEE Transaction on Parallel and Distributed Systems, 14(11), 1074–1084. Yau, S. S., & Karim, F. (2004a). A context-sensitive middleware for dynamic integration of mobile devices with network infrastructures. Journal of Parallel and Distributed Computing, 64(2), 301–317. Yau, S. S., & Karim, F. (2004b). An adaptive middleware for context-sensitive communications for real-time applications in ubiquitous computing environments. Real-Time Systems, 26(1), 29–61. Yee, G., Korba, L., Lin, N. H., & Shih, T. K. (2006). Context-aware privacy and security agents for distance education. International Journal of High Performance Computing and Networking, 3(5–6), 395–404. Yu, Z., Zhou, Z., Yu, Z., Zhang, D., & Chin, C. (2006). An OSGI-based infrastructure for context-aware multimedia services. Communications Magazine, IEEE, 44(10), 136–142. Yu, Z., Zhou, X., Zhang, D., Chin, C., Wang, X., & Men, J. (2006). Supporting context- aware media recommendations for smart phones. IEEE Pervasive Computing, 5(3), 68–75. Yuan, S. T., & Chen, F. Y. (2007). UbiSrvInt – A context-aware fault-tolerant approach toward wireless P2P service provision. Expert Systems with Applications, 32(3), 726–752. Yuan, S. T., & Peng, K. H. (2004). Location based and customized voice information service for mobile community. Information Systems Frontiers, 6(4), 297–311. Zhang, Q., Qi, Y., Zhao, J., Hou, D., & Niu, Y. (2007). Fuzzy privacy decision for context-aware access personal information. Wuhan University Journal of Natural Sciences, 12(5), 941–945. Zhong, Y., & Gilbert, J. E. (2005). A context-aware language model for spoken query retrieval. International Journal of Speech Technology, 8(2), 203–219. Zhu, F., Mutka, M. W., & Ni, L. M. (2005). Service discovery in pervasive computing environments. IEEE Pervasive Computing, 4(4), 81–90. Zimmermann, A., Specht, M., & Lorenz, A. (2005). Personalization and context management. User Modeling and User-Adapted Interaction, 15(3–4), 275–302. http://www.w3.org/TR/owl-ref Context-aware systems: A literature review and classification Introduction Procedures The selection criteria Data sources and procedures to extract articles The general characteristics of the articles Classification of articles by publication year Classification of articles by online database Classification of articles by journal Classification framework Abstract architecture of context-aware systems Classification framework of context-aware systems Result of article classification Concept and research layer Network layer Middleware layer Application and service layer User infrastructure layer Discussions and suggestions How to extract and use the cognitive context in context-aware application? What are design patterns of context-aware systems? Which is the best inferring algorithm to extract user context and provide service to user? How to deal with concurrently enormous data, information and knowledge having different format to serve suitable service to users? How to extract the best solution when the context of users is conflicted? How to reflect the preference of users for satisfying user needs? How to save users information in context-aware systems? How to evaluate performance of context-aware systems? Conclusions References 
work_jvdpob6gy5gnzmz3q6hjkvlpui	----	doi:10.1016/j.eswa.2004.10.004 An integrated approach for developing e-commerce applications Francisco Garcı́a-Sánchez, Rafael Valencia-Garcı́a, Rodrigo Martı́nez-Béjar* Departamento de Ingenierı́a de la Información y las Comunicaciones, Facultad de Informática, Universidad de Murcia, 30071 Espinardo (Murcia), Spain Abstract This paper presents a framework that merges various advanced information technologies for developing electronic commerce (e-commerce) applications. The use of e-commerce utilities provides several advantages to businesses. Intelligent agents can be used to facilitate some tasks from those that take place in a commercial transaction moving to a second generation of e-commerce applications. We present a prototype which integrates a multiagent system (composed by buyer and seller agents) with a Web application. q 2004 Elsevier Ltd. All rights reserved. Keywords: Electronic commerce; Intelligent agents; Natural language processing; Ontologies 1. Introduction Several definitions have been given to the term ‘e-commerce’ or ‘electronic commerce’. The one given by the Electronic Commerce Association 1 is: ‘electronic commerce covers any form of business or administrative transaction or information exchange that is executed using any information and communication technology’. We limit our approach to covering commercial activities conducted on the Internet. E-commerce offers opportunities to dramatically improve the way that businesses interact with both their customers and their suppliers, that is, to make business negotiations faster, cheaper, more personalized, and/or more agile. The number of web users who shop for or buy products online is continuously increasing (Silverman, Bachann, & Al-Akharas, 2001). However, searching and buying pro- ducts via on-line can be frustrating due to the lack of help or decision support given to the user. Nowadays e-commerce applications are being improved from a first generation stage where buyers are humans who browse through a catalog of commodities (e.g. books, computer components, films) and make purchases, often by means of a credit card transaction, to a second generation with a greater degree of 0957-4174/$ - see front matter q 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2004.10.004 * Corresponding author. Tel.: C34 968 364634; fax: C34 968 364151. E-mail addresses: fgs2@alu.um.es (F. Garcı́a-Sánchez), valencia@ um.es (R. Valencia-Garcı́a), rodrigo@dif.um.es (R. Martı́nez-Béjar). 1 http://www.theeca.org/. automation on both the buyer’s and the seller’s side (Wooldridge, 2002). The aim of this work was to develop a technology for facilitating ‘Business-to-Consumer’ (B2C) and ‘Business- to-Business’ (B2B) processes. For it, various methodologies including (multi)agents (which allow for the interaction among several manufacturers), decision support (used for rating and differentiation processes between the products found) and natural language processing (NLP) have been used. The previous objective can be decomposed into the following sub-objectives: (1) generation of a (computer) product ontology, (2) design of a database containing all relevant information about the products in question, (3) design and implementation of two agent models, one playing a buyer’s role and another playing a seller’s role, (4) integrating the agents into the Foundations for Intelligent Physical Agents-Open Source (FIPA-OS) framework so that communication among the agents registered at the platform may take place, and (5) design and implementation of a Web application that makes use of agents to offer e-commerce services (this will be the prototype). The solution described in this work is built on top of an integrated system elaborated on topics belonging to a number of disciplines, including knowledge management, NLP, decision support theory and multiagent technology. The experiments done with the system up to now make us to believe that its future is promising in helping solve some problems we can find in actual e-commerce applications. Expert Systems with Applications 28 (2005) 223–235 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa http://www.theeca.org/ F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235224 The structure of this paper is as follows. Section 2 offers a brief description of the methodology applied. In Section 3, the e-commerce framework used in this work is introduced. Section 4 describes the system prototype developed making use of the framework presented. The validation of the system prototype is shown in Section 5. Finally, some conclusions are put forward in Section 6. 2. Methodology In this work, we have applied a methodology that integrates different advanced information technologies with the aim of building a framework for developing e-commerce applications. These are: intelligent agents, which have been used to represent the different entities we find in a transaction process (buyers/sellers); decision support sys- tems (DSSs) theory, applied to help buyers in some decision processes; ontologies that establish a reusable, shareable semantics for the terminology, so that different types of agents are able to communicate among themselves; and NLP, which permits to better determine the users desires (since the users can use their own terminology). In the following lines, a brief description of each of them is realised. 2.1. Multiagent systems A multiagent system is seen as a system that consists of a group of agents that can potentially interact with each other (Vlassis, 2003). The term agent is defined as follows (Wooldridge, 2002): “an agent is a computer system that is situated in some environment, and that is capable of autonomous action in this environment in order to meet its design objectives”. For this author, an agent is intelligent if it fulfils the following properties: reactivity (it responds in a timely fashion when needed), pro-activeness (it satisfies internal goals; it takes actions when it seems it will be useful) and social ability (it interacts with other agents in order to satisfy goals). In order to design and construct agents it will be necessary to take the following points into account: (i) Agents theory (formal specification that describes what properties the agents must fulfil), (ii) Agent languages (tools for the design and construction of agent-based systems, e.g.: agent-oriented programming—Agent0—, concurrent METATEM) and, (iii) Agent architectures (agents’ internal structure which can be logic-based, reactive, Belief-Desire-Intention-based or layered). Agents can be useful as stand-alone entities that are delegated particular tasks on behalf of a user. However, in the majority of cases agents exist in environments that contain other agents. In these multiagent systems (MASs), the global behaviour derives from the interaction among the constituent agents. There are two main classes of MASs (Wooldridge, 2002): distributed problem solving systems, in which the component agents are explicitly designed to cooperatively achieve a given goal; and open systems, whose agents are not co-designed to share a common goal, and have been possibly developed by different people to achieve possibly different objectives. Moreover, the com- position of the system can dynamically vary as agents enter and leave the system. When a group of individual agents form a MAS, the presence of a mechanism to coordinate such a group as well as a communication language becomes necessary. Among the coordination mechanisms the cases of agents having common objectives (and, therefore, they cooperate) can be distinguished from those where agents are self-interested and have conflictive objectives with other agents. For the latter case we will need negotiation. The most usual cooperation mechanisms are: organizational structures, multiagent planning—centralized and distributed—, con- tract nets and functionally exact cooperation. Among the ones belonging to negotiation we may highlight coalitions formation, market mechanism, the bargaining theory, vote, auctions and task assignment between two agents. Agent communication languages allow agents to interact each other while they hide their internal work details by exchanging information and knowledge. Examples of those ones are KIF (Knowledge Interchange Format), KQML (Knowledge Query and Manipulation Language) and FIPA- ACL (Foundations for Intelligent Physical Agents-Agents Communication Language). 2.1.1. Multiagent platforms For this project, we have employed available implemen- tations of agent platforms which conform to FIPA (The Foundations for Intelligent Physical Agents) Specifications, namely, a group of normative rules that permit an agent society to operate among themselves. This model identifies some necessary agent’s roles for the platform management: the AMS (Agent Management System), the DF (Directory Facilitator), the ACC (Agent Communication Channel), the IPMT (Internal Platform Message Transport) and the IIOP (Internet Interface Object Protocol) (FIPA-OS Developers Guide, 2001). The communication language among agents that FIPA proposes is FIPA-ACL, which specifies a standard message passing language, establishing codification, semantics and message pragmatics. Finally, FIPA Standard includes the non-agent software integration specification, the agents mobility and security, the ontology service, and the human-agent interaction. Amongst the frameworks that satisfy the FIPA Standard, the more popular are ZEUS, JADE and FIPA-OS. We have selected this last one in this work. FIPA-OS is an agent framework developed with the aim of building hetero- geneous agent platforms, agents and services that adjust to FIPA Standard. In addition to the advantage to be an open system in terms of license of use and extensibility, like F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235 225 ZEUS and JADE, FIPA-OS presents some additional advantages: , Interoperability with other FIPA agents platforms not developed with FIPA-OS. , It allows for agents heterogeneity, that is, agents implemented with different programming languages. , It supports FIPA specification for agents management. , It provides abstractions and application program inter- faces (APIs) that permit extending and integrating an agents’s platform with other already existent software platforms. 2.1.2. Agents for electronic commerce According to Wooldridge (2002) agents make it possible the second-generation e-commerce systems, in which many aspects of a consumer’s buying behaviour is automated. One phased model that attempts to describe consumer- buying behaviour is the following (Guttman, Moukas, & Maes, 1998): 1. Need identification. This stage characterizes the con- sumer becoming aware of some need that is not satisfied. 2. Product brokering. A would-be consumer obtains information related to available products in order to determine what product to buy. 3. Merchant brokering. The consumer selects the supplier. This stage will typically involve examining offers from a range of different merchants. 4. Negotiation. The terms of the transaction are agreed between the would-be consumer and the would-be mer- chant. In some markets, the negotiation stage is empty— the terms of agreement are fixed and not negotiable. 5. Purchase and delivery. The transaction is actually carried out, and the good delivered. 6. Product service and evaluation. The post-purchase stage involves product service, customer service, etc. Agents have been widely promoted as being able to (partially) automate some of these stages. 2.2. Decision support systems (DSS) “In the rush to open their website, e-commerce sites too often fail to support buyer decision making and search, resulting in a loss of sale and the customer’s repeat business” (Silverman et al., 2001). DSSs cover a wide variety of systems, tools and technologies that support decision-making in semi-structured and unstructured situ- ations where no one knows exactly how the decision should be made. It provides information, models, and data manipulation tools to solve the structured parts of the problem and help isolating places where both judgment and experience are required. As Silverman remarks, in online shopping sites buyers have decisions to work out and tasks to perform that can be facilitated by means of DSS software. We help clients with part of this process using a natural language system (based on a morphological analysis) that allows the system to find out the user’s desires. Besides, a sorting process helps buyers to determine the most convenient product for them. 2.3. Ontologies Defining a common vocabulary (apart from the com- munication language selected) is necessary for the agents to communicate with one another. Thus, the communication is facilitated through one ontology that defines the concepts and the relations between the concepts of a particular domain. With all, an ontology is viewed in this work as a specification of a domain knowledge conceptualisation (Van Heijst, Schreiber, & Wielinga, 1997), and is represented through a set of concepts, while concepts are defined through sets of both attributes and interconceptual relations. So, the main ontological entity in the system developed is the concept but the use of the other ontological entities such as attributes is also possible in the model in order to provide the system with powerful representational capabilities. 2.4. Natural language processing (NLP) In the approach presented in this work, the methodology presented in Valencia-Garcı́a, Ruiz-Sánchez, Vivancos-Vi- cente, Fernández-Breis, and Martı́nez-Béjar (2004) has been used to get knowledge from text. This methodology, which uses ontologies and one incremental knowledge acquisition technique termed MCRDR (Kang, 1996), is based on the idea that relationships between concepts are usually associated to verbs in natural language. This methodology uses the mentioned technique and the grammar category of the words in the current sentence to infer other knowledge entities (e.g. concept, attributes and values) in order to create an ontology from a text fragment. 3. A framework for developing e-commerce applications 3.1. Overview of the system The system architecture, which is shown in Fig. 1, has four principal interconnected components. The most important component is the multiagent platform, where two agent types have been implemented, namely a buyer and a seller. These agents communicate with one another by means of ACL (Agent Communication Language) messages. These agents must register in the Agent Platform (AP), which specifies several types of agents that can facilitate the multiagent communication. In particular, the two Agent Platform types that will be necessary to get registered in are: the Directory Facil- itator (DF), which provides a yellow pages service to Fig. 1. Framework general structure. F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235226 the other agents, and the Agent Management System (AMS), which provides platform-typical management functions (such as life cycle monitoring, checking of all the entities’ correct behaviour and ‘white pages’). The buyer agent functions include to get external petitions, process them and return the obtained results. The petition processing may include NLP, product ontology matching, planning, plan execution and answer composition. Regard- ing the seller agent, this has been provided only with communication services with the buyer agents, so that agent is able to answer for two petitions types: questions about a particular type of product and orders. It can also communicate with a product database, which contains the product catalog of a manufacturer. The second system component is the database (one for each seller agent), which can be shared by the agents. The database can be acceded by the seller agents using the Data Access Object (DAO) pattern. The database is composed by the products a manufacturer wants to sell, so that the database contents are ontology-dependent. Another important component of the system is the product ontology. The ontology is saved in an XML file and represents the specific relevant domain knowledge we are dealing with, establishing a common set of words that allows different types of agents to communicate one another. The last main component is the Web application, which employs the FIPA-OS platform together with the above- referred agents. For the application design, we have made use of the Struts framework and Java technology. In the following paragraphs, all these components are detailed more precisely. 3.2. FIPA-OS and intelligent agents According to FIPA-OS Developers Guide (2001), FIPA- OS is a component-orientated toolkit for constructing FIPA compliant Agents using mandatory components (i.e. components required by all FIPA-OS Agents to execute), components with switchable implementations, and optional components (i.e. components that a FIPA-OS Agent can optionally use) (see Fig. 2). Agent shell (FIPAOSAgent). It provides a shell for Agent implementation to use by simply extending the FIPAOSA- gent class. The FIPAOSAgent shell is responsible for loading an agent’s profile, and initialising the other components which the Agent is composed by. Task manager. It provides the ability to split the functionality of an agent into smaller, disjoint units of work known as tasks. These can be defined as self-contained pieces of code that carry out some task and (optionally) return a result, have the ability to send and receive messages, and have little or preferably no dependence on the agent which they are executed within. Conversation manager. It provides the ability to keep track of the conversation state at the performative level, as well as a mechanism for grouping messages of the same conversation together. Message transport service. It provides the ability to send/receive messages to/from an agent implementation. 3.2.1. The buyer agent The buyer agent’s target is to look for the products that a customer requests for (starting from a sentence) as well as asking for a concrete product. Fig. 2. Components within FIPA-OS. F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235 227 Agent initialisation. It involves four main tasks. First, it is needed to initialise the NLP tool. Second, the agent must be registered in the FIPA Agent Platform (AMS and DF Agents). Third, messages buffers must be activated before conversa- tions take place. Finally, the agent has to get all products categories from the ontology (main concepts, synonymous, attributes, relationships,.); the agent will use this infor- mation when that one communicates with other agents. Products search. This process begins when a customer (i.e. the application user) writes a sentence denoting the product that (s)he wants to get, and finishes when the system returns the products matching what the user wants. In order to do this, it is necessary a planner that determines the steps that the agent must follow. For it, we have used an STRIPS- type planner and some operators (see Section 3.2.3 for details). When the system has the plan, it is necessary to execute it, so that the possible steps then are the following: , Sentence analysis. Through this step, the system obtains the concepts underlying the sentence introduced by the customer. For it, the sentences analyser referred to before in this paper is used. , Category obtaining. The system determines the product categories that the customer is looking for. For this, the system compares the concepts found in the sentence with the categories extracted from the product ontology. If the system does not find any category that matches any of the extracted concepts, then the plan execution is stopped. , Manufacturer finding. This is executed thanks to a task included in the agent’s implementation. The task target is to look for seller agents in the ‘yellow pages’. For it, the system incorporates the Directory Facilitator Agent—DF—, which returns a collection with all Fig. 3. Message format 1 for info the identifiers (ID) of the agents found. In order to match this ID with the agent transport address, the system uses the Agent Management System (AMS), i.e. white pages. If the DF does not find any agent, then the plan execution stops. , Information request. For each category and seller agent found, a task is created that starts a conversation with the found agent and returns all products found grouped by seller agent. The information returned has the following format (Fig. 3).Where ‘Cat-i’ represents the ith category, ‘Man-j’ the jth manufacturer and ‘Prod-ij’ is a product of the ith category offered by the jth manufacturer.For each category, the system will return all the products (and its respective manufacturers) found with respect to that category.It is important to emphasize that each seller agent works in parallel to find the products of each category. The buyer agent saves this information then and asks for the next category. , Offer selection. Once a buyer agent has the information described before, the only remaining thing it has to do is to sort the products of each category.The result of this process is a collection with the following format (Fig. 4).Where ‘BestPr’ represents the best product for the category under question, ‘ManBP’ represents the manufacturer for such a product, ‘2ndPr’ represent the second best product for the category, ‘Man2P’ is the manufacturer for the mentioned product, and so on.This will be returned to the customer. Asking for a product. In order to execute a product petition, a search must have taken place previously, because it is necessary to indicate the seller agent ID, in addition to other product-related features such as the amount of products, the product type and the product code. rmation request. Table 1 An operator for the blocks world problem Pick-up(block) Preconditions Clear(block) On-table(block) Arm-empty Add list Holding(block) Delete list On-table(block) Clear(block) Arm-empty Fig. 4. Message format 2 for offer selection. F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235228 The task to attend the request hence needs the product code (this must agree with the one that the seller agent has registered), the product type or category, the AgentID of the seller agent that offers the product and the number of units the customer requests. All this information is then sent to the seller through an ACL message. 3.2.2. The seller agent The seller agent is in charge of performing three tasks: one consists of attending petitions and determining which task to execute; the second is to keep the communication with the buyer agent which requests some products, and the last carries out the petition record of a given product. Therefore, this agent’s main objective is to receive petitions and questions from different buyer agents. Agent initialisation. This process is similar to the agent initialisation process in the buyer agent. This agent does not use NLP neither planning. Search task. This task is executed when a ‘Query’-type message is received. The aim of this task is to return all the products found in the database with the category indicated in the query. Petition task. This task is called when a message ‘Petition’ is received. The function of this task is to insert a record of the product sale into the database. This task can check whether the product is available in stock or it is necessary to make a petition to the providers. 3.2.3. The planner STRIPS (Fikes & Nilsson, 1971) is a linear planner that attempts to find a sequence of operators in a space of world models to transform a given initial world model into a model in which a given goal formula can be proven true. It represents a world model as an arbitrary collection of first-order predicate calculus formulae and works with models consisting of a large number of formulae. It uses a resolution theorem prover to answer questions of particular models and mean-ends analysis as a guide towards the desired goal-satisfying model. For every world model, we have a set of applicable operators, which transform the world model into another world model. The problem solver finds some composition of operators that transforms a given initial world model into one that satisfies a stated goal condition. In STRIPS, a world model is represented by a set of well formed first-order predicate calculus formulae. Each operator in a solution corresponds to an action routine whose execution causes a robot to take certain actions. Each operator in STRIPS is composed by: – A set of preconditions. To execute the action related to the operator it is necessary that preconditions are true before the operator can be applied. – Delete list, which is a set of formulae that will not be true after the operator has been applied (so that the planner has to delete them from the current world model). – Add list, which is a set of formulae that will be true after the operator has been applied (so that the planner has to add them to the current world model). For example, an operator for the Blocks World problem is shown in Table 1. In Table 1, clear(block), on-table (block), arm-empty and holding(block) are well formed formulae. STRIPS, like most of planners, has been applied to the blocks world problem as an effective benchmark (Slaney & Thiébaux, 2001). The blocks world problem consists of a finite number of blocks stacked into towers on a table large enough to hold them all. The towers’ positioning on the table is irrelevant. The Blocks World planning problem is to turn a blocks initial state into a goal state, by moving one block at a time from the top of a tower onto another tower or to the table. The optimal Block World planning problem consist in doing so in a minimal number of moves. Most of planners adopt the STRIPS representation and search forward the state space. The GRT planner (Refanidis & Vlahavas, 2001) is a domain-independent heuristic planner, which solves planning problems calculating in a first, phase the distances between the facts and the goals of the problem. The second phase consists of a simple best first search strategy using the distances calculated in the previous phase. This planner has been validated in several domains like blocks-world domain, in which GRT can easily solve problems with more than 20 blocks. Another planner, which adopts the STRIPS represen- tation is the Fast-Forward Planning System (Hoffmann, 2001), which attacks the planning problems by forward Table 2 The set of well formed formulae Formula Description PREFERENCE(S) The customer’s desires contained in the sentence S CONCEPT(CO, S) The concept CO in S CATEGORY(C, CO, S) The category C is one of the customer’s preferences related to CO in S MANUFACTURER(MA) The manufacturer MA is registered in the system, therefore, it is available for possible petitions INFO(OF, MA, C) The manufacturer MA presents the offer OF for C SORT(OF) The offer’s group is sorted out attending to the relation quality-price F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235 229 searching in the state space, guided by a heuristic function. This function is extracted from the domain description relaxing the planning problem by ignoring parts of its specification, concretely the delete lists of operators. In this work, we have used a STRIPS planner for allowing agents to determine dynamically the actions to be performed. More precisely, this planner has been implemented into the buyer agent to accomplish the product search task. The well formed formulae of first-order predicate calculus used in this planner are showed in Table 2. The initial world model is formed by the formulae which represent the user’s preferences. The formulae of the final world model represent an ordered set containing the products that the customer may be interested in (according to his/her likes and dislikes). The operators which have been designed can be viewed in Table 3. 3.2.4. Natural language processing The NLP application used in this approach is based on the work presented in Valencia-Garcı́a et al. (2004), where Table 3 The STRIPS operators Operator FINDMANUFAC- TURER( ) Preconditions Delete list Add list MANUFACTU ANALYZE(F) Preconditions PREFERENCE MANUFACTU Delete list MANUFACTU Add list CONCEPT(CO GETCATEGORY(CO) Preconditions CONCEPT(CO Delete list CONCEPT(CO Add list CATEGORY(C REQUESTINFO(C) Preconditions MANUFACTU CATEGORY(C Delete list CATEGORY(C Add list INFO(OF, MA SELECTOFFER(OF) Preconditions INFO(OF, MA Delete list INFO(OF, MA Add list SORT(OF) the authors described a software tool for ontology building from natural language texts. The principal idea sustaining our approach is that in natural language relationships between knowledge entities are usually associated to verbs. The knowledge acquisition process is divided into three sequential phases, which have been implemented into three separate modules: POS-tagging, Concept search and inference. POS-tagging. The main objective of this process is to obtain the grammatical category of each word in the current sentence. For this purpose, the POS-tagger described in Ruiz-Sánchez, Valencia-Garcı́a, Fernández-Breis, Martı́- nez-Béjar, and Compton (2003) is used. This is the first sub- process to be carried out. Concept search. Through this process, linguistic expressions representing concepts are identified. The associations between linguistic expressions and concepts have been stored in the conceptual knowledge base in a previous training of the system. This process is quite simple and as a result we obtain all the expressions of the fragment which are already in the conceptual knowledge base. Inference. In natural language, relationships between concepts are usually associated to verbs. Although the sub- process in which MCRDR acts is mainly concerned with obtaining relationships between concepts, it can also be used to get other knowledge categories like concepts, attributes, or values. This MCRDR component is formed by a knowledge-base that contains linguistic expressions representing generic conceptual relationships, and by an MCRDR subsystem that infers the participants in these relationships. We describe next the modus operandi of this process. Firstly, the verb in the current sentence is identified. Then, the user (of the system if the knowledge base is not empty) searches for the type of semantic relationship associated to that verb. Once the type of relation associated to the main verb in the current sentence has been found, the MCRDR sub-system is applied to extract knowledge by Action Return all the manufacturer registered RER(MA) (S) Obtain the relevant concepts within a sentence RER(MA) RER(MA) , S) ,S) Try to match a concept with the set of product categories ,S) , CO, S) RER(MA) Make a request to a manufacturer about a category , CO, S) , CO, S) , C) , C) Sort a set of offers related to the same category , C) F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235230 means of the grammatical category of the words, their position in the current sentence, and the type of relation associated to the verb, if any. The inputs of this tool are the product queries, in natural language, introduced by the customers. An example of these queries could be: “I need a 20 GB Hard Disk”. Then, the system would obtain, for instance, an ontology formed by the concept ‘HARD DISK’, so that the attribute ‘Capacity’ of this concept is ‘20 GB’. As it has been indicated before, the NLP system needs a previous training in the domain. This training is based on the study of a set of texts in natural language related to the domain. In our case, this domain has been a hardware domain. 3.3. Product ontology In order to maintain a conversation, agents must have a common language. This is established by means of an ontology, which contains the main concepts owning to the domain we are dealing with. In addition to this information, the ontology also includes attributes, values, relations between concepts and axioms so that consistency checking and inferences are done. These concepts are the ones the buyer agent utilizes to indicate the customer’s preferences to the seller agent and the ones the seller agent utilizes to access the database. An (incomplete) example of a product ontology related to the computer products domain is shown in Fig. 5. In this example, ontology shows how a computer is composed by several elements: monitor, keyboard, mouse, motherboard, processor, etc. 3.3.1. Ontology editor To specify the product ontology, a graphic tool prototype has been implemented that allows to introduce concepts, attributes and semantic relations in a intuitive way. It also generates an XML document with the designed ontology’s description. Fig. 5. Computer pro The XML file has the following characteristics: , ducts Concept: !concept commentZ“” nameZ“COMPUTER”O !alternative-namesO !nameOWORKSTATION!/nameO !nameOPC!/nameO !nameOLAPTOP!nameO !/alternative-namesO !specific-attributesO !attribute commentZ“” nameZ“TRADE- MARK” typeZ“STRING”O!/attributeO !attribute commentZ“” nameZ“MODEL” typeZ“STRING”O!/attributeO !attribute commentZ“” nameZ“DESCRIP- TION” typeZ“STRING”O!/attributeO !attribute commentZ“” nameZ“GUAR- ANTY” typeZ“INTEGER”O!/attributeO !attribute commentZ“” nameZ“PRICE” typeZ“INTEGER”O!/attributeO !/specific-attributesO !/conceptO ont It can be noticed that every concept, besides the ‘main name’, has also alternative names (that we will use like synonyms), in addition to its attributes. , Relation: !relation nameZ“COMPONENT_OBJECT”O !concept_nameOPROCESSOR! /concept_nameO !concept_nameOCOMPUTER! /concept_nameO !relation_typeO !propertyONonSymmetry!/propertyO !/relation_typeO !/relationO ology. F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235 231 In this case, in addition to the concepts taking part in the semantic relation under question, the relation will have a name with the relation type and eventually some other properties associated to that relation. 3.3.2. The analyser There are two possible parsers. The API SAX accom- plishes a sequential access to the document and provides an event based programming model (callbacks). On the other hand, the API DOM makes a tree structure from the document. The main advantages of API SAX are its simplicity and that it consumes less time of processing. So, we have chosen this option because we do not need to update the document and the API SAX is a more simple solution to do the queries in the ontology. Thus, every agent only needs to do one document analysis. The buyer agent, once the concepts of the customer sentence have been found, should compare these concepts with that found in the ontology in such a way that, if they match, we understand that it is a product the customer wants to buy. 3.4. Product database The product each manufacturer possesses has been stored in a database. For it, MySQL Database Server has been chosen because, apart from being Open Source (uses the GPL—GNU General Public License), it is very fast, reliable, and easy to use MySQL Reference Manual (2002). Besides, the MySQL Database Software is a client/server system that consists of a multi-threaded SQL server that supports different backends, several different client programs and libraries, administrative tools and a wide range of application programming interfaces (APIs). A JDBC–ODBC bridge driver has been utilized to access the database from Java due to its efficiently using native Fig. 6. Connection pool database client code. Besides, the drive may be changed easily in the Web application description file. In order to add flexibility to the database management the DAO pattern has been used, so that persistent information from different sources (relational databases, LDAP, XML, etc.) can be stored and recovered. The benefits from utilizing this solution are the following: transparency-oriented, easier component migration, complexity reduction and centraliza- tion of the access to data through a single layer. We also use a connection pool to improve the performance (see Fig. 6). The general structure of the DAO pattern is shown in Fig. 7. 3.5. Web application With the purpose of applying the designed agents in a real environment, the FIPA-OS framework has been integrated into a Web application. 3.5.1. Agents integration We have to distinguish three subjects: the AgentLoader creation, the buyer agent creation and the seller agent creation. AgentLoader creation. The environment initialisation is needed in order to start working with FIPA-OS. This is done by ‘Loader’, which reads configuration files and creates the AMS and DF agents (those that form the platform kernel). The AgentLoader must be created by the FrontController Servlet (see Section 3.5.2) at the application initialisation time. The buyer agent creation. Each customer has a buyer agent associated to him/her. When a customer starts to use the application, a buyer agent is created. The problem observed then was that the agent creation process is very slowly. implementation. Fig. 7. DAO pattern implementation. F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235232 The solution provided for that problem was to use the connection pool’s idea commented before for the database, so that when there are not enough agents to associate to customers new agents are created. The seller agent creation. For each manufacturer, we will have previously set up a database containing the stock together with a seller agent associated to it and registered in the platform. In this way, the customer through his buyer agent will have access to the stock of a large number of manufacturers being able to choose a better product. 3.5.2. Application design The design is based on the Apache Struts Web Application Framework. In this way, we only have one Servlet (FrontController) where all petitions go to. This Servlet studies every petition and executes the action that corresponds to the petition. 4. The system prototype A software tool based on the approach described in this work has been designed and implemented for facilitating e-commerce. This is a Web application implemented with Java technology and designed following Struts recommen- dations. The application domain on which it has been applied is computer hardware. The user has been provided with two ways to introduce a query. On the one hand, it is possible to write free text with the user’s desires and, on the other hand, the user can select specific product categories from those indicated in Fig. 8. We have trained the system with some sentences a user can write to ask for a hardware component. In this process, when we introduce a sentence to the tool, we have to establish the concepts, attributes, values and relations this sentence contains. For example, if we say I want a 60 GB hard disk, we will have to indicate that there is a concept, ‘hard disk’, with an attribute, Capacity, which value in this example is ‘60 GB’. Once the tool have learnt this, if we introduce a new sentence like I want a 256 MB memory, the tool will automatically determine that there is a concept ‘memory’, with the attribute Capacity which value is ‘256 MB’. Once the user has entered information into the system about the wanted products, the system creates a set with the product categories that match the user wishes. Once the communication with the seller agents is established, Fig. 8. Screenshot showing product categories. F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235 233 the application shows a page with the best products of each category. Then, the user is shown a list of all the products related to a found category sorted by an internal relation quality-prize. Finally, for each product the user is interested in (according to the information supplied by her/him to the system), the user will be able to ask for the amount (s)he wants (see Fig. 9). Internally, this request will be addressed Fig. 9. List of products so to the seller agent that represents the manufacturer which has the product under question in its catalog. 5. Validation of the system The system described in previous sections has been employed by eight users, who were given two hardware rted by categories. Table 4 System scoring Question Score (by user) U1 U2 U3 U4 U5 U6 U7 U8 Previous experience in suchlike application 0 5 8 6 7 4 1 3 Interface 7 6 6 3 5 5 6 5 Performance – 6 5 3 4 8 6 5 Application utility 8 8 – 6 5 8 7 8 Easy use 9 7 8 6 7 9 9 8 Agree with solution 7 6 6 5 5 8 7 6 Utility (Do they think new characteristics are important (DSS, NLP,.?)) 9 6 7 7 – 10 9 8 F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235234 product catalogs. Each catalog had products belonging to seven categories: hard disk, mouse, graphic card, sound card, monitor, motherboard and CD player. Moreover, each catalog possessed three items. The mode the system was deployed is the following. In one computer, we installed the Web application and one of the two seller agents (the one which accesses to its catalog). On the other hand, in a distinct computer we set up then the seller agent to make queries on the another available catalog. The users were asked to use the system during half an hour and to make at least 10 requests to the system. Utilities like search by all product categories to browse between products in an easy way were assessed as important for them. The main problem found by them was the performance (caused by the time consumed by the NLP module). They all think that the interface is intuitive but too simple and agree in the fact that the functionalities added to the e-commerce application are positive. Finally, the goodness of solution was well assed for most of them: they think that it is necessary to take additional attributes into account (price, brand, product characteristics, etc.) in order to minimize the space of search. These results are summarized in Table 4. 6. Discussion and conclusion The results presented in this paper contribute to the development of intelligent electronic-commerce appli- cations that are capable of simulating real world transactions. Concerning related work, a similar solution to ours can be found in Silverman et al. (2001). However, while in our solution we split the functionality into two different agent roles (buyer and seller), the quoted author makes use of an unique search agent that integrates all the functionality. Using intelligent agent technologies, NLP, decision support systems, ontologies and web technologies we have made a system that facilitates B2C (business-to-consumer) and B2B (business-to-business) transactions. The agents have been designed and implemented within a multiagent framework, FIPA-OS, that follow a broad standard such as FIPA (Foundation for Intelligent Physical Agents). It allows to develop a distributed application in a very simple way. All this makes the development process more realistic, simulating the interaction between a store customer and the seller, which attempts to help the buyer to find the product that (s)he is looking for. At the same time, the solution presented in this paper also gives the user an object pre-classification that supports the user decision. The system developed has made use of advanced technologies to build an intelligent application but there are some aspects that we have not taken into account such as security, auctions, efficiency, etc. In particular, a few interesting issues should be explored and improved in future system improvements. First, a more complete ontology (i.e. with more knowledge) must be created. The application must be able to take into account and to use all the knowledge contained in the ontology (axioms, attri- butes, etc.). Second, it is important to better train the NLP tool allowing more free text, so that the application be able to understand the customer’s wishes and to do a better search. Third, in the current approach the agents must follow a rigid plan created by an STRIPS-like planner. We can use a more flexible planner, which uses more knowledge about the environments and creates plans that take into account the possible contingencies that could take place. Fourth, we can make the application totally product- independent. This implies the intensive use of the ontology in all application fields. Fifth, both the buyer user and the seller user might be given the possibility of configuring their respective agents. So, a buyer agent could be configured with the buyer user’s preferences (e.g. the cheapest product, or the one with more quality,.) and something similar with the seller agent. Finally, it could be interesting to use XML messages for the agents to communicate with one another, as a number of manufacturers currently do. Acknowledgements We thank the Spanish Ministry for Science and Technology for its support for the development of F. Garcı́a-Sánchez et al. / Expert Systems with Applications 28 (2005) 223–235 235 the system through projects TIC2002-03879, FIT-110100- 2003-73 and FIT-150500-2003-503; the Regional Govern- ment of Murcia (Spain) through project 2I03SIU0039; and Seneca Foundation through project PI-16/0085/FS/01. We also thank the European Commission for its support under project ALFA II0092FA. References Fikes, R. E., & Nilsson, N. J. (1971). Strips: A new approach to the application of theorem proving to problem solving. Artificial Intelli- gence, 2, 189–208. FIPA-OS Developers Guide (2001). http://fipa-os.sourceforge.net/, February. Guttman, R. H., Moukas, A. G., & Maes, P. (1998). Agent-mediated electronic commerce: A survey. The Knowledge Engineering Review, 13(2), 147–159. Hoffmann, J. (2001). The fast-forward planning system. AI Magazine, 57– 61. Kang, B. (1996). Multiple classification ripple down rules. PhD Thesis, University of New South Wales. MySQL Reference Manual (2002). http://www.mysql.com/. Refanidis, I., & Vlahavas, I. (2001). The GRT planner. AI Magazine, 63–65. Ruiz-Sánchez, J. M., Valencia-Garcı́a, R., Fernández-Breis, J. T., Martı́nez Béjar, R., & Compton, P. (2003). An approach for incremental knowledge acquisition from text. Expert Systems with Applications, 25, 77–86. Silverman, B. G., Bachann, M., & Al-Akharas, K. (2001). Implications of buyer decisión theory for design of e-commerce websites. International Journal of Human Computer Studies, 55, 815–844. Slaney, J., & Thiébaux, S. (2001). Blocks World revisited. Artificial Intelligence, 125, 119–153. Valencia-Garcı́a, R., Ruiz-Sánchez, J. M., Vivancos-Vicente, P. J., Fernández-Breis, J. T., & Martı́nez-Béjar, R. (2004). An incremental approach for discovering medical knowledge from texts. Expert Systems with Applications, 26(3), 291–299. Van Heijst, G., Schreiber, A. T., & Wielinga, B. J. (1997). Using explicit ontologies in KBS development. International Journal of Human Computer Studies, 45, 183–292. Vlassis, N. (2003). A concise introduction to multiagent systems and distributed AI. http://carol.science.uva.nl/~vlassis/cimasdai/. Wooldridge, M. (2002). An introduction to MultiAgent systems. New York: Wiley. http://fipa-os.sourceforge.net/ http://www.mysql.com/ http://carol.science.uva.nl/~vlassis/cimasdai/ An integrated approach for developing e-commerce applications Introduction Methodology Multiagent systems Decision support systems (DSS) Ontologies Natural language processing (NLP) A framework for developing e-commerce applications Overview of the system FIPA-OS and intelligent agents Product ontology Product database Web application The system prototype Validation of the system Discussion and conclusion Acknowledgements References 
work_jx4y7pq42nhabgcfkfylz2q72e	----	Title On the Coordination of Multidisciplinary Design Optimization Using Expert Systems Andrew R. Price1, Andy J. Keane1, and Carren M.E. Holden2 1 School of Engineering Sciences, University of Southampton, Southampton, SO17 1BJ, UK 2 Airbus Operations Ltd., Bristol, BS99 7AR, UK a.r.price@soton.ac.uk 1 Introduction In the design of complex engineering systems involving multiple disciplines it is critical that the interactions between the subsystems of the problem are ac- counted for. Only by considering the fully coupled system can an optimal design emerge. Formal multidisciplinary design optimization (MDO) methods [1] fall into two broad categories; 1) monolithic formulations where a single optimizer addresses the whole problem and 2) multilevel methods where the problem is decomposed along disciplinary lines and optimization takes place at both a sys- tem and domain level. The single optimizer approach is simple to implement but can scale poorly for larger problems and increasing number of disciplines. It may also prove problematic in an industrial setting to bring all of the domain analysis tools under the control of a single optimizer. Multilevel architectures promote discipline autonomy. The system level is responsible for managing interactions between disciplines. Such an approach allows design teams to work in relative isolation based upon targets set at the system level. If MDO methods are to be accepted in an industrial context they must support this form of distributed design optimization for both organizational and computational reasons. In this work a related approach is proposed; that of replacing the formal system level optimizer with an expert system to reason over information from the domains and make decisions about changes to the common design variables vector or bounds. Such an approach sacrifices, possibly elusive, guarantees of convergence for potentially attractive returns in the enterprise. 2 Coordination of MDO Using an Expert System An investigative framework has been developed exploiting an expert system as the coordinating process for multidisciplinary design optimization. This system level “master” process has access to a central repository of information which details both the present state of the design and the history of the MDO search. Data mining is employed to analyze the content of this database to present the expert system with facts about features in the domain and system level opti- mization data. The expert system employs a rule base to make decisions about C. Blum and R. Battiti (Eds.): LION 4, LNCS 6073, pp. 212–215, 2010. c© Springer-Verlag Berlin Heidelberg 2010 On the Coordination of MDO Using Expert Systems 213 how the domain level design optimizations should proceed. The results of the reasoning of the expert system are written into the central database and the domains, acting asynchronously, perform the next local optimization as resource becomes available. The expert system controls the design process by specifying the bounds and parameters provided as input to the domain optimizers work- ing on their part of the decomposed problem. A rule base has been developed that solves the design problem by narrowing in on single values for the shared design variables through systematic reduction of their bounds, by managing the exchange and relaxation of the state coupling variables between the domains and by specifying the start points for the domain optimizers. In this work, the performance of the rule base is explored using two types of optimizer in the domains; a sequential quadratic program and a genetic algorithm. To assess the performance of the rule based coordination a number of stan- dard MDO algorithms from the literature have been implemented in Matlab using the SQP method fmincon. These include the methods: Multiple Disci- pline Feasible (MDF) [2], Individual Discipline Feasible (IDF) [2], All-At-Once (AAO), Collaborative Optimization [3], Bi-Level Integrated System Synthesis (BLISS) [4] and Multidisciplinary Design Optimization based on Independent Subspaces (MDOIS) [5]. A number of MDO problems have also been assembled from the literature ranging from simple numerical constructs, through relatively simple preliminary aircraft design problems to a cut-down and decomposed ver- sion of a commercial aircraft wing design tool. The problems have been imple- mented in both the rule base framework and the Matlab MDO framework to enable comparison of performance in both qualitative and quantitative terms. We present the results of the application of the MDO methods to two example MDO problems. The first numerical problem is taken from the third exam- ple study presented in Yi et al. [6] involving two disciplines. The second is a subsonic passenger aircraft design problem described by Lewis [7]. The prob- lem is also composed of two domains; an aerodynamics model and a weights model. 3 Results The results for the Yi3 problem are presented in Table 1. This minimization problem is solved by all methods and the rule base (RB) performs well in this instance. The global optimum value of the system objective function f = 0.5 is found exactly when using the SQP optimizer in the domains and is found less accurately when using the GA in the domains. However, it is noted that the problem does not have a unique global optimum and admits a number of solutions with the optimal system objective function value f = 0.5. The sin- gle optimizer methods all solve the problem using only two or three system level iterations. The bi-level methods need significantly more iterations for this problem. The MDO methods solve the problem to the tolerances set for the optimizers with the exception of CO which does not converge well. The rule base approach requires 8 and 18 system level iterations for the SQP and GA 214 A.R. Price, A.J. Keane, and C.M.E. Holden Table 1. Performance of MDO methods for the Yi et al. example 3 Method Objective Function Maximum Constraint (g ≤ 0) Number system iterations Domain-1 analysis calls Domain-2 analysis calls MDF 0.5000 0.0000 2(5) 281 281 IDF 0.5000 1.1102 × 10−16 2 19 19 AAO 0.5000 −2.2204 × 10−16 3 33 33 CO 0.4998 2.0609 × 10−4 148 19530 18376 BLISS 0.5000 2.6671 × 10−7 66(66) 4941 4950 MDOIS 0.5000 −4.2723 × 10−8 21(21) 1168 1184 RB (SQP) 0.5000 0.0000 8 280 225 RB (GA) 0.5008 −7.4939 × 10−4 18 25550 25550 domain level optimizers respectively and is competitive with the other bi-level methods. The performance of the algorithms is broadly comparable with the per- formance figures reported in Yi et al. [6] with the slightly greater number of func- tion calls required in our framework likely attributable to the higher tolerances used. Table 2 summarizes the results for the subsonic aircraft design problem. A con- sistent optimum is not found across the methods investigated but the rule base performs well compared to the other bi-level methods (CO, BLISS, MDOIS). The rule based approach (GA) and the MDF method find the best results. CO and BLISS exhibit poor performance for this problem and do not converge to an acceptable feasible solution. For BLISS it is possible that the trust region algorithm could be improved here but the performance of the algorithm on this problem, and others in our test suite, shows that it will often take the search to the bounds of the design variables. Conversely, the rule base in this case, finds a solution close to that of MDF. The broader search achieved using a GA in the domains provides an advantage over SQP for these problems. This also indicates that performance gains may be possible by improving the rules for managing the domain optimization start points. Table 2. Performance of MDO methods for the subsonic aircraft design problem Method Objective Function Maximum Constraint (g ≤ 0) Number system iterations Domain-1 analysis calls Domain-2 analysis calls MDF −2.0676 −5.7335 × 10−11 13(27) 668 668 IDF −2.0152 0.0 5 67 67 AAO −1.9629 −1.4627 × 10−4 5 127 127 CO −2.0139 3.1050 × 10−2 250∗ 156177 469726 BLISS −1.6035 6.8202 × 10−2 7(7) 148 148 MDOIS −1.9706 −6.9561 × 10−7 8(8) 297 308 RB (SQP) −1.9735 0.0 46 2406 923 RB (GA) −2.0549 −4.7834 × 10−5 70 86870 94024 On the Coordination of MDO Using Expert Systems 215 4 Discussion and Conclusions The rule base approach is found to work well and has the advantage that it is relatively straight forward to integrate into existing organizational infrastruc- ture. However, further work is required to assess whether the pragmatic rule base approach, that sacrifices formal guarantees of convergence, will be truly competitive across a large range of MDO problem. The relative ease with which a rule based system level control process can be implemented and managed is a significant advantage over methods like BLISS which can prove difficult to implement. Both BLISS and CO require domain experts to optimize constructs of the process rather than investigate the physics of the problem. Initial studies have involved a number of MDO problems ranging from simple numerical schemes, through basic aircraft sizing studies to a cut-down commer- cial in-house design tool. The initial rule base works by managing the bounds of the shared design variable vector until the enclosed hyper-volume converges to a specified tolerance and all domains are feasible. The performance of the rule base is found to be competitive for a range of MDO problems (of which only two are presented herein). Future work will extend the use of data mining of domain optimizers for improved feature recognition and development of a more sophisticated rule base to improve performance across the range of problems assembled. Acknowledgments The work is funded by the TSB funded NGCW/MDOW project and Airbus UK whose support is gratefully acknowledged. References 1. Sobieszczanski-Sobieski, J., Haftka, R.T.: Multidisciplinary aerospace design opti- mization: survey of recent developments. Structural and Multidisciplinary Optimiza- tion 14, 1–23 (1997) 2. Cramer, E.J., Dennis, J.J.E., Frank, P.D., Lewis, R.M., Shubin, G.R.: Problem Formulation for Multidisciplinary Optimization. SIAM Journal on Optimization 4, 754–776 (1994) 3. Braun, R.: Collaborative Optimization: An Architecture for Large-Scale Distributed Design. PhD Thesis, Stanford University (1996) 4. Sobieszczanski-Sobieski, J., Agte, J.S., Sandusky, R.R.: Bilevel Integrated System Synthesis. AIAA Journal 38, 164–172 (2000) 5. Shin, M.-K., Park, G.-J.: Multidisciplinary design optimization based on indepen- dent subspaces. International Journal for Numerical Methods in Engineering 64, 599–617 (2005) 6. Yi, S., Shin, J., Park, G.: Comparison of MDO methods with mathematical exam- ples. Structural and Multidisciplinary Optimization 35, 391–402 (2008) 7. Lewis, K.: An Algorithm for Integrated Subsystem Embodiment and System Syn- thesis. PhD Thesis, Georgia Institute of Technology, Atlanta (1997) On the Coordination of Multidisciplinary Design Optimization Using Expert Systems Introduction Coordination of MDO Using an Expert System Results Discussion and Conclusions References << /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Gray Gamma 2.2) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (ISO Coated v2 300% \050ECI\051) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Error /CompatibilityLevel 1.3 /CompressObjects /Off /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Perceptual /DetectBlends true /DetectCurves 0.1000 /ColorConversionStrategy /sRGB /DoThumbnails true /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams true /MaxSubsetPct 100 /Optimize true /OPM 1 /ParseDSCComments true /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo true /PreserveFlatness true /PreserveHalftoneInfo false /PreserveOPIComments false /PreserveOverprintSettings true /StartPage 1 /SubsetFonts false /TransferFunctionInfo /Apply /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true ] /NeverEmbed [ true ] /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 149 /ColorImageMinResolutionPolicy /Warning /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 150 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.50000 /EncodeColorImages true /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.40 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /ColorImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 149 /GrayImageMinResolutionPolicy /Warning /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 150 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.40 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /GrayImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 599 /MonoImageMinResolutionPolicy /Warning /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 600 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 >> /AllowPSXObjects false /CheckCompliance [ /None ] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description << /ARA <FEFF06270633062A062E062F0645002006470630064700200627064406250639062F0627062F0627062A002006440625064606340627062100200648062B062706260642002000410064006F00620065002000500044004600200645062A064806270641064206290020064406440637062806270639062900200641064A00200627064406450637062706280639002006300627062A0020062F0631062C0627062A002006270644062C0648062F0629002006270644063906270644064A0629061B0020064A06450643064600200641062A062D00200648062B0627062606420020005000440046002006270644064506460634062306290020062806270633062A062E062F062706450020004100630072006F0062006100740020064800410064006F006200650020005200650061006400650072002006250635062F0627063100200035002E0030002006480627064406250635062F062706310627062A0020062706440623062D062F062B002E0635062F0627063100200035002E0030002006480627064406250635062F062706310627062A0020062706440623062D062F062B002E> /BGR <FEFF04180437043f043e043b043704320430043904420435002004420435043704380020043d0430044104420440043e0439043a0438002c00200437043000200434043000200441044a0437043404300432043004420435002000410064006f00620065002000500044004600200434043e043a0443043c0435043d04420438002c0020043c0430043a04410438043c0430043b043d043e0020043f044004380433043e04340435043d04380020043704300020043204380441043e043a043e043a0430044704350441044204320435043d0020043f04350447043004420020043704300020043f044004350434043f0435044704300442043d04300020043f043e04340433043e0442043e0432043a0430002e002000200421044a04370434043004340435043d043804420435002000500044004600200434043e043a0443043c0435043d044204380020043c043e0433043004420020043404300020044104350020043e0442043204300440044f0442002004410020004100630072006f00620061007400200438002000410064006f00620065002000520065006100640065007200200035002e00300020043800200441043b0435043404320430044904380020043204350440044104380438002e> /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000410064006f006200650020005000440046002065876863900275284e8e9ad88d2891cf76845370524d53705237300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef69069752865bc9ad854c18cea76845370524d5370523786557406300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /CZE <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> /DAN <FEFF004200720075006700200069006e0064007300740069006c006c0069006e006700650072006e0065002000740069006c0020006100740020006f007000720065007400740065002000410064006f006200650020005000440046002d0064006f006b0075006d0065006e007400650072002c0020006400650072002000620065006400730074002000650067006e006500720020007300690067002000740069006c002000700072006500700072006500730073002d007500640073006b007200690076006e0069006e00670020006100660020006800f8006a0020006b00760061006c0069007400650074002e0020004400650020006f007000720065007400740065006400650020005000440046002d0064006f006b0075006d0065006e0074006500720020006b0061006e002000e50062006e00650073002000690020004100630072006f00620061007400200065006c006c006500720020004100630072006f006200610074002000520065006100640065007200200035002e00300020006f00670020006e0079006500720065002e> /ESP <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> /ETI <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> /FRA <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> /GRE <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a stvaranje Adobe PDF dokumenata najpogodnijih za visokokvalitetni ispis prije tiskanja koristite ove postavke. Stvoreni PDF dokumenti mogu se otvoriti Acrobat i Adobe Reader 5.0 i kasnijim verzijama.) /HUN <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> /ITA <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> /JPN <FEFF9ad854c18cea306a30d730ea30d730ec30b951fa529b7528002000410064006f0062006500200050004400460020658766f8306e4f5c6210306b4f7f75283057307e305930023053306e8a2d5b9a30674f5c62103055308c305f0020005000440046002030d530a130a430eb306f3001004100630072006f0062006100740020304a30883073002000410064006f00620065002000520065006100640065007200200035002e003000204ee5964d3067958b304f30533068304c3067304d307e305930023053306e8a2d5b9a306b306f30d530a930f330c8306e57cb30818fbc307f304c5fc59808306730593002> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020ace0d488c9c80020c2dcd5d80020c778c1c4c5d00020ac00c7a50020c801d569d55c002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /LTH <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> /LVI <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> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken die zijn geoptimaliseerd voor prepress-afdrukken van hoge kwaliteit. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <FEFF004200720075006b00200064006900730073006500200069006e006e007300740069006c006c0069006e00670065006e0065002000740069006c002000e50020006f0070007000720065007400740065002000410064006f006200650020005000440046002d0064006f006b0075006d0065006e00740065007200200073006f006d00200065007200200062006500730074002000650067006e0065007400200066006f00720020006600f80072007400720079006b006b0073007500740073006b00720069006600740020006100760020006800f800790020006b00760061006c0069007400650074002e0020005000440046002d0064006f006b0075006d0065006e00740065006e00650020006b0061006e002000e50070006e00650073002000690020004100630072006f00620061007400200065006c006c00650072002000410064006f00620065002000520065006100640065007200200035002e003000200065006c006c00650072002000730065006e006500720065002e> /POL <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> /PTB <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> /RUM <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> /RUS <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> /SKY <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> /SLV <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> /SUO <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> /SVE <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> /TUR <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> /UKR <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> /ENU (Use these settings to create Adobe PDF documents best suited for high-quality prepress printing. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.) /DEU <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> >> /Namespace [ (Adobe) (Common) (1.0) ] /OtherNamespaces [ << /AsReaderSpreads false /CropImagesToFrames true /ErrorControl /WarnAndContinue /FlattenerIgnoreSpreadOverrides false /IncludeGuidesGrids false /IncludeNonPrinting false /IncludeSlug false /Namespace [ (Adobe) (InDesign) (4.0) ] /OmitPlacedBitmaps false /OmitPlacedEPS false /OmitPlacedPDF false /SimulateOverprint /Legacy >> << /AddBleedMarks false /AddColorBars false /AddCropMarks false /AddPageInfo false /AddRegMarks false /ConvertColors /ConvertToCMYK /DestinationProfileName () /DestinationProfileSelector /DocumentCMYK /Downsample16BitImages true /FlattenerPreset << /PresetSelector /MediumResolution >> /FormElements false /GenerateStructure false /IncludeBookmarks false /IncludeHyperlinks false /IncludeInteractive false /IncludeLayers false /IncludeProfiles false /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe) (CreativeSuite) (2.0) ] /PDFXOutputIntentProfileSelector /DocumentCMYK /PreserveEditing true /UntaggedCMYKHandling /LeaveUntagged /UntaggedRGBHandling /UseDocumentProfile /UseDocumentBleed false >> ] >> setdistillerparams << /HWResolution [2400 2400] /PageSize [595.276 841.890] >> setpagedevice 
work_jxg34bzn2zhejdpe2f3ali3ntq	----	introduction Politeknik Dergisi Journal of Polytechnic Cilt:11 Sayı: 2 s.89-92, 2008 Vol: 11 No: 2 pp.89-92, 2008 89 Makale 30.01.2008 tarihinde gelmiş, 06.02.2008 tarihinde yayınlanmak üzere kabul edilmiştir. C. ELMAS, U. GÜVENÇ, Gazi Üniversitesi Teknik Eğitim Fakültesi Elektrik Eğitimi Bölümü, ANKARA, celmas@gazi.edu.tr, ugurguvenc@gazi.edu.tr R. DEMİRCİ, Düzce Üniversitesi Teknik Eğitim Fakültesi Elektrik Eğitimi Bölümü, DÜZCE recepdem@gmail.com Digital Object Identifier 10.2339/2008.11.2.89-92 Yön Bağımsız Yayınım Filtresinin Plaka Görüntülerindeki Gürültülerin Temizlenmesindeki Etkisi Çetin ELMAS, Recep DEMİRCİ, Uğur GÜVENÇ ÖZET Plaka karakter tanıma, araçların plakaları vasıtasıyla tanınmasına yarayan bir görüntü işleme uygulamasıdır. Günümüzde plaka karakter tanıma için geliştirilen birçok görüntü işleme teknikleri kötü hava koşulları ve kirli plakalar yüzünden etkili sonuçlar verememektedir. Bu olumsuz yönleri yok etmek ve etkili bir sonuç alabilmek için ortalama, ortanca, wiener gibi çeşitli görüntü filtreleme teknikleri kullanılmaktadır. Ancak kullanılan bu filtreleme teknikleri görüntüdeki gürültüyü yok ederken orijinal görüntünün kenarlarında kalınlaşma ve bozulmalar meydana getirmektedir. Bu çalışmada, görüntüdeki gürültünün temizlenmesi için yön bağımsız yayınım filtreleme yöntemi, dört farklı yayınım fonksiyonu ile test edilmiştir. Yapılan testlerde, yön bağımsız yayınım filtreleme sonucu görüntüdeki en az bozulmanın Perona-Malik’in ikinci fonksiyonu kullanıldığında ortaya çıktığı görülmüştür. Anahtar Kelimeler: Yön bağımsız yayınım, yayınım katsayısı, plaka karakter tanıma Effect of Anisotropic Diffusion Filter at the Reducing on Plate Images Noises ABSTRACT Plate number recognition is one of image processing application, which is used to identify cars. Nowadays, various types of image processing methods are used to recognize plate number. However, they could not produce satisfactory results because of bad weather conditions and dirty plates. Image filters such as mean, median, wiener are used in order to avoid from bad effects and to achieve effective results. On the other hand, the conventional filters destroy the edge information in the plate image while denoising the images. In this study, in order to filter the plate image, anisotropic diffusion filtering method has been tested with four different diffusion functions. Accordingly, it was observed that the minimum distortion occurred when it was realized by using Perona Malik’s second function. Keywords: Anisotropic diffusion, diffusion coefficients, plate character recognition 1. GİRİŞ Plaka karakter tanıma sistemlerini kullanım amaçları giriş/çıkış izni, özel alanların kontrolü, ücretlendirme, araç bulma ve raporlama olarak söylene- bilir. Plaka karakter tanıma askeri alanlar, süper mar- ketler, benzin istasyonlarında, havaalanlarında, gişe otomasyonlarında, sınır kapılarında, hastanelerde v.b. birçok alanda kullanılmaktadır. Hiç bir görüntü elde etme metodu gürültüden ba- ğımsız değildir. Ayrıca elde edilen görüntüler, oluşu- mundaki ve iletim sırasında meydana gelen bozulmalar- dan dolayı doğrudan kullanılabilecek durumda değildir (1). Bundan dolayı meydana gelen bu gürültülerin gö- rüntü üzerinden etkisi yok edilmeden temel görüntü iş- lemleri yapılamaz. Bu olumsuz yönleri yok etmek ve etkili bir sonuç alabilmek için görüntünün filtrelenmesi gerekmektedir. Ancak bir çok filtreleme teknikleri gö- rüntüdeki gürültüyü yok ederken orijinal görüntüde bo- zulmalar meydana getirmektedir. Görüntü işleme ile ilgili çalışmalar 1830’lu yıl- larda ilk fotoğrafların ortaya çıkmasıyla birlikte başla- mıştır. Ancak gerçek manada ilk görüntü işleme uygu- laması, 1964 yılında ABD tarafından aya gönderilen bir uydunun aydan görüntülediği fotoğraflarda meydana gelen görüntü bozukluklarını düzeltilmesidir (2). Bun- dan sonra gürültü yok edilmesi ve azaltılmasına dair birçok araştırmalar yapılmıştır. Görüntü filtreleme iş- leme için bir çok filtreleme teknikleri kullanılmaktadır (3-5). Son yıllarda kısmi diferansiyel denklemler temelli teknikler görüntü filtreleme için görüntü işlemede kul- lanılmaktadır. Koenderick (6) kısmi diferansiyel denk- lemler tabanlı görüntü işlemede ilk çalışanların başında gelmektedir. 1987’de Perona ve Malik (7) tarafından yön bağımsız yayınım adı verilen bir algoritma gelişti- rildi. Bu algoritma görüntü filtrelemede çok iyi sonuçlar mailto:celmas@gazi.edu.tr Çetin ELMAS, Recep DEMİRCİ, Uğur GÜVENÇ / POLİTEKNİK DERGİSİ, CİLT 11, SAYI 2, 2008 90 verdi. Bu algoritmayı diğerlerinden ayırt eden iki önemli özelliği vardır. Bunlar gürültü yok edilirken ke- narların korunmasıdır. Başka bir deyişle gürültüyü yok ederken aynı zamanda, filtreleme esnasında çoğu gö- rüntü işlemede tekniklerinde neredeyse imkânsız olan kenarları koruyabilme yeteneğine sahiptir. Bu çalışmada yön bağımsız yayınım filtreleme- nin sonuçlarını görmek için gerçek uygulama alanı ola- rak plaka karakter tanıma sistemleri seçilmiştir. Bunun sebebi gittikçe artan bir ilgiye sahip olan bu sistemlerin uygulanma alanlarının geniş olması ve uygulamalarda gürültüden meydana gelen dezavantajlarının olmasıdır. Yapılan uygulamada gürültünün temizlenmesi için yön bağımsız yayınım filtreleme için geliştirilen dört değişik fonksiyon test edilerek orijinal görüntüde en az bozul- mayı meydana getiren yayınım katsayısı fonksiyonu bulunmuştur. 2. YÖN BAĞIMSIZ YAYINIM Perona ve Malik (PM) tarafından geliştirilen yön bağımsız yayınım algoritma görüntü filtrelemede çok iyi sonuçlar verdi. Bu algoritmayı diğerlerinden ayırt eden iki önemli ve istenen yol olmuştur. Bunlar gürültünün yok edilmesi ve kenar koruma vazifesi için uygun ol- malarıdır. Perona-Malik’in öncülüğünü yaptığı araştır- mada yön bağımsız yayınım denklemi; ( )( ) ( ) ( ) ItyxcItyxc Ityxcdiv t I ∇⋅∇+Δ⋅= ∇⋅= ∂ ∂ ,,,, ,, (1) burada: I = piksel değerlerinin yoğunluğu, div = Diverjans operatörü, ∇ = Gradyant operatörü, Δ = Laplas operatörü ve ( )tyxc ,, = yayınım katsayısını ifade etmektedir. ( )tyxc ,, fonksiyonu, gradyant büyüklüğünün bir fonk- siyonu olarak seçilebilmektedir; ( ) ( )tyxIgtyxc ,,(,, ∇= (2) burada ( )•g sınır bölgelerinin tahmininde kullanılmak- tadır. PM bir jiI , görüntüsünün dört yakın komşuna bağlı olarak yön bağımsız yayınım eşitliğini aşağıdaki gibi geliştirdi: [ ] t jiBBDDGGKKt jit ji IcIcIcIcII ,,1, ∇⋅+∇⋅+∇⋅+∇⋅+=+ λ (3) BD,G,K,=k için ( )tyxIgc ktk ji ,,(, ∇= 3. YAYINIM KATSAYILARI Yön bağımsız yayınımın ayrık uyarlamasına ba- kıldığında filtrelemede kullanılan iki değişken bulun- maktadır. Bunlardan birincisi ve en önemlisi yayınım katsayısıdır. Yayınım katsayısı sonucu direkt olarak et- kilemektedir. Bundan dolayı yayınım katsayısının be- lirlenmesi yön bağımsız yayınımın en önemli problemi- dir. Şimdiye kadar birçok yayınım katsayısı eşitlikleri geliştirildi. Bunlardan ilk ikisi yön bağımsız yayınımın ayrık uyarlaması teorisini geliştiren Perona ve Malik’ tir. Bu eşitlikler: ( ) 2 1 1 ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ ∇ + =∇ K I Ig (4) ve ( ) ( ) ⎟ ⎠ ⎞⎜ ⎝ ⎛ ∇− =∇ 2/ KI eIg (5) dir. Denklem 4’de ifade edilen eşitlik Perona-Malik1 ve 5’de ifade edilen eşitlik Perona-Malik2 eşitlikleri olarak adlandırılmaktadır. Burada K pozitif değerli bir katsayı- dır. K değerini bir eşik değeri olarak düşünülebilir. Perona-Malik yön bağımsız yayınım filtrelemede ayrık uyarlamayı geliştirdikten sonra yayınım katsayısı- nın belirlenmesi için çeşitli fonksiyonlar geliştirildi. Bu- rada bunlardan en önemli olan iki yayınım fonksiyonla- rından birincisi 1994 yılında Charbonnier’in (8) geliştir- diği eşitlik 6’da verilen fonksiyondur. ( ) 2 1 1 ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ ∇ + =∇ K I Ig (6) Diğer bir yayınım fonksiyonu olan ve Weickert’in (9) 1996 yılında geliştirdiği fonksiyon eşitlik 7’de verilmiş- tir. ( ) ⎪ ⎪ ⎪ ⎩ ⎪ ⎪ ⎪ ⎨ ⎧ >∇ ⎟⎟ ⎟ ⎟ ⎟ ⎟ ⎠ ⎞ ⎜⎜ ⎜ ⎜ ⎜ ⎜ ⎝ ⎛ ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ ∇ − − =∇ =∇ 0 33666.2 exp1 01 4 Ieger K I Ieger Ig (7) 4. UYGULAMA SONUÇLARI Plakadaki karakterin tanımlanması genel olarak karakterlerin ayrıştırılması sonucunda farklı teknikler kullanarak belirlenmektedir. Burada en önemli işlem ka- rakterlerin doğru bir şekilde ayrıştırılmasıdır. Uygula- malarımızda kameradan alınan görüntünün filtrelenme- den önceki ve sonraki sonuçlarını görmek için otomatik görüntü ayrıştırılma(10) algoritması kullanılmıştır. Ya- pılan çalışmada kameradan alınan plaka görüntüsü Şekil 1(a)’da, yakınlaştırılmış plaka görüntüsü Şekil 1(b)’de ve ayrıştırılmış plaka görüntüsü Şekil 1(c)’de gösteril- miştir. YÖN BAĞIMSIZ YAYINIM FİLTRESİNİN PLAKA GÖRÜNTÜLERİNDEKİ GÜRÜ … / POLİTEKNİK DERGİSİ, CİLT 11, SAYI 2, 2008 91 a) b) c) Şekil 1. Plaka a) Orijinal görüntü b) Yakınlaştırılmış hali c) Ayrıştırılmış hali Kameradan alınıp yakınlaştırılan görüntüye, dört değişik eşitlik yön bağımsız yayınım filtreleme kullanılmıştır. Filtrelemeden sonra plakadaki her bir karakter ayrı ayrı ne kadar pikselden oluştuğu incelenmiştir. Böylece orijinal görüntü ile filtreleme sonrası görüntü arasındaki piksel sayısı karşılaştırılarak kenarlarda meydana gelen bozulma derecelendirilmiştir. Her bir ayrı eşitlikte elde edilen filtreleme sonuçları Şekil 2’de gösterilmiştir. a) b) c) d) Şekil 2. Plakanın filtrelemeden sonraki ayrıştırılmış durumları a) PM1 b) PM2 c) Charbonnier d) Weickert Uygulamadan alınan sonuçlara göre plakanın her bir ayrı karakterinin kaç piksel oluştuğunu gösteren değerler Çizelge 1’de verilmiştir. Çizelge 2’de ise orijinal görüntü ile filtrelemeden sonraki piksel sayıları arasındaki değişim oranları gösterilmiştir. Çizelge 2’deki veriler incelendiğinde orijinal görüntü ile filtrelemeden sonraki piksel sayıları arasındaki değişim en az %5.714 ile PM2 olduğu görülmektedir. Çizelge 1. Plakadaki her bir karakterin toplam piksel sayıları Orijinal PM1 PM2 Weickert Charbonnier 0 1921 2149 2098 2110 2233 6 1925 2037 2152 2103 2114 A 1681 1791 1746 1759 1728 N 2319 2544 2410 2409 2389 0 1786 2200 1960 1976 1925 5 1839 2055 1942 1946 2058 8 2113 2258 2063 2124 2175 2 1576 1677 1638 1649 1660 Çizelge 2. Plakada tüm karakterlerin ortalama piksel artış yüzdesi PM1 PM2 Weickert Charbonnier 10,27 5,714 6,152 7,479 5. SONUÇ Görüntü işlemede kullanılan bir çok filtreleme tekniklerinin görüntüdeki gürültüyü yok ederken orijinal görüntüde bozulmalar meydana getirmekte olduğu bilinmektedir. Bu çalışmada son zamanlarda görüntü işlemede oldukça fazla ilgiye sahip olan yön bağımsız yayınım filtreleme için geliştirilen dört değişik yayınım katsayısı fonksiyonlarının plaka karakterlerindeki gürültüler yok edildikten sonraki orijinal görüntü arasındaki değişim incelenmiştir. Yapılan filtrelemeler sonucu elde edilen görüntülerde kenarlarda kalınlaşma ve bozulmaların en az PM2 yayınım fonksiyonu ile gerçekleştirildiği görülmüştür. 6. KAYNAKLAR 1. Castleman, K. R., Digital Image Processing, Prentice Hall, NJ, 1996. 2. K.N. Plataniotis and A.N. Venetsanopoulos, Color Image Processing and Applications, Berlin: Springer Verlag, 2000. 3. I. Pitas and A. N. Venetsanopoulos, Nonlinear Digital Filters: Principles and Applications. Norwell, MA: Kluwer, 1990. Çetin ELMAS, Recep DEMİRCİ, Uğur GÜVENÇ / POLİTEKNİK DERGİSİ, CİLT 11, SAYI 2, 2008 92 4. J. Astola, P. Haavisto, and Y. Neuvo, “Vector median filter,” Proc.IEEE, vol. 78, pp. 678–689, 1990. 5. P. E. Trahanias and A.N. Venetsanopoulos, “Vector directional filters: Anew class of multichannel image processing filters,” IEEE Trans. Image Processing, vol. 2, pp. 528–534, Apr. 1993. 6. J. Koenderink, The Structure of Images, Biological Cybernetics, vol. 50, 1984, pp. 363-370. 7. P. Perona and J. Malik, “Scale-Space and Edge Detection Using Anisotropic Diffusion”, IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 12, 7 ,1990, pp. 629-939. 8. P. Charbonnier, L. Blanc-Feraud, G. Aubert and M. Barlaud, “Two deterministic half-quadratic regularization algorithms for computed imaging”, Proceedings of ICIP- 94., IEEE International Conference on Image Processing, vol. 2, 1994, pp.168–172. 9. J. Weickert, B.M. ter Haar Romeny, M. Viergever, Efficient and Reliable Schemes for Nonlinear Diffusion Filtering," IEEE Transactions on Image Processing, vol. 7, No. 3, p.398-410, 1998. 10. Demirci, R., Rule-based automatic segmentation of color images, International Journal of Electronics and Communications (AEU),60,435-442, 2006. 
work_jxvwxbkuffgs7czgluejhqlakm	----	iufro-gypsy.PDF A Web-based Expert System for Gypsy Moth Risk Assessment W.D. Potter, X. Deng, J. Li, M. Xu, Y. Wei, I. Lappas Artificial Intelligence Center, GSRC 111, University of Georgia, Athens, GA, 30605 M.J. Twery, D.J. Bennett USDA Forest Service, Northeastern Forest Research Station, Burlington, VT (Contact W.D. Potter at potter@cs.uga.edu) Abstract The gypsy moth is one of North America's most devastating exotic forest pests because it can cause the loss of valuable oak species, degraded aesthetics, loss of wildlife habitat, and detrimental effects on watersheds. Due to the increasingly wide infestation of the gypsy moth, it is important to develop decision aids that help assess the risks of this pest to our forests. Expert systems are a type of decision aid that could be applied to the area of risk assessment. We have developed the Gypsy Moth Expert System to estimate the risk that a forest stand faces from the gypsy moth based on the composition, structure, and management objectives of a particular forest. Risk assessment in this context is developed from forest susceptibility to infestation, vulnerability to damage caused by an infestation, and the hazard that management objectives for a forest may be affected if damage occurs. The system uses a straightforward set of if-then rules to classify risk. The development of a web-based expert system presented significant challenges to maintaining remote user processing integrity. Keywords: Gypsy Moth, Risk Assessment, Expert System 1. Introduction Gypsy moth (Lymantria dispar L.) is one of North America's most devastating introduced forest pests. The species originally evolved in Europe and Asia, and has existed there for thousands of years. In the late 1860’s, the gypsy moth was accidentally introduced near Boston by E. Leopold Trouvelot (Liebhold et al. 1989). About 10 years after this introduction, the first outbreaks began in Trouvelot's neighborhood. By 1890 state and federal governments began taking steps to eradicate the gypsy moth (Gerardi and Grimm 1979). These first attempts ultimately failed and since that time, the range of the gypsy moth has continued to spread (Elkinton and Liebhold 1990). It seems inevitable that the gypsy moth will continue to expand its range in the future because the gypsy moth is known for feeding on the foliage of hundreds of plant species in North America. Its most common hosts are oaks (Quercus spp.) and aspen (Populus spp.). Gypsy moth hosts are located throughout most of the continental United States (Montgomery and Wallner 1988). Gypsy moth populations are typically eruptive in North America. In a forest stand, population densities may fluctuate from 1 egg mass per hectare to over 10,000 per hectare. When densities reach very high levels, trees may become completely defoliated. Several successive years of defoliation, along with adverse contributions by other biotic and abiotic stress factors, may ultimately result in tree mortality. In some northeastern forests where trees are young and strong, few trees may die but occasionally tree mortality may be very heavy (Doane and McManus 1981). Due to the wide infestation of the gypsy moth, it is important to develop decision aids that help assess the risks of this pest to our forests. In addition to helping determine the risk of an infestation, these aids can also help when determining the type, duration, and intensity of a pest control management strategy. Expert systems are a type of decision aid that could be applied to the area of risk assessment. Expert systems are very popular today in a number of different domains (Benders and Manders 1993, Turban 1993, Stefik 1995, Holsapple and Winston 1996). Some of the most well known systems are used for diagnostic purposes. In a physician’s assistant system for example, the system receives as input information pertaining to a patient’s specific problem, and then provides as output one or more alternative diagnoses for the patient. The system possesses the knowledge of human experts including the manner in which an expert would determine a solution, so that it reaches the same conclusion that the expert would reach when presented with the same input information (Bobrow et al. 1986, Stefik 1995). We have developed a web-based expert system called GyMEs (Gypsy Moth Expert System, pronounced “jimmies”) to estimate the risk a forest stand faces from gypsy moth. GyMEs is an enhanced version of the GypsES expert system component (Thomas et al. 1998). GyMEs contains a revised and extended knowledge base as well as a more up-to-date inference mechanism. The expert system component of GypsES was originally developed for a MacIntosh computer (Twery and Elmes 1991). The output of these systems is a rough assessment of the risk to a forest stand due to gypsy moth, which is characterized either as "very low", "low", "moderate", "high" or "very high". The input is a series of environmental factors that have been observed in the stand for the last five years, as well as general information regarding the stand and management strategy. More specifically, the user of the system is prompted to answer a number of questions related to the stand itself, such as the age of the stand, its defoliation and drainage history etc. Then, the system processes the user's responses and determines the risk assessment. One of the important characteristics of our expert system, as well as almost all expert systems, is that it does not need to have the answers to every input question in order to reach a conclusion. Sometimes the conclusion can be drawn based on the user's answer to a single question, and at other times many or all of the questions must be answered to determine an assessment. This is a short-circuit feature characteristic of consulting expert systems. Another characteristic of GyMEs is that the system will not ask the same question twice. For example, if one line of reasoning is being followed by the system and the system needs to change to another line of reasoning, then responses given earlier will automatically be applied to this new approach. The system retains the information given to it by the user for future use. Thus, eliminating unnecessary user interaction. The most significant characteristic of GyMEs is that it is web-based, that is, it runs over the Internet. A forest manager can “surf” into the GyMEs web site and interact with the system to assess a particular forest stand. This is significant because each interaction over the web is disjoint or physically separate from every other interaction. In order to provide a (logically) cohesive interactive session, the system must bridge these gaps. In the following sections, we present our approach to the risk assessment problem, the knowledge that our system contains, and the conclusions that it may draw. A more precise description of the system's implementation is also presented. 2. Risk Assessment The risk assessment evaluation in GyMEs is based on a number of parameters, using a structure developed by Gottschalk and Twery (1989) and described in detail by Bennett (1995). Each parameter may have numerous acceptable values. But, what is the risk of a gypsy moth infestation? Risk is the chance we face of losing something of value combined with the expected loss based on our actions within the time frame of an appropriate planning horizon. The more valuable something is, the larger the loss we face if a loss-inducing event occurs. When managing a forest stand we take certain actions to achieve certain specified goals for the stand. Based on the conditions of the stand (i.e., how vulnerable the stand is to damage) and the estimated impact of an expected gypsy moth population on stand defoliation, we determine the level of damage (loss) that can be expected. The risk measures the extent of disruption the gypsy moth is expected to have on our management goals during the time of concern. This disruption corresponds to the loss we face if, in this case, we do not take any action to lessen the gypsy moth impact on the stand (Twery 1991, Bennett 1995). If the stand conditions and defoliation prediction indicate a very high risk, then there is a significant chance that the future condition of our stand will be different from what we want (our established goal). In GyMEs, risk is derived from a hazard rating combined with a defoliation prediction. To determine the hazard rating, we use the susceptibility and vulnerability of a stand. A defoliation prediction is an estimate of the amount of defoliation damage a stand will incur from some level of pest infestation. Associated with an infestation is the likelihood that the stand will in fact incur defoliation. This is called the stand susceptibility. The stand’s vulnerability is an assessment of the probable degree of mortality that would result from defoliation. Finally, combining susceptibility, vulnerability, and management objectives leads to a hazard rating. This, combined with a defoliation prediction, then brings us back to the risk assessment. Additional details of these concepts are discussed more fully in (Bennett 1995). Assessing the risk of a gypsy moth infestation requires knowledge and expertise due to the large number of parameters that must be considered. If we take all the combinations of parameter set values, we have the total solution space for this problem. If we add a single additional parameter value, the total number of combinations increases exponentially. In addition, the problem is aggravated by the fact that the gypsy moth not only causes damage to a forest’s health, but also interferes with the overall quality of the ecosystem. This is further complicated by the fact that a forest manager’s goals for the forest may have little to do with the actual health of the trees, but more, for example, to do with the happiness of hikers passing through the forest. Accurately assessing and predicting the risks of infestation by the pest is important for prescribing the proper short-term and long-term forest management practices. With a web-based expert system, the knowledge and expertise of the domain expert can be easily adopted and utilized by many users and at many different locations without the presence of the expert. 3. Expert Systems Development The following are the steps used in developing the GyMEs prototype. These include problem identification, knowledge base design, and knowledge base validation (Stefik 1995). An expert system requires a precise domain. The domain must be well organized and well understood. The selected application within the domain will need to require expert knowledge in order to solve specific problems within that domain. Otherwise, there would be no need for an expert system; a standard algorithmic search scheme would be more suitable in that case. This means that the types of problems encountered within the domain should exhibit combinatorially explosive solution spaces when using an exhaustive search scheme to find a solution. For example, in the diagnosis domain, as the number of disorders (diseases) increases linearly, the number of possible diagnoses increases exponentially (i.e., there are exponentially more combinations of diagnoses to consider). This type of growth in the total number of solutions is called combinatorial explosion. Clearly, gypsy moth risk assessment is an appropriate domain for expert systems development. The knowledge base is the core component of any expert system since it contains the knowledge acquired from an expert in the field. Typically, a knowledge engineer is responsible for working with an expert to build the knowledge base for the system. The knowledge engineer must perform a detailed analysis of the inference process and develop the prototype knowledge base. The tasks involved in developing the GyMEs knowledge base include knowledge acquisition, knowledge representation, knowledge programming, and knowledge refinement. The objective of knowledge acquisition is to obtain facts and rules from the expert that will allow the system to draw expert level conclusions. The process of knowledge acquisition is very time-consuming and difficult, especially if the knowledge engineer is unfamiliar with the domain. In our case, there were several knowledge engineers working on the project and among them were very knowledgeable foresters. For GyMEs, the expert knowledge was originally derived from co-author Twery of the USDA Forest Service (Bennett 1995). The resulting knowledge base was later expanded, analyzed and refined. As a result, additional rules were added and many were updated. After the acquisition of the knowledge begins, a prototype system implementation is begun to test the early stages of the system. This process involves encoding the expert knowledge into the proper format for the computer. Representing and encoding the facts and relationships that constitute knowledge is the next step in the system implementation. There are many established approaches of representing knowledge, for example, semantic networks, rules and logic expressions. "If - Then" style rules are, however, widely used because they are easy to understand and enhance. They facilitate the addition of an explanation facility early in the development process or as an add-on later in the development. GyMEs employs a typical rule- based knowledge representation. During the knowledge programming stage, we first design an overall framework and systematic representation scheme based on the rules derived from the expert. The knowledge is assembled into an organized rule base for the inference engine to interpret and use. The process involves coding facts, rules, objects and relationships found in the domain in the programming language for the system. The following are examples of rules taken directly from the GyMEs knowledge base. risk(very_low) :- Defolp =:= 1. % if the defoliation prediction is very low (i.e., 1), then the % risk is very low risk(very_low) :- Defolp =:= 2, % if the defoliation prediction is low (2) and hazard(very_low). % the hazard rating is very low, then the risk is very low risk(very_low) :- Defolp =:= 2, % if the defoliation prediction is low (2) and hazard(low). % the hazard rating is low, then the risk is very low risk(very_low) :- Defolp =:= 2, % if the defoliation prediction is low (2) and hazard(moderate). % the hazard rating is moderate then the risk is very low risk(very_low) :- Defolp =:= 3, % if the defoliation prediction is moderate (3) and hazard(very_low). % the hazard rating is very low, then the risk is very low In the above rules, the risk assessment is determined to be very low, if any one of the following conditions is satisfied: 1. the defoliation prediction equals to 1, OR 2. the defoliation prediction equals to 2 AND the hazard rating is very low, OR 3. the defoliation prediction equals to 2 AND the hazard rating is low, OR 4. the defoliation prediction equals to 2 AND the hazard rating is moderate, OR 5. the defoliation prediction equals to 3 AND the hazard rating is very low. Notice there are five different sets of conditions that could lead us to conclude the risk assessment is very low. Each condition set is different and is based on some input received from the user and some previous conclusion determined by the system. In this case, the first rule regarding the defoliation prediction requires an input value entered by the user. If the user rates the potential for defoliation as very low, then the system concludes that the estimated risk is likewise very low. On the other hand, all of the other rules are dependent on this user input and the additional condition related to the hazard rating. According to the expert, these value combinations for hazard rating and defoliation prediction lead us to conclude a very low risk assessment. Here, the hazard rating relates to the health of the trees in the stand, the stand management objectives, and the potential for gypsy moth habitation. These parameters are derived from other rules and user input data. 4. GyMEs – the Gypsy Moth Expert System The GyMEs knowledge base includes over 300 facts and rules for deciding the level of gypsy moth risk. The knowledge base is divided into six categories, including stand susceptibility, disturbance history, stand condition, site factors, intervention strategy and defoliation predication. Each category in the knowledge base is further subdivided into 3 to 5 levels and each level is determined by detailed information provided by the user. For example, stand susceptibility is determined by the stand age and species composition of the stand. An organization chart of the knowledge base is presented in Figure 1. Figure 1 illustrates that the user is prompted to input information on stand condition, crown condition, site factors, disturbances and species composition. The knowledge base uses these inputs to determine the severity of each of the six category values and eventually provides a gypsy moth risk assessment for the stand. The rules are compiled to represent the knowledge base for GyMEs. The following are additional examples of the rules (in an English format for easy reading) maintained by the system: IF the stocking percentage is greater than or equal to one percent AND the stocking percentage is less than or equal to eighty percent AND EITHER the crown condition is good OR the crown condition is fair THEN the stand condition is good. In the above rule, the stocking percentage is an index used to estimate the amount of growing space being utilized by trees within the stand. The crown condition is a rating determined using the amount of damaged or dead limbs in the trees’ crowns. These factors give us an indication of the stand’s condition or pre-existing stress level. The user provides the values for these factors as a result of a query that the system prompts the user with. As it turns out, this is a special rule where the English version shown above does not correspond exactly with the actual rule in the knowledge base. If the user satisfies the stocking percentage conditions yet does not know the crown condition (e.g., no data available) then the system concludes that the stand condition is good by default. Of course, the user may backtrack to this particular point in the prompt sequence once information is available on crown condition. Assuming a good or fair rating, the stand condition would be determined to be good. IF the defoliation prediction is greater than or equal to thirty percent AND the defoliation prediction is less than sixty percent AND the hazard rating is moderate THEN the risk is moderate. In this rule, the defoliation prediction value is determined using field data indicating gypsy moth egg mass counts. Based on egg mass counts, stands are categorized according to the level of defoliation expected in the coming year. The hazard rating is based on additional factors including stand vulnerability and management strategy (Bennett 1995). A hazard exists when conditions in the forest indicate damage that is likely to be incurred due to a gypsy moth infestation would impair the forest management objectives. As a web-based application, the GyMEs user interface was originally designed using the hypertext markup language (HTML) and the common gateway interface (CGI) to allow the user to access the system with a web browser such as Microsoft Internet Explorer or Netscape Communicator (Brown and Honeycutt 1998). Generally, the user is also referred to as the web surfer or the client while the computer system being accessed may be referred to as the web site or the server. The interface contains three main parts: a generic form that allows information to be obtained from the user, a mechanism that allows interaction with the inference engine in order to perform various computations on that information, and a generic form that allows results to be presented on the user’s screen. The main reason for using HTML as a front end was that by doing so we could take advantage of the fact that there exist web browsers for virtually every platform you might be interested in using for surfing the web. The screen layout consists of two horizontal frames or windows. The top horizontal window shows the current question being asked by the system. This window also contains a list of possible answers with radio buttons or an input box for the user’s response. The user can move the cursor to the appropriate answer and mouse click the "Submit" button located in the lower part of the input box. After the user selects or types in an answer and submits it, the system stores the information in a file and displays it on the second window. The risk assessment will be presented whenever the system receives enough information to reach an assessment conclusion. As mentioned earlier, GyMEs employs a backward chaining inference engine. Therefore, the system is given a specific goal or conclusion to determine based on the order of the rules in the knowledge base and then proceeds to search the knowledge base for rules that have that goal as its conclusion. Once the inference engine finds an appropriate rule, it recursively searches through the knowledge base to determine if it can satisfy the premise (conditions) of that rule. If this is not possible, GyMEs looks for another rule with that original goal as its conclusion (backtracks), and so on until a conclusion is found. GyMEs uses a prolog inference engine located on the server (i.e., the machine at our web site) to manipulate the assessment knowledge base. Since each interaction with the server from a client (the web surfer) is an independent event, the server also maintains a client specific file of known facts and derived conclusions for the current decision making session. After each interaction, the server consults this file before invoking the inference engine. Armed with the facts and derived results in the file, and the user’s latest input, the inference engine processes the rules in the knowledge base to determine the next conclusion. If this is a risk assessment, then the user is notified. If not, the server results file is updated and the prompt for the next piece of required information is determined and sent to the client’s browser. Prior to implementing this client specific file scheme, only one client at a time was allowed to access GyMEs. However, we devised a scheme where any number of clients could surf into the GyMEs web site and run the system. The scheme is based on the server keeping track of each client’s session file, and being able to relate these session files to the proper client. Before fully implementing our own session file scheme for tracking individual client interactions, we converted our prolog system to LPA Prolog with the Pro-Web feature (Shalfield 1998). Pro-Web handles all the client/server interactions automatically, making the GyMEs web site truly multi-user in an almost transparent fashion. That is, using Pro-Web gives us full multi- user capability so that any number of surfers may surf into the system. However, in order for Pro-Web to handle the problem of relating clients to sessions, there are some restrictions on how the web interface is developed. To the client, this is totally invisible, which is exactly what we wanted. 5. Future directions After compilation of the rules and making any necessary additions/changes, we partially verified the knowledge base to help ensure its accuracy. Verification is a necessary step to ensure that there are, for example, no dead-end lines of reasoning embedded within the knowledge base that would result in a “risk assessment unknown” conclusion derived via the inference process. There must also be a cross check of the results against the conclusions provided by the domain expert, if possible. For GyMEs, this step is currently in progress. One of the first issues for future research is the continued validation and expansion of the GyMEs knowledge base. One type of validation needed is criterion-related validity to determine the relationship between the expert system outcome and the advice from the domain expert. Another issue is the reasoning process used by the GyMEs inference engine. After a user enters information, the system should provide a way to track the logical reasoning process, and then output the information if the user requests it. Another issue to be addressed in the near future is the inclusion of an explanation facility. When a user is prompted for some information about the stand, it would be beneficial for them to have the ability to ask the system why it needs that information. Also, when an intermediate conclusion or risk assessment is derived, users may want to know how that conclusion was derived. This type of facility increases the utility of an expert system. A further goal is to develop GyMEs as a module that can be incorporated within the GypsES software system. This phase of development will soon be completed. However, our secondary goal is to enhance the web-based version. Our proposed approach is based on using java servlets (Moss 1998) to handle the server interaction to the inference engine. Servlets are applications running on the server to manage client requests. This approach will allow us more flexibility than currently provided by Pro-Web and will improve the server’s ability to handle client requests. References Bennett, D.B. (1995) Modeling a decision process: hazard and risk assessment for areas threatened by gypsy moth infestation. Master’s Thesis. West Virginia University. Morgantown, WV. Benders, J. and F. Manders (1993) Expert systems and organizational decision-making. Information and Management. 25: 207-231. Bobrow, D.G., S. Mittal and M. Stefik (1986) Expert systems: perils and promises. Communication of the ACM. 29: 880-894. Brown, M.R. and J. Honeycutt (1998) Using HTML 4, Fourth Edition. Que Corporation, Indianapolis, IN. Doane, C.C. and M.L. McManus, eds. (1981) The gypsy moth: research toward integrated pest management. USDA Forest Service Technical Bulletin 1584. Elkinton, J.S. and A.M Liebhold (1990) Population dynamics of gypsy moth in North America. Ann. Rev. Entomol. 35: 571-596. Gerardi, M.H. and J.K. Grimm (1979) The history, biology, damage, and control of the gypsy moth. Associated University Press, Inc. Cranbury, New Jersey. Gottschalk, K.W. and M.J. Twery (1989) Gypsy moth impacts in pine-hardwood mixtures. pp. 50-58, Proc. Pine-Hardwood Mixtures: A Symposium on Management and Ecology of the Type. 1989 April 18-19; General Technical Report. SE-58. USDA Forest Service, Southeastern Forest Experiment Station. Holsapple, C.W. and A.B. Winston (1996) Decision Support Systems – A Knowledge-Based Approach. West Publishing Company. New York. Liebhold, A., V. Mastro, and P.W. Schaefer (1989) Learning from the legacy of Leopold Trouvelot. Bulletin of the Entomological Society of America. pp. 20-22. Montgomery, M.E. and W.E. Wallner (1988) The gypsy moth – a westward migrant. p. 353-375. In: Berryman, A.A. (ed.) Dynamics of Forest Insect Populations. Plenum Publishing Corporation. Moss, K. (1998) JAVA Servlets, McGraw Hill Publishing, New York. Shalfield, R. (1998) LPA Pro-Web 1.0 User Guide. Logic Programming Associates, Ltd. London. England. Stefik, M. (1995) Introduction to Knowledge Systems. Morgan Kaufmann, California. Thomas, S.J., S.L.C. Fosbroke, and A.B. Cumming (1998) GypsES: Decision Support and Project Management – User’s Guide, Version 1.0. USDA Forest Service. Northeastern Forest Experiment Station. NA-TP-01-98. Turban, E. (1993) Decision Support and Expert Systems: Management Support Systems. 3rd ed., Macmillan, New York. Twery, M.J. (1991) Effects of defoliation by gypsy moth. In: Gottschalk, K.W.; Twery, M.J.; Smith, S.I., eds.; Proceedings USDA InterAgency Gypsy Moth Research Review; January 22-25, 1990; East Windsor, CT. Northeastern Forest Experiment Station, General Technical Report NE-146. Twery, M.J. and G.A. Elmes (1991) Hazard rating for gypsy moth on a macintosh computer: a component of the GypsES system. In: Gottschalk, K.W.; Twery, M.J.; Smith, S.I., eds.; Proceedings USDA InterAgency Gypsy Moth Research Review; January 22-25, 1990; East Windsor, CT. Northeastern Forest Experiment Station, General Technical Report NE-146. Figure 1. Flow chart of knowledge base Defoliation history during the most recent five years Drought history during the most recent three years Other nature disasters, such as fire, floods, ice storms Treatment history during the most recent five years Disturbance History Crown condition Stocking Soil type Management objectives Defoliation Prediction RISK ASSESSMENT Stand Condition Site Factors Intervention Strategy Stand Vulnerability Hazard Rating Stand age Species susceptibility Stand Susceptibility 
work_jy2fwvcsnbgfvk3ot7xdtsshhq	----	Microsoft Word - 6-1-2005.doc Bioinformation by Biomedical Informatics Publishing Group open access www.bioinformation.net Views & Proposal _______________________________________________________________________________________________________ ISSN 0973-2063 Bioinformation 1(1): 14-15 (2005) Bioinformation, an open access forum © 2005 Biomedical Informatics Publishing Group 14 A rapid identification system for metallothionein proteins using expert system Bhoopathi Praveen1*, Savariar Vincent3, Upadhyayula Suryanarayana Murty1, Amirapu Radha Krishna2 and Kaiser Jamil1 1Biology division, Indian Institute of Chemical Technology, Hyderabad-500 007, India; 2Computer division, Indian Institute of Chemical Technology, Hyderabad-500 007, India; 3Department of Zoology, Loyola College, Chennai - 600 034, India; Bhoopathi Praveen* - Email: bhoopathip@gmail.com; * Corresponding Author received April 15, 2005; revised April 16, 2005; accepted April 17, 2005; published online April 21, 2005 Abstract: Metallothioneins (MT) are low molecular weight proteins mostly rich in cysteine residues with high metal content. Generally, MT proteins are responsible for regulating the intracellular supply of biologically essential metal ions and they protect cells from the deleterious effects of non-essential polarizable transition and post-transition metal ions. Due to their biological importance, proper characterization of MT is necessary. Here we describe a computer program (ID3 algorithm, a part of Artificial Intelligence) developed using available data for the rapid identification of MT. Tissue samples contains several low molecular weight proteins with different physical, chemical and biological characteristics. The described software solution proposes to categorize MT proteins without aromatic amino acids and high metal content. The proposed solution can be expanded to other types of proteins with specific known characteristics. Keywords: metallothionein; rules; artificial intelligence; expert system; isolation; detection; tissue samples; purification Back ground: Metallothioneins (MT) constitute a super family of low molecular weight, cysteine-rich metalloproteins and metallopeptides responsible for regulating the intracellular supply of biologically essential zinc and copper ions. They have role in protecting cells from the deleterious effects of high concentration metal ions. [1] MT is now known to occur in all animal phyla examined so far as well as in certain fungi, plants and cyanobacteria. The function of MT is to detoxify non-essential metals such as mercury, cadmium and essential metals such as zinc and copper. MT is generally induced during stress conditions. [2] Hence, identification of MT from tissues/animal models is very important during analysis. Here, we describe a method for the identification of metallothionein proteins using expert system. An expert system is a set of computer programs that mimic a human expert. [3] Expert system has been used for the identification of proteins using mass spectrometry in peptide mapping by ProFound. [4] It has also been used for detecting the protein domain structure in sequence [5] and for the prediction of protein localization sites in Gram-negative bacteria. [6] Classification and characterization of proteins are generally laborious dealing with large datasets and experimentations. [4] Here, we describe a framework using expert system developed from a set of physical and chemical properties of known proteins, especially MT. This procedure utilizes knowledge organized as "IF-THEN" rules for metallothionein protein. This methodology finds application in protein recovery, isolation and detection of proteins. Methodology: The physical and chemical characteristics of MT proteins were used in the development of a rapid identification system in silico. These characteristics were derived based on the experiments conducted using animal models (wistar rats). Different groups of animals were exposed to different concentrations of heavy metals like Cr(VI), Cd, Zn and Hg. Tissue samples containing MT were isolated from various organs after a series of centrifugation steps with alcohol extraction and purified by Ion exchange followed by Gel filtration chromatography. Purified MT was further subjected for the estimation of metal content, molecular weight and non-aromatic amino acid content using atomic absorption spectro-photometry, SDS-PAGE and Bradford assay, respectively. The derived characteristics from experimental results were divided into a set of rules. These rules include, (1) protein with low molecular weight / high molecular weight, (2) proteins with metal content / without metal content, (3) aromatic amino acids / without aromatic amino acids and (4) sulphur content / without sulphur content. The derived rules at every step were trained in an expert system (ID3 algorithm) as described elsewhere. [7-9] A minimum number of rules were selected (each rule is a series of conditions consisting of attributes and value pairs, followed by a single conclusion that contains the class and the corresponding class value) using human expertise to maximize true positive identification. The rules were then formulated into an IF – THEN – ELSE algorithm and translated into VISUAL BASIC language statements. Thus a software solution is proposed for the identification of MT during protein recovery. The methodology for the identification of MT is illustrated in Figure 1. Bioinformation by Biomedical Informatics Publishing Group open access www.bioinformation.net Views & Proposal _______________________________________________________________________________________________________ ISSN 0973-2063 Bioinformation 1(1): 14-15 (2005) Bioinformation, an open access forum © 2005 Biomedical Informatics Publishing Group 15 Figure 1: Identification of MT proteins. The rules are given in boxes Limitation: Expert systems are useful tools for the identification of MT during protein characterization. Here, we propose the utilization of expert system for the identification of MT from a mixture of unknown proteins during a series of purification steps. A pilot development shows that the procedure is successful in identifying proteins with high specificity towards low or very low molecular weight components. The major limitation in the use of expert system is the need to train specific systems using specific sets of rules. This requires domain expert regarding the protein of interest. In other words, this procedure cannot help in the identification of unknown proteins other than metallothionein, unless the methodology is modified for other proteins. Acknowledgement: The authors are grateful to Director, IICT for providing financial assistance. References: [1] F. Silvestre, et al., Comp Biochem Physiol C Toxicol Pharmacol., 140:39 (2005) [PMID: 15792621] [2] P. Dziegiel, et al., Rocz Akad Med Bialymst., 49:43 (2004) [PMID: 15638370] [3] D. L. Rotin, et al., Arkh Patol., 66:47 (2004) [PMID: 15154385] [4] W. Zhang & B.T. Chait, Anal Chem., 72:2482 (2000) [PMID: 10857624] [5] E. Benatan, Proc Int Conf Intell Syst Mol Biol., 2:37 (1994) [PMID: 7584415] [6] K. Nakai & M. Kanehisa. Proteins, 11:95 (1991) [PMID: 1946347] [7] U. S. N. Murty, et al., Comput Appl Biosci., 12:491 (1996) [PMID: 9021267] [8] M. Edwards, et al., Comput Appl Biosci., 3:1 (1987) [PMID: 3330955] [9] R. Hofestadt, Medinfo., 2:964 (1995) [PMID: 8591601] Edited by P. Kangueane Citation: Praveen et al., Bioinformation 1(1): 14-15 (2005) License statement: This is an open-access article, which permits unrestricted use, distribution, and reproduction in any medium, for non-commercial purposes, provided the original author and source are credited. 
work_jy6w7cqljzdzjc6p3l5pcxtmnm	----	Thyroid Disease Diagnosis using Ontology based Expert System Thyroid Disease Diagnosis using Ontology based Expert System Vipula Rawte Student, M.E. (Computer Engineering) St. Francis Institute of Technology, Mumbai-400103, India. Bidisha Roy Associate Professor (Computer Engineering) St. Francis Institute of Technology, Mumbai-400103, India. Abstract—: This study shows a new method in building up an ontology based expert system to diagnose thyroid diseases. Ontology describes some domain in the form of concepts and relationships among them. It has an advantage of being computer and human readable. The acceptance of ontology in medical field has facilitated the domain experts and non-experts to execute the job of knowledge representation easily. This paper presents an ontology depicting the domain of thyroid diseases and its related symptoms. The thyroid gland is an important organ because it is responsible for metabolism in the body. Diseases related to thyroid gland can be difficult to diagnose because symptoms are confused with other possible conditions easily. This study proposes to build an expert system based on ontology for diagnosing thyroid diseases. It uses ontology to construct the domain knowledge and rules to infer the thyroid disease related diagnosis. Keywords—Thyroid; expert system; ontology I. INTRODUCTION The thyroid is a butterfly-shaped gland that lies wrapped around the windpipe and is positioned under the Adam's apple. The job of the thyroid gland is to secrete thyroid hormones that help regulate metabolism. The thyroid gland produces two necessary hormones–thyroxine (T4) and triiodothyronine (T3). Normally, most T3 in the blood stream gets converted from T4. These two similar hormones are called together thyroid hormone. When the degree of thyroid hormone is too small, the pituitary gland secretes a hormone called thyroid stimulating hormone (TSH) which stimulates the thyroid to produce more thyroid hormone. When the level of thyroid hormone goes back to normal, the pituitary gland ceases producing extra TSH. If thyroid hormone is present in excess, the pituitary gland makes less TSH and thus less production of thyroid hormone by the thyroid gland. The pituitary gland is stimulated by the hypothalamus to produce TSH by secreting thyrotropin releasing hormone (TRH). This entire system is called the hypothalamic-pituitary-thyroid axis because of the feedback loop between the pituitary gland and the thyroid gland which maintains normal levels of T4, T3 and TSH [1]. When thyroid gland fails to maintain the normal levels of TSH there arises mainly two types of thyroid related diseases called hyperthyroidism and hypothyroidism. First, hyperthyroidism, also known as over active thyroid disease, is the condition that occurs due to excessive production of thyroid hormone by the thyroid gland. Second, hypothyroidism, also called underactive thyroid disease, is a disorder in which the thyroid gland does not produce enough thyroid hormone [2]. Fig. 1.a. Thyroid Gland [1] Fig. 1.b. Hypothalamic-Pituitary-Thyroid axis [1] Depending only on one blood test, TSH, leaves millions of people undiagnosed and their symptoms are often confused with other conditions such as weakness, fatigue, etc. So, there is a necessity of an expert system which can consider the symptoms in addition to TSH levels and make the diagnosis respectively. An expert system is artificial intelligence based interactive computer application that performs a task that would otherwise be performed by a human expert. Expert systems typically consist of three parts: (1) a knowledge base International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS060415 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 06, June-2015 365 which contains the information acquired by questioning experts, and logic rules that control how data is applied; (2) an inference engine that understands the submitted problem against the rules and logic of information stored in the knowledge base and infers accordingly; and an (3) an interface that allows the user to show the problem in a human language such as English. Benefits of developing an Expert System:  Expert system is available any time.  The cost of offering expertise is cheaper.  The knowledge will last for an indefinite time.  Expert system can be made to have knowledge from multiple experts.  It is capable of developing the reasoning that directed to a conclusion.  It can respond fast because of the underlying benefits of computers over humans.  It is unemotional and responds at all times and works steadily during urgent situations. An Expert System for Thyroid Diagnosis is developed with the purpose of assisting the physician in diagnosing thyroid disease. Ontology is a body of knowledge describing some domain in form of concepts and relationships between them. Need for using ontology arises because of some drawbacks of Machine Learning Techniques such as requirements for training data, high computation for model learning, adapting with dynamic nature of data. Humans are capable of using the Web to carry out tasks. However, a computer cannot accomplish the same tasks without human direction because web pages are designed to be read by people and not machines. The semantic web is understandable by computers and so that they can do more of the complex work involved in searching, sharing, and merging information on the web. Benefits of using Ontology [3]:  Interaction among systems, among humans, and between humans and systems.  Computational inference.  Reuse and organization of knowledge. Combining the three concepts discussed above, i.e. Expert System (ES), ontology and thyroid disease into an expert system based on ontology shall diagnose thyroid disease. In this paper, an expert system based on ontology for diagnosis of thyroid diseases that uses the above methods is presented. The draft of this paper is as follows. Section II gives the literature survey. Section III shows the proposed approach. Section IV shows the results and analysis for the above system. Finally, Section V concludes along with the work which can be done afterward in future. II. LITERATURE SURVEY The occurrence of Hypothyroidism is nearly about 1 in 2,640 where almost 40,000 newborn babies were screened, which is much higher than the worldwide statistic of 1 in 3,800. Diagnostic time lag in hypothyroidism is because of lack of awareness among primary health care practitioners, family physicians and the cost, availability of laboratory investigations [4]. The study presents that Hypothyroidism among women was extremely high in this region [5]. 4543 children in the age group of 6-18 years from different schools in Delhi were screened during the research study for thyroid function and it was found that goiter is prevalent in children (17%) even after two decades of iodization [6]. Females above 35 years living in a randomly selected urban area Elamakkara in Cochin were surveyed for thyroid disease. This survey clearly showed that undetected thyroid disorders are very high in this community, demanding development of suitable screening strategies to detect and treat these conditions considering the level of health problems it can cause to the population [7]. According to various research studies on thyroid disease, it has been estimated that about 42 million people in India suffer from thyroid diseases [8]. [9] suggests that the UK is iodine-deficient, indicating need for examination of the UK iodine status and recommendations to safeguard public health. The subject population with 4409 adult members of resident welfare associations of 5 residential areas, from 18- 90 years of age participated in general health check-up camps. 30.6% of subjects with severe hypoechogenicity were diagnosed and this percentage was found higher in women [10]. The study was conducted in eight major cities – Mumbai, Bangalore, Delhi, Ahmedabad, Chennai, Kolkata, Hyderabad and Goa. "The prevalence of hypothyroidism was more, impacting roughly one in 10 adults. Females and old people were found to have substantial association with hypothyroidism," said the study [11]. The hospital based survey involved 4739 patients having undergone thyroid function test, in the central clinical biochemistry laboratory of Subharti Medical College, UP and its associated hospital. More prevalence of Hypothyroidism was observed in patients (especially females). The findings confirm the usefulness of screening of thyroid function mandatory for early detection and treatment to reduce the effects of thyroid dysfunctions [12]. Out of the 325 districts surveyed in India so far, 263 are IDD-endemic and iodine deficiency causes goiter [13]. More lately, estimates show that 27 million Americans have thyroid disease out of which about 13 million people go undiagnosed [14]. A correlation is present between iodine intake and thyroid diseases; high iodine intake may lead to the occurrence of thyroid diseases such as Hashimoto‟s thyroiditis, nodular goiter and Hyperthyroidism through a long-term mechanism. Females with high iodine intake had a close relationship with thyroid diseases than males [15]. International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS060415 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 06, June-2015 366 Gist of this literature survey states that the thyroid disease is one of the most prevalent diseases worldwide. Edward Feigenbaum [16] defined an Expert System (ES) as: “An expert system is an intelligent computer program that uses knowledge and inference procedures to solve problems that are difficult enough to require significant human expertise for the solution.” Medical diagnosis decision making becomes a very tedious task because the human experts make decisions and they can barely process the immense amounts of data. So an automated tool is required that should help them make an effective and right decision. An artificial neural network has the ability to imitate decision-making process, and use a knowledge base, and a training set, to learn to diagnose diseases. A medical decision support system based on the neural network is trained by applying an improved Back Propagation algorithm for Medical Diagnosis. The hidden layer of a neural network plays an important role for detecting the relevant features [17]. Case-based Reasoning (CBR) methodology provides a way to solving problems with recalled knowledge of solved cases. This paper gives a system model that employs the simplified medical knowledge: disease-symptom ontology for pre-diagnosis, given patient's symptoms as the input [18]. A new hybrid case-based reasoning approach for medical diagnosis is proposed to improve the accuracy of the CBR systems. The approach includes case-based reasoning and rule-based reasoning, and also carries out the adaptation process automatically by using adaptation rules. The rules viz., adaptation rules and reasoning rules are produced from the case-base. On solving a new case, the case-base is refined, and both adaptation and reasoning rules are updated [19]. A comparative thyroid disease diagnosis by using multilayer, probabilistic, and learning vector quantization neural networks is given in [20]. A three-stage expert system is proposed to diagnose thyroid disease. This system uses feature selection (FS), particle swarm optimization (PSO) and support vector machines (SVM). The PSO-based system costs a lot of computational time [21]. Linguistic Hedges Neural-Fuzzy Classifier with Selected Features (LHNFCSF) is given for diagnosis of thyroid diseases. The work to be done in the future includes improving the performance of NFCs using high-performance computing methods [22]. III. PROPOSED APPROACH Considering the drawbacks of the machine learning techniques, this paper proposes to diagnose thyroid related problems with the assistance of an expert system based on ontology. This is shown in Fig.2. Fig.2 . Block Diagram: OBESTDD [23] It consists of six main blocks. User is a person who wants to know diagnosis of thyroid disease. User uses Web application to interact with the computer. Domain Ontology represents concepts which belong to the real world. A reasoner is a tool which has the capability to infer logical consequences from a set of facts. Rule Base is the collection of rules which are used to infer diagnosis from the domain ontology. Database contains the patients records based on which the expert system can make the diagnosis. This can be MySQL database. Ontology for thyroid disease is created using an ontology editor called Protégé. The ontology acts as the knowledge base for this expert system. Rules are written in Semantic Web Rule Language (SWRL). Jena is used an inference engine to deduce diagnosis based on the knowledge and SWRL rules. MySQL database is used a backend to store users‟ details. Front end is a web application using JSP and acts as an interface through which user can enter the symptoms to know their diagnosis. Fig. 3.a. shows the Thyroid Ontology designed using ontology editor called Protégé. Fig 3.b. shows Thyroid Ontology Classes. Fig. 3.c. shows Thyroid Symptoms and Fig. 3.d. shows Thyroid Ontology Object Properties. Fig. 3.a. Thyroid Ontology There are 4 main classes, viz. Age, Person, Symptoms and Diagnosis and 9 subclasses, viz. Child, Teen, Adult, Male, Female, HyperSymptoms, HypoSymptoms, Hyperthyroidism, and Hypothyroidism in the thyroid ontology shown in Fig. 3.a. Symptoms are instances of the classes which are shown in Fig. 3.c. For example, Heat Intolerance is the symptom (instance) of class HyperSymtoms. International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS060415 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 06, June-2015 367 Fig. 3.b. Thyroid Ontology Classes Fig. 3.c. Thyroid Symptoms Fig. 3.d. Thyroid Ontology Properties Consider an example where a user named Rucha selects the age & gender and related symptoms. Fig. 4.a. User enters name and selects age & gender Fig. 4.b. User enters name and selects symptoms Fig. 4.c. Ontology based Expert System shows result Fig. 4.d. Data (in turtle format) for a user Fig. 4.e. Rule (in SWRL format) for a user Fig. 3.a. shows user selects gender “Female” and age “more than 20 years”. Fig. 3.b. shows that Rucha selects symptoms (1) Insomnia, (2) Heat Intolerance and More Sweating, (3) Feeling panicked and nervous, Increased Heart Rate and Hand Tremor, (4) Lighter Menstrual Periods or Missed Periods, (5) High libido(high estrogen level). Fig. 3.c. gives the diagnosis. Fig. 3.d. and Fig. 3.e. show user data in International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS060415 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 06, June-2015 368 turtle format and rules in SWRL format. Based on the data and rule, Jena reasoned draws inference and gives the diagnosis. Fig.5. shows the flowchart for the mechanism of the expert system based on ontology. User has to first register and then login if already registered. Then user should select „Age & Gender‟ and select the related symptoms accordingly. Finally, the result is given. Fig. 5. Flowchart for the working of the expert system based on ontology. Consider the following example: Data given is: Joseph p:isFatherof James John p:isBrotherof Joseph Rule given is: rule: (?f p:isFatherof ?a) (?u p:isbrotherof ?f) -> (?u p:isUncleof ?a) --------------------------------------------------------------------- ----- - John isUncleof James Similarly, Data for a user is: :RUCHA p:gender :female. :RUCHA p:age :adult. :adult p:hasHyperSymp :Insomnia. :adult p:hasHyperSymp :Heat_Intolerance. :adult p:hasHyperSymp :Feeling_panicked_and_nervous. :adult p:hasHyperSymp :Lighter_Menstrual_Periods_or_Missed_Periods. :adult p:hasHyperSymp :High_libido. :Insomnia p:hasHyperDiagnosis :HYPERTHYROIDISM. :Heat_Intolerance p:hasHyperDiagnosis :HYPERTHYROIDISM. :Feeling_panicked_and_nervous p:hasHyperDiagnosis :HYPERTHYROIDISM. :Lighter_Menstrual_Periods_or_Missed_Periods p:hasHyperDiagnosis :HYPERTHYROIDISM. :High_libido p:hasHyperDiagnosis :HYPERTHYROIDISM. Rule for this user is: rule:(?w p:gender ?x)(?w p:age ?y)(?y p:hasHyperSymp ?0)(?y p:hasHyperSymp ?1)(?y p:hasHyperSymp ?2)(?y p:hasHyperSymp ?3)(?y p:hasHyperSymp ?4)(?0 p:hasHyperDiagnosis ?z)(?1 p:hasHyperDiagnosis ?z)(?2 p:hasHyperDiagnosis ?z)(?3 p:hasHyperDiagnosis ?z)(?4 p:hasHyperDiagnosis ?z)->(?w p:HAS ?z)] -------------------------------------------------------------------------- -- RUCHA HAS HYPERTHYROIDISM where, 1. w is an instance of class Person. 2. x is an instance of class Female. 3. y is an instance of class Age. 4. 0,1,2,3,4 are instances of class HyperSymptoms 5. z is an instance of class Hyperthyroidism Likewise, based on the symptoms entered through the interface, data and rules get generated for the users. Then, a reasoned called Jena uses these data and rules to infer and make the diagnosis accordingly. IV. RESULTS AND ANALYSIS Implementation of the expert system for thyroid disease diagnosis based on neural network and ontology gave the following results. Total number of instances taken is 60. Neural network considers only TSH, T3, T4 levels in the blood to train the neural network whereas there is no training for the ontology based expert system. It makes diagnosis based on symptoms. TABLE I. NEURAL NETWORK Class Predicted Right Number of Instances Total Number of Instances Accuracy in % Normal 31 34 91.176 Hyper 1 11 9.091 Hypo 5 15 33.333 Male 16 23 69.565 Female 21 37 56.757 Child 4 6 66.667 Teen 8 11 72.727 Adult 27 43 62.791 Start Enter user login details Registered? Registration Form Enter user login details Enter name and select Age & Gender Select symptoms and submit Stop International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS060415 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 06, June-2015 369 TABLE II. ONTOLOGY BASED Class Predicted Right Number of Instances Total Number of Instances Accuracy in % Normal 32 34 94.118 Hyper 9 11 81.818 Hypo 13 15 86.667 Male 20 23 86.957 Female 35 37 94.595 Child 5 6 83.333 Teen 9 11 81.818 Adult 40 43 93.023 Fig. 6. Comparison of Neural Network and Ontology Based Expert System The analysis of the results presented in Fig.6 shows that expert system based on ontology gives more accurate results with lesser complexity than the one which uses neural network. V. CONCLUSION Thyroid disease diagnosis is difficult and often confused with symptoms of other diseases. If diagnosis is prognosticated, then by advising proper prevention methods and dosages to the patients, they can be recovered in the beginning. Thus, expert system based on ontology for diagnosis of thyroid disease has been proposed and implemented. Validation is done based on two parameters:  Accuracy  Time Complexity Thus, expert system based on ontology gives better results with lesser complexity. REFERENCES [1] “Thyroid disorders”, http://www.abc.net.au/health/library/thyroid_ff.htm. June 06, 2015 [June 16, 2005]. [2] James Norman, “Your Thyroid Gland,” http://www.endocrineweb.com/conditions/thyroid/your-thyroid-gland. [3] Bürger, Tobias, and Elena Simperl. "Measuring the benefits of ontologies." On the Move to Meaningful Internet Systems: OTM 2008 Workshops. Springer Berlin Heidelberg, 2008. [4] Meena P Desai, “Disorders of Thyroid Gland in India”, Symposium:Endocrinology-Part I Indian J Pediatr 1997;64: 11-20. [5] Rebecca Abraham, V Srinivasa Murugan, P Pukazhvanthen and S K Sen, “Thyroid Disorders In Women Of Puducherry”, Indian Journal of Clinical Biochemistry, 2009 / 24 (1) 52-59. [6] V.K. Chauhan, R.K. Manchanda, Archana Narang, R.K.Marwaha, Saurav Arora, Latika Nagpal, “Prevalence of Thyroid Disorders in School Children In Delhi – Post Iodization Scenario”, Indian Journal of Research in Homoeopathy Vol. 3, No. 3, July-September 2009. [7] DR V USHA MENON, “Spectrum Of Thyroid Disorders, Iodine Status And Diabetes Mellitus In Adult Kerala Population – A Community Study”. THESIS SYNOPSIS, Amrita Viswavidyapeetham University, 2011. [8] Ambika Gopalakrishnan Unnikrishnan and Usha V. Menon, “Thyroid disorders in India: An epidemiological perspective”, Indian J Endocrinol Metab. Jul 2011; 15(Suppl2): S78–S81. [9] “BTA Research Team: UK is now Iodine Deficient” Internet: http://www.btf-thyroid.org/index.php/thyroid/research-news/ukis- iodine-deficient. Sep 05, 2014 [April 15, 2011]. [10] Raman Kumar Marwaha, Nikhil Tandon, Mohd Ashraf Ganie, Ratnesh Kanwar,Aparna Sastry MK Garg, Kuntal Bhadra, Satveer Singh, “Status of Thyroid Function in Indian Adults: Two Decades After Universal Salt Iodization”, JAPI April 2012,VOL. 60. [11] Neetu Chandra, “India's cities in the grip of thyroid disease as new study reveals one in ten suffer from disorders”. Internet: http://www.dailymail.co.uk/indiahome/indianews/article2408114/India s-cities-grip-thyriod-problems-new-study-reveals-sufferdisorders.html Sep 05, 2011. [12] Naved Ahmad, Meenakshi Panthari, Akash Gupta, Prasnna Chandra, Sana Nafees, “Prevalence Of Hypothyroidism Among Patients Of Meerut, Uttar Pradesh: A Hospital Based Study,” Research Article, March,2013. [13] Chandrakant S. Pandav, Kapil Yadav, Rahul Srivastava, Rijuta Pandav* & M.G. Karmarkar, “Iodine deficiency disorders (IDD) control in India,” Indian J Med Res 138, September 2013, pp 418-433. [14] Mary Shomon, “Why Are So Many People Getting Thyroid Disease?,” http://thyroid.about.com/od/symptomsrisks/a/Why-Are-So-Many- People-Getting-Thyroid-Disease.htm Sep 05,2014 [Aug 18, 2014]. [15] Hengqiang Zhao & Yuan Tian & Zeming Liu & Xiaoyu Li & Mengyu Feng & Tao Huang, “Correlation Between Iodine Intake and Thyroid Disorders: A Cross-Sectional Study from the South of China,” Biological Trace Element Research, August 2014. [16] A. Barr and E.A. Feigenbaum, The Handbook of Artificial Intelligence, Vol. 1, William Kaufmann, 1981. [17] Ashish Dehariya, Vijay Choudhari, Ilyas Khan and Sourabh Karsoliya, “An Effective Approach for Medical Diagnosis Preceded by Artificial Neural Network Ensemble,” Proceedings of the 5th National Conference; INDIACom, 2011. [18] Hsien-Tseng Wang, Abdullah Uz Tansel, “Composite Ontology- Based Medical Diagnosis Decision Support System Framework,” Communications of the IIMA, Volume 13 Issue 2, 2013. [19] Dina A. Sharaf-El-Deen & Ibrahim F. Moawad & M. E. Khalifa, “A New Hybrid Case-Based Reasoning Approach for Medical Diagnosis Systems,” Springer Science+Business Media New York, 2014 [20] Feyzullah Temurtas, “A comparative study on thyroid disease diagnosis using neural networks,” Expert Systems with Applications 36 (2009) 944–949. [21] Hui-Ling Chen & Bo Yang & Gang Wang & Jie Liu & Yi-Dong Chen & Da-You Liu, “A Three-Stage Expert System Based on Support Vector Machines for Thyroid Disease Diagnosis”, J Med Syst (2012) 36:1953–1963. [22] Ahmad Taher Azar, Aboul Ella Hassanien and Tai-hoon Kim, “Expert System Based on Neural-Fuzzy Rules for Thyroid Diseases Diagnosis,” Communications in Computer and Information Science, 2012. [23] Vipula Rawte and Bidisha Roy, "OBESTDD: Ontology Based Expert System for Thyroid Disease Diagnosis." International Conference on Nascent Technologies in the Engineering Field (ICNTE), IEEE, 2015. International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS060415 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 06, June-2015 370 
work_jzz5hvqpavfevdzghtewyapvni	----	Kinship Algebra Expert System (KAES): A Software Implementation of a Cultural Theory UCLA UCLA Previously Published Works Title Kinship Algebra Expert System (KAES): A Software Implementation of a Cultural Theory Permalink https://escholarship.org/uc/item/9dz8q3sr Journal Social Science Computer Review, 24(1) Author Read, Dwight W Publication Date 2006 DOI 10.1177/0894439305282372 Peer reviewed eScholarship.org Powered by the California Digital Library University of California https://escholarship.org/uc/item/9dz8q3sr https://escholarship.org http://www.cdlib.org/ 10.1177/0894439305282372Social Science Computer ReviewRead / Kinship Algebra Expert System Kinship Algebra Expert System (KAES) A Software Implementation of a Cultural Theory Dwight W. Read University of California, Los Angeles The computer program Kinship Algebra Expert System (KAES) provides a graphically based framework for constructing, if possible, a generative algebraic model for the structure of a kin- ship terminology (the terms used to refer to one’s kin). The algebraic modeling is based on a the- ory of kinship terminologies elaborated through writing the software program. The theory relates the properties and structure of kinship terminologies to an underlying logic that the KAES program helps uncover and model as a generative structure. The program then relates the structural logic of a kinship terminology modeled by the KAES program to a genealogical space based on genealogical tracing of kin relations. The KAES program demonstrates the surpris- ingly logical character of kinship terminologies and challenges the received view of the primacy of genealogical relations in defining cultural kinship through showing how genealogical definitions of kin terms can be accurately predicted in the terminologies considered to date. Keywords: kinship; cultural theory; terminology; algebraic modeling The computer program, Kinship Algebra Expert System (KAES), is a graphical user inter-face application for investigating and modeling the structure of kinship terminologies and the ways in which these are instantiated (for further information and for downloading a copy of the application program, see Read & Fischer, 2005). It is based on a theory of kinship terminology structures elaborated through writing the KAES program. The program is an example of how a theory can be approached and evidence evaluated with respect to that the- ory to produce results that identify a future research trajectory rather than simply accounting for relations in a body of data. The results of research based on KAES will have far-reaching consequences for many other domains in the social sciences. The focus on kinship terminologies may strike many readers as a rather specialized and narrow application, especially following more than a decade in which the study of kinship has diminished in the curricula of major anthropology programs, only to be replaced by courses on gender and new reproductive technologies. However, KAES successfully imple- ments a theory of kinship terminology structures that is fundamental to our understanding of the kinship basis through which a newborn person initially obtains social identity crucial throughout his or her life. In addition, by providing a fully worked out theory and model of one culturally constructed symbolic system (symbolic in the linguistic or mathematical and 1 Social Science Computer Review Volume 24 Number 1 February 2006 1-25 © 2006 Sage Publications 10.1177/0894439305282372 http://ssc.sagepub.com hosted at http://online.sagepub.com not the interpretive sense), we gain insights for the critical task of understanding the relation- ship between culturally constructed symbolic systems and their cultural instantiation. The program owes its genesis in the 1980s to a problem with communicating the results of algebraic representation of the structure of kinship terminologies to an anthropological audi- ence with little or no mathematical background, especially in the area of algebraic structures. I had just published a paper (Read, 1984) showing how the American kinship terminology (AKT) could be structurally described in the form of what has been called a kin term map (Leaf, 1971) and, more importantly, how the resulting structure could be generated in the form of an algebraic structure known as a semigroup. It became evident that the number of readers in anthropology who could follow the algebraic argument was very limited, and one of my (then) graduate students, Dr. Clifford Behrens, suggested that we develop a software program for doing the structural analysis that would not require a mathematical background on the part of the user. The algebraic part, we decided, could be implemented using Prolog, and we could develop an expert system as a way to provide assistance to a user without back- ground in modeling using algebraic structures. We made the assumption that the algebraic part of the analysis would require, even with the algebraic machinery presented through a user-friendly interface, considerable guidance because the degree to which kinship terminol- ogies had a common, underlying logic was unknown at the time. I had demonstrated that the AKT did have a logic that allowed for the terminology structure displayed in the form of a kin term map to be isomorphically generated from first principles. Additional work on other ter- minologies made it evident that the same was likely to be true for terminologies very different from the AKT, such as the Trobriand kinship terminology used by the inhabitants of the Trobriand Islands. The Trobriand terminology has been the basis for considerable theoretical work on kinship terminologies using ideas such as rewrite rules borrowed from linguistic analysis (see Lounsbury, 1964) and provided a good test case for determining those features of a kinship terminology that could better be explicated through an algebraic or generative approach grounded in ethnographic observations about the way kin relations are calculated by individuals rather than linguistic methods. Consequently, the program was not simply a translation of an existing theory and application of that theory into a computer idiom. Instead the software program became the idiom through which the theory was developed (see Read & Behrens, 1990). The advantage of using a computer program as the basis for developing a theory of kinship terminology structures became evident as work progressed. Much of the algebraic analysis related to working out the structure of an algebra once its generating set and structural equa- tions have been specified. This part of the analysis could be programmed in a relatively straightforward manner using Prolog and obviated the need for a theorem-proof written for- mat. With the algebraic machinery developed in Prolog, the analytical and programming effort could then be focused on what precisely was meant by a terminology having a structure that could be generated from first principles and on what were those first principles. Many of the issues that arose during the writing of the software program had to do with developing the formal, structural equivalent of concepts that were used more intuitively in anthropological theorizing about kinship terminologies, such as reciprocals of kin terms, a key aspect of any kinship terminology. Briefly, the reciprocal of a kin term is the kin term that another person would use to refer to ego when ego refers to a person by that kin term. In the AKT, for example, if I refer to a person as uncle, then that person would refer to me as nephew (because I am a male person). The kin terms uncle and nephew are said to be reciprocal kin 2 Social Science Computer Review terms. Although this is a satisfactory notion from an ethnographic viewpoint, a structural analysis requires that the concept of reciprocity be given a structural definition (to be dis- cussed below) for the concept to be implemented in a software program that provides the machinery for modeling the structure of a kinship terminology as a generative, algebraic structure. An unexpected consequence of simultaneously working out a theory of kinship terminol- ogy structures while implementing that theory in the form of an algorithmically based com- puter program was that the expert system part of the original design for the computer program began to take on less importance as it was discovered that kinship terminology structures are highly logical and have few degrees of freedom in terms of how the structure is generated. Structural properties that are necessary for consistency in calculations with kin terms—such as consistency in sex marking of kin terms used in kin term calculations—and properties of kin terms such as “parent of sibling is parent” in the AKT can be automatically introduced by the algebraic machinery. As the theory became more completely developed, the algebraic analysis performed by KAES almost became a single, large composite algorithm. It cannot truly be a single algorithm as there are different terminology structures that begin with the same generating elements and initial structural equations. What distinguishes different kinship terminology structures are not differences in the algebraic methods of analysis (which would require expert system assistance for the naïve user) but structural differences in terminologies relevant to the ethnographic level of cultural differences in the specification of what constitutes kinship in a particular society. For exam- ple, the Trobriand terminology and the Tongan terminology (used by the inhabitants of the kingdom of Tonga in the Pacific Ocean) have the same initial starting conditions (set of gen- erating elements, structural equations satisfied by these generating elements) but differ at the point where ascending terms (terms used for persons in a generation older than the speaker) are combined with descending terms (terms used for persons in a younger generation) in accordance with structurally alternative ways these two sets of kin terms may be combined together (see Bennardo & Read, n.d.). The structural implications for alternative ways the ascending and descending terms are combined together to form a single structure relate directly to differences in the usages between the two terminologies for terms that otherwise are in similar structural positions in the two terminologies. The unexpected discovery that terminologies may be modeled almost as a single, compos- ite algorithm has major implications for our understanding of cultural constructs such as kin- ship terminologies and their relationship to kinship systems in human societies. Rather than providing support for the received view in anthropology that considered terminologies to be a labeling system for categories of kin distinguished for reasons external to the kinship termi- nology, the KAES program makes it evident that the structural properties of kinship termi- nologies arise from logical constraints on the generation of structure and differences among terminologies relate to places in the generation of a structure where alternative ways to develop the full terminology structure are logically possible. The current version of KAES has been implemented in Java with Michael D. Fischer (Read & Fischer, 2005). In the Java version of KAES, the underlying theory has been elabo- rated further, especially with regard to the relationship between a kinship terminology and genealogy. No attempt was made to translate the Pascal and Prolog code into Java as that code suffered, from a programming viewpoint, from the fact that it was written while developing the theory on which the current computer program is based and was written without having Read / Kinship Algebra Expert System 3 access to a fully object-oriented programming language. The latter has made possible a much more succinct and more precise representation of the underlying theory. Overview of the KAES Program The KAES program has three major components, with menu options including map oper- ations, algebra operations, and operations. The first component (map operations) provides the machinery for entering of kinship terms in the form of a kin term map (see Figure 3) that displays the manner in which kin terms are linked to one another through a kin term product based on ethnographic observations about the ways individuals calculate kinship relations. The kin term map can be represented in the form a table that shows the kin terms that result from taking the product of a kin term with one of the primary kin terms such as mother, father, son, or daughter in the AKT. The kin term map is also presented in the form of a directed graph linked to the table form for displaying the product of kin terms with the primary kin terms. When a kin term is entered by the user, its attributes such as sex marking and whether or not it is a generator kin term are entered by the user. The kin term map displays native knowledge about a kinship system expressed through a kinship terminology and provides the target structure for the algebraic modeling done in the second component (algebra operations) of the program. The algebraic modeling begins by structurally simplifying the kin term map through user- selected options for simplification of a kin term map. The simplification is run in reverse dur- ing the algebraic modeling as one elaborates on an initial algebraic model constructed in accordance with the simplified kin term map. A modified kin term map is defined to be sim- plified when the modified kin term map has a single, ancestral generating kin term and only consists of ancestral kin terms. The KAES program allows for the user to model a kin term map using user-selected options that activate an ensemble of individual steps composing a stage in the algebraic modeling, such as constructing the descendant structure from the ini- tial, ascendant structure (see Figure 5). Alternatively, the user may sequentially activate the individual steps making a stage in the algebraic modeling. As the modeling proceeds, the KAES program automatically introduces structural equations that are part of the distinctions made regarding kin terms in the kinship terminology. Each stage in the algebraic modeling of a kinship terminology structure is successful if the algebraic modeling ends up with an algebraic structure isomorphic to the kin term map from the initial simplification of the kin term map corresponding to the current stage in the alge- braic modeling. The algebraic construction stops when an algebraic structure that is isomor- phic with the kin term map for the complete kinship terminology has been generated. The algebraic structure may be locally modified by user-selected rules that implement the logic by which a portion of the generated structure may be culturally modified. For example, the generated algebraic model for the AKT has an undifferentiated “cousin space.” The cousin space is then modified by a logic applied consistently to kin term products to arrive at the cousin space with the features found in the kinship terminology and expressed through the kin terms of the form “ith cousin j-times removed.” Rules are largely terminology spe- cific. Rules for the terminologies considered so far are included in the KAES program. Addi- tional rules will be added to the KAES program as the analysis expands beyond the kinship terminologies considered so far (see Figure 6 for the algebraic model of the AKT generated by the KAES program, including the rule for the cousin terms). 4 Social Science Computer Review The third component (operations) constructs a genealogical instantiation of the algebraic structure. The instantiation is constructed by first linking the generating elements in the alge- braic structure to the genealogical space. In the AKT, for example, the kin terms parent and child are the primary, generating kin terms. These two kin terms are mapped to the sets of genealogical positions (genealogical mother, genealogical father) and (genealogical son, genealogical daughter), respectively. This mapping, along with the mapping of the identity element to the set (ego), is next extended to other kin terms through set products of the map- ping of the generating elements in corresponding to the primary elements in a kin term prod- uct. The result of carrying out the genealogical instantiation of kin terms can be displayed in tabular form or in the kind of genealogical diagrams used by anthropologists to show the mapping of kin terms onto a genealogical structure (see Figure 8). Anthropological Problem: What Constitutes Kinship? Unraveling what constitutes kinship relations and kinship behavior in human societies has been a central theme in anthropological theorizing ever since the publication by Lewis Henry Morgan of Systems of Consanguinity and Affinity of the Human Family in 1870. The central- ity of kinship in anthropological theorizing stems from the fact that all human societies are constructed, to one degree or another, around kinship and kinship relations. Although larger scale societies may have developed institutions for societal integration that are not directly based on kinship, such as the notion of citizenship and elected government in state democra- cies, no society has removed kinship as a fundamental basis for the social identity of individ- uals. Regardless of other aspects of societal organization, we are born into families that are formed on the basis of kin relations, and from birth onward we are provided, through kinship, with an identity through which we are linked to other individuals, and through these linkages we are provided with a social identity fundamental to our existence as social beings. Our kin- based social identity may coexist with other identities we also hold. In American culture, we are given individual identity at birth through a supposedly unique name—or at least a name unique within the domain of persons known to the parents of the newborn child. The notion of individual identity in American culture is in accord with the concept of a governmental system that also provides citizens with individual rights that recognize the uniqueness of each person as a biological entity. A child also takes on a kinship identity in American culture even though the latter has the opposite effect of constructing a social identity that overrides the biological uniqueness of each person. A newborn child may be a brother or a sister to any other offspring of her or his parents, a nephew or a niece to the siblings of her or his parents, a cousin to the children of the siblings of her or his parents, and so on, simply by virtue of birth and regardless of individual qualities. This kinship identity provides a social identity expressed (in American culture) through terms such as brother, sister, son, daughter, and so on and carries with it expected patterns of behavior by virtue of the kind of relationship involved. Americans have shared (though not identical) notions of how children should behave in their status as children or how siblings should interact in their status of siblings, regardless of their biological uniqueness. In small scale societies, kinship is the basis on which societal boundaries are formed and is a prerequisite to social interaction among individuals. Often the term used to identify the group to which one belongs translates into something like the meaning of the term ju/ ‘hoansi, used by a hunter-gatherer group in Botswana to refer to themselves, namely “we, the Read / Kinship Algebra Expert System 5 real people” (Lee, 2000). Social interaction between individuals does not take place unless they first recognize each other as kin. Nonkin, or strangers, are feared or perhaps considered potentially harmful because those who are not your kin may not share the same set of moral values and modes of behavior as do your kin. Nonkin are not predictable in terms of behavior in contrast to kin, whose behavior is constrained by patterns of behavior associated with the kinds of kin relations. Stranger, in the context of kin-based societies, refers not to those with whom one has not had prior interaction but to individuals outside of the domain of individu- als who share the same moral and social values. The world of individuals with a shared set of moral and social values is one’s kin. Kin, by virtue of being kin, share a common framework of behavior that is appropriate even if the persons in question have not previously met. With larger scale societies, a domain of kin-like relations that extends many of the properties of shared, expected patterns of behavior among kin to a larger domain of individuals has been added. Tribes, for example, include individuals who may not know their kin relationships but do know their connection to one another through the social groups to which they belong such as lineages and how these social groups are linked via connections derived from kin connec- tions. In state societies, patterns of expected behavior have been extended outside of the domain of kin and kin-like relations through institutions such as citizenship that define for a person the set of rules and laws that are to be obeyed and where infraction of those rules and laws may lead to punishment. But even citizenship has a quasikin basis through the notion that citizenship is passed on from parent to offspring. We may reasonably assert, then, that kinship is the fundamental concept on which human societies have been formed and constructed. Without understanding what constitutes kinship and how kinship articulates with, and provides the basis for, interaction in all other domains, we only have an incomplete picture of the nature of human societies and social systems. Try- ing to understand human societies without understanding the underlying kinship system is analogous to trying to understand the chemical properties of matter without understanding the underlying atomic and molecular properties from which those chemical properties derive. The kin relations that connect individuals provide the underlying framework for the interaction of individuals, and without understanding the properties and structural form that kinship relations may take on in a given society, we cannot fully understand how the individ- uals making up that society become social and not just individual beings and how the society is structured as a social system. Although the patterns of expected kin behaviors may vary from one society to another, all societies have a means for identifying or labeling some set of relationships between an indi- vidual and other individuals that is expressed linguistically through a set of terms known to anthropologists as a kinship terminology. For Americans, the kinship terminology is com- posed of the terms such as son, daughter, brother, sister, mother, father, wife, husband, and so on that are used by one individual as a way of referring to the kin relation one has with another individual. It is important to make a distinction between terms of reference (the terms one uses when referring to another individual, such as in the expression from an advertisement for Boys Town in the 1950s: “He ain’t heavy—he’s my brother.”) and terms of address that are used when talking to another individual. The latter is based on the former but may include terms or expressions that relate not simply to the kin relationship between the individuals in question but to patterns of speech that bring into play other aspects of the relationship between the individuals that is being conveyed by the way one person addresses another per- son, such as the use of honorific forms of speech. In American culture, we have a variety of 6 Social Science Computer Review expressions that a child may use when addressing her or his male parent, such as “dad,” “pop,” “father,” “pa,” or even the person’s first name, with each mode of address having a dif- ferent connotation. For our purposes here, we will focus on terms of reference only. Even when we limit ourselves to terms of reference, a complication enters through the fact that we have two distinct ways by which we may refer to someone’s kin relationship to speaker. One way is through kin terms, such as the expression, “This is my Aunt Sally.” The other is through specification of the genealogical relationship of the person in question to the speaker, as in the expression, “She is my mother’s sister.” Sometimes both ways of referring to a kin person may occur sequentially, as in the expression, “This is my Aunt Sally—she is my mother’s sister,” depending on the information that the speaker wants to convey about her or his relationship to the person in question. More generally, corresponding to each kin term, there is one or more genealogical relation for which the kin term may be substituted. For some kin terms, the kin term and the genealogical relations may not be linguistically distin- guished. In American culture the expression, “She is my mother,” is ambiguous in that the term mother can either be the kin term mother or the genealogical relation mother. Similar comments apply to the American or English kin terms father, brother, sister, son, and daugh- ter. This correspondence between kin terms and genealogical relations is not true of all termi- nologies. For many terminologies their term that we would transliterate as mother, for exam- ple, does not simply correspond to the genealogical relation mother but may also be properly used for other genealogical relations such as mother’s sister (mother’s daughter), mother’s mother’s daughter’s daughter, and so on. (A similar comment applies to the term we would transliterate as father.) For terminologies of this kind, the kin term mother is similar to kin terms such as uncle and aunt in the American or English kinship terminology in that the kin term can properly be used to refer to several, different genealogical relations. Kinship terminologies generally cannot be faithfully translated from one language into another as the relationships that are recognized in one society may not have an exact counter- part among the relationships recognized in another society. Terminologies also vary with respect to the specificity of a kin term with regard to genealogical relations. Whereas the terms aunt and uncle in American culture do not distinguish whether the relation is through the siblings or spouse of siblings of father versus mother, other terminologies use different terms depending on whether or not the relationship is through one’s (genealogical) mother or father as can be seen by comparing the kin term map for the AKT with the kin term map for the kinship terminology used by the Shipibo Indians of Peru (see Figure 1). The basis for the differences in the kin term relationships that are recognized by different societies has been a source of contention in anthropological theorizing. Equally problematic, though not as systematically addressed, has been identification of the basis for the particular relationships recognized within a given society. At one extreme is the presumption that the relationships linguistically identified through a kinship term are essentially biological in ori- gin and are to be understood as distinctions arising through the process of Darwinian evolu- tion. The biological argument relates the distinctions made within the kinship terminology to reproductive success, taking into account kin selection, inclusive fitness, and so on. A typical kind of argument would be the one advanced for why, in many lineage-based societies, the brother of a sister has a close and authoritative role vis-à-vis his sister’s children and lines of inheritance may be from him to his sister’s son. The emphasis on sister’s brother rather than (putative) genetic father stems, it is argued, arises when paternity is uncertain because of the fact that extramarital affairs are not severely sanctioned. Uncertainty of paternity does not Read / Kinship Algebra Expert System 7 extend to his sister’s children (assuming he and his sister are genetic siblings), and so direct- ing his parenting toward his sister’s children when paternity uncertainty is sufficiently high increases his reproductive success. More precisely, because his genetic correlation with his genetic children is one half and his genetic correlation with his genetic sister’s children is one fourth, then if the certainty of paternity is less than one half, his genetic correlation with his sister’s children is greater than his likely correlation with his wife’s children (½ � probability of being the genetic father), and so parenting directed toward his sister’s children gives him greater reproductive success than does parenting directed toward his wife’s children. A more middle-of-the-ground viewpoint, and one that characterizes most anthropological theorizing about kinship systems, has assumed that the kin terms are based on the biological relationship of reproduction but given cultural expression that may deviate from their biolog- ical grounding (Keesing, 1975, p. 13). A distinction is made between genetic father (the man who provides the sperm) and genetic mother (the woman who provides the ovum) and geni- tor (the man identified as the physical father but who need not be the man who provided the sperm) and genetrix (the woman identified as the physical mother but who need not be the woman who provided the ovum) because knowledge of genetics and its role in reproduction was not known until the work of Gregor Mendel in the mid-1800s. Because of the obvious link between pregnancy and birth, in most societies the woman recognized locally as the genetrix (assuming such an identification is made) is the genetic mother; however, adoption 8 Social Science Computer Review Figure 1 Comparison of the American Kinship Terminology With the Shipibo Kinship Terminology allows for a woman to be identified as mother even though she is not the genetic mother or the birth mother. In societies such as the Inuit, a woman adopting a child is as real a mother as a woman giving birth to a child, and no distinction is made between a child to whom she has given birth and a child she has adopted. In American culture, we give the adopting woman the kin designation mother but recognize that she is not the physical or biological (or birth) mother. The assumption in this perspective is that reproduction and marriage are the basis for kin relations and that the former may have local, cultural constructions that need not match actual genetic linkages. It is assumed that all societies have a notion of genealogy grounded in reproduction and marriage—however, these may be locally constructed—and the genea- logical relations are the basis for asserting a kinship connection. Kin terms are viewed as lin- guistic labels identifying the classification that the members of a society make of a genealogical space grounded in marriage and procreation. A third perspective, and the one underlying the KAES program, considers kin terms to constitute a system of interconnected concepts expressed through the conceptual connec- tions among the kin terms (Read, 2001). The terminology has a structure in its own right that may be identified without first referring to another domain such as a genealogical space. The structure is based on the manner by which we calculate kin relations using kin terms without first translating kin terms to genealogical representations for the kin terms. In this perspec- tive, the set of kin terms is not just a set of linguistic symbols but also a set of linked symbols that forms the kinship terminology into a structured set of symbols. The structure expressed through the linkages among the kin terms is generative in that the structure can be generated from a few primitive kin terms and structural equations that express the logic underlying the structural form for the terminology. Structural properties arise, in this perspective, through the logic by which the structure is generated from these primitive kin terms. Thus, a kinship terminology is a structured set of symbols that cannot be fully understood without reference to the way the system is generated as a conceptual construct. In this perspective, genealogical tracing, which is the conceptual basis underlying a genealogical space, is also recognized as generating a conceptual system—the genealogical space—distinct from the kin term concep- tual system but connected to it through a mapping of the primitive kin terms onto the genea- logical relations used in genealogical tracing. Although the previous perspective that posits kin terms to be semantic labels for the classes making up a classification of the genealogical space presumes a conceptual hierarchy going from a genealogical space to a kinship termi- nology, this third perspective views the genealogical space and the kinship terminology space to be two coexisting conceptual systems, each with a structure determined by the logic of how the space in question—genealogical or terminological—is generated and with a map- ping that connects the two spaces but that is necessarily from the kin terminology space onto the genealogical space. The mapping demonstrates that the classification made of the genea- logical space and the definition of kin terms using genealogical relations are both the consequence of a more fundamental set of concepts. Hence the definitions of kin terms using genealogical relations and assumed to be primitive concepts in the second perspective are not primitive but instead are derived concepts. The differences among these three perspectives—biological, genealogical, and cultural— are subject to verification and falsification as each perspective has different implications for what constitutes a kinship terminology. First of all, the biological perspective implies that the kin relations identified through the kin terms making up a kinship terminology should bear a close resemblance to the actual biological relatedness of individuals and differences in termi- Read / Kinship Algebra Expert System 9 nologies should relate to the pattern of reproductive success in the societies in question in accordance with terminological differences between kin terms and their referents. In this per- spective, there should be a close connection between what are effective patterns of social behavior with respect to reproductive success and normative patterns of interaction among kin distinguished through the kinship terminology. In brief, the biological perspective implies that the structure of a kinship terminology and its usage cannot be understood with- out reference to the biological relatedness of pairs of individuals, their behaviors, and the consequences of those behaviors for reproductive success. Second, the genealogical perspective implies that a terminology can primarily be under- stood through the criteria used to classify the possible genealogical relations between pairs of individuals and how different classifications of genealogical relations arise in accordance with linguistic, psychological, and/or ecological constraints. Kin relations in this perspective should be demonstrable as first being conceived genealogically and then given terminologi- cal expression according to the definition of kin terms using genealogical relations. The genealogical perspective implies that the structure of a terminology should reflect the criteria and constraints through which a classification of a genealogical space is formulated. In contrast to both of these perspectives, the cultural perspective implies that the terminol- ogy will have a structure whose features can be constructed or generated from a few funda- mental concepts along with a general process for the production of the structure of a kinship terminology. A key aspect of this perspective is the logic underlying the structure of the ter- minology. That logic implies that the kinship terminology should be a logically consistent structure, hence many of its features will be predictable from the logic of how the structure may be generated. In contrast, neither the biological perspective nor the genealogical per- spective imply that kinship terminologies should be highly logical in terms of how they con- stitute a symbolic system. Hence analysis of a kinship terminology with respect to it being constituted in accordance with an underlying logic provides a basis for distinguishing among these three perspectives on what constitutes kinship in human societies. Kinship Terminology Logic The KAES program is designed to explore the logic underlying the way the kinship terms form a structured system of symbols (with symbols taken in a linguistic and mathematical sense). But more than this, the program is simultaneously an idiom through which a theoreti- cal framework has been developed, a research tool aimed at resolving a fundamental question about a central feature of human societies, and an application program for the analysis of the structural properties of a kinship terminology and its relationship to the genealogical domain. The theoretical framework centers around working out the logic of kinship terminologies based on properties that are based on ethnographic observation. The key ethnographic obser- vation lies in the means by which individuals work out kin relationships. Kin Term Product The structural modeling portion of the KAES program is based on the common ethnographic observation that individuals usually do not determine kin relations by identify- ing their biological pedigree or their genealogical connectedness but by doing computations with kin terms. As the anthropologist Marshall Sahlins (1962) has noted with regard to Moala kinship: 10 Social Science Computer Review [Kin] terms permit comparative strangers to fix kinship rapidly without the necessity of elabo- rate genealogical reckoning—reckoning that typically would be impossible. With mutual rela- tionship terms all that is required is the discovery of one common relative. Thus, if A is related to B as child to mother, veitanani, while C is related to B as veitacini, sibling of the same sex, then it follows that A is related to C as child to mother although they never before met or knew it. Kin terms are predictable. If two people are each related to a third, then they are related to each other. (p. 155, italics added) Based on ethnographic observations similar to Sahlins’ comment about Moala kinship, we may define a kin term product informally as follows. Suppose one person (ego) properly refers to a second person (alter 1) by the kin term K, and alter 1 properly refers to a third person (alter 2) by the kin term L. The product of the kin terms K and L will be the kin term, M, if any, that ego properly uses for alter 2. For example and with respect to the American or English kin- ship terminology, when ego refers to alter 1 by the kin term aunt and alter 1 refers to alter 2 by the kin term daughter, then ego (properly) uses the kin term cousin to refer to alter 2. So the product of the kin terms daughter and mother is the kin term cousin in the AKT (see Figure 2). We need to make a clear distinction here between kin types and kin type products, which have to do with genealogical relations, and kin terms and kin term products, which have to do with the conceptual system we refer to as a kinship terminology and the structure formed from Read / Kinship Algebra Expert System 11 Figure 2 Illustration of a Kin Term Product for the Kin Terms Aunt and Daughter From the American Kinship Terminology Note: The product of these kin terms yields the kin term cousin. relations linking kin terms via kin term products. Kin term products express the ways in which the kin terms (viewed as linguistic symbols) form a structured set of symbols. In contrast, kin type products refer to the ways in which a genealogical relation can be constructed through concatenation of pairs of genealogical relations, such as the expression, “Mother’s brother’s daughter.” In the context of genealogy, the words mother, brother, and daughter refer to gene- alogical tracing based on genitors and genitrixes. Genealogical mother in American culture is the woman who begat ego, and if her identity is unknown, then genealogical tracing through a female parent cannot proceed. This contrasts (in American culture) from kin relations through adoption that are excluded for purposes of genealogical tracing. In both cases, the same kin term, mother, may be used for purposes of reference. Thus, if ego has been adopted, then ego has a mother from the perspective of the AKT but not a genealogical mother from the perspective of kin types. Anthropologists have used a technical distinction between social mother, or mater, and physical mother, or genetrix, to distinguish these two situations. The adopted mother, in American culture, is a social mother but not a genetrix. However, this dis- tinction is culturally salient only to the extent that it is a culturally recognized distinction. For some groups such as the Inui mentioned above, no distinction is made between genetrix and mater (or genitor and pater), and so a social mother is also a genetrix for the Inuit. From the perspective of kin terms, the matter is simpler because either the person in ques- tion is one for whom a kin term of reference is appropriate or not according to cultural rules regarding the usage of kin terms. In American culture, the kin term mother applies equally to a genetrix or to a social mother. Mother (in American culture) is thus a polysemic term that can have the kin type, genealogical mother, as its reference, that can have the social mother as its reference in the case of adoption, or that can be taken as a kin term with linkages to other kin terms through kin term products. When necessary, a distinction between a kin type versus a kin term will be made by capitalizing the kin term and using lower case when referring to a kin type. For kin terms, the kind of calculation discussed by Sahlins (1962) provides the conceptual basis for viewing kin terms as forming a structured set of symbols. More formally, we define a kin term product as follows: Definition: Let K and L be kin terms in a given kinship terminology, T. Let ego, alter1, and alter2 refer to three arbitrary persons, each of whose cultural repertoire includes the kinship terminol- ogy T. The kin term product of K and L, denoted K o L, is a kin term, M, if any, that ego may (prop- erly) use to refer to alter2 when ego (properly) uses the kin term L to refer to alter1 and alter2 (prop- erly) uses the kin term K to refer to alter2. Right-to-left product notation will be used as it permits reading the expressing K o L as “K of L” for a kin term product, and this contrasts with the reading of KL as “K’s L” for the genea- logical (kin type) product KL of kin types. For example, the symbolic expression S o M = A, where M stands for mother, S for sister, and A for aunt, would be read “sister of mother is aunt,” whereas the kin type product ms would be read “mother’s sister” and aunt would be the kin term that can be applied to a person who is one’s ms. Kin Term Map Kin term products are the bases on which the analysis of a kinship terminology structure proceeds. One part of the KAES program deals with representing the structure of a kinship 12 Social Science Computer Review terminology based on kin term products. A diagram displaying the structure for kin terms based on kin term products is called a kin term map and is the beginning point for the analysis of the structure of a kinship terminology. Figure 3 shows a kin term map for the AKT based on taking products with the kin terms parent, child, and spouse. A variant of this kin term map can be constructed using kin term products with the kin terms mother, father, son, daughter, wife, and husband. In general, more than one kin term map may be possible for the same terminology. Generating Terms Although the kin term product is defined for any pair of kin terms, we quickly discover that certain kin terms play central roles in constructing a kin term map. For the AKT, these are terms such as mother, father, and parent and their reciprocals daughter, son, and child. Other kin terms in the AKT can be constructed by taking products of these kin terms. For example, the kin term grandfather = father o parent (read, “Grandfather is father of parent,” or more completely, “The kin term grandfather results from taking the kin term product of the kin term father with the kin term parent”), where this is a statement about the three symbols, grandfather, father, and parent. The symbol o between father and parent indicates that we have a binary product (namely the kin term product) that maps a pair of symbols (in this instance, the father, parent pair) to another symbol (in this instance, the symbol grandfather). Read / Kinship Algebra Expert System 13 Figure 3 Kin Term Map for the American Kinship Terminology Based on Taking Products With the Kin Term Parent and its Reciprocal Term Child and the Kin Term Spouse Though this equation may look like a genealogical statement, nothing could be further from the meaning of this equation. The equation states that when ego properly refers to a per- son as parent (and regardless of the actual genealogical connection, if any, between the two persons) and that person properly refers to a third person as father (again, regardless of the actual genealogical connection between them, if any), then ego would properly refer to that third person by the kin term grandfather (again, regardless of the actual genealogical rela- tionship between them). Some of the persons in question might be adopted and so the actual genealogical relations that are involved are unknown; nonetheless the users of the AKT still know the proper usage of the terminology in cases of adoption by virtue of the conceptual relations that hold among the kin terms. Criteria for Generating Terms The kin term map is constructed by first deciding on the kin terms that will serve as the generating terms through which kin terms are linked to one another. The generating terms must account for the closest kin term linkages between a person and another person, and so they will likely be some variant on the kin terms that can be transliterated as mother, father (and possibly parent), and their reciprocals, daughter, son (and possibly child). In addition, terms that can be transliterated as brother or sister (or possibly older brother, younger brother, etc.) will be generating terms in some terminologies but not in other terminologies. In the AKT, the analysis of the logic underlying the structural form of the kin term map dis- played in Figure 3 establishes that the kin term brother is the product of the kin terms parent and son: brother = son o parent. In other terminologies, a kin term that can be transliterated as brother (or sister) or possibly as older brother (or older sister) is a generating term and not expressible as a kin term product. The distinction between a sibling term taken as a product versus a sibling term taken as a generating term provides the conceptual basis for the distinc- tion between so-called descriptive terminologies versus classificatory terminologies (see below). Kin Term Map as a Generated Structure The kin term map, even without any additional analysis, serves as an effective graphical device for displaying the structural differences among kinship terminologies. Each terminol- ogy has its own structural form (or variant on a structural form for closely related terminolo- gies) that expresses differences in the ways in which the kin terms of a terminology form a structure. The kin term map is more than just an effective descriptive means to display the conceptual structure formed by the kin terms in a kinship terminology. An obvious question is whether the map can be generated from a few simple structural properties, and if so, does this account for the features of the terminology (such as the odd way in which the aunt and uncle terms in the AKT are used both for individuals linked genealogically through parent and child links to ego and for individuals who are linked to ego only through marriage). If the kin term map is the graph of a generative structure, then it follows that a kinship terminology is neither simply a representation of biological kin relations nor a classification of a genealogical space by cri- teria that lie outside of the kinship terminology. Instead the kinship terminology is a cultural construct, and the meaning of kin terms cannot be fully understood without reference to that construct and the logic underlying how it is generated. 14 Social Science Computer Review The KAES program is designed to have a number of features aimed at helping construct a kin term map. Kin terms may be entered in any order and at any time. Any kin term that has been entered may be designated as a generating term, and the pattern of arrows that may be drawn showing the kin term product of a kin term with a generating term will be constrained by the features of the generating term such as its sex marking and whether it is an ascending term (i.e., refers to alters in an older generation with respect to ego). Generating Terms: Descriptive Versus Classificatory Terminologies One of the key discoveries about terminology structures that has been obtained to date relates to the distinction Morgan (1870) made between what he called descriptive terminolo- gies versus classificatory terminologies. The choice of labels was unfortunate as it led to sub- stantial debate about the validity of the distinction within anthropological theorizing about kinship terminological systems based on the labels descriptive versus classificatory rather than the structural differences that distinguish the two kinds of terminologies (e.g., Lowie, 1928). At the level of kin terms, some kin terms in a terminology may be descriptive in the sense that the kin term can be given a genealogical definition using a single kin type (such as mother is the term used in reference to the kin type mother), and other terms in the same ter- minology may be classificatory in the sense that a genealogical definition of the kin term will refer to more than a single kin type, such as uncle with a genealogical definition: mother’s brother, father’s brother, mother’s sister’s husband or father’s sister’s husband. The distinc- tion Morgan was attempting to identify had to do with structural properties such as whether or not lineal relations traced using either parent kin types only or child kin types only were kept distinct from collateral relations traced using first parent kin types and then child kin types. Thus in the AKT, parent is a lineal kin term, whereas aunt or uncle are collateral kin terms, and the collateral kin types to which the later two kin terms refer are kept distinct from the lineal kin type to which the kin term parent refers. In classificatory kinship terminologies that are common in Polynesia and other groups inhabiting islands in the Pacific Ocean, parent and parent’s siblings are not kept distinct, and a single kin term, mother, refers to mother and mother’s sister (and other females on the same generation as mother), and a single kin term, father, refers to father and father’s brother (and other males on the same generation as father). Morgan’s (1870) structural distinction based on lineal versus collateral only partially identifies the basis on which the lineal-collateral distinction arises. By working out the gener- ative implications of taking a sibling term as a primitive during the process of writing the KAES program, it became evident that the structural distinction Morgan was attempting to identify relates to whether or not the sibling kin terms are generating kin terms or are products of other kin terms. As noted above, in the AKT, the kin term brother is constructed via brother = son of parent (which can also be expressed via the two equations brother = son of father or brother = son of mother). The kin term map for the AKT is not consistent with a claim that the sibling terms are generating terms for the terminology, and so the AKT is a descriptive termi- nology. In contrast, classificatory terminologies such as the Trobriand terminology or the Ton- gan terminology in the Pacific Ocean area are found to be terminologies for which sibling terms are generating terms. The analysis of the structure of these terminologies demonstrates that the kind of genealogical equations that have been used to define classificatory terminolo- gies, namely f = fb and m = mz, are a consequence of the structural implications that arise from including the sibling kin terms among the generating terms for the terminology. Read / Kinship Algebra Expert System 15 Algebraic Model for a Kin Term Map The algebra part of KAES refers to the use of algebraic concepts to model the structure of a kinship terminology as displayed through a kin term map. By algebraic structure is meant the structure that is formed by a set of symbols (e.g., the kin terms making up a terminology), a binary product defined for all pairs of symbols in that set (e.g., the kin term product for kin terms), including whether or not the binary product is associative—for example, for the binary operation, +, the sum a + (b + c) is the same as the sum (a + b) + c; for kin terms, the product K o (L o M) = (K o L) o M—or commutative (e.g., for the binary number operation, +, a + b = b + a; however, commutativity does not hold for most kin term products; e.g., son of father is not the same as father of son), and certain equations that the binary product must sat- isfy. The equations give an algebra its particular structural form. For kinship terminologies, we have the following basic algebraic machinery: 1. A set of symbols—the terms making up the kinship terminology. 2. A binary product defined over those symbols—the kin term product, including 0 as a symbol indicating that the product does not yield a kin term; e.g., father o father-in-law = 0 in the AKT because there is no kin term representing the product father o father-in-law. 3. The binary product is associative—the kin term product is associative, except in certain special cases that are dealt with in the analysis on an individual basis. An algebraic system satisfying all three of the above is known in the mathematical litera- ture as a semigroup. The equations that are used when modeling a kinship terminology are used to introduce structural properties that characterize kin terms. The structural equations provide a way to express the structural properties of a kind of kin term with respect to the kin term map and thereby provide a formal definition for what otherwise are informal concepts associated with terminologies. These properties provide definitions for kinds of kin terms such as: (a) a sib- ling kin term, S, is a kin term for which S o S = S (e.g., in American culture, brother of brother is brother as a kin term product); (b) a spouse kin term, Sp, is a kin term for which Sp o Sp = self (with self a symbol representing the concept of self, a concept fundamental to any kin- ship system), and so on. The equation S o S = S, for example, follows from the fact that as users of our own terminology, we know what the kin term brother means. Among the various meanings associated with the term is the structural property that if ego properly refers to alter 1 as brother and if alter 1 properly refers to alter 2 as brother, then ego properly refers to alter 2 as brother, hence the structural equation S o S = S becomes the way to structurally define a kin term as a symbol that can be given instantiation with the concept of brother (or equiva- lently, with the concept of sister because sister of sister is sister). The structural equation does not provide an exhaustive definition of the meaning of, say, the term brother, as the instantiation (see below) of an (abstract) symbol may include properties that are not part of the kin terminology structure and bring to bear other cultural aspects associated with cultural constructs relevant to the cultural notion of what constitutes (in this example) brotherhood. The structural equations also express relations between kin terms such as reciprocity of kin terms. By reciprocity of kin terms is meant that if ego properly uses the kin term K to refer to alter, then the kin term L that alter properly uses to refer to ego is the reciprocal kin term for K. Because of sex marking of kin terms, a kin term need not have a unique reciprocal term. The AKT kin term mother, for example, has both son and daughter as reciprocal kin terms. 16 Social Science Computer Review The structural property that relates two kin terms K and L as reciprocal kin terms is either the structural equation, K o L = self or L o K = self, depending on whether K or L is the ascending kin term. If K is the ascending kin term, then K o L = self is the reciprocal equation that defines L as a reciprocal kin term for K. The notion of reciprocity is almost equivalent to the notion of an inverse element for algebraic structures in that in one direction—first using a descending term and then an ascending term—reciprocity refers back to ego, and self is the concept used by ego to refer to herself or himself. In the other direction, the product does not refer back to ego; otherwise the only kin terms, expressed genealogically, would be lineal kin terms. Thus, for the reciprocal kin terms father and son in the AKT, father of son is self as a kin term product, but son of father is not self and instead identifies a new kin term, brother. The KAES program generates these structural equations that express the structural properties of kin terms automatically as the analysis progresses. General Model for the Structure of a Kinship Terminology The following is an analytical sequence for generating a kinship terminology that provides the basic framework for the algebraic analysis of a kin term map. The algebraic portion of the KAES program is structured around this analytical sequence. 1. A structure isomorphic to a sequence of ascending kin terms, all with the same sex marking, is constructed based on an ascending generating term such as mother if the terms are to be marked female, father if the terms are to be marked male, and parent if the terms are to be neutral (see Figure 4). Included in this ascending structure is an identity element that can be designated male self, female self, or self, depending on whether the ascending terms are marked male or female or are neutral. By an identity element is meant an algebraic element I such that for any element K, K o I = I o K = K. The identity element will be designated by the symbol self (with or without a sex marker) on the grounds that a concept of self is both universal and necessary for there to be a kinship terminology; that is, all terminologies must make a distinction between self and other, whether or not there is a linguist symbol representing the concept self in the kinship terminology. 2. A structure of descending kin terms is constructed from the structure of ascending kin terms by making an isomorphic copy of the ascending kin term structure and then a larger algebraic structure made up of both ascending kin terms and descending kin terms, and cross products between ascending and descending kin terms are formed (see Figure 5). Included here is the structural equation that makes an ascending kin term—call it P—and its isomorphic kin term— call it C—into reciprocal terms, namely P o C = I. 3. The next step introduces kin terms with the other sex marking or, alternatively, bifurcation of neutral kin terms into male-marked and female-marked kin terms. First, when the kin terms constructed in steps 1 and 2 are already sex marked, then terms of the opposite sex are con- structed by taking an isomorphic copy of the structure generated in steps 1 and 2. The isomor- phic copy will be a structure of kin terms with the opposite sex. The original structure and this isomorphic structure are combined together, along with appropriate structural equations, to make a new, larger structure consisting of terms marked with both sexes. Second, when the kin terms constructed in steps 1 and 2 are all marked neutral, each kin term is bifurcated into two kin terms marked as male and female terms. Appropriate structural equations that relate male kin term and female kin term products are introduced as appropriate. Read / Kinship Algebra Expert System 17 4. Affinal kin terms are introduced next, along with appropriate equations that express the proper- ties of affinal kin terms. (See the double-headed, gray arrows in Figure 6 that represent taking products with the spouse generator. Figure 6 is the structure produced from steps 3 to 5.) 5. Rules are identified that refer to ways that portions of the overall structure are locally modified. For example, in the AKT, not all kin terms are marked as either male terms or female terms according to the sex marking rule: “A kin term, K, is marked neutral when the product of K with a spouse term is not a kin term.” Another rule in the AKT identifies the way in which cousin kin terms may be further distinguished according to properties such as “ith cousin” or “ith cousin j- times removed.” These rules relate to ways in which parts of the overall structure are modified for cultural reasons and are not modifications necessary for a terminology to have the structure of a kinship terminology (see the shape of the cousin space in the upper right portion of Figure 6). Criterion for the Validity of a Structural Model An algebraic model for a kinship terminology is a valid model for that terminology if and only if the algebraic model is isomorphic to the kin term map representation of the kinship terminology. Not all conceivable kin term maps can be validly represented by an algebraic model. Success in formulating a valid algebraic model for a kinship terminology map identi- fies the means through which the terminology structure can be logically generated in accor- dance with the general model for the construction of a kinship terminology. The kin term map 18 Social Science Computer Review Figure 4 An Ascending Structure Based on the Generator, P, and the Identity Element, I. Note: The structure will be structurally defined as an ascending structure by the structural properties that the gen- erating term, P, will satisfy when an isomorphic copy of this structure becomes the descending structure (see Figure 5). shown in Figure 3 is structurally isomorphic to the algebraic structure shown in Figure 6 (Read & Behrens, 1990), and so the algebraic construction has generated a valid algebraic model for the kin term map of the AKT. Properties of the algebraic model are properties that arise from the logic of how a kinship terminology can be generated and do not require reference to properties outside of the termi- nology. For example, the algebraic model that validly represents the AKT has the property, spouse o aunt = uncle and spouse o uncle = aunt, hence the lack of terms marked with the suf- fix -in-law such as aunt-in-law or uncle-in-law for these relations by marriage is a conse- quence of the logic underlying the generation of the AKT as a kinship terminology structure. In Figure 6, the double-headed, gray spouse arrow maps the uncle node onto the aunt node as a consequence of the logic of how the structure displayed in Figure 6 is generated. Rules that are introduced in step 5 identify structural properties that are not necessary for the structure to have the structural form of a kinship terminology; hence they are properties that may have been introduced for reasons extrinsic to the terminology structure per se. In other words, the analysis distinguishes between properties of the terminological structure that are necessary for the structure to be complete and consistent as a kinship terminology structure and properties that are parts of the terminology but are not necessary for its com- pleteness and consistency and hence are likely to have origin for reasons extrinsic to the generative logic of the kinship terminology structure. Instantiation of the Symbolic Structure The algebraic model of a kinship terminology structure consists of abstract symbols and the relations among these symbols as determined by structural equations and other properties guiding the generation of the symbolic structure. These abstract symbols are culturally instantiated (see Figure 7; see Read, 2002, for a discussion of cultural instantiation) with Read / Kinship Algebra Expert System 19 Figure 5 An Isomorphic Copy of the Ascending Structure, With Generating Terms C and I, is Constructed and Joined to the Ascending Structure, Including Products Between Terms in the Ascending Structure and the Descending Structure Note: The structural equation, P o C = I, is introduced to structurally define P and C as reciprocals. The structural equation also structurally identifies P as an ascending generating term and C as a descending generating term because an equation defining one term as the reciprocal of another term is of the form, ascending term � descend- ing term = self. respect to the individuals for whom the symbols may be properly used as terms of reference when using the kinship terminology to identify kin relationships between individuals. A pri- mary cultural instantiation is through genealogical concepts and is obtained by identifying the instantiation of the generation terms in the algebraic structure with kin types underlying the construction of a genealogical space. A male-marked kin term used in the construction of the ascending structure (e.g., it is instantiated with the kin type, father; more precisely, with genealogical father). The reciprocal male-marked kin term in the descending structure is instantiated with the kin type genealogical son and so on. Predicted Genealogical Definitions of Kin Terms Once the generating terms have been instantiated with genealogical kin types, then the instantiation of all other terms can be worked out from the instantiation of the generating kin 20 Social Science Computer Review Figure 6 The Algebraic Structure Generated From the Structure in Figure 5 After Steps 3 to 5 Are Implemented Note: The solid, single-headed arrows show the result of taking a product with the generator, P. The dashed, single- headed arrows show the result of taking a product with the reciprocal generator, C. The gray, double-headed arrows show the result of taking a product with the spouse generator, S. The oval around a pair of nodes indicates a node from the structure produced in step 2 (see Figure 5) that was bifurcated into a pair of sex-marked nodes in step 3. The gray nodes are the affinal (relations by marriage) nodes introduced in step 4 when a spouse-generating ele- ment was added to the structure. The gray nodes are labeled either with kin terms having an in-law suffix or are labeled with the spouse terms, husband and wife, when the algebraic structure is isomorphically mapped onto the kin term map shown in Figure 3. terms and the algebraic structure by identifying algebraic products with products of sets of kin types. For example, grandfather has the instantiation (genealogical father) � (genealogi- cal father) = (genealogical father’s father) because grandfather = father o father and father has the instantiation (genealogical father). This process of instantiation of the generating terms and construction of the instantiation of all other kin terms based on the algebraic structure leads to a predicted set of genealogical definitions of kin terms. For all terminologies considered to date, the predicted genealogical definitions match with 100% accuracy the genealogical definitions elicited by anthropolo- gists (see Figure 8). These genealogical definitions have been assumed to be the primary and irreducible data for the analysis of kinship terminologies. The KAES analysis demonstrates to the contrary that these definitions are predictable from the way in which the terminology is generated as a structure. Instantiation is not limited to genealogical instantiation. Instead, instantiation depends on cultural rules for mapping abstract symbols to individuals and can be based on other criteria Read / Kinship Algebra Expert System 21 Figure 7 Diagram Showing the General Process of Cultural Instantiation of the Algebraic Structure Note: The nodes in the algebraic structure have semantic meaning with respect to the structure in which they are embedded and are then given semantic meaning with regard to individuals to whom the kin terms are applicable through cultural rules of instantiation. One set of rules relates the algebraic structure to a genealogical space by mapping the generating terms for the algebraic structure onto the genealogical space and using the logic of the algebraic structure to determine the mapping of all other nodes in the algebraic structure onto the genealogical space. such as adoption or other means by which individuals are incorporated within the conceptual structure of a kinship terminology. Metaphoric Instantiation of Kin Terms The constraint imposed by the algebraic structure on instantiation lies in the structural properties of a kinship terminology that are mapped onto individuals through the process of instantiation, such as instantiation must be consistent with the reciprocity of kin terms and the generative properties of the kinship terminology structure. The latter implies that kin term products must also map over to individuals under the cultural rules of instantiation. Thus an adopted child is not only a child vis-à-vis the adopting parents, but the adopted child has the appropriate kin term relationship to other individuals in the kinship domain of the adopting parents. This contrasts with instantiation of individual kin terms that does not carry over to other kin terms, such as the use of the kin terms aunt and uncle by users of the AKT in an hon- orific sense (e.g., the use of the uncle and aunt terms for the friends of one’s parents). The dis- tinction between the more extended and the more limited forms of instantiation provides a 22 Social Science Computer Review Figure 8 The Mapping of the American or English Kinship Terminology Kin Terms Onto the Genealogical Space as Predicted From the Algebraic Analysis Performed by the Kinship Algebra Expert System Program Note: All of the predictions match the genealogical definitions of kin terms for the American or English Kinship Terminology. more formal way to express what has sometimes been referred to as metaphoric usages of kin terms (Scheffler, 1972) and distinguishes when instantiation is metaphoric and not primarily for the usage of the kinship terminology. The genealogical perspective on kin terms assumed, without ethnographic justification, that the primary meaning of kin terms is with respect to kin types, and so any application of kin terms outside of the genealogical domain or where the genealogical definition included noncomparable kin types became metaphorical extension. Classificatory terminologies, in this viewpoint, had kin terms with translation mother and father whose primary meanings were genealogical mother and genealogical father, respec- tively, and the usage of the same kin terms for mother’s sister and father’s brother (and other females and males) was taken to be metaphoric extension of the primary, genealogical mean- ing of the kin terms. The analysis developed through writing the KAES program construc- tively shows that the metaphoric extension hypothesis is unnecessary and was devised as a way to accommodate an unwarranted assumption about the nature of kinship terminologies to ethnographic facts that were not in accord with that assumption. There is no need to posit metaphoric extension as there is no need to posit extension from a primary, genealogical meaning of kin terms. The analysis produced by the KAES program constructively demon- strates how the range of kin types included under a kin term arises from the mapping of the generating kin terms to the genealogical domain and the structural equations that determine the structural form for the kinship terminology structure. There is no metaphoric extension as the full range of kin types included under a kin term is generated as part of the logic of the kinship terminology structure and how that structure relates to the genealogical domain. Conclusion The KAES program has successfully implemented a theory of kinship terminology struc- tures that makes evident the underlying logic giving a terminology its particular form and accounts for the way the kin terms relate to genealogical relations. Writing the KAES pro- gram has been central to working out the logic of kinship terminology structures and, through that logic, the broader implications for what we understand to constitute cultural constructs. The program clarifies the way in which we have two conceptual systems regarding kin rela- tions—the manner in which individuals are connected to one another through genealogical connections and the kin terms that identify the kin relation one individual has to another indi- vidual. The biological argument cannot account for the logic modeled through the KAES program; hence it is insufficient to consider kinship, as it occurs in human societies, through biological criterion alone. Human societies are based on what we can call cultural kinship in contrast to biological kinship. This is not to say that biological kinship is irrelevant; cultural kinship has its evolutionary origins in biological kinship but has transformed those origins. Similarly, the received view in anthropology that has viewed kinship as arising from mar- riage and reproduction is not so much incorrect as inadequate. Historically, some kind of pro- cess of genealogical tracing of connections must have preceded the “invention” of the con- ceptual structures we refer to as kinship terminologies. One possibility for the evolutionary origin of kinship terminology systems is that they provided a computational device for calcu- lating relations among individuals that was cognitively less demanding than memorizing extensive genealogical linkages. But even if it arose as a computational system, the computa- tional facility enabled the terminology to become the central means by which individuals determine if they have a kinship connection or not as indicated in the quote from Sahlins Read / Kinship Algebra Expert System 23 (1962) given above. One can calculate kin relations with the kinship terminological system without necessarily referring back to genealogical linkages, and hence the relative impor- tance of genealogical connections versus kin terminology connections has become society specific. Some Australian groups, for example, find it difficult to compute genealogical linkages as it is not a familiar way for determining one’s kinship status vis-à-vis another individual (Shapiro, 1982). The KAES program is a work in progress and will be modified to accommodate features of kinship terminologies not encompassed by the terminologies considered to date as additional terminologies are analyzed. The structure used to implement the algebraic modeling is not likely to be changed. However, if there is a terminology with structure that cannot be accom- modated within the structural sequence implemented in the KAES program, then that fact alone will be a striking discovery. The KAES program will also be expanded beyond its cur- rent architecture aimed at analyzing and modeling the structure of kin term maps. The instantiation of the algebraic model with respect to genealogical tracing provides the inter- face needed to embed the structural logic into an agent-based demographic model so as to follow out the logic of kinship terminology structures and their instantiation in terms of a group of individuals undergoing demographic change. Initial work in this direction has high- lighted the way in which implementation of an agent-based demographic model raises ques- tions about the constraints on the transmittal of information of kinship relations within a group to progeny. For example, if progeny learned in totality the set of parental kin relations, then the amount of information about kin being transmitted across generations expands at least exponentially and will outstrip memory capacity. Hence the amount of information transmitted must be limited, thereby raising the issue of how kin relations can be calculated for more distant potential kin and answered through the logic modeled by the KAES pro- gram. Coupling the KAES program with agent-based demographic simulation should make it possible to examine in more detail the consequences of marriage rules in conjunction with terminologies and how marriage rules and kin terminology structure relate to social organization in kin-based societies. The logic of terminologies raises questions about human cognition and why terminologies are highly logic. In part it relates to calculation of kin relations in a manner consistent across all individuals sharing the same terminology. In another part, it raises questions about the evolutionary trajectory that leads from behavior centered around biological kin to behavior centered around cultural kin and whether that trajectory can be accommodated in a Darwin- ian evolutionary framework or whether it signals a change to social organization based on culturally constructed conceptual systems for which the idiom of descent with modification driven by reproductive success is no longer the primary driver for evolutionary change (cf. Read, 2003). References Bennardo, G., & Read, D. (n.d.). The Tongan kinship terminology: Insights from an algebraic analysis. Unpub- lished manuscript. Keesing, R. (1975). Kin groups and social structure. New York: Holt, Rinehart & Winston. Leaf, M. (1971). The Punjabi kinship terminology as a semantic system. American Anthropologist, 73, 545-554. Lee, R. (2000). Dobe ju/‘hoansi. Fort Worth, TX: Harcourt Brace. Lounsbury, F. (1964). The structural analysis of kinship semantics. In H. Hunt (Ed.), Proceedings of the Ninth International Congress of Linguists (pp. 1073-1093). The Hague, the Netherlands: Mouton. 24 Social Science Computer Review Lowie, R. H. (1928). A note on relationship terminologies. American Anthropologist, 30, 263-267. Morgan, L. H. (1870). Systems of consanguinity and affinity of the human family. Washington, DC: Smithsonian Institution. Read, D. W. (1984). An algebraic account of the American kinship terminology. Current Anthropology, 25, 417- 440. Read, D. W. (2001). What is kinship? In R. Feinberg & M. Ottenheimer (Eds.), The cultural analysis of kinship: The legacy of David Schneider and its implications for anthropological relativism (pp. 78-117). Urbana: Uni- versity of Illinois Press. Read, D. W. (2002). Cultural construct + instantiation = constructed reality. Human complex systems (Paper DWR2002). Retrieved MONTH, DAY, YEAR?, from http://repositories.cdlib.org/hcs/ DWR2002 Read, D. W. (2003). From behavior to culture: An assessment of cultural evolution and a new synthesis. Complex- ity, 8(6), 14-41. Read, D. W., & Behrens, C. (1990). KAES: An expert system for the algebraic analysis of kinship terminologies. Journal of Quantitative Anthropology, 2, 353-393. Read, D. W., & Fischer, M. D. (2005). Kinship algebra expert system. Retrieved August 25, 2005, from http:// kaes.anthrosciences.net Sahlins, M. (1962). Moala: Culture and nature on a Fijian island. Englewood Cliffs, NJ: Prentice Hall. Scheffler, H. W. (1972). Kinship semantics. Annual Review of Anthropology, 1, 309-328. Shapiro, W. (1982). The place of cognitive extensionism in the history of anthropological thought. Journal of the Polynesian Society, 91, 257-297. Dwight W. Read, Ph.D., is a professor in the Department of Anthropology and the Department of Statistics at the University of California, Los Angeles. His research interests relate to the use of mathematical modeling as a way to extend and elaborate anthropological theorizing about the structure of human societies and the evolutionary origins of culture. Read / Kinship Algebra Expert System 25 
work_k252y2lh7fabhff6kzenb3gule	----	Developing a Web-Based Advisory Expert System for Implementing Traffic Calming Strategies Research Article Developing a Web-Based Advisory Expert System for Implementing Traffic Calming Strategies Amir Falamarzi,1 Muhamad Nazri Borhan,1 and Riza Atiq O. K. Rahmat2 1 Sustainable Urban Transport Research Centre (SUTRA), Faculty of Engineering, Universiti Kebangsaan Malaysia, 43600 Selangor Darul Ehsan, Malaysia 2 Department of Civil & Structural Engineering, Faculty of Engineering, Universiti Kebangsaan Malaysia, 43600 Selangor Darul Ehsan, Malaysia Correspondence should be addressed to Amir Falamarzi; amir.falamarzi@gmail.com Received 10 April 2014; Accepted 5 August 2014; Published 7 September 2014 Academic Editor: Cai W. Chang-Jian Copyright © 2014 Amir Falamarzi et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Lack of traffic safety has become a serious issue in residential areas. In this paper, a web-based advisory expert system for the purpose of applying traffic calming strategies on residential streets is described because there currently lacks a structured framework for the implementation of such strategies. Developing an expert system can assist and advise engineers for dealing with traffic safety problems. This expert system is developed to fill the gap between the traffic safety experts and people who seek to employ traffic calming strategies including decision makers, engineers, and students. In order to build the expert system, examining sources related to traffic calming studies as well as interviewing with domain experts have been carried out. The system includes above 150 rules and 200 images for different types of measures. The system has three main functions including classifying traffic calming measures, prioritizing traffic calming strategies, and presenting solutions for different traffic safety problems. Verifying, validating processes, and comparing the system with similar works have shown that the system is consistent and acceptable for practical uses. Finally, some recommendations for improving the system are presented. 1. Introduction Nowadays, along with the development of urbanism and an increasing number of vehicles, urban streets, and especially residential streets, suffer from different traffic safety problems [1]. Speeding, thru traffic, and other safety-related problems increase the risk of collisions between pedestrians and vehi- cles. Additionally, excessive numbers of vehicles in residential streets have lowered the quality of environmental factors in residential areas such as air and noise pollution. Furthermore, motorized transportation and lack of infrastructure can affect the perception of other road users toward the function of streets [2]. For example, traffic congestion and narrow side- walks can prevent pedestrians from walking as a means of transportation and, consequently, persuade them to use pri- vate vehicles [3]. The main purposes of traffic calming strategies are to reduce the speed of vehicles and the amount of nonlocal traffic volume entering residential streets [4]. Traffic calming strategies are designed to make streets safe and calm for nonmotorized transportation users including pedestrians, residents, and children [5]. They consist of physical and nonphysical engineering measures. Speed humps, chicanes, and traffic circles are common physical traffic calming mea- sures. Speed limit reduction and installing signs that prohibit turning are examples of nonphysical traffic calming measures [6]. Traffic calming strategies have a great impact on the safety of residential streets, the effect of which, in treated streets is highly dependent on the location of implementations, space between measures, and design considerations [7]. However, it has been proven that the impact of major traffic safety prob- lems such as speeding and thru traffic have been mitigated in residential streets after implementations have been made. Employing traffic calming strategies to deal with safety- related problems in residential streets requires experience and knowledge which can be achieved from traffic calming manuals and experts. Due to a wide range of strategies as well as problems in this field, deriving appropriate solutions Hindawi Publishing Corporation e Scientific World Journal Volume 2014, Article ID 757981, 16 pages http://dx.doi.org/10.1155/2014/757981 http://dx.doi.org/10.1155/2014/757981 2 The Scientific World Journal and effective mechanisms in traffic calming studies is essen- tial. Creating a framework for implementing traffic calming strategies can help both novice and experienced engineers to better recognize problems and accordingly apply appro- priate solutions. Expert systems are an interesting branch of artificial intelligence (AI) that are designed and structured to facilitate the decision making process for nonexperts or novice engineers. Expert systems are computer-based pro- grams which are developed to mimic and imitate problem- solving processes along with the reasoning of human experts in different knowledge fields [8–11]. Expert systems can assist humans in solving problems that require extensive knowledge or huge amounts of time. These systems are also applicable in dealing with problems related to computer science, agriculture, nutrition, medicine, engineering, educa- tion, geology, and so forth. One of their main advantages is that they are easy to access through computer technology [12]. Useful expert systems have already been developed in the field of transportation and safety. USLIMITS2 is a web-based expert system which aims to assist engineers in the selection of safe and appropriate speed limits in speed zones on all American roads [13]. Paver is a knowledge-based expert sys- tem developed for the management, maintenance, and reha- bilitation of pavement. Paver is applicable to military instal- lations, municipalities, airports, researchers in universities, and for the use of consultant companies [14]. COPRBU is a knowledge-based expert system developed to deal with prob- lems relating to public buses with respect to routes and sched- ule, level of service, and their reliability. Wen [15] designed an automatic and dynamic expert system for solving congestion problems at traffic lights. Logi and Ritchie [16] came up with a real-time knowledge-based system for managing and controlling traffic congestion within road networks. Castro et al. [17] developed a fuzzy expert system for forecasting collisions between pedestrians and vehicles in an attempt to avoid accidents. E-ASSIST is an expert system designed to assist engineers and decision makers in implementing TDM strategies. In this expert system, TDM strategies are classified and organised. Hence, users are enabled to find the strategies effectively based on the domain knowledge [18]. It has been proven that web-based expert systems can take on an important role in spreading knowledge among engineers and researchers because they are accessible any- where and at any time, only requiring an internet connection with no installation needed [19]. Furthermore, knowledge or solutions proposed by noncommercial expert systems can be delivered to users without any middlemen [20]. One of the valuable benefits of web-based expert systems is having the potential to be assessed globally through the internet. Also, developers of web-based expert systems are able to monitor the number of visitors and analyze their online feedback. In the following, Section 2 addresses the problem state- ment of the study. Section 3 discusses the importance of the proposed expert system. Section 4 describes the develop- ment process of the expert system including knowledge acquisition, selection of building tool, knowledge repre- sentation, CALMSYS structure, knowledge base, and user interface. Section 5 describes the evaluation of the system which consists of system verification, system validation, and comparison with similar works. Section 6 provides the con- clusion of the study and takes a look at future work which can strengthen the performance of the expert system. 2. Problem Statement Speeding and nonlocal traffic in residential streets are major safety problems in residential streets. Conventional traffic calming manuals normally deal with only speeding and thru traffic, while traffic calming strategies have the ability and potential to handle wider ranges of traffic safety problems in residential streets. Safety parameters such as the condition of nonmotorized transportation users, geometric design, public transportation, lack of infrastructure, special zones, and development factors are all absent in the decision making process of current traffic calming manuals and standards. Furthermore, the classifications of current traffic calming strategies are not well organized because they do not cover all available measures. In this regard, some measures are neglected and some strategies are not developed. 3. The Importance of CALMSYS An advisory expert system can be developed firstly to classify strategies and solutions in traffic calming subject. Strategies can be classified in physical, nonphysical, and combined categories. Policy making, psychological measures, traffic restriction and prevention, infrastructure improvement, and enforcement are suitable strategies which can be categorized in traffic calming studies. More importantly, an advisory expert system has the ability to assist engineers in finding proper traffic calming strategies toward traffic safety prob- lems. In this regard, users and engineers will be helped to employ proper solutions to related safety problems. Hence, this expert system has included all possible traffic safety problems which can be solved by employing traffic calming strategies. Moreover, designing traffic calming measures and also ranking traffic calming projects can be handled better and more successfully by computerized systems than by human experts due to the presence of large numbers of data and complex mathematical equations. 4. Development of CALMSYS In this section, different parts of the traffic calming expert sys- tem are described including knowledge acquisition, knowl- edge base, building tool, and user interface of CALMSYS. 4.1. Knowledge Acquisition. Knowledge for developing a traf- fic calming expert system can be collected and obtained from human expertise in traffic calming as well as written sources of knowledge. There are various sources of manuals, journals, and books existing in the field of traffic calming that can serve as the knowledge core for the proposed expert system. Written sources in traffic calming studies include different subjects and frameworks, but most of them contain descriptions of traffic calming measures, results from the impact of traffic calming measures, and ranking processes of The Scientific World Journal 3 Table 1: List of major sources used in knowledge acquisition. Number Title Year Publisher 1 A Policy on Geometric Design ofHighways and Streets [21] 2011 AASHTO 2 The Handbook of Road SafetyMeasures [22] 2009 Emerald 3 Manual on Uniform TrafficControl Devices 2003 FHWA 4 Traffic Calming: State of thePractice [4] 1999 FHEA 5 TDM encyclopedia [23] 1999 VTP 6 Alaska Traffic Calming Manual 2001 DOWL 7 Pennsylvania Traffic CalmingManual 2001 PDOT 8 Traffic Calming for Bus Routes 2005 TFL 9 International Approaches to Bicycles and Pedestrian Facility Design [24] 2006 FHWA 10 A Guide to Managing TruckTraffic on Local Streets [25] 1985 PVPC projects. Table 1 shows a list of major written sources in this field. The second source of knowledge for developing expert systems is domain experts. Selecting domain experts is one of the most essential parts of expert system development. Domain experts must be knowledgeable and have adequate experience in the field. The depth of experience and the type of experience (theoretical, practical, or a combination of both) must be taken into account when choosing domain experts to help develop traffic calming advisory expert systems [26]. Domain experts with greater experience and having practical experience are preferred, but it must be noted that involving younger experts who have innovative ideas to tackle complex problems can also be useful. In this study, 15 experts in the field of traffic calming and safety have been asked to identify and explain solutions for dealing with traffic safety problems in residential streets. Experts are divided into three groups. The first group includes young experts with 5 to 10 years of experience. The second group includes experts with experience ranging from 10 to 15 years. The third group includes experts with more than 15 years of experience. The average length of experience of all the experts is 15 years. 4.2. Selection of Building Tool. For developing the expert sys- tem, Microsoft Visual Basic.NET was used. The main advan- tage of this version over VB 6.0 is the capability of VB.NET to build a web-based expert system. Generally, Visual Basic software is an easy to learn and flexible language that enables developers to code and create GUIs (graphical user inter- faces) [27]. In a graphical user interface environment, users have icons, pictures, menus, and other useful elements which are not provided in basic expert systems shells. An expert sys- tem integrated with GUIs can make the system more accessi- ble for users who do not have a high level of proficiency [28]. Graphical user interface (GUI) Working memory (containing data and facts) Inference engine (matches facts and data against rules) Input data Output data Knowledge base rules in traffic calming domain Module for traffic calming strategies Module for problems and solutions Module for prioritizing traffic calming strategies Literature (including manuals, books, and article in domain knowledge) Experts in domain knowledge Knowledge engineer- computer programmer End users (transport engineers, students,. . .) Figure 1: The structure of CALMSYS. Furthermore, being simple allows other developers to modify or upgrade expert systems created by Visual Basic software. 4.3. Knowledge Representation. For developing the expert system in this study IF-THEN rules have been employed to represent the knowledge base. In total, more than 150 rules were generated. 100 rules were applied in the first module and the other 50 rules were applied in both the second and third modules. IF-THEN rules are an effective and useful type of forward-chaining inference engine where the decision making process is started from the data entered by users and ended with the attainment of a particular goal. In this method, for example, if the problem (A) is true, then the solution (B) will be recommended to users. 4.4. CALMSYS Structure. Figure 1 shows the structure of CALMSYS which have consisted of the relations between the main component of the expert system including working memory, inference engine, knowledge base, and user inter- face. 4.5. Knowledge Base. In this study, the traffic calming advi- sory expert system is composed of three main modules in- cluding a module for the classification of traffic calming 4 The Scientific World Journal strategies, a module for problems and solutions, and a module for prioritizing traffic calming strategies. The modules of this expert system can work independently but there is collaboration between the module of traffic calming strategies and the module of problems and solutions which have been described in Section 4.6.2. Descriptions related to the main modules of CALMSYS are provided in the following passages. 4.5.1. Module for Traffic Calming Strategies. An important step before finding solutions for traffic calming problems is to develop and explore traffic calming strategies. For this purpose, the main function of current traffic calming measures was gathered from available sources including traffic calming manuals, related books, and journal articles which are discussed in the knowledge acquisition section of this research paper. Furthermore, consulting with domain experts has improved overall knowledge of the subject and brought successful outcomes in traffic calming studies that have been applied in different places. It must be noted that implementation guidelines, advantages, disadvantages, and the design process can all enhance the knowledge base of an expert system. In addition, traffic calming measures must be categorized into specific strategies in order to employ them effectively according to their performance and characteristics. Inter- viewing with domain experts has been proven to help engineers in the field conduct classifications of traffic calming measures. Table 2 shows the classification of traffic calming measures according to eleven categories of strategies. In this module, descriptions of traffic calming strategies, including the measures mentioned in Table 2, have been provided and are available to end-users. In addition to the classification, designs of vertical and horizontal deflections have been included in this module. For example, according to the traffic calming manual published by ITE, there is a direct relation between length of vertical traffic calming measures and their crossing speed (design speed). In chicanes and lateral shifts, path angle has direct relation with the crossing speed of vehicles passing the measures. An example of employing this module is represented in the user interface section. 4.5.2. The Module for Problems and Solutions. There are var- ious safety problems in residential streets that can be treated by employing traffic calming strategies. In this study, elicita- tion of knowledge from both written sources and domain experts was carried out. Firstly, reviewing and examining written sources of the subject field to find traffic safety problems are essential. Although investigating traffic calming manuals and related books can give a general understanding to engineers about the safety problems that pedestrians and other road-users may face while traveling on residential streets, in most traffic calming manuals, capabilities of traffic calming strategies to solve various traffic safety problems are not defined clearly and only descriptions are presented to readers. The next step was to pinpoint traffic safety parameters which can cause safety problems in residential streets. To Table 2: Classification of traffic calming measures. Strategies Measures Vertical deflections Speed bumps, speed humps, speed tables, speed cushion, rumble strips, and raised crosswalks/intersections Horizontal deflections Chicane, lateral shift, central chicane, and traffic circle Narrowing Choker, neck-down, road-diet, sidewalk widening, pedestrian refuge island, hatched marking, turn lane, and median Pavement treatment Brick paving, stone paving, and colored surface Parking management Parking restriction/prohibition, nonparallel parking Volume control Half closures, full closure, diagonaldiverters, and turn prohibition Streetscaping Street furniture, tree planting, andgateway Changes in speed limit School zone, speed limit reduction, and truck speed limit Enforcement Police enforcement, increasedpunishment, and speed cameras Special zones Truck exclusion zone, shared space,pedestrian zone, and school zone Traffic signs Warning signs, regulatory signs, school signs, bicycle signs, pedestrian signs, residential signs, truck signs, special zone signs, and traffic calming signs Improvement of street infrastructure Crosswalk, sidewalk, bike lane, and street lighting Network analysis Changing street direction from one-way to two-way (or vice versa) and changing direction of a one-way street accomplish this, interviewing experts in the domain can be useful because these types of experts generally have a good deal of practical experience in facing these types of traffic safety problems. The duty of safety experts is to find solutions for different traffic safety problems. In this study several interviews with the domain experts were carried out. Experts were asked to express the parameters that they believe have the potential to be included in studies related to traffic calming and safety of residential streets. As a result, a list of the parameters categorized as traffic-related parameters, geometric design parameters, deficiencies in street infras- tructure, and land-use related parameters was compiled. Afterwards, two questionnaires were designed and dis- tributed among the experts. In the first questionnaire, experts were asked to select the importance of each parameter in traffic calming studies, which were elicited through the above methods, according to a 5-point Likert Scale. In this rating scale, 1 means strongly unimportant, 2 means unimportant, 3 means neutral, 4 means important, and 5 means strongly important. Furthermore, in this questionnaire, in the event that experts felt the given parameters could present safety problems and that employing traffic calming is necessary, The Scientific World Journal 5 Table 3: Results from the questionnaire about the importance of the parameters. Category Parameters Group 1 Group 2 Group 3 𝑃value Mean SD Mean SD Mean SD Traffic parameters Speeding in urban streets 3.80 0.45 4.60 0.55 4.40 0.55 0.074 Through traffic 3.60 0.55 4.20 0.84 4.20 0.45 0.262 Accident rate 4.20 0.84 3.80 0.84 4.00 0.71 0.735 Parking occupancy 4.20 0.84 4.40 0.55 4.00 0.70 0.679 Heavy vehicle 4.40 0.55 4.00 0.70 4.00 0.70 0.557 Geometric design parameters Width of streets 4.00 0.70 4.40 0.55 4.40 0.55 0.503 Length of streets 4.20 0.84 4.00 0.70 4.60 0.55 0.420 Grade 4.20 0.45 4.20 0.84 4.20 0.84 1.000 Curves 4.40 0.55 4.00 0.71 4.40 0.55 0.503 Infrastructure parameters Sidewalk 3.80 0.84 4.20 0.84 4.40 0.55 0.462 Bike lane 4.00 0.70 4.20 0.84 4.20 1.10 0.921 Crosswalk 4.20 0.84 4.60 0.55 4.00 1.00 0.516 Intersection 4.20 0.84 3.80 0.84 4.20 0.84 0.691 Street lighting 4.00 0.71 4.60 0.55 4.20 0.84 0.420 Bus stops 4.00 0.00 3.80 0.84 4.40 0.55 0.284 Land-use parameters Density 4.20 0.84 4.40 0.55 4.60 0.55 0.641 School 4.40 0.89 3.60 0.89 4.00 0.70 0.351 Trip generators 4.00 0.70 3.80 0.84 4.00 1.00 0.914 Residential complexes 4.20 0.84 4.40 0.55 4.60 0.55 0.641 they were asked to specify the associated threshold and conditions of the related traffic safety parameters. After collecting questionnaires, the data was analysed by SPSS software. Mean values and standard values were calculated as shown in Table 3. ANOVA was used in order to determine whether the answers of the three groups of experts were significantly different from each other or not. According to the𝑃 value in Table 3, no significant difference between the groups was found. This means that experts with different lengths of relevant experience have similar notions toward the importance of the traffic safety parameters contained in the expert system. Table 4 presents the thresholds of traffic safety parameters. The second questionnaire focuses on the solutions or strategies that experts have applied or recommended for dealing with the mentioned problems. To achieve the pur- pose, a list of traffic calming strategies with their associated measures and details was presented to them. Then, the experts could choose the measures under each strategy to match the safety problems according to their knowledge and experience. Experts could describe solutions for each problem and propose their own ideas and methods to tackle the problems. Finally, three experts with the highest level of experience were selected to evaluate the answers and finalize the solutions for each traffic safety problem. Table 5 summarizes the solutions that the domain experts have proposed. As an example, the process of running this module is illustrated in the user interface section. 4.5.3. Module for Prioritizing the Strategies. Functions of traf- fic calming measures in terms of different criteria including speed reduction, volume reduction, improvement of nonmo- torized transportation, environmental impacts, emergency access, and cost of implementation and maintenance must be compared in order to prioritize them. The analytic hierarchy process (AHP) is a multicriteria decision Making (MCDM) tool developed to make a correct decision with regard to the goal and proposed criteria [29]. The AHP technique has been used in a wide range of fields such as project management, software selection, and marketing [30]. The AHP technique can assist engineers in employing traffic calming strategies when considering and integrating different criteria are required. The first step for prioritizing strategies is to create a hierarchical structure of the model as in Figure 2. The goal is located at the top of the model. The second level contains the criteria in prioritizing strategies while the third level is for alternatives or traffic calming strategies. For developing the AHP technique, domain experts selected for knowledge acquisition and developing the expert system were asked to participate in this study. The Expert Choice (Version 11) software was used to prioritize traffic calming strategies with respect to different criteria. Due to a large number of calculations, using Expert Choice software can facilitate the group decision making process and reduce errors occurred in manual computation. Normalized scores of traffic calming strategies with respect to the different crite- ria and the weight of each criterion which was obtained from 6 The Scientific World Journal Table 4: Conditions of traffic safety parameters. Traffic safety parameters Condition Traffic speed Speed differential more than 8 km/h Through traffic Thru traffic more than 50% of total traffic Accident rate Number of accident resulted from speedingmore than three per year Parking occupancy Parking occupancy more than 50% innarrow street Heavy vehicle Heavy vehicle volume more than 25% oftotal traffic Width of streets More than one on each direction or thewidth of traffic lanes more than 3 meters Length of streets Unimpeded length more than 200 m in localstreets and 300 m in collector street Grade Grade more than 8% Horizontal or vertical curves Sharp curves or Visibility problems Sidewalk No sidewalk or usable shoulders Bike lane No bike lane in street with cycle volumemore than 200 per day Crosswalk Uncontrolled crosswalk with pedestrianvolume more than 200 per day Intersection Uncontrolled intersections with speedingproblems Street lighting No street lighting Bus stops Existence of bus stops Density Residential areas with more than 90dwelling units per hectare School Existence of primary schools orkindergartens Trip generators Existence of more than three major trip generators or commercial streets in 300 m length of a street Residential complexes Lack of safety measures pairwise comparison judgments matrices were calculated and determined. Finally, the composite priority weights of traffic calming strategies were determined as shown in Figure 3. The overall consistency of the model is 0.01 which indicates that the judgments of experts are carried out correctly. In this module end-users are enable to compare traffic calming strategies with respect to the criteria and employ them according to their needs. An example of running this module is provided in the user interface section. 4.6. User Interface. Making expert systems as user-friendly as possible and avoiding creating a complicated design can attract users to utilize expert systems in their fields. In advisory expert systems, users must be able to find their problems without being confused or easily frustrated from difficult procedures. The CALMSYS web-based expert system has provided different and useful functions for its users. Main modules of the system are displayed as toolboxes in the user interface. In addition, toolboxes for ranking traffic calming projects and developing complete streets are provided for assisting end-users. Figure 4 shows the main menu screen- shot. In this section, the function of toolboxes for traffic calming strategies, problems and solutions, and prioritizing strategies are demonstrated with practical examples. 4.6.1. Toolbox for Traffic Calming Strategies. In the toolbox for traffic calming strategies, different traffic calming meas- ures are classified in terms of the strategies that have been previously described. Users are able to select different traffic calming measures according to their purposes by clicking on them. For each measure, a description, typical example, advantages and design considerations are provided. Table 6 shows a screenshot of the toolbox for traffic calming strate- gies. In this toolbox, there are 52 hyperlinks that can serve as a useful source for engineers and students. For example, by clicking on the hyperlink for speed hump design, the designing process for the measure according to the ITE procedure is detailed (Figure 5). Speed humps are common and effective traffic calming measures which have been cate- gorized in vertical deflection strategies. Their effectiveness on speed reduction and reducing thru traffic is proven. Length of speed humps has an important role on their performance. Shorter lengths and greater heights can slow down vehicles drastically but they are not suitable for collector streets or streets with speed limits above 40 km/h. On this page, when users select the speed limit of the target streets from the dropdown list, the expert system inference engine provides the proper length, height, and distance between the measures accordingly in the dedicated textboxes. For example as shown in Figure 5, selecting speed limit of 45 km/h leads to design of a speed hump with height of 10 cm, length of 4.8 m and distance (between measures) of 100 m. In addition to the proposed geometric design, useful design considerations for implementing speed humps are recommended to users. In this regard, implementing speed humps on streets with bus routes is not recommended. Similarly, on streets with a grade of more than 8 percent implementation can cause accidents. 4.6.2. Toolbox for Problems and Solutions. In the toolbox for problems and solutions, a list of 19 traffic safety problems which were described previously are presented to users as shown in Figure 6. On this page users can select the problems they have faced. Key words of problems are highlighted with different colours which can facilitate the process of finding the right problem for users of the expert system. Descriptions of problems can help users and engineers know the problems and their effect on safety of road users. For example, if a street with speeding related problems is selected, users will be directed to the speeding problems page as illustrated in Figure 7 which describes the results and negative impacts of speeding in residential streets on road-users. In this page users will be informed that speeding can affect pedestrian safety and vulnerable road-users in two ways. Firstly speeding can increase a vehicle’s stopping distance exponentially. Secondly, excessive speed and higher The Scientific World Journal 7 Ta bl e 5: So lu tio ns fo rd ea lin g w ith tr affi c sa fe ty pr ob le m s. Pr ob le m s So lu tio ns Ve rt . de f. H or . de f. N ar ro w in g Pa ve m en t tr ea t. Pa rk in g m an ag . Vo lu m e co nt ro l St re et sc ap in g Sp d. lim it ch an g. En fo rc em en t Sp ec ia lz on es Tr affi c si gn s N et w or k an al ys is In fr as t. de ve lo p. Sp ee di ng ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ Th ru Tr . ∗ ∗ ∗ ∗ ∗ A cc id en t ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ Pa rk in g ∗ ∗ ∗ ∗ ∗ ∗ Tr uc ks ∗ ∗ ∗ W id th ∗ ∗ ∗ ∗ Le ng th ∗ ∗ G ra de ∗ ∗ ∗ ∗ ∗ ∗ C ur ve s ∗ ∗ ∗ ∗ ∗ Si de w al k ∗ ∗ ∗ ∗ ∗ ∗ Bi ke la ne ∗ ∗ ∗ ∗ ∗ C ro ss w al k ∗ ∗ ∗ ∗ ∗ ∗ U nc on .i nt er . ∗ ∗ ∗ ∗ ∗ ∗ ∗ St .l ig ht in g ∗ ∗ ∗ ∗ ∗ Bu ss to ps ∗ ∗ ∗ ∗ ∗ ∗ D en si ty ∗ ∗ ∗ ∗ ∗ ∗ ∗ Sc ho ol ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ Tr ip ge n. ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ R es .c om pl ex ∗ ∗ ∗ ∗ ∗ 8 The Scientific World Journal Safety Speed reduction Emergency access Cost Negative env. impacts Improving NMT H or iz on ta l d efl ec tio ns Ve rt ic al d efl ec tio ns N ar ro w in g Pa ve m en t t re at m en t St re et sc ap in g En fo rc em en t In fr as tr uc tu re C ha ng es to sp ee d lim it Tr affi c si gn s Ps yc ho lo gi ca l m ea su re s Ru m bl e st ri ps Volume reduction Figure 2: Hierarchical structure of the model. 0.131 0.118 0.117 0.108 0.103 0.096 0.084 0.071 0.065 0.056 0.053 0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 Vertical deflection Psycholigcal measures Horizontal deflection Narrowing Infrastructure Enforcement Changes to speed limit Traffic signs Streetscaping Rumble strips Pavement treatment Overall inconsistency = 0.01 Figure 3: The composite priority weights of traffic calming strategies. CALMSYS main menu Problems and solutions Traffic calming strategies Ranking projects Residential stress Complete street Prioritizing strategies Figure 4: A screenshot of the CALMSYS main menu. The Scientific World Journal 9 Table 6: A screenshot of the traffic calming strategies page. Strategy Measure Description Vertical deflection Speed bump Making slow points on residential complexes Speed hump Making slow points on roadways Speed table Making slow points on roadways Speed cushion Making slow points on roadways suitable for bus routes Rumble strip Alerting unaware drivers to the changes in traffic condition andenvironment Raised Crosswalk/intersection Enforcing drivers to give way to pedestrians at crosswalks to intersections Horizontal deflection Chicane Making slow points on roadways Lateral shift Making slow points on roadways Center island Chicane Making slow points on roadways Traffic circle Making slow points at intersections Narrowing Choker Reducing the width of roadways at midblock locations Bus bulb Curb extension to improve the safety of passengers Neck-down Reducing the width of roadways at intersections Road diet Reducing the number of lanes and the effective width Sidewalk widening Widening sidewalk to improve pedestrian activities Pedestrian refuge island Improving the safety of pedestrian when crossing wide streets Hatched markings Reducing the effective width of roadways Turn lane Assigning a turn lane in the middle of roadways to reduce effectivewidth of roadways Median Constructing raised median to reduce the effective width ofroadways Pavement treatment Pavement treatment Using alternative materials to alert drivers to changes in trafficcondition and environment Parking management Parking Restriction/prohibition Restricting or prohibiting on-street parking to improve the safety of pedestrians and other vulnerable road-users Nonparallel parking Different types of parking to reduce the effective width of roadways Volume control Full closure Closing streets to thru traffic Half closure Closing one-half of roadways Diagonal diverter Managing streets to discourage thru traffic Turn prohibition Closing streets at particular times to discourage thru traffic Streetscaping Street furniture Enhancing the social aspect of streets Tree planting Enhancing the environmental aspect of streets Gateway Inform drivers about changes in traffic zone Changes to speed limit Speed limit reduction Reducing current speed limit Heavy vehicle speed limit Different speed limits for trucks School zone Reducing speed limit at school times Network Analysis Converting one-way to two-way Reducing effective width of streets and encouraging drivers to lower their speed Changing direction of streets Managing streets to discourage thru traffic impact speed will result in higher severity of injuries. At the bottom of this page, users have to select the type of street (local or collector streets) and, according to their selection, the solutions which can prevent and reduce the impacts of speeding on residents and other road-users will be presented. Figure 8 shows a screenshot of the page related to solutions for speeding problems in residential streets. On the page of solutions for speeding, there are different traffic calming strategies which user can employ to deal with the problem. Traffic signs, pavement treatment, vertical deflections, horizontal deflections, streetscaping, and special zones are useful strategies that can eliminate or reduce prob- lem of speeding in residential streets. It must be mentioned that the toolbox of problems and solutions and the toolbox 10 The Scientific World Journal Speed hump Please select the proper speed limit and then press design button: Speed limit: Height: Length: meter Dist. bet. measures: meter H represents the height of speed humps L represents the length of speed humps D represent the distance between speed humps • Implementation speed humps on streets with grade more than 8% is not recommended • Implementation speed humps on streets which serves as bus or emergency routes is not recommended. Instead of speed humps, speed tables and speed cushion are more suitable Design Problem finding Length Distance 45 km/h 10 cm Speed hump is a common traffic calming measure which has been categorized in vertical deflection measures and has great impact on vehicle speed and discouraging cut through traffic. Speed hump is experienced in the US and around the word successfully. Common speed hump is segment and in series of implementations. 4.2 m and its design speed is 35 km/h. Speed humps must be implemented along the road H 4.8 100 Figure 5: A screenshot of speed hump design. of traffic calming strategies are connected to each other by means of hyperlinks. These hyperlinks have integrated different parts of the system effectively and facilitated the use of the system. As shown in Figure 8, each strategy have one or more measures that by clicking on them users of the system will be directed to the descriptions or design pages related to the toolbox of traffic calming strategies. 4.6.3. Toolbox for Prioritizing the Strategies. In this toolbox, as shown in Figure 9, two panels were designed for the purpose of comparison traffic calming strategies. In the first panel, by employing a dropdown list users are enabled to compare traffic calming strategies with respect to the criteria which have been described in Section 4.5.3. For example in Figure 9, when speed reduction is selected from the dropdown list, a diagram related to the comparison of traffic calming strate- gies with respect to the speed reduction criterion is displayed. According to the information provided, implementing verti- cal deflections is the most effective way to reduce traffic speed in residential streets. In the second panel, two radio button controls are provided. The first button is designed to display the diagram related to the normalized weights of criteria and the next is dedicated to display the diagram related to the composite priority weights of traffic calming strategies. 5. Evaluation of the System Evaluation of expert systems is an important task for system developers [31, 32]. Evaluation of this expert system was conducted through the verification of the system, validation process, and comparison with similar works as follows. 5.1. Verification. Verification of the expert system aims to check that the expert systems works as intended without any errors. Before using expert systems, they must be verified in order to investigate whether the system is consistent and whether or not it is stable [31, 33]. Verification of the system was carried out by questioning two groups of experts. The first group consisted of three computer professionals who are highly skilled in computer science and can give recommendations or comments toward the improvement of the system. The second group consisted of three domain experts with more than 10 years of experience. For verifying the system, two different questionnaires were distributed The Scientific World Journal 11 Please select the problem you have faced in your street from the following list and then press next button: Category Problem Traffic condition Street with cut through traffic (nonlocal land use trips) problem Speeding in urban streets Streets with a large number of accidents resulted from speeding Narrow streets with high parking occupancy Street with excessive truck traffic Geometric design Street with wide roadway or long crossing distance Street with excessive unimpeded block length (uninterrupted length) Street with steep downgrade Streets with sharp curves Infrastructure Streets without bike lane Street without sidewalk Midblock crosswalks Street without lighting Street with bus stops Safety problems at uncontrolled intersections Development Street with high residential density Commercial streets or streets with trip generators Street near a primary school kindergarten Safety problems in residential complexes Next Main menu Figure 6: A screenshot of the problems and solutions page. among them. In these questionnaires, the computer experts and domain experts were asked to rank the parameters in the questionnaires according to the 5 point Likert Scale (1 representing strong disagreement to 5 representing strong agreement). The parameters and results of the verification are represented in Table 7. Results show that the average of answers is higher than 4 (agree and strongly agree) which means that the experts were satisfied. In addition to the verification, the feedback collected from the end-users who analysed the system (such as online users) show that they are satisfied with both importance and performance of the systems, but they also made some useful suggestions for improvement. Thus, in order to increase user acceptance of the system, some modifications and adjust- ments were carried out in the system including: redesigning the menus and the format of web-pages, changing the size and type of fonts, increasing the use of hyperlinks (in order to access subjects faster than before), putting items in tables, and adding more images in different parts of the system that can enhance the learning process. 5.2. Validation. Validation of expert systems aims to confirm that the results derived from the system are compatible with the opinions of domain experts with regard to identical problems and situations. Validation of expert systems can ensure end-users of the reliability and credibility of systems during their decision making process. This process can also help developers or skilled engineers measure the accuracy of their knowledge base. In this study, the validation process was carried out for the three main modules of the expert system. 12 The Scientific World Journal Safety problems related to speeding If you are satisfied with this problem, please select the type of your street and then press solutions button: Local streets Collector streets Solutions Back Speeding is an offence which means driving over the speed limit. Vehicle with excessive speed can increase the risk of collisions and fatalities. In Great Britain, speeding related to more than 1000 deaths and around 35000 serious injuries every year. According to the statistics, speed is an essential contributing factor in more than 30 percent of crashes in the US speeding can affect pedestrian safety and vulnerable road-users in two ways. Firstly, speeding can increase the stopping distance progressively. Secondly, excessive speed and higher impact speed result in higher severity of injuries. it has been investigated that about 90 percent of pedestrian hit by cars driving at 30 km/h will be survived while 20 percent of pedestrians will be survived when hit by cars driving at 50 km/h Figure 7: A screenshot of the speeding problems page. Table 7: Experts’ responses to the system verification questions. Questions for computer professionals Scores 1 2 3 4 5 Av. (1) The user interface is user friendly √√ √ 4.3 (2) The system is easy to use √ √√ 4.67 (3) The system runs commands quickly √ √√ 4.67 (4) The system has no bugs √√ √ 4.3 (5) The system has correct codes √√√ 4 (6) Access to different parts of the system is easy √ √√ 4.67 Questions for domain experts Scores 1 2 3 4 5 Av. (1) The user interface is user friendly √ √√ 4.67 (2) The system is easy to use √√ √ 4.3 (3) The system runs commands quickly √√ √ 4.3 (4) The problems are well defined √ √√ 4.67 (5) The solutions are clear √√ √ 4.3 (6) Strategies are well organized √ √√ 4.67 (7) Measures are well described √ √√ 4.67 (8) The whole system is useful √ √√ 4.67 To perform the validation, three domain experts who are different from the experts who participated in the devel- opment of the expert system were asked to specify their experience and knowledge of the function (the purpose of implementing specific traffic calming measures), suitable place of implementation (local or collector streets), and the design process of the measures (in order to compare a measure designed by the system and a similar measure designed by experts). Also they were asked to prioritize traffic calming strategies according to the criteria which were used in this study. Furthermore, they were requested to express appropriate traffic calming strategies to match the safety problems which exist in the problems and solutions module. Finally, the percentages of answers which were the same as those provided by the system were calculated as shown in Table 8. According to this table, 71% of answers (average of answers) were matched with the output of the system. The result exceeds the level of performance which is expected between 50 and 60%. Therefore, it can be concluded that the system has achieved the goal for which it has been developed. The Scientific World Journal 13 Solutions for speeding in local streets Strategy Measures Purpose Traffic signs Caution sign Warning drivers of pedestrians and residents Speed limit sign Posting appropriate speed limit Residential advisory sign Warning drivers ofentering residential streets Pavement treatment strategy Pavement treatment Warning drivers of changes to traffic zone and alert drivers to vulnerable road-users Vertical deflections Speed hump Making slow points on roadways and alert drivers to the presence of vulnerable road-users Speed cushion Making slow points on roadways and suitable for streets carrying bus routes Horizontal deflections Chicane Making slow points to enforce drivers to slow down Lateral shift Making slow points to enforce drivers to slow down Central chicane Making slow points to enforce drivers to slow down Streetscaping strategy Installation street furniture Aesthetic approach to highlight the residential aspect of streets and discourage drivers from speeding Gateway Promoting the identity and increase the awareness of drivers toward residential activities Tree planting An aesthetic approach to highlight the residential aspect of streets and discourage drivers from There are different types of strategies that can be applied to mitigate the effects of speeding on vulnerable road-users in residential streets. Depending on the speed differential, deflection and nondeflection measures are applicable. If speed differential offensive drivers to slow down is required and implementing deflection measures can be useful. If speed differential is lower than 8 pavement treatment, and streetscaping can encourage drivers to slow down. Furthermore home zone scheme, which is considered as a psychological strategy, can be implemented to provide a calm place for residents and persuade drivers to reduce their speed. Useful strategies for dealing with speeding related problems are provided as follows: is greater than 8 km/h then Back Problem finding km/h then nondeflection measures such as traffic signs, enforcing Special zone Home zone Providing a safe place for pedestrians, cyclists, children, and residents and reduce the dominance of cars Network analysis Preventing offending drivers from dangerously overtaking Converting one- way to two-way Figure 8: A screenshot of the solutions for speeding page. 5.3. Comparison with Similar Works. There are various ex- pert systems in the field of transportation engineering but specifically in the field of traffic calming, there was the lack of development. Most expert systems are related to pave- ment engineering, transportation management, and intel- ligent transportation systems (ITS). As mentioned in the introduction section, current expert systems can deal with limited aspects of traffic calming studies. For example the function of USLIMITS and E-ASSIST can be classified into both traffic safety and transportation management, but they are unable to provide detailed recommendations for imple- menting, designing or assessing traffic calming measures. In this regard, USLIMITS can only propose speed limit in urban streets. E-ASSISST can provide general recommendations toward managing transportation demand and improving nonmotorized transportation. Helping engineers to design traffic calming measures, recommending the suitable strate- gies to deal with traffic safety problems, and enabling users 14 The Scientific World Journal Prioritizing traffic calming strategies using the AHP technique comparison of traffic calming strategies with respect to: Please select one of the following options and press show button: Normalized weights of criteria Composite priority weights of traffic calming strategies 0.032 0.042 0.043 0.047 0.063 0.08 0.087 0.091 0.154 0.17 0.191 0.00 0.05 0.10 0.15 0.20 Traffic signs Streetscaping Pavement treatment Changes to speed limit Rumble strips Infrastructure Psychological measures Narrowing Enforcement Horizontal deflections Vertical deflections 0.078 0.082 0.127 0.202 0.208 0.304 0.00 0.10 0.20 0.30 Cost Emergecny access Environmental impacts Improvement of NMT Volume reduction Speed redection Show Main menu Speed reduction Inconsistency = 0.01 Inconsistency = 0.01 Figure 9: A screenshot of the toolbox for prioritizing traffic calming strategies. Table 8: Summary of percentage of similarity between evaluators’ responses and those of the system. Items Similarity percentage Function of TC measures 75% Suitable places for TC measures 85% Design process of TC measures 65% Prioritization of TC strategies 60% Solutions for safety problems 70% Average 71% to compare traffic calming strategies are the benefits of CALMSYS compared to other relates systems. 6. Conclusion In this study, a web-based advisory expert system for imple- menting traffic calming strategies in residential streets was built, verified, and validated to assist end-users including transportation engineers, safety consultants, and students. The knowledge base of this expert system includes three use- ful modules; each one is capable of assisting users separately according to their needs or problems they have encountered. Using VB.NET software for building the expert system has made it globally accessible. Evaluation of the system has shown that the system is reliable and functional which can encourage end-users to employ it in their traffic calming deci- sion making processes. Feedback from reviewers includes some minor suggestions for improvement, but most of them expressed that working with the system has boosted their The Scientific World Journal 15 skills and creativity toward solving traffic safety problems. In order to satisfy users, some improvements as well as changes in style and format of the expert system web-pages were carried out. Furthermore, the following topic may be useful and will be added to the next version for improving the effectiveness and efficiency of the system. (i) With regard to uncertainty and vagueness in some topics of traffic calming studies such as the process of determining speed limit, distance between traf- fic calming measures, and ranking traffic calming projects, using Fuzzy logic is useful and can handle these type of problems. (ii) In addition to the evaluation process carried out in this research, conducting usability analysis can be worthwhile. The purpose of usability analysis is to measure the efficiency (taking less time to achieve results), the ease of use, and learnability of the system (the capability of an application to enable end-users to learn how to work with the application). (iii) Developing traffic calming strategies to enhance the performance of the system including adding new measures to the system, modifying current measures, and updating designing process. (iv) Incorporating other languages such as Persian, Bahasa Melayu, and Turkish to the system to attract more users. (v) Integrating the expert system with databases such as Microsoft SQL in order to store street data such as safety problems and implemented strategies. Conflict of Interests The authors declare that there is no conflict of interests regarding the publication of this paper. References [1] O. F. O. R. Economic, Speed Management, 2006. [2] A. Granà and M. Guerrieri, “Exploring effects of area-wide traffic calming measures on urban road sustainable safety,” Journal of Sustainable Development, vol. 3, no. 4, pp. 38–49, 2010. [3] N. Yiannakoulias and D. M. Scott, “The effects of local and non- local traffic on child pedestrian safety: a spatial displacement of risk,” Social Science & Medicine, vol. 80, pp. 96–104, 2013. [4] E. Reid, Traffic Calming State of the Practice, Institute of Transportation Engineers, Washington, DC, USA, 1999. [5] W. D. Cottrell, N. Kim, P. T. Martin, and H. J. Perrin Jr., “Effectiveness of traffic management in Salt Lake City, Utah,” Journal of Safety Research, vol. 37, no. 1, pp. 27–41, 2006. [6] DOWL, Traffic Calming Protocol Manual, Anchorage, 2001. [7] M. Pau and S. Angius, “Do speed bumps really decrease traffic speed? An Italian experience,” Accident Analysis and Prevention, vol. 33, no. 5, pp. 585–597, 2001. [8] S. Liao, “Expert system methodologies and applications—a decade review from 1995 to 2004,” Expert Systems with Applica- tions, vol. 28, no. 1, pp. 93–103, 2005. [9] A. Beerel, Expert Systems: Strategic Implications and Applica- tions, 1987. [10] F. Hayes-Roth, D. Waterman, and D. Lenat, “An overview of expert systems,” in Building Expert Systems, 1983. [11] K. Laudon and J. Laudon, Essentials of Management Information Systems, 2011. [12] A. Bianchini, “Fuzzy representation of pavement condition for efficient pavement management,” Computer-Aided Civil and Infrastructure Engineering, vol. 27, no. 8, pp. 608–619, 2012. [13] R. Srinivasan, D. Harkey, and D. Tharpe, Development of a Web- Based Expert System for Setting Speed Limits in Speed Zones, Transportation Research Board, 2008. [14] N. Ismail, A. Ismail, and R. Atiq, “An overview of expert sys- tems in pavement management,” European Journal of Scientific Research, vol. 30, no. 1, pp. 99–111, 2009. [15] W. Wen, “A dynamic and automatic traffic light control expert system for solving the road congestion problem,” Expert Systems with Applications, vol. 34, no. 4, pp. 2370–2381, 2008. [16] F. Logi and S. G. Ritchie, “Development and evaluation of a knowledge-based system for traffic congestion management and control,” Transportation Research C: Emerging Technologies, vol. 9, no. 6, pp. 433–459, 2001. [17] J. L. Castro, M. Delgado, J. Medina, and M. D. Ruiz-Lozano, “An expert fuzzy system for predicting object collisions. Its application for avoiding pedestrian accidents,” Expert Systems with Applications, vol. 38, no. 1, pp. 486–494, 2011. [18] R. Mansyur, Development of an Expert Advisory System for Implementing TDM Strategies, Universiti Kebangsaan Malaysia, Bangi, Malaysia, 2011. [19] H. M. S. Lababidi and C. G. J. Baker, “Web-based expert system for food dryer selection,” Computers and Chemical Engineering, vol. 27, no. 7, pp. 997–1009, 2003. [20] Y. Duan, J. S. Edwards, and M. X. Xu, “Web-based expert sys- tems: benefits and challenges,” Information and Management, vol. 42, no. 6, pp. 799–811, 2005. [21] AASHTO, A Policy on Geometric Design of Highways and Streets, American Association of State Highway and Transportation Officials, Washington, DC, USA, 2011. [22] R. Elvik, T. Vaa, A. Erke, and M. Sorensen, The Handbook of Road Safety Measures, 2009. [23] T. Litman, Traffic Calming: Benefits, Costs and Equity Impacts, 1999. [24] FHWA, “International Approaches to Bicycle and Pedestrian Facility Design,” 2006. [25] P. M. Lebeaux and M. Connor, A Guide to Managing Truck Traffic on Local Streets, Pioneer Valley, 1985. [26] C. C. Osuagwu and E. C. Okafor, “Framework for eliciting knowledge for a medical laboratory diagnostic expert system,” Expert Systems with Applications, vol. 37, no. 7, pp. 5009–5016, 2010. [27] Z. Jin, F. Sieker, S. Bandermann, and H. Sieker, “Development of a GIS-based expert system for on-site storm-water manage- ment,” Water Practice and Technology, vol. 1, no. 1, pp. 1–8, 2006. [28] A. M. Mosa, An Expert System to Control Construction Problems in Rigid Pavements, University Kebangsaan Malaysia, 2013. [29] T. L. Saaty, “How to make a decision: the analytic hierarchy process,” European Journal of Operational Research, vol. 48, no. 1, pp. 9–26, 1990. [30] E. W. T. Ngai, “Selection of web sites for online advertising using the AHP,” Information and Management, vol. 40, no. 4, pp. 233– 242, 2003. 16 The Scientific World Journal [31] R. M. Aguilar, V. Muñoz, M. Noda, A. Bruno, and L. Moreno, “Verification and validation of an intelligent tutorial system,” Expert Systems with Applications, vol. 35, no. 3, pp. 677–685, 2008. [32] E. Mosqueira-Rey and V. Moret-Bonillo, “Validation of intelli- gent systems: a critical study and a tool,” Expert Systems with Applications, vol. 18, no. 1, pp. 1–16, 2000. [33] Y. Qian, M. Zheng, X. Li, and L. Lin, “Implementation of knowledge maintenance modules in an expert system for fault diagnosis of chemical process operation,” Expert Systems with Applications, vol. 28, no. 2, pp. 249–257, 2005. 
work_k2oirynn6bebfocf5z5ht2j4pa	----	Dynamic Rescheduling Expert System for Hybrid Flow Shop with Random Disturbance Procedia Engineering 15 ( 2011 ) 3921 – 3925 1877-7058 © 2011 Published by Elsevier Ltd. doi: 10.1016/j.proeng.2011.08.734 Available online at www.sciencedirect.com Dynamic Rescheduling Expert System for Hybrid Flow Shop with Random Disturbance Jing Yin a, Tieke Lib ,Baojiang Chen a , Bailin Wangb a* aDepartment of electromechanical and vehicle engineering Beijing University of Civil Engineering and Architecture, Beijing, 100044, China bSchool of Economics and Management University of Science and Technology Beijing, Beijing, 100083, China Abstract For the problem of hybrid flow shop with random disturbance, a dynamic rescheduling expert system is established in the paper. As the core of the system, the rescheduling solver is detailed described and the heuristic algorithm is proposed on the base of random disturbance modeling. The system is developed on the platform of G2. The simulation and computational results given at last verified that the system proposed in the paper can well meet the requirements of both function and performance. © 2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of [CEIS 2011] Keywords: Dynamic rescheduling; Hybrid Flow Shop ; expert system, time repair 1. Introduction The key of dynamic rescheduling is to obtain the scheduling solution quickly that can satisfy the production constraints and minimize the production cost when workshop condition change. The key of dynamic rescheduling is to obtain the scheduling solution quickly that can satisfy the production constraints and minimize the production cost when workshop condition change. Usually, for the problem of dynamic rescheduling, the way is respectively to generate a totally new solution and on-line local * Corresponding author. Tel.: +086-010-68325943; fax: +086-010-68322514 E-mail address: yinjing@bucea.edu.cn. Open access under CC BY-NC-ND license. Open access under CC BY-NC-ND license. CORE Metadata, citation and similar papers at core.ac.uk Provided by Elsevier - Publisher Connector https://core.ac.uk/display/81986590?utm_source=pdf&utm_medium=banner&utm_campaign=pdf-decoration-v1 http://creativecommons.org/licenses/by-nc-nd/3.0/ http://creativecommons.org/licenses/by-nc-nd/3.0/ 3922 Jing Yin et al. / Procedia Engineering 15 ( 2011 ) 3921 – 3925 modify the initial scheduling solution. Because of the practical use and its difficulty, the dynamic scheduling method has always been the focus of the academic research and enterprise application [1]. A great deal of efforts has been made to the rescheduling problems in recent years. Vieira et al. describe a framework of rescheduling system [2] and present an analytical model to predict the performance of one-machine systems under periodic and event-driven rescheduling strategies [3]. Guo et al. study the properties of a three-machine flow shop scheduling problem and present a trigger mechanism to instruct the rescheduling [4]. References [5]-[6] address rescheduling problem on single machine, parallel machines, flow shops and job shops, respectively. Though variations of existing methods improve the efficiency of dynamic rescheduling from different angles of view, the most difficulty point still lies in the contradiction between the real-time and stability of solution performance and the corresponding contradiction between the real-time ability and complexity of computation. Focusing on these difficulties, this paper studies the hybrid flow shop scheduling (HFS) with the random disturbance which typically exists in the manufacture practice, such as machine failures, arrival of rush orders, material shortages, tool breakages and so on. The remainder of this paper is organized as follows: In next section, the dynamic rescheduling solver embedded in the expert system is introduced. Then, two heuristic algorithms for disturbance of processing abnormal and machine breakdown are proposed. The realization of dynamic rescheduling expert based on G2 and computational results are illustrated next after. We conclude at last. Nomenclature l iO the No. i operation of job lJ , 1, 2,...,l n l ip the normal processing time of l iO 'lip the abnormal processing time of l iO l ist the start time of l iO l iet the completion time of l iO l iSlack the slackness of l iO breakt the time of machine breakdown covre ert the time of machine repaired mcap the remaining capacity of machine jM , j = 1,2,...,m kjb the No. k breakdown of jM hjro the processing time for emergency rush order h on the jM jx the processing time change of jM 2. Design of dynamic rescheduling solver 2.1. The Semantic Network As the core of the whole system, dynamic rescheduling solver consists of rules and processes and describes the problem of HFS with random disturbance. As shown in the Fig.1, there are six class definitions in the conceptual model, including schedule solution, job, operation, machine, constraint and 3923 Jing Yin et al. / Procedia Engineering 15 ( 2011 ) 3921 – 3925 random disturbance. On this basis, the knowledge network can be structured. The underline in the Fig.1 represents the key attributes of the class and the semantic network inference is realized by correlation, focus and multiple inheritances between different classes provided by G2. Fig.1 The semantic network of dynamic rescheduling solver 2.2. The Solving Procedure The solving procedure of dynamic rescheduling: 1. Initialization: Create entity { }1 2 nJ = J , J ,..., J ; input the initial scheduling solution 0S ; Constraint check; initialize attribute value. 2. Real-time monitoring: Executing scheduling; monitoring the shop state; identifying the disturbance. 3. Choose proper program: If the disturbance belongs to processing abnormal, then call the heuristic algorithm of time parameter repair (TPR); if the disturbance belongs to machine breakdown or emergency rush order, then call the heuristic algorithm based on the remaining capacity of machine (RCM); else transfer to human-computer interaction (dynamic editor). 4. Constraint Check: If rescheduling solution is infeasible, then call the conflicts resolution program 5. Output the rescheduling solution. Schedule solution job-num machine-1-num machine-1-start-time machine-1-process-time machine-2-num Reschedule Operation job-num operation-num machine-num machine-start-time machine-process-time machine-end-time opwork-lst-i opwork-lst-j opwork-ff Job job-num job-status job-start-time job-process-time Machine operation-num machine-num machine-status job-num start-time process-time job-sequence Random diturbance event-describe happen-time job-num operation-num machine-num anti-last-time min-gather doing-gather Time-Delay Machine-Break conflict-gather Constraint job-num time-constraint machine-constraint process-constraint 1 1 1 1, 2, 3 n 1 m n 1 1 1 1 Pre-Schedule Graphic illustration Focus Correlation inheritance 3924 Jing Yin et al. / Procedia Engineering 15 ( 2011 ) 3921 – 3925 Considering the production requirement for both the real-time ability of rescheduling efficiency and stability of rescheduling quality, the strategy of local repair is adopted. We take the advantage of the powerful real-time capability and concurrent multi-thread capacity of G2 which is developed by the Gensym Corporation, USA. The solving procedure of dynamic rescheduling is shown as follows: According to different kinds of disruption, different heuristic algorithm is dynamically exploited. The heuristic algorithm of time parameter repair (TPR): 1. Calculate the minimum operation set as follows: 1 ( || ) ( ), l l l k k l k l l i i i j j i j i iA O O O A O O O st > et O A (1) 2. For liO A ,Calculate the l iSlack as follows: 1min , l l l l k l l i i i i j i iSlack st p st st p st (2) 3. If l ' l li i iSlack p - p , then Calculate the start time as follows: For 1 l iO , 1 1 1 1' max , ' l l l l l i i i i ist st st et st (3) For kjO , 1 1' max , ' l k k l l i j j i ist st st et st (4) 4. Constrain propagation and check. { }liA = A - O , if A = , then stop; else, go to step 2. The heuristic algorithm based on the remaining capacity of machine (RCM): 1. Calculate the minimum operation set as follows: cov cov l l l l l i i break i re er i break i re erA O st t st t et t et t (5) 2. For liO A , calculate the mcap of the other normal machine available as follows: ( ) max min , max , , 0k j l l k l k l m i i j i j iM O m cap et st et et st st (6) 3. Assign the machine of max mcap to process l iO , { } l iA = A - O , if A = , then go to step 4 ; else go to step 2. 4. Constrain propagation and check. 5. If conflict exists, TPR adopted to modify the time parameters of related rest operation; else stop. 3. Experimental result Table 1 The computation result of RCM A Initialsolution Rescheduling solution Constrain propagation and check 9 1O CF-1 CF-1 Confict exists in 91O and 13 1O on CF-1, 35f 5 1O CF-3 CF-3 13 1O CF-2 CF-1 3925 Jing Yin et al. / Procedia Engineering 15 ( 2011 ) 3921 – 3925 Random Create an entity of machine breakdown. Assign the attribute value as follows: event- describe=breakdown, job-num=13, operation-num=1, machine-num=2, happen-time = 125, anti-last- time=35. The rescheduling result is shown in table 1 and in table 2. Table 2 The computation result of TPR A Operation Rescheduling solution Operation Rescheduling solution 13 1O 14 1O 13 2O 15 1O 14 2O 13 3O 14 3O 12 1O 15 2O 6 2O 15 3O 12 2O 16 2O 6 3O 16 3O 12 3O 18 2O 17 3O 18 3O 13 1O 13 1 160st 14 3O 14 3 265st 14 1O 14 1 195st 6 2O 6 2 255st 13 2O 13 2 195st 6 3O 6 3 300st 13 3O 13 3 225st 15 3O 15 3 305st 15 1O 15 1 230st 16 3O 16 3 345st 14 2O 14 2 230st 17 3O 17 3 375st 15 2O 15 2 265st 18 3O 18 3 405st 4. Summary When scheduling solution executed in the workshop, dynamic disturbance often unpredictable occurs, such as task change, processing abnormal, machine breakdown and etc. These make rescheduling attach comprehensive attention to the solving mechanism, considering both the real-time ability and solution quality. Focus on this problem, the dynamic rescheduling solver is designed and developed on the platform-G2 in the paper, including the semantic network, the solving procedure and the heuristic algorithm of TPR and RCM. The simulation and computational results show that the system is efficient and effective. References [1] Peter Cowling, Djamila Oueelhadj, Sanja Petrovic. A multi-agent architecture for dynamic scheduling of steel hot rolling. Intelligent Manufacturing, no. 14, p. 457- 470, 2003. [2] G.E. Vieira, J.W. Herrmann, and E. Lin, Rescheduling manufacturing systems: a framework of strategies, policies, and methods, Journal of Scheduling, vol. 6, no. 1, pp. 35-58, 2003. [3] G.E. Vieira, J.W. Herrmann, and E. Lin, Analytical Models to Predict the Performance of a Single- machine System under Periodic and Event-driven Rescheduling Strategies, International Journal of Production Research, vol. 38, no. 8, p. 1899-1915, 2000. [4] B. Guo, Y. Nonaka, Rescheduling and optimization of schedules considering machine failures, International Journal of Production Economics, p. 503–513, 1997. [5] M.S. Akturk, E. Gorgulu, Match-up scheduling under a machine breakdown, European Journal of Operational Research, vol. 112, p. 81-97, 1999. [6] Guo Dong-fen, Li Tie-ke, Constraint Satisfaction-based Method for Steelmaking-Continuous Casting Production Scheduling Problem, Information & control, vol. 34, no.6, p. 753-758, 2005. 
work_k36ubeddt5fuvoo5aum6ofmeuq	----	Expert Systems With Applications. Vol. 2, pp. 47-58, 1991 0957-4174/91 $3.00 + .00 Printed in the USA. V 1991 Pergamon Press Plc Rule-Based Training of Neural Networks STAN C. KWASNY Center for Intelligent Computer Systems,' Department of Computer Science , Washington University, St. Louis, MO KANAAN A. FAISAL King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia Abstract-Rule-based expert systems either develop out of the direct involvement of a concerned expert or through the enormous efforts of intermediaries called knowledge engineers . In either case, knowledge engineering tools are inadequate in many ways to support the complex problem of expert system building. This article describes a set of experiments with adaptive neural networks which explore two types of learning, deductive and inductive, in the context of a rule-based, deterministic parser of Natural Language. Rule-based processing of Language is an important and complex domain. Experiences gained in this domain generalize to other rule-based domains. We report on those ex- periences and draw some general conclusions that are relevant to knowledge engineering activities and maintenance of rule-based systems. 1. INTRODUCTION IN THE FIELD of Artificial Intelligence (Al), one of the primary goals is to build intelligent systems . Since the 1970s, many such systems have been most easily built symbolically as rule-based systems. In many applica- tion areas, the most popular of these are known as rule-based expert systems. These systems have become ubiquitous in their application to everything from medical diagnosis to oil exploration to stock market trading. Typically, these systems develop from con- sultations with recognized experts who provide knowl- edge and experience during the process of rule devel- opment and debugging. Most expert systems use rules in some capacity. Usually, the rule formalism becomes the de facto "pro- gramming language" for the application and the rules Requests for reprints should be sent to Stan C . Kwasny, Center for Intelligent Computer Systems , Department of Computer Science, Washington University, Campus Box 1045, 1 Brookings Drive, St. Louis, MO 63130. S.C.W. gratefully acknowledges the support of King Fahd Uni- versity of Petroleum and Minerals. We thank the reviewers for their comments which have improved aspects of our presentation of this work. The authors also express gratitude to William E . Ball, Georg Dorffner, Marius Fourakis , David Harker, Dan Kimura , Barry Kal- man, Ron Loui, John Merrill, Gadi Pinkas , and Robert Port for thoughtful discussions and comments concerning this work. • The sponsors of the Center are McDonnell Douglas Corporation, Southwestern Bell Telephone Company, and Mitsubishi Electronics America, Inc. are executed like statements in the language . Building such systems is no trivial task, often occupying the full- time efforts of dozens of people . Much of the effort is spent talking with human experts , gaining their per- spective in an application domain , and teasing out the salient information that allows the expert to perform his task with a high degree of skill . Once some knowl- edge is acquired , it becomes the basis for system or- ganization and rules that mimic the expertise . However, acquisition of such knowledge is an ongoing concern even as refinements are made during maintenance. The task is not a simple one. Even with the most cooperative experts and best tools , rules are often dif- ficult to formulate . Ad hoc mechanisms that relate premise to conclusion have been invented to capture some of the vagueness and uncertainty of the relation- ships articulated by the expert. Certainty factors, for example , were invented for this purpose . Extensive ev- idence showing the difficulty of knowledge acquisition can be found in a recent special issue of SIGART (Westphal & McGraw, 1989). The task of Natural Language Understanding , in fact , has been perceived as a task of knowledge engineering ( Shapiro & Neal, 1982). Once formulated , rules are difficult to debug and maintain . In XSEL/XCON , for example, it is reported (Barker & O'Connor , 1989) that 40% of the rules re- quire changes each year to adapt to new products and for other reasons. Systems like TEIRESIAS (Davis, 1982) have been specifically designed to aid in the cri- 47 48 tiquing and reformulation of rule-based systems by the experts themselves. However, the expert is still required to operate at the level of rules, which may be very un- natural in many domains. Rules, no matter how so- phisticated, do not always make the best programming language. Even with better tools, once performance reaches a certain point it may prove difficult to achieve an im- proved level of performance. Upon analysis, some of the difficulty may be attributable to design decisions made early in the construction of the system which lose their validity as the system evolves. Such problems often require changes to the basic design and complete retooling of the system. Traditionally, rules also play an important role in the processing of Natural Language. In fact, the com- parison between rule-based expert systems and rule- based Natural Language systems is close indeed, es- pecially in examining some of the difficulties encoun- tered. Many linguistic rule systems have been proposed, although the rules in these systems may not always appear explicitly as in rule-based expert systems. Sym- bolic rules tend to be an important vehicle of specifi- cation for the symbolic processing requirements of both types of systems. The goal for expert systems is to symbolically de- scribe a domain in such a manner that relevant deci- sionmaking can occur and that a computer system can perform at the level of expert in this regard. But, is this always a realistic goal? While progress in Natural Lan- guage Processing (NLP), for example, continues to be made, why does this continue to be only a partially solved problem? Or, more specifically, why is there no operationally complete set of symbolic rules for English (i.e., no set of rules that perfectly describes all under- standable English utterances)? Certainly, it is neither for lack of experts nor lack of effort. Perhaps, if we agree with Winograd and Flores (1987, p. 107), the answer is that ". . . computers cannot understand language ." We do not believe computer systems are so impaired. The problems raised for rule-based expert systems as well as those mentioned for rule-based NLP could be attacked through learning. Learning would seem to be a solution, but traditional learning approaches too often require even deeper insights and understanding of the problem domain than the expert himself may possess. While strictly symbolic learning may be pos- sible, such systems typically need to invent new sym- bols to extend their capabilities symbolically and this imposes limitations on the approach. If learning could be conducted in a sufficiently flex- ible manner, a trainable system could be taught the rules articulated by the expert and then be led to adapt to those cases where the rules fail. In a similar fashion, the competence rules of a grammar can be taught in concert with the performance examples of NLP. This S. C Kwasny and K. A. Faisal distinction between easily articulated rules and the more difficult ones is analogous to the distinction be- tween textbook and experiential learning. The new in- tern, for example, is filled with book learning, but lacks the wisdom and knowledge of the domain which only comes through experience. Connectionist (neural) net- works hold promise for providing such a flexible learn- ing scheme. This article presents results from experimentation with a connectionist parsing system. The system is simply viewed as an example of a complex rule-based system. Conclusions reached in these experiments, therefore, have consequences for many other types of rule-based systems. 2. RULE-BASED CONNECTIONIST PARSING Natural Language processing is both symbolic and subsymbolic. It is symbolic in the role symbols play in writing systems and in the chunking of concepts, while it is subsymbolic because of the fuzziness of concepts, and the apparent high degree of parallelism in the ac- tivity. In building parsers, the subsymbolic aspect of language is usually lost, although there have been at- tempts to construct parsers that are totally subsymbolic (see Fanty, 1985; Selman & Hirst, 1985; Waltz & Pol- lack, 1985). Rules usually play a sacred role in parsing systems and, as mentioned earlier, are often executed as if fol- lowing instructions in a program. Whether we are only interested in building robust expert systems, or are in- terested in modeling human expertise, this method is incorrect. Rules should be permitted to play an advisory role only-that is, for guidance in typical situations and not as prescriptions for precise processing. In the case of English, if a complete set of rules for all meaningful English forms existed, then it might be satisfactory to rely on a rule-based approach. But no such set of rules exists, nor does it seem desirable or even possible to construct such a set. Any rule-based system that is based literally on rules tends to be brittle since there is no direct way to process inputs that are not anticipated to some extent. Furthermore, the ac- quisition of new rules often requires tedious retuning of existing rules. The only solution to these problems in a practical and realistic manner is through learning. We have developed a connectionist deterministic parsing system called CDP which offers solutions to these problems (Kwasny & Faisal, 1989). Training can be conducted either from rules or from examples of processing. The resultant network is then tested on a variety of novel sentence patterns and its generalization capabilities studied. A set of experiments are presented which support the claim that Natural Language can be syntactically processed in a robust manner using a connectionist deterministic parser. The model is trained based on a II Rule-Based Training of Neural Networks set of deterministic grammar rules and tested with sen- tences which are grammatical and ones that are not. Tests are also conducted with sentences containing lexically ambiguous items. 2.1. Deterministic Parsing For a complete understanding of the motivation and architecture of CDP, it is necessary to describe the work on which it is based. For a readable description of deterministic, "wait-and-see" parsing, see Winston (1984), or for a more thorough discussion, see Marcus (1980). The determinism hypothesis restricts Natural Lan- guage Processing to a deterministic mechanism. It states that Natural Language can be parsed by a mechanism that op- erates `strictly deterministically' in that it does not simulate a nondeterministic machine . . . (Marcus, 1980, p. 11) It follows from this hypothesis that NLP need not de- pend on backtracking, nor are any partial structures produced during parsing which fail to become part of the final structure. This is equivalent to prohibiting chains of reasoning in a rule-based expert system which ultimately do not contribute to the final answer. Ob- viously, processing is restricted in a major way under this assumption. PARSIFAL (Marcus, 1980) is a demonstration of a a deterministic, rule-based parser of Natural Language. Extensions to this system have been proposed for processing ungrammatical sentences [PARAGRAM (Charniak, 1983)], for resolving lexical ambiguities [ROBIE (Milne, 1986)], and for acquiring syntactic rules from examples [LPARSIFAL (Berwick, 1985)]. We have found it beneficial to combine these four tasks into one implementation which is partly symbolic and partly connectionist. The connectionist approach has particular advantageous in unifying these four sys- tems (Kwasny, Faisal, & Ball, in press). The system architecture of PARSIFAL has been reconfigured and the behavior of the rules and other mechanisms from these systems are being simulated using a neural net- work simulator. Learning in the network is achieved through backward propagation discovered indepen- dently by Werbos (1974) and Rumelhart, Hinton, and Williams (1986). As illustrated in Figure 1, the primary components of a deterministic parser are a buffer for lookahead in the sentence, a stack for processing embedded struc- tures, and a collection of rules for controlling the building and movement of constituents of the sentence being processed. Processing occurs essentially left-to- right. The absence of backtracking is an important ad- vantage in developing a connectionist-based parser since structures, once built, are never discarded. 49 Stack Buffer FIGURE 1 . Deterministic Parsing in Schematic Form. Rules are partitioned into rule sets . A standard rec- ognize-act cycle is used to achieve a parse . A consequent of a rule may be the activation or deactivation of a rule set thereby providing a simple conflict resolution strategy . Conflicts within rule sets are resolved from the static ordering (i.e., numeric priority) of the rules. Actions can effect changes to both the stack and the buffer. As buffer positions are vacated at the far end, new sentence components flow in sequentially. If a successful parse is found, a termination rule will fire leaving the final structure on top of the stack. 2.2. Connectionist Deterministic Parsing CDP provides a setting for experimentation with a va- riety of grammars and network designs. It combines the basic mechanism of deterministic parsing with an adaptive, feed-forward neural network to enable the generalization and robustness of connectionism to be evaluated in this domain. A backpropagation neural network simulator, which features a logistic function that computes values in the range of - I to + 1, is being used in this work. The ultimate goal is to construct a mechanism capable of learning to deal syntactically with language in a robust manner. As Gallant (1988) points out, there are important advantages to constructing rule-based systems using neural networks. Our focus is on building a connect tionist parser, but with more general issues in mind. How successfully can a connectionist parser be con- structed and what are the advantages? Success clearly hinges on the careful selection of training sequences. Our experiments have examined two different ap- proaches and compared them (Faisal & Kwasny, in press). The "deductive" strategy uses rule "templates" de- rived from the rules of a deterministic grammar. It is deductive in the sense that it is based on rules that are general (in the sense that they must be applicable in a wide variety of processing situations), but specific sen- tence forms must be processed. The "inductive" strat- egy derives its training sequence from coded examples 50 of sentence processing . It is inductive in the sense that it is based on specific sentence examples , although a potentially wider variety of sentence forms must be processed. The goal of both deductive and inductive learning is to produce a network capable of mimicking the rules or sentences on which its training is based and to do so in a way that generalizes to many addi- tional cases. Once initial learning has been accom- plished, simulation experiments can be performed to examine the generalization capabilities of the resulting networks. In our implementation , a moderate-sized grammar developed from PARSIFAL is used for training. The entire set of grammar rules is contained in Appendix A. The grammar used in CDP is capable of processing a variety of sentence forms which end with a final punctuation mark. Simple declarative sentences, yes- no questions, imperative sentences , and passives are permitted by the grammar. The model actually receives as input a canonical representation of each word in the sentence in a form that could be produced by a simple lexicon . Such a lexicon is not part of the model in its present form. Experiments are conducted to determine the effec- tiveness of training and to investigate whether the con- nectionist network generalizes properly to ungram- matical and lexically ambiguous cases. In comparison to the other deterministic parsing systems, CDP per- forms favorably. Virtually all of our examples have been drawn from previous work. Much of the perfor- mance depends on the extent and nature of the training, of course, but our results show that through proper training a connectionist network can indeed exhibit the same behavioral effect as the rules . Furthermore, once trained, the network is efficient , both in terms of representation and execution. Deductive training generally performs well on all generalization tasks and outperforms inductive training by scoring generally higher on all experiments. Reasons for this include the specificity of the inductive training data as well as the lack of a large amount of training data in the inductive case required to provide sufficient variety. 3. LEARNING A RULE -BASED GRAMMAR A deterministic parser applies rules to a stack and buffer of constituents to generate and perform actions on those structures. One of its primary features, as men- tioned earlier, is that it does not backtrack, but proceeds forward in its processing never building structures which are later discarded. Training of CDP proceeds by presenting patterns to the network and teaching it to respond with an appro- priate action using backpropagation. The input patterns represent encodings of the buffer positions and the top of the stack from the deterministic parser. The output S. C. Kwasny and K A . Faisal of the network contains a series of units representing actions to be performed during processing and judged in a winner-take-all fashion . Network convergence is observed once the network can achieve a perfect score on the training patterns themselves and the error mea- sure has decreased to an acceptable level (set as a pa- rameter). All weights in the network are initialized to random values between -0.3 and +0 . 3. Once the net- work is trained , the weights are saved so that various experiments can be performed. A sentence is parsed by iteratively presenting the network with coded inputs and performing the action specified by the network. Each sentence receives a score representing the average strength of responses during processing. The closer the sentence matches the training patterns, the lower the error and the greater the strength . Strengths are used for comparison purposes only. Strengths are computed as the reciprocal of the error. The strength for each individual step is simply the re- ciprocal of the error for that step . The step error is computed as the Euclidean distance between the actual output and an idealized output consisting of a -1 value for every output unit except the winning unit which has a +I value. The errors for each step are summed and averaged over the number of steps . The average strength is the reciprocal of the average error per step. As mentioned earlier, there are two distinct ap- proaches to training a network to parse sentences. Each of these training strategies results in a slightly different version of CDP. The differences in the derivation of the two types of training patterns are illustrated in Fig- ure 2. Deductive training begins with deterministic grammar rules which are coded into rule templates, one rule template representing one grammar rule. In- stantiation of a rule template leads to a training pattern which is presented during learning . Coding and in- stantiation are discussed below. Inductive training is based on traces of sentence processing itself. The coded training patterns derived in this way have in some sense r----------------t ----------- ----- Im'miWUic Grammes Rules U Coded Rule Templates U cgd Pa Training Panama ---------------- Sentence Traces U Ceded Training Pastes. I. J Deductive Training Inductive Training FIGURE 2. Extraction of Deductive and Inductive Training Pat- tems. Rule-Based Training of Neural Networks already been instantiated and, therefore, are suitable for learning with no further translation. 3.1. Deductive Learning Each grammar rule is coded as a training template which is a list of feature values, but templates are not grouped into rule packets as in PARSIFAL. Each con- stituent in the rule is represented by an ordered feature vector of +1 (on), -1 (off), or ? (do not care). Instan- tiation of the vector occurs by randomly changing ? to +1 or -1. Each template, therefore, can be instantiated into many patterns making each epoch of training slightly different. During training, the network learns the inputs which are highly correlated with expected outputs and those that are not. Training is arranged so that ? values are uncorrelated with the outputs from those training patterns. Each rule template containing n ?'s can generate up to 2" unique training cases. Some rule templates have up to 30 ?s which means they represent approximately 109 training cases. It is obviously impossible to test the performance of all these cases, so a zero is substituted for each ? in the rule template to provide testing pat- terns. While in actual processing a zero activation level for a feature will never be encountered, zero is a good test since it represents the mean of the range of values seen during training. Grammar rules are coded into rule templates by concatenating the feature vectors of the component constituents from the stack and buffer . Each grammar rule takes the following form: {<Stack>< Ist Item)(2nd Item ) (3rd item) -. Action} For example, a rule for Yes/No questions would be written: {(S node)("have")(NP)(VERB, -en) -0 Switch Ist and 2nd items} while a rule for imperative sentences would be written: {(S node)("have" )(NP)(VERB, inf) o Insert YOU) By replacing each constituent with its coding, a rule template is created. In the two rules above, rule tem- plates are created with a ?value for many of the specific verb features of the initial form "have" in each rule, but are carefully coded for the differences in the third buffer position where the primary differences lie. Be- cause different actions are required, these are also coded to have different teaching values during training. 51 Appendix A contains the grammar rules used as a basis for all deductive training experiments in this study. Our rules were derived from the grammar con- tained in Appendix C of Marcus (1980) which includes those rules specifically discussed in building a case for deterministic parsing. They can be taken as represen- tative of the mechanisms involved. To assure good performance by the network, training has ranged from 50,000 to 200,000 presentations cycling through train- ing cases generated from the rule templates. Once training is complete, the parser that uses the network correctly parses those sentences that the original rules could parse. 3.2. Inductive Learning Inductive training depends on training patterns derived from traces of processing of actual sentences. This pro- cessing is guided by application of the rules of a deter- ministic grammar as before. This form of training requires that the network demonstrate the correct rule-following behavior after training runs with a com- paratively small sample of sentence traces. Although no symbolic rules are learned, the behavior of the rules is captured within the weights of the network. Fur- thermore, the behavior is guaranteed to approximate the behavior required in the sample training sentences as closely as desired, depending on the convergence rate and the quantity of training employed. The primary difference between the two forms of learning is seen if we consider the space of possible patterns. With deductive training, that space is system- atically presented during learning in such a way that each major distinction to be made during processing is represented in each epoch. Inductive training hap- pens in a less systematic way with no guarantee of ap- propriate representation of cases. Thus, deductive training imposes an ordering on the training patterns that assures a completeness which is difficult to achieve with inductive training, but inductive training patterns reflect the frequency of rule occurrences seen in pro- cessing the actual samples of sentences. A small set of positive sentence examples were traced which resulted in 64 unique training patterns. These were used for all inductive experiments in this study. 3.3. Architecture of CDP As Figure 3 illustrates, CDP is organized into a sym- bolic component and a subsymbolic component. Ac- tions in CDP are performed symbolically on traditional data structures which are also maintained symbolically. The subsymbolic component is implemented as a nu- meric simulation of an adaptive neural network. The symbolic and numeric components cooperate in a tightly coupled manner since there are proven advan- tages to this type of organization (Kitzmiller & Ko- 52 Sub- Symbolic Symbolic Input Stream II Parse Structure FIGURE 3. CDP System Organization. walk, 1987). For CDP, the advantages are performance and robustness. It is the responsibility of the symbolic component to handle the input sentence coding it for presentation to the network . In the subsymbolic component, the network produces a designation of an action to be per- formed by producing an activation pattern across out- put units . Activation of output units is interpreted in a winner-take-all manner, with the highest activated unit determining the action to be taken . Actions them- selves are performed symbolically on conventional data structures. The whole process is very efficient in time and space, although learning itself occurs off-line and is a time-consuming process. In the set of experiments described here, the network has a three-layer architecture as illustrated in Figure 4, with 35 input units, 20 hidden units, and 20 output units. Each input pattern consists of three feature vec- tors from the buffer items and one stack vector. Each vector activates 14 input units in a pattern vector rep- resenting a word or constituent of the sentence. The stack vector activates seven units representing the cur- rent node on the stack . In our simplified version of the grammar, only two items are coded from the buffer and thus 35 input units are sufficient . One hidden layer has proven sufficient in all of these experiments. The output layer represents the 20 possible actions that can be performed on each iteration of processing. What the model actually sees as input during sen- tence processing is not the raw sentence but a canonical representation of each word in the sentence in a form that could be produced by a simple lexicon , although such a lexicon is not part of the model in its present form . Iteration over an input stream is performed by moving unprocessed sentence forms into the buffer as vacancies are created . Iteration ends when the buffer becomes empty and a stop action is requested by the network. At this point , it is instructive to follow an example with more details revealed as shown in Figure 5. When a sentence form like S. C. Kwasny and K A. Faisal appears in the input stream , the first three constituents fill the buffer. Note that in reality this is an ungram- matical sentence form . Later, it will be shown that CDP can correctly produce the structure shown . The con- tents of the three elements along with the contents of the top of the stack are coded into a feature vector and presented to the network producing a single action. The action is executed potentially producing changes to the buffer and stack . When processing stops, the final structure can be seen on the top of the stack, just as for PARSIFAL. 4. PERFORMANCE CDP is capable of successfully processing a variety of simple sentence forms such as simple declarative, pas- sive, and imperative sentences as well as yes-no ques- tions. For test and comparison between the inductive and deductive CDPs, several sentences are coded that would parse correctly by the rules of the deterministic parser . Also, several mildly ungrammatical and lexical ambiguous sentences are coded to determine if the net- work generalizes in any useful way. The objective is to test if syntactic context could aid in resolving such sit- uations. 4.1. Parsing Grammatical Sentences Experimentation with grammatical sentences dem- onstrates the ability of CDP to perform as PARSIFAL. Earlier we mentioned that convergence testing from the rule templates is possible by changing each ? to a zero value. Here we examine the performance of CDP with actual sentences. Grammatical sentences, by our definition, are those which parse correctly in the rule-based grammar from which we derived the training set. Table I shows several examples of grammatical sentences which are parsed successfully along with their response strengths in both deductive and inductive learning. Strengths are computed as the reciprocal of the av- erage error per processing step for each sentence. The error on each step is determined by taking the Euclid- ean distance between the actual vector of output unit activation values and an idealized vector with only the LAYERS: Output Cm un+u) Hidden (20 .mw) coded Actions ---------------------- l 4 \ Out (35 uniu) ------- ------------ r------oooooeoooooooo ;;occoeocmoooooo;; ooooooo; First K&; SxvM bnttt+ Suck 'John have should scheduled the meeting. FIGURE 4. Subsymbohc Component. Rule-Based Training of Neural Networks SUB-SYMBOLIC SYMBOLIC Buffer mhxi h,.^ should scheduled the meeting. Stack AUX 53 whcdulul ^ __ __ L - - - - - - - - - - - - - - - - J FIGURE S . CDP System Overview. winning unit turned on ( one comer of the hypercube in error space ). Strengths reflect the certainty with which actions for building structures are being selected, but should be used for relative comparisons only. Each example shows a relatively high average strength value, indicating that the training data has been learned well. Also, the deductive average strength value is higher , in almost all cases, than the corre- sponding inductive average strength . Although com- parisons are difficult to make due to variations in the number of unique training patterns and other factors, the deductively trained network exhibits uniformly more definitive decisions than the inductively trained network. Most of these sentences come from examples used in the deterministic parsing systems described earlier. Parse trees are developed which are identical with ones produced by those systems . Sentences ( 8)-(11), which contain ambiguous words, are presented to CDP un- ambiguously , but the lexical choices are provided in parentheses. Capabilities described thus far have only duplicated what is easily done symbolically . Of course , the feed- forward network does support very fast decisionmaking due to the feed-forward nature of the model. But what other features does the model possess? Importantly, how robust is the processing? 4.2. Lexical Ambiguity ROBIE extends PARSIFAL to address issues of lexical ambiguity. It requires additional rules and lexical fea- TABLE 1 Grammatical Sentences Used In Testing Sentence Form Deductive Avg. Strength Inductive Avg. Strength (1) John should have scheduled the meeting . 283.3 84.7 (2) John has scheduled the meeting for Monday. 179.3 84.2 (3) Has John scheduled the meeting ? 132.2 64.4 (4) John is sceduling the meeting . 294.4 83.5 (5) The boy did hit Jack. 298.2 76.2 (6) Schedule the meeting . 236.2 67.8 (7) Mary is kissed. 276.1 84.9 (8) Tom hit(v) Mary. 485.0 80.3 (9) Tom will(aux) hit(v) Mary. 547.5 78.7 (10) They can(v) fish(np). 485.0 80.3 (11) They can(aux) fish(v). 598.2 76.8 54 tures to handle these properly . In the deterministic ap- proach , it is essential that lexical items be properly dis- ambiguated to permit processing to proceed without backtracking. In a set of experiments with CDP, the parser is tested for this. Normal sentences are presented , except that selected words were coded ambiguously (here indicated by angle brackets < ) around the word) to represent an ambiguously stored word from the lexicon . Selected sentences are shown in Table 2. The numbers again indicate the average strength for each sentence. In the cases shown , the lexically ambiguous words are cor- rectly interpreted and reasonable structures result. CDP utilizes syntactic context to resolve these am- biguities . Again, the generalization capability of the network automatically works to relate novel situations to its training cases . For lexically ambiguous situations, some inputs may contain features which confuse its identity as expected by the parser . The context provided by the buffer and stack of the deterministic parser has proven to be sufficient to aid in resolving many am- biguities . An important fact is that, as before, no ad- ditional rules or mechanisms are required to provide this capability. For example, sentence ( 12) presents can ambigu- ously as an auxiliary modal , and main verb, white fish is presented uniquely as an NP. Can is processed as the main verb of the sentence and resulted in the same structure as sentence (10) of Table 1. This is shown in Figure 7 . Here, each word is presented unambiguously with can coded as a verb and fish coded as an NP. The same structure results in each case, with the average strength level much higher in the unambiguous case. Likewise, sentence ( 13), by coding fish ambiguously as a verb/NP and coding can uniquely as an auxiliary verb, produced the same structure as sentence (11). This is the structure in Figure 8. Sentence ( 14) contains the word will coded ambig- uously as an NP and an auxiliary modal verb. In the context of the sentence , it is clearly used as a modal auxiliary and the parser treats it that way. A similar result is obtained for sentence ( 15). In sentence (16), hit is coded to be ambiguous between an NP (as in playing cards) and a verb. The network correctly iden- tifies it as the main verb of the sentence. TABLE 2 Lexically Ambiguous Sentences Used in Testing Sentence Form Deductive Avg. Strength Inductive Avg. Strength (12) They <can) fish . 4.5 2.6 (13) They can fish ). 172.2 4.9 (14) <Will) he go? 83.6 14.3 (15) Tom will) hit Mary. 118.7 19.9 (16) Tom hit) Mary . 39.0 2.5 AUX \John have uld S. C. Kwasny and K. A. Faisal N scheduled the meeting yP\ FIGURE 6. Parse of "John have should scheduled the meet- .. In the cases shown, the lexically ambiguous words are disambiguated and reasonable structures resulted. Note that the overall average strengths are lower than comparable grammatical sentences discussed, as ex- pected . Also, the deductive average strength value is higher than inductive average strength. 43. Parsing Ungrammatical Sentences PARAGRAM extends PARSIFAL to handle ungram- matical sentences. This is accomplished by considering all rules in parallel and scoring each test performed on the left-hand side of a rule according to predefined weights. The rule with the best score fires. In this way, processing always has some rule to fire. Reported ex- perimentation with PARAGRAM shows this to be an effective method for stretching the inherent capabilities of the grammar. CDP attempts to accomplish this through general- ization. An important test of these capabilities is its responses to novel sentences. These are strictly depen- dent upon its training experiences since in deductive training no relaxation rules (Kwasny & Sondheimer, 1983), meta-rules (Weischedel & Sondheimer, 1983), or other special mechanisms are added to the original grammar rules to handle ungrammatical cases. Like- wise, in inductive training no ungrammatical sentences are used. Experiments consist of testing a few ungram- matical , but meaningful ; sentences that are close to the training data and within the scope of our encoding. For our purposes, most of the examples are derived from PARAGRAM and the parsing systems mentioned earlier. Selected ungrammatical sentences are properly pro- cessed as a direct result of the generalization properties yP MVB /jJP can fish FIGURE 7. Parse of "They can (r) fishtnp)." Rule-Based Training of Neural Networks TABLE 3 Ungrammatical Sentences Used in Testing Sentence Form Deductive Avg. Strength Inductive Avg. Strength ( 17) *John have should scheduled the meeting. 25.1 6.6 (18) Has John schedule the meeting? 38.1 18.2 ( 19) *John is schedule the meeting. 4.7 4.9t (20) *The boy did hitting Jack. 26.6 7.5; of learning. When presented an ungrammatical sen- tence, (i.e., one for which the rules of the grammar would fail to find a parse), the network automatically relates the situations arising during parsing to similar situations on which it has been trained . The tendency is for it to select the closest situation . Very often, this generates precisely the response required to relate the deviant form to one that is not. In Table 3, selected ungrammatical sentences are shown . These examples produce reasonable structures when presented to our system . They are shown in the table with their response strengths . Note that overall average strength is lower for ungrammatical sentences (in Table 3) when compared to similar grammatical ones (in Table 1). Sentence (17) is similar to sentence (1), except that the word order has been changed in a simple way. Since sequencing of items in the buffer is a seemingly crucial part of the model, it was felt that this would test the model significantly. Surprisingly, the structure pro- duced is identical to that produced while parsing sen- tence (1), but with lower strengths. The only difference is that the two auxiliary verbs, have and should, are reversed as shown in Figure 6. Sentence ( 18) contains a disagreement between the auxiliary has and the main verb schedule. This example investigates whether predictions made while processing the auxiliary can be only partially fulfilled and still lead to a successful parse . In this case, the comparable grammatical sentence (3) parses identically. Strengths are again lower in both deductive and inductive cases. Sentence ( 19) can be compared with sentence (4). In the deductive case, a structure similar to that built for sentence (4) is indeed constructed. However, in the /Ne They AUX I can VP MVB fish FIGURE S. Parse of "They can(aux) flsh(v)." 55 inductive case (marked with t) the network attempts to process 'is' as if it were indicating the past tense. Although this is incorrect for this sentence, it is not an unreasonable choice . On closer examination, our in- ductive training patterns seem to reflect passives more thoroughly than competing alternatives . This is the only case in which inductive exceeds deductive strength . In this case , two dissimilar processing se- quences are followed and, therefore, the two results are not comparable. Sentence (20) can be compared with sentence (5), but there is not one clear choice in how the sentence should appear if grammatical. The deductive-trained network processes sentence (20) as sentence (5), while the inductive result (marked with t) shows the sentence processed as if it were progressive tense ('The boy is hitting Jack'). In PARAGRAM , a nonsensical parse structure is produced for sentence (20), as reported by Charniak (1983). 4.4. Discussion of Results Using syntax-based approaches to resolving ungram- matical sentences has limitations as seen above due to the ambiguities of such sentences. These problems are well-known as discussed in Kwasny (1980). Impor- tantly, CDP is easily influenced to make one choice or another as a matter of competition. This is a funda- mental property of CDP as argued elsewhere (Kwasny & Faisal, 1989). While deductive training exhibits better perfor- mance than inductive training for all tasks, there are trade-offs in the two approaches. Deductive training requires rules as the basis for rule templates while in- ductive training requires a large amount of data to be successful. Fortunately there is a middle ground which allows mixtures of the two training strategies . Training can be performed using rule templates as well as patterns based on sentence traces . In a recent set of experiments in which the two types of training data were combined, the network was capable of generalizing in ways similar to deductive learning, but also showed particularly good performance on the specific cases reflected in the in- ductive data. 56 5. CONCLUSIONS What does this mean for rule-based expert systems? Where knowledge naturally exists in rule form and such rules can be reliably stated, rule templates can be formed which generate appropriate training cases. However, where knowledge only exists in the form of anecdotal cases, it can be expressed in the form of in- ductive training patterns. As new cases are discovered for which the rules do not apply, inductive data can be easily constructed and the network retrained. Con- trast this with the typical rule-based expert system in which each new rule may require rethinking an entire set of existing rules. Our work has shown some of the trade-offs between deductive and inductive learning. Both have a place in the construction of a neural network designed to per- form complex rule-based tasks such as parsing. De- ductive learning can be compared to the idealized form of learning one would typically get from textbooks, while the analogy to inductive learning is that which comes through practical experiences. How well do these ideas for a connectionist parsing system scale up to a larger grammar? The prospects for successful scaling up are very good. A grammar with more than 70 rules has been constructed and net- work convergence has been observed. The rules rep- resent some important new capabilities, including the processing of embedded sentences and processing per- formed by attention shifting rules in PARSIFAL. These new results are being evaluated and will be presented in a forthcoming PhD thesis ( Faisal , 1989). The potential for exploiting the notion of extensional programming is also very good. The dependency on inductive learning will be greatest in those domains which are difficult to analyze. These are exactly the domains in which standard expert system techniques break down. We have begun to examine an approach to NLP based almost exclusively on inductive training sequences (Kwasny & Faisal , 1989). This looks prom- ising, but is a topic for further research. REFERENCES Barker, V.E. & O'Connor, D.E. (1989). Expert systems for configu- ration at digital: XCON and beyond. Communications of the ACM, 32(3), 298-318. Berwick, R.C. (1985). The acquisition of syntactic knowledge. Cam- bridge, MA: MIT Press Charniak , E. (1983 ). A parser with something for everyone. In M. King (ed.), Parsing natural language (pp. 117-150 ). New York: Academic Press. Davis, R. (1982). Teiresias: Applications of meta-level knowledge. In R. Davis and D . B. Lenat (eds.), Knowledge-based systems in artificial intelligence. New York: McGraw-Hill. S. C. Kwasny and K A . Faisal Faisal, K.A. (1989, August). CDP. Connectionist deterministic parser, a dissertation proposal . Technical Report WUCS-89-33, Depart- ment of Computer Science, Washington University, St. Louis, MO. Faisal, K.A. & Kwasny, S.C. (1990, in press). Deductive and inductive learning in a connectionist deterministic parser. In Proceedings ofthe International Joint Conference on Neural Networks. Wash- ington, DC. Fanty, M. (1985). Context-free parsing in connectionist networks, (Technical Report 174). Computer Science Department, Uni- versity of Rochester, Rochester, NY, 1985. Gallant, S.I. (1988). Connectionist expert systems. Communications of the ACM, 31(2), 152-169. Kitzmiller, C.T. & Kowalik, J.S. (1987). Coupling symbolic and nu- meric computing in knowledge-based systems , AI Magazine, 8(2), 85-90. Kwasny, S.C. (1980, November). Treatment of ungrammatical and extra-grammatical phenomena in natural language understanding systems, Indiana University Linguistics Club, Bloomington, IN. Kwasny, S.C. & Faisal, K.A. (1989, November). Connectionism and determinism in a rule-based natural language system. (Technical Report WUCS-89-31). Department of Computer Science, Wash- ington University, St. Louis, MO. Kwasny , S.C. & Faisal, K.A. (1989, August) Competition and learning in a connectionist deterministic parser. In Proceedings of the 11th Annual Conference of the Cognitive Science Society (pp. 690- 697). Hillsdale, NJ: Erlbaum. Kwasny, S.C., Faisal, K.A., & Ball, WE. (1990, in press). Unifying several natural language systems in a connectionist deterministic parser. In Proceedings of the Western Multi Conference, San Diego, CA. Kwasny, S.C. & Sondheimer, N.K. (1981). Relaxation techniques for parsing ill-formed input . American Journal of Computational Linguistics, 7(2), 99-108. Marcus, M.P. (1980). A theory of syntactic recognition for natural language . Cambridge, MA: MIT Press. Milne, R. ( 1986). Resolving lexical ambiguity in a deterministic parser, Computational Linguistics, 12(1), 1-12. Rumelhart, D.E., Hinton, G., & Williams, R.J. (1986). Learning in- ternal representations by error propagation . In D.E. Rumelhart and J.L. McClelland (eds.), Parallel distributed processing (pp. 318-364) Cambridge, MA: MIT Press. Selman, B. & Hirst, G. (1985). A rule-based connectionist parsing system , In Proceedings ofthe 7th Annual Conference of The Cog- nitive Science Society (pp. 212-221). Hillsdale, NJ: Erlbaum. Shapiro, S.C. & Neal, J.G. (1982, June). A knowledge engineering approach to natural language understanding . In Proceedings of the 20th Annual Meeting of the Association for Computational Linguistics, 136-144. Waltz, D.L. & Pollack, J.B. (1985). Massively parallel parsing: A strongly interactive model of natural language interpretation. Cognitive Science, 9, 51-74. Weischedel, R.M. & Sondheimer , N.K. (1983). Meta-rules as a basis for processing ill-formed input , American Journal of Computa- tional Linguistics, 9(3-4), 161-177. Werbos, P. (1974). Beyond regression: New tools for prediction and analysis in behavioral science, (PhD Thesis). Harvard University, Cambridge, MA. Westphal, C.R. & McGraw, K.L. (1989, April). SIGART-Special issue on knowledge acquisition , New York: ACM Press. Winograd, T. & Flores, F. (1987). Understanding computers and cog- nition (p. 107). Reading , MA: Addison-Wesley. Winston, P.H. (1984). Artificial intelligence . Reading, MA: Addison- Wesley, 1984. t Rule-Based Training of Neural Networks 57 APPENDIX A Grammar Rules 1. Rule Initial if stack is empty then Crest a new S node 2. Rule Yes-No-Q If current node is S no attachments first item is an auxiliary verb second item is NP node then Switch 3. Rule Imperative If then 4. Rule Parse-Subj if then current node is S no attachments first item is a infinitive verb Insert YOU current node is S no attachments first item is NP node second item is a verb Attach the first item as NP 5. Rule Start-Aux if current node is S first item is a verb AUX node is not attached then Create AUX node 6. Rule Aux-Attach if current node is S first item is AUX node then Attach the first item as AUX 7. Rule Aux- If current node is AUX Perfective first item is auxiliary of root HAVE second is past participle verb then Attach the first item as PERF. 8. Rule Aux- If current node is AUX Progressive first item is auxiliary of root BE second is present participle verb then Attach the first item as PROG. 9. Rule Aux- If current node is AUX Passive first item is auxiliary of root BE second is past participle verb then Attach the first item PASSIVE 10. Rule Aux-Modal If current node is AUX first item is auxiliary of type Modal second is infinitive verb then Attach the first item as MODAL 11. Rule Aux-Do if current node is AUX first item is auxiliary of root DO second is infinitive verb then Attach the first item as DO 12. Rule Aux- If current node is AUX Complete first item is not auxiliary then Drop the current node into the buffer 13. Rule MVB 1 If current node is S first item is a verb AUX node is attached then Create VP node 14. Rule MVB 2 If current node is VP first item is a verb then Attach the first item as MVB 15. Rule Passive I If current node is VP auxiliary of the S node is passive then Create NP 16. Rule Passive 2 If the current node is NP then Drop the current node into the buffer 17. Rule Objects If current node is VP first item is NP node then Attach the first item as NP 18. Rule PP-Under- If current node is VP VP first item is PP node then Attach the first item as PP 19. Rule VP-Done I If current node is VP first item is FinalPunct. then Drop the current node into the buffer 20. Rule VP-Done 2 If current node is S first item is VP node then Attach the first item VP 21. Rule S-Done I If current node is S first item is Final Punct. then Attach the first item as FIN. 22. Rule S-Done 2 If current node is S buffer is empty then Stop, reporting a successful parse 58 S. C.Kwasny and K. A. Faisal APPENDIX B Grammar Rule Encoding 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Stack Nodes S - + + + + + - - - - - - + - - - - - - + + + NP - - - - - - - - - - - - - - - + - - - - - VP - - - - - - - - - - - - - + + + + + - - - AUX - - - - - + + + + + + - - - - - - - - - - No-attach + + + + AUX-PASSIVE - - - - - - - - - - - - - - + + ? ? ? ? ? ? AUX-ATTACHED - - - - - - - - - - - - + + + + + + + + + + 1st BUFFER NP ? - - + - VP ? - - - - PP ? - - - - AUX ? - - - - VERB ? + + - + Inf ? - + - ? ? ? ? ? ? Ing ? - - - ? ? ? ? 7 ? En ? - - - ? auxVerb ? + ? - ? Be ? ? ? - ? Have ? ? ? - ? Do ? ? ? - ? + ? ? ? ? 7 Modal ? ? - - ? + - ? - - ? ? FinalPunct ? - - - - 2nd BUFFER NP ? + ? - ? ? _ _ _ _ _ ? ? ? ? ? ? ? _ _ - VP ? - ? - ? ? - - - - - ? ? ? ? ? ? ? - - - - PP ? - ? - ? ? - - - - - ? ? ? ? ? ? ? - - - - AUX ? - ? - ? ? - - - - - ? ? ? ? ? ? ? - - - - VERB ? - ? + ? ? + + + + + ? ? ? ? ? ? ? - - - - Inf ? - ? ? ? ? - - - + + ? ? ? ? ? ? ? - - - - Ing ? - ? ? ? ? - + - - - 7 ? ? ? ? ? ? - - - - En ? - ? ? ? ? + - + - - ? ? ? ? ? ? ? - - - - auxVerb ? - ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? - - - - Be ? - ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? - - - - Have ? - ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? - - - - Do 7 - ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? - - - - Modal ? - ? ? ? ? - - - - - ? ? ? ? ? ? ? - - - - FinalPunct ? - ? - ? ? - - - - - ? ? ? ? ? ? ? - + - - Actions Attach Subj-NP - - - + - - - - - - - - - - - - - - - - - VP - - - - - - - - - - - - - - - - - - - + - AUX - - - - - + - - - - - - - - - - - - - - - FinalPunct - - - - - - - - - - - - - - - - - - - - + - MVB - - - - - - - - - - - - - + - - - - - - - Obj-NP - - - - - - - - - - - - - - - - + - - - - PP - - - - - - - - - - - - - - - - - + - - - - AU-Pert - - _ _ _ _ + _ _ _ _ _ _ _ AUX-Prog - - - - - - - + - - - - - - - - - - - - - AUX-Pass - - - - - - - - + - - - - - - - - - - - - - AUX-Modal - _ _ _ _ _ _ _ _ + _ _ _ _ AUX-DO - - - - - - - - - - + - - - - - - - - - - Create S + - - - - - - - - - - - - - - - - - - - - - NP - - - - - - - - - - - - - - + - - - - - - VP - - - - - - - - - - - - + - - - - - - - - AUX _ _ + Insert-You - - + - - - - - - - - - - - - - - - - - - - Switch - + - - - - - - - - - - - - - - - - - - - - Drop _ _ + + + Stop - - - - - - - - - - - - - - - - - - - - - + i page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 
work_k3tmdyj3vzekrjwk53daausqkm	----	On the Performance of Multi-GPU-Based Expert Systems for Acoustic Localization Involving Massive Microphone Arrays Jose A. Bellocha,∗, Alberto Gonzaleza, Antonio M. Vidalb, Maximo Cobosc aInstitute of Telecommunications and Multimedia Applications, Universitat Politècnica de València, Camino de Vera s/n, 46022 Valencia, Spain. bDSIC department, Universitat Politècnica de València, Camino de Vera s/n, 46022, Valencia, Spain. cComputer Science Department, Universitat de València,Poligon de la Coma s/n, 46100, Valencia, Spain. Abstract Sound source localization is an important topic in expert systems involving mi- crophone arrays, such as automatic camera steering systems, human-machine interaction, video gaming or audio surveillance. The Steered Response Power with Phase Transform (SRP-PHAT) algorithm is a well-known approach for sound source localization due to its robust performance in noisy and reverberant environments. This algorithm analyzes the sound power captured by an acoustic beamformer on a defined spatial grid, estimating the source location as the point that maximizes the output power. Since localization accuracy can be improved by using high-resolution spatial grids and a high number of microphones, accu- rate acoustic localization systems require high computational power. Graphics Processing Units (GPUs) are highly parallel programmable co-processors that provide massive computation when the needed operations are properly paral- lelized. Emerging GPUs offer multiple parallelism levels; however, properly managing their computational resources becomes a very challenging task. In fact, management issues become even more difficult when multiple GPUs are ∗Corresponding author: Phone Number +34-655436190 Email addresses: jobelrod@iteam.upv.es (Jose A. Belloch), agonzal@dcom.upv.es (Alberto Gonzalez), avidal@dsic.upv.es (Antonio M. Vidal), maximo.cobos@uv.es (Maximo Cobos) Preprint submitted to Journal of LATEX Templates February 25, 2015 involved, adding one more level of parallelism. In this paper, the performance of an acoustic source localization system using distributed microphones is analyzed over a massive multichannel processing framework in a multi-GPU system. The paper evaluates and points out the influence that the number of microphones and the available computational resources have in the overall system perfor- mance. Several acoustic environments are considered to show the impact that noise and reverberation have in the localization accuracy and how the use of massive microphone systems combined with parallelized GPU algorithms can help to mitigate substantially adverse acoustic effects. In this context, the pro- posed implementation is able to work in real time with high-resolution spatial grids and using up to 48 microphones. These results confirm the advantages of suitable GPU architectures in the development of real-time massive acoustic signal processing systems. Keywords: Sound Source Localization; Steered Response Power; Microphone Arrays; Graphics Processing Units 1. Introduction Microphone arrays are commonly employed in many signal processing tasks, such as speech enhancement, acoustic echo cancellation or sound source separa- tion (Brandstein & Ward, 2001). The localization of broadband sound sources under high noise and reverberation is another challenging task in multichannel5 signal processing, being an active research topic with applications in human- computer interfaces (Kodagoda & Sehestedt, 2014), teleconferencing (Wang et al., 2011) or emergency units (Calderoni et al., 2015). Microphone arrays may follow a given geometry, such as spherical arrays (Huang & Wang, 2014), or may be distributed. Algorithms for sound source localization can be broadly10 divided into indirect and direct approaches (Madhu & Martin, 2008). Indirect approaches usually follow a two-step procedure: they first estimate the Time Difference Of Arrival (TDOA) (Chen et al., 2006) between microphone pairs, and, afterwards, they estimate the source position based on the geometry of 2 the array and the estimated delays. On the other hand, direct approaches per-15 form TDOA estimation and source localization in one single step by scanning a set of candidate source locations and selecting the most likely position as an estimate of the real source location. Although the computation of TDOAs usually requires time synchronization, new approaches are being developed to avoid this limitation (Xu et al., 2013). Most localization algorithms are based20 on the Generalized Cross-Correlation (GCC) (Knapp & Carter, 1976), which is calculated by using the inverse Fourier transform of the weighted cross-power spectral density of the signals. The Steered Response Power - Phase Trans- form (SRP-PHAT) algorithm is a direct approach that has been shown to be very robust in adverse acoustic environments (DiBiase et al., 2001). The algo-25 rithm is usually interpreted as a beamforming-based approach that searches for the candidate position that maximizes the output of a steered delay-and-sum beamformer. The CUDA platform (CUDA, 2015) provides a computing framework that enables the use of Graphics Processing Units (GPUs) in applications beyond30 image processing (Liu et al., 2007; Zhao & Lau, 2013). GPUs are high parallel programmable co-processors that provide efficient computation when the needed operations are properly parallelized. Programming a GPU efficiently requires having good knowledge of both the underlying architecture and the mechanisms used by GPUs to distribute their tasks among their processing units. Since the35 appearance of CUDA programming, many researchers in different areas have made use of it to achieve better performances in their respective fields. For example, well-known computational cores have also been adapted to a GPU computing framework, such as LU factorization (Dazevedo & Hill, 2012), matrix multiplication (Matsumoto et al., 2011) or the Boltzmann equation (Kloss et al.,40 2010). In audio and acoustics, several works demonstrate the potential of GPUs for carrying out audio processing tasks. For example, the implementation of a multichannel room impulse response reshaping algorithm was carried out in (Mazur et al., 2011), and implementations of adaptive filtering algorithms were presented in (Schneider et al., 2012; Lorente et al., 2012, 2013, 2014). GPU-45 3 based room acoustics simulation was carried out in (Savioja, 2010; Southern et al., 2010; Webb & Bilbao, 2011; Hamilton & Webb, 2013). One of the main contributions within this field was carried out in (Savioja et al., 2011), where improved performances in additive synthesis, Fourier transform and convolution in the frequency domain were presented. A comparison between CPU and GPU50 performance for a simple crosstalk canceller is presented in (Belloch et al., 2011). Similarly, a binaural audio application with massive audio processing that was fully implemented on a GPU is presented in (Belloch et al., 2013a). GPUs are also used in (Vanek et al., 2012) and in (Bradford et al., 2011) for evaluating the likelihood function in automatic speech recognizers and for sliding phase55 vocoder, respectively. The use of GPUs for implementing sound source localization algorithms has also recently been tackled in the literature. The time performances of different localization algorithms implemented on GPU were reported in (Peruffo Minotto et al., 2012) and (Liang et al., 2012). In fact, although different implementations60 of the SRP-PHAT in the time-domain and frequency-domain are analyzed in (Peruffo Minotto et al., 2012), their results mainly focus on pure computational issues and do not discuss how localization performance is affected by using differ- ent numbers of microphones or a finer spatial grid. In (Seewald et al., 2014), the SRP-PHAT algorithm is implemented over two Kinects for performing sound65 source localization. In the same work, the algorithm only estimates the relative source direction instead of providing the absolute source position and the im- plementation is evaluated on different GPUs that belong to the old-fashioned Fermi(CUDA, 2015). One of our previous works (Belloch et al., 2013b) analyzed the performance of70 a 2-D SRP-PHAT implementation with different Nvidia GPU architectures. The present paper extends that work in various aspects. First, 3-D source localiza- tion is considered, leading to a significant increase in the required computational cost. Second, the system considered in this work makes use of multiple GPUs, facing new challenges in parallelization and resource management. Finally, this75 paper provides a deeper analysis of the influence of the acoustic environment 4 and the number of microphones in the final performance. As a result, this paper is aimed at demonstrating how localization systems using a high number of mi- crophones distributed within a room can perform sound source localization in real time under adverse acoustic environments by using GPU massive computa-80 tion resources. Specifically, the well-known SRP-PHAT algorithm is considered here. Note that coarse-to-fine search strategies have been proposed to overcome many of the processing limitations of SRP-PHAT (Do & Silverman, 2007; Said et al., 2013; Marti et al., 2013). However, while these strategies provide more efficient ways to explore the localization search volume, they only provide better85 performance than the conventional SRP-PHAT when the number of operations is restricted. Thus, the performance of the conventional SRP-PHAT with fine spatial grids is usually considered as an upper bound in these cases. Relevant parameters that affect the computational cost of the algorithm (number of microphones and spatial resolution) are analyzed, showing their in-90 fluence on the localization accuracy in different situations. We also discuss the scalability of the algorithm when multi-GPU parallelization issues are consid- ered. This paper highlights the need for massive computation in order to achieve high-accuracy localization in adverse acoustic environments, taking advantage of GPUs to fulfill the computational demand of the system.95 In comparison with the implementation presented in (Seewald et al., 2014), we design our application to achieve maximum performance on GPUs making use of the Kepler architecture GK110 (K20, 2014) (See Appendix A for de- tails). This architecture can be found on the Tegra K1 (TK1) systems-on-chip (SoC), embedded in the Jetson development kit (DevKit) (Jetson, 2015), and100 it is becoming widespread in current mobile devices such as Google’s Nexus 9 tablet (Nexus, 2015). Thus, the proposed implementation can be success- fully adapted to work properly on GPUs that are currently embedded in mobile devices. The paper is structured as follows. Section 2 briefly describes the basic SRP-105 PHAT localization algorithm that will be used throughout this paper. Section 3 presents the implementation of the algorithm on multi-GPU systems. The 5 proposed acoustic environments for real-time sound source localization are pre- sented in Section 4, describing the experiments conducted for studying the per- formance of the method in a real application context. The computational perfor-110 mance of the different multi-GPU implementations are also analyzed. Finally, Section 5 provides some concluding remarks. Two Appendixes are provided in order to facilitate the understanding of the parallelization techniques that are used throughout this article. 2. Sound Source Localization: SRP-PHAT Algorithm115 Consider the output from microphone l, ml(t), in an M microphone system. The Steered Response Power (SRP) at the spatial point x = [x,y,z]T for a time frame n of length TL can then be defined as Pn(x) ≡ ∫ (n+1)TL nTL ∣∣∣∣∣ M∑ l=1 wlml (t− τ(x, l)) ∣∣∣∣∣ 2 dt, (1) where wl is a weight and τ(x, l) is the direct time of travel from location x to microphone l. DiBiase (DiBiase, 2000) showed that the SRP can be computed120 by summing up the Generalized Cross-Correlations (GCCs) for all possible pairs of the set of microphones. The GCC for a microphone pair (k,l) is defined as Rmkml (τ) = ∫ ∞ −∞ Φkl(ω)Mk(ω)M ∗ l (ω)e jωτdω, (2) where τ is the time lag, ∗ denotes complex conjugation, Ml(ω) is the Fourier transform of the microphone signal ml(t), and Φkl(ω) is a combined weighting function in the frequency domain. The phase transform (PHAT) (Knapp &125 Carter, 1976) has been shown to be a suitable GCC weighting for time delay estimation in reverberant environments. The PHAT weighting is expressed as: Φkl(ω) ≡ 1 |Mk(ω)M∗l (ω)| . (3) Taking into account the symmetries involved in the computation of Eq.(1) and removing some fixed energy terms (DiBiase, 2000), the part of Pn(x) that 6 changes with x can be isolated as130 P ′n(x) = M∑ k=1 M∑ l=k+1 Rmkml (τkl(x)) , (4) where τkl(x) is the Inter-Microphone Time-Delay Function (IMTDF). This func- tion is very important since it represents the theoretical direct path delay for the microphone pair (k,l) resulting from a point source located at x. The IMTDF is mathematically expressed as (Cobos et al., 2011) τkl(x) = ‖x−xk‖−‖x−xl‖ c , (5) where c is the speed of sound (≈ 343 m/s), and xk and xl are the locations of135 the microphone pair (k,l). The SRP-PHAT algorithm consists in evaluating the functional P ′n(x) on a fine grid G with the aim of finding the point-source location xs that provides the maximum value: xs = arg max x∈G P ′n(x). (6) Figure 1 shows schematically the intuition behind SRP-PHAT localization.140 In this figure, an anechoic environment is assumed so that the GCC for each microphone pair is a delta function located at the real TDOA. Each TDOA de- fines a half-hyperboloid of potential source locations. The intersection resulting from all the half-hyperboloids matches the point of the grid having the greatest accumulated value.145 2.1. SRP-PHAT Implementation The SRP-PHAT algorithm is usually implemented on a grid by carrying out the following steps: 1. A spatial grid G is defined with a given spatial resolution r. The the- oretical delays from each point of the grid to each microphone pair are150 pre-computed using Eq.(5). 2. For each analysis frame, the GCC of each microphone pair is computed as expressed in Eq.(2). 7 -5 -4 -3 -2 -1 0 1 2 3 4 5 -5 -4 -3 -2 -1 0 1 2 3 4 5 Intersecting Half-Hyperboloids for a Point-Source x [m] y [m ] Mic 1 Mic 2 Mic 3 Figure 1: Intersecting half-hyperboloids for M = 3 microphones. Each half-hyperboloid corresponds to a TDOA peak in the GCC. 3. For each position of the grid x ∈ G, the contribution of the different cross-correlations are accumulated (using delays pre-computed in 1), as in155 Eq.(4). 4. Finally, the position with the maximum score is selected as in Eq.(6). The SRP-PHAT localization performance depends on the selected spatial resolution r. Figure 2 illustrates the algorithm performance when considering different spatial grid resolutions. The accumulated SRP-PHAT values for each160 spatial grid location are shown for a 2-D plane in a 4 × 6 m room with N = 6 microphones. Note how the location of the source is more easily detected when finer spatial resolutions are used, as in the case of r = 0.01 m. 8 2.2. Computational Cost The SRP-PHAT algorithm is usually implemented by performing a frequency-165 domain processing of the input microphone signals. Given M microphones, the number of microphone pairs to process is Q = M(M − 1)/2. For a DFT size of L (equal to the time-window size), an FFT takes 5L log2 L arithmetic oper- ations that result from L 2 log2 L complex multiplications and L log2 L complex additions. Note that one complex multiplication is equivalent to four real mul-170 tiplications and one real addition, while a complex addition is equivalent to two real additions. As a result, the signal processing cost for computing the GCC is given by: • DFT: Compute M FFTs, then, M × 5L log2 L. • Cross-Power Spectrum: A complex multiplication for L points, result-175 ing in 6L operations (4 real multiplications and 2 real additions). This is done for Q microphone pairs, resulting in a cost of 6QL. • Phase Transform: Magnitude of the L points of the GCC, which costs L operations. This is also done for Q pairs, resulting in QL operations. • IDFT: The IDFT for Q pairs must be performed, which requires Q5L log2 L180 operations. Moreover, for each functional evaluation, the following parameters must be calculated: • M Euclidean distances, ‖xm‖, requiring 3 multiplications, 5 additions and 1 square root (≈ 12 operations): 20M operations185 • Q TDOAs, requiring 2 operations (1 subtraction and 1 division by c) per microphone pair: 2Q operations. • The SRP requires truncating the TDOA values to the closest sample ac- cording to the system sampling frequency, multiplying the cross-power spectrum to obtain the phase transform for each microphone pair and190 adding up all the GCC values: 5Q operations. 9 0 1 2 3 4 0 1 2 3 4 5 6 SRP-PHAT values with r = 0.01 m 0 1 2 3 4 0 1 2 3 4 5 6 SRP-PHAT values with r = 0.1 m y [m ] y [m ] x [m] x [m] (a) (b) Figure 2: Accumulated SRP-PHAT values for a 2-D spatial grid (4 × 6 m and M = 6 microphones) with different spatial resolutions. (a) r = 0.01 m. (b) r = 0.1 m. As a result, the cost of the SRP-PHAT is given by: Cost = ( M + M2 2 ) 5L log2 L + 7M(M − 1) 2 L + ν ( 20M + 7M(M − 1) 2 ) , (7) where ν is the total number of functional evaluations. In the conventional full grid-search procedure, ν equals the total number of points of the grid G. Figure 3 shows the computational cost of the algorithm for different spatial resolutions195 and number of microphones, considering a 3D grid search space with a uniform spatial resolution of r meters. 3. Algorithm Parallelization for real-time GPU implementation The GPU-based implementation of the SRP-PHAT algorithm is applied to Nvidia hardware devices with Kepler architecture GK110 (K20, 2014). Ap-200 pendix A provides a detailed description of the GPU-parallelization techniques used throughout this section. 10 Computational Cost M (Number of microphones) 5 10 15 20 25 30 35 40 45 10 6 10 7 10 8 10 9 10 10 10 11 10 12 10 13 = 0.05 m = 0.1 m = 0.01 m = 0.005 m r r r r N u m b er o f o p er a ti o n s Figure 3: Computational cost when for different number of microphones M and spatial reso- lutions r. Since the localization is carried out in three dimensions, three different reso- lutions rx, ry, and rz define the spatial grid G. Taking a shoe-box-shaped room as a model room with dimensions lx × ly × lz, the size of the grid is ν = Px ×205 Py × Pz, where Px = lxrx , Py = ly ry and Pz = lz rz . The real-time implementation of the SRP-PHAT algorithm uses 50% time- window overlap, with audio sample buffers of size L. These L×M samples are transferred to the GPU first. A GPU buffer (denoted here as TGPU ) stores the audio samples in consecutive memory positions as they arrive to the GPU. One210 aspect that affects the performance for all audio signal processing applications on GPU is the transfer of audio samples from CPU to GPU. As mentioned in Appendix A.1, streams can be used to parallelize these transfers and overlap them with the computation. Since we use 50% overlap, the processing is carried out in blocks of size 2L, which are composed of the current audio-sample buffer215 and the previous one. Thus, a size of 2LM is used for TGPU . The SRP-PHAT 11 GPU implementation carries out the following steps: 1. M streams are created (one stream for each microphone in the system). The streams are launched consecutively in an asynchronous way. Stream l transfers L samples captured by microphone l to the GPU and stores them220 in TGPU , with l = 0, . . . ,M−1. Then, stream l launches Kernel A, which is responsible for grouping 2L elements of microphone l (L samples from previous buffers and L samples from current buffers). These 2L elements are also weighted using a Hamming window vector. For this purpose, the stream launches a kernel that is composed of 128-size thread blocks in a225 CUDA grid of dimensions ( 2L 128 × 1 ) (i.e., it is composed of 2L CUDA threads). Each thread computes one element of the 2L elements. The tasks carried out by Kernel A are simple. Each thread reads one value from global-memory, multiplies it by a float number (a value of Hamming window vector) and stores it in a different position of global-230 memory. The accesses to global-memory are totally coalesced, since audio samples are stored in consecutive memory positions both when reading and when writing (see Fig. 4). Also, L is power of 2 and is always larger than 1024. Thus, each thread block reads and writes in 128 consecutive memory positions. The selection of 128 for the block size was done experimentally235 among 64, 128, 256 and 512, with 128 being the one that requires less time. 2. Once Kernel A has finished, stream l uses the CUFFT library to per- form a 2L-FFT using these 2L elements. As a result of the computation performed by all the streams, M vectors that are composed of 2L fre-240 quency bins (denoted as fl, l = 0, . . . ,M − 1 ) are obtained. The use of streams allows us to overlap data transfers with computation. For exam- ple, while stream 1 is transferring samples from microphone 1, stream 0 can be executing Kernel A. However, the next steps involve operations among different channels. Thus, all the previous operations must finalize245 before continuing. This implies synchronization among all the streams. 12 b lo ck D im .y (ThreadIdx.x, ThreadIdx.y) Grid of CUDA threads blockDim.x . . .. . . .. .. .. . .. . TGPU L ML elements from previous audio samples ML elements from current audio samples f0 Hamming Vector 2L 128 128 Figure 4: Operations that are carried out by CUDA kernel 21 in case M=4. The following steps are computed by only one stream. 3. The GCC matrix is computed by means of another kernel (Kernel B). In this kernel, a GPU thread takes one value from each of two different fl buffers that are at the same vector position. It conjugates one of the250 values and multiplies it by the corresponding value of the other fl buffer. The phase of the complex number obtained by the multiplication is stored in the corresponding position in the GCC matrix. The accesses to the two fl buffers by GPU threads are totally coalesced since consecutive threads access consecutive memory positions (see Fig. 5).255 Kernel B is limited by the instruction bandwidth since GPU-native func- tions cosf, sinf, and atan2f are used and all of them require various clock cycles. Kernel B computes 2LQ values of the GCC matrix, where Q represents the number of microphone pairs. In order to define the size and the number of blocks to launch in Kernel B, different tests were exe-260 cuted. The best performance was achieved by using 128-size thread blocks in a CUDA grid with size 32×16. This implies launching 65536 threads, where each thread is responsible for computing 2LQ 65536 values of the GCC matrix. In this case, increasing the amount of work per thread block in Kernel B is more beneficial than launching more blocks with fewer oper-265 ations per GPU thread. Thus, the grid configuration applied to Kernel B 13 achieves maximum occupancy when 512 blocks are launched. This kernel does not require using shared-memory but preferably a large number of registers. Thus, we set L1 cache to 48 KB. As described in (K20, 2014), cache L1 is used for register spills, local memory, and stack, which are all270 private per-thread variables. b lo ck D im .y (ThreadIdx.x, ThreadIdx.y) Grid of CUDA threads blockDim.x . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. .idx f0 M f1 fM-1 Q GCC Multiplication and Phase Computation 2L 128 gridDim.y=16 gridDim.x=32 128 x 32 = 4096 Figure 5: Operations that are carried out by Kernel B. 4. Q inverse FFTs of size 2L are then carried out by using the CUFFT library. The GCC matrix is now composed of temporal (time delay) values (i.e., 2LQ real values). 5. The computation of a tridimensional matrix SRP storing the accumulated275 SRP values is carried out by Kernel C. This kernel also launches thread blocks of size 128 in a tridimensional CUDA grid whose dimension depends on the number of points of the grid G (ν). In total, ν threads are launched. In this kernel, each GPU thread is devoted to the computation of the total value of the SRP at each point of the grid. To this end, each thread280 computes and accumulates Q GCC values (it takes a value from each row of the GCC matrix and accumulates it). The computation of the SRP requires Q calculations of the IMTDF (see Eq. 5) at each point of the grid. The IMTDF of a pair of microphones specifies the column of the GCC matrix that should be selected and then accumulated in the SRP.285 Figure 6 illustrates these operations. Since the value of the IMTDF can indicate any position of the column of the GCC matrix, coalesced access to the global-memory is not guar- anteed. In fact, the most probable situation is that the accesses will be 14 . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . b lo ck D im .y (ThreadIdx.x, ThreadIdx.y, ThreadIdx.z) Grid of CUDA threads blockDim.x . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . . . .. . . .. .. .. . .. . blockDim.z idx SRP Px Py GCC Pz A GCC value is accumulated from each Row. This value corresponds to the column pointed out by the IMTDF. Computation of IMTDF Q Figure 6: Operations that are carried out by Kernel C. quite disordered, so that the kernel employs most of its time in memory290 accessing. However, this limitation can be reduced if we force the compiler to use the Kepler read-only data cache with the GCC matrix, since this cache does not require aligned accesses. This read-only cache memory has also been used in recent GPU-based audio research such as (Hamilton & Webb, 2013) and (Bilbao & Webb, 2013). Furthermore, as in Kernel B, we295 set L1 cache to 48 KB to favor possible register spills. In the accumulation loop of the SRP values, we have set a #pragma unroll to accelerate the computation. 6. The grid position corresponding to the maximum SRP value has to be searched. To this end, we launch Kernel D. This kernel exactly follows300 the reduction example in Harris’ implementation (Harris, 2014) that comes with the Nvidia GPU Computing SDK (Software development kit), but it changes the sum operation for a maximum operation. However, even though this code is optimized for finding the maximum value, it does not indicate its position. Thus, after obtaining the maximum, we launch an-305 other kernel (Kernel E). This kernel launches as many threads as elements of the SRP matrix and only performs a comparison operation with the maximum. If the comparison matches, the thread writes the value of its index in a variable. 15 3.1. Memory considerations310 The computation of the IMTDF could be carried out off-line since the grid resolutions and the microphone locations are static. However, this would imply storing a 4-dimensional data structure composed of ν ·Q elements. If we use a standard room size (such as 6.0 × 4.0 × 3.0 m), a resolution of r = rx = ry = rz = 0.01 m, and M = 48 microphones, this data structure would require using315 more than eight gigabytes of global-memory. This exceeds the global-memory size of most available GPU devices. Thus, every IMTDF value is computed for each group of processed buffers. There are also other variables that are used to compute the values of the GCC and SRP matrices, such as the room dimensions, the number of mi-320 crophones and their position. Since all of these read-only variables must be available for all of the threads, they are stored in the constant memory (with size 64 KB). 3.2. Multi-GPU Parallelization Distributing the above processing tasks among different GPUs is not straight-325 forward. The greatest computational load relies on Step 6, which consists in computing the maximum value of the SRP matrix. Table 1 shows the elapsed time corresponding to each step for M =48 microphones and a spatial grid resolution of r =0.01 m. The tasks from Kernels C, D and E can be easily distributed among NGPU330 GPUs (the number of GPUs present in the system): each GPU computes ν NGP U elements of the SRP matrix and locates the maximum among its computed elements. To this end, NGPU CPU threads are created at the beginning of a parallel region by means of openMP (see Appendix A.2 to know how openMP can deal with multiple GPUs). This strategy is only focused on multi-GPU335 parallelization of the SRP matrix. In Appendix B, there is a description of an alternative strategy that aims at parallelizing both the computation of the SRP matrix and the GCC matrix. 16 Table 1: Elapsed Time in each kernel with M=48 and spatial grid resolution r=0.01. Steps of the algorithm Time [ms] Transfer + Kernel A + FFT (steps 1 and 2 in Section III) 1.416 Kernel B (step 3: Computation of GCC) 0.015 IFFT of GCC (step 4) 0.006 Kernel C (Computation of SRP matrix) 0.007 Kernel D (Reduction: Computation of Maximum SRP value) 121.267 Kernel E (Localization of the Maximum) 0.009 Total elapsed time 122.720 This strategy uses also the UVA (Unified Virtual Addressing) feature for inter- GPU communication. This strategy requires different synchronization points340 that significantly penalize their performances, especially when compared to the parallelization presented in this article. 3.3. Basic Implementation using two GPUs As shown in Section 4, the performance of the SRP-PHAT algorithm is assessed in a system that is composed of two GPUs. Using all the parallelization345 techniques previously presented, the SRP-PHAT algorithm is implemented on two GPUs as follows: 1. A parallel region is created with two CPU threads. Each CPU thread is bound with a GPU. 2. Since different audio buffers are received in the system, each CPU thread350 independently and asynchronously sends all audio buffers to its GPU by using stream parallelization. The Kernels A and the FFTs are computed for each channel inside the streams. 3. As in step 2 of Section 3, stream synchronization is addressed. Only one stream is used to compute the rows of the GCC matrix.355 17 CPU master thread CPU thread 0 CPU thread 1 Transfer L Kernel A 2L-FFT Stream 0 Stream 1 Transfer L Kernel A 2L-FFT Stream 11 Transfer L Kernel A 2L-FFT Stream 12 Kernel A 2L-FFT Transfer L Kernel A 2L-FFT Stream 0 Stream 1 Transfer L Kernel A 2L-FFT Stream 11 Transfer L Kernel A 2L-FFT Stream 12 Transfer L Kernel A 2L-FFT Kernel B computes 66 rows of GCC matrix Kernel C computes half elements of SRP matrix Kernel B computes 66 rows of GCC matrix Kernel C computes half elements of SRP matrix Kernels D and E computes and locate the maximum of half elements of SRP matrix Kernels D and E computes and locate the maximum of half elements of SRP matrix Transfer maximum and localization to CPU Transfer maximum and localization to CPU CPU master thread Select Maximum value and its localization GPU0 GPU1 Transfer L Figure 7: Steps of the GPU-based SRP-PHAT implementation using two GPUs and openMP. 4. Since both of them have computed the GCC matrix, each GPU computes ν/2 elements of the SRP matrix and locates a maximum value among the computed elements. 5. Each GPU transfers back to the CPU its maximum value and its location inside the SRP matrix. Then, a synchronization barrier for both CPU360 threads is set followed by an openMP section that is only executed by the master thread. This thread compares the two maximum values and chooses the greatest one, getting its location. This location indicates the sound source position. Figure 7 illustrates the computation of the SRP- PHAT when M = 12.365 4. Experiments and Performance To analyze both the computing and localization performance of the above GPU implementations, a set of acoustic simulations using the image-source method (Allen & Berkley, 1979) have been considered. A shoe-box-shaped room with dimensions 4×6×3 m and wall reflection factor ρ (Kuttruff, 2000)370 was simulated using different numbers of microphones (M ∈ {6, 12, 24, 48}). 18 M = 6 M = 12 M = 24 M = 48 x y 6 m 4 m 3 m Figure 8: Microphone set-ups for M = 6, M = 12, M = 24 and M = 48. The black dots denote the actual active microphones in each configuration. The microphone set-up for the considered systems are shown in Figure 8. Note that the microphones are located on the walls of the room and are placed on eight different planes (z = {0.33, 0.66, 1.00, 1.33, 1.66, 2.00, 2.33, 2.66}) following hexagon-like shapes. Moreover, different reflection factors (ρ ∈ {0, 0.5, 0.9})375 were used to take into account different reverberation degrees. In all cases, in- dependent white Gaussian noise was added to each microphone signal in order to simulate different Signal to Noise Ratios (SNR ∈{0, 5, 10, 20}) (in dB). The audio card used in the real-time prototype uses an ASIO (Audio Stream Input/Output) driver to communicate with the CPU and provides 2048 samples380 per microphone (L=2048) every 46.43 ms at a sample frequency of 44.1 kHz. This time is denoted by tbuff . The time employed for the computation is denoted by tproc. This time takes into account all transfer times and measures the time from the first audio sample transferred to the GPU until the final source location 19 Table 2: Characteristics of the GPU K20c. Cuda Device Tesla K20c Architecture Kepler Capability 3.5 Number of SM 13 Total number of cores 2496 Max. dimension of a block 1024 x 1024 x 64 Max. dimension of a grid 231-1 x 65535 x 65535 Total amount of global memory 4 GB is estimated (at each time frame). The localization system works in real time385 as long as tproc < tbuff . Otherwise, microphone samples would be lost and the localization would not be correctly performed. The simulations were carried out in the Nvidia GPU K20c (K20, 2014), which has the characteristics shown in Table 2. Both computational and localization performances have been assessed taking into account three spatial grid resolutions (r ∈ {0.1, 0.05, 0.01}) in the390 XY plane (resolutions rx and ry are equal). The resolution rz is 0.33 m (resulting from dividing the height of the room into eight slots). 4.1. Localization Performance The source signal used in this study was a 5-second male speech signal with no speech pauses. Pauses were manually suppressed to evaluate localization395 performance only over frames where speech was present. The processing was carried out by using 50% overlap in time windows of length 4096 samples (size 2L), with sampling frequency fs = 44.1 kHz. For each frame, a source location x̂ = [x̂, ŷ, ẑ]T was estimated. A total number of 107 frames (Nf =107) per 40 different positions (Np=40) that were uniformly distributed over the room space400 were performed. Localization accuracy was computed by means of the Mean 20 Absolute Error, which is given by: MAE = 1 Nf 1 Np Nf∑ i=1 Np∑ j=1 |eij|2, (8) where eij=xij − x̂ij, with xij and x̂ij being the true and estimated source loca- tions at a given time frame i and source position j. Note that the above MAE was computed for each environmental condition (reflection factor and signal to405 noise ratio), microphone setup and spatial grid resolution. Figure 9 shows the results for different values of wall reflection factor ρ taking into account different spatial resolutions and number of microphones. It is important to point out that using a high number of microphones helps to substantially improve localization accuracy under high noise and reverberation.410 The error decreases as the SNR increases and/or reverberation decreases (lower ρ). It is important to see how the spatial resolution has an impact when there are few microphones. In this case, a coarse spatial grid is not sufficient to correctly find the minimum of the SRP search space, which is more easily detected when the SRP is enhanced by the contributions of additional microphone pairs. In415 fact, when the number of microphones is 12 or higher, the performance difference between r = 0.01 and r = 0.1 is almost negligible. Accuracy differences among different values of ρ are noticiable. It should be emphasized that, under favorable acoustic conditions (high SNR and low ρ), the experimental error is always below the maximum expectable error independently of the number of microphones.420 Note that the maximum error in anechoic conditions is given by the largest diagonal of the cuboids forming the 3D grid (≈ 0.179 m for r = 0.1 and ≈ 0.165 m for r = 0.01). In all cases, the use of a higher number of microphones significantly helps in reducing this error. 4.2. Computational Performance425 The spatial resolutions considered in this paper result in large-scale SRP matrices. Table 3 shows the processing times tproc for different combinations of r and M when using two GPUs. It can be observed that the only that does not obtain a tproc lower than 46.43 ms (tbuff ) is the configuration composed of 21 SNR 0 5 10 15 20 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Accuracy =0.01 ρ=0.0 SNR 6 Mic. 12 Mic. 24 Mic. 48 Mic. M A E [ m ] spr 0 5 10 15 20 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Accuracy =0.01 ρ=0.5 6 Mic. 12 Mic. 24 Mic. 48 Mic. M A E [ m ] spr Accuracy =0.01 ρ=0.9 0 5 10 15 20 0 0.5 1 1.5 2 2.5 SNR 6 Mic. 12 Mic. 24 Mic. 48 Mic. M A E [ m ] spr 0 5 10 15 20 0 0.5 1 1.5 2 2.5 Accuracy =0.1 ρ=0.9 SNR 6 Mic. 12 Mic. 24 Mic. 48 Mic. M A E [ m ] spr 0 5 10 15 20 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Accuracy =0.1 ρ=0.5 SNR 6 Mic. 12 Mic. 24 Mic. 48 Mic. M A E [ m ] spr 0 5 10 15 20 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Accuracy =0.1 ρ=0.0 SNR 6 Mic. 12 Mic. 24 Mic. 48 Mic. M A E [ m ] spr Figure 9: Localization accuracy for different wall reflection factors (ρ ∈ {0, 0.5, 0.9}) as a function of the SNR and the number of microphones M. Each column presents results for different spatial resolutions (r = 0.1 and r = 0.01 m). M = 48 and r = 0.01. Thus, real-time processing is not possible in this case.430 However, by looking at the results shown in Table 4, it is possible to observe 22 Table 3: Processing time tproc using two GPUs. r M = 6 M = 12 M = 24 M = 48 0.01 1.031 ms 3.578 ms 15,564 ms 60.108 ms 0.05 0.381 ms 0.758 ms 2.238 ms 6.433 ms 0.1 0.371 ms 0.650 ms 1.588 ms 4.588 ms Table 4: Processing time tproc using one GPU. r M=6 M=12 M=24 M=48 0.01 1.894 ms 6.731 ms 30.145 ms 122.720 ms 0.05 0.564 ms 1.132 ms 3.484 ms 11.203 ms 0.1 0.546 ms 0.926 ms 2.336 ms 7.493 ms that the influence of the second GPU becomes relevant. In the case of M = 48, the processing time is halved for any resolution. Real-time processing would be easily achieved for M = 48 and r = 0.01 by adding an additional GPU. Figure 10 shows more clearly the time differences among all the configurations by varying435 the number of microphones and the grid resolutions r ∈ {0.1, 0.05, 0.01}. Note that the time tbuff is marked by a solid black line. 5. Conclusion New emerging GPU architectures help to overcome different computational problems in acoustic signal processing algorithms involving many microphone440 channels. This paper has analyzed the specific case of sound source localization, where very fine spatial resolutions or having a high number of microphones have a deep impact in the performance of real-time applications. In this context, the following contributions have been presented in this paper. Firstly, we have proposed a scalable multi-GPU implementation of the well-445 known SRP-PHAT algorithm for source localization in three dimensions. To this end, two parallelization levels have been considered. On the one hand, 23 0.01 0.05 0.1 0 20 40 60 80 100 120 One GPU r [m s] M=06 M=12 M=24 M=48 tbuff 0.01 0.05 0.1 0 20 40 60 80 100 120 Two GPUs r [m s] M=06 M=12 M=24 M=48 tbuff Figure 10: Time tproc for different resolutions and number of microphones. multiple cores are included in a GPU device. On the other hand, the system is composed of several GPUs. The analyzed computational performance indicates that the algorithm is scalable, so that the time employed to estimate the source450 location is reduced with the number of GPU devices. Secondly, we have evaluated the relation existing among localization accu- racy, number of microphones and available computational resources. While most works take into account computational issues or performance issues only, we have approached both aspects from a practical implementation perspective.455 To this end, simulated experiments going from a noiseless anechoic case to a noisy and highly reverberant case have been analyzed, studying the overall per- formance with a varying number of microphones and spatial resolutions. Results show that thanks to the GPU computational resources, a fine spatial search grid and a high number of microphones can be used to improve the localization ac-460 curacy and the robustness of the system. Finally, the presented implementation exploits the resources of GPUs with Kepler architecture, which can be currently found in new generation mobile devices such as Google’s Nexus 9 tablet. Thus, the proposed implementation can be successfully run on embedded mobile devices.465 24 Acknowledgments This work has been partially funded by the Spanish Ministerio de Economı́a y Competitividad (TEC2009-13741, TEC2012-38142-C04-01, and TEC2012- 37945-C02-02), Generalitat Valenciana PROMETEO 2009/2013, and Univer- sitat Politècnica de València through Programa de Apoyo a la Investigación y470 Desarrollo (PAID-05-11 and PAID-05-12). References Allen, J. B., & Berkley, D. A. (1979). Image method for efficiently simulating small-room acoustics. J. Acoust. Soc. Am., 65 , 943–950. Belloch, J. A., Ferrer, M., Gonzalez, A., Martinez-Zaldivar, F., & Vidal, A. M.475 (2013a). Headphone-based virtual spatialization of sound with a GPU accel- erator. J. Audio Eng. Soc, 61 , 546–561. Belloch, J. A., Gonzalez, A., Martinez-Zaldivar, F. J., & Vidal, A. M. (2011). A real-time crosstalk canceller on a notebook GPU. In Proc. of IEEE Inter- national Conference on Multimedia and Expo (ICME) (pp. 1 –4). Barcelona,480 Spain. Belloch, J. A., Gonzalez, A., Vidal, A. M., & Cobos, M. (2013b). Real-Time Sound Source Localization on Graphics Processing Units. In Proc. of the International Conference on Computational Science, ICCS 2013 . Barcelona, Spain.485 Bilbao, S., & Webb, C. J. (2013). Physical modeling of timpani drums in 3D on GPGPUs. J. Audio Eng. Soc, 61 , 737–748. Bradford, R., Ffitch, J., & Dobson, R. (2011). Real-time sliding phase vocoder using a commodity GPU. In Proc. of ICMC 2011 . University of Huddersfield, United Kingdom.490 Brandstein, M., & Ward, D. (2001). Microphone arrays . Springer. 25 Calderoni, L., Ferrara, M., Franco, A., & Maio, D. (2015). Indoor localization in a hospital environment using Random Forest classifiers. Expert Systems with Applications , 42 , 125 – 134. doi:http://dx.doi.org/10.1016/j.eswa. 2014.07.042.495 Chen, J., Benesty, J., & Huang, Y. (2006). Time delay estimation in room acoustic environments: an overview. EURASIP Journal on Applied Signal Processing , 2006 , 1–19. Cobos, M., Marti, A., & Lopez, J. J. (2011). A modified SRP-PHAT functional for robust real-time sound source localization with scalable spatial sampling.500 IEEE Signal Processing Letters , 18 , 71–74. Cook, S. (2013). A Developer’s Guide to Parallel Computing with GPUs . Mor- gan Kaufmann. CUDA (2015). Nvidia CUDA Developer Zone. https://developer.nvidia.com/cuda-downloads. (accessed 2015 Febru-505 ary 02). Dazevedo, E., & Hill, J. (2012). Parallel LU Factorization on GPU Cluster. Procedia Computer Science, 9 , 67 – 75. doi:10.1016/j.procs.2012.04.008. Proceedings of the International Conference on Computational Science, ICCS 2012.510 DiBiase, J. H. (2000). A high accuracy, low-latency technique for talker local- ization in reverberant environments using microphone arrays . Ph.D. thesis Brown University Providence, RI. DiBiase, J. H., Silverman, H. F., & Brandstein, M. S. (2001). Robust localization in reverberant rooms. In M. S. Brandstein, & D. Ward (Eds.), Microphone515 Arrays: Signal Processing Techniques and Applications chapter 8. (pp. 157– 180). Berlin, Germany: Springer-Verlag. Do, H., & Silverman, H. F. (2007). A fast microphone array SRP-PHAT source location implementation using coarse-to-fine region contraction (CFRC). In 26 http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2014.07.042 http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2014.07.042 http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2014.07.042 https://developer.nvidia.com/cuda-downloads http://dx.doi.org/10.1016/j.procs.2012.04.008 Proc. of the IEEE Workshop on Applications of Signal Processing to Audio520 and Acoustics . New Paltz, USA. Hamilton, B., & Webb, C. J. (2013). Room acoustics modelling using GPU- accelerated finite difference and finite volume methods on a face-centered cubic grid. In Proc. of Conference on Digital Audio Effects (DAFx-13). Maynooth, Ireland.525 Harris, M. (2014). Optimizing Parallel Reduction in CUDA NVIDIA. http://developer.download.nvidia.com/assets/cuda/files/ reduction.pdf. (accessed 2014 August 27). Huang, Q., & Wang, T. (2014). Acoustic source localization in mixed field using spherical microphone arrays. EURASIP Journal on Applied Signal Process-530 ing , 90 , 1–16. Jetson (2015). Mobile GPU: Jetson. http://developer.download.nvidia.com/embedded/jetson/TK1/docs/ Jetson_platform_brief_May2014.pdf. (accessed 2014 November 22). K20 (2014). NVIDIA Kepler Architecture.535 http://www.nvidia.com/content/PDF/kepler/ NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf. (accessed 2014 August 27). Kloss, Y., Shuvalov, P., & Tcheremissine, F. (2010). Solving Boltzmann equa- tion on GPU. Procedia Computer Science, 1 , 1083 – 1091. Proceedings of540 the International Conference on Computational Science, ICCS 2010. Knapp, C. H., & Carter, G. C. (1976). The Generalized Correlation Method for Estimation of Time Delay. IEEE Transactions on Acoustics, Speech and Signal Processing , ASSP-24 , 320–327. Kodagoda, S., & Sehestedt, S. (2014). Simultaneous people tracking and mo-545 tion pattern learning. Expert Systems with Applications , 41 , 7272 – 7280. doi:http://dx.doi.org/10.1016/j.eswa.2014.05.019. 27 http://developer.download.nvidia.com/assets/cuda/files/reduction.pdf http://developer.download.nvidia.com/assets/cuda/files/reduction.pdf http://developer.download.nvidia.com/assets/cuda/files/reduction.pdf http://developer.download.nvidia.com/embedded/jetson/TK1/docs/Jetson_platform_brief_May2014.pdf http://developer.download.nvidia.com/embedded/jetson/TK1/docs/Jetson_platform_brief_May2014.pdf http://developer.download.nvidia.com/embedded/jetson/TK1/docs/Jetson_platform_brief_May2014.pdf http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2014.05.019 Kuttruff, H. (2000). Room acoustics . Abingdon, Oxford, UK: Taylor & Francis. 368 pages. Liang, Y., Cui, Z., Zhao, S., Rupnow, K., Zhang, Y., Jones, D. L., & Chen, D.550 (2012). Real-time implementation and performance optimization of 3D sound localization on GPUs. In DATE’12 (pp. 832–835). Liu, W., Schmidt, B., Voss, G., & Muller-Wittig, W. (2007). Streaming algorithms for biological sequence alignment on GPUs. IEEE Transac- tions on Parallel and Distributed Systems , 18 , 1270–1281. doi:http://doi.555 ieeecomputersociety.org/10.1109/TPDS.2007.1069. Lorente, J., Ferrer, M., , De Diego, M., & Gonzalez, A. (2014). GPU Implemen- tation of Multichannel Adaptive Algorithms for Local Active Noise Control. IEEE Transactions on Acoustics, Speech and Signal Processing , 22 , 1624 – 1635. doi:10.1109/TASLP.2014.2344852.560 Lorente, J., Ferrer, M., De Diego, M., Belloch, J. A., & Gonzalez, A. (2013). GPU implementation of a frequency-domain modified filtered-X LMS algo- rithm for multichannel local active noise control. In Proc. of the 52nd AES Conference. Guildford, United Kingdom. Lorente, J., Gonzalez, A., Ferrer, M., Belloch, J. A., De Diego, M., Piñero,565 G., & Vidal, A. M. (2012). Active noise conrol using Graphics Processing Units. In Proc of the International Congress on Sound and Vibration. Vilnius, Lithuania. Madhu, N., & Martin, R. (2008). Advances in digital speech transmission. chapter Acoustic Source Localization with Microphone Arrays. (pp. 135–166).570 New York, NY, USA: Wiley. Marti, A., Cobos, M., & Lopez, J. J. (2013). A steered response power iterative method for high-accuracy acoustic source location. Journal of the Acoustical Society of America, 134 . 28 http://dx.doi.org/http://doi.ieeecomputersociety.org/10.1109/TPDS.2007.1069 http://dx.doi.org/http://doi.ieeecomputersociety.org/10.1109/TPDS.2007.1069 http://dx.doi.org/http://doi.ieeecomputersociety.org/10.1109/TPDS.2007.1069 http://dx.doi.org/10.1109/TASLP.2014.2344852 Matsumoto, K., Nakasato, N., Sakai, T., Yahagi, H., & Sedukhin, S. G. (2011).575 Multi-level Optimization of Matrix Multiplication for GPU-equipped Sys- tems. Procedia Computer Science, 4 , 342 – 351. Proceedings of the In- ternational Conference on Computational Science, ICCS 2011. Mazur, R., Jungmann, J., & Mertins, A. (2011). On CUDA implementation of a multichannel room impulse response reshaping algorithm based on p-norm580 optimization. In IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) (pp. 305 –308). doi:10.1109/ASPAA.2011. 6082310. Nexus (2015). Google’s nexus 9. http://blogs.nvidia.com/blog/2014/10/17/585 nvidia-tegra-k1-google-nexus-9/. (accessed 2015 January 11). openMP (2014). openMP API Specifications. http://www.openmp.org. (accessed 2014 June 05). Peruffo Minotto, V., Rosito Jung, C., Gonzaga da Silveira, L., & Lee, B. (2012). GPU-based approaches for real-time sound source localization using the SRP-590 PHAT algorithm. International Journal of High Performance Computing Applications , . doi:10.1177/1094342012452166. Said, A., Lee, B., & Kalker, T. (2013). Fast steered response power computation in 3D spatial regions . Technical Report HPL-2013-40 HP Labs Palo Alto, USA.595 Savioja, L. (2010). Real-time 3D finite-difference time-domain simulation of low- and mid-frequency room acoustics. In Proc. of the Int. Conf. Digital Audio Effects . Graz, Austria. Savioja, L., Välimäki, V., & Smith, J. O. (2011). Audio Signal Processing using Graphics Processing Units. J. Audio Eng. Soc, 59 , 3–19.600 Schneider, M., Schuh, F., & Kellermann, W. (2012). The Generalized Frequency-Domain Adaptive Filtering Algorithm Implemented on a GPU for 29 http://dx.doi.org/10.1109/ASPAA.2011.6082310 http://dx.doi.org/10.1109/ASPAA.2011.6082310 http://dx.doi.org/10.1109/ASPAA.2011.6082310 http://blogs.nvidia.com/blog/2014/10/17/nvidia-tegra-k1-google-nexus-9/ http://blogs.nvidia.com/blog/2014/10/17/nvidia-tegra-k1-google-nexus-9/ http://blogs.nvidia.com/blog/2014/10/17/nvidia-tegra-k1-google-nexus-9/ http://www.openmp.org http://dx.doi.org/10.1177/1094342012452166 Large-Scale Multichannel Acoustic Echo Cancellation. Speech Communica- tion; 10. ITG Symposium; Proceedings of , (pp. 1 –4). Seewald, L. A., Gonzaga, L., Veronez, M. R., Minotto, V. P., & Jung, C. R.605 (2014). Combining SRP-PHAT and two Kinects for 3D Sound Source Lo- calization. Expert Systems with Applications , 41 , 7106 – 7113. doi:http: //dx.doi.org/10.1016/j.eswa.2014.05.033. Southern, A., Murphy, D., Campos, G., & Dias, P. (2010). Finite difference room acoustic modelling on a General Purpose Graphics Processing Unit. In610 Proc. of the 128th AES Convention. London, United Kingdom. Vanek, J., Trmal, J., Psutka, J., & Psutka, J. (2012). Optimized Acoustic Likelihoods Computation for NVIDIA and ATI/AMD Graphics Processors. IEEE Transactions on Audio, Speech, and Language Processing , 20 , 1818– 1828. doi:10.1109/TASL.2012.2190928.615 Wang, R., Wang, X., & Kim, M. J. (2011). Motivated learning agent model for distributed collaborative systems. Expert Systems with Applications , 38 , 1079 – 1088. doi:http://dx.doi.org/10.1016/j.eswa.2010.05.003. Intelligent Collaboration and Design. Webb, C. J., & Bilbao, S. (2011). Computing room acoustics with CUDA - 3D620 FDTD schemes with boundary losses and viscosity. In Proc of IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). Prague, Czech Republic. Xu, B., Sun, G., Yu, R., & Yang, Z. (2013). High-Accuracy TDOA-Based Localization without Time Synchronization. IEEE Transactions on Parallel625 and Distributed Systems , 24 , 1567–1576. doi:10.1109/TPDS.2012.248. Zhao, Y., & Lau, F. C. (2013). Implementation of Decoders for LDPC Block Codes and LDPC Convolutional Codes Based on GPUs. IEEE Transactions on Parallel and Distributed Systems , 99 , 1. doi:http://doi. ieeecomputersociety.org/10.1109/TPDS.2013.52.630 30 http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2014.05.033 http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2014.05.033 http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2014.05.033 http://dx.doi.org/10.1109/TASL.2012.2190928 http://dx.doi.org/http://dx.doi.org/10.1016/j.eswa.2010.05.003 http://dx.doi.org/10.1109/TPDS.2012.248 http://dx.doi.org/http://doi.ieeecomputersociety.org/10.1109/TPDS.2013.52 http://dx.doi.org/http://doi.ieeecomputersociety.org/10.1109/TPDS.2013.52 http://dx.doi.org/http://doi.ieeecomputersociety.org/10.1109/TPDS.2013.52 Appendix A. GPU and CUDA GPUs are wellknown for their potential in highly parallel data processing. A GPU is composed by multiple Stream Multiprocessors (SM) where, for 3.5 capability (Kepler architecture (K20, 2014)), there are 192 pipelined cores per SM1. In the CUDA model, the programmer defines the kernel function where the635 code to be executed on the GPU is written. A grid configuration, which defines the number of threads and how they are distributed and grouped, must be built into the main code (threads are grouped into thread blocks, and thread blocks configure a grid that is organized in three dimensions, denoted as BlockIdx.x, BlockIdx.y, and BlockIdx.z). The total number of threads launched in a640 kernel by means of thread blocks can exceed the number of physical cores. At runtime, the kernel distributes all the thread blocks among SMs. Each SM can host up to 16 thread blocks. If the number of blocks exceeds the resources of the GPU, these blocks wait until other blocks finish in order to be hosted later. A GPU device has a large amount of off-chip device memory (global-memory )645 and a fast on-chip memory (shared-memory, registers ). As its name indicates, the shared-memory is normally used when multiple threads must share data. There are also read-only cached memories called constant-memory and texture- memory. The first memory is optimized for broadcast (i.e., when all the threads read the same memory location), while the second one is more oriented to graph-650 ics. Figure A.1 shows how the GPU architecture is organized. Advanced GPU devices (beyond 2.x capability) come with an L1/L2 cache hierarchy that is used to cache global-memory. Cache L1 uses the same on-chip memory as shared- memory ; how much of the on-chip memory is dedicated to L1 is set for each kernel call.655 1At the time this paper was written, the most advanced GPU device was K20c with Kepler architecture which is the one considered throughout this work 31 SP SM GPU Shared M. / L1 Cache Registers SM Shared M. / L1 Cache Registers SM Shared M. / L1 Cache Registers Constant Memory/Texture Memory L2 Cache Global Memory SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SP SPSP SP SP SP SP Figure A.1: The GPU is configured by 16 Stream Multiprocessors (SMs), each of which has 192 pipelined cores (SP). Appendix A.1. Streams on GPU Streams are virtual work queues on the GPU. They are used for asynchronous operation, (i.e, the control of the program returns to the CPU immediately). Operations assigned to the same stream are executed in order and sequentially. Multiple streams can be defined on CUDA programming; however, up to 32660 streams are available to be independently run on the GPU thanks to the Hyper- Q technology that is presented in hardware with 3.5 capability (Cook, 2013). Different streams may execute their assigned operations out of order with respect to one another or concurrently. Thus, when a launched kernel does not require all the GPU resources, these could be used for another kernel that665 was launched from a different stream. Hence, streams allow multiple kernels to be launched concurrently. Following this idea, data transfer between CPU and GPU can also be overlapped with kernel computations and other transfers whenever they are carried out in different streams. If the data transfers are not assigned to any stream queue, they are executed synchronously and in an670 isolated way, (i.e., the CPU waits until all the previous operations have finished). GPU kernels are always launched asynchronously by the CPU (regardless of 32 whether or not they are scheduled on a stream queue). Thus, data transfers are usually used as a synchronization barrier. Figure A.2 illustrates the parallelization obtained using streams when M = 4675 (number of microphones) for steps 1,2, and 3 from Section 3 in the main article. Note that, in Fig.A.2, the alternative Kernel A’ would have a CUDA grid with dimensions ( 2L 128 × M ). Transfer L audio samples Kernel A 2L-FFT Transformation Transfer L audio samples Kernel A Stream 0 Stream 3 Transfer L audio samples Kernel A 2L-FFT Transformation Transfer L audio samples Kernel A 2L-FFT Transformation 2L-FFT Transformation Stream 1 Stream 2 CPU thread Transfer ML audio samples Kernel A’ M 2L-FFT Transformations CPU thread Figure A.2: Parallelization obtained with streams when M=4. Appendix A.2. Multi-GPU programming with multicore One of the standards that allows for multicore processing is openMP (openMP,680 2014). This standard works by using a fork/join pattern, that is, parallel regions are specified by the programmer. The CPU code runs sequentially and at some point hits a section where work can be distributed into several processors that perform the computations (CPU core spans several CPU threads). Afterwards, when all the computations are completed, all the CPU threads converge to a685 single thread again, which is called the master thread. If a machine has a multicore processor and several GPUs, the parallelization can be achieved by defining a number of threads in the parallel region equal to the number of GPUs. In this sense, each CPU thread deals with a GPU. This is very important since a CPU thread is bound with a GPU context.690 Thus, all subsequent CUDA calls (e.g. cudaMalloc) allocate memory only on its corresponding GPU (Cook, 2013). 33 PCI-E GPU 0 Communication through CPU memory space GPU 1 CPU Memory space PCI-E GPU 0 GPU 1 Memory space CPU Peer-to-Peer Communication Figure A.3: The UVA feature reduces data-transfer time among GPUs by using peer-to-peer communication (bottom). Recent CUDA releases (beyond 2.x capability and CUDA SDK 4.x) allow the time employed in data transfers among GPUs to be reduced by using the UVA (Unified Virtual Addressing) feature. That means that inter-GPU com-695 munication (peer-to-peer, P2P) can also be performed without routing the data through the CPU, saving PCI-E bandwidth. Before the appearance of these recent features, communication among GPUs had to be carried out through memory space in the CPU, as shown in Figure A.3. Appendix B. Multi-GPU Parallelization strategy involving GCC and700 SRP matrices The challenge of this strategy consists in parallelizing the computation of the GCC matrix. Initially, all the GPUs must have access to this matrix since each point of the SRP matrix requires a contribution from each pair of microphones (each row of the GCC matrix).705 The strategy that we present aims at achieving a good trade-off between the total operations carried out in each GPU and the number of transferred audio 34 buffers. For example, if the number of microphones is M = 12, the number of pairs to compute in GCC matrix is Q = 66. These pairs are distributed among the NGPU in a pseudo-triangular way. Figure B.1 shows the distribution of the710 computation and audio buffers among 2, 3 and 4 GPUs. The notation 01 x 05, indicates the element-wise multiplication of vector 1 and vector 5 of all computed vectors fl, l = 0, . . . ,M − 1 (see step 2 of Section 3). Note that the GPU that performs more multiplications deals with less audio buffers, minimizing the data transfers between CPU and GPU. This triangular structure can be considered715 independently of the number of microphones. Finally, after the distributed computation of the GCC matrix, all GPUs need all of the rows of the GCC matrix in order to compute their corresponding ν/NGPU elements of the SRP matrix. The use of UVA (see Appendix A.2) allows each GPU to access other GPU via peer-to-peer over the PCI-E bus720 rather than copying data back to the host and then to another GPU. Thus, each GPU transparently accesses the memories of other GPUs by just referencing a memory location. Appendix B.1. Basic Implementation using two GPUs Using all the parallelization techniques presented in Appendix A, the SRP-725 PHAT algorithm is implemented on two GPUs as follows: 1. A parallel region is created with two CPU threads. Each CPU thread is bound with a GPU. 2. Since different audio buffers are received in the system, each CPU thread independently and asynchronously sends its corresponding audio buffers730 to its GPU by using stream parallelization. The Kernels A and the FFTs are computed for each channel inside the streams. 3. As in step 2 of Section 3, stream synchronization is addressed. Only one stream is used to compute the rows of the GCC matrix. According to Figure B.1, in the case of M = 12, one GPU would compute 35 vectors735 and the other one would compute 31 vectors. 35 01 x 02 01 x 03 02 x 03 01 x 04 02 x 04 03 x 04 01 x 05 02 x 05 03 x 05 04 x 05 01 x 06 02 x 06 03 x 06 04 x 06 05 x 06 01 x 07 02 x 07 03 x 07 04 x 07 05 x 07 06 x 07 01 x 08 02 x 08 03 x 08 04 x 08 05 x 08 01 x 09 02 x 09 03 x 09 04 x 09 01 x 10 02 x 10 03 x 10 01 x 11 02 x 11 05 x 09 05 x 10 05 x 11 05 x 12 06 x 08 07 x 08 06 x 09 07 x 09 08 x 09 06 x 10 07 x 10 08 x 10 09 x 10 06 x 11 07 x 11 08 x 11 09 x 11 10 x 11 06 x 12 07 x 12 08 x 12 09 x 12 10 x 11 11 x 12 04 x 10 04 x 11 04 x 12 03 x 11 03 x 1202 x 12 GPU 1 GPU 0 uses 11 audio buffers and performs 35 element-wise multiplications GPU 1 uses 12 audio buffers and performs 31 element-wise multiplications 01 x 12 GPU 0 01 x 02 01 x 03 02 x 03 01 x 04 02 x 04 03 x 04 01 x 05 02 x 05 03 x 05 04 x 05 01 x 06 02 x 06 03 x 06 04 x 06 05 x 06 01 x 07 02 x 07 03 x 07 04 x 07 05 x 07 06 x 07 01 x 08 02 x 08 03 x 08 04 x 08 05 x 08 06 x 08 07 x 08 07 x 09 08 x 09 07 x 10 08 x 10 09 x 10 07 x 11 08 x 11 09 x 11 10 x 11 07 x 12 08 x 12 09 x 12 10 x 11 11 x 12 06 x 10 06 x 11 06 x 12 05 x 11 05 x 1204 x 12 01 x 09 02 x 09 03 x 09 04 x 09 05 x 09 06 x 09 01 x 10 02 x 10 03 x 10 04 x 10 05 x 10 01 x 11 02 x 11 03 x 11 04 x 11 01 x 12 02 x 12 03 x 12 GPU 0 GPU 1 GPU 2 GPU 0 uses 08 audio buffers and performs 28 element-wise multiplications GPU 1 uses 10 audio buffers and performs 18 element-wise multiplications GPU 2 uses 09 audio buffers and performs 20 element-wise multiplications 01 x 02 01 x 03 02 x 03 01 x 04 02 x 04 03 x 04 01 x 05 02 x 05 03 x 05 04 x 05 01 x 06 02 x 06 03 x 06 04 x 06 05 x 06 01 x 07 02 x 07 03 x 07 04 x 07 05 x 07 06 x 07 07 x 08 07 x 09 08 x 09 07 x 10 08 x 10 09 x 10 07 x 11 08 x 11 09 x 11 10 x 11 07 x 12 08 x 12 09 x 12 10 x 11 11 x 12 04 x 10 04 x 11 04 x 12 03 x 11 03 x 1202 x 12 01 x 08 02 x 08 03 x 08 04 x 08 05 x 08 01 x 09 02 x 09 03 x 09 04 x 09 01 x 10 02 x 10 03 x 10 01 x 11 02 x 11 01 x 12 GPU 0 GPU 1 GPU 3 GPU 0 uses 07 audio buffers and performs 21 element-wise multiplications GPU 1 uses 09 audio buffers and performs 15 element-wise multiplications GPU 2 uses 10 audio buffers and performs 15 element-wise multiplications GPU 3 uses 06 audio buffers and performs 15 element-wise multiplications 05 x 09 05 x 10 05 x 11 05 x 12 06 x 08 06 x 09 06 x 10 06 x 11 06 x 12 GPU 2 01 x 02 01 x 03 02 x 03 01 x 04 02 x 04 03 x 04 01 x 05 02 x 05 03 x 05 04 x 05 01 x 06 02 x 06 03 x 06 04 x 06 05 x 06 01 x 07 02 x 07 03 x 07 04 x 07 05 x 07 06 x 07 01 x 08 02 x 08 03 x 08 04 x 08 05 x 08 06 x 08 07 x 08 01 x 09 02 x 09 03 x 09 04 x 09 05 x 09 06 x 09 07 x 09 08 x 09 01 x 10 02 x 10 03 x 10 04 x 10 05 x 10 06 x 10 07 x 10 08 x 10 09 x 10 01 x 11 02 x 11 03 x 11 04 x 11 05 x 11 06 x 11 07 x 11 08 x 11 09 x 11 10 x 11 01 x 12 02 x 12 03 x 12 04 x 12 05 x 12 06 x 12 07 x 12 08 x 12 09 x 12 10 x 11 11 x 12 GPU 0 uses 12 audio buffers and performs 66 element-wise multiplications GPU 0 Figure B.1: Distribution of the audio buffers in order to compute the rows of the GCC matrix when NGP U is 1,2,3 and 4. 36 Table B.1: Speed up between strategies. rx,ry M = 6 M = 12 M = 24 M = 48 0.01 30.097 35.443 36.968 31.649 0.05 12.259 24.043 31.291 43.313 0.1 4.815 9.861 15.310 21.249 4. By using UVA, each GPU has access to the whole GCC matrix in order to compute ν/2 elements of the SRP matrix and locates a maximum value among the computed elements. 5. Each GPU transfers back to the CPU its maximum value and its location740 inside the SRP matrix. Then, a synchronization barrier for both CPU threads is set followed by an openMP section that is only executed by the master thread. This thread compares the two maximum values and chooses the greatest one, getting its location. This location indicates the sound source position.745 Appendix B.2. Comparison between strategies Table B.1 shows the speed up that the implementation strategy presented in section 3.3 achieves with respect to the strategy presented in Appendix B. Two important aspects significantly penalize the performance of this strategy in comparison with the strategy in section 3.3. First, since each GPU does not750 contain the whole GCC matrix, each GPU must access the global-memory of the other GPU in order to compute the SRP matrix; second, after computing the corresponding elements of the GCC matrix, both GPUs must be synchronized. 37 Introduction Sound Source Localization: SRP-PHAT Algorithm SRP-PHAT Implementation Computational Cost Algorithm Parallelization for real-time GPU implementation Memory considerations Multi-GPU Parallelization Basic Implementation using two GPUs Experiments and Performance Localization Performance Computational Performance Conclusion GPU and CUDA Streams on GPU Multi-GPU programming with multicore Multi-GPU Parallelization strategy involving GCC and SRP matrices Basic Implementation using two GPUs Comparison between strategies 
work_k6nztj3ypnbe7jzgie7uac7swq	----	Expositing stages of VPRS analysis in an expert system: Application with bank credit ratings Expositing stages of VPRS analysis in an expert system: Application with bank credit ratings Benjamin Griffiths, Malcolm J. Beynon* Cardiff Business School, Cardiff University, Colum Drive, Cardiff, CF10 3EU, Wales, UK Abstract The variable precision rough sets model (VPRS) along with many derivatives of rough set theory (RST) necessitates a number of stages towards the final classification of objects. These include, (i) the identification of subsets of condition attributes (b-reducts in VPRS) which have the same quality of classification as the whole set, (ii) the construction of sets of decision rules associated with the reducts and (iii) the classification of the individual objects by the decision rules. The expert system exposited here offers a decision maker (DM) the opportunity to fully view each of these stages, subsequently empowering an analyst to make choices during the analysis. Its particular innovation is the ability to visually present available b-reducts, from which the DM can make their selection, a consequence of their own reasons or expectations of the analysis undertaken. The practical analysis considered here is applied on a real world application, the credit ratings of large banks and investment companies in Europe and North America. The snapshots of the expert system presented illustrate the variation in results from the ‘asymmetric’ consequences of the choice of b-reducts considered. q 2005 Published by Elsevier Ltd. Keywords: Bank ratings; Data visualisation; Decision rules; Expert system; VPRS 1. Introduction Since the introduction of Rough Set Theory (RST) about twenty years ago (Pawlak, 1982, 1991), it has become a popular technique for the classification of objects (Tsumato, Slowinski, Komorowski, & Grzymala-Busse, 2004). Its popularity is a direct consequence of its operational processes, which adhere most closely to the notions of knowledge discovery and data mining (Li & Wang, 2004). These include; operating on the data to identify facts, only from that data which has been utilised and there are no externally imposed assumptions on the data (Jensen, 2004). There is for example no need for normally distributed attribute values as in multivariate discriminant analysis (see Lin & Piesse, 2004). These issues mitigate external concerns placed on a decision maker (DM), moreover they leave the decision maker to undertake their particular analysis (based on a known research theme). An illustration of the popularity of RST has been its nascent development (Alpigini, Peters, Skowron, & Zhong, 2002; Tsumato et al., 2004), which has included advances in the areas such as medical applications, bioinfornmatics, image recognition and information retrieval. Here the variable precision rough set (VPRS) approach is considered, which as its name may suggest allows for a level of miss- classification to exist in the decision rules constructed (see Ziarko, 1993a, b; Beynon, 2001). Moreover, central to this study is the exposition of an expert system that undertakes the various stages of VPRS analysis, with emphasis on the appropriate interaction with the DM throughout. The particular application of the described VPRS expert system, is in the area of bank ratings. Moreover, North American and European banks that have been assigned Moody’s Bank Financial Strength Rating (BFSR) are considered (Moody’s Europe, 2004). As with the general rating problem, there is a dearth of specific ‘public’ knowledge on how the credit agencies like Moody’s and Standard and Poor’s (S&Ps) make their classification decisions (Singleton & Surkan, 1991). This itself has encouraged analysis using a variety of techniques, including multiple regression models (Horrigan, 1966; Molinero, Gomez, & Cinca, 1996; West, 1970;), probit and logit models (Bouzouita & Young, 1998) and neural networks Expert Systems with Applications 29 (2005) 879–888 www.elsevier.com/locate/eswa 0957-4174/$ - see front matter q 2005 Published by Elsevier Ltd. doi:10.1016/j.eswa.2005.06.008 * Corresponding author. Tel: C44 029 2087 5747; fax: C44 029 2087 4419. E-mail address: beynonmj@cardiff.ac.uk (M.J. Beynon). http://www.elsevier.com/locate/eswa (Singleton & Surkan, 1991). With the use of VPRS producing sets of readable decision rules then a novel (initial) analysis in this area is also a part of this study. As an expert system, here its purpose is many-fold; firstly to adequately undertake the necessary analysis (in this case VPRS), secondly to present the results to a DM in a way that benefits them, thirdly control over the analysis is placed fully with the DM. This control issue, with respect to VPRS in particular includes the choice of b-reduct from those identified. Moreover, VPRS is one of a number of techniques that centers around knowledge reduction, the expert system places a level of choice on this reduction with the DM themselves. In the bank rating problem there needs to be details available on the individual banks and their correct or in-correct classification. The snapshots of the VPRS expert system presented in this study elucidate these issues. The structure of the rest of the paper is as follows: In Section 2, the rudiments of VPRS are described with initial snapshots of the expert system utilized on a small example data set. In Section 3, the bank data set is described. In Sections 4 and 5, the VPRS expert system is applied to the aforementioned bank data set, including snapshots describ- ing the three stages of VPRS outlined previously. In Section 6, conclusions are given as well as directions for future research. 2. VPRS and concomitant expert system (using a small example) This section briefly describes fundamental aspects of VPRS through a small example and use of the concomitant expert system, for further reading on VPRS and related methodologies see Ziarko (1993a, b), Beynon (2001); Tsumato et al. (2004). The information system, central to VPRS, is made up of objects described by sets of condition (C) and decision attributes and (D). Objects characterized and categorized by the same condition and decision attribute values are grouped into condition (E(C)) and decision (E(D)) classes, respectively. 1 The VPRS expert system described in this study includes the ability to view such an information system, see Fig. 1. In Fig. 1, the presented information system is made up of seven objects, described by six condition and one decision attribute. The 0 and 1 descriptor values 2 enable the construction of condition and decision classes. In summary, the condition classes of objects (using all the condition attributes, C) as groupings (X1, X2, X3, X4 and X5) of indiscernible objects are: X1 Z fo0g; X2 Z fo1; o4; o6g; X3 Z fo2g; X4 Z fo3g and X5 Z fo5g: Similarly, the decision classes (Y0 and Y1) associated with the decision attribute are: Y0 Z fo3; o4; o5; o6g and Y1 Z fo0; o1; o2g: The relationship (memberships of objects) between these condition and decision classes is fundamental to VPRS (and other RST based methodologies). Subject to a majority inclusion relation with threshold value b2(0.5, 1] (see An, Shan, Chan, Cercone, & Ziarko, 1996), VPRS looks at the condition classes that can be considered associated (classified) with a particular decision class. Moreover, those condition classes which have the property that the largest group proportion of objects classified to a decision class is at least b. The proportion of the objects in these classified condition classes is defined as the quality of the classification, more formally defined by Ziarko (1993a, b): g b ðP; DÞ Z cardgPrðZjXiÞRbfXi 2EðPÞg cardðUÞ ; where Z 2EðDÞ and P 4C; for a specified value of b. 3 An integral part of VPRS is rule construction through the use of b-reducts which are particular subsets of condition attributes (P) providing the classification of objects with the same gb(P, D) as the full set of attributes C (ibid.). Formally in VPRS, a b-reduct (RED b (P, D)), has the twin properties: (1) gb(C, D)Zgb(REDb (C, D), D), (2) No proper subset of RED b (C, D), subject to the same b value can also give the same quality of classification. Based on an identified b-reduct, a number of different sets of decision rules can be identified (maximal, minimal - see An et al., 1996). Here the minimal rule set is constructed, found through the identification of prime implicants and then further reduction in conditions (descriptor values) in the individual rules (ibid.). The details of VPRS presented so far, are its near original form (from Ziarko, 1993a, b), recent developments include those reported in Beynon (2001); Mi, Wu, and Zhang (2004) and Li and Wang (2004). The VPRS expert system exposited here incorporates the work in Beynon (2001), where information over the whole b domain (0.5, 1.0] is utilised. With a possible infinite number of different b values identified to choose from, more than one b-reduct (finite number) may be identified. For the information 1 These are equivalence classes E($), where a single member is made up of a group of objects indiscernible with each other based on the attribute values it is associated with (see Beynon, 2001). 2 VPRS works on a level of discretised data (rather than continuous values), hence may require the lowering of the level of granularity of the considered data set (see Beynon, 2004). 3 The full description of VPRS involves the identification of certain approximation regions (positive, boundary and negative), each containing groups of objects, see Beynon (2001) and Ziarko (1993a, b). For brevity this paper purposely limits the level of formal notation described herein. B. Griffiths, M.J. Beynon / Expert Systems with Applications 29 (2005) 879–888880 https://isiarticles.com/article/23289 
work_k6us6ymj6vhtnbvdivfxfp7nby	----	doi:10.1016/j.eswa.2006.08.011 www.elsevier.com/locate/eswa Expert Systems with Applications 33 (2007) 1063–1075 Expert Systems with Applications Development of an expert system for fault diagnosis in scooter engine platform using fuzzy-logic inference Jian-Da Wu a,*, Yu-Hsuan Wang a, Mingsian R. Bai b a Graduate Institute of Vehicle Engineering, National Changhua University of Education, 1 Jin-De Road, Changhua City, Changhua 500, Taiwan, ROC b Department of Mechanical Engineering, National Chiao-Tung University, Taiwan, ROC Abstract In the present study, a fault diagnosis system using acoustic emission with an adaptive order tracking technique and fuzzy-logic inter- ference for a scooter platform is described. Order tracking of acoustic or vibration signal is a well-known technique that can be used for fault diagnosis of rotating machinery. Unfortunately, most of the conventional order-tracking methods are primarily based on Fourier analysis with the revolution of the machinery. Thus, the frequency smearing effect often arises in some critical conditions. In the present study, the order tracking problem is treated as the tracking of frequency-varying bandpass signals and the order amplitudes can be cal- culated with high resolution. The order amplitude figures are then used for creating the data bank in the proposed intelligent fault diag- nosis system. A fuzzy-logic inference is proposed to develop the diagnostic rules of the data base in the present fault diagnosis system. The experimental works are carried to evaluate the effect of the proposed system for fault diagnosis in a scooter platform under various operation conditions. The experimental results indicated that the proposed expert system is effective for increasing accuracy in fault diag- nosis of scooters. � 2006 Elsevier Ltd. All rights reserved. Keywords: Fault diagnosis; Fuzzy logic inference; Adaptive order tracking; Acoustic emission 1. Introduction The technique of early fault diagnosis is used to prevent serious damages in a mechanical system. Rotating machin- ery such as internal combustion engines, cooling fans and air compressors can have their vibration and acoustic emis- sion signals monitored for early fault diagnosis. Conven- tional fault diagnosis using vibration and acoustic emission signals already exists in the form of techniques for applying the time and frequency domain of signals, and analyzing the amplitude difference in signals (Peng & Chu, 2004; Zheng, Li, & Chen, 2002). However, the con- ventional methods are not always effective for application under certain critical conditions such as using a fixed sam- pling frequency for fast Fourier transform (FFT) analysis 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.08.011 * Corresponding author. E-mail address: jdwu@cc.ncue.edu.tw (J.-D. Wu). in rapid fluctuation of the machinery speed. Order tracking is a well-known technique that can be used for fault diag- nosis of rotating machinery. Conventional order tracking technology is based on the frequency analysis, where nar- row band spectra of the frequency-varying signals utilize proper time-windowing, unfortunately, the smearing prob- lem that include spaced orders and crossing orders often arise critical conditions (Vold & Leuridan, 1993). In order to overcome this problem, a number of new techniques have been proposed, e.g. adaptive order tracking tech- niques (Bai, Huang, Hong, & Su, 2005; Bai, Jeng, & Chen, 2002), and adaptive wavelet analysis techniques (Lin & Zuo, 2003; Tse, Yang, & Tam, 2004). Most of the methods can extract the order of features embedded in the vibration and acoustic emission signals. In the present study, a high resolution adaptive order tracking technique is used to create different order ampli- tude figures of the scooter platform under various engine mailto:jdwu@cc.ncue.edu.tw 1064 J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 operation conditions as the data bank for the fault diagno- sis expert system (Wu, Huang, & Huang, 2004; Wu, Huang, & Chen, 2005). In the proposed technique, a more accurate description of signal feature can be drawn by con- verting frequency spectra into order spectra, and order is defined as the frequency normalized with various shaft speeds. The technique focuses on adaptive filter based on the Kalman filter. The continuous acoustic emission signal is measured and analyzed using adaptive Kalman filtering order tracking that can eliminate noise and reconstruct the acoustic emission signal in this system. High resolution y(n) Kalman gain computer One-step pre Riccati equation solver G(n) 1Z I Initial condition K(1,0) Initial conditio K(n,n-1) Fig. 1. Block diagram of a Fig. 2. Signal-flow graph representa adaptive order tracking is also used to practically analyze and track the energy of order signal from the dynamic sig- nal. Meanwhile, the order tracking is formulated in terms of state space models. The multiple harmonics of acoustic emission signals have been used as different order ampli- tude figures of the scooter in this system. The amplitudes of different order can be calculated for extracting the fea- ture of the order figures. After processing signal analysis that obtained order features, the fuzzy-logic inference is used to classify the faults automatically. Many machinery fault diagnostic approaches use automatic diagnosis in x̂( | )nn y 0x̂(1 | y ) dictor F(n,n+1) nx̂( 1 | )n y+ K(n+1,n) n daptive Kalman filter. tion of adaptive order tracking. 0 1 a m em be rs hi p 0.5 0 1 a m em be rs hi p a -n S.D. a +n S.D. a -n S.D. a +n S.D. Fig. 3. Membership function: (a) triangular membership function and (b) p membership function. J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 1065 order to increase accuracy caused by subjective human determination. Fuzzy logic is a powerful tool for complex numerical analysis and knowledge modeling. (Huang, Table 1 Rules of fuzzy logic inference for the fault diagnosis system If Then 1X 2X � � � 7X 8X Normal Low 1 Low 2 � � � Low 1 Low 1 • Low 1 High 3 � � � Low 2 Low 2 High 2 High 1 � � � Low 2 Low 1 Low 3 High 2 � � � Low 2 Low 1 High 2 High 1 � � � Low 1 Low 1 Tachometer signals Order tracking Fuzz Infere Features Signal processing Knowledge base Infere rul Microphone signals Fig. 4. Experimental arrangemen Yang, & Huang, 1997) proposed a fuzzy logic system for the fault diagnosis system. (Mechefske, 1998) also pro- posed a fuzzy logic inference for fault classification by machinery vibration signal. In the present study, a fuzzy-logic inference is proposed to develop the diagnostic rules of the data base in an order tracking amplitude features. The fuzzy logic inference can calculate complex numerical analysis and use a member- ship value easily explained by human. The membership function establishes the inference rules and the knowledge base to systematize the post-processing procedure in this system. Both the p and the triangular membership function will be compared for fault diagnosis by using order ampli- tude figures of the scooter with normal and four fault con- ditions. The structure and principle of high resolution adaptive order tracking technique and fuzzy logic inference are described in the following sections. 2. Principle of adaptive order tracking technique In this section, an adaptive order tracking technique based on the Kalman filtering algorithm is described (Hay- kin, 1996). The analysis of the order tracking with the acoustic emission signal and shaft speed can be described as the amplitude of the frequency-modulated (Bai et al., 2002; Bai et al., 2005). The acoustic emission signal x(t) of the rotating machinery containing k orders is usually represented as Fault 1 Fault 2 Fault 3 Fault 4 • • • • Decision Operator y nce nce es t and fault diagnosis system. Fig. 5. Revolution of scooter in run-up test condition. Table 2 Membership value of consequence with p membership function at idle condition with pulley damaged Faults n (multiples of S.D.) 1 2 3 4 5 Normal 0.0030 0.0116 0.0243 0.0396 0.0562 Pulley 0.5665 0.8202 0.9078 0.9451 0.9639 Belt 0.0853 0.1826 0.2572 0.3187 0.3712 Air 0.0892 0.1185 0.1316 0.1430 0.1552 Clutch 0.0250 0.0780 0.1309 0.1752 0.2115 Table 3 True membership values in the fault diagnosis system Test faults The result of faults Normal Pulley Belt Air Clutch Normal 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 1 0 0 0 Pulley 0 1 0 0 0 0 1 0 0 0 0 0 1 0 0 Belt 0 0 1 0 0 0 0 1 0 0 0 0 0 1 0 Air 0 0 0 1 0 0 0 0 1 0 0 0 0 0 1 Clutch 0 0 0 0 1 0 0 0 0 1 1066 J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 xðtÞ¼ A1 cos½hðtÞþ /1�þ A2 cos½2hðtÞþ /2�þ � � � � � � þ Ak cos½khðtÞþ /k�: ð1Þ Eq. (1) can be expanded as xðtÞ¼ A1I cos½hðtÞ�� A1Q sin½hðtÞ�þ �� �þ AkI cos½khðtÞ� � AkQ sin½khðtÞ�: ð2Þ By discrete time sampling, Eq. (2) can be represented as xðnÞ¼½cos½hðnÞ��sin½hðnÞ����cos½khðnÞ��sin½khðnÞ�� A1I A1Q .. . AkI AkQ 2 66666664 3 77777775 : ð3Þ The high-resolution order tracking based on Kalman filter has two equations. The first equation is called the process equation: Fig. 6. Order figures displayed for xðn þ 1Þ¼ Fðn þ 1; nÞxðnÞþ v1ðnÞ: ð4Þ engine speed at idle condition. Fig. 7. Performance of membership function versus the parameter n at idle condition: (a) triangular membership function and (b) p membership function. Fig. 8. Order amplitude figures of scooter acoustic emission in idle condition. A solid line depicts scooter without any fault; a dashed line depicts scooter with various faults: (a) pulley damaged, (b) belt damaged, (c) air leakage of intake and (d) clutch damaged. J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 1067 Meanwhile, the second equation is called the measurement equation: yðnÞ¼ CðnÞxðnÞþ v2ðnÞ ð5Þ The vectors v1(n) and v2(n) represent zero-mean and white noise. The noise vectors v1(n) and v2(n) are indi- vidually independent and E½v1ðnÞvH2 ðkÞ� ¼ 0 for all n and k. 1068 J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 The wave of the amplitude-modulated signal can track its frequency continually with respect to the fundamental and multiple frequencies in the rotating machinery. Using this form, the Kalman filtering algorithm is recommended to structure this model by the process and measurement equations. Measurement equation: yðnÞ� XN i¼�N AiðnÞexp½jhiðnÞ� ¼ v2ðnÞ: ð6Þ Process equation: r2AiðnÞ¼ v1ðnÞ: ð7Þ In common usage, Ai(n) can be assumed by the autoregres- sive (AR) model. HðzÞAiðzÞ¼ V 1ðzÞ: ð8Þ The state transition matrix and measurement matrix are described as Fig. 9. Order amplitude figures of scooter acoustic emission in 2000 rpm. A sol various faults: (a) pulley damaged, (b) belt damaged, (c) air leakage of intake Fðnþ1;nÞ¼ P 0 ��� 0 0 . . . 0 ��� 0 0 0 P 0 0 0 ��� 0 . . . 0 0 ��� 0 P 2 66666664 3 77777775 ; ð9Þ CðnÞ¼ exp½jhkðnÞ� 0 ��� exp½jh0ðnÞ� 0 ��� exp½jh�kðnÞ� 0½ �: ð10Þ The Kalman filtering procedure can be summarized as follows: GðnÞ¼ Fðn þ 1; nÞKðn; n � 1ÞCHðnÞ � ½CðnÞKðn; n � 1ÞCHðnÞþ Q2ðnÞ� �1 ; ð11Þ aðnÞ¼ yðnÞ� CðnÞx̂ðnjyn�1Þ; ð12Þ x̂ðn þ 1jynÞ¼ Fðn þ 1; nÞx̂ðnjyn�1Þþ GðnÞaðnÞ; ð13Þ Kðn þ 1; nÞ¼ Fðn þ 1; nÞKðnÞFHðn þ 1; nÞþ Q1ðnÞ: ð14Þ The representation of a block diagram in the Kalman filter is depicted in Fig. 1. The signal-flow graph of adap- tive order tracking and the identification of parameters by adaptive order tracking are represented in Fig. 2. By id line depicts scooter without any fault; a dashed line depicts scooter with and (d) clutch damaged. Fig. 10. Order amplitude figures of scooter acoustic emission in 2500 rpm. A solid line depicts scooter without any fault; a dashed line depicts scooter with various faults: (a) pulley damaged, (b) belt damaged, (c) air leakage of intake and (d) clutch damaged. J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 1069 adding the Kalman filtering algorithm we can eliminate noise and reconstruct the sound signals in this system. This proposed adaptive algorithm could increase the accuracy of the order amplitude figures and conveniently extract the features. 3. Fuzzy logic inference for fault diagnosis Fuzzy logic is a useful approach to simplify a complex system in engineering application. In the present study, a fuzzy-logic inference is used to calculate complex numerical analysis with a membership value easily interpreted by humans (Awad & Wafik, 1999). After extracting the fea- tures of the proposed adaptive order tracking amplitude figures, fuzzy logic is used to automatically diagnose the faults in the designed scooter experimental platform. The fuzzy-logic inference is proposed to establish the diagnostic rules of the data bank in this fault diagnosis system. The amplitudes of different order can be calculated for extract- ing the features of the order figures. The energy of order figure amplitude is calculated by using root-mean-square (rms) as the input values: W ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 H XH h¼1 YðhÞ !vuut ð15Þ where W is the rms value of order amplitude, Y is an order amplitude, and H is the number of the order amplitude. After obtaining the order feature of the order figures, the data analysis includes finding the mean value and standard deviation of each order after the extracting feature proce- dure has finished. The decision of the membership function is also a key point and especially punctilious for fuzzy-logic inference. The fuzzy membership function has represented variable sets that include S, p, triangular and gauss. In the present study, the triangular and p membership function will be used to develop the fuzzy logic inference (Mech- efske, 1998). The triangular and p membership functions are established as shown in Fig. 3. The equation of p mem- bership function can be described as 1070 J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 pðz; a; bÞ¼ 1 1 þðz�a b Þ2 ; ð16Þ where a is the mean value of each order, b is the n times the standard deviation (S.D.) in each order. The membership function establishes the inference rules and the knowledge base under fault decision conditions. The rules of fuzzy logic inference are defined as ±n times the standard deviation of each condition at each order. Those are used for the higher and lower limits of the fuzzy logic member- ship domain. The rules have been defined an inference rule variable x in universe of the discourse, and the equation follows as: ðA1 \ A2 \�� �\ AnÞðxÞ¼ minfA1ðxÞ; A2ðxÞ:::AnðxÞg; for all x 2 X ; ð17Þ where A(x) is a rule of the membership function. Eq. (17) can be represented the Table 1. The processing procedure must be systematized with the inference rules in this system. Fig. 11. Order amplitude figures of scooter acoustic emission in 3000 rpm. A so various faults: (a) pulley damaged, (b) belt damaged, (c) air leakage of intake The fuzzy membership domain is defined with various strength ranges from 0 to 1. The condition of maximal probability will be inferred according to the knowledge base and evaluation of the inference rules. The fuzzy logic inference could increase the efficiency of diagnosing faults in the present system. 4. Experimental investigation of fault diagnosis system In the experimental investigation, the scooter acoustic emission signals were analyzed to verify the proposed fuzzy logic inference fault diagnosis system. The fault diagnosis system consists of the signal analysis and the fault infer- ence. In the signal analysis, the adaptive order tracking with Kalman filter for extracting the order feature of acoustic emission signal is used, and fault inference utilizes a fuzzy logic approach to classify faults under different operating conditions. Experimental arrangement and the fuzzy logic system of the scooter defect diagnosis system lid line depicts scooter without any fault; a dashed line depicts scooter with and (d) clutch damaged. J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 1071 is shown Fig. 4. A scooter with an electronic fuel injection system, single-cylinder, four-stroke, 0.125-l internal com- bustion engine is used in this application. The sound emis- sion signal is measured using a condenser microphone (PCB 130D20) located close to the scooter engine transmis- sion. A fiber optical sensor (PW-PH02) is utilized to detect the revolution signal and angular displacement of the engine as reference signals of the order tracking procedures in the diagnostic system. The sampling frequency is chosen to be 20 kHz with the data record system (Hardware: NI 6024E; Software: Lab-view). In the present study, five conditions of the scooter are used in the experimental work. These include without fault, pulley damage, belt damage, air leakage of the intake man- ifold and clutch damage. In the experimental work, there are 10 pieces of data in each fault condition to establish the fuzzy-logic inference. The engine is operated in an idling condition (1700 rpm), 2000 rpm, 2500 rpm, 3000 rpm, 3500 rpm and run-up test conditions. The engine revolution of the run-up condition is shown as Fig. 5. In Fig. 12. Order amplitude figures of scooter acoustic emission in 3500 rpm. A so various faults: (a) pulley damaged, (b) belt damaged, (c) air leakage of intake particular, the run-up test is emphasized in this research because the adaptive order tracking can accurately track the energy of the order signal from the dynamic signal. After recording the signals, they would be used in order tracking analysis to track the energy of the order amplitude signal. Acoustic emission signal features could be drawn by converting the frequency spectrum into order spectrum, and order is defined as the frequency normalized with various shaft speeds. The first eight orders of the amplitude figures with adaptive Kalman filter under the idle condition without any fault in the scooter platform are shown as Fig. 6. The amplitudes of the order figures can be calcu- lated for the feature as input for fuzzy logic inference. After extracting the features, the fuzzy logic inference approach is used in the proposed fault diagnosis. First, the mean and the standard deviation of each condition at each order are estimated, and the higher and lower limits of the fuzzy logic membership domain as the fuzzy infer- ence rules are defined. There are eight orders in one fault condition which means there are eight inference rules in lid line depicts scooter without any fault; a dashed line depicts scooter with and (d) clutch damaged. Fig. 13. Order amplitude figures of scooter acoustic emission in run-up condition. A solid line depicts scooter without any fault; a dashed line depicts scooter with various faults: (a) pulley damaged, (b) belt damaged, (c) air leakage of intake and (d) clutch damaged. 1072 J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 the system. After calculation by the inference rules and knowledge base, the results have been represented by mem- bership values. The membership value of the fault diagno- sis system is between 0 and 1, where 0 means impossible and 1 means possible. When the membership value is close to 1, this means the condition is true. The inference rules and knowledge base have been established based on the membership function. In this research, the triangular and p membership functions are proposed to develop the fuzzy-logic inference. The triangular membership function applies when tolerance of fault is large and the p member- ship function is applied in critical conditions. Table 4 The optimum n values with two different membership function under various n Value Engine operation conditions Membership function Idle 2000 rpm 2500 rpm Triangular 4 3 3 p 2 2 2 Besides the membership function, choosing the opti- mum inference error versus n value could also increase the resolution of system. With fuzzy logic inference, the membership value of consequence with p membership function in the idling condition with pulley damage are cal- culated and summarized in Table 2. When the n is larger the membership domain is wider. The membership value of the accurate classification increases with an increasing n value. Nevertheless, the membership value of fault classi- fication also increases with an increasing n value. Defining optimum inference error versus n value is necessary. The optimum inference error versus n value is determined by: operation conditions 3000 rpm 3500 rpm Run-up 4 4 2 2 2 1 Table 5 Faults classification under various engine operation conditions with triangular membership function Engine operation Test faults Membership value of faults Normal Pulley Belt Air Clutch Idle Normal 0.7746 0.0413 0.1220 0.1953 0.0598 Pulley 0 0.7709 0.1412 0.1047 0.0462 Belt 0.1315 0.3227 0.8087 0.0929 0.1845 Air 0.1163 0.1629 0.0494 0.8626 0.1210 Clutch 0.1733 0.1653 0.1248 0.2085 0.6836 2000 rpm Normal 0.7258 0.0273 0.0991 0.1997 0.1155 Pulley 0 0.8760 0 0.0766 0.1163 Belt 0.1006 0.1474 0.7555 0.0380 0.2292 Air 0.1983 0.1846 0 0.7205 0.0187 Clutch 0.0975 0.1452 0.1899 0.0056 0.7119 2500 rpm Normal 0.7703 0.0676 0.0836 0.1122 0.2659 Pulley 0 0.6470 0.0640 0.2315 0 Belt 0.0805 0.2228 0.7233 0.1323 0.2442 Air 0 0.1207 0.0555 0.7630 0.0584 Clutch 0.2621 0.0715 0.2556 0.0744 0.7270 3000 rpm Normal 0.7551 0 0.0651 0 0.3291 Pulley 0 0.8143 0.1052 0.1298 0.0825 Belt 0 0.0972 0.7559 0.1223 0.3124 Air 0 0.1642 0.1134 0.8489 0.2022 Clutch 0.0341 0 0.2208 0.0749 0.7573 3500 rpm Normal 0.7441 0 0.1331 0.2122 0.2880 Pulley 0 0.7746 0.1247 0 0 Belt 0.0824 0.0886 0.8140 0.0324 0.4255 Air 0.1327 0 0.1542 0.7590 0.3111 Clutch 0.0644 0 0.2380 0.2244 0.8184 Run-up Normal 0.7265 0.0517 0.1525 0.2855 0.1209 Pulley 0.0118 0.4940 0.2737 0.1444 0.1097 Belt 0.1121 0.3728 0.6823 0.1996 0.1259 Air 0.1577 0.2550 0.2338 0.6522 0.0454 Clutch 0.1678 0.2025 0.1815 0.2047 0.7268 J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 1073 error ¼ Xl i¼1 Xm j¼1 ðd ij � d̂ ijÞ 2 ; ð18Þ where dij is the membership value of the ith calculation re- sult that is inferred to the jth fault. d̂ ij is the true member- ship value of the ith practical data relative to the jth fault. The parameters l and m are the numbers of the fault con- dition and test data, respectively. Assuming every fault condition has three data to establish a knowledge base, the membership values of the true consequence form d̂ ij in this system are summarized in Table 3. The optimum n value is the minimum error in estimation of the knowl- edge base. Fig. 7(a) is the performance of membership function versus the parameter n with the triangular mem- bership function under the idling engine operation condi- tion. The definition of optimum inference error versus n value with the p membership function under the idling operation condition is shown as Fig. 7(b). In the figure, the optimum n = 2 is relative to Table 2. When n = 2 the results of fault conditions have a higher resolution with the p membership function at an idling engine oper- ation schedule system. 5. Results and discussion This section will describe the results of the proposed fault diagnosis system in a scooter platform under various operating conditions. A conventional adaptive order track- ing fault diagnosis technique is carried out using visual inspection of the order amplitude figure under different fault conditions. Obviously, different faults in the mechan- ical system occur with different amplitudes in order figures. Unfortunately, the conventional inspection is not a precise approach for damage diagnosis. In the present study, the order amplitude figures of scooter acoustic emission are used for the data bank in the proposed fuzzy-logic infer- ence for the intelligent fault diagnosis system. The order amplitude figures of scooter acoustic emission in the idling condition with various faults in the platform are shown as Fig. 8. Although by observing Fig. 8(a), we can infer normal conditions and pulley damage conditions using visual comparison, however, for other fault conditions such as belt damage, air leakage of intake and clutch dam- aged as shown in Fig. 8(b)–(d), there are too many order figures to infer the fault conditions using only visual 1074 J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 comparison. Therefore, the fuzzy-logic inference is used to automatically increase the accuracy of the classification faults in the intelligent fault diagnosis system. The data bank also includes other amplitude figures of order track- ing under various operating conditions shown in Figs. 9– 13, which includes 2000 rpm, 2500 rpm, 3000 rpm, 3500 rpm and run-up engine operating conditions. Therefore, this research employs fuzzy-logic inference to establish data bank verification of the synthetic system under various operating conditions. The different data banks organize the inference rules and knowledge base of fuzzy logic infer- ence respectively. Estimating the optimum inference error versus n value can increase the resolution of the diagnostic system. Under various operating conditions with two different membership functions, inference rules evaluate the optimum n values which are summarized in Table 5. According to the results of Table 4, the membership values are calculated with fuzzy logic inference accurately and conveniently. Meanwhile, the triangular and p membership functions are compared in the fuzzy-logic inference under various engine operating conditions. The results of faults classifica- tion under various engine operating conditions using the triangular and p membership functions are summarized Table 6 Faults classification under various engine operation conditions with p membe Engine operation Test faults Membership value of fau Normal Pulle Idle Normal 0.8098 0.01 Pulley 0.1437 0.82 Belt 0.1846 0.18 Air 0.2287 0.11 Clutch 0.1379 0.07 2000 rpm Normal 0.8385 0.13 Pulley 0.0144 0.94 Belt 0.1183 0.32 Air 0.2941 0.30 Clutch 0.1246 0.30 2500 rpm Normal 0.8726 0.14 Pulley 0.0065 0.77 Belt 0.1143 0.35 Air 0.0785 0.23 Clutch 0.3635 0.19 3000 rpm Normal 0.7920 0.03 Pulley 0.0046 0.85 Belt 0.0319 0.15 Air 0.0175 0.20 Clutch 0.0762 0.04 3500 rpm Normal 0.7912 0.05 Pulley 0.0019 0.82 Belt 0.1065 0.16 Air 0.2133 0.06 Clutch 0.1153 0.07 Run-up Normal 0.7671 0.13 Pulley 0.0550 0.52 Belt 0.1757 0.43 Air 0.1740 0.28 Clutch 0.2276 0.34 in Tables 5 and 6. The data showed that both the triangular and p membership functions are effective for fault classifi- cation in various fault conditions. In particular, the run- up operation schedule is emphasized in this research, the scooter engine can be operated by running up or casting down under practical condition. The run-up operation schedule has fine result by inferring the faults successfully. The triangular membership function is applied when tol- erance of fault is large and the p membership function is applied in certain condition. From Table 4, we can find the p membership function with the same various engine operations has smaller n values than the triangular mem- bership function. This is because the features of adaptive order tracking have critical tolerance of faults in the same fault conditions. From Tables 5 and 6, the p membership function has higher membership values of accurate classifi- cation than the triangular membership function at each fault condition under the idling engine operation. In other operations, the p membership function also has high reso- lution. The p membership function is more capable of han- dling the qualities for this experimental work than the triangular membership function. But the results of the two membership functions provide evidence that fuzzy logic inference is a useful fault diagnosis system. rship function lts y Belt Air Clutch 16 0.1478 0.1391 0.2003 02 0.3903 0.2634 0.2757 26 0.8476 0.0968 0.2021 85 0.1092 0.9104 0.2538 80 0.2120 0.1643 0.7143 34 0.1617 0.3303 0.1933 57 0.0693 0.1061 0.1523 64 0.8637 0.0897 0.3306 46 0.0992 0.8417 0.1157 66 0.3683 0.1004 0.8129 40 0.1532 0.2633 0.3881 55 0.1146 0.3110 0.0321 65 0.8353 0.2370 0.2992 26 0.1079 0.8817 0.1158 09 0.3511 0.1980 0.8430 21 0.0997 0.0374 0.4237 55 0.1441 0.1634 0.1047 51 0.7932 0.1680 0.3456 44 0.1544 0.8962 0.2562 07 0.2520 0.1149 0.8030 91 0.1843 0.2560 0.3564 08 0.1609 0.0086 0.0330 53 0.8705 0.0828 0.4825 46 0.2111 0.8044 0.3707 62 0.3251 0.2481 0.8724 95 0.2233 0.2603 0.1850 06 0.2954 0.1866 0.2206 17 0.7016 0.2116 0.2671 49 0.3192 0.7006 0.1295 14 0.2555 0.2922 0.7214 J.-D. Wu et al. / Expert Systems with Applications 33 (2007) 1063–1075 1075 6. Conclusions In the present study, a prototype of an expert system for fault diagnosis in scooters platform using fuzzy-logic inter- ference with adaptive order tracking technique is devel- oped. The adaptive order tracking based on Kalman filter extracts the order features of the scooter test platform. They then are used for creating the data bank in the pro- posed intelligent fault diagnosis system. Fuzzy logic infer- ence is used for fault classification in the proposed system. The experimental results indicated that the pro- posed system is effective for increasing the accuracy of fault diagnosis under various operation conditions. Acknowledgement The study was supported by the National Science Coun- cil of Taiwan, the Republic of China, under project number NSC-94-2218-E-018-001. References Awad, S. H., & Wafik, B. L. (1999). A fuzzy logic approach to the selection of cranes. Automation in Construction, 8(5), 597–608. Bai, M., Huang, J., Hong, M., & Su, F. (2005). Fault diagnosis of rotating machinery using an intelligent order tracking system. Journal of Sound and Vibration, 280, 699–718. Bai, M., Jeng, J., & Chen, C. (2002). Adaptive order tracking technique using recursive least-square algorithm. American of Mechanical Engi- neers, Journal of Vibrations and Acoustics, 124(4), 502–511. Haykin, S. (1996). Adaptive filter theory. New York: Prentice-Hall. Huang, Y. C., Yang, H. T., & Huang, C. L. (1997). Developing a new transformer fault diagnosis system through evolutionary fuzzy logic. IEEE Transactions on Power Delivery, 12(2), 761–767. Lin, J., & Zuo, M. J. (2003). Gearbox fault diagnosis using adaptive wavelet filter. Mechanical Systems and Signal Processing, 17(6), 1259–1269. Mechefske, C. K. (1998). Objective machinery fault diagnosis using fuzzy logic. Mechanical Systems and Signal Processing, 12(6), 855–862. Peng, Z. K., & Chu, F. L. (2004). Application of the wavelet transform in machine condition monitoring and fault diagnostics: a review with bibliography. Mechanical Systems and Signal Processing, 18, 199–221. Tse, P. W., Yang, W. X., & Tam, H. Y. (2004). Machine fault diagnosis through an effective exact wavelet analysis. Journal of Sound and Vibration, 227, 1005–1024. Vold, H., & Leuridan, J. (1993). High resolution order tracking at extreme slew rates, using Kalman filters. The Society of Automotive Engineers, 931288, 219–226. Wu, J. D., Huang, C. W., & Chen, J. C. (2005). An order-tracking technique for the diagnosis of faults in rotating machineries using a variable step-size affine projection algorithm. NDT and E International, 38(2), 119–127. Wu, J. D., Huang, C. W., & Huang, R. W. (2004). An application of a recursive Kalman filtering algorithm in rotating machinery fault diagnosis. NDT and E International, 37(5), 411–419. Zheng, H., Li, Z., & Chen, X. (2002). Gear fault diagnosis based on continuous wavelet transform. Mechanical Systems and Signal Pro- cessing, 16(2–3), 447–457. Development of an expert system for fault diagnosis in scooter engine platform using fuzzy-logic inference Introduction Principle of adaptive order tracking technique Fuzzy logic inference for fault diagnosis Experimental investigation of fault diagnosis system Results and discussion Conclusions Acknowledgement References 
work_k7evadzfrff7tepoln7wxbshni	----	ESWA-D-14-04005R1 An evolutionary multi-objective optimization system for earthworks M. Parentea, P. Cortezb, A. Gomes Correiac a Corresponding author. ISISE Institute for Sustainability and Innovation in Structural Engineering / ALGORITMI Research Centre. University of Minho, School of Engineering, Azurém, 4800-058 Guimarães, Portugal (email: map@civil.uminho.pt; tel.: +351 253510200) b ALGORITMI Research Centre. University of Minho, School of Engineering, Azurém, 4800-058 Guimarães, Portugal (pcortez@dsi.uminho.pt) c ISISE Institute for Sustainability and Innovation in Structural Engineering. University of Minho, School of Engineering, Azurém, 4800-058 Guimarães, Portugal (agc@civil.uminho.pt) Abstract Earthworks involve the levelling or shaping of a target area through the moving or processing of the ground surface. Most construction projects require earthworks, which are heavily dependent on mechanical equipment (e.g., excavators, trucks and compactors). Often, earthworks are the most costly and time-consuming component of infrastructure constructions (e.g., road, railway and airports) and current pressure for higher productivity and safety highlights the need to optimize earthworks, which is a nontrivial task. Most previous attempts at tackling this problem focus on single-objective optimization of partial processes or aspects of earthworks, overlooking the advantages of a multi-objective and global optimization. This work describes a novel optimization system based on an evolutionary multi-objective approach, capable of globally optimizing several objectives simultaneously and dynamically. The proposed system views an earthwork construction as a production line, where the goal is to optimize resources under two crucial criteria (costs and duration) and focus the evolutionary search (non-dominated sorting genetic algorithm-II) on compaction allocation, using linear programming to distribute the remaining equipment (e.g., excavators). Several experiments were held using real-world data from a Portuguese construction site, showing that the proposed system is quite competitive when compared with current manual earthwork equipment allocation. Keywords: earthworks, evolutionary computation, multi-objective optimization, artificial intelligence 1. Introduction In Civil Engineering, a great majority of construction projects require earthworks activities prior to the construction of any structural element. Earthworks are engineering Manuscript Click here to view linked References M. Parente, P. Cortez, A. Gomes Correia 2 processes by which the ground surface in a target area is levelled or shaped through the moving or processing of the geomaterials that comprise it. It usually involves the excavation of these geomaterials, which can then be loaded and hauled to new areas to be spread and compacted into embankments, and may also include intermediate steps, such as material treatment or layer wetting. Nowadays, technical and environmental concerns require that, whenever possible, embankment fronts be built using mostly the material excavated from the construction site itself, in order to take maximum advantage of available materials and avoid the use of other materials brought in from outside borrowing areas. Earthwork tasks are reliant on heavy mechanical equipment, namely excavators (material excavation and loading to transportation equipment), dumper trucks (transportation between excavation and embankment fronts), bulldozers (material spreading so as to allow for compaction) and compactors (Zhang, 2008; Hola & Schabowicz, 2010). This is one of the main reasons that make earthworks to often incur the highest percentage costs and durations in road and railway construction projects, implying an increased importance regarding their automated optimization (Miao et al., 2011). Under this context, it is essential to optimize all available earthwork resources under two key objectives: cost and duration. Both depend on the availability of equipment and also on equipment allocation throughout the project. This is a nontrivial task due to several reasons. There is a vast number of possible equipment allocation combinations, thus the search space is large. And while there is often high competitiveness, where the pressure is to provide the least possible costs and durations (Moselhi & Alshibani, 2007), contractors and project designers often settle for an allocation solution that is mostly based on their own intuition and experience. Moreover, trade-offs need to be set between cost and duration, since these objectives can also conflict. For instance, a less expensive solution may use less amount of equipment, which in turn will result in higher project durations. Contrariwise, allocating more equipment will increase project costs, but decrease durations. However, a prolonged use of equipment implies higher fuel consumption and maintenance expenses. Furthermore, as an earthwork construction progresses, the equipment must be optimally reallocated in order to advance through successive construction phases, characterizing this as a dynamic multi- objective optimization problem. Ideally, earthworks should be optimized automatically. Considering the nontrivial characteristics of earthworks optimization (e.g., large search space and conflicting goals), pure conventional Operational Research (e.g., linear programming) and blind search methods M. Parente, P. Cortez, A. Gomes Correia 3 are infeasible. An interesting solution is to adopt metaheuristics, which are flexible optimization methods capable of searching interesting search space regions under a reasonable use of computational resources. Several studies have adopted metaheuristics to earthwork optimization, such as genetic algorithms (Marzouk & Moselhi, 2002; Moselhi & Alshibani, 2007; Xu et al., 2011) and swarm intelligence (Kataria et al., 2005; Miao, Sun et al., 2011; Nassar & Hosny, 2012; Zhang, 2008). However, many of these studies focus on single tasks or partial processes that comprise earthworks, such as excavation and hauling (Edwards & Griffiths, 2000; Kataria et al., 2005; Nassar & Hosny, 2012; Xu et al., 2011), in an attempt to deal with the high complexity of the problem. Therefore, these systems lack the advantages of a global optimization of execution durations and costs throughout all construction phases. In terms of optimization objectives, existent systems tend to be limited to single objective optimization, such as cost (Marzouk & Moselhi, 2002) or duration (Kataria et al., 2005), or attempt to consider both objectives via a weight-based optimization (Zhang, 2008). Although these solutions are considered effective in reducing computation effort requirements, they overlook the advantages of optimizing both objectives simultaneously. Even if it can be looked at as multi-criteria optimization, the weighted-based approach used in (Zhang, 2008) only outputs a single trade-off for a particular weight combination (e.g., 0.8 for first criteria and 0.2 for second). Yet, often there is not a single optimal trade-off solution, but rather a set of trade-offs with conflicting objectives. Thus, a much natural multi-criteria optimization approach is to optimize a Pareto front of solutions, where each solution is called non-dominated, or Pareto optimal, if none of the objectives can be improved in value without worsening the other (Bonissone et al., 2009). In the context of earthwork optimization, all Pareto-optimal solutions are considered equally good and the main choice criteria for selecting one solution over the other is often decided by the project designer based on the construction final deadline and/or budget. Providing a Pareto set of optimal solutions is valuable for the construction designer, as earthworks are inherently a dynamic and thus different cost-duration solutions might become better adjusted to the ever-changing site conditions as construction develops. Moreover, additional criteria could be used to support the final decision, such as environmental aspects, which can be assessed by the determination of carbon emissions in each solution. Indeed, environmental concerns and sustainability have been a rising trend in the past few years (Chang & Tsai, 2015; Guerin, 2014; Wang, Duan, Wu, & Yang, 2014). For this reason, a flexible methodology that allows for an easy adjustment of the optimization objectives and can deal with different trade-offs is clearly M. Parente, P. Cortez, A. Gomes Correia 4 advantageous. Taking into account that Pareto front multi-optimization requires the tracking of a population of solutions, population based Metaheuristics such as evolutionary multi-objective optimization (EMO), have become natural and popular solutions. Following this trend, this work adopts a popular EMO method, the non-dominated sorting genetic algorithm-II (NSGA- II) algorithm, to tackle the multi-criteria optimization problem associated with resource allocation in earthwork construction. In contrast with previous works, a true multi-objective approach is adopted, encompassing the whole earthwork construction phase and outputting a Pareto set of interesting trade-off solutions. Moreover, a novel representation of solutions is proposed, using first an EMO to allocate compaction equipment and then a linear programming to distributing the remaining equipment (e.g., excavators and trucks). Finally, the proposed system is validated by experimenting with real-world data from a construction site and compared against conventional manual earthwork design. This paper is organized as follows. Section 2 introduces the earthworks in the context of an optimization problem, including the definition of an objective function and associated constraints. Then, Section 3 presents the adaptation of the EMO algorithm to the earthwork domain, focusing on its representation, dynamic features and algorithmic flow. Next, Section 4 details the experiments conducted and analyses obtained results when validating the proposed system using real-world earthworks data. Finally, conclusions are summarized in Section 5, which also presents future research directions. 2. Earthwork Optimization From the optimization point of view, the earthworks process can be perceived as a production line based on resources (mechanical equipment) and dependency relations between sequential tasks. The available resources can be allocated to each task in the production line, ranging from excavation and transportation to spreading and compaction equipment. Depending on the amount and type of equipment allocated and other factors, such as material type, the work rate for each task in the production line can be easily computed, since it corresponds to the sum of the work rate of assigned equipment. Ideally, the added work rate of the equipment allocated for each task should be as close as possible to that of the equipment allocated for the next task, in order to allow a constant flow of material throughout the production line. On the one hand, this prevents idle M. Parente, P. Cortez, A. Gomes Correia 5 times from incurring on the allocated equipment, in cases where the work rate of the previous task is significantly inferior to that of the succeeding task. On the other hand, in cases where the rate of the previous task is significantly superior to the succeeding task, an excessive flow of material that can ultimately obstruct movement throughout the construction site is averted. Therefore, controlling the work rate in each task within a production line is paramount. Accordingly, the main variable associated with the earthworks optimization problem is the amount of equipment allocated in each task, for each construction phase. 2.1. Problem Definition In production lines with sequential interdependent jobs, the last job determines the speed at which the whole process progresses. Considering that compaction corresponds to the last job in the earthworks production line (Figure 1), it determines the development rate of the whole construction, thus having a direct influence on project durations. Maximizing the work rate in compaction fronts would correspond to the minimal execution duration solution, provided that there is enough equipment in the remaining tasks to support such allocation. In this point of view, an earthworks construction project is divided into a number of production lines, which correspond to the total number of compaction fronts. To each compaction front corresponds a potential production line and its associated equipment, ranging from excavation to compaction tasks. These production lines can work simultaneously and are independent from each other while progressing towards completion. However, whenever a compaction front is completed, the equipment associated with that production line becomes available once again. At this point, a construction phase is considered completed and a new one ensues, demanding a reallocation of the newly available equipment. Subsequently, any optimization attempt on this type of problem must take these factors into consideration, including task interdependence and variable site conditions over time. M. Parente, P. Cortez, A. Gomes Correia 6 Since resource allocation in earthworks is a time-evolving process (dynamic task), the solution should be time-adaptive as well. More specifically, the allocation of earthwork equipment has to constantly be updated whenever any work front has finished its projected work, hence ending the current construction phase. Consequently, the variables associated with this type of optimization are the amount and type of resources (equipment) to be optimally allocated for each task and also in each construction phase. The algorithmic flow of the general methodology used to solve the allocation and dynamic features of the problem is illustrated in Figure 2. Figure 1. Earthwork equipment aligned with excavation and compaction fronts. M. Parente, P. Cortez, A. Gomes Correia 7 2.2. Objective Function and Constraints The optimal allocation of earthwork equipment is carried with the main objective of minimizing both total construction times and costs. The ideal solution would allow for the fastest possible excavation, transportation, spreading and compaction of projected volume of geomaterials with minimum expenses. For a single work front, the execution duration, Ts, can be defined by the volume of material being handled (in m3), Vm, divided by the work rate of the corresponding allocated equipment, Qp, (in m3/h), as depicted in Equation 1. Figure 2. Algorithmic flow of dynamic earthwork resource allocation. M. Parente, P. Cortez, A. Gomes Correia 8 !! = !! !! (1) However, from the production line point of view, the work rate of each job is not completely independent of that of the adjacent jobs, but rather part of a global work rate of the entire production line. In essence, considering the interdependence of tasks in the production line, the work rate of a production line corresponds to the minimum work rate from between the tasks that comprise it. This occurs because the equipment of one job cannot work at a faster rate than that of the preceding job and, simultaneously, it should not have a higher productivity than the succeeding job. A clear example might be that a group of dumper trucks cannot possibly transport more material than the amount which has been made available by excavators. At the same time, the same dumper trucks should never bring more material to an embankment front than the amount spreaders can handle, since the accumulation of an excessive amount of material at that point can obstruct the remaining equipment. Hence, the duration of the production line work, Tpl, corresponds to the volume of material associated with the last job (compaction), Vc, divided by the minimum work rate within the production line, Qi, as illustrated in Equation 2. !!" = !! !"#!(!!,!!,!!,!!) (2) Concurrently, the associated cost can be divided into direct and indirect costs. While the former, dcosti, corresponds to constant expenses, such as equipment rental (when applicable) and manpower costs, the latter, icosti, is related to time dependent costs, such as fuel usage or maintenance cost. In this case, the associated cost can be obtained as shown in Equations 3 (cost of a single piece of equipment, Cs) and 4 (costs per production line, Cpl). !! = !"#$%!×!!" + !"#$%! (3) !!" = !! (4) Finally, the total duration of the whole earthwork project can be obtained by adding the durations of all non-simultaneous production lines, while total costs are related to the period each equipment is used during the project (this issue is further discussed in Section 3). M. Parente, P. Cortez, A. Gomes Correia 9 It should be noted that the optimization is constrained by the construction deadline, regarding total project duration, and budget, limiting the total costs obtained in the optimal solutions. One additional constraint regarding space restrictions on the construction site can be considered, limiting the maximum amount of equipment simultaneously at work on the same front, when applicable. 2.3. Solution Quality Assessment As referred in the previous section, total construction time corresponds to the accumulated time for each construction phase. Fundamentally, construction time in each front is a function of equipment productivity and material volume to be handled in that front. In turn, a high amount of factors have influence on the productivity of earthworks equipment for each case. For instance, the productivity of a compactor allocated to a specific front is a function of its type, the type material that is being compacted, and the conditions under which compaction takes place (e.g., layer thickness, atmospheric conditions). Similarly to total time, total cost equals to the accumulated costs for each construction phase. The latter is a function of the direct and indirect costs linked to each piece of equipment, as well as the amount of hours these are active. Since indirect costs are time dependant (e.g., the fuel usage depends on the amount of time the equipment is active), execution costs can only be calculated subsequently to the determination of construction time in each phase. The steps followed to determine the objective functions (total construction time and cost) are summarized in Table 1. As the allocation of equipment dictates the resulting construction time and cost, the usage of equipment to its full potential is paramount. In other words, the allocation of equipment takes into account the minimization of construction time and cost, but also the maximization of equipment efficiency. In turn, by using the equipment to its maximum efficiency, the subsequent allocation solutions will inherently reduce the environmental impact of the construction, for instance reducing carbon emissions. M. Parente, P. Cortez, A. Gomes Correia 10 Table 1. Steps for determination of total cost and duration of equipment allocation solutions Step Description Main associated variables/factors 1 Allocation of compactors to embankment fronts Available compactors (type and quantity); available embankment fronts 2 Determine individual productivity of the allocated compactors for each case Compactor type; material type; compaction conditions (e.g., layer thickness, meteorological conditions) 3 Calculate total productivity in each active compaction front Number of compactors of each front and individual productivity of each compactor 4 Allocate spreading equipment Total productivity in compaction task for associated embankment front; available spreaders/bulldozers (type and quantity); material type; work conditions 5 Allocate transportation equipment Minimum productivity in spreading and compaction tasks for associated embankment front; available trucks/dumpers (type and quantity); transportation distance; work conditions 6 Allocate excavation equipment Minimum productivity in transportation, spreading and compaction tasks for associated embankment front; available excavators (type and quantity), material type; work conditions 7 Calculate compaction duration in each embankment front Productivity of production line (minimum productivity amongst all tasks in a production line); required material volumes for completing each embankment front 8 Verify fastest production line to complete its work (corresponds to the duration of the current construction phase) Compaction duration of each production line; total volume of material required to complete each active embankment front 9 Calculate volumes of materials which have been excavated and compacted in each front during current phase Duration of current construction phase; productivity of each production line; volume of material available/required in each active front 10 Calculate cost according to the used equipment and the duration of current construction phase Direct and indirect costs of active equipment; duration of current construction phase 11 Verify if all embankment fronts have been completed. If not, initiate new construction phase (step 1), taking into account updated material volumes (calculated in step 9). Otherwise, output accumulated cost and duration. Available embankment fronts (if initiating new construction phase); individual cost and duration for each construction phase (if outputting results) In this context, and bearing in mind that earthwork construction can be interpreted as a series of production lines, global productivity will be at its highest rate when the productivity of the last task in these production lines (i.e., compaction task) is maximized. Given this premise, the allocation of equipment is firstly carried out for the compaction task, and then for each preceding task, as described in steps 1-6 of Table 1. Additionally, in order to guarantee maximum equipment efficiency, the allocation of the other tasks is performed in function of the productivity verified in embankment fronts. M. Parente, P. Cortez, A. Gomes Correia 11 For example, consider an initial construction phase where three compactors were distributed between two fronts. Lets assume the compactors, ci, are all of the same type, and the materials, as well as compaction conditions, in front 1 and 2 are the same, resulting in a compaction productivity of 400 m3/h per compactor. This means that total productivity in front 1 would be 400 m3/h, while front 2 would proceed at a rate of 800 m3/h. As such, the enough spreaders (e.g., bulldozers) must be allocated for each front so that the total productivity is as close as possible (equal or higher) to 400 m3/h in front 1 and 800 m3/h. That could mean that a higher amount of spreading equipment can be necessary for front 2 then for front 1, depending on the available spreader characteristics, their productivity when handling the material type, and the work conditions they are subjected to (e.g., atmospheric conditions). As previously stated, should the available spreaders not be enough to maintain the necessary work rate in one of the fronts, then the work rate of the compactors will be limited by the maximum productivity that those spreaders can achieve. The same allocation methodology is used for transportation and excavation equipment until each active embankment team is associated with a production line that can support its productivity. The methodology used for this allocation is described in Section 3. Having assembled the resulting production lines, the compaction time of each embankment front can be determined by applying Equation 2, as described in Section 2.2. Following the previous example, assuming that there is the necessary equipment to guarantee that the compactors in both fronts are working at their full potential (400 and 800 m3/h for front 1 and 2, respectively), and if the required volume of material to complete both embankments is 8000 m3, then the total duration of compaction should be 20 hours for front 1 and 10 hours for front 2. After the first 10 hours, the equipment associated with the production line of embankment front 1 will become idle, which means a new allocation must be performed. As such, construction phase 1 will be considered complete after 10 hours, and the compaction of all other fronts will be interrupted at that time, since a new optimal distribution of equipment can result in a reallocation of the equipment in that production line. The reason for redistribution all equipment after each construction phase is over is discussed in more detail in Section 3. As such, as construction phase 2 begins, the required volumes for completing embankment fronts 1 and 2 will be 0 m3 (completed) and 4000 m3, respectively. These correspond to steps 7-9 in Table 1. Finally, having the knowledge of the amount of hours each equipment has been active during construction phase 1 (in this case, 10 hours), it is possible to determine the time- M. Parente, P. Cortez, A. Gomes Correia 12 dependent cost (indirect costs) for each active piece of equipment. By adding the result to the direct costs of active equipment (Equation 3), the total cost for each piece of equipment can be calculated. The total cost for the current construction phase will correspond to the sum of the costs associated with of the active equipment (step 10 in Table 1). Although a construction phase is considered to end as soon as a compaction front is completed, each solution evaluated by the optimization algorithm is only complete when all fronts have been compacted. As such, this process is repeated for each construction phase, calculating the associated time and cost until all fronts have been compacted (step 11 in Table 1). In this case, since at least one embankment front is still not completed (embankment front 2), then at least one more construction phase will be necessary to complete the process. Subsequently, a new construction phase will begin from step 1 of Table 1, in which the required volume for embankment front 2 is now 4000 m3, whereas front 1 no longer is targeted for compactor allocation. Ultimately, the total time and costs are determined in the end of the process by adding the associated values for each construction phase. Section 3 further details some aspects of this process, framing it in the context of a chromosomal representation adopted for an EMO algorithm. 3. Evolutionary Multi-objective Optimization of Earthworks Different information technologies have been used for earthworks optimization, ranging from data mining techniques (Chapman et al., 2000; Fayyad et al., 1996), such as artificial neural networks (Hola & Schabowicz, 2010; Marques et al., 2008; Schabowicz & Hoła, 2008; Shi, 1999; Tam et al., 2002), multiple regressions (Edwards & Griffiths, 2000) and other techniques (Chou & Tseng, 2011; Liao, Chu, & Hsiao, 2012), to metaheuristcs, such as genetic algorithms (Burdett & Kozan, 2014; Marzouk & Moselhi, 2002; Moselhi & Alshibani, 2007; Xu et al., 2011), swarm intelligence (Jang & Topal, 2014; Kataria et al., 2005; Miao et al., 2011; Nassar & Hosny, 2012; Shishvan & Sattarvand, 2015; Zhang, 2008) and other heuristic algorithms (Mandow & Pérez-de-la-Cruz, 2004). The data mining based methods tend to focus on the prediction of equipment work rates rather than on direct resource allocation, which is directly approached using metaheuristics and other heuristics. The work presented in this paper is based on an EMO. Evolutionary computation methods, including genetic algorithms and EMO, work by maintaining a population of individuals (potential solutions), where a chromosome denotes individual data representation M. Parente, P. Cortez, A. Gomes Correia 13 of a solution and gene is a value position in such representation. And designing the chromosome is a key element when adopting evolutionary approaches, as it defines the search space of the problem. The majority of previous works that adopt metaheuristics only address one or two sequential steps of the earthwork construction (Edwards & Griffiths, 2000; Nassar & Hosny, 2012; Xu et al., 2011). Since we approach the whole earthwork process, a novel solution representation is proposed in this framework, where the individuals represent potential equipment allocations, which includes the front to which each piece of equipment is allocated. The key idea is to simplify solution representation by using domain knowledge and focusing solely on the optimal distribution of compaction equipment (the last task of the production line). The equipment for the other tasks, namely spreading, transportation and excavation, is then distributed according to the initial allocation of the compaction equipment (known as rollers), depending on the sum of work rates in each compaction front. Thus, for each construction phase, the solution is composed of a sequence of C integer genes: g1 g2 g3…gC, where gi denotes the position of the i-th compactor (or roller) in terms of its compaction front and C represents the total number of compactors. Genes can take any integer value from 0 to the maximum number of target compaction fronts, F (Figure 3). This representation aims firstly to allocate each roller ci to a compaction front ∈ {0, 1, 2, ..., F} (0 means that the roller is not allocated), followed by a validation of each solution acquired this way (and that involves a repair strategy). The whole individual (or chromosome) includes all construction phase gene sequences, thus the total number of genes corresponds to the number of available compactors times the number of necessary construction phases: C×F. In the particular case exemplified in Figure 3, there are C=3 rollers and F=3 compaction fronts, thus individuals are represented using nine genes. During fitness evaluation, gene values of 0 correspond to withholding the allocation of the specific roller to any front, remaining unused for the current construction front. This option is relevant since it allows the EMO to discover lower duration solutions and also to deal with (or discard) cases in which the available equipment plant for excavators, transporters and spreaders is not enough to support the allocation of all available rollers. Should the latter case be verified, and taking into account that the total rate of a production line corresponds to the minimum work rate obtained among the tasks comprised in it, the resulting solutions will have either infinite or very high durations with simultaneously high costs. This will cause the EMO to discard these solutions as non-optimal in early stages. M. Parente, P. Cortez, A. Gomes Correia 14 c1 c2 … cC c1 … cC … c1 … cC g1 g2 … gC gC+1 … g2C … gC×F-C+1 … gC×F Phase 1 Phase 2 … Phase F Figure 3. Chromosome representation for a generic distribution (top) and distribution example for a case with C=3 rollers and F=3 compaction fronts (bottom). When validating an individual, a repair strategy is used to assure that the solution is feasible. Such repair strategy involves three ordered steps: 1. Verification of completed compaction fronts in previous construction phases. When verified, any rollers allocated to the already completed front are instead allocated to the next active front. 2. Verification of maximum number of rollers in each compaction front. This constraint (when applicable) is set to deal with space restrictions on embankment fronts that exist in many construction sites. When verified, the rollers that are above the recommended limit of the current front are reallocated to the next active front according to the rule gi = (gi+1) mod (F+1), if possible, or otherwise assigned a value of 0 (not allocated). 3. Verification that at least one front must be compacted per construction phase. This verification is related with specific cases when the EMO generates a solution with a value of 0 to all roller allocations. In such cases, one gene is assigned with a value associated with one of the active fronts. The order of repair strategies takes into account the prevention of harmful interactions between the three verification steps. As an example, should a roller be reallocated to a front in which the roller limit has already been exceeded as a result of the original front being c1 c2 c3 c1 c2 c3 c1 c2 c3 2 1 2 1 1 3 3 0 3 Phase 1 Phase 2 Phase 3 M. Parente, P. Cortez, A. Gomes Correia 15 completed (repair #1), it will be reallocated once more either to a new front or to none (repair #2). Furthermore, the nature of genetic operators, namely crossover and mutation, guarantees the diversity of solutions, preventing any convergence to local optima as a result of the used repair strategies and assuring that a Pareto-optimal curve is found. It should be noted that these repair strategies were designed to be as least “invasive” as possible, in order to avoid influencing or hindering the algorithm convergence. After assigning the compactors, the allocation of the remaining equipment plant (excavators, transporters and spreaders) is then carried out by means of linear programming (LP) models. As suggested in (Liu & Lu, 2014), LP models are effective in the optimization of partial or secondary aspects of the process, in which less computation power is required. In this work, to each task and to each equipment type is associated a separate LP model that targets the work rate of the last task in the production line. These were executed as standard LP models, where the main objective function is the minimization of equipment costs (Equations 3 and 4). The associated constraints guarantee that the work rates of the allocated equipment for each task, Qs,i, are at least equivalent to that of the last task in the production line (compaction) when possible, or otherwise as close as possible, given the maximum amount of available equipment of each kind. The intention behind this type of distribution is to use the least expensive combination of available equipment that allows the succeeding task to operate at 100% work rate, by matching each task to the work rate of the last task in the production line. Equation 5 shows a generic LP model for any equipment kind, where xi corresponds to the number of allocated equipment units of a given type, with mi maximum available units. Minimize !!,!×!! !! ! Subject to !!,!×!! !! ! ≥ !!"!#$,!"#$%!&'"( !! ≥ !! (5) In a given construction phase, each production line is independent, but several production lines can be at work at the same time if there is enough equipment to support it. Thus, the number of simultaneous production lines is equal to the number of compaction fronts being compacted at the same time, which, in turn, is a consequence of the initial roller M. Parente, P. Cortez, A. Gomes Correia 16 distribution by the EMO. As such, the maximum number of simultaneous production lines at work is equivalent to the maximum number of rollers. Given the allocation of all equipment for the active productions lines, the compaction duration of all fronts is calculated according to Equation 2. From the resulting durations, the compaction front with minimum duration (the fastest compaction front) marks the end of the current construction phase. Accordingly, costs for the current construction phase are determined in function of its duration using Equations 3 and 4. Each construction phase corresponds to the completion of 1 embankment or compaction front. Whenever an embankment is completed, a new construction phase begins and thus a new equipment reallocation must follow, since the equipment corresponding to the completed production line has become idle/available. However, the construction site conditions may change after each construction phase and limiting the equipment reallocation to the idle equipment would not be a guarantee of optimal distribution. The solution is to treat each construction phase as a completely new optimal allocation of equipment, which must take into account the work already developed in previous phases, including site features and conditions, such as available and required material in excavation or compaction fronts. To achieve this, two memory lists are kept during the determination of construction phase durations and costs, concerning completed fronts and remaining fronts. Both lists are updated at the end of each construction phase. The former includes the fronts that have been completed in previous phases, entailing the duration associated with each, as well as the order of completion. The latter, FR, includes the actual quantities of material in each compaction front, in function of the duration of the previous construction phase and the work rate of the allocated equipment for each front, as illustrated in Equation 6. The determination of compaction durations in the construction phases (other than the first phase) is always carried out in function of the volumes, FR,i, maintained in FR. Subsequently, progress obtained in all the active compaction fronts is saved at the end of each construction phase, thus being accounted for when determining the compaction durations for the new equipment allocation, in the next phase. !!,! = !!,! − min!(!!",!)×Q! (6) Figure 4 shows the algorithmic flow of the EMO algorithm used to tackle the earthworks optimization problem. The initial population involves the random generation of M. Parente, P. Cortez, A. Gomes Correia 17 front indexes to each roller, representing their allocation, as shown in Figure 3. The fitness function includes the repair methodology described in the previous section, as well as the LP models for the allocation of equipment to the remaining tasks. These allow for the determination of construction phase durations and costs using the formulae discussed in Section 2.2, ultimately leading to the evaluation, in terms of total duration and cost, of each solution. The EMO will generate a new population and repeat this process until the target number of generations is achieved, at each point the Pareto-optimal set is presented, along with the outputs mentioned in Figure 4 for each solution. Figure 4. Algorithmic flow of the EMO algorithm. M. Parente, P. Cortez, A. Gomes Correia 18 4. Results 4.1. Real-World Construction Case To validate the proposed system, we adopt real-world data related with mechanical equipment of a Portuguese road construction site. Originally, the data included the description of several years of earthworks construction, broke down into the daily activities of the available mechanical equipment. The data attributes include date, daily work hours, number and distance of load trips (including daily volumes of transported material) and resource types, as depicted in Table 2. The data is related with 4 excavation fronts supplying two types of soil (a soil and a soil-rockfill mixture) to 5 embankment fronts. The total volume of transported material is approximately 89,356 m3. The distances from excavation to embankment fronts vary between 100 m to nearly 4,000 m, stressing the need for proper resource management. Table 2. Example of row values extracted from the available earthwork database. Date Work Hrs. (h) Nr of Load Associated excavator Load Zone Unload Zone Resource Type Transp. Volume (m3) 01/6 9 7+850 8+625 Excavator50T 01/6 8 7+850 8+625 Roller15T 01/6 10 37 20/871 13+750 12+250 Dumper40T 481 01/6 10 39 20/871 13+750 12+250 Dumper50T 634 01/6 9 13+750 12+250 Tractor40T The construction company available equipment included 12 excavators (with sizes ranging from 25 to 75 tons), 22 trucks (either 30 or 40 ton dumper trucks), 8 spreaders (between 20 and 50 ton) and 5 vibrating soil compactors. Both the materials and the compaction equipment were classified according to the Guide des Terrassements Routiers (GTR) (SETRA & LCPC, 2000), a well known and broadly used compaction guide. As such, the materials were classified as A1 and C2A1 for the soil and the soil-rockfill mixture, respectively. Accordingly, 4 out of 5 of the vibrating soil compactors were classified as V3, while the remaining one as a V4. The company designed resource distribution, set taking into account the construction deadlines and budget of the project, resulted in approximately 130 hours of work, spread through 21 days of work (approximately 1 month). All the available equipment was put to M. Parente, P. Cortez, A. Gomes Correia 19 use, with an average value of daily work hours per day equals to 9.88h. Considering this setup, the final cost obtained for these work phases was close to 135,199 €. The necessary inputs for the optimization system were retrieved from the available database and are summarized in Table 3. Equipment indirect costs and work rates derived from Caterpillar Performance Handbook (Caterpillar Inc., 1998) and Transportation Research Board NCHRP Report 744 (Skolnik, Brooks, & Oman, 2013), as well as the GTR in the case of compaction work rates. Direct costs (including equipment rental and manpower costs) were courtesy of the company. Table 3. Optimization system inputs Input Description fa Embankment front notation and material volume required for completion Material data fe Excavation front notation and material volume available for excavation and transport mt Material type in each excavation front (which will be compacted in embankment fronts after excavation and transport) cd Travel distance matrix from each excavation front to each embankment front EE Available excavation equipment, including type, number of available equipment of each type, work rate and direct and indirect costs E quipm ent data TE Available transportation equipment, including type, number of available equipment of each type, capacity and direct and indirect costs SE Available spreading equipment, including type, number of available equipment of each type, work rate and direct and indirect costs CE Available compaction equipment, including type, number of available equipment of each type, work rate and direct and indirect costs 4.2.Computational Experiments In this work, the equipment available was kept fixed to the one used by the construction company in their conventional allocation design. With a total of F=5 production lines working simultaneously and C=5 available compactors, the encoded individuals have 5x5=25 genes each, defining the search space for this problem. As the EMO search engine, the Non-dominated Sorting Genetic Algorithm-II (NSGA-II) (Deb, Pratap, Agarwal, & Meyarivan, 2002) was adopted due two main reasons. Firstly, NSGA-II is a popular and standard method for multi-criteria evolutionary optimization. Secondly, NSGA-II is easily M. Parente, P. Cortez, A. Gomes Correia 20 available for a computational use in the R statistical tool (R Development Core Team, 2011) via the package mco (Mersmann, Trautmann, Steuer, Bischl, & Deb, 2014), which is the same tool adopted for the development of our integrated optimization system. Moreover, the R tool includes several conventional optimization methods, such as the Linear Programming (LP) method that was used for individual fitness calculation. The default parameterization of NSGA-II method, as implemented in the R tool, was used, namely: population size of 100, stop after 100 generations, crossover probability of 0.7 and mutation probability of 0.2. The rationale is to focus more on assessing and validating the capabilities of the proposed integrated system when compared with current human design, rather than calibrating the optimization algorithm. Note that in preliminary tests, smaller population sizes (i.e., 20, 30, 40) were explored, but the obtained results were worse than the default population size of 100. Also, the fitness evaluation is computationally costly, as it requires several LP optimizations (for each front), thus a population size much larger than 100 individuals would increase the computational effort. Given that in the mco package of the R tool adopts a real value representation for the NSGA-II method, all genes were first rounded to the nearest integer as the first step of the fitness function. The method was executed with an exclusive access to an Intel Xeon 2.27GHz server under a Linux server. To get a more robust assessment of the quality of the results, 30 distinct runs of the NSGA-II algorithm were executed. The total computational effort (considering all 30 runs) was approximately 254,642.5 seconds (around 70.73 hours). To illustrate the multi-objective convergence, Figure 5 plots the evolution of solutions optimized by the NSGA-II towards the Pareto-optimal front according to cost and duration goals, and during a single run. In the plot, each point denotes a possible solution while line segments are used to join points that belong to the Pareto front. To facilitate the analysis, a colouring scheme is used, ranging from light grey (first generation) to black (last generation). Figure 5 shows that NSGA-II performs an initial fast convergence, with substantial movements of the Pareto front towards the bottom left region, and then the algorithm converges more slowly towards the returned Pareto front, which is non-convex near the bottom left region. M. Parente, P. Cortez, A. Gomes Correia 21 Figure 5. Example of the convergence of NSGA-II algorithm (x-axis in € and y-axis in hours). The obtained results were also compared with two other optimization algorithms. The purpose of the first comparison is to validate the quality and accuracy of obtained results by performing a comparative study with a different and EMO variant, namely the S metric selection evolutionary multi-objective algorithm (SMS-EMOA) (Emmerich, Beume, & Naujoks, 2005). The second comparison is aimed at demonstrating the advantages of multi- objective optimization when compared with a single-objective optimization method. These comparisons are depicted in Figure 6, showing the average Pareto fronts for both EMO algorithms, as well as the average point corresponding to a cost-only optimization solution using single-objective optimization. The average Pareto curves were achieved by performing a vertical averaging procedure (i.e., according to the Duration objective) of the Pareto curves outputted by each run, using the averaging method proposed by Fawcett (2006) for vertical averaging of ROC curves. The graph also includes “H” shape whiskers, on both the lines and the point, that denote the 95% confidence interval bars according to a t-student distribution. Regarding the comparison with a different EMO variant, the SMS-EMOA was chosen, for being established as a recent and well-known EMO algorithm, which makes use of the hypervolume measure as selection criterion. Furthermore, this algorithm is can be implemented under the R tool via a small portion of code that uses functions from the emoa package (Mersmann, 2012), facilitating the comparative analysis. Since the aim is to allow a direct comparison between results and performances of both EMO algorithms, the parameters used in SMS-EMOA are the same that were used for the NSGA-II optimization (population=100, stop after 100 generations, crossover probability=0.7, mutation probability=0.2). The analysis of Figure 6 reveals a good consistency between the results M. Parente, P. Cortez, A. Gomes Correia 22 obtained by both EMO algorithms. Nevertheless, NSGA-II slightly outperforms the SMS- EMOA in terms of the resulting Pareto curve results. Furthermore, the total computational effort for the SMS-EMOA to perform all 30 runs was approximately 270,605.1 seconds (around 75.17h), which is also above the 70.73 hours associated with the NSGA-II runs. The single objective algorithm is implemented in R by adopting the genalg package (Willighagen, 2015). The solutions for the single-objective optimization of costs (30 runs with the same optimization parameters as the EMO algorithms) seem to be consistent with the results obtained by the EMO algorithms in terms of the cost reduction goal. However, as previously referred, since the output is a single allocation solution, it lacks the flexibility of the Pareto curve optimization, providing to the decision maker just one cost-duration trade-off solution under one optimization execution, overlooking the advantages of multi-objective optimization. The computational effort for this algorithm (30 runs) was 212,000.9 seconds (58.89 h). Figure 6. Comparison between results obtained by different optimization methods (x-axis is in hours and y-axis in €) 4.3. Comparison with Manual Allocation The proposed system results were compared against the original manual allocation performed by the construction company, as shown in Figure 7. In the figure, the left graph shows the average (over all 30 runs) NSGA-II optimized Pareto front, while the right graph compares these results (black line) with the original solution obtained by manual allocation (white point). M. Parente, P. Cortez, A. Gomes Correia 23 When analysing Figure 7, it is clear that NSGA-II performs a substantial improvement (both in terms of cost and duration) when compared with the manual equipment allocation. The system output indicated several potential setups ranging from approximately 38 to 73 hours of construction duration, associated with approximate costs of 40,000 € to 43,000 €, respectively. This corresponds to a reduction of around 55% in duration and 70% in cost, if compared to the duration of 127 h and cost of 135,199 € that was obtained in the original allocation. The original human based allocation is far from the optimal solution, which can be explained by the lack of an efficient automated optimization and also by unforeseen delays. Figure 7. Optimization results (k=0.75) in terms of: left graph - vertically averaged Pareto front; and right - comparison between the optimized Pareto front (black curve) and the human based allocation solution (circle point) (in both graphs x-axis is in hours and y-axis in €). Regarding the lack of efficient automation, the original setup performs bottlenecks in the production lines in one of the tasks preceding compaction, resulting in several occurrences of idle equipment. To illustrate this, two distinct production lines are further analysed in Table 4. It is easy to infer that, for both production lines the work rates in each task of the original distribution setup are not homogeneous, as opposed to the work rates of the optimized solution. In these cases, the whole production line is limited by the work rate of excavators in the original setup, which means that the other tasks have to wait for material to be excavated in order to allow for its transport, spreading and finally compaction. This incurs in equipment idle time while waiting for material to be ready for handling, which represents wastes in terms of resources (since these do not work at full efficiency) and fuel (contributing to unnecessary costs), as well as an increase on unnecessary carbon emissions. Consequently, the total work rate of these production lines cannot be considered superior to that of the minimum work rate obtained in the production line tasks, in this case excavation (394 m3/h in production line 1 M. Parente, P. Cortez, A. Gomes Correia 24 and 540 m3/h in production line 2). In contrast, the work rates obtained in the proposed optimized solutions for each task that comprises the production line are as homogeneous as possible, given the available equipment. As such, a constant flow of material throughout tasks can be achieved, using the allocated resources to their full potential and efficiency. Table 4. Comparison between original and optimized (k=0.75) setups for both cases Original setup 1 Original setup 2 Average distance between excavation and compaction fronts: 175 m Average distance between excavation and compaction fronts: 500 m Hauled material volume: 29,753 m3 Hauled material volume: 10,647 m3 Equipment distribution and work rate Equipment distribution and work rate 1 Excavator (50T): 2 Dumper trucks (30T, 40T): 1 Spreader (20T): 1 Vibratory roller (15T): 394 m3/h 2,960 m3/h 413 m3/h 614 m3/h 1 Excavator (75T): 3 Dumper trucks (40T): 1 Spreader (40T): 1 Vibratory roller (19T): 540 m3/h 1,280 m3/h 675 m3/h 683 m3/h Minimal production line work rate: 394 m3/h (excavation) Minimal production line work rate: 540 m3/h (excavation) Duration: 75.5 h Duration: 19.7 h Cost: 24,462 € Cost: 7,996 € Optimized setup 1 Optimized setup 2 Average distance between excavation and compaction fronts: 175 m Average distance between excavation and compaction fronts: 500 m Hauled material volume: 29753 m3 Hauled material volume: 10647 m3 Equipment distribution and work rate Equipment distribution and work rate 2 Excavator (75T): 2 Tipper trucks: 2 Spreaders (20T, 50T): 1 Vibratory roller (19T): 1,080 m3/h 1,600 m3/h 1,239 m3/h 1,055 m3/h 2 Excavators (25T, 75T): 2 Dumper trucks (30T): 1 Spreader (50T): 1 Vibratory roller (19T): 743 m3/h 880 m3/h 820 m3/h 683 m3/h Minimal production line work rate: 1,080 m3/h (compaction) Minimal production line work rate: 683 m3/h (compaction) Duration: 28.2 h Duration: 15.6 h Cost: 12,718 € Cost: 5,740 € M. Parente, P. Cortez, A. Gomes Correia 25 It is noteworthy to emphasize that, besides optimizing the whole allocation in terms of costs and durations, the developed system is expected to always keep the allocated equipment working at full efficiency. It achieves this by focusing on the minimization of equipment idle time, which will also result in minimization of unnecessary carbon emissions. This last aspect is very challenging to accomplish by conventional design methodologies. Comparing the original setup with its optimized counterpart for production line 1, the human based allocation solution features a clear excess of work capacity regarding transportation equipment that is not contributing for its progress, as it is limited by the work rate of the excavation team. In order to counter this, the optimization system allocated smaller trucks (lower capacity, lower fuel consumption and, thus, lower operation costs) to fulfil this role instead, while investing its resources more heavily on the excavation, spreading and compaction teams. As a result, the optimized setup for this case is as homogeneous as possible in terms of work rate between tasks (taking into consideration the available equipment), resulting in a decrease of 50% in both duration and cost for this production line. Although the proposed optimized solution might allocate more equipment with a higher productivity, the reduction in duration is significant enough to reduce the operational costs, ultimately resulting in a reduction of both factors. In other cases, as one can see in the setups for production line 2, which features different fronts and material types, the setup achieved by the optimization module consists of slight differences when compared with the original allocation. The subtle changes to the class of equipment used in some tasks (e.g., spreading), as well as the removal of unnecessary equipment from others (e.g., transportation) allows a greater homogeneity of work rates, resulting in nearly 20% reduction in duration and 30% in cost. Moreover, the equipment that is dismissed from this production line is then available to be used in different work fronts as deemed necessary. The obtained results emphasize the importance of using intelligent computational tools for optimizing earthworks. In particular, it was shown how conventional human design allocation methodologies can be relatively counter-productive in some situations. During optimization, an efficiency factor of k=0.75 was assumed for compactors, as well as other equipment types. The efficiency factor is related to the amount of time that mechanical equipment spends in actual production. According to the technical guides used during the development of this work, namely GTR (SETRA & LCPC, 2000) and Caterpillar Performance Handbook (Caterpillar Inc.,1998), actual “on-the-job’’ productivity is influenced M. Parente, P. Cortez, A. Gomes Correia 26 by factors such as operator skill, personal delays, job layout and other delays. Since the goal is to maximize resource usage and minimize idle times, it makes sense to use a k value of 0.75, considering that this is the maximum recommended value by the GTR for compaction activities. Unforeseen delays are due to unpredictable situations that can occur in a real environment, such as presence of bad material or equipment malfunction. The huge difference between NSGA-II results and the manual design might be a consequence of a lower efficiency occurred in the real construction project. To attest the competiveness of the proposed system, additional NSGA-II experiments were conducted assuming a much lower efficiency factor of k=0.3 (Figure 8). When compared with previous experiments of k=0.75, the new k=0.3 NSGA-II Pareto front solutions require higher costs and duration. In particular, duration now ranges from 105 h to 155 h, a range on which the manual allocation solution falls on (as shown in the right of Figure 8). However, even when assuming a low efficient factor, NSGA- II still outperforms the manual solution, returning a Pareto front that is more interesting and that includes a trade-off point that is much less costly for the same duration. These new experiments show the versatility of the proposed system in respect to unpredictable events. If needed, the proposed system can be rerun in order to include updates of the site conditions and new restraints, such as altering the available equipment and/or productivity rates. Figure 8. Optimization results (k=0.3) in terms of: left graph - vertically averaged Pareto front; and right - comparison between the optimized Pareto front (black curve) and the human based allocation solution (circle point) (in both graphs x-axis is in hours and y-axis in €). M. Parente, P. Cortez, A. Gomes Correia 27 5. Conclusions Most civil construction structures (e.g., road, railway and airport construction) require earthworks activities, which involve levelling the ground surface and preparing the required foundation conditions. Under this context, the optimization of earthwork is a key factor for construction competitiveness, since earthworks account for a high percentage of the overall construction costs and duration. Conventional manual design, often based on experience and intuition, is limited. Moreover, current intelligent automated applications often target single tasks or partial processes that comprise earthworks and do not simultaneously optimize both cost and duration goals. This work presents a novel evolutionary multi-objective optimization (EMO) approach that addresses the whole earthwork construction phase and optimizes both cost and duration objectives. It is supported by an original representation of earthworks as production lines, a Pareto approach for dealing with conflicting objectives, and the integration of different optimization methods. The approach has been directed towards a practical and flexible expert intelligent system that can be used in any construction that includes large volumes of earthworks. The potential of the developed system is shown by its validation using real-world data from a Portuguese construction site, while identifying some significant limitations of conventional manual earthworks design. In fact, the obtained results show that the EMO system is quite competitive when compared with the conventional design and that it can be easily adapted to dynamic changes that are inherent to earthworks constructions. The development of this system opens several new insights, such as: (1) further experimentation with different optimization objectives, in order to keep up with current and future trends, such as environmental impact during construction (e.g., carbon emissions). For example, there has been some development regarding methodologies for controlling and reducing carbon emissions during construction, and such data could be used to test and validate its use as a minimization objective; (2) the optimization methodology regarding the production lines can be enhanced by taking into account other aspects, such as loading and dumping configurations; (3) the system can also profit from focusing on the tuning of the optimization parameters, as well as exploring other EMO variants (e.g., SPEA-2, IBEA, MOEA/D). Finally, the developed optimization system is expected to be enhanced with data mining and geographic information systems (GIS) technologies, in order to integrate a more M. Parente, P. Cortez, A. Gomes Correia 28 complex system. Data mining can be used to support parameter estimation, such as equipment productivity in a given situation, when the actual data is unavailable (i.e., early design phase), while GIS can provide a visual support for the designer, both in terms of data input (i.e., positioning of excavation and compaction fronts, or determination of shortest routes between fronts) and in terms of visualization of results (i.e., graphically representing the allocation of equipment throughout work fronts). ACKNOWLEDGEMENTS The authors wish to thank FCT for the financial support under the doctoral Grant SFRH/BD/71501/2010, as well as the construction company that kindly provided the real- world data. Also, we wish to thank Olaf Mersmann for kindly providing the R code for the SMS-EMOA algorithm. REFERENCES Bonissone, P., Subbu, R., & Lizzi, L. (2009). Multi Criteria Decision Making (MCDM): A Framework for Research and Applications. IEEE Computational Intelligence Magazine, 4(3), 48–61. Burdett, R. L., & Kozan, E. (2014). An integrated approach for earthwork allocation, sequencing and routing. European Journal of Operational Research, 238(3), 741–759. doi:10.1016/j.ejor.2014.04.036 Caterpillar Inc. (2011). Caterpillar Performance Handbook Edition 29. Chang, A. S., & Tsai, C. Y. (2015). Sustainable design indicators: Roadway project as an example. Ecological Indicators, 53, 137–143. doi:10.1016/j.ecolind.2015.01.036 Chapman, P., Clinton, J., Kerber, R., Khabaza, T., Reinartz, T., Shearer, C., & Wirth, R. (2000). CRISP-DM 1.0 Step-by-step data mining guide. Retrieved from http://www.citeulike.org/group/1598/article/1025172 Chou, J.-S., & Tseng, H.-C. (2011). Establishing expert system for prediction based on the project-oriented data warehouse. Expert Systems with Applications, 38(1), 640–651. Retrieved from http://www.sciencedirect.com/science/article/pii/S0957417410006329 Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002). A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2), 182–197. Retrieved from http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=996017 M. Parente, P. Cortez, A. Gomes Correia 29 Edwards, D. J., & Griffiths, I. J. (2000). Artificial intelligence approach to calculation of hydraulic excavator cycle time and output. Mining Technology, 109(1), 23–29. Emmerich, M., Beume, N., & Naujoks, B. (2005). An EMO algorithm using the hypervolume measure as selection criterion. Evolutionary Multi-Criterion Optimization, 62–76. doi:10.1007/978-3-540-31880-4_5 Fawcett, T. (2006). An introduction to ROC analysis. Pattern Recognition Letters, 27(8), 861–874. doi:10.1016/j.patrec.2005.10.010 Fayyad, U., Piatetsky-Shapiro, G., & Smyth, P. (1996). From Data Mining to Knowledge Discovery in Databases. American Association for Artificial Intelligence, 17(3), 1–18. Guerin, T. (2014). Root causes of fluid spills from earthmoving plant and equipment: Implications for reducing environmental and safety impacts. Engineering Failure Analysis, 45, 128–141. doi:10.1016/j.engfailanal.2014.06.011 Hola, B., & Schabowicz, K. (2010). Estimation of earthworks execution time cost by means of artificial neural networks. Automation in Construction, 19(5), 570–579. doi:10.1016/j.autcon.2010.02.004 Jang, H., & Topal, E. (2014). A review of soft computing technology applications in several mining problems. Applied Soft Computing Journal, 22, 638–651. doi:10.1016/j.asoc.2014.05.019 Kataria, S., Samdani, S. A., & Singh, A. K. (2005). Ant Colony Optimization in Earthwork Allocation. International Conference on Intelligent Systems, (7), 1–9. Retrieved from http://www.geocities.ws/saurabhsamdani/sourcecodes_files/icis-paper.pdf Liao, S., Chu, P.-H., & Hsiao, P.-Y. (2012). Data mining techniques and applications – A decade review from 2000 to 2011. Expert Systems with Applications, 39(12), 11303– 11311. doi:10.1016/j.eswa.2012.02.063 Liu, C., & Lu, M. (2014). Optimizing Earthmoving Job Planning Based on Evaluation of Temporary Haul Road Networks Design for Mass Earthworks Projects. Journal of Construction Engineering and Management. doi:http://dx.doi.org/10.1061/(ASCE)CO.1943-7862.0000940 Mandow, L., & Pérez-de-la-Cruz, J.-L. (2004). Sindi: an intelligent assistant for highway design. Expert Systems with Applications, 27(4), 635–644. doi:10.1016/j.eswa.2004.06.005 Marques, R., Gomes Correia, A., & Cortez, P. (2008). Data Mining Applied to Compaction of Geomaterials. In Eight International Conference on the Bearing Capacity of Roads, Railways and Airfields. Montreal, Canada. Marzouk, M., & Moselhi, O. (2002). Selecting Earthmoving Equipment Fleets Using Genetic Algorithms. In E. Yucesan, C.-H. Chen, J. L. Snowdon, & J. M. Charnes (Eds.), M. Parente, P. Cortez, A. Gomes Correia 30 Proceedings of the 2002 Winter Simulation Conference (pp. 1789–1796). Montreal, Canada. Mersmann, O. (2012). Package “emoa”: Evolutionary Multiobjective Optimization Algorithms. Retrieved from http://www.statistik.tu-dortmund.de/~olafm/software/emoa/ Mersmann, O., Trautmann, H., Steuer, D., Bischl, B., & Deb, K. (2014). Package “mco”: Multiple Criteria Optimization Algorithms and Related Functions. Retrieved from http://git.p-value.net/p/mco.git Miao, K., Sun, X., & Li, L. (2011). A roadbed earthwork allocation model based on ACO algorithm. Applied Mechanics and Materials, 44-47, 3483–3486. Moselhi, O., & Alshibani, A. (2007). Crew optimization in planning and control of earthmoving operations using spatial technologies. Journal of Information Technology in Construction, 12, 1–17. Retrieved from http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.98.1818&rep=rep1&type=pdf Nassar, K., & Hosny, O. (2012). Solving the Least-Cost Route Cut and Fill Sequencing Problem Using Particle Swarm. Journal of Construction Engineering and Management, 138(8), 931–942. Parente, M., Gomes Correia, A., & Cortez, P. (2014). Artificial Neural Networks Applied to an Earthwork Construction Database. In D. Toll, H. Zhu, A. Osman, W. Coombs, X. Li, & M. Rouainia (Eds.), Second International Conference on Information Technology in Geo-Engineering (pp. 200–205). Durham, UK: IOS Press. R Development Core Team. (2011). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Retrieved from http://www.r- project.org/ Schabowicz, K., & Hoła, B. (2008). Application of artificial neural networks in predicting earthmoving machinery effectiveness ratios. Archives of Civil and Mechanical Engineering, 8(4), 73–84. doi:10.1016/S1644-9665(12)60123-X SETRA, & LCPC. (2000). Guide des Terrassements Routiers - Réalisation des Semblais et des Couches de Forme. Paris, France: Laboratoire Central des Ponts et Chaussees. Shi, J. J. (1999). A neural network based system for predicting earthmoving production. Construction Management and Economics, 17(4), 463–471. Shishvan, M. S., & Sattarvand, J. (2015). Long term production planning of open pit mines by ant colony optimization. European Journal of Operational Research, 240(3), 825–836. doi:10.1016/j.ejor.2014.07.040 Skolnik, J., Brooks, M., & Oman, J. (2013). NCHRP Report 744 - Fuel Usage Factors in Highway and Bridge Construction. Washington, D.C. M. Parente, P. Cortez, A. Gomes Correia 31 Tam, C. M., Tong, T., & Tse, S. (2002). Artificial neural networks model for predicting excavator productivity. Journal of Engineering Construction and Architectural Management, 9(5-6), 446–452. Wang, X., Duan, Z., Wu, L., & Yang, D. (2014). Estimation of carbon dioxide emission in highway construction: a case study in southwest region of China. Journal of Cleaner Production. doi:10.1016/j.jclepro.2014.10.030 Willighagen, E. (2015). Package “genalg.” Retrieved from https://github.com/egonw/genalg Xu, Y., Wang, L., & Xia, G. (2011). Research on the optimization algorithm for machinery allocation of materials transportation based on evolutionary strategy. Procedia Engineering, 15, 4205–4210. doi:10.1016/j.proeng.2011.08.789 Zhang, H. (2008). Multi-objective simulation-optimization for earthmoving operations. Automation in Construction, 18(1), 79–86. doi:10.1016/j.autcon.2008.05.002 
work_kaayfpvw3rcxtdhrfkbkaaatpi	----	doi:10.1016/j.eswa.2007.01.018 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 1599–1608 Expert Systems with Applications An incremental cluster-based approach to spam filtering Wen-Feng Hsiao a,*, Te-Min Chang b a Department of Information Management, National Pingtung Institute of Commerce, Taiwan b Department of Information Management, National Sun Yat-sen University, Taiwan Abstract As email becomes a popular means for communication over the Internet, the problem of receiving unsolicited and undesired emails, called spam or junk mails, severely arises. To filter spam from legitimate emails, automatic classification approaches using text mining techniques are proposed. This kind of approaches, however, often suffers from low recall rate due to the natures of spam, skewed class distributions and concept drift. This research is thus to propose an appropriate classification approach to alleviating the problems of skewed class distributions and drifting concepts. A cluster-based classification method, called ICBC, is developed accordingly. ICBC contains two phases. In the first phase, it clusters emails in each given class into several groups, and an equal number of features (key- words) are extracted from each group to manifest the features in the minority class. In the second phase, we capacitate ICBC with an incremental learning mechanism that can adapt itself to accommodate the changes of the environment in a fast and low-cost manner. Three experiments are conducted to evaluate the performance of ICBC. The results show that ICBC can effectively deal with the issues of skewed and changing class distributions, and its incremental learning can also reduce the cost of re-training. The feasibility of the proposed approach is thus justified. � 2007 Elsevier Ltd. All rights reserved. Keywords: Email classification; Skewed class distribution; Concept drift; Incremental learning 1. Introduction With the rapid growth of the Internet, email has become one of the most common media for us to distribute information. More and more people depend on it to com- municate because of its properties of convenience, no restrictions in time and location, prompt delivery, and low cost. Some people, however, abuse it to spread large amount of information ranging from ads of drug, easy money, porn, to political promotion. As a result, our mail- boxes are usually filled with unsolicited emails (spam or junk mails). Spam filtering becomes one of the essential issues to most companies, governments, or even individual users. 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.01.018 * Corresponding author. Tel.: +886 8 7238700x6201. E-mail addresses: wfhsiao@mail.npic.edu.tw (W.-F. Hsiao), temin@ mail.nsysu.edu.tw (T.-M. Chang). To resolve the spam problem, traditional filters employed simple and straightforward methods, called fil- tering by rules, which classify spam emails by matching particular email fields (e.g. sender, recipient, date, subject, and attachment) with certain keywords. Although certain amount of spam can be filtered out, more complex spam (for example, those that can hide their own identities, including sending machines and/or senders) can survive easily with the method applied. More recent and advanced methods rely more on the content analyses of email bodies to deal with complex spam. This kind of methods extracts features (keywords) from both legitimate and spam emails, and then builds a classification model using techniques from text mining. However, their main problem is the poor performance in the recall rate resulting from the natures of spam itself (Fawcett, 2003), including skewed class distributions and concept drift. In the case of skewed class distributions, the number of spam emails received is far more than that mailto:wfhsiao@mail.npic.edu.tw mailto:temin@ mail.nsysu.edu.tw mailto:temin@ mail.nsysu.edu.tw Cl0 ClP ClN P P N N Cl 0 Cl P Cl N NP Fig. 1. Concept illustration of the refinement approach (Source: Wu et al., p. 210). 1600 W.-F. Hsiao, T.-M. Chang / Expert Systems with Applications 34 (2008) 1599–1608 of legitimate ones so that the class distributions of legiti- mate and spam are largely uneven. This often leads to a poor recall rate for the legitimate emails (the minority class). For the concept drift problem, users’ preferences may change with time or the topics of spam (or legitimate emails) may vary according to the fashionable trends. Therefore, the recall rates for both spam and legitimate emails can be low due to the changes. The learned structure of a classifier should adapt itself to accommodate these changes to classify new emails correctly. The purposes of this research are thus to propose an appropriate classification method to alleviate the above- mentioned problems, and to improve the effectiveness of spam filtering accordingly. We develop an adaptive cluster-based classification method, called incremental clus- tering-based classification (ICBC), for such purposes. Basi- cally, we deal with the class-skewed problem by first clustering emails in each class (spam and legitimate emails) into some coherent subsets. An equal number of keywords are then extracted from each subset. By doing so, rare but important keywords of minority classes can be extracted. To deal with the problem of concept drift, we capacitate ICBC with an inherent incremental learning mechanism. With the incremental learning capability, ICBC not only can personalize users’ email filtering preferences, but also can adjust itself over time for the changing environments, and thus can improve the classification accuracy under concept drift in a fast and low-cost fashion. The rest of this paper is organized as follows. The related research works on special spam natures, adaptive learning, cluster analysis, and spam filtering are reviewed in Section 2. The details of the ICBC method are elabo- rated in Section 3, followed by the presentation of experi- ments and results for evaluating the effectiveness of ICBC. Finally, the conclusions and future works are dis- cussed in Section 5. 2. Literature review 2.1. Special spam natures Emails with spam often exhibit the phenomenon of skewed class distributions. Namely, we receive spam emails far more than legitimate ones in our daily life. In such cases where the class distributions of the data are highly skewed, typical classifiers have difficulty in correctly predicting the classes with few data (minority classes) (Monard & Batista, 2002). Minority classes, however, may even deserve our attention more, for example, customer churn, credit card fraud, insurance fraud, and rare disorders. The cost of not identifying these minority classes correctly is often pro- hibitive. Typical approaches to deal with the problem of skewed class distributions can be classified into three cate- gories; the method of under-sampling for reducing majority data (Hart, 1968), the method of over-sampling for expanding minority data (Honda, Motizuki, Ho, & Okum- ura, 1997) and the method of multi-classifier committee (Chan, Fan, Prodromidis, & Stolfo, 1999). The under-sam- pling approach is to decrease the data of the majority class whereas the over-sampling approach is to increase the data of the minority class. The former may encounter the criti- cism of not fully making use of all the data, and the latter may induce noise. In contrast, the multi-classifier commit- tee approach proportionately partitions data into several subsets and generate multiple classifiers by training individ- ual subsets. The final prediction of the class is an aggregate from the outputs of all classifiers. On the other hand, concept drift refers to the varying concepts over time. A classifier predicts the class of an unknown example using the rules (boundaries) that are induced from training examples to discriminate the classes. If the learned concepts (the discrimination boundaries) are not changed, the classifier’s performance is about the same. However, when the concepts drift as time flies by, the ori- ginal classifier cannot perform well unless it keeps learning the emerging concepts and modifying its learnt structure accordingly. Spam emails are a good example of concept drift as users’ preferences change frequently and the topics of spam also change with contemporary trend. Based on the extent of the drift, three kinds of concept drift are usu- ally discussed in the literature: sudden drift, moderate drift and slow drift (Stanley, 2003). Based on the class distribu- tions, on the other hand, Forman (2006) classified the con- cept drift into three types: shifting class distribution, shifting subclass distribution, and fickle concept drift. To deal with concept drift, the classifiers should possess incre- mental learning capability that adapts itself to the environ- mental changes. 2.2. Adaptive learning in text categorization To improve the accuracy of a classifier, Wu, Pang, and Liu (2002) proposed a refinement approach to solving the problem of the misfit between data and models. Their method did not modify the classification algorithm itself; instead, it employed an incremental refinement procedure to deal with the model misfit problem. The concept is illus- trated in Fig. 1. It is an iterative process with two major steps. The first step is to train the classifier given a data W.-F. Hsiao, T.-M. Chang / Expert Systems with Applications 34 (2008) 1599–1608 1601 set. Since the classification result based on the trained clas- sifier is usually not perfect, the second step is to split the given data set based on the classifier’s current prediction results. The process then iterates to the first step to train a corresponding classifier for each split. It goes on repeatedly until a stop criterion is met (the F-measure in this case). Note that in this way, the data set in each split becomes purer and purer, which implies the concept represented at a more descending node would be more consistent. ICBC is motivated by Wu et al.’s work, which employed an adaptive refinement procedure to partition the data into sub-clusters until the concept represented within a cluster is consistent. However, unlike their computational effort to repeatedly train the classifier, ICBC clusters the data sim- ply by using clustering techniques. 2.3. Cluster analysis The clustering techniques can usually be divided into two categories: partitional and hierarchical techniques (Jain, Murty, & Flynn, 1999). Partitional clustering usually employs a measure (e.g. squared error) and attempts to optimize this measure (e.g. minimize the square error). The procedure is iterated several times until a stop criterion is met. Hierarchical clustering can be further divided into agglomerative and divisive clustering techniques. For agglomerative clustering, each data point is initially viewed as a cluster. Clusters are gradually combined together according to their similarities (with the minimal distance criterion). On the other hand, divisive clustering starts from a big cluster containing all the data points, followed by separating the cluster(s) iteratively. The measures employed to calculate the distances between clusters include single, complete, and average links. (1) Single link: the distance between two clusters can be obtained by finding the minimal distances between them, i.e., the distance between the two closest data points in the two clusters, respectively. (2) Complete link: the distance between two clusters is to calculate the maximal distances between them, i.e., the distance between the two farthest data points in the two clusters, respectively. (3) Average link: the distance between two clusters is the mean distance between all possible pairs of data points in the two clusters, respectively. 2.4. Spam filtering research With text-mining techniques adopted, a spam filter is usually established through the following steps. First, users manually classify their emails into legitimate and spam emails. A feature vector (keywords) is then formed based on TFIDF (term frequency · inverse document frequency) or its variants. Finally, a classifier is constructed using a classification algorithm that makes use of the feature vector to best discriminate the two categories (legitimate and spam). Payne and Edward (1997) employed a rule-based induc- tion method to automatically learn the classification rules. Their system, called ‘‘Magi’’, was a kind of agent program. It learned how to deal with new incoming emails through observing its user’s behaviors (e.g. forwarding, deleting, and storing) toward received emails. Magi used two thresh- olds, predictive threshold and confidence threshold, as the criteria for firing different level of delegated actions. The predictive threshold was used to determine whether Magi should provide its suggested action to the user for reference or not. On the other hand, the confidence threshold was set at a higher level for the system to automatically execute the action according to its prediction (e.g. forwarding, deleting, or storing). Bayes probabilistic model is another method that is commonly used in document classification (Lian, 2002). Sahami, Dumais, and Horvitz (1998) used Naı̈ve Bayes to analyze spam. Androutsopoulos, Koutsias, Chandrinos, and Spyropoulos (2000) employed Naı̈ve Bayes to con- struct email classifiers in spam filtering. Owing to its simple formulation and satisfactory performance, Naı̈ve Bayes classifiers were often used as the benchmark for compari- son purpose (Zhang & Oles, 2001). Naı̈ve Bayes, however, assumes that each document can only belong to one cate- gory (one-to-one correspondence) and the occurrence between individual terms is mutually independent (Nigam, McCallum, Thrun, & Mitchell, 2000), which are not realis- tic in the real world context. Therefore, recent researches of Bayesian Network tend to relax those assumptions. In addition, Drucker, Wu, and Vapnik (1999) employed support vector machine (SVM) to filter spam emails. Their results were compared to those of Ripper, Rocchio, and boosting decision tree. They found that with binary expres- sion in the feature vector of emails, SVM performs best. Overall, SVM was compatible with boosting decision tree but with shorter computational time. Luo and Zincir-Heywood (2005) designed an SOM-based sequence analysis (SBSA) system for spam filtering. The first part of SBSA was to represent document with a two-level hierarchical SOM architecture, and the second part was to employ a sequence-based k nearest neighbor (KNN) classi- fier for classification. SBSA was compared to the Naı̈ve Bayes classifier and showed superior performance. Delany and Cunningham (2006) proposed a spam filter system, ECUE, which employed instance-based learning technique to track concept drift. By learning from mis- classification emails (either spam or legitimate emails), ECUE could personalize to the specifics of users’ prefer- ences, and adapt to the changing natures of spam. How- ever, with its adoption of k nearest neighbor algorithm, ECUE had to re-perform the feature selection process and rebuild the instance-base every time the learning mech- anism was triggered by the misclassification of emails. Our proposed approach, ICBC, is compatible with ECUE on concept drift with a major exception that ICBC changes 1602 W.-F. Hsiao, T.-M. Chang / Expert Systems with Applications 34 (2008) 1599–1608 cluster structure incrementally rather than rebuild the instance base every time ECUE is triggered to learn. 3. Proposed approach The main purpose of ICBC is to handle the problem of skewed class distributions, and of concept drift. It basically includes two phases, each of which deals with the men- tioned two problems, respectively. Details of the phases, the classifier training phase and the incremental learning phase, are described as follows. 3.1. Classifier training phase The class-skewed problem is annoying because classes with more data (majority classes) usually dominate those with less data (minority classes). The consequence is that the classifiers give more weights on the majority classes to maintain higher prediction accuracy. To relieve this problem, one should find a way that can yield comparable weights to different classes while maintaining reasonable prediction accuracy. By inspecting the procedure of docu- ment classification task, we realize that a particular step that makes minority classes dominated is the feature selec- tion. The feature selection, commonly based on a variant of TFIDF (see, for example, Lee & Lee, 2006), is to extract a keyword vector that can discriminate different classes. Since it is done for the whole training document set, chances for keywords to be selected from the majority clas- ses are higher than from the minority classes. To overcome this problem, we propose ICBC, which bases itself on cluster analysis (as illustrated in Fig. 2). For each given class, ICBC groups the emails into several clusters (sub-concepts), and extract an equal number of keywords from each cluster. When a new email comes, ICBC compares the distances between this new email and the keyword vector extracted from each cluster, and assigns Feature Selection of TFIDF Fig. 2. The conceptual comparison of featu it to the cluster that has the minimal distance. By doing so, ICBC can effectively avoid the problem that the keywords in minority classes are usually insignificant to be selected such that documents from minority classes are often mis- classified to majority classes. The purpose of the classifier training phase is to resolve the problem of skewed class distributions. The details of this phase are shown in Fig. 3. Just like instance-based learning, the cluster-based ICBC simply performs feature selection task in this phase without training a classification model. ICBC first employs a stop word list (McCallum, 2002) and the Porter stemming algorithm (1980) to remove possi- ble noisy data and reduce the dimension of feature space. For each given class, ICBC applies TFIDF to extract key- words for clustering purpose. The clustering method adopted here is hierarchical with complete-linkage distance measure. Hierarchical, instead of partitional, clustering is used because the latter suffers from the instability that causes from different initial cluster centers and unknown number of clusters. The hierarchical clustering is also con- tributive to incremental learning, as explained later in the second phase. After the clustering is done, we re-extract an equal number of keywords from each cluster. The crite- rion to extract keywords now is TF without IDF since our aim is to find features that represent the cluster rather than those that have higher discrimination ability. Finally, the feature vector is determined for each cluster. To classify a new email, ICBC adopts the similarity comparison procedure as shown in Fig. 4. The essential key issue here is to select the cluster that most matches the new email, and assign the email the class that the matching cluster belongs to. It is referred to as using unsu- pervised learning to facilitate the supervised learning task. The similarity is calculated by comparing the keyword vec- tor of each cluster to the new email. The more keywords match, the more similar the email to the cluster. Finally, the cluster that has the largest number of matched key- Feature Selection of ICBC re selection between TFIDF and ICBC. Removed of stop words Porter’s stemmer Clustering Extracting keywords from each formed cluster Keywords extraction Training se t Removed of stop words Porter’s stemmer Clustering Extracting keywords from each formed cluster Keywords extraction Training set Fig. 3. The procedure of training. A new email Spam Calculate the similarity of the new email to each cluster (both legitimate and spam categories) Find the most similar cluster Legitimate email Which category does this cluster belong to? Feature vectors from each cluster Label the new email as spam and assign it into the cluster Label the new email as legitimate email and assign it into the cluster Fig. 4. The procedure to classify an unlabeled email. W.-F. Hsiao, T.-M. Chang / Expert Systems with Applications 34 (2008) 1599–1608 1603 words is selected as the most similar cluster, and the class of the new email is then determined. 3.2. Incremental learning phase The purpose of the incremental learning phase is to impose the incremental learning ability on ICBC. ICBC automatically triggers the incremental learning mechanism when it detects topic changes in spam (or legitimate emails), or when the users correct misclassified emails. In the former situation, ICBC will adapt itself by modifying the existent classification knowledge accordingly, whereas in the latter situation, ICBC quickly learns to avoid the same mistakes and personalizes the users’ email filtering preferences. The incremental learning ability of ICBC lies in its clus- ter structure of keywords. With new emails coming, the structure can be adjusted to accommodate new informa- tion. The procedure of this phase is illustrated in Fig. 5. With the same classification scheme for a new email in the previous phase, the incremental learning occurs when ICBC detects changes in its concepts (sub-clusters), or when users modify misclassified emails. As we classify a new email and store it to the corre- sponding cluster, we first consider whether the topic change in spam or legitimate emails is significant or not by check- ing the split condition of the cluster. If the coherence of the cluster no longer remains, the cluster will be automatically split into two or more sub-clusters, which denotes the generation of new topics. Fig. 6 illustrates the splitting process. Be it the original cluster or the split clusters, we re-extract the keyword vector of the cluster(s) to adapt to the changes. In addition, the cluster is split one at a time, if any, as we deal with the slow concept drift situation (Stanley, 2003) in incremental learning. Furthermore, users may trigger the incremental learning by modifying the classification results of the new emails. When the modification occurs, the misclassified email will A new email Spam Decide whether the cluster should be split after the new email is added. Re-extract keywords Calculate the similarity of the new email to each cluster (both legitimate and spam categories) Find the most similar cluster Legitimate email Which category does this cluster belong to? Feature vectors from each cluster Label the new email as spam and assign it into the cluster Label the new email as legitimate email and assign it into the cluster The user agree with the ICBC result Wait for another new email Yes No Re-extract keywords for this misclassification Fig. 5. The procedure of incremental learning in ICBC. Fig. 6. The illustration of cluster splitting process. 1604 W.-F. Hsiao, T.-M. Chang / Expert Systems with Applications 34 (2008) 1599–1608 be taken out of where it was stored and re-placed into the most matching cluster in the correct class. Again, for the matching cluster, we need to decide whether it should be split or not. Keywords of the misplaced cluster and the W.-F. Hsiao, T.-M. Chang / Expert Systems with Applications 34 (2008) 1599–1608 1605 matching cluster, or the split clusters are re-extracted to adapt to the changes. Note that ICBC changes the cluster structure before users correct the classification results. The main reason is that users usually process their emails in a batch mode, rather than one-by-one. For the efficiency purpose, ICBC will automatically activate possible structure change before users’ correction actions. 4. Experiments and results In this section we conduct three experiments to illustrate the feasibility of ICBC. We first explain our experimental design, followed by detailing the three experiments. 4.1. Experimental design We design three experiments to evaluate ICBC. Their main objectives are listed in Table 1. 4.1.1. Data collection To conduct the experiments, two data sets are collected. The first data set, used for Experiments I and II, is from Spamassassin’s1 non-spam email dataset and E.M. Canada’s2 spam email dataset. The non-spam emails from Spamassassin, with a total of 8624 records, serve as legiti- mate emails. The emails filtered from E.M. Canada with time frame from January 2002 to December 2003, with a total of 15,213 records, serve as spam. The second data set, used for Experiment III, is from Yahoo newsgroups. Six subjects (eight sub-subjects), dated from March 15, 2005 to March 21, 2005, are collected with statistics shown in Table 2. 4.1.2. Performance measure To evaluate the performance of the proposed classifier, we employ the measures that are widely used in email clas- sification. A confusion matrix can be formed according to the matching situations between the predicted classes by the classifier and their actual classes (as shown in Table 3). In this matrix, A represents the number that the classi- fier correctly classifies the legitimate mails; D is the number that the classifier correctly rejects the spam emails. B is the number that the classifier incorrectly classifies spam emails as legitimate ones, while C is the number that the classifier incorrectly rejects legitimate emails as spam. The common evaluation measures include Precision, Recall, F-measure, and Accuracy. Their corresponding definitions are as follows: Positive Precision ¼ A A þ B ð1Þ 1 Spamassassin email dataset: http://spamassassin.apache.org/ publiccorpus. 2 E.M. Canada spam email dataset: http://www.em.ca/~bruceg/spam. Negative Precision ¼ D C þ D ð2Þ Positive Recall ¼ A A þ C ð3Þ Negative Recall ¼ D B þ D ð4Þ F -measure ¼ 2pr p þ r ð5Þ Accuracy ¼ A þ D A þ B þ C þ D ð6Þ Positive precision represents the percentage of real legi- timate in the predicted legitimate emails. Negative preci- sion represents the percentage of real spam in the predicted spam. Positive recall is defined as the percentage of the true legitimate emails that are correctly predicted by the classifier. Negative recall is defined as the percentage of the true spam emails that are correctly predicted by the classifier. F-measure is the harmonic average of (positive or negative) precision and recall. Accuracy is the percent- age of all emails that are correctly classified by the classifier. 4.1.3. Benchmark comparison In the experiments, we employ the performance of the k nearest neighbors (KNN) approach (adopted in ECUE, Delany & Cunningham, 2006) as the benchmark for com- parison. KNN is a typical kind of instance-based learning methods, i.e., it simply performs feature selection task in the training phase without really training a classification model. When testing, it employs distance (dissimilarity) or similarity measures to find k instances in the training set that are most similar to the test data, and then aggre- gate the results of these k instances to be the prediction result for the test data. With no explicit knowledge structure during training, KNN is insignificantly influenced by the class-skewed problem. On the other hand, KNN does not have incre- mental learning mechanism. Whenever there are new com- ing instances, it simply stores them as the training data and re-perform feature selection task. When testing, it performs similarity matching of the test data with the entire training data. The computational cost of re-training, i.e., repetitive feature selection tasks, is expensive though. 4.2. Experiment I In Experiment I, under the condition of no significant skewed class distribution and of no concept drift, we exam- ine ICBC’s performance using the whole collected data from Spamassassin and E.M. Canada datasets. We ran- domly select 20% of the 8648 Spamassassin dataset and 20% of the 15,213 E.M. Canada dataset as the training set. The remaining 80% of each dataset is served as test data. The result is shown in Table 4. From Table 4, we first observe that both ICBC and KNN perform well under the large number of data http://spamassassin.apache.org/publiccorpus http://spamassassin.apache.org/publiccorpus http://www.em.ca/~bruceg/spam Table 1 Objectives of the experiments Objectives Experiment I To examine the performance of ICBC under the condition of no significant class-skewed problem and of no concept drift Experiment II To demonstrate the performance of ICBC under the situation of skewed class distributions Experiment III To justify the incremental learning capability of ICBC under the situation of a user’s preference changes Table 2 Statistics of collected data in Experiment III Subject Sub-subject Number Subject Sub-subject Number Sport Basketball 50 Entertainment Movie 60 Football 20 Health Weight loss 20 Business Economy 50 Technology Macintosh 20 Stock market 60 Weather Weather 20 Table 3 A confusion matrix in email classification Actual Legitimate Spam Predicted by the classifier Legitimate A B Spam C D Table 4 The classification result in Experiment I Precision Recall F-Measure Accuracy ICBC Legitimate 0.93177 0.92302 0.92738 0.94730 Spam 0.95762 0.96100 0.95931 KNN Legitimate 0.97538 0.92461 0.94931 0.96416 Spam 0.95927 0.98657 0.97273 1606 W.-F. Hsiao, T.-M. Chang / Expert Systems with Applications 34 (2008) 1599–1608 employed. Furthermore, it shows that ICBC performs slightly worse than KNN. The reason is that ICBC calcu- lates distances by comparing test data with the feature vec- tor of each cluster, while KNN compares test data with all training data. KNN, therefore, results in more accurate performance by enumerative comparison. Nevertheless, KNN suffers from its slowness in enumer- ative comparison. The time for classifying a new instance grows as the number of instances that KNN stores 0 20 60 40 100 80 120 140 160 180 10 20 30 40 50 60 70 80 90 100 No. of training instances S ec on ds ICBC KNN Fig. 7. The comparison of computational time between ICBC and KNN. increases. To illustrate, we perform a simple experiment that increases the number of training instances from 10 to 100, with an increment of 10, and fixes the number of testing instances to 200, and compute the computational time for ICBC with that for KNN. The result is shown in Fig. 7. From Fig. 7 it is obvious that the time for KNN grows as the number of training instances increases. On the con- trary, the time for ICBC seems to be steadily constant. ICBC largely decreases the influence of the number of training instances on the prediction speed. 4.3. Experiment II Experiment II is to examine ICBC on dealing with the issue of skewed class distributions. Traditional classifiers tend to prefer the majority classes and result in poor recall rates for minority classes. This problem may not even be salient if one employs the overall prediction accuracy as the performance measure, since the errors resulted from the minority classes would not have major impact on the overall prediction accuracy. To better account for the class-skewed problem, we employ the positive recall as the evaluation criterion in this experiment. To deliberately make significant skewed class distribu- tions, we select five training data sets, each of which con- sists of 200 spam and 10, 20, 30, 40, or 50 legitimate emails. The test set consisting of another 50 legitimate emails is used to evaluate the performance under each situation. The result is shown in Fig. 8. Overall, the positive recall rates for both ICBC and KNN deteriorate as the skewed class distribution phenomena become more significant. More importantly, ICBC outperforms KNN on all of the five cases. Originally, KNN should not be influenced by the problem of skewed class distributions. However, due to the global feature selection mechanism in text categori- zation, KNN still cannot avoid the feature selection prob- lem in minority classes. In contrast, ICBC selects equal number of features within clusters. This result illustrates Table 6 Recall rates from ICBC and KNN under incremental learning Size of test set ICBC KNN Movie Stock F- Measure Movie Stock F- Measure 0a 0.64000 0.74286 0.69143 0.46154 0.74074 0.60114 5 0.92857 0.93750 0.93304 0.70833 0.93548 0.82191 10 0.96774 0.96552 0.96663 0.93750 0.92857 0.93304 15 0.96774 0.96552 0.96663 0.92308 0.90909 0.91608 20 0.96774 0.96552 0.96663 0.85714 0.80000 0.82857 25 0.96774 0.96552 0.96663 0.88235 0.84615 0.86425 30 0.95082 0.94915 0.94999 0.85714 0.80000 0.82857 a Represents the initial framework without incremental learning. 10 20 30 40 50 The number of legitimate emails 1.0 0.8 0.6 0.4 0.2 0 0 P o si ti v e re ca ll ICBC KNN Fig. 8. The result of classification under skewed class distributions. W.-F. Hsiao, T.-M. Chang / Expert Systems with Applications 34 (2008) 1599–1608 1607 the effectiveness of the clustering strategy of ICBC on deal- ing with the issue of skewed class distributions. 4.4. Experiment III The objective of Experiment III is to evaluate the incre- mental learning capability of ICBC. Therefore, we deliber- ately create a scenario with significant changes of a user’s preference, a typical example of concept drift. The scenario is as follows. Suppose that an online user subscribes the Yahoo newsgroups and receives electronic news accord- ingly. At first, four subjects of news (Weather, Football, Weight Loss, and Macintosh) are sent to the user who con- siders emails of the former two subjects as legitimate and the latter two as spam. An initial classifier is trained according to the user’s original preferences. Gradually, another two subjects of news (Movie and Stock Market) are also sent to the user who begins to read the emails of Movie subject (legitimate) but still ignore the emails of Stock Market subject (spam). In this case, the classifier should be able to adapt itself to the new situation and make correct prediction. Abiding by this description, the data in this experiment are manipulated as shown in Table 5. The procedure to conduct this experiment is further explained as follows: (1) The initial classifier is built by a training data set that consists of 20 emails of Weather subject and 20 emails of Football subject as legitimate emails, and 20 emails of Weight Loss subject and 20 emails of Macintosh subject as spam. Table 5 Statistics of data manipulated in Experiment III Class Subject Number of emails Legitimate Weather 20 Football 20 Spam Weight loss 20 Macintosh 20 Incremental learning test data Legitimate Movie 30 Spam Stock market 30 Incremental learning validation data Legitimate Movie 30 Spam Stock market 30 (2) Movie and Stock are selected as the two targeted sub- jects emerging gradually, with emails of Movie sub- ject as the new legitimate emails and emails of Stock subject as the new spam. (3) To examine the effects of the incremental learning, we apply 30 emails of Movie subject and 30 emails of Stock subject as the incremental learning test data, in an increment of 5 emails fed into the classifier each time. We hold the rest 30 emails of Movie subject and the rest 30 emails of Stock subject as the validation data to evaluate the adapted classifier. The incremental learning results (recall rates) of ICBC and KNN are shown in Table 6. We observe that ICBC does not detect the concept drift at first because of few test data inputs. Nonetheless, it quickly adapts itself to the con- cept drift with more test data, and improves the recall rate accordingly. On the contrary, KNN may learn to adapt to the new situations, but in a much slow and unstable manner. We further plot the F-measure results from ICBC and KNN in Fig. 9 for visual comparison. It is clear that our ICBC outperforms KNN in the incremental learning pro- cess. In addition, KNN uses all the available data to re-per- form the feature selection task each time, while ICBC performs incremental learning and selects features within those varied clusters solely, and thus reduce the cost of re-training. 0 0.2 0.4 0.6 0.8 1 0 10 15 20 25 30 F -m ea su re ICBC KNN 5 Fig. 9. The F-measure comparison between ICBC and KNN. 1608 W.-F. Hsiao, T.-M. Chang / Expert Systems with Applications 34 (2008) 1599–1608 5. Concluding remarks With the prevailing of Internet, email has become an indispensable part of our daily life for communication. However, because of the prevalence of spam emails, how to detect and filter spam become a quite essential issue. In this study, we propose an incremental cluster-based clas- sification approach, ICBC, to improving the effectiveness of spam filtering. ICBC is a classification method based on cluster analysis. It first clusters concepts into several sub-concepts, and extract equal-sized representative features (keywords) from each sub-concept. With this procedure, concepts in minor- ity classes can have equal-sized features as those in majority classes, which lessens the class-skewed problem. At the same time, the cluster structure of ICBC can be adaptively changed, which imposes ICBC the capability of incremental learning. ICBC can accommodate changes in a quick and low-cost manner, and relieve the problem of concept drift. Three experiments are conducted to investigate the per- formance of ICBC. The result of Experiment I shows that, when the problem of skewed class distributions is not crit- ical, the performance of ICBC is comparable to that of KNN but its computational time is far less than that of KNN. In Experiment II, ICBC outperforms KNN when the problem of skewed classes is critical. The result shows that ICBC can deal with the issue of skewed class distribu- tions effectively. In addition, the result of Experiment III shows that the incremental learning of ICBC can accom- modate the concept drift, and thus reduce the prediction error. On the contrary, KNN, without incremental learn- ing, has slow-adapting and unstable performance and high re-training cost. To continue our research, we consider several potential future works. First, the initial cluster structure of ICBC has critical impact on the classification accuracy of its later modifications. A future research work is to consider approaches that can improve the initial cluster structure. WordNet or Ontology, for example, may be used to enforce the links between concepts and result in a more accurate initial clustering. Second, ICBC solely relies on the content analysis of emails without considering other useful information. A hybrid approach that includes con- tent analysis and useful email information (e.g. headers) may improve the effectiveness in spam filtering. Finally, it is worth implementing ICBC as a spam filtering system applicable in the real world. References Androutsopoulos, I., Koutsias, J., Chandrinos, K. V., & Spyropoulos, D. (2000). Learning to filter spam e-mail: a comparison of a naive bayesian and a memory-based approach. The 4th PKDD’s workshop on machine learning and textual information access. Chan, P. K., Fan, W., Prodromidis, A. L., & Stolfo, S. J. (1999). Distributed data mining in credit card fraud detection. IEEE Intelli- gent Systems, 14(6), 67–74. Delany, S. J., & Cunningham, P. (2006). ECUE: a spam filter that uses machine learning to track concept drift. Technical Report, Computer Science Department, The University of Dublin. Available from https:// www.cs.tcd.ie/publications/tech-reports/reports.06/TCD-CS-2006-05. pdf. Drucker, H., Wu, D., & Vapnik, V. N. (1999). Support vector machines for spam categorization. IEEE Transactions on Neural Networks, 10(5), 1048–1054. Fawcett, T. (2003). In vivo spam filtering: a challenge problem for KDD. ACM SIGKDD Explorations Newsletter, 5(2), 140–148. Forman, G. (2006). Tackling concept drift by temporal inductive transfer. SIGIR ’06 ACM. Hart, P. E. (1968). The condensed nearest neighbor rule. IEEE Transac- tions on Information Theory, IT-14, 515–516. Honda, T., Motizuki, H., Ho, T. B., & Okumura, M. (1997). Generating decision trees from an unbalanced data set. In Maarten vanSomeren., Gerhard Widmer (Eds.), Poster papers presented at the 9th European conference on machine learning (ECML) (pp. 68–77). Jain, A. K., Murty, M. N., & Flynn, P. J. (1999). Data clustering: a review. ACM Computing Surveys, 31(3), 264–323. Lee, C., & Lee, G. G. (2006). Information gain and divergence-based feature selection for machine learning-based text categorization. Information Processing and Management, 42(1), 155–165. Lian, Y. (2002). E-mail filtering, Master thesis. University of Sheffield, Department of Advanced Software Engineering. Luo, X., & Zincir-Heywood, N. (2005). Comparison of a SOM based sequence analysis system and naive bayesian classifier for spam filtering. In Proceedings of IEEE international joint conference on neural networks – IJCNN ’05 (Vol. 4, pp. 2571 – 2576). McCallum, A. (2002). Bow: a toolkit for statistical language modeling, text retrieval, classification and clustering. Available from http:// www.cs.cmu.edu/~mccallum/bow/. Monard, M. C., & Batista, G. E. A. P. A. (2002). Learning with skewed class distributions. In Advances in logic, artificial intelligence and robotics (LAPTEC’02). Nigam, K., McCallum, A. K., Thrun, S., & Mitchell, T. (2000). Text classification from labeled and unlabeled documents using EM. Machine Learning, 39, 103–134. Payne, T. R., & Edward, P. (1997). Interface agents that learn: an investigation of learning issues in a mail agent interface. Applied Artificial Intelligence, 1–32. Porter, M. (1980). An algorithm for suffix stripping. Program. Automated Library and Information Systems, 4(3), 130–137. Sahami, M., Dumais, S., Heckerman D., & Horvitz, E., (1998). A Bayesian approach to filtering junk e-mail. AAAI-98 workshop on learning for text categorization. Stanley, K. O. (2003). Learning concept drift with a committee of decision trees. Department of Computer Sciences Technical Report AI-03-302. Wu, H., Pang, T. H., Liu, B., & Li., X. (2002). A refinement approach to handling model misfit in text categorization. In Proceedings of the ACM SIGKDD international conference on knowledge discovery and data mining (pp. 23–26). Zhang, T., & Oles, F. J. (2001). Text categorization based on regularized linear classification methods. Information Retrieval, 4(1), 5–31. http://https://www.cs.tcd.ie/publications/tech-reports/reports.06/TCD-CS-2006-05.pdf http://https://www.cs.tcd.ie/publications/tech-reports/reports.06/TCD-CS-2006-05.pdf http://https://www.cs.tcd.ie/publications/tech-reports/reports.06/TCD-CS-2006-05.pdf http://www.cs.cmu.edu/~mccallum/bow/ http://www.cs.cmu.edu/~mccallum/bow/ An incremental cluster-based approach to spam filtering Introduction Literature review Special spam natures Adaptive learning in text categorization Cluster analysis Spam filtering research Proposed approach Classifier training phase Incremental learning phase Experiments and results Experimental design Data collection Performance measure Benchmark comparison Experiment I Experiment II Experiment III Concluding remarks References 
work_kavgx6kdrvhf3nliwd44ytwit4	----	International Journal of Engineering & Advanced Technology (IJEAT) International Journal of Recent Technology and Engineering (IJRTE) ISSN: 2277-3878, Volume-8, Issue-2S8, August 2019 178 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: B13430882S819/2019©BEIESP DOI: 10.35940/ijrte.B1343.0882S819  Abstract: The union of neighborhood has sent deletion coding, and current patterns recommend that the copying of the Ethernet will soon emerge. In certainty, couple of analysts would differ with the investigation of Byzantine faulttolerance. MegadynePuna, our new heuristic for SMPs, is the answer for these obstructions. Keywords :coupling.bayseian,models. I. INTRODUCTION The examination of journaling record frameworks has refinedonline algorithms, and current patterns recommend that the recreation of B-trees will soon emerge [5]. The basic tenet of this technique is the improvement of I/O [1],[ 3],[5]automata. Two properties make this arrangement particular: Mega-dyne Puna assesses the look aside cradle, and furthermore Megadyne Puna reenacts versatile configurations. The imitating of 802.11 work systems would enormously improve the advancement of open private key sets. So as to unravel this snag, we better see how working frameworks can be connected to the assessment of the World Wide Web. What's more, our technique examines compose ahead logging. Existing flimsy and decentralized frameworks utilize master frameworks to deal with the imitating of interferes. Positively, the impact on cyberinformatics of this has been viewed as hypothetical. The typical techniques for the development of steady hashing don't have any significant bearing around there. [2 ],[ 4],[6] Pushed by these recognitions, redundancy and A* search have been extensively improved by futurists. Actually, neural frameworks and I/O automata have a long history of partner thusly. Two properties make this game plan specific: our figuring is gotten from the evaluation of associated records, and besides MegadynePuna relies upon the norms of estimations. The fundamental guideline of this course of action is the examination of rasterization. Thusly, we see no Revised Manuscript Received on July 22, 2019. K. Yugendhar, Student Department of Information Technology, Bharath Institute of Higher Education and Research, Chennai, India Email: yugendhark887@gmail.com B.Sundarraj, Department of Computer Science and Engineering, Bharath Institute of Higher education and research, Chennai , IndiaEmail: sundarrajboobalan@gmail.com R.Velvizhi, Department of Computer Science and Engineering, Bharath Institute of Higher education and research, Chennai , IndiaEmail: velvizhisp@gmail.com reason not to use reliable time epistemologies to improve old style correspondence. [7],[ 9] ,[11] Our main contributions are as follows. To begin off with, we contend that albeit model checking and the look aside support are totally In consistent, the principal Bayesian calculation for the refinement of item arranged dialects by Davis and Raman keeps running in Θ(N) time. We focus our endeavors on contending that vacuum cylinders and online business are altogether inconsistent. [8],[ 10] ,[12] The remainder of the paper continues as pursues. We inspire the requirement for Web administrations. Thus, we contend the assessment of clog control. To understand this reason, we think about how hash tables can be connected to the comprehension of in-development recovery frameworks. Eventually, we conclude. [13], [15] ,[ 17] II. ARCHITECTURE In this fragment, we construct a designing for replicating Boolean reason. This seems to hold a great part of the time. We check that each fragment of our answer continues running in Ω(N) time, self-governing of each and every other part. See our related specific report [6] for nuances. [14],[ 16], [18] We guess that veritable advancement can get acquainted with the Turing machine without hoping to save the ordinary unification of the World Wide Web and recreated toughening. On a tantamount note, paying little mind to the results by J. Suzuki, we can abhorrent nearness strate that correspondence and DHCP can interface with location this issue. We use our heretofore examined outcomes as an explanation behind these suppositions. Figure 1: The relationship between MegadynePuna and Decoupling Expert Systems from Access Points in Cache Coherence K. Yugendhar, B.Sundarraj, R. Velvizhi Decoupling Expert Systems from Access Points in Cache Coherence 179 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: B13430882S819/2019©BEIESP DOI: 10.35940/ijrte.B1343.0882S819 active networks. Such a claim might seem unexpected but has ample historical precedence. [19],[21],[23] III. IMPLEMENTATION Regardless of the way that various skeptics said it was beyond the realm of imagination (most extraordinarily Zhou and Taylor), we assemble a totally working variation of our heuristic [7]. Similarly, paying little heed to the manner in which that we have not yet activity timized for security, this should be direct once we wrap up the hacked working structure. The client side library and the concentrated logging office must continue running in the identical JVM. Super dynePuna requires root get to in order to neutralize e – business. [20],[ 22], [24] IV. EVALUATION We by and by discussion about our evaluation. Our general appraisal attempts to show three hypotheses: (1) that tape drive space follows up on an extremely fundamental level dif-ferently on our work territory machines; (2) that simulated toughening has truly shown improved tenth percentile hit extent after some time; in conclusion that super pages have truly exhibited de-assessed square size after some time. A watchful peruser would now deduce that for clear reasons, we have decided not to envision floppy plate speed. Our evaluation attempts to make these centers indisputable. [25],[27],[29] Figure 3: The effective bandwidth of Mega-dynePuna, as a function of popularity Figure 4: These results were obtained by V. Ram-agopalan [8]; we reproduce them here for clarity. V. EXPERIMENTS AND RESULTS Is it possible to legitimize the phenomenal torments we took in our utilization? Correctly so. We ran four novel investigations: (1) we asked (and answered) what may happen if everything considered self-ruling article arranged lingos were used instead of association level confirmations; (2) we dogfooded our method in solitude work region machines, giving explicit thought to foreseen time since 1935; (3) we asked (and answered) what may happen if self-sufficiently regularly thorough cooling cess centers were used instead of hash tables; and [26],[28],[30] (4) we dogfooded MegadynePuna in solitude work zone machines, giving explicit thought to optical drive throughput. These experi-ments completed without noticable execution bottlenecks or the dim smoke that results from hardware frustration. [31],[33],[35] By and by for the climactic assessment of every one of the four ex-periments. Bugs in our system caused the unsta-ble direct all through the preliminaries. Cover up ther, the curve in Figure 2 should look surely understood; it is generally called G′ (N) = N. Goof bars have been excluded, since most of our data centers fell outside of 01 standard deviations from viewed suggests. [38],[40] Appeared in Figure 4, tests (1) and (4) enumerated above call attention to Mega-dynePuna's inertness. Note the staggering tail on the CDF in Figure 4, indicating exaggerated square size. Similarly, note that SCSI circles have less discretized expected throughput twists than do cemented compilers. Third, Gaussian electro-alluring disrupting impacts in our extensible overlay framework caused unreliable preliminary outcomes. At long last, we look at tests (1) and (4) recorded beforehand. Director botch alone can-not International Journal of Recent Technology and Engineering (IJRTE) ISSN: 2277-3878, Volume-8, Issue-2S8, August 2019 180 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: B13430882S819/2019©BEIESP DOI: 10.35940/ijrte.B1343.0882S819 speak to these results. Second, Gaussian electromagnetic aggravations in our event driven overlay framework caused wobbly preliminary outcomes [9]. The curve in Figure 3 should look conspicuous; it is generally called G−Y1(N) = N. Such a hypothesis is ordinarily a crucial ob-jective anyway has satisfactory chronicled need. [32],[34],[36] VI. CONCLUSION We nullified in our assessment that impedes and slight clients can cooperate to accomplish this yearning, and our framework is no exclusion to that standard. So additionally, one perhaps noteworthy drawback of MegadynePuna is that it can con-trol the advancement of destruction coding; we expect to address this in future work. We also presented an examination of Byzantine adjustment to non-basic disappointment. We expect to research more issues related to these issues in future work. [37],[39],[41] All in all, we talk about investigations (1) and (4) tallied beforehand. Executive goof alone can-not speak to these results. Second, Gaussian electromagnetic agitating impacts in our event driven overlay framework caused inconsistent test outcomes [9]. The twist in Figure 3 should look surely understood; it is generally called G−Y1(N) = N. Such a hypothesis is normally a central ob-jective yet has bounteous unquestionable need. REFERENCES [1] A., Rangarajan K.,Algorithm for automaton specification for exploring dynamic labyrinths,Indian Journal of Science and Technology,V-6,I-SUPPL5,PP-4554-4559,Y-2013 [2] P. Kavitha, S. Prabakaran “A Novel Hybrid Segmentation Method with Particle Swarm Optimization and Fuzzy C-Mean Based On Partitioning the Image for Detecting Lung Cancer” International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958, Volume-8 Issue-5, June 2019 [3] Kumaravel A., Meetei O.N.,An application of non-uniform cellular automata for efficient cryptography,2013 IEEE Conference on Information and Communication Technologies, ICT 2013,V-,I-,PP-1200-1205,Y-2013 [4] Kumarave A., Rangarajan K.,Routing alogrithm over semi-regular tessellations,2013 IEEE Conference on Information and Communication Technologies, ICT 2013,V-,I-,PP-1180-1184,Y-2013 [5] P. Kavitha, S. Prabakaran “Designing a Feature Vector for Statistical Texture Analysis of Brain Tumor” International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249-8958, Volume-8 Issue-5, June 2019 [6] Dutta P., Kumaravel A.,A novel approach to trust based identification of leaders in social networks,Indian Journal of Science and Technology,V-9,I-10,PP--,Y-2016 [7] Kumaravel A., Dutta P.,Application of Pca for context selection for collaborative filtering,Middle - East Journal of Scientific Research,V-20,I-1,PP-88-93,Y-2014 [8] Kumaravel A., Rangarajan K.,Constructing an automaton for exploring dynamic labyrinths,2012 International Conference on Radar, Communication and Computing, ICRCC 2012,V-,I-,PP-161-165,Y-2012 [9] P. Kavitha, S. Prabakaran “Adaptive Bilateral Filter for Multi-Resolution in Brain Tumor Recognition” International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-8 Issue-8 June, 2019 [10] Kumaravel A.,Comparison of two multi-classification approaches for detecting network attacks,World Applied Sciences Journal,V-27,I-11,PP-1461-1465,Y-2013 [11] Tariq J., Kumaravel A.,Construction of cellular automata over hexagonal and triangular tessellations for path planning of multi-robots,2016 IEEE International Conference on Computational Intelligence and Computing Research, ICCIC 2016,V-,I-,PP--,Y-2017 [12] Sudha M., Kumaravel A.,Analysis and measurement of wave guides using poisson method,Indonesian Journal of Electrical Engineering and Computer Science,V-8,I-2,PP-546-548,Y-2017 [13] Ayyappan G., Nalini C., Kumaravel A.,Various approaches of knowledge transfer in academic social network,International Journal of Engineering and Technology,V-,I-,PP-2791-2794,Y-2017 [14] Kaliyamurthie, K.P., Sivaraman, K., Ramesh, S. Imposing patient data privacy in wireless medical sensor networks through homomorphic cryptosystems 2016, Journal of Chemical and Pharmaceutical Sciences 9 2. [15] Kaliyamurthie, K.P., Balasubramanian, P.C. An approach to multi secure to historical malformed documents using integer ripple transfiguration 2016 Journal of Chemical and Pharmaceutical Sciences 9 2. [16] A.Sangeetha,C.Nalini,”Semantic Ranking based on keywords extractions in the web”, International Journal of Engineering & Technology, 7 (2.6) (2018) 290-292 [17] S.V.GayathiriDevi,C.Nalini,N.Kumar,”An efficient software verification using multi-layered software verification tool “International Journal of Engineering & Technology, 7(2.21)2018 454-457 [18] C.Nalini,ShwtambariKharabe,”A Comparative Study On Different Techniques Used For Finger – Vein Authentication”, International Journal Of Pure And Applied Mathematics, Volume 116 No. 8 2017, 327-333, Issn: 1314-3395 [19]M.S. Vivekanandan and Dr. C. Rajabhushanam, “Enabling Privacy Protection and Content Assurance in Geo-Social Networks”, International Journal of Innovative Research in Management, Engineering and Technology, Vol 3, Issue 4, pp. 49-55, April 2018. [20] Dr. C. Rajabhushanam, V. Karthik, and G. Vivek, “Elasticity in Cloud Computing”, International Journal of Innovative Research in Management, Engineering and Technology, Vol 3, Issue 4, pp. 104-111, April 2018. [21] K. Rangaswamy and Dr. C. Rajabhushanamc, “CCN-Based Congestion Control Mechanism In Dynamic Networks”, International Journal of Innovative Research in Management, Engineering and Technology, Vol 3, Issue 4, pp. 117-119, April 2018. [22] Kavitha, R., Nedunchelian, R., “Domain-specific Search engine optimization using healthcare ontology and a neural network backpropagation approach”, 2017, Research Journal of Biotechnology, Special Issue 2:157-166 [23]Kavitha, G., Kavitha, R., “An analysis to improve throughput of high-power hubs in mobile ad hoc network” , 2016, Journal of Chemical and Pharmaceutical Sciences, Vol-9, Issue-2: 361-363 [24] Kavitha, G., Kavitha, R., “Dipping interference to supplement throughput in MANET” , 2016, Journal of Chemical and Pharmaceutical Sciences, Vol-9, Issue-2: 357-360 [25] Michael, G., Chandrasekar, A.,”Leader election based malicious detection and response system in MANET using mechanism design approach”, Journal of Chemical and Pharmaceutical Sciences(JCPS) Volume 9 Issue 2, April - June 2016 . [26] Michael, G., Chandrasekar, A.,”Modeling of detection of camouflaging worm using epidemic dynamic model and power spectral density”, Journal of Chemical and Pharmaceutical Sciences(JCPS) Volume 9 Issue 2, April - June 2016 . [27] Pothumani, S., Sriram, M., Sridhar, J., Arul Selvan, G., Secure mobile agents communication on intranet,Journal of Chemical and Pharmaceutical Sciences, volume 9, Issue 3, Pg No S32-S35, 2016 [28] Pothumani, S., Sriram, M., Sridhar , Various schemes for database encryption-a survey, Journal of Chemical and Pharmaceutical Sciences, volume 9, Issue 3, Pg NoS103-S106, 2016 [29] Pothumani, S., Sriram, M., Sridhar, A novel economic framework for cloud and grid computing, Journal of Chemical and Pharmaceutical Sciences, volume 9, Issue 3, Pg No S29-S31, 2016 [30] Priya, N., Sridhar, J., Sriram, M. “Ecommerce Transaction Security Challenges and Prevention Methods- New Approach” 2016 ,Journal of Chemical and Pharmaceutical Sciences, JCPS Volume 9 Issue 3.page no:S66-S68 . [31] Priya, N.,Sridhar,J.,Sriram, M.“Vehicular cloud computing security issues and solutions” Journal of Chemical and Pharmaceutical Sciences(JCPS) Volume 9 Issue 2, April - June 2016 [32] Priya, N., Sridhar, J., Sriram, M. “Mobile large data storage security in cloud computing environment-a new approach” JCPS Volume 9 Issue 2. April - June 2016 [33] Anuradha.C, Khanna.V, “Improving network performance and security in WSN using decentralized hypothesis testing “Journal of Chemical and Pharmaceutical Sciences(JCPS) Volume 9 Issue 2, Decoupling Expert Systems from Access Points in Cache Coherence 181 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: B13430882S819/2019©BEIESP DOI: 10.35940/ijrte.B1343.0882S819 April - June 2016 . [34] Anuradha.C, Khanna.V, “A novel gsm based control for e-devices“ Journal of Chemical and Pharmaceutical Sciences(JCPS) Volume 9 Issue 2, April - June 2016 . [35] Anuradha.C, Khanna.V, “Secured privacy preserving sharing and data integration in mobile web environments “ Journal of Chemical and Pharmaceutical Sciences(JCPS) Volume 9 Issue 2, April - June 2016 . [36] Sundarraj, B., Kaliyamurthie, K.P. Social network analysis for decisive the ultimate classification from the ensemble to boost accuracy rates 2016 International Journal of Pharmacy and Technology 8 [37] Sundarraj, B., Kaliyamurthie, K.P. A content-based spam filtering approach victimisation artificial neural networks 2016 International Journal of Pharmacy and Technology 8 3. [38] Sundarraj, B., Kaliyamurthie, K.P. Remote sensing imaging for satellite image segmentation 2016 International Journal of Pharmacy and Technology 8 3. [39] Sivaraman, K., Senthil, M. Intuitive driver proxy control using artificial intelligence 2016 International Journal of Pharmacy and Technology 8 4. [40] Sivaraman, K., Kaliyamurthie, K.P. Cloud computing in mobile technology 2016 Journal of Chemical and Pharmaceutical Sciences 9 2. [41] Sivaraman, K., Khanna, V. Implementation of an extension for browser to detect vulnerable elements on web pages and avoid click jacking 2016 Journal of Chemical and Pharmaceutical Sciences 9 2. AUTHORS PROFILE K. Yugendhar, Student Department of Information Technology, Bharath Institute of Higher Education and Research, Chennai, India . B.Sundarraj, Assistant Professor, Department of Computer Science & Engineering, Bharath Institute of Higher Education and Research, Chennai, India R.Velvizhi ,Assistant Professor, Department of Computer Science & Engineering, Bharath Institute of Higher Education and Research, Chennai, India . . 
work_kb2btv7j35cqdf75a6hukvutga	----	Best Mortgages Canada | Valueland Mortgages Skip to main content Call 1-877-479-9998, Or Click to Contact Us NOW! X Main navigation Apply Now Apply Online Download Forms Purchases Mortgage Process Get Pre-Approval Compare Mortgages Mortgage FAQs Refinance Refinance Basics Debt Consolidation Refinance Process Mortgage FAQs Transfer Transfer Process Why to Transfer Mortgage FAQs Solutions Self Employed Net Worth Program New Immigrants Credit Problems Second Mortgages Rental Properties About Us Contact Us About Us Testimonials Blog Articles Welcome To Valueland!   Featured Mortgages 5 Year Variable at 1.40% 5 Year Variable Rate ... and more! read more No Stress Test Mortgages Get More Mortgage Than Your Bank! read more Net Worth Program Mortgages, Based on Total Assets! read more Why Choose Us Solutions. Great Rates. FREE Consultation. 18 Years of Excellent Service. Mortgage Solutions Mortgages for the Self Employed If you are self employed ... let us help you get the mortgage you want Get A Mortgage with Low Credit Scores Banks said No? We say Yes! Mortgage Pre-Qualification You get more value by getting a pre-qualification from Valueland. Testimonials Efficient and Professional Vanessa says, Quick turn around time, Simply great experience! read more Knowledge and Patience Philbert says, I've got Great Service, Patience and Knowledge. read more Extra Miles Ray says, Don't reimburse my appraisal cost. You deserve more! read more Valueland's Blog Tips for a mortgage renewal Valueland Fri, 03/26/2021 - 16:01 For most home owners, a mortgage renewal process is an inevitable occurrence in our life. Auto renew mortgage with your current bank (lender) seems convenient and easy, but is it the best option or right decision? The answer is “not always”. At renewal time, it is the perfect opportunity for financial review, adjustments and shopping for a lower rate. Here are five tips for homeowners to save money when looking to renew the mortgages:     1. Update your financial goal – your financial situation could change. It’s the perfect time for debts consolidation if you had other major debts. Maybe it’s time for equity take out if you intended to do some investments. Or maybe you planned to do title change adding/removing someone from title of property, it’s right time to do so. Start earlier – 3 or 4 month prior to mortgage renewal, start to ask for mortgage rates in the market. You will need time to do research and get yourself well educated in order to make best/right decision. Rate type – fixe or variable? Short term or long term? It is a decision you have to make depends upon your financial plan and personal situation. Research indicates that variable rates outperform fixed rates over the long-run, with fixed rates outperforming only a minority of the time and mainly at the trough of economic cycle. Get rate locked – if you decide to switch lender for your financial needs, it’s very important to start early to submit the application and lock in a rate. Generally you can lock in a rate 3 months prior to renewal date, if rates go down, you will have the lowered rate. If rates go up, your locked in low rate protected you from the increase.  Know the mortgage you get – interest rates are important for mortgage renewal, but it is not all. You should understand all the features of the mortgage such as penalty clause, fees incurred, special conditions (sale clause, high penalty clause, single mortgage or combined with HELOC as all in one product, etc.). For example, if you intended to sell the property during the next term, fixed rates are probably not a good option since penalty for breaking fixed rates are much higher than variable rates     Please click here for the best rate for mortgage transfers or if you would like to know what options you have, please reach out Valueland Mortgage at 1 877 479 9998 or contact us here. Residential Market Commentary Valueland Tue, 03/23/2021 - 11:55 The latest numbers from the Canadian Real Estate Association help to explain why worries about a bubble are on the rise. Sales activity in February jumped nearly 40% compared to a year earlier, setting a new record.  Sales rose nearly 7% compared to January.  The national average price surged by 25% year-over-year.  New listings rebounded month-over-month in February but inventories remain at record lows.  Nationally there is just 1.8 months of supply.  The Bank of Canada has expressed concerns about overheating.  Governor Tiff Macklem has noted that there are signs that real estate speculation is on the rise. "What we get worried about is when we start to see extrapolative expectations, when we start to see people expecting the kind of unsustainable price rises we've seen recently go on indefinitely, and they're basing their decision on those kinds of assumptions," Macklem warned earlier this month.   More Articles About Valueland Since 2002, Valueland has been the bridge between consumers and mortgage lenders. Valueland Mortgages is licensed by FSRA (Financial Services Regulatory Authority of Ontario) and Financial Institutions Commission of British Columbia. It services a number of provinces in Canada. Contact Information Valueland Mortgages 7800 Woodbine Ave, Suite 202 Markham, ON. L3R 2N7 Email: info@valueland.ca Local #: 416-238-2698 Phone: 1-877-479-9998 Fax: 1-877-479-9968   Questions? Call backs? Please ... Contact Us Now! Copyright © 2002-2021, Valueland Mortgages, Brokerage License #: 10133 
work_kbc6ibbv3fh53pefhc4tqr4plu	----	Microsoft Word - esa1999.doc Page 1 An Adaptive Knowledge Acquisition System using Generic Genetic Programming Man Leung Wong Department of Computing and Decision Sciences Lingnan University Tuen Mun Hong Kong mlwong@ln.edu.hk Page 2 An Adaptive Knowledge Acquisition System using Generic Genetic Programming Abstract The knowledge acquisition bottleneck greatly obstructs the development of knowledge-based systems. One popular approach to knowledge acquisition uses inductive concept learning to derive knowledge from examples stored in databases. However, existing learning systems cannot improve themselves automatically. This paper describes an adaptive knowledge acquisition system that can learn first-order logical relations and improve itself automatically. The system is composed of an external interface, a biases base, a knowledge base of background knowledge, an example database, an empirical ILP learner, a meta- level learner, and a learning controller. In this system, the empirical ILP learner performs top-down search in the hypothesis space defined by the concept description language, the language bias, and the background knowledge. The search is directed by search biases which can be induced and refined by the meta-level learner based on generic genetic programming. It has been demonstrated that the adaptive knowledge acquisition system performs better than FOIL on inducing logical relations from perfect or noisy training examples. The result implies that the search bias evolved by evolutionary learning is better than that of FOIL which is designed by a top researcher in the field. Consequently, Generic genetic programming is a promising technique for implementing a meta-level learning system. The result is very encourage as it suggests that the process of natural selection and evolution can successfully evolve a high performance learning system. Area: Evolutionary Computation, Genetic Programming, Knowledge Acquisition 1. Introduction The knowledge acquisition bottleneck greatly obstructs the development of knowledge-based systems. One popular approach to knowledge acquisition uses inductive concept learning to derive knowledge from examples stored in databases. The knowledge acquired can be expressed in different knowledge representations such as first-order logical relations, decision trees, decision lists, and production rules. Existing learning systems such as CART (Breiman et al. 1984), C4.5 (Quinlan Page 3 1992), ASSISTANT (Cestnik et al. 1987), AQ15 (Michalski et al 1986), and CN2 (Clark and Niblett 1989) use attribute-value language for representing the training examples and the induced knowledge and allow a finite number of objects in the universe of discourse. This representation limits them to learn only propositional descriptions in which concepts are described in terms of values of a fixed number of attributes. Dzeroski and Lavrac show that Inductive Logic Programming (ILP) can be used to induce knowledge represented as first-order logical relations (Dzeroski and Lavrac 1993, Dzeroski 1996). ILP is more powerful than traditional inductive learning methods because it uses an expressive first-order logic framework and facilitates the application of background knowledge. In this formalism, domain knowledge represented in the form of relations can be used in the induced relational descriptions of concepts. Moreover, ILP has a strong theoretical foundation from logic programming and computational learning theory. The task of inducing first-order logical relations can be formulated as a search problem (Mitchell 1982) in a hypotheses space of logical relations. Various approaches (Quinlan 1990; 1996, Muggleton and Feng 1990) differ mainly in the search strategy and the heuristics used to guide the search. The search space is extremely large, so strong heuristics are required to manage the problem. Most systems are based on a greedy search strategy. They generate a sequence of logical relations from general to specific (or from specific to general) until a consistent relation is found. Each relation in the sequence is obtained by specializing (or generalizing) the previous one. For example, FOIL (Quinlan 1990, 1996) applies a hill climbing search strategy guided by an information-gain heuristic to search relations from general to specific. But these strategies and heuristics are not always applicable because these systems may become trapped in local maxima. In order to overcome this problem, non greedy strategies should be adopted. Moreover, existing ILP systems cannot improve themselves automatically. In this paper, we describe an adaptive knowledge acquisition system that induces first-order logical relations and improves itself during learning. We formulate Page 4 the definitions of inductive concept learning and adaptive knowledge acquisition in the next section. The system is based on a generic genetic programming approach that is presented in Section 3. A generic top-down first-order learning algorithm is described in the next section. The fifth section contains a description of a meta-level learner that induces search bias. The experimentation and some evaluations of the system are reported in Section six. Finally, the conclusion is presented in the last section. 2. INDUCTIVE CONCEPT LEARNING AND ADAPTIVE KNOWLEDGE ACQUISITION The goal of machine learning is to develop techniques and tools for building intelligent learning machines. Machine learning paradigms include inductive, deductive, genetic-based, and connectionist learning. Multi-strategy learning integrates several learning paradigms. This section focuses on supervised inductive concept learning. If U is a universal set of observations, a concept C is formalized as a subset of observations in U. Inductive concept learning finds descriptions for various target concepts from positive and negative training instances of these concepts. In machine learning, formal languages for describing observations and concepts are called object and concept description languages respectively. Typically, object description languages are attribute-value pair descriptions and first-order languages of Horn clauses. Concepts can be described extensionally or intensionally. A concept is described extensionally by listing the descriptions of all of its instances (observations). Thus extensional concepts are represented in the object description language. On the other hand, intensional concepts are expressed in a separate concept description language that permits compact and concise concept descriptions. Typical concept description languages are decision trees, decision lists, production rules, and first-order logic. Inductive concept learning can be viewed as searching the space of hypothesis descriptions. A bias is a mechanism employed by a learning system to constrain the search for target hypotheses. A search bias determines how to conduct the search in Page 5 the hypothesis space while a language bias determines the size and structure of the hypothesis space. A strong search bias, such as the hill-climbing search strategy, employs existing knowledge about the size and structure of the hypothesis space to exploit promising solutions of the space, thus it can find the target concept quickly. But it may trap the system in a local maximum. A weak search bias, such as depth-first and breath-first search, explores the space completely; the learner is guaranteed to find the target concept that can be represented by the concept description language. Nevertheless, a weak bias is very inefficient. In other words, the search bias introduces the efficiency/completeness tradeoff into a learning system. A strong language bias defines a less expressive description language such as the propositional logic. The hypothesis space created by the bias is comparatively smaller and the learning can be performed more efficiently. But the learner may fail to find the target concept which is not contained in the small hypothesis space. A weak bias defines a larger space and thus the target concept is more likely to be expressible in the space. The disadvantage is that the learner is less efficient. The language bias introduces the efficiency/expressiveness tradeoff into a learning system. Background knowledge B is declarative prior knowledge that can be used by either the search bias to direct the search more efficiently, or the language bias to express the hypothesis space in a more natural and concise way. Background knowledge plays an important role in relational concept learning. Relational concept learning induces a new relation for the target concept (i.e., the target predicate) from training examples and known relations from the background knowledge. The training examples, the hypothesis space, and the background knowledge are represented in first order Horn clause languages (Muggleton 1992). Tradeoffs between expressiveness and efficiency are introduced by some additional restrictions on the three languages representing the examples, the hypothesis space, and the background knowledge. The background knowledge B provides definitions of known predicates qi which can be used in the definition of the target predicate p. It also provides Page 6 additional information to ease the learning. The information includes argument types, symmetry of predicates in pairs of arguments, input/output modes, rule models, predicate sets, parametrized languages, integrity constraints, determinations and any knowledge that can modify the operation of the search and language biases (Lavrac and Dzeroski 1994). An adaptive knowledge acquisition system is a relational concept learning system that can improve itself on the learning capability. It maintains various sets of background knowledge and biases. It improves itself by modifying its biases and background knowledge. Since a hypothesis space for learning is defined through the concept description language, the language bias, and the background knowledge, the size and structure of the hypothesis space can be modified by changing the language bias and the background knowledge. The search strategy and heuristics are changed if the system's search biases are modified. Here, we formulate the task of an adaptive knowledge acquisition system in Table 1. Given: -A set E of positive E+ and negative E- training examples of the target predicate p. Training examples are represented as ground atoms -A concept description language L -A set of learning biases BIASES -A set of various background knowledge BKs Find: -A modified set of learning biases BIASES' -A modified set of background knowledge BKs' -A concept definition H for the target predicate p expressible in L such that H is complete and consistent with respect to (w.r.t.) the training examples E and a background knowledge B' in BKs' H is complete if every positive example e+ in E+ is covered by H w.r.t. the background knowledge B. i.e. B ∪ H |= e+ H is consistent if no negative example e- in E- is covered by H w.r.t. the background knowledge B. i.e. B ∪ H |≠ e- Table 1: The definition of adaptive knowledge acquisition Page 7 External Interface Learning Controller BKbase Biases Base Example database Meta-Level Learner Empirical ILP Learner Data flow Control flow Figure 1: The logical organization of an adaptive knowledge acquisition system The logical organization of our system is depicted in Figure 1. Its components are introduced as follows: (1) External interface: It provides a user-friendly interface between the system and users. It accepts training examples, a set BKs of background knowledge, and a set BIASES of biases and transfers them through the learning controller to the example database, BKbase, and biases base respectively. The interface also provides commands for users to query about the results of an adaptive learning task and to directly control the operations of the learning controller. (2) Biases base: It is a knowledge base that stores all learning biases. Biases can be retrieved, added, deleted, and modified through the interface of this knowledge base. (3) BKbase: It stores various background learning knowledge that can be used in inductive learning. Background knowledge can be retrieved, added, deleted, and modified through the interface of BKbase. Since each entity of it is in fact a complex structure representing background knowledge, BKbase is implemented using object-oriented techniques. Page 8 (4) Examples database: It stores the training examples. (5) Empirical ILP learner: It induces first-order logical relations from the training examples, given a concept description language, a specific background knowledge, a search bias, and a language bias. A search of the hypothesis space can be performed bottom-up or top-down. Bottom-up techniques start from the training examples and search the space by employing various generalization operators. Top-down techniques start from the most general concept descriptions, and search the space by using various specialization operators. Top-down techniques are better suited for learning from imperfect examples because a large number of data are available in every specialization step and the system can employ various statistical techniques to decide how to perform the specialization. Moreover, top-down search can easily be guided by the search bias. In Section 4, a generic top-down first-order learning algorithm is described. (6) Meta-level learner: It learns search biases, language biases, and background knowledge. Search and language biases can be represented declaratively or procedurally. In this paper, we apply GGP to implement the meta-level learner that induces procedural biases. The description of GGP is presented in Section 3. (7) Learning controller: It is a knowledge-based system that controls the empirical ILP learner and the meta-level learner. The knowledge used by the learning controller can be updated by the meta-level learner. 3. Generic Genetic Programming (GGP) Generic Genetic Programming (GGP) is a novel approach that combines genetic programming (Koza 1992; 1994, Kinnear 1994) and inductive logic programming (Quinlan 1990; 1996, Muggleton 1992). Using GGP, programs in various programming languages can be evolved. The approach is also powerful enough to handle context-sensitive information and domain-dependent knowledge which can be used to accelerate the learning speed and/or improve the quality of the programs. Page 9 GGP can induce programs in various programming languages. This is achieved by accepting or choosing grammars of different languages to produce programs in these languages. Most modern programming languages are specified in the notation of BNF (Backus-Naur form) which is a kind of context-free grammars (CFGs). However, GGP is based on logic grammars because CFGs (Hopcroft and Ullman 1979, Lewis and Rapadimitrion 1981) are not expressive enough to represent context-sensitive information for some languages and domain-dependent knowledge of the target program being induced. This section first introduces the formalism of logic grammars followed by the descriptions of GGP. 3.1. Introduction to logic grammars Logic grammars are the generalizations of CFGs. Their expressivenesses are much more powerful than those of CFGs, but equally amenable to efficient execution. In this paper, logic grammars are described in a notation similar to that of definite clause grammars (Pereira and Warren 1980, Pereira and Shieber 1987, Sterling and Shapiro 1986). The logic grammar for some simple S-expressions in Table 2 will be used throughout this section. 1: start -> [(*], exp(W), exp(W), exp(W), [)]. 2: start -> {member(?x,[W, Z])}, [(*], exp-1(?x), exp-1(?x), exp-1(?x), [)]. 3: start -> {member(?x,[W, Z])}, [(+], exp-1(?x), exp-1(?x), exp-1(?x), [)]. 4: exp(?x) -> [(/ ?x 1.5)]. 5: exp-1(?x) -> {random(1,2,?y)}, [(/ ?x ?y)]. 6: exp-1(?x) -> {random(3,4,?y)}, [(- ?x ?y)]. 7: exp-1(W) -> [(+ (- W 11) 12)]. Table 2: A logic grammar A logic grammar differs from a CFG in that the logic grammar symbols, whether terminal or non-terminal, may include arguments. The arguments can be any term in the grammar. A term is either a logic variable, a function or a constant. A variable is represented by a question mark ? followed by a string of letters and/or digits. A function is a grammar symbol followed by a bracketed n-tuple of terms and Page 10 a constant is simply a 0-arity function. Arguments can be used in a logic grammar to enforce context-dependency. Thus, the permissible forms for a constituent may depend on the context in which that constituent occurs in the program. Another application of arguments is to construct tree structures in the course of parsing, such tree structures can provide a representation of the semantics of the program. The terminal symbols enclosed in square brackets correspond to the set of words of the language specified. For example, the terminal [(- ?x ?y)] creates the constituent (- 1.0 2.0) of a program if ?x and ?y are instantiated respectively to 1.0 and 2.0. Non-terminal symbols are similar to literals in Prolog, exp-1(?x) in Table 2 is an example of non-terminal symbols. Commas denote concatenation and each grammar rule ends with a full stop. The right-hand side of a grammar rule may contain logic goals and grammar symbols. The goals are pure logical predicates for which logical definitions have been given. They specify the conditions that must be satisfied before the rule can be applied. For example, the goal member(?x, [W, Z]) in Table 2 instantiates the variable ?x to either W or Z if ?x has not been instantiated, otherwise it checks whether the value of ?x is either W or Z. If the variable ?y has not been bound, the goal random(1, 2, ?y) instantiates ?y to a random floating point number between 1 and 2. Otherwise, the goal checks whether the value of ?y is between 1 and 2. Domain-dependent knowledge can be represented in logic goals. For example, consider the following grammar rule: a-useful-program -> first-component(?X), {is-useful(?X, ?Y)}, second-component(?Y). This rule states that a useful program is composed of two components. The first component is generated from the non-terminal first-component(?X). The logic variable ?X is used to store semantic information about the first component produced. The logic goal then determines whether the first component is useful according to the semantic information stored in ?X. Domain-dependent knowledge Page 11 about which program fragments are useful is represented in the logical definition of this predicate. If the first component is useful, the logic goal is- useful(?X, ?Y) is satisfied and some semantic information is stored into the logic variable ?Y. This information will be used in the non-terminal second- component(?Y) to guide the search for a good program fragment as the second component of a useful program. The special non-terminal start corresponds to a program of the language. In Table 2, some grammar symbols are shown in bold-face to identify the constituents that cannot be manipulated by genetic operators. For example, the last terminal symbol [)] of the second rule is revealed in bold-face because every S-expression must be ended with a ')'. The number before each rule is a label for later discussions. It is not part of the grammar. 3.2. Representations of programs One of the fundamental contributions of GGP is in the representations of programs in different programming languages appropriately so that initial population can be generated easily and the genetic operators such as reproduction, mutation, and crossover can be performed effectively. A program can be represented as a derivation tree that shows how the program has been derived from the logic grammar. GGP applies deduction to randomly generate programs and their derivation trees in the language declared by the given grammar. These programs form the initial population. For example, the program (* (/ W 1.5) (/ W 1.5) (/ W 1.5)) can be generated by GGP given the logic grammar in Table 2. It is derived from the following sequence of derivations: start => [(*] exp(W) exp(W) exp(W) [)] => [(*] [(/ W 1.5)] exp(W) exp(W) [)] => [(*] [(/ W 1.5)] [(/ W 1.5)] exp(W) [)] => [(*] [(/ W 1.5)] [(/ W 1.5)] [(/ W 1.5)] [)] => [(* (/ W 1.5) (/ W 1.5) (/ W 1.5))] This sequence of derivations can be represented as the derivation tree depicted in Figure 2. Page 12 In literature, the terms derivation trees and parse trees are usually used interchangeably. However, we will use the term derivation trees to refer to the tree structures in our framework and the term parse trees to refer to those in GP. The bindings of logic variables are enclosed in a pair of braces. The sub-trees enclosed in a dashed rectangular are frozen. In other words, they are generated by bold-faced grammar symbols and they cannot be modified by genetic operators. An advantage of logic grammars is that they specify what is a legal program without any explicit reference to the process of program generation and parsing. Furthermore, a logic grammar can be translated into an efficient logic program that can generate and parse the programs in the language declared by the logic grammar (Pereira and Warren 1980, Pereira and Shieber 1987, Abramson and Dahl 1989). In other words, the process of program generation and parsing can be achieved by performing deduction using the translated logic program. Consequently, the program generation and analysis mechanisms of GGP can be implemented using a deduction mechanism based on the logic programs translated from the grammars. [(*] exp(W) exp(W) exp(W) [)] start [(/ ?x 1.5)] {?x/W} [(/ ?x 1.5)] {?x/W} [(/ ?x 1.5)] {?x/W} Figure 2: A derivation tree of the S-expression in Lisp (* (/ W 1.5) (/ W 1.5) (/ W 1.5)) This method of translating a logic grammar into a logic program is common in the field of natural language processing (Pereira and Warren 1980, Pereira and Shieber 1987, Abramson and Dahl 1989). The original idea of this approach is to Page 13 rephrase the special purpose formalism of CFGs into a general purpose first-order predicate logic (Kowalski 1979, Colmerauer 1978, Pereira and Warren 1980). This approach is further refined and generalized to Definite Clause Grammars (DCGs) which can handle the properties of context-dependency of natural languages effectively. Since DCGs, a kind of logic grammars, can be translated into efficient logic programs automatically, parsers and generators for the corresponding natural languages can be obtained easily. In other words, researchers in the field of natural language processing only declare the grammar for a particular natural language, and the translation process will produce the corresponding parser and generator for them. Moreover, for some cases, the same logic program can be used as both a parser and generator at the same time. Alternatively, initial programs can be induced by other learning systems such as FOIL (Quinlan 1990; 1996) or given by the user. GGP analyzes each program and creates the corresponding derivation tree. 3.3. The evolution process of GGP In GGP, populations of programs are genetically bred using the Darwinian principle of survival and reproduction of the fittest along with genetic operations appropriate for creating programs. GGP starts with an initial population of programs generated randomly, induced by other learning systems, or provided by the user. Logic grammars provide declarative descriptions of the valid programs that can appear in the initial population. A fitness function must be defined by the user to evaluate the fitness values of the programs. Typically, each program is run over a set of fitness cases and the fitness function estimates its fitness by performing some statistical operations (e.g. average) to the values returned by this program. The initial programs in generation 0 are normally incorrect and have poor performances. However, some programs in the population will be fitter than others. Fitness of each program in the generation is estimated and the following process is iterated over many generations until the termination criterion is satisfied. The reproduction, sexual crossover, and asexual mutation are used to create new Page 14 generation of programs from the current one. The reproduction involves selecting a program from the current generation and allowing it to survive by copying it into the next generation. Either fitness proportionate or tournament selection can be used. The crossover is used to create offspring programs from two parental programs selected. Mutation creates a modified offspring program from a parental program selected. Unlike crossover, the offspring program is usually similar to the parent program. Logic grammars are used to constraint the offspring programs that can be produced by these genetic operations. This algorithm will produce populations of programs which tend to exhibit increasing average of fitness. Finally, GGP returns the best program found in any generation of a run as the result. A high level algorithm of GGP is presented in Table 3. 1. Generate an initial population of programs. 2. Execute each program in the current population and assign it a fitness value according to the fitness function 3. If the termination criterion is satisfied, terminate the algorithm. The best program found in the run of the algorithm is designated as the result. 4. Create a new population of programs from the current population by applying the reproduction, crossover, and mutation operations. These operations are applied to programs selected by fitness proportionate or tournament selections. 5. Rename the new population to the current population. 6. Proceed to the next generation by branching back to the step 2. Table 3: The high level algorithm of GGP 4. A generic top-down first-order learning algorithm This section presents a generic top-down first-order learning algorithm based on FOIL (Quinlan 1990, 1996). The algorithm is depicted in Table 4. The algorithm consists of three steps. In the pre-processing step, missing argument values in training examples are handled by assigning default or random values to them. A training example will be removed if it has too many missing values. If there are no or Page 15 inadequate negative examples in the training set, they can be generated. Different ways of creating negative examples have been proposed (Lavrac and Dzeroski 1994). Input: E: Training examples L: The concept description language BIASsearch: The search bias BIASlang: The language bias B: Background knowledge T: The target concept Output: A relation P which contains a set of clauses. Each clause C ∈ L. Function LEARNING(E, L, BIASsearch, BIASlang, B, T) (1) Pre-processing of the training examples E and producing a modified set of examples E': E' := Preprocessing(E). (2) Let Ecurrent := E'; Let P := {}; Repeat -Let C := T ←; -Find a specialization C' of C. This step constructs a clause C' from C by calling Clause-Construct(C, Ecurrent, B, L, BIASsearch, BIASlang); -If a specialization can be found -Add C' to P to produce a new relation P'. i.e. P' := P ∪ {C'}; -Remove all positive examples covered by P' from Ecurrent to get an updated training set E' E'current := Ecurrent - { positive examples in Ecurrent covered by P' w.r.t. the background knowledge B}; -Let Ecurrent := E'current; -Let P := P' Else -Set the flag No-More-Improvement to true; Until The Covering termination criterion is satisfied. i.e. covering-termination(P, No-More-Improvement, Ecurrent, B) returns true; (3) Post-processing the relation P and producing P'. i.e. P' := Post-processing(P); Return(P'); Table 4: A generic top-down first-order learning algorithm Page 16 The second step performs the construction of a logical relation. This step employs four local variables: Ecurrent (Current training examples set), E'current (Updated training examples set), P (Current relation) and P' (Modified relation). The main component of this step is the covering loop which implements Michalski's covering algorithm (Michalski et al. 1986a). The covering loop construct a relation by iteratively executing the following sub-steps: (a) Construct a clause that covers some positive examples in Ecurrent. (b) Append the clause to the current relation P and generate a modified relation P'. (c) Remove all positive examples from Ecurrent which are covered by P' with respect to the background knowledge B. The covering loop terminates if the terminating conditions are satisfied. A typical condition is that either all positive examples are covered or no more improvement can be achieved by searching for a new clause. The final step attempts to improve the accuracy of the relation induced when classifying unseen examples and to simplify the relation. The covering loop calls the 'Clause-Construct' function which is the core of the generic algorithm. A hill-climbing 'Clause-Construct' algorithm is presented in Table 5. The function constructs a clause Cn = T ← l1, l2, ..., ln starting from the most general clause C0 = T ← with an empty body. A sequence of clauses C0, C1, C2, C3, ...., Cn are generated by a number of specialization steps. At each step, the current clause Ci = T← l1,l2, ..., li is refined by appending a specific literal lj to its body. A literal lj is constructed from the background knowledge B restricted by the concept description language L and language bias BIASlang. The language may limit lj to be function-free while BIASlang may prevent new variable to be introduced in lj. The aim of the procedure is to find a clause which covers most positive examples while excludes all or most negative examples. In a hill-climbing search, the procedure keeps the current Page 17 best clause and refines it using the estimated best specialization at each step, until the stopping condition is satisfied The 'Clause-Construct' function calls the 'Find-Extension' function to find the extension Ei of the current training examples given the partially developed clause Ci = T(X1, X2, ..., Xn) ← l1, l2, ..., li and the background knowledge B. Each training example <x1, x2, ...,xn> is a n-tuple where xi, 1≤i≤n, are some constants. To find the extension, the function initializes a clause C0 = T(X1, X2, ..., Xn), then the literal l1 is added to the body of C0 to produce a new clause C1. The literal l1 is either of the form Xj = Xk, Xj ≠ Xk, pm(Y1, Y2, ...,Ysm ) or not pm(Y1, Y2, ...,Ysm ). If the literal contains k new variables, the arity of each tuple in the generated training set E1 increases to (n + k). E1 can be found by performing a natural join of Ecurrent with the relation corresponding to literal l1. The process is repeated for literals l2, l3, ..., li until the extension Ei is found. The most important component of the hill-climbing 'Clause-Construct' algorithm is the 'scoring' function that estimates the performance of each literal. An accurate estimation directs the search towards the global maxima while a misleading one traps the system into local-maxima. By providing different 'scoring' functions to the generic learning algorithm, various learning algorithms can be generated. The performances of a good and a bad learners can be significant different as shown in Sub-section 5.3. Page 18 Input: C: An initial clause C = T← Ecurrent: The current training examples B: Background knowledge L: The concept description language BIASsearch: The search bias BIASlang: The language bias Output: A clause that covers some positive examples in Ecurrent while excludes all or most negatives examples in Ecurrent Function Clause-Construct(C, Ecurrent, B, L, BIASsearch, BIASlang) There is a scoring function stored in BIASsearch, save this function to scoring; Repeat -Set BEST to a bad literal such as X = X where X is a variable appearing in the head of the clause; -Set Best-score to 0; -Find the extension Ei of Ecurrent using the clause C w.r.t. B. i.e. Ei := Find-Extension(C, Ecurrent, B); -Let ni + be the number of positive tuples in Ei; -Let ni − be the number of negative tuples in Ei; -Current-information := − log 2( ni + / (ni + + ni − )); -For all literal l from B that satisfy the constraints imposed by the language L and bias BIASlang -Set C ' = C ∪ {l}; -Find the extension Ei+1 of Ecurrent using the clause C ' i.e. Ei+1 := Find-Extension(C ', Ecurrent, B); -Let ni+1 + be the number of positive tuples in Ei+1; -Let ni+1 − be the number of negative tuples in Ei+1; -Let the number of positive tuples in Ei that have been represented by one or more tuples in Ei+1 be ni ++ ; -Find the score of the literal l by using the scoring function i.e. literal-score := scoring (ni ++ , ni+1 + , ni+1 − , Current-information); -If literal-score > Best-score then -BEST := l; -Best-score := literal-score; -If BEST == X=X then -No-More-Improvement := true; Else -Append BEST to the body of C; Until Clause-Termination(C, No-More-Improvement, Ecurrent, B) is true; Post-processing the clause C to find an improvement i.e. C' := Find-Improvement(C); If Acceptable(C') -Return(C'); Else -Return(No-Specialization-Can-Be-Found); Table 5: A hill-climbing 'Clause-Construct' algorithm Page 19 5. Inducing procedural search biases In this section, GGP is used in the meta-level learner to induce procedural search biases (i.e. the 'scoring' function). In order to employ GGP, a logic grammar must be defined (Table 6). In the grammar, the terminal symbols n-pos-i-plus-1, n-neg-i-plus-1, and n-pos-i represent respectively ni +1 + , ni +1 − and ni + + . With reference to the algorithms in Tables 4 and 5, assume that Ei is the extension of current training examples Ecurrent by current clause Ci, ni + and ni − are respectively the number of positive and negative tuples in Ei. Ei can be extended by using the literal l to Ei+1. ni+1 + and ni+1 − are respectively the number of positive and negative tuples in Ei+1. ni ++ is the number of positive tuples in Ei that have been represented by one or more tuples in Ei+1. The terminal symbol current-information is defined as − log 2( ni + / (ni + + ni − )) . start -> function. s-exp -> term. s-exp -> function. function -> [(], op1, s-exp, [)]. function -> [(], op2, s-exp, s-exp, [)]. op1 -> [ protected-log ]. op2 -> [ + ]. op2 -> [ - ]. op2 -> [ * ]. op2 -> [ % ]. op2 -> [ info ]. term -> [ n-pos-i-plus-1 ]. term -> [ n-neg-i-plus-1 ]. term -> [ n-pos-i ]. term -> [ current-information ]. term -> { random(-10, 10, ?a) }, [ ?a ]. Table 6: A logic grammar for learning procedural search bias The terminal symbols +, -, and * represent functions that perform ordinary addition, subtraction, and multiplication respectively. The symbol % represents Page 20 function that normally returns the quotient. However, if division by zero is attempted, the function returns 1.0. The symbol protected-log is a function that calculates the logarithm of the input argument x if x is larger than zero, otherwise it returns 1.0. The symbol info represents the basic function that calculates − log 2( X / ( X + Y )) given X and Y as inputs. The logic goal random(-10, 10, ?a) generates a random floating point number between -10 and 10 and instantiates ?a to the random number generated 5.1. The evolution process The evolution process of the adaptive knowledge acquisition system is depicted in Figure 3. Firstly, the Biases base is initialized with a population of different 'scoring' functions generated randomly using the logic grammar depicted in Table 6. To estimate the fitness of a specific 'scoring' function, it is combined with the generic top-down learner to produce a specific learner. The performance of this specific learner is then evaluated by using a fitness function. This measure is assigned as the fitness of the specific 'scoring' function. GGP employs selection, crossover, and mutation to generate potentially better functions. The modified functions are stored in the Biases base and the whole evolution process iterates until the best function is found or no computational resource is available Page 21 Biases Base GGP Empirical ILP Learner BKbase Example database Modified Biases Initial Biases Bias Performance of bias Figure 3: The evolution process of the adaptive knowledge acquisition system 5.2. The experimentation setup In this paper, learning curves are used to estimate the performances of various learning systems. The example space is divided randomly into disjoint training and testing sets. The learner is trained on progressively larger portions of the training set and the performance of the induced logical relation is estimated on the disjoint testing set. This process of dividing, training, and testing is repeated for 20 trials and the results are averaged to generate a learning curve. As a running example, we use a traditional problem discussed in the literature (Muggleton and Feng 1990). In the problem of learning the list predicate member, the data consist of all lists of lengths 0 to 3 defined over three constants. The background knowledge B contains definitions of list construction predicates: null which holds for an empty list and component which decomposes a list into its head and tail. The example space contains 75 positive and 45 negative examples. The training sets contain 20 to 52 examples, one-half of each training set is positive examples. The testing set consists of 45 positive and 15 negative examples. Page 22 5.3. Fitness calculation Adjusted and normalized fitness values are used as in Koza (1992). They are calculated from the raw fitness which is estimated by the fitness function. Various fitness functions have been tried and two of them are described here. The impact of fitness function on the generality of the evolved function is also demonstrated. The problem domain of learning the member predicate is used here. For the first fitness function, a random set of 24 positive and 21 negative examples is used. A specific 'scoring' function is combined with the generic top-down learner to produce a specific learner called Adapted-ILP hereafter. Adapted-ILP induces first-order logical relations using the random example set. The quality of the induced logical relations is evaluated by counting the total number of misclassified examples from the same training set. This measure is used as the raw fitness of the specific 'scoring' function. Using this fitness function, only poor 'scoring' functions have been evolved. The learning curve of a poor learner is depicted in Figure 4. For the second fitness function, the raw fitness is developed in several steps. At the beginning of each generation, four instances of the learning task are created randomly from the member domain. Each learning task has a training and a disjoint testing data. The training set contains 20 positive and 20 negative examples. For each learning task, a specific Adapted-ILP induces logical relations from the training set and the relations are evaluated by counting the number of misclassified examples from the testing set. The performance of the Adapted-ILP is the sum of numbers of misclassified examples for all learning tasks. This measure is then used as the raw fitness of the corresponding 'scoring' function. This fitness function can force the evolution of good 'scoring' functions. The learning curve of a good learner is shown in Figure 4. Page 23 20 24 28 32 36 40 44 48 52 Training size 0.7 0.75 0.8 0.85 0.9 0.95 1 A cc ur ac y Poor ILP learner Good ILP learner Figure 4: The learning curves of good and poor learner 6. Experimentation and evaluations This section compares the performance of our system with that of FOIL (Quinlan 1990; 1996). Standard learning tasks in the literature are used in these experiments (Quinlan 1990, Muggleton and Feng 1990). 6.1. The member predicate The learning curves for this problem are depicted in Figure 5. It is interested to find that our system has higher accuracy than FOIL. The difference is significant at 5% level of significance when the training size is less than 36. 20 24 28 32 36 40 44 48 52 Training size 0.75 0.8 0.85 0.9 0.95 1 A cc ur ac y Foil The Adaptive ILP system Figure 5: Learning curves for the member problem Page 24 6.2. The member predicate in a noisy environment Difference amount of noise is introduced into the training examples in order to study the performances of both systems in learning relations in noisy environment. To introduce n% of noise into the examples, n% positive examples are labeled as negative ones while n% negative examples are labeled as positive ones. In this experiment, the percentages of introduced noise are 10% (0.1) and 40% (0.4). Their learning curves are summarized in Figure 6. Our system performs better than FOIL at all noise level. 20 24 28 32 36 40 44 48 52 Training size 0.5 0.6 0.7 0.8 0.9 1 A cc ur ac y Foil (0.1) The Adaptive ILP system (0.1) Foil (0.4) The Adaptive ILP system (0.4) Figure 6: Learning curves for the member problem in a noisy environment 6.3. The multiply predicate In the problem of learning the arithmetic predicate multiply (Muggletion and Feng 1990), the data contain integers in the range from zero to ten. The background knowledge is composed of definitions for arithmetic predicates plus, decrement, zero, and one. The example space has 73 positive and 1258 negative examples respectively. The training sets consist of 400 to 500 examples, one-tenth of each training set is positive and the remainder is negative. The learning curves for multiply are presented in Figure 7. Our system performs better than FOIL when the size of training set is less than 460. The difference is significant at 5% level of significance. Page 25 400 420 440 460 480 500 Training size 0.6 0.7 0.8 0.9 1 A cc ur ac y Foil The Adaptive ILP system Figure 7: Learning curves for the multiply problem 6.4. The uncle predicate Another traditional testbed for relational learners is the domain of family relationships (Quinlan 1990). In this experiment, the uncle predicate is induced and the background predicates are parent, sibling, married, male, and female. The learning curves are presented in Figure 8. 50 70 90 110 130 150 Training size 0.75 0.8 0.85 0.9 0.95 1 1.05 A cc ur ac y Foil The Adaptive ILP system Figure 8: Learning curves for the uncle problem Page 26 7. Conclusion In this paper, we formulate an adaptive knowledge acquisition system which is composed of an external interface, a biases base, a knowledge base of background knowledge, an example database, an empirical ILP learner, a meta-level learner, and a learning controller. An implementation of the adaptive knowledge acquisition system has been developed. In the implementation, the empirical ILP learner performs top- down search in the hypothesis space defined by the concept description language, the language bias, and the background knowledge. The search is directed by search biases which can be induced and refined by a meta-level learner implemented by using generic genetic programming. Generic Genetic Programming (GGP) is a novel, general, and powerful approach of evolving search biases. It is also powerful enough to use logic grammars to represent context-sensitive information and domain-dependent knowledge. The idea of using formal grammars to direct search for knowledge in the hypothesis space or to reduce the size of the space has also been independently studied by other researcher recently (Cohen 1992, Gruau 1996, Whigham 1995; 1996). It has been demonstrated that the induced bias is better than that of FOIL on many standard learning tasks. From these experiments, it can be concluded that the adaptive knowledge acquisition system has superior learning ability compared to FOIL. Since they are different in their search biases only, the result implies that the search bias induced by GGP is better than that of FOIL for the learning problems. This result is surprising because the search biases of the adaptive knowledge acquisition system are initialized by a random process. These biases are normally poor, but the process of natural selection and evolution can successfully evolve a good bias. It is important to mention that the search bias is rather general because it has reasonable performance on many traditional learning problems using the same bias acquired automatically. This paper illustrates that GGP is a plausible approach for implementing a meta-level learning system. For future work, in order to find a Page 27 general, efficient, and effective bias, a large number of learning tasks of different kinds, such as the member, append, quick sort, ackermann, uncle, and grandfather problems, with various characteristics should be used. This adaptive learning approach, though computationally intensive, is rather exciting, as it opens up many opportunities for creating or improving learning algorithms. Page 28 Reference Abramson, H. and Dahl, V. (1989). Logic Grammars. Berlin: Springer-Verlag. Bremen, L., Friedman, J. H., Olshen, R. A. and Stone, C. J. (1984). Classification and Regression Trees. Belmont: Wadsworth. Cestnik, B., Kononenko, J. and Bratko, I. (1987). ASSISTANT 86: A knowledge elicitation tool for sophisticated users. In I. Bratko and N. Lavrac (Ed.), Progress in Machine Learning, 31-45. Wilmslow: Sigma Press. Clark, P. and Niblett, T. (1989). The CN2 induction algorithm. Machine Learning; 3, 261-283. Cohen, W. (1992). Compiling Prior Knowledge into an Explicit Bias. In Proceedings of the Ninth International Workshop on Machine Learning, 102-110. CA: Morgan Kaufmann. Colmerauer, A. (1978). Metamorphosis Grammars. In L. Bolc (Ed.), Natural Language Communication with Computers. Berlin: Springer-Verlag. Dzeroski, S. (1996). Inductive Logic Programming and Knowledge Discovery in Databases. In U. M. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy (eds.), Advances in Knowledge Discovery in Data Mining, 117-152. Menlo Park, CA: AAAI Press. Dzeroski, S. and Lavrac, N. (1993). Inductive Learning in Deductive Databases. IEEE Transactions on Knowledge and Data Engineering, 5, 939-949. Gruau, F. (1996). On Using Syntactic Constraints with Genetic Programming. In P. J. Angeline and K. E. Kinnear, Jr. (Eds.) Advances in Genetic Programming 2, 402-417. MA: MIT Press. Hopcroft, J. E. and Ullman, J. D. (1979). Introduction to automata theory, languages, and computation. MA: Addison-Wesley. Koza, J. R. (1994). Genetic Programming II: Automatic Discovery of Reusable Programs. Cambridge, MA: MIT Press. Koza, J. R. (1992). Genetic Programming: on the Programming of Computers by Means of Natural Selection. Cambridge, MA: MIT Press. Kinnear, K. E. Jr., editor (1994). Advances in Genetic Programming. Cambridge, MA: MIT Press. Kowalski, R. A. (1979). Logic For Problem Solving. Amsterdam: North-Holland. Page 29 Lavrac, N. and Dzeroski, S. (1994). Inductive Logic Programming: Techniques and Applications. London: Ellis Horword. Lewis, H. R. and Rapadimitrion, C. H. (1981). Elements of the theory of computation. NJ: Prentice Hall. Michalski, R. S., Mozetic, I., Hong, J. and Lavrac, N. (1986). The multi-purpose incremental learning system AQ15 and its testing application on tree medical domains. In Proceedings of the National Conference on Artificial Intelligence, 1041- 1045. San Mateo, CA: Morgan Kaufmann. Mitchell, T. M. (1982). Generalization as Search, Artificial Intelligence, 18, 203-226. Muggletion, S., editor (1992). Inductive Logic Programming. London: Academic Press. Muggleton, S. and Buntine, W. (1988). Machine invention of first-order predicates by inverting resolution. In Proceedings of the Fifth International Conference on Machine Learning, 339-352. CA: Morgan Kaufmann Muggleton, S. and Feng, C. (1990), Efficient induction of logic programs, In Proceedings of the First Conference on Algorithmic Learning Theory, 1-14. Pereira, F. C. N. and Warren, D. H. D. (1980). Definite Clause Grammars for Language Analysis - A Survey of the Formalism and a Comparison with Augmented Transition Networks Artificial Intelligence, 13, 231-278. Pereira, F. C. N. and Shieber, S. M. (1987). Prolog and Natural-Language Analysis. CA: CSLI. Quinlan, J. R. (1990). Learning logical definitions from relations. Machine Learning, 5, 239-266. Sterling, L. and Shapiro, E. (1986). The Art of Prolog. Cambridge, MA: MIT Press. Whigham, P. A. (1996). Search Bias, Language Bias and Genetic Programming. In Proceedings of the First Genetic Programming Conference, 230-237. Cambridge, MA: MIT Press. Whigham, P. A. (1995). Inductive Bias and Genetic Programming. In Proceedings of the First International Conference on Genetic Algorithms in Engineering Systems: Innovations and Applications, 461-466. UK: IEE. 
work_kdo3jxr3xvh7rbksrlfmoi3qxe	----	Re-sampling Strategies for Regression Lúıs Torgo1,2, Paula Branco1,2, Rita P. Ribeiro1,2, and Bernhard Pfahringer3 1LIAAD - INESC TEC 2 DCC - Faculdade de Ciências - Universidade do Porto 3 Department of Computer Science - University of Waikato ltorgo@dcc.fc.up.pt, paobranco@gmail.com, rpribeiro@dcc.fc.up.pt, bernhard@cs.waikato.ac.nz February 22, 2016 Abstract Several real world prediction problems involve forecasting rare val- ues of a target variable. When this variable is nominal we have a prob- lem of class imbalance which was thoroughly studied within machine learning. For regression tasks, where the target variable is continu- ous, few works exist addressing this type of problem. Still, important applications involve forecasting rare extreme values of a continuous target variable. This paper describes a contribution to this type of tasks. Namely, we propose to address such tasks by re-sampling ap- proaches that change the distribution of the given data set to decrease the problem of imbalance between the rare target cases and the most frequent ones. We present two modifications of well-known re-sampling strategies for classification tasks: the under-sampling and the Smote methods. These modifications allow the use of these strategies on re- gression tasks where the goal is to forecast rare extreme values of the target variable. In an extensive set of experiments we provide empir- ical evidence for the superiority of our proposals for these particular regression tasks. The proposed re-sampling methods can be used with any existing regression algorithm, which means that they are general tools for addressing problems of forecasting rare extreme values of a continuous target variable. 1 Introduction Forecasting rare extreme values of a continuous variable is very relevant for several real world domains (e.g. finance, ecology, meteorology, etc.). This problem can be seen as equivalent to classification problems with imbal- anced class distributions which have been studied for a long time within 1 machine learning (e.g. Domingos (1999); Elkan (2001); Zadrozny (2005); Chawla (2005)). The main difference is the fact that we have a target nu- meric variable, i.e. a regression task. This type of problem is particularly difficult because: i) there are few examples with the rare target values; ii) the errors of the learned models are not equally relevant because the user’s main goal is predictive accuracy on the rare values; and iii) standard pre- diction error metrics are not adequate to measure the quality of the models given the preference bias of the user. The existing approaches for the classification scenario can be cast into 3 main groups (Zadrozny, 2003; Ling and Sheng, 2010): i) change the evalu- ation metrics to better capture the application bias; ii) change the learning systems to bias their optimization process to the goals of these domains; and iii) re-sampling approaches that manipulate the training data distribution so as to allow the use of standard learning systems. All these three approaches were extensively explored within the classification scenario (e.g. Kubat and Matwin (1997); Chawla et al. (2002)). Research work within the regression setting is much more limited. Torgo and Ribeiro (2009) and Ribeiro (2011) proposed a set of specific metrics for regression tasks with non-uniform costs and benefits. Ribeiro (2011) described system ubaRules that was specif- ically designed to address this type of problem. Still, to the best of our knowledge, no one has tried re-sampling approaches on this type of regres- sion tasks. Nevertheless, re-sampling strategies have a clear advantage over the other alternatives - they allow the use of any existing regression tool on this type of tasks without the need to change it. The main goal of this paper is to explore this alternative within a regression context. We describe two possible methods: i) using an under-sampling strategy; and ii) using a Smote-like approach that performs both over- and under-sampling. The main contributions of this work are: i) presenting a first attempt at addressing rare extreme values prediction using standard regression tools through re-sampling approaches; and ii) adapting two well-known and suc- cessful re-sampling methods for regression tasks. The results of the empirical evaluation of our contributions provide clear evidence on the validity of these approaches for the task of predicting rare extreme values of a numeric tar- get variable. The significance of our contributions results from the fact that they allow the use of any existing regression tool on these important tasks by simply manipulating the available data set using our supplied code. All code and data used in this paper is provided in an associated web page1 to ensure easy reproducibility and re-use of our results and algorithms. 1 http://www.dcc.fc.up.pt/~ltorgo/ExpertSystems 2 2 Problem Formulation Predicting rare extreme values of a continuous variable is a particular class of regression problems. In this context, given a training sample of the prob- lem, D = {〈x,y〉}Ni=1, our goal is to obtain a model that approximates the unknown regression function y = f(x). The particularity of our target tasks is that the goal is the predictive accuracy on a particular subset of the do- main of the target variable Y - the rare and extreme values. As mentioned before, this is similar to classification problems with extremely unbalanced classes. As in these problems, the user goal is the performance of the models on a sub-range of the target variable values that is very infrequent. In this context, standard regression metrics (e.g. mean squared error) suffer from the same problems as error rate (or accuracy) on imbalanced classification tasks - they do not focus on the rare cases performance. In classification the solution usually revolves around the use of the precision/recall evaluation framework (Davis and Goadrich, 2006). Precision provides an indication on how accurate are the predictions of rare cases made by the model. Recall tells us how frequently the rare situations were signalled as such by the model. Both are important properties that frequently require some form of trade-off. How can we get similar evaluation for the numeric prediction of rare extreme values? On one hand we want that when our models predict an extreme value they are accurate (high precision), on the other hand we want our models to make extreme value predictions for the cases where the true value is an extreme (high recall). Assuming the user gives us information on what is considered an extreme for the domain at hand (e.g. Y < k1 is an extreme low, and Y > k2 is an extreme high), we could transform this into a classification problem and calculate the precision and recall of our models for each type of extreme. However, this would ignore the notion of numeric precision. Two predicted values very distant from each other, as long as being both extremes (above or below the given thresholds) would count as equally valuable predictions. This is clearly counter-intuitive on regression problems such as our tasks. A solution to this problem was described by Torgo and Ribeiro (2009) and Ribeiro (2011) that have presented a formu- lation of precision and recall for regression tasks that also considers the issue of numeric accuracy. We will use this framework to compare and evaluate our proposals for this type of tasks. For completeness, we will now briefly describe the framework proposed by Ribeiro (2011) that will be used in the experimental evaluation of our proposal2. 2The code implementing this evaluation framework that was used in our experiments is available at http://www.dcc.fc.up.pt/~rpribeiro/uba/. 3 2.1 Utility-based Regression The precision/recall evaluation framework we will use is based on the con- cept of utility-based regression (Ribeiro, 2011; Torgo and Ribeiro, 2007). At the core of utility-based regression is the notion of relevance of the target variable values and the assumption that this relevance is not uniform across the domain of this variable. This notion is motivated by the fact that con- trary to standard regression, in some domains not all the values are equally important/relevant. In utility-based regression the usefulness of a predic- tion is a function of both the numeric error of the prediction (given by some loss function L(ŷ,y)) and the relevance (importance) of both the predicted ŷ and true y values. Relevance is the crucial property that expresses the domain-specific biases concerning the different importance of the values. It is defined as a continuous function φ(Y ) : Y → [0, 1] that maps the target variable domain Y into a [0, 1] scale of relevance, where 0 represents the minimum and 1 represents the maximum relevance. To better illustrate the concept of the relevance function, let us consider a regression problem where the goal is to predict the total percentage of burnt area on forest fires. Figure 1 shows a possible relevance function for this problem. It was created with the goal of stressing the domain-specific knowledge that large scale forest fires are the most important events we should aim at forecasting3. In effect, the accurate prediction of such type of fires has high benefits as it allows prevention and planning actions to be carried out and also avoids the large costs of not being prepared for these events. 0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 Forest Fires Relevance Function φ (Y ) Y 0% 46% Figure 1: Relevance function φ for the percentage of burnt forest area. According to Ribeiro (2011), the only assumption regarding the shape 3The function φ was obtained by applying piecewise cubic Hermite interpolation to the box-plot (Cleveland, 1993) statistics of the target variable as described in Ribeiro (2011). Still, any other method could have been used to define the relevance function. 4 of the function φ() is that there are some ranges of its domain where it as- sumes a quasi-concave shape, designated as bumps of relevance (see Ribeiro (2011) for more details). These bumps correspond to the ranges of the target variable values the user deems as more important. Being a domain-specific function, it is the user responsibility to specify the relevance function. How- ever, Ribeiro (2011) describes some specific methods of obtaining auto- matically these functions when the goal is to be accurate at rare extreme values, which is the case of our applications. The methods are based on the simple observation that for these applications the notion of relevance is inversely proportional to the target variable probability density function. In our experiments we have used this strategy to come up with the relevance functions for the data sets we have used. The obtained relevance functions have essentially one4 or two bumps of relevance associated with the high and low extreme values, and then a large region in the middle with near zero relevance, corresponding to the non-extreme and frequent values of the target variable. The utility of a model prediction is related to the question on whether it has led to the identification of the correct type of extreme and if the predic- tion was precise enough in numeric terms. Thus to calculate the utility of a prediction it is necessary to consider two aspects: (i) does it identify the correct type of extreme? (ii) what is the numeric accuracy of the prediction (i.e. L(ŷ,y))? This latter issue is important because it allows for coping with different ”degrees” of actions as a result of the model predictions. For instance, in the context of financial trading an agent may use a decision rule that implies buying an asset if the predicted return is above a certain threshold. However, this same agent may invest different amounts depend- ing on the predicted return, and thus the need for precise numeric forecasts of the returns on top of the correct identification of the type of extreme. This numeric precision, together with the fact that we may have more than one type of extreme (i.e. more than one ”positive” class) are the key dis- tinguishing features of this framework when compared to pure classification approaches, and are also the main reasons why it does not make sense to map our problems to classification tasks. The concrete utility score of a prediction, in accordance with the original framework of utility-based learning (e.g. Elkan (2001); Zadrozny (2005)), results from the net balance between its benefits and costs (i.e. negative benefits). A prediction should be considered beneficial only if it leads to the identification of the correct type of extreme. However, the reward should also increase with the numeric accuracy of the prediction and should be dependent on the relevance of the true value. In this context, Ribeiro (2011) has defined the benefit of a prediction as a proportion of its maximum 4Some data sets only have one bump because they either only have high or low rare extreme values. 5 benefit that is given by the relevance of the true value, i.e. φ(y), Bφ(ŷ,y) = φ(y) · (1 − ΓB(ŷ,y)) (1) where ΓB is a bounded loss function for benefits. The bounded loss ΓB is a [0, 1] function that maps the unbounded range of a standard regression loss function (e.g. squared error) into a propor- tion5. ΓB is designed to give the value of 0 if we have a perfect prediction thus leading to the maximum benefits in the current situation, i.e. φ(y). The function smoothly increases up to 1 as we move away from a perfect prediction. ΓB definition ensures that the benefits will be zero if either the prediction is too inaccurate or the predicted value is not a similar type of extreme as the true value. Regarding costs, the rationale is that a prediction should entail a cost if it is too inaccurate and/or is unable to correctly identify the type of true value6. To quantify the cost of these erroneous predictions, it is necessary to determine how relevant the impact of such mistakes is. While the benefits of a prediction depend on the usefulness of its associated true value (i.e. φ(y)), costs depend on both the relevance of the true and predicted values, because they are of different type. This means that costs are proportional to the relevance of both the true and predicted values. The joint relevance function captures this notion by calculating a weighted average of these two factors, φp(ŷ,y) = (1 −p) ·φ(ŷ) + p ·φ(y) (2) where p ∈ [0, 1] is a factor differentiating the types of errors. In the context of our target applications we can distinguish three different types of mistakes: (i) false alarms where we predict an extreme value for an irrelevant/normal true value ; (ii) missed events where the model predicts an irrelevant value but the true value is either an extreme high or low value; or (iii) confusing events where we predict a high(low) extreme for a true low(high) extreme value. This third scenario is the most serious type of mistake and also the one where the value of φp(ŷ,y) will be higher because both φ(ŷ) and φ(y) will be high. Ribeiro (2011) defined the cost of a prediction as a proportion of the maximum cost that is given by the joint relevance of both true and predicted values. Hence, the cost function C p φ is defined as, C p φ(ŷ,y) = φ p(ŷ,y) · ΓC(ŷ,y) (3) where φp is the joint relevance function and ΓC is the bounded loss function for costs. 5See full details in Section 3.3 of Ribeiro (2011). 6Recall that we essentially have 3 types of values - extremes (high or low) or irrelevant values. 6 Similarly to ΓB, the bounded loss ΓC is a [0,1] function that determines the proportion of maximum cost assigned to a prediction, based on the prediction error measured by a standard loss function. It should be 0 for a perfect prediction and increase up to 1, as it moves away from a perfect prediction. Given the above definitions of benefits and costs the utility associated to any prediction can be calculated as the net balance of these two factors, U p φ(ŷ,y) = Bφ(ŷ,y) − C p φ(ŷ,y) = φ(y) · (1 − ΓB(ŷ,y)) − φp(ŷ,y) · ΓC(ŷ,y) (4) where Bφ(ŷ,y), C p φ(ŷ,y), ΓB(ŷ,y) and ΓC(ŷ,y) are functions related to the notions of costs and benefits of predictions that are defined in Ribeiro (2011). 2.2 Precision and Recall for Regression Precision and recall are two of the most commonly used metrics to estimate the performance of models in highly skewed domains (Davis and Goadrich, 2006) such as our target domains. The main advantage of these statistics is that they are focused on the performance on the target events, disregarding the remaining cases. In imbalanced classification problems, the target events are cases belonging to the minority (positive) class. Informally, precision measures the proportion of events signalled by the model that are real events, while recall measures the proportion of events occurring in the domain that are captured by the model. The notions of precision and recall were adapted to regression problems with non-uniform relevance of the target values by Torgo and Ribeiro (2009) and Ribeiro (2011). In this paper we will use the framework proposed by these authors to evaluate and compare our sampling approaches. We will now briefly present the main details of this formulation7. The key issue of the precision/recall framework is the notion of interest- ing events. The definition of the cases which are to be considered interest- ing for the user is not so straightforward in regression as in classification. Ribeiro (2011) suggests the use of the relevance function to define the notion of interesting events for regression problems. Namely, a case is considered an interesting event if, z := I(φ(y) ≥ tE) (5) where I is the indicator function giving 1 if its argument is true and 0 oth- erwise, and tE ∈ [0, 1] is a domain-specific event threshold on the relevance scores. 7Full details can be obtained in Chapter 4 of Ribeiro (2011). 7 In order to calculate precision and recall it is also necessary to establish when a prediction is accurate concerning a value being or not an event. In- tuitively, one could be tempted to consider a prediction as a correct forecast of the event if φ(ŷ) ≥ tE. However, as mentioned by Ribeiro (2011), there might be cases where the relevance of the predicted value is high and nev- ertheless the prediction is incorrect. For instance, if the true target value is an extreme high value and the predicted value is an extreme low value, both will have high relevance score and still this is a serious error. The origin of the problem lies on the fact that contrary to the classification set- up, in utility-based regression there may exist more than one interesting ”class”, which correspond to more than one bump of relevance. This is a fundamental difference to the imbalanced classification framework. The solution proposed by Ribeiro (2011) is to resort to the notion of utility of a prediction to decide its correctness. From the definition of utility (Equation 4) it follows that: (i) the maximum of U p φ(ŷ,y) is φ(y); and (ii) the minimum of U p φ(ŷ,y) is −φ p(ŷ,y). In this sense, the higher the precision of the predictions the closer is U p φ to φ(y). This means that the utility score is strongly correlated with the probability that a prediction ŷ is in the same bump as y, i.e. that it is a correct forecast of some event. Ribeiro (2011) proposes a method based on the calibration of the raw utility scores to reach the final decision on whether a predicted value is or not a correct ”event” prediction, i.e. the value of ẑ. Namely, this binary property is defined as, ẑ := I ( s > 1 −u 2 ) (6) where s is a calibrated utility score, u is the raw utility score (U p φ(ŷ,y)) and I is the indicator function giving 1 if its argument is true and 0 otherwise. Precision and recall are usually defined as ratios between the correctly identified events (usually known as true positives within classification), and either the signalled events (for precision), or the true events (for recall). Given that the notion of event was defined based on the concept of utility, Ribeiro (2011) also defines the two denominators of these ratios as functions of utility, finally leading to the following definitions of precision and recall for regression, recall = ∑ i:ẑi=1,zi=1 (1 + ui) ∑ i:zi=1 (1 + φ(yi)) (7) 8 and precision = ∑ i:ẑi=1,zi=1 (1 + ui) ∑ i:ẑi=1,zi=1 (1 + φ(yi)) + ∑ i:ẑi=1,zi=0 (2 − p (1 −φ(yi))) (8) where p is a weight differentiating the types of errors, while ẑ and z are binary properties associated with being in the presence of a rare extreme case. Through the use of utility in our proposed definitions of precision and recall we are able to penalize not only false alarms (FP) and missed op- portunities (FN), but also confusing events. In an application where both type of extremes, high and low, are of interest of the end user, a confusing event should be heavily penalized. In the stock market domain, for instance, if a model advises the user to buy when in fact should sell, it is a serious error. By using utility, that same prediction would be considered as a cost (i.e. negative utility value). Moreover, the amount of penalization is more sensitive, in the sense that it takes on a continuous scale depending on the difference between the true and predicted values because of the definition of utility. In the experimental evaluation of our sampling approaches we have used as main evaluation metric the F-measure that can be calculated with the values of precision and recall, F = (β2 + 1) ·precision ·recall β2 ·precision + recall (9) where β is a parameter weighing the importance given to precision and recall (we have used β = 1, which means equal importance to both factors). 3 Re-sampling Approaches for Regression The basic motivation for re-sampling approaches is the assumption that the imbalanced distribution of the given training sample will bias the learning systems towards solutions that are not in accordance with the user’s prefer- ence goal. This occurs because the goal is predictive accuracy on the data that is least represented in the original data set. Most existing learning systems work by searching the space of possible models with the goal of optimizing some criteria. These criteria are usually related to some form of average performance. These metrics will tend to reflect the performance on the most frequent cases, which are not the goal of the user. In this context, the goal of re-sampling approaches is to change the data distribution on the training sample so as to make the learners focus on cases that are of interest 9 to the user. The change that is carried out has the goal of balancing the distribution of the least represented (but more important) cases with the more frequent observations. Many re-sampling approaches exist within the imbalanced classification literature. To the best of our knowledge no attempt has been made to apply these strategies to the equivalent regression tasks - forecasting rare extreme values. In this section we describe the adaptation of two existing re-sampling approaches to these regression tasks. 3.1 Under-sampling common values The basic idea of under-sampling (e.g. Kubat and Matwin (1997)) is to decrease the number of observations with the most common target variable values with the goal of better balancing the ratio between these observations and the ones with the interesting target values that are less frequent. Within classification this consists on obtaining a random sample from the training cases with the frequent (and less interesting) class values. This sample is then joined with the observations with the rare target class value to form the final training set that is used by the selected learning algorithm. This means that the training sample resulting from this approach will be smaller than the original (imbalanced) data set. In regression we have a continuous target variable. As mentioned in Section 2.1 the notion of relevance can be used to specify the values of a continuous target variable that are more important for the user. We can also use the relevance function values to determine which are the observations with the common and uninteresting values that should be under-sampled. Namely, we propose the strategy of under-sampling observations whose tar- get value has a relevance less than a user-defined parameter. This threshold will define the set of observations that are relevant according to the user preference bias, Dr = {〈x,y〉 ∈D : φ(y) ≥ t} (10) where t is the user-defined threshold on relevance. Under-sampling will be carried out on the remaining observations Di = D\Dr. Regards the amount of under-sampling that is to be carried out the strategy is the following. For each of the relevant observations in Dr we will randomly select nu cases from the ”normal” observations in Di. The value of nu is another user-defined parameter that will establish the desired ratio between ”normal” and relevant observations. Too large values of nu will result in a new training data set that is still too unbalanced, but too small values may result in a training set that is too small, particularly if there are too few relevant observations. 10 3.2 SMOTE for regression Smote (Chawla et al., 2002) is a sampling method to address classifica- tion problems with imbalanced class distribution. The key feature of this method is that it combines under-sampling of the frequent classes with over- sampling of the minority class. Chawla et al. (2002) show the advantages of this approach when compared to other alternative sampling techniques on several real world problems using several classification algorithms. The key contribution of our work is to propose a variant of Smote for addressing regression tasks where the key goal is to accurately predict rare extreme values, which we will name SmoteR . The original Smote algorithm uses an over-sampling strategy that con- sists on generating ”synthetic” cases with a rare target value. Chawla et al. (2002) propose an interpolation strategy to create these artificial examples. For each case from the set of observations with rare values (Dr), the strat- egy is to randomly select one of its k-nearest neighbours from this same set. With these two observations a new example is created whose attribute values are an interpolation of the values of the two original cases. Regards the target variable, as Smote is applied to classification problems with a single class of interest, all cases in Dr belong to this class and the same will happen to the new synthetic cases. There are three key components of the Smote algorithm that we need to address in order to adapt it for our target regression tasks: i) how to define which are the relevant observations and the ”normal” cases; ii) how to create new synthetic examples (i.e. over-sampling); and iii) how to decide the target variable value of these new synthetic examples. Regarding the first issue, the original algorithm is based on the information provided by the user concerning which class value is the target/rare class (usually known as the minority or positive class). In our problems we face a potentially infinite number of values of the target variable. As we have mentioned in Section 2.1, our proposal is based on the existence of a relevance function and on a user-specified threshold on the values of this function, that leads to the definition of the set Dr (c.f. Equation 10). Our algorithm will over- sample the observations in Dr and under-sample the remaining cases (Di), thus leading to a new training set with a more balanced distribution of the values. Regarding the second key component, the generation of new cases, we use the same approach as in the original algorithm though we have introduced some small modifications for being able to handle both numeric and nominal attributes. Finally, the third key issue is to decide the target variable value of the generated observations. In the original algorithm this is a trivial question, because as all rare cases have the same class (the target minority class), the same will happen to the examples generated from this set. In our case the answer is not so trivial. The cases that are to be over- sampled do not have the same target variable value, although they do have a 11 high relevance score (φ(y)). This means that when a pair of examples is used to generate a new synthetic case, they will not have the same target variable value. Our proposal is to use a weighed average of the target variable values of the two seed examples. The weights are calculated as an inverse function of the distance of the generated case to each of the two seed examples. Algorithm 1 The main SmoteR algorithm. function SmoteR(D, tE,o,u,k) // D - A data set // tE - The threshold for relevance of the target variable values // %o,%u - Percentages of over- and under-sampling // k - The number of neighbours used in case generation rareL ←{〈x,y〉 ∈D : φ(y) > tE ∧y < ỹ} // ỹ is the median of the target Y newCasesL ← genSynthCases(rareL, %o,k) // generate cases for rareL rareH ←{〈x,y〉 ∈D : φ(y) > tE ∧y > ỹ} newCasesH ← genSynthCases(rareH, %o,k) // generate cases for rareH newRareCases ← rareL ⋃ newCasesL ⋃ rareH ⋃ newCasesH tgtNorm ←%u of |newRareCases| newNormCases ←sample of tgtNorm cases ∈D\{rareL ⋃ rareH} return newRareCases ⋃ newNormCases end function Algorithm 1 describes our proposed SmoteR sampling method. The algorithm uses a user-defined threshold (tE) of relevance to define the sets Dr and Di. Notice that in our target applications we may have two rather different sets of rare cases: the extreme high and low values. This is another difference to the original algorithm. The consequence of this is that the gen- eration of the synthetic examples is also done separately for these two sets. The reason is that although both sets include rare and interesting cases, they are of different type and thus with very different target variable values (extremely high and low values). The other parameters of the algorithm are the percentages of over- and under-sampling, and the number of neighbours to use in the cases generation. The key aspect of this algorithm is the gen- eration of the synthetic cases (the two calls to function genSynthCases). This process is described in detail on Algorithm 2. The main differences to the original Smote algorithm are: the ability to handle both numeric and nominal variables; and the way the target value for the new cases is generated. Regards the former issue we simply perform a random selection between the values of the two seed cases. A possible alternative could be to use some biased sampling that considers the frequency of occurrence of each of the values within the rare cases. Regarding the target value we have used a weighted average between the values of the two seed cases. The weights are decided based on the distance between the new case and these two seed cases. The larger the distance the smaller the weight. 12 Algorithm 2 Generating synthetic cases. function genSynthCases(D,o,k) newCases ←{} ng ←%o/100 // nr. of new cases to generate for each existing case for all case ∈D do nns ← kNN(k,case,D\{case}) // k-Nearest Neighbours of case for i ← 1 to ng do x ← randomly choose one of the nns for all a ∈ attributes do // Generate attribute values if isNumeric(a) then diff ← case[a] −x[a] new[a] ← case[a] + random(0, 1) ×diff else new[a] ← randomly select among case[a] and x[a] end if end for d1 ← dist(new,case) // Decide the target value d2 ← dist(new,x) new[Target] ← d2×case[T arget]+d1×x[T arget] d1+d2 newCases ← newCases ⋃ {new} end for end for return newCases end function 4 Experimental Evaluation The goal of our experiments is to test the effectiveness of our proposed sampling approaches at predicting rare extreme values of a continuous target variable. For this purpose we have selected 17 regression data sets whose main characteristics are described in Table 1. For each of these data sets we have obtained a relevance function using the automatic method proposed by Ribeiro (2011). This method assigns higher relevance for values above (below) the adjacent values of the target variable sample distribution. These are calculated as a function of the quartiles and the inter-quartile range and are well-known thresholds for considering a value as an outlier. The result of applying this method are relevance functions that assign higher relevance to high and low rare extreme values, which are the target of the work in this paper. Based on these relevance functions we have imposed a threshold of 0.7 on the values of φ(Y ) as the condition for a value to be taken as a rare extreme. As it can be seen from the information in Table 1 this leads to an average of 10-15% of the available cases having a rare extreme value for most data sets considered in our experiments. In order to avoid any algorithm-dependent bias distorting our exper- imental results, we have carried out our comparisons using a diverse set of standard regression algorithms. Moreover, for each algorithm we have 13 Data Set N p nRare %Rare Data Set N p nRare %Rare a1 198 12 32 0.162 dAiler 7129 6 617 0.087 a2 198 12 26 0.131 availPwr 1802 16 197 0.109 a3 198 12 34 0.172 bank8FM 4499 9 384 0.085 a4 198 12 37 0.187 cpuSm 8192 13 804 0.098 a5 198 12 22 0.111 dElev 9517 7 1109 0.117 a6 198 12 35 0.177 fuelCons 1764 38 225 0.128 a7 198 12 29 0.146 boston 506 14 74 0.146 Abalone 4177 9 882 0.211 maxTorque 1802 33 163 0.09 Accel 1732 15 106 0.061 Table 1: Used data sets and characteristics (N: n. of cases; p: nr. of predictors; nRare: nr. cases with φ(Y ) > 0.70; %Rare: nRare/N). Learner Parameter Variants R package MARS nk = {10, 17},degree = {1, 2}, thresh = {0.01, 0.001} earth (Milborrow, 2012) SVM cost = {10, 150, 300},gamma = {0.01, 0.001} e1071 (Dimitriadou et al., 2011) Random Forest mtry = {5, 7},ntree = {500, 750, 1500} randomForest (Liaw and Wiener, 2002) Table 2: Regression algorithms and parameter variants, and the respective R packages. considered several parameter variants. Table 2 summarizes the learning al- gorithms that were used and also the respective parameter variants. For instance, for the MARS regression algorithm we have considered the 8 vari- ants resulting from all combinations of parameter values reported in Table 2. To ensure easy replication of our work we have used the implementations of these algorithms available in the free and open source R environment (R Core Team, 2013), which is also the infrastructure used to implement our proposed sampling methods. Each of the 20 learning approaches (8 MARS variants + 6 SVM vari- ants + 6 Random Forest variants), were applied to each of the 17 regres- sion problems using 13 different sampling approaches, namely: i) carry- ing out no sampling at all (i.e. use the data set with the original imbal- ance); ii) 8 variants of our SmoteR method; and iii) 4 variants of under- sampling. The 8 SmoteR variants used 5 nearest neighbours for case gen- eration, a relevance threshold of 0.70 and all combinations of {200, 500}% and {50, 100, 200, 300}% for percentages of over- and under-sampling, re- spectively (c.f. Algorithm 1). For instance, in the data set availPwr, which contains 197 rare cases (c.f. Table 1), when using 200% oversampling and 50% under-sampling, we would generate 2 new cases for each existing rare case thus leading to 591 (= 197 + 197 × 2) rare cases and 295 normal cases (50% of 591), thus a training set of 886 cases, whilst the original data set contains 1802 cases. With respect to under-sampling the 4 variants used {50, 100, 200, 300}% as percentages of under-sampling and the same 0.70 14 0 5 10 15 20 25 30 35 0. 00 0. 05 0. 10 0. 15 Acceleration D en si ty φ ( Y ) 0. 0 0. 2 0. 4 0. 6 0. 8 1. 0 None U0.5 S.o2.u0.5 φ(Y) 100 200 300 400 0. 00 0 0. 00 5 0. 01 0 0. 01 5 0. 02 0 Available Power D en si ty φ ( Y ) 0. 0 0. 2 0. 4 0. 6 0. 8 1. 0 None U0.5 S.o2.u0.5 φ(Y) Figure 2: Distribution of the target variable before and after re-sampling. relevance threshold. This means that if applying under-sampling with 50% to the same availPwr data set, we would end up with a training set formed by 197 rare cases plus 98 (50 % of 197) normal cases. To have a better idea on the impact of these re-sampling strategies on the training set that is finally given to the learners, Figure 2 shows the distribution of the target variable8 on the original data and on the data sets resulting from applying two of the most successful variants of our re- sampling strategies, for 2 specific data sets. The graphs on this figure clearly illustrate the change in the target variable distribution that is carried out by these re-sampling strategies with the goal of biasing this distribution towards the areas where the relevance function has higher values. Moreover, as illustrated above the methods also change the number of cases used for training, which will obviously have an impact on the computation time taken to obtain the models. More specifically for the data sets in Figure 2, the original Acceleration data set contains 1732 observations and the S.o2.u0.5 configuration of SmoteR leads to a training set of 477, whilst the U0.5 under-sampling variant uses only 159 cases. With respect to the Available Power data set the original size is 1802 and the two re-sampling variants use 886 and 295, respectively. Given that the SmoteR variant uses a more complex algorithm to obtain the training set involving distance calculations to find the nearest neighbours of rare cases, and leads to a larger training set, we can immediately see that this re-sampling approach requires more computation time than the simpler under-sampling strategy. Still, under these two particular configurations the training set will always be smaller than the original data set. The goal of our experiments is to compare the 12 (8 SmoteR + 4 under- 8Approximated through a kernel density estimator. 15 Figure 3: Behaviour of the sampling strategies on 3 illustrative data sets (S - SmoteR ; U - under-sampling; ox - x×100% over-sampling; ux - x×100% under-sampling). sampling) re-sampling approaches against the default of learning with the given data, using 20 learning approaches on 17 data sets. All alternatives were evaluated according to the F-measure with β = 1 (c.f. Equation 2.2), which means that the same importance was given to both precision and recall scores that were calculated using the set-up described in Section 2.2. The values of the F-measure were estimated by means of 3 repetitions of a 10- fold cross validation process and the statistical significance of the observed paired differences was measured using the non-parametric Wilcoxon paired test. Given the amount of experimental variants it is not possible to show all results given the space restrictions of a paper. Figure 3 shows the results for 3 of our 17 data sets. The full results for all data sets can be seen in the web page associated with the paper. For each combination of data set and regression algorithm the graph provides 13 box-plots, one for each of the 12 mentioned sampling approaches plus the alternative of using the original data (tagged as none in the graphs). The box plots show the distribution of the F1 scores of all variants of each learner on each data set. All scores are estimated by means of a 3 × 10−fold cross validation process. These 16 Sampling Strat. Win (99%) Win (95%) Loss (99%) Loss (95%) Insignif. Diff. U0.5 243 31 17 5 44 U1 213 41 10 4 72 U2 184 36 7 2 111 U3 138 42 11 3 146 S.o2.u0.5 229 31 14 6 60 S.o5.u0.5 216 42 15 3 64 S.o2.u1 201 52 12 3 72 S.o5.u1 199 44 14 3 80 S.o2.u2 173 43 11 2 111 S.o5.u2 172 46 13 1 108 S.o2.u3 127 69 17 2 125 S.o5.u3 142 46 15 6 131 Table 3: Summary of the paired comparisons to the no sampling baseline. 3 particular data sets were chosen because they represent three different patterns of results. Results on data set a7 are among the best from the re-sampling approaches perspective. The boston data set can be regarded as an example of a domain where the advantage of re-sampling approaches is not so marked, while data set bank8FM is an example where re-sampling approaches behave poorly with the exception of random forest variants. When taking into consideration all 17 data sets, in most cases we have an advantage of the re-sampling approaches. This can be confirmed when looking at the overall results in terms of the statistical significance of the paired differences between each sampling approach and the alternative of using the original data (the baseline). Table 3 summarizes the results of these paired comparisons. Each sampling strategy was compared against the baseline 340 times (20 learning variants times 17 data sets). For each paired comparison we check the statistical significance of the difference between the average F score obtained with the respective sampling approach and with the baseline. These averages were estimated using a 3 × 10-fold CV process. We counted the number of statistically significant wins and losses of each of the 12 sampling variants on these 340 paired comparisons using two significance levels (99% and 95%). The results of Table 3 provide clear evidence of the advantage of using re-sampling approaches when the task is to predict rare extreme values of a continuous target variable. In effect, we can observe an overwhelming advantage in terms of number of statistically significant wins over the al- ternative of using the data set as given (i.e. no re-sampling). For instance, the particular configuration of using under-sampling with 50% (U0.5 ) was significantly better than the alternative of using the given data set on 80.5% ((243 + 31)/340) of the 340 considered situations, while only on 6.4% of the cases under-sampling actually lead to a significantly worse model. The 17 remarkable performance of this very simple re-sampling strategy is even re-enforced by the fact that it uses a much smaller training set than the other alternatives, which means lower computation costs. The SmoteR variant with 200% over-sampling and 50% under-sampling (S.o2.u0.5 ) also achieved very good results (76.5% significant wins). A pattern of results we have observed in our experiments is that higher values of the percent- age of under-sampling leads to worse results. The lower this percentage the higher the imbalance in favour of the rare cases (for instance with 50% under-sampling there will be twice as many rare cases as normal cases). This has improved performance in all our experimental set ups. Another conclu- sion from our experiments is that increasing the percentage of over-sampling (i.e. generating more synthetic cases) makes the results worse. This may be an indication that the process being used for generating new cases may be taking too many risks leading to a degradation of the accuracy of the models. In summary, the results of our experimental comparisons provide clear evidence on the validity of the re-sampling approaches we have considered, with particularly good results being obtained with training sets that bias the distribution towards a higher frequency of rare cases (i.e. do strong under-sampling of the normal cases). 5 Conclusions This paper has presented a general approach to tackle the problem of fore- casting rare extreme values of a continuous target variable using standard regression tools. The key advantage of the described re-sampling approaches is their simplicity. They allow the use of standard out-of-the-box regression tools on these particular prediction tasks by simply manipulating the distri- bution of the available training data. The key contributions of this paper are : i) showing that re-sampling approaches can be successfully applied to this type of regression tasks; and ii) adapting two of the most successful sampling methods (Smote and under- sampling) to regression tasks. The large set of experiments we have carried out on a diverse set of problems and using rather different learning algorithms, highlights the ad- vantages of our proposals when compared to the alternative of simply apply- ing the algorithms to the available data sets. Particularly noticeable are the results obtained with strong under-sampling of the normal cases. Moreover, in the case of under-sampling with 50% the good predictive performance is accompanied by lower computation costs. R code implementing both the SmoteR and under-sampling strategies described in Section 3 is freely provided at http://www.dcc.fc.up.pt/ ~ltorgo/ExpertSystems. This URL also includes all code and data sets 18 necessary to replicate the experiments in the paper as well as extra experi- mental results that we could not fit within the paper. Acknowledgements This work is part-funded by the ERDF - European Regional Development Fund through the COMPETE Programme (operational programme for com- petitiveness), by the Portuguese Funds through the FCT (Portuguese Foun- dation for Science and Technology) within project FCOMP - 01-0124-FEDER- 022701. References Chawla, N. V. (2005). Data mining for imbalanced datasets: An overview. In The Data Mining and Knowledge Discovery Handbook. Springer. Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. JAIR, 16:321–357. Cleveland, W. (1993). Visualizing Data. Hobart Press. Davis, J. and Goadrich, M. (2006). The relationship between precision-recall and roc curves. In ICML’06: Proc. of the 23rd Int. Conf. on Machine Learning, ACM ICPS, pages 233–240. ACM. Dimitriadou, E., Hornik, K., Leisch, F., Meyer, D., and Weingessel, A. (2011). e1071: Misc Functions of the Department of Statistics (e1071), TU Wien. Domingos, P. (1999). Metacost: A general method for making classifiers cost-sensitive. In KDD’99: Proceedings of the 5th International Confer- ence on Knowledge Discovery and Data Mining, pages 155–164. ACM Press. Elkan, C. (2001). The foundations of cost-sensitive learning. In IJCAI’01: Proc. of 17th Int. Joint Conf. of Artificial Intelligence, volume 1, pages 973–978. Morgan Kaufmann Publishers. Kubat, M. and Matwin, S. (1997). Addressing the curse of imbalanced training sets: One-sided selection. In Proc. of the 14th Int. Conf. on Machine Learning, pages 179–186. Morgan Kaufmann. Liaw, A. and Wiener, M. (2002). Classification and regression by random- forest. R News, 2(3):18–22. Ling, C. and Sheng, V. (2010). Cost-sensitive learning and the class imbal- ance problem. In Encyclopedia of Machine Learning. Springer. 19 Milborrow, S. (2012). earth: Multivariate Adaptive Regression Spline Mod- els. Derived from mda:mars by Trevor Hastie and Rob Tibshirani. R Core Team (2013). R: A Language and Environment for Statistical Com- puting. R Foundation for Statistical Computing, Vienna, Austria. Ribeiro, R. P. (2011). Utility-based Regression. PhD thesis, Dep. Computer Science, Faculty of Sciences - University of Porto. Torgo, L. and Ribeiro, R. P. (2007). Utility-based regression. In PKDD’07: Proc. of 11th European Conf. on Principles and Practice of Knowledge Discovery in Databases, pages 597–604. Springer. Torgo, L. and Ribeiro, R. P. (2009). Precision and recall in regression. In DS’09: 12th Int. Conf. on Discovery Science, pages 332–346. Springer. Zadrozny, B. (2003). Policy mining: Learning decision policies from fixed sets of data. PhD thesis, University of California, San Diego. Zadrozny, B. (2005). One-benefit learning: cost-sensitive learning with re- stricted cost information. In UBDM’05: Proc. of the 1st Int. Workshop on Utility-Based Data Mining, pages 53–58. ACM Press. 20 
work_khbyjrfnerfmdiy6tnotx4dfl4	----	Discovering During-Temporal Patterns (DTPs) in Large Temporal Databases ∗ Li Zhanga, Guoqing Chena,†, Tom Brijsb, Xing Zhanga a School of Economic and Management, Tsinghua University, Beijing, 100084, P.R.China b Transportation Research Institute, Hasselt University, Diepenbeek, B3920, Belgium Abstract Large temporal Databases (TDBs) usually contain a wealth of data about tem- poral events. Aimed at discovering temporal patterns with during relationship (during-temporal patterns, DTPs), which is deemed common and potentially valuable in real-world applications, this paper presents an approach to finding such DTPs by investigating some of their properties and incorporating them as desirable pruning strategies into the corresponding algorithm, so as to optimize the mining process. Results from synthetic reveal that the algorithm is efficient and linearly scalable with regard to the number of temporal events. Finally, we apply the algorithm into the weather forecast field and obtain effective results. Keywords: data mining; during relationship; temporal pattern 1 Introduction In recent years, discovery of association rules [14] and sequential patterns [13] has been a major research issue in the area of data mining. While typical association rules usually reflect related events occurring at the same time, sequential patterns represent commonly occurring sequences that are in a time order. However, real-world businesses often generate a massive volume of data in daily operations and decision-making processes, which are of a richer temporal nature. For instance, a customer could buy a DVD machine after TV was bought; the duration of an ERP project partially overlapped the duration of a BPR project; and a patient suffered from cough during the period of fever. Apparently, ∗The work was partly supported by the National Natural Science Foundation of China (70231010/70321001), Tsinghua University’s Research Center for Contemporary Management, and the Bilateral Scientific and Technological Cooperation between China and the Flanders. †Corresponding author.E-Mail: chengq@em.tsinghua. edu.cn 1 such temporal relationships (e.g., after, overlap, during, etc.) are kinds of real-world semantics that are, in many cases, considered meaningful and useful in practice. Usually, temporal relationships between events with different time stamps could be categorized into several types in forms of temporal comparison predicates such as after, meet, overlap, during, start, finish, and equal [10]. Though recent years have witnessed several efforts on discovering the after relationship [7,8,11,13], more in-depth investigations of the relationship are still badly needed, let along their explorations of other types of temporal relationships. Furthermore, results from the studies on the after relationship could hardly be simply extended to the case of some other relationships such as during, overlap, etc. This may be attributed to the fact that in the after relationship, events can generally be dealt with on a time point, whereas in other relationships, events are considered to be of a time interval nature. On the other hand, both Rainsford [3] and Hoppner [6] have recently discussed the issues of finding temporal relationships between time-interval-based events using temporal comparison predicates [10], but with different mining approaches. Rainsford introduced temporal semantics into association rules, in forms of X⇒Y∧P1∧P2∧...∧Pn (n≥0), where X and Y are itemsets, and X ∩ Y = ∅. P1∧P2∧...∧Pn is a conjunction of binary temporal predicates. While mining a database DT , a rule is accepted when its confidence factor 0≤c≤1 is equal to or larger than the given threshold. Similarly, each predicate Pi is measured with a temporal confidence factor 0≤tcPi ≤1. The algorithm firstly generates the traditional association rules without considering the temporal factors, and then finds all of the possible pairings of temporal items in each rule. Subsequently, these pairings are tested so that strong temporal relationships could be found. Obviously, the complexity of this sequentially executed algorithm rises rapidly as the number of typical rules grows. Differently, Hoppner proposed another technique for discovering temporal patterns in state sequences. He defined the supporting level of a pattern as the total time in which the pattern can be observed within a sliding window, which should be predetermined by the user. However, a major concern for this technique is how to decide a proper size for the sliding window, since the sliding window can affect the mining results. Furthermore, The changes of the sliding window will lead to a sub-patterns check. The check requires some backtracking mechanism, which is computationally expensive. Like many existing data mining algorithms, the algorithm needs to scan the database repeatedly, which would significantly lower its efficiency. This paper will focus on a particular type of temporal relationships, namely during, which rep- resents that one event starts and ends within the duration of another event. Notably, this during relationship could reflect the temporal semantics of during, start, finish and equal described in [10]. An approach will be proposed to discover the so-called during-temporal patterns (DTPs) in larger temporal databases, which are considered common and potentially valuable in real-world applications. One idea behind the approach is to design the corresponding algorithm so as to reduce the workload in 2 scanning the database. In doing so, the database is partitioned into some disjoined datasets with two operations when calculating the support level of each pattern, so that scanning the whole database could be avoided. Furthermore, some properties of DTPs are investigated and then incorporated into the algorithm as pruning strategies to optimize the mining process for efficiency purposes. The remainder of this paper is organized as follows. Section 2 formulates the problem and introduces related notions. In Section 3, the algorithmic details are provided, along with some of the related properties. The experiments on synthetic data and real weather data are discussed in Section 4, and Section 5 concludes the paper. 2 The problem formulation Let A={a1,a2,...,am} be a set of states, and DT a temporal database as shown in Table 1. Given a database DT with N records, each of which is in the form of {a,(st,et)} with respect to event e, i.e., e=(a,t), where a is the state involved in the event, t=(st,et) is the time interval which indicates starting time (st) and ending time (et) of state a in the event. A specific event is denoted as el =(ai ,tl ) (1≤l≤N and 1≤i≤m) and tl = (stl , etl ), i.e., S(el )=stl and E(el )=etl . For example, with a1=rain, e1=(a1,(1,20)) in Table 1 means that it began to rain at 1:00h and ended at 20:00h. Table 1: A Temporal Database Event State Starting Time Ending Time e1 a1 1 20 e2 a3 1 4 e3 a4 5 7 e4 a1 22 28 e5 a2 2 8 e6 a3 10 13 e7 a5 25 35 e8 a3 23 28 e9 a4 25 27 e10 a6 25 26 e11 a1 30 40 e12 a3 30 38 e13 a4 34 38 e14 a6 37 37 Definition 1 Let el =(ai ,tl ) and ek =(aj ,tk ) be two events in DT . We call el during ek (or ek contains el ), denoted as el <dek , if S (el ) ≥ S (ek ) and E (el ) ≤ E (ek ) Generally, given a set of events {e1, e2, ..., ek}, if ei+1 during ei is satisfied for all i=1,2,...,k-1, we have ek <d ek−1 <d ... <d e2 <d e1. 3 For any two states ai and aj , ai is called to be during aj , denoted as pattern ai⇒daj , if state ai occurs during the period of another state aj , which is a during-temporal pattern (DTP) with length 1. Generally, a DTP of length (k-1) (k≥1), namely DTPk−1, is of the form: ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d a1 When the length is 0 (i.e., DTP0), the pattern is a single state actually. More generally, given two patterns α and β, the form α⇒dβ is also a DTP (Aα ∩ Aβ = ∅, where Aα and Aβ are the sets of states included in pattern α and β respectively). As a special case, it retrogresses to ai⇒daj when the lengths of both patterns are 0. Given a pattern ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d a1 , ek <d ek−1 <d ... <d e2 <d e1 supports this pattern if ai is the state of ei for all i=1,2,...,k. In the case, ek <d ek−1 <d ... <d e2 <d e1 can be considered as an instance of this pattern. For example, in Table 1, e10<de9<de8 is an instance of the pattern a6⇒da4⇒da3, and e14<de13<de12 is another instance of this pattern. Further, a DTP α ak ⇒d ak−1 ⇒d ... ⇒d aj +1 ⇒d aj ⇒d aj−1 ⇒d ... ⇒d a2 ⇒d a1 is characterized by the number of its instances: ek <d ek−1 <d ... <d e2 <d e1, and a DTP β ⇒d γ, i.e., (ak ⇒d ak−1 ⇒d ... ⇒d aj +1) ⇒d (aj ⇒d aj−1 ⇒d ... ⇒d a2 ⇒d a1)(1 ≤ j ≤ k − 1) is characterized by the number of its instances: (ek <d ek−1 <d ... <d ej +1) <d (ej <d ej−1 <d ... <d e2 <d e1) which is ek <d ek−1 <d ... <d ej +1 <d ej <d ej−1 <d ... <d e2 <d e1 according to the associative law of events. Hence, β ⇒d γ equivalently reads ak ⇒d ak−1 ⇒d ... ⇒d aj +1 ⇒d aj ⇒d aj−1 ⇒d ... ⇒d a2 ⇒d a1 (1 ≤ j ≤ k − 1) Furthermore, for finding all instances of a DTP α, one may consider to scan the whole database. In fact, however, only a small part of the database, with respect to the set of states included in α, is useful. Thus, we can divide the database into m datasets (m is the number of the states in the database), each of which is the set of time intervals of a single state. Thus, when finding all instances of ak ⇒d ak−1 ⇒d ... ⇒d aj +1 ⇒d aj ⇒d aj−1 ⇒d ... ⇒d a2 ⇒d a1, only those datasets which include the sets of time intervals of ak , ak−1, ..., a2, and a1 are scanned. For this purpose, we define a time interval set g(α) and a state set h(α) to partition the database, and join the small datasets to count the support (which will be discussed in this and next sections). Definition 2 Let A and T be finite sets of states and time intervals respectively with respect to a temporal database DT . For pattern α, i.e., ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d a1 (k≥1), we define the set 4 of time intervals g(α) as the set of finest time intervals of all the instances of α. Formally, g(α) is the function mapping the set of patterns to the set of time intervals, g(α) = {t|(ai , t) ∈ DT }, if the length of α is 0, i.e., α is a single state ai . g(α) = {tk ∈ g(ak )|for all i = 1, 2, ..., k − 1, ai ∈ Aα, ∃ti ∈ g(ai ), such that ti+1 ∩ ti = ti+1}, (2-1) if the length of α is larger than 0. In the definition, tl ∩ tk = (max{stl , stk}, min{etl , etk}). Equivalently, we have g(α) = {tk|for each instance of α : ek (ak , tk ) <d ek−1(ak−1, tk−1) <d ... <d e1(a1, t1)} (2-2) g(ai ) includes all the intervals in which ai occurred, so all instances of α can be found using g(a1), g(a2),...,g(ak ). And ti+1 ∩ ti = ti means ei+1 <d ei , for i=1,2,...,k-1. That is, ti+1 ∩ ti = ti for i=1,2,...,k-1 means that ek (ak , tk ) <d ek−1(ak−1, tk−1) <d ... <d e2(a2, t2) <d e1(a1, t1). Hence, both (2-1) and (2-2) get the set of finest time intervals of all the instances of α. Support g(a1) h(a1) 1 (1,20) 2 (22,28) a2,a3,a4,a6 3 (30,40) Support g(a2) h(a2) 1 (2,8) a4 Support g(a3) h(a3) 1 (1,4) 2 (10,13) 3 (23,28) a4,a6 4 (30,38) (a) (b) (c) Support g(a4) h(a4) 1 (5,7) 2 (25,27) a6 3 (34,38) Support g(a5) h(a5) 1 (25,35) a4,a6 Support g(a6) h(a6) 1 (25,26) 2 (37,37) ∅ (d) (e) (f) Support g(a3 ⇒d a1) h(α) 1 (1,4) (10,13) 2 (23,28) a4,a6 3 (30,38) Support g(a4 ⇒d a1) h(α) 1 (5,7) 2 (25,27) a3,a6 3 (34,38) (g) (h) Figure 1: The examples of the sets g and h For example, given a temporal database DT as shown in Table 1. g(a1)={(1,20), (22,28), (30,40)} and g(a3)={(1,4), (10,13), (23,28), (30,38)}. According to ti+1∩ti = ti+1 in (2-1), we have (1,4)∩(1,20) =(1,4), (10,13)∩(1,20)=(10,13), (23,28)∩(22,28)=(23,28), and (30,38)∩(30,40)=(30,38), so g(a3 ⇒d a1)={(1,4), (10,13), (23,28), (30,38)}. According to (2-2), we need to find these intervals from original temporal database. The instances of a3 ⇒d a1 include e2 <d e1, e6 <d e1, e8 <d e4 and e12 <d e11, which result in time intervals (1,4), (10,13), (23,28), and (30,38) respectively. That is, g(a3 ⇒d 5 a1)={(1,4), (10,13), (23,28), (30,38)}. More examples can be found in Figure 1. Next, we will define the support degree of a DTP pattern (Definition 3). Definition 3 The support degree of a DTP α is the fraction of the support count of the pattern. That is, support(α) = |g(α)||g0| where |g(α)| is the number of time intervals in g(α) without double-counting those intervals of the instances in that an event contains several events with the same state, and |g0|= max{|g(ai )|, for i=1,2,...,m}. Actually, |g(α)| is the number of instances supporting α without double-counting. α is said to be frequent if the support degree is not less than the given threshold (i.e., minsupport). The support degree of pattern α in Definition 3 is the ratio of the number of time intervals included in all instances of α (without double-counting) over the maximum number of time intervals among |g(ai )| for all i. In other words, support(α) reflects the relative frequency of time intervals for α with respect to the number of time intervals for a most frequent state. In the first place, by ‘without double- counting’ we mean that the instances with an event containing several events having the same state will only be counted once, as they all support the same single pattern. In the second place, alternatively g0 may be defined as N= ∑m i=1|g(ai )| , for the same purpose. However, since N= ∑m i=1|g(ai )| is usually much larger than |g(α)|, it will result in too small values for support degrees. Therefore, a scale- down measure is often considered desirable. As a matter of fact, other forms of g0 could be possible, depending on the context and convenience. Notably, since g0 is a fixed number, the choice of it is a technical treatment and does not affect the properties of DTPs. Take Table 1 as an example again. For a DTP pattern α: a3⇒da1, we have |g(α)| =|{(1,4), (10,13), (23,28), (30,38)}|=3, and |g0|=4. Thus, support(α)=3/4=0.75. Here, we have an event that contains several events with the same state. That is, e1=(a1,(1,20)), e2=(a3,(1,4)), and e6=(a3,(10,13)), we have e2<de1 and e6<de1, which contribute to the same DTP a3⇒da1. This counting is also similar to that described in [2]. t a3 a2 a2 a1 a3 a2 Figure 2: A counterexample for pattern transitivity Note that the transitivity for during relationships between events exists. That is, if e <d ep and ep <d eq, then we have e <d eq. For instance, we have e10<de9 and e9<de8 in Table 1, and e10<de9<de8. However, it is worth mentioning that the during relationship are not transitive between 6 patterns in terms of support degree. Let us consider an example as follows. Given a temporal database DT ={(a3,(1,5)), (a2,(2,4), (a2,(6,15)), (a1,(8,15)), (a3,(10,20)), (a2,(20,20))} and minimal support count=1, with time intervals for the events being illustrated in Figure 2, we have |g(a1 ⇒d a2)|=1, |g(a2 ⇒d a3)|=2, but |g(a1 ⇒d a3)|=0. Thus, we cannot obtain longer patterns from short ones using transitivity. This gives rise to the effort to find other ways of generating longer patters. First, in examining whether DTPs are frequent, we may need to consider sub-patterns. Definition 4 introduces the notion. Definition 4 For a DTP α with length l, we say that α is a sub-DTP of another DTP β with length k, denoted as α¹β if l≤k and there exists an order-preserving mapping ϕ: {1,2,...,l}→{1,2,...,k} such that α(1) ⇒d α(2) ⇒d ... ⇒d α(l) is the same to β(ϕ(1)) ⇒d β(ϕ(2)) ⇒d ... ⇒d β(ϕ(l)) where α(i) is the ith state in the pattern α. For instance, a6⇒da4⇒da1 is one of sub-DTPs of the pattern a6⇒da4⇒da3⇒da1. Moreover, we say g(α)⊇g(β) if for every tk∈g(β) there exists a time interval tl∈g(α) such that tl ∩ tk = tk . In terms of events, g(α)⊇g(β) means that for every event ek with tk in g(β), there exists an event el with tl in g(α) such that ek <del . Note that g(α)⊇g(β) does not necessarily mean |g(α)|≥|g(β)|. For example, as shown in Figure 1, g(a1)⊇g(a3) but |g(a1)|<|g(a3)|. Importantly, for two DTPs α and β with α¹β, one could expect that a longer DTP in length will have a less chance of being supported than its sub-DTPs since the longer the length, the fewer its supporting instances in the database. Accordingly, the fewer the supporting instances for a longer DTP, the finer the set that contains the time intervals of the longer DTP. These statements are proved in Property 1. Property 1 if α¹β, then g(α)⊇g(β) and |g(α)|≥|g(β)| . Proof: Suppose that g(α)⊇g(β) does not hold. Therefore, there exists at least a time interval tk∈g(β), such that tl∩tk 6=tk for all tl∈g(α). According to the definition about the set g, every state in Aβ is active in tk . However, not all the states in Aα are active in tk since tk cannot be totally contained by any interval in g(α). That is, some states in Aβ are not active in tk since α¹β and Aα⊆Aβ . This is a contradiction with tk∈g(Aβ ). That is, there must be g(α)⊇g(β) if α¹β. Furthermore, each time interval tk∈g(β) is contained by the corresponding interval tl∈g(α) since g(α)⊇g(β). Some tl could contain several tk , and some could not contain any tk . Let |g(α)|=x, {T1,T2,...,Tx} be the set of time interval sets, in which Ti (i=1,2,...,x) represents the set of time intervals contributing only once for the support count. Let the variable yi be the flag which represents whether a new instance without double-counting is found in Ti . If a new instance without double- 7 counting has been found, then yi =1. Otherwise, yi =0. Thus, |g(β)| = ∑xi=1yi ≤ x × 1 = x = |g(α)| ¥ Property 2 All subsets of a frequent pattern are frequent. Proof: Let β be a frequent pattern, and α is a sub-DTP of β. Thus, we have support(β)≥ minsupport according to Definition 3, and |g(α)|≥|g(β)| according to Property 1. That is, support(α)≥support(β)≥minsupport, which means that α is frequent. ¥ The above two properties are very important in the process of finding frequent patterns. Effort in scanning the database and examining longer DTPs could then be largely saved by only concentrating on those frequent (sub-)patterns. This is because any DTP containing a non-frequent sub-DTP will not be frequent. A frequent pattern means that it occurs in forms of events in a sufficient level of frequency. Usually, one also needs to know how likely a pattern occurs given that the other pattern has already occurred. In other words, we are considering the notion of confidence. Concretely, given two DTPs β and γ, where β=ak ⇒d ak−1 ⇒d ... ⇒d aj +1 of length (k-j-1) and γ=aj ⇒d aj−1 ⇒d ... ⇒d a2 ⇒d a1 of length (j-1)for j={1,2,,k-1}. Then β during γ is a composition of these two DTPs as β⇒dγ which forms a DTP α as follows: α : (ak ⇒d ak−1 ⇒d ... ⇒d aj +1) ⇒d (aj ⇒d aj−1 ⇒d ... ⇒d a2 ⇒d a1) = ak ⇒d ak−1 ⇒d ... ⇒d aj +1 ⇒d aj ⇒d aj−1 ⇒d ... ⇒d a2 ⇒d a1 Definition 5 The confidence degree of a DTP α: β⇒dγ is defined as the fraction of the support of pattern α over the support of the consequent pattern γ. conf idence(α) = support(α) support(γ) = |g(α)||g(γ)| The DTP β⇒dγ is a valid DTP if the support and confidence degrees exceed the corresponding thresholds (minsupport and minconfidence). Consider Figure 1 again as an example with β and γ being of length 0, i.e., β=a3 and γ=a1. confidence(a3 ⇒d a1) = support(a3 ⇒d a1)/support({a1})=3/3=1. Finally, given a DTP α (i.e., ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d a1(k≥1)), let us consider a set of states occurring during the period a1 excluding those states already contained by Aα. The state in the set 8 shall be during a1 in a frequent manner. Concretely, such a set, h(α), is defined as: h(α) = {aj |aj 6∈ Aα, j =1, 2, ..., k , and if α is of a lengh ≥ 1 then aj ⇒d a1, support(aj ⇒d a1) ≥ minsuport} The set h(α) will be used conveniently in the process of generating candidate DTPs, which will be discussed in detail in the following sections. Merely for illustrative purposes, in Figure 1, we have, for example, h(a1)={a2,a3,a4,a6}. In comparison, we have h(a3 ⇒d a1)={a4,a6} where a2 is not included since a2⇒da1 is not frequent, and a3 is not included either since a3 is already in Aα. 3 The algorithm Straightforwardly, the problem of mining DTPs can be solved by traversing the relationship trees through the paths representing all possible during relationships. First, by scanning the temporal database, all frequent DTPs of length 1 (DTP1 ps) can be discovered. Each tree represents all the during relationships that have a public parent-state. Figure 3 shows an example, in which → represent contain. a2 a5 a6 a9 a10 a11 a12 a6 a5 a4 a2 a7 a9 a10 a11 a12 a11 a5 a6 a7 a9 a12 a13 Figure 3: trees of frequent DTP1 Next, we can traverse each tree using the following strategy: (1) Let Tl be a tree of frequent DTP1 ps. The root of the tree is R(Tl ). The set of leaf nodes of Tl is L(Tl )={L1,L2,...,Ln}. (2) Using the depth-first search, we completely solve one Tl (l =1,2,...,m) along one path before moving on to the next path. Search path Pj which represents the pattern α. The leaf node of Pj is Lj . Calculate the support degree support(α), and If Lj ∈ {R(Ti )|i =1,2,..,m, and i6=l}, If support(α) < minsupp, move on to the next path and repeat this step. Else calculate the confidence degree and get the valid DTPs, and then turn to (3). 9 Else move on to the next path and turn to (2). (3)Consider the leaf nodes of Ti which are child nodes of Lj in Ti , i.e., L(Tl (Lj ))={L(Ti}, the new tree T ′l is generated, T ′ l =Tl⊕L(Tl (Lj )), where ⊕ is an operation that treats all leaf nodes of Ti as the child nodes of Lj in and forms the new tree T ′l . Thus, path Pj is extended as a sub-tree, which represents several DTPs. Turn to (2). Stop it until no new tree is generated. Take Figure 3 as an example. For the tree with the root being a4, and Pj : a4→a2, we process the tree in the following order a4→a2, a4→a2→a5, a4→a2→a6,a4→a2→a6→a5..., a4→a2→a11, a4→a2→a11→a5, a4→a2→a11→a6, a4→a2→a11→a6 →a5, and so on. Note that this algorithm firstly finds the DTPs by traversing the trees in the way of depth-first search such that Property 1 can be applied. However, it needs to enumerate all possible candidate DTPs from DTP1 and will consume much time. The method is referred to as Tree algorithm for distinguishing and comparing with the optimized approach that is to propose in the next subsections. 3.1 DTP Algorithm To tackle the above problem, we propose a new algorithm, namely the DTP algorithm, which works in an iterative manner by alternating between the joining and pruning phases, after finding frequent DTP0 ps. As the first step, support levels of all single states DTP0 ps are calculated and frequent single state patterns are found. Meanwhile, the whole database DT is divided into mp (mp≤m) parts (mp frequent single states). Secondly, in the join phase of an iteration k, a collection CDTPk of new candidate patterns with length k is generated from the set of frequent DTPk−1 ps(FDTPk−1). Thirdly, the collection FDTPk of new frequent DTPk−1 ps are generated by computing the support degrees. Lastly, the patterns satisfying the confidence threshold are output after all iterations are terminated. More concretely, the algorithmic details are described as follows. (1). Finding frequent DTP0 By scanning the whole database DT only once and applying the function g(ai ) to each event, DT is divided into m parts and the support of each single state can be computed easily. Let mp=|FDTP0| be the number of frequent single states. So, the number of valid datasets is mp. In the second scanning, h(ai ) can be calculated for every frequent single state ai . The procedure is shown in Figure 4. FDTP0.Gen() 1. FDTP0 =∅; 2. for all ai∈A do 3. if |g(ai )|≥ minsupport then 4. FDTP0=FDTP0∪{ai}; 5. end if 6. end for Figure 4: The procedure used for F DT P0 10 (2). Join phase The algorithm employs an iterative approach known as a level-wise search, where FDTPk−1 is used to explore FDTPk (k≥1). First, the set of frequent DTP0(namely FDTP0) is found. FDTP0 is used to find FDTP1, which is used to find FDTP2, and so forth, until no more frequent pattern can be found. The generation of each frequent pattern carries a search cost. To improve the efficiency of generating frequent patterns, two properties are presented as follows. Let p(α,l) be a sub-pattern with the length (l-1) of the pattern α intercepted from left to right, and p(α,-l) be a sub-pattern with the length (l-1) of the pattern α from right to left. When l is 1, p(α,±1) is a single state essentially; when l is 0, p(α,0) returns an empty set. For example, given the pattern α, i.e., a3⇒da2⇒da1, we can get p(α,2)=a3⇒da2, p(α,-1)=a1, and p(α,0)=∅. Property 3 Let β be a frequent pattern, i.e., ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d a1. The pattern γ: ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d ap ⇒d a1 is not frequent if ap 6∈h(β). Proof: According to the notion of set h, ap⇒da1 is not frequent if ap 6∈h(β). As we know, all sub-DTPs of a frequent DTP are frequent, and ap⇒da1 is one sub-DTP of γ. So, the pattern γ must not be frequent. ¥ Thus, to find FDTPk , a set of candidate DTPs with length k is generated by adding ap∈h(β) into the frequent pattern β. This set of candidates is denoted CDTPk , which is a superset of FDTPk . That is, its members may or may not be frequent, but all of the frequent DTPk are included in CDTPk . In order to obtain a minimal candidate set, we introduce the following property to check whether more sub-DTPs are frequent. Property 4 Given a pattern β ∈FDTPk−1(k =1,2,...), for any ap∈h(β), if the two following conditions are satisfied: (i) pattern α: p(β, k )⇒dap, α∈FDTPk−1. (ii)p(α,-k )⇒dp(β,-1)∈FDTPk−1. we can join α and β, and add a new candidate pattern, α⇒dp(β,-1), into CDTPk . Proof: Let γ be p(α,-k)⇒dp(β,-1). All of the three patterns α, β and γ are the sub-DTPs of the pattern α⇒dp(β,-1). Once any of them is not frequent, the pattern α⇒dp(β,-1) must not be frequent (Property 2). That is, we add the pattern α⇒dp(β,-1) into the candidate set only when both α and γ are frequent. ¥ That is, in the kth iteration, for each ap∈h(β), check if both patterns α and γ, i.e., p(β,k )⇒dap and p(α,-k )⇒dp(β,-1) respectively, are frequent. Let us consider all the DTPs as a lattice as shown in Figure 5. And one sublattice is the set of the patterns with the same top element(e.g., a1,a3,a4, and a6). 11 For example, let β be a6⇒da4⇒da1 in the sublattice a1. For a3∈h(β), we can search the sublattices a3 and a1 to find the patterns satisfying Property 4, i.e., α a6⇒da4⇒da3 and γ a4⇒da3⇒da1. Hence, the new candidate pattern a6⇒da4⇒da3⇒da1 is obtained, which is added into sublattice a1. The lattice of frequent patterns is shown in Figure 5. a3=> d a1 a4=> d a1 a6=> d a1 a4=> d a3 a6=> d a3 a6=> d a4 a4=> d a3=> d a1 a6=> d a3=> d a1 a6=> d a4=> d a1 a6=> d a4=> d a3 a6=> d a4=> d a3=> d a1 a1 a3 a4 a6 {} Figure 5: The lattice of frequent patterns Next, the set g(α ⇒d p(β, −1)) for new candidate pattern α⇒dp(β,-1) is calculated and the cor- responding h(α ⇒d p(β, −1)) are obtained. We define the operation ∩d of the sets g(α) and g(β) as follows: g(α) ∩d g(β) = {tl ∈ g(α)|∃tk ∈ g(β), such that tl = tl ∩ tk} That is, if time interval tl ∈ g(α) is totally contained by a time interval tk ∈ g(β), then tl is an element of the set g(α) ∩d g(β). Proposition 1 Let β be ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d a1 , α be ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d ap and γ be ak−1 ⇒d ... ⇒d a2 ⇒d ap ⇒d a1. The three patterns can generate a longer pattern σ: ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d ap ⇒d a1. The set g for the longer pattern σ is g(σ) = g(α) ∩d g(γ) = g(β) ∩d g(γ). 12 Proof: According to the definition of g(α), we have g(α) = g(ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d ap) = {tk ∈ g(ak )|∃tp ∈ g(ap), tk ∈ g(ak ), ti ∈ g(ai ), such that ti+1 ∩ ti = ti+1, t2 ∩ tp = t2, for all i = 2, 3, ..., k − 1} g(γ) = g(ak−1 ⇒d ... ⇒d a2 ⇒d ap ⇒d a1) = {tpk−1 ∈ g(ak−1)|∃tpp ∈ g(ap), tpk−1 ∈ g(ak−1), tp1 ∈ g(a1), tpi ∈ g(ai ), such that tpi+1 ∩ tpi = tpi+1, tp2 ∩ tpp = tp2, tpp ∩ tp1 = tpp, for all i = 2, 3, ..., k − 2} g(σ) = g(ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d ap ⇒d a1) = {tqk ∈ g(ak )|∃tqp ∈ g(ap), tqk ∈ g(ak ), tq1 ∈ g(a1), tqi ∈ g(ai ), such that tqi+1 ∩ tqi = tqi+1, tq2 ∩ tqp = tq2, tqp ∩ tq1 = tqp, for all i = 2, 3, ..., k − 1} Firstly, we prove g(σ) ⊆ g(α)∩d g(γ). ∀tqk ∈ g(σ), from g(σ) we have tqi+1∩tqi = tqi+1 and tq2 ∩tqp = tq2 for tqi ∈ g(ai ), i=2,3,...,k-1. Let tk = tqk , tqk meets the conditions in g(α). So g(σ) ⊆ g(α). And then, we need to prove for any tqk ∈ g(σ), there exists tpk−1 ∈ g(γ), such that tqk ∩ tqk−1 = tqk . ∀tqk ∈ g(σ), there is a group of time intervals tqk−1, t q k−2,...,t q 2, t q 1, t q p satisfying the conditions in the form of g(σ). This group of time intervals can also meet the conditions in the form of g(γ), so tqk−1 ∈ g(γ). Take tpk−1 = t q k−1, we have t q k ∩ tpk−1 = tqk since tqk ∩ tqk−1 = tqk according to the conditions in g(σ). Hence, we get g(σ) ⊆ g(α) ∩d g(γ). Secondly, we prove g(α) ∩d g(γ) ⊆ g(σ). ∀tk ∈ g(α) ∩d g(γ) means ∃tk ∈ g(α), tpk−1 ∈ g(γ), such that tk ∩ tpk−1 = tk . tpk−1 ∈ g(γ) means that there exist corresponding tpk−2,...,tp2, tpp, tp1 satisfying tpi+1 ∩ tpi = tpi+1, tp2 ∩ tpp = tp2, and tpp ∩ tp1 = tpp for all i=2,3,...,k-2. Take tqk = tk , tqp = tpp, and tqi = tpi for all i=1,2,...,k-1, and then we have tqk ∩ tqk−1 = tk ∩ tpk = tk ; tqi+1 ∩ tqi = tpi+1 ∩ tpi = tpi = tqi+1 (i=2,3,...,k-2); tq2 ∩ tqp = tp2 ∩ tpp = tp2 = tq2; tqp ∩ tq1 = tpp ∩ tp1 = tpp = tqp which means that tk ∈ g(σ). Thus, for any tk ∈ g(α) ∩d g(γ), tk ∈ g(σ). That is, g(α) ∩d g(γ) ⊆ g(σ). In the same way, we can get g(σ) = g(β) ∩d g(γ). ¥ The proposition indicates that the set g for a new candidate can be computed by the frequent patterns g(α) and g(γ), or g(β) and g(γ), while the set h for a new candidate pattern can only be obtained from h(γ), 13 h(α ⇒d p(β, −1)) = h(α ⇒d p(γ, −1)) = h(γ) − {ak} Thus, we get the sets g and h for a new candidate pattern without scanning the original database and the sets g for the single states. The procedure of join phase is shown in Figure 6. CDTPk .Gen() 1. for all pattern β∈FDTPk−1 do 2. for all aj∈h(β) do 3. for α∈FDTPk−1,p(α,-1)=aj do 4. if Property 4 is satisfied then 5. CDTPk =CDTPk S{α⇒dp(β,-1)} 6. end if 7. end for 8. end for 9. end for Figure 6: The procedure used for CDTPk (3). Pruning phase In this phase, if the support degree of a candidate pattern in CDTPk is smaller than the user- specified threshold, then prune it from CDTPk . At last, FDTPk is obtained. Intercross Step 2 and Step 3 until Property 2 cannot be satisfied any longer. The procedure of pruning phase is shown in Figure 7. FDTPk .Gen() 1. FDTP0 known, FDTPk =∅ (k=1,2,...) 2. for (k=1;FDTPk−1 6=∅;k++) do 3. CDTPk =CDTPk .Gen(FDTPk−1); 4. for all candidate patterns β∈CDTPk do 5. FDTPk ={β∈CDTPk||g(β)|≥minsupport} 6. end for 7.end for 8. FDTP= S k FDTPk Figure 7: The procedure used for FDTPk (4). Generating valid DTPs Given a pattern α: ak ⇒d ak−1 ⇒d ... ⇒d a2 ⇒d a1, it is necessary to know how frequent α is when a1 has occurred, when a2 ⇒d a1 has occurred, when a3 ⇒d a2 ⇒d a1 has occurred, and so on. Thus, we can get the pattern β⇒dγ as mentioned in Section 3, (ak ⇒d ak−1 ⇒d ... ⇒d aj +1) ⇒d (aj ⇒d aj−1 ⇒d ... ⇒d a2 ⇒d a1) (1 ≤ j ≤ k − 1) That is, a frequent pattern can generate (k-1) valid DTPs at most. Calculating the confidence degree starting from the longest consequent pattern, i.e., from the pattern with j=k-1, the patterns meeting the confidence threshold will be valid DTPs. The pattern α will be stopped to compute if confidence(β⇒dγ) 14 is less than the threshold. Otherwise, the confidence degree of a shorter consequent pattern, i.e., j=j-1, is computed next. In discovering DTPs, a temporal database as shown in Tabel 1 is usually needed, which could be obtained either directly or by converting conventional databases. Since a DTP is acturally an event sequence in terms of time inclusion (i.e., during relationship), the records of a conventional database needs to be sorted by ascending start time primarily and descending end time secondarily. Consider the database in Table 1: DT ={e1,e2,e5,e3,e6, e4,e8,e7,e9,e10,e11, e12,e13,e14}, and in a sorted form as DT ={ep1,ep2, ep3,...,ep12,ep13,ep14}. Subsequently, the set of the resultant events {epi ,epi+1,...,ei+k} (k=1,2,...) is called a during-sequence if epi+j < depi+j−1 for all j=1,2,..,k and e p i+k+1≮ depi+k . For example, in the sorted DT , {e1,e2,e5,e3,e6} is a during-sequence since e6<de3<de5<de2<de1 and e6 is not during the next event e4. {e4,e7,e8,e9,e10} is the next one in the example. With data sorted in this way, the search space and the comparison of start time and end time can be reduced when the temporal database is scanned. Let el (ai ,tl ),ek (aj ,tk ), in which tl = (stl , etl ) and tk = (stk , etk ), be the lth and kth event about state ai and aj (i 6= j) in the sorted DT , respectively, and k is larger than l, el will not occur during the valid period of ek unless S(el )=S(ek ) and E(el )=E(ek ). Property 5 Assume that k is larger than l, and ek (aj ,tk )≮del (ai ,tl ). If there is another event ew (aj ,tw )(k<w≤N) with aj , ew must not occur during the period of el , i.e., ew≮del . Proof : since w>k and both ek and ew are the events with the same state aj , we have E(ek )<S(ew ). And, k>l and ek is not during the period of el , so we have E(ek )>E(el ). Thus, we have S(ew )>E(el ). That is, ew must not occur during the period of el . ¥ 3.2 An example Let us take an example to explain the DTP algorithm. We will execute the algorithm on the temporal database in Table 1 for minimal support count=2. From Figure 1 we know FDTP0 ={a1, a3, a4, a6}. Next, for each aj∈h(ai ), add aj⇒dai into CDTP1. Thus, we get CDTP1={a3⇒da1, a4⇒da1, a6⇒da1, a4⇒da3, a6⇒da3, a6⇒da4}, which corresponds to the sets shown in Figure 8. Subsequently, Step 2 and Step 3 of the DTP algorithm are carried out iteratively. For each aj∈h(β), search the corresponding α and γ mentioned in Property 4. For example, a3∈h(a4 ⇒d a1) and the support degree of a3 ⇒d a1 is not less than the support threshold, so we can join a3 ⇒d a1 and a4 ⇒d a3 if both patterns are frequent, and obtain the new candidata pattern a4⇒da3⇒da1 with the related sets g(a4⇒da3⇒da1) and h(a4⇒da3⇒da1) (as shown in Figure 9 (m)). Similarly, the sets g(a6⇒da4⇒da3⇒da1) and h(a6⇒da4⇒da3⇒da1) are obtained in Figure 10. 15 Support g(a6 ⇒d a1) h(α) 1 (25,26) 2 (37,37) a3,a4 Support g(a4 ⇒d a3) h(α) 1 (25,27) 2 (34,38) a6 (i) generated by (a) and (f) (j) generated by (c) and (d) Support g(a6 ⇒d a3) h(α) 1 (25,26) 2 (37,37) a4 Support g(a6 ⇒d a4) h(α) 1 (25,26) 2 (37,37) ∅ (k) generated by (c) and (f) (l) generated by (d) and (f) Figure 8: The sets of CDTP1 Support g(a4⇒da3⇒da1) h(α) 1 (25,27) 2 (34,38) a6 Support g(a6⇒da3⇒da1) h(α) 1 (25,26) 2 (37,37) a4 (m) generated by (g) and (j) (n) generated by (g) and (k) Support g(a6⇒da4⇒da1) h(α) 1 (25,26) 2 (37,37) a3 Support g(a6⇒da4⇒da3) h(α) 1 (25,26) 2 (37,37) ∅ (0) generated by (h) and (l) (p) generated by (j) and (k) Figure 9: The sets of CDTP2 Support g(a6 ⇒d a4 ⇒d a3 ⇒d a1) h(α) 1 (25,26) 2 (37,37) ∅ (q) generated by (m) and (p) Figure 10: The sets of CDTP3 Lastly, we calculate the confidence degree of the patterns from the bottom of every sublattice. In Figure 5, there are three sublittices. 4 Experiments To assess the relative performance of these two algorithms and study their scale-up properties, we performed several experiments on a computer with 512 RAM and Pentium4 2.6GHz for some synthetic datasets and a real data set with weather information, which was stored on a local 20G disk. 4.1 Generation of synthetic data To evaluate the performance of the algorithms over a large volume of data, we generated synthetic temporal events which mimic the events in the real word. We will show the experimental results from synthetic data so that the work relevant to data cleaning, which is in fact application dependent and 16 also orthogonal to the incremental technique proposed, is hence omitted for simplicity. For obtaining reliable experimental results, the method to generate synthetic data we employed in this study is similar to the ones used in [12]. Table 2 summarizes the meaning of the parameters used in the experiments. The number of input-events in the temporal database relies on |Q| and |T|. That is, the average number of events is |DT|=|Q|*|T|. The starting time and ending time of each event in a during-sequence are generated randomly based on the during relationship. We generated datasets by setting |L|=5, N=50 and P=25. Table 3 summarizes the dataset parameter settings. Table 2: Parameters |Q| Number of during-sequences |T| Average number of events per during-sequences |L| Average length of maximal potentially large patterns N Number of states P Number of maximal potentially large patterns Table 3: Parameters settings (Synthetic datasets) Name |Q| |T| |D| size(MB) Data1-Q10-T5 10000 5 50,000 0.83 Data2-Q10-T10 10000 10 100,000 1.79 Data3-Q20-T5 20000 5 100,000 1.82 Data4-Q20-T10 20000 10 200,000 3.72 Data5-Q30-T10 30000 10 300,000 5.65 Data6-Q40-T10 40000 10 400,000 7.58 Data7-Q50-T10 50000 10 500,000 9.51 4.2 The relative performance with synthetic datasets Figure 11 shows the execution times for the first four synthetic datasets given in Table 3 for decreasing values of minimum support. We did not plot the execution times of the Tree algorithm for some lower support values since they are too large compared to the execution times of DTP algorithm. As the support threshold decreases, the execution times of both the algorithms increase because of the increase in the total number of candidate and large patterns. When the support threshold is higher, there are only a limited number of frequent patterns with length 2 produced. So both algorithms consume less time. However, as the support threshold decreases, the performance difference becomes prominent in that DTP algorithm significantly outperforms the Tree algorithm. 4.3 The reduction of candidate patterns As explained previously, the DTP algorithm substantially reduces the number of candidate patterns generated. The experimental results in Table 4 and Table 5 show the number of candidate patterns of both algorithms for different supports on the two datasets. As shown in Table 5, the DTP algorithm 17 Data1-Q10-T5 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 3 5 0 0 . 2 5 % 0 . 5 0 % 1 % 2 . 5 0 % 5 % 10% 2 0 % Minimum Support T im e (s e c ) Tree DTP Data2-Q10-T10 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 1% 2 . 5 0 % 5% 10% 20% Minimum Support T im e (s e c ) Tree DTP Data3-Q20-T5 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 0 . 5 0 % 0 . 7 5 % 1 % 2 . 5 0 % 5 % 1 0 % 2 0 % Minimum Support T im e (s e c ) Tree DTP Data4-Q20-T10 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 2 . 5 0 % 5 % 1 0 % 2 0 % 3 0 % 4 0 % 5 0 % Minimum Support T im e (s e c ) Tree D T P Figure 11: Execution times leads to a 55%-75% candidate reduction rate compared to the Tree algorithm when Data3-Q20-T5 is used. In another dataset Data4-Q20-T10, the DTP algorithm can achieve higher reduction rate in generating candidate DTPs, such as 200%, as shown in Table 4. Similar phenomena were observed when other datasets were used. This feature of the DTP algorithm could help efficiently reduce the execution time (as mentioned in Section 5.1). Table 4: Reduction on candidate patterns with Data4-Q20-T10 candidate patternsminsupport DTP Tree frequent patterns 40% 34 112 27 30% 167 364 133 20% 552 1202 460 10% 2483 7074 1920 5% 10105 34046 8160 2.50% 35281 142009 29712 18 Table 5: Reduction on candidate patterns with Data3-Q20-T5 candidate patternsminsupport DTP Tree frequent patterns 20% 16 53 11 10% 118 351 68 5% 454 1029 251 2.50% 1223 2827 699 1% 3166 9334 2237 0.75% 4901 12845 3022 0.50% 7202 19090 4482 0.25% 13662 39555 8691 0.10% 30964 113144 21500 4.4 Scale-up Figure 12 shows how DTP algorithm scales up as the number of input-events is increased from 100,000 to 500,000 when the datasets Data2, Data4, Data5, Data6 and Data7 in Table 3 are used. The minimum support level was set to 5%. As shown in Figure 12, the algorithm is approximately linear scalable over the number of input events. Min-Support=5% 0 200 400 600 800 1000 1200 100000 200000 300000 400000 500000 The number of input-events T im e (s e c ) Figure 12: Scale-up 4.5 Experimtents with weather dataset The DTP algorithm has been applied to a weather dataset. The data set was obtained from a weather station in The Netherlands in 2002, which contains weather records with 17 attributes for each hour of the year. The attributes include wind direction, average wind speed, maximum wind gust, average hourly temperature, percentage relative humidity, global hourly radiation, hourly sunshine duration, hourly precipitation duration, hourly precipitation amount,horizontal visibility, fog, snow, etc. Most of the values are continuous. We discretized the data and then converted the database into a temporal one, in a form similar to Table 1. Figure 13 compares the execution time of the two algorithms on the real dataset, with different 19 minimum support degrees. From this figure, we can see that DTP algorithm is more efficient than Tree algorithm, especially at the lower support level. 0 100 200 300 400 500 600 700 800 900 1000 0.10% 0.25% 0.50% 0.75% 1% 2.50% 5% 10% 20% Minimum Support T im e (s e c ) Tree DTP Figure 13: Execution times on the weather dataset 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 0.25% 0.50% 0.75% 1% 2.50% 5% 10% 20% Minimum Support T h e n u m b e r o f v a li d D T P s Figure 14: The number of valid DTPs Figure 14 shows the number of valid DTPs with various support degrees. Some interesting results were generated by applying the algorithm. Firstly, we discovered that there were significant temporal relationships between wind and rain, humidity and radiation, and temperature, radiation and sunshine duration. For example, in terms of wind and rain, strong wind often came prior to rain and lasted until or beyond the end of the rain (rain during strong wind). Another example is that the horizontal visibility was usually weak during the time that fog was heavy or it snowed (weak horizontal visibility during heavy fog or snowing). Secondly, some more complex rules were also discovered. For instance, when the sunshine duration was shortened, the global hourly radiation would decrease and the hori- zontal visibility would become worse. In fact, before sunshine duration reached zero and the night fell, the temperature would have begun to decrease. This rule can be expressed as worse horizontal sight 20 during no global hourly radiation during no sunshine during very lower temperature. 5 Conclusions and future work In this paper, we have studied the problem of discovering during-temporal patterns between events and proposed the DTP algorithm. By analyzing the properties of the during relationship, we have developed an optimization technique with pruning strategies that enabled us to retrieve the patterns with minimal database scan. The experimental results have illustrated the effectiveness and efficiency of the algorithm. An ongoing effort centers on extending this algorithm to discovering the dynamic temporal database, since in the light of the fact that the content of a TDB always keeps growing and the discovered patterns need to be maintained periodically over time. While there are some studies on mining of dynamic database and maintenance of association rules [1,2,4,5,15], our focus is on the maintenance of temporal patterns, so as to reduce the overhead of rediscovering patterns in the presence of data updates. References [1] A. Veloso, W. Meira Jr., M.B.De Carvalho, S.Parthasarathy, and M.Zaki. (2003). Parallel, Incre- mental and Interactive Mining for Frequent Itemsets in Evolving Databases. In Proceedings of the Sixth SIAM Workshop on High Performance Data Mining, May 2003. [2] C. Jin, W. Qian, C. Sha, J. Yu, and A. Zhou. (2003). Dynamically Maintaining Frequent Items over A Data Stream. In Proceedings of the 12th ACM CIKM International Conference on Information and Knowledge Management, pages 287–294, 2003. [3] Chris P.Rainsford, and John F.Roddic. (1999). Adding Temporal Semantics to Association Rules. In Proc. of the 3rd European conf. on principles and practice of knowledge discovery in databases, pages 504-509, 1999. [4] D.W.Cheung, J.Han, V.T.Ng, and C.Y.Wong. (1996). Maintenance of Discovered Association Rules in Large Databases: An Incremental Updating Technique. Proc, Int’l Conf. Data Eng., (ICDE’96)S.Y.W.Su,ed., pp.106-114,1996. [5] D.W.Cheung, S.D.Lee, and B.Kao. (1997). A General Incremental Technique for Maintaining Dis- covered Association Rules. In Proc. Fifth Int’l Conf. Database Systems for Advanced Applications, 1997. 21 [6] F.Hoppner. (2001). Discovery of Temporal Patterns—Learning Rules about the Qualitative Be- haviour of Time Series. In PKDD’01, number 2168 of LNAI, Freiburg, Germany, pages 192-203, 2001. [7] Guoqing Chen, Jiang Ai, Wei Yu. (2002) Discovering Temporal Association Rules for Time-Lag Data. Proceedings of International Conference on E-Business (ICEB2002), Beijing, May 2002. [8] G.Das, K.I.Lin, H. Mannila. (1998). Rule discovery from time series. In Proceedings of the inter- national conference on KDD and Data Mining, 1998. [9] J. Chang and W. Lee. (2003). Finding Recent Frequent Itemsets Adaptively over Online Data Streams. In Proceedings of the 9th ACM SIGKDD International Conference on Knowledge Dis- covery and Data Mining, pages 487–492, 2003. [10] James F.Allen. (1983). Maintaining Knowledge about Temporal Intervals. Communication of the ACM, Volume 26, Number 11, P832-843, November 1983. [11] Mohammed.J.ZAKI. (2000). SPADE: An Efficient Algorithm for Mining Frequent Sequences. Machine Learning, 0, 1-31, 2000. [12] R.Agrawal, and R.Srikant. (1994). Fast algorithm for mining association rules. In Proc. of the 20th Int’l Conference on Very Large Databases, Santiago, Chile, September, 1994. [13] R.Agrawal, and R.Srikant. (1995). Mining Sequential Patterns. International Conference on Data Engineering(ICDE), March, Taiwan, 1995. [14] R.Agrawal, T.Imielinski, and A.Swami. (1993). Mining association rules between sets of items in large databases. In Proceedings of the 1993 ACM SIGMOD International Conference on Manage- ment of Data, pages 207-216, Washington DC, May 26-28, 1993. 10. [15] Y.Chi, H.J.Wang, Philip S.Yu, and Richard R. Muntz. (2004). Catch the Moment Maintaining Closed Frequent Itemsets. http://citeseer.ist.psu.edu. 22 Table 1: A Temporal Database Event State Starting Time Ending Time e1 a1 1 20 e2 a3 1 4 e3 a4 5 7 e4 a1 22 28 e5 a2 2 8 e6 a3 10 13 e7 a5 25 35 e8 a3 23 28 e9 a4 25 27 e10 a6 25 26 e11 a1 30 40 e12 a3 30 38 e13 a4 34 38 e14 a6 37 37 23 Table 2: Parameters |Q| Number of during-sequences |T| Average number of events per during-sequences |L| Average length of maximal potentially large patterns N Number of states P Number of maximal potentially large patterns 24 Table 3: Parameters settings (Synthetic datasets) Name |Q| |T| |D| size(MB) Data1-Q10-T5 10000 5 50,000 0.83 Data2-Q10-T10 10000 10 100,000 1.79 Data3-Q20-T5 20000 5 100,000 1.82 Data4-Q20-T10 20000 10 200,000 3.72 Data5-Q30-T10 30000 10 300,000 5.65 Data6-Q40-T10 40000 10 400,000 7.58 Data7-Q50-T10 50000 10 500,000 9.51 25 Table 4: Reduction on candidate patterns with Data4-Q20-T10 candidate patternsminsupport DTP Tree frequent patterns 40% 34 112 27 30% 167 364 133 20% 552 1202 460 10% 2483 7074 1920 5% 10105 34046 8160 2.50% 35281 142009 29712 26 Table 5: Reduction on candidate patterns with Data3-Q20-T5 candidate patternsminsupport DTP Tree frequent patterns 20% 16 53 11 10% 118 351 68 5% 454 1029 251 2.50% 1223 2827 699 1% 3166 9334 2237 0.75% 4901 12845 3022 0.50% 7202 19090 4482 0.25% 13662 39555 8691 0.10% 30964 113144 21500 27 Support g(a1) h(a1) 1 (1,20) 2 (22,28) a2,a3,a4,a6 3 (30,40) Support g(a2) h(a2) 1 (2,8) a4 Support g(a3) h(a3) 1 (1,4) 2 (10,13) 3 (23,28) a4,a6 4 (30,38) (a) (b) (c) Support g(a4) h(a4) 1 (5,7) 2 (25,27) a6 3 (34,38) Support g(a5) h(a5) 1 (25,35) a4,a6 Support g(a6) h(a6) 1 (25,26) 2 (37,37) ∅ (d) (e) (f) Support g(a3 ⇒d a1) h(α) 1 (1,4) (10,13) 2 (23,28) a4,a6 3 (30,38) Support g(a4 ⇒d a1) h(α) 1 (5,7) 2 (25,27) a3,a6 3 (34,38) (g) (h) Figure 1: The examples of the sets g and h 28 t a3 a2 a2 a1 a3 a2 Figure 2: A counterexample for pattern transitivity 29 a2 a5 a6 a9 a10 a11 a12 a6 a5 a4 a2 a7 a9 a10 a11 a12 a11 a5 a6 a7 a9 a12 a13 Figure 3: trees of frequent DTP1 30 FDTP0.Gen() 1. FDTP0 =∅; 2. for all ai∈A do 3. if |g(ai )|≥ minsupport then 4. FDTP0=FDTP0∪{ai}; 5. end if 6. end for Figure 4: The procedure used for F DT P0 31 a3=> d a1 a4=> d a1 a6=> d a1 a4=> d a3 a6=> d a3 a6=> d a4 a4=> d a3=> d a1 a6=> d a3=> d a1 a6=> d a4=> d a1 a6=> d a4=> d a3 a6=> d a4=> d a3=> d a1 a1 a3 a4 a6 {} Figure 5: The lattice of frequent patterns 32 CDTPk .Gen() 1. for all pattern β∈FDTPk−1 do 2. for all aj∈h(β) do 3. for α∈FDTPk−1,p(α,-1)=aj do 4. if Property 4 is satisfied then 5. CDTPk =CDTPk S{α⇒dp(β,-1)} 6. end if 7. end for 8. end for 9. end for Figure 6: The procedure used for CDTPk 33 FDTPk .Gen() 1. FDTP0 known, FDTPk =∅ (k=1,2,...) 2. for (k=1;FDTPk−1 6=∅;k++) do 3. CDTPk =CDTPk .Gen(FDTPk−1); 4. for all candidate patterns β∈CDTPk do 5. FDTPk ={β∈CDTPk||g(β)|≥minsupport} 6. end for 7.end for 8. FDTP= S k FDTPk Figure 7: The procedure used for FDTPk 34 Support g(a6 ⇒d a1) h(α) 1 (25,26) 2 (37,37) a3,a4 Support g(a4 ⇒d a3) h(α) 1 (25,27) 2 (34,38) a6 (i) generated by (a) and (f) (j) generated by (c) and (d) Support g(a6 ⇒d a3) h(α) 1 (25,26) 2 (37,37) a4 Support g(a6 ⇒d a4) h(α) 1 (25,26) 2 (37,37) ∅ (k) generated by (c) and (f) (l) generated by (d) and (f) Figure 8: The sets of CDTP1 35 Support g(a4⇒da3⇒da1) h(α) 1 (25,27) 2 (34,38) a6 Support g(a6⇒da3⇒da1) h(α) 1 (25,26) 2 (37,37) a4 (m) generated by (g) and (j) (n) generated by (g) and (k) Support g(a6⇒da4⇒da1) h(α) 1 (25,26) 2 (37,37) a3 Support g(a6⇒da4⇒da3) h(α) 1 (25,26) 2 (37,37) ∅ (0) generated by (h) and (l) (p) generated by (j) and (k) Figure 9: The sets of CDTP2 36 Support g(a6 ⇒d a4 ⇒d a3 ⇒d a1) h(α) 1 (25,26) 2 (37,37) ∅ (q) generated by (m) and (p) Figure 10: The sets of CDTP3 37 Data1-Q10-T5 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 3 5 0 0 . 2 5 % 0 . 5 0 % 1 % 2 . 5 0 % 5 % 10% 2 0 % Minimum Support T im e (s e c ) Tree DTP Data2-Q10-T10 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 1% 2 . 5 0 % 5% 10% 20% Minimum Support T im e (s e c ) Tree DTP Data3-Q20-T5 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 0 . 5 0 % 0 . 7 5 % 1 % 2 . 5 0 % 5 % 1 0 % 2 0 % Minimum Support T im e (s e c ) Tree DTP Data4-Q20-T10 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0 2 . 5 0 % 5 % 1 0 % 2 0 % 3 0 % 4 0 % 5 0 % Minimum Support T im e (s e c ) Tree D T P Figure 11: Execution times 38 Min-Support=5% 0 200 400 600 800 1000 1200 100000 200000 300000 400000 500000 The number of input-events T im e (s e c ) Figure 12: Scale-up 39 0 100 200 300 400 500 600 700 800 900 1000 0.10% 0.25% 0.50% 0.75% 1% 2.50% 5% 10% 20% Minimum Support T im e (s e c ) Tree DTP Figure 13: Execution times on the weather dataset 40 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 0.25% 0.50% 0.75% 1% 2.50% 5% 10% 20% Minimum Support T h e n u m b e r o f v a li d D T P s Figure 14: The number of valid DTPs 41 
work_khyn5sbapngmhcwljgay3bweaa	----	Microsoft Word - 05 Flexible Iterative Learning Control_Zuojun Liu.doc International Journal of Fuzzy Logic and Intelligent Systems, vol. 9, no. 3, September 2009 pp. 185-190 185 Flexible Iterative Learning Control Based Expert System and Its Application Liu Zuojun , Liu Zhihu, Zu Linan and Yang Peng Department of Automation, Hebei University of Technology, Tianjin, PR China 300130 Abstract A scheme of expert system based on iterative learning control is proposed. Iterative learning control can obtain control experience from the historical data to build the knowledge base of expert system. Considering some uncertain factors, a flexible measure is adopted in iterative learning control (ILC). Simulation proves the feasibility and effect of the air conditioning control expert system based on flexible iterative learning control (F-ILC). Finally, a feedback compensation unit is incorporated against irregular heavy disturbance. Key Words : Expert System, Iterative Learning Control, Flexible 1. Introduction Expert system is an artificial intelligence application that uses a knowledge base of human expertise for problem solving. Its success is based on the quality of the data and rules obtained from the human expert [1]. It derives its answers by running the knowledge base through an inference engine, which is software that interacts with the user and processes the results from the rules and data in the knowledge base. There are many successful examples of expert system in medical diagnosis, equipment repair, investment analysis, financial, estate and insurance planning, vehicle routing, contract bidding, production control and training[1-3]. In practice, expert systems perform both below and above that of a human, which is mainly due to the quality of human expert knowledge summary. Iterative learning control (ILC)[7-9] refers to the problem of finding an appropriate control input that causes the output to track a prescribed trajectory defined in a finite time interval by iteration trials without knowing the accurate plant model. Learning control applies naturally in processes where the same task is repeated many times, and the system is designed to return to the same initial conditions before commencing the next cycle or trial. For example, the machine resets automatically to a home position as one finished work piece moves out before the next work piece arrives. Robots performing conveying, assembling, welding and spray-painting all have the repetitive nature. Besides electro-mechanical systems, application of ILC can also extend to batch process system, such as the rapid thermal processing, batch reactors, and other batch chemical processes. Standard ILC can only be used in processes where the same desired parameters are repeated turn by turn, which means there should be no difference in each turn. However, in many systems with repetitive nature, the inevitable uncertain disturbances limit the application of ILC. So a Flexible ILC scheme is proposed to extent the application field of ILC. And in this paper, F-ILC is used for building knowledge base of expert system for the control systems with repetitive nature. The rest of the paper is organized as follows. Section 2 reviews standard ILC and presents the scheme of flexible ILC. Section 3 introduces the general conception of expert system. Section 4 presents the simulation results of expert system based on F-ILC used in the application of air conditioning control. Section 5 provides concluding remarks. 2. Flexible ILC 2.1 ILC Iterative learning control uses the historical control date in the past working turns. Like human learning experiences in repetitive trails, ILC obtains higher control precision turn by turn. And the error will decrease in the process of iterative learning. As shown in Fig.1, given a desired output trajectory yd(t) for an operation period [0, T], a control input u(t) is made out through iterative learning to keep the system output y(t) tracing yd(t). [6-9] ILC obtains u(t) in the way of iterative learning function, in which a correction learning function makes out a function array {uk(t)} from each repetitive turns, shown as below. 1( ) ( ) ( ( ), )k k ku t u t U e t t+ = + (1) ( ) ( ) ( )k d ke t y t y t= − (2) where k=0,1,2,… is iterative turns; yk(t) is output of the system; uk(t) is the output of ILC controller. uk(t) obtains the control experiences from the former repetitive turns. And U(ek(t),t) includes the error correction information in the k-th turn. uk+1(t) is so obtaining optimum turn by turn. Most existing ILCs are of either D-type, P-type, I-type or their variations[7-9], described as follows. '1 0( ) ( ) ( ) ( ) ( ) t kk k P k I k Du t u t K e t K e t dt K e t+ = + + +∫ (4) Where KP, KI and KD are constant gain matrixes. Manuscript received Jun. 14. 2009; revised Sep. 6. 2009. This work is supported by fundamental research plan of Tianjin Municipality (07JCYBJC05300) in China. Also, the aid from lab of intelligent robotics in Nanyang Technological University is appreciated. International Journal of Fuzzy Logic and Intelligent Systems, vol. 9, no. 3, September 2009 186 Fig. 1. The iterative learning control 2.2 F-ILC Scheme Flexibility is the ability of a control system to adapt the change in and out of the system. However, standard ILC applies in processes where the same desired parameters are repeated turn by turn, which means there should be no difference in each turn. Although there are repetitive natures in many control systems, the inevitable uncertain disturbances limit the application of ILC. For example, there is regular nature in the repetitive working process day by day in air conditioning system, but it is still not fit for the application of ILC because of some irregular factors, such as the outer temperature, the human flux distribution. The same issue also exists in the control of lower prosthesis, where the steps are basically same turn by turn with slim difference. So the key point is how to make the ILC more flexible to be used in these control system with acceptable error tolerance. So a F-ILC scheme is proposed below. At the beginning of iterative learning, PD-type ILC is used. The D factor may improve the learning rate and make the error convergent rapidly. '1( ) ( ) ( ) ( )kk k P k Du t u t K e t K e t+ = + + (5) Due to the irregular factor of AC system in each turn, the working processes are not strictly same in every repetitive turn. When the error reduces small, as shown in the equation below, the root mean square (rms) error in each day converges into a threshold value ( )k ae t e< (6) The D-type ILC will cause the error divergent even the disturbance is small in magnitude. So if Eq.(6) is met, D factor would not be used any more, and only P-type ILC is used for stable convergence. 1( ) ( ) ( )k k P ku t u t K e t+ = + (7) Standard ILC applies only in processes where the same desired parameters are repeated, including the start and end point, as well as all the working point in every cycle or turn. While some systems do not satisfy these conditions, so when the error converge into another much smaller threshold value, which may also be set though simulation or experiment, as below. ( )k be t e< (8) The system is basically stable and the error is very small and acceptable. The control precision has met with the tolerance of the system, so a control dead-zone is set to ILC. No more iterative learning is necessary, or there may be error divergence. The control experience data may be stored in a knowledge base of expert system. The knowledge base is created according to various typical sorts of working conditions. After the knowledge base is finished, the controller may obtain the corresponding control experience according to the present background conditions sort of the system. As the ILC presented above could be used in the systems not strictly same in every cycle or turn, and it may permit small scope error, so it is named flexible ILC, F-ILC. 3. Expert System An expert system is a computer program that simulates the judgment and behavior of a human or an organization that has expert knowledge and experience in a particular field. Typically, such a system contains a knowledge base containing accumulated experience and a set of rules for applying the knowledge base to each particular situation that is described to the program. A set of “if-then rules” are used to look up corresponding control experience in the knowledge base according to the practical working conditions. Generally, the performing effect of expert systems is mainly determined by the quality of human expert knowledge summary. However, generally, the human experience is not assured as the optima ones, and sometimes it might be improper and low efficient operating experience. Moreover, the process of building a knowledge base usually needs a long period, in which human expert operating skill is accumulated, extracted and summarized as a set of “if-then rules”. So the expert system is not an easy making thing, which greatly limits its application field. In practice, there are many systems working with repetitive and periodic nature. So iterative learning might be used for obtaining the knowledge in an automatic way that can be realized by computer program. As a result, it may make the expert system easy to realize. Also, the improper knowledge and tough process of knowledge accumulation and extraction can be avoided. Flexible Iterative Learning Control Based Expert System and Its Application 187 4. Application of Expert System Based on F-ILC 4.1 Control Plant Analysis Air conditioning system is usually controlled by means of PID control or Bang-bang control, which are all based on feedback[11-13]. Due to the inherit delay and lag of regulation process and actuators, the large range fluctuation and long time transient process of the regulation are inevitable, and accordingly the control effect and energy efficiency are not satisfied. In many cases, there is some kind of regularity in the working process of air conditioner in each day. That is because the load of AC is mainly determined by two factors, the outdoor temperature and the indoor activities of people. And the outdoor temperature runs similar curve each day of the same whether. The temperature is always lower in the early morning, and runs higher till the noon, then falls little in the dusk. Also, there is similar curve in the indoor activities of people each day. For example, the activities of passengers and staffs with their office electrical apparatuses in airport and railway station are associated with the schedules of planes and trains, so the AC load curves caused by it are similar. There are regularities in some other places too, such as hospital, supermarket and cinema. According to the praxeology, the activity of single people is random and disordered; while the activities of swarm people are self-organized and obey some kinds of statistics regularity. Human flux distribution of 4 common sunny workdays in a supermarket is statistically compared in Fig.2. The general statistics regularity does exist to some extend. 0 20 40 60 80 100 120 140 160 180 9 : 0 0 1 0 : 0 0 1 1 : 0 0 1 2 : 0 0 1 3 : 0 0 1 4 : 0 0 1 5 : 0 0 1 6 : 0 0 1 7 : 0 0 1 8 : 0 0 1 9 : 0 0 2 0 : 0 0 2 1 : 0 0 T/h H u m a n f l u x 1st day 2nd day 3rd day 4th day Fig. 2. The statistics regularity of human flux As there is regularity in the working process of AC system in each day, so the historical control data can provide experience for the control in next day, as well as for the building of knowledge base in expert system. However, there still are some irregularities in each day. As it is comparative small to the whole AC load, it may be considered as disturbance to be eliminated by feedback control. Take a supermarket as an example[11]. The regularity nature of AC system makes the energy consumption curve in each day similar, as shown in Fig.3. On one hand, the curves are similar due to the regularity of outdoor temperature and indoor activities of people. On the other hand, there are some differences caused by random factors of indoor activities and outdoor whether. As stated above, the standard ILC does not fit for such kind of systems, while the F-ILC may be used. In Fig.3, the 8 days load curves are chose under same whether and indoor activity conditions. The days in this ILC sample set are all sunny and fuzzy hot whether weekdays. That is to say, the rainy days or weekend days are in other ILC sample sets. And an expert system may be build based on F-ILC. 0 5 10 15 20 0 20 40 t/h T/ ℃ the 1st day 0 5 10 15 20 0 20 40 t/h T/ ℃ the 2nd day 0 5 10 15 20 0 20 40 t/h T/ ℃ the 3rd day 0 5 10 15 20 0 20 40 t/h T/ ℃ the 4th day 0 5 10 15 20 0 20 40 t/h T/ ℃ the 5th day 0 5 10 15 20 0 20 40 t/h T/ ℃ the 6th day 0 5 10 15 20 0 20 40 t/h T/ ℃ the 7th day 0 5 10 15 20 0 20 40 t/h T/ ℃ the 8th day Fig. 3. The energy consumption of air conditioning 4.2 Solution Scheme First of all, divide the working condition of the AC system into several fuzzy sets according to weather forecast, which may include strong hot, hot, weak hot, comfortable, cool, and so on. Next, each fuzzy set is divided into secondary subsets according to the human flux activities, which may include working days, weekend days, big sales days and so on. In another words, it can be processed as a Cartesian product between the weather fuzzy set and human flux category set. Then F-ILC is utilized for each fuzzy subset, and different control knowledge would be obtained to fill out knowledge base of expert system for all kinds of AC working background conditions. Fig.3 shows 8 days AC load curves in the same fuzzy subset. All the curves are generally same, but there are still some small differences with each other days. And F-ILC might be fit for this kind of plant to draw knowledge from it in iterative learning way. After the knowledge base of expert system is finished, the International Journal of Fuzzy Logic and Intelligent Systems, vol. 9, no. 3, September 2009 188 control system might be put into practice. The inference engine derives conclusion from the facts and rules contained in the knowledge base using the condition matching technique. And the knowledge obtained in F-ILC is called by means of “if… then” rules according to the AC working fuzzy subset information via the user interface of expert system. To sum up, the structure of flexible ILC based expert control system is shown as below. temperature Control experience Flexible ILC Air conditioner Plant Physical information feedback Expert system knowledge base b) The building of knowledge base c) The use of knowledge base weather human flux Expert system knowledge base If-then temperatureAir conditioner Plant weather fuzzy set Desired temperature - human flux category ╳ case A:strong hot weekend case B: cool big sales day case C: hot workday …… knowledge item 1 knowledge item 2 knowledge item 3 …… a) Structure of knowlede base If …… then… F-ILC Cartesian product weather human flux control experience Fig. 4. Structure of flexible ILC based expert control system 4.3 Control Effect The input of the AC control system is the error between the desired air temperature and return air temperature. And the output of system is the cooling capacity, which is temperature error between supply and return air multiplied by the air volume. The simulation is run in Matlab. The indoor temperature of AC F-ILC system is shown in Fig.5, in which the control precision is improved day by day based on F-ILC. With the decrease of error, there is no divergence trend which will occur in standard ILC. The error convergence is shown in Fig.6, which demonstrates the decrease of error rms of each day turn by turn. As the F-ILC archives high control precision in AC system, it could obtain better energy saving effect. So the final control parameters can be stored into the knowledge base of expert system. After the whole knowledge base is finished, each piece of knowledge obtained from F-ILC might be active according to the present weather and human flux conditions information via the user interface of expert system. And good control effects might be obtained. 0 10 20 0 20 40 t/h T/ ℃ the 1st day 0 10 20 0 20 40 t/h T/ ℃ the 2nd day 0 10 20 0 20 40 t/h T/ ℃ the 3rd day 0 10 20 0 20 40 t/h T/ ℃ the 4th day 0 10 20 0 20 40 t/h T/ ℃ the 5th day 0 10 20 0 20 40 t/h T/ ℃ the 6th day 0 10 20 0 20 40 t/h T/ ℃ the 7th day 0 10 20 0 20 40 t/h T/ ℃ the 8th day Fig. 5. The curve of energy saving 1 2 3 4 5 6 7 8 9 0 5 10 15 20 25 30 35 Turns/day T/ ℃ Fig. 6. Error curve of F-ILC 4.4 Feedback Compensation against Irregular Disturbance Although the load curves of AC system of each day with same weather and human flux condition are usually similar, some irregular heavy disturbances are inevitable. For example, maybe some day, many teenagers in a nearby school flow into the supermarket because they are all requested to buy a marker by their teachers. As the expert control system proposed in this paper is a kind of open loop control, it could not eliminate the error caused by the irregular heavy disturbances and poor control effect might occur. As a result, pure F-ILC based expert system is not enough for ideal control. Then a closed loop control unit based on real time feedback is incorporated as an essential compensation measure. The feedback compensation unit, a PID controller, will be active when the real time error ei(t) exceeds a predetermined Flexible Iterative Learning Control Based Expert System and Its Application 189 threshold value ec as below. ( )i ce t e> (9) ' 0 ( ) ( ) ( ) ( ) t kfbc P i I i D iu t K e t K e d K e t= + τ τ +∫ (10) The overall control output is composed of both the feedback part and the expert system part. So the irregular heavy disturbance can be eliminated and ideal control effect is obtained. The simulation is shown in Fig.7, in which the dashed line, denoted by y1, is pure expert control system, while the solid line, denoted by y2, is the ones with feedback compensation. 10 40 80 120 160 200 240 24 26 28 30 32 Sample number T /℃ y1 y2 Fig. 7. Feedback compensation against irregular disturbance 5. Conclusion A flexible ILC is proposed to build the knowledge base of expert system for the system with similar repetitive nature. It expends the application field of ILC. As F-ILC may learn the control experience to build the knowledge base in an automatic way, so it avoids the tough process of knowledge summary from experienced operators, as well as some possible wrong knowledge in the brain of operators. Furthermore, a closed loop control unit based on real time feedback is incorporated with expert control system as an essential compensation measure against irregular heavy disturbance. References [1] Tang Jianzhong, Wang Qingfeng, Bi Zhiyue. “Expert system for operation optimization and control of cutter suction dredger”. Expert Systems with Applications. vol. 34, no. 3, pp. 180-2192, 2008. [2] Reaz, M.B.I.; Choong, F.; Sulaiman, M.S.; Mohd-Yasin, F.; Kamada, M.; “Expert system for power quality disturbance classifier”, IEEE Transactions on Power Delivery. vol. 22, no. 3, pp. 1979-1988, 2007. [3] Men-Shen Tsai; “Development of an object-oriented service restoration expert system with load variations” IEEE Transactions on Power Systems, vol. 23, no. 1, pp. 219-225, 2008. [4] Leduc, R.J.; Lawford, M.; Pengcheng Dai; “Hierarchical interface-based supervisory control of a flexible manufacturing system”, IEEE Transactions on Control Systems Technology, vol. 14, no. 4, pp. 654-668, 2006. [5] Qing-lei Hu; Zidong Wang; Huijun Gao; “Sliding mode and shaped input vibration control of flexible systems”. IEEE Transactions on Aerospace and Electronic Systems. vol. 44, no. 2, pp. 503-519, 2008. [6] Mingxuan Sun, Danwei Wang. “Iterative learning control design for uncertain dynamic systems with delayed states”, Dynamics and Control. vol. 10, pp. 341-357, 2000. [7] R. W. Longman. Iterative learning control and repetitive control for engineering practice, International Journal of Control, vol. 73, no. 10, pp. 930-954, 2000. [8] HU Yue, ZHAI Chunyan. “Current status and expectation of iterative learning control”, Process Automation Instrumentation. vol. 26. no. 6, pp. 1-4, 2005 [9] Piao Fengxian, Zhang Qingling, Wang Zhefeng. “Analysis of convergence rate for iterative learning control”, Journal of Northe Astern University(Natural Science). vol. 27, no. 8, pp. 835-839, 2006 [10] Zuojun Liu, Junrui Zhang, Sufang Pang, Peng Yang. “The research and application of control method based on flexible ILC”. Proc. of IEEE WCICA 2008. Chongqing: pp. 5744-5749, 2008. [11] Xu Wangfa, Zhan Xu, “Dynamic simulation of energy consumption for large scale supermarkets”, Heating Ventilating & Air Conditioning. vol. 36, no. 11, pp. 106-109, 2006. [12] Aucevic M. Energy “Efficient office building design”. ASHRAE Journal, vol. 45, no. 1, pp. 26-29, 2005. [13] Liu Z., Qi W. “System integration control of HVAC in intelligent building”. Proc. of ICMLC2004. Shanghai: vol. 2, pp. 1125-1128, 2004. Liu Zuojun He received his B.S. and M.S. degree in the Department of Automation from Hebei University of Technology, Tianjin, China, in 1994 and in 2000 respectively, and the Ph.d degree in Department of Automation from Nankai University, Tianjin, China, in 2005. He is now the head of Department of Automation in Hebei University of Technology. His research interest includes intelligent robotics and intelligent building. International Journal of Fuzzy Logic and Intelligent Systems, vol. 9, no. 3, September 2009 190 Liu Zhihu He received his B.S. degree in the Department of Automation from North China Institute of Science and Technology, Langfang, China, in 2008. He is now a master student in Hebei University of Technology Zu Linan She received his B.S., M.S. and Ph.d degree in the Department of Automation from Jilin University, Changchun, China, in 1999, in 2002 and in 2006 respectively. She is now associated professor of Department of Automation in Hebei University of Technology. Her research interest includes intelligent robotics and control theory. Yang Peng He received his B.S. and Ph.d. degree in the Department of Automation from Hebei University of Technology, Tianjin, China, in 1982 and in 1998 respectively. He is now a professor of Department of Automation in Hebei University of Technology. His research interest includes intelligent robotics and intelligent control theory and application. 
work_km4lywou7zavbdgujqkajdp5ju	----	() Appl. Math. Inf. Sci. 9, No. 3, 1567-1574 (2015) 1567 Applied Mathematics & Information Sciences An International Journal http://dx.doi.org/10.12785/amis/090353 Trust-based Service Recommendation in Social Network Shuiguang Deng1, Longtao Huang1, Yuyu Yin2,∗ and Weitong Tang1 1 College of Computer Science and Technology, Zhejiang University,310027 Hangzhou, China 2 School of Computer, HangZhou DianZi University, 310027 Hangzhou, China Received: 23 Aug. 2014, Revised: 23 Nov. 2014, Accepted: 24 Nov. 2014 Published online: 1 May 2015 Abstract: With the number of Web services increasing constantly on the Internet, how to recommend personalized Web services for users has become more and more important. At present, there emerged some service recommendation systems utilizing influence ranking and collaborative filtering algorithms in service recommendation. However, they neither considered trust relationships among users, nor deal with the cold start problem very well. Fortunately, the popularity of social network in nowadays brings a good alternative for service recommendation to avoid those. In this study, we propose a social network-based service-recommendation method, which considers users’ history service invocation behaviors, users preferences as well as trust relationships among users implied in social network and users comments/reviews on services. We have applied this method in a data set extracted from www.epinions.com. A series of experiments on 86,719 users, 604,190 user trust-relationships and 963,591 reviews on 292,713 services/produces show that this recommendation method get better recall rate, precision, f-measure and rank score. Keywords: Web service, Service recommendation, Social network 1 Introduction Web service has played an important role in the fields of Enterprise Application Integration, E-commerce and Business Process Management [1]. It has achieved a great success both in academia and industry and its number is growing rapidly especially in the age of Cloud Computing and Big Data [2, 3]. This incurs the difficulty for users to find appropriate services from a large number of candidate services to meet their functional and non-functional requirements. Although service discovery method can help users to retrieve target services in some cases, it is a passive process requiring users to know their requirements clearly in advance. However, in many cases, users are neither able to know what they need clearly, nor know the existence of potential services on the Internet. As a result, service recommendation plays a more and more important role to help users out of the service overload predicament, and automatically recommend proper services for users. Currently, recommendation system has been widely used in various fields and its performance is mainly influenced by recommendation algorithms. The classical recommendation algorithms can be divided into two types: Influence ranking [4, 5] and collaborative filtering [6, 7]. Influence ranking recommendation is designed to find the most influential people in a social network. A service recommendation system based on influential ranking aims at recommending the most popular services to active users. Influential ranking mainly includes algorithms such as Reputation, Hits, and PageRank [8]. Collaborative filtering algorithm is more widely used in service recommendation, which is based on users service-invoking history and rating records. It first finds the similar users to the active user; then recommend the services with high evaluation value by the similar users. The algorithms can be subdivided into user-based and item-based collaborative filtering algorithm [9]. Although the current recommendation systems work well, they still have two main defects: (1) they do not consider the trust issue among users while recommending services; (2) they do not deal with the cold start problem very well. In this study, we propose a social-network based service recommendation method based on users trust-relationship and mainly consider: 1) Users preferences. We capture users preferences based on history records which contain the invocation of services, the reviews on services and feedbacks from other users as well; and 2) Trust relationship among users. In general, active user mainly prefers to invoke the services recommended by the person he trusts. The whole ∗ Corresponding author e-mail: yyy718@gmail.com c© 2015 NSP Natural Sciences Publishing Cor. http://dx.doi.org/10.12785/amis/090353 1568 S. Deng et. al. : Trust-based Service Recommendation in Social Network recommendation procedure runs according to the following 4 steps: 1) Building user characteristic vector; 2) Calculating similarity value between active user and other users; 3) Selecting Top-M similar users to the active user and calculating rank score of candidate services; 4) recommending Top-K services to the active user. We have applied this recommendation method in a data set extracted from www.epinions.com. The experiments on 86,719 users, 604,190 user trust-relationships and 963,591 reviews on 292,713 services/produces show that our recommendation method outperforms others. The rest of this paper is structured as follows: Section 2 introduces the background of service recommendation and reviews some related work. Section 3 presents a motivating scenario. Section 4 gives the social-network based service recommendation method in details. Section 5 illustrates the experiments and comparison results. Section 6 concludes this study and discusses the future work. 2 Related Work With more and more Web services emerged on the Internet, it becomes difficult to find a suitable Web service meeting users requirements from massive services. Traditional service discovery approaches are somewhat passive in discovering Web services for users since they highly depend on the clear users request description as well as services functional and non-functional description. Thus, as an important alternative to service discovery, service recommendation attracts more and more attention nowadays. Numerous research works have been done on Web service recommendation [9, 10, 11, 12, 13, 14, 15, 16]. Since our method is based on collaborative filtering, here we only concentrate on introducing related work on collaborative filtering based approaches for Web service recommendation. Collaborative filtering (CF) proposed by Rich has been proved to be a kind of successful approaches for recommendation systems [10]. Shao proposed a user-based CF algorithm using PCC (Pearson Correlation Coefficient) to compute similarity between users [13]. Users who have similar historical QoS experiences on Web services are deemed to be similar. For any active user, the missing QoS values of a Web service can be predicted by considering the corresponding QoS values of services used by his/her similar users. Finally, Web services with high-predicted QoS values are recommended. Zheng proposed a novel hybrid collaborative filtering algorithm for Web service QoS prediction by systematically combining both item-based PCC (IPCC) and user-based PCC (UPCC) [14]. However, they failed to solve the cold start problem. If a user never has QoS experiences on any service, the method cannot recommend any service to him. In order to improve the accuracy of recommendation for Web services, several new enhanced methods are proposed [18, 19, 20] Hao Ma, Haixuan Yang, et al, recognized data sparsity, scalability and prediction quality as the three most crucial challenges that every recommender system based on collaborative filtering algorithm confronts, and proposed a probabilistic matrix factorization method for social recommendation [17]. Chen Xi, Liu Xudong, et al. recognized the influence of user location in Web services QoS prediction and proposed a novel method [18]. The method groups users into a hierarchy of regions according to users locations and their QoS records, so that the users in a region are similar. When identifying similar users for a target user, instead of searching the entire set of users, the method only searches the regions that the target user belongs to. However, it does not consider service location in recommending Web services. Jiang proposed that the influence of personalization of Web service items should be taken into account when computing degree of similarity between users [19]. That is, more popular services or services with more stable QoS from user to user should contribute less to user similarity measurement. Zhang suggested that it was better to combine users QoS experiences, environment factor and user input factor to predict Web services QoS values [20]. But the environment factor and user-input factor can hardly represent users subjective opinions. In general, most of the current Web service recommendation approaches aim to predict the values of QoS attributes of Web services and use historic invocations to extract users interests or preferences, but consider little about the trust issue [21, 22]. As a consequence, they cannot be directly employed in a real Web service recommendation system to recommend services actively without users interference. Compared with prior related works, this paper makes the following major contributions. First, we consider not only QoS of services but also users trust relationships, and present a novel service-rating model. We show that both QoS of services and users trust relationships can be used in improving performance and accuracy of service recommendation significantly. Second, we carry out a series of experiments on a large-scale real dataset extracted from a social-network website www.epinions.com to show that considering users trust can improve the accuracy of recommending personalized services to target user. Furthermore, since the trust relationships between users are considered, the proposed method can partly solve the cold start problem. If a user do not invoked any services before, traditional approaches can hardly recommend services to him/her. But our approach can recommend services by utilizing the trust network of users according to his/her trusted users. c© 2015 NSP Natural Sciences Publishing Cor. Appl. Math. Inf. Sci. 9, No. 3, 1567-1574 (2015) / www.naturalspublishing.com/Journals.asp 1569 3 Motivation Scenario In order to show the importance of the trust factor in service recommendation, consider a recommendation scenario based on Table 1, which shows the records of service invocation by users. The numbers indicate the times of invocation of a service by a user. For example, it shows that Tom has invoked s1 4 times and s2 2 times, yet has not invoked the other three services before. Table 1: Service Invoking History s1 s2 s3 s4 s5 Tom 4 2 Merry 3 3 2 Bob 2 1 Jack 1 1 1 To recommend top-2 services for Tom, most of the current recommendation methods based on collaborative filtering will firstly calculate the similarity between Tom and other 3 users according to their service-invoking behaviors and then rank Merry, Bob and Jack based on the similarity values. Thus, Merry is ranked high among the three users because Merry have invoked the same services with Tom more than Bob and Jack; contrarily, Jack is ranked last because they do not invoke any same services with Tom. That means the preference of Tom is more similar with that of Merry than others. Thus, the services invoked by Merry tend to be the ones Tom will invoke. Accordingly, s4 and s3 are recommended to Tom as the top-2 services. The result is indeed good when we do not consider the relationship among users. However, supposing Jack is Toms friend and they have the same interests, in this case, Tom tend to use the services recommended by Jack more likely than the services recommended by Merry and Bob. Moreover, only considering the trust relationship among users is not enough; the reviews on services given by users will affect the recommendation result as well. Continue to consider the above scenario. In order to recommend Tom top-2 services invoked by Jack, the review or comment on s3 s4 and s5 given by Jack will determine the result. In general, the services with better comment are more likely recommended. With social-network websites attract more and more attention nowadays, people like to share their shopping experiences and their comments. It gives us a good platform to carry out products/services recommendation. In this paper, we propose a social-network based service recommendation method. It not only mines the trust relationships from social networks among users to help recommendation, but also considering users history service-invoking behaviors and users comments/reviews on products/services. 4 Social-network-based Service Recommendation 4.1 Social-network-based Recommendation Framework The framework for the social-network-based service recommendation is illustrated in Fig. 1. Firstly, the active users preference is built according to his service invocation history and trust relationships with other users. Then the similarity value between the active user and each user is calculated. The users with higher similarity value are picked out as the similar users of the active user. Then the rank scores of candidate services are calculated based on the similar users ratings. Finally, services with higher rank scores are recommended to the active user. Fig. 1: Social-network based Recommendation Framework. 4.2 Basic Definitions In order to make the service recommendation method understood easily, we give some notations and definitions. Definition 1 (Historic Service Invocation Records): A service recommendation system is made up of m users and n Web services, represented as U = u1,u2,...,um and S = s1,s2,...,sn, respectively. After invoking a service, the user will rate the service based on the invocation experience. The historic service invocation records of the ui(ui ∈ U) is represented as : IH i = (si1,category j,ratei1),...,(sin,category j,ratein) where: (1) si1 means user ui invoked service s1, (2) category j represents the category that the service belongs to; (3) ratei1 represents the rating value given by the user ui to the service s1. (4) the 3-tuple (si1,category j,ratei1) represents a service-invoking record which means user ui invoked service s1 and rated s1 as ratei1. c© 2015 NSP Natural Sciences Publishing Cor. www.naturalspublishing.com/Journals.asp 1570 S. Deng et. al. : Trust-based Service Recommendation in Social Network Definition 2 (Trust Relationship): The trust relationship between users can be divided into two types: 1) explicit trust relationship, which means users specify whom to trust definitely, the form goes like ’u1 trusts u2’; 2) implicit trust relationship, which means we can obtain the implicit trust relation between users through mining the association activities among users. Given two users u1 and u2, u1 trusts u2 if u1 has paid attention to u2s reviews on services and thinks they are helpful. For example, user posts a review targeted on the service which he/she invokes, then, other users publish feedbacks which express the idea of whether the review is helpful to them or not, at last, we can mine the implicit trust relation between feedback users and the user who posts the review. The explicit and implicit trust relationship constitutes complete trust relationships between users. Definition 3 (User Characteristics Vector): The tuple CVi =< category1,...,categoryp > constituted by each service category forms the prototype of user preference where p is the number of categories. 4.3 Key Steps 4.3.1 Building Users Preference According to Definition 3, for the initialization of a user characteristic vector, each service category is initialized as 0. Given a user ui, the building of ui’s user characteristic vector mainly includes the following steps: Firstly, traverse all the services from uis historic service invocation records IHi and extract the service categories which the services belong to, then plus 1 on the corresponding service category of the user characteristic vector. Besides, the user characteristic vectors of the users who ui trusts are merged into uis own user characteristic vector. The merging rule is: CV i = CV i + µ ∑ j∈Tui CV j, (1) where µ is a balancing factor (0 < µ ≤ 1), which used to adjust the impact on users characteristic vector and determined by experiment later. Tui represents the user set which user ui trusts. 4.3.2 Calculating Similarity Between Users The step of calculating similarity value between active and other users is the core step for the proposed recommendation framework. The Euclid Distance is adopted to calculate the similarity value, the formula is as follows. sim(ui,u j) = √ p ∑ k=1 (CV ik− CV jk) 2 , (2) (1) the Euclidean Distance also known as a Euclidean metric, is a commonly used distance definition. It is the true distance between the points in the m-dimensional space. (2) ui represents the active user and u j represents the trust user (u j ∈ Tui ,Tui represents the user set which user ui trusts). (3) the smaller the value of sim(ui,u j) is, the more similar of user ui and u j are. For example, given two users u1 and u2 whose user characteristic vectors are < 5,1,2 > and < 1,0,1 > respectively, then the similarity value between u1 and u2 is 4.242. 4.3.3 Calculate Rank Score of Candidate Services After calculating the similarity values of the active user and other users. We choose a number of similar users to calculate the rank score of the candidate services. The number of similar users is an adjusted parameter for our framework and in our experiment it is set as 100. The rank score of service is a comprehensive evaluation of the active user’s personal preference and the relationship of trust between active and other users. Calculation of the rank score mainly consists the following steps: Firstly, traversing each service of each similar user’s history service-invoking record and picking out the service which active user didn’t have invoked. Secondly, calculating the rank score of the service according to the following formula. rssk = ∑ j∈set sk (sim(ui, u j) × rate jsk ), (3) Where sk ∈ ( ⋃ j∈Tui IH j − IH i), Tui represents the user set which active user ui trusts, setsk is consisted of the user who has invoked service sk, rate jsk is a rating given by user u j to the service sk . Finally, these candidate services are classified according to the service category. After calculating the rank scores of the candidate services, it is intuitive to recommend the top-k services with higher scores in each category to the active user. 5 Experiment and Evaluation In order to verify the effectiveness of the proposed social- network-based service recommendation method, we carry out a series of experiments to measure 3 different metrics, i.e., recall rate, precision ,f-measure and rank score; and compare with other 5 popular recommendation methods, i.e., UserCF, ItemCF, Hits, Reputation and Pagerank. c© 2015 NSP Natural Sciences Publishing Cor. Appl. Math. Inf. Sci. 9, No. 3, 1567-1574 (2015) / www.naturalspublishing.com/Journals.asp 1571 5.1 Datasets Description In this paper, we select www.epinions.com as data source, which is operated by Shopping.com, Inc., a leading provider of comparison shopping services. Epinions is a premier consumer reviews platform on the Internet and provides a reliable source for valuable consumer insight, unbiased advice, in-depth product evaluations and personalized recommendations. We crawled two sets of data from this website, including user trust relationship and user reviews and feedback on reviews. a) User Trust Relationship: A user is allowed to build his personal trust network on the basis of specifying whom he trusts. Web of trust is a network of reviewers whose reviews and ratings a user has consistently found to be valuable. According to web of trust, the system predicts how helpful a review to a user and promotes the reviews of trusted members. So, a user can find what he/she is looking for more easily and gets the most out of his/her time on the website. We acquire 86,719 nodes and 604,190 edges from Epinions using a crawler. The nodes and the edges represent users and trust relationships between users, respectively. b) User Reviews and Feedback: Users are allowed to post reviews on a service/product he bought or used before. In addition, Epinions allows users to post feedback on these comments, indicating how valuable the reviews to them. The rating on a comment is divided into four levels shown in Table 2. Table 2: Feedback grade and value No Feedback Grade Grade Value 1 Very Helpful 3 2 Helpful 2 3 Somewhat Helpful 1 4 Not Helpful 0 Table 3 shows an example of user reviews. Userid represents different users and each line records one review posted by a user. We grabs 963,591 reviews which target on 292,713 different products by 57,301 users from this platform, about 17 reviews per each user. An example of feedback is illustrated in Table 4. Each line represents one feedback on a specific review (identified by the review title). At present, we crawl 23,142,103 feedbacks from Epinions and find that there are about 24 feedbacks for each review. 5.2 Measure Metrics A recommendation system is launched to generate a ranked list of recommendations of Top-K products. In general, the recommendation quality is measured based on the number of hits (recommendations that match the products actually purchased by the consumer as recorded in the testing set) and their positions in the ranked list, as shown in Table 5. The following recommendation-quality metrics regarding relevance, coverage, the combination of relevance and coverage as well as the ranking quality of the ranked list recommendation [20] are used in our experiments. Table 5: User feedback example Recommended Not Recommended Purchased True-Hits(TH) False-UnHits(FUH) Not Purchased False-Hits(FH) True-UnHits(TUH) 5.2.1 Recall Rate Recall rate refers to the ratio of correctly predicted (True-Hits) products to the number of all products purchased (True-Hits and False-UnHits) by the consumer in the testing set [23]. RR = #T H #T H +#FU H , (4) 5.2.2 Precision Recommendation precision refers to a ratio of correctly predicted (True-Hits) products to the number of all products predicted (True-Hits and False-Hits) [23]. RP = #T H #T H +#F H , (5) 5.2.3 F-Measure The F-Measure is a measure of a statistic experiment’s accuracy. It considers both recall and precision measures of the experiment to compute the score. We could interpret it as a weighted average of the recall and precision, where the best f-measure score has its value at 1 and worst score at the value 0 [23]. F − Measure = 2 × RR × RP RR + RP , (6) 5.2.4 Rank Score The recall, precision and f-measure are standard performance measures to argue the relevance and coverage of the recommended items relative to the consumers’ potential purchases. The f-measure is the harmonic mean of recall rate and precision. Because c© 2015 NSP Natural Sciences Publishing Cor. www.naturalspublishing.com/Journals.asp 1572 S. Deng et. al. : Trust-based Service Recommendation in Social Network Table 3: User review example Userid Title Product Category Rating Review Date 007girl Blue my mind! Here on earth Movies 1 2000-05-07 007girl Chicks dig maxim too Maxim Magazine Subscription Magazines & Newspapers 5 2000-10-22 00focus Ford Focus ZX3 2000 Ford Focus Cars & Motorsports 4 2000-07-09 00Shoe All’s Well but the hearing Sony Personal CD Player Electronics 3 2003-04-16 Table 4: User feedback example Userid Title Rating Feedback Date oofocus One of the most watched networks at my house 3 2001-01-25 01400ex Don’t let power rating deceive you 3 2001-12-25 092566 Chin-Chin is Yum-Yum 3 2001-07-28 1-2many A fun little Lefty treat 3 2006-02-22 recall and precision are potentially competing measures, the f-measure provides a single-index performance measure to balance them. The rank score measure [20] has adopted in many studies [24, 25] to evaluate the ranking quality of the recommendation list. Suppose the two recommendation lists with the same set of correct recommendations (matched with future purchases) for a target consumer. If the index of each of the correct recommendations in the first list is smaller than (closer to the top of the list by one position) that of correct recommendation in the second list, although the two lists would achieve the same recall, precision and f-measure, the first list is better than the second as the recommended items in the first list are ranked higher than those in the second list. In this case, the rank score of the first list is larger than that of the second and its value can be calculated according to the following formula. RS = 100 × ∑c RSc ∫ c RS max c (RSc = ∑ j qc j 2( j−1)/(h−1) ), (7) Where j is the index for the ranked list; h is the viewing half-life and qc j = { 1 i f j is in c′s testing set 0 otherwise (8) Where RSc max is the maximum achievable rank score for consumer c if all future purchases had been at the top of a ranked list [26]. 5.3 Comparison with other algorithms We divide user review and feedback datasets into training set and validation set according to a time-stamp. The choice of time-stamp based on the principle that the size of training set equals to the size of validation set. 5.3.1 Impact of Parameter µ In our experiement, the balancing factor µ is ranged from 0 to 1 with the interval 0.2. We set Top-K to be 20 and the number of most similar users to be 100. Fig. 2 illustrates the values of rank score, recall, priecision and f-measure. It shows that the value of recall, precision and f-measure increases as µ increases, and the value of rank score is the smallest when µ is 1. All of the four measure metrics increase as balancing factor becomes larger and arrives at the biggest value at the same time when µ goes up to 1. It indicates user’s own history service-invoking records and trust users’ history service-invoking records are equally important when building user characteristic vector. So, we assign µ = 1 in the following experiments. Fig. 2: Impact of parameter µ . 5.3.2 Comparisons with other algorithms In order to show the advantages of the social-network-based recommendation method, we compare it to the other five popular methods such as UserCF, ItemCF, Hits, Reputation and PageRank. We set top-K to be 10, 20 and 30. Fig. 3(a)-4(c) illustrates the results of the experiments. It indicates that the proposed social-network-based c© 2015 NSP Natural Sciences Publishing Cor. Appl. Math. Inf. Sci. 9, No. 3, 1567-1574 (2015) / www.naturalspublishing.com/Journals.asp 1573 recommendation method outperforms other in all the first four metrics. Fig. 3(d) shows the services recommended by the proposed model tends to be firstly invoked by the active user after time-stamp than others. (a) Recommendation recall rate (b) Recommendation precision (c) Recommendation f-measure (d) Recommendation rank score Fig. 3: Comparisons with different algorithms. 6 Conclusion and Future Work In this paper, we propose a social-network-based service recommendation method, which both considers users’ history service invocation behaviors, users preferences as well as trust relationships among users implied in social network and users comments/reviews on services. We apply this method in a real large-scale data set extracted from a popular social network www.epinions.com. A series of experiments show that this recommendation method gets better recall rate, precision, f-measure and rank score. Our future work focuses on the following two points: 1) distinguishing explicit and implicit trust relationship between active user and other users; 2) considering the dynamic update of user’s characteristic vector. Acknowledgement This research was supported by the National Key Technology R&D Program (No. 2013BAD19B10), National Natural Science Foundation of China under Grant (No. 61170033). References [1] Shuiguang Deng, Longtao Huang, Wei Tan, Zhaohui Wu. Top-k Automatic Service Composition: A Parallel Framework for Large-Scale Service Sets. IEEE Transactions on Automation Science and Engineering. 11(3):891-905, 2014. [2] X Li, F Zhou, X Yang, A multi-dimensional trust evaluation model for large-scale P2P computing, Journal of Parallel and Distributed Computing, 71, 837-847, Elsevier (2011). [3] Shuiguang Deng, Longtao Huang, Ying Li, Jianwei Yin. Deploying Data-Intensive Service Composition with a Negative Selection Algorithm. International Journal of Web Service Research. 11(1):75-92, 2014. [4] H. Kwak, C. Lee, H. Park, S. Moon, What is Twitter, a social network or a news media? in: Proceedings of 19th International World Wide Web (WWW) Conference, April 26C30, Raleigh NC (USA), 2010, pp. 591C600. [5] Z. Wang, Y. Tan, M. Zhang, Graph-based recommendation on social networks, in: Proceedings of International AsiaCPacific Web Conference, 2010, pp.116C122. [6] J. Konstan, D.B. Miller, D. Maltz, J. Herlocker, L. Gordon, J. Riedl, GroupLens: applying collaborative filtering to usenet news, Communications of ACM 40 (3) (1997) 77C87. [7] B. Sarwar, G. Karypis, J. Konstan, J. Riedl, Item- based collaborative filtering recommendation algorithm, in: Proceedings of the 10th International Conference on World Wide Web, 2001, pp. 285C295. [8] L. Page, S. Brin, R. Motwani, T. Winograd, The Pagerank citation ranking: bringing order to the web, Technical report 1999-66, Stanford InfoLab, November 1999. c© 2015 NSP Natural Sciences Publishing Cor. www.naturalspublishing.com/Journals.asp 1574 S. Deng et. al. : Trust-based Service Recommendation in Social Network [9] Shuiguang Deng, Longtao Huang, Guandong Xu. Social Network-Based Service Recommendation with Trust Enhancement. Expert Systems with Applications. 41(18):8075-8084, 2014 [10] Mehrbakhsh Nilashi, Othman Ibrahim, Norafida Ithnin: Hybrid recommendation approaches for multi-criteria collaborative filtering. Expert Systems with Applications. 2014,41(8):3879-3900 [11] E.Rich, ”User modeling via stereotypes,” Cognitive Science, Vol.3, No.4, 1979. [12] Mehrbakhsh Nilashi, Othman Ibrahim, Norafida Ithnin: Hybrid recommendation approaches for multi-criteria collaborative filtering. Expert Systems with Applications. 2014,41(8):3879-3900 [13] L. Shao, J. Zhang, Y. Wei, J. Zhao, B. Xie, and H. Mei, Personalized QoS prediction for Web services via collaborative filtering, in Proceedings of the 5th International Conference on Web Services (ICWS 2007), 2007, pp. 439- 446. [14] Z. B. Zheng, H. Ma, M.R. Lyu, and I. King, WSRec: a collaborative filtering based Web service recommendation system, in Proceedings of the 7th International Conference on Web Services (ICWS 2009), 2009, pp.437-444. [15] X. Y. Su, T. M. Khoshgoftaar, A Survey of Collaborative Filtering Techniques, Advances in Artificial Intelligence (ADVAI) 2009, 2009. [16] J. Breese, D. Heckerman, and C. Kadie, Empirical analysis of predictive algorithms for collaborative filtering, in Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence (UAI 98), 1998. [17] Hao Ma, Haixuan Yang, Michael R. Lyu and Irwin King, ”SoRec: social recommendation using probabilistic matrix factorization,” in Proc. of the 17th ACM Conf. on Information and Knowledge Management, pp. 931-940(2008). [18] Xi Chen, Xudong Liu, Zicheng Huang, and Hailong Sun, ”RegionKNN: A Scalable Hybrid Collaborative Filtering Algorithm for Personalized Web Service Recommendation,” Web Services (ICWS), 2010 IEEE International Conference on , vol., no., pp.9,16, 5-10 July 2010. [19] Yechun Jiang; Jianxun Liu; Mingdong Tang; Xiaoqing Liu, ”An Effective Web Service Recommendation Method Based on Personalized Collaborative Filtering,” Web Services (ICWS), 2011 IEEE International Conference on , vol., no., pp.211,218, 4-9 July 2011. [20] Qiong Zhang; Chen Ding; Chi-Hung Chi, ”Collaborative Filtering Based Service Ranking Using Invocation Histories,” Web Services (ICWS), 2011 IEEE International Conference on , vol., no., pp.195,202, 4-9 July 2011. [21] Shuiguang Deng, Bin Wu, Zhaohui Wu. Efficient Planning for Top-K Web Service Composition. Knowledge and Information Systems. 36(3): 579-605, 2013. [22] Shuiguang Deng, Longtao Huang, Bin Wu, Lirong Xiong. Parallel Optimization for Data-intensive Service Composition. Journal of Internet Technology.14(6): 817-824, 2013. [23] Nuno Antunes, Marco Vieira, ”Benchmarking Vulnerability Detection Tools for Web Services,” in Proceeding of International Conference on Web Service (ICWS 2010), 2010, pp.203-210. [24] Deshpande, M., G. Karypis. 2004. Item-based top-N recommendation algorithms. ACM Trans. Inform. Systems 22(1) 143C177. [25] Shuiguang Deng, Longtao Huang, Jian Wu, Ying Li, Jianwei Yin. Trust-based Personalized Service Recommendation: A Network Perspective. Journal of Computer Science and Technology,V29(1): 69-80, 2014. [26] Zan Huang, Daniel D.Zeng and Hsinchun Chen, ”Analyzing Consumer-Product Graphs: Empirical Findings and Applications in Recommender Systems,” in Journal of Management Science, 2007, pp.1146-1164. Shuiguang Deng received the Doctors degree in computer science from Zhejiang University, Hangzhou,China, in 2007. He is currently an associate professor in Zhejiang University. His research interests include service computing, cloud computing and middleware techniques. Longtao Huang received the Bachelors degree in software engineering from Zhejiang University, Hangzhou, China, in 2010. Now study the PhD of computer science and technology in Zhejiang University from 2010. His research interests include service computing and cloud computing. Yuyu Yin received the Doctors degree in computer science from Zhejiang University, Hangzhou, China, in 2010. He is currently an assistant professor in Hangzhou Dianzi University. His research interests include service computing, cloud computing and middleware techniques. Weitong Tang received the Master degree in computer science from Zhejiang University, Hangzhou, China, in 2014.His research interests focus on recommender systems. c© 2015 NSP Natural Sciences Publishing Cor. Introduction Related Work Motivation Scenario Social-network-based Service Recommendation Experiment and Evaluation Conclusion and Future Work 
work_kmeoqp42dvehznrq6273jg46be	----	International Journal of Engineering & Advanced Technology (IJEAT) International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-3, February, 2020 592 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: C5150029320/2020©BEIESP DOI: 10.35940/ijeat.C5150.029320  Abstract: Robotic planning to find the target our goal point/s is most important subject with the minimum distance and the fastest speed with obstacle avoidance expert system has been proposed. In this paper we try to compare and consider different scenario by taking two or more moving robot figure out the short path from the initial and the final point automatically through the map of many regular and irregular obstacles. Firstly, the adaptive fuzzy expert system is present where the fuzzy rule has been adaptive recursively through the robot moving, and then the potential field algorithm has been compared with the adaptive fuzzy system, the results demonstrated that the adaptive fuzzy is faster than the potential field but the accuracy moving of the potential field robotic path planning is much better. All the algorithms were failed when two robots moving from two different initial points to one final target point the why we have proposed particle swarm optimization (PSO) algorithm to solve such problem. Keywords : robot, path planning, optimization, fuzzy, PSO. I. INTRODUCTION At present portable robots have been effectively utilized in different engineering of designing for example, aviation examine, atomic research, generation designing and so forth. The significant target in the flow automated research field is to discover an impact freeway from a given beginning situation to predefined goal point. When all is said in done way arranging calculations are delegated neighborhood and worldwide relying on the encompassing condition. In worldwide way arranging the encompassing condition is totally known to the portable robot so the way went by the versatile robot is predefined, whereas in neighborhood way arranging the earth is totally obscure or halfway known to the versatile robot so different sensors are used to see the data about the encompassing condition and plan the movement in like manner. Numerous efforts have been paid in the past to improve different robot route calculations. In state of the art, there can be discovered a few scientists have been tended to on numerous canny strategies for portable robot route. Numerous creators have thought about a controller with complete data of the earth [1-2].Due to the unpredictability and vulnerability of the way arranging issue, old style way arranging strategies, for example, Visibility Graph [3], Voronoi outlines [4], Matrices [5], Cell Revised Manuscript Received on February 05, 2020. * Correspondence Author Baidaa M. Madlol*, Communication Technical Engineering Department ,Al Najaf Technical Engineering College, Al Furat Al Awast Technical University (ATU), Al Najaf(54001), Iraq,. Email:baidaa.alkhlidi@yahoo.com. Ahmad T. Abdulsadda, Communication Technical Engineering Department , Al Najaf Technical Engineering College, Al Furat Al Awast Technical University (ATU), Al Najaf(54001), Iraq,. Email:coj.abdulsad@atu.edu.iq. Ali A. Abdullah Albakry, Communication Technical Engineering Department ,Al Najaf Technical Engineering College, Al Furat Al Awast Technical University (ATU), Al Najaf, Iraq,. Email:aaali@yahoo.com. disintegration [6], counterfeit potential field [7], Rule based strategies [8], and Rules learning methods [9] are not fitting for way arranging in powerful conditions. The utilization of the above calculations for way finding for portable robot requires additional time and the finding of this way won't totally doable for constant development. There are numerous fluffy rationale techniques utilizing different executions or in blend with different systems [10-14]. In worldwide way arranging issue the earth will be acquainted with the robot while in nearby way arranging the robot is oblivious of the encompassing. For taking care of the advancement issue a socially roused famous meta-heuristic improvement calculation is utilized; which is known as Particle Swarm Optimization (PSO) method. PSO was motivated by the social conduct of feathered creatures and fish. Molecule Swarm Optimization (PSO) was created by Eberhart and Kennedy [15-20], propelled by social conduct of winged animal running or fish tutoring. Numerous past specialists have been utilized this calculation to comprehend way anticipating a portable robot. In [18, 21], creators have thought about a lot of versatile robots as swarm and they actualized the multi-robot-multi-target search with same wellness work in the work space. This article manages age of ideal impact free directions of different portable robots from their beginning to goal position with both static and dynamic obstructions. To achieve this ideal errand another wellness work is proposed in this paper which considers parameters like separation between obstruction – robot, robot-objective and robot-robot. Reenactment results are given to check the effectiveness of the created calculation. The rest of the paper is organized as: section II presents a mathematical model for the adaptive fuzzy, potential filed, and finally the adaptive PSO algorithms, respectively. Section III presents the simulation setup and the algorithms flowcharts. The simulation rsults would be discussed in section IV, finally the discussion and future works would be briefly discussed in section V. II. MATHEMATICAL FORMULATION A. Adaptive fuzzy expert system The perusers ought to please acquaint themselves with Fuzzy Logic before perusing this instructional exercise. Fluffy based route is a responsive arranging method, where the quick position and good Expert system for Robotic Path Planning Baidaa M. Madlol, Ahmad T. Abdulsadda, Ali A. Al Bakry Expert system for Robotic Path Planning 593 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: C5150029320/2020©BEIESP DOI: 10.35940/ijeat.C5150.029320 ways from impediments is considered to figure the prompt move, absent really any making a big deal about what's to come. In such a way prompt activities lead to movement of the robot, at last prompting the objective. So as to tackle the issue utilizing fluffy rationale, we first need to choose a couple of information sources which best speak to the circumstance that the robot is right now put in. The choice of movement is made simply based on these sources of info and not the genuine situation. For this issue 6 data sources are chosen. These are good ways from the snag in front, good ways from the deterrent at the front left inclining, good ways from the obstruction at the front right corner to corner, point between the heading course of robot and the objective, separation of the robot from the objective and favored turn. The various information sources are condensed in Figure 1. Fig.1: Adaptive fuzzy rules robotic path planning, adaptive, [3]. The last info, favored turn demonstrates whether it is helpful to turn clockwise or hostile to clockwise, all different sources of info disregarded. A straightforward guideline is utilized to set the parameter. In the event that the front impediment is far away, turn is in order to more face the objective. In the event that the front impediment is close and another front hindrance is experienced, turn utilizing the side of the objective is liked. On the off chance that the front hindrance is close and a similar impediment as experienced in the past advance is discovered, a similar turn as made beforehand is rehashed. The fluffy framework delivers a solitary yield, which is the controlling to make or the quick precise speed. The fluffy principles are composed to such an extent that the robot keeps away from the impediments and adjusts itself towards the objective. The fluffy framework is an aftereffect of a great deal of manual tuning of the guidelines and participation capacities, over a wide assortment of situations. Figure 2 is shown the an example of the three fuzzy rule for the one front obstacle map. Fig. 2: Three fuzzy front obstacle. B. Potential field expert system The reader ought to please acquaint themselves with Artificial Potential Fields before perusing this instructional exercise. Fake Potential Field based route is a responsive arranging procedure, where the prompt good ways from impediments are considered to register the quick move, absent really any fretting over what's to come. In such a way quick activities lead to movement of the robot, at last prompting the objective. All hindrances repulse the robot with a greatness conversely relative to the separation. The objective draws in the robot. The resultant potential, representing the appealing and terrible segments is estimated and used to move the robot. The potential field for an example situation is appeared in Figure 2. Bearings demonstrate the heading of the potential vector. The separation of the obstructions at all edges from the robot is estimated. In this instructional exercise just 5 separations at explicit points are estimated to figure the horrendous potential. These are forward, left side, right side, forward left slanting and forward right corner to corner. The various data sources are outlined in Figure 3. Fig. 3: Potential field path planning. C. Adaptive particle swarm optimization expert system As referenced before, Particle Swarm Optimization procedure is enlivened by the social conduct of feathered creatures and fishes. On account of fledgling running, when winged creatures move in bunch then each fowl in the herd is having its own position and speed. This is only the individual best situation of each winged creature. The best situation in the herd is measured dependent on an information parameter known as target capacity or wellness work. If there should arise an occurrence of feathered creature running, the wellness work parameter is scanning for nourishment in less good ways from the haven of winged animals. So each feathered creature is having some close to home best position dependent on this wellness work. The situation of the winged creature which has ideal wellness esteem is considered as the worldwide best position. As the winged animals share data among them, all fowls will in general move towards the worldwide best position. As PSO is propelled by the social idea of winged creature running, each fledgling can be considered as one molecule. So every molecule has its very own position and speed. The International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-3, February, 2020 594 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: C5150029320/2020©BEIESP DOI: 10.35940/ijeat.C5150.029320 situation of molecule which fulfill ideal wellness esteem measure is picked as the worldwide best position. Figure.4 beneath shows the structure of PSO. Fig. 4: PSO path planning,[18]. Consider at first, the robot Ri is set in the area at ( xi curr ,yi curr ). We need to locate the following area of the robot ( xi next , yi next )by joining of the two focuses between {( xi curr , yi curr ),( xi next , yi next )} and{( xi next , yi next ); ( xi goal , yi goal )} ought not contact the deterrent on the planet map, as spoke to in Figure 5, and limits the all out way length from current situation to an objective situation without contacting the obstruction by framing limitation. At that point the target work F1, which decides the direction way length for n, number of robots, is (1) Fig.5: Schematic diagram for robotic path planning,[12]. III. SIMULATION SETUP AND ALGORITHMS The main steps for the all expert systems can be defined by the following procedure : 1. Characterize robot and objective position. Characterize greatest number of cycles. 2. Move the robot from its source to goal by augmenting the estimation of level and vertical segment of the position vector. 3. In the event that there is no hindrance the robot will move to goal toward incline among source and goal since slant is the most limited way between any two focuses. 4. On the off chance that the robot detects any obstruction before it creates irregular particles and dependent on least wellness esteem criteria the worldwide best position is found. 5. The robot will move to these worldwide best situations until it arrives at the goal. At the point when it will arrive at the goal the emphasis stops. Firstly we have to define the mapping for the robot, this can be done by using the paint program to plot the obstacles and read the map directly from the robot sensors, or it can be generated by the matlab platform directly. Figure 6 is shown one example for the path robotic planning generated by the paint program. IV. Fig. 6: path planning map generated by paint program. V. SIMULATION RESULTS A. Adaptive fuzzy path planning So as to execute the code you have to execute the record 'fuzzy.m'. First make another bitmap record and draw any guide on it, over which you have to execute the calculation. Paint or some other straightforward drawing instrument can be utilized. Ensure you spare the document as a BMP. Spot the record in the undertaking envelope. You may want to open any of the provided maps and re-draw them. Change the name of the guide document in the code to call attention to the guide that you made. Supply the source and the objective positions and the underlying heading course (in radians). You can utilize paint or some other drawing utility which shows the pixel places of the focuses, to find the source and objective on your drawn guide. Robot static and kinematic parameters might be changed to suit the necessities. Anyway this will change the perfect fluffy surmising framework. Changes to the principles and the enrollment capacities can be conveyed at the fluffy editorial manager. Execute the calculation. You should take screen captures of the presentation, the instrument doesn't do that for you. The reenactment will stop when the robot arrives at its objective, impacts or is compelled to stop according to the wellbeing safeguards. The way length and execution time are given at the support. Figure 7 shown the final path find for the adaptive fuzzy expert system. Expert system for Robotic Path Planning 595 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: C5150029320/2020©BEIESP DOI: 10.35940/ijeat.C5150.029320 Fig. 7: robotic moving through the obstacles using adaptive fuzzy expert system. B. Potential field algorithm So as to execute the code you have to execute the record 'potential.m'. First make another bitmap record and draw any guide on it, over which you have to execute the calculation. Paint or some other straightforward drawing instrument can be utilized. Ensure you spare the document as a BMP. Spot the record in the undertaking envelope. You may want to open any of the provided maps and re-draw them. Change the name of the guide document in the code to call attention to the guide that you made. Supply the source and the objective positions and the underlying heading course (in radians). You can utilize paint or some other drawing utility which shows the pixel places of the focuses, to find the source and objective on your drawn guide. Robot static and kinematic parameters might be changed to suit the necessities. Anyway this will change the perfect fluffy surmising framework. Changes to the principles and the enrollment capacities can be conveyed at the fluffy editorial manager. Execute the calculation. You should take screen captures of the presentation, the instrument doesn't do that for you. The reenactment will stop when the robot arrives at its objective, impacts or is compelled to stop according to the wellbeing safeguards. Figure 8 shown the fail reach robotic moving using the potential field algorithm. Fig.8: Potential field doesn’t reach the goal point. C. PSO algorithm The recreation of robot way direction is plotted utilizing MATLAB 2013b. For producing diverse impact free directions the tuning parameter ȥ is tuned appropriately. The recreation is performed by picking c1=1 and c2=1.However various qualities for c1 and c2 can be picked. The estimation of r1 and r2 are picked as one for effortlessness. In the wake of tuning the parameter, it is discovered that the parameter at ȥ =0.36 gives better outcomes. The reproduction result approving the above articulations demonstrated as follows as shown in figure 9: Fig. 10: Robotic path planning using PSO. Generally, from the above results the adaptive fuzzy and the PSO is reach the target point through the obstacles, while the potential algorithms hit the obstacle and stop the program. Table I listed the numerical results of the three schemes. International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-3, February, 2020 596 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: C5150029320/2020©BEIESP DOI: 10.35940/ijeat.C5150.029320 Table- I: Performance time and accuracy. No. Table Column Head Adaptive fuzzy Potential PSO Path length 904 300 904 Execution Time (sec.) 1.08 2 1.16 VI. CONCLUSION AND FUTURE WORKS In this paper way arranging of portable robots have been tended to. The created movement organizer works for the automated condition with static and moving deterrents. So as to maintain a strategic distance from the detected obstructions (static and dynamic) and create briefest way, another wellness work has been presented. To approve the effectiveness of the proposed approach, reenactment results have been displayed in the different automated conditions. The adaptive fuzzy and the adaptive PSO can be detected and avoid the hard obstacles while the potential failed when the map has difficult irregular obstacles. In future the created calculation is to be executed in automated stage. Build a hard ware to demonstrate the simulation results, taken two moving synchronize robots. REFERENCES 1. Latombe J. C., Robot motion planning, 124, Springer Science and Business Media, 2012. 2. Brooks R. A., A robust layered control system for a mobile robot, IEEE Journal of Robotics and Automation, 2(1), 1986, 14-23. 3. Mitchell Joseph.S.B., An algorithm approach to some problems in terrain navigation, Artificial Intelligence, 37(1-3), 1988, 171-201. 4. Masehian E. and Sedighizadeh D., Classic and heuristic approaches in robot motion planning-a chronological review, World Academy of Science, Engineering and Technology, 23, 2007, 101-106. 5. Stoeter S.A. and Papanikolopoulos N., Kinematic motion model for jumping scout robots, IEEE Transactions on Robotics, 22(2), 2006, 398-403. 6. Hosoda K., Takuma T., Nakamoto A. and Hayashi S., Biped robot design powered by antagonistic pneumatic actuators for multi-modal locomotion, Robotics and Autonomous systems, 56(1), 2008, 46–53. 7. Wu X. and Ma S., CPG-based control of serpentine locomotion of a snake-like robot, Mechatronics, 20(2), 2010, 326–334. 8. Bayraktaroglu Z.Y., Snake-like locomotion: Experimentations with a biologically inspired wheel-less snake robot, Mechanism and Machine theory, 44(3), 2009, 591–602. 9. Guzman J.L., Berenguel M., Rodriguez F. and Dormido S., An interactive tool for mobile robot motion planning, Robotics and Autonomous systems, 56(5), 2008, 396–409. 10. Alexander J.C. and Maddocks J.H., On the kinematics and control of wheeled mobile robots, SRC TR 87-196, October 1987, 1-15. 11. Campion G., Bastin G., and AndrCa-Novel B.D., Structural properties and classification of kinematic and dynamic models of wheeled mobile robots, IEEE Transactions on Robotics and Automation, 12(1), 1996, 47-62. 12. Kim D.S., Kwon W.H. and Park H.S., Geometric kinematics and applications of a mobile robot, International Journal of Control, Automation, and Systems,1(3), 2003, 376-384. 13. Cariou C., Lenain R., Thuilot B. and Martinet P., Adaptive control of four-wheel-steering off-road mobile robots: Application to path tracking and heading control in presence of sliding, IEEE International conference on Intelligent Robots and Systems, France, September 22-26, 2008, 1759-1764. 14. Marcovitz F.R. and Kelly A., On-line mobile robot model identification using integrated perturbative dynamics, Experimental Robotics, Springer Berlin Heidelberg, January, 2014, 417-431. 15. Lee J.H., Kim B.K., Tanikawa T. and Ohba K., Kinematic analysis on omni-directional mobile robot with double-wheel-type active casters, IEEE International conference on Control, Automation and Systems, Seoul, Korea, October 17-20, 2007,1217-1221. 16. Chung W., Moon C., Jung C. and Jin J., Design of the dual offset active caster wheel for holonomic omni-directional mobile robots, International Journal of Advanced Robotic Systems, 7(4), 2010, 105‐110. 17. Li Y.P., Zielinska T., Ang Jr. M.H. and Lin W., Vehicle Dynamics of Redundant Mobile Robots with Powered Caster Wheels, Springer Vienna, 221-228. 18. Kim S. and Lee S., Local and global isotropy analysis of mobile robots with three active caster wheels, Advances in service robotics, ISBN 978-953-7619-02-2, 2008,117-126. 19. Indiveri G., Swedish wheeled omnidirectional mobile robots: kinematics analysis and control, IEEE Transactions on Robotics, 25(1), 2009, 164-171. 20. Indiveri G., Paulus J. and Ploger P.G., Motion control of swedish wheeled mobile robots in the presence of actuator saturation, LNCS 4434, Springer Berlin Heidelberg, 2007, 35-46. 21. Doroftei I., Grosu V. and Spinu V., Omnidirectional mobile robot-design and implementation, Bio-inspiration and robotics: walking and climbing robots, ISBN 978-3-902613-15-8, In-Tech publishers, 2007, 511-528. AUTHORS PROFILE Baidaa M. Madlol: received the B. Sc. degree in communication technical engineering from Al Furat Al Awast Technical University, Iraq in 2006. Currently he is a candidate to pursue a M. Sc. degree in communication technical engineering from Al Furat Al Awast Technical university since 2018. Dr.Ahmad T. Abdulsadda received the B. Sc. degree in electrical engineering from Tickrit University, Iraq in 1997, M. Sc. degree in electrical engineering from Baghdad University, Iraq in 2000, and the PHd in Electrical and Computer Engineering Department at Michigan State University, Michigan, USA. . From 2000 to 2006, he was a faculty member at Baghdad University, Iraq. Since 2006, he has been a faculty member at Technical Najaf College, Al Furat Al Awast University, Iraq. Currently, he is an assistance dean of Al Najaf Technical engineering college, Al Furat Al Awast Technical University. His research interests include robotics fish, feedback control systems, nonlinear estimation techniques, and control theory. Prof. Dr. Ali A. Abdullah Albakry: received the B. Sc. degree in electrical engineering from Tickrit University, Iraq in 1980, M. Sc. degree in electrical engineering from Al Technology University, Iraq in 2000, and the PHd in Electrical and Computer Engineering Department at , Al Technology University, 2008, Iraq. . 
work_kpdo4nx2xnahtfqwp3w53eq7hq	----	Simulation-optimisation based decision-support for coordinated disaster relief last mile distribution SIMULATION-OPTIMISATION BASED DECISION-SUPPORT FOR COORDINATED DISASTER RELIEF LAST-MILE DISTRIBUTION Christian Fikar (a) , Manfred Gronalt (b) , Johannes Goellner (c) , Patrick Hirsch (d) (a),(b),(d) University of Natural Resources and Life Sciences, Vienna, Institute of Production and Logistics, Feistmantelstraße 4, 1180 Vienna, Austria (c) Federal Ministry of Defence and Sports, National Defence Academy, Dept. of Central Documentation & Information, Stiftgasse 2a, 1070 Vienna, Austria (a) christian.fikar@boku.ac.at, (b) manfred.gronalt@boku.ac.at, (d) patrick.hirsch@boku.ac.at (c) johannes.goellner@bmlvs.gv.at ABSTRACT We present a simulation and optimisation based decision-support system to improve last-mile distribution of goods during disasters. In particular, the impact of road closures on supply and optimal transfer points is analysed by considering road, off-road and air transportation. To deal with panic buying, stockpiling and various consequences of a disaster over a rolling time-horizon, a limited number of transfer points have to be selected. At these points, relief organisations and retailers can transfer shipments to prevent stock-outs at regional stores. Furthermore, shipment requests are scheduled to vehicles, routed and optimised during each simulation run. Results show a significant reduction in average lead times when coordinating last-mile distribution during disasters. Keywords: humanitarian logistics, agent-based simulation, last-mile distribution, decision support system 1. INTRODUCTION Transportation highly impacts disaster relief operations (Berariu et al. 2015). Additionally, unexpected increases in demand during disasters are a huge challenge for private companies as well as relief organisations. Naohito et al. (2014) show a significant increase in household spending during the Great East Japan Earthquake of 2011; however, they also report that households were not able to stockpile goods to their desired levels. Holguín-Veras et al. (2014) note that demand also increased in areas not directly damaged by the disasters and that private companies were not able to sufficiently supply goods. Enabling a robust and efficient supply chain benefits retailers, who are able to sell their goods, and relief organisations as panic is reduced if households are able to purchase the demanded goods. For instance, during Hurricane Katrina in 2005, Horwitz (2009) points out that the private sector was able to supply the affected area long before the public agencies arrived. Additionally, retailers often have a substantial amount of demand data available and the logistic experience to deal with disruptions in supply and demand. Public agencies, in contrast, have the regulatory decision- making and special equipment to ship goods through affected areas. To support combining these strengths, we introduce a decision-support system (DSS) to coordinate shipments between private and public actors. Therefore, we model road closures as well as road, off-road and air transportation. In particular, we focus on the situation where relief organisations perform shipments between transfer points in the affected areas. This is shown in Figure 1. In contrast to classical last-mile distribution where the relief organisation delivers goods directly to the victims, by coordinating these actions, resources are utilised efficiently and victims are potentially served faster and more consistently. Therefore, retailers perform the first and the last part of the shipment where roads are still intact, while relief organisations ship between two transfer points through the affected area. Figure 1: Disaster relief last-mile distribution The DSS optimises routing and scheduling decisions, selects promising transfer points and allows analysing various disaster scenarios. Furthermore, users can interactively modify the analysed settings to test the impact of different policies. This enables improved decision-making and supports last-mile distribution. Consequently, the contribution of this paper is twofold, namely introducing a simulation-optimisation based DSS for disaster relief last-mile distribution and analysing the potential of coordinating shipments. 2. LITERATURE REVIEW High computational times and a lack of integrated systems with the right complexity level are the main challenges in humanitarian logistic models (Özdamar and Ertem 2014). The selection of optimal transfer points in our work equals an uncapacitated facility location problem. It aims to find an optimal subset from potential candidate locations derived from an objective Proceedings of the Int. Conf. on Harbor Maritime and Multimodal Logistics M&S, 2015 ISBN 978-88-97999-58-4; Bruzzone, Del Rio Vilas, Longo, Merkuryev, Piera Eds 15 mailto:christian.fikar@boku.ac.at mailto:manfred.gronalt@boku.ac.at mailto:patrick.hirsch@boku.ac.at mailto:johannes.goellner@bmlvs.gv.at function. For a survey on this topic, refer to Gao and Robinson Jr. (1994) and Verter (2011). Facility location in humanitarian relief is discussed by Balcik and Beamon (2008). In disaster relief routing, optimisation can lead to major improvements; however, only little work is found (de la Torre et al. 2012). Routing full truckloads between transfer points can be classified as a full truckload pickup and delivery problem (FT-PDP) as introduced by Gronalt et al. (2003). Location-routing problems consider routing decisions when optimising facility locations. For recent surveys on this topic, refer to Prodhon and Prins (2014) and Drexl and Schneider (2015). Only few articles investigate stochastic problems and the impact of network disruptions is rarely analysed. Longo and Ören (2008) note the importance of modelling and simulation to study supply chain vulnerability and to improve resilience. By combining simulation and optimisation, uncertainties and complex interactions within optimization models can be overcome (Glover et al. 1996). This allows one to investigate the impact of network disruptions as well as disasters on last-mile distribution and to further discuss the potential of coordination. 3. PROBLEM DESCRIPTION In an area affected by a disaster, roads may be closed for certain durations of time during a day. Potential reasons for this are floods or road blockages as a result of earthquakes or mudslides. Due to panic buying and unexpected demand, stores randomly request additional shipments. Full truck-loads are assumed, i.e. no split- deliveries are performed and shipments have exactly one supply and one demand point. The private actor, e.g., a retailer, can either ship these directly or use the service of a public actor, e.g., a relief organisation. In the latter case, the retailer transports the goods to a transfer point. At this point, the relief organisation takes over the shipment through the affected areas to a second transfer points from where goods are delivered to the store by the retailer. To perform this service, the relief organisation has a number of given heterogeneous helicopters and off-road vehicles available which are dedicated for this purpose. Each is assigned to one of multiple depots where the vehicle starts and ends its operations. If idle, the vehicle always returns to the depot. Private trucks of the retailers are assumed to be always immediately available. While the retailer is only able to drive on open roads, the relief organisation can either fly or utilise closed roads due to special permissions and equipment. Relief organisations only perform shipments between predefined transfer points, where a limited number of facilities have to be selected from a large set of potential candidates. This selection has to be done before the first request occurs and cannot be altered later due to the high organisational effort of moving transshipment equipment. Transfer points can either be large parking areas, open fields or any other suitable area where transshipments can be performed. All vehicles travel respecting speed limits on the individual arcs and are subject to loading and unloading times. The objective is to minimise the average lead time over all shipping requests, i.e. how long it takes on average to deliver shipments to stores from when the request occurred. Therefore, an optimal arrangement of transfer points, an optimal scheduling of requests to vehicles and optimal routes, all considering an uncertain environment and road closures, have to be derived. Figure 2 gives a simple sample problem with only one request, four potential transfer points, one public vehicle, one depot, one supply point and one demand point. Due to the closure of a bridge, the retailer can perform a detour crossing the next bridge or coordinate the shipment with the relief organisation, which has an off-road vehicle available to cross the closed bridge. As the detour is substantial, the shipment is coordinated. Therefore, two transfer points have to be selected and the shipment has to be scheduled to the public vehicle. As a result, both vehicles travel to the first transfer point where the goods get transhipped. After arrival of the public vehicle at the second transfer point, the retailer delivers the goods to the store. The lead time is denoted by the difference between the arrival time at the store and the time when the request occurred. Figure 2: Problem description 4. SIMULATION-OPTIMISATION BASED DSS An agent-based simulation, a heuristic optimisation procedure to improve vehicle routes and a Tabu Search metaheuristic to select promising transfer points are combined to analyse the introduced problem. Inputs are demand and supply points, vehicle depots, potential transfer points as well as available vehicles, network data and information about road closures. The user specifies a number of requests and a time horizon in which these shipments occur. The simulation stops as soon as the last request is delivered and outputs lead times, vehicle routes and coordination decisions. Proceedings of the Int. Conf. on Harbor Maritime and Multimodal Logistics M&S, 2015 ISBN 978-88-97999-58-4; Bruzzone, Del Rio Vilas, Longo, Merkuryev, Piera Eds 16 4.1. Agent-based Simulation The agent-based simulation consists of four main components: locations, vehicles, requests and closed roads. Each is modelled as an agent in a geographic space based on the respective network data and coordinates. Locations further differ in stores, depots, transfer points and supply nodes. Vehicles indicate public vehicles, while private trucks are represented within the request agent. As dynamic changes in the simulation can lead to a high number of substantial rerouting and rescheduling decision, vehicles are defined with the statechart shown in Figure 3. If a request gets scheduled, the vehicle moves to the pickup location where it waits if it arrives before the private truck. While moving or waiting, the current task may be aborted to reroute the vehicle to serve a different shipment request. After a loading time delay the vehicle travels to the delivery location. During this trip, we assume that no rescheduling is possible as the vehicle is already loaded. As a result, splitting a shipment in multiple stages to be performed by different vehicles is not enabled. Finally, at arrival at the delivery location, the shipment is transferred to the private truck and the public vehicle continues with its next scheduled request. If no further requests are scheduled, the vehicle returns to the depot. Figure 3: Statechart of public vehicles The given number of requests are uniformly randomly distributed over time and associated to randomly selected supply points and stores. To calculate the shortest path on the street network between two locations, a bidirectional implementation of the A* algorithm (Hart et al. 1968) is used. Helicopters are assumed to fly in a straight line; however, adding a special routing graph for this purpose is possible. Closed roads are specified as input and associated with a start and end time. The cost of traversing closed roads with private vehicles is set to infinity, forcing the routing algorithm to look for alternative paths. 4.2. Coordination Decision Each time a new request occurs or roads open or close, the solution procedure has to decide if a transshipment should be performed. If so, it decides at which transfer points and which public vehicle should at what time bridge the affected area. Figure 4 gives a simplified example of a decision if a transshipment should be performed considering loading and unloading times of 10 minutes. Due to a road closure, the direct delivery between supply point S1 and store D1 is not possible. As a consequence, the retailer can either select a detour or request a transshipment via transfer point T1 and T2 with a relief organisation. Due to loading times, the detour is faster, which results in no transshipment in this case. Figure 4: Routing and transshipment decision If coordinating is potentially beneficial, where to load and unload this request on a public vehicle has to be decided. Therefore, a cheapest insertion heuristic tests all promising transfer combinations on each position within each vehicle's schedule. One option is to select transfer points, which result in the shortest total travel duration, i.e. the shipment travels on the shortest path; however, depending on the current availability of public vehicles this potentially results in substantial wait times. Depending on which vehicle arrives later, either the shipment or the public vehicle waits. As a consequence, wait times are considered by selecting transfer points individually for each vehicle within the routing and scheduling algorithm. Therefore, all combinations of loading and unloading points where the shortest path is less than the best found combination so far for this request are evaluated. Figure 5 gives a simplified example ignoring loading and unloading times. Vehicle V1 first finishes its current shipment to transfer point T3 and then continues to T1. This trip requires 20 minutes. The private truck arrives after 15 minutes, resulting in 5 minutes of wait time for the request. After the loading operations, the request continues to the demand point with transshipment at T2. The request arrives at the final location after 35 minutes. This value is compared to the Proceedings of the Int. Conf. on Harbor Maritime and Multimodal Logistics M&S, 2015 ISBN 978-88-97999-58-4; Bruzzone, Del Rio Vilas, Longo, Merkuryev, Piera Eds 17 best found combination for the request so far and updated if improved. Changes in the route of the public vehicle can, however, lead to delays for other requests scheduled after the inserted change. These costs have to be additionally considered to evaluate transshipments. Figure 5: Lead time calculation to select transfer points The best combination is selected and scheduled. Whenever a road is closed or reopened, all previously made decisions are re-evaluated and improved if possible. To speed up computation, non-promising combinations are aborted if already worse than the best found combination so far. In the next step, the routing of the vehicle is optimised. 4.3. Trip Assignment and Improvement Each time after a request is scheduled or re-scheduled an intra-route optimisation operator aims to improve the schedule where the request is inserted. Therefore, swapping the order of two requests on this vehicle is tested. Additionally, an inter-route relocate operator checks if moving a request from one vehicle to another improves the objective value. Both operators follow a best improvement strategy, i.e. all potential improvements are tested and the best is applied. This is repeated until no further improvements are found. The occurrence of new shipments can, however, lead to a situation where it is advantageous to cancel earlier made coordination decisions as it beneficial to utilise public vehicles for other requests instead. To consider this fact, a drop operator is implemented and run after inter-route optimisation. It checks if removing a request from a public vehicle and delivering it directly with a private truck from its current position reduces the average lead time over all requests. Therefore, only public vehicles which were changed during the inter- route optimisation or as a result of an insertion of new requests are evaluated to reduce computational time. In case a request is dropped from a vehicle, the inter-route optimisation procedure is restarted. 4.4. Selection of Transfer Points As equipment to perform transshipment is limited and further time-consuming to relocate during disasters, the decision-makers may only be able to operate a limited number of transfer points. Selecting a good set of transfer points allows reducing lead times and improving the system’s performance. Nevertheless, this decision is not trivial due to network disruption and unknown demand. As a consequence, a Tabu Search (Glover 1989; Glover 1990) was implemented to select a limited number of transfer points out of a set potential candidate locations. Tabu Searches show promising results for multiple related optimisation problems such as the uncapacitated facility location problem (Ghosh 2003), the p-median problem (Rolland et al. 1997) and locating emergency facilities under random network damages (Salman and Yücel 2015). Each candidate location is associated with a binary variable yi, which equals 1 if open and 0 if closed. To generate an initial solution, a construction heuristic was developed which selects transfer points based on their attractiveness for transshipments. Therefore, the simulation is run for a pre-specified number of replications runs with all potential candidate transfer points being open and collects statistics about how often a shipment was loaded or unloaded at a certain candidate. In the next step, these candidates are sorted and the most utilised transfer points are added to the initial solution until the maximum number of transfer points is reached. The initial solution is further improved in the following iterations by a swap operator which opens one transfer point by simultaneously closing a previously open one. Each potential solution in this neighbourhood is evaluated by running the specified number of replication runs of the simulation. The best solution of the iteration is stored and acts as the incumbent solution for following iteration. The selected move is set tabu for τ iterations, i.e. it is not allowed to be undone. Nevertheless, if performing this move improves the overall best found solution so far, the tabu criteria is revoked. Furthermore, moves which lead to a worse objective value compared to the last iteration are penalised based on how often the respective candidate has been added to the solution in prior iterations. This helps to diversify the search procedure and is done by the following evaluation function 𝑔(𝑆) with 𝑓(𝑆) indicating the objective value of the solution, 𝑘𝑖 the times a transfer point has been previously added to a solution and ζ the penalty weight. g(S) = 𝑓(𝑆)(1 + ζki) (1) After either a given time limit or a maximum number of iterations, the Tabu Search terminates and returns the best found objective value as well as the best set of transfer points to operate. 4.5. Dynamic extensions The simulation further allows including sudden car breakage as well as changes in air weather, which enables or disables air transportation. Furthermore, roads and transfer points may be closed or open dynamically based on pre-specified stochastic distributions or by user interaction. In each of the cases, all affected requests are re-evaluated if transshipment Proceedings of the Int. Conf. on Harbor Maritime and Multimodal Logistics M&S, 2015 ISBN 978-88-97999-58-4; Bruzzone, Del Rio Vilas, Longo, Merkuryev, Piera Eds 18 should be adjusted considering the new situation. Allowing the user to dynamically perform such changes during a simulation run enables one to test different scenarios, run risk assessments and to visualise potential impacts. 5. COMPUTATIONAL EXPERIMENTS The DSS is developed with AnyLogic 7.1 (AnyLogic 2015) using OpenStreetMap (OpenStreetMap 2015) and a custom implementation of GraphHopper 0.3 (GraphHopper 2015) to route vehicles on different networks. The solution procedure is coded in Java. Computational experiments were run on an Intel Core i7-4930K, 64GB RAM, MS-Windows 7 and 6 threads operating in parallel. 5.1. Test Scenario Test area is Krems an der Donau, Austria, and the surrounding area, a region often affected by floodings of the Danube River Basin. To reach the city center of Krems, shipments either have to cross the Danube River on one of two bridges or circumvent the area to take a major highway bridge in the east of the region. As demand points, the geographic position of supermarkets, pharmacies and other major retailers in the area are selected. These stores are supplied from two supply points in the east, one north of the Danube and one south, both located at highway exits. To indicate potential transfer points, large parking lots as well as football pitches and industrial areas were selected. Therefore, it is assumed that all transfer points are accessible for both air and off-road vehicles. In total, 29 stores, two supply points and 80 potential transfer points are available. Figure 6 plots the studied region. Figure 6: Test scenario area Two helicopters and three off-road vehicles are stationed at five depots in the region. Vehicle depots were located based on fire-fighter stations in the area, a military barrack and a major hospital. Helicopters travel with an average speed of 135 km/h and require six minutes for loading and unloading operations. Off-road vehicles travel with 45 km/h and take three minutes for loading operations. Private trucks travel respecting individual speed limits of each road with a maximum travel speed of 95 km/h. Road closures are set as reported by the Austrian Press Agency (APA-OTS 2013) for the 7 th of June, 2013 when the region was struck by a major river flooding. Roads are closed for the entire period and travel times are considered to be certain, e.g., there is no random component in the time it takes to travel between two points. Based on opening hours of stores in the region, the simulation starts at 8am and shipment request may occur until 4pm. The average number of shipment requests per store in the computational experiments is varied between one and fifty requests over the eight hours simulation horizon. Furthermore, all dynamic extensions reported in Section 4.5 are disabled for the computational experiments to enable clearer comparisons of different policies by reducing stochastic impacts. 5.2. Parameter Setting To consider stochastics effects in demand, each evaluation is run with 100 replications and average results are reported in this work. The initial solution is constructed based on the attractiveness of transfer points after 6 runs. For the optimisation of facility locations, the tabu tenure τ is uniformly randomly selected after each iteration in [0,⌈√𝑛⌉] with n indicating the number of potential candidate facilities. The diversification penalty ζ is randomly set in [0,0.1] after each iteration to vary diversification. The time limit is set to 30 minutes. To speed up optimisation, replications are aborted for a single solution if the current mean is worse than the best found solution in the current iteration based on a confidence level of 95% and after a minimum of 6 replications. Additionally, memory techniques are used to store reoccurring evaluations, solutions are evaluated in parallel on multiple threads of the computer systems and all visual animations are disabled. 5.3. Preliminary Results and Discussion Figure 7 shows a solution with three transfer points to open and with on average five shipments requests occurring per store during the simulation horizon. Due to the road closures, two main connections within the city of Krems as well as one of the two bridges are not traversable for private vehicles. As a consequence, out of the 80 potential candidate locations, the Tabu Search selects three candidate locations for the coordination with the objective to reduce average lead times. Figure 7: Sample solution for the 07 th June 2013. Proceedings of the Int. Conf. on Harbor Maritime and Multimodal Logistics M&S, 2015 ISBN 978-88-97999-58-4; Bruzzone, Del Rio Vilas, Longo, Merkuryev, Piera Eds 19 Two transfer points are located closely to stores where direct connections are disrupted by road closures. The third one is located closely to the exit of the main bridge entering the area. At this point, the public actor takes over the shipment to bridge the affected area to one of the other two transfer points. 5.3.1. Impact of Coordination To analyse the impact of coordinating last mile distribution, three settings are compared. Therefore, the three prior selected transfer points are used. “Best case” indicates the situation where no roads are closed. It gives a natural lower bound for the system performance. Note that, due to a higher travelling speed and different routes of helicopters, coordination can potentially lead to a lower average lead time if no roads are closed; however, it can be assumed that no coordination between public and private actors is performed in such a setting. “Worst case”, in contrast, indicates the situation where no public vehicles are used and private trucks use the remaining open roads to deliver goods to stores. This setting acts as a natural upper bound. Both can be, assuming a uniform distribution of supply and demand points of requests and no changes in the street network, calculated by determining the average lead time μcj,m from all s supply points to all m stores under the given street network. μcij = 1 𝑠𝑚 ∑ ∑ 𝑐𝑖𝑗 𝑚 𝑗=1 𝑠 𝑖=1 (2) “Coordinated” allows transshipments of goods as studied in this paper and requires one to run the solution produce. Figure 8 plots the benefits of providing public vehicles to transfer goods through the affected areas. Figure 8: Impact of disaster-relief coordination Compared to the worst case scenario, average lead times are substantially reduced, especially if only few shipments occur. In this case, public vehicles are only little utilised and can efficiently transfer shipments. Nevertheless, the optimal scenario where no road closures occur is not reached due to detours to transfer points, additional loading times, waiting times for vehicles and varying travelling speeds of different vehicles. In general, the number of available public vehicles and the extend of road closures, as well as travelling speed of public vehicles and loading and unloading times have a major impact on the benefits of last-mile coordination in disaster relief. If the number of shipments per store increases, utilisation rises as well and, as a consequence, substantial wait times occur. High wait times further lead to more cases where a detour is faster compared to coordinating the shipment. As a result, with a higher number of shipments, average lead times converge to the worst case lead time without coordination as only few gains can be realised. In such a setting, adding more vehicles for the coordination is beneficial. 5.3.2. Impact of Vehicle Optimisation Optimisation within a dynamic setting is challenging as decisions which are “optimal” considering the current status of the problem might lead to negative effects later when the problem changes due to the occurrence of new requests. To analyse the impact of our optimisation procedure, it is compared to a simple scheduling rule, which adds a new shipment to the end of the vehicle schedule which results in the lowest average lead time. Nevertheless, all potential combinations of transfer points are evaluated in this step. Furthermore, no inter- and intra-route changes are performed. Results of both methods are compared in Figure 9 with the three prior selected transfer points. Figure 9: Impact of the optimisation procedure The analysis shows that given a small number of shipments per store, the optimisation procedure has nearly no impact. The main reason for this is the low number of shipments simultaneously scheduled to a single vehicle. This gives little room for improvement. With an increasing number of shipments, however, the number of simultaneously scheduled requests per vehicle increases, allowing the optimisation procedure to improve average lead times by altering vehicle routes. Based on these results, the user is given the option to deactivate the optimisation procedure prior to a simulation run to save on computational run time. 5.3.3. Impact of Number of Transfer Points Opening more transfer points increases organisational efforts and potentially costs; however, average lead Proceedings of the Int. Conf. on Harbor Maritime and Multimodal Logistics M&S, 2015 ISBN 978-88-97999-58-4; Bruzzone, Del Rio Vilas, Longo, Merkuryev, Piera Eds 20 times are expected to decrease as better coordination can be achieved. To plot the impact of the number of transfer points, the optimisation was run for each potential number of transfer points to open with 5 shipments requests per store. Results of this analysis are shown in Figure 10. Figure 10: Impact of the number of open transfer points As expected, average lead times are the highest if only a few transfer points are operated. Adding an additional transfer point decreases the lead time substantially if only few facilities are open, while little to no impact is achieved if many candidates are already selected. Furthermore, after 13 facilities, opening an additional transfer point is no longer beneficial as the additional facility is not utilised for transshipments and, as a consequence, does not improve the system’s performance. Analysing the problem in this way, helps to reduce the problem size substantially as non- promising candidates can be removed from the set of potential facilities to improve run times. Additionally, having fewer candidates to consider allows decision- makers to closer investigate the remaining candidates. 6. CONCLUSION To support coordinated last-mile distribution during disasters, DSSs are crucial to investigate and analyse potential strategies and resulting trade-offs. The introduced DSS offers this by combining simulation and optimisation techniques. The focus on modelling real- world transportation networks and the consideration of road closures enables one to utilise the developed tool in education and training activities. It helps to define promising locations of transfer points, analyse the impact of road closures and further plots the benefits of coordination. As a result, future decision-making in the context of disaster relief logistics can be improved to support victims effectively and to lower the impact of disasters. Future work includes facilitating the developed DSS to test the impact of certain road closure on the performance of the system. Additionally, embedding the DSS in the development of serious games to train decision-makers is one potential research direction of high interest. ACKNOWLEDGMENTS Financial support of the program KIRAS (project number 840 897) provided by the Austrian Federal Ministry for Transport, Innovation and Technology, is gratefully acknowledged. REFERENCES AnyLogic, 2015. AnyLogic - Multimethod Simulation Software. Available from: http://www.anylogic.com/ [accessed 25 June 2015] APA-OTS, 2013. Aktuelle Straßensperren wegen Hochwassers in Niederösterreich. Available from: http://www.ots.at/presseaussendung/OTS_201306 07_OTS0088/aktuelle-strassensperren-wegen- hochwassers-in-niederoesterreich [accessed 25 June 2015] Balcik, B., Beamon, B. M., 2008. Facility location in humanitarian relief. International Journal of Logistics, 11(2), 101-121. Berariu, R., Fikar, C., Gronalt, M., Hirsch, P., 2015. Understanding the impact of cascade effects of natural disasters on disaster relief operations. International Journal of Disaster Risk Reduction 12, 350-356. de la Torre, L. E., Dolinskaya, I. S., Smilowitz, K. R., 2012. Disaster relief routing: Integrating research and practice. Socio-Economic Planning Sciences 46 (1), 88-97. Drexl, M., Schneider, M., 2015. A survey of variants and extensions of the location-routing problem. European Journal of Operational Research 241 (2), 283-308. Gao, L.-L., Robinson Jr., E., 1994. Uncapacitated facility location: General solution procedure and computational experience. European Journal of Operational Research (3), 410-427. Glover, F., Kelly, J. P., Laguna, M., 1996. New advances and applications of combining simulation and optimization. In: Proceedings of the 28th Conference on Winter Simulation. WSC '96. IEEE Computer Society, Washington, DC, USA, pp. 144-152. Ghosh, D., 2003. Neighborhood search heuristics for the uncapacitated facility location problem. European Journal of Operational Research 150 (1), 150-162. GraphHopper, 2015. Graphhopper route planner. Available from: https://graphhopper.com [accessed 25 June 2015] Gronalt, M., Hartl, R. F., Reimann, M., 2003. New savings based algorithms for time constrained pickup and delivery of full truckloads. European Journal of Operational Research 151 (3), 520-535. Hart, P., Nilsson, N., Raphael, B., 1968. A formal basis for the heuristic determination of minimum cost paths. Systems Science and Cybernetics, IEEE Transactions on 4 (2), 100-107. Proceedings of the Int. Conf. on Harbor Maritime and Multimodal Logistics M&S, 2015 ISBN 978-88-97999-58-4; Bruzzone, Del Rio Vilas, Longo, Merkuryev, Piera Eds 21 Holguín-Veras, J., Taniguchi, E., Jaller, M., Aros-Vera, F., Ferreira, F., Thompson, R. G., 2014. The Tohoku disasters: Chief lessons concerning the post disaster humanitarian logistics response and policy implications. Transportation Research Part A: Policy and Practice 69, 86-104. Horwitz, S., 2009. Wal-Mart to the rescue: private enterprises response to Hurricane Katrina. The Independent Review 13 (4), 511-528. Longo, F., Ören, T., 2008. Supply chain vulnerability and resilience: a state of the art overview. Proceedings of European Modeling & Simulation Symposium, pp. 527-533. 17-19 September 2008, Campora S. Giovanni (CS) Italy. Naohito, A., Moriguchi, C., Noriko, I., 2014. The effects of natural disasters on prices and purchasing behaviors: The case of the Great East Japan Earthquake. RCESR Discussion Paper Series DP14-1, Research Center for Economic and Social Risks, Institute of Economic Research, Hitotsubashi University. OpenStreetMap, 2015. OpenStreetMap. Available from: www.openstreetmap.org [accessed 25 June 2015] Özdamar, L., Ertem, M. A., 2014. Models, solutions and enabling technologies in humanitarian logistics. European Journal of Operational Research 244 (1), 55-65. Prodhon, C., Prins, C., 2014. A survey of recent research on location-routing problems. European Journal of Operational Research 238 (1), 1-17. Rolland, E., Schilling, D. A., Current, J. R., 1997. An efficient tabu search procedure for the p-median problem. European Journal of Operational Research 96 (2), 329-342. Salman, F. S., Yücel, E., 2015. Emergency facility location under random network damage: Insights from the Istanbul case. Computers & Operations Research 62, 266–281. Verter, V., 2011. Uncapacitated and capacitated facility location problems. In: Eiselt, H. A., Marianov, V. (Eds.), Foundations of Location Analysis. Vol. 155 of International Series in Operations Research & Management Science. Springer US, pp. 25-37. AUTHORS BIOGRAPHY Christian Fikar is a Research Assistant at the Institute of Production and Logistics (PWL). He holds a master degree in supply chain management from Vienna University of Economics and Business, Austria. The main focus of his research is on humanitarian logistics and meta-/matheuristics in the context of innovative and sustainable transport concepts. Currently, he is working on his doctoral thesis in this field. Manfred Gronalt is a Professor and Head of PWL. His expertise lies within production management, logistics and simulation. Up to now, he authored over 100 scientific journal articles and conference contributions. He is a member of ÖGOR (Austrian society of operations research) and the Academic Association for Business Research (VHB). Johannes Goellner is head of the section Knowledge Management in the Department of Central Documentation and Information Service at the National Defence Academy of the Federal Ministry of Defence and Sports, Vienna and Expert in risk-, crises and knowledge management. His research areas include knowledge development and Foresight, organizational development, trend- and risk analysis as well as scenario planning and development, operations research and risk and crises management in supply chain management. He contributed several standardisation initiatives, like the Austrian Standards Institute, International Standards Organisation, European Committee and Guideline-Development for „Supply Chain Risk Management“ at the Risk Management Association-RMA, Munich, Germany. He has been director of the study program: Master of Business Administration “Environmental Threats and Disaster Management” and Head of Section Risk Management at the NBC-Defence School of the Austrian MoD. He is visiting professor for Knowledge Development at the Masaryk University, Brno, CZ and lecturer for risk and crises management at the University of Natural Resources and Life Sciences and member and Chairman of the Board of the Center of Risk- and Crises Management, Vienna. He has further research experience like in EU-Research-Project “FOCUS- Foresight Security Scenarios: Mapping Research to a Comprehensive Approach to Exogenous EU Roles”, 2009-2012 and various national security research projects. Patrick Hirsch is an Assistant Professor, Project Manager and Deputy Head of PWL. He acts as a reviewer for several international scientific journals and is area editor of the Journal of Applied Operational Research. His research interests are within transportation logistics, health care logistics and disaster management. He presented his work at several international conferences and published numerous book chapters as well as scientific journal articles. Proceedings of the Int. Conf. on Harbor Maritime and Multimodal Logistics M&S, 2015 ISBN 978-88-97999-58-4; Bruzzone, Del Rio Vilas, Longo, Merkuryev, Piera Eds 22 
work_krmvgatlr5ce3mvlxofc3t3e2i	----	Producing Modular Hybrid Rule Bases for Expert Systems Published in the International Journal of Artificial Intelligence Tools (IJAIT), Vol. 10, No. 1-2 (2001), 87-105.  Copyright World Scientific Pub 2000. All rights reserved. CONSTRUCTING MODULAR HYBRID RULE BASES FOR EXPERT SYSTEMS I. HATZILYGEROUDIS, J. PRENTZAS University of Patras, School of Engineering Dept of Computer Engin. & Informatics, 26500 Patras, Hellas (Greece) Email: ihatz/prentzas@ceid.upatras.gr & Computer Technology Institute, P.O. Box 1122, 26110 Patras, Hellas (Greece) Email: ihatz@cti.gr ABSTRACT Neurules are a kind of hybrid rules integrating neurocomputing and production rules. Each neurule is represented as an adaline unit. Thus, the corresponding neurule base consists of a number of autonomous adaline units (neurules). Due to this fact, a modular and natural knowledge base is constructed, in contrast to existing connectionist knowledge bases. In this paper, we present a method for generating neurules from empirical data. To overcome the difficulty of the adaline unit to classify non-separable training examples, the notion of ‘closeness’ between training examples is introduced. In case of a training failure, two subsets of ‘close’ examples are produced from the initial training set and a copy of the neurule for each subset is trained. Failure of training any copy, leads to production of further subsets as far as success is achieved. Keywords: hybrid knowledge representation, symbolic representation, connectionist representation, production rules, neural networks, modularity, naturalness. 1. Introduction Recently, there has been extensive research activity at combining (or integrating) the symbolic and the connectionist approaches1, 2, 3. There are a number of efforts at combining symbolic rules and neural networks for knowledge representation4, 5, 6, 7. What they do is a kind of mapping from symbolic rules to a neural network. Also, connectionist expert systems8, 9, 10 are a type of integrated systems that represent relationships between concepts, considered as nodes of a neural network. The strong point of those approaches is that knowledge elicitation from the experts is reduced to a minimum. A weak point of them is that the resulted systems lack the naturalness and modularity of symbolic rules. This is mainly due to the fact that those approaches give pre-eminence to connectionism. For example, the systems in8, 10 are more or less like black boxes and, to introduce new knowledge, one has to modify a large part of the network. Connectionist knowledge bases cannot actually be incrementally developed. We use neurules11, 12, which achieve a uniform and tight integration of a symbolic component (production rules) and a connectionist one (the adaline unit). Each neurule is considered as an adaline unit. However, pre-eminence is given to the symbolic component. Thus, the constructed knowledge base retains the modularity of production rules, since it consists of autonomous units (neurules), and their naturalness, since neurules look much like symbolic rules. A difficult point in this approach is the inherent inability of the adaline unit to classify non-separable training examples. In this paper, we describe a method for generating neurules directly from empirical (training) data. We overcome the above difficulty of the adaline unit by introducing the notion of ‘closeness’, as far as the training examples are concerned. That is, in case of failure, we produce two subsets of the initial training set of the involved neurule, which contain ‘close’ success examples and train a copy of the neurule for each subset. Failure of training any copy leads to production of further subsets as far as success is achieved. This paper is a revised and extended version of the one presented at FLAIRS’200013. The structure of the paper is as follows. Section 2 presents neurules and the corresponding expert system architecture. Section 3 presents the basic ideas introduced in the paper. In Section 4, the algorithm for creating a hybrid knowledge base directly from empirical data is described. In Section 5 the hybrid inference mechanism is presented. Section 6 contains examples and experimental results. Section 7 discusses related work and finally, Section 8 concludes. 2. Neurules 2.1 Structure Neurules (: neural rules) are a kind of hybrid rules. Each neurule is considered as an adaline unit (Fig.1a). The inputs Ci (i=1,...,n) of the unit are the conditions of the rule. Each condition Ci is assigned a number sfi, called its significance factor, corresponding to the weight of the corresponding input of the adaline unit. Moreover, each rule itself is assigned a number sf0, called the bias factor, corresponding to the weight of the bias input (C0 = 1, not illustrated in Fig.1 for the sake of simplicity) of the unit. Each input takes a value from the following set of discrete values: [1 (true), 0 (false), 0.5 (unknown)]. This gives the opportunity to easily distinguish between the falsity and the absence of a condition, in contrast to symbolic rules. The output D, which represents the conclusion (decision) of the rule, is calculated via the formulas: as usual14, where a is the activation value and f(x) the activation function, which is a threshold function (Fig.1b). Hence, the output can take one of two values, ‘-1’ and ‘1’, representing failure and success of the rule respectively. D f sf sf Ci i n i= + = ∑(a), a= 0 1 Fig.1. (a) A neu 2.2 Syntax and semantics The general syntax (structure occurrences and ‘< >’ denote <rule>::= (<bias-factor>) if < <conditions>::= <condition> <conclusions>::= <conclusion <condition>::= <variable> <l <conclusion>::= <variable> < where <variable> denotes a v domain, e.g. ‘sex’, ‘pain’ et concept’ etc in a tutoring do variable or an intermediate v in a conclusion can be eith variable takes values from t variables take values through conclusions respectively. <l- The symbolic predicates are =}. <r-predicate> can only be be a symbol or a number. significance (weight) of the neurules see Fig. 3, Section 6 We distinguish between inp is a neurule having only inpu variables in its conclusions. A intermediate variable in its c An output neurule is one havi C1 C2 (sf1) (sf2) (sfn) (sf0) D (a f(x) x 1 0 . . . rule as an adaline unit (b) the activation function ) of a rule is (where ‘{ }’ denotes zero, one or more s non-terminal symbols): conditions> then <conclusions> {, <condition>} > {, <conclusion>} -predicate> <value> (<significance-factor>) r-predicate> <value> ariable, that is a symbol representing a concept in the c, in a medical domain, or ‘learning-ability’, ‘related- main. A variable in a condition can be either an input ariable or even an output variable, whereas a variable er an intermediate or an output variable. An input he user (input data), whereas intermediate and output inference, since they represent intermediate and final predicate> denotes a symbolic or a numeric predicate. {is, isnot}, whereas the numeric predicates are {<, >, a symbolic predicate. <value> denotes a value. It can The significance factor of a condition represents the condition in drawing the conclusion(s). (For example ). ut, intermediate and output neurules. An input neurule t variables in its conditions and intermediate or output n intermediate neurule is a neurule having at least one onditions and intermediate variables in its conclusions. ng an output variable in its conclusions. Cn ) -1 (b) 2.3 The neurule based architecture In Fig.2, the architecture of a hybrid neurule-based expert system is illustrated. The run-time system (in the dashed rectangle) consists of three modules, functionally similar to those of a conventional rule-based system: the neurule base (NRB), the hybrid inference mechanism (HIM) and the working memory (WM). The NRB contain 4). The initial neur mechanism (TRM) a (training) data. The H the input data in the has the same forma value the special sym user or intermediate/ 3. The Basic Ideas The main objective of as fine granularity it is clear from the p to and look much component, we retai neurules are conside neural units, which possible neural granu The main problem classify training pat function. A well kn non-separability in m WM NRB HIM M Input data Final conclusions Training data Facts Neurules TR Fig.2 A neurule-based expert system architecture s neurules produced from empirical (training) data (see Section ules, constructed by the user, are trained using the training nd the training examples produced from the available empirical IM is responsible for making inferences by taking into account WM and the rules in the NRB. The WM contains facts. A fact t as a condition/conclusion of a rule, however, it can have as bol “unknown”. Facts represent either conditions given by the final conclusions produced during an inference course. of our approach is to produce modular hybrid knowledge bases as possible. Our main choice to this end is to use neurules. As revious section, neurules are hybrid rules that give pre-eminence like symbolic rules. By giving pre-eminence to the symbolic n the naturalness and modularity of symbolic rules. Internally, red as independent neural units, more specifically, as adaline are individually trained. As a single neural unit is the finest le, a neurule is the finest possible hybrid granule. with this is the inherent inability of a single neural unit to terns (examples) that correspond to a non-separable (boolean) own such function is the XOR function, which is the basis of any cases. The training examples that correspond to the XOR Initial neurules function with two input variables are presented in Table 1, where v1 and v2 represent the input values (‘0’ and ‘1’ mean ‘false’ and ‘true’ respectively). Also, d represents the output value (‘-1’ and ‘1’ mean ‘inactive’ and ‘active’ respectively). We call success examples the patterns with d=1 and failure examples those with d=-1. From a knowledge representation point of view, success examples result in a cell’s activation and hence in (positive) knowledge production, whereas failure examples act like a protection from misactivations (produce negative knowledge). As it is known, there is no single-cell network that can represent the XOR function14. For example, a single unit cannot correctly classify all four patterns of Table 1. However, it can classify any three of them. Hence, if we remove one of the success examples, the remaining examples can be used to train a single unit (neurule). Furthermore, to be able to represent the removed success example, we need a second unit (neurule). However, to avoid misactivations we should use the failure examples, alongside the removed success example, to train the second neural unit. Thus, two independent neural units (neurules) are needed to represent the two- input XOR function. The first unit is trained to classify the examples in the set {[0 0 -1], [0 1 1], [1 1 -1]} and the second those in {[0 0 -1], [1 0 1], [1 1 -1]}. In crea • Ea ex • Th sam The first activated, two units satisfy the examples training s that is the Similar four indep represents four succ Table 1. Two-input XOR examples Table 2. Three-input XOR examples v1 v2 d 0 0 -1 ting these training sets, we had two i ch unit (neurule) should be acti ample. ere should be no two units (neuru e success example. requirement assures that, there is that is will never produce an outpu (neurules) that will be activated by se requirements, contains one of th . This means that the success examp ubsets. Notice, here, that the two s y have no common input values. things hold for the three-input XO endent neural units (neurules) are n a subset of the eight training examp ess examples and all of the failure 0 1 1 1 0 1 1 1 -1 v1 v2 v3 d 0 0 0 -1 mplicit requirements in mind: vated by at least one success les) that can be activated by the no unit (neurule) that will never be t. The second assures that, there are no the same data. Each subset, in order to e success examples and all the failure les are the basis for the creation of the uccess examples are totally unrelated, R function (see Table 2). In this case, ecessary to represent it. Each of them les of Table 2 that includes one of the examples. Again, three of the success 0 0 0 1 1 1 1 0 1 1 0 0 1 1 1 0 1 0 1 0 1 1 1 -1 1 -1 -1 1 examples are totally unrelated, whereas the forth has only one common input value with the others. This analysis, which shows that totally or greatly unrelated success examples may cause non-separability, and an experience with a related problem12 led us to introduce the notion of ‘closeness’ between success examples (see Section 4.2) as a criterion for the creation of the training subsets with more than one success example. Closeness is used to guide distribution of success examples between subsets. In using the independent units (neurules) coming from the same training set, if one of them gets active, there is no need to evaluate the rest ones, since there is no possibility to have other activated units (neurules) for the same input data. 4. Neurule Base Construction The algorithm for constructing a hybrid rule base from empirical (training) data is outlined below: 1. Determine the input, intermediate and output variables and use dependency information to construct an initial neurule for each intermediate and output variable. 2. Determine the training set for each initial neurule from the training data, train each initial neurule using its training set and produce the corresponding neurule(s). 3. Put the produced neurules into the neurule base. In the sequel, we elaborate on each of the first two steps of the algorithm. 4.1 Constructing the initial neurules To construct initial neurules, first we need to know or determine the input, intermediate and output variables. Then, we need dependency information. Dependency information indicates which variables (concepts) the intermediate and output variables (concepts) depend on. If dependency information is missing, then output variables depend only on input variables, as indicated by the training data. In constructing an initial neurule, all the conditions including the input, intermediate and the output variables that contribute in drawing a conclusion (which includes an intermediate or an output variable) constitute the inputs of the initial neurule and the conclusion its output. So, a neurule has as many conditions as the possible input, intermediate and output variable-value pairs. Also, one has to produce as many initial neurules as the different intermediate and output variable- value pairs specified. Let’s assume that in the medical diagnosis domain there are four symptoms and two diseases. The symptoms are expressed by the conditions C1, C2, C3 and C4, and the diseases by the conclusions D1 and D2. We also know that D1 depends on C1, C2, C3 and D2 on C3, C4. Then, the following initial neurules are constructed: “(0) if C1 (0), C2 (0), C3 (0) then D1”, “(0) if C3 (0), C4 (0) then D2”. A zero initial value is assigned to each factor by default, except if the user assigns non-zero ones. For specific examples, see Section 6. 4.2 Training the initial neurules From the initial training data, we extract as many (sub)sets as the initial neurules. Each such set, called a training set, contains training examples in the form [v1 v2 … vn d], where vi, i= 1, …,n are their component values, which correspond to the n inputs of the neurule, and d is the desired output (‘1’ for success, ‘-1’ for failure). Each training set is used to train the corresponding initial neurule and calculate its factors. The learning algorithm employed is the standard least mean square (LMS) algorithm14. However, there are cases where the LMS algorithm fails to specify the right significance factors for a number of neurules. That is, the corresponding adaline units of those rules do not correctly classify some of the training examples in their training sets. This means that the training examples correspond to a non-separable (boolean) function. To overcome this problem, the training set of the initial neurule is divided into subsets in a way that each subset contains success examples, which are “close” to each other in some degree. The closeness between two examples is defined as the number of common component values. For example, the closeness of [1 0 1 1 1] and [1 1 0 1 1] is ‘2’. Also, we define as least closeness pair (LCP), a pair of success examples with the least closeness in a training set. There may be more than one LCP in a training set. Initially, a LCP in the training set is found and two subsets are created each containing as its initial element one of the success examples of that pair, called its pivot. Each of the remaining success examples are distributed between the two subsets based on their closeness to their pivots. More specifically, each subset contains the success examples which are closer to its pivot. Then, the failure examples of the initial set are added to both subsets, to avoid neurule misfiring. After that, two copies of the initial neurule, one for each subset, are trained. If the factors of a copy misclassify some of its examples, the corresponding subset is further split into two other subsets, based on one of its LCPs. This continues, until all examples are classified. This means that from an initial neurule more than one final neurule may be produced, which are called sibling neurules. So, step 2 of the algorithm for each initial neurule is analyzed as follows: 2.1 From the initial training data, produce as many initial training sets as the number of the initial neurules. 2.2 Train each initial neurule, by applying the LMS algorithm to its initial training set. If the calculated factors classify correctly all the examples, produce the corresponding neurule. Otherwise, find a LCP and produce two subsets of the initial set. In each subset put its pivot, the success examples of the initial training set which are closer to the pivot, and all the failure examples. In constructing the subsets, success examples with the same closeness to both pivots are put in the subset containing a success example with the greatest closeness to it, otherwise to the one with the least number of success examples. 2.3 For each subset do the following: 2.3.1 Perform training of a copy of the corresponding initial neurule and calculate its factors. 2.3.2 If the calculated factors misclassify examples belonging to the subset, further divide the subset into smaller subsets as in step 2.2 and apply step 2.3. 2.3.3 Produce the corresponding neurule. 5. The Inference Mechanism Although the focus of this paper is not on the inference of the system, for the sake of completeness, we concisely refer to it in this section. The hybrid inference mechanism (HIM) implements the way neurules co-operate to reach a conclusion. HIM gives pre-eminence to symbolic reasoning, which is based on a backward chaining strategy. As soon as the initial input data is given and put in the WM, the output neurules are considered for evaluation. One of them is selected for evaluation. Selection is based on textual order. A rule succeeds if the output of the corresponding adaline unit is computed to be ‘1’, after evaluation of its conditions (inputs). A condition evaluates to ‘true’, if it matches a fact in the WM, that is there is a fact with the same variable, predicate and value. A condition evaluates to ‘unknown’, if there is a fact with the same variable, predicate and ‘unknown’ as its value. A condition cannot be evaluated if there is no fact in the WM with the same variable. In this case, either a question is made to the user to provide data for the variable, in case of an input variable, or an intermediate neurule in NRB with a conclusion containing that variable is examined, in case of an intermediate variable. A condition with an input variable evaluates to ‘false’, if there is a fact in the WM with the same variable, predicate and different value. A condition with an intermediate variable evaluates to ‘false’ if additionally to the latter there is no unevaluated intermediate neurule in the NRB that has a conclusion with the same variable. Inference stops either when one or more output neurules are fired (success) or there is no further action (failure). In this process, because sibling neurules concern the same conclusion, if one of them fires, there is no need to evaluate any of the rest. To further increase inference efficiency, a number of heuristics are used12. 6. Examples and Experimental Results In this section, we present application of our algorithm for constructing neurule bases to two different sets of training data. More specifically, we present the construction process and the resulted neurule base in each case. Also, we compare three methods for choosing the LCP. 6.1 Fitting contact lenses The first data set was taken from a machine learning ftp repository15. It consists of 24 patterns (p1-p24) and concerns empirical data for fitting contact lenses (see Table 3). There are four input variables and one output variable. The input variables (with their possible values) are: age (young, pre-presbyopic, presbyopic), spectacle prescription (myope, hypermyope), astigmatic (no, yes), tear rate (reduced, normal). The output variable is: lenses-class (hard-lenses, soft-lenses, no-lenses). There are no intermediate variables, so there is no dependency information. Given that the output variable can take three possible values, we need three initial neurules, corresponding to the three possible conclusions. The output variable depends on all input variables, as empirical data shows. So, each initial neurule contains all the conditions related to the input variables. Each input variable produces as many conditions as its possible values. So, each neurule has nine conditions. The initial neurules are the same as the first three final neurules (see Fig. 3), except that the initial neurules have zero factors. Table 3. Data set for fitting contact lenses Pat. No age spect-pres astigmatic tear-rate lenses-class p1 young myope no reduced no-lenses p2 young myope no normal soft-lenses p3 young myope yes reduced no-lenses p4 young myope yes normal hard-lenses p5 young hypermetrope no reduced no-lenses p6 young hypermetrope no normal soft-lenses p7 young hypermetrope yes reduced no-lenses p8 young hypermetrope yes normal hard-lenses p9 pre-presbyopic myope no reduced no-lenses p10 pre-presbyopic myope no normal soft-lenses p11 pre-presbyopic myope yes reduced no-lenses p12 pre-presbyopic myope yes normal hard-lenses p13 pre-presbyopic hypermetrope no reduced no-lenses p14 pre-presbyopic hypermetrope no normal soft-lenses p15 pre-presbyopic hypermetrope yes reduced no-lenses p16 pre-presbyopic hypermetrope yes normal no-lenses p17 presbyopic myope no reduced no-lenses p18 presbyopic myope no normal no-lenses p19 presbyopic myope yes reduced no-lenses p20 presbyopic myope yes normal hard-lenses p21 presbyopic hypermetrope no reduced no-lenses p22 presbyopic hypermetrope no normal soft-lenses p23 presbyopic hypermetrope yes reduced no-lenses p24 presbyopic hypermetrope yes normal no-lenses The training sets of the initial neurules are extracted from the empirical data (Table 3) and are given in Table 4. In that table, the following correspondences are considered: age1 → ‘age is young’, age2 → ‘age is pre-presbyopic’, age3 → ‘age is presbyopic’, spec-pres1 → ‘spectacle-prescription is myope’, spec-pres2 → ‘spectacle-prescription is hypermetrope’, astig1 → ‘astigmatic is no’, astig2 → ‘astigmatic is yes’, tear-rate1 → ‘tear-rate is reduced’, tear-rate2 → ‘tear-rate is normal’, lenses-class1 → ‘lenses-class is hard-lenses’, lenses-class2 → ‘lenses-class is soft-lenses’, lenses-class3 → ‘lenses-class is no-lenses’. L1: (-2.4) if age is young (4.8), age is pre-presbyotic (-4.4), age is presbyotic (-4.5), spectacle is myope (-0.9), spectacle is hypermetrope (-2.1), astigmatic is no (-6.4), astigmatic is yes (3.3), tear-rate is reduced (-7.5), tear-rate is normal (4.8) then lenses-class is hard-lenses L2: (-2.4) if age is young (-0.5), age is pre-presbyotic (-0.3), age is presbyotic (-2.7), spectacle is myope (-4.0), spectacle is hypermetrope (0.9), astigmatic is no (2.9), astigmatic is yes (-6.4), tear-rate is reduced (-7.4), tear-rate is normal (4.4) then lenses-class is soft-lenses L3: (0.8) if age is young (-0.6), age is pre-presbyotic (-0.6), age is presbyotic (1.2), spectacle is myope (1.6), spectacle is hypermetrope (-0.2), astigmatic is no (2.7), astigmatic is yes (-2.0), tear-rate is reduced (4.4), tear-rate is normal (-4.6) then lenses-class is no-lenses L4: (-0.7) if age is young (-6.2), age is pre-presbyotic (1.6), age is presbyotic (3.2), spectacle is myope (-5.8), spectacle is hypermetrope (4.7), astigmatic is no (-4.1), astigmatic is yes (2.6), tear-rate is reduced (3.4), tear-rate is normal (-4.5) then lenses-class is no-lenses Table 4. Initial training sets for the contact lenses example lenses-class Ex. No age 1 age 2 age 3 spec- pres 1 spec- pres 2 astig 1 astig 2 tear- rate 1 tear- rate 2 1 2 3 p1 1 0 0 1 0 1 0 1 0 -1 -1 1 p2 1 0 0 1 0 1 0 0 1 -1 1 -1 p3 1 0 0 1 0 0 1 1 0 -1 -1 1 p4 1 0 0 1 0 0 1 0 1 1 -1 -1 p5 1 0 0 0 1 1 0 1 0 -1 -1 1 p6 1 0 0 0 1 1 0 0 1 -1 1 -1 p7 1 0 0 0 1 0 1 1 0 -1 -1 1 p8 1 0 0 0 1 0 1 0 1 1 -1 -1 p9 0 1 0 1 0 1 0 1 0 -1 -1 1 p10 0 1 0 1 0 1 0 0 1 -1 1 -1 p11 0 1 0 1 0 0 1 1 0 -1 -1 1 p12 0 1 0 1 0 0 1 0 1 1 -1 -1 p13 0 1 0 0 1 1 0 1 0 -1 -1 1 p14 0 1 0 0 1 1 0 0 1 -1 1 -1 p15 0 1 0 0 1 0 1 1 0 -1 -1 1 p16 0 1 0 0 1 0 1 0 1 -1 -1 1 p17 0 0 1 1 0 1 0 1 0 -1 -1 1 p18 0 0 1 1 0 1 0 0 1 -1 -1 1 p19 0 0 1 1 0 0 1 1 0 -1 -1 1 p20 0 0 1 1 0 0 1 0 1 1 -1 -1 p21 0 0 1 0 1 1 0 1 0 -1 -1 1 p22 0 0 1 0 1 1 0 0 1 -1 1 -1 p23 0 0 1 0 1 0 1 1 0 -1 -1 1 p24 0 0 1 0 1 0 1 0 1 -1 -1 1 Fig. 3. The neurule base for the contact lenses example Each of the three training sets consists of 24 examples. The examples in each of them have the same input patterns, but different output values, which are presented in Table 4 under the columns 1, 2 and 3 of the output variable ‘lenses-class’. For the first two initial neurules, the calculated factors successfully classified all training examples. The produced neurules L1 and L2 are presented in Fig. 3. However, it didn’t happen the same with the third initial neurule. So, from its initial training set two subsets were produced. There were six LCPs found (with closeness equal to ‘1’): (p1, p16), (p1, p24), (p7, p18), (p9, p24), (p15, p18) and (p16, p17). The first of them was chosen as the LCP. Then, from the third initial training set, two subsets were produced. The first, with pivot p1, included the success examples p3, p5, p9, p17, p18, p19 and p21, which were closer to p1 than to p16, and all failure examples. The second subset, with pivot p16, included the success examples p7, p11, p13, p15 and p24 and all the failure examples. Training of both copies of the third initial neurule was successful and two final neurules, L3 and L4, were produced (see Fig. 3). 6.2 Diseases of the sarcophagus The second set of training data was taken from14. It includes 8 patterns that concern acute theoretical diseases of the sarcophagus. According to the example in14, there are six symptoms (Swollen feet, Red ears, Hair loss, Dizziness, Sensitive aretha, Placibin allergy), two diseases (Supercilliosis, Namastosis), whose diagnosis is based on the symptoms and three possible treatments (Placibin, Biramibio, Posiboost). Table 5. Data set for the diseases of sarcophagus Pat. No Sym 1 Sym 2 Sym 3 Sym 4 Dis 1 Dis 2 Treat 1 Treat 2 Treat 3 p1 T F ×××× F T F T F T p2 F T T F F T T T F p3 T T F T T T F F F p4 F F T F F F F F F p5 ×××× T T T T T F T T p6 F T T F T T T T F p7 T F F T T F F F F p8 T F T T F T F F F For reasons that will become clear in Section 7, we omit the symptoms ‘Swollen feet’ and ‘Red ears’ as well as their related data. Thus, symptoms are reduced to four. Also, we consider that ‘Supercilliosis’ does not any more depend on symptoms, because they have been removed, but it is given as input information. The training data for our example are given in Table 5, where Sym1 → ‘Hair loss’, Sym2 → ‘Dizziness’, Sym3 → ‘Sensitive aretha’, Sym4 → ‘Placibin allergy’, Dis1 → ‘Supercilliosis’, Dis2 → ‘Namastosis’, Treat1 → ‘Placibin’, Treat2 → ‘Biramibio’ and Treat3 → ‘Posiboost’. Also, ‘T’and ‘F’ mean ‘true’and ‘false’ respectively and ‘××××’ means ‘unknown’. Finally, dependency information is provided (see Table 6), which shows the dependency between concepts. There is one input variable, namely symptom, which can take four possible values (the four symptoms). Also, there is one variable, called disease, which is both an input and an intermediate variable, depending on its value (see dependency information in Table 6). It can take two possible values (the two diseases). Finally, there is a last variable, treatment, which is both an intermediate and an output variable (see dependency information in Table 6) and can take three possible values (the three treatments). Because there are totally four possible values for the intermediate and output variables, four initial neurules are required. Table 6. Dependency information for the sarcophagus diseases problem Sym 1 Sym 2 Sym 3 Sym 4 Dis 1 Dis 2 Treat 1 Treat 2 Namastosis (Dis2) √√√√ √√√√ √√√√ Placibin (Treat1) √√√√ √√√√ √√√√ Biramibio (Treat2) √√√√ √√√√ √√√√ Posiboost (Treat3) √√√√ √√√√ The training sets for the four initial rules, which were extracted from the training data of Table 5, are presented in Tables 7-1 to 7-4. Notice that we didn’t use patterns including the ‘unknown’ value. Table 7-1. Training set for D1 HairLoss (Sym3) Dizziness (Sym4) Sensitive aretha (Sym5) Namastosis (Dis2) 1 0 1 1 0 1 1 1 1 0 0 -1 0 0 1 -1 1 1 0 1 Table 7-2. Training set for D2 Placibin allergy (Sym6) Supercilliosis (Dis1) Namastosis (Dis2) Placibin (Trea1) 0 0 1 1 0 1 0 1 1 1 1 -1 0 0 0 -1 0 1 1 1 1 1 0 -1 1 0 1 -1 Table 7-3. Training set for D3 HairLoss (Sym3) Superscilliosis (Dis1) Namastosis (Dis2) Biramibio (Trea2) 1 1 0 -1 0 0 1 1 1 1 1 -1 0 0 0 -1 0 1 1 1 1 0 1 -1 Table 7-4. Training set for D4-5 Placibin (Trea1) Biramibio (Trea2) Posiboost (Trea3) 1 0 1 1 1 -1 0 0 -1 0 1 1 Fig. 4. The neurule base for the sarcophagus diseases example D1: (-2.2) if Symptom is Dizziness (4.6), Symptom is SensitiveAretha (1.8), Symptom is HairLoss (0.9) then Disease is Namastosis D2: (-0.4) if Symptom is PlacibinAllergy (-5.4), Disease is Namastosis (1.8), Disease is Supercilliosis (1.0) then Treatment is Placibin D3: (-0.4) if Symptom is HairLoss (-3.6), Disease is Namastosis (1.8), Disease is Supercilliosis (1.0) then Treatment is Biramibio D4: (-0.4) if Treatment is Biramibio (-4.4), Treatment is Placibin (1.8) then Treatment is Posiboost D5: (-0.4) if Treatment is Placibin (-3.6), Treatment is Biramibio (1.0) then Treatment is Posiboost The calculated factors of all the initial neurules, except the last one, successfully classified all the training examples, even those containing the ‘unknown’ value, which were not used (generalization capability). Thus, three final neurules were produced (D1-D3 in Fig. 4). The fourth initial neurule, pertaining to the treatment Posiboost, failed to classify its respective set of training examples (Table 7-4), because they correspond to a non-separable function (XOR type). Thus, two subsets were created containing the first three and the last three examples respectively. Finally, two neurules were produced (D4, D5). 6.3 Choosing the least closeness pair A point of interest in training a neurule with a non-separable training set is how to choose a least closeness pair (LCP), in the process of producing the two subsets of the initial training set. Not all LCPs result in the same number of final neurules. So, we are looking for the pair that finally produces the minimum number of sibling neurules. We tried two heuristic methods for that. The best distribution method (BD) suggests choosing the pair that assures distribution of the two elements of the other pairs in different sets. So, examples with least closeness will be included in different sets, which may assure separability. The second, the mean closeness method (MC), computes the mean closeness of each of the two subsets to be created from each LCP. The mean closeness of a subset is the mean closeness of its examples. Then, calculates the mean closeness of the subsets created by each pair, which is the mean closeness of the two subsets, and chooses the pair with the greatest mean closeness. However, none of them faces all cases successfully. On the other hand, random choice method (RC) could be an alternative. In Table 8, results of using the above two heuristics and the random choice are presented. As random choice, we got the first LCP. The data used was indirectly taken from a medical rule base of 41 symbolic rules. We extracted training patterns from the symbolic rules by the method for training sets specification described in12. We did so, because we knew the optimal number of the neurules to be produced. In Table 8, Opt No indicates the optimal number of neurules and OLCPs the optimal LCPs, that is the LCPs that lead to the optimal number of final neurules. Table 8. Comparison of methods for the least closeness pair choice Conclusion Exam- ples Condi- tions LCPs OLCPs Opt No MC BD RC inflammation 72 8 7 5 2 2/3 2 3 arthritis 144 9 3 2 2 2 2/3 2 primary-malignant 120 10 8 3 2 2 2/3 2 secondary- malignant 72 7 2 1 2 2 2/3 3 early-inflammation 324 11 7 7 4 4 4 4 soft-tissue-early- bone- inflammation 288 11 2 2 2 2 2 2 early-soft-tissue- inflammation 270 11 2 2 3 3 3 3 As we can see from Table 8, none of the methods assures optimality of the number of the produced neurules in all cases. The expression ‘2/3’ means that the number of neurules can be 2 or 3. This is because there were more than one pairs that met the criterion of the method (e.g. had the same mean closeness or distributed the elements of the pairs in different subsets), but they weren’t all optimal. The MC method did a bit better than the BD method only in cases with a relatively small number of examples. However, random choice didn’t do bad, because,as a matter of fact, OLCPs is a large part, if not all, of LCPs. On the other hand, the MC method is computationally more expensive than the BD method and this latter than RC. So, given the computational effort required in the two heuristic methods, especially in the mean closeness, RC seems to be the best choice. Of course, some more experiments, with different rule bases, would give a more confident view on that. 7. Discussion and Related Work The main contribution of this work is the representation of non-separable empirical data in a hybrid, natural and modular way, for use in expert systems, in contrast to existing connectionist approaches. A method for representing non-separable training examples in a connectionist expert system is presented in14. It is called the “distributed method”. What that method does is to introduce a number of intermediate cells, between the inputs and the related output, called “distributed cells”, whose bias and weights are randomly generated. This extra layer between inputs and output makes representation of non-separable examples possible. The problem with that method is that those intermediate cells have no meaning, that is there are no concepts related to the problem assigned to them, as happens with the other nodes in a connectionist knowledge base. Also, there is no specific way to determine the number of the required intermediate cells. Thus, the resulted knowledge base is unnatural and complicated. Following the process for generating a connectionist knowledge base from empirical data described in14, we constructed the connectionist knowledge base corresponding to the data set for fitting contact lenses (see Section 6.1). According to the process, each node is individually trained. In case of non-separable training examples, the distributed method is used. The resulted (real) knowledge base is depicted in Fig. 5 and its corresponding neural network in Fig. 6, where we used real numbers for the weights and biases, instead of integers. The knowledge base in Fig. 5 is actually a matrix that represents the connections, their weights and the biases of the cells (concepts) in the network of Fig. 6. A zero at a position in the matrix shows that there is no connection between the input cell (variable) of its column and the intermediate or output cell (variable) of its row. In the network of Fig. 6, there are three output cells (the cyclic ones) representing the three outputs (conclusions), three intermediate (distributed) cells (the triangles) introduced for representing the non-separable training examples of the ‘no-lenses’ training set and nine input cells representing the nine input values. For readability reasons, we didn’t draw all the connections neither put all the weights on the net of Fig. 6. Actually, all inputs are connected to all intermediate and all output cells and the outputs of all the intermediate cells are connected to the output cells. Lenses class 1 -13.2 8.4 1.0 0.9 0.9 -2.1 -6.4 5.1 -5.7 4.8 0 0 0 Lenses class 2 -11.4 3.1 3.3 2.7 -4.0 2.7 2.9 -4.6 -7.4 6.2 0 0 0 Intermediate variable 1 3.2 0 -0.8 -3.6 7.8 -6.5 7.7 -6.7 -11 10.4 0 0 0 Intermediate variable 2 -4.0 -3.6 2.8 3.6 4.2 -2.9 4.1 -3.1 7.0 -7.6 0 0 0 Intermediate variable 3 -7.6 0 6.4 0 0.6 0.7 0.5 0.5 -0.2 -0.4 0 0 0 Lenses class 3 5.0 -5.4 -2.6 1.8 -1.2 2.5 -1.3 2.3 1.6 -2.2 -9.6 9.8 -5.3 Bias age 1 age 2 age 3 spec -pr1 spec -pr2 astig 1 astig 2 tear- r1 tear- r2 Inter 1 Inter 2 Inter 3 Fig. 5. The connectionist knowledge base for the contact lenses example age 1 age 2 age 3 spec- pr1 spec- pr2 astig 1 astig 2 tear- r1 tear- r2 Fig. 6. The neural network for the contact lenses example A comparison of the knowledge base in Fig. 5 to the one in Fig. 3 demonstrates the advantages of neurules. It is clear that the benefits of symbolic rule-based representation, such as naturalness and modularity are retained. Neurules are understandable, since significance factors represent the contribution of corresponding conditions in drawing the conclusion. On the other hand, the connectionist knowledge base is a multilevel network with some meaningless intermediate units. Thus, it lacks the naturalness of neurules. The corresponding connectionist knowledge base for the diseases of the sarcophagus example is depicted in Fig. 7. It is a modified version of that presented in14, to fit our modified example. The corresponding network is a multi-level network with three distributed cells, introduced by the training algorithm. Let’s suppose now that we get some new knowledge, which says that ‘Supercilliosis’ should not be given as input information, but it will be produced as an intermediate conclusion. Also, that it depends on ‘Hair loss’ and two new hard-lenses (lenses class 1) soft-lenses (lenses class 2) no-lenses (lenses class 3) Inter1 Inter2 Inter3 5.0 -11.4-13.2 -5.4 -2.6 -2.2 1.6 -0.8 3.2 -4 -7.6 symptoms (inputs), namely ‘Swollen feet’ and ‘Red ears’, according to the (training) data given in Table 9. Namastosis -1 3 3 3 0 0 0 0 0 0 0 0 0 Placibin -2 0 0 0 -4 2 2 0 0 0 0 0 0 Biramibio -1 -4 0 0 0 1 3 0 0 0 0 0 0 Intermediate variable 1 2 0 0 0 0 0 0 -4 5 0 0 0 0 Intermediate variable 2 3 0 0 0 0 0 0 -2 2 0 0 0 0 Intermediate variable 3 0 0 0 0 0 0 0 -1 -3 0 0 0 0 Posiboost 3 0 0 0 0 0 0 -3 1 -3 -3 -1 0 Bias Sym 1 Sym 2 Sym 3 Sym 4 Dis 1 Dis 2 Treat 1 Treat 2 Int 1 Int 2 Int 3 Treat 3 Fig. 7. The connectionist knowledge base for the diseases of the sarcophagus example. Table 9. Training data for Supercilliosis HairLoss (Sym1) SwollenFeet (Sym5) RedEars (Sym6) Supercilliosis (Dis1) 1 1 1 1 0 0 0 -1 1 0 0 1 0 1 1 -1 0 1 0 1 1 0 1 -1 To introduce this new knowledge to our neurule base, we train a neurule with three conditions, corresponding to the three symptoms of Table 8, and a conclusion related to ‘Supercilliosis’. The result is the final neurule D6 depicted in Fig. 8, which is put into the neurule base. To introduce that knowledge into the connectionist knowledge base of Fig. 7, not only training of a new unit is needed, but also modifications should be made to the knowledge base. More specifically, a new row (concerning ‘Superscilliosis’) and two new columns (concerning ‘Swollen feet’ and ‘Red ears’) should be added. Fig. 8. The new neurule concerning ‘Supercilliosis’. D6: (-0.4) if Symptom is RedEars (-4.4), Symptom is SwollenFeet (3.6), Symptom is HairLoss (2.7) then Disease is Supercilliosis Furthermore, one can easily add new neurules to or remove old neurules from a neurule base without making any changes to the knowledge base, since neurules are functionally independent units, given that they do not affect existing knowledge. Thus, a type of incremental development of the knowledge base is still supported, although by larger knowledge chunks. This corresponds to introducing one or more networks in an existing connectionist knowledge base sharing or not inputs and/or intermediate cell outputs. This is either difficult or impossible to do. 8. Conclusions In this paper, we introduce a method for generating neurules, a kind of hybrid rules, from empirical data of binary type. Neurules integrate neurocomputing and production rules. Each neurule is represented as an adaline unit. Thus, the corresponding rule base consists of a number of neurules (autonomous adaline units). In this way, the produced neurule base retains the modularity of symbolic rule bases. Also, it retains their naturalness, since neurules look much like symbolic rules. Furthermore, incremental development is still supported. This is in contrast to existing connectionist knowledge bases, which are not modular and thus do not actually offer incremental development. A difficult point in our approach is the inherent inability of the adaline unit to classify non-separable training examples. We overcome this difficulty by introducing the notion of ‘closeness’, as far as the training examples are concerned. That is, in case of failure, from the training set of the neurule two subsets of ‘close’ examples are produced and two copies of the neurule are trained. Failure of any copy training leads to further subsets production until success is achieved. A weak point of the neurules is the fact that we have multiple representations of the same knowledge, in case of sibling rules. Given the capability of producing neurules from empirical binary data and their advantages over symbolic rules, as far as inference efficiency and the rule base size are concerned12, we can argue that neurules are more suitable for representing knowledge in web-based intelligent tutoring systems than symbolic rules. This is our current continuation on this research. Acknowledgements This work was partially supported by the GSRT of Greece, Program ΠΕΝΕ∆’99, Project No Ε∆234. References [1] L. M. Fu (Ed), Proceedings of the International Symposium on Integrating Knowledge and Neural Heuristics (ISIKNH’94), Pensacola, FL (May 1994). [2] R. Sun and E. Alexandre (Eds), Connectionist-Symbolic Integration: From Unified to Hybrid Approaches, Lawrence Erlbaum (1997). [3] M. Hilario, An Overview of Strategies for Neurosymbolic Integration, ch.2 in [2]. [4] L-M Fu and L-C Fu, Mapping rule-based systems into neural architecture, Knowledge- Based Systems 3 (1990) 48-56. [5] F. Kozato and Ph. De Wilde, How Neural Networks Help Rule-Based Problem Solving, Proceedings of the ICANN’91 (1991) 465-470. [6] Tan C.L. and T. S. Quash, Implementation of rule-based expert systems in a neural network architecture, The World Congress on Expert Systems Proceedings (1991) 1843- 1851. [7] Kuncicky D. C., S. I. Hruska and D. C. Lacher, Hybrid Systems: The equivalence of rule- based expert system and artificial neural network inference, International Journal of Expert Systems, (1992) 4(3) 281-297. [8] S. I. Gallant, Connectionist Expert Systems, CACM, 31 (1988) 152-169. [9] B. Boutsinas and M. N. Vrahatis, Nonmonotonic Connectionist Expert Systems, Proceedings of the 2nd WSES/IEEE/IMACS International Conference on Circuits, Systems and Computers, Athens, Hellas (Oct. 1998). [10] A. Z. Ghalwash, A Recency Inference Engine for Connectionist Knowledge Bases, Applied Intelligence 9 (1998) 201-215. [11] I. Hatzilygeroudis, J. Prentzas, Neurules: Integrating Symbolic Rules and Neurocomputing, in D. Fotiades and S. Nikolopoulos (Eds), Advances in Informatics, World Scientific Pub., 2000, 122-133. [12] I. Hatzilygeroudis and J. Prentzas, Neurules: Improving the Performance of Symbolic Rules, International Journal on Artificial Intelligence Tools (IJAIT), 9(1) (2000) 113- 130. [13] I. Hatzilygeroudis and J. Prentzas, Producing Modular Hybrid Rule Bases for Expert Systems, Proceedings of the 13th International FLAIRS Conference, Orlando, FL (May 2000) 181-185. [14] S. I. Gallant, Neural Network Learning and Expert Systems, MIT Press (1993). [15] ftp://ftp.ics.uci.edu/pub/machine-learning-databases/ 
work_krrqblgh6fb2jce4fr2c4f4o7q	----	Analysis of healthcare coverage: A data mining approach Analysis of healthcare coverage: A data mining approach Dursun Delen *, Christie Fuller, Charles McCann, Deepa Ray William S. Spears School of Business, Department of Management Science and Information Systems, Oklahoma State University, North Classroom Building #378, 700 North Greenwood Avenue, Tulsa, OK 74106, USA Abstract The existing disparity in the healthcare coverage is a pressing issue in the United States. Unfortunately, many in the US do not have healthcare coverage and much research is needed to identify the factors leading to this phenomenon. Hence, this study aims to examine the healthcare coverage of individuals by applying popular machine learning techniques on a wide-variety of predictive factors. Twenty- three variables and 193,373 records were utilized from the 2004 behavioral risk factor surveillance system survey data for this study. The artificial neural networks and the decision tree models were developed and compared to each other for predictive ability. The sensitivity analysis and variable importance measures are calculated to analyze the importance of the predictive factors. The experimental results indicated that the most accurate classifier for this phenomenon was the multi-layer perceptron type artificial neural network model that had an overall classification accuracy of 78.45% on the holdout sample. The most important predictive factors came out as income, employment status, education, and marital status. Using two popular machine learning techniques, this study identified the factors that can be used to accurately classify those with and without healthcare coverage. The ability to identify and explain the reasoning of those likely to be without healthcare coverage through the application of accurate classification models can potentially be used in reducing the disparity in healthcare coverage. � 2007 Elsevier Ltd. All rights reserved. Keywords: Healthcare coverage; Data mining; Prediction; Classification; Neural networks; Decision trees 1. Introduction Healthcare coverage in general and the existing disparity in this coverage in specific is a pressing issue in the United States. Many in the US do not have healthcare coverage, and much research has been conducted, and much more is needed to identify the factors leading to this disparity in coverage. Previous work has identified two key situa- tions where understanding of these factors is beneficial (Cunningham & Ginsburg, 2001). First, given that these factors exist at both the state and local level, it is imperative that the individuals responsible for funding decisions cor- rectly interpret the reasons that the uninsured rates may be elevated. Second, it is important that the individuals are able to determine if the uninsured rates may be elevated simply due to unchangeable characteristics within the population. While existing research has identified factors which dif- fer between those with and without coverage, it has not progressed to building discriminatory models to separate them from each other. Further, despite identification of dif- ferences on some factors, the disparity has not been reduced (Glover, Moore, Probst, & Samuels, 2004). This study advances the current research by building classifica- tion models to identify those belonging to each group – those that do and those that do not have healthcare cover- age. Such a model may eventually be used to help reduce healthcare coverage disparity. Similar techniques have pre- viously been introduced in a more general manner for use in targeting customers in the insurance industry (Wu, Kao, Su, & Wu, 2005). Identifying those without healthcare coverage is impor- tant, as those without coverage may have reduced access to medical care (Monheit & Vistnes, 2000) and may have 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.10.041 * Corresponding author. Tel.: +1 918 594 8283; fax: +1 918 594 8281. E-mail address: dursun.delen@okstate.edu (D. Delen). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 36 (2009) 995–1003 Expert Systems with Applications mailto:dursun.delen@okstate.edu more preventable hospitalizations (Services USDoHaH, 2003). Lack of healthcare coverage has also been linked to poor health and early death. This issue has been exacer- bated by the clear increasing trend of the number of unin- sured in the last 20 years (Cunningham & Ginsburg, 2001; Herring, 2005). From 1989 to 1999, there was an 18.4% increase in the number of people without insurance. The level of uninsured has grown to approximately 16% of the population as of 2003 (Jonk et al., 2005). It has been estimated that 60 million people are uninsured at some point during a given year (Cunningham & Ginsburg, 2001). The factors leading to the increasing number of uninsured have been studied at both the state and local level (Cunningham & Ginsburg, 2001), and across many socio-demographic factors (Hendryx, Ahern, Lovrich, & McCurdy, 2002; Holtz-Eakin, 2002). Lifestyle factors may also contribute to this problem. Gender is one socio-demographic variable that has been found to be related to healthcare coverage (Carrasquillo, Carrasquillo, & Shea, 2000; Monheit & Vistnes, 2000; Shi, 2000). Many studies have found that females are less likely to be insured than males (Hendryx et al., 2002; Hol- tz-Eakin, 2002; Monheit & Vistnes, 2000), though the opposite has also been found (Carrasquillo, Himmelstein, Woolhandler, & Bor, 1999; Nelson, Bolen, Wells, Smith, & Bland, 2004). Race or ethnicity may also be a factor con- tributing to healthcare coverage disparity (Carrasquillo et al., 2000; Lucas, Barr-Anderson, & Kington, 2003; Mon- heit & Vistnes, 2000), with minorities generally having less healthcare coverage (Carrasquillo et al., 1999; Glover et al., 2004; Monheit & Vistnes, 2000). According to the results of multiple studies, those with lower incomes are less likely to have healthcare coverage (Cardon & Hendel, 2001; Car- rasquillo et al., 1999; Lucas et al., 2003), and having health- care coverage also has a relationship with employment status (Schmidt & Deichert, 1996) and type of employment (Krieger, Barbeau, & Soobader, 2005). There is also a dif- ference among states, region of the country, or even county in the rate of healthcare coverage (Cardon & Hendel, 2001; Carrasquillo et al., 1999; Nelson et al., 2004; Schmidt & Deichert, 1996). For example, in the Northeast and Mid- west, one is more likely to be insured while in the South or the West, one is more likely to be uninsured (Cardon & Hendel, 2001). Some studies suggest that younger adults have less insurance (Cardon & Hendel, 2001; Carrasquillo et al., 1999). Also, studies (Jonk et al., 2005; Woolhandler et al., 2005) that looked at differences between veterans and non-veterans found that fewer veterans had less insurance relative to the remaining population. Education and mari- tal status have also been found to have a relationship with this insurance disparity (Shi, 2000). Disabled Americans may lack coverage (Landerman et al., 1998). For some populations studied, the role of household size in this dis- parity has also been examined (Glover et al., 2004; Pol, Mueller, & Adidam, 2002). In addition to the socio-demographic factors described above, lifestyle may also play a part in the existing dispar- ity. Those that are already in poor health may have reduced coverage (Cunningham & Ginsburg, 2001; Glover et al., 2004). These studies did not specify whether poor health included mental and/or physical health. Smoking status has been studied in relation to type of insurance cov- erage held (King & Mossialos, 2005). Exercise and alcohol consumption are additional lifestyle variables that have previously been used in classification models for insurance policy purposes (Chae, Ho, Cho, Lee, & Ji, 2001). Further, the extent to which someone is a risk taker may affect whether they secure healthcare coverage (Cunningham & Ginsburg, 2001). Past studies have often looked at a subgroup, such as the near-elderly (Monheit, Vistnes, & Eisenberg, 2001) or immi- grants (Herring, 2005), of the overall US population and have succeeded in identifying variables that seem to contrib- ute to the disparity in healthcare coverage for the subgroup. This study will look at healthcare coverage disparity across the population of the US, rather than within a smaller sub- group of the population. It will also address the numerous possible contributing factors, both socio-demographic and lifestyle, and their contribution to the growing disparity in healthcare coverage. Further, the study will utilize machine learning techniques in building classification models. Previ- ously, the issue of healthcare coverage disparity has been studied using primarily statistical techniques such as logistic regression (Glover et al., 2004) and basic descriptive statis- tics (Shi, 2000; Woolhandler et al., 2005). For many years linear regression has been the primarily used technique in capturing and representing functional relationships between dependent and independent variables, largely because of its well-known statistically explainable optimiza- tion strategies. However, in many problem scenarios, the model accuracy suffers as the assumed linear approximation of a function is not valid. With current technology, machine learning techniques can easily model such scenarios as healthcare coverage, as is addressed in this paper. These techniques are not constrained by the Gauss–Markov assumption (such as multicollinearity and normality) which is a major concern for more traditional models (Uysal & Roubi, 1999). Previously, these techniques have been used to study other healthcare issues such as factors affecting inpatient mortality (Chae, Kim, Tark, Park, & Hoa, 2003) and influencing prenatal care (Prather et al., 1997). In this study, we have attempted to build an accurate classification model, using machine learning techniques. The model could then be used to predict whether or not an individual has healthcare coverage based on specific socio-demographic and lifestyle information as well as the importance of the various factors in the model. 2. Research method 2.1. Data The data source that was used for this project was the Behavioral Risk Factor Surveillance System 2004 Survey 996 D. Delen et al. / Expert Systems with Applications 36 (2009) 995–1003 https://isiarticles.com/article/21430 
work_krxbjzmlv5d4bkit6hg6hdi6mi	----	International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-4, April 2020 19 Retrieval Number: C6397029302/2020©BEIESP DOI: 10.35940/ijeat.C6397.049420 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Knowledge based Expert System for Predicting Diabetic Retinopathy using Machine Learning Algorithms J.Jayashree, Sruthi R, Ponnamanda Venkata Sairam, J.Vijayashree Abstract: Diabetic retinopathy (DR) is a medical condition that can affect the patient's retina and cause leaks in the blood due to diabetes mellitus. The increase in cases of diabetes limits existing manual testing capability. Today new algorithms are becoming very important for assisted diagnosis. Effective diabetes diagnosis can benefit the victims and reduce the negative harmful effects, including blindness. If not treated in a timely manner, this disorder can cause different symptoms from mild vision problems to total blindness. Early signs of DR are the hemorrhages, hard exudates, and micro-aneurysms (HEM) that occur in the retina. Timely diagnosis of HEM is important for avoiding blindness This paper presents PSO feature selection algorithms with three classifications for the detection of Diabetic retinopathy using python. Keywords: Diabetic retinopathy, feature selection, classification, Complications, Treatment, Prevention, Statistics. I. INTRODUCTION Diabetic retinopathy is the undergoing changes that take place in blood sugar levels throughout the capillary of the retinal system. Some vessels may swell up in some cases, and fluid leaks into the back of the eye. These could swell and drop in the capillary. Or they could close, blocking the flow of blood. Anomalous new capillary often grow up on the retina. All these improvements will rob your eyesight. DR was not leading symptoms initially, just low vision complications. Ultimately, that it will lead blindness. Whoever also type 1 and type 2 diabetic can develop the condition. The you have far more diabetes, and the less sugar on your blood regulated, the greater the probability that you will experience this eye complication. In other cases abnormal arteries will grow on the ground of the retina. Over time, too much blood sugar will contribute to blocking the tiny capillary that feed the retina and sever blood supply. As a consequence, the eye looks for new capillary development A. Types of diabetic retinopathy There are two types of diabetic!retinopathy: Early diabetic!retinopathy Commonly known as -non proliferative diabetic retinopathy (NPDR) which occurs when there isn’t growth/proliferating of new capillary. Revised Manuscript Received on April 25, 2020. J. Jayashree, School of Computer Science and Engineering, VIT,Vellore vijayashree.j@vit.ac.in Sruthi R, School of Computer Science and Engineering, VIT,Vellore Ponnamanda Venkata Sairam, School of Computer Science and Engineering, VIT,Vellore J. Vijayashree, School of Computer Science and Engineering, VIT,Vellore That is the initial phase of Diabetes eye disease. The walls of the capillary within the retina weaken when you have NPDR. Smaller bulges (microaneurysms) extend down from the walls of the smaller vessels, frequently withering fluid and blood into another retina. Greater retinal shafts, too, may start dilating and radius is abnormal. NPDR can switch from mild to severe, as more siege capillary. Capillary within the eye can also close off with NPDR. This is called ischemia macular. When this occurs, the macula cannot be penetrated by blood. Occasionally, small particles, termed exudates, could perhaps form in the retina. Nerve fibers in the retina can start swelling. Central segment of the retina (macula) sometimes starts swelling (macular edema), an ailment which needs medical attention. Advanced!diabetic retinopathy, known as proliferative diabetic!retinopathy, can progress to this more serious type. In this case, damaged capillary narrow off, brings new prosperity, irregular capillary of retina, and may dip into the transparent, fliud-like fluid that filling your (vitreous) middle of eye. PDR is the most advanced stage of eye disease for diabetics. It happens when new capillary start to grow in the retina. Neovascularization is called this. Often those delicate bleeding current vessels into the pigment particles. You might see some gloomy gnats, when they bleed a little. When it spills a lot, then all vision could be blocked. Finally scar tissue, eventually aroused by the development of new capillary, can induce the retina to divide from either the rear in your eye. Unless the new capillary interacts with ordinary fluid flow out from the eye, stress in the eye ball will accumulate. This can disrupt the nerves that brings stimuli (optic nerve) of your eye to your brain, contributing to macular degeneration.The body's effort to save its retina is proliferative retinopathy, but it can often lead to retina scarring and can cause the retina to detach itself, leading to blindness. Modern eye care can help prevent blindness from arising as a result of proliferative retinopathy. B. Stages of diabetic retinopathy STAGES DESCRIPTI ON IMAGE 1 Mild Nonproliferative Retinopathy Microaneurysms occur at this stage. These are small pockets of globular swelling in the relatively small capillary of the eye. mailto:vijayashree.j@vit.ac.in Knowledge based Expert System for Predicting Diabetic Retinopathy using Machine Learning Algorithms 20 Retrieval Number: C6397029302/2020©BEIESP DOI: 10.35940/ijeat.C6397.049420 Published By: Blue Eyes Intelligence Engineering & Sciences Publication 2 Moderate Nonproliferative Retinopathy This is the phase where blocking of capillary occur. 3 Severe Nonproliferative Retinopathy The capillary which helps for the nourishment of eye are blocked thus signalling the retina to grow new capillary. 4 Proliferative Retinopathy Fresh capillary were proliferating, expanding within the retina, and into the vitreous gel. C. Symptoms The initial stages of diabetic retinopathy will occur without signs or discomfort like many conditions of this nature. There will not be any direct effect on the vision when sickness progresses. Macular oedema might led to maculopathy and affects vision if leakage induces swelling of the macular fluid. Signs become obvious when the disease progresses, the normal retinopathy side effects to be observed also include:  Blurry vision  Spots and floaters in eye  Double vision  Eye pain D. Risk Factors Someone living with diabetes may develop diabetic retinopathy. This may increase the probability of cultivating the eye condition:  High blood sugar level  High blood pressure  Higher levels of protein content in urine  Raised obesity in the blood  High levels of cholesterol  Use of Tobacco Whoever has diabetes may grow diabetic retinopathy and other diabetes problems. The more a patient seems to get diabetes, the higher the risk of developing diabetic retinopathy. In addition, patients should always be informed that a rapid increase in blood glucose levels will result to retinopathy which is worse. In this scenario, a massive increase in blood sugar levels characterized by either a 30 mmol / mol or 3 percent reduction in HbA1c. E. Complications of DR Diabetic retinopathy includes the development of irregular blood vessels within that retina. Abnormalities can cause major problems regarding vision::  Vitreous!haemorrhage. The fresh capillary can bleed into the fresh, creamy-like stuff, covering the middle of your eye. Where the rate of leakage is low, you can only see some dark spots (floaters). Blood will fills the vitreous cavity in more severe cases and effectively block your vision. If the vitreous humor shrinks, these capillary can be weakened, causing them to bleed, which can contribute to the appearance of cobwebs in your eyes and make it harder to see. Blood from a vitreous haemorrhage can dissipate, but any complications would require medical attention. Vitreous haemorrhage on its own does not usually cause irreversible loss of vision. Often the blood clears within a few weeks or months from the eye. Without harm to your retina, the vision can revert to its former clarity.  Retinal!tightening. The enlarged capillary correlated with macular degeneration facilitates the development of scar tissue which may remove the retina from the back of the eye.  Glaucoma. New capillary must develop at the front of of your eye and collide to your eye's ordinary fluid flow, allowing pressure to build up in your eye (glaucoma). The above pressure may disrupt the nerve which carries images of your eye in your nervous system (antenna nerve).  Blindness. Diabetic retinal detachments, cataracts or both ultimately result in total vision lost. Treatment may include one or more of:  Laser therapy – To help new capillary rising  Anti-VEGF treatments – prevents the growth of new capillary but is a more expensive treatment. F. Prevention Diabetic retinopathy is not always preventable. Regular eye tests, handling stable blood glucose and heart rate, and early intervention of vision problems may help prevent serious loss of vision, however. Decrease the level your risk of diabetic retinopathy when you suffer from diabetes by doing the following:  Monitor your diabetes: Consider the daily routine part of healthy eating and physical activity. Seek minimum 150 minutes of medium aerobic exercise every week, for example walking. Take drugs or insulin for oral diabetes as prescribed.  Track the blood glucose level: You might need blood sugar monitoring and log them out many times a day— you may need more regular tests when you happen to be sick or under tension. Request your physician how frequently you need to check sugar in blood.  Keep levels of cholesterol control balance: Healthy eating, doing daily workouts and dropping muscle mass will help. Medication is also often required.  Quit Smoking: Smoking will increase the risk of several diabetes problems.  Beware of shifts in vision. Whether you notice sudden changes in vision, or your vision is blurred, spotty or hazy, call your eye doctor right away. International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-4, April 2020 21 Retrieval Number: C6397029302/2020©BEIESP DOI: 10.35940/ijeat.C6397.049420 Published By: Blue Eyes Intelligence Engineering & Sciences Publication G. Statistics Background retinopathy Among diabetes sufferers some type of retinopathy is normal. A 2002 Royal Liverpool University Hospital study examined 820 type 1 diabetes patients and 7,271 type 2 diabetes patients and made the following results:  Some form of retinopathy was found in 46 percent of individuals with type one diabetes  Some levels of the retinopathy was observed in 25.3 percent of individuals with type 2 diabetes The greater occurrence of retinopathy in seen among individuals having type 1 than those with type 2 diabetes. Table –II: Zone wise distribution of monitored diabetic patients, and area wise occurrence. At 194 centers, Diabetics found voluntarily assessed by citizens of society using a formal procedure that was given for evaluation by society. The findings were analyzed to assess the occurance of DR in the sample population and to classify age-related and historical risk factors such as length of diabetes, use of insulin, and other end-organ diseases using the Chi-square method. A total of 6218 diabetics known to have been screened. In total, 5130 forms of data entry were deemed suitable for further review. Approximately 61.2 percent were males, 88.6 percent were between the ages of 40 and 80, nearly two-thirds of patients were from the western and southern zones, and more than half had diabetes over 5 years. Table III. P Occurance of diabetic retinopathy about diabetes period mellitus. Table -IV Occurance of diabetic retinopathy in patients with other end--organ disease. The effect of diabetic retinopathy on people who have other organ diseases is quite common. Chart -I. Zone-wise occurance of diabetic retinopathy; This represents the commonness of DR in patients in India. Table -V. Distribution of DR II. LITERATURE REVIEW Franklin et al.(2014), This research work provides a technique for segmentation of retinal vessels that can be used in retinal image analysis of machines. This experimental technique can be a pre-screening tool to detect diabetic retinopathy early on. The methods used in analysis can be used anonymously in retinal images to classify and interpret vascular structures. Fleming et al,2007 , Computerized image processing is extensively followed to shorten the task of grading images resultant from the diabetic retinopathy screening programs. Attempting to correct exudates in retinal images is a primary goal for automatic identification being one of the indicators that the disease has advanced to a point that needs to be listened to as an ophthalmologist. Diabetic images and normal images are taken from a fundus camera, processed and analyzed for back-propagation on a computer using a neural network. The network had been Instructed in acknowledging functionalities of the retinal image. It evaluated the impacts of numerous network variables and automatic filtering strategies. Knowledge based Expert System for Predicting Diabetic Retinopathy using Machine Learning Algorithms 22 Retrieval Number: C6397029302/2020©BEIESP DOI: 10.35940/ijeat.C6397.049420 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Diabetic and normal images were then randomized to assess network performance. The model shows 88.4% sensitivity and 83.5% specificity .Gardner et al.1996) This paper attempted to detect exudates using neural network back propagation (Karegowda et al.,2011). The publicly available DIARETDB1 dataset for diabetic retinopathy was used in the assessment process. The optic disk is neglected to avoid the optic disk from interfering with detection of exudates. The model shows Sensitivity of 93.97 %, Specificity of 90% and Accuracy of 94.65%. Dupas et al.(2010) Automated microaneurysm and exudate identification was functionalise to two small image databases storage that manually marked those lesions on. A computer-based diagnostic system was then developed and tested for DR and ME risk recognition and rating, use of such a huge database comprising both ordinary and compulsive images and auto gradation comparison. Preprocessing of Comparison improvement is extended until four features are collected as input frequency variables, normal frequency differentiation, hue and several angle pixel resolution, to provide coarse segmentation as data variables utilizing FCM clustering tool (Sopharak et al.,2009).. The model showed a Sensitivity of 90.28%, specificity of 93.24% and Accuracy of 89.11%. This suggests an automated procedure for the identification of rough exudates, a lesion associated with diabetic retinopathy. Using a statistical description, the algorithm based on their color and their edges, adding an edge identifying to address them. In this way, we test the method's robustness to make it suitable to a clinical setting. The model showed a sensitivity of 79.62% (Sanchez et al., 2004) Rao et al.(2015) Among the difficult and important elements of managing primary open angle glaucoma (OAG) is identifying glaucoma progression. It is caused by pressure accumulation inside the eye. Detecting glaucomatous progression is important and demanding of handling firstly open angle glaucoma (OAG). The model showed an accuracy of 90.6%. An effective method for identifying exudates as hard and soft exudates in this article(Rajput et al., 2014). To remove noise, the retinal photograph in the color space of CIELAB is processed. Then after, the network of capillary is separated to allow the identification and removal of optic disks. Using Hough transform method, the optic disks are removed. The victim exudates are identified by using k- means clustering algorthim. The model showed a sensitivity of 95.92% and accuracy of 99.70%. Classification schemes are developed and tested to deduct the presence or absence of DR. The detection rule is depends up on the problem of binary-hypothesis testing which clarifies yes / no decisions with the question. It also shows an overview of the Bayes output optimality criterion applied to DR. On the real-world data, the proposed DSS is evaluated. The model showed specificity of 67%(Kahai et al., 2006) Prasad et al. (2015) This analysis suggests that usage of morphological techniques and methods of segmentation to identify the capillary, microaneurysms and hard & soft exudates. The representation of the retinal fundus is split into sub frames. Diverse attributes were derived through the image of the retinal fundus. On the extracted features hair wavelet transformations are applied. The main technique for the analysis of components is then applied for better selection of features. For the detection and classification of the images as diabetic or non-diabetic, neural network back propagation techniques were employed. The model showed an accuracy of 93.8%. This paper explores and suggests a optimally modified morphological operators technique to be used on the low- contrast images of patients with diabetic retinopathy for exudate detection (Sopharak et al.,2008). These automatically observed exudates are confirmed as compared with the hand-drawn ground-truths of professional ophthalmologists. The model showed a sensitivity and specificity is 80% and 99.5%. Kwasigroch et al.(2018) To improve system performance, We suggested a separate class coding methodology that enabled details to be included on the value of the discrepancy here between expected performance and the targeted performance in the subjective function monitored during neural network training. We used normal precision measurements and a quadratic weighted Kappa score to check classification capacity of the employed models. The model showed an accuracy of about 82%. We introduced a brief structure to the convolutionary neural network architecture by emerging a pre-processing layered and convolution layer for maximise the output for the convolutionary neural network classifier (Khojaste et al.,2018). Two image enhancement techniques such as Contrast enhancement technique and Adaptive histogram equalization with a minimal contrast technique. The model showed an accuracy of 87.6%. Priya et al.(2012)This paper proposed two models such as Probabilistic Neural Network and Support Vector Machine to diagnose diabetic retinopathy, and compares their efficiency. When diabetes progresses, a patient's vision can initialized to deteriorate and lead to diabetes retinopathy. Two classes are established, namely nonproliferative diabetes retinopathy and proliferative diabetes retinopathy. PNN has an accuracy of 92.5% and SVM has an accuracy of 93 %. The project's aim is to identify retinal micro-aneurysms and exudates using classifier for automated DR screening(Gupta et al., 2015). It is necessary to implement an automated DR screening system for detecting dark lesions and bright lesions in photographs of digital funds. To detect retinal micro-aneurysms and exude images from retinal funds. The model showed a sensitivity and specificity of 87% and 100%, accuracy of 86%. Geetharamani et al.(2017) Diabetic!Retinopathy, a primary leads of blindness, is discussed in this study. For define Diabetic Retinopathy, a two-tier system is adopted. The suggested technique is evaluated by the UCI Machine Learning Repository on Diabetic Retinopathy Drebechen Dataset. The evolved rules are evaluated and the best rules are created through 3 fold cross validation. Diabetic!Retinopathy is an eye disease caused by diabetes over the huge term (Athira et al.,2019). In our paper we suggest an approach for the diagnosis of DR from R-CNN (Regional Convolution Neural Network) digital fundus images. R-CNN is highly accurate and resistant to lesion detection. Throughout our work, we have implemented a new strategy, where the whole picture was segmented and only the regions of interest were taken for further analysis. International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-4, April 2020 23 Retrieval Number: C6397029302/2020©BEIESP DOI: 10.35940/ijeat.C6397.049420 Published By: Blue Eyes Intelligence Engineering & Sciences Publication In our process we used 10 layers of R-CNN, trained it on 129 fundus images and checked 111 images on them.Both images were divided into two categories, i.e., with DR and without DR. This R-CNN (Regional CNN) approach was found to be fast and accurate with an accuracy of approximately 94%. Sopharak et al. (2008) A identifying of lesions in digital fundus images is required for development of an automated diabetic retinopathy screening system. Microaneurysms are the first symptom of diabetic retinopathy in clinical treatment. Microaneurysm numbers are used to denote the situation of the condition. Early detection of microaneurysm may use to lower the risk of blindness. This study discusses a set of efficiently adapted morphological regulators which are used for microaneurysm detection on non-dilated pupils and significantly higher-contrast retinal images. The microaneurysms observed are checked when compared with the ophthalmologists ' hand-drawn ground-truth. As a outcome, 81.61, 99.99, 63.76 and 99.98 percent respectively were the sensitivity, specificity, precision and accuracy. Diabetic Retinopathy and blindness problems diabetic patients facing. As the number of patients with diabetes is steadily increasing, this also results in an increase in the data. Therefore, the use of data mining methods is important to obtain the useful information and unseen knowledge. DM plays a major role in DR as it can be utilise to society's better health. For retinal fundus images there are many techniques and algorithms which help to diagnose. This paper discusses, classifies and compares the previously proposed algorithms and methods with a view to creating better and more effective algorithms. This paper presents a summary view of different data mining techniques which shows that KNN and SVM have given the best accuracies. This review paper can act as a resource for future researchers to use data mining techniques to predict diabetic retinopathy.(Rathi, 2017) Ananthapadmanaban et al.(2014) The most frequent of eye disease is diabetic retinopathy, is affected by complications which occur when capillary in the retina weaken. When early detection is not achieved it results in vision loss. Depending on the modeling objective many data mining techniques serve different purposes. The results of the various techniques for classifying data mining were tested using a different method. To predict early detection of diabetic retinopathy eye disease, we used Naive bayes and Support Vector Machine, and the results indicate that the Naive bayes test was 89.47 percent accurate. Brief insight into the identification of DR in human eyes using various forms of preprocessing & segmentation techniques is provided in this research article (Kumaran et al.). The detection actually depends on the RNFL network region. If the total area of the nerve fiber is lower, it will be affected by diabetic retinopathy (DR) and if the region of the nerve network is larger, therefore diabetic retinopathy will not impact the eyes and is therefore normal. It is a well- known fact that diabetics play a critical role in the wellbeing of humans and affect all organs. One such organ that is in man's possession. The DR will lead to a loss of vision in the human eye as the optic nerve is connected to the brain. The retinal fundus images are widely used in disease-affected images to diagnose & interpret disease. Raw images of the retinal fundus are difficult to process with machine learning algos. It is in this very context that a survey is being given here. Kauppi et al.(2006) The advancement of image processing techniques to a high level where the finding can be transferred from research laboratories to practice includes the following: protocols approved and applied to test the techniques, protocols similar to the strict medical care regulations, and medicine research. They suggested the first step towards a systematic review of methods for detecting diabetic retinopathy findings. DIARETDB0 is at al complicated database in many respects but in reality it corresponds to the situation: the images are uncalibrated, the expert assessment is free form and the displays used to interpret the images are uncalibrated. Diabetic retinopathy is among Europe's most common causes of blindness. Effective therapies do exist however. In more than 50 per cent of all cases, correct and early diagnosis and proper treatment procedure will prevent blindness. As a screening tool for diabetic retinopathy, digital imaging is becoming available. In addition to providing a high-quality permanent retinal appearance record that can be used to track development or reaction to treatment and that can be checked by an ophthalmologist, digital images have the ability to be processed by predictive analytics systems. Identified the primary creation of a method for providing automated of digital photograph taken as part of our clinic's daily monitoring of diabetic retinopathy. A deep-learning enhanced DR detection algorithm reaches significantly better performance than a earlierly reported, but virtually similar, technique not using deep learning (Abramoff et al.,2016). Deep learning algorithms have the suitability to develop DR screening performance and prevent this devastating disease from vision impairment and blindness. Table VI: Related Work S/ N O. Author Year Feature Selection Techniques Used Machine Learning techniques used Performan ce Evaluation 1 S. Wilfred Franklin , S. Edward Rajan ( 2014) Segmentation technique Back Propagation algorithm Accuracy: 95.03% 2 Wong Li Yun(2007) Image processing techniques Three-layer feedforwar d neural network 90% sensitivity, 100% specificity 3 Alan D Fleming(2007 ) Multi-scale morphological process. Retinopath y screening programme s 95% sensitivity 84.6% specificity 4 G Gardner Digital filtering techniques Back propagation neural network. 88.4% sensitivity, 83.5% specificity 5 Asha Gowda Karegowda(2 011) Decision tree and GA-CFS Back propagation neural network 96.97% Sensitivity, 100%Speci ficity, 98.45% accuracy. 6 B. Dupas(2010) Embedded method (Regression) Grading of DR and risk of ME 72.8% sensitivity, 70.8% specificity. Knowledge based Expert System for Predicting Diabetic Retinopathy using Machine Learning Algorithms 24 Retrieval Number: C6397029302/2020©BEIESP DOI: 10.35940/ijeat.C6397.049420 Published By: Blue Eyes Intelligence Engineering & Sciences Publication 7 T. Teng(2005) PSO Image processing algorithms 100% sensitivity. 8 Gwenol´e Quellec(2017 ) Correlation Matrix CADe algorithms Group form 9 Akara Sopharak(200 9) Correlation Matrix with Heat maps Fuzzy CMeans (FCM) clustering PPV 42.77%,PL R 224.26%, accuracy 99.11% 10 C. I. Sánchez(2005 ) Univariate Selection statistical classificatio n 79.62% sensitivity 11 P.V.Rao(2014 ) Z Score Normalization Artificial Neural Network (ANN) 90.6% accuracy 12 Dr. G. G. Rajput(2014) Embedded methods k-means clustering technique. 95.92% sensitivity, 92.28% predictive value, 99.70%. accuracy. 13 P. Kaha(2005) Wrapper technique Decision support system (DSS) 67% specificity 14 Deepthi K Prasad(2015) Wavelet transformations Back propagation neural network 93.8% accuracy 15 Akara Sopharak(201 8) Wrapper method Exudate detection and classificatio n 80%sensiti vity 99.5%speci ficity 16 Manoj Raju(2017) Detecting the laterality of fundus image Deep learning application in classifying 80.28% sensitivity ,92.29% specificity , 93.28% accuracy 17 Arkadiusz Kwasigroch(2 018) Embedded(Regr ession) Deep convolution al neural networks (CNN) 82% accuracy 18 Xiaogang L(2017) Filter methods Convolutio nal Neural Networks (CNNs) Group form 19 P. Khojasteh Embedded(LAS SO Regression) Convolutio nal neural network architecture 87.6% accuracy 20 R.Priya(2012) SVM Probabilisti c Neural network (PNN) 89.60% accuracy 21 Kanika Verma Random Forests technique Density analysis and bounding box techniques. 90% accuracy 22 Swati Gupta(2015) Recursive Elimination Computatio nal techniques 87% sensitivity, 100%specif icity 86%accura cy 23 R. GeethaRaman i(2017) Wrapper and filter methods UCI Machine Learning Repository 96.14% accuracy 24 athira2019 Filter methods R-CNN (Regional CNN) 93.8% 25 priya2013 Binary patterns SVM 95.38% 26 chetoui2018 Local Ternary Pattern SVM with a Radial Basis Function kernel (SVMRBF) , 93.10% 27 wan2018 Filter and wrapper techniques Convolutio nary neural networks 95.68% 28 sopharak2012 LTP Bayes classifier 85.68- sensitivity, 99.99- specificity, 83.34 - precision 99.99- accuracy 29 rathi2017 Data mining techniques Artificial neural network 94.8% 30 sopharak2008 Wrapper methods SVM Classifier 99.98% 31 rathi2017 Filter methods SVM Classifier 90% 32 kandhasamy2 019 Local Binary Patterns, Colour Moments SVM classifier 98.01 33 ananthapadm anaban2014 Images by descriptors and Hu moment of GIST Naive baye s and Supp ort Vector Machine Rapid Miner Tool 83.37 34 kumaran RNFL network region Artificial neural network 85% 35 akram2014 NPDR lesions Gaussian Mixture Model 98.52% 36 kauppi2006 PSO Image database, ground truth and evaluation methodolog y Sensitivity 79% 37 ege2000 Mahalanobis classifier Bayes classifier, KNN classifier 93% 38 abramoff2016 lesion detectors CNN Sensitivity- 96.8% specificity 87.0% III. METHODOLOGY Particle swarm optimization has been applied to numerous areas in optimization and in combination with other existing algorithms. This method performs the search of the optimal solution through agents, referred to as particles, whose trajectories are adjusted by a stochastic and a deterministic component. The selected features are then classified using the following International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-4, April 2020 25 Retrieval Number: C6397029302/2020©BEIESP DOI: 10.35940/ijeat.C6397.049420 Published By: Blue Eyes Intelligence Engineering & Sciences Publication algorithms: Decision Tree, Random Forest, Support Vector Machine. Figure 3 depicts the System Architecture Diagram Fig. 3 System Architecture IV. RESULT ANALYSIS The performance of the proposed work is analyzed using the following metrics: accuracy, sensitivity and specificity. Table VII- Comparison results of classifiers with regard to Accuracy Sensitivity in % Classifiers No. of features (20) No. of features (25) No. of features (30) SVM 84.6 87 96.6 Random Forest 90.1 94.3 94 Decision Tree 90 94 95.4 Fig. 4 Comparison results of classifiers with regard to Accuracy Table VIII- Comparison results of classifiers with regard to Specificity Specificity in % Classifiers No. of features (20) No. of features (25) No. of features (30) SVM 82.5 95.5 96.5 Random Forest 85 91 95 Decision Tree 83.5 84 94.5 Fig. 5 Comparison results of classifiers with regard to Specificity Table 9 represents the comprasion between the classifiers terms of accuracy which is figured in fig.6 Table IX-Comparison results of classifiers with regard to Sensitivity ACCURACY in % CLASSIFIERS No. of features 20 No. of features 25 No. of features 30 SVM 96 95 98 Random Forest 96 97 98 Decision Tree 95 97 98 Fig. 6 Comparison results of classifiers with regard to Sensitivity Knowledge based Expert System for Predicting Diabetic Retinopathy using Machine Learning Algorithms 26 Retrieval Number: C6397029302/2020©BEIESP DOI: 10.35940/ijeat.C6397.049420 Published By: Blue Eyes Intelligence Engineering & Sciences Publication V. CONCLUSION Thus the paper examines the Diabetic retinopathy using PSO Feature Selection algorithm on three different Classifiers SVM Classifier accuracy (98), sensitivity (96.6) and specificity (96.5). SVM Classifier have got the maximum metrics percentage for PSO Feature selection algorithm. REFERENCES 1. Franklin, S. Wilfred, and S. Edward Rajan. "Computerized screening of diabetic retinopathy employing blood vessel segmentation in retinal images." biocybernetics and biomedical engineering 34.2 (2014): 117-124. 2. Yun, Wong Li, et al. "Identification of different stages of diabetic retinopathy using retinal optical images." Information sciences 178.1 (2008): 106-121. 3. Fleming, Alan D., et al. "Automated detection of exudates for diabetic retinopathy screening." Physics in Medicine & Biology 52.24 (2007): 7385. 4. Gardner, G. G., et al. "Automatic detection of diabetic retinopathy using an artificial neural network: a screening tool." British journal of Ophthalmology 80.11 (1996): 940-944. 5. Gowda Karegowda, Asha, et al. "Exudates Detection in Retinal Images using Back Propagation Neural Network." International Journal of Computer Applications 25.3 (2011): 25-31. 6. Dupas, Bénédicte, et al. "Evaluation of automated fundus photograph analysis algorithms for detecting microaneurysms, haemorrhages and exudates, and of a computer-assisted diagnostic system for grading diabetic retinopathy." Diabetes & metabolism 36.3 (2010): 213-220. 7. Teng, Thomas, Martin Lefley, and D. Claremont. "Progress towards automated diabetic ocular screening: a review of image analysis and intelligent systems for diabetic retinopathy." Medical and Biological Engineering and Computing 40.1 (2002): 2-13. 8. Quellec, Gwenolé, et al. "Deep image mining for diabetic retinopathy screening." Medical image analysis 39 (2017): 178-193. 9. Sopharak, Akara, Bunyarit Uyyanonvara, and Sarah Barman. "Automatic exudate detection from non-dilated diabetic retinopathy retinal images using fuzzy c-means clustering." sensors 9.3 (2009): 2148-2161. 10. Sánchez, Clara I., et al. "Retinal image analysis to detect and quantify lesions associated with diabetic retinopathy." The 26th Annual International Conference of the IEEE Engineering in Medicine and Biology Society. Vol. 1. IEEE, 2004. 11. Rao, P. V., R. Gayathri, and R. Sunitha. "A novel approach for design and analysis of diabetic retinopathy glaucoma detection using cup to disk ration and ANN." Procedia Materials Science 10 (2015): 446- 454. 12. Rajput, G. G., and Preethi N. Patil. "Detection and classification of exudates using k-means clustering in color retinal images." 2014 Fifth International Conference on Signal and Image Processing. IEEE, 2014. 13. Kahai, Pallavi, Kameswara Rao Namuduri, and H. Thompson. "A decision support framework for automated screening of diabetic retinopathy." International journal of biomedical imaging 2006 (2006). 14. Prasad, Deepthi K., L. Vibha, and K. R. Venugopal. "Early detection of diabetic retinopathy from digital retinal fundus images." 2015 IEEE Recent Advances in Intelligent Computational Systems (RAICS). IEEE, 2015. 15. Sopharak, Akara, et al. "Automatic detection of diabetic retinopathy exudates from non-dilated retinal images using mathematical morphology methods." Computerized medical imaging and graphics 32.8 (2008): 720-727. 16. Raju, Manoj, et al. "Development of a Deep Learning Algorithm for Automatic Diagnosis of Diabetic Retinopathy." MedInfo. 2017. 17. Kwasigroch, Arkadiusz, Bartlomiej Jarzembinski, and Michal Grochowski. "Deep CNN based decision support system for detection and assessing the stage of diabetic retinopathy." 2018 International Interdisciplinary PhD Workshop (IIPhDW). IEEE, 2018. 18. Li, Xiaogang, et al. "Convolutional neural networks based transfer learning for diabetic retinopathy fundus image classification." 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI). IEEE, 2017. 19. Khojasteh, Parham, et al. "Introducing a novel layer in convolutional neural network for automatic identification of diabetic retinopathy." 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 2018. 20. Saha, Rituparna, Amrita Roy Chowdhury, and Sreeparna Banerjee. "Diabetic retinopathy related lesions detection and classification using machine learning technology." International Conference on Artificial Intelligence and Soft Computing. Springer, Cham, 2016. 21. R.Priya, P. Aruna. (2012). SVM and Neural Network based Diagnosis of Diabetic Retinopathy. International Journal of Computer Applications. Volume 41– No.1, March 2012 22. Verma, Kanika, Prakash Deep, and A. G. Ramakrishnan. "Detection and classification of diabetic retinopathy using retinal images." 2011 Annual IEEE India Conference. IEEE, 2011. 23. Gupta, Swati, and A. M. Karandikar. "Diagnosis of diabetic retinopathy using machine learning." Journal of Research and Development 3.2 (2015): 1-6. 24. Athira, T. R., et al. "Automatic detection of Diabetic Retinopathy using R-CNN." (2019). 25. Priya, R., and P. Aruna. "Diagnosis of diabetic retinopathy using machine learning techniques." ICTACT Journal on soft computing 3.4 (2013): 563-575. 26. Chetoui, Mohamed, Moulav A. Akhloufi, and Mustanha Kardouchi. "Diabetic Retinopathy Detection Using Machine Learning and Texture Features." 2018 IEEE Canadian Conference on Electrical & Computer Engineering (CCECE). IEEE, 2018. 27. Wan, Shaohua, Yan Liang, and Yin Zhang. "Deep convolutional neural networks for diabetic retinopathy detection by image classification." Computers & Electrical Engineering 72 (2018): 274- 282. 28. Sopharak, Akara, Bunyarit Uyyanonvara, and Sarah Barman. "Fine microaneurysm detection from non-dilated diabetic retinopathy retinal images using a hybrid approach." Proc. of the World Congress on Engineering. Vol. 2. 2012. 29. Rathi, P., and Anurag Sharma. "A review paper on prediction of diabetic retinopathy using data mining techniques." Int. J. Innov. Res. Technol 4.1 (2017): 292-297. 30. Sopharak, Akara, et al. "Automatic detection of diabetic retinopathy exudates from non-dilated retinal images using mathematical morphology methods." Computerized medical imaging and graphics 32.8 (2008): 720-727. 31. Rathi, P., and Anurag Sharma. "A review paper on prediction of diabetic retinopathy using data mining techniques." Int. J. Innov. Res. Technol 4.1 (2017): 292-297. 32. Kandhasamy, J. Pradeep, et al. "Diagnosis of diabetic retinopathy using multi level set segmentation algorithm with feature extraction using SVM with selective features." Multimedia Tools and Applications (2019): 1-16. [33] Ananthapadmanaban, K. R., and G. Parthiban. "Prediction of chances-diabetic retinopathy using data mining classification techniques." Indian Journal of Science and Technology 7.10 (2014): 1498-1503. 33. Kumaran, Yogesh, and Chandrashekar M. Patil. "A Brief Review of the Detection of Diabetic Retinopathy in Human Eyes Using Pre- Processing & Segmentation Techniques." 7(4). 34. Akram, M. Usman, et al. "Detection and classification of retinal lesions for grading of diabetic retinopathy." Computers in biology and medicine 45 (2014): 161-171. [36] Kauppi, Tomi, et al. "DIARETDB0: Evaluation database and methodology for diabetic retinopathy algorithms." Machine Vision and Pattern Recognition Research Group, Lappeenranta University of Technology, Finland 73 (2006): 1-17. 35. Ege, Bernhard M., et al. "Screening for diabetic retinopathy using computer based image analysis and statistical classification." Computer methods and programs in biomedicine 62.3 (2000): 165-175. 36. Abràmoff, Michael David, et al. "Improved automated detection of diabetic retinopathy on a publicly available dataset through integration of deep learning." Investigative ophthalmology & visual science 57.13 (2016): 5200-5206. International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-4, April 2020 27 Retrieval Number: C6397029302/2020©BEIESP DOI: 10.35940/ijeat.C6397.049420 Published By: Blue Eyes Intelligence Engineering & Sciences Publication AUTHORS PROFILE J. Jayashree received UG degree from Anna University, Tamilnadu and received PG degree from VIT University, Tamilnadu and PhD from VIT University. She is working as Assistant Professor Senior at VIT University, Vellore, Tamilnadu, India. Her research interests include Data Mining, Machine Learning. She had published a good number of papers in reputed Scopus Indexed Journals. Sruthi R, I finished my schooling in Sishya School, Hosur. I am pursuing my B.Tech Computer Science with Information Security in VIT University Vellore Ponnamanda Venkata Sairam, I finished my schooling in St.Ann’s high school. I am pursuing my B.Tech Computer Science with Information Security in VIT University Vellore J.Vijayashree received PG degree and PhD from VIT University,Tamilnadu. She is working as Assistant Professor Senior at VIT University, Vellore, Tamilnadu, India. Her research interests include Data Mining, Machine Learning. She had published a good number of papers in reputed Scopus Indexed Journals. 
work_ksczsj2scra5peij5nkle3u6ei	----	Ontology design and individual cognitive peculiarities: A pilot study Expert Systems with Applications 42 (2015) 3883–3892 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Ontology design and individual cognitive peculiarities: A pilot study http://dx.doi.org/10.1016/j.eswa.2015.01.008 0957-4174/� 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). ⇑ Corresponding author. Mobile: +7 (911)777 8466. E-mail addresses: gavrilova@gsom.pu.ru (T.A. Gavrilova), leshcheva@gsom.pu.ru (I.A. Leshcheva). Tatiana A. Gavrilova a, Irina A. Leshcheva b,⇑ a St. Petersburg University, Graduate School of Management, Volkhovskiy Pereulok, 3, St. Petersburg 199004, Russia b St. Petersburg University, Graduate School of Management, Dekabristov Pereulok, 16, St. Petersburg 199155, Russia a r t i c l e i n f o Article history: Available online 13 January 2015 Keywords: Ontologies Knowledge engineering Cognitive style Conceptual structuring Collaborative design Cognitive ergonomics a b s t r a c t The paper presents the main results of the KOMET (Knowledge and cOntent structuring via METhods of collaborative ontology design) project, which aims to develop a novel paradigm for knowledge structuring based on the interplay between cognitive psychology and ontology engineering. By the knowledge struc- ture (a conceptual model) we define the main domain concepts and relations between them in the form of a graph, map or diagram. This approach considers individual cognitive styles and uses recent advances in knowledge engineering and conceptual structuring; it aims to create new, consistent and structurally holistic knowledge bases for various areas of science and technology. Two stages of research have been completed: research into correlations between the expert’s individual cognitive style and the peculiarities of the expert’s subject domain ontology development; and research into correlations between the expert’s individual cognitive style and the group ontology design (including design accomplished by groups of experts with either similar or different cognitive styles). The results of these research stages can be applied to organizing collaborative ontology design (especially for research and learning purposes), data structur- ing and other group analytical work. Implications for practice are briefly delineated. � 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1. Introduction One of the main objectives of research and learning processes is achieving maximal effectiveness from the creation, transfer and dissemination of new knowledge. This effectiveness can be mea- sured by the quality and speed of memorization of the principal concepts of a particular domain and of the relationship between these concepts. Wide evidence exists that the visual thinking used to address the subject of study is positively connected with the quality and speed of memorization, and thus with the effectiveness of knowledge dissemination. Visualization is working as a cogni- tive tool that facilitates communication both in teacher/learner interaction and within research communities. Special interest in such graphical forms of knowledge codifica- tion can be observed in education science, especially within learn- ing where the students are engaged in group knowledge sharing and co-creation processes with continuous feedback. Mutual understanding and mentalization in research is of spe- cial interest in collective study or discovery. One of the most pro- ductive methods of research and learning collaboration promises to be group ontology design. An ontology is a set of definitions we make in understanding and viewing the world (Gruber, 1993). The specific problem being addressed in this work deals with the problem of improving the quality of group or collective ontologies. We are also interested in filling the gaps in understanding the group ontology design process specifics, such as the causes of differentia- tion between the form and the content of individual ontologies. During the last decade, visual knowledge representation has become one of the key considerations in knowledge engineering methodology, and it is strongly associated with ontology design and development. These ontologies, which form a conceptual skel- eton of the modeled domain, might serve various purposes such as better understanding, knowledge creation, knowledge sharing and reusing, collaborative learning, problem solving, seeking advice, or developing competences by learning from peers (Chu, Lee, & Tsai, 2011; Jung, 2012). Recently, the ontological engineering perspec- tive has gained interest in many research domains, such as medi- cine, business and computer science (Brochhausen et al., 2011; Oltramari & Ferrario, 2009; Pfister & Eppler, 2012; Schnotz & Kurschner, 2008). These studies rely heavily on theory and tools from knowledge engineering analysis that already has a long-standing tradition in the knowledge-based systems domain (Mizoguchi, 2003; Mizoguchi & Bourdeau, 2000). The largest number of knowledge engineering research articles has been generated around the theme of descriptive logics and formal foundations of ontology design (Baader, Horrocks, & Sattler, 2005; Kuznetsov, Obiedkov, & Roth, 2007). Our work, however, emphasizes the informal approach http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.01.008&domain=pdf http://creativecommons.org/licenses/by-nc-nd/4.0/ http://dx.doi.org/10.1016/j.eswa.2015.01.008 http://creativecommons.org/licenses/by-nc-nd/4.0/ mailto:gavrilova@gsom.pu.ru mailto:leshcheva@gsom.pu.ru http://dx.doi.org/10.1016/j.eswa.2015.01.008 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 3884 T.A. Gavrilova, I.A. Leshcheva / Expert Systems with Applications 42 (2015) 3883–3892 based on human-centred ontology design processes, an aspect neglected by most of the existing approaches. Several attempts have been made to bridge this gap and ease the overall ontology development process, such as HCOME – Human-Centered Ontology Engineering MEthodology, by Kotis and Vouros (2006); and human- centred ontology design, by Iqbal, Murad, Mustapha, and Sharef (2013). The tools and techniques developed in the domain of ontology engineering can be applied fruitfully in the field of knowledge structuring and design (Dicheva & Aroyo, 2004; Dicheva, 2008; Knight, Gašević, & Richards, 2006; Schreiber, 2000) and semantic web applications (Davies, van Harmelen, & Fensel, 2002). The idea of using ontologies and visual structuring in research description and introduction has been discussed in many works (Fonesca, Davis, & Camara, 2003; Sherlock, 2000; Tansley & Tolle, 2009; Yudelson, Gavrilova, & Brusilovsky 2005) and is now being imple- mented in several research projects and software tools (Bard & Rhee, 2004; Hevner, 2007). This paper presents the main results of the KOMET (Knowledge and cOntent structuring via METhods of collaborative ontology design) project which was devoted to developing methods that use group visual ontology design in educational purposes, with regard to the respondents’ individual cognitive styles. The group ontology design was tested in the medical domain by a smaller group (Gavrilova, Ravodin, Bolotnikova, & Kotko, 2012) and com- puter science (informatics) domain by a larger group of partici- pants (Gavrilova, Leshcheva, Bolotnikova, Blagov, & Yanson, 2013). In the larger group of 79 respondents, all the participants were graduate students of the School of Computer Science of Saint Petersburg Polytechnic University. Almost all had 1–2 years’ expe- rience of research in computer science, and were in their fifth year of study, on the Masters programme. The domain ‘‘computer science’’ was chosen as all the students are young professionals in this area. We use the term synonymously with ‘‘informatics’’. The paper is organized as follows. First, it describes the concept of ontology, with an emphasis on the visual approach to ontology design. Section 2 concentrates on the theoretical background, with sub-Section 2.1 describing ergonomic metrics and their purpose and sub-Section 2.2 providing an overview of cognitive styles and the tests used to assess them. Section 3 presents our human-cen- tred research paradigm and framework, and Section 4 the results obtained in the study of the relationship between cognitive styles and the peculiarities of individual development of ontologies. Section 5 introduces the main results of collective ontology devel- opment, taking into account the cognitive styles of participants. Finally, some conclusions are drawn and future work is outlined. 2. Theoretical background of ontology engineering: visual bias The idea of using visual structuring of information to improve the quality of understanding and mentalization among research colleagues is not new (Shneiderman, 1996). For more than twenty years, concept mapping (Conlon, 1997; Grosslight, Unger, Jay, & Smith, 1991; Jonassen, 1998; Sowa, 1984) has been used to compile maps and mental models that support the process of knowledge sharing. Many scholars, especially those who teach science courses, operate as knowledge analysts or knowledge engineers by making visible the skeleton of the studied discipline and showing the domain’s conceptual structure (Kinchin, De-Leij, & Hay, 2005). This structure is frequently represented by a so-called ‘‘ontology’’. From a philosophical viewpoint, ‘‘ontology’’ is the branch of phi- losophy which deals with the nature and organization of reality. Ontologies aim at capturing domain knowledge in a generic way and providing a commonly agreed understanding of a domain, which may be reused and shared across applications and groups (Chandrasekaran, Josephson, & Benjamins, 1999). Neches and colleagues (Neches et al., 1991) gave the classical definition as follows, ‘‘An ontology defines the basic terms and relations com- prising the vocabulary of a topic area as well as the rules for com- bining terms and relations to define extensions to the vocabulary’’. There are numerous other definitions of this milestone term (Gruber, 1993; Guarino & Giaretta, 1998; Gómez-Pérez, Fernández-López, & Corcho, 2004). Together, these definitions clar- ify the ontological approach to knowledge structuring while giving sufficient freedom for open-ended, creative thinking. Many researchers and practitioners have argued about the differences between ontology and a conceptual model. We propose that ontology corresponds to the analyst’s view of the conceptual model, but is not de facto the formal model itself. The visual approach to presenting ontologies is not only com- pact but also comprehensive. It makes ontology a powerful mind tool (Jonassen, 1998). By definition, ontology is a declarative representation of a cer- tain precise domain specification, including the glossary of the domain terms and the logical expressions describing the meanings and the relationships of these terms, thus allowing structured sharing of knowledge related to the domain (Gruber, 1993). The relationships between the concepts in ontologies can be of different types, e.g. ‘‘is’’, ‘‘has part’’, ‘‘has a property of’’, etc. The concepts and relationships are universal for a certain class of objects in a subject area. Conceptual model visualization methods such as ontologies are also widely and effectively used in education, and many learn- ing ontologies have been developed for a number of disciplines (Barros, Verdejo, Read, & Mizoguchi, 2002; Chi, 2009; Dall’Alba and Barnacle, 2007; Fonesca et al., 2003; Gaeta, Loia, Mangione, Miranda, & Orciuoli, 2014; Gaeta, Loia, Orciuoli, & Ritrovato, 2015). However, the ontology-based approach to conceptual knowl- edge representation in research and pedagogy is a relatively new development. Ontologies are now considered as the most universal and shareable forms of such representation and modeling. There are more than a hundred techniques and notations that help to define and visualize conceptual models. Ontologies are useful structuring tools, in that they provide an organizing axis along which every researcher (or student) can men- tally mark his/her vision in the information hyper-space of domain knowledge. Frequently, it is impossible to express all the informa- tion as a single ontology. Accordingly, subject knowledge storage consists of a set of related ontologies. Some problems may occur when moving from one ontological space to another, but construct- ing group meta-ontologies may help to resolve these problems. For both formative and summative assessment purposes, crea- tion of ontologies and explanation of the processes involved can clearly indicate the extent and nature of the knowledge and under- standing. Knowledge entities that represent the static knowledge of the domain are stored in hierarchical order in the knowledge repository and can be reused by others. At the same time, those knowledge entities can be reused in descriptions of the properties or a methodological approach as applied in the context of another related knowledge entity. Of course, the ontologies are inevitably subjective to a certain extent, as knowledge by definition includes a component of per- sonal subjective perception; however, using the ontologies devel- oped by others is a convenient and compact means of acquiring new knowledge. At the same time, collective ontology develop- ment experience allows the participants in the process to gain the fullest possible understanding of the subject area. Meta-ontology provides a more general description dealing with higher-level abstractions (mind maps (Buzan, 2005) and con- cept maps (Novak, 1998; Novak & Cañas, 2006)). Fig. 1 (Gavrilova & Kudryavtsev, 2011) illustrates different ontology classifications in Genealogies Associations Partonomies Attributive Structures Computer science Person Group Organisation Nation Mankind Semiformal Formal Ontologies Methodology Relationship Formalisation Owner Domain Automated Semiautomated Manual Purpose Knowledge Based Systems Research Education Semantic Web Linguistic study Knowledge Management Descriptive and Informal Taxonomies Functional Biology Medicine Business Fig. 1. Summarizing the ontology classifications in a mind-map. T.A. Gavrilova, I.A. Leshcheva / Expert Systems with Applications 42 (2015) 3883–3892 3885 the form of the mind map. This representation may be called the knowledge map. Such maps are graphical tools for organizing and representing knowledge. Knowledge maps are now widely used for visualizing ontologies at the design stage, while ontology editors (like Protégé) facilitate the development stage. Research on knowledge mapping in the last 12 years has pro- duced a number of consistent findings (O’Donnell, Dansereau, & Hall, 2002). People recall more central ideas when they learn from a concept map than when they learn from text, and those with low verbal ability or low prior knowledge often benefit the most. It seems that knowledge maps reduce cognitive load. The use of knowledge maps also appears to amplify the benefits associated with scripted cooperation (Stephens, 2012). Learning from maps and communication via maps are enhanced by active processing strategies such as summarization or annotation and by designing maps according to gestalt principles of organization (Gavrilova, 2011; Wertheimer, 1938). Bearing in mind that ontologies are to be used not only as a knowledge component of knowledge management systems but also as a mind tool for comprehensiveness and better understand- ing, we have tried to follow the principle of good shape (or beauty) that is not new in basic scientific abstraction, and modeling (e.g. physics, chemistry, etc.). It is difficult to give a formal definition of this concept but it features the imprecise sense of harmonious or aesthetically pleasing proportionality and balance. The most substantial impulse to it was given by the German psychologist Max Wertheimer (Wertheimer, 1938). In a previous paper (Gavrilova, Leshcheva, & Strakhovich, 2014) we described how to apply this beauty-centred approach to business ontology design. In the KOMET project we consistently develop this method by com- bining expert assessment with formal assessment by ergonomic metrics, as presented in the next section. 2.1. Cognitive ergonomic metrics Many aspects affect the quality of an ontology from the cognitive ergonomic point of view, including the basic principle of evaluating visual perceptibility and understandability. Metrics of this kind were first proposed by the research group of Aldo Gan- gemi (Gangemi, Catennaci, Ciaramita, & Lehman, 2006). The ontol- ogy evaluation based on these metrics is formal but it helps to assess the quality of the ontology. The expanded list of metrics and the software tool for its assessment COAT (Cognitive Ontology AssessmenT) were presented in detail in two works (Bolotnikova, Gavrilova, & Gorovoy, 2011; Gavrilova, Bolotnikova, & Gorovoy, 2012). The COAT software environment provides the calculations for more than 30 metrics. COAT is implemented as an application in Java. Ontologies in OWL format are supported by the Jena Semantic Web Framework, a Java library class for working with RDF and OWL ontologies. The main metrics are calculated automatically. In evaluating the quality of the designed ontologies, the follow- ing two aspects are most important: (1) correctness and depth of reflection of the subject domain, and (2) ergonomic aspect of the ontology representation from the point of view of quality and human speed of perception. In the KOMET project the quality of the developed ontologies was assessed by two methods: � An expert method, where the ontology analyst and domain experts (both professors in computer science) evaluated the quality against such criteria as simplicity, completeness, imbalance and relevance. � A formalized method, where ontology (converted into OWL- representation) was assessed by a set of quantitative metrics using COAT software. The formalized method is preferable as it is free from experts’ and analysts’ subjective interpretations and had the potential to be automated. In our research the developed ontologies were assessed by an augmented set of metrics (such as minimal depth, absolute width, etc.) suggested in Bolotnikova et al. (2011). The notation used to describe the metrics is as follows: ‘‘g’’, a graph representing an ontology; the concepts (classes and exemplars) of the ontology are the graph nodes, and the rela- tionships between the concepts are the graph edges; ‘‘G’’, a set of all the nodes g; ‘‘E’’, a set of all the edges g. a. A minimal depth m ¼ Nj2P; 8iðNj2P 6 Ni2PÞ where Nj2P and Ni2P are the path lengths j and i from the set of paths P of the graph g. Fig. 2. Example of a narrow and deep ontology structure. Fig. 3. Example of a wide ontology structure. 3886 T.A. Gavrilova, I.A. Leshcheva / Expert Systems with Applications 42 (2015) 3883–3892 b. An absolute width m ¼ XL j Nj2L where Nj2L is the number of nodes of degree j from the set of nodes L of the graph g. c. An average width m ¼ 1 nL # g XL j Nj2L where Nj2L is the number of nodes of degree j from the set of degrees L of the graph g, nL # g , the number of all degrees of the graph (a maximal graph depth augmented by 1, if considering only a chosen dominant relationship). d. 90% line depth m ¼ P90ðNj2PÞ where P90ðNj2PÞ is a 90% percentile of the graph depth (possible value of the graph path length, not exceeding the length of 90% of the graph paths). e. Root-mean-square deviation of neighboring levels/degrees width ratio m ¼ XnL # g i¼2 Nli2L Nli�12L � 1nL # g�1 XnL # g i¼2 Nli2L Nli�12L !2 nL # g � 1 f. Complexity metric A number of nodes with multiple inheritance to the set of all the graph nodes: m ¼ Nv2MI nG where MI ¼ v 2 Gj9a1;a2ðisaðv;a1Þ^ isaðv;a2Þg is a set of all the graph nodes with more than one ‘‘is-a’’ relationship arc, Nv2MI is the number of all the elements of this set, nG is the number of graph nodes. g. Average number of the parent nodes of a graph node m ¼ 1 nG XG v NSv2G where Sv2G ¼fa 2 Gjisaðv; aÞg is a set of all the parents of the node v, NSv2G is the number of all the parent nodes of the node v, nG is the number of the graph nodes. These metrics can help in understanding what should be cor- rected in the description of the subject domain in order to improve it from the point of view of cognitive ergonomics or better percep- tion. Thus, it is supposed that each next version of the ontology will be better and can be perceived faster by users. The metrics can also be used in evaluating ontologies of the same subject domain produced by different people/teams. The cal- culated metrics help to estimate which of them is better from the point of view of cognitive ergonomics, and to choose the best of them if the evaluations of other important criteria differ insignifi- cantly. Figs. 2 and 3 show different perspectives of ontology structure. 2.2. About cognitive styles The cognitive-styles view has acquired great influence within the education field, and is frequently encountered at levels ranging from kindergarten to graduate school. There is a thriving industry devoted to publishing cognitive-style tests and guidebooks for teachers, and many organizations offer professional development workshops for teachers and educators built around the concept of learning styles (Peterson, Rayner, & Armstrong, 2009). However, we will use the concept of cognitive style only for the aim already stated. As the aim of the KOMET project was to develop a paradigm of structuring data and knowledge with regard to individual cognitive styles, we had to choose the appropriate parameters or features of cognitive style. The cognitive styles explain and describe how an individual acquires knowledge and how an individual processes information. The cognitive styles are related to problem solving, and generally to how information is acquired, structured and used. Among the main features of cognitive style (Hayes & Allinson, 1998) we can name: � field dependence versus field independence, � impulsivity versus reflection, � narrowness versus width of the categories, � rigidity versus flexibility, � levelling versus sharpening, � scope of cognitive equivalence, � visual/audio/kinesthetic preferences, etc. Three characteristics have been chosen from the plethora of cog- nitive style characteristics described in the literature (Groome, 2014; Mullany, 2001): field dependence/field independence (FD/ FID), impulsivity/reflection, and narrowness/width of the category. According to the definition by Witkin (Witkin, Moore, Goodenough, & Cox, 1977), FD/FID is ‘‘a structuring ability of per- ception’’. The field-independent style is defined by a tendency to separate details from the surrounding context. It can be compared to the field-dependent style, which is defined as a relative inability to distinguish detail from other information around it. The FD/FID characteristic can be interpreted as a proxy of the structuring T.A. Gavrilova, I.A. Leshcheva / Expert Systems with Applications 42 (2015) 3883–3892 3887 capability of an individual mind. The characteristic of this style does influence the structuring process as a whole (e.g. ontology development ‘‘from scratch’’), and even more it affects the restruc- turing process (the merging of individual ontologies). FD/FID exerts considerable influence on the collective problem-solving process. In dyads where members have cognitive styles differentiated by the FD/FID characteristic, the final solution is usually closer to the variant suggested by the FID participant. The FID dyads experi- ence difficulty in developing common decisions on arguable points, while the FD dyads are more successful in coming to agreement in collective problem solving. Psychologists in our research group were used to working with on-line test, based on the popular modification of Witkin’s method of ‘‘embedded figures’’ which aims to find a simple figure hidden within a complicated one (Witkin, Oltman, Raskin, & Karp, 1971). The impulsivity/reflection characteristic considers the amount of information collected prior to making a decision: impulsive indi- viduals are able to make decisions on a considerably bounded information basis, while reflective individuals are more inclined to make decisions considering full information concerning the respective situation. For assessing the respondents’ impulsivity/ reflection features, the ‘‘Matching Familiar Figures Test’’ (Kagan, 1966) was used. As for the narrowness/width of the category, the main differ- ence between the extreme poles of this characteristic is that nar- rowly categorizing individuals are inclined to restrict the area of application of a certain category, while the broad categorizers are, conversely, inclined to include a plethora of more-or-less related examples in a single category. The psychologists consulted on the experimental part of our work advised us to use a modifica- tion of the Pettigrew test (1958) by Fillenbaum (1959): the so-called ‘‘range width test’’. The procedure is based on respondents’ opinion on the minimum, average and maximum evaluations of a concept or category. 3. Research design: from individual to collective 3.1. Research paradigm: human-centred approach The KOMET project was targeted at developing a paradigm of data and knowledge structuring with regard to individual cognitive styles, using recent advances in knowledge engineering and conceptual structuring, aimed at creating structurally holistic knowledge bases for various areas of science and technology. We supposed that individual cognitive peculiarities may dramatically affect the shape and the content of domain ontology, as each ontol- ogy designer (knowledge engineer) has specific personal cognitive features. These features may affect such syntactic and semantic parameters of the domain ontology as: - Scope of categories, - Level of granularity, - Shape (depth, width and complexity), etc. We studied the medical case of a dermatology ontology design process, which took months to complete because of misunder- standings and contradictions within the development team, as a result of cognitive dissonance issues. The difficulties, problems and obstacles within another project on an optics ontology were similar, even in such a well structured domain. The actors– both knowledge analysts and experts – often disagreed over concept vocabulary, relations and even the principles of the designed structure. Most of the methods and methodologies for building ontologies are focused on the technological development activities, especially on ontology formalization and implementation, and they do not pay too much attention to other important aspects related to man- agement, learning, merging, integration, evolution and evaluation of ontologies (Corcho, Fernandez-Lopez, & Gomez-Perez, 2006). There are several methodologies for current ontology develop- ment. On the basis of the knowledge acquisition method, the methodologies can be classified into automated, semi-automated and manual (from ‘‘scratch’’). Even though automatic ontology learning methods, such as text mining and knowledge extraction (Navigli & Velardi, 2004), support ontology engineers by speeding up their task, significant amount of manual work is still required in the completion, consolidation, and validation of the automatically generated ontology. The manual method is based on the interaction between the knowledge analyst and the expert (as it was in classical expert systems: Adeli, 2003; Studer, Benjamins, & Fensel, 1998) and is an example of the collective ontology design process. The literature review identified over a dozen such approaches; for example, Iqbal and colleagues (2013) compared 15 methodologies, concluding that collaborative construction methods are not standardized and little attention is paid to this aspect. Other methods include UPON (Unified Process for ONtology building), based on a well established and widely used software engineering process, the Unified Process (UP) (De Nicola, Missikoff, & Navigli, 2009); and the classical and popular METHONTOLOGY, which has been tested on several knowledge domains. Of course, such methods of design and development are not without drawbacks, although collective construction mitigates the extreme differences between individual biases. However, none of these methods takes into consideration the individual peculiar- ities of the experts and analysts. These peculiarities comprise education, background, experience, temperament and personality, including the specific cognitive style parameters. The novelty of the KOMET project is that we propose to expand the existing palette of ontology, merging approaches from the human-centred viewpoint by preliminary psychological testing to identify the characteristics of individual team members. This makes it possible to propose a better way of collective working in ontology merging and alignment. The results allow different roles to be assigned to the team members. If several experts/ana- lysts are available they can be paired or grouped to optimize the knowledge structuring work. 3.2. Research framework The KOMET project’s objectives include: � research into correlations between the expert’s individual cog- nitive style and the peculiarities of the expert’s development of a specific domain ontology, � research into correlations between the expert’s individual cognitive style and group ontology design (including design performed in groups consisting of experts with either similar or different cognitive styles), and � identifying the implications for practical knowledge workers of the time-consuming and sophisticated ontology design process. The research was divided into two phases: A (individual) and B (collective). The phase-A research was composed of four consecu- tive steps: � A1: Identifying the significant characteristics of individual cognitive styles, based on the on-line test results, using the soft- ware ONTOmaster-TECOS (http://ontomaster.ru) developed in PHP and Java Script by Elena Kotova and Andrew Pisarev (Kotova, 2013). http://ontomaster.ru Table 1 Correlation matrix illustrating the correspondence between ontology metrics and the cognitive styles’ indices. Metrics Test results I/R NC/WC The time of the first answer The number of mistakes The size of the category Number of classes 0.44 3888 T.A. Gavrilova, I.A. Leshcheva / Expert Systems with Applications 42 (2015) 3883–3892 � A2: Creating the ‘‘computer science’’ research domain ontolo- gies using the Protégé (Protégé, 2013) tool. � A3: Informal assessment and formal automatic evaluation of the ontology metrics using the COAT software environment (Gavrilova et al., 2012). � A4: Performing statistical analysis in order to identify signifi- cant relationships between the characteristics of the young researchers’ (experts’) individual cognitive styles and the ontology metrics. A2 step performed using the same test sample of students as A1. All the students were given the task of developing a light- weight ontology for the computer science domain. They did this either by using a visual mapping approach with a pen-and-paper technique or by using the mind mapping software. Later they developed the same ontology with the Protégé tool and we assessed their ontologies in OWL-format. Phase B consisted of the following steps: � B1: Merging individual ontologies in face-to-face group ontol- ogy design. � B2: Expert evaluation of the merged ontologies and comparison with the individual ones. The KOMET-DILIGENT collective ontology development meth- odology was designed for step B2 (Gavrilova et al., 2013). This methodology enriches the findings of the Neon Project (Neon Project, 2010). All the individual and collective ontologies (both in pairs and in groups of 3–5) were analyzed. In the KOMET project, specificity of the collective ontology development was researched. The experi- ments also aimed to establish how the collective categorization style is developed. 4. Analysis of the individual construction affected by cognitive style As mentioned in the introduction, the research sample con- sisted of 79 students were enrolled in the Intelligent Systems Development course. They were given the task of developing an ontology for the computer science research domain. Due to the professional specificity of the sample, a bias toward narrow, reflec- tive and field-independent test persons was found in the sample. However, a statistically significant Spearman’s negative correlation between the FID score and the time of the first answer in the Kagan was calculated, showing that the sample was dominated by the fast FID and slow FD respondents. On the basis of the literature review and the practical ontology development experience, the following hypotheses were suggested: Hypothesis 1. Individuals at the extreme FID end of the FD/FID cognitive style spectrum tend to have highly developed cognitive structuring capabilities; thus, the quality of ontologies developed by FID individuals will be higher. Number of leaves 0.46 Absolute depth 0.39 Minimum depth 0.54 90th percentile depth 0.34 The average width 0.48 The standard deviation of the relative width 0.48 Average number of parents of a graph node 0.47 Hypothesis 2. Impulsive individuals tend to develop superficial ontologies lacking sufficient categorization in the upper level, while reflective individuals tend to develop deeper ontologies. The absolute cardinality of families 0.44 Branching factor 0.50 The absolute cardinality of leaves 0.46 The weighting factor branching leaves �0.39 Hypothesis 3. Ontologies developed by individuals described as ‘‘imprecise’’ in the Kagan impulsivity/reflectivity test results tend to be more complex (entangled). Hypothesis 4. The ‘‘narrowness/width of the category’’ cognitive style characteristic exerts a significant influence on the ontology width: the ‘‘wide categorizers’’ tend to develop broader ontologies. Table 1 presents part of the two series of test results. It describes the correlation coefficients for several metrics and the main parameters of cognitive style: � I/R — impulsiveness/reflexivity. � NC/WC — narrowness/width of category. The correlation between the cognitive style and ontology met- rics values was assessed by Spearman’s coefficient (rank correla- tion). The significant correlation between the metrics and such features as field dependence/field independence was not found, so it is not presented in the table. Empty cells in the table mean that no significant correlation was found. Hypothesis 1 was not confirmed, as no significant correlation between the FD/FID metric and the quality of the ontologies was found; this result gave rise to optimistic feelings about the whole project, as it shows that it is possible to teach any individual to develop ontologies of a high quality. Hypothesis 2 was partially confirmed: the ‘‘90% line depth’’ metric demonstrated significant positive correlations with the time of the first answer in the Kagan test, thus showing that reflec- tive test persons tend to develop deeper ontologies; however, no significant negative correlation between the time of the first answer and the ontology width was found. Hypothesis 3 was confirmed, as the number of mistakes in the Kagan test demonstrated a significant positive correlation with the values of the ‘‘average number of parents of a graph node’’ metric that characterizes the ontology complexity. Furthermore, the number of mistakes in the Kagan test demon- strated significant positive correlations with the metrics of the ‘‘minimal depth of the ontology’’ and the ‘‘families branching coefficient’’ and significant negative correlation with the weighted leaves branching coefficient. Hypothesis 4 was fully supported: the broad categorizers devel- oped larger ontologies in terms of the number of concepts, achieved mainly by the greater number of ‘‘children’’ of each parent concept. Respectively, the results of the ‘‘range width test’’ demonstrated significant correlations with such metrics as the ‘‘average ontology width’’, ‘‘number of leaves’’, ‘‘absolute cardinality of families’’, etc. These results also demonstrated significant correlation with the T.A. Gavrilova, I.A. Leshcheva / Expert Systems with Applications 42 (2015) 3883–3892 3889 root-mean-square deviation of the average ontology width. This result shows that the number of concepts belonging to the neigh- boring levels and to different branches is significantly different, indicating imbalance in the ontologies developed by the wide categorizers. Despite the objectivity of the quantitative metrics-based method of ontology assessment, this method has the significant drawback of being too formalized and lacking semantic analysis elements. Having augmented the quantitative metrics-based analysis by a semantic analysis performed manually, we found that the ontolo- gies developed by the field-independent test persons tended to have simpler and clearer structures. However, this simplicity and clarity tended to be achieved by truncating the concepts that did not fit into the developed ontology, thus sacrificing the ontology’s completeness and integrity for formal logical consistency. As for the collective ontology development, including wide categorizers in a single group with a FID individual can be useful, with the wide categorizers generating a plethora of sub-classes and the FID participant restructuring them. This hypothesis was tested on the stage of research dedicated to collective ontology development. Thus, the following relationships between the respondents’ individual cognitive styles and the peculiarities of their subject domain ontology development have been identified as follows: � Reflective individuals tend to develop deeper ontologies; � The ontologies developed by the imprecise individuals (as defined in the Kagan test) tend to be more complex; � The ‘‘narrowness/width of the category’’ cognitive style affects the ontology branching coefficient, i.e. the ontology width. 5. Collective ontology design Specificity of the collective ontology development was also studies, both in dyads and in groups of 3–5. One objective of this study was to establish how the collective categorization style was developed. The KOMET-DILIGENT collective ontology face-to-face design methodology proposed within the KOMET project uses the follow- ing algorithm (Gavrilova et al., 2013): 1. Preliminary individual ontology development by the partici- pants and consequent mutual ontology matching; 2. Ontology analysis, merging and alignment; 3. Ontology revision and redesign. Students first developed individual ontologies and were then asked to develop a collective common ontology on the same topic, ‘‘computer science’’. The time allowed was one hour. The experiments aimed to establish how a collective categoriza- tion style was developed. Two strategies were identified: Fig. 4. Example of individual A o S1 , a strategy of collective ontology development ‘‘from scratch’’, and S2 , a strategy of common ontology development on the basis of two or more individual drafts. These strategies were affected by the peculiarities of analyzing and merging individual ontologies in the collective ontology devel- opment methodology suggested and explained above. The second strategy, S2, is of greater practical interest. In this case the respondents effectively applied all three basic ontology engineering operations (matching, merging and alignment). The results showed that merging usually follows either of two scenarios: 1. Mechanistic scenario (60–70% of all the tested groups); 2. Synthesizing scenario (30–40% of the tested groups). We researched the mechanistic scenario in more detail, as this scenario was used more often than the others. The implications led to the design of two models: � A disjunctive model, in which the higher-power ontology absorbs the lower-power one, with further merging of the same-degree nodes in the resulting ontology. � A conjunctive model in which the reduction of nodes leads to the resulting ontology including only the conjunction or intersection of the same-degree nodes. Comparison of these scenarios with the cognitive styles of the test participants revealed the following relationships: � Field-independent (FID) test persons tend to prefer the conjunc- tive scenario; � Field-dependent (FD) test persons tend to prefer the disjunctive scenario. Figs. 4–6 illustrate the experiment and demonstrate the varia- tion from the synthesizing scenarios. This variation is the best alternative, with collective effort showing the synergetic effect of ‘‘jumping’’ to a higher level of abstraction. The higher level always demonstrates deeper knowledge and better understanding of the domain specifics. Although these ontologies can be criticized as they were developed by young researchers, our experiment was targeted at the study of collaborative ontology design, not the production of a serious domain ontology. Mind-mapping represen- tation is used only to illustrate the changes which occurred during the merging process. 6. Recommendations and future research directions This paper addresses the conceptual limitations of traditional research and learning communication and proposes using a visual ntology of computer science. Fig. 5. Example of individual B ontology of computer science. Fig. 6. Collaborative AB ontology of computer science. 3890 T.A. Gavrilova, I.A. Leshcheva / Expert Systems with Applications 42 (2015) 3883–3892 metaphor for illustration and presentation of essential pieces of knowledge. The first most obvious possible application of the results can be found in presenting the research findings to the participating groups of researchers; the collaboratively developed ontologies may be used for conference presentations and journal publications. Secondly, this approach may facilitate supervising Masters and PhD research and thesis writing. Indeed, team ontology develop- ment can lead the participating students to a deeper and more holistic understanding of their respective subject domains, espe- cially when the instructor takes the students’ cognitive styles into consideration when organizing groups for team ontology development. The results may also help in organizing brain storming and design thinking sessions (Eppler & Bresciani, 2013; Eppler & Burkhard, 2008). Finally, the impact of individual cognitive styles on team ontol- ogy development can be taken into consideration in all the areas where intellectual team work on ontology development might be used, e.g. fundamental science, R&D or management consulting. Using visual inspection of the ontology it is possible to detect gaps and misunderstandings in state-of-the-art knowledge and cognitive models of the domain knowledge. However, there is as yet little consensus on the useful design and orchestration of such structures. Furthermore, in many cases it is not known what the structure of socially legitimate knowledge patterns looks like, or how a particular instance of a knowledge model deviates from that ‘‘ideal’’ state (e.g. the guru’s view) (Cross, Parker, Prusak, & Borgatti, 2001). However, researchers are individuals, and they may dis- agree among themselves. The main focus of future research is concentrated around the deeper understanding of the cognitive issues of collaborative work. This discourse on the concept of cognitive style encompasses a number of interconnected topics, ranging from the impact of psy- chological theories to current perceptions about relationships between learning, understanding and teaching. As this journal focuses on expert systems development research, it may be interesting to aim further investigation at collaborative interdisci- plinary ontology design for mapping the emerging research areas of artificial intelligence and cognitive computing. It may be fruitful to discover new possibilities for such design and to investigate how ontologies can be arranged to support the process of knowledge-base development. Future research relates such issues to specific features of interdisciplinary collaboration. 7. Conclusion The chief result of the KOMET project may be formulated as its ‘‘human-centred approach to ontology design and development’’. It proposes not a tool but a methodology that can be easily and cheaply implemented for group knowledge engineering, in which the cognitive peculiarities of experts and knowledge analysts affect the form and content of the designed ontology. The study described here is only a first step in interdisciplinary research dedicated to enquiring into the effect of the expert/user/ learner’s individual cognitive style on group structuring design. Our results are therefore of a preliminary nature. Two stages of research have been completed: first, research into correlations between the expert’s individual cognitive style and the peculiarities of his/her development of the subject domain ontology; and secondly, research into correlations between the expert’s individual cognitive style and group ontology design (including design performed in groups consisting of experts either of similar or of different cognitive styles). The results of these research stages can be applied to organizing collaborative ontology design (especially for research and learning purposes), data structuring and other group analytical work. The implications for practice are briefly delineated. The results of the interplay of different cognitive styles can be observed as cognitive dissonance and discussions at different stages of ontology development. This is why these results may be used in the team-building phase of the ontology design lifecycle. The ‘‘recipe’’ for a good expert/analyst team may be defined as: T.A. Gavrilova, I.A. Leshcheva / Expert Systems with Applications 42 (2015) 3883–3892 3891 � All the team members undergo the cognitive style tests (this may take less than 30 min); � A project participant with reflective and narrow cognitive style parameters initiates the ontology A structure; � Wide-categorizing team members then generalize and expand this structure into ontology B; � The most field-independent participant (the visionary) re-engi- neers ontology B into the final ontology C; the team may discuss and orchestrate the resulting ontology also by using the formal metrics (see Section 2.1). If needed the last two steps may be repeated. Our empirical study was organized as individual and group visual ontology development. Group ontology design, both in dyads and larger groups, was performed either ‘‘from scratch’’ or on the basis of the drafts prepared by the members of the group individually. Merging of the individual drafts into a single group ontology follows either an absorption (uptake) or compromise (synthesis) scenario. The mechanistic scenario of the merging of ontologies can be implied from either a conjunctive or a disjunctive model. From the different cognitive styles, the field-dependent partici- pants tended to prefer the disjunctive model of merging ontolo- gies, with the higher-power ontology absorbing the lower-power one, and further merging of the same-level nodes in the resulting ontology. The field-independent participants tended to prefer the con- junctive model, with node reduction by conjunction or intercep- tion only of the same-level nodes. Despite the preliminary character of the research results, all the findings can be used in organizing collective ontology develop- ment, data structuring and other group analytical work. By taking into consideration their students’ cognitive styles when organizing them into groups, the instructor enhances her or his ability in guiding the ontology development process to reach deeper levels of students’ understanding of the subject domain. Our work presents a new human-centred viewpoint on collaborative ontology development. Knowledge of the personal cognitive peculiarities helps to leverage the subjectivity of the individually designed ontologies. This approach may lead to better results, through avoiding the pitfalls of personal incompatibility and conflict within the team of ontology developers. Eventually this will shorten the development period and lead to an increase in ontology quality. Finally, our approach does not require the construction of an entirely new technique; it may be embedded in the context of existing methods by separating the roles within the ontology development team. Acknowledgments The KOMET project was funded by the Russian Foundation for Basic Research (Grant 11-07-00140). Thanks to Elena Kotova and Andrey Pisarev for permission to use their software tool ONTOmas- ter-TECOS for testing cognitive style parameters. We are grateful to Katerina Bolotnikova and Vladimir Gorovoy for providing the COAT tool facilities. References Adeli, H. (Ed.). (2003). Expert systems in construction and structural engineering. CRC Press. Baader, F., Horrocks, I., & Sattler, U. (2005). Description logics as ontology languages for the semantic web. In D. Hutter & S. Werner (Eds.), Mechanizing mathematical reasoning, LNAI 2605 (pp. 228–248). Heidelberg: Springer-Verlag. Bard, J. B., & Rhee, S. Y. (2004). Ontologies in biology: design, applications and future challenges. Nature Reviews Genetics, 5(3), 213–222. Barros, B., Verdejo, M. F., Read, T., & Mizoguchi, R. (2002). Applications of a collaborative learning ontology. In C. A. Coello Coello et al. (Eds.), MICAI 2002. LNAI (2313, pp. 301–310). Heidelberg, Berlin: Springer-Verlag. Bolotnikova, E. S., Gavrilova, T. A., & Gorovoy, V. A. (2011). To a method of evaluating ontologies. Journal of Computer and Systems Sciences International, 50(3), 448–461. Brochhausen, M. M., Spear, A. D., Cocos, C. C., Weiler, G. G., Martín, L. L., Anguita, A. A., et al. (2011). The ACGT master ontology and its applications — Towards an ontology-driven cancer research and management system. Journal of Biomedical Informatics, 44(1), 8–25. Buzan, T. (2005). Mind map handbook. London: Thorsons. Chandrasekaran, B., Josephson, J. R., & Benjamins, V. R. (1999). What are ontologies, and why do we need them? Intelligent Systems and their Applications, 14(1), 20–26. Chi, Y. (2009). Ontology-based curriculum content sequencing system with semantic rules. Expert Systems with Applications, 36(4), 7838–7847. Chu, K.-K., Lee, C.-I., & Tsai, R.-S. (2011). Ontology technology to assist learners’ navigation in the concept map learning system. Expert Systems with Applications, 38(9), 11293–11299. Conlon, T. (1997). Towards Diversity: Advancing Knowledge-based Modelling with Knowledge Acquisition. In proceedings of 8th international PEG conference (pp. 379–386). Bulgaria, Sozopol. Corcho, O., Fernandez-Lopez, M., & Gomez-Perez, A. (2006). Ontological engineering: principles, methods, tools and languages. In O. Corcho et al. (Eds.), Ontologies for Software Engineering and Software Technology. Springer. Cross, R., Parker, A., Prusak, L., & Borgatti, S. (2001). Knowing what we know: Supporting knowledge creation and sharing in social networks. Organizational Dynamics, 30(2), 100–120. Dall’Alba, G., & Barnacle, R. (2007). An ontological turn for higher education. Studies in Higher Education, 32(6), 679–691. Davies, J., van Harmelen, F., & Fensel, D. (Eds.). (2002). Towards the semantic web: Ontology-driven knowledge management. New York, NY: John Wiley & Sons. De Nicola, A., Missikoff, M., & Navigli, R. (2009). A software engineering approach to ontology building. Information Systems, 34, 258–275. Dicheva, D., & Aroyo, L. (2004). Concepts and ontologies in web-based educational systems. International Journal of Continuing Engineering Education and Life-long Learning, 14(3), 187–190. Dicheva, D. (2008). Ontologies and Semantic Web for E-Learning. Handbook on information technologies for education and training. International handbooks on information systems. Springer (pp. 47-65). Eppler, M. J., & Bresciani, S. (2013). Visualization in management: From communication to collaboration. A response to Zhang. Journal of Visual Languages & Computing, 24(2), 146–149. Eppler, M. J., & Burkhard, R. A. (2008). Knowledge visualization. In M. Jennex (Ed.), Knowledge management: concepts, methodologies, tools, and applications (pp. 781–793). Hershey, PA: Information Science Reference. Fillenbaum, S. (1959). Some stylistic aspects of categorizing behavior. Journal of Personality, 27, 187–195. Fonesca, F., Davis, C., & Camara, G. (2003). Bridging ontologies and conceptual schemas in geographic information integration. GeoInformatica, 7(4), 355–378. Gaeta, M., Loia, V., Mangione, G.R., Miranda, S., & Orciuoli, F. (2014). Unlocking serendipitous learning by means of social Semantic web. CSEDU 2014 – In proceedings of the 6th international conference on computer supported education, 1, (pp. 285–292). Gaeta, M., Loia, V., Orciuoli, F., & Ritrovato, P. (2015). S-WOLF: Semantic workplace learning framework. Systems, Man, and Cybernetics: Systems, IEEE Transactions on, 45(1), 56–72. http://dx.doi.org/10.1109/TSMC.2014.2334551. Gangemi, A., Catennaci, C., Ciaramita, M., & Lehman, J. (2006). Modelling ontology evaluation and validation. In Y. Sure & J. Domingue (Eds.), The semantic web: Research and applications, LNCS 4011 (pp. 140–154). Heidelberg, Berlin: Springer-Verlag. Gavrilova, T. (2011). Knowledge engineering: structural view. In New trends in software methodologies, tools, and techniques frontiers. Artificial intelligence and applications series, proceedings of the tenth SoMet_11 (Vol. 231, pp. 219–227). IOS Press. Gavrilova, T., Bolotnikova, E., & Gorovoy, V. (2012). New ergonomic metrics for educational ontology design and evaluation. In H. Fujita & R. Revetria (Eds.), New trends in software methodologies, tools, and techniques frontiers. Artificial intelligence and applications series, proceedings of the tenth SoMet_12 (Vol. 246, pp. 361–378). IOS Press. Gavrilova, T., & Kudryavtsev, D. (2011). Diagrammatic knowledge modeling for managers–ontology-based approach. In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (pp. 26–29). Gavrilova, T., Leshcheva, I., Bolotnikova, E., Blagov, E., & Yanson, A. (2013). Measuring psychological impact on group ontology design and development: empirical approach. In P. Klinov & D. Mouromtsev (Eds.), Proceedings of 4th international conference knowledge engineering and the semantic web. Communications in computer and information science (Vol. 394, pp. 29–43). Springer. Gavrilova, T., Leshcheva, I., & Strakhovich, E. (2014). Gestalt principles of creating learning business ontologies for knowledge codification. Knowledge Management Research and Practice Journal, 20. http://dx.doi.org/10.1057/kmrp.2013.60. advance online publication. Gavrilova, T. A., Ravodin, R. A., Bolotnikova, E. S., & Kotko, A. V. (2012). Collaborative development of dermatovenerology Ontology. In Proceedings of international. conference. Med-e-Tel 2012. Global telemedicine and eHealth updates: Knowledge resources (Vol. 5, pp. 39–43). G.D. of Luxembourg: ISfTeH. http://refhub.elsevier.com/S0957-4174(15)00023-8/h0005 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0005 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0010 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0010 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0010 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0015 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0015 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0020 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0020 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0020 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0025 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0025 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0025 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0030 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0030 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0030 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0030 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0035 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0040 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0040 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0040 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0045 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0045 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0050 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0050 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0050 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0060 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0060 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0060 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0065 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0065 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0065 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0070 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0070 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0075 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0075 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0080 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0080 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0085 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0085 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0085 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0090 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0090 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0090 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0095 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0095 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0095 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0100 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0100 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0100 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0105 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0105 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0110 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0110 http://dx.doi.org/10.1109/TSMC.2014.2334551 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0125 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0125 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0125 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0125 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0130 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0130 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0130 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0130 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0135 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0135 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0135 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0135 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0135 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0145 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0145 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0145 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0145 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0145 http://dx.doi.org/10.1057/kmrp.2013.60 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0155 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0155 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0155 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0155 3892 T.A. Gavrilova, I.A. Leshcheva / Expert Systems with Applications 42 (2015) 3883–3892 Gómez-Pérez, A., Fernández-López, M., & Corcho, O. (2004). Ontological engineering with examples from the areas of knowledge management, e-Commerce and the semantic web. London: Springer-Verlag. Groome, D. (2014). An introduction to cognitive psychology (third edition). Psychology Press. Grosslight, L., Unger, C., Jay, E., & Smith, C. L. (1991). Understanding models and their use in science. Conceptions of middle and high school students and experts. Journal of Research in Science Teaching, 28(9), 799–822. Gruber, T. (1993). A translation approach to portable ontology specifications. Knowledge Acquisition, 5(2), 199–220. Guarino, N., & Giaretta, P. (1998). Ontologies and knowledge bases: Towards a terminological clarification. In Towards very large knowledge bases. Knowledge building & knowledge sharing (pp. 25–32). IOS Press. Hayes, J., & Allinson, C. W. (1998). Cognitive style and the theory and practice of individual and collective learning in organizations. Human Relations, 51(7), 847–871. Hevner, A. R. (2007). The three cycle view of design science research. Scandinavian Journal of Information Systems, 19(2), 87. Iqbal, R., Murad, M. A. A., Mustapha, A., & Sharef, N. M. (2013). An analysis of ontology engineering methodologies: A literature review. Research Journal of Applied Sciences, Engineering and Technology, 6(16), 2993–3000. Jonassen, D. H. (1998). Designing constructivist learning environments. In C. M. Reigeluth (Ed.), Instructional design models and strategies (2nd ed., pp. 215–239). Mahwah, NJ: Lawrence Erlbaum. Jung, J. J. (2012). Semantic annotation of cognitive map for knowledge sharing between heterogeneous businesses. Expert Systems with Applications, 39(5), 5857–5860. Kagan, J. (1966). Reflection-impulsivity: the generality and dynamics of conceptual tempo. Journal of Abnormal Psychology, 71(1), 17–24. Kinchin, I. M., De-Leij, F. A. A. M., & Hay, D. B. (2005). The evolution of a collaborative concept mapping activity for undergraduate microbiology. Journal of Further and Higher Education, 29(1), 1–14. Knight, C., Gašević, D., & Richards, G. (2006). An ontology-based framework for bridging learning design and learning content. Educational Technology & Society, 9(1). 23-3. Kotis, K., & Vouros, G. A. (2006). Human-centered ontology engineering: The HCOME methodology. Knowledge and Information Systems, 10(1), 109–131. Kotova, E. E. (2013). Modeling of individual student learning scenarios based on personal cognitive style parameters. Human Factors: Problems of Psychology and Ergonomics, 4(67), 124–128 (in Russian). Kuznetsov, S. O., Obiedkov, S., & Roth, C. (2007). Reducing the representation complexity of lattice-based taxonomies. In U. Priss, S. Polovina, & R. Hill (Eds.), Proceedings of 15th international conference on conceptual structures (ICCS 2007), LNAI (pp. 241–254). 4604: Springer. Mizoguchi, R. (2003). Part 1: Introduction to ontological engineering. New Generation Computing, 21(4), 365–384. Mizoguchi, R., & Bourdeau, J. (2000). Using ontological engineering to overcome common AI-ED problems. International Journal of Artificial Intelligence in Education, 11(2), 107–121. Mullany, M. J. (2001). Using cognitive style measurements to forecast user resistance. In Proceedings of the 14th Annual Conference of the NACCQ (pp. 95–100). Navigli, R., & Velardi, P. (2004). Learning domain ontologies from document warehouses and dedicated websites. Computational Linguist, 30(2), 151–179. Neches, R., Fikes, R. E., Finin, T., Gruber, T., Patil, R., Senator, T., et al. (1991). Enabling technology for knowledge sharing. AI Magazine, 12(3), 36–56. Neon Project (2010). From http://www.neon-project.org/. Novak, J. D. (1998). Learning, creating, and using knowledge: concept maps as facilitative tools in schools and corporations. Mahwah, NJ: Lawrence Erlbaum & Assoc. Novak, J. D., & Cañas, A. J. (2006). The Theory Underlying Concept Maps and How to Construct Them. Technical Report IHMC CmapTools 2006-01. Florida Institute for Human and Machine Cognition. Retrieved online December 12, 2010, from http://cmap.ihmc.us/Publications/ResearchPapers/ TheoryUnderlyingConceptMaps.pdf. O’Donnell, A. M., Dansereau, D. F., & Hall, R. H. (2002). Knowledge maps as scaffolds for cognitive processing. Educational Psychology Review, 14(1), 71–86. Oltramari, A., & Ferrario, R. (2009). Formal ontologies meet industry. Amsterdam: IOS Press. Peterson, E. R., Rayner, S. G., & Armstrong, S. J. (2009). Researching the psychology of cognitive style and learning style: Is there really a future? Learning and Individual Differences, 19(4), 518–523. Pettigrew, T. F. (1958). The measurement and correlates of category width as a cognitive variable. Journal of Personality, 26, 532–544. Pfister, R. A., & Eppler, M. J. (2012). The benefits of sketching for knowledge management. Journal of Knowledge Management, 16(2), 372–382. Protégé (2013). From http://protege.stanford.edu. Schnotz, W., & Kurschner, C. (2008). External and internal representations in the acquisition and use of knowledge: Visualization effects on mental model construction. Instructional Science. An International Journal of The Learning Sciences, 36(3), 175–190. Schreiber, G. (2000). Knowledge engineering and management: The commonkads methodology. Cambridge, MA: MIT Press. Sherlock, G. (2000). Analysis of large-scale gene expression data. Current Opinion in Immunology, 12(2), 201–205. Shneiderman, B. (1996). The eyes have it: a task by data type taxonomy for information visualizations. IEEE symposium on visual languages. IEEE Computer Society Press (pp. 336-343). Sowa, J. F. (1984). Conceptual structures: Information processing in mind and machine. Reading, MA: Addison-Wesley. Stephens, S. (2012). From tree to map: Using cognitive learning theory to suggest alternative ways to visualize macroevolution evolution. Education and Outreach, 5(4), 603–618. Studer, R., Benjamins, V. R., & Fensel, D. (1998). Knowledge engineering: Principles and methods. Data & Knowledge Engineering, 25(1), 161–197. Tansley, S., & Tolle, K. M. (Eds.). (2009). The fourth paradigm: data-intensive scientific discovery. Redmond, WA: Microsoft Research. Wertheimer, M. (1938). Gestalt theory. Barton Press. . Witkin, H. A., Oltman, P. K., Raskin, E., & Karp, S. A. (1971). A manual for the embedded figures test. Palo Alto, CA: Consulting Psychologists Press. Witkin, H. A., Moore, C. A., Goodenough, D. R., & Cox, P. W. (1977). Field-dependent and field-independent cognitive styles and their educational implications. Review of Educational Research, 47, 1–64. Yudelson, M., Gavrilova, T., & Brusilovsky, P. (2005). Towards user modeling meta ontology. In L. Ardissono, P. Brna, & A. Mitrovic (Eds.), UM 2005, LNAI 3538 (pp. 448–452). Heidelberg, Berlin: Springer-Verlag. http://refhub.elsevier.com/S0957-4174(15)00023-8/h0160 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0160 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0160 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0165 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0165 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0170 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0170 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0170 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0175 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0175 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0180 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0180 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0180 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0185 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0185 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0185 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0190 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0190 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0195 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0195 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0195 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0200 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0200 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0200 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0205 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0205 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0205 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0210 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0210 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0215 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0215 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0215 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0220 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0220 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0220 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0225 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0225 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0230 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0230 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0230 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0235 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0235 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0235 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0235 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0240 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0240 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0245 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0245 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0245 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0255 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0255 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0260 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0260 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0270 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0270 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0270 http://cmap.ihmc.us/Publications/ResearchPapers/TheoryUnderlyingConceptMaps.pdf http://cmap.ihmc.us/Publications/ResearchPapers/TheoryUnderlyingConceptMaps.pdf http://refhub.elsevier.com/S0957-4174(15)00023-8/h0280 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0280 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0285 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0285 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0290 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0290 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0290 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0295 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0295 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0300 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0300 http://protege.stanford.edu http://refhub.elsevier.com/S0957-4174(15)00023-8/h0310 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0310 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0310 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0310 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0315 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0315 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0320 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0320 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0325 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0325 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0325 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0330 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0330 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0335 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0335 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0335 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0340 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0340 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0345 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0345 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0350 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0355 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0355 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0360 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0360 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0360 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0365 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0365 http://refhub.elsevier.com/S0957-4174(15)00023-8/h0365 Ontology design and individual cognitive peculiarities: A pilot study 1 Introduction 2 Theoretical background of ontology engineering: visual bias 2.1 Cognitive ergonomic metrics a A minimal depth b An absolute width c An average width d 90% line depth e Root-mean-square deviation of neighboring levels/degrees width ratio f Complexity metric g Average number of the parent nodes of a graph node 2.2 About cognitive styles 3 Research design: from individual to collective 3.1 Research paradigm: human-centred approach 3.2 Research framework 4 Analysis of the individual construction affected by cognitive style 5 Collective ontology design 6 Recommendations and future research directions 7 Conclusion Acknowledgments References 
work_kswnvow3nfcytg7u3nxxhnzzsy	----	Website Archived SISLT COE Website Archived This website has been taken off-line. We are sorry that we have to retire a resource that has been very useful, and apologize for any inconvenience this may cause. However, the content for the site and the program code that runs the site is no longer being maintained due to lack of funding. Without proper maintenance the site cannot be kept secure and running. To view the courses we have offered since 2001, please visit MyZou and click on Schedule of Classes (Guest Access). If you have questions or concerns please contact sislt@missouri.edu. © 2019 Curators of the University of Missouri. DMCA and other copyright information. All rights reserved. An equal opportunity/access/affirmative action/pro-disabled and veteran employer. 
work_kti2eifb7rbd7peetif77jx5f4	----	A neural-AdaBoost based facial expression recognition system Expert Systems with Applications 41 (2014) 3383–3390 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A neural-AdaBoost based facial expression recognition system 0957-4174 � 2013 The Authors. Published by Elsevier Ltd. http://dx.doi.org/10.1016/j.eswa.2013.11.041 ⇑ Corresponding authors. Tel.: +86 18914557945 (E. Owusu), tel.: +13852914090 (Y. Zhan). E-mail addresses: kayowusueb@yahoo.com (E. Owusu), yzzhan@ujs.edu.cn (Y. Zhan), mao_qr@ujs.edu.cn (Q.R. Mao). Open access under CC BY-NC-ND license. Ebenezer Owusu ⇑, Yongzhao Zhan ⇑, Qi Rong Mao School of Computer Science and Communication Engineering, Jiangsu University, 301 Xuefu Road, 212301 Zhenjiang, Jiangsu, China a r t i c l e i n f o a b s t r a c t Keywords: Facial expression recognition Bessel transform Gabor feature AdaBoost MFFNN This study improves the recognition accuracy and execution time of facial expression recognition system. Various techniques were utilized to achieve this. The face detection component is implemented by the adoption of Viola–Jones descriptor. The detected face is down-sampled by Bessel transform to reduce the feature extraction space to improve processing time then. Gabor feature extraction techniques were employed to extract thousands of facial features which represent various facial deformation patterns. An AdaBoost-based hypothesis is formulated to select a few hundreds of the numerous extracted features to speed up classification. The selected features were fed into a well designed 3-layer neural network clas- sifier that is trained by a back-propagation algorithm. The system is trained and tested with datasets from JAFFE and Yale facial expression databases. An average recognition rate of 96.83% and 92.22% are regis- tered in JAFFE and Yale databases, respectively. The execution time for a 100 � 100 pixel size is 14.5 ms. The general results of the proposed techniques are very encouraging when compared with others. � 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY-NC-ND license. 1. Introduction Facial expression is the explicit transformation of the human face due to the automatic responses to the emotional instability. In most situations it is spontaneous and uncontrollable. The auto- matic facial expression involves the application of an artificial intelligent system to recognize the expressions of the face under any circumstance. Today, the studies of facial expressions have gained keen interest in pattern recognition, computer vision and its related fields. Mainly, such facial expressions are the seven pro- totypical ones, namely; anger, fear, surprise, sad, disgust, happy, and neutral. Research into automatic facial expression recognition is very important in this modern society of technological age. For instance, the technology is applied in a wide variety of contexts, including robotics, digital signs, mobile applications, and medicine. It is re- ported that ‘‘some robots can operate by first recognizing expres- sions’’ of humans (Bruce, 1993). The AIBO robot for instance is a biologically-inspired robot that can show its emotions via an array of LEDs located in the frontal part of the head (Breazeal & Scassel- lati, 2002). In addition to this, the robot can also display ‘happiness’ feeling when it detects a face. In behavioral sciences and medicine for instance, expression recognition is effectively applied for inten- sive care monitoring (Morik, Brockhausen, & Joachims, 1999). Cur- rently, there are developing systems that are capable of making routine examinations of facial behavior during pain in clinical set- tings. In infants the Neonatal Facial Coding System (NFCS) has been employed for real-time assessment within 32 to 33 week post-con- ceptional age infants who are undergoing a heel lance. The technol- ogy is being used in more advanced settings to reduce accidents through the implementation of automated detection of driver drowsiness in public transports. This system relays information about the drivers’ emotional states to observers for effective sur- veillance leading to necessary awareness. The hallmark of every facial expression system is accuracy and to some extent the speed of execution. However most of the exist- ing systems produce poor performances in terms of accuracy; as for execution speed, most of the systems are even silent to give a hint. Some few examples; Franco and Treves (2001) proposed a neural based facial expression recognition system that used princi- pal component analysis (PCA) to reduce the feature vectors. The features were fed into a feed-forward neural network that was trained by a back-propagation network. In this system an average recognition of 84.5% was reported on the Yale facial expression database – an achievement which is not very encouraging. Kumb- har, Jadhav, and Patil (2012) described a neural network classifica- tion facial expression recognition system that employs Gabor feature extraction and feature reduction by PCA to distinguish 7- class facial expression recognition on the JAFFE database. In this system they specified 20 inputs, 40 to 60 hidden layers and seven output feed-forward neural networks. Again, the 60–70% recogni- tion accuracy they obtained by their procedure is not encouraging http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2013.11.041&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.11.041 mailto:kayowusueb@yahoo.com mailto:yzzhan@ujs.edu.cn mailto:mao_qr@ujs.edu.cn http://dx.doi.org/10.1016/j.eswa.2013.11.041 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa http://creativecommons.org/licenses/by-nc-nd/3.0/ http://creativecommons.org/licenses/by-nc-nd/3.0/ Fig. 1. Sample face detection images (left), cropped detected face (right). 3384 E. Owusu et al. / Expert Systems with Applications 41 (2014) 3383–3390 to befit the expectations of a real-time system. Recently, Londhe and Pawar (2012) extracted features of the face using Affine Mo- ment Invariants and performed the classification using feed-for- ward neural network. The expression recognition obtained was 93.8% on the JAFFE database. Tai and Chung (2007) extracted the facial features using a Sobel filter. In their experiment they re- served the maximum connected component to reduce the wrinkles and noises and conducted 7-class classification on JAFFE facial expression database through the application of Elman network with two hidden layers, each layer containing fifteen neurons. With this approach the average accuracy of automatic facial expression recognition is 84.7%. Zhang and Tjondronegoro (2009) extracted the expressive face by using Gabor filters, feature reduc- tion by PCA and expression classification by neural network. In this method an average facial recognition of 93.4% was recorded in the JAFFE facial expression database. Dailey and Cottrell (1999) also extracted facial features by Gabor techniques and reduced the fea- tures by PCA. The expression classifier was neural network and the average expression recognition was 94.5% ± 0.7 on the seven proto- typical facial expressions, however the facial expression database was not mentioned. Most of these studies advocate the use of Neural Network as the expression classifier and extracted the facial features by Gabor fil- ters and reduced the features via PCA. The displeasing thing is that all the results were not very encouraging. This study persists in exploring the potentials of neural net- works to execute this kind of assignment, trying to esteem some biological constraints, utilizing the capabilities of modular systems. Though many techniques have been used to extract the facial features, Gabor feature extraction remains a high-quality choice; there are other alternatives but they are not very promising. Just a few examples: Satiyan and Nagarajan (2010) utilized the Haar technique to extract facial features which were used as input to the neural network for classifying 8 facial expressions. The Haar wavelet extraction is very fast (Satiyan & Nagarajan, 2010; Van, 2008), however the wavelets are too huge to result to effective classification when used as input to classifiers in facial expression recognition (Cemre, 2008); in other words it is a potential cause to misclassifications and poor performance. Distance-based feature extraction methods are also one of the largely applied techniques used for feature extraction in both 2D and 3D static faces. The idea behind these procedures is that the muscle deformations which are the major causes of changes in facial expression from normal expression results in variations of the Euclidean distances between facial landmarks or points. These points, as well as their distances, have been widely employed for static facial expression analysis (Sha, Song, Bu, Chen, & Tao, 2011; Soyel & Demirel, 2007; Tang & Huang, 2008). Among the most successful ones is feature extrac- tion based on the Bhattacharyya distance (Choi & Lee, 2003). How- ever, despite some advantages of this method, the degree of computational complexity is unacceptably high. The matching of even a small model shape with a normal image can take half an hour on an eight-processor Sun SPARCServer 1000 (Rucklidge, 1997; Zhang & Lu, 2004). The Patch based feature extraction meth- od is another alternative widely exploited for facial expression bio- metrics. Maalej, Amor, Daoudi, Srivastava, and Berretti (2010) for instance represented extracted patches from facial surfaces by sets of closed curves and then applied a Riemannian framework to ob- tain the shape analysis of the extracted patches. However, the patch-based features also have numerous drawbacks. First, partic- ular representations cannot be applied to other solutions without major modifications: the majority of the techniques have only been utilized to a single class. Also, most methods do not exploit the large amounts of available training data (Aghajanian et al., 2009). Thus on the basis of these we still considered Gabor features as the best approach, not because it does not have drawbacks, but the drawbacks can be easily managed. The Gabor filter is a superior model of simple cell receptive fields in cat striate cortex (Jones & Palmer, 1987), and it grants exceptional basis for object recogni- tion and face recognition (Lades et al., 1993; Wiskott, Fellous, Kru- ger, & vonderMalsburg, 1997). Again, the Gabor methods are superior to all the above-mentioned methods because it extracts the maximum information from local image regions (Deng, Jin, Zhen, & Huang, 2005), and it is invariant against, translation and rotations (Al Daoud, 2009). In this work, the data were reduced in dimensions by Bessel transform (Ganga, Prakash, & Gangashetty, 2011) and then after extraction of the face by Gabor methods, the features were further reduced via an AdaBoost-based (Freund & Schapire, 1995) feature reduction technique. The selected features which represented the facial deformation patterns were then fed into a 3-layer feed-for- ward neural network that is trained by a back-propagation algo- rithm. It is interesting to note that Bessel down-sampling techniques have never been adopted for facial expression recogni- tions. Again, the combinations of Bessel down-sampling and the formulated AdaBoost-based algorithm is an innovation that re- duces the expression dataset to enhance accuracy and speed. Final- ly, the construction of the feed-forward neural network is influential to bring about successful results. The rest of the work is arranged as follows. Section 2 discusses face detection and image down-sampling. Section 3 discusses Ga- bor feature extraction. Section 4 discusses feature selection. Sec- tion 5 discusses the multilayer feed-forward neural network (MFFNN). Results and analysis are presented in Section 6. The final conclusions of the work are drawn in Section 7. 2. Face detection and down-sampling The face detection component was implemented by the adop- tion of Viola and Jones system (Viola & Jones, 2004). Fig. 1 shows sample face detection by the Viola–Jones classifier. The size of the image is rescaled to a window of size 20 � 20 pixels by the use of Bessel down-sampling. Methods like bilinear interpolations have been utilized by several authors for this task in particular but interpolations are prone to aliasing problems (Munoz, Blu, & Unser, 2001). Bessel down-sampling reduces the size of the image and preserves the details and perceptual quality of the original image (Ganga et al., 2011). The down-sampling im- age signal xd (t1, t2) is expressed as: xdðt1; t2Þ¼ Xp n1¼1 Xq n2¼1 cðn1; n2ÞJ0 an1 p � r t1 � � J0 an2 q � s t2 � � ð1Þ where p and q refer to the respective image size, p � r and q � s are the required reduced size of the image; r and s are positive integers that represent the reduction values, n is the number of low-fre- quency DCT coefficients, J0ðan1Þ and J0ðan2Þ are zero order Bessel E. Owusu et al. / Expert Systems with Applications 41 (2014) 3383–3390 3385 functions, c(n1, n2) are Bessel coefficients computed from the first order Bessel function, t1 and t2 are chosen such that 0 6 t1 6 p � r and 0 6 t2 6 q � s. Interested readers are referred to (Al Daoud, 2009). 3. Gabor feature extraction The 2-D Gabor filters are spatial sinusoids localized by Gaussian window, and they can be created to be selective for orientation, localization, and frequency as well. It is very flexible to demon- strate images by Gabor wavelets because the details about their spatial relations are preserved in the process. Gðx; y; h; u; rÞ¼ 1 2pr2 exp � x2 þ y2 2r2 � � expf2piðR1 þ R2g ð2Þ where i is a complex number representing the square root of �1. R1 = ux cos h and R2 = uy sin h, u is the spatial frequency of the band pass, h is the spatial orientation of the function G, (x, y) spec- ify the position of light impulse in the visual field, r is the standard deviation of 2-D Gaussian envelop. In this Gabor family, we chose eight orientations 0; p8 ; 2p 8 ; . . . ; 7p 8 � � and five scales 4; 4 ffiffiffi 2 p ; 8; n 8 ffiffiffi 2 p ; 16g: In order to give added robustness to illumination we turned the Gabor filter to zero DC (direct current) by the expression ~Gðx; y; h; u; rÞ¼ Gðx; y; h; u; rÞ� 1 q Xn i¼�n Xn j¼�n Gðx; y; h; u; rÞ " # ð3Þ where, q is the size of the filter, given by q = (2n + 1)2. Fig. 2 shows the Gabor filter image. The sample points of the filtered image is coded to two bits, real bit x1 and imaginary bit x2 such that, G1 ¼ x1 ¼ 1; if R ~Gðx; y; h; u; rÞ h i � I n o P 0 x1 ¼ 0; if R ~Gðx; y; h; u; rÞ h i � I n o < 0 8>< >: ð4Þ G2 ¼ x2 ¼ 1; if I ~Gðx; y; h; u; rÞ h i � I n o P 0 x2 ¼ 0; if I ~Gðx; y; h; u; rÞ h i � I n o < 0 8>< >: ð5Þ where I is subimage of the expressional face, R and I are the real and the imaginary parts of each Gabor kernel, � is the convo- lution operator. With this coding, only the phase information in the facial expressions image is stored in the feature vector of size 256 bytes. The final magnitude response which is used to represent the feature vectors is computed by G ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi G21 þ G 2 2 q ð6Þ Fig. 3 shows the magnitude response of a template image. Fig. 2. Gabor filtered image: real (left) and imaginary (middle) parts of Gabor filter in 3 4. Feature selection Due to the large size of the Gabor wavelets, it is not practically possible to use all the wavelets as input to our classifier, for fear of misclassification and possible system crash. The AdaBoost feature reduction algorithms have special speed advantage in increasing classification process (Shen & Bai, 2004). Thus we formulated an AdaBoost-based algorithm to select a few deserving portions of the wavelets. Assuming the extracted Gabor features are represented by a total of i e (1, 2,..., N) appearance features. Then the image I is represented as Ui ¼fðxn; ynÞg N n¼�1 configured by the parameters z, l, v. The positive sets /(+) and the negative sets /(�) are denoted by /ðþÞ ¼ fðxn; ynÞg N n¼1 � R J �ð�1Þ and /ð�Þ ¼ fðxn; ynÞg N n¼�1 � R J� ð�1Þ respectively, where xn is the nth data sample containing J fea- tures, and yn is its corresponding class label. To train the vectors ||G||, which is denoted by /(u,v,z) over a distribution D, we simply determined the weights of all the feature vectors Ui ¼ fðxn; ynÞg N n¼�1 ¼ / ðþÞ þ /ð�Þ: This gives us a threshold k which indicated the decision hyperplane. k is computed as: k ¼ P 8i2/ðþÞ DðiÞ:/ðl;v;zÞ jj P 8i2/ðþÞ DðiÞ:/ðl;m;zÞjj þ P 8i2/ð�Þ DðiÞ:/ðl;v;zÞ jj P 8i2/ð�Þ DðiÞ:/ðl;v;zÞjj ð7Þ A sample is positive or client if it is located at the positive half of k (which is the majority decision), otherwise it is a negative or an imposter. The status is reversed if the minority of the positive in- stances is rather located in the positive half space. Let c be denoted by clients and p be the imposters. For a given training dataset con- taining both positive and negative samples, where each sample is (Si, yi); y e { ± 1} represents the corresponding class label, the fea- ture selection algorithm is formulated as follows: � Initialize sample distribution D0 by weighting every training sample equally such that the initial weight w1,i = 1/2c, 1/2p for y = 1 and �1, respectively. � For the iteration t = 1, 2,..., T, where T is the final iteration, do: (i) Normalize the weights, wt;i wt;iPN i¼1 wt;i , where wt is a probabil- ity distribution and N is the total number of features. (ii) Train a weak classifier ht for feature j, which uses a single feature. The training error nt is estimated with respect to wt such that: D, and t nt ¼ X t wt;ijhtðxiÞ� yij 2 ð8Þ (iii) Select the hypothesis h1t with the most discriminating infor- mation, that is to say, the hypothesis with the least classifi- cation error n1t ; on the weighted samples. (iv) Compute the weight xt that weights h 1 t by its classification performance as: xt ¼ 1 2 ln 1 n1t � 1 " # ð9Þ he real part of the Gabor kernels (2D) in the spatial and frequency domain. Fig. 3. A Gabor magnitude response of face image: a sample image (left), the magnitude response image of the whole Gabor filter bank of 40 Gabor filters (right). 3386 E. Owusu et al. / Expert Systems with Applications 41 (2014) 3383–3390 (v) The weight distribution is then updated and normalized by: wtþ1;i � wt;i:e�xt yi h 1 t ðS 1 t Þ ð10Þ � The final feature selection hypothesis H(S) which is a function of the selected features is denoted by: HðSÞ¼ sgn XT t¼1 xt h 1 t ðS 1 t Þ " # ð11Þ The selected features represent samples of the facial deforma- tion patterns of the expressive face. The datasets which were images from the JAFFE and Yale databases were partitioned into training and testing by leave-one-out cross validation (Wu, Bru- baker, Mullin, & Rehg, 2008). Fig. 4. A 3-layer feed-forward neural network. 5. Multilayer feed-forward neural network (MFFNN) classifier The selected features are fed into the constructed neural net- work to train it to identify the seven universal facial expressions. The architecture is a 3-layer feed-forward neural network and trained by a back-propagation algorithm (Bouzalmat, Belghini, Zarghili, Kharroubi, & Majda, 2011; Londhe & Pawar, 2012). The back propagation algorithm basically replicates its input to its out- put via a narrow conduit of hidden units. The hidden units extract regularities from the inputs because they are completely connected to the inputs. Every network was trained to give the maximum va- lue of 1 for exact facial expression and 0 for all other expressions. The construction is shown in Fig. 4. The input layer has 7 nodes, each for each facial expression whiles the hidden layer had 49 neu- rons; each expression for 7 neurons. We chose 7 neurons to com- pensate for the target output of seven facial expressions. This was the case for the seven prototypical facial expressions, which was validated by the use of the JAFFE facial expression database. Since the experiment was also validated using the Yale database, where four expressions were used there was a slight modification in the construction of the network for this application. Here, the hidden layer neurons were settled at 16 each facial expression dedicated for 4 neurons and 4 neurons in the output layer. The input vectors of the network represented by X = [x1, x2,..., xl] T. The output layers are denoted by Y = [y1, y2,..., yk] T. The optimi- zation model is formulated as X : h ? Y. The output dataset of each layer of the network is denoted by yj1; :::; y j k�1; y j k ; j ¼ 1; 2; :::; k � 1; k; where k � 1 corresponds to the total hidden layers and k represents the total output layers. We denote the tar- get datasets and its additive white noise by (t1, t2,..., tK) and g = (e1, e2,..., eK), respectively. The variable K represents the total patterns of the network. The corresponding vectors of the hidden units are denoted by V = (v1, v2,..., vk�1). The sigmoid activation function h = (h1, h2,..., hk) of each layer is h1, h2,..., hk�1. The weights of the network are updated by w1, w2,..., wk. The training epochs are 1000 and the target of error is 0.001. The training algorithm is modeled as: min h1;h2;v 1;w1;w2 XK j¼1 ðtj � yj2Þ 2 ð12Þ Subject to the constraints yj1 ¼ h1ðw1 x jÞ; w1 2 Rv 1�M; yj1 2 R v 1 ; xj 2 RM yj2 ¼ h2ðw2; y j 1Þ; w2 2 R 1�v 1 ; yj2 2 R 1 9>= >; ð13Þ The process of training involves weight initialization, calcula- tion of the activation unit, adjustment, weight adaptation, and test- ing for convergence of the network. Assuming vji represents the weight between the jth hidden unit and ith input unit; and wkj rep- resents the weight between the kth output and the jth hidden unit. The activation unit is then calculated sequentially, starting from the input layer. The activation of hidden and output unit is calcu- lated as: Table 2 Confusion matrix of 4-class facial expression recognition on Yale. Neural (%) Sad (%) Happy (%) Surprise (%) Neutral 86.16 9.81 0 4.03 Sad 8.35 86.79 0 4.86 Happy 1.7 0 97.60 0.7 Surprise 0.95 0.72 0 98.33 Average Recognition = 92.22% E. Owusu et al. / Expert Systems with Applications 41 (2014) 3383–3390 3387 yðpÞj ¼ h ðpÞ yj XI i¼1 v ji zi � v jo ! ð14Þ oðpÞk ¼ h ðpÞ ok XJ j¼1 wkjyj � wko ! ð15Þ where yðpÞj is the activation of the jth hidden unit and o ðpÞ k is the acti- vation of the kth output unit for the pattern, p. h is a sigmoid func- tion. k is the total number of output units, I is the total number of input units and J is the total number of hidden units. vjo is the weight connected to the bias unit in the hidden layer, zo = �1 and yo = �1. We adjusted the weights, starting at the output units and recursively propagated error signals to the input layer. The detected output oðpÞk is compared with the corresponding target value t ðpÞ k which is a facial image, over the entire training set using the sig- moid function to express the approximation error in the network’s target functions. EðpÞ ¼ 1 2 XK k¼1 tðpÞk � o ðpÞ k 2 ð16Þ The minimization of the error E(p), requires the partial deriva- tive of E(p) with respect to each weight in the network to be com- puted. The change in weight is proportional to the corresponding derivative. Dv jiðt þ 1Þ¼�g @EðpÞ @v ji þ aDv jiðtÞ ð17Þ Dwkjðt þ 1Þ¼�g @EðpÞ @wkj þ aDwkjðtÞ ð18Þ where, g is the learning rate, normally between 0 and 1, we set it to 0.9. The function a is also set to 0.9. The last term is a function of the previous weight change. @E @v ji ¼ @yj @v ji XK k¼1 �ðtk � okÞokð1 � okÞyj wkj ð19Þ @yj @v ji ¼ yjð1 � yjÞzi: Therefore, Dv ji ¼ g XK k¼1 ðtk � okÞokð1 � okÞyjwkj yjð1 � yjÞzi ð20Þ The weights are updated by, wkjðt þ 1Þ¼ wkjðtÞþ Dwkjðt þ 1Þ ð21Þ v jiðt þ 1Þ¼ v jiðtÞþ Dv jiðt þ 1Þ ð22Þ where, t is equal to the current time step. Dvji and Dwkj are the weight adjustments. We repeated the process once from the equa- tion (14) in order to achieve the desired output. Table 1 Confusion matrix of 7-class facial expression recognition on JAFFE. Neutral (%) Sad (%) Fear (%) Neutral 92.23 4.31 1.11 Sad 3.85 93.9 1.28 Fear 0 0.95 96.08 Anger 2.8 0 1.1 Disgust 0 0 0 Happy 0.21 0 0.07 Surprise 0 0 0.1 Average Recognition = 96.83% 6. Results and analysis The facial expression recognition was validated with the JAFFE and Yale facial expression databases. The JAFFE database contains 213 images of 10 female Japanese persons. Each respondent in the database posed three or four examples of each of the seven facial expression prototypes (happy (ha), sad (sa), anger (an), disgust (di), fear (fe), surprise (su), and neutral (ne)). 2 images of each individual from each class of expres- sion are randomly selected for training, leading to a total of 140 images corresponding to 65.7%, the rest was preserved for testing. The trial was performed using tenfold cross-validation to obtain the average recognition rate. In order to create distinct datasets for cross-validation, none of the sets in the training folder appear in any of the remaining folders. The Yale facial expression database contains 165 grayscale images in GIF format of 15 individuals. There are 11 images per subject. Each subject exhibited one of the six facial expressions; ha, ne, sa, sleepy (sl), surprise (su) and wink (wi). In this database we manually extracted 130 images corresponding to ha, ne, sa, and su. The datasets in this database were also partitioned into training and testing by using the same method described for the JAFFE. In due course about 77% of the images were used for training and the remaining, for testing. We recorded an average recognition rate of 96.83% in JAFFE and 92.22% in Yale on Intel(R) Core(TM) 2 Duo CPU P8400 @ 2.26 GHz (2 CPUs) – 2.3 GHZ and 2.0 GHz RAM computer running on Win- dows 7 Ultimate 64-bit (6.1, Build 7601) (see Tables 1 and 2 for the confusion matrices). The best recognitions were detected in ha, su and di, where we obtained almost 100% in the JAFFE database. We saw that facial images with extreme exhibited expressions recorded the best re- sults. Generally, the performance of ne was the weakest; about 92.23% in JAFFE and 86.16 in Yale. The results show that the defor- mations of the muscles around the mouth and the eyes are the most reliable determinants for automatic facial expressions. This accounts why recognition in the neutral face is poor. Thus the in- crease in these muscle deformations increases the accuracies of automatic recognitions. The execution time for a pixel of size 100 � 100 is 14.5 ms. Fig. 5 shows the comparative performance of execution time with other neural network classifiers. Anger (%) Disgust (%) Happy (%) Surprise (%) 2.35 0 0 0 0.97 0 0 0 0 1.83 0 1.14 96.10 0 0 0 0.61 99.91 0 0.29 0 0 99.72 0 0 0.03 0 99.87 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 x 10 5 0 100 200 300 400 500 600 700 800 Image Pixels P ro ce ss in g T im e (m s) Proposed method Ma & Khorasani classifier Kharat & Dudul classifier Kumbhar et al. classifier Fig. 5. Comparing execution time of different classifiers. Fig. 6. Comparing recognition rates of different methods in JAFFE database. Fig. 7. Comparing recognition rates of different methods in JAFFE database. Fig. 8. Comparing the average recognition rates of different methods in JAFFE and Yale databases. 3388 E. Owusu et al. / Expert Systems with Applications 41 (2014) 3383–3390 The proposed method was compared with three different clas- sifiers to access its performance in terms of recognition accuracy and execution speed. The system was also tested with some real life images from the World Wide Web. The results indicate that the proposed method is statistically better (p < 0.05) both in accu- racy and speed. The three methods are described as follows: Method I (same as Ma and Khorasani (2004) method): The fea- ture detector is by a discrete cosine transform (DCT), pruning tech- nique is used to reduce the input size of the network, the training algorithm is back-propagation procedure, the expression classifier is a feed-forward neural network with one hidden layer. Method II (same as Kumbhar et al. (2012) method): The feature detector is Gabor method, the feature dimensionality is principal component analysis (PCA) and the expression classifier is a feed- forward neural network. Method III (same as Kharat and Dudul (2009)): The feature detector is discrete cosine transform, the feature reduction is by principal component analysis (PCA), and the classifier is a feed-for- ward neural network. All the classifiers were trained and tested with the same data- sets used for the proposed method. Figs. 6 and 7 shows the average recognition rates of various expressions in JAFFE and Yale respec- tively; the intent is not to compare the performances of the two databases but to investigate for the robustness of the system in di- verse databases. Fig. 8 also shows the comparison of the overall average recognition rates of various descriptors in JAFFE and Yale. Fig. 9 shows sample real-time expression recognitions by the sys- tem. The average recognition rates are also compared with other methods that employed the same datasets in their experiment to give a general idea of the performance of the proposed method (see Tables 3 and 4 for details). However this does not signify a di- rect comparison because the experiments were not conducted un- der the same environment. The results show that the proposed method is very encouraging. Though performance in the Yale facial expression database is reduced as compared to that in the JAFFE fa- cial expression database, it is far better than all the performances in Yale we compared with. Fig. 9. Sample facial expression recognitions: happy (left), anger (middle), disgust (right). Table 3 Comparative performance of recognition rates in different methods on JAFFE database. Author Classifier/ method Database Rate (%) Lekshmi and Sasikumar (2009) SVM JAFFE 86.9 Kumbhar et al. (2012) Image feature JAFFE 60–70 Zhi and Ruan (2008) 2D-DLPP JAFFE 95.91 Zhao, Zhuang, and Xu (2008) PCA and neural network JAFFE 93.72 Lee, Huang, and Shih (2010) RDA JAFFE 96.7 Proposed method MFFNN JAFFE 96.81 Table 4 Comparitive performance of recognition rates in different methods on the Yale database. Author Classifier/method Database Rate (%) Lekshmi and Sasikumar (2009) SVM Yale 89.5 Franco and Treves (2001) Neural network Yale 84.5 Proposed method MFFNN Yale 91.52 E. Owusu et al. / Expert Systems with Applications 41 (2014) 3383–3390 3389 7. Conclusions This study employs many advanced techniques to improve the recognition rate and execution time of facial expression recogni- tion system. Face detection was carried out by the application of Viola–Jones descriptor. Detected faces were down-sampled by the Bessel transform. This approach reduced the image dimensions and preserved the perceptual quality of the original image. An Ada- Boost based algorithm was formulated to select a few hundreds of Gabor wavelets from the several thousands of the extracted fea- tures to reduce the computational cost and to avoid misclassifi- cation as well. The selected features were fed into a well- designed multilayer feed-forward neural network classifier. The network is thus trained with sample datasets from both JAFFE and Yale facial expression databases. The remaining datasets from the two databases and some images from the World Wide Web were used to test for the system. The execution time for a pixel of size 100 � 100 is 14.5 ms; the average recognition rate in JAFFE database is 96.83% and that in Yale is 92.22%. The proposed method is compared with several methods and the performance is out- standing. The results of the study also show that automatic expres- sion recognitions are very accurate in surprise, disgusts and happy; about 100%. Mild expressions like sad, fear and neutral have lower accuracies. However fear can be very accurate when it is at the peak because accuracies in recognitions largely depend on the magnitude of facial deformations around the mouth and eyes. To advance towards 100% efficiency we believe the development of natural databases would be of more help since many artificial dat- abases have many confused scenarios among facial expressions in sad, neutral and mild anger. Again future improvements of recog- nition accuracies will look at the possibility of increasing the num- ber of hidden neurons to expressions that recorded lower values. Acknowledgments This paper is supported by the National Nature Science Founda- tion of China (Nos. 61272211 and 61170126), the Natural Science Foundation of Jiangsu Province (No. BK2011521), and the Research Foundation for Talented Scholars of Jiangsu University (No. 10JDG065) References Aghajanian, J., Warrell, J., Prince, S. J., Li, P., Rohn, J. L., & Baum, B. (2009). Patch- based within-object classification. In: IEEE 12th International Conference on Computer Vision, 2009 (pp. 1125–1132). Al Daoud, J. E. (2009). Enhancement of the face recognition using a modified Fourier–Gabor filter. International Journal of Advanced Software Computer Applications, 1. Bouzalmat, A., Belghini, N., Zarghili, A., Kharroubi, J., & Majda, A. (2011). Face recognition using neural network based Fourier Gabor filters & random projection. International Journal of Computer Science and Security, 5, 376–386. Breazeal, C., & Scassellati, B. (2002). Robots that imitate humans. Trends in Cognitive Sciences, 6, 481–487. Bruce, V. (1993). What the human face tells the human mind: Some challenges for the robot–human interface. Advanced Robotics, 8, 341–355. Cemre, Z. (2008) Facial Expression Recognition. MSc Thesis, University of Surrey. Choi, E., & Lee, C. (2003). Feature extraction based on the Bhattacharyya distance. Pattern Recognition, 36, 1703–1709. Dailey, M. N., & Cottrell, G. W. (1999). PCA = Gabor for expression recognition. UCSD CSE TR CS-629. Deng, H. B., Jin, L. W., Zhen, L. X., & Huang, J. C. (2005). A new facial expression recognition method based on local Gabor filter bank and PCA plus LDA. International Journal of Information Technology, 11, 86–96. Franco, L., & Treves, A. (2001). A neural network facial expression recognition system using unsupervised local processing. In: IEEE Proceedings of the 2nd International Symposium on Image and Signal Processing and Analysis (pp. 628– 632). Freund, Y., & Schapire, R. E. (1995). A desicion-theoretic generalization of on-line learning and an application to boosting. In Computational Learning Theory (pp. 23–37). Berlin, Heidelberg: Springer. Ganga, M. P., Prakash, C., Gangashetty, S. V. (2011). Bessel transform for image resizing. In: IEEE 18th International Conference on Systems, Signals and Image Processing, (pp. 1–4). Jones, J. P., & Palmer, L. A. (1987). An evaluation of the two-dimensional Gabor filter model of simple receptive fields in cat striate cortex. Journal of Neurophysiology, 58, 1233–1258. http://refhub.elsevier.com/S0957-4174(13)00961-5/h0010 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0010 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0010 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0015 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0015 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0015 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0020 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0020 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0025 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0025 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0035 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0035 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0045 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0045 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0045 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0055 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0055 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0055 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0065 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0065 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0065 3390 E. Owusu et al. / Expert Systems with Applications 41 (2014) 3383–3390 Kharat, G. U., & Dudul, SV. (2009). Emotion recognition from facial expression using neural networks. Human–Computer Systems Interaction. Berlin, Heidelberg: Springer (pp. 207–219). Berlin, Heidelberg: Springer. Kumbhar, M., Jadhav, A., & Patil, M. (2012). Facial expression recognition based on image feature. International Journal of Computer and Communication Engineering, 1, 117–119. Lades, M., Vorbruggen, J. C., Buhmann, J., Lange, J., von der Malsburg, C., Wurtz, R. P., et al. (1993). Distortion invariant object recognition in the dynamic link architecture. IEEE Transactions on Computers, 42, 300–311. Lee, C. C., Huang, S. S., & Shih, C. Y. (2010). Facial affect recognition using regularized discriminant analysis based algorithms. EURASIP Journal on Advances in Signal Processing, 1. Lekshmi, V. P., & Sasikumar, M. (2009). Analysis of facial expression using Gabor and SVM. International Journal of Recent Trends in Engineering, 2. Londhe, R., & Pawar, V. (2012). Facial expression recognition based on Affine Moment Invariants. International Journal of Computer Science Issues, 9. Ma, L., & Khorasani, K. (2004). Facial expression recognition using constructive feed- forward neural networks. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, 34, 1588–1595. Maalej, A., Amor, B., Daoudi, M., Srivastava, A., Berretti, S. (2010). Local 3D shape analysis for facial expression recognition. In: IEEE 20th International Conference on Pattern Recognition (pp. 4129–4132). Morik, K., Brockhausen, P., & Joachims, T. (1999). Combining statistical learning with a knowledge-based approach – A case study in intensive care monitoring. In: ICML (pp. 268–277). Munoz, A., Blu, T., & Unser, M. (2001). Least-squares image resizing using finite differences. IEEE Transactions on Image Processing, 10, 1365–1378. Rucklidge, W. J. (1997). Efficient locating objects using Hausdorff distance. International Journal of Computer Vision, 24, 251–270. Satiyan, M., & Nagarajan, R. (2010). Recognition of facial expression using Haar-like feature extraction method. In: IEEE International Conference on Intelligent and Advanced Systems (pp. 1–4). Sha, T., Song, M., Bu, J., Chen, C., & Tao, D. (2011). Feature level analysis for 3D facial expression recognition. Neurocomputing, 74, 2135–2141. Shen, L., & Bai, L. (2004). AdaBoost Gabor feature selection for classification. In: Proceeding of Image and Vision Computing, (pp. 77–83), New Zealand. Soyel, H., & Demirel, H. (2007). Facial expression recognition using 3D facial feature distances. Image Analysis and Recognition. Berlin, Heidelberg: Springer (pp. 831–838). Berlin, Heidelberg: Springer. Tai, S. C., & Chung, K. C. (2007). Automatic facial expression recognition system using Neural Networks. In: TENCON IEEE Region 10 Conference (pp. 1–4). Tang, H., & Huang, T. (2008). 3D facial expression recognition based on properties of line segments connecting facial feature points. In: 8th IEEE International Conference on Automatic Face & Gesture Recognition (pp. 1–6). Van, P. F. (2008). Discrete Wavelet Transformations: An Elementary Approach with Applications. Hoboken, NJ: Wiley-Interscience. Viola, P., & Jones, M. J. (2004). Robust real-time face detection. International Journal of Computer Vision, 57, 137–154. Wiskott, L., Fellous, J. M., Kruger, N., & vonderMalsburg, C. (1997). Face recognition by elastic bunch graph matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19, 775–779. Wu, J. X., Brubaker, S. C., Mullin, M. D., & Rehg, J. M. (2008). Fast asymmetric learning for cascade face detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30, 369–382. Zhang, D., & Lu, G. (2004). Review of shape representation and description techniques. Pattern Recognition, 37, 1–19. Zhang, L., & Tjondronegoro, D. (2009). Selecting, optimizing, and fusing ‘salient’ Gabor features for facial expression recognition. Neural Information Processing. Berlin, Heidelberg: Springer (pp. 724–732). Berlin, Heidelberg: Springer. Zhao, L., Zhuang, G., & Xu, X. (2008). Facial expression recognition based on PCA and NMF. In: IEEE 7th World Congress on Intelligent Control and Automation (pp. 6826–6829). Zhi, R., & Ruan, Q. (2008). Facial expression recognition based on two- dimensional discriminant locality preserving projections. Neurocomputing, 71, 1730–1734. http://refhub.elsevier.com/S0957-4174(13)00961-5/h0070 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0070 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0070 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0075 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0075 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0075 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0085 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0085 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0085 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0090 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0090 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0095 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0095 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0100 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0100 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0100 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0115 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0115 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0120 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0120 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0130 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0130 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0140 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0140 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0140 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0155 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0155 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0160 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0160 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0165 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0165 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0165 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0170 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0170 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0170 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0175 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0175 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0180 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0180 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0180 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0190 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0190 http://refhub.elsevier.com/S0957-4174(13)00961-5/h0190 A neural-AdaBoost based facial expression recognition system 1 Introduction 2 Face detection and down-sampling 3 Gabor feature extraction 4 Feature selection 5 Multilayer feed-forward neural network (MFFNN) classifier 6 Results and analysis 7 Conclusions Acknowledgments References 
work_ktt4ulqwjjbihpp37txhu7g6we	----	Microsoft Word - AIKED REVISED HEADER.doc Customer pattern search for A/S association in manufacturing JIN SOOK AHN, SO YOUNG SOHN* Department of Information & Industrial Systems Engineering Yonsei University 134 Shinchon-dong, Seoul 120-749 KOREA Abstract: Manufacturing corporations aim to sell their goods while they try to keep a sound customer relationship by providing high quality after-sales service (A/S). This is because while such services have always been important in marketing and sales industries, they are currently gaining importance in the manufacturing industry as well. Therefore it is important to identify the needs of different customer groups and to provide respective A/S for each group accordingly. In this study, we propose a framework that consists of fuzzy clustering and an association rule to identify customer groups and their needs. We first carried out fuzzy clustering of customers in terms of indicators of CSI (Customer Satisfaction Index). Next, the association rule is used to grasp the kind of A/S that customers consider important. Our results identified three groups of customers and their needs: Group 1 represents those who have a high degree of satisfaction, loyalty and high number of complaints. This group considers the home visiting service most important. Despite the fact that the Group 1 has high degree of complaint occurrence, they show a high degree of loyalty. Group 2 has very high degree of satisfaction and loyalty with a low level of complaints. This group considers important the A/S factors in all of service sections, including at the call center, the home visiting service, and claim handling. Group 3 has average satisfaction, number of complaints, and loyalty. Group 3 customers put weight on A/S factors dealing with the call center and the home visiting service. We expect that manufacturing firms can strengthen CRM (Customer Relationship Management) by offering tailored A/S for each group accordingly. Keywords: A/S (After-sales Service), fuzzy clustering, association rule, CRM 1 Introduction Currently, service programs are gaining importance in the manufacturing sector. Manufacturing firms make an effort to sell their goods, keep customer loyalty and create growth opportunities in markets. Accordingly, ‘servicization’ has become prevalent in the manufacturing industry [1, 2]. Excellent service affects the retention of existing customers and inducement of new customers. As a result, it brings high customer retention and satisfaction [3]. A/S (after sales service) can play a very important role in manufacturing firms, because 7th WSEAS Int. Conf. on ARTIFICIAL INTELLIGENCE, KNOWLEDGE ENGINEERING and DATA BASES (AIKED'08), University of Cambridge, UK, Feb 20-22, 2008 ISSN: 1790-5109 Page 182 ISBN: 978-960-6766-41-1 manufacturers can get information about their product and service from customers through the processes of A/S. Therefore, many manufacturing firms concentrate on improving customer satisfaction via A/S. However, customers’ needs for A/S are not homogeneous throughout. The main objective of this paper is to find the satisfying A/S characteristics for specific groups of customers. In this manner, we expect to provide customer-tailored A/S in manufacturing. In order to accomplish this objective, we propose a framework that consists of the clustering of customers in terms of CSI (Customer Satisfaction Index) toward the A/S provided, as well as extracting the association of the preferred kind of A/S for each clustering of customers. The result of this research is expected to offer useful strategic information for each cluster of customers. The organization of this paper is as follows. In In Section 2, we introduce a proposed framework which consists of the two steps. In Section 3, we apply the proposed approach to the A/S data of manufacturing companies in Korea. Finally, in Section 4, we summarize the study results and offer valuable information for the effective management of customers by manufacturing industries. 2 Research Framework We propose a research framework to analyze customer patterns in terms of A/S. The methodology is composed of two steps as shown in Fig.1. At Step 1, we divide customers who use A/S of manufacturing firms into several groups using Fuzzy C-Means (FCM) clustering via customer satisfaction, loyalty, complaints that can be obtained as resulting indicators of CSI. Fig. 1 Proposed research framework Various CSI models which are ACSI (American Customer Satisfaction Index) and ECSI (European Customer Satisfaction Index) use multiple indicators to measure overall customer satisfaction, complaints, and loyalty via customer expectation, perceived quality, and perceived value [4, 5]. They are measured in terms of the variables described in Table 1. The advantage of clustering analysis is useful to identify groups of customers with similar characteristics. In Step 2, we use association rules to determine which A/S operations each group of manufacturing customers thinks are very important. A/S operations in the call center, the home visiting service, and claim handling. Identification of association rules will provide feedback for strategic customer management of manufacturing firms. In the next section, we apply the proposed methodology to the case of survey data regarding the A/S of manufacturing companies in Korea. 3 Case Study In this section, we apply the proposed analysis framework to A/S activities of manufacturing companies in order to examine what kind of A/S 7th WSEAS Int. Conf. on ARTIFICIAL INTELLIGENCE, KNOWLEDGE ENGINEERING and DATA BASES (AIKED'08), University of Cambridge, UK, Feb 20-22, 2008 ISSN: 1790-5109 Page 183 ISBN: 978-960-6766-41-1 operations affects each group of manufacturing customers. 3.1 Data collection and pre-analysis In order to attain our study objective, we surveyed customers who have experienced A/S from two manufacturing companies. One produces a water purifier and a water softener. Customers who rented or purchased its products receive a regular A/S visit. The other produces home appliances such as TV’s, refrigerators, and air conditioners. It provides A/S in the form of a home visit upon request when products are out-of-order. The questionnaire covers evaluation components of resulting indicators (customer satisfaction, loyalty, complaints) of CSI models and A/S operations along with customers’ personal information. The measurement variables described in Table 1 and A/S factors in Table 2 were the components of involved quality with a corporation’s service [4, 5]. Each variable was measured in a 10 point scale. For example, if the overall satisfaction measurement variable is 10 points that means the customer is greatly satisfied, and when the frequency of complaints measurement variable is 10 points that means the customer is not at all satisfied. The survey was conducted in 2006 during the month of December through a professional survey agency. This agency completed the survey with an internal panel consisting of three-hundred seventy-six people who responded to an on-line web survey. Based on the survey data for A/S quality indicators, we conducted CFA (Confirmatory Factor Analysis) to evaluate the validity of latent factors in terms of measurement variables, as shown in Table 1. Table 1. Result of confirmatory factor analysis Latent Factor Measurement Variable Loading Values Customer Expectation ·overall expectation of companies’ service ·expectation regarding customization ·expectation regarding reliability 0.83 0.83 0.88 Perceived Quality ·overall evaluation of companies’ service ·evaluation of customization experience ·evaluation of reliability 0.93 0.92 0.93 Perceived Value ·rating of quality given price ·rating of price given quality 0.92 0.93 Customer Satisfaction ·overall satisfaction ·expectancy disconfirmation ·received service versus ideal service 0.91 0.57 0.61 Customer Complaints ·level of customer complaints to a company directly about a service ·frequency of customer complaints 0.71 1.00 Customer Loyalty ·likelihood to recommend other people to purchase companies’ services 0.95 7th WSEAS Int. Conf. on ARTIFICIAL INTELLIGENCE, KNOWLEDGE ENGINEERING and DATA BASES (AIKED'08), University of Cambridge, UK, Feb 20-22, 2008 ISSN: 1790-5109 Page 184 ISBN: 978-960-6766-41-1 ·likelihood to repurchase companies’ services 0.90 The loading value is the correlations between the measurement variables and the latent factor. For example, Table 1 shows the overall expectation of a companies’ service, expectation regarding customization, and expectation regarding reliability scales load onto the customer expectation factor. Our results indicate that the loading values are mostly above 0.8, which suggests confirmed validity about the composition of latent variables and the selection of measurement variables. 3.2 Fuzzy clustering In this section, we carry out clustering to divide customers into groups using confirmatory factor analysis. We use MATLAB to do FCM (Fuzzy C- Means) clustering using the latent factors such as customer satisfaction, complaints, and loyalty. The K-means algorithm is required to define the number (c) of initial seeds of FCM clustering. The FCM clustering uses c for k that results from K- means clustering [6, 7, 8]. Accordingly, to the result of K-means clustering, 3 is suitable for c. So, we use 3 initial seeds for FCM clustering. The result of FCM clustering is that the 376 customers were segmented into three groups, with 106 customers belonging to Group 1, 166 customers belonging to Group 2, and 104 customers belonging to Group 3. The characteristics of each group are as follows: Group 1: The average customer satisfaction and loyalty are about 8 points, and complaint is about 7 points on average. It means that customers grumble about the company’s services very often, although the overall satisfaction level is high. Group 2: The average of customer satisfaction and loyalty are 8 points, but complaint is on average 2 points. Customers of this group are very positive about the service of the company. Group 3: The customer satisfaction, loyalty and complaint values are about 5 points on average. They are neutral about the service of the companies. 3.3 Association rule In this section, we extract the association rules from each cluster of FCM. We want to find out what kind of A/S operations customers of each group think are important. In Table 2, A/S operational characteristics are classified into call center, home visiting service, and claim handling. We analyze association rules using SAS Enterprise Miner. Because many A/S operations are considered, we consider four items in the association rules and use confidence levels of 80%. Table 2. The A/S operational characteristics Field Requirement A/S operations convenience of use conversation by phone at night and on holidays, various and simple methods of inquiry speed of access short call waiting time, short ARS opening line specialty of service representative understanding customer needs, knowledge of company rules, knowledge and technical complexity of product/service, using easy terminology, ability to communicate speed of answering and handling speed of handling problems and requirements, speed of connecting other A/S departments Call center positiveness of sufficient explanation, leading counseling, satisfactory call (making a call 7th WSEAS Int. Conf. on ARTIFICIAL INTELLIGENCE, KNOWLEDGE ENGINEERING and DATA BASES (AIKED'08), University of Cambridge, UK, Feb 20-22, 2008 ISSN: 1790-5109 Page 185 ISBN: 978-960-6766-41-1 service representative to check customer satisfaction after A/S) kindness of service representative kindness of speech and sound, proper pace and pronunciation, listening to customer needs carefully, sympathy and response to customer needs, sincere explanation convenience of use providing service at night and on holidays timing of appointment confirmation of visit appointment and visit, pre-contact of visit delay, abiding by visiting time and delayed visiting time visiting technician’s ability technical competence, performing exact A/S, using simple terminology, technical responsiveness speed of handling prompt visit, promptly performing A/S and supplying of parts visiting technician’s positiveness sufficient explanation before and after repair Field service visiting technician’s kindness and neatness showing identification, good attitude, cleanliness, listening to customer needs carefully, technician’s appearance convenience of claim handling simple method of inquiry, claim contacts, simple compensation procedure, providing information on handling procedure continuously speed of answer and handling speed of registration, providing adequate response, speed of compensation procedure Claim handling propriety of claim handling mediation from consumer protection board, possession of a certification for protecting customer rights, attentiveness to employee reaction to claim Table 3 shows the result of the association rules in each group. Table 3. The result of association rules in each group Group A/S factors Support (%) Confidence (%) Lift 1 providing adequate response, sufficient explanation after repair & promptly peA/S => providing service at holiday 20.72 100 1.36 2 attentiveness to employee reaction to claim, providing adequate response & providing service at holiday => sufficient explanation 28.92 82.76 1.02 3 technical competence, providing service at holiday & sufficient explanation => handling problems 20.19 84 1.3 Customers in Group 1 think the following A/S operations as important: providing an adequate response in claim handling, sufficient explanation after repair, promptly performing A/S, and providing service during a holiday. The important A/S operations in Group 2 are as follows: nipping employee’s attitude in the bud about a claim, adequate response in claim handling, and providing home visiting service during a holiday. Customers in this group also consider sufficient explanations from the service representative at the call center. Group 3 thinks that technical competence of the visiting technician, providing home visiting service during the holiday, and sufficient explanations from the call center are the most important A/S. Then customers in Group 3 concentrate on the speed of handling problems at the call center. Generally, a high rate of support and confidence has a strong mutual relation. But significant association rules are those which have a lift value higher than 1. All groups showed lift above 1. Therefore, we can say that all groups have a strong relationship. 4 Conclusion 7th WSEAS Int. Conf. on ARTIFICIAL INTELLIGENCE, KNOWLEDGE ENGINEERING and DATA BASES (AIKED'08), University of Cambridge, UK, Feb 20-22, 2008 ISSN: 1790-5109 Page 186 ISBN: 978-960-6766-41-1 Lately, manufacturing firms make an effort to sell their products, maintain customer relationships, and improve market competitiveness through A/S. These efforts of manufacturing firms contribute toward keeping existing customers and attracting new customers. As a result, they grow in the marketplace and receive high levels of return. Manufacturing companies can gain information about the design of products and service, quality, selling, and marketing activity from customers. Also, corporations achieve higher customer satisfaction and loyalty via A/S. So, many manufacturing corporations should concentrate on A/S. In this research, we used fuzzy clustering and association rules to find out which A/S operations each group of manufacturing customers thinks are very important. Group 1 considers the home visiting service most important. Customers in Group 1 grumble about the company’s service very often, but they have a high level of loyalty. This is because manufacturing firms provide better service for customers and do not neglect customer's requirements. Group 2 has a very high level of satisfaction and loyalty along with a low level of complaints. This group considers important A/S factors in all of the service sections, including at the call center, the home visiting service, and claim handling. Many firms think that the Group 2 is the best customers. However, customers in Group 2 do not give any critical feedback to firms. Companies need to be informed about customers’ inconvenience, complaints, and requirements. We expect that manufacturing firms can strengthen customer relationship by providing improved services and offering tailored A/S for each group. Previous studies pursued the enhancement of company competitiveness through improving the quality of products. But this research suggests that focusing on A/S in manufacturing corporations is of increasing importance to improve competitiveness. We expect that the methods and results of this research will be valuable not only for A/S in manufacturing but also for the whole service industry. Also, this research suggests a powerful strategy is one of keeping existing customers through paying careful attention to customer responses. Further research will apply these methods to an entire service area such as mobile communication or internet service. Acknowledgement: This work was supported by Korea Sanhak Foundation (2007) References [1] S. Brax, A manufacturer becoming service provider-challenges and a paradox, Managing Service Quality, Vol.15, No.2, 2005, pp.142-155. [2] J. B. Zhang, B. T. J. Ng, M. M. Wong, and L. Q. Zhuang, Manufacturing service negotiation and resource management: A QoS Approach, International Conference on Control and Automation (ICCA2005), Budapest, Hungary, 2005. [3] E. W. Anderson, C. Fornell, and D. R. Lehmann, Customer Satisfaction, Market Share and Profitability: Findings from Sweden, Journal of Marketing, Vol. 58, 1994, pp. 58-66. [4] C. Fornell, M. D. Johnson, E. W. Anderson, J. Cha, and B. E. Bryant, The American Customer Satisfaction Index: Nature, purpose and findings, Journal of Marketing, Vol.60, 1996, pp.7-18. [5] K. Kristensen, A. Martensen, and L. GrØnholdt, Customer satisfaction measurement at post Denmark: Results of application of the European Customer Satisfaction Index methodology, Total Quality Management, Vol.11, No.7, 2000, pp.S1007-S1015. [6] J.C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms, Plenum Press, New York, 1981. 7th WSEAS Int. Conf. on ARTIFICIAL INTELLIGENCE, KNOWLEDGE ENGINEERING and DATA BASES (AIKED'08), University of Cambridge, UK, Feb 20-22, 2008 ISSN: 1790-5109 Page 187 ISBN: 978-960-6766-41-1 [7] M. Roubens, Fuzzy clustering algorithms and their cluster validity, European Journal of Operational Research, Vol.10, 1982, pp.294-301. [8] S. A. Mingoti, and J. O. Lima, Comparing SOM neural network with Fuzzy c-means, K-means and traditional hierarchical clustering algorithms, European Journal of Operational Research, Vol.174, 2006, pp.1742-1759. 7th WSEAS Int. Conf. on ARTIFICIAL INTELLIGENCE, KNOWLEDGE ENGINEERING and DATA BASES (AIKED'08), University of Cambridge, UK, Feb 20-22, 2008 ISSN: 1790-5109 Page 188 ISBN: 978-960-6766-41-1 
work_ktzd2efr4jbidabo2vk3lxcvmq	----	Predicting Forest Regeneration in the Central Appalachians Using the REGEN Expert System Journal of Sustainable Forestry, 30:790–822, 2011 Copyright © Taylor & Francis Group, LLC ISSN: 1054-9811 print/1540-756X online DOI: 10.1080/10549811.2011.577400 Predicting Forest Regeneration in the Central Appalachians Using the REGEN Expert System LANCE A. VICKERS1, THOMAS R. FOX1, DAVID L. LOFTIS2, and DAVID A. BOUCUGNANI3 1Virginia Polytechnic Institute and State University, Department of Forest Resources and Environmental Conservation, Blacksburg, Virginia, USA 2USDA Forest Service, Southern Research Station, Bent Creek Experimental Forest, Asheville, North Carolina, USA 3SereneTech Consulting, Watkinsville, Georgia, USA REGEN is an expert system designed by David Loftis to pre- dict the future species composition of dominant and codominant stems in forest stands at the onset of stem exclusion following a proposed harvest. REGEN predictions are generated using compet- itive rankings for advance reproduction along with other existing stand conditions. These parameters are contained within modu- lar REGEN knowledge bases (RKBs). To extend REGEN coverage into hardwood stands of the Central Appalachians, RKBs were developed for four site classes (xeric, subxeric, submesic, mesic) based on literature and expert opinion. Data were collected from 48 paired stands in Virginia and West Virginia to calibrate the initial RKBs. Paired stands consisted of one mature uncut hard- wood stand adjacent to a regenerating clear-cut stand with similar Joint funding for this project was provided by the United States Forest Service and the Appalachian Hardwood Forest Research Alliance (AHFRA). The assistance of the fol- lowing AHFRA members is greatly appreciated: Curt Hassler of AHFRA, Jay Engle and Wesley Johnson of MeadWestvaco Corportation, Doug Toothman of Western Pocahontas Properties LP, and Theresa Henderson and Jason Weinrich of The Forestland Group LLC. The authors also thank Bob Radspinner of Plum Creek Timber Company Inc., Glen Jeurgens of the Monongahela National Forest, Thomas Schuler of the Fernow Experimental Forest, and Edward Leonard of the George Washington Jefferson National Forest. The comments and suggestions of the reviewers certainly improved the quality of this article and are appreciated. REGEN is available for public download at: http://regen.boucugnani.com. RKBs are available from the corresponding author. Address correspondence to Lance A. Vickers, Virginia Polytechnic Institute and State University, Department of Forest Resources and Environmental Conservation, 228 Cheatham Hall (0324), Blacksburg, VA 24061, USA. E-mail: vickersl@vt.edu 790 D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 791 site characteristics that was harvested within the previous 20 yr. Data from 17 additional paired stands was collected a year later to validate the performance of REGEN. Predicted values were within 4 percentage points of measured values on average, and model error was typically less than 20 percentage points for species groups. These results confirmed the suitability of REGEN to pre- dict the future species composition of stands regenerated using the clear-cut method in the Central Appalachians of Virginia and West Virginia. KEYWORDS hardwoods, Virginia, West Virginia, clear-cutting, harvesting, regeneration model, silviculture, species composition, oak, forest management, sustainable INTRODUCTION Appalachian Hardwood Regeneration The mixed hardwood forests of the Appalachians are extremely diverse (Miller & Kochenderfer, 1998). Individual stands can often contain more than 50 different species that vary in shade tolerance, growth rates, and regeneration strategies (Braun, 1950; Burns & Honkala, 1990; Smith, 1994). The long disturbance history of the Appalachians has increased the com- plexity of these forests (Yarnell, 1998). The regeneration potential of each species in a stand following disturbance is largely determined by shade tol- erance, individual growth rates, and the type of disturbance (Smith, Larson, Kelty, & Ashton, 1996). Fast growing shade-intolerant species such as black cherry (Prunus serotina Ehrh.) and yellow-poplar (Liriodendron tulipifera L.), often regenerate following a heavy disturbance from new seedlings that develop from seed recently dispersed and stored in the forest floor. In con- trast, slower growing, more shade-tolerant species such as oaks (Quercus spp.) rely on the buildup of large advance reproduction and timely release to compete successfully following a similar disturbance unless site factors offer an ecological advantage to these species (Sander, 1972; Loftis, 1983; Johnson, Shifley, & Rogers, 2002). Site productivity, primarily determined by moisture availability, is a sig- nificant driver of species composition and regeneration potential in the Appalachians (McNab, 1988). Naturally occurring species assemblages are often delineated by moisture availability (Eyre, 1980). Smith (1994) classified forests of the Southern Appalachians into four groups based on productivity (site index) that was largely determined by moisture availability, essen- tially occurring on sites with xeric, subxeric, submesic, and mesic moisture regimes. Site index for white oak (Quercus alba L.) at age 50 averaged less than about 17 m in the xeric communities and more than about 26 m in the mesic communities (Smith, 1994). D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 792 L. A. Vickers et al. Hardwood Regeneration Models Estimates of future species composition following harvest are important to forest managers and silviculturists to help ensure landowner objectives are met in a predictable and sustainable fashion. Because of the com- plexity of many Eastern hardwood forests, however, it is often difficult to reliably predict the future species composition that will exist following a pro- posed harvest using natural regeneration. Several modeling approaches have been used to predict regeneration in oak dominated ecosystems (Rogers & Johnson, 1998). These tools typically follow the model of forest succes- sion based on initial floristics (Egler, 1954), utilize stand stocking guides (e.g., Gingrich, 1967), and incorporate the concept that advance reproduc- tion size impacts regeneration success (Sander, 1972). Probabilistic models and management guidelines have been developed to evaluate regenera- tion potential in primarily oak dominated Central hardwood forests (Sander, Johnson, & Watt, 1976), the Missouri Ozarks (Sander, Johnson, & Rogers, 1984), the Southern Bottomlands (Belli, Hart, Hodges, & Stanturf, 1999), the Alleghenies (Brose et al., 2008), and the Central Appalachians (Steiner, Finley, Gould, Fei, & McDill, 2008). Other models that produce only quanti- tative estimates have been published as well (McQuilkin, 1975; Loftis, 1990; Gould, Steiner, Finley, & McDill, 2005). Loftis (1990) and McQuilkin (1975) developed single species models for some oaks. Gould, Steiner, McDill, and Finley (2006) described a methodology for modeling seed origin oak regeneration. In addition, models have been developed solely to quantify contributions to future composition from stump-sprouting (Johnson, 1977; Dey, Johnson, & Garrett, 1996; Gould, Fei, & Steiner, 2007). Most of these tools are limited to oak management. In some cases, multispecies, regional predictive models have been developed (Waldrop, Buckner, Shugart, & McGee, 1986; Dey, 1991), but such models are not available for all areas. REGEN Interest in a multispecies, regional predictive model for the Southern Appalachians fostered the development of the REGEN model (Loftis, 1989). The REGEN model is an expert system designed by David Loftis to pre- dict the species composition of upper canopy (dominant and codominant) stems at the onset of stem exclusion following a proposed harvest. REGEN uses numerical rankings of expected species competitiveness in conjunc- tion with existing stand characteristics to generate predictions. Boucugnani (2005) described the development of the computer adaptation of REGEN and documented the underlying framework of the model. A survey of advance reproduction prior to harvest is required in order to generate predictions in REGEN. Advance reproduction is categorized into five default size classes: germinants (newly germinated seedlings), small seedlings (<6-cm tall), medium seedlings (61- to 122-cm tall), large seedlings D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 793 (≥122-cm tall), and potential stump-sprouts (trees >122-cm tall and ≥5-cm dbh). REGEN allows for probabilistic establishment of stump-sprouts, root-suckers, and new seedlings following harvest using constant or logistic parameters along with multiple simulation runs of the input data. Each plot is stochastically populated with a number of root-suckers and new seedlings established postharvest in addition to the existing advance reproduction pool using those parameters during the prediction process. Advance reproduc- tion of a particular species must be present in the plot before root-suckers of that species are added into the regeneration pool. Those parameters are also used to determine which existing stems are expected to produce a stump-sprout for each simulation run. REGEN is a competition based model. Each species-size-source com- bination of reproduction is given a competitive ranking ranging from 1 to 8 decreasing in competiveness; i.e., individuals ranked 1 will outcompete those ranked 2, individuals ranked 2 will outcompete those ranked 3, and so forth. Competition is simulated at the plot level, and future upper canopy species composition is predicted based on the relative competitive rankings of the regeneration pool at each sample plot. REGEN populates each predic- tion plot by identifying up to six “winning” stems for a 0.01 ac (40.5 m2) plot from the reproduction pool based on competitive rank. Individuals ranked 1 are selected first. If the threshold of six is not met, the selection process moves on to individuals ranked 2, then 3, and so forth until all six “win- ning” stems have been selected. In the case of ties for the selection of the final “winning” stem, each equally ranked individual is selected proportion- ally (e.g., if there are five individuals of equal rank, each one is selected as 0.2 of an individual). This fragmentation is resolved when “winning” stems per plot are scaled to stems per acre (0.40 ha) in the summary output. On plots where stump-sprouts are chosen, fewer “winning” stems are added to compensate for the greater space requirements of stump-sprouts. If one stump-sprout is selected, only three additional individuals will be selected. If two stump-sprouts are selected, only one additional individual is selected. If three or more stump-sprouts are selected, the top two are chosen and the remaining “winning” stem is selected proportionally using the tiebreaker logic described earlier. The stochastic element and numerous simulation runs of input data allow summary statistics to be provided with model output. A REGEN Knowledge Base (RKB) contains all competitive rankings and parameters used to process input data. RKBs are modular, which allows REGEN to be adapted to different regions by creating custom RKBs. Efforts are currently underway to expand REGEN applicability through- out the Appalachians. This article documents work to adapt REGEN to hardwood forests in the Central Appalachians of Virginia and West Virginia. D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 794 L. A. Vickers et al. METHODS Model Building To adapt REGEN to the Appalachians of Virginia and West Virginia, four pre- liminary RKBs were created in an attempt to address species variability that resulted from site productivity differences associated with moisture availabil- ity. The four RKBs were designed to predict regeneration on xeric, subxeric, submesic, and mesic sites. These four site classes roughly approximate the four forest types described by Smith (1994) and were delineated using the methodology proposed by McNab, Loftis, and Shefield (2003). In this methodology, certain species serve as indicators of moisture availability and provide insight into relative site quality (McNab et al., 2003). The initial rankings and parameters for these four RKB’s were based on relevant liter- ature, including: silvical characteristics (Burns & Honkala, 1990), reported species composition following clear-cutting (e.g., Beck & Hooper, 1986; Ross, Sharik, & Smith, 1986; Loftis, 1989), and site productivity interactions (e.g., Doolittle, 1958). Stump-sprouting, root-suckering, and new seedling establishment prob- abilities were initially taken from the literature when possible. The Silvics Manual in particular served as a source of information for several species (Burns & Honkala, 1990). However, because data found in existing litera- ture was often from different forest types and climates, the information from these sources was modified or supplemented as necessary, based on the experience of the authors to create the four custom RKBs. In these RKBs, stump-sprouts were considered the most competitive source of regeneration for a species when applicable, and rankings subsequently decreased for smaller size classes. Although competition within a species (large vs. small seedling) was fairly straightforward, the rankings also had to maintain com- petitive relationships amongst species across all size classes. Therefore, the difference in competitive rank between a large stem of advanced reproduc- tion compared to a medium stem, for example, may not be uniform across all species. Because these rankings were designed to predict upper canopy species composition at the onset of stem exclusion, species that may be numer- ous in the lower canopy and possibly ascend into the upper canopy in later stages of development were not necessarily given a high rank. Conversely, early successional species, which were expected to be numer- ous, but ultimately short lived in the upper canopy, were ranked highly. Generally, more mesic species such as basswood (Tilia spp.) decreased in rank as moisture availability decreased, while xeric species such as Virginia pine (Pinus virginiana Mill.) increased in rank. Sweet birch (Betula lenta L.) and yellow-poplar were expected to be strong initial competitors and were given a superior rank in all site classes except for xeric. Black locust (Robinia pseudoacacia L.) and eastern white pine (P. strobus L.) D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 795 were also among the most competitive species across all site classes. Red maple (Acer rubrum L.) sprouts were considered very competitive across all four site classes. Oaks were among the most competitive species in the xeric and subxeric RKBs and decreased as moisture availability increased. Chestnut (Quercus prinus L.) and scarlet (Q. coccinea Muenchh.) oaks were more competitive in the xeric and subxeric classes and decreased as mois- ture availability increased, while northern red oak (Q. rubra L.) became increasingly competitive up to the submesic RKB. Following the development of the preliminary knowledge bases, field data were collected to evaluate the performance of the initial parameters and to serve as a database to improve those parameters with subsequent itera- tive trials. This model calibration data set was collected using a paired stand sampling approach. This approach utilized sample sites which consisted of a mature hardwood stand free from obvious recent disturbance adjacent to a regenerating clear-cut stand of similar site characteristics. According to available landowner records, the mature stands were no less than 70 yr old, while the regenerating clear-cut stands ranged in age from 5 to 20 yr old. This approach assumed that the two stands were once contiguous and of similar composition and productivity. It was further assumed that any mature stand would regenerate in a similar fashion to its paired regenerating clear-cut stand when harvested similarly. Slope, aspect, and landscape position were used to indirectly estimate site index using the Forest Site Quality Index (Meiners, Smith, Sharik, & Beck, 1984) to ensure similar site productivity between paired mature and regenerating clear-cut stands. This methodol- ogy provides estimates of site index in areas of similar rainfall based on moisture retention potential (Meiners et al., 1984). Paired stands were also classified into site classes using indicator species as proposed by McNab et al. (2003) in their ecological classification system. The species mois- ture weights from their methodology was applied to a list of sampled tree species that were ≥5-cm dbh in each mature stand to classify each paired stand into one of four moisture regimes: xeric, subxeric, submesic, and mesic. Sampling Sites and Procedures A total of 48 paired stands were located in Virginia and West Virginia to serve as a source of data to calibrate our RKBs (Figure 1). Forty-five of the 48 paired stands were located within the Appalachian Plateau Physiographic Province, while the remaining 3 were located in the Ridge and Valley Province (Fenneman, 1938). The difference in estimated oak site index between all paired mature and regenerating clear-cut stands was about 0.9 m on average. A maximum difference of about 4 m was estimated for one pair. All paired stands in this study fell on subxeric or submesic sites, with 38 paired stands on submesic sites and 10 paired stands on subxeric sites. D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 796 L. A. Vickers et al. FIGURE 1 Location of sample sites in the Central Appalachians. Shaded areas represent counties in which paired stands were located. For this reason, only the subxeric and submesic RKBs were tested with field data. Trends observed in these site classes were extended into the xeric and mesic RKBs according to the authors expectations, but those RKBs have not been field tested. The paired stands for model calibration were sampled between May and September, 2008. In the mature stands, about 2 fixed-radius plots per hectare were established with a maximum of 20 plots per stand. Plots were established using a systematic grid with a plot spacing of 50.3 m and a row spacing of 80.5 m. The distance from the stand boundary to the first plot was randomly determined but no less than 40.2 m. Advance repro- duction was measured in a 16.2 m2 fixed-radius circular plot by species and REGEN size class (Large: ≥122 cm, Medium: ≥61cm, Small: <61 cm, Germinant: newly germinated seedlings), and dbh of all stems ≥5-cm dbh within the plot was also measured. At each fixed-radius plot center, a basal area sampling point was also established on which the basal area of all stems ≥5-cm dbh was measured using a 2.3 BAF prism. Advance repro- duction from the mature stand inventories was later scaled up to represent 40.5 m2 (0.01 ac) to meet the input requirements of the REGEN computer D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 797 program. Regenerating clear-cut stands were sampled at a density of about 2 fixed-radius circular plots per hectare, using 4 m2 plots unless stands had developed such that plots of this size were frequently unpopulated. In this case, sample plots were reestablished as 16.2 m2 plots throughout the stand. In the regenerating stands, regeneration was tallied by species, stem ori- gin (seed or sprout), and crown class (dominant, codominant, intermediate, suppressed). An additional 17 paired stands were sampled in September and early October of 2009 to serve as a model validation data set. Twelve paired stands were located within the Appalachian Plateau Physiographic Province, while the remaining 5 were located in the Ridge and Valley Province (Fenneman, 1938). Sampling procedures were identical to those described above for the model calibration data set, with the exception of fixed-radius plot sizes. In the validation paired stands, advance reproduction was tallied on all mature stands using 40.5 m2 circular plots. On the regenerating clear-cut stands, regeneration was tallied on 16.2 m2 plots exclusively. The difference in estimated oak site index between these additional paired mature and regenerating clear-cut stands was about 0.3 m on average, with a maximum of about 0.3 m. In the validation data set, 14 of the stands were classified submesic, with the remaining 3 as subxeric. Because of the complex assemblages of species that occur in this area, eight species groups were created for the purposes of data analysis and presentation following data collection: (a) black cherry, (b) conifers, (c) maples, (d) midstory, (e) oaks, (f) other overstory, (g) pioneer, and (h) yellow-poplar (Table 1). Species groups were composed of either a single species that was numerous throughout the area or multiple species that are expected to occupy similar stand structural positions. The composition of the paired stands examined in this study was typical of Central Appalachian hardwoods. Mean basal area of the 48 mature stands in the model calibration dataset was 27.1 m2 per hectare dominated by oaks and maples. These stands had 914 stems per hectare ≥5-cm dbh, with maples making up the greatest proportion (Table 2). In the 48 regenerating clear-cut stands in the model calibration data set, maples and pioneer species were the most numerous in the upper canopy (dominant and codominant), while maples and midstory species were the most numerous in the lower canopy positions (intermediate and suppressed; Table 3). The composition of the paired stands in the validation data set was similar to those in the model calibration data set. Oaks and maples dominate the 28 m2 per hectare of basal area in the 17 mature stands in the validation data set (Table 2). In the 17 regenerating clear-cut stands in the validation data set, maples and pioneer species were the most numerous groups in the upper canopy, and maples and midstory were the most numerous groups in the lower canopy (Table 3). D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 798 L. A. Vickers et al. TABLE 1 Species Groupings Used in This Study for the Central Appalachians Group Species Black cherry Black cherry (Prunus serotina Ehrh.) Conifers E. white pine (Pinus strobus L.), hemlock (Tsuga spp.), pitch pine (Pinus rigida Mill.), red spruce (Picea rubens Sarg.), shortleaf pine (Pinus echinata Mill.), Table Mountain pine (Pinus pungens Lamb.), Virginia pine (Pinus virginiana Mill.) Maples Red maple (Acer rubrum L.), sugar maple (Acer saccharum Marsh.) Midstory Am. chestnut (Castanea dentata [Marsh.] Borkh.), Am. holly (Ilex opaca Ait.), beech (Fagus grandifolia Ehrh.), blackgum (Nyssa sylvatica Marsh.), dogwood (Cornus spp.), ironwood (Ostrya virginiana [Mill.] K. Koch.), sassafras (Sassafras albidum [Nutt.] Ness.), serviceberry (Amelanchier spp.), sourwood (Oxydendrum arboreum [L.] DC.), striped maple (Acer pensylvanicum L.) Oaks Black oak (Quercus velutina Lam.), chestnut oak (Quercus prinus L.), n. red oak (Quercus rubra L.), scarlet oak (Quercus coccinea Muenchh.), white oak (Quercus alba L.) Other overstory Ash (Fraxinus spp.), basswood (Tilia spp.), buckeye (Aesculus flava Ait.), cucumbertree (Magnolia acuminata L.), Fraser magnolia (Magnolia fraseri Walt.), hickory (Carya spp.), yellow birch (Betula alleghaniensis Britton) Pioneer Ailanthus (Ailanthus altissima [Mill.] Swingle), American sycamore (Platanus occidentalis L.), bigtooth aspen (Populus grandidentata Michx.), black locust (Robinia pseudoacacia L.), fire cherry (Prunus pensylvanica L.f.), royal paulownia (Paulownia tomentosa [Thunb.] Sieb. & Zucc. Ex Steud.), sweet birch (Betula lenta L.) Yellow-poplar Yellow-poplar (Liriodendron tulipifera L.) Model Calibration Preliminary model results from the advance reproduction survey in the mature stands using the initial rankings (predicted) were compared to the upper canopy species composition of the regenerating clear-cut stands (mea- sured) in the model calibration data set. The primary model output evaluated was the predicted species composition expressed as a proportion of total stand composition. That allowed predictions from REGEN to be uniformly compared to the measured species composition of regenerating clear-cut stands which varied somewhat in age, density, and development. Based on these intermediate results, competitive rankings for each species in each RKB were adjusted in an iterative process to achieve the best fit. Sprouting prob- abilities and other parameters were also amended according to the results of preliminary model runs. The final rankings developed for the four site classes are presented in Tables 4 and 5. Statistical Analysis On both the model calibration and validation data sets, tests for normality indicated a nonnormal distribution. Further analyses were conducted using D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 T A B L E 2 M e an C o m p o si ti o n o f th e M at u re Sa m p le St an d s in th e C e n tr al A p p al ac h ia n s M o d e l ca li b ra ti o n d at a se t M o d e l v al id at io n d at a se t O v e ra ll (n = 4 8 ) Su b x e ri c (n = 1 0 ) Su b m e si c (n = 3 8 ) O v e ra ll (n = 1 7 ) Su b x e ri c (n = 3 ) Su b m e si c (n = 1 4 ) Sp e ci e s g ro u p B A (S E ±) (m 2 / h a) T P H (S E ±) (s te m s/ h a) B A (S E ±) (m 2 / h a) T P H (S E ±) (s te m s/ h a) B A (S E ±) (m 2 / h a) T P H (S E ±) (s te m s/ h a) B A (S E ±) (m 2 / h a) T P H (S E ±) (s te m s/ h a) B A (S E ±) (m 2 / h a) T P H (S E ±) (s te m s/ h a) B A (S E ±) (m 2 / h a) T P H (S E ±) (s te m s/ h a) B la ck ch e rr y 1 .8 (0 .7 ) 3 0 (1 0 ) 0 (0 .0 ) 0 (0 ) 2 .3 (0 .7 ) 3 7 (1 2 ) 0 .5 (0 .2 ) 1 7 (5 ) 0 .0 (0 .0 ) 0 (0 ) 0 .7 (0 .2 ) 2 0 (5 ) C o n if e rs 0 .9 (0 .2 ) 4 0 (1 5 ) 0 .7 (0 .5 ) 2 5 (2 2 ) 1 .1 (0 .5 ) 4 2 (1 7 ) 1 .1 (0 .7 ) 2 (2 ) 0 .9 (0 .5 ) 2 0 (2 0 ) 1 .1 (0 .7 ) 0 (0 ) M ap le s 6 .0 (0 .7 ) 3 2 1 (2 5 ) 3 .0 (0 .7 ) 3 6 8 (7 2 ) 6 .9 (0 .7 ) 3 0 9 (2 5 ) 6 .4 (0 .9 ) 2 5 4 (2 2 ) 3 .7 (1 .1 ) 3 2 9 (4 7 ) 7 .1 (0 .9 ) 2 3 7 (2 5 ) M id st o ry 2 .3 (0 .2 ) 1 7 5 (2 0 ) 1 .4 (0 .5 ) 2 1 0 (5 7 ) 2 .5 (0 .5 ) 1 6 5 (2 0 ) 2 .1 (0 .5 ) 1 6 3 (3 2 ) 1 .8 (0 .7 ) 3 0 9 (1 3 1 ) 2 .3 (0 .7 ) 1 3 1 (2 5 ) O ak s 1 0 .8 (1 .1 ) 1 6 5 (2 2 ) 1 6 .8 (1 .4 ) 3 1 6 (4 0 ) 9 .4 (1 .4 ) 1 2 6 (2 0 ) 1 0 .6 (1 .6 ) 1 7 3 (4 0 ) 2 0 .0 (1 .1 ) 4 5 5 (7 9 ) 8 .5 (1 .6 ) 1 1 4 (2 5 ) O th e r o v e rs to ry 2 .3 (0 .2 ) 1 0 6 (1 7 ) 0 .7 (0 .2 ) 5 9 (3 2 ) 2 .8 (0 .5 ) 1 1 9 (2 0 ) 3 .9 (0 .9 ) 7 4 (1 5 ) 0 .7 (0 .5 ) 1 7 (1 0 ) 4 .4 (1 .1 ) 8 6 (1 7 ) P io n e e r 0 .7 (0 .2 ) 3 0 (7 ) 0 .2 (0 .2 ) 1 0 (7 ) 0 .9 (0 .2 ) 3 5 (1 0 ) 0 .9 (0 .2 ) 3 0 (1 0 ) 0 .0 (0 .0 ) 0 (0 ) 1 .1 (0 .2 ) 3 5 (1 2 ) Y e ll o w - p o p la r 2 .3 (0 .5 ) 5 2 (1 2 ) 0 .0 (0 .0 ) 2 (2 ) 2 .8 (0 .5 ) 6 4 (1 5 ) 2 .5 (0 .9 ) 3 2 (1 2 ) 1 .1 (0 .9 ) 2 0 (2 0 ) 2 .8 (1 .1 ) 3 7 (1 5 ) T o ta ls 2 7 .1 9 1 9 2 2 .8 9 9 0 2 8 .7 8 9 7 2 8 .0 7 4 5 2 8 .2 1 1 4 9 2 8 .0 6 6 0 N o te . B A = b as al ar e a; T P H = tr e ss p e r h e ct ar e . 799 D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 T A B L E 3 M e an C o m p o si ti o n o f th e R e g e n e ra ti n g C le ar -c u t Sa m p le St an d s in th e C e n tr al A p p al ac h ia n s M o d e l ca li b ra ti o n d at a se t M o d e l v al id at io n d at as e t O v e ra ll (n = 4 8 ) Su b x e ri c (n = 1 0 ) Su b m e si c (n = 3 8 ) O v e ra ll (n = 1 7 ) Su b x e ri c (n = 3 ) Su b m e si c (n = 1 4 ) Sp e ci e s g ro u p Lo w e r ca n o p y U p p e r ca n o p y Lo w e r ca n o p y U p p e r ca n o p y Lo w e r ca n o p y U p p e r ca n o p y Lo w e r ca n o p y U p p e r ca n o p y Lo w e r ca n o p y U p p e r ca n o p y Lo w e r ca n o p y U p p e r ca n o p y (s te m s/ h a) B la ck ch e rr y 9 4 4 (5 1 1 ) 6 1 8 (2 3 7 ) 8 6 (6 4 ) 9 9 (8 4 ) 1 ,1 6 9 (6 4 2 ) 7 5 3 (2 9 4 ) 3 7 (1 5 ) 9 6 (2 7 ) 0 (0 ) 0 (0 ) 4 4 (1 7 ) 1 1 9 (3 2 ) C o n if e rs 1 5 (7 ) 1 2 (7 ) 5 7 (3 0 ) 5 (5 ) 2 (2 ) 1 5 (1 0 ) 4 7 (2 7 ) 5 7 (5 2 ) 1 0 4 (1 0 4 ) 2 8 9 (2 8 9 ) 3 5 (2 7 ) 7 (5 ) M ap le s 2 ,4 9 2 (3 4 8 ) 1 ,0 7 9 (2 2 2 ) 1 ,3 7 1 (2 7 2 ) 1 ,1 2 6 (3 6 6 ) 2 ,7 8 9 (4 2 2 ) 1 0 6 7 (2 6 4 ) 1 ,8 4 0 (2 2 5 ) 5 4 3 (9 7 ) 8 0 3 (1 9 8 ) 4 7 4 (1 9 5 ) 2 ,0 6 5 (2 3 0 ) 5 5 8 (1 0 4 ) M id st o ry 2 ,1 9 3 (3 6 8 ) 8 1 8 (2 7 4 ) 1 ,3 5 1 (4 1 5 ) 4 2 2 (2 1 0 ) 2 ,4 1 6 (4 4 7 ) 9 2 1 (3 4 3 ) 1 ,2 1 0 (2 4 0 ) 2 7 4 (6 4 ) 1 ,2 0 3 (6 0 3 ) 2 0 7 (1 7 5 ) 1 ,1 0 (2 6 9 ) 2 8 9 (7 2 ) O ak s 1 ,6 4 7 (5 5 8 ) 6 2 2 (1 5 1 ) 2 ,1 6 9 (1 2 4 0 ) 1 ,2 7 5 (5 0 9 ) 1 ,5 0 9 (6 3 0 ) 4 5 2 (1 2 8 ) 5 9 8 (1 5 3 ) 4 4 0 (1 0 6 ) 1 ,0 5 0 (5 7 3 ) 7 8 3 (3 6 1 ) 5 0 1 (1 4 3 ) 3 6 6 (1 0 1 ) O th e r O v e rs to ry 7 6 1 (1 8 8 ) 4 5 7 (8 4 ) 2 3 7 (8 9 ) 3 2 6 (9 9 ) 8 9 9 (2 3 0 ) 4 9 2 (1 0 4 ) 4 7 2 (1 2 8 ) 2 0 5 (5 4 ) 3 5 1 (1 6 1 ) 1 6 5 (8 2 ) 4 9 6 (1 5 3 ) 2 1 2 (6 4 ) P io n e e r 9 8 6 (1 6 8 ) 1 ,0 3 0 (2 5 7 ) 3 4 8 (3 0 9 ) 6 5 2 (5 6 8 ) 1 ,1 5 3 (1 8 8 ) 1 1 2 9 (2 8 9 ) 5 0 9 (1 4 1 ) 5 4 3 (1 2 1 ) 0 (0 ) 0 (0 ) 6 1 8 (1 5 6 ) 6 5 9 (1 2 6 ) Y e ll o w - p o p la r 9 9 5 (1 6 3 ) 5 8 0 (9 1 ) 4 6 9 (1 9 8 ) 4 7 4 (1 9 0 ) 1 ,1 3 1 (1 9 5 ) 6 0 8 (1 0 4 ) 5 7 8 (2 9 1 ) 3 6 1 (1 4 8 ) 1 ,6 4 7 (1 6 4 7 ) 7 4 1 (7 4 1 ) 3 5 1 (1 1 9 ) 2 7 9 (1 0 6 ) T o ta l 1 0 ,0 3 3 5 ,2 1 6 6 ,0 8 8 4 ,3 7 9 1 1 ,0 6 8 5 4 3 6 5 ,2 9 1 2 ,5 1 9 5 ,1 5 8 2 6 5 9 5 ,3 2 0 2 ,4 8 9 N o te . Lo w e r ca n o p y = su p p re ss e d an d in te rm e d ia te st e m s; u p p e r ca n o p y = d o m in an t an d co d o m in an t st e m s. 800 D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 T A B L E 4 C o m p e ti ti v e R an k in g s fo r C o m m o n Sp e ci e s F o ll o w in g C le ar -c u t o n X e ri c an d Su b x e ri c Si te s in th e C e n tr al A p p al ac h ia n s R an k X e ri c Su b x e ri c 1 C h e st n u t o ak Sp , sc ar le t o ak Sp , so u rw o o d Sp , V ir g in ia p in e L B la ck lo cu st Sp , ch e st n u t o ak Sp , e . w h it e p in e L, re d m ap le Sp , sc ar le t o ak Sp , sw e e t b ir ch Sp , y e ll o w -p o p la r Sp 2 B la ck o ak Sp , ch e st n u t o ak L, e . w h it e p in e L, p it ch p in e L, re d m ap le Sp , sh o rt le af p in e L, T ab le M t. p in e L B la ck ch e rr y Sp , b la ck lo cu st L, b la ck o ak Sp , cu cu m b e rt re e Sp , e . w h it e p in e M , fi re ch e rr y Sp , sw e e t b ir ch L, w h it e o ak Sp , y e ll o w -p o p la r L 3 B la ck lo cu st Sp , b la ck o ak L, b la ck g u m Sp , e . w h it e p in e M , h ic k o ry Sp , n . re d o ak Sp , sa ss af ra s Sp , sc ar le t o ak L, so u rw o o d L, T ab le M t. p in e M , V ir g in ia p in e S, w h it e o ak Sp , B la ck lo cu st M , ch e st n u t o ak L, cu cu m b e rt re e L, fi re ch e rr y L, h ic k o ry Sp , n . re d o ak Sp , sc ar le t o ak L, sw e e t b ir ch M , y e ll o w -p o p la r M 4 B la ck lo cu st L, ch e st n u t o ak M , e . w h it e p in e S, p it ch p in e S, re d m ap le L, sc ar le t o ak M , sh o rt le af p in e S, so u rw o o d M , T ab le M t. p in e S, V ir g in ia p in e G , w h it e o ak L B la ck ch e rr y L, b la ck lo cu st S, b la ck o ak L, e . w h it e p in e S, fi re ch e rr y M , re d m ap le L, so u rw o o d Sp , sw e e t b ir ch S, V ir g in ia p in e L, y e ll o w -p o p la r S 5 B la ck lo cu st M , b la ck o ak M , b la ck g u m L, ch e st n u t o ak S, e . w h it e p in e G , h ic k o ry L, ir o n w o o d Sp , n . re d o ak L, p it ch p in e G , re d m ap le M , sa ss af ra s L, sc ar le t o ak S, se rv ic e b e rr y Sp , sh o rt le af p in e G , so u rw o o d S, T ab le M t. p in e G , w h it e o ak M , A sh Sp , b la ck ch e rr y M , b la ck lo cu st G , ch e st n u t o ak M , d o g w o o d Sp , e . w h it e p in e G , fi re ch e rr y S, ir o n w o o d Sp , n . re d o ak L, p it ch p in e L, sc ar le t o ak M , sh o rt le af p in e L, sw e e t b ir ch G , T ab le M t. p in e L, w h it e o ak L, y e ll o w -p o p la r G 6 B la ck ch e rr y Sp , b la ck lo cu st S, b la ck o ak S, b la ck g u m M , ch e st n u t o ak G , d o g w o o d Sp , h ic k o ry M , n . re d o ak M , re d m ap le S, sa ss af ra s M , sc ar le t o ak G , so u rw o o d G , w h it e o ak S, y e ll o w -p o p la r Sp B as sw o o d Sp , b e e ch Sp , b la ck ch e rr y S, b la ck o ak M , cu cu m b e rt re e M , fi re ch e rr y G , F ra se r m ag n o li a Sp , h ic k o ry L, ir o n w o o d L, sa ss af ra s Sp , se rv ic e b e rr y Sp , so u rw o o d L, st ri p e d m ap le Sp , V ir g in ia p in e M 7 A sh Sp , b e e ch Sp , b la ck ch e rr y L, b la ck o ak G , b la ck g u m S, d o g w o o d L, fi re ch e rr y Sp , h ic k o ry S, ir o n w o o d L, n . re d o ak S, re d m ap le G , sa ss af ra s S, se rv ic e b e rr y L, sw e e t b ir ch Sp , w h it e o ak G A sh L, b as sw o o d L, b e e ch L, b la ck g u m Sp , b u ck e y e Sp , d o g w o o d L, F ra se r m ag n o li a L, h e m lo ck L, ir o n w o o d S, p it ch p in e M , sa ss af ra s L, se rv ic e b e rr y L, sh o rt le af p in e M , st ri p e d m ap le L, su g ar m ap le Sp , T ab le M t. p in e M 8 A sh L, b as sw o o d Sp , b e e ch L, b la ck ch e rr y M , b la ck lo cu st G , b la ck g u m G , b u ck e y e SP , cu cu m b e rt re e Sp , d o g w o o d M , fi re ch e rr y L, F ra se r m ag n o li a Sp , h e m lo ck L, h ic k o ry G , ir o n w o o d M , n . re d o ak G , sa ss af ra s G , se rv ic e b e rr y M , st ri p e d m ap le Sp , su g ar m ap le Sp , sw e e t b ir ch L, y e ll o w b ir ch Sp , y e ll o w -p o p la r L A sh Sp , b as sw o o d Sp , b la ck ch e rr y G , b la ck o ak S, b la ck g u m L, b u ck e y e L, ch e st n u t o ak S, cu cu m b e rt re e S, d o g w o o d M , F ra se r m ag n o li a M , h e m lo ck M , h ic k o ry M , ir o n w o o d S, n . re d o ak M , re d m ap le M , sc ar le t o ak S, se rv ic e b e rr y M , so u rw o o d M , st ri p e d m ap le M , su g ar m ap le L, V ir g in ia p in e S, w h it e o ak M , y e ll o w -b ir ch Sp N o te . Sp = st u m p -s p ro u t; L = la rg e se e d li n g ; M = m e d iu m se e d li n g ; S = sm al l se e d li n g ; G = g e rm in an t. 801 D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 T A B L E 5 C o m p e ti ti v e R an k in g s fo r C o m m o n Sp e ci e s F o ll o w in g C le ar -c u t o n Su b m e si c an d M e si c Si te s in th e C e n tr al A p p al ac h ia n s R an k Su b m e si c M e si c 1 B as sw o o d Sp , b la ck ch e rr y Sp , b la ck lo cu st Sp , fi re ch e rr y Sp , sw e e t b ir ch Sp , y e ll o w -p o p la r Sp B as sw o o d Sp , b la ck ch e rr y Sp , fi re ch e rr y Sp , sw e e t b ir ch Sp , y e ll o w -p o p la r Sp 2 B as sw o o d L, b la ck ch e rr y L, b la ck lo cu st L, e . w h it e p in e L, fi re ch e rr y L, n . re d o ak Sp , sw e e t b ir ch L, y e ll o w -p o p la r L A sh Sp , b la ck ch e rr y L, b la ck lo cu st Sp , cu cu m b e rt re e Sp , e . w h it e p in e L, fi re ch e rr y L, su g ar m ap le Sp , sw e e t b ir ch L, y e ll o w b ir ch Sp , y e ll o w -p o p la r L 3 B la ck lo cu st M , b la ck o ak Sp , e . w h it e p in e M , fi re ch e rr y M , re d m ap le Sp , y e ll o w -p o p la r M A sh L, b as sw o o d L, b la ck ch e rr y M , fi re ch e rr y M , re d m ap le Sp , sw e e t b ir ch M , y e ll o w -p o p la r M 4 A sh Sp , b as sw o o d M , b la ck ch e rr y M , b la ck lo cu st S, ch e st n u t o ak Sp , e . w h it e p in e S, fi re ch e rr y S, sc ar le t o ak Sp , sw e e t b ir ch M , w h it e o ak Sp , y e ll o w -p o p la r S B e e ch Sp , b la ck ch e rr y S, b la ck lo cu st L, b la ck o ak Sp , b u ck e y e Sp , cu cu m b e rt re e L, d o g w o o d Sp , e . w h it e p in e M , fi re ch e rr y S, F ra se r m ag n o li a Sp , h e m lo ck L, h ic k o ry Sp , n . re d o ak Sp , st ri p e d m ap le Sp , su g ar m ap le L, sw e e t b ir ch S, y e ll o w b ir ch L, y e ll o w -p o p la r S, as h M , b as sw o o d M , b e e ch L 5 A sh L, b la ck lo cu st G , cu cu m b e rt re e Sp , e . w h it e p in e G , fi re ch e rr y G , h ic k o ry Sp , n . re d o ak L, st ri p e d m ap le Sp , sw e e t b ir ch S, y e ll o w b ir ch Sp , y e ll o w -p o p la r G B la ck ch e rr y G , b la ck lo cu st M , b u ck e y e L, ch e st n u t o ak Sp , cu cu m b e rt re e M , d o g w o o d L, fi re ch e rr y G , re d m ap le L, sc ar le t o ak Sp , su g ar m ap le M , sw e e t b ir ch G , w h it e o ak Sp , y e ll o w b ir ch M , y e ll o w -p o p la r G 6 A sh M , b as sw o o d S, b e e ch Sp , b la ck ch e rr y S, b la ck o ak L, b la ck g u m Sp , F ra se r m ag n o li a Sp , re d m ap le L, sa ss af ra s Sp , sc ar le t o ak L, se rv ic e b e rr y Sp , so u rw o o d Sp , sw e e t b ir ch G , w h it e o ak L, y e ll o w b ir ch L B e e ch M , b la ck lo cu st S, b la ck o ak L, b u ck e y e M , e . w h it e p in e S, F ra se r m ag n o li a L, h e m lo ck M , h ic k o ry L, n . re d o ak L, se rv ic e b e rr y Sp , so u rw o o d Sp , st ri p e d m ap le L, su g ar m ap le S, y e ll o w b ir ch S 7 A sh S, b as sw o o d G , ch e st n u t o ak L, cu cu m b e rt re e L, F ra se r m ag n o li a L, ir o n w o o d Sp , se rv ic e b e rr y L, su g ar m ap le Sp , y e ll o w b ir ch M B la ck lo cu st G , b la ck g u m Sp , ch e st n u t o ak L, cu cu m b e rt re e S, d o g w o o d M , e . w h it e p in e G , F ra se r m ag n o li a M , h ic k o ry M , ir o n w o o d Sp , re d m ap le M , sa ss af ra s Sp , sc ar le t o ak L, se rv ic e b e rr y L, st ri p e d m ap le M , w h it e o ak L, y e ll o w b ir ch G 8 A sh G , b e e ch L, b la ck ch e rr y G , b la ck o ak M , b la ck g u m L, b u ck e y e Sp , cu cu m b e rt re e M , d o g w o o d Sp , h e m lo ck L, h ic k o ry L, ir o n w o o d L, n . re d o ak M , p it ch p in e L, re d m ap le M , sa ss af ra s L, sc ar le t o ak M , se rv ic e b e rr y M , sh o rt le af p in e L, so u rw o o d L, st ri p e d m ap le L, su g ar m ap le L, T ab le M t. p in e L, V ir g in ia p in e L, w h it e o ak M , y e ll o w b ir ch S A sh S, b as sw o o d S, b e e ch S, b la ck o ak M , b la ck g u m L, b u ck e y e S, cu cu m b e rt re e G , d o g w o o d S, F ra se r m ag n o li a S, h e m lo ck S, h ic k o ry S, ir o n w o o d L, n . re d o ak M , p it ch p in e L, re d m ap le S, sa ss af ra s L, se rv ic e b e rr y M , sh o rt le af p in e L, so u rw o o d L, st ri p e d m ap le S, su g ar m ap le G , T ab le M t. p in e L, V ir g in ia p in e L, w h it e o ak M N o te . Sp = st u m p -s p ro u t; L = la rg e se e d li n g ; M = m e d iu m se e d li n g ; S = sm al l se e d li n g ; G = g e rm in an t. 802 D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 803 nonparametric tests. These analyses were conducted in SAS ® v. 9.2 software using the UNIVARIATE procedure. REGEN predicts the future upper canopy species composition; therefore, only the upper canopy measurements from the regenerating clear-cut stands were included in these analyses. To assess the overall performance of the model, values of predicted species compo- sition from the REGEN model (which were based on advance reproduction inventories from mature stands) were compared to values of measured species composition (based on upper canopy measurements from regener- ating clear-cut stand inventories) for each paired stand in each data set. For example, if the model predicted that black cherry will make up 25% of the future upper canopy stems following harvest based on the advance repro- duction inventory, but only 20% of the upper canopy stems measured in the regenerating clear-cut stand were black cherry, then the difference (or model error) at that sample site was 5 percentage points. The Wilcoxon Signed Rank test, a nonparametric alternative to a paired t test, was conducted to test for differences between measured and predicted sample distributions for each species group across all paired stands (Ott & Longnecker, 2001). Simple linear regression was employed on each data set to evaluate the ability of the model to explain the variability in future species composition. To conduct this analysis, paired values of species composition (predicted vs. measured) were analyzed without regard to their associated stands using the REG procedure in SASTM v. 9.2 statistical software. For example, in the model calibration data set, there were paired values of species composition for eight different species groups in 48 paired stands. Because no regard was given to the individual stand in this analysis, a total of 384 data points (8 species groups × 48 paired stands = 384 data points) were used to evaluate the model’s overall ability to explain future species composition. Assessment of performance was based on the coefficient of determination from the regression analysis. To evaluate the model performance for each species group individually, a regression analysis was conducted for the mea- sured and predicted proportions of each species group separately across all paired stands in each data set. Regression analysis was also conducted on the paired values of species composition for individual stands. The range and standard deviation of the model error (predicted– measured) were calculated for each species group using data from all paired stands in each data set to assess the variation in model predictions. One standard deviation of model error was both added and subtracted from the mean error for each species group to evaluate the spread and frequency of model error. RESULTS Model Calibration The rankings and results presented in this article represent the best fit achieved for the initial model calibration data set after several iterative trials D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 804 L. A. Vickers et al. with the revised RKBs. Using the final RKBs, REGEN appeared to reasonably describe the species composition of regenerating stands following clear- cutting in the 48 paired stands of the model calibration data set that was used to calibrate the RKBs (Fig. 2A). Across all 48 paired stands, mean species composition predicted by REGEN was within 4 percentage points of the measured species composition for all eight species groups used in this Black Cherry 9% Conifers 1% Maples 23% Midstory 8% Oaks 12% Other Overstory 8% Pioneer 20% Yellow- poplar, 19% PREDICTED Black Cherry 8% Conifers 1% Maples 20% Midstory 10% Oaks 12% Other Overstory 12% Pioneer 20% Yellow- poplar 17% MEASUREDA Black Cherry 5% Conifers 2% Maples 22% Midstory 9% Oaks 13% Other Overstory 8% Pioneer 24% Yellow- poplar 17% PREDICTED Black Cherry 4% Conifers 2% Maples 22% Midstory 11% Oaks 17%Other Overstory 9% Pioneer 22% Yellow- poplar 13% MEASUREDB FIGURE 2 Comparison of mean values for measured and predicted species composition in regenerating clear-cut stands from paired stands in the Central Appalachians. Graph A depicts the model calibration dataset, Graph B depicts the model validation data set. D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 805 study. Mean differences between measured and predicted samples indicated that black cherry, maples, and yellow-poplar species groups were generally slightly overpredicted. Midstory, and other overstory species groups were generally underpredicted. Tests for differences in sample distributions indi- cated that among species groups, only the other overstory and yellow-poplar measured distributions were statistically different than their predicted dis- tributions at an α level of .1 (Table 6). Across all 48 paired stands used to build the RKBs, one standard deviation of model error for any species group was no more than about ±20 percentage points from the measured value (Figure 3A). The error for the maples, midstory, pioneer, and yellow- poplar species groups tended to be greater as each had a standard deviation between 15 and 20 percentage points for model error. The predictions on the 10 subxeric sites were, on average, within about 6 percentage points of the measured values, with the greatest discrepancies observed for the maples, other overstory, and yellow-poplar species groups (Figure 4A) The predictions for the 38 submesic sites were within about 4 percentage points of the measured values on average (Figure 4B). The range of model error from the subxeric RKB appeared to be smaller than the submesic on average, with the exception of yellow-poplar on subxeric sites, which tended to be underpredicted to a greater extent (Figure 5). However, it is possible that this reduced error spread may be due to fewer subxeric sample sites rather than superior performance. On average, one standard deviation of the model error for a species group was not more than ±20 percentage points for both RKBs, except for subxeric yellow-poplar which was typically within about ±24 percentage points. The overall regression analysis for the 48 paired stands used to build the RKBs indicated a significant linear relationship between predicted and TABLE 6 Comparison of Mean Measured and Mean Predicted Species Composition of All Sample Stands in the Central Appalachians Model calibration data set Model validation data set Species group Measured Predicted p value Measured Predicted p value . . . . . . . .(%) . . . . . . . . . . . . . . . .(%) . . . . . . . . Black cherry 8 9 .3797 4 5 .2661 Conifers 1 1 .7819 2 2 .6250 Maples 20 23 .4060 22 22 1.0000 Midstory 10 8 .9419 11 9 .6777 Oaks 12 12 .8012 17 13 .4037 Other overstory 12 8 .0397 9 8 .7119 Pioneer 20 20 .1460 22 24 .8209 Yellow-poplar 17 19 .0466 13 17 .0413 Total 100 100 100 100 Note. The p values < .1 are considered significant differences between measured and predicted samples as calculated by the Wilcoxon Signed Rank test. D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 806 L. A. Vickers et al. –100% –80% –60% –40% –20% 0% 20% 40% 60% 80% 100% Black Cherry Conifers Maples Midstory Oaks Other Overstory Pioneer Yellow- poplar M O D E L E R R O R B Mean ± Standard Deviation Range –100% –80% –60% –40% –20% 0% 20% 40% 60% 80% 100% Black Cherry Conifers Maples Midstory Oaks Other Overstory Pioneer Yellow- poplar M O D E L E R R O R A | FIGURE 3 Mean model error spread for each species group from paired stands in the Central Appalachians. Model error was calculated as predicted–measured. Data points = mean species group model error, boxes = mean species group model error ± one stan- dard deviation, vertical lines = species group model error range. Graph A depicts the model calibration data set, Graph B depicts the model validation data set. measured species composition (Figure 6A). The coefficient of determina- tion indicated that the REGEN predictions explained 32% (R2 = .32) of the variation in species composition in regenerating clear-cut stands across both site classes. The subxeric RKB explained 42% (R2 = .42) and the submesic RKB explained 29% (R2 = .29) of the variation in species composition in regenerating clear-cut stands on those sites, respectively. In this data set, 86% of the individual predictions were within ±20 percentage points of their D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 807 Black Cherry 2% Conifers 2% Maples 27% Midstory 7%Oaks 24% Other Overstory 16% Pioneer 8% Yellow- poplar 14% MEASURED Black Cherry 1% Conifers 2% Maples 33% Midstory 11% Oaks 24% Other Overstory 9% Pioneer 12% Yellow- poplar 8% PREDICTEDA Black Cherry 10% Conifers 0% Maples 18% Midstory 11% Oaks 9% Other Overstory 11% Pioneer 23% Yellow- poplar 18% MEASURED Black Cherry 11% Conifers 0% Maples 20% Midstory 7% Oaks 9% Other Overstory 8% Pioneer 23% Yellow- poplar 22% PREDICTEDB FIGURE 4 Comparison of mean values for measured and predicted species composition by site class in regenerating clear-cut stands from paired stands in the Central Appalachians used for model calibration. Graph A depicts the subxeric stands, Graph B depicts the submesic stands. paired measured values, 66% of the predictions were within 10 percentage points, and about half were within 5 percentage points. The analysis of the model residuals, which plotted the model error (predicted–measured) against the measured values, suggested that the model consistently underpredicted future composition when a measured group was greater than about 30% of total stand composition (Figure 7A). In both the regression and residual plots, the subxeric and submesic data points appeared to be randomly distributed. In the model calibration data set, there were 46 instances where a species group comprised at least 35% of stand composition. Early successional species reached this level in 23 of the 46 instances (black cherry 3, pioneer 11, yellow-poplar 9). Maples were the next most frequent group, attaining a composition of at least D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 808 L. A. Vickers et al. Black Cherry Conifers Maples Midstory Oaks Other Overstory Pioneer Yellow- poplar –100% –80% –60% –40% –20% 0% 20% 40% 60% 80% 100% M O D E L E R R O R B –100% –80% –60% –40% –20% 0% 20% 40% 60% 80% 100% Black Cherry Conifers Maples Midstory Oaks Other Overstory Pioneer Yellow- poplar M O D E L E R R O R A Mean ± Standard Deviation Range | FIGURE 5 Mean model error spread for each species group by site class from paired stands in the Central Appalachians used for model calibration. Model error was calculated as predicted – measured. Data points = mean species group model error, boxes = mean species group model error ± one standard deviation, vertical lines = species group model error range. Graph A depicts the subxeric stands, Graph B depicts the submesic stands. 35% in 11 instances. The remaining 12 instances were by the midstory (3), oaks (4), and other overstory (5) groups. Although this level of stand occupancy was somewhat encouraged by grouping species, instances where individual species attained this level of composition were nearly as frequent (31 instances). Forty of the total 48 regenerating clear-cut stands had at least one group that occupied 35% or more of total composition, and 6 of the D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 809 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% P R E D IC T E D V A L U E MEASURED VALUE Submesic Subxeric A p < .0001 R2 = .32 n = 384 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% P R E D IC T E D V A L U E MEASURED VALUE p < .0001 R2 = .36 n = 136 B FIGURE 6 Regression analysis comparing measured and predicted values for each species group from paired stands in the Central Appalachians. Y-axis = predicted values, X-axis = measured values. Graph A depicts the model calibration data set, Graph B depicts the model validation data set. 40 stands had two different groups reaching this threshold. In all but one of the stands where two different species groups made up at least 35% of the species composition maples was one of those groups. Individual regression analysis for each species group indicated signifi- cant regressions for all species groups except midstory at an α level of .1 (Figure 8). The model explained the greatest amount of variation for black cherry (R2 = .45), and over 25% of the variability in individual species composition in the regenerating clear-cut stands for three other species D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 810 L. A. Vickers et al. –100% –80% –60% –40% –20% 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% M O D E L E R R O R MEASURED VALUE Submesic SubxericA 0% 20% 40% 60% 80% 100% –100% –80% –60% –40% –20% 0% 20% 40% 60% 80% 100% M O D E L E R R O R MEASURED VALUE B FIGURE 7 Residual analysis for each species group from paired stands in the Central Appalachians. Y-axis = model error (predicted–measured), X-axis = measured values. Graph A depicts the model calibration data set, Graph B depicts the model validation data set. groups: 39% (R2 = .39) for oaks, 30% (R2 = .30) for other overstory, and 28% (R2 = .28) for the pioneer species group. Along with the overall stand summary of species composition output, REGEN also reports the original source size (sprout, large, medium, small, germinant) of each winning stem at each plot. The ratio for winning stems of sprout origin in the predicted samples was compared to the sprout ratio in measured regenerating clear-cut stands across all 48 paired stands used to build the RKBs. Tests indicated that the distribution of the proportion of predicted stems of sprout origin were significantly different from measured distributions for the black cherry, midstory, oaks, and pioneer species groups at an α level of .1 (Figure 9A). D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 811 P R E D IC T E D V A L U E 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Black Cherry R2 = .45 p < .0001 0% 20% 40% 60% 80% 100% Conifers R2 = .19 p = .0022 0% 20% 40% 60% 80% 100%0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Maples R2 = .21 p = .0012 0% 20% 40% 60% 80% 100% Midstory R2 = .03 p = .2290 0% 20% 40% 60% 80% 100%0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Pioneer R2 = .28 p = .0001 0% 20% 40% 60% 80% 100% Yellow-poplar R2 = .11 p = .0210 0% 20% 40% 60% 80% 100%0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Oaks R2 = .39 p < .0001 0% 20% 40% 60% 80% 100% Other Overstory R2 = .30 p < .0001 FIGURE 8 Regression analysis comparing measured and predicted values for each species group individually from paired stands in the Central Appalachians used for model calibration. Y-axis = predicted values, X-axis = measured values. Model Validation When the final RKBs in REGEN were applied to the 17 paired stands col- lected to serve as an independent validation data set, the performance of the D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 812 L. A. Vickers et al. Black Cherry Conifers Maples Midstory Oaks Other Overstory Pioneer Yellow- poplar 0% 20% 40% 60% 80% 100% P R O P O R T IO N O F S P R O U T S p = .9219 p = .0026 p = .9780 p = .0580 p = .6355 p = .8501 p = .5771 B 0% 20% 40% 60% 80% 100% Black Cherry Conifers Maples Midstory Oaks Other Overstory Pioneer Yellow- poplar P R O P O R T IO N O F S P R O U T S p = 0.0317 p = 0.6317 p = .0030 p = 0.0791 p = .2019 p = 0.0243 p = 0.3799 A Measured Predicted FIGURE 9 Comparison of measured and predicted stump-sprout proportion of upper canopy stems for each species group individually from regenerating clear-cut stands in the Central Appalachians. Y-axis = % of dominant and codominant stems of sprout origin. The p values < .1 indicate significant differences between measured and predicted samples as calculated by the Wilcoxon Signed Rank test. Graph A depicts the model calibration data set, Graph B depicts the model validation data set. model in general was on par with its ability to explain the data used to cali- brate the RKBs. Across all 17 paired stands in the validation data set, mean species composition predicted by the REGEN model was within 4 percentage points of the measured species composition for all eight species groups used in this study (Figure 2B). Mean differences between measured and predicted samples indicated that black cherry, pioneer, and yellow-poplar species groups were generally overpredicted on average. Midstory, oaks, and other overstory species groups were generally underpredicted on average. Tests for differences in sample distributions indicated that, among species groups, only the yellow-poplar measured distribution was statistically different than D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 813 its predicted distribution at an α level of .1 (Table 6). Across all 17 paired stands in the validation data set, one standard deviation of the model error for any species group was not more than about ±20 percentage points (Figure 3B). The overall regression of predicted versus measured species composi- tion for the paired stands in the validation data set was highly significant (Figure 6B). The coefficient of determination indicated that REGEN predic- tions explained 36% (R2 = 0.36) of the variation in species composition in regenerating clear-cut stands across both site classes. About 89% of the pre- dictions were within ± 20 percentage points of their paired measured values in this data set. About 71% of the predictions were within 10 percentage points, and about half were within 5 percentage points. The analysis of the model residuals for the validation dataset suggested that the model consistently underpredicted future species composition when a measured species group reached greater than about 30% of total stand composition (Figure 7B). In the validation data set, a species group reached this level of site occupancy in 16 instances. In eight of these instances, early successional species groups (pioneer and yellow-poplar) were responsible. The maples and oaks species groups each reached this level three times, and the other overstory and midstory groups each reached this level once. Fourteen of the total 17 stands in the validation data set contained at least one species group that comprised at least 35% of total stand composition. Two stands had two different groups to reach this level. REGEN slightly overpredicted future species composition when a measured species group made up less than about 5% of total stand composition. Individual regression analyses for each species group in the validation data set indicated significant regressions for all groups except conifers and midstory at an α level of .1 (Figure 10). The model explained the greatest amount of variation in individual species composition in the regenerating clear-cut stands in the validation data set for oaks (R2 = .36), and over 25% for two other species groups: 34% (R2 = .34) for other overstory, and 26% (R2 = .26) for the pioneer species group. Tests on the validation data set indicated that mean predicted distributions for proportion of winning stems of sprout origin were significantly different from measured distributions of sprout origin regeneration for only the maples and oaks species groups using α = .1 (Figure 9B). Although the model performed well on average, and individual pre- dictions were often reasonable, there were considerable inconsistencies in the ability of the model to predict species composition for individual stands. Regression analyses conducted on each paired stand revealed that the amount of variation explained by the model for individual stands ranged from 0 to 98% (R2 = .00–.98) in the model calibration dataset and 0 to 93% (R2 = .00–.93) in the model validation data set. Further exploration into this condition did not reveal any obvious discrepancies in stand characteristics D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 814 L. A. Vickers et al. P R E D IC T E D V A L U E 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Black Cherry R2 = .19 p = .0811 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Conifers R2 = .06 p = .3469 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Maples R2 = .23 p = .0528 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Midstory R2 = .01 p = .6515 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Oaks R2 = .36 p = .0111 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Other Overstory R2 = .34 p = .0140 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Pioneer R2 = .26 p = .0381 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% Yellow-poplar R2 = .24 p = .0465 FIGURE 10 Validation regression analysis comparing measured and predicted values for each species group individually from paired stands in the Central Appalachians used for model validation. Y-axis = predicted values, X-axis = measured values. between those stands that were predicted with at least average results (from the overall regression analysis) and those that were below average. In the mature stands, mean basal area and composition, mean overstory density and composition, mean advance reproduction density and composition, site D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 815 quality, aspect, slope, and topographic position were all seemingly of similar distribution between stands that were predicted with at least average results and those that were below average. There were also no discernible major differences found in stand age, mean density and composition, site quality, aspect, slope, or topographic position between the regenerating clear-cut stands that were predicted with at least average results and those that were below average. DISCUSSION Rather than solely a collection of regression equations and probabilistic parameters, REGEN is more explanatory in design. It was built on a founda- tion of fundamental ecological and silvicultural concepts that acknowledge certain species possess characteristics that allow them to flourish following certain disturbances (Egler, 1954). The successes of REGEN in predicting the species composition following clear-cutting was likely attributable to its design and emphasis on local competition. Unlike many models, REGEN actually attempts to model the species competition for dominance in a stand. The often made assumption that competition can be considered a uni- form, constant variable is likely too broad in many cases. Regeneration and subsequent stand development is driven by local influences and microsite conditions (Oliver & Larson, 1996). During early stand development in Appalachian hardwoods, the composition of surrounding competition on a plot-sized scale likely influences the success or failure of individual stems to the greatest degree. This is evident in studies of precommercial crop-tree release (Trimble, 1974; Smith & Lamson, 1983; Heitzman & Nyland, 1991) that show increased levels of competition for site resources on some plots more than others. Previous research on regeneration dynamics in the Appalachians on stands of similar site quality and disturbance history (Trimble, 1973; Brashears, Fajvan, & Schuler, 2004; Gould, Steiner, Finley, & McDill, 2005), along with accepted theories of stand development (Egler, 1954; Oliver & Larson, 1996) enabled the development of rankings of species competi- tiveness. These rankings are a qualitative approximation of the ability of a species to capture growing space following disturbance and attempt to describe competitive relationships amongst species. In a similar manner, numerical points have also been used successfully to evaluate regeneration of red oaks and green ash (Fraxinus pennsylvanica Marsh.) in bottom- land hardwood ecosystems (Johnson, 1980; Johnson & Deen, 1993; Belli et al., 1999). The consideration of the advantages that more mature root systems typically provide to larger advance reproduction and stump-sprouts by incorporating competitive rankings for different size classes of repro- duction within a species in REGEN likely increased the potential accuracy D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 816 L. A. Vickers et al. of predictions of future species composition. The regeneration models developed for bottomland hardwoods (Johnson, 1980; Johnson & Deen, 1993; Belli et al., 1999), as well as those for upland oak by Sander et al. (1976, 1984), Steiner, Finley, Gould, Fei, and McDill, (2008), and Loftis (1990) all give a competitive advantage to larger stems of advance repro- duction. These stems have greater ability to capture site resources (Larsen & Johnson, 1998), and are the most competitive forms of regeneration for a given species following harvest (Sander, 1972). The competitive rankings and site quality delineations that were incor- porated into RKBs performed surprisingly well on the model calibration data set (Figures 2A, 3A) and warranted further investigation into the adoption of REGEN as a regeneration prediction model for the Central Appalachians. That performance was subsequently confirmed on an inde- pendent validation data set (Figures 2B, Figure 3B) collected a year later. As was expected, the model did not always perform as well for the vali- dation data set as it did for the model calibration data set (Figures 8, 10). Nonetheless, the results were encouraging due to a lack of statistical dif- ference for all species groups except yellow-poplar (Table 6). Difficulties predicting yellow-poplar was likely due to its successful reliance on buried seed in the forest floor as a regeneration strategy which is difficult and possibly impractical to quantify. In addition, many of the paired stands were located in the northern portions of the range of yellow-poplar where it occurs more sporadically than farther south in the Appalachians. The forests in this region tend to be transitional between central, Allegheny, and northern hardwoods (Braun, 1950; Stout, 1991) and as a result, may have made it more difficult to assess REGEN competitiveness compared to other regions. Waldrop et al. (1986) also experienced difficulty modeling early successional species, particularly yellow-poplar, on the Cumberland Plateau of Tennessee. Although the results indicate that yellow-poplar was slightly overpredicted on average (Table 6), the model error graph (Figure 3) shows potential for drastic underprediction if the ranking for yellow-poplar was reduced. Any change in rank for yellow-poplar could also adversely affect the model performance for other species as well. This type of scenario was encountered numerous times during the model calibration process. The poor results in the regression analyses for the midstory group (Figures 8, 10) was likely due to the developmental stage of the regenerat- ing stands rather than geography. As found in Table 3, the canopy position of this species group was in a state of flux as many stems were already being relegated to lower canopy positions, which made it difficult to assign competitive rankings to those species. Although species groups were used in analyses to simplify data pre- sentation, REGEN provides predictions for individual species. The model performance for individual species within groups was often similar. For D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 817 example, within the maples group, the model appeared to predict the regeneration of sugar maple slightly better than red maple. On average, predictions for sugar maple were within about 1 percentage point of the measured values while red maple was within about 2 percentage points in the model calibration data set. However, depending on ecological charac- teristics, some individual species within a group may exhibit more variation in their predictions than others. This may be the result of differences in the magnitude of regeneration across a landscape or a greater innate vari- ation in regeneration occurrence for a particular species. For example, one standard deviation of model error for sugar maple, which is more of a spe- cialist in the Appalachians and is often confined to mesic environments, was typically within about ±6 percentage points in the model calibration data set. Whereas for red maple, which is more of a generalist and occurs with varied prevalence across the Appalachian landscape, one standard devia- tion of model error was ±17 percentage points in the model calibration data set. Species groupings were composed of species that were expected to occupy similar ecological and structural niches in a mature forest where they occur. In most cases, individual species within these groupings prob- ably have similar commercial value as well. Even within a group such as the other overstory group, which contains a more miscellaneous assortment of species, all species should occupy similar positions in the canopy of a mature Appalachian hardwood stand when they occur and should col- lectively regenerate similarly to these results in spite of the variation for individual species predictions. Other regeneration prediction models for Appalachian hardwoods have performed better than the results from REGEN in this study (Loftis, 1990; Gould et al., 2006). However, these models were focused solely on oak regeneration and were developed from more extensive, long-term data sets that target only certain components of natural regeneration. While these more narrowly focused models serve as useful tools when the primary species of interest is oak, they do not provide a complete stand assess- ment of regeneration species composition. In contrast, REGEN provides a broad quantitative estimate of future species composition for a stand rather than an evaluation of regeneration adequacy for an individual species or species group. Consequently, REGEN could potentially provide greater insight into future stand dynamics and how species such as oak will com- pete with other species during the initial portion of the stem exclusion stage (Loftis, 2004). REGEN likely provides a more accurate portrayal of regeneration fol- lowing clear-cutting across a region than for a single stand. Although this phenomenon is to be expected with any effort to apply generalized pop- ulation trends to an individual, particularly given the natural variation in forest ecosystems, it is of particular interest in this line of research since managers most often prescribe silvicultural treatments to individual stands D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 818 L. A. Vickers et al. rather than across regions. Certainly poor model performance for a single species negatively impacted predictions for other species given their propor- tional relationship in some cases, but that explanation more appropriately describes an outcome of poor performance for a single species rather than an underlying cause. Part of the variability found in the results of individual stand predictions could be due to the sampling approach taken and the use of paired stands. While there was no visible reason that the assumptions of this paired stand approach were unwarranted, it is likely that some pre- harvest stand conditions in the regenerating clear-cut stands were different than those found in the mature stands. Because the selection of sites was based on the best available knowledge of the local forest manager, factors such as drought, frost, ice, browsing, or any other stochastic events that may have impacted species composition in the regenerating clear-cut stands cannot be accounted for. Considering this, it is possible that if the devel- opment of these mature stands were to be monitored following a clear-cut harvest to the onset of stem exclusion, the performance of REGEN may prove different from the results reported from this study. For this reason, proposed silvicultural treatments based on model predictions for any given stand must always be subject to the scrutiny of an experienced forester prior to implementation. While there is room for improvement in the accuracy of REGEN pre- dictions, the potential to improve these RKBs without additional data is limited. For modeling purposes, long-term trials conducted across the study region would be needed. In spite of the large amount of literature detail- ing the effects of clear-cutting on regeneration, there are few works that follow the progress of stand development to stem exclusion and include detailed characteristics of preharvest advance reproduction. This information would likely contribute a measurable improvement to the current model. However, species that are often the most successful immediately follow- ing clear-cutting have limited reliance on advance reproduction. Therefore, the greatest potential for improvements could come from modeling efforts to establish the probability for early successional species to attain upper canopy status from seed origin reproduction. An exception to this potential for improvement would likely be red maple. Red maple can function as a pioneer species in this region, but is most often regenerated from advance reproduction (Oliver & Larson 1996). Regeneration from seed sources tend to be highly stochastic events, and modeling attempts would likely be com- plex and inconsistent given the transitional tendency of the forests in much of this study region. Nonetheless, empirical results from such studies often can, and should be incorporated into the REGEN model as they become available. Insights from these types of efforts may begin to provide a more quantitative estimate of the competitive relationship amongst species and explain greater amounts of the seeming stochastic variation that is observed in regeneration dynamics. D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 819 Improvements in modeling seed origin regeneration could perhaps also remedy much of the lack of sensitivity in the RKBs for species occupying greater than 30% of future species composition and improve the accuracy of yellow-poplar and pioneer predictions. A suitable explanation for the consistent underprediction of species groups occupying greater than 30% of stand composition by REGEN was not apparent. Species groups reached this level of site occupancy about a third of the time in both data sets. Although a majority of these occurrences were by early successional species that can become established from seed sources following a major disturbance, the frequency of the maples group to reach this level discounts that trend as a comprehensive explanation. CONCLUSION Results of this study indicate that the REGEN model can be adapted to pre- dict regeneration of hardwood stands across the Appalachians. REGEN can be adapted quickly based on expert knowledge of local regeneration ecol- ogy and gradually improved as research becomes available. The detailed description of advance reproduction required by REGEN is likely more time-consuming in the field compared to other regeneration evaluations that utilize simpler metrics. Still, the REGEN methodology is straightforward and could likely be implemented into periodic stand evaluations with limited increases in labor, especially if an advance reproduction tally is already included in assessments of stands nearing rotation age. REGEN has the potential to be a useful tool in regeneration silviculture and should assist managers by identifying stands that may be candidates for preharvest manip- ulation or future ameliorative treatments necessary to meet management objectives. Further research is needed to fully realize that potential. REGEN should provide insight into regeneration trends to managers in the Central Appalachians of Virginia and West Virginia. However, any management deci- sions based on REGEN predictions should be subject to the judgment of an experienced forester. REFERENCES Beck, D. E., & Hooper, R. M. (1986). Development of a Southern Appalachian hardwood stand after clearcutting. Southern Journal of Applied Forestry, 10, 168−172. Belli, K. L., Hart, C. P., Hodges, J. D., & Stanturf, J. A. (1999). Assessment of the regeneration potential of red oaks and ash on minor bottoms of Mississippi. Southern Journal of Applied Forestry, 23(3), 133–138. Boucugnani, D. A. (2005). REGEN AGENT: A modular and generalized for- est regeneration agent (Unpublished master’s thesis). University of Georgia, Athens, GA. D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 820 L. A. Vickers et al. Brashears, M. B., Fajvan, M. A., & Schuler, T. M. (2004). An assessment of canopy stratification and tree species diversity following clearcutting in Central Appalachian hardwoods. Forest Science, 50(1), 54–64. Braun, E. L. (1950). Deciduous forests of eastern North America. Philadelphia, PA: Blackiston. Brose, P. H., Gottschalk, K. W., Horsley, S. B., Knopp, P. D., Kochenderfer, J. N., McGuinness, B. J., . . . Stout, S. L. (2008). Prescribing regeneration treatments for mixed-oak forests in the Mid-Atlantic region (General Technical Report NRS-33). Newton Square, PA: U.S. Department of Agriculture, Forest Service, Northern Research Station. Burns, R. M., & Honkala, B. H. (Eds.). (1990). Silvics of North America (Agriculture Handbook 654). Washington, DC: U.S. Department of Agriculture, Forest Service. Dey, D. C. (1991). A comprehensive Ozark regenerator (Unpublished doctoral dissertation). University of Missouri, Columbia, MO. Dey, D. C., Johnson, P. S., & Garrett, H. E. (1996). Modeling the regeneration of oak stands in the Missouri Ozark Highlands. Canadian Journal of Forest Research, 26 , 573–583. Doolittle, W. T. (1958). Site index comparisons for several forest species in the Southern Appalachians. Proceedings of the Soil Science Society of America, 22, 455–458. Egler, F. E. (1954). Vegetation science concepts. I. Initial floristic composition—A factor in old-field vegetation development. Vegetatio, 4, 412–417. Eyre, F. H. (Ed.). (1980). Forest cover types of the United States and Canada. Bethesda, MD: Society of American Foresters. Fenneman, N. M. (1938). Physiography of Eastern United States. New York, NY: McGraw-Hill. Gingrich, S. F. (1967). Measuring and evaluating stocking and stand density in upland hardwood forests in the Central States. Forest Science, 13(1), 38–53. Gould, P. J., Fei, S., & Steiner, K. C. (2007). Modeling sprout-origin oak regener- ation in the Central Appalachians. Canadian Journal of Forest Research, 37 , 170–177. Gould, P. J., Steiner, K. C., Finley, J. C., & McDill, M. E. (2005). Developmental pathways following the harvest of oak-dominated stands. Forest Science, 51(1), 76–90. Gould, P. J., Steiner, K. C., McDill, M. E., & Finley, J. C. (2006). Modeling seed-origin regeneration in the Central Appalachians. Canadian Journal of Forest Research, 36 , 833–844. Heitzman, E., & Nyland, R. D. (1991). Cleaning and early crop-tree release in north- ern hardwood stands: A review. Northern Journal of Applied Forestry, 8(3), 111–115. Johnson, P. S. (1977). Predicting oak stump sprouting and sprout development in the Missouri Ozarks (Research Paper NC-149). St. Paul, MN: U.S. Department of Agriculture, Forest Service, North Central Forest Experiment Station. Johnson, P. S., Shifley, S. R., & Rogers, R. (2002). The ecology and silviculture of oaks. Cambridge, MA: CABI. D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 Central Appalachian Forest Regeneration 821 Johnson, R. L. (1980, January). New ideas about regeneration of hardwoods. Paper presented at the Hardwood Regeneration Symposium, Southeastern Lumber Manufacturing Association, Forest Park, GA. Johnson, R. L., & Deen, R. T. (1993). Prediction of oak regeneration in bottomland forests. In D. L. Loftis & C. E. McGee (Eds.), Oak regeneration: Serious problems, practical recommendations (General Technical Report SE-84, pp. 146–155). Asheville, NC: U.S. Department of Agriculture, Forest Service, Southeastern Forest Experiment Station. Larsen, D. R., & Johnson, P. S. (1998). Linking the ecology of natural oak regeneration to silviculture. Forest Ecology and Management, 106 , 1–7. Loftis, D. L. (1983). Regenerating red oak on productive sites in the Southern Appalachians: A research approach. In E. P. Jones Jr. (Ed.), Proceedings of the 2nd Biennial Southern Silvicultural Research Conference (General Technical Report SE-24, pp. 144–150). Asheville, NC: U.S. Department of Agriculture, Forest Service, Southeastern Forest Experiment Station. Loftis, D. L. (1989). Species composition of regeneration after clearcutting Southern Appalachian hardwoods. In J. H. Miller (Ed.), Proceedings of the 5th Biennial Southern Silvicultural Research Conference (General Technical Report SO-74, pp. 253–257). New Orleans, LA: U.S. Department of Agriculture, Forest Service, Southern Forest Experiment Station. Loftis, D. L. (1990). Predicting post-harvest performance of advanced red oak reproduction in the Southern Appalachians. Forest Science 36 (4), 908–916. Loftis, D. L. (2004). Upland oak regeneration and management. In M. A. Spetich (Ed.), Proceedings of the Upland Oak Ecology Symposium: History, current con- ditions, and sustainability (General Technical Report SRS-73, pp. 163–167). Asheville, NC: U.S. Department of Agriculture, Forest Service, Southern Research Station. McNab, W. H. (1988). Hardwoods and site quality. In H. C. Smith, A. W. Perkey, & W. E. Kidd (Eds.), Guidelines for regenerating Appalachian hardwood stands (SAF Publication 88-03, pp. 226–240). Bethesda, MD: Society of American Foresters. McNab, W. H., Loftis, D. L., & Shefield, R. M. (2003, October). Testing tree indicator species for classifying site productivity in Southern Appalachian hard- wood stands. Paper presented at the Society of American Foresters National Convention, Society of American Foresters, Bethesda, MD. McQuilkin, R. A. (1975). Growth of four types of white oak reproduction after clearcutting in the Missouri Ozarks (Research Paper NC-116). St. Paul, MN: U.S. Department of Agriculture, Forest Service, North Central Forest Experiment Station. Meiners, T. M., Smith, D. W., Sharik, T. L., & Beck, D. E. (1984). Soil and plant water stress in an Appalachian oak forest in relation to topography and stand age. Plant and Soil, 80(2), 171–179. Miller, G. W., & Kochenderfer, J. N. (1998). Maintaining species diversity in the Central Appalachians. Journal of Forestry, 96 (7), 28–33. Oliver, C. D., & Larson, B. C. (1996). Forest stand dynamics. New York, NY: John Wiley & Sons. D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 822 L. A. Vickers et al. Ott, R. L., & Longnecker, M. (2001). Statistical methods and data analysis (5th ed.). Pacific Grove, CA: Thomas Learning. Rogers, R., & Johnson, P. S. (1998). Approaches to modeling natural regeneration in oak-dominated forests. Forest Ecology and Management, 106 , 45–54. Ross, M. S., Sharik, T. L., & Smith, D. W. (1986). Oak regeneration after clearfelling in Southwest Virginia. Forest Science, 32(1), 157–169. Sander, I. L. (1972). Size of oak advance reproduction: Key to growth following harvest cutting (Research Paper NC-79). St. Paul, MN: U.S. Department of Agriculture, Forest Service, North Central Forest Experiment Station. Sander, I. L., Johnson, P. S., & Rogers, R. (1984). Evaluating oak advance repro- duction in the Missouri Ozarks (Research Paper NC-251). St. Paul, MN: U.S. Department of Agriculture, Forest Service, North Central Forest Experiment Station. Sander, I. L., Johnson, P. S., & Watt, R. F. (1976). A guide for evaluating the adequacy of oak advance reproduction (General Technical Report NC-23). St. Paul, MN: U.S. Department of Agriculture, Forest Service, North Central Forest Experiment Station. Smith. D. W. (1994). The Southern Appalachian hardwood region. In J. W. Barrett (Ed.), Regional silviculture of the United States (pp. 173–225). New York, NY: John Wiley & Sons. Smith, H. C. & Lamson, N.I. (1983). Precommercial crop-tree release increases diam- eter growth of Appalachian hardwood saplings (Research Paper NE-534). Upper Darby, PA: U.S. Department of Agriculture, Forest Service, Northeastern Forest Experiment Station. Smith, D. M., Larson, B. C., Kelty, M. J., & Ashton, P. M. S. (1996). The practice of silviculture: Applied forest ecology (9th ed.). New York, NY: John Wiley & Sons. Steiner, K. C., Finley, J. C., Gould, P. J., Fei, S., & McDill, M. (2008). Oak regeneration guidelines for the Central Appalachians. Northern Journal of Applied Forestry, 25(1), 5–16. Stout, S. L. (1991). Stand density, stand structure, and species composition in tran- sition oak stands of Northwestern Pennsylvania. In L. H. McCormick & K. W. Gottschalk (Eds.), Proceedings of the 8th Central Hardwood Forest Conference (General Technical Report NE-148, pp. 194–206). Radnor, PA: U.S. Department of Agriculture, Northeastern Forest Experiment Station. Trimble, G. R., Jr. (1973). The regeneration of Central Appalachian hardwoods with emphasis on the effect of site quality and harvesting practice (Research Paper NE-282). Upper Darby, PA: U.S. Department of Agriculture, Forest Service, Northeastern Forest Experiment Station. Trimble, G. R., Jr. (1974). Response to crop-tree release by 7-year-old stems of red maple sprouts and northern red oak advance reproduction (Research Paper NE-303). Upper Darby, PA: U.S. Department of Agriculture, Forest Service, Northeastern Forest Experiment Station. Waldrop, T. A., Buckner, E. R., Shugart, H. H., Jr., & McGee, C. E. (1986). FORCAT: A single tree model of stand development following clearcutting on the Cumberland Plateau. Forest Science, 32(2), 297–317. Yarnell, S. L. (1998). The Southern Appalachians: A history of the landscape (General Technical Report SRS-18). Asheville, NC: U.S. Department of Agriculture, Forest Service, Southern Research Station. D ow nl oa de d by [ N at io na l A gr ic ul tu ra l L ib ra ry ] at 1 0: 46 2 0 A ug us t 20 12 
work_kuqj5kgfqzfkfoou7u523f5j2q	----	Protrader: An Expert System for Program Trading PROTRADER: An Expert System for Program Trading by K.C. Chen, Theodore Brix Professor of Finance, School of Business, California State University, Fresno, CA 93740-0007 and Ting-peng Liang, Assistant Professor of Accounting, Department of Accountancy, University of Illinois at Urbana- Chamapign, Chamapign, IL 61820 Managerial Finance Volume 15 Number 5 1989 1 Abstract Recently, program trading has allowed arbitrageurs to take advantage of the discrepancies in the futures market and the stock market. The key that enables program trading is computer technology. This article presents the design of PROTRADER - an expert system prototype for program trading implemented in M.1. In particular, a learning mech- anism that allows the system to adapt to the changes in the market is presented. Introduction In the past several years the introduction of stock index fu- tures and options has created a new way of making money for institutional investors in the stock market. This new game on Wall Street is known as "program trading." It allows arbitrageurs to take advantage of the disparities be- tween the futures and stock markets and provides oppor- tunities to make risk-free profits. Traditionally, investors buy or sell stocks for several fundamental reasons, including the prosperity of economy and the outlook of a company's earnings. Today, however, an increasing number of program traders have substan- tially changed the face of the stock market. By funneling billions of dollars in and out of stocks in concentrated doses, program trading has caused wild swings in the mar- ket. For example, program trading has been blamed for triggering an 87-point drop in the Dow Jones Industrial Average on September 11, 1986 and a 508-point drop on October 19, 1987. The key factor that enables program trading is com- puter technology. Because of the volatile nature of the fu- tures and stock markets, only computers can provide the power required for keeping track of the rapid changes in both markets, usually several times a minute. Designing computer software that executes transactions based on a predetermined return on investment is not very difficult. These kinds of traditional systems, however, provide only limited support to program traders. In order to take full ad- vantage of program trading, an intelligent system that can adapt itself to dynamic markets is essential. In other words, the system must be able to learn changes in the markets and plan its investment strategy accordingly. In this article, we will present an expert system for pro- gram trading - PROTRADER (PROgram TRADER) - and the learning mechanism adopted by this system. The system is implemented in M.1 expert system shell. It monitors the differences between two markets, determines the optimum strategy for investment, executes transactions when ap- propriate, and enhances its knowledge by learning. A key issue that makes an expert system useful is its ability to adapt to a dynamic environment. In other words, the sys- tem must have learning capabilities. The learning mechan- ism to be described is based on a parameter adjustment approach that adjusts several critical parameters based on the market situation. Since program traders have only a small set of possible actions and most of the decision rules are highly dynamic, program trading is different from tradi- tional expert system domains such as medical diagnosis. Program Trading: The Problem Domain There are two markets on which stocks are traded: stock market and futures market. On the stock market, investors buy or sell stocks; on the futures market, however, inves- tors buy or sell stock index contracts. Currently only se- lected stock indexes are available in the futures market, such as the Standard & Poor's 500 index (S&P500) based on 400 industrial firms, 40 utilities, 20 transportation firms, and 40 financial institutions, and the Major Market Index (MMI) based on 20 blue-chip stocks. In general, the price of a futures contract (or the index futures) is higher than that of its underlying stocks (or the index) because an ar- bitrageur, in order to sell a futures contract safely, must purchase the underlying stocks and hold them for delivery on the expiration. The difference between the index and the price of the futures contract is called "premium" or "spread". On the third Friday of the maturing month, the fu- tures contract will expire and the futures index and the stock index will have to be the same. The idea of program trading is simple. Theoretically, the premium is equal to the costs for holding the stocks, i.e., the interest expenses less the dividends received from the stocks. In the real world, however, different demand and supply forces in two different markets sometimes drive the index futures far above or far below their "theoretical" fair value. In this case, an arbitrageur can trade the same amount of stocks in both markets and converts the pre- mium into profits when the futures contract expires. If the premium is higher than the fair value, program traders buy stocks that match or mimic the index, usually in a "basket" of $6 million to $12 million, and at the same time hedge those stocks by selling short the overpriced stock-index futures contracts. This transaction is usually called a buy program. If the premium is sufficiently lower than the fair value, then program traders sell stocks in the stock mar- ket and at the same time buy an equivalent amount of fu- tures contracts. This is called a sell program. The following example illustrates how the premium can be converted into a risk- free return (more examples can be found in (2, 3, 4)). 2 Managerial Finance Volume 15 Number 5 1989 (Example 1) On October 10, 1986, the S&P stock index and a futures contract on the index expiring in December were selling for 235.48 and 236.87, respectively, i.e., the premium was 1.39 points. Suppose an arbitrageur invested $11.77 million to execute a buy program and successfully unwind the program on expiration (December 19, 1986) by selling the purchased stocks and settling the futures con- tracts sold on October 10. We also assume that the inves- tor received $123,600 dividends from the purchased stocks (equivalent to an annual dividend yield of 4.2%). Since the program trader had no control over the stock market, the actual S&P 500 index on the expiration date could be higher or lower than the purchased index. In either situation, however, the investor had the same return on investment. This is why we call it risk-free. Situation 1: The stock market was bullish and the index ex- pired at 249.73. Gain from selling the stocks $712,500 (bought at 235.48; sold at 249.73) Loss from settling futures contracts (643,100) (sold at 236.87; settled at 249.73) Dividends from stocks 123,600 Total profit 193,100 Rate of return on $11.77 million 1.64% Transaction costs 0.50% Annualized net rate of return 8.21 % Treasury bill yield 6.00% Bonus risk-free rate 2.21% Situation 2: The stock market was bearish and the index expired at 230.24. Loss from selling the stocks $ (262,000) (bought at 235.48; sold at 230.24) Gain from settling futures contracts 331,500 (sold at 236.87; settled at 230.24) Dividends from stocks 123,600 Total profit 193,100 Rate of return on $11.77 million 1.64% Transaction costs 0.50% Annualized net rate of return 8.21% Treasury bill yield 6.00% Bonus risk-free rate 2.21% The major decisions involved in program trading are when and how much to invest. Although program trading guarantees a risk-free return, the volatility in both the stock and the futures markets makes it difficult to optimize these decisions. For example, when the premium is higher than its fair value, there is always a chance that the premium could be even higher. Therefore, the decision maker must decide whether to invest immediately when the premium is profitable or to wait for a higher return. The major factor that affects this decision is the probability that the premium will be higher. This probability may be affected by various economic and political factors and usually changes every day. In other words, knowledge and judgment are import- ant in assessing the probability and an expert system is crucial to improving the performance of investment. Design of PROTRADER The return on investment of program trading is affected by two factors: the premium and the period an investor has to hold a buy or sell program. Since the premium varies from time to time, the major functions of PROTRADER in- clude the following: 1. Monitor the premium in the market; 2. Determine the optimum investment strategy, including - When to execute a buy or sell program, - When to unwind a buy program, and - When to unwind a sell program. 3. Execute transactions when appropriate, and 4. Modify the knowledge base through a learning mechanism. Although these functions indicate that PROTRADER falls into several generic categories of expert systems in- cluding monitoring and control (1,13), its design is differ- ent from many traditional applications. Figure 1 illustrates the differences among several typical domains in terms of the size of knowledge base and the volatility of decision rules. Figure 1: Different Domains Size of the Knowledge B a s e In MYCIN, for example, there are many rules repre- senting a substantial amount of possible actions in the problem space; each rule has an associated probability (9, 11, 14). Both the rules and the probability are reasonably stable. That is, they represent the best judgment of experts. Managerial Finance Volume 15 Number 5 1989 3 Designing PROTRADER, however, is facing a different situ- ation. First, its knowledge base includes only a small amount of decision rules. Only two major indexes in the futures market are appropriate for program trading. For each of these indexes, there are five possible actions: ex- ecute a buy program, execute a sell program, unwind a previously executed buy program, unwind a previously ex- ecuted sell program, and wait for better opportunities. Therefore, unlike MYCIN, PROTRADER needs to consider only a small set of possible actions. Second, the uncertainty associated with each rule is extremely dynamic. Depending upon the market situation, the same rule may have different degrees of certainty at different time. For example, the probability that the pre- mium will go higher than 1.39 point may be lower than 5% one day, but higher than 15% on another day. Different as- sessments on the probabilities may lead to different invest- ment decisions. Therefore, rule probabilities are variables to be dynamically determined in PROTRADER. The follow- ing is a sample rule in the knowledge base (the capital let- ter 'N' represents a variable): rule-51: if have money = yes and premium is positive and return is profitable and confidence = N then decision = execute buy program ef N. With these two features, PROTRADER is composed of three main components: a knowledge base, a deductive reasoning mechanism, and an inductive learning mechan- ism. Figure 2 illustrates the relationships among these components. The monitor is a sensor that keeps track of the changes in the markets. Figure 2: Architecture of the System The knowledge base contains three categories of rules that determine when and how much to invest. First, it includes rules that determine the best strategy. In general, there are two strategies for program trading: one suggests the investors to hold a buy or sell program until expiration; whereas the other requires the investors to con- tinuously monitor the market and to take actions based on a set of pre-determined rules. Because evidence indicates that the latter strategy usually outperforms the former, PROTRADER takes it as the primary strategy. In addition, the system also maintains a database that contains the re- sources on hand, including money, executed buy pro- grams, and executed sell programs. Second, it includes rules that convert a premium into return on investment. Although return and premium are highly related, decisions are made based on return. Be- cause the conversion does not involve complex computa- tion, the built-in arithmetic functions of M.1 are sufficient for this application. Finally, it includes rules that provide suggestions and command transactions. Rule-101 illustrated in the follow- ing is an example rule-101: if message-text (ID, LIST OF PARAMETERS) = SUGGESTION and display (SUGGESTION) then message (ID, LIST OF PARAMETERS). These rules are then integrated by a capstone rule as follows: rule-1: if resources is sought and market trend is sought and return on investment is sought and printed suggestions then consultation is over. Based on the rules in the knowledge base, the deduc- tive reasoning mechanism allows the system to draw con- clusions regarding the investment decision. Since PROTRADER is implemented in M.1, it takes advantage of the backward reasoning capabilities of the shell. The major reason for choosing a shell rather than developing a de- ductive reasoning mechanism in an Al language was the capability of quick prototyping. The inductive learning mechanism of the system examines the variance between the real outcome and the forecasted outcome for each transaction and adjusts the probability associated with each decision rule. It plays a central role in the system. Because the stock market is highly volatile, human experts are involved in the learning process. More detail about the mechanisms will be dis- cussed in the next section. In the decision process, the system first examines the decision environment. It checks the available resources, examines important information in the real markets, such as the general outlook of the market, and builds a theore- tical market model based on the rules in the knowledge base. Then, the monitor keeps track of the premiums, and the deductive reasoning mechanism works with the knowl- edge base to find opportunities for program trading. When a profitable opportunity is identified, the system com- mands a transaction. Since the market is extremely dy- namic, the actual outcome of the transaction may be better or worse than the forecasted one. In either situation, the variance will activate the inductive learning mechanism to adjust the rules in the knowledge base. Learning Mechanism in PROTRADER Learning denotes changes in the system that are adaptive 4 Managerial Finance Volume 15 Number 5 1989 in the sense that they enable the system to do the same task or tasks drawn from the same population more effi- ciently and more effectively the next time (12). It plays a key role in improving the performance of expert systems. Previous research in artificial intelligence has identified several learning strategies (5), such as: 1. Rote learning: no inference or other transformation of knowledge is required; 2. Learning from instruction: the learner performs some inference, but the instructor presents the knowledge; 3. Learning by analogy: the learner acquires new facts or skills by transforming and argumenting existing knowledge that bears strong similarity to the desired new concept or skill that fits the new situation; 4. Learning from examples: the learner induces a general concept from a set of positive examples and counterexamples; and 5. Learning from observation and discovery: the learner acquires new concept in an unsupervised environment. It requires more inference than the previous four approaches. Among these five approaches, two of them are appro- priate for PROTRADER to adjust the probabilities associ- ated with its decision rules. First, the system automatically identifies the market trend by analyzing historical data and then determines the probability that the premium will ex- ceed a certain value. This is an application of learning by observation. Second, the system acquires the market trend from human experts and then determine the prob- ability based on their judgment. This is an application of learning from instruction and is called "human-aided learn- ing" in this article. Since the volatility of the stock market makes it almost impossible for a system to identify future market trends, human-aided learning is more feasible in this case. The human-aided learning strategy implemented in PROTRADER is basically a parameter adjustment mech- anism. The major parameter to be adjusted is the prob- ability that the premium may exceed a particular value. This value determines the probability associated with a decision rule. In the example described in the first section, for in- stance, a premium of 1.39 results in a 2.21% bonus risk- free return. This is certainly a profitable opportunity. However, if the probability that the market will have a pre- mium higher than 1.39 today is 60%, then the optimal strategy would be investing half of the money now and holding the rest for better opportunities, rather than com- mitting all money right away. How does the mechanism work? The major factor that affects the value of premium but cannot be controlled by the system is market trend. Therefore, human experts are responsible for forecasting and observing the market trend. By feeding the probability determination function with expert judgment, the system formulates a theoretical market model and provides suggestions accordingly. Since the theoretical market model only represents a fore- cast, frequent calibration is required. After the market is closing or when significant variance between the theoreti- cal market model and the actual market is found, human experts are asked to provide their observation on the mar- ket. This information is then used to modify the standard patterns. Therefore, the major components in the mechan- ism are the formulation of the theoretical market model and pattern modification. Figure 3 illustrates this co-operation. Figure 3: Human - aided Learning in PROTRADER Formulation of the Theoretical Model In order to determine the appropriate probabilities, the sys- tem maintains three standard patterns: bull market, neutral market, and bear market. Each pattern represents a prob- ability distribution of premium in the corresponding mar- ket situation. They are determined every day based on the transaction data of the previous day. Figure 4 shows exam- ples of these patterns. Figure 4: Examples of Standard Patterns Every day the experts forecast whether today's mar- ket will be bullish, neutral, or bearish. An uncertain judg- Managerial Finance Volume 15 Numbers 1989 5 ment such as 30% bullish and 50% neutral is allowed. Based on this information and the standard patterns, the system creates the market model by computing the prob- ability that the premium will be higher than a certain value. This probability determines the probability associated with a decision rule. The major issue in creating the theoretical model is how to convert standard patterns into a theoretical market model; or in other words, how a statement that "the mar- ket will be50% bullish and 30% neutral" can be represented by a probability distribution of premiums. In designing PROTRADER, we use an approach based on a normality assumption on market patterns. Since the daily volume of transactions is very large, normal distribution is a reason- able assumption in this application. By assuming that the premium distribution for each of the three market patterns is normal, we can represent each pattern by its mean and variance. The mean and variance of a theoretical market model can be calculated by the fol- lowing two equations: Where: Wi = the probability that the market will be an i type of market, Mi = the mean of the premium distribution in the standard market pattern i, Vi = the variance of the premium distribution in the standard market pattern i, i = 1, 2 and 3 denote the bull, neutral, and bear market, respectively. The following is a numerical example illustrating how the system formulates the theoretical market model. (Example 2) Suppose the expert judgment on the market trend on April 1, 1987 is "50% bullish and 30% neutral" and the system has the following standard patterns: Patterns Bull Neutral Bear Mean 3.5 0.5 -3.5 Variance 1.0 0.09 1.0 Following equations (1) and (2), the mean and variance of the premium distribution for the theoretical market model are: With these mean and variance, when the system detects a 2.3 points profitable premium, it will suggest the manager "to invest half of the money to execute a buy program and hold another half for other opportunities because the prob- ability of having higher premium is higher than 50%." In summary, given the assumption of normal distribu- tion, this approach has successfully simplified the calcula- tion involved in the formulation of the theoretical market model and has made it possible to implement a prototype in M.1. Pattern Modification Standard patterns are the basis on which a market model can be developed. Therefore, it is very important to ensure that they are accurate. In the ideal case, the mean and vari - ance recorded from the actual market are equal to those calculated from the theoretical model. In the real world, however, this rarely happens. After the market is closing or when significant variance between the theoretical and the actual market is found, human experts must be asked to provide the system with their observations on the actual market. This information enables the system to adapt itself to the changes in the market. The method for calibrating standard patterns is simi- lar to the approach previously described. The system first asks human experts to present their observation on a 5 to -5 scale, as follows: What do you think would best describe today's mar- ket until now? Please indicate your observation on the fol- lowing scale: The value provided by human experts reflects a pos- terior observation which is likely to be more accurate than the prior forecast used to formulate the theoretical model. By modifying equations (1) and (2), the mean and variance of the modified standard pattern can be calculated. Since the system has only one observation, some of the patterns must remain unchanged in the calibration process. In selecting the pattern to be modified, the general rule is that the neutral case remains unchanged unless the ob- servation happens to be neutral (i.e., value = 0). In other words, if the value is higher than zero, then the system will modify the pattern of bull market. If the value is less than zero, then the system will modify the pattern of bear mar- ket. Equations (3) and (4) are used to adjust the means and variances of the bull and bear market patterns. If the value 6 Managerial Finance Volume 15 Number 5 1989 is equal to zero, then the system uses equations (5) and (6) to reassign the neutral pattern. Where: Mi = mean of standard pattern i Vi = variance of standard pattern i Ma = mean recorded from the actual market Va = variance recorded from the actual market Mo = mean of the neutral market V0 = variance of the neutral market b = value provided by human experts; -55 i = bull, if b > 0 = bear, if b < 0 . c = b , if b > 0 ; = -b, if b < 0 . (Example 3) This example continues the analysis shown in example 2. If at the closing of the market the system has recorded a mean of 2.1 and a variance of 0.7 and the ex- pert observation on the market is 2 which means "more or less bullish", then following equations (3) and (4), the new pattern for the bull market can be created. Next time the system is used this pattern will replace the old standard for generating the theoretical market model. Concluding Remarks The rapid progress of computer technology has enabled program trading which allows arbitrageurs to take advant- age of the disparities in the futures market and the stock market to make risk-free profits. Because of the volatile na- ture of both markets, the probability associated with each decision rule needs to be dynamically determined every time the system is used. This makes the development of an expert system for program trading different from tradi- tional applications such as MYCIN. In this article, we have presented a prototype system for program trading and the issues involved in designing this system. In particular, we have described a learning mechanism that integrates human observations and the knowledge of the system to improve the performance of the system. Although the work has been focused on pro- gram trading, the techniques described here can be gener- alised to any domain of high degree of volatility. References (1) Basden, A. (1983), "On the Application of Expert Sys- tems," International Journal of Man-Machine Studies, 19, pp. 461-477. (2) Business Week (1985), A New Breed of Investor is Whipsawing Wall Street, September 23, pp. 84-86. (3) Business Week (1986), Those Big Swings on Wall Street, April 7, pp. 32-36. (4) Business Week (1986), How Chicago Zaps Wall Street, September 29, pp. 32-36. (5) Carbonell, J.G., Michalski, R.S. and Mitchell, T.M. (1983), "An Overview of Machine Learning," in Michalski, R.S., Carbonell, J.G. and Mitchell, T.M. (eds), Machine Learning: An Artificial Intelligence Approach, Vol. I, Los Altos, CA: Morgan Kaufmann Publishers, Inc., pp. 1-24. (6) Coombs, M. and Alty, J. (1984), "Expert Systems: An Alternative Paradigm," International Journal of Man-Ma- chine Studies, 20, pp. 21-43. (7) Finnerty, J.E. and Park, H.Y. (1987), "Empirical Evi- dence on Stock Index Arbitrage: The Case of Program Trading," BEBR Faculty Working Paper 1321, College of Commerce and Business Administration, University of Illi- nois at Urbana-Chamapign. (8) Forsyth, R. (1984), "Fuzzy Reasoning Systems," in For- syth, R. (ed.), Expert Systems: Principles and Case Studies, London, U.K.: Chapman and Hall, pp. 51-62. (9) Hayes-Roth, F., Waterman, D.A., and Lenat, D.B. (1983), Building Expert Systems, Reading, MA: Addison- Wesley Publishing Co. (10) Schmucker, K.J. (1984), Fuzzy Sets, Natural Language Computations, and Risk Analysis, Rockville, MD: Com- puter Science Press. (11) Shortliffe, E.H. (1976), Computer-based Medical Con- sultations: MYCIN, Amsterdam, The Netherland: Elsevier Scientific Publishing Co. (12) Simon, H.A. (1983), "Why should Machines Learn," in Michalski, R.S., Carbonell, J.G. and Mitchell, T.M. (eds.), Machine Learning: An Artificial Intelligence Approach, Vol. I, Los Altos, CA: Morgan Kaufmann Publishers, Inc., pp. 25-37. (13) Stefik, M., et al (1982), "The Organization of Expert Systems: A Tutorial," Artificial Intelligence, 18:2, pp. 135- 174. (14) Waterman, D.A. (1986), A Guide to Expert Systems, Reading, MA: Addison-Wesley Publishing Co. (15) Zualkernan, I., Tsai, W.T. and Volovik, D. (1986), "Ex- pert Systems and Software Engineering: Ready for Mar- riage," IEEE Expert, 1:4, pp. 24-31. (16) Zadeh, LA. (1973), "Outline of a New Approach to the Analysis of Complex Systems and Design Pro- cesses,"IEEETransactions on Systems, Man, and Cyber- netics, Vol. 3, pp. 28- 45. 
work_kwxbifzmlrcnjkqrj2uy5wnbmi	----	Hindawi Publishing Corporation Mathematical Problems in Engineering Volume 2013, Article ID 343171, 8 pages http://dx.doi.org/10.1155/2013/343171 Research Article The Design and Implementation of an Intelligent Apparel Recommend Expert System A. H. Dong,1,2 D. Shan,1 Z. Ruan,1 L. Y. Zhou,3 and F. Zuo1,2 1 College of Information Science and Technology, Donghua University, Shanghai 201620, China 2 Engineering Research Center of Digitized Textile & Fashion Technology, Ministry of Education, Shanghai 201620, China 3 College of Textile, Donghua University, Shanghai 201620, China Correspondence should be addressed to A. H. Dong; dongaihua@dhu.edu.cn Received 11 January 2013; Accepted 5 February 2013 Academic Editor: Yang Tang Copyright © 2013 A. H. Dong et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Now with the rapid development of information science and technology, intelligent apparel recommend has drawn wide attention in apparel retail industry. Intelligent management and effective recommend are two issues of crucial importance for the retail store to enhance its corporate influence and increase its economic benefits. This paper proposes an intelligent recommend system design scheme for apparel retail which is based on expert system. By comprehensive utilization of database management and expert system technology, the proposed system provides a solid solution in improving the customer shopping experience. This paper presents a kind of object-oriented blackboard structure, which is applied in the apparel recommend expert system and establishes expert rule on the basis of apparel characteristic elements. Through the establishment of the rule base, the system generates personal recommend list by positive rule reasoning mechanism engine. The proposed method thus gives dress collocation scheme for the customer through the human-machine interaction from the point of view of the apparel experts. This design scheme avails the customers to experience targeted service with intellectualization, and personalization and it has certain reference significance for promoting apparel retail intelligence development. 1. Introduction Expert system is one of the most active and productive re- search in the artificial intelligence field now. Since Edward Feganbaum developed the first expert system, DENDRL, expert system has developed rapidly. It simulates the experts’ thinking process of solving the problem, and is widely used in medical diagnosis, geological exploration, the petroleum chemical industry, and so forth. Also, it brings great eco- nomic and social benefits in these fields [1]. With the expanding of apparel consumer market and increasingly fierce competition, apparel intelligent recom- mend has drawn wide attention in apparel retail industry. Intelligent recommend is an issue of crucial importance for the retail store to enhance its corporate influence and increase its economic benefits. Over the last several decades, apparel recommended technology has made some achievements at home and abroad. Recommender system firstly appeared in 1997 [2]. It refers to an application technology for knowledge discovery in a specific database. Some scholars applied intel- ligent agent and collaborative filtering technology in recom- mender service [3, 4]. Most of the recommendation systems are designed based on the analysis of product attributes, customer attributes, or customer behaviors. In 2003, Lam invented an apparatus for personalized apparel coordina- tion. This invention facilitates a user to select, coordinate, appraise, and harmonize apparel without actually putting on the clothing [5]. However, the apparatus is unable to give any evaluations about the coordination. In 2004, Kobayashi proposed a color planning method for coordinating apparel colors [6]. However, this method was from the perspective of apparel color to make dress collocation and recommend for users. In 2009, Wong et al. proposed a fashion mix-and- match expert system [7]. In their study, the expert system was developed to provide customers with professional and systematic mix-and-match recommendations automatically. 2 Mathematical Problems in Engineering This system could capture the knowledge and emulate the decisions of fashion designers on apparel coordination, and its knowledge base could store the literal form of informa- tion. This system thus generated effective mix-and-match recommendations for customers. Although expert system and apparel recommended technology matures, the case that intelligent recommend system applied in apparel retail is less. In this study, intelligent recommend system design scheme for apparel retail based on expert system technology is proposed to provide professional and systematic expert recommendations. As is known to all, owing to custom influence, seasonal factor, and fashion trend, apparel industry updates and changes quickly. Improving service quality is therefore essential for enhancing the brand image so as to obtain more economic benefits for the apparel retail. Dress is different from other commodities. It is related to many factors, such as color, style, and collar type. In view of the particularity of apparel goods, this system first consults the customer information such as the body characteristics and color of skin conditions. It then automatically generates recommended clothes which is suitable for the customer. In other words, it can simulate the thinking process of apparel recommend experts and use expert rules to recommend clothes for customers in the perspective of professionals. It avails the customers to experience targeted service with intellectualization and personalization, and it has certain ref- erence significance for promoting apparel retail intelligence development. The rest of the paper is organized as follows: Section 2 formulates the structure of apparel intelligent recommend expert system. It also analyzes how to represent and acquire the expert system rules. Section 3 presents expert system working process and principle. At the same time, it elaborates the system structure of blackboard in details. Section 4 describes the system design and implementation with exper- imental data. Section 5 draws the conclusion and the states of the possible further research direction. 2. Designing of Intelligent Apparel Recommend Expert System Among all artificial intelligence research fields, expert system has the most practical value, and recently its application technology becomes more mature. Expert system is a kind of computer programming system, which uses specialized knowledge and ratiocination to solve problem in some spe- cific fields [8]. That is to say, abundant specialized knowledge about the costume collocation has already been stored in the expert system illustrated in this paper. For those customers who do not qualify with this specialized knowledge, the proposed apparel recommend expert system can generate the recommendations sheet according to their personal informa- tion and give explanations on these recommendations. 2.1. Structure of Intelligent Apparel Recommend Expert System. The expert system illustrated in this paper is based on the expert rules. It can help customers selecting the most suitable clothes among the abundant apparel according to personal condition and increase customers’ satisfaction. The structure of apparel recommended expert system is shown in Figure 1. It contains human interface, inference machine, knowledge base, database, and explaining organization. The rule base and inference machine are the kernel of the system. Among all these modules in Figure 1, human machine interface (HMI) module is to realize the interaction between the system and customers. Firstly, the customer’s basic information is obtained from the customer via the HMI to generate the personal information elements (skin color preference, figure type, face type, etc.). On the other hand, the recommendation sheet is also provided to the customer via HMI. Inference machine is the “brain” of the system. Depth-first search mechanism is applied in the expert system [9]. It uses production rule (a kind of forward reasoning method) [10] to invoke the expert rule in the knowledge base, and finally generates the apparel recommendation sheet. Knowledge base module contains fact base and rule base. Fact base is designed to store the train data of clothes. Rule base is the main part of the knowledge base. The dynamic database stores the intermediate result of system ratiocination. The explaining organization gives reasons on the final recommendation according to the intermediate ratiocination. It makes the expert system more transparent, and further on, increases the customers’ satisfaction. Figure 2 shows the flow of the system working. 2.2. Presentation of Rule. Generally, human knowledge expresses and spreads via words and language. However, in order to be understood and ratiocinated by computer, it has to be formalized into a model which computer can recognize. This system introduced a definition of production rule. Usually, knowledge definition of production rules expressed as follows [11]: If 𝐴 Then 𝐵. (1) Which means if𝐴 set up,𝐵 also, established. It is simplified to 𝐴󳨀→𝐵. (2) 2.3. Rules Acquisition. Apparel is a kind of peculiar com- modity. From the perspective of costume design, each piece of apparel is composed of variety of characteristic elements. Whether each apparel characteristic element or the collocation between each factor matchs mainstream aesthetic is closely related to the physical characteristics of personal appearance, body conditions, or other factors. This paper combined with the research on expert advice on apparel merchandise put costume commodity into varieties of characteristic elements and characteristic elements of the decomposition factor. Table 1 shows the detailed apparel commodity characteristics elements, including color, style, collar type, pants type, skirt type, waist type, and sleeve type. Taking the advice of dress collocation expert, the pro- posed system is designed from the following aspects: the collocation of skin color and apparel color, the collocation of stature and apparel style, the fitness of collar type to the Mathematical Problems in Engineering 3 The expert system structure module Customer Generating recommended resultsDialogue HMI module Explanation module For explaining recommendation reason Fact database Inference engine Rule base Call Positive reasoning mechanism Knowledge base Search Apparel commodity training data Expert rules Deposit Deposit Backend database Expert Consultation and investigation from experts Figure 1: Structure of apparel intelligent recommend expert system. Table 1: Apparel merchandise characteristic elements of the decom- position table. Characteristic elements Name of characteristic elements of factoring Color Yellow, cream, orange, blue, shallow water, light gold, gold, orange, red, brown, black, white, gray, rose red, beige, green fruit, pink, blue, brown, pink Style Commuter, Korean, British, fresh, lady, sweet, sexy, leisure Collar type V-neck, low round neck, deep U-neck collar, round neck, POLO collar, stand-up collar, high collar Pants type Harem pants, leggings, slim pants, pencil pants Skirt type A-line skirt, pleated skirt bud skirt, pencil skirt,mini skirt, pleated skirt Waist type Waist, high waist, low waist Sleeve type Puff, conventional sleeves, dolman sleeves, raglansleeves face, and the apparel style of different situations. Then expert rules of this system are generated by selecting eight data types as characteristic elements, such as apparel color, style, collar type, and so forth. And different expert rules constitute different knowledge sources (KS). 3. Blackboard Structure of Intelligent Apparel Recommend Expert System 3.1. Blackboard Structure. Blackboard structure [12] is applied to the proposed apparel intelligent recommend expert system. It contains three parts: knowledge sources, blackboard and controlling program. Actually, blackboard structure is the global database which is used to store the original data, intermediate rati- ocination process, and ratiocination results; it acts as the medium in the interaction and communication of different knowledge sources. The information on the blackboard struc- ture is divided into many layers, and each layer represents the description of costume information in different detail levels [13]. High level apparel collation scheme evolves from low level ones, so their relationship can be expressed in a tree structure. Generally, only knowledge sources can modify, insert and delete information on the blackboard. In the proposed experts system, knowledge sources are the set of rules that generate apparel recommended program from the same point of view. Each knowledge source can complete apparel experts recommend work from different point of view of customers. Knowledge sources include both premise and action. Knowledge sources are independent of each other that cannot be called directly and can only communicate through the blackboard. Control mechanism is used to monitor the change of blackboard information, which can inspect the premise of each knowledge source continuously. Usually, once the premise of a knowledge source is established, this knowledge will be activated immediately, conduct the action of its part, and generate corresponding apparel recommendation scheme on the blackboard. And when change signal is sent by the information on blackboard, control mechanism can activate other knowledge sources, until the final recommen- dation scheme is generated [14]. The discipline is that when 4 Mathematical Problems in Engineering Y N N N Y Y Apparel commodity training data Deposit Backstage database Extraction Extraction Apparel characteristic elements Search next apparel characteristic elements Customer Customers personal information Dialogue Man-machine interaction module Generate Search next oneSearch next one Whether matching? Whether searching customers personal information is over? Apparel recommend results Show to customers Continue to screen by price and material factors, and so forth Whether the customer is satisfied? The final recommendation results Figure 2: Flow chart of apparel intelligent recommend expert system. the premises of multiple knowledge sources are coupled with the change state of blackboard information, the control mechanism can always choose the knowledge source which offers the best assistance to solve the problem to function first. 3.2. Task Tree of Blackboard Structure. The blackboard model is reasoning to random combinations to multiple knowledge sources multisub(task). The scheduler program between each subtask is a meta-level reasoning [15]. It make the use of the information generated from the blackboard to scheduling the reasoning of various knowledge sources; that is, the system is a meta-level knowledge and meta-level reasoning apparel rec- ommended expert system. Based on the blackboard structure model, the problem that needs to be solved can be divided into a quest tree [16]. Therefore, a problem consists of multiple tasks, and each task can be solved by taking advantage of different knowledge sources. As shown in Figure 3, the goal of apparel recommended expert system is to generate apparel recommended menu. This paper divides general task into six subtasks; each subtask took advantage of knowledge source to solve. The reasoning process is designed in positive reasoning mechanism engine [17]. For example, through invoking the knowledge source of apparel color and skin color collation system can generates an apparel color scheme which is suitable for this customer. In another word, system recommends this customer in the aspect of apparel color and skin color collation, and generate the color scheme. Based on this recommendation, system invokes next knowledge source, makes a recommendation in the aspect of the collation of pants/skirt type and stature, and generate, corresponding scheme. And the like, system will do the similar job until transversing all tasks. 3.3. Analysis of the Blackboard Structure. The blackboard structure of apparel recommended expert system is shown in Figure 4. Blackboard structure is the workspace of the global database which is used to store all kinds of statuses produced in the ratiocination. It is divided into different layers. Data used and modified by different knowledge sources is on different layers [18]. Blackboard structure of this system is divided into 7 layers; each layer dispatches this layer’s knowledge sources (KS) based on the upper layer, generates new recommendation scheme, and stores it to the next layer of the blackboard until all KS are dispatched. KS of this system Mathematical Problems in Engineering 5 To generate apparel recommend menu Color Pants/skirt type Sleeve type Waist type Collar type Style KS1 collocation of clothes color and skin color KS2 collocation of clothes type and stature KS3 collocation of collar type and face type KS4 collocation of clothes style and occasion Positive reasoning machine Figure 3: Task tree of blackboard structure. can be divided into 4 kinds: KS1 is about the collocation of clothes color and skin color, KS2 is about the collocation of clothes type and stature, KS3 is about the collocation of collar type and face type, and KS4 is about the collocation of clothes style and situation. Command is used to activate KS when scheduling sequence dispatch KS. The scheduling sequence is generated by the scheduling program, which is stored in the control database of this expert system. After the interaction between customer and expert sys- tem, system will store the result of interaction on layer 1 of the blackboard automatically. Scheduling program will generate the scheduling sequence, activate KS1, and generate the recommendation scheme in the aspect of color. Then it will store the scheme on layer 2 of black board. System will dispatch KS2, continue the similar preclude on layer 2, and generate the recommendation scheme in the aspect of clothes type. And the like, till invoking KS4, system will form the final clothes recommendation sheet in all the 4 aspects: clothes color, style, type, and dress occasion. Figure 5 illustrates the execution process of scheduling program from the bottom layer to the top layer of the black board. If the final recommendation sheet does not satisfy the customer successfully, the system could go back to the bottom layerand redo the whole procedure after more interactions with the customer, that makes the system have the feedback link of closed loop. 4. Realization of Intelligent Apparel Recommend Expert System 4.1. Development Environment. The common expert system developing languages include PROLOG and LISP, which are specially designed for the development of artificial intelli- gence system. The disadvantages of these languages include poor performance of visualization, difficulty in developing foreground, and bad controllability of reasoning process. In this paper, SQL Server 2005 is applied to store and process data while foreground program is developed in Visual C#. C# is objected-oriented language, and is the first choice for the development of foreground and various modules because of its excellent performance in the human-machine interaction and graphical display [19]. 4.2. System Realization. The principle of the proposed apparel intelligent recommend expert system is that it changes recommend process into a state space search by using the known facts that constitute the initial state to generate of target state. The realization of the recommendation process therefore can be simplified to search route which is from the initial state to a target state in the whole state space by inference machine. System testing data mainly sources from the opening apparel websites. In order to simulate the expert system of apparel shop, this system selects nearly 100 items of apparel information and stores them in the background database. And all these items form the training data set of the intelligent recommend expert system, including apparel name, apparel type, style, color, sleeve type, collar type, waist type, price, season, and material. Apparel intelligent recommend expert system can call the rules in the rule base to match the fact in the fact repository according to the information which the customer submits via HMI. It then gets expert recommendation results. For example, a customer submits her information to this system which includes shallow yellowness of skin color, standard body shape, inverted triangle face, and dress occasions for home leisure. The expert system could generate apparel recommendation results after running program and show it to the customer via HMI, as shown in Figure 6. 6 Mathematical Problems in Engineering Consultation customer physical characteristics Color recommend scheme Pants/skirt-type Sleeve-type Waist-type Collar-type recommend scheme recommend scheme recommend scheme recommend scheme Style recommended scheme KS of collocation of clothes color and skin color KS of collocation of clothes style and occasion KS of collocation of collar type and face type KS of collocation of clothes type and stature KS of collocation of clothes type and stature KS of collocation of clothes type and stature Positive reasoning machine Operation sequence Scheduling program Control database Blackboard Knowledge sources Data flow Control flow Figure 4: Blackboard structure of system. If the customer is not satisfied with the recommendations of the system, the system allows the user to further screening according to the price, material, and so forth on the basic of recommendation results. At the same time, expert system can explain the reason of generating recommendation results. If the customer is still not satisfied with recommendation results, this system will automatically recommend popular apparel in sales season. 5. Conclusions In this paper, an intelligent recommend expert system design scheme for apparel retail is proposed to provide professional and systematic apparel expert recommendations. The system designs a kind of object-oriented blackboard and establishes expert rule on the basis of apparel characteristic elements. Through the establishment of rule base, the system generates personal recommend list by positive rule reasoning mecha- nism engine. The proposed method thus gives dress collo- cation scheme for the customer through the HMI from the point of view of the apparel experts. The experimental results using the test data demonstrate that the proposed system is effective. In short, the intelligent recommend method based on expert system could recommend the suitable apparel for the customer and improve the customer experience level. It avails the customers to experience targeted service with intellectualization and personalization, and it has certain ref- erence significance for promoting apparel retail intelligence development. As the popular trend of apparel varies year by year, related recommend knowledge from the experts may be changed as well. How to make the recommend expert systems generate Mathematical Problems in Engineering 7 Setting activation condition of KS equals 1 Scheduling program Getting information from the first layer in blackboard by consulting customer physical characteristics Check the task activation conditions Which task of activation condition equals 1? Error task task of KS Extract information of current Start programs that current Reasoning by this KS Find next task name by the result of reasoning Send the result of reasoning to the blackboard of next task Whether the next task is empty ? Current task activation Put the recommend results into top of the blackboard End N N Y Y conditions equals 0 and the next equals 1 Figure 5: System scheduling program flow diagram. Figure 6: Effectiveness demonstration of system. new rules automatically, namely, machine learning, is one of the possible directions of future research. Acknowledgments The financial support from the National Natural Science Foundation of China (12K10130) and from the Shanghai Natural Science Foundation (12S10155) is gratefully acknowl- edged. References [1] K. W. Chau and F. Albermani, “An expert system on design of liquid-retaining structures with blackboard architecture,” Expert Systems, vol. 21, no. 4, pp. 183–191, 2004. [2] P. Resnick and H. R. Varian, “Recommender systems,” Commu- nications of the ACM, vol. 40, no. 3, pp. 56–58, 1997. [3] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, “Analysis of recommender algorithms for E-Commerence on electronic commerence,” in Proceedings of the 2nd ACM Conference on Electronic Commerce, pp. 158–167, 2000. 8 Mathematical Problems in Engineering [4] B. Mobasher, H. Dai, T. Luo, Y. Sun, and J. Zhou, “Integrating web usage and content mining for more effective personal- ization,” in Proceedings of the International Conference on E- Commence and Web Technologies, pp. 165–176, 2000. [5] P. A. F. Lam, “Personalized garment coordination apparatus,” United States Patent, vol. 6, pp. 629–644, 2003. [6] M. Kobayashi, “Studies on the color planning of clothing— existence of the ”ideal skin color” and the effect of the clothing color,” Journal of the Japan Research Association for Textile End- Uses, vol. 45, no. 3, pp. 56–63, 2004. [7] W. K. Wong, X. H. Zeng, W. M. R. Au, P. Y. Mok, and S. Y. S. Leung, “A fashion mix-and-match expert system for fashion retailers using fuzzy screening approach,” Expert Systems with Applications, vol. 36, no. 2, pp. 1750–1764, 2009. [8] R. Zhang, J. Lu, and G. Zhang, “A knowledge-based multi-role decision support system for ore blending cost optimization of blast furnaces,” European Journal of Operational Research, vol. 215, no. 1, pp. 194–203, 2011. [9] X. Y. Chi, H. J. Ma, Z. Zhao, and Y. H. Peng, “Research on hybrid expert system application to blanking technology,” Journal of Materials Processing Technology, vol. 116, no. 2-3, pp. 95–100, 2001. [10] J. Akoka and I. Comyn-Wattiau, “An expert system for software reverse engineering,” in Procedings of the 4th World Congress of Expert Systems, Application of Advanced Information Technolo- gies, vol. 1, pp. 209–217, 1998. [11] M. Weiss and F. Stetter, “Hierarchical blackboard architecture for distributed AI systems,” in Proceedings of the 4th Inter- national Conference on Software Engineering and Knowledge Engineering, pp. 349–355, June 1992. [12] C. Reidsema and E. Szczerbicki, “A blackboard database model of the design planning process in concurrent engineering,” Cybernetics and Systems, vol. 32, no. 7, pp. 755–774, 2001. [13] Y. Tang, Z. Wang, H. Gao, S. Swift, and J. Kurths, “A constrained evolutionary computation method for detecting controlling regions of cortical networks,” IEEE Transactions on Computa- tional Biology and Bioinformatics, vol. 9, pp. 1569–1581, 2012. [14] Y. Tang, H. Gao, J. Kurths, and J. Fang, “Evolutionary pinning control and its application in UAV coordination,” IEEE Trans- actions on Industrial Informatics, vol. 8, pp. 828–838, 2012. [15] Y. Tang, H. Gao, W. Zou, and J. Kurths, “Identifying controlling nodes in neuronal networks in different scales,” PLoS ONE, vol. 7, no. 7, Article ID e41375, 2012. [16] K. Eldrandaly and S. Naguib, “A knowledge-based system for GIS software selection,” International Arab Journal of Informa- tion Technology. In press. [17] G. Brzykcy, J. Martinek, A. Meissner, and P. Skrzypczyński, “Control aspects of the blackboard agent architecture for a mobile robot,” Control and Cybernetics, vol. 32, no. 4, pp. 851– 866, 2003. [18] J. G. Harvey and D. D. Harris, “Enhancing the blackboard concept to produce a flexible, extensible knowledge-based design tool,” Knowledge-Based Systems, vol. 9, no. 4, pp. 233– 243, 1996. [19] J. H. Liu, “The design of expert system and its realization based on C#,” Information and Technology, no. 4, pp. 150–153, 2006. Submit your manuscripts at http://www.hindawi.com Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Mathematics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Mathematical Problems in Engineering Hindawi Publishing Corporation http://www.hindawi.com Differential Equations International Journal of Volume 2014 Applied Mathematics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Probability and Statistics Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Mathematical Physics Advances in Complex Analysis Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Optimization Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Combinatorics Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Operations Research Advances in Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Function Spaces Abstract and Applied Analysis Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Mathematics and Mathematical Sciences Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Algebra Discrete Dynamics in Nature and Society Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Decision Sciences Advances in Discrete Mathematics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Stochastic Analysis International Journal of 
work_kxlz6lnai5bdbllarzrckrpaxa	----	Design and Application of a Multi-Variant Expert System Using Apache Hadoop Framework sustainability Article Design and Application of a Multi-Variant Expert System Using Apache Hadoop Framework Muhammad Ibrahim * and Imran Sarwar Bajwa Department of Computer Science & IT, The Islamia University of Bahawalpur, Bahawalpur 63100, Pakistan; imran.sarwar@iub.edu.pk * Correspondence: ibrahimbwp@gmail.com Received: 20 October 2018; Accepted: 12 November 2018; Published: 19 November 2018 ����������������� Abstract: Movie recommender expert systems are valuable tools to provide recommendation services to users. However, the existing movie recommenders are technically lacking in two areas: first, the available movie recommender systems give general recommendations; secondly, existing recommender systems use either quantitative (likes, ratings, etc.) or qualitative data (polarity score, sentiment score, etc.) for achieving the movie recommendations. A novel approach is presented in this paper that not only provides topic-based (fiction, comedy, horror, etc.) movie recommendation but also uses both quantitative and qualitative data to achieve a true and relevant recommendation of a movie relevant to a topic. The used approach relies on SentiwordNet and tf-idf similarity measures to calculate the polarity score from user reviews, which represent the qualitative aspect of likeness of a movie. Similarly, three quantitative variables (such as likes, ratings, and votes) are used to get final a recommendation score. A fuzzy logic module decides the recommendation category based on this final recommendation score. The proposed approach uses a big data technology, “Hadoop” to handle data diversity and heterogeneity in an efficient manner. An Android application collaborates with a web-bot to use recommendation services and show topic-based recommendation to users. Keywords: recommender systems; opinion mining; SentiWordNet; polarity scores 1. Introduction Since the advent of web intelligence, artificial intelligence-based services, frameworks and products have become popular in the World Wide Web. One of the key services of such web intelligence applications is a recommendation system. In recent times, recommendation systems have become popular in the domain of movies, music, books, restaurants, garments, mobile applications and many other fields of life. Such recommendation systems filter huge amount of structured and unstructured data and predict the preference of a user that one would give to an item. In the last decade, a few movie recommendation systems have been presented using conventional methods [1–3]. However, previous movie recommender systems lack various features and/or accuracy of true recommendation. A majority of these recommender systems use quantitative variables (likes or ratings) and a few others use qualitative variables (polarity score, etc.) [4–7]. This paper proposes an intelligent and automated recommendation system that provides two-fold novelty. First, our recommender system uses a multi-variant popularity matrix to recommend a suitable movie to a user on the basis of both quantitative and qualitative variables to achieve true recommendations. Secondly, a fuzzy logic-based module provides the final recommendation of movies in a particular field of user’s choice (such as comedy, action, horror, fiction, etc.), whereas the currently available systems give general recommendations. In our multi-variant recommendation system, one of the challenges was opinion mining of users’ reviews to calculate a polarity score that shows the degree of likeness or dis-likeness of a movie by Sustainability 2018, 10, 4280; doi:10.3390/su10114280 www.mdpi.com/journal/sustainability http://www.mdpi.com/journal/sustainability http://www.mdpi.com https://orcid.org/0000-0002-5161-6441 http://www.mdpi.com/2071-1050/10/11/4280?type=check_update&version=1 http://dx.doi.org/10.3390/su10114280 http://www.mdpi.com/journal/sustainability Sustainability 2018, 10, 4280 2 of 21 a user [8–12]. Such polarity scores provide a qualitative aspect of user’s opinions about a particular movie. Another challenge was how to handle the diversity of heterogenous data as the presented approach uses both quantitative and qualitative data. To handle this issue a big data solution involving Hadoop was used in our approach because it efficiently handles data heterogeneity and data diversity in a better way. Nowadays society has changed, people own smartphones and they are highly dependent on mobile applications such as recommendation systems, which need to communicate with smartphone Apps so that users can easily interact with the services and efficiently select the recommended items [9]. Therefore, in this paper, a recommender system is coupled with an Android application and a web-bot offering open web services and merging movie data from linked data composed with different external resources. Big data is defined by four dimensions represented by four V’s (volume, variety, velocity, and veracity). Volume is represented by the amount of text data that we use to generate recommendation. Variety represents the different types of data extracted from different sources like blogs, Facebook, and Twitter as well as different review and opinion sites. Reviewers can write their reviews, remarks, and feedback in any format-like structure, semi-structured, or unstructured and these should be handled by the system. Velocity represents the speed of data generation on the internet. Veracity represents the trust worthiness of the data. A multi-variant recommendation system can get benefit of a NoSQL environment in reducing complexity and to handling the sparsity by factorization, ensuring scalability by using an empowered server machine and dealing with heterogeneity by using Hadoop platform to handle the big data issues [13–15]. Ontology and linked data can be information sources for movies’ descriptions and are available among Internet applications and are provided through the semantic queries of standard web technologies, such as URIs, RDF, HTTP and the semantic web. The linked data from Google Places, Trovacinema, Wikipedia and netflix or linked movie databases (http://linkedmdb.org) are useful for the recommender systems [16]. The work presented in [13] discusses a movie recommendation system that uses movie ratings to recommend movies only in a general category. However, this work seriously lacks accuracy of true recommendations. The reason for less accuracy in [13] is the use of only numeric data such as likes and ratings. Such numeric features only cover the quantitative aspect of the users’ likeness. However, the qualitative aspect of likeness of users is totally ignored in this work, which makes the results of this work questionable. Here, it is important to mention that quantitative and qualitative aspects of likeness of users can provide us with true recommendations. The qualitative aspect of likeness can be achieved from text reviews that were not covered by the approach used in [13]. Moreover, this approach is tested on a single small dataset. Other issues with [13] are tabulated in Table 1. There is need of a multivariate approach that involves both quantitative and qualitative aspects to finalize a recommendation and achieve highly accurate results. Table 1 represents differences in previous approaches and our multi-variant approach. Table 1. Deviation in different approaches. Source Multi-Variants Ratings Votes Likes Polarity Scores Tf-Idf Fuzzy Logic Multi-Data Sources [13] × √ × √ × × × × [17] × √ × √ √ × √ × [18] × × × × √ × × × [19] × × × × × × × × Multi-variant approach √ √ √ √ √ √ √ √ http://linkedmdb.org Sustainability 2018, 10, 4280 3 of 21 During the literature review of modern recommender systems, Hsieh’s movie recommender system [13] was identified as the relevant. This system has used the benefits of big data solutions and also provides a mobile app to interact with the recommender system. However, the key short-coming in this work is the limited approach used to recommend movies. Major issues with this recommendation system are discussed and comparison with our approach is given in Table 2. Table 2. Difference to Hsieh’s work [13]. # Hsieh’s Work Our Approach 1 It is a general recommendation system for movies. Our approach supports topic vise recommendation of movies such as drama, comedy, action, horror, etc. 2 This approach only uses quantitative data (ratings and likes) for recommendation that provides less accuracy. . Our approach uses both quantitative (votes, likes, etc.) and qualitative data (polarity score) for true recommendation of movies. 3 This approach banks on simplistic calculation in the base of similarity measures. No real decision making approach is used that makes quality of results questionable. Our approach uses Fuzzy Logic approach for better decision making and true recommendations of movies. 4 True likeness of users is not reflected by this approach. Our approach reflects true likeness of the users as qualitative aspects of likeness is also considered. 5 This approach is tested only on one limited dataset. Our approach is tested on three large datasets. 6 This approach calculates recommendation on two quantitative variables. Our approach calculates recommendation on three quantitative variables and one qualitative variable. The recommender systems field has made significant progress with many new techniques proposed and new systems developed. However, modern systems still require significant improvements to provide better recommendations. The major contribution to knowledge and novelty of the work is outlined below: i. A topic (action, comedy, horror, etc.) based recommendation is supported. ii. Multi-variant (ratings, votes, likes and polarity score) parameters are used. iii. Both quantitative and qualitative data is used for movie recommendations. iv. Three external data resources are used for datasets (Metacritics, IMDB, and Fandango). v. A web-bot used to fetch the web contents collaborates with the server. vi. Filters and integrates movie descriptions from linked data or ontology (linkedmdb). vii. Recommender system is developed using NoSQL environment with apache Hadoop. viii. Fuzzy sets are established for movie ranking categorization. ix. Front end App collaboration with the movie recommender through web services is supported. The rest of this paper is organized as follows. In Section 2 related work is discussed, where recommender systems running for different subjects could clear up the native problems of user’s data processing. A multi-variant ranking model is presented, for movie recommender using a mobile application, and Apache Hadoop in Section 3. Experiments and results, and evaluation of the system are discussed in Sections 4 and 5, respectively. The conclusion and future works are presented in Sections 6 and 7, respectively. 2. Related Work Recommendation services typically rely on customer reviews or customer ratings and such recommendations can provide a useful service for new customers. Emotional expressions, Sustainability 2018, 10, 4280 4 of 21 social interaction and behavior changes of the users are studied on the Twitter, allowing management to distinguish clients who do or do not return [20]. Vox Civitas obtains responses from social media (e.g., Twitter), which can support journalistic investigations in more effective ways [21]. Anomalies are removed and pure data is obtained which reflect the United Kingdom’s worst influenza [22]. Detection of noise in text (tweets) from micro-blogs is discussed in [23]. Sentiment analysis approaches can be used to extract sentiments associated with positive or negative polarities for specific subjects from a document, instead of classifying the whole document as positive or negative [24]. An NLP-based methodology of sentiment evaluation on user’s comment has been used as a way to retrieve the best and perfect YouTube videos. The process works in four steps. First, a review collection and preprocessing component extracts data (comments) from the particular YouTube video and language preprocessing is undertaken to prepare for the next process. Second, the processed text goes through NLP-based methods to generate data sets. Subsequently, the sentiment classifier (Sentistrength) is applied on the data sets to calculate the positivity and negativity ratings. Finally, the standard deviation applied to get the rating result [25]. Features level sentiment analysis, which is based on the idea that an opinion consists of a sentiment (positive or negative) and a feature of movies is another approach. Each short comment is represented as a sequence of sentiment words and underlying states [9,12]. A linear regression model (LRM), a supervised machine learning technique to classify twitter gossip (positive and negative) has been used to predict the box-office revenue for different movies [8]. Neural networks (NNs) classification of sentiment analysis of large movie reviews has been handled by introducing a method. Recursive neural networks wrap the previous sentence-level-sentiment classification and are used with recurrent neural networks. Recursive neural networks are used for sentence-level analysis and a recurrent neural network is used for whole passage analysis to create better results [26]. The vector space model (VSM) was used to implement the instance-based learning (IBL) classification method. Text documents were treated as vectors in IBL algorithms to identify the class (positive or negative review) of the document [27]. Sentiment analysis is negative when the text includes some negative words, such as “bad acting, stilted dialog.” It is positive if the text includes some positive words such as “it’s funny”. A suggestion instead of an exact rating is done by sentiment classification of the comments (polarity), and then aggregated into a rating score selected as the recommended list of popular movies [11,17]. For example, the hotel management of Starwood Hotels and Resorts use social media’s strength to stay connected with their guests, to guide them, and seek responses to the services they provide. [28]. Other related work includes typical recommender frameworks that construct calculations in light of different fuzzy set theoretic likeness measures (the fuzzy set augmentations of the Jaccard list, cosine, closeness or relationship similitude measures), and aggregation techniques for figuring suggestion certainty scores (the maximum-minimum or weighted-whole fuzzy set theoretic accumulation strategies) for recommendation [4]. The strategy for ranking in light of the content involves building a sentiment graph from the collocation of adjectives, PageRank algorithm and a very small set of adjectives (such as ‘good’, ‘excellent’, etc.) that rank different movies using reviews of box office movies by users of a popular movie review site [18]. With regard to the utilization of labels with the end goal of recommendation of movies, the German movie website, Moviepilot uses viewers, and movie ratings, and all out labels are marked to every movie. Labels are allotted by a group of moderators and viewer are then able to rate how well the labels fit every motion picture [3]. This collaborative filtering was first applied elsewhere in filtering the information in Usenet news [29]. Music recommender systems provide personalized music recommendations and Ringo Agent was one of the first applications [30]. Content-based filtering recommends movies based on a comparison between user profile data and content of movies. Content-based filtering is also called cognitive-filtering. The recommendations are generated by matching users and movie content [4]. Collaborative filtering is also called social filtering. The fundamental rule behind collaborative filtering is that if a user likes a certain category of movie in the past, then they may like similar movies in the Sustainability 2018, 10, 4280 5 of 21 future. This information is used in deciding which movie to suggest [19,29,30]. Hybrid filtering is a combined technique of content filtering and collaborative filtering [31]. The previous work discussed above suggests that most of the approaches used for recommendation services, especially for movie recommendation are uni-variant and use varaiables such as ratings, which tend to provide results with low accuracy. There are other variables including likes, number of reviews, and the sentiment score of a review that can help in achieving an accurate and efficient recommendation, and in this paper we aim to use these new variables for the proposed movie recommendation service. 3. Multi-Variant Expert System The proposed approach works on the fetched data (scores and reviews) from a set of movie websites and databases. The collected data is heterogeneous in nature, such as numeric data (for example, number of votes, number of likes and number of ratings) and text data (for example, user reviews of movies). A web crawler was developed to fetch structured and unstructured data and store the fetched data in a NoSQL database on a server machine for further processing. The used approach works in two parallel streams. In the first stream, the text data (such as movie reviews) is preprocessed using NLP modules, tf-idf algorithms and the SentiWordNet auxiliary database to identify polarity scores of terms (lexicons) in the form of negative and positive scores. All the movies reviews are processed for an aggregate polarity score for each movie from each participating external data source. In the second stream, all the numeric scores and weights (rating, votes and likes) of the movies are normalized and computed to achieve weighted aggregate of polarity scores. The result of a search query of the movies is shown in the user interface of an Android app (see Figure A1). The user query interacts with the server and the server processes the request by forwarding it to the web crawler. The web crawler module responds to the server’s request by crawling the web for keywords (lexicons) matching and downloading the webpages to the server, and then the server processes the data to generate a recommendation as shown in Figure 1. Sustainability 2018, 10, x FOR PEER REVIEW 5 of 21 The previous work discussed above suggests that most of the approaches used for recommendation services, especially for movie recommendation are uni-variant and use varaiables such as ratings, which tend to provide results with low accuracy. There are other variables including likes, number of reviews, and the sentiment score of a review that can help in achieving an accurate and efficient recommendation, and in this paper we aim to use these new variables for the proposed movie recommendation service. 3. Multi-Variant Expert System The proposed approach works on the fetched data (scores and reviews) from a set of movie websites and databases. The collected data is heterogeneous in nature, such as numeric data (for example, number of votes, number of likes and number of ratings) and text data (for example, user reviews of movies). A web crawler was developed to fetch structured and unstructured data and store the fetched data in a NoSQL database on a server machine for further processing. The used approach works in two parallel streams. In the first stream, the text data (such as movie reviews) is preprocessed using NLP modules, tf-idf algorithms and the SentiWordNet auxiliary database to identify polarity scores of terms (lexicons) in the form of negative and positive scores. All the movies reviews are processed for an aggregate polarity score for each movie from each participating external data source. In the second stream, all the numeric scores and weights (rating, votes and likes) of the movies are normalized and computed to achieve weighted aggregate of polarity scores. The result of a search query of the movies is shown in the user interface of an Android app (see Figure A1). The user query interacts with the server and the server processes the request by forwarding it to the web crawler. The web crawler module responds to the server’s request by crawling the web for keywords (lexicons) matching and downloading the webpages to the server, and then the server processes the data to generate a recommendation as shown in Figure 1. Figure 1. Multi-variant expert system for movie recommendation. Figure 1. Multi-variant expert system for movie recommendation. Sustainability 2018, 10, 4280 6 of 21 3.1. NLP Module Real-world data is generally incomplete and noisy, and is likely to contain irrelevant and redundant information or errors. By pre-processing, raw unstructured data can be converted into a structured, understandable form as shown in Figure 2. Since the real-world data can contain ambiguity and anomalies, it is necessary to remove these abnormalities before the actual analysis of data. The data was pre-processed to remove anomalies and to identify the abbreviated language of the actual English and also remove the reviews which were in other languages other than English [32–34]. Sustainability 2018, 10, x FOR PEER REVIEW 6 of 21 3.1. NLP Module Real-world data is generally incomplete and noisy, and is likely to contain irrelevant and redundant information or errors. By pre-processing, raw unstructured data can be converted into a structured, understandable form as shown in Figure 2. Since the real-world data can contain ambiguity and anomalies, it is necessary to remove these abnormalities before the actual analysis of data. The data was pre-processed to remove anomalies and to identify the abbreviated language of the actual English and also remove the reviews which were in other languages other than English [32–34]. Figure 2. Preprocessing data using NLP module. 3.1.1. Tokenization Then, the given character stream is separated into units called tokens. The tokens might be words or numbers or highlighting check. Tokenization does this by finding word limits. For example, “The message of this film is simple” string is tokenized as [The] [message] [of] [this] [film] [is] [simple] [35,36]. 3.1.2. Stemming (Lemmatization) This is optional; most stemming usesthe Porter Stemmer. English words like “look” can be arched with a morphological suffix to deliver “looks, looking, looked”. These have a similar stem, “look” [37,38]. 3.1.3. Stop Word Evacuation Most regularly used words do not convey much significance.For example: “the, an, of, for, in ...”. We used a small corpus based library to exclude stp words from the input data. This library is developed in Java. 3.1.4. POS-Tag Generation The query was then analyzed and POS tags were generated of all the words in the query. Then, the resulting string of words and their relevant POS (Parts of Speech) tags are tokenized on the basis of space. A good example of an English sentence, “This movie is so riddled”, is Pos-tagged as [this/DT movie/NN is/VBZ so/RB riddled/JJ]. The Treebank Project (URL: https://catalog.ldc.upenn.edu/docs/LDC95T7/treebank2.index.html) shows 36 POS-tags [39], e.g., determiner [DT], adjective [JJ] and adverb [RB], etc. 3.2. Polarity Computation 3.2.1. Lexical Frequency Measuring For this purpose, we applied a tf-idf frequency measure [40,41]; we first calculated the frequency of the valid/important terms, which represents the number of times that term ‘t’ occurs in the document (review) ‘d’, as in the following Equation (1): tf(t, d) = f(t, d) (1) Figure 2. Preprocessing data using NLP module. 3.1.1. Tokenization Then, the given character stream is separated into units called tokens. The tokens might be words or numbers or highlighting check. Tokenization does this by finding word limits. For example, “The message of this film is simple” string is tokenized as [The] [message] [of] [this] [film] [is] [simple] [35,36]. 3.1.2. Stemming (Lemmatization) This is optional; most stemming usesthe Porter Stemmer. English words like “look” can be arched with a morphological suffix to deliver “looks, looking, looked”. These have a similar stem, “look” [37,38]. 3.1.3. Stop Word Evacuation Most regularly used words do not convey much significance.For example: “the, an, of, for, in ...”. We used a small corpus based library to exclude stp words from the input data. This library is developed in Java. 3.1.4. POS-Tag Generation The query was then analyzed and POS tags were generated of all the words in the query. Then, the resulting string of words and their relevant POS (Parts of Speech) tags are tokenized on the basis of space. A good example of an English sentence, “This movie is so riddled”, is Pos-tagged as [this/DT movie/NN is/VBZ so/RB riddled/JJ]. The Treebank Project (URL: https://catalog.ldc. upenn.edu/docs/LDC95T7/treebank2.index.html) shows 36 POS-tags [39], e.g., determiner [DT], adjective [JJ] and adverb [RB], etc. 3.2. Polarity Computation 3.2.1. Lexical Frequency Measuring For this purpose, we applied a tf-idf frequency measure [40,41]; we first calculated the frequency of the valid/important terms, which represents the number of times that term ‘t’ occurs in the document (review) ‘d’, as in the following Equation (1): tf(t, d) = f(t, d) (1) https://catalog.ldc.upenn.edu/docs/LDC95T7/treebank2.index.html https://catalog.ldc.upenn.edu/docs/LDC95T7/treebank2.index.html Sustainability 2018, 10, 4280 7 of 21 After calculating tf we calculated idf (inverse document frequency) of the terms to obtain information about how rare or common that term is in the documents (reviews). We used the Equation (2): idf(t, D) = logN/dt (2) where as d ∈ D and t ∈ d, N is total number of documents in the corpus N = |D|. The end results are then obtained by applying the Equation (3): tfidf(t, d, D) = tf(t, d)∗ idf(t, D) (3) 3.2.2. Polarity Identification SentiWordNet 3.0 automatically annotates all WordNet 2.0 (synsets) according to their degrees of positivity, negativity and neutrality. In this step, the SentiWordNet score was used in the sentiment analysis of the documents (reviews) [42,43]. For this purpose, we applied the Equation (4): Polarity_term_score = SentiWordNetScore ∗ Frequency(tfidf) (4) The SentiWordNetScore (positive or negative) of the term and its frequency were computed to get the overall sentiment of the terms in the documents. The SentiWordNetScore of each term for all reviews of the movie is calculated and the score (negative or positive) tells us how many terms are positively or negatively important in the review. Then, all the positive terms scores are added to obtain the positive term’s weight, and also all the negative terms scores are combined to obtain the negative term’s weight in a review. The polarities of all the reviews of the movies from each participating website are calculated as follows. Polarity of a Term By applying sign P(t) on term, if SentiWordNetScore (Sti ) of terms (ti) of the review (ri) of movie (mi) is less than zero then the term lies in the negative poll (nt), if greater than zero than it lies in the positive poll (pt) and if it is equal to zero than it lies in the neutral poll (tn) as shown in equation (5). P(t) =   −Sti , Sti < 0 ( nt = negative term ) Sti , Sti = 0 (tn = neutral term) +Sti , Sti > 0 (pt = positive term ) (5) Polarity of a Document (Reviews) For calculating the polarity of a document (review) polarityri (r), positive terms ptri and negative terms ntri are aggregated for each document (review) from each participating websites’ negative_termri (x) and negative_termri (y) and then take their differences are taken to find the polarity of each documents (reviews) by applying the sign function f(r) as shown in Equations (6) and (7). positive_termri (x) = n ∑ i=0 ptri (6) negative_termri (y) = n ∑ i=0 ntri (7) where as ptri ∧ ntri ∈ ri and ri ∈ mj Polarity of a review is calculated as shown in Equations (8) and (9): polarityri (R) = sgn [ |xri|− ∣∣∣yri∣∣∣ ] (8) Sustainability 2018, 10, 4280 8 of 21 f(r) =   −pri , pri < 0 ( nr = negative review ) pri , pri = 0 (rn = neutral review) +pri , pri > 0 (pr = positive review) (9) by applying sign f(r) on each review, If the difference of aggregated positive_termri (x) of review (ri) of the movie (mj) from website (wk) and aggregated negative_termri (y) is less than zero then review is sentimentally lie in negative poll (nr), if greater than zero than lie in positive poll (pr) and if equal to zero than lie in neutral poll (rn). Polarity of a Collection (Movie Reviews) review_positive_scoremj (a) is the aggregated polarity score of positive reviews pr and review_negative_scoremj (b) is the aggregated polarity score of negative reviews nr of a particular movie mj from a particular website (wk) used to calculated the polarity of the movie from participating sites as given in Equations (10) and (11). review_positive_scoremj (g) = n ∑ i=0 pri (10) review_negative_scoremj (h) = n ∑ i=0 nri (11) where as pri ∧ nri ∈ mj and mj ∈ wk. Polarity of a collection is calculated using Equation (12): polarityri (p) = sgn [ ∣∣∣gmj∣∣∣− ∣∣∣hmj∣∣∣ ] (12) where as gmj ∧ hmj ∈ mj and mj ∈ wk. Here wk is a movie website such as IMDB. 3.3. Weighted Polarity Manipulation Opinion mining determines the emotions (positive or negative) of textual communication on social media, and examines the positive or negative emotions by simply extracting polarity scores from the review (number of stars or thumbs up/down and votes etc.). However, we used both the polarity score and weight score (rating, votes and likes) of the movies. First, we computed the aggregated polarity score of each movie from each participating site, and then we took the average of the aggregated polarity by total reviews of the respective movie and their site. Again, we take aggregation of average polarity score. Also, total likes of the movie were combined with weighted_average_polarity to find the aggregated_weighted_average_polarity. After that, the final score of the movie was rescaled to get the ranked score and category of the movie. In this computation, Equations (13)–(18) are used. aggregate_polarity_Scoremj (g) = n ∑ i=1 pi (13) whereas gmj ∈ mi , mi ∈ wk. weightmj = [ (voteswk ) + ( ratingwk )] (14) whereas weightmj ∈ mi , mi ∈ wk. weighted_average_polaritymj (a) = gmj n + ( weightmj ) (15) Sustainability 2018, 10, 4280 9 of 21 whereas n is number of reviews of movie (mj) from movie website (wk). aggregated_weigted_average_polaritymj (G) = N ∑ k=1 awk + likesmj (16) average_aggregated_weighted_average_polaritymj (A) = Gmj N (17) where as awk ∈ mj , mj ∈ wk , wk ∈ N. (N) is the number of movie websites (Metacritic, IMDBand Fandango) which has huge collection of material. Rank_scoremj (R) = rmj − min(rmj ) max(rmj )− min(rmj ) ∗ 10 (18) Here R is the rescale value of the normalized average aggregated score (rmj ) of the movie (mi) from movie websites (N) to rank the top five movies (M). 3.4. Categorization Movie genres are various forms or identifiable types, categories, classifications or groups of movies (genre comes from the French word meaning “kind”, “category”, “or “type”). http://www. filmsite.org/filmgenres.html. The user can directly query for content (e.g., “news London”, “golf 1940”, “documentary Alfred Hitchcock”, “movies tonight”). The querying is done on different fields describing the content (e.g., title, creator, year, genre, language, location). The search results are presented as organized according to concepts from common vocabularies (e.g., Time Ontology, Geo Ontology, WordNet) [29,44–46]. 3.4.1. WordNet The WordNet library was used in our approach to find synonyms and alternative forms of query terms, e.g., “weather” = {“weather report”, “weather forecast”, etc.}. Identification of synonyms in data helps in obtaining better and more accurate results. 3.4.2. Geo Ontology Finds related geographical areas, e.g., “London” = {City of London, Camden, Westminster, Greenwich, Greater London, England, and UK}. 3.4.3. Time Ontology Determines temporal context e.g., “tonight” = {18:00–24:00}, or “this week” = {10/02–17/02}. 3.4.4. TVA-CS Finds related genres (e.g., “sports” = {sport reports, sport live, sport news, sport documentary, football game, etc.} Documentary, football game, etc.). 3.5. Recommendation The final recommendation is achieved by using the fuzzy logic approach on the following fuzzy set to evaluate the final score and find the category of the movie as follows. Step 1 if final score ≥ 8 then Category: “A: Recommended” Step 2 else if final score ≥ 6 then Category: “B: Top Recommended” Step 3 else if final Score ≥ 4 then Category: “C: Recommended Average” http://www.filmsite.org/filmgenres.html http://www.filmsite.org/filmgenres.html Sustainability 2018, 10, 4280 10 of 21 Step 4 else if Final Score ≥ 2 then Category: “D: Least recommended” Step 6 else Category: “F: Not recommended” Figure 3 shows the final recommendations of the movies in a particular category (such as comedy, horror, fiction, etc.) in one of the five different classes. The user interface showing the output is discussed in Appendix A and Figure A1. Sustainability 2018, 10, x FOR PEER REVIEW 10 of 21 Category: “B: Top Recommended” Step 3 else if final Score ≥ 4 then Category: “C: Recommended Average” Step 4 else if Final Score ≥ 2 then Category: “D: Least recommended” Step 6 else Category: “F: Not recommended” Figure 3 shows the final recommendations of the movies in a particular category (such as comedy, horror, fiction, etc.) in one of the five different classes. The user interface showing the output is discussed in Appendix A and Figure A1. Figure 3. Multi-variant ranked category. 4. Experimental Setup 4.1. NoSQL for Big Data Stroage The number of publicly available test corpora is quite limited and comparatively of small size with respect to the number of texts documents in a corpus. Thus, producing adequately precise comparisons between reported performances is difficult. So, we decided to build a new corpus and for this purpose, we used three different external data source websites to extract a large number of reviews, votes, ranking and likes. The data for our corpus was retrieved by a web bot implemented in PHP. We wrote a webpage (web-bot) scraping scripts which extract movie URLs with matching user’s queries, if the query keywords (lexicons) are matched then the crawler downloads the page in a server machine NoSQL environment using Hadoop, otherwise this page is discarded [13,14,47–52]. This procedure is depicted in Figure 4. The process of data extraction uses the following steps. Step 1 Receive URLs for a movie type i.e., comedy, horror, fiction, etc. Step 2 Matches the keywords from the query to the page If Step 3 Keywords matched Then Step 4 Download the web page Step 5 Send it for storage Step 6 Discard the page Step 7 Repeat the step 2 to 6 until all the matched web pages are found. Figure 3. Multi-variant ranked category. 4. Experimental Setup 4.1. NoSQL for Big Data Stroage The number of publicly available test corpora is quite limited and comparatively of small size with respect to the number of texts documents in a corpus. Thus, producing adequately precise comparisons between reported performances is difficult. So, we decided to build a new corpus and for this purpose, we used three different external data source websites to extract a large number of reviews, votes, ranking and likes. The data for our corpus was retrieved by a web bot implemented in PHP. We wrote a webpage (web-bot) scraping scripts which extract movie URLs with matching user’s queries, if the query keywords (lexicons) are matched then the crawler downloads the page in a server machine NoSQL environment using Hadoop, otherwise this page is discarded [13,14,47–52]. This procedure is depicted in Figure 4. The process of data extraction uses the following steps.Sustainability 2018, 10, x FOR PEER REVIEW 11 of 21 Figure 4. Data collection process from websites using Web Crawler. Web crawler (web-bot) downloads the webpages (crawled pages) by which it extracts more contents (Meta tags) like movie reviews, rating, votes and likes and other irrelevant pages discarded as shown in Figure 4. Computational processing can occur on data stored either in a file-system (unstructured) or in a database (structured) as shown in Figure 5. Apache Hadoop is the de facto data operating system. It is an open-source software framework processing of big data on clusters of commodity hardware. Figure 5. Multi-variant recommendation system. A multi-variant web agent is implemented in the Hadoop environment to handle big data generated by recommendation systems in order to improve the scalability and efficiency. The above- mentioned Figure 5 shows the interaction and computation in the Hadoop environment between the Android user app, web bot and the external participating sites for data. In each Mapper, there are various algorithms porterStemmer(), tokenizer(), POStager(), polarityComputation(), weightedRanking(), Webcrawler(), etc. and in Table A1, the Server machine’s specification and Android device specification are presented. The hardware used in implementation is discussed in Appendix B and mentioned in Table A1. 4.2. Experiment and Results We used three repositories whose reviews should be trusted, that is, IMDB, Metacritic and Fandango, which contain all the required data (reviews and scores). These repositories contain movie Figure 4. Data collection process from websites using Web Crawler. Sustainability 2018, 10, 4280 11 of 21 Step 1 Receive URLs for a movie type i.e., comedy, horror, fiction, etc. Step 2 Matches the keywords from the query to the page If Step 3 Keywords matched Then Step 4 Download the web page Step 5 Send it for storage Step 6 Discard the page Step 7 Repeat the step 2 to 6 until all the matched web pages are found. Web crawler (web-bot) downloads the webpages (crawled pages) by which it extracts more contents (Meta tags) like movie reviews, rating, votes and likes and other irrelevant pages discarded as shown in Figure 4. Computational processing can occur on data stored either in a file-system (unstructured) or in a database (structured) as shown in Figure 5. Apache Hadoop is the de facto data operating system. It is an open-source software framework processing of big data on clusters of commodity hardware. Sustainability 2018, 10, x FOR PEER REVIEW 11 of 21 Figure 4. Data collection process from websites using Web Crawler. Web crawler (web-bot) downloads the webpages (crawled pages) by which it extracts more contents (Meta tags) like movie reviews, rating, votes and likes and other irrelevant pages discarded as shown in Figure 4. Computational processing can occur on data stored either in a file-system (unstructured) or in a database (structured) as shown in Figure 5. Apache Hadoop is the de facto data operating system. It is an open-source software framework processing of big data on clusters of commodity hardware. Figure 5. Multi-variant recommendation system. A multi-variant web agent is implemented in the Hadoop environment to handle big data generated by recommendation systems in order to improve the scalability and efficiency. The above- mentioned Figure 5 shows the interaction and computation in the Hadoop environment between the Android user app, web bot and the external participating sites for data. In each Mapper, there are various algorithms porterStemmer(), tokenizer(), POStager(), polarityComputation(), weightedRanking(), Webcrawler(), etc. and in Table A1, the Server machine’s specification and Android device specification are presented. The hardware used in implementation is discussed in Appendix B and mentioned in Table A1. 4.2. Experiment and Results We used three repositories whose reviews should be trusted, that is, IMDB, Metacritic and Fandango, which contain all the required data (reviews and scores). These repositories contain movie Figure 5. Multi-variant recommendation system. A multi-variant web agent is implemented in the Hadoop environment to handle big data generated by recommendation systems in order to improve the scalability and efficiency. The above-mentioned Figure 5 shows the interaction and computation in the Hadoop environment between the Android user app, web bot and the external participating sites for data. In each Mapper, there are various algorithms porterStemmer(), tokenizer(), POStager(), polarityComputation(), weightedRanking(), Webcrawler(), etc. and in Table A1, the Server machine’s specification and Android device specification are presented. The hardware used in implementation is discussed in Appendix B and mentioned in Table A1. 4.2. Experiment and Results We used three repositories whose reviews should be trusted, that is, IMDB, Metacritic and Fandango, which contain all the required data (reviews and scores). These repositories contain movie (2016 and 2017) data for 1000 of the most popular movies (with a significant number of votes and ratings) and their reviews were released in 2016 and 2017, and as of 22 March 2017. We computed the polarity score of the text data (reviews) by computing the movie’s reviews corpus which was fetched from each participating external data source sites. This procedure was followed by data preprocessing, tf-idf classification and polarity identification using SentiWordNet to compute the polarity scores of each term of each document from each participating data source sites. The following tables illustrate the values of movies fetched data which were processed and evaluated. Here, Table 3 shows the movie’s title and the corresponding movie ID as follows. Sustainability 2018, 10, 4280 12 of 21 Table 3. Movie Id of movies used in the experiments. Movie Title Movie ID Avengers: Age of Ultron (2015) m1 Cinderella (2015) m2 Ant-Man (2015) m3 Do You Believe? (2015) m4 Hot Tub Time Machine 2 (2015) m5 Table 4 shows the popular external data source sites and their corresponding movies sites ID for better formulation. Table 4. Movie database sites ID. Movie Database Site Movie Database Sine ID Metacritic w1 IMDB w2 Fandango w3 Some computed values such as polarity scores of movie reviews from participating sites, which are already labeled are shown in Table 5. Table 5. Calculated polarity cores. Polarity Scores Movie ID w1 w2 w3 m1 21 29 3 m2 −17 16 26 m3 2 2 4 m4 −2 2 13 m5 1 19 7 Here we normalized the scores “likes” metascore and IMDB rating scores are normalized ratings to a (0–5) scale because Metacritic and IMDB user rating is out of ten stars, but Fandango rating is out of five stars so we normalized the Metacritic and IMDB to five stars. These normalized values are given in the Table 6. Table 6. Normalized and un-normalized rating. Un-Normalized Rating Scores Normalized Rating Scores Movie ID w1 w2 w3 w1 w2 w3 m1 7.1 7.8 5 3.55 3.9 5 m2 7.5 7.1 5 3.75 3.55 5 m3 8.1 7.8 5 4.05 3.9 5 m4 4.7 5.4 5 2.35 2.7 5 m5 3.4 5.1 3.5 1.7 2.55 3.5 Here Metacritic_votes, IMDB_votes and Fandango_votes are the number of the votes, which are allotted by users to the particular movies from specific movie websites and are represented in the Table 7. After calculating and aggregating the polarity score from each participating movie site and taking an average of polarity scores by total movies, and taking the weighted average polarity by adding the weights (normalized ranking and votes) to each average polarity, Likes may also represent the Sustainability 2018, 10, 4280 13 of 21 behavior of users, which impact the movie rating. That is why we selected the likes in our model to present the multi-variant approach. We added Facebook likes to them to take the aggregated weighted average polarity for better recommendations. The classified scores are shown in Table 8. Table 7. Movie votes. Movie Votes Movie ID w1 w2 w3 m1 1330 271,107 14,846 m2 249 65,709 12,640 m3 627 103,660 12,055 m4 31 3136 1793 m5 88 19,560 1021 These multi variants (votes, ranking and likes) are computed according to our model, which ranked the movies as mentioned from the corpora of different movie data, the final score and category are represented in Table 9. The multi-variant score of m1 movie “Avengers: Age of Ultron” is greater than six which is why it is categorized “B”, the m2 movie “Cinderella” score is greater than four so it lies in Category “C”, m3 and m4 “Ant-Man” and “Do You Believe?”, respectively, are greater than six so these are categorized “B”, and the m5 “Hot Tub Time Machine 2” movie score is greater than four so it is also lies in “C” category. Figure 6 represents the final ranking score of a movie’s particular category such as fiction, horror, drama, etc. Figure 7 shows the rating scores of a particular movie from three specific movie websites We compared their normalized rating scores among these sites, and we observed that rating by Fandango is so high that is not significant individually. Sustainability 2018, 10, x FOR PEER REVIEW 15 of 21 These multi variants (votes, ranking and likes) are computed according to our model, which ranked the movies as mentioned from the corpora of different movie data, the final score and category are represented in Table 9. Table 9. Final scores and movie category. Movie ID Likes Aggregated Weighted Average Polarity Final Ranking Score Movie Category m1 308,130 403,895.2977 7.44944873 B m2 331 26,534.68444 5.878236708 C m3 140,000 178,785.0504 6.979146632 B m4 97,000 98,656.85966 6.636052698 B m5 14,000 20,892.44087 5.740277373 C The multi-variant score of m1 movie “Avengers: Age of Ultron” is greater than six which is why it is categorized “B”, the m2 movie “Cinderella” score is greater than four so it lies in Category “C”, m3 and m4 “Ant-Man” and “Do You Believe?”, respectively, are greater than six so these are categorized “B”, and the m5 “Hot Tub Time Machine 2” movie score is greater than four so it is also lies in “C” category. Figure 6 represents the final ranking score of a movie’s particular category such as fiction, horror, drama, etc. Figure 6. Multi-variant movie ranking. Figure 7 shows the rating scores of a particular movie from three specific movie websites We compared their normalized rating scores among these sites, and we observed that rating by Fandango is so high that is not significant individually. Figure 7. Rating difference. In Figure 8, voting for movies also represents the user’s interest in movies from different participating websites, which indicates a huge difference if we select only one site. One site is not adequate for a ranking approach, which is the reason we selected multi-variants from different sites. Figure 6. Multi-variant movie ranking. Sustainability 2018, 10, x FOR PEER REVIEW 15 of 21 These multi variants (votes, ranking and likes) are computed according to our model, which ranked the movies as mentioned from the corpora of different movie data, the final score and category are represented in Table 9. Table 9. Final scores and movie category. Movie ID Likes Aggregated Weighted Average Polarity Final Ranking Score Movie Category m1 308,130 403,895.2977 7.44944873 B m2 331 26,534.68444 5.878236708 C m3 140,000 178,785.0504 6.979146632 B m4 97,000 98,656.85966 6.636052698 B m5 14,000 20,892.44087 5.740277373 C The multi-variant score of m1 movie “Avengers: Age of Ultron” is greater than six which is why it is categorized “B”, the m2 movie “Cinderella” score is greater than four so it lies in Category “C”, m3 and m4 “Ant-Man” and “Do You Believe?”, respectively, are greater than six so these are categorized “B”, and the m5 “Hot Tub Time Machine 2” movie score is greater than four so it is also lies in “C” category. Figure 6 represents the final ranking score of a movie’s particular category such as fiction, horror, drama, etc. Figure 6. Multi-variant movie ranking. Figure 7 shows the rating scores of a particular movie from three specific movie websites We compared their normalized rating scores among these sites, and we observed that rating by Fandango is so high that is not significant individually. Figure 7. Rating difference. In Figure 8, voting for movies also represents the user’s interest in movies from different participating websites, which indicates a huge difference if we select only one site. One site is not adequate for a ranking approach, which is the reason we selected multi-variants from different sites. Figure 7. Rating difference. Sustainability 2018, 10, 4280 14 of 21 Table 8. Weighted average polarity scores. Reviews Aggregated Polarity Average Polarity Weighted Average Polarity w1 w2 w3 w1 w2 w3 w1 w2 w3 w1 w2 w3 Movie ID Reviews Reviews Reviews Aggre. Polarity Aggre. Polarity Aggre. Polarity Average Polarity Average Polarity Average Polarity Weighted Average Polarity Weighted Average Polarity Weighted Average Polarity m1 66 1168 30 21 29 3 0.318 0.0248 0.1 1333.86 271,110 14,851.1 m2 67 363 27 −17 16 26 0.253 0.0440 0.962 252.49 65,712.5 12,645.9 m3 64 605 24 2 2 4 0.031 0.003 0.166 631.08 103,663 12,060.1 m4 22 69 22 −2 2 13 0.090 0.0289 0.590 33.25 3138.72 1798.59 m5 29 101 20 1 19 7 0.034 0.1881 0.35 89.73 19,562.7 1024.85 Sustainability 2018, 10, 4280 15 of 21 Table 9. Final scores and movie category. Movie ID Likes Aggregated Weighted Average Polarity Final Ranking Score Movie Category m1 308,130 403,895.2977 7.44944873 B m2 331 26,534.68444 5.878236708 C m3 140,000 178,785.0504 6.979146632 B m4 97,000 98,656.85966 6.636052698 B m5 14,000 20,892.44087 5.740277373 C In Figure 8, voting for movies also represents the user’s interest in movies from different participating websites, which indicates a huge difference if we select only one site. One site is not adequate for a ranking approach, which is the reason we selected multi-variants from different sites. Sustainability 2018, 10, x FOR PEER REVIEW 16 of 21 Figure 8. Movie votes difference. Here, Figure 9 represents the differences in weighted average polarity scores by computing the multi-variant to show the categories. Figure 9. Comparison scores of movies. Time complexity was computed and also the watched time at different machines was observed. Time complexity in the worst case of our approach is O(n) because n number of datasets are used, and all following operation take one unit time, so time complexity of following is operations O(1). The computation watched time details are presented in Table 10. Table 10. Computational watched time at different machines. CPU Clock Move Add. Sub. Mul. Div. Comp. Speed TMS320C30 (16.67 MHz) 22 3n 4 3 4 19 6 ms MC68000 (16 MHz) 5 20 20 70 160 15 59 ms 22 3n 4 3 4 19 59 ms Z80 (8 MHz) 22 3n 4 3 4 19 280 ms 5. Evaluation Recall is defined as the number of relevant movies retrieved by a search divided by the total number of existing relevant movies, while precision is defined as the number of relevant movies retrieved by a search divided by the total number of movies retrieved by the search. The precision is the proportion of recommendations that are good recommendations, Precision = tp/(tp + fp) (19) and recall is the proportion of good recommendations that appear in top recommendations. Recall = tp/(tp + fn) (20) tp: predicted positive interested movie it is true, it is really interested. 0.00 5,000.00 10,000.00 15,000.00 20,000.00 25,000.00 30,000.00 w1 w2 w3 Movie Id m1 m2 m3 m4 m5 Figure 8. Movie votes difference. Here, Figure 9 represents the differences in weighted average polarity scores by computing the multi-variant to show the categories. Sustainability 2018, 10, x FOR PEER REVIEW 16 of 21 Figure 8. Movie votes difference. Here, Figure 9 represents the differences in weighted average polarity scores by computing the multi-variant to show the categories. Figure 9. Comparison scores of movies. Time complexity was computed and also the watched time at different machines was observed. Time complexity in the worst case of our approach is O(n) because n number of datasets are used, and all following operation take one unit time, so time complexity of following is operations O(1). The computation watched time details are presented in Table 10. Table 10. Computational watched time at different machines. CPU Clock Move Add. Sub. Mul. Div. Comp. Speed TMS320C30 (16.67 MHz) 22 3n 4 3 4 19 6 ms MC68000 (16 MHz) 5 20 20 70 160 15 59 ms 22 3n 4 3 4 19 59 ms Z80 (8 MHz) 22 3n 4 3 4 19 280 ms 5. Evaluation Recall is defined as the number of relevant movies retrieved by a search divided by the total number of existing relevant movies, while precision is defined as the number of relevant movies retrieved by a search divided by the total number of movies retrieved by the search. The precision is the proportion of recommendations that are good recommendations, Precision = tp/(tp + fp) (19) and recall is the proportion of good recommendations that appear in top recommendations. Recall = tp/(tp + fn) (20) tp: predicted positive interested movie it is true, it is really interested. 0.00 5,000.00 10,000.00 15,000.00 20,000.00 25,000.00 30,000.00 w1 w2 w3 Movie Id m1 m2 m3 m4 m5 Figure 9. Comparison scores of movies. Time complexity was computed and also the watched time at different machines was observed. Time complexity in the worst case of our approach is O(n) because n number of datasets are used, and all following operation take one unit time, so time complexity of following is operations O(1). The computation watched time details are presented in Table 10. Sustainability 2018, 10, 4280 16 of 21 Table 10. Computational watched time at different machines. CPU Clock Move Add. Sub. Mul. Div. Comp. Speed TMS320C30 (16.67 MHz) 22 3n 4 3 4 19 6 ms MC68000 (16 MHz) 5 20 20 70 160 15 59 ms 22 3n 4 3 4 19 59 ms Z80 (8 MHz) 22 3n 4 3 4 19 280 ms 5. Evaluation Recall is defined as the number of relevant movies retrieved by a search divided by the total number of existing relevant movies, while precision is defined as the number of relevant movies retrieved by a search divided by the total number of movies retrieved by the search. The precision is the proportion of recommendations that are good recommendations, Precision = tp/(tp + fp) (19) and recall is the proportion of good recommendations that appear in top recommendations. Recall = tp/(tp + fn) (20) tp: predicted positive interested movie it is true, it is really interested. tn: predicted positive uninterested movie it is true, it’s really uninterested. fp: predicted positive interested movie but wrong, it is actually interesting. fn: predicted negative uninterested movie but wrong, it is actually uninteresting. In the recommendation domain, a perfect precision score of 1.0 means that every movie recommended in the list was good (although this says nothing about if all good recommendations were suggested) whereas a perfect recall score of 1.0 means that all good recommended movies were suggested in the list. Typically, when a recommender system is tuned to increase precision, recall decreases as a result (or vice versa). F-Score = 2. (precision.recall)/(precision + recall) (21) Table 11 shows some outcomes of our recommendation system. Table 11. Outcomes of multi-variant recommendation system. Evaluating Parameters TP TN FP FN Aggregated Polarity 220 356 278 146 Weighted Average Polarity 486 244 196 74 aggregated weighted Average Polarity 640 138 97 125 Final Ranking Score 953 11 16 5 In Table 12 and Figure 10 for ecommendations in this domain, a single value is obtained by combining both the precision and recall measures and indicates the overall utility of the recommendation list. One thousand movies were used as exemplary data sets. Evaluations are really important in the recommendation engine building process, which can be used to empirically discover improvements to a recommendation algorithm. This research used the MovieLens 1K dataset. There are 943 users and 1000 movies; we used the 1000 ratings, votes, likes and views from the users on the films to test the performance of proposed method. The results of the f-measure differentiated the accuracy of our work from others. If we use the multi-variants system it provided an accuracy of about 98.6%. Sustainability 2018, 10, 4280 17 of 21 Table 12. Weighted average polarity scores. Evaluating Parameters Aggregated Polarity Weighted Average Polarity Aggregated Weighted Average Polarity Multi-Variant Precision 0.3819 0.6658 0.8226 0.9886 Recall 0.4418 0.7126 0.8684 0.9835 F-measure 0.4097 0.6884 0.8449 0.9860 Sustainability 2018, 10, x FOR PEER REVIEW 17 of 21 tn: predicted positive uninterested movie it is true, it’s really uninterested. fp: predicted positive interested movie but wrong, it is actually interesting. fn: predicted negative uninterested movie but wrong, it is actually uninteresting. In the recommendation domain, a perfect precision score of 1.0 means that every movie recommended in the list was good (although this says nothing about if all good recommendations were suggested) whereas a perfect recall score of 1.0 means that all good recommended movies were suggested in the list. Typically, when a recommender system is tuned to increase precision, recall decreases as a result (or vice versa). F-Score = 2. (precision.recall)/(precision + recall) (21) Table 11 shows some outcomes of our recommendation system. Table 11. Outcomes of multi-variant recommendation system. Evaluating Parameters TP TN FP FN Aggregated Polarity 220 356 278 146 Weighted Average Polarity 486 244 196 74 aggregated weighted Average Polarity 640 138 97 125 Final Ranking Score 953 11 16 5 In Table 12 and Figure 10 for ecommendations in this domain, a single value is obtained by combining both the precision and recall measures and indicates the overall utility of the recommendation list. One thousand movies were used as exemplary data sets. Evaluations are really important in the recommendation engine building process, which can be used to empirically discover improvements to a recommendation algorithm. This research used the MovieLens 1K dataset. There are 943 users and 1000 movies; we used the 1000 ratings, votes, likes and views from the users on the films to test the performance of proposed method. Table 12. Weighted average polarity scores. Evaluating Parameters Aggregated Polarity Weighted Average Polarity Aggregated Weighted Average Polarity Multi- Variant Precision 0.3819 0.6658 0.8226 0.9886 Recall 0.4418 0.7126 0.8684 0.9835 F-measure 0.4097 0.6884 0.8449 0.9860 The results of the f-measure differentiated the accuracy of our work from others. If we use the multi-variants system it provided an accuracy of about 98.6%. Figure 10. Performance comparison of movie recommendation system. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 average polarity weighted average polarity aggregated weighted average polarity Multi-Variant precision recall F-measure Figure 10. Performance comparison of movie recommendation system. 6. Conclusions This paper presented an intelligent and automated recommender system to provide topic (action, comedy, horror, etc.) based, accurate recommendations of movies to users. The used approach relies on both quantitative and qualitative data for achieving authentic recommendations. The used quantitative data includes ratings, votes, likes, etc., and the quantitative data is the polarity score that is calculated from user reviews using NLP and opinion mining techniques. The developed application was tested on three external data sources such as Metacritics, IMDB, and Fandango and achieved better results in terms of true recommendations as compared to previous approaches. The presented recommender system was developed using a NoSQL environment with Apache Hadoop to filter and integrate movie descriptions from linked data or ontology (linkedmdb). Our approach used a Fuzzy logic approach for movie ranking categorization. A front-end application is designed in Android for the interaction of a user with the movie recommender through web services. When users search for a movie through a mobile app then the server effectively responds to users with a recommended list of movies. Thus, users can take a decision before and secure a watch time for the movie and can conserve other important resources, like money and energy, etc. 7. Future Work Further work is required to enhance the system for both registered and unregistered viewers or users of apps by adding more parameters, such as showbiz industry influence, movie quality, movie trends and a user’s profile-based movie recommendation system in a NoSQL distributed environment. Semantic and sentiment computation are required to find the semantic relation between the movies and users as well as the psychological influence of movies. Author Contributions: M.I. designed the algorithm and conducted the experiments. I.S.B. supervised the research work. Funding: This research received no external funding. Conflicts of Interest: The authors declare no conflict of interest. Sustainability 2018, 10, 4280 18 of 21 Appendix A. One android app provides the following listed features in the user interface for users’ usage, by which users can request or query a movie and app will respond with movie list, the list is provided by the server machine by interaction of the app. The following Figure A1. illustrates the specification of the mobile application. Sustainability 2018, 10, x FOR PEER REVIEW 18 of 21 6. Conclusions This paper presented an intelligent and automated recommender system to provide topic (action, comedy, horror, etc.) based, accurate recommendations of movies to users. The used approach relies on both quantitative and qualitative data for achieving authentic recommendations. The used quantitative data includes ratings, votes, likes, etc., and the quantitative data is the polarity score that is calculated from user reviews using NLP and opinion mining techniques. The developed application was tested on three external data sources such as Metacritics, IMDB, and Fandango and achieved better results in terms of true recommendations as compared to previous approaches. The presented recommender system was developed using a NoSQL environment with Apache Hadoop to filter and integrate movie descriptions from linked data or ontology (linkedmdb). Our approach used a Fuzzy logic approach for movie ranking categorization. A front-end application is designed in Android for the interaction of a user with the movie recommender through web services. When users search for a movie through a mobile app then the server effectively responds to users with a recommended list of movies. Thus, users can take a decision before and secure a watch time for the movie and can conserve other important resources, like money and energy, etc. 7. Future Work Further work is required to enhance the system for both registered and unregistered viewers or users of apps by adding more parameters, such as showbiz industry influence, movie quality, movie trends and a user’s profile-based movie recommendation system in a NoSQL distributed environment. Semantic and sentiment computation are required to find the semantic relation between the movies and users as well as the psychological influence of movies. Author Contributions: M.I. designed the algorithm and conducted the experiments. I.S.B. supervised the research work. Funding: This research received no external funding. Conflicts of Interest: The authors declare no conflict of interest. Appendix A One android app provides the following listed features in the user interface for users’ usage, by which users can request or query a movie and app will respond with movie list, the list is provided by the server machine by interaction of the app. The following Figure A1. illustrates the specification of the mobile application. Figure A1. Illustration of multi-variant recommendation. Appendix B All the experiments were performed to test the performance and accuracy of the proposed approach using Intel i7 @ 3.4 GHz, operating on Linux/Ubuntu 14.04, 64-bit with 8 GB memory. The nltk tool kit is written in Python under the GPL open source license, the Stanford CoreNLP Natural Language Processing Toolkit and libraries [50–52] as well as Apache Hadoop 2.0 are used for the Figure A1. Illustration of multi-variant recommendation. Appendix B. All the experiments were performed to test the performance and accuracy of the proposed approach using Intel i7 @ 3.4 GHz, operating on Linux/Ubuntu 14.04, 64-bit with 8 GB memory. The nltk tool kit is written in Python under the GPL open source license, the Stanford CoreNLP Natural Language Processing Toolkit and libraries [50–52] as well as Apache Hadoop 2.0 are used for the deployment of the NoSQL environment for our movie recommendation system. Table A1 represents the hardware and software specifications. Table A1. Server machine and Android device specification. Resources Specification of Server Machine Specification of Android Device Processor INTEL i7 processor with 3.4 GHz clock rate Qualcomm Snapdragon 835, 2.45 GHz octa-core Kryo 280 CPU, Adreno 540 GPU RAM 8 GB/machine 4 GB NoSQL Hadoop 2.0, Apache Cassandra. Operating System Linux/Ubuntu 14.04 Android 7.1.1 Storage 1 TB 128 GB (UFS 2.1) Topology Connected by gigabit Ethernet cable 802.11ac Wi-Fi with MIMO, Bluetooth 5.0 LE APP web-bot android app References 1. Raigoza, J.; Karande, V. A Study and Implementation of a Movie Recommendation System in a Cloud-based Environment. Int. J. Grid High Perform. Comput. 2017, 9, 25–36. [CrossRef] 2. Christakou, C.; Vrettos, S.; Stafylopatis, A. A hybrid movie recommender system based on neural networks. Int. J. Artif. Intell. Tools 2007, 16, 771–792. [CrossRef] 3. Said, A.; Kille, B.; de Luca, E.W.; Albayrak, S. Personalizing Tags: A Folksonomy-Like Approach for Recommending Movies. In Proceedings of the 2nd International Workshop on Information Heterogeneity and Fusion in Recommender Systems, Chicago, IL, USA, 27 October 2011; pp. 53–56. 4. Zenebea, A.; Norciob, A.F. Representation, similarity measures and aggregation methods using fuzzy sets for content-based recommender systems. Fuzzy Sets Syst. 2009, 160, 76–94. [CrossRef] 5. Singh, D.K.; Gangwar, A.; Sharma, A. Movie Recommendation System. Volume 4. Available online: www.ijariit.com (accessed on 23 July 2018). 6. Wang, Z.; Yu, X.; Feng, N.; Wang, Z. An improved collaborative movie recommendation system using computational intelligence. J. Vis. Lang. Comput. 2014, 25, 667–675. [CrossRef] 7. Jain, K.N.; Kumar, V.; Kumar, P.; Choudhury, T. Movie Recommendation System. In Intelligent Computing and Information and Communication; Springer: Singapore, 2018; pp. 677–686. http://dx.doi.org/10.4018/IJGHPC.2017010103 http://dx.doi.org/10.1142/S0218213007003540 http://dx.doi.org/10.1016/j.fss.2008.03.017 www.ijariit.com http://dx.doi.org/10.1016/j.jvlc.2014.09.011 Sustainability 2018, 10, 4280 19 of 21 8. Yessenov, K.; Misailovic, S. Sentiment Analysis of Movie Review Comments. Methodology 2009, 17, 1–7. 9. Bhuiyan, H.; Ara, J.; Bardhan, R.; Islam, R. Retrieving YouTube Video by Sentiment Analysis on User Comment. In Proceedings of the 2017 IEEE International Conference on Signal and Image Processing Applications (IEEE ICSIPA 2017), Kuching, Malaysia, 12–14 September 2017. 10. Singh, V.K.; Piryani, R.; Uddin, A.; Waila, P. Sentiment analysis of movie reviews: A new feature-based heuristic for aspect-level sentiment classification. In Proceedings of the 2013 International Multi-Conference on Automation, Computing, Communication, Control and Compressed Sensing (iMac4s), Kottayam, India, 22–23 March 2013. 11. Alsaqer, A.F.; Sasi, S. Movie Review Summarization and Sentiment Analysis using RapidMiner. In Proceedings of the 2017 International Conference on Networks & Advances in Computational Technologies (NetACT), Thiruvanthapuram, India, 20–22 July 2017. 12. Ouyang, C.; Liu, Y.; Zhang, S.; Yang, X. Features-level Sentiment Analysis of Movie reviews. Adv. Sci. Technol. Lett. 2015, 81, 110–113. 13. Hsieh, M.Y.; Chou, W.K.; Li, K.C. Building a mobile movie recommendation service by user rating and APP usage with linked data on Hadoop. Multimed. Tools Appl. 2017, 76, 3383–3401. [CrossRef] 14. Godhani, G.; Dhamecha, M. A Study on Movie Recommendation System Using Parallel Map Reduce Technology; V.V.P. Engineering College: Rajkot, India, 2017; Volume 5. 15. Reza, M.; Sinha, A.; Nag, R.; Mohanty, P. CUDA-enabled Hadoop cluster for Sparse Matrix Vector Multiplication. In Proceedings of the 2015 IEEE 2nd International Conference on Recent Trends in Information Systems (ReTIS), Kolkata, India, 9–11 July 2015. 16. Castells, P.; Fernández, M.; Vallet, D. An Adaptation of the Vector-Space Model for Ontology-Based Information Retrieval. IEEE Trans. Knowl. Data Eng. 2007, 19, 161–272. [CrossRef] 17. Wang, J.; Liu, T. Taiwan Improving Sentiment Rating of Movie Review Comments for Recommendation. In Proceedings of the 2017 IEEE International Conference on Consumer Electronics—Taiwan (ICCE-TW), Taipei, Taiwan, 12–14 June 2017. 18. Wijaya, D.T.; Bressan, S. A Random Walk on the Red Carpet: Rating Movies with user reviews and pagerank. In Proceedings of the 17th ACM Conference on Information and Knowledge Management, Napa Valley, CA, USA, 26–30 October 2008; ACM: New York, NY, USA, 2008; pp. 951–960. 19. Chang, A.; Liao, J.F.; Chang, P.C.; Teng, C.H.; Chen, M.H. Application of artificial immune systems combines collaborative filtering in movie recommendation system. In Computer Supported Cooperative Work in Design (CSCWD). In Proceedings of the 2014 IEEE 18th International Conference on Computer Supported Cooperative Work in Design (CSCWD), Hsinchu, Taiwan, 21–23 May 2014; pp. 277–282. 20. Tumasjan, A.; Sprenger, T.O.; Sandner, P.G.; Welpe, I.M. Predicting Elections with Twitter: What 140 Characters Reveal about Political Sentiment. ICWSM 2010, 10, 178–185. 21. He, W.; Zha, S.; Li, L. Social media competitive analysis and text mining: A case study in the pizza industry. Int. J. Inf. Manag. 2013, 33, 464–472. [CrossRef] 22. Murnane, E.L.; Counts, S. Unraveling abstinence and relapse: Smoking cessation reflected in social media. In Proceedings of the 32nd Annual ACM Conference on Human Factors in Computing Systems, Toronto, ON, Canada, 26 April–1 May 2014; ACM: New York, NY, USA, April 2014; pp. 1345–1354. 23. Diakopoulos, N.; Naaman, M.; Kivran-Swaine, F. Diamonds in the rough: Social media visual analytics for journalistic inquiry. In Proceedings of the 2010 IEEE Symposium on Visual Analytics Science and Technology, Salt Lake City, UT, USA, 25–26 October 2010; pp. 115–122. 24. Baldwin, T.; Cook, P.; Lui, M.; MacKinlay, A.; Wang, L. How Noisy Social Media Text, How Diffrnt Social Media Sources? In IJCNLP; The Association for Computational Linguistics: Stroudsburg, PA, USA, October 2013; pp. 356–364. 25. Corley, C.D.; Cook, D.J.; Mikler, A.R.; Singh, K.P. Text and structural data mining of influenza mentions in web and social media. Int. J. Environ. Res. Public Health 2010, 7, 596–615. [CrossRef] [PubMed] 26. Asur, S.; Huberman, B.A. Predicting the future with social media. In Proceedings of the 2010 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, Toronto, ON, Canada, 31 August–3 September 2010; Volume 1, pp. 492–499. 27. Timmaraju, A.; Khanna, V. Sentiment Analysis on Movie Reviews using Recursive and Recurrent Neural Network Architectures. Available online: https://cs224d.stanford.edu/reports/TimmarajuAditya.pdf (accessed on 14 November 2018). http://dx.doi.org/10.1007/s11042-016-3833-0 http://dx.doi.org/10.1109/TKDE.2007.22 http://dx.doi.org/10.1016/j.ijinfomgt.2013.01.001 http://dx.doi.org/10.3390/ijerph7020596 http://www.ncbi.nlm.nih.gov/pubmed/20616993 https://cs224d.stanford.edu/reports/TimmarajuAditya.pdf Sustainability 2018, 10, 4280 20 of 21 28. Sarker, A.; Ginn, R.; Nikfarjam, A.; O’Connor, K.; Smith, K.; Jayaraman, S.; Upadhaya, T.; Gonzalez, G. Utilizing social media data for pharmacovigilance: A review. J. Biomed. Inform. 2015, 54, 202–212. [CrossRef] [PubMed] 29. Resnick, P.; Iacovou, N.; Suchak, M.; Bergstrom, P.; Riedl, J. GroupLens: An open architecture for collaborative filtering of netnews. In Proceedings of the 1994 ACM Conference on Computer Supported Cooperative Work (CSCW ’94), Chapel Hill, NC, USA, 22–26 October 1994; ACM: New York, NY, USA, 1994; pp. 175–186. 30. Shardanand, U.; Maes, P. Social information filtering: Algorithms for automating “word of mouth”. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ’95); Katz, I.R., Mack, R., Marks, L., Rosson, M.B., Nielsen, J., Eds.; ACM Press/Addison-Wesley Publishing Co.: New York, NY, USA, 1995; pp. 210–217. 31. Lekakos, G.; Caravelas, P. A Hybrid Approach for Movie Recommendation; Springer Science + Business Media: New York, NY, USA, 21 December 2006. 32. Tumsare, P.; Sambare, A.S.; Jain, S.R. Sentiment Analysis Approach for Movie Reviews of Natural Language. Int. J. Res. Comput. Commun. Technol. 2014, 3, 256–261. 33. Kreutzer, J.; Witte, N. Opinion Mining Using SentiWordNet Semantic Analysis; HT 2013/14; Uppsala University: Uppsala, Sweden, 2013. 34. Haddia, E.; Liua, X.; Shib, Y. The Role of Text Pre-processing in Sentiment Analysis. Procedia Comput. Sci. 2013, 17, 26–32. [CrossRef] 35. Webster, J.J.; Kit, C. Tokenization as the initial phase in NLP. In Proceedings of the 14th conference on Computational linguistics, Nantes, France, 23–28 August 1992. 36. Vijayarani, S.; Janani, M.R. Text mining: Open source tokenization tools—An analysis. Adv. Comput. Intell. Int. J. 2016, 3, 37–47. 37. Issac, B.; Jap, W.J. Implementing spam detection using bayessian and porter stemmer keyword stripping approaches. In Proceedings of the TENCON 2009–2009 IEEE Region 10 Conference, Singapore, 23–26 November 2009; pp. 1–5. 38. Porter, M.F. An Algorithm for Suffix Stripping. J. Program. 1980, 14, 130–137. [CrossRef] 39. Alphabetical List of Part-Of-Speech Tags Used in the Penn Treebank Project. Available online: http://www. ling.upenn.edu/courses/Fall_2003/ling001/penn_treebank_pos.html (accessed on 11 June 2018). 40. Tf-idf Weighting. Available online: https://nlp.stanford.edu/IR-book/html/htmledition/tf-idf-weighting- 1.html (accessed on 7 April 2018). 41. Hakim, A.A.; Erwin, A.; Eng, K.I.; Galinium, M.; Muliady, W. Automated document classification for news article in Bahasa Indonesia based on term frequency inverse document frequency (TF-IDF) approach. In Proceedings of the 2014 6th International Conference on Information Technology and Electrical Engineering (ICITEE), Yogyakarta, Indonesia, 7–8 October 2014; pp. 1–4. 42. Esuli, A.; Sebastiani, F. Sentiwordnet: A Publicly Available Lexical Resource for Opinion Mining. Available online: http://nmis.isti.cnr.it/sebastiani/Publications/LREC06.pdf (accessed on 21 September 208). 43. Baccianella, S.; Esuli, A.; Sebastiani, F. Sentiwordnet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining. LREC 2010, 10, 2200–2204. 44. Penalver-Martinez, I.; Garcia-Sanchez, F.; Valencia-Garcia, R.; Rodriguez-Garcia, M.A.; Moreno, V.; Fraga, A.; Sanchez-Cervantes, J.L. Feature-based opinion mining through ontologies. Expert Syst. Appl. 2014, 41, 5995–6008. [CrossRef] 45. Miller, G.A. WordNet: A Lexical Database for English. Commun. ACM 1995, 38, 39–41. [CrossRef] 46. Pedersen, T.; Patwardhan, S.; Michelizzi, J. WordNet::Similarity—Measuring the Relatedness of Concepts; The Association for Computational Linguistics: Stroudsburg, PA, USA, 2004. 47. Bird, S.; Loper, E. NLTK: The Natural Language Toolkit; The Association for Computational Linguistics: Stroudsburg, PA, USA, 2004. 48. Manning, C.D.; Surdeanu, M.; Bauer, J.; Finkel, J.; Bethard, S.J.; McClosky, D. The Stanford CoreNLP Natural Language Processing Toolkit. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations 2014, Baltimore, MD, USA, 22–27 June 2014. 49. Atserias, J.; Casas, B.; Comelles, E.; Gonzàlez, M.; Padró, L.; Padro, M. FreeLing 1.3: Syntactic and Semantic Services in an Open-Source NLP Library; TALP Research Center Universitat Politècnica de Catalunya: Barcelona, Spain, 2006. http://dx.doi.org/10.1016/j.jbi.2015.02.004 http://www.ncbi.nlm.nih.gov/pubmed/25720841 http://dx.doi.org/10.1016/j.procs.2013.05.005 http://dx.doi.org/10.1108/eb046814 http://www.ling.upenn.edu/courses/Fall_2003/ling001/penn_treebank_pos.html http://www.ling.upenn.edu/courses/Fall_2003/ling001/penn_treebank_pos.html https://nlp.stanford.edu/IR-book/html/htmledition/tf-idf-weighting-1.html https://nlp.stanford.edu/IR-book/html/htmledition/tf-idf-weighting-1.html http://nmis.isti.cnr.it/sebastiani/Publications/LREC06.pdf http://dx.doi.org/10.1016/j.eswa.2014.03.022 http://dx.doi.org/10.1145/219717.219748 Sustainability 2018, 10, 4280 21 of 21 50. Tiwari, J.; Pawar, M.; Pandey, A. A hadoop based collaborative filtering recommender system accelerated on gpu using opencl. Int. J. Eng. Sci. Res. Technol. 2017, 6, 195–209, Retrieved 5 September 2017. 51. Thangavel, S.K.; Thampi, N.S.; Johnpaul, C.I. Performance Analysis of Various Recommendation Algorithms Using Apache Hadoop and Mahout. Int. J. Sci. Eng. Res. 2013, 4, 279–287. 52. Jose, A.V.; Jini, K.M. Personalized Movie Recommender System using Rank Boosting Approach on Hadoop. IJIRST Int. J. Innov. Res. Sci. Technol. 2015, 2, 2349–6010. © 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/). http://creativecommons.org/ http://creativecommons.org/licenses/by/4.0/. Introduction Related Work Multi-Variant Expert System NLP Module Tokenization Stemming (Lemmatization) Stop Word Evacuation POS-Tag Generation Polarity Computation Lexical Frequency Measuring Polarity Identification Weighted Polarity Manipulation Categorization WordNet Geo Ontology Time Ontology TVA-CS Recommendation Experimental Setup NoSQL for Big Data Stroage Experiment and Results Evaluation Conclusions Future Work References 
work_kxnytaeovfgx5ka7hu2mrtlgwu	----	EMA(A4) Ad_JRM International Journal of Rotating Machinery 2000, Vol. 6, No. 5, pp. 363-373 Reprints available directly from the publisher Photocopying permitted by license only (C) 2000 OPA (Overseas Publishers Association) N.V. Published by license under the Gordon and Breach Science Publishers imprint. Printed in Malaysia. Rotating Machinery Predictive Maintenance Expert System Through M. SARATH KUMAR and B.S. PRABHU* Department of Applied Mechanics, Machine Dynamics Laboratory, Indian Institute of Technology, Madras 600 036, India (Received 24 November 1998;In final form 19 February 1999) Modern rotating machines such as turbomachines, either produce or absorb huge amount of power. Some of the common applications are: steam turbine-generator and gas turbine- compressor-generator trains produce power and machines, such as pumps, centrifugal compressors, motors, generators, machine tool spindles, etc., are being used in industrial applications. Condition-based maintenance of rotating machinery is a common practice where the machine’s condition is monitored constantly, so that timely maintenance can be done. Since modern machines are complex and the amount of data to be interpreted is huge, we need precise and fast methods in order to arrive at the best recommendations to prevent catastrophic failure and to prolong the life of the equipment. In the present work using vibration characteristics of a rotor-bearing system, the condition of a rotating machinery (electrical rotor) is predicted using an off-line expert system. The analysis of the problem is carried out in an Object Oriented Programming (OOP) framework using the finite element method. The expert system which is also developed in an OOP paradigm gives the type of the malfunctions, suggestions and recommendations. The system is implemented in C++. Keywords: Condition monitoring, Vibration, Finite element method, Expert system, Object oriented programming, Rotating machinery 1. INTRODUCTION Condition monitoring programs offer an innova- tive means for modern rotating equipment to implement and schedule predictive maintenance, as opposed to relying on conventional preventive maintenance techniques. Vibration, wear and temperature are the three important condition monitoring techniques to predict the health of the rotating machinery. Sohre (1972) has discussed the analysis and prediction of the reliability of turbo- machinery installations and influence of various components on overall reliability. He has also presented, the prediction and identification of faults in the turbomachinery. Renwick (1984) has presented a condition monitoring program for plant maintenance and diagnosis of heavy machinery problems in which he discussed the importance of machinery dynamics in the design of a rotating machinery. Zimmer et al. (1986) have *Corresponding author. Tel." (044)235 1365, ext. 8180. Telex: TECHMAS 041-8926 IITM IN. Fax: (044) 2350509. E-mail: bsprabhu@hotmail.com. 363 364 M. SARATH KUMAR AND B.S. PRABHU reviewed predictive maintenance programs for the rotating machinery using vibration signature anal- ysis. Recently, the literature has given emphasis on expert systems for condition-based maintenance. An expert system is an Artificial Intelligence system created to solve problems in a particular domain. Hill et al. (1988) have presented the expert system for health monitoring of a rotating machinery. They discussed the expert system components for condition monitoring and design considerations. Lewis et al. (1989) have described a rule-based expert system for fault identification and predictive maintenance of turbomachinery using vibration oriented diagnosis. Sakthivel and Kalyanaraman (1993) have presented a knowledge based expert systems (KBES) approach to an engineering data. Rao (1993) discussed the various aspects of condi- tion monitoring based on the vibration signature analysis and development of an off-line expert system for fault diagnosis in rotating machines. Sarath Kumar and Prabhu (1997) presented a rule- based off-line expert system for condition moni- toring of a rotating machinery. Bettig and Han (1998) have discussed the usage of rotordynamic models in predictive maintenance, in which vari- ables characterizing the state of deterioration mechanism are trended to determine the rate of deterioration to predict the machine life or the maintenance period (the time period between major over haul shutdowns). In finite element modeling of rotor-bearing systems, Gmfir and Rodrigues (1991) proposed a C-compatible linearly tapered shaft element to model the rotor system. This element consists of four degrees of freedom per nodal point (i.e., two lateral displacements and two total cross-sectional rotations), which includes the effects oftranslational inertia, rotary inertia, gyroscopic moments, internal viscous and hysteretic damping, shear deformation and mass eccentricity. Singh et al. (1977) have made parametric studies on the performance character- istics (load-carrying capacity, oil flow, attitude angle, stiffness coefficients) of a hydrostatic oil journal bearing using the finite element method. Birembaut and Peigney (1980) predicted the dynamic characteristics of rotor supported on hy- drodynamic bearings using FEM and validated the theoretically obtained results with the experimental results. Lund (1987) has reviewed the concept of calculating the spring and damping coefficients ofjournal bearings. When a rotor is operated beyond a certain rota- tional speed, a very high and unstable vibration com- ponent with a fractional frequency ofthe rotor speed will be developed. This rotor speed is referred to as the instability threshold. Lund (1965) presented a theoretical analysis for predicting stability of a sym- metrical, flexible rotor supported in journal bear- ings. Rao (1983) and Muszynska (1988) presented instability criteria of the rotor supported on journal bearings. Majumdaret al. (1988) investigated the sta- bility ofa rigid rotor in oiljournal bearings including the effect of elastic distortion in the bearing liner. The method of implementation whereby pro- grams are organized as a cooperative collections of objects, in which each object represents an instance ofsome class is called Object Oriented Programming (OOP). In rotor dynamic analysis a physical object is a constituent of the rotor, representing its geome- trical shape, material, etc. and also the linkages with other rotating components. The main features of OOP are data encapsulation, operator overloading and inheritance. Each object encapsulates (hides) its data attributes from other objects and only displays its behavior, which enhances the portability of the application codes. Operator overloading refers to multiple usage of the same function name, with differing argument characteristics. A class may have several derived classes that may inherit some of its attributes and behavior from the parent class. This enables efficient and reusability of codes (e.g., coupling of the shaft or bearing may inherit char- acters of the class shaft). In OOP, both data and subroutines are linked intrinsically which leads to clear thinking of program design, so that the errors are minimized. Allocation of less storage space, less computational time, allowing programs to be substantially altered, without the need to change existing code, are the advantages of the OOP compared to FORTRAN language. Forde et al. PREDICTIVE MAINTENANCE 365 (1990) have discussed the implementation of an object-oriented program for the numerical analysis of two-dimensional solid and structural mechanics, emphasizing on knowledge-based expert system. Zeglinski et al. (1994) have discussed the concepts of OOP for the finite element method using C++ language in a generalized matrix library and showed that the efficiency, flexibility and maintainability of the C+/ program is superior to a comparable version written in a non-OOP language, such as FORTRAN. Bettig et al. (1997) described the concepts of OOP for predictive maintenance of a rotating machinery, in which trend variables (e.g. bearing rotordynamic coefficients) were trended to predict the machine life/maintenance period using a finite element model in a graphical framework. An expert system for condition monitoring pur- poses consists of three key components. They are (i) knowledge base (ii) inference engine and (iii) data base. The architecture ofthe expert system is shown in Appendix A. The system comprises hierarchical levels of generic rules (surface knowledge) and generic analytical simulation models (deep models). For a rotating machinery some surface knowledge is required about vibration, bearings, lubricant, seals, coupling etc. coupled to a much deeper knowledge ofvibration analysis (i.e., critical speeds, resonance, responses, stability, etc.), bearing analysis (static and dynamic characteristics), lubricant flow analysis, seal analysis, crack initiation/propagation etc. In order to predict the type of the malfunction exactly, the detailed models knowledge is used. In the present study the deep knowledge regarding vibration analysis and bearing analysis is used to predict the malfunctions in rotating machinery. Knowledge of the machine is stored in the form of IF-THEN rules. Therefore rule has the following form IF (antecedent AND (antecedent2) AND (antecedentn) THEN (consequent AND (consequent2) AND (consequentm) The antecedents to n are known as premises, which are given facts, and the consequents to rn are known as conclusions, which are deduced facts. The inference engine analyses the base and actual data in the data base and detects the discrepancies by using rules. The data base stores base values (from data books, reference manuals, past experi- ence of personnel and experts in particular field) and actual values (experimentally measured values, analytically calculated values etc.). The rule compiler converts the rules about a machine con- dition to a form expected by the inference engine. The report generator, creates a consolidated report about the health of the machine based on infer- ences. The system gives the type of malfunction, cause of the problem, recommendations and sug- gestions regarding the machine. Very little literature is available which relates to combining theoretical rotor dynamic analysis with expert system for pre- dicting the malfunctions in a rotating machinery. In the present work a theoretical model to predict the dynamic characteristics of the rotor-bearing system using the finite element method and a rule-based off-line expert system, for fault diagnosis of a rotating machinery, are presented and discussed in an object oriented framework. 2. ANALYSIS A true object oriented solution to an engineering problem arises from a thorough analysis of the problem domain and the development of an application that emulates the existence of objects. A rotor analysis using the finite element method consists of four different classes of objects. They are: shaft, material, matrices and eigenvalue. Class shaft consists of details of the number of elements, number of nodes, type of element, type of shape function, speed of rotor, number of point masses etc. Figure corresponds to the C/+ program in an OOP paradigm and it illustrates how the class shaft data has been accessed through member functions in the main program without passing the data directly through functions. 366 M. SARATH KUMAR AND B.S. PRABHU class shaft{//class declaration protected: int ne, nn, npt;//No, of elements, nodes, point masses float len, dia, speed;//shaft length, diameter, speed char eletyp[10], shafunc[ 10];//element type, shape function public: void input );//functions than can access the private or protected data void printshaft ); //class method definitions follow void shaft :: inputl ( ){ cout << "Enter ne:" << "npt:" << "speed:" << endl; cin >> ne >> npt >> speed; for (int i= 1; <= ne; ++){ cout << "i:" << "len:" << "diG:" << endl; cin >> >> len >> diG; cout << "Enter eletyp:" << "shafunc:" << endl; cin >> eletyp >> shafunc; nn ne + 1;//In beam element} void shaft :: printshaft { cout << setw (3) << ne << setw (3) << nn << setw (3) << npt << endl; cout << setw (5) << speed << setw (10) << eletyp << setw (10) << shafunc << endl; for (int i= 1; i<=ne; i++){ cout << setw (3) << << setw (5) << len << setw (5) << dia << endl;} } //MAIN PROGRAM void main (void) shaft sl;//shaft object s .input );//variables encapsulated within methods s .printshaft ); //remaining source code follows FIGURE Passing of data from class shaft to main pro- gram via member functions. The program shown in Fig. consists of class shaft, two member functions input (), printshaft and the main program. The class shaft consists of two access specifiers "protected" and "public". The data declared under protected were called as member data and the functions declared in public are called member functions or methods, because they are the members of the class. The data attributes declared under protected of the class shaft are encapsulated with in the object. The data declared under "protected" were only known to the class and its descendent classes in the hierarchy, whereas "public" data are known to all the users of the class. The two functions input and printshaft defined in the public of the class shaft is used to access the input from the class shaft and print the same. The data declared in "private" is known only to the class and not accessible to the other classes. Messages are the means of invoking the methods supported by the class of objects and they are the primary way in which objects communicate and interact with each other. Messages are generally accompanied by arguments. The scope resolution operator "::" is used to tie the function definition to the class shaft. In Fig. 1, input and printshaft are the two member functions tied up to the class shaft. Member function inputl is for giving the input details like number of elements, number of point masses, speed of rotation, length, diameter of the shaft element etc., in the finite element analysis of rotor. Member function printshaft ( to print the input details given in member function input (). The main program, consists of object shaft "s 1". s 1. input and s 1.printshaft corresponds to class member access operator, which connects the object name and the member functions, so that, the member functions input and printshaft of the s object will be executed. Class material consists of physical and geometri- cal properties (p, E, G, A, Ip, Ia, etc.) of shaft and disks. Class bearing is inherited from the class shaft which consists of the number of bearings, member functions to calculate static and dynamic character- istics of the bearing etc. The inherited class bearing is shown in Fig. 2. In a similar way as the class shaft above, the two classes, material and bearing are framed to access the data to the main program. Class matrices consist of element stiffness matrices and global stiffness matrices (stiffness matrix, mass PREDICTIVE MAINTENANCE 367 class bearing: public shaft { private: float theta, z, **p,*h, L, U, eta; //eta =-r/ float c, eps, W,Q,F; //eps e float mu,trise, **K,**C; //**K,**C-stiffness & damping coefficients public://functions to calculate steady state & dynamic, characteristics. void calcsteady(float**p,float*h); void calcdyna(float **K,float **C);}; FIGURE 2 Inherited bearing class. matrix, gyroscopic matrix, damping matrix and unbalance force matrix, etc.). Class eigenvalue is used for determining the critical speeds ofthe rotor. The different classes of objects for the expert system are Class input, Class rules and Class inference. Class input consists of details of actual and base values of machine. Class rules consists of the knowledge base rules of the rotating machinery, and Class inference for details of interpreting the data. The information transfer from one class to another class was achieved through messages. 2.1. Bearing Modeling A schematic diagram ofjournal bearing is shown in Fig. 3. The two-dimensional Reynolds equation gov- erning the flow ofthe lubricant in the clearance space ofjournal bearing under steady-state loading is R2 00 - + z -z 200"By using the variation principles of calculus theabove equation can be expressed in pressurefunctional as Iao=o 24R2 + + p d0 dz. (2) The fluid film has been divided into triangular elements as in the reference Booker and Huebner X Y w FIGURE 3 Journal bearing. (1972). Applying Reynolds boundary conditions p---0 at 0=0 and at z=0, L; p=(Op/O0)=O at The load carrying capacity, leakage of lubricant and frictional force are calculated by integrating the Eqs. (3)-(5) which are given in Cameron (1966). w-[ fI pOdO(0+ cos dz) 7r+c L 2 I 1/2 +(/oo /oo l sinOdOdz) (3) Q--ZR dO+ L Uh / dz, 27r h ( z)dl (4) f0 L F- 7-dA Coefficient of friction (#) F/W, (6) Temperature rise At FU JmC where rn (pQ). (7) 368 Bearing M. SARATH KUMAR AND B.S. PRABHU 2585mm 200mm 620 FIGURE 4 Electrical rotor. Dynamic characteristics: Stiffness and damping coefficients of the journal bearing are calculated using the perturbation analysis presented by Lund (1987). 2.2. Rotor-Bearing System Modeling The electrical rotor-bearing system taken for analysis is shown in Fig. 4. The schematic diagram of mathematical equivalent model is shown in Fig. 5. The rotor has been descritized into 50 beam elements. Equations of motion for a linearly damped rotor-bearing system proposed by Gmtir and Rodrigues (1991 can be expressed in the matrix form as follows: M0 + (G + r/1K)0 + (rllK + rl2C)q r (8) where r represents the 4N dimensional nodal unbalance force vector. The matrices M and K are symmetric, whereas G and C are skew- symmetric. Since the equations of motion are written in discrete form, the effects of thin rigid disks and short journal bearings can be incorpo- rated directly into the formulation by adding adequate nodal contributions. FIGURE 5 Mathematical equivalent model of rotor-bearing system. developed in C++ in an object oriented paradigm. The formulated matrix equations are solved by Gauss elimination algorithm to obtain the bearing pressures in the clearance space of the bearing. Hessenberg algorithm is used to obtain the eigen values of the rotor bearing system. The stability limit of the rotor bearing system was obtained by plotting the values of growth factor versus rota- tional speed. The expert system code has been devel- oped in C++ in an object-oriented framework. Figure 6 shows the flow chart for rotor dynamic analysis and its linkage to the expert system shell. 4. RESULTS AND DISCUSSION 3. SOLUTION PROCEDURE The problem has been formulated by using the finite element method. A computer code was For a given rotor bearing system the program calculates the static and dynamic characteristics of the journal bearing and critical speeds, unbalance response, stability speed limit of the rotor. These values are taken as input to the actual value data PREDICTIVE MAINTENANCE 369 Solving Reynolds equation using FEM Calculation of static & dynamic characteristics of joumal bearing Rotor -bearing system modelling using FEM Solving dynamic global matrix by Hessenberg algorithm to get eigen values Investigation of stability limit ofrotor bearing system To Expert system actual value data base FIGURE 6 Flow chart showing rotor dynamic analysis and linkage to expert system shell. base. The rotor shown in Fig. 4 is a rigid rotor supported on oil lubricated journal bearings. Total mass of the rotor is 3500 kg. The bearing details used for the analysis are D--0.2m, L/D=0.675, f/- 0.025 Pa s, c 0.48, c 160 lam, N 1500 rpm. The steady state characteristics results of the journal bearing obtained from the equations (3-7) are W-- 17.44 kN, Q-7.17 10-4m3/s, F= 251 N, #--0.014, At--5.18C. The coefficient of friction due to aerodynamic forces and unbalance whirling is assumed to be negligible. When the journal is perturbed by a small displacement and a small velocity, the new journal forces can be expressed by a Taylor series expansion, from which one can establish the stiffness and damping coefficients of 25.00 _20.00 (13 0 15.00 10.00 c 0 5.00 0.00 0.00 1ooo.oo ooo.oo ooo.oo ,,ooo.oo ooo.oo Rotor Speed (RPM) FIGURE 7 Rotational speed vs response. the bearing. The values obtained are K- 1.88 108, K12 6.83 x 107, K. 9.07 106, K22 2.36 108 N/m, C]1-2.33105 C2- 2.10105 C2]- 1.66 105, C22- 4.48 105N s/re. The static performance characteristic equations (3)-(7) and dynamic characteristics have compared with the software package REYNOLDS developed by the second author for bearing modeling using the finite element method in FORTRAN language. The results obtained by using the above software package for the same input values are W= 17.50 kN, Q- 7.25 x 10-4 m3/s, F-- 260 N, #-0.0148, At-3.13C, Kl--l.92x10 6.73 107, K21- 8.97 x 106, K22--2.48 x 108 N/m, C-2.13x105 C:-1.83105 C2-1.34x 105, C22 4.24 105 N s/m. The former stiffness and damping coefficients values are used in the calculation of critical speeds, unbalance response and the stability of the rotor bearing system. The operating speed of the rotor is 1500 rpm. The rotor system is stable and free from subsynchronous vibration at operating speed. The unbalance eccen- tricity of the rotor is taken as 0.01 mm. Figure 7 shows the unbalance response ofthe rotor at the left end bearing point. The peak of the curve corre- sponds to first critical speed of the rotor, which is equal to 1780.2rpm. The critical speed of the rotor is compared with the software package ROTODYNA developed by the second author for rotor-bearing modeling using the finite element 370 M. SARATH KUMAR AND B.S. PRABHU method in FORTRAN. The calculated value of first critical speed of the rotor using ROTODYNA is 1772.8 rpm. The results obtained through OOP program were compared with the two software packages and found to be in good agreement. The amplitude of vibration at critical speed is 24.0 lain. From the ISO 2372 vibration severity chart for stable operation of the rotor, the vibration displace- ment peak-to-peak at the bearing cap at 1500 cpm is 10 l.tm, which is taken as the base value displace- ment ofthe rotor. These values can be directly fed to the data base. The available knowledge is imple- mented in the form of a set of rules. Some of the rules are shown in Appendix B, which will predict the type of malfunction of the rotating machinery through the inference engine. The system uses data to match with the ante- cedents of every rule. If any rule matches, then that rule is said to be fired or triggered and the cor- responding inference is added to the report gene- rator. Suggestions and recommendations will be displayed by the system. For example, the actual vibration amplitude level at bearing point. at frequency I*RPM is 24 lam, which exceeds the base value amplitude level by 14tm. The ruli R1 in Appendix B corresponds to a rotor unbalance malfunction. In practice, rotors can never be perfectly balanced because of manufacturing errors, such as inhomogeneties in material, manu- facturing tolerances and gain or loss of material during operation, which increases the vibrational response of the rotor-bearing system. A state of imbalance occurs when the center of mass of the rotating system does not coincide with the center of rotation. If the frequency of excitation coincides with one of the natural frequencies of the system and the unbalance response envelop is narrow (In rules of Appendix B:--1 corresponds to true) and predominantly in the radial plane, and amplitude of vibration of journal at bearing point exceeds the base value, then rule R1 is triggered. Since all the antecedents and consequents of the rule R1 satisfies the above input data, rule R1 gets triggered and deduces that shaft unbalance is true. This new inference causes the rule R4 gets triggered, and 10.00 8.00 2o= d.d Rotor Speed (rpr -2.00 6.O0 4.00 2.00 0 L_9 o.oo --4.00 FIGURE 8 Rotational speed vs growth factor. deduces "shaft unbalance detected". Check for eccentricity and mass unbalance; if necessary, balancing is to be done, is the recommendation and suggestion given by the system from rule R4. Fig. 8 shows growth factor versus speed of the rotor, from which one can find that the threshold speed of the rotor is 3740.5rpm. If the rotor reaches this speed and the amplitude of vibration exceeds the base value at bearing point 1, rule R3 gets triggered and deduces that the rotor has reached threshold speed. If 1X, 2X order frequen- cies are existing with 1X amplitude of vibration greater than 2X amplitude, the phase angle is equal to 180, the envelop is narrow, the dominant plane is radial, and the actual vibration amplitude levels at bearing point at frequency I*RPM exceeds their respective allowable percentage change of base value amplitude level, then rule R5 gets triggered, and deduces that shaft misalignment has occurred. This new inference causes the rule R6 gets triggered, and deduces "misalignment ofthe shaft is detected". Check alignment is the recom- mendation given by the system. If frequencies 0.25*RPM, 0.33*RPM, 0.66*RPM, I*RPM exist in the radial dominant plane with bearing lubricating oil temperature exceeding base value temperature 50C by 60.0% (i.e. 50 x 1.6=80C, At=30C) and bearing coefficient of friction exceeds 0.1, then R15 gets triggered and deduces that the malfunc- tion is "light rotor rub in bearings". Similarly, the system uses data to match the compiled rules and PREDICTIVE MAINTENANCE 371 gives the type ofmalfunction, such as a rotor crack, looseness, gear box problem, etc. and necessary action should be taken. The expert system should be run each time to take into account any effect due to changes in the parameters. 5. CONCLUSIONS In the present work the static and dynamic characteristics of journal bearing and the vibra- tion characteristics of a rotor-bearing system were obtained using the finite element method in an OOP framework. Using detailed deep knowledge of vibration analysis and bearing analysis one can predict the condition of the machine accurately. An off-line expert system for condition monitoring of a rotating machinery has been developed in an object oriented programming frame work. An object orientation is a promising paradigm for handling complexities in the large scale software applica- tions. The development of a fault diagnostic system for rotating machinery requires a resolution of a number of issues involving expert systems, analy- tical methods and user interaction. NOMENCLATURE English Symbols A C C E G G h cross section area of rotor, m2 radial clearance of bearing, m circulation or pseudo-gyroscopic matrix damping coefficients of bearing, N s/m (i,j= to 2) specific heat of oil, kCal/kg K diameter of the journal, m distance between bearing center and journal center, m Young’s modulus, N/m2 gyroscopic matrix shear modulus, N/m2 film thickness [c (1 + e cos 0)], m cross sectional moment of inertia, kgm2 polar moment of inertia, kgm2 K L N N P PT1 q Q R At U W W z 1X stiffness matrix stiffness coefficients of bearing N/m (i,j= to 2) length of the bearing, m mass flow rate of the oil, kg/s mass matrix number of gear teeth in meshing rotor speed, rev/min number of nodes hydrodynamic pressure, Pa location of bearing generalized displacements, m total leakage of oil through bearing, m3/s nodal unbalance force vector, N radius of the journal, m temperature rise C velocity component in x direction, m/s weight of the rotor, N load carrying capacity of bearing, N axial coordinate of bearing, m l*rpm Greek Symbols 711 7] 711,7]2 # 0 P p 7- film extent, rad eccentricity ratio (e/c) viscous loss factor, s viscosity of oil, Pa s auxiliary damping factors coefficient of friction circumferential coordinate, rad mass density of the shaft, kg/m3 density of the oil (900), kg/m shear stress, Pa angular velocity, rad/s References Bettig, B.P. and Han, R.P.S. (1997) An Object-oriented program machine condition monitoring scheme for rotating machinery, Proc. of ASME Design Engineering Technical Conference, Sacramento, California, pp. 1-13. Bettig, B:P. and Han, R.P.S. (1998) Predictive maintenance using the rotordynamic model of a hydraulic turbine-generator rotor, ASME Transactions, Journal of Vibration and Acoustics, 120, 441-448. Birembaut, Y. and Peigney, J. (1980) Prediction of dynamic properties of rotor supported by hydrodynamic bearings using the finite element method, computers and structures, 12, 483-496. 372 M. SARATH KUMAR AND B.S. PRABHU Booker, J.F. and Huebner, K.H. (1972) Application of finite element methods to lubrication: an engineering approach, ASME Transactions, Journal of Lubrication Technology, 313-323. Cameron, A. (1966) The Principles of Lubrication, Longmans, London. Forde, B.W.R., Foschi, R.O. and Stiemer, S.F. (1990) Object- oriented finite element analysis, Computers and Structures, 34 (3), 355-374. Gmfir, T.C. and Rodrigues, J.D. (1991) Shaft finite elements for rotor dynamics analysis, ASME Transactions, Journal of Vibration and Acoustics, 113, 482-493. Hill, J.W. and Baines, N.C. (1988) Application of an expert system to rotating machinery, Proc. of I Mech. E, C307, 449-454. Lewis, D.W. and Sheth, P. (1989) ROMEX An expert system testbed for turbomachinery diagnostics, Proc. of 19 th Turbomachinery Symposium, pp. 135-148. Lund, J.W. (1965) Stability ofan elastic rotor in journal bearings with flexible, damped supports, ASME Transactions, Journal ofApplied Mechanics, 911-920. Lund, J.W. (1987) Review of the concept of dynamic coefficient for fluid filmjournal bearings, ASME Transactions, Journal of Tribology, 109, 37-41. Majumdar, B.C., Brewe, D.E. and Khonsari, M.M. (1988) Stability of a rigid rotor supported on flexible oil journal bearings, ASME Transactions, Journal of Tribology, 110, 181-187. Muszynska, A. (1988) Stability of whirl and whip in rotor- bearing systems, Journal ofSound and Vibration, 127, 49-64. Rao, J.S. (1983) Instability of rotor in fluid film bearings, ASME Transactions, Journal of Vibration, Acoustics, Stress and reliability in Design, 105, 274-279. Rao, J.S. (1993) Future Perspective Condition Monitoring with Special Emphasis on Expert Systems, Frontiers of Tribology and Condition Monitoring, Tata McGraw Hill, New Delhi, 87-101. Renwick, J.T. (1984) Condition monitoring using computerized vibration signature analysis, IEEE Transactions on Industry Applications, IA-20, 519-527. Sakthivel, T.S. and Kalyanaraman, V. (1993) A knowledge based expert system for integrated engineering, Engineering with Computers, 9, 16. Sarath Kumar, M. and Prabhu, B.S. (1997) Condition moni- toring of rotating machinery using expert systems, Proc. of International Conference on Industrial Tribology, Calcutta, India, pp. 403--411. Singh, D.V., Sinhasan, R. and Ghai, R.C. (1977) Performance characteristics of hydrostatic journal bearings with journal rotation by a finite element method, Wear, 44, 41-55. Sohre, J.S. (1972) Turbomachinery analysis and protection, Proc. 1st Turbomachinery Symposium, pp. 1-9. Zeglinski, G.W., Han, R.P.S. and Aitchison, P. (1994) Object oriented matrix classes for use in a finite element code using C+/, International Journal for Numerical Methods in Engi- neering, 37, 3921-3937. Zimmer, S. and Bently, D.E. (1986) Predictive maintenance programs for rotating machinery using computerized vibra- tion monitoring systems, Proc. of 1st European Turboma- chinery Symposium, Brunel University, pp. 91-107. APPENDIX A: THE ARCHITECTURE OF EXPERT SYSTEM Report Recommendations for Maintenance Report Generator Rules--Rulecompiler Inference Engine Knowledge Base Data Base Base value Values Database f Actual-value Database Values PREDICTIVE MAINTENANCE 373 APPENDIX B: RULES FOR ROTATING MACHINERY RI" IF FREQUENCY (I*RPM) AND NARROW BAND ENVELOP AND DOMINANT PLANE RADIAL AND AMPLITUDE AT PTI > BASE_VALUE_AT_P 1.20) THEN SHAFT UNBALANCE R2: IF FREQUENCY (I*RPM) AND AMPLITUDE AT PT1 > BASE_VALUE_AT_P 1.20) AND PHASE ANGLE 90 THEN CRITICAL SPEED R3: IF RPM 2.1*CRITICAL SPEED AND AMPLITUDE AT PT1 > (BASE_VALUE_AT_PI* 1.20) THEN THRESHOLD SPEED R4: IF SHAFT UNBALANCE THEN "SHAFT UNBALANCE DETECTED" AND "CHECK FOR ECCENTRICITY AND MASS UNBALANCE" AND "IF NECESSARY BALANCING IS TO BE DONE" RS: IF FREQUENCY (1 *RPM) IF FREQUENCY (2*RPM) AND 1X AMPLITUDE > 2X AMPLITUDE AND PHASE ANGLE- 180 AND NARROW BAND ENVELOP AND DOMINANT PLANE RADIAL AND AMPLITUDE AT PT1 > (BASE_VALUE_AT_P 1.20) THEN SHAFT MISALIGNMENT 116: IF SHAFT MISALIGNMENT THEN "MISALIGNMENT OF THE SHAFT IS DETECTED" AND "CHECK ALIGNMENT" 117: IF FREQUENCY (0.48*RPM) IF FREQUENCY (I*RPM) AND 0.48X AMPLITUDE > 1X AMPLITUDE AND DOMINANT PLANE RADIAL AND AMPLITUDE AT PT1 > (BASE_VALUE_AT_P1" 1.20) THEN ROTOR INSTABILITY R8: IF ROTOR INSTABILITY THEN "ROTOR INSTABILITY DETECTED" AND "LOWER OR RAISE OIL VISCOSITY" IF FREQUENCY (n*RPM) AND DISCRETE ENVELOP= AND AMPLITUDE AT PT1 > (BASE_VALUE_AT_P 1.20) THEN GEAR BOX PROBLEM RI0: IF FREQUENCY (I*RPM) AND FREQUENCY (2*RPM) AND FREQUENCY (3*RPM) AND AMPLITUDE AT PT1 > (BASE_VALUE_AT_P 1.20) AND DECREASE IN CRITICAL SPEED THEN CRACK IN THE ROTOR Rll: IF FREQUENCY (1 *RPM) AND AMPLITUDE AT PT1 > BASE_VALUE_AT_P 1.20) AND SPLIT IN CRITICAL SPEED AND ONE CRITICAL SPEED INCREASES AND A NOTHER DECREASES THEN GYROSCOPIC EFFECT RI2: IF GEAR BOX PROBLEM THEN "PROBLEM WITH GEAR BOX DETECTED" AND "CHECK FOR RUN OUT ECCENTRICITY AND CHECK ALIGNMENT" R13: IF FREQUENCY (I*RPM) AND FREQUENCY (2*RPM) AND FREQUENCY (3*RPM) AND FREQUENCY (4*RPM) AND 2X AMPLITUDE > 1X AMPLITUDE AND AMPLITUDE AT PT1 > (BASE_VALUE_AT_P1" 1.20) THEN LOOSENESS R14: IF LOOSENESS THEN "PROBLEM WITH LOOSENESS OF PARTS DETECTED" AND "CHECK FOR LOOSENESS OF BEARINGS, PEDESTAL, FOUNDATION ETC." AND "CHECK ALIGNMENT" R15: IF FREQUENCY (0.25*RPM) AND FREQUENCY (0.33*RPM) AND FREQUENCY (0.66*RPM) AND FREQUENCY (I*RPM) AND DOMINANT PLANE RADIAL: AND AMPLITUDE AT PT1 > (BASE_VALUE_AT_PI* 1.20) AND ACTUAL VALUE OIL TEMP AT PT1 > BASE_VALUE_OIL_TEMP_AT_Pl*l.60) AND COEF FRICTION > 0.1 THEN ROTOR RUB. EENNEERRGGYY MMAATTEERRIIAALLSS Materials Science & Engineering for Energy Systems Economic and environmental factors are creating ever greater pressures for the efficient generation, transmission and use of energy. Materials developments are crucial to progress in all these areas: to innovation in design; to extending lifetime and maintenance intervals; and to successful operation in more demanding environments. Drawing together the broad community with interests in these areas, Energy Materials addresses materials needs in future energy generation, transmission, utilisation, conservation and storage. The journal covers thermal generation and gas turbines; renewable power (wind, wave, tidal, hydro, solar and geothermal); fuel cells (low and high temperature); materials issues relevant to biomass and biotechnology; nuclear power generation (fission and fusion); hydrogen generation and storage in the context of the ‘hydrogen economy’; and the transmission and storage of the energy produced. As well as publishing high-quality peer-reviewed research, Energy Materials promotes discussion of issues common to all sectors, through commissioned reviews and commentaries. The journal includes coverage of energy economics and policy, and broader social issues, since the political and legislative context influence research and investment decisions. SSUUBBSSCCRRIIPPTTIIOONN IINNFFOORRMMAATTIIOONN Volume 1 (2006), 4 issues per year Print ISSN: 1748-9237 Online ISSN: 1748-9245 Individual rate: £76.00/US$141.00 Institutional rate: £235.00/US$435.00 Online-only institutional rate: £199.00/US$367.00 For special IOM3 member rates please email ssuubbssccrriippttiioonnss@@mmaanneeyy..ccoo..uukk EEDDIITTOORRSS DDrr FFuujjiioo AAbbee NIMS, Japan DDrr JJoohhnn HHaalldd, IPL-MPT, Technical University of Denmark, Denmark DDrr RR VViisswwaannaatthhaann, EPRI, USA FFoorr ffuurrtthheerr iinnffoorrmmaattiioonn pplleeaassee ccoonnttaacctt:: Maney Publishing UK Tel: +44 (0)113 249 7481 Fax: +44 (0)113 248 6983 Email: subscriptions@maney.co.uk or Maney Publishing North America Tel (toll free): 866 297 5154 Fax: 617 354 6875 Email: maney@maneyusa.com For further information or to subscribe online please visit wwwwww..mmaanneeyy..ccoo..uukk CCAALLLL FFOORR PPAAPPEERRSS Contributions to the journal should be submitted online at http://ema.edmgr.com To view the Notes for Contributors please visit: www.maney.co.uk/journals/notes/ema Upon publication in 2006, this journal will be available via the Ingenta Connect journals service. To view free sample content online visit: wwwwww..iinnggeennttaaccoonnnneecctt..ccoomm//ccoonntteenntt//mmaanneeyy NN EEWW FFOO RR 22000066 Maney Publishing on behalf of the Institute of Materials, Minerals and Mining International Journal of Aerospace Engineering Hindawi Publishing Corporation http://www.hindawi.com Volume 2010 Robotics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Active and Passive Electronic Components Control Science and Engineering Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Rotating Machinery Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Journal of Engineering Volume 2014 Submit your manuscripts at http://www.hindawi.com VLSI Design Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Shock and Vibration Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Civil Engineering Advances in Acoustics and Vibration Advances in Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Electrical and Computer Engineering Journal of Advances in OptoElectronics Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Sensors Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Modelling & Simulation in Engineering Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Chemical Engineering International Journal of Antennas and Propagation International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Navigation and Observation International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Distributed Sensor Networks International Journal of 
work_kyquynut5jc2bbfsalcv26na4q	----	Microsoft Word - INS_EprintsCoverSheet_Apr08.doc This is the author version published as: This is the accepted version of this article. To be published as : This is the author version published as: Catalogue from Homo Faber 2007 QUT Digital Repository: http://eprints.qut.edu.au/   Ho, T.K. and Wong, K.K. (2003) Peak power demand reduction  under moving block signalling using an expert system. IEE  Proceedings ‐ Electric Power Applications, 150(4). pp. 471‐482.  Copyright 2003 The Institution of Engineering and Technology  1 Peak Power Demand Reduction Under Moving Block Signalling Using Expert System T.K. Ho and K.K. Wong Abstract: The concept of moving blocking signalling (MBS) has been adopted in a few mass transit railway systems. When a dense queue of trains begins to move from a complete stop, they can re-start in very close succession under MBS. The feeding substations nearby will likely be overloaded and the service will inevitably be disturbed unless substations of higher power rating is facilitated. By introducing starting time delays among the trains or limiting the trains’ acceleration rate to certain extent, the peak energy demand can be contained. However, delay is introduced and quality of service is degraded. We present an expert system approach here to provide a supervisory tool for the operators. As the knowledge base is vital for the quality of decisions to be made, this study focuses on its formulation with the balance between delay and peak power demand. 1 Introduction Fixed-block signalling (FBS) has been widely adopted in railway systems for more than a century because of its simple and safety-effective concept of one physical block of track occupied by no more than one train at a time [1-2]. To increase line capacity with FBS, it is possible to have shorter block lengths but the installation and maintenance cost of the signalling and track equipment may not be justified by the increased line capacity. The coarse positional resolution of the trains is another drawback. The demand on the headway becomes so heavy in the metro systems of some major cities that FBS is not able to handle without operating with the full capacity of the infrastructure. Moving- block signalling (MBS) was proposed a few decades ago [3] to provide more room for headway reduction. Theoretically, two successive trains are separated by a distance equivalent to the braking distance for the train behind to brake to a complete stop from its current speed, as well as a safety margin. The separation is reduced to the bare minimum and hence the headway is improved to the limit for the given operating speed and train characteristics, such as train length 2 and braking rate. MBS operations rely on the continuous bi-directional communication links between trains and controllers which can be distributed at track-side locations as well as being centralised. The positional resolution of the trains is therefore much higher than that under FBS. Most successful implementations of MBS systems are not exactly utilising the concept in its original form [4-7]. The communication is not absolutely continuous, but the ‘sampling frequency’ is adequately high to provide near-continuous communication with respect to the maximum train speed. It is regarded as pseudo-moving-block signalling. The communication links are realised by track conductor loops of a certain length. The boundary of two loops is identified by transposition and the loop length defines the resolution of train position, and hence accuracy of speed restriction for the train behind and minimum headway. Because the trains can get closer to each other under MBS, a dense queue may form when the leading train stops at a station, or other reasons, for an unexpected period of time. With the assumptions that traction controllers are not capable of ramping down their demand to create a self-regulated condition and no specific traffic-regulation strategies are imposed from the ATS control centre, the trains behind may start when its separation from the rear of the accelerating train ahead becomes greater than the minimum ATP safety distance. As a result, soon after the leading train has started to move again, the whole queue of trains accelerates in close succession. It is feasible from the operational point of view but the instantaneous increase of power demand will be too much for the power system to bear. The feeding substations in the vicinity may be overloaded. In the worst case, the circuit breakers, set for earth-fault detection, will be tripped and the service will be disturbed as a result. One possible solution is to raise the power ratings of the feeding substation but the additional cost cannot be justified by the fact that overloading is not expected to occur very often and the substations are usually operated well below its ratings. Indeed, the peak demand can be reduced by regulating the train movement and spreading the demand to the adjacent substations. Takeuchi and Goodman have investigated the starting behaviour of a queue of trains within simple metro systems under pure MBS [8]. Two peak demand reduction strategies: starting time delay (STD) by which the starting of each train behind is delayed; and accelerating rate limit (ARL) where the acceleration of each train behind is limited to certain extent, were proposed. Simulation results show ARL achieves better peak demand reduction under certain traffic conditions. Further improvement can be attained when different values of ARL are used for the trains behind according to how far they are at the back of the queue. This study was carried on 3 further to evaluate the combination of STD and ARL [9] with performance indices including peak demand reduction, total energy saving, delay propagation and total arrival time delay. Even though the exclusive use of the graded ARL technique is still suggested to be the best solution to this re-starting problem from the viewpoints of the defined performance indices, the combination of both techniques has the merit of better total energy saving. Power demand and delay are two conflicting criteria in railway operation and the tilt of balance varies with traffic conditions, service demands and time of operation. Unsurprisingly, no single technique can provide the optimal solution for this re-starting problem in general. To complicate matter further, MBS has so far been applied in metro systems in which only a single type of trains serves the passengers and the track topology is simpler. Some new mainline systems, where mixed traffic must be allowed and complex rail network is needed, are investigating the possibility of using MBS. The West Rail of KCRC in Hong Kong is one example. Because different types of trains have their own equipment characteristics and operation schedules, more variables are introduced to the selection of peak energy demand and delay minimisation techniques to restart a queue of trains. An analytical model to relate various system parameters, such as traction equipment, train distribution, substation locations, to peak energy demand and delay is anything but simple. Intelligently combining different techniques in accordance with the current operational demand is an alternative to maximise the advantages of MBS. This study is to investigate the performance of various techniques and their combinations on peak power demand and delay reduction; and their balance under various operational requirements and traffic conditions by computer simulation. An advisory system is then built to assist the operations, automatic or otherwise, to restart a queue of trains under MBS in either metro or mainline systems while satisfying the current operational requirements and traffic conditions. The time given to find the solution is limited as the trains may start to move very soon. A large amount of parameters are involved in the decision-making process and some may come with uncertainty and ambiguity, expert knowledge and common sense on railway operation will therefore be needed to hasten the solution-searching process. An expert system approach is therefore adopted to develop this advisory system. Expert system is one of the early products from the artificial intelligence and has been used extensively in numerous areas. The deployment of expert systems is not yet very common in railway applications but more examples have emerged in recent years. Successful applications have been found in adapting different fixed-block signalling specifications to a multi-train simulator [10], AC supply system control on 4 an electrified line [11], scheduling and timetabling [12-13], as well as other applications on traffic management [14] and service distribution [15] in road transportation. We describe the application of an expert system to find the appropriate train control measures for re-starting trains under MBS while reducing the peak power demand. The formulation of the knowledge base for the expert system from the consideration of various system parameters will be presented and the performance will be evaluated by computer simulation. 2 Restarting with fixed block and moving block signalling With FBS, the track is divided into a number of sections called blocks and each block cannot be occupied by more than one trains at any one time. The presence of a train within a block, or the occupancy of a block, is detected by means of electrical track circuits or similar form of train detection. With speed-signalled systems, a train may proceed to a vacant block, subject to the speed restriction relating to the number of vacant blocks ahead. The line capacity is thus determined by the minimum number of blocks between two successive trains while they do not interact with each other via the signalling system. The resolution of a train’s position is rather coarse as it depends upon the length of block it occupies. When a train stops for a station or whatever reason, train(s) behind, if any, will stop at least a block away, so is the separation between any two trains further behind. A queue forms but it spreads over a long section of track because of this compulsory block-length separation. Even when the first train starts to move, the second train cannot start until the first one clears the block it was occupying. The same applies to the trains behind. Hence, there is an intrinsic time delay imposed to the starting process of the queue of trains. With this time delay, the train in front has established a certain speed and its power demand may have subsided before the train behind starts and accelerates with the maximum power demand. The instantaneous power demand comes in a number of phases as each train moves on in turn and there is no need for the supply system to cater for the need of the simultaneous starting of a few trains. Besides, when the queue is longer, it is more likely that the power demand is shared by more feeding substations. On the other hand, two trains are only separated by a safety margin when they stop under MBS. A dense queue is resulted if more trains are brought to a stop. The time interval of the starting 5 and then drawing the maximum power among the trains will be very short. More trains on the queue will only add to the toll of the instantaneous power demand. It does not take too many trains to push this sudden rise of power demand to exceed the ratings of feeding substations in the vicinity. The installation and maintenance costs of the substations are directly linked to the power ratings. The provision of higher ratings to solve this particular problem with MBS is not necessarily justified when there may be other possible solutions. It is definitely not a feasible solution for the existing lines which are to be re-signalled to MBS. Takeuchi et al conducted a thorough investigation on FBS and MBS under steady-state and perturbed traffic conditions through mathematical analysis and simulation studies [16]. Pure Moving Block signalling has been proven to produce the best performance in both cases. However, MBS suffers from the inherit problem of high peak demand for a queue of re-starting trains. To tackle this problem, starting time delay (STD) and acceleration rate limit (ARL) have been adopted to reduce the peak demand with certain degree of success [8-9, 16]. They impose delays to successive trains during the re-starting process through manipulating the time schedule and tractive effort respectively. The reduction of peak demand is however at the expense of headway because of the delays introduced, which may reduce the ability of MBS to recover from a disruption. As a result, the relationship between peak demand reduction and headway deterioration under both STD and ARL has to be investigated before their advantages can be fully taken. Nevertheless, an analytical model to link peak demand and delay together may not be a technically viable option because there are so many inter-dependent parameters and uncertainties involved. Some of the parameters are system-related, such as train weight, tractive equipment characteristics, power system ratings and service headway whilst the others vary in different re- starting situations. The numbers of trains on the re-starting queue, mixture of the trains and feeding substation locations may have different implications on the application of STD and ARL. The relative position of the train queue to the substations is another crucial factor. When the trains stop mid-way between two substations, the peak demand will be evenly shared and overloading at the substations may be avoided. It is of course a different scenario when the trains have to start very close to a particular substation. As the re-starting process may start any time after a queue forms, a complicated model, however accurate, is not particularly helpful for the quest of a quick solution, which may be a graded 6 application of STD or ARL or even a combination of both. Human experts of substantial relevant background and experience are capable of making such real-time decisions, just like traffic warden at a road junction or even a signalman in the past. The optimal solution may not be forthcoming all the time from a human expert. A reasonable, sensible and consistent solution is what it is required as a supervisory tool for the operators. While it is impossible to post human experts along the rail line, an expert system approach is employed to emulate the expert’s decision-making capability with a software program. 3 Expert system An expert system is a computer program that utilises knowledge and inference procedures to solve problems which require human expertise in a specific domain of applications [17-18]. For complex systems where simple ‘blind’ search technique is inadequate, knowledge-guided search has emerged and led to the idea of expert system. The knowledge required in an expert system consists of facts and heuristics. The facts are the well-known and widely accepted information or practices of a particular field. The heuristics are some rules of good judgment or good guess for decision making in that field. They are normally less publicised but they usually work well and result to a quicker and/or better solution in most cases. Different experts may have their own sets of heuristic rules based upon their experience in the field. There are two basic components in an expert system, a knowledge base and an inference engine. The domain knowledge required for the expert system is placed in the knowledge base. It may be in the form of rules or other appropriate formats, depending on the knowledge representation. The inference engine is the problem solving strategy which organises and controls the steps taken towards the solution. The most popular paradigms are the top-down (goal driven) and the bottom-up (data driven) approaches [19-20]. It is also desirable to have a user-friendly interface to enable the users to communicate with the system under a comprehensible and comfortable environment. A natural language interface would be particularly helpful for non-expert users. However, a window environment with dialogue boxes, menu-driven controls and sufficient help messages is enough to make the interaction simple and smooth. An explanation module may also be included as an option, allowing users to examine the reasoning underlying the solution given by the expert system. Fig. 7 1 gives an overview of the functional blocks within an expert system. The expert system in this study is developed with Visual Basic and run on IBM compatible PCs. 3.1 Knowledge base Most of the expert systems have the knowledge represented in one of the three forms, production rules, semantic network and predicate logic. They apply a number of rules, graphs and clauses respectively to match a particular pattern denoting the problem to be solved. In this expert system, the most popular format, production rule, has been adopted as it resembles better with the experts’ experience and hence provides a natural representation. The basic principle of production rules is to formulate the relations between the patterns of data presented to the system and the resulting action(s) the system should take. In addition to the knowledge base, the expert system also consists of a global database to keep a record of the problem status and a rule interpreter to decide when and how to apply the rules. The rules are made up of two components, conditions and actions: If 1A & …… & mA , then 1B & …… & nB It means that ‘if the conditions 1A and … and mA are satisfied, then perform the actions 1B and … and nB . The production rules are tried in turn to match the pattern in the global database. The appropriate rule is then executed and the actions taken generate another pattern in the global database. The procedure then repeats until a solution is found. As the number of rules increases, simply examining every rule becomes a tedious job. Meta-rules may be introduced and they can select a particular set of rules to execute next. Meta-rules are distinguished from ordinary production rules that they guide the reasoning towards the solution, rather than performing the reasoning. 3.2 Inference engine An inference engine executes the procedures of applying the knowledge. With the production rules, the inference engine compares the IF part of the rules against known facts in the global database in order to determine if the THEN part, i.e. new facts, can be inferred. In this expert system, the initial facts are the system constraints and operating conditions and the ultimate 8 inferred action is a re-starting policy. The inference strategy adopted is therefore forward chaining with data-driven rule searching. Unlike most conventional software programs in which the algorithms and control of the program flow are mixed together, the knowledge base and inference control within the expert system must be physically and functionally independent so that they can be developed, modified and refined without imposing any limitations or alterations on one another. It is indeed in consistent with how a human expert operates. 3.3 User interface Communication between the users/experts and the expert system is made possible through the user interface. A user supplies the information and data of the problem and then obtains the solution with the aid of the interface. Fig. 2 and 3 give the examples of the input and output interfaces respectively. The former allows the system requirements and operational conditions to be inputted whilst the latter displays the peak power demand reduction measure recommended by the expert system. 4 Acquisition of knowledge base The rules and facts in the knowledge base are the core of the expert system. They are the knowledge usually acquired from human experts. However, human experts are not available in this application because no signal engineer has ever been engaged in this capacity. As a result, the knowledge base has to be established through experiences with practical traffic conditions and knowledge, as well as common sense, of railway operation. In this section, a number of tests are carried out with the aid of a whole system simulator [21] in order to investigate the effects of various system parameters and peak power demand reduction measures on the power demand during the re-starting process. A specific test-bed is used to generate the rules here to demonstrate the application of expert system. Different systems may produce different sets of rules similarly and the rules can be inserted in the knowledge base directly. 4.1 System constraints and operational parameters As illustrated in Fig. 4, a queue of three trains 1T , 2T and 3T is intentionally put to a halt between two substations AS and BS . It is a dc supply system with a nominal voltage of 1.5kV 9 and AS and BS are 7km apart. The peak power demands at AS and BS and the time required for the last train (i.e. 3T ) to clear BS during the re-starting process under the following conditions are attained. The results are then used to formulate the rules and facts in the knowledge base. 4.1.1 Weights of trains: With mixed traffic, trains for different services are running on the same line and the weights they carry may vary significantly. A lighter train can go away from a re-starting queue quicker so that the peak demand may spread over more substations whilst a heavy one may limit the acceleration of the trains behind. This test takes the case of equal- weight on the three trains as a reference and assumes identical traction equipment characteristics. To exaggerate the possible effects of train weight on peak demand, one of the train weights is reduced to half or doubled. The results are summarised in Table 1 and the case of equal weight is used as the reference. The peak demand increases when the weight distribution is uneven (as reflected by the negative values on percentage reduction). If the heaviest train is not the first on the queue, the peak demand goes further up. A heavy train tends to accelerate slowly and the signalling influence from the train in front gradually diminishes. The train then accelerates freely without any limitation and draws the maximum necessary power. It is therefore recommended that a queue of trains with mixed weights should be divided into two so that the heaviest train heads the second queue. A time delay can be introduced between the two queues to spread the peak demand over farther distance, but of course it may impose delays on the trains. Indeed, the heaviest train is usually the one heading a queue in metro systems as more passengers board the train after longer period of waiting. 4.1.2 Re-starting location: The re-starting location of the first train with respect to the two sandwiching substations plays an important role on the distribution of the peak power demand to the substations nearby. In this test, the relative position of 1T to AS and BS is defined by YXK / so that 10  K , as illustrated in Fig. 4. The peak demand distributions on the two substations are shown in Table 2 with K=0 as the reference. As the distribution for the first half of K should be roughly a mirror image of that for the second half, only the cases with K ranging from 0 to 0.5 are given here. 10 Unsurprisingly, imbalance of peak demand on AS and BS is the most apparent when K=0 (reference case) and rapidly subsides when K approaches 0.5. In order to denote the extent of urgency of peak demand reduction according to the relative position of trains within the expert system, the inter-substation distance is divided into a number of regions which are defined in Table 3. More regions may lead to better control resolution, but only when such resolution is required. 4.1.3 Safety margin: Table 4 (the first case taken as the reference) shows that safety margin between trains does not carry significant impact on peak power demand. There is a slight tendency of peak demand reduction when safety margin increases because the trains are farther apart and peak demand is spread around. However, excessive safety margin introduces unnecessary delays to the train service. 4.1.4 Train length: From Table 5 where the case of equal length is taken as the reference, there is no clear relationship between peak power demand and lengths of trains as the trains still draw the maximum power simultaneously during re-starting. Longer trains only make the queue longer and possibly allow more substations to share the peak demand. 4.2 Peak power reduction measures Performance of the two peak demand reduction measures, STD and ARL, are investigated with this 3-train, 2-substation test-bed. As combinations of the two measures provide more control options for the operators, the effects of various combinations on peak demand reduction are also examined. 4.2.1 Starting time delay: Peak demand reduction at AS and BS under different combinations of starting delay times on 2T and 3T , with K=0, are summarised in Table 6. The case of no starting time delay is taken as reference. Peak demand keeps falling with increasing starting time delays and the demand reduction is particularly obvious on AS . While the starting times of the trains 2T and 3T , which are closer to AS , are well separated, the demand on AS does not pile up simultaneously and hence the peak demand reduction on AS is more apparent. However, the overall run-time suffers as a result of extra time delay, which may lead to unwanted deterioration of quality of service. 11 With different values of K, Tables 7, 8 and 9 show similar pattern. As K gets closer to 0.5, the peak demand is already better balanced over AS and BS . Even though further reduction on peak demand is possible, it is justified by the extra time delay. 4.2.2 Acceleration rate limit: Table 10 illustrates the peak demand reduction when the acceleration rate of 2T and 3T are refrained to various fractions of full motoring. The case of all trains at full power is used as the reference. Demand reduction is possible but not as significant as when STD is adopted. However, a reasonable demand sharing between AS and BS is maintained, as indicated by a positive reduction on AS and a negative one on BS . Although the peak demand from each train decreases with subdued acceleration, a train spends more time on motoring before it reaches the maximum permissible speed. It is more likely to have trains motoring simultaneously and hence the accumulative power demand is still considerable. Besides, the acceleration rate cannot be set in practice as freely as in the simulation because of its dependency upon the traction drive system characteristics. The control space of this peak demand reduction measure may be rather limited. On the other hand, the delays introduced to the trains with ARL are not too excessive as the overall run-time is only extended slightly. As illustrated in Tables 11 and 12, further increase of K produces similar results. However, the peak demand sharing is already quite even when K approaches 0.5, substantial reduction on acceleration rates only sees the peak demand shifting toward BS and induces unbalance peak demand. 4.2.3 Combined measures: Combinations of STD and ARL to different extents are applied to various operational conditions and the results for K=0 are given in Table 13. The case of no STD and ARL is used as the reference. Peak demand is reduced considerably in most cases and the demand on the two substations is better balanced. Having attained control on time delay and acceleration rate, the control space becomes two-dimensional. It is therefore more flexible for the operators to exert the necessary actions according to the operational conditions. The only drawback is that the overall run-time is stretched extensively because both measures impose delays to trains. 4.3 Formulation of knowledge base According to the results in previous sections, rules in the knowledge base can be built. The rules are organised in five sets, designated to various functions in this application. Priority is given in 12 the order of when they should be triggered during the process of formulating a solution. Supported by the facts on train positions, number of trains, train characteristics and service requirement, the rule-sets are initiated and the appropriate rules are fired to bring the reasoning toward the possible solution. The 5 rule-sets are classified as in Table 14 and the inference flow is illustrated in Fig. 5. The expert system thus starts with identifying the heaviest train, followed by splitting the queue into two groups according to the total number of trains in the queue and usually the heaviest train leads the second group. Introduction of time delay between the train groups is possible and it is at the discretion of the user. Having allowed the user to indicate preference on either shorter time delay or more peak demand reduction as a result of the recommended measure, the expert system goes on to devise the appropriate measures within the ‘Time-delay Assignment’ and ‘Acceleration-rate Assignment’ rule-sets. The ‘Time-delay Assignment’ and ‘Acceleration-rate Assignment’ rule-sets are divided into a number of classes and groups, which contain various extents of time-delay and acceleration-rate, in order to suit operation conditions and user requirements on time delay and peak demand reduction. They can be updated and/or deleted flexibly with respect to any changes in the system conditions and operation requirements. 5. Results and discussions 5.1 Testing This section demonstrates the functions and versatility of the expert system by putting it through similar traffic conditions with different operational requirements. The tests have been undertaken in the same set-up as for the test-bed because the rules are generated from the ‘experiences’ there. Demand reduction percentages at the substations and overall run-time are the basic performance indicators. The performance of the expert system has been investigated when the operational concern focuses on peak power demand reduction, time delay reduction or both, as envisaged in three tests here. There are three levels of time delay and peak demand reduction: high, medium and low, to allow the user to indicate the requirement on the two performance concerns. They will be used to determine the extents of application of any re- starting measures. The case of K=0 for the leading train with no action from expert system, 13 which is the worst case with highly unbalance demand on AS and BS (case A in the tests), is taken as the reference. The operational requirements of the three tests are listed below and the results of various cases are given in Tables 15, 16 and 17 respectively. Additional time delay can be introduced between train groups by the user if the train queue is divided into two groups and more. The results also show its impact. As the reduction of accelerating rate of a train depends upon traction equipment, it may not be set to any arbitrarily low value. The user may specify such a threshold that the acceleration rate should not fall below whenever ARL measure is adopted. Test 1 Performance concern: Peak demand reduction Level of peak demand reduction: High Priority to the heaviest train: No Total number of trains in the queue: 5 Number of trains in group (if grouping required): 3 Test 2 Performance concern: Time delay reduction Level of time delay reduction: High Priority to the heaviest train: No Total number of trains in the queue: 6 Number of trains in group (if grouping required): 3 Test 3 Performance concern: Both peak demand & time delay reduction Level of peak demand reduction: High Level of time delay reduction: High Priority to the heaviest train: No Total number of trains in the queue: 5 Number of trains in group (if grouping required): 3 In Test 1, peak demand reduction is the prime concern. The five trains are divided into two groups, with two trains in the first group and three in the second. From the results, only STD measures are recommended and they lead to either peak demand reduction on both AS and BS or a spread of loading from AS to BS (as reflected by a negative peak demand reduction on BS , coupled with a positive reduction on AS ). The overall run-time is however longer as a result. When a time delay is introduced between the two train groups (cases D, E, G and I), the peak and unbalance loadings are further improved in general at the expense of extra time delay. The expert system recommends no action (cases G, H & I) when the train queue re-starts mid-way 14 between AS and BS . The power demand should be quite evenly distributed between the two substations in these cases. On the other hand, the operational focus is placed on time delay reduction in Test 2. The six trains are also divided into two groups, with three trains each. ARL measures are therefore the resulting actions. Peak demand reduction is not evident in most cases but unbalance loading on AS and BS has been alleviated with significant increase of loading on BS . The overall run-time is very comparable in all cases and even the introduction of time delay between the two train groups only imposes slight delay. However, such delay between the two groups (cases B, D, F and H) does not help reduce the peak demand, if not otherwise. Test 3 entertains the two conflicting requirements in time delay and peak power reduction. Various combinations of ARL and STD measures are recommended for different cases. Unsurprisingly, the two ends cannot be met simultaneously and no overwhelming superiority is attained. Compromise has to be made instead. Peak demand reduction is not always possible as in Test 1 and even when it is possible, only a limited extent of reduction is realised. Nevertheless, this lesser concession on peak demand reduction is well compensated by a better sharing of peak demand between AS and BS ; and a more acceptable overall time delay imposed on the trains. The combination of ARL and STD also allows any time delay between the two trains groups (cases B, D, F and H) to achieve further peak demand reduction while keeping the overall time delay down. 5.2 Implementation The expert system is an independent supervisory tool with direct interface with the ATS control centre. It is not safety critical and its recommended actions, if adopted, are safeguarded by the ATP. Two sets of input are required, static and dynamic. The former consists of track layout, substation locations and ratings which can be inserted as facts to the knowledge base; whilst the latter contains traffic conditions and operational requirements which are attained from the ATS via the bi-directional communication links between trains and controllers. As numerical calculation is kept to minimal within the inference engine, a modest microprocessor platform with a fair size of memory for the knowledge base is adequate to satisfy the hardware requirement. An AI-specific programming language or any advanced high-level 15 language can be used for software implementation. An expert system shell will of course quicken up the development process. 6. Conclusions We have presented an expert system approach to reduce peak power demand in a railway system when re-starting a queue of trains under moving block signalling scheme. Based on the studies with STD and ARL measures, their adoption and the extent they should be applied have to be assigned with respect to the system conditions and operation requirements in order to take their full advantages. An expert system has been built to make such assignments and its knowledge base is derived through the simulated ‘re-starting experiences’ in a trial traffic scenario. As peak demand reduction is often accompanied by an extra time delay imposed on the train queue, the balance between demand reduction and time delay is the key performance indicator. The expert system is then asked to make recommendations in a number of tests with different priorities on performance. The results show that it is capable of providing appropriate advices according to the operational requirements in all cases of the tests. The expert system also enables better sharing of peak demand between substations and holds the balance between peak demand and time delay reduction in response to the user’s request. The re-starting queue of trains is usually divided into groups before the expert system is applied to each group. A deliberate time delay between these train groups provides another possible means to further reduce peak demand. This study reveals the feasibility of applying expert system on this railway operation problem with a simple and small-scale test-bed. A more complicated re-starting case with more trains and system constraints involved should however not hinder the effectiveness of an expert system approach. Even though every railway system is unique on its own characteristics and the re- starting problem under MBS is heavily system-dependent, the same expert system is still a generic supervisory tool for the railway operators because of its intrinsic separation of knowledge and inference. Different system conditions and operation requirements merely mean a change of knowledge base while the basic structure of the expert system remains. Further works therefore include development of an expert system shell, with which any system- dependent knowledge base can be slotted in flexibly. 16 7. Acknowledgements The authors are grateful to the Research Committee of the Hong Kong Polytechnic University for its financial support. 8. References [1] NOCK, O.S., (Ed.): ‘Railway Signalling’, (A. & C. Black, London, 1990). [2] HILL, R.J.: ‘Electrical Railway Traction: Part 4 – Signalling and Interlockings’, IEE Power Engineering Journal, pp. 201-206, August, 1995. [3] PEARSON, L.V.: ‘Moving Block Railway Signalling’, PhD thesis, Loughborough University of Technology, 1973. [4] LOCKYEAR, M.J.: ‘Changing Track – Moving-block Railway Signalling’, IEE Review, pp. 21-25, January 1996. [5] LOCKYEAR, M.J., and NORGROVE, N.: ‘Transmission Based Signalling Systems’, Lecture Notes of IEE 6th Vacation School – Railway Signalling and Control Systems, July, 1996. [6] LOCKYEAR, M.J.: ‘The Application of a Transmission Based Moving Block Automatic Train Control System on Docklands Light Railway’, International Conference on Developments in Mass Transit Systems, pp. 51-61, 1998. [7] COX, G.: ‘A Study of the Three Present Linienzugbeeinflussung (LZB) Systems, their Functionality and the Possibility of Defining a Universal LZB System’, MEng thesis, University of Bath, 1993. [8] TAKEUCHI, H., and GOODMAN, C.J.: ‘Simulation Study of Peak Demand Reduction Strategies When Starting Under Moving Block Signalling’, COMPRAIL’96, vol.2, pp. 187-196, 1996. [9] TAKEUCHI, H., and GOODMAN, C.J., ‘Peak Demand Reduction Techniques When Starting under Moving Block Signalling’, International Conference on Developments in Mass Transit Systems, pp. 280-285, 1998. [10] HO, T.K., ALLAN, J., DIGBY, G., and GOODMAN, C.J.: ‘Modelling of Signalling for an Interactive On-line Rapid Transit Railway Simulator’, Second International Conference on Software Engineering for Real Time Systems, pp. 209-213, 1989. 17 [11] CHANG, C.S., CHAN, T.T., HO, S.L., and LEE, K.K.: ‘AI Applications and Solution Techniques for AC-Railway-System Control and Simulation’, IEE Proc-B, 140(3), pp. 166-176, 1993. [12] IIDA, Y.: ‘Timetable Preparation by AI Approach’, Proceedings of the European Simulation Multiconference, pp. 163-168, 1988. [13] MERCURI, A.: ‘An Expert System for Train Traffic Optimisation’, International Colloquium on Railway Applications of Expert System, 1989. [14] STACK, R.: ‘Boston Central Artery/Tunnel Traffic Management Using An Expert System’, IEEE Conference on Intelligent Transportation Systems, pp. 76-81, 1997. [15] LEVINE, P., and POMEROL, J.C.: ‘Railcar Distribution at the French Railways’, IEEE Expert, pp. 61-69, October 1990. [16] TAKEUCHI, H., GOODMAN, C.J. and SONE, S.: ‘Moving Block Signalling Dynamics: Performance Measures and Re-starting Queued Electric Trains’, IEE Proc.-Electr. Power Appl., to appear. [17] DUDA, R.O., and GASCHING, J.G.: ‘Knowledge-Based Expert System Come of Age’, Byte, pp. 238-278, Sept. 1981. [18] JACKSON, P.: ‘Introduction to Expert System’, (Addison Wesley, 1999). [19] GEVARTER, W.B.: ‘Artificial Intelligence, Expert System, Computer Vision and Natural Language Processing’, (Noyes Publications, 1984). [20] LUCAS, P., and VAN DER GAAG, L.: ‘Principles of Expert Systems’, (Addison Wesley, 1990). [21] HO, T.K., MAO, B.H., YUAN, Z.Z., LIU, H.D., and FUNG, Y.F., ‘Computer Simulation and Modelling in Railway Applications, Computer Physics Communications, 143(1), pp. 1-10, 2002. 18 User Interface Inference Engine Knowledge Base Explanation module User Fig. 1 Structure of an expert system Fig. 2 Input interface 19 Fig. 3 Output interface SA SB T3 T2 T1 Y X Direction of Travel Fig. 4 Locations of trains and substations 20 Identify the heaviest train and divide the queue into groups, if necessary, so that the heaviest train heads a train Facts Determine if peak demand reduction measure is required Facts Identify the dominant performance factor End Determine levels of time delay applying to trains Determine acceleration rate reduction applying to trains Facts FactsFacts Yes No Fig. 5 Inference flow within the expert system 21 Ratio of train weight T1:T2:T3 Peak demand reduction on SA (%) Peak demand reduction on SB (%) Time for T3 to clear SB (min) 1:1:1 0 (6122.3 kW) 0 (3545.5kW) 13:35 2:1:1 -7.4 -10.4 14:06 0.5:1:1 -8.5 14.9 13:22 1:2:1 -14 -12.4 13:34 1:1:2 -15.5 -7.6 13:35 Table 1 Peak demand reduction with different train weights K Peak power demand reduction on SA (%) Peak demand reduction on SB (%) 0 0 (6028.5 kW) 0 (3235.2 kW) 0.178 -2.3 -17.2 0.338 7.6 -21.9 0.5 26.6 -40.2 Table 2 Peak demand distribution with relative positions of the trains K Region 0.45 < K  0.55 1 0.05 < K  0.25 2 0.25 < K  0.45 3 0.55 < K  0.75 4 0.75 < K  0.95 5 0 < K  0.05 6 0.95 < K  1 7 Table 3 Regions between two substations Safety margin (m) Peak demand reduction on SA (%) Peak demand reduction on SB (%) Time for T3 to clear SB (min) 20 0 (6086 kW) 0 (3530 kW) 11:49 40 -0.4 12.4 11:52 60 1.6 7.1 11:54 80 1.5 7.5 11:56 100 2 7.1 12:00 120 -10.7 -1.3 12:04 140 1.8 17 12:05 160 3.2 17.3 12:07 180 3.6 17.9 12:10 200 4 18.5 12:12 Table 4 Peak demand with different safety margins 22 Ratio of train length T1:T2:T3 Peak power demand reduction on SA (%) Peak power demand reduction on SB (%) Time for T3 to clear SB (min) 1:1:1 0 (6122.3 kW) 0 (3545.5 kW) 12:45 0.5:1:1 1.88 -6.2 12:40 1:0.5:1 -0.35 -15.5 12:38 1:1:0.5 -0.2 1.3 12:42 Table 5 Peak demand with different train lengths Starting time delay on T2 and T3 (sec) Peak power demand reduction on SA (%) Peak power demand reduction on SB (%) Time for T3 to clear SB (min) 0-0 0 (6028.5 kW) 0 (3235.23 kW) 11:53 10-10 -17.81 16.98 11:58 10-20 7.27 17.51 11:56 10-30 7.27 4.62 11:57 10-40 7.27 -2.62 11:59 10-50 5.86 -4.13 12:04 10-60 7.27 -1.62 12:15 20-20 2.98 1.86 11:56 20-30 10.53 -0.98 11:58 20-40 16.68 1.2 12:05 20-50 18.65 2.08 12:15 20-60 22.9 -3.73 12:25 30-30 12.5 4.58 12:05 30-40 17.96 6.36 12:15 30-50 22.38 0.11 12:25 30-60 25.89 -2.81 12:35 40-40 19.82 3.13 12:25 40-50 24 -2.81 12:35 40-60 31.76 13.71 12:45 50-50 33.89 16.63 12:45 50-60 33.89 29.74 12:55 Table 6 Peak power demand reduction with STD (K=0) 23 Starting time delay on T2 and T3 (sec) Peak power demand reduction on SA (%) Peak power demand reduction on SB (%) Time for T3 to clear SB (min) 0-0 0 (6169.5 kW) 0 (3793.1 kW) 11:54 10-10 -15 -14.6 11:56 10-20 13.9 13.8 11:55 10-30 13.9 6.6 11:56 10-40 1.3 4.4 11:56 10-50 7.7 6 12:04 20-20 6 6.4 11:56 20-30 13 5.9 11:57 20-40 14 7.2 12:04 20-50 18 8.3 12:14 30-30 13.6 10 12:04 30-40 17.7 11 12:14 30-50 20.9 7.3 12:24 40-40 19.1 9.5 12:24 40-50 24.7 10.8 12:34 50-50 37.2 24.3 12:44 Table 7 Peak power demand reduction with STD (K=0.178) Starting time delay on T2 and T3 (sec) Peak power demand reduction on SA (%) Peak power demand reduction on SB (%) Time for T3 to clear SB (min) 0-0 0 (5570.3 kW) 0 (3944 kW) 11:55 10-10 -18.9 -22 11:56 10-20 11.4 10.8 11:56 10-30 11.4 -0.7 11:56 10-40 6.1 -9.1 11:58 10-50 4.8 -7.4 12:04 20-20 3.4 1 11:56 20-30 9.8 1.1 11:58 20-40 16.6 0.2 12:04 20-50 16.5 -2.4 12:14 30-30 11.5 3.4 12:04 30-40 16 0.9 12:14 30-50 19.7 -3 12:24 40-40 17.6 -1.3 12:24 40-50 21.1 -2.5 12:34 50-50 35.7 15.9 12:44 Table 8 Peak power demand reduction with STD (K=0.338) 24 Starting time delay on T2 and T3 (sec) Peak power demand reduction on SA (%) Peak power demand reduction on SB (%) Time for T3 to clear SB (min) 0-0 0 (4427.5 kW) 0 (4534.5 kW) 11:54 10-10 -20.4 -19.7 11:56 10-20 9.8 6.3 11:56 10-30 9.8 -0.9 11:57 10-40 6.3 -12.5 11:57 10-50 5.1 -11.2 12:04 20-20 3.8 1.4 11:57 20-30 9.7 1.9 11:58 20-40 16.1 -2.8 12:04 20-50 16 -4.9 12:14 30-30 11.3 -0.3 12:04 30-40 15.6 -2.2 12:14 30-50 19.1 -5 12:24 40-40 17.1 -3.8 12:24 40-50 20.4 -4.6 12:34 50-50 35.1 16 12:44 Table 9 Peak power demand reduction with STD (K=0.5) Acceleration factors on T2 and T3 Peak power demand reduction on SA (%) Peak power demand reduction on SB (%) Time for T3 to clear SB (min) 1-1 0 (6122.3 kW) 0 (3545.5 kW) 11:55 0.9-0.81 2.9 -7.4 11:54 0.9-0.72 3.4 -8.4 11:56 0.9-0.63 8.2 -2.5 11:53 0.9-0.54 5.4 -8.5 11:55 0.9-0.45 8 -11.7 11:55 0.8-0.64 5 -12.5 11:55 0.8-0.56 4.2 -7.1 11:54 0.8-0.48 7.2 -15.6 11:57 0.8-0.4 12.3 -16.1 11:56 0.7-0.49 4.1 -22.5 11:59 0.7-0.42 3.5 -25.1 11:58 0.7-0.35 0.1 -28.6 12:01 0.6-0.36 5.7 -23.8 11:58 0.6-0.3 -10.2 -19.6 11:56 0.5-0.25 1.2 -43.4 12:10 Table 10 Peak power demand reduction with ARL (K=0) 25 Acceleration factors on T2 and T3 Peak power demand reduction on SA (%) Peak power demand reduction on SB (%) Time for T3 to clear SB (min) 1-1 0 (5554 kW) 0 (3973 kW) 11:55 0.9-0.81 8 -10.4 11:55 0.9-0.72 7.2 -11.3 11:54 0.9-0.63 11.6 -11.5 11:56 0.9-0.54 10.7 -9.3 11:56 0.9-0.45 14 -14.1 11:55 0.8-0.64 5.1 -15.3 11:56 0.8-0.56 7.6 -12.8 11:55 0.8-0.48 4 -20.4 11:57 0.8-0.4 11.8 -19.9 11:56 0.7-0.49 11 -25.2 11:59 0.7-0.42 12.7 -27.2 11:58 0.7-0.35 14.1 -31.5 12:01 0.6-0.36 19.3 -32.5 11:58 0.6-0.3 3 -27.1 11:56 0.5-0.25 15.1 -49.5 12:10 Table 11 Peak power demand reduction with ARL (K=0.338) Acceleration factors on T2 and T3 Peak power demand reduction on SA (%) Peak power demand reduction on SB (%) Time for T3 to clear SB 1-1 0 (4410.9 kW) 0 (4560 kW) 11:54 0.9-0.81 8.9 -4.9 11:55 0.9-0.72 8.3 -4.1 11:56 0.9-0.63 11.6 -6.7 11:55 0.9-0.54 11.1 -7.7 11:55 0.9-0.45 14.2 -12.4 11:57 0.8-0.64 4.8 -8.6 11:56 0.8-0.56 8 -8 11:57 0.8-0.48 3 -20.3 11:57 0.8-0.4 12.3 -15.1 11:55 0.7-0.49 10.4 -17.2 11:59 0.7-0.42 12.2 -20.6 11:58 0.7-0.35 14 -24.9 12:01 0.6-0.36 18.6 -29.9 11:59 0.6-0.3 0.2 -27.7 11:56 0.5-0.25 14.1 -41 12:10 Table 12 Peak power demand reduction with ARL (K=0.5) 26 Combined measures Peak power demand reduction on SA (%) Peak power demand reduction on SB (%) Time for T3 to clear SB (min) STD ARL 0-0 1-1 0 0 11:55 10-10 0.9-0.81 -5.9 -12.6 11:56 10-20 0.9-0.72 13.4 10.7 11:56 10-30 0.9-0.63 32.9 20.1 12:58 10-40 0.9-0.54 12.8 1.85 12:11 10-50 0.9-0.45 13.4 13.2 12:31 20-30 0.8-0.64 14.7 4.47 12:05 20-40 0.8-0.56 35.9 31.82 13:26 20-50 0.8-0.48 30.1 14.6 12:38 30-40 0.7-0.42 29.9 18.7 12:45 30-50 0.7-0.35 27.9 20.6 13:05 Table 13 Peak power demand reduction with combined STD and ARL Rule-set Fact(s) required Priority Train-weight identification & train grouping - Weight of each train - Number of trains in a queue 1 Strategy-activation - Restarting location (i.e. K) of the first train 2 Performance selection - Headway or timetable - Power system rating 3 Time-delay assignment - K of the first train - Level of quality of service required 4 Acceleration-rate assignment - K of the first train - Level of power reduction required 4 Table 14 Classifications of the rule-sets Case K of the 1st train Expert system recommendations Time delay between train groups (sec) Peak demand reduction on SA (%) Peak demand reduction on SB (%) Time for last train to clear SB (min) A 0 N/A 0 0 (7213.4kW) 0 (4279.9kW) 14:48 B 0 1st group: no action 2nd group: time delay C4 0 5.1 5.7 15:24 C 0.25 1st group: no action 2nd group: time delay B3 0 3.7 -3.1 14:48 D 0.25 1st group: time delay B3 2nd group: time delay B3 20 9.7 -9.5 15:05 E 0.25 1st group: time delay B3 2nd group: time delay B3 40 12 0.35 15:24 F 0.3 1st group: no action 2nd group: time delay A2 0 7.6 -6.6 14:48 G 0.3 No action recommended 60 14.7 -2.4 15:22 H 0.5 No action recommended 0 24.3 -35.3 14:48 I 0.5 No action recommended 10 24.3 -33.9 14:53 Table 15 Recommended actions from expert system and their results in Test 1 27 Case K of the 1st train Expert system recommendations Time delay between train groups (sec) Peak demand reduction on SA (%) Peak demand reduction on SB (%) Time for last train to clear SB (min) A 0 N/A 0 0 (7213.4kW) 0 (5087.6kW) 17.05 B 0 1st group: acc rate C4 2nd group: acc rate C4 20 -18.4 -15 17.13 C 0.25 1st group: no action 2nd group: acc rate B3 0 3.7 13.3 17:04 D 0.25 1st group: acc rate B3 2nd group: acc rate C4 20 -37.3 -26.4 17:16 E 0.3 1st group: no action 2nd group: acc rate C4 0 -19.4 -4.7 17:06 F 0.3 1st group: time delay A2 2nd group: time delay C4 20 -12.7 -25.8 17:11 G 0.5 1st group: no action 2nd group: acc rate B3 0 8.7 -14.4 17:06 H 0.5 1st group: no action 2nd group: acc rate B3 20 3.3 -21.4 17:09 Table 16 Recommended actions from expert system and their results from Test 2 Case K of the 1st train Expert system recommendations Time delay between train groups (sec) Peak demand reduction on SA (%) Peak demand reduction on SB (%) Time for last train to clear SB (min) A 0 N/A 0 0 (7212.7 kW) 0 (4279.9 kW) 14:48 B 0 1st group: time delay C4 & acc rate C4 2nd group: time delay C4 & acc rate C4 0 10.3 -6.06 16.52 C 0.25 1st group: time delay B3 & acc rate B3 2nd group: time delay B3 & acc rate B3 0 -3.68 -3.1 14:48 D 0.25 1st group: time delay B3 & acc rate B3 2nd group: time delay B3 & acc rate B3 30 -0.13 -10.37 16:06 E 0.3 1st group: time delay A2 & acc rate A2 2nd group: time delay C4 & acc rate C4 0 12.13 -47.3 15:41 F 0.3 1st group: time delay A2 & acc rate A2 2nd group: time delay C4 & acc rate C4 30 20.29 -20.46 15:44 G 0.5 1st group: no action 2nd group: time delay B3 & acc rate B3 0 10.74 -31.72 15:42 H 0.5 1st group: no action 2nd group: time delay B3 & acc rate B3 30 11.17 -38.54 15:40 Table 17 Recommended actions from expert system and their results from Test 3 
work_kz5fdhrbfvhljh6q7vxobdhzk4	----	Microsoft Word - BDS_final version.docx VISITORS OF TWO TYPES OF MUSEUMS: A SEGMENTATION STUDY Juan Gabriel Brida; Marta Disegna*; Raffaele Scuderi School of Economics and Management JuanGabriel.Brida@unibz.it; marta.disegna@unibz.it; Raffaele.Scuderi@unibz.it Free University of Bozen-Bolzano Universitätsplatz 1 - piazza Università, 1 39100 Bozen-Bolzano – Italy Abstract Market segmentation comprises a wide range of measurement tools that are useful for the sake of supporting marketing and promotional policies also in the sector of cultural economics. This paper aims to contribute to the literature on segmenting cultural visitors by using the Bagged Clustering method, as an alternative and effective strategy to conduct cluster analysis when binary variables are used. The technique is a combination of hierarchical and partitioning methods and presents several advantages with respect to more standard techniques, such as k-means and LVQ. For this purpose, two ad-hoc surveys were conducted between June and September 2011 in the two principal museums of the two provinces of the Trentino-South Tyrol region (Bolzano and Trento), Northern Italy: the South Tyrol Museum of Archaeology in Bolzano (ÖTZI), hosting the permanent exhibition of the “Iceman” Ötzi, and the Museum of Modern and Contemporaneous Art of Trento and Rovereto (MART). The segmentation analysis was conducted separately for the two kinds of museums in order to find similarities and differences in behaviour patterns and characteristics of visitors. The analysis identified three and two cluster segments respectively for the MART and ÖTZI visitors, where two ÖTZI clusters presented similar characteristics to two out of three MART groups. Conclusions highlight marketing and managerial implications for a better direction of the museums. Keywords: Bagged clustering, Logit models, Museum, Segmentation, Motivation. *Corresponding Author 1. Introduction Museums are the most popular cultural attractions, usually followed by art galleries and monuments (McKercher, 2004). For a long time visitors of cultural attractions were treated as a homogeneous mass of people. The tendency of the recent tourism literature is instead to consider them as a heterogeneous market with different characteristics, perceptions and needs (Hughes, 2002). Brida et al. (2012) showed that visitors of Christmas Markets in Northern Italy clustered into three groups according to a set of motivational factors that drove them to make the visit. Other studies showed that tourists who visited art museums presented different socio-demographic characteristics (in particular regarding the level of education, income and occupation) than those who engaged in festivals, musical activities, theme parks, amusements parks, local fairs, and events (Kim et al., 2007; Bennett, 1994; Schuster, 1991). Most research on tourism considered different types of museums (like art museum, stamps, history, science, and even children’s museums) as a unique cultural attraction with the same “label”. However, MacDonald and Alsford (1995) suggested they are heterogeneous, by affirming that “all museums are products of their particular cultural and historical experiences”. Each museum exhibits its peculiarity by offering visitors different kinds of involvements (Dicks, 2003) and experiences, which are suitable for different kinds of tourists. Furthermore, an art museum, a history museum, an opera, or an outdoor festival might produce different experiences in visitors (Stylianou-Lambert, 2011). For these reasons research should analyse cultural attractions, and in particular museums, separately according to the subject matter and the experiences that they offer (Stylianou-Lambert, 2011). Profiling museum visitors by taking into consideration also the different characteristics of the museums can be of crucial importance for managers and marketing analysts of the museum. Identifying homogeneous clusters of consumers-visitors can be in fact an essential step for planning and developing appropriate strategies, in order to satisfy the needs of each segment of guests. In this context clustering proposes a set of widely used unsupervised techniques with the aim to discover hidden associations among statistical units and identifying segments (Saarenvirta, 1998). Given a set of selected segmentation variables, these methodologies aggregate the units in groups, in such a way that each aggregation contains the most similar units, and at the same time is dissimilar from the remainder. The supervision means that “membership of data points which can illustrate the general structure of the group is required in order to derive the classification rules”, therefore the supervision implies that there is no rule for initiation of classification (Budayan, Dikmen & Birgonul, 2009). This implies that the empirical distribution and characteristics of the data will determine the cluster membership. Since the introduction of market segmentation in the late 1950s, the number and type of approaches for segmentation has grown enormously (Liao, Chu, & Hsiao, 2012; Dolnicar & Leisch, 2004). Unfortunately, as emphasized by many researchers, no absolutely “correct” way to segment a market exists in the literature (Brida, Disegna, & Osti, 2012; Kotler, Bowen, & Makens, 2010; Dolnicar et al., 2008; Beane & Ennis, 1987). On the contrary, the researcher intervenes in different moments of the estimation process, which of course involves the final results. This implies that “clustering is exploratory data analysis and different methods present different views of data” (Leisch, 2006). The degrees of freedom in the clustering algorithm concern, among other things, the variables selection, the choice of a measure of dissimilarity between units, the final number of clusters, the test of the clustering solution as not purely random, the interpretation of final results for addressing management and marketing. Moreover, one has to bear in mind that “in the case of no clear cluster structure there is no “correct” solution” (Leisch, 2006). The most popular clustering techniques are partitioning and hierarchical methods. The standard partitioning procedures aim to group the observations around a centre in order to find a segmentation of a set of units in an a priori fixed number of clusters. In the marketing and tourism literature k-means is the most commonly used algorithm that falls into this category. Hierarchical methods instead obtain the final clusters solution by repeatedly joining the “closest” clusters composed of one or more observations (agglomerative clustering), or repeatedly splitting the “further” clusters (divisive clustering). This study instead makes use of Bagged Clustering, which combines both hierarchical and partitioning methods. It was proposed by Leisch (1999) and has the advantages of overcoming many of the limitations of the two methods. This method has been used successfully in the past by Leisch himself or his research team, for the sake of tourism market segmentation (Dolnicar & Leisch, 2000, 2003; Dolnicar et al., 2008), but it has been applied infrequently by other researchers in the same field or in others (Huang, Chang, & Wu, 2009). Its application to the field of culture aims to study the profiles of tourists with respect to their motivations in visiting two different types of museums. This can shed light on investigating whether museums offering different experiences are visited by heterogeneous types of tourists, or on the contrary if segments with common characteristics can be detected. This objective is pursued by using a dataset from ad- hoc surveys. These were conducted from June to September 2011 in the two main museums of Trento and Bolzano, the two provinces of the Trentino-South Tyrol region. The South Tyrol Museum of Archaeology (shortened to ÖTZI) is located in the Province of Bolzano and hosts the permanent exhibition of the mummy Ötzi, “the Iceman”, whereas the Museum of Modern and Contemporary Art (shortened to MART) is places in the province of Trento and owns one of the most important collections in Italy for what concerns this artistic period. The article will first proceed by outlining the research objectives, overviewing the clustering technique adopted, presenting the sample and questionnaire employed, and discussing the clustering results combined with binary and multiple Logit analysis. Both academic and practical implications, limitations of the research and future perspectives are provided. 2. Research Objective The focus of this paper is to find and describe groups of visitors with similar motivational characteristics in visiting an archaeological and a modern and contemporary art museum. This work constitutes a first attempt in studying whether there can be detected heterogeneous profile of visitors in two different types of museums. The set of motivational factors for segmentation were measured as binary variables, i.e. “Yes/No”, in the ad-hoc survey used in this study. When binary data are used for the sake of clustering observations, it is a common practice in literature to use one of the following approaches: applying a hierarchical clustering method using a dissimilarity measures, such as Jaccard, Russell/Rao, Matching, or Dice, computed on the original data (Řezanková, 2009; Finch, 2005); applying the k-means method using the Euclidean measure on the original data (Leisch, 2006); transforming the binary variables into continuous ones through a Factor Analysis, like the Correspondence Analysis, and then use the results as input of a clustering method using the Euclidean distance (Bouguila, 2010). When k-means is applied on the original data, the centres give the conditional marginal probabilities of observing a “1” (i.e., “YES” in one of the segmentation variables) given the cluster membership. It is important to underline that the dissimilarity measures for binary data have been extensively used and analysed with hierarchical methods but not with partitioning one, in which these types of measures are less common. In this context the Bagged Clustering method can be viewed as a useful solution for two reasons: it allows segmenting visitors by using the original binary data, and it overcomes the main limitations of the traditional segmentation methods. 3. Methodology In this study, the Bagged Clustering method proposed by Leisch (1999) was adopted. This method is a combination of partitioning and hierarchical procedures and consists of the following steps: 1. First of all, B bootstrap sample !!! ,… ,!!! were constructed by drawing with replacement from the original sample !!, were N is the sample size. 2. A partitioning method, called base method, is chosen by the researcher (e.g., k-means) and applied to each bootstrapped sample. From this procedure, !×! centres !!!,… ,!!! ,!!!,… ,!!!,!!!,… ,!!! are obtained, where K is the number of centres used in the base clustering method and !!! is the jth centre (! = 1,… ,!) of !!! , which is the ith bootstrap sample (! = 1,… ,!). 3. All the centres are combined into a new dataset !!×!. 4. A hierarchical cluster algorithm is applied to the !!×! dataset in order to produce a partition of the centres. 5. The final outcome is displayed through the usual dendrogram of classical hierarchical methods, where the best partition of centres is obtained by simply investigating it. Finally, the partition of the original observations results from assigning the ! ∈ !! observations to their closest centres. In this way each observation is assigned to the cluster containing the centre to which it is associated. Figure 1 schematically represents the steps that characterize the Bagged Clustering. <INSERT FIGURE 1 BY HERE> This method can be interpreted as both a complexity-reducing pre-processing stage for the hierarchical methods and a combination procedure of several partitioning results (Kang, Zhang, & Fan, 2008; Leisch, 1999). It has a better performance in comparison to other standard clustering methods for both continuous and binary data sets (Leisch, 1999). Furthermore, the Bagged Clustering technique overcomes many limitations of both partitioning and hierarchical algorithms. Partitioning methods are more flexible and perform better with large dataset than hierarchical methods (Everitt et al., 2011). The latter have the disadvantage that once observations are merged with others in a group, they cannot be removed from that cluster. However, many partitioning algorithms depend strongly on the starting selected centres because they are based on iterative stochastic procedures. Thus running the k-means algorithm twice on the same dataset with different starting centres may result in two different solutions, for the less clear the hidden data structure, the higher the difference between the two solutions. From this point of view k-means is an unstable algorithm, though widely used. The reason is related to the possibility of finding at each run only a local and unstable solution rather than a global one (Jain, 1999). In addition, the latter can be absent. The Bagged Clustering technique instead is more stable and it is less dependent on the starting solution. In fact, running the base method (i.e. the partitioning methods chosen) on B bootstrapped sample is equivalent to running B times the base method, and then obtaining the final solution from summarizing all these results properly. Another limitation in using k-means is that it is necessary to select the number of clusters in advance (Buttrey & Karo, 2002; Jain, 1999). In tourism studies, while using non- hierarchical algorithm it is common practice to decide the number of clusters on the basis of practical and subjective preferences (Choi, 2011; Konu, Laukkanen, & Komppula, 2011; Albalate & Bel, 2010; Pérez & Nadal, 2005) or derive this information from applying a hierarchical cluster method (Claver-Cortés, Molina-Azorín, & Pereira-Moliner, 2007; Bigné & Andreu, 2004; Chen & Hsu, 1999; Punj & Steward, 1983). Although many internal validity indices were developed in order to overcome this problem and to drive the researchers to select this number properly (see for example Handl, Knowles, & Kell, 2005), none has yet been globally accepted, and in the tourism field they have not been widely applied (see Brida, Disegna, & Osti 2012 for an example of application). Furthermore, as underlined by Vesanto & Alhoniemi (2000), in practice the value of these indices must be interpreted as a mere guideline. The use of the Bagged Clustering algorithm overcomes also the problem of selecting the “optimal” number of groups. Although an initial choice of a number of groups is required, it does not affect the final results. The “final” number of clusters is in fact obtained a posteriori as a result of the hierarchical algorithm. Finally, only when binary data is used the interpretation of the results from Bagged Clustering is easier and more exhaustive than traditional methods. In fact, a k-means algorithm applied to a binary dataset produces K centres, one per each segment. Each centre is a d-dimensional vector (where d denotes the number of variables used for segmentation) in which each value can be interpreted only in term of the frequency (i.e. mean value) with which the value 1 occurred among the units observed in the selected segment for each variable. Each segment created by the Bagged Clustering algorithm is composed by several d-dimensional centres. Therefore the empirical distribution of the centres can be easily checked and represented through the standard Box-plot for each variable and segment. 4. Data and structure of the questionnaire 4.1 The museums The research involved the two main museums of Trentino-South Tyrol, an Italian region located in the North-East. The ÖTZI museum is situated in Bolzano, the main city of South Tyrol. It hosts the permanent exhibition a mummy from the Neolithic period Ötzi. The mummy was discovered in the region in 1991 as one of the oldest mummies in the world. The good preservation status of it and its belongings has attracted researchers and visitors from around the world, and made the museum the most important cultural attraction in Bolzano. The MART museum, that is the second museum under consideration, is placed in Trento and Rovereto, the two main cities of the province of Trento. It hosts both a permanent collection of modern art, where works are displayed on a rotating basis, and a temporary exhibition. It has the most important collections in Italy of modern and contemporary art works, in particular futurism. As pointed by Brida, Pulina, and Meleddu (2012), the idea of a museum for modern and contemporary art was born in the late 1970s, against the background of industrial and unemployment crisis. 4.2 Research design The research is based on a survey conducted from June to September 2011 among the visitors of the ÖTZI and MART museums. A total of 1,288 interviews were successfully collected almost equally divided between MART (46%) and ÖTZI (54%). In order to encourage cooperative behaviour respondents were informed that the research had exclusively scientific aims, and that impartiality in the data analysis was guaranteed. Furthermore, a pilot survey was carried out for testing the questionnaire before conducting the full survey, in order to avoid biases related to its structure and wording. Interviews were held to visitors exiting the museums after their visit, in selected working and weekend days of the four months analysed, and during different times of the day. Only one person per travel party was selected. The questionnaires were anonymous and self- administered in three languages (Italian, German and English), though a research team member was present to respond if questions or doubts emerged. A convenience sampling method (Cochran, 1977) was used, as there was no sufficient information on the characteristics of museums visitors in order to apply a probabilistic design. The sections of the questionnaire are presented schematically in Table 1. <INSERT TABLE 1 BY HERE> 5. Discussion of the results As reported above, the set of variables for segmenting interviewees referred to their motivations to visit each of the two museums. The questionnaire asked the respondents if they agreed or not (dichotomous answer) with a set of push factors that motivated the visit to the museum. The set of factors included: satisfying a curiosity (“curiosity”), resting/relaxing (“relax”), a specific interest in such an attraction (“interest”), accompanying friend/family member with a specific interest in such an attraction (“friend”), learning something new (“learn”), telling friends about the visit (“tell”), doing something that one ought to do (“do”), contributing to preserving this attraction for future generation (“future”), revisiting this museum (“revisit”), showing the museum to friends or relatives (“show”), professional or academic reasons (“work”), doing something worthwhile (“worthwhile”), occupying some leisure time (“leisure”), visiting temporary exhibition (“temporary”), seeing the building (“building”). The Bagged Clustering algorithm considered the k-means as base method, with K=20 centres and 10,000 iterations was used as base method. A number of B=50 bootstrap samples were considered, resulting in a total of 1,000 !×! centres, which were then hierarchically clustered using Euclidean distance and Ward’s agglomerative linkage method. These parameters were chosen because they provided the best performances in previous studies, which used simulated artificial datasets with similar characteristics to the one of this paper (Dolnicar & Leisch, 2004, 2000). Results of the Bagged Clustering method for the MART analsys are graphically showed in Figures 2 and 4, whereas Figures 3 and 5 display those of ÖTZI. <INSERT FIGURE 2 BY HERE> <INSERT FIGURE 3 BY HERE> The top part of the graph in Figures 2 and 3 display the dendrograms deriving from the procedure, respectively for MART and ÖTZI. The plot under each dendrogram shows the distance of aggregation for each cluster, where the black line reports standardized absolute hights and the grey one stands for first differences. The accentuated bend in the grey line suggests that for MART the correct number of clusters is three, whereas for ÖTZI it is two. These correspond to cutting the dendrogram where the longest distance between two consecutive aggregations appear. <INSERT FIGURE 4 BY HERE> <INSERT FIGURE 5 BY HERE> The Box-plots in Figures 4 and 5 report the distribution of the centers for each segmentation variable and within each group. In addition the red line that runs across all the Box-plot indicates the sample mean, i.e. the average frequency of “Yes” for each motivational factor. For the sake of interpretation, it is important to emphasize that the higher the height of the box (i.e. the Interquartile Range), the smaller the homogeneity of the segment with respect to the variable considered. This implies that segments are better characterised by those variables presenting low dispersion, and that a strong dispersion within a variable indicates non homogeneity of the units of the segment with respect to that characteristic. The Chi-square test on the centres was calculated in order to test the null hypothesis of independence of each motivation variable from the observed segment. MART results revealed that only telling friends about the visit at the museum (“tell”) is independent from the segments identified (p-value = 0.744), whereas all other ones significantly depend from the observed segments (p-value < 0.01). On the other hand ÖTZI reported that only the following motivational factors depended significantly from the identified segments (p-value < 0.01): satisfying a curiosity (“curiosity”), resting/relaxing (“relax”), accompanying friend/family member with a specific interest in such an attraction (“friend”), learning something new (“learn”), telling friends about the visit (“tell”), doing something that one ought to do (“do”), doing something worthwhile (“worthwhile”), and occupying some leisure time (“leisure”). The Box-plots of MART in Figure 4 revealed the presence of a niche segment (cluster 3, which contains 7.6% of the whole MART sample) and two segments of almost equal size (42.8% and 49.6% of the total visitors of MART respectively in the cluster 1 and 2). Visitors of cluster 3 visited the museum mainly because for the sake of satisfying a curiosity (“curiosity”), learning something new (“learn”) and doing something worthwhile (“worthwhile”). Therefore, these visitors have been named “Knowledge seeker”. Visitors of cluster 2, named “Interested”, seemed to be strongly attracted by the temporary exhibition (“temporary”) and by a specific interest in such an attraction (“interest”). Cluster 1 collected instead all the remainder visitors. They had in common the fact that they declared that their motivation in visiting MART was “not” one of those considered. Moreover, compared to the other two groups, cluster 1 exhibits lower median values in the majority of the factors, which implies that a big part of these visitors did not select the motivation items proposed in the questionnaire. In particular, it resulted that they did “not” visit MART because it is something that one ought to do (“do”), or to contribute to preserve this attraction for future generation (“future”), revisiting this museum (“revisit”), doing something worthwhile (“worthwhile”), occupying some leisure time (“leisure”). Therefore, cluster 1 is labelled “Non-motivated”. Clusters of ÖTZI visitors recall similar segments as those identified for MART. About 25% of respondents are grouped in cluster 1 where the main motivations are satisfying a curiosity (“curiosity”) and learning something new (“learn”), such as the “Knowledge seeker” visitors identified as niche in the MART dataset. If “Knowledge seekers” represented a niche for MART, this segment is bigger for ÖTZI. The remainder 75% was clustered into the “Non-motivated” cluster similarly to MART dataset. This cluster groups visitors that did not report any of the proposed motivations. In particular, they did not visit ÖTZI for the sakes of resting/relaxing (“relax”), telling friends about the experience (“tell”), or, as for MART’s similar group, doing something that one ought to do (“do”), doing something worthwhile (“worthwhile”), occupying some leisure time (“leisure”). 5.1. Clusters description The additional information collected through the survey were used to characterize the clusters identified in each museum in terms of socio-demographic (gender, age, level of education, origin, occupation, and visiting party) and economic (household income, total expenditure per person per night, and expenditure at the shop of the museum per person) variables. Table A1 in the Appendix A reports the complete list of these profiling variables with a brief description of them. <INSERT TABLE 2 BY HERE> Some statistically significant dependency emerged between clusters and the profiling variables for each museum (Table 2). Among the visitors of MART, the “Interested” was on average older (47 years old) and mainly employed, whereas the “Knowledge seeker” was of younger age (39 years old on average) though it reports the highest percentage of retired. “Non-motivated” was instead more in an autonomous worker or in other position than the remainder. The origin of the visitors was main distinctive variable between the two clusters for the ÖTZI museum. The “Knowledge seeker” came mainly from abroad (Germany or other cities), whereas “Non-motivated” were mainly Italian. This result was not surprising because the marketing policy of this museum is actually more oriented to the foreign countries than to Italy. Furthermore, the National Geographic Magazine dedicated it articles written in English or German, thus contributing to promote this museum in a good way. <INSERT TABLE 3 BY HERE> 5.2. Profiling tourists: Logit models Logit models were then estimated in order to assess whether the set of socio-demographic and economic characteristics was a significant predictor of the likelihood to be part of one of the groups. Due to the presence of 3 groups for MART and 2 for ÖTZI, Multinomial Logit and Logit models were respectively adopted. For both museums the “Non-motivated” cluster was used as baseline. Table 3 reports the coefficients for the resulting models. MART’s Multinomial Logit result leads to the conclusion that “Interested” visitors were significantly discriminated from the “Non-motivated” group only because they were willing to spend more at the shop of the MART museum. It is relevant to remind that “Interested” visitors were highly interested in the temporary exhibition (“temporary”) and in the museum as attraction (“interest”). Therefore, from these empirical results one can derive that they were likely to be interested in spending a higher amount of money at the shop, and probably buying souvenirs like books that were the most expensive ones. “Knowledge seeker” were instead more likely to be male, come from the Northern Italy, and with higher levels of education (university degree or postgraduate), whereas in comparison to them the “Non-motivated” cluster was likely to come from abroad. A higher education of knowledge seekers is in line with the main motivations that push these visitors to make the visit: curiosity, learning and doing something worthwhile. The age, the origin, and the visiting party were instead the variables that significantly discriminate the “Knowledge seeker” from the “Non-motivated” of ÖTZI. The lower the age of the visitors, the higher (less than proportionally) the probability of being a “Knowledge seeker”. Foreign visitors were more likely to be grouped in the “Knowledge seeker” and the members of this group of visitors were more likely to visit ÖTZI alone or with their families, instead in groups of people that were not member of their families. 6. Conclusions The Bagged Clustering method proposed by Leisch (1999) was adopted in this research as a useful exploratory technique and a suitable alternative to the more traditional clustering methods. In particular, this method doesn’t require the a priori definition of the number of clusters and the result is more stable than those obtain by a more traditional partitioning method, like the k-means. Even if dichotomous variables are used, this method allows using the Box-plots to easily represent the empirical distribution of each segmenting variable among the clusters identified. Furthermore, in contrast with other segmentation procedures commonly used in the tourism literature for binary datasets (like the use of the factor analysis followed by a hierarchical cluster method), the Bagged Clustering allows considering the original set of variables using all available information. As underlined by Chen and Hsu (1999), there are two main reasons to use all variables for segmenting purposes: firstly, results of the factor analysis could be different when researcher uses different rotation methods; secondly, the original variables may be more interpretable than derived constructs with factor labels because the naming and the interpretation of the construct involve personal judgments that it is generally preferred to avoid. This algorithm was presented in this research both from a theoretical and an applied point of view. The dataset used were collected from two ad-hoc surveys conducted from June to September 2011 at the two main museums of the two provinces of the Trentino-South Tyrol region (Trento and Bolzano): the South Tyrol Museum of Archaeology (shortened to “ÖTZI”) and the Museum of Modern and Contemporary Art (shortened to “MART”). The aim was to segment the visitors of these two different types of museums with respect to the motivational factors that push them to make the visit. Furthermore, a comparison between the cluster solutions obtained for the two museums was developed. The visitors of MART museum were grouped into three clusters while those of the ÖTZI museum into two clusters. In both cases: • A group of “Knowledge seeker” emerged, having heterogeneous socio-demographic and economic characteristics between the museums. In fact, the “Knowledge seeker” of MART was a group of visitors living near the museum (in Northern Italy) with high level of education, whereas the “Knowledge seeker” of ÖTZI was a group of young, foreign visitors who preferred to visit the museum with their families or alone. Therefore, MART museum must driving its promotional and marketing efforts in order to attract highly educated and foreign visitors. Vice versa, the promotional policy of ÖTZI museum should be more oriented to capture the Italian visitors, mainly young people or families. • A large group of visitors with any particular push-motive (the “Non-motivated”) was found in each museum and must be analysed more in depth in future researches. The last, and largest, group identified among the MART visitors, i.e. the “Interested”, seemed to spend more than the other groups at the shop of the museum. In fact, they visited the museum because they are interested in such an attraction and in visiting the temporary exhibition. These two push factors can lead to buy more books, or other expensive souvenirs, than other visitors. So that, the museum should propose more books or souvenirs regarding its exhibitions and its attractions in order to increase the revenues of the museum shop. The main limitation of this study is that the segmentation analysis is based on a non- random sampling technique. Thus, to verify if the results of this research are valid for other museums, a future study will be required in other museums, in other years, and/ or other towns. Future researches must focus on the looking for a more suitable distance measure for binary data to apply both in the base method and in the hierarchical analysis, in order to improve the Bagged Clustering method. Finally, this research could be extended in the future by comparing the Bagged Clustering results with other segmentation techniques, both linear and non-linear techniques, evaluating which of them is better for segmenting visitors of museums and cultural attractions. References Albalate, D., & Bel, G. (2010). Tourism and urban public transport: Holding demand pressure under supply constraints. Tourism Management, 31(3), 425–433. Beane, T. T., & Ennis, D. M. (1987). Market segmentation: A review. European Journal of Marketing, 21(5), 20–42. Bennett, T. (1994). The reluctant museum visitor: A study of non-goers to history museums and art galleries. Sydney: Australia Council. Bigné, J. E., & Andreu, L. (2004). Amotions in Segmentation. An Empirical Study. Annals of Tourism Research, 31(3), 682–696. Bouguila, N. (2010). On multivariate binary data clustering and feature weighting. Computational Statistics and Data Analysis, 54, 120-134. Brida, J. G., Disegna, M., & Osti, L. (2012). Segmenting visitors of cultural events by motivation": A sequential non-linear clustering analysis of Italian Christmas Market visitors. Expert Systems with Application, 39, 11349–11356. Brida, J.G., Pulina, M. and Meleddu, M. (2012). Factors influencing the intention to revisit a cultural attraction: the case of MART of Rovereto. Journal of Cultural Heritage, 13(2), 167–174. Budayan, C., Dikmen, I., & Birgonul, M.T. (2009). Comparing the performance of traditional cluster analysis, self-organizing maps and fuzzy C-means method for strategic grouping. Expert System with Application, 36, 11772-11781. Buttrey, S. E., & Karo, C. (2002). Using K-nearest-neighbor classification in the leaves of a tree. Computational Statistics and Data Analysis, 40, 27-37. Chen, J. S., & Hsu, C. H. C. (1999). The use of logit analysis to enhance market segmentation methodology. Journal of Hospitality & Tourism Research, 23(3), 268–283. Choi, A. S. (2011). Implicit prices for longer temporary exhibitions in a heritage site and a test of preference heterogeneity: A segmentation-based approach. Tourism Management, 32, 511–519. Claver-Cortés, E., Molina-Azorín, J. F., & Pereira-Moliner, J. (2007). Competitiveness in mass tourism. Annals of tourism Research, 34(3), 727–745. Cochran, W. G. (1977). Sampling techniques. Wiley series in probability and mathematical statistics (Third Edn.). John Wiley & Sons. Dicks, B. (2003). Culture on display: The production of contemporary visitability. Berkshire: Open University Press. Dolnicar, S., & Leisch, F. (2000). Getting More Out of Binary Data: Segmenting Markets by Bagged Clustering. Working Paper Series 71 SFB “Adaptive Information Systems and Modeling in Economics and Management Science”, http://www.wu-wien.ac.at/am, August 2000. Dolnicar, S., & Leisch, F. (2003). Winter Tourist Segments in Austria: Identifying Stable Vacation Styles Using Bagged Clustering Techniques. Journal of Travel Research, 41(3), 281–292. Dolnicar, S., & Leisch, F. (2004). Segmenting Markets by Bagged Clustering. Australasian Marketing Journal, 12(1), 51–65. Dolnicar, S., Crouch, G. I., Devinney, T., Huybers, T., Louviere, J. J., & Oppewal, H. (2008). Tourism and discretionary income allocation. Heterogeneity among households. Tourism Management, 29(1), 44–52. Everitt, B.S., Landau, S., Leese, M., & Stahl, D. (2011). Cluster Analysis. Wiley series in Probability and Statistics, London. Fingh, H. (2005). Comparison of distance measures in cluster analysis with dichotomous data. Journal of Data Science, 3, 85-100. Handl, J., Knowles, J., & Kell, D. B. (2005). Computational Cluster Validation in Post- Genomic Data Analysis. Bioinformatics, 21(15), 3201–3212. Huang, S. C., Chang, E. C., & Wu H. H. (2009). A case study of applying data mining techniques in an outfitter’s customer value analysis. Expert System with Application, 36, 5909-5915. Hughes, H. L. (2002). Culture and tourism: A framework for further analysis. Managing Leisure, 7(3), 164–175. Jain, A. K., Murty, M. N., & Flynn, P. J. (1999). Data Clustering: A review. ACM Computing Surveys, 31(3), 264-323. Kang, K., Hua-Xiang, Z., & Ying, F. (2008). A novel Cluster Ensemble Algorithm Based on Dynamic Cooperation. Fifth International Conference on Fuzzy Systems and Knowledge Discovery, IEEE Computer Society, 32–35. Kim, H., Cheng, C. K., & O’Leary, T. J. (2007). Understanding participation patterns and trends in tourism cultural attractions. Tourism Management, 28(5), 1366–1371. Konu, H., Laukkanen, T., & Komppula, R. (2011). Using ski destination choice criteria to segment Finnish ski resort customers. Tourism Management, 32(5), 1096–1105. Kotler, P., Bowen, J. T., & Makens, J. C. (2010). Marketing for hospitality and tourism (5th Edn.). Upper Saddle River, New Jersey: Pearson Prentice Hall. Leisch, F. (1999). Bagged Clustering. Working paper 51, SFB “Adaptive Information Systems and Modeling in Economics and Management Science”, http://www.wu- wien.ac.at/am, August 1999. Leisch, F. (2006). A toolbox for K-centroids cluster analysis. Computational Statistics & Data Analysis, 51, 526–544. Lia, S. H., Chu, P. H., & Hsiao P. Y. (2009). Data mining techniques and applications – A decade review from 2000 to 2011. Expert System with Application, 39, 11303-11311. MacDonald, G., & Alsford, S. (1995). Canadian Museums and the Representation of Culture in a Multicultural Museum. Cultural Dynamics, 7(1), 15–36. McKercher, B. (2004). A comparative study of international cultural tourists. Journal of Hospitality and Tourism Management, 11(2), 95–107. Pérez, E. A., & Nadal, J. R. (2005). Host Community Perceptions. A Cluster Analysis. Annals of Tourism Research, 32(4), 925–941. Punj, G., & Steward, D. W. (1983). Cluster analysis in marketing research: Review and suggestions for application. Journal of Marketing Research, 20(2), 138–148. Řezanková, H. (2009). Cluster analysis and categorical data. Statistika, 3, 216-232. Saarenvirta, G. (1998). Mining customer data. DB2 Magazine, 3(3), 10–20. Schuster, M. D. (1991). The audience for American art museums. National endowment for the arts research division report 23. Washington: NEA. Stylianou-Lambert, T. (2011). Gazing from home: cultural tourism and art museums. Annals of Tourism Research, 38(2), 403–421. Vesanto, J., & Alhoniemi, E. (2000). Clustering of the self-organizing map. IEEE Transactions on Neural Networks, 11(3), 586–600. Figure 1. The bagged clustering algorithm. Original sample (N = observations) . . . . . . . . . . . .. . . . . . .. . . . . . . B Bootstrap samples XN 1 XN B . . . . Partitioning method (K = centers) c1 1,…,cK 1 c1 B,…,cK B . . . . Hierarchical method on B × K centres Partition of the centres by cutting conveniently the dendrogram. Partition of the original data by assigning each observation to the cluster containing the centre closest to x. XN x ∈ XN cj i Figure 2. MART museum: Bagged Clustering dendrogram together with the plot regarding the relative height of aggregation (black line) and the first differences (grey line). 0 20 40 60 80 Dendrogram hclust (*, "ward") H ei gh t 0 20 40 60 80 Cluster Dendrogram hclust (*, "ward") d H ei gh t 5 10 15 20 0. 0 0. 2 0. 4 0. 6 0. 8 1. 0 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Figure 3. ÖTZI museum: Bagged Clustering dendrogram together with the plot regarding the relative height of aggregation (black line) and the first differences (grey line). 0 20 40 60 80 10 0 12 0 Dendrogram hclust (*, "ward") H ei gh t 0 20 40 60 80 10 0 12 0 Cluster Dendrogram hclust (*, "ward") d H ei gh t 5 10 15 20 0. 0 0. 2 0. 4 0. 6 0. 8 1. 0 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Figure 4. MART museum: box-plot for the three clusters solution. ●●●●● ● ● ● ●●●●● ● ●●●●●●●● ● ●●● ● ● ●●●●●●●●● ● ●● ● ● ● ●●● ● ●●●●●● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●● ● ● ● ● ● ●●● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ●●●●●● ● ● ●● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ●● ● ● ●●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●●●●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ●●● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ●●●● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ●●● ● ●● ● ●●● ● ● ● ● ●● ● ● ● ●●●●●●●● ● ●●● ● ●0. 0 0. 4 0. 8 cur ios ity rela x inte res t frie nd lea rn tell do futu re rev isit sho w wo rk wo rth wh ile leis ure tem por ary bui ldin g Cluster 1: 517 centers, 253 data points ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●●●● ● ●● ● ● ● ● ● ●●●● ● ●● ● ●●● ● ● ● ●●●●●● ● ●●●● ● ●●●● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●●●● ● ●●●●●●●●●● ● ● ● ●●●●●●● ● ●●●●●●●●●●●●●● ● ●●● ● ●●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● 0. 0 0. 4 0. 8 cur ios ity rela x inte res t frie nd lea rn tell do futu re rev isit sho w wo rk wo rth wh ile leis ure tem por ary bui ldin g Cluster 2: 332 centers, 293 data points ●●●●●●● ● ●●●●●●● ● ● ●● ● ● ●●●● ●●● ● ● ● ● ●● ● ●● ● ● ●●●●●●● ● ● ●●●●● ● ●● ● ●●●●●●●●●●●●●●●●●● ●● ● ●● ● ●● ● ● ●●●●●●● ● ● ●●●●●●●0. 0 0. 4 0. 8 cur ios ity rela x inte res t frie nd lea rn tell do futu re rev isit sho w wo rk wo rth wh ile leis ure tem por ary bui ldin g Cluster 3: 151 centers, 45 data points Figure 5. ÖTZI museum: box-plot for the two clusters solution. ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ●●●● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ●● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●●●●● ● ● ● ● ● ●●●● ● 0. 0 0. 4 0. 8 cur ios ity rela x inte res t frie nd lea rn tell do futu re rev isit sho w wo rk wo rth wh ile leis ure tem por ary bui ldin g Cluster 1: 361 centers, 173 data points ●● ● ● ● ● ● ●● ●● ● ●● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●●●●●● ● ● ● ●●● ● ●●● ● ● ● ●●●●●● ● ● ●● ●●●● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ●●●●●●●●● ● ● ●●●●●● ● ●●● ● ●●●●●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ●●● ● ●● ●● ●● ● ● ● ●● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ●● ●●●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ●● ● ● ● ●● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●●● ● ●●● ● ●● ● ● ● ● ●●● ● ●● ●●● ● ● ● ●● ● ● ● ●●● ● ●●●● ● ●●● ● ● ● ● ● ●●●●●●●● ● ● ● ● ● ● ●●●●●● ● ● ● ●● ● ●● ● ●●●● ●● ●●● ● ● ●● ● ●● ● ● ● ●● ● ●● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ●● ● ● ● ●● ●● ● ● ●●●●●● ● ● ●●● ●● ● ● ● ●●●●● ●● ●● ● ●● ● ● ● ● ●● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●●●● ● ● ●● ● ● ●● ● ●● 0. 0 0. 4 0. 8 cur ios ity rela x inte res t frie nd lea rn tell do futu re rev isit sho w wo rk wo rth wh ile leis ure tem por ary bui ldin g Cluster 2: 639 centers, 506 data points Table 1. Structure of the questionnaire. Sections Description Categories of variables I Museum information Repeat visiting; number of museums visited in the last year; factors that stimulated the visit*; rating of factors that describe the visit**; shopping expenditure at the museum; authenticity perception*. II Trip information Motives of the trip; number of nights, total expenditure per person per night. III Interviewees’ profile Some socio-demographic and economic characteristics of interviewees and their families. Notes: * dichotomous variables have been used; **A 5-points Likert scale has been used. Table 2. Socio-demographic and economic characteristics. MART ÖTZI Interested Knowledge Non- p-value Knowledge Non- p-value seeker motivated seeker motivated Male (%) 41.30 55.56 45.63 0.17 53.66 50.61 0.49 Age (mean) 46.94 39.38 42.07 *** 44.58 44.12 0.68 University (%) 81.57 93.33 82.14 0.14 68.48 68.74 0.95 Origin of visitors (%) 0.41 *** Abroad 3.08 0.00 3.97 25.60 16.09 Germany 3.42 2.22 4.37 42.86 31.57 Centre/South of Italy 7.19 11.11 11.51 5.95 17.92 North–East of Italy 48.97 60.00 41.67 8.33 15.68 North–West of Italy 13.70 11.11 14.68 14.88 14.46 Local resident 23.63 15.56 23.81 2.38 4.28 Occupation (%) ** 0.94 Autonomous worker 17.41 13.33 18.65 21.95 19.96 Employed 49.83 37.78 46.83 58.54 60.29 Retired 15.70 37.78 9.52 12.20 12.89 Other occupations 17.06 11.11 25.00 7.32 6.86 Visiting party (%) 0.85 0.31 Alone 8.19 6.67 7.91 8.33 5.69 Couple 12.97 11.11 34.39 41.67 36.79 Children 35.84 28.89 15.42 33.93 40.24 Group 43.00 53.33 42.29 16.07 17.28 Household annual income (%) 0.34 0.14 0 -| 25,000 19.45 15.56 20.63 5.45 10.86 25,000 -| 50,000 40.27 51.11 36.11 27.88 26.64 50,000 -| 75,000 9.90 15.56 13.49 13.33 16.19 > 75,000 7.51 0.00 8.33 20.00 14.55 Missing income 22.87 17.78 21.43 33.33 31.76 Expenditure (mean) Total expenditure 30.38 23.62 33.42 0.54 55.21 48.66 0.13 Shopping at the museum 13.78 8.18 9.18 0.12 5.55 10.08 0.48 Notes: Chi-square test was used for qualitative variables and continuous variables recoded in classes. ANOVA test and t-test were used in order to test whether the mean value of the quantitative variables significantly differ among the three clusters identified for the MART museum and between the two clusters identified for the ÖTZI museum. All test results are not significant unless indicated otherwise: ***Significant at p ≤ 0.01, **Significant at p ≤ 0.05, *Significant at p ≤ 0.1 Table 3. Multinomial Logit (MART) and Binary Logit (ÖTZI) coefficients. MARTA ÖTZIB Interested Knowledge seeker Knowledge seeker Male -0.25 (0.19) 0.66 (0.33)** 0.05 (0.20) Age >0.01 (0.06) -0.01 (0.08) -0.13 (0.06)** Age2 >0.01 (>0.01) -0.01 (>0.01) >0.01 (>0.01)** University -0.01 (0.24) 1.41 (0.61)** 0.14 (0.22) Origin of visitors Abroad -0.50 (0.50) -29.79 (0.66)*** 1.05 (0.59)* Germany -0.09 (0.48) 0.63 (1.18) 0.94 (0.56)* Centre/South of Italy -0.50 (0.38) 0.86 (0.70) -0.51 (0.65) North–East of Italy 0.11 (0.23) 1.06 (0.46)** 0.03 (0.62) North–West of Italy -0.10 (0.32) 0.63 (0.63) 0.71 (0.60) Occupation Autonomous worker 0.16 (0.32) -0.66 (0.61) 0.34 (0.35) Employed 0.36 (0.28) -0.65 (0.44) 0.21 (0.30) Visiting party Alone -0.03 (0.35) -0.51 (0.71) 0.74 (0.43)* Couple -0.12 (0.21) -0.25 (0.39) 0.52 (0.26)** Children -0.32 (0.29) -0.28 (0.57) 0.50 (0.28)* Household annual income Income -0.01 (>0.01) -0.01 (0.01) >0.01 (>0.01) Missing income -0.07 (0.30) -0.88 (0.51)* 0.26 (0.30) Expenditure Total expenditure -0.01 (>0.01) -0.01 (0.01) >0.01 (>0.01) Missing total expenditure -0.01 (0.21) -0.21 (0.38) -0.23 (0.25) Shopping at the museum 0.03 (0.02)** 0.01 (0.03) -0.03 (0.03) Missing expenditure at the shop of the museum -0.32 (0.21) 0.32 (0.37) -0.15 (0.24) Constant -0.33 (1.12) -2.44 (1.75) 0.21 (1.22) Notes: All test results are not significant unless indicated otherwise: ***Significant at p ≤ 0.01, **Significant at p ≤ 0.05, *Significant at p ≤ 0.1. Robust Std. Err. in brackets. A Multinomial logit: N = 588; Wald chi2(40) = 9207.03; Prob > chi2 = 0.00; Pseudo R2 = 0.0655; McFadden R2 = 0.066; Cox & Snell R2 = 0.112; Nagelkerke R2 = 0.134. B Binary Logit: N = 631; Wald chi2(20) = 40.40; Prob > chi2 = 0.00; Pseudo R2=0.0626; McFadden R2 = 0.063; Cox & Snell R2 = 0.068; Nagelkerke R2 = 0.101. Appendix A Table A1. Description of the explanatory variables. Independent variables Descriptions Socio-demographic and economic characteristics Male 1= male; 0= female Age Age of the respondent (continuous) Age2 Squared age of the respondent (continuous) University 1 = education level is university degree or postgraduate; 0= otherwise Origin of visitors Abroad 1= Abroad (excluding Germany); 0= otherwise Germany 1= Germany; 0= otherwise Centre/South of Italy 1= Centre, South, or islands of Italy; 0= otherwise North–East of Italy 1= North-East of Italy (excluding the province in which the museum is located); 0= otherwise North–West of Italy 1= North-West of Italy; 0= otherwise Local resident 1= province that host the museum; 0= otherwise (reference category) Occupation Autonomous worker 1= autonomous worker; 0= otherwise Employed 1= employed (full-time or part-time); 0= otherwise Retired 1= retired; 0= otherwise Other occupation 1= student/unemployed/housewife/working occasional or on project/teacher/other; 0= otherwise (reference category) Visiting party Alone 1= alone; 0 = otherwise Couple 1= partner/spouse; 0 = otherwise Children 1= children between 0 and 12 years; 0 = otherwise Group 1= friends/colleagues/organized group/other relatives; 0 = otherwise (reference category) Household annual income Income Central value of each income category (see the list reported in Table 2); 0 if the respondent does not declare her income (continuous) Missing income 1 = respondent does not declare his/her income; 0 = otherwise Expenditure Total expenditure Individual expenditure for accommodation, food and beverage, shopping in the shops of the city, pharmacy, tour guide services, other expenditures linked to the visit (excluding expenditure for transportation) per night in Euros; 0 if respondent does not state their expenditure (continuous) Missing total expenditure 1 = respondent does not declare his/her total expenditure; 0 = otherwise Shopping at the museum Individual expenditure at the shop of the museum in Euros; 0 if respondent does not state their expenditure (continuous) Missing expenditure at the shop of the museum 1 = respondent does not declare his/her expenditure at the shop of the museum; 0 = otherwise 
work_l4gu7odznvbzbmx7s3mssyjs7m	----	Calibrated fuzzy AHP for current bank account selection Calibrated fuzzy AHP for current bank account selection Alessio Ishizaka ⇑, Nam Hoang Nguyen University of Portsmouth, Portsmouth Business School, Richmond Building, Portland Street, Portsmouth PO1 3DE, United Kingdom a r t i c l e i n f o Keywords: Fuzzy AHP Membership functions Customisation Current account Banking a b s t r a c t Fuzzy AHP is a hybrid method that combines Fuzzy Set Theory and AHP. It has been developed to take into account uncertainty and imprecision in the evaluations. Fuzzy Set Theory requires the definition of a membership function. At present, there are no indications of how these membership functions can be constructed. In this paper, a way to calibrate the membership functions with comparisons given by the decision-maker on alternatives with known measures is proposed. This new technique is illustrated in a study measuring the most important factors in selecting a student current account. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction Despite the popularity and simplicity of the Analytic Hierarchy Process (AHP), it is often criticised for its inability to adequately han- dle the uncertainty of a decision maker’s preferences. In classic AHP, the judgements are represented by exact values on a scale of 1–9 (Saaty, 1977, 1980). However, in many real cases, the linguistic assessments of human evaluations are often vague, and it is not real- istic to represent them with crisps values. To overcome these short- comings, fuzzy AHP has been developed to take into account this uncertainty and imprecision. It is essentially the combination of two methods: fuzzy set theory and AHP (Van Laarhoven & Pedrycz, 1983). Fuzzy set theory requires the definition of a membership function for each verbal judgement. However, in all papers re- viewed, there was no indication of how the membership functions have been selected. This paper proposes a way to calibrate the mem- bership functions with comparisons given by the decision-maker on alternatives with known measures. In this case, we asked to com- pare the surface of geometrical figures and as a result, the member- ship function was personalised for each participant. Then, the fuzzy AHP with the customised membership functions is applied to a case study in order to establish the most important factors in the selec- tion of a student current account. We found that service is the most weighted criteria when selecting a bank account. 2. Fuzzy AHP Fuzzy AHP was first proposed by Van Laarhoven and Pedrycz (1983) and is an extension of AHP combined with fuzzy set theory (Zadeh, 1965). The main advantage of this combination is that it makes allowances for the vagueness and imprecision of human preference. The key idea is that a certain degree of an element be- longs to a fuzzy membership set, which is given by a function de- picted on a two-axis diagram. The horizontal axis consists of the domain elements of the fuzzy sets and the vertical axis the degree of membership on a scale of 0–1. These membership functions can take several shapes: linear, S-curves, triangular or trapezoidal rep- resentations. In practice, triangular and trapezoidal membership functions are the most frequently used. They can be denoted by à = (l, ml, mu, u), where l 6 ml 6 mu 6 u correspond to lower, mod- al-lower, modal-upper and upper bound, i.e. the trapezium’s angle points. If the membership is triangular, then ml = mu (Fig. 1). The membership of à is defined by: leAðxÞ¼ x�l m�l ; l 6 x 6 m 1; x ¼ m u�x u�m ; m < x < u 0; otherwise 8>>>< >>>: ð1Þ Fuzzy AHP is based on 4 steps: (a) For each linguistic term of the evaluation scale, a member- ship function is constructed. (b) Criteria/alternatives are pair-wise compared in comparison matrix à eA ¼ ~a11 ~a12 . . . ~a1n ~a21 ~a22 . . . ~a2n . . . . . . . . . . . . ~an1 ~an2 . . . ~ann 2 6664 3 7775 ð2Þ where ãij is the fuzzy comparison between criterion/alternative i and j. (c) Fuzzy priorities are derived from comparison matrix Ã. This is done using the eigenvalue method (3) or any other method used in traditional AHP (Ishizaka & Labib, 2011) 0957-4174/$ - see front matter � 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2012.12.089 ⇑ Corresponding author. Tel.: +44 23 92 844171. E-mail addresses: Alessio.Ishizaka@port.ac.uk (A. Ishizaka), nhnam1988@gmail. com (N.H. Nguyen). Expert Systems with Applications 40 (2013) 3775–3783 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2012.12.089 mailto:Alessio.Ishizaka@port.ac.uk mailto:nhnam1988@gmail.com mailto:nhnam1988@gmail.com http://dx.doi.org/10.1016/j.eswa.2012.12.089 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa eA � ~p ¼ k � ~p ð3Þ (d) As these fuzzy priorities must be ranked, they need to be translated into real numbers to make the ranking more obvi- ous than fuzzy numbers. Several methods exist including the weighted average approach, the centre of area, the mean- max membership and the first (or last) of maxima. The most popular is the centre of area or centroid (Van Leekwijck & Kerre, 1999). Except for the fuzzy representation of the judgement scale, the steps of fuzzy AHP are the same as traditional AHP (Ishizaka & Labib, 2011). Therefore, this paper will concentrate on the fuzzy membership function that represents the judgement scale. In the literature review, we found 27 different representations of fuzzy membership functions (Table 1), however none have been justified. Li and Kuo (2008) are the only one to ask the decision-maker to construct their own membership function but do not give any guidance on how to fulfil this task. This paper presents a new way to construct a personalised membership function. The meth- odology will be illustrated by a case study in banking. 3. Membership function calibration The calibration of the membership function is performed through a comparison of measurable alternatives. In our case, we used geometrical figures but it is possible for other items to be used (Fig. 2). The participants were asked to compare their surface with the verbal scale given in Table 2. They were also informed that the figures were in an increasing order, so the questionnaire only had one scale direction (Table 3), e.g. A is necessarily smaller than B. Not all comparisons are required for the calibration; therefore only a subset was asked to avoid overwhelming the participants. The measured pairwise comparisons of the figures are given in Table 4. The verbal judgements (Table 3) given by the decision-maker are matched with the real values (Table 4). For example, suppose that the decision-maker evaluates a ‘‘very strong’’ difference between figures G and A, D and A and also between Fig. I and B. The real values of these three evaluations (i.e. 7, 4, 4.5) are entered into the match- ing table (Table 5). Therefore, it can be deduced that the decision maker values outcomes of between 4.5 and 7 as ‘‘very strong’’. All the judgements matched with the real measures are entered into a Table 5. For each verbal judgement, the minimal, mean and maximal values are calculated. They correspond to the angle points of the customised membership function. Fig. 3 represents the cus- tomised membership functions of all verbal judgements. Notice that these membership functions are not similar (e.g. the wideness of the membership function ‘‘very strong’’ is much larger than ‘‘moderate’’) because they depend on the person’s interpretation of verbal judgements. 4. Case study 4.1. Introduction The development of an appealing product may have a long-term impact on the profitability of companies. This is especially true in the banking sector, where students often remain with the same bank when they leave education. Students are not a profitable seg- ment of the market because their income is low, however they are the potentially high earner in the future. As a result, it is in the best interests of the bank to attract and retain these customers early. This explorative study will give an insight into the most impor- tant criteria in selecting a student bank account using calibrated fuzzy AHP, described in Section 3. 4.2. Criteria description In the literature there are several studies for bank selection in different countries: Romania (Katircioglu, Tumer, & Kılınç, 2011a); Ghana (Hinson, Owusu-Frimpong, & Dasah, 2011; Mahmoud, Tweneboah-Koduah, & Danku, 2011); USA (Lee & Marlowe, 2003), Northern Cyprus (Katircioglu, Unlucan, & Dalci, 2011b; Safakli, 2007); Malaysia (Ahmad, Rustam, & Dent, 2011; Amin, 2008; Mokhlis, Salleh, & Mat, 2011); Greece (Lymperopou- los, Chaniotakis, & Soureli, 2006); Bahrain (Al-Ajmi, Abo Hussain, & Al-Saleh, 2009; Almossawi, 2001); United Kingdom (Devlin & Gerrard, 2005; Farquhar & Panther, 2008; Thwaitesa & Verea, 1995); Singapore (Ta & Har, 2000), Poland (Kennington, Hill, & Rakowska, 1996); Hong Kong (Denton & Chan, 1991); India (Gupta & Dev, 2012). Each study has its own list of criteria. As the utilisation of AHP becomes difficult with a large number of criteria, similar factors were grouped together (Table 6) and structured into a hierarchy (Fig. 4). This also avoids the problem of overweighting dependent criteria (e.g. internal and external bank appearance). Some criteria have not been considered because: � They are out-dated, for example, ATM service. Banks have a con- sensus scheme to share ATM information systems, therefore; a person can withdraw cash either free of charge or for a small fee from any ATM belonging to another bank. � They are outside the control of the banks, such as recommenda- tions from friends and relatives. Some studies also suggest that these criteria are negligible in bank account selection (Almossawi, 2001; Ta & Har, 2000). 4.3. Demography of the participants Forty participants of the University of Portsmouth were re- cruited in a sample of equal gender and nationality proportions (Table 7). Participants are aged between 19 and 30 (Table 8). Twenty- three students are on a bachelor course and 17 on a master’s level course. Only participant P36 had full-time work experience of more than six months. 4.4. Questionnaire collection mode To increase the response rate, different collection channels were used: � E-mail: This collection mode has a low associated cost (no printing and postage) and is timesaving as a large population can be targeted at once. The questionnaire was sent to 75 stu- dents. Twenty-five questionnaires were returned but only 14 were correctly completed. The perceived disadvantage of this Fig. 1. Trapezoidal membership function. 3776 A. Ishizaka, N.H. Nguyen / Expert Systems with Applications 40 (2013) 3775–3783 https://isiarticles.com/article/6320 
work_l543wnt5qnf2jfz2knphpdvvsm	----	LUND UNIVERSITY PO Box 117 221 00 Lund +46 46-222 00 00 An Architecture for Expert System Based Feedback Control Årzén, Karl-Erik 1988 Document Version: Publisher's PDF, also known as Version of record Link to publication Citation for published version (APA): Årzén, K-E. (1988). An Architecture for Expert System Based Feedback Control. (Technical Reports TFRT- 7399). Department of Automatic Control, Lund Institute of Technology (LTH). Total number of authors: 1 General rights Unless other specific re-use rights are stated the following general rights apply: Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal Read more about Creative commons licenses: https://creativecommons.org/licenses/ Take down policy If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim. https://portal.research.lu.se/portal/en/publications/an-architecture-for-expert-system-based-feedback-control(f7b7cc68-31fb-4e83-b338-d48dffd9cf01).html CODEN: IUTFD2/(TFRr-73 ee) I L-07 I (1es8) An Architecture for Expert System Based Feedback Control Karl-Erik Årzén Department of Automatic Control Lund Institute of Technology September 1988 Department of Automatic Control Lund fnstitute of Technology P.O. Box 118 5-221 00 Lund Sweden Documcnt na¡rc¡c Report Datc of itøuc September L988 Documcnt ÀIumbc¡ CoDEN: tUTFD2/(TFRT-73ee)/1-07l( 1e88) Author(s) Karl-Erik Årzén Supcrviso¡ Sponcoring org¡anísation STU Titlc dnd subtìtle An architecture for expert system based feedback control. Absttæt It is a recognized problem that many industrial control loops are badly tuned or run in manual mode. Two expert system approaches have been suggested for this problem. Fuzzy, rule-based control replace control algorithms by linguistic rules which model the operators manual control strategy. Knowledge-based control extends the range of conventional controllers by encoding general control knowledge and heuristics concerning tuning and adaptation in a supervisory expert system. An architecture for knowledge-based control is described where two concurrent processes are used for the knowledge-based system and the numerical algorithms. A modular, blackboard-based approach is used. This allows the decomposition of the problem into subtasks which are implemented as separate knowledge sourcee that can be rule-based with different inference strategies or procedural. The framework can be compared with a real-time operating system and has similar real-time primitives. The system has been implemented on a VAX tL/780 and used with good experiences. Key words Real-time expert systems, Feedback control Classífic¿tíon systcrm and./or indcx tcrms (íf any) S up plc ment ary b iblio graphic al ínfo tmat io n ISSN and kcy title ISBN Languagc English Numbc¡ oÍ pages 7 R.ccípìcnt's notes S ccuríty classìfrcat íon The report may bc o¡de¡ed from thc Department of Automatíc Cont¡ol or bortowcd úårough thc llnivercíty Librury 2, Box 7010, 3-227 O3 Lund, Swcdcn, Tclcx: 33248 lubbís lund. Paper to be presented at AI in real-time control, the IFAC Vüorkshop on 21-23 September, Wales. An architecture for expert systern based feedback control Kart-Erik Ä.rzén Department of Automatic Control Lund Institute of Technology Box 118 5-221 00 Lund, Swcden Abstract. It is a recognized problem that many industrial control loops are badly tuned or ru¡l in manual modc. Two cxpcrt systcm approachcs havc bccn suggcstccl ftrr tlris problem. Fuzzy, rule-based control replace control algorithms by linguistic rules rv¡ic[ model the operators m¿nual control strategy. Knowledge-based control extends the rangc of conventional controllers by encoding general control knowledge and heuristics concerning tuning and adaptation in a supervisory expert system. An architecture for knorvledge- based control is described where two concurrent processes are used for the knowledge-basecl system and the numerical algorithms. A modular, blackboard-based approach is used. Tiris allows the decomposition of the problem into subtasks which are implemented as separate knowledge sources that can be rule-based with different inference strategies or procedural. The framework can be compared with a real-time operating system and has similar real- time primitives. The system has beer- implemented on a VAX I1/780 and used rvith goocl experiences. Keyuords: Real-time expert systems, Feedback control 1. Introduction There is currently a significant interest in expert sys- tem techniques in the process control community, Applications of many different types have been pro- posed, implemented and a few also fielded. This pa- per considers the use ofexpert system, or knowledge- based system, techniques in the closed control loop. It is a recognized problem that many industrial con- trol loops are badly tuned or run in rnanual ¡node. This decreases the quality of the end product ancl thus increases cost. The manual control task also adds to thc already high cognitivc burclc¡r that pro- cess operators are exposed to in modern control sys- tems. The reasons for the poor control arc many. Onc could be that the control loop is badly tuned from the beginning. Another could be that the operating conditions have changed since the initialization of the controller. This could, c.g. bc duc to operation at dillerent operating points or time-varying dynamics. The conventional solution to the problern of poorly tuned control loops is to use adaptive controllcrs. Adaptive controllers, e.g., (,4.ström and Wittenmarrk, 1989), are currently beginning to be used in i¡rdustrial practice. There are, however, problems. Even though an explicit self-tuning regulator periodically updates the coefñcie¡rts of a process model thcre still a¡e rrÌany parameters that must be set explicitly. Examples are model orders and time scales. Such information can be diflìcult to provide and process operators typicall¡' lack the intuitive understanding that they have with conventional PID controllers- Two expert system approaches have been suggested for the described problem. Both i¡volve using the expert system as a part ol the feedback loop. In the u'ell-known fuzzy or rule-based approach, e.g., (Tong, 1984), the attempt is to model the manuai control strategy of the process operator. It is exprcssed as qualitative, Iinguistic rules lor horv to choose the con- trol signal in different situations, The rules replace conventional control algorithms. The intended appli- cations are control ofcomplcx proccsscs such as, c.g., cement kilns, for which either appropriate models do not exist or are inadequate. The sccond approach, from now on rcfcrrc<l f,o as knowledge-based control (Åström and Anton, 1g84; Ãström et al, 1986i Ârzén, 1987), instead uses ex- pert system techniques to extend the range of con- vcntional control ir.lgorithrns by crrcoding gcrrcral control knowlcdgc and heuristics rcgirrding tuning and adaptation in a supcrvisory expert system. Tlie kno*'ledge-based control approach is closer in spirit to co¡n'entional aclap[ivc control than fuzz¡' conlrol is. The approach is also motivaied by shortconrings of adaptive cont¡ollers. 'Ihe recursive identification algorithrns and the co¡r- trol algorithms of adaptive controlle¡s c¿r¡r l¡e sr:cn Supervlslon algorlthmsKnowledge Bas€d System algorlthme Conlrol algorllhms Process A prlorl informatlon Figure 1. A knowledge-based controller as the final algorithmic representation of an large amount of undcrlying thcoretical as v/ell as practi- cal control knowledgc. This reprcsent,ation is howevcr not enough. It has to be combined with heuristic logic that assures the controller performance under non-standard conditions. These conditions include switching between different operating modes, insuf- ficient process excitation, control signal saüuration etc, The concept safety jacket or saÍety nef (Isermann and Lachmann, 1985; Warvick, 1g88) has been estab- lished for this logic. The safety net part of a controller is often much larger than the actual algorithms and is designed mai¡ly from intuition, experience, ancl simulation. Safety nets tend to be very complex. Ex- perience has shown that design and testing is quite time consu:rring. The approach in knowledgc-based control is to use an expert system to represent the heuristic safety net. The controller consists of the combination of the expert system and a set of control algorithms, identification algorithms and supervision algorithms, as shown in Fig. 1. The topic of this paper is the organization and arch! tecture of a knowledge-based controller. I{nowledge- based control is a real-time expert system applica- lion and, as suclì, contains several difEcult problems such as non-monotoning reasoning, representation of time and temporal reasoning, reasoning under time constraints, responsiveness to asynchronous events etc. For an overview of these issues see Laífey et al (1988) or Chantler (198S). Some of these issues are still unsolved and will be probably never be com- pletely solved. In several cases, however, practical approaches exist that to some degree solve the prob- lems. There exist a widely spread mis-unclerstanding that, adding intelligent behaviour to a controller is simply a matter of generating a few rules and implementing them in an off-the-shelf expert system shell. This is far from the case. There is a strong interplay bctween the architecture of the expert system and the type of knowledge that ca¡r be naturally expressed in it. The majority of the expert system softrvare is st,ill intended for stand alone, off-line applications. Real- time capacities a¡e only available in ferv cases such as, e.9., G2 (Gensym, 1987), PICON (Moore el aI, 1985), and Muse (CCL, 1987). Section 2 describes the overall organization of an ex- pcrt system framework that has been developed and ,-/-\ Figure 2. Ove¡all implementation structure. implemented on a VAX lI1780. The frame-*'ork rvas developed for knowledge-based controi applications but is not restricted to it. The system has been uscd for design of intelligent tuning controllers (,4.r2én, 1987). In Section 3, the architecture of the expert system part of the controller is described t,ogether. Implementational issues are described in Section 4. Finally, Section 5 contains a discussio:r .rbout the sys- tem and a comparison *'ith other systems. 2. Overall architecture The knowledge-based controller consists of two major parts: the numerical algorithms and the knowledgc- based system. To assure that the execution of the nu- merical algorithms are not delayed by the knowledge- based system the parts are implemented as two com- municating concurrent VMS processes wlìere the nu- merical algorithms have the highest priority. The man-machine interface is implemented as a sepa- rate process. From tiris process, the user can interact directly with the knorvledge-based system and incli- rectly with the algorithms. The timer process is used t,o implement certain real-time interrupts described in the next section. The overall structure is shown in Fig 2. The knorvledge-baseci syste¡n and thc rnan-nrachi¡re interface are both writtcn in Lisp. Thc Lisp uscd is the UnLx dialect Franz Lisp, (Foderaro et aI, I?BB). The software package EUNICE, (Kashtan, 1g82), is used to create a Unix environment under Vi',{S. The reason for using Lisp is mainly its powerful syrnbolic processing capabilities. A drawback rvith Lisp in a real-time system is the garbage collection. The probiem can be avoided in t.rvo rvays. By using a Lisp system with i¡rcremental garbage coliection the garbage collection activily is spread uniformly in time and performed in the background. Thc seconcì approaclÌ is to write Lisp code that does not generate any garbage. This is rvhat is done in G2. Tlm. bo Process R.. u lt bor Timer System Basod Numerical algorithm Man-mach¡n€ communlcatio The nurnerical algorithms The numerical algorithm process is written in Pas- cal. It consists of a library of different algorithms such as PID algorithms, pole-placement algorithms, discrete filters, relays, recursive least-squares algo- ritlrrns, lcvcl crossing dctccl,ors, ctc. Thc algorilluns are uniformly coded and have well-defined interfaces. It is relatively straight forward to add new types of algorithms to the system. The process is connected to A/D and D/A converters. The algorithms can principally be divided into three groups: control algorithms, identification algorithms and monitoring algorithms. The cont,rol algorithms all compute a control signal based on command and measurement signals. Only one control algorithm can be running at a time. The identification and moni- toring algorithms all in some sense extract informa- tion from the numerical signal ff.ow. This informa- tion is sent to the knowledge-based system. The al- gorithms in these two groups can be viewed as filt,ers or feature extractors that send information to the knowledge-based system only when something signif- icant has happened. During steady-state operation, the knowledge-bascd system is not i¡volved and t,he system resembles a conventional controller. The sep- aration between the numerical algorithms and the knowledge-based system is favourable from the point of information flow. If a knowledge-based system was interfaced directly to a physical process or to an exist- ing control system, numerical information would have to be senú forth and back again at a high rate. The knowledge-based system also had to itself extract all useful symbolic i¡formation from the signals. This is a task that often is expressed in the form of numer- ical algorithms. Using expert system techniques for such tasks is often inefficient. Inter-pro cess cornnlunication The processes communicate by sending messages through mailboxes shown as rectangles in Fig. 2. The messages that are sent to the algorithms via Outbox are confìguration commands, parameter changes, and information requests from the knowledge-based sys- tem. The messages t,hat go to the knowledge-based system contains results obtained by an algorithm, alarrns that have been detected, answers to infor- mation requests, user cornmands, and timer inter- rupts. Messages to the knowledge-based system are normally sent to Inbox which is a standard first-in, first-out VMS ¡nailbox. Messagcs wilh prioritics in- dicating their importance are allowed, This is made possible by having an inte¡nal maill¡ox inside ùhe knowledge-based system process into which the mcs- sages in Inbox are inserted according to their pri- ority. Important messages such as alarms and timer interrupts have a high priority and are thus taken care of as fast as possible. Answerbox is used for re- sponses to information requests that, has been made by ihe knowledge-based system. Resultbox is used by the knowledge-based system to rcturn results to Lltc man-machine intcrfacc. The use of Lisp has interesting consequences for the communication between the knowledge-based system and thc ¡nan-¡n¿rchinc intcrfocc. Muilboxcs can bc seen as text files whcre a text line corrcsponcls to a message. In Lisp there is no syntactical difierence between dat,a objects and program code, i.e., Lisp functions. Lists are r¡sc<l to rcprcscnt both. This makes it possible to implement remote evaluation of Lisp functions. An arbitrary Lisp function can l¡e sent to the knowlcdgc-based systern from thc rna¡r- machine interface wherc it is cvaluatcd and the rcsull is returned in Resultbox, erronous Knowledge-based system architecture A standard ofÏ-the-shelf expert system framework, OPS 4, was used to implement the knorvledge-based part of the system in a first prototype (.Â.rzén, 1g86a, 1980b). OPS 4 (Forgy, 1979) is a simple rule-based, forward-chaining expert system framework. The rea- son for this choice was the data-driven nature of knowledge-based control. Data in the form of signif- icant events detected by the algorithms are sent to the knowledge-based system which should react and generate some response. The framework OPS 4 uses the incremental pattern-matching algorithm RETE (Forg¡ 1982) and is therefore also reasr¡nably fast. Another reason for the choice was simply that the system was alrailable to us and that rve rvanted to test the basic ideas rapidl¡'. Experiences of a prototype The first prototype was used to implement a relay- based PID auto-tuner (Åström and Hägglund, 1984). Ðxperiments with the prototype gave many results concerning both the feasibility of the approach and the demands on a expert system framex'ork for knowledge-based control. We were reassured in that the approach is feasible. The response times for the knowledge-based system were acceptable. The sam- pling rate for the numerical aigorithm process was 1 sccond. It took approximatively 2-3 sampling pe- riods from that a message tvas sent to the until a responding message was returned. A second positive result was a clean implementation of a relay auto- tuner that clearly benefited from the separation of logic and algorithms. The tin.re and ellort to make extensions to the controller were significantly smaller tha¡r for comparablc irnplcrncntaLions i¡r co¡rvc¡r{,i<.¡¡r¿rl languages. The negative experiences all concerncd OPS4. It be- camc clear that a sirnple forward-chaining rule sys- tem is not sufficient. An expert system framework must allow for a modular decomposition of both the rulebase and the database. For the database, this could be achieved by a frame system. The rulebase must be decomposable into rulc groups for the dilTer- ent subtasks. Furthcr it was found thaü one knowl- edgc represcntation technique is not enorrgh. Somc subtasks contai¡rs large sequential elements. 'Ihcse are mo¡e naturally represented procedurally tlran 3 Global Database (Blackboard) Schedu ler Knowledge sources TJ Figure 3. Knowledge based control st¡ucture with rules. It was also clear that backward chaining inferencing would be useful for some problems. Thc most, important drawback with OPS 4 was, however, that it is not designed for real-time operation. It has, e.9., no possibilities to havc timc-outs associatcd with database elements, no means for halting the rule exe- cution for a certain time, and no possibilities to check rules at given time intervals. A blackboard system Based on the experiences of the prototype, a real- time expert system framework has been developed. The reasoning model chosen as the basis for frame- work is the blackboard model, (Nii, 1986). A globat database, the blackboard, is available to dilïerent, co- operating knowledge sources. The database allows for frame structures for storing associated information. The knowledge sources can be thought ofas different actors, each ofwhich solves some subtask ofthe prob- lem. The knowledge sources also have thei¡ own local databases. Knowledge sources can be rule-based with either forward or backwa¡d chaining and procedural, The structure of the framework is shown in Fig. B. A knowledge source implements the domain knowl- edge for a certain task. It is ofien associated with one or more numerical algorithms. It could for example contain the hcrrristic logic surrounding an algorithm. The knowledge sources have primitives for adding, modifying, and deleting frames both gtobally and lo- cally. They also have primitives to halt their execu- tion for a certain time or until a certain dat¿base element is added to the blackboard. It is possible to have forward chaining rules that are tested with spe- cific time i¡terv¿ls and to associate validiiy intervals with database elements. The operation of the knowledge-based controller in- volves the actiyation of different knowledge sources both in sequence and in parailel. A typical case when knowledge sources are active in parallel is during the steady state control of the process. One knowl- edge source takes care of the actual control algorithm while other knowledge sources implernent diferent monitoring aspects. A separate rule-based module schedules the selcc- tion of knowledge sourccs at two different levels. The first level involves the sequential activation of dif- ferent knowledge sources. The second level involves the scheduling between different knowledge sources that are active simultaneously. This resembles the scheduling in an ordinary real-time, multi-tasking operating system where the knowledge sources are thc equivalents of concurrcnt proccsscs, Â knos,lcclgc source runs until it explicitly rcturns control to thc scheduler, e.g,, if it has to wait for some information or if it is finished. A simple extension rvhich allo.*,s interrupts among the knowledge sources is described in Årzén (i987). With this extension, priorities can be associated with knowledge sources. The scheduìer contains frames with inlormation of the state of the different knowledge sourccs and fra¡ncs rvhich co¡r_ tains information about the conditions on *,hich a knowledge source is waiting. Thc prinritivcs that involvcs waLil,ing ¿ ccrt¿¡in l,i¡nc are implemenied with the help of the timer process. A primitive that causes a knowledge source to .rvait a certain time gives rise to a mcssagc to thc timcr process. The message contains the desired wakeup time and a unique identifier for the rvaittime request. A high-priority message is returned to the scheduler when the waiting time has elapsed. This message causes the state of the waiting knowledge source to be changed to ready. Knowledge source combination The operation ofthe knowledge-based controller typ- ically consists of a sequence, with parallel parts, of knowledge source activations. Three diffe¡ent meth- ods for combiniag knowledge sources into scquences have been implemented. The most straightforward way is to use primitives that let knowledge sources acti'r'ate and deactivate each other. A knorvledge source has the possibility to wait until another knowledge source is finished. Procedural knowledge sources also have t,he possibil- ity to call other procedural knorvledge sources, and await and use their returned result. Anothcr alLcrnativc is to havc ¿l nunrL¡c¡ of ¡-rrc-sl,orcd sequc¡rces. One example of a sequence could be the i¡ritial tuning sequence. Othcr sequences could be used to ¡eturn to steady-state control when different alarm conditions have been de tected. Combination of knowledge sou¡ces into sequences is b¿sically a pro- cedural operation. It is therefore natural to express it with procedural knowledge sources. In order for this to be possible, rvait primitivcs that allows *'aiting for conjunctions and disjunctions of multiple events have been implemented. Tire last, and most complex mcthod is to dynanri- cally generate sequences. This is accom¡rlished by as- sociating goal states, i.e., post-conditions, and ini- tial states, i.e., pre-conditions, with each knowledge source. Each knowledge source can be vierved as an opcrator that transforms tìrc staie of the system front its initial statc to its goal statc. A sequence is recu¡- sivcly generated b¡. comparing the desircd goal and the current state u'ith the pre- and post-conditions of the operators. This formulation turns the problem into a planning problem. The scheduler generates a plan which then is executed. The possibility for diferent knowledge representa- tion techniques allows the user to choose the tech- nique most natural for each sub-problem. The v¿rious methods of combining knowledge sources give a rich and flexible structure. For instance, it is possible to have one knorvledge source that contains monitoring rulcs which are checked periodically. If something er- roneous is detected the rules can invoke other knorvl- edge sources that focuses on the problem. These knowledge sources could, e.g. be backward chainers that tries t,o verify some hypothesis concerning the error or proceclural knowledge sources that performs some procedural tests. Meta-knowledge sourccs with knowledge about, the applicability of othcr knowlcdge sources are also easy to implement. A knowledge source in a knowledge-based control application could be, e.g., contain knowledge about design of different controllers. Another knowledge source could contain knowledge about modelling ¿¡rcl ¡noclcl vnlidation. Othcr cxam¡rlcs cor¡lrl co¡rlni¡r knowledge of dillerent monitoring aspects of the con- troller. The possibility to refer to past signal values is important in a real-time environment. This is pos- sible through statistics knowledge sources that com- putes signal statistics over different time horizons. These knowledge sources are associated witîr numer- ical algorithms that collect the signal values. The described framework has been used for the de- sign of elaborate extensions of relay auto-tuning. This is described in ,A,rzén (1987). 4, Implementation The implementation of the expert system framework is built on the object-oriented system Flavors (Can- non, 1982) and the forward-chaining production sys- tem YAPS (Allen, 1983). The YAPS system is a pattern-matching system i¡ the same spirit as the OPS family with a similar optimized, incremental matching algorithm. The important difference is that YAPS is v¡ritten in Flavors and allows Flavor in- stances in its database. These Flavor instances can be instances of other YAPS systems. The YAPS system originally only allows arbitrarily nested list structures of containing numbers, atoms, and Flavor instances as database elements. The sys- tem has been modified to allow frame structures. The system has also been extended to allow for auto- matic explanations of how database elements have l¡een added to the system. The Scheduler is implement,ed as a flavor which in- herits a YAPS flavor. The scheduling strategy is rep- resented with rules. The different types of knowledge sources are implemented as diflerent flavors. Each in- dividual knowledge source is an instance of the cor- responding flavor. Each knorvledge source is repre- sented as a frame in the scheduler database. The frarne contains slots for the type of knorvledge source, Scheduler - YAPS system Figure 4. Implementation stnrcture e.g., forward or procedural, for the state of the knorvl- edge source, and for the actual flavor instance that implements the knowledge source. The implementation structure is illustrated in Fig, 4. The actual interface between the knorvledge sources and the schedrrler consists of a relat.ively small sct of messages for which the knowledge sou¡ce flavors should supply methods. This makes it easy t,o add new types of knowledge sources to the system. A slightly simplified example of a rule in the sched- uler is given below. (p schsdul€l 'rIf a knorledgo sourco is ready a¡d no othor knoçledge sourco is runaing thsn ru¡ this knosl_edge sourcs" (f ra¡e k¡o¡ledgo-source status activo 6tato t6ady instance -x) (' (fra¡oe krrorledge-source stato running)) (nodify 1 stat€ running) (<- -¡ 'run)) The forward chailing knowiedge sources are imple- mented as instances of a flavor that inherits a YAPS flavor. This givcs a st¡ucture where several YAPS sys- tcms residc as datab¿rsc clements insidc tirc Schcdulcr YAPS system. The backward chaining knowledge sources are based on an small expert systcm example in Winston and Horn, (1981), which has bcen extended and embed- ded in Flavors. Currently they can only ask qucs- tions to the operator when an answer cannot be au- tomatically deduced. A possible extension rvould be to, in that case, allow a backward chaining knowÌedge source to invoke a lorward or procedural knowledge source that returns the needed atìswer. The procedural knowledge sources cor.rsists of Lisp functions. In order to allow interrupts of these func- tions at arbitrary places the entire state of the Lisp computation must be saved. This is not possiblc rvith the Franz Lisp of the basic system. Inste¿Ld Lisp has been used to write a register nìacllirìe basecl intcr- preter for a procedural Lisp-like language that allorvs the computation to be suspended. ,._ìì I Knowledge ü:Þ Global Database Rules Scheduling rules 5. Summary Àrr gcrrcrul cx¡rcrL systcrrr frur¡rcwork fcrr rcul-tirrrc u¡_l- plications has becn prescntcd. It has becn dcvelopcd for knowledge-based control applications but is not restricted to it. The framework has real-time facilities. It is modu- larized into knowledge sources that can be compared with concurrent processes. This is similar to thc Muse system. In the current version, however, the knorvl- edge sources cannot interrupt each other. With a sim- ple extension this is possible. The knowledge sources have primitives to wait a certain time or for a cer- tain database element. These primitives are used to implement periodic rule testing in a way similar to G2 and Picon. Validity intervals can be used to indicate how long database elements remain valid. In contrast v¡ith G2 and Picon the validity interv¿ls are not propagated to inferred facts. History values of important signal values are maintained. It is not possible to store history va.lues of arbitrary frame attributes. The system allows for both rule-based and procedural representation which is very important, The flexible means of combining knowledge sources gives a rich structure. 1ìrere are many similarities between real-time op- erating systems and real-time knowledge-based sys- tems. Real-time operating systems for process control have evolved over a long period of time. This paper indicatcs a new system architccturc where real-time operating systems, databases, object-oriente(i pro- gramrning, and knowledge-based systems are com- bined. The exccution speed of the system is of the orcler of one forward chaining rule per second, The system is currently being ported to a Symbolics - IBM PC en- vironment where the knowledge-based system resides on the Symbolics and the numerical algorithms reside on the IBM PC. Preliminary results indicate a factor of 10 in increased speed. Other possible candidates for migration are systerns where powerful symbolic processing capacity is combined with conventional cornputing. One example of this is the ¿z-Explorer. ,4. cknorvle dge rncnts The author rvould like to thank Professor I{arl Johan Âström and Sven Erik Mattsson for many useful discussions. This work has been supported by the National Swedish Board for Technical Development (STU) under contract 85-3084. Refe¡ences ALLnu, E.M. (1983): ,,yApS: yet another production system," TR-1146, Department of Computer Science, University of Maryland. ÂnzÉN, K-8. (19S'6a): ,,Expert systems for process control,' in D. Sri¡am and R. Adey (trds.): proc. of Fj¡st Intetnational Conference on Applications of Artifrcial lntclligence in Engineering Practice, Springer Verlag, Berlin, pp. 1127-1138. .ÅnzÉx, K-8. (1986b): "Usc of expert svstems in closed loop feedback control," Proc. of American Control Conference, Seattle, Wr\. ÅnzÉx, K-8. (19S7): "Realization of cxpert system bascd feedback control," Ph.D. thesis CODEN: LUTFD2/ TFRT-1029, Department of Automatic Control, Lund Llul,iùutc of 'l'cchnology, Lund, Swcdcr¡. ÂsrRölt, K.J. and J..J. ANro¡ (1984): ,,Experr control,,, Proc. 9'th IFAC World Congress, Budapest, IIungar5,. Ås'rRöu, K.J. and 'I'. IIÃccLUND (1984): ,,Autorrratic tuning of simple regulators," Proc. IFAC g'th l!¡orld Co ngress, B udapest, Ilungary. Åsrnö¡'¡ K.J. and B. W¡TTnu¡¿lnx (fOSO): Adaptit,e Control, To appear, Addison-lVesley, Reading, lr,fA. ÅsrRölr, K.J., J.J. Ar.row and K.-E. ÅnzÉN (1986): "Expert control," Automatica, 22, 3, 277-286. CaNuoll, H.I. (1982); "Flavors: A non-hierarchical ap- proach to object-oriented programming,," unpublished paper. CCL (1987): "Muse product description,', Cambridge Consultants Liroited, 4 pages. Cu.lxrr,er, M.J. (1988): ,'Real-time aspects of experr systems in process control," Colloquium on Expert Systems jn P¡ocess Control, IEE, Savoy Place, Lonclon UK. FoouR,o.Ro, J.K., K.L. Sxr,ownR and K. Lay¡R (19g3): "The Franz Lisp Manual," ',JC Berkeley, California. FoRcY, C.L. (1979): "OPS4 IJser's manual," Technical report CMU-CS-79-132, Department of Computer Sci- ence, Carnegie-Melion University, FoRcY, C.L. (1982): "Rete: À fast algorithm for the many pattern/many object pattern match problem," Artifr.cial Intelligence, 75, l, 17-37. GnNsyl,t (1987): G2 User's manual, Gens¡.m Corp.. Cambridge, MA. IsERlaaNN, R. and K.H. Llcuuaun (1985): ,,pa¡ame- ter-adaptive control with configuration aids and sr.rper- vision functions," Äuf,omalica, 2L, 6, 625-638. K.a,snT,nN, D.L. (1982): "EUNICE: A system for porring UNIX programs to VAX/VMS," Artificial Intelligence Center, SRI Inte¡national, \,Ienlo Park, California. L,rrrnv, 1'.J., P.A. Cox, J.L. Scurvrrot, S.tr,I. K¡o. and J. Y READ (1988): "Real-time knowledge-based systems," AI Magazine, I, I, 27-45, MooRrì, R.L., L.R. IIrrwxrwsor,r, M.lì. i.¡;vtN;r¡rrì C.G. I(xrcxnRBocKÐR (1gS5): ,,Expert control,', proc. American Conttol Conf., Boston, \,f Ä, pp. ES5-887. Ntt, lI.P. (i986): "Blackboa¡d sy6tems: t he Ì¡lackl¡oarrl model of probiem solving and the evolution of black- board a¡chitectures," AI Magazine, 7, 2, 38-53. ToNc, R.M. (198a): "A retrospective vicw of ftzzy con- trol systems," Fuzzy Sets and ,Sysúerns, 14, lgg-210. WaRvtcr, K. (1988): Implementation of self-tuning controllers, Peter Peregrinus, London. 'Wnts,oN, P.H. and B.K.P. HoRN (1981): Iisp, Ad<ìi- son-\\Iesley, Reading, MA, 
work_l56qdmdo5ffbheakeupbix4od4	----	paper.dvi GRASP with Path Relinking for Commercial Districting Hugo Jair Escalante Computer Science Department Instituto Nacional de Astrof́ısica, Óptica y Electrónica (INAOE) Tonantzintla, Puebla, Mexico hugojair@inaoep.mx Roger Z. Ŕıos-Mercado1 Graduate Program in Systems Engineering Universidad Autónoma de Nuevo León (UANL) San Nicolás de los Garza, NL, Mexico roger.rios@uanl.edu.mx March 2014 1Corresponding author Abstract The problem of grouping basic units into larger geographic territories subject to dispersion, connec- tivity, and balance requirements is addressed. The problem is motivated by a real-world application from the bottled beverage distribution industry. For addressing dispersion minimization, a more robust measure based on the diameter of the formed territories is used. For solving this particular territory design problem, a greedy randomized adaptive search procedure (GRASP) that incorpo- rates a novel construction procedure where territories are formed simultaneously in two main stages using different criteria is proposed. The GRASP is further enhanced with two variants of forward- backward path relinking, namely static and dynamic. The proposed algorithm and its components have been extensively evaluated over a wide set of data instances. Experimental results reveal that the construction mechanism produces feasible solutions of acceptable quality, which are improved by an effective local search procedure. In addition, empirical evidence indicate that the two path relinking strategies have a significant impact on solution quality when incorporated within the GRASP framework. The ideas and components of the developed method can be further extended to other districting problems under balancing and connectivity constraints. Keywords: Service industry; Districting; Metaheuristics; GRASP; Path relinking. 1 Introduction The territory design problem (TDP) may be viewed as the problem of grouping small geographic basic units (BUs) into larger geographic clusters, called territories, in a way that the territories are acceptable (or optimal) according to relevant planning criteria. Territory design or districting has a broad range of applications such as political districting, sales territory design, school districting, power districting, and public services, to name a few. The reader can find in the works of Kalcsics, Nickel, and Schröder [17] and Duque, Ramos, and Suriñach [11] state of the art surveys on models, algorithms, and applications to districting problems. The problem addressed in this paper is a commercial territory design problem (CTDP) moti- vated by a real-world application from the bottled beverage distribution industry. The problem, introduced by Ŕıos-Mercado and Fernández [28], considers finding a design of p territories with minimum dispersion subject to planning requirements such as exclusive BU-to-territory assign- ment, territory connectivity, and territory balancing with respect to three BU attributes: number of customers, product demand, and workload. An important criterion in territory design problems is compactness. Typically this is achieved by minimizing a dispersion function. In commercial territory design, several models based on dispersion functions from the well-known p-center and p-median location problems have been studied in the past. These are center-based dispersion functions, that is, the dispersion is measured with respect to a centroid of a territory. However, there are other non-center-based measures of dispersion that can be used. Center-based functions rely heavily on the location of the centers; if the centers are “badly” located, the resulting design may cause a serious deterioration in objective function. In addition, in location problems, the centers represent a physical entity or facility that provides some service; however, in CTDPs the centers are artificially located as no facility is actually placed there, it is just a reference for the dispersion measure. These limitations motivate the study on other ways of measuring dispersion. For instance, a measure such as the diameter, which measures the longest distance between any two basic units in a territory, is a more robust function since it does not depend on a center location, providing more flexibility. Even from the algorithmic perspective, heuristic methods for tackling TDPs under center-based dispersion functions need to constantly 1 update and recompute as centers keep moving along every time the territory suffers a change. This time-consuming task can be avoided if other measures such as the diameter are used. In this work, we focus our study in a commercial territory design problem that seeks to minimize territory dispersion based on a diameter dispersion measure. To the best of our knowledge, this type of problem has not been addressed before in the territory design literature. Since the aim is to target large instances, we present a Greedy Randomized Adaptive Search Procedure (GRASP) with Path Relinking for this NP-hard CTDP. The algorithm is denoted as GPR CTDP. In our proposed GRASP we develop a procedure that builds exactly p territories at once simultaneously, that is, we start with p node seeds and start associating nodes to the seeds until all of them are assigned. By growing the territories simultaneously rather than one at a time one expects that the violation of the balancing constraints be considerably lower. In addition, we develop two path relinking (PR) strategies, one dynamic and one static, motivated by the work of Resende et al. [21], who successfully applied it to the max-min diversity problem. In our work, these PR strategies rely on finding a “path” between two different territory designs. To this end, an associated assignment subproblem for finding the best match between territory centers is solved. The solution to this problem provides a very nice way of generating the trajectory between two given designs. This idea is novel in any districting of territory design application to the best of our knowledge. To assess its efficiency, the proposed GPR CTDP with many of its components and strategies, has been extensively evaluated over a wide set of data instances. We have found, for instance, that building territories simultaneously results in feasible solutions of acceptable quality. The two PR variants implemented in GPR CTDP allowed us to obtain better solutions than those obtained when using straight local search; although, the static strategy resulted more helpful. The main algorithmic ideas incorporated in the developed algorithm can be extended so as to handle other districting problems with similar structure. The paper is organized as follows. In Section 2 we describe the problem in detail and present a combinatorial optimization model. Section 3 gives an overview of relevant previous related work. Section 4 describes in detail the components of the proposed heuristic, and Section 5 presents the empirical evaluation of the method. We end the paper in Section 6, with some conclusions and 2 final remarks. 2 Problem Description Let G = (V,E) denote a graph where V is the set of city blocks or basic units (BUs), and E is the set of edges representing adjacency between blocks, that is, (i,j) ∈ E if and only if BUs i and j are adjacent blocks. Let dij denote the Euclidean distance between BUs i and j, with i,j ∈ V . For each BU i ∈ V there are three associated parameters. Let wai be the value of activity a at node i, where a = 1 (number of customers), a = 2 (product demand), and a = 3 (workload). The number of territories is given by the parameter p. A p-partition of V is denoted by X = (X1, . . . ,Xp), where Xk ⊂ V is called a territory of V . Let w a(Xk) = ∑ i∈Xk wai denote the size of territory Xk with respect to activity a ∈ A = {1,2,3} and k ∈ K = {1, . . . ,p}. The balancing planning requirements are modeled by introducing a user-specified tolerance parameter τa that measure the allowable relative deviation from the target average size µa, given by µa = wa(V )/p, for each activity a ∈ A. Another planning requirement is that all of the nodes assigned to each territory are connected by a path contained totally within the territory. In other words, each of the territories Xk must induce a connected subgraph of G. Finally, we seek to maximize territory compactness or, equivalently, minimize territory dispersion, where dispersion is given by the largest diameter over all territories, that is maxk=1,...,p maxi,j∈Xk{dij}. Let Π be the collection of all p-partitions of V . The combinatorial optimization model is given as follows. Model (CTDP) min X∈Π f(X) = max k∈K max i,j∈Xk {dij} (1) subject to wa(Xk) µa ∈ [1 − τa,1 + τa] k ∈ K, a ∈ A (2) Gk = G(Vk,E(Vk)) is connected k ∈ K (3) Objective (1) measures territory dispersion. Constraints (2) represent the territory balance with 3 respect to each activity measure as it establishes that the size of each territory must lie within a range (measured by tolerance parameter τa) around its average size. Constraints (3) guarantee the connectivity of the territories, where Gk is the graph induced in G by the set of nodes Xk. Note that there is an exponential number of such constraints. The model can be viewed as partitioning G (the contiguity graph representing the BUs) into p connected components (contiguous districts) under the additional side constraints on balancing product demand, number of customers, and workload of each territory, and minimizing a dispersion measure of the BUs in a territory. The basic contiguity graph model for the representation of a territory divided into elementary units has been adopted in political districting [26]. 3 Related Work Territory design or districting has a broad range of applications such as political districting [4, 15, 3, 18, 27, 20], sales territory design [10, 32], school districting [6], power districting [1, 8], and public services [2, 7, 19], to name a few. The reader can find in the works of Kalcsics, Nickel, and Schröder [17] and Duque, Ramos, and Suriñach [11] state of the art surveys on models, algorithms, and applications to districting problems. Zoltners and Sinha [33] present a survey focusing on sales districting and Ricca et al. [26] present a survey on political districting. Here we discuss the related work on commercial territory design. Ŕıos-Mercado and Fernández [28] introduced the commercial TDP by incorporating a territory compactness criterion and a fixed number of territories p. They seek to maximize this compactness criterion subject to planning requirements such as exclusive BU-to-territory assignment, territory connectivity, and territory balancing with respect to three BU attributes: number of customers, product demand, and work- load. In their work, the authors consider as a minimization function a dispersion function based on the objective function of the well-known p-Center Problem. After establishing the NP-completeness of the problem, the authors propose a Reactive GRASP for obtaining high-quality solutions to this problem. The core of their GRASP is a three-phase iterative procedure composed by a construction phase, an adjustment phase, and a local search phase. In the construction phase a solution with q territories, where q is usually larger than p, satisfying the connectivity constraints is built. Then 4 an adjustment phase based on a pairwise merging mechanism is applied to obtain a solution with p territories. Afterwards, a local search phase attempting both to eliminate the infeasibility with respect to the balancing requirements and to improve the dispersion objective function is applied. One interesting observation is that the construction and adjustment phases produce solutions with very high degree of infeasibility. This is very nicely repaired by the local search, at a very high computational cost though. The reason for this is that attempting to merge two territories into one in the adjustment phase may result in a high violation of the upper bound of the balancing constraints. Aguilar-Salazar et al. [30] present an exact optimization framework for tackling relatively small instances of several CDTP models. They studied two linear models that differ in the way they measure dispersion, one model uses a dispersion function based on the objective of the p-Median Problem (MPTDP) and the other is based on the p-Center Problem (CPTDP). They can success- fully solve instances of up to 100 BUs for the CPTDP and up to 150 BUs for the MPTDP. This concludes that p-center-based dispersion measures yield more difficult models as they have weaker LP relaxations than the median-based models. Salazar-Acosta and Ŕıos-Mercado [29] present a heuristic based on GRASP and adaptive mem- ory programming for a CTDP that considers the minimization of a p-Center Problem function subject to additional budget routing constraints. One of the most popular methods for addressing districting problems is the location-allocation technique [17]. However, this technique is not applicable to our problem mainly because the nature of the dispersion objective function is different. As it has been show, the location-allocation method seems to work well when a p-Median Problem-based objective function is used. As stated previously, CTDPs with diameter-based dispersion functions have not been studied in the past. One of the reasons is that center-based functions yield well-structured mixed-integer programming problems which in turn can lead to relatively good optimization algorithms. However, this structure is somewhat lost when addressing a problem from the heuristic perspective. In fact, using non-center-based functions such as the diameter may be more convenient since no time consuming center updating operations are needed. 5 4 Proposed Heuristic This section introduces the proposed GRASP heuristic with path relinking for the commercial terri- tory design problem (GPR CTDP). GRASP is a well known meta-heuristic based on greedy search and random construction mechanisms [14] that has been successfully used for many combinatorial optimization problems, including CTDP [28]. We propose a GRASP improved with path relinking (PR). The heuristic comprises a new construction procedure and a very effective PR mechanism. The construction procedure intelligently handles a strategy for building terriories simultaneously, while the PR formulation allows us to obtain better solutions than those obtained when using straight local search, see Section 5. The rest of this section describes in detail the components of the GPR CTDP approach, which receives as input an instance of the CTDP and a set of parameters as described below. 4.1 GRASP A GRASP is an iterative process in which each major iteration consists of two phases: construction and local search [14]. The construction phase attempts to build a feasible solution and the local search phase attempts to improve it. This process is repeated for a fixed number of iterations and the best overall solution is returned as the result. GRASP incorporates greedy search and randomization mechanisms that allow it to obtain high quality solutions to combinatorial problems in acceptable times. Despite the simplicity of this multi-start heuristic it has proved to be very effective in a wide range of problems and applications [22]. Previous work on GRASP for the CTDP is presented in Section 3. In this paper we propose procedure GPR CTDP, which is in essence a GRASP augmented with PR mechanisms, accordingly, in this section we describe the particular construction and local search procedures of the GRASP and the next subsection presents the PR strategies. Construction phase At a given iteration, the construction phase consists of building p territories, X1, . . . ,Xp, simulta- neously in such a way that connectivity is always satisfied while infeasibility in terms of dispersion 6 and balance is allowed to some extent. Each territory Xk is formed by a subset of BUs or nodes such that ∪k=1,...,pXk = V and Xk ∩ Xl = ∅, for all k 6= l. Under the proposed procedure each territory Xk is associated to a center, c(k). This is not a requirement of the problem but a feature of the proposed formulation that was adopted for convenience when measuring dispersion of territories. Procedure 1 grasp construction( δ, L, α ) Input: δ: fraction of nodes assigned by the distance criteria; L: interval for updating centers; α: RCL quality parameter; Output: X: A p-partition of V ; (c(1), . . . ,c(p)) ← max disp( p ); {Compute p initial centers} i ← 0; V̄ ← V ; while ( n−|V̄ | ≤ δn ) do for all ( k ∈{1, . . . ,p} ) do Nq(Xk) ← q nearest (unassigned) neighbors of Xk; Xk ← Xk ∪Nq(Xk); V̄ ← V̄ \Nq(Xk); end for i ← i + 1; if ( i module L = 0 ) then c(k) ← min(maxdv,w), ∀v,w ∈ Xk, k = 1, . . . ,p; {Update centers} end if end while open(k) ← TRUE, k = 1, . . . ,p; while ( |V̄ | > 0 and ∃k such that open(k) == TRUE ) do for all ( k = 1, . . . ,p ) do if ( open(k) = TRUE ) then Compute φk(v) in Eq. (4), ∀ v ∈ N(Xk); Φmin ← min{φk(v)}; Φmax ← max{φk(v)}; RCL ←{h ∈ N(Xk) : φk(h) ≤ Φmin + α(Φmax −Φmin)}; Choose v ∈ RCL randomly; Xk ← Xk ∪{v}; V̄ ← V̄ \{v}; if ( N(Xk) = ∅ or w a(Xk) > (1 + τ a) for any a ) then open(k) ← FALSE; {Close this territory} end if end if end for end while if (|V̄ | > 0) then for all ( v ∈ V̄ ) do Xv ← Nearest territory to node v; Xv ← Xv ∪{v}; V̄ ← V̄ \{v}; end for end if return X = {X1, . . . ,Xp}; Procedure 1 presents the construction phase of the proposed GPR CTDP. V̄ denotes the set of nodes that have not been assigned to any territory and n = |V | the number of BUs. The process starts by selecting p seeds or centers, {c(1), . . . ,c(p)}, which are the first nodes assigned to each territory; that is, c(k) ∈ Xk, k ∈ {1, . . . ,p}. Territories are then built iteratively in two main stages followed by a postprocessing stage. In the first stage q BUs are iteratively assigned to each territory 7 Xk. For each territory Xk, we iteratively assign the q (unassigned) nearest neighboring nodes of that territory, v ∈ Nq(Xk). The BUs in Nq(Xk) that are assigned to Xk must be connected by an edge to a BU already assigned to Xk. The latter process is iterated until a fraction δ of the total of BUs have been assigned to one of the p territories (i.e. ⌊δn⌋ BUs have been assigned), where the centers c(1), . . . ,c(p) are updated every L iterations. One should note that the notion of centers is only used for this very-first phase of the construction procedure and it is not used elsewhere. Figure 1 shows the BUs assigned after stage one of the construction phase for an instance of the CTDP considered for experimentation. From this stage the p territories have been simultaneously built by using a neighborhood criteria completely ignoring the balance constraints. The rationale behind this is that nodes that belong to the same territory must be close to each other, hence a portion of nodes can be assigned with a closeness criterion. The remaining nodes will lie at boundaries among territories, therefore, balance and dispersion information is taken into account for assigning those nodes. 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 X − coordinates Y − c o o rd in a te s Figure 1: First stage of the proposed construction procedure for an instance of the CTDP. An important aspect of stage one is that of selecting seed centers. Clearly, randomness must be considered for this process as we want to generate fairly different centers at each iteration of the GPR CTDP approach. To this end, we view the problem of choosing an appropriate set of p initials seeds as a p-Dispersion Problem [12], which is a combinatorial optimization problem that places p points in the plane as far way of each other as possible by using an appropriate measure 8 for maximizing dispersion. In our procedure, we used an approach that selects centers randomly with a maximum dispersion criteria. The particular strategy starts with a randomly selected node as the center for the first territory and the rest of centers are obtained by using a greedy heuristic for the p-dispersion problem [12]. The second stage of the construction phase consists of assigning the remaining n − ⌊δn⌋ nodes that were not assigned in stage one. For this stage BUs are assigned to territories using a greedy randomized adaptive procedure that takes into account both balance and dispersion constraints. For each territory Xk, the cost of assigning every neighboring node v ∈ N(Xk) to Xk is evaluated according to Equation (4). Then a restricted candidate list (RCL) is formed, from which a single BU is randomly selected and assigned to the current territory Xk. This RCL is restricted by a quality parameter α, that is, RCL is formed by those BUs whose greedy function evaluation falls within α percent from the best evaluation. Equation (4) determines the cost incurred when assigning node v to a territory Xk. This cost is determined by a linear combination of the weights assigned to nodes in territory Xk ∪{v}, as determined by the term Gk(v), and the dispersion of those nodes, as estimated by the term Fk(v), with Gk(v) and Fk(v) defined in Equations (5) and (6), respectively φk(v) = λFk(v) + (1 − λ)Gk(v), (4) Gk(v) = ∑ a∈A gak(v), (5) Fk(v) = ( 1 dmax ) f(Xk ∪ {v}) = ( 1 dmax ) max { f(Xk), max i∈Xk,v {div}) } , (6) where f(Xk) = maxk∈K maxi,j∈Xk {dij} is the dispersion measure (as dictated by the objective function) and gak(v) = 1 µa max{wa(Xk ∪ {v}) − (1 + τ a)µa,0} accounts for the sum of relative infeasibilities for the balancing constraints. Here dmax = maxi,j∈V {dij}, the maximum distance between any pair of nodes, is used for normalizing the objective function. One should note that gak(v) represents the infeasibility with respect to the upper bound of the balance constraint for activity a. Both factors dispersion and balancing are weighted by a parameter λ in expression (4). The process is repeated for every territory k. If a territory exceeds the expected average weight for a territory it is considered closed (i.e., open(j) = false) and no further node can be assigned 9 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 X − coordinates Y − c o o rd in a te s Figure 2: Second and third stages of the proposed construction procedure for an instance of the CTDP. to it. The latter process iterates until either every node has been assigned to a territory or every territory is considered closed. Since stage two of this construction phase does not guarantee that all nodes will be assigned to a territory, a third stage is applied in which each unassigned node gets assigned to its nearest territory. Figure 2 shows the distribution of BUs for an instance of the CTDP after stages two and three of the construction procedure. Local search After a solution is build a postprocessing phase consisting of local search is performed. The goal in this phase is to improve the objective function value and recovering feasibility (if violated) in the constructed solution, X. In this local search, a merit function that weights both the infeasibility with respect to balancing constraints and the objective function value is used. This function is indeed similar to the greedy function used in the construction phase with the exception that now the sum of relative infeasibilities take into consideration lower and upper bound violation of the balancing constraints. Specifically, the merit function for a given territory design X = {X1, . . . ,Xp} is given by ψ(X) = λF(X) + (1 − λ)G(X) (7) 10 where F(X) = ( 1 dmax ) max k∈K max i,j∈Xk {dij} (8) and G(X) = p ∑ k=1 ∑ a∈A ga(Xk), (9) with ga(Xk) = 1 µa max{wa(Xk) − (1 + τ a)µa,(1 − τa)µa − wa(Xk),0} being the sum of the relative infeasibilities of the balancing constraints. The quality of solutions is then determined by Expression (7), we now describe the mechanism for exploring solutions around the constructed territory design. Let t(i) denote the territory node i belongs to, i = 1, . . . ,n. A move move(i,j) is defined as moving a node i from its current territory to a territory t(j), where t(j) 6= t(i). Only moves move(i,j) where (i,j) ∈ E and t(i) 6= t(j) are allowed. Thus, move(i,j) transforms a solution X = (X1, . . . ,Xt(i), . . . ,Xt(j), . . . ,Xp) into X T = (X1, . . . ,Xt(i) \ {i}, . . . ,Xt(j) ∪ {i}, . . . ,Xp). If connectivity must be kept, only moves where Xt(i) \ {i} remains connected are allowed. Note that in general move(i,j) is asymmetric. Procedure 2 local search( X ) Input: X: A solution to the CTDP; Output: X: Improved solution to the CTDP; nmoves ← 0; local optima ← FALSE; k ← 1; {starting territory} while ( nmoves ≤ limit evals AND ¬local optima ) do improvement ← FALSE; while ( |N(Xk)| > 0 and ¬improvement) do move(i,j) ← Choose valid move from N(Xk); N(Xk) ← N(Xk) \{(i,j)}; Evaluate ψ(XT ) using Expression (7); if ( ψ(XT ) < ψ(X) ) then X ← XT ; {perform move} nmoves ← nmoves + 1; improvement ← TRUE; kend ← k; k ← (k + 1) mod p; end if end while if ( ¬improvement ) then k ← (k + 1) mod p; end if if ( k = kend ) then local optima ← TRUE; end if end while return X 11 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 X − coordinates Y − c o o rd in a te s Figure 3: Solution found after applying the local search procedure for an instance of the CTDP. The basic idea of the local search is to start the search with a given territory, say territory k, and then consider first the moves emanating from territory k, that is, if we let N(Xk) denote the feasible moves move(i,j) with t(i) = k evaluate first all the moves in N(Xk), and take the best that improves the current solution, if any. If none found, proceed with territory (k + 1) mod p. As soon as a better move is found, perform the move, and restart the search from this new solution XT but setting k + 1 as the starting territory, where k was the last territory examined, that is, in a new move the starting territory is k + 1 and the final territory to be examined is k. By using this cyclic strategy for starting territory we avoid performing many unnecessary move evaluations. A move is performed using a different territory each time until no improvements can be found. In practice an additional stopping criterion: the maximum number of allowed evaluations of the fitness function (limit evals), is added to avoid performing an extensive search for long periods of time. Therefore, the postprocessing step stops when either a local optima is found or the number of moves exceeds limit evals. The postprocessing phase is described in Procedure 2. Figure 3 shows a solution obtained after applying the local search procedure. 4.2 Path relinking Path Relinking (PR) was originally proposed by Glover and colleagues as a way of incorporating intensification and diversification strategies in tabu search [16]. PR consists of exploring the path 12 of intermediate solutions between two selected solutions called starting (XS) and target (XT ) with the hypothesis that some of the intermediate solutions can be either better than XS and XT (in- tensification) or comparable but different enough from XS and XT (diversification). Intermediate solutions are generated by performing moves from the starting solution in such a way that these moves introduce attributes that are present in the target solution. Successful applications of PR in the context of Tabu and Scatter Search are reported in Resende et al. [23]. Despite the fact that PR was originally proposed for Tabu and Scatter search, it has been successfully used with GRASP as well [22, 21]. In the context of GRASP, PR can be considered as a way of introducing memory into the search process. To the best of our knowledge PR has not been used in the context of territory design, although it has been recently applied to the related problem of capacitated clustering by Deng and Bard [9]. Whereas both problems are related, the proposed formulations differ significantly. For example, Deng and Bard did not consider centers in their PR approach and they proposed a single PR variant (at a cluster-level basis). Deng and Bard report experiments with less than 90 nodes and 5 clusters, while in Section 5 we report instances of up to 500 nodes and 10 territories. Different PR variants have been proposed so far each having benefits and limitations in terms of efficiency and efficacy. In this work we consider two variants of forward-backward PR, namely static and dynamic, that have proven very effective in related problems [21]. For excellent surveys on applications of GRASP with PR we refer the reader to the work of Resende and Ribeiro [22]. The so called, forward-backward PR strategies explore the paths between XS and XT in two different ways (i.e., from XS to XT and viceversa) [22]. The main benefit of these strategies is that more and different solutions can be generated, although it has been found that there is little gain over one-way strategies [25]. This can be due to the greediness of usual PR methods, which evaluate every possible solution that can be generated by making a move from a initial solution and choose the move that results in the best intermediate solution [25, 21]. Thus, these methods explore a large number of solutions and, therefore, forward-backward PR does not help to improve the quality of final solutions. In this work we select moves in such a way that a single move is evaluated for generating intermediate solutions. This form of PR is more efficient at the expense of 13 sacrifying the benefit of greedy strategies. Nevertheless, we believe that in the considered setting the use of a forward-backward PR strategy is advantageous. Besides the direction of the search, there are other aspects that make PR strategies different [22, 21]. For example, greedy randomized PR methods form a RCL with candidate moves and select a move randomly as in GRASP [13]. Truncated PR techniques explore partially the tra- jectory between XS and XT . Evolutionary PR consists of evolving a reference set of solutions in a similar way as the reference set is evolved in scatter search [24]. In this work we developed static and dynamic PR strategies that resulted very effective for the CTDP. Both strategies have been successfully used in other combinatorial optimization problems [21]. The rest of this section describes the PR strategies incorporated in GPR CTDP. Recall each solution of the CTDP is an assignment of every node i ∈ V to one of p territories X1, . . . ,Xp. Let t(X,i) ∈ {1, . . . ,p} denote the index of the territory to which node i is assigned according to solution X. Given two particular solutions XS and XT , PR aims at generating inter- mediate solutions or p-partitions in the path starting at XS and finishing at XT . In GPR CTDP intermediate solutions are created by changing t(XS, i), the territory to which node i is assigned in solution XS into the corresponding territory t(XT , i). Because both XS and XT solutions are created independently, and the territory ordering may be arbitrary, it is not clear what territory in XS corresponds to what territory in XT . Hence, a correspondence between territories must be obtained before starting the search process. The problem of finding the best match between territo- ries can be set as an Assignment Problem (AP) by considering the territory centers only. Let C(X) the set of p node centers corresponding to solution X. Then a complete bipartite graph is formed with sets C(XS) and C(XT ), where the cost between node i ∈ C(XS) and j ∈ C(XT ) is given by dij. The AP can be solved in polynomial time. We use one of the most recent implementations of the Hungarian algorithm [5]. A solution to the AP represents a minimum cost assignment between territory centers, and therefore a match between territories. Let M be the solution to AP given by M = {(i1,j1), . . . ,(ip,jp)}. The idea of the PR is then to “transform” each territory Xt(ik) to territory Xt(jk) for each (ik,jk) ∈ M. The rationale for this matching stems from the fact that it is expected that relatively close territories (from different designs) will have many BUs in common. 14 This scheme is illustrated in Figure 4. One should note that the notion of centers is adopted at this stage for convenience, as centers allows us to establish a correspondence between territories in an efficient way. AP subproblem S AP solution Design X TDesign X Figure 4: Illustration of how to set up a search trajectory from two given designs (top) by solving an associated Assignment Problem (bottom). Once that correspondence between territories has been established it is possible to perform moves from one solution XS to another XT . As a consequence, in order to arrive at solution XT starting from XS, every node in XS such that t(XS, i)) 6= t(XT , i) must be moved to its associated territory in XT . We define a PR move, movePR(X S,XT , i), as a function that moves or reassigns a node i from territory t(XS, i) to territory t(XT , i). The move is valid as long as t(XS, i) 6= t(XT , i) and the resulting p-partition remains connected, that is, if and only if Xt(XT ,i)∪{i} is connected and Xt(XS,i)\{i} remains connected. One should note that moves are always made between boundary nodes as it is not possible to exchange a non-boundary node from one territory to another territory in a single move because loss of connectivity. Intermediate solutions between XS and XT are generated by making moves from XS to XT and updating the solution XS accordingly. Clearly, the order in which nodes i are selected may give rise to different trajectories between XS and XT . In this work we chose nodes i in lexicograph- ical order, we also tried a random node selection approach although no difference in performance was obtained. After an intermediate solution is created it is evaluated using Formula (8). The generation-evaluating process is repeated for every node with t(XS, i) 6= t(XT , i) and the process 15 stops when t(XS, i) = t(XT , i) for all i ∈ V . Thus, the PR procedure receives as input a pair of solutions XS and XT , generates and evaluates all of the intermediate solutions from XS to XT and the best intermediate solution XR is returned as output. In the following we denote with PR(XS,XT ) the application of PR starting at solution XS and finishing at solution XT . Procedures 3 and 4 present the static and dynamic variants of PR implemented in GPR CTDP, respectively. Both static and dynamic variants maintain a set of b elite solutions B = {B1, . . . ,Bb}. B is initialized by running the construction and local search procedures for b times. Solutions in B are always kept sorted in ascending order of their objective function value estimated with Equation (8). Procedure 3 grasp pr static( imax ) Input: imax: number of global iterations; Output: Xbest: A p-partition of V ; for all ( i ∈{1, . . . ,b} ) do X ← grasp construction(); Bi ← local search( X ); end for Sort B from best to worst; for all ( iter = 1, . . . , imax ) do XS ← grasp construction(); XS ← local search( XS ); if ( (ψ(XS) < ψ(B1)) or (ψ(X S) < ψ(Bb) and d sol µ (X S,B) > θ) ) then Ej ← closest solution to X S in B with ψ(XS) < ψ(Bj); Ej ← X S; Update B; end if end for Xbest ← B1; for all ( i ∈{1, . . . ,b−1} ) do for all ( j ∈{i + 1, . . . ,b} ) do Apply PR(Bi,Bj) and PR(Bj,Bi) and let X S ← best solution found; XS ← local search( XS ); if ( ψ(XS) < ψ(Xbest) ) then Xbest ← XS; end if end for end for return Xbest; 4.2.1 Static GPR CTDP In the static variant, PR is performed at the end of imax iterations of a typical GRASP. In each iteration of the GRASP a solution is constructed and improved with local search, XS. This solution is compared with the solutions in B. If XS is better than the best solution in B (i.e., B1) or if X S is 16 better than the worst solution in B (i.e., Bb) and is at a distance larger than a given threshold θ from solutions in B, then the most similar solution to XS in B is replaced by XS. Solutions in B are then sorted from best to worst. After imax iterations the static PR starts. Every path between solutions in B is evaluated and the best solution is returned. The distance between XS and solutions in B is estimated as dsolµ (X S,B) = 1 b ∑b i=1 g(X S,Bi), where g(X S,Bi) is the fraction of nodes in X S and Bi that are assigned to different territories; that is, d sol µ (X S,B) is the average number of nodes assigned to different territories in XS and Bi. Alternative measures of similarity/distance between territory designs have been described before, see for example the work by Tavares Pereira and Rui Figueira [31]. However, such measures do not take advantage of the information we have available when solving the AP. That is, those measures do not know the correspondence between territories beforehand. Besides, distance measures described in [31] are defined in terms of a single attribute and it is not clear how to extend the similarity measure to incorporate information of more than one attribute (e.g., the three activities considered in this work). For that reasons we adopted a simple, yet very informative, measure for computing the distance between territory designs. The pseudocode of the static variant of PR is shown in Procedure 3. θ ∈ [0,1] is a scalar that is set empirically. 4.2.2 Dynamic GPR CTDP The dynamic PR variant differs from the static one in that in each iteration of the GRASP the solution XS is compared to a randomly selected solution from B, say B′. The intermediate solutions between XS and B′ are evaluated, and the best solution found in the path is denoted XR. Then if XR is better than B1 or if X R is better than Bb and it is at a distance of at most θ from the solutions in B, then the closest solution in B to XR is replaced with XR. Then solutions in B are sorted from best to worst. After imax iterations the best solution, namely B1, is returned. The pseudocode is shown in Procedure 4. A number of parameters are associated with GPR CTDP in both variants, namely δ the fraction of nodes assigned with a distance criterion, k the number of neighbors that are considered for building a territory, λ the tradeoff parameter of the objective function, α the GRASP quality 17 Procedure 4 grasp pr dynamic( imax ) Input: imax : number of global iterations; Output: Xbest; A p-partition of V ; for all ( i = {1, . . . ,b} ) do XS ← grasp construction(); Bi ← local search( X S ); end for Sort B in ascending order; for all ( iter = 1, . . . , imax ) do XS ← grasp construction(); XS ← local search( XS ); Randomly select B′ from B; Apply PR(XS,B′) and PR(B′,XS) and let XR ← best solution found; if ( (ψ(XR) < ψ(B1)) or (ψ(X R) < ψ(Bb) and d sol µ (X R,B) > θ) ) then Bj ← closest solution to X R in B with ψ(XR) < ψ(Bj); Bj ← X R; Update B; end if end for return Xbest ← B1; parameter for the RCL, limit evals the maximum number of evaluations for the local search, b the number of solutions in the elite set B and θ the distance threshold in PR. In this work we have fixed all of these parameters based on preliminary experimentation. The next section reports experimental results with the proposed GPR CTDP. 5 Computational experiments This section reports experimental results obtained with GPR CTDP. The proposed method was implemented in MatlabR. The code and data sets are publicly available for research purposes from the authors upon request. All of the experiments were run in a 64-bit workstation with a Corei7 processor at 3.4GHz and 8 GB in RAM. 5.1 Experimental setting For the experiments we used the data base from [28]. These are randomly generated instances based on real-world data. Data sets DS and DT are considered for experimentation. The former generate the BU weights from a uniform distribution and the latter uses a triangular distribution. Data set DT more closely resembles real-world instances. These data sets are fully described in [28]. For each of DS and DT data sets there are 20 different instances of size n = 500 and p = 10. 18 For all of the instances in both DS and DT data sets we use a tolerance level τa = 0.05, a ∈ A. Recall that τa measures the allowable relative deviation from the target average size µa for activity a. Hence, a value of τa = 0.05 implies that instances are tightly constrained in all activities and therefore the problem is more difficult to solve than instances that use a larger value of τa. In previous work [28], experiments have been reported with other values for τa ∈ [0.05,0.30]. Here we focus on the most difficult instances. Throughout the evaluation, the GRASP is run with imax = 500. Based on preliminary ex- perimentation for fine-tuning the algorithmic parameters for GPR CTDP, we will use the values reported in Table 1. Showing the fine-tuning of these parameters is out of the scope of this paper. Table 1: Summary of values used for the algorithmic parameters of GPR CTDP. Parameter Value Description δ 0.5 Fraction of nodes assigned with a distance criterion. k 3 Number of neighbors that are considered for growing a territory. λ 0.7 Weight parameter in the meritfunction. α 0.3 RCL quality parameter. limit evals 1,000 The maximum number of fitness function evaluations in the local search. b 20 The number of solutions in the elite set E. θ 0.6 The distance threshold in PR. imax 500 Number of global iterations for GPR CTDP. In the following sections we report the obtained experimental results. We have divided experi- mental results in three sections that aim at assessing different aspects of the GPR CTDP. 5.2 Assessing the construction and local search procedures within a GRASP framework This section describes results of experiments designed to evaluate the effect of the proposed con- struction and local search procedures. To this end we apply the new construction phase within a GRASP framework, that is, no PR phase is applied in this experiment. First, we apply the GRASP with construction phase only and then we apply the complete GRASP with both construction and local search phases. For each of these, we tested the two different data sets. Figures 5 and 6 show the performance of the construction and local search procedures for DT and DS data sets, respectively. In each figure we plot the values of the objective, F(S), and infeasibility, G(S), for each instance and for each mechanism. As expected, from these figures we can see that local search 19 improves significantly the construction procedure, in terms of both infeasibility and dispersion. For both data sets, local search (triangle marker) obtains feasible solutions (i.e., G(S) = 0) for most of the instances starting from the highly infeasible solutions generated by the construction mechanism (diamond marker). Besides, there are considerable improvements in terms of F(S) for all of the instances in the DT data set, see Figure 5. Lower improvements in terms of dispersion are observed for the DS data set, see Figure 6; although local search always obtained returned better solutions. It is expected that the PR mechanisms further improve the dispersion of solutions obtained with plain local search. 0 2 4 6 8 10 12 14 16 18 20 200 250 300 350 400 O b je ct iv e f u n ct io n ( d is p e rs io n ) Instances 0 2 4 6 8 10 12 14 16 18 20 0 0.02 0.04 0.06 0.08 In fe a si b ili ty F(S) − Local search G(S) − Local search F(S) − Construction G(S) − Construction Figure 5: Performance of the construction and local search mechanisms for instances in the DT data set. We show the values of F(S) (left y−axis) and G(S) (right y−axis). Table 2 summarizes the performance of the construction and local search procedures across all instances of both DT and DS data sets. For the dispersion term F(S), we show the relative deviation between the solution obtained with each procedure and the best known solution for each instance RDB = (F(S)−F(Sbest)/F(Sbest). The column labeled “local search” indicates that both construction and local search phases are applied. From this table we can see that the average of the sum of relative infeasibilities is maintained low in the construction procedure for both data sets. This result shows that the proposed procedure is able to obtain acceptable solutions in terms of the degree of satisfaction of the balance constraints despite the fact part of the construction procedure is based on a purely distance-based criterion. 20 0 2 4 6 8 10 12 14 16 18 20 40 60 O b je ct iv e f u n ct io n ( d is p e rs io n ) Instances 0 2 4 6 8 10 12 14 16 18 20 0 0.5 In fe a si b ili ty G(S) − Construction G(S) − Local search F(S) − Construction F(S) − Local search Figure 6: Performance of the construction and local search mechanisms for instances in the DS data set. We show the values of F(S) (left y−axis) and G(S) (right y−axis). Table 2: Evaluation of the construction and local search procedures of GPR CTDP. Data set DT DS Measure / Mechanism Construction Local search Construction Local search RDB Best 5.81% 0.00% 0.00% 0.00% Average 34.12% 1.51% 20.91% 14.51% Worst 81.45% 6.04% 58.28 56.76% G(S) Best 0.00E + 00 0.00E + 00 2.60E −01 0.00E + 00 Average 2.37E −02 0.00E + 00 3.61E −01 3.01E −04 Worst 7.06E −02 0.00E + 00 5.24E −01 3.55E −03 After applying local search to the constructed solutions, the dispersion measure F(S) is im- proved as it shows a reduction in the relative deviation with respect to the best dispersion value. In the case of the DT data set the solutions obtained with local search are very close to the best ones in terms of dispersion (average deviation of 1.51%), while for DU there is much more room for improvement (average deviation of 14.51%). For the DT data set the objective function is improved in average by 32.61%, while for the DS data set the improvement is of 6.4%. These are rather important differences that evidence the effectiveness of the proposed local search mechanism. It is very important to emphasize that dispersion is improved by also considerably reducing G(S). 5.3 GRASP vs. GPR CTDP This section reports experimental results on the improvements of the PR strategies over the straight GRASP implementation described in Section 4.1. Table 3 shows the performance of GPR CTDP 21 under both static (GPR-ST column) and dynamic (GPR-DY column) PR strategies for DT and DS data sets. In the table, we compare the performance of GPR CTDP when using PR and when only GRASP without PR is adopted. We show the relative deviation between the best solution obtained with each method and the best known solution for each instance. Table 3: Evaluation of GPR CTDP with static and dynamic PR. Data set DT DS Measure GRASP GPR-ST GPR-DY GRASP GPR-ST GPR-DY RDB Best 0% 0% 0% 0% 0% 0% Average 1.51% 0.51% 1.27% 14.51% 0.76% 13.92% Worst 6.04% 3.09% 3.91% 56.76% 11.44% 56.76% G(S) Best 0.00E + 00 0.00E + 00 0.00E + 00 0.00E + 00 0.00E + 00 0.00E + 00 Average 0.00E + 00 0.00E + 00 0.00E + 00 3.01E −04 2.53E −04 2.84E −04 Worst 0.00E + 00 0.00E + 00 0.00E + 00 3.55E −03 5.07E −04 3.55E −03 As we can see, for the DT data set, the improvements obtained with PR over local search are small yet non-negligible. We believe this result can be due to the fact that we are approaching to the global optimum for this data set and since the local search procedure provides very competitive solutions by itself the improvements due to PR are rather small. However, it is important to emphasize that all of the solutions found with GRASP and GPR CTDP are feasible for this data set. For this data set the static PR strategy outperformed the dynamic one by less than 1% in terms of the objective function. For the DS data set the improvements due to PR are larger. GPR CTDP with static PR outperforms the results of local search by an average of ≈ 13% in terms of the objective function, whereas the dynamic strategy outperforms local search by less than 1%. The static variant of PR achieve important improvements in terms of the dispersion objective (F(S)), while also reducing the infeasibility term. Finally, it is important to point out that even in the case when GRASP is allowed to run by itself for an amount of time equal to the total amount of time employed by GPR CTDP, the results reported by the later are still better. This is due to the fact that the GRASP seems to converge within the first iterations, thus a better solution is hardly found by GRASP afterwards. Figures 7 and 8 show the territories obtained with the construction, local search, and PR GPR CTDP procedures for a particular instance of the DT data set. Figure 7 shows the solution from a run of the static PR GPR CTDP and Figure 7 shows the corresponding solution for the 22 0 100 200 300 400 500 0 100 200 300 400 500 X − coordinates Y − c o o rd in a te s F(S) = 259.2441 / G(S) = 0.037334 0 100 200 300 400 500 0 100 200 300 400 500 X − coordinates Y − c o o rd in a te s F(S) = 222.8781 / G(S) = 0 0 100 200 300 400 500 0 100 200 300 400 500 X − coordinates Y − c o o rd in a te s F(S) = 216.1026 / G(S) = 0 Figure 7: Solutions obtained by the construction, local search and static PR procedures for a particular instance of the DT data set. 0 100 200 300 400 500 0 100 200 300 400 500 X − coordinates Y − c o o rd in a te s F(S) = 244.6376 / G(S) = 0.020266 0 100 200 300 400 500 0 100 200 300 400 500 X − coordinates Y − c o o rd in a te s F(S) = 243.6608 / G(S) = 0 0 100 200 300 400 500 0 100 200 300 400 500 X − coordinates Y − c o o rd in a te s F(S) = 222.0514 / G(S) = 0 Figure 8: Solutions obtained by the construction, local search and dynamic PR procedures for a particular instance of the DT data set. dynamic PR GPR CTDP. These figures illustrate the advantages of GPR CTDP over the construc- tion and local search mechanisms. Territories generated after the construction procedure present infeasibilities. The local search process eliminates infeasibilities and reduces the dispersion objec- tive. However, the dispersion is further minimized with both PR variants. Visually, it can be seen that territories generated with local search (center plots) are more disperse than those generated with GPR CTDP (rightmost plots). For this particular instance, a better solution was obtained with the static version of PR, which agrees with results presented in this section. 5.4 Static vs. dynamic path relinking This section elaborates on the difference in performance between the static and dynamic PR variants of GPR CTDP. From Table 3 we can see that the improvements of static and dynamic GPR CTDP over local search are of 1% and 0.24% for the DT data set and of 13.75% and 0.59% for the DS data set (in terms of the objective function). Thus, despite the fact both strategies resulted effective, the use of the static one is advantageous. We think this can be due to the fact that static GPR CTDP 23 explores all of the paths between elite solutions at the end of the search process. Hence a global picture of the search process is considered during the execution of static GPR CTDP. Dynamic GPR CTDP on the other hand, explores the paths between every solution processed by local search and a random solution from the elite set. Since it is not guaranteed that PR is performed over two competitive solutions, it is less likely that an effective solution can be found after exploring the paths. Figure 9 shows the relative deviation of the solutions found with each tested method and the best known solution for each instance for DT data set. This figure give us more insight into the performance of the different methods across the instances, it is rather clear that the static PR strategy obtained the best solutions for most of the instances (those instances for which the relative deviation is zero), followed by the dynamic PR approach. 0 5 10 15 20 0 1 2 3 4 5 6 Instances R e la tiv e d e vi a tio n f ro m t h e b e st s o lu tio n f o u n d ( % ) Local search GPR CTDP−ST GPR CTDP−DY Figure 9: A comparison among the methods in terms of relative deviation from best objective function value for the DT data set on each individual instance. Table 4 reports the processing time for each variant of GPR CTDP and for each data set. In general terms a new territory design can be obtained with either variant of GPR CTDP in a few hours. This processing times are acceptable, since from the practical standpoint, this decision is taken every 3-4 months. One final comment, it was observed that, in the GPR CTDP method, around 80% of the time is spend in the GRASP and 20% doing the Path Relinking. Therefore, we have empirically observed 24 Table 4: CPU time (min) comparison for static and dynamic GPR CTDP. DT DS GPR-ST GPR-DY GPR-ST GPR-DY Best 124.84 119.93 179.74 179.37 Average 136.25 133.70 204.42 200.24 Worst 152.96 150.34 240.30 227.46 that this additional amount of effort pays off significantly. 6 Conclusions We have introduced a new model in commercial territory design. The new model makes use of a diameter-based dispersion function instead of the traditional center-based functions. We have described a GRASP with path relinking (GPR CTDP) for this CTDP. The problem, motivated by a real-world application, consists of grouping commercial units into geographic terri- tories subject to dispersion, connectivity and balance constraints. A novel construction procedure was developed and two variants of PR were explored in GPR CTDP, namely, static and dynamic PR. The components of GPR CTDP were evaluated and compared extensively in instances that are known to be very challenging from previous work. Experimental results show that the proposed construction procedure is able to construct very competitive solutions, mainly in terms of the dispersion criterion. The local search of the GPR CTDP improves solutions in terms of both dispersion and balance requirements. Both versions of PR improve the performance of the application of the construction and local search mechanisms, con- firming previous work on the combination of GRASP and PR. In particular we found that with the static PR variant better solutions can be obtained for the TDP. This can be due to the fact that the PR process is applied over elite instances, which increases the chances of finding a better solution. In general terms the processing time of both PR variants lies in reasonable ranges for the application. We have identified several future work directions in the context of GPR CTDP. In particular we would like to explore other variants of PR that are known to be very effective, for example, evolutionary PR. Further, we are interested in the development of an adaptive filtering step that 25 allows us to identify pairs of solutions that can be potentially improved by applying PR. This is in addition to the rules used for updating the set of elite solutions. We think that such a filtering strategy will have a very positive impact in the efficiency of GPR CTDP. Since we found evidence that maintaining a set of elite solutions can be beneficial for TDP, we would like to explore the use of other “population-based” metaheuristics such as scatter search. It is important to note that the method developed in this work can also be extended and applied to other districting problems under balancing and connectivity constraints. The presence of the connectivity constraints make the path relinking process more challenging. For instance, path relinking has been applied in a different manner in related partitioning problems such as capacitated clustering [9]. In this particular work, we have successfully exploited the problem structure by solving an associated Assignment Problem whose solution will guide the relinking process in a more intelligent fashion. To the best of our knowledge this PR idea is novel and worthwhile for further exploration in other districting or clustering problems under connectivity constraints. Acknowledgements: The first author was supported by PROMEP under grant 103.5/11/4330. The second author was supported by the Mexican National Council for Science and Technology (CONACYT) under grants CB-2005-01/48499–Y and CB-2011-01/166397, abnd by Universidad Autónoma de Nuevo León under its Scientific and Technological Research Support through grants UANL-PAICYT CE012–09, IT511-10, and CE728-11. References [1] P. K. Bergey, C. T. Ragsdale, and M. Hoskote. A simulated annealing genetic algorithm for the electrical power districting problem. Annals of Operations Research, 121(1):33–55, 2003. [2] M. Blais, S. D. Lapierre, and G. Laporte. Solving a home-care districting problem in an urban setting. Journal of the Operational Research Society, 54(11):1141–1147, 2003. [3] B. Bozkaya, E. Erkut, and G. Laporte. A tabu search heuristic and adaptive memory procedure for political districting. European Journal of Operational Research, 144(1):12–26, 2003. 26 [4] M. H. Browdy. Simulated annealing: An improved computer model for political redistricting. Yale Law and Policy Review, 8:163–179, 1990. [5] R. E. Burkard, M. Dell’Amico, and S. Martello. Assignment Problems. SIAM, Philadelphia, 2009. [6] F. Caro, T. Shirabe, M. Guignard, and A. Weintraub. School redistricting: Embedding GIS tools with integer programming. Journal of the Operational Research Society, 55(8):836–849, 2004. [7] S. J. D’Amico, S.-J. Wang, R. Batta, and C. M. Rump. A simulated annealing approach to police district design. Computers & Operations Research, 29(6):667–684, 2002. [8] L. S. de Assis, P. M. Franca, and F. L. Usberti. A redistricting problem applied to meter reading in power distribution networks. Computers & Operations Research, 41:65–75, 2014. [9] Y. Deng and J. F. Bard. A reactive GRASP with path relinking for capacitated clustering. Journal of Heuristics, 17(2):119–152, 2011. [10] A. Drexl and K. Haase. Fast approximation methods for sales force deployment. Management Science, 45(10):1307–1323, 1999. [11] J. C. Duque, R. Ramos, and J. Suriñach. Supervised regionalization methods: A survey. International Regional Science Review, 30(3):195–220, 2007. [12] E. Erkut, Y. Ülküsal, and O. Yeniçerioğlu. A comparison of p-dispersion heuristics. Computers & Operations Research, 21(10):1103–1113, 1994. [13] H. J. Faria, S. Binato, M. G. C. Resende, and D. J. Falcao. Transmission network design by a greedy randomized adaptive path relinking approach. IEEE Transactions on Power Systems, 20(1):43–49, 2005. [14] T. A. Feo and M. G. C. Resende. Greedy randomized adaptive search procedures. Journal of Global Optimization, 6(2):109–133, 1995. 27 [15] S. L. Forman and Y. Yue. Congressional districting using a TSP based genetic algorithm. In E. Cantú-Paz, J. A. Foster, K. Deb, L. David, and R. Rajkumar, editors, Genetic and Evo- lutionary Computation – GECCO 2003, volume 2723 of Lecture Notes in Computer Science, pages 2072–2083. Springer, Berlin, 2003. [16] F. Glover. Tabu search and adaptive memory programming – Advances, applications, and challenges. In R. S. Barr, R. V. Helgason, and J. L Kennington, editors, Interfaces in Computer Science and Operations Research, chapter 1, pages 1–75. Kluwer, Dordrecht, 1996. [17] J. Kalcsics, S. Nickel, and M. Schröder. Toward a unified territorial design approach: Appli- cations, algorithms, and GIS integration. TOP, 13(1):1–56, 2005. [18] A. Mehrotra, E. L. Johnson, and G. L. Nemhauser. An optimization based heuristic for political districting. Management Science, 44(8):1100–1114, 1998. [19] L. Muyldermans, D. Cattryse, D. Van Oudheusden, and T. Lotan. Districting for salt spreading operations. European Journal of Operational Research, 139(3):521–532, 2002. [20] F. Pukelsheim, F. Ricca, B. Simeone, A. Scozzari, and P. Serafini. Network flow methods for electoral systems. Networks, 59(1):73–88, 2012. [21] M. G. C. Resende, R. Mart́ı, M. Gallego, and A. Duarte. GRASP and path relinking for the max-min diversity problem. Computers & Operations Research, 37(3):498–508, 2010. [22] M. G. C. Resende and C. C. Ribeiro. Greedy randomized adaptive search procedures: Ad- vances, hybridizations, and applications. In M. Gendreau and J.-Y. Potvin, editors, Handbook of Metaheuristics, volume 146 of International Series in Operations Research & Management Science, chapter 10, pages 283–319. Springer, Berlin, 2010. [23] M. G. C. Resende, C. C. Ribeiro, F. Glover, and R. Mart́ı. Scatter search and path-relinking: Fundamentals, advances, and applications. In M. Gendreau and J.-Y. Potvin, editors, Hand- book of Metaheuristics, volume 146 of International Series in Operations Research & Manage- ment Science, chapter 4, pages 87–107. Springer, Berlin, 2010. 28 [24] M. G. C. Resende and R. F. Werneck. A hybrid heuristic for the p-median problem. Journal of Heuristics, 10(1):59–88, 2004. [25] C. C. Ribeiro, E. Uchoa, and R. F. Werneck. A hybrid GRASP with perturbations for the Steiner problem in graphs. INFORMS Journal on Computing, 14(3):228–246, 2002. [26] F. Ricca, A. Scozzari, and B. Simeone. Political districting: From classical models to recent approaches. Annals of Operations Research, 204(1):271–299, 2013. [27] F. Ricca and B. Simeone. Local search algorithms for political districting. European Journal of Operational Research, 189(3):1409–1426, 2008. [28] R. Z. Ŕıos-Mercado and E. A. Fernández. A reactive GRASP for a commercial territory design problem with multiple balancing requirements. Computers & Operations Research, 36(3):755– 776, 2009. [29] R. Z. Ŕıos-Mercado and J. C. Salazar-Acosta. A GRASP with strategic oscillation for a commercial territory design problem with a routing budget constraint. In I. Batyrshin and G. Sidorov, editors, Advances in Soft Computing, volume 7095 of Lecture Notes in Artificial Intelligence, pages 307–318. Springer, Heidelberg, Germany, 2011. [30] M. A. Salazar-Aguilar, R. Z. Ŕıos-Mercado, and M. Cabrera-Ŕıos. New models for commercial territory design. Networks & Spatial Economics, 11(3):487–507, 2011. [31] F. Tavares Pereira, J. Rui Figueira, V. Mousseau, and B. Roy. Comparing two territory partitions in districting problems: Indices and practical issues. Socio-Economic Planning Sciences, 43(1):72–88, 2009. [32] A. A. Zoltners and P. Sinha. Sales territory alignment: A review and model. Management Science, 29(11):1237–1256, 1983. [33] A. A. Zoltners and P. Sinha. Sales territory design: Thirty years of modeling and implemen- tation. Marketing Science, 24(3):313–331, 2005. 29 
work_l6q4o7kqwnh3jmqca4gmv4sqxm	----	Intelligent tourism recommender systems: A survey Expert Systems with Applications 41 (2014) 7370–7389 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Review Intelligent tourism recommender systems: A survey http://dx.doi.org/10.1016/j.eswa.2014.06.007 0957-4174/� 2014 Elsevier Ltd. All rights reserved. ⇑ Corresponding author at: Science & Technology Park for Tourism and Leisure, C/ Joanot Martorell, 15, 43480 Vila-Seca, Catalonia, Spain. Tel.: +34 977394754. E-mail addresses: joan.borras@pct-turisme.cat (J. Borràs), antonio.moreno@urv. cat (A. Moreno), aida.valls@urv.cat (A. Valls). Joan Borràs a,b,⇑, Antonio Moreno b, Aida Valls b a Science & Technology Park for Tourism and Leisure, C/Joanot Martorell, 15, 43480 Vila-Seca, Catalonia, Spain b Intelligent Technologies for Advanced Knowledge Acquisition (ITAKA), Departament d’Enginyeria Informàtica i Matemàtiques, Universitat Rovira i Virgili, Av. Països Catalans, 26, 43007 Tarragona, Catalonia, Spain a r t i c l e i n f o a b s t r a c t Article history: Available online 9 June 2014 Keywords: Recommender systems Tourism Ontologies Planning Clustering Recommender systems are currently being applied in many different domains. This paper focuses on their application in tourism. A comprehensive and thorough search of the smart e-Tourism recommend- ers reported in the Artificial Intelligence journals and conferences since 2008 has been made. The paper provides a detailed and up-to-date survey of the field, considering the different kinds of interfaces, the diversity of recommendation algorithms, the functionalities offered by these systems and their use of Artificial Intelligence techniques. The survey also provides some guidelines for the construction of tour- ism recommenders and outlines the most promising areas of work in the field for the next years. � 2014 Elsevier Ltd. All rights reserved. 1. Introduction The amount of information available in the World Wide Web and its number of users have experienced an enormous increase in the last decade. All this information may be particularly useful for those users who plan to visit an unknown destination. Informa- tion about travel destinations and their associated resources, such as accommodations, restaurants, museums or events, among oth- ers, is commonly searched for tourists in order to plan a trip. How- ever, the list of possibilities offered by Web search engines (or even specialised tourism sites) may be overwhelming. The evaluation of this long list of options is very complex and time consuming for tourists in order to select the one that fits better with their needs. Personalization techniques (Gao, Liu, & Wu, 2010) aim to provide customised information to users based on their preferences, restrictions or tastes. They are particularly relevant in recom- mender systems (Adomavicius & Tuzhilin, 2005), whose objective is to filter irrelevant options and to provide personalised and rele- vant information to each particular user. In the tourism field, travel recommender systems (Ricci, 2002) aim to match the characteris- tics of tourism and leisure resources or attractions with the user needs. These systems are especially useful if they can automati- cally learn the user’s preferences through the analysis of her expli- cit or implicit feedback (Sieg, Mobasher, & Burke, 2007). Explicit data may be given by the user in different ways, for instance whenever she specifies her cultural interests by filling in a form. Implicit interests can be inferred by the system through the analysis of the behaviour of the user. In many cases recommender systems not only take into account the preferences of the tourist but they also analyze the dynamic context (Dey & Abowd, 1999) of the trip. This is especially useful when tourists are already at the destination and they are willing to use their mobile devices to customise their trips in real time. The context can include aspects like the tourist’s location, the time of the visit or the current weather (Lamsfus, Alzua-Sorzabal, Martin, Salvador, & Usandizaga, 2009). Approaches that take con- text into account can send suggestions proactively, depending on the current state of the tourist. For example, a museum that was planned to be visited today may be too far from the visitor and she may not have enough time to reach it, so the plan scheduled for tomorrow may be changed to include the visit to the museum. The last ten years have witnessed an explosive increase in the use of mobile technology among tourists. Therefore, e-Tourism systems (Buhalis, 2003) provide a good opportunity for mobile ser- vices that help visitors by offering recommendations based on their preferences and their current context. This paper focuses on tourism recommender systems that employ Artificial Intelligence (AI) techniques at some point. Some examples can be the following: � Intelligent autonomous agents may analyse the behaviour of a user, learn automatically the user profile and provide proactive recommendations depending on the current context (Batet, Moreno, Sánchez, Isern, & Valls, 2012). http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.06.007&domain=pdf http://dx.doi.org/10.1016/j.eswa.2014.06.007 mailto:joan.borras@pct-turisme.cat mailto:antonio.moreno@urv.cat mailto:antonio.moreno@urv.cat mailto:aida.valls@urv.cat http://dx.doi.org/10.1016/j.eswa.2014.06.007 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. Interfaces used in the reviewed works (in%). J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7371 � Some systems go beyond offering a list of recommended tourist attractions and use automated planners to schedule these recommendations within a route that can span several days (Vansteenwegen, Souffriau, Vanden Berghe, & Van Oudheusden, 2010). � Other approaches take into account the opening and closing times of the attractions, or the time needed to go from one point of interest to another, hence offering a detailed timetable of the visit. However, this is a very complex planning and scheduling problem for which it is difficult to guarantee an optimal solu- tion. Some systems solve this complexity with the use of AI opti- mization techniques, such as ant colony or meta-heuristic iterative methods (Lee, Chang, & Wang, 2009). � Automatic clustering algorithms may be applied to classify tour- ists with similar tastes or similar features (Gavalas & Kenteris, 2011). � Approximate reasoning methodologies, like fuzzy logic or Bayesian networks, may be used to manage the imprecision related to the inference of the preferences of the user (Huang & Bian, 2009). � Reasoning procedures, like rule-based systems, are also employed to deduce the user’s preferences (Lamsfus, Alzua-Sorzabal, Martin, & Smithers, 2011). � Formalisms developed in the knowledge representation field of AI, like ontologies, are commonly used to represent (and reason about) the tourism domain knowledge (Moreno, Valls, Isern, Marin, & Borràs, 2013). The contributions of this article are twofold. On the one hand, it provides a comprehensive review of the tourism recommender systems published in scientific journals and conferences since 2008, with an especial focus on the ones that employ AI tech- niques. Commercial products are not considered in this review, as it is usually not possible to know how they have been designed and implemented. Several aspects of these systems are analysed, such as their interface, their functionalities, the recommendation mechanisms and the AI methods and techniques employed. This review provides an up-to-date state of the art of the field of intel- ligent tourism recommenders, which may be useful not only to the scientists working in this field but to designers and developers of intelligent recommender systems in other domains. On the other hand, the paper also provides guidelines to be followed in the design of this kind of intelligent systems and an outline of some of the most promising research lines that may be pursued in the near future. The reminder of this review is structured as follows. In the next section we analyze which interfaces are commonly used by tour- ism recommender systems to interact with users, discussing espe- cially the differences between mobile and Web-based approaches. After that we survey the main functionalities offered by these sys- tems, ranging from the recommendation of a tourist destination to Table 1 Review of user interfaces. Interface References Web + mobile Venkataiaha et al. (2008),(Lamsfus et al. (2009), Niaraki and Kim (20 (2011), Borràs et al. (2012a), Ruotsalo et al. (2013) and Umanets et a Only web Coelho, Martins, and Almeida (2009), Huang and Bian (2009), Lucas, L Montes (2009), Jannach et al. (2010), Mínguez et al. (2010), Sebastià Colomo-Palacios, González-Carrasco, and Ruiz-Mezcua (2011), Linaza et al. (2011), Montejo-Ráez et al. (2011), Sebastià et al. (2009), Garci (2013), Kurata and Hara, (2013), Lucas et al. (2013) and Savir et al. ( Only mobile Castillo et al. (2008), Ceccaroni et al. (2009), García-Crespo et al. (2009 Martínez-Santiago et al. (2012), Noguera et al. (2012), Braunhofer et Yang and Hwang (2013) the automatic construction of a detailed complex schedule of a visit of several days to a certain area. Section 4 comments the rec- ommendation methods employed by e-Tourism recommenders, focusing on content-based and collaborative approaches. The next section exposes the use of AI techniques from different fields like multi-agent systems, approximate reasoning, knowledge represen- tation, etc. A comparison with previous surveys on tourism recom- menders is given in Section 6. The paper concludes with a global analysis of the surveyed systems and some suggestions of lines of future work in the field. 2. Interface This section analyses the user interfaces of recent tourism rec- ommender systems. Most of them offer a Web-based interface and/or an interface specifically designed to be used in mobile devices. Table 1 classifies the most relevant e-Tourism recom- menders in these two broad categories, and Fig. 1 shows the per- centage of surveyed systems in each of them. A Web-based interface is the option chosen by most of the systems, since it per- mits an easy access from any computer connected to the Web without any kind of downloading, installation and configuration. However, due to the enormous increase in the use of smart phones connected to the Web in the last years, more than half of the reviewed systems have specific interfaces for mobile devices. There are some recommender systems that have been designed as desktop applications and do not offer any of the two usual kinds of interfaces (e.g., Kurata, 2011). This kind of applications can usu- ally be implemented more quickly than the mobile or Web-based ones; however, they require downloading and installing the pro- gram, which is not comfortable to most of the tourists that want to get recommendations as simply as possible without being both- ered by technical details. 09), Vansteenwegen et al. (2010), Gavalas and Kenteris (2011), Rey-López et al. l. (2013) aurent, Moreno, and Teisseire (2009), Lee et al. (2009), Ruiz-Montiel and Aldana- et al. (2010), Yang (2010), Borràs et al. (2011), García-Crespo, López-Cuadrado, , Aguirregoikoa, García, Torres, and Aranburu (2011), Lorenzi et al. (2011), Luberg a et al. (2011), Wang et al. (2011), Koceski and Petrevska (2012), Gyorodi et al. 2013) ), Yu and Chang (2009), Ricci et al. (2010), Martin et al. (2011), Batet et al. (2012), al. (2013), Garcia et al. (2013), Meehan et al. (2013), Rojas and Uribe (2013) and Fig. 2. Personalized route through Tainan City (from Lee et al., 2009). 7372 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 The following subsections review some approaches based on Web or mobile interfaces. 2.1. Web-based recommenders The use of a Web-based interface is the most common option adopted by e-Tourism recommenders. This kind of interfaces allows tourists to look for information in a user-friendly manner. Users normally have a rich interaction with the system using a wide screen which allows displaying a large amount of data extended with maps, images or even high quality videos. More- over, the mouse permits to interact easily with the computer and move through maps, perform zoom actions, select items or even drag and drop them. This is very useful for tourists when they are still in the planning stage of their trips. Nevertheless, Web- based applications are usually not designed to be used during the stay since most of the tourists will not have easy access to comput- ers with Internet connection. Although an increasing number of tourists have mobile handsets or tablets with Internet connection, the information-ridden Web pages usually shown by recommend- ers cannot be easily read or manipulated on such small screens. In the remainder of this section we comment some interesting fea- tures exploited in Web-based interfaces to improve the interaction with the users. Venkataiaha, Shardaa, and Ponnadaa (2008) report the design of two visualisation systems (called discrete and continuous) for a tourism recommender and compare the interaction of the users in both cases. The former provides a high quantity of information in the screen at the same time, and it was determined that users needed too much time and effort to understand it. The latter aggre- gates all the information into a single video clip that combines the most relevant media content, including text, photographs and videos. The approach shown in Lee et al. (2009) is one of the firsts that embeds Google Maps Services1 in their Web pages (Fig. 2) in order to plot the travel route on a map, so that tourists can follow the per- 1 https://developers.google.com/maps/ (last access March 2014). sonalised itinerary to enjoy cultural heritage and local gourmet food during their stay at Tainan City. Other Web-based recommender systems that display in a map the places scheduled to be visited in a single day are e-Tourism (Sebastià, Garcia, Onaindia, & Guzman, 2009), City Trip Planner (Vansteenwegen et al., 2010), Otium (Montejo-Ráez, Perea- Ortega, García-Cumbreras, & Martínez-Santiago, 2011) and EnoSig- Tur (Borràs et al., 2012a). In this last system the user introduces her socio-demographic information and general preferences (left image of Fig. 3), and then she will receive recommendations of attractions spread out over the province of Tarragona, a wide region of Spain (right image of Fig. 3). The last Web page shows a high quantity of information, such as geo-localised attractions classified by categories, a route indicating driving times between the suggested activities, their approximate visiting times, etc. The VIBE virtual spa advisor (Jannach, Zanker, & Jessenitschnig, 2010) keeps an avatar-based conversation with the tourist in order to acquire the user’s visit requirements through personalised forms. The main point of this approach is its dynamicity. If a new attribute has to be added to the product catalogue, it is automati- cally taken into account not only in the recommendation and pref- erence elicitation processes, but also in the Web interface which is changed accordingly. The Web site has a section for domain experts, in which they can add or modify logical conditions that govern the conversational and recommendation procedures. Wang, Zeng, and Tang (2011) show how Semantic Web technol- ogies may be integrated with Web 2.0 services to leverage each other’s strengths. To do so they proposed an ontology-based tour- ism recommender that allows the automatic and dynamic integra- tion of heterogeneous on-line travel information. The platform is built in Ruby on Rails with view extensions to create rich Ajax Web-based applications. They also use third party services to pro- vide additional features, such as Google Map, Yahoo Weather, and WikiTravel. 2.2. Mobile recommendations Systems that offer mobile interfaces have increased consider- ably in the last few years, due to the large number of users acquir- https://developers.google.com/maps/ Fig. 3. Web version of EnoSigTur: introduction of initial user profile and route planner (from Borràs et al., 2012a). J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7373 ing mobile devices with Internet connection or, more recently, the well-known smartphones. Mobile devices are small and their Internet connection is usually slow; thus, the quantity of informa- tion that can be shown in these devices cannot be compared with a standard Web page. Therefore, mobile tourism recommender sys- tems have to make an effort to provide only the information that is essential for the user, and it must be well structured in order to be displayed correctly in small screens. Moreover, the user’s interaction with the system is limited, since even the basic actions made in Web-based interfaces (scrolling, introducing text) are not that easy. However, it is fair to say that the latest smartphones with bigger touchscreens provide a better user interaction. Fur- thermore, the main advantage of mobile devices is that they allow the use of the system in any place with an Internet connection, so that tourists may access information, discover places or modify their trips during the stay. Besides, most mobile systems are equipped with GPS and the recommender may know the present location of the user and it may offer geo-referenced information, advice or recommendations based on this knowledge. One of the first approaches in the field that used mobile systems was reported in Yu and Chang (2009). This system, designed for PDAs, offers location-based recommendation services to support personalised tour planning. Recommendations are based on Fig. 4. Prototype system for Windows mo tourists’ preferences, location and time. Fig. 4 shows the mobile user interface in four separated screenshots. The first one shows the different mobile tourism services (restaurant, hotel, sightsee- ing spot, user profile, and tour plan recommendation). The second image illustrates the interface for setting user preferences. The third screenshot shows the recommended tour plan with informa- tion about the places to visit, such as names, descriptions, photos or visiting time frames. Finally, the last image displays the tour plan on Google Maps. Another approach compatible with PDAs is MTRS (Gavalas & Kenteris, 2011). The authors argue that tourists may have prob- lems to connect with the Internet, either because they are in a rural area or because they are foreigners and cannot afford the roaming costs abroad. They propose to solve this problem by installing an infrastructure to support proximity detection and a cost-effective means for remote content update. In fact, they propose to use small- to medium-scale wireless sensor networks. Through this infrastructure, they introduce the concept of ‘con- text-aware rating’, in which user ratings uploaded through fixed Internet connection infrastructures (located at the rated places) are weighted higher to differentiate them from users that pro- vide an evaluation using the Internet away from the visited place. bile devices (from Yu & Chang, 2009). 7374 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 Another product using mobile devices is MapMobyRek (Ricci, Nguyen, & Averjanova, 2010) that exploits quite well its interface by showing recommendations in lists and on maps. This system permits to compare two items with their characteristics displayed side-by-side in order to decide the one that is preferred. GeOasis (Martínez-Santiago, Ariza-López, Montejo-Ráez, & Ureña-López, 2012) acts as a tourist guide that describes the places to visit while the tourist approaches the recommended locations. The system uses the mobile GPS device to know the tourist location and speed in order to estimate the available time to give the expla- nations. Users can interact with the system in two ways: using a tactile interface or using a voice-based interface (voice recognition and text-to-speech software). Despite the existence of several mobile tourism recommenders, not many of them use the newest technologies in mobile devices, such as the Android or iPhone platforms. Some examples that use these popular and rising platforms are reviewed below. Fig. 6. Preference elicitation from the mobile interf Fig. 5. Mobile version of EnoSigTu The moreTourism (Rey-López, Barragáns-Martínez, Peleteiro, Mikic-Fonte, & Burguillo, 2011) Android-based platform provides information about tourist resources through the use of mashups, integrating images, videos, augmented reality services, geo-loca- tion, guide services, access to urban networks, etc. Another approach that uses Android platforms is EnoSigTur (Borràs et al., 2012a). Fig. 5 shows some screenshots of the mobile version of EnoSigTur: a list of recommended places, the route of the trip and detailed information of an attraction. LiveCities (Martin, Alzua, & Lamsfus, 2011) uses the notification service of Android systems to provide push information according to the user context. This information can be plain text, audio, video or HTML. The STS system (Braunhofer, Elahi, Ricci, & Schievenin 2013) is a powerful Android application with a good design inter- face that permits the user to enter accurate information about her interests and opinions on the trip and the visited attractions (see Fig. 6). ace of the STS system (Braunhofer et al. 2013). r (from Borràs et al., 2012a). Fig. 7. iPad version of the GUIDEME system (Umanets et al. 2013). Fig. 8. Functionalities applied in the reviewed approaches (in%). J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7375 The recent GUIDEME system (Umanets, Ferreira, & Leite 2013) features a good implementation for mobile devices since its designers have not only developed an app for phones but also for tablet devices. In particular, the app is built for the iOS platform and it is adaptive to the screen sizes with specific adjustments for both iPhone and iPad devices. Fig. 7 shows screenshots of its iPad version. REJA (Noguera, Barranco, Segura, & Martínez, 2012) also works for iOS platforms. 3. Functionalities In this section we describe the general functionalities provided by the reviewed tourism recommender systems. Table 2 catalogues the approaches in four broad groups, depending on the services they offer: suggestion of a destination and construction of a whole tourist pack, recommendation of suitable attractions in one specific destination, design of a detailed multi-day trip schedule, and social capabilities. Fig. 8 gives a visual estimation of the percentage of systems that offer each of them. These aspects are commented in more detail in the following subsections, with examples of the most prominent proposals. Table 2 Review of user functionalities. Functionalities References Destination/tourist packs Seidel et al. (2009), Yu and Chang (2009), Lorenzi et al. (2011 Suggest attractions Castillo et al. (2008), Coelho et al. (2009), Ceccaroni et al. (2009 et al. (2009), Lee et al. (2009), Niaraki and Kim (2009), Ruiz-M Mínguez et al. (2010), Ricci et al. (2010), Sebastià et al. (2010), Gavalas and Kenteris (2011), Kurata (2011), Linaza et al. (201 et al. (2011), Rey-López et al. (2011), Sebastià et al. (2009), G Koceski and Petrevska (2012), Martínez-Santiago et al. (2012) 2013, Lucas et al. (2013), Meehan et al. (2013), Rojas and Uribe and Hwang (2013) Trip planner Castillo et al. (2008), Coelho et al. (2009), Ceccaroni et al. (2009 (2009), Niaraki and Kim (2009), Yu and Chang (2009), Míngue Linaza et al. (2011), Luberg et al. (2011), Montejo-Ráez et al. et al. (2011), Batet et al. (2012), Borràs et al. (2012a), Kurata Social aspects Coelho et al. (2009), Ceccaroni et al. (2009), García-Crespo et (2009), Garcia et al. (2011), Garcia et al. (2013), Meehan et al 3.1. Travel destination and tourist packs Some of the reviewed systems focus on the recommendation of a destination that suits the user’s preferences. This is the case of systems like PersonalTour (Lorenzi, Loh, & Abel, 2011), Itchy Feet ) and Koceski and Petrevska (2012) ), García-Crespo et al. (2009), Huang and Bian (2009), Lamsfus et al. (2009), Lucas ontiel and Aldana-Montes (2009), Yu and Chang (2009), Jannach et al. (2010), Vansteenwegen et al. (2010), Yang (2010), Borràs et al. (2011), Fenza et al. (2011), 1), Lorenzi et al. (2011), Luberg et al. (2011), Martin et al. (2011), Montejo-Ráez arcia et al. (2011), Wang et al. (2011), Batet et al. (2012), Borràs et al. (2012a), , Braunhofer et al. 2013, Garcia et al. (2013), Gyorodi et al. 2013, Kurata & Hara (2013), Ruotsalo et al. (2013), Savir et al. (2013), Umanets et al. (2013) and Yang ), García-Crespo et al. (2009), Huang and Bian (2009), Lucas et al. (2009), Lee et al. z et al. (2010), Sebastià et al. (2010), Vansteenwegen et al. (2010), Kurata (2011), (2011), Rey-López et al. (2011), Sebastià et al. (2009), Garcia et al. (2011), Wang & Hara 2013, Lucas et al. (2013) and Savir et al. (2013) al. (2009), Vansteenwegen et al. (2010), Rey-López et al. (2011), Sebastià et al. . (2013), Umanets et al. (2013) and Yang and Hwang (2013) 7376 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 (Seidel, Gärtner, Pöttler, Berger, & Dittenbach, 2009) and MyTravelPal (Koceski & Petrevska, 2012). PersonalTour is used for travel agencies to help their costumers to find the best travel pack- ages according to their preferences. Once the recommendation process is finished, a rated list of options is presented to the cos- tumer. Table 3 shows an example of the hotel recommendation service. After that, the customer can rate each item of each travel service. Itchy Feet not only recommends tourism destinations but also provides purchasing services for booking a trip and assistance from professional travel agents. Users make search requests, which are handled by autonomous agents that search for information in the internal database as well as in external data sources. The results are shown to the user through the interface, where recommended items (flights and hotels) can be selected and purchased. MyTravelPal (Koceski & Petrevska, 2012) first recommends areas of interest over a region graphically (see Fig. 9), where the size of the circle indicates the level of affinity with the user. Once the user focuses on a particular area, their tourist resources are also shown and sized depending on the affinity to the user profile. 3.2. Ranked list of suggested attractions Most Tourist recommender systems tend to suggest places once the user has decided the destination of the trip or she is already there. These systems are more complex, since they can suggest a large number of attractions, accommodations, restaurants or even temporal events. In this context the capability of recommenders to Fig. 9. MyTravelPal – recommend Table 3 Example of hotel recommendation in PersonalTour (adapted from (Lorenzi et al. (2011)).) Id Hotel name City Hotel category 1 Libertel Paris Economic 2 Palladium Punta Cana Resort 3 Amadeus Milan Economic 4 Riu Palace Cancun First 5 WestIn Aruba Economic classify and rank only those elements considered important for a particular user among the huge quantity of available information is very useful. With the support of these systems the users can find interesting places in an efficient way and even discover unex- pected ones that may be of their interest. The activities to be rec- ommended are normally stored in a static database, although some systems (e.g. Otium, Montejo-Ráez et al., 2011) extract auto- matically information about events from the Web to ensure that they always provide updated information. This kind of recommender systems (e.g. Borràs et al., 2011; Fenza, Fischetti, Furno, & Loia, 2011; Ruiz-Montiel & Aldana- Montes, 2009; Sebastià et al., 2009) usually provide a list of activ- ities that match the user profile, have been visited and/or posi- tively evaluated by similar users in the past, or are similar to activities previously enjoyed by the user. Thus, they include mech- anisms to compare the user preferences with the features of an object, or to compare the similarities between two users or two objects (more details on the recommendation techniques are given in Section 4). The selection of the recommended items may also take into account contextual factors, like the present location of the user (Noguera et al., 2012). Some systems are also capable of justifying the provided recommendations (e.g. Jannach et al., 2010). SMARTMUSEUM (Ruotsalo et al., 2013) is an example of a more complex recommendation system, which detects automatically if the user is outdoors or indoors, based on her location. For the first case, it can display the recommendations on a map. For indoor sce- narios, it gives a list of the most relevant objects according to the ation of regions of interest. . Room category Room type Swimming Pool WiFi Standard Double No Yes Luxe Double Yes Yes Standard Single No Yes Luxe Double Yes Yes Luxe Double Yes Yes Fig. 10. List of recommendations of specific objects for indoor scenarios. J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7377 user’s preferences. This is useful, for instance, in museums, where the number of objects to see may be relatively high (Fig. 10). 3.3. Planning a route There are several projects that not only provide a list of the places that fit better with the user’s preferences but also help tour- ists to create a route through several attractions. CT-Planner (Kurata, 2011; Kurata & Hara, 2013) offers tour plans, as shown in Fig. 11, that are refined gradually as the user’s expresses her preferences and requests (duration, walking speed, reluctance to walk, etc.). It displays a radar chart that represents the user’s preferences and a cartoon character as a navigator, in order to enrich the sense of user-friendliness and interactivity. There are several systems that provide an initial set of recom- mended activities (or an initial plan), with which the user can directly interact to add more activities, remove activities, select an activity to be visited, change the order of visit, etc. The planning component of the recommender system takes into account impor- tant factors like the expected duration of the visit, the opening and closing times of the attractions and the distance between them. Some relevant examples include EnoSigTur (Borràs et al., 2012a), City Trip Planner (Vansteenwegen et al., 2010), CRUZAR (Mínguez, Berrueta, & Polo, 2010), Smart City (Luberg, Tammet, & Järv, 2011), Otium (Montejo-Ráez et al., 2011) and e-Tourism (Sebastià et al., 2009). A more detailed review of trip planning functionalities is available in (Vansteenwegen & Souffriau, 2011). Some advanced recommenders, like SAMAP (Castillo et al., 2008) and PaTac (Ceccaroni, Codina, Palau, & Pous, 2009), are capable of analysing the connection possibilities between the activities using different means of transport (walking, by bike, by car, or by public transport). Some of these systems incorporate more complex Geographical Information Systems (GIS) to manage the geographical data associ- ated to the touristic points and events. (Huang & Bian, 2009) argued that it is computationally unfeasible to maintain large amounts of spatial data and use them in planning procedures. Hence, they used existing geospatial Web service technologies, in concrete the ESRI ArcWeb Service,2 to obtain the location of the attractions, the distance between them given their street address, and driving directions between two attractions. GeOasis (Martínez- Santiago et al., 2012) continually calculates the position and the speed of the user. The estimated time to reach a place is considered in order to create the plan in real time. The key aspect is the predic- tion of where the user will be in the immediate future: in a city, near a city or on the road. If the user is already in a city, the planning algo- rithm checks the nearest places to the user without taking into 2 http://www.esri.com/news/arcuser/0403/arcweb.html (last access March 2014). account the route or the speed, since it is considered that the user is close to them. If the user is near a city, the planning algorithm checks the most relevant attractions in it. If the user is on the road, but not near a city, then the planning algorithm is more complex because it considers temporal constraints. The plan is not computed by the server but by the client application, since it is constantly checking the location by GPS. Routes are computed using Google Maps as an external resource. Once the visit plan has been completely defined, the user may wish to retrieve the full schedule to follow the route. This retrieval can take different forms. Systems like SAMAP (Castillo et al., 2008) or EnoSigTur (Borràs et al., 2012a) allow downloading a PDF file that contains a geo-referenced map with a detailed explanation of the plan. In others, like City Trip Planner (Vansteenwegen et al., 2010) and Otium (Montejo-Ráez et al., 2011), the user can download the route to a mobile phone. 3.4. Social aspects Several projects (e.g. Ceccaroni et al., 2009; Garcia, Torre, & Linaza, 2013; Umanets et al., 2013; Vansteenwegen et al., 2010) have paid special attention to the inclusion of social functionalities that allow users to share material (pictures, comments, evalua- tions) and interact with other tourists. These aspects may be very interesting to help to promote the use of a recommender among the visitors of a particular city. Recommenders like moreTourism (Rey-López et al., 2011) and Itchy Feet (Seidel et al., 2009) allow users not only to interact over popular social networks but also to create location-based activity groups that can be employed to post comments, join groups for doing common activities or interact with other users. The system e-Tourism (Garcia, Sebastià, & Onaindia, 2011) allows to create plans that accommodate the pref- erences of a whole group of visitors. In iTravel (Yang & Hwang, 2013) users communicate among them with mobile peer-to-peer communications to send ratings of attractions. Their navigation map not only displays the location of attractions but also the position of near-by users with which it is possible to communicate. Fig. 12 shows a map with recommended attractions (green pins) and nearby users (blue pins). The VISIT system (Meehan, Lunney, Curran, & McCaughey, 2013) applies sentiment analysis techniques (using the Alchemy API3) to analyse the updates about a given attraction in Twitter and Facebook and identify if users are expressing positive or negative comments about it. This information is shown with green and red colours by the system in its interface, so that the user may easily identify those places visitors are loving most today and which not. 4. Recommendation techniques in e-Tourism Recommender systems have been classically classified, accord- ing to the way in which they analyze the information of the user and filter the list of items, into content-based, collaborative and demographic systems (Burke, 2002; Manouselis & Costopoulou, 2007; Montaner, López, & de la Rosa, 2003). In this section we introduce these three paradigms, analyzing its use in current tour- ism recommender systems. In addition, we present the different user models that have been employed in those systems. 4.1. Approaches to recommendation Content-based (CB) systems calculate a degree of similarity between the users and the items to be recommended. The process 3 AlchemyAPI (2014). Alchemy API: Transforming text into knowledge. [Online] http://www.alchemyapi.com/. http://www.esri.com/news/arcuser/0403/arcweb.html http://www.alchemyapi.com/ Fig. 11. CT-Planner2 user interface (from (Kurata, 2011)). Fig. 12. Navigation map of iTravel (from Yang & Hwang, 2013). 7378 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 is carried out by comparing the features of the item with respect to the user’s preferences. So, it is assumed that both users and alter- natives share a common representation (e.g., they are composed of the same set of attributes or keywords). The output of the compar- ison process is usually an overall performance score, which indi- cates the degree of matching between the user’s profile and each alternative. The higher the score is, the higher the performance of the alternative for a given user. Sometimes these methods also take into account the rating history of the user. In this approach, the recommendation system relies on having an accurate knowl- edge of the user’s preferences to be able to select the appropriate items. This kind of approaches may suffer from the ‘‘cold start’’ problem when a new user enters in the system, because we can elicit poor knowledge about the user in an initial stage. Some solutions to this problem are explained later in this section. In CB systems the recommendation process is mainly focused on defining an appropriate measure to compare a user and an item. The two most common approaches are the aggregation of ratings and distance calculation. � When the user profile is represented as a rating vector with the degree of interest of the user in each attribute, each rating can be interpreted as a performance score that can be used to eval- uate an alternative. The goal is then to calculate an overall inter- est score for a certain alternative. The simplest approach consists of using an aggregation operator to combine the user ratings on the concepts that define a certain alternative (Batet et al., 2012). More sophisticated aggregation methods have also been applied, like AHP (Analytic Hierarchy Process) (Huang & Bian, 2009; Niaraki & Kim, 2009). � When the items and the users are described by a list of key- words, some similarity measures can be applied. For example in Lamsfus et al. (2009) items and users are described using concepts from an ontology, which defines archetypes of tourists (e.g. cultural, sportive or adventurous), and the cosine similarity between the two vectors (user and item) is calculated. A similar approach is proposed in Gyorodi, Gyorodi, and Dersidan (2013) with ad-hoc hierarchies of tags for locations that are rated by users. The locations ratings are then compared to the user’s tags. In García-Crespo et al. (2009) a feature-based similarity algorithm is applied, using several ontologies as reference. In Fenza et al. (2011) classification rules are automatically gener- ated and later used to define the matching degree between the user and the item. J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7379 Other methods of recommendation exploiting AI techniques are explained in Section 5, such as the ones using rules or probabilistic information. Collaborative (CL) systems make recommendations based on groups of users with similar preferences. The similarity between users is normally computed by comparing the ratings that they give to some of the items. When the system identifies who are the people that share similar interests with the current user, then the items that those people liked are recommended to this user. In this approach, some feedback about the provided recommenda- tions is necessary, in order to know which items the user has liked or disliked (e.g. which places she has enjoyed visiting). Two types of CL methods are distinguished: user-based and item-based. The former finds neighbours of a target user by matching her opinions with the ones of the other users in the system. The latter builds groups by finding similarities on the items that the users liked (or disliked) in the past. Two weak points are recognised in CL systems: ‘‘data sparsity’’ and ‘‘grey sheep’’. The former occurs when the number of ratings from users is small in comparison with the total number of items, so that the probability of finding users that rate the same items is too low to make good estimations. The latter, ‘‘grey sheep’’, refers to a user with a profile different from the rest of users of the sys- tem. In this case, it is difficult to find appropriate items to recom- mend because we do not have information about similar users. Finally, this approach also suffers from the scalability problem if the community of users is large. Demographic-based (DM) systems rely on the demographic data of the user (e.g. age, country of origin, level of studies, etc.). In this case, the recommendation is not based on the user’s interests and preferences but on her personal characteristics. In this approach, users are usually assigned to a certain stereotypical class depend- ing on their demographic data, so that the members of the same group share a common demographic profile. The system has inter- nal knowledge about the standard preferences of each stereotype, which is used to provide the recommendations to the users. The definition of stereotypes of tourists is not new in this field. Many studies have defined segments of tourists according to their behav- iour in different cities or territories (Brewer, 1984; Marques, 2009; Tsung-Chiung, Chyong-Ru, & Wan-Chen, 2012). Specific stereo- types provide precise descriptions of what tourists want and how they act in different situations. This information is normally used as a guide to conduct business with tourists, but it can also be exploited in recommender systems. Since each of the approaches has some drawbacks, the combi- nation of different techniques is also a widespread practice. Fig. 13 shows the distribution of the types of recommendation techniques used in the field of tourism recommenders, in percent- ages. Half of the works use a mixture of techniques (53%), combining Fig. 13. Recommendation techniques used in the reviewed works (in%). mainly CB methods with CL filtering or with DM techniques. The rest of the systems apply a single approach, having a clear predominance for the techniques based exclusively on the description of the con- tent of the alternatives (38% of the reviewed papers). Hybrid systems can integrate these techniques in different ways. Three approaches can be distinguished: 1. Selection of the method: the system incorporates DM, CB and CL methods, but only one of them is applied depending on the par- ticular situation of each user. For example, the first time the user arrives, a method based on demographic data is used. Later on, if similar users can be found, a CL recommendation is made. Otherwise, a CB procedure is applied. This is the case of Martínez, Rodríguez, and Espinilla (2009), Huang and Bian (2009) or Noguera et al. (2012). 2. Sequential use: each recommendation technique is used in dif- ferent stages of the process. For example, SPETA (García-Crespo et al., 2009) has four steps: first, contextual information such as the location or time are used to make the first selection of appropriate options; second, a more fine grained set of results is obtained using knowledge-based filtering techniques, by cal- culating the semantic similarity between the user preferences and the touristic services; third, preferences and CL techniques are used to refine the set of options; finally, in the fourth step, a vector of preferences is used to make the final selection. In Braunhofer et al. (2013) CL filtering with DM and personal information is applied in a training phase to build a prediction model in different contexts. Once the model has been trained, CB techniques generate the list of recommendations by com- puting ratings for each item based on the current and predicted values. 3. Integrated use: both CB and CL techniques are combined during the execution. For example, in SigTur (Borràs et al., 2011) differ- ent ratings are calculated to estimate the interest of an activity for a target user. Ratings are obtained using DM-clustering, CL-clustering and CB similarity; afterwards, these ratings are merged to find an overall quality rating for each item and make the filtering of the best tourist attractions. In Lucas et al. (2013) the users are classified into groups using simultaneously per- sonal demographic data (DM), information about the content of the items previously selected by the user (CB) and the infor- mation of other users (CL). Then a set of fuzzy rules is automat- ically generated so that new users can be automatically classified into several groups (with different membership degrees). The list or recommended items is finally derived from a prediction based on the groups the user belongs to. In the survey we have observed an increasing trend in the exploitation of CL filtering techniques since 2012, mainly in hybrid systems. More precisely, from 2008 to 2011 only 25% of systems used such method, whereas since 2012 the percentage has increased to 75% (see Table 4). 4.2. Representation of the user profile A key component in recommender systems is the user profile, which stores the information related to the user preferences and permits to make personalised recommendations. Different user profile models have been developed. The simplest model associates to each user a list of preferred keywords or categories in which items are pre-classified. However, this information is usually too general to provide accurate recommendations. A more widespread approach consists of storing a vector with numerical ratings corre- sponding to the attributes of the items. This rating indicates the degree of interest of the user with respect to each attribute. This vector approach facilitates the inclusion of other types of features Table 4 Review of recommendation methods used. Recommendation method Reference ALL Lucas et al. (2009), Ruiz-Montiel and Aldana-Montes (2009), Borràs et al. (2011, 2012a), Batet et al. (2012), Koceski and Petrevska (2012), Braunhofer et al. 2013, Garcia et al. (2013), Lucas et al. (2013) and Meehan et al. (2013) CB + CL Castillo et al. (2008), García-Crespo et al. (2009), Fenza et al. (2011), Rey-López et al. (2011), Noguera et al. (2012) and Rojas & Uribe 2013 CB + DM Coelho et al. (2009), Ceccaroni et al. (2009), Lamsfus et al. (2009), Niaraki and Kim (2009), Mínguez et al. (2010), Yang (2010), Martin et al. (2011), Sebastià et al. (2009) and Garcia et al. (2011) CL + DM Gavalas and Kenteris (2011) CB Huang and Bian (2009), Lee et al. (2009), Seidel et al. (2009), Yu and Chang (2009), Jannach et al., (2010), Ricci et al. (2010), Sebastià et al. (2010), Vansteenwegen et al. (2010), García-Crespo et al. (2011), Kurata (2011), Linaza et al. (2011), Lorenzi et al. (2011), Luberg et al. (2011), Montejo- Ráez et al. (2011), Martínez-Santiago et al. (2012), Gyorodi et al. (2013), Kurata and Hara (2013) and Ruotsalo et al. (2013) CL Savir et al. (2013), Umanets et al. (2013) and Yang and Hwang (2013) DM Wang et al. (2011) 7380 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 in the profile, such as demographic information, as can be seen in Codina and Ceccaroni (2010), Garcia et al. (2011), Mínguez et al. (2010) and Kurata and Hara (2013). These basic models can be extended by means of some AI knowledge representation techniques. One possibility is to use semantic models, in which the domain knowledge usually takes the form of an ontology. For example, the keyword list may contain names of classes in the ontology (i.e. concepts) and the ratings vec- tor can also be defined in the space of the concepts in the ontology. Uncertainty models have also been incorporated into some recom- mender systems to be able to handle the credibility associated to the information stored in the profile. The rating values are uncer- tain in many situations, especially when they are not given explic- itly but have to be inferred from the user interaction with the domain items. A confidence degree can be associated to each rating in the profile and can be used as a weighting factor in the exploi- tation stage. More details about these techniques are given in Sections 5.5 and 5.4, respectively. In order to produce up-to-date personalised recommendations to the same user along time, the user model has to be updated. Feedback information is used to modify the profile when any change on the user’s preferences is detected. This information can be collected explicitly or implicitly. Explicit feedback is obtained by means of the direct interaction with the user. The deci- sion maker is requested to fill in some form (giving her opinion on different values of the criteria or indicating her location) or to rate a set of alternatives. This approach provides precise knowledge because the data is given directly by the user. However, it is usually considered quite an intrusive way of elicitation, and many users are not keen on spending time in answering this kind of questions. Techniques based on implicit feedback aim at collecting the user information by analysing her behaviour in the system, such as Table 5 Review of the AI techniques used. AI techniques References Multi-agent systems Castillo et al. (2008), Ceccaroni et al. (2009), Lee et al. (200 Optimization techniques Castillo et al. (2008), Lee et al. (2009), Garcia et al. (2010), Va (2013) Automatic clustering Castillo et al. (2008), García-Crespo et al. (2009), Martínez e Noguera et al. (2012), Lucas et al. (2013), Kurata & Hara 20 Management of uncertainty García-Crespo et al. (2009), Huang and Bian (2009), Lamsfu et al. (2012), Lucas et al. (2013), Meehan et al. (2013) and R Knowledge representation Castillo et al. (2008), Ceccaroni et al. (2009), Lamsfus et al. ( (2011), García-Crespo et al. (2011), Wang et al. (2011), Alon (2013), Ruotsalo et al. (2013) and Moreno et al. (2013) the alternatives that are selected, purchased or viewed. More sophisticated tools study the sequence of actions done by the user on a certain alternative, or even the amount of time spent with each one. The main advantage of these methods is that an addi- tional effort from the user is not required. For example, in Savir, Brafman, and Shani (2013), the travellers modify a trip plan and the system observes the modifications to adjust the budget, daily time, maximum travel distance, transportation constraints, etc. Other works (Albanese, Chianese, d’Acierno, Moscato, & Picariello, 2010; Albanese, d’Acierno, Moscato, Persia, & Picariello, 2013; Moscato, Picariello, & Rinaldi, 2013) analyze the sequence of acces- ses to information, to deduce which is the most accessed object after viewing a specific one. They use both low level (colour, tex- ture) and high level (author, type) features of objects to compare the usage behavioural patterns. Despite the advantages of implicit feedback, this kind of data is more uncertain than explicit informa- tion, so less confidence must be given to it when the profile is mod- ified. In this review we observed that around 60% of the works use only explicit feedback, whereas the rest combine explicit and implicit feedback. The most common approach in explicit feedback collection consists of requiring the user to vote the different options proposed by the system. The most sophisticated approach is presented in the VIBE system (Jannach et al., 2010), which builds personalised dialogues to gather new knowledge about the user preferences. 5. Use of AI techniques in tourism recommender systems This section makes a brief review of the main AI techniques and tools employed in tourism recommender systems in the last years, which are summarised in Table 5. 9), Seidel et al. (2009), Sebastià et al. (2010) and Lorenzi et al. (2011) nsteenwegen et al. (2010), Garcia, Vansteenwegen, et al. (2013) and Meehan et al. t al. (2009), Fenza et al. (2011), Gavalas and Kenteris (2011), Batet et al. (2012), 13, Ruotsalo et al. (2013) and Moreno et al. (2013) s et al. (2009), Lamsfus et al. (2011), Pinho et al. (2011), Wang et al. (2011), Hsu uotsalo et al. (2013) 2009), Lee et al. (2009), Sebastià et al. (2009), Sebastià et al. (2010), Garcia et al. so et al. (2012), Batet et al. (2012), Martínez-Santiago et al. (2012), Lucas et al. J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7381 5.1. Multi-agent systems Agents are autonomous and proactive software entities capable of obtaining information from their environment and acting in an intelligent way upon it in order to try to accomplish a set of goals or objectives. Multi-agent systems are groups of agents that com- municate between themselves to share information and resources, coordinate their activities and cooperate in the joint efficient solu- tion of a distributed problem (Wooldridge, 2009). Turist@ (Batet et al., 2012) is an agent-based system that pro- vides personalised recommendations on cultural activities. The architecture of the system is shown in Fig. 14. There is one agent for each kind of cultural activity, which maintains a small database with the events of that type available in the city (museums are the exception, as there is one specific agent for each museum in the city). The user interacts with the system through a graphical inter- face provided by a User Agent. A Broker Agent mediates the commu- nication between the User Agents and the cultural activities agents. The user can make specific queries, can evaluate an activity that she has attended, or can ask for a personalised recommendation. The core of Turist@ is the Recommender Agent, which maintains a user profile for each tourist. This profile is initialised with some basic information on high-level cultural interests provided by the user when she uses the system for the first time. The Recommender Agent dynamically and automatically refines this initial knowledge about the user preferences by analysing the user’s queries and evaluations. The User Agent can also provide proactive recommen- dations, because it knows the position of the user in the city and can suggest cultural activities that fit the user’s preferences and are located in the vicinity. The system uses both CB and CL recom- mendation techniques. The idea of having an initial profile and refining it by analysing the explicit (evaluations) and implicit (actions) activities of the tourist is also given in Ceccaroni et al. (2009). That work proposes to have a Profile Management Agent, which not only initializes the profile (by fitting the user into stereotyped classes) but also mod- ifies it depending on the feedback provided by the tourist. In this agent-based proposal there are Information Service Agents that retrieve touristic information from databases and ontologies, and a Personalization Agent that, given the user profile and the available touristic data, applies CB recommendation techniques to select the items that should be suggested. In PersonalTour (Lorenzi et al., 2011) there is a set of Travel Agents, and each of them is specialised in the recommendation of flights, hotels or attractions. When a new costumer arrives and Fig. 14. Multi-agent based architecture expresses her preferences, these agents collaborate among them- selves in order to propose a travel package to the tourist. The user can later evaluate each of the components of the package, provid- ing a feedback to the system so that the degree of expertise of each Travel Agent can be conveniently updated. Some recommenders (e.g. Castillo et al., 2008; Lee et al., 2009; Sebastià, Giret, & Garcia, 2010) ‘‘agentify’’ the different compo- nents of the system (the interface with the user, the capture of her requirements and preferences, the analysis of the suitability of each attraction, the creation of a route among the selected points of interest), although there is not any kind of complex communica- tion or coordination between them. In all these systems the agents seem to work in a sequential fashion, without any kind of coordi- nated effort. Therefore, the full potential of distributed, concurrent and coordinated behaviour of agents is not employed. 5.2. Optimization techniques Many tourism recommender systems have to solve complex planning and scheduling problems, which are well known to be NP complete and, therefore, cannot be optimally solved in an effi- cient way. In some cases, researchers have opted for the use of dif- ferent kinds of optimization techniques which, although in many cases they do not guarantee the optimal solution, offer an afford- able computational cost. One example is the agent-based travel route recommender for Tainan (Lee et al., 2009), that uses ant colony optimization tech- niques. In these methods a set of autonomous entities (which rep- resent the ants) cooperate through pheromone-mediated indirect and global communication to find a good solution to the travelling salesman problem (in this case, to plan a route that goes through different points of interest around the city). CT-Planner4 (Kurata & Hara, 2013) uses a genetic algorithm to construct the plan to visit a city. In each iteration of a cyclic process it considers a population of different possible plans, which are evaluated according to their utility for the user; the best ones are mutated and recombined via crossover to generate another population for the next iteration. After a certain number of iterations, the best plan is finally selected. The authors of the VISIT system (Meehan et al., 2013) pro- pose to make recommendations adapted to the context of the user, that is composed of different factors (location, time, weather, social media sentiment and user preferences). In that work, they suggest the idea of using an artificial neural network to assess the relevance of each context component for each user. of Turist@ (from Batet et al., 2012). 7382 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 Some heuristic procedures to build travel itineraries were explored in the City Trip Planner system and related works (Garcia, Arbelaitz, Linaza, Vansteenwegen, & Souffriau, 2010; Garcia, Vansteenwegen, Arbelaitz, Souffriau, & Linaza, 2013; Vansteenwegen et al., 2010). One possibility is the use of Iterated Local Search, a meta-heuristic iterative method that builds sequences of solutions generated by a local search. The heuristic perturbs the solution found by the local search (a route to visit some city attractions) to create a new solution. Then, it takes the best solution as the new starting solution for the local search. The process is repeated until a termination criterion is met. Another option that was studied is the use of meta-heuristic itera- tive Greedy Randomised Adaptive Search methods (Souffriau, Vansteenwegen, Vanden Berghe, & Van Oudheusden, 2011). In each iteration a list of possible visits is generated from an initial solution which contains only the start and end of each tour. Those visits that have a heuristic value below a certain threshold are eliminated. A random visit from the remaining list is selected and applied to the current solution. Most of the tourism recommender systems that build personalised routes or itineraries implement an ad-hoc planning mechanism, but some of them apply more classical domain- independent AI planning techniques. For instance, in the SAMAP system (Castillo et al., 2008) the use of heuristic, A⁄ and hierarchi- cal temporal planners was explored. 5.3. Automatic clustering Many tourism recommenders employ techniques based on CL filtering, in which the users of the system are partitioned into groups that share some common characteristics. The basic idea of these methods is that it can be appropriate to recommend to the Fig. 15. Portion of the SAMAP domain ontol user those items that have been positively valued by similar tour- ists. The concept of similarity employed to group users may be based on demographic information, on the general preferences of the users over diverse types of touristic activities, or on the explicit ratings of individual activities. In any case, the automatic cluster- ing tools developed in AI may be successfully used to classify the tourists. This section comments different alternatives that have been used in touristic recommender systems. A very simple way of associating a new user with similar past users of the system is to employ the k-nearest neighbours approach (Dasarathy, 1991), calculating which are the k past users of the sys- tem who were more similar to the current one (e.g. Martínez et al., 2009; Noguera et al., 2012). Having done that, the information on those users may be employed to provide recommendations (e.g., the activities that were more highly valued for them). In SAMAP (Castillo et al., 2008) the similarity between users is based on the preferences expressed over the concepts of a domain ontology (a portion of it may be seen in Fig. 15). For instance, the system could easily infer that a user that likes Cinema is more similar to a user that enjoys Theatre than to another that prefers Sport activities. Scalability is one of the main problems to be addressed when using this method. A common option to group the users into different classes is to use the k-means algorithm (e.g. Gavalas & Kenteris, 2011; Pinho, da silva, Moreno, de Almeida, & Lopes, 2011). The initial seeds of the k desired clusters are established in some application-dependent way. Then there is an iterative process in which, in every step, the objects are sorted into the nearest cluster and the cluster pro- totypes are recalculated. The method converges to a solution when the objects belong to the same clusters in two consecutive itera- tions. In Moreno et al. (2013) the k-means algorithm is applied with three different purposes: to obtain a set of initial tourist segments, ogy (adapted from Castillo et al., 2008). J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7383 to obtain classes of users with similar demographic characteristics, and to classify users according to the explicit ratings they have pro- vided. In the first case, the historical data of 30,000 questionnaires was used to compute a set of 100 generic tourist types (segments). Each of them had an associated prototype and a preference level for each kind of activity. New users fill a small form with some basic personal and preference data and it is possible to compute the segment to which they belong, guiding the initial recommen- dations. The demographic classification is mainly based on the country of origin, the travel group composition, the travel budget and the accommodation type. Aggregation operators like OWA (Yager, 1988) and LSP (Dujmovic & Nagashima, 2006) are used to join the similarities on these different kinds of data. The classifica- tion of the users according to their ratings uses the well-known Pearson correlation as similarity measure. The recommender system described in Fenza et al. (2011) pro- poses the use of the uncertain version of k-means, fuzzy c-means. The result of this algorithm is a fuzzy partition of a set of objects into clusters, so that each object has a degree of membership between 0 and 1 to each cluster, and the addition of the degrees of membership to all the clusters is 1. This algorithm is both applied to users and to touristic points of interest (POIs). After the definition of clusters of users and POIs, the system is able to derive rules that characterise them, that are used to integrate new users and new POIs to the clusters in which they fit better. This work also proposes to build association rules, which explain the relationship between clusters of users (plus contextual infor- mation) and clusters of POIs. These rules permit to determine the kind of touristic activities that should be recommended to a certain type of users. Very similar techniques are employed in the PSIS (Personalised Sightseeing Information System) recommender (Lucas et al., 2013). Turist@ (Batet et al., 2012) also employs CL filtering recommen- dation techniques that require the definition of classes of similar users. The clustering is applied every time that 10 new users join the system, so classes are periodically recomputed. The employed clustering system is ClusDM (Valls, 2003), which builds a hierarchy of classes taking into account the interests of the users in general kinds of activities and their demographic data. The tree generated by the algorithm can be cut at different levels to generate parti- tions with the desired number of classes. The use of Support Vector Machines (SVMs) as a classification technique in tourism recommender system is suggested in the SPETA system (García-Crespo et al., 2009). Tourist preferences on several kinds of tourist activities are stored in a vector, and the characteristics of each activity are also stored in the same way. Thus, SVMs may be used to compute the distance between the Fig. 16. Use of a Bayesian network to detect the preferred user’s preferences and the recommendable items, so that the most appropriate ones can be efficiently found. 5.4. Management of uncertainty The task of recommending activities to a tourist is not simple, as there is not any clear and precise relationship between the charac- teristics and preferences of a visitor and the POIs available at a given destination. Some of the techniques developed in the AI field of approximate reasoning have been proposed to represent and rea- son about this uncertain relationship. One possibility is to use Bayesian networks (Pearl, 1988). A Bayesian network is an acyclic graph in which edges represent relationships of causality or influence between nodes. Nodes that do not have any parent have an associated probability table, indi- cating how likely they are to occur. Nodes that have n parents have a conditional probability table of 2n nodes, indicating how likely they are to occur depending on the presence (or absence) of their parents. A very simple use of Bayesian networks is presented in Hsu, Lin, and Ho (2012), where a number of attributes (age, nation- ality, occupation, income, travel motivation, etc.) influence directly on the probability that a certain touristic point is interesting for the user. The initial Bayesian network was built after the analysis of more than 2400 questionnaires. A more complex application of this kind of networks is given in Huang and Bian (2009) and Wang et al. (2011). They propose a network (see Fig. 16) in which the age, occupation and personality influence the type of user which, along with the travel motivation, influences the probability of the user liking a certain kind of touristic destinations. Specific touristic events are not included in the network. Another common option to manage uncertainty is the use of fuzzy logic. A fuzzy variable make take as values a series of linguistic labels. Each linguistic label has an associated fuzzy set, in which every value in the domain of reference is assigned a membership value to the set between 0 and 1. In that way, fuzzy logic provides a generalisation of standard logic. Fuzzy sets and fuzzy reasoning may be used to represent the preferences of the user and to calcu- late how they fit with the characteristics of a tourist attraction (García-Crespo et al., 2009; Lee et al., 2009), to obtain the degree of membership of each user to different groups of users (Pinho et al., 2011) or to represent contextual aspects of the journey (Meehan et al., 2013). For instance, if the weather conditions are represented with a value between 0 and 1, instead of using a simple Boolean value for good/bad weather, it is possible to make a more fine grained analysis of the weather conditions and reason about its influence on the recommendation of each cultural activity. kind of tourist activities (from Huang & Bian, 2009). 7384 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 Some touristic recommender systems also employ a rule-based approach, but without the addition of a fuzzy component. For instance, in the CONCERT system (Lamsfus et al., 2009; Lamsfus et al., 2011) there are rules that detect the events to be recom- mended depending on the user preferences and the context, such as this one: hasFoodPreferencesRule: (v? red:type dcl:Visitor), (?v dcl:hasPre- ferences ?p), (?p red:type dcl:FoodPreferencesDemographics), (?v dcl:usesDevice ?d), (?d dcl:isConnectedToNetwork ?n), (?n dcl:hasLo- cation ?l), (?l dcl:hasEnvironment ?e), (?e dcl:offersKindOfTourism- Concepts ?s), (?s dcl:isRestaurantOfTypeVegetarian ?r) => print (?r dcl:isTourismServiceOfferedToVisitor ?v) 5.5. Knowledge representation Recommender systems in e-Tourism need, as any knowledge- based intelligent system, a way to represent in an efficient way the domain knowledge, so that it can be used in their reasoning processes. The knowledge representation and reasoning tech- niques developed in AI are adequate tools for this purpose. In par- ticular, nowadays the most common way of representing domain knowledge is the use of ontologies. An ontology describes a shared and explicit formal conceptualization of a given domain. Its main components are classes (representing concepts, usually organised in some kind of hierarchical structure), taxonomical and non-taxo- nomical relationships (Sánchez & Moreno, 2008a, 2008b), axioms (Sánchez, Moreno, & Del Vasto-Terrientes, 2012) and instances (representing specific objects). There are several tourism recommenders that employ ontolo- gies to formalize the domain knowledge. Most systems have gen- eric ontologies that store information about different aspects that have to be taken into account in the recommendation of cultural activities. For example, the system described in Wang et al. (2011) includes a generic travel ontology (with information about accommodation, restaurants, transport, shopping, culture, etc.) and a user ontology in which the demographic characteristics of the tourists and their preferences are modelled. Other systems like GeOasis (Martínez-Santiago et al., 2012), SAMAP (Castillo et al., 2008) (see Fig. 15), SMARTMUSEUM (Ruotsalo et al., 2013) and the one proposed in Alonso et al. (2012) also include ontologies to model the different kinds of touristic activities and to be able Fig. 17. Portion of the tourism ontology devel to reason on them in a semantic fashion. These systems use ontol- ogy-based similarity measures to deduce if two kinds of activities are similar, and this knowledge may also be used to compute the similarity between users and provide recommendations based on CL filtering techniques. Two systems that make a heavy use of ontologies are SigTur (Borràs et al., 2011; Moreno et al., 2013) and e-Tourism (Garcia et al., 2011; Sebastià et al., 2009; Sebastià et al., 2010). They use domain ontologies that detail the different kinds of leisure activi- ties available in the city. For instance, Fig. 17 shows a portion of the SigTur ontology, which has more than 200 concepts organised in a hierarchy of 5 levels. SigTur stores in each node of the ontology the preference of the user (and the confidence of the system on that preference). These values are initialised with a small questionnaire filled by the user when entering the system (preferences on the most generic con- cepts are transmitted through their children to all the hierarchy). When the user performs actions on the recommended activities (e.g. adds an activity to the current travel plan) the system updates the preferences on the concepts associated to that activity and spreads this information through their parents in the ontology. Thus, there is an ontology-based dynamic management of the user preferences (Borràs, Valls, Moreno, & Isern, 2012b). The e-Tourism system follows a similar approach, in which the user preferences are updated after analysing the explicit ratings she has provided. There are systems that propose the use of a set of ontologies, instead of having a single integrated ontology. PaTac (Ceccaroni et al., 2009) includes separate ontologies about cultural activities, restaurants, entertainment, hotels, etc. The authors propose to link those ontologies with standard temporal and geo-location ontolo- gies provided by the W3C consortium and with a User Model ontology that contains different kinds of touristic stereotypes. CONCERT (Lamsfus et al., 2009) models all the context associated to a travel in a network of ontologies called ContOlogy, which includes 11 separate ontologies that deal with aspects like tourism services, preferences, activities or travel motivations. Normally the ontologies used by recommender systems are designed ad-hoc for a specific application and built manually. A way to reduce the cost of the construction of the ontology (Ruíz- Martínez, Miñarro, Castellanos, García, & Valencia, 2011) is to pop- ulate it in automatic fashion, by analysing electronic resources (e.g. oped in SigTur (from Moreno et al., 2013). J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7385 Web pages), extracting the appropriate information about tourist activities and creating the associated instances. A similar proposal was made in Vicient, Sánchez, and Moreno (2013). 6. Related reviews There have been some previous papers explaining the applica- tion of recommender systems in the tourism area, which are chro- nologically mentioned in this section. Ricci (2002) and Staab and Werthner (2002) explain in a very generic way the characteristics of travel recommender systems with some examples, without making an exhaustive review or comparing different approaches. Werthner (2003) gives also a very generic description of technolog- ical approaches applied to tourism, where some examples related to Artificial Intelligence are mentioned. However, it does not pro- vide any review or comparative analysis of different systems. Berka and Plößnig (2004) provides a brief guide on how to design recommmender systems for tourism, but it does not attempt to make a survey of the area either. The survey that is more similar to this work is Kabassi (2010). It is mainly a classification of tourism recommender systems (until early 2009) under different criteria: kind of objects they recommend (hotels, flights, restau- rants, etc.), hardware support (computer or handheld device), indi- vidual/group recommendations, explicit/implicit acquisition of information from the user, recommendation technique (content- based, collaborative or hybrid) and personalization techniques (mainly decision-making tools and Bayesian networks). The authors of that paper basically group the systems in these catego- ries, without making a deep analysis or explanation of all these possibilities. It does not provide any guideline on how to build this kind of systems and it does not consider the latest advances in the last five years, which are the basis of our study (advanced geolocal- isation capabilities of mobile phones and tablets, context-aware recommendations, semantic management of preferences, use of social networks, etc.). Gretzel (2011) makes an analysis of tourism recommenders from the point of view of social sciences, not from the technological perspective. The author of this paper argues that intelligent systems are necessary in the tourism domain because there are many complex aspects to be managed: the mobility of tourists, the increased risk and uncertainty experienced in unfa- miliar environments, the distributed nature of information sources, the idiosyncratic quality of tourism decision-making, the multi- faceted nature of tourism experiences, and the interdependency of suddecisions. A description of some systems that tackle those issues is done. The author also comments the main issues on the design and the evaluation of those systems, focusing on the user interaction, the context, the social perspective and the decision making process to maximise tourists utility; however, this work does not cover the use of intelligent techniques. The main recommendation methods applied in tourism are reviewed in Felfernig et al. (2007). This paper presents some exam- ples of the use of these techniques, but they are not deeply described nor compared. This paper emphasises some interesting topics like group recommendation and context-aware recommen- dations in mobile devices. Finally, it is worth mentioning the paper (Vansteenwegen & Souffriau, 2011), that makes a deep overview of systems built between 2001 and 2011 that compose trip plans, although they only comment this single functionality. The authors compare each of the reviewed references in terms of these plan- ning functionalities: personal interest estimation, selection and routing, mandatory points of interest, dynamic recalculation (update plan in real time when unexpected events occur), multiple day decision support (enable plans for multiple days), opening hours, budget limitations, max-n Type (limitation of activity types per day), mandatory types, weather dependency, scenic routes (build paths with beautiful views rather than the shortest ones), hotel selection, public transportation and group profiles. This paper describes how the orienteering problem and its extensions can be used to model trip planning functionalities. In summary, as far as we know, there is not any recent survey of tourism recommenders with the technological focus, novelty and breadth of coverage of this paper. 7. Conclusions This final section contains a brief summary of the work pre- sented in this paper, describing some points that should be taken into account by scientists aiming to design and develop tourism recommender systems, and an outline of several lines of future work. 7.1. Summary and guidelines to develop tourism recommenders Tourism recommender systems give personalised and relevant suggestions to tourists whenever they visit unknown places. They provide support tools to make the process of deciding what to do more manageable. In this article we have reviewed tourism recom- mender systems published mainly in AI-related scientific journals and conferences since 2008. We first analysed the interfaces used by these systems and we pointed out the predominance of Web-based approaches, which are especially useful for tourists when they are planning a visit before the stay. However, lately the usage of mobile platforms has widely increased, since they allow a direct access to the infor- mation about attractions during the stay. Moreover, they also per- mit to personalise and contextualise the gathered information, for instance taking the current location of the user into account. How- ever, we have noticed that new mobile platforms such as Android or iPhone have been weakly exploited. Since these platforms are currently being widely used for tourists when visiting places, it is necessary to address the development of applications for those systems and to create responsive Web designs that permit to adapt the content to any viewing device. Tourism recommender systems, as we have seen, not only manage textual information, but most of them use images, pictures and interactive maps. We have also analysed the main functionalities of the reviewed tourism recommender systems. Most of the approaches suggest points of interest in a destination according to the user prefer- ences. These suggestions may be shown in a ranked list ordered by importance or they may be integrated in a scheduled plan. Usu- ally the system provides only a list and a planner that the user can employ to build manually the detailed plan. However, some approaches can make this process automatically, hence providing an almost complete route taking the context of the user into account, such as location, time and opening hours, among others. This is a novel issue and it may be extended in the future with more functionalities and contextual elements. Another aspect added in the last years is the use of social features that allow tourists to share information among friends. The recommendation process is a crucial aspect in tourism advi- sory systems, hence we have analysed the main mechanisms used in the reviewed articles. The most popular approaches use content- based, collaborative and demographic-based techniques. These techniques suffer from several problems when applied individu- ally. Hence, a good practice is the combination of several techniques together to overcome the mentioned drawbacks. We observed that the most recent approaches follow this trend and propose hybrid recommendation methods, including also contex- tual information. However, there are still a large number of approaches that apply only content-based methods. In tourism 7386 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 recommender systems it is also a good practice to include a diver- sification mechanism to avoid content-based problems and to offer unexpected and off-the-beaten-track alternatives to tourists. Recommenders need to learn and manage the user profile to make accurate personalised suggestions. This user profile can be repre- sented by a vector of words or be extended with semantic and uncertainty models which offer a richer representation of the user’s preferences. These preferences can be acquired explicitly or implicitly. The most common method is the acquisition of explicit information. However, we consider interesting to apply a combination of both methods. Even though implicit information is inherently more uncertain, it is also less intrusive for users and it is easy to collect it directly by monitoring the interaction of the users with the system. We have also shown how a wide range of Artificial Intelligence techniques are already being employed in e-Tourism recommend- ers. Knowledge representation and reasoning techniques are com- monly used to represent and reason about the tourism domain. In this field it is especially promising the introduction of semantic measures of similarity between users and items (or between users), taking advantage of detailed ontological representations of the domain. Intelligent autonomous agents may be used to provide proactive recommendations depending on the context; therefore, their use may be especially relevant in the case of mobile-based recommenders. Approximate reasoning techniques are applied to manage the uncertainty on the user’s preferences, making them an ideal option when the user feedback is only impli- cit. Reasoning procedures of different kinds may be employed to infer the user’s preferences. Automatic clustering algorithms may be successfully used to classify tourists with similar tastes or sim- ilar features. Finally, optimization techniques offer cost-effective solutions to complex planning and scheduling problems when the system wants to build automatically a multi-day route. As a result of this analytic review, some basic guidelines that can be followed in the design and development of Tourism recom- mender systems may be given: � The use of both a Web and a mobile version of a recommender system can be suggested as a good practice, since the Web ver- sion can be comfortably used at the home of the tourist to select the destination, the activities to be carried out and a global planning of the routes, whereas the mobile version (that should run on the latest and most widely spread platforms) may permit to follow the route in situ, receive more detailed information about the visited places, receive proactive recommendations depending on the user preferences, or even share the travel experience through social networks. The mobile version should provide a rich graphical interaction with the user, to enhance the tourist experience (e.g. augmented reality approaches) and also exploit heavily contextual information (e.g. the current time, the location of the user, etc.) to improve the accuracy of the recommendations and adapt it to the dynamically changing circumstances of the trip. It is sensible to design dedicated Web interfaces for mobile devices, or even better to create respon- sive Web designs that provide an optimal viewing experience for any device (from personal computers to mobile phones or tablets). A very interesting topic is to extend the current formats with videos, 3-dimensional objects or augmented reality, as some approaches already start to do. � Another way of getting more information from the user and to improve the information that the recommender has on her interests is the exploitation of all the data provided by social networks and other Web 2.0 applications, including the rela- tionships between users and the different kinds of content they provide (comments, pictures, ratings). This aspect is certainly very relevant in the tourism field, due to its highly social nature. Thus, it is important to include in tourism recommenders as many possibilities of sharing information (pictures, videos, comments, ratings, localisation, etc.) as possible. The analysis of the social relationships of the users is a recent area of work that can surely lead towards the discovery of more accurate rec- ommendations that fit better with the user’s tastes, by taking into account the opinions of her closest friends, weighting the opinions depending on the strength of the relationship with the acquaintance, etc. � In any kind of recommender systems it is essential to have the precise information about the user’s interests stored in her pro- file. The particular characteristics of the tourism field offer the possibility to define new mechanisms to analyse the interaction of the user not only with the recommender system but also with the related environment, in order to learn automatically (and dynamically adapt) the user’s preferences. In particular, contextual recommendations are key in the success of any recommender of tourist activities. � It is also becoming increasingly clear that, in order to provide precise recommendations, it is necessary to move away from purely textual information and represent in a semantic way (e.g. through the use of ontologies) both the preferences of the user and the features of the different kinds of cultural and leisure activities. Having this structured information, it is possi- ble to define and use complex semantic similarity techniques to compare users, compare objects or compare the preferences of the user with the characteristics of the objects. 7.2. Lines of future work This section briefly comments three relevant issues that are cur- rently being studied in the development of e-Tourism recommend- ers: the diversification of the suggestions provided to the user, the use of social data available in current Web 2.0 applications, and the improvement of the recommendations by leveraging the extra capabilities of mobile devices. � Content-based systems focus on recommending items similar to the user’s profile, which may cause overspecialized results, leaving aside other items that might be interesting for the user. This is an important issue in some applications in the field of tourism. Some recommender systems aim at making publicity of ‘‘different’’ or new sorts of activities which may be ignored by most visitors (e.g. a new restaurant or a new guided tour). It has also been argued that a smart recommender should pro- vide a diversified list of recommendations (e.g., even if the sys- tem knows that the user is interested in going to the beach, it is not very exciting to show a list of ten different beaches and not to suggest other kinds of related activities). In Savir et al. (2013) a measure of balance between the number of attractions of a certain type and the minimum rating threshold is proposed in order to keep a fixed diversity level in the activities proposed in a trip. The SigTur recommender system Borràs et al. (2011) also includes a diversification mechanism that aims to widen the range of suggested activities. In Ruotsalo et al. (2013) the objects of a museum are gathered in clusters sharing the same features so that the recommendation procedure picks a repre- sentative number of objects from each cluster to increase the diversity of the proposal made to the visitor. � Recent recommender systems exploit the power of new Web- based applications, like social networks. In addition to offering social functionalities, these tools facilitate the use of collabora- tive filtering techniques, since this kind of technologies permit new forms of rating items or collecting user information at an individual level or at a social level. They are known as social rec- ommender systems (Noel et al., 2012). These tools can be used J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7387 both to identify groups of items and to build groups of users. For example, in moreTourism (Rey-López et al., 2011) the users have an associated tag cloud with terms relevant to their profile, and a new tag is created for each attraction based on the tags of the users who liked it. This information is used to compare the tag clouds of users and items and find coincidences. TasTicWiki obtains information about the user interactions with the items by analyzing the searches, readings and editions in a wiki (Ruiz- Montiel, Molina-Castro, & Aldana-Montes, 2010). This informa- tion is used to calculate the satisfaction degree that an article in the wiki has for a certain user. Another example is found in SPETA (García-Crespo et al., 2009), which maintains a social net- work profile of the user, so that the user’s contact data is taken into account in order to analyze the interactions between the users. Trust is another component that appears when dealing with social recommenders. It has been argued that ratings of credible users should be treated with higher weights than oth- ers (Gavalas & Kenteris, 2011). � A special characteristic in tourism, which distinguishes it from other domains in which recommenders have been applied, is the mobility of the users, which may need recommendations in different moments and in different places. For this reason, this particular type of recommender systems has started to incorporate context-aware techniques. The success of this approach is due to the widespread use of mobile devices, as introduced in Section 2.2. Many tourism recommenders run on phones, so the user’s location can be used to guide the filter- ing of the items to be shown (Kurata, 2011; Lamsfus et al., 2009; Yang, 2010). Not only the current location of the user is impor- tant, but also the places that have already been visited (Gavalas & Kenteris, 2011; Umanets et al. 2013). Other features that are considered as contextual information in tourism recommender systems are, for instance, the current weather to decide if it is more appropriate to recommend indoor or outdoor activities (Braunhofer et al. 2013; García-Crespo et al., 2009; Gavalas & Kenteris, 2011) or the motion speed and time to generate plans (Noguera et al., 2012). In the system described in Niaraki and Kim (2009) a complex model of the context is considered for constructing personalised route plans. The context information is organised on a hierarchy, including aspects related to the traf- fic, weather, safety (like telephone booth, side road parking, medical centre, etc.), facilities (gas station, etc.) and tourist attractions (fishing zone, recreation place, seaside, etc.). In Amato, Mazzeo, Moscato, and Picariello (2013b) four main parameters for the context are set: (i) time (time needed by the user to reach the place, the opening/closing times, etc.); (ii) location of the user and the place; (iii) weather and environ- mental conditions (e.g. temperature, humidity, rainfall degree, wind, season, moment of the day, etc.); (iv) social factors (num- ber of users close to the place and number of positive/negative feedbacks). Moreover, the same authors extended the work (Amato, Mazzeo, Moscato, & Picariello, 2013a) to indoor scenar- ios to analyze room crowd, room fitness, network performances, location and time interval. They use a pre-filtering strategy to select those alternatives that satisfy the user’s needs and a post-filtering strategy to arrange the recommended items based on their contextual values. This dimension is devised as a cru- cial point in the success of recommender systems in tourism, due to the inherent mobile behaviour of the users in this specific application domain. Acknowledgements This work has been supported by the Spanish research projects DAMASK (Data Mining Algorithms for Semantic Knowledge, TIN2009-11005), SHADE (Semantic and Hierarchical Attributes in Decision Making, TIN2012-34369), and the collaboration agreement between the ITAKA research group and the Science & Technology Park for Tourism and Leisure in the area of technological applications for Tourism mobility management. References Adomavicius, G., & Tuzhilin, A. (2005). Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17, 734–749. Albanese, M., Chianese, A., d’Acierno, A., Moscato, V., & Picariello, A. (2010). A multimedia recommender integrating object features and user behavior. Multimedia Tools and Applications, 50(3), 563–585. Albanese, M., d’Acierno, A., Moscato, V., Persia, F., & Picariello, A. (2013). A multimedia recommender system. ACM Transactions on Internet Technology (TOIT), 13(1), 3. Alonso, K., Zorrilla, M., Confalonieri, R., Vázquez-Salceda, J., Iñana, H., Palau, M., Calle, J., & Castro, E. (2012). Ontology-based tourism for all: Recommender and information retrieval system for interactive community displays. In Proceedings of eighth international conference on information science and digital content technology (pp. 650–655). Amato, F., Mazzeo, A., Moscato, V., & Picariello, A. (2013b). Exploiting cloud technologies and context information for recommending touristic paths. In Intelligent distributed computing VII – proceedings of the seventh international symposium on intelligent distributed computing, IDC 2013, Prague, Czech Republic (pp. 281–287). Amato, F., Mazzeo, A., Moscato, V., & Picariello, A. (2013a). A recommendation system for browsing of multimedia collections in the internet of things. In Internet of things and inter-cooperative computational technologies for collective intelligence (pp. 391–411). Berlin, Heidelberg: Springer. Batet, M., Moreno, A., Sánchez, D., Isern, D., & Valls, A. (2012). Turist@: Agent-based personalised recommendation of touristic activities. Expert Systems with Applications, 39, 7319–7329. Berka, T., & Plößnig, M. (2004). Designing recommender systems for tourism. In Proceedings of the 11th international conference on information technology in travel & tourism (ENTER’2004). Borràs, J., de la Flor, J., Pérez, Y., Moreno, A., Valls, A., Isern, D., et al. (2011). SigTur/E- destination: A system for the management of complex tourist regions. In R. Law, M. Fuchs, & F. Ricci (Eds.), Information and communication technologies in tourism 2011 (pp. 39–50). Springer. Borràs, J., Moreno, A., Valls, A., Ferré, M., Ciurana, E., Salvat, J., Russo, A., & Anton- Clavé, S. (2012a). Uso de técnicas de inteligencia artificial para hacer recomendaciones enoturísticas personalizadas en la Provincia de Tarragona. In IX Congreso turismo y tecnologías de la información y las comunicaciones, Málaga, Spain (pp. 217–230). ISBN: 978-84-615-9946-2. Borràs, J., Valls, A., Moreno, A., & Isern, D. (2012b). Ontology-based management of uncertain preferences in user profiles. In S. Greco, B. Bouchon-Menier, B. Bouchon-Menier, G. Colletti, M. Fedrizzi, B. Matarazzo, & R. Yager (Eds.), Information processing and management of uncertainty in knowledge-based systems part II. Communications in computer and information science (CCIS) (Vol. 298, pp. 127–136). Springer-Verlag. Braunhofer, M., Elahi, M., Ricci, F., & Schievenin, T. (2013). Context-aware points of interest suggestion with dynamic weather data management. In Information and communication technologies in tourism 2014 (pp. 87–100). Springer International Publishing. Brewer, J. D. (1984). Tourism and ethnic stereotypes: Variations in a Mexican town. Annals of Tourism Research, 11(3), 487–510. Buhalis, D. (2003). ETourism: Information technology for strategic tourism management. Englewood Cliffs: Prentice Hall. Burke, R. (2002). Hybrid recommender systems: Survey and experiments. User Modeling and User-Adapted Interaction, 12(4), 331–370. Castillo, L., Armengol, E., Onaindía, E., Sebastiá, L., González-Boticario, J., Rodríguez, A., et al. (2008). Samap: An user-oriented adaptive system for planning tourist visits. Expert Systems with Applications, 34, 1318–1332. Ceccaroni, L., Codina, V., Palau, M., & Pous, M. (2009). PaTac: Urban, ubiquitous, personalized services for citizens and tourists. In Third international conference on the digital society (ICDS 2009), February 1–7, 2009, Cancun, México (pp. 7–12). Codina, V., & Ceccaroni, L. (2010). Taking advantage of semantics in recommendation systems. In R. Alquézar, A. Moreno, & J. Aguilar (Eds.) (pp. 163–172). Amsterdam, The Netherlands, The Netherlands: IOS Press. Coelho, B., Martins, C., & Almeida, A. (2009). Adaptive tourism modeling and socialization system. In CSE ‘09 proceedings of the 2009 international conference on computational science and engineering (Vol. 4, pp. 645–652). Dasarathy, B. (1991). Nearest neighbour (NN) norms: NN pattern classification techniques. IEEE Computer Society Press. Dey, A., & Abowd, G. (1999). Towards a better understanding of context and context-awareness. In HUC ’99: Proceedings of the first international symposium on handheld and ubiquitous computing. Dujmovic, J., & Nagashima, H. (2006). LSP method and its use for evaluation of java IDEs. International Journal of Approximate Reasoning, 41, 3–22. Felfernig, A., Gordea, S., Jannach, D., Teppan, E., & Zanker, M. (2007). A short survey of recommendation technologies in travel and tourism. OEGAI Journal, 25(7), 17–22. http://refhub.elsevier.com/S0957-4174(14)00343-1/h0005 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0005 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0005 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0010 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0010 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0010 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0015 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0015 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0015 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0025 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0025 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0025 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0025 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0035 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0035 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0035 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0045 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0045 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0045 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0045 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0535 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0535 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0535 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0535 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0535 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0535 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0490 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0490 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0490 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0490 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0065 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0065 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0070 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0070 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0075 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0075 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0080 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0080 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0080 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0090 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0090 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0090 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0100 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0100 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0110 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0110 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0115 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0115 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0115 7388 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 Fenza, G., Fischetti, E., Furno, D., & Loia, V. (2011). A hybrid context aware system for tourist guidance based on collaborative filtering. In Proceedings of IEEE int. conference on fuzzy systems (pp. 131–138). Gao, M., Liu, K., & Wu, Z. (2010). Personalisation in web computing and informatics: Theories, techniques, applications, and future research. Information Systems Frontiers, 12, 607–629. Garcia, A., Arbelaitz, O., Linaza, M. T., Vansteenwegen, P., & Souffriau, W. (2010). Personalized tourist route generation. In F. Daniel, F. M. Facca (Eds.), 10th International conference on web engineering 2010, ICWE2010: Vol. 6385. LNCS (pp. 486–497). Garcia, I., Sebastià, L., & Onaindia, E. (2011). On the design of individual and group recommender systems for tourism. Expert Systems with Applications, 38, 7683–7692. Garcia, A., Torre, I., & Linaza, M. T. (2013). Mobile social travel recommender system. In Information and communication technologies in tourism 2014 (pp. 3–16). Springer International Publishing. Garcia, A., Vansteenwegen, P., Arbelaitz, O., Souffriau, W., & Linaza, M. T. (2013). Integrating public transportation in personalised electronic tourist guides. Computers & Operations Research, Special issue: Transport Scheduling, 3, 758–774. García-Crespo, A., Chamizo, J., Rivera, I., Mencke, M., Colomo-Palacios, R., & Gómez- Berbís, J. M. (2009). SPETA: Social pervasive e-tourism advisor. Telematics and Informatics, 26, 306–315. García-Crespo, A., López-Cuadrado, J. L., Colomo-Palacios, R., González-Carrasco, I., & Ruiz-Mezcua, B. (2011). Sem-fit: A semantic based expert system to provide recommendations in the tourism domain. Expert Systems with Applications, 38(10), 13310–13319. Gavalas, D., & Kenteris, M. (2011). A web-based pervasive recommendation system for mobile tourist guides. Personal and Ubiquituous Computing, 15, 759–770. Gretzel, U. (2011). Intelligent systems in tourism: A social science perspective. Annals of Tourism Research, 38(3), 757–779. Gyorodi, R., Gyorodi, C., & Dersidan, M. (2013). An extended recommendation system using data mining implemented for smart phones. International Journal of Computers and Technology, 11(3), 2360–2372. Hsu, F. M., Lin, Y. T., & Ho, T. K. (2012). Design and implementation of an intelligent recommendation system for tourist attractions: The integration of EBM model, Bayesian networks and google maps. Expert Systems with Applications, 39, 3257–3264. Huang, Y., & Bian, L. (2009). A Bayesian network and analytic hierarchy process based personalized recommendations for tourist attractions over the internet. Expert Systems with Applications, 36(1), 933–943. Jannach, D., Zanker, M., & Jessenitschnig, M. (2010). Developing knowledge-based travel advisor systems: A case study. In N. Sharda (Ed.), Tourism informatics: Visual travel recommender systems, social communities, and user interface design (pp. 38–53). Hershey, PA: Information Science Reference. Kabassi, K. (2010). Personalizing recommendations for tourists. Telematics and Informatics, 27(1), 51–66. Koceski, S., & Petrevska, B. (2012). Empirical evidence of contribution to e-tourism by application of personalized tourism recommendation system. Annals of the Alexandru Ioan Cuza University – Economics, 59(1), 363–374. Kurata, Y., & Hara, T. (2013). CT-planner4: Toward a more user-friendly interactive day-tour planner. In Information and communication technologies in tourism 2014 (pp. 73–86). Springer International Publishing. Kurata, Y. (2011). CT-planner2: More flexible and interactive assistance for day tour planning. ENTER 2011. In Information and communication technologies in tourism (pp. 25–37). Innsbruck, Austria. Lamsfus, C., Alzua-Sorzabal, A., Martin, D., & Smithers, T. (2011). An evaluation of a contextual approach to visitor information system. In R. Law, M. Fuchs, & F. Ricci (Eds.), Proceeding of the ENTER conference (pp. 191–202). Innsbruck, Austria: Springer-Verlag (pp.179–189). Lamsfus, C., Alzua-Sorzabal, A., Martin, D., Salvador, Z., & Usandizaga, A. (2009). Human-centric ontology-based context modelling in tourism. In Proceedings of the international conference on knowledge engineering and ontology development, Funchal, Madeira, Portugal, October 6–8, 2009 (pp. 424–434). Lee, C. S., Chang, Y. C., & Wang, M. H. (2009). Ontological recommendation multi- agent for Tainan city travel. Expert Systems with Applications, 36, 6740–6753. Linaza, M. T., Aguirregoikoa, A., García, A., Torres, J. I., & Aranburu, K. (2011). Image- based travel recommender system for small tourist destinations. In Information and communication technologies in tourism 2011 (pp. 1–12). Innsbruck, Austria. Lorenzi, F., Loh, S., & Abel, M. (2011). PersonalTour: A recommender system for travel packages. In Proceedings of the 2011 IEEE/WIC/ACM international conference on intelligent agent technology, IAT 2011, Lyon, France, August 22–27. IEEE/WIC/ACM international conferences on web intelligence and intelligent agent technology. Luberg, A., Tammet, T., & Järv, P. (2011). Smart city: A rule-based tourist recommendation system. In R. Law, M. Fuchs, & F. Ricci (Eds.), Information and communication technologies in tourism 2011. ENTER 11, January 26th–28th 2011, Innsbruck, Austria. Springer-Verlag. Lucas, J., Laurent, A., Moreno, M., & Teisseire, M. (2009). Fuzz-CBA: Classification à base de règles d’association floues et systèmes de recommandation. Rencontres Francophones sur la Logique Floue et ses Applications, 283–290. Lucas, J., Luz, N., Moreno, M., Anacleto, R., Figueiredo, A., & Martins, C. (2013). A hybrid recommendation approach for a tourism system. Expert Systems with Applications, 40(9), 3532–3550. Manouselis, N., & Costopoulou, C. (2007). Analysis and classification of multi- criteria recommender systems. World Wide Web, 10(4), 415–441. Marques, S. S. (2009). Fabrication of image-destination: The impact of stereotypes at the level of tourism. Revista Turismo & Desenvolvimento, 12. Martin, D., Alzua, A., & Lamsfus, C. (2011). A contextual geofencing mobile tourism service. In R. Law, M. Fuchs, & F. Ricci (Eds.), Proceeding of the ENTER conference (pp. 191–202). Innsbruck, Austria: Springer-Verlag. Martínez, L., Rodríguez, R. M., & Espinilla, M. (2009). REJA: A georeferenced hybrid recommender system for restaurants. In Proceedings of IEEE/WIC/ACM international conference on web intelligence and intelligent agent technology (pp. 187–190). Martínez-Santiago, F., Ariza-López, F., Montejo-Ráez, A., & Ureña-López, A. (2012). GeOasis: A knowledge-based geo-referenced tourist assistant. Expert Systems with Applications, 39(14), 11737–11745. Meehan, K., Lunney, T., Curran, K., & McCaughey, A. (2013). Context-aware intelligent recommendation system for tourism. In Proceedings of the 11th IEEE international conference on pervasive computing and communications (pp. 328–331). Mínguez, I., Berrueta, D., & Polo, L. (2010). CRUZAR: An application of semantic matchmaking to e-tourism. In Y. Kalfoglou (Ed.), Cases on semantic interoperability for information systems integration: Practices and applications (pp. 255–271). Hershey, PA: Information Science Reference. Montaner, M., López, B., & de la Rosa, J. L. (2003). A taxonomy of recommender agents on the internet. Artificial Intelligence Review, 19(3), 285–330. Montejo-Ráez, A., Perea-Ortega, J. M., García-Cumbreras, M. A., & Martínez- Santiago, F. (2011). Otium: A web based planner for tourism and leisure. Expert Systems with Applications, 38, 10085–10093. Moreno, A., Valls, A., Isern, D., Marin, L., & Borràs, J. (2013). SigTur/e-destination: Ontology-based personalized recommendation of tourism and leisure activities. Engineering Applications of Artificial Intelligence, 26, 633–651. Moscato, V., Picariello, A., & Rinaldi, A. M. (2013). Towards a user based recommendation strategy for digital ecosystems. Knowledge-Based Systems, 37, 165–175. Niaraki, A. S., & Kim, K. (2009). Ontology based personalized route planning system using a multi-criteria decision making approach. Expert Systems with Applications, 36, 2250–2259. Noel, J., Sanner, S., Tan, K. -N., Christen, P., Xie, L., Bonilla, E. V., Abbasnejad, E., & Della Penna, N. (2012). New objective functions for social collaborative filtering. In Proceedings of the 21st int. conf. on world wide web (pp. 859–868). Noguera, J. M., Barranco, M. J., Segura, R. J., & Martínez, L. (2012). A mobile 3D-GIS hybrid recommender system for tourism. Information Sciences, 215, 37–52. Pearl, J. (1988). Probabilistic reasoning in intelligent systems. Networks of plausible inference. Morgan Kaufmann. Pinho, J., da silva, B., Moreno, M., de Almeida, A. M., & Lopes, C. (2011). Applying recommender methodologies in Tourism sector. Highlights in practical applications of agents and multiagent systems. In Advances in Intelligent and Soft Computing (pp. 101–108). Springer. Rey-López, M., Barragáns-Martínez, A. B., Peleteiro, A., Mikic-Fonte, F. A., & Burguillo, J. C. (2011). moreTourism: Mobile recommendations for tourism. In IEEE int. conf. on consumer electronics (pp. 347–348). Ricci, F. (2002). Travel recommender systems. IEEE Intelligent Systems, 17(6), 55–57. Ricci, F., Nguyen, Q. N., & Averjanova, O. (2010). Exploiting a map-based interface in conversational recommender systems for mobile travelers. In N. Sharda (Ed.), Tourism informatics: Visual travel recommender systems, social communities, and user interface design (pp. 73–93). Hershey, PA: Information Science Reference. Rojas, G., & Uribe, C. (2013). A conceptual framework to develop mobile recommender systems of points of interest. In International workshop on advanced software engineering. Proceedings jornadas chilenas de computación, Chile, November 2013. Ruíz-Martínez, J. M., Miñarro, J. A., Castellanos, D., García, F., & Valencia, R. (2011). Ontology population: An application for the e-tourism domain. International Journal of Innovative Computing, Information and Control, 7(119), 1–19. Ruiz-Montiel, M., & Aldana-Montes, J. (2009). Semantically enhanced recommender systems. On the move to meaningful internet systems: OTM 2009 workshops (Vol. 5872, pp. 604–609). Springer. Ruiz-Montiel, M., Molina-Castro, J. J., & Aldana-Montes, J. F. (2010). TasTicWiki: A semantic Wiki with content recommendation. In Proceedings of the fifth workshop on semantic wikis-linking data and people. Seventh extended semantic web conference (pp. 31–40). Ruotsalo, T., Haav, K., Stoyanov, A., Rochee, S., Fanid, E., Deliaic, R., et al. (2013). SMARTMUSEUM: A mobile recommender system for the web of data. Web semantics: Science, services and agents on the world wide web, 20, 50–67. Sánchez, D., & Moreno, A. (2008a). Pattern-based automatic taxonomy learning from the web. AI Communications, 21(1), 27–48. Sánchez, D., & Moreno, A. (2008b). Learning non-taxonomic relationships from web documents for domain ontology construction. Data and Knowledge Engineering Journal, 64, 600–623. Sánchez, D., Moreno, A., & Del Vasto-Terrientes, L. (2012). Learning relation axioms from text: An automatic web-based approach. Expert Systems with Applications, 39, 5792–5805. Savir, A., Brafman, R., & Shani, G. (2013). Recommending improved configurations for complex objects with an application in travel planning. In Proceedings of the seventh ACM conference on recommender systems (pp. 391–394). ACM. Sebastià, L., Garcia, I., Onaindia, E., & Guzman, C. (2009). E-tourism: A tourist recommendation and planning application. International Journal on Artificial Intelligence Tools, 18(5), 717–738. http://refhub.elsevier.com/S0957-4174(14)00343-1/h0125 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0125 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0125 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0155 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0155 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0155 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0495 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0495 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0495 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0150 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0150 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0150 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0130 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0130 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0130 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0135 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0135 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0135 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0135 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0160 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0160 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0165 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0165 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0170 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0170 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0170 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0175 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0175 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0175 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0175 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0180 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0180 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0180 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0185 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0185 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0185 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0185 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0190 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0190 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0195 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0195 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0195 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0500 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0500 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0500 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0505 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0505 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0505 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0505 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0220 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0220 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0510 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0510 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0510 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0510 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0510 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0240 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0240 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0240 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0245 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0245 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0245 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0250 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0250 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0255 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0255 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0515 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0515 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0515 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0270 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0270 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0270 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0280 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0280 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0280 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0280 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0285 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0285 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0290 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0290 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0290 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0295 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0295 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0295 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0300 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0300 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0300 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0305 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0305 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0305 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0315 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0315 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0320 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0320 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0325 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0325 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0325 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0325 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0335 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0335 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0340 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0340 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0340 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0340 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0350 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0350 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0350 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0520 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0520 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0520 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0365 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0365 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0365 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0370 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0370 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0375 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0375 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0375 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0380 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0380 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0380 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0525 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0525 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0525 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0390 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0390 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0390 J. Borràs et al. / Expert Systems with Applications 41 (2014) 7370–7389 7389 Sebastià, L., Giret, A., & Garcia, I. (2010). A multi agent architecture for Tourism recommendation. Trends in practical applications of agents and multiagent systems. Advances in Intelligent and Soft Computing, 71, 547–554. Seidel, I., Gärtner, M., Pöttler, M., Berger, H., & Dittenbach, M. (2009). Itchy feet: A 3D e-tourism environment. In N. Sharda (Ed.), Tourism informatics: Visual travel recommender systems, social communities, and user interface design. Hershey, PA: IGI Global. Sieg, A., Mobasher, B., & Burke, R. (2007). Learning ontology-based user profiles: A semantic approach to personalized web search. IEEE Intelligent Informatics Bulletin, 8(1), 7–18. Souffriau, W., Vansteenwegen, P., Vanden Berghe, G., & Van Oudheusden, D. (2011). The planning of cycle trips in the province of East Flanders. Omega, 39, 209–213. Staab, S., & Werthner, H. (2002). Intelligent systems for tourism. IEEE Intelligent Systems, 17(6), 53–64. Tsung-Chiung, W., Chyong-Ru, L., & Wan-Chen, Y. (2012). Segmenting indigenous tourists from a serious leisure perspective. Journal of Vacation Marketing, 18(1), 17–29. Umanets, A., Ferreira, A., & Leite, N. (2013). GUIDEME – A tourist guide with a recommender system and social interaction. In Conference on electronics, telecommunications and computer (CETC), Lisboa, December 2013. Valls, A. (2003).ClusDM: A multiple criteria decision method for heterogeneous data sets (Ph.D. thesis). AI Communications 16 (pp. 129–130). Vansteenwegen, P., & Souffriau, W. (2011). Trip planning functionalities: State-of- the-art and future. Information Technology and Tourism, 12(4), 305–315. Vansteenwegen, P., Souffriau, W., Vanden Berghe, G., & Van Oudheusden, D. (2010). The city trip planner: An expert system for tourists. Expert Systems with Applications, 38(6), 6540–6546. Venkataiaha, S., Shardaa, N., & Ponnadaa, M. (2008). A comparative study of continuous and discrete visualisation of tourism information. ENTER 2008 conference, international federation for information technology and travel & tourism (IFITT), Innsbruck, Austria, 23–25 January 2008. Vicient, C., Sánchez, D., & Moreno, A. (2013). An automatic approach for ontology- based feature extraction from heterogeneous textual resources. Engineering Applications of Artificial Intelligence, 26(3), 1093–1106. Wang, W., Zeng, G., & Tang, D. (2011). Bayesian intelligent semantic mashup for tourism. Concurrency and Computation: Practice and Experience, 23, 850–862. Werthner, H. (2003). Intelligent systems in travel and tourism. In IJCAI (Vol. 3, pp. 1620–1625). Wooldridge, M. (2009). An introduction to multiagent systems. John Wiley and Sons. Yager, R. (1988). On ordered weighted averaging operators. Operators in multi- criteria decision making. IEEE Transactions on Systems, Man and Cybernetics, 18, 183–190. Yang, W. S., & Hwang, S. Y. (2013). ITravel: A recommender system in mobile peer- to-peer environment. The Journal of Systems and Software, 86(1), 12–20. Yang, Y. (2010). Semantic user model inferences for travel recommender systems. In N. Sharda (Ed.), Tourism informatics: Visual travel recommender systems, social communities, and user interface design (pp. 23–37). Hershey, PA: Information Science Reference. Yu, C. & Chang, H. (2009). Personalized location-based recommendation services for tour planning in mobile tourism applications. In Proceedings of the 10th international conference on e-commerce and web technologies, Linz, Austria, September 1–4: Vol. 5692. Lecture notes in computer science (pp. 1–49). http://refhub.elsevier.com/S0957-4174(14)00343-1/h0395 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0395 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0395 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0400 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0400 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0400 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0400 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0405 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0405 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0405 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0410 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0410 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0415 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0415 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0420 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0420 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0420 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0440 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0440 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0435 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0435 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0435 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0450 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0450 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0450 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0455 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0455 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0465 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0470 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0470 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0470 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0480 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0480 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0475 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0475 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0475 http://refhub.elsevier.com/S0957-4174(14)00343-1/h0475 Intelligent tourism recommender systems: A survey 1 Introduction 2 Interface 2.1 Web-based recommenders 2.2 Mobile recommendations 3 Functionalities 3.1 Travel destination and tourist packs 3.2 Ranked list of suggested attractions 3.3 Planning a route 3.4 Social aspects 4 Recommendation techniques in e-Tourism 4.1 Approaches to recommendation 4.2 Representation of the user profile 5 Use of AI techniques in tourism recommender systems 5.1 Multi-agent systems 5.2 Optimization techniques 5.3 Automatic clustering 5.4 Management of uncertainty 5.5 Knowledge representation 6 Related reviews 7 Conclusions 7.1 Summary and guidelines to develop tourism recommenders 7.2 Lines of future work Acknowledgements References 
work_l7g5umz4bnfeflcudj6itwakna	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy A multi-agent system to construct production orders by employing an expert system and a neural network Omar López-Ortega *, Israel Villar-Medina Universidad Autónoma del Estado de Hidalgo, Instituto de Ciencias Básicas e Ingenierı́a, Centro de Investigación en Tecnologı́as de Información y Sistemas, Carr. Pachuca-Tulancingo Km. 4.5 Pachuca, Hidalgo, C. P. 42083, México Abstract The authors describe the implementation of a multi-agent system, whose goal is to enhance production planning i.e. to improve the construction of production orders. This task has been carried out traditionally by the module known as production activity control (PAC). However, classic PAC systems lack adaptive techniques and intelligent behaviour. As a result they are mostly unfit to handle the NP Hard combinatorial problem underlying the construction of right production orders. To overcome this situation, we illustrate how an intelligent and collaborative multi-agent system (MAS) obtains a correct production order by coordinating two different tech- niques to emulate intelligence. One technique is performed by a feed-forward neural network (FANN), which is embedded in a machine agent, the objective being to determine the appropriate machine in order to fulfil clients’ requirements. Also, an expert system is provided to a tool agent, which in turn is in charge of inferring the right tooling. The entire MAS consists of a coordinator, a spy, and a scheduler. The coordinator agent has the responsibility to control the flow of messages among the agents, whereas the spy agent is constantly reading the Enterprise Information System. The scheduler agent programs the production orders. We achieve a realistic MAS that fully auto- mates the construction and dispatch of valid production orders in a factory dedicated to produce labels. � 2008 Elsevier Ltd. All rights reserved. Keywords: Multi-agent systems; Production planning; Expert systems; Artificial neural networks 1. Introduction Manufacturing companies, regardless their size, com- pete fiercely in the current globalised marketplace. There- fore, supporting software for manufacturing activities must help firms achieving flexibility and adaptation to changes. Production management, particularly, is sup- ported by a system known as production activity control (PAC), which links planning and manufacturing together (Browne, Harhen, & Shivnan, 1992). This union is materi- alized by the generation of production orders addressing values for available machines and right tooling, among other variables. However, as input data is not necessarily homogeneous, classical PAC modules are mostly unfit to deal with the exponential number of combinations to gen- erate the correct production order. Hence, flexible approaches open immense opportunities to improve this manufacturing activity, as reported in Wang, Yung, and Ip (2005), where a genetic algorithm has been employed to schedule orders in a dispersed manufacturing setting. It has been claimed that multi-agent systems (MAS) have emerged as the major swift to enhance performance in manufacturing: Authors such as Mônch and Stehli (2006) report that key activities of industrial enterprises are being performed by multi-agent systems. Such is the case of process planning (López-Morales & López-Ortega, 2005; López-Ortega & López-Morales, 2006), holonic con- trol, and production management (Marı́k & Lazanský, in press), which have already incorporated agent technology. The addition of capabilities to emulate intelligence in agents is no longer an aspiration but a reality, although some topics regarding cognitive abilities are left to be 0957-4174/$ - see front matter � 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2008.01.070 * Corresponding author. E-mail addresses: lopezo@uaeh.reduaeh.mx, olopez27@prodigy.- net.mx (O. López-Ortega), villar_israel@yahoo.com (I. Villar-Medina). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 36 (2009) 2937–2946 Expert Systems with Applications Author's personal copy explored. Such is the case for algorithms that mirror the supervised learning process, which is achieved primarily through feed-forward artificial neural networks (FANN), with the backpropagation algorithm. Even though this technique provides flexible ways of making inferences, it has not been widely exploited in multi-agent systems to make production planning more flexible. Modelling and implementing a MAS that employs co- ordinately learning and decision making capabilities, becomes a great challenge to provide flexible mechanisms to cope with changing market conditions. In this setting, our main contribution is the design and implementation of a multi-agent system that coordinates an expert system and a FANN, whose global is to construct production orders, hence improving the key activity of production planning. The article is organized as it is explained next. A brief analysis of the related work is presented in Section 2. Sec- tion 3 contains the knowledge engineering process that was carried out. Next, we present the design and implementa- tion of the intelligent and collaborative system. Finally, conclusions and future research perspectives are outlined. 2. Related work We behold the continuous incorporation of agent tech- nology into manufacturing systems. Marı́k (op cit) coined the revealing term agentification of manufacturing systems to summarize this issue. Specifically, agent technology has been implemented to improve communication of data among supporting applications. An example is given by Feng (2005), who created a MAS to improve collaboration between design and manufacturing departments. Planning activities are improved by using AI techniques and agent- based systems. Kornienko, Kornienko, and Priese (2004) presents a MAS aimed at optimizing resource assignments by analyzing process plans. Wang and Shen (2003) tackles the process planning problem, and extends the results to the scheduling problem (Wang, Shen, & Hao, 2006), in an effort to optimize decisions before manufacturing execu- tion takes place. In (Feng, Stouffer, & Jurrens, 2005) a rule- based system is provided to agents to plan manufacturing processes. Frey, Nimis, Worn, and Lockemann (2003) implements a MAS to calculate the lot size, the operations to be performed, and the time-span between jobs. The information is fed to a simulator, which determines the best production plan according to the data provided by the MAS. Also, a multi-agent system including data mining techniques to set up the profile of resources in a supply chain is described thoroughly (Symeonidis, Kehagias, & Mitkas, 2003). On the other hand, neural networks and agents have been recently applied in production planning. Paternina- Arboleda and Das (2005) reports learning algorithms to schedule multiple products on a single server. The solution is validated via simulation. Fichtner et al. (2006) describes an unsupervised learning technique to determine the appropriate NC machine, having as input unknown design features coming from a CAD system. Thus, the agentification of production planning helps replacing centralized systems for distributed and more flex- ible architectures. Also, the incorporation of agent technol- ogy must be grounded on the ability to include intelligence. Although the related research found in literature is highly valuable, this present paper pioneers on how a MAS employs co-ordinately an Expert System and a FANN to better production planning. Our proposal is not only inno- vative, but it also shows practical benefits, because it has already been tested on a real-life setting. 3. Knowledge engineering 3.1. Part I. Analysis of the case study The environment is a factory dedicated to manufacture labels. Labels are used almost everywhere: On bottled products such as wine or sodas, candies, jeans, bar-coding, CDs, etc. They provide information about a given product. Consequently, their demand on the market place is high, yet prone to changes. Minor variations in size, colour, or raw material impact on the entire manufacturing system. Few companies have the capacity and expertise to produce them, and those that manufacture labels face tremendous production planning problems, because traditional plan- ning systems are not suitable to handle such a complex task. The typical flow of data in our case study is presented in Fig. 1. The labels are produced according to clients’ requirements. An external application is used by the sales department to acquire clients’ data, such as label design, colour, dimensions, material and specific attributes. These data are stored in the Enterprise Information System (EIS). Once the client’s order is processed by the sales department, the production planning department must elaborate a production order on which the information to produce the label is comprised. The production order must be accurate and error-free so that the manufacturing department can perform adequately. The possible combinations to produce a label are numerous. For example, the following raw material can be employed: glued-paper, non-glued paper, thermal paper, nylon, polyester, plastic, and cardboard to name but only a few. Then, the raw material is grouped into families to sim- plify the planning process and to determine how to acquire it, which can be in the form of continuous cylinders or in batches (also called master). Another complexity arises when a client sets the colours to be used (i.e. blue letters on white, glued-paper). Thus, the tints combination must be fixed according to the client’s requirements. The produc- tion planning manager is responsible for establishing what the operations (SU1, SU2 and SU3) that are to be per- formed to comply with the client’s design. On such infor- mation, the manager must determine the number of tools (up to three). Once these pieces of data are available, the 2938 O. López-Ortega, I. Villar-Medina / Expert Systems with Applications 36 (2009) 2937–2946 Author's personal copy production planning manager must provide the right machine (out of three options) on which the label will be produced. The machine depends on the following variables: raw material family, number of tints, resolution (dots per inch), type of finishing, and number of tools. This creates great difficulty to construct consistent production orders. Normally, the production planning manager relies on his/ her experience; however, production orders normally con- tain imprecise data, as the manager shows inconsistent behaviour about his/her decisions. The production man- ager either assigns the wrong number of tools for a given machine, or launches a production order for a machine that is not suitable to deal with a specific design require- ment. This situation might occur due to fatigue, or the inability of a human-being to manage larger number of combinations. Therefore, more accuracy and flexibility can be achieved by automasing the process of generating a production order. 3.2. Part II. The rule base The first decision to be made is to establish the number of tools, the raw material family, and the way to purchase such raw material. These decisions are possible to be achieved by setting a rule base. Therefore, the rule base receives the following data: length, width, raw material, and the content of three variables known as SU1, SU2, and SU3. In this way, the rule base provides the number of tools, based on the values of SU1, SU2 and SU3. This is exemplified by the following excerpt: ruleT1: IF SU1 ! = ‘‘‘‘none”” AND SU2 ! = ‘‘‘‘none”” AND SU3 ! = ‘‘‘‘none”” THEN number_tools = 3 ruleT2: IF SU1 = ‘‘‘‘none”” AND SU2 = ‘‘‘‘none”” AND SU3 = ‘‘none” THEN number_tools = 1 ruleT3: IF SU1 ! = ‘‘none” AND SU2 = ‘‘none” AND SU3 � = ‘‘none” THEN number_tools = 1 ruleT4: IF SU1 = ‘‘none” AND SU2 ! = ‘‘none” AND SU3 = ‘‘none” THEN number_tools = 1 ruleT5: IF SU1 = ‘‘none” AND SU2 = ‘‘none” AND SU3 ! = ‘‘none” THEN number_tools = 1 ruleT6: IF SU1 ! = ‘‘none” AND SU2 ! = ‘‘none” AND SU3 = ‘‘none” THEN number_tools = 2 ruleT7: IF SU1 ! = ‘‘none” AND SU2 = ‘‘none” AND SU3 ! = ‘‘none” THEN number_tools = 2 ruleT8: IF SU1 = ‘‘none” AND SU2 ! = ‘‘none” AND SU3 ! = ‘‘none” THEN number_tools = 2 Based on the raw material, the raw material family is estab- lished. Some rules read: ruleRMF01: IF RAW_MATERIAL = ‘‘kimno” THEN RAW_MATERIAL_FAMILY = ‘‘kimdura” ruleRMF02: IF RAW_MATERIAL = ‘‘kimsi” THEN RAW_MATERIAL_FAMILY = ‘‘kimdura” ruleRMF03: IF RAW_MATERIAL = ‘‘kimte” THEN RAW_MATERIAL_FAMILY = ‘‘kimdura” ruleRMF04: IF RAW_MATERIAL = ‘‘valer” THEN RAW_MATERIAL_FAMILY = ‘‘fabric” ruleRMF05: IF RAW_MATERIAL = ‘‘poliester” THEN RAW_MATERIAL_FAMILY = ‘‘fabric” Fig. 1. Illustration of the case study. O. López-Ortega, I. Villar-Medina / Expert Systems with Applications 36 (2009) 2937–2946 2939 Author's personal copy ruleRMF06: IF RAW_MATERIAL = ‘‘nylon” THEN RAW_MATERIAL_FAMILY = ‘‘fabric” ruleRMF07: IF RAW_MATERIAL = ‘‘auttr” THEN RAW_MATERIAL_FAMILY = ‘‘paper” ruleRMF08: IF RAW_MATERIAL = ‘‘orono” THEN RAW_MATERIAL_FAMILY = ‘‘paper” ruleRMF09: IF RAW_MATERIAL = ‘‘oroag” THEN RAW_MATERIAL_FAMILY = ‘‘paper” ruleRMF10: IF RAW_MATERIAL = ‘‘cartd” THEN RAW_MATERIAL_FAMILY = ‘‘cardboard” ruleRMF11: IF RAW_MATERIAL = ‘‘carts” THEN RAW_MATERIAL_FAMILY = ‘‘cardboard” ruleRMF12: IF RAW_MATERIAL = ‘‘bopbc” THEN RAW_MATERIAL_FAMILY = ‘‘bopp” ruleRMF13: IF RAW_MATERIAL = ‘‘boppm” THEN RAW_MATERIAL_FAMILY = ‘‘bopp” ruleRMF14: IF RAW_MATERIAL = ‘‘boptr” THEN RAW_MATERIAL_FAMILY = ‘‘bopp” Then, the type of purchase is set. Should raw material be purchased on continuous cylinders or on batches, it all depends on the raw material family. An excerpt of the rules to decide what type of purchase must be made, follows here: ruleTOP1: IF RAW_MATERIAL_FAMILY = ‘‘fab- ric”AND RAW_MATERIAL ! = ‘‘nylon” THEN TYPE_OF_PURCHSE = ‘‘master” ruleTOP2: IF RAW_MATERIAL_FAMILY = ‘‘fab- ric” AND RAW_MATERIAL = ‘‘nylon” THEN TYPE-OF_PURCHASE = ‘‘continuous cylinder” ruleTOP3: IF RAW_MATERIAL_FAMILY = ‘‘card- board” AND RAW_MATERIAL ! = ‘‘cartd” THEN TYPE_OF_PURCHASE = ‘‘master” ruleTOP4: IF RAW_MATERIAL_FAMILY = ‘‘card- board” AND RAW_MATERIAL = ‘‘cartd” THEN TYPE_OF_PURCHASE = ‘‘continuous cylinder” ruleTOP5: IF RAW_MATERIAL_FAMILY = ‘‘kimdura” THEN TYPE_OF_PURCHASE = ‘‘master” ruleTOP6: IF RAW_MATERIAL_FAMILY = ‘‘various” THEN TYPE_OF_PURCHASE = ‘‘master” ruleTOP7: IF RAW_MATERIAL_FAMILY = ‘‘paper” AND RAW_MATERIAL ! = ‘‘dual” THEN TYPE_OF_PURCHASE = ‘‘master” ruleTOP8: IF RAW_MATERIAL_FAMILY = ‘‘paper” AND RAW_MATERIAL = ‘‘dual” THEN TYPE_OF_PURCHASE = ‘‘continuous_cylinder” ruleTOP9: IF RAW_MATERIAL_FAMILY = ‘‘bopp” THEN TYPE_OF_PURCHASE = ‘‘master” Consequently, one agent must possess the previous knowledge, receive input data, and provide values for the number of tools, the way to purchase raw material, and the raw material family. This agent is called tool agent (see Section 4 for more details). The purchasing depart- ment is informed about the type of purchase, whereas the number of tools and the raw material family are used along with other data to determine the machine that will be in charge of producing the label. 3.3. Part III. The FANN and the machine We provide details of the feed-forward artificial neural network to determine the right machine. As it has been sta- ted before, obtaining the appropriate machine depends on the combination of five different data sets, as follows: Raw Material FamilyX TintsX ResolutionX Finishing TypeX Number Tools� > Machine Each input data set contains the following elements: Raw Material Family ¼ fFabric; Paper; Kimdura; Cardboard; BOPP; Otherg Number Tools ¼f1; 2; 3g Tints ¼f1; 2; 3; 4; 5; 6g Resolution ¼fnotspecified; 280; 360; 400; 500; 600; 700; 800g Finishing Type ¼fnotspecified; laminated; sulfatedg The output data set is formed by Machine ¼f830; 2200; Nilpeterg The raw material family and the number of tools are obtained by the rule base previously described; the rest of the input variables are provided by the planning depart- ment based on the client’s specifications. The entire range of possibilities to determine a right machine represents an NP-Hard combinatorial problem. We chose the feed-forward artificial neural network archi- tecture because such neural networks are known to be good classifiers. They are trained by backpropagating the error signal, based on the gradient descent technique: the error value diminishes as the gradient of a given function i.e. the logistic function, is calculated (Jang, Sun, & Mizutani, 1997). In each training pass, the weights of the net are updated. When the error value is less than a given target value, the training phase is complete and the resultant weights are stored. A fiable training matrix is employed to train the feed-forward artificial neural network (Fig. 2), which is a subset of the entire range of possibilities to set a machine. The actual configuration of the FANN is related to the training matrix. Specifically, one processing unit is created for each value contained in the training set. For example, 2940 O. López-Ortega, I. Villar-Medina / Expert Systems with Applications 36 (2009) 2937–2946 Author's personal copy as input 1 of the training matrix has six different values, six processing units are built. However, the actual value is rep- resented by a string, which is not a suitable input type for the FANN. Thus, such values are converted to a stream of 0’s and 1’s. Table 1 illustrates the codification for the raw material family. The prior codification is necessary because FANNs only handle values within the closed interval [0, 1]. Therefore, the actual input and output values that the net receives and obtains are 0’s and 1’s. This codification–decodifica- tion is done by the encoder class attached to the machine agent (see Fig. 5). Therefore, the FANN has six processing units in the input layer, which are in charge of dealing exclusively with the raw material family. The totality of dis- crete values contained in the input and output sets were codified in a similar way. Tables 2–6 show the resultant codification. Consequently, the number of processing units in the input layer of the FANN equals the number of codified input values. For this case, 25 processing units in the input Fig. 2. The training matrix. Table 1 Codification of the raw material family set Input stream Raw material family 0 0 0 0 0 1 Paper 0 0 0 0 1 0 Fabric 0 0 0 1 0 0 Kimdura 0 0 1 0 0 0 Cardboard 0 1 0 0 0 0 BOPP 1 0 0 0 0 0 Other O. López-Ortega, I. Villar-Medina / Expert Systems with Applications 36 (2009) 2937–2946 2941 Author's personal copy layer are fixed. Once the processing units for the input layer are generated, then the intermediate layer is set. The Java code assigns the same number of processing units in the input layer to the intermediate layer. Only one intermediate layer is created, since a FANN with a single intermediate layer is powerful enough to deal with non-linear problems. The output layer is then constructed. The number of processing units in the output layer also depends on the number of values that are provided as valid outcomes in the training set. In our training matrix, the output layer is given three different values, and three pro- cessing units are constructed. According to the previous codification, the resultant FANN consists of 3 layers, 25 processing units in both, the input and intermediate layers, and three processing units in the output layer. This is represented in Fig. 3. The processing units on the input layer receive streams of 0’s and 1’s, which result after codifying the incoming sym- bol. The reverse process is carried out for the output layer, where the processing units provide 0’s and 1’s, all of which Table 2 Codification of the tints set Input stream Number of tints 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 1 0 2 0 0 0 0 1 0 0 3 0 0 0 1 0 0 0 4 0 0 1 0 0 0 0 5 0 1 0 0 0 0 0 6 1 0 0 0 0 0 0 7 Table 3 Codification of the resolution set Input stream Resolution 0 0 0 0 0 0 0 1 Not specified 0 0 0 0 0 0 1 0 280 0 0 0 0 0 1 0 0 360 0 0 0 0 1 0 0 0 400 0 0 0 1 0 0 0 0 500 0 0 1 0 0 0 0 0 600 0 1 0 0 0 0 0 0 700 1 0 0 0 0 0 0 0 800 Table 4 Codification for the finishing type set Input stream Finishing type 0 0 Not specified 0 1 Laminated 1 0 Sulfated Table 5 Codification for the number of tools set Input stream Number of tools 0 0 1 0 1 2 1 0 3 Table 6 Codification of the machine set Input stream Machine 0 0 1 830 0 1 0 2200 1 0 0 Nilpeter Fig. 3. Configuration of the resultant FANN. 2942 O. López-Ortega, I. Villar-Medina / Expert Systems with Applications 36 (2009) 2937–2946 Author's personal copy are further converted to the symbol that represents a machine. Although there are countless simulators of neural net- works, none of them is suited to be used by a multi-agent system. Hence, it was necessary to incorporate supervised learning in the machine agent by means of Java program- ming. To comply rightly with the requirements to imple- ment backpropagation, a number of technological innovations were carried out, adapting the work of (Bigus & Bigus, 2002) on Java coding of backpropagation. Fig. 4 presents the module that we created to set training matrices for FANNs. 4. Design and implementation of the multi-agent system The primary goal of the MAS is to construct a produc- tion order. Our solution is to decompose the decision pro- cess carried out by the human manager in a series of tasks performed by software agents. These tasks are: 1. To read the EIS for newly created sales orders, and to acquire relevant variables. 2. To determine all the necessary tooling information. 3. To determine the correct machine. 4. To establish a sequence for launching production orders to the manufacturing department. 5. To maintain data consistency and robustness along the process. According to the previous requirements the general design of the multi-agent system is presented. The AUML (Bauer & Odell, 2005) has been employed extensively to model the MAS, whereas JADE is employed as the imple- mentation platform. The class diagram of the system is pre- sented in Fig. 5. The MAS has the following agents: � coordinator agent � machine agent � tool agent � spy agent � scheduler agent Fig. 4. The software module to train the FANN. Fig. 5. Structure of the MAS. O. López-Ortega, I. Villar-Medina / Expert Systems with Applications 36 (2009) 2937–2946 2943 Author's personal copy In order to actually implement intelligent agents on the JADE platform, external tools must be employed. In the proposed design both, the machine agent and the tool agent possess intelligent capabilities. The tool agent has a rule- based system to determine the right tooling, whereas, the machine agent incorporates a FANN. In the previous sec- tion we already discussed about the specifics of the rule base and the FANN. The coordinator agent is responsible for maintaining data consistency during the process by controlling the flow of messages. A spy agent must read the Enterprise Information System in order to acquire clients’ orders, and send the information to the coordinator. The machine agent obtains the appropriate machine, whereas the tool agent is responsible for providing the right tooling. As soon as the machine, tools and other data are established, a priority is assigned to the production order. To do so, we include a scheduler agent. When the production order gets the priority, it is then launched to the manufacturing department. The behaviours of the agents were set up properly in order to avoid interferences while the agents are exchanging information. Interested readers may con- sult (Bellifemine, Caire, & Greenwood, 2007) for insights on how to program agents on the JADE platform. In the following section we present the results achieved by the multi-agent system. 4.1. The MAS exemplified We illustrate the series of tasks and partial decisions achieved by the software agents. We deliberately use the command line to illustrate the entire process of message exchange between the agents. The process starts as soon as the sales department enters a client’s requirement. The spy agent, which possesses a cyclic behaviour, realizes that a new order has just entered the EIS, and it sends a series of messages to the coordinator, informing about the values of the following variables: In the example, the client demands labels made of a material called ‘‘BOPBC”, with a laminated finishing type, and a resolution of 360 dots per inch. The planning depart- ment feeds the following values to the EIS: Sales order: 595 Width: 110 Length: 1500 Raw material: BOPBC SU1: R107 SU2: ‘‘ ” SU3: ‘‘ ” Number of tints: 5 Resolution: 360 Type of finishing: laminated Fig. 6. Inferences and communication occurring inside the MAS. 2944 O. López-Ortega, I. Villar-Medina / Expert Systems with Applications 36 (2009) 2937–2946 Author's personal copy Fig. 6a shows how the spy agent informs the coordinator agent about the previous data, organized in the sales order 595. When the coordinator acknowledges reception of mes- sages from the spy, then it informs the tool agent regarding width, length, raw material, SU1, SU2 and SU3. The tool agent assesses that it has enough data to initiate its reasoning process (‘‘Agente herramienta con suficientes datos”), then it runs the expert system and obtains a con- clusion (Fig. 6a, bottom). The following rules are triggered according to the input data of the current example: ruleRMF12: IF RAW_MATERIAL = ‘‘bopbc” THEN RAW_MATERIAL_FAMILY = ‘‘bopp” ruleTOP9: IF RAW_MATERIAL_FAMILY = ‘‘bopp” THEN TYPE_OF_PURCHASE = ‘‘master” ruleT3: IF SU1 ! = ‘‘none” AND SU2 = ‘‘none” AND SU3 = ‘‘none” THEN number_tools = 1 Therefore, the tool agent reaches the next conclusion: RAW MATERIAL FAMILY: BOPP NUMBER OF TOOLS: 1 TYPE OF PURCHASE: Master (batch) The tool agent informs the coordinator agent about its conclusions. On the reception of messages from the tool agent, the coordinator sends the following data to the machine agent (Fig. 6b, bottom – ‘‘la cadena entrante es: BOPP 5 360 Laminado 1”): RAW MATERIAL FAMILY: BOPP NUMBER OF TINTS: 5 RESOLUTION: 360 TYPE OF FINISHING: laminated NUMBER OF TOOLS: 1 Based on this type of information, the FANN embedded within the machine agent, which is set into execution mode, obtains a valid machine, and sends a message to the coor- dinator (Fig. 6c – ‘‘Agente coordinador recibi mensaje MR = Nilpeter”). In this example, the obtained machine is known as ‘‘Nilpeter”, which is a valid machine for this combination of values. For this production order, the assigned priority is 1, since it was the only order set the day of executing this example. The bottom side of Fig. 6c summarizes the results. This process is repeated entirely every time the spy agent reads a new sales order. Thus, the construction and release of a valid production order is fully automized by the intelligent and collaborative sys- tem that we developed. 5. Conclusions and research perspectives Based on theoretical and practical results, we sustain that multi-agent systems represent a major swift in sup- porting systems for manufacturing. We contribute to agent-based production planning with the collaborative and intelligent MAS presented here. Marı́k (op cit) argues that the agentification process of the production planning department provides an elegant mechanism for system inte- gration, and supports the migration from centralized plan- ning towards distributed and flexible architectures. The MAS we developed contributes to achieve this flexibility. To the best of the authors’ knowledge, this is the first report on how to coordinate a rule-based system and super- vised learning, making more flexible the production plan- ning activity. Although production planning is improved by the col- laborative and intelligent system presented here, it is neces- sary to develop a solid scheduling policy to provide more realistic plans. Our scheduler agent employs a FIFO policy, even though it is the simplest ordering/queuing mechanism. We can asses that integrating planning and scheduling is not a trivial task, yet it opens up opportunities to actually close the loop between production planning and manufac- turing execution. The current configuration of the MAS that we developed enhances production planning, yet pro- duction control is not fully covered. Thus, a dynamic con- trol policy might be obtained when relevant variables, such as Work In Process (WIP) inventory, backorder penalty cost, setup time and cost, processing and transportation cost and time, or machine disruption time, are sent back to planning for achieving production control. The loop is closed when events within the manufacturing department, such as machine failure or a delay in the setup time, are communicated in real-time to the planning department. Therefore, communication must be bidirectional, from production planning to manufacturing execution and vice versa. We suggest to design a hierarchy of agents, or a hol- archy (Walter, Brennan, & Norrie, 2006), to accomplish agent-based PAC. Communication, intelligence, and the capability to integrate data from external applications are key features that must be exploited thoroughly. Achieving such a necessary feedback is an opportunity to continue research in the field of agent-based manufacturing. References Bauer, B., & Odell, J. (2005). UML 2.0 and agents: How to build agent- based systems with the new UML standard. Engineering Applications of Artificial Intelligence, 18, 141–157. Bellifemine, F. L., Caire, G., & Greenwood, D. (2007). Developing multi- agent systems with JADE. USA: John Wiley and Sons. Bigus, J., & Bigus, J. (2002). Constructing intelligent agents using java (2nd ed.). NY: John Wiley and Sons. Browne, J., Harhen, J., & Shivnan, J. (1992). Production management systems. A CIM perspective. UK: Addison – Wesley. Feng, S. C. (2005). Preliminary design and manufacturing planning integration using web-based intelligent agents. Journal of Intelligent Manufacturing, 16, 423–437. Feng, S. C., Stouffer, K. A., & Jurrens, K. K. (2005). Manufacturing planning and predictive process model integration using software agents. Advanced Engineering Informatics, 19, 135–142. Fichtner, D., Nestler, A., Dang, T., Schulze, A., Carlsten, U., Schreiber, S., et al. (2006). Use of agents and neural networks for acquisition and preparation of distributed NC information to support NC planning. O. López-Ortega, I. Villar-Medina / Expert Systems with Applications 36 (2009) 2937–2946 2945 Author's personal copy International Journal of Computer Integrated Manufacturing, 19, 581–592. Frey, D., Nimis, J., Worn, H., & Lockemann, P. (2003). Benchmarking and robust multi-agent production planning and control. Engineering applications of Artificial Intelligence, 16, 307–320. Jang, J.-S. R., Sun, C.-T., & Mizutani, E. (1997). Neuro-fuzzy and soft computing. A computational approach to learning and machine intelli- gence. USA: Prentice – Hall. Kornienko, S., Kornienko, O., & Priese, J. (2004). Application of multi-agent planning to the assignment problem. Computers in Industry, 54, 273–290. López-Morales, V., & López-Ortega, O. (2005). A distributed semantic network model for a collaborative intelligent system. Journal of Intelligent Manufacturing, 16, 515–525. López-Ortega, O., & López-Morales, V. (2006). Cognitive communication in a multi-agent system for distributed process planning. International Journal of Computer Applications in Technology, 26, 99–107. Marı́k, V., & Lazanský, J. (in press). Industrial applications of agent technology. Control Engineering Practice, doi: doi:10.1016/ j.conengprac.2006.10.001. Mônch, L., & Stehli, M. (2006). ManufAG: A multi-agent system framework for production control of complex manufacturing systems. Information Systems and Electronic Business. doi:10.1007/s10257-005-0030-5. Paternina-Arboleda, C. D., & Das, T. K. (2005). A multi-agent reinforcement learning approach to obtaining dynamic control policies for stochastic lot scheduling problems. Simulation Modelling Practice and Theory, 13, 389–406. Symeonidis, A. L., Kehagias, D., & Mitkas, P. (2003). Intelligent policy recommendations on enterprise resource planning by the use of agent technology and data mining techniques. Expert Systems with Applica- tions, 25, 589–602. Walter, S. S., Brennan, R. W., & Norrie, D. H. (2006). Experience and reflection on the development of a holonic job-shop scheduling system. International Journal of Computer Applications in Technology, 26, 15–27. Wang, Q., Yung, K.-L., & Ip, W.-H. (2005). An order dispatcher for dispersed manufacturers solved by a mixed variable hybrid genetic algorithm. International Journal of Computer Integrated Manufactur- ing, 18, 41–52. Wang, L., & Shen, W. (2003). DPP: An agent-based approach for distributed process planning. Journal of Intelligent Manufacturing, 14, 429–439. Wang, L., Shen, W., & Hao, Q. (2006). An overview of distributed process planning and its integration with scheduling. International Journal of Computer Applications in Technology, 26, 3–14. 2946 O. López-Ortega, I. Villar-Medina / Expert Systems with Applications 36 (2009) 2937–2946 
work_lazgfeppefb75jh4qegfhihxte	----	A major purpose of the Techni- cal Information Center is to provide the broadest dissemination possi- ble of information contained in DOE’s Research and Development Reports to business, industry, the academic community, and federal, state and local governments= Although a small portion of this report is not reproducible, it is being made available to expedite the availability of information on the research discussed herein. ------ -. . , Los Ammos Nallonal LbMfsW 10 ODW~W by MO Unwwry of Cahfomm for IM Uml.d Slw.t Doparmont cd Energy undar conlrsct w. 7405 .ENG. 36 LA-UR--88-3917 DE89 003587 TITLE. THE GRAPHICS ADVISOR: A PROTOTYPICAL ADVISORY EXPERT SYSTEM AUTHORfS): K. P. Berkbigler, C-10 P. A. Max, C-10 SUBMITTEDTO IMACS Conference on Expert Systems for N~merical Computing. Purdue Lfnivcrsicy, West Lafaycttr, lndla~a, D~cember 5-7, 198(+. DIS(:l.AIMER Thin rqmrl wm prepred usmt rnwunl{)f wurksplnmrdhynn ngcncyoflhe (lnllcd S1mcc (irrvernme.11 Neither the I Jmlcd Stma (iovcrnmcnt nor my ngency Iherw(, nor uny of Ifmr cmploycck, m~kcs trny w:wrtinly, express m Imphuf. or -ssumcs urry Iegul Ilnhllily or rcsprci- hlhty fur Ihc uccurmcy,completencw+.or uscfulncu of -ny mformnlmn, uppmrnlu%prwfucl, or prmxw dlwl(mcd, or rcprcncrm Ihsi IIS use would noI Infrlngc prwmlcly owned rlghm Refer. crwc hcrcm III Any rnpcafic cornmurcml pruduc!. prwcas, or .wrwcc hy Irtidc numc, Irrnrkcmark. manufxlurcr, M olhcrwi= d(xs nol ncrxss~rily constitute or Imply IIX cndtmumcnl, rccom mcndwlmr, or fnvormB hy Ihc [ Jnltal SItIcs (irwcrnm?nl or uny tigerrcy !hcreo( “I”hcviews and opmmnn of wthorrn cnpred harm do not rrcccnssnly USIC or reflect Ih(= III Ihe I In,tcd Slnlcs ( iovernmcnl or uny ngcncy Ifrcrcnf BV ~CC091WKP 0! Ihm ●rl#clm Iho publmtlm Iocognllan Ihal Iho L) S GOVC#nmUll rwlmns ● nommclualwa royally.lf~ hcongg 10 publ,mh or 19DroduLa lho publ, shed lrrfm 01 Ih, s Conlf!bul,On of 10 ●now Olhsra lo do go Iof L S Gcrvarn~l DuIpM. lLOSNlaUTilOSL..AI.. .NevvfVlexi..87545Los Alamos National Laboratory u!. , , ~ bsTH!L ~~~~.,~wtim(l#mm 1,*I1(4,:,II!”,}.,1~31mlmnlm, .“ About This Report This official electronic version was created by scanning the best available paper or microfiche copy of the original report at a 300 dpi resolution. Original color illustrations appear as black and white images. For additional information or comments, contact: Library Without Walls Project Los Alamos National Laboratory Research Library Los Alamos, NM 87544 Phone: (505)667-4448 E-mail: lwwp@lanl.gov The (haphics .Advisor: .4 Prototypical Ad~.isor~- Expert S~xten)., . K. P. Berkbigler and P. A. Max Computing and Communications Division Los Alamos National Laboratory Abstract llIe Graphics Advisor is an expert system that lILJllJS uw=rs of tlw Los Alamos Integrated Com- l~lllillg Xclworii select the graphics ]ibrary t]la[ is best suited to the c]]aracteristics Of a specific graphics application The implemental ion of tt, r system, using a commercial expert system devel- opment tool, is described. Delivery option6 are discussed. Although the domain knowledge of [lie Graphics Advisor is Griented toward libraries that arc supported try the Computing and Communi- cations Division at Los Alamos, it exemplifies a class of knowl~dge-based .sys[cms that arc polcIl- tiitll~ useful iII a varir[y of advisory situatiwls. 1 Introduction ‘1 11~’ Gri+l~llics Advisor is an expert system tl:at 11111,IS llsor,~ of tlIII I,crs A]alna Intrgratml (;oIn- pIII”IIIg x~,twnrk (](’N) sr]cct thr graphics Iihrary t Ili\t IS iwst ~ulted to tht= characteristics of a spr- clfic graphich npplicatioll. TIIC expert Ryst.rm de- trrlllineh Ihr p:efr-rred library hy mntching thr graphic features requirrd by the usrr wilt) tlw fra- t urwi supportrd i~ each library usiug r.lles fur- nislwd by an ●xpcr( on the graphics libraries. ‘III(’ Graphim Ad\ism project WJM undertaken wit h several goals in rrlind, OIIF wwr to gain aonir rx]wricllw with untlmercial rxpwt SyRtrIII devrl- 0~111101111001s and ascrrlain thrir nuitahility for dP- vrl(]l~ing ~ystrwm in our rnvlronrrwno Another r-d) JIII IVI) wns to Cxnl]lillr tllr ufirfu]llrns of ?xprrt h}hlq,llb twl ImdL)gy ill fiul.lmrl of tllr ctm~ult illg n~-l Ivlllrs 1~(.rforttm[l ill t lie ( k)llll)llllll~ nn(l ( ‘f)tlI- Illllllirrrllollb l)lvlsir~ll ((’ I)lvifii(m). ‘Ii) t.hi~ endl wo rwhrchrf! for all n~)].licnti,m thrd ui ty}>lral of t)llr rImkIIlt IIIg nelivitirh ftIId mmwal)lr tfl iml~lr- tllf.lllnt li~tI * all rxlwrl nyntrtll ‘1’11(” Illr)drl Fll) l,licnli[~tl slif:hld II,, l;wKr rIIIIuglI t[~ ~xrwrirw rllurb of the functionality available in commercial tools but small enough to be implemented using several tools so that LOOIS can be compared, The prorcs~ of choosing an appropriate graphics library mtj~h[> t!lese criteria and was w’lected for implemental UJII. Hence the Gra])hics Advisor svrvm as a protrlt-w~,( for other advisory syst-nrs that might he built ill C Divisio]], 2 Background Tile I.os Alamos I(”N consists of computers witllill the Central (’cmlputing Facility ({’{’F) al~{.1 a IIIt - worli linking tbcse computers to distribu[~,{l pr,.,. cessors outsldc the CCF and to tcmnillals ii~l{l TCP/l P-ba*ed workstations ill users’ offircs.[1 ] kf”i[ hin the (’(’~ arc many compul ers frmll srvrr;li vendors, including (3ay, CD(”. and L) E(’,l runl]lllg a varivty of opmat ing systrms, CT!X$, B l(~r-al s}>- tenl, runfi on the (’ray cornputrrsi NOS, a (“[)(’ operating systenl runs CrII thr CD(’ n]wllinvs: aII(l the I)E(”’ computcr~ run ritlwr \’NIS or a versir,ll ,,f [IN IX. ThtI I(’N is uscrl hy morr than 8,000” lwtq)b. of which a~~proxilllatrly 5,500” an. Lnl)orator} VIII, ployces and 2,500” arr lor-at rd at ot her inst alliit II IIIS throughout tllc [Jnitr,d S[atw, ‘1’hww usrrs arc,w thiI ICN through tm~tlinaf+, dnitrihutrd procmw,r+, or workst ationri, wit II renmfr usrrs connrct lllg \I; I le]cphonr- cfi~lull, I{*wAxI Iinm+, the I),,f(,lls(, I);it,t Nrt work, or ‘1’rlrlwt, OI)r rr=sponsillillty of t IIC ( ‘<~llll~utor [Is(,r S{,r vices (;roul~ i~ to o~wr~tr thr ( ‘f HIHUll IIIg ofli(f ‘1’his gr(ul)) of higlil! exporlrnrrv{ c~nqmt Ing ]~r, I fwriotlal H rotlrullt~ (Ill n wdr vi+rir-ty of l(,j~i{s, III rludl[lg progrmnlilling lmIgur IgrK, grrrphlc~ nll(l fiys tel II Iil)rnrirfi, and III Illt Ivh ‘]’ll,.~ ~si, Il,lvlNr uw,r> ‘(’lay II FWWIII. III, ( “Inllnd I)mla ( “tq)tm.sllt,th 1)IRII al l~~uipnlriil { ‘url,(]l ● l it,II 1 about what soflware is available for solving cer- tain problems. how to use the softwarf . and what document a[ion describes the software. C Division supports a common set of software across all systems to allow users to move conve- niently from one system to ancther. “l’his soft- ware includes Fortran graphics lib~aries, two of which, the Common Graphics Systm! (CGS) and CGSHIGH, were developed at Los A; .mos, Three arc vendor-supplied, DISSPLA and Graphical Ker- nel Sjsteln (GKS) fron] Computer Associates and ](CAR fronl t]le National Center ff r Atmospheric Research. Consulting on these libra lies is provided by the Consulting Office. Howew:r. the Consulting Office is not responsible for coi)slilting on third- par{)” graphics software that users might instail on individual distributed processors m workstations CGS, a library of device drivcr6 and prirnitivcs. is the basis for the other graphics libraries. It sup- ports a device-independent file, the (XS metafile. that may kw processed on ally graphics device ill tl)e tXF or may be viewed on any suppolted grapl]ics ter]l]inal, SUCIT as the Tektronix 4125, TIIC cGS1ll(; lI library pro~ides high-level subroutines t hat call prclducc a conlpletr x-y plot from one call. II al-u 6upIwrts subroutines that control plotting dd:lu]ts, such as color and line style, that are set l~! Ill,, higll-lrvr] routinrs. ‘l” IIv I) ISSI)LA librtiry ; ;ovides ntiddlc-lcvrl cii- j,ill~llilics, illcluriillg x-y tllrer-dinlellsiollal, and rrmtcjur pluts, grapllirs arts chara~ter fonts. intvr- polat ion and curve smoothing. pie and bar charts, and Illaps of coast lilw and political boulldarivs, I) ISSPLA is the only !ihrary at Los Alamorj tllal supllorts graphics art’. character fonts, so it is umd for producing publication quality graphic~ GKS, a Iihrary of b~ic funrt i,ms, provides mow than 200 subroutine Illat olill~ul graphical primltivcti, WI ntlriljulrkl control worlwtntiollsl pcr[ornl transfer. II Ia[IrNIS, ol)tuin gr,)l,llical input, and lIandl P crr(jls ‘1’111*R( ‘Al{ Iibl;try proviflos high-l?vcl rapnhili- ti{,s, IIlcludlllg x-;?, t l~rrr-dirllrllsiollal, nnd cwnt~mr lIIIrl S, lWt>{lilllrl,siOllill vrlt)city fields, Rtrfanllinr rt’~llc%(’illnliclllh of n flow fwld, half- tolw l}lc[ure6, ;III{I gt,t~grnl~llir II I,ajm ‘1 lIr ill]portanrr lo tlwr comftlunilim of arI n(l- wisf.jry hrrvirr in wrll rfil ni~liNhrd[2], I)(II M rmm)t r c{,llli~utiti~ Ivwmlwk nmrc prcvah-nt thr arIwIunt (If fnw Iwfam IOIIRUII it:g is drrlllling ThI~ hm hmII t 1P’ trcrl{; Id I (M Alnmf~ for BIJIIW tilllr Innfn’nl Ivr nl)l)rfmctlrs I(I providing gIIIdmICr to lJnrrs nrr r(. qulrr~l, nlld ~x~mrt riyN~enlti Ilmvr Iwrn ~uggr~t.r(l M n l~(vi~il}lc hlltlltl(lll [3] Tlie advisory function is being actively studi~.i! from bo~h the Behavioral science and com~)u[(’r science perspectives. (Set= [4] for a comprell~ll- sive overview of advice-gi..’ing syste,rns. inciudi]lg a review C! the literature. ) Much of tht, currl.llt research is oriented toward embedding tll~, fi(ll i- sory ~ystem within t IJe soflware system being used. e.g., Lhe operating system or word processing sys- tem. These advisory systems are designed to ei~ lwr passively or actively help the user dbring interac- tive use of the software. A user model in wllicll the user’s goals are dynamically inferred from tll~ user’s interactions with the >ystern is generally inc- luded. Advice can thus bc given in context atld geared to the user’s apparent level of sophist i cat ic,ll with the system At this stage tlw scope of the Graphics Advisor is more modest. It operat~s M a separa(r systt’]1), so no opportunity exists for observing lllr user at- tempting to usr t!:c graphics librarivs. Furtlll. r- more becaus~ its p!Jrpose is to help the user st~ltI-I a library, the Graphics Advisor will typi( ally 1,,, consulted before the user gains any experience with the libraries. The exp~:ted users of the Graplllcs Advisor are primarily new or novice users who nrc~l general information on tlw availah!~ librarirs an~l tllcir c,apabilitim relative to sprcific applications. The inlplicit user model in tlw ~; raphics Advisor is t.argctecl at this audicllcr The C;raphirfi Advisor has nmre i~l ccmrlnlon with the results of the protocol study cm onr-sllt)l di- alogs reported by Artronson and ( ‘arroll .[.’); \li\l)} of thr Wrategics they found human ndviw)rs Prll. pio~ in ri situation whrr( no follow-u~) is ]~ossil,l(. havr also been used in tlw (;rapbirs Advisor, !+)lllt. ?xanlplm arc reeking rmsulllptiot:s al}out III(. Ilw,r’s goal%, [)rovidillg R!tcrllat ivc fiolllt lolls, 1)011)1IIIg III rrfmcncc murcrs, and pr(~$’idillg ill) CXlllilllilt if)]) tion of II]C (lrapllics Advisor took place on a Texas Ill+tru])wnls Exl)lorer Lisl> nlacllilm. 3.1 KEE KEE is a hybrid expert syslem development tool offering a variety of knowledge representation ar,d reasoning possibilities[8], many of which are used in the Graphics Advisor. KEE is available for nu- merous hardware platforms in both development and delivery versions. Interactive arid progranl- n)atic interfaces to its capabilities are provided. Tile frame system allows the definition of hierar- cl)it’s of objects or concepts. At the top of the IIierarclly arc classes that contain slots describing [he attrihu(cs and behaviors of an object and serve as templates for the creation of individual objects Iinmvil as instances. Instances inherit the slots of their parent classe6 and may define their own val- UM for these attributes, Tile rule system implements both data-drivel) reasoning (forward chaining) and goal-driven rea- soning ‘backward chaining) using a single represen- tatiol~ for the rules, Rules may be partitioned into classes for efficient handling of subproblenm. \“ari- ables arv Iwr]nit[ed ill the rules and arc bound to values at runtinw when the rules are tested, ‘I”hv ol~j~’ct-oriclltrci programlnitlg and acc~ss-oricll(r(! prograllln):l~g (r-lemons) paradign~ ar~ available l): Ivri[il)g rod(, ill conlllloll Lisil, [he inlr)lt~lnell[atit)~l li~llgllag(’ fur I(EE. Support for nonrnonotonic rrn- smlillg is IIrcrvidrd in tlw fornl of tllrKEEworl~!\ and lrulll llliiilltCliJllC(’facilili(:s, I{es(’arcll in Ilunlan-compulrr intcrfare t~ cll- Ili[lu(,s lIas lIOgUII to produce guidclilms for WIICII 10 usr Kllec]fic ]lltrr~r(]crn sty]es. [~,lo] KEE pro- Vldvs uficr-inlrrfare drvrlopnlcnt facilil ics for rwv- crnl typeh of interfarcs, Muttiplr witldcwfi are tyll- ically usrd in an application to separatr diffrrrnt t ypm of information, Mrnus cmu be generated ~ly - nnl~lirnlly, if drsirrd, and tllr Activc]magrs fnril- ity for Ijuildll]g dirrct nlanilmlation interfmw call I)r ru~tol]uzrd. (irnl]llical illt,crfacr~ Illay he CIIII hlrucletl Ilhillg colllljolmnts frorll li~;l;pirt urvs. 3.2 Graphics Advisor i(II(mlr{lgI ill tlw (irallllic~ Advi~{~r is rrl)rcncllt[.tl usill~ l)f)tlI tlIr franw siyntorll and III(. rule nystrltl l:l~llro 1 klI[m+ n (lIklJ]ay of [11~ linow]vdg,,IHWWfor 111~*f;r:~l,lll(,~ Atlvl~{)r. ShJII. It) fIm IIVh hrr usrd to r~’l)rmmrlt [nrls rrl~(llll ~rn[~lllrb IIl)rnrms kurlI I* wholllrr fI ~jnrllrlll;lr (1.;I am-mii:nams Ma-~AL-nm-mTTas aAs-~MruAToas RAs-smLs-rALLrY- msPsATs us-sLoT-~MKs MS--W- WLT-H~S SAD-~~- WLt- MITI@S-CUSTa-S~~ SAs-usrsu-n!sa-mnms , Sussl 4 ‘ .s0ss2 ,, ‘ ‘ .Sossa,, ,.. - ,...s0ss4 I SO’SSIBLMILUtj :-- - Sosss,: . I .. .! . ‘SUSS6 / ‘, ‘s0ss7 ?0ss8 I / .?mrl ,,mtrls MD, ~LES \ .’.’ mmi ,,, \ ,’,’,J,Pmmz ,,, , ,,, , .rscr~a,,, ,, ‘“’,’, rmi4,,, , ,,, ,., “,. . , ?ml$ . ‘.--,,. ~ -., , - ?mi6 ~:, .-. PBtrnus, mus4 i -. .-. . PmIr17 d:. . , Css .;, casalm Imuxssa : - .- DISSPSJ . ...m(s ‘ Mu 10T ,~L ,M L(S. ~t ,AV lLss. PWsm w’. --. p,trz . . . . . “.. .. . . . . . ‘. mr3... . . . tui~ is supplied by the library, definable by the user, or not definable. Attributes are defined for the libraries class and are inherited by individual libraries and assigned appropriate values at the in- slance level. information about the characteristic of the user’s application also is stored in slots in a similar frame as the user enters those characteristics. Con- straints upon allowable values have been specified and are enforced by KEE. For example, it makes no sense for publication quality to bc both required and not required in a siug]e application, and this inconsistency is not permil led. Customized facets have betn defined in the frame used to describe the user’s problem: they arc used to store references to defi]litions and examples of certain plotting fea- tures. Rules are used to represent the conclusive as- pects of the expert “s krlo\vlcdg?, The rules ir, tllc Cral)hics Advisor are diVided inlo two classes. One class of ru]cs determines \~llich libraries could pos- sibly h used by conlparillg facts about libraries \vill] requirements defined by Lhe user, Libraries IIIa I do no[ suppor[ rI required feature are elilni- Ila[ecl fronl consideration. Another class of ruh,s represents the expert’s ht=uristlcs for when to use ~ii~]l library. This incl~dcs sul)jeclive judgrnlt,ll[+ such a~ whe[]ler a library provides su]]crlor funr- 1 iulliilily in it particular area, case of use, and docu- 1111,111~11011 (Illiilil)’. 7“IIC rxpcrt S~SIClll oll C7illf ’S US- ill~ forward cllp.illing. \vilh some \vt?igllling f[irtors ill,l~ll(’(1 \vi[llill tl~v second rule class. ‘T”llc cal]irbil- ily fur varialjlr l)iu Jing within rulvs is hcavi]y used and is rmpolisil~lr for [he rihility to rxpress knmvl- (,Il&r ill {II(, Gral)llics Arlvis~,r i!) so frw rulm. It is oslill~;lt~vl tl];~l il woIll(l Iako 1-2 cmicrs (Jf nlilglli I IId C IIIrx(’ rulr+ to lllll~lrllwnl the sallw knt~wlrtlg(, using a rulr riy KICIII tnat dow Iiot support varial)lcs. Iricllcl]y Uscrh WIIO l]avc exalllild tlh= (ir;l])];ics Advlsrx tilld ilR usrr interface vrry attract iw’, “1’l)v illtrrfnrr iri in~plenmntcrl in ~ dirrct rnani}~ulno icm slyh. lJFlllg t.]IF Ac[ivrlllmgcri suhyslcm of ~ltl~. ‘1.110 primary illmgr paurl is shrJwlI ill Iigurr 2 ‘1’llr urwr Billlply I)wnbi willl the mouse to Ihfwr frn[ urrs IIInl tlIt- ,tl)])li~iiti(}l) IIrrd~ ILIIrl cllckp, IIIr Irft III(NW 1)11111)11 t!: rrglstrr 1110 rlmirrw. No uw of 1111 Iwyl)onrd is mndI’ rxcr-lIt ill thr optimml rf)ltr 1111’111 filrl]lt~ ‘]’tlr’ hlj-: ‘]”J’l)mlcrill! Wllldo W iR IJW(I III rol].llillct Lu) wit II t III, (IIIt put jrotl] Art ivrlllla~r~ I(J IIlovi,lr ndditl(ulnl trx Iuml itlfi)rllmlitul, murh M n rrfvrrtlrr to (lf,rlllllt’lltntl~lll !ilr 1.1]0 rf,cltllllllrlltl(,tl Illlrilry tw ml rxlllrrtlnti(jll {)f 111(, rruwlllillg mml ill ;lr~ li lllg xl Ilw rcrl]llll,l(,ll(l:illllll Some of the operatioll.i ill tllr (;ral)llir+ Ad\i - sor are implemented as nlethods. thal I’ . prfJC!dll- ral Lisp code that is executed under o11P ol,j,lct - oriented paradigm when il recrivcs a rlws<agf. Messages are sent by the user through IIIC us, of method actuator images. For exalnpll. tvlwli the Make ficoznmendation image is sriectod, a message is sent to the Lisp met hod t hat invokes t h,, rules and displays the possible and recomrrmndf.,.l libraries on the screen. Other mc[llods arc lis~:d Ii, initialize the knowledge basr to its prc-ccmsult :iII(III state and to display explanatlor]s of (1w expert sY~- tem’s re~oning process. Demons, known as Active l:alues ill Kill;, itr, used to control the display of plo! option> tt) [III. user. General in forrnalioll, such as plot tyl~~+ r, quircd and operating sys[t’m requir~”d. is dlsl~l:l}til in the primary image panel that is alwa}s prt,wrl! on tile screen. Options having 10 do will] [Ill, ii])- pearallrr of the plot arc on a srparale Ij,lnt,l tll;,l pops up only if the user r(,qurwts limro (Io[aill,(l control over plot appear anrc, Silllilarly, I Iw fnlll,lll device to bc usrd is otlly relevant for onf, plot [yl)(. so it is shown only wllcll t]lal plot tyl)t, is WIITIIOII The USI of Inultiplc, autoirla[ica]ly rli+lllay’lt l,A]!. rls reciucrs I])(, complexity of llIr illforll]al 1o11 111:11 tilt’ user nwst deal wit]) a[ OIIV tilll~’. ‘1’110 cxplanttioli n.mipollrnl f,f at) FIll\IY ,r} ,s prrl Sy”slrnl iti al) CMlllllliil t’lIillI[,lIl I’m I II, s)sI(III”> accrplancf~; usf’rs oflvll Walll to kllilu \vl]! it r(,i, onlllmndtit ior, wns n)adr, All Il[nlgll Kl:l.1 I]rlwl,ll.< gOOd rll)f, tritrlllg ~apal)ll~llt,s fllr 1]](- k}’~11,111lllI\ll oprr, \w, feel that lhiq cnI)trlJllily d(ws 11,)1 llr(~rlll,, an adrquit(rly uhw-oriwtrri CXlIl:III;IIIIJII III I III, PN pcrl syslcIII’s rc;wolling Accord lllgly, a t(,t 1111111111 was adnlllc~l whrrrl)y VJIIgl IslI St iltl’111(’111~ (If t 11~, rc(lso]w tll:ll IIl)r;lrim wrrf, clill]ill;\lf,,l fr,}lll t(tll\i,l crat ion rirc riirvml ill a sh)t ill II)(, kn~nv!wlg,, I);!-,, ru+ I Ilr rr[L6clllllg pr(lgrrs~(,s ‘1’llf’w’ ilf’(’llll llll;llr(l ~tatclllcllts arc III(SII dlsllln~wl III t lIr I]sf’r \YIII~II ill) rxplfinnliml i~ -rqumtr(l ‘1’11(’ I’xplilllallllll (“a]lilllll ‘t)’ lh il]lJkitriltllll ill ll~llrl, :] Just .M tllr (’xl~{”rt ~y~trlll IINlSI i,, :11,1, i,, ,x l~]nill Itsrlf 10 ttlr Uhrr, N) II II.111111.,+ 1111, IIw.r IV, III III like It) cxl~rrHs cotllt]lrllt~ III tllr S! SII, III II) 1111 (;ral~llir~ Arjvimlr, {JIW lllny {lo ~,, I,y usllIg II I,, ,,,1,, rlwnt fmclllty ‘1’hr user I% I)rotlq)lr(l to t~l),. 1,, ;, frrv-forlll rwnmrwl thnt ifi rm~r{l III n Iilr f~ir KIII, nrqllrllt Iimlls;ll Ity thv Oxl)crl ~ystt,lll (lr\,f.1ol)t.1 ‘1’llin cnl}nl~illly ifi Irivlnl t[~ l~r,}iltl,. ntl(l I)rrwr(l Ii, Itr umf’ul ftjr tllllltilllllg fvrIll Imli wllI-11 IIW.11 III .I l>rrvlt~lm ~y~[,,l)l IIvvoI,,I),uI ;~t 1,,):. Al;ItIl(,~ I I I ] ‘1’llc Illlrnr.w (It hl.rlllllly, llIr ~.,r;ll,lllls ft., It III,.< 4 Left click the mmse on each required plotline featu”e Then left c!ick on Make Recommendation lis24 Middle click for tin example of the plo:tine feature Right click for information on th”: plotting feature lfT’=71 1 I Selections I -=-’lI-=l II m CIJXVE II I Pla Pmwous II La MAP PXOJLC1 ION MFALMISm VELOrll V v1C1ORS II FMM41WFMINTS J rUnknown CGSHIGH LIBRLWY CANNOT DO MAP PLOTS. CGSH!GH LIBRARY CXNN13T DO CONTOLIRFYOTS. GKS LIBRARY IS NOT AV41LAllLE ON TtK CTSS WERATING SYSTEM. CGS LIBRARY CANNOT DO PUBLICATION IJUALITY. NC4R LIBRARY CANNOT DO PUBLICATION iJllALITY. GUS LIBRAfw CANN(IT 00 PllBLICATlnN OIIALITY. CGSHIGH LIBRARY LhNNOl 00 PUBLILA1lUN QLIAI ITY. t lint III(* u..rr niay rrquirc arr mwwsnrlly Ivrw 1 mm I User’sComments 1 I Explaln Possibl(’ I-ibrarles Figure2: Graphics Advise: imagcpallcl I zxa~la a-D Cuswo I 5 icl)[ state of technology does not permit this, at least not as tl~e Graphics Advisor is presently in~- plenlen$.ed- One of the strengths of KEE is its built-ill capabilities for designing direct manipu- lation user interface Such interfaces are easy to learn and use and allow tk ~ user much more flex- ibility in the order in which operations are per- formed Placing the locus of control with the user permits specification of input values in whatever order seems most natural, rather than in some pre-determined order dictated by the programmer. The user interface to the Graphics Advisor requires a high-resolution bit-mapped display with a win- dow system and a mouse, features not available to users with “dumb” terminals. The technical op- IicI-rs avaiialr!c today for expert system delivery to lC~~ users represet]t varying degrees of compromise bet\veen availability} to et-ery user and a powerful, flexil.11(’ user interface, C Division is working on a project aimed at distributing computing tas!is bet\vcen workstations and mainframes, which may alleviate this problem in Ihe future. More than one delivery option may eventually IJP implenlented, but t]c first delivery platform for tlw (Jraphics Advisor \vill Lre to use Run[ime KEE 011 a 5[” S \\orkslatioil.2 The expert system will lJe available in the Qomputer User Services Group and possibly at selected user sites that have a large CUIII ingcnt of new users. The feedback obtained fron~ use of Ille system at these locations will guidr tllc dlrr=ction of future efforts, Olllrr options tl;at were considered and dis- card(’d for now inclr-rdcd using a distritulcd dc!iv- cr} sy\lcIIl fron] Inlcll]corp wllcre the expert sys- tctl) functionality is dis[ril~ulcd bet.t;een a VAX ]naillfranw a:ld I13M/PC front ends. Thi~ option is Ilul cml]patihle with much of our current tcr- tr; illal network, lh=irnplemellting t he system 011 an ](’if ll(Jd!’ SUCII u a VAX and redoing the user in- tr-rfacr ..0 operate on any V’TIOO-compatible t,ermi- nal also was considered. This idea was rej~cted m too titlm-consenting and too compromising of the desirable umr interface ill the current implementa- tion Reinlplcmenting tlip sy~telll in a less ●xpen- sivr rx}wrt systcnl tool that could be distributed wit ]1 tile app!icalioll WM at~amdonrd M too difficult to ndlliil~iHtcr. Although d!.livrritlg 011 SIIN workntatitms has a dmdval)tagr, Ilallwly the ability to rracll only a sul Iwl of our Utirr Crmll;mnity, it al)prars tr) tw tl)o III*SI nlt~rll:ltivc RI this t.ill~r Accordingly, rI I{UII ‘SIIII hlicrtmynlrlllm. 1111 time KEE license for the SUN was obtained. and the Graphics Advisor has Lcvl) pcrtcd to that pliit - form. The system has no! yet been dislribulcd, I,ul we expect to do so shortly. q Future Directions Many possibilities exist for enhancements to t hv Graphics Advisor. The knowledge stored for tlw graphics librarief could be extended to include ill- formatirm about when and how to use specific suh- rou:irws to accomplish particular plotting opera- tions. Such a system would be appropriate fc,r users having a broader range of skills and nm} require a more sophisticated user model. As II I,, foundation for dis(rihuted computing ill IIIC 1(’S becomes more rohusl, \ve may try to ilnl}lcnwllt the (lraphics Advisor as a distributed appllra(iol] It has also Lecll suggested that ,t similar sys[cl]l It, advise on tile matllematica] libraries a\ai IaLl,, al Los A]amos would be useful. Further wo:h \vill III governed in part by the feedback received 01] [II(: prototype of the Graphics Ad~isor that is al,out [~, be rcleiuicd. Acknowledgment, This work was done und~:r III(I auspices of the ~ .S, Depart rrlent of Fnergy, References [1] [2] [3] [4] [5] N, R. Morse C-Dlul.ston Annual R(lr(u CIId Optmiirrg Pla II. LA-11216-MS, Los A]irIIICM National Laboratory, 198S M. J. Coombs and J. L. Alt::. (.’umpu/lnfl Skill.+ attd fh~ /lsrr fn/crfarr, Acacfr-]llit, ]’r,,ss, LoIIdoII, 1981. P. Anstry. Computing Advice at a l)istal\((,: The ‘Remet e Advisory” Concept. So/tuJ4 n Prachcc ~nd Erpcrlrncr, 16(11):1045 1052, November 1986 John M. (;arroll and Jpan McKerrrlree, III. tcrfam Design Issues for Advice-giving k:x pert Systems, (~ommun:raflons of thr ,4{”,1/, 30(1):14 31, January 1987 Anly Aarontum at]d John hf. (~arroll, lntrl ligrnt I{f?lp III a ollr’-shol l)lrIhJg: A I’rf,- torol St urfy In Prwrrding.s )Iuvtan l“ar-(or.+ m (.’umpullng .~’ystrm.i all d (;wpl)fr.+ /H/r r- facr, l)npi I(M l(iH, A(’M/sl(;( ’111, ‘I”,w,,l,,,,, (’arncla, April $!) I!M7 [6] \\”illian~ B. Gevartet, The Nature and Eval- uat ion of Commercial Expert System Build- ing Tools. lEEE Computer, 20(5):24–41, May 19d7. [71 \Villiam ?dettwy. An Assessment of Tools for Building Lmge Knowledge-baaed Systems. A I J?aga:tne, d(4):81-89, \Yinter 1987. [i”: I(EE User’s Cti Idc. Intel licorp. Inc.. hfoun- tain \:iew, California, K3.1-VG-1, hlay 1988. [{)] Ben Shneiderman. Designing fh~ Us~r lnferj’ace: Stmt~gtts /or Eflcctire Human- Computer ~nttracilon. Addison-M:esley Pub- lishing Company, Reading, hlassachusetts, 19s7. [10] James A. Hendler, editor. Expert Systems. The (;scr /nt(rfac(. Ablex Publishing Com- pany. Norwood, New Jersey, 1988. [11] hlary Stoddard. Kathj Berkbigler. Bob \\-lleat. and Eva Peter. l-ser Behavior l-pen introduction of a Network Help System. SfGCill Bulletln. 16(3):25-31, January 1985. 
work_ldcd3yq7wverxmugytokxwwahe	----	Risk-based Test Case Prioritization Using a Fuzzy Expert System Charitha Hettiarachchi North Dakota State U. charitha.hettiarachc@ndsu.edu Hyunsook Do U. of North Texas hyunsook.do@unt.edu Byoungju Choi Ewha Womans U. bjchoi@ewha.ac.kr August 25, 2015 Context: The use of system requirements and their risks enables software testers to identify more important test cases that can reveal the faults associated with system components. Objective: The goal of this research is to make the requirements risk estimation process more systematic and precise by reducing subjectivity using a fuzzy expert system. Further, we provide empirical results that show that our proposed approach can improve the effectiveness of test case prioritization. Method: In this research, we used requirements modification status, complexity, security, and size of the software requirements as risk indicators and employed a fuzzy expert system to estimate the requirements risks. Further, we employed a semi-automated process to gather the required data for our approach and to make the risk estimation process less subjective. Results: The results of our study indicated that the prioritized tests based on our new approach can detect faults early, and also the approach can be effective at finding more faults earlier in the high-risk system components compared to the control techniques. Conclusion: We proposed an enhanced risk-based test case prioritization approach that estimates requirements risks systematically with a fuzzy expert system. With the proposed approach, testers can detect more faults earlier than with other control techniques. Further, the proposed semi-automated, systematic approach can easily be applied to industrial applications and can help improve regression testing effectiveness. 1 Introduction Software products change over time due to feature updates or user demand changes. When software functionalities change, software engineers need to retest the software to ensure that the changes did not affect software quality. Regression testing is one of the important maintenance activities, but it requires a great deal of time and effort. Often, software companies have pressures with time and budget, so expensive and time-consuming regression testing could be a major burden for them. To overcome these schedule and cost-related concerns with regression testing, many researchers have proposed various cost-effective regression testing techniques [20, 33, 35]; in particular, test case prioritization techniques have been actively studied because they provide appealing benefits, such as flexibility for testers who need to adjust their testing efforts for the limited time and budget [10, 13, 42, 44]. While the majority of test case prioritization ap- proaches utilize source code information, some researchers have investigated using other software artifacts, such as © 2015. This manuscript version is made available under the Elsevier user license http://www.elsevier.com/open-access/userlicense/1.0/ system requirements and design documents, produced during early development phases [7, 25, 41]. For instance, Kr- ishnamoorthi and Mary [25] proposed a system-level test case prioritization approach using the information obtained from the requirements specification, such as requirements completeness and implementation complexity. Srikanth et al. [41] also introduced a system-level test case prioritization technique that analyzes and evaluates the requirements in terms of requirement volatility, complexity, customer priority, and fault proneness. In addition to utilizing requirements information for test case prioritization, some researchers used risk information that can help identify more important test cases that are likely to detect defects associated with the system’s risks (e.g., safety or security risks) [43, 48]. The results of previous research work empirically showed that the effectiveness of test case prioritization could be improved by using requirements risks. However, these risk-based test case prioritization techniques did not consider the direct relationship between requirements risks and test cases [48] or only used one type of risk, such as fault information obtained from preceding versions [43]. Further, these studies evaluated the approaches by measuring how fast the reordered test cases detected faults; the approaches utilized risk information to prioritize tests, so they should be evaluated by measuring whether the detected faults are, indeed, from the locations where risks reside in the product. To address these limitations, our previous work [21] proposed a new requirements risk-based test case prioritization approach by considering the direct relationship between requirements risks and the test case. We also introduced a new evaluation method to measure how effective test case prioritization approaches were for detecting defects in risky components of software systems. While our previous research was shown to be promising, the approach required human experts’ involvement dur- ing the risk estimation process. Human involvement with the risk estimation process is important, but it makes the estimation process subjective and imprecise. To avoid the possible imprecision introduced by human judgment, a more systematic approach should be considered. Often, fuzzy expert systems have been utilized to address such problems because they can provide a mechanism to simulate the judgment and reasoning of experts in a particular field. To date, many researchers have used fuzzy expert systems in different application areas to help with complex decision-making problems, such as the diagnosis of disease [4] and risk estimation in aviation [18]. These studies have shown that fuzzy expert systems can be used to systematically represent human expertise in a particular domain and to deal with imprecision and subjectivity-related issues of the decision-making process while making the decision-making process more effective. In this research, we propose a systematic risk estimation approach using a fuzzy expert system to address the limitations of our previous approach. We also reduced the number of risk items used for the risk estimation and simplified the prioritization approach so that we can perform test case prioritization with less effort. Further, from the results of our previous requirements risk-based approach, we learned that incorporating code information with the requirements could improve the rate of fault detection. Therefore, in this study, we used code information to extract requirements risks with respect to a few risk indicators in addition to the information obtained from requirements specifications written in natural language. Because we use code information, the proposed approach is applied during 2 the testing phase after coding is done. To evaluate our approach, we used one open source application and one industrial application developed in Java. The results of our study indicate that the systematic, risk-based test case prioritization approach has the capability to find faults earlier compared with other test case prioritization techniques, including our previous requirements risk- based approach. Moreover, the new approach is also better at finding more faults earlier in high-risk components than other techniques. The rest of the paper is organized as follows. Section 2 describes the fuzzy expert system used in this research and the related work. Section 3 describes our new prioritization technique in detail. Section 4 describes our experiment including the research questions. Section 5 presents the results and analysis. Section 6 discusses our results and their implications. Section 7 presents the conclusions and discusses future work. 2 Fuzzy Expert Systems and Related Work In this section, we provide background information on the fuzzy expert system and the existing work related to test case prioritization techniques, mainly focusing on techniques that use requirements, risks, and fuzzy expert systems which are most closely related to our work. 2.1 Fuzzy Expert Systems In this research, we use a fuzzy expert system to derive requirements modification status (RMS) and potential security threats (PST) values. The fuzzy expert system used in this work simulates human expert’s reasoning to derive the RMS and PST values for each requirement in a similar way that a human expert would estimate these values using the same input values. Existing empirical studies [19, 46] indicate that fuzzy expert systems can improve the effectiveness of decision making process in many different application areas including regression testing. Moreover, fuzzy expert systems can handle ambiguity, which in turn produce output values much closer to realistic values. To provide a better understanding about the process of acquiring the RMS and PST values, we summarize the mechanism for a fuzzy expert system used with our approach. A fuzzy expert system is comprised of fuzzy mem- bership functions and rules. It contains four main parts: fuzzification, inference, composition, and defuzzification. Figure 1 shows the typical architecture of a fuzzy expert system. The fuzzification process transforms the crisp input into a fuzzy input set. The inference process uses the fuzzy input set to determine the fuzzy output set using rules for- mulated in the knowledge base and the membership functions. The composition process aggregates all output fuzzy sets into a single fuzzy set. Finally, the defuzzification process calculates a crisp output using the fuzzy set produced by the composition process. The knowledge base shown in Figure 1 contains the selected fuzzy rule set. In a fuzzy expert system, fuzzy rules play a vital role because they are formulated based upon the experts’ knowledge about the domain of interest. The fuzzy rule’s antecedent defines the fuzzy region of the input space, and the consequent defines the fuzzy region of 3 the output space. Fuzzy rules can not only support multiple input variables, but also multiple output variables. The following equation shows an example of a fuzzy rule. if x is A and y is B then z is C where x and y are input variables, and z is the output variable. A is a membership function defined on variable x; B is a membership function defined on variable y, whereas C is a membership function defined on output variable z. Fuzzification Knowledge Base Defuzzification Fuzzy Inference Engine Inference Composition Crisp Inputs Crisp Outputs Rules Input Fuzzy Set Output Fuzzy Set Fuzzy Inputs Fuzzy Outputs Figure 1: Architecture of a fuzzy expert system Fuzzy Set Theory. To understand the fuzzification process, some knowledge of fuzzy set theory is necessary. Fuzzy set theory was first introduced by Zadeh [49] in 1965, and it defines fuzzy sets as an extension of conventional sets. In a conventional set, an element either belongs or does not belong to the set. Equation 1 defines a conventional set where µA(x) shows the membership of an element, x, of a conventional set, A. µA(x) = { 1, if x ∈ A 0, if x /∈ A (1) Unlike conventional sets, in fuzzy sets, the membership of elements in the sets can be partially represented. Be- cause fuzzy sets permit partial membership, the degree of membership is determined by using membership functions for unit interval [0, 1]. Hence, the imprecision of input data is handled by obtaining degree of membership values for each membership function defined in the fuzzy expert system. A fuzzy set is defined as shown in Equation 2. A = (x,µA(x))|x ∈ X,µ(x) : X → [0,1] (2) where A is the fuzzy set, µA is the membership function, and X is the universe of discourse. Fuzzification. In the fuzzification step, the input variables’ values are used to determine the degree to which these values fit into each membership function used by the fuzzy rules. There are several types of membership functions 4 available, such as triangular, trapezoidal, and gaussian. In this research, we utilize the triangular membership function, which is widely used for fuzzy expert system based research and applications. Various empirical studies [39, 46] indicate that the triangular membership function is easy and simple to apply compared to other membership functions. As an initial investigation of the use of a fuzzy expert system, we chose the triangular membership function due its application simplicity. The triangular membership function is specified in Equation 3. µA(X) =   0, x < 0 (x−a)/(b−a), a ≤ x ≤ b (c−x)/(c− b), b ≤ x ≤ c 0, x > 0 (3) where A is the fuzzy set, µA is the membership function, X is the universe of discourse, a is the lower limit, b is the modal value, and c is the upper limit. Table 1 shows the membership functions used for the input variables in our experiment. Table 1: Input variable membership functions Linguistic Value Triangular Fuzzy Numbers (a,b,c) Low (0,0,5) Medium (0,5,10) High (5,10,10) Inference. In the inference step, the fuzzified inputs (i.e., the degree of appropriate membership functions for input variable values) are applied on each rule antecedent to determine the degree of truth for each rule. The degree of truth is applied to the consequent of each rule, thus each output variable obtains an appropriate membership function. The output variable membership functions defined in our experiment are shown in Table 2. Our fuzzy expert system is comprised of nine rules that represent experts’ knowledge and experiences regarding the risks associated with software requirements. A sample set of rules used by the fuzzy inference process to estimate RMS is shown in Table 3. With the first rule (R1), the membership functions of two input variables, RML and PRV are low, and the membership function for the output variable, RMS is also low. Table 2: Output variable membership functions Linguistic Value Triangular Fuzzy Numbers (a,b,c) Low (0,0,5) Medium (0,5,10) High (5,10,10) Composition. The composition step is to combine all membership functions obtained from each rule for each output variable and to form a single membership function for each output variable. 5 Table 3: Fuzzy rules for RMS R1. If RML is Low and PRV is Low then RMS is Low R2. If RML is Medium and PRV is Low then RMS is Low R4. If RML is Medium and PRV is Medium then RMS is Medium R9. If RML is Medium and PRV is High then RMS is High Defuzzification. The last step of fuzzy expert system is defuzzification which is an optional process. The defuzzi- fication process produces crisp output for each output variable. In our experiment, we need exact, crisp output to quantitatively measure the RMS and PST for each requirement. Therefore, in this step, the resulting membership function of the previous step is defuzzified into a single number. There are several methods to perform the defuzzifi- cation. In our fuzzy expert system, we follow the Mamdani [30] type fuzzy inference process. Therefore, we use the center of gravity (COG) method which is considered as more accurate and widely used with the Mamdani-type fuzzy expert system. Equation 4 shows how to compute the COG that represents the crisp output for a particular output variable. y∗ = ∫ b a µA(y)ydy∫ b a µA(y)dy (4) where y∗ is the crisp output; µA(y) is the aggregated membership function, A, on interval ab; and y is the output variable. To illustrate the aforementioned processes, we provide a simple example. Suppose we estimate a requirements modification status for a requirement using two inputs: requirements modification level value of 10 and potential requirement volatility value of 8. The fuzzification process fuzzifies these values to produce the degree of membership for each input membership function. The fuzzy rules defined in the expert system use the fuzzified input and the inference process produces the degree of memberships for the output variable (requirements modification status). All membership functions obtained from each rule are combined and the resulting membership function is defuzzified at the defuzzification process to obtain the final crisp value of 8.4 as the requirements modification status value for this requirement. 2.2 Related Work Test case prioritization provides a way to schedule test cases so that testers can run more important or critical test cases early. Various prioritization techniques have been proposed [47], and some of them have been used by several software organizations [28, 42]. The majority of prioritization techniques have used the information obtained from software source code [10, 36, 40]. For instance, one technique, total statement coverage prioritization reorders the test cases in the order of the number of statements they cover. One variation of this technique, additional statement coverage prioritization reorders the test cases in the order of number of new statements they cover. Other types of code information for aiding prioritization include code change history, code modification, and fault proneness of 6 code [31, 40]. Beyond code-based information, other software artifact types, such as software requirements and design information, have also been utilized. For example, Srikanth et al. [41] proposed a test case prioritization approach using several requirements-related factors, such as requirements complexity and requirements volatility, for the early detection of severe faults. Krishnamoorthi and Mary [25] also proposed a model to prioritize test cases using the requirements specification to improve the rate of severe fault detection. Arafeen and Do [7] proposed an approach that clusters requirements based on similarities obtained through a text-mining technique and that prioritizes test cases using the requirements-tests relationship. These studies reported that using requirements information improved the effectiveness of prioritization. In addition to requirements and design information, other researchers have used software risk information to pri- oritize test cases in order to run test cases to exercise code areas with potential risks as early as possible [43, 48]. Many risk-based testing techniques have adopted Amland’s [6] risk model that estimates risk exposure as a product of probability of faults in software components and the impact (e.g., cost or damage) of the corresponding fault if it occurs in the operational environment. In our approach, we also used the risk exposure of both requirements and risk items to prioritize tests. Stallbaum et al. [43] proposed a technique, RiteDAP (risk-based test case derivation and prioritization), that can automatically generate test case scenarios from activity diagrams and can prioritize test cases using the risks associated with fault information. In this RiteDAP approach, to quantify the risk, probability of failure for each action is estimated by the usage frequency of each action, whereas the damage (impact) caused by that par- ticular failure is estimated through its financial losses. Yoon et al. [48] used the relationship among requirements risk exposure, risk items, and test cases to determine the order of test cases. Another paper [27] proposed a value-based software engineering framework to improve the software testing process. The proposed multi-objective feature prior- itization strategy prioritizes the new features by considering the business importance, quality risks, testing costs, and the market pressure. Further, Felderer and Schieferdecker [17] presented a framework that organizes and categorizes the risk-based testing to aid the adoption of appropriate risk-based approaches according to the circumstances. Erdo- gan et al. [14] conducted a systematic literature review on the combined use of risk analysis and testing. This survey identified, classified, and discussed the existing approaches in terms of several factors such as main goals and the ma- turity levels of the approaches. For example, the survey discusses a model-based security testing approach proposed by Zech [51] using risk analysis for cloud computing environments. In Zech’s proposed approach, misuse cases are used on a model-driven approach for test code generation. These existing papers on risk-based testing demonstrate that the use of risks in the software systems can help find critical functional defects that may cause severe security or safety related issues. Another research area that is relevant to our work is fuzzy expert systems. Fuzzy expert systems have been used in areas that require expert knowledge to make decisions while minimizing several issues, such as uncertainties and subjectivity, in the decision-making process. In general, fuzzy expert systems are applied to various domains, such as diagnosing diseases in the medical field [4, 23], risk assessment in aviation [18], risk assessment in construction 7 projects [9], and selecting superior stocks on the stock exchange [15]. For instance, Adeli and Neshat [4] proposed a fuzzy expert system to diagnose heart disease. Recently, fuzzy expert systems have been used in software engineering areas such as software development effort prediction [5], software cost estimations [24], and risk analysis for e- commerce development [32]. For instance, Ahmed et al. [5] developed a fuzzy expert system to obtain accurate software cost and schedule estimation by managing the uncertainties and imprecision that exist in the early stages of software development. More recently, some researchers applied fuzzy expert systems to regression testing. Schwartz and Do [39] used a fuzzy expert system to determine the most cost-effective regression testing technique for different testing environments by addressing the limitations of existing, adaptive regression testing strategies. Xu et al. [46] applied a fuzzy expert system to deal with the inaccurate and subjective issues present during the test case selection process of regression testing. In this work, we used a fuzzy expert system with requirements and their risk information to improve risk estimation processes, thus improving the effectiveness of test case prioritization. 3 Proposed Approach In this section, we describe the proposed approach. Our new method consists of four main steps. 1. Estimate risks by correlating with requirements 2. Calculate the risk exposure for the requirements 3. Calculate the risk exposure for risk items 4. Prioritize requirements and test cases Figure 2 gives an overview of the proposed approach. The main steps are shown in light blue boxes, and the inputs and outputs for each step are shown in the ovals. The first three steps are used to calculate the requirements priorities. In the last step, test cases are prioritized using the results produced by the first three steps. The following subsections describe each step in detail by using an example that we excerpted from our experimental data which are fully described in Sections 4 and 5. 3.1 Estimate Risks by Correlating with the Requirements In order to perform risk assessment for the requirements, we identify four risk indicators that have been used by previous requirements and risk-based regression testing research [6, 21, 25, 41]: requirements complexity (RC), re- quirements size (RS), requirements modification status (RMS), and potential security threats (PST). These previous studies indicate that these risk indicators can be effective in finding defects in software systems, thus we focus on these four risk indicators in this study, but we consider to use other risk indicators, such as usage rate [16], as we evaluate our approach in the future. While obtaining the first two risk indicators is straightforward, the last two risk indicators can be subjective, so we utilize a fuzzy expert system to reduce the subjectivity and possible errors made by 8 human judgment. To calculate the first two risk indicators (RC and RS), we used both source code and requirements information, and for others (RMS and PST), we used requirements information. We explain each of these indicators in detail. Requirements Complexity (RC) Requirements that need complex functionalities during implementation tend to introduce more faults. A case study conducted by Amland [6] showed that requirements that need complex function- alities at the coding phase tend to introduce a higher number of defects. In addition, the study indicated that functions with a higher number of faults have a higher McCabe complexity. Therefore, in this research, we used McCabe complexity to measure the requirements complexity (RC). We measured the McCabe complexity values of software functionalities (source code) using Eclipse IDE. The complexity of requirements was determined by examining the relationships between requirements and functionalities. For the specific applications that we used with our experiment, the complexity values ranged from 1 to 12, and we normalized these values into a range from 0 to 10. A value of 0 indicated the lowest complexity, whereas 10 indicated the highest complexity for the requirements. The eighth column of Table 4 shows the example for the complexity values used for our experiment. Requirements Size (RS) To measure the requirements size (RS), the size of functions associated with the require- ments is used, and it is measured through lines of code (LOCs). Functions with higher LOC numbers tend to contain more defects. Amland’s case study [6] shows that the size of the functions could affect the number of faults in a system. In the experiment we performed, requirements size values range from 16 to 508, and these values are normalized into a range from 0 to 10, where a value of 0 indicates the lowest size and 10 indicates the highest size for the requirements. Requirements Calculate Req-Risks Calculate Risk Exposure for Risk Items Prioritize Req & Test Cases Prioritized Test Cases Weighted- Risk Exposure Risk Indicator Values Req-Risk Assessment RM L PVL CIA PAA Fu zzy Expert System RMS PST Complexity Size Req-Test Tracebility Req-1 Req-2 Req-n Test 1 Test 2 Test n Req-Risk Items Correlation Req-Risks Risk Items Figure 2: Overview of the risk-based approach 9 The last column of Table 4 shows the example for the RS values. Requirements Modification Status (RMS) To estimate the RMS, two aspects of requirements modifications are considered: requirements modification level (RML) and potential requirements volatility (PRV). These two variables are the inputs for the fuzzy expert system that we develop for this research, and the fuzzy expert system produces RMS using these two input variables. RMS reflects the overall modification status of each requirement. RML represents the degree of a requirement’s modification by comparing the same requirement in the previous version. However, manual comparison of requirements can be a subjective and time-consuming process. Therefore, to reduce the amount of time and subjectivity, we developed a program that uses cosine similarity [26] to measure similarities between requirements. The program compares two given requirements and produces the requirements modification level values (RML). This semi-automated approach helps eliminate some mistakes that may occur in the manual requirements- comparison process. Here, requirements modification includes existing requirements change and new requirements addition. The RML values are normalized into a range from 0 to 10, where 10 indicates the highest requirements modification level and 0 indicates no modification. New requirements for a subsequent version are automatically assigned the value of 10 because functionalities associated with the new requirements have a high possibility of introducing new faults to the system. After finishing this process, we examine the final RML values to ensure that the RML values obtained from the automated process reflect the actual modification levels for the requirements. The requirement’s PRV values are used to quantify the possibility of having requirements changes in later versions of the system. PRV values are measured through experts’ knowledge and experiences with requirements engineer- ing. Several requirements characteristics, such as functional instability, possible interface changes, and other factors that may influence the requirements modifications, are taken into account to estimate the PRV values. For example, consider a healthcare application. For such an application, requirements that calculate an insurance premium would change whenever the insurance policy changes. Thus, these requirements are assigned a higher PRV value compared to the requirements that handle patients’ information. The second and the third columns in Table 4 show the example for the RML and PRV values used in our experiment, and the fourth column shows the RMS values provided by the fuzzy expert system which is explained in Section 2.1. Potential Security Threats (PST). Today, software security is a major concern for software applications (in par- ticular, for web applications) due to the rapid growth of malicious activities, such as SQL injection, eavesdropping, etc., against software applications. Security flaws for a software application lead to severe consequences unless the security-related issues are identified and properly handled as early as possible. Therefore, in this approach, potential security threats (PST) is used as an indicator of security-related risks that reside in the requirements. Again, the fuzzy expert system is employed with a different rule set to estimate the PST values. The two input variables contain sets of software security objectives that are identified in the software security field [1, 34]. The first input variable contains the major security objectives, such as confidentiality, integrity and availability (CIA), and the second variable consists 10 of secondary security objectives, such as privacy, authentication and accountability (PAA). To estimate these input variable values for each requirement, we developed a term-extraction tool to find the number of security key words in a particular requirement that were related to the input variables. For example, suppose we identified 50 security key words associated with the first input variable (i.e., CIA) and a particular requirement contains 5 of the 50 keywords, then the CIA variable value of the requirement is 0.1 (5/50). Then, the input variable values obtained from the tool are normalized into a range from 0 to 10. A value of 0 indicates no association with CIA, whereas 10 indicates the highest association. The same technique is used to obtain the PAA values for each requirement. The fifth and sixth columns of Table 4 show the CIA and PAA values, respectively, and the seventh column shows the PST values provided by the fuzzy expert system. Table 4: Risk indicators and fuzzy input-output values Fuzzy input Fuzzy output Fuzzy input Fuzzy output Requirement RML PRV RMS CIA PAA PST Complexity Size UC1S1 0 2 4.6 7 6 8.1 1.9 5.8 UC2S1 0 3 4.6 7 9 8.3 1.9 4.3 UC8S1 5 6 5.0 5 7 6.0 1.9 2.8 UC10S1 6 5 5.0 6 8 8.3 1.0 1.6 UC11S1 2 5 5.0 5 7 4.5 2.8 9.4 UC23S3 3 5 5.0 6 6 5.7 1.9 10.0 UC26S3 0 3 4.6 4 5 3.2 2.8 4.9 ... ... ... ... ... ... ... ... ... UC34S6 10 8 8.4 5 6 5.2 1.0 1.7 3.2 Calculate the Risk Exposure for the Requirements Risk assessment for software requirements is performed by considering the probability of the risk occurrence (risk likelihood) for the risk indicators and the degree of possible damage (risk impact) with each indicator. We consider the four risk indicators explained in the previous step, and the weight for each indicator is determined using the analytic hierarchy process (AHP) that supports pairwise comparisons [37]. The comparison’s result values are normalized into a range from 1 to 5, and we obtain the final weight values for each indicator. In Table 5, columns two to five show the comparison values of risk indicators; sixth column shows the total; and the last column shows the priority vector (PV) values calculated for each indicator. Then, the PV values obtained from each expert are averaged. (In this example, two experts perform pairwise comparison.) The first column of Table 6 shows the risk indicators; the second and third columns show the PV values obtained from each expert; and the fourth column shows the averaged PV values. The last column shows the normalized weight values for the risk indicators. Equation 5 is used to calculate the risk exposure of a requirement (RE(Req)) by following a common risk- assessment practice used by several researchers (i.e., multiplication of risk likelihood and risk impact). Each risk 11 Table 5: Risk indicator comparison RMS Complexity PST Size Total Priority Vector First Expert’s Comparison RMS 0.37 0.37 0.38 0.35 0.37 0.37 Complexity 0.37 0.37 0.38 0.30 0.37 0.36 PST 0.18 0.18 0.19 0.30 0.18 0.21 Size 0.05 0.06 0.03 0.05 0.05 0.04 Second Expert’s Comparison RMS 0.38 0.32 0.48 0.38 1.56 0.39 Complexity 0.38 0.32 0.24 0.29 1.22 0.31 PST 0.19 0.32 0.24 0.29 1.03 0.26 Size 0.05 0.05 0.04 0.05 0.19 0.05 Table 6: Risk indicators and weights Indicator PV1 PV2 Average Weight Requirements Modification Status 0.37 0.39 0.38 5 Requirements Complexity 0.36 0.31 0.34 4 Potential Security Threats 0.21 0.26 0.24 3 Size 0.04 0.05 0.05 1 indicator value corresponds to the risk likelihood of a requirement, and the weight of the indicator corresponds to the risk impact. RE(Reqj) = n∑ i=1 (Wi ∗Rji) (5) where n is the number of indicators; Wi is the weight of the indicator; and RIji is the risk value of the requirement, Reqj , in terms of indicator i. The following example shows the risk calculation for the UC1S1 requirement used in our experiment (“The health- care personnel create a patient as a new user of the system.”). The last column of Table 7 shows the risks for a sample set of requirements. RE(UC1S1) = (5*4.6) + (4*1.9) + (3*8.1)+ (1*5.8)= 60.8 The risk values of the requirements RE(Req) range from 0 to 130, where 0 indicates the lowest risk and 130 indicates the highest risk. In this example, the risk of UC1S1 is 60.8, implying an average risk level. 3.3 Calculate the Risk Exposure of Risk Items In the previous subsection, the process for requirements risk exposure estimation was explained. These requirements risk exposure (RE(Req)) values indicate how risky each requirement is from the system requirements’ point of view. Although, the (RE(Req)) value of a requirement reflects the risk residing in the requirement, it is not sufficient to obtain information about the association between the requirement and potential defect types of a software system. When a particular requirement is associated with multiple defect types, that requirement has a high tendency to expose 12 Table 7: Risk indicator values and the risk exposure for requirements Requirement RMS Complexity PST Size RE(Req) 5 4 3 1 UC1S1 4.6 1.9 8.1 5.8 60.8 UC2S1 4.6 1.9 8.3 4.3 59.8 UC8S1 5.0 1.9 6.0 2.8 53.5 UC10S1 5.0 1.0 8.3 1.6 55.6 UC11S1 5.0 2.8 4.5 9.4 59.2 UC23S3 5.0 1.9 5.7 10.0 59.8 UC26S3 4.6 2.8 3.2 4.9 48.9 ... ... ... ... ... ... UC34S6 8.4 1.0 5.2 1.7 63.1 more defects. Thus, the association information between requirements and potential defect types help identify the requirements with a higher defect density which eventually cause common software failures. Therefore, high priorities can be assigned to the test cases that are correlated to such requirements with a higher defect density. To prioritize the test case using the association information between requirements and potential defect types, we adopted a process defined by Yoon et al. [48]. We improved this process by introducing weights for the risk items (RiIM). The weight values for the risk items were obtained from our previous requirements risk-based research [21], and for this research, we further calibrated the weight values by using the risk exposure values for risk items in Yoon et al.’s approach [48]. The risk items denoted potential defect types for a software system. The process of calculating risk exposure for risk items involved the following activities: (1) Identify Risk Items: The risk items employed in this research were derived from studies that used different applications and standards [8, 22, 45]. The identified risk items are shown in Table 8. The input problem was an example for risk items because input data can cause several problems for an operating software system. For instance, input data that were not validated before the execution or input data that were beyond the valid boundary resulted in operational failures for the system or a system crash. (2) Calculate risk exposure values for the risk items (RE(RiIM)): The risk level for a particular risk item was quantified by using the risk exposure values. To find the risk exposure values for risk items, first, we need to find the probability of fault occurrence by con- sidering the association between risk items and requirements. When we consider a set of requirements, some of them may be associated with a particular risk item. The probability of fault occurrence depends on the number of associated requirements. When a particular risk item is associated with multiple requirements, its probability for fault occurrence will increase. Second, we need to estimate the impact of the risk items. For this purpose, we consider the risk exposure of requirements (RE(Req)). When a particular risk item is associated with requirements that have higher risk exposure, the risk impact of that particular item increases. In this research, we ignored the Startup/ShutDown risk item which was used for our previous risk-based ap- proach [21] because Startup/ShutDown risk item only associated with a small number of requirements and also 13 did not make a significant impact on the test case prioritization. Additionally, in Yoon et al.’s [48] approach, the Startup/ShutDown risk item indicated a relatively low exposure value. Eliminating one risk item caused an approxi- mate 17% work reduction for this risk items’ risk exposure calculation subprocess. Table 8: Software product-risk items Risk Item Abbreviation i Input Problem IP ii Output Problem OP iii Calculation Calc iv Interactions Inac v Error Handling ErHa Table 9: Risk exposures and weighted risk exposure matrix Risk Items Requirement RE(Req) RiIM1 . . . RiIMx RiIMn W-RE (SV1) . . . (SVx) (SVn) Req1 RE(Req1) C11 C1x C1n W-RE(Req1) ... ... ... . . . ... ... ... Reqy RE(Reqy) Cy1 . . . Cyx Cyn W-RE(Reqy) Reqm RE(Reqm) Cm1 . . . Cmx Cmn W-RE(Reqm) RE(RiIM1) . . . RE(RiIMx) RE(RiIMn) Table 9 shows a matrix to calculate the risk exposure (RE(RiIM)) values for risk items and the weighted risk exposure (W-RE) values for requirements. The matrix lists requirements, risk exposure for requirements, and a set of risk items. Each risk item (RiIMx) has a severity value (SVx) that indicates how risky the item is. The severity values are defined in Table 10. If a risk item is associated with a certain requirement, the Cmx value is 1; otherwise, the value is 0. Again, multiple associations between risk times and requirements can exist. The last row in the matrix shows the RE values for risk items, and the RE values are calculated using Equation 6; the last column shows the W-RE values that are calculated by incorporating RE(RiIM) values with the severity values of risk items for each requirement. The final outcome of this step, the W-RE values, is used to prioritize requirements. Equations 6 and 7 are used to calculate the risk exposure values for risk items and the weighted risk exposure values for requirements. RE(RiIMx) = m∑ y=1 (RE(Reqy) ∗Cyx) (6) where m is the number of requirements, RE(Reqy) is the risk exposure of Reqy, and Cyx is 1 when requirement Reqy is associated with risk item RiIMx or 0 otherwise. 14 W −RE(Reqy) = n∑ x=1 (RE(RiIMx) ∗Cyx ∗SVx) (7) where n is the number of risk items for the system; RE(RiIMx) is the risk exposure value of the risk item, RiIMx; Cyx indicates the correlation between the requirement, Reqy, and the risk item, RiIMx; and SVx is the severity value of the risk item, RiIMx. Table 10: Severity of risk items Severity Value Description 1 Least critical risk item 2 Slightly critical risk item 3 Moderately critical risk item 4 Very critical risk item 5 Most critical risk item Table 11: Example of risk exposures and weighted risk exposures of iTrust Requirement TC ID RE(Req) OP ErHa Inac IP Calc W-RE Risk Items Severity Values 4 5 5 2 3 UC1S1 TC2 60.8 1 1 0 1 0 61384.6 UC2S1 TC5 59.8 1 1 1 1 0 88440.5 UC2S1 TC6 57.2 1 1 1 1 0 88440.5 UC8S1 TC13 53.5 1 1 0 1 1 68855.7 UC10S1 TC19 55.6 1 1 1 1 1 95911.6 UC11S1 TC21 59.2 1 1 1 0 0 79497.2 UC23S3 TC31 59.8 1 1 1 1 1 95911.6 UC26S3 TC33 48.9 1 1 1 0 0 79497.2 ... ... ... ... ... ... ... ... ... UC34S6 TC141 63.1 1 1 1 1 0 88440.5 Risk Exposure (RE(RiIM)) 6146.0 5571.5 5411.2 4472.0 2490.0 Table 11 shows a sample data set collected from our experiment. In this example, we can see that the output problem risk item is associated with all requirements. Thus, the risk exposure value of the output problem risk item can be calculated using Equation 6 as follows: RE(OP ) = (60.8*1) + (59.8*1) + (53.5*1) + (55.6*1) + . . . + (63.1*1) = 6146.0 Because the output problem has a high RE value compared to other risk items, it implies that the output problem is a high risk area for this product. After calculating the RE values for risk items, we calculate the W-RE values for each requirement by utilizing Equation 7. For instance, we obtain the W-RE value for UC1S1 as follows. W-RE(UC1S1) = (6146.0*1*4) + (5571.5*1*5) + (5411.2*0*5) + (4472.0*1*2) + (2490.0*0*3) = 61384.6 15 After calculating the W-RE values for each requirement, we prioritize requirements by their W-RE values (in descending order). Table 12 shows a portion of our data. From the table, we can see that each requirement has one or more corresponding test cases (The last subsection explains how to map these two.) and that the requirements appear by their W-RE values (in descending order). When multiple requirements have the same W-RE value (e.g., UC33S1 and UC26S2 have the same value, 7250.3, in the table.), we used the RE(Req) values as the second factor to prioritize the requirements. The next section describes, in detail, the prioritization of the test cases. Table 12: Example for a prioritized test suite: iTrust Requirement TC-ID W-RE Re(Req) UC33S1 TC117 7250.3 76.5 UC26S2 TC94 7250.3 70.3 UC21 TC30 7090.6 80.2 UC21 TC80 7090.6 70.1 ... ... ... ... UC12 TC26 4735.4 80.0 UC30S3 TC100 4236.6 67.2 UC30S1 TC99 4236.6 56.3 UC3S3 TC43 2947.6 33.3 UC24 TC91 2860.2 42.5 UC25 TC92 2860.2 20.8 ... ... ... ... UC2S2 TC77 1787.7 66.5 3.4 Prioritize the Requirements and Test Cases In this final stage, we prioritize requirements using the W-RE values and then by the the risk exposure of requirements (RE(Req)). To prioritize test cases, we need to know the relationship between the test cases and the requirements. In this research, we use the traceability matrices created by the object programs’ (i.e., iTrust and Capstone) developers. Unavailability of traceability matrices could require more effort to adopt our proposed approach. However, often, the documents that are created in the early software development stages, such as business requirements document and functional specification document, are used when creating test cases, which can help construct traceability informa- tion between requirements and test cases. Further, some industrial tools such as Qmetry [2] and Microsoft Testing Manager [3] can aid establishing traceability. For most software applications, including our experimental programs, a single requirement is tested with multiple test cases. As a result, test cases associated with the same requirement have the same W-RE value. However, the different functions used to manipulate a particular requirement have different complexity levels and LOCs, hence producing different RE(Req) values. Therefore, after mapping test cases to their corresponding requirements and functions, the test cases with similar W-RE values can be re-prioritized using the RE(Req) values. If test cases still 16 have the same W-RE and RE(Req) values, then we randomly order those test cases. 4 Empirical Study To investigate the effectiveness of our new risk-based test case prioritization approach, we designed and performed a controlled experiment. The following subsections describe research questions, the objects of analysis, independent variables, dependent variables and measures, experimental setup and procedure, and threats to validity. 4.1 Research Questions In this study, we investigate the following research questions: RQ1: Can systematic risk-based test case prioritization improve the rate of fault detection for test suites? RQ2: Can systematic risk-based test case prioritization find more faults in the risky components early? 4.2 Objects of Analysis In order to evaluate the new approach, we utilized two applications: one open source and one industrial application developed for graduate-student project. iTrust is the open source application used for this experiment. The iTrust program is a patient-centric electronic health record system which was developed by the Realsearch Research Group at North Carolina State University. In this experiment, four versions of the iTrust application (versions 0, 1, 2, and 3) were used. We considered the functional requirements of each version, and the test cases were used to test the system functionalities associated with the system requirements. For iTrust, all the test cases used in the experiment were developed by the iTrust system developers. The industrial application, Capstone, was developed by computer science graduate students at North Dakota State University in collaboration with a local software company. Capstone is an online application that is used to automate the company’s examination procedure. The information about the system components for all versions of the two applications used in this experiment is shown in Table 13. For each system, version 0 (base version) is not listed in the table because regression testing starts with the second version. However, the information obtained from version 0 is utilized to obtain the mutants for version 1. Table 13: Experiment objects and associated data Object Version Requirements Test Cases Size (KLOCs) Mutation Faults Mutation Groups iTrust V1 91 122 24.42 54 13 V2 105 142 25.93 71 12 V3 108 157 26.70 75 12 Capstone V1 21 42 6.82 118 23 17 4.3 Variables and Measures Independent Variable The test case prioritization technique is the independent variable manipulated in this study. We consider seven control techniques, including a requirements risk-based technique and one heuristic prioritization technique, as follows: • Control Techniques – Original (Torig): The object program provides the testing scripts. Torig executes test cases in the order in which they are available in the original testing script. – Total statement coverage (Tsc): This technique prioritizes test cases based on the total number of state- ments exercised by test cases. – Code metric (Tcm): This technique uses a code metric that we defined in our previous study [7]. The code metric is calculated using three types of information obtained from the source code, Line of Code (LOC), Nested Block Depth (NBD), and McCabe Cyclomatic Complexity (MCC), which are considered good predictors for finding error-prone modules [38, 50].1 – Requirements-based clustering: We consider another set of techniques proposed by Arafeen and Do [7] as control techniques that follow the requirements-based clustering approach. Because all requirements are not equally important to clients, requirements clustering is important to prioritize the requirements based on their importance to the client so that the tester can pay more attention to the test cases associated with high-priority requirements. With the previous approach, test cases are clustered based on requirements and test case association. Within clusters, test cases are prioritized using code metric information or use the original test case order. Thus, all techniques are categorized into two broad categories based on the original test case order and the code metric test case order. In this research, we only considered the code metric based category which produced relatively better results compared to original order category. Then, the clusters are ordered in three ways: original cluster order, random cluster order, and prioritized cluster order. Clusters are prioritized using both requirements commitment (i.e., the priority of requirements to be implemented by the developers) and code modification information. After ordering the test cases and the clusters, test cases are selected from the clusters in a round-robin fashion for the prioritization. The three requirements-based clustering techniques are as follows: ∗ Tcl-cm-orig (Tcco): It uses the code metric based test case order within clusters, and the clusters are ordered according to the original order of the clusters. ∗ Tcl-cm-rand (Tccr): This technique uses the code metric based test case order within clusters, and the clusters are ordered randomly. ∗ Tcl-cm-prior (Tccp): This technique uses the code metric based test case order within clusters, and 1Tcm = NBD Max(NBD) + MCC Max(MCC) + LOC Max(LOC) 18 the clusters are ordered according to requirements commitment and code modification information. This requirements cluster-based approach [7] used five different cluster sizes for the iTrust application. In this study, we only considered cluster sizes 10 and 20 which produced moderate and the best results, respectively. In the case of Capstone, the only cluster size used in requirements cluster-based approach [7] was cluster size 6. – Requirements risk-based (Trrb): This technique prioritizes test cases based on the risks residing in the requirements and the association between a system’s requirements and potential defect types. • Heuristic (Tfrrb): The proposed technique uses a fuzzy expert system to estimate requirements risks and priori- tizes the test cases as described in Section 3. Dependent Variable and Measures We considered two dependent variables: • Average Percentage of Fault Detection (APFD): The APFD [12, 29] value represents the average for the percentage of fault detection during the execution of a particular test suite. APFD values range from 0 to 100. The prioritization techniques are being considered as better techniques when their APFD values are closer to 100, and the technique that obtains the highest APFD value is considered to be the best prioritization technique. • Percentage of Total Risk Severity Weight (PTRSW): PTRSW is used to measure the effectiveness of test suites in terms of finding more faults for risky components in a particular system as early as possible. The PTRSW values range from 0% to 100%. Equation 8 shows how to calculate the PTRSW value. When a test suite can detect all faults for risky components in a system at a particular test case execution rate, then the PTRSW becomes 100% at that particular execution rate. A test case prioritization technique can be considered as effective in finding more faults for risky components if the test suite produced by that technique is capable of achieving a higher PTRSW value for a lower test case execution rate. For example, if the PTRSW value of a test suite produced by prioritization technique A is 100% when half of the test cases are executed (i.e., 50% test execution rate), whereas the prioritized test suite for technique B produces a 70% PTRSW value for the same 50% test execution rate, then technique A is considered the most effective technique to find more faults in the risky components. PTRSW = (STRSW/GTRSW)∗100% (8) The sub-total of the total risk severity weight (STRSW) and the grand total of total risk severity weight (GTRSW) are explained, in detail, in the next subsection. 19 4.4 Experimental Setup and Procedure In this study, for each requirement, we estimated the risk exposure using a fuzzy expert system and then estimated the weighted risk exposure (W-RE) values for the requirements as described in Section 3. We prioritized the requirements based on the W-RE values and then by the risk exposure of requirements (RE(Req)) values of the requirements. To obtain the prioritized test cases, we used the requirements-test mapping information which was provided by the object programs’ developers. To empirically evaluate the proposed approach, we also need fault data. Due to the absence of the faults with the applications, we used a set of mutation faults created for each object program during previous research [7]. The second-last column in Table 13 lists the total number of mutation faults for each program version. In actual testing scenarios, however, programs do not typically contain this many faults. Thus, to reflect more realistic situations, our previous study [11] introduced mutant groups which were formed by randomly selecting mutants from the pools of mutants created for each version, ranging from 1 to 10 mutants per group. For the iTrust application, 13, 12, and 12 mutation groups were created for version 1, 2, and 3, respectively, whereas 23 mutation groups were created for Capstone version 1. To perform the risk estimation process, we followed the steps described in Section 3. As a human expert, one graduate student who has several years of software industry experience performed the risk estimation process. To obtain the requirements modification level (RML); confidentiality, integrity, and availability (CIA) values; and privacy, authentication, and accountability (PAA) values, the student followed the process explained in Section 3. When the final RML, CIA, and PAA values were obtained, those values were reviewed by the student to validate the inaccuracy. On average, only 3% of the requirements needed slight adjustments for their RML values, and less than 2% of the requirements required minor adjustments for the CIA and PAA values. The potential requirements volatility (PRV) values for the requirements were estimated based on the expert’s knowledge and experience. The weight values used for the risk indicators were determined through the priority vector (PV) values of analytic hierarchy process (AHP) that was performed by two experts (graduate students). We averaged and normalized the priority vector values obtained from each expert for every indicator to obtain the final weight values for risk indicators. Test cases were prioritized using W-RE and RE(Req), and with test cases that had the same W-RE and RE(Req) values, those test cases were randomly prioritized. For all the applications and versions used for this experiment, on average, only 3% of the test cases required random orders. To calculate the PTRSW values for each test case prioritization technique, we need know the classes and methods modified by the mutation faults. Therefore, we used a mutation analysis tool, ByteME (Bytecode Mutation Engine, a tool from the software-artifact infrastructure repository (SIR) [12]), to locate these altered classes/methods and devel- oped the faults-classes/methods trace files for each version of the object programs. Further, we used the requirements- classes/methods trace files to identify the association relationship between requirements and classes/methods. We estimated the risks of classes/methods caused by mutation faults, risk severity weight (RSW), using the information 20 provided by ByteME and the requirements risks. The RSW values estimated for classes/methods were assigned to the associated mutation faults. The RSW values ranged from 0 to 10. The RSW value was 0 if there was no risk caused by a mutation fault for a class/method, whereas 10 indicated the highest risk caused by a mutation fault on a critical system component. The faults-tests traceability matrix is used to identify the relationship between mutation faults and test cases. One mutation fault can be detected by one or more test cases. In that case, the same RSW value is assigned to all test cases associated with the same mutation fault. On the other hand, one test case can detect several mutation faults, and all RSW values for the detected mutation faults are added together to obtain the total risk severity weight (TRSW) for that test case. The sum of the TRSW values of all test cases produces the grand total risk severity weight (GTRSW) for the test suite. When developing software applications, software companies often face time and budget constraints, and typically, the companies cut back on testing activities to ensure a timely release for their product. When testing activities are cut short, faults will slip through testing. If testing techniques can detect riskier defects earlier, then the companies could reduce potential severe consequences. To examine this situation, we consider four different test execution rates as shown in Table 14. For example, the STRSW values for the original test order of iTrust version 2 are 4 and 19 for test execution rates of 12.5% and 25%, respectively. Table 14 shows sample RSW, TRSW, STRSW, GTRSW, and PTRSW values for the original test case order of iTrust version 2. The first column shows the test cases; columns 2, 3, and 4 show the sample set of RSW values for the test cases; column 5 shows TRSW and STRSW; and the last column shows the PTRSW values for different test execution rates. For example, the STRSW for the 50% test execution rate is 225, and the PTRSW is 46.68%. When we compare these data with the results of other techniques shown in Table 18, these values are relatively low, meaning that the original order of test cases is unable to detect faults in the risky components effectively. After gathering all required data, we obtained the prioritized test suite by following the proposed approach. We calculated the APFD and PTRSW values for the prioritized test suite. 4.5 Threats to Validity This section describes the internal, external, and construct threats to the validity of our study, and the approaches we used to limit their effects. Internal Validity: In this study, we estimated the risks residing in the functional requirements in terms of product risks and did not consider other types of software risks, such as project and process risks. This threat can be minimized by adding more appropriate risk indicators in the requirements risk estimation process. We estimated the complexity and the size of a particular requirement using the McCabe Cyclomatic Complexity (MCC) and Lines of Code (LOC) for the class/method utilized to implement that particular requirement. There are other alternatives available to measure 21 Table 14: Example test cases, PTRSW, and associated data for iTrust original test order-version 2 Test cases (Orig v2) RSW of Mutations TRSW Percentage of Total Risk Severity Weight M1 ... M71 (PTRSW) TC7 3 0 1 4 ... ... ... ... ... TC18 0 0 0 0 ... ... ... ... ... Sub-total of TRSW (STRSW) after 12.5% test case execution 4 0.83% ... ... ... ... ... TC20 0 1 8 10 Sub-total of TRSW (STRSW) after 25% test case execution 19 3.94% ... ... ... ... ... TC53 0 2 2 5 Sub-total of TRSW (STRSW) after 50% test case execution 225 46.68% ... ... ... ... ... TC106 0 9 8 42 Sub-total of TRSW (STRSW) after 75% test case execution 276 57.26% ... ... ... ... ... TC136 0 8 8 24 ... ... ... ... ... Grand total after execution the entire test suite (GTRSW) 482 the complexity and the size, and the results could be affected by the choice of different alternatives. However, much of the previous research has shown that MCC and LOC are good indicators to measure the complexity and size with ease. Moreover, the crisp output obtained from the fuzzy expert system may vary due to several factors, such as the number of input/output membership functions, the types of waveforms, and the defuzzification methods. The fuzzy expert system used for this research is based on three triangular membership functions for both input and output variables because many previous fuzzy expert systems used a similar configuration successfully. We utilized the center of gravity (COG) method for the defuzzification because it is a widely used defuzzification method and is also considered as accurate. However, further studies can be performed to investigate the effectiveness of the proposed approach with different configurations. External Validity: The object programs we used in this research are a small, industrial application (Capstone) and a mid-size, open source application (iTrust). Therefore, our findings cannot be interpreted in the context of large applications. This limitation can be addressed by applying our approach to large, open source and industrial applications. Construct Validity: Estimating the potential requirements volatility (PRV) for requirements and determining the 22 correlation between requirements and risks items needed human involvement. Because human judgment is based on several factors, such as human experts’ knowledge and experiences, the results could vary from person to person. We formulated and validated the rules of our fuzzy expert system based on our experiences and knowledge, but a fuzzy expert system with fewer or more rules could be developed and potentially change the results. 5 Data and Analysis In this section, we present the results of our study and the data analyses for each research question. We discuss further implications of the data and results in Section 6. 5.1 The effectiveness of risk-based prioritization with a fuzzy expert system for improving the fault detection rate of test cases (RQ1) In our first research question (RQ1), we consider whether the risk-based approach that incorporates a fuzzy expert system can help improve the effectiveness of test case prioritization. To answer this question, we compared techniques based on the results shown in Tables 15, 16, and 17. Table 15 shows the APFD values for our heuristic technique (Tfrrb). Columns 2 to 4 show the APFD values for iTrust versions 1, 2, and 3, respectively, and the last column shows the APFD value for the Capstone application. The first column of Table 16 shows the control techniques. CS10 and CS20 indicate the two cluster sizes. All APFD values shown in Tables 16 and 17 are averaged values for the 13, 12, and 12 data points of iTrust versions 1, 2, and 3, respectively, and the 23 data points for Capstone version 1. Table 16 and 17 also show the improvement rates for our proposed approach over the control techniques. The results in Table 16 indicate that our approach (Tfrrb) outperformed the original (Torig), statement coverage (Tsc), code metric (Tcm), and our previous requirements risk-based approach (Trrb) across all versions. The improve- ment rates ranged from 5% to 175.83%. However, when we compared our approach to the cluster-based approaches, the trend varied across versions. For version 1, the heuristic outperformed the control techniques with a cluster size of 10 (The improvement rates ranged from 59.45% to 63.81%.), but it was not better than the techniques with a cluster size of 20. In the case of version 2, the results were reversed. The heuristic performed slightly better than the control techniques with a cluster size of 20, and it performed slightly worse than the control with a cluster size of 10. In the case of version 3, the heuristic outperformed all cluster-based control techniques. The results for Capstone shown in Table 17 indicate that the proposed approach outperformed all control tech- niques except for the Tccp technique. The heuristic and Tccp produce the same results. For this application, the improvement rates range from 0.00% to 55.26%. To show our results visually, we present them in boxplots. Figure 3 presents the boxplots that show APFD values for the control techniques and heuristic for all iTrust and Capstone versions. For iTrust version 1, each boxplot has 13 data points, and for versions 2 and 3, each has 12 data points. In the case of Capstone, each boxplot has 23 data points. 23 Table 15: Heuristic APFD: iTrust and Capstone itrust Capstone Version V1 V2 V3 V1 APFD (Tfrrb) 63 65.78 79.44 67.86 Table 16: APFD comparison and improvement over controls: iTrust Control iTrust - V1 iTrust - V2 iTrust - V3 Technique APFD Improvement Over APFD Improvement Over APFD Improvement Over Control (%) Control (%) Control (%) Torig 43.70 44.16 47.80 37.62 28.80 175.83 Tsc 53.02 18.82 56.75 15.91 69.26 14.70 Tcm 45.80 37.55 48.80 34.81 44.84 77.18 Trrb 60.00 5.00 60.60 8.55 56.00 41.86 Tcco -CS10 38.46 63.81 68.26 -3.63 68.48 16.00 Tccr -CS10 38.88 62.04 68.30 -3.69 64.54 23.09 Tccp -CS10 39.51 59.45 57.78 13.85 75.31 5.48 Tcco -CS20 65.58 -3.93 62.67 4.96 66.69 19.12 Tccr-CS20 66.22 -4.86 63.65 3.35 67.68 17.38 Tccp-CS20 75.45 -16.50 65.33 0.69 72.41 9.71 Each subfigure for iTrust contains boxplots for ten prioritization techniques; the first nine boxplots present data for the control techniques, and the last one presents the heuristic technique. The last subfigure, Capstone, contains boxplots for seven prioritization techniques; the first seven boxplots present data for the control techniques, and the last one presents the heuristic technique. When we examine the boxplots for iTrust, in version 1, the results for requirements risk-based approaches (Trrb and Tfrrb) show a wider distribution of data points than the other two versions. The results with other techniques for all versions show similar data-distribution patterns except for a couple cases. For version 3, the control techniques that used clustering with a cluster size of 10 show a wider distribution than other techniques. Overall, the heuristic shows better results compared to the controls across all versions of iTrust except for a few cases (clustering-based approaches for version 2). In the case of Capstone, all techniques show a similar data-distribution pattern, and the heuristic produces the best median value (indicated with a line in the box) compared to the control techniques and the best average value (indicated with a diamond) except for one case (Tcm). 5.2 The effectiveness of risk-based prioritization with a fuzzy expert system to find faults in risky components (RQ2) For the first research question, we consider whether the risk-based approach that incorporates a fuzzy expert system can find faults earlier. The results are encouraging, but we do not know whether the early detected faults are, indeed, the faults that reside in the risky components. Thus, our second research question (RQ2) considers whether the risk- based approach with a fuzzy expert system can be more effective at finding faults associated with risky components early compared to the control techniques. To evaluate RQ2, we consider five control techniques: Torig, Tsc, Tcm, Trrb, 24 Table 17: APFD comparison and improvement over controls: Capstone Control Technique Capstone - V1 APFD Improvement Over Control (%) Torig 43.70 55.26 Tsc 55.19 22.96 Tcm 67.80 0.07 Trrb 62.88 7.90 Tcco 61.19 10.89 Tccr 61.47 10.37 Tccp 67.86 0.00 iTrust version 1 iTrust version 2 iTrust version 3 Capstone version 1 Figure 3: APFD boxplots for all controls and the heuristic: iTrust and Capstone and a clustering technique that produced the best results for RQ1 (i.e., Tccp-CS20 for iTrust and Tccp for Capstone). To address this research question, we measured the percentage of the total risk severity weight (PTRSW) values that we described in Section 4.3, showing how fast the technique can detect faults in the risky components. Tables 18 and 19 show the PTRSW values for all versions of iTrust and Capstone, respectively. The first column of Tables 18 and 19 shows the version number, and the second column categorizes the control techniques and heuristic. The third column shows the prioritization techniques. The subsequent columns show the PTRSW values when we foreshorten 25 the test execution process by the specified percentage for each column. For example, the sixth column shows the PTRSW values when 50% of the test cases are executed, meaning that, for a 50% cutting ratio, we simulate the effects of having the testing process halted halfway through. The results showed that our approach can detect more faults in the risky components earlier than the controls except for a few cases (Trrb in version 1 for iTrust). For version 1, the heuristic (Tfrrb) and the requirements risk-based approach (Trrb) showed very similar results, but for versions 2 and 3, the trend changed. For version 2, the heuristic produced relatively high fault detection rates when test execution rates were low, and for version 3, the differences between these two techniques were more outstanding. For instance, at a 25% execution rate, the heuristic produced 70.64%, but Trrb produced only 18.94%. In the case of Capstone (Table 19), our approach outperformed all control techniques. These results indicated that using requirements risks with a fuzzy expert system during prioritization was effective in locating faults that reveal risks early. Further, even when companies need to cut their testing process short due to their product release schedule, they can still identify and fix more important faults with the limited time and budget than otherwise. Table 18: Percentage of total risk severity weight (PTRSW) for different test execution levels: iTrust Version Technique Percentage of Total Risk Severity Weight (PTRSW) for Different Test Execution Levels: iTrust Execution Rate 12.5% Execution Rate 25% Execution Rate 50% Execution Rate 75.0% V1 Controls Torig 0.00 0.00 26.49 70.90 Tsc 12.31 26.07 70.51 76.49 Tcm 1.87 5.22 29.85 52.24 Tccp-CS20 7.46 7.46 30.97 71.27 Trrb 21.64 30.22 67.91 82.09 Heuristic Tfrrb 21.64 26.12 70.52 79.48 V2 Controls Torig 0.83 3.94 46.68 57.26 Tsc 42.53 43.36 58.51 95.13 Tcm 11.00 26.97 67.63 74.90 Tccp-CS20 0.83 17.22 67.43 74.48 Trrb 10.37 26.35 68.46 95.44 Heuristic Tfrrb 42.74 43.78 73.24 95.44 V3 Controls Torig 0.00 0.85 25.32 29.36 Tsc 46.98 46.98 47.44 99.07 Tcm 0.43 10.43 19.36 65.32 Tccp-CS20 45.53 45.53 70.85 90.00 Trrb 0.00 18.94 66.38 86.38 Heuristic Tfrrb 49.79 70.64 97.66 99.57 Figure 4 presents the results graphically. The first three subgraphs show the results for iTrust, and the last graph shows the PTRSW comparison for Capstone. As we observed from Tables 18 and 19, our approach outperforms all control techniques for version 1 except for one technique (Trrb), and for versions 2 and 3, our approach outperforms the controls at all test execution rates. The trend for version 3’s results is different from other versions. The results of the techniques for version 3 vary widely across test execution rates. In particular, the clustering-based approach performs 26 Table 19: Percentage of total risk severity weight (PTRSW) for different test execution levels: Capstone Version Technique Percentage of Total Risk Severity Weight (PTRSW) for Different Test Execution Levels: Capstone Execution Rate 12.5% Execution Rate 25% Execution Rate 50% Execution Rate 75% V1 Controls Torig 11.51 16.43 54.26 75.30 Tsc 24.15 37.36 54.68 77.45 Tcm 24.33 39.35 66.04 83.67 Tccp 11.20 23.55 52.59 79.75 Trrb 22.87 37.10 58.03 84.67 Heuristic Tfrrb 24.49 40.08 66.77 85.45 Figure 4: PTRSW comparison graphs for all versions of iTrust and Capstone better than all other control techniques. In the case of Capstone, the heuristic outperforms all control techniques across all test execution rates. Similar to version 1 of iTrust, the heuristic and the requirements risk-based approach (Trrb) produce very similar results. 6 Discussion and Implications In this section, we present more insight about the findings of our research and possible implications. Through the results we obtained with our study, we draw the following observations. First, the proposed systematic, risk-based approach is capable of outdoing the original test order, code-metric-based approach and our previous requirements 27 risk-based approach for all versions of the iTrust application. Cluster-based approaches with a cluster size of 20 out- perform our proposed approach for version 1 while two cluster-based approaches with a cluster size of 10 outperform our approach for version 2. However, our proposed approach outperforms all cluster-based approaches for version 3. The source code of version 3 is significantly affected by the requirements modifications for version 3. In our previous requirements risk-based approach [21], we did not get better results for version 3 compared to cluster-based approaches because we only used one code-related risk factor, line of code (LOC), to estimate requirements risk. However, in this research, we use both LOC and McCabe Complexity, and we speculate that using more source code data to extract risk information has affected the better results obtained for version 3 that underwent major code modifications. In the case of Capstone, our approach outperforms all control techniques while producing the same result as the Tccp approach. Overall, the proposed approach produces very effective results across all versions of the iTrust and Capstone programs in spite of imprecise, inconsistent, and complex conditions that may occur when extracting the risk information from software requirements. Using a fuzzy expert system contributes to handling such circumstances successfully and fa- cilitates making more realistic risk estimations. Further, we emulate expert thinking in the risk estimation procedure by using our fuzzy expert system and minimize subjectivity in the risk estimation procedure. Additionally, we employ a semi-automated process to assess the risk factors for our risk estimation procedure. Hence, having a systematic risk estimation procedure is the most possible reason for getting better results with all versions of every object program. Second, the results indicated that our proposed approach has the ability to detect more faults early in risky com- ponents of software applications. In particular, for the third version of iTrust, our proposed approach detected a significantly higher number of faults in risky components at low test execution rates compared to the first and the second versions. Again, the requirements modifications in version 3 and their effect on the source code are a possi- ble reason for this outcome. In the case of Capstone, our new approach produced the best results over other control techniques. In the proposed approach, requirements were primarily prioritized by considering their direct relationship with critical risk items, and then, the requirements were further prioritized using requirements risks which were esti- mated through a less subjective, fuzzy expert system based approach. Because the test cases are prioritized using the association between requirements and test cases, the top test cases in the prioritized test suite were able to detect faults in the high-risk components early. Therefore, in this research, considering the direct relationship between critical risk items and test cases was the major reason for the early detection of more faults in high-risk components. The results of this research indicate very important implications for the software industry. Modern software sys- tems which are developed in today’s software industry are very complex and are vulnerable to malicious attacks; frequent changes are inevitable to maintain the systems’ quality. Software systems which are intended to be available online (web-based) and the systems which are considered as mission critical are even more complex, undergo frequent changes, and have much bigger concerns for their security-related issues. The proposed approach pays much attention to these factors in terms of fault detection. Hence, by adopting our proposed approach, software development compa- nies can detect faults in their modern applications within a short time frame. Furthermore, if a company happens to 28 shorten their regression testing process due to any constrains (time, budget, etc.), they can still detect more faults in their applications, including faults in high-risk components, with the limited resources. In particular, early detection of more defects in critical systems is very important because such defects can eventually lead to severe failures, such as life-threatening conditions or huge financial loss. Therefore, using our proposed approach, companies can develop their software systems with more confidence and cost-effectiveness while meeting their tight production deadlines. 7 Conclusions and Future Work In this paper, we presented a systematic risk estimation approach using a fuzzy expert system which can minimize the subjectivity, imprecision, and inconsistency issues confronted by the requirements risks estimation process. We empirically evaluated the new approach using two Java applications with multiple versions. The results of this study demonstrated that our new systematic, risk-based approach can detect faults earlier and is even better at finding faults in the risky components earlier than the control techniques. With the proposed approach, software companies can manage their testing and release schedules better by providing early feedback to testers and developers so that the development team can fix the problems as soon as possible. While we addressed the limitations of our previous approach, as discussed in Section 4.5, there are still some limitations that need to be overcome. For example, determining the relationships between requirements and risk items is done by human experts, but this process can be improved by reducing human involvement through semi-automated approaches (e.g., semantic analysis of natural language). Another limitation involves the choices of membership functions and the defuzzification method. We considered three triangular membership functions for both the input and output variables and used the centroid defuzzification method to defuzzify output variables, but the choice of these functions and methods could affect our results. Therefore, we plan to conduct more studies by considering different wave types (i.e., trapezoidal, gaussian, etc.), different numbers of membership functions, and different defuzzification methods to see how these variations affect the requirements risk estimation and final results. Further, in this work, we only used four risk indicators, but we plan to use other risk indicators, such as usage rate. We also plan to investigate the use of other fuzzy technologies, such as the fuzzy clustering approach, to improve risk-based regression testing approaches, and plan to evaluate these approaches considering their efficiency and effectiveness. Acknowledgments This work was supported, in part, by NSF CAREER Award CCF-1149389 to University of North Texas. This work was also funded by the MSIP (Ministry of Science, ICT, and Future Planning), Korea, under the ITRC (Informa- tion Technology Research Center) support program (NIPA-2014-H0301-14-1023) supervised by NIPA (National IT Industry Promotion Agency) 29 References [1] Minimum Security Requirements for Federal Information and Information Systems . http://csrc.nist.gov/publications/fips/fips200/FIPS-200-final-march.pdf, 2006. [2] QMetry Test Management. http://www.qmetry.com, 2014. [3] Quick Start Guide for Manual Testing using Microsoft Test Manager. https://msdn.microsoft.com/en- us/library/vstudio/dd3807632014. [4] A. Adeli and M. Neshat. A fuzzy expert system for heart disease diagnosis. International MultiConference of Engineers and Computer Scientists (IMECS), 1:1–7, 2010. [5] M.A. Ahmed, M.O. Saliu, and J. AlGhamdi. Adaptive fuzzy logic-based framework for software development effort prediction. Information and Software Technology, 47(1):31–48, 2005. [6] S. Amland. Risk-based testing: Risk analysis fundamentals and metrics for software testing including a financial application case study. Journal of Systems and Software, 53(3):287–295, 2000. [7] M.J. Arafeen and H. Do. Test case prioritization using requirements-based clustering. International Conference of Software Testing, Verification and Validation (ICST), March 2013. [8] J. Bach. Risk and requirements-based testing. IEEE Computer, 32(6):113–114, 1999. [9] V. Carr and J.H.M. Tah. A fuzzy approach to constuction project risk assessment and analysis: construction project risk management system. Advances in Engineering Software, 32(10-11):847–857, 2001. [10] H. Do, S. Mirarab, L. Tahvildari, and G. Rothermel. The effects of time constraints on test case prioritization: A series of controlled experiments. IEEE TSE, 26(5), September 2010. [11] H. Do and G. Rothermel. An empirical study of regression testing techniques incorporating context and lifecycle factors and improved cost-benefit models. In Proceedings of the ACM SIGSOFT Symposium on Foundations of Software Engineering, November 2006. [12] H. Do and G. Rothermel. On the use of mutation faults in empirical assessments of test case prioritization techniques. IEEE Transactions on Software Engineering, 32(9):733–752, September 2006. [13] S. Elbaum, A. G. Malishevsky, and G. Rothermel. Test case prioritization: A family of empirical studies. IEEE TSE, 28(2):159–182, February 2002. [14] G. Erdogan, Y. Li, R. Runde, F. Seehusen, and K. Stlen. Approaches for the combined use of risk analysis and testing: a systematic literature review. International Journal on Software Tools for Technology Transfer, 16(5):627–642, 2014. 30 [15] M. Fasanghari and G.A. Montazer. Design and implementation of fuzzy expert system for tehran stock exchange portfolio recommendation. Expert Systems with Applications, 37(9):6138–6147, 2010. [16] M. Felderer and R. Ramler. Integrating risk-based testing in industrial test processes. Software Quality Journal, 22(3):543–575, 2014. [17] M. Felderer and I. Schieferdecker. A taxonomy of risk-based testing. International Journal on Software Tools for Technology Transfer, 16(5):559–568, 2014. [18] M. Hadjimichale. A fuzzy expert system for aviation risk assessment. Expert Systems with Applications, 3:6512– 6519, 2009. [19] V. Hajipour, A. Kazemi, and S. M. Mousavi. A fuzzy expert system to increase accuracy and precision in measurement system analysis. Measurement, 46(8):2770–2780, 2013. [20] M. J. Harrold, D. Rosenblum, G. Rothermel, and E. Weyuker. Empirical studies of a prediction model for regression test selection. IEEE TSE, 27(3):248–263, March 2001. [21] C.S. Hettiarachchi, H. Do, and B Choi. Effective regression testing using requirements and risks. In Eighth International Conference on Software Security and Reliability, pages 157–166, June 2014. [22] IEEE. IEEE Guide to Classification for Software Anomalies. Std 1044.1-1995. Institute of Electrical and Elec- tronics Engineers, Inc, 1996. [23] M.A. Kadhim, M.A. Alam, and H. Kaur. Design and implementation of fuzzy expert system for back pain diagnosis. International Journal of Innovative Technology and Creative Engineering, 1:16–22, 2011. [24] M. Kazemifard, A. Zaeri, N. Ghasem-Aghaee, M.A. Nematbakhsh, and F. Mardukhi. Fuzzy emotional cocomo ii software cost estimation (fecsce) using multi-agent systems. Applied Soft Computing, 11(2):22602270, 2011. [25] R. Krishnamoorthi and S.A. Sahaaya Arul Mary. Factor oriented requirement coverage based system test case prioritization of new and regression test cases. Information and Software Technology, 51(4):799–808, 2009. [26] G. Kumaran and J. Allan. Text classification and named entities for new event detection. In Proceedings of the 27th annual international ACM SIGIR conference on Research and development in information retrieval, pages 297–304, July 2004. [27] Q. Li, Y. Yang, M. Li, Q. Wang, B. W. Boehm, and C. Hu. Improving software testing process: feature prioritiza- tion to make winners of successcritical stakeholders. Journal of Software: Evolution and Process, 24(7):783–801, 2012. 31 [28] M. J. Harrold and A. Orso. Retesting software during development and maintenance. In ICSM: Frontiers of Software Maintenance, pages 88–108, September 2008. [29] A. Malishevsky, G. Rothermel, and S. Elbaum. Modeling the cost-benefits tradeoffs for regression testing tech- niques. In ICSM, pages 204–213, October 2002. [30] E.H. Mamdani and S. Assilian. An experiment in linguistic synthesis with a fuzzy logic controller. International Journal of Man-Machine Studies, 7(1):1–13, 1975. [31] S. Mirarab and L. Tahvildari. A prioritization approach for software test cases on Baysian Networks. In FASE, pages 276–290, March 2007. [32] E.W.T. Ngai and F.K.T. Wat. Fuzzy decision support system for risk analysis in e-commerce development. Decision Support Systems, 40(2):235–255, 2005. [33] J. Offutt, J. Pan, and J. M. Voas. Procedures for reducing the size of coverage-based test sets. In Proc. Int’l. Conf. Testing Comp. Softw., pages 111–123, June 1995. [34] M. Riaz, J. King, J. Slankas, and L. Williams. Hidden in plain sight: Automatically identifying security require- ments from natural language artifacts. In 22nd IEEE International Conference on Requirements Engineering Conference (RE), pages 183–192, August 2014. [35] G. Rothermel and M. J. Harrold. Analyzing regression test selection techniques. IEEE TSE, 22(8):529–551, August 1996. [36] G. Rothermel, R. Untch, C. Chu, and M. J. Harrold. Prioritizing test cases for regression testing. IEEE TSE, 27(10):929–948, October 2001. [37] T. L. Saaty. The Analytic Hierarchy Process. McGraw-Hill, 1980. [38] N.F. Schneidewind and H.M. Hoffman. An experiment in software error data collection and analysis. IEEE TSE, 5(3):276–286, May 1979. [39] A. Schwartz and H. Do. A fuzzy expert system for cost-effective regression testing strategies. In 29th IEEE International Conference on Software Maintenance (ICSM), pages 1–10, September 2013. [40] Mark Sherriff, Mike Lake, and Laurie Williams. Prioritization of regression tests using singular value decompo- sition with empirical change records. In ISSRE, pages 81–90, November 2007. [41] H. Srikanth, L. Williams, and J. Osborne. System test case prioritization of new and regression test cases. In ESE, pages 64–73, August 2005. 32 [42] A. Srivastava and J. Thiagarajan. Effectively prioritizing tests in development environment. In Proceedings of the International Symposium on Software Testing and Analysis, pages 97–106, July 2002. [43] H. Stallbaum, A. Metzger, and K. Pohl. An Automated Technique for Risk-based Test Case Generation and Prioritization. In Proceedings of the 3rd International Workshop on Automation of Software Test, pages 67–70, May 2008. [44] A. Walcott, M. L. Soffa, G. M. Kapfhammer, and R. S. Roos. Time-aware test suite prioritization. In Proceedings of the International Conference on Software Testing and Analysis, pages 1–12, July 2006. [45] D.R. Wallace and D.R. Kuhn. Failure Modes in Medical Device Software: An Analysis of 15 Years of Recall Data. Reliability, Quality and Safety Engineering, 8(4):301–311, 2001. [46] Z. Xu, K Gao, and T.M. Khoshgoftaar. Application of fuzzy expert system in test case selection for system regression test. In IEEE International Conference on Information Reuse and Integration, pages 120–125, August 2005. [47] S. Yoo and M. Harman. Regression testing minimization, selection and prioritisation: A survey. JSTVR, pages 67–120, March 2010. [48] M. Yoon, E. Lee, M. Song, and B. Choi. A test case prioritization through correlation of requirement and risk. Journal of Software Engineering and Applications, 5(10):823–835, 2012. [49] L.A. Zadeh. Fuzzy sets. Information and Control, 8(3):338–353, 1965. [50] W.M. Zage and D.M. Zage. Evaluating design metrics on large-scale software. IEEE TSE, 10:75–81, 1993. [51] P. Zech. Riskbased security testing in cloud computing environments. In Fourth IEEE International Conference on Software Testing, Verification and Validation, pages 411–414, March 2011. 33 
work_ldtjjd6yffebdlzeevskcuf6xu	----	Evaluating new product development performance by fuzzy linguistic computing Evaluating new product development performance by fuzzy linguistic computing Wen-Pai Wang * Department of Industrial Engineering and Management, National Chin-Yi University of Technology, 35, Lane 215, Section 1, Chung-Shan Road, Taiping City, Taichung 411, Taiwan, ROC a r t i c l e i n f o Keywords: New product development Group decision making Heterogeneous information 2-Tuple fuzzy linguistic computing a b s t r a c t New product development (NPD) is indeed the cornerstone for companies to maintain and enhance the competitive edge. However, developing new products is a complex and risky decision-making process. It involves a search of the environment for opportunities, the generation of project options, and the evalu- ation by different experts of multiple attributes, both qualitative and quantitative. To perceive and to measure effectively the capability of NPD are real challenging tasks for business managers. This paper presents a 2-tuple fuzzy linguistic computing approach to deal with heterogeneous information and information loss problems during the processes of subjective evaluation integration. The proposed method which is based on the group decision-making scenario to assist business managers to measure the performance of NPD manipulates the heterogeneous integration processes and avoids the informa- tion loss effectively. Finally, its feasibility is demonstrated by the result of NPD performance evaluation for a high-technology company in Taiwan. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction Product design has been long recognized as an opportunity for differential advantage in the market place. A number of companies successfully focus on product design as a competitive tool (Creusen & Schoormans, 2005). Nowadays, more requirements for enter- prises have been put forward, such as more product variety, short- er time-to-market, lower product cost and higher quality. The globalization of competition in the manufacturing industry and the diversification of customers’ demands as well as rapid techno- logical developments continue to spur technology-based innova- tions at a frenetic pace. Product design innovation therefore has developed quickly and has gradually become one of mainstream production modes of manufacturing industries in the 21st century. Therefore, improving product development performance is becom- ing increasingly important and challenging. New product development (NPD) is undeniably vital in deter- mining the economic success of manufacturing companies. Firms need to create and sustain competitive advantages in order to sur- vive in today’s highly competitive business environment. One ma- jor determinant of sustaining competitive advantage is the ability of the firms to develop and launch successful new products. Differ- entiation through NPD is therefore one of the most effective strat- egies for achieving success. As competition in global markets has become intense, firms have begun to recognize the importance of NPD and innovation issues. Through innovation and the introduc- tion of new products, new markets and growth possibilities can be created. Increasing international competition accentuates the importance of the NPD process which is secure and accurate (Ozer, 2005; Sherman, Berkowitz, & Souder, 2005). Gemser and Leenders (2001) conclude that being innovative with respect to design and design strategy can enhance competitiveness regardless of indus- try evolution. Timely, correct and responsive NPD has become even more critical in the highly competitive global environment. The need to respond quickly to these dynamic global market forces re- quires the firm to establish a specialized evaluation mechanism and platform for NPD performance. However, the decision-making domain of NPD is highly com- plex and uncertain due to a demanding environment characterized by increased globalization and segmentation of markets, increased levels of product complexity, changing customer needs, and short- er product life cycles (Belecheanu, Pawar, Barson, Bredehorst, & Weber, 2003). New product introduction in today’s technology-dri- ven markets carries significant risk. New product failure rates can be as low as one of every three products or as high as the 90% of new grocery products which are withdrawn within a year of their introduction. New technology, improved communications, in- creased profit demands and shorter product life cycles have added to the inherent risk. Yet, without the introduction of new products, deterioration of the firm’s market position is inevitable. Without new products, firms will inevitably stagnate (Yelkur & Herbig, 1996). In order to evaluate the performance of NPD more appropri- ately, the firms should consider not only quantitative index but also qualitative dimensions or factors which are evaluated by multiple decision-makers or experts. Thus, the evaluation of NPD performance should be regarded as a group multiple criteria decision-making problem as well. 0957-4174/$ - see front matter � 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2009.02.034 * Tel.: +886 4 23924505x6018; fax: +886 4 23921742. E-mail address: wangwp@ncut.edu.tw Expert Systems with Applications 36 (2009) 9759–9766 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a mailto:wangwp@ncut.edu.tw http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Experts devote themselves to judge the NPD performance mea- surement by their experiential cognition and subjective perception in the decision-making process. However, there exists a consider- able extent of uncertainty, fuzziness and heterogeneity (Hwang & Yoon, 1981). This is not a seldom situation. In addition, it is prone to information loss happening during the integration processes, and gives rise to the evaluation result of the performance level which may not be consistent with the expectation of the evalua- tors. Consequently, developing an easy way to calculate the perfor- mance ratings while the processes of evaluation integration and to manipulate the operation of qualitative factors and expert judg- ment appropriately in the evaluation process of NPD could brook no delay. In this paper we propose a suitable model based on 2-tu- ple fuzzy linguistic information to evaluate the NPD performance. The proposed approach not only inherits the existing characters of fuzzy linguistic assessment but also overcomes the problems of information loss of other fuzzy linguistic approaches (Herrera- Viedma, Herrera, Martinez, Herrera, & Lopez, 2004). This paper is organized as follows. In Section 2 the measure- ment dimensions of NPD are described. In Section 3 we introduce the basic definitions and notations of the fuzzy number, linguistic variable and 2-tuple fuzzy linguistic representation and operation, respectively. In Section 4a NPD performance measurement meth- od based on 2-tuple fuzzy linguistic information is proposed . The proposed model is then illustrated with an example for a high-technology company in Taiwan. In Section 5 conclusion is given. 2. Literature review A contemporary NPD process usually consists of hundreds or thousands of activities, where the activities may be dependent or interdependent on one another. A rapidly changing competitive landscape and dynamic customer expectations require manufac- turers to seek flexibility in product development. Unlike the man- ufacturing processes, product development is a creative and discovering process that tends to create something new from trial-and-error and learning from the errors made (Wang & Lin, forthcoming). The purpose of NPD is to accumulate the knowledge and capability necessary to determine an appropriate new product. Superior product design, potential for breakthrough innovation, low project and product cost, shorter lead time, better communica- tion of cross-functional teamwork, and increased customer satis- faction and market share are among many other advantages for successful NPD. Suchlike concerns enable firms in making NPD decisions while ensuring full knowledge of the customer, the tech- nology, and with the team’s support. In view of this, a performance evaluation method or approach that is capable of systematically analyzing and accurately quantifying those subjective experiences and judgments of the NPD team is highly required. Ozer (2005) indicated that the quality of new product evalua- tion decisions is affected by four major sets of factors, namely the nature of the task, the type of individuals who are involved in the decisions, the way the individuals’ opinions are elicited and the way the opinions are aggregated. The main drivers of NPD include: quality and speed to market; widening customer choice and expectation; competitive priorities of responsiveness, delivery, flexibility, concern for the environment and international competitiveness. For example, Wang and Lin (forthcoming) pointed out that the introduction timing of new products is impor- tant for high-technology industries to gain premium pricing and higher sales volume. A NPD project in nature should possess four latent abilities: delivering value to the customer; being ready for change; valuing human knowledge and skills; and forming virtual partnerships (McCurry & McIvor, 2002). NPD is thus a key factor for survival for business firms. Most of the fast-growing companies achieve above 50% of their total sales from the new products developed within 5 years (Lee, Lee, Koo, & Yan, 1996). Not only is the technology changing rapidly, but the process of the commercialization of technological change–the industrial innovation process–is also changing. Nowadays due to the increasingly competitive climate, more and more managers are forcing themselves to update on the range of factors that deter- mine product innovation success. Takeuchi and Nonaka (1986) indicated that all profits of new products would account for 30– 40% of total sales. Griffin (1997) represented a substantial antici- pated increase in the profit impact of new products. Sales from establishments which were part of the business five years earlier represented 32.4% of total annual sales. Especially, high-technol- ogy industries attained a great percentage of 42.3% and this in- creased continuously. Even so, the average failure rate of new products also reached a great percentage of 41%. In sum, the firms are in urgent need of developing a specialized NPD performance evaluation mechanism and platform for their effective manage- ment and for enhancing business competitiveness further. It is however difficult and laborious to measure NPD perfor- mance using traditional crisp value directly as the process of NPD performance measurement possesses many intangible or qualitative factors and items. Linguistic variable representation is therefore favorable for experts to express and evaluate the ratings of NPD under such a situation. The fundamentals of 2-tuple fuzzy linguistic approach are to apply linguistic variables to stand for the difference of degree and to carry out processes of computing with words easier and without information loss during the integra- tion procedure (Herrera-Viedma et al., 2004). That is to say, deci- sion participants or experts can use linguistic variables to estimate measure items and obtain the final evaluation result with proper linguistic variable. It is an operative method to reduce the decision time and mistakes of information translation and avoid information loss through computing with words. 3. Fuzzy linguistic computing approach Many aspects of different activities in a real world cannot be as- sessed in a quantitative form, but rather in a qualitative one, i.e., with vague or imprecise knowledge. Whereas characteristics of the fuzziness and vagueness are inherent in various decision-mak- ing problems, a proper decision-making approach should be capa- ble of dealing with vagueness or ambiguity (Yager, 1995). Fuzzy set theory is a very feasible method to handle the imprecise and uncertain information in a real world. Especially, it is more suitable for subjective judgment and qualitative assessment in the evalua- tion processes of decision making than other classical evaluation methods applying crisp values (Lin & Chen, 2004; Wang & Chuu, 2004). Basic definitions and concepts of fuzzy sets are briefly re- viewed as follows; and further, notations given below will be used throughout the paper until otherwise stated. 3.1. Fuzzy number A positive triangular fuzzy number (PTFN) eA can be denoted aseA ¼ða; b; cÞ, where a 6 b 6 c and a > 0, which are illustrated in Fig. 1. The membership function, leAðxÞ, is defined as (Zimmer- mann, 1991) leAðxÞ¼ ðx � aÞ=ðb � aÞ; a 6 x 6 b ðx � cÞ=ðb � cÞ; b 6 x 6 c 0; otherwise 8>< >: ð1Þ where x takes its values on the real line. A larger leAðxÞ means a stronger degree of belongingness for x in X. Triangular fuzzy num- 9760 W.-P. Wang / Expert Systems with Applications 36 (2009) 9759–9766 https://isiarticles.com/article/2748 
work_lfs4kidgmbhqnnedxcv2k6o6m4	----	NASA Technical Reports Server (NTRS) 
work_lgr5suct7rhcdlkzwade3g4fhm	----	Microsoft Word - 47_PE_08_14_199-204_latkowski PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 8/2014 199 Tomasz LATKOWSKI1, Stanisław OSOWSKI1,2 Military University of Technology (1), Warsaw University of Technology (2) Feature selection methods in application to gene expression: autism data Abstract. The paper presents the application of several different feature selection methods for recognizing the most significant genes and gene sequences (treated as features) stored in dataset of gene expression microarray related to autism. The outcomes of each method have been examined by analyzing gene expression profiles of selected genes. In the next step fusion of the most relevant features selected by different methods, has been implemented. The optimal number of features has been defined as the set providing the best clustering purity. Streszczenie. Praca prezentuje badanie wybranych metod selekcji cech diagnostycznych w celu wyodrębnienia najbardziej znaczących sekwencji genowych z mikromacierzy ekspresji genów dotyczącej autyzmu. Dla wyselekcjonowanych cech przeanalizowano wartości poziomów ekspresji genów. W kolejnym etapie dokonano fuzji wyselekcjonowanych cech. Optymalny zbiór cech wyznaczono na podstawie czystości przestrzeni klasteryzacji. (Metody selekcji cech diagnostycznych w zastosowaniu do ekspresji genów: baza danych autyzmu). Keywords: feature selection, gene expression microarrays, clustering, autism. Słowa kluczowe: selekcja cech, mikromacierze ekspresji genów, klasteryzacja, autyzm. doi:10.12915/pe.2014.08.47 Introduction Gene expression microarray is a sophisticated technique used in molecular biology. DNA microarrays are mainly applied for analyzing the expression of genes in specific cells at given time and under certain conditions [4]. This technique generates a huge number of features (genes and gene sequences) which create a large information dataset. The main problem from the data mining point of view is a limited number of observations related to very large number of gene expressions. Number of observations is usually in the range of hundreds and number of genes tens of thousands. Figure 1 shows the organization scheme of the DNA microarray. Fig.1 . DNA microarray organization in the form of matrix. Each row in the above figure corresponds to one observation (patient) and each column to one gene or gene sequence (feature). Because of the large imbalance of the number of genes and patients the selection is an ill conditioned problem. Moreover, data stored in medical databases are typically noisy and some gene sequences have large variance [19]. It makes the gene selection in DNA microarrays very difficult task. Progress in bioengineering and data mining, which has been observed in recent years, has created the solid foundations for discovering the genes which are the best associated with the particular disease. Data analysis of microarrays is widely examined and introduced in literature [2,9-11,13,18-22] starting with pioneering Golub investigation in 1999 [7]. This area of science is being studied due to the fact that is strictly connected with early disease prediction (especially in tumors). It is thought that some diseases, i.e., autism, breast cancer, leukemia, etc., have their reflections in genetic changes [1]. From the practical point of view biologists need to identify only a small number of the most significant genes that can be used as biomarkers in the disease tracing. In the next step these selected markers can be observed in order to understand or find association with the disease. Present data mining methods allow to look at the gene selection problem from many points of view. It is known that the selection algorithms may provide various results for different datasets (diseases) [19]. It makes selection issue more complicated due to the fact that the model developed for one problem may not be suitable for another dataset. Many publications have examined gene selection problem using only one method of ranking. If we take into account that DNA microarrays create ill conditioned problem such approach may not provide the globally optimal result. Moreover, examining different methods on one dataset we can observe various outcomes [20]. In this paper we follow the direction pointed in [20] and present more complex approach based on using several methods simultaneously. The results of individual selection methods are fused, leading to the final set of genes. The application of several methods gives opportunity to look at the selection problem from different points of view. In this way we increase the probability of proper selection of the most important genes. Developed model is verified on the publicity available database related to autism. Expression levels of the selected genes have been examined and compared to the randomly chosen ones. The cluster purity idea has been applied for obtaining the optimal number of the most relevant genes. The results of selection have been illustrated in a graphical way using the PCA mapping. Dataset description This paper presents the problem of gene selection applied to the dataset related to the autism. The database is publicity available and was downloaded from GEO (NCBI) repository [23]. Autism is a severe neurodevelopmental disorder with characteristic social and communication deficits and ritualistic or repetitive behaviors [1]. Many etiologies have been suggested and numerous risk factors identified. Autism is associated with a high degree of heritability. Few specific genetic mutations have been identified accounting for a minority of cases [1], while the majority of cases are considered sporadic. 200 PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 8/2014 Number of observations in this dataset equals 146 and number of genes 54613. The database consists of two classes: the first one is related to children with autism (n=82) and the second to control (healthy) children (n=64). Blood draws for all subjects were done between the spring and summer of 2004. Total RNA was extracted for microarray experiments with Affymetrix Human U133 Plus 2.0 39 Expression Arrays. The important challenge is to find a small subset of genes with good class discriminative abilities. This problem is resolved by using several gene selection methods combined into one system. Feature selection methods Feature selection is an important operation in processing the data stored in gene microarrays. The most relevant genes (treated as the features) increase our understanding of the mechanism of disease formation and allow to predict the potential danger of being affected by such disease. The application of feature selection methods allows to identify a small number of important genes that can be used as biomarkers of the appropriate disease. In this paper different feature selection methods will be examined and integrated in the final system. Using the set of methods well increase the probability of finding the globally optimal set of genes which are the best associated with the particular disease. The paper will study the following methods: Fisher discriminant analysis, ReliefF algorithm, two sample t-test, Kolmogorov-Smirnov test, Kruskal-Wallis test, stepwise regression method, feature correlation with a class and multi-input linear Support Vector Machine network. These methods rely their operation principles on different foundations and thank to this allow to look on the selection problem from different points of view. The following sections present short description of each method used in investigation. Fisher discriminant analysis In Fisher approach the greatest wage is assigned to feature which is characterized by a large difference of the mean values in two studied classes and a small value of standard deviations within each class. The two class discrimination measure of the feature f is defined in the form [14] : (1) 21 21 12 )(     cc fS where: c1 and c2 represent the mean values for classes 1 and 2, respectively, while σ1 and σ2 are the appropriate standard deviations. A large value of S12(f) indicates good class discriminative ability of the feature. On the other hand small value is an indication of the insignificance of the feature in the recognition of these two classes. ReliefF algorithm The reliefF algorithm ranks the features according to the highest correlation with the observed class. It can be implemented for incomplete and noisy data. According to the reported results [15] it represents an important approach to gene selection. The main idea of the ReliefF algorithm is to estimate the quality of the features according to how well their values distinguish between observations that are near to each other. ReliefF selects randomly an instance Ri and then searches for k of its nearest neighbors from the same class, called nearest hits Hj, and also k nearest neighbors from each of the different classes, called nearest misses Mj(C). It updates the quality estimation W[A] for all attributes A depending on their values for Ri, hits Hj and misses Mj(C). If instances Ri and Hj have different values of the attribute A then the attribute A separates two instances with the same class which is not desirable. So the quality estimation W[A] is decreased. If instances Ri and Mj have different values of the attribute A then this attribute separates two instances of different class values which is desirable. So the quality estimation W[A] is increased. The algorithm averages the contribution of all hits and misses. The contribution for each class of misses is weighted with the prior probability of that class P(C) which is estimated from the training set. The process is repeated m times, where m is a user-defined parameter. The detailed description of the procedure can be found in [15]. Two-sample t-test The next used selection method is a two-sample t-test. One explicit assumption of t-test is that each of two compared populations should follow a normal distribution. The null hypothesis of t-test is that data in the class 1 and 2 are independent random samples of normal distributions with equal means and equal but unknown variances against the alternative hypothesis that the means are not equal. The test statistic is formulated in the form (2) mn cc t 2 2 2 1 21     where n and m represent the sample sizes of both classes [12]. Two sample t-test is implemented in MATLAB as ttest2 function [12]. The test result returns h, which is equal 1 or 0. The value of 1 indicates a rejection of the null hypothesis at the 5% significance level, while h=0 indicates a failure to reject the null hypothesis at the same significance level. The function returns also the p-value of the test. Low value of p indicates that the compared populations are significantly different. Figure 2 illustrates the histogram of the distribution of values of the randomly selected gene for the autism and references classes. Fig. 2. The histograms and their Gaussian approximations for randomly selected gene representing two classes (autism and controls) In the above case, the null hypothesis at the 5% significance level was rejected (h=1). It means that the distributions of these classes differ. This randomly selected gene can be accepted as a candidate for the relevant feature. Checking the condition of normality distribution of genes we found that in about 80% cases it was fulfilled. Kolmogorov-Smirnov test The other statistical feature selection method applied in the research was the Kolmogorov-Smirnov (KS) test. It compares the medians of the groups of data to determine if the samples come from the same population [12].The null PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 8/2014 201 hypothesis is that both classes are drawn from the same continuous distribution. The alternative hypothesis is that they are drawn from different distributions. The KS test statistic is based on the relation (3)  )()(max 21 xFxFKS  where F1(x) and F2(x) are the cumulative distribution of samples of feature f belonging to class 1 and 2. High value of this coefficient indicates that the feature has good class discriminative ability. On the other hand, a small value of this factor indicates that feature should be rejected at selection stage. Figure 3 illustrates the result of KS test for randomly selected gene. Fig. 3. The cumulative distribution function for two classes (autism and controls) for randomly selected gene in Kolmogorov-Smirnov test. From the discriminative point of view this gene is not significant (p=0.8136) and should not be taken into account in further selection procedure. Kruskal-Wallis test In this method medians of the samples are compared, but test uses ranks of the data rather than the numeric values. It finds ranks by ordering the data from the smallest to the largest across all groups and taking the numeric index of this ordering. The Kruskal-Wallis test does not make any assumptions about normality. It assumes that the observations in each group come from populations with the same shape of distribution. It returns the p value for the null hypothesis that all samples are drawn from the same population. The Kruskal-Wallis test is implemented in MATLAB as kruskalwallis function. The returned value of p indicates kruskalwallis test rejects the null hypothesis that all data from two classes come from the same distribution at 1% significance level. Stepwise regression method Stepwise regression is a systematic method for adding and removing features to the set of input attributes based on their statistical significance in a regression. The method begins with an initial model and then compares the explanatory power of incrementally larger and smaller models. At each step, the p value of an F-statistic [14] is computed to test models with and without selected feature. Based on the statistic result algorithm makes a decision whether feature should be included in a model or not. If a feature is not currently in the model, the null hypothesis is that the term would have a zero coefficient if added to the model. If there is sufficient evidence to reject the null hypothesis, the feature is added to the model. Conversely, if a feature is currently in the model, the null hypothesis is that the term has a zero coefficient. If there is insufficient evidence to reject the null hypothesis the term is removed from the model. The method proceeds as follows [12]: 1. Fit the initial model. 2. If any terms not in the model have p-values less than an entrance tolerance add the one with the smallest p value and repeat this step; otherwise go to step 3. 3. If any terms in the model have p-values greater than an exit tolerance remove the one with the largest p value and go to step 2; otherwise end. The algorithm is interrupted if none of steps leads to the increase of the model accuracy. Presented method may build different sets of features depending on initial model and order of adding and removing features from the set of attributes. Considering that outcomes may not be reproducible the stepwise regression method provides locally optimal result. Feature correlation with class In this method, the correlation of the feature values with a class is examined. The discriminative value S(f) of the feature f for recognizing one class from the other K classes is defined as follows [14]: (4)        K k kk K k kk PPf mmP fS 1 2 1 2 )1()( )( )(  where m is a mean value of feature for all data, mk is a mean value of the feature for the kth class data, σ2(f) is a variance of feature, Pk is a probability of kth class occurrence in dataset (the uniform distribution is assumed). In this paper number of classes is 2 (K=2). At the uniform distribution of both classes the above equation can be simplify to the following formula: (5) )(2 ))()(())()(( )( 2 2 2 2 1 12 f fmfmfmfm fS    The large value of S12(f) indicates good discriminative ability of feature for recognition of two classes. Multi-input linear SVM network SVM network is mainly used in regression and classification tasks [8]. It can be also configured for solving selection problem. In this approach SVM network with linear kernel is used. Network is learned applying all available features used simultaneously as an input. The sign (sgn) function is added for matching the input values to the appropriate class label [14]. The output y at presentation of the features organized in the form of vector f is defined by the following equation (6) )sgn()sgn()( buy T  fwf where w=[w1, w2,..., wn] T is the weight vector, f=[f1, f2,...,fn] T is a vector of features and b is a bias. Large absolute value of weight connecting feature f with the network denotes strong ability of this feature to distinguish two classes. In practice the recursive feature elimination (RFE) approach is used [20]. In RFE, the features are eliminated step by step according to an assumed criterion related to their support in the discrimination of the classes. The SVM is re-trained at each step using smaller and smaller population of features. In first step the linear SVM network is learned using all features. The weights are adapted and then sorted in descending order. In the next steps the features associated with the smallest absolute weights are eliminated. The process is repeated until we obtain an appropriate number of the most important features. Numerical results of experiments This section describes the numerical experiments using the autism data. We will present the results concerning the selection, clusterization and PCA transformation. In the first 202 PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 8/2014 stage eight feature selection methods are used to discover the sequence of the most significant genes. In the next stage, we consider only 100 top selected genes from each method. These subsets are fused into one reduced common set. To find the optimal number of genes we use the notion of cluster purity. The final outcome of the clusterization is illustrated and examined by using PCA transformation. Gene selection results Feature selection methods described in the previous sections have been applied to get the order of genes, sorted in a decreasing fashion. As a result of these experiments eight subsets of 100 the most relevant genes are selected. The following abbreviations were applied: FD - the Fisher discriminant analysis, RF - the ReliefF algorithm, TT - the two-sample t-test, KS - the Kolmogorov-Smirnov test, KW - the Kruskal-Wallis test, SWR - the stepwise regression method, COR - the feature correlation with a class, SVM - the multi-input linear SVM network. As was expected the methods have selected different sets of genes. Table 1 shows how many identical genes among the first 100 of the most important have been selected by different methods. Table 1. The percentage of the same genes among top 100 selected by different methods. FD RF TT KS KW SWR COR SVM FD 100 46 90 30 66 3 90 1 RF 46 100 46 33 46 4 46 1 TT 90 46 100 30 68 3 100 0 KS 30 33 30 100 46 3 30 1 KW 66 46 68 46 100 3 68 1 SWR 3 4 3 3 3 100 3 1 COR 90 46 100 30 68 3 100 0 SVM 1 1 0 1 1 1 0 100 The contents of the selected sets differ from method to method. Analyzing them we may find that few methods identified a large number of the same genes. For example the correlation feature with a class and t-test produced the same sets of genes. The overlapping results between the Fisher and correlation with a class methods cover 90% of genes. On the other hand some of them have resulted in very different sets, i.e., stepwise regression and t-test outcomes are overlapping only in 3%. The quality of the selection processes has been confirmed by analyzing the expression profiles for the identified genes. Fig. 4 illustrates the expression levels of all patients for the most important gene selected by the Fisher method. As we can see the mean value of the observations belonging to the autism class differs significantly from the reference class. Fig. 4. Expression levels for gene belonging to the most significant group. For comparison the appropriate expression levels representing gene selected randomly from the least significant subset are shown in Fig. 5. This time the difference between the means of both classes is unnoticeable. Fig. 5. Expression levels for gene representing the least significant group. In the case of the carefully selected genes (Fig. 4) we got the mean equal 60.94±9.50 (autism) and 53.74±7.00 (control group). The difference of the mean values is 7.20 (11.81% in relative terms). For the gene representing the least significant subset (Fig. 5) we got 9.13±1.13 (autism) and 9.16±1.49 (control group). The spread of mean values in this case is equal only 0.03 (0.33%). These differences confirm the advantage of the proposed selection procedure. The next important step is fusing the results of the individual methods into one common outcome. To find the most important genes valid for all analyzed methods we have to assign the global weight to each selected gene. The following formula has been proposed for assigning the global weight w of the gene f (7) )()( 1 fwfw k i k   The index k is the number of applied selection methods (k=8 in this research), wk is the position of genes in the appropriate method of selection. The best gene is the one of the smallest value of the global weight w. Analyzing the contents of all selected sets we have found 427 different genes. They have been sorted according to their global weights in an increasing order (the best gene is the one of the smallest weight). To the best 10 genes selected by the fusion algorithm belong: 16274 (206827_s_at), 46183 (236933_at), 38133 (228878_s_at), 335 (1552729_at), 50099 (240849_at), 18235 (208819_at), 45248 (235998_at), 2923 (1556314_a_at), 26879 (217593_at), 4077 (1558154_at). Clusterization of gene space Good way for assessing the quality of the selected genes is to apply the clustering of data in the multidimensional space. Good set of genes should provide the clusters of high purity with respect to the class membership. Different approaches to clustering are possible: K-means, fuzzy c-means or expectation maximization algorithm [14]. In this work we apply the simplest K-means. It is a method of vector quantization, that aims to partition n observations into K clusters (K<n). Each N-dimensional observation belongs to the cluster with the nearest mean (centroid) serving as a prototype of the cluster. This aim is achieved by minimizing the squared sum distances between centroids and the vectors within each cluster [14]. K-means can be developed in two approaches: off-line and on-line. We use the off-line Matlab version with batch updates. Each iteration consists of reassigning all points to their nearest cluster followed by recalculation of the cluster centroids. PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 8/2014 203 In our paper, we assume the number of clusters equal two (K=2), the same as the number of investigated classes. The K-means will be used by us to find the number of the most significant genes providing the highest class purity of the clustered space. The procedure consists of repeating the K-means algorithm at varying number of the genes. In each step we increase this number by one. The clusters are assessed using their purity index, defined as follows (8) i ij i n n p max In this definition nij is a number of observations of jth class inside of ith cluster and ni is a number of observations forming ith cluster (i,j =1,..,K). In the next step the total purity of the clustered space is calculated using the following equation: (9)    K i i i p n n p 1 where K is the number of clusters. In this way we can calculate the total purity of the clustered space at varying dimension (number of the most significant genes) of the representative vectors. Figure 6 presents the change of the total purity index versus the number of the most relevant genes after final fusion procedure. We can observe that the best purity is obtained for the top 24 features. Its value in our experiments was equal 0.84. Fig. 6. Total purity index of clustered space versus number of the most significant genes. The best result obtained at application of the fusion approach has been compared to the outcomes of 8 individual selection methods. Table 2 presents the highest values of the total purity index for all investigated selection methods and the number of genes at which these maxima happened. Table 2. The highest values of the total purity corresponding to the set of genes selected individually by different methods and after their fusion. FD RF TT KS KW SWR COR SVM Fusion purity 0.84 0.77 0.83 0.77 0.82 0.70 0.83 0.62 0.84 number of genes 52 75 34 53 24 17 34 9 24 We can notice that total purity of the clustered space at application of the investigated methods differs significantly. Moreover, the highest purity is obtained at different number of genes. For instance, in the stepwise regression method the best purity occurs at only nine the most significant genes, whereas in the ReliefF algorithm at seventy-five. The best clusterization of the space corresponds to the fusion approach at presence of only 24 best genes. These 24 genes have been taken into account in further experiments. To illustrate these results in a graphical form we have presented the expression levels of the selected genes in the form of image. Figure 7 shows the image of the expression profiles for the top twenty four genes using the colormap of hot. There is a visible border between the 82 observations of the autism group and the remaining 64 representing the reference one. For comparison the image of the expression profiles for 24 genes chosen randomly from the base is presented in Figure 8. There is a significant difference confirming good performance of the proposed selection procedure. Fig. 7. The color image of the expression profiles for 24 the most significant genes selected by the fusion procedure. Fig. 8. The color image of the expression profiles for 24 randomly chosen genes. Illustration of selection results using PCA The next considerations are concerned with the graphical representation of the multidimensional observation vectors using the Principal Component Analysis (PCA). PCA is a statistical method mapping the original vectors x from the large space to vectors y in the space of the reduced dimension. The transformation is done through the linear relation (10) Wxy  in which the transformation matrix W is formed from chosen number of eigenvectors corresponding to the largest eigenvalues of the correlation matrix of the original data x. Fig. 9. The distribution of the two-class samples mapped on two the most important principal components at representation of vectors x by twenty four most significant genes. 204 PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 90 NR 8/2014 We have mapped the multidimensional observations into 2-dimensional space formed by two most important principal components. Two cases have been investigated. In the first approach the original vectors contained only the selected 24 genes. Figure 9 depicts the case in which we use only the best representative genes in the vectors x. For comparison we have repeated PCA on the full size original vectors containing all genes. The graphical results of sample distribution are presented in Fig. 10. Fig. 10. The distribution of the two-class samples mapped on two most important principal components at application of all genes. We can observe a significant difference in distribution of the samples belonging to both classes. In the first case (Fig. 9) we can see two compact regions relatively pure with respect to the classes. In the second case (Fig. 10) the situation is different. The observations belonging to different classes interlace each other. It is practically impossible to separate them into two classes. To confirm this graphical impression the Euclidean distance between the observations and their centroids have been computed. Table 3 shows the values of their average distance to the appropriate centroids and standard deviation of the observations for both investigated cases. They correspond to the data mapped to the 2-dimensional space. Table 3. The average distance and standard deviation of the observations from their centroids. Number of genes Autism Control group Top 24 genes 0.05±0.03 0.04±0.03 All genes 1.70±0.90 1.64±1.34 The results show that the total dispersion in the first case (24 genes forming the vectors x) is much smaller than in the second (all genes). At the same time we see significant difference of distances between the centroids representing both classes of data in 2-dimensional space. At representation of data by 24 genes this distance related to the maximum range of data was equal 0.226. In the second case this distance was only 0.034. Conclusions The paper has examined several data mining methods for selection of the most important genes in the expression microarray of autism. The most relevant genes have been selected using two stage approach. In the first step we applied eight different feature selection methods working independently. The final set has been identified by fusing all obtained subsets. Next, the expression levels of the selected genes have been investigated. We applied different tools and methods, including the clusterization of the data and the measures of its quality, principal component analysis and statistical characterization of the clustered space. The notion of the cluster purity has allowed us to identify the optimal number of the genes. Good quality of the presented selection approach has been confirmed by projecting the selected (limited) number of genes on the 2-dimensional space formed by two most important principal components using PCA algorithm. The results presented in the paper form the first step of exploring the microarray data related to autism. They allow to identify the genes which are the best associated with the disease. In the next step we can use them in early recognition of this disease by applying the predictive classifier system. REFERENCES [1] Alter M., Kharkar R. Ramsey K., Craig D., Melmed R., Grebe T., Curtis-Bay R., Ober-reynolds S., Kirwan J., Jones J., Blake- Turner J., Hen R., Stephan D., Autism and increased patternal age related changes in global levels of gene expression regulation, Plos One, 2011, vol. 6, pp. 1-10. [2] Baldi P., Long A.D., A Bayesian framework for the analysis of microarray expression data: regularized t-test and statistical inference of gene changes, Bioinformatics, 2001, vol. 17, pp. 509-519. [3] Fan R.E., Chen P.H., Lin C.J., Working set selection using second order information for training SVM, Journal of Machine Learning Research, ., 2005, vol. 6, pp.1889-1918. [4] De Rinaldis E., DNA microarrays: current applications, Horizon Scientific Press, Norfolk, 2007. [5] Duda R.O., Hart P.E., Stork P., Pattern Classification and Scene Analysis, Wiley, New York. 2003. [6] Eisen M., Spellman P., Brown P., Cluster analysis and display of genome wide expression patterns, Proc. Natl. Acad. USA, 1998, vol. 95, pp. 14863-14868. [7] Golub T. et al., Molecular classification of cancer: class discovery and class prediction by gene expression monitoring, Science, 1999, vol. 286, pp. 531-537. [8] Guyon I., Weston A.J., Barnhill S., Vapnik V., Gene selection for cancer classification using SVM, Machine Learning, 2002, , vol. 46, pp. 389-422. [9] Guyon I., Elisseeff A., An introduction to variable and feature selection, Journal of Machine Learning Research, 2003, vol. 3, pp. 1158 – 1182. [10] Hewett R., Kijsanayothin P., Tumor classification ranking from microarray data, BMC Genomics, 2008, vol. 9(2), pp. 1-11. [11] Huang X., Pan W., Linear regression and two-class classification with gene expression data, Bioinformatics, 2003, vol. 19, pp. 2072-2078. [12] Matlab user manual – Statistics toolbox, MathWorks, Natick, USA, 2012. [13] Mitsubayashi H., Aso S., Nagashima T., Okada Y., Accurate and robust gene selection for disease classification using a simple statistics, Biomedical Informatics, 2008, v. 391, 68-71. [14] Osowski S, Methods and tools in data mining (in Polish), BTC, Warsaw, 2013. [15] Robnik-Sikonja R, Kononenko I., Theoretical and empirical analysis of ReliefF and RReliefF, Machine Learning, 2003, vol. 53, pp. 23-69. [16] Sprent P., Smeeton N.C., Applied Nonparametric Statistical Methods, Boca Raton: Chapman & Hall/CRC, 2007. [17] Tan P. N., Steinbach M., Kumar V., Introduction to data mining, Pearson Education Inc., Boston, 2006. [18] Wang X., Gotoh O., Cancer classification using single genes, Genom Informatics, 2009, vol. 23 (1): pp. 179-188.. [19] Wang X., Gotoh O., A Robust Gene Selection Method for Microarray-based Cancer Classification, Cancer Informatics, 2010, vol. 9, pp. 15-30. [20] Wiliński A., Osowski S., Ensemble of data mining methods for gene ranking, Bulletin of the Polish Academy of Sciences, 2012. vol. 60, pp. 461-471. [21] Woolf P. J., Wang Y., A fuzzy logic approach to analyzing gene expression data, Physiological Genomics, 2000, v. 3, pp. 9-15. [22] Yang F., Robust feature selection for microarray data based on multicriterion fusion, IEEE Trans. Computational Biology and Bioinformatics, 2011, vol. 8(4), pp. 1080-1092. [23] http://www.ncbi.nlm.nih.gov/sites/GDSbrowser?acc=GDS4431 Authors: mgr inż. Tomasz Latkowski, Military University of Technology, Email: tlatkowski@wat.edu.pl; prof. dr hab. inż. Stanisław Osowski, Warsaw University of Technology, Military University of Technology, Email: sto@iem.pw.edu.pl. 
work_lguabr33l5h7jekr56bsun2o64	----	Discrimination discovery in scientific project evaluation: A case studyI Andrea Romei, Salvatore Ruggieri and Franco Turini Dipartimento di Informatica, Università di Pisa Largo B. Pontecorvo 3, 56127 Pisa, Italy Abstract Discovering contexts of unfair decisions in a dataset of historical decision records is a non-trivial problem. It requires the design of ad-hoc methods and techniques of analysis, which have to comply with existing laws and with legal argumenta- tions. While some data mining techniques have been adapted to the purpose, the state-of-the-art of research still needs both methodological refinements, the consolidation of a Knowledge Discovery in Databases (KDD) process, and, most of all, experimentation with real data. This paper contributes by presenting a case study on gender discrimination in a dataset of scientific research proposals, and by distilling from the case study a general discrimination discovery process. Gender bias in scientific research is a challenging problem, that has been tackled in the social sciences literature by means of statistical regression. However, this approach is limited to test an hypothesis of discrimination over the whole dataset under analysis. Our methodology couples data mining, for unveiling previously unknown contexts of possible discrimination, with statistical regression, for testing the significance of such contexts, thus obtaining the best of the two worlds. Keywords: Discrimination discovery, Gender bias, Case study, Situation testing, Data mining, KDD process 1. Introduction Discrimination refers to an unjustified distinction of in- dividuals based on their membership, or perceived mem- bership, in a certain group or category disregarding indi- vidual merits. Unfair behaviors have been observed in a number of settings, including credit, housing, insurance, personnel selection and worker wages. Civil rights laws prohibit discrimination against protected groups defined on the grounds of gender, race, age, nationality, marital status, personal opinion, and so on. One crucial prob- lem from legal, economic and social point of view is dis- crimination discovery, that is defining methods capable of providing evidence of discriminatory behavior in activities such as the ones listed above. In the socio-economic field the problem has been addressed by analysing data with a statistical approach. The basic idea is to see, by means of regression analysis, whether sensitive features, like gen- der and race, are correlated with a less favorable treat- ment of individuals. In our opinion such an approach can highlight only macroscopic situations, while missing to dig out situations of deep discrimination in (small) subsets of the population, i.e., niches of individuals with a particular combination of characteristics. As an example, consider the case of loan applications to a bank. The discrimina- tory behavior of a single branch office against applicants from a local minority can readily be hidden in the much IA preliminary version of the results of this paper appeared in Romei et al. (2012). larger set of decisions of the whole bank. In a few words, the statistical approach tends to find a general model char- acterizing the whole population, whereas discrimination often arises in specific contexts. We maintain that a data mining approach, that is the search for particular patterns in the data, must be coupled with statistical validation of the patterns found as a thorough strategy for discovering (unexpected or unknown) contexts of discrimination. Data mining approaches to discrimination discovery have recently gained momentum, but, in our opinion, they still need major advancements: first, experimentation with real data; second methodological refinements; and third, the consolidation of a KDD process of discrimination dis- covery. Solving these issues is essential for the acceptance of discrimination discovery methods based on data mining in practice. In this paper we contribute to the advance- ment of the state-of-the-art in all those aspects. First, we describe a large experiment on a real case study concerning the challenging problem of discovering gender discrimination in the selection of scientific projects for funding. Data refers to an Italian call for research proposals with 3790 applications nationwide. Second, we couple a recently developed discrimination discovery method (Luong et al., 2011), based on data min- ing, with statistical validation of its results, thus reconcil- ing the statistical and the data mining methodologies. The data mining method is unsupervised in the sense that there are no examples of discriminatory or non-discriminatory situations to learn from. Rather, the method discovers sets of situations in which the comparison of the features Expert Systems with Applications May 1, 2013 and the decision may suggest a possible discrimination ac- cording to the legal methodology of situation testing. Sta- tistical regression analysis is then used to prove or disprove them as an hypothesis. Third, the description of the steps followed in the case study provides us with the basis for distilling a general KDD process for discrimination discovery. The process abstracted is rather complex. It contains both automated and semi-automated steps, the possibility of iterating sub- processes, and the need of tuning the parameters of the analyses. The paper is organized as follows. Section 2 offers a survey of the background material, including a multi- disciplinary introduction to the gender bias in scientific research, a survey of data mining approaches for discrim- ination discovery, and details on the approach based on situation testing. Sections 3 presents the case study of an Italian national research funding call, the available data, and the data preparation steps. Sections 5 and 6 report the core of the experiment, consisting of first extracting clas- sification models describing contexts of possible discrim- ination, and then of validating such contexts by means of logistic regression. Four interesting contexts of possi- ble discrimination are discussed. Section 7 generalizes the phases of the case study to a generic discrimination dis- covery process. Finally, the conclusions summarize the contributions of the paper and discuss some future work. 2. Background The problem of gender differences in research is of concern to all major funding institutions. The European Union (EU) regularly publishes a report on the status of gender funding in the member states (European Commis- sion, 2009), and it promotes gender equality in scientific research.1 EU legislation includes an explicit resolution on women and science (Council of the E.U., 1999), which no- tably preceded the resolutions on racial and employment equality. The National Science Foundation (NSF) in the United States (US) supports the development of systemic approaches to increase the representation and advance- ment of women in academic science, technology, engineer- ing, and mathematics through the ADVANCE program.2 Broader overviews of studies and findings on gender in (sci- entific and technological) research have been conducted by the European Commission (2012) and by UNESCO (2007). In the next subsection, we review the existing literature on gender bias in scientific research, which basically relies on statistical regression as the basic tool for data analy- sis. Then, we briefly review recent approaches that use data mining for discrimination discovery and prevention. Finally, a deeper introduction is reported on the data anal- ysis technique of situation testing, and to its implementa- tion as a data mining algorithm. 1http://www.yellowwindow.be/genderinresearch 2http://www.portal.advance.vt.edu 2.1. Gender bias in scientific research Forms of gender discrimination may explain women’s under-representation in academia, both past and present. The surveys by Bentley and Adamson (2003) and Ceci and Williams (2011) cover multi-disciplinary literature on gender differences in manuscript reviewing, grant fund- ing, university admission, and hiring and promoting in re- search. We focus here on grant and fellowship funding. The influential paper by Wenner̊as and Wold (1997) reports a study on post-doctoral fellowship applications to the Swedish Medical Research Council (MRC) in 1995. A total of 62 applications were submitted by men and 52 by women: 16 men were funded (25.8%) versus 4 women (7.7%). Applicant’s sex and scientific competence are con- sidered as independent variables in a linear regression model estimating the score assigned by the reviewers. Scientific competence as a control factor is measured in terms of the number of published journal articles, their citation count, and total impact of those journals.3. The regression shows that “a female applicant had to be 2.5 times more produc- tive than the average male applicant to receive the same score”. However, subsequent studies by several funding so- cieties in Europe and North America fail to show evidence of sex bias in approval rates (Ceci and Williams, 2011). In fact, Sandström and Hällsten (2008) analysed data on applications to the MRC in 2004 and found a reversed gen- der bias, namely a small but significant effect in favor of funding women’s grants compared to men with the same scientific competence score. Let us recall here a few large scale studies. RAND (2005) investigates grant applications in the period 2001- 2003 to the NSF, the National Institutes of Health (NIH), and the Department of Agriculture. No evidence of gender bias was generally found after controlling for age, academic degree, institution, grant type, institute, and application year. There were two exceptions, partly explainable by the lack of further control variables. First, women received only 63% of the amount of funding awarded to men by the NIH. Secondly, women who applied in 2001 were less likely than men to submit applications in the next two years. Similar findings as in the first exception are also reported by Larivière et al. (2011) with reference to 9074 professors at universities in Quebec (Canada). The lower amount of funding received by women is not necessarily ev- idence of discriminatory decisions. Wilson (2004) explains the lower amount of funding granted to women by their marginalization within the scientific community, by their segregation to lower rank positions, and by their smaller social networks – all of these factors affecting their chances of funding possibilities. 3The problem of measuring and analysing science is the subject of scientometrics Indicators of scientific productivity of researchers have been debated for long time. See Bornmann and Daniel (2009) for a discussion on H-index and its variants, and for a comparison with the Impact Factor index. 2 Ley and Hamilton (2008) highlight that, whilst there is now generally gender equality between students and in- structors, there is still a striking drop in the roles of as- sistant professors and professors – i.e., a glass ceiling in science. The authors analysed more than 100,000 appli- cations between 1996 and 2007 to NIH grant programs to determine whether gender differences occur at some stage of a researcher’s career, which may explain the observed attrition. While they found a decrease in female applicants for grants throughout a researcher’s career, there is sub- stantial equity of the rates of funded applications between males and females at all stages of the process. Similar results are observed also by Brouns (2000) on grants awarded to 809 individual applicants by the Dutch Organization for Scientific Research. In this case, how- ever, the stratification by discipline exhibits a higher vari- ability of the success rate for women, from 26% up to 84%, compared to men, from 46% to 76%. Women ap- pear very successful in “hard” sciences (Physics, Mathe- matics, and Astronomy) and surprisingly unsuccessful in the “soft” natural sciences (Biology, Oceanography, and the Earth Sciences). Bornmann and Daniel (2005) anal- yse 1954 applications for doctoral and 743 applications for post-doctoral fellowships to a German foundation for the promotion of research in Biomedicine. The odds ratio of the approved doctoral fellowships for females (7%) against males (16%) is found statistically significant after checking for applicant’s age, grade, mobility, number of recommen- dation letters, ratings of reviewers. Marsh et al. (2008) summarize the major findings of an eight year research program on the analysis of peer reviews in grant applications to the Australian Research Council. Their dataset includes 2331 proposals rated by 6233 exter- nal assessors, out of a total of 10023 reviews. They con- sider issues such as: reliability of reviews, in the sense of an agreement of reviewers across individual proposals and across disciplines; trustworthiness of reviewers nominated by applicants; bias of national reviewers, who give more favorable evaluations than international ones; the positive influence of academic rank, in the sense that professors are more likely to be funded due to their experience and successful research track records; and the positive influ- ence of the prestige of the university affiliation and of the applicant’s age. They also consider the influence of an ap- plicant’s gender, finding that 15.2% of funded applications were led by females, which was exactly the same percent- age as female applicants. Once again, although women are under-represented in the applicants pool, they are equally represented in the funded pool. Their experiments also re- ject the “matching hypothesis” that reviewers give higher ratings to applicants of their same sex. Regarding the analytical methodology, research on peer review studies has carried out statistical analyses mainly by means of variants of correlation (Brouns, 2000), Z- tests of proportions (Ley and Hamilton, 2008), regression and more rarely by analysis of variance and discriminant function analysis. Multi-stages peer review processes have been also analysed with latent Markov models (Bornmann et al., 2008). The variants of regression adopted include multiple regression (Wenner̊as and Wold, 1997; Sandström and Hällsten, 2008), multi-level regression4 (Jayasinghe et al., 2003; Mutz et al., 2012), logistic regression (Born- mann and Daniel, 2005). The coefficient of the indepen- dent variable coding the applicant’s gender is taken as a measure of how gender affects the dependent variable, which is typically the score received by the application or its probability (or its logit) of being funded. Other in- dependent variables control factors such as scientific per- formance, scientific field, age, position, and institution. In this sense, “discrimination is the remaining racial [in our context, gender] difference after statistically account- ing for all other race-related [gender-related] influences on the outcome” (Quillian, 2006). However, it is difficult to know that all important characteristics of individuals have been taken into account: a recurring problem known as the omitted-variable bias. The inclusion of an omitted control variable may then explain (part of) the remaining gender differences. 2.2. Data mining for discrimination data analysis Discrimination discovery from data consists in the ac- tual discovery of discriminatory situations and practices hidden in a large amount of historical decision records. The aim is to unveil contexts of possible discrimination on the basis of legally-grounded measures of the degree of discrimination suffered by protected-by-law groups in such contexts. The legal principle of under-representation has inspired existing approaches for discrimination discovery based on pattern mining. A common tool for statistical analysis is provided by a 2 × 2, or 4-fold, contingency ta- ble, as shown in Figure 1. Different outcomes between groups are measured in terms of the proportion of people in each group (p1 for the protected group, and p2 for the unprotected one) with a specific outcome (benefit denial). Differences and rates of those proportions are commonly adopted as the formal counterpart of the legal principle of group under-representation. They are known in statistics as risk difference (RD = p1 − p2), also known as absolute risk reduction; risk ratio or relative risk (RR = p1/p2); relative chance (RC = (1 − p1)/(1 − p2)), also known as selection rate; odds ratio (OR = p1(1 − p2)/(p2(1 − p1))). Starting from a dataset of historical decision records, Pe- dreschi et al. (2008); Ruggieri et al. (2010a) propose to extract classification rules such as for instance: race=black, purpose=new_car => credit=no 4In addition to a measurement level random variable, multi-level regression (Goldstein, 2011) includes a subject level random variable modelling variations in a cluster of data. For instance, (Jayasinghe et al., 2003) adopt multi-level regression to take into account corre- lation in the cluster of ratings of a reviewer, and in the cluster of ratings of a same field of study. 3 benefit group denied granted protected a b n1 unprotected c d n2 m1 m2 n p1 = a/n1 p2 = c/n2 RD = p1 −p2 RR = p1 p2 RC = 1−p1 1−p2 OR = RR RC = a/b c/d Figure 1: 4-fold contingency table and discrimination measures. called potentially discriminatory rules, to unveil contexts (here, people asking for a loan to buy a new car) where the protected group (here, black people) suffered from under or over-representation with respect to the decision (here, over-representation w.r.t. credit denial). The approach has been implemented by Ruggieri et al. (2010b) on top of an Oracle database by relying on tools for frequent itemset mining. The main limitation of such an approach is that measuring group representation by aggregated values over undifferentiated groups results in no control of the char- acteristics of the protected group, versus or as opposed to others in this context. The high value of a discrimination measure from Figure 1 can be justified by the fact that proportions p1 and p2 mix decisions for people that may be very different as per characteristics that are lawful to obtain the required benefit (e.g., skills required for a job position). This results in an overly large number of rules that need to be further screened to filter out explainable discrimination. Luong et al. (2011) overcome this lim- itation by exploiting the legal methodology of situation testing, which will be presented in Section 2.3. The approach described so far assumes that the dataset under analysis contains an attribute that denotes the pro- tected group under analysis. The case when data do not contain such an attribute (or it is not even collectable at a micro-data level, e.g., as in the case of sexual orienta- tion) is known as indirect discrimination analysis (Ruggieri et al., 2010a), where ‘indirect’ refers to the exploitation of a known correlation with some other attribute, which can be used as a proxy for group membership. A well-known ex- ample is redlining discrimination analysis, occurring when the ZIP code of residence is correlated with the race of individuals in highly segregated regions. In this paper, we restrict to consider direct discrimination analysis. Finally, we mention the related research area of dis- crimination prevention in data mining and machine learn- ing (Calders and Verwer, 2010; Kamiran and Calders, 2012; Hajian and Domingo-Ferrer, 2012), where the problem is to design classification algorithms that trade off accuracy for non-discrimination in making predictions. Discrimina- Discovery(t) { for r ∈P { if( benefit(r) = denied and diff (r) ≥ t ) disc(r) ← true else disc(r) ← false } build a classifier on P where the class is disc } Figure 2: Left: example of risk difference diff (r) for k = 4. Women are the protected group, knnsetwomen(r) (resp., knnsetmen(r)) is the set of female (resp., male) k-nearest neighbors of r. Red labels benefit denied, green labels benefit granted. Right: pseudo code of k-NN as situation testing. Individuals r from the protected group P are first labeled as discriminated or not, and then a classifier is induced for describing those discriminated. tory predictions may be the result of a bias of the classifier induction algorithm, or of learning from training data tra- ditional prejudices that are endemic in reality. Summaries of contributions in discrimination discovery and prevention are collected in a recent book (Custers et al., 2013). In par- ticular, Romei and Ruggieri (2013) present, in a chapter of that book, a multi-disciplinary annotated bibliography of statistical tools, legal argumentations, economic mod- els, and experimental approaches for discrimination data analysis. 2.3. Situation testing and k-NN In a legal setting, situation testing is a quasi-experi- mental approach to investigate for the presence of discrim- ination by checking the factors that may influence decision outcomes. Pairs of research assistants, called testers, un- dergo the same kind of selection. For example, they apply for the same job, they present themselves at the same night club, and so on. Within each pair, applicant characteris- tics likely to be related to the situation (characteristics related to a worker’s productivity on the job in the first case, look, age and the like in the second case) are made equal by selecting, training, and credentialing testers to appear equally qualified for the activity. Simultaneously, membership to a protected group is experimentally ma- nipulated by pairing testers who differ in membership – for example, a black and a white, a male and a female, and so on. Observing significant difference in the selec- tion outcome between testers is a prima facie evidence of discrimination, i.e., a proof that, unless rebutted, would be legally sufficient to prove the claim of discrimination. For applications of situation testing, we refer to Bendick (2007), covering employment discrimination in the US; to Rorive (2009), covering the EU member States context; and to Pager (2007), including an appendix on the design of situation testing experiments. In Luong et al. (2011), the idea of situation testing is exploited for discrimination discovery just inverting the 4 Figure 3: The two-phases review process of the FIRB “Future in Research” call. point of view. Given past records of decisions taken in some context, for each member of the protected group with vector of attributes r suffering from a negative de- cision outcome (someone who may claim to be a victim of discrimination), we look for 2k testers with similar char- acteristics. Such characteristics are legally admissible in affecting the decision, apart the one of being or not in the protected group. Similarity is modeled via a distance func- tion between tuples. If we can observe significantly dif- ferent decision outcomes between the k-nearest neighbors of r belonging to the protected group and the k-nearest neighbors belonging to the unprotected group, we can as- cribe the negative decision to a bias against the protected group, hence labeling the individual r as discriminated. This approach resembles the k-nearest neighbor (k-NN) classification model, where the class of an individual is predicted as the most frequent class among its k-nearest neighbors. Difference in decision outcomes between the two groups of neighbors is measured by any of the func- tions from Figure 1, calculated over the proportions for the two sets of testers. Throughout the paper, we con- sider risk difference diff (r) = p1 − p2 with the intuitive reading that it represents the difference in the frequency p1 of negative decisions in the neighbors of the protected group with respect to frequency p2 in the neighbors of the unprotected group (see Figure 2 for an example). A value diff (r) > t, for a maximum threshold t, implies that the negative decision for r is not explainable on the basis of the (legally-grounded) attributes used for distance mea- surement, but rather it is biased by group membership. diff (r) is then a measure of the strength of such a bias. When diff (r) > t, individual r is labeled as discriminated by setting a new attribute disc to true. Starting from this labeling procedure, the actual learning of the conditions of discrimination is then modeled as a standard classification problem, where the class is the attribute disc. The overall procedure is reported in Figure 2. 3. Case study: data understanding & preparation In this section, we start the analysis of the case study of an Italian national call for research projects. We introduce the call and its evaluation process, the available data, and the features selected to form the dataset in input to the discrimination discovery analysis. Figure 4: Input data on research units. 3.1. The FIRB “Future in Research” call In 2008, the Italian Ministry of University and Re- search published a call for scientific research projects un- der the Basic Research Investment Fund (FIRB) reserved to young scientists – the FIRB “Future in Research” call. The scientific scope of the call is very broad, ranging from social sciences and humanities, to physical sciences and engineering, to life sciences. Research proposals are sub- mitted by a consortium of one or more research units, with a principal investigator (from now on, PI) and zero or more associate investigators heading each unit and affiliated to an Italian university or to a public research organization. Research proposals are distinguished in two programs, de- pending on whether the PI holds a non-tenured position and she/he is at most 33 years old at the time of the call (program P1), or she/he holds a tenured position and she/he is at most 39 years old (program P2). Each pro- gram has its own total budget, but the submission forms and the evaluation procedures are the same for both. The submitted proposals consist of a description of work and a budget for each research unit, called the B forms, and of a description of work and a budget for the whole proposal, called the A form. The global budget is basically the sum of the budgets of the research units par- ticipating in the project. The A form also contains the curriculum vitae of the PI, a list of her/his main publi- cations, and an abstract of the proposal. The hiring of at least one young researcher (defined as “a post-doc or a post-degree of at most 32”) per project proposal is required by the call. Invitation of good reputation researchers from abroad to spend some period working on the project is instead an option. Project proposals underwent a two-steps evaluation pro- cess, as shown in Figure 3. The first step consisted in a blind peer-review by national and international reviewers resulting in four scores about: (S1) scientific relevance of the proposal (score 0–8); (S2) impact of the proposal (score 0–7); 5 Figure 5: Input data on research proposals and evaluation results. (S3) scientific and technical value of the proposal (score 0–15); (S4) quality of the partnership (score 0–10). Only project proposals that received the best score for all of the four evaluation criteria (i.e., a total score of 40) were admitted to the second step, which consisted in an audition of the PI in front of a panel of national experts. The panel ranked the proposals into three classes: “to be funded”, “to be funded if additional budget were available”5 and “not to be funded”. In the first two cases, the panel also decided a budget cut with respect to the budget requested. 3.2. Data sources Anonymized data on project proposals and evaluation results were made available to us as an Oracle relational database. Proposals are identified by unique IDs CODE A. Similarly, research units have unique IDs CODE B. Data on research units (see Figure 4) are retrieved from B forms of project proposals. Table FORM B contains, for each research unit, attributes on the associate investigator leading the unit (gender, age, title, institution) and on the planned effort of the research unit in person/months. For each research unit, three other groups of data are available: • participants to the research unit, whose gender and age attributes are stored in the PARTICIPANT B table; • detailed costs of the research unit, stored in the DE- TAILED COST B table, including costs of: tenured per- sonnel, personnel to be hired, equipment, overhead, travel and subsistence, consulting and other costs; • aggregated costs, stored in the COST B table, includ- ing: the costs of hiring young researchers and good reputation researchers, the total budget of the re- search unit, and the eligible costs to be funded by the call. 5At that time, an increase of the budget of the call was under consideration. Data on research proposals (see Figure 5) are retrieved from A forms. Table FORM A contains data on the PI (gen- der, age, title, institution) and the research program of the proposal (P1 or P2). A few auxiliary tables follow: • PUBLICATION stores the list of publications of the PI. Authors’ names have been removed, but the number of authors is recorded; • ERC CLASSIFICATION stores the scientific area of the research project according to the European Research Council (ERC) classification; more than one area could have been chosen for a research proposal, e.g., in case of multi-disciplinary topics, with the first one representing the main area6; • KEYWORDS and ABSTRACT store respectively an or- dered list of keywords and the textual abstract of the proposal, in English; • COST A stores the duration of the project in months, and aggregated budget data: total effort in person/- months, total cost of the project, total eligible costs to be funded, number of young researchers and their total cost, number of good reputation researchers and their total cost. Data on the evaluation results is shown in the right- most tables of Figure 5. The scores obtained by a pro- posal over the four evaluation criteria of the first step of the evaluation process are stored in the SCORE table. Each criterium is coded with an ID. The ranks assigned by the commission of national experts to proposals in the second step of the evaluation process is stored in the AUDITIONS table. For proposals ranked as “to be funded” or “to be funded if additional budget were available”, the total cost and grant assigned to the project after budget cut is stored in the GRANTS table. 6The main area is used in the first step of the evaluation process to select the peer-reviewers of the proposal from a pool of area experts. 6 Name Description Type Features on principal and associate investigators gender Gender of the PI Nominal region Region of the institution Nominal city City of the institution Nominal inst type Type of the institution Nominal title Title of the PI Nominal age Age of the PI Numeric pub num No. of publications of the PI Numeric avg aut Average number of authors in pubs Numeric f partner num No. of females among principal or associate invest. Numeric Project costs (absolute values in e) tot cost Total cost of the project Numeric fund req Requested grant amount Numeric fund req perc fund req over tot cost Numeric yr num No. of young researchers Numeric yr cost Cost of young researchers Numeric yr perc yr cost over tot cost Numeric grr num No. of good reputation researchers Numeric grr cost Cost of good reputation researchers Numeric grr perc grr cost over tot cost Numeric Research areas program Program P1 or P2 of the proposal Nominal d1 lv1, d2 lv1, d3 lv1 1st, 2nd and 3rd area at the 1st ERC level Nominal d1 lv2, d2 lv2, d3 lv2 1st, 2nd and 3rd area at the 2nd ERC level Nominal d1 lv3, d2 lv3, d3 lv3 1st, 2nd and 3rd area at the 3rd ERC level Nominal Evaluation results S1 Score S1 assigned by the peer-reviewer Numeric S2 Score S2 assigned by the peer-reviewer Numeric S3 Score S3 assigned by the peer-reviewer Numeric S4 Score S4 assigned by the peer-reviewer Numeric peer-review Passed or rejected at the peer-review Nominal audition Passed or rejected at the audition (i.e., proposal funded) Nominal grant Amount granted after budget cut Numeric Table 1: Attributes of the dataset of the case study. 3.3. Data preparation The data preparation phase produced a dataset for the discrimination analysis in the form of a single relational table, including both source and derived features for each project proposal. Table 1 summarizes four groups of fea- tures. Features on the principal and associate investigators. These include gender, age, and title of the PI; number of publications and average number of authors in publica- tions of the PI; region (North, Center, South of Italy), city and type of her/his institution (University, Consortium or Other); and number of female principal or associate inves- tigators in the project proposal. Project costs. Several costs are considered: total cost of the project, requested grant (both absolute and in pro- portion to the total cost), number and cost of young re- searchers, number and cost of good reputation researchers. Research areas. In addition to the research program a proposal is submitted to (P1 or P2), up to three research areas are included, the first of which is the main area, according to the ERC classification. Such a classification consists of a three-level hierarchy. The top level includes Social sciences and Humanities (SH), Physical sciences and Engineering (PE), and Life Sciences (LS). The second and third levels (coded, e.g., as PE n and PE n m) include 25 and 3792 sub-categories respectively. Evaluation results. The following attributes are in- cluded: the scores (S1)-(S4) received at the peer-review, whether the project passed the first evaluation phase (i.e., the peer-review), whether the project passed the second evaluation phase7 (i.e., the audition), the actual amount granted after budget cut. 4. Case study: risk difference analysis Since research proposals of programs P1 and P2 are evaluated in isolation (due to distinct budgets), from now on, we act as if there were two datasets, one per program. Program P1 received 1804 applications, 923 of which are from female PIs; program P2 received 1986 ones, 792 of which from female PIs. 4.1. Exploratory data analysis of gender differences Table 2 summarizes the proportion of genders in the two phases of the evaluation process: peer-review and audition. It is readily checked that, for both programs, the proportion of female PIs progressively decreases when moving from applicant proposals to proposals passing the peer-review up to those passing the audition decision. Let us quantify such a decrease by means of discrimination measures. Figure 6 shows the 4-fold contingency tables of 7Since no additional budget was available for the call, proposals ranked as “to be funded if additional budget were available” are considered as not passing the audition. 7 PIs Peer-review passed Audition passed Male Female Male Female Male Female P1 881 923 43 31 25 17 (48.8%) (51.2%) (58.1%) (41.9%) (59.5%) (40.5%) P2 1194 792 100 31 51 12 (60.1%) (39.9%) (76.3%) (23.7%) (81%) (19%) Table 2: Aggregate data on gender differences. Peer-review P1 Audition P1 Applic. Rejected Passed Rejected Passed Female 892 31 923 14 17 31 Male 838 43 881 18 25 43 1730 74 1804 32 42 74 p1 = 892/923 = 0.966 p1 = 14/31 = 0.452 p2 = 838/881 = 0.951 p2 = 18/43 = 0.419 RD = 0.015 RR = 1.02 RD = 0.033 RR = 1.08 RC = 0.69 OR = 1.48 RC = 0.94 OR = 1.15 Peer-review P2 Audition P2 Applic. Rejected passed Rejected passed Female 761 31 792 19 12 31 Male 1094 100 1194 49 51 100 1855 131 1986 68 63 131 p1 = 761/792 = 0.961 p1 = 19/31 = 0.613 p2 = 1094/1194 = 0.916 p2 = 49/100 = 0.49 RD = 0.045 RR = 1.05 RD = 0.123 RR = 1.25 RC = 0.46 OR = 2.24 RC = 0.76 OR = 1.65 Figure 6: 4-fold contingency tables and discrimination measures. passing the peer-review and of being funded for propos- als in programs P1 and P2. Consider first the peer-review phase. Recall that the measures of risk difference (RD) and risk ratio (RR) compare the proportions of rejected proposals. Due to the small fraction of projects passing the phase, it turns out that RD and RR cannot highlight differences in the outcomes. Overall, the vast majority of proposals were rejected. In fact, RR is only 1.02 for P1 and 1.05 for P2; RD is only 1.5% for P1, and a modest 4.5% for P2. On the other hand, since relative chance (RC) com- pares the success rates, it highlights major differences: the chance of passing the peer-review for a female is only 69% of the chance of a male for program P1, and only 46% for program P2. Finally, since the odds ratio (OR) is the ra- tio between RR and RC, it highlights differences in both rejection and success rates. Consider now the audition phase. Rejected and funded projects are now more evenly distributed. The discrimination measures highlight no sig- nificant difference for program P1. Differences in program P2 are lower than the first phase, yet still moderately high. Are the different success rates of males and females due to legitimate characteristics or skill differences of the gender of applicants? In order to answer such a question, Figure 7 reports the distributions of the age of the PIs, of the number of her/his publications, and of a few costs of project proposals (young researchers, good reputation researchers, total cost, request grant). Distributions are distinguished for gender of the PI and for program of the project proposal. The distribution of age across gender highlights no difference for both programs. Notice that the distributions across programs are clearly distinguished due to the requirements of each program in the call for proposals. However, the plot of the number of publica- tions shows that males have a slightly higher productivity than females, for both programs P1 and P2. As an exam- ple, about 37% of females in program P2 have more than 20 publications, against a percentage of 48% for males. Turning our attention on project costs, we observe that proposals led by females require slightly lower costs for young researchers than proposals led by males, in both programs. The situation is similar for the total cost and the requested grant: the average total cost is e 980K for females and e 1080K for males. The distributions of costs for good reputation researchers are, instead, similar. No- tice that such costs are non-zero for only 19% of proposals in program P1 and only 10% in P2. Summarizing, even though an analysis of distributions provides some hints on gender differences, it is still too gross grained to draw any conclusion about the presence of discrimination. Aggregations at the level of the whole dataset may hide differences in smaller niches of data. Un- veiling these niches is precisely the objective of the discrim- ination discovery analysis. 4.2. Distributions of gender risk difference Let us instantiate the approach of situation testing (see Section 2.3) by exploring risk differences. Let r be a project proposal led by a female PI that did not pass the peer-review phase. The function diff (r) = p1 − p2 mea- sures the risk difference between the rejection percentage p1 of its k-nearest neighbor proposals headed by female PIs and the rejection percentage p2 of its k-nearest neighbor- ing proposals headed by male ones. Distance is measured on the basis of proposal’s characteristics that are (legally) admissible in affecting the (first or second phase) decision. We consider here all the features of Table 1 apart from the project evaluation results and the gender of the PI. Simi- larity is modeled via the distance function adopted in the experiments by Luong et al. (2011), which consists of the Manhattan distance of z-scores for continuous attributes in r, and of the percentage of mismatching attributes for discrete ones. The higher diff (r) is, the more the negative decision on proposal r is unexplainable by differences in the compared characteristics. The residual explanation is then the gender of the PI, which implies a prima facie evidence of gender discrimination, or the lack of further explana- tory variables – the omitted variables. A critical choice concerns how to set the k constant. Figure 8 (a,b) shows the distributions of diff () for k = 4, 8, 16, 32 with reference to proposals from programs P1 and P2. As k increases, the distributions tend to flatten (for k sufficiently large, the risk differences of all proposals collapse to a unique value). From now on, we fix k = 8, which means compar- ing each proposal with 0.9% of the proposals in program P1 (= 16/1804, where 16 is 2k, and 1804 is the overall number of proposals), and with 0.8% of the proposals in program P2. 8 0 20 40 60 80 100 26 28 30 32 34 36 38 40 % Age gender = Female, program = P1 gender = Male, program = P1 gender = Female, program = P2 gender = Male, program = P2 0 20 40 60 80 100 0 10 20 30 40 50 % Number of Publications gender = Female, program = P1 gender = Male, program = P1 gender = Female, program = P2 gender = Male, program = P2 0 20 40 60 80 100 100K 300K 500K 700K % Young Researcher Cost gender = Female, program = P1 gender = Male, program = P1 gender = Female, program = P2 gender = Male, program = P2 0 20 40 60 80 100 0 50K 100K 150K 200K 250K % Int. Good Reput Researcher Cost gender = Female, program = P1 gender = Male, program = P1 gender = Female, program = P2 gender = Male, program = P2 0 20 40 60 80 100 300K 700K 1.1M 1.5M 1.9M % Total Cost gender = Female, program = P1 gender = Male, program = P1 gender = Female, program = P2 gender = Male, program = P2 0 20 40 60 80 100 200K 600K 1.0M 1.4M % Requested Grant gender = Female, program = P1 gender = Male, program = P1 gender = Female, program = P2 gender = Male, program = P2 Figure 7: Cumulative distributions across gender of PIs and program of proposals. 0 0.2 0.4 0.6 0.8 1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 F ra c t. o f tu p le s w it h d if f ≥ t t program=P1 peer-review=rejected gender=Female k=4 k=8 k=16 k=32 (a) 0 0.2 0.4 0.6 0.8 1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 F ra c t. o f tu p le s w it h d if f ≥ t t program=P2 peer-review=rejected gender=Female k=4 k=8 k=16 k=32 (b) 0 0.2 0.4 0.6 0.8 1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 F ra c t. o f tu p le s w it h d if f ≥ t t program=P2 k=8 gender=Female (S1) scientific relevance <> 8 (S2) impact <> 7 (S3) technical value <> 15 (S4) quality of partnership <> 10 (c) Figure 8: Cumulative distributions of diff (). Recall that only the project proposals that receive the highest scores (S1)-(S4) pass the peer-review phase. It is interesting then to look at the distributions of diff () separately for each score. This is shown in Figure 8 (c), where the “benefit denied” decision is set as receiving a score lower than the maximum. The impact of the project (S2) appears as the most biased criteria. Distributions might also unveil forms of multiple dis- crimination. Figure 9 shows the distributions of risk dif- ference for two possibly discriminated groups in isolation, namely female PIs and PIs affiliated to institutions from the South of Italy (an historically disadvantaged region of Italy), and for PIs belonging to both groups. The two groups in isolation exhibit some risk difference, with gen- der bias being more prominent than bias against people from the South. However, no increase in risk difference can be observed for the sub-group of female PIs from the South when compared to the whole group of female PIs. 5. Case study: discrimination model extraction In this section, we start applying a discrimination dis- covery approach to the pre-processed datasets of propos- als in program P1 (resp., P2) with reference to the peer- review decision. We will not be considering the decision of the audition phase due to three motivations. First, the number of proposals involved in the second phase of the reviewing process is rather small, hence we run the risk of drawing no (statistically) significant conclusion. Second, the discrimination measures in Figure 6 highlight higher gender differences in the peer-review decisions than in the audition decisions, so we expect higher chances of finding non-negligible contexts of clear discrimination. Third, and most important, the set of features available in Table 1 appear adequate as control factors for the decision of the first phase only. In fact, peer-reviewers had access to the proposal text, to the curriculum and list of publications 9 0 0.2 0.4 0.6 0.8 1 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 F ra c t. o f tu p le s w it h d if f ≥ t t program=P2 k=8 gender=Female region=South gender=Female ∧ region=South Figure 9: Cumulative distributions of diff (). of the PI, and to the budget data. This is about the set of features listed in Table 1. On the contrary, the panel of national experts “entered in personal contact” with the PIs during the auditions, so their decisions are affected by additional factors not recorded in the data, e.g., physical characteristics of the PI, proficiency in speaking, motiva- tion, and appropriateness of answers to questions. The omitted-variable bias in analysing data with reference to the decision of the panel of national experts would then be considerably high. Consequently, any finding of possible discrimination would be questionable. 5.1. Before data mining: what can regression tell us? Data analysts from economic and social sciences have typically adopted logistic regression as a tool for testing an hypothesis of possible discrimination. Before starting our data mining analysis, let us then follow such an approach and discuss what conclusions can be made without apply- ing data mining methods. Logistic regression is a form of multiple linear regression: logit(P(Y = 1)) = α + N∑ i=1 βiXi where the logit of the dependent variable value Y = 1 is estimated as a linear function of the independent vari- ables X1, . . . , Xn. The logit function logit(P(Y = 1)) = log(P(Y = 1)/(1 − P(Y = 1))) is the log of the odds of the probability P(Y = 1). By exponentiating the equation sides, we obtain: P(Y = 1) 1 − P(Y = 1) = eα+ ∑N i=1 βiXi = eα N∏ i=1 eβiXi The value βi can then be interpreted as the variation co- efficient of the logarithm of the odds of the event Y = 1 due to a linear variation of the factor Xi, all other control factors being constant. A nominal feature X with values v1, . . . , vk is modeled in this framework by k − 1 indepen- dent indicator variables X = v1, . . . X = vk−1. The coeffi- cients of these features model the variation of the logit of P(Y = 1) with respect to the default value X = vk. Model for P1 Model for P2 Variable Coeff. (Std. err) Coeff. (Std. err) gender = Female -0.33 (-0.39) -0.87 (0.30) [***] region = North 0.14 (0.30) 0.22 (0.22) region = South -0.02 (0.35) -0.42 (0.28) inst type = Univ -0.42 (0.62) -0.43 (0.40) inst type = Cons -0.36 (0.50) 0.04 (0.45) age -0.03 (0.09) 0.01 (0.04) pub num 0.01 (0.01) -0.01 (0.01) avg aut 0.01 (0.01) 0.03 (0.02) f partner num -0.02 (0.25) 0.12 (0.16) tot cost 0.10 (0.35) 0.03 (0.26) fund req -0.13 (0.50) -0.05 (0.37) fund req perc 0.51 (0.37) 0.39 (0.28) yr num -0.11 (0.26) -0.56 (0.19) [***] yr cost -0.09 (0.35) -0.03 (0.26) yr perc 0.31 (0.26) 0.27 (0.19) grr num -0.28 (0.46) 0.22 (0.24) grr cost -0.09 (0.35) -0.03 (0.26) grr perc 0.16 (0.32) 0.26 (0.21) d1 l1 = PE -0.42 (0.33) -0.34 (0.25) d1 l1 = SH 0.08 (0.33) 0.03 (0.29) Table 3: Logistic regression models for datasets of proposals for pro- grams P1 and P2. The dependent variable is peer-review = passed. Coefficients marked by [***] are statistically significant at the 99% confidence level. Table 3 shows logistic regression models for the datasets of proposals in program P1 and P2. The event Y = 1 is here peer-review = passed. Standard errors and statisti- cal significance of regression coefficients are also shown. In both models, the regression coefficient of the indicator variable gender = Female is negative, which means that, all other factors being equal, female PIs have lower odds of passing the peer-review phase: by a factor of e−0.33 = 0.72 for program P1, and of e −0.87 = 0.42 for program P2 w.r.t. the odds of male PIs. For program P2, the null hypothesis that the coefficient is zero is rejected at the 99% level of statistical significance. The region of the in- stitution of the PIs affects the odds of passing the peer review as well, particularly for proposal in program P2: PIs from the North of Italy have higher chances, whilst those from the South have lower ones. Variables on age, number of publications, average number of authors in pub- lications have coefficients close to zero. Concerning cost variables, proposals with higher percentage of costs cov- ering young researchers (variables yr perc, fund req perc) have higher odds of passing the peer-review. This is not unexpected, since the call explicitly states the objective of funding start-up research groups of young researchers. However, proposals with large (yet, young) groups (vari- able yr num) are disfavored in the peer-review decision. Moreover, competition appears to be harder in the area of Physical sciences and Engineering (PE) rather than in Humanities (SH), and Life Sciences (LS); and for PIs from the University (variable inst type = Univ) compared to PIs from other institutions. Finally, the literature on discrimination analysis accounts for an included-variable bias (Killingsworth, 1993), namely for control variables that incorporate some form of gender discrimination. One such variable is f partner num, i.e., the number of female principal or associate investigators. Since its coefficients 10 Accuracy Precision Recall f-Measure Size Id Set Alg CS P1 P2 P1 P2 P1 P2 P1 P2 P1 P2 1 D13 Jrip Y 48.4 45.2 33.1 40.2 97.0 93.4 0.49 0.56 35 23 2 D19 C4.5 N 77.6 70.7 54.3 58.7 83.1 74.5 0.66 0.66 302 184 3 D19 C4.5 Y 56.4 53.0 36.6 44.0 93.5 92.3 0.53 0.60 78 73 4 D19 Jrip Y 53.5 45.4 35.4 40.4 97.0 95.5 0.52 0.57 21 11 5 D19 Part N 72.3 66.6 47.9 54.5 77.9 67.8 0.59 0.60 74 33 6 D19 Part Y 61.9 57.3 39.7 46.4 90.9 88.8 0.55 0.61 74 87 7 D27 C4.5 N 82.2 74.5 62.3 63.4 78.8 75.9 0.7 0.69 2051 4645 8 D27 C4.5 Y 50.7 37.6 33.5 37.6 92.2 100 0.49 0.55 9 325 9 D27 Jrip Y 50.0 46.5 33.7 41.0 96.1 96.9 0.5 0.58 12 11 10 D27 Part Y 61.2 53.1 38.1 43.7 80.1 86.7 0.52 0.58 128 113 Table 4: Top 10 classification models of discriminated proposals. are small, it appears that gender discrimination bias, if present, is directed mainly against the PI, and not against the group of investigators. The conclusions drawn from Table 3 are certainly more informative than the explorative data analysis of Section 4. However, whilst they reveal some gender bias at the level of the whole datasets, there is no indication of whether this is uniformly distributed or whether there are some contexts with a very high bias. Unfortunately, the sta- tistical regression approach is limited to the verification of an hypothesis. Thus, one should explicitly figure out a possible context and re-compute a regression model for proposals in such a context. The purpose of our data min- ing approach is precisely to let such contexts emerge as a result of the analysis. 5.2. A classification model of the discriminated proposals The number of proposals led by female PIs that did not pass the peer-review phase amounts at 892 for pro- gram P1 and at 761 for program P2. Now, we intend to extract from these two datasets a global description of pro- posals whose negative decision is discriminatory according to the legal methodology of situation testing. We proceed as follows. First, we set a threshold value t to the maxi- mum admissible risk difference. Values of risk differences greater than 0.05 (i.e., 5%) have been considered prima facie evidence of discrimination in some legislations and law cases, and Figure 8 (a) supports this choice in prac- tice. In order to be more stringent, we assume from now on the higher threshold t = 0.10. Second, a proposal r is labeled as discriminated if its risk difference is greater or equal than t – technically, we introduce a binary attribute disc defined as disc(r) = true iff diff (r) ≥ 0.10. These two steps allow us for reducing the problem of characterizing discriminatory decisions to the standard problem of induc- ing a classification model, where the class attribute is the newly introduced attribute disc. The resulting datasets have a distribution of disc = true and disc = false values of 26-74% for program P1, and of 38-62% for P2. Since the intended use of the classification models is to describe global conditions under which a proposal led by a female PI is rejected at the peer-review phase with a risk difference of 0.10 or above, we restrict the search space to classification models that are readily interpretable (e.g., before a court in a law case). We experimented classification rule models (RIPPER by Cohen (1995), and PART by Frank and Witten (1998)) and decision trees (C4.5 by Quinlan (1993)). Classification models are eval- uated by the objective interestingness measures of accu- racy, precision, recall and f-measure for the class value disc = true using a 10-fold cross validation. The actual classification model is extracted from the whole dataset. Other settings that have been experimented are mainly concerned with tackling the unbalanced distribution of class values, and they include standard approaches: uni- form resampling of the training folds, cost-sensitive in- duction of classifiers,8 and meta-classification approaches (bagging and boosting). We relied on the Weka tool by Witten and Frank (2011) for algorithms and as experi- mental environment. Finally, we also varied the set of predictive attributes in order to evaluate the explanatory power of different subsets: • set D13 includes a subset of features of the PI (age, title, pub num, avg aut), of project costs (yr num, yr cost, grr num, grr cost, tot cost, fund req), of the research area (d1 lv1, d1 lv2) as well as the class attribute disc; • set D19 adds features of project costs (fund req perc, yr perc, grr perc), of the PI (inst type, region) and of the participants (f partern num); • set D27 also includes the remaining attributes of the ERC hierarchy and the attribute city of the PI. Table 4 reports the top 10 classification models obtained. For each model, the table includes: the set of predictive at- tributes, the extraction algorithm, whether cost-sensitive (CS) classification is adopted, and performance measures for both program P1 and P2. All of the top 10 clas- sifiers use resampling, whilst none of them adopt meta- classification. The size of a model measures its structural complexity: for classification rules, it is the number of rules; for decision trees, it is the number of leaves. A few comments follow on the lessons learned in tuning models and parameters. First, resampling of the training 8The best performance is obtained with a cost of misclassifying disc = true set to 2.5 times the cost of misclassifying disc = false. 11 folds reveals itself as an effective technique, improving per- formances both in term of accuracy and f-measure, and irrespectively of the model type and subset of attributes. Using in addition cost-sensitive classification does not im- prove further. Compare for instance rows 2 vs 3, 5 vs 6, 7 vs 8 from Table 4, where the only difference between the pairs is in the use or not of misclassification costs. Second, the impact of the set of predictive attributes is dependent on the classification model. Jrip and PART both benefit from larger sets as per accuracy and f-measure when mov- ing from D13 to D19, but then the additional attributes in D27 worsen the performances. Contrast for example rows 1 vs 4 vs 9, and 6 vs 10. This also holds for C4.5 models when using misclassification costs (see rows 3 vs 8). How- ever, when using resampling only, there is an improvement from D19 to D27 (rows 2 and 7). C4.5 with resampling on D27 (row 7) turns out to be the best model with respect to both accuracy and f-measure. Third, we highlight the im- portance of extracting models that trade off performance with simplicity. The best model (row 7) is, unfortunately, the most complex one. The global description of discrim- inatory decisions it provides is accurate but sparse in too many conditions, whose validation, e.g., by a legal expert, is impractical. This motivates the search for a few local contexts of possible discrimination. 6. Case study: rule reasoning and validation In this section, we report four rules filtered out from the top 10 classifiers. They have been ranked in the top po- sitions on the basis of both objective measures (precision, recall, average diff (), odds ratio) and subjective ones (in- terpretability, relation with known stereotypes). The an- tecedent of a rule unveils a context of prima facie evidence of gender discrimination. Proposals led by female PIs in such a context observe different decisions (with risk differ- ence of at least 0.10) of peer-reviewers between projects with similar characteristics led by male PIs and projects with similar characteristics led by female PIs. We validate the statistical significance of such a context by means of logistic regression. This way, we merge the capability of the k-NN as a situation testing approach for discovering contexts of possible discrimination with the capability of statistical regression for testing hypothesis of possible dis- crimination – thus obtaining the best of the two worlds. A technical note is in order. For each of the four rules, all proposals led by female PIs in the context of the rule result to have been rejected at the peer-review. As a conse- quence, the coefficient of the independent variable gender = Female in a logistic regression model cannot be calcu- lated – this is known as the separation problem (Heinze, 2006). We will then apply the Firth logistic regression (Firth, 1993), also called penalized maximum likelihood method, which takes into account such a problem. For the same reason, we will calculate the odds ratio (OR) of a rule by applying the plus-4 correction, which consists of 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 -0.4 -0.2 0 0.2 0.4 F ra c t. o f tu p le s w it h d if f ≥ t t (d1 − lv1 = LS) and (yr − cost ≥ 244,000) and (yr − num ≥ 2) and (avg − aut ≥ 8.4) and (pub − num ≤ 12) gender=Female gender=Male Figure 10: Cumulative distributions of diff () for proposals satisfying the antecedent of rule R1. adding a fictitious +1 to each cell in the contingency table of Figure 1. 6.1. Rule R1: life sciences in program P1 The first rule concerns proposals in program P1. It highlights a possible discrimination against female PIs in the area of Life Sciences (LS). R1: (d1_lv1 = LS) and (yr_num >= 2) and (yr_cost >= 244,000) and (pub_num <= 12) and (avg_aut >= 8.4) => disc=yes [prec=1.0] [rec=0.095] [diff=0.165] [OR=11.14] The antecedent of the rule points out a context of research proposals requiring two or more young researchers, having a cost for them of e 244K or more, and such that the PI has at most 12 publications with a mean number of authors of 8.4 or more. There are 33 proposals in the contex: 8 led by male PIs, 2 of which passed the peer-review, and 25 led by female PIs, none of which passed the peer-review. All of the 25 proposals led by female PIs have been labeled as discriminated, i.e., the precision of the rule is 100%. Among the proposals led by female PIs that are labeled as discriminated, 9.5% satisfy the antecedent of the rule, i.e., the recall of the rule is 9.5%, which makes the context rather relevant for the anti-discrimination analyst. With reference to proposals of the LS area only, recall lifts up to 27%. The average risk difference measure of the 25 pro- posals led by female PIs is 16.5%, which is much higher than the discrimination threshold value of 10%. Figure 10 reports the cumulative distributions of diff () for proposals satisfying the antecedent of the rule R1 distinguishing fe- male and male led projects. This is more informative than simply the average risk difference. Moreover, it highlights the other side of discrimination, namely favoritism: pro- posals led by males exhibit very low or even negative risk differences, or stated otherwise, they have been favored in comparison to similar projects led by female PIs. Finally, the (corrected) odds risk is 11.14: the odds of proposals led by female PIs of being rejected at the peer-review is 11.14 times the odds of those led by male PIs. Rule R1 unveils then a possible gender discrimination in the Life Science area when proposals are ambitious (more 12 Rule R1 Rule R2 Rule R3 Rule R4 Variable Coeff. (Std. err) Coeff. (Std. err) Coeff. (Std. err) Coeff. (Std. err) gender = Female -1.64 (1.40) [*] -1.37 (1.38) [**] -0.86 (1.22) -1.20 (0.63) [***] inst type = Univ 0.11 (2.41) -1.58 (1.28) inst type = Cons. -0.70 (1.63) [**] age -0.05 (0.46) 0.03 (0.42) 0.01 (0.44) 0.01 (0.20) pub num 0.12 (0.22) 0.01 (0.10) 0.01 (0.06) -0.01 (0.02) avg aut 0.03 (0.16) 0.01 (0.46) -0.13 (0.24) 0.04 (0.06) tot cost 0.87 (1.83) 0.01 (0.01) 0.14 (1.15) -0.19 (0.68) fund req -1.25 (2.62) 0.01 (0.01) -0.21 (1.64) 0.27 (0.97) fund req perc 0.40 (1.83) -1.7 (2.48) 0.83 (0.65) yr num 0.58 (1.38) 0.03 (0.66) 1.49 (1.83) -0.93 (0.51) yr cost -0.87 (1.83) 0.01 (0.01) -0.15 (1.15) 0.19 (0.68) yr perc 0.22 (1.41) -1.0 (1.5) -0.12 (0.70) 0.52 (0.46) grr num 0.27 (0.81) 0.58 (0.43) grr cost -0.15 (1.15) 0.19 (0.68) grr perc -0.11 (0.70) 0.24 (0.51) d1 l1 = PE 0.24 (1.93) 0.78 (0.66) d1 l1 = SH 0.25 (1.85) 0.59 (0.83) d1 l2 = LS2 -1.17 (1.88) d1 l2 = LS3 -0.72 (2.24) d1 l2 = LS4 -0.17 (1.63) d1 l2 = LS6 -0.21 (1.70) d1 l2 = LS7 0.14 (2.10) Table 5: Firth logistic regression models for the datasets of proposals satisfying the antecedent of rules R1-R4. The dependent variable is peer-review = passed. Coefficients marked by [*], [**], and [***] are statistically significant at the 90%, 95% and 99% confidence level respectively. Blank cells are due to control variables with unique values (e.g., d1 lv 1 is always “LS” in rule R1), or to control variables omitted due to high standard errors. f partner num is not part of the model in order to account for the included-variable bias (see Section 5.1). than one young researcher to be hired) but the PI has a low productivity record of publications (at most 12) and high uncertainty on the PI’s effective contribution (large number of co-authors). The peer-reviewers of the LS area appear to have compensated the lack of knowledge on the skills of the PIs by some prior knowledge or stereotype resorting to the gender of the PI, with females being dis- advantaged. This phenomenon is known as statistical dis- crimination or rational racism (Romei and Ruggieri, 2013) – as opposed to taste-based discrimination which is moti- vated by prejudice. Table 5 reports the Firth logistic regression model for the proposals satisfying the antecedent of rule R1. All other factors being equal, female PIs have e−1.64 = 0.194 the odds of male PIs of passing the peer-review.9 The coefficient −1.64 is greater (in absolute value) than −0.33, the one computed over the whole dataset of proposals in program P1 (see Table 3). More important, the coefficient is now significantly non-zero at 90% confidence level. 6.2. Rule R2: physical and analytical chemical sciences in program P2 A context of possible discrimination for proposals in program P2 is unveiled by the following rule: R2: (d1_lv2 = PE4) and (tot_cost >= 1,358,000) and (age <= 35) => disc=yes [prec=1.0] [rec=0.031] [diff=0.194] [OR=4.50] where PE4 is the Physical and Analytical Chemical Sci- ences panel, at the second level of the ERC hierarchy. The 9Here we deal with the odds of passing, and low values denote high burden. The odds ratio (OR) deals with the odds of being rejected, and high values denote high burden. context of rule R2 concerns proposals with high budget led by young PIs. There are 9 proposals led by male PIs, 2 of which passed the peer-review, and 11 proposals led by fe- male PIs, none of which passed it. The recall of rule R2 is 3.1%, i.e., the context covers 3.1% of the proposals labeled as discriminated. The precision of rule R2 is 100%, mean- ing that all of the 11 proposals in the context led by female PIs have been labeled as discriminated. The average risk difference is 19.4%, and the odds ratio is 4.5. Table 5 shows the Firth logistic regression model for the proposals in the context of rule R2. All other factors being equal, female PIs have e−1.37 = 0.254 the odds of male PIs of passing the peer-review. The coefficient −1.37 is greater (in absolute value) than −0.87, the one computed over the whole dataset (see Table 3), yet it is significantly non-zero at the lower confidence level of 95%. Summarizing, rule R2 unveils a niche of proposals with a gender bias higher than the average bias of the whole dataset of proposals in program P2. 6.3. Rule R3: expensive projects in program P1 A second rule about proposals in the program P1 is the following: R3: (yr_cost >= 187,000) and (grr_cost >= 70,000) => disc=yes [prec=0.86] [rec=0.052] [diff=0.161] [OR=5.77] The antecedent of the rule concerns proposals with high budget for young researchers and for good reputation re- searchers. We checked that such a context is disjoint from the one of rule R1, where all proposals had no budget for good reputation researchers. There are 16 proposals led by male PIs in the context, 4 of which passed the peer-review, 13 Figure 11: Scatter plot of grr cost over yr cost for proposals satis- fying the antecedent of R3. and 14 proposals led by female PIs, none of which passed it. Precision of the rule is 86%, meaning that 12 out of the 14 proposals led by female PIs have been labeled as discriminated, with an average risk difference for the 14 proposals of 16.1%. The recall of the rule is 5.2%, hence rules R1 and R2 cover together 14.7% of the proposals la- beled as discriminated. The odds ratio of the proposals in the context is 5.77. Intuitively, the peer-reviewers of program P1 seem to trust more male PIs than female PIs in managing projects with high costs of personnel, namely young and good rep- utation researchers. Does rule R3 unveil then a case of actual discrimination? Firth logistic regression on the dataset of proposals of the context (see Table 5) shows a coefficient for gender=Female of -0.86, which, however, is not statistically significant at 90% confidence level – i.e., the null hypothesis that the coefficient is actually zero cannot be rejected. We proceed by analysing further the proposals in the context. The scatter plot in Figure 11 reports the costs of good reputation researchers over the costs of young researchers. It highlights that proposals led by female PIs tend to a higher proportion of good reputa- tion researcher costs over proposals led by male PIs. This could somehow be in contrast with the intended objectives of the call for proposals, which require a substantial hiring of young researchers. Therefore, it may well be the case that peer-reviewers have scored worse those proposals re- lying too much on the hiring of senior researchers. This could be argued as a legitimate and objective justification for the disparate treatment of female PIs, an exception admitted by the anti-discrimination laws. Whether this is the case or not, however, is a matter of legal argumenta- tion. Strictly speaking, the call for proposals did not set an explicit maximum threshold on the proportion of good reputation researcher costs over young researcher costs. 6.4. Rule R4: young PIs in program P2 A second rule for proposals in the program P2 is the following: R4: (age <= 32) and (fund_req >= 310,000) => disc=yes [prec=0.52] [rec=0.12] [diff=0.07] [OR=9.6] The antecedent of the rule concerns younger PIs with a fund request greater or equal than e 310K. Intuitively, this can be interpreted as a negative bias against younger female PIs who require a medium-high grant. There are 201 proposals in such a context: 131 with male PIs, 16 of which passed the peer-review; and 70 with female PIs, only 1 of which passed the peer-review. The odds ratio is 9.6. Precision of the rule is moderately higher than 38% – the overall percentage of proposals labeled as discrimi- nated in program P2. That is, about half of the 69 female PIs whose proposal was rejected showed a risk difference greater or equal than 10%. In fact, the average risk dif- ference is only 7%. However, recall is rather high: 12% of the proposals labeled as discriminated in program P2 are in the context of rule R4. Finally, we checked that the overlap of proposals in the contexts of both rule R4 and R2 is minimal, with only 3 proposals led by male PIs and 1 led by a female PI. Such overlap originates from the fact that rules R2 and R4 are selected from two different classification models. Consider the logistic regression model for proposals in the context of rule R4 (see Table 5). All other factors being equal, female PIs have e−1.20 = 0.301 the odds of male PIs of passing the peer-review. The coefficient −1.20 is smaller (in absolute value) than the one of rule R2, but significant at the higher confidence level of 99%. Moreover, it is greater (in absolute value) than the one of the whole dataset (see Table 3). Summarizing, rule R4 highlights a context with higher gender bias than in the whole dataset of proposals in program P2. This and the contexts of the other rules were not previously known as possible stereo- types of discriminatory behaviors. Rather, they have been the result of a discrimination discovery investigation. 7. A KDD process in support of discrimination discovery Since personal data in decision records are highly di- mensional, i.e., characterized by many multi-valued vari- ables, a huge number of possible contexts may, or may not, be the theater for discrimination. In order to extract, select, and rank those that represent actual discrimina- tory behaviors, an anti-discrimination analyst should ap- ply appropriate tools for pre-processing data, extracting prospective discrimination contexts, exploring in details the data related to the context, and validating them both statistically and from a legal perspective10. Discrimination discovery consists then of an iterative and interactive pro- cess. Iterative because, at certain stages, the user should have the possibility of choosing different algorithms, pa- rameters, and evaluation measures or to iteratively repeat some steps to unveil meaningful discrimination patterns. 10As observed by Gastwirth (1992), the objectives of science and the law often diverge, with rigorous scientific methods conflicting with the adversarial nature of the legal system. 14 Figure 12: The KDD process of situation testing for discrimination discovery. Interactive because several stages need the support of a domain expert in making decisions or in analysing the re- sults of a previous step. We propose here to adopt the process reported in Figure 12, which is specialized in the use of the situation testing for extracting contexts of possi- ble discrimination. The process has been abstracted from the case study presented in the previous sections, and it consists of four major steps. Data Understanding and Preparation. The availability of historical data concerning decisions made in socially- sensitive tasks is the starting point for discovering dis- crimination. We assume a collection of data sources stor- ing historical decisions records in any format, including relational, XML, text, spreadsheets or any combination of them. Standard data pre-processing techniques (selec- tion, cleansing, transformation, outlier detection) can be adopted to reach a pre-processed dataset consisting of an input relation as the basis for the discrimination analysis. The grain of tuples in the relation is that of an individual (an applicant to a loan, to a position, to a benefit). Three groups of attributes are assumed to be part of the relation: protected group attributes: one or more attributes that identify the membership of an individual to a pro- tected group. Attributes such as sex, age, marital status, language, disability, and membership to po- litical parties or unions are typically recorded in ap- plication forms, curricula, or registry databases. At- tributes such as race, skin color, and religion may be not available, and must be collected, e.g., by survey- ing the involved people; decision attribute: an attribute storing the decision for each individual. Decision values can be nominal, e.g., granting or denying a benefit, or continuous, e.g., the interest rate of a loan or the wage of a worker; control attributes: one or more attributes on control fac- tors that may be (legally) plausible reasons that may affect the actual decision. Examples include attributes on the financial capability to repay a loan, or on the productivity of an applicant worker. Risk Difference Analysis. For each tuple of the input relation denoting an individual of the protected group, the additional attribute diff is calculated as the risk difference between the decisions of its k nearest-neighbors of the pro- tected group and the decisions for its k nearest-neighbors of the unprotected group (see Section 2.3). We call the output of the algorithm the risk difference relation. The value k is a parameter of the algorithm. A legitimate ques- tion is how to choose the “right” k? A large k means that every instance is a neighbor, hence the distribution of diff tends towards a unique value. Conversely, for a small k, we run the risk that the distribution is affected by ran- domness. As a consequence, a study of the distribution of diff for a few values of k is required. This means iterating the calculation of the diff attribute. Exploratory analy- sis of diff distributions may also be conducted to evaluate risk differences at the variation of: the protected group under consideration, e.g., discrimination against women or against youngsters; the compound effects of multiple discrimination grounds, e.g., discrimination against young women vs discrimination against women or youngsters in isolation; the presence of favoritism towards individuals of a dominant group, e.g., nepotism. Once again, this requires iterating the calculation of diff by specifying a different protected group attribute to focus on. Discrimination Model Extraction. By fixing a thresh- old value t, an individual r of the protected group can be labeled as discriminated or not on the basis of the condi- tion diff (r) ≥ t. We introduce a new boolean attribute disc and set it to true for a tuple r meeting the condition above, and to false otherwise. A global description of who 15 has been discriminated can now be extracted by resorting to a standard classification problem on the dataset of indi- viduals of the protected group, where the class attribute is the newly introduced disc attribute. Accuracy of the clas- sifier is evaluated with objective interestingness measures, e.g., precision and recall over the disc = true class value. The intended use of the classifier is descriptive, namely to provide the analyst with a characterization of the individ- uals that have been discriminated. The choice of the value t should then be supported by laws or regulators.11 For in- stance, the four-fifths rule by the US Equal Employment Opportunity Commission (1978) states that a job selec- tion rate (the RC measure from Figure 1) lower than 80% represents a prima facie evidence of adverse impact. Since the intended use of the extracted classifier is de- scriptive, classification models that are easily interpretable by (legal) experts and whose size is small should be pre- ferred. In other words, one should trade accuracy for sim- plicity. Classification rules and decision trees are natural choices in this sense, since rules and tree paths can eas- ily be interpreted and ranked. The extracted classification models provide a global description of the disc class val- ues. They are stored in a knowledge base, for comparison purposes and for the filtering of specific contexts of dis- crimination – as described next. Rule Reasoning and Validation. The actual discovery of discriminatory situations and practices may reveal it- self as an extremely difficult task. Due to time and cost constraints, an anti-discrimination analyst needs to put under investigation a limited number of contexts of pos- sible discrimination. In this sense, only a small portion of the classification models can be analysed in detail, say the top N rules or the top N paths of a decision tree. We propose to concentrate on classification rules of the form: (cond_1) and ... and (cond_n) => disc=yes [prec] [rec] [diff] [OR] where (cond 1) and ... and (cond n) is obtained from a classification model (from a rule or from a path of a de- cision tree). Rules are ranked on the basis of one or more interestingness measures, including: precision [prec] (proportion of discriminated individuals among those of the protected group which satisfy the antecedent), recall [rec] (proportion of the overall discriminated individuals covered by the antecedent), average value of diff [diff] (a measure of the degree of discrimination observed by individuals of the protected group which satisfy the an- tecedent), and odds ratio [OR] (a measure of the burden of negative decisions on the individuals of the protected group when compared to those of the unprotected group satisfying the antecedent of the rule). Notice that [diff] and [OR] may rank rules differently because they con- trast distinct sets of groups (the 2k nearest neighbors, and 11A relevant question is the other way round – namely, can data mining help law makers and regulators in the definition of appropri- ate values for t? the members of the unprotected group satisfying the an- tecedent of the rule). Statistical validation is accounted for in our approach by relying on logistic regression, which is a well-known tool in the legal and economic research com- munities. Readability and interpretability should also be taken into account by preferring rules with fewer items in the antecedent, thus trading interestingness with simplic- ity. As an alternative approach to the selection of rules from the classifiers extracted in the previous step of the process, one could mine all classification rules of the form above by means of association rule mining. Unfortunately, this results in a huge number of rules covering overlapping contexts of possible discrimination. This is what occurs, for instance, in the rule extraction and filtering approach of Ruggieri et al. (2010a). The rules selected in this step of the process, however, are still subject to further consid- eration, e.g., by a legal expert, who may require further data exploration and, possibly, iteration of previous steps of the process. Therefore, the number of selected rules must be reasonably low. Selecting the best rules from the best performing classification models is then a means to keep the number of (overlapping) rules to a minimum. 8. Conclusions The contribution of this paper has been threefold. First, we have presented a complex case study in the context of scientific project funding using real data from an Italian national call for proposals. The application of dis- crimination discovery methodologies based on data mining to real case studies was lacking in the existing literature. So far, experiments and analyses have been conducted on “general purpose” datasets, not explicitly collected or pro- cessed for discrimination analysis. As a consequence, the reported analyses have been necessarily partial, typically being limited to summary statistics (e.g., number of possi- bly discriminatory contexts found), to artificial examples, and to generic argumentations on the results found. This is a serious drawback that limits the acceptance of knowl- edge discovery methods in practice. Second, we have proposed and applied a methodology that couples legal methods (situation testing) for the def- inition of cases of possible discrimination, data mining methods (a variant of k-NN plus standard classification) for the search of contexts of possible discrimination, and regression analysis for the statistical validation of such contexts. This approach overcomes the statistical anal- ysis12 of discrimination conducted in the social sciences, economics, and legal literature, which is limited to the verification of an hypothesis of possible discrimination on 12The contrast between the two approaches above is an instance of the two general “cultures” in the use of statistical modeling (Breiman, 2001): data modeling vs the algorithmic modeling. 16 the whole set of past decision records. Such an analysis re- veals to be inadequate to cope with the problem of search- ing for unknown or unforeseen contexts of discriminatory decisions hidden in a large dataset. On the contrary, the rules discussed in Section 6 unveil prima facie evidence of discrimination when certain project costs are above a threshold value. Both the cost attribute and the threshold value, however, come as the result of the analysis – they were not an a priori hypothesis to be verified. The extrac- tion of contexts of discrimination is precisely the objective of discrimination discovery. Third, from the specific case study, we have abstracted a general process of discrimination discovery. The adopted methodology relies on an implementation of the legal prac- tice of situation testing using a variant of k-NN, and then on extracting and reasoning about a classification model. The steps of the methodology have been described in the process of Figure 12, which represents a guidance for re- searchers and anti-discrimination analysts. We believe that this contribution can provide higher confidence about the replicability of the analyses and their applicability to real cases. Some issues remain open for future investigation. With reference to the case study, further analysis will be made possible by enriching the available dataset with additional control features, e.g., some accurate measures of the sci- entific productivity of applicants and of their professional network. This was not possible in our study, since our in- put data were anonymized. Concerning the tools adopted, while the k-NN algorithm remains the core component of the proposed process, of particular interest is the formal- ization of the deductive component, in which the extracted classification models are filtered, refined, transformed and validated into useful knowledge. We aim at designing a post-processing tool, by adapting the XQuake system (Romei and Turini, 2010), able to support the user in the deductive part of the process. Finally, throughout the pa- per, we have assumed the availability of a feature denoting the protected group under analysis – in the case study, the gender of the PIs. In indirect discrimination discovery, this assumption does not hold, e.g., because race, ethnicity, or sexual orientation may be not recorded in data. In such cases, a different approach must be devised. References Bendick, M., 2007. Situation testing for employment discrimination in the United States of America. Horizons Stratégiques 3 (5), 17– 39. Bentley, J. T., Adamson, R., 2003. Gender differences in the careers of academic scientists and engineers: A literature review. Special report, National Science Foundation, http://www.nsf.gov. Bornmann, L., Daniel, H.-D., 2005. Selection of research fellowship recipients by committee peer review: Reliability, fairness and pre- dictive validity of Board of Trustees’ decisions. Scientometrics 63 (2), 297–320. Bornmann, L., Daniel, H.-D., 2009. The state of h index research. EMBO reports 10 (1), 2–6. Bornmann, L., Mutz, R., Daniel, H.-D., 2008. Latent markov mod- eling applied to grant peer review. Journal of Informetrics 2 (3), 217–228. Breiman, L., 2001. Statistical modeling: The two cultures. Statistical Science 16 (3), 199–231. Brouns, M., 2000. The gendered nature of assessment procedures in scientific research funding: The Dutch case. Higher Education in Europe 25 (2), 193–199. Calders, T., Verwer, S., 2010. Three naive bayes approaches for discrimination-free classification. Data Mining & Knowledge Dis- covery 21 (2), 277–292. Ceci, S. J., Williams, W. M., 2011. Understanding current causes of women’s underrepresentation in science. Proc. of the National Academy of Sciences 108 (8), 3157–3162. Cohen, W. W., 1995. Fast effective rule induction. In: Proc. of Int. Conf. on Machine Learning (ICML 1998). Morgan Kaufmann, pp. 115–123. Council of the E.U., 1999. Resolution 1999/C 201/01 on Women and Science. http://eur-lex.europa.eu. Custers, B. H. M., Calders, T., Schermer, B. W., Zarsky, T. Z. (Eds.), 2013. Discrimination and Privacy in the Information So- ciety. Vol. 3 of Studies in Applied Philosophy, Epistemology and Rational Ethics. Springer. Equal Employment Opportunity Commission, 1978. Uniform guidelines on employee selection procedure. 43 FR 38295, http://www.gpo.gov. European Commission, 2009. The gender challenge in research fund- ing: Assessing the European national scenes. Directorate Gen- eral for Research, Science, Economy and Society, Unit L.4, http://ec.europa.eu. European Commission, 2012. Meta-analysis of gender and science research. Directorate General for Research and Innovation, Sector B6.2, http://www.genderandscience.org. Firth, D., 1993. Bias reduction of maximum likelihood estimates. Biometrika 80 (1), 27–38. Frank, E., Witten, I. H., 1998. Generating accurate rule sets without global optimization. In: Proc. of Int. Conf. on Machine Learning (ICML 1998). Morgan Kaufmann, pp. 144–151. Gastwirth, J. L., 1992. Statistical reasoning in the legal setting. The American Statistician 46 (1), 55–69. Goldstein, H., 2011. Multilevel Statistical Models, 4th Edition. Wi- ley. Hajian, S., Domingo-Ferrer, J., 2012. A methodology for direct and indirect discrimination prevention in data mining. IEEE Transac- tions on Knowledge and Data Engineering, to appear. Heinze, G., 2006. A comparative investigation of methods for logistic regression with separated or nearly separated data. Statistics in Medicine 25 (24), 4216–4226. Jayasinghe, U. W., Marsh, H. W., Bond, N. W., 2003. A multilevel cross-classified modeling approach to peer-review of grant propos- als. Journal of the Royal Statistical Society 166 (3), 279–300. Kamiran, F., Calders, T., 2012. Data preprocessing techniques for classification without discrimination. Knowledge and Information Systems 33, 1–33. Killingsworth, M. R., 1993. Analyzing employment discrimination: From the seminar room to the courtroom. American Economic Review 83 (2), 67–72. Larivière, V., Vignola-Gagné, E., Villeneuve, C., Gélinas, P., Gin- gras, Y., 2011. Sex differences in research funding, productivity and impact: an analysis of Québec university professors. Sciento- metrics 87 (3), 483–498. Ley, T. J., Hamilton, B. H., 2008. The gender gap in NIH grant applications. Science 322 (5907), 1472–1474. Luong, B. T., Ruggieri, S., Turini, F., 2011. k-NN as an implementa- tion of situation testing for discrimination discovery and preven- tion. In: Apté, C., Ghosh, J., Smyth, P. (Eds.), Proc. of the ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining (KDD 2011). ACM, pp. 502–510. Marsh, H. W., Jayasinghe, U. W., Bond, N. W., 2008. Improving the peer-review process for grant applications: Reliability, validity, bias, and generalizability. American Psychologist 63 (3), 160–168. 17 Mutz, R., Bornmann, L., Daniel, H.-D., 2012. Does gender matter in grant peer review? An empirical investigation using the example of the Austrian Science Fund. Journal of Psychology 220, 121–129. Pager, D., 2007. The use of field experiments for studies of employ- ment discrimination: Contributions, critiques, and directions for the future. The ANNALS of the American Academy of Political and Social Science 609 (1), 104–133. Pedreschi, D., Ruggieri, S., Turini, F., 2008. Discrimination-aware data mining. In: Li, Y., Liu, B., Sarawagi, S. (Eds.), Proc. of the ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining (KDD 2008). ACM, pp. 560–568. Quillian, L., 2006. New approaches to understanding racial prejudice and discrimination. Annual Review of Sociology 32 (1), 299–328. Quinlan, J. R., 1993. C4.5: Programs for Machine Learning. Morgan Kaufmann, San Mateo, CA. RAND, 2005. Is there gender bias in federal grant programs? RAND Infrastructure, Safety, and Environment Brief RB-9147- NSF, http://rand.org. Romei, A., Ruggieri, S., 2013. Discrimination data analysis: A multi- disciplinary bibliography. In: Custers, B. H. M., Calders, T., Schermer, B. W., Zarsky, T. Z. (Eds.), Discrimination and Privacy in the Information Society. Vol. 3 of Studies in Applied Philoso- phy, Epistemology and Rational Ethics. Springer, pp. 109–135. Romei, A., Ruggieri, S., Turini, F., 2012. Discovering gender dis- crimination in project funding. In: Proc. of the IEEE ICDM 2012 Int. Workshop on Discrimination and Privacy-Aware Data Mining (DPADM). IEEE Computer Society, pp. 394–401. Romei, A., Turini, F., 2010. XML Data Mining. Software: Practice and Experience 40 (2), 101–130. Rorive, I., 2009. Proving Discrimination Cases - the Role of Situa- tion Testing. Centre For Equal Rights & Migration Policy Group, http://www.migpolgroup.com. Ruggieri, S., Pedreschi, D., Turini, F., 2010a. Data mining for dis- crimination discovery. ACM Trans. on Knowledge Discovery from Data 4 (2), Article 9. Ruggieri, S., Pedreschi, D., Turini, F., 2010b. DCUBE: Discrimina- tion discovery in databases. In: Elmagarmid, A. K., Agrawal, D. (Eds.), Proc. of the ACM SIGMOD Int. Conf. on Management of Data (SIGMOD 2010). ACM, pp. 1127–1130. Sandström, U., Hällsten, M., 2008. Persistent nepotism in peer- review. Scientometrics 74 (2), 175–189. UNESCO, 2007. Science, Technology and Gender: An International Report, 4th Edition. UNESCO Publishing. Wenner̊as, C., Wold, A., 1997. Nepotism and sexism in peer-review. Nature 387 (5), 341–343. Wilson, R., 2004. Where the elite teach, it’s still a man’s world. The Chronicle of Higher Education 51 (15). Witten, I. H., Frank, E., 2011. Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations., 3rd Edition. Morgan Kaufmann, San Francisco. 18 
work_lhzdvddi6jelbnsefl2b735u2a	----	. I C O N L % ~ o 6 a i l --z Smartweld : A Knowledge-based Approach to Welding J. L. Mitchiner, S. D. Kleban, B. V. Hess, K. W. Mahin Sandia National Laboratories Albuquerque, New Mexico 87185-0722 jlmitch@ sandia.gov D. Messink IntelliCorp Mountain View. California 94040-22 16 Abstract Smartweld is a concurrent engineering system that integrates product design and processing decisions within an electronic desktop engineering environment. It is being developed to provide designers, process engineers, researchers and manufacturing technologists with transparent access to the right process information, process models, process experience and process experts, to realize “right the first time” manufacturing. Empirical understanding along with process models are synthesized within a knowledge-based system to identify robust fabrication procedures based on cost, schedule, and performance. Integration of process simulation tools with design tools enables the designer to assess a number of design and process options on the computer rather than on the manufacturing floor. Task models and generic process models are being embedded within user friendly GUI’s to more readily enable the customer to use the Smartweld system and its software tool set without extensive training. The integrated system architecture under development provides interactive communications and shared application capabilities across a variety of workstation and PC-type platforms either locally or at remote sites. Introduction Welding in the US is a $50B per year business which generates upwards of $7B of waste due to rework and scrap. Many of the problems associated with welding arise due to the high degree of ”trial and error” that typically accompanies the development of weld processes and schedules, weld joint designs, and material selection. Sandia material scientists and engineers have decades of experience in welding. They have written hundreds of memos and papers addressing welding issues from partlproblem specific memos to refereed journal and conference publications. In addition, Sandia analysts have developed weld process models and analysis tools that can predict thermal response and residual stress due to welding. This knowledge should be an important contributor toward reducing scrap and rework, however, it was only embodied in a select group of experts and was not routinely integrated into the design of welded assemblies. The goal of the Smartweld project is to collect and package legacy welding information, develop new welding models that advance welding from art to science, and to make that legacy information and new science available at the fingertips of designers, weld engineers, and analysts. The Smartweld system supports users from conceptual design through final, detailed design by incorporating models of varying fidelity and complexity. During conceptual design, it is sufficient for the designer to know that acceptable weld processes exist. During final design, the details of the weld schedule are necessary and may have to be simulated within a 3D Finite Element Analysis code to quantify the design margin. The Smartweld system addresses two classes of welding problems. First, for a few, important Sandia product families we use object models to specify default attributes such as alternative weld locations, weld function, candidate materials, part topology, and a finite element mesh that can be changed parametrically by changing part dimensions. For this class of problem, Smartweld can assist users from conceptual design through visualization of a 3D finite element analysis of heat dissipation during welding. Second, users want advice on generic welds, such as welding a lid to a housing or welding two flat plates together. In these cases, the user must provide information that is defaulted in the object models of the modeled product families. Smartweld advises on weld joint geometries, weld process, and weld schedules, but cannot aid in analysis without a user provided mesh. In the first case, Smartweld is more useful, but for a limited class of problems. In the second case, Smartweld provides less, but to a far broader class of problems. In both cases, the goal of Smartweld is to provide users with the best available data, information and knowledge to assist them in the design and analysis of welded assemblies. Smartweld can be accessed either directly or through a product-focused design environment to provide the welding perspective. In the standalone mode, a designer can enter Smartweld early in the design conceptualization stage and rapidly perform “what-if’ testing to estimate the impact of welding on the design. The user can access the weld advisor, the weld schedule database or the weld schedule optimization tools at any time during the design process without having to enter the overall Smartweld http://sandia.gov DISCLAIMER Portions of this document may be illegible in electronic image products. Images are produced from the best available original document. Smartweld Environment Figure 1. SmartWeld Architecture architecture. Later, in a more detailed design mode, the designer can enter the system through the product design environment with many of the design decisions complete. Smartweld then assists the user in developing a more detailed design of the weld. It automatically incorporates the weld and the joint geometry into the product solid model, and creates an executable weld specification. Weld Design Issues The main issues that must be addressed to design a predictably good weld include: Choosing a weld location. Within product families, usually there are a few locations that are acceptable based on issues such as, reachability by welding machines and location of thermally sensitive regions. Choosing a weld process and joint design. Different weld processes and joint designs have different properties, such as, depth of metal penetration, heat input, residual stress, distortion, sensitivity to loading and sensitivity to corrosive environments. Choosing the weld process and joint geometry is a core problem. Currently, we are considering fusion welds and solid state welds consisting of six arc and beam welding options, six resistance and friction welding options, and approximately forty weld joint geometries. Weld Schedule Development. The details of the weld process, and geometry including speed, amps and volts, filler material feedrates, and a full geometric description of the weld joint are are represented in the weld schedule. Evaluating the effect of welding on product performance. Welding joins metal by melting the metal of each piece and creating a fusion zone which is a composite of the original materials. The intense heat required for welding can cause both thermal and mechanical problems. One major concern during welding is damage to thermally sensitive regions near the weld. Changes in the mechanical properties of the part in the weld zone due to this heat treatment is a second major concern. System Design Smartweld is being built using a three-tiered client server architecture, see Figure 1. All communications between the user interface, the integrating application layer, and each of the tools are done through COR13Am objects. The goal is to create a system that allows teams of developers to work together without interfering with each other. For example, the user interface can now be either Motif and X- windows or JavaTM and Netscapem. The user interface is independent of the application layer. We want to negotiate an interface between the application layer developers and the analysis agent developers that can remain relatively invariant. The analysis agent developers can then improve and modify their code with no adverse effects on the application layer developers. Figure 2 . SmartWeld Graphical User Interface The system consists of: A user interface layer - Motif/X-windows or JavaTM/Netscapem an intelligent, integrating application layer, which integrates, within a single desktop environment, the tools and information sources needed by the user to make robust decisions a Weld Advisor, which captures Sandia welding knowledge in an object base a Weld Schedule database that stores the details of weld schedules and joint geometries used successfully in the past a Weld Schedule Optimization tool built on Matlabm an integrated Pro/Engineerm for CAD modeling P A T W for meshing a simple 2D steady state heat flow model (Rosenthal’s equation) that allows users to quickly and analytically evaluate temperatures near the weld new laser weld model for predicting thermal contours and residual stress using 3D FEA tools new GTA weld models for predicting thermal contours and residual stress for multi-pass welds using 3D FEA tools integrated visualization tools to see the results of the FEA analyses. The SmartWeld GUI, shown in Figure 2, has three major components: The upper section records important attributes of the current problem solving state. The middle section is highly structured to guide the user through the weld design and analysis tools required to produce a predictably good weld. The lower section allows experts direct access to information and tools, such as the World- Wide-Web, hypertext design guides, CAD tools, Matlabm, and finite element analysis tools. The middle section of the GUI is a network of nodes representing the steps required to design the weld, design the weld schedule, simulate the temporal temperature distribution during the welding process, and visualize the temperature distribution. This structure allows information to be passed seamlessly from one step to the next. The code in this area is able to both syntactically and semantically link all of the weld design and analysis codes. Each node is color-coded to immediately let the user know the status of the current problem - green nodes are available for execution; red ones cannot be performed because at least one precursor condition is not satisfied; blue ones represent completed tasks. Figure 3. Define Part/Weld Window Using Smartweld The first step in the weld design and analysis process is the initial definition of the part and weld. When the user clicks on it, the window shown in Figure 3 appears. The user can either import a design at this point or graphically choose a product family by clicking on an icon at the top of the window. In this example the product family has six different welds, Default dimensions (or dimensions passed from the product design environment) are shown on the screen and can be edited for “what-if‘ testing. The second step in the design and analysis process is the Weld Advisor. The Weld Advisor is an object-oriented knowledge base. Information about the weld location and function specified in Define Part/Weld starts the reasoning required to determine plausible weld scenarios. The welds are determined to be plausible based on knowledge about compatibility between materials, weld processes and joint geometries. The Advisor knows the default attributes of the product family when a part of that type is chosen in Define Part/Weld. The Advisor has both knowledge about welding in general and specific knowledge pertaining to product families. At this point, it begins to reason in the knowledge base. During the reasoning process, if information is required that is not available to the knowledge base, the user is queried. When all information requirements are met, the advisor ranks and scores each plausible weld scenario based on performance criteria. " Figure 4. The Results Screen from the Weld Advisor Each ranking criterion has an "importance" and a "weight" associated with it. "Required" criteria are hard constraints, and "desired criteria" are soft constraints. If a user wishes to omit one of the normal criteria from the evaluation function, the "don't care" importance for that criteria is chosen. Initially by default, all criteria have "required"importance as opposed to "desirable" or "don't care" and each has a weight of 50. Both the importance and weight of each criterion can be changed by the user to develop an understanding of the sensitivity of the weld scenario rankings to the importance and weighting of each criterion. This allows the user to rapidly perform trade-off assessments. In addition, an explanation facility exists which explains in detail how each weld scenario was evaluated. Each ranked weld process and joint geometry are classified into one of three color-code categories. Green welds satisfy all of the hard and soft constraints. Yellow welds satisfy the hard constraints, but fail one or more soft constraints. Red welds fail at least one hard constraint. Once the user selects a weld from the rank-ordered welds, the user can transparently access appropriate weld schedules from either a database or from optimization codes. The database of historical weld schedules describes in extensive detail all of the information needed to fully specify the new weld on the new part. The SQL selection is based on the material, weld process, joint geometry and material thickness. Information is stored about the component being welded, its piece parts, details on the joint geometry, and extensive information on the machine parameters during the weld. The database is an implementation of a highly normalized relational schema developed using the information modeling methodology, NIAM. Once a weld schedule is selected, the information is mapped to objects for further usage downstream of the advisor. At times, a validated weld schedule may not exist in the database, or the user may not want to use the weld schedules that do exist. In either case, Smartweld includes validated optimization models of several weld processes and joint geometries that can produce the necessary weld schedule information. The next step in the weld design and analysis process is to create the solid model of the part, if necessary, and to paste the selected weld joint into it. Pro/Engineer is our solid modeling CAD tool. A complex CAD model can either have been input into the Smartweld system from a design environment or can be automatically generated from decisions the user made in the Define PartJWeld step. If generated from the specification in the Define Part/Weld step, the model is created on-the-fly and saved. Next, for finite element analysis the part, if complex, Figure 5. A meshed cylinder. should be simplified to minimize meshing and thermal analysis calculations around features that do not impact the temperature and stress distributions generated by the weld. The system cannot automatically do this simplification, but does automatically bring up the model in Pro/EngineerTM for simplification and then saves it in the appropriate place. Next, a finite element mesh must be installed in the solid model. PATRANm is used for mesh generation. An example mesh for a cylindrical part is shown in Figure 5. For each product family, an expert analyst designs the first parameterized mesh. Once this mesh exists, the user can modify the mesh parametrically within bounds for "what-if'' testing. The user can also specify boundary conditions. For example, the user can decide to use a water-cooled fixture if heating is expected to cause a problem. With a water-cooled fixture, the surface temperature boundary condition under the fixture can be held to 30°C for the duration of the weld. The next step in the weld design and analysis process is to validate the selected weld schedule by performing a 3- dimensional thermal analysis. To do this the user must choose the appropriate material model, and must specify the weld process schedule and setup conditions. For different weld processes, process specific heat flux models are applied. These models are typically represented as either heating the surface under the arc or beam or heating the whole volume of metal under the arc or beam. For both material and process models, the most likely model is presented as a default, but the user must confirm the choice and, if knowledgeable, can override it. When these conditions are satisfied, the thermal analysis is performed automatically. All of the input to the thermal analysis code is modified on-the-fly based on the user's decisions. The thermal analysis code resides on a several computers on our network so each computer bids and the one with the estimated fastest turnaround is given the job. We are using JACQ, an internally developed Sandia code for thermal analysis. JACQ simulates the evolution of the temperatures generated by the welding process over time, including the addition of filler material in the weld. The final step in the weld design and analysis process is visualization, as illustrated in Figure 6 . The output of the thermal analysis is automatically input into another Sandia code, FEAVR (Finite Element Analysis Viewer) for automatic creation of a movie of the evolving temperature distribution in the part. FEAVR is written in the Advanced Visual Systems (AVSm) code. The user can also pick a node in the mesh and watch the temperature for that specific point evolve over time. This is important if there is a specific thermally sensitive area that is of concern. Another component of Smartweld is random, direct access to information and tools. Expert users want to exploit the structure of the problem, but do not want to be locked in to that structure. They need random access to all of their tools and information sources. Access to the World-Wide-Web, the Weld Advisor, the Weld Schedule Database, and an HTML Weld Design Guide is built into the system. Pro/Engineerm, AVSm, P A T W , X P , JACQ, Matlabm, video-teleconferencing, and application sharing are currently available from the Tool Bar. PAGE: 1 MASTER RECORD REPORT DUPCHECK-ID-NUMBER: 96001607265 PERMANENT-DOCUMENT-NUMBER: M96011842 000 <013> DATE-COMPLETED 960803 <014> DATE-OF-RECORD-ENTRY 960701 <015> DATE-RECEIVED 960701 <016> COPIES-RECEIVED 2 <020> DOCUMENT-TYPE R <022> MEDIUM-CODE H <030> CLASSIFICATION-CODE Uncl <040> LITERARY-INDICATOR K <050> GPO-SUPERINTENDENT-OF-DOCUMENTS Y <072> PERSONAL-AUTHOR-AND-AFFILIATION Mitchiner, J.L.; Kleban, S.D.; Hess, B.V.; Mahin, K.W. [Sandia National Labs., Albuquerque, NM (United States)]; Messink, D. [IntelliCorp, Mountain View, CA (United States) 1 <080> SPONSORING-ORGANIZATION-CODE DOE/DP <110> TITLE-ENGLISH Smartweld: A knowledge-based approach to welding <150> PRIMARY-REPORT-NUMBER SAND--96-1372C <210> SECONDARY-REPORT-NUMBER CONF-9606211--2 <240> CONTRACT-NUMBER-DOE AC04-94AL85000 <241> ABBREV-CONTRACT-NUMBER-DOE AL8 5 00 0 <242> AWARDING-OFFICE-CODE 04 <243> BUDGET-REPORTING-CODE GBO 103 0 12 <245> LEGIBILITY-CODE 0 <246> DOE-INITIATING-OFFICE-CODE AL <247> MICROFICHE-DISTRIBUTION-CODE 4 <248> VENDOR-ID-CODE 206460-0001-2 <249> VENDOR-NAME SANDIA CORP <251> REPORTING-REQUIREMENT DD <276> DUPCHECK-BYPASS-FLAG N <291> PACKED-PRIMARY-REPORT-NUMBER SAND9 6 13 72C <293> PREFIX DE ~ 2 9 5 ~ INDEX-DOCUMENT-NUMBER M96011842 <370> PUBLICATION-DATE [1996] 4376s REPORT-TYPE-CODE-AND-FREQUENCY C/O6 / 96 <390> PAGES-BIBLIOGRAPHIC 6 <400> REPORT-DISTRIBUTION-CODE A <421> LANGUAGE-CODE EN <425> AUDIENCE-CODE 01 <426> LIMITATION-CODE UNL <449> CONFERENCE-NUMBER 9606211 <510> DISTRIBUTION-CATEGORY M -700 <520> PROJECT-STATUS P <530> ANNOUNCEMENT-CODE EDB;ERA;ETD;NTS < 5 4 0 > EDB-SUBJECT-CATEGORIES 360101;320303 <550> SOURCE-OF-BIBLIOGRAPHIC-INPUT IMS <560> COUNTRY-OF-INTELLECTUAL-ORIGIN us <570> COUNTRY-OF-PUBLICATION us <686> DOCUMENT-STATUS-CODE 000 ~7002 CORPORATE-CODE 9511100 <748> TAPE-VOL-ISSUE 96R15 <749> TAPE-INCOMING-SERIAL-NUMBER AHC2 96 15% % 10 9 <801> SUBJECT-DESCRIPTORS COMPUTER-AIDED MANUFACTURING:Ql,T2;EXPERT SYSTEMS:Q2;COMPUTER GRAPHICS; WELD1NG;WASTE MANAGEMENT;MATHEMATICAL M0DELS;WELDED JO1NTS:Tl; COMPUTER-AIDED DES1GN;KNOWLEDGE BASE;MESH GENERATION MASTER RECORD REPORT PAGE: 2 DUPCHECK-ID-NUMBER: 96001607265 PERMANENT-DOCUMENT-NER: M96011842 000 <931> AVAILABILITY-CODE 0s;NT <950> ABSTRACT Smartweld is a concurrent engineering system that integrates product design and processing decisions within an electronic desktop engineering environment. It is being developed to provide designers, process engineers, researchers and manufacturing technologists with transparent access to the right process information, process models, process experience and process experts, to realizel'right the first time" manufacturing. Empirical understanding along with process models are synthesized within a knowledge-based system to identify robust fabrication procedures based on cost, schedule, and performance. Integration of process simulation tools with design tools enables the designer to assess a number of design and process options on the computer rather than on the manufacturing floor. Task models and generic process models are being embedded within user friendly GUI's to more readily enable the customer to use the Smartweld system and its software tool set without extensive training. The integrated system architecture under development provides interactive communications and shared application capabilities across a variety of workstation and PC-type platforms either locally or at remote sites. . ' e . Y DISCLAIMER This report was prepared as an accouht of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsi- bility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Refer- ence herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recom- mendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. 
work_lizh6fmihnd4venudgttrcwqmi	----	challenges-08-00001.pdf challenges Article Expert System for Bomb Factory Detection by Networks of Advance Sensors Carlotta Ferrari 1 , Alessandro Ulrici 2 and Francesco Saverio Romolo 1,3, * 1 Institut de Police Scientifique (IPS), Université de Lausanne, Dorigny, 1004 Lausanne, Switzerland; carlotta.ferrari23@libero.it 2 Department of Life Sciences, University of Modena and Reggio Emilia, Via Amendola, 2, 42122 Reggio Emilia, Italy; alessandro.ulrici@unimore.it 3 Legal Medicine Section, Department Saimlal, Sapienza Università di Roma, Viale Regina Elena, 336, 00161 Rome, Italy * Correspondence: francescosaverio.romolo@uniroma1.it; Tel.: +39-06-4991-2581 Academic Editor: Palmiro Poltronieri Received: 31 October 2016; Accepted: 26 December 2016; Published: 3 January 2017 Abstract: (1) Background: Police forces and security administrations are nowadays considering Improvised explosives (IEs) as a major threat. The chemical substances used to prepare IEs are called precursors, and their presence could allow police forces to locate a bomb factory where the on-going manufacturing of IEs is carried out. (2) Methods: An expert system was developed and tested in handling signals from a network of sensors, allowing an early warning. The expert system allows the detection of one precursor based on the signal provided by a single sensor, the detection of one precursor based on the signal provided by more than one sensor, and the production of a global alarm level based on data fusion from all the sensors of the network. (3) Results: The expert system was tested in the Italian Air Force base of Pratica di Mare (Italy) and in the Swedish Defence Research Agency (FOI) in Grindsjön (Sweden). (4) Conclusion: The performance of the expert system was successfully evaluated under relevant environmental conditions. The approach used in the development of the expert system allows maximum flexibility in terms of integration of the response provided by any sensor, allowing to easily include in the network all possible new sensors. Keywords: improvised explosive; precursor; network; chemometrics; security; forensic science 1. Introduction The behaviour of terrorists using explosives changed in last decade of bombings: improvised explosives (IEs) produced with chemical substances available on the market substituted commercial and military explosives [1–4]. The chemical substances used in home-made preparations of IEs are called precursors. The European Parliament and the Council adopted the Regulation (EU) No. 98/2013 on the marketing and use of explosives precursors on 15 January 2014 [5]. According to this regulation, seven precursors shall not be available to the general public anymore in concentrations greater than their limit values listed in Table 1 and other precursors, listed in Table 2, will be monitored to report suspicious transactions when purchased by the public. The traces of precursors used in IEs production (particulates, vapours and/or waterborne) present in the environment surrounding the vicinity of a “bomb factory” could allow police forces to locate sites, where the on-going manufacturing of IEs is suspected. This approach for protecting citizens from bombings is expected to be more effective than simply patrolling a possible target, because the production time of IEs is much longer than the time needed to transport an improvised explosive device (IED) close to the target from the manufacturing site [6]. Between 2009 and 2011, the LOTUS project developed sensors to detect precursors of both drugs of abuse and IEs in the Challenges 2017, 8, 1; doi:10.3390/challe8010001 www.mdpi.com/journal/challenges Challenges 2017, 8, 1 2 of 18 environment. Another project, called “Bomb factory detection by Networks of Advanced Sensors” (BONAS) [7], recently studied a network of wireless sensors to detect precursors of IEs outside explosives manufacturing sites. Results from the BONAS project about the analysis of the precursor hydrogen peroxide have been recently published [8]. The different sensors of the BONAS project were specifically designed to be deployed in sensitive locations and easily camouflaged. This network allows an early threat alarm thanks to an expert system, which is a system allowing the detection of one precursor based on the signal provided by a single sensor, the detection of one precursor based on the signal provided by more than one sensor, and the production of a global alarm level based on data fusion from all the sensors of the network. The aim of the present work is to show how the expert system was successfully developed and tested to play its key role in protecting the security of citizens. Table 1. Precursors listed in the Annex I of the European Parliament and the Council adopted the Regulation (EU) No. 98/2013 on the marketing and use of explosives precursors on 15 January 2014 [5], which are available to the public only in limited concentration and in Annex II the eight substances that are required to be reported as suspicious transactions when purchased by the public. ANNEX I Regulated Substance Limit Value (w/w) 1 Hydrogen peroxide 12% 2 Nitromethane 30% 3 Nitric acid 3% 4 Potassium chlorate 40% 5 Potassium perchlorate 40% 6 Sodium chlorate 40% 7 Sodium perchlorate 40% Table 2. Precursors listed in the Annex II of the European Parliament and the Council adopted the Regulation (EU) No. 98/2013 on the marketing and use of explosives precursors on 15 January 2014 [5], which are required to be reported as suspicious transactions when purchased by the public. ANNEX II Regulated Substance 1 Hexamine 2 Sulphuric acid 3 Acetone 4 Potassium nitrate 5 Sodium nitrate 6 Calcium nitrate 7 Calcium ammonium nitrate 8 Ammonium nitrate (in concentration of 16 % by weight of nitrogen in relation to ammonium nitrate or higher) 2. Expert System for Advanced Sensors Network Data Analysis Sensors Network Expert System Overview The BONAS expert system was aimed to perform pattern recognition of the data collected by each network sensor, estimating the predicted probability of detection for each target. The system was then designed to integrate the responses from each sensor to provide the end-user with a global alarm level, summarizing information from the entire network into a unique evaluation of the possible criminal threat of an IE production site. Sensors able to detect explosives were considered for inclusion in the network too. The expert system follows a three step workflow while providing a real time response for any new measurement: 1. The first step is estimating the predicted probability of the presence of each target for each sensor included in the BONAS network. The classification models adopted are computed by means of Challenges 2017, 8, 1 3 of 18 a pattern recognition technique on a dataset of experimental data acquired from laboratory and field conditions. These classification models are then applied to any new data received to obtain the corresponding predicted probability value. 2. The expert system integrates the output given from multiple sensors measuring the same target to provide the user with a unique alarm value for each compound. In this context, sensor location is also considered and only output from sensors located within a user-defined distance range are integrated, giving a positive detection within a reasonable time delay. 3. At the last step, the output from the previous steps is further integrated to provide the user with a defined global alarm. This alarm can be triggered by the detection of a single explosive compound, such as trinitrotoluene (TNT) or RDX, because these substances are generally forbidden to the general public. When precursors are detected, the alarm value is triggered when specific couples of targets defined by the user are detected from sensors located within a user-defined distance range and giving a positive detection within a reasonable time delay. A compact representation of the information provided by the expert system is shown in the dedicated user interface developed in collaboration with TEKEVER [7]. A schematic representation of the three step approach developed for the BONAS expert system is shown in Figure 1. Figure 1. Schematic representation of the three-step approach developed for the BONAS expert system. The alarms due to individual targets are provided as output of Step 2 while the global alarm level is obtained as output of Step 3. Challenges 2017, 8, 1 4 of 18 3. Materials and Methods 3.1. Advanced Sensors Network and IE Precursors The BONAS project developed a network of advanced sensors able to detect traces of precursors used in IEs’ production present in the environment surrounding the vicinity of a “bomb factory”. Sensors based on a wide range of analytical methods were selected in order to enable the detection of precursors in different forms, i.e., particulates, vapours and/or waterborne. The sensors developed and tested throughout the project include a quartz-enhanced photo-acoustic spectroscopy (QEPAS) sensor, an electrochemical (EC) sensor, a light detection and ranging (LIDAR)/differential absorption LIDAR detection system (DIAL) sensor, and a surface-enhanced Raman spectroscopy (SERS) sensor. The substances considered in the project were selected among the IE precursors taken into account by the Regulation (EU) No. 98/2013 on the marketing and use of explosive precursors. Overall, classification models for the classification of five precursors with different forms were computed. These precursors are listed in Table 3 as codes, because their names are Confidential EU. Table 3. Precursors considered as targets of the different sensors included in the network. Sensor Precursor Code Precursor Form QEPAS B02 Vapor B10 Vapor EC B01 Waterborne B08 Waterborne B15 Waterborne LIDAR B02 Vapor B10 Vapor SERS B15 Particulate Three key requirements were identified in the development of the present expert system: (i) minimization of false positive detections at sensor level (described in Section 3.2); (ii) flexibility of the system in terms of type and number of sensors included at any time in the network (described in Section 3.3); (iii) flexibility of the system in terms of choice of the parameters that determine the global alarm activation (described in Section 3.4), in order to allow the user to easily customize the system according to the condition-specific needs. All data analysis were performed using the PLS Toolbox ver. 7.5 and ver. 7.8.2 (Eigenvector Research Inc., Manson, WA, USA) and all routines were written in Matlab© platform 7.11 R2010b (The Mathworks Inc., Natick, MA, USA). 3.2. STEP 1: Supervised Pattern Recognition at Sensor Level At the first step, the expert system is requested to estimate the predicted probability of presence of each target for each sensor included in the sensor network. To this aim, different classification rules were chosen, based on the nature of the sensor data. In particular, for QEPAS and EC, Partial Least Squares-Discriminant Analysis (PLS-DA) [9] models were developed using experimental data acquired from laboratory and field conditions. The PLS-DA classification models were validated by means of an external test set, and their performance was evaluated on the basis of the following parameters: • sensitivity (SENS): the percentage of objects of each modelled class correctly accepted by the class model; • specificity (SPEC): the percentage of objects of the other classes correctly rejected by the class model; • efficiency (EFF): the geometric mean of sensitivity and specificity. Challenges 2017, 8, 1 5 of 18 Once the optimal model for each target substance has been selected, for any new measurement the probability of the presence of the different target substances can be computed and stored. An alarm is then triggered if this value exceeds the corresponding probability threshold. In order to better exploit the collected analytical information, two different thresholds are considered: • a lower threshold (maxeff_thresh): probability threshold which provides the best compromise between false positive and false negative results; • a higher threshold (maxspec_thresh): probability threshold which allows to minimize false positive results. Following the double threshold approach, two different alarm levels are considered: a lower alarm level is provided when the predicted probability exceeds the value of maxeff_thresh but is lower than maxspec_thresh, because it means that the predicted probability value is below the threshold that minimizes the false positive results, while a higher level alarm is given when the predicted probability exceeds the value of maxspec_thresh as well. In the case of the SERS sensor, a simpler classification rule was used, based on the correlation coefficient between the signal measured on the sample and the corresponding signal measured on the pure target. Concerning the LIDAR sensor, it provided a concentration value of the precursor B10 [7]. In addition to the display of the detected target names and of the obtained predicted probability values, a simple colour-based code has been implemented to summarize this information in the expert system human interface and facilitate the interpretation of results. In particular, a green bar is used to indicate that no detection has occurred while a blue or yellow bar is displayed when a lower or higher alarm has been triggered, respectively. A numeric value of 1, 2 and 3 has been assigned to green, blue and yellow colour codes, respectively (Table 4). Table 4. Interpretation of the colour code used for the output of the expert system after the Step 1. Colour Code Interpretation Associated Value Grey The target substance is not monitored by any of the sensors included at that moment in the network. 0 Green The target substance can be detected by the network but is not detect at that moment. 1 Blue The target substance is detected by the sensor at the lower threshold. 2 Yellow The target substance is detected by the sensor at the higher threshold. 3 A schematic overview of the approach used in the first step of the expert system is provided in Figure 2. Figure 2. Scheme of the expert system step for pattern recognition at sensor level. The colour code obtained by each sensor for each target substance at this step is stored in a matrix, which is passed to the function of the second step of the expert system. In this way, this information can then be integrated with that of the other sensors of the network and a unique alarm level for each target substance can be therefore provided to the end-user. Challenges 2017, 8, 1 6 of 18 3.2.1. QEPAS Sensor For the BONAS project, the CREO team developed a QEPAS sensor allowing IR analysis of vapour [10]. At the first step, a database of experimental signals was created in order to calculate the PLS-DA classification model at the basis of the developed expert system. In order to obtain a representative spectra database, signals were acquired both in laboratory and on field conditions. In particular, a laboratory dataset was created in laboratory by acquiring spectra of four target substances, i.e., B01, B02, B03 and B10 and four interferents, i.e., Int02, Int05, Int10 and Int11 at two different concentration levels have been acquired. In order to evaluate the reproducibility of the system, at least two samples (replicates) of each substance at each concentration were considered. The laboratory spectra database was initially investigated by means of Principal Component Analysis (PCA) considering several data pre-treatments (data not shown for conciseness reasons). The results of this explorative analysis suggested focusing on two main target molecules, i.e., B02 e B10. Considering that the QEPAS sensor is able to analyze vapors, the database was then updated with data acquired during on-field tests carried out according to realistic IE preparation scenarios. During this test campaign, two target substances included in the priority list for BONAS and one interferent were tested, i.e., target B02, target B10 and Int10. The composition of the updated datasets used for classification model computation of both target B02 and target B10 are reported in Table 5. Table 5. Composition of the QEPAS dataset for target B02 and target B10 classification model computation. Target Class # Spectra Training Set # Spectra Test Set Total Number of Spectra B02 Class TARGET B02 139 90 229 Class NO TARGET B02 340 363 703 B10 Class TARGET B10 94 59 153 Class NO TARGET B10 385 395 780 Subsequently, these datasets have been used to compute the classification models for the two targets and the selected classification model was then used to analyze the data acquired during on field tests carried out in realistic scenarios. 3.2.2. EC Sensor For the BONAS project, the UCBL team [7] developed an electrochemical sensor dedicated to the achievement of simultaneous monitoring of different explosive precursors. The BONAS EC sensor uses electrodes to oxidize or reduce molecules soluble in water using portable electrodes, easily hidden in the sewage system, operating stand-alone by battery, with wireless communication capability. The electrochemical chips used for the acquisition of the database signals were of one counter electrode, one pseudo-reference electrode and eight working electrodes. These electrodes were divided into two series of four electrodes each; in each series, three electrodes were modified with different electrodeposited metals (gold, palladium and platinum), resulting in four voltammograms per analysis. In addition to the analysis of the individual signals of the different electrode surfaces, another approach based on the creation of a unique signature for each sample was applied so as to take full advantage of the information provided by the different electrodes on the same sample. To this aim, the signals acquired with the four electrode surfaces were merged together as a sequence and each block of data then was scaled to unit variance, in order to assign equal importance to information provided by the different electrodes. A schematic representation of this approach is shown in Figure 3. Challenges 2017, 8, 1 7 of 18 Figure 3. Electrochemical data analysis approach. In addition to the three main target substances analysed by this sensor, i.e., B01, B08 and B15, data were acquired also on additional target substances, i.e., B04, B05 and B11, and on the interferent Int09, selected as components of common cleaning products. The measurements were realized in: (i) NaCl 0.1 M electrolyte solution; (ii) tap drinking water; (iii) tap non-drinkable water (iv) soap water (GEH dish washing soap 0.1% v/v in tap drinking water) and (v) artificial sewage water prepared, using tap drinking water. The composition of the dataset used for the calculation of the PLS-DA classification models of target B01, target B08 and target B15 is reported in Table 6. Table 6. Composition of the training set and test set used to compute and validate the classification models of the EC sensor. Target Class # Spectra Training Set/Electrode # Spectra Test Set/Electrode Total Number of Spectra/Electrode B01 Class TARGET B01 25 11 36 Class NO TARGET B01 187 122 309 B08 Class TARGET B08 33 22 55 Class NO TARGET B08 179 111 290 B15 Class TARGET B15 33 21 54 Class NO TARGET B15 179 112 291 Challenges 2017, 8, 1 8 of 18 Similarly to the QEPAS sensor, also in this case these datasets have been used for the computation of classification models to be applied on the data acquired during on field tests carried out in realistic scenarios. 3.2.3. SERS Sensor For the BONAS project, the Serstech team [7] developed a SERS based sensor, used for detecting particles and/or vapours in the air surrounding a potential IED factory. Briefly, the sensor system consists of a miniaturized Raman spectrometer in combination with a sampling system for collecting particles and/or vapours from the surrounding air. The sampling system collects air through an inlet fan and directs this to a cooled SERS-surface were particles are trapped onto the surface. The SERS-substrate is then dried by heating and placed in position for measurement by the Raman spectrometer. The resulting spectral information is transferred to the command centre for data analysis. In order to test the possibility to successfully integrate this sensor in the expert system, a very preliminary data analysis approach was applied during the on field test campaign. The approach was based on the calculation of the linear correlation coefficient (R) between any new acquired spectrum and a reference spectrum, after having applied a pre-processing step by means of linear detrend in order to remove baseline shifts. Furthermore, on the basis of a meeting with the specialists in charge of the sensor, only the data in spectral range between 812 and 880 nm have been considered for data analysis. For alarm triggering, it was set a criteria of correlation coefficient above a threshold of 0.8 with a significance level of 0.05. 3.2.4. Lidar/DIAL Remote Sensor For the BONAS project, the ENEA [7] and CSM teams [7] developed a lidar/DIAL remote sensor. As far it is concerned, no pattern recognition analysis has been required, since the sensor output is not provided as raw data or signals but directly as target B10 concentration values. In this case, therefore, the expert system just compares the concentration value in output with a concentration threshold defined experimentally that can be easily modified by the user. This information is then integrated with the response of the other sensors during the second step of the expert system in order to trigger an alarm. In order to take into account the limited selectivity of the sensor response, at this step the concentration values obtained for target B10 were also converted to equivalent concentration values of the other target substance the sensor can detect, i.e., target B02 . This is achieved by applying a correction factor defined according of the optical absorption characteristic of target B02. Therefore, whenever lidar/DIAL remote sensor has a positive detection, alarms are triggered for all the target substances it can detect (i.e., B10 and B02). However, this aspect is considered at the second step of the expert system and, when other sensors confirm the detection of only one of the lidar/DIAL target substances, the alarms triggered for the other one is set back to “no detection”. 3.3. STEP 2: High-Level Data Fusion for Single Target Substance Detection At the second step, the expert system is requested to integrate the output given from multiple sensors measuring the same target to provide the user with a unique alarm value for each compound. Over the years, a considerable number of data fusion methods have been developed, which can be roughly subdivided into three different levels as follows [11]: • low-level fusion, where the raw signals provided by the different sensors are combined before performing any data pre-processing; • mid-level fusion, which is based on the combination of features extracted from the data of each sensor; • high-level fusion, where the response of the different sensors (e.g., detection decision) are combined to obtain a unique response. Challenges 2017, 8, 1 9 of 18 Considering the need of having a flexible expert system able to include any new sensor as well as to consider the possibility of temporary inactivity (e.g., battery drained waiting for replacement) of some sensors, the high level data fusion approach has been selected to be used in the development of the expert system. In particular, high level data fusion was used at this stage to enable the expert system to integrate the information about the alarms triggered for the same target substance by all the sensors which are included at that moment in the network and which are able to detect it. To this aim, a dedicated routine was developed, which runs over all the target substances the sensors network is monitoring at that moment and it checks if any alarm has been triggered for each of them. If this is not the case, nothing happens and the green colour bar is displayed in the user interface. Similarly, when only one sensor is detecting that target substance, the same alarm level and the same color code given by that sensor is kept. Whenever more than one sensor detects the same target substance, the information about their relative distance is first considered. For this reason, the distances between the positions of all those sensors are computed and compared with the defined threshold of maximum distance. For the stand-off sensors, the coordinates of the actual sampling volume are computed and considered in the calculation of their relative distance to the other sensors. At this point, if all the sensors are closer to each other than the distance threshold, the final alarm level is calculated as a weighted sum of the alarm levels of the individual sensors. The alarm of each sensor is thus weighted according to the efficiency value of the model used at Step 1 for its definition so as to take into account also the ability of that model to correctly discriminate samples belonging or not to the target class. If this value is greater than 4 (orange colour code) then the alarm is set to the maximum level, i.e., 5, corresponding to a red colour code. The colour codes used to represent the Step 2 output in the human interface are summarized in Table 7. Table 7. Interpretation of the colour code used to represent the Step 2 output. Colour Code Interpretation Associated Value Grey The target substance is not monitored by any of the sensors included at that moment in the network. 0 Green The target substance can be detected by the network but is not detect at that moment. 1 Blue The target substance is detected by the network at the lowest threshold by one sensor only. 2 Yellow The target substance is detected by one sensor only at the highest threshold or by two sensors at the lower threshold. 3 Orange The target substance is detected by at least two sensors at the .higher threshold 4 Red The target substance is detected by at least two sensors at the higher threshold. 5 It has to be underlined that the approach used in the development of the expert system allows maximum flexibility in terms of integration of the response provided by any sensor, even if not considered in the previous step of the expert system. This characteristic is particularly important since it allows to easily include in the network any possible new sensor by just communicating to the expert system the information about the code of the detected target substance, the concentration/probability measured, the concentration/probability threshold and its location. The flexibility of the expert system in easily integrating the response of sensors not considered at Step 1 has been successfully tested in the integration of the lidar/DIAL remote sensor response. 3.4. STEP 3: Data Fusion for Multiple Target Substances Detection At the last step, the expert system further integrates the outputs of Step 1 and Step 2 to provide the user with a unique evaluation of the possible criminal threat of an IE production site, defined Challenges 2017, 8, 1 10 of 18 taking into account the information obtained from the entire network of sensors. The global alarm level provided as output of this third step, in fact, considers sensor location, the total number of detected targets as well as the detection of specific targets known to be used together to prepare a specific IE in a reasonably short time window. In this context, particular attention has been paid to the development of a system characterized by maximum flexibility in terms of choice of the parameters that determine the alarm activation, in order to allow the user to easily customize the system according to the condition-specific needs. This means that the end user can easily modify the conditions that define an alarm triggering event as well as the global alarm threshold used to indicate the presence of a possible threat. In the set up adopted, the global alarm ranges from 1 to 100 and the alarm threshold was set to 50. In case explosive compounds are being monitored by the sensors network, the detection of just one of them is considered as an alarm triggering event and the global alarm score is increased above the threshold according to its probability of detection. When only precursors are monitored, this threshold can be exceeded only in case specific couples of target substances are detected, while otherwise the alarm level is lower or equal to 50. The reason of this approach is to avoid an excessive number of false positives, because precursors can be used in legal activities. The first part of the score (from 1 to 50) is in fact just a weighed sum of the information provided by Step 2, where the weight to be given to each detected target substance is calculated according to the total number of target substances monitored at that time. The second part of the score (from 51 to 100) is instead defined according to the number of specific couples of target substances, which are being detected, and to their probability level of detection. The names of all the couples of target substances are displayed when detected together. We stress once again the point that the system has maximum flexibility and that therefore the final user, according to the specific needs, can easily implement different choices about the parameters, which define the alarm activation. The user can thus easily modify the list of individual targets whose detection determine the alarm initiation in order to include, for example, not only explosive compounds but also precursors of particular interest. Similarly, also the list of couples of targets substances can be updated without requiring any modification of the routine code. As mentioned above and similarly to what described in Step 2 (see Section 3), also in this context the information about the relative distance between sensors is taken into account before combining the information about couples of target substances, i.e., the alarm is raised above the threshold only if the sensors which have detected the couple of target substances are located within the distance threshold defined by the user. 4. Discussion The discussion of the results is organized in a first section (4.1), where we report about applying the PLS-DA classification models developed for the QEPAS and EC sensors to the relevant test set signals in laboratory conditions (Step 1), and a second section (4.2) where we discuss the results of the final validation in real scenario both at single sensor level (Step 1) and by the whole network of sensors (Step 2 and Step 3). 4.1. Validation of PLS-DA Classification Models in Laboratory Conditions With regards to QEPAS sensor, efficiency values in prediction equal to 98.13% and 100.00% were obtained for target B10 and target B02, respectively (see Table 8). Table 8. PLS-DA classification results obtained in prediction on QEPAS spectra acquired for models validation in laboratory conditions. B10 B02 SENS SPEC EFF SENS SPEC EFF 100.00 96.30 98.13 100.00 100.00 100.00 Challenges 2017, 8, 1 11 of 18 The PLS-DA classification model for the EC sensor has been calculated for the three main targets, i.e., B01, B08 and B15, and the results obtained for the prediction of the test set samples are reported in Table 9. Table 9. PLS-DA classification results obtained in prediction on the EC signals acquired for models validation in laboratory conditions. B01 B08 B15 SENS SPEC EFF SENS SPEC EFF SENS SPEC EFF 90.90 100.00 95.34 100.00 99.10 99.55 100.00 100.00 100.00 4.2. Expert System Validation in a Real Case Scenario A final validation of the whole expert system structure was carried out during two demos of the project, one at the Italian Air Force base of Pratica di Mare (Italy) and one at the Swedish Defence Research Agency (FOI) in Grindsjön (Sweden). During this demo, real case scenario tests were in fact performed in order to verify the detection abilities of the individual sensors as well as the expert system ability to successfully integrate their response. No details related to test protocols followed during the demo are described as they are Confidential EU. 4.2.1. Validation of Pattern Recognition Models at Sensor Level Considering that these tests were carried out on field and in real case scenario, target concentrations were not constant during the whole test due to several factors such as weather conditions. It can be however noticed that positive detections were obtained during all the tests while no false positive detections occurred (Tables 10 and 11). As for the QEPAS sensor, three tests were performed for each of the two target substances, i.e., target B10 and target B02. During these trials, the QEPAS sensor was placed inside a dumpster used for camouflage and used to analyse vapours of IE precursors generated during an IE production (or the simulation of an IE production). The results are reported in terms of alarm colour codes triggered after Step 1 (please refer to Table 4 for colour code interpretation). Table 10. Results of the three tests performed at FOI for target B10. # Test # Measurement Cycle Target B10 Target B02 Test 1 1 3 1 2 1 1 3 1 1 4 2 1 5 1 1 Test 2 1 3 1 2 3 1 3 3 1 4 3 1 Test 3 1 3 1 2 3 1 3 3 1 4 3 1 5 1 1 As for the QEPAS sensor, also the capability of the electrochemical sensor to detect IE precursors directly in sewage water drains from simulated bomb factories has been performed during the final demonstration in Sweden. Three different scenarios have been tested in this context, each of them involving the use of one the three main targets at one step of an IE preparation. It has to be underlined Challenges 2017, 8, 1 12 of 18 that all the experiments performed during this test campaign were carried out using local tap water and a sink system found in loco (Figure 4). Table 11. Results of the three tests performed at FOI for target B02. # Test # Measurement Target B10 Target B02 Test 4 1 1 3 2 1 3 3 1 3 4 1 3 5 1 3 Test 5 1 1 1 2 1 3 3 1 3 4 1 3 5 1 3 Test 6 1 1 3 2 1 1 3 1 3 4 1 3 5 1 3 6 1 1 Figure 4. EC sensor experimental set-up used during the final demonstration carried out at FOI (Grindsjön). The results obtained by applying the final classification models (see Section 4.1) to the measurements acquired during the experiments performed in the final demonstration are shown in Table 12. It can be noticed that true positive alarms were obtained for the target substances during all the tests. However, in some of the tests carried out on target B01, false positive alarms were obtained also for target B08. In order to investigate the reason for this anomalous behaviour not observed during the laboratory test, EC signals acquired during the tests on target B01 were compared to the average signals previously acquired in tap water on target B01 and target B08. It was observed that, despite the signals acquired during the final demonstration showed most of the typical features of the average B01 signals, the intensity registered especially using the Pd and Pt electrodes was lower than usual and Challenges 2017, 8, 1 13 of 18 more similar to the intensity registered in the presence of target B08. A possible explanation for this behaviour could be the degradation of the solution used during the tests due to the high instability of target B01. Furthermore, it has to be underlined that a more limited number of signals acquired in a lower number of experimental conditions was available for this target substance. For these reasons, a further training of the expert system with a larger number of samples, acquired, e.g., at lower concentrations, could allow for significant improvement of the expert system performance. Table 12. Results obtained for the EC sensor during the tests performed during the final demonstration at FOI. # Test Target Alarm Colour Code B01 B08 B15 Test 1 B15 1 1 3 Test 2 B01 3 1 1 Test 3 B08 1 3 1 Test 4 B01 3 3 1 Test 5 B15 1 1 3 Test 6 B01 3 3 1 Test 7 B15 1 1 3 Test 8 Blank 1 1 1 Test 9 Blank 1 1 1 Test 10 B15 1 1 3 Test 11 B08 1 3 1 Test 12 Blank 1 1 1 Test 13 B15 1 1 3 Test 14 B01 3 3 1 Test 15 Blank 1 1 1 Test 16 B08 1 3 1 Test 17 Blank 1 1 1 Test 18 B08 1 3 1 Test 19 B08 1 3 1 Test 20 Blank 1 1 1 As mentioned in Section 3.2, in the case of the SERS sensor, no pattern recognition was applied and the possibility to successfully integrate the sensor in the network was evaluated by triggering the alarm on the basis of the correlation coefficient value calculated between the signal measured on the sample and the corresponding signal measured on the pure target. During the final demo, four tests were performed by applying a real case scenario involving the use of Target B15. The spectra obtained during these tests are reported in Figure 5, where it can be observed that positive detections by the Raman spectrometer were obtained in all experiments. Despite that positive detections were observed in all experiments, the initiated alarms obtained from the analysis of the pretreated data (Table 13) show that false negative detections were also observed. In particular, the false alarm obtained for the Test 4 spectrum acquired was likely due to a slight misalignment of the main absorption peak. The application of different data pretreatment methods as well as of different data analysis approaches is expected to greatly help in solving this issue. In conclusion, the tests performed during the demo at FOI confirmed the capability of the SERS sensor in sampling particles from the atmosphere and successfully acquiring spectra on the collected analytical data. The promising results obtained suggest that the acquisition of a more representative database of experimental data could allow developing a more reliable and robust data analysis step as well as to extend the list of target substances for this sensor. Challenges 2017, 8, 1 14 of 18 Figure 5. Raman spectra pretreated by means of linear detrend. Table 13. Results of the analysis of the spectra pretreated by linear detrend. # Test Target Alarm Color Code Test 1 B15 2 Test 2 B15 1 Test 5 B15 1 Test 6 B15 2 Blank B15 1 4.2.2. Validation of Expert System Data Fusion at Sensors Network Level In addition to validating the classification models developed for the individual sensors, an extensive validation of the capability of the expert system to effectively integrate the information provided by the different sensors has also been carried out during the two test campaigns at the Italian Air Force base of Pratica di Mare and in Sweden at FOI. During both test campaigns, the command and control (C2) centre has shown its ability to communicate with all the sensors included in the network and to process, by means of the expert system, the data received. The live alarm monitor end-user application developed by Tekever allowed a real-time monitoring of the information provided by the expert system. In fact, the user interface reports: • on the left, the alarm obtained for each target substance considering the information of all the sensors able to monitor it (Step 2 output); • in the middle, the global alarm which summarizes the information from the entire network into a unique evaluation of the possible criminal threat of an IE production site (Step 3 output); • on the right, a map showing the location of the different sensors is also reported. To indicate the sensors’ positions, markers coloured according to the colour code of Step 2 are used in order to show a possible positive detection of the sensor; • a circle with a radius equal to the value maximum distance threshold for data integration is also reported for each sensor. In this way, the interface allows to monitor the data based on which sensors are integrated. As an example, the user interface obtained during an experiment performed at FOI is shown in Figure 6. From the interface, it can be noticed that only one sensor is showing positive detections and that two targets are being detected. The location of the sensor corresponded to the lidar/DIAL remote sensor, while the two detected targets were target B02 and target B10. The double detection provided Challenges 2017, 8, 1 15 of 18 by the sensor is due to its limited selectivity. Since two target substances are being detected at the lower alarm level, a low global alarm level is provided as output of Step 3. Figure 6. User interface during an experiment performed at FOI, which shows the results for the two possible targets giving positive detection by Lidar. Another example of a user interface obtained during the test campaign at FOI is shown in Figure 7. In this case, in addition to the detection described above, also another sensor, i.e., the electrochemical sensor, showed a positive detection for target B08 with a probability above the threshold which minimizes false positives (colour code yellow, probability of 100%). However, since no couples of targets known to be used together in the preparation of IEs have been detected, the global alarm still shows a low value. Figure 7. Example of user interface obtained during the test campaign at FOI, which shows the positive detections for three targets obtained by two sensors. No couples of targets are known to be used together in the preparation of one IE. Challenges 2017, 8, 1 16 of 18 A third example of a user interface obtained during the test campaign at FOI is shown in Figure 8. It can be noticed that, as in the previous case, three target substances are being detected by two sensors. However, since target B02 and target B15 are one of the couples of targets known to be used together in the preparation of one IE, a much higher global alarm level is obtained in this case as output of Step 3. Figure 8. Example of a user interface obtained during the test campaign at FOI, which shows the positive detections for three targets obtained by two sensors. Since a couple of target detected is known to be used together for the preparation of one IE, the global alarm has increased above the threshold of 50. Finally, we want to emphasize that during the final demonstration at the Swedish Defence Research Agency, also the sensors developed in another project (EMPHASIS) [12] were able to communicate with the BONAS Central Command Unit (CCU) in order to have their data processed by means of the BONAS expert system. In particular, the positive detections of the EMPHASIS sensors were successfully integrated in the second step of the BONAS expert system and did participate to determine both the individual target substance alarms and the global alarm of the sensor network. This test campaign, therefore, confirmed that the approach used in the development of the BONAS expert system allows maximum flexibility in terms of integration of the response provided by any sensor, even if not considered at the beginning of the expert system development. It is important to note that, despite the use of IR and Raman sensors, any positive detection produced by the network is not yet a forensic identification. To obtain a forensic identification of a chemical substance (e.g., the IE TATP) a suitable number of laboratory analyses must be carried out [13,14]. The capability to integrate the response any possible new sensor is the basis to establish a large network supporting safer smart cities, where many existing sensors can be networked in a grid allowing new elements of new types to collect information to be provided to the security authority [15]. 5. Conclusions This paper describes the innovative expert system approach developed in the BONAS research project to exploit the information provided by the whole sensor network and to provide the end-user with a unique global evaluation of the possible criminal threat of an IE production site. Three key requirements have been identified in the development of the present expert system: (i) minimization of false positive detections at the sensor level (developed within Step 1); (ii) flexibility Challenges 2017, 8, 1 17 of 18 of the system in terms of type and number of sensors included at any time in the network (developed within Step 2); (iii) flexibility of the system in terms of choice of the parameters that determine the global alarm activation (developed within Step 3), in order to allow the user to easily customize the system according to the condition-specific needs. In order to meet these requirements, an expert system which follows a three-step workflow has been created. During the first step, the expert system estimates the predicted probability of presence of each target for each sensor included in the BONAS network. To this aim, classification models were computed by means of a PLS-DA algorithm with a dataset of experimental data acquired by means of QEPAS and EC sensors both in laboratory and in field conditions. The computed classification models were then applied to any new data received to obtain the corresponding predicted probability value. Two different probability thresholds were then considered for alarm activation, one which allows to maximize classification efficiency and one which allows to maximize classification specificity (i.e., to minimize false positive results). During the second step, the expert system integrates the output given from multiple sensors measuring the same target to provide the user with a unique alarm value for each compound. In this context, the application of a high-level data fusion approach allows to meet the requirement of having a flexible expert system able to include at this stage any new sensor (e.g., LIDAR sensor) as well as to consider the possibility of temporary inactivity of some sensors. During the last step, the expert system further integrates the outputs from the previous steps to provide the user with a unique global evaluation of the possible criminal threat of an IE production site. The global alarm level provided as output of this third step is, in fact, defined considering the sensors’ locations, the short time, if any, between the positive detections of the total number of detected targets as well as the detection of specific targets known to be used together to prepare a specific IE. At this stage, the end user can easily customize the system according to the condition-specific needs by modifying the parameters that determine the alarm activation (e.g., the list of individual targets or the list of couples of targets substances whose detection determine the alarm). Finally, a dedicated user interface has been developed in collaboration with TEKEVER (reference) to provide the user with a compact representation of the information given by the expert system. The performance of the developed sensor network was evaluated under relevant environmental conditions, i.e., in the presence of interferents and pollutants in both air and water, in the Italian Air Force base of Pratica di Mare and in Swedish Defence Research Agency facility in Grindsjön. An extensive validation of the capability of the expert system was carried out to effectively integrate the information provided by the different sensors showing that the end-users were effectively provided with a global alarm level, summarizing information from the entire network into a unique evaluation of the possible criminal threat of an IE production site. The expert system during the validation showed its capability to handle information not only from the BONAS sensors, but also from the sensors developed in another project, EMPHASIS. This flexibility in terms of integration of the response provided by any sensor allows easily including in the network any possible new sensor and establishing a large network supporting safer smart cities. Acknowledgments: The project BONAS was funded under the 7th Framework Programme of the European Commission, grant agreement No. 261685. We acknowledge Christophe A. Marquette, Directeur Adjoint of the Institut de Chimie et Biochimie Moléculaires et Supramoléculaires Equipe Génie Enzymatique, Membranes Biomimétiques et Assemblages Supramoléculaires (GEMBAS) of the University Lyon 1 (France) for the tests carried out with the electrochemical sensor, Pedro Santo Antonio from TEKEVER, Obidos (Portugal), for developing the software connecting the sensors of the BONAS network. Author Contributions: Francesco Saverio Romolo proposed the general approach for the expert system to handle the network of sensors (Step 1, Step 2 and Step 3), and contributed to the writing of this article. Carlotta Ferrari developed and validated chemometric models, and contributed to the writing of this article. Alessandro Ulrici contributed to the validation of the results of the expert system, and supervised the writing of this article. Challenges 2017, 8, 1 18 of 18 Conflicts of Interest: The authors declare no conflict of interest. The founding sponsors had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the decision to publish the results. References 1. Marshall, M.; Oxley, J.C. (Eds.) Aspects of Explosives Detection; Elsevier: Amsterdam, The Netherlands, 2009; p. 21. 2. Department of Justice, Office of Public Affairs. Umar Farouk Abdulmutallab Sentenced to Life in Prison for Attempted Bombing of Flight 253 on Christmas Day 2009. Available online: http://www.justice.gov/opa/ pr/2012/February/12-ag-227.html (accessed on 26 October 2016). 3. Transportation Security Administration. UK 2006 Liquid Explosives Plot Trial Overview. Available online: http://www.tsa.gov/press/releases/2008/09/08/uk-2006-liquid-explosives-plot-trial-overview (accessed on 26 October 2016). 4. Intelligence and Security Committee. Report into the London Terrorist Attacks on 7 July 2005. Available online: www.fas.org/irp/world/uk/isc_7july_report.pdf (accessed on 26 October 2016). 5. Regulation (EU) No. 98/2013 of the European Parliament and of the Council on the Marketing and Use of Explosives Precursors was Adopted on 15 January 2013. Available online: http://eur-lex.europa.eu/legal- content/EN/TXT/?uri=uriserv:OJ.L_.2013.039.01.0001.01.ENG (accessed on 26 October 2016). 6. Önnerud, H.; Wallin, S.; Östmark, H.; Menning, D.; Ek, S.; Ellis, H.; Kölhed, M. Localisation of threat substances in urban society—LOTUS: A viable tool for finding illegal bomb factories in cities. In Proceedings of the SPIE 8019, Sensors, and Command, Control, Communications, and Intelligence (C3I) Technologies for Homeland Security and Homeland Defense X, 80190Y, Orlando, FL, USA, 25 April 2011. [CrossRef] 7. Final Report Summary—BONAS (Bomb Factory Detection by Networks of Advanced Sensors). Available online: http://cordis.europa.eu/result/rcn/169207_en.html (accessed on 30 October 2016). 8. Romolo, F.S.; Connell, S.; Ferrari, C.; Suarez, G.; Sauvain, J.J.; Hopf, N.B. Locating bomb factories by detecting hydrogen peroxide. Talanta 2016, 160, 15–20. [CrossRef] [PubMed] 9. Barker, M.; Rayens, W. Partial least squares for discrimination. J. Chemom. 2003, 17, 166–173. [CrossRef] 10. Viola, R.; Liberatore, N.; Luciani, D.; Mengali, S. Quartz enhanced photoacoustic spectroscopy for detection of improvised explosive devices and precursors. Adv. Opt. Technol. 2016, 2016, 1–12. [CrossRef] 11. Kemp, M.C. A review of sensor data fusion for explosives and weapons detection. In Proceedings of the SPIE Defense, Security, and Sensing, International Society for Optics and Photonics, Baltimore, MD, USA, 29 April 2013. 12. Final Report Summary—EMPHASIS (Explosive Material Production (Hidden) Agile Search and Intelligence System). Available online: http://cordis.europa.eu/result/rcn/169001_en.html (accessed on 30 October 2016). 13. Romolo, F.S.; Cassioli, L.; Grossi, S.; Cinelli, G.; Russo, M.V. Surface-sampling and analysis of TATP by gas chromatography/mass spectrometry. Forensic Sci. Int. 2013, 224, 96–100. [CrossRef] [PubMed] 14. Sigman, M.E.; Clark, C.D.; Caiano, T.; Mullen, R. Analysis of triacetone triperoxide (TATP) and TATP synthetic intermediates by electrospray ionization mass spectrometry. Rapid Commun. Mass Spectrom. 2008, 22, 84–90. [CrossRef] [PubMed] 15. Connell, Toward a “Smart City”, IH and Security Professionals Collaborate on Early Detection of Bombs. Available online: http://synergist.aiha.org/201605-toward-a-smart-city (accessed on 30 October 2016). © 2017 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/). 
work_llenip2c6jb7zoiyterzi2w6mu	----	Rule-based expert systems to support step-by-step guidance in algebraic problem solving: The case of the tutor PAT2Math Rule-based expert systems to support step-by-step guidance in algebraic problem solving: The case of the tutor PAT2Math Patricia A. Jaques a,⇑, Henrique Seffrin a, Geiseane Rubi a, Felipe de Morais a, Cássio Ghilardi a, Ig Ibert Bittencourt b, Seiji Isotani c a PIPCA – UNISINOS, Av. Unisinos, 950 Bairro Cristo Rei, CEP 93.022-000 São Leopoldo, Brazil b NEES – IC – UFAL, Campus A.C. Simões, BR 104, Norte, km 97, Cidade Universitária, CEP 57072-970 Maceio, AL, Brazil c ICMC – University of São Paulo, Avenida Trabalhador São-carlense, 400 Centro, CEP 13566-590 Sao Carlos, SP, Brazil a r t i c l e i n f o Keywords: Intelligent tutoring systems Expert systems Math learning Algebra Equations a b s t r a c t In order for an Intelligent Tutoring System (ITS) to correct students’ exercises, it must know how to solve the same type of problems that students do and the related knowledge components. It can, thereby, com- pare the desirable solution with the student’s answer. This task can be accomplished by an expert system. However, it has some drawbacks, such as an exponential complexity time, which impairs the desirable real-time response. In this paper we describe the expert system (ES) module of an Algebra ITS, called PAT2Math. The ES is responsible for correcting student steps and modeling student knowledge compo- nents during equations problem solving. Another important function of this module is to demonstrate to students how to solve a problem. In this paper, we focus mainly on the implementation of this module as a rule-based expert system. We also describe how we reduced the complexity of this module from O(nd) to O(d), where n is the number of rules in the knowledge base, by implementing some meta-rules that aim at inferring the operations students applied in order to produce a step. We evaluated our approach through a user study with forty-three seventh grade students. The students who interacted with our tool showed statistically higher scores on equation solving tests, after solving algebra exercises with PAT2Math during an approximately two-hour session, than students who solved the same exercises using only paper and pencil. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction Intelligent Tutoring Systems (ITSs) have shown promising results when applied as a supplemental classroom learning tool (Koedinger, Anderson, Hadley, & Mark, 1997; Nicaud, Bittar, Chaachoua, Inam- dar, & Maffei, 2006; Nicaud, Bouhineau, & Huguet, 2002). Large- scale experiments in high-schools demonstrated that ITSs can im- prove students learning (Koedinger et al., 1997; Koedinger & Sueker, 1996). The success of this type of educational software is due to the fact that it can offer important features to personalize the learning processes such as one-on-one learning, immediate personal feed- back, demonstration of problem solving when students are having difficulty, and assessment of students’ skills. Vanlehn (2006) describes the tutor as having two loops. The outer loop is responsible for deciding the sequence of exercises or problems for students to work on. The inner loop provides step- by-step guidance during problem solving activity. In order to provide immediate feedback in the inner loop, an ITS’ architecture is generally composed of an expert system module (ES) that is able to solve the same type of exercises that students should do in multiple ways. Thus, for each step of the problem the system compares the answer provided by a student with the expert system’s solutions (an answer may have several correct solutions) and checks whether they are equivalent. If the ES can generate or test all possible solutions to a given prob- lem, then it can identify when a student has taken an incorrect path to solve the given problem and offer immediate feedback (generally colored labels are presented to indicate whether or not the student’s answer for a given step is correct) (Heffernan, Koedinger, & Razzaq, 2008). Some tutors provide additional re- sources (e.g. explanations or hints) when students are having difficulty or solving the problem or arriving at the correct answer. In fields such as math and physics, the knowledge is usually implemented as a rule in the form: ‘‘if hcondition is truei then hdo action Ai’’. Each rule represents an operation that can be applied in a step to solve a problem. The ES inference engine scans the base searching for rules to be triggered, i.e. rules whose conditions are satisfied by the current step of the solution. 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.04.004 ⇑ Corresponding author. Tel.: +55 51 3591 1100x1626. E-mail addresses: pjaques@unisinos.br, patricia.jaques@gmail.com (P.A. Jaques), hseffrin@hotmail.com (H. Seffrin), geiserubi@gmail.com (G. Rubi), felipedemor- aisfm@hotmail.com (F. de Morais), cassioghilardi@hotmail.com (C. Ghilardi), ig.ibert@ic.ufal.br (I.I. Bittencourt), sisotani@icmc.usp.br (S. Isotani). Expert Systems with Applications 40 (2013) 5456–5465 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.dyndns.org/dialog/?doi=10.1016/j.eswa.2013.04.004&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.04.004 mailto:pjaques@unisinos.br mailto:patricia.jaques@gmail.com mailto:hseffrin@hotmail.com mailto:geiserubi@gmail.com mailto:felipedemoraisfm@hotmail.com mailto:felipedemoraisfm@hotmail.com mailto:cassioghilardi@hotmail.com mailto:ig.ibert@ic.ufal.br mailto:sisotani@icmc.usp.br http://dx.doi.org/10.1016/j.eswa.2013.04.004 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa In addition to providing appropriate feedback to students, the rules also contain information about the related knowledge com- ponents. For example, a rule can represent the process (knowledge component or skill) of ‘‘subtracting an integer b on both sides of the equation’’. Thus, when the ES module corrects a student step solution, it is able to provide the student model with information about the skills necessary to solve that step. The student model uses this information to infer which skills students have mastered and which they need to practice more. This allows the tutor to cre- ate personalized hints in the inner loop and also to select more appropriate exercises for students to solve (outer loop). Although there are well known algebra tutors (Chaachoua, Ni- caud, Bronner, & Bouhineau, 2004; Cohen, Beal, & Adams, 2008; Koedinger & Sueker, 1996; Melis, Goguadze, Libbrecht, & Ullrich, 2009), few of them have an expert system module that is able to solve exercises and provide step-by-step guidance (Chaachoua et al., 2004; Koedinger & Sueker, 1996), an essential feature for a learning system to be classified as an ITS, according to (Vanlehn, 2006). Furthermore, previous work does not explore in detail how to implement an expert system module, which artificial intel- ligence knowledge representation format needs to be used, when an inference mechanism should be trigged and how to solve some inherent computational complexity problems. This paper presents the ES of the algebra tutor PAT2Math. PAT2- Math is an intelligent tutor system that teaches students how to solve linear and quadratic equations. It is a web system imple- mented in Java, which allows students to use it in any computer or platform with Internet access. PAT2Math is composed of an algebra editor (PATequation), which assists students in solving equations. The ES has an essential role in PATequation; it is responsible for providing immediate feedback to students at every step of their problem solving. Our main goal is to present this module knowl- edge and explain how to the ES implements problem solving and provides the student with step-by-step guidance. We describe how to reduce the complexity of this module from O(nd) to O(d), where n represents the number of rules in the knowledge base, by using meta-rules that guide the inference of the operations stu- dent applied to produce a step. We finish this paper by presenting the results of a user study we conducted with forty-three 7th grade students who interacted with PATequation for three classes. This paper is organized as follows. Section 2 describes problem solving under a pedagogical perspective. Section 3 presents the current state of the art in Algebra Intelligent Tutoring Systems. In Section 4, we explain the main artificial intelligence techniques used to develop ITS expert systems. In Section 5, we describe PAT2- Math, the Algebra Tutor that our research group is developing. PATequation, the problem solving editor of PAT2Math, is presented in Section 6. The ES responsible for providing step-by-step guid- ance in PATequation is described in Section 7. The experiment de- sign and results are reported in Section 8. Finally, Section 9 presents our conclusions. 2. Solving algebraic problems An Algebra task is generally a word problem for the student to solve. An algebraic word problem consists of one or more sen- tences representing a situation or a story, where the student needs to understand the elements in order to generate a mathematical model to represent it. The model consists of one or more equations that the pupil should solve in order to obtain the numerical values that are the solution of the problem (Gama, 2004). Take for exam- ple the word problem below (adapted from Munem & West (2003, p. 107)): ‘‘A computer store sells desktop and laptop computers. Due to space considerations, the number of laptops in inventory is seven less than twice the number of desktops in stock. How many desktops does the store have if it has a total of 272 computers?’’ The process of solving a word problem has two phases (Mayer, 1999; Polya, 2004): (i) the Problem Representation (also called Symbolization (Heffernan, Koedinger, & Razzaq, 2008)), and (ii) the Problem Solution. While the former concerns to the transfor- mation of algebra word problems into a system of equations, the second encompasses the process of solving these equations using algebraic operations. For example, for the word problem that we previously pre- sented, the student could provide the following solution: x þð2x � 7Þ¼ 272 ð1Þ 3x � 7 ¼ 272 ð2Þ 3x ¼ 279 ð3Þ x ¼ 93 ð4Þ In the example above, line (1) refers to the process of Problem Rep- resentation, and lines (2–4) to the Problem Solution phase. As shown in the above solution, solving a task involves several steps. Each line provided by the student in the above solution is a step. A step can involve the correct use of one or more Knowledge Components (KC) (also called knowledge units (Aleven, McLaren, Sewall, & Koedinger, 2009)). It comprises any unit into which the knowledge can be broken down, such as rules, concepts, facts, and procedures (Vanlehn, 2006). For instance, in order to arrive at line (2) in the above example, the student applied the operation (or KC) ‘‘add variable coefficients’’ in line (1) of the equation. In the next section, we will describe the main Algebra Intelli- gent Tutoring Systems and the tools and types of feedback they of- fer to help students solve algebra word problems in these two phases. 3. Algebra Intelligent Tutoring Systems The field of ITS has shown significantly improvements since the emergence of the first systems in the eighties (Woolf, 2009). The evolution of the Internet, the increasing performance of computers, and improvements of artificial intelligence techniques and tools have furthered the development of ITSs in several domains, such as Physics, Math, Medicine and others (see (Woolf, 2009) for an overview). Previous work has largely been applied in classroom settings, demonstrating they can improve student performance on stan- dardized and experimenter-designed tests by one-half to two stan- dards deviation (Cohen et al., 2008; Koedinger & Sueker, 1996; Shelby et al., 2000). Some of the most known research focused on the Algebra content domain. This is the case of Cognitive Alge- bra Tutor (previously PAT) (Koedinger & Sueker, 1996), Aplusix (Nicaud et al., 2006, Nicaud, Bouhineau, & Huguet, 2002), Active- Math (Goguadze & Melis, 2008; Melis, Goguadze, Libbrecht, & Ull- rich, 2009) and AnimalWatch (Birch & Beal, 2008; Cohen et al., 2008). We believe there are two main reasons for this. First, Alge- bra is a content domain (or task domain (Vanlehn, 2006)) in which a great number of students experience poor achievement (Carpen- ter, Kepner, Corbitt, Lindquist, & Reys, 1982; National Commission on Excellence in Education, 1983). Secondly, it requires less effort to formalize math content into computer algorithms, because it is mainly composed of procedural content, which can be easily rep- resented by computer algorithms. In the end of this section, we de- scribe the main algebraic tutors proposed by the Artificial Intelligence and Education community. P.A. Jaques et al. / Expert Systems with Applications 40 (2013) 5456–5465 5457 https://isiarticles.com/article/52573 
work_llg5s5igsrckbapxj33ntbcgb4	----	A proposed expert system for nursing practice Journal of Medical Systems, Vol. 9, Nos. 1/2. 1985 A Proposed Expert System for Nursing Practice A Springboard to Nursing Science J u d y G. O z b o l t , S a m u e l S e h u l t z II, M a r y A n n S w a i n , a n d I v o L . A b r a h a m The knowledge on which nursing practice is based comes largely f r o m traditional sources, expert nurses passing on the wisdom o f their experience to novices. Nursing research, although in- creasing, is usually parallel to nursing practice, and its findings, at best, are implemented only after long delays. Consequently, the most effective nursing responses to a particular client problem may be undiscovered or unknown. Nursing #formation systems reflect the nature and usage o f nursing knowledge. They offer standard care plans, but the knowledge and decision structures f o r individualizing care remain exclusively in the mind o f the nurse. Nurses may have great freedom to enter information into the information system, but the information is rarely retrievable in a form suitable f o r evaluation or research. Nursing practice, and the knowledge on which it is based, could be enhanced through the use o f a novel expert system. This paper describes how such a system could be developed, with examples from the authors' prototype programs. Taxonomies o f data, diagnoses, objectives, and interventions would make it possible to compare patients and to determine the relative effectiveness o f nursing interventions. A built- in evaluation component would provide feedback and correction. Everyday nursing practice would become a f i e l d f o r research, and the knowledge gained from research would immediately be f e d back into practice. In its development and in its implementation, this kind o f system would help to build nursing science. I N T R O D U C T I O N For more than a century, nurses have been developing and communicating the knowledge on which their practice is based. Significant scientific research on nursing practice, however, has been under way only since the 1960s. Most nursing knowledge, therefore, has been gleaned from the clinical experience of practitioners and passed down from experts to novices. As a result, it is often idiosyncratic in conceptualization and untested From the Center f o r Nursing Research, The University o f Michigan, Ann Arbor, Michigan 48109, and the School o f Nursing, Case Western Reserve University, Cleveland, Ohio 44106. 57 0148-5598/85/0400-0057504.50/0 © 1985 Plenum Publishing Corporation 58 O z b o i t et al. in validity and reliability. Early information systems for nursing care have reflected the largely unsystematic nature of nursing knowledge. That is, they have allowed nurses to share and communicate the kinds of information they believe are useful, but they have provided only the most rudimentary decision s u p p o r t - - i . e . , standard care plans associ- ated with common problems. Decisions about how to derive diagnoses from data and how to individualize the care plan rest with the cognitive and intuitive processes of the nurse. Developing the next generation o f nursing information systems will require that nursing knowledge be codified for incorporation into the systems. Once such systems are in place, they will provide a mechanism for systematically testing existing knowledge and feeding back new knowledge into practice. T H E N A T U R E O F N U R S I N G K N O W L E D G E W h a t K i n d o f K n o w l e d g e Does N u r s i n g Need? Before Florence Nightingale, nursing was assumed to consist of the intuitive, nur- turant action " n a t u r a l " to any woman. Nightingale I established that nursing was the application of specialized knowledge and skills, and that the knowledge and skills had to be learned. Nightingale saw nursing as " t h e proper use o f fresh air, light, warmth, cleanliness, quiet, and the proper selection and administration o f d i e t - - a l l at the least expense of vital power to the patient. ''1 In her view, then, nurses required knowledge of basic principles o f hygiene, comfort, and energy conservation, and skill in imple- menting them. Since Nightingale's time nursing has become increasingly complex and sophisti- cated. Although Nightingale's principles have not been abandoned, many nurses have attempted to give a more comprehensive description of nursing as its is practiced today. One of the more influential descriptions, that o f Virginia Henderson, z has been adopted by the International Council o f Nurses and by many other nursing organizations, schools, and agencies worldwide: Nursing is primarily assisting the individual (sick or well) in the performance of those activities contributing to health, or its recovery (or to a peaceful death) that he would perform unaided if he had the necessary strength, will, or knowledge. It is likewise the unique contribution of nursing to help the individual to be independent of such assistance as soon as possible. 2 According to Henderson, nurses need to know which activities will contribute to health or to its recovery or to a peaceful death. They must be skilled in performing these activities and in helping others to become able to perform them. All this implies that nurses are able to distinguish between health and illness, to set appropriate goals, to select those actions most likely to help the individual, and to foster the strength, will, and knowledge that promote independence. In the 1960s and 1970s many nurses developed descriptions o f their concepts of nursing. Orem's 3,4 description of nursing as the provision and management o f self-care seems to be a direct descendant of Henderson's, 2 and so would require similar kinds of knowledge. Roy, 5,6 on the other hand, said that the purpose of nursing was to promote adaptation. Nurses practicing within this framework would need to know how to rec- Proposed System for Nursing Practice 59 ognize positive adaptation and maladaptation and how to identify and modify the stimuli impinging on a patient. Rogers 7 offered yet another perspective, saying that the focus o f nursing is "unitary m a n , " an indivisible whole that is greater than the sum o f its parts and that is in continuous interaction with the universe. According to Rogers, nurses must interact with the whole person rather than focusing on parts o f the person, but she offered little guidance about how to do this or with what objectives, or about the kinds o f knowledge that would be needed. More recently, Erickson, Tomlin, and Swain 8 have developed a theory and paradigm of nursing that, while original in perspective, incorporates many elements o f earlier descriptions of nursing. According to their definition, Nursing is the holistic helping o f persons with their self-care activities in relation to health. This is an interactive, interpersonal process that nurtures strengths to enable development, release, and channeling o f resources for coping with o n e ' s circumstances and environment. The goal is to achieve a state o f perceived o p t i m u m health and contentmentfi Nurses who use this paradigm would require knowledge o f healthy functioning and human development as well as the clinical sensitivity and interpersonal skills to identify each client's strengths and needs and to select appropriate responses to promote health. All these definitions depict nursing as nurturant, as focusing on care rather than cure. Although there is considerable overlap among conceptual frameworks in the content of the knowledge needed for nursing practice, there are differences in perspective, em- phasis, and terminology. Nurses have attempted to apply knowledge derived from prac- tice or borrowed from other disciplines to one or another of these frameworks. Only recently, however, have some researchers begun to develop knowledge within particular nursing frameworks, describing nursing phenomena from a consistent perspective and empirically defining relationships among them. The lack o f a well-developed episte- mological system of nursing knowledge hinders nursing practice and inhibits the devel- opment of computer systems to facilitate nursing's cognitive tasks. Progress in nursing would be advanced by programs of research within selected frameworks to develop coherent bodies o f knowledge. What Kinds of Knowledge Does Nursing Have? Although there is great diversity in the conceptual models o f nursing, most nurses agree that they provide care via a problem-solving process that includes collecting and analyzing data, formulating diagnoses, setting objectives, choosing and implementing interventions, and evaluating outcomes. This process is depicted in Figure 1. The general model of the nursing process is shown as an open system with a feedback loop. When interventions have been implemented, this information is fed back into the system. New data are collected and new diagnoses are synthesized. These new diagnoses are, in effect, outcomes of the care that has already been given. They are compared with the original diagnoses and the objectives to determine whether the client has progressed, regressed, or stayed the same. The next step is to analyze the causes o f success or failure. If the intervention was effective in helping the client to achieve the objective or to move toward the objective, this is noted for future reference. If the client did not achieve the objective, the nurse must ask why. Was the diagnosis incorrect? Was the objective inappropriate or unrealistic? Was the intervention ineffective? Was the target date too 60 Ozbolt e t al. J~] COLLECTING DATA I AN+z'NG °A'A I ÷ SYNI HESIZING NURSING DIAGNOSES ,L " FORMULATING OBJECTIVES J~ P ,~' ESTABLISHING PRIORITIES FOR OBJECTIVES + ESTABLISHING TARGET DATES FOR EVALUATING ACHIEVEMENT OF OBJECTIVES YES 4 SELECTING NURSING INIERVENTIONS J EVALUATING CARE GIVEN ANALYZING CAUSES OF SUCCESSES AND FAILURES IMPtE/V, ENIING NURSING INTERVENTIONS j H O YES + Figure. 1. General model of the nursing process. early? The nurse investigates, makes a judgment, and continues the process, setting new objectives and target dates, choosing and implementing new interventions (or continuing with the same ones), and evaluating the results. This goes on, ideally, until the diagnoses indicate that the client has no problems requiring nursing assistance. The nurse and client Proposed System for Nursing Practice 61 choose objectives and interventions related to termination o f the professional relationship and exit from the system. Although this model is a simplified and idealized representation o f the complexities of nursing practice, the process it depicts is generally accepted in professional circles as the means by which nurses o u g h t to d e l i v e r care. I n d e e d , the A m e r i c a n N u r s e s ' s Association 9 has set standards for each step o f the nursing process. Yet nurses often have difficulty implementing the process, even where information systems have been set up to support the process. 10 One source of difficulty is the fragmentary nature o f knowledge at each step o f the process. Note that in the general model shown in Figure l, the content o f the boxes is not defined. What data are to be collected? What are possible diagnoses, objectives, and interventions? Nurses have used each of four general '°methods o f k n o w i n g " 11 to answer these questions. These methods are tenacity, authority, intuition, and science. In the method o f tenacity people hold firmly to what they know to be true because they have always known it to be true. In the nursing process the method of tenacity results in ritualistic action. Nurses collect certain data because they believe they must have the data because they have always collected them, even if they don't know what to do with the data once collected. The method o f tenacity can also result in ritualistic interventions, always treating a problem in the same way because '°that's the way it's done h e r e . " When nurses operationalize the nursing process using knowledge derived from the method of tenacity, they risk perpetuating false knowledge and overlooking valid but less traditional knowledge. Nurses who use the method of authority accept certain knowledge because it is declared to be true by a recognized authority or authorities. Nurses are using the method of authority when they choose particular interventions because a nursing textbook rec- ommends them, or when they carry out a treatment according to the instructions in a procedure manual. The method of authority is a useful way o f knowing since it would be impossible for every nurse personally to conduct empirical tests o f all knowledge before accepting it. Like the method of tenacity, however, the method o f authority may result in accepting invalid knowledge and ignoring valid knowledge. The third method of knowing, the a priori method or the method o f intuition, relies on common sense to determine what is true. Nurses use the a priori method when they review a mass of patient data and make an intellectual leap to diagnoses, or when they consider alternative objectives and interventions and, on the basis o f careful reasoning (but no scientific data), select those they believe to be most appropriate for their clients. The hazard in this method is that reasonable people can disagree, and the a priori method itself cannot resolve the dispute by determining who is correct. Finally, nurses are increasingly using the method o f science. This method o f knowing is characterized by objectivity, as propositions are put to empirical tests, and is capable of self-correction on the basis o f the results o f those tests. Particularly since the 1960s, an increasing number o f studies have described nursing phenomena and re- lationships and provided evidence of the relative effectiveness o f nursing interventions. Because nurse researchers conduct their investigations from a variety o f perspectives on the nature of nursing or from no explicit nursing perspective at all, however, it is difficult or impossible to synthesize their findings into a coherently growing body o f knowledge. How, for example, does one relate one researcher's findings about how to promote 62 Ozbolt et al. adaptation to another researcher's description o f self-care assets and deficits? And how can these diverse bits of knowledge be used to operationalize nursing data, diagnoses, objectives, and interventions? Such difficulties have hindered the implementation o f research knowledge in nursing practice. The problem is compounded by the fact that most researchers are not practicing clinicians, and most nurses are not researchers; in 1978 only 5% of nurses held a master's or doctoral degree. 12 Consequently, research often goes on parallel to the world of nursing practice, and findings are implemented after long delay, if at all. If nursing knowledge is an amalgam o f ritual, tradition, intuition, and science, how is it organized and used? In some settings, nurses have identified nursing problems that occur frequently on their units and have devised standard care plans for those problems. A variety of classification schemes for patient problems have been developed and im- plemented in hospitals and other health care agencies. In reviewing some of these schemes and considering the North American Nursing Diagnosis Association's efforts to classify its own set of diagnoses, Kritek 13 proposed that nurses need to differentiate between data and diagnoses, which are confounded in many attempts at classification; to find ways o f classifying that are consistent with holistic nursing, rather than splitting persons into minds and bodies; to identify the common denominators in existing classification schemes and assess what is useful; and to find organizing principles appropriate to nursing. If nurses need to find useful and appropriate ways to classify nursing data and diagnoses, the same is true for nursing objectives and interventions. Meanwhile, for most nurses in most situations, the organization o f nursing knowledge is personal and idiosyn- cratic. Each nurse decides, on the basis o f his or her own synthesis o f knowledge, the data to collect, the problems to attack, the objectives to pursue, and the interventions to implement. One nurse's experience remains highly individualized and difficult to inte- grate with another's. Thus, although nurses have learned a great deal about practice from experience and from research, nursing knowledge is disunified and fragmentary, k will remain so until nurses find organizing principles to permit its systematic testing, syn- thesis, and codification. E X I S T I N G N U R S I N G I N F O R M A T I O N S Y S T E M S Representative Systems The first comprehensive hospital information system was the Technicon system at E1 Camino Hospital in Mountain View, California. Its nursing component reflects the state of nursing knowledge at the time of its development, the late 1960s and early 1970s. Charged with devising a way to computerize nursing notes, the traditional narrative charting of patient conditions and nurse actions, a committee of nurses recognized that their task was a more fundamental one, to develop a comprehensive, integrated system for processing nursing care data, a system that would include care planning, charting, and feedback for audit. 14 Relying primarily on the methods o f authority and intuition, they identified common nursing problems frequently associated with certain medical diagnoses and developed standard care plans for the nursing problems.15 Ultimately, a menu-driven system was developed wherein nurses use a light pen and a keyboard to Proposed System for Nursing Practice 63 enter patient data, select and modify nursing problems and care plans, and record nursing actions and patient progress. For example, a medical diagnosis o f cholecystitis or cholelithiasis will cause the system to produce a nursing care plan that includes four "usual problems": pain, nausea, potential abdominal distention, and potential misunderstanding o f diet at home in relation to diagnosis. For each problem, the system also produces a list o f expected outcomes, deadlines, and nursing orders. 15 This standard care plan will be produced for any patient with a medial diagnosis of cholecystitis or cholelithiasis. The nurse can then modify the care plan or add "unusual problems" and related outcomes, deadlines, and nursing orders. Note that the system itself produces a standard care plan in response to the cue of a medical diagnosis. It is not constructed to compare patient data to a nursing knowl- edge base and reach decisions about nursing problems or nursing care. Those activities must be done by the nurse. This fact allows nurses in different settings using the Technicon system to organize their thinking about nursing in their own ways. In the Technicon installation at the NIH Clinical Center, for example, Mayers's 16 common nursing problems and standard care plans are not used. Instead, the nursing process has been operationalized in a framework based on Maslow's 17 hierarchy o f needs, with a different set o f standard care plans. Still, it is up to the nurse, not the computer system, to classify data according to need cate- gories, to determine the nursing problems, and to select and modify care plans appro- priately, i0 Although developed later, the nursing care plan module of IBM's Patient Care system bears significant similarities to the nursing component of the Technicon system. The system stores a limited number o f nursing diagnoses (one hospital reports having 21 nursing diagnoses on file) with associated care plans. It is up to the nurse to decide which if any of the stored nursing diagnoses applies to a particular patient. The nurse may then call up the relevant diagnoses and care plans and edit the care plans to corre- spond to the patient's situation. Components of the diagnosis and care plan include the following: (1) statement of the diagnosis (e.g., '"decreased mobility/immobility"); (2) identifying information about the patient (age, sex, room number, physician); (3) nursing m e a s u r e s - - a s s e s s m e n t s (what to assess, e . g . , mobility, capabilities, weakness); (4) nursing measures--interventions (what to do, e.g., mobilize by hail walk with crutches Q.I.D.); (5) nursing m e a s u r e s - - t e a c h i n g (what to teach, e.g,, appropriate activity and rest); (6) goals ( e . g . , m a x i m u m mobility, ambulates independently); (7) evaluation (whether or not goals are met, whether care plan is to be continued, revised, or discon- tinued).~a It is apparent that the diagnostic statements are quite broad, and the care plans are very general. These features permit a small number o f diagnoses and care plans to apply to a large number o f patients. The question might be raised as to whether such broad and general care plans are useful if the goal is to provide individualized care. The computer, however, provides no assistance in individualizing the care plan. If the care plan is to be individualized, the nurse must do it. Nursing Knowledge and Existing Systems The Technicon and IBM systems, like other commercially available nursing infor- mation systems, reflect many of the limitations in nursing knowledge described earlier. 64 Ozbolt et al. They are adaptable to many different conceptualizations of nursing, or may be used with no explicit or consistent conceptualization at all. The nursing knowledge needed to produce standard problem lists and care plans may be derived from tenacity, authority, intuition, or research, but once placed in the system it remains static. The more complex and fluid knowledge and decision structures necessary to design and evaluate care ap- propriate to individual patients remains exclusively in the mind of the nurse. As repos- itories of standard care plans, existing nursing information systems are little more than filing cabinets for the most simplified forms o f nursing knowledge. They neither capture the complexity and richness of nursing knowledge nor help to advance it. A PROPOSED EXPERT SYSTEM The Knowledge Base It is possible to develop an expert system for nursing practice with a much richer and more complex knowledge base, one that can be used by the computer for artificially intelligent reasoning and one that will grow with experience. Such a knowledge base would require a high degree of internal consistency, since its elements and their rela- tionships would have to be clearly and explicitly defined. How is this possible, given the "disunified and fragmentary" nature o f nursing knowledge described previously? It can be achieved through a process of bootstrapping, starting from the knowledge that is currently available and strengthening and enriching it over time. First, to achieve internal consistency in the knowledge base it is useful to select a conceptual model of nursing and to deduce taxonomies for each step of the nursing process. There is no compelling reason to choose one of nursing's conceptual models over another, but because of its practical emphasis on care and its own internal consis- tency our group has been working with Orem's. 3'4 For future work we are considering a change to the Erickson, Tomlin, and Swain s model because o f its greater richness and complexity, but this paper will use examples from our work with Orem's model. According to Orem, each person requires certain kinds of self-care to maintain or restore health or to promote normal development. When the person is unable to provide the necessary self-care because of illness, injury, or disability, or because the needed care requires specialized knowledge and skills, assistance in the form o f nursing is called for. The nurse thus collects data to obtain evidence of the degree to which self-care requisites are met and the degree to which the client is able to meet them independently. Orem 4 defined 14 categories of self-care requisites. These are as follows: (1) maintaining sufficient intakes o f air, water, food; (2) providing care associated with eliminative processes and excrements; (3) maintaining a balance between activity and rest; (4) main- taining a balance between solitude and social interaction; (5) preventing hazards to life, functioning, and well-being; (6) promoting normalcy; (7) bringing about and maintaining living conditions that support life processes and promote the process of development; (8) providing care to prevent or mitigate the deleterious effects o f conditions that can affect human development; (9) seeking and securing appropriate medical assistance when needed; (10) being aware of and attending to effects o f pathological conditions or states; (11) effectively carrying out medically prescribed actions; (12) being aware o f and at- tending to or regulating the discomforting or deleterious effects of medical care; (13) Proposed System for Nursing Practice 65 modifying the self-concept to accept o n e ' s state o f health and need for specific forms o f health care; (14) learning to live with the effects o f illness, injury, and treatment in a way that promotes continued personal development. 4 Specific data items can be identified in each category, and the relation o f each item or combination o f items to the status o f self-care can be described. For example, items related to maintaining a sufficient intake o f air include assessment for the presence of wheeze, stridor, and adventitious breath sounds. These items are related specifically to patency o f airways. I f the data tell us that airways are patent and if a number o f other conditions are similarly satisfied, then we can conclude that the self-care requisite for maintaining a sufficient intake o f air is met. For m a n y categories o f self-care requisites, valid and reliable measures for nursing assessment can be borrowed from the research- based Horu-Swain 19 instrument, which was also developed in the Orem framework. Other items will need to c o m e from the literature or from nurses' experience or to be invented as a part o f systems development. Nursing diagnoses are easy to define within O r e m ' s conceptual framework. Since the focus is on self-care, nursing diagnoses are statements o f the degree to which self- care requisites are met and o f the degree to which the client is able to meet them independently. These statements are specific to the individual client and are based on the data collected. Because the relationships between data and diagnoses are clem'ly delineated, it is possible for the computer to derive diagnoses from the data. Thus, nurses using this sytem would no longer collect data to no purpose, or overlook important data and miss diagnoses. The ability to generate client-specific diagnoses and care plans from the nursing assessment data has not been realized in any commercially available nursing information system. That this is possible, however, has been demonstrated b y our prototype pro- grams, 2°,21 which produce diagnoses like this: Maintaining a sufficient intake of air is impaired by a. infection or irritation in the lungs or airways b. increased airflow resistance Self-care agency for maintaining a sufficient intake o f air is impaired by a. inadequate knowledge related to alterations in respiratory function appropriate actions to take when alterations occur b. inability to adapt activity so that oxygen supply is equal to demand c. embarrassment re clearing secretions in social situations Nursing objectives are related to diagnoses and are statements to the effect that self- care requisities will be met and that the ability to meet them independently will be maintained or augmented to the degree feasible. Like diagnoses, objectives are specific to the individual client. An example o f objectives related to the diagnoses above would be: Client will maintain a sufficient intake o f air, with no further evidence o f a. infection or irritation in lungs o f airways b. increased airflow resistance 66 Ozbolt e t al. Client's self-care agency for maintaining a sufficient intake of air will be augmented to include a. adequate knowledge related to alterations in respiratory function appropriate actions to take when alterations occur b. ability to adapt activity so that oxygen supply is equal to demand c. resolution of embarrassment re clearing secretions in social situations. Objectives are derived from the client data base, the diagnoses, and knowledge of what it is possible to achieve. Some of this knowledge is available from the research literature, but much o f it would have to come from the opinions of expert nurses (the a priori method of knowing). Once an expert system was operational, however, it could aggregate data about the circumstances under which objectives were achieved and correct its own structures to propose more appropriate and achievable objectives. Interventions, in this conceptual model, have to do with supplying the self-care the client is unable to do unassisted and promoting the client's self-care agency. They are also specific to the diagnoses and objectives. Since the research on effectiveness o f nursing interventions does not begin to cover the number and variety o f possible inter- ventions as defined within this framework, the knowledge base for system development would be largely a priori. Again, however, evaluations of the effectiveness o f interven- tions could be aggregated over time so that the system could learn from its own expe- rience. T h e D e c i s i o n S t r u c t u r e s Once the elements of the knowledge base and their relationships have been defined, it is possible to develop decision mechanisms. Both the knowledge base and the client data base are extensive, and complex decisions must be made throughout the nursing process. Decision structures are needed that will produce accurate and reliable decisions without requiring inordinate time for data processing. Statistical decision models used in some systems for medical diagnosis are reliable but slow, overlooking associative pro- cesses in data acquisition and storage that could streamline search and retrieval. On the other hand, human inferential strategies, even as mimicked in some computerized de- cision-support systems, are subject to biases. In an effort to overcome these limitations, we have proposed a multivariate math- ematical approach to clinical decision making that permits the incorporation of infor- mation-processing strategies experts typically used while maintaining the statistical rigor of data-based systems. 22,23 In this approach, both item information and associative in- formation are used for clinical inference. Item information, the actual data that are acquired, is represented as a vector in which the features o f the item constitute the elements. As each item is acquired, it is immediately integrated into a larger network of information. The integration of two items o f information into a new item (mathematically, the convolution of two item vectors into a composite vector) results in associative in- formation. The clinical condition of a patient at any time is therefore defined as a vector composed of both item and associative information acquired to that time, with appropriate weighting and with a defocusing factor indicating loss of information (since information acquired earlier may be partially lost by the time of a subsequent assessment). Because Proposed System for Nursing Practice 67 clinical inference includes not only a description of the patient's condition at any time, but also a judgment of the normality or abnormality of the condition, the clinical condition vector is next compared to upper- and lower-limit vectors representing the population criterion interval. These population criteria are stored in a normative data base, or knowl- edge base. If the patient condition vector falls outside these limits, the condition is considered abnormal. The proposed decision structures thus combine the associative processes of data acquisition and storage generally found in heuristic approaches with the normative comparisons of normality and abnormality commonly found in statistical approaches. Although we have not yet tested this approach in a prototype program, we are hopeful that it will provide an effective and efficient tool for clinical decision support in nursing. Toward the Development of Nursing Science The creation and implementation of an expert system such as the one described here would do much to advance the development of nursing science. Work already completed on prototype programs demonstrates the feasibility of organizing and codifying nursing knowledge, even when some of the knowledge is based on tradition, experience, or common sense rather than research. A system so organized and codified, with defined and traceable decision structures, would be an invaluable resourse for research and eval- uation. A taxonomy of patient data with valid and reliable measures combined with a taxonomy of nursing diagnoses reliably derived from the data would make it possible to compare patients to determine similarities and differences in their problems, strengths, and responses. Taxonomies of objectives and interventions would make it possible to determine the relative effectiveness of interventions in achieving the objectives, given particular patterns of nursing diagnoses. A built-in evaluation component that took note of successes and failures and accordingly adjusted the system's decision making would provide quick feedback and correction. The day-to-day activities of nursing practice would automatically and continuously be the subject of research, as the sytem aggregated data and interpreted successes and failures. Furthermore, the knowledge from that re- search would be immediately fed back into practice. An expert system, then, would contribute to nursing knowledge first by requiring that existing knowledge be organized and codified for the development of the system. It would make a still greater contribution by making the everyday practice of nursing amenable to research and by putting back into practice the knowledge gained from that research. The development of expert systems for nursing need not await fuller emergence of nursing science; on the contrary, expert systems can become a springboard to nursing science, REFERENCES 1. Nightingale, F., Notes on nursing: What it is and what it is not, Appleton, New York, 1860. Republished Dover, New York, 1969. 2. Henderson, V., ICN basic principles o f nursing care, ICN (International Council of Nurses) House, London, 1960. 68 Ozbolt e t al. 3. Orem, D. E., Nursing: Concepts of practice, McGraw-Hill, New York, 1971. 4. Orem, D. E., Nursing: Concepts o f practice, 2nd ed., McGraw-Hill, New York, 1980. 5. Roy, C., Sr., Adaptation: A conceptual framework for nursing. Nursing Outlook 18(3):42-45, 1970. 6. Roy, C., Sr., Adaptation: A basis for nursing practice. Nursing Outlook 19:254-257, 1971. 7. Rogers, M. E., An introduction to the theoretical basis o f nursing, Davis, Philadelphia, 1970. 8. Erickson, H. C., Tomlin, E. M., and Swain, M. A. P., Modeling and role-modeling; a theory and par- adigm for nursing, Prentice-Hall, Englewood Cliffs, New Jersey, 1983. 9. American Nurse's Association, Nursing: A social policy statement. Author, Kansas City, Missouri, 1980. 10. Romano, C., Documentation of nursing practice using a computerized medical information system. Pro- ceedings o f the fifth annual symposium on computer applications in medical care, IEEE Computer Society, Los Angeles, 1981. 11. Kerlinger, F. N., Foundations of behavioral research, 2nd ed., Holt, Rinehart & Winston, New York, 1973. 12. Moses, E. A., and Roth, A., Nursepower: What do statistics reveal about the nation's nurses. Am. J. Nursing 79:1745-1756, 1979. 13. Kritek, P. B., Current nomenclature and classification systems. Classification o f nursing diagnoses: Pro- ceedings ofthefifth national conference (M. J. Kim, G. K. McFarland, and A. M. McLane, eds.), Mosby, St. Louis, 1984. 14. Gall, J. E., Demonstration and evaluation o f a total hospital information system (DHEW Pub. No. (HRA) 77-31880), U.S. Government Printing Office, Washington, D.C., 1977. 15. Mayers, M. G., and E1 Camino Hospital, Standard nursing care plans, Vol. 1, K.P. Medical Systems, Palo Alto, 1974. 16. Mayers, M. G., A systematic approach to the nursing care plan, Appleton-Century-Crofts, New York, 1972. 17. Maslow, A., Toward a psychology o f being, Van Nostrand, Princeton, 1962. 18. Light, N., Computers in nursing practice: On-line nursing care plans. Comput. Nursing 1(3):4, 1983. 19. Horn, B. J., and Swain, M. A., Criterion measures of nursing care quality: A health status instrument (PHS Publ. No. 78-3187), U.S. Government Printing Office, Washington, D.C., 1978. 20. Ozbolt, J. G., A prototype information system to aid nursing decisions. Proceedings of the sixth annual symposium on computer applications in medical care, IEEE Computer Society, Silver Spring, Maryland, 1982, 21. Ozbolt, J. G., Schultz II, S., Swain, M. A. P., Abraham, I. L., and Stein, K. F., Developing an expert system for nursing practice. Proceedings o f the eighth annual symposium on computer applications in medical care, IEEE Computer Society, Siver Spring, Maryland, 1984. 22. Abraham, I. L., Schultz II, S., Ozbolt, J. G., and Swain, M. A. P., A multivariate algorithm for diag- nostic information systems: Data acquisition and storage procedures. Proceedings of the Fifth Congress of the European Federation f o r Medical Informatics, Belgian Society for Medican Informatics, Brussels, 1984. 23. Abraham, I. L., Schultz II, S., Ozbolt, J, G., and Swain, M. A. P., A mutivariate mathematical algorithm for diagnostic information systems: Procedures for clinical inference. Proceedings o f the Eighth Annual Symposium on Computer Applications in Medical Care, IEEE Computer Society, Silver Spring, Maryland, 1984. 
work_lmef6ienqzcznh2ngbbeber26a	----	Expert environments: machine intelligence methods for ambient intelligence Expert environments: machine intelligence methods for ambient intelligence The ideas put forward by Donald A. Norman (1999) in his monograph entitled The Invisible Computer can be considered the main source of inspiration of a new research area, called ambient intelligence. Ambient intelligence, com- monly abbreviated as AmI, is primarily con- cerned with human–environment interactions. An environment is seen anthropomorphically, as an intelligent agent able to interact with users, creating for them processes to interpret, inform, communicate and dialogue (Abowd & Mynatt, 2000; Remagnino & Foresti, 2005). The history of ambient intelligence starts in Europe in 2001 with the Fifth European Frame- work Program. At that time, the IST Program Advisory Group (ISTAG) of the European Commission (Directorate General on Informa- tion Society and the Media) introduced the concept of ambient intelligence by publishing the report Scenarios for Ambient Intelligence in 2010 (Ducatel et al., 2001). Since then, ambient intelligence has been recognized in Europe as one of the key concepts related to the informa- tion and communication technology society. An updated version of the report was published in 2003 under the title Ambient Intelligence: from Vision to Reality (Ducatel et al., 2003). Ambient intelligence’s emphasis is on support to human interactions with the environment, user-friendliness, ubiquitous accessibility etc. and requires competences from many research areas, ranging from computer vision, machine learning, distributed computing and middleware, context awareness systems, sensor networks etc. Enticing illustrative scenarios have been pub- lished, in which the user wears technology that communicates with systems and devices present in the environment in order to provide informa- tion and receive services. Since its inception, ambient intelligence has inspired the design and implementation of sys- tem prototypes for intelligent spaces, developed and tested in controlled environments. The scope of this special issue is to publish innova- tive ideas on a selection of topics related to ambient intelligence. Cameras and computer vision algorithms, for instance, can be used to unobtrusively acquire awareness of the environment. In the paper entitled ‘A multi-camera vision system for fall detection and alarm generation’, Cucchiara et al. propose a system for using cameras to detect people’s falls. The use of cameras is preferred in ambient intelligence to acceler- ometers or other devices because they are not intrusive, people tend to get used to their pre- sence and people’s actions are less affected. In this paper, a system of multiple cameras is used to monitor people’s movements in house envir- onments and detect changes in people’s posture with the specific goal to identify falls. In the paper entitled ‘Understanding inten- tion of movement from electroencephalograms’, Lakany and Conway analyse people’s inten- tions, using electroencephalogram waves. Their method is non-intrusive and it uses a brain– computer interface. Support vector machines are used to select features and perform classifi- cation of intentions, and tests are performed to detect the direction of users’ movements. While the previous papers address ambient intelligence from the point of view of sensing Guest editorial__________________ c� 2007 The Authors. Journal Compilation c� 2007 Blackwell Publishing Ltd Expert Systems, November 2007, Vol. 24, No. 5 293 technologies and algorithms, the following two papers are mainly focused on knowledge representation for context awareness. The paper entitled ‘Knowledge representation for ambient security’ by Snidaro and Foresti proposes an ontology-based methodology for representing the interaction between the user and the en- vironment with specific reference to security scenarios. Similarly, in the paper entitled ‘Con- text-aware environments: from specification to implementation’ Reigner et al. are concerned with the problem of implementing a context model for a smart environment. Their paper proposes interesting approaches based on ‘net- works of situations’, introducing a comparison of the use of Petri nets and hidden Markov models. Finally, in the paper entitled ‘Collection, storage and application of human knowledge in expert system development’ Balch et al. propose an analysis of the knowledge engi- neering flow, encompassing knowledge acqui- sition, representation and inference. This flow analysis is presented and applied to the petro- leum industry application domain, and specific software tools that use fuzzy logic are utilized. Paolo Remagnino, 1 Andrea Prati, 2 Gian Luca Foresti 3 and Rita Cucchiara 4 (1) Faculty of Computing, Information Systems and Mathematics, Kingston University, UK E-mail: p.remagnino@kingston.ac.uk (2) Dipartimento di Scienze e Metodi dell’ Ingegneria, Universita’ di Modena e Reggio Emilia, Italy (3) Department of Mathematics and Computer Science, University of Udine, Italy (4) Dipartimento di Ingegneria dell’ Informazione, Universita’ di Modena e Reggio Emilia, Italy References ABOWD, G.D. and E.D. MYNATT (2000) Charting past, present, and future research in ubiquitous com- puting, ACM Transactions on Computer–Human Interaction, 7 (1), 29–58. DUCATEL, K., M. BOGDANIWICZ, F. SCAPOLO, J. LEIJTEN and J.-C. BURGELMAN (2001) Scenarios for Ambient Intelligence in 2010, Technical Report 10, ISTAG, February. DUCATEL, K., M. BOGDANIWICZ, F. SCAPOLO, J. LEIJTEN and J.-C. BURGELMAN (2003) Ambient Intelligence: from Vision to Reality, Technical Report, ISTAG. NORMAN, D.A. (1999) The Invisible Computer, Boston, MA: MIT Press. REMAGNINO, P. and G.L. FORESTI (2005) Ambient intelligence: a new interdisciplinary paradigm, IEEE Transactions on Systems, Man and Cybernetics: Part A, Systems and Humans, 35 (1), 1–7. 294 Expert Systems, November 2007, Vol. 24, No. 5 c� 2007 The Authors. Journal Compilation c� 2007 Blackwell Publishing Ltd 
work_ln3yl3e2jjezbhzw2ekrdhgmmq	----	Declarative Process Mining in Healthcare Marcella Rovania, Fabrizio M. Maggib, Massimiliano de Leonic,d, Wil M.P. van der Aalstd, Ronny S. Mansd, Alessandro Pepinoa aUniversity of Naples Federico II, Naples, Italy bUniversity of Tartu, Tartu, Estonia cUniversity of Padua, Padua, Italy dEindhoven University of Technology, Eindhoven, The Netherlands Abstract Clinical guidelines aim at improving the quality of care processes based on evidence-based insights. However, there may be good reasons to deviate from such guidelines or the guidelines may provide insufficient support as they are not tailored towards a particular setting (e.g., hospital policy or patient group characteristics). Therefore, we propose an approach to mediate between event data reflecting the clinical reality and clinical guidelines describing best-practices in medicine. Concretely, we repair and enrich declarative process models based on event logs. Initial models based on clinical guidelines are improved by observing the real process. Declarative models are used as they allow for more flexibility. Process mining techniques are used to check conformance, analyze deviations, and enrich models with conformance-related diagnostics. The approach has been applied in the urology department of the Isala hospital in the Netherlands. The results demonstrate that the approach is feasible and that our toolset based on ProM and Declare is indeed able to provide valuable insights related to process conformance. Keywords: Healthcare Processes, Process Mining, Declarative Modeling Languages 1. Introduction Healthcare processes can be classified either as medical treatment processes – concerning diagnostic and therapeutic activities performed to treat a patient – or as generic organizational processes – more related to administrative aspects [1]. The latter are generally handled according to well-established procedures; conversely, medical treatment processes must be flexible enough to be adapted to the variability of the healthcare environment and to the doctors’ freedom in their execution. Medical treatment processes are universally performed according to clinical guidelines, thus translating the evidence-based insights into actions [2, 3] that have proven to be effective [4]. Email addresses: marcella.rovani@unina.it (Marcella Rovani), f.m.maggi@ut.ee (Fabrizio M. Maggi), m.d.leoni@tue.nl (Massimiliano de Leoni), w.m.p.v.d.aalst@tue.nl (Wil M.P. van der Aalst), R.S.Mans@tue.nl (Ronny S. Mans), pepino@unina.it (Alessandro Pepino) Preprint submitted to Elsevier October 29, 2014 Guidelines are expected to be followed as they formalize the best practices [2]. Nonetheless, it is often observed, by monitoring process executions, that a ’gap’ exists between these recommendations and the actual clinical practice [5, 6, 3], i.e., there are differences between the executed activities and the activities recommended in the guidelines. Therefore, a crucial challenge for the healthcare management sector is to address the gap between the actual clinical processes and the recommendations given in guidelines. There is an urgent need for methods that • can measure the adherence of the actual process behavior with respect to the expected behavior, • pinpoint where deviations happen most frequently, • produce results that can easily be understood by doctors to highlight the most common causes of the identified deviations. This is the realm of process mining (see [7] for examples), which emerged a little over a decade ago. Process mining aims at extracting process knowledge from event logs. These logs may originate from different types of systems such as generic enterprise information systems as well as from Hospital Information Systems (HIS). Typically, event logs contain information about the start/completion of process tasks together with related context data (e.g., actors and resources) and timestamps. In a relatively short timespan, process mining has proven to be capable of providing deep insight into process-related problems that contemporary institutions face, including hospitals. Through the application of process mining, institutions can discover the processes as they are conducted in reality, check whether certain practices and regulations were really followed and gain insight into bottlenecks, resource utilization, and other performance-related aspects of processes. Most of the process mining techniques, both for process discovery and checking the conformance to regulations, rely on the idea that processes are described by procedural languages where the completion of a task may enable the execution of other tasks. While such a high degree of support and guidance is certainly an advantage when processes are repeatedly executed in the same way, in healthcare settings, this is considered to be too restrictive. As illustrated, e.g., in Rebuge et el. [8], medical treatment processes are, in fact, highly dynamic, highly complex, increasingly multidisciplinary and often ad hoc. Moreover, due to the complexity of the environment in which they are executed, they usually involve several variables that can be handled differently depending on the specific patient being treated [9]. Therefore, these process models need to provide more freedom and should not restrict users in taking appropriate actions. As suggested by Mulyar et al. [10], declarative process modeling languages provide more flexibility with respect to imperative languages when describing process behavior. Imperative languages can lead to unreadable and complex models in turbulent and variable 2 environments. Conversely, declarative languages do not model the allowed behavior, but rather what is disallowed. In this way, they offer more possibilities for executions since everything is allowed unless explicitly forbidden [11]. This is achieved by representing the process as a set of rules so that it can be executed in all possible ways as long as these rules are respected. In this paper, we present a methodology for analyzing medical treatment processes based on declarative languages and process mining techniques. The aim of this methodology is to investigate the discrepancies between expected and actual process behaviors of clinical processes. The identification of these discrepancies can be used to adapt the expected process behavior to the current practice or to find information about how the actual process execution can be modified to better fit the expected behavior. In order to demonstrate the applicability of our methodology, we show how the declarative modeling language Declare1 and the process mining tool ProM 2 can be used for acquiring insights about processes handled in the urology department of the Isala hospital. The hospital is one of the top-clinical hospitals in the Netherlands treating more than half a million patients per year distributed over five locations in and around Zwolle. The paper is structured as follows. Section 2 motivates the work and compares it with the state of the art. The proposed methodology is described in Section 3. Section 4 shows how recently developed plug-ins of the process mining tool ProM support our methodology. Section 5 illustrates the case study. Section 6 discusses the results obtained and concludes the paper. 2. Motivation and Related Work The gap between evidence-based guideline approaches and the actual execution of the guidelines is one of the key concerns of healthcare management [12, 5, 13, 14]. This gap has been highlighted for fields, such as dermatology [15], psychology [16], urology [17], etc. In [18], the authors investigate the reasons why clinicians do not always implement the best evidence. Through structured interviews, they found that personal experiences, backgrounds, relations with individual patients, logistic and practical aspects usually influence clinical decisions: as consequence, the actual practice deviates from the prescriptions. Sometimes, clinical decisions can also be influenced by the cost of a treatment and by the patients’ socioeconomic status [19]. Furthermore, as reported in [20], clinical guidelines may not adequately address issues relevant to everyday’s patient care. For example, patients can exhibit allergies that do not allow for the prescribed way of treating the pathology from which they are affected. Surgery, even if prescribed by the clinical guidelines, can be sometimes critical for patients having coagulation defects, diabetes, or heart problems. The coexistence of many pathologies can require further examinations or treatments not listed by the guideline. Hence, the expected treatment can be delayed or even avoided in some cases. Therefore, 1http://www.win.tue.nl/declare/ 2http://www.processmining.org 3 in many circumstances this gap is patient-related, since the doctors have to deal with the particular patient conditions which may require specific and exceptional treatments. In addition, there are some deviations that can be attributed to the difficulties of translating research findings into practice. As reported in [21], sometimes, patients cannot benefit of treatments of proven validity because of the time needed to incorporate research results into practice. This is the case, for example, of some experimental treatments tested in advanced medical centers responsible of the clinical trials execution. Although in those centers the effectiveness of a particular treatment has been demonstrated, and it is well recognized by the medical community, it requires time to transfer the adoption of the same technique to other contexts, so that misalignments can exist in the treatment of the same pathology in different realities. To alleviate this issue, the ‘Knowledge Translation’ approach [22] has been introduced. It has the specific aim of translating scientific knowledge into practical guidelines in the healthcare sector. However, the knowledge translation approach does not quantitatively investigate whether such guidelines, once formalized, are then applied in the actual context. In this paper, we approach the issue by suggesting a novel methodology to verify the correspondence of the guidelines statements to the reality and to highlight any possible deviations. Indeed, if some deviations can positively affect the patient process care, others represent errors that can compromise the patient recovery. Therefore, it is crucial to detect these deviations and to investigate the corresponding reasons in order to assess which of them can be classified as positive - and therefore can be taken into account to update the clinical guidelines statements - and which ones should be avoided and corrected in the clinical practice. As reported in [23], safety can be improved by learning from incidents and near misses. The need of monitoring the patient safety has becoming even more widespread in the healthcare sector, both in order to protect patients from adverse events [24] and to avoid the related costs in terms of money [25]. Moreover, one would like to learn from “positive deviants”. As shown in this paper, process mining techniques can be used to address all these issues. According to [16], our goal should be to address this gap with data rather then with debates or subjective arguments. Chart abstractions, observational studies, chart audits, results of clinical trials, structured interviews have proven to be useful instruments to investigate this gap in various clinical fields [15, 13, 26, 16, 3]. Furthermore, in [27], the authors highlight how the frequency and consequences of errors in medical care can be reduced by monitoring the activities through Hospital Information Systems. In general, as summarized by Lenz et al. [28], various studies have reported positive effects when using process-oriented IT systems in healthcare to support and monitor the processes execution [29]. In addition, several case studies have proven the usefulness of process mining in healthcare [8, 30, 31]. De Weerdt et al. [32] stressed the importance of process mining to acquire knowledge about past clinical experiences in order to avoid medical errors in the future. However, these techniques are based on ad hoc workflow models. In this paper we propose a new approach in order to investigate how to deal with the turbulence in healthcare environments. The approach is based on declarative process models. Conformance checking based on such models helps 4 to pinpoint deviations and learn from them. Moreover, we show that our declarative models can be used to better mediate between practice and guidelines. The resulting more balanced models can be used for understanding deviations and analyzing performance. 3. Methodology In this section, we describe a methodology for the analysis of medical treatment processes. In short, we first create a reference model representing the expected behavior of the process. Then, we compare the actual behavior of the process as recorded in a log with the expected one. The discrepancies can be used to highlight errors in the process execution or to adapt expected model to better reflect the reality expressed in the event log. The proposed methodology is outlined in Figure 1 and described below in detail. Event Log Test Log Training Log Clinical Guidelines Model Design (Hand made) Model Repair De Jure Model De Facto Model Split Event Log for Cross Validation Conformance Checking Record Event Data Healthcare Information System Conformance Analysis Report HIS Figure 1: Methodology for the analysis of medical treatment processes. 3.1. Model Design Based on Clinical Guidelines In the model design phase, a process model is created. The aim of this phase is to obtain a standard representation of a medical treatment process as described by the corresponding clinical guidelines. Since this representation is based on theoretical statements established by the scientific literature, we will refer to it as de jure model. In recent years, many different approaches aiming at representing clinical guidelines through formal models have been proposed, e.g., in the context of simulations for medical decision-support systems [10, 33, 34]. One of the crucial points of these representations concerns how to deal with the variability characterizing healthcare processes: as reported in [10] and motivated in Section 1, we use a declarative process modeling language to represent healthcare processes based on guidelines statements. As shown in Figure 1, the actual behavior of a medical treatment process can be analyzed using the information recorded by process oriented IT Systems recently introduced in the healthcare sector [1]. These 5 systems typically log the activities performed during the process execution in an event log. An event log contains information about activities and cases (process instances); it can also contain time information (timestamps) and indications about the actors involved in a specific case. Process mining techniques [7] use event logs to investigate the actual behavior of a process. Model repair, conformance checking, performance analysis and (unsupervised) process discovery, all belong to this family of techniques. 3.2. Model Repair and Cross Validation Medical treatment processes have some basic characteristics (as stated in the clinical guidelines) but also some characteristics that can differently be interpreted and handled according to the background and the experience of the end-users as discussed in Section 2. The model repair phase aims at revising some aspects of the de jure model based on information derived from the actual execution of the process as recorded in the logs. Of course, these modifications must be justified from a medical point of view and still be consistent with the intent of the guidelines. The model derived from the repair procedure, called de facto model, represents how the process should be executed in the clinical practice, according to both the guideline and everyday reality. Once the de facto model has been created, conformance checking techniques [7] can be applied to measure the adherence of future executions of the process with respect to this model. As shown in Figure 1 the event data are split into a large training set for repairing the model and a smaller test set for validating the result. The de facto model is confronted with process executions that have not been used for model repair. 3.3. Using the Model for Detecting Deviations The declarative process model based on the guidelines and event data is subsequently used for contin- uously monitoring deviations. Conformance checking techniques make it possible to detect violations in the process executions. Starting from the identified violations, these techniques also define how the process should be executed to perfectly fit the modeled behavior (using so-called trace alignments [35]). Confor- mance checking can be applied both in offline and in online modality. The offline analysis is performed post mortem for process instances that have completed. It is used to verify if cases comply with the de facto model and to determine how processes can be improved. For example, whenever a clinical trial is performed to test the effects of a particular medicine or treatment, the validation of the results requires to check whether the rules provided for the experimentation have been respected. Online conformance checking can be used during the execution of processes that need to be continuously monitored, like in the context of critical care. For example, within an intensive care unit, the detection of deviations may be crucial to allow for interventions when needed [36]. In this study, in particular, we focused on the offline analysis. The methodology described in Figure 1 is supported by various ProM plug-ins as described in the next section. 6 4. Operationalization This section describes the tools that have been used to perform each phase of the methodology proposed in Section 3. A combination of Declare3 and ProM 4 is used. Declare is used for the manipulation and enactment for declarative process models. ProM is a “pluggable” open-source framework for process mining. ProM can be used to discover declarative process models from event logs, check the conformance of such models, and enhance declarative models. 4.1. Model Design To perform the model design, we use Declare. Declare is an LTL-based declarative language [37, 38] supported by a toolset that includes a designer, a workflow engine, a worklist handler, and various analysis tools [39]. The designer can be used to support this study, since it allows us to obtain a formal representation of clinical processes by designing the activities and the rules characterizing the process behavior. Declare is based on an extensible set of templates, i.e., abstract entities defining parameterized classes of rules specified through Linear Temporal Logic (LTL) with semantics based on finite traces5 and each one equipped with its own graphical representation. Templates can be classified according to four groups: ex- istence, relation, negative relation and choice. Existence templates involve only one activity and define the allowed cardinality or position of an activity in a process instance. Relation/negative relation templates de- fine a dependency/negative dependency between two activities. Choice templates define alternative relations between activities. Constraints are concrete instantiations of templates and inherit the graphical represen- tation and LTL semantics from the corresponding templates [37]. A Declare model is a set of constraints that should hold in conjunction during the process execution. Figure 2 shows an example of Declare model. The model describes a process for gastric cancer surgical treatments. The first hospital admission requires the registration of the patient’s data when he/she is admit- ted to the hospital. Usually, this activity precedes all diagnostic/therapeutic treatments. The preoperative screening, according to the World Health Organization Guidelines for Safe Surgery, 6 is usually performed before any surgical treatment in order to assess whether the patient’s conditions are good enough for the surgery to be performed and to estimate potential risks. As far as the surgical technique is concerned, nowa- days, the gastric resection for malignant diseases can be performed by using either a laparoscopic surgery or a traditional open approach [40]. Furthermore, according to the clinical guidelines, in both cases a nursing period is needed to monitor the patient after the operation. 3http://www.win.tue.nl/declare/ 4http://www.promtools.org 5For compactness, we simply write LTL to refer to LTL on finite traces. 6http://apps.who.int/iris/bitstream/10665/44185/1/9789241598552_eng.pdf 7 Figure 2: An example of a Declare model modeling the surgical treatment of gastric cancer. The Declare model for this process involves five activities: First Hospital Admission, Preoperative Screen- ing, Open Gastrectomy, Laparoscopic Gastrectomy and Nursing, and seven constraints: three derived from the precedence template, one from the not succession template, one from the exclusive choice template and two from the response template. A precedence constraint between two activities A and B indicates that B can only be executed after that A has been executed at least once. In the model shown in Figure 2, a precedence constraint is used to indicate that activity Preoperative Screening can be executed only after that First Hospital Admission has occurred before. Also, precedence constraints are used to specify that the surgical treatment, either performed by using the open or the laparoscopic technique, can occur only if the Preoperative Screening has been performed before. A not succession constraint between two activities A and B indicates that whenever A occurs, B cannot follow. In the model shown in Figure 2, a not succession constraint specifies that after Preoperative Screening, First Hospital Admission cannot occur again, since the patient has already been registered through a univocal ID. An exclusive choice constraint between two activities A and B indicates that one of them must be ex- ecuted, but not both. In the model shown in Figure 2, an exclusive choice constraint indicates that the surgical intervention can be executed either by using the Laparoscopic Gastrectomy or the Open Gastrec- tomy; these techniques cannot be both applied to the same patient (unless emergent conditions require the conversion from the laparoscopic to the open approach [40]). A response constraint between two activities A and B states that B has to be executed at least once after that A has been executed. In the model shown in Figure 2, response constraints are used to indicate that both activities Laparoscopic Gastrectomy and Open Gastrectomy must be followed by a Nursing period. Referring to the example above, Table 1 shows the semantics for each constraint involved in the model. Each constraint is defined in terms of an LTL formula (here, First Hospital Admission is indicated as FHA, Preoperative Screening as PS, Laparoscopic Gastrectomy as LG, Open Gastrectomy as OG, Nursing as N). �(LG ⇒ ♦N) is the LTL constraint stating that every occurrence of LG should eventually be followed by 8 Table 1: Semantics of the Declare constraints from Figure 2 and their “activating” activities (� means “always”, ♦ means “eventually”, X t Y means “X should hold until Y ” (strong until). Constraint LTL semantics Activation response(LG, N) �(LG ⇒ ♦N) LG response(OG, N) �(OG ⇒ ♦N) OG precedence(F HA, P S) (¬P S t F HA) ∨ �(¬P S) P S precedence(P S, LG) (¬LG t P S) ∨ �(¬LG) LG precedence(P S, OG) (¬OG t P S) ∨ �(¬OG) OG not succession(F HA, P S) �(F HA ⇒ ¬(♦P S)) F HA, P S N. (¬PS tFHA) ∨�(¬PS) models that PS should be preceded by FHA. �(FHA ⇒¬(♦PS)) specifies that FHA cannot be eventually followed by PS. Table 1 also shows the activities activating a constraint. Note that, for example, a response constraint between two activities X and Y is satisfied in a trivial way in a process instance if X never occurs. In this case, we say that the constraint is vacuously satisfied [41]. In [42], the authors introduce the notion of behavioral vacuity detection according to which a constraint is non-vacuously satisfied in a process instance when it is activated in that process instance. An activation of a constraint in a process instance is an event whose occurrence imposes, because of that constraint, some obligations on other events in the same process instance. For example, if we consider the model shown in Figure 2, the occurrence of activity Laparoscopic Gastrectomy is an activation for the response between Laparoscopic Gastrectomy and Nursing, because the execution of Laparoscopic Gastrectomy forces Nursing to be also executed eventually. An activation of a constraint can be a fulfillment or a violation for that constraint. For example, the execution of Laparoscopic Gastrectomy is a fulfillment for the response between Laparoscopic Gastrectomy and Nursing if Laparoscopic Gastrectomy is followed by Nursing. The execution of Laparoscopic Gastrectomy is a violation for the response between Laparoscopic Gastrectomy and Nursing if Laparoscopic Gastrectomy is not followed by Nursing. Table 1 also shows activities that activate each constraint listed in the table. The example presented in Figure 2 shows that, using Declare, it is possible to design formal models that are easily understandable due to its simple graphical representation. Moreover, since Declare models are based on formal LTL semantics, they are verifiable and executable [37]. If events referring to activity executions are stored in an event log, it is possible to apply process mining techniques [7] for further analysis on the process. Table 2 shows an example of event log related to the surgical treatment of gastric cancer. Each line refers to the occurrence of an event with a unique id, a corresponding activity, a resource, a cost, and a timestamp. Events are grouped per process instance, where each instance represents an individual patient. An event log can be used to apply three types process mining techniques like process discovery, model repair and conformance checking. Each one of these techniques is supported by a ProM plug-in as shown below. 9 Table 2: Fragment of event log including for each case: case ID, event ID, timestamp, activity, resource, cost. Case ID Event ID Timestamp Activity Resource Cost 1 7785621 30-11-2011:08.27 First Hospital Admission Carol 90 1 7785624 2-12-2011:13.24 Preoperative Screening Susanne 350 1 7785625 4-12-2011:8.30 Laparoscopic Gastrectomy Andrew 500 1 7785631 4-12-2011:13.30 Nursing Paul 250 2 7785631 1-12-2011:11.00 Preoperative Screening Giuseppe 350 2 7785634 2-12-2011:15.28 Laparoscopic Gastrectomy Simon 500 2 7785638 2-12-2011:16.35 Nursing Clare 250 2 7785640 3-12-2011:13.00 Laparoscopic Gastrectomy Paul 500 2 7785641 3-12-2011:15.00 Nursing Andrew 250 2 7785661 4-12-2011:9.00 First Hospital Admission Victor 90 3 7785654 7-12-2011:10.00 First Hospital Admission Jane 90 3 7785624 8-12-2011:13.24 Laparoscopic Gastrectomy Giulia 500 3 7785625 9-12-2011:16.35 Nursing Paul 250 4 7785640 6-12-2011:14.00 First Hospital Admission Gianluca 90 4 7785667 8-12-2011:13.24 Preoperative Screening Robert 350 4 7785671 10-12-2011:16.35 Preoperative Screening Giuseppe 350 4 7785685 13-12-2011:11.00 Laparoscopic Gastrectomy Simon 500 4 7785698 13-12-2011:16.00 First Hospital Admission Jane 90 5 7785701 7-12-2011:15.00 First Hospital Admission Carol 90 5 7785711 9-12-2011:7.30 Preoperative Screening Susanne 350 5 7785723 13-12-2011:11.00 Laparoscopic Gastrectomy Simon 500 5 7785728 13-12-2011:13.50 Nursing Clare 250 5 7785732 13-12-2011:19.50 Nursing Vivianne 250 6 7785701 7-12-2011:15.00 First Hospital Admission Gianluca 90 6 7785711 9-12-2011:7.30 Preoperative Screening Susanne 350 6 7785723 13-12-2011:11.00 Laparoscopic Gastrectomy Andrew 500 7 7785744 11-12-2011:7.00 First Hospital Admission Jane 90 7 7785754 12-12-2011:9.30 Preoperative Screening Susanne 350 7 7785757 13-12-2011:16.00 Laparoscopic Gastrectomy Andrew 500 7 7785761 13-12-2011:18.50 Nursing Vivianne 250 7 7785763 14-12-2011:8.30 Laparoscopic Gastrectomy Gabriel 500 8 7785765 11-12-2011:7.00 Preoperative Screening Gianluca 350 8 7785766 12-12-2011:9.00 Preoperative Screening Susanne 350 8 7785768 13-12-2011:7.40 Preoperative Screening Gianluca 350 8 7785771 14-12-2011:9.30 Laparoscopic Gastrectomy Giulia 500 8 7785777 14-12-2011:16.00 Nursing Viviana 250 9 7785785 16-12-2011:12.00 Preoperative Screening Robert 350 9 7785795 17-12-2011:9.00 Preoperative Screening Susanne 350 9 7785797 19-12-2011:9.30 Laparoscopic Gastrectomy Andrew 500 9 7785891 19-12-2011:18.50 First Hospital Admission Carol 90 9 7785893 20-12-2011:18.00 Nursing Paul 250 10 7786185 19-12-2011:13.00 First Hospital Admission Jane 90 10 7786189 20-12-2011:9.30 Preoperative Screening Robert 350 10 7786193 22-12-2011:8.50 Laparoscopic Gastrectomy Andrew 500 10 7786198 22-12-2011:10.30 Nursing Clare 250 10 7786223 23-12-2011:14.30 Nursing Vivian 250 4.2. Model Repair In the model repair phase the de jure model is adjusted and repaired based on the information extracted from an event log. The result is a de facto model, which represents, in fact, how the process should be executed in the daily clinical practice. To repair the model, the Declare Miner plug-in in ProM can be used. It takes as input a Declare model and an event log and produces a new Declare model that better reflects how the process is executed in the reality [7]. Model repair can be performed through the Declare Miner. 10 Three possible modalities can be used to repair the model using event data: 1. The first modality generates a Declare model that includes constraints defined starting from only those templates that have been used in the input Declare model but over all the activities contained in the event log. 2. The second modality generates a Declare model that includes constraints defined starting from only those templates and activities that have been used in the input Declare model. 3. The third modality replaces a constraint in the input Declare model with a stronger constraint if this constraint holds in the event log (for example, a response constraint between A and B could be replaced by a chain response constraint indicating that B must immediately follow A). Figure 3: Surgical treatment of gastric cancer: repaired model. A minimum support can be specified as input of the repair algorithm [43]. Through this parameter, it is possible to leave out those constraints that are rarely activated in the log (i.e., constraints with a low support). Using different combinations of configuration options, the input Declare model can be modified by removing or changing existing constraints or adding new constraints, so that several variants of the original model can be generated. To select the most appropriate one, doctors as expert in the specific field should be consulted. Consider the Declare model shown in Figure 2 and the event log in Table 2. The Declare model shows that, according to the clinical guidelines, only one of activities Laparoscopic Gastrectomy and Open Gastrectomy should be applied for each patient. However, in the event log in Table 2, we can observe that Open Gastrectomy is never executed. Therefore, the model repair algorithm would generate a new Declare model in which activity Open Gastrectomy and all the constraints connected to it are removed, as shown in Figure 3. Indeed, these constraints are trivially satisfied, i.e., they are never really activated being the frequency of open gastrectomy equal to zero. This result could be justified with the fact that, due to the specific clinical environment (e.g., doctors’ expertise, type of patients treated, medical instruments available), the doctors prefer to apply the laparoscopic approach. Note that, if in a different department the open technique is preferred, the same Declare model would be repaired by removing activity Laparoscopic Gastrectomy. These differences are related to particular aspects of the real clinical environment and can be included in the de 11 facto models, since they do not violate the clinical guidelines. The de facto models can be used to perform further investigations through other process mining techniques as described below. 4.3. Conformance Checking Once the de facto model has been identified, conformance checking techniques [7] can be applied to investigate the deviations between the modeled behavior and the process execution [44]. Conformance checking can be applied both in an offline mode (i.e., when the process has been terminated) and in an online mode (i.e., when the process is still running). In this paper, we focus on the offline analysis of constraints. Two ProM plug-ins, the Declare Analyzer and the Declare Checker, provide concrete tool support for offline conformance checking. The Declare Analyzer takes a Declare model and an event log as input and pinpoints the deviations between the actual behavior as recorded in the event log, and the admissible one as described in the Declare model. In particular, it shows activations, fulfillments and violations of each constraint involved in the model. Note that when a process instance is compliant with respect to a constraint, every activation of that Figure 4: Surgical treatment of gastric cancer: the response constraint between Laparoscopic Gastrectomy and Nursing is fulfilled twice for case 2. Figure 5: Surgical treatment of gastric cancer: the response constraint between Laparoscopic Gastrectomy and Nursing is fulfilled once for case 1. 12 Figure 6: Surgical treatment of gastric cancer: the response constraint between Laparoscopic Gastrectomy and Nursing is violated for case 7. constraint leads to a fulfillment. Consider the model shown in Figure 2 and the event log in Table 2. In case 2 of the event log, the response constraint between Laparoscopic Gastrectomy and Nursing is activated and fulfilled twice (see Figure 4), whereas, in case 1, the same constraint is activated and fulfilled only once (see Figure 5). When a process instance is non-compliant with respect to a constraint, an activation of a constraint can lead to a fulfillment but also to a violation (and at least one activation leads to a violation). See, for example, the response constraint between Laparoscopic Gastrectomy and Nursing in the model in Figure 2 and case 7 in the event log in Table 2 (Figure 6). In this case, the constraint is activated twice, but the first activation leads to a fulfillment (eventually Nursing occurs) and the second activation leads to a violation (Nursing does not occur eventually). Another example of violation can be found in the event log if we consider the precedence constraint between First Hospital Admission and Preoperative Screening. Indeed, for cases 2, 8, and 9 we can find respectively one, three and two violations of this precedence constraint. Since each process instance is referred to an individual patient treatment process, using the information provided by the Declare Analyzer at the case level, it is possible to investigate the deviations based on the characteristic of the specific patient. For example, when First Hospital Admission is missing in a process instance, it is possible that the patient has already been registered in a different department of the same hospital for a different pathology. For each constraint in the de facto model, the Declare Analyzer also provides the total number of fulfillments and violations on the entire log. In the considered event log, for example, for the response constraint between Laparoscopic Gastrectomy and Nursing, there are 9 fulfillments and 3 violations. As already mentioned, the conformance of an event log with respect to a de facto model can also be checked using the Declare Checker. In particular, the Declare Checker estimates the fitness [7] between the model and the log, and, also, shows how non-compliant process instances should be modified to be aligned 13 with the desired behavior, leading to the identification of reparative actions. The Declare Checker is composed of two different plug-ins: the Declare Replayer and the Declare Diag- noser. The Declare Replayer takes as input a Declare model and an event log and measures for each process instance in the log the fitness of the process instance with respect to the model in a range from 0 (worst fitness) to 1 (best fitness) [35, 44]. If, for some process instance, the corresponding fitness value is less than 1, an alignment between the process instance and the model is created to show how the process instance should be modified to perfectly fit the model. To establish an alignment between a process model and an event log we need to relate “moves” in the log to “moves” in the model. In case of deviations, there are moves in the log cannot be mimicked by the model or vice versa. We explicitly denote a “no move” by �. Consider, for example, case 9 of the event log in Table 2 and the model in Figure 2. To align this process instance with the model, activity First Hospital Admission should be moved to the first position. Table 3 shows the alignment. The first line denotes a move on model only (see �) due to the missing First Hospital Admission event before Preoperative Screening. The fifth line denotes a move on log only: the � symbol in the last column shows that First Hospital Admission cannot follow Preoperative Screening. Starting from the results derived from the Declare Replayer, the Declare Diagnoser generates a map to give a helicopter view of the discrepancies between the log and the model. In particular, constraints and activities in the reference Declare model are annotated with numbers (ranging from 0 to 1) representing their degree of conformance [44]. The degree of conformance of an activity in the model is an indication of the number of times the activity has been removed or added to compute the alignments between all the process instances in the log and the reference model [35]. The degree of conformance of a constraint in the model is an indication of the number of times the constraint has been violated in the event log. In the map, activities and constrains of the reference model are colored with colors ranging from red to green to indicate a degree of conformance varying from 0 (red) to 1 (green). 5. Case Study In this section, we describe the outcomes of a case study in order to demonstrate the applicability of our approach in practice. In particular, we apply our methodology for processes handled in the urology Table 3: Example of a trace alignment: There is one move on model and one move on log due to the wrong position of the First Hospital Admission event in case 9. Case ID Event ID Timestamp Activity in Log Activity in Model 9 - - � First Hospital Admission 9 7785785 16-12-2011:12.00 Preoperative Screening Preoperative Screening 9 7785795 17-12-2011:9.00 Preoperative Screening Preoperative Screening 9 7785797 19-12-2011:9.30 Laparoscopic Gastrectomy Laparoscopic Gastrectomy 9 7785891 19-12-2011:18.50 First Hospital Admission � 9 7785893 20-12-2011:18.00 Nursing Nursing 14 department of the Isala hospital. The Isala hospital consists of multiple clinics and is the largest top-clinical hospital in the Netherlands. Isala aims at providing efficient and safe care to their patients. Therefore, they are looking for innovative ways to analyze and improve their care processes. Within Isala, the urology department is involved with diagnosing and treating patients suffering from illnesses on the male and female urinary tract, and on the male reproductive system. We have applied our methodology to the process for diagnosis and treatment of a particular urogenital disease: the cryptorchidism. The first step of our study focused on the case study investigation, in order to find the information required to perform the model design. We carried out a literature analysis aimed at finding the main scientific evidences about the process: we examined some of the main articles related to the topic stored in the Database PubMed 7, as [45, 46, 47, 48, 49, 50, 51, 52] and also the Guidelines of the European Association of Urology (EAU).8 Based on this information, we designed the corresponding Declare model. Then, we performed conformance checking in offline mode –as described in the previous section, through the Declare Analyzer and the Declare Checker plug-ins in ProM– to verify how the de jure model based on the literature instructions complies with the actual execution of the process. Information about the actual execution of the process came from event logs registered in the urology department of the Isala hospital. We performed the model repair by using the Declare Miner plug-in in ProM. Using the repaired the model, we applied again the conformance checking to verify whether the repaired model has a better fitness with respect to the log. The steps of the study and the results obtained are described in the following subsections. 5.1. Case Study Investigation Cryptorchidism is the absence of one or both testis in the scrotum caused by a deficient or irregular testicular descent; it is one of the most common disorders of childhood, affecting 0.8 – 1.8 percent of infants at 1 year of age, 3 percent of full-term newborns, and 21 percent of premature babies [45]. It has been assessed that it can also lead to an increasing risk of infertility and cancer [46]. Thus, early diagnosis and treatment of undescended testis are needed to preserve fertility and to prevent testicular malignancy. According to the European guidelines, although 15 – 20 of retained testes descend during hormonal treatment, the intake of such substances may be harmful to future spermatogenesis by increasing the apoptosis of germ cells; it was also assessed that the descended testes, when pharmacologically treated, often reascend later. Thus, the hormonal treatment is no longer recommended. On the other hand, the success rate of surgical treatment for undescended testes, named ’Orchidopexy’ is 70 – 90 percent [47]. Also Gapany et Al. [50] state that the surgery is the cornerstone of treatment. To surgically treat the cryptorchidism, either the open or the laparoscopic approach can be applied, even if it 7http://www.ncbi.nlm.nih.gov/pubmed 8http://www.urology-textbook.com 15 is still open the debate related to which one of them should be preferred [48]. The laparoscopic approach is more used because, according to Gapany et al. [50], to date, no imaging technique for non-palpable testes has proven to be superior to laparoscopy, which allows both the diagnosis and the surgical treatment if there is a pathology affecting testes. Whenever the operation cannot be completed through the laparoscopic approach, a switching to the open technique can be applied. This decision can be made by surgeons either before or during the actual operation, in those cases in which it is retained opportune for patients safety, for example in case of inability to visualize structures [53, 54]. Generally, as reported in the guidelines, a staged orchidopexy (Fowler-Stephenson) procedure can be performed, using the open surgery, or the mini-invasive approach as laparoscopy or microsurgery. A biopsy at the time of orchidopexy can be performed to reveal intratubular germ cell neoplasia of unclassified type, which can be removed thereby preventing development of a malignant tumor. 9 An association between cryptorchidism and testis cancer has been known for more than one century [55]. As reported by [52] and [51], when unilateral cryptorchidism with postpubertal male is identified through diagnostic procedures, the orchiectomy – a surgical procedure to remove a testicle and the full spermatic cord through an incision in the lower lateral abdomen – should be performed instead of the orchidopexy. 5.2. Declare Model Design The model of the process has been designed through the Declare Designer, mentioned in Section 4. Information derived from the European guidelines and from some scientific articles found in PubMed [45, 46, 47, 48, 49, 50, 51, 52], has been used to define activities and constraints characterizing the process. It has been payed attention on considering, among the activities characterizing the process, only those that appeared at least once in the event log extracted from the information system of the urology department of the Isala hospital in Zwolle. The model is shown in Figure 7. The model is composed of 10 activities and 13 constraints using 4 different templates. The LTL semantics of some Declare templates have been provided in Section 4. Hereby, the particular meaning of each constraint in the context of the process of interest will be illustrated. • Precedence between Radiological Examination and First Hospital Admission and between Lab Tests and First Hospital Admission: the hospital admission should be usually preceded by the examinations required to assess the presence of the pathology. Among these, radiological examination and lab tests are the most common ones needed for cryptorchidism. • Precedence between First Hospital Admission and Preoperative Screening: to undergo the preoperative screening, the patient must be first admitted to the hospital. 9http://www.uroweb.org/fileadmin/guidelines/EAU_GO_Manual_November_28th_2012.pdf 16 Figure 7: The cryptorchidism treatment according to scientific literature: representation through a Declare model. • Precedence between Preoperative Screening and Open Surgery, Preoperative Screening and Orchi- dopexy, Preoperative Screening and Inguinal Hernia Treatment: according to the World Health Or- ganization clinical guidelines for surgery, every surgical treatment cannot be executed without pre- liminary examinations required in order to prevent risks for patients. It is important to outline that, according to the Isala Hospital denomination of the activities, the word orchidopexy is refer to the testes treatment carried out through the laparoscopic approach, while the open surgery indicates those treatments performed through the open one. • Response between Orchidopexy and Nursing, Inguinal Hernia Treatment and Nursing and between Open Surgery and Nursing: according to the World Health Organization Guidelines mentioned above, a nursing period must follow every surgical treatment, in order to monitor the patient conditions after the operation. • Responded Existence between Histologic Examination and Orchidopexy: according to the European Guidelines of Urology, 10 the biopsy with the histologic examination can be performed when the orchidopexy is performed whenever it is retained recommendable to analyze the nature of the tissue. Therefore, if the histologic examination occurs, the orchidopexy should occur as well in the process execution. • Succession between Nursing and Follow Up: after the nursing period, patients must undergo the follow up check in order to define whether the treatment has been effective and the patient can be discharged, 10http://www.uroweb.org/fileadmin/guidelines/EAU_GO_Manual_November_28th_2012.pdf 17 or further examinations and interventions are required. Also, the follow up is usually preceded by a nursing period. • Responded Existence between Inguinal Hernia Treatment and Orchidopexy: according to the European Guidelines of Urology, the inguinal hernia treatment can be performed at the time of the orchidopexy; therefore, if the inguinal hernia treatment is performed, the orchidopexy is also performed. • Exclusive Choice between Orchidopexy and Open Surgery: as explained above, to surgically treat the cryptorchidism, it is possible to choose either the open or the laparoscopic approach, but not both [48]. 5.3. Conformance Checking (First Iteration) Once the de jure model has been designed based on the information derived from the scientific literature, conformance checking has been applied in order to measure the adherence of such model with the actual behavior of the process. To do so, part of event log was extracted from the information system of the Isala Hospital (the training set mentioned in Figure 1). The selected event log was analyzed in conjunction with the Declare model. For this we used the Declare Checker and the Declare Analyzer plug-ins of ProM (see Section 4). The results obtained with the Declare Replayer and the Declare Diagnoser are shown in Figures 8 and 9. Figure 8: Conformance checking between the de jure model and a fragment of event log: the Declare Replayer. As shown in Figure 8, the average fitness between the de jure model and the actual behavior of the process is approx. 0.831. Figure 9 shows that the most severe conformance problems between the theoretical model and actual practice are related to the diagnostic activities preceding the first hospital admission. This 18 Figure 9: Conformance checking between the de jure model and a fragment of event log: the Declare Diagnoser. can be explained by considering that patients can undergo diagnostic examinations outside the hospital and then they can be admitted to the hospital if these examinations highlight pathologies as cryptorchidism that may require surgical treatments. A low degree of conformance is also observed with reference to the open surgery that is performed only four times in the entire event log. Through the Declare Analyzer plug-in it has been possible to gather a global overview about fulfillments, violations and conflicts for each constraint in the model. The results are outlined in Table 4 (here, First Hos- pital Admission is indicated as FHA, Preoperative Screening as PS, Laparoscopic Gastrectomy as LG, Open Gastrectomy as OG, Nursing as N, Orchidopexy as OR, Inguinal Hernia Treatment as IHT, Radiological Examination as RE, Lab Tests as LB, Follow Up as FU). The Declare Analyzer confirmed that, in most of cases, the First Hospital Admission is not preceded by Lab Tests and Radiological Examinations (a screenshot of the Declare Analyzer showing this violation in one of the traces in the log is shown in Figure 8. Indeed, when patients have already undergone the exams required for the diagnosis - and the surgical treatment has been recommended from the doctor - the repetition of the diagnostic examination is not required at the time of the admission to the Hospital. Therefore, in these cases, the process starts with the First Hospital Admission, and the Lab Tests and Radiological Examination are not present. It is also interesting to observe that there are some traces in which the precedence between First Hospital Admission and Preoperative Screening is not respected. To investigate this anomaly, the doctors of the Isala Hospital were consulted. The case IDs (i.e., the patients) for which such violations were observed were discussed. According to the doctors, the observed violations correspond to patients admitted to different departments of the hospital and then transferred to the urology department because urinary pathologies such as cryptorchidism were suspected. The same considerations hold for the cases in which the precedence between Preoperative Screening and Orchidopexy is violated. It 19 Table 4: The cryptorchidism treatment: conformance checking between the de jure model and the event log through the Declare Analyzer. Constraint Activations Violations Fulfillments Precedence [PS - OR] 197 7 190 Precedence [PS - OS] 4 0 4 Precedence [PS - IHT] 7 0 7 Precedence [RE - FHA] 172 162 10 Precedence [FHA - PS] 264 86 178 Precedence [LT - FHA] 172 168 4 Responded Existence [HE, OR] 3 2 1 Responded Existence [IHT, OR] 7 6 1 Exclusive Choice [OR-OS] 201 0 201 Succession [N-FU] 667 150 517 Response [OS-N] 4 0 4 Response [OR-N] 197 9 188 Response [IHT-N] 7 1 6 Figure 10: Conformance checking between the de jure model and a fragment of event log: the Declare Analyzer. is indeed possible that patients have already undergone different surgical treatments in other departments of the hospital and the data of the preoperative examinations were transferred to the urology department. In these cases, the orchidopexy was directly performed, without repeating the exams. Patients can move from one department to another of the hospital. This also explains the violations of the succession constraint between nursing and follow up and between orchidopexy and nursing, since, after the treatment within the urology department, patients can be moved to other departments if other pathologies arise, and monitored 20 there. The obtained results also show that the responded existence between Orchidopexy and Inguinal Hernia Treatment is violated in some cases. This can be explained by considering that, based to the clinical guidelines, the inguinal hernia treatment can be performed at the time of the orchidopexy. 11 According to the doctors there are cases in which inguinal hernia treatment is performed either in association with other treatments or alone, without performing an orchidopexy. 5.4. Model Repair In order to obtain a model that takes into account also the actual aspects of the process execution, the de jure Declare model has been matched with the event log through the function Repair a Declare Model of the Declare Miner plug-in in ProM, described in Section 4. Different combinations of configuration options have been set to generate different models. All these models have been shown to the doctors in order to choose the one that better reflects the cryptorchidism treatment process in the Isala Hospital. The chosen repaired model is shown in Figure 11. By comparing this model with the original one, important insights can be derived. First, the repaired model does not show the diagnostic activities performed before the First Hospital Admission, i.e., Lab Tests and Radiological Examination. This can be explained by considering that, as already mentioned before, there are several cases in which the patients undergo diagnostic examinations outside the hospital: therefore, if the diagnosis is established, it is not mandatory to repeat the examinations when the patients are admitted to the hospital, and they are directly prepared for surgical treatments. Moreover, the precedence constraint indicating that Preoperative Screening must be preceded by First Hospital Admission, is removed. This can be due to the fact that, when patients are admitted to different departments of the hospital, and then moved to the urology department for the testis treatment, they undergo directly the screening for the surgical treatment, without a new registration. However, the precedence is replaced by a response. This indicates that in the actual process, if the First Hospital Admission occurs, the preoperative screening must happen afterwards. For what concerns the surgical activities, the responded existence between Inguinal Hernia Treatment and Orchidopexy is removed. This can be justified by considering that hernia treatment can either be performed in the actual practice or not according to the specificity of the cases treated. Such a situation represents a typical example of those aspects of the clinical guidelines that are strongly dependent on the specific cases to be treated: in the considered event log, indeed, it is possible that most of the patients were not affected from inguinal hernia, so that the treatment was not performed. The same considerations hold for activity Histologic Examination, removed in the repaired model. Since the Open Surgery is performed only in three traces in the event log, it does not appear in the repaired model. This is in accordance to [50], in which it is stated that the laparoscopic approach represents 11http://www.uroweb.org/fileadmin/guidelines/EAU_GO_Manual_November_28th_2012.pdf 21 Figure 11: The cryptorchidism treatment: the repaired model. the best method to treat cryptorchidism, since, with this approach, the zone to be treated becomes more visible; in addition, nowadays, the mini-invasive approaches are likely to replace the traditional open surgery in several medical sectors. Nonetheless, there are still particular cases, less frequent, in which sometime the open approach may be preferred, or switching from the laparoscopy to the open surgery becomes necessary [53]. In the repaired model it is also shown that the precedence relation between Preoperative Screening and Orchidopexy is respected, in line with the clinical guidelines statements. There is also a response that has been added between them, indicating that usually, if preoperative screening is carried out, the orchidopexy occurs afterwards. Furthermore, the response relation between Orchidopexy and Nursing appears to be reinforced in the actual practice: in the repaired model, indeed, it is replaced by a succession constraint, which indicates that not only after the surgery the nursing must be provided, as stated by the clinical guidelines, but also that in the actual process Nursing is always preceded by Orchidopexy. Finally, a response constraint between Nursing and Follow Up is shown instead of succession: it indicates that after the nursing period, the follow up is needed for patients, but it is not mandatory that the nursing occurs before the follow up. A responded existence links Follow Up to Orchidopexy, meaning that, in the actual process, if Follow Up occurs, Orchidopexy occurs as well. 5.5. Conformance Checking (Second Iteration) In this step of the study, the repaired model has been matched with the actual process by applying conformance checking between this (de facto) model and the event log, with the aim of verifying whether it really better fits the reality with respect to the de jure model. As shown in Figure 12, the Declare Replayer indicates a fitness value between the repaired and the actual model of 0.884, which is higher, as expected, than the one obtained by applying conformance checking between the de jure model and the event log. 22 Figure 12: The cryptorchidism treatment: conformance checking between the repaired model and the event log (Declare Replayer). Figure 13: The cryptorchidism treatment: conformance checking between the repaired model and the event log (Declare Diagnoser). The results returned by the Declare Diagnoser are shown in Figure 13. It is possible to notice that all the activities are represented in green, indicating a good degree of conformance. The results obtained with the Declare Analyzer are represented in Table 5 (here, First Hospital Admission is indicated as FHA, Preoperative Screening as PS, Laparoscopic Gastrectomy as LG, Open Gastrectomy as OG, Nursing as N, Orchidopexy as OR, Follow Up as FU). It is interesting to notice that there are 14 cases in which Nursing is not followed by Follow Up and 43 cases in which the succession between Orchidopexy and Nursing is violated. According to the doctors’ opinion, these violations are due to the fact that there are cases in which, after the testis treatment, patients are moved to different departments. The same consideration holds for the precedence between Preoperative Screening and Orchidopexy, which was violated in 7 cases. 23 There are also cases in which the responded existence between Follow Up and Orchidopexy is not respected, indicating that Follow up is performed whereas Orchidopexy is not. Examples of violations detected from the Declare Analyzer are shown in Figure 14. Figure 14: Conformance checking between the repaired model and a fragment of event log: the Declare Analyzer. Table 5: The cryptorchidism treatment: conformance checking between the repaired and the event log through the Declare Analyzer. Constraint Activations Violations Fulfillments Succession [OR-N] 446 43 403 Precedence [PS - N] 513 48 465 Responded Existence [FU - OR] 418 80 338 Response [PS - OR] 264 58 206 Response [FHA - PS] 172 16 156 Precedence [PS - OR] 197 7 190 Response [N - FU] 249 14 235 In summary, the obtained results allow us to demonstrate that through the repair procedure, it is possible to obtain a model of the clinical process of interest that takes into account not only the guidelines statements, but also the actual process execution. Conformance checking, indeed, shows that the repaired model better complies with the actual process. 6. Conclusion In this paper we presented a novel methodology to analyze medical treatment processes thereby mediating between guidelines and reality. The methodology aims at: 24 1. discovering and pinpointing the deviations between the de jure model, i.e., the model that complies with existing medical protocols, and the actual behavior as observed in the event log extracted from the hospital’s system, and 2. creating a de facto model, i.e., the model that represent the right compromise between the de jure model and what practically observed. The methodology is made possible by recent developments in process mining. Process mining is gaining more and more interest since companies and institutions are interested to improve the processes that they are executing. Of course, hospitals also share the same interests since they want to treat patients efficiently while making them satisfied. The difference between health-care processes and more business-oriented processes is related to the fact that the former are characterized by being much more dynamic and turbulent. Often, similar patients need to be treated quite differently, and variability is the norm. Traditionally, process mining relies on the fact that models are modeled with procedural languages, such as BPMN. Procedural languages are very good when the process instances tends to be repetitive, whereas they fail miserably when processes exhibit a high degree of flexibility. As such, existing process-mining methodologies have problems when applied in the context of health-care processes, as discussed in Section 2. Our methodology is based on a declarative language, namely Declare, where models are simply sets of constraints. As discussed, this guarantees a higher degree of flexibility when defining the prescribed process’ behavior in a procedural way. As reported on in Section 5, the methodology has been validated through a real health-care case study in the Isala hospital. The results have shown to be quite insightful, thus being helpful for the process analysts to improve how patients are treated. We provided a way to measure the adherence of the actual process to the guidelines statements and propose a solution to improve it, by suggesting a model that could be the right compromise between theory and practice. In particular, we addressed this aspect by creating models in which both the de jure knowledge and the practice-based evidence (i.e., event logs) are taken into account in the clinical process design. To this aim, repair techniques have been developed to match the modeled behavior with logs in order to include in the model design also behaviors that are not mentioned in the guidelines but that are still allowed in practice. A typical example is related to the possibility to prescribe biopsy or not whenever the surgical treatment for the cryptorchidism is performed [46]. The clinical guidelines indeed state that a biopsy at the time of orchidopexy can reveal intratubular germ cell neoplasia of an unclassified type, which can be removed thus preventing the development of a malignant tumor.12 Surgeons, therefore, can decide to either require or not the biopsy based on what they observe during the operation. The same considerations hold for surgical hernia treatment. The clinical guidelines state that hernia treatment “can” be performed at the time of orchidopexy. This means that surgeons can as 12http://www.uroweb.org/fileadmin/guidelines/EAU_GO_Manual_November_28th_2012.pdf 25 well postpone this treatment based on the patient conditions. Based on the most frequent actual executions of the process, model repair helps to decide on the required trade-off. After the model repair, all deviations detected through the conformance checking can be monitored systematically without many “false alarms” due to contract between clinical practice and guidelines. The interaction with doctors and nurses also provided valuable input for future work. Doctors, nurses and process analysts are familiar with procedural modeling languages, even though these models tend to be large and barely readable. In the case study, we had to make the stakeholders acquainted with Declare. Although Declare has a simple graphical notation and succinct semantics, quite some time is required to fully grasp the notation. Therefore, it is good to investigate how to tailor the notation to the domain. The main input for our methodology is an event log that records the executions of process instances. Healthcare information systems are typically poorly integrated: many systems coexist within the same hos- pital, thus returning different event logs that cannot be easily merged and that often contain inconsistencies. As future work, we aim to come up with a technique to merge several event logs, containing inconsistencies. Some genetic approaches exist, such as [56], but they seem to not properly scale with logs of increasing size. Finally, we aim to apply the same technique in additional case studies in other hospitals. References [1] R. Lenz, M. Reichert, It support for healthcare processes - premises, challenges, perspectives, Data & Knowledge Engi- neering 61 (1) (2007) 39–58. [2] J. N. Lavis, E. J. Paulsen, A. D. Oxman, R. Moynihan, Evidence-informed health policy 2: Survey of organizations that support the use of research evidence, Implementation Science 3 (2008) 54. [3] W. Y. Lew, A. N. DeMaria, The divergence between clinical guidelines and practice, Journal of the American College of Cardiology 61 (1) (2013) 41–43. [4] D. M. Eddy, Evidence-based medicine: a unified approach, Health Affair 24 (1) (2005) 9–17. [5] M. C. Hay, T. S. Weisner, S. Subramanian, N. Duan, E. J. Niedzinski, R. L. Kravitz, Harnessing experience: exploring the gap between evidence-based medicine and clinical practice, Journal of evaluation in clinical practice 14 (2008) 707–713. [6] L. J. Cochrane, C. A. Olson, S. Murray, M. Dupuis, T. Tooman, S. Hayes, Gaps between knowing and doing: Understanding and assessing the barriers to optimal health care, Journal of Continuing Education in the Health Professions 27 (2) (2007) 94–102. [7] W. van der Aalst, Process Mining: Discovery, Conformance and Enhancement of Business Processes, Springer, 2011. [8] A. Rebuge, D. R. Ferreira, Business process analysis in healthcare environments: A methodology based on process mining, Information System 37 (2) (2012) 99–116. [9] K. Anyanwu, A. Sheth, J. Cardoso, J. Miller, K. Kochut, et al., Healthcare enterprise process development and integration, Journal of research and practice in information technology 35 (2) (2003) 83–98. [10] N. Mulyar, M. Pesic, W. van der Aalst, M. Peleg, Declarative and procedural approaches for modelling clinical guidelines: addressing flexibility issues, in: Business Process Management Workshops, Springer, 2008, pp. 335–346. [11] F. M. Maggi, A. J. Mooij, W. van der Aalst, User-guided discovery of declarative process models, in: Computational Intelligence and Data Mining (CIDM), 2011 IEEE Symposium on, IEEE, 2011, pp. 192–199. 26 [12] P. Bansal, S. Bertels, T. Ewart, P. MacConnachie, J. O’Brien, Bridging the research-practice gap, Academic of Management Perspective 26 (1) (2012) 73–92. [13] T. R. Dresselhaus, J. W. Peabody, M. Lee, M. M. Wang, J. Luck, Measuring compliance with preventive care guidelines, Journal of General Internal Medicine 15 (2000) 782–788. [14] N. E. Cahill, D. K. Heyland, Bridging the guideline practice gap in critical care nutrition: A review of guideline imple- mentation studies, Journal of Parenteral and Enteral Nutrition 34 (6) (2010) 653–659. [15] F. Hajjaj, M. Salek, M. Basra, A. Finlay, Nonclinical influences, beyond diagnosis and severity, on clinical decision making in dermatology: understanding the gap between guidelines and practice, British Journal of Dermatology 163 (2010) 789–799. [16] E. A. Newnham, A. C. Page, Bridging the gap between best evidence and best practice in mental health, Clinical Psy- chology Review 30 (1) (2010) 127–142. [17] J.-N. Cornu, O. Cussenot, F. Haab, B. Lukas, A widespread population study of actual medical management of lower urinary tract symptoms related to benign prostatic hyperplasia across europe and beyond official clinical guidelines, European Urology 58 (2010) 450–456. [18] A. Freeman, K. Sweeney, Why general practitioners do not implement evidence: qualitative study, BMJ 323 (2001) 1–5. [19] S. M. B. Bernheim, J. S. Ross, H. M. Krumholz, E. H. Bradley, Influence of patients’’ socioeconomic status on clinical management decisions: a qualitative study, Annals of Family Medicine 6 (2008) 53–59. [20] R. Grol, Has guideline development gone astray? yes, BMJ 340 (2010) 394–395. [21] I. D. Graham, J. Logan, M. B. Harrison, S. E. Straus, J. Tetroe, W. Caswell, N. Robinson, Lost in knowledge translation: Time for a map?, Journal of Continuing Education in the Health Professions 26 (1) (2006) 13–24. [22] J. N. Lavis, Research, public policymaking, and knowledge-translation processes: Canadian efforts to build bridges, Journal of Continuing Education in the Health Professions 26 (2006) 37–45. [23] R. Mahajan, Critical incident reporting and learning, Br J Anaesth 105 (2010) 69–75. [24] S. M. Berenholtz, P. J. Pronovost, Monitoring patient safety, Critical Care Clinics 23 (3) (2007) 659–673. [25] S. Bolsin, A. Patrick, M. Colson, B. Creatie, L. Freestone, New technology to enable personal monitoring and incident reporting can transform professional culture: the potential to favourably impact the future of health care, Journal of Evaluation in Clinical Practice 11 (2005) 499–506. [26] A. Petzold, N. Korner-Bitensky, A. Menon, Using the knowledge to action process model to incite clinical change, Journal of Continuing Education in the Health Professions 30 (3) (2010) 167–171. [27] D. Bates, M. Cohen, L. Leape, J. M. Overhage, M. M. Shabot, T. Sheridan, Reducing the frequency of errors in medicine using information technology, Journal of the American Medical Informatics Association 8 (4) (2001) 299–308. [28] R. Lenz, M. Reichert, It support for healthcare processes: Premises, challenges, perspectives, Data & Knowledge Engi- neering 61 (1) (2007) 39–58. [29] R. Fichman, R. Kohli, R. Krishnan, The role of information systems in healthcare: Current research and future trends, Information Systems Research 22 (3) (2011) 419–428. [30] R. Mans, M. Schonenberg, M. Song, W. van der Aalst, P. Bakker, Application of process mining in healthcare: A case study in a dutch hospital, in: A. Fred, J. Filipe, H. Gamboa (Eds.), Biomedical Engineering Systems and Technologies, Vol. 25 of Communications in Computer and Information Science, Springer, 2009, pp. 425–438. [31] R. Mans, H. Schonenberg, G. Leonardi, S. Panzarasa, A. Cavallini, S. Quaglini, W. van der Aalst, Process mining techniques: an application to stroke care (2008). [32] J. Weerdt, F. Caron, J. Vanthienen, B. Baesens, Getting a grasp on clinical pathway data: An approach based on process mining, in: Emerging Trends in Knowledge Discovery and Data Mining, Vol. 7769 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2013, pp. 22–35. 27 [33] M. A. Grando, D. W. Glasspool, J. Fox, Petri nets as a formalism for comparing expressiveness of workflow-based clinical guideline languages, in: Business Process Management Workshops, Springer, 2009, pp. 348–360. [34] A. Seyfang, B. Mart́ınez-Salvador, R. Serban, J. Wittenberg, S. Miksch, M. Marcos, A. Ten Teije, K. Rosenbrand, Maintaining formal models of living guidelines efficiently, Artificial Intelligence in Medicine (2007) 441–445. [35] W. van der Aalst, A. A., van Dongen B., Replaying History on Process Models for Conformance Checking and Performance Analysis, WIREs Data Mining and Knowledge Discovery 2 (2) (2012) 182–192. [36] Z. Tang, L. Weavind, J. Mazabob, E. J. Thomas, M. Y. L. Chu-Weininger, T. R. Johnson, Workflow in intensive care unit remote monitoring: A time-and-motion study, Critical care medicine 35 (9) (2007) 2057–2063. [37] W. van der Aalst, M. Pesic, H. Schonenberg, Declarative workflows: Balancing between flexibility and support, Computer Science-Research and Development 23 (2) (2009) 99–113. [38] M. Pesic, W. van der Aalst, A declarative approach for flexible business processes management, in: Business Process Management Workshops, Springer-Verlag, Berlin, Heidelberg, 2006, pp. 169–180. [39] M. Westergaard, F. M. Maggi, Declare: A tool suite for declarative workflow modeling and enactment, in: BPM (Demos), 2011. [40] J. Qiu, P. Pankaj, H. Jiang, Y. Zeng, H. Wu, Laparoscopy versus open distal gastrectomy for advanced gastric cancer: a systematic review and meta-analysis, Surgical Laparoscopy, Endoscopy & Percutaneous Techniques 23 (1) (2013) 1–7. [41] O. Kupferman, M. Y. Vardi, Vacuity detection in temporal model checking, International Journal on Software Tools for Technology Transfer (2003) 224–233. [42] A. Burattin, F. M. Maggi, W. van der Aalst, A. Sperduti, Techniques for a posteriori analysis of declarative processes, in: Enterprise Distributed Object Computing Conference (EDOC), 2012 IEEE 16th International, 2012, pp. 41–50. [43] F. M. Maggi, R. Bose, W. van der Aalst, Efficient discovery of understandable declarative process models from event logs, in: Advanced Information Systems Engineering, Springer, 2012, pp. 270–285. [44] M. de Leoni, F. M. Maggi, W. van der Aalst, Aligning event logs and declarative process models for conformance checking, in: Business Process Management, Vol. 7481 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2012, pp. 82–97. [45] M. Smolko, G. Kaplan, W. Brock, Location and fate of the nonpalpable testis in children, The Journal of Urology 129 (6) (1983) 1204–1206. [46] F. Ferro, A. Lais, L. Gonzalez-Serva, Benefits and afterthoughts of laparoscopy for the nonpalpable testis, The Journal of Urology 156 (6) (1996) 795–798. [47] P. Jones, Approaches to orchidopexy, British Journal of Urology 75 (6) (1995) 693–696. [48] J. Guo, Z. Liang, H. Zhang, C. Yang, J. Pu, H. Mei, L. Zheng, F. Zeng, Q. Tong, Laparoscopic versus open orchidopexy for non-palpable undescended testes in children: a systemic review and meta-analysis, Pediatric Surgery International 27 (9) (2011) 943–952. [49] F. T. Denes, F. J. Saito, F. A. Silva, A. M. Giron, M. Machado, M. Srougi, Laparoscopic diagnosis and treatment of nonpalpable testis, International braz j urol 34 (3) (2008) 329–335. [50] C. Gapanya, P. Freya, F. Cachatb, F. Gudinchetc, P. Jichlinskie, B.-J. Meyrata, P. Ramseyera, G. Theintzd, B. Burnandf, Management of cryptorchidism in children:guidelines, Risk 16 (33-34) (2008) 19. [51] E. Rogers, S. Teahan, G. Hugh, B. M. R., R. Grainger, T. McDermott, T. J. A., The role of orchiectomy in the management of postpubertal cryptorchidism, The Journal of Urology 159 (1998) 851–854. [52] E. Rogers, S. Teahan, H. Gallacher, M. R. Butler, R. Grainer, T. McDermott, J. A. Thornill, The role of orchiectomy in the management of postpubertal cryptorchidism, The Journal of urology 159 (3) (1998) 851–854. [53] C. O. Dominguez, Expertise in laparoscopic surgery: anticipation and affordances, Linking expertise and naturalistic decision making (2001) 287–301. 28 [54] M. J. Mathers, H. Sperling, H. R̈ı¿ 1 2 bben, S. Roth, The undescended testis: Diagnosis, treatment and long-term conse- quences, Dtsch Arztebl Int 106 (33) (2009) 527–532. [55] H. M. Wood, J. S. Elder, Cryptorchidism and testicular cancer: separating fact from fiction, The Journal of urology 181 (2) (2009) 452–461. [56] J. Claes, G. Poels, Merging computer log files for process mining: An artificial immune system technique, in: Business Process Management Workshops, Vol. 99 of Lecture Notes in Business Information Processing, Springer Berlin Heidelberg, 2012, pp. 99–110. 29 
work_loxphymenzbbnpykm2xiscsjpi	----	DESIGNING EXPERT SYS'lXMS IN A BUSINESS ENVIRONMENT James Clifford, Matthias Jarke and Henry C. Lucas, Jr. May 1985 Center for Research on Information Systems Computer Applications and Information Systems Area Graduate School of Business Administration New York University Working Paper Series CRIS #99 GBA f85-62 To appear in : L.F. Pau (ed.): Artificial Intelligence in Economics and Management, North-Holland 1985. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page i Table of Contents 1. INTRODUCTION 2. EXPERT SYSTEMS IN THE BUSINESS ENVIRONMENT 3. THE UNDERWRITING EXPERT SYSTEM 3.1. The Information-Gathering Decision 3.2. The Underwriting Decision 4. KNOWLEDGE ACQUISITION 4.1. General Issues in a Business Environment 4.2. Underwriting System in Particular 5. UNDERWRITING IN CONTEXT 6. RELATIONSHIP TO TRADITIONAL SYSTEMS ANALYSIS AND DESIGN 6.1. -ditional Systems Analysis and Design 6.2. Prototyping 6.3. Expert Systems Design 7. SUMMARY AND CONCLUSIONS Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 1 Abstract The integration of an ES into a business environment presents a different set of problems to the designer. First, it can be difficult to isolate and draw boundaries around the domain of the business problem. Second, there is often a need to generate a large number of expert decisions in a short period of time. That is, there can be many transactions requiring some expertise to process, such as applications for life insurance. Finally, there may exist a number of computerized transactions processing systems which interact with very large databases and there may be a need to integrate the ES with these existing systems. This paper discusses these general issues involved in developing expert systems for business applications, with particular examples drawn from the domain of insurance underwriting. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 2 1. INTRODUCTION An expert system (ES) is a computer application that provides decision support similar to that of a human expert in solving a problem. The systems analyst who builds an expert system is called a knowledge engineer; his or her role is to learn the expert's approach to problem solving. Usually the knowledge engineer can formulate the expert's techniques as a series of rules. Then, an Artificial Intelligence language like Prolog can be used to design a computer system that exhibits some of the decision making characteristics of the expert. Most expert systems to date incorporate knowledge that is primarily scientifically oriented; for example, applications exist in medicine, chemistry and'biology [CJV 831. These systems can be characterized as having a clear task and expertise that is relatively easy to locate. Generally the problems to which these applications are applied require substantial time, such as the diagnosis of a patient's diseases. Frequently, existing ES's are used in environments where there is no current computer support. The conditions imposed by a business environment differ from those described above. First there appears to be extensive interaction among different tasks; drawing the boundaries for what the expert does is not easy. Second, there may be substantial expertise required to perform a given task, but the expert is often functioning in a high volume, high pressure environment. He or she often has to generate a large number of expert decisions in a relatively short period of time. Finally, successful business utilization of expert systems technology will require that these systems interact heavily with existing business subsystems, such as transactions processing systems and database management sys tems . This paper reports on the authors' experiences and reflections on the design of an expert system for life insurance underwriting. The particular application is underwriting applications for life insurance for a large insurance company. The problem, as in many ES, is essentially one of classification: determining into which insurance category an applicant falls for insurance purposes. The problem is an important one, and is a routine part of the business. Transactions volumes are high and the underwriter makes use of a large volume of information in coming to a decision. In addition, the underwriter must interpret facts and frequently apply judgment based more on experience or rules of thumb than on clear-cut policies. While insurance firms are some of the largest users of computer-based processing systems, the fact that there are very few systems (for an exception, see the CAUSE underwriting DSS [A1 801) to aid in the underwriting decision provides evidence that significant human judgment is required for this task. It should be noted that this research is part of a larger NYU Expert Systems Project which is also concerned with developing a general interface mechanism between expert systems and database management systems [JV 841. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 2. EXPERT SYSTEMS IN THE BUSINESS ENVIRONMENT One generally accepted macro view of the managerial control processes in an organization defines the structure as consisting of a hierarchy of three distinct but cooperating levels of activity and decision-making [An 65, Si 60, GS 71 1. Operational control, the lowest level in the hierarchy, is concerned with all of the details of the individual transactions which constitute the activity of the business. One level higher is the managerial control level, where more summary information about the business is needed in order to support medium-level decisions affecting the management of the business functions. Finally, the top level consists of the strategic planning and policy making function which sets the goals and defines the overall nature of the business organization. Traditional data processing systems have primarily concerned themselves with the lower two levels of this hierarchy. When one thinks of business processing systems one thinks of such transaction processing systems as order-entry or inventory management, and such managerial reporting systems as accounts receivable and general ledger. More recently, a class of somewhat less-defined systems known as Decision Support Systems have begun to be discussed and researched [SC 821. These systems are primarily aimed at supporting the kinds of decisions that are required at the higher levels of this hierarchy. We believe that expert systems and expert systems technology have a role to play at all levels of this organizational hierarchy. The reason for this is that, orthogonal to this hierarchic view of the nabre of the business activity is a view of the complexity of the activity being carried on. Accepting this classification of business activities along a complexity scale ranging from vstructuredtf to "semi-structured" to ttunstructured", these two parameters, level of activity and complexity of activity, produce a matrix of business activity types. We might view the situation therefore as in Figure 1, and ask ourselves (a) where in this matrix does a particular activity lie, but, more importantly for this discussion, (b) where in this matrix are traditional DP techniques and the newer ES techniques most suited? Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 4 STRUCTURED SEMI-STRUCTURED UNSTRUCTURED I TRANSACTION I DP ES ES* I I MIDDLE-LVL I DP /DSS ES/DSS ES* I I STRATEGIC 1 ES/DSS ES* ? DP = traditional Data Processing System DSS = traditional Decision Support System ES = Expert System, Current Technology ES* = Expert System, Future Technology ( Integrated with DSS Concepts) Figure 1: Taxonomy of Computer Applications We believe that traditional DP techniques have been and will continue to be applied successfully to both transaction and middle-level information systems, but primarily at problems of a routine level of difficulty. What is beginning to happen, however, is that the techniques of ES and DSS technology are being applied at some routine problems at the strategic level, and also at complex problems at all three levels. For the really difficult problems at all three levels, it seems that the jury is still out; we will have to wait both for some more powerful ES/DSS techniques, languages and tools, and also for an evaluation of the success/failure of these techniques at the complex level. What this analysis, if correct, means, is that business have barely begun to automate or semi-automate processes, decision-making, and planning in their organizations -- they have simply gone through the first stages of a computer revolution, which have helped to solve some of their easier problems. Now that techniques of A 1 have been developed that enable the automation of much more intelligent processes, businesses will begin to re-examine the whole range of their activities, from transaction processing through strategic planning, searching for new candidates for automation, from among those problems previously thought too difficult. The underwriting expert system project can be viewed as an attempt to apply ES technology in a complex transaction processing system. Before discussing some of the particular characteristics of this problem, some general remarks about transaction processing and ES technology are in order. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 5 The vast majority of the ESfs that have been built have been highly interactive, and intended for an environment in which there was a long interaction with a human and the system in an effort to come to a reasonable solution or decision for a single problem. Whether this problem was the diagnosis of a patient's illness (MYCIN [Sh 761, INTERNIST [P 847, PUFF [Kea 781, etc.), the synthesis of a promising new chemical compound (SYNCHEM [Gea 77]), the likelihood of a mineral deposit in a particular geographical area (PROSPECTOR [DR 841), or whatever, the striking characteristic is one of prolonged interactive problem-solving.' R1 E M 821 is probably the first ES reported in the literature which moved closer to the transaction processing realm. It also appears to be one of the few expert systems in actual routine use. This system advises systems engineers at Digital Equipment in configuring VAX computers. Although advertised as "thew major success in expert systems to date, it should be noted that its development costs have reportedly far exceeded the current or expected benefits of the system. We can conclude that (a) ways must be found to reduce the costs of building expert systems, and (b) one has to be very careful in selecting an application domain with sufficient potential payback. As noted above, the latter can be achieved by supporting either important single decisions or mass- processing of transaction-like, medium-complexity decisions. What are some of the characteristics of transaction processing systems? The first is that they must be able to handle a high volume of transactions per unit of time, i.e., they should not spend a long time agonizing over the decision for any particular case. A second is that they typically have a high level of interaction with other systems in the business, e.g. a DBMS, other transaction systems and systems at the middle and strategic levels of the organization. Finally, a third characteristic is that they must be able to track transactions through various stages of processing, for example from initial processing through an "awaiting action" state through a "final disposition" state, since in many applications not all transactions can be completely processed in one pass. Each of thes factors should influence the shape of business ES's that might be built for complex problems at the transaction level. The first issue, that of high volume, dictates that the ES needs to be organized to be extremely fast, and probably would best operate in a batch rather than an interactive mode. However, many people feel that ES's will not be trusted to make decisions unilaterally in complex tasks, but rather will only be effective when built to operate in conjunction with human experts in the domain. The inherent tension between these two issues involved in complex transackion level systems, namely high volume with its need for speed, and decision importance with its need for confidence, will require creative systems design if ES's are to succeed in this business area. As we shall see, the underwriting system addressed this issue by partitioning the transactions into disjoint classes of transactions with different characteristics. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 6 The second issue is a crucial one in determining the scope of the ES and its relationship to its environment. Most ES's to date have been built by systems designers who were free to construct their knowledge base in isolation from any existing organizational databases, policies, or constraints. They have typically chosen a knowledge representation scheme appropriate to the problem to be solved and to the language in which the system was being built. By contrast, a business ES will undoubtedly have to rely upon an existing DBMS for the major part of its factual knowledge about the enterprise in question, and probably upon existing business subsystems -- transaction systems, mathematical modelling systems, decision support systems, etc. -- for some of the data processing that will be relevant to its proper functioning [JV 841. No business could tolerate the indiscriminate replication of data, of modelling algorithms or processing elements such as summary reports and statistical data on its operations. Not only would such replication be costly, but it would open the door to data and processing inconsistency and lack of security -- all of the things that modern DP techniques and policies, and DBMS's have worked hard to prevent. These components of the business -- data and processing information -- have attained the status of an important corporate resource, and ES technology, if it is to succeed in this environment, will have to accommodate itself to this view and the policies and structures that it has generated. In practice this means that the ES will have to rely heavily -- as more and more a 2 business systems do -- upon the DBMS as the source of most of its data. One of the major thrusts of the research that involved us in the insurance underwriting ES was the exploration of a general mechanism for just such ES-DBMS interaction, in particular in the context of a Prolog-based ES and a relational DBMS [JCV 84, VCJ 851. However it is clear that the DBMS is not the only existing system with which the ES will have to contend. For instance, the underwriting function in insurance is but one step in the process of processing an insurance application. This process begins with the salesperson-client interaction, which generates an application. Some of the information on the application is then coded into an order-entry subsystem. This generates a file of pending applications, which together with the entire application is sent to the underwriters. After the underwriting is completed -- which may involve a pending file awaiting the arrival of additional information, as we shall see -- the application is then either sent on to policy issue, where the actuarial subsystem computers the premium and creates an entry in the corporate database, or it is sent to the reject subsystem which must generate a file and a letter explaining the reason for the rejection. If an ES is to be built to help perform some or all of the tasks of the underwriting function, clearly it must fit into an elaborate, pre-existing structure of systems, subsystems, policies, procedures, and legal strictures. Before discussing some of the more general issues involved in business ES deployment, the next section provides a context by discussing the particular application of insurance underwriting. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 7 3. THE UNDERWRITING EXPERT SYSTEM An insurance policy transaction begins when the sales representative and the customer complete an application form and forward it to a regional office for underwriting. Here the application is assigned to an underwriter based on the size of the requested insurance policy and the type of medical information provided on the form. At this point the underwriting process begins. An insurance underwriter is basically charged with determining whether or not the company should accept a customer. This overall task involves examining the potential customer's application and all other accompanying material to determine either (a) accept the customer in such and such a rating category, or (b) reject the customer for such and such reasons, or (c) withhold a final decision pending the acquisition of additional supporting information. In the company cooperating in this research, the type of product involved is life insurance of any kind, term or whole life. Underwriting is not a single decision at one point in time. Rather, there are two distinct decision styles. The underwriter first makes a decision on what information to request, for example, a medical examination, report from the applicant's personal physician or several other pieces of information. The application is placed in a file pending the arrival of the information needed for underwriting. When all data are complete, the underwriter examines the information and makes a decision that classifies the applicant into one of a number of categories, each having a different price for a given level of insurance coverage. 3.1. T& Information-Gathering Decision Figure 2 shows the different sources of information available to the underwriter. The applicant signs a release giving consent for all of this information to be made available to the company, and FOR the results of the underwriting process to be reported to a computerized database. Much of this information costs the insurance company money to obtain; in addition, the more information requested, the longer the underwriting process is likely to require, resulting in the potential loss of the customer to the competition. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 8 A check of computerized databases MIB/AIF of applicants for insurance for most large carriers in the U.S. A medical examination by varying levels of medical personnel from a paramedic to a specialist. A report from the applicant's attending physician. A mercantile report or credit and community investigation. A driving report from the Department of Motor Vehicles. Tax returns and/or a financial statement from the applicant. Figure 2: Sources of Underwriting Information One of the most important sources of information for the underwriting decision is the applicant's previous insurance history. This information can come from one of two sources. There is an industry-wide database maintained by the major U.S. insurance companies. Each time a company rejects an applicant for insurance, it reports that fact, and the reasons for the rejection, to this shared database.. When an individual applies for insurance at the company in this study, there is an automatic search request issued to this database in certain instances, for example is the amount of coverage applied for is over a certain threshold, or the applicant is over a certain age. A second source of similar, though more detailed, information is the company's own internal database. Again, certain policy characteristics trigger a query to this database. The next information that might be requested is a medical examination. Such factors as the age of the applicant and the amount of insurance requested will determine the scope of a medical examination. In addition, there are questions on the application about medical history, If the applicant reports one or more of a number of conditions, the underwriter will generally ask for a medical examination. This decision is complicated by the fact that there are several types of medical exams; the applicant may need a paramedical exam, a doctor, or a company-appointed specialist. Certain answers on the application cause the underwriter to ask for a report from the applicant's attending physician. Usually the doctor photocopies the most recent part of the patient's record and sends it to the company. Naturally the complexity of reading a handwritten report, difficult enough for the human "expert," were sufficient to place the examination of these reports outside the scope of the proposed system. A mercantile report is like a credit investigation. However, it focuses on the behavior of the applicant and may be requested for a number of reasons. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 9 Again, a major reason for authorizing this type of report is the face amount of the requested policy. Other triggers include such factors as the occupation of the applicant, a history of motor vehicle problems, or perhaps a medical disability. A mercantile report always includes a query to the department of motor vehicles to determine the applicant's driving record. Of particular interest are any arrests for driving while intoxicated. If the applicant does not require a mercantile, but is considered a high-risk driver for some other reason, especially age, and if additionally the policy amount warrants it, the company will query the department of motor vehicles. Finally, for certain very large policies, the applicant will be requested to submit a financial statement, usually in the form of a copy of a tax return. 3.2. Underwriting Decision Given the requested information, the underwriter can evaluate the application. The underwriter assigns scores for various conditions and diseases; the more points assigned to an application, the higher the cost of the insurance. There are a number of possible categories into the which the application will be placed, each one representing a different premium schedule. There can be a substantial dollar difference in premiums between the various categories, so the underwriting decision is an important one for the applicant. In addition to the classification, certain medical conditions will warrant a surcharge of so many dollars for every thousand dollars of insurance for a period of years. What is the source of the underwritersq scoring system? The medical and actuarial departments furnish underwriters with a medical guide. The guide's point totals assigned for different conditions reflect the results of actuarial studies and represent the risk the applicant poses to the insurance company. However, not all risks are additive, as assumed in the scoring system. In such cases, the underwriter applies judgmental rules, For example, an application is rejected if the applicant has a history of mental problems and is unemployed, even though none of these properties have a high score by themselves. Because the agent in the field has done some screening and since the company must insure customers to remain in business, there are relatively few absolute rejects, on the order of 5% of applications. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 10 4. KNOWLEDGE ACQUISITION 4.1. General &sues a Business Environment Business expertise is seldom localizable. In most businesses it is distributed among various people at different levels, and among various forms of documentation and processing elements. Most expert systems have been built in a different environment, where the expertise of orl_e person was evoked during knowledge acquisition, systematized, and incorporated into the knowledge base. In a business system, the knowledge acquisition process is complicated by the distributed nature of the expertise. Among the possible sources for relevant expertise are existing applications programs and other computerized systems; manuals, rule books, policy statements or other written materials representing compiled expertise of the organization; human experts from various departments or functional areas. Among the problems that this diversity of knowledge sources poses are the following: inconsistency across the sources; incompatibility of the view of the problem space; potential compromising of a human expert who may disagree with a stated policy or rule, The issue of resolving conflicts or disparities among multiple experts is one that will increasingly be an issue in the development of expert systems [H 851. The security of the organization's rules is an added issue in business ES's. The power and flexibility of the knowledge representation tools used to build ES's, i.e., the ability both to represent rules and policies declaratively and to access and modify these rules during program execution comes with a price. Just as data in a DBMS is seen as a corporate resource which must be protected from mis-use, so too (perhaps even more so?) business policies, strategies and rules encoded into the knowledge base of the expert system must be protected. One offshoot of this necessity to preserve the confidentiality of business rules is the difficulty of publishing results in the area of business expert systems; most of the real progress involves the development of prototype systems in confidential application areas. 4.2. Underwritin& System i . Particular For the underwriting expert, the design team interviewed various underwriters. At first there was interest in an expert for the high value policies, those with face amounts in the millions of dollars. However, it turns out that these policies are underwritten by a committee rather than a single individual. Construction of an ES for the committee decision did not look promising because of the number of individuals involved and the complexity of the decision process. Below the level of policies requiring a decision by committee, there is a large volume of underwriting at the district offices. The human expert used in the knowledge acquisition and prototyping for this system is a senior lay medical underwriter. He can underwrite policies up to about a quarter of a million Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 1 1 dollars, and other underwriters send to him any applications that need some medical judgement. The design team met with the expert for a period of six months prior to writing any code. The meetings began with the expert describing his job and decisions. Soon, the designers asked the expert to bring various applications and talk through his decisions on each one. A protocol analysis of these applications reviews provided enough understanding of the process to involve the expert in design. The designers and human expert worked together to complete Figure 2 and to describe the rules that apply for requesting each type of information, The system was then developed piece by piece, first designing the data structures needed to represent all of the information relevant to the decision, and then building the rules that determine when a request for individual pieces of additional information are necessary. Moreover, there were rules for developing a composite "scoren of the application, which could potentially contribute both to a request for additional information and ultimately to the insurance classification of the application. The development proceeded in the usual ES development mode, namely as a sequence of successive refinements to the quality of the system's decision making. Each meeting we would review the performance of the system on an application with our human expert, and discover where and why it had errors or been incomplete. In the next meeting we would have developed new rules to handle the previously discovered problems, and then proceed to discover new ones. At this point, an expert system is working for the first underwriting decision, on what information to acquire. In a limited "testw of the system against eight applications chosen by the underwriter to reflect a broad spectrum of application types, the system's performance was comparable to our advisor. In a production environment the system would need substantially more thorough testing, in a carefully planned and controlled test against many more cases and also involving other underwriters. Current plans call for developing a full expert system to underwrite either a class of applicants or applicants with a particular disease to learn how an expert can assist in the actual classification of an application. 5. UNDERWRITING IN CONTEXT The insurance firm is organized into regions and there is a steady flow of applications. However, business does tend to be seasonal with a definite peak in December as agents try to achieve their sales goals for the year. In contrast to systems that might advise on unsticking an oil well drill bit or on where a valuable body of one might exist, there are literally hundreds of underwriting decisions made each day. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 12 In contrast with other ES, this system involves a number of relatively low value decisions, the cumulative effect of which is of high importance to the firm. It should also be noted that there is a lack of a fine grained performance measure for the underwriter. These individuals are not given feedback on how the cases they underwrite perform, though such information is available in the company. An ES must be designed to work with the information provided on the insurance application and from the sources described above. There can be no lengthy discussion with the underwriter or it will take too long to process transactions. In fact, one of the attractions of an expert system for underwriting is the potential to move part of the decision into the field. An underwriting expert accessible by the sales agent could speed response to a customer. With government deregulation, insurance companies are beginning to offer more financial services. Customers, however, are used to relatively fast response from current service firms. An ES that provided a fast underwriting decision for a large percentage of applicants would be a competitive advantage for an insurance company. Exactly how such a system would be implemented, and its potential impact upon employment, are questions that need further examination. There is an existing system at the company which accepts limited input from the application and requests reports from both the internal and industry-wide databases when the applicant falls under the rules requiring such information. Further, some of the firm's existing systems are used by the actuarial and medical staff to prepare the medical guide, a manual which covers all of the underwriting policies regarding the medical condition of the applicant. One could eventually envision a series of expert systems, each making extensive use of part of the firm's database. An actuarial expert (cf. [SJ 851) could be queried by the underwriting expert and could produce a customized risk analysis for each applicant. An overall picture of some of the system interconnections desirable in the insurance company environment is shown in Figure 3. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 13 I I I INDUSTRY- I 1 WIDE I 1 DATABASE I 1- I I I I /UNDERWRITING I I ACTUARIAL I I EXPERT I<-------- I SUBSYSTEM I / I SUBSYSTEM I I I / - I \ I 1-1 I I \- I I I ORDER I I CORPORATE1 I POLICY I I ENTRY I ------------ > I DBMS I------- I ISSUE I I SUBSYSTEM I I I I SUBSYSTEM I Figure 3: Underwriting System in Context More immediately, it is not clear yet what kind of interaction the underwriting expert will require with databases. The information gathering part of the expert system accesses very large databases but does not appear to need extensive secondary storage capabilities of its own. However, systems that make actual underwriting decisions must interact with the large set of rules and tables in the medical guide. Design is not sufficiently completed as yet to determine whether the coupling between the database and expert will need to be loose or tight [JV 841. 6. RELATIONSHIP TO TRADITIONAL SYSTEMS ANALYSIS AND DESIGN To what extent is the design of an expert system different from that of a conventional, computer-based system in a business firm? Table 1 shows the stages in the conventional life cycle for a business information system [L 851. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 14 LIFE TRANSACTION CYCLE PROCESSING SYSTEM PROTOTYPE EXPERT SYSTEM Incept ion Idea for system Idea for system Idea for system Feasibility Cost/Benefit; Working prelim. R & D; Study Mgmt. approval version Set boundaries Sys terns Details of Work with user Listen to Analysis current procs. group single expert Design Create new sys tem Iterative enhancement Begin to code single expert Specifications Detailed design Evolving prototype, Expert critique; or prototype as new rules added model for system Programming Code specs. Code specs. Add rules Testing Debugging; Debugging; Continued Acceptance testing Acceptance testing enhancement; Problems when exposed to others Training Conversion Operations For users For users For non-experts? Scheduled Gradual evolution Gradual Maintenance & Maintenance & Evolving expertise enhancements enhancements Table 1: System Life Cycle Comparisons 6.1. Traditional Systems Analysis and Design Usually someone in the firm has an idea for how a system might improve information processing or solve a problem. The systems staff conducts a feasibility study showing various costs and benefits (including intangible benefits) and management is asked to approve developing the application. During systems analysis a design team of users and analysts document the details of present information processing procedures. The team creates the design and specifications for a new system. Programmers Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 15 develop necessary code or modify packages. The team tests the system, trains users and converts to the new application. During maintenance, the information services department makes changes and enhancements to the system, an activity that consumes on the average 50% of the discretionary resources of the department. 6.2. Prototyping What is different in the development process for an expert system based on the experience with the underwriting expert? A new approach to conventional design called prototyping has received a great deal of recent attention [L 851. Table 1 shows that a prototype differs significantly from traditional development approaches. Elasically, the prototype is a working model of a system or some part of it. In the conventional approach, the user does not receive feedback until too late in the process; often the user does not understand the system, but yet agrees to its specifications. With a prototype, the user has something concrete to which he or she can respond; the prototyping process also tends to draw the user into the design process. It is hard to avoid user involvement with a prototype. 6.3. Expert Systems Design The design of an expert system comes closest to the prototyping process. Some of the major differences between prototyping and ES design result from the current R&D nature of expert system development. So few of these systems exist in a business environment that a clear methodology has not emerged for their design. Based on experience with the underwriting expert system, six issues will be addressed, below: reliance on a single expert, staffing, program structuring, user training, systems maintenance, and accountability. Based on the experience with the underwriting expert, one significant difference with other design approaches is the reliance on a single expert. ES designers have generally been unanimous that one cannot currently design an expert system if it is necessary to adjudicate among different experts with differing opinions. In conventional design, most organizations try to include as many potential users in the design process as feasible. An issue of major concern to businesses weighing the potential costs and benefits of pursuing ES technology is one of staffing. DP departments do not currently have staff who are familiar with A1 languages or with the process and techniques of knowledge engineering. A commitment to ES development will require substantial investment either in outside consultants, training for current staff members or the hiring of new staff trained in ES development. However, this expertise is itself still rare, making it difficult to locate and then keep a staff of knowledge engineers. A final point is that the maintenance of ES1s is Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 16 more akin to the learning and maturation of a human expert than to the maintenance and enhancement of traditional DP systems. Staffing over the lifetime of the usefulness of an ES, which is assumed to be rather lengthy, poses an additional problem. With an ES the designers are using a different type of language than COBOL or even fourth generation, nonprocedural languages. There is relatively little experience with rule-based languages in the business environment. One advantage of these languages is purported to be the ease with which rules can be added as designers acquire new knowledge. Experience with the underwriter shows that some new knowledge can be significant enough to require restructuring of parts of the program even though it is written in Prolog, a rule-based language. With an ES training is a potential problem; what happens when other experts are asked to evaluate the system? If it was designed with a single expert, will it necessarily be robust enough to pass tests designed by colleagues? How does one approach the training of users? How does the system respond when a nonexpert user presents a situation beyond its domain of expertise? Designers of commercial systems are used to developing applications with extensive error checking, sometimes well over half of the code exists to detect user errors, Will expert systems designers be able to detect and prevent user errors? Maintenance is a major problem in the information processing industry. Expert Systems will add to this problem, not lessen it. ES's will probably become members of a class of enduring systems, applications so expensive or so complex that one could not envision replacing them completely. Good examples of this type of systems are those used by airlines for making passenger reservations. Two U.S. carriers have invested over $200 million on their systems [CS 841; it is inconceivable that these systems will ever be totally replaced. Instead, they will grow and be modified over time. Expert Systems will have to receive careful maintenance. New rules will be discovered so there will have to be a staff to tend the system and incorporate new knowledge. The domain and nature of the problem addressed by the ES may change as well. The care of an Expert System will therefore probably be more demanding than the maintenance of a conventional application. A final issue relates to the notion of auditing and accountability. Expert systems will presumably become more and more involved in making and/or assisting in the making of complex, high-risk decisions. In many applications, e.g. insurance underwriting, drug synthesis, etc., there are corporate and legal responsibilities for maintaining complete records about the decisions made and the environment in which they were made. Increasingly, lawsuits are won (or lost) with the help of such documentation. The recording and maintenance of such information about a decision as when, how, why, and under what assumptions about the "state of the worldft was a decision made will increasingly become part and Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 17 parcel of the expected behavior of expert systems as they begin to become involved in the making of crucial organizational decisions. 7. SUMMARY AND CONCLUSIONS ES in business differ from many of the expert systems created to date. One application for business ES1s is at the transaction level, for processing more sophisticated transactions than those amenable to DP systems. Here one often finds a high volume of transactions with each transaction being of relatively low value. However, all transactions must be processed and there will be many expert decisions made in a short period of time. The insurance underwriter is an example of -such a system. We believe that ES1s will begin to be seen at this organizational level first, as business take preliminary steps toward the acceptance of the new technology. If they are successful at this level, ES1s will begin to be applied to higher-level, more critical decisions in the organization. The design process for an Expert System is sufficiently different from conventional design that it will be difficult for traditionally-oriented staff members to develop an ES without assistance. There is a need for educating knowledge engineers and for a methodology for ES design. An approach that follows prototyping looks appealing based on the experiences reported in this paper. Because there are few ES in business, no one has addressed the question of implementation for these systems. In the underwriting environment, many overall implementation questions arise. Will an underwriting expert support or replace the human underwriter? Will it be used for the thousands of policy applications that require modest human expertise, eliminating a layer of underwriters at the level below the senior lay underwriter in this study? These questions will naturally arise in other business areas where ES1s are being introduced; they will need to be considered carefully if business is to seriously accept the introduction of this technology. In light of the high development costs currently incurred in the development of ES1s, strategies for multiple systems using a central or "corew knowledge base need to be considered in order to make business ES1s more cost effective. For example, one additional use of the underwriting expert system would be to move some of the simpler underwriting into the field. Another possibility would be to develop an underwriting training system which would draw upon the same knowledge base of expertise, but could be used to generate test case applications and score the decisions made by the underwriting trainees. In summary, the introduction of Expert System technology into a business environment presents new and challenging problems as well as opportunities, Our experience in developing the underwriting expert illustrates both of these facets of this technology transfer. The opportunities involved in computerizing Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 18 activities that have until now resisted automation, with the potential for an increase in speed, accuracy, and competitive edge, are very real and attractive. The problems are no less real: how to incorporate techniques of knowledge engineering into environments more comfortable with traditional systems analysis and design techniques; how to minimize the cost of the knowledge acquisition and knowledge engineering process;how to integrate ES's into the wide family of cooperating systems that are today the norm in most businesses; how to manage the overall human impact of this technology on the employees and their relationship to their job environment, their colleagues, and their careers. Clearly much more research is needed on the design and implemention of ESts if this technology is to become widely adopted in the business environment. REFERENCES [An 651 Anthony, R.N. Planning Control Systems: A Framework for Analysis, Harvard University GSBA, Studies in Management Control, Cambridge, Mass., 1965. [CJV 831 Clifford, J., Jarke, M. and Vassiliou, Y. "A Short Introduction to Expert Systems," Database Engineering Bulletin, vol. 8, no. 4, December 1983. [DR 841 Duda, R.O., and Reboh, R. "A1 and Decision-Making: The PROSPECTOR Experiencett, in W.Reitman (ed.), Artificial Intelliaence Applications for Business, Ablex, Norwood, NJ, 1984, pp.lll-147. [~ea 771 Gelernter, H.L., Sanders, A.F., Larsen, D.L., Agarwal, K.K., Boivie, R.H., Spritzer, G.A., and Searleman, J.E., "Empirical Explorations of SYNCHEM," Science v. 197, 1977, pp. 1041-1049. [GS 841 Gifford, D. and Spector, A. "The TWA Reservation System," Communications of the ACM, 27,7, July 1984, 650-685. [H 851 Hewitt, C. "Implications of Open Systemst+, in M.Jarke (ed.), Managers, Micros, a* Mainframes-: Proceegi.~ of the N x agosium Integrating Systems for End Users, New York University, May 1985. --- [ JCV 841 Jarke, M., Clifford, J., and Vassiliou, Y. "An Optimizing Prolog Frontend to a Relational Query System", Proceedings ACM-SIGMOD Conference, Boston, Mass., pp.296-306. [Jv 841 Jarke, M. and Y. Vassiliou. "Coupling Expert Systems with Database Management Systems, in W. Reitman (ed. ) Artificial I~telligenw A_pplications for Business, Ablex, Norwood, N.J. 1984, pp.65-85. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 Page 19 [Kea 781 Kunz, J.C. et al. "A Physiological Rule-based System for Interpreting Pulmonary Function Test Rules," Heuristic Programming Project Memo HPP-78-19. [L 851 Lucas, H.C., Jr. Analysis and Design of Information Systems, 2nd ed., McGraw-Hill , 1985. [M 821 McDermott, J. "R1: A Rule-Based Configurer of Computer Systems", Artificial Intelligence, vol. 19, no. 1, pp.39-88. [P 841 Pople, H.E., jr. "Knowledge-Based Expert Systems: The Build or Buy Decision", in W.Reitman (ed.), Artificial Intelligence Applications f ~ r Business, Ablex, Norwood, NJ, 1984, pp.23-40. [Sh 761 Shortliffe, E.H. Computer Based Medical Consultations: MYCIN, American Elsevier, New York, 1976. [Si 601 Simon, H.A. The New Science of Management Decision, Harper & Row, New York, 1960. [SJ 841 Sivasankaran, T., and Jarke, M. "Logic-Based Formula Management Strategies in an Actuarial Consulting System", Decision Support Systems, vol. 1, no. 3, 1985. [VCJ 851 Vassiliou, Y., Clifford, J., and Jarke, M. "Access to Specific Declarative Knowledge by Expert Systems: The Impact of Logic Pr~gramming,~~ Decision Support Systems, vol. 1, no. 2., 1985. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-62 
work_lpc476slxvcqfexd5moavsefpy	----	An Algebraic Approach to Rule Based Expert Systems Available on line at www.rac.es/racsam REVISTA DE LA REAL ACADEMIA DE CIENCIAS Computational Sciences EXACTAS, FISICAS Y NATURALES. Survey SERIE A: MATEMATICAS Madrid (España / Spain) RACSAM 104 (1), 2010, 19–40. DOI:10.5052/RACSAM.2010.04 An algebraic approach to rule based expert systems Eugenio Roanes-Lozano, Luis M. Laita, Antonio Hernando and Eugenio Roanes-Macı́as Abstract This article presents a survey of the authors’ research on knowledge extraction and verification of Rule Based Expert Systems (RBES) using algebraic inference engines and based on Gröbner bases theory. A shell, including a graphic user interface and inference engines for different logics (both classic and modal multi-valued) as well as in different computer algebra systems, is also presented here. The shell distinguishes three levels: at the lower level, we provide the computer algebra system code of the algebraic inference engines; at the intermediate level, the RBES developer has to detail the rules and integrity constraints of a certain RBES; and, finally, at the upper level, the final user deals with a simple GUI, where he can perform knowledge extraction or verify the RBES, after choosing the logic and inputing a consistent set of facts. We believe that this shell can be really useful for teaching and quick RBES design. Una aproximación algebraica a los sistemas expertos basados en reglas Resumen. Este artı́culo presenta una panorámica de la lı́nea de investigación de los autores en ex- tracción de conocimiento y verificación de Sistemas Expertos Basados en Reglas (RBES) usando motores de inferencia algebraicos y basada en la teorı́a de bases de Gröbner. Se presenta también una shell, que incluye una interfaz gráfica de usuario y motores de inferencia para distintas lógicas (tanto clásicas como modales multivaluadas) y en distintos sistemas de cómputo algebraico. La shell distingue tres niveles: en el más bajo proporcionamos el código del motor de inferencia para el sistema de cómputo algebraico elegido; en el intermedio el desarrollador del RBES tiene que detallar las reglas y las restricciones de integridad de un cierto RBES; y, finalmente, en el nivel superior, el usuario final trata con una sencilla interfaz gráfica de usuario, en la que puede llevar a cabo extracción de conocimiento o verificar el RBES, después de elegir la lógica y de introducir un conjunto consistente de hechos. Creemos que esta shell puede ser realmente útil para la enseñanza y para el rápido diseño de RBES. Submitted by David Rı́os Insua. Received: August 2, 2009. Accepted: October 19, 2009 Keywords: Rule based expert systems, logic and symbolic computing, Groebner bases Mathematics Subject Classifications: 03B05, 13P10, 03B80 c© 2010 Real Academia de Ciencias, España 19 http://www.rac.es/racsam E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 1 Introduction This article presents a survey of the authors’ research on knowledge extraction and verification of Rule Based Expert Systems (RBES) using algebraic inference engines. These inference engines are based on the use of Gröbner bases. The leader of the research team is Luis Laita, and the first works on the topic are dated in the early 90’s, although he had already been working in the field using other techniques. The present paper is self-contained, in such way that introductory well-known sections 2, 5, 7may be skipped by an acquainted reader, while the other sections convey a view of new results and approaches by the authors. The article begins with a brief introduction to lattices, lattice orders, Boolean algebras and Boolean rings (Section 2). It is followed by the construction of the polynomial model for Boolean logic due to these authors (Sections 3 and 4). Then a brief introduction to modal multi-valued logics can be found (Section 5). Section 6 introduces the main result of this theory, relating to be a tautological consequence with a polynomial ideal membership. A brief introduction to RBES is included afterwards (Section 7). Section 8 details the adaptation of the algebraic model for logic to RBES. In Section 9 it is shown how two levels of inconsistency can be distinguished in the multi-valued case (including an interpretation in algebraic geometry). Section 10 introduces how (in RBES whose underlying logic is classic Boolean logic), given a set of facts, it is possible to find out which new fact can make a certain goal be inferred, just computing a Gröbner basis. Implementations in the computer algebra systems (CAS) CoCoA and Maple are included in Section 11. Finally, a GUI for keeping the CAS hidden from the final user of the RBES is presented in Section 12. In view that quite a number of topics are treated, many proofs are omitted for the sake of brevity (in fact, some of the proofs are very long, for instance, the one connecting the concept of being tautological consequence in multi-valued logic with an ideal membership in a polynomial residue class ring). Anyway, references where the reader can find details on the topics treated can be found along the paper. 2 An introduction to Boolean algebras and Boolean rings An elementary introduction to Boolean algebras may be found in [21]. The proofs omitted in this section can be found in [18, 28] (these references also extend both Sections 3 and 4). For a deeper study of Boolean algebras see, for instance, [10, 11, 12, 22, 36]. 2.1 Lattices and lattice orders Definition 1 A set L where two binary operations t and u are defined is said to be a lattice if and only if both operations are commutative and associative and the absorption laws holds, i.e., if and only if for all elements a, b, c ∈ L: at b = bta au b = bua at (bt c) = (at b) t c au (bu c) = (au b) u c at (bua) = a au (bta) = a Proposition 1 If (L,t,u) is a lattice, both operations verify idempotency, i.e., for every element a ∈ L: ata = a aua = a Surprisingly, idempotency, although redundant, is often required as a fourth condition for a structure to be a lattice. Definition 2 A non-strict partial order defined over the set L is said to be a lattice order if and only if every pair of elements of L has a unique infimum and a unique supremum. 20 An algebraic approach to rule based expert systems Proposition 2 From the two lattice operations, t and u , a lattice order v can be defined: for all elements a, b ∈ L: a v b ⇔ at b = b (⇔ au b = a) Conversely, from a lattice order v , the two operations of a lattice can be defined: for all elements a, b ∈ L: at b = supv(a,b) au b = infv(a,b) Example 1 Let us consider the lattice of “subsets of E”: (P(E),∪,∩). From (P(E),∪,∩), (P(E),⊆) is obtained. Conversely, from (P(E),⊆), (P(E),∪,∩) is obtained. 2.2 Boolean algebras Definition 3 A lattice (L,t,u) is said to be distributive if and only if both operations are distributive w.r.t. the other operation, i.e., if and only if for all elements a, b, c ∈ L: at (bu c) = (at b) u (at c) au (bt c) = (au b) t (au c) Definition 4 A lattice is said to be bounded if and only if it has a greatest element (top), g, and least element (bottom), l. Definition 5 A bounded lattice (L,t,u) is said to be complemented if and only if for every element a ∈ L, there is an element a′ ∈ L such that: ata′ = g aua′ = l The top and bottom of a lattice are traditionally denoted 0 and 1. However this seems not to be an adequate notation, as the top and bottom of (L,t,u) and those of (L,u,t) are exchanged. Definition 6 A Boolean algebra is a distributive and complemented lattice. Example 2 (P(E),∪,∩, ′) is a Boolean algebra: • ∪ is distributive w.r.t. ∩ and viceversa. • The least and the greatest of the lattice (for ⊆) are ∅ and E, respectively. Moreover, the usual complement of a set, ′ , works as expected: ∀A ∈P(E) : A∪A′ = E ; A∩A′ = ∅ Let us underline that being distributive is completely independent from being complemented, as shown below. Example 3 The lattice of the linear subspaces (also denoted “vector subspaces”) of R2, the sum of linear subspaces and intersection is a complemented, but not distributive, lattice: • the complement of a line is any other (different) line, and the complement of the whole plane is {0}, • if R, S, T are different lines, R + (S ∩T) 6= (R + S) ∩ (R + T)). Example 4 The lattice (N, lcd, gcd) (where lcd and gcd stand for “least common multiple” and “greatest common divisor”, respectively) is distributive, but it is not complemented: 1 and 0 are the least and greatest of the lattice (respectively), but, for instance, 3 has no complement. Example 5 The lattice of the convex subsets of the plane, the convex union ∗ ∪ (i.e., the least convex set that contains the usual union) and intersection is a lattice, but it is neither distributive nor complemented: • for instance, if A, B, C are three disjoint and aligned squares of equal size and A is the one in the middle A∩ (B ∗ ∪ C) = A, meanwhile (A∩B) ∗ ∪ (A∩C) = ∅ ∗ ∪∅ = ∅, • there is no convex subset of the plane that behaves as the complement of a line. 21 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 2.3 Boolean rings Definition 7 A Boolean ring is a ring with unit such that the second operation is idempotent. Let us include below some well-known results that will be useful when obtaining the polynomial model for Boolean logic. Let us denote by (L, ∆,u) a general Boolean ring, by 0 its neutral element, by ′ the symbol for the opposite and by 1 its unit. Proposition 3 In any Boolean ring (L, ∆,u) any element is its own opposite. PROOF. For any elements a, b ∈ L: (a ∆ b) ∆ 0 = a ∆ b = (a ∆ b) u (a ∆ b) = (aua) ∆(au b) ∆(bua) ∆(bu b) = a ∆((au b) ∆(bua)) ∆ b = (a ∆ b) ∆((au b) ∆(bua)) but (L, ∆) is a group, so the opposite is unique, and consequently: 0 = (au b) ∆(bua). Therefore, in the particular case that a = b, from the idempotency of u, we would obtain: 0 = a ∆ a. � Proposition 4 Any Boolean ring (L, ∆,u) is a commutative ring. PROOF. It follows from the equality 0 = (au b) ∆(bua) in the previous proof and the facts that (L, ∆) is a group (and therefore the opposite is unique) and that any element is its own opposite. � Proposition 5 A Boolean ring (L, ∆,u) can be defined from a Boolean algebra (L,t,u, ′): for all elements a, b ∈ L: a4b = (au b′) t (a′ u b) Conversely, a Boolean algebra (L,t,u, ′) can be defined from a Boolean ring (L, ∆,u): for all elements a, b ∈ L: at b = (a4b)4(au b) and if 0 and 1 are the neutral elements of ∆ and u (respectively), then 0 is the least, 1 is the greatest, and: a′ = a ∆ 1 Example 6 (P(E),∪,∩, ′) is a Boolean algebra (least: ∅; greatest: E). If we define the symmetric difference of two elements of P(E), A and B, as: A4B = (A∩B′) ∪ (A′ ∩B) we have that (P(E),4,∩) is a commutative ring with unit: all properties are a straightforward conse- quence of the same property of the Boolean algebra or can be obtained from a simple symbolic manipulation (like the commutativity and associativity of 4 ). Moreover: • ∅ is the neutral element for 4 (and the absorbent element of ∩ ), • E is the neutral element for ∩ , • ∀A ∈P(E), A4A = ∅, so any element in P(E) is its own opposite. Example 7 (P(E),4,∩) is a Boolean ring. If we define in this ring the following operation: AtB = (A4B)4(A∩B) the operation ∪ is obtained. If we define A′ = A ∆ E the usual complement of a set is obtained. 22 An algebraic approach to rule based expert systems 3 A polynomial model for propositional logic We can obtain a polynomial model for the Boolean algebra associated to (propositional) Boolean Logic, as suggested in Figure 1. propositional Boolean algebra oo Boolean algebra isomorphism // OO �� polynomial Boolean algebra OO �� propositional Boolean ring oo ring isomorphism // polynomial (Boolean) ring Figure 1. Isomorphisms between polynomial and propositional Boolean algebras and Boolean rings. In logic, we usually work in the structure in the upper left corner of the diagram of Figure 1. Meanwhile, in commutative algebra we normally work in the structure in the lower right corner of this diagram. But we can move from the propositional Boolean algebra to the polynomial ring following the two alternative paths in this diagram. The advantages of working in a polynomial ring are that: • there is a powerful tool that solves the ideal membership problem: Gröbner bases, • there are many implementations of Buchberger’s algorithm for computing Gröbner bases, i.e., per- forming effective calculations in algebra. In fact, most CAS, like Maple, Mathematica, Axiom, Mu- PAD, Reduce, Derive, CoCoA, Singular, Risa/Asir,. . . include implementations of this algorithm. As a consequence, we shall be able to implement very easily an algebraic tool to perform effective compu- tations in propositional Boolean logic. Moreover, it will be possible to extend this tool to perform effective computations in modal multi-valued logics (like Łukasiewicz, Kleene and Bochvar). Finally, it will be pos- sible to use the same computational tool to perform verification and knowledge extraction in RBES based on both classic Boolean and modal multi-valued logics. 3.1 Building the Taylor-made polynomial (Boolean) ring (A, +, ·) Let us build a polynomial Boolean ring (A, +, · ) so that: • every element is its own opposite, i.e., for all a ∈A, a + a = 2 ·a = 0, so it would be reasonable to consider A as a ring over a field with characteristic 2, • idempotency holds for the second operation, so it would be reasonable to consider the residue class ring over the polynomial ideal 〈p2 − p,q2 − q, . . . ,r2 − r〉, where p, q, . . . , r are the polynomial variables. Therefore, we shall consider: A = Z2[p,q, . . . ,r]/〈p2 −p,q2 −q, . . . ,r2 −r〉 Proposition 6 The polynomial product in A is idempotent. PROOF. It is enough to take into account that any polynomial a ∈A can be written: a = δ0 + δp ·p + δq ·q + · · · + δr ·r + δpq ·p ·q + · · · + δpqr ·p ·q ·r + · · · where all the δ belong to Z2, and compute a2. � It is well-known that an algebraic structure like (A, +, · ) is a ring (denoted residue class ring) and, therefore, we have the following: 23 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as Corollary 1 The ring (A, +, · ) is a Boolean ring. Proposition 7 All elements in (A, +, ·) are zero divisors (in fact an analogous result holds in any Boolean ring). PROOF. a · (1 + a) = a + a2 = a + a = 2 ·a = 0. � 3.2 The polynomial Boolean algebra (A, +̃, · , 1+) corresponding to the po- lynomial (Boolean) ring (A, +, ·) Let us define a (polynomial) Boolean algebra (A,t, · , ′) from the (polynomial) Boolean ring (A, +̃, ·), as shown in Proposition 5. For all a, b ∈A: at b = (a + b) + (a · b) a′ = a + 1 The first operation will be hereinafter denoted +̃ and 1+ will denote the addition of 1. 0 is the least and 1 is the greatest of the Boolean algebra (A, +̃, · , 1+): for all a, b ∈A, a+̃0 = a + 0 + a · 0 = a a · 0= 0 a+̃1 = a + 1 + a · 1 = 1 a · 1= a Unlike what happens in a ring, it is not usual to establish priorities for the operations in a Boolean algebra (as the structure is completely “symmetrical”). Nevertheless, we shall consider, hereinafter, that · is prioritary w.r.t. +̃. Proposition 8 The lattice order defined in (A, +̃, · , 1+) as suggested in Proposition 2: ∀a,b ∈ A, a ≤ b ⇔ a · b = a is, precisely, “is a multiple” (a is a multiple of b ⇔∃k ∈A : a = b ·k). PROOF. ⇒) ∀a,b ∈A: a ≤ b ⇔ a · b = a ⇒ a is a multiple of b ⇐) ∀a,b ∈A: a is a multiple of b ⇔∃k ∈ A : a = b·k ⇒ a·b = (b·k)·b = b2 ·k = b·k = a ⇔ a ≤ b. � Proposition 9 For all a, b ∈A: (1) a · b = a (2) a+̃b = b (3) (1 + a)+̃b = 1 (4) a · (1 + b) = 0 are equivalent (an analogous result holds in any Boolean algebra). 4 The Boolean algebra isomorphism ϕ Let ∨, ∧, ¬, → denote the logic disjunction, conjunction, negation and implication, respectively. Let (C,∨,∧,¬,→) be the Boolean algebra of the propositions that can be constructed using a finite number of propositional variables P , Q, . . . , R. Let us denote tautology by 1 and contradiction by 0. Let us consider the Boolean algebra (A, +̃, · , 1+, “is a multiple”), where A is the residue class ring A = Z2[p,q, . . . ,r]/〈p2 −p,q2 −q, . . . ,r2 −r〉 24 An algebraic approach to rule based expert systems We define: ϕ: (C,∨,∧,¬,→) −→ (A, +̃, · , 1+, “is a multiple”) in the following way. For propositional variables P −→ p Q −→ q . . . . . . R −→ r and for any A, B ∈C A∨B −→ a+̃b ¬A −→ 1 + a Therefore, as an immediate consequence of the De Morgan laws: A∧B −→ a · b Moreover, as the (lattice) order can be obtained from the operations of the lattice, ϕ preserves the ordering. Proposition 10 ϕ is well defined. PROOF. For all propositions A, B: A ↔ B ⇒ A → B and B → A ⇒ ⇒ ϕ(A) is a multiple of ϕ(B) and ϕ(B) is a multiple of ϕ(A) ⇔ ⇔ a is a multiple of b and b is a multiple of a ⇔ a = b. � Corollary 2 ϕ is an order-preserving homomorphism, and ϕ(1) = 1,ϕ(0) = 0. Remark 1 In order to save space, we shall sometimes write b|a (“b divides a”) instead of “a is a multiple of b”. Proposition 11 ϕ is surjective. PROOF. The elements of the (Boolean) ring A (they are the same as those of the Boolean algebra A) are a linear algebraic combination of the polinomial variables, and: ϕ(P ∧Q) = p ·q ϕ((¬P ∧Q) ∨ (P ∧¬Q)) = p + q. � Proposition 12 ϕ is injective. PROOF. Let us suppose that for two propositions, A and B, we have: ϕ(A) = a, ϕ(B) = b and a = b. As a = b, we have: a|b and b|a. But ϕ preserves the ordering (i.e., → and “is a multiple” do correspond), so: b|a ⇒ A → B; a|b ⇒ B → A and therefore A ↔ B. � 25 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as Remark 2 To be precise, we really consider C/ ↔ and A/ = (Lindenbaum algebras) in order for the relations → and ≤ to be anti-symmetric. Example 8 Let P and Q be the propositional variables of C. Then C has 16 elements and (C,→) is the transitive closure of the diagram in Figure 2. It corresponds in ϕ with (A, “is a multiple”), where A = Z2[p,q]/〈p2 −p,q2 −q〉 and “is a multiple” is the transitive closure of the diagram in Figure 3. 0 � � �� @ @ @R �� �1 PPPq P ∧¬Q ¬P ∧¬Q P ∧ Q ¬P ∧ Q �� �� ���: - � � � � � ��* XXXXXXXz -XXXXXXXz -XXXXXXXz A A A AU B B B B BBN � �� � � � � � � ��� (P ∨¬Q) ∧ (¬P ∨ Q) ¬Q P Q ¬P (P ∧¬Q) ∨ (¬P ∧ Q) �� �� ���: - �� �� ���: XXXXXXXz �� �� ���: - - H H H H H HHj A A A AU B B B B B BN � �� � � � � � � � �� P ∨¬Q ¬P ∨¬Q P ∨ Q ¬P ∨ Q � � ��� ��1 PPPq @ @ @R 1 Figure 2. (C,→) when there are two propositional variables. 0 � � �� @ @ @R �� �1 PPPq p · (1 + q) (1+p)·(1+q) p · q (1 + p) · q �� ���1 - � � � ��3 PPPPPq - PPPPPq - PPPPPq J J J Ĵ A A A A A AU � � ��* � � � � � � �� 1 + q p q 1 + p 1 + (p + q) p + q �� ���1 - �� ���1 PPPPPq �� ���1 - - Q Q Q QQs @ @ @ @R J J J J J Ĵ � � � � � � � �� � � ��* p+̃(1 + q) (1 + p)+̃(1 + q) p+̃q (1 + p)+̃q � � ��� ��1 PPPq @ @ @R 1 Figure 3. (A, “is a multiple”) when there are two polynomial variables. 4.1 Ideals of a Boolean algebra and ideals of a ring Definition 8 Let (R, +, ·) be a commutative ring. A subset I ⊆ R is said to be an ideal of R if and only if: 1) ∀i, i′ ∈ I : i + i′ ∈ I 2) ∀i ∈ I,∀a ∈R : a · i ∈ I (i.e. I is a subring of A such that the product of an element of the ring by an element of the subring always belongs to the subring). 26 An algebraic approach to rule based expert systems Definition 9 Let (R, +, ·) be a commutative ring. Let S 6= ∅, S ⊆ R. The ideal generated by S = {p1, . . . ,pm}, denoted 〈p1, . . . ,pm〉, is the smallest ideal containing S. If S = {q}, the ideal generated by q, 〈q〉, is said to be a principal ideal, and results to be: 〈q〉 = {a ∈A : q|a} Despite the concept of ideal is usually associated with rings, ideals are also defined in Boolean algebras. In such case, the definition is different (based on the lattice order), and only ideals generated by a single element are considered. Definition 10 In the propositional Boolean algebra (C,∨,∧,¬,→), the (principal) ideal generated by Q is defined as EQ = {X ∈C : X → Q} Similarly, in the polynomial Boolean algebra (A, +̃, · , 1+, “is a multiple”), the (principal) ideal gener- ated by q ∈A is: Eq = {a ∈A : a is a multiple of q} Proposition 13 As ϕ preserves the ordering, the ideals of the Boolean algebra (C,∨,∧,¬,→) do corre- spond in ϕ with the ideals of the Boolean algebra (A, +̃, · , 1+, “is a multiple”). Proposition 14 The ideals of the Boolean algebra (A, +̃, · , “is a multiple”) are obviously ideals of the ring (A, +, · , “is a multiple”). Theorem 1 (A, +, ·) is a ring where all ideals are principal ones. PROOF. Let s1, s2, . . . , sn ∈A. We shall prove that 〈s1,s2, . . . ,sn〉 = 〈s1+̃s2+̃ · · ·+̃sn〉 in two steps: i) +̃ is an internal operation in the ideal (because + and · also are internal operations) and therefore: s1+̃s2+̃ · · ·+̃sn ∈ 〈s1,s2, . . . ,sn〉⇒ 〈s1+̃s2+̃ · · ·+̃sn〉⊆ 〈s1,s2, . . . ,sn〉 ii) That s1+̃s2+̃ · · ·+̃sn|si; i = 1, . . . , n can be easily proven. But, then: si ∈ 〈s1+̃s2+̃ · · ·+̃sn〉; i = 1, . . . , n and, consequently, the minimum ideal that contains {s1,s2, . . . ,sn}, is contained in 〈s1+̃s2+̃ · · ·+̃sn〉, i.e.: 〈s1,s2, . . . ,sn〉⊆ 〈s1+̃s2+̃ . . . +̃sn〉. � Remark 3 Let us observe that (A, +, ·) is not a “Principal Ideals Domain” because it is not an integrity domain (see Proposition 7). Corollary 3 As a consequence of Proposition 13, Proposition 14 and Theorem 1, the ideals of the polyno- mial Boolean algebra (A, +̃, · , “is a multiple”) are precisely the ideals of the polynomial ring (A, +, ·). 5 Introductory note about multi-valued propositional logics A simple introduction to the topic can be found in [37]. 27 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 5.1 Kleene’s and Łukasiewicz’s three-valued logic Kleene’s and Łukasiewicz’s multi-valued logics are two of the most commonly used multi-valued logics. We shall begin introducing these three-valued logics, in which a third truth value is considered. In Kleene’s logic this third valued is “undecided”. If an assertion is assigned the truth value “undecided” that means that at present, we can’t assign a truth-value to this conjecture (but the conjecture is either true or false). For instance, if we are reasoning in a medical diagnostic RBES, until we receive the result of a test, its truth value is “undecided”, but it will become either true or false in the future (when we receive the results). Kleene’s A → B is equivalent to ¬A∨B. Two modal unary connectives, necessary (�) and possible (♦) are usually considered in multi-valued logics. Kleene’s three valued logic truth-tables are detailed in Tables 1 and 2 (note that 0 represents “false”, 1 represents “undecided” and 2 represents “true”). It would also be possible to assign numbers to truth values in a different way. For instance, in Kleene’s three valued logic, it is also common to consider that 0 represents “false”, that 2 represents “undecided” and that 1 represents “true”. In such case the corresponding truth tables would obviously change. ¬ 0 2 1 1 2 0 � 0 0 1 0 2 2 ♦ 0 0 1 2 2 2 Table 1. Truth-tables of Kleene’s three-valued logic unary connectives. ∧ 0 1 2 0 0 0 0 1 0 1 1 2 0 1 2 ∨ 0 1 2 0 0 1 2 1 1 1 2 2 2 2 2 → 0 1 2 0 2 2 2 1 1 1 2 2 0 1 2 ↔ 0 1 2 0 2 1 0 1 1 1 1 2 0 1 2 Table 2. Truth-tables of Kleene’s three-valued logic binary connectives. In Łukasiewicz’s logic the third truth value considered is “indeterminate”. If a statement is considered to be indeterminate, it is neither true not false, but indeterminate in a metaphysical sense. All connectives have the same truth tables as in Kleene’s logic, except → and ↔ (see Table 3). → 0 1 2 0 2 2 2 1 1 2 2 2 0 1 2 ↔ 0 1 2 0 2 1 0 1 1 2 1 2 0 1 2 Table 3. Truth-tables of the Łukasewicz’s three-valued logic binary connectives whose truth-tables are different from Kleene’s ones. Remark 4 It is also possible to consider more than three truth values, so that the confidence in the truth- fulness of statements can be graded. Moreover, there are other multivalued logics aside from Kleene’s and Łukasewicz’s. 28 An algebraic approach to rule based expert systems 5.2 Tautological consequence Definition 11 A propositional formula A0 is a tautological consequence of the propositional formulae A1, A2, . . . , Am, denoted {A1,A2, . . . ,Am} |= A0, if and only if whenever the formulae A1,. . . ,Am hold (i.e., take the truth value “true”), then A0 also holds (i.e., takes the truth value “true”). Remark 5 Let us observe that: • in Boolean logic, there are formulae that can only take the truth value “false”, such as P ∧¬P , • in multi-valued logic, there are also formulae that can only take the truth value “false”, such as �P ∧�¬P . • in multi-valued logic, there are formulae that, although can take other truth values different from “false”, they can never take the truth value “true”, such as P ∧¬P (observe that not only �P ∧ �¬P |= P ∧¬P , but also P ∧¬P |= �P ∧�¬P ). 5.3 The isomorphism ϕ in the multi-valued case Let (C,∨,∧,¬,→) be a p-valued logic, where p is a prime number (we require p to be a prime in order for Zp to be a field) and C is the set of propositions that can be constructed using the propositional variables X1, X2, . . . , Xn. If in the residue class ring A = Zp[x1,x2, . . . ,xn]/〈x p 1 −x1,x p 2 −x2, . . . ,x p m −xm〉 adequate polynomial translations of the logical connectives are defined (so that they behave as suggested by the truth-tables), the natural correspondence ϕ, defined as in the Boolean case (Section 4), is also an isomorphism. 6 The main Theorem 6.1 A brief note about Gröbner bases In the early 60’s, both Heisuke Hironaka and Bruno Buchberger independently proved that, for each poly- nomial ideal, a basis completely identifying it always exists. They denoted their bases as standard bases and Gröbner bases (GB) [4, 5], respectively. The latter’s great advantage was that it provided a constructive method (Buchberger’s algorithm). Some GB are particularly important: we call them reduced Gröbner bases. We say that a Gröbner basis is reduced if and only if the leading coefficient of all its polynomials is 1 and we cannot simplify any of its polynomials by adding a linear algebraic combination of the rest of the polynomials in the basis. The input to Buchberger’s algorithm is a polynomial set, a term order (for instance, “total degree” or “pure lexicographical”), and a variable order (for instance, x > y > z) and its output is the ideal’s reduced GB with respect to the specified term and variable orders. The key point is that, once the term order and the variable order are fixed, such a reduced GB completely characterizes the ideal: any ideal has a unique reduced GB. As a consequence we have two key results: i) two sets of polynomials generate the same ideal if and only if their reduced GB are the same, ii) {1} is the only reduced GB for the ideal that is equal to the whole ring. An elementary introduction to the use of GB in algebraic system solving can be found in [33, 34]. For detailed introductions see, for instance, [1, 2, 9, 16, 38]. There are interesting applications in other fields like transportation engineering [32, 30]. 29 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 6.2 Connecting “To be a tautological consequence” with an ideal member- ship The main theorem is the following: Theorem 2 A logical formula, A0, is a tautological consequence of the set of formulae {A1,A2, . . . ,Am} in a certain p-valued modal logic (where p is a prime number) if and only if the polynomial translation of ¬A0 in A = Zp[x1,x2, . . . ,xn]/〈x p 1 − x1,x p 2 − x2, . . . ,x p m − xm〉, ϕ(¬A0), belongs to the ideal 〈ϕ(¬A1),ϕ(¬A2), . . . ,ϕ(¬Am)〉 of A. The key point is that this theorem can take advantage of the two important results stated at the end of the previous section. The proof of this theorem is really long and can be found in [19, 27]. 7 Some introductory notes about RBES A RBES consists of an “input”, an “output”, a “knowledge base” and an “inference engine”. A graphic user interface is frequently provided. The input of a RBES is concerned with the information related to the environment of the RBES. Since the environment may change, this information is also subject to change as times goes by. This information is described by means of a finite number of different atomic propositions: such propositions are termed as input atomic propositions. Input atomic propositions determine what we will call potential facts (see Definition 13). The output of a RBES is concerned with the information inferred by the RBES. It is described by means of a finite number of atomic propositions which we shall call output atomic propositions. The knowledge base (KB) of the RBES is concerned with the information contained in the system, which is used along with the input of the RBES to infer the output of the system. In a RBES based on propositional Boolean logic, this knowledge-base will be mostly represented by a finite set of rules. Remark 6 We shall use hereinafter the following notation: z may mean: • no symbol at all or “negation” (denoted ¬), if the underlying logic is classic Boolean, • no symbol at all, “negation”, “possibility” (denoted ♦), “necessity” (denoted �), or a combination of these symbols, like “necessarily-not” (denoted �¬), equivalent to ¬♦, if the underlying logic is a modal multi-valued one. 7.1 Rules Definition 12 The KB consists of a certain number of logical formulae of the form: ∧ki=1zX[i] −→∨ l j=1zX[j] known as production rules (and usually refereed to as, simply, rules). For example, formula X[1] ∧X[2] ∧¬X[3] ∧♦X[4] ∧¬X[5] −→ X[13] ∨�¬X[17] is read as: “IF X[1] and X[2] and not X[3] and possibly X[4] and not X[5] HOLD, THEN X[13] or necessarily-not X[17] HOLDS”. Let us underline that there is no constraint in the structure of the rules: the ocurrences of “∧” in the consequent and of “∨” in the antecedent of a rule can be avoided (if such symbols ocurred in a rule, the rule could be split into rules without these symbols). 30 An algebraic approach to rule based expert systems 7.2 Potential facts and facts. Integrity constraints Definition 13 A formula A is a potential fact if and only if A ≡ zX[i], where X[i] is an input atomic proposition. Definition 14 That a given set of facts (i.e., a subset of the set of potential facts that is stated as true) is consistent means that, for each propositional variable appearing in the potential facts of the system, X[k], at most one expression of the form zX[k], is chosen. That a given set of facts is maximal means that, for each propositional variable appearing in the potential facts of the system, X[k], exactly one expression of the form zX[k], is chosen. Definition 15 An integrity constraint (IC) is a piece of knowledge added by the experts and expressing that some potential facts cannot hold at the same time. For instance an IC could be: man ∧ pregnant. The negation of the ICs (sometimes denoted “NICs”) have to be added to the KB. 7.3 The inference engine. RBES inconsistency The inference engine (IE) is an automated tool (in our case, a program in the CAS), that verifies consistency and draws tautological consequences from the information contained in the KB. Definition 16 That a formula A be obtained by forward firing of the formulae in the union, G, of a con- sistent set of facts and the set of all production rules and the set of the negation of the integrity constraints of a RBES means that G |= A. Definition 17 A RBES is said to be inconsistent if there is a consistent set of facts verifying that con- tradiction can be obtained by forward firing of these facts and the rules and the negation of the integrity constraints in the KB. Otherwise the RBES is said to be consistent. 8 Knowledge extraction in RBES through the use of an alge- braic inference engine This extension of the algebraic method to RBES was already mentioned in [18] and detailed in [26]. 8.1 A polynomial model for the propositional Boolean algebra associated to RBES The Boolean algebra associated to a RBES whose underlying logic is classic Boolean logic is a structure (C∗,∨,∧,¬,→) where → is the relation obtained applying the rules of logical deduction to the implications of C and the rules, facts and the negations of the integrity constraints of the RBES (consequently, → is not the usual implication), and C∗ is the set of equivalence classes defined by this enlarged equivalence relation ↔ in C. If the rules Rule1, . . . , Rulev , the facts Fact 1, . . . , Fact m and the integrity constraints IC 1, . . . , IC u of a RBES are added as true to the Boolean algebra C of the previous section, the structure obtained is isomorphic to the image of A in the natural surjective homomorphism ψ : A−→A/J where J is the ideal 〈ϕ(¬Rule1), . . . ,ϕ(¬Rulev),ϕ(¬Fact 1), . . . ,ϕ(¬Fact m),ϕ(¬IC 1), . . . ,ϕ(¬IC u)〉 31 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 8.2 The main theorem adapted to knowledge extraction in RBES The main theorem detailed above can be adapted to treat the logic underlying a RBES: Theorem 3 A formula A0 can be obtained by forward firing of the formulae in the union of a consistent set of facts and the set of all production rules and the negation of the integrity constraints of a RBES, if and only if, the polynomial translation of the negation of A0 belongs to the ideal J of A = Zp[x1,x2, . . . ,xn]/〈x p 1− x1,x p 2 −x2, . . . ,x p n −xn〉, generated by the polynomial translations of the negations of the given potential facts, of the negations of all the production rules, and the integrity constraints of the RBES. Remark 7 Observe that, denoting I = 〈xp1 − x1,x p 2 − x2, . . . ,x p m − xn〉, we could alternatively check that A0 belongs to the ideal I + J of Zp[x1,x2, . . . ,xn]. We usually work this way as most CAS are not prepared to perform computations in a residue class ring. We have extensively used this approach in RBES design and verification. [20] is an interesting example in medicine (regarding techniques for coronary surgery), where we could find a situation that could lead to an inconsistency and was not taken into account by the panel of experts. 9 RBES strong and weak inconsistency. Algebraic interpreta- tion An introduction to verification can be found in [17]. This section summarizes [31]. We shall make the following distinction in the multi-valued case: Definition 18 We shall say that there is strong inconsistency if and only if the conjunction of the facts in a certain consistent set of facts, the rules and the negations of integrity constraints of the RBES, can only take the truth value “false”. Definition 19 We shall say that there is weak inconsistency if and only if all formulae can be obtained by forward firing of the formulae in the union of a certain consistent set of facts, the set of all production rules and the set of the negations of integrity constraints of the RBES. Proposition 15 Independently of the number of truth values (p) of the underlying logic: strong inconsistency ⇒ weak inconsistency. PROOF. If there is strong inconsistency, we have a consistent set of facts {Fact 1, . . . , Fact m} such that: Fact 1 ∧ . . .∧ Fact m ∧ Rule1 ∧ . . .∧ Rulev ∧ NIC 1 ∧ . . .∧ NIC u can only take the truth values false. But {Fact 1, . . . ,Fact m, Rule1, . . . , Rulev, NIC 1, . . . , NIC u} |= Fact 1 ∧ . . .∧ Fact m ∧ Rule1 ∧ . . .∧ Rulev ∧ NIC 1 ∧ . . .∧ NIC u so a contradiction can be obtained by forward firing of the formulae in {Fact 1, . . . , Fact m, Rule1, . . . , Rulev, IC 1, . . . , IC u}, and, therefore, we have weak inconsistency. � 32 An algebraic approach to rule based expert systems Proposition 16 If the underlying logic is Boolean: weak inconsistency ⇒ strong inconsistency. PROOF. If there is weak inconsistency, we have a consistent set of facts {Fact 1, . . . , Fact m} such that any formula follows of the formulae in {Fact 1, . . . , Fact m, Rule1, . . . , Rulev, NIC 1, . . . , NIC u} by forward firing. In particular, {Fact 1, . . . , Fact m, Rule1, . . . , Rulev, NIC 1, . . . , NIC u} |= contradiction But we are considering that the underlying logic is Boolean, so: Fact 1 ∧ . . .∧ Fact m ∧ Rule1 ∧ . . .∧ Rulev ∧ NIC 1 ∧ . . .∧ NIC u → contradiction and therefore Fact 1 ∧ . . .∧ Fact m ∧ Rule1 ∧ . . .∧ Rulev ∧ NIC 1 ∧ . . .∧ NIC u can only take the truth value “false”. � Remark 8 The previous result does not hold if the underlying logic is a multi-valued one. Example 9 Let us consider a small RBES whose KB consists of the rules: R1 :X[1] → X[3] R2 :X[2] →¬X[3] and whose underlying logic is Kleene’s three-valued logic. The potential facts are X[1] and X[2]. If we draw consequences from the maximal set of potential facts {X[1],X[2]}, we obtain X[3] ∧¬X[3]. As this formula can only take the truth values “undecided” or “false”, any formula can be obtained by forward firing of it, and, consequently, by forward firing of the formulae in {X[1],X[2],R1,R2}. Therefore we have weak inconsistency. Meanwhile, we cannot obtain a formula that is always “false” by forward firing of the conjunction of the facts in any consistent set of facts and R1 ∧R2. Therefore, we do not have strong inconsistency. Corollary 4 There is weak inconsistency if and only if 1 ∈ 〈f¬(ϕ(Rules)),f¬(ϕ(Facts)),f¬(ϕ(NICs))〉. But, what is the meaning of strong inconsistency in the polynomial model? Definition 20 The algebraic variety corresponding to the polynomial ideal I is the set of points of the respective affine space which satisfy all polynomials in the ideal I. We shall denote it by V (I). Lemma 1 For any positive prime integer p > 1, we have: i) There are non-zero ideals in Zp[x1,x2, . . . ,xm] whose algebraic variety is the whole affine space. ii) The only ideal in A = Zp[x1,x2, . . . ,xm]/I; I = 〈x p 1 −x1,x p 2 −x2, . . . ,x p m −xm〉 whose variety is the whole affine space is the zero ideal. PROOF. i) For example x · (x − 1) = x2 − x ∈ Z2[x] is not the null polynomial but its value is always 0 in {0, 1}. Therefore 〈x2 + x〉 verifies condition i). The same occurs with x · (x− 1) · (x− 2) ∈ Z3[x]. Its value in {0, 1, 2} is always 0. This construction can be obviously generalized to Zp[x]. 33 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as ii) Case I: Univariate polynomials. If 0, 1, . . . , p− 1 are roots of a polynomial in A = Zp[x]/〈xp −x〉, then the polynomial is either 0 or it can be factored as x · (x− 1) · · · · · (x− (p− 1)) · · · · Therefore, if this polynomial is different from 0, its total degree is greater or equal than p. But in A all polynomials are simplified to polynomials of degree lesser than p, so this polynomial simplifies to 0. Case II: Multi-variate polynomials. It can be proved by applying the previous case to the intersections of the hypersurface by hyperplanes parallel to the “vertical” axis. � Theorem 4 There is strong inconsistency if and only if V (〈ϕ(Rules ∧ Facts ∧ NICs)〉) is the whole (respective) affine space. PROOF. Rules ∧ Facts ∧ NICs can only take the truth value “false” ⇔ ⇔ polynomial ϕ(Rules ∧ Facts ∧ NICs) can only take the value 0 ⇔ ⇔ V (〈ϕ(Rules ∧ Facts ∧ NICs)〉) is the whole (respective) affine space. � But we can express the previous result in terms of ideals instead of algebraic varieties, what makes it easily decidable using Gröbner bases: Corollary 5 There is strong inconsistency if and only if 〈ϕ(Rules ∧ Facts ∧ NICs)〉 is the zero ideal of A = Zp[x1,x2, . . . ,xm]/I. 10 Backward reasoning in rule based expert systems whose underlying logic is classic Boolean logic Details on the topic of this section, our most recent result, can be found in [25]. Based on the classical works of automatic discovery of geometric theorems [14, 23] (a topic already advanced in [8]), we have now adapted this idea to a specific task in RBES whose underlying logic is classic Boolean logic: given a set of facts and a goal, finding a new fact such that if it also held, this goal would be inferred. The main result proven in [25] (Theorem 5 below) shows how it is enough to compute one single reduced GB in order to detect them: Theorem 5 Let A1, . . . , Ak, B ∈C∗ be formulae such that {A1, . . . ,Ak} 6|= B. Let C be a potential fact. Let G be the reduced GB of the ideal 〈ϕ(¬A1), . . . ,ϕ(¬Ak),ϕ(B)〉 + I. We have that: {A1, . . . ,Ak,C} |= B ⇔ ϕ(C) ∈ G Observe that the membership relation in the right hand side of the equivalence refers to a membership to the set of polynomials in the reduced GB, not to the ideal generated by that basis. 11 Implementation in a Computer Algebra System As most CAS include an implementation of GB and of “normal forms” (reductions modulo an ideal), to develop an inference engine in a CAS following the method detailed above (based on checking polynomial ideal memberships) is straightforward and surprisingly brief. 34 An algebraic approach to rule based expert systems 11.1 Polynomial translation of the logical connectives For instance, the polynomial expressions corresponding to the basic logical formulae in Łukasiewicz’s three-valued logic (if 2 is assigned to “true”, 1 to “undetermined” and 0 to “false”) are detailed afterwards. • ¬M is translated into the polynomial: 2 −m • ♦M is translated into the polynomial: 2m2 • �M is translated into the polynomial: m2 + 2m • M ∨N is translated into the polynomial: m2n2 + m2n + mn2 + 2mn + m + n • M ∧N is translated into the polynomial: 2m2n2 + 2m2n + 2mn2 + mn • M → N is translated into the polynomial: 2m2n2 + 2m2n + 2mn2 + mn + 2m + 2 • M ↔ N is translated into the polynomial: m2n2 + m2n + mn2 + 2mn + 2m + 2n + 2 (note that the coefficients of these polynomials belong to Z3). The polynomials corresponding to the different logics can be obtained by solving an algebraic system that is obtained from the truth tables. 11.2 CoCoA implementation These polynomial translations can be input almost directly in CoCoA 4.3 [6, 35], which provides a “Normal Form” command, NF(pol,I), that returns the reduction of polynomial pol modulo ideal I. We have to define the polynomial ring and ask CoCoA to use it: USE Z/(3)[x[1..15]]; and then we can define the ideal MEMORY.I, generated by the expressions x3i −xi (denoting it this way, it is a global variable, and it can be an input to NF): MEMORY.I:=Ideal([x[K_]ˆ3-x[K_] | K_ In 1..15]); and introduce the polynomial expressions for the three-valued connectives of Łukasiewicz modal logic (the operators are prefix ones)1: NEG(M):=NF(2-M,MEMORY.I); POS(M):=NF(2*Mˆ2,MEMORY.I); NEC(M):=NF(Mˆ2+2*M,MEMORY.I); OR(M,N):=NF(Mˆ2*Nˆ2+Mˆ2*N+M*Nˆ2+2*M*N+M+N,MEMORY.I); AND(M,N):=NF(2*Mˆ2*Nˆ2+2*Mˆ2*N+2*M*Nˆ2+M*N,MEMORY.I); IMP(M,N):=NF(2*Mˆ2*Nˆ2+2*Mˆ2*N+2*M*Nˆ2+M*N+2*M+2,MEMORY.I); IFF(M,N):=NF(Mˆ2*Nˆ2+Mˆ2*N+M*Nˆ2+2*M*N+2*M+2*N+2,MEMORY.I); For instance, the CoCoA implementation of the polynomial expressions of a 7-valued modal logic can be found in [27]. Whether A0 is a consequence of a set of formulae or not, can be checked with CoCoA just typing: NF(NEG(A[0]),MEMORY.I+J); (where J is the polynomial ideal generated by the polynomial expressions of the negation of the formulae in the given set). If the output of the command is 0, the answer is “yes”, otherwise the answer is “no”. 1In CoCoA 4.7 these operators would be defined using Define...EndDefine instead of directly :=. 35 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as It can also be directly checked using the Boolean command IsIn, that tests ideal memberships, by typing: NEG(A[0]) IsIn MEMORY.I+J; Whether the RBES is inconsistent or not can also be checked using Gröbner bases by typing in CoCoA: GBasis(MEMORY.I+J); If the output is 1, the RBES is inconsistent; otherwise (the output is normally a large set of polynomials), that set of facts doesn’t lead to an inconsistency. Using IsIn makes it even simpler, as it directly returns “true” or “false”, just typing: 1 IsIn MEMORY.I+J; 11.3 Maple implementation Let us implement the model in Maple. We have developed different implementations in Maple, but the one presented afterwards is the most modern and fastest [29]. In Maple (version ≥ 10) it is possible to define a polynomial ring over a finite field using Maple’s Ore Algebra. In this way, the NormalForm command is computed in such ring. We load the Gröbner and Ore algebra packages first, and then define the list of variables, the polynomial ring and the order that will be used by the GB-related commands: with(Groebner): with(Ore_algebra): SV:=x[1],x[2],x[3]: A:=poly_algebra(SV,characteristic=3): Orde:=MonomialOrder(A,’plex’(SV)): and the ideal I (denoted iI, as I is a reserved word in Maple), using map in order to save time: fu:=v->vˆ3-v: iI:=map(fu,[SV]); 3 3 3 iI := [x[1] - x[1], x[2] - x[2], x[3] - x[3]] We can now define the functions that associate to the logical connectives their polynomial expressions, as done in CoCoA above. NEG :=(m::algebraic) -> NormalForm(2-m,iI,Orde): NEC :=(m::algebraic) -> NormalForm(expand(mˆ2+2*m),iI,Orde): POS :=(m::algebraic) -> NormalForm(expand(2*mˆ2),iI,Orde): ‘&AND‘:=(m::algebraic,n::algebraic) -> NormalForm(expand(2*mˆ2*nˆ2+2*mˆ2*n+2*m*nˆ2+m*n), iI,Orde): ‘&OR‘ :=(m::algebraic,n::algebraic) -> NormalForm(expand(mˆ2*nˆ2+mˆ2*n+m*nˆ2+2*m*n+m+n), iI,Orde): ‘&IMP‘ :=(m::algebraic,n::algebraic) -> NormalForm(expand(2*mˆ2*nˆ2+2*mˆ2*n+2*m*nˆ2+m*n+2*m+2), iI,Orde): ‘&IFF‘ :=(m::algebraic,n::algebraic) -> NormalForm(expand(mˆ2*nˆ2+mˆ2*n+m*nˆ2+2*m*n+2*m+2*n+2), iI,Orde): 36 An algebraic approach to rule based expert systems 12 Description of the shell A GUI was presented at FLINS’2008 conference [24], and a complete shell is now provided. When working with the shell we could distinguish three levels. At the lower level, we provide the CAS code for performing knowledge extraction and verification in different logics (i.e., we provide the algebraic inference engines). For instance, to work with the CAS CoCoA 4.3 in Boolean logic with the propositional variables x1, . . . , x5, the code that contains the algebraic model of the IE should be included in the file named CODE CoCoA.TXT. It would look like: USE Z/(2)[x[1..5]]; MEMORY.I:=Ideal([x[K_]ˆ2-x[K_] | K_ In 1..5]); NEG(M):=NF(1+M,MEMORY.I); OR(M,N):=NF(M+N+M*N,MEMORY.I); ... (similar to the code shown in Section 11.2). Some lines regarding I/O between the GUI and the CAS follow. At the intermediate level, the RBES developer details the concrete potential facts, rules and integrity constraints of a certain RBES (the code must be stored in file RBES.TXT). Example 10 For instance, the production rules and IC of a tiny RBES whose underlying logic is classic Boolean and their CoCoA translations can be (note that in a real RBES there would be dozens of rules): R1 : x[1] → x[2] R1:=IMP(x[1],x[2]); R2 : x[2] ∧¬x[3] → x[4] R2:=IMP(AND(x[2],NEG(x[3])),x[4]); R3 : x[1] ∧x[4] → x[5] R3:=IMP(AND(x[1],x[4]),x[5]); NIC 1 : ¬(x[3] ∧x[5]) NIC1:=NEG(AND(x[3],x[5])); At the upper level, the final user does not have to deal directly with the CAS but with a simple GUI. Both the number of truth values of the logic (e.g., 3) and the desired logic (e.g., Kleene’s) are to be chosen beforehand. Then consistency can be checked and knowledge extraction performed. To do so, a list formed by a proposition and a (consistent) list of facts has to be introduced. The GUI calls the CAS and checks firstly if they lead to inconsistency. Afterwards, it checks if the given proposition follows from the given set of facts. Example 11 Let us consider the tiny RBES of Example 10. If the list of given facts is [x[1],¬x[3]], an inconsistency is not obtained. Moreover, x[4] ∧x[5] follows from them (see Figure 4). This GUI has been implemented in Visual-BASIC and runs on computers using the most common windows-based operating system. Porting it to other operating systems is under consideration now. It can call (at least) the mathematical systems: CoCoA, Maple, Maxima, Singular, Risa-Asir and Octave. 13 Conclusions We believe that this algebraic approach is flexible and powerful (as several logics can be easily and effec- tively simulated) and that the whole shell now put together can be really useful for teaching and for quick RBES designing. This approach has the advantage over that of Kapur-Narendran [15] and Hsiang [13] for Boolean logic (or Alonso et al. [3] and Chazarain et al. [7] in the multi-valued case) of providing an algebraic structure that is isomorphic to the propositional algebra, instead of only performing effective com- putations in logic using Gröbner bases. This is the key for going one step further and obtaining an algebraic structure isomorphic to a RBES based on one of these logics, i.e., from A = Zp[x,y, . . . ,z]/〈xp −x,yp −y, . . . ,zp −z〉 37 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as Figure 4. Screenshot of a GUI for RBES consistency checking and knowledge extraction. to A/J where J is the ideal generated by the negation of the facts, rules and integrity constraint, so that knowledge extraction and verification of RBES can be performed. Acknowledgement. This work was partially supported by the research projects TIN2006-06190 (Government of Spain) and UCM2008-910563 (UCM - BSCH Gr. 58/08, Research Group ACEIA, Spain). We would also like to thank the anonymous referees for their most valuable comments and sugges- tions. References [1] ADAMS, W. W. AND LOUSTAUNAU, P., (1994). An Introduction to Gröbner Bases. Graduate Studies in Mathe- matics 3, AMS. [2] AKRITAS, A. G., (1989). Elements of Computer Algebra with Applications, Wiley - Interscience. [3] ALONSO, J. A. AND BRIALES, E., (1989). Lógicas Polivalentes y Bases de Gröbner, en M. Vide ed., Procs. of the V Congress on Natural Languages and Formal Languages, Barcelona Univ. Press, 307–315. 38 An algebraic approach to rule based expert systems [4] BUCHBERGER, B., (1988). Applications of Gröbner Bases in non-linear Computational Geometry. In J. R. Rice (editor), Mathematical Aspects of Scientific Software, IMA Volumes in Math. and its Applications, 14, Springer- Verlag. [5] BUCHBERGER, B., (2006). Bruno Buchberger’s PhD thesis 1965: An algorithm for finding the basis elemen- tals of the residue class ring of a zero dimensional polynomial ideal, J. Symbolic Comput., 41, 3-4, 475–511. DOI: 10.1016/j.jsc.2005.09.007 [6] CAPANI, A. AND NIESI, G., (1996). CoCoA User’s Manual (v. 3.0b), Depto. de Matemáticas, Universidad de Genova. [7] CHAZARAIN, J.; RISCOS, A.; ALONSO, J. A. AND BRIALES, E., (1991). Many-valued Logic and Gröbner Bases with Applications to Modal Logic, J. Symbolic Comput., 11, 181–194. [8] CHOU, S.-C., (1987). Mechanical Geometry Theorem Proving. Kluwer Academic Publishers. [9] COX, D.; LITTLE, J. AND O’SHEA, D., (1992). Ideals,varieties, and algorithms, Springer-Verlag. [10] HALMOS, P. R., (1974). Lectures on Boolean Algebras, Springer-Verlag. [11] HALMOS, P. AND GIVANT, S., (1998). Logic as Algebra, The Mathematical Asociation of America. [12] HERMES, H., (1963). La teorı́a de retı́culos y su aplicación a la lógica matemática, Conf. Mat. VI., CSIC- Madrid. [13] HSIANG, J., (1985). Refutational Theorem Proving using Term-rewriting Systems, Artificial Intelligence, 25, 255–300. DOI: 10.1016/0004-3702(85)90074-8 [14] KAPUR, D. AND MUNDY, J. L., (1989). Wu’s method and its application to perspective viewing. In D. Kapur and J. L. Mundy, eds., Geometric Reasoning. MIT Press, Cambridge MA, 15–36. [15] KAPUR, D. AND NARENDRAN, P., (1984). An Equational Approach to Theorem Proving in First-Order Predicate Calculus, 84CRD296, General Electric Corporate Research and Development Report (Schenectady, NY, March 1984, rev Dec 1984). Also in A. K. Joshi, ed., Proceedings of IJCAI-85. Morgan Kaufmann, 1146–1153. [16] KREUZER, M. AND ROBBIANO, L., (2000). Computational Commutative Algebra. Springer-Verlag. [17] LAITA, L. M. AND DE LEDESMA, L., (1997). Knowledge-Based Systems Verification. In Kent, J. G. Williams (editors), Encyclopedia of Computer Science and Technology. Marcel Dekker, 253–280. [18] LAITA, L. M.; DE LEDESMA, L.; ROANES-LOZANO, E. AND ROANES-MACÍAS, E., (1995). An Interpretation of the Propositional Boolean Algebra as a k-algebra. Effective Calculus. In J. Campbell, J. Calmet (editors), Proceedings of AISMC-2, Springer-Verlag LNCS 958, 255–263. [19] LAITA, L. M.; ROANES-LOZANO, E.; DE LEDESMA, L. AND ALONSO, J. A., (1999). A Computer Approach to Verification and Deduction in Many-valued Knowledge Systems, Soft Comput., 3, 1, 7–19. DOI: 10.1007/s005000050086 [20] LAITA, L. M.;ROANES-LOZANO, E.; MAOJO, V.; DE LEDESMA, L. AND LAITA, L., (2001). An expert system for managing medical appropriateness criteria based on computer algebra techniques, Comput. Math. Appl., 42, 5, 12, 1505–1522. DOI: 10.1016/S0898-1221(01)00258-9 [21] MENDELSON, E., (1970). Theory and Problems of Boolean Algebras and Switching Circuits, Schaum’s, MacGraw-Hill. [22] MONK, D., (1989). Handbook of Boolean Algebras, North-Holland. [23] RECIO, T. AND VÉLEZ, M. P., (1999). Automatic Discovery of Theorems in Elementary Geometry, J. Automat. Reason., 23, 63–82. [24] ROANES-LOZANO, E.; HERNANDO, A.; LAITA, L. M. AND ROANES-MACÍAS, E., (2008). A Shell for Rule- Based Expert Systems Development Using Groebner Bases-Based Inference. In D. Ruan, J. Montero, J. Lu, L. Martı́nez, P. D’Hondt, E. E. Kerre (editors), Computational Intelligence in Decision and Control. Proceedings of the 8th International FLINS Conference, World Scientific Proceedings Series on Computer Engineering and Information Science 1, 769–774. DOI: 10.1142/9789812799470 0126 39 http://dx.doi.org/10.1016/j.jsc.2005.09.007 http://dx.doi.org/10.1016/0004-3702(85)90074-8 http://dx.doi.org/10.1007/s005000050086 http://dx.doi.org/10.1016/S0898-1221(01)00258-9 http://dx.doi.org/10.1142/9789812799470_0126 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as [25] ROANES-LOZANO, E.; HERNANDO, A.; LAITA, L. M. AND ROANES-MACÍAS, E., A Groebner Bases- based Approach to Backward Reasoning in Rule Based Expert Systems, Ann. Math. Artif. Intell., to appear. DOI: 10.1007/s10472-009-9147-4 [26] ROANES-LOZANO, E.; LAITA, L. M. AND ROANES-MACÍAS, E., (1995). Maple V in A.I., The Boolean Alge- bra Associated to a KBS, CAN Nieuwsbrief, 14, 65–70. [27] ROANES-LOZANO, E.; LAITA, L. M. AND ROANES-MACÍAS, E., (1998). A Polynomial Model for Many- valued Logics with a Touch of Algebraic Geometry and Computer Algebra, Math. Comput. Simulation, 45, 1 83–99. [28] ROANES-LOZANO, E.; LAITA, L. M. AND ROANES-MACÍAS, E., (2003). Cálculos efectivos en Lógica Proposi- cional Booleana interpretada como un anillo de clases residuales (polinomial) sobre Z2, Boletı́n de la Sociedad “Puig Adam” de Profesores de Matemáticas, 65, 17–42. [29] ROANES-LOZANO, E.; LAITA, L. M. AND ROANES-MACÍAS, E., (2008). A Groebner Bases Based Many- valued Modal Logic Implementation in Maple. In S. Autexier et al. (editors), AISC / Calculemus / MKM 2008, LNAI 5144 Springer-Verlag, 170–183. DOI: 10.1007/978-3-540-85110-3 14 [30] ROANES-LOZANO, E.; MUGA, R.; LAITA, L. M. AND ROANES-MACÍAS, E., (2005). A terminal area topology-independent GB-based conflict detection system for A-SMGCS, RACSAM Rev. R Acad. Cienc. Exactas, Fı́s. Nat. Ser. A Mat., 98, 1-2, 229–237. [31] ROANES-LOZANO, E.; ROANES-MACÍAS, E. AND LAITA, L. M., (1999). Geometric Interpretation of Strong Inconsistency in Knowledge Based Systems. In V. G. Ganzha, E. W. Mayr, E. V. Vorozhtsov (editors), Computer Algebra in Scientific Computing. Proceedings of CASC’99). Springer-Verlag, 349–363. [32] ROANES-LOZANO, E.; ROANES-MACÍAS, E. AND LAITA, L. M., (2000). Railway Interlocking Systems and Groebner Bases, Math. Comput. Simulation, 51, 5, 473–481. DOI: 10.1016/S0378-4754(99)00137-8 [33] ROANES-LOZANO, E.; ROANES-MACÍAS, E. AND LAITA, L. M., (2004). The Geometry of Algebraic Sys- tems and Their Exact Solving Using Groebner Bases, Computing in Science and Engineering, 6, 2, 76–79. DOI: 10.1109/MCISE.2004.1267612 [34] ROANES-LOZANO, E.; ROANES-MACÍAS, E. AND LAITA, L. M., (2004) Some Applications of Gröbner Bases, Computing in Science and Engineering, 6, 3, 56–60. DOI: 10.1109/MCISE.2004.1289309 [35] ROBBIANO, L. ET AL., (2002). CoCoA 4.2 Online Help, (electronic file companying CoCoA v.4.2). [36] STONE, M., (1940). The Theory of Representations for Boolean Algebras, Trans. Amer. Math. Soc., 40, 1, 37– 111. DOI: 10.2307/1989664 [37] TURNER, R., (1984). Logics for Artificial Intelligence, Ellis Horwood. [38] WINKLER, F., (1996). Polynomial Algorithms in Computer Algebra. Springer-Verlag. Eugenio Roanes-Lozano Luis M. Laita Depto. de Álgebra Depto. CC. de la Computación e I.A. Facultad de Educación Facultad de Informática Universidad Complutense de Madrid Universidad Politécnica de Madrid c/ Rector Royo Villanova s/n Campus de Montegancedo 28040-Madrid Boadilla del Monte, 28660-Madrid Spain Spain eroanes@mat.ucm.es laita@fi.upm.es Antonio Hernando Eugenio Roanes-Macı́as Depto. de Sistemas Inteligentes Aplicados Depto. de Álgebra Escuela Universitaria de Informática Facultad de Educación Universidad Politécnica de Madrid Universidad Complutense de Madrid Carretera de Valencia km 7 c/ Rector Royo Villanova s/n 28031-Madrid 28040-Madrid Spain Spain ahernando@eui.upm.es roanes@mat.ucm.es 40 http://dx.doi.org/10.1007/s10472-009-9147-4 http://dx.doi.org/10.1007/978-3-540-85110-3_14 http://dx.doi.org/10.1016/S0378-4754(99)00137-8 http://dx.doi.org/10.1109/MCISE.2004.1267612 http://dx.doi.org/10.1109/MCISE.2004.1289309 http://dx.doi.org/10.2307/1989664 Introduction An introduction to Boolean algebras and Boolean rings Lattices and lattice orders Boolean algebras Boolean rings A polynomial model for propositional logic Building the Taylor-Made Polynomial (Boolean) Ring The polynomial Boolean algebra corresponding to the polynomial (Boolean) ring The Boolean Algebra Isomorphism varphi Ideals of a Boolean algebra and ideals of a ring Introductory note about multi-valued propositional logics Kleene's and Łukasiewicz's three-valued logic Tautological consequence The isomorphism varphi in the Multi-Valued Case The main Theorem A brief note about Gröbner bases Connecting ``To be a tautological consequence'' with an ideal membership Some introductory notes about RBES Rules Potential facts and facts. Integrity constraints The inference engine. RBES inconsistency Knowledge extraction in RBES through the use of an algebraic inference engine A polynomial model for the propositional Boolean algebra associated to RBES The main theorem adapted to knowledge extraction in RBES RBES strong and weak inconsistency. Algebraic interpretation Backward reasoning in rule based expert systems whose underlying logic is classic Boolean logic Implementation in a Computer Algebra System Polynomial translation of the logical connectives CoCoA implementation Maple implementation Description of the shell Conclusions 
work_lpfqcnje75gvjh7fqjryfk2gou	----	[PDF] A connectionist expert system for fault diagnosis | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1016/0378-7796(92)90076-D Corpus ID: 36916022A connectionist expert system for fault diagnosis @article{Chen1992ACE, title={A connectionist expert system for fault diagnosis}, author={J. Chen and C. Chen}, journal={Electric Power Systems Research}, year={1992}, volume={24}, pages={99-103} } J. Chen, C. Chen Published 1992 Engineering Electric Power Systems Research Abstract The design of a connectionist expert system for the fault diagnosis of a distribution system is proposed. The heuristic rules acquired from distribution dispatchers' experiences are incorporated into the connectionist network which can be utilized as the expert system knowledge base. The data sets used to train the connectionist network are the current waveforms recorded upon the occurrence of miscellaneous faults. The interconnection weights of this network are adjusted by the back… Expand View via Publisher ntur.lib.ntu.edu.tw Save to Library Create Alert Cite Launch Research Feed Share This Paper 6 CitationsHighly Influential Citations 1 Background Citations 1 Methods Citations 1 View All Figures and Tables from this paper figure 2 table 2 figure 3 figure 4 figure 5 figure I View All 6 Figures & Tables 6 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency A new connectionist expert system scheme for distribution system fault diagnosis N. Rao, J. Chen, W. C. Chan Engineering 1994 Proceedings of Canadian Conference on Electrical and Computer Engineering 1994 2 Save Alert Research Feed Intelligent Approach for Fault Diagnosis in Power Transmission Systems Using Support Vector Machines R. B., Thukaram Dhadbanjan, H. Khincha Engineering 2007 8 Save Alert Research Feed A symptom-driven fuzzy system for isolating faults J. Chen, R. Tsai, H.-W. Tzeng Computer Science [Proceedings] 1992 IEEE International Conference on Systems, Man, and Cybernetics 1992 3 Save Alert Research Feed Online expert systems for fault diagnosis in technical processes C. Angeli Computer Science Expert Syst. J. Knowl. Eng. 2008 32 Save Alert Research Feed Object-oriented expert systems for fault diagnosis Shao-Hung Chang, Muh-Shenq Lin, Hung-Fa Sun, J. Chen Computer Science Proceedings of IEEE Systems Man and Cybernetics Conference - SMC 1993 1 Save Alert Research Feed Hybrid expert systems: A survey of current approaches and applications Seda Sahin, M. Tolun, R. Hassanpour Computer Science Expert Syst. Appl. 2012 112 Highly Influenced PDF View 11 excerpts, cites methods and background Save Alert Research Feed References SHOWING 1-10 OF 17 REFERENCES SORT BYRelevance Most Influenced Papers Recency A logic programming approach to fault diagnosis in distribution ring networks K. Wong, C. Tsang Engineering 1988 18 Save Alert Research Feed A neural network approach to the detection of incipient faults on power distribution feeders S. Ebron, D. Lubkeman, M. W. White Engineering 1990 195 Save Alert Research Feed An Expert System for Fault Section Estimation Using Information from Protective Relays and Circuit Breakers C. Fukui, Junzo Kawakami IEEE Transactions on Power Delivery 1986 251 Save Alert Research Feed Artificial Neural-Net Based Dynamic Security Assessment for Electric Power Systems D. Sobajic, Y. Pao Engineering IEEE Power Engineering Review 1989 397 Save Alert Research Feed Fault diagnosis for an HVDC system: a feasibility study of an expert system application P. Kalra Engineering 1988 10 Save Alert Research Feed Development of a Knowledge Based System for Power System Restoration T. Sakaguchi, K. Matsumoto Engineering IEEE Power Engineering Review 1983 231 Save Alert Research Feed Neural-net based real-time control of capacitors installed on distribution systems N. Santoso, O. Tan Engineering 1990 225 Save Alert Research Feed The Operator's Assistant--An Intelligent, Expandable Program for Power System Trouble Analysis S. N. Talukdar, E. Cardozo, Ted Perry Computer Science IEEE Transactions on Power Systems 1986 107 Save Alert Research Feed Distribution High Impedance Fault Detection Utilizing High Frequency Current Components B. M. Aucoin, B. Russell Engineering IEEE Power Engineering Review 1982 123 Save Alert Research Feed Connectionist expert systems S. I. Gallant Computer Science CACM 1988 689 Save Alert Research Feed ... 1 2 ... Related Papers Abstract Figures and Tables 6 Citations 17 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_lpsq3va57netlcnxnhp4pcz2zq	----	FACULTEIT ECONOMIE EN BEDRIJFSKUNDE   TWEEKERKENSTRAAT 2 B-9000 GENT Tel. : 32 - (0)9 – 264.34.61 Fax. : 32 - (0)9 – 264.35.92 WORKING PAPER   Mining Ideas from Textual Information Dirk Thorleuchter1 Dirk Van den Poel2 Anita Prinzie3 November 2009 2009/619 hofer.de 1 Fraunhofer INT, Appelsgarten 2, 53879 Euskirchen, Germany & PhD Candidate, Ghent University, dirk.thorleuchter@int.fraun 2 Prof. Dr. Dirk Van den Poel, Professor of Marketing Modeling/analytical Customer Relationship Management, Faculty of Economics and Business Administration, dirk.vandenpoel@ugent.be; more papers about customer relationship management can be obtained from the website: www.crm.UGent.be, more papers about text mining can be downloaded from www.textmining.UGent.be 3 Prof. Dr. Anita Prinzie is visiting professor at Ghent University D/2009/7012/71 mailto:dirk.thorleuchter@int.fraunhofer.de mailto:dirk.vandenpoel@ugent.be http://www.crm.ugent.be/ http://www.textmining.ugent.be/ Mining Ideas from Textual Information Dirk Thorleuchter1, Dirk Van den Poel2, and Anita Prinzie2 1Fraunhofer INT, Appelsgarten 2, 53879 Euskirchen, Germany 2Ghent University, Faculty of Economics and Business Administration, Tweekerkenstraat 2, 9000 Gent, Belgium Abstract This approach introduces idea mining as process of extracting new and useful ideas from unstructured text. We use an idea definition from technique philosophy and we focus on ideas that can be used to solve technological problems. The rationale for the idea mining approach is taken over from psychology and cognitive science and follows how persons create ideas. To realize the processing, we use methods from text mining and text classification (tokenization, term filtering methods, Euclidean distance measure etc.) and combine them with a new heuristic measure for mining ideas. As a result, the idea mining approach extracts automatically new and useful ideas from a user given text. We present these problem solution ideas in a comprehensible way to support users in problem solving. This approach is evaluated with patent data and it is realized as a web-based application, named 'Technological Idea Miner' that can be used for further testing and evaluation. Keywords Idea Mining, Text Mining, Text Classification, Technology Introduction Overview  An idea is an image existing or formed in the mind but it can be written down as textual information. In the last years, we see a continually increasing amount of information. About 80 % of all this information is stored in textual form [9]. Examples are research papers, articles in technical periodicals, reports, documents, web pages etc. These texts possibly contain many new ideas. A new idea is often needed to discover unconventional approaches e.g. to create a technological breakthrough. However, a manual extraction of new ideas from these masses of texts is time consuming and costly. Therefore, it is useful to search for new problem solution ideas automatically. Text mining or knowledge discovery from texts refers generally to the process of extracting interesting information and knowledge from unstructured text [12]. Referring to this, we introduce idea mining as an automatically process of extracting new and useful ideas from unstructured text using text-mining methods. 2 Creating ideas is a well-known topic that is related to psychology and cognitive science. There, we find many approaches dealing with how persons create ideas especially for problem solution. Therefore, in Sect 2 we focus on a general process of creating problem solution ideas and use it as rationale for the idea mining approach. In recent years, data and text mining techniques explore and analyze huge amounts of available textual data [4]. Idea mining uses known methods from these techniques and combine them with a new method to create text patterns and a new heuristic measure for mining ideas to realise the rationale. Therefore, we present the processing of the idea mining approach in Sect. 3 and we introduce this new idea mining measure in Sect. 4. A further task of idea mining is to present the extracted ideas in a comprehensible way to the user. Therefore, we focus on results of comprehensibility research and their relations to our task (see Sect. 5). Additionally, we provide an extensive evaluation to show the success of the idea mining approach and specifically the heuristic idea mining measure (see Sect. 6). Idea Definition  We limit our approach to the technological language because of two reasons. Firstly, the technological language is much more standardized than the colloquial language [11,16]. Therefore, we get better results by analyzing technological texts with text mining approaches. Secondly, our idea definition is taken over from technique philosophy [21]. There, an idea is defined as a combination of two things: a mean and an appertaining purpose. An example for an idea is a transistor. A transistor is a semiconductor device. It can be used to amplify or switch electronic signals. Here, we have a mean (a semiconductor device) and an appertaining purpose (to amplify or switch electronic signals). In general, we talk about a new idea if a know mean is related to an unknown purpose or if a known purpose is related to an unknown mean [1]. Then, a new idea is a nanomagnet because a nanomagnet is a miniaturized magnet that also can be used to amplify or switch electronic signals. Here we have an unknown mean (a miniaturized magnet) appearing together with a known purpose. This new idea could be useful to humans who are working in the field of electronic signals because in future nanomagnetic technology possibly could replace transistor technology. Therefore, we define a new and probably useful idea as a text phrase. This text phrase consists of domain specific terms that occur together in textual information. These terms can be divided up into two subsets. The first subset should represent a known mean (or a known purpose) and the second subset should represent an unknown purpose (or an unknown mean). Additionally, all terms in the first subset should occur together in a text phrase of the technological problem description. 3 Rationale behind Idea Mining Creating ideas is a well-known topic that is related to creativity in psychology and cognitive science. One of the first descriptions of the creative process was published by Wallas [23]. His stage model explains creative insights and illuminations for finding a problem solution. This model consists of a four stages process. In stage one 'preparation', the problem is analyzed so that a person recognizes the problem's dimensions. The stage two 'incubation / intimation' and the stage three 'illumination' transfer the problem from the conscious to the unconscious mind. The unconscious mind works on the problem continuously and it probably finds a solution by creative insights and illuminations. This solution is transferred to the conscious mind, which means after some time the person suddenly gets an idea that is new for him and that probably solves the problem. In the last stage 'verification', the idea is tested for novelty and usefulness. One of the best-known pragmatic approaches of using practical creativity is brainstorming from Osborn [17]. The first step in brainstorming is to define the problem e.g. by creating descriptions of the problem. Then, persons generate new ideas using creativity methods like idea association etc. The last step in the brainstorming process is to cluster the generated ideas and to evaluate it for novelty and usefulness. Beside this, there are several further approaches dealing with the creation of new ideas. We can learn from all these approaches that for creating ideas three steps are necessary. The first step is to focus on a problem, the second step is to generate some new ideas specific for this problem with creative methods and the third step is to evaluate the generated ideas for novelty and usefulness concerning the problem. Referring to these approaches, we build an adequate rationale for the idea mining process. Therefore, idea mining also consists of three steps. In the first step, we focus on the problem. Here, the user of our idea mining approach has to provide textual information where he describes his specific problem (a problem description). In the second step, the user has to provide further textual information where he supposes the existence of new and useful ideas (a new text) that probably can solve his problem [20]. Ideas are contained in text phrases inside this new text as described in Sect. 1.2. Therefore, with an automatically process, we extract a very large number of overlapping text phrases from the new text. In the remainder of this paper, text phrases will be named text patterns. In the third step, all extracted text patterns are evaluated for novelty and usefulness. This means, they are compared to the problem description by using a specific idea mining measure. With this measure, text patterns can be classified as new and useful idea. Therefore, idea mining identifies new and useful ideas in three steps: 1. Preparation of a problem description 2. Extraction of text patterns from a new text and 3. Evaluation of text patterns for novelty and usefulness concerning problem description. Idea Mining Process 4 Fig. 1 shows the processing of the idea mining approach in different steps based on the rationale for the idea mining process (see Sect. 2). Figure 1: Processing of our idea mining approach in different steps: After tokenization and term filtering, text patterns are created and term vectors are built representing these text patterns. Term vectors from the new text are compared to term vectors from the problem description using the Euclidean distance measure. Then, term vectors from the new text are compared to their most similar term vectors from the problem description using the idea mining measure. As a result, we get term vectors from the new text that represent new and useful ideas. With tokenization [3], texts are separated in terms and the term unit is word. The set of different terms in a text is reduced by using stop word filtering methods and stemming [12]. For this, a general list of stop words is used as well as the well-known Porter stemming algorithm [18]. A related problem to the use of stemming is to identify synonyms and homonyms. Synonyms are different words with identical or at least similar meanings. Homonyms are groups of words with the same spelling but with different meanings. With stemming synonyms and homonyms cannot be identified because stemming does not use knowledge of the context of a term. In this idea mining approach, we do not identify synonyms and homonyms. This is because the approach always considers the context of a term by working on text patterns containing several co- occurring terms as described below. Here, we show how to create these text patterns automatically. Around each appearance of each term in the new text, we create a text pattern containing the selected term and all terms, which occur in the left and right context of the selected term. To reduce the number of text patterns, we only create text patterns around non-stop words and around terms that occur both in the new text and in the problem description. One important decision to be taken is to determine the length of a text pattern. Text patterns should not be too small so that they contain all terms representing a new idea. Further text patterns should not be too large so that only terms occur in the text patterns that are related to the new idea. For example if we set the length of the text patterns to l then a text pattern contains the 5 selected term, l terms from its left context and also terms from its right context. The cardinality of the set of stop word filtered and stemmed terms from this pattern is normally smaller than because some terms are stop words, some terms occur twice and some terms have the same stem. l 1l2 +* Nn ∈ In this paper, we do not use a constant length l for all patterns but a variable length of text patterns based on a dynamic adaptation of its context. This is realized by using a term weighting scheme based on the difference between stop words and non-stop words because the importance of a stop word in a text pattern is not as high as the importance of a non-stop word. If an author formulates an idea very briefly by joining catchwords together then he normally does not use many stop words and the text pattern length can be small. If an author formulates an idea in a flowery style that means his writing is not expressed in a clear and simple way then he normally uses more stop words and the text pattern length has to be larger. In the idea mining application the value of text pattern length and the percentage of the importance of stop words and of non-stop words v can be provided by the user. l u To compute the variable length of a text pattern, we firstly define the term weighting scheme. Definition 1. Let (a text) be a list of terms (words) in order of appearance and let be the number of terms in ],..,[ 1 nwwT = iw T and i ]n,..,1[∈ . Let ]~,..,~[ mw1w=Σ be a set of domain specific stop terms [15] and let be the number of terms in Σ . Let the percentage be a term weighting coefficient for stop words. Let the percentage be a term weighting coefficient for non-stop words. Then, we define as term weighting scheme: Nm∈ Nu ∈ N∈v N∈w(f ig ) =)( ig wf ⎪⎩ ⎪ ⎨ ⎧ Σ∉ Σ∈ i i wv wu { }),..,1( ni ∈∀ (1) We give an example for this. The text pattern 'components for frequency conversion of infrared lasers' is built around the word 'conversion'. It contains the word conversion itself, three terms from its left context (components for frequency), and three terms from its right context (of infrared lasers). Here, we use a constant length and a term weighting scheme with 100 %. This means the importance of a stop word is equal to the importance of a non-stop word. The next text pattern is an example for a variable length: 'In a 1st phase, known but so far not available materials and technologies such as layer systems and crystals'. This text pattern is built around the word 'technologies'. Here we use a constant length and a term weighting scheme with 10 % and 100 %. As a result, this text pattern contains six terms from the right context and eleven terms from the left context of the term 'technologies'. In this example, non- stop words are phase, materials, technologies, layer, systems, and crystal. We compute the number of terms from the left and right context as described below: 3l = == βα 3l = =u Nl left ∈ =v Definition 2. Let be a constant length of text patterns. Let be the number of terms from the left context of a text pattern that is built around the term . Let be the number of terms from the right context of a text pattern that is built around the term . Then, we define and as: Nl ∈ N∈ left il iw right il iw i l right i 6 )())(( 1 min njilwfl j k kig j right i =+∨≥= ∑ = + { n,..,1i ∈ }∀ (2) )1())(( 1 min =−∨≥= ∑ = − jilwfl j k kig j left i { n,..,1i ∈ }∀ (3) After computing and , we can build a text pattern around the term from the text . left il right il iT iw ],..,[ 1 nwwT = ],...,,..,[ right i left i liili wwwT +− = (4) For each text pattern from the new text, we create a term vector in vector space model. The size of the vector is defined by the number of different stemmed and stop word filtered terms in the new text. For text pattern encoding, we use binary term vectors that means a vector element is set to one if the corresponding unstemmed term is used in the text pattern and to zero if the term is not. We also build text patterns from the problem description and create term vectors as described above. To identify new and useful ideas, we create a specific idea mining measure. This idea mining measure is described in Sect. 4. By comparing a vector from the new text to one from the problem description, we can compute a result value always between 0 % and 100 % using this measure. The greater the result value the more is the probability that the vector from the new text represents a new and useful idea concerning a vector from the problem description. We use this measure for comparing vectors from the new text to their most similar vectors from the problem description but not to all vectors. This is because result values from comparing a vector to its most similar vectors predominate result values from comparing a vector to its further vectors. For example if a vector from the new text is similar to one from the problem description then the idea is not new to the user regardless whether result values from comparing this vector to further vectors from the problem description are greater than zero. Therefore, we can be sure that a vector represents a new and useful idea only if it gets a great result value from idea mining measure concerning one of its most similar vectors. Further, the computing of the idea mining measure is time consuming. Therefore, it is necessary to limit the number of comparisons with idea mining measure for implementing an idea mining application. We choose a two-step classification way. In the first step, we compare each vector from the new text to all vectors from the problem description by using the well-known Euclidean distance measure. Fortunately, the computing of the Euclidean distance measure is not time consuming so that it is suited for implementing in an idea mining application. In detail, for each vector from the new text, we identify all vectors from the problem description where the Euclidean distance result value is the lowest that means we identify the most similar vectors. In the second step, we compare each vector from the new text to its most similar vectors using the idea mining measure. 7 Each vector from the new text - that is compared to several similar vectors - gets the highest result value from idea mining measure as result value. To identify a new and useful idea we use alpha-cut method. An alpha-cut of the idea mining measure result value is the set of all vectors from the new text such that the appertaining result value is greater than or equal to alpha (α~ ). In the idea mining application, the user can provide the value of . α~ Idea Mining Measure With the idea mining measure, we compare a vector that represents a text pattern from the new text to its most similar vectors from the problem description to identify a new and useful idea inside the text pattern from the new text. In detail, we have to find text pattern from the new text where all terms representing a mean (purpose) and no terms representing a purpose (mean) occur in a text pattern from the problem description. If all terms in the text pattern from the new text are known, which means all terms also occur in a text pattern from the problem description then the idea is not new to the user. Furthermore, the idea is not useful if all terms in the text pattern from the new text are unknown because there is no relation to the problem. It is shown in [22] that to find new and useful ideas the number of known terms (e.g. representing a mean) and the number of unknown terms (e.g. representing an appertaining purpose) shall be well balanced. Definition 3. Let be a the set of stemmed and stop word filtered terms representing a text pattern with number i from the new text. Let be a set of stemmed and stop word filtered terms representing a text pattern with number iα jβ j from the problem description. Let be the set of all stemmed and stop word filtered terms from the new text. Let γ γx = be the cardinality of . Let be a term vector in vector space model concerning . Let be a term vector in vector space model concerning . Let γ { } xi 10ω ,∈ iα { 1, } x j 0ρ ∈ jβ ∑ === x 1k ki,i ωαp be the number of all (known and unknown) terms in text pattern with number . Let i ∑ x 1 ω =k kiji ρβαq •=∩= , kj , be the number of known terms in text pattern with number i concerning a text pattern with number j from the problem description. Then, we define as measure for well-balanced known and unknown term distribution. 1m ⎪ ⎪ ⎩ ⎪⎪ ⎨ ⎧ ⋅ −⋅ = p q p qp m 2 )(2 1 ) 2 ( ) 2 ( p q p q < ≥ (5) The known terms in the text pattern from the new text should occur in the problem description more frequently than other terms. This is because they represent a known mean or a known purpose that is a central part of the problem. In the problem description, terms that represent the 8 problem occur more frequently than other terms. For this, we define these frequent terms by using a percentage z as parameter and we compute as the number of known and frequent terms over the number of all known terms. 2m Definition 4. Let z be a percentage. Let be a set of δ z % most frequently stemmed and stop word filtered terms in the problem description. Let ξ be a term vector in vector space model concerning δ . Let { } x1,0∈ k x 1kji δβαr == = ∑∩∩ kjki ξω •,, ρ• be the number of known terms, which occur frequently in the problem description. We define as measure for frequently occurrence of known terms in the problem description. 2m q r m =2 (6) The unknown terms in the text pattern from the new text represent a new approach (an unknown mean or purpose), which is a central part of the new idea. These terms normally occur more frequently than other terms in the new text because this text deals about the new idea. For this, we also define these frequent terms by using a percentage z as parameter and we compute as the number of unknown and frequent terms over the number of all unknown terms. 3m Definition 5. Let φ be a set of % most frequently stemmed and stop word filtered terms in the new text. Let be a term vector in vector space model concerning . Let z {0 } x1τ ,∈ φ k x 1k kjki x 1k kiji τρωτωφβαs •••= == ∑∑∩∩ ,,, nknown k= be the number of unknown terms, which occur frequently in the new text. We define as measure for frequently occurrence of u terms in the new text. 3m qp s m − =3 (7) There are often characteristic terms (higher, quicker, integrated, minimized etc.) that occur together with new ideas. They point to a changing purpose or a changing mean and can be an indicator for new ideas. Definition 6. Let be a set of these characteristic terms (stemmed and stop word filtered). Let be a term vector in vector space model concerning . Let λ { } x10θ ,∈ λ ∑∩ x 1k kkii θωλαt = •== , be the number of these characteristic terms in text pattern with number i . We define as measure for changing means and purposes. 4m ⎩ ⎨ ⎧ = > = )0(0 )0(1 4 t t m (8) 9 The idea mining measure bases on all four heuristic sub measures. Definition 7. Let and let be weighting factors with ∑ . Let the idea mining measure be the sum of all four sub measures multiplied by weighting factors in case of . { 41h ,..,∈ } 0g h ≥ = = 4 1h h 1g hg qp ≠ ⎩ ⎨ ⎧ = ≠+++ = )(0 )(44332211 qp qpmgmgmgmg m (9) Idea Mining and Comprehensibility Research The aim of idea mining is to find new and useful ideas but also to present these ideas in a comprehensible way to the user. To realize this, we focus on comprehensibility research. Up to the 1960s comprehensibility was a property of the text. It was measured in an objective way by analysing text parameters like word length, sentence length, word-usability, relationship between number of different words and number of words. The well-known approach in this time was the 'Reading Ease'-formula from Flesch [8]. Later research in this field focuses on cognitive effects by doing textual production and reception. The results of this research are presented by two approaches: the 'Hamburger Verständlichkeitsmodell' [14] and the 'Groebener Modell' [10]. Both approaches describe four dimensions of comprehensibility: simplicity, structure-organization, brevity-shortness and interest-liveliness. Figure 2: We present the new text back to the user with text patterns in bold print that represent new and useful ideas. A further approach from cognition research is named text excerption. If a human expert finds new and useful ideas in texts he highlights all corresponding text phrases e.g. with text marking. This behaviour is described by Puppe et al. [19]. In the idea mining application, text excerption is used to present the extracted ideas to the user (Fig. 2 shows an example). For the 'Groebener Modell' marking text pattern is important for structure-organization and this leads directly to comprehensibility. In this point, there are differences between the 'Groebener Modell' and the 'Hamburger Verständlichkeitsmodell' in which structure-organization is not so important for comprehensibility. 10 As a result, the presentation of ideas in the idea mining application based on text excerption. It is comprehensible after the 'Groebener Modell' and it is less comprehensible after the 'Hamburger Verständlichkeitsmodell'. Results and Discussions In a study for the German Ministry of Defence (MoD), we use this approach to identify new technological ideas for the German defence research program. In detail, we have to identify new solution ideas to solve current problems in German defence based research projects. We extract new ideas from 300 descriptions of research projects granted in 2006 by the National Institute of Standards and Technology (NIST) in the United States Small Business Innovation Research (SBIR) Program. We use textual information from current defence based research projects of the German MoD as problem description. As a result, we extract several new ideas that are useful for German defence research planners and that now are used as starting point for collaboration projects or for new defence based research projects. A proper selection of these ideas is a strategic issue and - together with the weapon selection problem [5] - it has significant impacts to the efficiency of future defence systems. The results are published in [6]. Here, we show some successful examples: A modified focal plane array technology is identified that can be used to create a detector for the far ultraviolet spectrum. It leads to an improvement of military reconnaissance. This idea is new because up to now focal plane array technology is only used in the infrared, visual and near ultraviolet area. Further, the approach identifies personnel ultrasonic locating equipment that was originally developed to make orientation possible for fire fighters in dense smoke. It also can be used to improve the location and navigation of soldiers in urban warfare (e.g. in buildings). Additionally, the approach shows that the use of avalanche photodiode (APD) technology can improve the internal gain and the dark current of infrared detectors. This also leads to an improvement of military reconnaissance. This study shows that some of the automatically extracted ideas are useful for technological research planners from the German MoD. Unfortunately, the used problem description (textual information about current defence based research projects) is classified as German restricted (Verschlusssache - Nur für den Dienstgebrauch) that means it is not allowed to distribute it to the scientific community. Therefore, we cannot use the results of this study to evaluate this idea mining approach. However, a separate evaluation (see Sect. 6) is done using (unclassified) patent data that allows re-computing of the evaluation. 11 Evaluation The idea mining measure as central point in the idea mining approach consists of four heuristic sub-measures that are not theoretically founded. Therefore, it is crucial to provide an extensive evaluation to show their success. We compare this approach to a baseline because we are not aware of other approaches for idea mining. As measure for the baseline, we use Jaccard's coefficient [7] as well-known heuristic similarity measure. The idea mining approach is evaluated by using our idea mining application (see Sect. 8). There the web based application and all texts that are used for evaluation are presented. Additionally, we create an alternative idea mining application, based on Jaccard's coefficient instead of the idea mining measure for the sole purpose of comparison to the baseline. For evaluation, we use patent data because in patent descriptions, we normally can find new ideas, which include a considerable part of scientific and technological knowledge [13]. We use the abstract of a patent as new text. A patent often bases on further patents. We aggregate abstracts of theses references as problem description. Then we identify new and useful ideas from this patent concerning its patent references using the idea mining applications. We use abstracts from 40 randomly selected patents and from their references, a general stop word list and Porter stemmer for evaluation. Then we determine the parameters of the idea mining measure ( , , , , , and 1g 2g 3g 4g α~ z ) as well as the parameters for the length of the text patterns ( , , and ). l u v For this, we use further patent data and their references as new text and as problem description. The results are evaluated by a human expert and compared to each single sub measure , , and alone. We find out that using the first sub measure alone is successful. If this sub measure is small then the corresponding text pattern normally does not contain a new and useful idea. If this sub measure is large then the probability that the text pattern contains a new idea is also high. We also find out that using the further sub measures alone is not successful. This means, they are successful only if the result value of the first sub measure is medium to high. Therefore, they only can be used in addition to the first sub measure. 1m 2m 3m 4m The results of the second and third sub measures depend on the parameter . This parameter is used to define frequent terms by building a set of % most frequently stemmed and stop word filtered terms. We heuristically think that this parameter should be between 10 % and 30 % to get good sub measures. This is because if is greater than 30 % then we probably classify several terms, which only occur once as frequent terms. If is smaller than 10 % then we only identify high frequently terms for the set. In this case, the result values of the second and third sub measures are small regardless weather known terms occur frequently in the problem description or unknown terms occur frequently in the new text. Therefore, we determine to the mean value (20 %). Additionally, we see that the second and third sub measure is nearly equally successful and that the fourth sub measure is less successful. Therefore, we heuristically determine the parameters of to 50 %, to 20 %, to 20 % and to 10 %. z z z z z 1g 2g 3g 4g 12 We also have used other values to optimize the combination of these four sub measures. However, we do not find a combination that is generally superior to the selected combination. This is because the success of these value combinations depends on the quality of the user given textual information. Then, we determine the alpha cut value of the idea mining measure . If the percentage α~ is small then we get many result items. This leads to a small precision value because many extracted text patterns do not contain a new and useful idea. If is large then we only get a very small number of results and probably our recall value is small because we do not find most of the new and useful ideas in the new text. A human expert checks the results of several patent descriptions for an optimal value of . He gets the experience that 60 % is a good compromise. Therefore, we set to 60 %. We also determine the alpha cut value of Jaccard's coefficient as measure for the baseline to 20 % by using the same way of evaluation as described above. α~ m α~ α~ α~ After this, we determine the length of the text patterns. The length depends on the parameter l and on , a term weighting scheme that is based on the difference between stop words and non-stop words (see Sect. 3). Text patterns should not be too small so that they contain all terms representing a new and useful idea. Additionally, text patterns should not be too large so that further terms occur in the text patterns that are not related to the new and useful idea. To find out an optimal size of text patterns, we create text patterns from several patent descriptions by using different values for and for the percentages u and . A human expert checks the different length of these text patterns for an optimal size. He gets the best results by setting the value of text pattern length l to 7 terms and the percentage to 50 % and to 100 % . )( ig wf l v u v Then, the approach extracts automatically about 200 new ideas from the 40 randomly selected patents. To cluster these results, means and purposes are assigned to scientific categories in the science citation index and examples are presented below. Several ideas are identified that uses methods from 'Artificial Intelligence' (mean) for applications in 'Health Care Sciences and Services' (purpose). We also identify new ideas using 'Imaging Science and Photographic Technology' (mean) for 'Medical Informatics' purposes. Further ideas use techniques from 'Remote Sensing' (mean) in the field of 'Tropical Medicine' (purpose). Additionally, several ideas use 'Computer Science, Theory and Methods' (mean) for applications in 'Psychiatry' (purpose). Furthermore, methods from 'Artificial Intelligence' (mean) are used for 'Automation and Control Systems' purposes. To evaluate these results, we use precision and recall measures commonly used in information retrieval based on true positives, false positives and false negatives. For this, we have to define the ground truth for our evaluation. Therefore, a human expert also identifies new and useful ideas from these patents manually that means without using our idea mining approach. He uses the idea definition in Sect. 1.2. This means, he checks each text pattern for finding terms representing a known mean (purpose) and terms representing an unknown purpose (mean). These results are the ground truth for the evaluation. For each patent, we compute its precision and recall values by using the idea mining measure and by using the Jaccard's coefficient. Then, we compute the average precision and recall values. As a result, we get a precision value of 40 % and a recall value of 25 % by using the idea mining 13 approach with the idea mining measure. A precision value of 40 % means that if the idea mining approach extracts ten text patterns then four of them represent a new and useful idea. A recall value of 25 % means that if there are four new and useful ideas in the new text then the idea mining approach extracts only one of them. In contrast to this, we get a precision value of 30 % and a recall value of 20 % by using Jaccard's coefficient. This is because in some texts Jaccard's coefficient extracts text patterns from the new text that are similar to text patterns from the problem description. This represents probably a known idea but not a new idea. Beside Jaccard's coefficient, we also test other well-known heuristic measures like overlap-index, cosine-similarity and dice-similarity [7] as baseline. However, we get nearly the same results for the precision (30 %) and for the recall (20 %) value. The Idea Mining Application The idea mining application focus on users without extensive knowledge in the text mining field as well as on text mining experts. We give them the possibility to extract specifically problem solution ideas for their own needs using this idea mining approach. They can access to the web- based application via the internet. It is available under http://www.text-mining.info and it is programmed in perl and ruby. A user has to provide two textual files, a problem description and a new text that probably consists of problem solution ideas. These files can be formatted in various ways e.g. as plain text, html, xml etc. However, scripting code, (html- or xml-) tags, and images are discarded that means the application extracts plain text from the provided files. Then, the user has to select the language of these texts to integrate a general stop word list of this language. The application offers general stop word lists in English, German, Dutch, Spain and French. After determining the parameters of the application the automatically extraction of new and useful ideas from the new text starts as described in the idea mining process (see Sect. 3 and Sect. 4). As a result, new ideas are presented as described in Sect. 5. Conclusions and Future Research This study shows the success of an automatic approach for finding new ideas from textual information. For this, the study transforms creativity approaches from psychology and cognitive science to text mining approaches. One main finding here is to redefine an abstract term (an idea) in a concrete way that it can be used for computing with text mining methods. In detail, it is shown that a technological idea represents a combination of a purpose and a mean and that purposes and means are defined by a combination of terms, which co-occur. 14 Additionally, it is shown that problems and problem solution ideas can be represented as term vectors in vector space model. For this, the study contributes a new (idea mining) measure. This measure identifies new ideas by comparing vectors that represent a problem to vectors that represent a problem solution idea. Last, it is shown that approaches from comprehensibility research can be adopted to this approach to present the new ideas in a comprehensible way to the user. As further main finding, it is demonstrated that this theoretical approach can be realized by a web-based application. The success of the idea mining measure is proved by comparing it to further heuristic measures (overlap-index, cosine-similarity and dice-similarity). Directions for future research are given by the fact that nowadays there is a large amount of textual information available on the internet and this information probably contains many new technological ideas. Enlarging this approach to a web idea mining approach that automatically identifies problem solution ideas from the internet is an interesting topic for further research. Additionally, the parameters of the approach can be optimized and the idea mining measure can probably be enlarged with further aspects to improve its quality that means to get better results for the precision and recall values. A further aspect is to transform this idea mining approach to the colloquial language. For this, it is necessary that the idea definition also contains new product ideas from the consumers. Then, new product ideas can be identified to support marketing activities. Last, the approach can be extended with innovation-related aspects. Then, extracted ideas can be classified as innovative ideas and might be used as starting point for the new product development. Acknowledge We thank Joachim Schulze and Jörg Fenner for constructive technical comments. References [1] Albers, S., & Gassmann, O. (2005). Handbuch Technologie- und Innovationsmanagement: Strategie- Umsetzung- Controlling (p.196). Wiesbaden: Gabler Verlag. [2] Baeza-Yates, R., & Ribeiro-Neto, B. (1999). Modern Information Retrieval. New York: ACM Press. [3] Coussement, K., & Van den Poel, D. (2008). Integrating the voice of customers through call center emails into a decision support system for churn prediction. Information & Management 45, 165. [4] Coussement, K., & Van den Poel, D. (2009). Improving customer attrition prediction by integrating emotions from client/company interaction emails and evaluating multiple classifiers. Expert Systems with Applications 36, 6127-6134. [5] Dagdeviren, M., Yavuz, S., & Kilinc, N. (2009). Weapon selection using the AHP and TOPSIS methods under fuzzy environment. Expert Systems with Applications 36, 8150. [6] Fenner, J., & Thorleuchter, D. (2009). Textmining-Analyse von Forschungsvorhaben des National Institute of Standards and Technology. Euskirchen: Fraunhofer INT Edition. [7] Ferber, R. (2003). Information Retrieval (p. 74-80). Heidelberg: dpunkt.verlag. [8] Flesch, R. (1948). A new readability yardstick. Journal of Applied Psychology 32, 221-233. [9] Gentsch, P., & Hänlein, M. (1999). Text Mining. WISU 12, 1646. [10] Groeben, N. (1982). Leserpsychologie: Textverständnis - Textverständlichkeit. Münster: Aschendorff. 15 16 [11] Hoffmann, L., Kalverkämper, H., & Wiegand, H.E. (1998). Fachsprachen - Languages for Special purposes: Ein internationales Handbuch zur Fachsprachenforschung und Terminologiewissenschaft - an international Handbook of Special-language and Terminology Research (p. 1602). Berlin: Walter de Gruyter. [12] Hotho, A., Nürnberger, A., & Paaß, G. (2005). A Brief Survey of Text Mining. LDV Forum 20(1), 19-26. [13] Li, Y.R., Wang, L.H., & Hong, C.F. (2009). Extracting the significant-rare keywords for patent analysis. Expert Systems with Applications 36, 5200-5204. [14] Langer, I., Schulz v. Thun, F., & Tausch, R. (1974). Verständlichkeit in Schule und Verwaltung. München: Ernst Reinhardt. [15] Lustig, G. (1986). Automatische Indexierung zwischen Forschung und Anwendung (p. 92). Hildesheim: Georg Olms Verlag. [16] Martin-Bautista, M.J., Sanches, D., Serrano, J.M., & Vila M.A. (2004). Text Mining using Fuzzy Association Rules. In V. Loia, M. Nikravesh, & L.A. Zadeh (Eds.), Fuzzy Logic and the Internet (p. 173). Berlin, Springer-Verlag. [17] Osborn, A.-F. (1948). Your Creative Power. New York: C. Scribner's sons. [18] Porter, M.F. (1980). An algorithm for suffix stripping. Program 14(3), 130-137. [19] Puppe, F., Stoyan, H., & Studer, R. (2003). Knowledge Engineering. In G. Görz, C.R. Rollinger, & J. Schneeberger (Eds.), Handbuch der Künstlichen Intelligenz (p. 611). München: Oldenbourg. [20] Ripke, M., & Stöber, G. (1972). Probleme und Methoden der Identifizierung potentieller Objekte der Forschungsförderung. In H. Paschen & H. Krauch (Eds.), Methoden und Probleme der Forschungs- und Entwicklungsplanung (p. 47). München, Oldenbourg. [21] Rohpohl, G. (1996). Das Ende der Natur. In L. Schäfer, & E. Sträker (Eds.), Naturauffassungen in Philosophie, Wissenschaft und Technik (pp. 143-163). Freiburg, München: Alber. [22] Thorleuchter, D. (2008). Finding Technological Ideas and Inventions with Text Mining and Technique Philosophy. In C. Preisach, H. Burkhardt, L. Schmidt-Thieme, & R. Decker (Eds.) Data Analysis, Machine Learning, and Applications (pp. 413-420). Berlin: Springer-Verlag. [23] Wallas, G. (1926). The Art of Thought. New York: Harcourt Brace. FACULTEIT ECONOMIE TWEEKERKENSTRAAT 2 B-9000 GENT WORKING PAPER November 2009 Introduction Overview Idea Definition Rationale behind Idea Mining Idea Mining Process Idea Mining Measure Idea Mining and Comprehensibility Research Results and Discussions Evaluation The Idea Mining Application Conclusions and Future Research 
work_lu4ys3vilvgi3lctc6a3hvkgxy	----	PII: 0004-3702(96)80446-2 Artificial ELSEVIER Artificial Intelligence 82 (1996) 393-394 Intelligence Forthcoming Papers A. Becker and D. Geiger, Optimization of Pearl’s method of conditioning and greedy-like approximation algorithms for the vertex feedback set problem We show how to find a small loop curser in a Bayesian network. Finding such a loop cutset is the first step in the method of conditioning for inference. Our algorithm for finding a loop cutset, called MGA, finds a loop cutset which is guaranteed in the worst case to contain less than twice the number of variables contained in a minimum loop cutset. The algorithm is based on a reduction to the weighted vertex feedback set problem and a new approximation of the latter problem. The complexity of MGA is O(NI + n logn) where m and n are the number of edges and vertices respectively. A greedy algorithm, called GA, for the weighted vertex feedback is also analyzed and bounds on its performance are given. We test MGA on randomly generated graphs and find that the average ratio between tbe number of instances associated with the algorithm’s output and the number of instances associated with an optimum solution is 1.22 for the graphs tested. R Walley, Measures of uncertainty in expert systems This paper compares four measures that have been advocated as models for uncertainty in expert systems. The measures are additive probabilities (used in the Bayesian theory), coherent lower (or upper) previsions, belief functions (used in the Dempster-Shafer theory) and possibility measures (fuzzy logic). Special emphasis is given to the theory of coherent lower previsions, in which upper and lower probabilities, expectations and conditional probabilities are constructed from initial assessments through a technique of natural extension. Mathematically, all the measures can be regarded as types of coherent lower or upper previsions, and this per- spective gives some insight into the properties of belief functions and possibility measures. The measures are evaluated according to six criteria: clarity of interpretation; ability to model partial information and imprecise assessments, especially judgements expressed in natural language; rules for combining and updating uncer- tainty, and their justification; consistency of models and inferences; feasibility of assessment; and feasibility of computations. Each of the four measures seems to be useful in special kinds of problems, but only lower and upper previsions appear to be sufficiently general to model the most common types of uncertainty. Y. Moses and M. Tennenholtz, Off-line reasoning for on-line efficiency: knowledge bases The complexity of reasoning is a fundamental issue in AI. In many cases, the fact that an intelligent system needs to perform reasoning on-line contributes to the difficulty of this reasoning. This paper considers the case in which an intelligent system computes whether a query is entailed by the system’s knowledge base. It investigates how an initial phase of off-line preprocessing and design can improve the on-line complexity Elsevier Science B.V. 
work_luhg4p4curaffnoxsi3hvlg2lm	----	SURE: A method for decision-making under uncertainty This is a repository copy of SURE: A method for decision-making under uncertainty. White Rose Research Online URL for this paper: http://eprints.whiterose.ac.uk/135170/ Version: Accepted Version Article: Hodgett, RE orcid.org/0000-0002-4351-7240 and Siraj, S orcid.org/0000-0002-7962-9930 (2019) SURE: A method for decision-making under uncertainty. Expert Systems with Applications, 115. pp. 684-694. ISSN 0957-4174 https://doi.org/10.1016/j.eswa.2018.08.048 © 2018 Elsevier Ltd. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http://creativecommons.org/licenses/by-nc-nd/4.0/. eprints@whiterose.ac.uk https://eprints.whiterose.ac.uk/ Reuse This article is distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs (CC BY-NC-ND) licence. This licence only allows you to download this work and share it with others as long as you credit the authors, but you can’t change the article in any way or use it commercially. More information and the full terms of the licence here: https://creativecommons.org/licenses/ Takedown If you consider content in White Rose Research Online to be in breach of UK law, please notify us by emailing eprints@whiterose.ac.uk including the URL of the record and the reason for the withdrawal request. mailto:eprints@whiterose.ac.uk https://eprints.whiterose.ac.uk/ For consideration in Expert Systems with Applications. SURE: a method for decision-making under uncertainty Richard Edgar Hodgett Leeds University Business School, The University of Leeds, LS2 9JT, United Kingdom - r.e.hodgett@leeds.ac.uk Sajid Siraj Leeds University Business School, The University of Leeds, LS2 9JT, United Kingdom - s.siraj@leeds.ac.uk Managerial decision-making often involves the consideration of multiple criteria with high levels of uncertainty. Multi-attribute utility theory, a primary method proposed for decision-making under uncertainty, has been repeatedly shown to be difficult to use in practice. This paper presents a novel approach termed Simulated Uncertainty Range Evaluations (SURE) to aid decision makers in the presence of high levels of uncertainty. SURE has evolved from an existing method that has been applied extensively in the pharmaceutical and speciality chemical sectors involving uncertain decisions in whole process design. The new method utilises simulations based upon triangular distributions to create a plot which visualises the preferences and overlapping uncertainties of decision alternatives. It facilitates decision- makers to visualise the not-so-obvious uncertainties of decision alternatives. In a real-world case study for a large pharmaceutical company, SURE was compared to other widely-used methods for decision-making and was the only method that correctly identified the alternative eventually chosen by the company. The case study demonstrates that SURE can perform better than other existing methods for decision-making involving multiple criteria and uncertainty. Key words: Simulated Uncertainty Range Evaluations; MCDM; Uncertainty; Simulations; AHP; ELECTRE III. History: This paper was first submitted on 13th December 2017. Revisions were submitted on 9th May 2018. Revisions were submitted on 2nd August 2018. Paper was accepted on 28th August 2018. mailto:r.e.hodgett@leeds.ac.uk mailto:s.siraj@leeds.ac.uk 2 1. Introduction It is often the case in managerial decision-making that alternatives are assessed in terms of several criteria. These assessments are not so straightforward due to the uncertainty present in real-life situations. Most multi-criteria decision-making (MCDM) methods have been developed or adapted in one way or another to handle uncertainty, often focusing on the uncertainty of the criteria weights. Many of these methods are founded on multi-attribute utility theory (MAUT) (Keeney & Raiffa, 1976) which is primarily designed to handle trade-offs among multiple criteria for a given situation. MAUT is one of the most well-known MCDM methods that was explicitly developed to deal with uncertain information (Belton & Stewart, 2002). It requires the selection of utility functions which represent the risk attitude of the decision-maker for each criterion in a decision problem. It has been extensively discussed in the decision-making literature and is generally valued for its axiomatic foundations. However, MAUT is also known to be difficult to use in practice (Polatidis, et al., 2006; Kumar, et al., 2017) as it specifies uncertain outcomes by means of probability distrubutions which are not typically known (Schaetter, 2016). Excessive time and a high cognitive load is required to derive an accurate representation of an individual╆s utility function ゅLumby ┃ Jones, 2003; Cinelli, et al., 2014). Perhaps as a result, there are few real-world examples of MAUT being used in the literature in comparison to its theoretical development (Durbach & Stewart, 2012b). In this context, Multi-Attribute Range Evaluations (MARE) (Hodgett, et al., 2014) is recomended for handling uncertain decisions. Although MARE was primarily proposed for decision-making in whole process design in the manufacturing industry, the technique is applicable to any decision problem involving multiple criteria and 3 uncertainty. As a result, MARE has been further developed into a number of proprietory software tools as well as open-source libraries like the MCDA package for R (Bigaret, et al., 2017). MARE requires the decision-maker to provide a range in the form of a minimum, most likely and maximum value for each alternative with respect to each criterion. Using a range to capture preferences has become more common in medical applications (Peleg, et al., 2012), survey design (Schwarz, 1999; Bruine de Bruin, et al., 2012) and software development (Wagner, et al., 2017). Peleg et al. (2012) identified that some factors are difficult to be represented by a single value and that ranges can be relatively easy to agree upon by experts. This indicates that asking for ranges is beneficial for both single and group decision-making environments. Therefore it is important to investigate and incorporate the use of ranges in MCDM techniques. In this paper, we propose a new MCDM methodology, termed as Simulated Uncertainty Range Evaluations or SURE, which allows decision-makers to provide their preferences in ranges and the technique utilises triangular distributions to account for uncertain information. SURE offers a more theoretially sound methodology and an improved output for visualising the uncertainty associated with each decision alternative. The value of the proposed method is assessed using a real-life case study from a large pharmaceutical company where it is compared against other widely-used methods for decision-making. In the next section, we give a detailed overview of MARE and the issues associated with it in order to make the case for SURE discussed in the following section. 2. Overview and limitations of Multi-Attribute Range Evaluations MARE was initially proposed as a methodology for handling uncertain decisions in whole process design. Whole process design considers the optimisation of the entire product development process, from raw materials to end product, rather than focusing on each 4 individual unit operation. The complexity involved with the implementation of whole process design requires decision-making, often with limited or uncertain information. A survey sent to management in the speciality chemical and pharmaceutical industries who utilise whole process design identified that the majority (69%) of respondents would spare an hour (or even less) for decision making tasks, and many respondents (89%) prefered to have a decision-making system that guides the user in the right direction quickly, as opposed to those producing exact results but with a very long entry procedure (Hodgett, et al., 2014). These findings meant that a decision-making methodology was needed that could handle uncertain information but could also be used quickly. Most literature that discusses uncertainty in MCDM focuses on probabilities, decision weights, explicit risk measures, fuzzy numbers and scenarios (Durbach & Stewart, 2012). The search for a balance between theoretically rich and complex methods that can handle uncertainty and those that are transparent and easily understood (yet may not conform to the prescriptive principles of rationality) has polarized much of MCDM research. This led to the development of MARE. MARE is based on aggregation-based methods like weighted sums of judgments and preferences, one of the widely-used and simple MCDM approaches. The weighted sum method calculates a score for each alternative Ai by summing the products of each decision variable with its corresponding criterion weight. The decision-maker can provide values (vj) for the importance of each criterion or directly provide criteria weights (wj) that sum to one. If values are provided, the summation ratio normalisation method (sometimes referred to as additive normalisation or distributed normalisation) is typically used to convert the values into weights: 拳珍 噺 塚乳デ 塚乳韮乳転迭 (1) 5 where a decision problem has n fixed criteria, vj denotes the criterion value of the jth criterion and wj denotes the criterion weight of the jth criterion. As the weighted sum method assumes every criterion is maximised, any minimising criterion should have their corresponding decision variables inverted. In the likely event that the decision criteria use different measurement units, the summation ratio normalisation method (Eq. 1) is often utilised before calculating the alternative scores with: 畦沈 噺 デ 拳珍 欠沈珍津珍退怠 for i サ な┸に┸┼┸m. (2) where a decision problem has m fixed alternatives and n fixed criteria, wj denotes the weight of each criterion and aij is the score for the ith alternative with respect to the jth criterion. The primary difference between the weighted sum method and MARE is that the decision-maker can assign up to three scores for each alternative (Aj) in terms of each criterion (Cj). These scores represent the most likely values (aij), the lowest possible values (欠沈珍陳沈津 ) and the highest possible values (欠沈珍陳銚掴 ). If the most likely decision variable is certain, then all the three values converge together and hence no values are required for 欠沈珍陳沈津 or 欠沈珍陳銚掴 as aij is sufficient to represent all three values. In any case, the most likely value should always remain between the lowest and highest values, and the lowest value must remain less than or equal to the highest value, that is, 欠沈珍陳沈津 ズ aij ズ 欠沈珍陳銚掴 . There are two major limitations of the MARE methodology. The first stems from an issue with using three values for each alternative. Using three values means that the summation ratio normalisation method (Eq. 1) which divides by the sum of all the decision variables cannot be applied. This is because there has to be an equal scale length for the minimum, most likely and maximum decision variables. This issue with 6 maintaining an equal scale length also applies to vector normalisation: 欠沈珍茅 噺 銚日乳謬デ 銚鉄日乳尿日転迭 (3) where 欠沈珍茅 is the normalised decision variable for the ith alternative with respect to the jth criterion, 欠沈珍 is the decision variable for the ith alternative with respect to the jth criterion. Consequently, the max scale normalisation procedure that utilises the largest decision variable (欠沈珍陳銚掴 ) for normalisation is proposed: 欠沈珍茅 噺 銚日乳銚乳尿尼猫 (4) where 欠珍陳銚掴 is the largest decision variable with respect to the jth criterion. Max scale normalisation was chosen based on the results of a simulated experiment by Chakraborty & Yeh (2007) who identified that vector normalisation and max scale normalisation are more suitable to be used with the weighted sum method when criteria measurement units are diverse in range and there are a small number of alternatives to be assessed. There are of course other normalisation methods such as the Max-Min method which could also be used: 欠沈珍茅 噺 銚日乳貸 銚乳尿日韮銚乳尿尼猫貸 銚乳尿日韮 (5) where 欠珍陳沈津 is the smallest decision variable with respect to the jth criterion. Once normalisation is performed the decision variables are represented by a value between 0 and 1 and are used to calculate the alternatives scores with respect to minimum, most likely and maximum, with scores closer to 1 being better. The choice of normalisation procedure has been shown to affect the outcome of a MCDM method (Çelen, 2014; Gardziejczyk & Zabicki, 2017) and limiting the number of potential normalisation procedures available is a weakness of the MARE methodology. Especially considering the most widely applied normalisation approach of summation ratio normalisation isn╆t compatible with MARE (Hodgett, et al., 2014). 7 The second major issue with MARE is the level of preference understood between the lowest/highest and most likely values. Figure 1 shows the output of MARE for an equipment selection decision made by Fujifilm Imaging Colorants Ltd (Hodgett, 2016). The circles indicate the most likely values for each alternative and the error bars represent the minimum and maximum outputs for each alternative. The problem with this output is that you cannot interpret the strength of preference between the minimum/maximum and the most likely values. This makes it difficult for selecting an alternative based on the output, particularly if there are many overlapping ranges. Both of the issues with MARE described above are overcome in the SURE methodology which is presented in the next section. Figure 1 Output of MARE for an equipment selection decision (Hodgett, 2016) 8 3. The Simulated Uncertainty Range Evaluations Methodology The SURE methodology is also based on the weighted sum method and requires the same information from the decision-maker as MARE, i.e. criteria weights and the minimum, most likely and maximum values for each alternative in respect to each criterion. However, instead of independently calculating single values for the minimum, most likely and maximum for each alternative, random deviates are generated based upon triangular distributions. The methodology for SURE can be summarised in the following five steps: 1. Set number of decision tables to be simulated (s). 2. Generate s number of simulated decision tables using the minimum, most likely and maximum values as the input parameters to the triangular distributions. 3. Normalise decision tables using summation ratio normalisation. 4. Calculate the results of s decision tables using the weighted sum method. 5. Plot the results using a kernel density plot. It may be possible to implement SURE without using simulations, using convolution- based formulas for mathematically deriving the densities but this would be challenging for a decision with many criteria. Using simulations with SURE offers a quick and simple solution on a modern computer. There are a number of possible distributions that can be used to generate simulated decision tables, however, we suggest the use of triangular distributions. A triangular distribution is a continuous probability distribution that has three parameters, a lower limit, an upper limit and mode. With SURE, 欠珍陳沈津 is used as the lower limit, 欠珍陳銚掴 as the upper limit and 欠沈珍 as the mode. Although other distributions are proposed in theory to account for uncertainty, in practice the triangular distribution is used more frequently as it is easier to turn decision-makers viewpoints into the parameter estimates needed for triangular distributions (Stein & Keblis, 2009). 9 Triangular distributions are often used in areas such as reliability analysis (Ormon, et al., 2002), project scheduling (Vanhoucke, 2016) and corporate finance (Nersesian, 2004) where there is a high amount of uncertainty present. Figure 2 shows various possible triangular distributions with varying levels of skewness. The case on the far left is a left triangular density while the one on the far right is a right triangular density. The case in the centre is a symmetric triangular density as 欠沈珍 = (欠珍陳沈津 + 欠珍陳銚掴 )/2. Figure 2 possible cases of triangular distributions Random deviates are generated based upon triangular distributions which form s simulated decision tables. These decision tables contain various possible scenarios based on the uncertainty in the minimum/maximum ranges given by the decision-maker. The higher the number chosen for s, the more random scenarios generated but also the more computational time required to generate the random numbers. Any minimising criterion should have their corresponding decision variables inverted in all the simulated decision tables. As the simulated decision tables only have one value for each alternative with respect to each criterion, any normalisation method can be used. This overcomes the first major issue with MARE, regarding the limited choice of applicable normalisation methods, as the summation ratio normalisation can be used. The weighted sum method is then used to calculate a score for each alternative for the s simulated decision tables. The output can be shown using a kernel density plot which visualises the overlapping 10 distribution of possible outcomes and the uncertainty present. This overcomes the second major limitation with MARE. The next section presents an illustrated example of the SURE methodology with a real-world case study built in collaboration with a large pharmaceutical company. 4. A case study in the pharmaceutical industry The case study was developed with a process engineering manager at a large pharmaceutical company who had over にど years╆ experience in industrial process engineering, an honours degree in chemical engineering and is a fellow of the Institution of Chemical Engineers. The decision was to select an appropriate degasification technology for a new chemical development process. Details of the product and the process are withheld for reasons of confidentiality. The decision-maker identified five criteria (shown in Table 1) on which to base the decision. The underlying philosophy for the company was to select a technology that was inexpensive, available and straightforward to implement. For all the criteria except Technically Possible (c3), the decision maker chose to use a 0- 100 scale which meant the decision maker provided all inputs using a slider bar where numeric values were not visible but rather textual descriptions were provided from ╉Extremely poor╊ to ╉Excellent╊. For the criterion Technically Possible (c3), only two values (0 or 1) were possible because the technique was either possible at the time or not. The decision-maker selected the term ╉Available Now╊ for cね which can essentially be considered as the level of availability at that time. 11 Table 1 Criteria for the decision problem ID Name Scale Rationale (from the decision-maker) c1 Minimises Hold Up 0-100 ╉Supports the economics of the process and ease of operation┻╊ c2 Simple to Build 0-100 ╉Simplicity in build will speed up development. Must increase robustness of the solution and make the equipment easier to clean. This will contribute to a lower cost┻╊ c3 Technically Possible 0 or 1 ╉The solution has to be capable of removing the gas from the solution to a low enough level┻╊ c4 Available Now 0-100 ╉Need to test and place orders now┸ solutions not off the shelf need to be excluded┻╊ c5 Low Cost 0-100 ╉Lower the cost┸ the better the project payback┻╊ The decision-maker also identified five alternatives; Packed Column (a1), Membrane (a2), Duty Standby CSTR ‒ Vacuum (a3), Duty Standby CSTR with Sparge (a4) and Ultrasonic (a5). From the five alternatives, four were declared technically viable as Ultrasonic (a5) was not capable of removing enough gas from the solution. However, this was included in the analyses as it could become a viable alternative in the future, for example if advances are made in the technology. The least expensive alternatives were Packed Column (a1) and Membrane (a2). However, these options were not readily available to implement quickly within the company. The best options in terms of availability were the two Duty Standby CSTR alternatives (a3 and a4). The data with a 0-100 scale was collected through a user interface with slider bars that had one thumb for certain input (where 欠沈珍陳沈津 = 欠沈珍 = 欠沈珍陳銚掴 ) or three thumbs for uncertain input on a scale of 0-100. As the scores represent the decision-maker╆s preference for 12 each alternative (higher the better), the data collected were all maximising. The 0-1 scale data was collected through a user interface which accepted the input of numerical values. The weights where also collected through slider bars with one thumb. The data collected are shown in Table 2. When using simulations, it is important to consider correlations between the criteria as to not misinform the decision-maker. As the decision-maker mentioned that Simple to Build ゅcにょ ╉will contribute to a lower cost╊ (in Table1) it is possible that this criterion is highly correlated with Low Cost (c5). Therefore, we performed two versions of the analysis. In the first analysis, we calculated the results using independent simulations assuming no correlations between the criteria and in the second analysis we calculated the results assuming that the criteria Simple to Build (c2) and Low Cost (c5) are highly correlated. Table 2 Data collected for the case study Criteria Weights Packed Column (a1) Membrane (a2) Duty Standby CSTR Ȃ Vacuum (a3) Duty Standby CSTR with Sparge (a4) Ultrasonic (a5) Minimises Hold Up (c1) 71 Minimum 49 56 25 6 45 Most Likely 61 88 40 40 50 Maximum 75 97 48 48 60 Simple to Build (c2) 26 Minimum 58 58 29 29 4 Most Likely 62 70 35 36 50 Maximum 66 75 51 52 93 Technically Possible (c3) 96 Most Likely 1 1 1 1 0 Available Now (c4) 61 Minimum 87 25 74 74 0 Most Likely 91 76 85 85 17 Maximum 100 83 97 97 39 Low Cost (c5) 50 Minimum 69 25 34 28 3 Most Likely 80 80 50 50 50 Maximum 91 91 59 59 75 13 There are different simulation methods based on triangular distributions which have been compared on factors such as speed, algorithmic code length, applicability and simplicity (Stein & Keblis, 2009). Josselin & Maux (2017) and Nguyen & McLachlan (2016) both suggest the use of the rtriangle function in the triangle package for R (Carnell, 2017) for the generation of triangular random variates. Initially, this function was used to generate 500,000 decision tables using the data in Table 2. This was then repeated with 1 million decision tables to assess whether there was any significant change in the output. There was no significant change/improvement found as we increased the number of simulations further, so we stopped and calculated the final results using 1 million simulations. As all the data are maximising, there is no need to invert any of the data. Summation ratio normalisation is used on all the simulated decision tables and then the 1million results for each alternative are calculated using the weighted sum method. The results can then be visualised using a kernel density plot. The alternatives can be plotted together as shown in Figure 3 or separately as shown in Figure 4. The vertical lines shown in Figure 4 represent the mean result of each alternative. The performance of alternatives can be judged by the horizontal position of their distribution, with alternatives to the right performing better. The width and height of the distributions illustrate the uncertainty present, with wider and shorter distributions having greater uncertainty. Figures 3 & 4 show the results of independent simulations assuming no correlations between the criteria. However, we also want to evaluate what happens when Simple to Build (c2) and Low Cost (c5) are highly correlated. We implemented this in R using the datasynthR package (Knowles, 2018). Using this package, we can simulate two vectors of 1million uniform random numbers that are highly correlated which can then be used as inputs to the rtriangle function to simulate results (where c2 and c5 are considered highly 14 correlated). The results for this analysis are shown in Figure 5 where the primary difference from the previous analysis (see Figure 3) is that the overlapping distributions for Packed Column (a1) and Membrane (a2) have reduced and the distribution for Ultrasonic (a5) has narrowed. Although the results from the two analyses are similar, we feel it is important to investigate whether correlations between the criteria have any impact on the results as to not misinform the decision-maker. Figure 3 Case study results for SURE on the same plot where simulations are independent The data in Table 2 can also be used with MARE to create the plot shown in Figure 6. Unlike the SURE outputs shown in Figures 3, 4 and 5, in MARE the performance is judged by the vertical position with the dots representing the most likely values and the error bars represent the maximum and minimum values. 15 Figure 4 Case study results for SURE where simulations are independent on separate plots with means shown Figure 5 Case study results for SURE on the same plot where c2 and c5 are highly correlated 16 Figure 6 Case study results for MARE To evaluate SURE against more traditional MCDM approaches, data for this decision were also collected in the form of pairwise comparisons for AHP (Saaty, 1980) and the values for indifference, preference and veto with respect to each criterion was also collected so that ELECTRE III (Roy, 1978; Roy, 1968) can be evaluated. AHP was chosen as it is arguably the most widely used MCDM method while ELECTRE III was included as it has been identified as a superior method for its ability to directly deal with uncertainty (Sayyadi & Makui, 2012; Salminen, et al., 1998). AHP was proposed as a method to solve decision problems using a hierarchical structure of criteria and alternatives. It uses pairwise comparisons as input on the scale of 1-9. 1 infers equal importance, 3 for moderate importance, 5 for strong importance, 7 for very strong importance and 9 for extreme importance. The values of 2, 4, 6 and 8 are compromises between the previous definitions. Pairwise comparisons given by the decision-maker are placed into reciprocal matrices and priorities are identified as the principal eigenvectors of the matrices. These priorities form a decision table where the weighted sum method can then be used to calculate a score for each alternative. ELECTRE III is an outranking approach which uses 17 concordance and discordance indices that are calculated for every possible pair of alternatives. A concordance index expresses how many criteria are in favour of each alternative and a discordance index expresses how many criteria are not in favour of each alternative. Using threshold values provided by the decision-maker, it is possible to determine if each alternative pair is preferred, indifferent or incomparable. ELECTRE III uses pseudo criteria to derive the concordance and discordance indices. Pseudo criteria are a fuzzy representation of each criterion thus the method is capable of dealing with uncertain and limited information. Pseudo criteria are incorporated through the use of indifference, preference and veto thresholds. The indifference threshold is a value below which the decision-maker is indifferent in terms of two alternatives whilst the preference threshold is a value above which the decision-maker prefers one alternative to another. Finally, the veto threshold is the value at which the decision-maker ultimately prefers one alternative over another and wishes to select that alternative with total certainty. ELECTRE III results are calculated through two distillations procedures, one in descending order (finding the best to worst alternatives) and the other in ascending order (finding the worst to best alternatives). The final ranking order is produced by taking an intersection of the descending and ascending orders. The result for AHP is shown in Figure 7 using a bar chart. Alternatives with higher scores are better. The pairwise comparisons given for AHP can be found in the supplementary material along with the R code for creating Figures 3 to 7. All pairwise matrices given were consistent according to the consistency ratio rule propsoed by Saaty (1980) (i.e. CR < 0.1). The most likely values in Table 2 along with the threshold values in Table 3 are used to calculate the results for ELECTRE III shown in Table 4. Unlike the other methods, ELECTRE III results are given in the form of an ordinal ranking. The ascending distillation placed Membrane (a2) and Packed Column (a1) as joint best alternatives while the 18 descending distillation placed Membrane (a2) as the single best alternative. Consequently, the final order classification placed Membrane (a2) as the best alternative in the final rank. The credibility matrix in Table 5 shows that Packed Column (a1) outranked Membrane (a2) by 0.75 while Membrane (a2) outranked Packed Column (a1) by 0.87 resulting in Membrane (a2) achieving a better rank in the descending distillation and subsequently, the final rank. Table 5 also shows that the two Duty Standby CSTR options (a3 and a4) are not comparable as their outranking relationships are both 1 which is why they have been ranked together in the descending, ascending and final rankings in Table 4. Figure 7 Case study results for AHP Table 3 Thresholds for ELECTRE III Indifference Threshold Preference Threshold Veto Threshold c1 5 20 80 c2 5 20 80 c3 0.1 0.9 1 c4 5 20 80 c5 5 20 80 19 Table 4 Case study results for ELECTRE III Rank Descending Order Ascending Order Final Order 1  Membrane  Membrane  Packed Column  Membrane 2  Packed Column  Duty Standby CSTR with Sparge  Duty Standby CSTR - Vacuum  Packed Column 3  Duty Standby CSTR ‒ Vacuum  Duty Standby CSTR with Sparge  Ultrasonic  Duty Standby CSTR ‒ Vacuum  Duty Standby CSTR with Sparge 4  Ultrasonic  Ultrasonic Table 5 Case study credibility matrix for ELECTRE III Packed Column Membrane Duty Standby CSTR Ȃ Vacuum Duty Standby CSTR with Sparge Ultrasonic Packed Column 0.75 1 1 1 Membrane 0.87 0.95 0.95 1 Duty Standby CSTR Ȃ Vacuum 0.5 0.52 1 0.87 Duty Standby CSTR with Sparge 0.5 0.52 1 0.87 Ultrasonic 0 0 0 0 After conducting the analyses, the decision-maker made time to review his experiences and to discuss the results. The decision-maker preferred the methods that allowed the user to ╉spread their answers╊ as ╉it was much more useful in terms of seeing the uncertainty behind the membrane option╊┻ (e explained that Membrane ゅaにょ would have been the favoured alternative internally within the company if it had been possible to 20 reduce the uncertainty associated with it. However, post analysis, he favoured Packed Column (a1), as that alternative was more certain to perform well. In terms of data entry, the decision-maker found the AHP consistency check ╉somewhat disconcerting╊ and he stated that straight data entry was faster in contrast to pairwise comparisons. Nevertheless, when asked about the differences in the criteria weights, the decision-maker said the weights produced by AHP were more representative. His reasoning was that cぬ ゅ╅technically possible╆ょ was a ╉veto type attribute╊ and A(P weighted this criterion much higher. Considering the analysis output, the decision-maker disliked ELECTRE III. He explained that the credibility index (Table 5ょ was ╉confusing╊ and that he disliked output in the form of an ordinal ranking as the differences between the alternatives were not clear. The next section provides a detailed discussion of the case study results which is followed by a list of conclusions, further work and limitations. 5. Discussion and Conclusions This paper presented a new multi-criteria decision-making method termed Simulated Uncertainty Range Evaluations (SURE) which improves upon multi-attribute range evaluations (MARE), a method that has been found to be very effective in supporting uncertain decisions in whole process design. We have shown how SURE is superior in allowing for any form of normalisation procedure and in its ability to visualise the strength of preference and uncertainty associated with each alternative. The practicality of the new approach was illustrated in a real-world case study for a large pharmaceutical company. SURE was compared against MARE, AHP and ELECTRE III and was the only method to identify the alternative chosen by the company as the best alternative. As ELECTRE III provides results in the form of an ordinal ranking, the outputs of the four 21 analyses were not comparable on a numerical scale. The company selected Packed Column (a1) as the best alternative due to the large uncertainty associated with the Membrane (a2) option. AHP and ELECTRE III failed to identify Packed Column (a1) as the best alternative. The uncertainty associated with the Membrane (a2) option was only identified with the MARE and SURE analyses. One could argue however that MARE didn╆t explicitly identify Membrane (a2) as the best option. The output provided by MARE requires the decision-maker to make a choice in terms of which alternative to select given the overlapping uncertainty of the alternatives. In fact, MARE identified Membrane (a2) as the best option in terms of the most likely value but you can also see (in Figure 6) the large uncertainty associated with Membrane (a2) and the smaller uncertainty associated with Packed Column (a1) which informed the decision-maker to select Packed Column (a1). This uncertainty is also clearly visible in the output for SURE through the overlapping distributions (Figures 3-5). The output clearly shows Packed Column (a1) as the furthest distribution to the right signifying it as the best alternative. Furthermore, in the separated plots output for SURE (Figure 4) you can see that the mean result for Packed Column (a1) is higher than all other alternatives. Therefore, SURE has outperformed the other methods with regards to identifying the correct result. The dissimilar results for MARE and SURE that are both based upon the weighted sum method will be a consequence of the different normalisation procedures used. MARE is unable to utilise summation ratio normalisation, the most common approach to normalisation which is utilised in SURE. With respect to criteria weights, the decision-maker favoured the weights calculated by AHP as it gave a higher weighting to cぬ ゅ╅technically possible╆ょ. AHP has been known to exaggerate weights in comparison to direct weighting methods (Hodgett, 2016). This can 22 be seen in Figure 8 which shows the normalised weights calculated by AHP beside the weights used by the other methods in the case study. Figure 8 Comparison of the weights for AHP and the other methods in the case study Clearly┸ cぬ ゅ╅technically possible╆ょ has been exaggerated which in this case the decision- maker wanted. Therefore, for certain applications it might be worth using AHP to calculate the criteria weights for SURE rather than normalising direct weightings. Figure 9 shows the output for SURE if the normalised criteria weights for AHP are used rather than the weights given in Table 2. As expected, there is a much greater divide between Ultrasonic (a5) and the other alternatives as the weighting has increased for c3 ゅ╅technically possible╆ょ where Ultrasonic (a5) performs much worse than the other alternatives. This has also reduced the difference between the other alternatives as with Ultrasonic (a5) getting a lower overall score the other alternatives will achieve a higher overall score. Nevertheless, Packed Column (a1) remains as the highest performing alternative which confirms the result that SURE is the best method to use for this particular case study. Another way to assess the weight given to cぬ ゅ╅technically possible╆ょ is to conduct a sensitivity analysis to investigate how sensitive the results are with respect to a change in an input parameter. 0 0.2 0.4 0.6 0.8 Minimises Hold Up Simple to BuildTechnically PossibleAvailable Now Low Cost AHP Other Methods 23 Figure 9 Case study results for SURE on separate plots using AHP weights SURE handles the uncertainty in evaluations using triangular distributions and therefore minor changes of evaluations scores should not significantly alter the results as random deviates will be between the minimum and maximum values. However, we propose the use of a one-at-a-time (OAT) sensitivity analysis where one evaluation can be modified at a time to see its impact on the results. The uncertainty in criteria weights can also be analysed using the same approach. To illustrate this. we used OAT changing the weight for cぬ ゅ╅technically possible╆ょ from 96 (see Figure 10a) to 1 (see Figure 10b) to see how this impacted the distributions. We consider the evaluation of other possible ways to perform sensitivity analysis for SURE to be an important area of future work. 24 (a) cぬ ゅ╅technically possible╆ょ weight = 96 (b) cぬ ゅ╅technically possible╆ょ weight = 1 Figure 10 illustrative example of OAT sensitivity analysis 25 Further work is also needed to evaluate SURE with other decisions, particularly decisions faced in other sectors outside of the pharmaceutical industry. There are also other well- known MCDM methods such as MAUT, TOPSIS (Hwang & Yoon, 1981) and PROMETHEE (Brans, 1982) which can be compared and assessed against SURE. The procedure for assigning criteria weights also needs further investigation. It is unclear from the case study whether AHP or a direct weighting approach is best for assigning the criteria weights for SURE. There are, therefore, several opportunities for further research and practice. To assist with further work in this area and to make this work as open, transparent and replicable as possible, the R code used to create Figures 3-7 in this paper is included as supplementary material. Acknowledgments The authors would like to thank the large (anonymous) Pharmaceutical company for providing the case study and Britest Limited (http://www.britest.co.uk) for their support. We would also like to thank Prof. Alan Pearman and Prof. Wändi Bruine de Bruin in the Centre for Decision Research for their helpful comments and suggestions. References Belton, V. & Stewart, T. J., 2002. Multiple Criteria Decision Analysis: an Integrated Approach. Kluwer Academic Publisher. Bigaret, S., Hodgett, R.E., Meyer, P. et al. , 2017. Supporting the Multi-Criteria Decision Aiding process: R and the MCDA package. EURO Journal on Decision Processes, 5 (1-4) p. 169-194 Brans┸ J┻ P┻┸ なひぱに┻ L╆ingénierie de la décision┺ élaboration d╆instruments d╆aide à la décision┻ La méthode PROMET(EE┻ Presses de l╆Université Laval┻ http://www.britest.co.uk/ 26 Bruine de Bruin┸ W┻ et al┻┸ にどなに┻ The effect of question wording on consumers╆ reported inflation expectations. Journal of Economic Psychology, 33 p. 749‒757. Carnell, R., 2017. The triangle Package. [Online] Available at: https://cran.r- project.org/web/packages/triangle/triangle.pdf Çelen, A., 2014. Comparative Analysis of Normalization Procedures in TOPSIS Method: With an Application to Turkish Deposit Banking Market. Informatica, 24(2) p.185-208. Chakraborty, S. & Yeh, C.-H., 2007. A Simulation Based Comparative Study of Normalization Procedures in Multiattribute Decision Making. Corfu Island, Greece, Proceedings of the 6th WSEAS Int. Conf. on Artificial Intelligence, Knowledge Engineering and Data Bases. Cinelli, M., Coles, S. R. & Kirwan, K., 2014. Analysis of the potentials of multi criteria decision analysis methods to conduct sustainability assessment. Ecological Indicators, 46 p. 138-148. Durbach, I. N. & Stewart, T. J., 2012b. A comparison of simplified value function approaches for treating uncertainty in multi-criteria decision analysis. Omega,40 p. 456‒ 464. Durbach, I. N. & Stewart, T. J., 2012. Modelling uncertainty in multi-criteria decision analysis. European Journal of Operational Research, 223 p. 1‒14. Gardziejczyk, W. & Zabicki, P., 2017. Normalization and variant assessment methods in selection of road alignment variants ‒ case study. Journal of Civil Engineering and Management, 23(4), p. 510-523. Gass, S. I., 2005. Model World: The Great Debate: MAUT versus AHP. Interfaces, 35 (4) p. 308-312. 27 Hodgett, R. E., 2016. Comparison of multi-criteria decision-making methods for equipment selection. The International Journal of Advanced Manufacturing Technology, 85 (5-8) p. 1145‒1157. Hodgett, R. E., Martin, E. B., Montague, G. & Talford, M., 2014. Handling uncertain decisions in whole process design. Production Planning & Control, 25(12) p.1028-1038. Hwang, C. L. & Yoon, K., 1981. Multiple Attribute Decision Making Methods and. In: a state of the art survey. Springer-Verlag. Josselin, J.-M. & Maux, B. L., 2017. Statistical Tools for Program Evaluation: Methods and Applications to Economic Policy, Public Health, and Education. Springer. Keeney, R. L. & Raiffa, H., 1976. Decisions with Multiple Objectives: Preferences and Value Trade-offs. New York: Wiley. Knowles, J., 2018. The datasynthR Package. [Online] Available at: https://github.com/jknowles/datasynthR Kumar, A. et al., 2017. A review of multi criteria decision making (MCDM) towards sustainable renewable energy development. Renewable and Sustainable Energy Reviews,69 p. 596‒609. Lumby, S. & Jones, C., 2003. Corporate Finance: Theory & Practice. Cengage Learning EMEA. Nersesian, R. L., 2004. Corporate Financial Risk Management: A Computer-based Guide for Nonspecialists. Greenwood Publishing Group. Nguyen, H. D. & McLachlan, G. J., 2016. Linear mixed models with marginally symmetric nonparametric random effects. Computational Statistics and Data Analysis, 103 p. 151- 169. Ormon, S. W., Cassady, C. R. & Greenwood, A. G., 2002. Reliability prediction models to support conceptual design. IEEE Transactions on Reliability, 51(2) p. 151 - 157. 28 Peleg, M., Normand, M. D. & Corradini, M. G., 2012. A method to estimate a person or group health risks and benefits from additive and multiplicative factors. Trends in Food Science & Technology, 28 p. 44-51. Polatidis, H., A. Haralambopoulos, D. A., Munda, G. & Vreeker, R., 2006. Selecting an Appropriate Multi-Criteria Decision Analysis Technique for Renewable Energy Planning. Energy Sources, Part B: Economics, Planning, and Policy, 1:2p. 181-193. Roy, B., 1968. Classement et choix en présence de points de vue multiples (la méthode ELECTRE). La Revue d'Informatique et de Recherche Opérationelle, 8 p. 57‒75. Roy, B., 1978. ELECTRE III: Un algorithme de classements fondé sur une représentation floue des préférences en présence de critéres multiples┻ Cahiers du Centre d╆Etudes de Recherche Opérationnelle, 20, p. 3-24. Saaty, T. L., 1980. The Analytic Hierarchy Process. New York: McGraw-Hill. Salminen, P., Hokkanen, J. & Lahdelma, R., 1998. Comparing multicriteria methods in the context of environmental problems. European Journal of Operational Research, 104(3), p. 485-496. Sayyadi, M. K. & Makui, A., 2012. A new view to uncertainty in Electre III method by introducing interval numbers. Decision Science Letters, 1 p. 33-38. Schaetter, F., 2016. Decision support system for a reactive management of disaster- caused supply chain disturbances. KIT Scientific Publishing. Schwarz, N., 1999. Self-reports: how the questions shape the answers. American psychologist, 54(2). Stein, W. E. & Keblis, M. F., 2009. A new method to simulate the triangular distribution. Mathematical and Computer Modelling, 49 p. 1143‒1147. Stein, W. E. & Keblis, M. F., 2009. A new method to simulate the triangular distribution. Mathematical and Computer Modelling, 49(5-6), p. 1143-1147. 29 Stewart, T. J., 2005. Dealing with Uncertainties in MCDA. In: Multiple Criteria Decision Analysis: State of the Art Surveys. New York: Springer, p. 445-466. Vanhoucke, M., 2016. Integrated Project Management Sourcebook - A Technical Guide to Project Scheduling, Risk and Control. Springer International Publishing. Wagner, M., Rind, A., Thür, N. & Aigner, W., 2017. A knowledge-assisted visual malware analysis system: Design, validation, and reflection of KAMAS. Computers & Security, 67 p. 1-15. 30 Supplementary Material # Install and add all packages that are needed install.packages("triangle") library("triangle") install.packages("plyr") library(plyr) install.packages("ggplot2", dependencies = T) library("ggplot2") install.packages("MCDA") library(MCDA) install.packages("devtools") library(devtools) install_github("jknowles/datasynthR") library("datasynthR") # Enter MARE / SURE data weights <- c(71, 26, 96, 61, 50) c1min <- c(49, 56, 25, 6, 45) c1 <- c(61, 88, 40, 40, 50) c1max <- c(75, 97, 48, 48, 60) c2min <- c(58, 58, 29, 29, 4) c2 <- c(62, 70, 35, 36, 50) c2max <- c(66, 75, 51, 52, 93) c3 <- c(1, 1, 1, 1, 0) c4min <- c(87, 25, 74, 74, 0) c4 <- c(91, 76, 85, 85, 17) c4max <- c(100, 83, 97, 97, 39) c5min <- c(69, 25, 34, 28, 3) c5 <- c(80, 80, 50, 50, 50) c5max <- c(91, 91, 59, 59, 75) ### Use SURE ŷ assuming independence # Set number of simulations s <- 1000000 ### Normalise Weights using Linear Scale Sum Normalisation total <- sum(weights) for (i in 1:length(weights)) { weights[i] <- weights[i] / total } names(weights) <- c("Minimises Hold Up","Simple to Build","Technically Possible","Available Now","Low Cost") # Create performance table to save simulations performanceTable <- array(NA, c(5,5,s)) 31 # Simulations based on triangular distributions performanceTable[1,1,] <- rtriangle(s, a=c1min[1], b=c1max[1], c=c1[1]) performanceTable[1,2,] <- rtriangle(s, a=c1min[2], b=c1max[2], c=c1[2]) performanceTable[1,3,] <- rtriangle(s, a=c1min[3], b=c1max[3], c=c1[3]) performanceTable[1,4,] <- rtriangle(s, a=c1min[4], b=c1max[4], c=c1[4]) performanceTable[1,5,] <- rtriangle(s, a=c1min[5], b=c1max[5], c=c1[5]) performanceTable[2,1,] <- rtriangle(s, a=c2min[1], b=c2max[1], c=c2[1]) performanceTable[2,2,] <- rtriangle(s, a=c2min[2], b=c2max[2], c=c2[2]) performanceTable[2,3,] <- rtriangle(s, a=c2min[3], b=c2max[3], c=c2[3]) performanceTable[2,4,] <- rtriangle(s, a=c2min[4], b=c2max[4], c=c2[4]) performanceTable[2,5,] <- rtriangle(s, a=c2min[5], b=c2max[5], c=c2[5]) performanceTable[3,1,] <- c3[1] performanceTable[3,2,] <- c3[2] performanceTable[3,3,] <- c3[3] performanceTable[3,4,] <- c3[4] performanceTable[3,5,] <- c3[5] performanceTable[4,1,] <- rtriangle(s, a=c4min[1], b=c4max[1], c=c4[1]) performanceTable[4,2,] <- rtriangle(s, a=c4min[2], b=c4max[2], c=c4[2]) performanceTable[4,3,] <- rtriangle(s, a=c4min[3], b=c4max[3], c=c4[3]) performanceTable[4,4,] <- rtriangle(s, a=c4min[4], b=c4max[4], c=c4[4]) performanceTable[4,5,] <- rtriangle(s, a=c4min[5], b=c4max[5], c=c4[5]) performanceTable[5,1,] <- rtriangle(s, a=c5min[1], b=c5max[1], c=c5[1]) performanceTable[5,2,] <- rtriangle(s, a=c5min[2], b=c5max[2], c=c5[2]) performanceTable[5,3,] <- rtriangle(s, a=c5min[3], b=c5max[3], c=c5[3]) performanceTable[5,4,] <- rtriangle(s, a=c5min[4], b=c5max[4], c=c5[4]) performanceTable[5,5,] <- rtriangle(s, a=c5min[5], b=c5max[5], c=c5[5]) ### Normalise decision tables using summation ratio normalisation for (i in 1:s){ sumj <- c(1:5) for (j in 1:5) { sumj[j] <- sum(performanceTable[j,,i]) } for (j in 1:5) { performanceTable[j,,i] <- performanceTable[j,,i] / sumj[j] } } ### Calculate the results 32 results <- array(NA, c(1,5,s)) dimnames(results)[[2]] <- c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic") for (k in 1:s){ for (i in 1:5) { result <- 0 for (j in 1:5) { result <- result + (performanceTable[j,i,k] * weights[j]) } results[1,i,k] <- result } } # Figure 3 - Greyscale plot <- data.frame(data = c(results[1,1,], results[1,2,], results[1,3,], results[1,4,], results[1,5,]), lines = rep(c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic"), each = s)) ggplot(plot, aes(x = data, fill = lines)) + geom_density(alpha = 0.5) + scale_fill_grey() # Figure 3 - Colour plot <- data.frame(data = c(results[1,1,], results[1,2,], results[1,3,], results[1,4,], results[1,5,]), lines = rep(c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic"), each = s)) ggplot(plot, aes(x = data, fill = lines)) + geom_density(alpha = 0.5) # Figure 4 - Greyscale plot <- data.frame(data = c(results[1,1,], results[1,2,], results[1,3,], results[1,4,], results[1,5,]), lines = rep(c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic"), each = s)) mu <- ddply(plot, "lines", summarise, grp.mean=mean(data)) ggplot(plot, aes(x = data, fill = lines)) + geom_density(alpha = 0.5, fill="grey") + facet_grid(lines ~ .) + theme(legend.position="none") + geom_vline(data=mu, aes(xintercept=grp.mean),color="black", linetype="solid", size=1) # Figure 4 - Colour plot <- data.frame(data = c(results[1,1,], results[1,2,], results[1,3,], results[1,4,], results[1,5,]), lines = rep(c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic"), each = s)) mu <- ddply(plot, "lines", summarise, grp.mean=mean(data)) ggplot(plot, aes(x = data, fill = lines)) + geom_density(alpha = 0.5) + facet_grid(lines ~ ., labeller = labeller(lines = label_wrap_gen(10))) + theme(legend.position="none") + geom_vline(data=mu, aes(xintercept=grp.mean),color="black", linetype="solid", size=1) ### Use SURE ŷ where c2 and c5 are correlated 33 # generate s uniform random numbers for c2 and c5 that are correlated cormatrix <- cor(data.frame(c2,c5), method="pearson") rc2 <- runif(s) rc5 <- runifcor.cor(rc2, cormatrix[2]) cor(rc2,rc5) # a modified rtriangle function that accepts custom random numbers in crunif mrtriangle <- function (n = 1, a = 1, b = 100, c = 10^((log10(a) + log10(b))/2), logbase = 10, crunif = NULL) { stopifnot(length(n) == 1) if (n < 1 | is.na(n)) stop(paste("invalid argument: n =", n)) n <- floor(n) if (any(is.na(c(a, b, c)))) return(rep(NaN, times = n)) if (any(a > c | b < c)) return(rep(NaN, times = n)) if (any(is.infinite(c(a, b, c)))) return(rep(NaN, times = n)) if (any(c(a, b, c) == 0)) return(rep(-Inf, times = n)) if (any(c(a, b, c) < 0)) return(rep(NaN, times = n)) if (is.null(crunif)) { lp <- runif(n) } else { stopifnot(nrow(crunif) >= length(n)) lp <- crunif[n] } stopifnot(length(logbase) == 1) if (logbase == 10) { la <- log10(a) lb <- log10(b) lc <- log10(c) } else { la <- log(a)/log(logbase) lb <- log(b)/log(logbase) lc <- log(c)/log(logbase) } if (a != c) { i <- which((la + sqrt(lp * (lb - la) * (lc - la))) <= lc) j <- which((lb - sqrt((1 - lp) * (lb - la) * (lb - lc))) > lc) 34 } else { i <- which((la + sqrt(lp * (lb - la) * (lc - la))) < lc) j <- which((lb - sqrt((1 - lp) * (lb - la) * (lb - lc))) >= lc) } if (length(i) != 0) lp[i] <- la + sqrt(lp[i] * (lb - la) * (lc - la)) if (length(j) != 0) lp[j] <- lb - sqrt((1 - lp[j]) * (lb - la) * (lb - lc)) p <- logbase^lp return(p) } performanceTable <- array(NA, c(5,5,s)) # Simulations based on triangular distributions - c2 and c5 correlated performanceTable[1,1,] <- rtriangle(s, a=c1min[1], b=c1max[1], c=c1[1]) performanceTable[1,2,] <- rtriangle(s, a=c1min[2], b=c1max[2], c=c1[2]) performanceTable[1,3,] <- rtriangle(s, a=c1min[3], b=c1max[3], c=c1[3]) performanceTable[1,4,] <- rtriangle(s, a=c1min[4], b=c1max[4], c=c1[4]) performanceTable[1,5,] <- rtriangle(s, a=c1min[5], b=c1max[5], c=c1[5]) performanceTable[2,1,] <- mrtriangle(s, a=c2min[1], b=c2max[1], c=c2[1], crunif=rc2) performanceTable[2,2,] <- mrtriangle(s, a=c2min[2], b=c2max[2], c=c2[2], crunif=rc2) performanceTable[2,3,] <- mrtriangle(s, a=c2min[3], b=c2max[3], c=c2[3], crunif=rc2) performanceTable[2,4,] <- mrtriangle(s, a=c2min[4], b=c2max[4], c=c2[4], crunif=rc2) performanceTable[2,5,] <- mrtriangle(s, a=c2min[5], b=c2max[5], c=c2[5], crunif=rc2) performanceTable[3,1,] <- c3[1] performanceTable[3,2,] <- c3[2] performanceTable[3,3,] <- c3[3] performanceTable[3,4,] <- c3[4] performanceTable[3,5,] <- c3[5] performanceTable[4,1,] <- rtriangle(s, a=c4min[1], b=c4max[1], c=c4[1]) performanceTable[4,2,] <- rtriangle(s, a=c4min[2], b=c4max[2], c=c4[2]) performanceTable[4,3,] <- rtriangle(s, a=c4min[3], b=c4max[3], c=c4[3]) performanceTable[4,4,] <- rtriangle(s, a=c4min[4], b=c4max[4], c=c4[4]) performanceTable[4,5,] <- rtriangle(s, a=c4min[5], b=c4max[5], c=c4[5]) performanceTable[5,1,] <- mrtriangle(s, a=c5min[1], b=c5max[1], c=c5[1], crunif=rc5) performanceTable[5,2,] <- mrtriangle(s, a=c5min[2], b=c5max[2], c=c5[2], crunif=rc5) performanceTable[5,3,] <- mrtriangle(s, a=c5min[3], b=c5max[3], c=c5[3], crunif=rc5) performanceTable[5,4,] <- mrtriangle(s, a=c5min[4], b=c5max[4], c=c5[4], crunif=rc5) 35 performanceTable[5,5,] <- mrtriangle(s, a=c5min[5], b=c5max[5], c=c5[5], crunif=rc5) for (i in 1:s){ sumj <- c(1:5) for (j in 1:5) { sumj[j] <- sum(performanceTable[j,,i]) } for (j in 1:5) { performanceTable[j,,i] <- performanceTable[j,,i] / sumj[j] } } results <- array(NA, c(1,5,s)) dimnames(results)[[2]] <- c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic") for (k in 1:s){ for (i in 1:5) { result <- 0 for (j in 1:5) { result <- result + (performanceTable[j,i,k] * weights[j]) } results[1,i,k] <- result } } plot <- data.frame(data = c(results[1,1,], results[1,2,], results[1,3,], results[1,4,], results[1,5,]), lines = rep(c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic"), each = s)) ggplot(plot, aes(x = data, fill = lines)) + geom_density(alpha = 0.5) # Figure 5 - Greyscale plot <- data.frame(data = c(results[1,1,], results[1,2,], results[1,3,], results[1,4,], results[1,5,]), lines = rep(c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic"), each = s)) ggplot(plot, aes(x = data, fill = lines)) + geom_density(alpha = 0.5) + scale_fill_grey() # Figure 5 - Colour plot <- data.frame(data = c(results[1,1,], results[1,2,], results[1,3,], results[1,4,], results[1,5,]), lines = 36 rep(c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic"), each = s)) ggplot(plot, aes(x = data, fill = lines)) + geom_density(alpha = 0.5) ### Use MARE performanceTableMin <- matrix(c(c1min,c2min,c3,c4min,c5min),nrow=5,ncol=5, byrow=TRUE) performanceTable <- matrix(c(c1,c2,c3,c4,c5),nrow=5,ncol=5, byrow=TRUE) performanceTableMax <- matrix(c(c1max,c2max,c3,c4max,c5max),nrow=5,ncol=5, byrow=TRUE) row.names(performanceTable) <- names(weights) colnames(performanceTable) <- c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic") row.names(performanceTableMin) <- names(weights) colnames(performanceTableMin) <- colnames(performanceTable) row.names(performanceTableMax) <- names(weights) colnames(performanceTableMax) <- colnames(performanceTable) criteriaMinMax <- c("max", "max", "max", "max", "max") MAREResults <- MARE(performanceTableMin, performanceTable, performanceTableMax, weights, criteriaMinMax) # Figure 6 plotMARE(MAREResults) ### Use AHP criteriaWeightsPairwiseComparisons <- t(matrix(c(1,4,1/9,1,3,1/4,1,1/9,1/3,1,9,9,1,9,9,1,3,1/9,1,2,1/3,1,1/9,1/2,1),nrow=5,ncol=5)) colnames(criteriaWeightsPairwiseComparisons) = names(weights) rownames(criteriaWeightsPairwiseComparisons) = names(weights) ac1 <- t(matrix(c(1,1/4,3,4,1/2,4,1,6,4,2,1/3,1/6,1,1,1/2,1/4,1/4,1,1,1/2,2,1/2,2,2,1),nrow=5,ncol=5)) colnames(ac1) <- c("Packed Column", "Membrane", "Duty Standby CSTR - Vacuum", "Duty Standby CSTR with Sparge", "Ultrasonic") rownames(ac1) <- colnames(ac1) ac2 <- t(matrix(c(1,1/3,3,3,1/3,3,1,6,6,2,1/3,1/6,1,1,1/3,1/3,1/6,1,1,1/3,3,1/2,3,3,1),nrow=5,ncol=5)) colnames(ac2) <- colnames(ac1) rownames(ac2) <- colnames(ac1) ac3 <- t(matrix(c(1,1,1,1,9,1,1,1,1,9,1,1,1,1,9,1,1,1,1,9,1/9,1/9,1/9,1/9,1),nrow=5,ncol=5)) colnames(ac3) <- colnames(ac1) rownames(ac3) <- colnames(ac1) ac4 <- t(matrix(c(1,5,2,2,5,1/5,1,1/4,1/4,3,1/2,4,1,1/2,4,1/2,4,2,1,4,1/5,1/3,1/4,1/4,1),nrow=5,ncol=5)) colnames(ac4) <- colnames(ac1) rownames(ac4) <- colnames(ac1) ac5 <- t(matrix(c(1,1,3,3,5,1,1,1,1,1,1/3,1,1,1,1,1/3,1,1,1,1,1/5,1,1,1,1),nrow=5,ncol=5)) 37 colnames(ac5) <- colnames(ac1) rownames(ac5) <- colnames(ac1) alternativesPairwiseComparisonsList <- list(c1=ac1, c2=ac2, c3=ac3, c4=ac4, c5=ac5) # Check consistency measures pairwiseConsistencyMeasures(criteriaWeightsPairwiseComparisons) pairwiseConsistencyMeasures(ac1) pairwiseConsistencyMeasures(ac2) pairwiseConsistencyMeasures(ac3) pairwiseConsistencyMeasures(ac4) pairwiseConsistencyMeasures(ac5) AHPResults <- AHP(criteriaWeightsPairwiseComparisons, alternativesPairwiseComparisonsList) # Figure 7 barplot(AHPResults) 
work_lv4egn54gbaijckh5yvmruh64y	----	Expert Systems in Service Operations By: Luvai F. Motiwalla and Vidyaranya B. Gargeya. Motiwalla, L. F. and Gargeya, V. B. (1992). Expert systems in service operations. Industrial Management & Data Systems, 92(8), 14-19. Made available courtesy of Emerald Group Publishing: http://dx.doi.org/10.1108/02635579210019820 ***© Emerald Group Publishing. Reprinted with permission. No further reproduction is authorized without written permission from Emerald Group Publishing. This version of the document is not the version of record. Figures and/or pictures may be missing from this format of the document. *** This article is (c) Emerald Group Publishing and permission has been granted for this version to appear here (https://libres.uncg.edu/ir/uncg/). Emerald does not grant permission for this article to be further copied/distributed or hosted elsewhere without the express permission from Emerald Group Publishing Limited. Abstract: Despite previous work on expert systems in the manufacturing sector very few articles relate to a classification scheme for the application of Expert Systems (ES) in the service operations area. Based on a review of more than 200 abstracts from the ABI/INFORM database, seeks to provide a categorization of ES in the service sector and also to detail future needs in the application of this vital information technology in service operations. Enhances the manager′s understanding of ES technology and provides a perspective on how ES have been used in service operations. Keywords: Databases | Expert systems | Information management | Information technology | Service industries Article: Several survey, classification and prospect articles have been written on the use of Expert Systems (ES) technology in the manufacturing sector[1,2]. However, very few papers present a classification of the use of ES in the service sector. As the world economy moves to becoming more and more service-oriented, it is expected that ES technology would be utilized extensively in service operations. This article takes a first step by providing a categorization of ES in service operations. This categorization allows the reader to identify service areas where ES are used extensively and areas where more ES need to be developed. Even though service operations and manufacturing operations appear to be similar on a superficial level, there does exist some difference between the two on more than one dimension[3]. First, a study of the characteristics of service operations reveals that services, compared with manufacturing, are people-oriented. This means that there is not only a greater https://libres.uncg.edu/ir/uncg/clist.aspx?id=775 http://dx.doi.org/10.1108/02635579210019820 https://libres.uncg.edu/ir/uncg/ interaction between the employees of the firm and the customers, but there is greater customer participation in the service delivery process. For example, in some restaurants, after placing an order for the entrees, the customers could go over to the salad bar and prepare a salad for themselves. This indicates that customers get actively involved in the service provision. Second, there is simultaneity of service provision and consumption in service operations as compared to manufacturing operations. This indicates that while manufactured products can be inventoried, services cannot be "stocked up" for use at a later point in time. In a service operation, the customers' "consume" the service while it is being provided. Finally, services by default are more intangible than manufactured goods. The intangibility in services leads to unstructured decision- making processes, which require the use of a subjective heuristic on the part of the service provider, to solve intangible problems quickly and effectively. Analogies between services and manufacturing can be superficial and misleading because they tend to ignore the unique, "people-changing" and "people-processing" nature of services[3]. "People-changing" organizations attempt to alter directly the characteristics or behaviour of their customers through the application of various modification and treatment technologies. "People- processing" organizations, on the other hand, attempt to change their clients, not by altering basic personal attributes, but by conferring upon them a public status and relocating them in a new set of social circumstances. Hence, within this service context, not only is there a need to find new and creative approaches to the study of service operations management, but there is also a need to develop innovative uses of computer-based information systems, for both on-line and off-line operations in services depending upon the nature of services provided. A categorization of all service operations has been provided in Figure 1. This categorization is a modified version of the Fitzsimmons and Sullivan model[3]. Expert Systems and their Importance in Service Operations Over the last decade, no other area in information systems has received more attention than ES. ES are a branch of artificial intelligence (AI) that focuses on leveraging the human expertise available within an organization to gain competitive advantage. They do so by capturing the problem-solving know-how of experts, to automate corporate expertise and make it available on a constant basis throughout the organization[4]. The basic philosophy of expert systems has been well represented by Turban[5] as follows: The basic idea behind ES is simple. Expertise, which is the vast body of task-specific knowledge, is transferred from the human to the computer. This knowledge is then stored in the computer and users call on the computer for specific advice as needed. The computer can make inferences and arrive at a specific conclusion. Then, like a human consultant, it advises the non-experts and explains, if necessary, the logic behind the advice. Organizations can use ES to leverage scarce human expertise to improve customer services due to the fact that ES have a portion of the expert's knowledge which can be used to solve well- defined to semi-structured problems in a consistant manner. For example, American Express's expert system (AMEX) for credit approval provides the novice and inexperienced telephone operators with a round-the-clock access to the credit approval knowledge of experienced credit officers[6]. Thus, just as the ATMs provide a 24-hour access to banking services, ES provide a 24-hour access to the expert's knowledge within the organization. There are several reasons why ES are ideally suited for service operations. First, ES are highly interactive programs. That is, an end-user can demand a must-run explanation or clarification at any time while running an application. This makes ES an ideal technology for most service operations, which are people-oriented, as service consumers or clients play an active role, in terms of providing information and asking questions, in the facilitation of the service. This characteristic of ES is ideally suited for the simultaneous provision and consumption component of service operations. For example, the American Express credit-card holder can demand an explanation from AMEX as to why his credit limit was not extended for a credit-card sale. AMEX then displays from its working memory, the heuristic used in the decision-making process. This also helps American Express in training its new operators. Second, ES are much faster in service delivery than their human counterparts. This facet of ES makes them more suitable for services where speed and timeliness are critical to the provision of the service. For example, American Express receives thousands of calls every minute for credit limit extensions, 24-hours a day. Prior to the introduction of the AMEX system, the operator would have to provide customer information to the credit officer on duty and get his/her approval for every call. During peak load hours a customer would have to wait several minutes to get an approval. Now, with the AMEX system installed on every operator's terminal, the response time has been reduced to a matter of seconds. Third, ES do not fatigue and hence can provide consistent and round-the-clock service to service providers and consumers. For example, AMEX can operate consistently at any hour of the day without taking any breaks. Finally, ES use knowledge of the best "experts" in the organization and hence ES guarantee the best quality service to clients all the time. For example, AMEX uses credit approval knowledge of the best loan officers, not just any loan officer. The above-mentioned reasons suggest that there is as much need, if not more, for the use of ES in the decision-making process of service operations as in manufacturing operations. Since there is a greater degree of subjectivity and intangibility in service operations (as compared to manufacturing operations), it is all the more essential that important decision-making tools such as ES be made use of in service operations. There is not only a great deal of subjectivity in the decision-making process of the service provider, but there is an enormous amount of subjectivity and naivety on the part of the service consumer. This aspect of the service sector underscores the fact that the ES could be extremely useful in the decision-making process of both the service provider and the service consumer. The service-providing company can disseminate enhanced decision-making skills right down to the level of the customer. This is due to the fact that expert's knowledge can be easily duplicated and distributed through an ES to all employees who are involved in directly serving the customer. For example, Toyota Motors service department has an ES which can be accessed by all its service centres for diagnosing car-care problems[6]. This ensures best quality, consistent and friendly service on the part of the provider. ES also benefit the service consumer because the consumers can seek answers from the ES to, what the consumers may consider, even apparently naive questions without being concerned about embarrassing themselves and at the same time receive consistent explanations of how and why the system made a certain recommendation. These explanations can serve as a training (people- changing) device for the naive consumer in areas such as auto repairs and home improvement where a consumer has to keep up with the new tools and technologies. Quite similar to manufacturing operations, the decision-making process in service operations also involves a wide variety of criteria, and the decisions tend to be "fuzzy" or based on intuition[3]. Thus, there is a greater requirement for developing ES based on heuristic rather than optimizing algorithms to solve problems in service operations. The usage of the latter process no doubt would enable the system to come up with more accurate decisions but would entail the development and processing of data based on a large number of criteria and constraints. The development and processing of information based on a large number of criteria and constraints may not be feasible due to the "fuzzy" nature of the problem-solving processes in service operations. Categorization of Expert Systems in Service Operations In the past few years, several articles have been written on the use of ES in various service- oriented organizations. The use of ES in services can be classified on the basis of the specific area of application within the service sector. Service operations management applications of ES presented in articles catalogued in published sources[6,7,8], and abstracts of articles listed in ABI/INFORM (a CDROM library search facility from UMI Corporation), have been categorized using the framework shown in Figure 1. The published sources provide a list of ES applications before 1986 while the ABI/INFORM source provides a list of ES applications from 1986 to 1990. As the number of sources from which the list of EX applications has been generated is limited, it is acknowledged by the authors that this categorization may not include all the ES applications in service operations. However, it is the belief of the authors that this categorization is representative of the total population of ES applications in service operations. In all over 200 articles of ES applications were reviewed from both the sources. A brief outline of the categorization is presented in Table I. A general description of the findings from this categorization is presented in the remainder of this section. Transport Services Transport services, which constitute one of the important sectors of the economy, include services related to air, road, rail and sea transport. There are several engineering applications of ES (such as engine vibration diagnosis) in the transport industry. However, there are no applications of ES that relate to the supply of transport services to consumers from a managerial/business standpoint. There are a few ES in airline, motor freight, and car rental agencies, but no ES were found in the shipping and railroad industries. Communication Services In recent years, communication services have been fast developing into a vital sector of the service economy. The communications industry includes television services, telecommunications and computer services, postal services, and news media services. ES have been extensively used in the telecommunications and computer services, postal services and news media services but no ES were found in television services. Warehousing and Retailing Services Warehousing and retailing services constitute a vital link between the manufacturers and the consumers. Value is added by the warehousing and retailing firms through the process of storage and exchange. In recent years, a few ES have been developed in the retailing industry, but no ES were found in the warehousing industry. Financial Services One of the most popular sectors of the service economy where ES have been developed extensively is in insurance, finance, banking and accounting. Three noteworthy knowledge systems under development for mainframe use include: (1) the Charge Card Approval knowledge system from American Express Co., (2) Financial Product Advisor from the Travelers Corporation, and (3) the Lending Advisor from Syntelligence Inc.[9]. Nearly 25 per cent of the nation's 100 leading insurance companies surveyed by Coopers & Lybrand are now using ES. The survey also showed that 65 per cent of the firms are involved in some way with ES. The study revealed that some 70 per cent of life and health insurance companies are involved in ES technology as compared to 56 per cent of their property-casualty counterparts[10]. Most ES applications in the insurance industry are custom-designed. Some in the field predict that some time in the future, off-the-shelf systems will be available[11]. Over the next decade, ES are expected to be applied to virtually all the areas of life insurance information services which require expert knowledge and decision making. The growth of ES technology in banking services has been slow, and most bankers seem to be reluctant to embrace it fully. One of the major reasons for the sluggish growth of ES technology in banking applications is a shortage of mentors or bankers to teach their skills to computers [12]. One other reason is that the research and development efforts and test runs are taking much longer than had been anticipated. It has been reported that banks in the UK are also slowly developing ES for dealing room applications such as: (1) spotting arbitrage opportunities; (2) hedging block equity trading; and (3) predicting price movements[13]. It has been predicted that, by 1993, about 80 per cent of all banks will be using ES in some form[14]. Some banks have started to use small-scale ES to enhance the credit analysis and loan review processes[15]. Accommodation Services Accommodation services provide physical (in most cases) and psychological (in some cases) comfort to customers. Some of the important service organizations which fall under this category are hotels and restaurants. While ES have found their way into hotel services, none have been identified in the restaurant services. As people tend to have more and more time devoted to games and sporting events, this sector of society has no limits in terms of growth and potential. The games and sporting industry includes a variety of organizations ranging from ski resorts to sporting event franchises. It has been noted that ES were developed to schedule events to meet a multi-criteria objective at the 1992 Barcelona Summer Olympic Games [16] Scientific Services Organizations that offer scientific services vary from research and development establishments to weather bureaus. ES are being developed for application in these organizations as well. ES have been developed in a large UK industrial research and development establishment to prioritize the projects to allocate the scarce resources[17]. Due to the fact that the prediction of the weather is an inexact science, several weather bureaus are considering the use of ES technology in their operations. Legal Affairs and Government In recent times, a large amount of computerization has been undertaken in government organizations. With this, it is very well expected that the next step would be the development and usage of ES in the functioning of the government. A few ES are being considered in the legal affairs area, but none have been found in government offices. People-processing Services As stated in an earlier section above, people-processing services attempt to change their clients by conferring upon them a public status and relocating them into a new set of social circumstances. Services that are included under this category are employment services, credit- rating services, and medical diagnostic services. While there are only a few ES in employment services, and credit-rating services, there are many ES in medical services. In fact, the medical services area has been one of the pioneering areas for ES. MYCIN was one of the first commercially used ES to diagnose patients with bacteraemia, meningitis and cystitis infections. There are 43 ES developed and commercially used in diagnostic services[7]. A few have been listed to give the reader a flavour of the kinds of ES in this area. People-changing Services As stated in an earlier section, people-changing services attempt to alter directly the attributes or behaviour of their clients through the application of various modification and treatment technologies. Services that fall under this category could be sought by customers on a voluntary, or an involuntary basis. In voluntary receipt, people-changing services, services are sought by the clients of their own volition. Services that fall under this category are educational/library services, and health/medical services. There are a few ES being developed in the educational/library services and at least ten ES have been identified in the health/medical services[7]. On the other hand, the involuntary, receipt peoplechanging services are services that are sought by the customers on an involuntary basis, that is, they are not sought by the clients of their own volition, but are provided to them by society for their wellbeing. Services that fall under this category are services provided by the hospital for the mentally ill/retarded and prisons. ES have not been developed in this area. Conclusions This article has presented a categorization of the applications of ES in service operations management. It should be noted, however, that the above categorization does not represent an exhaustive list of ES applied to service operations. The authors acknowledge that several applications of ES may have been omitted in this categorization, due to any one or a combination of following reasons. Some of the applications of ES may be: (1) not reported as ES but rather as decision support systems, etc; (2) proprietary in nature and hence might not have been reported publicly; (3) in the prototype stage and may not have been fully tested and implemented and therefore not reported. Despite the above limitations, this categorization serves as a basis for examining and comparing the ES applications in the service sector. This categorization can be used to identify new application opportunities for ES in service operations management. The categorization of ES in service operations indicates that in some areas of the service economy there is an abundance of ES applications. Some of those areas are telecommunication and computer services, finance, banking and insurance services, and medical services (both diagnostic and health care provision). On the other hand, very little has been done in applying ES in the areas of railroad, shipping, television and people-changing involuntary services, such as health services for the mentally ill or retarded, and prison services. It is quite likely that the above-mentioned sectors of the service economy may not be suitable for the application of ES. It is suggested that future research should focus on understanding the reasons for non-development of ES in these sectors of the economy. The categorization carried out here may turn out to be beneficial to service operations managers and ES specialists in more ways than one. First, this categorization helps service operations managers (in a specific sector of the economy) to identify the kinds of ES applications that have been developed in their area of interest. Second, using this categorization, ES specialists can identify opportunities for developing ES applications. Last of all, operations managers involved with a particular type of service could consider adopting ES applied in a different sector of the economy. References 1. Mertens, P. and Kanet, J.J., "Expert Systems in Production Operations: An Assessment", Journal of Operations Management, Vol. 6 No. 4, 1986, pp. 393-404. 2. Rao, H.R. and Lingaraj, B.P., "Expert Systems in Production and Operations Management: Classification and Prospects", Interfaces, Vol. 18 No. 6, 1988, pp. 80-91. 3. Fitzsimmons, J.A. and Sullivan, R.S., Service Operations Management, McGraw-Hill, New York, NY, 1982. 4. Peppard, J., "Corporate Knowledge-Based Systems: Implication and Impact", Management Decision, Vol. 27 No. 5, 1989, pp. 48-52. 5. Turban, E., Decision Support and Expert Systems Managerial Perspectives, Macmillan, New York, NY, 1990. 6. Feiganbaum, E., McCorduck, N. and Nii, L., The Rise of Expert Company, Times-Mirror, New York, NY, 1989. 7. Waterman, D.A., A Guide to Expert Systems, AddisonWesley, Reading, MA, 1986. 8. Harmon, P., Maus, R. and Morrissey, W., Expert Systems: Tools & Applications, John Wiley & Sons, New York, NY, 1988. 9. Vacca, J.R., "The Birth of Banking's Knowledge Systems", Banking Software Review, Vol. 12 No. 4, 1987, pp. 68-71. 10. Haggerty, A.G., "Future is Now for Washington WC Insurer", National Underwriter (Property/Casualty/ Employee Benefits), Vol. 92 No. 31, 1988, p. 15. 1 1. Benham, B.T., "The Promise of Technology", Best's Review (Property/Casualty), Vol. 88 No. 7, 1987, pp. 40-6, 121. 12. Fitch, T.P., "Teaching Computers to be Bankers", Bankers Monthly, Vol. 105 No. 5, 1988, pp. 51-4. 13. Kanji, S., "When is an Expert not an Expert?", Banker (UK), Vol. 138 No. 746, 1988, pp. 86- 8. 14. Belser, M.F., "Expert Systems put Expertise in Everyone's Hands", Financial Managers' Statement, Vol. 10. No. 5, 1988, pp. 43-5. 15. McGinn, C., "Computer Mentors give Loan Officers a Hand", ABA Banking Journal, Vol. 82 No. 11, 1990, pp. 45-53. 16. Andreu, R. and Corominas, A., "SUCCES92: A DES for Scheduling the Olympic Games", Interfaces, Vol. 19 No. 5, 1989, pp. 1-12. 17. Grainge, N.J. and Pearson, A.W., "Managing an In-House R&D Service Department", R&D Management (UK), Vol. 19 No. 1, 1989, pp. 27-45. 
work_lvv6byn3mnbh3pdw2x5apxx3be	----	ORIGINAL ARTICLE Bayes pulmonary embolism assisted diagnosis: a new expert system for clinical use Davide Luciani, Silvio Cavuto, Luca Antiga, Massimo Miniati, Simona Monti, Massimo Pistolesi, Guido Bertolini . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . See end of article for authors’ affiliations . . . . . . . . . . . . . . . . . . . . . . . . Correspondence to: Dr G Bertolini, GiViTI Coordinating Center— Istituto di Ricerche Farmacologiche ‘‘Mario Negri’’, Centro di Ricerche Cliniche per le Malattie Rare Aldo e Cele Daccò, Ranica (Bergamo) 24020, Italy; bertolini@marionegri.it Accepted 29 September 2006 . . . . . . . . . . . . . . . . . . . . . . . . Emerg Med J 2007;24:157–164. doi: 10.1136/emj.2006.037440 Background: The diagnosis of pulmonary embolism demands flexible decision models, both for the presence of clinical confounders and for the variability of local diagnostic resources. As Bayesian networks fully meet this requirement, Bayes Pulmonary embolism Assisted Diagnosis (BayPAD), a probabilistic expert systems focused on pulmonary embolism, was developed. Methods: To quantitatively validate and improve BayPAD, the system was applied to 750 patients from a prospective study done in an Italian tertiary hospital where the true pulmonary embolism status was confirmed using pulmonary angiography or ruled out with a lung scan. The proportion of correct diagnoses made by BayPAD (accuracy) and the correctness of the pulmonary embolism probabilities predicted by the model (calibration) were calculated. The calibration was evaluated according to the Cox regression–calibration model. Results: Before refining the model, accuracy was 88.6%. Once refined, accuracy was 97.2% and 98%, respectively, in the training and validation samples. According to Cox analysis, calibration was satisfactory, despite a tendency to exaggerate the effect of the findings on the probability of pulmonary embolism. The lack of some investigations (like Spiral computed tomographic scan and Lower limbs doppler ultrasounds) in the pool of available data often prevents BayPAD from reaching the diagnosis without invasive procedures. Conclusions: BayPAD offers clinicians a flexible and accurate strategy to diagnose pulmonary embolism. Simple to use, the system performs case-based reasoning to optimise the use of resources available within a particular hospital. Bayesian networks are expected to have a prominent role in the clinical management of complex diagnostic problems in the near future. D espite the recent improvements in diagnostic methods for thromboembolism, the diagnosis of pulmonary embolism remains challenging.1–8 Reasons include the high cost of accurate examinations, the different risk perception with techniques based on contrast media (like pulmonary angio- graphy,9 phlebography and spiral computed tomographic scan10 11), and the variability in terms of practical availability and even the performance of qualified people.12 13 Moreover, some observations may have a negative or positive effect on the value of further ascertainments, explaining why, for instance, pulmonary embolism can be hard to diagnose in patients with other cardiorespiratory diseases.14–17 Thus, the diagnosis of pulmonary embolism cannot be made without combining and interpreting a collection of investigations.18 One innovative approach to see how different findings influence medical hypotheses is offered by Bayesian or probabilistic networks.19 20 These can assist a decision flexibly, depending on the examinations that are available among a large range of choices.21 22 To exploit these innovations in the diagnosis of pulmonary embolism, we developed Bayes Pulmonary embo- lism Assisted Diagnosis (BayPAD), an evidence-based expert system23 composed of a probabilistic network focused on thromboembolic disease (fig 1). Once a patient’s findings are entered, the model provides the probability of pulmonary embolism and the information content of examinations still to be carried out. The information content is related to the ability of an examination to reduce the uncertainty about a diagnostic hypothesis, and can be assessed by the mutual information measure.24 The BayPAD suggestions are tailored to each patient’s characteristics and each centre’s facilities. The system was extensively validated through different steps, covering face and content validity, and qualitative comparison with inde- pendent experts’ suggestions. Here we used the data collected in the Prospective Investigative Study of Acute Pulmonary Embolism Diagnosis (PISA-PED) study on the diagnosis of pulmonary embolism25 to quantitatively validate and further improve the system. As BayPAD was designed to help doctors in correctly classifying suspected thromboembolic events, the primary index of performance is diagnostic accuracy, split into its two dimen- sions: sensitivity and specificity. Given that the model also indicates the probability of the disease, we analysed the ‘‘calibration’’, which evaluates the degree of correspondence between the estimated probability produced by the model and the patient’s true disease status. METHODS Data The PISA-PED study was a prospective observational study completed at an Italian tertiary hospital in 1996 on 750 consecutive patients. Eligibility was based on the clinical judgement of suspected pulmonary embolism according to six on-call pulmonary doctors, all of whom had experience with diagnosis of the disease.25 From the whole study population, a first group of 500 patients were randomly selected to further develop the model (training sample), with the second group of 250 patients employed only to evaluate the validity of the refined model (validation sample). In all patients pulmonary Abbreviations: BayPAD, Bayes Pulmonary embolism Assisted Diagnosis; PISA-PED, Prospective Investigative Study of Acute Pulmonary Embolism Diagnosis 157 www.emjonline.com o n A p ril 5 , 2 0 2 1 b y g u e st. P ro te cte d b y co p yrig h t. h ttp ://e m j.b m j.co m / E m e rg M e d J: first p u b lish e d a s 1 0 .1 1 3 6 /e m j.2 0 0 6 .0 3 7 4 4 0 o n 9 M a rch 2 0 0 7 . D o w n lo a d e d fro m http://emj.bmj.com/ embolism was confirmed by pulmonary angiography, or was ruled out when a lung scan showed no perfusional defects. In the training sample, perfusion lung scans were normal or near normal in 105 of 500 (21%) patients and abnormal in 395 (79%). Angiograms were positive for pulmonary embolism in 200 (40%) patients and negative in 191 (38%). In four patients who died before angiography could be done, the diagnosis was established at autopsy. Two of these patients had pulmonary embolism. Patients were aged 63.8 (14.5) years (range 15–91 years); 243 (49%) of them were men. Most patients (85%) were hospitalised at the time of study entry (table 1).26 Table 2 displays the variables of diagnostic interest the BayPAD system is able to cope with, as well as those available in the PISA-PED database. As the BayPAD system had 48 variables and the PISA-PED 34, the model could not be validated completely. Regardless of this limit, the bayesian characteristic of the network enables us to deal with missing information, without the need to input any observation merely because it is contemplated by the model. This feature explains the adaptability of the system to the diagnostic resources of a particular hospital or ward, and it is also what makes this retrospective analysis methodologically feasible. Overall, six variables were present in the PISA-PED study but not in our model. Five variables (‘‘chest pain’’, ‘‘pregnancy’’, ‘‘surgery’’, ‘‘electrocardiographic signs of right heart overload’’ and ‘‘pulmonary embolism’’) were expressed with more details in the network than in the study (three-level v two-level variables). To use the PISA-PED data, we simplified the network in these variables. The opposite happened for ‘‘perfu- sion lung scan’’, which was a three-level variable in the PISA- PED Study (‘‘no perfusional defects’’, ‘‘segmental defect’’ and ‘‘not segmental defect’’) but binary in the original network (‘‘normal’’ and ‘‘abnormal’’). The model was extended accord- ingly. Eventually, 28 variables were available for the validation analysis (table 2). Data were processed following the algorithm through which BayPAD implements its strategy (fig 2). The diagnostic criteria are based on two probabilistic cut-offs: probability .95% to confirm pulmonary embolism and ,5% to exclude pulmonary embolism, provided that all the examinations whose costs divided by the probability of pulmonary embolism do not exceed a conventional boundary of J3500 were already done.23 Once applied, these criteria become clearly asymmetric and more sensitive to false negatives, as many diagnostic procedures remain cost-sustainable even for probabilities ,5%. BayPAD Figure 1 A simplified Bayesian network showing the main pathophysiological relationships in the diagnostic reasoning focused on a suspected pulmonary embolism event. For the sake of clarity, observable findings have been given a generic label of symptom, sign, test or risk factor, and their association with the rest of the variables has been omitted. 158 Luciani, Cavuto, Antiga, et al www.emjonline.com o n A p ril 5 , 2 0 2 1 b y g u e st. P ro te cte d b y co p yrig h t. h ttp ://e m j.b m j.co m / E m e rg M e d J: first p u b lish e d a s 1 0 .1 1 3 6 /e m j.2 0 0 6 .0 3 7 4 4 0 o n 9 M a rch 2 0 0 7 . D o w n lo a d e d fro m http://emj.bmj.com/ suggests pulmonary angiography when the information content of the examinations not carried out is too low to allow a diagnostic conclusion (fig 2). As a result, we had the final BayPAD diagnosis for each patient and one predicted probability of pulmonary embolism for each patient step (fig 3, where the algorithm has been applied to two real clinical examples). Patients whose diagnosis was reached by asking for this ‘‘standard’’ were identified. Hardware requirements for the analysis consist of a central processing unit of at least 233 MHz and 64 Mb random access memory. Diagnostic performance BayPAD’s ability to classify cases correctly was assessed from sensitivity, specificity and overall diagnostic accuracy. We computed these indicators for the whole sample and for the subsample of patients whose diagnosis was reached without pulmonary angiography. To measure how well the model was calibrated, we adopted the approach introduced by Cox.27 Labelling p the probabilities computed by the model, a logistic regression was done with the outcome of pulmonary embolism as the dependent variable and the natural logarithm transformation of the ratio p/(1-p) as the independent regressor (ie, the ordinary logit transformation). If the accuracy of p is faultless, the estimates of the intercept and the slope are 0 and 1, respectively. An intercept different from 0 would show a systematic disagreement between the probabil- ities produced by the model and the proportion of pulmonary embolism cases. A positive slope lower or higher than 1 would show the predicted probabilities, respectively, increasing or decreasing compared with the observed density of pulmonary embolism events. A null slope would mean that predicted probabilities are completely independent of the outcomes, whereas a negative slope would prove a negative association.28 Finally, a calibration curve was drawn by plotting deciles of predicted probabilities against the corresponding proportion of pulmonary embolism events. Refinement and validation of the model As diagnostic accuracy is sensitive to biases affecting predicted probabilities, the network was refined by looking at its calibration in the 500 cases representing the training sample. We followed the approach suggested by Miller et al.28 Firstly, a sensitivity analysis was performed by excluding cases with identical characteristics. As a result, each case was associated with a measure of its effect on calibration, considering both Cox parameters. Secondly, these measures played the part of dependent variable in two separate linear regression models, where the influential patient’s characteristics (on calibration) were identified. Finally, the network parameters related to these characteristics were reappraised and tuned, keeping a modification only when it conformed to the medical literature and enhanced the validity of the model. As the structure of the network represents the cause–effect relationships among variables (fig 1), if an improvement was attainable with structural changes, these were discussed and allowed only if consistent with the pathophysiological understanding of pulmonary embolism. Such a conservative approach in chan- ging the probabilistic network was adopted to reduce the chance of overfitting the PISA-PED data. The process was repeated until influential variables and convincing refinements could be detected. Diagnostic performance was evaluated on both the original and the refined model, with the additional employment of a validation (250 cases) sample as it concerns the latter. RESULTS The original network Among the 500 cases provided by the training sample, there were 40.4% of pulmonary embolism cases. BayPAD made a correct diagnosis in 88.6% of the cases (accuracy), with 17 false negative and 40 false positive cases. This figure can be divided into 91.6% (95% CI 86.9 to 94.7) cases of correct diagnosis among true pulmonary embolism cases (sensitivity) and 86.6% (95% CI 82.2 to 90) cases of correct diagnosis among truly non- pulmonary embolism cases (specificity); 152 (30.4%) cases required pulmonary angiography to reach the final diagnosis. In the subgroup in which pulmonary angiography was not used, accuracy was 83.6%, sensitivity 88.8% and specificity 79.6%. When the BayPAD strategy was applied to the data, each case passed through an average of 3.3 further ascertainments before the final diagnosis, producing a total of 1660 steps where predicted probabilities were computed. Cox analysis indicated that the intercept was 20.234 (95% CI –0.364 to –0.104), significantly different from 0 (p,0.001), and the slope 0.2091 (95% CI 0.182 to 0.236), significantly different from 1 (p,0.001). The refined network Several phases of parameter tuning were done to increase the validity of the model. Most of them affected the quantitative strength of associations between variables, and others the increase in the number of discrete intervals for continuous variables (like paO2 and paCO2). Structural changes were introduced to account for previously neglected associations, like that for ‘‘bone fractures’’, found to be an extra explanation of ‘‘unilateral oedema’’. After refinement, BayPAD showed 97.2% accuracy in the 500 cases of the training sample, with six false-negative and eight false-positive cases. Sensitivity and specificity were 97.0% (95% CI 94.2 to 98.7) and 97.3% (95% CI 95.2 to 98.7); 187 (37%) cases required pulmonary angiography to reach the final diagnosis. Accuracy was 96.0%, with 96.0% sensitivity and 95.9% specificity in the subgroup in which pulmonary angiography was not needed. Concerning calibration, the intercept and slope of the Cox approach were, respectively, –0.05 (95% CI –0.08 to 0.18), not significantly different from 0 (p = 0.47), and 0.62 (95% CI 0.56 to 0.68), significantly different from 1 (p,0.001). Figure 4 shows the calibration curve of the refined model. Each decile had more than five expected cases of pulmonary embolism. The validation sample showed a proportion of pulmonary embolism cases (41.6%) comparable to that which emerged from the training sample. BayPAD delivered an accurate Table 1 Characteristics of the first 500 patients enrolled in the Prospective Investigative Study of Acute Pulmonary Embolism Diagnosis Study n (%) Dyspnoea 345 (69) Chest pain 240 (48) Fainting 91 (18) Palpitations 82 (16) Hospitalisation 437 (85) Surgery* 197 (39) Bone fractures (lower limbs)* 82 (16) Pre-existing diseases Cardiovascular 149 (30) Pulmonary 86 (17) Neoplastic 79 (16) Endocrine 53 (11) *Within 4 weeks of study entry. The diagnosis of pulmonary embolism 159 www.emjonline.com o n A p ril 5 , 2 0 2 1 b y g u e st. P ro te cte d b y co p yrig h t. h ttp ://e m j.b m j.co m / E m e rg M e d J: first p u b lish e d a s 1 0 .1 1 3 6 /e m j.2 0 0 6 .0 3 7 4 4 0 o n 9 M a rch 2 0 0 7 . D o w n lo a d e d fro m http://emj.bmj.com/ Table 2 Variables selected in the validation analysis Variables BayPAD included in the networkAvailable in the PISA-PED Study Selected for the validation Pre-existing disease Chronic cardiovascular disease Yes Yes Yes Chronic pulmonary disease Yes Yes Yes Cancer Yes Yes Yes Hereditary DVT predisposing factors Yes No No Risk factors Age Yes Yes Yes Sex Yes Yes Yes Immobilisation Yes Yes Yes Cigarette smoking Yes No No Central line Yes No No Surgery Yes Yes Yes Pregnancy Yes Yes Yes Chronic venous insufficiency Yes Yes Yes Oestrogen intake Yes Yes Yes Bone fractures Yes Yes Yes DVT prophylaxis Yes No No Symptoms Dyspnoea Yes Yes Yes Orthopnoea Yes Yes Yes Fainting Yes Yes Yes Chest pain Yes Yes Yes Haemoptysis Yes Yes Yes Agitation Yes No No Lower limb discomfort Yes No No Cough Yes No No Signs Lower limb unilateral oedema Yes Yes Yes Cyanosis Yes No No Fever.38 C̊ No Yes No Bronchospasmus Yes Yes Yes Systolic blood pressure Yes No No Tachypnoea Yes Yes Yes Tachycardia No Yes No Cardiac failure signs Yes Yes Yes Shock Yes No No Laboratory findings D-dimer test Yes No No paO2 Yes Yes Yes paCO2 Yes Yes Yes CK-MB enzymes Yes No No ECG findings ECG right heart overload signs Yes Yes Yes ECG acute myocardial infarction signs Yes No No Chest x rays Consolidation (infarction) Yes Yes Yes Consolidation (no infarction) No Yes No Plate-like atelectasis No Yes No Pulmonary oedema Yes Yes Yes Elevation of half of the diaphragm Yes Yes Yes Unilateral pleural effusion Yes No No Pulmonary oligaemia No Yes No Amputation of hilar artery No Yes No Pneumonia Yes No No Imaging Doppler ultrasounds of the lower limbs Yes No No Echocardiographic signs of pulmonary embolism Yes No No Ventilation/perfusion lung scan Yes No No Perfusion lung scan Yes Yes Yes Phlebography Yes No No Spiral CT scan Yes No No Pulmonary angiography Yes Yes Yes BayPAD, Bayes Pulmonary embolism Assisted Diagnosis; CK-MB, Creatine Kinase; CT, Computed tomography; DVT, Deep venous thrombosis; ECG, electrocardiogram; PISA-PED, Prospective Investigative Study of Acute Pulmonary Embolism Diagnosis. 160 Luciani, Cavuto, Antiga, et al www.emjonline.com o n A p ril 5 , 2 0 2 1 b y g u e st. P ro te cte d b y co p yrig h t. h ttp ://e m j.b m j.co m / E m e rg M e d J: first p u b lish e d a s 1 0 .1 1 3 6 /e m j.2 0 0 6 .0 3 7 4 4 0 o n 9 M a rch 2 0 0 7 . D o w n lo a d e d fro m http://emj.bmj.com/ diagnosis in 98% of the 250 cases, with four false-negative and one false-positive cases. Sensitivity and specificity were 96.1% and 99.3%; pulmonary angiography was required for 79 (31%) cases. Sensitivity and specificity were 94.5% and 98.9% in the subgroup in which pulmonary angiography was not needed. The intercept and slope of the Cox regression–calibration model were, respectively, 0.31 (95% CI 0.10 to 0.52), significantly different from 0 (p = 0.003), and 0.70 (95% CI 0.61 to 0.80), significantly different from 1 (p,0.001). DISCUSSION BayPAD is an expert system based on a probabilistic network whose structure represents the ‘‘state of the art’’ on the pathophysiological understanding of thromboembolism (fig 1). Thus, within its automated diagnostic reasoning, the system acknowledges which part of the network is relevant to a decision, given the specific patient’s findings. On the basis of the causal relationships between events, a virtually infinite number of clinical scenarios can be dealt with, whereas the computation to identify the most appropriate examination takes just a few milliseconds.19 Conversely, in guidelines based on decision trees, whether or not an examination is appropriate is established through a limited set of predefined patient’s findings. So if the performance of these algorithms proves to be satisfactory at a population level, they are often inappropriate when applied to the clinical investigation of individual problems.29 The other popular aid to medical diagnosis is scoring systems, indicating the need for further ascertainments according to the predicted probability of the diagnosis. However, they overlook the possible influence of available findings on the results of the examinations suggested. As an example, recent surgery increases the risk of pulmonary embolism, but it also reduces the specificity of the D-dimer test 30; again, previous cardio- pathy makes an echocardiography or an ECG more sensitive to an embolic episode, because in these patients a haemodyna- mically relevant pulmonary embolism is more likely.17 31 In the BayPAD model, these phenomena are taken into account. Finally, the Bayesian nature of the probabilistic network allows BayPAD to deal with missing information.32 The consequent flexibility makes it possible to deal with the problem of optimal exploitation of available diagnostic resources. This is important, as resources are usually so differently distributed among the clinical settings where pulmonary embolism is a challenge that it is hard to expect widespread acceptance of any guideline on the basis of a fixed set of examinations.13 The potential value of probabilistic networks in this field has been theoretically accepted,21 22 33 but their successful application obviously depends on their diagnostic accuracy over different contexts. Usually, to validate a prediction model, all the variables considered must be collected, without missing. However, a Bayesian Network can be regarded as the assemblage of smaller networks, allowing independent validation of each part. Such an approach is safe Figure 2 The Bayes Pulmonary embolism Assisted Diagnosis (BayPAD) strategy: the algorithm implementing the BayPAD diagnostic strategy on the PISA-PED cases. The diagnosis of pulmonary embolism 161 www.emjonline.com o n A p ril 5 , 2 0 2 1 b y g u e st. P ro te cte d b y co p yrig h t. h ttp ://e m j.b m j.co m / E m e rg M e d J: first p u b lish e d a s 1 0 .1 1 3 6 /e m j.2 0 0 6 .0 3 7 4 4 0 o n 9 M a rch 2 0 0 7 . D o w n lo a d e d fro m http://emj.bmj.com/ as long as the structure of the network represents the causal relationships among the events included in the model.34 Consequently, even a quite old database like PISA-PED is of value for a validation purpose. Indeed, although some important tests are lacking (eg, computed tomographic scan or D-dimer), the PISA-PED data allowed the assessment of the most complex part of the model. Moreover, since in a Bayesian Network the overall proportion of an event depends on the conditional observations included, the large set of variables considered in BayPAD makes the possible difference between the validation sample and another population of interest relatively unimportant. As a matter of fact, BayPAD, after its refinement based on the PISA-PED data, still returns a proportion of pulmonary embolism cases before any observa- tion is introduced which is around 1%. This resembles the prevalence of pulmonary embolism in a general hospital, rather than the prevalence expected given a clinical suspicion of pulmonary embolism, like in the PISA-PED study. Here the assessment of the system passed through the evaluation of two major properties: predicting a reliable probability of pulmonary embolism for a specific patient, and distinguishing between true and false cases of pulmonary embolism. These are closely related, as the second is obtained through a couple of probabilistic cut-offs that are expected to be reliable. Therefore, the calibration of BayPAD has been examined first, providing the only evidence used to refine the model. Rather than looking at a qualitative response in the calibration of our model, like within a significance testing framework,35 36 our aim was to see how accurate the prob- abilities predicted by the model were. The approach introduced Figure 3. Bayes Pulmonary embolism Assisted Diagnosis (BayPAD) working with two real clinical examples. The algorithm steps presented in fig 2 are indicated in brackets. Figure 4 The calibration curve with observed probabilities (frequencies) of pulmonary embolism events plotted against predicted probabilities of the same event. Data are grouped into deciles according to the predicted probabilities. The dotted line represents perfect calibration, whereas the green (with triangles) and the blue (with squares) curves, respectively, refer to the training and validation samples. 162 Luciani, Cavuto, Antiga, et al www.emjonline.com o n A p ril 5 , 2 0 2 1 b y g u e st. P ro te cte d b y co p yrig h t. h ttp ://e m j.b m j.co m / E m e rg M e d J: first p u b lish e d a s 1 0 .1 1 3 6 /e m j.2 0 0 6 .0 3 7 4 4 0 o n 9 M a rch 2 0 0 7 . D o w n lo a d e d fro m http://emj.bmj.com/ by Cox provided us with this quantitative insight.27 Once applied to the original network, the analysis gave an intercept of –0.23, meaning that, on average, the probability of pulmonary embolism exceeded its observed frequency. Instead, the slope of 0.21 showed a tendency of the low probabilities to be too low and that of the high probabilities to be too high, resulting in overconfidence in the diagnosis. Looking at the effect of this calibration on the diagnostic accuracy in the studied population, we found correct classifica- tion for 88.6% of the suspected cases. The complete independence of the PISA-PED study from the development of the model, the heterogeneity of the sources in the medical literature fuelling the network’s knowledge37 and the qualitative assessment of the network’s behaviour as its sole former validation23 make this figure encouraging. However, this estimation is also inconsistent with the probabilistic values adopted to authorise a diagnosis, which would allow ,10% of diagnostic errors. As expected, this inadequacy was treated as a model recalibration problem. After the refinement analysis, the slope parameter moved near to its ideal value of 1, both in the training and in the validation sample (0.62 and 0.70, respectively). This means that the original tendency to exaggerate the effect of the observations on the pulmonary embolism hypothesis has been greatly reduced. The intercept parameter, on the other hand, shows that, in the validation sample, the predicted probabilities tend to be overall too low. Particularly, an intercept of 0.3 entails for predicted probability around 50% to be undersized of 7%. How the residual bias affects the diagnostic accuracy is easily visualised on the calibration curve (fig 4). Misclassified cases became ,10% (2.8% and 2%, respectively, in the training and validation samples), and this remains true when accuracy was measured in the 347 cases where pulmonary angiography was not suggested (4%). The lack of important investigations in the pool of available data (table 2) often prevented BayPAD from suggesting minimally invasive but still informative tests, with the result that pulmonary angiography was suggested for about one third of the patients. The many investigations not contemplated in the PISA-PED study include Doppler ultra- sound of the lower limbs, echocardiography, D-dimer and spiral computed tomographic scan. As these procedures are available in most hospitals, the resulting proportion of pulmonary angiography seems not to reflect the behaviour of BayPAD, once introduced in the clinical setting (eg, fig 3 shows two real cases where BayPAD evaluates the appropriateness of D-dimer and spiral computed tomographic scan). Moreover, given the progress achieved by the latest multidetector computed tomo- graphic scans,38 this technique is expected to replace pulmonary angiography in most instances, while still supporting the diagnostic engine with similar precision. To confirm this, we simulate the effect of the availability of a computed tomo- graphic scan with 0.92 sensibility and 0.98 specificity,3 6 obtaining a correct diagnosis in 96% of all cases. The literature offers a number of diagnostic strategies for the diagnosis of pulmonary embolism,25 39–44 but it is hard to compare their performances. Studies to validate the proposed algorithms apply different procedures to check the true diagnosis25 43: some are focused on the prognostic outcome rather than the diagnostic classification.42 44 They mostly differ in terms of available diagnostics,25 39–41 43 and classify patients according to a variable number of pulmonary embolism probability levels.25 40 Without perfusion lung scan, but with a clinical assessment extended to chest x rays, ECG and arterial blood samples, Miniati et al found their algorithm correct in 90% of suspected cases.25 On the basis of similar findings, but with the inclusion of the ventilation/perfusion lung scan and bilateral leg vein ultrasonography, the Wells score correctly identified 96% of cases.39 Another study showed a predictive accuracy for the Geneva score and a simplified version of the Wells score, both based on clinical findings alone, of 78% and 74%, respectively.40 41 The BayPAD system deals with most of the issues raised by a complex diagnostic problem like pulmonary embolism. To the best of our knowledge, this is the first validated clinical model where the choice of a new ascertainment depends on its information content.24 45 Moreover, the probabilistic network underlying the expert system covers the largest set of examinations ever contemplated for the disease. This already enables BayPAD to deal with hypotheses other than pulmonary embolism. Such a possibility will be fully exploited in a future version, where the most appropriate ascertainment will depend on a broad spectrum of possible diagnoses, without privileging pulmonary embolism. Growing to be a multi-purpose system, BayPAD could be accepted even in the emergency department, where time constraints and overcrowding often hamper the use of computer-based systems. Although a prospective investigation is still needed to evaluate BayPAD in its full potential, the results obtained with this model, even at an initial phase of its development, reserve for Bayesian networks a prominent role in the next generation of clinical decision models. This prospect sounds like a proper reply to some farsighted authors who, like Feinstein, 10 years ago advocated ‘‘appropriate scientific analyses for the unique and fundamental characteristics of clinical activities that still occur as ‘clinical judgement’’’.46 ACKNOWLEDGEMENTS We thank Professor Phil Dawid for his constructive comments, and Judy Baggot for the revision of the manuscript. Authors’ affiliations . . . . . . . . . . . . . . . . . . . . . . . Davide Luciani, Silvio Cavuto, Unit of Clinical Knowledge Engineering, Laboratory of Clinical Epidemiology, ‘‘Mario Negri’’ Institute of Pharmacological Research, Clinical Centre for Rare Diseases Aldo e Cele Daccò, Ranica (Bergamo), Italy Luca Antiga, Laboratory of Biomedical Technologies, ‘‘Mario Negri’’ Institute of Pharmacological Research, Clinical Centre for Rare Diseases Aldo e Cele Daccò, Ranica (Bergamo), Italy Massimo Miniati, Simona Monti, Department of Clinical Physiology, Italian National Research Council, Pisa, Italy Massimo Pistolesi, Department of Critical Care, University of Florence, Firenze, Italy Guido Bertolini, Laboratory of Clinical Epidemiology, Istituto di Ricerche Farmacologiche ‘‘Mario Negri’’, Centro ‘‘Aldo e Cele Daccò’’, Ranica (Bergamo), Italy Funding: This study was partially supported by Sanofi-Aventis Italy. Competing interests: None. REFERENCES 1 Chunilal SD, Eikelboom JW, Attia J, et al. Does this patient have pulmonary embolism? JAMA 2003;290:2849–58. 2 Laack TA, Goyal DG. Pulmonary embolism: an unsuspected killer. Emerg Med Clin North Am 2004;22:961–83. 3 Revel MP, Petrover D, Hernigou A, et al. Diagnosing pulmonary embolism with four-detector row helical CT: prospective evaluation of 216 outpatients and inpatients. Radiology 2005;234:265–73. 4 Rosen MP, Sands DZ, Morris J, et al. Does a physician’s ability to accurately assess the likelihood of pulmonary embolism increase with training? Acad Med 2000;75:1199–205. 5 van Beek EJ, Wild JM, Fink C, et al. MRI for the diagnosis of pulmonary embolism. J Magn Reson Imaging 2003;18:627–40. 6 Reinartz P, Wildberger JE, Schaefer W, et al. Tomographic imaging in the diagnosis of pulmonary embolism: a comparison between V/Q lung scintigraphy in SPECT technique and multislice spiral CT. J Nucl Med 2004;45:1501–8. 7 Smith TP. Pulmonary embolism: what’s wrong with this diagnosis? Am J Roentgenol 2000;174:1489–97. 8 Calder KK, Herbert M, Henderson SO. The mortality of untreated pulmonary embolism in emergency department patients. Ann Emerg Med 2005;45:302–10. The diagnosis of pulmonary embolism 163 www.emjonline.com o n A p ril 5 , 2 0 2 1 b y g u e st. P ro te cte d b y co p yrig h t. h ttp ://e m j.b m j.co m / E m e rg M e d J: first p u b lish e d a s 1 0 .1 1 3 6 /e m j.2 0 0 6 .0 3 7 4 4 0 o n 9 M a rch 2 0 0 7 . D o w n lo a d e d fro m http://emj.bmj.com/ 9 Stein PD, Athanasoulis C, Alavi A, et al. Complications and validity of pulmonary angiography in acute pulmonary embolism. Circulation 1992;85:462–8. 10 Hall-Craggs MA, Hine AL. Ascending lower-limb phlebography: a comparative study of Hexabrix 320, iohexol 300 and iohexol 240. Br J Radiol 1986;59:685–7. 11 Parry RA, Glaze SA, Archer BR. The AAPM/RSNA physics tutorial for residents. Typical patient radiation doses in diagnostic radiology. Radiographics 1999;19:1289–302. 12 Stein PD, Henry JW, Gottschalk A. Reassessment of pulmonary angiography for the diagnosis of pulmonary embolism: relation of interpreter agreement to the order of the involved pulmonary arterial branch. Radiology 1999;210:689–91. 13 Fennerty T. Pulmonary embolism. Hospitals should develop their own strategies for diagnosis and management. BMJ 1998;317:91–2. 14 Kearon C. Diagnosis of pulmonary embolism. CMAJ 2003;168:183–94. 15 Hartmann IJ, Hagen PJ, Melissant CF, et al. Diagnosing acute pulmonary embolism: effect of chronic obstructive pulmonary disease on the performance of D-dimer testing, ventilation/perfusion scintigraphy, spiral computed tomographic angiography, and conventional angiography. ANTELOPE Study Group. Advances in New Technologies Evaluating the Localization of Pulmonary Embolism. Am J Respir Crit Care Med 2000;162:2232–7. 16 Perrier A, Perneger T, Cornuz J, et al. The COPD-PE study: prevalence and prediction of pulmonary embolism in acute exacerbations of chronic obstructive pulmonary disease. Rev Mal Respir 2004;21(Part 1):791–6. 17 Stein PD, Terrin ML, Hales CA, et al. Clinical, laboratory, roentgenographic, and electrocardiographic findings in patients with acute pulmonary embolism and no pre-existing cardiac or pulmonary disease. Chest 1991;100:598–603. 18 Kelly J, Hunt BJ. The utility of pretest probability assessment in patients with clinically suspected venous thromboembolism. J Thromb Haemost 2003;1:1888–96. 19 Pearl J. Probabilistic reasoning in intelligent systems: networks of plausible inference. San mateo, CA: Morgan Kaufman, 1988. 20 Jensen FV. Bayesian networks and decision graphs. New York: Springer Verlag, 2001. 21 Andreassen S, Jensen FV, Olesen KG. Medical expert systems based on causal probabilistic networks. Int J Biomed Comput 1991;28:1–30. 22 Spiegelhalter D. Probabilistic expert systems in medicine. Stat Sci 1987;2:3–44. 23 Luciani D, Marchesi M, Bertolini G. The role of Bayesian networks in the diagnosis of pulmonary embolism. J Thromb Haemost 2003;1:698–707. 24 Shannon C. A mathematical theory of communications. Bell Systems Tech J 1948;27:379–423. 25 Miniati M, Pistolesi M, Marini C, et al. Value of perfusion lung scan in the diagnosis of pulmonary embolism: results of the Prospective Investigative Study of Acute Pulmonary Embolism Diagnosis (PISA-PED). Am J Respir Crit Care Med 1996;154:1387–93. 26 Miniati M, Prediletto R, Formichi B, et al. Accuracy of clinical assessment in the diagnosis of pulmonary embolism. Am J Respir Crit Care Med 1999;159:864–71. 27 Cox D. Two further applications of a model for binary regression. Biometrika 1958;45:562. 28 Miller ME, Hui SL, Tierney WM. Validation techniques for logistic regression models. Stat Med 1991;10:1213–26. 29 Zorman M, Stiglic MM, Kokol P, et al. The limitations of decision trees and automatic learning in real world medical decision making. J Med Syst 1997;21:403–15. 30 Brown MD, Vance SJ, Kline JA. An emergency department guideline for the diagnosis of pulmonary embolism: an outcome study. Acad Emerg Med 2005;12:20–5. 31 Bova C, Greco F, Misuraca G, et al. Diagnostic utility of echocardiography in patients with suspected pulmonary embolism. Am J Emerg Med 2003;21:180–3. 32 Gelman A, Carlin J, Stern H, et al. Bayesian Data Analysis, 2nd edn.Chapman Hall/CRC, London, 2003. 33 Spiegelhalter DJ. Alternative formalisms for the representation of medical knowledge and clinical judgement. Hong Kong: First Hong Kong Medical Informatics Conference, 1991. 34 Heckerman D. A tutorial on learning with bayesian networks. Technical Report MSR-TR-95-06, Microsoft Research, Redmond, Washington, 1995, revised June 1996. 35 Hosmer D, Lemeshow S. Applied logistic regression. New York: John Wiley and Sons, Inc, 1989. 36 Hosmer DW, Hosmer T, Le Cessie S, et al. A comparison of goodness-of-fit tests for the logistic regression model. Stat Med 1997;16:965–80. 37 Druzdzel MJ, Diez JF. Criteria for combining knowledge from different sources in probabilistic models. Sixteenth Annual Conference on Uncertainty in Artificial Intelligence, Stanford, CA 2000. 38 van Belle A, Buller HR, Huisman MV, et al. Effectiveness of managing suspected pulmonary embolism using an algorithm combining clinical probability, D-dimer testing, and computed tomography. JAMA 2006;295:172–9. 39 Wells PS, Ginsberg JS, Anderson DR, et al. Use of a clinical model for safe management of patients with suspected pulmonary embolism. Ann Intern Med 1998;129:997–1005. 40 Wicki J, Perrier A, Perneger TV, et al. Predicting adverse outcome in patients with acute pulmonary embolism: a risk score. Thromb Haemost 2000;84:548–52. 41 Chagnon I, Bounameaux H, Aujesky D, et al. Comparison of two clinical prediction rules and implicit assessment among patients with suspected pulmonary embolism. Am J Med 2002;113:269–75. 42 Musset D, Parent F, Meyer G, et al. Diagnostic strategy for patients with suspected pulmonary embolism: a prospective multicentre outcome study. Lancet 2002;360:1914–20. 43 Donkers-van Rossum AB. Diagnostic strategies for suspected pulmonary embolism. Eur Respir J 2001;18:589–97. 44 Kruip MJ, Leclercq MG, van der Heul C, et al. Diagnostic strategies for excluding pulmonary embolism in clinical outcome studies. A systematic review. Ann Intern Med 2003;138:941–51. 45 Benish WA. The use of information graphs to evaluate and compare diagnostic tests. Methods Inf Med 2002;41:114–18. 46 Feinstein AR. ‘‘Clinical Judgment’’ revisited: the distraction of quantitative models. Ann Intern Med 1994;120:799–805. BNF for Children 2006, second annual edition In a single resource: N guidance on drug management of common childhood conditions N hands-on information on prescribing, monitoring and administering medicines to children N comprehensive guidance covering neonates to adolescents For more information please go to bnfc.org 164 Luciani, Cavuto, Antiga, et al www.emjonline.com o n A p ril 5 , 2 0 2 1 b y g u e st. P ro te cte d b y co p yrig h t. h ttp ://e m j.b m j.co m / E m e rg M e d J: first p u b lish e d a s 1 0 .1 1 3 6 /e m j.2 0 0 6 .0 3 7 4 4 0 o n 9 M a rch 2 0 0 7 . D o w n lo a d e d fro m http://emj.bmj.com/ 
work_lydzvx52yjdtdgjrphenz3g4uy	----	Title: 1 Does segmentation always improve model performance in credit scoring? Katarzyna Bijak a * and Lyn C. Thomas a a School of Management, University of Southampton, Southampton SO17 1BJ, UK * Corresponding author. Tel.: +44 23 80598964. E-mail address: K.Bijak@soton.ac.uk. Abstract Credit scoring allows for the credit risk assessment of bank customers. A single scoring model (scorecard) can be developed for the entire customer population, e.g. using logistic regression. However, it is often expected that segmentation, i.e. dividing the population into several groups and building separate scorecards for them, will improve the model performance. The most common statistical methods for segmentation are the two-step approaches, where logistic regression follows Classification and Regression Trees (CART) or Chi-squared Automatic Interaction Detection (CHAID) trees etc. In this research, the two-step approaches are applied as well as a new, simultaneous method, in which both segmentation and scorecards are optimised at the same time: Logistic Trees with Unbiased Selection (LOTUS). For reference purposes, a single-scorecard model is used. The above-mentioned methods are applied to the data provided by two of the major UK banks and one of the European credit bureaus. The model performance measures are then compared to examine whether there is improvement due to the segmentation methods used. It is found that segmentation does not always improve model performance in credit scoring: for none of the analysed real-world datasets, the multi-scorecard models perform considerably better than the single-scorecard ones. Moreover, in this application, there is no difference in performance between the two-step and simultaneous approaches. Keywords Credit scoring; Segmentation; Logistic regression; CART; CHAID; LOTUS mailto:K.Bijak@soton.ac.uk 2 1. Introduction Thomas et al. (2002) define credit scoring as “the set of decision models and their underlying techniques that aid lenders in the granting of consumer credit” (p. 1). These models and techniques are used to assess the credit risk of bank customers (individuals as well as small and medium enterprises). Depending on the data used to build models, there are different types of scoring. Application scoring is based on data from loan application forms while behavioural scoring is based on data on customers’ behaviour stored in bank databases. A special type of the latter is credit bureau scoring. Credit bureaus are institutions that collect and analyse data on loans granted by banks operating in a given country (Anderson, 2007; Van Gestel and Baesens, 2009). Such data enable tracking the credit history of a customer in the banking sector. Credit bureau scoring is based on data on customers’ credit histories. Application scoring can also be enriched with data from a credit bureau. As a rule, using such data increases performance of a scoring model (Van Gestel and Baesens, 2009). A scoring model describes the relationship between customer’s characteristics (independent variables) and his or her creditworthiness status (a dependent variable). A customer’s status can be either “good” or “bad” (and sometimes also “indeterminate” or “other”). The most common form of scoring models is referred to as a scorecard. According to Mays (2004), the scorecard is “a formula for assigning points to applicant characteristics in order to derive a numeric value that reflects how likely a borrower is, relative to other individuals, to experience a given event or perform a given action” (p. 63). Scorecards are used to calculate scores and/or probabilities of default (PD). They are sometimes scaled to obtain a required relationship between scores and PD. A scoring model can consist of one or more scorecards. In the latter case, it can be referred to as a suite of scorecards. In order to develop such a multi-scorecard model, segmentation has to be applied. It is commonly expected that segmentation will improve the model performance. Segmentation is often carried out using the two-step approaches, where logistic regression follows Classification and Regression Trees (CART) or Chi-squared 3 Automatic Interaction Detection (CHAID) trees. In this research, these approaches were applied as well as Logistic Trees with Unbiased Selection (LOTUS). The latter is a new, simultaneous method, in which both segmentation and scorecards are optimised at the same time. A single-scorecard logistic regression model was used as a reference. All these methods were applied to the data provided by two of the major UK banks and one of the European credit bureaus. Once the models were developed, the obtained results were analysed to examine whether there is improvement in the model performance due to the segmentation methods used. Moreover, the segmentation contribution was assessed. The paper is structured as follows. In the next section, the theoretical background of segmentation is presented as well as segmentation methods and other researchers’ findings on its impact on the model performance. In the third section, the basics of logistic regression, CART, CHAID and LOTUS are introduced. In the fourth section, the datasets are described. The fifth section is on the research results. The sixth section is a discussion and the last section includes the research findings and conclusions. 2. Segmentation 2.1. Theoretical background In credit scoring, segmentation can be defined as “the process of identifying homogeneous populations with respect to their predictive relationships” (Makuch, 2001, p. 140). The identified populations are treated separately in the process of a scoring model development, because of possible unique relationships between customer’s characteristics and a dependent variable. Nowadays segmentation is widely used in banking. There are various segmentation drivers, i.e. factors that can drive the division of a scoring model into two or more scorecards. Anderson (2007) classifies them into: marketing, customer, data, process and model fit factors. The first four factors reflect, respectively, the special treatment of some market segments, or customer groups, data issues (such as data availability) and business process requirements (e.g. different definitions of a dependent variable). 4 The model fit relates to interactions within the data and using segmentation to improve the model performance. In this research, the focus is on segmentation which is driven by the model fit factors. As far as segmentation is concerned, there are two key concepts: a segmentation basis and a segmentation method. A segmentation basis is a set of variables that allow for the assignment of potential customers to homogeneous groups. Segmentation bases can be classified as either general or product-specific, and either observable or unobservable (Wedel and Kamakura, 2000). As far as scorecard segmentation is concerned in this research, there is an unobservable product-specific basis. Once the segmentation is implemented, customers are grouped on the basis of their unobservable behavioural intentions to repay their loans or the relationship between their intentions and characteristics. On the date of grouping, it is not known whether the customers are going to repay or not. According to Wedel and Kamakura (2000), there are six criteria for effective segmentation. It seems that three of them are especially important in credit scoring: identifiability (customers can be easily assigned to segments), stability and responsiveness (segments differ from each other in their response/behaviour). Unobservable product-specific bases, which contain behavioural intentions, are characterised by good identifiability, moderate stability and very good responsiveness (Wedel and Kamakura, 2000). The above-mentioned advantages make these bases promising as far as scorecard segmentation is concerned. Segmentation methods can be classified as either associative (descriptive) or regressive (predictive) approaches (Aurifeille, 2000; Wedel and Kamakura, 2000). Since the ultimate goal is to assess the credit risk, the latter are applied in this research. There are two types of regressive approaches: two-step (a-priori) and simultaneous (post-hoc) methods (Aurifeille, 2000; Wedel and Kamakura, 2000). In the two-step approaches, segmentation is followed by the development of a regression model in each segment. In the simultaneous methods, both segmentation and regression models are optimised at the same time. 5 The two-step approaches are not designed to yield optimal results in terms of the prediction accuracy but rather to aid the understanding of overall strategy. On the other hand, the simultaneous methods give priority to a low, tactical level rather than to a high, strategic level of decision: the optimisation objective is to obtain the most accurate prediction, and not necessarily a meaningful and easily understandable segmentation (Desmet, 2001). 2.2. Segmentation methods There is not much literature on segmentation methods in credit scoring. According to Siddiqi (2005), segmentation methods can be classified as either experience-based (heuristic) or statistical. As far as the experience-based methods are concerned, one approach is to define segments that are homogeneous with respect to some customers’ characteristics. This allows for the development of segment-specific variables. For example, creating a segment of customers, who have a credit card, enables construction of such characteristics as credit limit used. Another approach is to define segments that are homogeneous with respect to the length of customers’ credit history (cohorts) or data availability (thin/thick credit files). For instance, creating a segment of established customers allows building behavioural variables based on the data from the last 12 months, the last 24 months etc. Furthermore, if there is a group (e.g. mortgage loan owners or consumer finance borrowers) that is expected to behave differently from other customers, or for whom the previous scoring model turned out to be inefficient, it is worth creating a separate segment for such a group. Moreover, customers can be grouped into segments in order to make it easier for a bank to treat them in different ways, e.g. by setting different cut-offs, i.e. score thresholds used in the decision making (Thomas, 2009). Finally, segmentation can be based on variables (e.g. age) that are believed to have strong interactions with other characteristics (Thomas, 2009). This is a heuristic approach but it has been developed into statistical methods based on interactions. An alternative to segmentation based on a selected variable is to include all its interactions with the other variables in a single-scorecard model (Banasik et al., 1996). 6 However, such a model has a large number of parameters and is less understandable than a multi-scorecard one. The experience-based segmentation methods can help achieve various goals such as improving the model performance for a certain group of customers or supporting the decision making process. The experience-based segmentation may also allow for better risk assessment for the entire population of customers. However, there is no guarantee that segmentation, which intuitively seems reasonable, will increase the model performance (Makuch, 2001). As far as statistical methods are concerned, segmentation is obtained using statistical tools as well as data mining and machine learning techniques. One approach is to do the cluster analysis (Siddiqi, 2005). The cluster analysis can be conducted using hierarchical clustering, the k-means algorithm or Self-Organizing Maps (SOMs). Regardless of the algorithm applied, clustering is based on customers’ characteristics. Therefore, customers with different demographic or behavioural profiles are classified into different segments. The resulting groups are homogeneous with respect to the characteristics but, since the customers’ status is not used in segmentation, they do not need to differ in risk profiles. Another approach is to use tree-structured classification methods such as CART or CHAID (VantageScore, 2006). In this approach, grouping is based on the customers’ status, and thus segments differ in risk profiles. Both the cluster analysis and classification trees can constitute the first step in the two-step regressive approaches. However, the classification trees often yield sub-optimal results (VantageScore, 2006). In 2006 VantageScore introduced a new, multi-level segmentation approach: combining experience-based segmentation (at higher levels) and segmentation based on a dedicated score (at lower levels). This score must be calculated using an additional scoring model which has to be built first. The split points on the score are determined using CART. Using the score enables dividing customers in such a way that in each segment, customers are similar to one another as far as their risk profile is concerned. There is an assumption that different risk profiles are associated with different relationships between a dependent variable and customer’s characteristics. 7 The VantageScore approach makes it easier for a bank to treat subprime and prime customers in different ways, but it seems that this approach does not have to be always optimal in terms of the model performance. There were also some attempts to develop methods that would allow for the optimal segmentation, i.e. a segmentation that would maximise the model performance. Their results can be classified as the simultaneous methods. Hand at el. (2005) suggested a method for the optimal division into two segments. In both segments, the same set of variables is used to develop a scorecard. The optimal division into the two groups is found using exhaustive search (each possible split point is examined on each variable or the linear combination of variables). For each possible pair of segments, two logistic regression models are built. The fit of the two-scorecard model is assessed using its overall likelihood, i.e. a product of likelihoods of the scorecards, and that division is chosen which gives the highest overall likelihood. However, the adopted assumptions (only two segments, the same variables) result in limited usefulness of the suggested method. In banking practice, customers are usually divided into at least a few segments, in which different sets of variables are used. Another approach to the optimal segmentation is Fair Isaac’s Adaptive Random Trees (ART) technology (Ralph, 2006). In this approach, the trees are not built level by level as in most tree-structured classification methods. In the beginning, the trees are randomly created using some predefined split points on the possible splitting variables. Then a genetic algorithm is applied to find the best tree, i.e. the tree that gives the highest divergence in the system of scorecards in its leaves, where the scorecards are naïve Bayes models. In all of them, there is the same set of characteristics as in the parent scorecard which is built on the entire sample. The ART technology has fewer drawbacks than other methods. It should allow for the maximisation of the model performance (measured using divergence). The number of segments is not predetermined. The use of the genetic algorithms avoids the exhaustive search that is both expensive and time-consuming. However, there is still a serious disadvantage, since – as in Hand at el. (2005) – the same set of variables is used in all scorecards. 8 2.3. Impact of segmentation It is commonly asserted by scorecard developers that a suite of scorecards allows for better risk assessment than a single scorecard used for all customers. According to Makuch (2001), segmentation usually increases performance by 5 to 10 percent in comparison with a single-scorecard system. It is also believed that segmentation can significantly contribute to performance of a scoring model. Impact of segmentation on the model performance measures can be assessed using simulated results of random scorecards applied to the identified segments (Thomas, 2009). The segmentation contribution to the model performance can also be assessed using difference between a performance measure of the model and the weighted average among the scorecards. This average is calculated using weights equal to percentages of customers classified to the segments. Banasik et al. (1996) analysed impact of some experience-based divisions on discrimination of a model. They set a few cut-offs and measured the discrimination in terms of errors that occur on a holdout sample. As a result, they found that “it is not the case that creating scorecards on separate subpopulations is necessarily going to give better discrimination than keeping to one scorecard on the full population”. For a suite of scorecards, it is difficult to choose cut-offs that are independent, good and robust at the same time. However, if cut-offs are chosen in the same way for all models, multi-scorecard models reject less applicants than single-scorecard ones. This may also be considered an advantage of segmentation. 3. Models 3.1. Logistic regression Logistic regression is the most commonly used method for developing scoring models. Since there is a binary dependent variable (either good or bad), binomial logistic regression is applied. In binomial logistic regression, a dependent variable y is equal to the cumulative distribution function F of a logistic distribution: 9 βx βx    e Fy 1 1 )( , where x is a vector of independent variables (covariates) and  is a vector of model parameters (Greene, 2000, p. 815). The parameters are usually estimated using the maximum likelihood (ML) method. The estimated value of a dependent variable lies between 0 and 1. Thus, it can be interpreted as probability of a dependent variable being equal to one. In credit scoring, this is probability of a customer being bad (probability of default). In scorecards, covariates are often used in the form of Weights of Evidence (WoE). If a discrete or discretised variable X takes K values, then the Weight of Evidence for its nth value (n ≤ K) is computed using the following formula (Anderson, 2007, p. 192):                            K k kn K k knn BBGG BnP GnP WoE 11 ln )|( )|( ln , where Gn (Bn) is a number of goods (bads) for whom X takes the nth value. A ratio of goods to bads is referred to as the odds in credit scoring. The population odds is a ratio of the proportion of goods pG to the proportion of bads pB in the population. It is often assumed that there is a linear relationship between the score and the log odds (Mays, 2004). Using the Bayes’ rule, it can be shown that the log odds sn among customers, for whom X takes the nth value, are equal to a sum of the population log odds spop and the Weight of Evidence for the nth value of X (Thomas, 2009, p. 33): npop B G B G n WoEs BnP GnP p p pBnP pGnP nBP nGP s                              )|( )|( lnln )|( )|( ln )|( )|( ln . Weights of Evidence allow for the assessment and comparison of the relative credit risk associated with different values of a variable (attributes of a characteristic). There is sometimes no theory that would support the choice of covariates. Therefore, the best set of covariates is often identified using the stepwise selection of variables (Hosmer and Lemeshow, 2000). The stepwise selection is a procedure of alternate inclusion and exclusion of variables from a model based on the statistical significance of their coefficients that is measured with a p-value. In logistic regression, the 10 likelihood ratio test or the Wald test are used to assess significance of the coefficients. In both cases there are the chi-square test statistics. In a forward selection step, the variable is included that, once added to the model, has the most significant coefficient. In a backward elimination step, the variable, which has the least significant coefficient, is excluded from the model. The stepwise selection is especially useful in case of a large number of possible covariates. Therefore, it is popular in behavioural scoring. The goodness-of-fit of a logistic regression model can be measured e.g. using the deviance. In logistic regression, the deviance plays the same role as the residual sum of squares in linear regression. It is calculated according to the following formula:                            n i i i i i i i y p y y p yD 1 1 ˆ1 ln)1( ˆ ln2 , where i y is the dependent variable value and i p̂ is the estimated probability of 1 i y for the ith observation, i = 1, …, n (Hosmer and Lemeshow, 2000, p. 13). In credit scoring, it is important how well the model fits the data but it is even more important how effectively it separates the goods and the bads. The separating ability is often referred to as the discriminatory power. There is a wide selection of discriminatory power measures (Thomas, 2009), with the Gini coefficient and the Kolmogorov-Smirnov (KS) statistic being the most commonly used ones. Both the Gini coefficient and the KS statistic can be calculated using the cumulative distribution functions (CDFs) of scores, computed separately for goods and bads (Thomas, 2009). The KS statistic is equal to the maximum difference between these CDFs. In order to calculate the Gini coefficient, the Receiver Operating Characteristic (ROC) curve is usually constructed. The ROC curve can be drawn by plotting the above-mentioned CDFs against each other. The Gini coefficient is equal to the double area under the ROC curve (AUROC) less one. Similarly to the KS statistic, it takes values between 0 and 1 with higher values meaning the stronger discriminatory power. Among other discriminatory power measures, there are the Somers D-concordance statistic and the Mann-Whitney U-statistic. The relationship between these statistics, the Gini coefficient and AUROC is as follows (Thomas, 2009, p. 113 and p. 120): 11 121AUROC2GINI  BG S nn U D , where nG and nB are numbers of good and bad customers, respectively. 3.2. CART Classification and Regression Trees (CART) are a popular nonparametric statistical method (Breiman et al, 1998). In this research, the focus is on classification trees, i.e. trees with a nominal dependent variable. In CART, predictors can be both continuous and categorical while splits are binary. All possible splits on all variables are examined and assessed. In order to measure quality of a split, the impurity function values are calculated for both child nodes. The impurity is often assumed to take the form of the entropy: )1log()1(log)( ppppNI  or the Gini index: )1(2)( ppNI  , where p is a fraction of observations with a positive response in the node N (Izenman, 2008, p. 288). Once all splits are assessed, such a split of the node N into N1 and N2 is selected that results in the largest decrease in impurity (Breiman et al, 1998, p. 32): )()()(),( 2 2 1 1 21 NI N N NI N N NINNIG  . The tree is grown using the recursive partitioning, i.e. each child node is split in the same way (Berk, 2008). The growing process continues until no more nodes can be split. In order to avoid excessively large structures and overfitting, the tree is then pruned back. The pruning process consists in minimising the cost-complexity measure that is defined as follows: TTRTR   )()( , where R(T) is an estimate of the misclassification cost of the tree T and  is the complexity parameter while |T| denotes the number of leaves (Breiman et al, 1998, p. 66). For each value of the complexity parameter, the smallest tree can be identified that minimises the cost-complexity measure. As a result, there is a sequence of nested subtrees. The best subtree is selected using a test sample or cross-validation. In this research, test samples were used. The trees were created in SAS Enterprise Miner and 12 served as the first step in the two-step approach. Splits were selected using the Gini index as the impurity function. The CART method is often compared to the C4.5 algorithm, another popular method for building classification trees (Hand et al, 2001; Larose, 2005). However, there are some important differences between them, e.g. the latter allows splitting into three or more child nodes (multi-way splits). Moreover, in the C4.5 algorithm, the split selection is always based on the information gain, i.e. reduction in entropy. 3.3. CHAID Chi-Square Automatic Interaction Detection (CHAID) is also a tree-structured classification method (Kass, 1980). It belongs to a family of methods known as Automatic Interaction Detection (AID). As its name suggests, the AID allows for the detection of interactions between variables. Thus, the segmentation is based on the interactions. The AID requires that predictors are categorical, i.e. either discrete or discretised (if originally continuous). The original categories of the predictors are grouped into a number of classes using a stepwise procedure that includes both merging and splitting steps. In a merging step, all categories or classes are compared to one another using some tests. The least significantly different ones are then grouped into a new class. In a splitting step, all possible binary divisions of a class are analysed and such a division is selected that leads to the most significantly different classes. Only classes, which consist of 3 or more categories, can be divided. The resulting grouping is then selected to split the node. There can be multi-way splits (Hawkins and Kass, 1982). In CHAID, the dependent variable has to be nominal, and the split selection is based on the chi-square tests of independence between the grouped predictors and the dependent variable. In order to account for multiple testing, the Bonferroni correction is used (Hawkins and Kass, 1982). The Bonferroni correction adjusts the test significance level for many tests that are performed at the same time. 13 Once a node is split, the grouping and testing process is repeated for each child node. Growing the tree continues until there are no more nodes that can be split. No pruning is carried out. However, in this research, CHAID was used as the first step in the two- step approach. Thus, manual pruning was performed to ensure that in each leaf, there are enough bads to build a logistic regression model. The trees were produced in SAS Enterprise Miner. Classification trees, including CART and CHAID, can be used not only for segmenting customers but also for developing scoring models (Thomas et al., 2002; Yobas et al., 2004). They can be applied instead of e.g. logistic regression. In such an application, each customer can be assigned probability of default equal to the bad rate in the leaf that he or she falls into. 3.4. LOTUS Chan and Loh (2004) noticed that there is selection bias in CART (but not CHAID) and in all other methods where exhaustive search is used for variable selection: if all possible splits based on all variables are considered, then variables with more unique values are more likely to be selected to split the node. The selection bias problem is overcome in the Logistic Tree with Unbiased Selection (LOTUS) algorithm (Chan and Loh, 2004; Loh, 2006). This algorithm allows for the development of classification trees with logistic regression models in their leaves. Since the trees are built together with the models, this is a simultaneous method. The algorithm starts with a regression model developed using the entire training sample (at the root). Once a node is split, new models are built in the child nodes. In order to avoid the bias, the split selection is divided into two separate steps: variable selection and split point selection (Chan and Loh, 2004). For all variables, which are analysed in the first step, the chi-square statistics are computed. The statistic used depends on whether the analysed variable serves as a regressor in the parent node, i.e. the node to be split. For non-regressors the ordinary chi-square statistic is calculated while for regressors the trend-adjusted chi-square statistic is computed. The latter tests whether there are any nonlinear effects after adjusting for a linear trend (Armitage, 1955). The variable with the lowest p-value is selected to split the node. In the second 14 step, the split point is selected that minimises the total deviance, i.e. the sum of deviances of regression models built in the child nodes. The algorithm stops when there are too few observations to split a node or to develop a model. The CART pruning method is then used to prune the tree. The cost- complexity measure is based on the total deviance (summed over all leaves). Finally, the subtree with the lowest total deviance is selected (Chan and Loh, 2004). The LOTUS algorithm is implemented in the LOTUS software (Chan, 2005). In this research, the LOTUS software was used with the following options: logistic regression models with stepwise selection were built in all nodes, and the pruning process was based on test samples. 4. Data In this research, three real-world datasets are used. The data describes individual customers. There are two datasets containing application data and one dataset with behavioural (credit bureau) data. The datasets are referred to as A1, A2 and B, respectively. In order to get unbiased results, each dataset was randomly divided into training, validation and test samples. In all these samples, the bad rate is the same as in the original dataset. The datasets A2 and B were divided into the samples that contain ca 50, 30 and 20 per cent of customers, respectively. The samples, which were created as a result of the dataset A1 division, include ca 50, 25 and 25 per cent of customers (there would be an insufficient number of bads in a smaller test sample). The training samples were used to develop models. The validation samples served as holdout ones, i.e. they were not used in the model development. Once a model was built, its stability was evaluated through the comparison of its discriminatory power on the training and validation samples. The smaller the difference, the more stable the model. The test samples were only used to prune the trees. 15 4.1. Dataset A1 The dataset A1 was provided by one of the major UK banks. There is data on 7,835 applicants, of whom 6,440 were goods and 1,395 were bads. Originally, there was also data on some rejected applications but they were then excluded from the dataset. The applications were made between April and September 1994. Customers applied for personal loans for different purposes. Loan amounts ranged from £500 to £50,000 while repayment periods varied from 6 months to 5 years. The characteristics are listed in Appendix A. They describe both a customer and a loan that he or she applied for. There are also some credit bureau variables in the dataset. 4.2. Dataset A2 The dataset A2 was provided by another major UK bank. There is data on 39,858 customers, including 38,135 goods and 1,723 bads. Originally, there were also some indeterminates who were then eliminated from the dataset. The loans were opened between May 1994 and August 1996. Loan amounts ranged from £300 to £15,000 while loan terms (durations) varied from 6 months to 10 years. In the original dataset, there were 111,946 customers. There was not only application but also credit bureau data (see Appendix A). However, the additional data was provided only for a part of the dataset. There are reasons to assume that the bank had such data for other customers, too. In order to account for this, the bad rate should be the same among customers with and without the credit bureau data (4.32%). All goods and bads, for whom there is the additional data, are included in the dataset. As far as customers without the credit bureau data are concerned, all bads are included as well as such a number of randomly sampled goods that the bad rate is equal to 4.32%. The resulting numbers of goods and bads are mentioned above. 16 4.3. Dataset B The dataset B was provided by one of the European credit bureaus. There is data on 186,574 customers, of whom 179,544 were goods and 7,030 were bads. In the original dataset, there was also data on some indeterminates but they were then excluded. Since the data was sampled from the credit bureau database, the customers had different credit products with different banks. There are 324 characteristics based on the customer’s credit history. However, they cannot be listed since this is proprietary information. Some examples include: worst payment status within the last 12 months, number of credit inquiries within the last 12 months, number of open accounts, number of past loans, total credit limit etc. The characteristics are as of the 1st of July 2008 (observation point) and the customer’s status is as of the 1st of July 2009 (outcome point). Thus, the outcome period length is exactly equal to twelve months. 5. Results In this research, suites of scorecards were developed based on the above-mentioned datasets. Both the two-step and simultaneous approaches were adopted. In the two- step approaches, segmentation was performed using CART and CHAID, and scorecards were built for the identified segments. In the simultaneous approach, the LOTUS algorithm was used to develop both segmentation and scorecards. For reference purposes, a single-scorecard model was estimated based on each dataset. All the scorecards were built using logistic regression with stepwise selection. No interaction variables were allowed in the scorecards. The variable grouping process was performed in the Interactive Grouping node in SAS Enterprise Miner. Categories of discrete variables were grouped into classes while continuous variables were discretised (binned) first. For each variable, such a division was selected that maximises reduction in entropy on the entire training sample. No more than five classes were allowed. The groupings were sometimes modified manually to put them in line with the banking experience. 17 In all the adopted approaches, only grouped variables and those original ones, which are categorical, were allowed to split the nodes. If necessary, the CART and CHAID trees were pruned back manually until there were at least a minimum number of bads in each leaf. This number was assumed to be equal to 100 for the datasets A1 and A2 and 500 for the dataset B. The same minimum numbers of bads were set as an option in the LOTUS algorithm. The CART, CHAID and LOTUS trees are presented in Appendix B. In each leaf, numbers represent: the number of bads and the bad rate in the leaf, as well as the number of all customers and their share in the training sample. In the CHAID tree for the dataset B, there is one leaf with only 16 bads (marked with an asterisk). It was not possible to prune the tree more because this leaf is a child node of the root. However, with such a number of bads, it was not possible to build a scorecard, either. Therefore, in this leaf all customers were assigned the same probability of default that is equal to the bad rate (0.3%). As a result, there is no separating ability and both the Gini coefficient and the KS statistic are equal to 0 in this leaf. For each dataset, there is at least one variable that was selected to split nodes in most trees based on this dataset. Time with Bank was used in all trees for the dataset A1. For the dataset A2, all nodes were split using either Loan Amount or Loan Purpose. For the dataset B, Var2 was used in both the CART and the LOTUS trees. The variables Var1, Var2 and Var3 are based on the payment statuses of customer’s loans (describing delinquencies etc.). In all the developed scorecards, characteristics were used in the form of WoE (based on the entire training sample). It was assumed that no scorecard could consist of more than 10 characteristics since in a credit scoring application, there are usually between 6 and 15 best variables (Anderson, 2007). In Appendix A, the characteristics, which were used in the reference logistic regression models based on the datasets A1 and A2, are marked with a bold font. In the reference scorecard based on the dataset B, there are, among other variables, Var1, number of credit inquiries within the last 9 months and age of the oldest loan. Some variables were used both in the reference models and in the trees: Time with Bank and Insurance (A1), Loan Amount and Loan Purpose (A2) as well as Var1 (B). 18 In each suite, the scorecards are consistent in terms of scale, i.e. there is the same relationship between scores and PD. This enables the calculation of discriminatory power measures for the entire model. The Gini coefficients and KS statistics are presented in Tables 1 and 2, respectively. There are values obtained on the training, test and validation samples. Only for the dataset A1, do the multi-scorecard models perform slightly better than the reference logistic regression model on a training sample: both the Gini coefficients and the KS statistics are higher by 2-3 percentage points. For the other datasets, the differences in the Gini coefficient do not exceed one percentage point, what makes them negligible. All the models for the dataset B are perfectly stable: the Gini coefficients and the KS statistics are very similar on the training and validation samples. The perfect stability is probably due to the size of a training sample and the power of credit bureau variables. The models for A2 are still stable while those for A1 cannot be considered stable: the Gini coefficients are lower by more than 10 percentage points on the validation sample as compared to the training sample. For both A1 and A2, logistic regression models are the most stable, probably due the smallest number of parameters and the simplest structure. The Gini coefficients and the KS statistics, which were obtained on the validation samples for the datasets A2 and B, are similar for single- and multi-scorecard models. However, on the validation sample for the dataset A1, the discriminatory power measures are higher by 3-5 percentage points for the logistic regression than for the CART- and CHAID-based models. For each approach, the segmentation contribution to the model performance was assessed using difference between the Gini coefficient or the KS statistic of the model and the weighted average among the scorecards on the training sample (see Tables 3 and 4). For comparison purposes, the discriminatory power measures were also calculated for the CART, CHAID and LOTUS trees. In order to compute these measures, it was assumed that each customer was assigned a probability of default equal to the bad rate in his or her segment. The results are presented in Tables 3 and 4. There are the Gini coefficients and the KS statistics of the entire models (“Model”) 19 and scorecard averages calculated using weights equal to percentages of customers classified to the segments (“Scorecards”). There are also differences between the former and the latter (“Difference”) as well as the discriminatory power measures of the trees (“Tree”). For the dataset A1, the trees are much weaker than the scorecards, the segmentation contribution does not exceed 9 percentage points and the scorecards are comparable to the logistic regression. As a result, the multi-scorecard models slightly outperform the single-scorecard one. For the datasets A2 and B, both the Gini coefficients and the KS statistics of the trees are high, often higher than those of the scorecards. The segmentation contribution is up to even 20 percentage points. However, the scorecards, which were built for the identified segments, are much weaker than the logistic regression models developed on the entire training samples. Therefore, there is no difference in performance between the single- and multi-scorecard models. 6. Discussion It can be surprising that segmentation does not improve the model performance, especially on the credit bureau dataset. As far as the credit bureau is concerned, the population is heterogeneous because there are customers of different banks, using different products etc. It could be expected that segmentation would bring an improvement in risk assessment for this population. It is worth seeing, in what situations segmentation can improve the model performance and the simultaneous approach can perform better than the two-step approaches. In order to show an example of such a situation, an artificial dataset was constructed. It is assumed that there is a random variable X and two simple logistic regression models based on this variable. In the first model, the parameter coefficient is equal to  while in the second model it is equal to –. It means that the relationship between X and a binary dependent variable Y is positive in the former and negative in the latter model. Values of Y are randomly generated using these two models. As a result, there are two groups of customers: G1 and G2. Their sizes do not have to be equal but should not differ much. In G1, the bad rate is higher than in G2. Subsequently, G1 is 20 split into G11 and G12 so that G12 is similar to G2 in terms of the bad rate. Ultimately, there are three groups of customers: G11 (the first model, high bad rate), G12 (the first model, low bad rate) and G2 (the second model, low bad rate). In order to distinguish them from one another, a new variable Z is created. For different groups, Z takes random values from different, non-overlapping intervals, e.g. (a, b) for G11, (b, c) for G12 and (c, d) for G2. It is determined for each customer separately. The artificial dataset contains three variables (X, Y and Z). There are training, validation and test samples having at least a few thousand customers each. The two-step approaches based on CART and CHAID as well as the LOTUS algorithm and logistic regression were applied to an artificial dataset that was constructed in the above-described way. The results (Gini coefficients and KS statistics) are presented in Tables 5 and 6. The single-scorecard model performance is relatively poor since both X and Z are weak variables on the entire sample. CART and CHAID produced the same segmentation: the sample is split on Z equal to b so that G11 is in one node and G12 and G2 are in another node. The high-bad-rate group was separated from the low-bad-rate ones (this is how the classification trees work). However, it was difficult to build a good scorecard for the node, which contains both G12 and G2, since the data were generated using the completely different models. As a result, the entire model performs only slightly better than the single-scorecard one. The LOTUS algorithm split the sample on Z equal to c so that G11 and G12 are in one node and G2 is in another node. The groups, whose data were generated using the different models, were separated from each other. This allowed for the development of good scorecards in both nodes. Therefore, the simultaneous approach outperforms the two-step approaches on the artificial dataset. This is an example of a situation in which segmentation improves the model performance and the simultaneous approach outperforms the two-step approaches. However, it seems rather unusual in banking practice that the same characteristic affects the score positively in one group and negatively in another. Provided that there 21 is such a characteristic in a real-world application, will it make a difference in a ten- or-more-characteristic scorecard? 7. Conclusions For none of the analysed real-world datasets, the multi-scorecard models perform considerably better than the logistic regression. Thus, the first and most important finding is that segmentation does not always improve model performance in credit scoring. The performance improvement is not necessary to occur even if it is going to be the only goal of segmentation, as in this research. This is in line with findings of Banasik et al. (1996) which were confirmed here also for the statistical methods of segmentation. Secondly, there is no difference in performance between the two-step and simultaneous approaches. Classification trees (CART and CHAID) followed by logistic regression in their leaves yield similar results to the LOTUS algorithm, in which both segmentation and scorecards are optimised at the same time. The LOTUS algorithm had seemed promising as a method for the optimal segmentation. However, it outperforms neither the two-step approaches nor the logistic regression. Thirdly, for a large sample including strong characteristics, all the models have the same separating ability and are equally stable. In this case, the two-step and simultaneous approaches as well as the logistic regression perform very similarly. For smaller samples and/or weaker characteristics, the logistic regression models are the most stable since they have less parameters and a simpler structure than the multi- scorecard models. Fourthly, segmentation contribution can be up to 20 percentage points. The discriminatory power measures of the trees, which are used for segmentation, can be even higher than those of the scorecards developed in their leaves. This means that segmentation itself can be a very powerful tool. However, it seems that such strong segmentation does not leave much space for the scorecards to further discriminate customers. Thus, the scorecards on average are weaker than the single-scorecard model. 22 Fifthly, it is possible to show an example of a situation in which segmentation improves the model performance and the simultaneous approach outperforms the two- step approaches on an artificial dataset. However, such a situation as in the example seems rather unusual in credit scoring practice. Building more than one scorecard requires more time and resources to be allocated to development, implementation, maintenance, monitoring and validation of the model. These additional costs should be compensated for by the improvement in performance, if it is the goal of segmentation. As this research shows, such improvement is not necessary to occur. If it does not occur, it makes sense to use a single-scorecard model. In banking practice it is common not to compare the developed multi-scorecard model with a single-scorecard one. Building the latter is usually considered a waste of time since there is a strong belief that segmentation allows for better risk assessment. However, maintaining several scorecards, which perform like a single one, seems to be a much greater waste of resources. In light of this research, it is strongly recommended to develop a single-scorecard model for comparison purposes. Although the model performance is very important, it is not the only criterion of the model choice. It is possible that a multi-scorecard model is similar to a single- scorecard one in terms of the performance but e.g. the ROC curve has a better shape for the former than for the latter. Then it makes sense to choose the multi-scorecard model since there are better cut-off levels. In this research, the focus is on segmentation, which is driven by the model fit factors, but it should not be forgotten that segmentation is sometimes driven by other factors. Then the model performance improvement is not the goal. Further analysis of segmentation in credit scoring could also include using other simultaneous approaches, e.g. Logistic Model Trees (Landwehr et al., 2005). 23 Training sample Test sample Validation sample Dataset A1 CART 0.527 0.374 0.359 CHAID 0.531 0.392 0.351 LOTUS 0.520 0.425 0.386 Logistic regression 0.499 0.404 0.397 Dataset A2 CART 0.663 0.623 0.618 CHAID 0.664 0.621 0.622 LOTUS 0.664 0.634 0.634 Logistic regression 0.657 0.640 0.635 Dataset B CART 0.807 0.813 0.808 CHAID 0.807 0.814 0.805 LOTUS 0.805 0.817 0.803 Logistic regression 0.801 0.818 0.807 Table 1. The Gini coefficient values for training, test and validation samples Training sample Test sample Validation sample Dataset A1 CART 0.389 0.296 0.267 CHAID 0.386 0.320 0.283 LOTUS 0.379 0.344 0.298 Logistic regression 0.362 0.317 0.316 Dataset A2 CART 0.516 0.479 0.477 CHAID 0.520 0.469 0.489 LOTUS 0.502 0.491 0.487 Logistic regression 0.497 0.505 0.485 Dataset B CART 0.705 0.704 0.701 CHAID 0.705 0.712 0.696 LOTUS 0.702 0.710 0.700 Logistic regression 0.692 0.708 0.698 Table 2. The KS statistic values for training, test and validation samples 24 Model (1) Scorecards (2) Difference (1) – (2) Tree Dataset A1 CART 0.527 0.442 0.086 0.328 CHAID 0.531 0.453 0.077 0.295 LOTUS 0.520 0.485 0.036 0.164 Dataset A2 CART 0.663 0.502 0.161 0.567 CHAID 0.664 0.499 0.165 0.563 LOTUS 0.664 0.554 0.110 0.397 Dataset B CART 0.807 0.671 0.136 0.634 CHAID 0.807 0.635 0.172 0.619 LOTUS 0.805 0.608 0.197 0.572 Table 3. The Gini coefficient values of models, scorecards and trees Model (1) Scorecards (2) Difference (1) – (2) Tree Dataset A1 CART 0.389 0.353 0.036 0.261 CHAID 0.386 0.355 0.031 0.234 LOTUS 0.379 0.370 0.009 0.164 Dataset A2 CART 0.516 0.395 0.121 0.443 CHAID 0.520 0.389 0.130 0.443 LOTUS 0.502 0.433 0.070 0.384 Dataset B CART 0.705 0.514 0.190 0.615 CHAID 0.705 0.496 0.209 0.595 LOTUS 0.702 0.546 0.156 0.517 Table 4. The KS statistic values of models, scorecards and trees 25 Training sample Test sample Validation sample Artificial Dataset CART/CHAID 0.528 0.519 0.517 LOTUS 0.636 0.635 0.633 Logistic regression 0.482 0.479 0.469 Table 5. The Gini coefficient values for training, test and validation samples (artificial dataset) Training sample Test sample Validation sample Artificial Dataset CART/CHAID 0.392 0.388 0.380 LOTUS 0.486 0.497 0.499 Logistic regression 0.335 0.344 0.330 Table 6. The KS statistic values for training, test and validation samples (artificial dataset) 26 References Anderson, R. (2007). The Credit Scoring Toolkit. New York: Oxford University Press. Armitage, P. (1955). Tests for Linear Trends in Proportions and Frequencies. Biometrics, 11(3), 375-386. Aurifeille, J.M. (2000). A bio-mimetic approach to marketing segmentation: Principles and comparative analysis. European Journal of Economic and Social Systems, 14(1), 93-108. Banasik, J.L., Crook, J.N. and Thomas, L.C. (1996). Does scoring a subpopulation make a difference. The International Review of Retail, Distribution and Consumer Research, 6(2), 180-195. Berk, R.A. (2008). Statistical Learning from a Regression Perspective. New York: Springer. Breiman, L., Friedman, J.H., Olshen, R.A. and Stone, C.J. (1998). Classification and regression trees. Boca Raton, Florida: Chapman & Hall/CRC. Chan, K.-Y. and Loh, W.-Y. (2004). LOTUS: An algorithm for building accurate and comprehensible logistic regression trees. Journal of Computational and Graphical Statistics, 13, 826-852. Chan, K.-Y. (2005). LOTUS User Manual (version 2.2). Available at: http://www.stat.wisc.edu/~kinyee/Lotus/manual.pdf. (Accessed: 15 January 2010). Desmet, P. (2001). Buying behavior study with basket analysis: pre-clustering with a Kohonen map. European Journal of Economic and Social Systems, 15(2), 17-30. Greene, W.H. (2000) Econometric Analysis. Upper Saddle River: Prentice Hall. Hand, D., Mannila, H. and Smyth, P. (2001). Principles of Data Mining. Cambridge, Massachusetts: MIT Press. Hand, D.J., Sohn, S.Y. and Kim, Y. (2005). Optimal bipartite scorecards. Expert Systems with Applications, 29(3), 684-690. Hawkins, D.M. and Kass, G.V. (1982). Automatic Interaction Detection. In: D.M. Hawkins (ed) (1982). Topics in Applied Multivariate Analysis. Cambridge: Cambridge University Press, 269-302. Hosmer, D.W. and Lemeshow, S. (2000). Applied Logistic Regression. New York: Wiley. http://www.stat.wisc.edu/~kinyee/Lotus/manual.pdf 27 Izenman, A.J. (2008). Modern Multivariate Statistical Techniques: Regression, Classification, and Manifold Learning. New York: Springer. Kass, G.V. (1980). An Exploratory Technique for Investigating Large Quantities of Categorical Data. Applied Statistics, 29(2), 119-127. Landwehr N., Hall M. and Frank E. (2005). Logistic Model Trees. Machine Learning, 59(1-2), 161-205. Larose, D.T. (2005). Discovering knowledge in data: an introduction to data mining. Hoboken, New Jersey: Wiley. Loh, W.-Y. (2006). Logistic Regression Tree Analysis. In: H. Pham (ed) (2006) Handbook of Engineering Statistics. London: Springer, 537-549. Makuch, W.M. (2001). The Basics of a Better Application Score. In: E. Mays (ed) (2001) Handbook of Credit Scoring. Chicago: Glenlake Publishing Company, 127-148. Mays, E. (2004). Credit Scoring for Risk Managers. The Handbook for Lenders. Mason, Ohio: Thomson South-Western. Ralph, C. (2006). Using Adaptive Random Trees (ART) for optimal scorecard segmentation. A Fair Isaac White Paper. Available at: http://www.computerworlduk.com/cmsdata/whitepapers/5126/UsingAdaptiveRan domTreesARTforoptimalscorecardsegmentationwp0406.pdf. (Accessed: 15 January 2010). Siddiqi, N. (2005). Credit Risk Scorecards: Developing and Implementing Intelligent Credit Scoring. New York: Wiley. Thomas, L.C., Edelman, D.B. and Crook, J.N. (2002). Credit Scoring and Its Applications. Philadelphia: SIAM. Thomas, L.C. (2009). Consumer Credit Models: Pricing, Profit, and Portfolios. New York: Oxford University Press. Van Gestel, T. and Baesens, B. (2009). Credit Risk Management. Basic concepts: financial risk components, rating analysis, models, economic and regulatory capital. New York: Oxford University Press. VantageScore (2006). Segmentation for Credit Based Delinquency Models White Paper. Available at: http://www.vantagescore.com/docs/ segmentation.pdf. (Accessed: 22 January 2010). Wedel, M. and Kamakura, W.A. (2000). Market Segmentation: Conceptual and Methodological Foundations. New York: Springer. http://www.computerworlduk.com/cmsdata/whitepapers/5126/UsingAdaptiveRandomTreesARTforoptimalscorecardsegmentationwp0406.pdf http://www.computerworlduk.com/cmsdata/whitepapers/5126/UsingAdaptiveRandomTreesARTforoptimalscorecardsegmentationwp0406.pdf http://www.vantagescore.com/docs/segmentation.pdf http://www.vantagescore.com/docs/segmentation.pdf 28 Yobas, M.B., Crook, J.N. and Ross, P. (2004). Credit scoring using neural and evolutionary techniques. In: Thomas, L.C., Edelman, D.B. and Crook, J.N. (eds) (2004) Readings in Credit Scoring: Foundations, Developments, and Aims. New York: Oxford University Press, 277-293. 29 Appendix A. Customer’s characteristics Dataset A1 Dataset A2 Age Age a Marital Status Marital Status Residential Status Number of Children MOSAIC Classification Residential Status Time at Current Address Time at Current Address Time at Previous Address Home Phone Home Phone Time with Current Employer Occupation Gross Income Time with Current Employer FiNPiN Classification Time with Previous Employer Loan Type Net Income Loan Amount Pension Scheme Loan Purpose Time With Bank Insurance Number of Credit Cards Payment Frequency Amex / Diners Card Holder Number of Searches for Exact Name (Current Address) Loan Amount Time since Last CCJ for Exact Name (Current Address) Loan Term Number of Write-offs for Exact Name (Current Address) Loan Purpose Time since Last CCJ for Similar Name (Current Address) Total Cost of Goods Number of Write-offs for the Same Surname (Current Address) Insurance Number of Bad Events for the Same Surname (Current Address) Payment Frequency Number of Bad Events at the Postal Code (Current Address) Payment Method Number of Bad Events Which Have Turned Good at the Postal Code (Current Address) Number of Searches in the Last 6 Months Percentage of Bad Events Which Have Turned Good at the Postal Code (Current Address) Value of CAIS (Bad Debts, Same Surname, Other Initial, Current and Previous Address) Number of Dormant Events at the Postal Code (Current Address) Value of CAIS (Bad Debts, Same Surname, Same Initial, Current and Previous Address) Electoral Roll Status for the Same Surname (Current Address) Value of CCJ (Same Surname, Other Initial, Current and Previous Address) Time on Electoral Roll (Current Address) Value of CCJ (Same Surname, Same Initial, Current and Previous Address) Number of Searches for Exact Name (Previous Address) Time since Most Recent CAIS (Bad Debt, Same Surname, Other Initial, Current and Previous Address) Time since Last CCJ for Exact Name (Previous Address) 30 Dataset A1 Dataset A2 Time since Most Recent CAIS (Bad Debt, Same Surname, Same Initial, Current and Previous Address) Number of Write-offs for Exact Name (Previous Address) Time since Most Recent CCJ (Same Surname, Other Initial, Current and Previous Address) Time since Last CCJ for Similar Name (Previous Address) Time since Most Recent CCJ (Same Surname, Same Initial, Current and Previous Address) Number of Write-offs for the Same Surname (Previous Address) Number of CAIS (Bad Debts, Same Surname, Other Initial, Current and Previous Address) Number of Bad Events for the Same Surname (Previous Address) Number of CAIS (Bad Debts, Same Surname, Same Initial, Current and Previous Address) Number of Bad Events at the Postal Code (Previous Address) Number of CCJ (Same Surname, Other Initial, Current and Previous Address) Number of Bad Events Which Turned Good at the Postal Code (Previous Address) Number of CCJ (Same Surname, Same Initial, Current and Previous Address) Percentage of Bad Events Which Have Turned Good at the Postal Code (Previous Address) Number of Dormant Events at the Postal Code (Previous Address) Electoral Roll Status for the Same Surname (Previous Address) Time on Electoral Roll (Previous Address) a The characteristics, which were used in the reference logistic regression models, are marked with a bold font. Table 7. Customer’s characteristics 31 Appendix B. Tree structures Time with Bank Insurance Bads: 152 (9.4%) All: 1612 (41.0%) Time at Current Add. Bads: 146 (16.1%) All: 908 (23.1%) Bads: 185 (23.5%) All: 786 (20.0%) Bads: 216 (34.5%) All: 626 (15.9%) < 3 years >= 3 years Yes No < 5.2 years >= 5.2 years Figure 1. The CART tree for the dataset A1 Time with Bank Bads: 216 (34.5%) All: 626 (15.9%) Bads: 168 (21.6%) All: 777 (19.8%) Bads: 193 (14.7%) All: 1314 (33.4%) Bads: 122 (10.0%) All: 1215 (30.9%) < 3 years 3-6 years 6-12 years >= 12 years Figure 2. The CHAID tree for the dataset A1 Time with Bank Bads: 577 (21.2%) All: 2717 (69.1%) Bads: 122 (10.0%) All: 1215 (30.9%) < 12 years >= 12 years Figure 3. The LOTUS tree for the dataset A1 32 Loan Amount Loan Amount Bads: 104 (1.1%) All: 9904 (49.7%) Loan Purpose Bads: 202 (3.6%) All: 5625 (28.2%) Bads: 110 (7.8%) All: 1402 (7.0%) Loan Purpose Loan Amount Bads: 145 (9.2%) All: 1577 (7.9%) Bads: 116 (16.2%) All: 714 (3.6%) Bads: 185 (26.1%) All: 708 (3.6%) < £ 3400 >= £ 3400 < £ 5200 >= £ 5200 < £ 1400 >= £ 1400 RF RF other other Figure 4. The CART tree for the dataset A2 Loan Amount Bads: 104 (1.1%) All: 9904 (49.7%) Bads: 154 (3.4%) All: 4512 (22.6%) Bads: 158 (6.3%) All: 2515 (12.6%) Loan Purpose Bads: 120 (8.9%) All: 1350 (6.8%) Bads: 118 (18.1%) All: 652 (3.3%) Bads: 208 (20.9%) All: 997 (5.0%) < £ 1400 £ 1400 - £ 2500 £ 2500 - £ 3400 £ 3400 - £ 5200 >= £ 5200 other RF or HI Figure 5. The CHAID tree for the dataset A2 Loan Amount Bads: 416 (2.5%) All: 16931 (85.0%) Loan Purpose Bads: 149 (9.9%) All: 1502 (7.5%) Bads: 297 (19.8%) All: 1497 (7.5%) < £ 3400 >= £ 3400 other RF, MP or HI Figure 6. The LOTUS tree for the dataset A2 33 Var1 Var2 Bads: 1085 (1.3%) All: 84040 (90.1%) Bads: 500 (11.7%) All: 4279 (4.6%) Bads: 1929(38.8%) All: 4967 (5.3%) Figure 7. The CART tree for the dataset B Var3 Bads*: 16 (0.3%) All: 6156 (6.6%) Bads: 1090 (1.4%) All: 76595 (82.1%) Bads: 525 (9.4%) All: 5614 (6.0%) Bads: 1883(38.3%) All: 4921 (5.3%) Figure 8. The CHAID tree for the dataset B Number of Loans Var2 Bads: 969 (1.3%) All: 72181 (77.4%) Bads: 1423(28.0%) All: 5077 (5.4%) Bads: 1122 (7.0%) All: 16028 (17.2%) < 3 >= 3 Figure 9. The LOTUS tree for the dataset B 
work_lykvmffhsvao5iyxubj4bhlgne	----	HQE: A hybrid method for query expansion Expert Systems with Applications 36 (2009) 7985–7991 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a HQE: A hybrid method for query expansion Lixin Han a,b,c,*, Guihai Chen b a College of Computer and Information Engineering, Hohai University, Nanjing, China b State Key Laboratory of Novel Software Technology, Nanjing University, Nanjing, China c Department of Mathematics, Nanjing University, Nanjing, China a r t i c l e i n f o Keywords: Information retrieval Collaborative filtering Data mining Optimization 0957-4174/$ - see front matter � 2008 Elsevier Ltd. A doi:10.1016/j.eswa.2008.10.060 * Corresponding author. Address: College of Com neering, Hohai University, Nanjing, Jiangsu 210098, C E-mail address: lixinhan2002@hotmail.com (L. Ha a b s t r a c t Query expansion methods have been extensively studied in information retrieval. This paper proposes a query expansion method. The HQE method employs a combination of ontology-based collaborative filter- ing and neural networks to improve query expansion. In the HQE method, ontology-based collaborative filtering is used to analyze semantic relationships in order to find the similar users, and the radial basis function (RBF) networks are used to acquire the most relevant web documents and their corresponding terms from these similar users’ queries. The method can improve the precision and only requires users to provide less query information at the beginning than traditional collaborative filtering methods. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction Query expansion methods have a long history in traditional information retrieval. Its study may date back to, at least, the stud- ies of Sparck Jones and Needham (1968), Sparck Jones (1971) in which the collection is analyzed to provide a similarity thesaurus of word relationships. A number of query expansion methods have been proposed. Generally the existing query expansion methods can be classified into the two classes of global analysis and local analysis (Xu & Croft, 2000). (1) Global analysis is one of the first techniques to produce con- sistent and effective improvements through query expansion. Global analysis analyzes the entire document corpus to discover word relationships. A known global technique is Latent Seman- tic Indexing (LSI) (Xu & Croft, 2000). Latent Semantic Indexing (LSI) (Deerwester, Dumais, Landauer, Furnas, & Harshman, 1990), which is a way to represent data, has explored the use of semantics for information retrieval and reduces the retrieval and ranking problem to one of significantly lower dimensions. Other global techniques are term clustering, similarity thesauri and Phrasefinder (Xu & Croft, 2000). Generally, global analysis needs to compute term correlation only once at system devel- opment time. However, term ambiguity may introduce irrele- vant correlated terms. (2) In contrast to global analysis, local feedback methods (Sal- ton & Buckley, 1990) are generally more effective. However, ll rights reserved. puter and Information Engi- hina. n). these local feedback methods are not relatively robust. Local analysis uses only top-ranked retrieved documents for further query expansion. Local analysis needs to compute term correla- tion for every query at run time. The above query expansion methods require the systems to analyze a large amount of information that the users provide. In addition, it is too difficult to ensure the completeness and correct- ness of the existing documents. Collaborative filtering goes some way to addressing these issues. Collaborative filtering (CF) meth- ods (Resnik, Iacovou, Suchak, Bergstrom, & Riedl, 1994) work by assessing similarities among users, then recommending given users more useful documents that like-minded users accessed pre- viously. Therefore users’ query expansion information can be ac- quired from other similar users. However, if the users provide little information at the beginning, collaborative filtering cannot discover any correlations between users and their similar users and further collaborative filtering cannot find their similar users. If external ontological sources including users’ personal informa- tion can be employed, some similar users may be found. Initial knowledge about all users and their interests can be provided from these external ontologies. Accordingly, based on the idea above, we propose a novel query expansion method called HQE (hybrid query expansion). The HQE method employs collaborative filtering com- bined with the Semantic Web and neural networks to improve query expansion. The HQE method derives ontology-based user similarity and then finds the similar users, and further constructs the training data of relevant documents retrieved by similar users, at last predict document relevancy and discover the most relevant web documents and their corresponding terms. The HQE method is more convenient for users to expand new query than the above traditional query expansion methods. Be- mailto:lixinhan2002@hotmail.com http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 7986 L. Han, G. Chen / Expert Systems with Applications 36 (2009) 7985–7991 cause of using ontology-based collaborative filtering, the HQE method only requires users to enter some keywords and these users do not require provide a large amount of information that is acquired by analyzing these retrieved documents at the beginning. 2. Related work Various methods of query expansion have been explored in pre- vious work. Xu and Croft (2000) propose a local context analysis method, which combines both local analysis and global analysis. Expansion terms employed are not based on frequencies in the top-ranked documents, similarly to local feedback, but on co- occurrences with the query terms. Therefore, the local context analysis method can overcome the difficulty of local analysis to some extent. The method is based on the hypothesis that a fre- quent term from the top-ranked relevant documents will tend to co-occurrence with all query terms within the top-ranked docu- ments. However, the hypothesis is not always true. Cui, Wen, Nie, and Ma (2002) propose a new method for query expansion based on query logs. In the method, probabilistic correlations be- tween query terms and document terms are extracted by analyzing query logs. These correlations are then used to select high-quality expansion terms for new queries. Kamps (2004) explores a new feedback technique that re-ranks the set of initially retrieved doc- uments based on the controlled vocabulary terms assigned to the documents. The approach uses the combination of global and local feedback techniques. On the one hand, the approach uses a global feedback technique to analyze the usage of controlled vocabulary in the collections. The rationale for this is that the approach does not want to rely on the availability of special dictionaries or the- sauri. The approach is similar to latent semantic indexing. On the other hand, the approach uses a local feedback technique for re-ranking the set of initially retrieved documents. Han and Chen (2006) use ontology information to extend the keywords and then employ an algorithm for mining association rules to find the terms that are closely related to these keywords and their synonyms. These keywords and their synonyms are restricted by these rele- vant words to reduce ambiguity and improve precision. Pôssas, Ziviani, Wagner, and Ribeiro-Neto (2002) propose a model referred to as set-based model for computing index term weights and for ranking documents. The components in set-based model are no longer terms, but termsets. The model computes term weights using association rules technique. El-Kahlout and Oflazer (2004) present the design and implementation of a system called Meaning to Word (MTW) to find the appropriate Turkish words that closely match the definition entered by the user. The approach for extract- ing words based on meanings checks the similarity between the user’s definition and each entry of the Turkish database that does not consider the semantics or the context. Various methods of collaborative filtering have been proposed in previous work. Sarwar, Karypis, Konstan, and Riedl (2001) pro- pose an effective item-based approach to dealing with both the scalability and sparsity problems in the context of collaborative fil- tering. In this approach, the user-item representation matrix is analyzed to identify the relations between items, rather than rela- tions between users, and then to use these relations to indirectly compute recommendations for users. Because the relations be- tween items are relatively static, item-based algorithms may have less online computation than the user-based algorithms. Huang, Chen, and Zeng (2004) present an effective collaborative filtering approach to exploring relational user and item similarities for deal- ing with the sparsity problem. They apply associative retrieval techniques and three related spreading activation algorithms including a constrained spreading activation algorithm based on the Leaky Capacitor Model (LCM), a branch and bound serial sym- bolic search algorithm (BNB), and a Hopfield net parallel relaxation search algorithm (Hopfield) to explicitly generate associations among consumers and products in collaborative filtering. Rashid, Karypis, and Riedl (2005) present a measure of influence that is algorithm-independent, in order to ensure that it can be applied to any ratings-based recommender system. Sarwar, Karypis, Kon- stan, and Reidl (2000) present and experimentally evaluate an ap- proach to applying the combination of the classical association rule algorithms, nearest-neighbor collaborative filtering algorithms, and algorithms based on dimensionality reduction to CF-based rec- ommender systems. Hofmann (2004) presents a model-based ap- proach to collaborative filtering. The approach introduces a statistical latent class model to discover user communities and build user profiles. In addition, the approach employs overlapping user communities to decompose user preferences. Lemire (2005) presents a scale and translation invariant as being a desirable prop- erty for collaborative filtering. The paper takes into consideration such criteria as the amplitude and the mean, and the number of ratings to normalize users, in order to improve accuracy. Here, we only have space to contrast our work with the most di- rectly related to Middleton, Shadbolt, and De Roure (2004); Hust (2004); Robin and Ramalho (2003); Ciaramita, Hofmann, and John- son (2003) Akrivas, Wallace, Andreou, Stamou, and Kollias (2002). Middleton et al. (2004) present a general ontology-based recom- mendation system approach. Their two experimental systems, Quickstep and Foxtrot, are built. Quickstep improves user profiles’ accuracy via inference and Quickstep is integrated with an external ontology built from a publication database and personnel database in order to bootstrap a recommender system for easing the cold- start problem. Foxtrot enhances Quickstep by visualizing user pro- files in order to acquire direct user profiles’ feedback. In contrast to Middleton et al. (2004), the HQE method applies the combination of using an ontology and collaborative filtering to query expansion instead of user profiling within recommender systems. In addition, the HQE method presents some algorithms to improve query expansion. Hust (2004) introduces collaborative filtering tech- niques to query expansion in a restricted collaborative information retrieval environment. The information retrieval system acquires the relevant documents from previous search processes carried out by one or several users to improve retrieval performance for the current query. The method does not acquire users’ knowledge from ontologies. In contrast to Hust (2004), HQE combines Seman- tic Web with collaborative filtering to perform the analysis of semantic relationships for finding the similar users, and acquires some relevant web documents from these similar users’ queries using the Radial Basis Function (RBF) networks. Various uses of ontology as a linguistic knowledge resource to improve IR effec- tiveness have been tried in previous work. Robin and Ramalho (2003) argue search engines access to wide-coverage linguistic ontologies such as WordNet, can improve web search effectiveness in retrieving a set of purely textual documents with no semantic annotation. They present an experiment that empirically measured the impact on retrieval effectiveness of automatically expanding the query keywords with their synonyms or immediately neigh- boring concepts. In the experiment, the individual impact of such expansion under environment typical web searches is evaluated. Ciaramita et al. (2003) introduce a hierarchical learning architec- ture based on the multiclass perceptron for lexical semantic classi- fication problems that integrates task-specific and general information. They propose to generate additional training data by extracting training sentences from a dictionary ontology-WordNet. Akrivas et al. (2002) propose a query expansion method, which takes into account the query context based on the Inclusion rela- tion. The query context is defined as a fuzzy set of semantic enti- ties. They integrate the method with the user’s profile. The user’s L. Han, G. Chen / Expert Systems with Applications 36 (2009) 7985–7991 7987 preferences are used to change this context to provide the capabil- ity for query personalization. In contrast to Robin and Ramalho (2003), Ciaramita et al. (2003), and Akrivas et al. (2002), the HQE method uses ontology information to find the similar users, ac- quires the relevant web documents from these similar users’ que- ries, and the relevant web documents are used by the RBF networks in order to expand new query. In contrast to the above work, the HQE method employs ontol- ogy-based collaborative filtering combined with neural networks to improve query expansion. In addition, the HQE method presents some algorithms to improve query expansion. In the HQE method, collaborative filtering is used to analyze semantic relationships that are acquired from the constructed ontologies in order to find the similar users, and the radial basis function (RBF) networks are used to acquire the most relevant web documents and their corresponding terms from these similar users’ queries. 3. The HQE method of improving query expansion In this section, we propose a new query expansion method called the HQE method. The HQE method consists mainly of the CCIO (combines complete-link clustering with inference for con- structing of ontologies) algorithm, the BBFU (branch and bound for finding similar users) algorithm and the RENQ (RBF to expand new query) algorithm. The CCIO algorithm is described in Section 3.1, The BBFU algorithm is described in Section 3.2, and the RENQ algorithm is described in Section 3.3. The HQE Method can be described as the following steps: {the CCIO algorithm is used to construct ontologies; based on the above ontologies, the BBFU algorithm is used to perform the analysis of semantic relationships for finding the similar users; some relevant web documents are acquired from the above similar users’ queries; the RENQ algorithm is used to rank these web documents and discover the most relevant web documents and their corre- sponding terms; } 3.1. The CCIO algorithm for constructing ontologies An ontology is an explicit specification of a conceptualization (Gruber, 1993). Form a human standpoint, an ontology is a concep- tualisation of a domain; Form a machine standpoint, it consists of entities, attributes, relationships, and axioms (Guarino & Giaretta, 1995). The ontologies are constructed to identify the association rela- tionships between concepts that can be extracted by users’ per- sonal information, and to be shared among all users. In addition, the existing relationships in the knowledge base provide a scope for discovering new relationships. For example, if such entities as teaching, institution, staff, project, student, and paper are con- structed, we can expect to find all users’ university, department, research interests, research projects, publications, course, graduate student etc. Accordingly, based on the idea above, we propose the CCIO algorithm for automatically constructing ontologies. The CCIO algorithm can exploit the desirable properties of both com- plete-link clustering algorithms and inference mechanism to con- struct the hierarchically structured ontologies, that is, based on the existing association entities acquired from the Complete-link Clustering algorithm, new meaningful relationships that cannot di- rectly observed can be inferred through Jena2’s inference mechanism. Formula (1) denotes the similarity between two entities repre- sented as sets A and B. SðA; BÞ¼ jA \ Bj=jA [ Bj ð1Þ where ax is an interest of user x, by is an interest of user y, jj is the cardinality of a set, and A = {ua1, . . . , uan} and B = {ub1, . . . , ubn}, where ua1, . . . , uan are the terms of a x, and ub1, . . ., ubn are the terms of b y. The larger S(A, B) is, the larger the similarity between A and B is. The CCIO algorithm can be described as follows: { i = 0; using formula (1), calculate the proximity matrix Di contain- ing the distance between ax and by; treat each interest as a cluster, that is Dpq = dpq; // Dpq is the distance between each pair of clusters, dpq is the distance between each pair of terms while all terms are not in one cluster { i = i + 1; find the most similar pair of Dpq in the proximity matrix and merge these two clusters Cp and Cq into one cluster Cr, that is Cr = {Cp, Cq}; //Cr, Cp, Cq is cluster using formula (1) and the formula Drk = max{Dpk, Dqk}, calcu- late the proximity matrix Di to perform a merge operation; a class hierarchy can be constructed; generate some rules from the class hierarchy; store these rules into the knowledge base; the widely used Jena2’s inference mechanism (Wilkinson, Sayers, Kuno, & Reynolds, 2003 & Carroll et al., 2003) is used to infer semantic associations from the existing rules within the knowledge base in order to discover more meaningful association entities; a class hierarchy can be reconstructed; } } In contrast to the widely used complete-link clustering algo- rithm (Jain, Murty, & Flynn, 1999), the CCIO algorithm provides the inference mechanism to find more useful association entities that are not directly observed in the users’ behaviour. 3.2. The BBFU algorithm for finding the similar users Branch and bound algorithms are methods for global optimiza- tion in non-convex problems (Lawler & Wood, 1966 & Moore, 1991). They are non-heuristic, in the sense that they maintain a provable upper and lower bound on the globally optimal objective value; they terminate with a guarantee proving that the indeed suboptimal point found is suboptimal. In the worst case they re- quire effort that grows exponentially with problem size, but in some cases we are lucky, and the methods converge with much less effort. The methods are useful to solve small instances of hard problems. We propose an efficient branch and bound search algorithm BBFU that introduces a branch and bound method to find the sim- ilar users. The BBFU algorithm employs optimization techniques to evaluate the nodes close enough to the specified node. Specifically, the BBFU algorithm examines the connectivity of entities in ontol- ogies, traverses the semantic relationships between entities before a given threshold is reached, according to their semantic distance, identifies a set of close entities, and finds the similar users. In the following description, k stands for the iteration index. Listk denotes the list of nodes. Uk denotes the upper bound, at the end of k iterations. x* denotes the optimal solution. The bounding function F(x) = max(P). The Pearson correlation P ¼ Pn i¼1 ðxi��XÞðyi��YÞ ðn�1ÞSX SY , where X and Y are two entities vectors, �X and �Y are their means, respectively, SX and SY are their standard deviations, respectively. 7988 L. Han, G. Chen / Expert Systems with Applications 36 (2009) 7985–7991 The Pearson correlation is often used to the similar entities whose features are known. The BBFU algorithm can be described as the following steps: { given ontologies and a user; x* = Ø; k = 0; List0 = {root node}; U0 = +1; do {maximizing choose the class node x with bounding func- tion F(x), where x 2 Listk; Listk+1 = Listk � {x}; if f(x) P Uk then prune the node; else {x* = x; Uk = f(x); split x into x1 and x2; Listk+1 = Listk [ {x1, x2}; } k = k + 1; } while Listk – Ø and Uk � f(x) > e determine optimal solution x*, that is find the most similar user; a set of the similar users are identified from the most similar user’ subclass and superclass in ontologies; } 3.3. The RENQ algorithm for expanding new query The RENQ algorithm introduces RBF networks to rank web doc- uments and discover the most relevant web documents and their corresponding terms. In the RBF networks, the relevant web pages are obtained from a set of the similar users. The algorithm is based on the hypothesis that two web pages are more similar in content if there are more common keywords in their web pages. The algo- rithm begins by training known relevant and irrelevant web docu- ments. Relevant web documents are obtained from a set of the similar users and irrelevant web documents are obtained from a set of the dissimilar users. During training, the connection weights of the RBF networks are initialized with some random values. The training samples in the training set are then presented to the RBF networks in random order and the connection weights are ad- justed using the SC method. This process is repeated until the least-squares error between the predicted outputs from the net- work and the observed output is small or the predefined maximum iterative number is achieved. After network training has finished, the neural network is built using keywords from these web pages. Each keyword maps into a 0 or 1, so the length of the input vector equals the number of the keywords chosen. The neural network can assign all the web pages to a relevancy value that is a fraction between 1 and 0. These relevant values may be used to rank these relevant web pages. The RENQ algorithm can be described as follows: {the training set is obtained from a set of the similar users and the dissimilar users; cstep = 0; // cstep is the current iterative step; While (the least-squares error > tnumber) and step < maxstep // tnumber is a predefined tolerance number, maxstep is a iter- ative maximum number; {using the following the SC method, the training samples are trained; step = step + 1; } using the trained neural networks, rank a set of web docu- ments acquired from these existing similar users’ queries, and discover the most relevant web documents and their cor- responding terms; } In contrast to the widely used RBF Networks (Yang & Zheng, 2003 & Ibnkahla, 2000), a SC method has been proposed to update RBF networks. 3.3.1. The SC method for updating RBF networks The SC (Combine SACU with CTSB) method has been proposed to update RBF networks. The method is composed of the SACU algorithm for centers update and the CTSB algorithm for the output weights update. For clarity of presentation, the basic concepts about RBF net- works are briefly recalled that are relevant to this paper. The RBF network (Ibnkahla, 2000) is a two-layer feedforward network, whose input-layer activation function can be described as follows: f(x) = /(kxk), where / is a radial basis function. k � k is the Euclidean norm on Rn. In contrast to MLNN (multi-layer neural network), the activation function is a compact one, instead of the usual sigmoid function. The activation function can be described as follows: /ðTÞ¼ 1ffiffiffiffi 2p p d expð�TÞ, T ¼ ðkx�ckkÞ 2 2d2 . The outputs of the first layer neurons are described as follows: x1k ¼ 1ffiffiffiffi2pp d exp � kx�ckk 2 2d2 � � , where x is the input vector and c is the weight associated to neuron k. The network outputs are described as follows: yj ¼ PN1 k¼1 wkj 1ffiffiffiffi 2p p d exp �kx�ckk 2 2d2 � � , where wkj is the weights associ- ated with the output layer. 3.3.1.1. The SACU algorithm for centers update. Simulated annealing is a generalization of a Monte Carlo method for examining the equations of state and frozen states of n-body systems (Metropolis, Rosenbluth, Rosenbluth, Teller, & Teller, 1953). The concept is based on the manner in which liquids freeze or metals recrystalize in the process of annealing. By analogy the generalization of this Monte Carlo approach to combinatorial problems is straight for- ward (Cerny, 1985; Kirpatrick, Gellat, & Vecchi, 1983). The major difficulty in implementation of the algorithm is that there is no obvious analogy for the temperature T with respect to a free parameter in the combinatorial problem. Furthermore, avoidance of entrainment in local minima is dependent on the ‘‘annealing schedule”, the choice of initial temperature, how much iterations are performed at each temperature, and how much the tempera- ture is decremented at each step as cooling proceeds (Gray et al., 1997). We propose the SACU (Simulated Annealing to Centers Update) algorithm to update centers. The SACU algorithm introduces simu- lated annealing for centers update. The simulated annealing algo- rithm can be viewed as globally optimizing an objective function. The SACU algorithm can ensure to break away current local opti- mal solution and to reduce the computing workload. The SACU algorithm is based on the following solution form: 1. The solution space is a training set; 2. The objective function is f ¼ min Pk r¼1 Sr ¼ min Pk r¼1 Pnr i¼1 ðxri � �xrÞ 0ðxri � �xrÞ, where k is the number of clusters, nr is the number of training samples in cluster r, Sr is the squared error, xri is training sample i in cluster r, and x�r is the centroid of clus- ter r; 3. A new solution is produced by randomly choosing another training sample from neighboring region N(x); L. Han, G. Chen / Expert Systems with Applications 36 (2009) 7985–7991 7989 4. The objective function difference is Df = f (new training sam- ple)-f (local optimal solution); 5. The accept criterion is P = 1 for Df > 0 or P = exp(�Df/t) for Df 6 0. The SACU algorithm can be described below: {given a set of the training samples, initial solution x0, initial optimal objective function value f*, initial temperature t0, initial step q0, monotone increment function I(x); x* = x0; // initialize optimal solution x* q = q0; // initialize step q t = t0; // initialize temperature t, t0 = 100 i = 1; // initialization at = 1; // initialization while jRi � Cj > e // e is a given enough small positive number, where e = 2 � 10�3. C is an threshold, where C = 0.75 { Ri�1 = at/as; // at is the acceptable times of new solution, as is iterative steps number, Ri�1 is acceptable rate at = at + 1; as = 0; // initialization while J < >U//U is the upper limit of step length { as = as + 1; while x R N(x)//N(x) is neighboring region {step length is q and the BFGS optimizer in the quasi- Newton method is used to compute local optimal solution; } xi is produced by randomly choosing another training sam- ple from neighboring region N(x) and objective function dif- ference Df = f(xi) � f(x) is computed; if Df 6 0 or exp(�Df/t) > random(0, 1) then {x = xi; if f < f* then {x = x*; f = f*; } if at is equal to the given acceptable times then break; } q = I(Ri)�q; // adjustment function I(x) is monofonic increasing function, where I(x) = (x-0.5)2 + 1 } if f < f* then {x = x*; f = f*; } Ri = at/as; if i > 2 then // compute self-adjusting temperature {if Ri�1 < C and Ri < C then t = t+TC;// TC is a given constant if Ri�1 P C and Ri P C then t = t-TC; if Ri�1 6 C and Ri P C thent = t-TC/2; if Ri�1 > C and Ri < C then t = t + TC/2; } i = i + 1; } } In contrast to the widely used simulated annealing algorithm (Jonhson, Aragon, Mcgeoch, & Schevon, 1991; Kirpatrick et al., 1983), there are three features in the SACU algorithm. Firstly, the SACU algorithm can self-adaptingly adjust temperature. Therefore it is regarded as time after time optional process. On the one hand, the Boltzmann factor exp(�Df/t) increases with t. This is useful to break away from the current local optimal solution and to increase step length in order to look for a better new solution. On the other hand, when acceptable rate meets the given condition, the pro- gram can end so as to reduce the computing workload. Secondly, local optimization techniques instead of choosing randomly a new solution are employed in order to compute a local optimal solution. Thirdly, the local optimal solution is recorded during annealing process, in order to protect the current optimal solution from being thrown away. 3.3.1.2. The CTSB algorithm for updating the output layer weights. We propose the CTSB (combine TS with BFGS) algorithm to update the output layer weights. The CTSB algorithm is a hybrid algorithm of exploiting the desirable properties of both the widely used globally optimal algorithms Tabu Search (Glover, 1986; Gray et al., 1997) and locally optimal algorithms BFGS (Xie, Han, & Lin, 1997). The CTSB algorithm can be described below: {xj is normalized, that is, x0j ¼ ajxj þ bj, where aj ¼ 2 xjmax�xjmin , bj ¼ 1 � ajxjmax, and xj 2 (xjmin, xjmax); count = 1; while count < maxno // maxno is the iterative maximum number {DWijk(n � 1) = Wijk(n) � Wijk(n � 1), n = 0, 1, . . .;// Wijk(n) is a weight vector DW ijkðnÞ¼�l oEðW ijkðnÞÞ oW ijkðnÞ þ mDW ijkðn � 1Þ; n ¼ 0; 1; . . . ; // l is the learning rate, m is a law of inertia coefficient if Pk2�1 n¼0 ðEðW ijkðn þ 1ÞÞ� EðW ijkðnÞÞÞ > kEðW ijkðn1ÞÞ// move a local optimal solution to a better local optimal solution, where k is a given appropriate small positive number between 0.15 and 0.25, k2 is a current value, and Wijk(n1) is a initialized weight vector, the squared error energy function E(n) is the error power between the network output vector at time n and the desired output then the BFGS algorithm is used to determine a local optimal solution and the weights are updated; if DEðW ijkðnÞÞj j EðW ijkðnÞÞ < e and EðW ijkðnÞÞ > L // indulge in a local optimal solution, where L is a given appropriate large number, and e is a given enough small positive number then the Tabu Search algorithm is used to break away current local optimal solution and the weights is updated; count = count + 1; } } In contrast to the supervised learning rules BP (Ibnkahla, 2000; Yang & Zheng, 2003) for the output weights update in RBF net- works, the SACU algorithm exploit the desirable properties of both Tabu Search and BFGS. Therefore the SACU algorithm is easy to find global optimal solution and to bring a less the squared error en- ergy. In contrast to the BP algorithm (Ibnkahla, 2000), DWijk(n) in the SACU algorithm consider not only learning rate but also a law of inertia coefficient in order to have faster convergence at iter- ative steps and avoid fluctuating fiercely around convergence point. 4. Experimental results and discussion In this section, we present the experiments results of the HQE method. The HQE method employs OWL (Dean & Schreiber, 0 0.2 0.4 0.6 0.8 1 M ea n a v er ag e p re ci si o n All Tops HQE method Hust's method keywords method Fig. 2. MAP f or all tops. 7990 L. Han, G. Chen / Expert Systems with Applications 36 (2009) 7985–7991 2004; McGuinness & Harmelen, 2004) to describe resources and their inter-relations. Jena2’s inference mechanism is used to infer semantic associations from the existing rules. We select 251 Web pages to acquire ontology information. These selected Web pages are related to personal information. Some entities are extracted from these Web pages and stored in the OWL files. Jena2’s infer- ence mechanism allows additional facts to be inferred from these entities. Inference mechanism uses forward chaining, backward chaining and a hybrid execution model provided by the Jena2 rea- soner. The connectivity of entities in ontologies is examined, and the semantic relationships between entities are traversed before a given threshold is reached, and according to their semantic dis- tance, a set of close entities are identified, and eighteen similar users are found. Some relevant web documents are acquired from these similar users’ research interests. After training these relevant web documents, a set of term for query expansion are discovered. We submit the user’s queries that are related to our research to the existing third party search engine Google. The search engine re- sults are the ten top-ranked retrieved documents. These queries belong to such topics as information retrieval, data mining, pattern recognition, semantic web, optimization and network computing. The query results are divided into relevant and irrelevant docu- ments. The ranked results are analyzed to check whether the rele- vant results are contained in the answer, in order to meet the need of a quantitative measure. We adopt mean average precision (MAP) to evaluate the HQE retrieval performance. Mean average precision, which is a standard IR evaluation measure, is used to find the mean of the average precisions over a set of queries, where average precision is the mean of the precision scores after each rel- evant document is retrieved. Hust’s method is the most relevant to our work. We compare the results of the HQE method with the re- sults of Hust’s method (Hust, 2004) and entering a set of keywords without using the HQE method for six topics. Six topics are closely related to these similar users. Six topics are described below: topic 1 is information retrieval, topic 2 is data mining, topic 3 is pattern recognition, topic 4 is semantic web, topic 5 is optimization and to- pic 6 is network computing. Mean average precision of these meth- ods for six topics are shown, respectively, in Fig. 1. The experiment result in Fig. 1 shows that the HQE method can perform better in topic 1, topic 2, topic 4, and topic 5. The experiment result shows that if a set of term for query is scientific terms without ambiguity, these methods led to similar higher mean average precision, and if a set of term for query suffer the usual problem of synonymy and ambiguity, that HQE can perform better. Hust’s method does not take into consideration acquiring relevant knowledge from ontolo- gies. Hust’s method is a method of introducing collaborative filter- ing techniques to query expansion. The method can help to improve query expansion. However, if similar users provide too many terms for a query and this causes too many expansion terms, the method can not improve query expansion. In contrast to Hust’s 0 0.2 0.4 0.6 0.8 1 M ea n av er ag e pr ec is io n top1 top2 top3 top4 top5 top6 Every top HQE method Hust's method keywords method Fig. 1. MAP f or every top. method, the HQE method can acquire the most relevant terms from these similar users’ queries and reduce the number of expansion terms. For example, we look for ranking algorithm in information retrieval. When keyword ranking is entered, such useless informa- tion as best college rankings and best graduate schools rankings could be found from query results. Thus, its average precision is only 0.297. After Hust’s method is used, its average precision is 0.757. After the HQE method is used, its average precision is 0.953. It is because ranking algorithm can be acquired from ontol- ogy information. Some similar users in ontology user do research in information retrieval area. Thus, ranking may be regarded as a kind of ranking algorithm in ontology information retrieval. Mean average precision of these methods for all tops are shown, respec- tively, in Fig. 2. The experiment result in Fig. 2 shows that HQE can perform better. 5. Conclusion Nowadays, the amount of information in the Internet is increas- ing dramatically. Facilitating users to get useful information has become more and more important to information retrieval sys- tems. The traditional query expansion methods require users to analyze a large amount of information. In addition, it is too difficult to ensure the completeness and correctness of the existing docu- ments. To solve these problems, in this paper, we propose a query expansion method called HQE. In the HQE method, the CCIO algo- rithm is proposed to construct ontologies. The BBFU algorithm is proposed to analyze semantic relationships for finding the similar users. The RENQ algorithm is proposed to acquire the most rele- vant web documents and their corresponding terms from these similar users’ queries. The method can improve the precision and only requires users to provide little query information at the begin- ning than traditional collaborative filtering methods. Acknowledgements The authors thank Professor Linping Sun for technical help. This work is supported by the National Natural Science Foundation of China under Grants 60673186, 60571048 and 60736015, the na- tional 863 program under Grants 2007AA01Z178, and the State Key Laboratory Foundation of Novel Software Technology at Nan- jing University under Grant A200604. The work is partly supported by China NSF grants (60573131, 60673154, 60721002, 60825205), Jiangsu High-Tech Research Project of China (BG2007039), and China 973 project (2006CB303000). References Akrivas, G., Wallace, M., Andreou, G., Stamou, G., & Kollias, S. (2002). Context- sensitive semantic query expansion. In 2002 IEEE international conference on artificial intelligence systems (ICAIS’02). Carroll, J., Dickinson, I., Dollin, C., Reynolds, D., Seaborne, A., & Wilkinson, K. (2003). The Jena semantic Web platform: Architecture and design. HP Laboratories Technical Report HPL-2003-146. L. Han, G. Chen / Expert Systems with Applications 36 (2009) 7985–7991 7991 Cerny, V. (1985). Thermodynamical approach to the traveling salesman problem: An efficient simulation algorithm. Journal of Optimization Theory and Applications, 45(1), 41–51. Ciaramita, M., Hofmann, T., Johnson, M., 2003. Hierarchical semantic classification: Word sense disambiguation with world knowledge. In IJCAI 2003 (pp. 817–822). Cui, H., Wen, J.-R., Nie, J.-Y., & Ma, W.-Y.(2002). Probabilistic query expansion using query logs www.2002. (pp. 325–332). Dean, M., & Schreiber, G. (2004). OWL Web ontology language reference. W3C recommendation 10 February 2004. <http://www.w3.org/TR/2004/REC-owl-ref- 20040210/>. Deerwester, S. C., Dumais, S. T., Landauer, T. K., Furnas, G. W., & Harshman, R. A. (1990). Indexing by latent semantic analysis. JASIS, 41(6), 391–407. El-Kahlout, I. D. & Oflazer, K. (2004). Use of Wordnet for retrieving words from their meanings. In Proceedings of the global Wordnet conference (GWC2004), (pp. 118– 123). Glover, F. (1986). Future paths for integer programming and links to artificial intelligence. Computers and Operations Research, 13(5), 533–549. Gray, P., Hart, W., Painton, L., Phillips, C., Trahan, M., & Wagner, J. (1997). A survey of global optimization methods. <http://www.cs.sandia.gov/opt/survey/ main.html>. Gruber, T. R. (1993). A translation approach to portable ontology specifications. Knowledge Acquisition, 5(2), 199–220. Guarino, N., & Giaretta, P. (1995). Ontologies and knowledge bases: Towards a terminological clarification. In N. Mars (Ed.), Towards very large knowledge bases: Knowledge building and knowledge sharing (pp. 25–32). IOS Press. Han, L., & Chen, G. (2006). The HWS hybrid web search. Information and Software Technology, 48(8), 687–695. Hofmann, T. (2004). Latent semantic models for collaborative filtering. ACM Transactions on Information Systems (TOIS), 22(1), 89–115. Huang, Z., Chen, H., & Zeng, D. (2004). Applying associative retrieval techniques to alleviate the sparsity problem in collaborative filtering. ACM Transactions on Information Systems, 22(1), 116–142. Hust, A. (2004). Introducing query expansion methods for collaborative information retrieval. Reading and Learning, 252–280. Ibnkahla, M. (2000). Applications of neural networks to digital communications – A survey. Signal Processing, 80(2000), 1185–1215. Jain, A. K., Murty, M. N., & Flynn, P. J. (1999). Data clustering: A review. ACM Computing Surveys, 31(3), 264–323. Jonhson, D. S., Aragon, C. A., Mcgeoch, L. A., & Schevon, C. (1991). Optimization by simulated annealing: an experimental evaluation – Part II (graph coloring and number partition). Operations research, 39(3), 378–406. Kamps, J. (2004). Improving retrieval effectiveness by reranking documents based on controlled vocabulary. ECIR 2004, 283–295. Kirpatrick, S., Gellat, C. D., & Vecchi, M. P. (1983). Optimization by simulated annealing. Science, 4598, 671–680. Lawler, E. L., & Wood, D. E. (1966). Branch-and-bound methods: A survey. Operations Research, 14, 699–719. Lemire, Daniel (2005). Scale and translation invariant collaborative filtering systems. Information Retrieval, 8(1), 129–150. McGuinness, D. L., & Harmelen, F. v. (2004). OWL Web ontology language overview. W3C recommendation 10 February 2004. <http://www.w3.org/TR/2004/REC- owl-features-20040210/>. Metropolis, N., Rosenbluth, A., Rosenbluth, M., Teller, A., & Teller, E. (1953). Equation of state calculations by fast computing machines. Journal of Chemical Physics, 21, 1087–1091. Middleton, S. E., Shadbolt, N. R., & De Roure, D. C. (2004). Ontological user profiling in recommender systems. ACM Transactions on Information Systems (TOIS), 22(1), 54–88. Moore, R. E. (1991). Global optimization to prescribed accuracy. Computers and Mathematics with Applications, 21(6/7), 25–39. Pôssas, B., Ziviani, N., Jr., Wagner, M., & Ribeiro-Neto, B. A. (2002). Set-based model: A new approach for information retrieval. SIGIR 2002, 230–237. Rashid, Al., Karypis, G., & Riedl, J. (2005). Influence in ratings-based recommender systems: An algorithm-independent approach. In Proceedings of SIAM international conference on data mining, 2005. Resnik, P., Iacovou, N., Suchak, M., Bergstrom, P. & Riedl, J. (1994). GroupLens: An open architecture for collaborative filtering of netnews. In Proceedings of ACM 1994 conference computer supported cooperative work (pp. 175–186). ACM Press. Robin, J., Ramalho, F., 2003. Can ontologies improve web search engine effectiveness before the advent of the semantic web? In SBBD 2003 (pp. 157–169). Salton, Gerard, & Buckley, Chris (1990). Improving retrieval performance by relevance feedback. JASIS, 41(4), 288–297. Sarwar, B., Karypis, G., Konstan, J., & Reidl, J. (2000). Analysis of recommendation algorithms for e-commerce. In Proceedings of the ACM conference on electronic commerce (pp.158–167). New York: ACM. Sarwar, B. M., Karypis, G., Konstan, J. A., & Riedl, J. T. (2001). Item-based collaborative filtering recommendation algorithms. In Proceedings of the tenth international World Wide Web conference (pp. 285–295). Sparck Jones, K. (1971). Automatic keyword classification for information retrieval. London: Butterworths. Sparck Jones, K., & Needham, R. M. (1968). Automatic term classification and retrieval. Information Processing & Management, 4(1), 91–100. Wilkinson, K., Sayers, C., Kuno, H. A., & Reynolds, D. (2003). Efficient RDF storage and retrieval in Jena2. In The first International Workshop on Semantic Web and Databases, Co-located with VLDB 2003 (SWDB’03), Humboldt-Universität, Berlin, Germany, September 7–8, 2003 (pp. 131–150). Xie, K., Han, L., & Lin, Y. (1997). Optimization method. Tientsin: University Press. Xu, J., & Croft, B. W. (2000). Improving the effectiveness of informational retrieval with local context analysis. ACM Transactions on Information Systems, 18(1), 79–112. Yang, X., & Zheng, J. (2003). Artificial neural network and blind signal processing. Tsinghua University Press. http://www.w3.org/TR/2004/REC-owl-ref-20040210/ http://www.w3.org/TR/2004/REC-owl-ref-20040210/ http://www.cs.sandia.gov/opt/survey/main.html http://www.cs.sandia.gov/opt/survey/main.html http://www.w3.org/TR/2004/REC-owl-features-20040210/ http://www.w3.org/TR/2004/REC-owl-features-20040210/ HQE: A Hybrid hybrid method for Query Expansionquery expansion Introduction Related work The HQE method of improving query expansion The CCIO algorithm for constructing ontologies The BBFU algorithm for finding the similar users The RENQ Algorithm algorithm for Expanding New Queryexpanding new query The SC Method method for updating RBF networks The SACU algorithm for centers update The CTSB algorithm for updating the output layer weights Experimental results and discussion Conclusion AcknowledgementAcknowledgements References 
work_lzposws5kvh67ch4d4vnaurek4	----	PII: S0957-4174(02)00149-5 What makes expert systems survive over 10 years—empirical evaluation of several engineering applications Jukka K. Nurminen*, Olli Karonen, Kimmo Hätönen Nokia Research Center, P.O. Box 407, FIN-00045 Nokia Group, Finland Abstract This case study analyzes eight expert system applications that have successfully been in industrial use for a long time. We have personally been involved in the development of these applications and are thus in a good position to analyze what is important for a successful application and what kind of mistakes can be made. Since the development of the applications started in 1986 – 1990 and some of them are still in use we are able to observe what has happened to those applications during their lifetime. Our key observations are related to the scope of the applications, to the trade-off between usability and automation, to the role of human experts in the use and development of expert systems, on the technical solutions used, on aspects of the operation of the expert system and on the similarities between expert systems and information systems. The key findings are expressed as 20 hypotheses for successful expert systems. The support of each application to the hypotheses is discussed. q 2002 Elsevier Science Ltd. All rights reserved. Keywords: Expert systems; Industrial applications; Lessons learned; Success factors; Implementation issues 1. Introduction In late 1980s there was a period of high activity on expert systems, with a lot of commercial interests and expectations. As very few of the overdimensioned expectations were met the interest declined (Harmon, 1995) but now the activity seems to be increasing again. The goal of this paper is to look back at some expert systems that were developed at that time in Nokia Research Center, discuss their lifecycle, and see where those systems are today. This experience allows us to extract several lessons learned that are important for the success and longevity of expert systems. In this paper we would like to emphasize the aspects of practical, industrial usage of the expert system technology. The applications are real-life, deployed applications that have been in production use within Nokia or its partners for years. We have personally been working on the develop- ment, deployment, and use of these applications. We feel that this is a good time for a post mortem analysis since now over a decade has passed since the applications were first deployed. It is thus possible to review what happened to those applications, what aspects were important for the future development and maintenance, and which applications or features were able to stand the test of times. Application descriptions in the literature commonly discuss either prototypes or solutions that have been in use for a short time. Such descriptions illustrate the possibilities of new technologies but they fail to cover issues that become visible after a longer period of serious use. Grandon Gill (1995) statistically studied the status of the early successful expert systems to see what had happened to those systems in a decade. Only about a third of the systems that were initially reported as successes were still in use. The reasons were found to be individual and organizatorial rather than technical or economical. The important issues highlighted in that study were: coordinating ES develop- ment with the IT and business strategies of the firm, understanding the task to be performed by the system, recognizing the legal implication of systems, identifying user concerns and expectations, and managing developers and development responsibilities. Guimaraes, Yoon, and Clevenson (1996) tested several hypotheses for expert system success using a sample of 130 expert systems at E.I. DuPont de Memours and Company, Inc. Most of the systems were very small and developed by one person. The study found support for seven of the nine hypothesized determinants of ES success. The important factors were: problem importance, developer character- istics, shell characteristics, and with less confidence: end- user characteristics, ES desirable impact on end-user jobs, and user involvement. The two factors that did not find 0957-4174/03/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 7 - 4 1 7 4 ( 0 2 ) 0 0 1 4 9 - 5 Expert Systems with Applications 24 (2003) 199–211 www.elsevier.com/locate/eswa * Corresponding author. Tel.: þ358-7180-36442; fax: þ358-7180-36855. E-mail address: jukka.k.nurminen@nokia.com (J.K. Nurminen). http://www.elsevier.com/locate/eswa support were problem difficulty and managerial support. Another very similar statistical study (Guimaraes, Yoon, & O’Neal, 1995) focuses on manufacturing expert systems withing IBM. Duchessi and O’Keefe (1995) analyzed six cases (three successes, three failures) within three companies focusing mainly on engineering applications. The success factors found in their study include management support, demon- strable business benefits, problem urgency, degree of organizational change, organizational support, and users’ personal stake in the system. Millett and Powell (1996) focused on the organizational aspects. They identified measures for the success of expert systems and used these measures to evaluate and benchmark the performance of one anonymous company. The critical success factors found in that study include: the type of application, importance of the system to business strategy, an early, key, successful development, existence of project champion, existence of an information technology strategy, and organizational culture. Main factors contributing to failure were found to be poor project control, insufficient resources for maintenance, and financial constraints. Stylianou, Madey, and Smith (1992) performed a field survey among expert system shell users to identify the most important features needed for expert system shells. The most critical characteristics found in that study were embeddability, rapid prototyping, backward chaining, explanation facility, ability to customize explanations, linkages to databases, and documentation compressiveness and readability. Other authors have focused not only on expert systems but AI research and applications in general. Hayes-Roth (1997) discussed from his subjective perspective different fields of artificial intelligence, tried to find the most successful areas, and presented practical issues for real- life use. Allen (1998) also uses a subjective perspective to discuss the opportunities of AI underlining the connection between AI research and real world needs. Doyle and Dean (1996) summarize the results of the Strategic Directions in Computing Research AI Working group. Among a variety of topics the paper highlights the need to integrate AI research and solutions to other research disciplines and computing systems, collaborative problem solving, and usability of the computers. The problem with all case-based research is that the results are specific to the applications and companies studied. However, the similarity of results from two companies, IBM (Guimaraes et al., 1995) and E.I. DuPont de Memours and Company, Inc. (Guimaraes et al., 1996), indicates that the conclusions from one organization can be applicable to others. A problem, especially with the statistical studies covering a large group of systems, (Grandon Gill, 1995; Guimaraes et al., 1996, 1995; Parpola, 1995), is that the tested hypotheses are quite abstract. Studies with smaller sample or narrower focus give more concrete, practical advice. Measuring and collecting detailed data of application development and lifecycle without disturbing it is proble- matic. One suggested way is to retrospectively look at the data available and consider what kind of hypotheses could be tested with the data (Fenton, 2001). This is the empirical approach we have followed here. This paper complements the previous studies by giving support to some earlier observations, refuting some other ones, and suggesting a set of new observations. The rest of this paper is structured in the following way. In Section 2 we provide a more detailed overview of the applications. Section 3 discusses our experiences and the lessons learned. Section 4 summarizes the most important observations of this study and suggests ideas for further research. 2. Applications Starting from 1986 Nokia Research Center had a group focused on knowledge technology, which was later merged with the Software Technology department. The mission of the group was to develop knowledge-based applications for Nokia internal use and for key customers. The emphasis was strongly on practical applications; very little basic AI research was done. Table 1 lists the most influential and widely used ones of these applications, some of which have been in use for close to 15 years. The table also shows the domain, the lifetime, the software and hardware environment (which may have changed during the lifetime), and the current status of each application. As can be seen from Table 1, five applications deal with planning types of activities, two with diagnostics, and one is a decision-making system. All of the applications except one (MATIAS) work on the engineering domain. 2.1. NICAD NICAD (Karonen, 1987) was a knowledge-based system for designing plants and other complex entities. Its first target application was the design of the soda recovery boilers for paper industry. A boiler plant contains thousands of components and hundreds of kilometers of pipes. The design of a new unique plant can take over 30 person years. In NICAD, the components of the plant (e.g. valves, pipes) were modeled as frames using an object-oriented approach. Each frame contained attributes that described the needed features of the corresponding component. Most of the attributes were geometric but there were also other types such as materials and standards. Each attribute could be associated with a calculation rule that tells how the value of the attribute depends on other attributes. The local dependencies between attributes defined global dependency graphs. These graphs were J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211200 two-directional and NICAD behaved like a 3D spreadsheet. On one hand, the dependencies were used in the calculation of the value of an attribute. On the other hand, if you changed a value, the resulting modifications were propa- gated in the opposite direction in the dependency graphs. The whole plant was described as a hierarchical product structure. E.g. a power plant consists of the boiler and the building, the boiler consists of the furnace and the economizer, etc. The product structure contained also calculation rules for the numbers of components (e.g. the number of support pillars), which changed dynamically during the design based on the dependencies. The calculation rules were expressed by means of a simple language, including Lisp functions and keywords like right, left, behind and above: object C is to the right of object A by 10 units and under object B by five units. NICAD was developed initially in the KEE expert system environment but was later ported to C and Sun. The current version of the tools, under the name Designþþ, runs on Windows NT platform. Interfaces to ODBC compliant databases and APIs for C, Cþþ and Visual Basic are supported. In 1989 a separate spin-off company, Design Power Inc., located in Silicon Valley, was founded to handle the commercialization and further development of the tool. 2.2. RFT RFT was a design support system for analog and especially radio frequency (RF) circuits. It was developed mainly for the needs of Nokia Mobile Phones. RFT is discussed in greater detail in Ketonen, Lounamaa, and Nurminen (1988). RFT was in a pilot usage but the complete system never reached a wide user space. One of the reasons was the Lisp implementation language, which tied the system to an expensive engineering workstation. Later the intelligent user interface of the system was reimplemen- ted in Cþþ for the Windows platform. The user interface is still used today together with the APLAC simulator. RFT was based on three main techniques. First, the graphical user interface allowed the direct manipulation of circuit diagrams. This was the most successful part of the system and is still in use today. Secondly, a rule-based layer was used to hide the details of the tools from the users. A number of separate tools were needed for RF design at that time. Since each tool had their own relatively unfriendly user interfaces the idea was to create a knowledge-based layer that would take care of transforming a user-specified problem to the appropriate Table 1 The main expert system applications developed at Nokia Research Center in 1986 – 1990 Name Domain Lifetime Platform Status Planning and configuration NICAD Intelligent CAD for engineering applications 1986 – today KEE, Symbolics ! C, Sun, Microsoft Windows Since 1989 continued within an independent company design power (http://www.dp.com) by the name Designþþ RFT Electronic CAD system for mobile phones 1987 – today Lisp, HP ! Cþþ, Microsoft Windows Intelligent GUI in use as part of the Aplac simulator package. Since 1998 continued within an independent company Aplac Solutions Corporation (http://www.aplac.com) Nokia planner Project planning 1987 – today(?) Lisp, Apollo ! C, Microsoft Windows Since 1992 continued within an independent company ViSolutions by the name Visual Planner. Merged with ICL 1996 CabCon Cable configuration 1989 – 1996 Xi Plus, Genesis, NICAD ! C In trial use in Nokia Cables until 1996 DXCON Telephone exchange configuration 1990 – 1997 SQLWindows, Microsoft Windows In production use in Nokia Networks until ,1997 Diagnostics DMG Diagnostics of radio links 1988 – 1992 Lisp, apollo, Sun, HyperExpert, PC In production use in Nokia Networks until ,1992 DXLib Documentation management 1990 – today Lisp, Sun and Cþþ, Microsoft Windows ! www In production use in Nokia Networks. DXLib model is still part of product shipment. Decision-making MATIAS Loan application analysis 1986 – today(?) Xi Plus/GURU, PC In production use at OKO bank Finland J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211 201 http://www.dp.com http://www.aplac.com form for each tool. In this way the user would be able to work with familiar schematic circuit diagrams. The symbolic manipulation at the knowledge-based layer would take care of the details, constraints, and complex syntaxes of the underlying tools. Although this concept worked well on the pilot cases it was not flexible enough. A further problem was that hiding too many details was confusing to the user. The third focus area was the design synthesis. The idea was to collect design rules from experienced engineers and thus allow the distribution of the knowledge to more inexperienced persons and to geographically different locations. This turned out to be extremely difficult. Creating a prototype system was relatively easy but updating and maintaining the design rules became a problem. Additionally, automated design synthesis was not flexible enough since, as observed also in Simon (1995), the goals and constraints are changing during the design process. The creativity of the user cannot be substituted by an automated solution. 2.3. Nokia planner Nokia Planner was a project management tool. The initial goal was to automate key project planning and project tracking tasks within Nokia. After the initial prototype was built in Lisp on Symbolics the project focus shifted from automation to visualization. With a graphical user interface running on early versions of Microsoft Windows Nokia Planner provided a graphical user-interface that allowed easy manipulation of activities and their interdependencies as well as resources. An automatic feature was available for scheduling the tasks to minimize the project duration. However, it turned out that the users mostly relied on manual manipulation of the activities since this allowed them more flexibility. Nokia Planner was 1992 moved to an independent company ViSolutions that continued the development and marketing of the tool under the name Visual Planner. ViSolutions was merged with ICL in 1996. 2.4. CabCon CabCon was a system for cable configuration (Karonen, 1989). The design of a cable is much more complex than the simple outer sheath lets us believe. The topology (what and how many components), materials, and geometry (also 3D) of the cable must be determined to meet the requirements of the client including transmission capacity and quality, purpose of usage, strength, maximum allowed weight, latest time of delivery, etc. In addition to the mechanical, thermal, and electronical properties of the cable itself, also the overall production situation has an influence on the priority of alternatives. The general goal is to find out a feasible solution involving minimum costs. The objective of the CabCon tool was to increase productivity and quality of cables design and manufacturing at Nokia. The subgoals were: prompt and high quality responses to customer inquires, lower costs in production, documentation of the design knowledge, standardization of delivered products, and reduced time-lag between design and delivery. The project started in 1988, and three prototypes were implemented using three different expert system shells, respectively. Exploiting the results of the complementary prototypes, a small production version was implemented in C in 1989. The system was in limited, mainly experimental, use until 1996. CabCon did not fully satisfy the users’ needs for two main reasons. First, the repertoire of cable structures needed is so large that the CabCon rules were not able to generate versatile enough solutions. Secondly, a graphical user interface was missing which made it difficult to update, customize, and manipulate the configurations proposed by the system. These two shortcomings amplified each other: automatically generated solutions were not satisfactory and their manual improvement and fine-tuning was too clumsy. 2.5. DXCON DXCON was a system for configuring Nokia DX 200 telephone exchanges. Configuration of a DX 200 switching system was a tedious task, and only a few experts were available. Equipping required both brute force and good command of the frequently evolving design rules. DXCON was meant to be an expert tool for design, sales, and production personnel to increase the productivity and the quality of equipping and price setting of DX 200 switching systems. Some of the equipping phases were totally automated, but for the more sophisticated design tasks interactive tools were considered as the most optimal compromise. A switching system consists of rack rows, racks, subracks, plug-in units, and cables. A central concept in DXCON was an equipping model, e.g. for each rack type there was a template describing what kind of subracks can be equipped and what are the feasible locations for those subracks. For each cable we had to define its type, source and destination, routing, and length. In full-scale production use the dominant way was to reuse and incrementally modify earlier complete switching system models created by the end-users themselves or by more senior experts. DXCON was running on a standard 386 personal computer with a relational database package called SQLWindows by Gupta Technologies. One of the main problems of DXCON was its speed. Even after several optimization rounds some operations required about half an hour to complete. Complex DB schema, queries with several joins, and big DB size were main reasons for the slow operation. During its lifetime starting from 1990 DXCON was in wide use and several hundreds of switching system J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211202 deliveries were configured with it. Its use was terminated when the importance of the fixed telephone exchange business decreased. 2.6. DMG Fault diagnosis of complex technical equipment is always a big challenge. One solution for the problem has been model-based expert systems, where a predefined decision system guides the analysis and is able to identify the fault. DMG (Tyrväinen, 1991) was used to create and edit decision trees that were used in the core of the diagnosis system. The DMG system had two modules: an engineering workstation, where the equipment model was created and where a decision tree was extracted from the model. The tree was then transferred to the end user environment, which was a standard PC with an expert system shell that was able to execute the decision tree. The system was used at radio link manufacturing. The link equipment was tested and fixed before they left the factory. 2.7. DXLib DXLib (Tyrväinen, Saarinen, & Hätönen, 1993) was a system for document management and information retrieval from documentation of DX 200 based network elements. Its main components were a text analyzer that produced SGML tagged formatted documents and a browser that linked documentation to the logical model of the element. The objective of the system was to assist human expert in finding relevant information. Like DMG the system consisted of two modules. An engineering workstation, where a logical model was created, was used to link the model and relevant texts together. The model and the documentation were then transferred to a PC environment, where they could be browsed. The system has been used ever since its first trial in 1993. The model is an optional feature of the electronic network element documentation package that accompanies network deliveries. 2.8. MATIAS MATIAS (Kontio, 1991) was an expert system for loan application analysis that was developed for OKO bank, which at that time was an important customer of Nokia Data. It contained knowledge about farming loans in Finland. The regulations for such loans including state subsidies were quite complex and MATIAS was meant to speed up the loan decision process. Containing a few thousand if-then rules divided into multiple modules MATIAS was a large expert system at its time. Most of the knowledge used by MATIAS existed already on documents about the loan policies and regu- lations. However, a large number of rules was needed to codify this knowledge. The large size of the rulebase resulted into the usual problems of verification and validation of the rulebase as well as difficulties in maintaining the rulebase. 3. Characteristics of successful systems We have collected our main observations in Table 2. The ‘Hypothesis’ column contains properties or statements that are characteristic of successful expert systems. These hypotheses are either derived from literature or are based on our own observations. The ‘Source’ column provides references to literature that has suggested a given hypothesis. It is difficult to track the ultimate origin of each observation or hypothesis in the literature. Therefore we have mentioned those studies where we have encountered a mentioning of a given hypothesis. Also different authors have formulated their findings in different ways. We have tried to understand what has been the spirit of the observation when listing which authors have discussed each hypothesis. The ‘Evidence of our case study’ column indicates general level of support this study gives for each hypothesis. The following columns contain more detailed analysis for each application. A ‘ þ ’ means that the application supports the hypothesis, a ‘ 2 ‘ means it refutes the hypothesis. An empty value means that the aspect was not relevant or observed for the application or that the conclusion is not clear. The observations are divided into four categories: domain, development, operation, and general. The hypoth- eses and the case study evidence for each of them are discussed in greater detail in the following sections. 3.1. Domain 3.1.1. Narrow scope Hayes-Roth (1997) suggests that narrowing the scope is important for expert system success. Naturally it is easier to develop a system with a narrow scope. However, the impact of such a system tends to be low because the solution is too limited. Fig. 1 shows the difference between narrow and wide scope applications. With a narrow scope it is possible to develop a system that is able to generate high quality solutions. These solutions can often do things better than average experts. However, it seems that the law of diminishing returns applies: the higher the solution quality is, the bigger marginal effort is needed to increase it. This means that pushing the solution quality above the acceptable level requires much more development than settling with a more modest level. The drawback of narrow scope is that it is possible to cover only a small part of the problem domain with such systems. Although a large enough number of narrow J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211 203 Table 2 Case study evidence for the hypotheses Hypotheses Source Evidence of our case study NICAD RFT CabCon DXCON DXLib DMG MATIAS Planner Domain Narrow scope Hayes-Roth (1997) Mixed 2 2 2 2 2 þ þ 2 Focused objective Hayes-Roth (1997) No 2 2 2 2 2 2 2 2 Stability of environment Hayes-Roth (1997) No 2 2 2 2 2 2 2 2 High degree of automation Hayes-Roth (1997) Mixed þ 2 2 2 þ þ 2 High degree of repetition Hayes-Roth (1997) Yes þ þ þ þ þ þ þ þ Small project Hayes-Roth (1997) Yes þ þ þ þ þ þ þ þ Development Users prefer usability over automation Nurminen, Karonen and Hätönen (2003) Yes þ þ 2 þ þ 2 þ Expert systems complement rather than replace human experts Millett and Powell (1996) Yes þ þ þ þ þ þ 2 þ Early benefits to experts themselves are important Nurminen, Karonen and Hätönen (2003) Yes þ þ þ þ þ þ 2 AI should be an embedded part of a bigger system Duchessi and O’Keefe (1995) and Millett and Powell (1996) Yes þ þ þ þ þ þ 2 þ Simple, straightforward solutions work best Nurminen, Karonen and Hätönen (2003) Yes þ þ þ þ þ þ þ If-then rules considered harmful Nurminen, Karonen and Hätönen (2003) Yes þ þ þ Lots of custom work Hayes-Roth (1997) Yes þ þ þ þ þ þ 2 þ Fast and agile development is important Duchessi and O’Keefe (1995) Yes þ þ þ þ þ þ þ Knowledge-based applications are based on domain specific application engines and generators Nurminen, Karonen and Hätönen (2003) Yes þ 2 þ þ þ Monolithic applications are sometimes better than knowledge-based applications Nurminen, Karonen and Hätönen (2003) Yes þ þ þ Rules of normal SW development apply Duchessi and O’Keefe (1995) and Millett and Powell (1996) Yes þ þ þ þ þ þ þ Operation Move towards mainstream software and hardware platforms Nurminen, Karonen and Hätönen (2003) Yes þ þ þ þ þ þ Successful systems tend to move out from the company Nurminen, Karonen and Hätönen (2003) Some þ þ 2 2 2 þ General The difference between expert systems and information systems is small Duchessi and O’Keefe (1995) Yes þ þ þ þ þ þ þ J .K . N u rm in e n e t a l. / E x p e rt S y ste m s w ith A p p lic a tio n s 2 4 (2 0 0 3 ) 1 9 9 – 2 1 1 2 0 4 systems could in theory cover the whole problem domain, in practice there are not enough development resources. When the scope is wider the level of expertise is typically smaller. In that case the expert system is not targeting complete solutions. Instead, it relies on humans to supplement the draft or partial solutions provided by the system. In this way it typically is easier to cover a much larger area of the problem domain. A comparison of the two graphs in Fig. 1 shows that total work for human experts is less when the expert system covers a wider problem area with rough level of detail. At the same time the average quality of solutions tends to be higher. The consequence is that the system cannot work autonomously and a human expert is continuously needed. Most of the applications of this study had a wide scope. Only the loan decision-making tool (MATIAS) and the radio link diagnostics tool (DMG) had a narrow scope. When the problem domain is large, like for the planning applications, the wide scope allows some support for most of the tasks. 3.1.2. Focused objective Hayes-Roth (1997) states that focusing on a single objective, such as to reduce time, to reduce cost, or to produce a better design, is beneficial. Our experience strongly contradicts this. In all of our applications we have had multiple goals. Typically we have tried to improve quality, increase speed and save time at the same time. The need to emphasize multiple goals is especially evident in planning tasks. Simon (1995) observes that design work has multiple goals and constraints and that their priority and importance changes during the design process. Our experience is very much in line with this. For any reasonable complex planning task it is impossible to list all of the objectives in advance. Focusing the design on the optimization of a single criterion tends to create plans that are not feasible in practice. 3.1.3. Stability of environment Stability of environment is another issue where our experience strongly contradicts (Hayes-Roth, 1997). A stable environment helps the implementation but the assumption of a stable environment is too idealistic. The shortening release cycles of all kinds of products make the environment even more dynamic than what is has been in the past. It is essential that the expert systems are useful in changing environments. Otherwise they would either be outdated or their maintenance would be a constant problem. Systems with a narrow scope have a high risk in this respect. As can be seen in Fig. 1 if the problem domain moves slightly (e.g. with the introduction of a new technology) it can happen that the narrow scope system falls out from the problem domain. For the wide scope systems this risk is much smaller. 3.1.4. High degree of automation High degree of automation may seem like a desirable feature. As described in Hayes-Roth (1997) this would cause savings in the operation of the system. In this issue our evidence is mixed. The narrow scope systems, MATIAS and DMG, provide a high degree of automation. In DXCON and NICAD some parts are highly automated while others require a lot of user involvement. After the planner had created the configuration DXCON took care of hardware details and thus saved work and reduced errors. NICAD automated a lot of the calculation even if human experts made the design decisions. In the other applications the general level of automation was not very high. We can see several explanations for these observations. First, MATIAS and DMG were intended for non-expert users. All the other applications aimed at improving the performance of an expert user, which seems to be common in engineering applications. Secondly, increasing the degree of automation requires increasingly more development effort. When the problem Fig. 1. Narrow vs. wide scope. J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211 205 complexity and width increases reaching a high degree of automation may become infeasible goal. In terms of Fig. 1 the question is how much width is sensible to sacrifice to increase automation. One practical way is to develop some modules with high degree of automation to automate routine tasks and rely on the end-user in the more difficult tasks. Modules containing calculation or detailed but straightforward computation are good candidates. 3.1.5. High degree of repetition High level of repetition is clearly useful. It allows the benefits from the system to be drawn continuously and thus makes the investment in the system worthwhile. In addition to the benefits that are gained from using the system the high degree of repetition is also useful for the development. More repetition means more feedback to the development, more need for new features, and more incentive to fund the further development and maintenance. 3.1.6. Small project Our expert system projects have been small with less than 10, typically 2 – 3, full-time developers plus part-time experts. We do not have hands-on experience with the larger expert system projects. With our data we can only conclude that small projects work but the data does not say anything about large project. Our subjective feeling, based on experience with expert system and large-scale software development projects, is that large-scale expert system projects should not be more difficult than any large software projects as long as the expert system parts are developed as compact, well- defined, and reasonably independent components. 3.2. Development 3.2.1. Users prefer usability over automation Finding the proper balance between usability and automation seems to be an issue in most projects. Since the applications are developed with limited resources it is important to focus the development to the most useful direction. The experience with MATIAS and with some modules of DXCON indicates that when the degree of automation is high the usability is not considered very important. If the automatically generated solutions are fully correct the user has no need to analyze and interpret the results. The situation is different in applications where full automation is not feasible. In that case usability is a major issue since users need to control the application, understand the solutions, analyze the trade-offs, and manually adjust the solutions until they satisfy the needs. This seems to be the most common case. Automation is valued but only up to a certain level. The users want ultimately to be in control and understand the solution before they can accept it. This division also mirrors the difference between expert and ordinary users. MATIAS brought the loan knowledge to the customer agents in bank offices. DXCON had two classes of user: senior experts who defined the component configurations and less experienced expert users who utilized the expertise. In the other applications the users were typically the experts themselves. It seems that in tools that are intended for the experts the usability is very important. The expert can compensate for missing auto- mation by using his experience and competence. The ordinary user is more or less blindly relying on the results of the tool and therefore is not very interested, or competent, to analyze the rational behind the results. Graphical user-interfaces naturally have a major impact on usability. In fact, the graphical user interfaces seem to be one of the most important results of long-term practical value of the AI boom in late 1980s. Today the graphical user interfaces are commonplace but 15 years ago the expert systems were among the first applications with intuitive user-friendly graphical interfaces. 3.2.2. Expert systems complement rather than replace human experts A key characteristic of our applications has been that they are intended for expert users such as product designers or electrical engineers. In such a role the tools support, rather than replace, the expert. According to our experience for any non-trivial planning task it is not possible to create completely automatic solutions. An expert user is essential to handle changing and unexpected needs, multiple objectives, and other similar aspects where the human flexibility and a wide under- standing are important. Systems which combine the strengths of both humans and computers, intelligence amplification systems (Brooks, 1996), seem thus to be the most promising direction for practically successful applications. Millett and Powell (1996) made a similar observation that ‘one reason for success is that expert system has not attempted to undertake the whole task, merely a part of it that is time consuming for staff. This frees experts for more complex tasks not handled by the system’. Requiring the user involvement may seem like a failure from a puritanist AI perspective but from the practical perspective it solves many problems and results into less complex systems. Requirements for the completeness and correctness of the system can be less strict. The system can consist of a set of basic tools that are controlled by the user. Changes and new demands do not necessarily require a modification to the system but can be handled by the end user. As a result the application development is easier and the applications are more adaptive to external changes. 3.2.3. Early benefits to experts themselves are important The experts with the relevant knowledge and under- standing of the problem are key to expert system success. Typically the experts are busy, which is one reason why it makes sense to develop the expert system. There is often J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211206 a conflict between the short-term and long-term priorities when the expert schedules his time between different tasks. Our experience suggests that one very useful practice to involve the expert is to develop the system in such a way that it offers benefits almost immediately to the experts themselves. It is not enough to satisfy the end-users. It is equally, and sometimes more, important to satisfy the expert to be able to create a good enough system to be delivered to other experts and end-users. When the experts can directly influence the development their attitudes tend to be more positive towards the new systems. Early participation in development reduces the often-perceived risk that the expert will be replaced by the expert system. Instead active involvement emphasizes the importance of the expert and the tool makes the expert more competent. Duchessi and O’Keefe (1995) state a more general observation that a system that directly benefits a user fosters operational use. While we agree with their observation we feel that emphasizing early benefits and to the experts themselves is important. Otherwise the development easily slows down and stops. At least one of the successful cases that (Duchessi & O’Keefe, 1995) analyzed seems to confirm this. 3.2.4. AI should be embedded part of a bigger system All of the applications, except MATIAS, show that expert system modules cannot be handled in isolation. To be useful they have to be embedded to larger software systems. The expert system modules perform key tasks and provide the intelligence of the application but without the other parts the intelligent parts would not be useful alone. Millett and Powell (1996) found that stand-alone systems made users unhappy by requiring existing data to be re-keyed. Two of the three successful expert systems in Duchessi and O’Keefe (1995) were actually database applications where the expert system module was ‘icing on the cake’. The first implication of this is that expert system modules are most useful when implemented in the same environment and same tools as the rest of the application. Often this means the use of a standard programming language like Cþþ or database centric solutions, like in DXCon. It still makes sense to specify the expertise in special format in definition files that are either loaded to the application or compiled and linked as part of the application. Another implication is that in the development a lot of effort has to be spent on the non-AI parts. This work involves, for instance, the graphical user interface as well as interfaces to external systems, and can easily exceed the effort needed for the expert system parts. In the project planning it is thus important to allocate enough effort to all parts of the system and avoid the mistake of focusing only on the key expert system parts. 3.2.5. Simple, straightforward solutions work best If the ambition level is too high it easily happens that the system uses sophisticated techniques and solves limited subsets of the target problem very well, but it is not able to handle the real size problems. In almost all of the applications in this study the experience was similar. The objectives for the degree of automation were at the start much higher than what was accomplished in the end. The advanced AI modules were pruned and the mainstream technology modules started to dominate the development. Better results can be achieved when the system focuses on solving the practical problem. Millett and Powell (1996) found that successful organizations were less concerned with pushing the technology and more inclined to concentrate on the application. A typical sin on the expert system area has been to try to force a certain solution to a problem. Instead it is important to consider which solution approach would the optimal one for a particular problem. Although expert systems are good solutions for some problems, for other ones e.g. the use of optimization and other mathematical algorithms could be better choices. Developers with a wide understanding of possible solutions techniques have a key role. Naturally finding developers with such broad experience is difficult. 3.2.6. If-then rules considered harmful The analysis of the applications of this study indicates that if-then rules are not the right way to represent engineering knowledge. This finding is in contrast with the study (Stylianou et al., 1992) that found support for backward chaining to be the third most important feature in expert system shell selection. MATIAS, the only non-technical expert system of this study, was the only deployed application that successfully used if-then rules. A few of the other applications also experimented with using if-then rules on some parts but these trials did not give good results. Technical knowledge deals with artifacts and their relationships, formulas, and numeric values that are more easily expressed as classes and objects. Rules are more suitable for describing knowledge on the area of legal reasoning, such as the terms of loan approval. Another observation is that the maintenance of if-then rulebases is more difficult than for most other knowledge representations. Instead of the general if-then format a special application specific structure is useful for easier development and maintenance. The preexisting solutions available in application development tools are seldom suitable as such. Finding the right structure is very similar to software architecture design, which is a key step for successful software development. 3.2.7. Lots of custom work The large individual differences between developers (see e.g. McConnell (1993)) are likely to be even bigger in expert system development than in normal software development. The reason for this is that, being in the intersection of AI and J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211 207 information technology, expert system development requires more creative solutions than normal software development. The task does not only require good software development competence but also good understanding about the applicability, limitations, and usefulness of different solutions techniques. Because many expert systems are pushing the state of the art it is not possible to blindly follow the solutions used in other systems. 3.2.8. Fast and agile development is important In the engineering domain the expertise is developing on a fast speed. This has two implications. First, rapid development is important to bring the solution into use at the right time. If the development time is too long then the system may not be up to date anymore when it is ready. Secondly, without proper maintenance and adequate tools for it the system very soon becomes obsolete. Fast and agile development process is thus important. At early development phases flexibility is needed to discover the way to approach the problem and to find the right architectural solution. Also, as discussed above, it is important to bring the tool early into use to ensure the expert commitment and feedback and continue the devel- opment in an iterative and incremental fashion. Although Duchessi and O’Keefe (1995) do not explicitly state the importance of this it seems that a user driven phased development has been used at least in the successful systems that they analyzed. The agile development is another area besides graphical user interfaces where the expert system community has done pioneering work. The incremental, customer driven style of expert system development has only recently become accepted, and increasingly popular, in software community at-large (see e.g. Cusumano and Yoffie (1999)). 3.2.9. Knowledge-based applications are based on domain specific application engines and generators As shown in Fig. 2 the applications of this study can be divided into three groups † Monolithic applications have no clear difference between knowledge and application logic. Monolithic applications are very close to general IS applications. † Knowledge-based applications separate the knowledge and the processing of knowledge (engine). This is the traditional model for expert system applications. The same engine can be used with different knowledge to create different applications. Maintenance of the appli- cation is supposed to be easy since the changes are mainly made to the knowledge part. † Application generators are not intended for end-users but for developers. Their main purpose is to assist in the application development. Since the development and maintenance of a knowledge-base is typically difficult, the generator applications are often essential in making the knowledge bases more robust, easier to understand, and faster to develop. The generators typically work by representing the knowledge to the developer and to the expert in a form that is easier to understand and manipulate. In the application the primary goal of knowledge representation is efficient execution, in the generator it is human comprehension. This study suggests that the textbook solution to use a general-purpose expert system shell and to concentrate on the knowledge base in the application development does not work. There is a need to create very specialized, domain specific tools for the engine (NICAD) or for knowledge base development and maintenance (DMG, DXLib). Such tools take advantage of the special structure of the knowledge in a domain and make the knowledge more manageable by constraining the knowledge to a more compact, consistent and modular format. MATIAS was the only application developed with an expert system shell but the experience with it also triggered the development of knowledge engineering tools (Parpola, 1995) to better handle the knowledge acquisition process. DXCON was the only application that did not support this observation. An explanation for this is that DXCON was a database centric application and stored the knowledge in Fig. 2. Applications and their types. J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211208 the form of configurations components to relational database tables. This kind of open architecture allowed easy addition of new modules developed with different technologies. Own development effort is needed to find the right structure for the knowledge but after that initial effort the rest of the development becomes a lot easier. Developing the engines and generators requires major work. The knowledge-based system creation is thus only partly about the knowledge acquisition. It is also about software development to create suitable tools that can be efficiently used to build the applications. For instance, at the beginning most of the work on the NICAD application had to be spent on developing the engine. The situation may have changed and today there are more tools that can be used as building blocks for the applications. Rather than expert system tools such software is marketed as domain specific automation tools. 3.2.10. Monolithic applications are sometimes better than knowledge-based applications For most end-users and applications the difference between monolithic applications and knowledge-based applications is small. A major difference can arise if the normal operation of the system requires that experts dynamically update the rules. The division has a bigger effect for the development and maintenance of the applications. Knowledge-based appli- cations are often seen as the orthodox way to develop expert systems. Also in our case that was the starting point of all the applications. However, as the development progressed Planner, RFT, and CabCon shifted to the monolithic style. This was mainly caused by the fact that the importance of application specific knowledge turned out to be less important than initially expected. Knowledge-based application have their price. The development of a knowledge-based application, in our case, required also the development of the engine and generator tools. Reaching the needed generality with such tools requires development effort. Separating the knowledge from the engine also tends to reduce the execution speed. Our observation is that the selection between monolithic and knowledge-based application depends on the type of knowledge. If the knowledge is quite static and not growing, if the amount of knowledge is not very large, and if its representation with mainstream information technology is easy then a monolithic application can be a good choice. Proper software architecture makes it also possible to modify the knowledge. Naturally it is harder than in the case of a knowledge-based application. Further characteristics of successful monolithic appli- cation seem to be a wide and general application domain, high importance, and expert users. 3.2.11. Rules of normal SW development apply Our experience suggests that there is very little difference between software development and expert system develop- ment. Most of the successful solutions in our applications are very close to normal software development. Likewise the process of developing an expert system and developing normal software is very similar. There are differences in emphasizing the importance of different steps rather than the actual work. Our experiences confirm those of Duchessi and O’Keefe (1995) who stated that factors leading to success in information system and decision support system develop- ment are also important for expert system implementation. The normal software process activities like development, testing, and maintenance easily become very difficult when a larger knowledge base is needed. This is especially problematic if the modularity of the knowledge is low. If the knowledge is stored in rulebases with little internal structure the number of alternative execution paths becomes very high. It is no longer possible to understand how the rules interact, is the rulebase complete, or what happens when something unexpected happens. Testing all of the execution paths is clearly beyond practical limits. Moreover, the maintenance work easily is very difficult. Maintenance is the biggest consumer of software develop- ment resources. Some estimates assume that 70% of software lifetime cost is spent on maintenance (Boehm & Basili, 2001). Millett and Powell (1996) observe that expert systems are no different from conventional systems in that they require maintenance. To us it seems that the maintenance of knowledge bases is even more expensive. First, the expertise tends to change frequently. Secondly, the complexity of expertise and its presentation in knowledge bases makes incremental modifications difficult. Finally, low modularity may result into the need to retest the whole application after each small modification. It can also happen that different experts view the functionality from different angles. It is thus possible that a rulebase that is developed with one expert and is later maintained by another is no longer consistent. One of the big successes in software engineering has been object-oriented programming and the software components. They have made the software reuse a practical issue. How is the reuse of expertise? For successful reuse the expertise would need to be packaged into some kind of component. Moreover, a mechanism would be needed to control, expand, and modify the behavior of the components from the outside. At least rulebases are quite sensitive to small changes. Already a change in a single parameter value in one rule requires a thorough understanding about the possible effects of the change. Other domain specific formats are potentially better but still there are major difficulties for black-box reuse. Domain specific engines and generators are more applicable for reuse than the knowledge itself. 3.3. Operation 3.3.1. Move towards mainstream software and hardware platforms The hardware platform tended to shift from specialized computers, Symbolics, to Unix workstations and personal J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211 209 computers running Microsoft Windows. The applications that were initially developed with Lisp or expert system shells were converted to C or Cþþ. There were several reasons for this shift. For widely used applications the software and hardware platform cost is essential. Furthermore, with the rise of the personal computer, users tended to prefer to do all of their work on single computer using tools that had a familiar look-and- feel. The mainstream environment also allowed the creation of open interfaces to link the expert system applications with other applications. It also turned out that the platform support for expert system tools was not adequate. The companies providing specialized AI platforms, Lisp programming environments, and application generators were working on a niche market. The volume of the market was not large enough to guarantee the high level of robustness, development speed, and support that is needed for serious products. Moreover, most of the expert system environments were closed systems and did not easily allow the use of normal, mainstream development tools such as version control systems, makefiles, debuggers, etc. Even if porting the application to new environments caused extra work this approach was, in fact, very productive. The developer-oriented environments and application generators allowed a fast start and enabled early end-user involvement. When the user needs were understood and the appropriate solution found in this way, it was relatively straightforward to reimplement the relevant parts of the solution in another environment. Another benefit of the platform change was that it was a good opportunity to improve the speed and robustness of the application. When the needs were clear it was possible to choose architectural solutions that allowed efficient and future-proof applications to be written. 3.3.2. Successful systems tend to move out from the company Our data gives some support to the observation that successful systems often lead into a spin-off company that focuses specially on the further development and mainten- ance of the expert system or the platform tools. Even if the balance is even, three cases supporting and three refuting the claim, the applications supporting the claim are the most successful ones in terms of user base and longevity. There can be a result of numerous factors in the operating environment and in strategy of our company that explain this observation. Expert systems, in their role as support tools, improve the operational efficiency but are not the primary products of a larger company. Their benefits for the main business are indirect and hard to quantify. Therefore, with increasing pressure to focus on the core business, the development and maintenance of such system may not satisfy the profit requirements. Application development requires that a lot of effort is spent on monolithic applications or on domain specific engines and generators. It thus becomes attractive to reuse the platform in order to divide the platform development cost between several applications. Such a generalization, which e.g. happened in NICAD, results into the creation of application specific environments that can be used as a basis of a product family. In our case the monolithic application covered a wide domain (project planning) and important special area (analog circuit design) and therefore had a large market and good business potential. A dedicated spin-off company may be in a position to handle the further development and marketing of such tools. While the mother company is mainly interested in the applications, the spin-off company can focus on the expert system technology, develop a variety of solutions around the platform, and market them to different customers. It also seems that the platform technology is more stable and easier to maintain than the actual applications. 3.4. General 3.4.1. The difference between expert systems and information systems is small Duchessi and O’Keefe (1995) expected that many of the factors that lead to successful implementation of other systems are also important in expert system implementation. According to our experience many of the organizational and system development issues are common between expert systems and information systems. First, it is important to understand that the applications are essential, not the technology. Secondly, it is important to develop the applications to solve the right problems at the right time. Millett and Powell (1996) state that it is important that an expert system project tackles a major business problem at an early stage. Duchessi and O’Keefe (1995) found that problem urgency and growth of the underlying application area are important factors for expert system success. Thirdly, the management support, matching the solution to the organizations needs, and right organization support are important (Duchessi & O’Keefe, 1995). Our experiences support all of these observations. For the expert system development it would be important to utilize all the accumulated experience of information system development. Actually, we feel that one of the mistakes of expert system development may have been to view it as a distinctive area and develop own processes, tools, and methods independently from the information systems. Potentially, better results would have been achieved if the problems had been approached from the information system perspective and the special expert system techniques would have been utilized only when appropriate. J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211210 4. Conclusions The domain of engineering applications is complex: technology is changing, tasks are open-ended, and multiple, often implicit, goals are common. Traditional, standalone expert systems have problems to cope with these challenges. Contradicting previous studies, this study suggests that wide scope, multiple objectives, unstable environment, and moderate degree of automation are important considerations for expert system development. The user of an engineering expert system typically is and has to be an expert by him/herself. Expert systems should complement rather than replace human experts. Particular attention should be given to usability, which expert users often consider more important than automation. Part of good usability is that the expert systems are embedded and integrated parts of larger information systems. Implementation of such systems has a lot in common with normal software development. In fact, most of the systems were ported to mainstream software and hardware platforms during their lifecycle. Specially developed domain specific application engines and generators turned out to be highly useful. General-purpose knowledge representations, in particular if-then rules, did not suit well for engineering applications. Fast and agile development was found to be important to cope with the changing environment. Experts, both as developers and users of the systems, are in a crucial role. Organizing the development in such a way that early versions of the systems gave benefits to the experts themselves, was a key means of getting the experts involved deeply enough in spite of their overload with other urgent business-critical tasks. The problem with empirical evaluation is that a single study provides only a piece of evidence. The analysis of multiple systems and the long time period increase the confidence of the observations of this study. However, numerous factors, like the culture of our development group, may bias the results. Moreover, it is not clear how sensitive the results are to the type of industry, application domain, or point-of-time. The findings of this study are based on observations about eight successful applications. It would be interesting to deepen the analysis and look for more fundamental reasons behind these observations. The development of such theory would give more confidence to the observations and provide a general framework where the results of case studies could be positioned. References Allen, J. F. (1998). AI growing up: The changes and opportunities. AI Magazine, Winter, 13 – 23. Boehm, B., & Basili, V. R. (2001). Software defect reduction top 10 list. IEEE Computer, (January), 135 – 137. Brooks, F. P. (1996). 1994 ACM Allen Newell Award acceptance speech. Communications of the ACM, (March). Cusumano, M. A., & Yoffie, D. B. (1999). Software development on internet time. IEEE Computer, (October), 60 – 69. Doyle, J., & Dean, T. (1996). Strategic directions for artificial intelligence. ACM Computing Surveys, 28(4), 653 – 670. Duchessi, P., & O’Keefe, R. M. (1995). Understanding expert systems success and failure. Expert Systems with Applications, 9(2), 123 – 133. Fenton, N. (2001). Conducting and presenting empirical software engineering. Empirical Software Engineering, 6, 195 – 200. Grandon Gill, T. (1995). Early expert systems: Where are they now? MIS Quarterly, 19(1), 51 – 81. Guimaraes, T., Yoon, Y., & Clevenson, A. (1996). Factors important to expert systems success—A field test. Information and Management, 30(3), 119 – 130. Guimaraes, T., Yoon, Y., & O’Neal, Q. (1995). Success factors for manufacturing expert system development. Computers and Industrial Engineering, 28(3), 545 – 559. Harmon, P. (1995). Object-oriented AI: A commercial perspective. Communication of the ACM, 38(11), 80 – 86. Hayes-Roth, F. (1997). Artificial Intelligence—What works and what does not? AI Magazine, (Summer), 99 – 113. Karonen, O. (1987). Intelligent CAD. Norwegian Artificial Intelligence Society (NAIS) Seminar. Karonen, O. (1989). Experiences with different application generators in cable configuration (in Finnish, original title Projekti jossa jouduttiin vaihtamaan kehitintä). BLANKO’89 conference. Ketonen, T., Lounamaa, P., & Nurminen, J. K. (1988). An electronic design CAD system combining knowledge-based techniques with optimization and simulation. In J. S. Gero (Ed.), Artificial intelligence in engineering: design. Amsterdam: Elsevier. Kontio, J. (1991). Matias: Development and maintenance of a large but well-defined application. Expert Systems with Applications, 3(2), 241 – 248. McConnell, S. (1993). Code complete. Redmond: Microsoft Press. Millett, D., & Powell, P. (1996). Critical success factors in expert system development: A case study. Proceedings of the 1996 Conference on ACM SIGCPR/SIGMIS Conference, 214 – 222. Parpola, P (1995). Object-oriented knowledge acquisition. Licentiate Thesis, University of Helsinki. Simon, H. A. (1995). Problem forming, problem finding, and problem solving in design. In A. Collen, & W. W. Gasparski (Eds.), Design and systems: General applications of methodology (pp. 245 – 257). New Brunswick: Transaction Publishers. Stylianou, A. C., Madey, G. R., & Smith, R. D. (1992). Selection criteria for expert system shells. Communications of the ACM, 35(10), 30 – 48. Tyrväinen, P. (1991). DMG—Object-oriented iterative knowledge acqui- sition for the generation of diagnostic hypertext systems. LISP Pointers, 4(1). Tyrväinen, P., Saarinen, P., & Hätönen, K. (1993). Domain modelling for technical documentation retrieval. In H. Kangassalo, H. Jaakkola, K. Hori, & T. Kitahushi (Eds.), Information modelling and knowledge bases IV (pp. 388 – 399). IOS Press. J.K. Nurminen et al. / Expert Systems with Applications 24 (2003) 199–211 211 What makes expert systems survive over 10 years-empirical evaluation of several engineering applications Introduction Applications NICAD RFT Nokia planner CabCon DXCON DMG DXLib MATIAS Characteristics of successful systems Domain Development Operation General Conclusions References Copyright: © 2003 Elsevier Science. Reprinted with permission from Expert Systems with Applications 24, No. 3, pages 199-211. 
work_m2cwseludvhjrmju2eamkqcyby	----	doi:10.1016/j.eswa.2007.03.013 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 2334–2341 Expert Systems with Applications Novel yield model for integrated circuits with clustered defects Lee-Ing Tong, Li-Chang Chao * Department of Industrial Engineering and Management, National Chiao Tung Uninversity, 1001 Dah-Hsei Road, Hsin-Chu 300, Taiwan, ROC Abstract As wafer sizes increase, the clustering phenomenon of defects increases. Clustered defects cause the conventional Poisson yield model underestimate actual wafer yield, as defects are no longer uniformly distributed over a wafer. Although some yield models, such as neg- ative binomial or compound Poisson models, consider the effects of defect clustering on yield prediction, these models have some draw- backs. This study presents a novel yield model that employs General Regression Neural Network (GRNN) to predict wafer yield for integrated circuits (IC) with clustered defects. The proposed method utilizes five relevant variables as input for the GRNN yield model. A simulated case is applied to demonstrate the effectiveness of the proposed model. � 2007 Elsevier Ltd. All rights reserved. Keywords: Clustered defects; General regression neural network; IC; Pattern; Yield model 1. Introduction Wafer yield is an important index of success used in inte- grated circuits (IC) manufacturing. Wafer yield is defined as the probability that a chip on a wafer has no defect. Defects are physical anomalies which result in circuit faults; dirt particles are the primary source of defects in IC manufacturing (Ferris-Prabhu, 1992). Numerous mathematical models have been developed for predicting wafer yield in the last 40 years (Cunningham, 1990; Stapper, 1991; Stapper & Rosner, 1995; Tyagi & Bayoumi, 1992). Most of these models treat wafer yield as a function of chip size, mean number of defects per chip and the average number of defects per unit area; the Poisson model, compound Poisson models and negative binomial model are examples such models (Cunningham, 1990). The Poisson model is the simplest model to use; however, to successfully predict wafer yield, defect must occur independently with constant probability of occurring in any small area on a wafer (Albin & Friedman, 1991). If these assumptions hold, defects are uniformly scattered over a wafer. However, Stapper (1985) reported that 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.03.013 * Corresponding author. Tel.: +886 35 731 896; fax: +886 35 733 873. E-mail address: lichang.iem91g@nctu.edu.tw (L.-C. Chao). defects are typically clustered rather than dispersed ran- domly over a wafer, and this distribution becomes more evident as wafer size increases. Clustered defects usually violate the independence assumption of the Poisson model. The Poisson model, therefore, underestimates actual yield when defects cluster. Under this scenario, numerous yield models obtain more accurate yield predictions than the Poisson model. Compound Poisson yield models are complicated and only evaluate the relationship between chip size and yield (Cunningham, 1990). The cluster parameter a of the nega- tive binomial model can be very scattered and negative when the model is applied to predict yield (Cunningham, 1990). Consequently, these mathematical yield models have particular problems in predicting wafer yield. Dupret and Kielbasa (2004) use the partial least square (PLS) regres- sion methods to model the yield from measurements obtained during the production. However, an advanced statistics is needed to use the PLS regression methods. Neural networks can handle problems such as recognizing complicated patterns and fitting nonlinear functions. Back- Propagation Neural Network (BPNN), known for its general pattern-mapping capability, can be applied to numerous prediction problems and always performs well (Bishop, 1994; Fausett, 1994). However, obtaining good mailto:lichang.iem91g@nctu.edu.tw L.-I. Tong, L.-C. Chao / Expert Systems with Applications 34 (2008) 2334–2341 2335 prediction network requires substantial effort to identify BPNN’s parameters, such as the number of hidden layers, number of hidden units, learning rate and momentum. Furthermore, the BPNN model has certain problems such as local optimal solution, overtraining and undertraining. Compared with BPNN, the General Regression Neural Network (GRNN) model has numerous advantages: learn- ing is fast; only one parameter is required; there are no overtraining or undertraining problems; and, the likelihood of obtaining a global optimal solution is higher than with BPNN. This study presents a novel yield model which employs GRNN to predict wafer yield with clustered defects. The proposed GRNN yield model utilizes five relevant variables as input variables to predict the wafer yield: number of defects; chip size; mean number of defects per chip; mean number of defects per unit area; and, clustering index. The prediction accuracy of the proposed approach is com- pared with those of the negative binominal yield model and the BPNN yield model. A simulated case is presented to demonstrate the effectiveness of the proposed approach. 2. Yield models The Poisson yield model (Ferris-Prabhu, 1992), which is based on the Poisson distribution, is Y 1 ¼ Pðk ¼ 0Þ¼ e�k0; ð1Þ where k represents the number of defects in a chip and k0 represents the mean number of defects per chip. The Pois- son yield model was sufficiently effective for small chip sizes and tended to underestimate yields for larger chip sizes (Cunningham, 1990). To identify the clustering properties of defects in the yield model, some spatial distributions, including compound Poisson distributions, have been con- sidered (Raghavachari, Srinivasan, & Sullo, 1997). The compound Poisson yield model replaces defect density, which is assumed to be a constant in the Poisson yield mod- el, with a probability density function. The compound Poisson yield model can be described as Y ¼ Z 1 0 e�DAfðDÞdD; ð2Þ where D represents the defect density, A represents the chip size and f(D) represents the probability density function of defects. The compound Poisson yield model is complex and only considers relations between chip size and yield. The negative binomial yield model, which is a widely applied yield model, employs a gamma function for the dis- tribution of defect density (Okabe, Nagata, & Shimada, 1972; Stapper, 1973). The negative binomial distribution can be described as PðkÞ¼ Cðk þ aÞð�k=aÞk k!CðaÞð1 þ �k=aÞkþa ; k ¼ 0; 1; 2; . . . ; ð3Þ where �k and a are parameters of the negative binomial dis- tribution. The negative binomial yield model is Y 2 ¼ 1 ð1 þ �k=aÞa : ð4Þ Parameter a, called the cluster parameter, can be calculated as a ¼ �k2 ðr2 � �kÞ ; ð5Þ where �k is the mean number of defects per chip and r2 is the variance. The negative binomial model has been shown to be a powerful prediction model in IC manufacturing (Cunningham, 1990). However, reports also show that the cluster parameter a in the negative binomial model can be very scattered and negative when the model is used to predict yield (Cunningham, 1990). Langford, Liou, and Raghavan (2001) presents a simple robust windowing method for the Poisson yield model to extract the systematic and random components of yield from wafer probe bin map data. Liou et al. (2002) presents a statistical modeling of MOS devices for parametric yield prediction. Skinner et al. (2002) discuss two classes of tradi- tional multivariate statistical methods and a classification and regression tree (CART) method for modeling and anal- ysis of wafer probe test data to determine the cause of low yield wafers. Meyer and Park (2003) present a center-satel- lite model to Predicting defect-tolerant yield in the embed- ded core context. Dupret and Kielbasa (2004) presents partial least square (PLS) regression methods to model the yield from measurements obtained during the produc- tion. Hong, Milor, Choi, and Lin (2005) utilize two models which are derived from the Poisson yield model and the neg- ative binomial yield model for the effect of area scaling on IC reliability. Kim and Baldwin (2005) present a theoretical yield model for assembly process of area array solder inter- connect process. Other yield models used in various compa- nies are summarized in Stapper and Rosner (1995). In summary, existing wafer yield models have significant limitations: clustered defects cause the conventional Pois- son yield model to underestimate wafer yield; the com- pound Poisson yield model is too complex; the cluster parameter a of the negative binomial model can be sub- stantially scattered and sometimes negative; and, many parameters must be set when applying the BPNN model. Such drawbacks affect performance when these models are employed to predict yield. The most accurate models are the negative binomial (Cunningham, 1990; Stapper, 1973) and BPNN network (Bishop, 1994; Fausett, 1994) yield models. Only these two models, therefore, were selected for comparison in this study. 3. General regression neural network implementation The major difference between GRNN and other super- vised neural networks is that GRNN can treat continuous valued outputs and categorize data, and there are fewer training parameters are required, such as the number of hidden layers, number of hidden units, learning rate and 2336 L.-I. Tong, L.-C. Chao / Expert Systems with Applications 34 (2008) 2334–2341 momentum, than in BPNN. Moreover, GRNN can be used for any regression problem in which a linearity assumption is violated, and it converges fast on the optimal regression surface as the number of samples becomes substantially large. The GRNN model, then, is used in this study to pre- dict wafer yield. Fig. 1 shows the three-layer network of the GRNN model (Specht, 1991). Input units are merely distribution units which forward measurement variables to the pattern units in the second (hidden) layer. This hidden layer con- sists of one neuron for each pattern in the training pattern. The GRNN is essentially trained after one pass of the training patterns and its activation function normally uses an exponential function. The unique parameter of GRNN is the smoothing factor r which influences the output value; that is, high smoothing factors produce increased relaxed surface fits throughout the data. Unlike the conventional regression model, GRNN can be defined through its joint continuous probability density function, rather than utilizing a specified function that must be determined in advance. Assume that f(x, y) repre- sents the known joint continuous probability density func- tion of a vector variable, x, and a scalar random variable, y; the regression of y on X, then, is E½yjx ¼ X� ¼ R1 �1 yfðX; yÞdyR1 �1 fðX; yÞdy : ð6Þ When the density f(x, y) is unknown, it must be estimated by observations of x and y. The GRNN model utilizes a Parzen (1962) window, which is a nonparameter approach to estimating the joint continuous probability density func- tion f(x, y). The estimator can be represented as f̂ðX; YÞ¼ 1 ð2pÞðpþ1Þ=2rðpþ1Þ � 1 n Xn i¼1 exp � ðX � XiÞTðX � XiÞ 2r2 " # � exp � ðY � Y iÞ2 2r2 " # ; ð7Þ Fig. 1. GRNN block dia where p is the dimension of x, r is the smoothing parame- ter, Xi and Yi are sample values of observations x and y; and n is the number of sample observations. Combining Eq. (6) and (7), Eq. (8) can be obtained as bEðyjXÞ¼ Y_ðXÞ ¼ Pn i¼1 exp � ðX�XiÞTðX�XiÞ 2r2 h iR1 �1 y exp � ðy�Y iÞ2 2r2 h i dyPn i¼1 exp � ðX�XiÞTðX�XiÞ 2r2 h iR1 �1 exp � ðy�Y iÞ2 2r2 h i dy : ð8Þ Eq. (8) can be further simplified as Eq. (9), which is given by bY ðXÞ¼ Pn i¼1Y i exp � D 2 i 2r2 h i Pn i¼1 exp � D2i 2r2 h i ; ð9Þ where D2i ¼ðX � X iÞTðX � XiÞ. The GRNN model utilizes Eq. (9) to estimate y. Typically, the activation function of GRNN network is exponential, as shown in Eq. (10): fðD2i Þ¼ exp � D2i 2r2 � � : ð10Þ A new vector X is subtracted from the stored pattern vector when it enters the network. The squares of the difference are summed and input into the activation function in Eq. (10). Those values which pass through the activation func- tion are the pattern unit outputs and are forwarded to the summation units. The summation units proceed to sum the dot product between a weight vector and the pattern unit outputs to generate an estimate as shown in Eq. (11). Xn i¼1 exp � D2i 2r2 � � : ð11Þ gram (Specht, 1991). L.-I. Tong, L.-C. Chao / Expert Systems with Applications 34 (2008) 2334–2341 2337 Conversely, the summation units sum the dot product between the samples Yi and the pattern unit outputs to generate an estimator as shown in Eq. (12).Xn i¼1 yi exp � D2i 2r2 � � : ð12Þ Finally, the output unit divides Eq. (12) by Eq. (11) to ob- tain the desired estimate of y, which is the same as Eq. (9). GRNN measures how far a given sample pattern is from patterns in the training set. When a new pattern is pre- sented to the network, the input pattern is compared to all of the patterns in the training set to determine how far it is from those patterns. The output that is predicted by the network is a proportional amount of all of the out- puts in the training set. The proportion is based upon how far the new pattern is from the given patterns in the train- ing set. GRNN uses an algorithm to find appropriate indi- vidual smoothing factors for each input as well as an overall smoothing factor. The algorithm proceeds in two parts. The first part trains the network with the data in the training set. The second part tests a whole range of smoothing factors. The method will produce networks which work much better on the test set. The performance of neural networks can be measured by a root-mean squared error (RMSE), which can be calcu- lated as RMSE ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPn i¼1ðAi � OiÞ 2 n s ; ð13Þ where n represents the number of patterns, Ai represents the actual value of output and Oi represents the predicted value. Another indicator for measuring the strength of the relationship between the actual and predicted outputs Fig. 2. A simple representation of is the Pearson’s linear correlation coefficient r. In this study, RMSE and r are applied to evaluate the perfor- mance of the negative binomial, BPNN and the proposed GRNN yield model. 4. Proposed approach 4.1. Defect clustering patterns A major cause affecting yield is the degree to which defects are clustered (Friedman, Hansen, Nair, & James, 1997; Stapper, Armstrong, & Saji, 1983). Hence, the defects clustering phenomenon must be integrated when construct- ing a yield model. In this study, Borland Delphi program- ming language is employed to simulate a variety of defect clustering patterns for 8-in. wafers. Fig. 2 presents a simple representation of defect clustering patterns. Three design factors are employed in this study to simulate defect clus- tering patterns: the cluster pattern; percentage of defects located on grey regions; and, chip size. The following is a brief description of these three design factors. (1) Cluster pattern: Fig. 2 presents one random pattern and four clustering patterns (Friedman et al., 1997). The defects in a random pattern are distributed ran- domly over the entire wafer. Distribution of defects in the four clustering patterns depends on the per- centage of defects located in grey region. Grey region represents the defect-dense areas. (2) Percentage of defects located on grey regions: In the four clustering patterns, four percentages, 60%, 70%, 80% and 90%, of the total number of defects are located in grey regions, and the remaining defects are distributed randomly. the defect clustering patterns. 2338 L.-I. Tong, L.-C. Chao / Expert Systems with Applications 34 (2008) 2334–2341 (3) Chip size: Six chip sizes are considered: 1(1 · 1), 1.44(1.2 · 1.2), 1.96(1.4 · 1.4), 2.56(1.6 · 1.6), 3.24(1.8 · 1.8), and 4(2 · 2) cm2. From these three design factors, a total of 102 simula- tion trials can be obtained. For each simulation trial, up to 292 defects are obtained randomly. The number of defects obtained randomly creates simulations that are sig- nificantly close to real IC manufacturing conditions. The maximum number of chip that can be cut in an 8-in. wafer is 292. Therefore, the number of defects are simulated up to 292 defects to reduce the possible influence of outliers. Each simulation trial represents one defect clustering pattern. 4.2. Procedure of the proposed approach Models of predicting yield can be classified into macro yield modeling and micro yield modeling. Macro yield modeling uses die size, device density, and other large-scale factors to predict yields for new designs. Micro yield mod- eling uses critical device area, parametric sensitivity, redun- dancy effect, and other factors to predict yields (Mullenix, Zalnoski, & Kasten, 1997). Gruber’s general yield model (Gruber, 1994) is the most recognized model in Macro yield modeling, and can be described as Y ¼ Y 0ðD; A; hÞLðYÞ; ð14Þ whereY0 represents the asymptotic yield, which is a real- valued function of D, A, h, and; D represents point defect density per unit area, A represents chip area, h represents a set of parameters unique to the specific yield model, L(Y) represents a real-valued function describing learning effects. In this study, there are numerous attributes that can be obtained from each defect clustering pattern by sim- ple calculation: number of defects; chip size; mean number of defects per chip; mean number of defects per unit area; and, clustering index CI (Jun, Hong, Kim, Park, & Park, 1999). The clustering index CI can be calculated as CI ¼ min s2v �v2 ; s2w �w2 � � ; ð15Þ where vi and wi are a sequence of defect intervals on the x axis and y axis defined as vi ¼ xðiÞ � xði�1Þ; i ¼ 1; 2; . . . ; n; wi ¼ yðiÞ � yði�1Þ; i ¼ 1; 2; . . . ; n; where x(i) and y(i) denote the ith smallest defect coordinates on the x axis and y axis, respectively; �v and s2v represent the sample mean and the sample variance of vi, respectively; �w and s2w denote the sample mean and the sample variance of wi, respectively. The value of CI is close to 1 if the defects are randomly scattered, and the value of CI is expected to be greater than 1 if defects are clustered. These attributed values are input into the GRNN yield model, whereas the only output of GRNN yield model is the actual wafer yield. The number of replications for each simulation is 10; hence, a total of 1020 pairs of input–output data are obtained to train and test the GRNN yield model. The Poisson yield model is easy to apply; however, the effect of defect clustering is not considered by the conven- tional model. This study, proposes a GRNN yield model to predict wafer yield. The proposed approach assumes that wafer yield is affected by each wafer defect. Under this assumption, the proposed approach for the wafer yield pre- diction in IC manufacturing can be described as follows: Step 1: Determine the defect clustering pattern. Obtain the simulated defect wafer map. Utilize Borland Del- phi programming language to simulate all possible defect clustering patterns for 8-in. wafers. Step 2: Calculate all attributed values for each pattern. For each defect clustering pattern on a wafer, cal- culate the following attributed values of patterns: number of defects; chip size; mean number of defects per chip; mean number of defects per unit area; and, clustering index CI. Step 3: Build a GRNN yield model. Input the attributed values in Step 2 into the GRNN yield model. The actual yield of the wafer is the only output of the GRNN yield model. The percentage of the chip without defects on a wafer is used as the actual yield value of the wafer. In this study, the neural networks package NeuroShell 2 is employed to train and test the GRNN network. A trained GRNN network can be obtained after a few training patterns have been input. Finally, the trained GRNN network produces the net- work’s prediction for each pattern in the test set. Step 4: Calculate predicted yields. Input the attributed values in Step 2 into the negative binomial yield model to derive the predicted yields of the model. Then build the BPNN yield model as in Step 3 to obtain the predicted yields for the BPNN yield model. Step 5: Predict and analyze the wafer yield. Utilize the pre- dicted yields obtained by the negative binomial yield model, BPNN yield model and the proposed GRNN yield model to predict the actual yields for the wafer and compare these three yields. 5. Implementation 5.1. A simulation study This section presents a simulation study to demonstrate the effectiveness of the proposed approach. The data required in this simulation study are obtained by employ- ing the Borland Delphi programming language to simulate a variety of defect clustering patterns for 8-in. wafers. A total of 102 combinations for simulation trials are obtained by combining the three design factors outlined in Section 4. Fig. 3. The relationships between the predicted and actual yields for the negative binomial, BPNN and proposed yield models. Table 1 The comparisons of RMSE and correlation coefficients between predicted and actual yields Yield model RMSE Correlation coefficient Negative binomial yield model 0.1203 0.8838 BPNN yield model 0.0960 0.9030 Proposed yield model 0.0914 0.9127 L.-I. Tong, L.-C. Chao / Expert Systems with Applications 34 (2008) 2334–2341 2339 The number of replications is 10 in each simulation; hence, a total of 1020 pairs of input–output wafer attributed values are obtained and used to train and test the GRNN network. These wafers are divided into two parts: one part contains 816 wafers which are used to train the GRNN network; and, the second part contains 204 wafers which are employed to test the accuracy of the GRNN network. These attributed values and the actual yields of 1020 wafers are, respectively, utilized as the inputs and output for the proposed GRNN network. The percentage of the chip without defects on a wafer is used as the actual yield value of the wafer. NeuroShell 2 is utilized to train and test the GRNN network. Trained GRNN network are obtained after inputting the 1020 training patterns. Finally, the trained GRNN network is utilized to produce the net- work’s prediction of yields for these 204 test patterns. In this study, the unique parameter of GRNN network, that is, the smoothing factor r, is set at 0.076, and training is terminated when the error with no improvement of 1%. Substitute the attributed values of these identical 204 simulated wafers, respectively, into the negative binomial yield model to calculate the predicted yields of the model. Then build a BPNN yield model to obtain the predicted yields for the same wafers. The BPNN network in this study are constructed as three layers (Hornick, Stinch- combe, & White, 1990), with input and output units the same as those in the proposed GRNN network and 17 hid- den units (Widrow, Winter, & Baxter, 1987). The learning rate and momentum are 0.6 and 0.9, respective, which are the defaults in NeuroShell 2. The learning epochs are 10,000. To evaluate the performance of these yield models, the relationships between the predicted and actual yield are evaluated. The proposed GRNN yield model (Fig. 3) effec- tively estimates the actual yield. Table 1 presents the com- parisons of RMSE and correlation coefficients r for these three predicted yields and the actual yields. A low RMSE value and high r value indicates that the yield model per- forms better than the other models. The RMSE of the pro- posed approach is 0.0914, which is the smallest value of these three RMSEs, and the correlation coefficient is 0.9127, which is the largest value of these three correlation coefficients. These findings reveal that the proposed approach precisely estimates wafer yield, and more accu- rately predicts yield than the other models. The influences of each of the following three designed factors on the wafer yield are analyzed: cluster pattern; per- centage of defects located on grey regions; and, chip size. The RMSE and correlation coefficients r are applied to analyze the performance of the negative binomial, BPNN and the proposed GRNN yield model. Table 2 shows the RMSE and correlation coefficients of these three yield models for three designed factors. Table 2 reveals that the proposed approach produces the best prediction for wafer yield of the three yield models. This study varies the simulated wafer sizes from 6 to 12 in., and compares the prediction results for wafer yield from the three yield models. The simulation procedure is the same as that in the previous simulation. Table 3 sum- marizes the prediction results for three wafer sizes. Table 3 reveals that the proposed approach performs best of these three yield models regardless of wafer size. This simulation obtains the same result as obtained in the 8-in. wafer sim- ulation; that is, the proposed model performs best, and is followed by the BPNN yield model and the negative bino- mial yield model. The performance of the proposed approach and the BPNN yield model are very close and are better than that of the negative binomial yield model. Although the performance of the proposed GRNN yield model is slightly better than the BPNN yield model, the proposed GRNN yield model has numerous advanta- ges over the BPNN yield model. The GRNN yield model learns quickly and requires only one parameter. Table 2 The RMSE and correlation coefficients of these three yield models for three design factors Design factor Level Negative binomial BPNN Proposed GRNN RMSE Correlation coefficient RMSE Correlation coefficient RMSE Correlation coefficient Cluster pattern Random 0.0989 0.9465 0.0506 0.9716 0.0214 0.9947 Bull’s eye 0.1225 0.7785 0.0622 0.927 0.0386 0.9732 Crescent Moon 0.1315 0.8221 0.0884 0.847 0.0848 0.8599 Bottom 0.1218 0.7833 0.0977 0.7951 0.0977 0.8066 Edge 0.1463 0.831 0.0722 0.9371 0.0691 0.9466 Percentage 60% 0.1136 0.9239 0.0433 0.9801 0.0464 0.9807 70% 0.1312 0.8686 0.0511 0.9687 0.0496 0.9701 80% 0.1287 0.8332 0.0481 0.9511 0.0462 0.9558 90% 0.1359 0.9009 0.0622 0.9256 0.0577 0.9324 Chip size 1 0.1626 0.9529 0.0671 0.9191 0.0626 0.9255 1.44 0.098 0.9018 0.1191 0.7168 0.084 0.8766 1.96 0.0628 0.9503 0.1075 0.8212 0.0839 0.9002 2.56 0.0819 0.9197 0.1381 0.7185 0.0939 0.8824 3.24 0.1526 0.8155 0.1118 0.8047 0.0848 0.9039 4 0.153 0.9078 0.1476 0.7409 0.1128 0.8616 Table 3 The prediction results for three wafer sizes Wafer size The actual yield Yield model The predicted yield RMSE Correlation coefficient Average Std. dev. Average Std. dev. 6-in. 0.6610 0.2070 Negative binomial 0.6350 0.2348 0.1225 0.8597 BPNN 0.6538 0.1939 0.0854 0.9114 Proposed GRNN 0.6593 0.1906 0.0836 0.9145 8-in. 0.5662 0.2207 Negative binomial 0.5784 0.2561 0.1203 0.8838 BPNN 0.5599 0.2142 0.0960 0.9030 Proposed GRNN 0.5666 0.1852 0.0914 0.9127 12-in. 0.7007 0.1584 Negative binomial 0.7517 0.1772 0.0902 0.9073 BPNN 0.6955 0.1365 0.0667 0.9085 Proposed GRNN 0.7045 0.1393 0.0529 0.9448 2340 L.-I. Tong, L.-C. Chao / Expert Systems with Applications 34 (2008) 2334–2341 Furthermore, the proposed GRNN yield model has no overtraining or undertraining problems and the likelihood of obtaining a global optimal solution is higher than that of a BPNN yield model. 6. Conclusion As wafer size increases, the clustering of defects increases. Under this scenario, the conventional Poisson yield model cannot predict wafer yield. In this study, a pro- posed neural network-based approach is presented for defect clustering patterns to predict the wafer yield in IC manufacturing. The GRNN network is used to construct the yield model that can accurately predict wafer yield. The merits of the proposed approach are as follows: 1. The proposed approach utilizes five relevant variables as input variables to predict the wafer yield, rather than utilizing only some of those variables as do the Poisson yield model, compound Poisson yield models and the negative binomial yield model. Therefore, the proposed model is more accurate than both the negative binomial and BPNN yield model. 2. The influences of each of the three designed factors on the wafer yield are analyzed. The RMSE and correlation coefficients of these three yield models for three designed factors reveals that the proposed approach produces the best prediction for wafer yield of the three yield models. 3. This study varies the simulated wafer sizes from 6 to 12 in., and compares the prediction results for wafer yield from the three yield models. This simulation obtains the same result as obtained in the 8-in. wafer simulation regardless of wafer size. 4. The proposed GRNN yield model is fast learning and requires only one parameter to identify for the learning. 5. The proposed approach does not need to construct a complex mathematical yield model and more advanced statistics skill – it only requires a neural network pack- age to predict wafer yield. References Albin, S. L., & Friedman, D. J. (1991). Clustered defects in IC fabrication: impact on process control charts. IEEE Transactions on Semiconductor Manufacturing, 4(1), 36–42. L.-I. Tong, L.-C. Chao / Expert Systems with Applications 34 (2008) 2334–2341 2341 Bishop, C. M. (1994). Neural networks and their applications. Review of Scientific Instrumentation, 65(6), 1803–1832. Cunningham, J. A. (1990). The use and evaluation of yield models in integrated circuit manufacturing. IEEE Transactions on Semiconductor Manufacturing, 3(2), 60–71. Dupret, Y., & Kielbasa, R. (2004). Modeling semiconductor manufactur- ing yield by test data and partial least squares. In Proceedings of 16th International Conference on Microelectronics (pp. 404–407). France. Fausett, L. (1994). Fundamentals of neural networks architectures, algorithms, and applications. Englewood CLiffs, NJ: Prentice Hall. Ferris-Prabhu, A. V. (1992). Introduction to semiconductor device yield modeling. Boston: Artech House. Friedman, D. J., Hansen, M. H., Nair, V. N., & James, D. A. (1997). Model-free estimation of defect clustering in integrated circuit fabri- cation. IEEE Transactions on Semiconductor Manufacturing, 10(3), 344–359. Gruber, H. (1994). Learning and strategic product innovation: Theory and evidence for the semiconductor industry. Amsterdam, Netherlands: Elsevier. Hong, C., Milor, L., Choi, M., & Lin, T. (2005). Study of area scaling effect on integrated circuit reliability based on yield models. Micro- electronics Reliability, 45(9–11), 1305–1310. Hornick, K., Stinchcombe, M., & White, H. (1990). Universal approxi- mation of an unknown mapping and its derivatives using multilayer feedforward networks. Neural Networks, 3(5), 551–560. Jun, C. H., Hong, Y., Kim, S. Y., Park, K. S., & Park, H. (1999). A simulation-based semiconductor chip yield model incorporating a new defect cluster index. Microelectronics Reliability, 39(4), 451–456. Kim, C., & Baldwin, D. F. (2005). A theoretical yield model for assembly process of area array solder interconnect packages with experimental verification. IEEE Transactions on Electronics Packaging Manufactur- ing, 28(4), 344–354. Langford, R. E., Liou, J. J., & Raghavan, V. (2001). The application and validation of a new robust windowing method for the Poisson yield model. In Advanced Semiconductor Manufacturing Conference, IEEE/ SEMI (pp. 157–160). Germany. Liou, J. J., Zhang, Q., McMacken, J., Thomson, J. R., Stiles, K., & Layman, P. (2002). Statistical modeling of MOS devices for parametric yield prediction. Microelectronics Reliability, 42(4), 787–795, 9. Meyer, F. J., & Park, N. (2003). Predicting defect-tolerant yield in the embedded core context. IEEE Transactions on Computers, 52(11), 1470–1479. Mullenix, P., Zalnoski, J., & Kasten, A. J. (1997). Limited yield estimation for visual defect sources. IEEE Transactions on Semiconductor Manufacturing, 10(1), 17–23. Okabe, T., Nagata, M., & Shimada, S. (1972). Analysis of yield of integrated circuits and a new expression of the yield. Electrical Engineering in Japan, 92(12), 135–141. Parzen, E. (1962). On estimation of a probability density function and mode. The Annals of Mathematical Statistics, 33(3), 1065–1076. Raghavachari, M., Srinivasan, A., & Sullo, P. (1997). Poisson mixture yield models for integrated circuits: A critical review. Microelectronics Reliability, 37(4), 565–580. Skinner, K. R., Montgomery, D. C., Runger, G. C., Fowler, J. W., McCarville, D. R., Rhoads, T. R., et al. (2002). Multivariate statistical methods for modeling and analysis of wafer probe test data. IEEE Transactions on Semiconductor Manufacturing, 15(4), 523–530. Specht, D. F. (1991). A general regression neural network. IEEE Transactions Neural Networks, 2(6), 568–576. Stapper, C. H. (1973). Defect density distribution for LSI yield calcula- tions. IEEE Transactions on Electron Devices (Correspondence), 20(7), 655–657. Stapper, C. H. (1985). The effects of wafer to wafer defect density variations on integrated circuit defect and fault distributions. IBM Journal of Research Development, 29(1), 87–97. Stapper, C. H. (1991). On Murphy’s yield integral. IEEE Transactions on Semiconductor Manufacturing, 4(4), 294–297. Stapper, C. H., Armstrong, F. M., & Saji, K. (1983). Integrated circuit yield statistics. Proceedings of the IEEE, 71(4), 453–470. Stapper, C. H., & Rosner, R. J. (1995). Integrated circuit yield manage- ment and yield analysis: Development and implementation. IEEE Transactions on Semiconductor Manufacturing, 8(2), 95–102. Tyagi, A., & Bayoumi, A. M. (1992). Defect clustering viewed through generalized Poisson distribution. IEEE Transactions on Semiconductor Manufacturing, 5(3), 196–206. Widrow, B., Winter, R.G., & Baxter, R.A. (1987). Learning phenomena in layered neural networks. In Proceedings of the First IEEE International Conference on Neural Networks (pp. 411–429). San Diego. Novel yield model for integrated circuits with clustered defects Introduction Yield models General regression neural network implementation Proposed approach Defect clustering patterns Procedure of the proposed approach Implementation A simulation study Conclusion References 
work_m334rjrvzvff7a74iqx4arnla4	----	The Development of ALADIN, and Expert System for Aluminum Alloy Design The Development of ALADIN, an Expert System for Aluminum Alloy Design Martha L. Farinacci., Mark S. Fox, lngemar Hulthage, and Michael D. Rychener CMU-RI-TR-86-5 Intelligent Systems Laboratory The Robotics Institute Carnegie Mellon University Pittsburgh, Pennsylvania 1521 3 *Process Control and Computer Technology Division Alcoa Laboratories Alcoa Center, Pennsylvania 15069 January 1986 Copyright @ 1986 Carnegie-Mellon University This research was funded by the Aluminum Company of America. This report has appeared in Proceedings, Third International Conference on Advanced Information Technology. Gottlieb Duttweiler Institute, Zurich, Switzerland, November 1985. i Table of Contents 1. Introduction 2. Assessment of Artificial Intelligence for Alloy Design 3. Alloy Design Reasoning 4. Case Study of ALADIN 4.1. First Iteration 4.2. Second Iteration 5. Conclusions 6. References 1 1 3 4 5 8 11 11 ii List of Figures Figure 3-1: Alloy design graph Figure 4-1: First iteration flow of control 4 7 A b s t r a c t ALADIN is an expert system that aids metallurgists in the design of new aluminum alloys. The project began in January 1984, and two major iterations of the system design have been completed. The current system is a hybrid of several approaches and artificial intelligence techniques. Declarative representations are used for metallurgical data and concepts. The basic design procedure is encoded in a rule-based, with the potential for flexible control and participation by the user. The system is based on the hypothesize-and-test model, and a constraint-based search is used to find alternative designs. Strategies are included for resolving the interactions and trade-offs among conflicting property targets. Metallurgical relationships cover a broad range of symbolic and numerical types of reasoning, which are integrated in t h e system. In this article, an o\vervicw of the project is given, including discussions of knowledge acquisition techniques, design decisions and implementation. 1 1. Introduction ALADIN (ALuminum Alloy Design INventor) is an expert system that aids metallurgists in the design of new aluminum alloys. The system can be operated in several modes. As a decision support system, it will accept alloy property targets as input and suggest alloying additives, processing methods or microstructural features to meet the targets. As a design assistant, it can evaluate designs supplied by a metallurgist, or provide information that is useful for design from a knowledge bank. ALADIN was developed through the combined efforts of three organizations: The Intelligent Systems Lab of the Robotics Institute of Carnegie-Mellon University, the Artificial Intelligence group at Alcoa Laboratories, and the Alloy Technology Division of Alcoa Laboratories. Design and development of the system began in January, 1984 and continues at the time of this writing. While it is difficult to evaluate a system that is under development, it is hoped that ALADIN will provide many important benefits when completed. The amount of knowledge required to successfully develop new materials is so great that individuals often must supplement their private knowledge with information from books, journals and specialized consultants. ALADIN, when used as a knowledge bank, will provide another source of valued information. There is some hope that by fusing together multiple sources of knowledge from different experts, a system will be developed that exceeds the capabilities of individual experts. The development of new alloys now requires testing many alternatives before one adequate solution is found. An attempt is being made to develop a more complete understanding of the behavior of alloys through the development of quantitative models. These models will , in turn, enable development to proceed with fewer tests. ALADIN will serve as a collection point for these models and also the older empirical methods. Hence ALADIN will aid in the development of new materials at a reduced cost and time. In this report, a case study of the ALADIN project to date is given. 2. Assessment of Artificial Intelligence for Alloy Design Computers have long been used to solve engineering and science problems in the metals industries. For example, finite element models of numerous manufacturing processes, including rolling, extrusion, forming, forging, casting and smelting, have been developed [9]. With these models, the effects of processing on the shape of the material, internal stresses, temperatures, and currents (in the case of smelting) are predicted. Simulation methods are usually used to study processing on a larger scale. Like finite element methods, these models are predictive although they often deal with multiple processing steps. They can determine the effects of material handling practices, plant layout, equipment design and operating practices on product characteristics and equipment performance. Another major application of computing in the metals industry is process or numerical control. Tool adjustments are made in order to control the shape and quality of the product being produced. Often the control models are based, in part, on the equations underlying the predictive models mentioned earlier [2]. The classical applications of computing just described share several important features: 0 The problems are primarily numerical. 2 0 The processes are well understood quantitatively, and equations are available to describe effects. e Well defined algorithms are available to solve the equations and make the appropriate predictions. Engineering problems with these qualities are usually solved with models implemented in a procedural language such as FORTRAN. Alloy design, in contrast to these problems, is characterized by the following features: .Quantitative models for calculating the properties of a proposed design are often not available, or insufficient data prevent their use. 0 Models that are available contain a high degree of uncertainty. *The design process, as practiced by people, often involves reasoning about images appearing under a microscope, abstract concepts, interpretations and heuristic rules of thumb based on experience. The reasoning is highly symbolic as opposed to numeric. 0 The alloy design process, as practiced by people, appears to be similar to the diagnostic problems discussed in artificial intelligence literature. When given a new set of alloy specifications, designers identify an existing alloy, determine in what way the existing alloy fails to meet specifications, and try to find a cause for the discrepancy. Alloy design problems are more complex than diagnostic problems, however, since a great deal of reasoning is required to construct a solution even after the cause of the problem is identified. 0 No fixed algorithm seems to be used in design. Instead, metallurgists seem to search through the space of design variables with guidance from heuristic rules and models. The flow of reasoning depends on what knowledge is available and how reliable that knowledge is at the time the design problem is specified. It was because of these features that an artificial intelligence approach was used in the alloy design problem. Although alloy design has a few features that make it a natural artificial intelligence application, there are other features that make automation difficult. For example, the solution of a single problem takes weeks or often months. This is because the reasoning is complex, many options must be explored, and vital missing information must be searched for. Because of this, it was recognized that knowledge acquisition would be difficult. Furthermore; design involves reasoning about time. Processing of the metal involves changes to the structure of the material which influence later production steps. In other words, the time and sequence in which changes are made is important. It is known that planning problems of this sort are difficult to solve. However, it was felt that the benefits of an alloy design system would outweigh the risks involved in its development. 3 3. Alloy Design Reasoning Before beginning a technical discussion of the ALADIN system, an overview of the reasoning used in alloy design will be given. This overview does not require any knowledge of metallurgy. The introduction of a few basic scientific concepts will make later descriptions of the system clearer. The alloy designer usually begins his problem with a set of constraints on the physical properties of the material he is to make, The constraints are often one-sided, as in: strength must be higher than x and density must be lower than y. Sometimes the constraints involve properties that are not numerical such as machinability or surface appearance. Specifications usually involve many properties and correspond to a material that does not yet exist. This is because the design problem comes from a customer or the corporation who would like something that exceeds the capabilities of everything on the market. The design problem may be overconstrained - there may be no material that can meet the specifications given. It is difficult to know in advance which problems are impossible to solve. The designer is expected to do the best he can and come as close as possible to meeting the targets. Once property constraints are specified, designers usually select some baseline alloy to begin their search for a solution. Design will then involve finding alterations to the baseline that change the characteristics of the material in the direction of the targets. There are several different strategies that people use for selecting a baseline. Some look for a commercial alloy that comes as close as possible to meeting the targets, or the experimental alloy from the most recent iteration in the current design problem. Still others begin with a simple, commercially pure aluminum alloy and design from basic principles. After the baseline is selected, designers look for changes that can be made to the material in order to improve the properties. The things that can be changed directly are: 0 COMPOSITION - the elements added and the amounts 0 PROCESS - the fabrication steps used, their sequence, and such specifications as There are rules that indicate the effects of these design choices on properties. Some examples are: method, temperature or time 0 I f an element with low atomic number is added THEN density will decrease 0 IF Mg is added THEN strength will increase However, the most powerful rules involve reasoning about the microstructure of the alloy. Some examples of these rules are: 0 I f The aging process is done for a long time THEN equilibrium precipitates will form 0 I f equilibrium precipitates are present THEN they are usually incoherent 0 /f incoherent precipitates are present THEN they may form on the grain boundary 0 IF precipitates are on the grain boundary and strength is medium or high THEN elongation and fracture-toughness are low 4 Metallurgists often illustrate the knowledge they work with in a graph (figure [3-11). Nodes correspond to the four classes of knowledge that characterize an alloy: the composition, processing, structure and properties. Arrows indicate the relationships that are used so heavily in design. The rules just mentioned are examples of these relationships. For instance, the first rule: "IF an element with low atomic number is added THEN density will decrease", corresponds to the arrow from composition to property. \\ composition structure R__+ properties process Figure 3-1: Alloy design graph Designers seem to apply rules in an opportunistic fashion. Whenever rules are identified that will make some progress in solving the problem, those rules are applied. There are regularities to the search process, however. Rules that involve reasoning about structure are given a preference over rules that do not. Also, some property targets are viewed as being more important than others, and rules that deal with important targets are used first. Because metallurgists are asked to develop materials with properties outside of the range of existing alloys, they must reason about designs which have never been tested and whose behavior is not known with certainty. Designers try to fill this gap in existing knowledge with general models, speculation, extrapolation of known trends, and analogy with existing materials. Even after applying these methods, however, it is usually impossible to determine exactly what composition and processing specification will meet the targets. The designer identifies a family of alloys that is likely to meet the targets, and makes these alloys in the laboratory. Sometimes, one member in the family will have the desired characteristics. Often, all tests will fail to meet targets exactly, but analysis of the data results in more accurate rules that can form the basis for a better set of experiments in the next iteration. 4. Case Study of ALADIN ALADIN was designed and developed using a rapid prototype approach. During the first year, 1984, metallurgists were interviewed, the alloy design methods were characterized, a system was designed and the program written. At the end of 1984, a demonstration of the system was given to the metallurgists and an assessment of the project was made. This assessment involved technical design decisions as well as staffing levels and the approach used for knowledge acquisition. Several changes were made to the design and to the organization of the project during 1985. The design changes are now being incorporated into the software, and it is expected that the second iteration will be completed by the end of 1985. 5 4.1. First Iteration During the first iteration, many alloy designers were involved in the knowledge acquisition process. During the first two months, five people were interviewed who had done work on a variety of different alloy systems. A n attempt was made to find the common approaches used by all designers. By focusing on these, it was hoped that the system could be developed in a general way and could handle many different design strategies. After the initial interviews, a single prototype problem was identified and short term goals were established to develop a system based on the prototype. The problem selected was the first iteration in the design of aluminum lithium alloys. These alloys are now under development at Alcoa and are targeted for future aerospace applications. This choice was made for several technical and nontechnical reasons. The aluminum lithium alloys are important to Alcoa, so management support was expected to be high. Furthermore, many experts were available and anxious to help on the project. The reasoning used on the aluminum lithium project covered a broad range of empirical and theoretical ideas, so a system based on the prototype could be extended to other problems easily. Finally, since the company had only recently begun research on this alloy system, it was speculated that there were many design alternatives, not yet explored by human designers, that the ALADIN system could study. After the prototype problem was selected, a few more experts were interviewed, and a team of five metallurgists was finally selected for more detailed knowledge acquisition. During the summer and autumn of 1984, the first iteration ALADIN system was designed. Because of tight deadlines for the implementation, several simplifications were made. The solution method was modelled as a variant of generate & test [7]. This involved iteration of three steps in a fixed sequence: 0 Hypothesis-Generation - find actions to move the alloy in the direction of targets 0 Hypothesis-Selection - select the best option 0 Hypothesis-Evaluation - predict the properties of the alloy and identify deviations from target Metallurgy rules defining the relationships between comosition, processing, structure and properties provided the operators for generation and evaluation. In cases where operators were not available, regression analysis was used to identify useful trends from the alloy data base [8]. The available generation operators often dealt with property targets independently. So, the design problem was decomposed into subproblems - one for each target. The effects of subproblem interaction were not addressed, although it was known that properties were interdependent. The search space was limited to composition choices. Few microstructure and no process rules were built into the first system. Because of this decision, the system contained almost no advanced scientific knowledge. But the developers were able to test and evaluate the use of various problem solving paradigms on this problem before spending too much time learning difficult metallurgical concepts. A hierarchical planning approach was used: an abstract plan containing the types. of elements to add was developed before the details of additive percents were derived. A representation was developed for a database of alloy information and other metallurgical knowledge. The representation was based on the concept of frames. The database included known 6 alloys with information about alloying and impurity elements, physical properties, product forms and applications. Alloy families and series were also stored and linked to the individual alloys. Inheritance across links was used for default reasoning. The alloy database was used to determine standard practices and to identify useful relationships between composition and properties. Representations were also developed for important metallurgical tables and diagrams. For example, phase diagrams indicate the types of equilibrium phases that are present in an alloy for a given composition and temperature. These diagrams are used extensively by alloy designers when selecting alloying additions and processing conditions. Regions of the diagram and important characteristics of the associated phases were represented in the database. Relations between regions were defined and corresponded to transformations between phases. Two types of data elements were created to hold information about the alloy design problem - the constraint and the hypothesis. Property targets specified by the user were represented as constraints. Although in real problems, early decisions constrain choices that are available later in the search, no attempt was made to represent these dynamic constraints in the first iteration. A representation was also developed for what was called a hypothesis tree. All design choices were posted on this tree, which was developed in a depth first manner. This hypothesis tree was slightly different from the standard search graphs used for artificial intelligence problem solvers [lo]. Separate trees were developed for the abstract and the detailed plans. Links were created between nodes in the two trees to indicate dependencies between the two levels of decisions. Three types of control elements were created for the system - contexts, goals and tasks. A context corresponded to a major phase of the design problem [l]. Six contexts were identified for the first iteration system - problem-definition, search-setup, hypothesis-generation, hypothesis-selection, hypothesis-evaluation and search-termination. Contexts could have a status of active or suspended, and they were activated in a fairly rigid control sequence, as indicated in the flow chart in figure [4-11. All metallurgy rules checked for the presence of some active context as the first precondition. Hence, the context was used to decompose the system into major subsystems. Goals were used to represent actions to be performed or already attempted [l]. Goals could be arranged into fairly general and/or/all trees and were built dynamically as the domain rules identified required subtasks. In other words, any goal could have several subgoals, and the activation of subgoals would depend on the type of the supergoal. If the supergoal was an "or" type, subgoals would be activated until one succeeded, and the success of any one subogal would result in the success of the supergoal. If the supergoal was an "and" type, subgoals would be activated until one failed, and the success of all subgoals would be required for the success of the supergoal. Finally, if the supergoal was and "all" type, all subgoals would be activated, regardless of success or failure. In this case, the supergoal would always succeed. Sequence restrictions could be imposed on goals that shared the same supergoal, although these specifications were not required. Several status values were allowed for each goal, including posted (an action that should be performed in the future), active (an action that should be performed now), success (an action that was completed successfully) or failure ( an action that was attempted but not completed). All metallurgy rules checked for the presence of an active goal as the second precondition, so the goals allowed for a finer decomposition of the rule set into knowledge sources. A general set of goal management rules was written to update status values throughout the tree whenever the status of a leaf goal node was changed by the domain rules. Finally, task elements were used to represent simple actions to be performed. Unlike goals, tasks could not be linked to form a complex tree structure. 7 pro blem-definition ‘7 I search-setup ] hypothesis-generation + l n o I r 1 1 -1 hypothesis-evaluation 1 - sea rch-termination Figure 4-1: First iteration flow of control Several conventions and standards were established for the use of control and data elements. Some examples were: 0 at any time, only one context and goal should be active 0 hypotheses must designate at what level of detail they reside in the hierarchical plan Of course, when developing a system, it is always difficult to guarantee that conventions are followed, and failure to follow the practices can lead to subtle bugs that are hard to find. To make the debugging task easier, a set of diagnostic checks was formulated. These rules checked for violations from the standards and reported the errors to the developers. During the last few months of 1984, the first iteration ALADIN system was implemented. Three languages were integrated to form the system. OPS5 [6], a forward chaining production system, was used for the overall search control, the representations of goals, hypotheses, and constraints and the metallurgy rules. CRL [4], a frame-based knowledge representation system was used to store long term declarative knowledge about alloys, their composition and properties, and also metallurgical tables and diagrams. FRANZ LISP [5] was used for the interface between OPS5 and CRL, for the user interface and also for some numerical procedures. The system was implemented on a Vax 750 running under the Unix operating system. During the first few months of 1985, an assessment of the ALADIN project was made. Most of the earlier decisions were still found to be correct ones. For example, after a year of knowledge acquisition, it was still felt that an artificial intelligence-based approach was a reasonable one to use on the alloy design problem. Similarly, the choice of languages was still felt to be a good one. Even some of the design decisions were still felt to be appropriate. The goal management system, the decomposition of the rules base with contexts and goals, the generate and test approach, the idea of using some problem decomposition according to targets, and the use of hierarchical planning were retained for the second iteration. The fixed sequence of context activations was unable to respond to problems such as lack of information or failure of goals. The system had to be extended to deal with decisions about structure and process. The first iteration representation of the hypothesis tree was unable to record all of the important details about structure and process decisions. The representation of constraints also had to be generalized to deal with restrictions on composition, process and structure - decisions that could be proposed, evaluated and then withdrawn during the problem solving process. Finally, some problems were identified in the organization of the team and the approach used in knowledge acquisition. Gathering information from five metallurgists was extremely confusing and it was difficult to organize the knowledge in a coherent fashion. The primary method used to teach metallurgists about ALADIN system capabilities was to run demonstrations of the program. However, this proved to be confusing as well since at any time, most planned capabilities were not implemented and also, the user interface was not well developed. However, several problems were also identified. 4.2. Second Iteration The second iteration of the ALADIN system began after the assessment phase. A conversion was made to a newer version of the CRL language and Common LISP [l 11, running on a VAX VMS system. Then all steps were repeated - knowledge acquisition, system design and implementation. However, the approach was changed based on the first year’s experiences. Major changes were.made to the organization and staffing of the project. The aluminum lithium prototype problem was replaced by an even more specific problem that was called the training case. This case consisted of four experiments that were identified and tested during the first iteration of the aluminum lithium project. The ALADIN project goal for 1985 was to characterize and model the reasoning that was used to select those four experiments. The alloy design staff was reduced to two people, who were each allowed more time to spend on the project and given more responsibilities. One expert had a great deal of experience with the aluminum lithium training case. He was given the responsibility of helping to prepare a description of the reasoning used on the problem. One 9 computer scientist assisted him with this assignment. The metallurgist tried to describe the sequence in which ideas were considered and rejected and to characterize the rules that were used to make the decisions. The computer scientist tried to reformulate this reasoning in terms of the ALADIN design concepts. In other words, the flow of metallurgical ideas and the justifications for decisions were classified as hypothesis evaluation, hypothesis generation using search operators, backtracking, etc. New contexts and goals were identified during this process, and strategies for context switching were formulated. A written report was prepared in which the reasoning for the training case was described in detail. The report was partitioned into sections - approximately half of the sections contained descriptions using metallurgical language; the other half contained formulations of the same ideas using computing concepts. During each meeting, the report was reviewed and corrections were made before work continued. After six iterations, the report was reasonably complete. Another expert was chosen for his experience in working with computers and for his research interest in developing general models of metallurgical structure/property relationships. His main responsibilities were to help design the knowledge representation system for microstructure and processing and to help incorporate his models into the system. Another computer scientist worked closely with him to identify ways that the models could be smoothly integrated into the existing system. Written reports were again used to ensure that Me& had been understood correctly. This expert was also asked to use the system and to comment on its functionality and accuracy. Because he had more experience in using computers than many of his colleagues, it was expected that this expert would be more tolerant of the simple user interface. The design of the system was changed in four major ways. A meta space was designed that provided a more rational sequencing of contexts and goals. The hypothesis representation was generalized to deal with decisions about processing and microstructure. A least-commitment decision-making strategy, using multidimensional constraints, was incorporated into the system. Finally, several provisions were made to deal with interactions among subproblems. The meta space was responsible for controlling the activation of knowledge sources. It developed plans for how to solve the alloy design problem by building context elements and high level goal trees. Once the meta space activated a context and goal, control was transferred to the appropriate domain knowledge source, which could build subgoals, develop the hypothesis tree or update the status of its goals to success or failure. The meta space could then build new context and goal elements based on the status of previous goals and the reasons for their success or failure. The meta level rules were formulated to reproduce the design strategy that was used in the aluminum lithium training case. The hypothesis representation was also generalized so that decisions about processing and microstructure could be made as well as those about composition. This extension required that new links be invented to indicate the dependencies between decision nodes. Some of these dependencies were cause and effect relationships. For example, in order to meet a property target, the program would determine the required microstructure features and post those decisions as hypotheses. During a later iteration, the program would seek composition or processing options that were likely to produce the desired microstructure features. Another type of dependency came from multiple decisions that had to be selected together in order to achieve the desired effects. The first iteration hypothesis tree was essentially an "or" tree - all children of a single node were considered to be independent options. The second iteration representation allowed for "and" branches when 10 multiple decisions had to be grouped together and treated as a unit. The restriction of building hypotheses using only depth first search was abandoned. A breadth first search of structure space, followed by an evaluation and pruning of poor options, followed by a depth first search of composition and process spaces seemed to be a better model of the human design process. A general representation of multidimensional constraints was developed and used to implement a least-commitment decision-making strategy. Using this approach, decisions about exact amounts of additives or processing times and temperatures were postponed. Instead, the system derived linear or piecewise linear constraints in composition/ process space to define regions where properties were acceptable. As more property targets were considered, the sizes of these regions would decrease. If the initial problem was overconstrained or incorrect decisions were made during the search, conflicting constraints would be generated and detected. This would cause the system to backtrack or relax some of the property targets. On the other hand, if the system lacked sufficient knowledge of the problem (generation operators were lacking), or initial targets were too vague, the search procedure would terminate with a fairly large region remaining. In this case, the system would identify a family of points in the region and propose them for experimentation. In summary, the multidimensional constraints served four important purposes: 0 Backtracking was reduced through the use of a least-commitment decision-making strategy. 0 The interactions between search variables were represented. 0 The constraint regions provided a criteria for success or failure of a design proposal. 0 Multiple solutions (or one optimal solution) could be found for alloy design problems that lacked sufficient specification. During the first iteration, design problems were decomposed according to property targets. However, most property targets were known to interact. In other words, operators that improved one physical property would also degrade other properties. Without taking into account these interactions, there was a good chance that the system would loop indefinitely. Since the nature of some types of interactions were known and understood by metallurgists, this knowledge was incorporated into the hypothesis-generation operators. These operators fired on the existence of an active goal to improve some physical property. However, the operators also checked for the existence of posted goals related to interacting properties in their preconditions. In this way, preference was given to operators that did not adversely affect other property targets being pursued in the same design problem. Other more subtle property interactions were detected during the hypothesis-selection context. Hypothesis-selection was preceded by an heuristic evaluation of all target properties. Selection was based on a measure of the distance between target and estimates in multidimensional property space. Hence, the effects of hypotheses on other property targets not being directly pursued in the current iteration were at least considered. The second iteration design has been formulated and partially implemented. It is expected that implementation will be complete by the end of 1985. The alloy designers will then aid in another evaluation of the system. 11 During 1986, the ALADIN system will be converted to run on a Symbolics machine at ALCOA. After that, a more suitable iiscr interface will be developed so that more alloy design experts will have easy access to the system. The Carnegie-Mellon team will complete its work and an ALCOA project team will prepare for continued development of the system. Major changes to the architecture or design are not expected, but a great deal of work remains in order to build the knowledge bank to an acceptable level. Also, the rule system must be generalized to deal with a broader class of problems. Several new training cases will be identified, and the reasoning will be characterized using the procedures established for the first training case. The rule system will then be augmented to reason correctly for all training cases. It is expected that the Artificial Intelligence group will play a gradually decreasing role on the project and that the Alloy Technology Division will increase the level of responsibility. This is because future work will require an ever-increasing level of knowledge of metallurgy. 5. Conclusions Several insights were gained by our work on the ALADIN project. These insights pertain to knowledge acquisition techniques, design, architecture and implementation. With respect to knowledge acquisition, it is very important to keep the teams small and give each individual specific responsibilities. Writting a detailed script of the knowledge and reasoning used on training cases serves several purposes and will be used on other projects. The script helps to uncover errors in communication before plans proceed too far, it helps the expert to understand some of the major computing concepts used on his problem and finally, it defines requirements during system design and implementation. Some of the design ideas contained in ALADIN can also be extended to other systems. The use of control elements to decompose a production system into smaller units and to control rule firing has been suggested and used before [l]. The goal management subsystem is particularly powerful and is flexible enough to handle a variety of situations. Similarly, the idea of using multi-dimensional linear constraints to gradually converge on the solution to a problem with a reduction in the amount of backtracking can be applied to other application areas. The diagnostic checking rules have saved a great deal of debuging time and are recommended for complex systems requiring more than one developer. Finally, the rapid prototype method is a good one to use, and it is most effective when the prototype and the initial design are kept as simple as possible. 6. References 1. Brownston, L., Farrell, R., Kant, E., Martin, N., "Programming Expert Systems in OPS5 - An Introduction to Rule-Based Programming", Addison-Wesley Publishing, 1985. 2. Bryant, G.F., "Automation of Tandem Rolling Mills", 1973, The Iron and Steel Institute. 3. Buchanan, B.G., Shortliffe, E.H., "Rule-Based Expert Systems", Addison-Wesley Publishing Co, 1984. 4. Carnegie Group Inc, "Knowledge Craft Beta Release Notes", 1985. 12 5. Foderaro, J.K., "The Frant LISP Manual", University 1980. of California Technical Report, 6. Forgy, C.L., "OPS5 User's Manual", Department of Computer Science Technical Report, Carnegie-Mellon University, 1981 , CMU-CS-81-135. 7. Hayes-Roth, F., Waterman, D.A., Lenat, D.B., "Building Expert Systems" , Addison- Wesley Publishing Co, 1983. 8. Hulthage, I., Rychener, M., D., Fox, M., S., Farinacci, M., L., "The Use of Quantitative Databases in Aladin, an Alloy Design System", Robotics Institute Technical Report, Carnegie-Mellon University, 1985, CMU-RI-TR-85-19. 9. Kobayashi, S., "A Review of the Finite Element Method and Metal Forming Process Modelling", Journal of Applied Metalworking, vol2, #3, July 1982, p163-169. 10. Nilsson, N.J., "Problem Solving Methods In Artificial Intelligence", McGraw-Hill Book Company, 1971. 11. Steele, G.L., Jr., "Common LISP the Language", Digital Press, 1984. 
work_m34d4rlyzbeabool2o2ptqd6lu	----	PII: 0270-0255(87)90579-3 IDES: INFLUENCE DIAGRAM BASED EXPERT SYSTEM Alice M. Agogino & Ashutosh Rege Department of Mechanical Engineering, University of California, 5 I36 Etcheverry Hall. Berkeley. California 94720,U.S.A. Abstract. Infhrence diagrams have been used effectively in applied decision analysis to model complex ~. systems, identify probabilistic dependence and characterize the flow of information. Their graphical representation and intuitive framework are particularly effective in representing knowledge from experts with diverse backgrounds and varying degrees of technical proficiency. They allow both a symbolic representation of the system interrelationships and a quantitative measure that can be of discrete or con- tinuous functional form. By exploiting this abstraction hierarchy. successive degrees of specification can be made by several individuals. each encoding his or her expert knowledge of the problem and bounds on critical parameters. It is proposed that an interactive computer program that automates this influence diagram technology would provide an excellent tool for building expert systems. This paper describes such a modeling tool: the IDES (Influence Diagram Based Expert System) developed at the University of California at Berkeley as a modeling tool for building expert systems requiring reasoning with uncertain or incomplete information. The Diagnostician’s Problem is presented as a tutorial for describing the IDES solution procedure. Keywords. Expert systems: influence diagrams; probabilistic modeling: diagnostic reasoning: probabilis- tic inference. INTRODUCTION WHAT ARE INFLUENCE DIAGRAMS? Human beings possess an unparalleled ability for processing complex types of information and for making inferences based on this knowledge. First generation expert systems try to capture human knowledge by means of logic predi- cates or If-Then statements that involve measurable and quantifiable parameters. Consider the following example from MYCIN, a well known rule-based medical diagnostic expert system (Buchanan et al.. 1984:82). MYCIN RULE NO. 37: IF: I) The idenrirv of’rhe organism is no! known nith ccrruintr und 2) The stuln of rhc organtsm IS gamneg and 3) The morphotqw ofrhr organism is rod, and 4) The aerohiclrl, of rhr or,anism is aerobic THEN: There is .srrongI~~ .srcgg:‘.rrr ve evidence (0.8) that the class of /he orgun~sm IS mrerobacleriaceae. Following this rule. a conclusion about the class of organism present can be drawn (with a certain degree of confidence) from satisfaction of all of the antecedents. Influence diagrams were conceived of and developed by researchers in the Decision Analysis Group at SRI Interna- tional in order to automate the modeling of complex deci- sion problems involving several uncertain variables (Miller et a1.,1976). Knowledge of the interrelationships between variables is represented in a compact graphical and numeri- cal framework which identifies the critical variables and explicitly reveals any conditional independence between them. Although the original objective in the development of influence diagrams was for computer-aided modeling. they have been used effectively in improving communication among people. Use of influence diagrams in participative modeling of complex decision problems is described in Owen (1978) and Howard and Matheson (1984). A direct solution procedure to automate influence diagrams has been developed by Olmsted (1984) nd formalized in algorithmic form by Shachter (I 984a&b). The application of influence diagrams as a development tool for building expert systems has been proposed by Agogino (1985) and Holtzman ( 1985). In addition to explicit rules of Inference like those in MYCIN, human beings also have unique skills in holistic reasoning, in the use of intuition. in deciding what parame- ters or variables are important to a problem. and in know- ing what questions to ask when more information is needed, We are able to synthesize the interrelationshtps of complex problems and gam quahtative insights about them. If a computerized system is expected to capture any of this deeper human knowledge. it should at least be able to represent and process expert knowledge in both a symbolic and numeric fashion. As with the example from MYCIN. the inference mechanism and representation must be able to work with the uncertainties and imprecise data of both the material world and the human mind. The approach presented in this paper is based on the intuitive graphical framework of influence diagrams coupled with the powerful Bayesian inference engine and automation algorithms sup- porting them. In the next section. we review the theoretical framework for the IDES (Influence Diagram Based Expert System) An influence diagram can be interpreted at several levels: (I) relational, (2) functional, and (3) numerical. On the rela- tional or symbolic level the nodes in the influence diagram represent the important variables in the system being modeled and the directed arcs identify conditional influences between them. The nature of these influences is specified at the functional level and further quantified at the numerical level. Relational Level This is perhaps the most powerful level of the diagram. An influence diagram at the relational level is a graphical representation of the interrelationships between the vari- ables involved in the problem under examination. It reveals visually the flow of information. influences, and the overall structure of the problem. Three different kinds of nodes will be considered in building an influence diagram: 1) state (circular nodes), 2) decision or control (rectangular nodes). and 3) value nodes (diamond- shaped). Addition or deletion of nodes is analogous to simi- lar operations on rules in rule-based expert systems but the 227 22.3 5th ICMM effect of the same is much more apparent and intuitive with influence diagrams. State Variables Consider the following two-state variable influence diagram shown in Figure 1. The circular nodes represent two state variables, x and y, and the arc can be considered informally to signify that a conditional influence may exist. On the relational level, one interpretation of the influence diagram in Figure 1 is : “x may influence J’“. A variable x is said to influence a state variable )’ if given information about x tells you something about J’. However, this should non be interpreted to mean that x causes .r_ No sense of causality is implied. Conversely, if knowing x gives you no new information about state variable I’, then variables x and .r are said to be independenf and no influence exists between them. This strong information about the interrelationship (or rather lack of) between variables x and y is signified by the absence of an arc between the corresponding nodes as illus- trated in the influence diagram in Figure 2. It should be noted that the lack of an arc is in some ways a stronger statement of our knowledge than the existence of an arc. The presence of an arc indicates that a possible dependency exists, while the lack of an arc states that no dependency exists. In other words. an arc can be thought of as representing our state of ignorance rather than our state of knowledge. Decision Variables In addition to state variables, influence diagrams represent decision or control variables by square nodes, following the same symbolism used in decision trees. The nature of an influence arc going into a state node and that going into a control node is conceptually and mathematically distinct. Arcs leading into state nodes represent conditional injluences as discussed previously. Two examples of influence arcs are shown in Figure 3: probabilistic and fuzzy. Arcs going into decision or control nodes represent informational influences and show exactly which variables will be known by the decision maker or controller at the time that the decision is made. An arc from a decision node to a chance node indicates causality in the sense that the selection of one decision option over the other choices restricts, in general, the space or the universe of values the state variable can take on. Thus the meaning of an arc must be viewed relative to the types of nodes it connects. Control nodes are the only nodes where the directions of the arcs (to and from control nodes) are immutable aspects of the structure of an influence diagram and should not be reversed unless the intent is to reformulate the problem. Although not always shown explicitly. it is assumed that once a decision has been made. that decision is not forgot- ten. Hence the term no-jhgnring arcs is used to signify the invisible but implicit arcs between decision nodes which must always appear sequentially in time. Value Function The value jimction, signified by a diamond-shaped value node is a model of the objectives of the decision maker or computer system user. There can be at most only one value node in an influence diagram. However, as will be applied in the diagnostic expert system example. the location of the value node will vary depending on the question being asked of the expert system. Structural Matrix In a more rigorous sense, but still at the symbolic or rela- tional level, the influence diagram can be defined as an acy- clic directed graph consisting of N variables represented by N nodes and a set of directed arcs between them. This rela- tional information may be captured in a structural matrix that relates direct predecessors to each node in the diagram. A node j is called a direct predecessor of node i if and only if (itI) there is an influence arc directed from node j into node i. 0 FIG. 1. x influences J a 0 FIG. 2. x is independent of y Probabilistic o-----: 0 Fuzzy Informational CallSal 04 ‘No-Forgetting’ FIG. 3. Five Interpretations of Arcs The N xN structural matrix A for an influence diagram of N nodes (each representing a distinct state or control variable) is a Boolean matrix of elements n,, defined as follows: o’J = 1 I if the ith node is a direct predecessor of the jth node 0 if the ith node is not a direct predecessor of the jth node A node j is called a direct successor of node i iff there is an influence arc directed from node i into node j. The tran- spose of the structural matrix, A’, represents the Boolean matrix of direct successors in an influence diagram. Although this relational information may appear trivial, the existence or lack of influence and independence between variables is an important part of understanding any complex problem. Too often expert systems will assume indepen- dence at a certain level or between levels in hierarchical data structures, such as semantic nets, to simplify the data storage and computational complexity. Various forms of conditional independence were assumed in PROSPECTOR (Duda et al., 1978 and Pednault et al., 1981) and MYCIN (Buchanan et al., 1984). The philosophy behind modeling with influence diagrams, is that this structural information is important domain knowledge of the problem, and should be captured along with more in-depth rules and numerical measures. This structural information is represented in a graphical fashion that experience has shown to be intuitive to many experts regardless of their level of mathematical proficiency. A knowledge of the mathematical calculus behind influence diagrams is not needed to model at the relational level. Functional Level In what follows, we shall outline the Bayesian or probabilis- tic approach to manipulating influence diagrams. It should be pointed out that influence diagrams on the relational level do not need a probabilistic basis to justify themselves. Rather probabilistic analysis is one approach to the deeper level interpretation of an influence diagram. Although not discussed in this paper, it is possible to incorporate other logics or inference techniques into the framework of the influence diagram (e.g., fuzzy logic and first-order predicate calculus) when appropriate. IDES 229 Mathematical Definition of Influence conditional independence between the variables involved. An influence between two random variables x and y is said to exist iff (y ) x,ff) # (y 1 H), where the inferential notation (y ( x, H) represents the probability distribution for y condi- tioned on x and the history or state of information H (Owen, 1978). Arc Reversals If the functional influence of y given x in Figure 1 is proba- bilistic, the influence diagram represents one expansion of the joint probabilities of state variables x and y: (x,y 1 H). Regardless of whether x and y are probabilistically indepen- dent, the probabilistic expansion represented by the influence diagram in Figure I can be written: The framework for arc reversals was established in previous sections. We shall now study a heuristic justification of the rules for the same. Arc reversals in an influence diagram should be performed such that they are consistent with the probabilistic information the diagram represents at the numerical and functional levels. Any rules for arc reversals must be therefore justified in terms of the calculus of condi- tional probabilities. As an example, consider the arc rever- sal between nodes x and z in the influence diagrams in Fig- ures 5a and 5b. (X,Y IN) = (Y lx,H)(x IH) (1) This implies that we know both the unconditional or margi- nal probability distribution on x and the probability of y conditioned on X. Suppose, however, that our human expert or statistical data base can not directly produce the marginal probability distribution on X, but does have the conditional distribution of x given y. Suppose also that we can obtain the marginal distribution on y. Using the laws of conditional probability , we can still obtain the joint dis- tribution on x and y using the following expansion: The expansion of the joint probability distribution represented in Figure 5a is: (x.Y.~ IH) = (x lff)(y IH)(z P,Y>H) (3) As explained before, if we know the terms on the right side of (3). we can construct the joint distribution on the left. Consider now the influence diagram in Figure Sb. The expansion in this case is: (X,V IH) = (x IY,H)(Y IH) (2) This expansion is represented by the influence diagram shown in Figure 4. Although both Figures 1 and 4 are equivalent in the joint system they represent, they differ in their suitability for probability assessment. Suppose now that x and y are independent in the sense that having information about x gives no additional information about y. Then the conditional probability of x given y, (X ly,H), is equal to the marginal distribution on x, (X ( H). The expansion of the joint probability distribution, (x,y 1 H), is then the product (x I H) (y (H) corresponding to the influence diagram shown previously in Figure 2. The lack of an arc between the state variables in Figure 2 graphi- cally illustrates their probabilistic independence at the func- tional level. (X,Y,Z IH) = (Y lH)(z ly,Hl(x l~.z.Hl (4) The problem facing us now is: Using only the terms on the right-hand side of (3), can we arrive at the terms in the right-hand side of (4)? Since (y I H) is known, our problem is now reduced to expressing (z 1y.H) and (x Iy,;,H) in terms of the quantities in (3). This can be done as follows (q, is the sample space for x): (2 IY,H) = /,,(c.r ly,H)~ = Jn,(z Ix,y.ff).(x IY.H)dx- = @ lx,~.H).(x IH)dx Using Bayes’ rule: Acyclicity = (z Ix,.v,H)‘(x IH) (2 1y.H) Cycles are not permitted in influence diagrams. This is because a cyclic diagram does not represent any expansion of the joint distribution of the variables involved. On the functional level it would lead to an infinite loop in which information on a particular variable (say X) is needed to get information on x itself. On the other hand, one might try to reverse the arc directly in Figure Sa to obtain the representation in Figure 5c. Here the expansion is: Numerical Level The numerical form of the probability distributions can be of either discrete or continuous form. The numerical data base for any node, however, must be conditioned on all of the nodes with arcs directed into it. (~,Y,=IH)=(~IH~~(~I~.H)~(~Is,H~ However, it would then not be possible to obtain the condi- tional probability (x 1 z.H) from the quantities in (3) without violating the conditional independence shown in Figure 5a. Thus we see heuristically the consequence of the arc reversal is the addition of an arc from node y to node x. This brings us to the following rules of arc reversal. RULE 1 for arc reversals: An arc from state node y to state node x can be reversed iff y is not a multi-path predecessor of x . TOPOLOGICAL TRANSFORMATIONS In the functional level description, we defined the proba- bilistic interpretation of the topology of influence diagrams, Now let us consider how Bayesian inference relates to topo- logical transformations on an influence diagram. RULE 2 for arc reversals: On reversal of an arc from y to x, node y inherits the direct predecessors of x and vice-versa. Arc Addition Recall that the existence of an arc signifies that an influence may exist. Therefore an arc can always be added between nodes as long as no cycles are introduced. A cycle implies that one can condition one’s knowledge of a variable on that variable itself, resulting in an endless loop; a sort of proba- bilistic Klein bottle. The addition of an arc, however, results in a loss of (possibly vital) information regarding the The purpose of Rule 1 is to disallow any arc reversal that may cause a cycle to be introduced in the diagram. A node y is called a mulfi-pafh predecessor of node x if it has more than one path to y, e.g. see Figure 5b. Clearly, the reversal of the arc from y to x without first reversing the arc from y n n n o---o (b) FIG. 4. y influences x FIG. 5. Examples of Arc Reversals 230 5th ICMM to z in Figure 5b will cause a cycle, which as explained before, invalidates the diagram. The example cited above for Rule 2 is of course not sufficient justification for the same. However, a little thought will show that the crux of obtaining the termgin the expansion represented by the new diagram is the joint distri- bution of the variables involved (x and z in our example), conditioned on the set of nodes formed by the union of the sets of their direct predecessors. In other words, this set of nodes will always precede the variables involved in the reversal in the hierarchy of the expansion. Hence they will always occur as conditioned on the aforementioned set of predecessors. A formal proof for the same can be found in Olmsted (I 984:Chapter 2) and Shachter (I 984b: 14- 16). Furthermore, it should be noted that arcs to and from deci- sion nodes cannot be reversed. This arises from the infor- mational nature of the arc to a decision node and the causal nature of the arcs from decision nodes to state nodes. Reversal of the former would violate chronological pre- cedence and that of the latter, causality. Barren Node Removal Although strictly speaking barren node removal need not form an integral part of the transformations to the influence diagram, we will refer to this operation since it is a con- venient technique to visualise the flow of information graph- ically. A node is said to be barren if it is not a member of the set of total predecessors of the node designated as the value node. In other words, the node is not concerned with the inference problem in hand. Furthermore, a node can become barren due to arc reversals. Such nodes (as a result of the topological transformations of the diagram) can be removed, and the diagram redrawn without changing the solution. Repeated application of this procedure leads to a shrinking infl.uence diagram concluding in a representation of the solution to the inference or decision problem under consideration. Details of the procedures for solving influence diagrams will be given in the following section. In expert system applications, barren node removal can be used as a mechanism to delete nodes and influences that are irrelevant to the specific question that is being asked of the system. The question and topological transformations needed to answer it. create what will be referred to as the currenl slruclure of the influence diagram. SOLVING INFLUENCE DIAGRAMS Two kinds of uses of influence diagrams will be described herein: I) probabilistic inference and 2) decision making under uncertainty. In either case there must be some ques- tion to be answered as represented by the diamond-shaped value node. Probabilistic Inference Probabilistic inference is equivalent to determining the pro- bability distribution (G(X)) Y.H) where X and Y are sets of nodes given in the diagram. The function G(X) is analogous to a value function and is called the goal-value function for probabilistic inference. This new goal-value node G(X) is added to the diagram with conditioning arcs from the set X. Because an influence diagram can contain at most one value node at a time, all previous goal-value nodes are converted to state nodes. The procedure involves manipulating the influence diagram to obtain a representation of the problem under consideration. The relevant arcs are reversed until the resulting diagram includes the goal: the desired represen- tation (G(X) 1 Y,H). Decision Analysis The transformations or operations involved in decision analysis are the same as that for probabilistic inference; that is arc reversals and barren node removal. However the difference is that here instead of determining a probability distribution. we are computing the expected value (or util- ity) of a sequence of decisions, comparing them and assess- ing the optimal sequence by maximizing the expected value of the utility function represented by the value node. The topological or symbolic solution procedure is equivalent to ordering the nodes in the correct sequence(s) of integration of state nodes or optimization over decision nodes. As pointed out by Shachter (1984a:l4) the steps involved at the numerical level are those used in evaluating a stochastic dynamic program. Note that the order so determined would correspond to one of the orderings that might be used in a decision tree based technique in which the variables in an expanded decision tree are either integrated or optimized over. The principle of dynamic programming allows pruning of suboptimal branches in this implicit decision tree. However. the influence diagram algorithm allows solution without the need of an intermediate tree representation. AI1 intermedi- ate calculations are performed directly from the influence diagram knowledge base and thus retain their original mean- ing. Node Removal Concept An equivalent way of looking at the above procedure is to consider node removal as a topological transformation that corresponds to manipulations of probabilistic functions. Nodes are removed in the order cf integration/optimization described above. Rules can be formulated that allow one to develop a topologicat solution to the problem, without expli- citly performing functional or numerical evaluations. If the topological transformations are to be useful, these rules need to have a basis in the probabilistic analysis of the problem. Given an acyclic influence diagram the optimal sequence of decisions can be found by ordering or removing nodes from the diagram according to the following rules: RULE 1 for State Node Removal: A state node i can be absorbed into the value node iff it directly precedes the value node and no other node. On its removal the value node inherits all the direct predecessors of i. Absorption of the node is equivalent to integration over the corresponding state variable, i.e., conditional expecta- tion. RULE 2 for Decision Node Removal: A decision node i can be removed from the influence diagram iff it directly precedes the value node and directly succeeds ail other direct predecessors of the value node. Removal of the node is then equivalent to maximization of the (condi- tioned) expected utility. Therefore, combining these rules with appropriate arc rever- sals, the topological solution procedure can be determined. The justification of the rules can be found in the literature on decision making under uncertainty (see Raiffa, 1970 or Bunn, 1984). COMPUTER ASSEMBLY DIAGNOSTICS TUTORIAL As an illustration of the use of influence diagrams, we present the Diagnostician’s Problem in the context of the automated assembly of microcomputers. In order to sim- plify the presentation, we will assume that during the final testing of a microcomputer, a system failure can be traced to failures in either of two components: (1) the logic board or (2) the I/O board. Let us further assume that a sensor is available that gives rudimentary information about the operating status of the assembled microcomputer and is a function of the operating status of the logic board and the I/O board. The associated influence diagram is shown in Figure 6a. This represents the Diagnostician’s Problem on the relational level. In the present structure of the influence diagram, the joint probability can be obtained from the conditional probablil- ity of the system status 3” and the marginal distributions on the state of the logic board “L” and the I/O Board “I/O” as given in the expansion below. (S,L,IIO IH) = (S I L,IIO,H)(L I H)(IIO 1 H) Note that L and I/O are conditionally independent (there is no arc between them) and thus their joint probability distri- bution is the product of each marginal distribution. IDES 231 FIG. 6. Probabilistic Injluence Diagram FIG. 7. Solvejor (L 1 E.H) The nature of the influences is specified at the functional level. The influence diagram in Figure 6a implies that the conditional distribution on S is known along with the margi- nal distributions on L and I/O. Let us assume for illustra- tion that the test for system status gives a deferministic result based on the status of the I/O and logic boards: if any of these boards (or both) is in a failure state the system status will show a failed system state. Using a subscript of “0” for a failed state and “1” for the operational state, this implies on the numerical level that the conditional distrihu- tion of the system status, S is: I for S, given L, and I/O, I for So given L, and I/O0 (S 1 L,IIO,H) = 1 for SO given Lo and I/O, (5) I for SO given Lo and I/O, 0 elsewhere Let us further assume that the marginal probability of failure is 5% for the logic board and 1% for the I/O board, giving the marginal probability distributions: (I’OIH) = 1 0.99 for I/O = I/O, 0.01 forl/O=l/O~ I 0.95 for L = L, (L 1 H) = 0.05 forL =LO (6) Probabilistic Reasoning The diagram m Figure 6a represents the diagnostician’s knowledge of the problem and system interrelationships. In diagnostic reasoning. however. we are more apt to want to know the cause of the failure given information about the system status. One question that might be asked is: What is the likelihood that the logic board has failed given a system failure has occurred? In this case, the evidence E = S = S,,. the event that there is a system failure. The joint probabil- ity distribution on L and f/O given the evidence can be obtained mathematically by use of conditional probability: (7) Equation (7) is summed (or integrated for continuous proba- bility distributions) over all possible states of the I/O board to find the conditional probability of the logic board status given a system failure, (L / E = So.H): I (L IE,H) = ~(L.llO 1E.H) w 0 where R,,o is the sample space for the possible states of the II0 board. This is an example of probabilistic reasoning. On the rela- tional level of the associated influence diagram. equation (7) corresponds to a sequence of arc reversals leading to the influence diagrams shown in Figure 7a. The logic board node, L, is overlayed with a diamond to signify that this is the goal of our probabilistic reasoning. If the evidence E is known with certainty the probabilistic node S is now deter- ministic and updated with the new information on the sys- tem status. Because S is no longer probabilistic, we are only interested in events that are conditioned on this new infor- mation, i.e., we want to reverse all arcs leading into S. In Figure 7a the arc between L and S is reversed and node L inherits all of the direct predecessors of S, which in this case is the I/O node. Next the arc between I/O and S is reversed in Figure 7b without the addition of any new arcs. Finally the I/O node is absorbed in Figure 7c by summing over the states of the I/O board according to Rule 1 for state node removal. In numerical terms, the topological transformation shown in Figure 7a can be considered a combination of two steps as shown in Figure 6h & 6c: (1) Aggregation of the two nodes on each side of the arc to be reversed. Aggregation corresponds to calculating the joint probability distribution of the nodes condi- tioned on the union of the predecessors of both. In this case, aggregation means finding the joint probability dis- tribution of S and L conditioned on I/O (Fig. 6b). (2) Disaggregalion of the joint distribution by condition- ing it on the node that the new arc will be directed away from. In this case, the joint distribution should be con- ditioned on S given I/O (Fig. 6~). The denominator in the disaggregation step is obtained by summing (or integrating for continuous distributions) the joint distri- bution from the aggregation step over the outcome space of the node the new arc will be directed towards (in this case, L). Although shown graphically in Figure 6 as separate steps, all information in the original diagram must be stored until the arc reversal transformation is complete. At the numerical level, the aggregation step corresponds to the following expansion using equations (5) and (6): {L,SIIIO,H) = {SIL,IIO.H)~{L lll0.H) (8) = (S(L.IIO,H)‘(L \H) 0.95 for S, and L I given I/O, 0.95 for SO and L , given I /Oa = 0.05 for So and Lo given I/O, 0.05 for SO and Lo given I/O0 0 elsewhere The disaggregation step involves conditioning this joint dis- tribution on the conditional distribution on S. The numerator in equation (9) is the joint probability distri- bution in equation (8). The denominator (S (II0.H) is obtained by “integrating” the joint probability (L,S ) IIO,H) from equation (8) over the state space of L. Substituting in the numerical values from equations (6) and (8) into equation (9) gives the conditional distribution on L : I for L, given S, and I/O, 0.95 for L, given So and I/O, (L ) IIO,S,H) = I for Lo given So and I/O, 0.05 for L, given S0 and I/O,, 0 elsewhere (10) The influence diagram in Figure 7b represents an arc rever- sal between S and I/O. The new conditional probability (II0 1S.H) is obtained from (S ll/O,H) and (I/O ( H) as represented in Figure 6c or 7a. This can also be thought of as aggregating nodes II0 and S and then disaggregating by 232 5th ICMH conditioning the joint distribution on S. = (S.IIO I H) (S(H) _- Figure 7c represents the conditional probability distribution (L 1.7, H) and the marginal distribution (S 1 H). (L IS,H) = z(L l.S,~IO,H)(~IO 1.S.H) Q,0 On a topological level, summation or integration corresponds to absorption of nodes. In this example, after the arc reversals shown in Figure 7a & 7b are performed, the I/O node is absorbed leaving the influence diagram in Figure 7c. Suppose that we are given that a system failure has occurred i.e., S = St,. Let us define this as event E = S = So. Thus the likelihood that the logic board is the cause of the failure may be quantified by (L =LO ) E =S =S,,, H). From equa- tions (IO) and (I 1): (L=L,,JE-S=S,,H) = (LoIE,I/Oo,H).(lIOo(E,H) +(LoIE,IIOI.H)~(IIO, 1E.H) = (0.05)~(0.1681) + (1).(0.8319) = 0.8403 Decision Problem Let us now look at a decision problem. The Diagnostician’s Problem is to decide which component should be tested for failure first and repaired if appropriate. In an automated assembly operation, this could mean deciding where the computer should be sent for “rework”. If the diagnostician is wrong in his or her decision for rework, valuable time would have been wasted by the rework technician and hence in producing the final product. For purposes of illustration, let us assume that the costs of rework, which involves open- ing the computer and pulling out the appropriate board, testing, and repairing it, is a constant for each board. The inguence diagram in Figure 8 has been modified to show these decision and value nodes. Recall that there is an implicit arc between all decision nodes and the value node. This “No-Forgetting Arc” has been added to the influence diagram in Figure 8. It has been assumed that a system failure has occurred and this information is known at the time that the rework decision is made. The diagnostician has two choices given that the system status shows a failure (E =S =.S,): (1) DL = Send the logic board to rework first and (2) D,,* = Send the l/O board to rework first Each of the two decisions has three possible outcomes corresponding to the possible states of the boards given that a system failure has occurred S = So: (1) LdlOo, (2) L ,1/O,, and (3) &J/O,. If the wrong board is sent to rework, then debug time has been wasted on that board and the other board must be sent to be debugged and reworked. It is also possible that the board sent to rework has failed but the other board has failed also and must be repaired. The two decisions, possible outcomes, and cost consequences are shown in the decision tree in Figure 9. From the influence diagram in Figure 8 and the decision tree in Figure 9, we see that the rework decision influences the value node. If this were not the case the decision would be irrelevant (a “worry” over which the decision maker has no control). Furthermore, the outcome of the rework deci- sion also influences the cost. This is due to the fact that though we pick one of the two boards (say I/O) for rework, it could well turn out that the other board (L in this case) has really failed. In such a case, the cost of debugging and reworking the second board would have to be added to the original expenditure. Given that a system failure has occurred. which board should be tested first? Diagnostician’s Problem Solved Applying the rules for influence diagram manipulation, dis- cussed previously, we get the topological or symbolic solu- FIG. 8. Diagnostician S Problem peckzion_ Results “R” CostValue l&O1 $Lkbw(L) + WepaW) $Debug(L+I/O) + $Repair(VO) $Debug(L+VO) + $Repak(L+VO) $Debug(l/O+L) + $Aepair(L) $Debug(VO)+ $Repair(llO) $Debug(l/O+L) + $Repair(l/O+L) FIG. 9. Decision Tree D c v L 110 S (8) D C v I/O s (c) D C w s (e) 0 C D C v I/O W S B-4 (1) Optimal Expected Utility FIG. 10. Topological Solution of the Diagnostician S Problem. tion as the following sequence of node removals: R , L, I/O, S, and D. The modified intluence diagram in the above . stages of node removal is shown in Figure 10. A numerical example of the solved diagnostician’s problem can be found in Agogino and Rege (1985:30-37). Value of Imperfect Information (Value of Testing) A generalization of the Diagnostician’s Problem is the addi- tion of another test or sensor that gives more information IDES 233 lNrERRcGAx)Fy INTERFACE SOL”T72Y%%_E ? I FIG. 1 I. IDES Structure about the state of the boards without having to go through expensive debugging procedures. Unfortunately the test is imperfect and the diagnostician must still make a decision under uncertainty after viewing the results of this imperfect test. To make matters worse, this test is not free and thus the costs of the additional test must be weighed against its benefits in reducing the uncertainty in the decision. The decrease in expected value due to the additional sensor is called the expected value of additional testing. The influence diagram solution procedure remains !he same, except the optimal result will now include two decisions: (I) Dr = testing decision and (2) DR = rework decision. AUTOMATION OF INFLUENCE DIAGRAMS The IDES (Influence Diagram Based Expert System) is an expert system development tool based on the automation of influence diagram technology described previously. Because the underlying interpreter is written entirely in the compiled language C, IDES is transportable to a large number of mainframe and microcomputer systems. The IDES system provides several levels of automation and inference from the graphical influence diagram representation: (1) Bayesian inference, (2) optimization, and (3) sensitivity analysis. Principles of dynamic programming and influence diagram logic allow efficient numerical optimization of the decision or control problem without ever having to consider inter- mediate representations such as fault trees, simulation models, or decision trees. An upper bound on the value of additional information or testing can be evaluated through use of IDES’s unique features for simulation and sensitivity analysis. Figure 11 shows a schematic_outline of the various modules of IDES and their interaction with each other. SUMMARY As illustrated in the Diagnostician’s Problem, influence diagrams can be applied to expert systems requiring both reasoning under uncertainty (probabilistic inference) and advising (or controlling at the machine level). The lnterro- gator module or the IDES (Influence Diagram Based Expert System) elicits information from the expert/user on the topological structure of the influence diagram which represents a model of the problem under consideration as well as the functional relationships and numerical probabili- ties associated with it. The topological solution module in IDES finds a symbolic solution without any numerical com- putauons. The topological solution is equivalent to ordering the state and decision variables in a sequence same as that of the corresponding decision tree. Once the order in which to integrate chance nodes and optimize over decision nodes is known, the actual computations are relatively straightfor- ward and calculated in the numerical computational module of the IDES program. Because the control structure is independent of the knowledge base, new knowledge can be added or subtracted from the data base without adjusting the basic Bayesian control structure. The sensitivity analysis module can be used to add new evidence and test failure hypotheses in real-time diagnostic applications in process control and automated manufacturing environ- ments. REFERENCES Agogino, A.M. and A. Rege (1985). A Tutorial on influence diagrams. Working Paper #85-0702-2, Expert Systems Laboratory, Department of Mechanical Engineering, University of California. Berkeley 94720. Agogino, A.M. (1985). Use of probabilistic inlluence in diagnostic expert systems, Proceedings of the 1985 ASME International Computers in Mechanical Engineering, Boston, MA, Aug. 4-8, Vol. 2, pp. 305- 310. Buchanan, B.G. and E.H. Shortliffe (1984). Rule-Based Expert Sysfems, Addison-Wesley Pub. Co., Menlo Park. CA. Bunn, D.G. (1984). Applied Decision Anal,vsis. McGraw- Hill. New York. N.Y. Duda, R.i)., P. E. I&t, and N. J.,Nilsson, (1978). Semantic network representations in rule-based inference sys- tems. Pattern Directed Inference Systems, Waterman, D.A. & F. Hayes-Roth, eds., pp. 1075-1082, Academic Press, London LTD. Holtzman, S. (1985). Intelligent decision systems. Ph.D. Dissertation, Department of Engineering-Economic Systems, Stanford University. Howard, R.H. and J.E. Matheson (1984). Influence diagrams. Readings on the Principles and Applications of Decision Analysis, Strategic Decisions Group, Menlo Park, CA. Miller, A.C., M.M. Merkhofer and R.A. Howard (1976). Development of automated aids for decision analysis. Stanford Research Institute, Menlo Park, CA. Olmsted, S.M. (1984). On representing and solving decision problems. Ph.D. Dissertation, Engineering-Economic Systems Department, Stanford University. Owen, D.L. (1978). The use of influence diagrams in struc- turing complex decision problems. Proceedings, Second Lawrence Symposium on Systems and Decision Sciences, Oct. 3-4. Also in Bunn. D.W. (1984:2 12). Pednault, E.P.D., S.W. Zuker, and L.V. Muresan (1981). On the independence assumption underlying subjective Bayesian updating. Artificial Intelligence, Vol. 16, pp. 2 13-222. Raiffa, H. (1970). Decision Anai.vsis: Introductoql Lectures on Choices Under Uncertainty, Addison-Wesley Pub- lishing Company, Menlo Park, CA. Shachter, R.D. (1984a). Evaluating influence diagrams, Department of Engineering-Economic Systems, Stan- ford University. Shachter, R.D. (1984b). Automating probabilistic inference. Department of Engineering-Economic Systems, Stan- ford Univ., presented at the Nineteenth Actuarial Research Conference, University of California, Berke- ley. 
work_m3aeltcz3rgqhipsknajgheqm4	----	From data mining to knowledge mining: Application to intelligent agents From data mining to knowledge mining: Application to intelligent agents Amine Chemchem ⇑, Habiba Drias * USTHB-LRIA, BP 32 El Alia Bab Ezzouar, Algiers, Algeria a r t i c l e i n f o Article history: Available online 10 September 2014 Keywords: Knowledge mining Induction rules Classification Clustering Cognitive agent a b s t r a c t The last decade, the computers world became a huge wave of data. Data mining tasks were invoked to tackle this problem in order to extract the interesting knowledge. The recent emergence of some data mining techniques provide also many interesting induction rules. So, it is judicious now to process these induction rules in order to extract some new strong patterns called meta-rules. This work explores this concept by proposing a new support for induction rules clustering and classification. The approach invokes k-means and k-nn algorithms to mine induction rules using new designed similarity measures and gravity center computation. The developed module have been implemented in the core of the cognitive agent, in order to speed up its reasoning. This new architecture called the Miner Intelligent Agent (MIA) is tested and evaluated on four public benchmarks that contain 25,000 rules, and finally it is compared to the classical one. As foreseeable, the MIA outperforms clearly the classical cognitive agent performances. � 2014 Elsevier Ltd. All rights reserved. 1. Introduction Nowadays, the induction rules have become inseparable pattern of the artificial intelligence thanks to their existence as the basis for many disciplines, such as the agent technology, data mining and knowledge discovery. . .This paper is about how to extend data mining techniques to induction rules in order to extract meta-rules. There are many data mining tasks for instance: clustering, classification, association rules mining, regression, pre- diction,. . .We are interested through this work in the first two tasks which are used on many applications (the image processing, the intrusion detection,. . .etc) and can be solved by different algorithms (k-means, HCA, fuzzy c-means. . .for clustering, KNN, SVM,ID3,. . .for classification). K-nn and K-means are in the top ten of data mining algorithms (Wu et al., 2008). The latter are extended to induction rules by introducing new version of similar- ity measure and gravity center computation. The algorithms called K-NN-IR and K-means-IR are developed and demonstrated on a public large scale benchmark including 25,000 induction rules. The whole idea behind this work is to improve the reasoning process by integrating the knowledge mining module in today’s intelligent agent in order to speed up the reasoning engine process. The rest of this paper is organized as follows: Next section shows a short history of data mining. Section 2 summarizes related works. In Section 3: induction rules representation are presented, followed by proposing mathematical preliminaries. In Section 5, the suggested algorithms are described and followed by the defini- tion of a new architecture for intelligent agent. Then, experimental results are shown in Section 7 compared to the previously proposed algorithms. Finally we conclude by making some remarks and talking about future works. 2. Data mining overview The generation of models from a large number of data is not a recent phenomenon. Egypt Pharaoh Amasis organizing the census of the population in the fifth century BC Rocchi (Rocchi, 2003). This is the seventeenth century we begin to analyze the data to find common characteristics. In 1662, John Graunt published his book ‘‘Natural and Political Observations Made upon the Bills of Mortal- ity’’ in which he analyzed the mortality in London and trying to predict the appearances of the bubonic plague. In 1763, Thomas Bayes shows that we can determinate not only probabilities from observations derived from experience, but also the parameters for these probabilities. Legendre published in 1805 an essay on the least squares method for comparing a set of data with a math- ematical model. From 1919 to 1925, Ronald Fisher develops the analysis of variance as a tool for its proposed medical statistical http://dx.doi.org/10.1016/j.eswa.2014.08.024 0957-4174/� 2014 Elsevier Ltd. All rights reserved. ⇑ Corresponding authors. E-mail addresses: aminechemchem@gmail.com (A. Chemchem), hdrias@usthb. dz (H. Drias). Expert Systems with Applications 42 (2015) 1436–1445 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.08.024&domain=pdf http://dx.doi.org/10.1016/j.eswa.2014.08.024 mailto:aminechemchem@gmail.com mailto:hdrias@usthb.dz mailto:hdrias@usthb.dz http://dx.doi.org/10.1016/j.eswa.2014.08.024 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa inference. The 1950s saw the advent of computer technology and computer calculation. Same methods and techniques are emerging such as segmentation, neural networks and genetic algorithms, and then in the 1960s, the decision tree, the method of mobile centers, these techniques allow researchers to exploit and discover models more accurate. The advent of the microcomputer stimulates research and statistical analyzes are more numerous and precise. The term ‘‘data mining’’ had a negative connotation in the early 1960s, expressing contempt for statisticians research approaches without correlation assumptions. it fell into oblivion, and Rakesh Agrawal employed again in the 80s when they were beginning research on databases with a volume of 1 Mb. The concept of data mining makes its appearance – according Pal (2007) – when the IJCAI1 conferences took place in 1989. Then, in the 1990s, came the machine learning techniques such as SVM in 1998, complement- ing the tools of the data analysis. At the turn of the century, a company like Amazon uses these tools to offer our customers products that may interest. Actually, There are many tasks of data mining such as: Supervised and unsupervised classification, association rule mining, prediction and regression. 2.1. Supervised classification Classification of a collection consists of dividing the items that make up the collection into categories or classes (Kotsiantis, 2007; Jain, Murty, & Flynn, 1999). In the context of data mining, classification is done using a model that is built on historical data. The goal of predictive classification is to accurately predict the tar- get class for each record in new data, that is, data that is not in the historical data. A classification task begins with build data (also known as training data) for which the target values (or class assignments) are known. Different classification algorithms use different techniques for finding relations between the predictor attribute’s values and the target attribute’s values in the build data. K Nearest Neighbor (K-NN from short) is one of those algorithms that are very simple to understand, furthermore, it works incredi- bly well in practice, especially in the anomaly detection domain like Liao and Vemuri (2002), also for text categorization like in the work Guo, Wang, Bell, Bi, and Greer (2006). Also it is surpris- ingly versatile and its applications range from vision to proteins to computational geometry to graphs and so on. With KNN algo- rithm, we can obtain a satisfactory results, in addition, its basic principle is very simple, and easy to implement. It also might sur- prise many to know that K-NN is one of the top 10 data mining algorithms. K-NN is an non parametric learning algorithm, it is used when the data set does not obey a defined function as (gaussian mixtures, linearly separable etc). K-NN algorithm can be explained as follows, in the first time, training data that are already classified are considered, and then to classify the new data, we have to compute the similarities distance between this new data and all training data. After that the k nearest neighbors are extracted. In the end the new data is assigned to the most frequent class of these neighbors. 2.2. Clustering data technique Clustering data mechanism consist to put the homogeneous data into the same group or class in order to dispatch the hetero- geneous data into different groups. In the literature, it exists differ- ent manner to group the data, the two principals are: the hierarchical and the partitioning clustering. For the hierarchical clustering, the clusters are inside each others. This category of clustering is used when data can be separated in different levels. Also, CHA is the most known hierarchical algorithm, it starts by putting each instance in one cluster after that it computes the dis- similarities for all two instances to combine the clusters that have the lower distance. This process is repeated until we get one cluster (Steinbach, Ertöz, & Kumar, 2004; Han, Kamber, & Pei, 2006). In the contrary of the partitioning clustering, it consists to cluster the data separately. K-means is one of the simplest pure partitioning learning algorithms that solves the well known clustering problem (Han et al., 2006; MacQueen et al., 1967). The procedure follows a simple and easy way to classify a given data set through a certain number of clusters (assume k clusters) fixed initially. The main idea is to define k gravity centers, one for each cluster. The cen- troids should be placed in a cunning way because the clustering result depends on their location in the clusters. In order to optimize the efficacy of the outcomes, it is judicious to place them as much as possible far away from each other. The next step is to take each point belonging to a given data set and associate it to the nearest centroid. When no point is pending, the first step is completed and an early grouping is done. At this point we need to recalculate k news centroids of the clusters resulting from the previous step, and iterates the process. The latter stops when no more changes of the clusters are observed, in other words when the centroids do not move any more. 3. Related works Our interest in this study revolves around two main subjects: scalable cognitive agent and knowledge mining in general which involves induction rules mining. As for first subject, we found very few papers with ideas about the notion of scalable cognitive agent like Cao, Gorodetsky, and Mitkas (2009), and nothing about the paradigm that we would like to cover in this article. However, we notice that biologists and psychologists are showing interest in the study of scalable brain (Eliasmith, 2013). What can be said about the second topic is that the literature offers a large spectrum of detailed research on knowledge Mining. Mining knowledge including simple data and other patterns have been examined intensively over the last decade. In the following, we will talk about some knowledge mining. Many works are about mining association rules in order to obtain meta rules whose purpose is to reduce the large number of discovered rules. The CLOSET algorithm was proposed in Strehl, Gupta, and Ghosh (1999) as a new efficient method for mining closed itemsets. CLOSET uses a novel frequent pattern tree (FP-tree) structure, which is a compressed representation of all the transactions in the database. Moreover, it uses a recursive divide-and-conquer and database projection approach to mine long patterns. Another solution for the reduction of the number is introduced by Hahsler and Chelluboina (2011) used an item- set-tid set search tree and pursued with the aim of generating a small non redundant rule set. To this goal, the authors first found minimal generator for closed itemsets, and then, they generated non redundant association rules using two closed itemsets. A new algorithm to group rules via hierarchical clustering has been developed in Berrado and Runger (2007) to visualize the large number of rules. The clustering of rules is done by defining a new distance called dJaccard that represents the number of items of the two rules divided by the number of unique items. Saneifar, Bringay, Laurent, and Teisseire (2008) were interested in discover- ing sets of data. In their paper, they have developed a new similar- ity measure between two rules and extended k-means algorithm to cluster them. In literature some works about induction rules anal- ysis have been proposed: In Poongothai and Sathiyabama (2012b), eh authors have developed a new algorithm to select the interest- ing induction rules from all the discovered rules in web mining1 International Joint Conference on Artificial Intelligence. A. Chemchem, H. Drias / Expert Systems with Applications 42 (2015) 1436–1445 1437 https://isiarticles.com/article/46043 
work_m4gttcplwneqteedgvssjo6oy4	----	An expert system for detecting automobile insurance fraud using social network analysis An expert system for detecting automobile insurance fraud using social network analysis Lovro Šubelj *, Štefan Furlan, Marko Bajec Faculty of Computer and Information Science, University of Ljubljana, Tržaška 25, SI-1001 Ljubljana, Slovenia a r t i c l e i n f o Keywords: Fraud detection Automobile insurance Social network analysis Link analysis Assessment propagation a b s t r a c t The article proposes an expert system for detection, and subsequent investigation, of groups of collaborating automobile insurance fraudsters. The system is described and examined in great detail, several technical difficulties in detecting fraud are also considered, for it to be applicable in practice. Opposed to many other approaches, the system uses networks for representation of data. Networks are the most natural representation of such a relational domain, allowing formulation and analysis of complex relations between entities. Fraudulent entities are found by employing a novel assessment algo- rithm, Iterative Assessment Algorithm (IAA), also presented in the article. Besides intrinsic attributes of entities, the algorithm explores also the relations between entities. The prototype was evaluated and rig- orously analyzed on real world data. Results show that automobile insurance fraud can be efficiently detected with the proposed system and that appropriate data representation is vital. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction Fraud is encountered in a variety of domains. It comes in all dif- ferent shapes and sizes, from traditional fraud, e.g. (simple) tax cheating, to more sophisticated, where entire groups of individuals are collaborating in order to commit fraud. Such groups can be found in the automobile insurance domain. Here fraudsters stage traffic accidents and issue fake insurance claims to gain (unjustified) funds from their general or vehicle insurance. There are also cases where an accident has never oc- curred, and the vehicles have only been placed onto the road. Still, the majority of such fraud is not planned (opportunistic fraud) – an individual only seizes the opportunity arising from the accident and issues exaggerated insurance claims or claims for past damages. Staged accidents have several common characteristics. They oc- cur in late hours and non-urban areas in order to reduce the prob- ability of witnesses. Drivers are usually younger males, there are many passengers in the vehicles, but never children or elders. The police is always called to the scene to make the subsequent acquisition of means easier. It is also not uncommon that all of the participants have multiple (serious) injuries, when there is al- most no damage on the vehicles. Many other suspicious character- istics exist, not mentioned here. The insurance companies place the most interest in organized groups of fraudsters consisting of drivers, chiropractors, garage mechanics, lawyers, police officers, insurance workers and others. Such groups represent the majority of revenue leakage. Most of the analyses agree that approximately 20% of all insurance claims are in some way fraudulent (various resources). But most of these claims go unnoticed, as fraud investigation is usually done by hand by the domain expert or investigator and is only rarely computer supported. Inappropriate representation of data is also common, making the detection of groups of fraudsters extremely difficult. An expert system approach is thus needed. Jensen (1997) has observed several technical difficulties in detecting fraud (various domains). Most hold for (automobile) insurance fraud as well. Firstly, only a small portion of accidents or participants is fraudulent (skewed class distribution) making them extremely difficult to detect. Next, there is a severe lack of la- beled data sets as labeling is expensive and time consuming. Be- sides, due to sensitivity of the domain, there is even a lack of unlabeled data sets. Any approach for detecting such fraud should thus be founded on moderate resources (data sets) in order to be applicable in practice. Fraudsters are very innovative and new types of fraud emerge constantly. Hence, the approach must also be highly adaptable, detecting new types of fraud as soon as they are noticed. Lastly, it holds that fully autonomous detection of automobile insurance fraud is not possible in practice. Final assess- ment of potential fraud can only be made by the domain expert or investigator, who also determines further actions in resolving it. The approach should also support this investigation process. Due to everything mentioned above, the set of approaches for detecting such fraud is extremely limited. We propose a novel ex- pert system approach for detection and subsequent investigation 0957-4174/$ - see front matter � 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.07.143 * Corresponding author. Tel.: +386 1 4768 186. E-mail addresses: lovro.subelj@fri.uni-lj.si (L. Šubelj), stefan.furlan@fri.uni-lj.si (Š. Furlan), marko.bajec@fri.uni-lj.si (M. Bajec). Expert Systems with Applications 38 (2011) 1039–1052 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2010.07.143 mailto:lovro.subelj@fri.uni-lj.si mailto:stefan.furlan@fri.uni-lj.si mailto:marko.bajec@fri.uni-lj.si http://dx.doi.org/10.1016/j.eswa.2010.07.143 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa of automobile insurance fraud. The system is focused on detection of groups of collaborating fraudsters, and their connecting acci- dents (non-opportunistic fraud), and not some isolated fraudulent entities. The latter should be done independently for each particu- lar entity, while in our system, the entities are assessed in a way that considers also the relations between them. This is done with appropriate representation of the domain – networks. Networks are the most natural representation of any relational domain, allowing formulation of complex relations between enti- ties. They also present the main advantage of our system against other approaches that use a standard flat data form. As collaborat- ing fraudsters are usually related to each other in various ways, detection of groups of fraudsters is only possible with appropriate representation of data. Networks also provide clear visualization of the assessment, crucial for the subsequent investigation process. The system assesses the entities using a novel Iterative Assess- ment Algorithm (IAA algorithm), presented in this article. No learn- ing from initial labeled data set is done, the system rather allows simple incorporation of the domain knowledge. This makes it applicable in practice and allows detection of new types of fraud as soon as they are encountered. The system can be used with poor data sets, which is often the case in practice. To simulate realistic conditions, the discussion in the article and evaluation with the prototype system relies only on the data and entities found in the police record of the accident (main entities are participant, vehicle, collision,1 police officer). The article makes an in depth description, evaluation and anal- ysis of the proposed system. We pursue the hypothesis that auto- mobile insurance fraud can be detected with such a system and that proper data representation is vital. Main contributions of our work are: (1) a novel expert system approach for the detection of automobile insurance fraud with networks; (2) a benchmarking study, as no expert system approach for detection of groups of automobile insurance fraudsters has yet been reported (to our knowledge); (3) an algorithm for assessment of entities in a rela- tional domain, demanding no labeled data set (IAA algorithm); and (4) a framework for detection of groups of fraudsters with net- works (applicable in other relational domains). The rest of the article is organized as follows. In Section 2 we discuss related work and emphasize weaknesses of other proposed approaches. Section 3 presents formal grounds of (social) net- works. Next, in Section 4, we introduce the proposed expert system for detecting automobile insurance fraud. The prototype system was evaluated and rigorously analyzed on real world data, descrip- tion of the data set and obtained results are given in Section 5. Dis- cussion of the results is conducted in Section 6, followed by the conclusion in Section 7. 2. Related work Our work places in the wide field of fraud detection. Fraud ap- pears in many domains including telecommunications, banking, medicine, e-commerce, general and automobile insurance. Thus a number of expert system approaches for preventing, detecting and investigating fraud have been developed in the past. Re- searches have proposed using some standard methods of data min- ing and machine learning, neural networks, fuzzy logic, genetic algorithms, support vector machines, (logistic) regression, consoli- dated (classification) trees, approaches over red-flags or profiles, var- ious statistical methods and other methods and approaches (Artis et al., 2002; Bolton and Hand, 2002; Brockett et al., 2002; Estevez et al., 2006; Furlan and Bajec, 2008; Ghosh and Schwartzbard, 1999; Hu et al., 2007; Kirkos et al., 2007; Perez et al., 2005; Quah and Sriganesh, 2008; Rupnik et al., 2007; Sanchez et al., 2009; Viae- ne et al., 2005; Viaene et al., 2002; Weisberg and Derrig, 1998; Yang and Hwang, 2006). Analyses show that in practice none is signifi- cantly better than others (Bolton and Hand, 2002; Viaene et al., 2005). Furthermore, they mainly have three weaknesses. They (1) use inappropriate or inexpressive representation of data; (2) de- mand a labeled (initial) data set; and (3) are only suitable for larger, richer data sets. It turns out that these are generally a problem when dealing with fraud detection (Jensen, 1997; Phua et al., 2005). In the narrower sense, our work comes near the approaches from the field of network analysis, that combine intrinsic attributes of entities with their relational attributes. Noble et al. (2003) pro- posed detecting anomalies in networks with various types of ver- tices, but they focus on detecting suspicious structures in the network, not vertices (i.e. entities). Besides that, the approach is more appropriate for larger networks. Researchers also proposed detecting anomalies using measures of centrality (Freeman, 1977, 1979), random walks (Sun et al., 2005) and other (Holder and Cook, 2003; Maxion and Tan, 2000), but these approaches mainly rely only on the relational attributes of entities. Many researchers have investigated the problem of classifica- tion in the relational context, following the hypothesis that classi- fication of an entity can be improved by also considering its related entities (inference). Thus many approaches formulating inference, spread or propagation on networks have been developed in various fields of research (Brin and Page, 1998; Domingos and Richardson, 2001; Kleinberg, 1999; Kschischang and Frey, 1998; Lu and Getoor, 2003b; Minka, 2001; Neville and Jensen, 2000). Most of them are based on one of the three most popular (approximate) inference algorithms: Relaxation Labeling (RL) (Hummel and Zucker, 1983) from the computer vision community, Loopy Belief Propagation (LBP) on loopy (Bayesian) graphical models (Kschischang and Frey, 1998) and Iterative Classification Algorithm (ICA) from the data min- ing community (Neville and Jensen, 2000). For the analyses and comparison see (Kempe et al., 2003; Sen and Getoor, 2007). Researchers have reported good results with these algorithms (Brin and Page, 1998; Kschischang and Frey, 1998; Lu and Getoor, 2003b; Neville and Jensen, 2000), however they mainly address the problem of learning from an (initial) labeled data set (supervised learning), or a partially labeled (semi-supervised learning) (Lu and Ge- toor, 2003a), therefore the approaches are generally inappropriate for fraud detection. The algorithm we introduce here, IAA algorithm, is almost identical to the ICA algorithm, however it was developed with different intentions in mind – to assess the entities when no la- beled data set is at hand (and not for improving classification with inference). Furthermore, IAA does not address the problem of classi- fication, but ranking. Thus, in this way, it is actually a simplification of RL algorithm, or even Google’s PageRank (Brin and Page, 1998), still it is not founded on the probability theory like the latter. We conclude that due to the weaknesses mentioned, most of the proposed approaches are inappropriate for detection of (auto- mobile) insurance fraud. Our approach differs, as it does not de- mand a labeled data set and is also appropriate for smaller data sets. It represents data with networks, which are one of the most natural representation and allow complex analysis without simpli- fication of data. It should be pointed out that networks, despite their strong foundations and expressive power, have not yet been used for detecting (automobile) insurance fraud (at least according to our knowledge). 3. (Social) networks Networks are based upon mathematical objects called graphs. Informally speaking, graph consists of a collection of points, called 1 Throughout the article the term collision is used instead of (traffic) accident. The word accident implies there is no one to blame, which contradicts with the article. 1040 L. Šubelj et al. / Expert Systems with Applications 38 (2011) 1039–1052 https://isiarticles.com/article/17725 
work_m567gcghv5cdla5fggs3lo4kra	----	RESEARCH REPOSITORY This is the author’s final version of the work, as accepted for publication following peer review but without the publisher’s layout or pagination. The definitive version is available at: http://dx.doi.org/10.1007/s11673-016-9756-7     Attwell, K., Leask, J., Meyer, S.B., Rokkas, P. and Ward, P. (2017) Vaccine rejecting parents’ engagement with expert systems that inform vaccination programs. Journal of Bioethical Inquiry, 14 (1). pp. 65-76.             http://researchrepository.murdoch.edu.au/id/eprint/34803/   Copyright: © 2016 Journal of Bioethical Inquiry Pty Ltd. It is posted here for your personal use. No further distribution is permitted. 1 [issue] 14(1) [category] Symposium: Public Trust in Expert Knowledge [title] Vaccine Rejecting Parents’ Engagement With Expert Systems That Inform Vaccination Programs [subtitle] [authors] Katie Attwell, Julie Leask, Samantha B. Meyer, Philippa Rokkas, Paul Ward [author details] K. Attwell [corresponding author] Sir Walter Murdoch School of Public Policy and International Affairs, Murdoch University South Street, Murdoch, WA 6150, AUSTRALIA Immunisation Alliance of Western Australia, Cockburn Integrated Health and Community Facility Suite 14, 11 Wentworth Parade Success, WA 6164, AUSTRALIA e-mail: k.attwell@murdoch.edu.au J. Leask School of Public Health, Faculty of Medicine, Faculty of Nursing University of Sydney, NSW 2006, AUSTRALIA e-mail: julie.leask@sydney.edu.au S.B. Meyer University of Waterloo, 200 University Ave West Waterloo, Ontario, N2L3G1, CANADA e-mail: samantha.meyer@uwaterloo.ca P. Rokkas Faculty of Medicine, Nursing and Health Sciences, Flinders University, Adelaide, SA 5001 e-mail: philippa.rokkas@flinders.edu.au P. Ward Department of Public Health, Flinders University, Adelaide, SA 5001 2 e-mail: paul.Ward@flinders.edu.au Keywords Vaccination; Vaccine hesitancy; Qualitative; Trust; Giddens; Modernity Abstract In attempting to provide protection to individuals and communities, childhood immunization has benefits that far outweigh disease risks. However, some parents decide not to immunize their children with some or all vaccines for reasons including lack of trust in governments, health professionals, and vaccine manufacturers. This article employs a theoretical analysis of trust and distrust to explore how twenty-seven parents with a history of vaccine rejection in two Australian cities view the expert systems central to vaccination policy and practice. Our data show how perceptions of the profit motive generate distrust in the expert systems pertaining to vaccination. Our participants perceived that pharmaceutical companies had a pernicious influence over the systems driving vaccination: research, health professionals, and government. Accordingly, they saw vaccine recommendations in conflict with the interests of their child and “the system” underscored by malign intent, even if individual representatives of this system were not equally tainted. This perspective was common to parents who declined all vaccines and those who accepted some. We regard the differences between these parents—and indeed the differences between vaccine decliners and those whose Western medical epistemology informs reflexive trust—as arising from the internalization of countering views, which facilitates nuance. Introduction Childhood vaccination is one of public health’s salient achievements (Larson et al. 2014). Alongside environmental public health measures, it constitutes the most successful and cost effective global public health measure to reduce disease-related mortality (Andre et al. 2008). Like any medical intervention, vaccines can cause common minor side effects (e.g. fever) and very rare serious ones (e.g. Guillain-Barré syndrome) (ATAGI 2015). Accordingly, the immunization process induces complex decisions, both rational and emotional, in some parents faced with balancing the welfare of the community with a “do no harm” ethos for their own child. While parental rejection of vaccines is complex and context-specific, varying across time, place, and vaccine (MacDonald and SAGE Working Group on Vaccine Hesitancy 2015), key themes recur in studies, with safety a predominating concern (Casiday et al. 2006; Mills et al. 2005; Smith et al. 2011). Other widely held concerns include the number of vaccinations and 3 perceptions they may overload the immune system (Hilton, Petticrew, and Hunt 2006). Some parents believe alternative medicines may suffice in place of vaccines (Zuzak et al. 2008), and some are unsure why their children still need to be vaccinated against diseases that are now rare (Janko 2012). Occasionally, parents may also recognize that if there is a high proportion of individuals who are already vaccinated, their own child can “hide in the herd” (Offit and Moser 2009). Trust arises as an issue in numerous studies of why people do and do not vaccinate their children (Mills et al. 2005; Dube, Vivion, and MacDonald 2015; Yaqub et al. 2014), but the extent to which trust and distrust—as distinct concepts—shape vaccination decisions remains underexplored. In this study, we theoretically analyse trust and distrust to investigate parents’ perceptions of the expert systems central to vaccination policy and practice in Australia, where approximately 3.3 per cent of children are not up to date with their vaccinations due to their parent or caregiver’s active rejection of some or all vaccines (Beard et al. 2016). We start with an important observation: trust and distrust are not binary categories whose qualities can be populated by the converse of the other. They are conceptually and semantically distinct from each other. Trust may be considered to fall somewhere on a spectrum between complete trust to complete distrust (Brown and Meyer 2015; Gambetta 1988). Distrust is often a focus of study; since distrust correlates with vaccine refusal, policymakers want to understand it in order to address it. Our work is situated in this realm. However, what trust and distrust share is that both are rooted in the cognitive (rational) and affective (emotional) and embedded in broader social contexts, health system understandings, socioeconomic structures, and illness vulnerability and chronology (Brown and Meyer 2015), as well as within health experiences and narratives. We follow a particular narrative here— that most of us appear to trust (for reasons relating to late modernity’s complexity), and some of us distrust (for reasons we further elucidate through empirical study). However, we do not suggest that the cognitive and affective drivers of vaccine refusers’ distrust are mirrored (in the opposite) by vaccine acceptors. We argue, on the contrary, that there are many pathways to the active acceptance of (as well as passive compliance with) vaccination, and trust need not be written on their signposts. When it comes to distrust, however, the strong relationship between distrust and refusal demands our attention, and is the focus of the present paper. First, then, we explore “the rest of us,” in order to contextualize our analysis of vaccine refusers’ distrust. Given that the vast majority of parents in developed countries with access to childhood vaccines make use of them, one might deduce that these parents trust those who make, recommend, and administer them. Trust can be seen as a function to overcome the 4 uncertainties parents face in the decision to vaccinate, since informed medical decision- making requires that an individual grasps the condition and clinical context and has evaluated his or her preferences (Rimer et al. 2004, 1214). Complete knowledge is not always possible (or even perhaps desired) and therefore, accordingly, parents are called to place trust in healthcare providers to act in the interest of their child. Since providers are embedded within broader systems of service organization, professional expertise, and knowledge development (Mollering 2001; Mollering 2006; Luhmann 1979), trust extends to include health systems and broader social systems (e.g., economic, political, judicial) that shape knowledge and assumptions of health and healthcare (Luhmann 1995). Conceptually, disentangling trust in individuals from trust in systems is challenging, since individuals (e.g. GPs, midwives) are seen to represent institutions and thus as “part” of them rather than distinct from them. Giddens (1990) refers to the “meeting places” for interpersonal and institutional trust as “access points”—the doctor for the medical system, the researcher for the scientific system, the politician for the political system, the news reporter for the media and so on. Although everyone is aware that the real repository of trust is in the abstract system, rather than the individuals who in specific contexts “represent” it, access points carry a reminder that it is the flesh and blood people (who are potentially fallible) who are its operators. (Giddens 1990, 85) According to this reasoning, we arrive at the centrality of both the individuals (“access points”) and institutions which together comprise expert systems. Expert systems penetrate nearly all aspects of social life in conditions of modernity (Giddens 1991; Habermas 1997; Scambler and Britten 2001). We trust in expert systems as a way of managing the limited technical knowledge that most of us possess about the information that routinely affects our lives (Giddens 1991). We need not have direct experience with that which we trust or distrust; social or cultural norms underpin the decision to trust and are based on a constructed characterization or stylized notion of the institution (Govier 1998). When we are faced with the uncertainty and risk central to health and medical decision-making, trust in expert systems reduces complexity; empirical research has demonstrated a “will to trust” that veils patients’ anxieties regarding a treatment or medical condition (Brown 2009). Continuing with this narrative, a key marker of contemporary society is that trust can no longer be simply taken for granted or expected (Giddens 1994) and distrust (or at least healthy scepticism) has become the norm (Sztompka 1999, 6). To apply this reasoning to the present context, the bewildering array of individuals professing (or indeed holding) expertise 5 on vaccination in high income countries is in many ways emblematic of late modernity. Today, western medical epistemology is not the only system appealing to parents as a repository for their trust—alternative modalities also entice with critiques of that epistemology. Accordingly, parents choose “what” or “whom” they view as the expert system and, placing trust in that system, make the decision to vaccinate or not (Brownlie and Howson 2005). It is here that we can start to see the relationship between distrust in the dominant expert system and vaccine refusal. Current literatures consistently link institutional distrust in government, pharmaceutical companies, healthcare professions, and medical science and technology to vaccine rejection (Dube, Vivion, and MacDonald 2015). In a British study, parents of under-vaccinated children found it difficult to know where to place their trust and did not trust the government (Austin et al. 2008). An American study identified trust as playing a key role in where parents were situated along the vaccine acceptance spectrum (Benin et al. 2006). Another found that parents who sought exemptions from vaccination were more likely than other parents to report “little or no trust in health information provided by … government agencies, health provider groups or organizations and to distrust local doctors” (Gaudino and Robison 2012, 1135). Parents’ experiences with health professionals (“access points” of the vaccination expert system) are also significant, with insufficient, biased, poorly communicated advice from healthcare providers provoking distrust (Donovan and Bedford 2013). Current social trends towards patient advocacy, empowerment, and choice are heightening some parents’ distrust in vaccines. While a conventional public health approach might concur with a Giddens-style (1994) narrative of an erosion of trust (“once we trusted, now we don’t”), we are cautioned by research on healthcare and trust to take a more critical view of compliance. Patients often follow the instruction of healthcare providers, not because they explicitly trust them but because they feel dependent upon them, particularly in high risk situations or when the public health system precludes choice (Ward et al. 2015). They may also feel obliged to follow directions—to “do as they’re told”—when these instructions come from government. This links to working class or specific culturally mediated notions of duty to the state (Ward, Coffey, and Meyer 2015). Such accounts of compliance remind us that we should not assume that the flipside of vaccine refusing parents’ reflexive distrust is a coherent and deeply held sense of trust experienced by the rest of us. Nor, indeed, can we nostalgically look back to a past where such a deeply held trust determined societal acceptance of vaccination. Lack of choice, lack of knowledge, and obedience to authority may explain this historical compliance as well. 6 Accordingly, it is more helpful to conceptualize a “contemporary health consciousness” (Crawford 2004, 50) driving many parents today to undertake protective action for their children. Protection from perceived risk of vaccines jostles against protection from disease. Parents pursue health information through unfiltered channels such as the Internet (Yaqub et al. 2014), which offers quick access and the advantages of interactivity, information tailoring and anonymity (Cline and Haynes 2001). Health information is one of the most frequently sought topics online (MuMullan 2006), with access growing rapidly through globalization, the diffusion of the news media, and social networking. Although moderated by other factors, information we encounter affects us—one study found that exposure to (what the authors described as) anti-vaccination conspiracy theories reduced the likelihood of intent to vaccinate (Jolley and Douglas 2014). When engaging with conflicting material it can be impossible to know the quality of the evidence chosen by those who assert knowledge. Brownlie and Howson (2005) argue that “when parents are identifying good reasons to either vaccinate or not, they are acting as bricoleurs, piecing together different knowledges” (226), a troublesome process because there are “good reasons” on both sides of the argument (224). In addition, current research identifies the role of social relationships and identities in vaccination decision-making, emphasizing that refusal is not simply a rejection of an offered benefit, but rather a process through which parents navigate social norms and articulate social inclusion and exclusion in relation to those around them (Sobo 2016). The perceptions of members of this research team, as “experts” whose epistemological standpoint enables us to trust those whom we regard as experts, cannot negate the experiences of these parents whose knowledge is otherwise (socially) constructed. Accordingly, we take seriously our participants’ formulation of logic regarding vaccination even as they differ from our own assessments. While our approach privileges scientific knowledge, we acknowledge the body of literature critiquing evidence-based public health as being misappropriated by vested interests, lending support to calls for greater emphasis on lay experience and knowledge to guide policy and practice. Greenhalgh et al. (2014, 1) argue that the “evidence- based brand” is currently distorted as “the drug and medical devices industries increasingly set the research agenda” and “define what counts as a disease and pre-disease ‘risk states.’” Accordingly, critical reflection on the means by which knowledge (and our consequent labelling of “experts”) is constructed requires that we engage seriously with our participants’ questioning of the values underpinning Australia’s vaccination schedule and policies, and the interests being served, in order to better understand what informs their decision-making. 7 These considerations inform our analysis of Australian vaccine rejecting parents’ construction of “what” or “whom” they trust and distrust when making decisions about vaccines. We asked questions about institutions with which we assumed they, as parents, would have direct experience in relation to childhood vaccinations. However, the parents may also be drawing from media representations and/or shared cultural values of the institutions, rather than direct experience. Our results and previous research (Gaudino and Robison 2012; Benin et al. 2006; Leask and Chapman 2002) depict competing expert systems at work, with allopathic healthcare (buttressed by scientific research, government and industry) challenged by the alternative epistemology of complementary and alternative medicine (CAM). Our aim is to understand what it is about allopathic healthcare—the expert system behind vaccination—that leads some parents to distrust both its “access points” (healthcare professionals in particular) and the system as a whole. Methodology We interviewed parents in Fremantle, Western Australia (WA) and Adelaide, South Australia (SA), who declined some or all vaccines for their children. Each recruitment site involved distinct research projects undertaken in different years but with convergent aims and methods sufficient to justify a pooling of data. Both studies aimed to better understand vaccine hesitancy; used a qualitative methodology with semi-structured interviews; and explored in- depth the underpinning perceptions of vaccinations, healthcare professionals, and social systems influencing the structure and function of immunization programmes. In Adelaide, interviews specifically sought to explore trust and distrust, while in Fremantle, interviews were conducted as part of a pro-vaccination campaign development and evaluation (Attwell and Freeman 2015). While each study explored these a priori interests, both focused on parental influences and experiences with vaccination expert systems. The qualitative in-depth interviews enabled us to provide rich accounts of the unique practices and experiences of each parent. In Fremantle, parents were recruited between September 2013 and April 2014 from postcodes surrounding the City of Fremantle, which at the time recorded full vaccine coverage rates at below 87 per cent for children under five compared to the Australian average of just over 90 per cent (National Health Performance Authority 2014). Parents were recruited through posters, advertisements in the local newspapers, social media, and snowballing. Participants were screened prior to interviews to ensure that they met study inclusion criteria of delay or refusal of recommended vaccines and had a child aged five or under. In Adelaide, parents 8 were recruited between October and December 2015 from areas within postal codes identified as having low immunization coverage rates; less than the South Australian average 91.3 per cent at sixty to sixty-three months of age on the Australian Childhood Immunisation Register. Parents were recruited by the researcher at a suburban organic farmers’ market and by snowballing, then screened to ensure that they met the study inclusion criteria of either delaying or refusing vaccination for their children. The University of Western Australia and Flinders University Social and Behavioural Research Ethics Committee provided ethical approval for the project under permit RA 4/1/5890 and project number 6976 respectively. To explore our participants’ engagement with vaccination expert systems in late modernity, we applied deductive social theoretical reasoning (Willis et al. 2007) with a narrative analysis approach to our data. We worked from key principles of narrative analysis: recognizing overlapping stories from participants, interpretive explanations, and the future conclusions our readers would construct (Riessman 2008, 6). Lacking access to our participants’ “unmediated experience,” we were conscious of narratives constructed by “socially situated individuals” within the interview setting (23) and the impact of our interviewing, transcription, and analysis in the stories ultimately told (50). We focused on participants’ interactions with the allopathic healthcare system and other systems and notions of trust and distrust. KA and PR interviewed participants face to face. Following transcription and narrative-reading, KA developed an initial coding tree using QSR International’s NVivo 10 Software. This involved developing a visual representation of expert systems as constructed by participants. PW and JL evaluated a sub-set of the transcripts with continued input from PR, and the team agreed on a final coding structure based on the expert systems connected to vaccination policy and practice. During this process, personal reflection and open discussion during team meetings facilitated reflection on how our knowledge, experiences, beliefs, and backgrounds influenced our reading of the data. Results Semi-structured interviews were conducted with a total of twenty-seven parents: nine from Fremantle and eighteen from Adelaide. Twenty-four participants were women and three were men. Seventeen parents were aged between thirty-six and forty-two; the youngest was twenty-five and the oldest was fifty. The Fremantle parents were younger because of the age requirements of the youngest child. Over half of the parents had university qualifications. The sample included ten parents who had never vaccinated, five who commenced but ceased, seven who were currently delaying or partially vaccinating, and five former delayers now up 9 to date. We recognize differences between parents who decline all vaccines and those willing for their children to receive some or all vaccines eventually. However, in terms of how the participants in our study viewed the vaccination expert system, their fundamental perceptions were similar; for most it was only the strength of these perceptions that determined whether they could vaccinate despite their reservations. For some participants, the willingness to vaccinate despite their distrust developed over time, while for others the converse was true; their distrust of vaccinations increased as time progressed. Unequivocally, parents’ distrust specifically pertained to the pharmaceutical industry and the means by which it “infiltrated” other systems, thereby diminishing trust in allopathic healthcare. This concurs with Luhmann’s (1979) theory of relational trust in social systems. In such a “web of trust” (Meyer et al. 2008), trust or distrust in one system impacts upon trust or distrust in others. As our parents described their interactions with, and perceptions of, the expert systems pertaining to childhood vaccination, their broad distrust of the pharmaceutical industry cast a shadow over other institutions and individuals. Everything, it seemed, was for sale, and this meant the parents could not trust the “experts.” Pharmaceutical Industry as “Puppet Master” Our participants’ distrust of the pharmaceutical industry (and the related systems in the “web”) can be explained by their perception that it acts as a “puppet-master.” Almost all our participants saw the pharmaceutical industry as the dominant force behind the expert systems pertaining to vaccination, “tainting” systems of research, the motives of health professionals, and the operation of government. Many framed this distrust of the pharmaceutical industry within a broader objection to capitalism. It is on this basis that we emphasize similar worldviews between those parents who outright declined all vaccines and those who, despite their dim view of the vested interests, factored in other considerations enabling them to accept some or all vaccines eventually. Almost all our participants depicted companies as responsible to shareholders rather than societies, with the profit motive itself suspect. Participants spontaneously linked the pharmaceutical industry to the resources sector, with its use of fracking; and agribusiness, with its use of genetic modification. “They don’t care about us. It’s all a big moneymaking scam and we are the cattle that is having money made off of us” (female, forty, SA, ceased vaccinator). Participants recounted events like those surrounding the pesticide DDT (dichlorodiphenyltrichloroethane) and the birth defects caused from the anti-nausea drug, thalidomide, in the 1950s and 1960s as part of what one called the “horrific history of the industry” (female, twenty-five, WA, partial vaccinator). 10 …[T]he whole pharmaceutical thing, to me it’s just a big money making scheme…They’re a business. They make money from people being sick, so they don’t want to find a cure for this or that, because that means people won’t rely on their drugs to help them get through their disease, or whatever it is that’s making them sick. They want people to stay sick so that they keep buying their drugs, essentially (female, thirty-two, SA, non-vaccinator). The pharmaceutical industry’s untrustworthiness involved a perception of companies limiting information about risks and inefficacies. In explaining her distrust, the partial vaccinator above described “missing data, unpublished stuff when things don’t turn out how they want them to turn out, the regulations around drug development and testing, and the information on that” (female, twenty-five, WA). Another participant referred to an alleged cover-up, “They’re not going to advertise any negative about their stuff. They bury it. They go to great lengths to make sure that people don’t actually see the negative effects” (male, forty-one, SA, non-vaccinator). The overall sentiment of most participants can be summed up thus: “Vaccines are run by Big Pharma companies; to me it’s all about making money” (female, thirty-nine, SA, delayed vaccinator). Rarely was this given a nuanced interpretation. However, the partial vaccinator who had only declined the varicella vaccine for her son rejected the notion that all pharmaceuticals were untrustworthy. The notion that Big Pharma is only ever out to do evil is a bit naive and delusional. Is it evil to make money? … [T]he Marxist in me probably thinks “Well, yes,” but the pharmaceutical industry has developed drugs that treat cancer (female, twenty-five, WA, partial vaccinator). Another was even more positive: Look, a lot of people are pretty nasty about the pharmaceutical industry and they don’t understand … the research that goes into these drugs and the expenses and the costs and everything that happens, and they’ve got to make good money out of it once they get it. When you pay for a vaccine you’re not actually paying for the production of the vaccine, you’re paying for the twenty years of research and testing that went before it, so I don’t have a problem with that’ (female, forty- seven, SA, partial vaccinator). The only other parent to be equanimous stated, “I don’t trust them in certain areas, I guess. I do think they’re all about making money, and clearly we’ve seen evidence of it, especially in the US, the way they’re involved in politics, those companies … ” However, she went on to 11 reflect, “It’s a funny one, isn’t it, because obviously I trust them to a degree because I give my children medicines or immunisations that come from these companies … ” (female, thirty-eight, SA, full but delayed vaccinator). This last quote encapsulates the difference in degree between partial vaccinators and complete decliners. Partial vaccinators could “get on board” with vaccination to an extent, but still ultimately distrusted its agents on the basis of the pharmaceutical industry’s profit motive. Complete decliners’ more extreme perceptions ruled this out: I see that the pharmaceutical companies … were founded by war criminals from Nazi Germany and they were taken over to America and they started pharmaceutical companies. So are they practising Eugenics on us beyond our— like, without our knowledge? (female, forty, SA, ceased vaccinator). Such distrust of the influence of the pharmaceutical industry was so pervasive that we now track it through the other systems of expert power that emerged from the interviews. We have adopted this approach because in seeking to answer the question: “What makes the vaccination expert system so untrustworthy for parents?” almost every road from the data led to the pharmaceutical industry. By illustrating exactly how the parents perceive its role, with tendrils curling into multiple facets of “expertise,” we can demonstrate how this expertise— the very expertise that we ourselves privilege as researchers—is contingent; socially constructed through cues to trust that do not reach all members of the public. The Pharmaceutical Industry’s Influence on Research Most participants were very suspicious of studies into vaccination, reflecting an assumption that all were funded by vaccine manufacturers and the results, or the dissemination of results, were biased by the profit motive. One respondent, who reported meticulous reading of numerous studies, complained, “Most research seems to be funded by pharmaceutical companies” (female, thirty-six, WA, unvaccinated). Another stated, “There is a lot of conflict of interest within the scientific field of vaccinations because the researchers have shares in outcomes so they only publish favourable outcomes” (female, forty, SA, ceased vaccinator). Participants were also concerned that comparative studies between vaccinated and unvaccinated populations, which might expose some of the risks they feared, would never be conducted. You are not going to put a heap of money into something to prove something that is going to affect you negatively, and when you have got one kind of, like, main corporation, or a few corporations controlling everything; that’s what they do. 12 They are not going to do any kind of research … (male, forty-one, SA, non- vaccinator) Non-vaccinators are often accused of cherry-picking data to support stories of harm, cover- up, and collusion. However, our parents’ accounts highlight two key concerns that cannot be so easily dismissed. The first concern is that funding of research does matter to perception. The political economy of university research—in particular the prevalence of researchers on “soft money”—renders researchers reliant on commercial funds (indeed, this is true of one of the studies in this paper). The second is that, given this political economy, certain research questions are unlikely to ever be asked, because even public money orients towards intervention-focused studies that can employ technocratic expertise in problem solving—with vaccine refusal a “problem” to be “solved.” This is the world with which we, as researchers, make accommodations, but those outside can perceive its limitations. They recognize, with an acuity not always recognized by “experts,” that researchers can pursue objectivity and neutrality but research context is anything but (Leach and Fairhead 2007, 23). Pharmaceutical Companies’ Influence over Doctors Health professionals in general, and doctors in particular, also emerged as important “access points” for parents encountering the vaccination expert system. Participants were concerned about the access pharmaceutical companies had to influence doctors through presentations, “kickbacks,” and junkets, damaging their trustworthiness. [Pharma] are a huge, billion dollar industry. Their shareholders want to make a profit … they’re giving doctors world class, first class flights to somewhere all over the world. If they sell enough of their drug, we will fly you first class, and your wife, to a lecture on this drug. (Female, thirty-nine, SA, non-vaccinator) Participants perceived this financial reach stretching right back to doctors’ training. The pharmaceutical companies supply the universities with funding … so it comes down to the doctors that have, you know, their information that they have and … their scope of operation comes from their training which is funded by, you know. They have got a certain way of being taught and that’s it. (Male, forty-one, SA, non-vaccinator) These data point broadly to the perception of collusion between the pharmaceutical industry and medical professionals, whereby doctors may recommend vaccines that conflict with the interests of the recipient. However, one participant noted, “I have quite a few family members in the medical community, and I find the complete distrust of medical community 13 to be offensive, really...” (female, twenty-five, WA, partial vaccinator). This participant’s personal connection to medical professionals she regarded as trustworthy appeared to influence her perception, though she immediately followed this statement with an unsolicited declaration of distrust in drug companies. We will return to the significance of this later. The Pharmaceutical Industry’s Influence over Government and Bureaucracy A particular concern for participants was the influence of pharmaceutical companies over the state. It was in these responses, in particular, that the perception of the pharmaceutical industry as ultimate “puppet-master” became clear. The participant who described doctors’ “junkets” went on to question (rhetorically), “What are they [the pharmaceutical industry] doing to the government?” She queried whether advisors who help Parliamentarians make decisions were paid or biased (female, thirty-nine, SA, non-vaccinator). Another declared outright: “I think that they [the pharmaceutical industry] own the government” (female, forty, SA, non-vaccinator), and a third stated, “[T]he leaders of the government actually have shares in the pharmaceutical companies that these vaccines represent, so it’s a very big conflict of interest” (female, thirty-six, SA, non-vaccinator). Participants saw the pharmaceutical industry’s government influence playing out in two key ways. First, through the technocratic decisions of highly specialized bureaucrats and advisors who compile the vaccination schedule. Participants were suspicious that the large number of vaccines on the schedule was driven by profit. Even the participant who was most equanimous about the industry’s morality, “wonder[ed] about over immunisation, if it is really about moneymaking” (female, twenty-five, WA, partial vaccinator). Concerns about the schedule also contributed to broader fears of government power, informed by vested interests: [T]hey’re bringing out all these schedules and all this stuff and it’s like, “Give us a break.” … I don’t trust these people to—you know, because it’s all interlinked with the power of government and the agendas. (Male, fifty, SA, non-vaccinator) Second, participants perceived pharmaceutical companies’ reach in governments’ use of policy levers to coerce parents to vaccinate. After the WA data collection, the Australian Federal Government initiated a more “hard line” policy towards vaccine exemptions. “No Jab No Pay” withholds access to financial benefits, including childcare subsidies, from those who decline vaccines by removing a previously available “conscientious objection” (Parliament of Australia 2015). South Australian respondents interviewed after the policy change responded 14 strongly to it. More than one queried its logic, given that the small numbers of parents affected would not produce savings for the government. Rather, it was a matter of control: It’s just knowing that the pharmaceutical companies have got such a strong hold on the government that they make those kinds of decisions … [The government are] not going to benefit [financially] … so the only reason I can see for them doing it is the pharmaceutical companies are taking a stronger hold on our lives. (Female, forty-two, SA, non-vaccinator) This participant continued that she did not see every politician as implicated in “individually think[ing] about it.” Rather, “there are much higher levels in the government and higher people who are pulling the strings.” “Ultimately,” concluded another participant, “money drives decisions” (male, forty-one, SA, non-vaccinator). Another participant elaborated how she saw this policy change playing out: I feel like the pharmaceutical company is getting … a huge say in … how the government chooses to go with those things and if they are going to get a lot of money … or there is going to be a payoff for certain things. I believe that that is also an issue. I think there is a, you know, underhanded deals that, “Yeah, let’s make it so they all have to do it now.” (Female, forty-two, SA, non-vaccinator) Construction of Consensus Part of the role of the pharmaceutical industry as puppet-master was its capacity to construct and sustain ideological hegemony with regard to vaccination in broader society, through sheer financial muscle. I just always wondered about the power of the pharmaceutical companies and the people who make money out of this and the huge power they might have in— which would be a pretty big opponent for people who had other ideas. I could see that just being a very scary environment to step into and present any other ideas or any other research, even. (Female, thirty-nine, SA, delayed vaccinator) The media was part of this picture for respondents who presented a web of profiteering relationships between pharmaceutical companies, magnates, and the government: … When you see things like Rupert Murdoch [owner of many Australian newspapers] was friends with Tony Abbott [Prime Minister at the time] and Rupert Murdoch’s son sits on the board of GlaxoSmithKline [pharmaceutical company producing vaccines], it just doesn’t—you know, your mind goes “Why? Why are you doing this?” (Female, fifty, SA, non-vaccinator) 15 Other Features of Expert Systems Although the dominance of the pharmaceutical industry was the clearest theme to emerge from the research, participants described other quasi-malignant features of the expert systems they had encountered on their (non)vaccination journeys. Several participants used “the system” in their discussion of Western medicine, and one-sidedness, authoritarianism, and rigidity were its key perceived failings, echoing other studies (Brown et al. 2010). Many distrusted Western medicine because they saw it as narrow, symptom-focused (with links back to pharmaceutical companies) and the opposite of holistic. Often, perceptions related to negative experiences with specific health providers, highlighting again how “access points” are central to perceptions of institutions as a whole—in this case the entire institution of Western medicine. Discussion Our study found that the discourse of parents who reject vaccines is dominated with notions of the pharmaceutical industry and the profit motive. This finding echoes those of a number of other studies (Sobo 2015; Benin et al. 2006; Brownlie and Howson 2005). The parents we interviewed saw the pharmaceutical industry as exerting a wide-ranging influence on vaccination research, the motives of health professionals, and the construction of government policy consensus. This discourse was readily mobilized in their rationale for their vaccination decisions. There was a general lack of specificity regarding the agencies and mechanisms through which industry explicitly influenced vaccination policy, although we did not probe for this. We could draw one of two conclusions. Suspicion of the pharmaceutical industry may indeed form a genuine basis for decision-making, but it possibly also serves as a salient post-hoc rationale within a mainstream Australian framing of non-vaccination as deviant. However, the parents in both Fremantle and Adelaide were actively choosing alternative lifestyles— eating organic food and utilizing alternative healthcare and education. A distrust in big business to look after the concerns of the public—coined the Greedy Bastard Hypothesis by Graham Scambler (2001)—is indicative of a left-wing ideology and “off the grid” philosophy. Rejecting capitalism would fit with our parents’ Habitus and their social identities, with vaccine refusal in valued social settings generating bonds and affirmation (Sobo et al. 2016). Accordingly, we take seriously the parents’ representations of the 16 pharmaceutical industry as a significant driver of the distrust that induces vaccine refusal, and recognize its socio-political elements. Our study demonstrates that perceptions of expertise with regard to vaccination are socially constructed, and that judgements of whom or what is trustworthy arise in specific contexts that may facilitate or block the development of trust. We learn from our participants to look to our own social contexts and methods of reasoning, and to ask ourselves how we differ from them in our acceptance of childhood vaccination programs as a public good. Analysing these differences serves two purposes. First, it reminds us to focus on what it actually means to “trust the experts,” something we might otherwise gloss over, given the consensus, at least within our research team, to consciously and reflexively trust expert systems with regard to vaccination. (We do not assume that such trust is a driving factor for the rest of the vaccinating population, since they may be motivated by other factors.) Second, distrust in vaccination can manifest in a “public bad,” with refusal correlating with outbreaks of preventable disease (see Omer et al. 2008). Since the political economy of research inevitably draws us to considering technocratic responses to this “public bad” (to restore the “public good” of herd immunity), it is useful to assess whether interventions address how these parents differ from those who reflexively “trust the experts.” To engage in this comparison, we first return to the political economy of research itself, and our intimacy with government, the public sector, and vaccine manufacturers, as researchers (and sometimes advocates) addressing non-vaccination. As researchers, we are familiar with both the objectivity of the research process and the political and economic drivers of the research we do. Our intimacy with research (including the meaning of systematic reviews, the distinction between correlation and causation and the synthesis of numerous methods to arrive at the same result) provides some of the basis for our trust. Second, we can enhance the comparison between those who trust and those who do not by returning to the distinction between partial vaccinators and complete non-vaccinators. As we noted above, the parents who ultimately consented to some vaccinations still distrusted the industry and its profit motive and did not follow the schedule on this basis. Their acceptance of some vaccines, however, demonstrates mitigation of this distrust with other factors, such that they came to be sufficiently comfortable with nuance. The pharmaceutical industry remained untrustworthy, with its tendrils infiltrating research, medical professionals, and government negatively, yet parents incorporated contradictory beliefs into their assessments. Pharmaceutical companies make drugs that treat cancer. There are good and trustworthy doctors. Parents accept vaccines despite their distrust. 17 We suggest, then, that ultimately the difference between those who trust “enough” to vaccinate and those who do not, is the absolutism of the worldview that the profit motive facilitates only bad. Here, absolutism is not only the strength of the belief; but also the extent to which it is untempered by other considerations, the comprehension of complexity, and sitting with dissonance. It may not be possible to counter all-encompassing worldviews of corporations that seek to profit through harm—inducting and corrupting governance and healthcare in their mission—head on. One possible answer derives from our reflections on how we and some of our research participants arrive at nuance, supported by research that shows selectively vaccinating parents oscillate between numerous standpoints and tend towards indeterminate positions on vaccination (Sobo et al. 2016). We should avoid attempts to drastically re-orient worldviews, given that those initiating them would be distrusted anyway. Instead, we could creatively develop ways of priming for nuance, capacity building for dissonance, and sharing our own insights into how we see what the parents see, but see it differently from inside the system. Additionally, we need to explore mechanisms for increasing trust and consider whether the expert systems under study in this article can do better in this regard. Conclusion Our actively non-vaccinating parents conveyed a perception of vaccination expert systems as tainted by the profit motive of pharmaceutical companies. This “taint” linked to broader anti- capitalist critiques of big business, with parents perceiving the pharmaceutical industry’s reach over research, doctors, governance, and society. With reflection on our own relationship to the vaccination expert system, we suggest that degrees of distrust appear to be determined by the incorporation of countering views. Accordingly, one set of strategies to address the deep distrust that provokes vaccine refusal could involve openly acknowledging the pharmaceutical industry’s sometimes pernicious role, while sharing experiences and perspectives to facilitate nuanced views that can sit alongside it. Competing Interests See under “Funding.” Funding The Fremantle data was gathered by Katie Attwell while working for the Immunisation Alliance of Western Australia, a not-for-profit immunization advocacy organization. The 18 Alliance received funding from Sanofi Pasteur in the form of a $20,000 unrestricted grant to develop and evaluate the “I Immunise” campaign, which itself was funded by the Department of Health, Western Australia. Neither external organization contributed to the study design; data collection, analysis, interpretation, or writing, nor did they influence manuscript submission decisions. The data collection in South Australia was funded by the Flinders Medical Centre Foundation as a Seeding Grant, with no input into the decisions and processes outlined above. Julie Leask receives funding from the Australian Government Department of Health and the National Centre for Immunisation Research and Surveillance for research in addressing vaccine hesitancy. She is in receipt of an NHMRC Career Development Fellowship. References Australian Technical Advisory Group on Immunisation (ATAGI). 2015. The Australian immunisation handbook, 10th ed (2015 update). Canberra: Australian Government Department of Health. Andre, F., R. Booy, H. Bock, et al. 2008. Vaccination greatly reduces disease, disability, death and inequity worldwide. Bulletin of the World Health Organization 86(2): 140– 146. Attwell, K., and M. Freeman. 2015. I Immunise: An evaluation of a values-based campaign to change attitudes and beliefs. Vaccine 33(46): 6235–6240. Austin, H., C. Campion-Smith, S. Thomas, and W. Ward. 2008. Parents’ difficulties with decisions about childhood immunisation. Community practitioner : The journal of the Community Practitioners’ & Health Visitors’ Association 81(10): 32–35. Beard, F.H., B.P. Hull, J. Leask, A. Dey, and P.B. McIntyre. 2016. Trends and patterns in vaccination objection, Australia, 2002–2013. Medical Journal of Australia 204(7): 275. Benin, A.L., D.J. Wisler-Scher, E. Colson, E.D. Shapiro, and E.S. Holmboe. 2006. Qualitative analysis of mothers’ decision-making about vaccines for infants: The importance of trust. Pediatrics 117(5): 1532–1541. Brown, K.F., J.S. Kroll, M.J. Hudson, et al. 2010. Factors underlying parental decisions about combination childhood vaccinations including MMR: A systematic review. Vaccine 28(26): 4235–4248. 19 Brown, P.R. 2009. The phenomenology of trust: A Schutzian analysis of the social construction of knowledge by gynae-oncology patients. Health, Risk & Society 11(5): 391–407. Brown, P.R., and S.B. Meyer. 2015. Dependency, trust and choice? Examining agency and “forced options” within secondary-healthcare contexts. Current Sociology doi: 10.1177/0011392115590091. Brownlie, J., and A. Howson. 2005. “Leaps of faith” and MMR: An empirical study of trust. Sociology 39(2): 221–239. Casiday, R., T. Cresswell, D. Wilson, and C. Panter-Brick. 2006. A survey of UK parental attitudes to the MMR vaccine and trust in medical authority. Vaccine 24(2): 177–184. Cline, R.J.W., and K.M. Haynes. 2001. Consumer health information seeking on the Internet: the satate of the art. Health Education Research 16(6): 671–692. Crawford, R. 2004. Risk ritual and the management of control and anxiety in medical culture. Health, An Interdisciplinary Journal for the Social Study of Health, Illness and Medicine 8(4): 505–528. Donovan, H., and H. Bedford. 2013. Talking with parents about immunisation. Primary Health Care 23(4): 16–20. Dube, E., M. Vivion, and N.E. MacDonald. 2015. Vaccine hesitancy, vaccine refusal and the anti-vaccine movement: influence, impact, and implications. Expert Review of Vaccines 14(1): 99–117. Gambetta, D. 1988. Trust: Making and breaking cooperative relations. Oxford, Cambridge: Basil Blackwell. Gaudino, J.A., and S. Robison. 2012. Risk factors associated with parents claiming personal- belief exemptions to school immunization requirements: Community and other influences on more skeptical parents in Oregon, 2006. Vaccine 30(6): 1132–1142. Giddens, A. 1990. The consequences of modernity. Cambridge: Polity Press. ———. 1991. Modernity and self-identity: Self and society in the late modern age. Stanford: Stanford University Press. ———. 1994. “Risk, trust, reflexivity.” In Reflexive modernization: Politics, tradition and aesthetics in the modern social order, edited by U. Beck, A. Giddens and S. Lash, 194–197. Cambridge: Polity Press. Govier, T. 1998. Dilemmas of trust. Montreal: McGill-Queen’s University Press. Greenhalgh, J., J. Howick, and N. Maskrey. 2014. Evidence based medicine: A movement in crisis? BMJ 348: g3725. 20 Habermas, J. 1997. The theory of communicative action, Vol. 1. Cambridge: Polity Press. Hilton, S., M. Petticrew, and K. Hunt. 2006. “Combined vaccines are like a sudden onslaught to the body’s immune system”: Parental concerns about vaccine “overload” and “immune-vulnerability”. Vaccine 24(20): 4321–4327. Janko, M. 2012. Vaccination: a victim of its own success. American Medical Association Journal of Ethics 13(1): 3–4. Jolley, D., and K.M. Douglas. 2014. The effects of anti-vaccine conspiracy theories on vaccination intentions. PloS one 9(2): e89177. Larson, H.J., C. Jarrett, E. Eckersberger, D.M.D. Smith, and P. Paterson. 2014. Understanding vaccine hesitancy around vaccines and vaccination from a global perspective: A systematic review of published literature, 2007–2012. Vaccine 32(19): 2150–2159. Leach, M., and J. Fairhead. 2007. Vaccine anxieties: Global science, child health and society. London; Stirling, VA: Earthscan. Leask, J., and S. Chapman. 2002. “The cold hard facts” immunisation and vaccine preventable diseases in Australia's newsprint media 1993–1998. Social Science & Medicine 54(3): 445–457. Luhmann, N. 1979. Trust and power: Two works by Niklas Luhmann. Translated by Howard Davis, John Raffan and Kathryn Rooney. Brisbane: John Wiley and Sons. ———. 1995. Social systems, Writing science. Stanford, Calif.: Stanford University Press. MacDonald, N.E., and SAGE Working Group on Vaccine Hesitancy. 2015. Vaccine hesitancy: Definition, scope and determinants. Vaccine 33(34): 4161–4164. Meyer, S., P.R. Ward, J. Coveney, and W. Rogers. 2008. Trust in the health system: An analysis and extension of the social theories of Giddens and Luhmann. Health Sociology Review 17(2): 177–186. Mills, E., A.R. Jadad, C. Ross, and K. Wilson. 2005. Systematic review of qualitative studies exploring parental beliefs and attitudes toward childhood vaccination identifies common barriers to vaccination. Journal of Clinical Epidemiology 58(11): 1081– 1088. Mollering, G. 2001. The nature of trust: From Georg Simmel to a theory of expectation, interpretation and suspension. Sociology 35(2): 403–420. ———. 2006. Trust: Reason, routine, reflexivity. Oxford: Elsevier. 21 MuMullan, M. 2006. Patients using the Internet to obtain health information: How this affects the patient-health professional relationship. Patient Education and Counseling 63(1– 2): 24–28. National Health Performance Authority. 2014. Healthy communities: Immunisation rates for children in 2012–2013. Sydney, NSW: National Health Performance Authority. Offit, P.A., and C.A. Moser. 2009. The problem with Dr. Bob’s alternative vaccine schedule. Pediatrics 123(1): e162–e169. Omer, S.B., K.S. Enger, L.H. Moulton, N.A. Halsey, S. Stokley, and D.A. Salmon. 2008. Geographic clustering of nonmedical exemptions to school immunization requirements and associations with geographic clustering of pertussis. American Journal of Epidemiology 168(12): 1389–1396. Parliament of Australia. 2015. Social services legislation amendment (No jab, no pay) bill Australia. Riessman, C.K. 2008. Narrative methods for the human sciences. Los Angeles, London, New Delhi, Singapore: Sage Publications. Rimer, B.K., P.A. Briss, P.K. Zeller, C.C.Y. Chan, and S.H. Woolf. 2004. Informed decision making: What is its role in cancer screening? Cancer 101(S5): 1214–1228. Scambler, G. 2001. Class, power and the durability of health inequalities. In Habermas, critical theory and health, edited by G. Scambler, 86–118. London: Routledge. Scambler, G., and N. Britten. 2001. System, lifeworld and doctor–patient interaction: Issues of trust in a changing world. In Habermas, critical theory and health, edited by G. Scambler, 45–67. London: Routledge. Smith, P.J., G.H. Sharon, E.K. Marcuse, et al. 2011. Parental delay or refusal of vaccine doses, childhood vaccination coverage at 24 months of age, and the health belief model. Public Health Reports (1974-) 126: 135–146. Sobo, E.J. 2015. Social cultivation of vaccine refusal and delay among Waldorf (Steiner) school parents. Medical Anthropology Quarterly 29(3): 381–399. ———. 2016. Theorizing (vaccine) refusal: Through the looking glass. Cultural Anthropology 31(3): 342–350. Sobo, E.J., A. Huhn, A. Sannwald, and L. Thurman. 2016. Information curation among vaccine cautious parents: Web 2.0, pinterest thinking, and pediatric vaccination choice. Medical Anthropology: Cross Cultural Studies in Health and Illness: 1–18. Sztompka, P. 1999. Trust: A sociological theory. Cambridge: Cambridge University Press. 22 Ward, P.R., C. Coffey, and S. Meyer. 2015. Trust, choice and obligation: A qualitative study of enablers of colorectal cancer screening in South Australia. Sociology of Health & Illness 37(7): 988–1006. Ward, P.R., P. Rokkas, C. Cenko, et al. 2015. A qualitative study of patient (dis)trust in public and private hospitals: The importance of choice and pragmatic acceptance for trust considerations in South Australia. BMC Health Services Research 15(1): 297. Willis, K., J. Daly, M. Kealy, et al. 2007. The essential role of social theory in qualitative public health research. Australian and New Zealand Journal of Public Health 31(5): 438–443. Yaqub, O., S. Castle-Clarke, N. Sevdalis, and J. Chataway. 2014. Attitudes to vaccination: A critical review. Social Science & Medicine 112: 1–11. Zuzak, T.J., I. Zuzak-Siegrist, L. Rist, G. Staubli, and A.P. Simões-Wüst. 2008. Attitudes towards vaccination: Users of complementary and alternative medicine versus non- users. Swiss Medical Weekly 138(47–48): 713–718. Author Version (with link) Accepted Manuscript in form for downloading from the internet 
work_m56ahqpewbavvkcyoocqiqs75m	----	doi:10.1016/j.eswa.2006.12.030 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 1220–1226 Expert Systems with Applications Input data for decision trees Selwyn Piramuthu * Decision and Information Sciences University of Florida, Gainesville, FL 32611–7169, United States Abstract Data Mining has been successful in a wide variety of application areas for varied purposes. Data Mining itself is done using several different methods. Decision Trees are one of the popular methods that have been used for Data Mining purposes. Since the process of constructing these decision trees assume no distributional patterns in the data (non-parametric), characteristics of the input data are usu- ally not given much attention. We consider some characteristics of input data and their effect on the learning performance of decision trees. Preliminary results indicate that the performance of decision trees can be improved with minor modifications of input data. � 2006 Elsevier Ltd. All rights reserved. Keywords: Decision trees; Data characteristics 1. Introduction Data Mining has been successful in a wide variety of application areas, including marketing, for varied purposes (Adomavicius & Tuzhilin, 2001; Kushmerick, 1999; van der Putten, 1999; Shaw, Subramaniam, Tan, & Welge, 2001; Thearling, 1999). Data Mining itself is done using several different methods, depending on the type of data as well as the purpose of Data Mining (Ansari, Kohavi, Mason, & Zheng, 2000; Cooley, Tan, & Srivastava, 1999; Srivas- tava, Cooley, Deshpande, & Tan, 2000). For example, if the purpose is classification using real data, feed-forward neural networks might be appropriate (Ragavan & Pira- muthu, 1991). Decision trees might be appropriate if the purpose is classification using nominal data (Quinlan, 1993). Further, if the purpose is to identify associations in data, association rules might be appropriate (Brijs, Swin- nen, Vanhoof, & Wets, 1999). Decision Trees are one of the popular methods that have been used for Data Mining purposes. Decision trees can be constructed using a variety of methods. For example, C4.5 (Quinlan, 1993) uses information-theoretic measures and CART (Breiman, Friedman, Olshen, & Stone, 1984) uses 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.12.030 * Tel.: +1 352 392 8882; fax: +1 352 392 5438. E-mail address: selwyn@ufl.edu statistical methods. The usefulness as well as classification and computational performance of Data Mining frame- works incorporating decision trees can be improved by (1) appropriate preprocessing of input data, (2) fine-tuning the decision tree algorithm itself, and (3) better interpreta- tion of output. There have been several studies that have addressed each of these scenarios. Input data can be preprocessed (1) to reduce the com- plexity of data for ease of learning, and (2) to reduce effects due to unwanted characteristics of data. The former includes such techniques as feature selection and feature construction as well as other data modifications (see, for example, Brijs & Vanhoof, 1998; Kohavi, 1995; Ragavan & Piramuthu, 1991). The latter includes removal of noisy, redundant, and irrelevant data used as input to decision tree learning. We consider some characteristics of input data and its effect on the learning performance of decision trees. Specif- ically, we consider the effects of non-linearity, outliers, het- eroschedasticity, and multicollinearity in data. These have been shown to have significant effects on regression analy- sis. However, there has not been any published study that deals with these characteristics and their effects on the learn- ing performance of decision trees. Using a few small data sets that are available over the Internet, we consider each of these characteristics and compare their effects on regression mailto:selwyn@ufl.edu S. Piramuthu / Expert Systems with Applications 34 (2008) 1220–1226 1221 analysis as well as decision trees. The results from regres- sion analysis are from the Internet. The contribution of this paper is in studying the effects of these characteristics on decision trees, specifically See-5 (2001). Preliminary results suggest that the performance of decision trees can be improved with minor modifications of input data. The rest of the paper is organized as follows: Evaluation of some input data characteristics and their effects on the learning performance of decision trees is provided in the next section. Experimental results are also included in Sec- tion 2. Section 3 concludes the paper with a brief discussion of the results from this study and their implications as well as future extensions to this study. 2. Evaluation of input data characteristics for decision trees Traditional statistical regression analysis assumes cer- tain distribution (e.g., Gaussian) of input data, as well as Source SS df MS Number of obs = 100 F(3,96) = 2.21 Model 5649.25003 3 1883.08334 Prob > F = 0.0915 Residual 81668.75 96 850.716146 R2 = 0.0647 Adj R2 = 0.0355 Total 87318.00 99 882.00 Root MSE = 29.167 y Coef. Std. Err. t P > jtj (95% Conf. interval) x1 .1134093 .3626687 0.313 0.755 �.06064824 .833301 x2 �.0089643 .3757505 �0.024 0.981 �.7548232 .7368946 x3 .5932696 .2560351 2.317 0.023 .0850430 1.101495 -cons 20.09967 11.61974 1.730 0.087 �2.965335 43.16468 other characteristics of data such as the data being inde- pendent and identically distributed. In most real world data, some of these assumptions are often violated. And, there are several means to at least partially rectify some of the consequences that arise from these violations. We consider a few of these situations: non-linearity in the data, the presence of outliers in the data, the presence of heteroschedasticity in the data, and the presence of multi- collinearity in the data. The data sets used in this study are known to have these characteristics. The follow- Source SS df MS Model 76669.37633 4 19167.3441 Residual 10648.6237 95 112.0907 Total 87318.00 99 88.2 y Coef. Std. Err. t x1 .1706278 .1316641 1.296 x2cent �.0262898 .1363948 �0.193 x2centsq .2954615 .11738 25.171 x3 .2584843 .0938846 2.753 -cons �.8589132 1.3437 �0.639 ing subsections address each of these scenarios in turn. We use See-5 as the decision tree generator throughout this study. 2.1. Non-linearity in input data Non-linearity is a problem in linear regression simply because it is hard to fit a linear model on a non-linear data. Therefore, non-linear transformations are made to the data before running regressions on these data. We con- sider the effects of non-linear data on decision trees both before and after the appropriate data transformations are made. This data set contains four variables – the inde- pendent variables x1, x2, and x3 and the dependent variable y. Result using ordinary least squares (OLS) regression to predict y using x1, x2, and x3 are provided below. The presence of higher order trend effects are identified to be present in the data using the omitted variable test (ovt- est with the rhs option in the statistical analysis software Stata). Higher order trend effects are also present. On inspection of scatter plots of the data, the presence of non-linear trend patterns in the data in the variable x2 is confirmed. We substitute x2 with its centered (x2cent) value (i.e., subtract its mean from every value) and the square (x2centsq) of the centered value. The results from this regression are provided below: Number of obs = 100 F(4,95) = 171.00 Prob > F = 0.0000 75 R2 = 0.08780 Adj R2 = 0.8729 Root MSE = 10.587 P > jtj (95% Conf. interval) 0.198 �.0907856 .4320141 0.848 �.2970677 .244488 0.000 .2721586 .3187645 0.007 .0720997 .4448688 0.524 �3.526496 1.808669 Table 1 Results using non-linear data Input variables Decision tree size Prediction error (%) x1, x2, x3 3.2 (0.6) 28.0 (4.2) x1, x2cent, x2centsq, x3 3.0 (0.5) 11 (2.8) 1222 S. Piramuthu / Expert Systems with Applications 34 (2008) 1220–1226 On further testing for higher order terms, the result turns out to be negative. Here, by including the squared term, it is shown that the new term is indeed statistically significant. The resulting model also fits the data better as shown by the increase in the R2 value. Now, let us consider the same two sets of data, both before and after incorporating the squared term, and eval- uate its effects on the performance of decision tree learned. We use 10-fold cross-validation in See-5 to reduce any bias due to sample selection. Both the mean values and the stan- dard deviation values (in parentheses) are provided for the resulting decision trees. Source SS df MS Number of obs = 100 F(3,96) = 14.12 Model 6358.64512 3 2119.54837 Prob > F = 0.0000 Residual 14406.3149 96 150.06578 R2 = 0.3062 Adj R2 = 0.2845 Total 20764.96 99 209.747071 Root MSE = 12.25 y Coef. Std. Err. t P > jtj (95% Conf. interval) x1 .1986327 .1523206 1.304 0.195 �.1037212 .5009867 x2 .576853 .1578149 3.655 0.000 .2635928 .8901132 x3 .3533915 .1075346 3.286 0.001 .1399371 .5668459 -cons 32.33932 1.229643 26.300 0.000 29.8985 34.78014 Here (Table 1), the addition of the two transformed x2 variables has resulted in a small reduction in the size of the decision trees and a significant decrease in the predic- tion error. The prediction error is the classification error on unseen (during generation of decision trees) examples. 2.2. Presence of outliers in input data The presence of outliers in input data is a problem in any learning application because most methods used for Source SS df MS Number of obs = 100 F(3,96) = 20.27 Model 5863.70256 3 1954.56752 Prob > F = 0.0000 Residual 9255.28744 96 96.4092442 R2 = 0.3878 Adj R2 = 0.3687 Total 15118.99 99 152.717071 Root MSE = 9.8188 y Coef. Std. Err. t P > jtj (95% Conf. interval) x1 .325315 .1220892 2.665 0.009 .08297 .5676601 x2 .4193103 .126493 3.315 0.001 .1682236 .670397 x3 .3448104 0.861919 4.000 0.000 .1737208 .5159 -cons 31.28053 .9855929 31.738 0.000 29.32415 33.23692 learning patterns in the data over-fit the outliers. Depending on how serious the outliers are the resulting patterns learned could be significantly different from the actual patterns with- out the outliers. We consider the effects of outliers on deci- sion trees both before and after the outliers are removed. This data set used here contains four variables – the independent variables x1, x2, and x3 and the dependent var- iable y. Result using OLS regression to predict y using x1, x2, and x3 are provided below. We start out to examine for the presence of outliers by looking at the scatter plot of the data. On inspec- tion, we identify a single point that stands out from the rest. After several attempts at evaluating this point, we find that this point indeed is an outlier. This data point is due to a typographical error during data input, and therefore the error that led to this situation is corrected. The result from the OLS regression run on the data with the outlier problem fixed is provided below. Now, let us consider the same two sets of data, both before and after fixing the problem of outlier data, and evaluate its effects on the performance of decision tree learned. Again, we use 10-fold cross-validation in See-5 to reduce any bias due to sample selection. Table 2 Results using outlier data Input variables Decision tree size Prediction error (%) x1, x2, x3 (with outlier) 4.8 (0.6) 46.0 (5.0) x1, x2, x3 (without outlier) 4.4 (0.6) 38 (4.2) Table 3 Results using heteroschedasticity data Input variables Decision tree size Prediction error (%) x1, x2, x3 (dep var: y) 4.9 (0.6) 27.0 (3.0) x1, x2, x3 (dep. var: ln(y)) 5.5 (0.4) 26 (4.0) S. Piramuthu / Expert Systems with Applications 34 (2008) 1220–1226 1223 Here (Table 2), the removal of the input error has resulted in a small reduction in the size of the decision trees and a significant decrease in the prediction error. 2.3. Heteroschedasticity in input data Data where the error term variance is not constant has Heteroschedasticity. The presence of heteroschedasticity in input data is a problem in regression analysis because the modeling depends on the assumption that that the error term variance is a constant. We consider the effects of het- eroschedasticity on decision trees both before and after the problem has been alleviated in the input data. This data set contains four variables – the independent variables x1, x2, and x3 and the dependent variable y. Result using OLS regression to predict y using x1, x2, and x3 are provided below. Source SS df MS Number of obs = 100 F(3,96) = 65.68 Model 8933.72373 3 2977.90791 Prob > F = 0.0000 Residual 4352.46627 96 45.3381903 R2 = 0.6724 Adj R2 = 0.6622 Total 13286.19 99 134.203939 Root MSE = 6.7334 y Coef. Std. Err. t P > jtj (95% Conf. interval) x1 .2158539 .83724 2.578 0.011 .0496631 .3820447 x2 .7559357 .086744 8.715 0.000 .5837503 .9281211 x3 .3732164 .0591071 6.314 0.000 .2558898 .490543 -cons 33.23969 .6758811 49.180 0.000 31.89807 34.5813 We use the hettest command in Stata to test for hetero- schedasticity, and find that the results are indeed hetero- schedastic. We then try to stabilize the variance by using a natural logarithmic transformation of the dependent var- iable. The OLS regression results after this transformation is provided below: Source SS df MS Model 8.17710164 3 2.7257 Residual 3.74606877 96 .0390 Total 11.9231704 99 .1204 y Coef. Std. Err. t x1 .0054677 .0024562 2.22 x2 .0230303 .0025448 9.05 x3 0.118223 .001734 68.18 -cons 3.445503 .0198285 173.76 Now, let us consider the same two sets of data, both before and after fixing the problem of heteroschedasticity, and evaluate its effects on the performance of decision tree learned. Again, we use 10-fold cross-validation in See-5 to reduce any bias due to sample selection. Here (Table 3), the transformation of the dependent var- iable has resulted in a small increase in the size of the decision trees and an insignificant decrease in the prediction error. 2.4. Multicollinearity in input data Data where the independent variables are highly corre- lated is said to have multi-collinearity. In regression analysis, multicollinearity is a problem when we are interested in the exact values of the coefficients of the independent variables. When multicollinearity is present, this is not possible. Mul- ticollinearity is identified by (1) the presence of high pair- wise correlation among independent variables, (2) a high R 2 value with low t-statistics, and (3) the coefficients change when variables are added and dropped from the model. Multicollinearity is not a problem when the only purpose of regression analysis is forecasting. However, if the analysis is to determine and evaluate the coefficients, multicollinearity Number of obs = 100 F(3,96) = 69.85 0055 Prob > F = 0.0000 2155 R2 = 0.6858 Adj R2 = 0.6760 36065 Root MSE = .19754 P > jtj (95% Conf. interval) 6 0.028 .0005921 .0103432 0 0.000 .0179788 .0280817 0.000 .0083803 .0152643 5 0.000 3.406144 3.484862 Table 4 Results using multicollinearity data Input variables Decision tree size Prediction error (%) x1, x2, x3, x4 3.2 (0.5) 34.0 (3.1) x1, x2, x3 4.8 (0.5) 45.0 (2.7) 1224 S. Piramuthu / Expert Systems with Applications 34 (2008) 1220–1226 is a problem. One of the ways to alleviate this problem is to drop the variable with the highest pair-wise correlation val- ues among the independent variables. We consider the effects of multicollinearity on decision trees both before and after the problem has been alleviated in the input data. This data set contains five variables – the independent variables x1, x2, x3, and x4 and the dependent variable y. Result using OLS regression to predict y using x1, x2, x3, and x4 are provided below. Source SS df MS Number of obs = 100 F(3,96) = 16.37 Model 5995.66253 4 1498.91563 Prob > F = 0.0000 Residual 8699.33747 95 91.5719733 R2 = 0.4080 Adj R2 = 0.3831 Total 14695.00 99 148.434343 Root MSE = 9.5693 y Coef. Std. Err. t P > jtj (95% Conf. interval) x1 1.118277 1.024484 1.092 0.278 �.9155806 3.152135 x2 1.286694 1.042406 1.234 0.220 �.7827429 3.356131 x3 1.191635 1.05215 1.133 0.260 �.8971469 3.280417 x4 �.8370988 1.038979 �0.806 0.422 �2.899733 1.225535 -cons 31.61912 .9709127 32.566 0.000 29.69161 33.54662 Here, the R2 value is significant while the t-statistics are not. We then consider the pair-wise correlations among the independent variables. The resulting matrix is given below: x1 x2 x3 x4 x1 1.0000 x2 0.3553 1.0000 x3 0.3136 0.2021 1.0000 x4 0.7281 0.6516 0.7790 1.0000 Clearly, x4 is highly correlated with the rest of the inde- pendent variables. This variable is then removed from the data. The results from the OLS regression run without x4 is given below: Source SS df MS Number of obs = 100 F(3,96) = 21.69 Model 5936.21931 3 1978.73977 Prob > F = 0.0000 Residual 8758.78069 96 91.2372989 R2 = 0.4040 Adj R2 = 0.3853 Total 14695.00 99 148.434343 Root MSE=9.5518 y Coef. Std. Err. t P > jtj (95% Conf. interval) x1 .298443 .1187692 2.513 0.014 .0626879 .5341981 x2 .4527284 .1230534 3.679 0.000 .2084695 .6969874 x3 .3466306 .0838481 4.134 0.000 .1801934 .5130679 -cons 31.50512 .9587921 32.859 0.000 29.60194 33.40831 Here, all the variables turn out to be significant. Now, let us consider the same two sets of data, both before and after removing x4 from the data, and evaluate its effects on the performance of decision tree learned. Again, we use 10-fold cross-validation in See-5 to reduce any bias due to sample selection. Here (Table 4), the transformation of the dependent var- iable has resulted in a significant increase in the size of the decision trees and a significant increase in the prediction error. Here, removal of the variable does not seem to help the performance of decision trees. 2.5. Data reduction Data reduction is an important preprocessing step in any pattern recognition method. Benefits of data reduction include removal of irrelevant attributes from data, alleviate effects due to outliers in data, reduce effects due to noise, resulting parsimony in decision-making, reducing data complexity for learning algorithms, among others. These are even more critical when dealing with huge data sets, as is the case with most data mining applications. Data reduction essentially involves dimensionality reduction and/or example reduction, among others. In a majority of Table 5 Summary results for data reduction # of training examples Training error Testing error Tree size 120 2 (0.94) 4.66 (3.81) 4.6(0.55) 60 3.68 (2.17) 0 (0) 3.4 (0.55) 30 3.32 (2.37) 4 (2.81) 3 (0) 15 0 (0) 0 (0) 3 (0) 9 0 (0) 0 (0) 3 (0) 6 0 (0) 5.34 (7.69) 3 (0) S. Piramuthu / Expert Systems with Applications 34 (2008) 1220–1226 1225 cases, when dealing with data used as input to learning algo- rithms, the former deals with reducing the number of rele- vant attributes and the latter involves effectively reducing the number of examples. In this study, we present and eval- uate a framework to reduce the effective number of exam- ples that are used as input to pattern learning algorithms. Most example reduction methods use some form of sam- pling (e.g., random, stratified) to select examples to be con- sidered further. The underlying assumptions are different for each of these sampling methods. For example, in simple random sampling, it is assumed that every unit in the pop- ulation provides the same amount of information. There is a vast amount of literature on example reduction methods (e.g. Ishibuchi, Nakashima, & Nil, 2001; Liu & Motoda, 2001; Provost & Kolluri, 1999; Provost, Jensen, & Oates, 1999). We utilize clustering as a pre-processing step for learn- ing applications. Specifically, we use (k-means) clustering for data reduction in the number of example dimension, and as a pre-processing step for decision trees. We use the fuzzy thresholds option in See-5 since we are using real-numbered variables in the data set. See-5 partitions real-numbered variable at a given threshold point, and small movements in the variable value near the threshold can change the branch taken. The Fuzzy thresholds option softens this knife-edge behavior for decision trees by con- structing an interval close to the threshold. Within this interval, both branches of the tree are explored and the results combined to give a predicted class. We use the Iris plants database (Anderson, 1935; Fisher, 1936) to illustrate the proposed method. We chose this database simply because it is among the best known pat- tern recognition databases (e.g. Duda & Hart, 1973), its simplicity, and any results generated using this database can be readily compared with those generated through other methods. The database consists of 150 examples cov- ering three iris plant types (Iris-Setosa, Iris-Versicolor, and Iris-Virginica), with 50 examples from each type. Data from one of the type is linearly separable from the other two, and the latter two are not linearly separable from each other. The four independent attributes (sepal length, sepal width, petal length, and petal width) are numeric. There are no missing attribute values. Since we want to compare the results for the case where clustering is used with that where clustering was not used as a pre-processing step, we randomly divided the data set into five parts (a,b,c,d,e). We did this to alleviate problems due to sampling errors. The five data sets were then used in a leave-one-out fashion. I.e., in the first case, the first four parts were used to generate the clusters (abcd), which were then used as input to See-5. The resulting decision tree was tested using the last part (e). We repeated this five times for each case, where different parts were used to test the result- ing decision tree. Table 5 summarizes the results based on the number of examples used as input to the decision tree generator (See- 5). For training error, testing error and tree size, the mean value is followed by standard deviation values in parenthe- ses. As can be seen neither the training error nor the testing error seem to be uni-modal functions. However, there is a significant decrease in both training (pair-wise two-tailed t- test with p < 0.005) and testing (pair-wise two-tailed t-test with p = 0.026) error as the number of clusters decreases. This could possibly be because the benefits of clustering become apparent as the number of examples used to form any given cluster is increased. In the example data set used, the best results are for data sets with 15 and 9 clusters. Then again, as the number of clusters is reduced to its min- imum (here, two for each iris type, resulting in 6 clusters overall), there tends to be a loss in information content simply because of the complexity of data dictating the necessity for more points (here, clusters) to draw bound- aries among examples belonging to different categories (here, type of iris). The decision tree size also decreases (statistically significant with a pair-wise two-tailed t-test with p < 0.002) as the number of clusters is decreased. 3. Discussion Even though decision trees constructed using informa- tion-theoretic measures are considered non-parametric, the distribution of data does influence the classification performance of these decision trees. Preliminary results indicate that the performance of decision trees can be improved by considering the effects due to non-linearity, outliers, heteroschedasticity, and multicollinearity in input data as well as data reduction. Both non-linearity and the presence of outliers did affect the classification performance of decision trees. The presence of heteroschedasticity did not affect the classification performance of decision trees significantly. And, the presence of multicollinearity is not of concern for decision trees. The attempt to remove mul- ticollinearity resulted in poor classification performance. Data reduction resulted in improved performance both in terms of the resulting tree-size and classification. We are currently in the process of evaluating the results we have thus far using larger and more data sets. We are also in the process of studying why these data characteris- tics affect the classification performance of decision trees. Also, in this study, we were only interested in the size of the decision trees and their classification accuracy. The computational cost of this process is also important, and this is left as an exercise for a future study. In addition to the characteristics presented in this paper, we are 1226 S. Piramuthu / Expert Systems with Applications 34 (2008) 1220–1226 also evaluating other data characteristics including non- independence and non-normality of data. We presented one possible means to improve the classi- fication performance of decision trees. This, along with other pre-processing methods (such as feature selection and feature construction), methods for fine-tuning decision trees, and those that enhance interpretability of results, would help improve the overall performance of these deci- sion support tools incorporating decision trees. References Adomavicius, G., & Tuzhilin, A. (2001). Using data mining methods to build customer profiles. IEEE Computer (February), 74–82. Anderson, E. (1935). The Irises of the Gaspe Peninsula. Bulletin of the American Iris Society, 59, 2–5. Ansari, S., Kohavi, R., Mason, L., & Zheng, Z. (2000). Integrating E- commerce and data mining: architecture and challenges. WEB- KDD’2000 workshop on Web mining for E-commerce – challenges and opportunities, August. Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and regression. Wadsworth. Brijs, T., & Vanhoof, K. (1998). Principles of data mining and knowledge discovery. In J. M. Zytkow (Ed.), PKDD ’98. Lecture notes in artificial intelligence (Vol. 1510, pp. 102–110). Berlin: Springer. Brijs, T., Swinnen, G., Vanhoof, K., & Wets, G. (1999). Using association rules for product assortment decisions: a case study. In Proceedings of the 5th international conference on knowledge discovery and data mining (KDD’99) (pp. 254–260). San Diego, CA, August 15–18, 1999. Cooley, R., Tan, P.-N. & Srivastava, J. (1999). Discovery of interesting usage patterns from Web data. Technical Report, Department of Computer Science and Engineering, University of Minnesota. Duda, R. O., & Hart, P. E. (1973). Pattern classification and scene analysis. John Wiley & Sons (p. 218). Fisher, R. A. (1936). The use of multiple measurements in taxonomic problems. Annual Eugenics, 7(Part II), 179–188. Ishibuchi, H., Nakashima, T., & Nil, M. (2001). Genetic-algorithm-based instance and feature selection. In H. Liu & H. Motoda (Eds.), Instance selection and construction for data mining. Kluwer Academic. Kohavi, R. (1995). Wrappers for performance enhancement and oblivious decision graphs, Ph.D. Dissertation, Computer Science Department, Stanford University. Kushmerick, N. (1999). Learning to remove Internet advertisements. Third International Conference on Autonomous Agents. Liu, H., & Motoda, H. (Eds.). (2001). Instance selection and construction for data mining. Kluwer Academic. Provost, F., Jensen, D., & Oates, T. (1999). Efficient progressive sampling. In Proceedings of the 5th international conference on knowledge discovery and data mining (pp. 23–32). AAAI Press. Provost, F., & Kolluri, V. (1999). A survey of methods for scaling up inductive algorithms. Data Mining and Knowledge Discovery, 3(2), 131–169. Quinlan, J. R. (1993). C4.5 programs for machine learning. Morgan Kaufman. Ragavan, H., & Piramuthu, S. (1991). The Utility of Feature Construction in Back-propagation. Proceedings of the twelfth IJCAI, 844–848. See-5. (2001). Rulequest research data mining tools. Shaw, M., Subramaniam, C., Tan, G. W., & Welge, M. E. (2001). Knowledge management and data mining for marketing. Decision Support Systems, 31, 127–137. Srivastava, J., Cooley, R., Deshpande, M., & Tan, P.-N. (2000). Web usage mining: discovery and applications of usage patterns from Web data. SIGKDD Explorations, 1(2), 12–23, January. Thearling, K. (1999). Data mining and CRM: zeroing in on your best customers. dmDitrect (December), 20. van der Putten, P. (1999). Data mining in direct marketing databases. In W. Baets (Ed.), Complexity and management: a collection of essays. Singapore: World Scientific Publishers. Input data for decision trees Introduction Evaluation of input data characteristics for decision trees Non-linearity in input data Presence of outliers in input data Heteroschedasticity in input data Multicollinearity in input data Data reduction Discussion References 
work_malg4cvzmfb23i2pyh6r3trb5u	----	Hybrid recommendations for mobile commerce based on mobile phone features Hybrid recommendations for mobile commerce based on mobile phone features Chuen-He Liou and Duen-Ren Liu Institute of Information Management, National Chiao Tung University, Hsinchu, Taiwan Email: dliu@iim.nctu.edu.tw Abstract: Mobile data communications have evolved as the number of third generation (3G) subscribers has increased. The evolution has triggered an increase in the use of mobile devices, such as mobile phones, to conduct mobile commerce and mobile shopping on the mobile web. There are fewer products to browse on the mobile web; hence, one-to-one marketing with product recommendations is important. Typical collaborative filtering (CF) recommendation systems make recommendations to potential customers based on the purchase behaviour of customers with similar preferences. However, this method may suffer from the so-called sparsity problem, which means there may not be sufficient similar users because the user-item rating matrix is sparse. In mobile shopping environments, the features of users’ mobile phones provide different functionalities for using mobile services; thus, the features may be used to identify users with similar purchase behaviour. In this paper, we propose a mobile phone feature (MPF)-based hybrid method to resolve the sparsity issue of the typical CF method in mobile environments. We use the features of mobile phones to identify users’ characteristics and then cluster users into groups with similar interests. The hybrid method combines the MPF-based method and a preference-based method that uses association rule mining to extract recommendation rules from user groups and make recommendations. Our experiment results show that the proposed hybrid method performs better than other recommendation methods. Keywords: mobile web, one-to-one marketing, product recommendation, collaborative filtering, mobile phone features, association rules 1. Introduction In the last decade, mobile communications have evolved from 2G=2.5G to 3G=3.5G. As a result, the data transfer rate has been progressively upgraded from 64 Kbps (2.5G=GPRS) to 384 Kbps (3G=WCDMA), and 3.5 Mbps (3.5G= HSDPA), which is comparable to the wired Internet. The evolution has triggered an increase in the use of mobile devices, such as mobile phones, to conduct mobile commerce (m- commerce) on the mobile web (Chae & Kim, 2003; Venkatesh et al., 2003; Ngai & Guna- sekaran, 2007). M-commerce covers a large num- ber of services, one of which is mobile shopping. Retailers have also increased their investment in mobile shopping channels to deliver dedicated products, content and promotions to customers. Recommender systems have emerged in m-commerce or e-commerce applications to support product recommendation, which pro- vide individual product recommendation for each customer. Recommender systems assist business in implementing one-to-one marketing strategies, relying on customer purchase history to determine preferences and identify products that a customer may purchase. Recommender DOI: 10.1111/j.1468-0394.2010.00566.x Article _____________________________ c� 2010 Blackwell Publishing Ltd Expert Systems, May 2012, Vol. 29, No. 2 108108 c� 2010 Blackwell Publishing LtdExpert Systems, May 2012, Vol. 29, No. 2 systems increase the probability of cross-selling, establish customer loyalty and fulfill customer needs by discovering products in which they may be interested in purchasing (Schafer et al., 2001). Recommender systems are widely used to recommend various items, such as consumer products, movies and music, to customers based on their interests (Hill et al., 1995; Shardanand & Maes, 1995). Generally, recommender sys- tems can be classified as collaborative or con- tent-based filtering techniques. Collaborative filtering (CF), which has been used successfully in various applications, utilizes preference rat- ings given by customers with similar interests to make recommendations to a target customer (Resnick et al., 1994; Linden et al., 2003; Lee, 2004; Cho et al., 2005; Liu & Shih, 2005). In contrast, content-based filtering derives recom- mendations by matching customer profiles with content features (Mooney & Roy, 2000; Martinez et al., 2007). A number of product recommendation sys- tems have been developed for m-commerce on the mobile web (Kim et al., 2004; Choi et al., 2007; Lee & Park, 2007). For example, VISCOR is a mobile recommender that combines colla- borative and content-based filtering to provide better wallpaper recommendations (Kim et al., 2004). MCORE considers users’ context data to recommend mobile services (Choi et al., 2007). In addition, mobility information about user locations obtained from global positioning sys- tems (GPS) is usually used in m-commerce. A number of mobile recommendation systems use customers’ mobility patterns to make recommen- dations (Brunato & Battiti, 2003; Yang et al., 2008). Mobile phone features (MPF) such as Bluetooth and card slots have been used as product attributes to recommend mobile phone products. iTVMobi recommends mobile phone products based on the users’ preferences for MPF (Virvou & Savvopoulos, 2007). Existing works use MPF as product attributes to recom- mend mobile phone products, instead of using the features of mobile phones as the users’ characteristics (profiles) to recommend products. The typical CF method relies on finding users with similar interests to make recommendations. However, it may suffer from the so-called spar- sity problem because users only rate a few items. As a result, the user-item rating matrix is very sparse, so the recommendation quality is poor due to the difficulty of finding users with similar interests. In mobile shopping environments, active users may only browse=purchase a few items on the mobile web; thus, it is difficult to find users with similar interests based on the product preferences derived from users’ brow- sing=purchasing histories. In this study, we propose a MPF-based hy- brid method to resolve the sparsity issue of the typical CF method used in mobile environ- ments. The MPF-based method uses the fea- tures of users’ mobile phones as user profiles to cluster users into groups with similar character- istics and then makes recommendations. The MPF indicate users’ motivations for using mo- bile services; thus, they can be used to identify users with similar product preferences. For example, the profiles of businessmen or sales representatives who own mobile phones with intelligence and GPS features may indicate a strong interest in high-tech 3C (Computer, Communication and Consumer) products. Thus, we consider MPF as user characteristics to help find users with similar interests. How- ever, some users who own mobile phones with the similar features may not have the similar product preferences. Hence, we still need to refer users’ product preferences for making recom- mendations. Thus, we propose a hybrid method which combines the MPF-based method and the preference-based method to improve recom- mendation quality by considering both MPF and product preferences. Similar to the MPF- based method, the preference-based method makes recommendations based on user groups that are clustered according to the users’ pro- duct preferences. Experiments were conducted to compare the performance of the proposed hybrid method with that of MPF-based, prefer- ence-based and typical CF methods. The results show that the hybrid method outperforms the other methods. The remainder of this paper is organized as follows. In Section 2, we illustrate the c� 2010 Blackwell Publishing LtdExpert Systems, May 2012, Vol. 29, No. 2109 109c� 2010 Blackwell Publishing Ltd Expert Systems, May 2012, Vol. 29, No. 2 background of related methods. In Section 3, we describe the proposed MPF-based, preference- based and hybrid recommendation methods. In Section 4, we present the evaluation metrics and the experiment results. Then in Section 5, we summarize our findings and draw some conclusions. 2. Background Our proposed method is based on MPF, and uses association rule-based and most frequent item-based recommendation methods. In this section, we briefly introduce the concepts and methods that are used in our research. This section also illustrates the typical CF method that is compared with our approach in experi- ment evaluation. 2.1. MPF Mobile phones have evolved from the tradi- tional voice communication model to advanced digital convergence platforms with various fea- tures, such as Bluetooth technology, cameras, card slots, flash lights, as well as java, MP3, radio, touch screen, video and Wi-Fi functions (Ojanpera, 2006). These features enable users to access related mobile services, for example, download MP3 files, upload photos to blogs, video streaming and on-line shopping (Ko et al., 2007). Ling et al. (2006) investigated the impact of MPF on user satisfaction and analysed the feature preferences of diverse ethnic groups as well as preferences based on gender. Virvou and Savvopoulos (2007) developed an intelligent application called iTVMobi, which recommends mobile phone products on an interactive televi- sion. The system uses K-means clustering to group users based on their preferences for the attributes of mobile phones. The system then applies an association rule-based approach to recommend mobile phones based on the users’ preferences. Existing works use MPF as product attributes to recommend mobile phone products, instead of using the features of mobile phones as user characteristics (profiles) to recommend products. The features of different types of mobile phones can be obtained from the respec- tive websites. In this study, we log users’ mobile phone types when they browse products on a mobile shopping website. Then, we derive the phone features preferred by each user and use them to compile MPF-based user profiles. 2.2. Users clustering Clustering techniques, which are usually used to segment users (Punj & Stewart, 1983; Chen et al., 1996), seek to maximize the variance among groups while minimizing the variance within groups. Many clustering algorithms have been developed, such as K-means, hierarchical and fuzzy c-means algorithms (Omran et al., 2007). K-means clustering (MacQueen, 1967) is a similarity grouping method widely used to partition a dataset into k groups. The K-means algorithm assigns instances to clusters based on the minimum distance principle, which assigns an instance to a cluster such that the distance to the centre of the cluster is the minimum over all k clusters. 2.3. Association rule-based recommendation method Association rule mining tries to find the associa- tions between two sets of products in a transac- tion database. Agrawal et al. (1993) formalized the problem of finding association rules that satisfy the minimum support and the minimum confidence requirements. For example, assume that a set of purchase transactions includes a set of product items I. An association rule is an implication of the form: X ) Y, where X � I, Y � I and X T Y ¼ F. X is the antecedent (body) and Y is the consequent (head) of the rule. Two measures, support and confidence, are used to indicate the quality of an association rule. The support of a rule is the percentage of transactions that contain both X and Y, whereas the confidence of a rule is the fraction of transactions that contain X and also contain Y. Sarwar et al. (2000) described the association rule-based recommendation method as follows. For each customer, a customer transaction is c� 2010 Blackwell Publishing Ltd Expert Systems 3110 c� 2010 Blackwell Publishing LtdExpert Systems, May 2012, Vol. 29, No. 2 created to record all the products he=she pur- chased previously. An association rule mining algorithm is then applied to find all the recom- mendation rules that satisfy the given minimum support and minimum confidence. The top-N products to be recommended to a customer u, are then determined as follows: Let Xu be the set of products previously purchased by u. The method first finds all the recommendation rules X ) Y in the rule set. If X � Xu then all products in Y � Xu are deemed candidate pro- ducts for recommendation to the customer u. The candidate products are then sorted and ranked according to the associated confidence of the recommendation rules, and the top-N candidate products are selected as the top-N recommended products. 2.4. Most frequent item-based recommendation method The most frequent item-based recommendation method (Sarwar et al., 2000) counts the pur- chase frequency of each product by scanning the products purchased=browsed by the users in a cluster. Next, all the products are sorted by the purchase frequency in descending order. Final- ly, the method recommends the top-N products that have not been purchased=browsed by the target customer. 2.5. Typical CF method CF (Resnick et al., 1994; Shardanand & Maes, 1995) utilizes the nearest-neighbour principle to recommend products to a target audience. The neighbours are identified by computing the similarity of customers’ purchase behaviour or tastes. The similarity is measured by Pearson’s correlation coefficient, which is defined as fol- lows: corrPðCi; CjÞ ¼ P s2I ðrCi;s � �rCiÞðrCj;s � �rCjÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP s2I ðrCi;s � �rCiÞ2 P s2I ðrCj;s � �rCjÞ2 q ð1Þ where �rCi and �rCj denote the average number of products purchased by customers Ci and Cj, respectively; variable I denotes the mix of the set of products; and rCi;s and rCj;s indicate, respectively, that customers Ci and Cj purchased product item s. The typical CF method utilizes k-nearest neighbours (k-NN) to recommend N products to a target user (Sarwar et al., 2000). The k-NN are identified by computing the similarity of customers’ purchase behaviour or tastes. The similarity is measured by Pearson’s coeffi- cient, as shown in equation (1). After the neighbourhood has been formed, the N recom- mended products are determined by the k-NN as follows. The frequency count of products is calculated by scanning the data about the products purchased=browsed by the k-NN. The products are then sorted based on the frequency count, and the N most frequent pro- ducts that have not been purchased by the target customers are selected as the top-N recommen- dations. 3. Proposed MPF-based hybrid recommendation method In this section, we describe the proposed hybrid recommendation method, which combines an MPF-based method and a preference-based method, as shown in Figure 1. First, the MPF- based method extracts the features of users’ mobile phones from the respective phone web- sites, as shown on the left-hand side of the figure. The features of users’ mobile phones are taken as user profiles to identify users with similar characteristics. The system then applies the K-means clustering method to cluster users into groups based on the similarity of users’ MPF. Next, the association rules and frequently browsed products are extracted from each clus- ter. The system then recommends products based on the association rules and frequently browsed products. However, there may be very few products recommended according to the association rules because of the limited number of products that can be browsed on the mobile web. If the association rule-based recommenda- tions are not sufficient, the most frequent item- based recommendations are used to recommend 4 Expert Systems c� 2010 Blackwell Publishing Ltd111c� 2010 Blackwell Publishing Ltd Expert Systems, May 2012, Vol. 29, No. 2 products to users. Similar to the MPF-based method, the preference-based method, shown on the right-hand side of Figure 1, clusters users by the K-means clustering method based on Pearson’s correlation coefficient of users’ pro- duct preferences. It then recommends products based on the association rules and the most frequent items. Finally, the hybrid recom- mendation scheme combines the MPF-based recommendations and preference-based recom- mendations with the hybrid ratio determined by the preliminary analytical data to recommend products. We discuss the recommendation phase of the MPF-based, preference-based and hybrid recommendation schemes in Sections 3.2 and 3.3, respectively. 3.1. Data pre-processing and clustering We obtained the features of each mobile phone from one of the mobile phone websites. There are more than 100 features on a mobile phone. It is hard to analyse all of them. Therefore, we selected the features based on the following three criteria. (1) Advertisements of a mobile phone retailer: The advertisements of a mobile phone retailer often list the important features for users’ preferences and comparison; (2) Fea- tures with too many missing values are not suitable for analysis and thus are not selected; and (3) Features with values that can discrimi- nate the differences of mobile phones. Table 1 lists the selected features, including Bluetooth Figure 1: An overview of the proposed hybrid recommendation scheme. c� 2010 Blackwell Publishing Ltd Expert Systems 5112 c� 2010 Blackwell Publishing LtdExpert Systems, May 2012, Vol. 29, No. 2 technology, cameras, card slots, flash lights, as well as java, MP3, radio and video functions. The price feature is complicated for analysis, since the prices of mobile phones may vary under different subscription fees provided by various service providers. Thus, we do not select the price feature. The price feature has been somewhat implicitly considered and depends on the selected eight features, because mobile phones with more features are often more ex- pensive. The display feature is not listed in the advertisements of the mobile phone retailer and is a combination of three discrete data type features including screen size, colour and mate- rial. These values of the display features are missing and are difficult to collect. Thus, we do not select the display feature. We calculate the similarity of users based on the selected features. The camera quality feature, which is a discrete data type, and the other seven features are Boolean data types. The camera resolution pixels (3.2, 2.0 and 1.3 mega-pixel resolution) need to be normalized to the seman- tic values of high, medium and low, as shown in equation (2) (Lin et al., 2003). Therefore, we use three Boolean operators to represent high, med- ium and low quality camera resolution (1, 0, 0) represents high quality (0, 1, 0) represents med- ium quality, and (0, 0, 1) represents low quality. Zcamera ¼ Xcamera � MðXcameraÞ sXcamera ð2Þ where Xcamera is the camera quality; and M(Xcamera) and sXcamera are, respectively, the mean value and the standard deviation of the camera quality. Next, we identify all the users’ mobile phones and expand the phones’ features to form a user-mobile phone feature matrix, as shown in Table 2. In the matrix, the values of the camera resolution mega-pixels are transformed into semantic values based on equation (2), with Zcamera < � 0.8, � 0.8 % Zcamera % 0.8 and Zcamera > 0.8, representing low-level, medium- level and high-level quality cameras, respec- tively. We then use the matrix to cluster the users into groups. The MPF-based method clusters users by the K-means clustering method with Pearson’s correlation coefficient based on the users’ preferred MPF. User product preference clustering is more intuitive than user MPF clustering, as it clusters users directly based on the user-product prefer- ence matrix. The preference-based method clus- ters users by the K-means clustering method with Pearson’s correlation coefficient based on users’ product preferences. 3.2. The MPF-based and preference-based recommendation phase After clustering users into groups based on similar MPF or product preferences, the asso- ciation rules and the most frequent items in each group (cluster) are generated for the next step of the recommendation phase. The steps of the MPF-based and preference-based recommenda- tion phase are shown in Figure 2 and described as follows. Let Xu represent the set of products browsed previously by a user u. For each asso- ciation rule X k ! Yk, if Xk � Xu then all pro- ducts in Y k � Xu, denoted by Yuk, are regarded as candidate products for recommendation to the user u. Let Yu AR be the set of all candidate products generated from all association rules that satisfy X k � Xu. The products in YuAR are ranked according to c(Yu k ), that is, the associated confidence of the association rule (AR) X k ! Yk. We compare the number of candidate pro- ducts jYARu jand the top-N recommendations. If the former is greater than the latter, the system Table 1: Mobile phone features No Feature Data type Value 0 Bluetooth Boolean (0, 1) 1 Camera quality Discrete (Low, Medium, High) 2 Card slot Boolean (0, 1) 3 Flash light Boolean (0, 1) 4 Java Boolean (0, 1) 5 MP3 Boolean (0, 1) 6 Radio Boolean (0, 1) 7 Video Boolean (0, 1) 6 Expert Systems c� 2010 Blackwell Publishing Ltd113c� 2010 Blackwell Publishing Ltd Expert Systems, May 2012, Vol. 29, No. 2 Table 2: User-mobile phone feature matrix User ID Phone type Bluetooth Camera Card slot Flash light Java MP3 Radio VideoH M L 1 MOTO V191 0 0 0 1 0 0 1 1 0 1 2 Nokia N70 1 1 0 0 1 1 1 1 1 1 3 SAMSUNG SGH-Z238 1 0 1 0 1 0 1 1 0 1 4 Sony Ericsson K800i 1 1 0 0 1 1 1 1 1 1 Figure 2: The MPF-based and preference-based recommendation phase. c� 2010 Blackwell Publishing Ltd Expert Systems 7114 c� 2010 Blackwell Publishing LtdExpert Systems, May 2012, Vol. 29, No. 2 recommends the top-N products among the products in Yu AR . On the other hand, if the number of candidate products jYARu j is less than the number of top N recommendations ðjYARu j < NÞ, the remaining N � jYARu j pro- ducts for recommendation are selected from Yu MF . The selected products are the most fre- quent items ranked according to the frequency count of products browsed by the users in the target user’s cluster. Then, products in Yu MF that have not been browsed by the user u and have not been included in Yu AR are added to the recommended product list so that the number of top-N recommendations is sufficient. 3.3. The hybrid recommendation phase The hybrid recommendation phase combines the MPF-based method and the preference- based method, as shown in Figure 3. Similar to the MPF-based method, the hybrid method first recommends products based on the association rules (AR); and then recommends products based on the most frequent item (MF) count. Let X Mi ! YMi and XPj ! YPj be the associa- tion rules extracted from an MPF-based cluster (M) and a preference-based cluster (P), respec- tively; and let their associated confidence scores be c Mi and c Pj , respectively. In addition, let Xu represent the set of products previously browsed by the target user u; and let Yu AR be the set of all candidate products generated from all associa- tion rules that satisfy X Mi � Xu or XPj � Xu. The products in Yu AR are ranked according to the weighted sum of their confidence scores. cH ¼ wM � cMi þ wP � cPj ð3Þ where wM and wP are the weights assigned to the MPF-based approach and the preference-based approach, respectively. Similar to the MPF-based method and the preference-based method, if the number of can- didate products jYARu jis less than the number of top N recommendations ðjYARu j<NÞ, the re- maining N � jYARu j products for recommenda- tion are selected from Yu MF . The selected products are the most frequent items, which are ranked according to the frequency count of products browsed by the users in the target user’s MPF cluster and preference cluster. The most frequent items are ranked as follows. Let Y MF � M and Y MF � P denote the set of most frequent items derived from the target user’s MPF-cluster and preference cluster, respec- tively; and let fc M and fc P represent the fre- quency count of an item in Y MF � M and Y MF � P , respectively. Products in Yu MF that have not been browsed by the target user u and have not been included in Yu AR are recommended based on the ranking order of the weighted sum of their frequency counts. fcH ¼ wM � fcM þwP � fcP ð4Þ where wM and wP are the weights assigned to the MPF-based approach and the preference-based approach, respectively. The relative effects of the MPF-based ap- proach and preference-based approach on the recommendation quality may be different for different top-N recommendations; therefore, we set different values of wM and wP. We discuss the effects in detail in the next section. 4. Experiment evaluation 4.1. Experiment setup and data sets Data for the mobile web log was collected between October 2006 and January 2007. The dataset, which contained information about 1692 users, 1416 products and 184 mobile phones, was divided as follows: 80% for training and 20% for testing. The training set was also used as the data set in the preliminary analytical experiment. Specifically, 55% of the data set was used to derive recommendation rules; and 25% was used as a preliminary analytical data set to determine the number of clusters, the feature combinations, and the hybrid weights assigned to the MPF-based and preference-based meth- ods based on the quality of the recommenda- tions. There were 1353 users and 165 mobile phones in the training data set, and 339 users and 93 mobile phones in the test data set. The 8 Expert Systems c� 2010 Blackwell Publishing Ltd115c� 2010 Blackwell Publishing Ltd Expert Systems, May 2012, Vol. 29, No. 2 minimum support was set at 0.004, and the minimum confidence level was set at 0.6. 4.2. Evaluation metrics Two metrics, precision and recall, are com- monly used to measure the quality of a recom- mendation. They are also used in the field of information retrieval (Van Rijsbergen, 1979; Salton & McGill, 1986). Product items can be classified into products that customers are inter- ested in browsing, and those that they are not interested in browsing. The recommendation method then recommends products to the Figure 3: The hybrid recommendation phase. c� 2010 Blackwell Publishing Ltd Expert Systems 9116 c� 2010 Blackwell Publishing LtdExpert Systems, May 2012, Vol. 29, No. 2 customers accordingly. The recall metric indi- cates the effectiveness of a method for locating interesting products, while the precision metric represents customers’ level of interest in the recommended product items. Recall is the fraction of interesting product items that can be located. Recall ¼ number of correctly recommended items number of interesting items ð5Þ Precision is the fraction of the recommended products that customers find interesting. Precision ¼ number of correctly recommended items number of recommended items ð6Þ The items that were considered interesting to customers were the products the customers browsed in the test set. Correctly recommended items were those that matched interesting items. Because increasing the number of recommended items tends to reduce the precision and increase the recall, the F1 metric is used to balance the tradeoff between precision and recall (Van Rijs- bergen, 1979). The F1 metric, which assigns equal weights to precision and recall, is given by F1 ¼ 2 � recall � precision recall þ precision ð7Þ Each metric was computed for each user, and the average value was computed for each clus- ter. The overall average (i.e. of all users) was calculated to measure the quality of the recom- mendations. 4.3. Experiment results 4.3.1. MPF and cluster number selection Al- though we selected eight MPF in Section 3.1, the recommendation quality may not be the best if we combine all the features. Therefore, we try all possible combinations of the eight features to determine the best combination and the number of clusters. We cluster users by the K-means clustering method with Pearson’s correlation coef- ficient based on the selected features. Using the MPF-based method described in Section 3.2, we try various MPF combinations and various num- bers of clusters between two and eight. The best recommendation quality of the preliminary analy- tical data is derived by combining five features, namely the Bluetooth, card slot, flash light, java and video functions, and the number of clusters is three. Hence, we use these five features with three clusters as the parameters for MPF-based recom- mendation. 4.3.2. Mobile phone and product preference clus- ter identification Based on the results derived in the previous section, we divide the training users into three clusters according to the five selected phone features, as shown in Table 3. The Bluetooth function enables users to connect to the other Bluetooth devices, including ear- phones and notebooks. The card slot function expands a mobile phone’s data storage capacity for music, photos and movie files. The flash light function improves the quality of photographs taken in certain environments. Mobile phones with the java function can run java applications, including games; while phones with a video function are becoming increasingly popular for playing MP4 and movies in 3GP format. Based on Table 3, we can calculate the feature frequency of each cluster by considering the frequency count and the representative MPF of each cluster that are above the frequency thresh- old of 50%. The frequency count of a feature is defined as the number of users’ phones that have the feature divided by the number of users. According to the feature frequency of each cluster, the users are classified into three types of MPF. Users in cluster 0 prefer camera phones with advanced features, such as Bluetooth, card slot and flash light functions; users in cluster 1 prefer simple phones with basic java features and video functions; and users in cluster 2 prefer feature phones with Bluetooth and card slot functions for device connectivity and data storage. The preference-based method clusters users according to their product preferences, that is, products browsed by users. We cluster users into four groups based on the best recommendation 10 Expert Systems c� 2010 Blackwell Publishing Ltd117c� 2010 Blackwell Publishing Ltd Expert Systems, May 2012, Vol. 29, No. 2 quality achieved using the preliminary analyti- cal data set. We use the product category fre- quency count with at threshold of 20% to identify the characteristics of each product pre- ference cluster, as shown in Table 4. The fre- quency count of a product category is defined as the number of users that browse the product category divided by the total number of users. 4.3.3. Determining the MPF and preference weights of the hybrid recommendation sche- me The hybrid recommendation scheme is based on the hybrid weighting ratios wM and wP (wP ¼ 1 � wM) of the mobile phone and product preference clusters. Hybrid recommen- dation becomes pure preference-based recom- mendation when wM equals zero, and pure MPF-based recommendation when wM equals one. The top-N recommendations are divided into two segments. One segment is from the top-1 to the top-10 recommendations and the other is from the top-11 to the top-20 recommendations. We choose the top-5 and top-15 recommenda- tions to represent the first and second segments, respectively. The quality of the top-5 and top-15 hybrid recommendations with different MPF weights (wM) is shown in Figure 4. The best recommendation quality for the top-5 and top- 15 occurs when wM ¼ 0.9 and wM ¼ 0.6, respec- tively. We use these weights as the hybrid weighting ratios of the hybrid recommendation scheme in Section 4.3.4. 4.3.4. Evaluation of MPF-preference hybrid re- commendation methods We compare two pro- posed methods, namely, MPF-based and Hybrid MPF-Preference methods, with the other two methods, preference-based and CF methods. MPF-based method cluster users into groups based on users’ MPF and recommend products according to the association rules and most frequent items extracted from user groups. Preference-based method makes recommenda- tions based on user groups that are clustered according to the users’ product preferences.T a b le 3 : M o b il e p h o n e cl u st er cl a ss ifi ca ti o n C lu st er ID U se rs B lu et o o th F re q u en cy ( % ) C a rd sl o t F re q u en cy ( % ) F la sh li g h t F re q u en cy ( % ) J a va F re q u en cy ( % ) V id eo F re q u en cy ( % ) M o b il e p h o n e fe a tu re s P h o n e ty p e 0 4 4 9 4 4 2 9 8 3 6 8 8 2 4 0 0 8 9 4 4 2 9 8 4 3 9 9 8 Ja v a , V id eo , B lu et o o th , C a rd sl o t, F la sh li g h t C a m er a p h o n e 1 3 1 9 0 0 7 2 1 3 4 3 1 8 1 0 0 3 0 1 9 4 Ja v a , V id eo S im p le p h o n e 2 1 6 2 1 5 6 9 6 1 6 2 1 0 0 8 5 1 5 6 9 6 1 4 6 9 0 Ja v a , V id eo , B lu et o o th , C a rd sl o t F ea tu re p h o n e T o ta l 9 3 0 5 9 8 6 4 5 3 7 5 8 4 2 1 4 5 9 1 6 9 8 8 8 6 9 5 c� 2010 Blackwell Publishing Ltd Expert Systems 11118 c� 2010 Blackwell Publishing LtdExpert Systems, May 2012, Vol. 29, No. 2 Hybrid MPF-Preference recommendations are generated by a combination of the MPF-based and preference-based recommendation schemes with the hybrid weighting ratio described in Section 3.3. The hybrid weighting ratio de- scribed in Section 4.3.3 is set at wM ¼ 0.9 for the first top-N segment (top1-10) and wM ¼ 0.6 for the second top-N segment (top11–20). The CF method is a typical k-NN CF method that recommends the top-N most frequently occur- ring products of the k-NN (similar users). Be- cause the average number of users in the product clusters is 232.5( ¼ 930=4), we choose k ¼ 200 as the number of nearest neighbours. Table 5 presents the precision, recall and F1- metric evaluation of k-NN CF, Preference- based, MPF-based and Hybrid MPF-Preference methods. The F1 values of all methods are low, since the user-item matrix of our experiment data is very sparse. Although the F1 values of our proposed methods are still low, our methods can achieve better improvement over conven- tional methods. For example, as listed in Table 5, the average F1-metric of the MPF- based method is 11% better than the prefer- ence-based method. Furthermore, the average F1-metric of the hybrid MPF-Preference meth- od, which combined MPF-based and prefer- ence-based methods, is 33% better than the preference-based method. The F1-metric of the hybrid MPF-Preference, MPF-based, Prefer- ence-based and k-NN CF methods are shown in Figure 5. As shown in Figure 5, the recommendation quality of all the methods declines after the top-4 recommendations, as the number of re- commended products increases. Recall that as- sociation rule-based recommendations are based on the items users browsed previously. There are only a few recommended products because the average number of previously browsed products was 3.87. Therefore, the most frequent item recommendations are used to support the association rule recommendations if the number of recommended products is not sufficient. However, most frequent item-based recommendations are not better than associa- tion rule-based recommendations, so the recom- mendation quality deteriorates after the top-4 recommendations. 4.3.5. The effect of the hybrid method on mobile phone and product preference clusters Figure 6 shows that the mobile phone cluster 0 (camera phones with java, video, Bluetooth, card slot and flash light functions) achieves the best recommendation quality, followed by cluster 1 (simple phones with java and video functions), Table 4: Product preference cluster identification Cluster ID Users Product category 0 185 Lingerie, pants 1 336 Mobile phones, cordless phones, digital cameras 2 179 Hotels, travel coupons, food, domestic travel 3 230 Skin care, MP3, cosmetics, living products Total 930 Hybrid recommendations 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 wM F 1- m et ri c top-5 top-15 Figure 4: The weighting ratio wM of the hybrid recommendations. 12 Expert Systems c� 2010 Blackwell Publishing Ltd119c� 2010 Blackwell Publishing Ltd Expert Systems, May 2012, Vol. 29, No. 2 and cluster 2 (feature phones with java, video, Bluetooth and card slot functions). Among all the phone types, camera phones with the flash light feature yield the best recommendation quality. The owners of camera phones like to browse for digital cameras and travel products because they like to travel and take photo- graphs. We also evaluate the effect of the hybrid method on the recommendation quality of MPF-based clusters. Figure 6 shows that the recommendation quality of the hybrid method (hybrid 0–2) is better than that of the MPF- based method (mphone 0–2) for each MPF- based cluster. In other words, the effect of combining MPF-based recommendations with preference-based recommendations is positive. From Figure 7, we observe that product cluster 0 (lingerie, pants and skincare products) achieves the best recommendation quality in terms of product preferences, followed by pro- duct cluster 2 (hotels, travel coupons, food and domestic travel), product cluster 1 (mobile phones, digital cameras, cordless phones and notebooks), and product cluster 3 (skincare, mp3, cosmetics and consumer products). Users who prefer lingerie and underwear products receive better quality recommendations than users who prefer other products. We also Table 5: Evaluation of k-NN CF, Preference-based, MPF-based and hybrid methods TopN k-NN CF Preference-based MPF-based Hybrid MPF-preference Precision Recall F1 Precision Recall F1 Precision Recall F1 Precision Recall F1 2 0.015 0.004 0.006 0.153 0.085 0.092 0.161 0.088 0.099 0.176 0.100 0.110 4 0.026 0.017 0.017 0.104 0.113 0.089 0.122 0.128 0.106 0.140 0.157 0.125 6 0.036 0.055 0.035 0.080 0.124 0.080 0.096 0.156 0.100 0.113 0.186 0.118 8 0.039 0.098 0.045 0.072 0.146 0.081 0.079 0.165 0.091 0.092 0.195 0.106 10 0.035 0.107 0.044 0.063 0.156 0.076 0.067 0.172 0.082 0.081 0.212 0.100 12 0.030 0.109 0.040 0.057 0.165 0.072 0.058 0.178 0.075 0.072 0.221 0.094 14 0.027 0.111 0.036 0.051 0.171 0.066 0.051 0.180 0.069 0.064 0.227 0.087 16 0.023 0.112 0.033 0.046 0.174 0.062 0.045 0.181 0.063 0.058 0.236 0.081 18 0.021 0.112 0.030 0.042 0.179 0.059 0.043 0.197 0.062 0.053 0.244 0.077 20 0.019 0.112 0.028 0.040 0.187 0.057 0.041 0.210 0.061 0.049 0.248 0.073 Average 0.027 0.084 0.032 0.071 0.150 0.073 0.076 0.165 0.081 0.090 0.203 0.097 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Top-N F 1- m et ri c Hybrid MPF-Preference MPF-based Preference-based k-NN CF Figure 5: Evaluation of hybrid, MPF-based, preference-based and k-NN CF methods. c� 2010 Blackwell Publishing Ltd Expert Systems 13120 c� 2010 Blackwell Publishing LtdExpert Systems, May 2012, Vol. 29, No. 2 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Top-N F 1- m et ri c hybrid0 mphone0 hybrid1 mphone1 hybrid2 mphone2 Figure 6: Effect of the hybrid method on the recommendation quality for mobile phone clusters. 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Top-N F 1- m et ri c hybrid0 product0 hybrid2 product2 hybrid1 product1 hybrid3 product3 Figure 7: Effect of the hybrid method on the recommendation quality for product clusters. 14 Expert Systems c� 2010 Blackwell Publishing Ltd121c� 2010 Blackwell Publishing Ltd Expert Systems, May 2012, Vol. 29, No. 2 evaluate the effect of the hybrid method on the recommendation quality of preference-based clusters. Figure 7 shows that the recommenda- tion quality of the hybrid method (hybrid 0–3) is better than that of the preference-based method (product 0–3) for each preference-based cluster. Hence, combining the preference-based method with the MPF-based method can improve the quality of recommendations. 5. Conclusion We have proposed a MPF-based hybrid method to resolve the sparsity issue of the typical CF method in mobile environments. We assume that the MPF preferred by users indicate their interest in particular m-commerce products and services; thus, they can be used to group users with similar interests. The hybrid method com- bines the MPF-based method and preference- based method, which uses association rule mining to extract recommendation rules from user groups and make recommendations. Experiment results show that the quality of MPF-based recommendations is better than that of the preference-based method and the typical k-NN CF scheme. However, the hybrid method outperforms the MPF-based, prefer- ence-based and the typical k-NN CF methods. According to the cluster analysis results, mobile phone cluster 0 (camera phones with Bluetooth, card slot, flash light, java and video functions) yields the best recommendation quality among the mobile phone clusters; product cluster 0 (lingerie, pants and skincare products) achieves the best recommendation quality in terms of product preference clusters. The hybrid method, which combines recommendations derived by the MPF-based and preference-based methods, im- proves the recommendation quality of MPF- based clusters and preference-based clusters. Acknowledgements This research was supported in part by the National Science Council of the Taiwan under grant NSC 96-2416-H-009-007-MY3. References AGRAWAL, R., T. IMIELIŃSKI and A. SWAMI (1993) Mining Association Rules between Sets of Items in Large Databases, Washington, DC, USA: ACM SIGMOD International Conference on Manage- ment of Data, 207–216. BRUNATO, M. and R. BATTITI (2003) PILGRIM: A location broker and mobility-aware recommenda- tion system. Proceedings of the First IEEE Interna- tional Conference on Pervasive Computing and Communications (PerCom 2003). Fort Worth, TX, USA, pp. 265–272. CHAE, M. and J. KIM (2003) What’s so different about the mobile Internet?, Communications of the ACM, 46, 240–247. CHEN, M.S., J. HAN and P.S. YU (1996) Data Mining: an overview from a database perspective, IEEE Transactions on Knowledge and Data Engineering, 8, 866–883. CHO, Y.B., Y.H. CHO and S.H. KIM (2005) Mining changes in customer buying behavior for collabora- tive recommendations, Expert Systems with Applica- tions, 28, 359–369. CHOI, J.Y., H.S. SONG and S.H. KIM (2007) MCORE: a context-sensitive recommendation system for the mobile Web, Expert Systems, 24, 32–46. HILL, W., L. STEAD, M. ROSENSTEIN and G. FURNAS (1995) Recommending and Evaluating Choices in a Virtual Community of use. Proceedings of the SIG- CHI Conference on Human Factors in Computing Systems, Denver, CO, USA: ACM Press=Addison- Wesley Publishing, 194–201. KIM, C.Y., J.K. LEE, Y.H. CHO and D.H. KIM (2004) Viscors: a visual-content recommender for the mobile web, IEEE Intelligent Systems, 19, 32–39. KO, S.M., Y.G. JI and D. KIM (2007) Understanding influence of mobile internet services on life behavior of mobile users, Lecture Notes in Computer Science, 4553, 944–953. LEE, H.J. and S.J. PARK (2007) MONERS: a news recommender for the mobile web, Expert Systems with Applications, 32, 143–150. LEE, W.-P. (2004) Applying domain knowledge and social information to product analysis and recom- mendations: an agent-based decision support sys- tem, Expert Systems, 21, 138–148. LIN, Q., Y. CHEN, J. CHEN and Y. CHEN (2003) Mining inter-organizational retailing knowledge for an alli- ance formed by competitive firms, Information and Management, 40, 431–442. LINDEN, G., B. SMITH and J. YORK (2003) Amazon. com recommendations: item-to-item collaborative filtering, IEEE Internet computing, 7, 76–80. LING, C., W. HWANG and G. SALVENDY (2006) Diver- sified users’ satisfaction with advanced mobile phone c� 2010 Blackwell Publishing Ltd Expert Systems 15122 c� 2010 Blackwell Publishing LtdExpert Systems, May 2012, Vol. 29, No. 2 features, Universal Access in the Information Society, 5, 239–249. LIU, D.-R. and Y.-Y. SHIH (2005) Integrating AHP and data mining for product recommendation based on customer lifetime Value, Information & Manage- ment, 42, 387–400. MACQUEEN, J. (1967) Some Methods for Classification and Analysis of Multivariate Observations. Fifth Ber- keley Symposium on Mathematical Statistics and Probability, Berkeley, CA: Statistical Laboratory of the University of California, Berkeley, University of California Press, 281–297. MARTINEZ, L., L.G. PEREZ and M. BARRANCO (2007) A multigranular linguistic content-based recommen- dation model, International Journal of Intelligent Systems, 22, 419–434. MOONEY, R.J. and L. ROY (2000) Content-based book recommending using learning for text categorization, in Proceedings of the Fifth ACM Conference on Digital Libraries, San Antonio, TX, USA: ACM, 195–204. NGAI, E.W.T. and A. GUNASEKARAN (2007) A review for mobile commerce research and applications, Decision Support Systems, 43, 3–15. OJANPERA, T. (2006) Convergence transforms internet, Wireless Personal Communications, 37, 167–185. OMRAN, M.G.H., A.P. ENGELBRECHT and A. SALMAN (2007) An overview of clustering methods, Intelligent Data Analysis, 11, 583–605. PUNJ, G. and D.W. STEWART (1983) Cluster analysis in marketing research: review and suggestions for application, Journal of Marketing Research, 20, 134–148. RESNICK, P., N. IACOVOU, M. SUCHAK, P. BERG- STROM and J. RIEDL (1994) GroupLens: an open architecture for collaborative filtering of netnews, in Proceedings of the 1994 ACM Conference on Compu- ter Supported Cooperative Work, Chapel Hill, NC, USA: ACM, 175–186. SALTON, G. and M.J. MCGILL (1986) Introduction to Modern Information Retrieval, NY, USA: McGraw- Hill. SARWAR, B., G. KARYPIS, J. KONSTAN and J. RIEDL (2000) Analysis of recommendation algorithms for e-commerce, in Proceedings of the Second ACM Conference on Electronic Commerce, Minneapolis, MN, USA: ACM, 158–167. SCHAFER, J., J. KONSTAN and J. RIEDL (2001) E- commerce recommendation applications, Data Mining and Knowledge Discovery, 5, 115–153. SHARDANAND, U. and P. MAES (1995) Social information filtering: algorithms for automating ‘‘word of mouth’’, in Proceedings of the SIGCHI conference on Human factors in computing systems, Denver, CO, USA: ACM=Addison-Wesley Publish- ing, 210–217. VAN RIJSBERGEN, C.J. (1979) Information Retrieval, MA, USA: Butterworth-Heinemann Newton. VENKATESH, V., V. RAMESH and A.P. MASSEY (2003) Understanding usability in mobile commerce, Com- munications of the ACM, 46, 53–56. VIRVOU, M. and A. SAVVOPOULOS (2007) An intelli- gent TV-shopping application that provides recom- mendations. ICTAI 2007, 19th IEEE International Conference on Tools with Artificial Intelligence, Patras, Greece, pp. 366–373. YANG, W.S., H.C. CHENG and J.B. DIA (2008) A location-aware recommender system for mobile shopping environments, Expert Systems with Appli- cations, 34, 437–445. The authors Chuen-He Liou Chuen-He Liou is a PhD student of the Institute of Information Management at the National Chiao Tung University of Taiwan. He received the MS degree in Physics from the National Taiwan Normal University. His research inter- ests include information systems, recommender systems, mobile commerce, channel marketing and strategic planning. Duen-Ren Liu Dr Duen-Ren Liu is a Professor of the Institute of Information Management at the National Chiao Tung University of Taiwan. He received the BS and MS degrees in Computer Science from the National Taiwan University and his PhD in Computer Science from the University of Minnesota. His research interests include information systems, knowledge engineering and management, workflow systems, electronic commerce and recommender systems. 16 Expert Systems c� 2010 Blackwell Publishing Ltd123c� 2010 Blackwell Publishing Ltd Expert Systems, May 2012, Vol. 29, No. 2 
work_mcnxmtrdrrgu5kghkduzzfdf54	----	Eigenspace method for spatiotemporal hotspot detection Article DOI: 10.1111/exsy.12088 Eigenspace method for spatiotemporal hotspot detection Hadi Fanaee-T and João Gama Laboratory of Artificial Intelligence and Decision Support (LIAAD), University of Porto INESC TEC, Rua Dr Roberto Frias, Porto, Portugal E-mail: hadi.fanaee@fe.up.pt Abstract: Hotspot detection aims at identifying sub-groups in the observations that are unexpected, with respect to some baseline information. For instance, in disease surveillance, the purpose is to detect sub-regions in spatiotemporal space, where the count of reported diseases (e.g. cancer) is higher than expected, with respect to the population. The state-of-the-art method for this kind of problem is the space–time scan statistics, which exhaustively search the whole space through a sliding window looking for significant spatiotemporal clusters. Space–time scan statistics makes some restrictive assumptions about the distribution of data, the shape of the hotspots and the quality of data, which can be unrealistic for some non-traditional data sources. A novel methodology called EigenSpot is proposed where instead of an exhaustive search over the space, it tracks the changes in a space–time occurrences structure. The new approach does not only present much more computational efficiency but also makes no assumption about the data distribution, hotspot shape or the data quality. The principal idea is that with the joint combination of abnormal elements in the principal spatial and the temporal singular vectors, the location of hotspots in the spatiotemporal space can be approximated. The experimental evaluation, both on simulated and real data sets, reveals the effectiveness of the proposed method. Keywords: hotspot detection, spatiotemporal data, eigenspace, SVD, outbreak detection 1. Introduction Eigenspace techniques are very popular ones, which encompass many applications in data mining, signal processing, infor- mation retrieval and other domains. A famous instance of such an application relates to the current success of Google search engine. As described in an article entitled $25000000000 Eigenvector (Bryan & Leise, 2006), Google search engine is largely attributed to the eigenspace techniques. Another success story is related to the application of the singular value decom- position (SVD) (Klema & Laub, 1980) in collaborative filtering. In the year 2008, the BellKor team won the $1000000 prize for the improvement of the system of the Netflix movie recommendation. The report later (Bell et al., 2007) stated that SVD was the key data analysis tools that they used. However, the application of the eigenspace techniques is not restricted to the aforementioned instances. There are several different examples in other areas and sciences. For instance, in face recognition (Turk & Pentland, 1991), the specific set of the largest eigenvectors can be used to approximate the images of the human face. In structural engineering, both the eigenvalue and eigenvectors are used to estimate the vibration of structures. In control engineering, the eigenvalues of the linear system are used to assess the stability and response of the system (wikibooks, 2012). Nevertheless, despite the merits of the eigenspace tech- niques, they have not been applied yet to some potential problems, such as hotspot detection. Hotspot detection that may come with different terminologies, such as outbreak detection, cluster detection or event detection, is somehow related to the clustering and anomaly detection; however, it is distinct from these two. In clustering, the entire data set is partitioned into some groups, but in the anomaly detection, the anomalous points are searched for and sought after. Hotspot detection addresses the same problem but with this difference that anomalous instances are recognized given some baseline information. In other words, looking into the data set, everything might seem normal; however, when the cases along the baseline are considered, some points might be considered unexpected. A realistic scenario of the hotspot detection is in disease surveillance. Suppose that we have the population of different postal codes, during a range of years, as the baseline information and the count of the reported diseases in a range of postal codes, through- out different years as the cases data set. The goal is to detect those spatiotemporal regions that contain unexpected counts. For instance, the output such as zones S1, S2 and S3 during the years T1 to T5 might be considered a spatiotemporal hotspot. The detection of such hotspots enables the officials to better understand their target of interest for essential medical care and preventive measures. The current methods for hotspot detection are twofold: clustering-based techniques and the scan statistics-based ones. Clustering-based techniques, such as (Levine, 2006), infer some thresholds from the population data and then apply the thresholds for clustering of data points in the cases set. Their prominent benefit, as opposed to the other methods, is that they provide the exact shape of the clusters. However, handling complex data, such as spatiotemporal data, is not straightforward for these techniques. Besides, the clustering methods do not consider chance and randomness issues, © 2014 Wiley Publishing Ltd454 Expert Systems, June 2015, Vol. 32, No. 3 which are very important in sensitive applications, such as security and public health. Moreover, there is not a standard clustering method for hotspot detection, which is widely accepted by the community. The methods are mostly outspread and diverse, in terms of the technical details. The clustering methods also require restrictive input parameters, which make their usage limited in automatic settings. The second group of techniques that relies on scan statistics are widely used and accepted by the epidemiological com- munity. These groups of techniques exhaustively scan the whole space to find interesting spatial and spatiotemporal clusters. A specific statistic is computed for each possible window, and then, potential clusters are sorted, on the basis of the obtained statistics. Thereafter, the statistical signi- ficance of the top-k clusters is simulated via a Monte Carlo technique. Because these methods scan an entire space, they are extremely computationally expensive. Spatial scan sta- tistics (Kulldorff, 1997) requires computation time of O(N3) and space–time scan statistics (STScan) (Kulldorff & Nagarwalla, 1995; Kulldorff, 1997, 1999; Kulldorff et al., 1998) requires O(N4). Some recent efforts are made to reduce this complexity. For instance, (Agarwal et al., 2006) propose a method that requires O 1εN 2Log2N � � for spatial scan, which is more efficient than O(N3). However, the minimum complexity for STScan in the best condition has not reached less than O(N3). This high computational cost practically has restricted their use in real-time appli- cations or large-scale data sets. Besides, scan statistics-based techniques are highly associated with the strong parametric model assumptions (e.g. Poisson or Gaussian counts) (Neill, 2006). These assumptions mitigate the performance when the models are incorrect for non-traditional data sources. Additionally, scan statistics-based methods are not efficient for detection of irregular shape clusters (Duczmal & Assunção, 2004; Tango & Takahashi, 2005) apart from the circles (spatial scan) and cylinders (space–time scan). They also assume that data is presented in a high quality format, hence are vulnerable against the noise and outliers (Neill & Sabhnani, 2007). Our proposed method is a solution to some of the afore- mentioned issues, in scan statistics-based methods. An effi- cient method is proposed (linear with both space and time dimensions) for approximation of hotspots in the spatio- temporal space, without the need for exhaustive search. Instead of looking for deviations in the assumed parametric model, we track changes in the space–time occurrences structure, using the eigenspace techniques. This approach enables us to detect irregular shape hotspots from even noisy data sets, without any prior knowledge about the data nature or hotspot characteristics. To the best of our knowledge, this problem has not already been addressed by other researchers. Our approach also differs from those of ones that focus on the improvement of scan statistics-based methods efficiency (e.g. Neill & Moore, 2004; Agarwal et al., 2006; Neill & Cooper, 2010). We do not improve the efficiency of scan statistics-based methods; rather, we propose and examine a new methodology, which follows a different aim. Hence, this is not an ‘apples to apples’ comparison, as both groups of approaches have inherent differences and subsequently their own applications. STScan can be more helpful for retrospective and sensitive applications, when some prior knowledge exists about the nature of hotspot and data. On the other hand, our presented approach focuses more on real-time applications, where neither the nature of data nor the hotspot characteristics is known in advance. In such circumstances, a computationally feasible approximation method that rapidly identifies the alarming areas without any prior knowledge might be very useful. The rest of the paper is organized as follows. Section 2 describes the problem, the proposed solution and algorithm, as well as an illustrative example. Section 3 includes an experimental evaluation and results of the simulation study and the real case study. The last section concludes the exposition presenting the final remarks. 2. The proposed approach 2.1. The problem Given a spatiotemporal count matrix for the cases, needed for the detection of those spatiotemporal regions (hotspots) that seem unexpected, given the baseline spatiotemporal matrix. Each cell in each matrix represents a count corresponding to a specific region and time. In particular, for disease outbreak detection, each cell in the baseline matrix represents the population corresponding to a region in a specific period. Each cell in the matrix cases also represents the count of reported disease in a specific region, within a given time period, as well. The purpose is to determine those sub-groups of the spatio- temporal space whose reported cases are unexpected. A baseline method that can be applied to the problem is to compute the ratio of the cases to the population for all possible spatiotemporal regions (each cell in the spatiotemporal matrix) and then compute the z-score of the ratios. Then, the null hypothesis Ho: There is no hotspot is rejected, in case some spatiotemporal regions with z-score greater than a threshold are found. This approach theoretically and practically, as will be illustrated later, imposes too many false alarms, because for a n×m matrix, it’s required to perform n×m comparison tasks. In this paper, an approach that performs only n+m comparisons is clearly proposed. An unsupervised method may be needed to use with the clustering on the ratios. However, it suffers from the same problem of the baseline method (it requires n×m comparisons and not n+m). Besides, the requirement for determination of appropriate cut-point or number of clusters adds more complexity and user involvement to the system. We are interested in developing a system, which has the following characteristics: (1) does not require any input parameter; and (2) weighs all the possible hotspots, on the basis of a standard metric such as statistical significance (p-value). The benefit is this that the output can be compared with relevant systems or methods. The alpha threshold is also easy to estimate (usually alpha= 0.01 or 0.05). © 2014 Wiley Publishing Ltd Expert Systems, June 2015, Vol. 32, No. 3 455 2.2. The method In this section, the logic used in the method deployment is thoroughly described. Assume that we have two identical n× m matrices B (baseline) and C (cases) such that n be the number of components in the spatial dimension and m be the number of components in the temporal dimension. The SVD of the n× m matrix is a factorization of the form M =UΣV*. The n columns of U and the m columns of V are called the left-singular vectors and right-singular vectors of the matrices, respectively. The left- singular vectors correspond to spatial dimension, whilst the right-singular ones correspond to the temporal dimension. In order to clarify and elaborate, a new terminology spatial singular vector is used along with temporal singular vector that, respectively, refers to the principal left-singular vector and the principal right-singular one. Note that we take only the singular vector corresponding to the largest eigenvalue for the com- parison, because the first principal singular vector accounts for major directions of data. Hence, it explains or extracts the largest part of the inertia of the data (Abdi & Williams, 2010). Now, let us denote the spatial singular vector of baseline (B) and cases (C), respectively, with → sb = (sb1, sb2, …, sbn) and → sc = (sc1, sc2, …, scn). Then, lets denote temporal singular vector of B and C, respectively, with → tb = (tb1, tb2, …, tbm) and → tc = (tc1, tc2 …, tcm). If we hypothetically assume that B= C, then → sb ¼ →sc and →tb ¼ →tc. In this condition, the angles between → sb and → sc and between → tb and → tc would be equal to zero. Now assume that some change occurs in C, and this change corresponds to a specific region and period. Therefore, the matrices are no longer identical, and subsequently, the angles between their singular vectors rise up in value. From this angle change, we can only infer that some changes occur, but we do not know what sub-group of data is affected by this change. If we could identify those vector elements from → sc and → tc that caused this change, we would be able to identify the spatial and temporal components of the affected area. For instance, assume that through a hypothetical method we could identify that sc1 from the → sc and tc1 from → tc correspond to the affected area. If we remove the region corresponding to sc1 and time related to tc1 from both baseline and cases data sets, the matrices should again become identical. Hence, we would have the angles between the pair singular vectors equal to almost zero. Here, (sc1, tc1) is called hotspot, and sc1 and tc1 are, respectively, called the spatial and temporal components of the hotspot. The process of finding these components is also called hotspot detection. Note that in this work, the angle between the singular vectors is not computed. In the previous texts, the angle concept is only used to explain the rationale behind the proposed method. Some assumptions were made earlier in the text, which were only for simplification of explanation. In practice, we rarely find two identical baseline and cases matrix. However, we are able to assume that in a normal condition, where no hotspot exists, both baseline and cases set should have a same space– time occurrences structure. In this case, the pair singular vectors of baseline and cases sets, regardless of the data distribution should stay in a constant distance. Now, if a hotspot starts to grow in the cases set, this change can be directly observed from the changes in the singular vectors elements. In such cases, some distances between the singular vector elements become abnormal for elements corresponding to the affected areas in both the spatial and temporal dimension. We exploit this idea to develop our algorithm for hotspot detection. According to the aforementioned explanation, two kinds of tools are required, a tool for obtaining the singu- lar vectors of a non-square matrix and a process control tool for monitoring distances between singular vector elements. SVD and statistical process control are two powerful techniques, which for several years have been successfully applied to many problems in different domains and are state-of-the-art methods to these requirements. In the next section, we explain how these techniques are exploited in the deployment of the solution. © 2014 Wiley Publishing Ltd456 Expert Systems, June 2015, Vol. 32, No. 3 2.3. The EigenSpot algorithm In this section, our proposed algorithm EigenSpot is explained in detail. The inputs of the algorithm 3 are n × m matrices for the baseline and cases where n represents the number of regions and m represents the number of temporal instants. We start by decomposing both matrices, using one-rank SVD. The one-rank SVD gives us the principal singular vector corresponding to the spatial and temporal dimensions (lines 1–2). The reason why the low-rank SVD is applied versus the full-rank SVD is that our approach requires only the principal singular vector for each matrix. The full-rank SVD is a more expensive method, because of the fact that a N × N matrix requires O(N3), whilst the low-rank SVD requires O(kN2) where in our case k = 1, and therefore, we require only O(N2) for each one-rank SVD. The principal singular value accounts for the major directions of data in both cases and baseline; therefore, it is appropriate for matching purposes. In the next step, we subtract each element of the pair singular vectors together (lines 3–8). If we denote the spatial singular vector for baseline with (sb1, sb2, …, sbn) and spatial singular vector for cases with (sc1, sc2, …, scn), the subtract vector would be →ds = (ds1 = sc1 � sb1, ds2 = sc2 � sb2, …, dsn = scn � sbn). Similarly for the temporal dimension, we have →dt = (dt1 = tc1 � tb1, dt2 = tc2 � tb2, …, dtm = tcm � tbm). Subsequently, in order to identify the spatial and temporal components of the hotspot, a z-score control chart is applied on vectors →ds and →dt with significant level α. To do so, the standardized vector of z-scores is first computed for ds and dt. Thereafter, we obtain the equivalent two- tailed p-value for each z-score. Finally, those components of →ds and →dt that obtain p-value lower than α are considered abnormal. Finally, a joint combination of all spatial and temporal components to the original space gives us the approximation of hotspots. For instance, assume that → sb ¼ 0:25; 0:10; 0:75; 0:20ð Þ be the spatial singular vector of baseline and →sc ¼ 0:30; 0:90;ð 0:80; 0:15Þbe the spatial singular vector of cases. Each element in the spatial singular vector corresponds to a specific region. For instance, 0.30 and 0.25 in the first element corresponds to region 1. Similarly, the second, third and the fourth element corresponds to the region 2, 3 and 4, respectively. The angle between the two singular vectors →sb and →sc is equal to 68° in this example. This angle does not tell us what elements of singular vector have contributed to this difference. However, if in the aforementioned example, we remove region 2 from the system, we have two vectors → sb ¼ 0:25; 0:75; 0:20ð Þ and →sc ¼ 0:30; 0:80; 0:15ð Þ where the angle between them is equal to 0.09, which is almost equal to zero. Region 2 in this example is equivalent to the spatial component of the hotspot. In order to identify the region 2 in this example, a z-score control chart is applied on the subtract vector → ds ¼ 0:25 � 0:30; 0:10�ð 0:90; 0:75 � 0:80; 0:20 � 0:15Þ ¼ �0:05; �0:80; �0:05; 0:05ð Þ. Afterwards, we compute the standardized z-scores of the subtract vector, which in this case is zds = (0.4119, � 1.4893, 0.4119, 0.6654). As shown, z-score of �1.4893 is equivalent to the left-tailed p-value of 0.06. If we define α = 0.10, region 2 would be identified as hotspot spatial component. This is because its p-value is lower than 0.10. However, if we define α = 0.05, region 2 is not detected as hotspot spatial component. 2.3.1. The algorithm complexity If we assume that we have N regions and N time instants, EigenSpot requires two O(N2) for two one-rank SVD for cases and baseline matrices, and two O(N) for elements matching corresponding spatial and temporal dimensions. This makes the EigenSpot require only O(2N2) + O(2N) = O(N2), which is much more efficient than the STScan. Because, the STScan requires O(NlogN) and O (N2logN) for finding the relevant time and space cylinders and O(N4) for finding the space–time cylinders as intersections of space and time cylinders (Assunção et al., 2004). Therefore, a single execution of the STScan procedure takes O(NlogN) + O(N2logN) + O(N4) = O(N4). 2.4. Illustrative example Figure 1 demonstrates an illustrative example of how a hotspot can be identified by the EigenSpot algorithm. We are given two sets of baseline and cases that encompass three regions within four time windows. Each region can be a postal code or a city. In addition, each temporal window can be a period, such as a year (e.g. T1 = 2010). If we represent these two sets as a matrix, we have two sets of 3 × 4 matrices such that each cell represents the count. For instance, b11 represents the population of region 1 at time window T1, and c32 represents the count of reported disease in region 3 within the temporal window T2. The shaded area in the cases matrix (conjunction of third row with first–second columns) is the hotspot of interest that is required to be detected by the method. As demonstrated, the principal singular vector corresponding to the spatial and temporal dimensions is obtained via one-rank SVD. As a result, we have two singular vectors corresponding to the spatial and temporal dimensions for each set. In the next step, we subtract elements of each singular vectors pairs together. Therefore, we would have two vectors →dt and → ds that represent subtract vectors for the temporal and spatial dimensions, respectively. As demonstrated, dt has four elements, and ds has three elements, each of which corresponds to the original regions and temporal windows (e.g. dt1 corresponds to T1, and ds1 corresponds to region 1). In the next step, we apply a z-score control chart with significance level α (e.g. α = 0.05) on both of these vectors to identify their abnormal elements. As it is hypothetically shown in the example, T1 and T2 are identified as temporal hotspot components, and region 3 is identified as the spatial hotspot component. We only need to combine spatial components with temporal components to approximate the hotspots in the spatiotemporal space. © 2014 Wiley Publishing Ltd Expert Systems, June 2015, Vol. 32, No. 3 457 As shown, the identified hotspot of region 3, T1, T2 is equivalent to the shaded area in cases matrix (the target). 3. Experimental evaluation In this section, the effectiveness of our proposed approach is assessed, through an experimental study. The data sets used in evaluation of hotspot detection techniques are usually threefold (Buckeridge et al., 2005): (1) wholly simulated: both baseline and cases and hotspots are simulated; (2) semi-realistic: baseline is taken from a real population, but cases and hotspots are simulated; and (3) real data: both baseline and cases are real and hotspots are verified by a domain specialist. In this paper, we evaluate the proposal, using the latter two strategies. We evaluate the algorithm performance, via the simulation study (Section 3.1) and a real-world data (Section 3.2). All experiments are conducted on a PC with Intel Core 2 Duo CPU and 3 GB Ram. We use MATLAB 7 for the algorithm implementation and experiments and SatSan 9.2 (Kulldorff, 2012b) for expe- rimenting with STScan. Our method is compared with two other techniques, including the STScan and a baseline method. STScan (Kulldorff & Nagarwalla, 1995; Kulldorff, 1997, 1999; Kulldorff et al., 1998) exhaustively moves a varying radius and height cylinder over the whole spatiotemporal space. The height of the cylinder represents the time dimension, and the surface corresponds to the space dimension. Furthermore, it scores each possible cylinder, on the basis of likelihood ratio statistics. Next, it sorts cylinders on the basis of an order of the highest to the lowest score. Finally, a randomization test is performed for obtaining the cylinders statistical significance. The cylinder whose p-value is lower than α (e.g. 0.05) is returned as hotspots. In the baseline method, we compute the ratio of the count of cases to the corresponding population for each matrix cell; then, we compute the z-score of the obtained ratios and obtain the p-value from z-score. Afterwards, we signal an alarm, when p-value for a cell goes lower than α. 3.1. Simulation study Here, we describe how the simulated data is generated and subsequently present the obtained result. 3.1.1. Data generation We generate 1500 sets of semi-real data, on the basis of the extracted baseline data set from (Kulldorff et al., 1998). The baseline set includes the spatiotemporal distribution of population in New Mexico, USA, during 1973–1991. In order to simulate the cases count, we initially obtain the maximum likelihood of the parameter of the Poisson distribution, λ from the first year of the baseline set. Let the vector of counts for the first year be (c1, c2,.., ci) where 1 ≤ i ≤ n (n: number of spatial items). λ simply can be obtained by computing the means of the vector. Then, we multiply λ by a fixed constant of 1.2% for subsequent years (1.2% is the average population growth rate). Next, we generate random numbers from the Poisson distribution with corresponding estimated parameters for each year. In order to inject the hotspot into the cases, we select a matrix window with size H × H (hotspot size) and multiply the counts inside the window by a fixed value of I (hotspot impact). We then vary H from 1 to 5 and select I from 1.5, 2 to 2.5. Because we generate data sets on the basis of the random numbers, we generate 100 data sets for each setting to reduce the effect of randomness. The next section explores the evaluation results. 3.1.2. Performance evaluation Hotspot detection can be considered a binary classification problem, because the detection approach marks each spatiotemporal window with hotspot or non-hotspot. However, in any approach, we determine a decision threshold to distinct hotspots from non-hotspots. Determination of this threshold becomes more important in sensitive applications, such as security and Figure 1: EigenSpot algorithm, an illustrative example. The goal of the approach is the identification of the shaded area in the cases matrix. The values c and b in the baseline and cases matrix are counts corresponding to a spatiotemporal window. The process is composed of the following four steps: (1) matrix decomposition; (2) subtraction of pair singular vectors elements; (3) applying the z-score control chart on the subtract vector; and (4) combining the spatial and temporal hotspots components. © 2014 Wiley Publishing Ltd458 Expert Systems, June 2015, Vol. 32, No. 3 public health. For such applications, the evaluation of methods has to be evaluated within different ranges of decision thresholds. The receiver operating characteristic (ROC) curve (Zweig & Campbell, 1993) is a widely accepted method for such evaluation tasks. However, because of two reasons, ROC curve cannot be used as an appropriate strategy for the evaluation in this simulation study. On one hand, we want to evaluate the method performance on 100 random data sets for each 15 setting. Therefore, we have 1500 data sets, which require the analysis of 1500 ROC curves, which is infeasible. We also cannot reduce the number of data sets to one, because we are generating random sets, and if we rely only on one data set, then our results would be highly dependent on the chance and randomness. At the first glance, the area under the ROC curve (AUC) seems to be an appropriate choice, as the AUC does not have user-defined parameters. Besides, it is a summarized scalar and seems to be appropriate for mass comparison of the methods. However, the main criticism about use of AUC in applications, such as hotspot detection and outbreak detection is that AUC considers all thresholds equal, which is not true in many applications. In practice, in sensitive applications, such as epidemiology where we deal with the human lives, we are not interested in knowing how a method performs in high alpha values. The operational and practical p-value used is always low values. In other words, the alpha of interest is not between 0 and 1, rather is limited to lower values. Besides, AUC as a summarized scalar hides the real ROC curve behind the evaluation. In fact, AUC can give potentially misleading results if ROC curves cross (Hand, 2009). Some detailed criticisms against AUC can be found in (Lobo et al., 2008; Hand, 2009; Hand & Anagnostopoulos, 2013). For this reason, instead of AUC, we opt to use an averaging strategy for operation thresholds (Wong et al., 2005). We compute the average accuracy for a range of operational significance levels, such as alpha from 0.20 to 0.01 for each data set and then the average obtained values for all 100 data sets for each setting. The range of alpha is obtained as follows: We vary z-score from 1.28 to 3 (equivalent to two-tailed p-value of 0.2005–0.0027) and then increase z-score 0.1 in each loop. We compare our method performance against both the STScan and the baseline approach, via control chart on ratios, described in Section 2.1. The accuracy of methods in the identification of simulated hotspots is used as the criterion for the performance evaluation. The results are presented in Table 1. As seen, EigenSpot presents a better performance in almost all settings, except low-impact hotspots. The baseline method also as expected, because of the high rate of false positives, presents the lowest accuracy. The superiority of EigenSpot over STScan possibly relies on two reasons. One reason is related to the inherent methodological difference between the EigenSpot and STScan. STScan search the whole space to find some spatiotemporal windows that the data distribution inside them has some deviation to the standard distribution models (e.g. Poisson). This strict assumption makes this approach less effective, when the data in each of sets does not exactly follow the standard distribution model or some deviation occurs by the chance. EigenSpot instead of putting this strict restriction search for changes in the occurrence patterns and therefore is less sensitive to the deviations in data distri- bution. The second reason could be that the EigenSpot is a shape-free method and does not search for a particular shape hotspot, whilst STScan looks for specific shape hotspots. Some accuracy loss in STScan relates to different shapes of the simulated hotspot. STScan looks for cylinder-shape hotspots, whilst the simulated hotspots are in fact cubic. The results also show the performance of each method against noise. We intentionally design some low-impact and size settings for evaluating the ability of the methods in handling noise and outliers. A low-size and low-impact region such as impact of 1.5 and size 1 × 1 more seem to be an outlier or anomaly, rather than a realistic hotspot. Therefore, we expect that the methods do not detect that region as hotspot and ignore that. In other words, the detection of such hotspot shows how a method wrongly identifies the outliers and noise as hotspots. Hence, the lower accuracy in this setting reveals the better performance of the method in dealing with noise and outliers. Because the EigenSpot is a spectral method, it definitely ignores such outliers and does not report them as hotspots, whilst STScan is vulnerable against such circumstances. For this reason, it presents a higher performance for low-size and impact Table 1: The mean accuracy for 173 α in the range of 0.20–0.01 averaged for 100 data sets Size Method Impact 1 × 1 2 × 2 3 × 3 4 × 4 5 × 5 EigenSpot 1.5 0.7011 0.7670 0.8124 0.8574 0.8263 Baseline 1.5 0.7270 0.7417 0.7523 0.7663 0.7669 STScan 1.5 0.7966 0.7984 0.8008 0.8030 0.8005 EigenSpot 2.0 0.8751 0.9588 0.9588 0.9492 0.9498 Baseline 2.0 0.7259 0.7453 0.7510 0.7662 0.7741 STScan 2.0 0.8034 0.8130 0.8171 0.8273 0.8267 EigenSpot 2.5 0.9393 0.9718 0.9725 0.9675 0.9555 Baseline 2.5 0.7321 0.7511 0.7588 0.7783 0.7879 STScan 2.5 0.8069 0.8314 0.8578 0.8629 0.8723 Bold items implies the outperforming method in the corresponding impact and size. © 2014 Wiley Publishing Ltd Expert Systems, June 2015, Vol. 32, No. 3 459 regions. In the experiment, we presumed that a hotspot with small size of 1 × 1 or 2 × 2 and low impact of 1.5 is more a noise and not a real hotspot. However, because the hotspots are simulated, this is only an assumption. We may interpret the result in other way. If we assume that impact 1.5 is not noise and reveals a real hotspot, then we can infer that STScan outperforms EigenSpot for hotspots with low impacts and sizes. 3.1.3. Effect of hotspot size and impact on the performance In the previous section, we evaluated the methods per- formance for some limited important settings. In this section, with the same method of data generation and evaluation criterion, we study the stability of two algorithms STScan and EigenSpot for a wider range of hotspot sizes and impacts. We do not study the baseline method at this stage, because it is defeated in all previous cases. Hence, this time the hotspot size (H) varies from 1 to 10, and we test 16 different hotspot impacts (I) from 1.25 to 5 by step of 0.25. As formerly used, for each setting, we generate 100 random data sets. Therefore, we generate 16000 = 10 × 16 × 100 data sets. We then apply STScan and EigenSpot on all data sets and measure their average accuracy on 100 data sets for each setting in the operational significance levels (p-values) from 0.20 to 0.01. Figure 2 shows the result of this comparison. In order to see that whether this improvement is obtained by chance or is statistically significant, we per- form a paired student’s t-test between two sets of obtained performances for STScan and EigenSpot. The t-test con- firms that the obtained improvement is statistically sig- nificant with p � value = 3.3591 × 10� 89 ≈ 0. Figure 3 shows the performance of methods against different hotspot sizes and impact. The lowest performance for EigenSpot is obtained for impacts of 1.25 and 1.5, which is more related to the noise (as was already discussed). However, we can observe that both methods relatively are robust for a hotspot impact greater than a threshold. For instance, EigenSpot is robust for impacts over 1.75, and STScan is robust for impacts over 2.5. Regarding the hot- spot size, EigenSpot has a descending trend by increasing the hotspot size. This implies that by increasing the hotspot size, we should expect lower performance from EigenSpot. It makes sense, as by increasing the size of hotspot, the affected areas gradually start to seem normal and are left undetected, via a spectral method such as EigenSpot. Eigen- Spot, however, exhibits more regular behaviour comparing STScan in this matter. The variance of performance is almost zero for EigenSpot during different size of hotspots, whilst it can vary up to 0.20 for STScan. STScan, also opposed to EigenSpot, experiences both ascending and Figure 2: Mean accuracy for 16000 data sets averaged for 173 α from 0.20 to 0.01. Figure 3: Mean accuracy of STScan and EigenSpot corresponding to different settings for 173 α from 0.20 to 0.01 averaged for 100 data sets. © 2014 Wiley Publishing Ltd460 Expert Systems, June 2015, Vol. 32, No. 3 descending trend. For hotspot sizes 1 × 1 to 6 × 6 have an ascending trend and then tend to decrease for bigger sizes. For hotspot 9 × 9, it has relatively the same performance as 1 × 1. To understand whether the hotspot size and impact affect the performance of the methods, we perform an analysis of variance (ANOVA) test (Montgomery et al., 1984) on the obtained performance for different hotspot sizes and impacts. The null hypothesis H0 is that that the mean accuracy does not change for different sizes and impacts. The test result (Table 2) confirms our initial guess that both STScan and EigenSpot become independent of hotspot impact when impact goes upper than a specific threshold. However, very low p-values for hotspot size indicate that the performance of both EigenSpot and STScan is dependent on the hotspot size. However, as observed, both methods do not differ in their dependence on hotspot impact and size. 3.1.4. The effect of singular value decomposition implementation The central technique used in EigenSpot is the SVD. Two kinds of SVD can be used for this purpose: a full-rank SVD and a low-rank SVD. Here, four of SVD implementations are chosen, two from each category, and their effect is studied on the EigenSpot performance. Table 3 demonstrates the average accuracy for 1500 data sets for the range of p-values from 0.20 to 0.01. As it is seen, the ARPACK implementation (Sorensen, 1997) (the default SVD implementation we use in the experiments) outperforms other methods. However, because both ARPACK and IncPACK (Brand, 2006) have the same computational cost, we perform an ANOVA test (Montgomery et al., 1984) to see whether using IncPACK affects the performance or not. The ANOVA test shows that two sets of accuracy obtained from these two implementations are not statistically different (p-value = 0.26). Therefore, we can conclude that low-rank SVD implementation used does not affect the EigenSpot performance. Concerning full-rank SVD, we observe that LAPACK (Anderson, 1999) outperforms PROPACK (Larsen, 1998); however, full-rank SVD, because of its computational cost, is not of our interest. 3.2. Experiment with real data In this section, we study the performance of EigenSpot on a real data set. The data set that is publicly available (Kulldorff, 2012a) is provided by surveillance, epidemiology and end results programme of the National Cancer Institute and collected by the New Mexico Tumor Registry between the years of 1973 and 1991 for 32 sub-regions of the New Mexico State, USA. There are 1175 reported cases of malignant neoplasm of the brain and the nervous system. The goal of the initial study was a response to a serious concern in 1991 in the New Mexico resident community about the correlation of wartime nuclear activities in Los Alamos with the recent brain tumour deaths in the neighbourhood. The concern rapidly emerged at the local and national level and therefore became a centre of attention by the local health departments. The data set was gathered, via a comprehensive review of the reported brain cancer incidence rates for the year 1973 through 1991 in order to identify the statistically significant spatiotemporal affected areas. STScan is already applied to this data (Kulldorff et al., 1998). The conclusion made from the previous study shows that excess of brain cancer in Los Alamos falls within the realm of chance, which confirms the final conclusion of the New Mexico Health Department. In order to compare the EigenSpot with STScan, the EigenSpot is applied on the same data set. In addition to the initial study (Kulldorff et al., 1998), we use the adjusted incidences for temporal trends, age, race and sex. The results obtained via STScan and EigenSpot are shown in Table 4. STScan reports only Santa Fe and Los Alamos in the years 1986–1989 with a relative high p-value = 0.45, which indicates that there is no significant hotspot. Applying EigenSpot with α = 0.05, we could find a significant hotspot, including areas of Santa Fe, Bernalillo and Valencia as spatial components and the years 1981 and 1987 as the temporal components. However, for α = 0.01, EigenSpot does not find any hotspot. If we look the hotspot spatial positions in Figure 4, we can find a meaningful relationship between STScan and EigenSpot results. Both candidate areas are found close the Los Alamos, the region of nuclear Table 2: The effect of hotspot size and impact on the per- formance (one-way analysis of variance test) Factor STScan EigenSpot Hotspot size p = 1.6826 × 10� 10 p = 5.9713 × 10� 13 Hotspot impact p = 0.1834 (for impacts ≥ 2.5) p = 0.9337 (for impacts ≥ 1.75) Table 3: Average accuracy for 1500 data sets averaged for 173 α from 0.20 to 0.01 Method Cost Implementation Average accuracy One-rank SVD O(N2) ARPACK 0.8975 IncPACK 0.8387 Full SVD O(N3) LAPACK 0.8429 PROPACK 0.8177 SVD, singular value decomposition. Bold items implies the outperforming method in terms of accuracy. Table 4: Comparison of STScan versus EigenSpot in detection of hotspots Method Affected Regions Temporal period p-value STScan Santa Fe and Los Alamos 1986–1989 0.45 EigenSpot Santa Fe, Bernalillo and Valencia 1981, 1987 0.05 Incidence rates were adjusted for temporal trends, age, race and sex. © 2014 Wiley Publishing Ltd Expert Systems, June 2015, Vol. 32, No. 3 461 activities. The temporal component of 1987 also appeared in the EigenSpot, which is located in the period detected by STScan. On the basis of EigenSpot result, it seems that in addition to the area close to Los Alamos and Santa Fe, more areas were affected by the nuclear activities. These areas are Bernalillo and Valencia where the neighbours of Santa Fe and Los Alamos are. The very interesting point about EigenSpot is that EigenSpot opposed to STScan was not aware about the geographic relationship of the regions. STScan knows in advance that for instance whether Santa Fe and Bernalillo are neighbours or not, whilst EigenSpot does not have this prior knowledge. On the basis of the EigenSpot result, we can infer that the effect of nuclear activities in the neighbourhood has experienced two peaks during the years 1981 and 1987. It makes sense, because the initial concerns about the effect of the nuclear activities started in 1991, four years right after 1987 (second detected temporal component). Indeed, EigenSpot has truly approximated the hotspot spatially and temporally very close the nuclear activity area. Most interestingly, no other meaningless hotspots are detected by EigenSpot. However, lack of strong p-value for the recognized area reveals that the neighbourhood has been under a low effect of nuclear activities in Los Alamos but has not had an enough support to be considered alarming. If we do not consider α = 0.05 as significant, we can confirm initial conclusion about the random incidence of brain cancers in Los Alamos. 4. Conclusion and future works A new methodology for hotspot detection is proposed, which is based on two robust techniques, including matrix factorization and process control. We evaluate and compare the performance of the algorithm for detection of a single hotspot against the state-of-the-art and the baseline methods through a comprehensive simulation study. The obtained results indicate a statistically significant improvement over the state-of-the-art method STScan. This improvement comes from the inherent methodological differences of the two approaches. The STScan uses the deviation in pro- bability model as the criteria for identification of hotspots, whilst our approach tracks the changes the correlation patterns in spatial and temporal dimension to approximate the hotspot location. Besides, our approach is a shape-free method, and contrary to STScan, it is robust to the noise and outliers. Our approach is also much more efficient than the scan statistics-based approaches. The main benefit of our ap- proach is that it has linear complexity, in terms of both space and time. The comprehensive comparison of scan statistics-based methods in (Agarwal et al., 2006) reveals that any algorithms that even provide approximately optimal answers to the problem must use space linear in the input. EigenSpot provides an approximate optimal answer and is linear with space and time dimensions, and it therefore meets this requirement. We also study the effect of hotspot size and impact in the methods performance. On the basis of this result, both the STScan and EigenSpot are independent of the hotspot impact in some specific ranges. However, both methods are dependent on the hotspot size. Nevertheless, EigenSpot exhibits a more regular trend against changes in hotspot size and impact. We also study the effect of SVD implemen- tation on the EigenSpot performance. The study shows that there is no statistical difference between two low-rank SVD implementations ARPACK and IncPACK. Therefore, SVD implementation does not affect the performance of Eigen- Spot. Finally, we apply EigenSpot to a real data set and compare its performance to STScan. EigenSpot, as well as the STScan, recognizes the affected area close to the nuclear activity area, both in space and time, however, as well as STScan, cannot provide the strong statistical evidence to identify this area as hotspot. EigenSpot can be used as an important component in surveillance systems in particular biosurveillance systems. Some estimations (Kaufmann et al., 1997) show that the Figure 4: Detected hotspot via STScan (left) and EigenSpot (right). © 2014 Wiley Publishing Ltd462 Expert Systems, June 2015, Vol. 32, No. 3 timely detection of hotspots can save the live of thirty thousands of people per day during a bio-agent release. It also prevents the economic cost of $250m per hour during a disease outbreak. Therefore, any early knowledge of hotspots plays an important role in improving response effectiveness. It is estimated (Morrison et al., 2010) that diagnosing and controlling abnormal situations has an economic impact of at least $10bn annually in the United States. Although EigenSpot is an ideal solution in terms of both accuracy and computational cost for single hotspot de- tection, there is a doubt that this result is valid when multiple hotspots exist. In this work, we did not evaluate the performance of EigenSpot for multiple hotspot de- tection. However, theoretically, we expect that STScan performs better for that purpose. Because, combining the spatial and temporal components of different hotspots together raise many false positives, which reduce the method performance, even though this may not be considered a serious issue in disease surveillance, where in practice the most likely cluster is desired. There are two directions for future works. In the first direction, we intend to find a solution for adapting EigenSpot for multiple hotspot detection, and in the second direction, we are going to apply EigenSpot along with visualization tools for online and real-time moni- toring purposes. Acknowledgements This research was supported by the Projects NORTE-07- 0124-FEDER-000059/000056 that is financed by the North Portugal Regional Operational Program (ON.2 O Novo Norte), under the National Strategic Reference Framework (NSRF), through the European Regional Development Fund (ERDF), and by national funds, through the Portuguese funding agency, Fundação para a Ciência e a Tecnologia (FCT). Authors also acknowledge the support of the European Commission through the project MAESTRA (grant number ICT-750 2013-612944). References ABDI, H. and L. J. WILLIAMS (2010). Principal component analysis. Wiley Interdisciplinary Reviews: Computational Statistics (4), 433–459. AGARWAL, D., A. MCGREGOR, J. M. PHILLIPS, S. VENKATASUBRAMANIAN, and Z. ZHU (2006). Spatial scan statistics: approximations and performance study. In Proceedings of the 12th ACM SIGKDD international conference on knowledge discovery and data mining, 24–33. ACM. ANDERSON, E. (1999). LAPACK Users’ Guide, Volume 9. Philadelphia, PA, USA: Siam. ISBN:0-89871-447-8 ASSUNÇÃO, R., A. TAVARES, and M. KULLDORFF (2004). An early warning system for space-time cluster detection. Technical report, Laboratorio De Estatística Espacial. BELL, R. M., Y. KOREN, and C. VOLINSKY (2007). The Bellkor solution to the Netflix prize. KorBell team’s report to Netflix . BRAND, M. (2006). Fast low-rank modifications of the thin singular value decomposition. Linear algebra and its applica- tions (1), 20–30. BRYAN, K. and T. LEISE (2006). The $ 25,000,000,000 eigenvector: the linear algebra behind Google. Siam Review (3), 569–581. BUCKERIDGE, D. L., H. BURKOM, M. CAMPBELL, W. R. HOGAN, and A. W. MOORE (2005). Algorithms for rapid outbreak detection: a research synthesis. J. Biomed. Inform. (2), 99–113. DUCZMAL, L. and R. ASSUNÇÃO (2004). A simulated annealing strategy for the detection of arbitrarily shaped spatial clusters. Computational Statistics & Data Analysis (2), 269–286. HAND, D. J. (2009). Measuring classifier performance: a coherent alternative to the area under the roc curve. Machine learning (1), 103–123. HAND, D. J. and C. ANAGNOSTOPOULOS (2013). When is the area under the receiver operating characteristic curve an appropriate measure of classifier performance? Pattern Recognition Letters (5), 492–495. KAUFMANN, A. F., M. I. MELTZER, and G. P. SCHMID (1997). The economic impact of a bioterrorist attack: are prevention and postattack intervention programs justifiable? Emerg. Infect. Dis. (2), 83. KLEMA, V. and A. LAUB (1980). The singular value decomposition: its computation and some applications. Automatic Control, IEEE Transactions on (2), 164–176. KULLDORFF, M. (1997). A spatial scan statistic. Communications in Statistics-Theory and methods (6), 1481–1496. KULLDORFF, M. (1999). Spatial scan statistics: models, calculations, and applications. In Scan Statistics and Applications, 303–322. Boston , USA: Birkhauser. KULLDORFF, M. (2012a). Brain cancer incidence in New Mexico. http://www.satscan.org/datasets/nmbrain/index. html. [Accessed: December 2012.] KULLDORFF, M. (2012b). Satscan – software for the spatial, temporal, and space–time scan statistics. http://www.satscan. org. [Accessed: December 2012]. KULLDORFF, M., W. ATHAS, E. FEURER, B. MILLER, and C. KEY (1998). Evaluating cluster alarms: a space-time scan statistic and brain cancer in Los Alamos, New Mexico. Am. J. Public Health (9), 1377–1380. KULLDORFF, M. and N. NAGARWALLA (1995). Spatial disease clusters: detection and inference. Stat. Med. (8), 799–810. LARSEN, R. M. (1998). Lanczos bidiagonalization with partial reorthogonalization. DAIMI Report Series 27(537). LEVINE, N. (2006). Crime mapping and the crimestat program. Geographical Analysis (1), 41–56. LOBO, J. M., A. JIMÉNEZ-VALVERDE, and R. REAL (2008). AUC: a misleading measure of the performance of predictive distribution models. Glob. Ecol. Biogeogr. (2), 145–151. MONTGOMERY, D. C., D. C. MONTGOMERY, and D. C. MONTGOMERY (1984). Design and Analysis of Experiments, Volume 7. Wiley New York. MORRISON, D., W. FOSLIEN, W. MACARTHUR, P. JOFRIET, and P. ENG (2010). The early event detection toolkit. Honeywell Process Solutions14. NEILL, D. B. and G. F. COOPER (2010). A multivariate Bayesian scan statistic for early event detection and characterization. Machine Learning 79, 261–282. NEILL, D. B. and A. W. MOORE (2004). Rapid detection of significant spatial clusters. In Proceedings of the tenth ACM SIGKDD international conference on knowledge discovery and data mining, KDD’04, New York, NY, USA, 256–265. ACM. NEILL, D. B. and M. R. SABHNANI (2007). A robust expectation- based spatial scan statistic. Advances in Disease Surveillance 61. NEILL, D.B. (2006). Detection of spatial and spatio-temporal clusters. Ph.D. thesis, Department of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA. © 2014 Wiley Publishing Ltd Expert Systems, June 2015, Vol. 32, No. 3 463 http://. http://www.satscan.org. http://. http://www.satscan.org. SORENSEN, D. C. (1997). Implicitly Restarted Arnoldi/Lanczos Methods for Large Scale Eigenvalue Calculations. Dordrecht, Netherlands: Kluwer Academic Publishers. TANGO, T. and K. TAKAHASHI (2005). A flexibly shaped spatial scan statistic for detecting clusters. Int. J. Health Geogr. (1), 11. TURK, M. and A. PENTLAND (1991). Eigenfaces for recognition. J. Cogn. Neurosci. (1), 71–86. Wikibooks (2012). Control systems/eigenvalues and eigenvectors – wikibooks, the free textbook project, wikibooks. [Online] http://en. wikibooks.org/w/index.php?title=Special:Cite&page=Control_Systems %2FEigenvalues_and_Eigenvectors&id=2672614 [accessed 28-October- 2013]. WONG, W.-K., A. MOORE, G. COOPER, and M. WAGNER (2005). What’s strange about recent events (wsare): an algorithm for the early detection of disease outbreaks. The Journal of Machine Learning Research 1961–1998. ZWEIG, M. H. and G. CAMPBELL (1993). Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clin. Chem. (4), 561–577. The authors Hadi Fanaee-T received his BSc degree in industrial metallurgy from Ferdowsi University of Mashhad, 2008, and his MSc in IT from Shiraz University, 2010. He has worked over 10 years as a programmer and developer in various governmental and industrial projects in Iran. For almost 2 years, he was a teacher at the University of Applied Science and Technology in Iran. He also was a research fellow at IEETA — Aveiro for 1 year. Currently, he is working toward a PhD at the Laboratory of Artificial Intelligence and Decision Support, University of Porto. His main research interest are in data mining using tensor and matrix decomposition. He also has been a PC member and a reviewer for over 20 conferences (e.g., ECML- PKDD) and journals. His publication records mostly are in the area of event detection and tensor decomposition applications. João Gama is an associate professor at the University of Porto and a senior researcher at LIAAD/INESC TEC. He received his PhD degree in computer science from the University of Porto, Portugal. His main interests are in machine learning and data mining, mainly in the context of data streams. He published more than 200 papers in major international conferences and journals and served as chair at ECML05, DS09, ADMA09, IDA11, and ECMLPKDD 2015. He co-organized a series of workshops on learning from data streams in conjunction with ECML-PKDD, KDD, SAC, and ICML. He is a member of the editorial board of MLJ, DAMI, NGC, and PAI; and he is the author of a recent book in Knowledge Discovery from Data Streams. © 2014 Wiley Publishing Ltd464 Expert Systems, June 2015, Vol. 32, No. 3 
work_mdbxzh27xbgwveaawrkit36ry4	----	Reasoning Techniques for Diabetics Expert Systems Procedia Computer Science 65 ( 2015 ) 813 – 820 Available online at www.sciencedirect.com 1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of Universal Society for Applied Research doi: 10.1016/j.procs.2015.09.030 ScienceDirect Reasoning Techniques for Diabetics Expert Systems IBRAHIM M.AHMEDa, MARCO ALFONSEb, MOSTAFA AREFc ,ABDEL-BADEEH M.SALEMd Computer Science Department, Faculty of Computer and Information Sciences, Ain Shams University, Cairo, Egypt Abstract The field of reasoning methodologies is very important in the area of knowledge computing and engineering. Reasoning methodologies has been one of the standard and improving techniques with strong methods for health expert systems industry. Reasoning techniques has provided greatest support for predicting diagnosing and treatment of disease with correct case of results. Diabetes needs greatest support of reasoning techniques for diagnosis and treatment. This paper focus on the main technical characteristics of four reasoning methodologies which are commonly used in developing diabetic expert systems, namely; reasoning with production rules, fuzzy reasoning, case-based reasoning, and ontological-case based reasoning. In addition, this paper proposes the best reasoning technique for diabetic expert systems. The main result of this study covers a variety of four reasoning methodologies, which reveals that case based reasoning paradigm is the best reasoning technique methodology regarding to the issues of maintenance ,powerful and knowledge representations . © 2015 The Authors. Published by Elsevier B.V. Peer-review under responsibility of Universal Society for Applied Research. Keywords: Diabetes;Rule based reasoning;case based reasoning; fuzzy reasoning; ontological reasoning; medical expert systems; 1. Introduction Studies in the field of medical decision support systems have been established and due to the high success rate of these studies, interest in this field is increasing every day. These systems frequently use various artificial intelligence techniques. Reasoning techniques is one of major branches of artificial intelligence and, indeed, it is one of the most rapidly developing subfields of AI research. Reasoning techniques were from the very beginning designed and used to analyse medical data sets. Today reasoning techniques provides several indispensible tools for intelligent data analysis. Especially in the last few years and in particular there is a lot of work done in medical diagnosis in small specialized diagnostic problems. © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of Universal Society for Applied Research CORE Metadata, citation and similar papers at core.ac.uk Provided by Elsevier - Publisher Connector https://core.ac.uk/display/82222687?utm_source=pdf&utm_medium=banner&utm_campaign=pdf-decoration-v1 http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.09.030&domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.09.030&domain=pdf 814 Ibrahim M. Ahmed et al. / Procedia Computer Science 65 ( 2015 ) 813 – 820 There is a greater interest in the study of diseases that are common throughout the world. Diabetes is one of them [1]. Diabetes mellitus is a syndrome with disordered metabolism and inappropriate hyperglycemia due to either a deficiency of insulin secretion or to combination of insulin resistance and inadequate insulin secretion to compensate. Fortunately, diabetes can be managed very effectively through healthy lifestyle choices, primarily diet and exercise. Mostly, Type 2 diabetes is strongly connected with obesity, age, and physical inactivity [2]. Most medical resources reported that 90 to 95% of diabetic is diagnosed as type-2. Simply, in these cases the pancreas is not able to produce enough insulin to keep the blood sugar level within normal ranges. In addition, the majority of this type diabetics do not know they are suffering from it. Over 80-90% of Type 2 diabetes is overweight. Therefore, reducing daily carbohydrates and fats intake and the commitment to a healthy diet with a simple waling keeps your glucose within normal ranges and help dropping those extra pounds [3]. Reasoning techniques has been an excellent support for making prediction of a particular from the data which is provided. Reasoning techniques in recent years have been the evolving, reliable and supporting tool in medical domain. Automatic learning has fetched a greater amount of interest in medical domain due to less amount of time for detection and less interaction with patient, saving time for patients care. There are many methods used in the reasoning technique rule based, case based ontology case based and fuzzy based. Among these methods, the case based is widely used. On the other hand, the development of computer technology and tools has provided a valuable assistance for Medicare [4]. An expert system is a computer program that provides expert advice as if a real person had been consulted where this advice can be decisions, recommendations or solutions [5]. The intention of our research is to provide self- monitor for patient of type 2 diabetes to get proper amount of daily calories with list of proper diet satisfy the amount of the calories. In this paper we propose the comparison results using reasoning techniques in diabetes exert systems. The paper goes through the reasoning techniques identifications, related works, comparison, discussion, results and conclusion. 2. Reasoning techniques in expert systems The abilities of inference, reasoning, and learning are the main features of any expert system. The research area in this field covers a variety of reasoning methodologies, e.g.; automated reasoning, case-based reasoning, commonsense reasoning, multi-model reasoning, fuzzy reasoning, geometric reasoning, non-monotonic reasoning, model-based reasoning, probabilistic reasoning, causal reasoning, qualitative reasoning, spatial reasoning and temporal reasoning [6,7]. In this section we focus our discussion about the main characteristics of three of the reasoning methodologies which are commonly used in developing diabetic expert systems, namely; reasoning with production rules, fuzzy-rules, and case-based reasoning. 2.1 Reasoning with Production Rules Production rules are the most commonly technique used in developing the inference engine of expert system. Forward chaining can be used to produce new facts (hence the term “production” rules), and backward chaining can deduce whether statements are true or not. Rule-based systems were one of the first large-scale commercial successes of artificial intelligence research. An expert system or knowledge-based system is the common term used to describe a rule-based processing system. It consists of three major elements, a knowledge base (the set of if-then rules and known facts), a working memory or database of derived facts and data, and an inference engi ne, which contains the reasoning logic used to process the rules and data[8,9]. Rule-based systems solve problems by taking an input specification and then “chaining” together the appropriate set of rules from the rule base to arrive at a solution. Given the same exact problem situation, the system will go through exactly the same amount of work to come up with the solution. In other words rule-based systems don’t inherently learn. In addition, given a problem that is outside the system’s original scope, the system often can’t 815 Ibrahim M. Ahmed et al. / Procedia Computer Science 65 ( 2015 ) 813 – 820 render any assistance. Finally, rule-based systems are very time-consuming to build and maintain because rule extraction from experts is labor-intensive and rules are inherently dependent on other rules, making the addition of new knowledge to the system a complex debugging task. Forward chaining is a data-driven reasoning process where a set of rules is used to drive new facts from an initial set of data. It does not use the resolution algorithm used in predicate logic. The forward-chaining algorithm generates new data by the simple and straightforward application or firing of the rules. As an inferencing procedure, forward chaining is very fast. Forward chaining is also used in real-time monitoring and diagnostic systems where quick identification and response to problems are required. Backward chaining is often called goal-directed inferencing, because a particular consequence or goal clause is evaluated first, and then we go backward through the rules. Unlike forward chaining, which uses-rules to produce new information, backward chaining uses rules to answer questions about whether a goal clause is true or not. Backward chaining is more focused than forward chaining, because it only processes rules that are relevant to the question. It is similar to how resolution is used in predicate logic. However, it does not use contradiction. It simply traverses the rule base trying to prove that clauses are true in a systematic manner. Backward chaining is used for advisory systems, where users ask questions and get asked leading questions to find an answer. One advantage of backward chaining is that, because the inferencing is directed, information can be requested from the user when it is needed. Some expert systems also provide a trace capability which allows the user to ask the inference engine why it asking for some piece of information, or why it came to some conclusion. 2.2 Reasoning with Fuzzy Rules In the rich history of rule-based reasoning in AI, the inference engines almost without exception were based on Boolean or binary logic. However, in the same way that neural networks have enriched the AI landscape by providing an alternative to symbol processing techniques, fuzzy logic has provided an alternative to Boolean logic - based systems [7]. Unlike Boolean logic, which has only two states, true or false, fuzzy logic deals with truth values which range continuously from 0 to 1. Thus something could be half true 0.5 or very likely true 0.9 or probably not true 0.1. The use of fuzzy logic in reasoning systems impacts not only the inference engine but the knowledge representation itself [7]. For, instead of making arbitrary distinctions between variables and states, as is required with Boolean logic systems, fuzzy logic allows one to express knowledge in a rule format that is close to a natural language expression. The difference between this fuzzy rule and the Boolean-logic rules we used in our forward- and backward- chaining examples is that the clauses “temperature is hot” and “humidity is sticky” are not strictly true or false. Clauses in fuzzy rules are real-valued functions called membership functions that map the fuzzy set “hot” onto the domain of the fuzzy variable “temperature” and produce a truth-value that ranges from 0.0 to 1.0 (a continuous output value, much like neural networks). Reasoning with fuzzy rule systems is a forward-chaining procedure. The initial numeric data values are fuzzified, that is, turned into fuzzy values using the membership functions. Instead of a match and conflict resolution phase where we select a triggered rule to fire, in fuzzy systems, all rules are evaluated, because all fuzzy rules can be true to some degree (ranging from 0.0 to 1.0). The antecedent clause truth values are combined using fuzzy logic operators (a fuzzy conjunction or and operation takes the minimum value of the two fuzzy clauses). Next, the fuzzy sets specified in the consequent clauses of all rules are combined, using the rule truth values as scaling factors. The result is a single fuzzy set, which is then defuzzified to return a crisp output value. 816 Ibrahim M. Ahmed et al. / Procedia Computer Science 65 ( 2015 ) 813 – 820 2.3 Reasoning with Cases Case-Based Reasoning (CBR) means reasoning from experiences (old cases) in an effort to solve problems, critique solutions and explain anomalous situations. The CBR approach to problem-solving and learning has got a lot of attention within the Artificial Intelligence’s community over the last few years, because as an intelligent-systems’ method enables information managers to increase efficiency and reduce cost by substantially automating processes. People tend to be comfortable using the CBR methodology for decision making, in dynamically changing situations and other situations were much is unknown and when solutions are not clear. The case is a list of features that lead to a particular outcome. (e.g. The information o n a patient history and the associated diagnosis). Determining the appropriate case features is the main knowledge engineering task in developing medical case-based expert systems [10]. This task involves defining the terminology of the domain and gathering representative cases of problem solving by the expert. Representation of case can be in any of several forms (predicate, frames). The CBR systems’ expertise is embodied in a collection (library) of past cases rather, than being encoded in classical rules. Each case typically contains a description of the problem plus a solution and/or the outcomes. The knowledge and reasoning process used by an expert to solve the problem is not recorded, but is implicit in the solution. A case-library can be a powerful corporate resource allowing everyone in an organization to tap in the corporate library, when handling a new problem. CBR allows the case-library to be developed incrementally, while its maintenance is relatively easy and can be carried out by domain experts. The methodology of CBR is becoming popular in developing expert systems because it automates applications that are based on precedent or that contain incomplete causal models [6].In a rule-based systems an incomplete mode or an environment which does not take into account all variables could result in either an answer built on incomplete data or simply no answer at all. Case-based methodology attempt to get around this shortcoming by inputting and analyzing problem data. Case-based expert systems solve new problems by adapting solutions that were used for previous and similar problems. Adaptation is one of the main difficulties in developing Case-Based expert systems. 2.4 Ontological Case Base reasoning Methodology for Diabetes Management Ontology is a formal and explicit specification of a shared conceptualization. Ontology defines a common vocabulary for researchers who need to share information in a domain. It includes machine-interpre definitions of basic concepts (classes) in the domain, properties of each concept describing various features and attributes of the concept (slots, relationships) and restrictions on slots (facets or role restrictions) . Ontology together with a set of individual instances of classes constitutes a knowledge base [20]. Domain knowledge ontology supports the implementation of intelligent Case Based Reasoning (CBR) systems. Standardized terminologies support efficient indexing and processing of patient data. It is an essential element for the implementation of knowledge-based clinical decision support by exploiting pre-defined semantic relationships, both hierarchical and non-hierarchical in nature. Systemized Nomenclature of Medicine-Clinical Terms (SNOMED CT) is the most comprehensive and complete terminology. For semantic retrieval purpose in CBR system and for the semantic interoperability of data collected from distributed EHR systems, a common terminology is needed to encode case-base knowledge. The unique identifiers and the conceptual structure representing each concept in a terminology system allow an unambiguous interpretation of the concept meaning across systems [21]. 3. Literature Review and Related Work M. Beulah et. al (2007) [11] introduced the ability to access diabetic expert system from any part of the world. They collect, organize, and distribute relevant knowledge and service information to the individuals. The project was designed and programmed via the dot net framework using rule based. The system allows the availability to 817 Ibrahim M. Ahmed et al. / Procedia Computer Science 65 ( 2015 ) 813 – 820 detect and give early diagnosis of three types of diabetes namely type 1, 2, gestational diabetes for both adult and children. S. Kumar and B. Bhimrao (2012) [12] developed a natural therapy system for healing diabetic, they aim integrate all the natural treatment information of diabetes in one place using ESTA (Expert System Shell for Text Animation) as knowledge based system. Their system begins with Consultation asking the users to select the disease (Diabetes) for which they want different type of natural treatment solution then describes the diabetes diseases and their symptoms. After that describes the Natural Care (Herbal /Proper Nutrition) treatment solution of diabetes disease. C. Marling et. Al (2014) [13] Presented systems are for CARE-PARTNER, which supports the long-term follow-up care of stem-cell transplantation patients; diabetes Support System, which aids in managing patients with type 1 diabetes on insulin pump therapy; renal disease; diagnosis and treatment of stress-related disorders using case-based reasoning. N. Nnamoko et.al (2013) [14] proposed a fuzzy expert system framework that combines case-based and rule- based reasoning effectively to produce a usable tool for Type 2 Diabetes Mellitus (T2DM) management, to produce crisp outputs to patients in the form of low-risk advice. The extended framework features a combined reasoning approach for simplified output in the form of decision support for clinicians. With seven operational input variables and two additional pre-set variables for testing, the results of the proposed work will be compared with other methods using similarity to expert’s decision as metrics. Using MATLAB as a tool. Tawfik. S. Zeki et.al (2012) [15] designed an expert system for diagnosis all types of diabetes. After data acquisition and designing a rule-based expert system, there system has been coded with VP_Expert Shell and tested in Shahid Hasheminezhad Teaching Hospital affiliated to Tehran University of Medical Sciences and final expert system has been presented. Margret Anouncia S. et.al (2013) [16] proposed a diagnosis system for diabetes. The system is implemented to diagnose the type of diabetes with the input symptoms given by the user. The system proves to be advantageous in aspects, such as accuracy and time consumption due to the rough set based knowledge representation. The system is adaptable for any number of symptoms and is evaluated with respect to the rule based. The inference engine interacts with the knowledge base which is constructed using rough sets for the process of diagnosis. The system is implemented using Java and JSP. Abdelhak.M et. al (2013) [17] proposed an approach based on using a multi-criteria decision guided by a case- based reasoning (CBR) approach. The study is intended to experiment with a multiple criteria decision approach to medical care in the diagnosis and the proposed therapy for diabetic patients. The system is developed in JAVA with an interconnecting module to the JCOLIBRI system. M. Kalpana and A.V Senthil Kumar (2011) [18] proposed a fuzzy expert system framework which constructs large scale knowledge based system effectively for diabetes. The knowledge is constructed by using the fuzzification to convert crisp values into fuzzy values. By applying the fuzzy verdict mechanism, diagnosis of diabetes becomes simple for medical practitioners. The proposed fuzzy expert system for diabetes application was implemented with the MATLAB using rules for knowledge representation. Cindy. M et.al (2009) [19] presents a case-based decision support system prototype to assist patients with Type 1 diabetes on insulin pump therapy, detect common problems in blood glucose control, and retrieval metrics were developed to find the most relevant past cases for solving current problems to control blood glucose levels.. The system is developed in JAVA. Shaker H. El-Sappagh et. al (2014) [20] Proposed an ontology engineering methodology to generate case bases in the medical domain. they mainly focuses on the research of case representation in the form of ontology to support 818 Ibrahim M. Ahmed et al. / Procedia Computer Science 65 ( 2015 ) 813 – 820 the case semantic retrieval and enhance all knowledge intensive CBR processes A case study on diabetes diagnosis case base had been provided to evaluate the proposed methodology, the ontology is created and implemented using OWL 2 language and Protégé tool. Jian-xun .C et.al (2013) [21] proposed a system which combines CBR and ontology to generate personalized care plan. they choose diabetes care as an example to use this system. When there is no matched case, embedded in there system diabetes care ontology can be consulted, they choose OWL DL(OWL with Description Logics) as the language of designing there ontology and builted by Protégé. D. Forbes and J. Singh (2012) [22] introduces a novel approach using new technology to improve understanding between the patient and healthcare practitioner they developed a framework that links medical information with different language and cultural information to provide ease of understanding and communication between the patient from a minority group with a healthcare practitioner from a different cultural group. The key component of this framework is the Type-2 Diabetes Management Patient-Practitioner Assistive Communication Ontology, The tool used in the implementation process is protégé 4.2. 4. Results and Discussion Tables 1, 2 , 3 and 4 show the results of our analysis of the expert systems in diabetes domain, based on the type of the reasoning methodology, during the last five years. Our frame work includes the flowing features: • The reasoning based of the system, it is either rule or case or fuzzy or ontology. • The system’s purpose. • The developing tool, it is either langue or shells. • The knowledge representation technique of the system’s knowledge-base, it is, frame, ontology, rule or case or graphs. Table 1: Rule based expert systems for diabetes System Purpose Knowledge technique Developing tool Authors Fully Automated Real- Time Web- Centric Expert System detect and give early diagnosis of three types of diabetes Rule Dot net P. M. Beulah et.al (2007) [11] knowledge Base Expert System for Natural treatment of Diabetes disease describes the Natural Care (Herbal /Proper Nutrition) treatment solution of diabetes disease. Rule ESTA S. Kumar and B. Bhimrao (2012) [12] An Expert System for Diabetes Diagnosis diagnosis all types of diabetes Rule VP_Expert Shell Tawfik. S. Zeki et.al (2012) [15] Table 2: Case based reasoning expert systems for diabetes System Purpose Knowledge technique Developing tool Authors case-based reasoning in medical domains managing patients with type 1 diabetes Case NA C. Marling et. Al (2014) [13] Toward case based reasoning for diabetes management control blood glucose levels Case JAVA Cindy. M et.al (2009) [19] Decision support system application on healthcare diagnosis for diabetic patients Case & graphs JAVA + JCOLIBRI Abdelhak .M et. al (2013) [17] 819 Ibrahim M. Ahmed et al. / Procedia Computer Science 65 ( 2015 ) 813 – 820 Table 3: Fuzzy-based expert systems for diabetes System Purpose Knowledge technique Developing tool Authors Design of a Diabetic Diagnosis System Using Rough Sets diagnose the type of diabetes Rule Java and JSP Margret Anouncia S. et.al (2013) [16] Fuzzy Expert System for Type 2 Diabetes Mellitus management of Type 2 Diabetes Rule & case MATLAB N. Nnamoko et.al (2013) [14] Fuzzy Expert System for Diabetes constructs large scale knowledge based system effectively for diabetes Rule MATLAB M.Kalpana and A.V Senthil Kumar (2011) [ 18] Table 4: Ontology case based reasoning expert systems for diabetes System Purpose Knowledge technique Developing tool Authors Ontology for Type2 Diabetes Management Understanding between Type-2 Diabetes Patient and healthcare practitioner Ontology protégé 4.2 D.Forbes and J. Singh (2012)[22] Diabetes Care Decision Support System generate personalized care plan for diabetes Ontology Protégé Jian-xun .C et.al(2013) [21] An Ontological Case Base Engineering Methodology for Diabetes Management diagnosis for diabetic patients Ontology Protégé Shaker H. El-Sappagh et. Al(2014) [20] From our analysis presented in tables 1,2,3, and 4 as well as the reported research mentioned in section 3 , it can seen the following observations and results: 1- Most of the case based reasoning systems were developed in Java, most of fuzzy based reasoning systems were developed in MATLAB, and rule based reasoning were developed with dot net/ESTA/VP expert shell. Reasoning techniques used for deferent purpose in diabetes domain. 2- All of the ontology case based reasoning systems developed in Protégé tool and the ontology presented by OWL, because OWL (Web Ontology Language) is easier to express meaning than XML, RDF, RDF-S. 3- Rule-based systems are very time-consuming to build and maintain because rule extraction from experts is labor-intensive and rules are inherently dependent on other rules, making the addition of new knowledge to the system a complex debugging task. 4- Case based reasoning is a powerful methodology regarding to the issues of maintenance and knowledge representations over the rule based system. 5. Conclusion and future work The research in diabetic systems is important for both medical industry and diabetes patients. Reasoning techniques for diagnosing and treatment diabetes is urgently needed for helping both specialist doctors and patients. The abilities of inference, reasoning, and learning are the main features of any expert system. The research area in this field covers a variety of reasoning methodologies, e.g.; case-based reasoning, ontology case-based reasoning, fuzzy reasoning and rule reasoning. Case based reasoning is the more efficient, powerful and less cost.Our research was motivated by the need of such techniques, therefore the reasoning techniques for diabetics expert system has been presented in this paper as platform towards designing and implementation expert systems for diabetes. Now we are developing expert system for diabetes diet that intended to be used in Sudan and Arab countries we finished the knowledge acquisition stage and going to design and implement an intelligent mobile base expert system in Arabic language interface using the case based reasoning as reasoning technique. In future we need develop a new approach and then compare it with the current methods in this paper. 820 Ibrahim M. Ahmed et al. / Procedia Computer Science 65 ( 2015 ) 813 – 820 References 1. Rebecca Shafield,"Case-Based Reasoning", Logic Programming Associates Ltd, England, 2004. 2. Audrey Mbogho et al, "Diabetes Advisor a Medical Expert System for Diabetes Management ", University of Cape Town, pp 84-87. 2005. 3. Diabetic Diet Plan and Food Guide http://www.diabeticdietfordiabetes.com/foods.htm, 22 may, 2014. 4. Huiqing H. Yang and Sharnei Miller, "A PHP-CLIPS Based Intelligent System for Diabetic Self-Diagnosis", Department of Math & Computer Science, Virginia State University Petersburg, pp., 2006. 5. Byoung-Ho Song, Kyoung-Woo Park and Tae Yeun Kim. "U-health Expert System with Statistical Neural Network", Advances on Information Sciences and Service Sciences. vol. 3, no.1, pp 54-61, 2011. 6. Kolonder, J., Case-Based Reasoning, Morgan Kaufmann, 1993. 7. Zadeh, L. A., “Fuzzy Sets”, Information and Control, 8, 338-353, 1965. 8. Salem, A-B.M., Roushdy, M. and HodHod, R. A., “A Rule-Base Expert System for Diagnosis of Heart Diseases”, Proceedings of 8th International Conference on Soft Computing MENDEL. Brno, Czeeh Republic, June 5-7, 2002, pp. 258-263. 9. Greer, J. Proceedings of AI-ED 95, World Conference on Artificial Intelligence in Education, Assocaition for Advancement of Computing in Education (AACE), 1995. 10. Abdel-Badeeh M. Salem and Michael Gr.Voskoglou, “Applications of CBR Methodology to Medicine “, Egyptian Computer Science Journal, ISSN 1110-2586, Special Issue for EMMIT 9th Int. Conf. for Scientific and Social Development in Mediterranean Countries ,Nador, Morocco, October 21-23,2013, Vol.37,No.7,pp.68-77,2013. 11. P. M. Beulah Devamalar, V. Thulasi Bai, and Srivatsa S. K. "An Architecture for a Fully Automated Real-Time Web- Centric Expert System", World Academy of Science, Engineering and Technology, pp11-23, 2007. 12. Sanjeev Kumar and Babasaheb Bhimrao, "Development of knowledge Base Expert System for Natural treatment of Diabetes disease", (IJACSA) International Journal of Advanced Computer Science and Applications, vol. 3, no. 3, pp 44-47, 2012. 13. C. Marling, S.Montani , I. Bichindaritz and Peter Funk' Synergistic case-based reasoning in medical domains", Expert Systems with Applications, Volume 41, Issue 2, 1, Pages 249–259, February 2014. 14. N.Nnamoko, F.Arshad, D.England and J. Vora,"Fuzzy Expert System for Type 2 Diabetes Mellitus (T2DM) Management Using Dual Inference Mechanism", AAAI Spring Symposium, p 67-70, 2013. 15. Tawfik .S, Mohammad .V, Yousef. A and S.Talayeh," An Expert System for Diabetes Diagnosis", American Academic & Scholarly Research Journal Vol. 4, No. 5, Sept 2012. 16. Margret Anouncia S., Clara Madonna L., Jeevitha P.and Nandhini R.," Design of a Diabetic Diagnosis System Using Rough Sets", CYBERNETICS AND INFORMATION TECHNOLOGIES • Volume 13, No 3, pp 124-139,Sofia ,2013. 17. Abdelhak .M, Baghdad .A, and Sofia. B," A HYBRID DECISION SUPPORT SYSTEM :APPLICATION ON HEALTHCARE", CoRR, 2013. 18. M.Kalpana and A.V Senthil Kumar," Fuzzy Expert System for Diabetes using Fuzzy Verdict Mechanism ", Int. J. Advanced Networki ng and Applications Volume: 03, Issue: 02, Pages: 1128-1134 (2011). 19. Cindy.M, Jay. Sand Frank. S," TOWARD CASE BASED REASONING FOR DIABETES MANAGEMENT", Computational Intelligence Journal, Volume 25, Number 3,9p 165-179, 2009. 20. Shaker H. El-Sappagh & Samir El-Masri &Mohammed Elmogy & A. M. Riad & Basema Saddik," An Ontological Case Base Engineering Methodology for Diabetes Management", Springer Science+Business Media New York 2014. 21. Jian-xun .C. Shih-Li Su and Che-Ha, "Diabetes Care Decision Support System", Chang://tweb.cjcu.edu.tw/conference_abstract/2013_03_04. 22. D.Forbes1 and J. Singh2," Development of Patient-Practitioner Assistive Communications (PPAC) Ontology for Type2 Diabetes Management", CIHealth, Sydney, pp. 43-54, 2012. 
work_mely4t5o2vcybjs6l7fhhfyyb4	----	               City, University of London Institutional Repository Citation: Arunkumar, S., Sensoy, M., Srivatsa, M. and Rajarajan, M. (2015). Reasoning with Streamed Uncertain Information from Unreliable Sources. Expert Systems with Applications, 42(22), pp. 8381-8392. doi: 10.1016/j.eswa.2015.04.031 This is the other version of the paper. This version of the publication may differ from the final published version. Permanent repository link: http://openaccess.city.ac.uk/8678/ Link to published version: http://dx.doi.org/10.1016/j.eswa.2015.04.031 Copyright and reuse: City Research Online aims to make research outputs of City, University of London available to a wider audience. Copyright and Moral Rights remain with the author(s) and/or copyright holders. URLs from City Research Online may be freely distributed and linked to. City Research Online: http://openaccess.city.ac.uk/ publications@city.ac.uk City Research Online http://openaccess.city.ac.uk/ mailto:publications@city.ac.uk Creative Commons Legal Code Attribution­NonCommercial­NoDerivatives 4.0 International Official translations of this license are available in other languages. Creative Commons Corporation (“Creative Commons”) is not a law firm and does not provide legal services or legal advice. Distribution of Creative Commons public licenses does not create a lawyer­ client or other relationship. Creative Commons makes its licenses and related information available on an “as­is” basis. Creative Commons gives no warranties regarding its licenses, any material licensed under their terms and conditions, or any related information. Creative Commons disclaims all liability for damages resulting from their use to the fullest extent possible. Using Creative Commons Public Licenses Creative Commons public licenses provide a standard set of terms and conditions that creators and other rights holders may use to share original works of authorship and other material subject to copyright and certain other rights specified in the public license below. The following considerations are for informational purposes only, are not exhaustive, and do not form part of our licenses. Considerations for licensors: Our public licenses are intended for use by those authorized to give the public permission to use material in ways otherwise restricted by copyright and certain other rights. Our licenses are irrevocable. Licensors should read and understand the terms and conditions of the license they choose before applying it. Licensors should also secure all rights necessary before applying our licenses so that the public can reuse the material as expected. Licensors should clearly mark any material not subject to the license. This includes other CC­licensed material, or material used under an exception or limitation to copyright. More considerations for licensors. Considerations for the public: By using one of our public licenses, a licensor grants the public permission to use the licensed material under specified terms and conditions. If the licensor’s permission is not necessary for any reason–for example, because of any applicable exception or limitation to copyright–then that use is not regulated by the license. Our licenses grant only permissions under copyright and certain other rights that a licensor has authority to grant. Use of the licensed material may still be restricted for other reasons, including because others have copyright or other rights in the material. A licensor may make special requests, such as asking that all changes be marked or described. Although not required by our licenses, you are encouraged to respect those requests where reasonable. More considerations for the public. Creative Commons Attribution­NonCommercial­NoDerivatives 4.0 International Public License By exercising the Licensed Rights (defined below), You accept and agree to be bound by the terms and conditions of this Creative Commons Attribution­NonCommercial­NoDerivatives 4.0 International Public License ("Public License"). To the extent this Public License may be interpreted as a contract, You are granted the Licensed Rights in consideration of Your acceptance of these terms and conditions, and the Licensor grants You such rights in consideration of benefits the Licensor receives from making the Licensed Material available under these terms and conditions. Section 1 – Definitions. http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensees http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensors a.  Adapted Material means material subject to Copyright and Similar Rights that is derived from or based upon the Licensed Material and in which the Licensed Material is translated, altered, arranged, transformed, or otherwise modified in a manner requiring permission under the Copyright and Similar Rights held by the Licensor. For purposes of this Public License, where the Licensed Material is a musical work, performance, or sound recording, Adapted Material is always produced where the Licensed Material is synched in timed relation with a moving image. b.  Copyright and Similar Rights means copyright and/or similar rights closely related to copyright including, without limitation, performance, broadcast, sound recording, and Sui Generis Database Rights, without regard to how the rights are labeled or categorized. For purposes of this Public License, the rights specified in Section 2(b)(1)­(2) are not Copyright and Similar Rights. c.  Effective Technological Measures means those measures that, in the absence of proper authority, may not be circumvented under laws fulfilling obligations under Article 11 of the WIPO Copyright Treaty adopted on December 20, 1996, and/or similar international agreements. d.  Exceptions and Limitations means fair use, fair dealing, and/or any other exception or limitation to Copyright and Similar Rights that applies to Your use of the Licensed Material. e.  Licensed Material means the artistic or literary work, database, or other material to which the Licensor applied this Public License. f.  Licensed Rights means the rights granted to You subject to the terms and conditions of this Public License, which are limited to all Copyright and Similar Rights that apply to Your use of the Licensed Material and that the Licensor has authority to license. g.  Licensor means the individual(s) or entity(ies) granting rights under this Public License. h.  NonCommercial means not primarily intended for or directed towards commercial advantage or monetary compensation. For purposes of this Public License, the exchange of the Licensed Material for other material subject to Copyright and Similar Rights by digital file­sharing or similar means is NonCommercial provided there is no payment of monetary compensation in connection with the exchange. i.  Share means to provide material to the public by any means or process that requires permission under the Licensed Rights, such as reproduction, public display, public performance, distribution, dissemination, communication, or importation, and to make material available to the public including in ways that members of the public may access the material from a place and at a time individually chosen by them. j.  Sui Generis Database Rights means rights other than copyright resulting from Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases, as amended and/or succeeded, as well as other essentially equivalent rights anywhere in the world. k.  You means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning. Section 2 – Scope. a.  License grant. 1.  Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty­free, non­sublicensable, non­exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to: A.  reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only; and B.  produce and reproduce, but not Share, Adapted Material for NonCommercial purposes only. 2.  Exceptions and Limitations. For the avoidance of doubt, where Exceptions and Limitations apply to Your use, this Public License does not apply, and You do not need to comply with its terms and conditions. 3.  Term. The term of this Public License is specified in Section 6(a). 4.  Media and formats; technical modifications allowed. The Licensor authorizes You to exercise the Licensed Rights in all media and formats whether now known or hereafter created, and to make technical modifications necessary to do so. The Licensor waives and/or agrees not to assert any right or authority to forbid You from making technical modifications necessary to exercise the Licensed Rights, including technical modifications necessary to circumvent Effective Technological Measures. For purposes of this Public License, simply making modifications authorized by this Section 2(a)(4) never produces Adapted Material. 5.  Downstream recipients. A.  Offer from the Licensor – Licensed Material. Every recipient of the Licensed Material automatically receives an offer from the Licensor to exercise the Licensed Rights under the terms and conditions of this Public License. B.  No downstream restrictions. You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, the Licensed Material if doing so restricts exercise of the Licensed Rights by any recipient of the Licensed Material. 6.  No endorsement. Nothing in this Public License constitutes or may be construed as permission to assert or imply that You are, or that Your use of the Licensed Material is, connected with, or sponsored, endorsed, or granted official status by, the Licensor or others designated to receive attribution as provided in Section 3(a)(1)(A)(i). b.  Other rights. 1.  Moral rights, such as the right of integrity, are not licensed under this Public License, nor are publicity, privacy, and/or other similar personality rights; however, to the extent possible, the Licensor waives and/or agrees not to assert any such rights held by the Licensor to the limited extent necessary to allow You to exercise the Licensed Rights, but not otherwise. 2.  Patent and trademark rights are not licensed under this Public License. 3.  To the extent possible, the Licensor waives any right to collect royalties from You for the exercise of the Licensed Rights, whether directly or through a collecting society under any voluntary or waivable statutory or compulsory licensing scheme. In all other cases the Licensor expressly reserves any right to collect such royalties, including when the Licensed Material is used other than for NonCommercial purposes. Section 3 – License Conditions. Your exercise of the Licensed Rights is expressly made subject to the following conditions. a.  Attribution. 1.  If You Share the Licensed Material, You must: A.  retain the following if it is supplied by the Licensor with the Licensed Material: i.  identification of the creator(s) of the Licensed Material and any others designated to receive attribution, in any reasonable manner requested by the Licensor (including by pseudonym if designated); ii.  a copyright notice; iii.  a notice that refers to this Public License; iv.  a notice that refers to the disclaimer of warranties; v.  a URI or hyperlink to the Licensed Material to the extent reasonably practicable; B.  indicate if You modified the Licensed Material and retain an indication of any previous modifications; and C.  indicate the Licensed Material is licensed under this Public License, and include the text of, or the URI or hyperlink to, this Public License. For the avoidance of doubt, You do not have permission under this Public License to Share Adapted Material. 2.  You may satisfy the conditions in Section 3(a)(1) in any reasonable manner based on the medium, means, and context in which You Share the Licensed Material. For example, it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information. 3.  If requested by the Licensor, You must remove any of the information required by Section 3(a)(1)(A) to the extent reasonably practicable. Section 4 – Sui Generis Database Rights. Where the Licensed Rights include Sui Generis Database Rights that apply to Your use of the Licensed Material: a.  for the avoidance of doubt, Section 2(a)(1) grants You the right to extract, reuse, reproduce, and Share all or a substantial portion of the contents of the database for NonCommercial purposes only and provided You do not Share Adapted Material; b.  if You include all or a substantial portion of the database contents in a database in which You have Sui Generis Database Rights, then the database in which You have Sui Generis Database Rights (but not its individual contents) is Adapted Material; and c.  You must comply with the conditions in Section 3(a) if You Share all or a substantial portion of the contents of the database. For the avoidance of doubt, this Section 4 supplements and does not replace Your obligations under this Public License where the Licensed Rights include other Copyright and Similar Rights. Section 5 – Disclaimer of Warranties and Limitation of Liability. a.  Unless otherwise separately undertaken by the Licensor, to the extent possible, the Licensor offers the Licensed Material as­is and as­available, and makes no representations or warranties of any kind concerning the Licensed Material, whether express, implied, statutory, or other. This includes, without limitation, warranties of title, merchantability, fitness for a particular purpose, non­infringement, absence of latent or other defects, accuracy, or the presence or absence of errors, whether or not known or discoverable. Where disclaimers of warranties are not allowed in full or in part, this disclaimer may not apply to You. b.  To the extent possible, in no event will the Licensor be liable to You on any legal theory (including, without limitation, negligence) or otherwise for any direct, special, indirect, incidental, consequential, punitive, exemplary, or other losses, costs, expenses, or damages arising out of this Public License or use of the Licensed Material, even if the Licensor has been advised of the possibility of such losses, costs, expenses, or damages. Where a limitation of liability is not allowed in full or in part, this limitation may not apply to You. c.  The disclaimer of warranties and limitation of liability provided above shall be interpreted in a manner that, to the extent possible, most closely approximates an absolute disclaimer and waiver of all liability. Section 6 – Term and Termination. a.  This Public License applies for the term of the Copyright and Similar Rights licensed here. However, if You fail to comply with this Public License, then Your rights under this Public License terminate automatically. b.  Where Your right to use the Licensed Material has terminated under Section 6(a), it reinstates: 1.  automatically as of the date the violation is cured, provided it is cured within 30 days of Your discovery of the violation; or 2.  upon express reinstatement by the Licensor. For the avoidance of doubt, this Section 6(b) does not affect any right the Licensor may have to seek remedies for Your violations of this Public License. c.  For the avoidance of doubt, the Licensor may also offer the Licensed Material under separate terms or conditions or stop distributing the Licensed Material at any time; however, doing so will not terminate this Public License. d.  Sections 1, 5, 6, 7, and 8 survive termination of this Public License. Section 7 – Other Terms and Conditions. a.  The Licensor shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed. b.  Any arrangements, understandings, or agreements regarding the Licensed Material not stated herein are separate from and independent of the terms and conditions of this Public License. Section 8 – Interpretation. a.  For the avoidance of doubt, this Public License does not, and shall not be interpreted to, reduce, limit, restrict, or impose conditions on any use of the Licensed Material that could lawfully be made without permission under this Public License. b.  To the extent possible, if any provision of this Public License is deemed unenforceable, it shall be automatically reformed to the minimum extent necessary to make it enforceable. If the provision cannot be reformed, it shall be severed from this Public License without affecting the enforceability of the remaining terms and conditions. c.  No term or condition of this Public License will be waived and no failure to comply consented to unless expressly agreed to by the Licensor. d.  Nothing in this Public License constitutes or may be interpreted as a limitation upon, or waiver of, any privileges and immunities that apply to the Licensor or You, including from the legal processes of any jurisdiction or authority. Creative Commons is not a party to its public licenses. Notwithstanding, Creative Commons may elect to apply one of its public licenses to material it publishes and in those instances will be considered the “Licensor.” The text of the Creative Commons public licenses is dedicated to the public domain under the CC0 Public Domain Dedication. Except for the limited purpose of indicating that material is shared under a Creative Commons public license or as otherwise permitted by the Creative Commons policies published at creativecommons.org/policies, Creative Commons does not authorize the use of the trademark “Creative Commons” or any other trademark or logo of Creative Commons without its prior written consent including, without limitation, in connection with any unauthorized modifications to any of its public licenses or any other arrangements, understandings, or agreements concerning use of licensed material. For the avoidance of doubt, this paragraph does not form part of the public licenses. Creative Commons may be contacted at creativecommons.org. Additional languages available: Nederlands, Norsk, suomeksi. Please read the FAQ for more information about official translations. http://creativecommons.org/licenses/by-nc-nd/4.0/creativecommons.org/zero/1.0/legalcode http://creativecommons.org/policies http://wiki.creativecommons.org/FAQ#officialtranslations http://creativecommons.org/licenses/by-nc-nd/4.0/legalcode.fi http://creativecommons.org/licenses/by-nc-nd/4.0/legalcode.nl http://creativecommons.org/ http://creativecommons.org/licenses/by-nc-nd/4.0/legalcode.no 
work_meumwudbhvaj5p2sxziaief7iq	----	Estimation Medicine for Diseases System to Support Medical Diagnosis by Expert System (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 Estimation Medicine for Diseases System to Support Medical Diagnosis by Expert System Noor T. Mahmood1 Computer Science Department Al- Mustansiriyah University Baghdad, Iraq Abstract—Researches confirmed that 70 thousand cases of death, which happen yearly in the world, were because of the misprescribing of the drug itself or its dose (overdose or lower dose). Choosing the wrong alternative drug inspired the professionals in the healthcare field to the importance of assigning the best technologies to decrease the percentages of the therapeutic methods in giving the drug to prevent mistakes in prescribing the suitable drug. A system based on Rete Algorithm is proposed where the best-chosen medicine is offered through the suggested system. Selection of Estimation Medicine for Diseases (EMD) System is introduced where the diagnosis is made basically according to the symptoms and the medical history of the patient. This research aims to acquire a good model using this algorithm to obtain more accurate choices of medicine. The system (EMD) is tested by the doctors in Iraqi hospitals and it has been found that there is no other systems that can be compared to EMD system. The accuracy of estimating the appropriate medicine for heart diseases is approximately (87.26%). Keywords—Diagnosis; Disease; Medicine; Rete Algorithm; Expert System; Intelligent System I. INTRODUCTION Many of the patients have some troubles when they visit the doctor for identifying the suitable treatment for them. After the diagnosis of the disease have been made by the doctor, another problem is appeared, the drug might be written in a nonspecific way. Since the choice of the drug depends on the disease history of the patient and whether he had any chronic diseases, consequently careful assessment should be done as the drug is given. The Estimation Medicine for Diseases (EMD) is an intelligent system (Expert System) to identify the suitable treatment to the patient by knowing the disease or the case and the disease history of the patient and the symptoms that the patient is suffering. This process may prevent the mistakes that effects the wrong prescription of the drug which might not cause a significant health damage to many thousands of the patients only but also might lead to death to some of them. What’s done in the proposed approach is a process to enter the information into the system which identifies the correct diagnosis and obtains the accurate diagnosis of patient's diseases as quickly as possible with less cost. In our system selection, the best medicine is proposed for the patient and the doctor. It helps in determining medicine by disease or conditions(symptoms) according to some concepts, rules, and algorithms which are expressed in the next sections As a comparative view between previous researches and that one; previous researches usually choose two or three side effects and then contrast them to reach drugs with less interference and are less risky for the patient. On the other hand, this research is tested by an expertise doctor in diagnosing the disease depending on the rules of choosing the right drug for the intended patient. Also to realize the side effects of each drug or medication in addition to the previous symptoms taken from the patient prohibits causing pharmacological interference. The system picks the selected medicine that takes less error ratio and higher priority of accuracy for the preferring drugs that have fewer side effects and less conflicts with other medicines. Collecting information about both diseases and medicines along with the side effects of each drug is made by the assistance of expert doctors from Iraqi hospitals and from reliable international resources in [1][2][3][4]. II. LITERATURE SURVEY The side effects of the prescribed medications are a common occurrence in the medical history of patients. Jenna, M. R., et. al [5] presented the opportunity to identify new side effects efficiently through electronic healthcare databases. A confirmation of concept method is suggested. It learns common associations and uses this knowledge to automatically refine exposure-outcome associations (i.e. side effect signals). A novel measure is determined. It is named the confounding-adjusted risk value, an accurate absolute risk value of a patient. They show that signals filtering might be possible at a patient level according to association rules acquired by taking into account patients’ medical histories. Chen, Jie, et. al [6] developed an association classification algorithm in which rules discovered can be used to alert medical practitioners when prescribing drugs. They propose two kinds of probability trees that can present clearly the risk of specific adverse drug reactions to prescribers. Chen, Yu, et. al [7] used association rule mining approach to derive possible side effects due to exposure to multiple drugs at different durations of the pregnancy. They derived sequential temporal rules to discover new information that would not be detected by the traditional analysis method that is currently used by pharmacists. The optimizing drug dose is a major factor in improving the sustainability of health care. Daughton, C. G., and Ruhoy, I. S. [8] presents the first critical examination of the multi-faceted role of drug dose in reducing the ambient levels of active 140 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 pharmaceutical ingredients (APIs) in the environment. Personalized adjustment of drug dose holds the potential for enhancing therapeutic outcomes while simultaneously lowering the incidence of adverse drug events and in lowering patient healthcare cost. III. EXPERT SYSTEM An expert system is an application that tries to react like a human expert on a special subject area. This system is used to advise non-experts in situations where a human expert is unavailable [9]. An expert system consists of three parts [9]: (see Figure 1) A user interface - that accepts a non-expert user and it asks questions of the expert system, and is conferred with advice. The interface should be a simple and easy to use [9]. A knowledge base - It is a group of facts and rules. It is built with information that is supplied by human experts [9]. An inference engine - this acts as a search engine, inspecting the knowledge base for a particular arrangement that matches the user's query [9]. Fig. 1. An Expert system works in block diagram [9] An example of expert systems is diseased diagnosis (the input would contain medical history information, the symptoms of the patient. Those might be the query, and the answer is a diagnose for the patient’s mental or physical pain). However, Nowadays, “conventional” computer programs may achieve some of the typical expert system functionalities. When using, expert systems problems are solved using heuristics or approximate methods which, unlike traditional applications, are not guaranteed to present an optimal solution [9]. IV. GENERAL DESCRIPTION FOR ESTIMATING THE BEST MEDICINE SYSTEM This section describes the proposed Estimation Medicine for Diseases (EMD) System framework, algorithms, and the applied rules. Figure (2) illustrates a block diagram for selection medicine for each patient's disease. Fig. 2. Block Diagram Estimation Medicine for each Patient's Disease The method used in the system makes use of dataset for heart diseases along with medicine dataset for medical data. The obtained medical data are utilized in the formation of the expert system. The system utilizes in a convenient way the concepts and algorithms introduced in the following subsections. A. Reliable rules for safer drug use This section explains how the doctor selects the best medicine for each patient's disease for conditions by some rules basing on safer drug use; these rules are expressed in [10]. Those rules were compiled for more dependable drug use (or nonuse). They were listed particularly from the World Health Organization’s General Prescribing Principles for the Elderly. Nevertheless, these rules might be applied to all ages. Doctors and patients involved in drug therapy should know them. The guiding principle is to use those rules as guidelines for patient treatment according to symptoms and patient’s medical history. Among those rules are [10]: • Additional drugs used may make the earlier drugs much more dangerous [10]. • The primary doctor ought to coordinate patient’s care and drug use [10]. • The patient should make sure that drug therapy is needed [10]. 141 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 • Important, believable, and adverse effects of the drug in addition to plausible drugs interactions should be foreseeable in advance [10]. • Any adverse drug reactions patient may have should be taken into consideration early [10]. B. Recognize-Act Cycle Since the production system is a formulation, that the expert system’s inference engine can process efficiently, they can be used to generate new information. The basic recognize-act cycle consists of four steps [11]: 1) Recognize (get) matches. 2) Choose a match. 3) Act (perform the selected production). 4) Return to step (1). It matches the premise patterns of the rules against elements in the working memory. If there is more than one rule that can be applied, use a conflict resolution strategy to choose one to apply. Early production systems spent over 90% of their time doing pattern matching, but there is now a solution to this efficiency problem which is the Rete algorithm [11]. C. Rete Algorithm The traditional approach for recognize-act cycle matches all Elements (E) in working memory along with all premises (P) in all rules (R). It then checks E×P×R possibilities in each cycle. Nevertheless, the rules have structural similarity, since parts of their conditions are common. Where the application of any similar rule changes the working memory slightly (temporal redundancy) [11]. Rete Algorithm uses these facts to improve efficiency (‘rete’ is Latin for ‘net’). First, the initial conflict set is generated through the matching algorithm by connecting the network for all condition parts. After that, only the changed elements in working memory are fed into the network to distinguish the changes in the conflict set [11]. D. Estimation Medicine for Diseases (EMD) algorithm The above concepts and foresight are consolidated into a single unified algorithm called Estimation Medicine for Diseases (EMD) which performs the main system functionality. Algorithm: Estimation Medicine for Diseases (EMD) Input: Diseases Table (D), Medicine Table (M), Disease Name for Doctor Diagnosis (DN). Output: Medicines Name (MsN), Error Ratio(ER). Begin: • Input patient's History and Symptoms (PHS). • For each (D)That has name (DN) • Find Diseases Symptoms (DS) • Find Similar Symptoms between(PHs) and (DS) and put it on new array (SS) • (If SS is found in M Then MsN( )→ Count +1)//Compare (SS) with (M) for (DN) • Sort MsN() • Select the minimum Error Ratio • If the value of MsN is equal, then select the high priority from (M) End V. MEDICAL DATASET (DISEASES, MEDICINE) The data was collected from an Iraqi hospital. Patients' files and information were extracted from the Statistics Division at Al- Kindi Teaching Hospital. The data for each disease consists of medical history, symptoms, laboratory analysis, etc. Those have been entered into the database. Each disease has ten medicines that were fed into the system. (See Table I). In this table, each column represents a single disease with its well-known ten medicines. The following are some of the heart diseases: • High Blood Pressure (I10). • Angina pectoris (I20). • Myocardial Infarction (I21). • Atrial Fibrillation (I48). • Heart Failure (I50). Note that I10, I20, I21, I48, and I50 represent the statistical identifiers of the diseases. TABLE I. DISEASES AND MEDICINES ID I10(HighBl ood Pressure) I20(Angina pectoris) I21(Myocar dial Infarction) I48(Atrial Fibrillation) I50(Heart Failure) 1 Diazoxide Glyceryl_trinitrate Atenolol Digoxin Captopril 2 Hydralazine Aspirin Metoprolol Dronedarone Cilazapril 3 Sodium Nitroprussid e Clopidogrel Acebutolol Warfarin Enalapril maleate 4 Methyldopa Prasugrel Propranolol Hydrochlorid e Acenocoum arol Lisinopril 5 Captopril Heparin Labetalol Hydrochlorid e Phenindione Valsartan 6 Lisinopril Streptokinase Pindolol Diltiazem Telmisartan 7 Moexipril Hydrochlori de Urokinase Sotalol Hydrochlorid e Verapamil Olmesartan Medoxomil 8 Fosinopril Sodium Tenecteplase Timolol maleate Propranolol Hydrochlori de Losartan potassium 9 Enalapril Propranolol Hydrochloride Carvedilol Acebutolol Irbesartan 10 Cilazapril Atenolol Bisoprolol Atenolol Eprosartan A large amount of data is required to train the system since poor of data does not give high accuracy in diagnosis. Each represented disease has its specific table. Each table consists of some attributes. These attributes are different from one table to another. All data are taken from the patient's file, which was read by doctors, who helped me and taught me how to read and understand each patient's file. Training datasets consist of two groups: the first is made of five tables where a table is specified for each medical case in 142 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 which the patient data are taken from the patient files. These are the symptoms and patient medical history. While the second category contains five tables, where a table is designated for each disease. This table has ten drugs commonly identified with that disease, yet each drug has a set of side effects. The name of the medicine is the scientific name as it is an international standard. Most of the researches are different from this one because in these researches the acquisition of data is made from the internet and they are almost based on one disease. The results of these researches indicate that the disease may exist or not and they don't give the accurate medicine for the patient and as a result they cannot tell all the symptoms that the patient suffers from and hence they cannot prevent the misprescribing of drugs" VI. TYPICAL SYSTEM OPERATIONS One user interface of the system explained in this section is shown in Figure (3). This window is called as one of some sub- windows. Learning in the system depends on the data stored in the system (training dataset) and data tested by the system. As an example, the system works to diagnose that the patient has heart diseases by using functional (EMS) strategies. The system deals with heart disease, which is diagnosed as the patient's disease. Doctor selects any heart disease, as shown in Figure (3) where the patient is suffering from one of them, but it is unknown for his doctor at this moment. A sample medical history data is (patient's age: more than fifty years (age), Increase of blood cholesterol (cholesterol), Hypertension (HT), Respiration Rate (RespRate) and Blood Pulse High (BPH)). The symptoms are (Shortness of Breath, Dizziness, Weakness, Rapid heartbeats, Palpitation, Headache and General malaise). The doctor in the hospital diagnosis that patient is suffering from High Blood Pressure and Myocardial Infarction. When his medical history and symptoms are entered into the system, the result would be selecting the best medicine for diagnosis to High Blood Pressure (I10) and Myocardial Infarction (I21). Hence, the doctor has selected medicine by the system (EMS) depending on all information and side effects of that medicine, symptoms, and history of this patient. Fig. 3. A heart disease diagnosis and selected medicine by (EMD) VII. EXPERIMENTAL RESULTS This section illustrates the experimental results of the proposed approach described in the previous sections. A. Settings The experimental results also include Rete Algorithm for Medicine. The results are organized into a table for heart diseases. Brief descriptions and the diagnosis were given by the doctors and also by the proposed system—Estimation Medicine for Diseases (EMD) that offers selecting the best medicine for a given disease. These approaches are running on a laptop (running Core i5 CPU@ 1.7MHz, RAM 4GHz on Windows 7 Home Basic), SQL Server 2012 for building Database and Visual Basic.NET 2012,2013 (VB.NET) environment as the programming language used in the implementation. B. Results The proposed system deals with many types of heart diseases. System experiments on the applied fed data. Each stored table is used to test different disease cases. The system was configured over five categories representing the selected heart diseases. Table I demonstrate this training dataset. The system has been tested by entering symptoms and medical history of each patient. Data are entered into the built structure, then calculations of the probability of occurrences of the disease for each patient are made. This result is the likelihood of disease for each patient along with the selected medicine. The doctors can choose two of the top up possibilities of these medicines (i.e. the higher priority among them). The performance of the system is expressed in Table II. Where we can see twenty- two patients for heart diseases. After that accuracy is calculated by (true (cases) / (true (cases) + false (cases))) to get the results for the system model of (EMD). The performance of the EMD can be seen for each disease. Hence, the accuracy rate of the given model of (EMD) is approximately (87.26%) in heart diseases. It can be deduced that the rate of accuracy of the system (EMD) is the best average results. These results are shown in table II while the running time is presented in Table III. TABLE II. EMD PERFORMANCE FOR THE SELECTED MEDICINE Disease code Disease name Number of patients used in testing The selected medicine for disease Accuracy (%) Error Ratio (%) I10 High Blood Pressure 22 81.8% 18.2% I20 Angina pectoris 22 86.36% 13.64% I21 Myocardial Infarction 22 90.9% 9.1% I48 Atrial Fibrillation 22 90.9% 9.1% I50 Heart Failure 22 86.36% 13.64% Average 87.26% 12.74% 143 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 TABLE III. TRAINING DATASET RUNNING TIME FOR HEAR T DISEASES AND MEDICINE DATASET Model for Heart Diseases Time (Sec) SBM 0.30 Training dataset running time for heart diseases and medicine dataset is (0.30) sec because it is only a time to choose the best medicine in the system yet it is not for diagnosing the disease. The system is an assistant to the doctor, and it is a complementary tool for the medical diagnosis. Figure (4) demonstrates a chart indicating the Performance of (EMD) for Heart Diseases. Training datasets consist of two groups: the first is made of five tables where a table is specified for each medical case in which the patient data are taken from the patient files. These are the symptoms and patient medical history. While the second category contains five tables, where a table is designated for each disease. This table has ten drugs commonly identified with that disease, yet each drug has a set of side effects. The name of the medicine is the scientific name as it is an international standard. Fig. 4. The performance of (EMD) for Heart Diseases VIII. CONCLUSIONS In this paper, an approach based on Rete Algorithm is used where the best-chosen medicine through the system (EMD) is suggested. We want to reach into the best model using the algorithm to obtain a more accurate selection for an optimal medicine for diagnosis. The doctor can select medicine by the system (EMD) depending on all information and side effects of that medicine, symptoms and medical history for a patient. The system (EMD) is tested by the doctors in Iraqi hospitals. The accuracy of selecting the medicine to heart diseases is approximately (87.26%). This accuracy is obtained because the principles of nominating drugs by a doctor for a given patient are based on two factors. The first factor is the symptoms experienced by the patient along with his medical history, while the second one is the chemical interactions that happen inside the patient's body. Our system depends on the first factor in which the selection of medicine is made to avoid the interfering with other medicaments for a certain disease. It doesn’t concentrate on more of other suspicious diseases but simply treats the current symptoms that are experienced by the patient. REFERENCES [1] U.S. Pharmacopeia National Formulary, “United States Pharmacopeia (U.S. Pharmacopeia National Formulary),” Us Pharmacopeia ,Pck Slp edition, ISBN-13, January 1, 2015. [2] U K Stationery Office, “British Pharmacopoeia (BP) 2016,” Stationery Office Books (TSO), ISBN-13, August 24, 2015. [3] Sid M. Wolfe, "Worst Pills, Best Pills: A Consumer's Guide to Avoiding Drug-Induced Death or Illness," Gallery Books; Revised ed. edition, ISBN- 13, January 4, 2005. [4] BMJ Group , Pharmaceutical Press , " BNF 69: March 2015 - September 2015 (British National Formulary)," Pharmaceutical Pr, 1 edition, ISBN- 13, March 31, 2015. [5] Jenna M. Reps, Uwe Aickelin, Jiangang Ma, Melbourne, Yanchun Zhang, "Refining Adverse Drug Reactions using Association Rule Mining for Electronic Healthcare Data," arXiv:1502.05943v1 [cs.DB] , 20 Feb 2015. [6] Jie Chen , Hongxing He ,JiuyongLi , Huidong Jin , Damien McAullay ,Graham Williams, Ross Sparks and Chris Kelman, "Representing Association Classification Rules Mined from Health Data,” R. Khosla et al., Eds. Springer-Verlag Berlin Heidelberg, 2005, pp. 1225–1231. [7] Yu Chen, Lars Henning Pedersen, Wesley W. Chu, Jorn Olsen, “Drug Exposure Side Effects from Mining Pregnancy Data,” ACM SIGKDD Explorations Newsletter - Special issue on data mining for health informatics, Vol. 9 Issue 1, June 2007 pp. 22 - 29 . [8] Christian G. Daughton , Ilene Sue Ruhoy, "Lower-dose prescribing: Minimizing “side effects” of pharmaceuticals on society and the environment," Elsevier B.V. , CC BY-NC-ND license,2012. [9] Nwigbo Stella N , Agbo Okechuku Chuks,"Expert System: A Catalyst in Educational Development in," Human Resource Management Academic Research Society, Proceedings of the 1st International Technology, Education and Environment Conference, 2011. [10] Brian R. Walker BSc MD FRCPE FRSE , Nicki R Colledge BSc (Hons) FRCPE , Stuart H. Ralston MD FRCP FMedSci FRSE , Ian Penman BSc MD FRCPE, "Davidson's Principles and Practice of Medicine: With STUDENT CONSULT Online Access, 22e (Principles & Practice of Medicine (Davidson's)) 22nd Edition," Churchill Livingstone, ISBN-13, 22 edition ,February 15, 2014. [11] Robert B. Doorenbos," Production Matching for Large Learning Systems," Robert B. Doorenbos, Thesis for the degree of Doctor of Philosophy, CMU-CS-95-113 ,January 31, 1995. 144 | P a g e www.ijacsa.thesai.org https://www.amazon.com/s/ref=dp_byline_sr_book_1?ie=UTF8&text=U+K+Stationery+Office&search-alias=books&field-author=U+K+Stationery+Office&sort=relevancerank https://www.amazon.com/s/ref=dp_byline_sr_book_1?ie=UTF8&text=Sid+M.+Wolfe&search-alias=books&field-author=Sid+M.+Wolfe&sort=relevancerank https://www.amazon.com/s/ref=dp_byline_sr_book_1?ie=UTF8&text=BMJ+Group&search-alias=books&field-author=BMJ+Group&sort=relevancerank https://www.amazon.com/s/ref=dp_byline_sr_book_2?ie=UTF8&text=Pharmaceutical+Press&search-alias=books&field-author=Pharmaceutical+Press&sort=relevancerank https://www.amazon.com/s/ref=dp_byline_sr_book_1?ie=UTF8&text=Brian+R.+Walker+BSc+MD+FRCPE+FRSE&search-alias=books&field-author=Brian+R.+Walker+BSc+MD+FRCPE+FRSE&sort=relevancerank https://www.amazon.com/s/ref=dp_byline_sr_book_2?ie=UTF8&text=Nicki+R+Colledge+BSc+%28Hons%29+FRCPE&search-alias=books&field-author=Nicki+R+Colledge+BSc+%28Hons%29+FRCPE&sort=relevancerank https://www.amazon.com/s/ref=dp_byline_sr_book_2?ie=UTF8&text=Nicki+R+Colledge+BSc+%28Hons%29+FRCPE&search-alias=books&field-author=Nicki+R+Colledge+BSc+%28Hons%29+FRCPE&sort=relevancerank https://www.amazon.com/s/ref=dp_byline_sr_book_3?ie=UTF8&text=Stuart+H.+Ralston+MD+FRCP+FMedSci+FRSE&search-alias=books&field-author=Stuart+H.+Ralston+MD+FRCP+FMedSci+FRSE&sort=relevancerank https://www.amazon.com/s/ref=dp_byline_sr_book_4?ie=UTF8&text=Ian+Penman+BSc+MD+FRCPE&search-alias=books&field-author=Ian+Penman+BSc+MD+FRCPE&sort=relevancerank https://www.amazon.com/s/ref=dp_byline_sr_book_4?ie=UTF8&text=Ian+Penman+BSc+MD+FRCPE&search-alias=books&field-author=Ian+Penman+BSc+MD+FRCPE&sort=relevancerank I. Introduction II. Literature Survey III. Expert System IV. General Description for Estimating the Best Medicine System A. Reliable rules for safer drug use B. Recognize-Act Cycle 1) Recognize (get) matches. 2) Choose a match. 3) Act (perform the selected production). 4) Return to step (1). C. Rete Algorithm D. Estimation Medicine for Diseases (EMD) algorithm Algorithm: Estimation Medicine for Diseases (EMD) V. Medical Dataset (Diseases, Medicine) VI. Typical System Operations VII. Experimental Results A. Settings B. Results VIII. Conclusions References 
work_mgmhrb4wzjephif7zdvx6xha3m	----	Carnegie Mellon University research repository - Browse Browse Discover research from Carnegie Mellon University Follow AllCategoriesgroupsSEARCH Hide footerAboutHow to DepositPreparing DataDMP ToolDeposit SupportContactDeposit AgreementTermsPrivacyToolsFAQsDisclaimerSitemap figshare. credit for all your research. 
work_mgqmivrbtnge3fttkeec5zdg24	----	Edinburgh Research Explorer Evolutionary FineTuning of Automated Semantic Annotation Systems Citation for published version: Cuzzola, J, Jovanovic, J, Bagheri, E & Gasevic, D 2015, 'Evolutionary FineTuning of Automated Semantic Annotation Systems', Expert Systems with Applications, vol. 42, no. 20, pp. 6864-6877. https://doi.org/10.1016/j.eswa.2015.04.054 Digital Object Identifier (DOI): 10.1016/j.eswa.2015.04.054 Link: Link to publication record in Edinburgh Research Explorer Document Version: Peer reviewed version Published In: Expert Systems with Applications General rights Copyright for the publications made accessible via the Edinburgh Research Explorer is retained by the author(s) and / or other copyright owners and it is a condition of accessing these publications that users recognise and abide by the legal requirements associated with these rights. Take down policy The University of Edinburgh has made every reasonable effort to ensure that Edinburgh Research Explorer content complies with UK legislation. If you believe that the public display of this file breaches copyright please contact openaccess@ed.ac.uk providing details, and we will remove access to the work immediately and investigate your claim. Download date: 06. Apr. 2021 https://doi.org/10.1016/j.eswa.2015.04.054 https://doi.org/10.1016/j.eswa.2015.04.054 https://www.research.ed.ac.uk/portal/en/publications/evolutionary-finetuning-of-automated-semantic-annotation-systems(a441cfe2-919d-49a2-8037-47256495b81e).html   Evolutionary Fine­Tuning of Automated Semantic Annotation Systems    John Cuzzola​a​ (jcuzzola@ryerson.ca), Jelena Jovanovic​b​ (jeljov@gmail.com),   Ebrahim Bagheri​a​ (bagheri@ryerson.ca), Dragan Gasevic​c​ (dgasevic@acm.org)  a​ ​Laboratory for Systems, Software and Semantics (LS​3​), ​http://ls3.rnet.ryerson.ca​, Ryerson University, Ontario, Canada  b​ ​Faculty of Organizational Sciences (FOS), ​http://www.fon.bg.ac.rs​, University of Belgrade, Belgrade, Serbia  c ​Schools of Education and Informatics​, ​http://www.ed.ac.uk/schools-departments/informatics​, University of Edinburgh, Scotland, United                        Kingdom    Abstract. ​Considering the ever­increasing speed at which new textual content is generated, an efficient and effective                                use of large text corpora requires automated natural language processing and text analysis tools. A subset of such tools,                                      namely automated semantic annotation tools, are capable of interlinking syntactical forms of text with their underlying                                semantic concepts. The optimal performance of automated semantic annotation tools often depends on tuning the                              values of the tools’ adjustable parameters to the specificities of the annotation task, and particularly to the                                  characteristics of the text to be annotated. Such characteristics include the text domain, terseness or verbosity level, text                                    length, structure and style. Since the default configuration of annotation tools is not suitable for the large variety of                                      input texts that different combinations of these attributes can produce, users often need to adjust the annotators’ tunable                                    parameters in order to get the best results. However, the configuration of semantic annotators is presently a tedious and                                      time consuming task as it is primarily based on a manual trial­and­error process. In this paper, we propose a Parameter                                        Tuning Architecture (PTA) for automating the task of configuring parameter values of semantic annotation tools. We                                describe the core fitness functions of PTA that operate on the quality of the annotations produced, and offer a solution,                                        based on a genetic algorithm, for searching the space of possible parameter values. Our experiments demonstrate that                                  PTA enables effective configuration of parameter values of many semantic annotation tools.    Keywords.​ ​semantic annotation, automated configuration, genetic algorithm, parameter learning    I. Introduction  The quantity and variety of unstructured textual content has rapidly increased over the last few years,                                leading large and small organizations towards seeking solutions that enable effective and efficient use                            of both the internally produced textual content, and the content originating from the Web . Considering                             1 the amount of textual content and the speed at which it has to be processed, it is gradually becoming                                      evident that automated machine comprehension of text is a necessity, if the objectives of efficiency and                                effectiveness were to be reached. This has led to an increased research focus, both in academia and                                  industry, on text mining, natural language processing and other related Artificial Intelligence fields                          (Hovy, Navigli, & Ponzetto, 2013​), and resulted in numerous proposals and specific software solutions                            for addressing some aspects of text comprehension through, for example, named entity extraction                          (Ratinov & Roth, 2013; Atdağ & Labatut, 2013), relation extraction (Yan, Okazaki, Matsuo, Yang, &                              Ishizuka, 2009; Weston, Bordes, Yakhnenko, & Usunier, 2013), and sentiment analysis (Liu, 2012).  Automated semantic annotation of textual content addresses an important aspect of text comprehension,                          namely, the extraction and disambiguation of entities and topics mentioned in or related to a given                                piece of text (​Uren et al., 2005​). Each identified entity is disambiguated, i.e., unambiguously defined,                              by establishing a link to an appropriate entry (concept or instance) in a knowledge base that uniquely                                  1http://blog.digitalinsights.in/social­media­users­2014­stats­numbers/05205287.html    1  http://ls3.rnet.ryerson.ca/ http://www.fon.bg.ac.rs/ http://www.ed.ac.uk/schools-departments/informatics http://blog.digitalinsights.in/social-media-users-2014-stats-numbers/05205287.html identifies the entity and provides further information about it. This task, also known as ​entity linking                                (​Hachey, Radford, Nothman, Honnibal, & Curran, 2013​)​, typically relies on large, general­purpose,                        Web­based knowledge bases, such as Wikipedia and other more structured knowledge bases such as                            DBpedia (​http://dbpedia.org​), YAGO (​http://www.mpi­inf.mpg.de/yago­naga/yago/​), and Wikidata            (​http://wikidata.org​).  Tools and services for automated semantic annotation of text are offered by a constantly increasing                              number of companies and research groups (​Jovanovic et al., 2014​). Major Internet players are also very                                active in this area. For instance, to fulfill its well known mission of “organizing the world’s                                information”, Google is continuously evolving its proprietary knowledge base – the Knowledge Graph                          – and according to one Google executive, “every piece of information that we [Google] crawl, index, or                                  search is analyzed in the context of Knowledge Graph” . In addition, Google has been working on a                                 2 probabilistic knowledge base, named Knowledge Vault, that combines automated extraction of facts                        from the Web and prior knowledge derived from existing knowledge bases (​Dong ​et al., 2014).                              Similarly, Microsoft is developing its own knowledge repository called Satori and using it to                            semantically index content and thus improve both its search engine Bing and the applications running                              on Windows . 3 In (​Jovanovic et al., 2014​), we have provided a comprehensive descriptive comparison of the                            state­of­the­art semantic annotation tools by considering numerous features, especially those that could                        be relevant for selecting the right tool(s) to use in a specific application case. One common                                characteristic of all the reviewed tools is that they need to be optimally configured in order to give their                                      best results when working with different kinds of texts – such as texts of diverse level of formality,                                    length, domain­specificity, and use of jargon. While the examined annotators provide default                        configuration of their parameters suitable for some annotation tasks, to our knowledge, no single                            annotator can reach its best performance on all kinds of text with one single configuration.                              Furthermore, the quality of an annotator’s output is not a category that could be assessed in absolute                                  terms; instead, it depends on the application case, i.e., on the specificities of the requirements that stem                                  from a particular context of use (​Maynard, 2008​). For instance, in some cases a very detailed                                annotation would be required and highly valued, whereas in other cases a terse annotation of only the                                  most relevant entities would be considered the best output. This indicates that in order to get the best                                    from a semantic annotation tool, one should configure it according to the specificities of the intended                                context of use, including both the characteristics of the text to be annotated and the requirements of the                                    annotation task (e.g., precision/recall trade­off).  Configuration of semantic annotators is not an easy task, for at least two reasons. First, since an                                  annotator’s configuration parameters are closely tied to the tool’s internal functioning, it is difficult to                              expose them in a manner that would enable users to effectively and efficiently use the tool without                                  having to know the details of the tool’s inner logic. In other words, the first challenge is in enabling                                      users to tune the annotator with respect to the key issues such as specificity and comprehensiveness of                                  annotations, without them being concerned with the details of the tool’s parameters. The second                            challenge stems from the fact that those configuration parameters are not mutually independent but                            2 ​http://goo.gl/mZ7a9H   3 ​http://goo.gl/iDwP2x   2  http://dbpedia.org/ http://www.mpi-inf.mpg.de/yago-naga/yago/ http://wikidata.org/ http://goo.gl/mZ7a9H http://goo.gl/iDwP2x interact with one another, so that one has to find an optimal combination of parameter values for a                                    specific application case. Moreover, annotators may have many parameters, and some of those                          parameters are continuous variables, thus making the tuning task very time consuming. As the                            state­of­the­art annotators do not provide support for finding an optimal parameter combination for a                            specific annotation task, it is often done manually, through a trial­and­error process. For example,                            consider the commercial semantic annotator ​TextRazor ​whose best practices state the following:    “Experiment with different confidence score thresholds...If you prefer to avoid false­positives                      in your application you may want to ignore results below a certain threshold. The best way to                                  find an appropriate threshold is to run a sample set of your documents through the system and                                  then manually inspect the results.”   4 To our knowledge, no solution to the above stated problem of parameter configuration has been                              reported in the literature. Therefore, in this paper, we make the following contributions:  ● Parameter Tuning Architecture (PTA) to automate the task of parameter value selection for a                            user­supplied testing set; thus resulting in performance that is better or at least equal to the                                tool’s performance with its default parameter values.  ● Five variations of the fitness function that emphasize different aspects of annotation quality                          (namely most annotations produced, most known correct annotations, least unknown                    annotations, best recall/precision), and a means to identify which variation performed the best                          for a particular testing set.   ● A method to efficiently search the solution space of possible parameter values using a Genetic                              algorithm.  The proposed Parameter Tuning Architecture (PTA) is applicable to a variety of automated semantic                            annotators, since its core component of the fitness function is not concerned with any textual or                                annotator­specific features but rather metrics based on known correct, known incorrect, or unknown                          annotations produced. To search the space of possible solutions, i.e., possible configurations of                          parameter values, we rely on a Genetic algorithm (for the reasons given in Section IV), although PTA                                  can also be applied with other methods for searching a large solution space (e.g., evolutionary                              algorithms or probabilistic methods). Our experiments with PTA have demonstrated that PTA can be                            used as an effective configurator for automated semantic annotators.    After more precisely defining and illustrating the problem of parameter tuning in the context of                              semantic annotation tools (Sect. II), and the associated challenges (Sect. III), in Section IV, we present                                PTA in detail. Section V reports on the experiments that we performed in order to evaluate the PTA’s                                    ability to find a set of parameter values that provides an adequate level of the annotator’s output while                                    minimizing annotation errors. The experimental results and the overall proposal are further critically                          discussed and summarized in Section VI, while Section VII positions the contributions of our work                              with respect to related research work. Lastly, we acknowledge the limitations of our solution and                              propose future experiments before we conclude our paper (Sections VIII / IX).       4 ​https://www.textrazor.com/docs/rest#optimization   3  https://www.textrazor.com/docs/rest#optimization II. Problem Definition     As indicated in the Introduction, today’s automated semantic annotators offer a variety of tunable                            parameters in order to produce results accordant with the desired level of granularity, precision, and                              recall. There is no single “best” configuration as this is a function of various factors: is the text we are                                        annotating restricted to a specific topic or domain such as history, food, or politics? Is the input text                                    descriptive with verbose and meaningful wording or is it terse with numerous empty (stop) words?                              What is the length, style, and structure of the input text: paragraph, single sentence, or tweet?                                Therefore, we must decide on a ​gold-standard​, a manually labelled training or testing set, that contains                                such factor­specific target questions that the annotator will be exposed to. Observe, as in the TextRazor                                introduction example, that an annotator would already be trained with default parameter values; thus,                            PTA uses the gold standard as an evaluation/testing set to tailor the annotator’s parameters to the kind                                  of input text represented by the gold standard.      Further difficulties arise when a parameter configuration comprises many individual or continuous                        floating point parameters resulting in an exponential number of possible combinations that make a                            complete search of this space unrealistic. To illustrate this, consider Table 1 showing how the ​TagME                                semantic annotator (​Ferragina & Scaiella, 2012​) choses to disambiguate the same spot (“​field​”), when                            different values of the tool’s two tunable parameters (​epsilon ​and ​long_text​) ​are used. As the table                                indicates, even slight changes to the values of epsilon and long_text can have significant impact on the                                  disambiguation process and lead to absurd results.     Table 1​. Disambiguation of the spot “​field​” in the sentence “​Gesture or salute? A soccer star who made the sign on the                                            field says otherwise.​” for different values of TagMe’s tunable parameters ​epsilon​ and ​long_text​.  Configuration  Disambiguation of the spot “field” by TagMe  epsilon = 0.39  long_text = 0  Wikipedia Reference​: Field (agriculture)  Abstract​: In agriculture, the word field refers generally to an area of land enclosed or                              otherwise and used for agricultural purposes.  epsilon = 0.30  long_text = 1     Wikipedia Reference​: Field (mathematics)  Abstract​: In abstract algebra, a field is a ring whose nonzero elements form a                            commutative group under multiplication.  epsilon = 0.52  long_text = 4  Wikipedia Reference​: Academic discipline  Abstract​: An academic discipline, or field of study, is a branch of knowledge that is                              taught and researched at the college or university level.  epsilon = 0.55  long_text = 6  Wikipedia Reference​: Field (physics)  Abstract​: In physics, a field is a physical quantity associated with each point of                            spacetime.     It is this difficulty of choosing appropriate parameter values that provides the motivation for the work                                presented in this paper. To our knowledge, the problem of selecting appropriate values for                            4  configuration parameters of semantic text annotators has not yet been reported in the literature. In the                                following sections, we first further explain the challenges associated with tuning parameters of                          semantic annotators, and then proceed to introduce our Parameter Tuning Architecture (PTA) and a                            method to find suitable parameter values in reasonable time.    III. Challenges in Tuning Parameters of Semantic Annotators     Our proposed method involves a fitness function to evaluate the output of an annotator given an                                arbitrary configuration. PTA’s fitness function scores the annotator’s output against a testing set                          containing oracle­identified correct (C) and/or incorrect (E) annotations. Table 2 gives the four possible                            combinations of labelled annotations within the testing set. NOT(¬) indicates that any correct or                            incorrect annotations are absent from the testing set.     Table 2​. The four kinds of testing sets based on the presence/absence of correct (C) and incorrect (E) labels.  Combination (1,2,3,4)  Correct Annotations  (positive labels)  Incorrect Annotations    (negative labels)  Testing Model  #1  C  E  two­class supervised  #2  C  ¬E  one­class supervised  positive labels only  #3  ¬C  E  one­class supervised  negative labels only  #4  ¬C  ¬E  unsupervised    Combination #4 assumes a testing set with no labels, and thus is an unsupervised model that PTA                                  cannot take advantage of; hence, it will not be further considered. In contrast, combinations #1, #2, and                                  #3 are supervised models that variants of the PTA fitness function can utilize. However, supervised                              configuration of semantic annotators have caveats when relying on testing sets. Namely:    caveat (i)​: A testing set that depends on negative labels (#1, #3) cannot feasibly cover all                                possible mistakes that an annotator could output for an input text.     caveat (ii)​: Due to (i), publicly available testing sets with labelled errors are difficult to find,                                particularly one­class (error­only) testing sets (#3).    caveat (iii)​: Testing sets with correct labels (#1, #2) often specify only a few correct                              annotations that the annotator should identify from the input text. However, the output could                            contain many more correct annotations not mentioned in the testing set.    5  caveat (iv)​: An oracle may specify correct but not necessarily ideal annotations. The annotation                            tool may find an annotation that is better than or of equal semantic quality to that of the                                    oracle­recommended annotation.     caveat (v)​: Regardless of the form of the testing set used (#1, #2, #3), an annotator will often                                    produce annotations that are not explicitly identified in the testing set, thus forcing any method                              of evaluation to make assumptions for the unidentified annotations (assume correct, assume                        incorrect, or ignore). Consequently, the quality of the annotator’s output in relation to the                            testing set provided is questionable.    To illustrate the above given statements, we obtained the ​wiki-annot30 dataset of A​3 labs from the                                University of Pisa Computer Science Department available under a Creative Commons License . This                         5 set contains 186,000 short text fragments from Wikipedia with correct­only identified annotations                        (combination #2) and was constructed using the procedure described in ​Ferragina and Scaiella ​(2012).                            Table 3 summarizes the output on one of these text fragments produced by the TagME semantic                                annotator with default configuration. The output is given alongside the A​3 dataset gold standard. The                              table gives the Wikipedia pages that are linked to the identified spot. The TagME column also shows                                  whether the linked Wikipedia page for the spot is true correct (C) or a true error (E).    Table 3​. Wikipedia entities produced by TagME and compared with the A​3 gold standard for the text fragment “​It is                                        home to the American University Eagles basketball and volleyball teams. The arena, named for Washington DC                                philanthropists, Howard and Sondra Bender​”. TagME links are identified as (C)orrect or (E)rror.   Spot  A​3​ dataset  TagME  volleyball   wikipedia:Volleyball  wikipedia:Volleyball (C)  washington dc  wikipedia:Washington,D.C.  wikipedia:Washington,D.C. (C)  basketball   wikipedia:Basketball  wikipedia:Basketball (C)  the american university (eagles)  wikipedia:American  University  wikipedia:American Eagles (C)  arena   wikipedia:Arena  wikipedia:Arena (C)  home  Not available  wikipedia:Home(sports) (C)  teams  Not available  wikipedia:NHL (E)  howard  Not available  wikipedia:Howard University (E)  bender  Not available  wikipedia:Gary Bender (E)    From the output we can see that TagME correctly linked more spots than what was listed in the A​3                                      dataset (caveat ​iii​). Specifically, the ‘home’ spot is a correct entity mention absent from A​3​. Further,                                absent from the A​3 dataset is the correct spot for ‘team’ that was linked incorrectly by TagME along                                    with incorrect links for ‘howard’, and ‘bender’. Consequently, to assess the accuracy of the output, one                                must first consider how these unknown spots should be treated (caveat ​v​). Lastly, consider the spot ‘the                                  5 http://acube.di.unipi.it/tagme­dataset  6  american university (eagles)’. A​3 recommends the Wikipedia link of ‘American University’ while                        TagME prefers the link ‘American Eagles’. Both suggestions are correct since the official sports team                              of American University is in fact the American Eagles. However, in the context of the text fragment, it                                    is clear that the link to American Eagles is a better (i.e., more precise) choice than the gold standard                                      (caveat ​iv​). A similar situation occurs with the ‘arena’ spot in which TagME agrees with the                                wiki­annot30 dataset although the Wikipedia entity for ‘Bender Arena’ would clearly be superior.    In Section IV we detail our five variations of the Parameter Tuning Architecture (PTA) and how they                                  match to the testing set combinations given in Table 2. We also explain how each variation of PTA                                    addresses the caveats.    IV. The Parameter Tuning Architecture (PTA)    In this section, we derive our Parameter Tuning Architecture as a mathematical model. ​Let ​p​i be a                                tuneable parameter for a semantic annotator. Let be a vector of ​n tuneable parameters              vn                 for a semantic annotator configuration, and let V be the space of all possiblep p ..., ]vn = [ 1, 2, pn                               configurations. Let T be a testing or validation set of any of the supervised combinations of Table 2                                    (#1, #2, or #3). We wish to find a vector   that maximizes the following fitness function:υ                                 eq. 1IT(T) rg max f (T)   F = a νεV[ c ] 2 − f (T)[ e ] 2       here FIT(T) in [− , ] and f (T), (T) in [0, ].  w 1 1 c fe 1     FIT(T) is a real­valued ranking function in the interval [­1,1] consisting of a reward component                              (T)fc   and a penalty The reward is a real­valued scaled measure in the interval [0,1] representing the      (T).fe                             correct annotations discovered by an annotator. Conversely, is a measure of the incorrect              (T)fe               annotations within the same interval [0,1]. Consequently, FIT(T) is a trade­off between the correct                            annotations found by the annotator and the errors the annotator produces. We square and to place                         fc  fe     more importance on many correct and incorrect annotations over a few correct answers and infrequent                              mistakes. The exact formula for computation of and is specific to the variant of FIT(T) proposed              fc   fe                 and is given in sections IV.1 through IV.5.     However, equation 1 faces a key obstacle. Namely, the space of all possible solution vectors V is very                                    large making an exhaustive search of this space intractable. Furthermore, this space may be infinite if                                any of the tunable parameters are real­valued. Consequently, we need to decide on how to efficiently                                search a potentially enormous solution space. Our approach is to use a genetic algorithm (GA) for this                                  task. Genetic algorithms are a class of evolutionary algorithms inspired by nature (​Chiong & Beng,                              2007​). They are an easy to implement technique that works well in optimization problems where                              potential solutions to the given problem can be abstracted to the “chromosomes” model of GA. In our                                  case, the vector of configuration parameters ​v​ can be seen as the required chromosome.      7  Figure 1 outlines our PTA method with GA which begins with a set of randomly created configurations                                  (a). In the next step (b), we randomly select a subset of samples of size ​n from the validation set, then                                          annotate the n­samples (c) using each of the configurations initially created in the step (a). The                                configurations are evaluated against the samples by our fitness function (d). The highest ranking                            configurations generate offspring through crossover (g) and gene mutation (h). The initial n­samples                          from the validation set is updated by removing the oldest x­samples and replacing them with another                                random set of x­samples (i). We partially re­sample the validation set at each generation as a form of                                    bootstrap aggregation to minimize overfitting and improve generalization. The process repeats with the                          next generation of configurations until some stopping condition, such as convergence, is satisfied (f).     Figure 1​. The Genetic Algorithm applied in the Parameter Tuning Architecture (PTA)    Although GA does not guarantee an optimal solution, it often converges to near optimal answers in a                                  relatively short period of time (​Szczerbicka, Becker & Syrjakow​, 1998). It also has the capability to                                search multiple regions of the solution space simultaneously since each member of a population                            occupies a different area than its peers (​Grefenstette, 1992​). This is particularly useful when there are                                many global optimal solutions that could satisfy the problem. Finally, GA is an incremental method                              that allows for starting/stopping the search from its currently best known solutions without beginning                            from scratch. This trait is useful in reinforcement learning or re­tuning the annotator as more examples                                become available (​Moriarty, ​Schultz​ & ​Grefenstette​, 1999).     It is worth noting that our emphasis is on the fitness functions of PTA and less on the choice of                                        evolutionary algorithm for the traversal of the search space. Our future work will investigate the                              comparative effectiveness of other evolutionary algorithms such as swarm optimization. In order to                          provide the basis for a comparative analysis and provide GA implementation details, we have made the                                source code and data available for use under an open source license at:                          http://ls3.rnet.ryerson.ca/annotator/PTA​.    8  http://www.informatik.uni-trier.de/~ley/pers/hd/s/Syrjakow:Michael.html http://www.informatik.uni-trier.de/~ley/pers/hd/s/Schultz:Alan_C=.html http://www.informatik.uni-trier.de/~ley/pers/hd/g/Grefenstette:John_J=.html http://ls3.rnet.ryerson.ca/annotator/PTA http://ls3.rnet.ryerson.ca/annotator/PTA Algorithm 1 outlines the PTA process joining our fitness function (equation 1) with the GA of Figure 1.    Input: i) A one­class positive­labeled validation set from the gold standard. ii) A one­class                            positively­labeled testing set from the gold standard. iii) A semantic annotator.    Output:​ A recommended configuration for input annotator (iii).    Algorithm:    1. For each of these PTA variants: pessimistic (eq. 5), apathetic (eq. 6), delta (eq. 7), optimistic                                (eq. 8) and stochastic (eq. 9)  1.1. Find recommended configuration on the validation set using GA of Figure 1.  2. For each recommended configuration from 1.1:  2.1. Annotate the testing set using the recommended configuration.   2.2. Calculate F​1 ​area under the curve (F​1​­AUC) (figures 2,3,4,5).   3. Best recommended solution is the configuration with highest F​1​­AUC.  Algorithm 1: PTA algorithm for automated evolutionary fine­tuning of semantic annotators.     We now derive various forms of our PTA fitness function, for use within step 1 of Algorithm 1, We                                      focus on the one­class positive labels testing set (Table 2, combination #2) since this is the most                                  prevalent type among the available gold­standard datasets for the entity linking task.     To begin, we define A(X) as the output of a semantic annotator using text fragments from set X as                                      input. Let function C(X) return the set of correct annotations from the set of input text X. Similarly, let                                      E(X) return incorrect annotations from the input text set X. We use the generic term ‘annotation’                                broadly to refer to any kind of output produced by a semantic annotation system (an entity link, a                                    related topic, a keyword) as per the examples provided within the validation or testing set. We define                                  function as a numerical score for the correct annotations of set X, while does the same for  (X)fc                         (X)fe           the errors of X. With these definitions in place, we can identify subsets such as: (i) annotations that                                    match the gold standard A(T)∩C(T), (ii) gold standard annotations the annotator failed to identify                            C(T)\A(T), and (iii) additional annotations with unknown label A(T)\C(T). In the following                        subsections, we present five adaptations of PTA (pessimistic, apathetic, delta, optimistic, stochastic)                        that pair different combinations of these subsets to meet the challenge of constructing a suitable fitness                                function in the presence of uncertainty A(T)\C(T). In section V, we examine the effectiveness of these                                five PTA forms in tuning a variety of annotators.     IV.1 Pessimistic PTA    The basic form of our fitness function is:        eq. 2IT(T)  (C(A(T)))   (E(A(T)))F = fc 2 −  fe 2     9  This basic form is suitable for a two­class supervised testing set (Table 2 #1) because it assumes that                                    we are provided with known correct and incorrect annotations that can be exploited:       eq. 3IT(T)  f (C(A(T)))   (E(A(T)))   (A(T) (T))   (A(T) (T))F =   c 2 − fe 2 = fc ⋂C 2 − fe ⋂E 2     Specifically, we find the correct annotations by considering only those annotations that are            (A(T))C                 explicitly known to us as correct within the testing set . We do the same for the error                    (T) (T)A ⋂C                 component of equation 2 with . In this variant of PTA the annotations that are not part of          (T) (T)A ⋂E                           the testing set (unrecognized) are ignored (Section II caveat ​v​). Although this form operates for a                                two­class labeled testing set, it cannot be used for the commonly encountered one­class positive labels                              testing set (Table 2 #2) due to the absence of any defined errors. To compensate, we modify to                                  fe   assume that unrecognized annotations are most likely errors by default:       eq. 4IT (T)  (A(T) (T))   (A(T)∖ C(T))F P = fc ⋂C 2 − fe 2     We call equation 4 pessimistic because of the assumption that annotations not explicitly labeled as                              correct should be considered erroneous. The final step in the formulation is to define  and   as:f  c f  e      and     eq. 5(A(T) (T)) fc ⋂C =   | A(T) ⋂ C(T) | | A(T)  ⋂ C(T) |  max (A(T)∖C(T)) fe =   | A(T) ∖ C(T) | | A(T)  ∖ C(T) |  max     where the denominator |Y|​max of is the maximum observed count from all of the previously          ( )f |Y | |Y |max                     encountered sets of Y; it is used to normalize the numerator into a value between 0 and 1 inclusive.    A shortcoming of pessimistic PTA is the possibility that the testing set identifies a smaller number of                                  correct annotations than the annotator might generate (Section III caveat ​iii​), as demonstrated in Table                              3. Consequently, those unidentified correct annotations would be missed by thus undermining the                    fc       true quality of the output produced, and would be unfairly counted against the annotator in the error                                  calculation of   (equation 5) or ignored entirely (equation 3).fe     IV.2 Apathetic PTA    In this section, we introduce a version of PTA that allows for ignoring unknown annotations when only                                  a one­class positive testing set is available. Simply, we change the penalty component of PTA to                              fe     look at the number of missed known gold standard annotations rather than the total number of                                unrecognized annotations generated. Specifically:     eq. 6IT (T)  (A(T) (T))   (C(T)∖A(T))F A = fc ⋂C 2 − fe 2     where is defined as the normalized number of missed gold standard annotations and  fe                         |C(T)∖A(T)| |C(T)∖A(T)|max    fc   is the normalized number of matching answers as before . We call this variant ​apathetic                  |A(T) ⋂ C(T)||A(T) ⋂ C(T)|max             10  because its concern is only with the number of known correct and recognized versus correct but missed                                  annotations, regardless of the number of uncertain annotations.     IV.3 Delta PTA    With apathetic PTA, the reward component of matching annotations from the gold standard is paired                              with the penalty of missing annotations from the same. In this adaption, we compute as the absolute                           Δ         difference between the total number of annotations and the number of gold­standard annotations                          desired.      |A(T)| C(T)| |Δ = | − |     We want to be as close as possible to zero to minimize the number of unrecognized annotations.    Δ                                 Consequently, we define the reward component to encourage this result . Furthermore, we                     fc = 1 − ΔΔmax       want A(T) to have many matching C(T) answers thus we penalize missing gold standard annotations                              using the same  as in apathetic PTA.fe       eq. 7IT (T)   F D = 1[ − ΔΔmax] 2   −  [ |C(T)∖A(T)||C(T)∖A(T)|max] 2     IV.4 Optimistic PTA    The former pessimistic, apathetic, and delta versions of PTA reward known correct annotations                          (equations 3,4,6) while penalizing known errors (equation 3), missed answers (equations 6,7) and                          uncertainty (equation 4). With optimistic PTA we assume (optimistically) that unknown annotations are                          more often right than wrong. Consequently, we reward configurations that produce more annotations                          than those configurations that output less. To this end we redefine the reward component of the                                function  as the normalized ratio of the number of annotations produced :fc fc = | A(T) | | A(T) |max        eq. 8IT (T)   (E(A(T)))  F O   =  [ | A(T) || A(T) |max] 2 −  fe 2     Equation 8 is available for a two­class testing set and one­class negative label testing set (Table 2 #1,                                    #3). However, for the one­class positive labeled testing set, the term needs to be calculated without                    fe               assistance from negative testing samples. The term of equation 5 could be used in this circumstance            fe                       but works against the optimistic assumption that unknown annotations are most likely correct. In                            section IV.5, we compensate for this rigidness.    IV.5 Stochastic Optimistic PTA    Optimistic PTA in its current form (equation 8) cannot use a one­class positive label testing set without                                  ‘borrowing’ from its pessimistic counterpart (equation 5). While equation 4 assumes 100% of the  fe                           11  unknown (unlabeled) annotations are errors, stochastic optimistic PTA tries to estimate the expected                          number of errors within this unknown set.     Stochastic Optimistic PTA assumes if oracle­specific annotations are missing from the annotator’s                        output, then this is an indicator of poor performance of the annotation tool. However, it might also                                  happen that a tool produces more correct annotations than what the testing set mentions, with perhaps                                better semantic quality (Section III caveats iv,v and Table 3). On the other hand, one might assume that                                    although the gold standard might not list all the possible annotations (caveat iii), those annotations that                                are listed are likely to be the ​most important ​in the context of the given text fragment. In other words,                                        the annotations that form the gold standard are the results the annotator ​should ​produce. Consequently,                              Stochastic Optimistic PTA introduces an adjustment factor to equation 5 to soften the 100% error              φ)(                   assumption while still emphasizing the importance of the gold standard. First, precision (P) and recall                              (R) are computed:       and      eq. 9(T)P = | A(T) | |  A(T) ⋂ C(T)  | (T)R = | C(T) | |  A(T) ⋂ C(T)  |     Then, a ​likelihood of error​ is calculated using:          eq.10    The ratio in brackets [] is the familiar F​β​­score metric often used to balance precision and recall. When                                    β<1, unknown annotations are more likely considered errors, whereas β>1 puts the emphasis on                            matching answers to the gold standard. We use β=1 to equally weight precision and recall. To avoid a                                    division by zero error, we consider the case when precision is zero. In such a circumstance the                                  likelihood of error is 1. Next, the expected number of errors is determined for an estimate of the actual                                      number of errors of A(T).       eq. 11(T) (T)  A(T) ∖ C(T) |φ = Lφ × |     To illustrate, consider the example of Table 3. TagME produced nine annotations (|A(T)|) of which four                                matched the testing set ( ) out of a possible five known correct (|C(T)|). What remained         A(T) (T) || ⋂C                       was five annotations of unknown classification {​washington dc, home, teams, howard, bender​} (                        ). Substituting these values for equations 9, 10 and 11 gives a likelihood of error of A(T)∖C(T) ||                                 with an expected error count of . The true error count of the unknown set is(T) .429 Lφ = 0               (T) .15 φ = 2                     3 {​teams, howard, bender​}.    The last step is to compute   by normalizing   then substituting into equation 8.fe (T)φ      12     eq. 12IT (T)    F S   =  [ | A(T) || A(T) |max] 2 −[ φ(T)φ(T)max] 2        V. Experimentation    In this section, we evaluate the different forms of PTA against our one­class positive labeled gold                                standard under varying assumptions. Specifically, we examine how each form of PTA performs when                            the unrecognized annotations are believed to be mostly wrong (pessimistic), mostly right (optimistic),                          partially correct (stochastic/delta), or unimportant (apathetic).     We used a genetic algorithm (GA) to search the solution space as outlined in Figure 1 of Section IV.                                      The GA parameters were defined as follows: an initial population of 10 random configurations, with a                                maximum surviving population of 30 configurations per generation and a mutation rate of 0.05. We                              began with a testing set of 15 randomly selected text fragments from our gold standard set of                                  wiki-annot30 ​(Section III), with a sample replacement rate of 5 text fragments per generation​. The GA                                would terminate once the configurations within the surviving population stabilize. The top­ranked                        configurations for each PTA variant were then evaluated against a random set of 1000 text fragments                                from the gold standard and compared to the default out­of­the­box configuration of four popular                            semantic annotators: ​TagME, Wikipedia Miner, DBpedia Spotlight, and ​Yahoo Content Analysis​. We                        also considered numerous other semantic annotators that were ultimately not evaluated due to                          unavailability of a Web service or restrictive end­user license agreement. Lastly, we focused our tests                              around Wikipedia­centric annotation tools leaving other knowledge based annotation tools for future                        work. Table 4 lists all semantic annotators we considered for PTA evaluation.    Table 4​. List of tested and considered semantic annotators for PTA evaluation.  Annotator  Tested  URL and notes for considered but not tested annotators  TagME  YES  http://tagme.di.unipi.it/tagme_help.html#tagging  DBPedia Spotlight  YES  https://github.com/dbpedia­spotlight/dbpedia­spotlight/wiki/Web­service#Can didates  Wikipedia Miner  YES  http://wikipedia­miner.cms.waikato.ac.nz/services/?wikify  Yahoo Content Analysis  YES  https://developer.yahoo.com/contentanalysis/  AIDA  NO  Authors Web service is for demonstration purpose only. They explicitly  request not to use their service for research purposes.  http://www.mpi­inf.mpg.de/departments/databases­and­information­systems/r esearch/yago­naga/aida/webservice/  Denote  NO  Denote RESTful Web service is in beta and not available for use at the time  of writing this paper. http://inextweb.com/denote_demo  AlchemyAPI  NO  Restrictive Terms of Use. Alchemy can not be used for the purpose of  benchmarking and/or comparing with other annotators, especially for those  13  having a competing annotation service.  http://www.alchemyapi.com/api/register.html  Aylien  NO  Restrictive Terms of Service. Terms prohibits publication of any results  obtained using their annotator without permission. No response from Aylien  when contacted to obtain consent. http://aylien.com/text­api­tos  TextRazor  NO  Trial account allows for only 500 annotations per day which is insufficient for  training and testing, https://www.textrazor.com/plans  OntoText  NO  Formerly known as ​KIM - The Semantic Annotation Platform​. No accessible  Web service without first consultation with the commercial company.  http://www.ontotext.com/products/ontotext­semantic­platform/      V.1 PTA with TagME  The TagME annotation service offers fast execution and high accuracy particularly with short text                            fragments (​Chiong and Beng, 2007; ​Cornolti, Ferragina and Ciaramita, 2013​). Table 5 provides                          summary statistics and the recommended configuration for TagME using each of the five variants of                              PTA including TagME’s default configuration for side­by­side comparison. Summary statistics include                      the average of the following measures on the 1000 gold standard text fragments: number of annotations                                A(T), number of annotations matching to the gold standard A(T)∩C(T), number of unmatched gold                            standard annotations C(T)\A(T), and count of unrecognized annotations A(T)\C(T). Derived from these                        measures are precision, recall, and F​1​­score, also included in Table 5. These measures provide                            competing dimensions that impact each variant of PTA differently. For instance, TagME’s default                          configuration produces the highest number of annotations with an average of 9.56, but also produces                              the highest number of unrecognized annotations, 6.22 on average. Comparatively, the pessimistic PTA                          solution produces the least number of unrecognized annotations (0.228 on average), but does so by                              allowing only a few annotations (1.63 on average). Table 5 also includes an EMPTY% metric defined                                as the percentage of text fragments that returned no annotations and/or no gold standard annotations                              A(T)∩C(T)=∅. Pessimistic appears to do well in the precision dimension, but its scarce annotator                            output produces a large number of empty annotations at 33.8%.      Figure 2 provides graphs of the recall, precision, and F​1 scores on all individual 1000 text fragments,                                  sorted from lowest to highest score, instead of the average­only values shown in Table 5. Mean values                                  of recall, precision, and F​1 equate to the scaled [0,1] area­under­the­curve (AUC) for Figure 2.                              Precision and recall (equation 9) were calculated as the ratio of matched gold standard annotations to                                the total number of annotations produced (precision), and to the total number of gold standard                              annotations (recall).    The recall graph shows that TagME’s default configuration and apathetic PTA performed identically as                            the best configuration for matching the most number of gold standard annotations followed closely by                              stochastic, delta, and optimistic PTA. Pessimistic PTA performed the worst with respect to matching                            gold standard annotations due to its conservative nature of only providing 1.63 annotations on average.                              In regards to precision, the solutions offered by optimistic, delta, and stochastic PTA outperform the                              14  default and apathetic configurations. Finally, when both precision and recall are considered together                          through F​1 measure we see that stochastic, delta, and optimistic configurations perform the best with                              average F​1​­scores of 0.683, 0.677, and 0.672, respectively, compared with apathetic and default scores                            of 0.521 and 0.493, respectively.     Table 5​. Default and recommended values for TagMe’s tunable parameters (first 3 rows), followed by summary                                statistics (mean values) for comparison metrics computed on 1000 text fragments gold­standard using tunable                            parameters’ default values (1st column) and values recommended by different forms of PTA.      Default  Pessimistic  : ↑ A(T)∩C(T) fc   : ↓ ​A(T)\C(T) fe   Apathetic  : ↑A(T)∩C(T) fc   : ↓​C(T)\A(T) fe   Optimistic  : ↑ A(T) fc   : ↓​A(T)\C(T) fe   Stochastic  : ↑ A(T) fc   : ↓ E[​A(T)\C(T)] fe   Delta  : ↑ ±A(T)­C(T) fc   : ↓​C(T)\A(T) fe   *epsilon  0.30  0.494  0.282  0.156  0.427  0.357  *long_text  0  6  0  10  10  7  *rho  0.0  0.551  0.0221  0.2662  0.1613  0.2429  C(T)  4.031  4.031  4.031  4.031  4.031  4.031  A(T)  9.567  1.631  8.778  3.721  4.915  4.029  A(T)∩C(T)  3.347  1.403  3.346  2.757  3.128  2.861  C(T)\A(T)  0.684  2.628  0.685  1.274  0.903  1.17  A(T)\C(T)   6.22  0.228  5.432  0.964  1.787  1.168  PRECISION  0.375  0.607  0.407  0.728  0.659  0.710  RECALL  0.827  0.324  0.827  0.673  0.768  0.699  F­SCORE  0.493  0.401  0.521  0.672  0.683  0.677  EMPTY%  1.6%  33.8%  1.6%  6.9%  2.8%  5.6%      Figure 2​. Graphs of recall (top left), precision (bottom left), and F​1​­score (right) for each variant of PTA including                                      default configuration used by TagME on 1000 text fragments.    15  The results indicate that if the objective is to match as many as possible gold standard annotations,                                  without concern for unrecognized annotations, TagME’s default or PTA’s apathetic solution would be                          the best option, as they most closely achieve C(T) ⊆ A(T). However, if the goal is to avoid                                    uncertainty, i.e., A(T) ⊆ C(T), then either optimistic, delta, or stochastic PTA would be the best                                candidates. Finally, for the behaviour that closely resembles the gold standard output, i.e., C(T)=A(T),                            stochastic PTA would be a good choice.     V.2 PTA with WikipediaMiner  WikipediaMiner is a toolkit that provides semantic services, including semantic annotation through a                          downloadable software library or Web service. The default configuration along with the solutions                          recommended by the five variants of PTA are given in Table 6. Pessimistic PTA discovered a solution                                  very similar to WikipediaMiner’s default configuration. Apathetic PTA produced the highest number of                          annotations per text fragment A(T) with an average of 14.23 annotations, but at the expense of the                                  highest number of unknowns A(T)\C(T) (11.2 on average). This resulted in a high recall of 0.75, but                                  low precision of 0.35.     Table 6​. Default and recommended values for WikipediaMiner’s tunable parameters (first 3 rows), followed by                              summary statistics (mean values) for comparison metrics computed on 1000 text fragments gold­standard using tunable                              parameters’ default values (1st column) and values recommended by different forms of PTA.      Default  Pessimistic  : ↑ A(T)∩C(T) fc   : ↓ ​A(T)\C(T) fe   Apathetic  : ↑A(T)∩C(T) fc   : ↓​C(T)\A(T) fe   Optimistic  : ↑ A(T) fc   : ↓​A(T)\C(T) fe   Stochastic  : ↑ A(T) fc   : ↓ E[​A(T)\C(T)] fe   Delta  : ↑ ±A(T)­C(T) fc   : ↓​C(T)\A(T) fe   *minProbability  0.50  0.53  0.023  0.33  0.23  0.46  *disambiguation  Policy  strict  strict  loose  strict  loose  strict  *weight  0.0  0.205  0.0427  0.427  0.663  0.199  C(T)  4.031  4.031  4.031  4.031  4.031  4.031  A(T)  3.862  3.862  14.229  4.435  6.027  4.185  A(T)∩C(T)  2.099  2.099  3.029  2.303  2.718  2.217  C(T)\A(T)  1.932  1.932  1.002  1.728  1.313  1.814  A(T)\C(T)   1.763  1.763  11.2  2.132  3.309  1.968  PRECISION  0.525  0.525  0.245  0.519  0.468  0.519  RECALL  0.511  0.511  0.755  0.566  0.674  0.543  F­SCORE  0.485  0.485  0.350  0.508  0.525  0.498  EMPTY%  15.3%  16.8%  2.9%  11.6%  5.1%  13.4%     The recall graph of Figure 3 demonstrates that apathetic and stochastic PTA perform the best in this                                  metric, followed by optimistic, delta, default, and pessimistic performing relatively the same. The                          results were opposite for the precision, with pessimistic, default, delta, and optimistic performing well                            followed closely by stochastic then lastly apathetic due to its high unknown count. Stochastic PTA                              gave the best overall solution with comparable precision to the default configuration, but with                            significantly better recall. In addition, stochastic PTA generated the second­least number of empty                          annotations with 5.1% compared to the second­worst score of 15.3% from the default configuration.    16        Figure 3​. Graphs of recall (top left), precision (bottom left), and F​1​­score (right) for each variant of PTA including                                      default configuration used by WikipediaMiner on 1000 text fragments.    V.3 PTA with DBpedia Spotlight  We tested PTA on the DBPedia Spotlight ​candidates ​Web service. The candidates service is an                              annotation service that returns a ranked list of annotations per mention. This ranking includes output                              statistics entitled ​contextual score, percentage of second rank, and ​final score ​with the intent that the                                application using the annotator service will prune the ranked list based on the chosen thresholds for                                these statistics. Consequently, the goal of PTA is to discover what threshold values to use. Therefore, it                                  was not surprising that the default configuration gave the best results for recall (0.52, Table 7) but poor                                    precision (0.19), and the highest number of annotations per text fragment (14.8). Pessimistic PTA had                              the most difficulty with the lowest F​1​­score (0.02) caused by a high number of unrecognized                              annotations and low number of matching gold standard answers. Stochastic PTA would have provided                            the best precision if not for the many empty solutions it gave (Figure 4 and Table 7). Overall, optimistic                                      PTA gave the second best recall with the best precision and thus is the recommended solution with the                                    top F​1​­score of 0.38.     V.4 PTA with Yahoo Content Analysis (YCA)  The Yahoo Content Analysis API provides named entity linking to Wikipedia entities through its Web                              service. PTA mean statistics show apathetic PTA leads in seven of the eight metrics (Table 8).                                However, the difference is statistically inconsequential and did not translate to a significant                          improvement over the default configuration. The AUC graphs of Figure 5 reveal stochastic, apathetic,                            17  and delta forms of PTA perform equally well as YCA’s default configuration using different local                              maxima solutions. Optimistic PTA was a close second while pessimistic struggled.     Table 7​. Default and recommended values for Spotlight’s tunable parameters (first 5 rows), followed by summary                                statistics (mean values) for comparison metrics computed on 1000 text fragments gold­standard using tunable                            parameters’ default values (1st column) and values recommended by different forms of PTA.      Default  Pessimistic  : ↑ A(T)∩C(T) fc   : ↓ ​A(T)\C(T) fe   Apathetic  : ↑A(T)∩C(T) fc   : ↓​C(T)\A(T) fe   Optimistic  : ↑ A(T) fc   : ↓​A(T)\C(T) fe   Stochastic  : ↑ A(T) fc   : ↓ E[​A(T)\C(T)] fe   Delta  : ↑ ±A(T)­C(T) fc   : ↓​C(T)\A(T) fe   *confidence  0.20  0.14  0.029  0.29  0.62  0.020  *support  20  48  5  1  1  355  *contextualScore  0.0  0.132  0.0508  0.0  0.0  0.115  *percentageOf  SecondRank  0.0  0.0290  0.596  0.0  0.0  0.459  *finalScore  0.0  0.491  0.128  0.0  0.228  0.004  C(T)  4.031  4.031  4.031  4.031  4.031  4.031  A(T)  14.89  0.072  4.227  5.042  1.496  3.153  A(T)∩C(T)  2.085  0.055  1.477  1.8  0.878  1.022  C(T)\A(T)  1.946  3.976  2.554  2.231  3.153  3.009  A(T)\C(T)   12.8  0.017  2.75  3.242  0.618  2.131  PRECISION  0.190  0.0373  0.403  0.406  0.403  0.334  RECALL  0.522  0.0157  0.372  0.451  0.218  0.263  F­SCORE  0.251  0.0211  0.345  0.384  0.259  0.262  EMPTY%  12.4%  96%  22.6%  17.6%  48.5%  36%            Figure 4​. Graphs of recall (top left), precision (bottom left), and F​1​­score (right) for each variant of PTA including                                      default configuration used by DBpedia Spotlight on 1000 text fragments.  18  Table 8: Default and recommended values for YCA’s tunable parameters (first 2 rows), followed by summary statistics                                  (mean values) for comparison metrics computed on 1000 text fragments gold­standard using tunable parameters’                            default values (1st column) and values recommended by different forms of PTA.      Default  Pessimistic  : ↑ A(T)∩C(T) fc   : ↓ ​A(T)\C(T) fe   Apathetic  : ↑A(T)∩C(T) fc   : ↓​C(T)\A(T) fe   Optimistic  : ↑ A(T) fc   : ↓​A(T)\C(T) fe   Stochastic  : ↑ A(T) fc   : ↓ E[​A(T)\C(T)] fe   Delta  : ↑ ±A(T)­C(T) fc   : ↓​C(T)\A(T) fe   *max  100  9  18  6  99  98  *score  0.0  0.917  0.455  0.594  0.192  0.007  C(T)  4.031  4.031  4.031  4.031  4.031  4.031  A(T)  1.439  0.243  1.457  1.233  1.444  1.443  A(T)∩C(T)  1.071  0.199  1.082  0.938  1.075  1.070  C(T)\A(T)  2.960  3.832  2.949  3.093  2.956  2.961  A(T)\C(T)   0.368  0.044  0.375  0.295  0.369  0.373  PRECISION  0.520  0.160  0.529  0.511  0.525  0.521  RECALL  0.268  0.0491  0.271  0.238  0.268  0.266  F­SCORE  0.331  0.0719  0.335  0.306  0.332  0.329  EMPTY%  39.9%  83.6%  38.7%  42.4%  39.2%  39.6%          Figure 5: Graphs of recall (top left), precision (bottom left), and F​1​­score (right) for each variant of PTA including                                      default configuration used by YCA on 1000 text fragments.    VI. Result Summary     Table 9 provides a summary of the experimental results. For each examined annotator, the table                              presents the forms of PTA that evaluated as the best(+) and the worst(­) performing for mean measures                                  of: most annotations, most annotations matching the gold standard, least unrecognized annotations,                        least empty answers and best precision/recall. Also included is the overall (recommended) PTA variant                            based on AUC graphs. Noteworthy is that the best PTA variant per individual dimensions does not                                19  necessarily indicate best performing overall solution. As an example, consider WikipediaMiner.                      Apathetic PTA scored highest for most annotations, most matched, least empty, and best recall;                            however, the recommended solution is stochastic which was not the top (nor bottom) of any single                                measure. The table also shows that no single variant of PTA is best as this is a function of annotator                                        behavior, testing set, and user assumptions (unknown annotations are mostly right, mostly wrong, or                            partially correct). Nonetheless, it would appear pessimistic PTA is excessively conservative and is                          often surpassed by its probabilistic counterpart: stochastic. Consequently, the pessimistic PTA variant                        should be avoided. From Table 9, the following summary observations can  be made:    ● Apathetic PTA emerged as a good choice for most annotations, most matched and best recall.   ● Optimistic PTA performed well when precision is the primary concern.   ● Stochastic PTA was the prevailing general­purpose strategy with an overall best solution for                          three of the four annotators tested.       Table 9: Summary table indicating the best(+) and the worst(­) solutions for individual measures by averages plus                                  overall recommended solution based on AUC graphs​.     most  Annotations  A(T)  most  matched   A(T)∩C(T)  least   unknown  A(T)\C(T)  least empty  answers  EMPTY %  best   recall  best  precision  Overall  best AUC­F​1  TagME  +default ­pessimistic  +default  +apathetic  ­pessimistic  +pessimistic  ­default   +default  +apathetic  ­pessimistic  +default  +apathetic  ­pessimistic  +optimistic  ­default  +stochastic  ­pessimistic  WikipediaMiner  +apathetic  ­default  ­pessimistic  +apathetic  ­default  ­pessimistic  +default  +pessimistic  ­apathetic  +apathetic  ­pessimistic  +apathetic  ­default  ­pessimistic  +default  +pessimistic  ­apathetic  +stochastic  ­apathetic  DBpedia  Spotlight  +default  ­pessimistic  +default  ­pessimistic  +pessimistic  ­default  +default  ­pessimistic  +default  ­pessimistic  +optimistic  ­pessimistic  +optimistic  ­pessimistic  Yahoo (YCA)  +apathetic ­pessimistic  +apathetic  ­pessimistic  +pessimistic  ­apathetic  +apathetic  ­pessimistic  +apathetic  ­pessimistic  +apathetic  ­pessimistic  +default  +apathetic  +stochastic  +delta  ­pessimistic    Our experimental results demonstrate that PTA is effective. In all annotators tested, PTA either: 1)                              suggested an improved configuration, or 2) validated the default configuration as a local maximum                            solution. PTA successfully found alternative configurations that better fitted the testing set than the                            default parameters of TagME, WikipediaMiner, and DBpedia Spotlight.     VII. Related Literature    Semantic annotation tools offer the possibility of deeper analysis of textual content by disambiguating                            terms and phrases present in the text and linking them to appropriate concepts from a knowledge base,                                  hence enabling more efficient and accurate classification, organization, search and retrieval of textual                          content. The research community has already developed a critical mass of both research prototypes and                              usable software products in this area. Automated semantic annotation tools examined in this paper:                            TagMe (​Ferragina & Scaiella, 2012​), Denote (​Cuzzola et al., 2013​), DBPedia Spotlight (​Mendes,                          20  Jakob, García­Silva & Bizer, 2011​), Wikipedia Miner (​Milne & Witten, 2013​), among others, provide                            means to automatically semantically process textual content and identify relevant semantic concepts.     Recently, these tools have been increasingly referred to as entity linking tools since they link entity                                mentions in the text with the corresponding entry/entries in a knowledge base. Shen, ​Wang & Han                                (2015) have provided a very comprehensive and detailed survey of the state­of­the­art automated entity                            linking (i.e., semantic annotation) tools. They have identified three modules that such tools consist of,                              namely candidate entity generation, candidate entity ranking and unlinkable menton prediction                      modules, and for each module presented and analyzed different methods and techniques that were                            proposed in the literature and applied in existing annotation tools. They also reported that the examined                                tools “differ along multiple dimensions and are evaluated over different data sets”.    Large majority of today’s semantic annotators rely on a general purpose knowledge base (KB) such as                                Wikipedia or more structured and semantically rich KBs like DBpedia, YAGO, and Wikidata.                          However, there are also domain­specific semantic annotators, many of them in the biomedical domain,                            e.g., MetaMap (Aronson & Lang, 2010) and NCBO annotator (Whetzel et al., 2013), that rely on                                domain­specific KBs such as Unified Medical Language System (UMLS) (Bodenreider, 2004),                      DrugBank (Law et al., 2014) or medical ontologies available at NCBO BioPortal (Whetzel et al., 2011).                                In order to assure good domain coverage, both in terms of breadth and depth of the entities covered,                                    some annotation tools rely on more than one KB. For instance, the semantic annotation method                              proposed by Berlanga, ​Nebot & Pérez (2015) relies on the use of several KBs of arbitrary size and                                    domain specificity, and is also independent of any specific characteristic of a KB (e.g., disambiguation                              pages, internal links and other Wikipedia specific features). Some semantic annotators, such as TagMe                            and the one developed by WalmartLabs (Gattani et al., 2013), are specifically designed and developed                              for semantic annotation of social media content that is typically short, fragmented, poorly spelled and                              grammatically incorrect​. ​Since such an annotator requires a global and almost real­time KB, Gattani et                              al. (2013) expanded Wikipedia with data from ​various structured sources (e.g., Adam (health),                          MusicBrainz (albums), City DB, and Yahoo Stocks), as well as new interesting events extracted from                              Twitter data stream.    The algorithms developed as a part of semantic annotation systems, e.g. spot identification and                            disambiguation, often rely on user­dependent fine­tuning of input parameters. These parameters allow                        the user to customize the algorithms for the specific topic or textual content type (​Jovanovic et al.,                                  2014​). While annotation tools do provide a suggested default value on these parameters, they are not by                                  any means optimal for all scenarios. To our knowledge, no existing related work or implemented                              software has attempted to address the problem of automated parameter fine­tuning for semantic                          annotation tools; therefore, our proposed PTA framework serves as a first foundational step in this                              direction.    Having said that, it is important to point out that while researchers have not addressed the problem of                                    parameter fine­tuning for semantic annotators, there have been a few fruitful work on the systematic                              evaluation of semantic annotation systems. For instance, Cornolti, ​Ferragina & Ciaramita ​(2013) have                          developed a framework consisting of metrics and standard datasets for measuring the efficiency and                            21  effectiveness of existing semantic annotation tools. The major contribution of their work is its novel                              approach to the systematic classification of different tasks of a semantic annotator and the development                              of suitable metrics for each of these specific tasks. The work by Steinmetz, ​Knuth & Sack (2013) is                                    also focused on the evaluation of semantic annotation tools; however, their attention is centered more                              on the statistical analysis of different benchmark and dictionary datasets that can be used in the                                evaluation process. Heuss, ​Humm, Henninger & Rippl (2014) compared the performance of several                          state­of­the­art semantic annotation tools on domain specific texts (namely texts about museum                        collections). The study found that, on average, each tool achieved roughly just a third of its F1 score on                                      texts covering general/common topics. The results also showed very high standard deviations for all                            performance measures (recall, precision and F1), indicating not only lower performance than in a                            common case, but wider distribution of the results. This is consistent with the findings of Shen, ​Wang,                                  & Han (2015) who reported that the tools examined in their survey tend to perform very differently for                                    different data sets and domains. It is worth pointing out that existing comparative studies of semantic                                annotation tools have all relied on the suggested default parameter settings for the compared systems;                              therefore, they do not necessarily reflect the best case performance of the annotator tool on the objects                                  of the experiment. It could very well be the case that if the optimal parameter values were chosen (as                                      opposed to the default values) that the obtained results could be significantly different.    It should be mentioned that besides automated semantic annotation tools, there are also semi­automated                            semantic annotators that allow for user’s intervention during the annotation process. This intervention                          often takes the form of choosing the best option from a list of candidate annotations, or removing some                                    of the proposed annotations that the user considers incorrect or irrelevant. While considerable research                            and development efforts were put into the design and development of semi­automated annotation tools                            (​Uren et al., 2005​), (Oliveira & Rocha, 2013), their reliance on human active participation impacts their                                efficiency, and thus they have been largely superseded by fully automated tools. Still, there are some                                usage scenarios, e.g., scholarly reading, where, due to their mixed­initiative annotation approach,                        semi­automated annotators are preferred. For instance, based on a study of scholarly annotation                          practices, Müller­Birn, Klüwer, Breitenfeld, Schlegel, & Benedix (2015) have designed and developed                        Neonion, a lightweight annotation tool for creating, sharing and reusing annotation data. Neonion users                            can accept, reject or modify annotations recommended by the tool; annotations take the form of                              references to appropriate Wikidata (Vrandečić & Krötzsch, 2014) entities. This feedback that users                          provide is leveraged for improving subsequent recommendations. Annotation of medical texts is                        another domain where human involvement is often needed to assure the accuracy of automatically                            produced annotations. For example, RapTAT is a semi­automated semantic annotation tool based on an                            interactive and iterative machine learning approach, and aimed at assisting end users with annotation of                              various kinds of medical texts (Gobbel et al., 2014). In each iteration, the tool annotates potentially                                relevant phrases within a document, presents the annotations to a reviewer for correction, and then uses                                the obtained feedback (i.e., corrected annotations) to re­train its machine learning model before                          annotating subsequent document.    Although work on finding optimal parameter values for semantic annotation tools is novel, it is                              important to point out that the use of evolutionary algorithms as Genetic Algorithms for optimal                              parameter estimation in control systems has been a commonplace (​Chang, 2006​). For instance, Seng et                              22  al (1999) used Genetic Algorithms to simultaneously tune the parameters of a self­tuning fuzzy logic                              control system. Similarly, Yao & ​Sethares (1994) used Genetic Algorithms for optimizing the structure                            and parameters of feedforward and recurrent neural networks and shown to be able to reduce estimation                                error in probability to zero. Genetic Algorithms have also been widely used for parameter tuning in                                Proportional­Integral­Derivative (PID) controllers (​Panda, 2011​), power system optimization (​Kothari,                  2012​) and HVAC systems (​Kusiak, Tang & Xu​, 2011). They were also used to deal with challenges of                                    image annotation (​Bahrami & Abadeh, 2014​). In the Semantic Web domain, Genetic Algorithm­based                          approaches can be observed in a variety of applications (​Chen, Wu & Cudré­Mauroux​, 2012) such as                                finding optimal ontology alignments between multiple ontologies (​Martinez­Gil, Alba &                    Aldana­Montes​, 2008); identification and alignment of datasets of the Linked Open Data cloud                          (Gunaratna, Lalithsena & Sheth, 2014); RDF query answering (​Oren, Guéret & Schlobach​, 2008); and                            semantic Web service discovery (Sangers, Frasincar, Hogenboom & Chepegin, 2013) and composition                        (​Fanjiang & Syu, 2014​), among others, and have shown to be effective for such optimization problems.     VIII. Limitations and Future Work    We have proposed a method, called Parameter Tuning Architecture (PTA), for finding suitable values                            for tunable parameters of semantic annotators (Section IV) and demonstrated this method on four                            semantic annotators (Section V). Note that the annotators themselves were not trained by PTA; instead,                              alternative parameter values were proposed that better suited the targeted evaluation/testing set                        (​wiki-annot30​). PTA could be extended for semantic annotator training, namely, as a tool to decide on                                the fallback parameter defaults when no alternative values are supplied. This would simply require the                              use of a training set rather than a testing set with PTA. However, since we were unable to acquire the                                        exact training set used for each of the four annotators in our experiments, we were not able to compare                                      how general purpose default values suggested by PTA would perform relative to the defaults currently                              chosen by the annotators’ designers. Even if the training sets were available, there may be ​internal                                tunable parameters that are only accessible to the designers, and not exposed through the annotator’s                              public interface. Moreover, since each of the tested annotators were trained with different gold                            standards on (most­likely) different versions of Wikipedia, subsequent work should include                      experimentation with other datasets such as Microsoft’s “​Entity Recognition and Disambiguation                      Challenge​”  and/or ​ClueWeb  for further validation of our PTA framework.  6 7   Future work may also include validation with other knowledge based annotation systems                        (non­Wikipedia) that rely on life sciences, biomedical, or other ontologies. Additionally, we propose an                            extension to our fitness function (equation 1) with the introduction of weights :δ , )( c δe     IT(T) rg max δ  where δ  F = a νεV c f (T)[ c ] 2 − δe f (T)[ e ] 2 c + δe = 1     This would allow for an emphasis toward either finding more correct annotations or less annotations                       fc       in error . The weights themselves could be part of the tunable parameter vector in which    fe         δ , )( c δe                  v     6 http://web­ngram.research.microsoft.com/ERD2014/  7 http://lemurproject.org/clueweb09/FACC1/  23  PTA would be tasked to not only find suitable parameter values for the annotator in question but also                                    suggesting parameter values for itself. Finally, our future work will investigate other evolutionary                          algorithms such as swarm optimization and compare its performance with the currently used genetic                            algorithm (Figure 1).    IX. Conclusion    In this paper we present Parameter Tuning Architecture (PTA), a general method to determine                            “best­fit” values of configuration parameters for semantic annotators. We explain the caveats of                          supervised testing specific to semantic annotators and devised PTA variants to tackle the uncertainty of                              unlabeled annotations. We tested these variants on four well­known semantic annotators and provided a                            method for selecting the best solution using a genetic algorithm and area­under­the­curve metric.                          Experimental results indicate that PTA is capable of suggesting configurable parameters that improve                          upon specific individual areas of most annotations, most matched gold standard answers, and least                            uncertainty. Finally, our tests demonstrate that PTA can consistently find a configuration that provides                            an overall best solution, i.e., solution with the best precision versus recall trade­off for many semantic                                annotators. We balance our findings by acknowledging the limitations of our work and propose five                              future research directions for further study (Section VIII).    Since the PTA fitness function, the core component of the proposed method, does not rely on any                                  annotator­specific feature, our PTA method is applicable to any semantic annotator, and can be used to                                enhance the annotator’s performance on any specific annotation task. Besides being directly beneficial                          for semantic annotation tools, the proposed method might also be indirectly useful to ​any intelligent                              system that relies on semantic­rich annotation of textual content, such as text search and retrieval                              systems, various content­based recommender systems, systems that rely on semantics of textual content                          to support personal or business decision making and the like.      References    A. R. Aronson and F.­M. Lang. (2010). ​An overview of MetaMap: historical perspective and recent advances​, 17(3),                                  pp.229­236, JAMIA. http://metamap.nlm.nih.gov/    S. Atdağ, and V. Labatut. (2013). ​A Comparison of Named Entity Recognition Tools Applied to Biographical Texts​,                                  (pp. 228­233). 2nd International Conference on Systems and Computer Science.    S. Bahrami and M.S. Abadeh, (2014). ​Automatic image annotation using an evolutionary algorithm (IAGA)​, pp.320 ­                                325, 7th International Symposium on Telecommunications (IST 2014).    R. Berlanga, V. Nebot, M. Pérez. (2015). ​Tailored semantic annotation for semantic search​, 30, pp. 69­81, Web                                  Semantics: Science, Services and Agents on the World Wide Web.    O. Bodenreider. (2004). ​The Unified Medical Language System (UMLS): integrating biomedical terminology​, 32                          (Database­Issue), pp. 267­270, Nucleic Acids Research. http://www.nlm.nih.gov/research/umls    24  W.D. Chang, (2006). ​An improved real-coded genetic algorithm for parameters estimation of nonlinear systems, 20(1),                              pp. 236­246. Mechanical Systems and Signal Processing.    H. Chen, Z. Wu & P. Cudré­Mauroux, (2012). ​Semantic Web Meets Computational Intelligence: State of the Art and                                    Perspectives,​ 7(2), pp. 67­74, IEEE Computational Intelligence Magazine.    R. Chiong, and O.K. Beng, (2007). ​A Comparison between Genetic Algorithms and Evolutionary Programming based                              on Cutting Stock Problem.​ vol. 14, pp. 72­77.  Engineering Letters.    M. Cornolti, P. Ferragina, M. Ciaramita, (2013), ​A Framework for Benchmarking Entity-Annotation Systems, pp.                            249­260. 22nd International World Wide Web Conference​.    J. Cuzzola, D. Gasevic, E. Bagheri, Z. Jeremic, J. Jovanovic, R. Bashash, (2013). ​Semantic Tagging with Linked Open                                    Data,​ vol. 1054, pp 52­53. 4th Canadian Semantic Web Symposium (CSWS 2013).    X. Dong, E. Gabrilovich, G. Heitz, W. Horn, N. Lao, et al. (2012), ​Knowledge vault: a web-scale approach to                                      probabilistic knowledge fusion, pp. 601­610, 20th ACM SIGKDD international conference on Knowledge discovery                          and data mining (KDD '14).    Y. Fanjiang and Y. Syu, (2014). ​Semantic-based automatic service composition with functional and non-functional                            requirements in design time: A genetic algorithm approach​, 56(3), pp. 352­373, Information and Software Technology.    P. Ferragina, U. Scaiella, (2012). ​Fast and Accurate Annotation of Short Texts with Wikipedia Pages. 29(1), pp. 70­75,                                    IEEE Software.    A. Gattani, D. S. Lamba, N. Garera, M. Tiwari, X. Chai, et al. (2013). ​Entity extraction, linking, classification, and                                      tagging for social media: a wikipedia-based approach​,  pp. 1126­1137, Proc. VLDB Endow. 6, 11 (August 2013).    G. Gobbel, J. Garvin, R. Reeves, R.M. Cronin, J. Heavirland et al. (2014). ​Assisted annotation of medical free text                                      using RapTAT​, 21(5), pp.833­841, J Am Med Inform Assoc.    J. Grefenstette, (1992). ​Genetic Algorithms for Changing Environments, pp. 139­146, ​Parallel Problem Solving from                            Nature 2.    K. Gunaratna, S. Lalithsena, and A.P. Sheth, (2014). ​Alignment and Dataset Identification of Linked Data in Semantic                                  Web​, 4 (2), pp. 139­151, Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery.    B. Hachey, W. Radford, J. Nothman, M. Honnibal, and J.R. Curran, (2013). ​Evaluating Entity Linking with Wikipedia​,                                  194, pp. 130­150, Journal of Artificial Intelligence.    T. Heuss, B. Humm, C. Henninger, and T. Rippl. (2014). ​A comparison of NER tools w.r.t. a domain-specific                                    vocabulary​, pp.100­107, In Proceedings of the 10th International Conference on Semantic Systems (SEM '14), H. Sack,                                A. Filipowska, J. Lehmann, and S. Hellmann (Eds.). ACM, New York, NY, USA.    E. Hovy, R. Navigli, S.P. Ponzetto, (2013). Collaboratively built semi-structured content and Artificial Intelligence:                            The story so far​, 194, pp. 2­27.  Journal of Artificial Intelligence.    25  J. Jovanovic, E. Bagheri, J. Cuzzola, D. Gasevic, Z. Jeremic, R. Bashash, (2014). ​Automated Semantic Annotation of                                  Textual Content​,  16(6), pp. 38­46. IEEE IT Professional.    D.P. Kothari, (2012). ​Power system optimization, pp. 18­21, 2nd National Conference on Computational Intelligence                            and Signal Processing (CISP).    A. Kusiak, F. Tang, & G. Xu, (2011). ​Multi-objective optimization of HVAC system with an evolutionary computation                                  algorithm,​ ​36(​5), pp. 2440­2449, Energy.    V. Law, C. Knox, Y. Djoumbou, T. Jewison, A.C. Guo, et al. (2014). ​DrugBank 4.0: shedding new light on drug                                        metabolism​, 42(1), pp.D1091­7, Nucleic Acids Res.    B. Liu, (2012). ​Sentiment Analysis and Opinion Mining (Synthesis Lectures on Human Language Technologies)​.                            Morgan & Claypool Publishers.    J. Martinez­Gil, E. Alba, & J.F. Aldana­Montes, (2008). ​Optimizing ontology alignments by using genetic algorithms,                              Workshop on Nature based Reasoning for the Semantic Web.    D. Maynard, (2008). ​Benchmarking Textual Annotation Tools for the Semantic Web. 6th International Conference on                              Language Resources and Evaluation.    P.N. Mendes, M. Jakob, A. García­Silva, and C. Bizer, (2011). ​DBpedia spotlight: shedding light on the web of                                    documents.,​ pp. 1­8, 7th International Conference on Semantic Systems. ACM.    D. Milne, & I.H. Witten, (2013). ​An open-source toolkit for mining Wikipedia. ​vol. 194, pp. 222­239, Journal of                                    Artificial Intelligence​.    D. Moriarty, A. Schultz​, J. Grefenstette​, (1999), ​Evolutionary Algorithms for Reinforcement Learning, 11, pp. 241­276,                              Journal Artificial Intelligence Research (JAIR).    C. Müller­Birn, T. Klüwer, A. Breitenfeld, A. Schlegel, and L. Benedix. (2015). ​Neonion: Combining Human and                                Machine Intelligence​, pp. 223­226, Proceedings of the 18th ACM Conference Companion on Computer Supported                            Cooperative Work & Social Computing (CSCW'15 Companion).     P. Oliveira, and J. Rocha. (2013). ​Semantic annotation tools survey​, pp.301­307, IEEE Symposium on Computational                              Intelligence and Data Mining (CIDM).     E. Oren, C. Guéret, & S. Schlobach, (2008). ​Anytime query answering in RDF through evolutionary algorithms, pp.                                  98­113, 7th International Semantic Web Conference (ISWC 08).    S. Panda, (2011). ​“Multi-objective PID controller tuning for a FACTS-based damping stabilizer using Non-dominated                            Sorting Genetic Algorithm-II,​, 33(7), pp. 1296­1308, International Journal of Electrical Power & Energy Systems.    L. Ratinov and D. Roth, (2009). ​Design challenges and misconceptions in named entity recognition​, (pp. 147­155).                                Thirteenth Conference on Computational Natural Language Learning (CoNLL '09).    J. Sangers, F. Frasincar, F. Hogenboom, and V. Chepegin, (2013). ​Semantic Web service discovery using natural                                language processing techniques​, 40(11), pp. 4660­4671, Expert Systems with Applications.  26  http://www.informatik.uni-trier.de/~ley/pers/hd/s/Schultz:Alan_C=.html http://www.informatik.uni-trier.de/~ley/pers/hd/g/Grefenstette:John_J=.html   T.L. Seng, M. Bin Khalid, & R. Yusof, (1999). ​Tuning of a neuro-fuzzy controller by genetic algorithm, 29( 2), pp.                                        226­236, Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Trans.,.    W. Shen, J. Wang, and J. Han. (2015). ​Entity Linking with a Knowledge Base: Issues, Techniques, and Solutions, ​27(2),                                      pp.443­460, IEEE Transactions on Knowledge & Data Engineering.    N. Steinmetz, M. Knuth & H. Sack, (2013). ​Statistical Analyses of Named Entity Disambiguation Benchmarks, pp.                                21­25, 1st International Workshop on NLP and DBpedia.    H. Szczerbicka, M. Becker, M. Syrjakow​, (1998). ​Genetic Algorithms: a Tool for Modelling, Simulation, and                              Optimization of Complex Systems,​ 29(7), pp. 639­659, Cybernetics and Systems​.    V. Uren, P. Cimiano, J. Iria, S. Handschuh, M. Vargas­Vera, E. Motta, S. Ciravegna, (2005). ​Semantic annotation for                                    knowledge management: Requirements and a survey of the state of the art. ​ 4(1). pp. 14­28. Journal of Web Semantics.    D. Vrandečić, M. Krötzsch. (2014). ​Wikidata: A Free Collaborative Knowledge Base​. 57(10). pp. 78–85.                            Communications of the ACM.    J. Weston, A. Bordes, O. Yakhnenko, N. Usunier, (2013). ​Connecting Language and Knowledge Bases with                              Embedding Models for Relation Extraction, (pp.1366­1371), Conference on Empirical Methods in Natural Language                          Processing.    P.L. Whetzel, N.F. Noy, N.H. Shah, P.R. Alexander, C. Nyulas, et al. (2011). ​BioPortal: enhanced functionality via                                  new Web services from the National Center for Biomedical Ontology to access and use ontologies in software                                  applications​, 39, pp.W541­5. Nucleic Acids Res. ​http://bioportal.bioontology.org/    P.L. Whetzel and NCBO Team. (2013). ​NCBO Technology: powering semantically aware applications​, 4 (Suppl 1),                              S8, J. Biomed. Semantics.    Y. Yan, N. Okazaki, Y. Matsuo, Z. Yang, and M. Ishizuka, (2009). ​Unsupervised relation extraction by mining                                  Wikipedia texts using information from the web​. vol. 2, pp. 1021­1029. Joint Conference of the 47th Annual Meeting of                                      the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP.    L. Yao, & W. A. Sethares, (1994). ​Nonlinear parameter estimation via the genetic algorithm,​, 42(4), pp. 927­935,                                  Signal Processing, IEEE Transactions on.    27  http://www.informatik.uni-trier.de/~ley/pers/hd/s/Syrjakow:Michael.html http://bioportal.bioontology.org/ 
work_mgyp4v6pkbfj7hwkkuu5qbafai	----	doi:10.1016/j.eswa.2007.02.018 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 2019–2028 Expert Systems with Applications Apply ontology and agent technology to construct virtual observatory Ruey-Shun Chen a, Duen-Kai Chen b,* a Department of Information Management, China University of Technology, Taipei 116, Taiwan, ROC b Institute of Information Management, National Chiao Tung University Hsinchu 300, Taiwan, ROC Abstract The need to deal with abundance and heterogeneous information is apparent in the astronomy community. The virtual observatory (VO) concept is the astronomical community’s response to alleviate this problem and Web services serve as one of the most important VO enabling technologies. However, one of the limitations of Web services is the lack of semantic description of its content, thus pro- hibits its ability to understand the queries and its inference capabilities. This study proposes to develop a conceptual framework based on multi-agent systems and ontology technology, in order to create a VO with semantically enriched Web services. Intelligent agents rep- resent: (1) users to submit requests (2) perform semantic matching in between users’ requests and Web services registered within agent platform, and (3) activate a serial of Web services. The capabilities offered by multi-agent systems to query and invoke semantically enriched Web Services is also exploited in this study. To validate the proposed framework, an illustration example is implemented in JADE agent platform to demonstrate how the proposed framework operates and how it benefits the research regarding to auroral images. � 2007 Elsevier Ltd. All rights reserved. Keywords: Multi-agent systems; Ontology; Web services; Virtual observatory 1. Introduction The need to retrieve and process exponentially increas- ing of information is becoming apparent and is having a huge impact on professions that rely on distributed archived information. This is true especially for the auroral physicists who depend largely on the shared image infor- mation content, for example, images from different loca- tions, with different sizes, intensity, and morphology of the auroral oval. Therefore, how to retrieve needed infor- mation, access to language independent software for image processing and furthermore, share the research results becomes an essential topic. The virtual observatory (VO) concept is the astronomical community’s response to the information abundance. VO is an emerging, open, web- 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.02.018 * Corresponding author. Tel.: +886 3 5712121x57427; fax: +886 3 5723792. E-mail addresses: rschen@iim.nctu.edu.tw (R.-S. Chen), cdk@iim. nctu.edu.tw (D.-K. Chen). based, distributed research environment for astronomy with massive and complex data sets. Web services serve as one of VO enabling technologies; however, there are some limitations of the traditional Web services. One of the limitations is that Universal Descrip- tion, Discovery and Integration (UDDI) and Web Services Description Language (WSDL) do not provide semantic description of its content. UDDI is characterized for its lack of semantic description mechanisms, such as semantic interoperability, explicit semantic models to understand the queries and inference capabilities (Baousis et al., 2006). The same limitation also applies to WSDL, for WSDL does not contain information about the capabilities of the described Web services. This paper proposes to develop a conceptual framework based on multi-agent systems and ontology technology to assist auroral physicists work with semantically enriched web services and realize the concept of VO. The above mentioned framework can be used for retrieving necessary information, dynamically allow users to define descriptors and operations most appropriate for their purposes, for mailto:rschen@iim.nctu.edu.tw mailto:cdk@iim.nctu.edu.tw mailto:cdk@iim.nctu.edu.tw Fig. 1. Conceptual architecture of virtual observatory (Schlenoff et al., 2005). 2020 R.-S. Chen, D.-K. Chen / Expert Systems with Applications 34 (2008) 2019–2028 example, comparison of physical features between auroral images, regardless of their source or data representation. In other words, the proposed framework complies with the VO concept. The goal of the proposed framework is to assemble data archives and services, as well as data exploration and analysis tools. Recently, ontology and agent technology have been widely applied to various domains. For example, Jing, Bi, and Wu (2005) pointed out that Geospatial information services may provide the data and function services to han- dle with basic geo-tasks, in the meanwhile, user and soft- ware agent should be able to discover, invoke, compose and monitor geospatial resources offering particular services and having particular properties. Web services are turning into the dominant means for connecting remo- tely executing programs via Internet and commonly used machine readable representations such as XML. The Web Ontology Language for Services (OWL-S) is imported to express the process ontology and process control ontol- ogy. In virtue of the Web services semantic markup, develop the agents to support the automatic reasoning and fulfill the composition of Web Services and inter-oper- ation is possible. The contributions of this paper are twofolds. First is to provide an environment which is easy to obtain and share information with the assistance of ontology and multi- agent systems via dynamically discover and invoke seman- tically enriched Web services. Furthermore, it is believed that the proposed framework could easily be extended for similar domains. Second, the proposed framework utilizes multi-agent systems and ontology for share, retrieve and process information. Adopting multi-agent systems to assist information processing can achieve advantages such as robustness and scalability. The proposed framework enhances flexibility and extensibility owing to its modular design. Modular design means that the module within the architecture can be substituted by better and newer mod- ules when necessary, thus ensuring the flexibility and exten- sibility. To facilitate information sharing among different research projects, the proposed framework develop ontol- ogy which includes standardized, community-accepted descriptions of the information. Such ontology served as the basis for information content definitions that can be used repeatedly. The rest of the paper is organized as follows. Section 2 introduces the background on Web services, VO, and space-based auroral image research; ontology technology; multi-agent systems-especially focuses on the agent com- munication and coordination. Section 3 presents how the proposed framework assists auroral physicists in retrieving and processing images from heterogeneous data sources by incorporated ontology and multi-agent systems technol- ogy. Section 3 also gives the overview of the proposed framework; the types of agents consisted in the framework, the functionality of each agent. Section 4 highlights the software solution and implementation to the proposed framework and provides practical examples to illustrate how images from different sources (e.g. ultraviolet imager (UVI), far-ultraviolet imager (FUV)) can be retrieved and processed. Section 5 gives the conclusion and identifies future works. 2. Literature review 2.1. Space-based auroral image, virtual observatory and Web services A conceptual architecture of a VO is shown in Fig. 1. From a space physicists’ viewpoint, one should be able to discover the available data for their study, which generally reside in distributed archives, federate them, and pipe the output into a set of data mining/knowledge discovery in databases analysis and discovery tools. These services may be implemented as Web services, and may involve use of AI and machine learning tools, coupled with sophis- ticated visualization environments (Djorgovski, 2005). W3C (2004) defines web service as ‘‘. . . a software sys- tem designed to support interoperable machine-to-machine interaction over a network. It has an interface described in a machine-processable format (specifically WSDL). Other systems interact with the Web service in a manner pre- scribed by its description using simple object access proto- col (SOAP) messages, typically conveyed using HTTP with an XML serialization in conjunction with other Web- related standards’’. A Web service is an accessible application that other applications and humans can discover and invoke. Web services are nowadays emerging as a major technology for deploying automated interactions between distributed and heterogeneous applications. Various standards back this deployment, including • WSDL – support the definition of Web services. • UDDI – support advertisement to the community of potential users. • SOAP – binding for invocation purposes. R.-S. Chen, D.-K. Chen / Expert Systems with Applications 34 (2008) 2019–2028 2021 One of the examples for facilitating Web services as an enabling technology for VO described in Ohishi et al. (2004). The authors in Ohishi et al. (2004) depicted how UDDI, which serves as yellow page services, is adopted to establish Japanese virtual observatory (JVO). 2.2. Ontology Ontology is an agreement about a shared conceptualiza- tion, which includes frameworks for modeling domain knowledge and agreements about the representation of particular domain theories (Huhns & Singh, 1997). Ontol- ogy is an esoteric concept borrowed from philosophy by the Artificial Intelligence/Information Technology commu- nities. In this study, we adopt the definition of ontology as ‘‘a way of describing all things in common across a collec- tion of similar objects’’. Ontology served as standardized, community-accepted descriptions of the information con- tent. There existed previous attempts to apply ontology in the field of image processing, in Soo, Lee, Yeh, and Chen (2002), a framework of utilizing sharable domain ontology and thesaurus to help the retrieval of historical images of the First Emperor of China’s terracotta warriors and horses was presented. Ontology has also been used in Earth Sciences and Geographic Information Systems (GIS) appli- cations to organize and allow inter-comparison of remote sensing data sets from disparate detectors (Fonseca, Egen- hofer, Agouris, & Camara, 2002). The reason why ontology is essential to the intelligent agents is that it can provide a shared virtual world in which each agent can ground its beliefs and actions. That is why ontology is becoming increasingly recognized as a crucial element of scalable multi-agent system technology. Recently, there are plenty of literatures (Bailin & Trusz- kowski, 2001; Bravo, Perez, Sosa, Montes, & Reyes, 2005; Obitko & Marik, 2002) which highlight the crucial role ontology played in the development of multi-agent systems. 2.3. Agent technology According to Jennings and Wooldridge’s (1998) defini- tion, an agent is a computer system situated in some envi- ronment, and that is capable of autonomous action in this environment in order to meet its design objectives. Agent systems can be further classified either single-agent or multi-agent systems. Agent systems, particularly multi- agent systems, are subfields of distributed artificial intelli- gence (DAI) research, and have existed under AI for two decades (Chen, Chen, & Lin, 2005). Generally, DAI is bro- ken into distributed problem solving (DPS) and multi- agent systems. Research on designing and developing multi-agent systems focuses on the interaction between agents. Topics frequently discussed in this area of research include agent action, the relationship between agents, multi-agent system architecture and the environment in which the multi-agent systems exist; interactions among agents within the multi-agent system, and agent adapta- tion ability. Coordination, negotiation, cooperation are three common interactions among agents in multi-agent systems. Regarding the application of software agents, Jennings and Wooldridge (1998) noted that ‘‘agent are being used in an increasingly wide variety of applications, ranging from comparatively small systems such as email filters to large, open, complex, mission critical systems such as air traffic control’’. For example, Lee, Yun, and Jo (2003) pro- posed an auction agent system using a collaborative mobile agent and a brokering mechanism called Mobile collabora- tive auction agent system (MoCAAS), which mediates between the buyer and the seller and executes bidding asyn- chronously and autonomously. In the past decade, academic communities have wit- nessed a proliferation of software agents with widely vary- ing specialties. Agent technology is a key area in artificial intelligence research. Today, a software agent generally means a software program that accomplishes a task on behalf of its user. Multi-agent systems along with ontology have been applied for supporting distributed decision mak- ing in several fields, such as manufacturing, business and engineering. For example, a cooperative multi-agent plat- form was introduced in Soo, Lin, Lin, and Cheng (2005) to support the invention process based on the patent doc- ument analysis. There are also studies abundant in inte- grate ontology with agent technology. Wu, He, and Jin (2005) proposed to build an open, large-scale and interop- erable distributed intelligent medical diagnosis and therapy system in the context of a wide-area network such as the Internet, the paper studies integrating mobile agent and ontology with Web Services, using ontology as a standard web service, which avoids misunderstandings in agent com- munication. Furthermore, it assures the correctness of matching agent services according to semantics accurately in service register server like the UDDI registry. Schlenoff, Washington, and Barbera (2005) developed intelligent ground vehicle (IGV) ontology. The goal of their effort was to develop a common, implementation-independent, extendable knowledge source for researchers and develop- ers in the intelligent vehicle community. Malucelli, Palzer, and Oliveira (2005) in their paper, combines the use of ontology and agent technologies to help in solving the semantic heterogeneity problem in e-commerce negotia- tions. The proposed approach aims at creating a methodol- ogy that assesses lexical and semantic similarity among concepts represented in different ontologies without the need to build an a priori shared ontology. Day et al. (2005) described an intelligent tutoring agent (ITA) that uses the ontology, INFOMAP, and question answering techniques through the Instant Messaging platform for the ‘‘operating system’’ course. In virtue of the semantically enriched Web services, develop the intelligent agents to support the automatic rea- soning and decision making, as well as fulfill the composi- tion of Web services and inter-operation is possible. Fig. 2. Proposed system architecture. 2022 R.-S. Chen, D.-K. Chen / Expert Systems with Applications 34 (2008) 2019–2028 3. Apply ontology and agent-based architecture to build virtual observatory In this section, we first describe how the proposed framework operates. Afterwards, the functionality and internal structure of the agents and components are depicted. As illustrated in Fig. 2, the proposed framework mainly consists of the web server and the agent platform for all the agents to reside in. In a typical operational scenario, the user may send the request specifying some user defined cri- teria through the browser, as same as using any of cur- rently existing web application. The advantages of using browser to connect to the proposed framework are easy to use and to ensure that users can be accommodated with- out having agent platform or any additional programs (e.g. Java Runtime Environment) installed on their client side. The web application also provides services such as user account creation, user login, logout and creation of user profile. Web server receives the request and then sends it to the Gateway Agent. The role of the Gateway Agent here is to connect web application and agent platform in a transpar- ent manner; it serves as an intermediary service. Gateway Agent allows the web application to invoke agent; translat- ing the HTTP compliant messages to the Foundation for Intelligent Physical Agents (FIPA) compliant Agent Com- munication Language (ACL) – and vice versa – in order to let Web services and agent technology work together. Per- sonal Agent, as describe in Fig. 3, should be capable of interpreting user’s intention. Once the Personal Agent decomposes the user’s request into subtasks, it then send the request to the Broker Agent and searching for Service Provider Agents capable of accomplish the aforementioned subtasks. Broker Agent allows Service Provider Agents to post its services. After finding appropriate agents to per- form the subtask, Personal Agent then sends request to the target agent and starts to negotiate with the target agent directly. When all of the subtasks had been fulfilled, then Personal Agent will send result to Gateway Agent and Gateway Agent then return the result to user via web page. After introducing the overview of the proposed frame- work; the types of agents consist in the framework and its functionalities are describes as follows. 3.1. Personal agent (PA) The PA serves as a representative for users in agent plat- form. The main purpose of a PA is finding and executing services and delivering results to the user. A PA is com- posed of the following components: • Agent communication module, agents should be able to communicate with each other via ACL. Agent commu- nication module facilitates interaction with Gateway Agent, Broker Agent, and all potential Service Provider Agents. • Domain knowledge base specifies the autonomous behavior of the PA. It contains the information to guide the PA to accomplish the task delegated by the user. • Ontology database is used to store ontology for PA to facilitate query and invoke semantically enriched Web services. • User intention module can be further broken down into User log database, Learning module and User prefer- ences database. The aim of this module is to provide Fig. 3. Personal Agent internal architecture. R.-S. Chen, D.-K. Chen / Expert Systems with Applications 34 (2008) 2019–2028 2023 customize services for each user and attempt to be able to provide the user with more satisfactory services in the future. To be able to assist end user to complete its goal. This is done by ontology and image processing knowl- edge base, as well as the learning capability build within the PA. Being intelligent agent, the PA should also able to improve itself through self-learning. The focus of the learning module is to learn user’s preferences and pro- files in order to provide more accurate assistance. The learning source is provided by the user’s log, which is the history of the user’s actions on the agent platform. Ontology let allows the agent share the same concept of as the domain expert. The learning mechanism will assist agent in identifying the user’s preferences. 3.2. Service provider agent (SPA) The service provider agent serves as a representative of an existing web service in the agent platform. It provides the web service to interested users and maintains a descrip- tion of the web service expressed in WSDL and ontology, such as Web Ontology Language (OWL). The WSDL description is facilitated to find the necessary definitions for Web services’ successful invocation, while OWL is used to enhance the expressiveness of WSDL in terms of seman- tic information. 3.3. Gateway agent This framework presents a Gateway agent for connect- ing intelligent agents and Web services in an automatic, transparent manner. This Gateway agent allows Web ser- vices to invoke agents by translating Web services’ requests to agent communication language encodings, and enable automatic, transparent connection between these two tech- nologies. In other words, Gateway agent serves as Web ser- vices and agent platform’s bridge. Integrating Web services and intelligent agents generates the foreseeable benefits of connecting these two application domains. Once the inter- connection is established, intelligent agent concepts and technologies will help enable new, advanced operational and usage modes of Web services. Service invocation by the Gateway agent depends on the OWL description of the service requests. OWL is used to enhance the expres- siveness of WSDL in terms of semantic information. OWL is used to specify the input and output ontologies, enabling an advanced service capability search, and find the necessary definitions for successful Web services invocation. 3.4. Broker agent Service matching along with ontology and ACL assure the correctness of matching agents and Web services according to semantics in service register server like the UDDI registry. Brokering is one of the most significant discovering mechanisms among autonomous, intelligent agents. The main task of Broker agent is to interpret que- ries from requester and matches capabilities provided by service provider. The brokering process can be categorized into two steps. First, broker agent extracts required capa- bilities from the query send by PA. Then it compares and matches these required capabilities with what service pro- viders can accomplish. Broker agent extendeds the tradi- tional matching mechanism by finding service providers by capability or functionality rather than simply by name. Service providers register their capabilities within Broker agent, which stores all the registries in a local knowledge base. If capabilities change or the service providers no longer provide the service, then broker agent maintains Fig. 4. Service request and service provider matching. Table 2 Comparison of imaging characteristics for contemporary auroral imagers. Imager Frame rate (s) Image size (pixels) Imaging method Spatial resolution (Apogee) Polar/UVI 37 200 · 228 Snapshot 40 km Image/FUV-WIC 120 256 · 256 Time Delayed Imaging 100 km 2024 R.-S. Chen, D.-K. Chen / Expert Systems with Applications 34 (2008) 2019–2028 or deletes the service registries. Fig. 4 illustrates the broker- ing process. 4. Illustration example and discussion 4.1. Illustration example In this section, an illustration example is provided to demonstrate how the proposed framework operates. In a typical scenario, an auroral physicist compares two images from two different imagers, UVI and FUV. The detector- specific structural details are sufficiently different between imagers that comparison of the information content such as the size of the oval can be difficult. The steps to accom- plish the comparison are listed in Tables 1 and 2 summa- rizes imaging characteristics of these two imagers. The typical workflow can be described by Fig. 5. During the past, physicists must be able to master all the image databases management systems, related image processing algorithm and be able to manipulate the image according to the specification requirement. Image processing algo- rithms could be implemented in different programming lan- guages, for example, backgrounds remove and flat field correction mentioned in the workflow were implemented Table 1 Image comparison processes User selects time range and spatial location Access/find data files for Imager 1 (I1) Same for Imager 2 Select records within time range for I1 Same for Imager 2 Process I1 records to convert from stored format to science format. Includes background removal, calibration, flat-field correction, and any other instrument-specific corrections Same for Imager 2 Interpolate image to requested time Same for Imager 2 Select region of interest in interpolated image that matches the requested spatial location Same for Imager 2 Summarize ROI, e.g. mean and standard deviation of all pixels within ROI Same for Imager 2 Compare values from both cameras in IDL, and calibration was implemented in C++. Thus, physicists sometimes need to install or even recompile the program and make it work in a local environment. In our proposed framework, first of all, the workflow is according to the following steps. Step 1, physicist gets to select time range and spatial location of the images to be compared through a browser. The client system is implemented in JSP and Servlet technology. Step 2, if the user is the first time user, Gateway Agent will create a new account and profile, as well as a new PA for this specified user. Then the auroral physicist’s PA decomposes the task ‘‘compare two images’’ into the process mentioned above through the help of Auroral Image Knowledge Base. Step 3, PA sends the request to the Broker Agent for matching process, to see if there exists any needed Data and Functional Agents (In this illustration example, SPA can be further categorized as data agent and functional agent. The former provides data retrieval and the latter provides variety of other functionalities.). Step 4, after the matching services provided by Broker Agent, the PA first sends a Call-For-Proposal to all the potential Data Agent (representing which imager) to deter- mine which Data Agents have the image that satisfies the user’s request. Some of the Data Agents respond with Pro- pose-essentially, detail description of the imager and its capability to fulfill the user’s request. The PA then chooses those imagers match user’s request. Step 5, PA then sends the user’s query to Data Agents as a service request. Now it is the Data Agent’s turn to inter- pret and try to fulfill the request. After receiving the result from Data Agents, the PA can apply image processing algorithms, such as image background removal, calibra- tion, flat field correction, to the images. Step 6, in order to do so, again, the PA sends a Call-For- Proposal to potential Functional Agent (representing image processing algorithm) which can be helpful in satis- fying the user’s request. Some of the Functional Agents respond with Propose, and PA decide which Functional Agent is most appropriate to perform the image processing according to the user’s preferences (such as processing time or accuracy). After receiving the response from the Func- tional Agent, the PA then represents the result to Gateway Agent, and Gateway Agent then sends the result to user via JSP page. Such process can be described by the flowchart listed in Fig. 6. Fig. 5. Image comparison workflow. Send request to PA First time user Create new PA through gateway PA interprets user request PA find appropriate data agent in DF PA send fail message to user and End PA request DA to perform data retrieval DA accepts Pa s request or not DA performs data retrieval and send result back to PA Yes No Yes No PA find appropriate functional agent in DF No FA performs requested function and send result back to PA Yes Fig. 6. Workflow of how illustrated example is accomplished. R.-S. Chen, D.-K. Chen / Expert Systems with Applications 34 (2008) 2019–2028 2025 We implement the prototype with JADE and Tomcat web server to demonstrate the efficacy of this framework. Currently, JADE is perhaps the most pervasive agent sys- tem in use, especially in the context of research driven applications. JADE is an open source, FlPA compliant agent platform designed to be a middleware solution for developing agent-based software. Fig. 7 illustrates how agents in the prototype communicate with each other via FIPA compliant ACL. We use Protégé as our ontology edi- tor. Protégé, developed by Stanford University, is an ontol- ogy editor, a knowledge-base editor, an open-source, Java tool. One of the advantages of Protégé is that it provides an extensible architecture for the creation of customized knowledge-based applications. The reason for chosen Pro- tégé was due to its strong user community and its ability to support the OWL as ontology representation. The Sniffer agent, a JADE tool which is basically a FIPA-compliant Agent with sniffing features in order to track messages exchanged between agents in a JADE based environment, was employed to verify the validity of the model proposed in this paper. Every message exchanged between agents implies the execution of the model, can be tracked to view in runtime. Fig. 8 drawn by Sniffer agent tool from JADE depicts the message exchange between agents. 4.2. Discussion Comparing the user behaviors-in both traditional means to handle the image retrieval, image processing and via proposed framework – in the scenario described above, it is believed that the workload of the users can be alleviated. Fig. 9 describes the steps to be accomplished in order to perform analysis for two images from two imagers. Note that in the proposed framework, the tasks to be accomplished within the dotted rectangle are handled by the multi-agent systems. This implies the following advan- tages for the users. (1) Web services solve the problem of obtaining (includ- ing find, install or even recompile) the appropriate software to retrieve and process the images. (2) The auroral image process knowledge base can be facilitated for the PA to choose appropriate Web ser- vices to perform image retrieval and image process for different images from different imagers and with different image specifications. In other words, users may invoke a set of services. By this means, users do not have to be familiar with all the imagers’ spec- ification to choose the appropriate tools to accom- plish the pre-processes for the analysis. This is an explicit advantage, since collaboration between vari- ous image teams is common and users might not be familiar with other imager’s specification and related image processing algorithm. The proposed frame- work allowed users to focus on their research and Fig. 7. Agent communication sequence. 2026 R.-S. Chen, D.-K. Chen / Expert Systems with Applications 34 (2008) 2019–2028 analysis instead of mastering image processing algo- rithm, software, and tools. (3) Apply ontology to facilitate semantic enriched Web services can provide better request-service matching. Service matching performed by Broker Agent along with ontology can assure the correctness of matching service requester and service provider. In other words, the capabilities offered by agents to query and invoke semantically enriched Web services is based on enhanced registries enriched with semantic informa- tion that provide semantic matching to service queries and published service descriptions. (4) New services, agents, and users can be easily added to the framework, thus providing an extendable and flexible open system. 5. Conclusion In this study, we present a framework that provides easy access to heterogeneous data sources and image pro- cessing algorithms through ontology and multi-agent Sys- tems; meanwhile, we also exploit the capabilities offered by multi-agent systems to query and invoke semantically enriched Web services. All the users, data sources and algo- rithms are represented by different types of agents. Service providers register themselves to the yellow page service provided by broker agents with semantic information. PA represent users and send user’s requests to broker agents. Broker agent then performs semantic matching in between services registered within yellow page services and service request submitted by Personal Agent. The advantages of Fig. 8. Snapshot of agent communication sequence in JADE. Fig. 9. Illustration example workflow. R.-S. Chen, D.-K. Chen / Expert Systems with Applications 34 (2008) 2019–2028 2027 the proposed framework can be summarized as follows: (1) develop ontology which includes standardized, commu- nity-accepted descriptions for the auroral images. (2) Auro- ral scientists can focus on their research instead of mastering image processing algorithm, software, tools. (3) Users may invoke a set of Web services with the integration of Web services and intelligent agents. This generates the foreseeable benefits of connecting these two application domains, once the interconnection is established, intelligent agent concepts and technologies will help enable new, advanced operational and usage modes of Web services. (4) Provide an environment which is easy to obtain and share research data and image processing algorithms with the assistance of ontology and multi-agent systems. (5) New services, agents, users and service registries can be eas- ily added to the framework, thus providing an expandable and flexible open system. The framework could easily be extended for similar domains, such as other satellite images and Earth Sciences. The prototype implementation indi- cates that the proposed framework is flexible, robust and extendable. References Bailin, S. C., & Truszkowski, W. (2001) Ontology negotiation between scientific archives. In 2001 Proceedings of thirteenth international conference on scientific and statistical database management (pp. 245– 250). Virginia. 2028 R.-S. Chen, D.-K. Chen / Expert Systems with Applications 34 (2008) 2019–2028 Baousis, V., Zavitsanos, E., Spiliopoulos, V., Hadjiefthymiades, S., Merakos, L., & Veronis, G. (2006) Wireless web services using mobile agents and ontologies. In 2006 ACS/IEEE international conference on pervasive services (pp. 69–77). Beijing. Bravo, M. C., Perez, J., Sosa, V. J., Montes, A., & Reyes, G. (2005) Ontology support for communicating agents in negotiation processes. In 2005 Fifth international conference on hybrid intelligent systems (pp. 482–487). Rio de Janeiro. Chen, R., Chen, D., & Lin, S. (2005). ACTAM: Cooperative multi-agent system architecture for urban traffic signal control. IEICE Transac- tions on Information and Systems, E88-D(1), 119–126. Day, M., Lu, C., Yang, J., Chiou, G., Ong, C., & Hsu, W. (2005) Designing an ontology-based intelligent tutoring agent with instant messaging. In ICALT 2005 fifth IEEE international conference on advanced learning technologies (pp. 318–320). Kaohsiung. Djorgovski, S. G. (2005) Virtual astronomy, information technology, and the new scientific methodology. In Proceedings of seventh international workshop on computer architecture for machine perception, CAMP (pp. 125–132). Terrasini-Palermo. Fonseca, F., Egenhofer, M., Agouris, P., & Camara, C. (2002). Using ontologies for integrated geographic information systems. Transactions in GIS, 6(3), 231–258. Huhns, M. N., & Singh, M. P. (1997). Ontologies for agents. Internet Computing, IEEE(6), 81–83. JADE, http://jade.tilab.com/. Jennings, N. R., & Wooldridge, M. J. (1998). Applications of intelligent agents. In Agent technologies: Foundations, applications, and markets (pp. 3–28). Springer. Jing, D., Bi, S., & Wu, F. (2005) Geospatial information services on the basis of agent and OWL-S. In Proceedings of 2005 IEEE international geoscience and remote sensing symposium, IGARSS ’05 (pp. 885–888). Seoul. Lee, K., Yun, J., & Jo, G. (2003). MoCAAS: Auction agent system using a collaborative mobile agent in electronic commerce. Expert Systems with Applications, 24, 183–187. Malucelli, A., Palzer, D., & Oliveira, E. (2005) Combining ontologies and agents to help in solving the heterogeneity problem in e-commerce negotiations. In Proceedings of international workshop on data engineering issues in e-commerce (pp. 26–35). Tokyo. Obitko, M., & Marik, V. (2002) Ontologies for multi-agent systems in manufacturing domain. In Proceedings of 13th international workshop on database and expert systems applications (pp. 597–602). Aix-en- Provence. Ohishi, M., Mizumoto, Y., Yasuda, N., Shirasaki, Y., Tanaka, M., Honda, S., & Masunaga, Y. (2004) A prototype toward Japanese virtual observatory (JVO). In 2004 International symposium on applications and the internet workshops (pp. 591–595). Tokyo. Schlenoff, C., Washington, R., & Barbera, T. (2005) An intelligent ground vehicle ontology to enable multi-agent system integration. In Interna- tional conference on integration of knowledge intensive multi-agent systems (pp. 169–174) Waltham. Soo, V., Lee, C., Yeh, J., & Chen, C. (2002) Using sharable ontology to retrieve historical images. In Proceedings of the 2nd ACM/IEEE-CS joint conference on digital libraries (pp. 197–198). Portland. Soo, V., Lin, S. Yang, S., Lin, S., & Cheng, S. (2005) A cooperative multi- agent platform for invention based on ontology and patent document analysis. In Proceedings of the ninth international conference on computer supported cooperative work in design (pp. 411–416). W3C, Web Services Architecture, http://www.w3.org/TR/2004/NOTE- ws-arch-20040211/, 2004. Wu, Z., He, Y., & Jin, H. (2005) A model of intelligent distributed medical diagnosis and therapy system based on mobile agent and ontology. In Proceedings of eighth international conference on high-performance computing in Asia-Pacific region (pp. 582–587). Beijing. http://jade.tilab.com/ http://www.w3.org/TR/2004/NOTE-ws-arch-20040211/ http://www.w3.org/TR/2004/NOTE-ws-arch-20040211/ Apply ontology and agent technology to construct virtual observatory Introduction Literature review Space-based auroral image, virtual observatory and Web services Ontology Agent technology Apply ontology and agent-based architecture to build virtual observatory Personal agent (PA) Service provider agent (SPA) Gateway agent Broker agent Illustration example and discussion Illustration example Discussion Conclusion References 
work_midcwaprtbgflduuek4gea36su	----	Staff-line removal with Selectional Auto-Encoders Antonio-Javier Gallego, Jorge Calvo-Zaragoza∗ Department of Software and Computing Systems University of Alicante, Carretera San Vicente del Raspeig s/n, 03690 Alicante, Spain Abstract Staff-line removal is an important preprocessing stage as regards most Optical Music Recognition systems. The common procedures employed to carry out this task involve image processing techniques. In contrast to these traditional methods, which are based on hand-engineered transformations, the problem can also be approached from a machine learning point of view if representative ex- amples of the task are provided. We propose doing this through the use of a new approach involving auto-encoders, which select the appropriate features of an input feature set (Selectional Auto-Encoders). Within the context of the prob- lem at hand, the model is trained to select those pixels of a given image that belong to a musical symbol, thus removing the lines of the staves. Our results show that the proposed technique is quite competitive and significantly out- performs the other state-of-art strategies considered, particularly when dealing with grayscale input images. Keywords: Staff-line removal, Optical Music Recognition, Auto-encoders, Convolutional Networks 1. Introduction Music is an important vehicle for cultural transmission, which is a key ele- ment as regards understanding the social, cultural and artistic trends of each ∗Corresponding author: Tel.: +349-65-903772; Fax: +349-65-909326 Email addresses: jgallego@dlsi.ua.es (Antonio-Javier Gallego), jcalvo@dlsi.ua.es (Jorge Calvo-Zaragoza) Preprint submitted to Expert Systems with Applications March 7, 2018 Usuario Texto escrito a máquina This is a previous version of the article published in Expert Systems with Applications. 2017, 89: 138-148. doi:10.1016/j.eswa.2017.07.002 http://dx.doi.org/10.1016/j.eswa.2017.07.002 period of history. A large number of musical documents have, therefore, been carefully preserved over the centuries and are scattered throughout cathedrals,5 libraries or historical archives (Fujinaga et al., 2014). A significant effort has been made in recent decades to digitize these docu- ments by means of scanners for their storage and distribution (Choudhury et al., 2000). However, in order to make the music contained in the documents truly accessible, it is necessary for the images to be transcribed to a structured digital10 format that makes it possible to encode the content (notes, musical symbols, tonality, etc.) of the document. This also makes it possible to perform other interesting tasks, such as large-scale computational music analysis, search and retrieval by content, or transcription between different musical notations (Han- kinson et al., 2012).15 The process of converting a scanned document into a musical structured digital format can be carried out manually by a user. The disadvantage is that it involves costs in terms of both resources and time. In addition, this process is especially tedious—because of the burdensome software for score edition— and very prone to introducing errors.20 The research field known as Optical Music Recognition (OMR), which fo- cuses on detecting and storing the musical content of a score from a scanned image (Raphael & Wang, 2011), has therefore been postulated as an important alternative that mitigates the aforementioned disadvantages of manual tran- scription. The objective of the OMR process is to import a scanned musical25 score and export its musical content to a machine-readable format 1), typically MusicXML or MEI. The OMR task is similar to the recognition of text, typically known as Optical Character Recognition. However, unlike the text scenario in which words are analyzed sequentially with a single element to identify at each time30 instant, musical notation is considerably more difficult to recognize. This is principally owing to the possible existence of simultaneous matching elements, as in the case of polyphonic pieces with multiple notes that sound at the same time, thus resulting in several musical symbols being placed on the same interval. But 2 (a) Example of input piece for an OMR system (b) Symbolic representation of the piece Figure 1: The task of Optical Music Recognition (OMR) is to analyze an image containing a musical score in order to export its musical content to a machine-readable format. it is also because of the presence of marks of expression, dynamics, articulations35 or even text to be sung in works with a vocal presence, among others. Most current systems employ segmentation and classification approaches (Rebelo et al., 2010; Wen et al., 2015). The first important obstacle that the OMR process must overcome is, therefore, the staff (or pentagram) lines: the set of five parallel lines on which musical symbols are located depending on40 their pitch. The staff-line removal stage is usually performed after the binariza- tion of the document in the OMR workflow (Rebelo & Cardoso, 2013). This binarization step helps to reduce the complexity of the problem, and it is ad- visable to apply strategies based on histogram analysis, connected components, or morphological operators.45 Despite being necessary for musical readability, staff lines complicate the au- tomatic detection and segmentation of symbols. Some specific works have taken advantage of specific features of printed and/or ancient notation to approach the 3 problem of maintaining the staff lines (Ramirez & Ohya, 2014; Calvo-Zaragoza et al., 2015); however, the established OMR pipeline includes their detection50 and removal (Rebelo et al., 2012). This process must remove the staff lines while maintaining as much of the symbol information as possible (Fig. 2). (a) Example of input score for an OMR system (b) Input score after staff-line removal Figure 2: Example of a perfect staff-line removal process. This paper proposes a framework with which to remove staff lines that is based on machine learning, that is, labeled examples can be used to train a model to perform the task. This allows using the same approach in a wide55 range of scores. We make use of a new approach of auto-encoder, which is trained to select only those pieces of the image that belong to musical symbols. The remainder of the paper is structured as follows: Section 2 presents the background to staff detection and removal; Section 3 describes our approach with which to model the process; Section 4 contains the experimentation per-60 formed and the results obtained; and finally, the current work concludes in Section 5. 4 2. Background This section presents the background to our approach. First, methods pro- posed for staff-line removal are introduced, after which the principles of auto-65 encoders are briefly described, given their importance as regards the present work. 2.1. Staff-line detection and removal Staff-line removal has been an active research field for many years. A com- prehensive review and comparison of the first attempts considered for this task70 can be consulted in the work of Dalitz et al. (2008), who divided the staff-line removal strategies proposed until then into four categories: the Line Track- ing (Randriamahefa et al., 1993; Bainbridge & Bell, 1997), Vector Field (Martin & Bellissant, 1991; Roach & Tatem, 1988), Runlength (Carter & Bacon, 1992; Fujinaga, 2005) and Skeleton (Ng, 2001) methods. However, given the interest75 in the task, many other methods have been proposed more recently. Dos Santos Cardoso et al. (2009) proposed a method that considers the staff lines to be connecting paths between the two margins of the score. The score is then modeled as a graph so that staff detection is solved as a maximization problem. This strategy was later improved and extended to be used on grayscale80 scores (Rebelo & Cardoso, 2013). Dutta et al. (2010) developed a method that considers the staff line segment as a horizontal connection of vertical black runs with a uniform height. These segments are validated using neighboring properties before removing them. Su et al. (2012) started by estimating the properties of the staves such as85 height and space. An attempt is then made to predict the direction of the lines and fit an approximate staff, which is subsequently adjusted. Géraud (2014) developed a method that entails a series of morphological operators: first, a permissive hit-or-miss with a horizontal line pattern, followed by a horizontal median filter and a dilation operation. A binary mask is then90 obtained with a morphological closing. Finally, a vertical median filter is applied 5 to the largest components of the mask. The procedure is directly applied to the image, which eventually removes staff lines. Notwithstanding all the efforts made, the staff-removal stage is still inaccu- rate and often produces noise—staff lines not completely removed. Although it95 is possible to use more aggressive methods that minimize this noise, they may cause the partial or total loss of some musical symbols. The trade-off between these two aspects, in addition to the accuracy of the techniques, has hitherto led to the inevitable production of errors during this stage. Moreover, the differences among score style, sheet conditions and scanning100 processes have led researchers to develop ad-hoc method for staff-line detection and removal. This results in methods that are not robust when applied to differ- ent types of document (from different eras, or with different notations or styles). In modern notation, a staff is composed by five black parallel lines over a white background. However, that is not always true in old music notation because105 (i) staff lines may appear with different ink colors, even closer to background color than to symbol color, (ii) handwritten staff lines are not totally straight, (iii) the thickness of the lines is irregular because of quill leakage, and (iv) the staff does may have less than five lines. In addition, lyrics in modern scores are always far enough from the staff, whereas there is much overlapping be-110 tween music and lyrics in old notation, and so it could also hinder the staff-line removal process. Therefore, many of the assumptions for staff-line removal in modern music are not always fulfilled in different eras, thereby being extremely difficult to develop methods that are able to work on any kind of scores with an acceptable accuracy.115 Here we introduce a new and more generalized framework based on ma- chine learning that can be applied to a wide variety of musical notation styles and musical documents. The main advantage of using machine learning lies in its ability to be generalized when compared with hand-crafted systems. In this respect, a machine learning strategy for staff-line removal has recently been pro-120 posed, which consists of training a classifier that discriminates between whether a given pixel belongs to a symbol or to a staff line (Calvo-Zaragoza et al., 2016). 6 The foreground pixels of the image are queried so that those classified as staff are removed. Nevertheless, it has two important disadvantages, the first being that this strategy does not improve the performance of other state-of-the-art125 algorithms for staff-line removal. The second is its high computational cost, which results from having to classify each pixel in the image. The aim of this work is to alleviate the aforementioned issues by using spe- cialized auto-encoders. The following section presents a brief introduction to auto-encoders and their related extensions in order to provide the reader with130 a better understanding of the proposed framework. 2.2. Auto-encoders Auto-encoders consist of feed-forward neural networks for which the input and output must be exactly the same. The network typically consists of two stages that learn the functions f and g, which are called encoder and decoder135 functions, respectively. Formally speaking, given an input x, the network must minimize the diver- gence L(x, g(f(x))). The hidden layers of the encoder perform a mapping of the input —usually decreasing its dimension— until an intermediate representation is attained. The same input is then subsequently recovered by means of the140 hidden layers of the decoder function. An auto-encoder might initially appear to be useless because it is trained to learn the identity function. Nevertheless, the encoder function f is typically forced to produce a representation with a lower dimensionality than the input. The encoder function therefore provides a meaningful compact representation145 of the input, which might be of great interest as regards feature learning or dimensionality reduction (Wang et al., 2014). The idea of auto-encoders was proposed decades ago (Hinton & Zemel, 1994), and it has since been an active research field (Deng et al., 2010; Lauly et al., 2014). As auto-encoders are feed-forward networks, they can be trained by using150 conventional optimization algorithms such as gradient descent. In some applications, it is assumed that an input x̂ is received after passing 7 through a noisy process that corrupted the actual input x. In this case, a denois- ing auto-encoder might be taught to minimize the divergence L(x, g(f(x̂))). The network, therefore, not only focuses on copying the input but also on removing155 the noise (Vincent et al., 2010; Bengio et al., 2013). In the context of staff-line removal, we could have formulated the problem by assuming that these lines are the result of noise and a denoising auto-encoder could, therefore have been taught to remove them. However, in this paper we go one step further and propose a new type of auto-encoder especially designed160 for the problem at hand. In this case, we wish neither to learn the identity function nor an underlying error but rather a codification that maintains only those input features that are relevant. That is, the output of our auto-encoder is a codification with the same dimension as the input that indicates the original features that must be maintained. Note that, in the staff-line removal formula-165 tion, the relevant features will be those pixels of the image that depict symbols, while those containing staff lines must be discarded. 3. Staff-line removal with Selectional Auto-Encoders As mentioned above, the traditional formulation for the staff-line removal task considers a binary image as input; that is, a binarization process is assumed170 to be performed before this step. The binary nature of modern musical scores (black ink on white paper) has, to some extent, justified this pipeline. However, it should be borne in mind that document binarization is not a trivial question — especially when dealing with ancient documents (Ntirogiannis et al., 2014). Furthermore, the staff-line removal stage is actually a process that attempts175 to leave only the musical symbols in the image. From this point of view, our proposal focuses on selecting the pixels of the input image that correspond to musical symbols, regardless of the nature of the input (see Fig. 3), thus leading to a more generalizable approach that can be applied to binary, grayscale or color images.180 This has been done by making use of a new type of auto-encoder. As 8 Input score image Binarization Staff-line removal Symbol-only score image (a) Traditional approach for isolating musical symbols in score images Pixel selectionInput score image Symbol-only score image (b) Proposed approach for isolating musical symbols in score images Figure 3: Comparison of traditional and proposed approaches for isolating symbols in musical score images. mentioned above, the model is trained to select which features of the input layer are relevant for the task at hand (that is, the pixels that belong to mu- sic symbols). From here on, we shall refer to this model as the Selectional Auto-Encoder (SAE). The SAE is trained to perform a function such that185 s : R(w×h) → [0, 1](w×h). In other words, it learns a binary map over a w × h image that preserves the input shape. Following the idea of auto-encoders, however, the function is further divided into encoding and decoding stages. The topology of an SAE can be quite varied. However, we have restricted ourselves to considering convolutional models. Convolutional models have been190 applied with great success to the detection, segmentation and recognition of objects and regions in images, and have even come close to human performance in some of these tasks (LeCun et al., 2015). They can take advantage of local connections, shared weights, pooling and the use of many connected layers. The hierarchy of layers of our SAE consists of a series of convolutional plus195 pooling layers, until an intermediate layer in which meaningful representations of the input are attained. As these layers are applied, filters are able to relate parts of the image that were initially far apart. It then follows a series of convolutional plus upsampling layers that reconstruct the image up to the same input size. The last layer consists of a set of neurons with sigmoid activation200 that predict a value in the range of [0, 1], depending on the selectional level for the corresponding input feature. The scheme of our hierarchy is illustrated in Fig. 4. The specific configuration, along with a suitable size of the input/output 9 Convolutions & Poolings 5x5x96 5x5x96 5x5x96 5x5x96 5x5x96 5x5x96 Convolutions & UpSampling Figure 4: General overview of a SAE used for staff-line removal. The output layer consists of the activation level assigned to each input feature (white signifies activated) layer, has been adjusted by means of a grid search of the network configuration205 hyper-parameters, as will be detailed in the next section. As a preliminary proof-of-concept test, we also carried out some research with non-convolutional models, which proved to be less suitable for the task at hand. Since an SAE is a type of feed-forward network, the training stage is carried out by means of back-propagation, considering the cross-entropy loss function210 between each output activation and its expected activation. Let n be the number of training examples, and d be the dimensionality of the input (and output). Let us denote the activation of the jth neuron of the last layer for the example i as aij and its desired activation as yij. The loss L for the training set can therefore be computed as:215 L = − 1 nd n∑ i=1 d∑ j=1 [yij ln aij + (1 −yij) ln(1 −aij)] The learning of the network parameters is performed by means of stochastic gradient descent (Bottou, 2010) with a mini-batch size of 8 samples, considering the adaptive learning rate proposed by Zeiler (2012). This training stage consists of providing the SAE with examples of images and their corresponding ground- truth, that is, binary maps over the pixels that belong to musical symbols (see220 Fig. 4). Once the SAE has been properly trained, removing staff lines from the image of a musical score consists of querying the image patch by patch. Each patch 10 is passed through the SAE, which outputs the selection level assigned to each input pixel. Those pixels whose selection value exceeds a certain threshold are225 considered to belong to a musical symbol, whereas the others are discarded. Figure 5 illustrates this process. (a) Score cut into patches of fixed size (b) Selection values obtained (c) Score after thresholding Figure 5: Example of staff-line removal task using an SAE. The input image is parsed patch by patch by means of the network. A selection value is predicted for each pixel in a patch (shown here as grayscale levels). Finally, a thresholding is applied in order to select the pixels that will eventually be maintained. 11 4. Experiments This section presents the experiments carried out to evaluate the goodness of our proposal 1. We took advantage of the ICDAR / GREC 2013 Competition230 on Music Scores (Fornés et al., 2013) staff-line removal contest by making use of the same dataset to allow reproducible research and future comparisons with our results. This corpus contains trios consisting of a grayscale image of a score and its corresponding binarized version with and without staff lines (see Fig. 6). This235 dataset therefore provides readily-available data with which to train the model, along with testing data for evaluation. The corpora is organized into train and test sets, containing 4 000 and 2 000 samples, respectively. (a) Grayscale score (b) Binary score (c) Ground-truth Figure 6: Example of sample from the GREC/ICDAR 2013 Staff-Line Removal Competition dataset. In our case, the train set is further subdivided into training and validation partitions. These partitions are distributed in 80 % and 20 % out of the total240 set of available training scores, respectively. Training data is used to perform 1For the sake of reproducible research, the code of the experiments is available at http: //github.com/ajgallego/staff-lines-removal available under the conditions of the GNU General Public License version 3. 12 http://github.com/ajgallego/staff-lines-removal http://github.com/ajgallego/staff-lines-removal the optimization of the network weights by means of gradient descent during a maximum of 200 epochs, with a mini-batch size of 8 samples. Validation data is used to monitor the training process and prevent over-fitting. Thus, the training process is stopped if the validation performance do not increase during245 5 epochs. In order to evaluate the results, the F-measure (F-m) metric will be consid- ered, following the guidelines of this contest to select the best method. Let TP be the number of true positives (symbol pixels correctly classified), FP be the number of false positives (symbol pixels incorrectly classified) and FN be the250 number of false negatives (staff-line pixels incorrectly classified). Therefore, F-m = 2 · TP 2 · TP + FN + FP We first present the results obtained with the different topologies proposed for the SAE, with a study of the influence of certain hyper-parameters. We then compare our proposal with other methods for solving the same task that participated in the latest contest.255 We also include further results concerning related issues, such as the robust- ness of the model with regard to the threshold. Before carrying out that analysis, we should state that the results shown below are obtained by assuming the best threshold for each case: 0.3 for binary images and 0.1 for the grayscale. The amount of data required to train the model successfully and the representational260 power of the model are also analyzed. 4.1. Hyper-parameter selection This section presents the results obtained with different topologies of the SAE, namely the number of layers and input sizes. In order to reduce the search space, we have restricted ourselves to: i) considering symmetric models,265 that is, those with the same number of coding and decoding layers (from 1 to 3 for each stage); ii) considering only square images of sides 64, 128, 256, 384 and 512. 13 There are also a number of parameters to be tuned, such as the kernel size of the convolutions, the number of filters per layer and the batch size selected270 during training. We have performed comprehensive experimentation in order to tune these parameters by means of a grid search of: • The number of filters per convolutional layer: 16, 32, 64, 96 • The kernel size of each convolutional layer: 3, 5, 7, 9, 11 Since the number of configurations becomes huge, this first experiment fo-275 cus on finding the optimal hyper-parameters only for the binary format of the dataset. The idea of doing this hyper-parameter tuning only with binary images is twofold: on the one hand, it is the case that has been traditionally studied and, on the other hand, we do not want the grayscale results, the more complex scenario, depending on an excessive tuning of the model that best fits. Our280 premise is that the goodness of our work lies in the proposed approach, and not in selecting the optimum topology for every case. In addition, claiming that a topology is the optimal one entails a very exhaustive search, and so we believe that exploiting that avenue is not of actual interest in this work. We there- fore assume that the best configuration for this binary case will also achieve285 competitive results in any other domain. Table 1 shows the best result attained by each different SAE configuration on the validation set. For the sake of readability, it merely reports the best configuration as regards the number of filters and the kernel size of the con- volutional layers obtained for each combination of number of layers and input290 size. As an initial remark, we should state that our approach is relatively robust to the different number of configurations, since all of the results consist of very accurate figures. With regard to the window size, the results do not yield any magic number but the best figures appear to indicate that either smaller295 or bigger window sizes would not significantly improve the results. On the contrary, the number of layers has a higher impact on the performance, leading to a variation in the F-m attained of up to 2 units. 14 Window size # layers 64 128 256 384 512 Average 1 96.99 97.08 97.21 98.40 96.94 97.32 2 97.73 98.70 97.79 97.73 97.79 97.95 3 97.80 97.84 99.13 97.81 97.61 98.01 Table 1: F-m (%) attained by the different number of layers considered (rows) in combination with different values of input window size (columns) for binary images. The values in bold type highlight the best results in each row. The model with 3 layers and 256 × 256 input patches performed best in this experimentation. Specifically, the complete configuration that reported the300 best accuracy comprises 96 filters per layer, and kernel sizes of 5. A technical description is given in Table 2. In the following, we shall assume that this configuration is a representative of our proposal for both binary and grayscale images. 4.2. Comparison with state-of-the-art305 Taking advantage of the aforementioned contest, we tested our method against state-of-the-art staff-line removal strategies. Since the conditions and participants are very different, the binary and grayscale format of the contest are analyzed separately. The test set provided is further divided into three subsets (TS1, TS2, and310 TS3) in order to measure the robustness of the participants with regard to the deformations applied to the scores: 3D distortions in TS1 (three types of distortions with 166, 167, and 167 samples; 500 scores in total), local noise in TS2 (two types of distortions with 250 samples each; 500 scores in total), and both 3D distortion and local noise in TS3 (6 specific distortions, equally315 distributed, as a result of the combination of the previous ones; 1 000 scores in total). Table 3 shows the results obtained by the participants in the contest for the binary case. The main idea of each method was described in Sect. 2 — unlike 15 Input Encoding Decoding Output Conv(96,5,5,ReLU) Conv(96,5,5,ReLU) MaxPool(2,2) UpSamp(2,2) Conv(96,5,5,ReLU) Conv(96,5,5,ReLU) MaxPool(2,2) UpSamp(2,2) [0, 255]256×256 Conv(96,5,5,ReLU) Conv(96,5,5,ReLU) [0, 1]256×256 MaxPool(2,2) UpSamp(2,2) Conv(96,5,5,ReLU) Conv(96,5,5,ReLU) MaxPool(2,2) UpSamp(2,2) Conv(1,5,5,Sigmoid) Table 2: Detailed description of the selected SAE architecture. Conv(f,h,w,a) stands for a convolution operator of f filters, with h × w pixel kernels with an a activation function; Max- Pool(h,w) stands for the max-pooling operator with a w × h kernel and stride; UpSamp(h,w) denotes an up-sampling operator of h rows and w columns; ReLU and Sigmoid denote Rec- tifier Linear Unit and Sigmoid activations, respectively. TAU, which was a method specifically designed to participate in the contest.320 Those readers who require more detailed information about the participants are referred to the competition report (Visaniy et al., 2013), as some of the aforementioned strategies were slightly tuned for the contest. Moreover, we include the work of Calvo-Zaragoza et al. (2016) (Pixel), since their results were obtained under the same conditions of the contest. The figures of our approach325 are those obtained by the selected configuration topology, based on the results depicted in the previous section. As can be seen, most participants are able to achieve good performance fig- ures, being some of them really close to the optimum. However, our approach is able to improve on all the results obtained previously in all cases considered.330 This improvement is especially remarkable in TS3, when distortions are more aggressive. The comparison with Pixel method is also illustrative of the good- 16 TS1 TS2 TS3 Whole TAU (Visaniy et al., 2013) 85.72 81.72 82.29 83.01 NUS (Su et al., 2012) 69.85 96.25 67.43 75.24 NUASi-lin (Dalitz et al., 2008) 94.99 94.86 94.00 94.29 NUASi-skel (Dalitz et al., 2008) 94.25 93.80 92.92 93.34 LRDE (Géraud, 2014) 97.73 96.86 96.98 97.14 INESC (Dos Santos Cardoso et al., 2009) 89.29 97.72 88.52 91.01 Pixel (Calvo-Zaragoza et al., 2016) 94.10 98.11 94.00 95.04 Our approach 99.11 99.36 99.03 99.13 Table 3: F-m (%) comparison of the participants in the ICDAR / GREC 2013 staff removal contest (binary format) and our approach based on SAE. The values in bold type represent the best result in each set, whereas the underlined values represent the best result without considering our model. The column Whole refers to the weighted average obtained for the three test sets. ness of our proposal, since it demonstrates that the performance is not only achieved by using a supervised learning scheme (Pixel also does so) but because of the adequacy of the proposed model. Taking the test corpora as a whole,335 our approach is able to improve up to 2 points with respect to the best result achieved so far. As mentioned previously, the dataset provided in the contest also contains a grayscale version of the scores. Our approach can easily be extended to deal with grayscale images with no further effort but to change the training data. In340 this case, only two of the methods submitted to the contest dealt with grayscale images: LRDE and INESC. Table 4 shows the results obtained by these partic- ipants, when compared to those obtained by our SAE, including again those of Pixel method. It is clear that the participants performance decreases remarkably, particu-345 larly as regards the INESC method. LRDE maintains a fair accuracy in TS1, but its performance is much worse in TS2 and TS3. Pixel method has a more robust behavior, and achieves similar figures regardless of the distortions ap- plied to the images. With more room for improvement, our method achieves a 17 TS1 TS2 TS3 Whole LRDE (Géraud, 2014) 92.17 79.47 79.88 82.85 INESC (Dos Santos Cardoso et al., 2009) 38.50 52.11 38.87 42.09 Pixel (Calvo-Zaragoza et al., 2016) 92.56 88.84 89.76 90.24 Our approach 99.14 99.34 98.94 99.09 Table 4: F-m (%) comparison of the participants in the ICDAR / GREC 2013 staff removal contest (grayscale format) and our approach based on SAE. The values in bold type represent the best result in each set, whereas the underlined values represent the best result without considering our model. The column Whole refers to the weighted average obtained for the three test sets. performance far superior to the participants. It also undergoes a certain drop in350 accuracy with grayscale images but results still consist of very accurate figures, clearly outperforming the other methods. In this case, our approach is able to improve up to almost 10 points the state-of-the-art figure. Results reported above only reflect the average performance of the methods. In order to minimize the possibility that the differences are due to chance varia-355 tion, we perform a pairwise, non-parametric Wilcoxon signed-rank test Demsar (2006). We considered the 11 independent results (one per specific distortion applied) to perform these tests. It resulted in p-values below 0.01 in all pairwise comparisons between our approach and the other methods, in both binary and grayscale formats. Therefore, our approach proves to outperform the rest of the360 configurations with an alpha confidence level of 99%. 4.3. Analysis The objective of this section is to analyze some of the characteristics of the SAE for the task of staff-line removal. The quantitative results obtained by our method are very close to the op-365 timum, especially for binarized images. Thus, it is interesting to look into the actual impact of such room for improvement. Figure 7 shows a qualitative exam- ple of the performance obtained by the SAE for a piece of image in both binary and grayscale formats. These examples depict an F-m of 99.06 % and 98.17 18 %, respectively, therefore being good representatives of the results reported in370 the previous section. As can be observed, the differences between the outputs and the ground-truth are hardly perceptible. The mistakes might be related to pixels very close to the boundaries of symbols, which in any case does not seem to make a noticeable difference on the results. (a) Binary format (b) Grayscale format (c) Ground-truth (d) Selection for the binary format (e) Selection for the grayscale format Figure 7: Example of a staff-line removal using the proposed SAE configuration for both binary and grayscale formats, depicting an F-m of 99.06 % and 98.17 %, respectively. At this point, it would be interesting to measure the impact of the accuracy of375 this staff-line removal in a functional OMR system, with respect to the accuracy obtained by considering previous works. Note that in musical notation it is important to be as precise as possible in this process: regions of FP (staff- line segments maintained) may involve the incorrect detection of small musical 19 symbols such as the dot, or changing the meaning of some notes (quarter note380 confused with an eighth note); similarly, regions of FN (symbol information removed) may make the system mislabel a quarter note as a half note. Taking into account that music notation must fulfill strong grammatical constraints, these hypothetical isolated mistakes may cause many errors in other regions of the score. Unfortunately, there are no reproducible strategies for complete OMR385 workflow, and so carrying out this experiment would imply a study beyond the scope of this paper. However, we place this issue at future work considerations. Furthermore, having demonstrated the accuracy of the approach, it is impor- tant to stress its principal differences from conventional methods for staff-line removal. As stated above, one interesting property is that the SAE follows a su-390 pervised learning approach, so its extension as regards dealing with any type of input image is feasible. This applies not only in the case of a different nature of the image —such as RGB images— but also in that of different types of features of the score, different notational styles, heterogeneous document conditions and so on. It is true, however, that the approach cannot be immediately applied to395 any domain but our claim is that it is much easier to set up the SAE environ- ment (basically, training data and some parameter tweaking) than developing a completely new staff-line removal strategy. Pixel method is also based on supervised learning but there are two im- portant differences. The first is that our method achieves significatively better400 results, while the second, which is indeed the most relevant from a practical point of view, is that Pixel method is a computationally expensive technique, since it has to classify every single pixel in the image. In this respect, once the SAE has been trained, the computation time needed to process a score is in the order of seconds, while that of Pixel method is in the order of hours.405 In the following sections, we delve into interesting aspects of the proposed model. We first study the influence of the threshold used to convert SAE predic- tions into the binary selectional output. An incremental learning scenario is also considered to check the number of samples needed for the model to learn the task. Finally, we show the representational capabilities of the SAE by analyzing410 20 the intermediate codification of the input patch. 4.3.1. Threshold As stated in Section 3, the SAE predicts a value for each input feature, which can be understood as the level of selection. In this respect, for the results shown above it was assumed that a threshold equal to 0.3 (binary images) and415 0.1 (grayscale images) would discriminate between whether or not the feature was eventually activated. Here we report the reason why these values were chosen and the real impact that other configurations have. Figure 8 shows the F-m obtained for different threshold values — within the range of [0, 1] — considering the binary and grayscale format of the images. The420 values obtained by the best state-of-the-art algorithms have also been included for the sake of comparison. It can be observed that, although the thresholds considered are effectively those that obtain the best performances, any other threshold would have outper- formed the state-of-the-art strategies with a fair margin. These figures reinforce425 the robustness of the activations predicted by the SAE. These values tend to be close to either 0 or 1, thereby decreasing the importance of the threshold. 4.3.2. Training set size We shall now focus on assessing the impact of the amount of data used to train the model on the performance figures. To this end, an additional430 experiment was performed in which the number of training scores was iteratively increased. The results of this study are shown in Fig. 9. In the case of binary images, in which the variability is much lower, the SAE is able to perform very com- petitively with a relatively small number of scores. The model outperforms the435 state-of-the-art result with 50 scores. On the contrary, the case of grayscale images is more complex, and more samples are needed to reach stable results. Note, however, that the state-of-the-art is reached with a few number of training scores. 21 90 92 94 96 98 100 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 F -m Threshold SAE LRDE Pixel (a) Binary format 80 85 90 95 100 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 F -m Threshold SAE LRDE Pixel (b) Grayscale format Figure 8: Influence of the threshold parameter on the performance of the SAE. Performance of state-of-the-art strategies are also included for a better comparison. Furthermore, it should be emphasized that, in both cases, the curve shows440 an upward tendency when reaching the total number of training scores. It seems that a greater number of training samples could still improve the results obtained, especially in the case of grayscale. 4.3.3. Intermediate representation As a final analysis, this section studies the type of intermediate representa-445 tion that the SAE encodes. In the case of the model that we have determined is the most appropriate for this task (see Section 4.1), the intermediate repre- sentation comprises 96 × 32 × 32 features. In other words, it codifies 96 images of 32 × 32. A full example of intermediate representation can be seen in Fig. 10. In450 22 90 92 94 96 98 100 0 500 1000 1500 2000 2500 3000 3500 4000 F -m Training size SAE LRDE Pixel (a) Binary format 80 85 90 95 100 0 500 1000 1500 2000 2500 3000 3500 4000 F -m Training size SAE LRDE Pixel (b) Grayscale format Figure 9: Performance of the SAE with regard to the amount of training scores used to train. Performance of state-of-the-art strategies are also included for a better comparison. our grid search, which was carried out to determine the best model parameters, the number of intermediate codes did not lead to an excessive variation in the performance figures, which is why many of the intermediate representations are quite similar: there is a high redundancy of information in the case of 96 codes, signifying that similar results could be obtained with a fewer number of them.455 The actual utility of all the codes might only arise in the most complex cases. Furthermore, in order to compare the difference between binary and grayscale input images, Table 5 shows examples of the intermediate codifications obtained from both formats (enlarged for visualization). It is interesting to note that many of these intermediate representations are also quite similar, regardless of460 the type of input image. This could evidence that the SAE is actually learning a good internal representation to perform this task, discarding the information 23 (a) Input patch (b) Intermediate representations Figure 10: Illustration of the 96 intermediate representations for the depicted input and output patches. that refers to the specific characteristics of the input image. 4.4. Experiment with old documents A new experiment is described in this section with the aim at verifying the465 adaptability of the approach model to a different type of document images. It is obvious that, in this case, data-driven strategies have advantages because they are provided with information of the specific domain on which they are going to be applied. Traditional methods for staff-line removal have not taken into account the great heterogeneity that can be found in musical documents, thus470 leading to solutions that are not generalizable. We want to show here that the SAE entails a more adaptable approach, and it also behaves better than other supervised learning approaches. For this experiment we use a set of 20 staff sections from sacred music com- posed during the second half of the XVIIth century. These compositions were475 handwritten in a music book by a copyist of that time. In addition, they do not depict common modern notation, but the so-called mensural notation. Since 24 Input Intermediate representations (enlarged) Output Table 5: Example of intermediate representation of the SAE. music was intended to be sung, the sections depict both music notation and ac- companying text (lyrics), which might hinder the staff-line removal algorithms. We have manually created a binary version of the considered staff sections,480 which are originally depicted in grayscale, as well as their corresponding ground- truth (without staff lines). An example of this corpus is illustrated in Table 6. Source Binary Ground truth Table 6: Examples of the corpus created for staff-line removal on mensural notation manuscripts. Unlike the previous case, for which there exist a large dataset with thousands of images to train the models, this experiment shows a more real scenario. In 25 such a situation, it cannot be assumed that a large corpus can be effortlessly485 obtained, but the network must learn with a limited amount of examples. How- ever, we resort to fine-tuning to alleviate this situation. That is, a pre-trained model for other types of scores (ie. that obtained in the previous experiment) is initially considered, and then examples of the new domain are used to re- estimate the parameters of the model. In addition, the extracted patches from490 the training images are not totally disjoint but overlapping is considered in order to have a larger number of training examples from the same number of labeled documents. In this experiment, we first focus on selecting the most appropriate strategy to approach a new domain (mensural notation). The first option is to directly495 use the pre-trained network obtained from the previous experiment with modern notation. However, we also measure the performance obtained assuming that we have a limited number of tagged data of the new domain. Thus, we include in the comparison the performance of a trained network from scratch with these new data, as well as starting from the pre-trained network and performing a500 fine tuning process. To take the most out of the available data, we have established a 5-fold cross- validation scheme. That is, for each fold, 16 images are considered for training (20% of which are used to monitor the training and prevent over-fitting) and 4 images are used to measure the performance. The average results of this505 experiment can be checked in Table 7. In this case we can observe relevant differences between binary and grayscale formats. In the formed, the network trained with other types of scores is quite reliable—even better than training it from scratch with the new data—as the staff lines are similar to those depicted in the previous corpora. However, it is510 appreciated that adding data from the new domain allows boosting the recog- nition. In the grayscale format, it is observed that training with domain data, even with a limited number of examples, is more beneficial than directly consider the pre-trained network with data of another type of score. The difference with the binary case is that here the model has to deal also with the background of515 26 Method Binary Grayscale Pre-trained 96.16 84.05 Trained from scratch 95.05 91.45 Pre-trained + fine-tuning 97.98 95.71 Table 7: F-m (%) comparison among the different strategies for training the SAE in a new domain (mensural notation). Results report average performance considering a 5-fold cross validation experiment in both binary and grayscale format. Values in bold represent the best average accuracy in each set. the score, which presents relevant differences with respect to the background of the previous corpus. Nevertheless, it is also reported that the best performance is achieved when combining a pre-trained network with the new data. Finally, we compare the best case obtained by our model (that is, starting with a pre-trained model and fine-tuning) with the algorithms previously pro-520 posed. In particular, we consider the 3 methods that obtained the best result in the benchmark established by the aforementioned staff-line removal contest: LRDE, INESC and Pixel. The first two were specially designed for binary im- ages, and so we only consider their most favorable case. Furthermore, Pixel is easily usable in both contexts. Thus, the results of this comparative study can525 be consulted in Table 8. It can be observed that the SAE is able to outperform the results obtained with other approaches, even with a greater difference than in the previous ex- periment, in both binary and grayscale. Although the supervised approach is clearly an advantage in this context, it is also reported that our approach no-530 ticeably improves the results obtained by Pixel. An illustrative detail of these results is given in Fig. 11. A statistical significance test is performed again, considering each complete document as an independent sample. These tests resulted in p-values below 0.01 when comparing our method against the rest of the strategies, implying535 therefore that our method significantly improves their performance with an 27 Method Binary Grayscale LRDE 92.81 — INESC 90.89 — Pixel 91.24 86.64 SAE 97.98 95.71 Table 8: F-m (%) comparison among LRDE, INESC, Pixel, and SAE methods for the staff-line removal over mensural notation manuscripts. Results report average performance considering a 5-fold cross validation experiment in both binary and grayscale format. Values in bold represent the best average accuracy in each set. A dash mark (—) is used when the method in the row is not applicable. alpha confidence level of 99%. 5. Conclusions This work has studied the removal of staff lines from musical scores by considering a machine learning approach. Our proposal consists of using a540 new type of model called the Selectional Auto-Encoder (SAE), a convolutional neural network that learns which characteristics of the input should be selected by means of coding and decoding stages. In the context of the problem to be addressed, the model is trained to select only those pixels of the input that belong to a musical symbol. The process of removing the staff lines can therefore545 be solved, once the model is trained appropriately, by iterating the image of a score patch by patch. An activation in the range of [0, 1] is obtained for each pixel, indicating the level of selection predicted. Those pixels whose activation surpass a certain threshold are considered to be part of a musical symbol, and are otherwise discarded.550 Our comprehensive experimentation on a standard dataset has demonstrated the goodness and robustness of the approach. Regardless of the specific features of the chosen model — such as the size of the input window or the depth of the network — the results obtained were competitive. The model that obtained the 28 (a) Source (b) Binary (c) Ground-truth (d) INESC (e) LRDE (f) Pixel (bin) (g) Pixel (gr) (h) SAE (bin) (i) SAE (gr) Figure 11: Qualitative detail of the performance achieved by the different staff-line removal strategies for mensural notation from old manuscripts. Grayscale format (source), manual binarization, and ground-truth are also included for a better illustration. best figures was specifically that which takes a window of size 256×256, with 3555 convolutional layers per stage, 96 filters per layer and a 5×5 convolution kernel per filter. When compared to other algorithms proposed for the same task, the SAE performs significatively better for both binary and gray input images. The goodness of our approach is more evident in the latter case (up to more than 10 points of improvement in F-m).560 We have also studied the behavior of our model with regard to the chosen threshold, showing that this value does not have a particular impact on the results. However, it has been determined that a value of 0.3 is the most ap- propriate for dealing with binary inputs, and 0.1 for grayscale. An incremental study has also been carried out, and we have observed that the model attains565 a competitive performance with few samples. However, in the case of grayscale 29 images, the best values are only achieved with a large number of examples. Finally, we have shown the intermediate representations that the SAE learns, regarding which two main observations can be made: these representations have a high redundancy of information, which could mean that the whole capacity570 of the model is only needed in the most complex cases; the intermediate rep- resentations for the binary and grayscale cases are quite similar, which could indicate that the SAE is actually learning how to attain a deep representation of the task, independently of the characteristics of the input image. As prospects for future work, the intention is twofold. On the one hand, we575 intend to include our staff-line removal process in a functional OMR system so as to study its impact in a goal-directed way. We are especially interested in measuring the final performance of the system as regards the staff-line removal accuracy. On the other hand, we are interested in learning the task with a limited number of labeled samples. Note that the reference corpus for the prob-580 lem of staff-line removal is expensive to obtain. The idea is, therefore, to follow semi-supervised approaches in which the task can be learned in an unsupervised manner and a fine-tuning process with a small labeled corpus is carried out. It would be of great interest to carry out studies by increasing the variability of the input images. Musical scores can be highly heterogeneous, and we thus wish585 to obtain a model that can adapt to any type of them. Our idea is, therefore, to develop models that can process scores of a very different style from those seen during training by means of Transfer Learning (Pan & Yang, 2010) or Domain Adaptation (Patel et al., 2015) strategies. Acknowledgments590 This work was partially supported by the Spanish Ministerio de Educación, Cultura y Deporte through a FPU fellowship (AP2012-0939) and the Span- ish Ministerio de Economı́a y Competitividad through Project TIMuL (No. TIN2013-48152-C2-1-R, supported by UE FEDER funds). 30 References595 Bainbridge, D., & Bell, T. C. (1997). Dealing with superimposed objects in optical music recognition. In Proceedings of the 6th International Conference on Image Processing and Its Applications (pp. 756–760). Bengio, Y., Yao, L., Alain, G., & Vincent, P. (2013). Generalized denoising auto- encoders as generative models. In Advances in Neural Information Processing600 Systems (pp. 899–907). Bottou, L. (2010). Large-scale machine learning with stochastic gradient de- scent. In Proceedings of COMPSTAT’2010 (pp. 177–186). Springer. Calvo-Zaragoza, J., Barbancho, I., Tardón, L. J., & Barbancho, A. M. (2015). Avoiding staff removal stage in optical music recognition: application to scores605 written in white mensural notation. Pattern Anal. Appl., 18 , 933–943. Calvo-Zaragoza, J., Micó, L., & Oncina, J. (2016). Music staff removal with supervised pixel classification. International Journal on Document Analysis and Recognition, 19 , 211–219. Carter, N., & Bacon, R. (1992). Automatic recognition of printed music. In610 H. Baird, H. Bunke, & K. Yamamot (Eds.), Structured Document Image Anal- ysis (pp. 454–465). Springer. Choudhury, G. S., Droetboom, M., DiLauro, T., Fujinaga, I., & Harrington, B. (2000). Optical music recognition system within a large-scale digitization project. In ISMIR 2000, 1st International Symposium on Music Information615 Retrieval, Plymouth, Massachusetts, USA, October 23-25, 2000, Proceedings . Dalitz, C., Droettboom, M., Pranzas, B., & Fujinaga, I. (2008). A comparative study of staff removal algorithms. IEEE Trans. Pattern Anal. Mach. Intell., 30 , 753–766. Demsar, J. (2006). Statistical comparisons of classifiers over multiple data sets.620 Journal of Machine Learning Research, 7 , 1–30. 31 Deng, L., Seltzer, M. L., Yu, D., Acero, A., Mohamed, A., & Hinton, G. E. (2010). Binary coding of speech spectrograms using a deep auto-encoder. In INTERSPEECH 2010, 11th Annual Conference of the International Speech Communication Association, Makuhari, Chiba, Japan, September 26-30, 2010625 (pp. 1692–1695). Dos Santos Cardoso, J., Capela, A., Rebelo, A., Guedes, C., & Pinto da Costa, J. (2009). Staff Detection with Stable Paths. IEEE Trans. Pattern Anal. Mach. Intell., 31 , 1134–1139. Dutta, A., Pal, U., Fornes, A., & Llados, J. (2010). An Efficient Staff Removal630 Approach from Printed Musical Documents. In 2010 20th International Con- ference on Pattern Recognition (ICPR) (pp. 1965–1968). Fornés, A., Kieu, V. C., Visani, M., Journet, N., & Dutta, A. (2013). The IC- DAR/GREC 2013 Music Scores Competition: Staff Removal. In 10th Inter- national Workshop on Graphics Recognition, Current Trends and Challenges635 GREC 2013, Bethlehem, PA, USA, August 20-21, 2013, Revised Selected Pa- pers (pp. 207–220). Fujinaga, I. (2005). Staff detection and removal. In S. George (Ed.), Visual Perception of Music Notation (pp. 1–39). Hershey, PA: Idea Group Inc. Fujinaga, I., Hankinson, A., & Cumming, J. E. (2014). Introduction to SIMSSA640 (single interface for music score searching and analysis). In Proceedings of the 1st International Workshop on Digital Libraries for Musicology, DLfM@JCDL 2014, London, United Kingdom, September 12, 2014 (pp. 1–3). Géraud, T. (2014). A Morphological Method for Music Score Staff Removal. In Proceedings of the 21st International Conference on Image Processing (ICIP)645 (pp. 2599–2603). Paris, France. Hankinson, A., Burgoyne, J. A., Vigliensoni, G., & Fujinaga, I. (2012). Creating a large-scale searchable digital collection from printed music materials. In 32 Proceedings of the 21st World Wide Web Conference, WWW 2012, Lyon, France, April 16-20, 2012 (Companion Volume) (pp. 903–908).650 Hinton, G. E., & Zemel, R. S. (1994). Autoencoders, minimum description length and helmholtz free energy. In Advances in Neural Information Pro- cessing Systems (pp. 3–10). Lauly, S., Larochelle, H., Khapra, M., Ravindran, B., Raykar, V. C., & Saha, A. (2014). An autoencoder approach to learning bilingual word representations.655 In Advances in Neural Information Processing Systems (pp. 1853–1861). LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521 , 436–444. Martin, P., & Bellissant, C. (1991). Low-level analysis of music drawing images. In First International Conference on Document Analysis and Recognition (pp.660 417–425). Ng, K. (2001). Music manuscript tracing. In International Workshop on Graph- ics Recognition (pp. 330–342). Springer. Ntirogiannis, K., Gatos, B., & Pratikakis, I. (2014). ICFHR2014 competition on handwritten document image binarization (H-DIBCO 2014). In 14th Inter-665 national Conference on Frontiers in Handwriting Recognition, ICFHR 2014, Crete, Greece, September 1-4, 2014 (pp. 809–813). Pan, S. J., & Yang, Q. (2010). A survey on transfer learning. IEEE Transactions on knowledge and data engineering , 22 , 1345–1359. Patel, V. M., Gopalan, R., Li, R., & Chellappa, R. (2015). Visual Domain670 Adaptation: A survey of recent advances. IEEE Signal Processing Magazine, 32 , 53–69. Ramirez, C., & Ohya, J. (2014). Automatic recognition of square notation symbols in western plainchant manuscripts. Journal of New Music Research, 43 , 390–399.675 33 Randriamahefa, R., Cocquerez, J. P., Fluhr, C., Pepin, F., & Philipp, S. (1993). Printed music recognition. In Document Analysis and Recognition, 1993., Proceedings of the Second International Conference on (pp. 898–901). IEEE. Raphael, C., & Wang, J. (2011). New approaches to optical music recognition. In Proceedings of the 12th International Society for Music Information Retrieval680 Conference, ISMIR 2011, Miami, Florida, USA, October 24-28, 2011 (pp. 305–310). Rebelo, A., Capela, G., & Cardoso, J. S. (2010). Optical recognition of music symbols - A comparative study. International Journal on Document Analysis and Recognition, 13 , 19–31.685 Rebelo, A., & Cardoso, J. (2013). Staff Line Detection and Removal in the Grayscale Domain. In 2013 12th International Conference on Document Anal- ysis and Recognition (ICDAR) (pp. 57–61). Rebelo, A., Fujinaga, I., Paszkiewicz, F., Marçal, A. R. S., Guedes, C., & Cardoso, J. S. (2012). Optical music recognition: state-of-the-art and open690 issues. International Journal of Multimedia Information Retrieval , 1 , 173– 190. Roach, J., & Tatem, J. (1988). Using domain knowledge in low-level visual pro- cessing to interpret handwritten music: an experiment. Pattern recognition, 21 , 33–44.695 Su, B., Lu, S., Pal, U., & Tan, C. (2012). An Effective Staff Detection and Removal Technique for Musical Documents. In 2012 10th IAPR International Workshop on Document Analysis Systems (DAS) (pp. 160–164). Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., & Manzagol, P.-A. (2010). Stacked denoising autoencoders: Learning useful representations in a deep700 network with a local denoising criterion. Journal of Machine Learning Re- search, 11 , 3371–3408. 34 Visaniy, M., Kieu, V., Fornes, A., & Journet, N. (2013). ICDAR 2013 Music Scores Competition: Staff Removal. In 2013 12th International Conference on Document Analysis and Recognition (ICDAR) (pp. 1407–1411).705 Wang, W., Huang, Y., Wang, Y., & Wang, L. (2014). Generalized autoencoder: A neural network framework for dimensionality reduction. In The IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR) Workshops (pp. 490–497). Wen, C., Rebelo, A., Zhang, J., & Cardoso, J. S. (2015). A new optical music710 recognition system based on combined neural network. Pattern Recognition Letters , 58 , 1–7. Zeiler, M. D. (2012). ADADELTA: an adaptive learning rate method. CoRR, abs/1212.5701 . 35 Introduction Background Staff-line detection and removal Auto-encoders Staff-line removal with Selectional Auto-Encoders Experiments Hyper-parameter selection Comparison with state-of-the-art Analysis Threshold Training set size Intermediate representation Experiment with old documents Conclusions 
work_minrsa6e45eqtd5mwmrhjxf2hq	----	PII: 0950-7051(88)90071-8 Socrates: a flexible toolkit for building logic-based expert systems R Corlett*, N Davies*, R Khan*, H Reichgeltt and F van Harmelent This paper describes the architecture o f the Socrates tool- kit ./'or building expert systems. The authors analyse the problems associated with existing expert system tools and propose a solution based on the use o f logic and meta-level inference. The abstract architecture f o r the toolkit is described which embodies this combination o f logic and meta-level inference. This architecture can be instantiated to create a system that is specialized for a particular application. This specialization process can be seen as a methodology f o r building expert systems. The three stages o f this methodology are discussed in detail, along with descriptions o f how the Socrates toolkit supports it. The current implementation o f Socrates, plus a number o f applications o f the toolkit are described, and the open problems are discussed. Keywords: expert systems, toolkit, architecture, building methodology Most o f the tools that are currently available for constructing expert systems fall into two categories, namely expert system shells and high level programming language environments. Both o f these types o f tool suffer from a number o f shortcomings which limit their useful- ness. The first type o f available tools, 'shells', are usually constructed by abstraction from an existing expert system. Thus, a shell normally consists o f an inference engine and an empty knowledge base, and some de- bugging and explanation facilities. Buyers o f a shell often believe, and manufacturers often claim, that the shell is appropriate for a number o f different applications. However, a large number o f people have expressed dis- satisfaction with expert system shells. The two most frequently heard complaints (e.g. A l v e y ' ) are first, that the view that the inference engine which was successful *GEC Research, Marconi Research Centre, Chelmsford, UK tDepartment of Artificial Intelligence, University of Edinburgh, Edinburgh EH8 9YL, UK 0950-7051/88/030132-11 $03.00 © 1988 132 in one application will also be successful in other applications is unwarranted, and second, that the knowledge representation scheme often makes the expression o f knowledge in another d o m a i n awkward, if not impossible. F o r example, a production rule language that was designed for solving a classification problem using backward chaining would probably not be very suitable for solving a design or planning problem. On the other hand, proponents o f high level program- ming language environments (sometimes known as 'hybrid systems'), such as LOOPS 2, K E E 3 or A R T 4, can be seen as taking a more pluralistic position: instead o f providing knowledge engineers with a single pre- fabricated inference engine, one provides them with a large set o f tools, each o f which has proven useful in other applications. LOOPS, for example, provides object-oriented programming with inheritance, a pro- duction rule interpreter and active values, as well as Lisp. While we accept that hybrid systems are useful as tools for program development, we would claim that they are less useful as tools for building expert systems. Their main problem is that they provide the knowledge engineer with a bewildering array o f possibilities, and little, if any, guidance as to the circumstances in which any o f these possibilities should be used. Unless used by experienced programmers, high level programming environments encourage an ad hoc programming style in which no attention is paid to a principled analysis o f the problem at hand to see which strategy is best suited for its solution. The conclusion that is drawn from the problems associated with shells and high level programming language environments is that a number o f different "models o f rationality' are needed, and that different applications require different models o f rationality. When constructing an expert system, the knowledge engineer then has to decide which model o f rationality is appropriate to the application at hand. A model o f rationality has both a 'static' and a 'dynamic' aspect. These two aspects correspond to the knowledge about Butterworth & Co (Publishers) Ltd Knowledge-Based Systems the application area plus a strategy that describes how to use this knowledge when solving a problem. The 'interpretation models' from Breuker and Wielinga s correspond closely to our models o f rationality. The distinction between the static and the dynamic aspects of a model of rationality corresponds to a dis- tinction one can make between two different aspects of an expert system. First, there is the 'domain' in which the expert system is to solve problems. For example, the domain of an expert system may be electronics, or internal medicine. Secondly, there is the 'task' which the knowledge engineer wants the expert system to perform. For example, the task of a system can be diagnosing a faulty electronic circuit or designing a new circuit. It is interesting to note that the problems with shells reflect these two aspects of an expert system. The first complaint about shells concerning the expressive- ness of the knowledge representation language is related to the structure of the domain. The second complaint concerning the rigidity of the inference engine is related to the task of the expert system. As pointed out by Chandrasekaran ~, typical expert system tasks such as diagnosis, planning, monitoring, etc. seem to require particular control regimes. Socrates allows the use of a variety of logical languages and a variety of control regimes to solve the problem o f the lack of flexibility associated with shells. By using logic as its unifying framework, and by provid- ing guidelines for the choice of both the representation language and the control regime, Socrates avoids the unstructured richness of the hybrid systems. These points are discussed in detail below. U S I N G L O G I C F O R K N O W L E D G E R E P R E S E N T A T I O N Socrates uses a 'logical language' as the main formalism to implement the static aspects of the required models of rationality for different domains. On the one hand, logical formalisms are rich enough to provide different models of rationality, while on the other hand the use of logic provides a unifying framework for the system which saves it from the unstructured richness of the hybrid systems. This choice o f logic as the main formalism implies that logical languages will serve as the representational scheme, while logical deduction will be the paradigm for the inference engine. Many advantages come with the use of logic as the main knowledge representation formalism. Logic comes with a formal semantics, it has well understood proper- ties regarding completeness, soundness, decidability, etc., and it has great expressive power. For a further elaboration on these arguments, see Corlett, Davies, Khan, Reichgelt and van Harmelen 7. U S I N G M E T A - L E V E L I N F E R E N C E F O R C O N T R O L The correspondence between the two aspects of a model of rationality and the problems with shells suggests that a model of rationality should be computationally realized as a knowledge representation formalism plus a control regime for using this formalism. A number of arguments can be given for the explicit and separate representation of control knowledge. First, a system with I - . . . . . . . . . . . . U ~ r - f ' r o n ' t end . . . . . . . . . . . . . . "I L . . . . . . . . - I / A - , , [ 2 - _ - - - - 2 - - _- ; _ . . . . . . . . . . . . ] Figure 1. Socrates architecture an explicit representation of its control strategies is easier to develop, debug and modify, as argued by Davis s, Bundy and Welham 9, Clancey ~° and Aiello and Levi 11. Second, the separation of control knowledge from domain knowledge (or 'object-level knowledge') increases the reusability of the system: the same object- level knowledge can be used for different purposes, and the same control knowledge can be used in different domains s. ~ o. 12 Finally Clancey ~ 3 stresses the importance of explicit control knowledge for the purpose of explanation. The separation of control knowledge from domain knowledge allows domain knowledge to be purely declarative in nature. While formulating domain knowledge we do not have to worry about efficiency, only about the 'representational adequacy' (in the words of McCarthy). In the control knowledge, on the other hand, the efficiency of the problem solving process is the most prominent aspect ('computational adequacy'). A B S T R A C T A R C H I T E C T U R E O F S O C R A T E S As argued in many other places in the literature (see ~0.~2.~4-20 among others), an architecture with a separate object-level and meta-level interpreter can be used to implement the required separation of domain and control knowledge. This architecture is shown in Figure 1. The two-layer architecture of object-level and recta-level interpreter is extended in Socrates with a third level, the scheduler, discussed below. Each of the three interpreters communicates with the knowledge base in order to store and retrieve logical propositions. As described below, the knowledge base can be organized into a number of partitions, each of which can be hierarchically organized into subpartitions. The Socrates architecture distinguishes between front- ends to be used by the 'end-user' and the 'knowledge engineer'*. Knowledge engineers and end-users will need different tools for communication with the system. For example, an end-user, when asking for an explanation, will not want to see the entire proof tree, but rather 'edited highlights'. A knowledge engineer, on the other hand, may want to be able to monitor the reasoning *The end-user is the person who uses the expert system application built with the Socrates toolkit, whereas the knowledge engineer is the person building such an application system. Vol 1 No 3 June 1988 133 process much more closely and m a y require the full p r o o f tree. Similarly, a knowledge engineer will need tools for adjusting the b e h a v i o u r o f the interpreters and for edit- ing the knowledge base, whereas an end-user only needs to browse in a read-only m a n n e r t h r o u g h the knowledge base. F u r t h e r m o r e , the levels o f abstraction at which the system communicates will differ between end-user and knowledge engineer. This general, application independent architecture needs to be "configured' for a particular expert systems application, given the characteristics o f the d o m a i n and the task o f that application. The use o f logic as the underlying framework for the system gives a very specific meaning to the process o f configuring the system for a particular application. R a t h e r than having to choose between a r b i t r a r y representational schemes and inference m e t h o d s (as is the case in the hybrid systems), the configuration process now consists o f three well defined stages; i.e. the declaration o f a 'logical language', the declaration o f a ' p r o o f theory" for that logical language, and the declaration o f a ' p r o o f strategy'. L O G I C A L R E P R E S E N T A T I O N L A N G U A G E D E C L A R A T I O N As the first step in the process o f configuring a system for a particular application, the knowledge engineer defines the logical language that will be used for representing the d o m a i n knowledge. In the context o f a knowledge-based system, where the logical language is to be used for representational purposes, the 'syntactic" declaration o f the language has to be augmented with a mechanism for 'storage and retrieval'. Each o f these aspects will be discussed in this section. R e p r e s e n t a t i o n l a n g u a g e As the first part o f the declaration o f the logical represen- tation language, the knowledge engineer has to declare the "vocabulary' o f the language. T h e predicate symbols, function symbols and constants that are going to be part o f the representation language have to be defined. Socrates uses ' m a n y - s o r t e d logics' o f the kind pro- posed by Waither 2~, making use o f typed quantification, where the sorts are organized in a hierarchy. Each sort is used to represent a n o n - e m p t y set o f individuals belonging to that sort. In such logics the sort hierarchy is defined by a partial ordering on the set o f sorts, corres- ponding to a lattice structure with the universal sort as the maximal, and the e m p t y sort* as the minimal element in the set. This lattice is m o r e general than a tree structure, since it allows sorts to have more than one supersort, thereby increasing the expressiveness o f the sort o f system. T h e use o f m a n y - s o r t e d logics over unsorted logics brings a n u m b e r o f advantages. First, complicated unsorted expressions can be reformulated in a much simpler m a n y - s o r t e d form. F o r example, the unsorted expression: V x 3 y [human(x) ~ h u m a n ( y ) & mother(y, x)] * T h e e m p t y s o r t , n o t a t i o n ~,~, is t h e o n l y s o r t r e p r e s e n t i n g a n e m p t y set o f individuals, a n d a s s u c h d o e s n o t fulfill a n y r e p r e s e n t a t i o n a l role. T h e r e a s o n w h y s o r t s m u s t c o r r e s p o n d t o n o n - e m p t y sets is d i s c u s s e d below. t T h e n o t a t i o n x . t is used to indicate t h a t v a r i a b l e x is o f s o r t t. can be rewritten, after the declaration o f human as a sort, as V x: h u m a n 3 y: h u m a n [mother(y, x)]'t This reduction in complexity o f axioms not only improves the readability o f the knowledge base, but it also causes a significant decrease in the size o f the search space for proofs. C o h n 2-" (in the context o f a resolution based theorem prover), and Davies 23 (in the context o f a natural d ed u ct i o n system) show that this reduction can a m o u n t to as much as an o r d e r o f magnitude. Second, the sort hierarchy allows the knowledge engineer to represent what Clancey 13 calls "structural knowledge', representing the taxonomical hierarchy o f the application domain. In sorted logics, the unification algorithm has to take the sort hierarchy into account. Th e rules for unification in a sorted logic are: a constant c,:t~ unifies with a co n st an t C2:/2 if c~ = c2 and ti = I 2 a variable x.'t~ unifies with a co n st an t c.'t 2 if t 2 = tt a variable x.'q unifies with a variable y:t2 if tj Th e third rule results in restricting the sorts o f both variables to t3 = t~ c~ t2. In o rd er to g u aran t ee that t 3 can be c o m p u t e d and is itself a legal sort, it is necessary that the system knows the values o f all pairwise intersec- tions o f all sorts, and that sort-intersection is closed o v er sorts. In o t h e r words, if T is the set o f all sorts (including ~ ) , then V t~. t=:tt • T & l 2 • T - , t~ c~ t2 • T F u r t h e r m o r e , in o r d e r to g u aran t ee that 13 is uniquely defined, we require that no two sorts are used to repre- sent the same set o f individuals. I f this were allowed, the unification algorithm would no longer be able to return a unique most general unifier. Th e knowledge engineer declares the sorts o f the logic by declaration first the set o f sorts, and then the hierarchy o f sorts using set-theoretic primitives, such as equality, subset, superset, intersection, union, set- difference, c o m p l e m e n t a t i o n an d disjointness. After this has been done, the system automatically checks whether the sort hierarchy satisfies the following criteria: • No two sorts are equal. • N o sort except .~,, is empty. • Intersection is closed o v er sorts. If an y o f these constraints is violated, or if the constraints are not satisfied by the current declarations (because the knowledge engineer has underspecified the sort hierarchy), the system asks the knowledge engineer for different o r additional declarations. T h e sorts o f the constants o f the logical language can be declared in two ways. T h e simplest way is by simply en u m erat i n g the constants o f a particular sort ('extensional' definition o f types). H o w ev er, for some sorts e n u m e r a t i o n is not feasible, either because the set o f constants o f that sort is not k n o w n in advance, or because this set is t o o large, o r indeed infinite. F o r these cases it is possible to define constants by declaring a "recognition procedure" for the sort. Every object for which the recognition p r o c e d u r e succeeds will be assumed to be a co n st an t o f t h at particular sort ('inten- sionai' definition o f types). T h e definitions o f the 134 Knowledge-Based Systems recognition procedures must be supplied in the imple- mentation language of the system. This allows an inter- face between the logical representation language and the computational data types supplied by the implemen- tation language o f the system. An example of a sort that can be declared via this mechanism is the sort of all integers, or as a second example, for any given sort T, the sort T L i s t consisting of all lists of elements of sort T. As part of declaring the vocabulary of the logical representation language, it is possible to exploit what Weyhrauch 24 calls 'semantic attachment'. This is done by declaring a special class of predicates called 'evaluable predicates'. When one of these predicates is encountered in a proof, it is possible (depending on the control decisions made by the meta-level interpreter, discussed below) to execute a procedure defined for this predicate to determine its truth value, and possibly provide bindings for any variables. These predicates provide an interface between the logical representation language and the computational environment of the system, enabling external systems interaction, input/output for interacting with the end-user, and the access of the facili- ties provided by the implementation language of the system. At this stage the representation language consists only of a sort hierarchy and a set of predicates, constants and functions. We still need to extend our language with 'logical connectives', such as implication, dis- junction, etc., and possibly non-standard operators*, such as modal and temporal operators. As part of the declaration of these logical connectives the knowledge engineer can declare their properties with respect to commutativity and associativity. These declarations will be used to configure a unifier for the defined logical language (see the section below on the declaration of the proof theory for a more detailed discussion of this subject). Socrates distinguishes between different kinds of logical connectives: 'Unary' connectives take one argument; binary connectives take two arguments, and can be divided on the basis o f their commutativity. Implication ( ~ ) is an example of a 'non-commutative binary' connective, whereas equivalence (4--,) is an example of a 'commutative binary' connective. If con- nectives are both commutative and associative (such as, for instance, conjunction), they are treated as so-called 'set-connectives'. This amounts to them taking any number o f arguments in any order. Furthermore, set connectives are assumed to be idempotent: all multiple occurrences of an argument can be reduced to only one occurrence. Other possible connectives are not currently implemented in Socrates: 'bag' operators, which are associative and commutative but not idempotent (i.e. the order of arguments does not matter, but multiple occurrences o f an argument cannot be reduced), and 'sequence' operators, where the order of the arguments is significant, and multiple occurrences cannot be reduced. S t o r a g e and retrieval mechanism Now that the vocabulary of the logical language is complete we need to declare a storage and retrieval *The terms logical connective and logical operator are used synonymously mechanism for expressions in the language. This is what is typically called the 'knowledge base' of the system. The knowledge base of Socrates is not a flat space of assertions in the declared logical language, but can be divided into a number of 'partitions'. Each o f these partitions can have a separate logical language asso- ciated with it (but see below for a restriction on this). This facility serves a number of different purposes. First, it allows the knowledge engineer to use mixed language representations of the domain. Different types of knowledge or different aspects of the domain can be represented in separate languages. A particular example of this is described below, where we discuss the control knowledge of the system. This meta-level knowledge is expressed in a different language from the object-level knowledge. Nevertheless, it can be stored in the same knowledge base, using the partitioning mechanism to separate the two. Second, the partition hierarchy can be used to reduce the search that needs to be done both by the retrieval mechanism and by the inference machinery. In many problem solving situations, only a subset of all the avail- able knowledge is applicable at any one time to a given problem. Partitioning allows useful subsets to be applied while others are ignored. In this way, Socrates can be used to model a blackboard architecture, with each of the partitions simulating the contents of a knowledge s o u r c e . An important aspect of the partitions in the knowledge base is that they can be recursively divided into sub- partitions, and that these subpartitions are organized hierarchically. Partitions lower down in this hierarchy inherit all the propositions from partitions higher up in the hierarchy, but not vice versa. A particular example of this could be the use of a single working memory for a number of different subsets of domain knowledge, where the working memory would be represented in the top node, and the subsets of domain knowledge repre- sented in subpartitions. The results o f reasoning done within one subpartition can in this way be communicated to the other partitions via the working memory. Again, this feature can also be used in the modelling of a black- board system in Socrates, since the blackboard needs to be visible from all knowledge sources, but not vice versa. Since partitions can be created dynamically at runtime, another application for the partition hierarchy in the knowledge base is hypothetical reasoning, some- times known as 'what-if' reasoning. Each time a new hypothesis is generated, a new subpartition can be created containing this hypothesis, and inheriting all existing propositions from higher partitions. The inheritance mechanism puts restraints on the logical languages that can be used within subpartitions. Because subpartitions inherit propositions from super- partitions, all the partitions in a partition hierarchy must be associated with the same logical representation language. (Strictly speaking, it is only necessary that the language of a subpartition is an extension of the language of its superpartition, but Socrates does not support this type of language inheritance.) Thus, the knowledge base can be seen as a set of trees of partitions. Within each tree, all partitions must have the same language, but each separate tree can be associated with a different language. A final facility supported by the knowledge base is Voi 1 N o 3 June 1988 135 the a n n o t a t i o n o f p r o p o s i t i o n s . Each p r o p o s i t i o n can be a n n o t a t e d with a r b i t r a r y i n f o r m a t i o n using a slot- value m e c h a n i s m . This allows us to associate extra- logical i n f o r m a t i o n with propositions. This i n f o r m a t i o n can be used for a wide range o f purposes, o f which a few e x a m p l e s are: n a t u r a l language descriptions to be used in explanations, c o n t r o l i n f o r m a t i o n to be used by the meta-level interpreter, c e r t a i n t y values, etc. This slot-value m e c h a n i s m could be used to model a truth- m a i n t e n a n c e system in Socrates, by storing each derived p r o p o s i t i o n in the k n o w l e d g e base t o g e t h e r with a n n o t a - tions that c o n t a i n the premises t h a t were used in its derivation. Socrates uses a discrimination net technique for retrieval o f propositions. T h e o p t i m a l criteria that the net should use for discrimination d e p e n d on the parti- cular application, a n d Socrates therefore allows the knowledge engineer to adjust these discrimination criteria to suit the p a r t i c u l a r ways in which the k n o w - ledge base retrieval m e c h a n i s m will be used. F o r example, if the m a i n control regime for a p a r t i c u l a r application is s o m e f o r m o f b a c k w a r d chaining, then p r o p o s i t i o n s o f the f o r m "left-hand-side ~ r i g h t - h a n d - side' will be retrieved f r o m the knowledge base on the basis o f the p a t t e r n s in the ' r i g h t - h a n d - s i d e ' a r g u m e n t . This implies t h a t the discrimination net should first dis- c r i m i n a t e implications on the basis o f their consequent, r a t h e r than their antecedent. As a second example, we note that m a n y predicates use a certain n u m b e r o f their a r g u m e n t s as ' i n p u t s ' , while others are used as "output" a r g u m e n t s . A predicate such as ' s u f f e r s - f r o m (patient, disease)' will typically be used to associate a given patient with a disease, r a t h e r t h a n to try a n d find all patients suffering f r o m a given disease. This p a r t i c u l a r use o f a r g u m e n t s indicates t h a t the discrimination net should use the first a r g u m e n t as the m a i n discrimination criterion, r a t h e r t h a n the second a r g u m e n t . Because these p r o p e r t i e s are specific to a p a r t i c u l a r application, they can only be adjusted by the k n o w l e d g e engineer w h o configures the system, r a t h e r t h a n being hardwired into the retrieval m e c h a n i s m o f the knowledge base. PROOF THEORY DECLARATION T h e declaration o f a logical language t o g e t h e r with a storage and retrieval m e c h a n i s m allows us to represent knowledge, but in o r d e r to m a n i p u l a t e this k n o w l e d g e to derive new conclusions we need rules that tell us what the legal d e r i v a t i o n s will be. Since we have chosen logic as the r e p r e s e n t a t i o n language for the system, we are c o m m i t t e d to logical d e d u c t i o n as the inference process. W e therefore need to define a p r o o f theory, i.e. a set o f inference rules* t h a t tell us h o w logical p r o p o s i t i o n s can be m a n i p u l a t e d in o r d e r to p e r f o r m a p r o o f . In Socrates we follow Biedsoe 2s in using the natural d e d u c t i o n style o f p e r f o r m i n g p r o o f s r a t h e r than, for instance, resolution. A l t h o u g h n a t u r a l d e d u c t i o n systems use a relatively large n u m b e r o f inference rules (as o p p o s e d to the single inference rule o f resolution based systems), and thereby create a p o t e n t i a l c o n t r o l *We use the term "inference rule' in the logician's sense: an inference rule is a rule that describes how true formulas can be inferred from other true formulas, and is of the form . f t . • • f , ~- g , where g is inferred from J ~ . . . f , . The term 'inference rule' is often incorrectly used to denote formulas of the form .f--, g (i.e. logical implications). p r o b l e m , a n u m b e r o f reasons can bc given in f a v o u r o f n a t u r a l deduction. T h e inference rules o f a n a t u r a l d e d u c t i o n system are m o r e intuitively meaningful than, for instance, the resolution rule. F u r t h e r m o r e , no n o r m a l f o r m s are required for the f o r m u l a s used in a proof. As a result, the p r o o f s p e r f o r m e d by a n a t u r a l deduction system are easier to follow for a h u m a n reader, thereby i m p r o v i n g the possibilities for e x p l a n a t i o n facilities. T h e naturalness o f the p r o o f d e v e l o p m e n t also m a k e s it easier to identify heuristics to c o n t r o l the p r o b l e m solving pro- cess, in the m a n n e r discussed below. W h e n expressing inference rules, the k n o w l e d g e engineer can m a k e use o f extra-logical variables that range o v e r well f o r m e d f o r m u l a s f r o m the object-level representation language. F o r instance, the rules P , P ~ Q ~ Q P ~ - P V Q represents the rules f o r M o d u s Ponens a n d Disjunction I n t r o d u c t i o n . It is i m p o r t a n t to stress that these inference rules can be used in b o t h a f o r w a r d a n d a b a c k w a r d direction. F o r instance, M o d u s Ponens can be used to d e t e r m i n e t h a t P a n d P -~ Q have to be p r o v e d in o r d e r to p r o v e Q ( b a c k w a r d use), o r the rule c a n be used to infer that Q is true when we k n o w t h a t b o t h P and P --* Q are true. F u r t h e r m o r e , because o f the associativity and c o m m u t a t i v i t y o f s o m e o f the logical connectives, a single inference rule c a n often be applied in m o r e t h a n one way. F o r instance, if disjunction (V) has been declared as a ' s e t - o p e r a t o r ' , Disjunction I n t r o - d u c t i o n can be applied b a c k w a r d to (/'V g) in t w o differ- ent ways, binding P to e i t h e r . f or g, thereby generating either f o r g as a sub-goal for p r o v i n g 0"V g). H o w e v e r , which o f these possible a p p l i c a t i o n s o f an inference rule should be used is a control decision, a n d is therefore a meta-level issue, which is not decided as p a r t o f the p r o o f theory. N o t all the inference rules t h a t the system uses are declared as p a r t o f the p r o o f strategy. First o f all, there is a set o f inference rules t h a t tell the system h o w to deal with t y p e d quantification. These rules: V x:t~ P[x] ~ P[c:t2] for an a r b i t r a r y c o n s t a n t c, a n d all sorts t~ and t2 with t~ _~ t2 ~ ~' P[c:t~] ~ 3 X:tz P[x] for an a r b i t r a r y c o n s t a n t c, a n d all sorts t~ and t2 with t 2 ~ t~ D ~)* 3 x:tt V y:t2 P[x,y] I'-- V y:t2 q x:t~ P[x,y] for all sorts t~ a n d t2 are taken to be o f universal validity (that is: across differ- ent a p p l i c a t i o n areas o f the system), a n d are therefore hardwired into the retrieval m e c h a n i s m . A second set o f rules, t h a t is p a r t o f the retrieval m e c h a n i s m in the knowledge base r a t h e r t h a n the p r o o f theory, are the results that deal with the c o m m u t a t i v i t y a n d associativity o f certain logical connectives. F o r a n y o p e r a t o r ~ that has been declared as " b i n a r y - c o m m u t a t i v e ' , the inference rule P d p Q ~ Q ~ P is hardwired into the knowledge base retrieval mechanism, as are, for every ' s e t - o p e r a t o r ' , the additional rules P F - Pdp P (P dp (Q dp R)) I-- ((P ~ Q) dp R). *it is exactly because of this rule that in the section on logical represen- tation declaration we required all sorts except ~ to be non-empty. 136 K n o w l e d g e - B a s e d Systems By taking all these rules out o f the explicit declaration o f the p r o o f strategy and transferring them to the retrieval mechanism o f the knowledge base, we have given a limited deductive capability to the knowledge base. The retrieval mechanism comprises two sequential stages, a syntactic 'pattern matcher' and a 'unifier'. Syntactic pattern matching is effected using the discrimi- nation net technique mentioned above. Given a formula as a query to the knowledge base, this process will retrieve all formulas with the same pattern o f operators and predicates, subject to the commutativity and associa- tivity rules as declared for the particular language. This means that unification will be attempted only on patterns that have a strong possibility o f succeeding. The patterns that can be specified as input to the syntactic pattern maching phase are allowed to contain 'recta-logical (propositional) variables' that can match with formulas o f the logical representation language rather than with terms. These meta-logical variables will be bound to the corresponding components o f the matching expression, using pattern matching under the inference rules governing commutativity and associa- tivity. For example, the pattern (& f ?P) will match with a formula like (& g f ) , binding ?P with g after applying commutativity. To facilitate the retrieval o f expressions containing set-operators, a special version o f these meta-logical variables is available, the so-called 'segment-variables'. These segment variables do not match with formulas, but with lists o f logical formulas. For example, a formula like (& f g h) will match with a pattern like (& g ?P'segment), binding ?P with the list ( f h). Because o f the meta-logical variables, the patterns sent to the knowledge base for retrieval are actually schemata, representing whole families o f queries rather than just a single query. In the second phase o f the retrieval process the unifier will construct bindings o f the logical variables occurring in the query. This is necessary since in the context o f expert systems we are interested in performing construc- tive proofs, and we therefore need the values for the existentially quantified variables in the query for which the query succeeds. In other words, for a query such as 3 x:ttp(x), we want not only a yes/no response, but also the values o f x which can be deduced from the contents o f the knowledge base. Notice that these bind- ings might only consist o f restrictions on the sort o f x, rather than o f bindings o f x to terms. The sorted unification algorithm might tell us that the query succeeds for all values o f x o f a certain sort t2, with t2 c t,. In this way, taxonomic reasoning is performed at retrieval time. A final point to be made about the declaration o f the p r o o f theory concerns the soundness and complete- ness o f the set o f inference rules. In order to guarantee soundness o f the p r o o f theory, the knowledge engineer should not be allowed to declare arbitrary inference rules, but only to select inference rules from a predefined (and sound) set*. This selection process will o f course affect the completeness o f the system. However, the loss o f completeness in the context o f expert systems is not serious, since one does not want to infer a / / f a c t s that follow from the available knowledge, but only those facts that one is interested in. *Such a selection procedure has not been provided in the current implementation o f Socrates. P R O O F S T R A T E G Y D E C L A R A T I O N At this point, the knowledge engineer has declared both a logical language and a corresponding p r o o f theory. From a purely logical point o f view, no further declara- tions are necessary. The combination o f language and p r o o f theory determines all the possible inferences that the system can make. However, in order to create a practical computer system, one more step has to be taken. Proving statements in any non-trivial logical language is a search intensive problem. The logical language and p r o o f theory together define a search space for the p r o o f process. What remains to be done is the specification o f the strategy that the system should use to traverse this space while searching for a proof. For this task, Socrates provides a declarative language for representing such a control strategy, described below. Because such a declarative language has certain dis- advantages associated with it, a more procedural language has also been investigated (see below). Declarative representation of control knowledge Socrates allows the knowledge engineer to explicitly specify a control strategy. This control strategy provides the system with a description o f its desired behaviour, and is interpreted at run time by the meta-level inter- preter. As a result, the meta-level interpreter executes this control strategy, and thereby guides the search through the space o f all possible proofs. The language that is used to express the control strategy is a many sorted version o f Horn Clause Logic. This language, although also a logical representation language, should be distinguished from the logical languages used to represent the domain knowledge. Unlike the object-level languages, the language used at the meta-level has a fixed set o f logical connectives, namely exactly those connectives needed in Horn Clause Logic: conjunction, implication and negation, plus disjunction. All these connectives are declared as non-commutative, non- associative. This is done because the procedural interpre- tation (i.e. the way in which the meta-level interpreter executes expressions o f the language) is also fixed. The procedural interpretation o f the language is the standard interpretation for Horn Clauses, the standard depth-first p r o o f procedure as found in Prolog systems. The reason why Socrates does not allow the knowledge engineer to change the control regime o f the meta-level interpreter (which would amount to providing a meta-meta-level interpreter~f) follows from the analysis o f models o f rationality, sketched above. As indicated, typical expert system tasks such as diagnosis, planning, monitoring, etc. are related to particular control regimes. The meta- level controls the behaviour o f the object-level interpreter according to the expert system task. The variation in control is achieved by changing the meta- level knowledge base. There is no need to change the interpreter, which always has the same task, namely con- trolling the behaviour o f the object-level interpreter by using the data in the meta-level knowledge base. The parts o f the meta-level language that are still subject to declarations made by the knowledge engineer tThe notion of a meta-meta-level interpreter should not be confused with the scheduler; the relation between meta-level interpreter and scheduler is very different from the relation between object-level and meta-level interpreter. Voi I N o 3 June 1988 137 are therefore the set o f constants, predicates and function symbols, the sort hierarchy, and the set o f "evaluable predicates'. Figure 2 shows an example o f a description o f a local-best-first non-exhaustive backward chaining control strategy. F o r this control strategy a sort hierarchy was defined containing the sorts formula, list- of-formulas, substitution and list-of-substitutions. T h e sort formula was further subdivided into c o m p o u n d - , evaluable- and non-evaluable-formula. A further specialization o f list-of-formulas was non-empty-list-of- formulas. In the example shown in Figure 2, clause (1) states that in order to prove a n o n - c o m p o u n d expression F on the basis o f the contents o f knowledge base partition P giving a substitution S as a result, the system should either try to see if the formula is a know n fact in the knowledge base, try to infer the formula on its own, or it should ask the end-user. Trying to infer the formula means generating all possible inferences, select- ing some o f these possible inferences, and continuing with clause (2). T h a t clause chooses the best o f all selected possible steps, and tries to continue the p r o o f with this selection. I f this succeeds, the p r o o f terminates (i.e. non-cxhaustive), if this fails, the p r o o f continues with the next best step. Clause (3) states the criterion used in the best-first search, while clause (4) describes what needs to be done in order to prove a c o m p o u n d expression: prove both left- and right-hand sides o f the c o m p o u n d expression, and c o m b i n e the results. Clause (5) states that all evaluable predicates e n c o u n t e r e d in a p r o o f should be evaluated without any further control scheduling. This example shows how the different aspects o f this strategy can be changed if needed for a particular appli- cation. F o r example, the order o f the disjuncts in clause (1) might be changed to ask the end-user for solutions before the system tries a p r o o f itself, or the ask-user (I) (v F:non-evaluable-formula, P:partition, S:substitution, Next:non-empty-list-of-formulas. SomeNext:non-empty-list-of-formulas) [knowledge-base-lookup(F, P, S) V ( object-level-interpreter(F, P, backward. Next) & ~lect-inferences(Next, SomeNext) & infer(SomeNext, P, S)) V ask-uscr(F, P, S)) -oproof(F. P, S)] (2) (V inferences:non-empty-list-of-formulas, P:partition, S:substitution, Best:formula, Rest:list-of-formulas) [best(Inferences. Best,Rest) & (proof(Best, P. S) V infer(Rest, P, S)) -,infer(Inferences, P. S)] (3) (V List-non-empcy-list-of-formulas, Best:formula, Rest:list-of-formulas) [higher-certainty-value(List, Best,Rest) --.best(List, Best,Rest)l (4) (V F:compound-formula. P:partition, S:substitution, Lhs:formula, Rhs:formula, LhsSubst:substitution, RhsSubst:substitution) [split-compound-expression(F, Lhs, Rhs) & proof(Lhs;, P, LhsSubst) & proof(Rhs, P, RhsSubst) &-combine(LhsSubst. RhsSubst, S) --* proof(F, P, S)] (5) V F:evaluable-formula, P:partition, S:substitution. [evaluate(F. P, S) -, proof(F, P, S)] Figure 2. Declarative speciJ~cation of control in Socrates disjunct might be deleted altogether. T h e criterion used for the best-first scheduling could be changed, or a new decision for scheduling the o rd er in which conjuncts are proved in clause (4) could be introduced. M o r e t h o r o u g h changes to the strategy could also be made, but they would a m o u n t to writing a completely new p r o o f strategy rather than changing the one shown in this example. O f the evaluable predicates in Figure 2 (such as ask- user, knowledge-base-lookup, combine) the predicate object-level-interpreter is the most i m p o r t a n t one. This predicate encapsulates the interface between the meta- level interpreter executing the co n t ro l strategy, and the object-level interpreter handling the logical representa- tion language and the c o r r e s p o n d i n g p r o o f theory. The input o f this predicate is an object-level formula F, the name o f a knowledge base partition P and a direction in which to apply inference rules (either forward or back- ward), and returns as o u t p u t the result o f applying all inference rules specified as part o f the p r o o f theory for partition P to the input fo rm u l a F in the indicated direc- tion. In terms o f the p r o o f search space, this a m o u n t s to generating all nodes that are accessible from the current n o d e as represented by F. Thus. the predicate object-level-interpreter allows the meta-level interpreter to access an explicit representation o f the object-level search space, and to choose which branches o f the object-level p r o o f tree will be ex p an d ed on the basis o f the control regime provided by the knowledge engineer. Unlike m a n y logic-based meta-level architectures pro- posed in the literature, (such as Silver t°, or the Prolog system described inZ°), Socrates completely separates the languages used at the object-level from the language used at the meta-level. Even when the object-level repre- sentation language h ap p en ed to be defined as sorted H o r n Clause Logic, the two languages would still be syntactically separate. Th e meta-level and object-level languages are connected t h ro u g h a 'n am i n g relation'. Th e meta-level language contains names for all object- level expressions. In Socrates, the n am e o f an object-level sentence c o r r e s p o n d s to a co n st an t in the meta-level language. O t h e r meta-level constants are used to denote bindings for object-level variables. If this is required, meta-level expressions could range o v er any o f the extra- logical properties o f object-level expressions, such as truth values, certainty factors, justifications, etc. In this way, for instance, Socrates could be configured to deal with certainty values by specifying as part o f the control strategy how certainty values should be used in a proof. This c o r r e s p o n d s to the a p p r o a c h suggested by Shapiro 2;, with the i m p o r t a n t difference that Socrates makes a correct distinction between meta-level and object-level languages, whereas Shapiro confuses the two and uses Prolog for both. A n u m b e r o f reasons can be given why it is i m p o r t a n t for the meta-level language to be separated from the object-level languages. First, there is an epistemological reason: as argued in 2~, different d o m a i n s require different representation languages, and the object-level and the meta-level o f Socrates deal with widely different d o m a i n s (the object-level deals with the application d o m a i n o f the system, while the meta-level deals with the issue o f controlling the object-level). A second argu- ment concerns the m o d u l a r i t y o f the system: it should 138 Knowledge-Based Systems be possible to vary control knowledge and domain knowledge independently. The third argument is one o f explanation: in order to enable the system to include control knowledge explicitly in its explanations, it is important for both the human reader and the automated explanation generator that control knowledge can be syntactically distinguished from domain knowledge. Procedural representation of control knowledge The approach to the specification o f a p r o o f strategy described above is based on the use of a declarative meta-level language. Although the declarative style o f control o f reasoning has its attractions, there are also disadvantages. Two problems in particular are caused by the use o f a declarative language, and in an attempt to overcome these problems Socrates provides an alter- native, more procedural language for specifying control regimes. First, the extra layer o f interpretation that is incurred by the explicit meta-level interpreter is expen- sive, because o f the declarative nature o f the meta-ievel language. A procedural meta-level language would be much closer to the underlying implementation language and machine architecture, and therefore cheaper to execute. Second, much o f the knowledge expressed in the meta-level language is procedural, rather than declarative. For example, one often wishes to apply knowledge o f the form 'try method-I before method-2', or 'in order to achieve goal-l, achieve sub-goal-I to subgoal-n'. In the declarative meta-level language this type o f procedural knowledge has to be expressed by either relying on the hardwired control regime for the meta-levei interpreter, or by using semantic attachment. Neither o f these ways o f expressing procedural knowledge is very desirable, since they encode knowledge implicitly rather than represent it explicitly. A more pro- cedural language provides a more natural medium for expressing the procedural control knowledge. Our approach to procedural control is therefore to provide a 'meta-level programming language' rather in the vein o f M L 3° in its relationship to LCF. That is, we provide high level primitives that make the writing o f control regimes easier. It is important to note that a procedural control language is only an alternative way o f implementing the separation o f control knowledge from object-level knowledge. The procedural approach still clearly separates control knowledge from object- level knowledge. The difference is that rather than put- ting the control knowledge into a declarative knowledge base with its own interpreter, we propose implementing a special purpose meta-level interpreter that incorporates the control knowledge. The three stage process o f build- ing systems is retained. One particular point to note is that we retain the explicit declaration o f inference rules and provide primitives to apply inference rules. The procedural meta-levei language consists o f the implementation language o f the system (Common Lisp), extended with primitives implementing standard artifi- cial intelligence techniques that have been found useful in writing interpreters. Central to the system is the use o f lazy evaluation. The procedural meta-level language provides facilities for the manipulation o f lazily evaluated lists, which form the basis for backtracking and coroutining in the control regimes. The second important component is agenda- (defun proof (GoalList Partition Subst &aux NewGoals NewSubst) (if GoalList (foreach (NewGoals. New Subst) in (or (any-of (order-inferences (evaluate (first GoalList) :partition Partition :subst Subst)) (order-inferences (lookup (first GoalList) :partition Partition :subst Subst)) (select-inferences (generate-backward-inferences (first GoalList) :partition Partition :subst Subst))) (ask-user (first GoalList) :partition Partition :subst Subst)) generate-each (proof (append NewGoals (cdr GoalList)) Partition NewSubst)) (list Subst))) Figure 3. Procedural specification o f control in Socrates based reasoning. This technique allows the knowledge engineer to experiment with several different control strategies, often only changing the way the agenda is handled without changing the rest o f the control regime code. The third component is a pattern matcher which provides the basis o f pattern directed invocation of p r o o f methods. It is important to realize that in this procedural approach it is still the case that the only inference rules are those declared explicitly during the declaration o f the p r o o f theory (as described above). Language primi- tives are provided to apply the declared inference rules. Figure 3 shows an example o f a control regime formu- lated in the procedural meta-level language. This code is the procedural equivalent o f the declarative prover given in Figure 2. The main function is called 'proof' and performs the same function as the predicate o f that name. Being stream-based, p r o o f returns a stream o f substitutions that prove the goals in the 'GoalList' argu- ment. Thus, if the GoalList is empty there is one such proof, given by the Subst argument. If the GoalList is not empty then the inference rules are applied to the first goal in GoalList. The ' a n y - o f macro is a way o f lazily combining streams. Thus, in this example, we generate all the possible inferences using 'evaluate' (and put them in a preferred order), then all the possible inferences using 'lookup' (again in preferred order), finally followed by a selection o f the possible inferences generated using all the inference rules. ' I f and only if' no possible inferences are generated by this procedure, then 'ask-user' is applied to obtain possible inferences. Each inference rule application generates three things: a set of new goals to be proved in order to prove the goal (bound to the variable 'NewGoals'); a new substi- tution, bound to the variable 'NewSubst'; and a justifi- cation, which merely describes which inference rule has been applied and is effectively ignored by this prover. The 'foreach' macro itself generates a stream o f answers. Thus, if at some later stage in the p r o o f there is a need to backtrack, the next inferences will be generated only then. Note that, unlike the declarative prover, which generates only the first proof, the procedure ' p r o o f returns a stream o f proofs, and thus generates the set o f all proofs (lazily). It was noted at the beginning o f this section that much Vol 1 No 3 June 1988 139 control knowledge seems very procedural in nature (such as the execution o f a sequence o f goals). However, some- times control knowledge is declarative in nature (e.g. a set o f criteria used in the ordering o f conjuncts). The architecture o f Socrates is such that even with procedural control it is still possible to invoke a declarative meta- level interpreter that is implemented using the techniques described above. Although the above features provide a basis for a procedural language for formulating control regimes, certain problems remain. First, the current language may not be powerful enough, and further additions may be needed. Second, although in one sense the current proce- dural meta-level language might not be powerful enough, in a n o t h e r sense it might be too powerful. As described, the procedural meta-level language consists o f extensions to C o m m o n Lisp, thereby making all the general purpose expressive power o f Lisp available to the knowledge engineer when writing control regimes. It might well be the case that this provides too powerful a language, since it does not restrict the knowledge engineer in any way. The scheduler A third level in the architecture o f Socrates (see Figure 1) is the 'scheduler'. This third level is not actually imple- mented in the current Socrates architecture, but it can be added to the current system with little effort. The main notion that is treated at this level is that o f a "subtask'. As shown in Reichgelt and van Harmelen 29, m a n y expert systems perform not just one simple task, but a composite one that can be thought o f as consisting of a number o f elementary tasks ( M Y C I N , RI and VM are a m o n g the systems discussed in that paper). It is unlikely that one appropriate control regime can be found that would be suitable for these composite tasks. Rather, the composite task should be split up into its constituent subtasks, and a proper control regime can then be chosen for each o f the subtasks. The subtasks that would result from this decomposition process are the kind o f prototypical tasks proposed in Reichgelt and van Harmelen 29, Chandrasekaran 6, and Breuker and Wielinga 5, like classification, monitoring, simulation, design, etc. The scheduling level o f the Socrates architecture is meant to deal with this sub- division o f the major task into prototypical subtasks. Each o f these prototypical subtasks can then be solved using the appropriate meta-level control strategy. (By using the knowledge base partition mechanism, it is possible to equip a Socrates configuration with more than one control strategy.) F o r engineering purposes it would be easiest to equip the scheduling level with a language similar to (but again syntactically separate from) the language used to describe the control strategy at the meta-level. However, early experience indicates that the type o f knowledge to be expressed at the schedul- ing level is o f a very procedural nature (even more so than the knowledge expressed at the meta-level), and therefore a language with more conventional procedural primitives, such as sequences, conditionals, loops and subroutines, might be more appropriate. PRACTICAL P R O G R E S S A N D A C H I E V E M E N T S A version o f the Socrates architecture as described above has been implemented in approximately 15K lines o f C o m m o n Lisp code. (Notice the distinction between "implementation language" and "representation language'of a system: although the knowledge represen- tation and inference process o f Socrates are both logic based, the system is n o t implemented in Prolog or any other logic programming language.) The system currently runs in a number o f C o m m o n Lisp implemen- tations on U N I X based systems. In one o f these C o m m o n Lisp systems, the Poplog system, a graphics- based knowledge engineer interface, has been con- structed, allowing interaction with the system via menus, browsers, graphers, etc. A substantial set o f different control strategies has been written as meta-level programs, including backward and forward chaining, exhaustive and non-exhaustive search, user guided or automatic conflict resolution, best-first, depth-first, breadth-first search, branch and bound type algorithms, generate and test procedures, elimination and confirmation strategies, etc. A number o f demonstration systems have been built using Socrates, including an expert system in the domain o f personal investment advice and a route planning system. More significantly. Socrates has been used to reimplement an existing expert system under the name o f DOCS, developed by G E C Research in collaboration with Westminster Hospital, London. This system con- sists o f 130 rules divided over two knowledge base partitions. The partition hierarchy is organized so that these two partitions can use data from a c o m m o n work- ing memory partition. A sort hierarchy o f over 80 sorts was used to model the taxonomic hierarchy o f the medical domain. A generate and test control strategy was specified tbr this system, consisting o f some 20 clauses in the meta-level knowledge base. In the area o f theorem proving Socrates has been used to solve the problem described by Walther 2', known as 'Schubert's steamroller'. This problem was originally formulated because o f the huge search space that it generates. Using the sort hierarchy to model the taxonomic part o f the problem, a breadth-first search strategy, implemented using our procedural control tech- niques, resulted in the first natural deduction style p r o o f for this problem. A number o f other procedural control strategies were implemented to solve the problem, and a comparison o f these different strategies showed the importance o f the exploitation o f meta-knowledge to guide the search for a proof. This application o f Socrates is described in detail by Davies 23. Current areas o f activity are: • The use o f meta-level interpretation for dealing with extra-logical issues, such as uncertainty, truth maintenance, explanation, etc. (how the slot-value a n n o t a t i o n mechanism o f Socrates' knowledge base opens up these possibilities is described above). • The implementation o f modal logics through the use o f reification, as described by Reichgelt 3 ~. • The classification o f domains and subtasks, as stated by Reichgelt and van Harmelen 2s'29, in order to provide the knowledge engineer with guidelines that indicate how to choose the appropriate representation language given a particular application domain, and how to choose the appropriate control regime given a particular prototypical task. • The development o f a library o f control regimes that 140 Knowledge-Based Systems can be executed by the system. Rather than having to write a new control strategy from scratch every time, the knowledge engineer can use a library o f preprogrammed control strategies. The knowledge engineer can then either use one o f the strategies directly from the library, or use one o f the library elements as the basis for his own control strategy by making small changes to the preprogrammed strategy. O P E N P R O B L E M S O f the three stages o f the configuration process, the third one (declaring a p r o o f strategy) is by far the most problematic. An important open question here is the choice o f a good language for specifying such a control strategy. As described above, the system primarily uses a declarative logical language to do this, but a procedural language has also been investigated. Apart from the type o f language used at the meta-level, a related problem is the required vocabulary o f such a language. At the moment, the vocabulary o f the system is specified by the knowledge engineer, and although this is to a certain extent inevitable, since part o f the vocabulary will be application-specific, one would hope that at least a central core vocabulary can be distinguished that can be preprogrammed into the system. The example o f a control regime discussed above suggests predicates to do with manipulating substitutions and formulas, and with generating the object-level search space, but a more extensive and more exactly defined vocabulary is needed in order to alleviate the task o f the knowledge engineer. A second problem associated with the explicit meta- level interpreter is that o f meta-level overhead. Although the flexibility in defining the appropriate control strategy at the meta-level can considerably reduce the object-level search space, the price we have to pay for this is the fact that the object-level inference process is completely simulated by the meta-level interpreter. This is obviously much more expensive than an object-level interpreter that has the appropriate control strategy hardwired into it. This problem could be solved by taking the explicit formulation o f a control regime, and compiling it into an interpreter that has the particular control regime hardwired into it. This compilation process (whose first stages could be similar to that described by Altman and Buchanan 3z) has been simulated in Socrates by hand coding a number o f hardwired control strategies. Experi- ence with these hardwired strategies in both the DOCS system and in solving Schubert's steamroller indicates that the meta-level overhead can indeed be reduced to an acceptably small amount. C O N C L U S I O N The work described in this paper attempts to create an environment for building expert systems based on the following principles: • An epistemological analysis o f the domain and task o f a particular application guides the choice o f the appropriate knowledge representation language and the appropriate control regime. • Logic is used as the main underlying formalism. • Control knowledge is represented explicitly and is separated from the domain knowledge. When configuring the Socrates environment into a particular expert system, a knowledge engineer can vary the architecture along three dimensions: • The representation language: a knowledge engineer can define his own logical representation language, including first order logics (possibly many-sorted), modal logics, temporal logics, etc. • The inference rules for the logical language: the set o f rules that determine the possible inferences made in the logical language can be changed by the knowledge engineer. • The control regime under which the inference rules will be used to perform proofs in the logical represen- tation language. An implementation o f the Socrates abstract architecture and a number o f applications o f the system have proved the feasibility o f this approach. Due to limitations o f space, some o f the arguments and descriptions in this paper are rather terse. A long version o f this paper can be found in Corlett, Davies, Khan, Reichgelt and van HarmelenL A C K N O W L E D G E M E N T S As the grant holder, Peter Jackson has made many con- tributions to the work described above. Ray McDowell and Dave Brunswick have contributed to the work done at GEC Research. The research was carried out as part of Aivey Project IKBS/031 in which G E C Research, the University o f Edinburgh and G E C Avionics were partners. The university work was supported by SERC grant GR/D/17151. R E F E R E N C E S 1 AIvey, P 'Problems o f designing a medical expert system' 3rd Tech. Conf. British Comput. Soc. Specialist Group on Expert Syst. (December 1983) pp 20-42 2 Bobrow, D and Stef'~, M The LOOPS Manual Rank Xerox, U K (1983) 3 The knowledge engineering environment Intellicorps, California, USA (1984) 4 Williams, C ART, the advanced reasoning tool, conceptual overview Inference Corporation, Los Angeles, California, USA (1983) 5 Breuker, J A and Wiefinga, B J 'Models o f Expertise' Proc. 7th Europ. Conf. on Artif. Intel. (July 1986) pp 306-318 6 Chandrasekaran, B 'Towards a functional architec- ture for intelligence based on generic information processing tasks' 7th Int. Joint Conf. on Artif. lntel. (August 1987) pp 1183-1192 7 Corlett, R, Davies, N, Khan, R, Reichgelt, H and van Harmelen, F 'Socrates: a flexible toolkit for building logic based expert systems' in Jackson, P, Reichgelt, H and van Harmelen, F (eds) Logic-based knowledge representation MIT Press, USA (1988) 8 Davis, R 'Meta rules: reasoning about control' Artif. lntel. Vol 15 No 2 (1980) pp 179-222 9 Bundy, A and Welham, B 'Using meta-level inference for selective application o f multiple rewrite rules in algebraic manipulation' Artif. Intel. Vol 16 N o 2 (1981) pp 189-212 Vol 1 No 3 June 1988 141 10 Clancey, W "The advantages of abstract control knowledge in expert system design' Proc. 3rd Annual Meeting o f th ~ American Assoc..for A rt([~ Intel. (1983) pp 74-78 11 Aiello, L and Levi, G 'The uses o f metaknowledge in AI systems' Proc. European Conf. on Artif. Intel. (September 1984) pp 707-717 12 Clancey, W 'Representing control knowledge as abstract tasks and metarules' in Cooml~, M and Bolc, L (eds) Springer-Verlag, F R G (1985). Also: Stanford Knowledge Systems Laboratory, Working Paper No KSL-85-16 (April 1985) 13 Claneey, W 'The epistemology of a rule-based expert system: a framework for explanation' Artif. lntel. Vol 20 (1983) pp 215-251. Also: Stanford Heuristic Programming Project, Memo HPP-81-17, STAN-CS- 81-896 (November 1981 ) 14 Bundy, A, Byrd, L, Luger, G, Mellish, C, Milne, R and Palmer, M 'Solving mechanics problems using meta-level inference" Proe. Int. Joint Conf. on Artif. lntel. (August 1979). Also in Michie, D (ed.) Expert S.vstems in the Micro Electronic Age Edinburgh University Press, UK (1979) pp 50 64 15 Davis, R 'TEIRESIAS: applications of meta-level knowledge' in Davis, R and Lenat, D B (eds) Knowledge-Based Systems in Artificial Intelligence McGraw-Hill, New York, USA (1982) pp 227~,90 16 Pereira, L M "Logic control with logic' Proc. 1st Int. Logic Programming Conf. (1982) pp 9-18. Also in: Campbell, J A (ed) Implementations ~?fProlog Ellis Horwood, UK (1984) 17 Genesereth, M and Smith, D 'An overview o f meta- level architecture' Proc. 3rd Annual Meeting o f the American Assoc. for Art(f. Intel. (1983) pp 119-124 18 Sterling, L 'Implementing problem solving strategies using the meta-level" DAI Research Paper No 209, Department o f Artificial Intelligence, University of Edinburgh, UK (1984) 19 Silver, B Meta-level InJerence Studies in Computer Science and Artificial Intelligence, North Holland, Amsterdam, The Netherlands (1986) 20 Welham, B 'Peclaratively programmable interpreters and meta-level inferences' Hewlett-Packard Labora- tories Bristol Research Centre Technical Memo No HPL-BRC-TM-86-027 (September 1986). Also in: Maes, P and Nardi, D (eds) Meta-h, vel architectures an'l r~/h, ction North Holland. Amsterdam. The Netherlands (1987) 21 Waltber, C "A mechanical solution of Schubert's steamroller by many-sorted resolution" Proc. 4th Annual Meeting of the American Assoc. Art(f. lntel. (1984) pp 330- 334 22 Cohn, A G 'OL the solution of Schubert's steamroller in many sorted logic" Pro,'. 9th Int. Join t. Con/i on Art(/: lntel. (August 1985) pp 345-352 23 Davies, N "Schubert's steamroller in a natural deduction theorem prover" Proc. 7th Tech. Conf. o f the British Comput. Soc. Specialist Group on Expert Syst. (December 1987) 24 Weyhraueh, R 'Prolegomena to a theory of mechanised formal reasoning' Art(/: hztel. Vol 13 No 1 (1980)pp 133 170 25 Bledsoe, W 'Non-resolution theorem proving" Artif. Intel. Vol 9 No I (1977) pp I--35 26 Gallaire, M and Lassetre, C 'Meta-level control for logic programs" in Clark, K and Tarnlund, S (eds) Logic Programming Academic Press. UK (1982) pp 173 188 27 Shapiro, E "Logic programs with uncertainties: a tool for implementing rule-based systems" Proc. 8th Int. Joint Conf. on Art(f. lntell. (August 1983) pp 529-532 28 Reichgelt, H and van Harmelen, F "Relevant criteria for choosing an inference engine in expert systems" Proc. 5th Tech. Conf. o f the British Comput. Soc. Specialist Group on Expert Syst. (December 1985) pp 21--30 29 Reichgelt, H and van Harmelen, F "Criteria for choos- ing representation languages and control regimes for expert systems" The Knowledge Eng. Rev. Vol I No 4 (December 1986) pp 2-- 17 30 Gordon, M, Milner, R and Wadsworth, C Edinburgh LCF." a mechanized logic o f computation Springer- Verlag Lecture Notes in Computer Science, Vol 78 Springer-Verlag, F R G (1979) 31 Reichgelt, H 'Semantics for a reified temporal logic" Proc. 1987 A I S B Conf. (April 1987) pp 49 62 32 Altman, R B and Bwehanan, B G 'Partial compilation of strategic knowledge' Proc. 6th Annual Meeting o[ the American Assoc. /or ,4rt(/~ lntel. (July 1987) pp 399- 404 142 Knowledge-Based Systems 
work_mjudm5ydjvcn7oi53shtzs4cqu	----	An Algebraic Approach to Rule Based Expert Systems Available on line at www.rac.es/racsam REVISTA DE LA REAL ACADEMIA DE CIENCIAS Computational Sciences EXACTAS, FISICAS Y NATURALES. Survey SERIE A: MATEMATICAS Madrid (España / Spain) RACSAM 104 (1), 2010, 19–40. DOI:10.5052/RACSAM.2010.04 An algebraic approach to rule based expert systems Eugenio Roanes-Lozano, Luis M. Laita, Antonio Hernando and Eugenio Roanes-Macı́as Abstract This article presents a survey of the authors’ research on knowledge extraction and verification of Rule Based Expert Systems (RBES) using algebraic inference engines and based on Gröbner bases theory. A shell, including a graphic user interface and inference engines for different logics (both classic and modal multi-valued) as well as in different computer algebra systems, is also presented here. The shell distinguishes three levels: at the lower level, we provide the computer algebra system code of the algebraic inference engines; at the intermediate level, the RBES developer has to detail the rules and integrity constraints of a certain RBES; and, finally, at the upper level, the final user deals with a simple GUI, where he can perform knowledge extraction or verify the RBES, after choosing the logic and inputing a consistent set of facts. We believe that this shell can be really useful for teaching and quick RBES design. Una aproximación algebraica a los sistemas expertos basados en reglas Resumen. Este artı́culo presenta una panorámica de la lı́nea de investigación de los autores en ex- tracción de conocimiento y verificación de Sistemas Expertos Basados en Reglas (RBES) usando motores de inferencia algebraicos y basada en la teorı́a de bases de Gröbner. Se presenta también una shell, que incluye una interfaz gráfica de usuario y motores de inferencia para distintas lógicas (tanto clásicas como modales multivaluadas) y en distintos sistemas de cómputo algebraico. La shell distingue tres niveles: en el más bajo proporcionamos el código del motor de inferencia para el sistema de cómputo algebraico elegido; en el intermedio el desarrollador del RBES tiene que detallar las reglas y las restricciones de integridad de un cierto RBES; y, finalmente, en el nivel superior, el usuario final trata con una sencilla interfaz gráfica de usuario, en la que puede llevar a cabo extracción de conocimiento o verificar el RBES, después de elegir la lógica y de introducir un conjunto consistente de hechos. Creemos que esta shell puede ser realmente útil para la enseñanza y para el rápido diseño de RBES. Submitted by David Rı́os Insua. Received: August 2, 2009. Accepted: October 19, 2009 Keywords: Rule based expert systems, logic and symbolic computing, Groebner bases Mathematics Subject Classifications: 03B05, 13P10, 03B80 c© 2010 Real Academia de Ciencias, España 19 http://www.rac.es/racsam E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 1 Introduction This article presents a survey of the authors’ research on knowledge extraction and verification of Rule Based Expert Systems (RBES) using algebraic inference engines. These inference engines are based on the use of Gröbner bases. The leader of the research team is Luis Laita, and the first works on the topic are dated in the early 90’s, although he had already been working in the field using other techniques. The present paper is self-contained, in such way that introductory well-known sections 2, 5, 7may be skipped by an acquainted reader, while the other sections convey a view of new results and approaches by the authors. The article begins with a brief introduction to lattices, lattice orders, Boolean algebras and Boolean rings (Section 2). It is followed by the construction of the polynomial model for Boolean logic due to these authors (Sections 3 and 4). Then a brief introduction to modal multi-valued logics can be found (Section 5). Section 6 introduces the main result of this theory, relating to be a tautological consequence with a polynomial ideal membership. A brief introduction to RBES is included afterwards (Section 7). Section 8 details the adaptation of the algebraic model for logic to RBES. In Section 9 it is shown how two levels of inconsistency can be distinguished in the multi-valued case (including an interpretation in algebraic geometry). Section 10 introduces how (in RBES whose underlying logic is classic Boolean logic), given a set of facts, it is possible to find out which new fact can make a certain goal be inferred, just computing a Gröbner basis. Implementations in the computer algebra systems (CAS) CoCoA and Maple are included in Section 11. Finally, a GUI for keeping the CAS hidden from the final user of the RBES is presented in Section 12. In view that quite a number of topics are treated, many proofs are omitted for the sake of brevity (in fact, some of the proofs are very long, for instance, the one connecting the concept of being tautological consequence in multi-valued logic with an ideal membership in a polynomial residue class ring). Anyway, references where the reader can find details on the topics treated can be found along the paper. 2 An introduction to Boolean algebras and Boolean rings An elementary introduction to Boolean algebras may be found in [21]. The proofs omitted in this section can be found in [18, 28] (these references also extend both Sections 3 and 4). For a deeper study of Boolean algebras see, for instance, [10, 11, 12, 22, 36]. 2.1 Lattices and lattice orders Definition 1 A set L where two binary operations t and u are defined is said to be a lattice if and only if both operations are commutative and associative and the absorption laws holds, i.e., if and only if for all elements a, b, c ∈ L: at b = bta au b = bua at (bt c) = (at b) t c au (bu c) = (au b) u c at (bua) = a au (bta) = a Proposition 1 If (L,t,u) is a lattice, both operations verify idempotency, i.e., for every element a ∈ L: ata = a aua = a Surprisingly, idempotency, although redundant, is often required as a fourth condition for a structure to be a lattice. Definition 2 A non-strict partial order defined over the set L is said to be a lattice order if and only if every pair of elements of L has a unique infimum and a unique supremum. 20 An algebraic approach to rule based expert systems Proposition 2 From the two lattice operations, t and u , a lattice order v can be defined: for all elements a, b ∈ L: a v b ⇔ at b = b (⇔ au b = a) Conversely, from a lattice order v , the two operations of a lattice can be defined: for all elements a, b ∈ L: at b = supv(a,b) au b = infv(a,b) Example 1 Let us consider the lattice of “subsets of E”: (P(E),∪,∩). From (P(E),∪,∩), (P(E),⊆) is obtained. Conversely, from (P(E),⊆), (P(E),∪,∩) is obtained. 2.2 Boolean algebras Definition 3 A lattice (L,t,u) is said to be distributive if and only if both operations are distributive w.r.t. the other operation, i.e., if and only if for all elements a, b, c ∈ L: at (bu c) = (at b) u (at c) au (bt c) = (au b) t (au c) Definition 4 A lattice is said to be bounded if and only if it has a greatest element (top), g, and least element (bottom), l. Definition 5 A bounded lattice (L,t,u) is said to be complemented if and only if for every element a ∈ L, there is an element a′ ∈ L such that: ata′ = g aua′ = l The top and bottom of a lattice are traditionally denoted 0 and 1. However this seems not to be an adequate notation, as the top and bottom of (L,t,u) and those of (L,u,t) are exchanged. Definition 6 A Boolean algebra is a distributive and complemented lattice. Example 2 (P(E),∪,∩, ′) is a Boolean algebra: • ∪ is distributive w.r.t. ∩ and viceversa. • The least and the greatest of the lattice (for ⊆) are ∅ and E, respectively. Moreover, the usual complement of a set, ′ , works as expected: ∀A ∈P(E) : A∪A′ = E ; A∩A′ = ∅ Let us underline that being distributive is completely independent from being complemented, as shown below. Example 3 The lattice of the linear subspaces (also denoted “vector subspaces”) of R2, the sum of linear subspaces and intersection is a complemented, but not distributive, lattice: • the complement of a line is any other (different) line, and the complement of the whole plane is {0}, • if R, S, T are different lines, R + (S ∩T) 6= (R + S) ∩ (R + T)). Example 4 The lattice (N, lcd, gcd) (where lcd and gcd stand for “least common multiple” and “greatest common divisor”, respectively) is distributive, but it is not complemented: 1 and 0 are the least and greatest of the lattice (respectively), but, for instance, 3 has no complement. Example 5 The lattice of the convex subsets of the plane, the convex union ∗ ∪ (i.e., the least convex set that contains the usual union) and intersection is a lattice, but it is neither distributive nor complemented: • for instance, if A, B, C are three disjoint and aligned squares of equal size and A is the one in the middle A∩ (B ∗ ∪ C) = A, meanwhile (A∩B) ∗ ∪ (A∩C) = ∅ ∗ ∪∅ = ∅, • there is no convex subset of the plane that behaves as the complement of a line. 21 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 2.3 Boolean rings Definition 7 A Boolean ring is a ring with unit such that the second operation is idempotent. Let us include below some well-known results that will be useful when obtaining the polynomial model for Boolean logic. Let us denote by (L, ∆,u) a general Boolean ring, by 0 its neutral element, by ′ the symbol for the opposite and by 1 its unit. Proposition 3 In any Boolean ring (L, ∆,u) any element is its own opposite. PROOF. For any elements a, b ∈ L: (a ∆ b) ∆ 0 = a ∆ b = (a ∆ b) u (a ∆ b) = (aua) ∆(au b) ∆(bua) ∆(bu b) = a ∆((au b) ∆(bua)) ∆ b = (a ∆ b) ∆((au b) ∆(bua)) but (L, ∆) is a group, so the opposite is unique, and consequently: 0 = (au b) ∆(bua). Therefore, in the particular case that a = b, from the idempotency of u, we would obtain: 0 = a ∆ a. � Proposition 4 Any Boolean ring (L, ∆,u) is a commutative ring. PROOF. It follows from the equality 0 = (au b) ∆(bua) in the previous proof and the facts that (L, ∆) is a group (and therefore the opposite is unique) and that any element is its own opposite. � Proposition 5 A Boolean ring (L, ∆,u) can be defined from a Boolean algebra (L,t,u, ′): for all elements a, b ∈ L: a4b = (au b′) t (a′ u b) Conversely, a Boolean algebra (L,t,u, ′) can be defined from a Boolean ring (L, ∆,u): for all elements a, b ∈ L: at b = (a4b)4(au b) and if 0 and 1 are the neutral elements of ∆ and u (respectively), then 0 is the least, 1 is the greatest, and: a′ = a ∆ 1 Example 6 (P(E),∪,∩, ′) is a Boolean algebra (least: ∅; greatest: E). If we define the symmetric difference of two elements of P(E), A and B, as: A4B = (A∩B′) ∪ (A′ ∩B) we have that (P(E),4,∩) is a commutative ring with unit: all properties are a straightforward conse- quence of the same property of the Boolean algebra or can be obtained from a simple symbolic manipulation (like the commutativity and associativity of 4 ). Moreover: • ∅ is the neutral element for 4 (and the absorbent element of ∩ ), • E is the neutral element for ∩ , • ∀A ∈P(E), A4A = ∅, so any element in P(E) is its own opposite. Example 7 (P(E),4,∩) is a Boolean ring. If we define in this ring the following operation: AtB = (A4B)4(A∩B) the operation ∪ is obtained. If we define A′ = A ∆ E the usual complement of a set is obtained. 22 An algebraic approach to rule based expert systems 3 A polynomial model for propositional logic We can obtain a polynomial model for the Boolean algebra associated to (propositional) Boolean Logic, as suggested in Figure 1. propositional Boolean algebra oo Boolean algebra isomorphism // OO �� polynomial Boolean algebra OO �� propositional Boolean ring oo ring isomorphism // polynomial (Boolean) ring Figure 1. Isomorphisms between polynomial and propositional Boolean algebras and Boolean rings. In logic, we usually work in the structure in the upper left corner of the diagram of Figure 1. Meanwhile, in commutative algebra we normally work in the structure in the lower right corner of this diagram. But we can move from the propositional Boolean algebra to the polynomial ring following the two alternative paths in this diagram. The advantages of working in a polynomial ring are that: • there is a powerful tool that solves the ideal membership problem: Gröbner bases, • there are many implementations of Buchberger’s algorithm for computing Gröbner bases, i.e., per- forming effective calculations in algebra. In fact, most CAS, like Maple, Mathematica, Axiom, Mu- PAD, Reduce, Derive, CoCoA, Singular, Risa/Asir,. . . include implementations of this algorithm. As a consequence, we shall be able to implement very easily an algebraic tool to perform effective compu- tations in propositional Boolean logic. Moreover, it will be possible to extend this tool to perform effective computations in modal multi-valued logics (like Łukasiewicz, Kleene and Bochvar). Finally, it will be pos- sible to use the same computational tool to perform verification and knowledge extraction in RBES based on both classic Boolean and modal multi-valued logics. 3.1 Building the Taylor-made polynomial (Boolean) ring (A, +, ·) Let us build a polynomial Boolean ring (A, +, · ) so that: • every element is its own opposite, i.e., for all a ∈A, a + a = 2 ·a = 0, so it would be reasonable to consider A as a ring over a field with characteristic 2, • idempotency holds for the second operation, so it would be reasonable to consider the residue class ring over the polynomial ideal 〈p2 − p,q2 − q, . . . ,r2 − r〉, where p, q, . . . , r are the polynomial variables. Therefore, we shall consider: A = Z2[p,q, . . . ,r]/〈p2 −p,q2 −q, . . . ,r2 −r〉 Proposition 6 The polynomial product in A is idempotent. PROOF. It is enough to take into account that any polynomial a ∈A can be written: a = δ0 + δp ·p + δq ·q + · · · + δr ·r + δpq ·p ·q + · · · + δpqr ·p ·q ·r + · · · where all the δ belong to Z2, and compute a2. � It is well-known that an algebraic structure like (A, +, · ) is a ring (denoted residue class ring) and, therefore, we have the following: 23 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as Corollary 1 The ring (A, +, · ) is a Boolean ring. Proposition 7 All elements in (A, +, ·) are zero divisors (in fact an analogous result holds in any Boolean ring). PROOF. a · (1 + a) = a + a2 = a + a = 2 ·a = 0. � 3.2 The polynomial Boolean algebra (A, +̃, · , 1+) corresponding to the po- lynomial (Boolean) ring (A, +, ·) Let us define a (polynomial) Boolean algebra (A,t, · , ′) from the (polynomial) Boolean ring (A, +̃, ·), as shown in Proposition 5. For all a, b ∈A: at b = (a + b) + (a · b) a′ = a + 1 The first operation will be hereinafter denoted +̃ and 1+ will denote the addition of 1. 0 is the least and 1 is the greatest of the Boolean algebra (A, +̃, · , 1+): for all a, b ∈A, a+̃0 = a + 0 + a · 0 = a a · 0= 0 a+̃1 = a + 1 + a · 1 = 1 a · 1= a Unlike what happens in a ring, it is not usual to establish priorities for the operations in a Boolean algebra (as the structure is completely “symmetrical”). Nevertheless, we shall consider, hereinafter, that · is prioritary w.r.t. +̃. Proposition 8 The lattice order defined in (A, +̃, · , 1+) as suggested in Proposition 2: ∀a,b ∈ A, a ≤ b ⇔ a · b = a is, precisely, “is a multiple” (a is a multiple of b ⇔∃k ∈A : a = b ·k). PROOF. ⇒) ∀a,b ∈A: a ≤ b ⇔ a · b = a ⇒ a is a multiple of b ⇐) ∀a,b ∈A: a is a multiple of b ⇔∃k ∈ A : a = b·k ⇒ a·b = (b·k)·b = b2 ·k = b·k = a ⇔ a ≤ b. � Proposition 9 For all a, b ∈A: (1) a · b = a (2) a+̃b = b (3) (1 + a)+̃b = 1 (4) a · (1 + b) = 0 are equivalent (an analogous result holds in any Boolean algebra). 4 The Boolean algebra isomorphism ϕ Let ∨, ∧, ¬, → denote the logic disjunction, conjunction, negation and implication, respectively. Let (C,∨,∧,¬,→) be the Boolean algebra of the propositions that can be constructed using a finite number of propositional variables P , Q, . . . , R. Let us denote tautology by 1 and contradiction by 0. Let us consider the Boolean algebra (A, +̃, · , 1+, “is a multiple”), where A is the residue class ring A = Z2[p,q, . . . ,r]/〈p2 −p,q2 −q, . . . ,r2 −r〉 24 An algebraic approach to rule based expert systems We define: ϕ: (C,∨,∧,¬,→) −→ (A, +̃, · , 1+, “is a multiple”) in the following way. For propositional variables P −→ p Q −→ q . . . . . . R −→ r and for any A, B ∈C A∨B −→ a+̃b ¬A −→ 1 + a Therefore, as an immediate consequence of the De Morgan laws: A∧B −→ a · b Moreover, as the (lattice) order can be obtained from the operations of the lattice, ϕ preserves the ordering. Proposition 10 ϕ is well defined. PROOF. For all propositions A, B: A ↔ B ⇒ A → B and B → A ⇒ ⇒ ϕ(A) is a multiple of ϕ(B) and ϕ(B) is a multiple of ϕ(A) ⇔ ⇔ a is a multiple of b and b is a multiple of a ⇔ a = b. � Corollary 2 ϕ is an order-preserving homomorphism, and ϕ(1) = 1,ϕ(0) = 0. Remark 1 In order to save space, we shall sometimes write b|a (“b divides a”) instead of “a is a multiple of b”. Proposition 11 ϕ is surjective. PROOF. The elements of the (Boolean) ring A (they are the same as those of the Boolean algebra A) are a linear algebraic combination of the polinomial variables, and: ϕ(P ∧Q) = p ·q ϕ((¬P ∧Q) ∨ (P ∧¬Q)) = p + q. � Proposition 12 ϕ is injective. PROOF. Let us suppose that for two propositions, A and B, we have: ϕ(A) = a, ϕ(B) = b and a = b. As a = b, we have: a|b and b|a. But ϕ preserves the ordering (i.e., → and “is a multiple” do correspond), so: b|a ⇒ A → B; a|b ⇒ B → A and therefore A ↔ B. � 25 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as Remark 2 To be precise, we really consider C/ ↔ and A/ = (Lindenbaum algebras) in order for the relations → and ≤ to be anti-symmetric. Example 8 Let P and Q be the propositional variables of C. Then C has 16 elements and (C,→) is the transitive closure of the diagram in Figure 2. It corresponds in ϕ with (A, “is a multiple”), where A = Z2[p,q]/〈p2 −p,q2 −q〉 and “is a multiple” is the transitive closure of the diagram in Figure 3. 0 � � �� @ @ @R �� �1 PPPq P ∧¬Q ¬P ∧¬Q P ∧ Q ¬P ∧ Q �� �� ���: - � � � � � ��* XXXXXXXz -XXXXXXXz -XXXXXXXz A A A AU B B B B BBN � �� � � � � � � ��� (P ∨¬Q) ∧ (¬P ∨ Q) ¬Q P Q ¬P (P ∧¬Q) ∨ (¬P ∧ Q) �� �� ���: - �� �� ���: XXXXXXXz �� �� ���: - - H H H H H HHj A A A AU B B B B B BN � �� � � � � � � � �� P ∨¬Q ¬P ∨¬Q P ∨ Q ¬P ∨ Q � � ��� ��1 PPPq @ @ @R 1 Figure 2. (C,→) when there are two propositional variables. 0 � � �� @ @ @R �� �1 PPPq p · (1 + q) (1+p)·(1+q) p · q (1 + p) · q �� ���1 - � � � ��3 PPPPPq - PPPPPq - PPPPPq J J J Ĵ A A A A A AU � � ��* � � � � � � �� 1 + q p q 1 + p 1 + (p + q) p + q �� ���1 - �� ���1 PPPPPq �� ���1 - - Q Q Q QQs @ @ @ @R J J J J J Ĵ � � � � � � � �� � � ��* p+̃(1 + q) (1 + p)+̃(1 + q) p+̃q (1 + p)+̃q � � ��� ��1 PPPq @ @ @R 1 Figure 3. (A, “is a multiple”) when there are two polynomial variables. 4.1 Ideals of a Boolean algebra and ideals of a ring Definition 8 Let (R, +, ·) be a commutative ring. A subset I ⊆ R is said to be an ideal of R if and only if: 1) ∀i, i′ ∈ I : i + i′ ∈ I 2) ∀i ∈ I,∀a ∈R : a · i ∈ I (i.e. I is a subring of A such that the product of an element of the ring by an element of the subring always belongs to the subring). 26 An algebraic approach to rule based expert systems Definition 9 Let (R, +, ·) be a commutative ring. Let S 6= ∅, S ⊆ R. The ideal generated by S = {p1, . . . ,pm}, denoted 〈p1, . . . ,pm〉, is the smallest ideal containing S. If S = {q}, the ideal generated by q, 〈q〉, is said to be a principal ideal, and results to be: 〈q〉 = {a ∈A : q|a} Despite the concept of ideal is usually associated with rings, ideals are also defined in Boolean algebras. In such case, the definition is different (based on the lattice order), and only ideals generated by a single element are considered. Definition 10 In the propositional Boolean algebra (C,∨,∧,¬,→), the (principal) ideal generated by Q is defined as EQ = {X ∈C : X → Q} Similarly, in the polynomial Boolean algebra (A, +̃, · , 1+, “is a multiple”), the (principal) ideal gener- ated by q ∈A is: Eq = {a ∈A : a is a multiple of q} Proposition 13 As ϕ preserves the ordering, the ideals of the Boolean algebra (C,∨,∧,¬,→) do corre- spond in ϕ with the ideals of the Boolean algebra (A, +̃, · , 1+, “is a multiple”). Proposition 14 The ideals of the Boolean algebra (A, +̃, · , “is a multiple”) are obviously ideals of the ring (A, +, · , “is a multiple”). Theorem 1 (A, +, ·) is a ring where all ideals are principal ones. PROOF. Let s1, s2, . . . , sn ∈A. We shall prove that 〈s1,s2, . . . ,sn〉 = 〈s1+̃s2+̃ · · ·+̃sn〉 in two steps: i) +̃ is an internal operation in the ideal (because + and · also are internal operations) and therefore: s1+̃s2+̃ · · ·+̃sn ∈ 〈s1,s2, . . . ,sn〉⇒ 〈s1+̃s2+̃ · · ·+̃sn〉⊆ 〈s1,s2, . . . ,sn〉 ii) That s1+̃s2+̃ · · ·+̃sn|si; i = 1, . . . , n can be easily proven. But, then: si ∈ 〈s1+̃s2+̃ · · ·+̃sn〉; i = 1, . . . , n and, consequently, the minimum ideal that contains {s1,s2, . . . ,sn}, is contained in 〈s1+̃s2+̃ · · ·+̃sn〉, i.e.: 〈s1,s2, . . . ,sn〉⊆ 〈s1+̃s2+̃ . . . +̃sn〉. � Remark 3 Let us observe that (A, +, ·) is not a “Principal Ideals Domain” because it is not an integrity domain (see Proposition 7). Corollary 3 As a consequence of Proposition 13, Proposition 14 and Theorem 1, the ideals of the polyno- mial Boolean algebra (A, +̃, · , “is a multiple”) are precisely the ideals of the polynomial ring (A, +, ·). 5 Introductory note about multi-valued propositional logics A simple introduction to the topic can be found in [37]. 27 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 5.1 Kleene’s and Łukasiewicz’s three-valued logic Kleene’s and Łukasiewicz’s multi-valued logics are two of the most commonly used multi-valued logics. We shall begin introducing these three-valued logics, in which a third truth value is considered. In Kleene’s logic this third valued is “undecided”. If an assertion is assigned the truth value “undecided” that means that at present, we can’t assign a truth-value to this conjecture (but the conjecture is either true or false). For instance, if we are reasoning in a medical diagnostic RBES, until we receive the result of a test, its truth value is “undecided”, but it will become either true or false in the future (when we receive the results). Kleene’s A → B is equivalent to ¬A∨B. Two modal unary connectives, necessary (�) and possible (♦) are usually considered in multi-valued logics. Kleene’s three valued logic truth-tables are detailed in Tables 1 and 2 (note that 0 represents “false”, 1 represents “undecided” and 2 represents “true”). It would also be possible to assign numbers to truth values in a different way. For instance, in Kleene’s three valued logic, it is also common to consider that 0 represents “false”, that 2 represents “undecided” and that 1 represents “true”. In such case the corresponding truth tables would obviously change. ¬ 0 2 1 1 2 0 � 0 0 1 0 2 2 ♦ 0 0 1 2 2 2 Table 1. Truth-tables of Kleene’s three-valued logic unary connectives. ∧ 0 1 2 0 0 0 0 1 0 1 1 2 0 1 2 ∨ 0 1 2 0 0 1 2 1 1 1 2 2 2 2 2 → 0 1 2 0 2 2 2 1 1 1 2 2 0 1 2 ↔ 0 1 2 0 2 1 0 1 1 1 1 2 0 1 2 Table 2. Truth-tables of Kleene’s three-valued logic binary connectives. In Łukasiewicz’s logic the third truth value considered is “indeterminate”. If a statement is considered to be indeterminate, it is neither true not false, but indeterminate in a metaphysical sense. All connectives have the same truth tables as in Kleene’s logic, except → and ↔ (see Table 3). → 0 1 2 0 2 2 2 1 1 2 2 2 0 1 2 ↔ 0 1 2 0 2 1 0 1 1 2 1 2 0 1 2 Table 3. Truth-tables of the Łukasewicz’s three-valued logic binary connectives whose truth-tables are different from Kleene’s ones. Remark 4 It is also possible to consider more than three truth values, so that the confidence in the truth- fulness of statements can be graded. Moreover, there are other multivalued logics aside from Kleene’s and Łukasewicz’s. 28 An algebraic approach to rule based expert systems 5.2 Tautological consequence Definition 11 A propositional formula A0 is a tautological consequence of the propositional formulae A1, A2, . . . , Am, denoted {A1,A2, . . . ,Am} |= A0, if and only if whenever the formulae A1,. . . ,Am hold (i.e., take the truth value “true”), then A0 also holds (i.e., takes the truth value “true”). Remark 5 Let us observe that: • in Boolean logic, there are formulae that can only take the truth value “false”, such as P ∧¬P , • in multi-valued logic, there are also formulae that can only take the truth value “false”, such as �P ∧�¬P . • in multi-valued logic, there are formulae that, although can take other truth values different from “false”, they can never take the truth value “true”, such as P ∧¬P (observe that not only �P ∧ �¬P |= P ∧¬P , but also P ∧¬P |= �P ∧�¬P ). 5.3 The isomorphism ϕ in the multi-valued case Let (C,∨,∧,¬,→) be a p-valued logic, where p is a prime number (we require p to be a prime in order for Zp to be a field) and C is the set of propositions that can be constructed using the propositional variables X1, X2, . . . , Xn. If in the residue class ring A = Zp[x1,x2, . . . ,xn]/〈x p 1 −x1,x p 2 −x2, . . . ,x p m −xm〉 adequate polynomial translations of the logical connectives are defined (so that they behave as suggested by the truth-tables), the natural correspondence ϕ, defined as in the Boolean case (Section 4), is also an isomorphism. 6 The main Theorem 6.1 A brief note about Gröbner bases In the early 60’s, both Heisuke Hironaka and Bruno Buchberger independently proved that, for each poly- nomial ideal, a basis completely identifying it always exists. They denoted their bases as standard bases and Gröbner bases (GB) [4, 5], respectively. The latter’s great advantage was that it provided a constructive method (Buchberger’s algorithm). Some GB are particularly important: we call them reduced Gröbner bases. We say that a Gröbner basis is reduced if and only if the leading coefficient of all its polynomials is 1 and we cannot simplify any of its polynomials by adding a linear algebraic combination of the rest of the polynomials in the basis. The input to Buchberger’s algorithm is a polynomial set, a term order (for instance, “total degree” or “pure lexicographical”), and a variable order (for instance, x > y > z) and its output is the ideal’s reduced GB with respect to the specified term and variable orders. The key point is that, once the term order and the variable order are fixed, such a reduced GB completely characterizes the ideal: any ideal has a unique reduced GB. As a consequence we have two key results: i) two sets of polynomials generate the same ideal if and only if their reduced GB are the same, ii) {1} is the only reduced GB for the ideal that is equal to the whole ring. An elementary introduction to the use of GB in algebraic system solving can be found in [33, 34]. For detailed introductions see, for instance, [1, 2, 9, 16, 38]. There are interesting applications in other fields like transportation engineering [32, 30]. 29 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 6.2 Connecting “To be a tautological consequence” with an ideal member- ship The main theorem is the following: Theorem 2 A logical formula, A0, is a tautological consequence of the set of formulae {A1,A2, . . . ,Am} in a certain p-valued modal logic (where p is a prime number) if and only if the polynomial translation of ¬A0 in A = Zp[x1,x2, . . . ,xn]/〈x p 1 − x1,x p 2 − x2, . . . ,x p m − xm〉, ϕ(¬A0), belongs to the ideal 〈ϕ(¬A1),ϕ(¬A2), . . . ,ϕ(¬Am)〉 of A. The key point is that this theorem can take advantage of the two important results stated at the end of the previous section. The proof of this theorem is really long and can be found in [19, 27]. 7 Some introductory notes about RBES A RBES consists of an “input”, an “output”, a “knowledge base” and an “inference engine”. A graphic user interface is frequently provided. The input of a RBES is concerned with the information related to the environment of the RBES. Since the environment may change, this information is also subject to change as times goes by. This information is described by means of a finite number of different atomic propositions: such propositions are termed as input atomic propositions. Input atomic propositions determine what we will call potential facts (see Definition 13). The output of a RBES is concerned with the information inferred by the RBES. It is described by means of a finite number of atomic propositions which we shall call output atomic propositions. The knowledge base (KB) of the RBES is concerned with the information contained in the system, which is used along with the input of the RBES to infer the output of the system. In a RBES based on propositional Boolean logic, this knowledge-base will be mostly represented by a finite set of rules. Remark 6 We shall use hereinafter the following notation: z may mean: • no symbol at all or “negation” (denoted ¬), if the underlying logic is classic Boolean, • no symbol at all, “negation”, “possibility” (denoted ♦), “necessity” (denoted �), or a combination of these symbols, like “necessarily-not” (denoted �¬), equivalent to ¬♦, if the underlying logic is a modal multi-valued one. 7.1 Rules Definition 12 The KB consists of a certain number of logical formulae of the form: ∧ki=1zX[i] −→∨ l j=1zX[j] known as production rules (and usually refereed to as, simply, rules). For example, formula X[1] ∧X[2] ∧¬X[3] ∧♦X[4] ∧¬X[5] −→ X[13] ∨�¬X[17] is read as: “IF X[1] and X[2] and not X[3] and possibly X[4] and not X[5] HOLD, THEN X[13] or necessarily-not X[17] HOLDS”. Let us underline that there is no constraint in the structure of the rules: the ocurrences of “∧” in the consequent and of “∨” in the antecedent of a rule can be avoided (if such symbols ocurred in a rule, the rule could be split into rules without these symbols). 30 An algebraic approach to rule based expert systems 7.2 Potential facts and facts. Integrity constraints Definition 13 A formula A is a potential fact if and only if A ≡ zX[i], where X[i] is an input atomic proposition. Definition 14 That a given set of facts (i.e., a subset of the set of potential facts that is stated as true) is consistent means that, for each propositional variable appearing in the potential facts of the system, X[k], at most one expression of the form zX[k], is chosen. That a given set of facts is maximal means that, for each propositional variable appearing in the potential facts of the system, X[k], exactly one expression of the form zX[k], is chosen. Definition 15 An integrity constraint (IC) is a piece of knowledge added by the experts and expressing that some potential facts cannot hold at the same time. For instance an IC could be: man ∧ pregnant. The negation of the ICs (sometimes denoted “NICs”) have to be added to the KB. 7.3 The inference engine. RBES inconsistency The inference engine (IE) is an automated tool (in our case, a program in the CAS), that verifies consistency and draws tautological consequences from the information contained in the KB. Definition 16 That a formula A be obtained by forward firing of the formulae in the union, G, of a con- sistent set of facts and the set of all production rules and the set of the negation of the integrity constraints of a RBES means that G |= A. Definition 17 A RBES is said to be inconsistent if there is a consistent set of facts verifying that con- tradiction can be obtained by forward firing of these facts and the rules and the negation of the integrity constraints in the KB. Otherwise the RBES is said to be consistent. 8 Knowledge extraction in RBES through the use of an alge- braic inference engine This extension of the algebraic method to RBES was already mentioned in [18] and detailed in [26]. 8.1 A polynomial model for the propositional Boolean algebra associated to RBES The Boolean algebra associated to a RBES whose underlying logic is classic Boolean logic is a structure (C∗,∨,∧,¬,→) where → is the relation obtained applying the rules of logical deduction to the implications of C and the rules, facts and the negations of the integrity constraints of the RBES (consequently, → is not the usual implication), and C∗ is the set of equivalence classes defined by this enlarged equivalence relation ↔ in C. If the rules Rule1, . . . , Rulev , the facts Fact 1, . . . , Fact m and the integrity constraints IC 1, . . . , IC u of a RBES are added as true to the Boolean algebra C of the previous section, the structure obtained is isomorphic to the image of A in the natural surjective homomorphism ψ : A−→A/J where J is the ideal 〈ϕ(¬Rule1), . . . ,ϕ(¬Rulev),ϕ(¬Fact 1), . . . ,ϕ(¬Fact m),ϕ(¬IC 1), . . . ,ϕ(¬IC u)〉 31 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as 8.2 The main theorem adapted to knowledge extraction in RBES The main theorem detailed above can be adapted to treat the logic underlying a RBES: Theorem 3 A formula A0 can be obtained by forward firing of the formulae in the union of a consistent set of facts and the set of all production rules and the negation of the integrity constraints of a RBES, if and only if, the polynomial translation of the negation of A0 belongs to the ideal J of A = Zp[x1,x2, . . . ,xn]/〈x p 1− x1,x p 2 −x2, . . . ,x p n −xn〉, generated by the polynomial translations of the negations of the given potential facts, of the negations of all the production rules, and the integrity constraints of the RBES. Remark 7 Observe that, denoting I = 〈xp1 − x1,x p 2 − x2, . . . ,x p m − xn〉, we could alternatively check that A0 belongs to the ideal I + J of Zp[x1,x2, . . . ,xn]. We usually work this way as most CAS are not prepared to perform computations in a residue class ring. We have extensively used this approach in RBES design and verification. [20] is an interesting example in medicine (regarding techniques for coronary surgery), where we could find a situation that could lead to an inconsistency and was not taken into account by the panel of experts. 9 RBES strong and weak inconsistency. Algebraic interpreta- tion An introduction to verification can be found in [17]. This section summarizes [31]. We shall make the following distinction in the multi-valued case: Definition 18 We shall say that there is strong inconsistency if and only if the conjunction of the facts in a certain consistent set of facts, the rules and the negations of integrity constraints of the RBES, can only take the truth value “false”. Definition 19 We shall say that there is weak inconsistency if and only if all formulae can be obtained by forward firing of the formulae in the union of a certain consistent set of facts, the set of all production rules and the set of the negations of integrity constraints of the RBES. Proposition 15 Independently of the number of truth values (p) of the underlying logic: strong inconsistency ⇒ weak inconsistency. PROOF. If there is strong inconsistency, we have a consistent set of facts {Fact 1, . . . , Fact m} such that: Fact 1 ∧ . . .∧ Fact m ∧ Rule1 ∧ . . .∧ Rulev ∧ NIC 1 ∧ . . .∧ NIC u can only take the truth values false. But {Fact 1, . . . ,Fact m, Rule1, . . . , Rulev, NIC 1, . . . , NIC u} |= Fact 1 ∧ . . .∧ Fact m ∧ Rule1 ∧ . . .∧ Rulev ∧ NIC 1 ∧ . . .∧ NIC u so a contradiction can be obtained by forward firing of the formulae in {Fact 1, . . . , Fact m, Rule1, . . . , Rulev, IC 1, . . . , IC u}, and, therefore, we have weak inconsistency. � 32 An algebraic approach to rule based expert systems Proposition 16 If the underlying logic is Boolean: weak inconsistency ⇒ strong inconsistency. PROOF. If there is weak inconsistency, we have a consistent set of facts {Fact 1, . . . , Fact m} such that any formula follows of the formulae in {Fact 1, . . . , Fact m, Rule1, . . . , Rulev, NIC 1, . . . , NIC u} by forward firing. In particular, {Fact 1, . . . , Fact m, Rule1, . . . , Rulev, NIC 1, . . . , NIC u} |= contradiction But we are considering that the underlying logic is Boolean, so: Fact 1 ∧ . . .∧ Fact m ∧ Rule1 ∧ . . .∧ Rulev ∧ NIC 1 ∧ . . .∧ NIC u → contradiction and therefore Fact 1 ∧ . . .∧ Fact m ∧ Rule1 ∧ . . .∧ Rulev ∧ NIC 1 ∧ . . .∧ NIC u can only take the truth value “false”. � Remark 8 The previous result does not hold if the underlying logic is a multi-valued one. Example 9 Let us consider a small RBES whose KB consists of the rules: R1 :X[1] → X[3] R2 :X[2] →¬X[3] and whose underlying logic is Kleene’s three-valued logic. The potential facts are X[1] and X[2]. If we draw consequences from the maximal set of potential facts {X[1],X[2]}, we obtain X[3] ∧¬X[3]. As this formula can only take the truth values “undecided” or “false”, any formula can be obtained by forward firing of it, and, consequently, by forward firing of the formulae in {X[1],X[2],R1,R2}. Therefore we have weak inconsistency. Meanwhile, we cannot obtain a formula that is always “false” by forward firing of the conjunction of the facts in any consistent set of facts and R1 ∧R2. Therefore, we do not have strong inconsistency. Corollary 4 There is weak inconsistency if and only if 1 ∈ 〈f¬(ϕ(Rules)),f¬(ϕ(Facts)),f¬(ϕ(NICs))〉. But, what is the meaning of strong inconsistency in the polynomial model? Definition 20 The algebraic variety corresponding to the polynomial ideal I is the set of points of the respective affine space which satisfy all polynomials in the ideal I. We shall denote it by V (I). Lemma 1 For any positive prime integer p > 1, we have: i) There are non-zero ideals in Zp[x1,x2, . . . ,xm] whose algebraic variety is the whole affine space. ii) The only ideal in A = Zp[x1,x2, . . . ,xm]/I; I = 〈x p 1 −x1,x p 2 −x2, . . . ,x p m −xm〉 whose variety is the whole affine space is the zero ideal. PROOF. i) For example x · (x − 1) = x2 − x ∈ Z2[x] is not the null polynomial but its value is always 0 in {0, 1}. Therefore 〈x2 + x〉 verifies condition i). The same occurs with x · (x− 1) · (x− 2) ∈ Z3[x]. Its value in {0, 1, 2} is always 0. This construction can be obviously generalized to Zp[x]. 33 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as ii) Case I: Univariate polynomials. If 0, 1, . . . , p− 1 are roots of a polynomial in A = Zp[x]/〈xp −x〉, then the polynomial is either 0 or it can be factored as x · (x− 1) · · · · · (x− (p− 1)) · · · · Therefore, if this polynomial is different from 0, its total degree is greater or equal than p. But in A all polynomials are simplified to polynomials of degree lesser than p, so this polynomial simplifies to 0. Case II: Multi-variate polynomials. It can be proved by applying the previous case to the intersections of the hypersurface by hyperplanes parallel to the “vertical” axis. � Theorem 4 There is strong inconsistency if and only if V (〈ϕ(Rules ∧ Facts ∧ NICs)〉) is the whole (respective) affine space. PROOF. Rules ∧ Facts ∧ NICs can only take the truth value “false” ⇔ ⇔ polynomial ϕ(Rules ∧ Facts ∧ NICs) can only take the value 0 ⇔ ⇔ V (〈ϕ(Rules ∧ Facts ∧ NICs)〉) is the whole (respective) affine space. � But we can express the previous result in terms of ideals instead of algebraic varieties, what makes it easily decidable using Gröbner bases: Corollary 5 There is strong inconsistency if and only if 〈ϕ(Rules ∧ Facts ∧ NICs)〉 is the zero ideal of A = Zp[x1,x2, . . . ,xm]/I. 10 Backward reasoning in rule based expert systems whose underlying logic is classic Boolean logic Details on the topic of this section, our most recent result, can be found in [25]. Based on the classical works of automatic discovery of geometric theorems [14, 23] (a topic already advanced in [8]), we have now adapted this idea to a specific task in RBES whose underlying logic is classic Boolean logic: given a set of facts and a goal, finding a new fact such that if it also held, this goal would be inferred. The main result proven in [25] (Theorem 5 below) shows how it is enough to compute one single reduced GB in order to detect them: Theorem 5 Let A1, . . . , Ak, B ∈C∗ be formulae such that {A1, . . . ,Ak} 6|= B. Let C be a potential fact. Let G be the reduced GB of the ideal 〈ϕ(¬A1), . . . ,ϕ(¬Ak),ϕ(B)〉 + I. We have that: {A1, . . . ,Ak,C} |= B ⇔ ϕ(C) ∈ G Observe that the membership relation in the right hand side of the equivalence refers to a membership to the set of polynomials in the reduced GB, not to the ideal generated by that basis. 11 Implementation in a Computer Algebra System As most CAS include an implementation of GB and of “normal forms” (reductions modulo an ideal), to develop an inference engine in a CAS following the method detailed above (based on checking polynomial ideal memberships) is straightforward and surprisingly brief. 34 An algebraic approach to rule based expert systems 11.1 Polynomial translation of the logical connectives For instance, the polynomial expressions corresponding to the basic logical formulae in Łukasiewicz’s three-valued logic (if 2 is assigned to “true”, 1 to “undetermined” and 0 to “false”) are detailed afterwards. • ¬M is translated into the polynomial: 2 −m • ♦M is translated into the polynomial: 2m2 • �M is translated into the polynomial: m2 + 2m • M ∨N is translated into the polynomial: m2n2 + m2n + mn2 + 2mn + m + n • M ∧N is translated into the polynomial: 2m2n2 + 2m2n + 2mn2 + mn • M → N is translated into the polynomial: 2m2n2 + 2m2n + 2mn2 + mn + 2m + 2 • M ↔ N is translated into the polynomial: m2n2 + m2n + mn2 + 2mn + 2m + 2n + 2 (note that the coefficients of these polynomials belong to Z3). The polynomials corresponding to the different logics can be obtained by solving an algebraic system that is obtained from the truth tables. 11.2 CoCoA implementation These polynomial translations can be input almost directly in CoCoA 4.3 [6, 35], which provides a “Normal Form” command, NF(pol,I), that returns the reduction of polynomial pol modulo ideal I. We have to define the polynomial ring and ask CoCoA to use it: USE Z/(3)[x[1..15]]; and then we can define the ideal MEMORY.I, generated by the expressions x3i −xi (denoting it this way, it is a global variable, and it can be an input to NF): MEMORY.I:=Ideal([x[K_]ˆ3-x[K_] | K_ In 1..15]); and introduce the polynomial expressions for the three-valued connectives of Łukasiewicz modal logic (the operators are prefix ones)1: NEG(M):=NF(2-M,MEMORY.I); POS(M):=NF(2*Mˆ2,MEMORY.I); NEC(M):=NF(Mˆ2+2*M,MEMORY.I); OR(M,N):=NF(Mˆ2*Nˆ2+Mˆ2*N+M*Nˆ2+2*M*N+M+N,MEMORY.I); AND(M,N):=NF(2*Mˆ2*Nˆ2+2*Mˆ2*N+2*M*Nˆ2+M*N,MEMORY.I); IMP(M,N):=NF(2*Mˆ2*Nˆ2+2*Mˆ2*N+2*M*Nˆ2+M*N+2*M+2,MEMORY.I); IFF(M,N):=NF(Mˆ2*Nˆ2+Mˆ2*N+M*Nˆ2+2*M*N+2*M+2*N+2,MEMORY.I); For instance, the CoCoA implementation of the polynomial expressions of a 7-valued modal logic can be found in [27]. Whether A0 is a consequence of a set of formulae or not, can be checked with CoCoA just typing: NF(NEG(A[0]),MEMORY.I+J); (where J is the polynomial ideal generated by the polynomial expressions of the negation of the formulae in the given set). If the output of the command is 0, the answer is “yes”, otherwise the answer is “no”. 1In CoCoA 4.7 these operators would be defined using Define...EndDefine instead of directly :=. 35 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as It can also be directly checked using the Boolean command IsIn, that tests ideal memberships, by typing: NEG(A[0]) IsIn MEMORY.I+J; Whether the RBES is inconsistent or not can also be checked using Gröbner bases by typing in CoCoA: GBasis(MEMORY.I+J); If the output is 1, the RBES is inconsistent; otherwise (the output is normally a large set of polynomials), that set of facts doesn’t lead to an inconsistency. Using IsIn makes it even simpler, as it directly returns “true” or “false”, just typing: 1 IsIn MEMORY.I+J; 11.3 Maple implementation Let us implement the model in Maple. We have developed different implementations in Maple, but the one presented afterwards is the most modern and fastest [29]. In Maple (version ≥ 10) it is possible to define a polynomial ring over a finite field using Maple’s Ore Algebra. In this way, the NormalForm command is computed in such ring. We load the Gröbner and Ore algebra packages first, and then define the list of variables, the polynomial ring and the order that will be used by the GB-related commands: with(Groebner): with(Ore_algebra): SV:=x[1],x[2],x[3]: A:=poly_algebra(SV,characteristic=3): Orde:=MonomialOrder(A,’plex’(SV)): and the ideal I (denoted iI, as I is a reserved word in Maple), using map in order to save time: fu:=v->vˆ3-v: iI:=map(fu,[SV]); 3 3 3 iI := [x[1] - x[1], x[2] - x[2], x[3] - x[3]] We can now define the functions that associate to the logical connectives their polynomial expressions, as done in CoCoA above. NEG :=(m::algebraic) -> NormalForm(2-m,iI,Orde): NEC :=(m::algebraic) -> NormalForm(expand(mˆ2+2*m),iI,Orde): POS :=(m::algebraic) -> NormalForm(expand(2*mˆ2),iI,Orde): ‘&AND‘:=(m::algebraic,n::algebraic) -> NormalForm(expand(2*mˆ2*nˆ2+2*mˆ2*n+2*m*nˆ2+m*n), iI,Orde): ‘&OR‘ :=(m::algebraic,n::algebraic) -> NormalForm(expand(mˆ2*nˆ2+mˆ2*n+m*nˆ2+2*m*n+m+n), iI,Orde): ‘&IMP‘ :=(m::algebraic,n::algebraic) -> NormalForm(expand(2*mˆ2*nˆ2+2*mˆ2*n+2*m*nˆ2+m*n+2*m+2), iI,Orde): ‘&IFF‘ :=(m::algebraic,n::algebraic) -> NormalForm(expand(mˆ2*nˆ2+mˆ2*n+m*nˆ2+2*m*n+2*m+2*n+2), iI,Orde): 36 An algebraic approach to rule based expert systems 12 Description of the shell A GUI was presented at FLINS’2008 conference [24], and a complete shell is now provided. When working with the shell we could distinguish three levels. At the lower level, we provide the CAS code for performing knowledge extraction and verification in different logics (i.e., we provide the algebraic inference engines). For instance, to work with the CAS CoCoA 4.3 in Boolean logic with the propositional variables x1, . . . , x5, the code that contains the algebraic model of the IE should be included in the file named CODE CoCoA.TXT. It would look like: USE Z/(2)[x[1..5]]; MEMORY.I:=Ideal([x[K_]ˆ2-x[K_] | K_ In 1..5]); NEG(M):=NF(1+M,MEMORY.I); OR(M,N):=NF(M+N+M*N,MEMORY.I); ... (similar to the code shown in Section 11.2). Some lines regarding I/O between the GUI and the CAS follow. At the intermediate level, the RBES developer details the concrete potential facts, rules and integrity constraints of a certain RBES (the code must be stored in file RBES.TXT). Example 10 For instance, the production rules and IC of a tiny RBES whose underlying logic is classic Boolean and their CoCoA translations can be (note that in a real RBES there would be dozens of rules): R1 : x[1] → x[2] R1:=IMP(x[1],x[2]); R2 : x[2] ∧¬x[3] → x[4] R2:=IMP(AND(x[2],NEG(x[3])),x[4]); R3 : x[1] ∧x[4] → x[5] R3:=IMP(AND(x[1],x[4]),x[5]); NIC 1 : ¬(x[3] ∧x[5]) NIC1:=NEG(AND(x[3],x[5])); At the upper level, the final user does not have to deal directly with the CAS but with a simple GUI. Both the number of truth values of the logic (e.g., 3) and the desired logic (e.g., Kleene’s) are to be chosen beforehand. Then consistency can be checked and knowledge extraction performed. To do so, a list formed by a proposition and a (consistent) list of facts has to be introduced. The GUI calls the CAS and checks firstly if they lead to inconsistency. Afterwards, it checks if the given proposition follows from the given set of facts. Example 11 Let us consider the tiny RBES of Example 10. If the list of given facts is [x[1],¬x[3]], an inconsistency is not obtained. Moreover, x[4] ∧x[5] follows from them (see Figure 4). This GUI has been implemented in Visual-BASIC and runs on computers using the most common windows-based operating system. Porting it to other operating systems is under consideration now. It can call (at least) the mathematical systems: CoCoA, Maple, Maxima, Singular, Risa-Asir and Octave. 13 Conclusions We believe that this algebraic approach is flexible and powerful (as several logics can be easily and effec- tively simulated) and that the whole shell now put together can be really useful for teaching and for quick RBES designing. This approach has the advantage over that of Kapur-Narendran [15] and Hsiang [13] for Boolean logic (or Alonso et al. [3] and Chazarain et al. [7] in the multi-valued case) of providing an algebraic structure that is isomorphic to the propositional algebra, instead of only performing effective com- putations in logic using Gröbner bases. This is the key for going one step further and obtaining an algebraic structure isomorphic to a RBES based on one of these logics, i.e., from A = Zp[x,y, . . . ,z]/〈xp −x,yp −y, . . . ,zp −z〉 37 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as Figure 4. Screenshot of a GUI for RBES consistency checking and knowledge extraction. to A/J where J is the ideal generated by the negation of the facts, rules and integrity constraint, so that knowledge extraction and verification of RBES can be performed. Acknowledgement. This work was partially supported by the research projects TIN2006-06190 (Government of Spain) and UCM2008-910563 (UCM - BSCH Gr. 58/08, Research Group ACEIA, Spain). We would also like to thank the anonymous referees for their most valuable comments and sugges- tions. References [1] ADAMS, W. W. AND LOUSTAUNAU, P., (1994). An Introduction to Gröbner Bases. Graduate Studies in Mathe- matics 3, AMS. [2] AKRITAS, A. G., (1989). Elements of Computer Algebra with Applications, Wiley - Interscience. [3] ALONSO, J. A. AND BRIALES, E., (1989). Lógicas Polivalentes y Bases de Gröbner, en M. Vide ed., Procs. of the V Congress on Natural Languages and Formal Languages, Barcelona Univ. Press, 307–315. 38 An algebraic approach to rule based expert systems [4] BUCHBERGER, B., (1988). Applications of Gröbner Bases in non-linear Computational Geometry. In J. R. Rice (editor), Mathematical Aspects of Scientific Software, IMA Volumes in Math. and its Applications, 14, Springer- Verlag. [5] BUCHBERGER, B., (2006). Bruno Buchberger’s PhD thesis 1965: An algorithm for finding the basis elemen- tals of the residue class ring of a zero dimensional polynomial ideal, J. Symbolic Comput., 41, 3-4, 475–511. DOI: 10.1016/j.jsc.2005.09.007 [6] CAPANI, A. AND NIESI, G., (1996). CoCoA User’s Manual (v. 3.0b), Depto. de Matemáticas, Universidad de Genova. [7] CHAZARAIN, J.; RISCOS, A.; ALONSO, J. A. AND BRIALES, E., (1991). Many-valued Logic and Gröbner Bases with Applications to Modal Logic, J. Symbolic Comput., 11, 181–194. [8] CHOU, S.-C., (1987). Mechanical Geometry Theorem Proving. Kluwer Academic Publishers. [9] COX, D.; LITTLE, J. AND O’SHEA, D., (1992). Ideals,varieties, and algorithms, Springer-Verlag. [10] HALMOS, P. R., (1974). Lectures on Boolean Algebras, Springer-Verlag. [11] HALMOS, P. AND GIVANT, S., (1998). Logic as Algebra, The Mathematical Asociation of America. [12] HERMES, H., (1963). La teorı́a de retı́culos y su aplicación a la lógica matemática, Conf. Mat. VI., CSIC- Madrid. [13] HSIANG, J., (1985). Refutational Theorem Proving using Term-rewriting Systems, Artificial Intelligence, 25, 255–300. DOI: 10.1016/0004-3702(85)90074-8 [14] KAPUR, D. AND MUNDY, J. L., (1989). Wu’s method and its application to perspective viewing. In D. Kapur and J. L. Mundy, eds., Geometric Reasoning. MIT Press, Cambridge MA, 15–36. [15] KAPUR, D. AND NARENDRAN, P., (1984). An Equational Approach to Theorem Proving in First-Order Predicate Calculus, 84CRD296, General Electric Corporate Research and Development Report (Schenectady, NY, March 1984, rev Dec 1984). Also in A. K. Joshi, ed., Proceedings of IJCAI-85. Morgan Kaufmann, 1146–1153. [16] KREUZER, M. AND ROBBIANO, L., (2000). Computational Commutative Algebra. Springer-Verlag. [17] LAITA, L. M. AND DE LEDESMA, L., (1997). Knowledge-Based Systems Verification. In Kent, J. G. Williams (editors), Encyclopedia of Computer Science and Technology. Marcel Dekker, 253–280. [18] LAITA, L. M.; DE LEDESMA, L.; ROANES-LOZANO, E. AND ROANES-MACÍAS, E., (1995). An Interpretation of the Propositional Boolean Algebra as a k-algebra. Effective Calculus. In J. Campbell, J. Calmet (editors), Proceedings of AISMC-2, Springer-Verlag LNCS 958, 255–263. [19] LAITA, L. M.; ROANES-LOZANO, E.; DE LEDESMA, L. AND ALONSO, J. A., (1999). A Computer Approach to Verification and Deduction in Many-valued Knowledge Systems, Soft Comput., 3, 1, 7–19. DOI: 10.1007/s005000050086 [20] LAITA, L. M.;ROANES-LOZANO, E.; MAOJO, V.; DE LEDESMA, L. AND LAITA, L., (2001). An expert system for managing medical appropriateness criteria based on computer algebra techniques, Comput. Math. Appl., 42, 5, 12, 1505–1522. DOI: 10.1016/S0898-1221(01)00258-9 [21] MENDELSON, E., (1970). Theory and Problems of Boolean Algebras and Switching Circuits, Schaum’s, MacGraw-Hill. [22] MONK, D., (1989). Handbook of Boolean Algebras, North-Holland. [23] RECIO, T. AND VÉLEZ, M. P., (1999). Automatic Discovery of Theorems in Elementary Geometry, J. Automat. Reason., 23, 63–82. [24] ROANES-LOZANO, E.; HERNANDO, A.; LAITA, L. M. AND ROANES-MACÍAS, E., (2008). A Shell for Rule- Based Expert Systems Development Using Groebner Bases-Based Inference. In D. Ruan, J. Montero, J. Lu, L. Martı́nez, P. D’Hondt, E. E. Kerre (editors), Computational Intelligence in Decision and Control. Proceedings of the 8th International FLINS Conference, World Scientific Proceedings Series on Computer Engineering and Information Science 1, 769–774. DOI: 10.1142/9789812799470 0126 39 http://dx.doi.org/10.1016/j.jsc.2005.09.007 http://dx.doi.org/10.1016/0004-3702(85)90074-8 http://dx.doi.org/10.1007/s005000050086 http://dx.doi.org/10.1016/S0898-1221(01)00258-9 http://dx.doi.org/10.1142/9789812799470_0126 E. Roanes-Lozano, L. M. Laita, A. Hernando and E. Roanes-Macı́as [25] ROANES-LOZANO, E.; HERNANDO, A.; LAITA, L. M. AND ROANES-MACÍAS, E., A Groebner Bases- based Approach to Backward Reasoning in Rule Based Expert Systems, Ann. Math. Artif. Intell., to appear. DOI: 10.1007/s10472-009-9147-4 [26] ROANES-LOZANO, E.; LAITA, L. M. AND ROANES-MACÍAS, E., (1995). Maple V in A.I., The Boolean Alge- bra Associated to a KBS, CAN Nieuwsbrief, 14, 65–70. [27] ROANES-LOZANO, E.; LAITA, L. M. AND ROANES-MACÍAS, E., (1998). A Polynomial Model for Many- valued Logics with a Touch of Algebraic Geometry and Computer Algebra, Math. Comput. Simulation, 45, 1 83–99. [28] ROANES-LOZANO, E.; LAITA, L. M. AND ROANES-MACÍAS, E., (2003). Cálculos efectivos en Lógica Proposi- cional Booleana interpretada como un anillo de clases residuales (polinomial) sobre Z2, Boletı́n de la Sociedad “Puig Adam” de Profesores de Matemáticas, 65, 17–42. [29] ROANES-LOZANO, E.; LAITA, L. M. AND ROANES-MACÍAS, E., (2008). A Groebner Bases Based Many- valued Modal Logic Implementation in Maple. In S. Autexier et al. (editors), AISC / Calculemus / MKM 2008, LNAI 5144 Springer-Verlag, 170–183. DOI: 10.1007/978-3-540-85110-3 14 [30] ROANES-LOZANO, E.; MUGA, R.; LAITA, L. M. AND ROANES-MACÍAS, E., (2005). A terminal area topology-independent GB-based conflict detection system for A-SMGCS, RACSAM Rev. R Acad. Cienc. Exactas, Fı́s. Nat. Ser. A Mat., 98, 1-2, 229–237. [31] ROANES-LOZANO, E.; ROANES-MACÍAS, E. AND LAITA, L. M., (1999). Geometric Interpretation of Strong Inconsistency in Knowledge Based Systems. In V. G. Ganzha, E. W. Mayr, E. V. Vorozhtsov (editors), Computer Algebra in Scientific Computing. Proceedings of CASC’99). Springer-Verlag, 349–363. [32] ROANES-LOZANO, E.; ROANES-MACÍAS, E. AND LAITA, L. M., (2000). Railway Interlocking Systems and Groebner Bases, Math. Comput. Simulation, 51, 5, 473–481. DOI: 10.1016/S0378-4754(99)00137-8 [33] ROANES-LOZANO, E.; ROANES-MACÍAS, E. AND LAITA, L. M., (2004). The Geometry of Algebraic Sys- tems and Their Exact Solving Using Groebner Bases, Computing in Science and Engineering, 6, 2, 76–79. DOI: 10.1109/MCISE.2004.1267612 [34] ROANES-LOZANO, E.; ROANES-MACÍAS, E. AND LAITA, L. M., (2004) Some Applications of Gröbner Bases, Computing in Science and Engineering, 6, 3, 56–60. DOI: 10.1109/MCISE.2004.1289309 [35] ROBBIANO, L. ET AL., (2002). CoCoA 4.2 Online Help, (electronic file companying CoCoA v.4.2). [36] STONE, M., (1940). The Theory of Representations for Boolean Algebras, Trans. Amer. Math. Soc., 40, 1, 37– 111. DOI: 10.2307/1989664 [37] TURNER, R., (1984). Logics for Artificial Intelligence, Ellis Horwood. [38] WINKLER, F., (1996). Polynomial Algorithms in Computer Algebra. Springer-Verlag. Eugenio Roanes-Lozano Luis M. Laita Depto. de Álgebra Depto. CC. de la Computación e I.A. Facultad de Educación Facultad de Informática Universidad Complutense de Madrid Universidad Politécnica de Madrid c/ Rector Royo Villanova s/n Campus de Montegancedo 28040-Madrid Boadilla del Monte, 28660-Madrid Spain Spain eroanes@mat.ucm.es laita@fi.upm.es Antonio Hernando Eugenio Roanes-Macı́as Depto. de Sistemas Inteligentes Aplicados Depto. de Álgebra Escuela Universitaria de Informática Facultad de Educación Universidad Politécnica de Madrid Universidad Complutense de Madrid Carretera de Valencia km 7 c/ Rector Royo Villanova s/n 28031-Madrid 28040-Madrid Spain Spain ahernando@eui.upm.es roanes@mat.ucm.es 40 http://dx.doi.org/10.1007/s10472-009-9147-4 http://dx.doi.org/10.1007/978-3-540-85110-3_14 http://dx.doi.org/10.1016/S0378-4754(99)00137-8 http://dx.doi.org/10.1109/MCISE.2004.1267612 http://dx.doi.org/10.1109/MCISE.2004.1289309 http://dx.doi.org/10.2307/1989664 Introduction An introduction to Boolean algebras and Boolean rings Lattices and lattice orders Boolean algebras Boolean rings A polynomial model for propositional logic Building the Taylor-Made Polynomial (Boolean) Ring The polynomial Boolean algebra corresponding to the polynomial (Boolean) ring The Boolean Algebra Isomorphism varphi Ideals of a Boolean algebra and ideals of a ring Introductory note about multi-valued propositional logics Kleene's and Łukasiewicz's three-valued logic Tautological consequence The isomorphism varphi in the Multi-Valued Case The main Theorem A brief note about Gröbner bases Connecting ``To be a tautological consequence'' with an ideal membership Some introductory notes about RBES Rules Potential facts and facts. Integrity constraints The inference engine. RBES inconsistency Knowledge extraction in RBES through the use of an algebraic inference engine A polynomial model for the propositional Boolean algebra associated to RBES The main theorem adapted to knowledge extraction in RBES RBES strong and weak inconsistency. Algebraic interpretation Backward reasoning in rule based expert systems whose underlying logic is classic Boolean logic Implementation in a Computer Algebra System Polynomial translation of the logical connectives CoCoA implementation Maple implementation Description of the shell Conclusions 
work_ml2dtmgkd5fgjimjvj4xtxtzqa	----	Convolutional Neural Network approaches to granite tiles classification Expert Systems With Applications 84 (2017) 1–11 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Convolutional Neural Network approaches to granite tiles classification Anselmo Ferreira a , ∗, Gilson Giraldi b a Shenzhen Key Laboratory of Media Security, College of Information Engineering, Shenzhen University, Nanhai Ave 3688, Shenzhen, Guangdong 518060, PR China b National Laboratory for Scientific Computing (LNCC), Av. Getúlio Vargas 333, Quitandinha, Petrópolis, Rio de Janeiro 22230 0 0 0, Brazil a r t i c l e i n f o Article history: Received 1 February 2017 Revised 26 April 2017 Accepted 27 April 2017 Available online 4 May 2017 Keywords: Granite classification Convolutional Neural Networks Deep learning a b s t r a c t The quality control process in stone industry is a challenging problem to deal with nowadays. Due to the similar visual appearance of different rocks with the same mineralogical content, economical losses can happen in industry if clients cannot recognize properly the rocks delivered as the ones initially pur- chased. In this paper, we go toward the automation of rock-quality assessment in different image reso- lutions by proposing the first data-driven technique applied to granite tiles classification. Our approach understands intrinsic patterns in small image patches through the use of Convolutional Neural Networks tailored for this problem. Experiments comparing the proposed approach to texture descriptors in a well- known dataset show the effectiveness of the proposed method and its suitability for applications in some uncontrolled conditions, such as classifying granite tiles under different image resolutions. © 2017 Elsevier Ltd. All rights reserved. 1 i u b d m v c d d t c a s a i t q G m g i b u t p t e w a n a B t t s t i b a d a p h 0 . Introduction Natural Stone is the first building material used by the human- ty and it continues to be used on new structures. Among the nat- ral stones, the granite has received a particular interest due to its eauty and strength. Granite is an intrusive igneous rock widely istributed throughout Earth’s crust at a range of depths up to 31 iles (50 km) according to University of Tennessee (2008) . Since there are different colors of granite, their denomination aries depending on the country. Even with some standards being reated such as the one from the European Committee for Stan- ardization (CEN) (2009) and, more recently, from the Stone In- ustry SA Standards Board (2015) , there are some cases on which he visual appearance of different granites with same mineralogi- al content may not differ significantly, making it time consuming nd economically expensive to the stone industry to check, slab by lab, if the product colors to be delivered to clients are the same s the colors purchased. Most of the times, the quality assessment s done by human-skilled professionals in a subjective procedure hat can fail. The international operations of stone industries re- uire this procedure to be faster and more accurate. Automated systems can play an important role in this scenario. ranite companies are interested in computer vision allied with achine learning procedures able to identify patterns on granite ∗ Corresponding author. E-mail addresses: anselmo@szu.edu.br , anselmo.ferreira@gmail.com (A. Ferreira), ilson@lncc.br (G. Giraldi). d t b r ttp://dx.doi.org/10.1016/j.eswa.2017.04.053 957-4174/© 2017 Elsevier Ltd. All rights reserved. mages, using these patterns to sort and grade granite products ased on their visual appearance, helping in this way the prod- ct traceability and warehouse management. Also, an on-site tool hat can help to identify the real type of a product can help solving ossible misunderstoods between customers and manufacturers at he time of a delivery. This challenge has received some attention by the scientific lit- rature in the past years, and some studies of the application of ell-known colors and texture descriptors were performed, such s in the works made by Kurmyshev, Sánchez-Yáñez, and Fer- ández (2013) , Bianconi and Fernández (2006) , Lepisto, Kunttu, nd Visa (2005) , Araújo, Martínez, Ordóñez, and Vilán (2010) and ianconi, González, Fernández, and Saetta (2012) . However, most of hem are general purpose descriptors and were created for a con- rolled scenarios, where the slabs used for experiments have the ame size and images acquired have the same resolution. Descrip- ors applied on granite slabs classification to date generate what s called hand-crafted features created by feature engineering, built ased on a behavior that is expected to happen in all granite im- ges. Finally, there is no study regarding the application of these escriptors in slabs pieces. In this paper, we go beyond the use of hand-crafted features nd propose what is, as far as we know, the first data-driven ap- roach to identify granite patterns. The proposed technique uses eep Convolutional Neural Networks (CNNs) with different archi- ectures, trained to classify the intrinsic patterns of granite tiles ased on texture and color. Instead of designing a very deep neu- al network to be applied on high resolution images, a process that http://dx.doi.org/10.1016/j.eswa.2017.04.053 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2017.04.053&domain=pdf mailto:anselmo@szu.edu.br mailto:anselmo.ferreira@gmail.com mailto:gilson@lncc.br http://dx.doi.org/10.1016/j.eswa.2017.04.053 2 A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 a f m q d o c s t p k l p s t e t w o d 3 e c c t e r t H t r a t t f e n a f a s f requires too many layers in the architecture and also thousands of images to train the network, our approach uses lightweight neu- ral networks on small patches of granite images, taking into ac- count the majority voting of patches classification for the images classification, making the proposed approach able to classify slabs of different sizes, image resolutions and even slab pieces. Experi- ments comparing our approach against some hand-crafted descrip- tors and pre-trained networks show the effectiveness of the pro- posed technique. In summary, the main contributions of this paper are: 1. Design and development of ad-hoc CNNs for granite tiles clas- sification. 2. Application of CNNs on multiple image data, represented by small patches located in regions of interest from granite slab images, to learn features that lead to high recognition accuracy by using majority voting on patches classification. 3. Validation of the proposed methodology with other descriptors not used before for granite classification. The remainder of this paper is organized as follows: Section 2 discusses related work about color and texture de- scriptors in the literature and also some studies applying them to the granite classification. Section 3 shows the basic concepts of CNNs, which are necessary to understand the proposed approach. Section 4 presents our approach for granite color classification and Section 5 reports all the details about the experimental methodology used for validating the proposed method against the existing counterparts in the literature. Section 6 shows the performed experiments and results and Section 7 reports our final considerations and proposals for future work. 2. Related work The problem of classifying materials by their type can be re- garded as a problem of classifying textures, colors and geometrical features. To create solutions in this regard, the scientific literature focused on applying computer-vision based approaches allied with machine learning to grade different materials. In the specific case of granite rocks, Ershad (2011) presented a method based on Primitive Pattern Units, a morphological oper- ator which is applied to each color channel separately. Statistical features are then used to discriminate different classes of natural stone such as granite, marble, travertine and hatchet. Kurmyshev et al. (2013) used coordinated clusters representation (CCR) for classifying granite tiles of the type “Rosa Porriño”. Bianconi and Fernández (2006) employed different Gabor filter banks for gran- ite classification. Lepisto et al. (2005) used a similar approach, but applied in each color channel separately. Araújo et al. (2010) em- ployed spectrophotometer to capture spectral data at different re- gions of interest in granite tiles and used Support Vector Machines to grade them. Fernández, Ghita, González, Bianconi, and Whelan (2011) studied how common image descriptors such as Local Bi- nary Patterns, Coordinated Clusters Representation and Improved Local Binary Patterns acted in granite image classification after ro- tation conditions. Most approaches described in literature are based on textural features alone. According to the study of Bianconi et al. (2012) this is somewhat surprising, since the visual appearance of granite tiles strongly depends on both texture and color. To this end, they per- form a study using several textures and color descriptors in five different classifiers, showing that combining color and texture fea- tures and classifying them on Support Vector Machines classifiers outperforms previous methods based on textural features alone. Finally, in a recent work, Bianconi, Bello, Fernández, and González (2015a) investigated the problem of choosing the ade- quate color representation for granite grading. They discussed pros nd cons of different color spaces for granite classification by per- orming experiments using very simple color descriptors: mean, ean + standard deviation, mean + moments from 2nd to 5th, uartiles and quintiles of each color channel. They showed that, epending on the classifier used, some color spaces are better than thers for classification, such as Lab and Luv spaces for the linear lassifier. As can be noticed, solutions presented for granite color clas- ification are based only on feature engineering. In other words, hese approaches are driven by patterns that are supposed to hap- en over all investigated images and also require different expert nowledge. Veering away from these methods, we propose a deep earning based approach, which extracts meaningful discriminative atterns straight from the data by using small image patches in- tead of using ordinary feature engineering in whole images. To do his, our approach exploits the advantages of back-propagation of rror used in Convolutional Neural Networks to automatically learn hese discriminant features present on the image textures. Before e discuss our proposed method to perform granite grading based n deep learning, it is worth discussing some basic concepts about eep neural networks in the next section. . Basic concepts Convolutional Neural Networks have been attracting a consid- rable attention by the computer vision research community re- ently, mainly because of their effective results in several image lassification tasks, outperforming even humans in certain situa- ions according to He, Zhang, Ren, and Sun (2015) , winning sev- ral image classification challenges, such as the ImageNet image ecognition contest as described by Krizhevsky, Sutskever, and Hin- on (2012) . Pioneered by the work of Lecun, Bottou, Bengio, and affner (1998) , CNNs are regarded as a deep learning application o images and, as so, they simulate the activity in layers of neu- ons in the neocortex, the place where most of thinking happens, ccording to Hof (2013) . So, the network learns to recognize pat- erns by highlighting in its layers the edges and pixel behaviors hat are commonly found in different images. The main benefit of using CNNs with respect to traditional ully-connected neural networks is the reduced amount of param- ters to be learned. Convolutional layers made of small size ker- els allow an effective way of extracting high-level features that re fed to fully-connected layers. The training of a CNN is per- ormed through back-propagation and stochastic gradient descent s Rumelhart, Hinton, and Williams (1986) describes. The misclas- ification error drives the weights update of both convolutional and ully-connected layers. The basic layers of a CNN are listed below: 1. Input layer : where data is fed to the network. Input data can be either raw image pixels or their transformations, whichever better emphasize some specific aspects of the image. 2. Convolutional layers : contain a series of filters with fixed size used to perform convolution on the image data, generating what is called feature map . These filters can highlight some pat- terns helpful for image characterization, such as edges, textures, etc. 3. Pooling layers : these layers ensure that the network focuses only on the most important patterns. They summarize the data by sliding a window across the feature maps, applying some lin- ear or non-linear operations on the data within the window, such as local mean or max, reducing the dimensionality of the feature maps used by the following layers. 4. Rectified Linear Unit (ReLU) : ReLU layers are responsible for ap- plying a non-linear function to the output x of the previous layer, such as f (x ) = max (0 , x ) . According to Krizhevsky et al. (2012) , they can be used for fast convergence in the training of A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 3 Fig. 1. Common architecture arrangement of a CNN. The input image is transformed into feature maps by the first convolution layer C1. A pooling stage S1 reduces the dimensions across the feature maps. The same process is repeated for layers C2 and S2. Finally a classifier is trained with the data generated by layer S2. r d C f d t t ( f 4 s r s n l n d 4 c p c I c C w t i 4 i a i d C f i o c i c C t i n 4 e p i b a B f t a l b n f 5 u l o t p p CNNs, speeding-up the training as they deal with the vanishing gradient problem by keeping the gradient more or less constant in all network layers. 5. Fully-connected layers : used for the understanding of patterns generated by the previous layers. Neurons in this layer have full connections to all activations in the previous layer. They are also known as the inner product layers. After trained, transfer learning approaches can extract features in these layers to train another classifier. 6. Loss layers : specify how the network training penalizes the de- viation between the predicted and true labels and is normally the last layer in the network. Various loss functions appropriate for different tasks can be used: Softmax, Sigmoid Cross-Entropy, Euclidean loss, among others. Fig. 1 depicts one possible CNN architecture. The type and ar- angement of layers vary depending on the target application. Although very powerful at representing patterns present in the ata, the main drawback of deep learning is the fact that common NNs normally need thousands or even millions of labeled data or training. This is an unfeasible condition in many applications ue to the lack of training data and to the amount of time needed o train a model. In this work, we present an alternative approach hat deals with this requirement by considering using image tiles multiple data) of the input images in the networks, as described urther in Section 4 . . Proposed method In this paper, we propose a methodology for granite tiles clas- ification through the application of CNNs. The proposed pipeline elies solely on the data to learn the texture patterns therein. We olve the CNN requirement of large training sets by applying the eural network on small patches of images, this way the CNN will earn discriminative features, instead of relying on feature engi- eering. Fig. 2 shows the pipeline of the proposed approach. We iscuss each step of this pipeline in the next paragraphs. .1. Conversion to grayscale As some proposed CNNs in this paper include networks that lassify grayscale images to identify textures, in this optional pre- rocessing step, the image is converted to gray-level, removing olor information from the analysis. Given an input RGB image , with one pixel represented by I(i, j) = (R (i, j) , G (i, j) , B (i, j) , its onversion to graylevels happens using Eq. (1) below: (i, j) = 0 . 2989 R (i, j) + 0 . 5870 G (i, j) + 0 . 1140 B (i, j) , (1) here C is the image converted to grayscale and used as input to he network and R, G and B are the color channels from the initial mage. .2. Patching In the first step of our proposed approach (labeled as step A n Fig. 2 ), the image is divided into tiles of interest. These tiles re out of image borders (no artificial pixel filling is done in the nput images) and will be used as input to the CNNs. This proce- ure is useful for the following reasons: (i) generate more data to NN training (ii) use majority voting in image tiles classification or testing a given input image and (iii) possibility to use different mage resolutions to train and test the CNN, as only image tiles f interest will be used in the network. This is an important pro- edure because, normally, CNNs are designed for given resolution nput images. However, this is not necessary in our approach be- ause only small granite blocks are used as input. In this step the NN is investigating several regions of interest (which we call mul- iple data) to classify the entire image. We use 28 × 28 grayscale mages and 32 × 32 color images in our CNNs, described in the ext subsection. ( Figs. 5 and 6 ) .3. Recognition by Convolutional Neural Networks In the second step, labeled as step B in Fig. 2 , we use differ- nt Convolutional Neural Networks to recognize patterns of image atches. For this, in a training step, we apply tiles from training mages through the networks to estimate the filter weights for a etter classification. Then, we do what is called transfer learning s described by Johnson and Karpathy (2016) and Yosinski, Clune, engio, and Lipson (2014) , using the already trained network as a eature extractor in the train and test stages. To do that we extract he last layer (the softmax layer) of the already trained network nd then the new output will be feature vectors generated by the ast convolutional layer. Finally, these image characterizations will e used by another classifier in the training and testing steps. The umber of networks used are four and their architectures are as ollows. • MNIST1 network: based in a design used in the MNIST dataset digit recognition challenge ( VLFEAT, 2016 ), it has an input layer which requires 28 × 28 grayscale input images and is com- posed by eight other layers: four convolutional layers, two pool- ing layers, one RELU layer and one fully-connected layer. • MNIST2 network: we extended MNIST1 network, creating a new one composed by eleven layers: one input layer, five convolu- tional layers, three pooling layers, one RELU layer and a final fully connected layer. • MNIST3 network: we extended even more our initial MNIST1 network, creating a new one composed by thirteen layers: one input layer, six convolutional layers, four pooling layers, one RELU layer, and a final fully connected layer. • CIFAR network: based in a design used in the CIFAR image recognition challenge ( VLFEAT, 2016 ), it has an input layer which requires 32 × 32 RGB input images and is composed by twelve other layers: five convolutional layers, three pooling lay- ers, four RELU layers and one fully-connected layer. These networks will generate, respectively, feature vectors with 00, 256, 256 and 64 dimensions respectively, which we aim to se in another classifier. Fig. 3 shows the learned filters in the first ayer of each network. Figs. 4 –7 show the output of these filters n networks MNIST1, MNIST2, MNIST3 and CIFAR respectively and heir power to discriminate by convolutions some granite blocks resent on the dataset built in the work of Bianconi et al. (2015a) . Additionally to the use of these networks individually, we pro- ose the fusion of them using for that two approaches: 1. Early fusion: The feature vector of a block will be the fusion of descriptions (feature vectors) from the three networks. Con- catenating such feature vectors will yield a final feature vector 4 A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 Fig. 2. Pipeline of the proposed approach based on Convolutional Neural Networks for granite classification. Fig. 3. Filter weights of the first layer from (a) MNIST1 network (b) MNIST2 network (c) MNIST3 network and (d) CIFAR network. a M n of 500 + 256 + 256 = 1012 dimensions, which will be used in a given classifier. 2. Late fusion: the classification of a tile will happen by majority voting of the three classifiers outputs, fed by the feature vectors generated by the three networks applied on that tile. a Figs. 8 and 9 show, respectively, the early and late fusion pproaches pipelines for granite image classification using the NIST-based CNNs considered in this paper. If the first approach (early fusion) is chosen, the next step is ormalizing (or scaling) the concatenated feature vectors. There re several approaches to do this, but the one we choose is the A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 5 Fig. 4. MNIST1 first layer convolutions used to identify (a) “Giallo Veneziano” granite block (b) “Sky Brown” granite block and (c) “Verde Oliva” granite block. Fig. 5. MNIST2 first layer convolutions used to identify (a) “Giallo Veneziano” granite block (b) “Sky Brown” granite block and (c) “Verde Oliva” granite block. Fig. 6. MNIST3 first layer convolutons used to identify (a) “Giallo Veneziano” granite block (b) “Sky Brown” granite block and (c) “Verde Oliva” granite block. Fig. 7. CIFAR first layer convolutons used to identify (a) “Giallo Veneziano” granite block (b) “Sky Brown” granite block and (c) “Verde Oliva” granite block. Convolutions are represented in gray for a better visualization. s t | w e c V U i i f m a 4 g S F implest, which is dividing each vector by its norm. Given a fea- ure vector � V , its norm p is calculated as: | � V || p = ( n ∑ i =1 | � V (i ) p | ) 1 p , (2) here i is the i th vector element and n is the number of vector lements. For our application, we choose p = 2 and the norm generated is alled Euclidean Norm . The final feature vector � V f is � f = � V || � V || 2 . (3) sing this approach the feature vectors components will be scaled n such a way that each vector magnitude is always one. This is mportant because, as the feature vectors to be concatenated come rom different sources, the range of all features should be nor- alized so that each feature contributes approximately proportion- tely to the final feature vector. .4. Image patches classification In the training and testing steps, we apply the feature vectors enerated by the previously trained networks in another classifier. o, in the image patches classification step (labeled as step C in ig. 2 ), we choose the 1st Nearest Neighbor (1NN) classifier to clas- 6 A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 Fig. 8. Pipeline of the early fusion of Convolutional Neural Networks for granite classification. In this fusion approach, three pre-trained networks with different architectures (MNIST1, MNIST2 and MNIST3) are applied in the same tiny input data (blocks), generating feature vectors that are concatenated and normalized to classify a given 28 × 28 block using information from these three networks. The final classification for an image uses the majority voting of its blocks. Fig. 9. Pipeline of the late fusion of Convolutional Neural Networks for granite classification. In this fusion approach, three pre-trained networks with different architectures are applied in the same tiny input data (blocks), extracting feature vectors and classifying them individually. The final classification of a block is the majority voting of labels generated by the three networks. The same happens on all the classified blocks to classify the whole image. i c w 5 s m s o sify the tiny image blocks from input images. Basically, this classi- fier works by assigning a testing sample to the class of its nearest neighbor defined in a training step. In the end of this process, each valid block from the image is classified. 4.5. Image classification After the individual patches classification, in step D in Fig. 2 ) we will classify the image by analyzing, in the blocks classifi- cation, which is the most predicted class. Given a vector � x = { out put , . . . , out put } which contains the classification of b blocks 1 b n the image, the predicted class of a granite image G is lass (G ) = mode ( � x ) , (4) here mode ( � x ) is the most predicted class in vector � x . . Experimental setup Before we discuss our experimental results, we present in this ection the materials and methods used to compare the proposed ethod against its counterparts in the literature. In the following ubsections we show the granite tile dataset used, the methodol- gy and metrics used in the experiments to assess the classifica- A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 7 t t 5 B 1 p c i i a w p T t v t s q o w t 5 n m n w t a D i d a a a a b i g 1 i fi r b ( m t b c a t b t c t w t A w y a m fi a g t r t d p t e w e w j p c l e o e 5 b a t w m w i f p t v n p f { w T i t v t m W a ion performance, parameters used in our proposed approach and he state-of-the art implementations used. .1. Image dataset We used the same granite tiles dataset applied in the work of ianconi et al. (2015a) . It contains 10 0 0 RGB images with 150 0 × 500 dimensions, subdivided in 25 classes containing 40 images er class. The first 100 images were acquired naturally by a spe- ific scanner. The other 900 were created by rotating the initial 100 mages in 10 °, 20 °, 30 °, 40 °, 50 °, 60 °, 70 °, 80 ° and 90 °. To be suitable for our proposed approach, we subdivide these mages in 28 × 28 valid blocks ( i.e. , outside image borders) to be pplied on MNIST1, MNIST2 and MNIST3 networks, creating this ay 2809 valid blocks per image. So, the dataset used in the ex- eriments contains 2809 × 10 0 0 = 2809 . 0 0 0 tiny grayscale images. o use the proposed CIFAR network on RGB images we subdivide he dataset images on 32 × 32 valid blocks, creating this way 2116 alid RGB blocks per image and a total of 2116 × 10 0 0 = 2116 . 0 0 0 iny color images. Doing these procedures we artificially create a great amount of mall images to be used in small networks, eliminating the re- uirement of big networks applied in a big amount of high res- lution images. By subdividing the test image into these blocks, e can increase classification accuracy by doing majority voting of hese blocks, classifying a big image by its small parts. .2. Methodology and metrics We validate the proposed method using two experimental sce- arios: assessing the performance of classifying whole images (by ajority voting of blocks) and also small blocks. For the first sce- ario, we consider a 5 × 2 cross-validation protocol, on which e replicate the traditional 1 × 2 cross-validation protocol five imes (thus 5 × 2). In each of them, we divided the set of im- ges D randomly into equal subsets of images D 1 and D 2 , on which 1 ∪ D 2 = D and D 1 ∩ D 2 = ∅ . Then, the classifier is trained with D 1 mages as input and tested on D 2 images and then the inverse is one. If the process is repeated five times, metrics can be reported fter 10 rounds of experiments. The experimental results using our pproaches will be in terms of high resolution images classification ccuracy after majority voting of low resolution image blocks. Using the 5 × 2 cross validation in our scenario, if our proposed pproach acts on 2889 28 × 28 valid (outside borders) grayscale locks and 2116 32 × 32 valid RGB blocks from each of the 10 0 0 mages in our dataset, we will end up with 500 × 2809 = 1404 . 500 rayscale blocks (for MNIST based networks) and 500 × 2116 = 058 . 0 0 0 RGB blocks (for the CIFAR based network) used for train- ng the classifier, and the same number of blocks to test the classi- er. This happens in all 10 rounds of experiments. So, our training atio will always be 50% of the data and the other 50% of data will e the testing ratio. According to a study conducted by Dietterich 1998) , the 5 × 2 cross-validation is considered an optimal experi- ental protocol for learning algorithms. For the second experimental scenario, we show the classifica- ion performance of image blocks. For this case, we used one com- ination of training and test images. We intend to show using this onfiguration how the proposed approaches and some state of the rt behave on classifying only small image resolution images con- aining granite patterns. In this case results are reported on block asis (without majority voting). We used the accuracy method to evaluate the performance of he algorithms tested. In a multi-class problem with c classes, the lassification results may be represented in a c × c confusion ma- rix M . In this case, the main diagonal contains the true positives hile the other entries contain either false positives or false nega- ives. So, the accuracy of an experiment round is calculated as ccuracy = ∑ c i =1 M(i, i ) ∑ c i =1 ∑ c j=1 M(i, j) . (5) here i and j are line and column indexes of M , respectively. In the 5 × 2 cross validation protocol, one confusion matrix is ielded per experiment. Therefore, we present results by averaging ccuracies from these matrices. The other set of metrics shown in the experiments are com- only applied in CNN image recognition tasks and will be used to nd the number of epochs to train our proposed networks. These re called top-1 and top-5 errors and are widely used in the Ima- eNet Large Scale Visual Recognition Challenge and also for other asks as reported by Russakovsky et al. (2015) . Commonly, CNNs eturns for one image i the j most probable classes { c i 1 , . . . , c i j } in he output of its last layer as the prediction for that image. If we efine c ij the prediction of an image i and C i its ground truth, the rediction is considered correct if c i j = C i for some j . If we define he error of a prediction d i j = d(c i j , C i ) as 0 if c i j = C i and 1 oth- rwise, the error of an algorithm is the fraction of test images on hich the algorithm makes a mistake,: rr = 1 N N ∑ i =1 min j (d(c i j , C i )) , (6) here N is the total of images used in the validation of a CNN. The difference of top-1 and top-5 errors is the value used for : in top-1 j = 1 and top-5 j = 5 . In other words, top-1 error com- ares if the top class (the one having the highest probability, or i 1 ) is the same as the target label C i and top-5 scores if the target abel is one of the top 5 predictions (the 5 ones with the high- st probabilities, or { c i 1 , . . . , c i 5 } ). The top-5 error is normally used nly for images with more than one object and will not be consid- red in our scenario (we use only top-1 error). .3. Implementation aspects of the proposed approaches To implement our proposed approaches we used the CNNs li- rary for MATLAB available at VLFEAT (2016) . We used the CIFAR nd MNIST1 architectures, training them from scratch. We also ex- ended layers from network MNIST1, creating this way new net- orks MNIST2 and MNIST3. To create feature extractors, we re- ove the last layer from the trained networks and feed them again ith training images, generating training feature vectors to be the nput for the 1NN classifier. This same last process happens again or testing images. For the early fusion approach, applied only to grayscale in- ut images networks (MNIST related CNNs), we used the three rained networks from scratch to extract train and test feature ectors, concatenating the vectors from these three networks and ormalizing them by Eq. 3 . Finally, these vectors are the in- ut of the 1NN classifier and, to classify the image we per- orm majority voting of blocks. We denominate this approach as M NIST 1 , M NIST 2 , M NIST 3 } _ concat . Finally, we test the complementarity of the three grayscale net- orks (MNIST related networks), using majority voting for a block. o do this we apply, in the test phase, each network individually n the same block. The three feature vectors will be then fed into hree independent 1NN classifiers, previously trained with feature ectors from their corresponding CNNs. The result for a block is he majority voting of the three 1NN classifications. Then, a final ajority voting of blocks will define the class of the test image. e label this approach as { M NIST 1 , M NIST 2 , M NIST 3 } _ v ote . The MNIST based networks were trained using batches of im- ges containing 100 images, with learning rate fixed on 0.001. We 8 A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 e a t n c a i t g l 6 g i e m fi fi I a t a d c w a n t ( 1 f F v 1 b i o t i i o a c i t 6 p h t i t M p c r used the stochastic gradient descent with momentum equals to 0.9 and weighting decay equals to 0.0 0 05 without dropout. The CI- FAR network was trained using batches of images containing 100 images, with stochastic gradient descent and momentum equals to 0.9. The weighting decay is 0.0 0 01 without dropout. Upon accep- tance of this paper, all the source code of the proposed approaches will be available at GitHub 1 5.4. Baselines To compare our approach against the state of the art we initially chose ten texture descriptors to be used as baselines, some of them were already used for Granite Image classification and others were applied for other texture recognition applications. The first five baselines approaches can be regarded as general purpose texture descriptors used in several applications. The first three chosen are the statistics of Gray-Level Co-occurrence Matrices (GLCMs) from Haralick, Shanmugam, and Dinstein (1973) (we label this approach as GLCM in the experiments), the Local Binary Patterns (we label this approach as LBP in the experiments) from Ojala, Pietikäinen, and Harwood (1996) and the Histogram of Gradients (we label this approach as HOG in the experiments) from Dalal and Triggs (2005) . The GLCM texture description approach build statistics calculated over matrices of neighborhood relations only in given directions and offsets (distances between pixels); LBP can be regarded as a histogram of neighborhood relations between a pixel and its eight neighbors and HOG is a histogram of gradient orientation on re- gions of interest of images. The other two descriptors are the Dom- inant Local Binary Patterns of Bianconi, González, and Fernández (2015b) (labeled as DLBP in the experiments) and the Rotation In- variant Co-Occurrences of patterns from González, Fernández, and Bianconi (2014) without feature selection (labeled as C ri 1 in the ex- periments). These two last approaches were validated in the same granite image dataset we are using, which was first presented in the paper of Bianconi et al. (2015a) . Other five descriptors used as baselines in the experiments were originally proposed for specific texture description applica- tions and were never used for texture characterization of granite tiles. The first three approaches are based on the Convolutional Texture Gradient Filter (CTGF), proposed by Ferreira, Navarro, Pin- heiro, dos Santos, and Rocha (2015) to identify texture patterns of printed letters to attribute the laser printer source of a doc- ument. This descriptor measures the texture of an image as his- tograms of convolved textures in low gradient areas, using for the convolution matrices of different sizes. We label approaches based on CTGF in the experiments as CTGF 3 × 3 , CTGF 5 × 5 and CTGF 7 × 7 . The other two approaches came from the same work of Ferreira et al. (2015) . These are multidirectional extensions of GLCMs (we call this GLCM − MD in the experiments) and multi-directional and multi-scale extensions of GLCMs (called GLCM − MD − MS). Finally, we also used existing pre-trained CNNs to compare against our networks trained from scratch. These are the origi- nal networks used in the CIFAR general-purpose image recognition challenge and MNIST digit recognition challenge and are available to download in the MatConvNet library from VLFEAT (2016) . As we do with our approach, we removed the last layer and applied the feature vectors in the nearest neighbor classifier. We call these ap- proaches CIFAR original and MNIST original , respectively. 6. Experiments Now we focus on showing experimental results. For this, we start applying the methodology and calculated the metrics dis- cussed in Section 5.2 using the dataset showed in Section 5.1 . All 1 www.github.com/anselmoferreira/granite- cnn- classification . T t p xperiments were performed on a cluster node with 11 processors nd 131 GB RAM and graphics card NVIDIA GeForce 210. This sec- ion is described as follows: firstly, we show how we choose the umber of epochs used to train our networks. Then, we start the omparison of our proposed approaches against the state of the rt techniques described in Section 5.4 in two scenarios: using big mages and small blocks. These two last experiments were chosen o show the effectiveness of the proposed approaches to classify ranite rocks images of different resolutions (very high and very ow). .1. Defining epochs to train CNNs Before start training our proposed CNNs from scratch using ranite images, the number of epochs to train the network, which s the number of one forward and one backward pass of all training xamples through the network, must be defined. For this experi- ental scenario we show the result of classification of using the rst combination of train and validation using for that the classi- er attached to the end of the network ( i.e., a soft-max classifier). n this case, we further subdivide each image on small-sized blocks nd use them to train and validate the classifier. One natural solu- ion would be using the whole image containing the granite tile s the input for CNNs, but choosing this solution would require eeper networks, which will require a larger amount of data, more omputational time and more memory resources to train the net- orks. Using smaller areas as input do not require as many layers s using the whole image and also can lead to a faster learning of etwork parameters and weights. Using our proposed procedure of classifying small blocks, we hen found the number of epochs to be used in each network MNIST1, MNIST2, MNIST3 and CIFAR) as the ones with less top- validation error ( valtop1e ) after 18 epochs and are, respectively or each network, 10, 15 and 15 epochs as Fig. 10 shows. For CI- AR network, we choose the number of epoch with the less top-1 alidation error ( valtop1e ) after 150 epochs, which is found in the 42th epoch. As we used so many epochs in our CIFAR seek for est validation epoch, it is not so clear to see its lowest top1-error n the graph, so we decide to exclude it from Fig. 10 for the sake f clarity. After finding the number of epochs to train our feature extrac- or, we find the model used to extract feature vectors by first train- ng the networks using the epochs found in this validation exper- ment. Then, we can use these networks to extract feature vectors f images, by applying training and testing images in the networks nd use the output of the last but one layer to feed a multiclass lassifier based on the nearest neighbor classification, as described n Section 4.4 . The following experiments of this paper consider his scenario and will be discussed as follows. .2. Comparison against baselines in high resolution images Now we focus on the experiments comparing our proposed ap- roaches against the state of the art considering classification of igh resolution images of the dataset considered. Table 1 shows he results. For this task, our approaches (highlighted in light gray n this table) perform the classification of test images after doing he majority voting of the 2809 valid 28 × 28 grayscale blocks for NIST-based networks and 2116 valid 32 × 32 RGB blocks in the roposed CIFAR-based network. As can be seen from Table 1 , our deeper network applied on olor tiles, called CIFAR , is the best network to classify granite ocks, showing perfect detection in all 10 rounds of experiments. his happens because the layers in this network better decompose he color information from these images, highlighting important atterns to be used in the classification. http://www.github.com/anselmoferreira/granite-cnn-classification A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 9 0 5 10 15 20 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 training epoch e rr o r error traintop1e valtop1e (a) 0 5 10 15 20 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 training epoch e rr o r error traintop1e valtop1e (b) 0 5 10 15 20 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 training epoch e rr o r error traintop1e valtop1e (c) Fig. 10. Validation error results after 18 training epochs of the proposed (a) MNIST1 (b) MNIST2 and (c) MNIST3 CNNs for granite images classification. The errors in these three figures were measured for 28 × 28 granite blocks classification. From this figure the smallest validation error ( valtop 1 e ) can be found in the 10th, 15th and 15th epoch respectively. Table 1 Results comparing the best configurations of the proposed method to the existing meth- ods in the literature after 5 × 2 validation. Proposed CNNs methods and CNNs fusion approaches are highlighted in light gray. Method Mean ± Std. dev. Min Max CIFAR 100% ± 0.00 10 0.0 0% 10 0.0 0% CT GF _ 3 X 3 ( Ferreira et al., 2015 ) 100% ± 0.00 10 0.0 0% 10 0.0 0% C ri 1 ( González et al., 2014 ) 99.96% ± 0.12 99.60% 10 0.0 0% CT GF _ 5 X 5 ( Ferreira et al., 2015 ) 99.92% ± 0.10 99.80% 100% CT GF _ 7 X 7 ( Ferreira et al., 2015 ) 99.70% ± 0.14 99.60% 10 0.0 0% DLBP ( Bianconi et al., 2015b ) 99.88% ± 0.16 99.60% 10 0.0 0% MNIST 2 99.32% ± 0.70 97.80% 10 0.0 0% LBP ( Ojala et al., 1996 ) 98.96% ± 0.64 97.80% 99.80% MNIST 3 98.16% ± 0.56 97.00% 99.00% { M NIST 1 , M NIST 2 , M NIST 3 } _ v ote 96.50% ± 0.99 94.80% 97.60% CIFAR original 94.84% ± 0.52 94.00% 96.00% GLCM − MD − MS ( Ferreira et al., 2015 ) 92.04% ± 1.09 90.00% 93.80% GLCM − MD ( Ferreira et al., 2015 ) 90.18% ± 1.43 88.40% 93.20% MNIST 1 86.62% ± 1.61 84.20% 88.80% HOG ( Dalal & Triggs, 2005 ) 78.04% ± 1.50 76.20% 80.80% GLCM ( Haralick et al., 1973 ) 72.74% ± 1.97 68.80% 74.60% MNIST original 62.98% ± 2.12 60.80% 66.40% { M NIST 1 , M NIST 2 , M NIST 3 } _ concat 26.44% ± 25.89 0.00% 67.00% o s a t t c e f M s p l One interesting point to be noticed is the good performance of ther texture descriptors that were proposed for other applications, uch as the ones from Ferreira et al. (2015) . The best state of the rt approach, CT GF _ 3 X 3 , also showed a perfect detection in all of en rounds of experiments. This happens because CTGF builds his- ogram of low pass textures that are in flat areas of the images, onsidering only the texture of these flat areas instead of using t dges information and some abnormal imperfections that can dif- erentiate the same class of granite slabs. Table 1 also shows that the proposed methods MNIST1 and NIST3 are not showing good results if compared with most tate of the art. This severely impacts the proposed fusions ap- roaches by early fusion ( { M NIST 1 , M NIST 2 , M NIST 3 } _ concat ) and ate fusion ( { M NIST 1 , M NIST 2 , M NIST 3 } _ v ote ), because feature vec- ors and outputs from MNIST 1 and MNIST 3 networks do not de- 10 A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 Fig. 11. Results considering the classification of 32 × 32 granite image blocks. l u i t f p p a v t p i b t t o e i i f d w 7 s c a a e t t n d t l p e o r t g l t t e p p a a t s a p t r d c a f m scribe the granite images effectively as MNIST 2 does. One good step towards the solution could be training new and more deeper net- works to be used in the fusion. Finally, it is worth discussing the results of pre-trained mod- els in classifying granite slabs. As seen in Table 1 , the pre-trained models CIFAR original and MNIST original show poor classification re- sults. This happens because these pre-trained models were tough to identify different patterns from the ones present on granite tiles, so their parameters (or filter weights) were not found to discrim- inate granite slabs patterns. This affects extremely their experi- ments results. 6.3. Comparison against baselines in low resolution images For the final round of experiments considering the comparison with the state of the art, we now focus on the experiments regard- ing small 32 × 32 blocks classification. For this, we consider the classification result without performing the majority voting, using for this the first split of training data and test data. This will result in a total of 10 0 0 (images) × 2116 (blocks) = 2,116,0 0 0 blocks of size 32 × 32, half of them are used to train and the other half used to test the classifier. We show, in Fig. 11 , the accuracy result of our best proposed approach against the best state of the art methods considered in the experiments of Section 6.2 . Results showed in Fig. 11 highlights the difficulty of classify- ing very tiny images. This can be explained because they contain a very small color information that can be confused with other gran- ite classes, decreasing this way the classification accuracy. Even in this difficult scenario, our proposed CNN CIFAR , trained to classify these small blocks with backpropagation of error and more layers, showed a classification accuracy higher than 85% using the same nearest neighbor classifier used in the experiments of Section 6.2 , classifying a total of 923,231 blocks out of 1,058,0 0 0 blocks. The high classification accuracy of low resolution blocks explains why the majority voting of individual blocks helped our proposed ap- proach to reach the 100% accuracy on the experiments results shown in Table 1 . The poor results of the feature engineering approaches of Ferreira et al. (2015) and González et al. (2014) are due the in- ner characteristic of these feature engineering approaches to be global descriptors, not being thought to classify low resolution im- ages. For example, the CT GF _ 3 X 3 approach of Ferreira et al. (2015) , which classifies correctly only a total of 351,760 blocks, creates histogram of textures on convoluted areas with a 3 × 3 win- dow. As the areas used to describe the granite tiles contain only 32 × 32 = 1024 pixels, the final histogram used as feature vector will contain several bins unused, and the accurate description of this small granite tile will be affected. The low accuracy results of iterature solutions in Fig. 11 seriously decrease the possibility of sing these approaches to do image classification by majority vot- ng of image patches, as our proposed approach does. Other impor- ant aspect of the proposed approach is the time for extracting the eature vectors. While our trained from scratch network took ap- roximately one hour to extract 1,058,0 0 0 feature vectors, the ap- roaches from Ferreira et al. (2015) and González et al. (2014) took pproximately 6 days each to extract the same number of feature ectors. This highlights the efficiency of the proposed method in his multiple data scenario. With the results presented in this section, we show that the roposed approach can achieve perfect detection of high resolution mages, showing comparable results with the state of the art and, y considering low resolution granite rocks slabs, it outperforms he state of the art comfortably. This indicates, firstly, the poten- ial of our proposed method to be a complementary approach to thers in the literature for industrial applications with controlled nvironment. Moreover, it can be a good starting point for help- ng classifying rocks on uncontrolled environments, such as us- ng smartphones with different resolutions as acquisition devices or granite recognition, acting as a possible expert to solve misun- erstoods between customers and manufacturers when a possible rong delivery happens. . Conclusion The creation of quality-control procedures for sorting natural tones such as granites is a promising task, as they avoid economi- al losses due to delivery of wrong granite packages to clients. The utomation of such process is either important to reduce errors nd eliminate the use of a subjective process involving expensive xperts. However, most of the applications proposed in the litera- ure thus far for this task rely on feature engineering approaches hat investigate texture and color behaviors which are supposed to ot change. Additionally, they do not validate the approaches in a ifficult task of classifying natural stones of different image resolu- ions. In this paper, we address these issues by proposing a deep earning based approach to granite rocks classification. Our ap- roaches are based on Convolutional Neural Networks of differ- nt architectures applied on small image tiles, analyzing textures f each one and using majority voting of them to classify high- esolution images. We also investigate the performance of some of he networks when applied together. Experimental results showed ood results of the proposed approaches for classifying high and ow resolution images if compared to some other texture descrip- ors proposed in the literature. Although there is a long path to go toward the classifica- ion of natural stones in real-world situations, in which differ- nt resolutions and lightning conditions of granite rocks can hap- en, we believe that the direction considers deep learning ap- roaches through Convolutional Neural Networks. As our proposed pproaches deal with tiny patches, they have potential to be invari- nt to most of acquisition resolutions, once they are bigger than he input tiny patches that fit our networks. This is an important tep to make the granite slabs recognition tasks available to other cquisition devices, such as smartphones. Also, the proposed ap- roaches can also be complimentary to other approaches in indus- ry applications of slabs recognition, as they showed comparable esults to the state of the art in classifying high resolution images. The work started in this paper opens a set of future work to be one, and involves (i) the proposal of new and deeper network ar- hitectures for this task; (ii) use of Convolutional Neural Networks pplied on other image representations; (iii) testing new ways of using different network architectures; (iv) studying the comple- entarity of these data-driven approaches with featuring engineer- A. Ferreira, G. Giraldi / Expert Systems With Applications 84 (2017) 1–11 11 i o w i A a a t p M s R A B B B B D D E E F F G H H H J K K L L O R R S S U V Y ng techniques and (v) application of experiments considering the pen set scenario of Scheirer, Rocha, Sapkota, and Boult (2013) , on hich the classifier is designed to also consider unknown samples n the training process. cknowledgments This work was supported partly by NSFC ( 61332012 , U1636202 ) nd the Shenzhen R&D Program (JCYJ2016032814 4 421330). We lso thank the support by the Brazilian National Council for Scien- ific and Technological Development (Grant #312602/2016-2 ) and rofessors Nuria Fernández, Bruno Montandon Noronha Barros and illena Basílio da Silva for the discussions that originated this re- earch. eferences raújo, M. , Martínez, J. , Ordóñez, C. , & Vilán, J. A. (2010). Identification of granite varieties from colour spectrum data. Sensors, 10 (9), 8572–8584 . ianconi, F. , Bello, R. , Fernández, A. , & González, E. (2015a). On comparing colour spaces from a performance perspective: Application to automated classification of polished natural stones. In New trends in image analysis and processing . In Lecture notes in computer science: Vol. 9281 (pp. 71–78). Genoa, Italy: Springer . ianconi, F. , & Fernández, A. (2006). Granite texture classification with Gabor fil- ters. In Proceedings of international congress on graphical engineering (INGEGRAF), Sitges, Spain . ianconi, F. , González, E. , & Fernández, A. (2015b). Dominant local binary patterns for texture classification: Labelled or unlabelled? Pattern Recognition Letters, 65 , 8–14 . ianconi, F. , González, E. , Fernández, A. , & Saetta, S. A. (2012). Automatic classifi- cation of granite tiles through colour and texture features. Expert Systems with Applications, 39 (12), 11212–11218 . alal, N. , & Triggs, B. (2005). Histograms of oriented gradients for human detection. In Proceedings of International conference on computer vision & pattern recogni- tion, California, USA: Vol. 2 (pp. 886–893) . ietterich, T. G. (1998). Approximate statistical tests for comparing supervised clas- sification learning algorithms. Neural Computation, 10 , 1895–1923 . rshad, S. F. (2011). Color texture classification approach based on combination of primitive pattern units and statistical features. Multimedia and its Applications, 3 (3), 1–13 . uropean Committee for Standardization (CEN) (2009). EN 12440:2008 natural stone: Denomination criteria. Report . European Committee for Standardization (CEN) . ernández, A. , Ghita, O. , González, E. , Bianconi, F. , & Whelan, P. F. (2011). Evaluation of robustness against rotation of LBP, CCR and ILBP features in granite texture classification. Machine Vision and Applications, 22 (6), 913–926 . erreira, A. , Navarro, L. C. , Pinheiro, G. , dos Santos, J. A. , & Rocha, A. (2015). Laser printer attribution: Exploring new features and beyond. Forensic Science Inter- national, 247 , 105–125 . onzález, E. , Fernández, A. , & Bianconi, F. (2014). General framework for rotation in- variant texture classification through co-occurrence of patterns. Journal of Math- ematical Imaging and Vision, 50 (3), 300–313 . aralick, R. M. , Shanmugam, K. , & Dinstein, I. (1973). Textural features for im- age classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3 (6), 610–621 . e, K. , Zhang, X. , Ren, S. , & Sun, J. (2015). Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of IEEE in- ternational conference on computer vision (ICCV), Santiago, Chile (pp. 1026–1034) . of, R. (2013). Deep learning. With massive amounts of computational power, ma- chines can now recognize objects and translate speech in real time. Artificial in- telligence is finally getting smart. https://www.technologyreview.com/s/513696/ deep- learning/ . ohnson, J., & Karpathy, A. (2016). CS231n convolutional neural networks for visual recognition. http://cs231n.github.io/transfer- learning/ . rizhevsky, A. , Sutskever, I. , & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. In Proceedings of neural information pro- cessing systems (NIPS), Nevada, USA (pp. 1106–1114) . urmyshev, E. V. , Sánchez-Yáñez, R. E. , & Fernández, A. (2013). Colour texture clas- sification for quality control of polished granite tiles. In Proceedings of the third IASTED international conference on visualization, imaging and image processing: Vol. 2 (pp. 603–608). ACTA Press . ecun, Y. , Bottou, L. , Bengio, Y. , & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86 (11), 2278–2324 . episto, L. , Kunttu, I. , & Visa, A. (2005). Rock image classification using color features in Gabor space. Journal of Electronic Imaging, 14 (4), 040503-1–040503-3 . jala, T. , Pietikäinen, M. , & Harwood, D. (1996). A comparative study of texture mea- sures with classification based on feature distributions. Pattern Recognition, 29 , 51–59 . umelhart, D. E. , Hinton, G. E. , & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323 , 533–536 . ussakovsky, O. , Deng, J. , Su, H. , Krause, J. , Satheesh, S. , Ma, S. , . . . Fei-Fei, L. (2015). ImageNet large scale visual recognition challenge. International Journal of Com- puter Vision, 115 (3), 211–252 . cheirer, W. J. , Rocha, A. , Sapkota, A. , & Boult, T. E. (2013). Towards open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36 , 1757–1772 . tone Industry SA Standards Board (2015). Stone standards. Report . Stone Industry SA Standards Board . niversity of Tennessee (2008). Granite dimensional stone quarrying and process- ing. Report . University of Tennessee – Center for Clean Products . LFEAT (2016). MatConvNet: CNNs for MATLAB. http://www.vlfeat.org/matconvnet/ . osinski, J. , Clune, J. , Bengio, Y. , & Lipson, H. (2014). How transferable are features in deep neural networks? In Advances in neural information processing systems (pp. 3320–3328). Montreal, Canada: Curran Associates, Inc . http://dx.doi.org/10.13039/501100001809 http://dx.doi.org/10.13039/501100003593 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0007 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0007 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0008 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0008 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0013 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0013 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0013 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0013 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0013 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0014 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0014 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0014 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0014 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0014 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0014 https://www.technologyreview.com/s/513696/deep-learning/ http://cs231n.github.io/transfer-learning/ http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0023 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0023 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0024 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0024 http://www.vlfeat.org/matconvnet/ http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30303-2/sbref0025 Convolutional Neural Network approaches to granite tiles classification 1 Introduction 2 Related work 3 Basic concepts 4 Proposed method 4.1 Conversion to grayscale 4.2 Patching 4.3 Recognition by Convolutional Neural Networks 4.4 Image patches classification 4.5 Image classification 5 Experimental setup 5.1 Image dataset 5.2 Methodology and metrics 5.3 Implementation aspects of the proposed approaches 5.4 Baselines 6 Experiments 6.1 Defining epochs to train CNNs 6.2 Comparison against baselines in high resolution images 6.3 Comparison against baselines in low resolution images 7 Conclusion Acknowledgments References 
work_mlu43b75bnbodkglnzba6yrhuq	----	GagneAndersenNorford2011_ExpertSystem_BAE_PrePrint Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 1 Title: An Interactive Expert System for Daylighting Design Exploration Authors: Jaime M. L. Gagne1 (Corresponding Author) Email: jaimelee@mit.edu MIT Room 5-418 77 Massachusetts Ave. Cambridge, MA 02139, USA Phone: 1-617-835-0950 Fax: 1-617-253-6152 Marilyne Andersen1, 2 Email: marilyne.andersen@epfl.ch Leslie K. Norford1 Email: lnorford@mit.edu 1Building Technology Program, Massachusetts Institute of Technology, USA 2Present Address: Interdisciplinary Laboratory of Performance-Integrated Design (LIPID), School of Architecture, Civil and Environmental Engineering (ENAC), Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland Abstract: (150-250 words) Architects increasingly use digital tools during the design process, particularly as they approach such complex problems as designing for successful daylighting performance. However, while simulation tools may provide the designer with valuable information, they do not necessarily guide the user towards design changes which will improve performance. This paper proposes an interactive, goal-based expert system for daylighting design, intended for use during the early design phase. The expert system consists of two major components: a daylighting knowledge-base which contains information regarding the effects of a variety of design conditions on resultant daylighting performance, and a fuzzy rule-based decision-making logic which is used to determine those design changes most likely to improve performance for a given design. The system gives the user the ability to input an initial model and a set of daylighting performance goals in the form of illuminance and daylighting-specific glare metrics. The system acts as a “virtual daylighting consultant,” guiding the user towards improved performance while maintaining the integrity of the original design and of the design process itself. Two sets of case studies are presented: first, a comparison of the expert system results to high performing benchmark designs generated with a genetic algorithm; and second, an evaluation of the expert system performance based on varying levels of aesthetic constraints. The results of these case studies indicate that the expert system is successful at finding designs with improved performance for a variety of initial geometries and daylighting performance goals. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 2 Keywords: Expert system, daylighting, design process 1 Introduction Designers have long considered daylight as an important aid for architectural expression. In recent decades, we have come to understand that daylighting may provide additional benefits, such as reduced energy consumption and improved occupant health and well- being [1,2,3]. Nevertheless, simply providing daylight in a building will not always result in positive results. Daylighting is only as good as its delivery system, so careful design is necessary to ensure that enough light is available and that glare, shadows, and reflections are reduced [4]. Unfortunately, it is often a challenge to create a successfully daylit building. Digital tools offer new ways of helping architects create or find designs with high levels of daylighting performance using efficient and intelligent guided design exploration methods. Optimization algorithms are a common solution, largely because they have the capabilities necessary to find or generate successful solutions; however, these methods generally allow only limited amounts of user interaction during the actual optimization or decision making process. Optimization algorithms also produce solutions based on performance criteria, not based on any understanding of design. As it is highly unlikely for a designer to simply accept a design generated by an optimization algorithm, an alternative approach would be a more interactive search method, which would accept input from a designer and grant the designer a larger degree of control. An example of such an approach is a knowledge-based or expert system. An expert system is one in which human expert knowledge about a specific domain is encoded in an algorithm or computer system [5]. In the daylighting domain, such a system would function as a virtual lighting consultant, guiding the designer towards design modifications which improve overall daylighting performance. Knowledge-based systems have already been successfully implemented for artificial lighting scenarios [6,7]. For daylighting, a few simple expert systems exist. The Leso-DIAL tool uses an expert system based on fuzzy logic rules to provide users with a “qualitative diagnosis” (for example, it might diagnose the light levels in a space as “Very Low” and suggest modification of certain design characteristics) [8]. The NewFacades approach considers energy and visual comfort based on a prescription energy code for hot climates to suggest a range of façade solutions to the designer [9]. These systems represent first steps in expert systems for daylighting in design, but they do not allow for a comprehensive understanding of daylighting or a large amount of user interactivity regarding design choices. This paper describes a user-interactive expert system approach which includes two climate-based performance metrics, one for illuminance and one for daylighting-specific glare, in order for the designer to have an understanding of the amount of light and the visual comfort in the space. The method begins with a designer's own initial design and performance goals. It then evaluates the performance of the design and creates a series of Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 3 suggestions for design changes which are likely to result in improved performance, thus enabling a search process that is highly specific to the user's design problem. The expert system is comprised of a pre-calculated database of daylighting-specific information connected to a set of fuzzy daylighting expert rules. Any design decision that the designer chooses to allow will be automatically generated in the original model and the new performance will be calculated. The designer is allowed to interact with the system through an iterative search process that is both agreeable to the designer and likely to improve the performance of the design. The effectiveness of the proposed expert system as a design-making algorithm has been assessed through a series of case studies which compare the performance of designs found by the expert system to a set of reference designs generated by a genetic algorithm, a known optimization method. The behavior of the expert system was also evaluated based on façade design constraints, including the initial façade designed by the user and the selected window uniformity scheme (which describes whether the expert system must constrain all windows on the façade to have the same dimensions or whether windows may differ from one another). The results of these case studies are presented in this paper and indicate that the expert system is successful at finding good solutions for a variety of performance and design conditions. 2 Expert System Development The expert system presented in this paper is a fuzzy rule-based system combined with an external database of previously computed daylighting simulation data. This external database serves as a “knowledge base” of information about how various changes to façade design elements, such as window size and external shading devices, affect the illuminance and glare in the interior of a space. The fuzzy rule-based system uses information from this database in addition to information about the geometry and daylighting performance of a given design to create a list of suggested façade design changes that should improve overall daylighting performance. This section describes the major components and logic used by the expert system. 2.1 Overall System Structure The expert system was developed as an extension of the Lightsolve program, an intuitive rendering and simulation tool aimed to help designers consider daylighting performance in the early design stages [10]. A schematic of the overall expert system is shown in Figure 1. The process begins when the user inputs information about his or her specific design problem into Google SketchUp. This data includes a 3d model, location information, and specific daylighting performance goals for illuminance and glare. The next steps of the process are to populate a simple building data model based on the 3d model which will be used by the expert system as well as to determine the performance of the design using a daylighting simulation program. The user’s specific performance goals are taken into account using goal-based metrics which are calculated using the simulation data. The expert system component of the system consists of a series of fuzzy rules which use information about the current goal-based performance as well as the Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 4 geometry and materials used within the design to create a list of façade-specific design changes to suggest to the user. The user is allowed to view this list, along with the current performance of the design, in an interactive interface that was developed specifically for the expert system. The user may choose a design suggestion to try from within the interface, and the system will automatically make the selected change to the original 3d model. The process then repeats until the designer is satisfied with the design. Each of the major components of the expert system are described in further detail in the following sub-sections: the 3d modeler (section 2.1.1), the user inputs (section 2.1.2), the simulation engine and daylighting metrics used by the system (section 2.1.3), the building data model (section 2.1.4), the user interface (section 2.1.5), the knowledge-base of daylighting-specific information used by the expert system (section 2.1.6), and the major assumptions and logic used within the fuzzy logic rule bases for decision making (section 2.2). Figure 1: Overall system diagram of the expert system. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 5 2.1.1 3d modeler The 3d modeler currently used by the system is Google SketchUp [11]. This program is an intuitive and robust modeling tool with an embedded Ruby application programming interface (API), which was used to develop the majority of the expert system functionalities. The expert system process can be initiated from within SketchUp after the user creates a 3d model of his or her design. 2.1.2 User Inputs The expert system requires a number of initial user inputs that describe the design problem. The major user input is a 3d model of an original design in SketchUp. Sensors for illuminance and/or glare are modeled as 2d planes. The user may elect to have any number of illuminance and/or glare sensor planes and they may be any size. Sensor planes may be oriented vertically or horizontally, and they may be opaque or transparent. Materials of opaque and glass surfaces must be specified within SketchUp. An example model with horizontal illuminance sensor planes and vertical glare sensor planes is shown in Figure 2. The Ruby API embedded within SketchUp was used to create pop-up interfaces that allow the user to enter such additional inputs as performance goals for each sensor plane. For each illuminance sensor plane, the user must specify a desired illuminance goal range in lux, including the actual desired range and a buffer zone of acceptable values. For example, the user may desire the illuminance of a given sensor plane to fall between 400 lux and 1200 lux, but he or she will also accept illuminances as low as 200 lux and as high as 1500 lux. For each glare sensor plane or glare sensor group, the user must choose a glare tolerance. The glare tolerance options are “zero” (i.e. no glare is tolerated), “medium”, and “high” (i.e. a high amount of glare is allowed). Figure 2. Example 3d model that meets expert system modeling criteria. Interior illuminance and glare sensors are shown as horizontal and vertical planes, respectively. Additional inputs allow the user to customize the behavior of the expert system. One set of inputs is the set of priority levels for each performance goal. The priority level is a number from 1 to n, where n is the total number of sensors. The user may choose to prioritize one or more goals over others, or he or she may elect for all goals to have the same priority. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 6 The user may constrain the expert system aesthetically by selecting a window uniformity scheme. Three choices are allowed: “All windows in the model should look the same,” “All windows on a façade should look the same,” or “Windows can look different from other windows on the same façade.” A location and weather file must be specified to provide climate data to the simulation engine. Weather data in an EnergyPlus weather data format (.epw) can be used by the system. Finally, the user must indicate his or her times and seasons of interest (the choices are: winter, fall/spring, summer, morning, mid-day, and afternoon). The expert system will only consider performance during those times of interest selected by the user. 2.1.3 Simulation Engine and Daylighting Metrics The engine used to calculate daylighting performance is the Lightsolve Viewer (LSV) [12], the rendering and simulation engine native to the Lightsolve program. LSV is a hybrid global rendering method that combines forward ray tracing with radiosity and shadow volumes rendering. It is a stand-alone executable which is called directly from within the SketchUp/Ruby environment and simulates the performance of 3d models created in SketchUp. For all illuminance and glare sensor planes within the 3d model, the LSV engine calculates annual goal-based performance metrics using the 3d model, the location and weather information, the performance goals (illuminance ranges and goal thresholds), and the times of interest. To calculate the goal-based illuminance, the LSV engine first triangulates each sensor plane into small patches, and then calculates climate-based illuminance [12] on each patch over 56 time periods that represent a whole year. For a single patch, the goal-based illuminance metric is defined as the percentage of the user’s times and seasons of interest during which daylight provides an illuminance within the user’s specified range. The final goal-based illuminance for a sensor plane is an average of the performance over all patches on a sensor plane. Partial credit is given for illuminance levels that fall between the “acceptable” and “desired” values (Figure 3). A value of 100% indicates that the entire area of the sensor plane sees an illuminance in the user’s desired range over all periods of day and seasons of interest. Similarly, goal-based glare is calculated on each glare sensor over 56 time periods that represent a whole year. The glare metric used by the expert system is Daylight Glare Probability (DGP), which indicates the percent of occupants disturbed by a daylighting glare situation [13]. DGP was chosen because it is a daylighting-specific glare metric that considers windows as glare sources. DGP has also been found to yield the most plausible results for glare due to daylighting when compared to other glare indices [14]. The LSV engine calculates a model-based approximation of the DGP known as DGPm, which performs within a 10% error of the DGP over 90% of the time for rectangular models that do not include window frames [15]. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 7 To evaluate glare risks, the expert system uses the DGP threshold values described by Wienold [16], where any value below 0.33 (imperceptible glare) is considered a “no glare” situation and given a glare credit of 100%. The threshold values for these three glare tolerance levels (“zero”, “medium”, or “high”) correspond to the three glare ratings of “perceptible”, “disturbing”, and “intolerable” glare [16]. Any calculated glare value above the upper threshold is given a glare credit of 0% (Figure 3). These glare credits are averaged across all glare sensors in each glare sensor group within the model. A value of 100% indicates that the specified view direction is unlikely to see glare due to daylighting. Figure 3: Functions for goal-based performance metrics for illuminance and glare 2.1.4 Simple building data model In addition to performance, it is necessary for the expert system to understand the geometry and materials of the design. To accommodate this, a simple building data model was developed whose values are automatically assigned once the process is initiated. The model contains information about each building element (floor, wall, ceiling, window, shading device) and the relationships between them. Each building element object contains information about its location, geometry, orientation, and material. The general structure of the data model is indicated in Figure 4. The building data model was implemented using the SketchUp Ruby API and is created using 3d models in SketchUp. The logic used to populate the building data model is described further in [17]. The model allows the expert system to understand which walls have windows, how large those windows are and where they are oriented relative to each Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 8 other, as well the shading devices and glazing associated with each window. It also allows the system to understand the locations of each illuminance or glare sensor relative to each façade and to each individual window on the façades. Figure 4: Structure of simple building data model automatically generated by the expert system based on a 3d model 2.1.5 User Interface A stand-alone interface was developed to allow the user to interact with the expert system. The interface indicates the current performance of the design based on the goal- based illuminance and glare metrics and displays the design changes suggested by the expert system to the user. The interface also allows the user to select design changes to implement based on the expert system’s suggestions. The user’s selected design changes are automatically applied to the user’s 3d model in Google SketchUp and simulated using LSV. Once the simulations are completed, the interface is updated with the performance of the new designs and a new list of design suggested is displayed. The interface will continue to update to display the performance of the designs over multiple iterations (Figure 5). This interface was implemented using Adobe Flash. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 9 Figure 5: Performance analysis and decision making interface for the expert system. Views of the current design are shown (top left) along with annual performance in temporal map form (top right). Performance over multiple iterations is shown in the interactive graph (lower left). Expert system design suggestions are given in the lower right. 2.1.6 Daylighting Knowledge-Base To aid in the decision-making process, the expert system uses a knowledge-base of pre- calculated, climate-specific data [18]. This database was populated with simulation data, using the Design of Experiments method [19]. It contains information about the relative effects of 10 different façade parameters on both illuminance and glare from the various zones and views within the space. The 10 different façade parameters considered are: window area, window height-to-width ratio, vertical and horizontal location of windows on the façade, window distribution (how close or far apart windows are to each other), total number of windows, length of horizontal overhangs and/or vertical fins, glass transmissivity, and glass type (regular or translucent). The expert system can potentially suggest 20 different façade design changes, which correspond to two directions of change for each of the 10 façade parameters considered in the knowledge base (for example, window area can be made larger or smaller). The expert system uses the information within the general daylighting knowledge-base to create a customized database which includes only the data corresponding to the seasons and times of interest given by the user. This subset is further customized for each individual sensor based on the zones in which each sensor is located. Additionally, only Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 10 the relevant views are included for glare sensors. For example, a user may create a 3d model in which glare sensors are placed facing south-west from within the core and south perimeter zones. The user might also specify that his or her goals are relevant only during the school schedule. In this situation, the expert system would only use information from the knowledge-base which is relevant to glare sensors facing south and west from core and south zones during autumn, winter, and spring, from early morning through early afternoon only. All additional information, such as data about glare sensors facing north, data about the summer months, or data about illuminance sensors, would be ignored by the system. 2.2 Fuzzy Logic System The expert system rule base is a decision-making algorithm that assesses specific design situations and creates lists of suggested design changes that should improve the current performance, based on user-defined daylighting goals. The rule base uses fuzzy logic [20], which allows it to better emulate the human thought process than classical logic. It has been developed to be a flexible algorithm that can accommodate a wide variety of initial design scenarios. 2.2.1 Assumptions and Logic This section provides an overview of the general assumptions and logic used within the expert system to determine which design changes to recommend to the user to improve performance. Selecting Which Windows to Target The expert system assumes that design changes made to the façade closest to a given sensor will affect that sensor more than design changes made to façades further from the sensor. Similarly, the expert system assumes that on a given façade, some windows will be closer to a sensor plane than other windows, and changes to those windows will have a greater effect on the sensor plane than other windows on the same façade. Dealing with Multiple Performance Goals If there are multiple sensors within a model, the expert system will attempt to find design changes that are likely to improve the performance of all sensors at once. However, in situations where the user’s goals are conflicting, the expert system will choose to improve one sensor at a time, perhaps at the expense of performance of another sensor. In these scenarios, the expert system uses the following logic: user-specified high priority goals take precedence over lower priority goals, and sensor planes that have the lowest current performance have priority over sensor planes with good current performance.   Dealing with Illuminance Goal Ranges Illuminance goals are based on user-specified lower and upper bounds. As a result, an illuminance sensor plane may see illuminance that is too low, within range, or too high. Dealing with a sensor plane which sees illuminance that is both too high and too low at the same time is a complicated problem. The expert system chooses to deal with this problem in two ways: it determines if other sensors would benefit more from moving Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 11 towards higher illuminance or lower, and it takes into account whether the amount of illuminance that is too high is greater than or smaller than the amount of illuminance that is too low.   Determining an Appropriate Magnitude of Change A problem similar to that of dealing with illuminance goal ranges is selecting an appropriate magnitude of change. For example, the system may conclude that a small increase to the illuminance on a sensor plane will bring performance closer to the user’s goal range; however, an increase which is too large will result in decreased performance due to the illuminance on the sensor plane’s being too high. The expert system deals with this issue by determining whether a change should be “small” or “large,” and by selecting design changes from the daylighting knowledge base which are deemed most appropriate for that magnitude. The system then creates design changes in three increments and allows the user to select the version he or she prefers based on the resultant performance and aesthetics. 2.2.2 Fuzzy Sets and Rules During the expert system process, the goal-based illuminance and glare performance of the design, along with the original user inputs, is used to assign values to sets of fuzzy variables, which help to describe the current scenario. These fuzzy sets are: userPriority (high and low), sensorPerformance (good and bad), illuminanceSensorPerformance (too high and too low), glareSensorPerformance (too high), and distanceFromPerformanceGoal (close and far). In addition to these fuzzy variables, the system also uses information from the model’s customized knowledge-base (section 2.1.6) to determine values of the fuzzy set actionResult (Figure 6) for each potential design change. These fuzzy variables refer to the likely result of a given design action on a given sensor, for example “Large Increase in Illuminance”. Each sensor in the model will have a unique actionResult fuzzy set, with different values for each possible design change. Figure 6: Membership functions for ActionResult fuzzy set. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 12 Once created, the fuzzy variables are used to fire a series of fuzzy rules. The result of this process is a ranked set of design actions that are most likely to improve the performance of the current design based on the user’s goals and preferences. Each individual fuzzy rule is an if-then statement which uses fuzzy variables for both the antecedents and consequents. The fuzzy rules have been divided into four sets of “rule bases” which are fired in the order indicated in Figure 7. Each rule base contains a series of fuzzy logic if-then statements which are fired in sequence. The purposes of each rule base, along with example if-then statements, are listed below: Rule Base 1: Determine priority of each sensor. For example, one rule within this rule base is: IF SensorPerformance is Bad AND UserPriority is High, THEN SensorPriority is High. Rule Base 2: Determine which change(s) will improve performance, based on the current scenario. For example, IF SensorPriority is High AND SensorType is Illuminance AND IlluminancePerformance is TooLow: (a) IF distanceFromGoal is Far, THEN DesiredChange is “Increase Illuminance by a Large Amount”; (b) IF distanceFromGoal is Close, THEN DesiredChange is “Increase Illuminance by a Small Amount”. Rule Base 3: Evaluate each possible design action in the customized database using the desired changes determined in Rule Base 2. For example, IF DesiredChange is “Increase Illuminance by a Large Amount” AND ActionResult is LargeIncrease, THEN action is GoodForSensor. These rules are fired once per potential action, and once per sensor. Rule Base 4: Each potential action is ranked based on how likely it is to improve each sensor and the sensor priorities. The final step is to sort the set of design actions from highest to lowest rank. The first design actions in the list will be those actions most likely to produce positive performance results in the current design, while those actions at the end of the list are likely to decrease overall performance. The expert system interface presents the design suggestions to the user one at a time in this order. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 13 Figure 7: Flow chart for fuzzy logic rules fired by the expert system with inputs and fuzzy variables indicated. Each rule base represents a series of fuzzy if-then statements which are fired sequentially. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 14 3 Evaluation of the Expert System The main function of the expert system presented in this paper is to effectively guide a user towards improved daylighting performance of an original design. It is of critical importance that users have confidence in the advice given by the system, so a high level of performance is essential. Although the expert system differs from a traditional optimization algorithm due to its domain-specific and user-interactive nature, it should be equally capable of finding successful solutions in a best-case scenario. In order to assess the behavior of the expert system, a set of studies was performed to compare the performance of designs found using the expert system to high performing benchmark designs generated using a genetic algorithm (GA). The GA was chosen to create the benchmark cases because it is an algorithm known to find optimal or near- optimal solutions for a variety of solution spaces; however, it should be noted that other optimization algorithms could have been used with the similar results. The GA used in these case studies was a micro-genetic algorithm (micro-GA) [22], which is a GA with a very small population size. For each case study, the objective function for the micro-GA was to maximize the goal-based performance of all sensor planes within the model. For comparison purposes, the micro-GA was implemented within the Lightsolve system and uses the same 3d models and performance metrics as the expert system. The same 10 façade variables were considered, encoded into a 30-bit string. These variables were the same variables considered by the expert system. For these case studies, no constraints were considered for either the micro-GA or the expert system. Further details about the micro-GA system can be found in [17]. Section 3.1 describes the results of a selection of case studies that demonstrate the behavior and performance of the expert system across a variety of scenarios. A more complete set of case studies can be found in [21]. Section 3.2 presents the results of two additional studies that were performed to determine how the behavior of the expert system is influenced by initial façade constraints, the façade design of the initial model and the window uniformity scheme selected by the user. All case studies were sited in Boston, MA. 3.1 Comparison Case Studies A set of study procedures was developed to better compare results from the expert system to the GA, given their differences in algorithm type. While a GA generates designs, the expert system always assumes that an initial design is given and suggests design changes based on the current design. The following procedures were used: Micro-GA procedure An initial massing model with no windows was used to generate a new model of each generated design. The algorithm was run for 10 generations before stopping. If a solution that met all goals was not found, the highest performance found over all generations was considered to be the best design. As the goal of the study was to generate a high- Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 15 performing benchmark design and not necessarily an optimal design, it was assumed that 10 generations were sufficient. Expert system procedure An initial model, designed to be of mediocre performance, was created with generic rectangular windows. For these case studies, a “perfect user” was assumed. The “perfect user” was defined as someone who would select the first suggested design change at each iteration and the best performing magnitude of each design change. This scenario is illustrated more clearly in Figure 8, which shows the first four design stages suggested by the expert system for case study #1 (section 3.1.2.1). The “perfect user” scenario was also one in which the process continued even if performance decreased after a given design iteration. The algorithm was run for 10 design iterations before stopping. As with the GA study, if a solution which met all goals was not found, the best design was considered to be that with the highest performance over all completed iterations. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 16 Figure 8 The performance and designs of the first four design steps of an example problem. The performance goal considered is a wide illuminance range (described further as case study #1). The “perfect user” selections are 1c, 2a, 3c, and 4a. It was not possible to select a “best” performing design from either the GA or the expert system for case studies involving multiple conflicting goals. In these cases, an approximate Pareto front was created by the multi-objective GA to demonstrate the range of possible designs and their performances for each of the conflicting goals. The designs Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 17 produced by the expert system were compared with those along the approximated Pareto front. For all case studies presented in section 3.1, the “uniform window” scheme was selected. The behavior of the expert system for the “non-uniform window” scheme will be discussed in section 3.2. 3.1.2 Comparison Case Studies Results This section presents five case studies that demonstrate the range of problems that the expert system can handle successfully. The authors also initially completed simpler studies in which single goals were considered with either minimum or maximum illuminance values or a maximum glare threshold. Although they are not presented here, in all simple case studies, the expert system was able to find a solution within 10 design iterations that met the performance goal over 100% of the sensor plane area and over all times of year. Figure 9 Comparison of best performing final designs from the expert system and micro- GA for case studies #1 through #4 Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 18 3.1.2.1 Case Study #1: Wide Range Illuminance Goal The first two case studies consider a simple box model with a single illuminance sensor plane located in the core zone and two external façades with windows oriented towards the east and south. The first case study considers a single performance goal with an illuminance range: 300 lux minimum preferred (100 lux accepted) and 1500 lux maximum preferred (2500 lux accepted). This is a relatively simple problem to solve, given that the range of acceptable illuminance values is fairly wide. For this case study, only the school schedule was considered (morning through mid-day, autumn through spring). In this case study, the micro-GA was able to find a solution that was essentially “perfect” (99.9% in range) after 10 generations. The expert system was also able to find a near- perfect solution (99.8% in range) after 10 design iterations. The best performing designs for both algorithms are shown in Figure 9. Both final designs feature smaller windows on the south façade and larger windows on the east façade, and both designs have small or no shading devices on either façade. 3.1.2.2 Case Study #2: Narrow Range Illuminance Goal In this case study, the same initial model as the previous case study was used with a narrower illuminance range as a performance goal: 300 lux minimum preferred (100 lux accepted) and 800 lux maximum preferred (1200 lux accepted). Because the illuminance range is stricter than the previous case study, the problem is more difficult to solve. Like the previous case study, a school schedule was considered for this problem. In this case study, the micro-GA was able to find a design with excellent performance (97.3% in range, for the times and seasons considered) after 10 generations. The expert system was also able to find a design with very good performance (94.4% in range). The final designs produced by the two algorithms (Figure 9) both have large shading devices on the south facing windows; however, the east façades are visually very different. This difference may be the cause of the 3% difference in performance between the final designs found by the two systems. 3.1.2.3 Case Study #3: Two Illuminance Goals – Sensors Parallel to Façades This case study considers an L-shaped space with two illuminance goals. The two façades of interest are oriented towards the south and west, and the two illuminance sensors are located parallel to these façades (Figure 7). The illuminance goals for each sensor are: • South zone: 400 lux minimum preferred (200 lux accepted); No maximum. • West zone: No minimum; 500 lux maximum preferred (800 lux accepted). Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 19 Figure 10 L-shaped initial massing model for case study #3 with two illuminance sensors indicated Based on these goals, the known design solutions to this problem featured small, shaded windows on the west façade and larger windows on the south façade. For this case study, the goals were considered non-conflicting and the total performance of each design was calculated as the average performance of both sensors. The expert system was able to find a design solution with an average of 96.1% in range. The micro- GA found a similarly good solution with an average performance of 95.3%. As expected, both “best” designs have either very small or highly shaded windows on the west façade with larger or less shaded windows on the south façade (Figure 9). 3.1.2.4 Case Study #4: Two Illuminance Goals – Sensors Perpendicular to Façades The second non-conflicting goals case study considers a trapezoidal space with a sloped roof. The two façades of interest are oriented towards the south and north, and the two illuminance sensors are located perpendicular to these façades in the east and west ends of the space (Figure 10). In this case study, the height of the north façade is twice the height of the south façade. The illuminance goals for each sensor are: • East zone: 200 lux minimum preferred (100 lux accepted); 800 lux maximum preferred (1200 lux accepted) • West zone: 400 lux minimum preferred (200 lux accepted); No maximum. Figure 11 Trapezoidal initial massing model for case study #4 with two illuminance sensors indicated Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 20 As with the previous case study, total performance is considered as the average performance of the two illuminance goals. The micro-GA was able to find a solution with an average performance of 87.0% while the expert system’s best design had an average performance of 82.6% (Figure 9). It is clear that both systems struggled with this particular case study, which indicates that the performance goals may have been somewhat conflicting. This case study represents the largest difference (4.4%) between performance found by the expert system and that found by the micro-GA. 3.1.2.5 Case Study #5: Conflicting Illuminance and Glare Goals This case study is a conflicting goals scenario which features one illuminance goal with a desired range of high illuminance values and one glare goal. The goals in this case are conflicting because achieving the illuminance goal is likely to cause glare to increase for the views considered. This case study considers a Z-shaped floorplan, and the two façades of interest face east and west (Figure 12). Two illuminance sensors planes are located in the east and west zones and glare arrays are located within the same zones with views facing outwards. Figure 12 Initial Z-shaped massing model (a) and façade design (b) for case study #5 with illuminance and glare sensors shown. The performance goals for the two sensor groups are: • Illuminance: 200 lux minimum preferred (0 lux accepted); No maximum. • Glare: Zero glare tolerance (only imperceptible glare allowed). This case study had an additional constraint that all façades must be uniform and that all façades must be identical to ensure that the performance goals would be conflicting. Because there cannot be a single “best” solution to a problem with conflicting goals for this case study, an approximated Pareto front was generated using a multi-objective micro-GA [16]. The approximated Pareto front demonstrates the range of possible solutions that are considered non-dominated. By examining the set of all non-dominated solutions, one can begin to understand the relationship between the two conflicting performance goals. To compare the results of the expert system to those generated by the multi-objective micro-GA, the expert system was run three different times, each for five design iterations, Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 21 with a different sensor priority given for each of the three runs. When all the designs generated by the expert system over the three runs are compared to those generated by the micro-GA over 50 generations (Figure 13), it is clear that the expert system designs cover a wide area within the solution space and offer a way of approximating the Pareto front using fewer total simulations than those required by the micro-GA. Figure 13 Case study #5 - Conflicting illuminance and glare goals: Performance for the expert system for three different goal priority scenarios over the entire solution space (upper) and over the approximated Pareto front (lower). Although the expert system does not generate a Pareto front itself, it is interesting to compare three designs generated by each algorithm: the design with the best average performance, the design with the best illuminance sensor performance, and the design with the best glare sensor performance (Figure 14). In this case study, the micro-GA was able to find a design with an average performance that is over 5% better than the design found by the expert system, and in general, it is clear that the expert system designs Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 22 tended to have slightly lower glare performance than the micro-GA designs in the middle area of the Pareto front. However, the expert system is still able to effectively provide the user with a rough approximation of the Pareto front and a set of designs that explores the trade-offs between illuminance and glare performance.       Figure 13 Case study #5 - Conflicting illuminance and glare goals: Comparison of designs for best average, glare, and illuminance sensor performance for the expert system and the micro-GA./   3.2 Façade Constraint Case Studies While the previous set of case studies was able to demonstrate that the expert system can indeed produce designs that are comparable to those produced by a micro-GA, the behavior and performance of the expert system is also dependent on several variables that do not affect the micro-GA. One important difference between the micro-GA and the expert system is that the micro-GA generates its own designs while the expert system Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 23 begins with an initial design and suggests changes to be made to that specific design based on its current performance. The best design produced by the expert system is thus highly dependent on the initial design given to the system. The expert system also allows the user to select a uniformity scheme for the windows in his or her design. In the micro-GA comparison studies, both the micro-GA and the expert system designs were constrained to have uniform façades, which meant that all windows on a single façade had the same characteristics. However, the expert system includes the option of having non-uniform façades. Selecting this option would allow the system to make changes to individual windows on a façade instead of all of them at once. For certain types of design scenarios, selection of the non-uniform window option will result in greater performance improvement than the uniform window option. Two brief case studies examine the effects of the initial façade design and the window uniformity scheme on the overall improvement found by the expert system. 3.2.1 Effect of Initial Façade Design To examine the relationship between the performance of the expert system and the initial façade design, the problems in the micro-GA case studies #1 and #2 (sections 3.1.2.1 and 3.1.2.2) are considered: a simple box model with an illuminance sensor in the core zone, external façades on the east and south, and an illuminance range as a performance goal. The two case studies differ as the performance goal for one is a wider illuminance goal range than the other, which is therefore an easier problem to solve. These case studies enable a comparison of the expert system behavior for different façade types on two different levels of goal difficulty. In the original case studies presented, the expert system process began with a relatively generic façade design with mediocre performance (around 50%). In this study, four different starting façades are shown in Figure 15 and have varying levels of specificity: the first is the generic façade with square windows, the second has slightly more elongated windows, the third has extremely elongated windows, and the fourth has elongated windows clustered towards one end of each façade. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 24 Figure 15 Starter and best designs for four levels of façade specificity and two levels of goal ranges (wide and narrow) In this study, the expert system process was run for each of the four initial façade types and for each of the two illuminance goal ranges. The expert system was run for 10 design iterations in all cases, and the “uniform façade” scheme was selected. The best performing final designs for all cases are shown in Figure 15. For both cases, it is clear that the starting design clearly influences the aesthetic of the final best performing design; however, in all cases, good solutions were found. The wide-range goal case study was more successful, with performances ranging from 94.5% to 99.8%. This result was expected as this goal is easier to meet. For the narrow range goal case, solutions with performance ranging from 88.4% to 94.4% were found. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 25 3.2.2 Effect of Selected Window Uniformity Scheme In addition to being heavily influenced by the initial façade design, the expert system’s search process is also dependent on the window uniformity scheme selected by the user at the beginning of the process. In this section, the models from case studies #3 and #4 from the micro-GA comparison studies were considered (sections 3.1.2.3 and 3.1.2.4), once with the uniformity of the façade maintained and once with non-uniform façades allowed. It was hypothesized that the uniformity of the façade would be more influential on designs with more than one goal than for single-goal scenarios. These two design problems each have two illuminance sensor planes with different performance goals, but they are considered non-conflicting goals because reasonable solutions exist which meet both goals at once. For the L-shaped room case study, the known good solutions featured small windows with shading devices on the west façade and larger windows on the south façade. For the trapezoidal case study, the known good solutions feature windows clustered towards the west end of both façades. The best performing designs found by the expert system for both uniform and non-uniform façades for both case studies are shown in Figure 16. Figure 16 Comparison of best performing final designs from the expert system for L-shaped and trapezoidal models (from case studies #6 and #7) with two different uniformity schemes Figure 16 indicates that the non-uniform window scheme produces significantly better results for the trapezoidal case study, where the sensor planes are located perpendicular to the façades of interest. The non-uniform window scheme allows the expert system to target the two illuminance sensors individually by making different changes to the windows that are located closest to each one instead of making the same design change to all windows. The final design for the non-uniform scheme still resembles the initial façade design, but the façades have each been divided into two, based on the locations of the two sensors. This final design has an average performance of 96.4%, which is 13.8% better than the performance of the uniform façade design. In the L-shaped case study, Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 26 where the sensor planes are located parallel to the façades of interest, the non-uniform façade scheme produced slightly better performance, but the final best designs for both schemes are within 1.1% of each other. This set of case studies demonstrates that the use of the non-uniform window scheme can produce dramatically better results for certain types of designs. 4. Discussion Based on the results presented in section 3, the expert system was found to be successful at making design decisions that improved the daylighting performance of five case study designs. In these case studies, the performances of designs using the expert system were compared to baseline examples generated by a micro-genetic algorithm (micro-GA). The purpose of the comparison studies was to evaluate the performance of the expert system relative to a known optimization algorithm which could be relied upon to consistently generate designs with very good, if not globally optimal, performance. The results of these case studies indicated that the expert system was successful at improving the performance of designs for a variety of initial conditions and performance goal scenarios. In some situations, the micro-GA was able to find designs which performed slightly better than those found using the expert system, but this difference in performance was small (4.4% at most for all case studies considered) and acceptable given the fact that the expert system was designed with user interactivity in mind, while the micro-GA was not. Two additional short studies were completed which investigated the effect of initial façade conditions and the effect of user-selected window uniformity constraints on the performance of the expert system. In the first study, it was found that although the initial façade design may affect the expert system performance, the system was still able to improve performance for even highly designed façades. For more stylized initial façades, it was found that the final improved designs were similar in appearance to the original designs, which demonstrates that the expert system preserves the original design intent. In the second study, it was found that the window uniformity scheme selected by a user can have a significant effect for certain types of model geometries. For models in which the sensor planes are located perpendicular to the façades of interest, the non-uniform window scheme selection was found to improve expert system performance (by over 10% in the case study considered). The expert system developed for this paper was a prototype tool which has several limitations. Due to the structure of the tool, the expert system requires the use of a pre- computed database (“knowledge-base”) for its decision-making logic. Although the system presented in this paper used a knowledge-base specifically for Boston, the addition of new climates or locations would be straightforward, requiring only the creation of new climate-specific knowledge bases based on the method described in [18]. However, the simulations required to create such knowledge-bases are time-consuming and would likely prove too complex for a casual user. A more robust version of the expert system would feature a selection of pre-computed knowledge-bases available for a variety of locations. Alternatively, more climates could be considered using a more Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 27 generic meta-database, which could potentially be designed to be applicable to multiple locations using weight factors based on climate and latitude. The current expert system was also limited in its use of geometrical forms. Only designs with orthogonal components can be considered by the expert system, and the number of possible façade design changes it can suggest was limited to those which could be implemented using existing functionality in the SketchUp Ruby API. These geometrical limitations constrained the number and type of designs that could be tested in section 3. Given the wide range of complex forms that designers can now create using digital tools, these limitations also restrict the potential for the expert system to be used in actual design scenarios. One partial solution is to expand the system to include a larger set of design changes and geometries, which would require an expanded knowledge base. Such a database could be populated using a method similar to the current knowledge base, with a larger set of variables. Because the system reads the knowledge base data from a text file kept separate from the coded logic, an expanded knowledge base could be added into the current system with only a small amount of editing to the code, mostly to accommodate the addition of new automated design changes. The addition of a significantly larger knowledge base could potentially increase the time required for the expert system to make decisions; however, such an increase would likely be negligible compared to the time necessary for simulations. The expansion of the system to include more varied geometrical forms would require integration of the expert system into a more flexible 3d modeling environment, for example a NURBS-based tool such as Rhinoceros. 5. Conclusions This paper presents a new user-interactive expert system approach which enables architects to consider daylighting goals in the early design stages by engaging them in a performance-driven design exploration process. One of the goals of this research was to introduce a new method for performance-driven design that allows a user to receive design suggestions that are specific to his or her original model and performance criteria. This goal required that the expert system be a highly flexible system that can produce positive results for a wide variety of possible inputs. The results of the case studies indicate that the expert system may find designs that perform similarly to those generated by an optimization algorithm; additionally, it may retain some of the qualities of the user’s original design, such as stylized façades. Based on these results, a potential user can have confidence that the design changes suggested by the system will improve the performance of his or her initial design if a number of design iterations are completed. If a true optimal design is not required, the expert system may be used in lieu of a traditional optimization method to find design with improved performance. However, as the expert system relies on an entirely different type of algorithm, it may also be used as a valuable complement to the optimization process, providing feedback in the form of design suggestions that may inform the purely performance-driven results of optimization. While the current version of the tool is limited by the geometries, locations, and performance metrics that it can consider, the expert system has a flexible structure which Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 28 could act as a framework for future expansion in areas that would enable the system to become more viable for use in actual design practice. Such an expansion could include additional locations, geometries, and design suggestions, as discussed in section 4. Additionally, while the current system considers only daylighting performance, it could be possible to expand the system to consider performance in other domains by considering solar thermal gains or building energy use. The solar thermal gains metric can already be calculated using the LSV engine, while building energy use could be calculated using an existing simulation engine from within SketchUp, such as EnergyPlus [23]. The addition of new metrics would require a substantial expansion of the current system, including the development of new knowledge bases and new fuzzy logic rules to work with the additional information. More research is necessary to assess the feasibility of such a scheme. In addition to the case studies presented in this paper, the expert system has also been tested by a group of designers who were asked to complete a design task with the system and to evaluate their experiences using the tool. The results of these user studies were positive and will be presented in a future paper. Acknowledgements The authors were supported by the Massachusetts Institute of Technology. The first author gratefully acknowledges financial support from the Martin Family Society of Fellows for Sustainability. The second author was also supported by the Ecole Polytechnique Fédérale de Lausanne. The authors wish to thank Sian Kleindienst for her work on the Lightsolve project, Phil Seaton for his help implementing the user interface, and Luisa Caldas for her valuable advice. References [1] Rashid M, Zimring C. A review of the empirical literature on the relationships between indoor environment and stress in health care and office settings: Problems and prospects of sharing evidence. Environment and Behavior 2008;40(2):151–190. [2] Edwards L, Torcellini P. A literature review of the effects of natural light on building occupants. National Renewable Energy Laboratory; 2002. [3] Boyce P, Hunter C, Howlett O. The benefits of daylight through windows. Capturing the Daylight Dividend Program, U.S. Department of Energy; 2003. [4] Boyce P, Heerwagen J, Jones C, Veitch J, Newsham G. Lighting quality and office work: A field simulation study. Pacific Northwest National Laboratory; 2003. [5] Luger G. Artificial Intelligence: Structures and Strategies for Complex Problem Solving. 5th ed. Boston: Addison-Wesley; 2004. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 29 [6] Jung T, Gross M, Do E. Light pen: sketching light in 3D. Proceedings of 10th International Conference on Computer Aided Architectural Design Futures, Taiwan; 2003. [7] Guo B, Belcher C, Roddis W. RetroLite: an artificial intelligence tool for lighting energy-efficiency upgrade. Energy and Buildings 1993;20(2):115–120. [8] Paule B, Scartezzini J. Leso-DIAL, a new computer based daylighting design tool. Right Light 1997;4(1):93–97. [9] Ochoa C, Capeluto I. Advice tool for early design stages of intelligent façades based on energy and visual comfort approach. Energy and Buildings 2009;41(5):480–488. [10] Andersen M, Kleindienst S, Yi L, Lee J, Bodart M, Cutler B. An intuitive daylighting performance and optimization approach. Building Research & Information, 2008;36(6):593–607. [11] Google SketchUp. 2011. Available at: http://sketchup.google.com. Accessed February 2011. [12] Cutler B, Sheng Y, Martin S, Glaser D, Andersen M. Interactive selection of optimal fenestration materials for schematic architectural daylighting design. Automation in Construction 2008;17(7):809–823. [13] Wienold J, Christoffersen J. Evaluation methods and development of a new glare prediction model for daylight environments with the use of CCD cameras. Energy and Buildings 2006;38(7):734-757. [14] Jakubiec J, Reinhart C. The ‘adaptive zone’ – A concept for assessing glare throughout daylit spaces. Submitted to Lighting Research & Technology. [15] Kleindienst S, Andersen M. The adaptation of daylight glare probability to dynamic metrics in a computational setting. Proceedings of LuxEuropa, Istanbul; 2009. [16] Wienold J. Dynamic daylight glare evaluation. Proceedings of Building Simulation, Glasgow; 2009. [17] Gagne J, Andersen M. A generative façade design method based on daylighting performance goals. Journal of Building Performance Simulation; 2011 (in press). [18] Gagne J, Andersen M. A daylighting knowledge-base for performance-driven façade design exploration. Submitted to LEUKOS. [19] Montgomery D. Design and Analysis of Experiments. 6th ed. John Wiley & Sons; 2004. Published as: J.M.L. Gagne, M. Andersen, L.K. Norford, An Interactive Expert System for Daylighting Design Exploration, Building and Environment, vol. 46 (11), pp. 2351-2364, 2011. 30 [20] Siler W, Buckley J. Fuzzy expert systems and fuzzy reasoning. John Wiley & Sons; 2005. [21] Gagne J. An interactive performance-based expert system for daylighting in architectural design. Ph.D. thesis, Massachusetts Institute of Technology, 2011. [22] Krishnakumar K. Micro-genetic algorithms for stationary and non-stationary function optimization/ SPIE Proceedings: Intelligent Control and Adaptive Systems; 1989. [23] OpenStudio. OpenStudio plug-in for google SketchUp. 2010. Available at: http://apps1.eere.energy.gov/buildings/energyplus/openstudio.cfm. Accessed February 2011. 
work_mm2jw6n2xzfxbi23fnvwvpvelq	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220994486 Params is empty 220994486 exception Params is empty 2021/04/06-03:05:39 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220994486 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:39 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_mm5rvxauovhfrhbonux35xb6me	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220990381 Params is empty 220990381 exception Params is empty 2021/04/06-03:05:34 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220990381 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:34 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_modzlbani5cnjeb3dhb4uw2ij4	----	Expert Systems of Multivariable Predictive Control of Oil and Gas Facilities Reliability Procedia Engineering 113 ( 2015 ) 312 – 315 1877-7058 © 2015 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Omsk State Technical University doi: 10.1016/j.proeng.2015.07.271 ScienceDirect Available online at www.sciencedirect.com International Conference on Oil and Gas Engineering, OGE-2015 Expert systems of multivariable predictive control of oil and gas facilities reliability Zemenkov Yu.D.a, Shalay V.V.b, Zemenkova M.Yu.a* a Tyumen State Oil and Gas University, 38, Volodarskogo St., Tyumen 625000, Russian Federation b Omsk State Technical University, 11, Mira Pr., Omsk 644050, Russian Federation Abstract The authors conducted a set of studies concerning the development of methodological support of multivariate predictive control reliability system for oil and gas industry. The algorithms, the innovative methods of calculation and the mathematical models of reliability factors, compatible with the modern production technological maintenance system, the system of dispatcher data registration, non-destructive testing diagnostics, and automated process control systems are developed. The mathematical software is designed to meet the technological features of the specific facilities, with applying the theory of process analysis, theory of reliability and fluctuation analysis elements. The developed models of reliability factors provide the possibility of predicting the parameters of technical facilities in a real time mode or for a fixed period, the structural and factor analysis function of the system in order to plan its optimal maintenance. © 2015 The Authors. Published by Elsevier Ltd. Peer-review under responsibility of the Omsk State Technical University. Keywords: reliability; safety; pipeline transport; reliability prediction 1. Introduction Monitoring technologies of MРC (multivariable predictive control) meet modern economical, production, and safety requirements based on the special continuous control of quantitative and running values of various factors: controllable, regulated, (set-points) and performance ones [1, 2]. * Corresponding author. Tel.: +7-345-220-1931. E-mail address: zemenkov@tsogu.ru © 2015 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Omsk State Technical University http://crossmark.crossref.org/dialog/?doi=10.1016/j.proeng.2015.07.271&domain=pdf 313 Yu. D. Zemenkov et al. / Procedia Engineering 113 ( 2015 ) 312 – 315 For the prediction and assessment of the reliability of pipeline transportation and hydrocarbon storage facilities in real time, complex study of all the factors, phenomena and processes that determine various properties of the system reliability is necessary. The analysis of existing studies indicates that, despite the urgency of the problem and a significant amount of research, existing techniques allow only a one-time assessment of reliability and are not focused on the use of immediate assessment and prediction in real time with the use of modern computer technologies [3- 6]. 2. Study subject The subject of the study is the technology of reliability prognostic control, which requires the development of algorithmic and mathematical set. 3. Methods Based on the analysis of the operating experience of the industrial facilities, it seems advisable to divide the sequence of the basic processes of design, implementation, and monitoring systems operation into several stages (Fig. 1). Stages 1, 2, 5, 6, and 7 are based on the results of theoretical and industrial research, while stages 3 and 4 are undertaken by organizations, developing information systems, as well as involved directly in construction and operating of the hazardous facility. international, state, regional safety requirements defining monitoring goals (decision-making) development of mathematical and algorithm system complex development of a tool kit: software and hardware development of monitoring network at the production facility System operation: data obtaining and analysis processing of monitoring results (decision making) Periodical critical analysis and system modernization economical and production requirements of manufacturers 1 22 33 44 de ve lo pm en t de ve lo pm en t op er at io n op er at io n im pppl em en ta tio n im pl em en ta tio n modernizationmodernization mm rr 55 66 77 Fig. 1. Algorithm of developing the technological processes monitoring system It is necessary to note, that possibilities and prospects of the implementation of the developed MРC system are not limited by only technological sphere of production: they are also focused on the scientific research and project institutes, regulatory authorities while evaluating the risk, choosing the complex of technological equipment for diagnostic and planned preventive maintenance, declaring organizations safety and so on. Almost every system of technological or industrial monitoring and control contains information databank. The system of source data collection is constructed in such a way that can be generated at existing organizations using both current and new hardware. For the analysis and control the system reliability, are formed structural and factor analysis algorithms, which, on a real-time basis, allow: 314 Yu. D. Zemenkov et al. / Procedia Engineering 113 ( 2015 ) 312 – 315 evaluating reliability indices of every system element at any level, and identifying the most vulnerable elements; determining system settings, which have the greatest impact on the facility reliability. It should be noted that the basis for developing a methodology for monitoring the reliability of pipeline transportation facilities systems and sub-systems is a complex of requirements, including the following: multivariable models of the system reliability should provide unambiguous, correct operational quantitative assessment of a particular hazardous facility; a system model and a set of universal logical-mathematical models should allow to make comprehensive evaluation and analysis of reliability, taking into account the factors and parameters that change in long periods of operation and storage, and constantly changing parameters in real-time modes; correlations between factors and indicators (often dependent or not uniquely determined) affecting the safety and reliability of the technical system as a whole (in particular, e.g., retention) should be clearly defined; methods for evaluating the reliability of the technological system should be compatible with existing systems technological processes ACS and allow the parameters to be adjusted under the influence of various factors; MPC system should provide functions for predicting (using the theory of probability and mathematical statistics) process parameters, planning, optimization and safe and effective control of the technical system as a whole. That is why the developed monitoring is based on the system analysis method being practically the only one providing decision-making under conditions of a large amount of information of different nature. MPC implements the principle of continuous scanning of reliability factors (in the system, it is a controlled variable CV) and comparing the values obtained with the critical one (Fig. 2). Upon reaching the critical values of the controlled factors, the area and the type of technical intervention is determined. N 3 N 2 Ncrit1 1 N per System lifetime ,Тsys Nsys N 4 5.Emergency condition 4. General maintenance 3 . Mid-life repair The system is reliable0 . 1. Maintenance 2. Routine repairs Т m Тrr Тmr Тgm Тec crit crit crit Fig. 2. Schematic diagram of a technical system reliability factor analysis With the purpose of implementing the estimation and analysis algorithms, we developed a set of mathematical models and calculation methods. Let the example be a methodology, based on the coefficient model of reliability factor. There are various models for estimating technical facilities reliability: experimental models; “parameter – dimensional limits”; “load – strength”; “load – bearing capacity”; model coefficients. There is no doubt, that coefficient methods are the most efficient universal methods for the implementation under conditions of long equipment operating period, under uncertainty and lack of information, because they represent a simplified set of complex particular models. The connection of reliability factors and operating conditions is provided by coefficients, considering the influence on the reliability factors: production technology characteristics; climatic factors; load factors: mechanical, electrical, etc.; facilities complexibility; maintenance characteristics etc. Thus, the problem formulated above can be solved by using coefficient models in combination with expert assessment of the coefficients significance. 315 Yu. D. Zemenkov et al. / Procedia Engineering 113 ( 2015 ) 312 – 315 The mathematical model of the monitoring system according to operational factors is based on the following assumptions and correlations: system reliability Nsys is characterized by a complex of reliability factors Si, domain of functions 0;1,0;1sy s iSN ; factors Si, are characterized by changing system operating parameters, i.е.: ;1321 ), n...S,S, Sf(SN nsys (1) perper iii ;yy), yf(yS maxmin or f(t)Si (2) General reliability factor is an integral value, found by using the method of linear convolution for every technological system facility considering the hierarchy: n i mimim kSN 11 (3) where Nm is function of the system reliability of complexity degree m; Sim-1 is reliability factor at hierarchy level (m-1) (by structural and factorial scheme); i is factor number ni ;1 ; ki is significance weight factor for the i -th factor. Qualitative and quantitative parameters composition Si (e.g., retention) is determined individually for every system at various hierarchical levels, based on the functional characteristics of the technological scheme of the facility. Values of ki are calculated by mathematical multivariable models or expert estimates. 4. Results and discussion The practical importance of the research includes the development of the set of mathematical models and prediction methods for the decision support and reliability factors monitoring system, operating on a real-time basis and providing the transition from the “after-failure” system of maintenance and repair to the preventative one, based on predictive reliability data. References [1] V.Kurushina, Y. Zemenkov, Innovative cyclical development of the Russian pipeline system. WIT Transactions on Ecology and the Environment 190, 2 (2014) 881 [2] Y.N Antip'ev, A.P. Nevolin, Yu.D Zemenkov, Operation of intermediate pumping stations on transport of gas saturated oils. NEFT. KHOZ. 10, (1981), pp. 46 [3] V.N Blinov, V.V., Kositsin, V.I. Ruban, V.V. Shalay, The research of ammonia electrothermal microengines for small spacecrafts Dynamics of Systems, Mechanisms and Machines, Dynamics 2014 - Proceedings 7005640. [4] V. Trushlyakov, V. Shalay, J. Shatrov, M. Jakovlev, A. Kostantino, Active de-orbiting onboard system from LEO of upper stages of launchers, European Space Agency, (Special Publication) ESA SP 672 SP (2009). [5] A.S. Bacherikov, M.I. Pashkov, Y.V. Bogatenkov, Y.D. Zemenkov, V.V. Shalay, G.G. Vasiliev, A.D. Prokhorov, S.N. Sidelnikov, S.Y. Podorognikov, V.F. Kehl, Y.D. Zemenkov, Ed. Diagnosis in the maintenance of pipeline transportation. St.Peterburg: Nedra, (2007) 380 p. [6] V.V. Shalay, V.I. Trushlyakov, V.Y. Kudentsov, Simulation of heat and mass transfer processes during gasification of liquid remaining fuel in the tanks missiles, Thermal processes in the art. T. 6. no 2 (2014) 67-74. 
work_mp2xxbjyqbgjddrwfxas6epqkq	----	The Role of Semantics in Legal Expert Systems and Legal Reasoning* Ratio Juris. Vol. 4 No. 2 July 1991 (219-44) copyright 0 Ronald Stamper 1991 NOTES DISCUSSION BOOK REVIEWS The Role of Semantics in Legal Expert Systems and Legal Reasoning* RONALD K. STAMPER 1. Semantics is Central to Legal Reasoning The consensus among legal philosophers is probably that rule-based legal expert systems leave much to be desired as aids in legal decision-making. Why? What can we do about it? A bureaucrat administering some set of complex rules will ascertain the facts and apply the rules to them in order to discover their consequences for the case in hand. This process of deductive reasoning is characteristically bureaucratic.' If the client or subject of the decision, after he has checked the deductions, does not like the apparent consequences of the rules, he will question their interpretation. This is not a deductive process. He will examine the meanings of the words in the rules and those used to characterise his case, looking for adjustments that lead to a more favourable decision. * This paper is based upon an invited contribution in May 1989 to the Conference on Legal Expert Systems organised by Prof. Enrico Pattaro as a part of the Ninth Centennial Celebrations of the University of Bologna. The contributions to this research programme by past and present members of the team are gratefully acknowledged, especially those by Peter Mason, Susan Jones, Clare Tagg, Sandra Cook, Martin Kolkmann and Liu Kechen. Weber's characterisation of bureaucracy is found in his Essays in Sociology where, in the EngIish translation (1946), he says: "The fully developed bureaucratic mechanism compares with other organisations exactly as does the machine with the non-mechanical modes of production." This allows the administration to be performed "precisely, unambiguously, continuously and with as much speed as possible." He stresses the "objective [. . .] discharge of business according to calculable rules" which are "of paramount importance for modem bureaucracy. " 220 Ronald K . S t a m p e r If he turns to an expert for advice, he will call in a lawyer. Of course it may be worth rechecking the deductions but that is a relatively trivial part of his skilL2 The lawyer applies his real expertise when he looks for variant reading of the rules, from those readings already made by the bureaucracy, to readings that might be made by a judge or tribunal. The dispute will only be resolved when the inclusion or exclusion of different rules has been settled and when the precise meanings of the included rules are agreed. Of course, the lawyer will also help his client by managing the procedural options in his favour. The procedures are also governed by rules and we have but another decision problem that divides into issues of bureaucratic choice and issues of interpretation. In his shoes, I should not be pleased with a lawyer who can reason about the deductive part of these problems but not the interpretive issues. So-called “legal expert systems” that fail to handle the problems of interpretation do not deserve the epithet “ e ~ p e r t . ” ~ At best they can be called “bureaucratic expert systems,” which is not to deny their potential value, only to recognise honestly their limitations. We shall examine the extent to which expert systems can handle meanings, the root of all problems of interpretation. We need to uncover the semantics of the system, those principles, tacit or explicitly stated, that link the elements of a knowledge-base or the text of a body of rules to the features of the world they signify. 2 . Misleading Metaphors In the paper, “Expert Systems: Lawyers Beware!”, Stamper (1988) draws attention to several metaphors regularly and un uestioningly employed in the discussion of 1 . The conduit metaphor of language treats words, expressions and sentences as carriers of meanings; detached from people, words go from place to place, or are ”stored” in books and computers, “carrying” with them this abstract “content“ we call “meaning.“ Whenever we talk about language this metaphor tends to be used (see Reddy 1979; Stamper 1985). 2 . The chemical engineering metaphor of data-processing reveals itself in such commonplace descriptive phrases as ”the extraction of information from data” and “the distillation of meaning from information.” They lull one into thinking that data, information and meaning, like chemical materials, exist independently of their users.‘ 3 . The set metuphor of reality, basic to the treatment of meaning in classical logic, regards the world as composed of individuals which can be assembled into sets, for example, the set of all red individual^.^ This metaphor is appropriate in the mathematical world of timeless abstractions but is not justified in the world of practical affairs where individuality and class membership can be open to dispute. expert-systems and the related fields o 4 problem analysis. In their book How To Do Things With Rules, Twining and Myers (1976), whilst acknowledging the importance of deductive reasoning in legal argumentation, nevertheless devote only six pages to that topic in their 270 pages volume. See Dreyfus and Dreyfus (1986) for a thorough criticism of the misuse of the word ”expert” in this context. In the expert system field this metaphor is used. Michie and Johnston (1984) write of ”a novel type of industrial plant, the ’knowledge refinery’, which would take in specialist knowledge in its existing form . . . and turn out knowledge that is precise, tested and certified correct.” For examples of mathematical logic spilling over into non-mathematical domains see Whitehead and Russell (1910127). Carnap (1942) is wholehearted in confirming this extension and Tarski (1965) in formulating it more precisely. One of the important points later in this paper is that to use mathematical logic in the domain of social and legal affairs requires justification of a non-mathematical kind. Semantics in Legal Expert Systems 221 4 . The correspondence metaphor of meaning6 is also needed if one defines the meaning of “red” as that set of a l l red individuals - an idea that leads to making another metaphysical assumption about the existence of ”possible worlds” without which some expressions (”your next contract”) would have nothing to correspond with. This illustrates 5. The platonic metaphor which most of us accept readily enough as a result of the indoctrination we receive in our mathematicslessons. We do not balk at assuming the existence of worlds populated by abstract objects and mappings between them.’ Finally, bringing together all these metaphors, we have number 6 . 6. The information-processing metaphor of mind, which equates minds and information processing devices, this helps us to accept the notion of knowledge as a saleable commodity packaged in an expert system.’ So commonplace are these metaphors that they seem l i e common sense. Even when someone’s attention is drawn to them and their shortcomings have been acknowledged, they can be difficult to abandon. Always the question is asked: What can we put in their place? We cannot answer by offering to eliminate metaphors but by supplying others, as we shall explain below. In order to deal more successfully with social systems, a new kind of common sense, new metaphors, a new paradigm, can be introduced. 3. Propositional Logic as a Basis for Legal Expert Systems To illustrate the semantic problems that arise in constructing legal expert systems, we shall examine rule-based systems built upon classical logics as their theoretical foundation. Equal attention should, perhaps, be given to systems based on some form of semantic net but their diversity and relative informality makes such systems difficult to target critically. In every case, the basis of the analysis should be to ask what is the underlying theory of meaning employed to relate the character-strings in the computer to the reality in which the systems serve their users. Logic-based systems are familiar enough to serve our purposes. The simplest logical tool is Propositional Logic. The smallest meaningful character- strings are elementary propositions, P,Q,R . . . etc. Their meanings are not analysed any further than to determine whether they are true or false (t or f ) . So we can characterise the underlying semantics by a function, u, which provides a truth value for each proposition. a(l‘)=t, u(Q)=f, a(R)=t, . . . . . The logic allows one to construct compound propositions which have meanings, in the rudimentary sense of the function u, computable from the truth-functional meanings of their constituents. Thus Z = (PorQ)&R is a compound proposition for which .(Z) = t . This is fully worked out in “Montague Semantics” to which Dowty et a l . (1981) provide an excellent introduction. ’ For a discussion of the value of platonism as an aid to the mathematical imagination, see Davis and Hersh (1983). Psychology needs metaphors in its search for a scientific understanding of the mind. Computer science provides the cognitive psychologist with a useful modern mechanical analogy, but the computer scientist cannot then use the presumption that minds are like machines to justlfy his quest for artificial intelligence. 222 Ronald K. Stamper Using Propositional Logic knowledge of the law can be stored in the form of rules such as (PorQ)&R- S so that, when the complex condition (PorQ)&R is met, the conclusion, S, can be asserted to be true also. This simple idea is just a re-writing of the familiar decision-table. This kind of legal expert system is described by Susskind (1987). Here is a slightly simplified fragment from his analysis of the Scottish law of divorce. legal production number 1 IF AND ONLY IF the marriage has broken down irretrievably A.s.l(l) of The Divorce (Scotland) Act 1976 AND NOT decree is to be withheld in respect of action under s.l(Z)(e) A.s.1(5) of The Divorce (Scotland) Act 1976 THEN the court may decide that [permission] d e a e e of divorce is to be granted legal production number 2 IF the defender has committed adultery time limitation: since the date of the marriage A.s.l(Z)(a) of The Divorce (Scotland) Act 1976 AND NOT the adultery has been connived at in such a way as to raise the defence of lenocinium A.s.1(3) of The Divorce (Scotland) Act 1976 AND NOT the adultery had been condoned A.s.1(3) of The Divorce (Scotland) Act 1976 THEN it shall be established that [obligation] the marriage has broken down irretrievably A&-B- [PIC A B C (D&D')%(-E&-F)- [o]G D D' E F G The basic structure here comes from Propositional Logic with some important additional features, in particular the deontic operators (permission and obligation) and the references to the relevant authorities, which we shall examine later. What does such a system "know" about the meanings of the words used in the law? "Not much" must be the answer. Whoever assigns the truth-values to the elementary propositions makes the interpretations from which the consequences are mechanically determined. This kind of system has virtually no semantics, nor does it pretend to have any. Its strength lies in the analysis of the constituent propositions and their organisation into rule structures that draw upon a mixture of statute law, case law and legal principles, supported by a clear statement of the authority for each element. In addition to the logical structure, such a system could provide an indexing scheme to a diversity of relevant material for use by the person who is deciding on the truth- values of the propositions. Thus the semantic knowledge that is the most important part of the lawyers' expertise has to be left out of an expert system based on propositional logic because Semantics in Legal Expert Systems 223 such a logic has no structural machinery to embody semantic information except the truth values of whole propositions. Propositional logic does not recognise smaller units than whole propositions, certainly not individual words. 4. Predicate Logic for Expert Systems Predicate Logic goes much further in the structures it recognises so it can exploit a more sophisticated semantics. Expert systems with this foundation appear to embody a substantial amount of knowledge of meanings. We shall examine the extent of this knowledge. Predicate logic deals with the inner structure of each proposition. Each proposition has the simple grammatical structure < subject > is c predicate > the subject can be the name of a single individual and the predicate the name of a property, or the subject can be a pair, a triple, . . . of names of individuals and the predicate names a relationship between them: John is a lawyer John and Mary are married. The crucial idea is that this logical system can distinguish not only propositions but also names of individuals and names of predicates. But what do these names mean? In order that predicate logic should not be an abstract mathematical tool, we have to provide it with a semantics that can relate its symbolic expressions to entities in the world of practical affairs. This requires two basic assumptions. First an ontological assumption that the world is composed of individuals each of which can be iden- tified; and also an epistemological assumption that we can know to which individual each name applies. The semantics of predicate logic uses the idea of a truth function, as does propositional logic. The meaning of a proposition is given by its mapping onto a truth-value: u(John is a lawyer) = true for example. But this truth value can be determined by referring to a more fundamental information. "John" is the name of an actual person (let us signlfy him by john) whilst "lawyer" is the name of the class of all individuals of that kind (let us signify this class of real individuals by LAWYER). The above proposition is true if and only if, referring to actual set membership, we find john E LAWYER We do something similar to give a precise meaning to "married" by making it mean the set of all ordered pairs of individuals of the appropriate kind, called MARRIED. s o < john,mary > E MARRIED has to be tested to find the truth of the proposition "John and Mary are married." Therefore, to know the meanings of the words in the formulae of predicate logic, we have to assume a knowledge of which individuals have the properties or fall'into the classes named by the predicates, and which pairs, triples, . . ., of individuals are related in ways named by the higher order predicates. Each allocation of individuals to the relevant classes is called a "model." 224 Ronald K. Stamper The legal rules will be regarded as an axiomatic theory which can have any number of interpretations supplied by different models, that is, sets of individuals and their groupings into sets of individuals, sets of ordered pairs, sets of ordered triples . . . etc. This elegant mathematical theory can draw upon all the sophisticated machinery of set theory. In its richest form, initiated by Richard Montague (1974; see Dowty et al. 1981), it can handle the semantics of a logical formalism that approximates to natural language. How then does it serve our purposes in building legal expert systems? 5. Theoretical Semantic Problems We have no need to look at the application of this kind of logic to law to see that it will raise some important semantic problems. Let us make a note of them before examining the additional problems arising in applications. Propositional logic, which is a component of predicate logic, employs the concept of truth-functional semantics. It assumes that, at the level of whole propositions, you need only know the truth-values of each of them in any situation to know their meanings and the meanings of the logical expressions derived from them. This reliance on truth as a simple, basic concept for a theory of meaning seems to work in domains which are free from dispute (routine engineering, or routine natural science) where we can perform a reliable operation of checking the truth of a fact-statement. In the case of a legal dispute, truth is what we arrive at, not what we start from. In such circumstances the operational basis of the semantics must be different so that we can treat truth as the point we reach following negotiation (among the members of a jury or between the parties to a dispute). Predicate logic uses the concept of an individual. Outside the mathematical domain, this concept is rather a sophisticated one which gives rise to the kinds of paradoxes that suggest we should be cautious about making it the basis of a semantics. What is and is not one and the same individual has been the subject of paradoxes since ancient times (the river of Heraclitus). The semantics of predicate logic does not elucidate these problems, is merely expects that the user has solved them to his own satisfaction. There are many important entities, such as water, gold and other substances that may be difficult to treat within the model of the world as a collection of individuals. Disputed identity may cause legal conflict, as in the case, for example, of the written-off car which appears on the road again - do we have one car that has been repaired or two cars, the second having been constructed from raw materials taken from the first when that one went out of existence. The meaning of each predicate is defined as a set of individuals, or pairs of individuals, or triples . . . etc. If the membership of the set changes, then the mean- ing of the predicate changes, unless one is prepared to abandon the set-theoretic definition of identical sets in terms of the one-to-one identity of their members (by extension) in favour of an appeal to defining properties (by intension). But doing this requires a different ontological position in which the set is independent of the membership. The problem has to be solved: What happens as individuals are born and die? We have to live with an every-changing meaning of “person”! To escape from this dilemma, we can take as our defining set, the set of all persons past and future and in any imaginable world. (This illustrates the comfort to be derived from a faith in the Platonic Reality.) In its turn, this entails our treating all instants of time as identifiable individuals. There is an escape. We can forget about a semantic theory that maps linguistic expressions onto sets of real objects. Instead we can put our trust in a purely syntactic theory. The logic contains rules of inference. These enable us, given one set of propositions (premises) to deduce, by mechanical operations on those premises, any Semantics in Legal Expert Systems 225 number of conclusions (theorems). To bother with what lies outside the system is regarded as irrelevant. This route simply dismisses semantics as not a real problem (Kowalski 1979,9, for example) - a justifiable point of view if you are only interested in the closed world of deductive processes and have no wish to devote time to justifying their relationship with the untidy world outside. To escape into the study of purely syntactic problems does not make the semantic problems disappear. On deciding to apply abstract logical formalisms to practical affairs, assumptions about their semantic justification should be made explicit. As we shall see, there is often too much reliance upon the users supplying the semantics intuitively. 6 . A Fragment of an Expert System To illustrate the semantics of expert systems based on predicate logic, a fragment of a rule-system developed by Sergot et al. (1986) will be examined. It deals with the British Nationality Act 1981, and it typifies the kind of system that interests us. The first sub-section of the British Nationality Act 1981, l.-(l) A person born in the United Kingdom after commencement shall be a British citizen if at the time of birth his father or mother is (a) a British citizen; or (b) settled in the United Kingdom. They express as follows, captured by the knowledge engineer in a number of Horn Clauses in Prolog x acquires British citizenship by section 1.1 on date y if x is born in the U.K. and x was born on date y and y is after or on commencement and x has a parent who qualifies under 1.1 on date y if z is a parent of x and z is a British citizen on date y if z is a parent of x and z is settled in the U.K. on date y if z is the mother of x if z is the father of x x has a parent who qualifies under 1.1 on date y x has a parent who qualifies under 1.1 on date y z is a parent of x z is a parent of x A system based on such rules can either give us their logical consequences if we feed in the facts e.g.: Matthew was born in the United Kingdom Matthew was born on 10 January 1987 Ronald is the father of Matthew 10 Jan 87 is after commencement Ronald was a British citizen on 10 Jan 1987 or, given a goal such as finding out whether Matthew has British citizenship or not, the system can generate the questions needed to elicit the facts before chaining forward to compute the answer. 226 Ronald K. Stamper Most of us would describe this kind of application of the law as a function of an ideal Weberian bureaucracy. Straightforward facts generate unequivocal consequences. The process does cot em loy much of a lawyer’s powers of reasoning although the bureaucracy may employ P awyers to perform this kind of deduction, especially to deal with very complex and seldom used chains of rules. Normal legal practice could not be performed only using the skills of deductive reasoning; neither could real bureaucrats do their work using only such banal reasoning processes. What is missing? 7. Semantic Problems Arising in Expert Systems Based on Predicate Logic The bureaucrat has to put the facts into words. Even when his job is limited to mechanical, deductive decision-making, he soon encounters some obvious semantic problems. For example, if I gave him the facts relevant to my son‘s claim for citizen- ship as I did above, then the system could not reason from those postulates. They would have to be revised as follows: not: Matthew was born in the United Kingdom but, to match “ x i s born in the U.K.”: Matthew is born in the U.K. not: Matthew was born on 10 January 1987 but, to match ”x was born on date y ” : Matthew was born on date 10 January 1987 not: 10 ]an 87 is after commencement but, to match ” y is after or on commencement”: 10 January 1987 is after or on commencement not: Ronald is Matthew’s father but, to match “z is the father of x”: Ronald is the father of Matthew not: Ronald was a British citizen on 10 Jan 1987 but, to match “z is a British citizen on date y”: Ronald is a British citizen on date 10 January 1987 Of course the system would not have been designed to accept the input in any form but rather to ask for the facts to be ”filled in” using the formats prescribed by the analyst. This solution means that this system has its own private way of collecting relevant data - just the problem that caused so much trouble in the early days of data- processing when dozens of stand-alone applications could not exchange data. In administration this is a serious fault. 8. Semantics and Large Systems The problems caused by inventing a private language for each application not only affects the exchange of data but it makes it impossible to integrate legal expert systems by the obvious, simple method of merging rules. This would not be a serious problemif different areas of the law could be placed in water-tight compartments, but as the law is a single fabric, this is a serious issue when considering strategy of developing integrated systems. Semantics in Legal Expert Systems 227 At the root of this problem is the fact that predicate logic is weak methodologically. Analysis of a legal text and its expression for manipulation in predicate logic does not elucidate meanings, but rather obscures them. Nothing in the structures d i s h ished b this logic will lead two different knowledge engineers to the same paraprase of t i e original text. Thus, the same original concept appearing in two different pieces of legislation analysed by two different engineers will only result in the identical predicate expression by a happy accident. We should not have to rely upon such chance agreements to preserve semantic integrity across a growing corpus of legal knowledge-bases. Returning to the example above, we can observe another semantic problem. The predicate names (in bold type) must have exactly the same form whoever uses them because they function as single symbols. A more honest representation of the above ten predicates created by the analyst would use single symbols, or at least strings that do not masquerade as natural language: for x acquires British citizenship by section 1.1 on date y for x is born in the U.K. for x was born on date y for y is after or on commencement for x has a parent who qualifies under 1.1 on date y for z is a parent of x for z is a British citizen on day y for z is settled in the U.K. on date y for z is the mother of x for z is the father of x The analyst who ran out of symbols could use longer strings, following the general practice in programming. This simple change emphasizes that we should not forget the fact that the analyst invents a new, artificial language when he creates an expert system using a language such as Prolog. 9. The Humpty Dumpty Syndrome These "fat predicates," the elongated symbols that look like natural language, are constructed by trial and error by the anal ~ t . ~ He invests them with meaning. The and other symbols precisely the meanings he wants them to have. The Humpty Dumpty syndrome is not only a disease transmitted by Prolog, it infects virtually all computer applications. It has some serious consequences. First, by actually allowing the analyst to invent an artificial language, over and above the already complex language of the law, we reduce the chance of anyone understanding correctly the contents of the "knowledge-base." Second, "fat predicates" serve to deceive the naive customer looking for an expert system. He may imagine - as an unscrupulous expert system vendor may intend - that the computer understands the meanings of the natural language words. It does not. We should not hesitate to criticise a mechanical engineer who supplies cardboard boxes painted grey leaving the customer to assume that they are steel girders, so let us demand similar standards from the knowledge engineer. Third, the analyst assumes that the original text of the statute contains a meaning which he has "captured" (note the metaphor) equally well in his formal version. Sergot et al. (1986) explain and appear to recommend this procedure in the section entitled analyst behaves like Humpty Dumpty i n A r ice Through the Looking Glass, giving words "Formalisation by Trial and Error." 228 Ronald K. Stamper Lawyers, however, take great care to preserve forms of words that have withstood the test of scrutiny by courts over many years. They will recognise the danger in this casual attitude towards language. The same problem will arise if the original text is analysed into its constituents by even the most sophisticated semantic analysis method, but, in this case, there is a reasonable chance that the analytical process will tend to improve the drafting of legal texts from a semantic point of view. We have noted the ease with which semantic confusions can be introduced to the knowledge-base in this kind of system. What can we do to remove the potential confusions introduced by a dozen Humpty Dumpties working on different parts of a huge corpus of formalised law? The only notion of meaning accessible to the computer depends upon the purely syntactic equivalence of one sign to another. Of course, the analyst working on a lengthy problem finds it difficult to remember the exact form of each elaborate predicate expression he has defined. Where we fear that several different formal predicates express the same legal concept, we might attempt to uncover these accidental semantic confusions by searching predicate strings for common terms. This cannot guarantee success. One can easily find paraphrases that contain no common significant terms! For example, the original x acquires British citizenship by section 1.1 on date y x acquires British citizenship on date y by sect. 1.1 British citizenship of x acquired on y under sub-section l.-(l) x is a citizen from y by virtue of British statute 1981c61,1.-(1) x is a natural born Briton commencing on y (ref UK 1981~61) may appear elsewhere in the following guises Unfortunately, as one can see from the last version, paraphrases need not include any common term. Clearly, this problem of matching different versions of what are supposed to be the same legal concept will give rise to errors and confusions when systems have to be extended and amended. It will prove an obstacle to the difficult process of unifying legal knowledge-bases created at different times by different analysts. And it will even obstruct the integration of the efforts of a team of analysts engaged on the same large legal domain. Quite another attitude towards paraphrasing might be justified - why not admit different interpretations? Dealing with law in the European Community it may be wise to allow the same piece of text to be treated in several different ways in a knowledge-base, to take account of national differences of behaviour. Interpretations can be localised to individuals or to groups. Unfortunately, this logic has no place to record the agent providing the interpretation of the meaning of the text. Meanings, it seems, are assumed to belong objectively to the natural language text, and the analyst only has to perceive and record them in another, but formal language. Paraphrasing problems are explained by the shallowness of the semantic analysis required for predicate logic. It is quite easy to shift between different forms when, with equal validity, one can accommodate a concept by treating it as a distinct individual, as the date is treated in x acquires British citizenship by section 1.1 on date y x acquires British citizenship at birth by section 1.1 whilst it can also be incorporated into a prolix predicate, as in This illustration demonstrates the looseness of the ontology behind predicate logic. What is deemed to exist or not exist is decided at the whim of the analyst! Semantics in Legal Expert Systems 229 One can only draw the conclusion that predicate logic, whilst being wonderfully rigorous from a syntactic viewpoint, is sloppy and informal semantically. 10. Predicate Logic Suits Bureaucratic Systems Bureaucratic systems are expected to operate as impartial machinery giving effect to policies worked out by the political system where value-judgments are paramount. Logical systems devised to serve mathematics may suit the ideal bureaucratic system. Legal systems are more than bureaucratic ones. They may give rise to systems of routine office work to take care of the commonest cases but the creation of the legal norms and their interpretation in the difficult cases involves the examination and re- examination of the values which the norms are supposed to embody. Differences of value-judgments become exposed as disputes about meanings. A logic for legal systems must give semantics pride of place. Mathematics, the paradigm of deductive reasoning, despite the claims made on its behalf (Susskind 1987, 185), does not necessarily provide the ideal model for reasoning in the domain of practical legal and business affairs. A logic inspired by mathematics will take advantage of simplifications that have no justification in the world of human relationships. Despite all the above noted objections to the rule-based expert systems built on the foundation of classical logic, they do have a place in the automation of routine administrative tasks defined by laws. I have no objection to such systems but I would hesitate to call them "legal expert systems" rather than "bureaucratic expert systems." Why should we expect a logic devised by mathematicians for their work to transfer comfortably to the domains of law and other practical social affairs that do not share the special properties of timelessness, abstraction, precise formality, independence from human judgment, desires and intentional action? It seems to me potentially a poor candidate for expressing legal issues but, faufe de mieux, people have embraced it too enthusiastically and quite uncritically. "What can we put in its place?" you will rightly ask. 11. Escaping from a Misleading Paradigm In aresearch project aimed at discovering ways of modelling organisations as systems of social norms,'" my colleagues and I have attempted to escape from the frame,of reference within which classical logic was created. Leaving a familiar framework can be very difficult, especially if there does not exist a ready-made alternative. Sometimes even more daunting is the passionate, irrational opposition with which a new framework will probably be greeted, revealing that the underlying metaphysics of a formal system owes as much, perhaps, to blind faith as any religious fundamentalism! This work on developing a legally orientated language, Legol, was conducted at the London School of Economics over many years but has now relocated at the University of Twente. Several versions of this language were formulated and implemented. Throughout, the law itself acted as the main source of ideas. Instead of aiming to apply logic to the law, we hoped to explicate the logic of the law. Whereas classical logic embodies the structures observed in mathematical systems, the law (as ancient a discipline as mathematics) might be expected to have evolved very different structures. Instead of dealing with timeless abstractions for which no person is A strategy which, we are pleased to note, Bob Kowalski's team has now embraced. 230 Ronald K. Stamper responsible, the law concerns a strict time-frame where specific people, as far as possible, are held responsible for the concrete events and actions to which its norms apply. This led us to incorporate into the formalism structures that require the analyst to take account of a number of features of obvious importance in every legal problem we had investigated. The result was an informal, partial answer to some of the questions about semantics raised in this paper. This partial solution" can be illustrated using the same Nationality Act example. Our empirical workmade it clear that the analysis would have to penetrate to the level of individual words or expressions that function as the semantic units in natural language. It was also clear that time had a special role; there was no point in treating it as either a class of individual instants or a class of intervals. Also, there had to be a place for the authority that determined the existence of every legally significant state of affairs. The result, from a semantic point of view, is indicated in the table below. Entity nation nation person citizenship citizenship citizenship territory territory located in act name act "1981~61" commencement fatherhood motherhood settled in Characteristic Antecedents name "Britain" name "1981~61" citizen, nationality "1981c61ssl(l)" "1981c61ss1(2)" name "United Kingdom" ? "1981~61~53" father child mother child "1981c61s50" ? ? Iden tifierslExistence StartFinish < time > < time .: birthdeath person, nation person, nation person, nation <time > <time > person, territory <time > < time > act person. 1, person. 2 person.1, person.2 person, territory Most of this table corresponds to a schema (in database terminology) which specifies the universals or entity types, that is, the classes of particular instances. In bold type are a few particulars. Every particular has associated with it a period of existence and the corresponding universal can be annotated to indicate how the start and finish times should be selected (birth and death for a person, for example). The chnracferistic can be name, in some cases, or it can be a criterion for the existence of a particular. For example, citizenship exists by virtue of a number of different criteria, giving rise potentiall to a number of different meanings of the concept. In the cases of "located in," "&herhood" and "motherhood" we do not know the criterion, and this is an indication that the analyst should clarlfy this point in order to complete his semantic schema.'* Identifiers are the names of the types of individuals being qualified or related by the entity. Later these were more accurately described as antecedents when they were generalised to be any other entity. The importance of antecedents is that their "See Jones et al. (1979), Tagg (1979) and Stamper (1980) for accounts of Legol-2.0 and Legol-2.1, the latter having been demonstrated in the 1979 Symposium on Computers and Law at Swansea. l2 Note how the syntax of this formalism guides the analyst. Semantics in Legal Expert Systems 231 co-existence was a necessary precondition for the existence of an instance of the entity, except in the case of what we call “things” which could have an independent existence. The antecedents also commonly have role names which can be inserted in the definitions, as above for citizen, nationality, father, mother, child. This is not the place to explain the Legol language for manipulating these structures and for expressing norms. An illustration will be enough. It would not have been difficult to provide an infix notation that would have made the first subsection of the Nationality Act 1981 read as follows. Citizenship @Britain of a person [criterion: 1981c61ssl(l)] starts when the birth of that person occurs - after commencement the British Nationality Act 1981~61 while his father is a British citizen 0’ settled in the UK or while his mother is a British citizen 0’ settled in the UK The words underlined are operators that function as logical operators that in every case take account of time. Some words here are just syntactic sugar and would be ignored by the interpreter. However, those words that appear in the table above can be manipulated as separate concepts (and so are in a sense “understood”) by the computer. This intermediate solution is full of defects. In particular, it does not function as a logic although it does allow the forward and backward chaining of rules that a legal expert system requires. In that respect it does not have the theoretical underpinning enjoyed by Prolog. Nevertheless, it has the advantage of taking care of some of the semantic problems catalogued earlier. In particular 1. fat predicates disappear and the semantic units of the formalism are those of the natural language; 2. semantic analysis no longer depends totally on the analyst’s skills in para- phrasing the original text; 3. indeed, semantic analysis is guided by the quite rich structure awaiting his results - each relevant word in the text generates a string of questions that must be answered about the start, finish, criterion, antecedents and role names; 4. accidental synonyms or multiple interpretations cease to be a danger because the arbitrary solutions that bedevil most systems specifications are greatly reduced, if not entirely eliminated when the level of analysis is reduced to the word. Despite these advances, this intermediate solution fails to break away from the old, defective framework. It still assumes the existence of individuals although it does not force one to treat instants or intervals of time among them. It still ado ts the stance and what is false. Nevertheless, it is loaded with signposts directing us towards a new framework. For example, - existence is clearly a candidate to replace truth as a primitive concept; - the idea of a criterion suggests that each instance of anything existing has to be determined by an authority, either in the form of an agent or a norm; - there appears to be a unlfying structure that eliminates the distinctions between individuals and predicates, or particulars and universals; - there seems no reason to treat norms as another kind of entity. The signposts begin to make sense once one abandons the old metaphysical that we can describe an objective reality using our language and so deci 1 e what is true assumptions. 232 Ronald K. Stamper 12. A New Framework The goal is to create a logical language incorporating a theory of meaning appropriate for domains of practical human affairs. The work is not yet complete. At present the result is a formalism, capable of mechanical interpretation, and, we believe, capable of development into a fully functioning logical system despite its many rather unusual features. It has become convenient to employ two formalisms, one, still called Legol, for manipulating the knowledge-base and another, called Norma, for knowledge representation. Norma is an attempt to capture in a logical syntax what we have discovered about the structures underlying social and legal norms. The theory has been developed through the analysis of very large numbers of legal problems and business information systems. This had led to the evolution of a method of analysis which has proved to be effective in designing systems that are cheaper to build and far cheaper to maintain than those developed using the conventional methodologies. This practical aspect of the work is emphasized in the Appendix to this paper where a concrete example is presented. This is also included in the hope that it will help the reader to follow some of the more abstract concepts introduced below. The metaphysical assumptions on which Norma is based are radically different from those that seem tacitly to be adopted by builders of rule-based expert systems today. I hope that you will agree that they are not difficult to accept, indeed that they are less mysterious than the metaphysical assumptions upon which orthodox logics are based. In our view there is no need to depend on the notions of individuality, identity, truth, and possible worlds, for example. These are all very sophisticated concepts of which we are innocent at birth. However, we start our lives with a desire to act, and a rudimentary system of values that defines the boundaries between states of behaviour that we like and dislike. During our most innocent months of life, we live in the here-and-now and we act directly on the world; only later do we begin to construct realities distant in space and time, and learn to act indirectly through communication. That very simple beginning is also our starting point for constructing our new framework. A belief in the existence of at least one agent (yourself) and hislher behaviour are the most basic ontological assumptions. We need symbols to represent these two kinds of things. Two epistemological principles govern the syntactic structure: all knowledge entails a knowing agent the agent only gains his knowledge through action. The agent, an or anism that has acquired (genetically, presumably, though it really does not matter i ow) a rich enough structure (an issue that can be examined later using the theory to be created) stands at the centre of reality. For the organism, reality unfolds through its action (see Gibson 1979; Michaels and Carello 1981). Each of us is responsible for the knowledge we have but the social system enables us to share our efforts and so live in a much richer reality. The concept of an agent can be extended to include groups, or social agents such as committees, teams, companies, nation- states even. Their collective behaviour is what we are interested in when we study norms.13 The idea of responsibility is basic. In the crudest sense of having to live with their consequences, the agent takes responsibility for its actions; by organising its behaviour system well, the agent has a better chance of surviving as an organism in the physical world or as a social agent in a social environment. In a more sophisticated way, responsibility arises in a human society as a result of the expectations produced l 3 The whole social agent is much more than its individual membership because of the addition of its norms. Every norm exists in relation to a particular social grouping. Semantics in Legal Expert Systems 233 by the agents’ intentional acts involving one another and the capacity of the social system to compel its members to endure the social consequences of their actions. The two fundamental principles result in a syntactic structure that mirrors the form of every pattern of behaviour in the shape of well-formed-formulas (wff) having two components: <agent > <behavioral invariant > Such a formula we call a “realisation” because it represents a realised or actual state of affairs, here and now. An agent that attains some behaviourally invariant state may be regarded as a modified form of agent which can enter as the agent expression into another realis- ation (wff). So we find the early concept of an antecedent in Legol incorporated into this language (for example, a state that has created an Act can then perform the behaviour it calls “commencement” in the example above). The logical operation of building a new pattern of behaviour on the foundation of its antecedents is represented by the concatenation of invariants: AXY where the agent, A, realising x, has become the agent (Ax) and so has been able to realise the dependent behaviour y. Because the existence of y can only occur during the existence of x, we call y the ontological dependent of x, its ontological antecedent. What are these behavioural invariants? The easiest way of recognising them is by means of our natural language vocabulary. Most of our words (certainly our nouns, verbs, prepositions, adjectives, adverbs) are names of invariants. It may seem puzzling to treat the names of objects this way but just consider what you know about any ordinary object such as a cup. All your knowledge of a cup comes from what you do with it, or what someone else tells you about what he does with a cup. To speak of a cup is to speak of a repertoire of behaviour that the cup makes avail- able - provided that you have the invariant cup, you have available to you all the things you can do with the cup.“ Vocabulary (rather neglected by linguistic theory (Aitchison 1987) ) is thus linked to our collective choice of where we draw boundaries to mark biologically or socially significant shifts in the value of our behaviour. Words are names for zones in our behavioural space, zones where some important features are invariant. To this, the most basic of the logic structures, ontological dependency, we add a number of logical constants to account for our ability to combine behaviour patterns at the same level. These are the analogues of the familiar operators and, or, and not of propositional logic with the important difference that they do not combine sentences but behaviours; one may do one thing while doing another, onvhile doing another, or whilenot doing another (notice that we cannot have negative actions, so whilenot is dyadic). The agent only exists in the present, but its present behaviour (overt and internal) can become extremely complex if it is to accommodate its beliefs about the l4 This repertoire of behaviour is associated with any number of amount of cups. Given an indeterminate quantity of cups, we can engage in the important and very basic behaviour of dividing them up and discarding part. This behaviour can continue until we reach a minimum amount of cups, and then the partitioning behaviour will cause our behaviour repertoire to change - our one remaining cup is broken! In such a manner we do arrive at the notion of an individual as a rather sophisticated pattern of behaviour. 234 Ronald K. Stamper past and the future. These operators play a large part in constructing the necessary descriptions. For example:I5 In which the composite expressions also refer to behavioural invariants having existences that are functionally dependent on their components. The correlate of material implication in propositional logic is the norm, for example: A ((x while y) whilenot (z onohile w)) A((x,y);(z:w)) A (y whenever x) A (x then y) the existence of which is not functionally dependent upon the existence of its components x and y. A norm specifies a mechanism, physical or social. This is also an invariant. Another way of representin a mechanism is to do so without sa ing anything about its structure by naming t%e ability to behave in a certain way. &r example: In a similar way, we can name two other related behaviours, the beginning and the ending of an invariant: A walk able Aw* A walk begin A walk end Aw < Aw> which are OUT names, not necessarily A's names for those behaviours. The agent A may not be aware that x and x < are related. In order for the agent to behave intentionally, he will have to recognise the connection. So far, although we have introduced the notion of existence into the discussion, nothin has been said about time. Time is a rather sophisticated construct that represent another. The model of behaviour introduced so far is sufficient to account for the appearance of semiological behaviour. (But that is another long story.) All we need to recognise is that the agent has the capacity to behave in a way that signifies something else: A"0x" represents A realising a sign, "Bx," meaning Bx. The precise nature of "Bx" is often of indifference to us in modelling behaviour, the use of voice, print, picture, electronic signals, and so would all be covered by the same expression provided that the agent A were using any of them to mean "Bx." One important result of this analysis is that we can easily go on to show the communication acts that the agent can perform with the sign A "Bx" asserts A "Bx" commands A "Bx" suggests for example. The interpretation or misinterpretation of a physical signal is not considered here but this more detailed level of analysis is readily accommodated. Once the agent has acquired the ability to use signs helshe can do more than live only in the here-and-now (see Mead 1932). The well-formed formula, Ax, represents A bein in the invariant state x. The syntax only allows one to represent the here-and- now. l%e start and finish of the invariant cannot be known directly, one is always in the past and the other in the future but the agent can here and now use signs to represent them. We use these symbols: depen % s upon the agent having, first, the ability to use one form of behaviour to A walk start Aw + A walk finish AW- l5 The abbreviations that we use for the logical operators (e.g., :; for while, onuhile, and whilenot) are introduced without other explanation than the presentation of the formula in the two versions. Semantics in Legal Expert Systems 235 Thus we are able, in our current behaviour, to accommodate worlds that are past, future and distant. The importance of signs in social and legal behaviour cannot be overestimated. We should distinguish, of course, between the internalised norms and the signs representing norms, "rules" are what we usually call them. The fundamental notion of responsibility has also to be given an explicit place in the formalism. Every realisation has an associated authority for its existence, symbolised thus Ax@ This may be an agent, often it is a subagent of a group agent. It may also be a norm expressing the combined judgments of several agents. Ultimately, the authority always reduces to a set of human judgments. In other words, this logic will always point to agents in the social system who are responsible for the use of the constructs expressed in the formalism. Every realisation has an associated authority. so every compound pattern of behaviour has its authority, so the logical combinations that in classical logic we are free to assemble at will, and even assemble mechanically, need to be authorised in Norma, strictly speaking. From a mathematical point of view this is a dreadful restriction but for modelling the social world, the restriction is realistic. In particular, although we have accepted such facts and norms as Ax A (y whenever x) Ax A(Y - X) we do not take it for granted in Norma that we are allowed to make combinations of norms and facts, such as A(x while (y whenever x)) A(X4Y -4) Such an act is not automatically permitted by the logic of the system, as in practice we assume that a competent judge will take responsibility for invoking the norm. It would be unwise to make the authority, as in classical logic, some universal rule permitting the unrestricted combination of expressions recorded in the formal system. Of course, we may wish to make use of this kind of mechanical linking of formulas but we should not make the mistake of assuming that the laws of logic are divinely authorised for all places and all times. They are human constructs with human authority behind them. All too often, the dogmatic logician (like the fundamentalist believer) insists upon pushing logical rules beyond their limits of practical validity. This language is very strange from a mathematical point of view. It can never be fully represented as a closed formal system. The agent responsible for the formal expressions is needed to complete the well-formed formula. (The author of the expert system is one factor in its expression.) In Norma, one has no alternative but to represent an open system. One is always referred, by the authority, to the containing social system. In Norma one cannot deal with a finite system because every behaviour it represents has other behaviours necessarily associated with it - the ability, the beginning, the ending, the sign for it, its authority - and these, recursively, have theirs. No matter how deep his analysis, the knowledge-engineer (dreadful, misleading expression!) has ultimately to pin his faith upon the informal system. Norma shows where to put the pins. 13. Semantic Problems Solved? The kind of logical language outlined above is potentially capable of capturing the complexities of real social behaviour. As ointed out, if you want to simpllfy, you can do so. But, if you decide to simpllfy, you K ave no alternative but to take responsibility 236 Ronald K . Stamper for imposing your simplification. At least we may, in the new framework, choose to describe our legal system as a mechanical bureaucratic structure but the new kind of logic proposed will not force us to do so. The mechanical nature of the classical logic approach, as observed above, intro- duces semantic problems. They were classified as either theoretical or as by-products of applications to expert systems. Let us look again at those points and see how far the proposed new framework might solve them. The theoretical problems identified relate to the concepts of truth, individuality, identity, extension, and the retreat into syntax. Truth is to be abandoned as a primitive concept. Truth, in the context of social behaviour is a very sophisticated notion that depends upon some responsible agents reaching a consensus on the matter in hand. Existence here and now takes the place of truth as a primitive concept, with the benefit of being a concept that can readily be operationalised. Individuality and the identification of individuals are also regarded as sophisticated concepts that an agent discovers only after a few years of experience which includes playing such games as peek-a-boo! In the legal domain it would be foolish to assume that all questions of individuality and identity are intrinsically resolved in some objective reality, whilst all we have to do is to uncover these facts. Heraclitus's river is the same river if all he does is to walk along its banks with his dog or gaze at it from the same bridge, but if, employed by the local water authority, he has the job of tracing its contents on their way to the sea, his different behaviour repertoire in this different context gives the word "river" a different meaning. If the concept of an individual is lost, so is the concept of a set of individuals defined extensively by enumerating its members. The idea of a behavioural invariant is close to the idea of an intension. Finally, we spurn the evasion of semantic issues by retreating into syntactics. All these theoretical advantages are justified by their contribution to giving us an operational semantics. The application problems that commonly arise when predicate logic is used for building expert systems were seen to be numerous. They are almost entirely caused by the freedom to formulate fat predicates. This liberty leads the analyst to invest a private language, introducing complexity, system rigidity, artificially enforced uniformity, the illusion that the machine uses natural language, and explains many of the difficulties in maintaining and merging systems. The new framework requires us to analyse legal norms down to the level of individual words or equivalent semantic elements and to place these in an ontological structure that goes a long way towards removing the arbitrariness of the analysis. This analysis gives us flexibility, a genuine uniformity to which diversity can be added by linking different inter- pretations to different agents, and insight into the operational meanings of our natural vocabulary. These benefits arise from an insistence on having an operational semantics and ruling out-of-order mathematical shortcuts, however elegant they may appear. 14. Jurisprudence in Support of the New Framework As a coda, we may ask whether the quest for a radically different kind of logic is justified by the arguments of legal philosophers who cast doubt on the notion that the law is a system of rules. This task is made easier by Susskind's excellent book (Susskind 1987). He has marshalled their principal arguments preparatory to dismissing each of them in turn. Susskind rejects them on the grounds that, despite their force, they still leave open the possibility of useful practical applications for mechanical devices that can simulate deductive reasoning within a system of explicit rules. I agree with his thesis but prefer to classlfy this residual set of applications as Semantics in Legal Expert Systems 237 “bureaucratic expert systems.” However, the arguments he dismissed provide strong support for pursuing the very different line of thinking advanced here which can lead to systems supporting the decision-maker or adviser faced with the semantic problems that the bureaucratic expert system leaves untouched. The courts seldom concern themselves with disputes about the accuracy of a chain of deductive argument, but rather adjudicate on disputes about meaning or the assignment of liability either at the substantive level or at the level of procedure. Legal problems that go before the courts may certainly entail a measure of deductive reasoning but much of the required reasoning can only be carried out by the exercise of creative thinking combined with judgment and accountability. This kind of reasoning cannot be mechanised. However, we can support the person engaged in reasoning about meaning or responsibility if we use a logic that deals with the semantic elements of natural language and which does so by clarifying who has the authority to convert words into actions either directly (the hangman) or at several removes (the clerk of the court). I propose that we explore the line of thinking described here as a new framework in order to build systems to support the lawyer in his quintessential legal work. The names of the arguments cited below are those coined by Susskind and the quotations are of his words or from the authors he has quoted. Just two of his arguments are considered but the others would also support the position advanced here. Argument from Act of Will (Susskind 1987, 174). “Hart captures the thrust of the argument in the comment that ’rules cannot provide for their own application, and even in the clearest case a human being must apply them’.” Susskind also quotes Hart saying ”fact situations do not await the judge neatly labelled with the rule applicable to them” and ”rules cannot claim their own instances.” No self-contained logical system can supply the will that Hart refers to or the “judgment” that Kant, an earlier proponent of this view, points out that logic does not provide. However, the proposed logic of action, Norma, is being devised to account for open systems. It cannot be treated as a closed system and its syntax is designed to force the analyst to draw attention to the agents who must supply the ”will” and the “judgment.” In fact, because it presupposes a responsible authority for every realisation, Norma provides for a link reaching from the formal structure held by the computer to the informal, human system, for each of the cases cited by Hart requiring an agent responsible for: 1. existence of the norm - the agent who successfully began the norm’s existence 2. expressing the facts in terms employed in the norm - the author of the sentences expressing the facts 3. invoking the norm - the agent responsible for conjoining the facts and the norm 4. communicating the result with due authority - the one who is empowered to perform the communication act A given body of legal norms (some tax law, for example) may not deal with these issues directly, relying on other statutes. The Norma syntax provides a place for this information, and it is up to the analyst to use when appropriate. The third case has been touched upon already. It was earlier noted what a grave assumption is made in the belief that the laws of logic apply categorically in m y formalised legal knowledge-base. Should this assumption be made, then someone must explicitly take responsibility for it. For example, if you want to use the rule of modus ponens then you may, but you will have to incorporate that rule quite explicitly 238 Ronald K. Stamper into the systemanducknowledge your responsibilityfor doing so, together with your indirect responsibility for the results it generates. Judges do not indiscriminately link facts and rules to draw "the logical conclusions" because the results can be unjust. Logic is a mechanical model of language; there is no licence to drive it everywhere except perhaps in the Platonic world of mathematics.I6 It is interesting to note that Susskind's expert system for the Scottish law of divorce does have an informal treatment of the notion of authority. This provides a partial answer to the question of who supplies the will to make the norm system function. I am proposing that a formal treatment of authority should be a necessary component of any logic to model social or legal affairs. Argument from principles (Susskind 1987, 169). "Dworkin argues that lawyers and judges, while solving legal problems, as well as reasoning with rules, often also have recourse to non-rule standards that he terms 'principles' ." These principles do not lead to necessary conclusions but lend weight or probability to certain possible conclusions. Often they are formulated at a far higher level of generality, with unlimited exceptions hence incapable of being reduced to rules. "There is no such thing as 'the law' as a collection of discrete propositions." Hart goes so far as to say that "it remains the judge's duty, even in hard cases, to discover what the rights of the parties are, not to invent new rights retrospectively." Principles are essential to this task. This argument points to an essential openness in legal reasoning. A mechanical, deductive system will sooner or later break down. Comparing it with a machine in a factory, the operative may have done all he can to make it work, but a point comes when he must hand over to the engineer or manager to take actions that have never been prescribed in the handbooks. Engineering and managerial expertise, like the lawyers' expertise, goes beyond the books to the principles they have learned. A solution to the problem has to be found; they cannot opt out. If one accepts Dworkin's views, one might attempt to accommodate these principles as additional rules together with a host of illustrative exceptions. To do so would be to fall into the Artificial Intelligence trap: The ridiculous attempt to usurp the roles of people.17 Applying classical logic to expert systems perhaps encourages the tendency to fall into the Artificial Intelligence (AI) trap by compelling one to work in a closed system. The same computing techniques (the genuine, non-ideological contribution of AI) can be employed to offer social groups a medium in which they can collaborate more effectively. That is the spirit in which the new framework is proposed. It permits any desired degree of automation but it links every relevant decision point to the human agent involved and so acknowledges the social context of legal and other practical affairs. However, it will not permit closed models to be formulated, but always it involves the informal system which sustains the principles to which Dworkin has rightly drawn our attention. l6 Even in mathematics, there is not a universal agreement about the logical rules which can be applied legitimately (Bloor 1976; Haak 1978; see the discussion of Brouwer in Kneale and Kneale 1978, and Lakatos 1979). " Very few AI practitioners among themselves fall into this trap but they are not always scrupulous about warning the less sophisticated where to tread. Some are even prepared, in their advertising copy, to express themselves in a manner that might mislead the layman into a belief that the expert systems they are selling can, in effect, place the relevant expert on the user's desk. Susskind himself(1989), sadly, falls into this excess in the opening paragraph of an article on his Latent Damage expert system: "Imagine the Chairman of the Law Faculty of Oxford University sitting on the desk of every lawyer in the land and always ready for consultation. " Semantics i n Legal Expert Systems 239 As Dworkin indicates, the judges uncover the relevant law that exists within the informal legal system. He argues that we cannot make a complete, explicit representation of a system of general, probabilistic standards that have unlimited exceptions. There is no need to pretend that we can. A logic which enshrines openness in its syntax denies the pretence. The openness of the legal system means that, no matter how much of the law and the given situation you have reduced to explicit formulas and rules, much more remains unsaid. But a responsible agent can invoke the unspoken principles within the informal world of law-in-action, and stop a deductive chain that would otherwise lead to a miscarriage of justice. Argument from Particularity of Facts. Susskind (1987, 181) quotes Hart again: “Fact situations do not await us neatly labelled, creased, and folded; nor is their legal classification written on them to be simply read off by the judge. Instead, in applying legal rules, someone must take the responsibility of deciding what words do or do not cover some case in hand, with all the practical consequences involved in this decision.” We have already dealt with the responsibility issue but the new issue raised by this quotation concerns the problem of choosing descriptive terms in which to report the facts of a particular case. “The facts of any case can always be characterised in a vast number of different, highly specific ways . . .” but the vocabulary we employ will be related to the norms invoked. Susskind appears to favour a scheme in which the user and the law relating to his case are kept so far apart that they do not interact until the deductive process begins. “It is possible, in rinciple, to have a system with such extensive knowledge would be so far removed from judgments of law that it would be obtuse to withhold the term ‘deductive’ in relation to the process of legal reasoning executed by the computer” (Susskind 1987, 181). The proposed new paradigm incorporates the principle that meanings are relationships between linguistic structures and behaviour, hence we are led to adopt the position that the reporting of the facts will justifiably always be coloured by the norms. It is obvious that the terms used to characterise a case must take some of their meaning from the norms that are considered relevant, simply because the behaviour induced by using the chosen descriptors will be mediated by the norms. Children are fond of playing this kind of semantic trick: “What’s in my hand? Get it right and I’ll give you a penny.” The prankster is safe, the answer “Nothing” will bring the riposte “Wrong! I’ve got air in my hand” but the answer “Air“ leads to “No. My hand is in the air but nothing is in my hand.” The same kind of semantic manipu- lation forms the basis of many a confidence trick and dubious sales technique. Unless you know the behaviour that will result from the various formulations from which one might reasonably select a description of a case, one does not know the meanings of the words to be employed in that socie reveal the links between the choice of words and their legal consequences. By attempting to lock legal knowledge inside a black box we do the opposite. Knowledge, justice and other important social constructs have to be understood and constantly rehearsed in a society even to exist. Expertise obscured from critical appraisal inside computer packages should not be considered acceptable. that any judgments ma B e in relation to questions actually answered by the user in that legal context. For a computer system to help with this 1 ind of semantic problem-solving it must Appendix A Simple Example of the Semantic Analysis and Norm Analysis Techniques of MEASUR: Canadian Unemployment lnsurance Law - Revised Statutes of Canada (2985), Ch U-1 To illustrate the method of Semantic Analysis outlined in this paper, here is an example taken from Mackaay et al. (1990). 240 Ronald K . Stamper Work on the formalisation of legal systems to handle the law has tended to neglect the problems of methodology that become serious issues when one scales up “desk” research to the size of real systems. The LEGOLlNORMA research programme has generated a methodology suitable for such problem areas. It is called MEASUR. It incorporates a variety of analytical techniques that take one from the articulation of somewhat vaguely stated needs (for example, at the earliest stages of legal drafting), through the analysis of meanings, to the formulation of norms. The second and third phases are illustrated below. Two further stages deal with the elaboration of relevant procedures and appropriate sanctions. Of all the techniques included in MEASUR, that of Semantic Analysis has proved to be the most important. Phase I - Semantic Analysis Step I: Generate list of candidate semantic units The problem domain has to be reduced to semantically atomic units each of which corresponds to an invariant in the behaviour of some responsible agent. The analyst begins by listing candidate semantic units for the problem or application domain. Here we select the significant items from the small sample of the act covered by this example. Great care is not important at this stage. We could select all the words in the text not accounted for by functions within the system. Thus: Sec.l8.(1) [. . .]the qualifyingperiod of an insured person is the shorter of (a) the period of fifty-two weeks that immediately precedes the commencement of a benefit period under subsection (1) of section 20, and (b) the period that begins on the commencement date of an immediately preceding benefit period and ends with the end of the week preceding the commencement of a benefit period under subsection (1) of section 20. Sec.20. (1) A benefit period begins on the Sunday of the week in which (a) the interruption of earnings occurs, or (b) the initial claim for benefit is made whichever is the later. Step 2: Classify the candidates This is not a significant task for this example but it helps to look for the following classes, among others. 1. responsible agents, e.g., person 2. roles indicative of agents (e.g., claimant) 3. relationships indicative of missing components, e.g., insured which requires two 4. periods used in formulating conditions, e.g., qualifying period, week, date 5 . particular values of measuring or classification frames, called “determinants, ” 6. measurements or classification frames, called “determiners,” e.g., day (or day of Our attention is now drawn to week as a unit of measurement of a duration (a determiner applicable to every invariant) 52 weeks may be a determinant for a duration of a period. Other questions are suggested. Is week, in the sense of a period named by the current state of a calendarlchronometer, different in meaning from week in the expression “52 weeks”? We need to check the interpretation of week - does it have agents person and stute so we add state to our candidate list e.g., Sunday which may be indicative of missing determiners week). Semantics in Legal Expert Systems 241 a standard meaning as an interval lasting 7 days of duration starting on a Sunday or does it have a different meaning in the context of the act? [We shall treat 52 weeks as a determinant.] 7. semiological behaviour - e.g., claim which is a "speech act" performed by a person using a sign, the meaning of which is the benefit he or she wants. Step 3: Create an ontology chart Time is central to our understanding of meaning in the context of this method. Every word or expression that is a semantic unit corresponds to an invariant in the flux of events or behaviour of the agents. Thus, to know the meaning of a word entails knowing when the corresponding invariant starts and finishes. The existence of each invariant depends upon the existence of others, hence we can clanfy meanings by constructing an ontology chart to demonstrate these dependencies. See Fig. 1. Figure 1: A specific ontology char! for a fragment of legislation Notes: 1. The responsible agents are given in capital letters. 2. Each element names a universal for which there will be numerous particulars. 3. Determiners are prefixed by #, their determinants are omitted here just as names of individuals are omitted. 4. Signs, such as "benefit" have the relevant linguistic community as their antecedent. 5. The invariants inbehaviour are often dependent o n others (see the discussion in the first few paragraphs dealing with the "New Framework" in the main body of the paper, above). Thus benefit canonly exist as a joint invariant of PERSON and STATE. It may appear that the insurance contract between PERSON and STATE should be an antecedent of benefit but the entitlementlobligation referred to as the benefit, though created by the legislation, will not necessarily cease to exist if the legislation is repealed, it may be paid subsequently and cease to exist then. 6. Notice that PERSON, STATE and the sign "benefit" have no stated antecedents. As the root agent we may assume thecommunity at large, as the custodian of commonsense meanings. We may supplement this with other ontology charts to clanfy meanings of certain elements. Some such fragments of analysis are very general and belong in a library of semantics. For example 242 Ronald K. S t a m p e r payments # benefit- premium- Figure 2: A generic ontology chart for "benefit-payment" and "premium-payment" Notes: 1. The antecedent for money is most probably the STATE but it could be the responsibility of another agent (e.g., ECUS in the EEC). 2. The box under payment lists specific forms including benefit. 3. Notice that payment is a process taking place during the co-existence of receiving and relinquishing acts. 4. Ownership enters here - the processes of beginning ownership = receiving 5. Beginning and ending processes are quite different ontologically from start to finish times which are signs indicating when, respectively, a beginning and an ending were successfully completed. ending ownership - relinquishing. Phase 11 - Norm Analysis Having made it clear what we are talking about by constructing an ontology chart showing the relevant words organised as a structure of behavioural invariants, we can begin to spec+ the authorities governing behaviour in our problem domain. I n the case of an informal system, we should probably only indicate the agents responsible for deciding when instances of each type of invariant start and finish their existence. In the case of a formal or legal system, many of the authorities will be norms or agents empowered (conditionally) to take these decisions. In this example we have two legal norms. One defines the start of a benefit period and the other, consisting of two logical norms, defines the start and finish of a benefit period. In LEGOL, they would be expressed as follows: Context insurance (STATE,PERSON) Start ofbenefit period [20.(1)] <- day of( # Sunday(day) while (week while lust of(finish of earning omhile start of claim)) Note that the context operator specifies that the norm is applied to each individual case of its operand. This defines the antecedents of benefit period, week, earning, and claim. This simplifies the expression of the norm. Also note that all the logical operators take account of time. The system has always been implemented with Semantics in Legal Expert Systems 243 one-sided open time intervals, the start being part of the period but not the finish. One needs to check the formal interpretation. The benefit period is tagged with the reference to the norm defining it. Of course, start and finish may have different authorities. The other two norms can be specified within the additional context of a single benefit period. Context add benefit period Finish of qualifying period [18.(1)] <- start ofbenefit period Start of qualifying period [1&3.(1)] <- lust of ((finish of quallfylng period minus 52 weeks) orwhile start of prior benefit period) The efficiency of evaluation would be improved by defining a function the previous to select only last of the prior instances of an invariant type. Norms themselves have periods of existence and so do the universals in the ontological structure (the analo e a database schema). Thus changes of meaning can be handled easily. The onto Y ogical structure, however, tends to be very stable, only changing with major paradigm shifts. Hence this method has advantages in the design of large administrative systems where restructuring massive files is often too great a task to contemplate. Once an unsuitable database structure has been implemented one is struck with it and with the resulting inflexibility of the system. Using the method of semantic analysis, one obtains a structure that can easily be extended in scope or detail, leaving the associated norms unaffected. The amount of thought that must go into the preparation of the semantic scheme before the system is built is the extra "up-front" investment, but this thought can help to uncover some suspected legal problems before drafting errors are made, and so bring immediate dividends. University of Twente Department of Information Management P.O. Box 227 NL-7500 A€ Enschede The Netherlands References Aitchison, Jean. 1987. Words in theMind: An Introduction to theMental Lexicon. Oxford: Blackwell. Bloor, David. 1976. Knowledge and Social Imagery. London: Routledge and Kegan Paul. ---. 1983. Wittgenstein: A Social Theory of Knowledge. London: Macmillan. Camap, R. 1942. lntroduction to Semantics. Cambridge, Mass.: Harvard University Press. Davis, P. J. and R. Hersh. 1983. The Mathematical Experience. Harmondsworth: Pelican Books. Dowty, D. R., R. E. Wall and S. Peters. 1981. Introduction to Montague Semantics. Dordrecht: Dreyfus, J. L. and S. E. Dreyfus. 1986. Mind over Machine. Oxford: Blackwell. Gibson, James, J. 1979. The Ecological Approach to Visual Perception. Boston: Houghton Mfiflin Haak, Susan. 1978. Philosophy of Logics. Cambridge: Cambridge University Press. Jones, S., P. Mason andR. K. Stamper. 1979. Legol-2.0: ARelational Specification Language for Kneale, W. and M. Kneale. 1975. The D a t e l o p e n t of Logic. Oxford: Oxford University Press. Kowalski, R. 1979. Logic for Problem Solving. Amsterdam: North Holland. Lakatos, I. 1976. Pro@ and Refutations. Cambridge: Cambridge University Press. Lakoff, G. and M. Johnson. 1980. Metaphors We Live By. Chicago: University of Chicago Press. Mackaay, Ejan, Daniel Poulin, Jacques Frbmont, Paul Bratley, and Costant Deniger. 1990. The Reidel. Company. Complex Rules. Information Systems 4. Logic of Time in Law and Legal Expert Systems. Ratio Juns 3: 254-71. 244 Ronald K. Stamper Mead, G. H. 1932. The Philosophy of the Present. Chicago: University of Chicago Press. Michaels, C. F. and C. Carello. 1981. Direct Perception. Englewood Cliffs: Rentice-Hall. Michie, D. and R. Johnston. 1984. The Crentiw Computer. New York: Viking. Montague, R. 1974. Formal Philosophy. Ed. R. H. Thomason. New Haven, Conn.: Yale Niblett, B., ed. 1980. Computer Science and Law. Cambridge: Cambridge University Press. Ortony, A., ed. 1979. Metaphor and Thought. Cambridge: Cambridge University Press. Reddy, M. J. 1979. Conduit Metaphor - A Case of Frame Conflict in Our Language About Language. In Metaphor and Thought. Ed. A. Ortony. Cambridge: Cambridge University Press. Ser ot, J. J . , F. Sadri, R. Kowalski, F. Kriwawek, P. Hammond and H. T. Co . 1986. The gritish Nationality Act as a Logic Programme. Communications of the A.C.M. 2 7 Stamper, R. K. 1991. Expert Systems - Lawyers Beware. In Law, Decision-Making and Microcomputers. Ed. S . S . Nagel. Westport, Conn.: Quorum Books. Stamper, R. K. 1980. Legol: Modelling Legal Rules by Computer. In Computer Science and Law. Ed. B. Niblett. Cambridge: Cambridge University Press. -. 1985. Information: Mystical Fluid or a Subject for Scientific Enquiry. The Computer Susskind, R. E. 1987. Expert Systems in Law: A Jurisprudential Inquiry. Oxford: Oxford University University Press. ]ouml28. Press. .1989. An Expert Lawyer On Every Desktop. The Independent London. Tagg, Clare. 1979. The Impact of Data Modelling on LEGOL. (Internal Report.) London: London Tarski, A. 1965. Introduction to Logic. New York: Oxford University Press. Twining, W. and D. M. Myers 1976. How to Do Things with Rules. London: Weidenfeld and Weber, Max. 1946. Essays in Sociology. Trans. and ed. H. H. Gerth and C. Wright Mills. Oxford: Whitehead, A. N. and B. Russell. 191011927. Principia M a t h m t i c a . Cambridge: Cambridge School of Economics. Nicolson. Oxford University Press. University Press. 
work_mpmu7qcvmbgg3mkqpinkvrs7km	----	Personal Identification With Iris Patterns Raf. J. of Comp. & Math’s. , Vol. 7, No. 2, 2010 129 Design a Fuzzy Expert System for Liver and Pancreas Diseases Diagnosis Baydaa S Bhnam baydaa_sulaiman@uomosul.edu.iq College of Computer Sciences and Mathematics University of Mosul, Mosul, Iraq Received on: 12/04/2010 Accepted on: 19/08/2010 ABSTRACT Fuzzy logic is a branch of artificial intelligence techniques, it deals with uncertainty in knowledge that simulates human reasoning in incomplete or fuzzy data. Fuzzy relational inference that has applied in medical diagnosis was used within the medical knowledge base system to deals with diagnostic activity, treatment recommendation and patient's administration. In this research, a medical fuzzy expert system named (Liv&PanFES) has been developed for diagnosis and decision making of general Liver and Pancreas diseases. The (Liv&PanFES) is a rule based fuzzy expert system, results of laboratory analysis are inserted into the system. This system can define the probable diagnosis on these data, and later on it can pick out the most probable one for disease. Keywords: Artificial Intelligence, Fuzzy logic, Fuzzy Expert system تصميم نظام خبير مضبب لتشخبص امراض الكبد والكلى بيـداء سليمان بهنام العراق ،جامعة الموصل، كلية علوم الحاسبات والرياضيات 19/08/2010 تاريخ قبول البحث: 12/04/2010تاريخ استالم البحث: الملخص المنطققا المبققو رققو ققري مققل كانيققات القق فاو اسصققطناعم، مققو لمققبة م المعلومققات المبققوبة و ققر إن العالقققات المبققوبة كمقققبة م ققم البطةيةققات الطويقققة اعبمققا ا علققق إن. اإلنمقققانعمققل أسققلو الكاملققة ققم محاكقققا ات الةاصة بالطةص المعنم.المعلومات الممبةلةة مل الفحوص نظقام مبقو يو قر لق يل قم الملقاي الطوقم ويمقاع قم (Liv&PanFES) م ر ا البحث كم كةميم الكو والونكرياس. أمراضاكةاذ الارار فيما لبعلا ببطةيص قققم كح لققق نوميقققة المقققرض المةقققا بققق (Liv&PanFES)لعبمققق كطقققةيص النظقققام الةو قققر المبقققو إل الطائعة البم كعو األمراضائج البحال ل المةبورية، ويمبة م النظام م اكةاذ الارار وكطةيص المريض عل نب لكل مما. أوالونكرياس أوالكو .، نظام يو ر مبو ال فاو اسصطناعم، المنطا المبو الكلمات المفتاحية: 1. Artificial Intelligence in Medicine Artificial Intelligence (AI) is a study to emulate human intelligence into computer technology. The potential of AI in medicine has been expressed by a number of researchers [8][10][17] and [19] summarized the potential of AI techniques in medicine as follows: ▪ Proved a laboratory for the examination, organization, representation and cataloging of medical knowledge. ▪ Produces new tool to support medical decision-making, training and research. ▪ Integrates activities in medical, computer cognitive and other sciences. Baydaa S. Bhnam 130 ▪ Offers a content-rich discipline for future scientific medical specialty. 2. Fuzzy Control in Medicine Fuzzy control techniques have recently been applied in various medical processes. Fuzzy control compared to classical control theory, which is a fuzzy logic approach to control, offers the following advantages: ▪ The main advantage of using fuzzy control system is the simplicity of the approach , the capacity of dealing with the complex data and solves a problem-control system Methodology [7][14]. ▪ It can be used in systems, which don't need to measure the parameters of the model and cannot be easily modeled mathematically. In particular systems with non-linear responses that are difficult to analyze may respond to a fuzzy control approach. Hence, many complex systems can be controlled without knowing the exact mathematical model of them [3][12]. ▪ Using fuzzy control to identify the diseases form the symptoms which helps to develop Fuzzy rules that can be stored in the knowledge base and can be fired during further decision process [4]. ▪ Continuous variables may be represented by linguistic constructs that are easier to understand, making the controller easier to implement and modify [11]. ▪ Fuzzy controllers may be less susceptible to system noise and parameter changes, in other words, they will be more robust. ▪ Complex process can be controlled by relatively few logical rules permitting an easily comprehensible controller design and faster computation [12]. ▪ In other word, fuzzy control can be best applied to production tasks, that heavily rely on human experience and intuition, and which therefore rule out the application conventional control methods [11][14]. 3. Medical Knowledge as a Fuzzy Relation The knowledge base, which is the central to any Fuzzy Logic System, represents the domain knowledge. A knowledge base contains two types of knowledge: facts and rules. The facts represent various aspects of specific domain that are known before consultation of the system. A rule is an if-then structure that logically relates information contained in the if part to other information contained in the then part [10]. The relationship between symptoms and diagnosis by the concept of medical knowledge was introduced. "In a given pathology, we denote S a set of symptoms, D a set of diagnoses and P a set of patients. What we call medical knowledge is a fuzzy relation, generally denoted by R, from S to D expressing of diagnoses". Zadeh's max- min compositional rule adopted as an inference mechanism. It accepts fuzzy descriptions of the patient's symptoms and infers fuzzy descriptions of the patient's diseases by means of the fuzzy relationships. If a patient's symptom is Si then the patient's state in terms of diagnoses in a fuzzy set Dj with the following membership function [7][15][16]:   DdSsdssd Rs Ss D ij =  ,,),();(minmax)(  ),( ds R  is the membership function of the fuzzy relation "medical knowledge". With P, a set of patients, and a fuzzy relation Q from P to S, and by "max-min" composition" we get the fuzzy relation T=Q×R with the membership function [3][21]:   DdSsPpdsspdp RQ Ss T =  ,,,),();,(minmax),(  Design a Fuzzy Expert System for Liver and Pancreas Diseases Diagnosis 131 4. The Designed Liver & Pancreas Fuzzy Expert system(Liv&PanFES) The design structure, application and working principles of Liver& Pancreas Fuzzy Expert system (Liv&PanFES) has been described for diagnosis of general Liver & Pancreas diseases. The (Liv&PanFES) is parallel rule-firing system, in which all fireable rules are fired effectively at one time. During the process with the medical expert system (Liv&PanFES), laboratory test results are converted into fuzzy compatibility values reaching from zero to unity by consideration of the linguistic medical concepts. These fuzzified data are used to infer diagnosis with knowledge contained in a knowledgebase. Fuzzy relations were calculated for all linguistic medical concepts between test results and diagnosis by using the obtained fuzzy sets with the given set of patient data. 4.1 Medical Knowledgebase The examination and laboratory data of the system assigned for the Liver&Pancreas patients were collected from Al-Jamhoree Teaching Hospital. The data then inserted into the system (Liv&PanFES), for the design process: The Total Serum Bilirubin (TSB), (TSB-direct), Serum Glutamic Pyruvic Transaminase (SGPT), Serum Glutamic Oxaloacetic Transaminase (SGOT) and Serum Alkaline Phosphataste (SAP) examinations are used for diagnosis Hemolytic Anemia (HA), Jaundice (JAU), Hepatitis (HEP), Myocardial Infarction (MYI) and Osteomalacia (OST) diseases; which are related with liver diseases. The Fasting Blood Sugar (FBS) and Random Blood Sugar (RBS) examinations are used for diagnosis Hypoglycemia (HYPOGL), Hyperglycemia (HYPERGL) diseases; which are related with pancreas diseases. The unit that used for all examinations is (mg/dl ). 4.2 Inference Methodology The fuzzy control application structure as shown in Figure (1) has seven crisp input variables , these input parameters represent the patients data. The mamdani model of inference was used. All (Liv&PanFES) fuzzy rules were of the form: " If x1 is A1and y1 is B1 then z1 is C1 ", Where A1 , B1 and C1 are fuzzy sets. The max-min operators were used for implication throughout for implementing (Liv&PanFES). It was necessary to obtain crisp output for the purposes of evaluation of the fuzzy model , center-of- gravity (centroid) defuzzification was used to produce crisp values on an arbitrary scale of the fuzzy output variable (OutFES) . 4.3 Linguistic Variables and Fuzzy Terms Each of the seven input parameters was assigned a linguistic variable and examination of the data and rules showed that each could naturally be divided into two or three fuzzy terms corresponding to meanings of Low (L), Normal (N), and High (H) for the types laboratory examinations. One output fuzzy variable was used, from the rules it was determined that the output fuzzy variable (OutFES) has eighteen linguistic terms: (Normal, HA, JAU, HEP, MYI, OST, HYPOGL, HYPERGL, HA&HYPOGL, HA&HYPERGL, JAU&HYPOGL, JAU&HYPERGL, HEP&HYPOGL, HEP&HYPERL, MYI&HYPOGL, MYI&HYPERGL, OST&HYPOGL, OST&HYPERGL), these terms represent the general Liver and pancreas diseases. Baydaa S. Bhnam 132 4.4 Rule Set Liv&PanFES must contain a set of rules that can deal with fuzzy tags, fuzzy sets and relations and must provide an appropriate output which corresponds to a particular input. The rules for the fuzzy expert system were obtained by means of knowledge expert-doctor, and had been carefully refined to form a complete and consistent set of classifiers. Part of the developed (Liv&PanFES) fuzzy knowledge base rules are shown in Table (1), total of 90 rules are formed as shown in the appendix. Table(1): Liv&PanFES fuzzy rules OutFES RBS FBS SAP SGOT SGPT TSB (direct) TSB Rule No. Normal N N N N N N N 1 … … … … … … … … … MYI N L N H N N N 26 … … … … … … … … … OST N H H N N N N 36 … … … … … … … … … HEP&HYPOGL L L N H H H H 40 … … … … … … … … … OST&HYPERGL H H H N N N N 90 `For example Rule 40 can be interpreted as follows (as in Table (1)): Rule 40: if patient's TSB is High, patient's TSB-Direct is High, patient's SGPT is High, patient's SGOT is High and patient's SAP is Normal, patient's FBS is Low, patient's RBS is Low, then patient has Hepatitis (HEP) and Hypoglycemia (HYPOGL) diseases; which are related with liver and pancreas diseases. 4.5 Fuzzy Operations The probabilistic family of operators was chosen for the reason that all the rules features conjunction of the input linguistic variables use the fuzzy conjunction operators. The min operator is used for conjunction , the overall truth value of such a rule will obviously by determined solely by the reasoning of the clinician in considering all the parameters. After the truth degree of each rule is determined, we calculate the truth degree of all rules by taking max between working rules. 4.6 Membership functions In this paper, triangular and trapezoidal membership functions are used. Fuzzy membership functions for the medical factors for the mentioned laboratory parameters are calculated by the following functions as shown in Figures(2) and (3). FBS SGPT TSB(direct) RBS TSB TSB(indirect) SAP SGOT SGO T Figure(1): The structure of the Liver&Pancreas fuzzy expert system(Liv&PanFES) Design a Fuzzy Expert System for Liver and Pancreas Diseases Diagnosis 133 Normal (0,10) ; OST (50,,60) ; JAU&HYPOGL (100,110) ; MYI&HYPERGL(150,160) HA (10,20) ; HYPOGL (60,70) ; JAU&HYPERGL (110,120) ; OST&HYPOGL(160,170) JAU (20,30) ; HYPERGL (70,80) ; HEP&HYPOGL (120,130) ; OST&HYPERGL(170,180) HEP (30,40) ; HA&HYPOGL (80,90) ; HEP&HYPERGL (130,140) MYI (40,50) ; HA&HYPERGL (90,100) ; MYI&HYPOGL (140,150) Figure(3): The membership functions of output term (OutFES) for Liv&PanFES Figure(2):The membership functions of input terms for Liv&PanFES Baydaa S. Bhnam 134 5. Experimental Results For the output factor OutFES, the linguistic expressions represent the patients disease. The truth degree of the rules are determined for each rule by aid of the min and then by taking max between working rules. For example, for the input patient laboratory tests values TSB=13.5, TSB(direct)=4, SGPT= 42, SGOT =22.5, SAP=65, FBS =78 and RBS= 120. The rules (4) and (5) is fired as shown in Figure (4). Figure(4): Calculation of the value OutFES represents the diagnosis of a patient affected by HEP depending on the following laboratory test values (TSB=13.5, TSB(direct)=4 , SGPT=42 , SGOT =22.5 , SAP =65 , FBS =78 , RBS= 120 ). From Mamdani max-min inference we will obtain the membership function of our system, max (rule 4, rule 5) = rule 5, then we calculate the crisp output. The crisp value of the output is calculated by the method of center of gravtity defuzzifier. As it is seen from Figure (4), the value of OutFES=35, this means that the patient has the Hepatitis (HEP) disease (see Figure(3)). Another examples for implementation of the (Liv&PanFES) system for patients disease diagnosis shown in figures (5,6,7,8,9,10,11,12 and 13) with patients laboratory tests values, fired rules (16), (54), (1), (3 and 29 then rule3 is max), (89), (19), (72 and 81 then rule 72 is max), (71) and (56) are values of output term (OutFES) that refer to the diagnosis of the patient complaint. The performance of the model is tested for 100 patients, the results show that the all diagnosis using (Liv&PanFES) system is correct. Figure(5): Calculation of the value OutFES represents the diagnosis of a patient affected by HYPERGL depending on the following laboratory test values (TSB=0.5, TSB(direct)=0.35, SGPT=6.5, SGOT =11, SAP =31, FBS=52, RBS=190) Design a Fuzzy Expert System for Liver and Pancreas Diseases Diagnosis 135 Figure(6): Calculation of the value OutFES represents the diagnosis of a patient affected by OST&HYPOGL depending on the following laboratory test values (TSB=0.5 , TSB(direct) = 0.3 , SGPT = 12 , SGOT =14, SAP =92 , FBS =80, RBS= 72). Figure(7): Calculation of the value OutFES represents the diagnosis of a patient affected by NORMAL depending on the following laboratory test values (TSB=0.6, TSB(direct)= 0.15 , SGPT =9, SGOT =7.5, SAP =46, FBS =75 , RBS=110 ). Figure(8): Calculation of the value OutFES represents the diagnosis of a patient affected by JAU depending on the following laboratory test values (TSB=3, TSB(direct)= 0.85, SGPT =21, SGOT =15.5, SAP =62, FBS =115, RBS=96). Baydaa S. Bhnam 136 Figure(9): Calculation of the value OutFES represents the diagnosis of a patient affected by HEP&HYPERGL depending on the following laboratory test values (TSB=13.5, TSB(direct)= 6, SGPT=34, SGOT =23, SAP =775, FBS=120, RBS=200). Figure(10): Calculation of the value OutFES represents the diagnosis of a patient affected by HA depending on the following laboratory test values (TSB=16.5, TSB(direct)= 0.2, SGPT=12.5, SGOT =9, SAP =37, FBS=46, RBS=89). Figure(11): Calculation of the value OutFES represents the diagnosis of a patient affected by OST&HYPERGL depending on the following laboratory test values (TSB=0.9, TSB(direct)=0.3, SGPT=19, SGOT=2 , SAP=101, FBS=60, RBS=180). Design a Fuzzy Expert System for Liver and Pancreas Diseases Diagnosis 137 Figure(12): Calculation of the value OutFES represents the diagnosis of a patient affected by MYI&HYPERGL depending on the following laboratory test values (TSB=0.7, TSB(direct)=0.25, SGPT=17, SGOT=31.5, SAP=83, FBS=52, RBS=220). Figure(13): Calculation of the value OutFES represents the diagnosis of a patient affected by JAU&HYPOGL depending on the following laboratory test values (TSB=6.5, TSB(direct)=14.5, SGPT=19, SGOT=7.5, SAP=71, FBS=160, RBS=76 ). 6. Conclusions The developed diagnosis module (Liv&PanFES) consists of expert system and fuzzy logic techniques to perform diagnostic tasks. A set of rules will be defined using the patients disease database as well as the expert knowledge on the disease domain (from doctors). The designed expert system uses the rules to diagnose patient's illness base on their laboratory tests. In addition , fuzzy logic is integrated to enhance the reasoning when dealing with fuzzy data. The combination of expert system and fuzzy logic that forms a hybrid (expert-fuzzy) system could increase the system performance and has been implemented successfully. Baydaa S. Bhnam 138 REFERENCES [1] Adlassing K.P. (1986), "Fuzzy set theory in medical diagnosis", IEEE Transactions Systems Man Cybernetics , Vol. 16, 2, pp. 260 – 265. [2] Adlassing K.p. and G.Kolarz, (1982), "CADIAC -2: computer – assisted medical diagnosis using fuzzy subsets", in Approximate Reasoning in Decision Analysis, MM Gupta and E Sanchez, Eds., pp. 219 – 247, North-Holland, New York . [3] Balwinder, S., Rajneesh, G., Rakesh ,K. and Singh, R.P.P, (2009), "Design and VLSI implementation of Fuzzy Logic Controller", IJCNS) International Journal of Computer and Network Security, Vol. 1, No. 3, December, pp.77-81. [4] Chinniah, P. and Muttan, S., (2009), "ICD 10 Based Medical Expert System Using Fuzzy Temporal Logic", (IJCSIS) International Journal of Computer Science and Information Security, Vol. 6, No. 3, pp.84-89. [5] Cohen M.E. and Hudson D.L., (1988), "The use of fuzzy variables in medical decision making", in Fuzzy computing M.M. Gupta and T. Yamakawa, Eds., pp. 263 – 271. Elsevier Science, North Holland. [6] Halim M., K. M. Ho, and A.Liu, (1990), "Fuzzy logic for medical expert systems", Annals Academy Medicine, Vol. 19, no. 5, pp. 672- 683. [7] Hamid, M., Dan, I., Jerome, B. and Bernadette, D., (2009)., "Human Activities of Daily Living Recognition Using Fuzzy Logic For Elderly Home Monitoring", FUZZ-IEEE, Korea, August 20-24, pp. 2001-2006. [8] Hoong, N.K., (1998), "Medical Information Science - Framework and Potential", International Seminar and Exhibition computerization for Development -the Research Challenge, University Pertain Malaysia: Kuala Lumpur, pp. 191 – 198. [9] Hughes, C., (1989)., "The representation of uncertainty in medical expert systems", Medical Informatics, Vol.14, pp. 269-279. [10] Khalil, S., (2008), "Performance Tuning of Novell Netware Based on Fuzzy Reasoning", International Journal of Computers, Issue 1, Vol. 2, pp. 80-86. [11] Klement, E.P. and Slany, W., (1995), "Fuzzy Logic in Artificial Intelligence", Proceedings of the 8th Austrian Artificial Intelligence Conference FLAI 93, Berlin. [12] Mamdani, E.H., Qstergaard, J.J. and Lembessis, E., (1989), "Use of Fuzzy Logic for Implementing Rule-Based Control of Industrial Processes", In: P.P.Wang: Advances in Fuzzy Sets, Possibility theory, and Applications, New York: Plenum Press, pp. 307-323. [13] Nguyen Hoang Phuong and Vladik Kreinovich, (2000), "fuzzy logic and it's applications in medicine", Institute of Information Technology, National Center for Natural Science and Technology of Vietnam. [14] Oscar, M., Roberto, S. and Patricia, M., Oscar, C., Miguel A. P. and Iliana, M.M, (2007),“Performance of a Simple Tuned Fuzzy Controller and a PID Controller on a DC motor”, In Proceedings of the IEEE Symposium on Foundations of Computational Intelligence, pp. 531-537. Design a Fuzzy Expert System for Liver and Pancreas Diseases Diagnosis 139 [15] Sanchez, E., (1976), "Resolution of Composite Fuzzy Relation Equations", Information and Control. 30, pp.38-48. [16] Sanchez, E., (1979), "Medical Diagnosis and Composite Fuzzy Relations", Advances in Fuzzy Set Theory and Applications, Amsterdam: North-Holland, pp. 437-444. [17] Serena, H. C., Anthony J. J. and John P. N., (2008), "Artificial Intelligence techniques: An introduction to their use for modelling environmental systems", Mathematics and Computers in Simulation, Vol.78, pp. 379-400. [18] Shivaraj, G., Prakash, B.D., Vinayak, V.H., Avinash, A K.M., Sonal, N.,V. and Shruthi, S., K., (2009), "A review on laboratory liver function tests", PanAfrican Medical Journal, Published 22 November, pp. 3175-3181. [19] Ulrich H. L.,(2009),"Humanoid Knowledge-Based Decision Support Systems (KBDSS)- Subjective or Objective Data Query", Journal of Advanced Computattional Intelligence and Intelligent Informatics, vol.13, No.1,pp.10-15. [20] Watanabe H., W.J. Yakowenko, Y. M. Kim, J.Ande, and T.Tobi, (1994), "Application of a fuzzy discrimination analysis for diagnosis of valvular heart disease", IEEE Transactions Fuzzy Systems, vol.2, no.4, pp. 267 - 276. [21] William Siler and James J. Buckley, (2005), " Fuzzy Expert Systems and Fuzzy Reasoning", Wiley Inerscience. [22] Zadeh A.L.,(1983),"The role of fuzzy logic in the management of uncertainty in expert systems ", Fuzzy Set Systems, Vol. 11, pp. 199- 227. Baydaa S. Bhnam 140 Appendix: Rule-base Knowledge for (Liv&PanFES) OutFES RBS FBS SAP SGOT SGPT TSB (direct) TSB Rule No. Normal N N N N N N N 1 HA N N N N N N H 2 JAU N N N N N H H 3 HEP N N N N H H H 4 HEP N N N H H H H 5 HEP N N H H H H H 6 HEP N N N N N H N 7 HEP N N N N H N N 8 MYI N N N H N N N 9 OST N N H N N N N 10 Normal N L N N N N N 11 Normal N H N N N N N 12 HYPOGL L N N N N N N 13 HYPOGL L L N N N N N 14 HYPOGL L H N N N N N 15 HYPERGL H N N N N N N 16 HYPERGL H L N N N N N 17 HYPERGL H H N N N N N 18 HA N L N N N N H 19 JAU N L N N N H H 20 HEP N L N N H H H 21 HEP N L N H H H H 22 HEP N L H H H H H 23 HEP N L N N N H N 24 HEP N L N N H N N 25 MYI N L N H N N N 26 OST N L H N N N N 27 HA N H N N N N H 28 JAU N H N N N H H 29 HEP N H N N H H H 30 HEP N H N H H H H 31 HEP N H H H H H H 32 HEP N H N N N H N 33 HEP N H N N H N N 34 MYI N H N H N N N 35 OST N H H N N N N 36 HA&HYPOGL L L N N N N H 37 JAU&HYPOGL L L N N N H H 38 HEP&HYPOGL L L N N H H H 39 HEP&HYPOGL L L N H H H H 40 HEP&HYPOGL L L H H H H H 41 HEP&HYPOGL L L N N N H N 42 HEP&HYPOGL L L N N H N N 43 MYI&HYPOGL L L N H N N N 44 OST&HYPOGL L L H N N N N 45 HA&HYPOGL L N N N N N H 46 JAU&HYPOGL L N N N N H H 47 HEP&HYPOGL L N N N H H H 48 HEP&HYPOGL L N N H H H H 49 HEP&HYPOGL L N H H H H H 50 HEP&HYPOGL L N N N N H N 51 HEP&HYPOGL L N N N H N N 52 MYI&HYPOGL L N N H N N N 53 OST&HYPOGL L N H N N N N 54 HA&HYPOGL L H N N N N H 55 JAU&HYPOGL L H N N N H H 56 HEP&HYPOGL L H N N H H H 57 Design a Fuzzy Expert System for Liver and Pancreas Diseases Diagnosis 141 HEP&HYPOGL L H N H H H H 58 HEP&HYPOGL L H H H H H H 59 HEP&HYPOGL L H N N N H N 60 HEP&HYPOGL L H N N H N N 61 MYI&HYPOGL L H N H N N N 62 OST&HYPOGL L H H N N N N 63 HA&HYPERGL H L N N N N H 64 JAU&HYPERGL H L N N N H H 65 HEP&HYPERGL H L N N H H H 66 HEP&HYPERGL H L N H H H H 67 HEP&HYPERGL H L H H H H H 68 HEP&HYPERGL H L N N N H N 69 HEP&HYPERGL H L N N H N N 70 MYI&HYPERGL H L N H N N N 71 OST&HYPERGL H L H N N N N 72 HA&HYPERGL H N N N N N H 73 JAU&HYPERGL H N N N N H H 74 HEP&HYPERGL H N N N H H H 75 HEP&HYPERGL H N N H H H H 76 HEP&HYPERGL H N H H H H H 77 HEP&HYPERGL H N N N N H N 78 HEP&HYPERGL H N N N H N N 79 MYI&HYPERGL H N N H N N N 80 OST&HYPERGL H N H N N N N 81 HA&HYPERGL H H N N N N H 82 JAU&HYPERGL H H N N N H H 83 HEP&HYPERGL H H N N H H H 84 HEP&HYPERGL H H N H H H H 85 HEP&HYPERGL H H H H H H H 86 HEP&HYPERGL H H N N N H N 87 HEP&HYPERGL H H N N H N N 88 MYI&HYPERGL H H N H N N N 89 OST&HYPERGL H H H N N N N 90 
work_mptuiuoxkvdybb4jzfcviv56je	----	NASA Technical Reports Server (NTRS) 
work_mqycojfotfaktdybajm2kyww34	----	PII: S0957-4174(94)E0001-B Pergamon Expert Systems With Applications, Vol. 8, No. 1, pp. 89-99, 1995 Copyright © 1994 Elsevier Science Ltd Printed in the USA. All rights reserved 0957-4174/95 $9.50 + .00 0957-4174(94)E0001-B An Expert System for Homeopathic Glaucoma Treatment (SEHO) F. ALONSO-AMO, A. G r M E Z PI~REZ, G . L O P E Z G O M E Z , A N D C. M O N T E S Facultad de lnform~itica--Universidad Politrcnica de Madrid, Campus de Montegancedo--Boadilla del Monte, Madrid, Spain Abstract--In this article, an Expert S y s t e m f o r Homeopathic Glaucoma Treatment ( S E H O ) is pre- sented, the task o f which is to assist ophthalmologists in selecting the most appropriate therapy f o r a patient diagnosed as having glaucoma. It is based on techniques proper to homeopathic medicine, a trend that is gaining more and more supporters all over the world, but in which real experts are f e w and f a r between. After a brief overview o f the state o f the art, the authors describe in detail on the development o f the system, f o r which the I D E A L methodology, designed f o r knowledge-based system development, was used. 1. I N T R O D U C T I O N LAMENTABLY, GLAUCOMA is o n e o f the most signifi- cant pathologies affecting sight, the n u m b e r o f cases making it one o f the m a j o r visual disorders. This disease often leads to irreversible blindness and is accompanied by an inherited t r a u m a t i c process, m a k i n g g l a u c o m a sufferers especially sensitive to the m e t h o d s used in their t r e a t m e n t . A m a r k e d partiality toward the field o f alternative m e d i c i n e has been observed within the blind com- m u n i t y , o n some occasions as a c o m p l e m e n t to tra- ditional m e d i c i n e a n d o n others due to their openness to a n d acceptance o f these branches o f medicine, o f which h o m e o p a t h y is a clear example, which they con- sider to be less aggressive. H o m e o p a t h i c m e d i c i n e is based o n the activation o f the organism's healing mechanisms by administering h o m e o p a t h i c dilutions c o r r e s p o n d i n g to specific doses o f these medicines. It should n o t be forgotten that a blind person faces o t h e r p r o b l e m s besides sight i m p a i r m e n t , such as dif- ficulties in social integration, adaptation, and mobility, etc., as well as other underlying illnesses which in m a n y cases originally i n d u c e d the process leading to blind- This paper was sponsored and backed by the ONCE (Spanish National Organization of the Blind) and the CICYT (Interministerial Com- mission of Science and Technology) grant no. TIC1235/12-E and prepared and written with the collaboration of CETTICO (Centre o f Technology Transfer in Knowledge Engineering, Madrid, Spain). Requests for reprints should be sent to F. Alonso-Amo, Facultad de Inform~tica--Universidad Politrcnica de Madrid, Campus de Montegancedo--Boadilla del Monte, 28660 Madrid, Spain. 89 ness. As a result, these people can hardly be treated medically w i t h o u t their physical an d m en t al charac- teristics being t ak en into account. H o m e o p a t h y is in fact o ri en t ed in this direction, and, as far as visual disorders are concerned, h o m e o - pathic t r e a t m e n t depends o n four factors (Rubio, 1988): the threat to vision, the k i n d o f disorder a n d m a n n e r in which it evolves, the level o f individual sensitivity, a n d the subject's potential for response. A n d such t r e a t m e n t is c o m p o s e d o f • Rem ed i es prescribed for local s y m p t o m s , an d • Rem ed i es for the g l a u c o m a t o u s field (patient's phys- ical a n d m e n t a l constitution). T h e secretion o f a q u e o u s h u m o u r (an i m p o r t a n t factor in glaucoma) is closely linked with the sensitivity o f the neurovegetative system a n d that the stress o f ev ery d ay life a n d e m o t i o n s affect pressure in the eye- ball. Psychological tests have revealed t h at g l au co m a sufferers are anxious, susceptible, a n d meticulous a n d that the g l a u c o m a t o u s subject's hypersensitivity affects the t r e a t m e n t o f the patient. H o m e o p a t h i c t r e a t m e n t is very often associated with traditional therapies, often permitting a r e d u c t i o n in the dosage a n d an increase in the patient's tolerance to the latter. So, we can say t h at the different therapies, far f r o m being opposed to each other, are co m p l em en t ary , or as Blamentier states (Rubio, 1988), " i t is the patients an d n o t the therapies t h at differ." It is i m p o r t a n t to observe h o w sensitive the patient is to a t h e r a p y that stimulates his defence potential. However, it is n o t easy to appreciate this potential, as it d ep en d s o n m a n y factors, such as age, the degree o f sensitivity o f the lesions, a n d the patient's character- istics an d constitutional data. These m a k e it possible 90 F. A l o n s o - A m o et al. for the homeopathic doctrine and clinical expertise to decide what is the most suitable treatment. However, it is not possible to establish a general rule to cover the wide range of possibilities that crop up in the homeopathic doctor's surgery and, at all events, the only suitable guide he or she has personal experi- ence. Therefore, it was thought necessary to employ knowledge engineering techniques to deal with this problem. This led to the development of the expert system described, which was developed at CETTICO (Centre of Technology Transfer in Knowledge Engi- neering) and is able to assign homeopathic treatment to glaucoma sufferers. There are no similar expert systems on record. Vi- sual disorders have been little dealt with in the past, and the therapy model has always been based on al- lopathic medicine. The CASNET system (Kulikowski & Weiss, 1982), oriented to the diagnosis and treatment of different kinds of glaucoma, is an example of this. In addition, there is a knowledge gap in homeopathic glaucoma treatment, due both to the lack of experts in this kind of therapy and to the fact that they are con- centrated in half a dozen countries. Therefore, the system entails a qualitative advance in the treatment of the blind, bringing innovative tech- niques into an alternative approach to medicine that is held in high esteem by the blind community. 2. C O M P U T E R SYSTEMS IN H O M E O P A T H I C MEDICINE Until recently, the computer models used to simulate medical decision making in a computer system have been mainly based on probabilistic methods, as their machine representation is easy to obtain (Gorry, Sil- verman, & Pauker, 1978; Solomon & Papert, 1976). Despite the fact that these programs have come up with very interesting results, the doctor does not iden- tify his or her reasoning and manner of arriving at a diagnosis with theirs, and it is also difficult to evaluate the quality of a diagnosis proposed in this way. In addition, it has been noted (Ledley & Lusted, 1979) that the majority of clinical errors are made by omission, that is, errors due to a failure to take into account all of the possibilities playing an important role in determining the illness suffered by the patient so as to arrive at the correct diagnosis and treatment. Therefore, a doctor needs assistance in establishing the diagnosis and a suitable therapy, especially in the case of unusual illnesses or when the patient's symptoms may lead to different interpretations. Considering that all of the information required on a patient can be stored and classified in a computer, along with the symptoms of the illnesses of a domain, it follows that, in such circumstances, a computer may come up with a more precise and rapid response than a doctor (Barr & Feigenbaum, 1982), especially when the knowledge of the symptoms of an illness has been elicited from an expert doctor and incorporated into a knowledge base that interacts with an expert system. A complete and adequate homeopathic study of a patient depends on the skill of the homeopath and his or her ability in identifying, storing, recording, refer- encing, analysing and evaluating any class of data or group of data. This requires a system of classification, according to which the concepts and relevant infor- mation are organized, and a coding, which facilitates their use. Originally, the homeopath's traditional prescription and, later, data bases, which were and are of great help in homeopathic surgeries, were used for classification and coding. However, it has been noted on several oc- casions that doctors are generally somewhat reluctant to use computers as a tool and consider them to be little suited to establishing repertories, that is, what medicines cover the patient's symptoms, the number of symptoms, and to what extent. On the other hand, medical reasoning is related to judgement problems, problem solving, decision mak- ing, and knowledge (Fieschi, 1987), which is why it has come to be a traditional working domain in knowledge engineering. The introduction of ES into medical diagnosis and treatment has done away with initial scepticism, and they have come into more widespread use. Examples of this are MYCIN (Shortliffe, 1976), TEIREISIAS (Davis, 1976), INTERNIST (Pople, 1977), PIP (Pauker & Szolovitz, 1977), DIGITALIS THERAPY ADVI- SOR (Gorry et al., 1978), CENTAURI (Aikins, 1980), SAM (Gascuel, 1981), ATTENDING (Miller, 1988), CASNET (Kulikowski & Weiss, 1982), NESTOR (Cooper, 1984), KARDIO (Bratko, 1989). Unfortunately, however, few ES have gained access to homeopathic surgeries, though several computer systems based on this alternative medicine have been developed, such as the following. • RADAR (Shrogens, 1982), which contains several pharmacopoeias, including those by Allen, Hering, Heneman, and Boerick. It has access to 2,000 ho- meopathic remedies and their corresponding phar- macopoeias. • HINEIRO (Bachelerie, 1986) contains 2,535 Boen- ninghausen therapy rubrics (symptoms). Each of these rubrics is associated with a blackboard with the most common remedies. • ABIES (Benson, 1980) is a clinical information sys- tem. In addition to carrying out medical treatment, it locates notes and treatments for patients using the RCC system (Real Clinical Classification) (Read & Benson, 1986), which is a hierarchical statistical classification of a nomenclature with four detail levels. • STAPHISE (SalaiJn & Simonet, 1989), an infor- mation system using the Ken repertory, composed Homeopathic Glaucoma Treatment 91 of some 20,000 rubrics taken from different phar- macopoeias. Its information base may be custom- ized. As regards ES in homeopathic medicine, we should mention VES (Vithoulkas, 1988), developed by the homeopath George Vithoulkas. His working philoso- phy can be situated within the unitarian current of homeopathy. The VES system returns the best remedy with a given scoring and certainty factor, indicating to what extent any alternative remedies presented can be administered. VES is integrated into the RADAR sys- tem and takes advantage of its potential. It is a general medical application and is not specialized in any class of illness. 3. SEHO (ES FOR HOMEOPATHIC GLAUCOMA TREATMENT) Considering homeopathy as complementary to tradi- tional treatments of visual disorders causing blindness and taking into account its acceptance among the blind community, it was thought necessary to research and develop an ES to treat glaucoma. This would assist the homeopath in inference tasks and, finding the medi- cines most suited to each patient, would come up with the appropriate dilutions. The end result was the pro- totype system SEHO (Cristrbal & Ortiz Latierro, 1991). As opposed to other systems, NEOMYCIN for ex- ample, which mainly use the description of the patient's illness to select the therapy (Clancey, 1981; Clancey & Shortliffe, 1984), SEHO compiles extensive and com- plete information on the patient through homeopathic questioning before suggesting any remedy and thus does not point the doctor in any particular direction that might lead him or her to overlook important data on the patient during the session. The results of the questioning session are sent to the homeopathic techniques of the expert, which attempts to put together the most complete profile of the patient possible and to establish his or her characteristics and symptoms so as to apply the most suitable medicines. Optionally, the system can provide information on other possible, though less suited, medicines together with the patient's symptoms that each of them covers. Like other ES, SEHO provides information on its reasoning process, explaining the intermediate conclu- sions and why it selects and incorporates each medicine into the working memory. There are two different trends in homeopathic med- icine: the unitarian trend, which suggests only one medicine as a remedy (e.g., VES) and the pluralist trend, to which SEHO belongs, which proposes differ- ent dilutions of several medicines. So, SEHO's dilutions contain several medicines. SEHO is, therefore, the first ES for treating a visual disorder from the point of view of alternative medicine. Another important characteristic differentiating it from other systems is the fact that it has been designed to be used by the blind, available Braille adaptations hav- ing been incorporated into the prototype system. Its line of reasoning, based on the techniques of the chosen expert, the Chairman of the Association of Homeo- paths of Madrid, leads to scaled dilutions, that is, to the assignation of three or more medicines in most cases, some with low, some with intermediate, and others with high dilutions, except when any of these are superfluous. The inference process passes through several stages before arriving at these recommendations. The first stage, medicine determination, covers all of the pa- tient's symptoms, both those related with specific symptoms and their field characteristics. It selects the lowest possible number of the most suited medicines. Provisional dilutions are assigned in the second stage, and the scaled dilutions are established in the final one. SEHO was developed using GURU, and its knowl- edge base is composed of seven rule bases, which con- tain a total of 72 rules based on public and expert knowledge on glaucoma, and five data bases, contain- ing medicines and symptoms, categories and field. Its inference mechanism is backward chaining. The prototype has been designed in such a way that its knowledge base and data bases, containing phar- macopoeias including the latest findings with respect to homeopathic remedies for glaucoma, can be ex- panded. 4. SPECIFICATION AND DEFINITION OF SEHO The IDEAL methodology (Mat6 & Pazos, 1988), one of the most prestigious methodologies for ES devel- opment today, was used to define and develop the pro- totype system. This methodology brings together the most relevant ones in this area. Some of these meth- odologies are (Hayes-Roth, Waterman, & Lenat, 1983; Liebowitz & De Salvo, 1989; Waterman, 1986). Requirements definition was based on the following general concepts: • Interface faithfully reflecting the content and extent of homeopathic questioning on the patient's char- acteristics and constitutional data and his or her symptoms, and adapted for the blind. • Storage of complete pharmacopoeias on each med- icine. • Assignation of scaled dilutions of the remedies in the final treatment: low (4 CH), intermediate (7 CH), and high (15 CH) dilutions. • Assignation of the smallest possible number of nec- essary medicines. The adequacy test was carried out to evaluate the application, and the characteristics were grouped in four dimensions according to the IDEAL methodology: plausibility, justification, success, and adequacy, using 92 F. Alonso-Amo et aL Dimension TABLE 1 Evaluation of the Application Dimension Value (geometric mean) (Vci) M a x i m u m Dimension Value (geometric mean) (Vcmi) Plausibility 62 89.1 Justification 29.8 64.1 A d e q u a c y 48.4 61.9 Success 46.8 66.5 t = 4 Vg (General value) = ~ Vci/4 = 46.75. I=1 1=4 Vm (Maximum value) = ~ Vcmi/4 = 70.4. i=1 Vf (Final application value) = V ~ x 10 = 6.64 > 5. variable threshold values to accept or reject a charac- teristic a n d the geometric m e a n (Mat6, 1988; Pazos, 1989) to evaluate the task. T h e application was f o u n d to be suitable for t r e a t m e n t using an ES (threshold val- ues o f 6.64 > 5) (Table 1). S E H O is a decision s u p p o r t system for assigning h o m e o p a t h i c resources or remedies, which was con- sidered central for its definition. T h e r e is an essential difference in the m e t h o d used for assigning the appro- priate treatment: it does n o t directly take a c c o u n t o f the patient's pathology or type o f glaucoma, as allo- pathic medicine would; the disease's manifestation is truly decisive, that is, the patient's s y m p t o m s a n d con- dition are determined by his or her field characteristics. S E H O exclusively considers the s y m p t o m s typical o f glaucoma and the psychological profile and personal factors o f the subjects suffering from glaucoma. Surgical affections are excluded f r o m this framework, t h o u g h c o m p l e m e n t a r y t r e a t m e n t or t h e r a p y to i m p r o v e tol- erance to allopathic medicines m a y be assigned. Figure 1 shows the flow chart o f the S E H O prototype. T h e r e are two types o f knowledge i m p l e m e n t e d in the system: • Public knowledge, based o n h o m e o p a t h i c p h a r m a - copoeias (Barraza, 1980; L a t h o u d , 1988) in 5 data- bases, including lists o f medicines along with specific symptoms, the categories a n d the field, as well as a list o f antagonistic categories. • Expert knowledge, c o n t a i n e d in 7 p r o d u c t i o n rule bases, elicited f r o m the expert a n d based o n the op- erative a n d heuristic p r o c e d u r e used by the expert himself to prepare treatment. T h e rules will be fired using a backward chaining control strategy, until all the premises in its a n t e c e d e n t are true or false. • T h e response t i m e has to be short (less t h a n 5 min- utes). • During execution, the system offers i n f o r m a t i o n o n the medicines that are being considered to cover the s y m p t o m s , category, a n d field, as well as c o m m e n t s o n how it arrives at the scale o f dilutions. • Once executed, the system offers i n f o r m a t i o n on medicines in the final solution, the dilution in which the m ed i ci n e should be administered, dosage, length o f the t r e a t m e n t until the n ex t visit, s y m p t o m s a n d characteristics that the selected medicines cover, a n d medicines selected t h at are n o t included in the s o - l u t i o n along with a list o f the s y m p t o m s t h at they cover. An ideal solution is o n e that assigns a scaled dilution o r a single m ed i ci n e per dilution. This ideal situation does n o t necessarily have t o o c c u r in every case: this depends o n the patient a n d his o r h er characteristics. 5. C O N C E P T U A L I Z A T I O N A N D F O R M A L I Z A T I O N O F S E H O Knowledge acquisition for the conceptualization a n d subsequent formalization o f the knowledge base was based o n n o n s t r u c t u r e d interviews in the first stage. Cases were presented to the expert a n d the p ro t o co l was analysed d u ri n g this process in the second stage. In the final stage, very explicit structured interviews were a m e a n s o f solving the problems that arose. It was f o u n d t h at a h o m e o p a t h views a patient f r o m three different b u t c o m p l e m e n t a r y angles, w h en estab- lishing a therapy: • Specific o r local sy m p t o m s, related with the organ o r organs affected b y the illness (eyes, in the case o f glaucoma). S y m p t o m s such as congestive p h e n o m - ena, increase in ocular pressure, sight i m p a i r m e n t , alteration o f the optical nerve, etc. • Categories, as a m e a n s o f classifying the sy m p t o m s. T h e y specify the i m p r o v e m e n t s o r deterioration o f a s y m p t o m o r o f the patient in general. • Field, which establishes h o w the patient reacts to the illness an d which is characterized b y the patient's characteristics: physical, mental, an d constitutional data, such as anxious, susceptible, meticulous, for example. T h e result was the conceptual m o d el shown in Fig- ure 2 a n d the knowledge m a p shown in Figure 3. P at ie n t M ed ic in e Q u es ti o n in g o n P at ie n t S y m p to m s Q u es ti o n in g o n P at ie n t C at eg o ry Q u es ti o n in g o n P at ie n t F ie ld C at eg o ry L o o k f o r S u it ab le M ed ic in es P ro d u ce S o lu ti o n M ed ic in es S o lu ti o n F IG U R E 1 . S E H O f lo w c h ar t. M ed ic in e B as e D el et e A ss ig n P ro v is io n al D il u ti o n s M ed ic in es a n d P ro v is io n al D il u ti o n s A ss ig n S ca le d D il u ti o n s J 94 F. Alonso-Amo et al. PATIENT 1 1 1 I I sY T° S I I ATIENTI I I MEDICINE , , SYMPTOMS SYMPTOMS I I CATEGORY n ~ I PAT1ENT CATEGORY I I MEDICINE CATEGORY I FIELD I~ n I I n I PATIENT MEDICINES I FIGURE 2. Conceptual model. Knowledge formalization required that a distinction be m a d e between two stages in the expert process: medicine d e t e r m i n a t i o n a n d dilution assignment. 5.1. Selection of Medicine Before medicines are selected, the patient's s y m p t o m s are elicited t h r o u g h questioning o n individual symp- toms, establishing their i m p o r t a n c e or otherwise, a n d o n the patient's categories a n d fields. Considering t h e x , y , z SI(x) nSI(x) n M O D ( x ) n M O S A N T A ( x ) n SNI(x) n T O T S I N T ( x ) ST(x) A B following definitions: Medicines Set o f i m p o r t a n t s y m p t o m s cov- ered b y m ed i ci n e x N u m b e r o f i m p o r t a n t s y m p t o m s co v ered b y medicine x N u m b e r o f categories covered b y m ed i ci n e x N u m b e r o f antagonistic categories o f m ed i ci n e x N u m b e r o f u n i m p o r t a n t specific s y m p t o m s co v ered by medicine x To t al n u m b e r o f s y m p t o m s cov- ered b y x (specific + category + field) Set o f field s y m p t o m s covered by x Set o f medicines selected for spe- cific sy m p t o m s. It changes as rules are fired a n d some medicines are selected an d others excluded. At the beginning o f the process, the set covers an y i m p o r t a n t sy m p t o m . Set o f medicines selected for the field. Like A, it changes w h en rules are fired. It initially covers a n y o f the patient's characteristics or sy m p t o m s. I Patient Name A Name Category Name A Pharmacopoeial Category 1 Med. Cat. Name Category Name Medicine Name Field Name Field Present in Pharmacopoeia Name, Symptom Importance Name. Important Name. Unimportant Symptom Name " ' " ~ 1 Name Symptom Name Medicine Symptom Name Medicine Name I Medicine N. Dilution / ] I Med. F. Name I " j ~v~e~c?aernI~la me I FIGURE 3. Knowledge map. Homeopathic Glaucoma Treatment 95 • T h e p r o c e d u r e followed b y the e x p e r t for selecting the m e d i c i n e to c o v e r given s y m p t o m s is based o n the following rules: I f x, y ~ A a n d SI(x) C SI (y), t h e n select y I f x, y ~ A a n d SI(x) = SI (y) a n d n M O D A N T A ( x ) > n M O D A N T A ( y ) , t h e n select y I f x, y E A a n d SI(x) = SI (y) a n d n M O D ( x ) > n M O D ( y ) , t h e n select y I f x, y ~ A a n d SI(x) = SI(y) a n d nSNI(x) > nSNI(y), t h e n select x I f x, y ~ A a n d SI(x) = SI (y) a n d nST(x) > nST(y), t h e n select x. T h e o b t e n t i o n o f the m i n i m u m set in A is based o n the rules below: I f x, y E A a n d [x,y] covers the s y m p t o m s o f Ix, y, z], t h e n A = A - [z]. I f there is m o r e t h a n o n e m i n i m u m set, select the o n e whose m e d i c i n e s c o v e r m o r e , i m p o r t a n t s y m p - t o m s . • T h e selection o f m e d i c i n e s for the field is centered o n the following rules: I f x, y E B a n d ST(x) C ST(y), t h e n select y I f x, y @ B a n d ST(x) = S T (y) a n d n T O T S I N T ( x ) > n T O T S I N T ( y ) , t h e n select x. T h e o b t e n t i o n o f the m i n i m u m set in B is b a s e d o n the s a m e rules as for the o b t e n t i o n o f the m i n i m u m A, t h a t is: I f x, y, z E B a n d [x, y] covers the s y m p t o m s o f [x, y, z], t h e n B = B - [z]. Select the m i n i m u m set t h a t covers m o r e , i m p o r t a n t s y m p t o m s . 5 . 2 . A s s i g n a t i o n o f D i l u t i o n s T h e goal p u r s u e d in this stage is the a s s i g n m e n t o f scaled dilutions. First a p r o v i s i o n a l a s s i g n m e n t o f the dilutions o f the m e d i c i n e s o b t a i n e d is taken, a n d the case-related o p t i m u m is sought. If: M H Set o f high-dilution m e d i c i n e s M L n H nL ML1 M L 2 nLi M H 1 M H i nHi Set o f low-dilution m e d i c i n e s N u m b e r o f m e d i c i n e s in M H N u m b e r o f m e d i c i n e s in M L Set o f m e d i c i n e s in M L t h a t c o v e r categories Set o f M L I m e d i c i n e s t h a t c o v e r m o s t cate- gories N u m b e r o f M L i m e d i c i n e s with i = l, 2. Set o f M H medicines that c o v e r field s y m p t o m s Set o f M H l m e d i c i n e s t h a t c o v e r m o s t s y m p - t o m s (i -- 2 . • . 9 ) N u m b e r o f M H i m e d i c i n e s with i = l • • • 9. T h e provisional dilutions are assigned according to the following rules: I f x E A a n d x E B, assign a low dilution to x a n d x @ M L I f x E A, assign a low dilution to x a n d x E M H I f A :~ 0 a n d B 4 = 0, t h e n provisional dilutions -- low I f B :~ 0 a n d M L v ~ 0, t h e n provisional dilutions = low a n d high I f A = B a n d B ~ 0 a n d A ~ 0, t h e n provisional dilutions = high T h e different p a t h s t a k e n t o arrive at the goal (the suit- able dilution), once the provisional dilutions have b e e n obtained, are represented in the shape o f a tree in Fig- ures 4, 5, a n d 6. T h e final t r e a t m e n t is b a s e d o n the following rules: I f x is assigned a low dilution, prescribe 3 doses o f a 4 C H dilution o f x p e r day. I f x is assigned a n i n t e r m e d i a t e dilution, prescribe 4 doses o f a 7 C H dilution o f x p e r day. I f x is assigned a high dilution, prescribe 5 doses o f a 15CH dilution o f x p e r day. 6. S E H O S Y S T E M I M P L E M E N T A T I O N T h e system p r o t o t y p e has b e e n i m p l e m e n t e d using G U R U as a d e v e l o p m e n t tool. In addition, a n interface, i n c o r p o r a t i n g a voice synthesizer a n d Braille line, has b e e n designed t o e n a b l e the blind to use the system. PROVISIONAL DILUTIONS = LOW [ nL = 1 nL>l / nL1 >_ 1 [ / nL> 1 I R14 Intermed, Dil, ] nL2>l J RI5 ] [Intermed. Dil. I x . , I nL1 = 1 ] R16 R17 FIGURE 4. Dilution assignation I. i INoc"an e'iodi' I bchaogesindi' I 96 F. Alonso-Amo et aL PROV. DILUTIONS = HIGH I FIGURE 5. Dilution assignation II. Its design and implementation matches the Struc- tures Map in Figure 7, in which the modules describe the following actions: MSINS Generates the interface for recording patients' symptoms MMODS Generates the category interfaces MTERRS Generates the field interfaces MLEESIN Searches the data base for medicines that cover any patient symptom MLEEMOD Searches the data base for medicines that cover any patient category MLEETERR ORDENA MINIMOSA ORDENB MINIMOSB MBUSDILU MDILBYA Searches the data base for medicines that cover any patient field Obtains final medicines for particular symptoms Obtains minimum sets of medicines for symptoms Obtains final medicines for field Obtains minimum sets of medicines for field Defines provisional dilution Defines suitable dilution I PROV. DILUTIONS = HIGH & LOW I / FIGURE 6. Dilution assignation III. "N l M A IN M O D U L E se h o D A T A C O L L E C T IO N M/ II M M~ O~ i I M TER RS C H O IC E O F S O L U T IO N M E D IC IN E S S E A R C H II F IG U R E 7 . S tr u ct u re s m a p . 98 M R E S U L T Assigns the most appropriate treat- m e n t , including medicines, their di- lutions, a n d the way o f administering them. 7. E V A L U A T I O N S E H O was evaluated in three phases in line with tra- ditional methodological orientations. 7.1. Validation of System Decisions by the E x p e r t Fifteen test cases that had been selected by the expert a n d set o u t in the project success criteria, along with a n o t h e r 20 o f the m o s t frequent cases p u t forward by the expert were used for this purpose. O f the examples selected, 10% were e x t r e m e cases a n d generated arti- ficially, 10% were ambiguous, a n d the remaining were typical cases. T h e expert a p p r o v e d the system's pro- cedure in all o f the cases. As regards dilution assignation, the different mod- ules were verified as follows. TABLE 2 Typical Case, C a s e No. 6 SS (Specific Symptoms) Nebulas Photophobia Orbital pain Reduced vision Dilated pupils Stigmata on the cornea Categories Worsens when lying down Worsens in the morning Worsens with changes in the weather Improves in the open air Improves with warmth FT (field symptoms or characteristics) Depression Obesity Apathy Pessimism Shyness Skin irritations Fatigue Constipation Egotism Recommended medicines FLUORIC CALCAREA--Iow dilution SULPHUR--intermediate dilution CARBONIC CALCAREA--high dilution CAUSTICUM--high dilution Optimal treatment with GELSEMIUM--only high NUX VOMICA--only high AURUM METALICUM~only high COMOCLADIA~only high Important (I) Unimportant (U) I I I I F. A l o n s o - A m o et al. A) Step f r o m provisionally low dilutions to low an d i n t e r m e d i a t e dilutions B) Step f r o m provisionally high dilutions to low a n d high dilutions C) Step f r o m provisionally low an d high dilutions t o low, intermediate, a n d high dilutions. It was f o u n d that, as required b y the expert, the dilu- tions are n o t fully scaled in two situations: • W h e n there are n o field characteristics and, therefore, low dilutions are n o t assigned; • It is impossible to assign intermediate dilutions o n the basis o f high a n d low dilutions, as the set o f med- icines c o u n t e d in this case is equal to the total. This case is considered extreme. Finally, it was f o u n d that the n u m b e r o f medicines assigned with a given dilution is n ev er greater t h a n 2, j u st as the expert stipulated. A typical case is illustrated in Table 2, indicating the f o r m t h ey take. 7.2. Validation of Typical Cases by Experts Not Involved With System Development Fifteen typical cases were p u t to t h em , an d they o n l y disagreed o n o n e a m b i g u o u s case. This was d u e to the fact t h at there were two equally acceptable forms o f t r e a t m e n t , a n d this was therefore a question o f pref- erences. Moreover, the system had indicated the second possibility as an optional treatment. T h e system has n o w b een transferred t o the Spanish N at i o n al Organization for the Blind (ONCE) as an aid for therapists n o t specialized in h o m eo p at h i c medicine. 8. FUTURE RESEARCH WORK Although the results provided b y S E H O are an im- p o r t a n t ad v an ce in the a u t o m a t e d t r e a t m e n t o f glau- c o m a using h o m e o p a t h i c techniques, we should n o t o v e r l o o k the fact t h at S E H O is a p r o t o t y p e requiring f u r t h e r d e v e l o p m e n t . Th i s will involve two courses o f action: o n e regard- ing the system's knowledge a n d the other, its c o m p u t e r structure. As regards the knowledge at present incor- p o r a t e d i n t o SEHO, it is p l a n n e d to ex t en d the medi- cine an d expert knowledge data bases. F o r this purpose, a n o t h e r expert in h o m e o p a t h i c medicine, likewise a m e m b e r o f the pluralist school, will j o i n the research t eam , with a view to adding to the knowledge a n d co m p ari n g his approach with that o f the f o r m e r expert. With respect t o the S E H O ' s c o m p u t e r structure, t h ree basic measures are envisaged. • T h e new knowledge acquisition stage will m a k e it possible to identify new rules an d new fo rm s o f pro- cessing an d handling the m o st useful aspects o f the individual cases. F o r this stage, it is p l a n n e d to in- c o r p o r a t e an a u t o m a t e d knowledge acquisition m o d u l e that will equip the system with the capability H o m e o p a t h i c G l a u c o m a T r e a t m e n t 99 of analyzing any new information that comes to light, as well as facilitating the knowledge acquisition task. There are also plans to equip the final system with an intelligent interface capable o f selecting the ques- tions to be put to the expert. • Adapt the user interface to a more flexible graphic environment for use by sighted personnel, while the present interface will be kept for the blind user. For this purpose, the screens o f yes/no questions (62 on symptoms, 57 on category, and 66 on field charac- teristics) will be replaced by 4 multiple choice win- dows with a scroll bar (for symptoms, category, and field, and to select the important symptoms from those chosen). The information will be output to an independent window controlled by the application, and new phases and explanations will be incorporated to make it easier to follow the system's logic. • With a view to creating an easily extendible and reusable expert system, in place of the present GURU-based configuration that is difficult to alter, the SEHO prototype will be adapted to an object- oriented environment by transforming the present medicine and knowledge bases and rules into a class structure. The CLASER (Set o f Open Libraries for the Development and Query of Reusable Expert Systems) (Garcia, 1993), recently developed at CET- TICO in C + + , will be employed to this end. From experience with other systems, this measure should make the future SEHO system approximately 10 times faster than the present system, in addition to the above-mentioned advantages o f extendibility and reusability. REFERENCES Aikins, J.S. (1980). Prototypes and Production Rules: A Knowledge Representation for Computer Consultations. Stanford Heuristic Project, Department of Computer Science Report No. STAN- C5-80-84. Bachelerie, R. (1986). HOMEOREP. Clermont-Fen'and. Barr, A., & Feigenbaum, E.A. (1982). The handbook of artificial in- telligence. Reading, MA: Addison Wesley. Barraza, J. (1980). Elementos de terapedaica homeopdtica. Buenos Aires: Editorial Albatros. Benson, T. (1986). Developing Tools for On Line Medical Reference Works. In R. Solomon, B. Blum and M. Jorgensen (Eds.), "MED1NFO 86." Amsterdam: Elsevier North Holland. Bratko, T. (1989). KARDIO: A study in deep and qualitative knowl- edge for expert systems. Cambridge, MA: The MIT Press. Clancey, W.J. (1981). NEOMYCIN: Reconfiguring a Rule-Based Expert System for Application to Teaching. In Proceedings of the Seventh IJCAI (pp. 829-836). Clancey, W.J., & Shortliffe, E. (Eds.). (1984). Readings in medical artificial intelligence. The first decade. Reading, MA: Addison- Wesley. Cooper, G. (1984). NESTOR, A Medical Decision Support System that Integrates Causal Temporal and Probabilistic Knowledge. PhD thesis, Medical Information Sciences, Stanford University, La Jolla, CA. Cristrbal, M,R., & Ortiz Latierro, N. (1991). Sistema Experto para el Tratamiento Homeopdtico del Glaucoma. MSc thesis, Facultad de lnformfitica--Universidad Politrcnica de Madrid, Madrid. Davis, R. (1976). Applications of metalevel knowledge to the con- struction, maintenance and use of large knowledge bases (AIM- 283). Computer Science Department, Stanford University. Fieschi, M. (1987). Inteligencia artificial en medicina. Barcelona: Masson S.A. Garcia (1993). C L A S E R Conjunto de Librer{as Abiertas para el De- sarrollo y Consulta de SS.EE. reutilizables. Final-year project. Facultad de lnformfitica, Madrid, Universidad Politrcnica de Madrid, Spain. Gascuel, O. (1981). SAM: Un systrme expert dans le domaine med- ical. In Proceedings of Congr~s AFCET RF-IA, Nancy, France. Amsterdam: North Holland. Gorry, G.A., Silverman, H., & Pauker, S.G. (1978). Capturing clinical expertise: A computer program that considers clinical response to digitalis. American Journal of Medicine, 64, 452-460. Hayes-Roth, F., Waterman, D.A., & Lenat, D.B. (1983). Building expert systems. Reading, MA: Addison-Wesley. Kulikowski, C.A., & Weiss, S.M. (1982). Representation of expert knowledge for consultation: The CASNET and expert projects. In P. Szolovits (Ed.), Artificial Intelligence in Medicine. Boulder, CO: Westview Press. Lathoud, J. (1988). Materia M~dica Homeopdtica. Buenos Aires: Editorial Albatros. Ledley, R., & Lusted, L. (1959). Reasoning formulations of medical diagnosis. Science 130, 9-24. Liebowitz, J., & De Salvo, D.A. (1989). Structuring expert systems: Domain, design and development. Englewood Cliffs, N J: Yourdon. Matr, J.L., & Pazos, J. (1988). Ingenier{a del conocimiento. Disoqo y Construccirn de sistemas expertos. Crrdoba, Argentina: SEPA. Miller, P. (1988). Selected topics in medical artfwial intelligence. New York: Springer Verlag. Pauker, S.G., & Szolovits, P. (1977). Analyzing and simulating taking the history of the present illness context formation. In Schneider and Sagwall Hein (Eds.), Computational linguistics in medicine. Amsterdam: North Holland. Pazos, J. (1989). Metodologfa construcci6n sistemas expertos. Notes MSc in Knowledge Engineering, Facultad de Informfitica-Univ- ersidad Potitrcnica de Madrid, Madrid. Pople, H. (1977). The formation of composite hypotheses in diagnostic problem solving. An exercise in synthetic reasoning. In Proceed- ings of the Fifth 1JCAL MIT, Cambridge, MA. Read, J., & Benson, T. (1986). Computer coding. British Journal of Healthcare Computing. Rubio, C. (1988). Indicaciones y contraindicaciones de la homeopat[a en oftalmolog{a. Notes C. Rubio, Asociaci6n de Home6patas de Madrid. Salatin, P., & Simonet, M. (1989). La medicina homeopathe. Gazette of the National Union of French Homeopathic Doctors. Shortliffe, E.M. (1976). Computer-based medical consultations: MY- CIN. New York: Elsevier. Shrogens, D. (1982). Specific and technical details about RADAR. Ghent, Belgium: ARCHIMEDE, Association pour la recherche en informatique medicale. Solomon, C., & Papert, S. (1976). A case study of a young child doing turtle graphics in LOGO. In Proceedings of the American Fed- eration of lnformation Processing Societies National Computer Conference. Vithoulkas, G. (1988). Vithoulkas expert system Universit6 de Na- mun, Institut d'lnformatique, Belguim. Waterman, B. (1986). A guide to expert systems. New York: Addison- Wesley. 
work_mrdnczzbxzf43moxvih3bdskei	----	Expert System of Fault Diagnosis for Gear Box in Wind Turbine Systems Engineering Procedia 4 (2012) 189 – 195 2211-3819 © 2011 Published by Elsevier Ltd. Selection and peer-review under responsibility of Desheng Dash Wu. doi:10.1016/j.sepro.2011.11.065 Available online at www.sciencedirect.com Systems Engineering Procedia Systems Engineering Procedia 00 (2011) 000–000 www.elsevier.com/locate/procedia The 2nd Expert System of Fault Diagnosis for Gear Box in Wind Turbine International Conference on Complexity Science & Information Engineering YANG Zhi-Ling∗ North China Electric Power University, Beijing,102206, China , WANG Bin, DONG Xing-Hui, LIU Hao Abstract Gear box is one of the key components in wind turbine. To make timely and accurate diagnosis for gear box, an expert system based on fault tree analysis is developed in this paper. Firstly, a fault tree model is established in accordance with the structural features of wind turbine gearbox. On basis of the fault tree model, qualitative analysis and quantitative analysis of the gearbox faults are carried out afterwards. Finally，a Web-oriented expert system is developed by C# on the .NET platform which will save the fault diagnosis time significantly and make the expert solution for the fault of gear box to achieve the precise and quick maintenance more effectively. © 2011 Published by Elsevier Ltd. Selection and peer-review under responsibility of Desheng Dash Wu Keywords: expert system, fault diagnosis, industrial engineering, gear box, wind turbine; 1. Introduction Gear box is one of the key components in wind turbine and the fault of gear box will lead to wind turbine shut down inevitably. So what must be done is making timely and accurate diagnosis to gear box. However, due to rapid development of wind power industry, there is no fault diagnosis system specifically for wind turbine gearbox on the market in China. Currently, during the wind turbine gearbox fault diagnosis, workers generally analyze the phenomenon by rule of thumb. This fault diagnosis approach will spend more time and the accuracy is low. Statistics show that the determining time for the fault diagnosis takes up 70% to 90% of the total time, while the repair time takes up only about 10% to 30%[1]. Fault diagnosis systems have been successfully applied in many kinds of technical processes to improve operation reliability and safety. Wu introduced an expert system for fault diagnosis in internal combustion engines using wavelet packet transform and neural network[2]. Lei proposed a new multidimensional hybrid intelligent diagnosis method to identify different categories and levels of gear damage automatically[3]. Saravanan dealt with the effectiveness of wavelet-based features for fault diagnosis using support vector machines and proximal support vector machines[4]. In another paper, he introduced the use of decision tree for selecting best statistical features that discriminate the fault conditions of the gear box from the signals extracted and the decision tree was also used to generate the rules automatically from the feature set [5]. ∗ Corresponding author. E-mail address: yzhil@ncepu.edu.cn. © 2011 Published by Elsevier Ltd. Selection and peer-review under responsibility of Desheng Dash Wu. Open access under CC BY-NC-ND license. Open access under CC BY-NC-ND license. http://creativecommons.org/licenses/by-nc-nd/3.0/ http://creativecommons.org/licenses/by-nc-nd/3.0/ 190 YANG Zhi-Ling et al. / Systems Engineering Procedia 4 (2012) 189 – 195 2 YANG Zhi-Ling / Systems Engineering Procedia 00 (2011) 000–000 Fault diagnosis systems are relatively new and are flourishing on a rapid scale in wind energy technology, especially for the development of large-scale wind turbines. Barszcz presents the application of the spectral kurtosis technique for detection of a tooth crack in the planetary gear of a wind turbine[6]. Tang presents a wind turbine fault diagnosis method based on the Morlet wavelet transformation and Wigner-Ville distribution[7]. Jiang applied a new denoising method based on adaptive Morletwavelet and singular value decomposition (SVD) to feature extraction for wind turbine vibration signals[8]. Hussain proposed a novel method for real time fault detection in gearboxes by using adaptive features extraction algorithm to deal with non-stationary faulty signals[9]. This paper applies the fault diagnosis system on wind turbine gear box, in order to make timely, accurate diagnosis for gear box and provide timely solutions. In this way, it can not only meet the requirements of rapid response to wind turbine fault diagnosis, but also prevent the emergence of human error, the valuable experience of experts can be retained to prevent lost, too. Fault tree analysis is an intuitive, clear, logical, and applicable system compares to the relatively simple ones. The gear box of wind turbine has less hierarchical structure, so fault tree is used to analyze the component faults and express their reasons, and an expert system based on fault tree analysis is chosen to make diagnosis on wind turbine gear box in this paper. 2. Fault Tree of Gear Box in Wind Turbine 2.1. Fault Tree Analysis Fault tree analysis is a graphical interpretation method which can refine the system fault causes from whole to part by dendrite progressively. It is also an important method of analyzing system reliability and safety. Fault tree is drawn through analyzing a variety of factors (including hardware, software, environment, human factors, etc.) which should cause system fault, refining system fault event from whole to part by dendrite progressively, identifying the cause of system fault, determining the probability of fault, and evaluating the important degree related to the cause[10]. Fault tree model is a behavior and qualitative causal model based on studying the structure and functional characteristics of fault. The top event is the event which is to be avoided firstly. The basic events are those can cause top event while the intermediate events are the nodes. Fault tree is a dendritic logic diagram which shows relevance between events by using logic gates. As shown in Fig. 1. System fault Module A fault Module B fault C1 C2 and C3 C4 or or Fig.1 Fault tree diagram 191 YANG Zhi-Ling et al. / Systems Engineering Procedia 4 ( 2012 ) 189 – 195 YANG Zhi-Ling / Systems Engineering Procedia 00 (2011) 000–000 3 In Fig. 1, the top event is system fault caused by intermediate events of component A fault or component B fault. Component A fault is caused by basic event of part 1 fault or part 2 fault. The basic events of part 3 fault and part 4 fault happen will cause the component B fault. 2.2. Fault Tree of Gear Box in Wind Turbine From the current structural characteristics of gear box in wind turbine and its actual fault conditions, faults usually occur in parts such as gears, shafts, bearings and box. But in order to perfect the fault tree, fastener fault and oil seal fault are added to the fault tree. In the fault tree of gear box, wind turbine gear box fault is the top event; the intermediate events include gear fault, shaft fault, bearing fault, the box fault, fasteners fault, seals fault and other events which can be further refined; the damage of parts or improper operation is taken as the basic event. As fault of any component of the gear box will cause a system fault, the main logical relationship between the fault is "or". The final fault tree is shown in Figure 2(M stands for intermediate event, while C stands for basic event). Wind turbine gear box fault Seal faultFastener faultBox faultBearing faultShaft faultGear fault M1 Shaft misalignment Shaft bending Shaft loose C7 C9C8C6C5C3 C4...C2C1 C10 C10C11 ... w1 w5w4w2 w3 assembling error Shaft deformation Uneven settlement Box wear or deformation Bearing wear or deformation Load too large Improper placement Speed ​​too fast Unequal selection Heating less Fig.2 Wind turbine gear box fault tree diagram In order to get an extended fault tree, auxiliary information including the weight and solution of fault is added to each basic event. Fig.2 only shows corresponding weight of intermediate event M3. The relationship of weights is: (1) In which, ωi Finally the diagnostic knowledge base is obtained by using production rules to put the fault tree information into knowledge base [11]. According to the diagnostic knowledge base, not only the specific locations and causes of fault can be obtained, but also the reference solution can be given by experts. stands for the corresponding weight of basic event C. 192 YANG Zhi-Ling et al. / Systems Engineering Procedia 4 (2012) 189 – 195 4 YANG Zhi-Ling / Systems Engineering Procedia 00 (2011) 000–000 3. Fault Analysis Based on Fault Tree 3.1. Qualitative Analysis The task of qualitative analysis is to identify the possible causes of system top event fault, or to find out the minimal cut sets of the fault tree [12]. The method of finding the minimum cut set is mainly upstream and downstream. According to logical relationship between faults of gear box in wind turbine, this paper adopts downstream. The idea of downstream is: according to the two adjacent terms in the fault tree, logic "or" gate increases the number of cut sets and logic "and" gate increases the capacity of the cut set [13]. Starting from the top event, the cut sets of fault tree is obtained by replacing the previous events into the next level event, writing event vertically when meeting "or" gate, until the gates are all converted into the top event. Fig. 3 shows the minimal cut sets are gotten by qualitative analysis which take the gearbox fault, axis fault and axis bending as the main line. The minimal cut sets of other components fault can be obtained by the same method. Due to the structural characteristics of gear box, the resulting cut sets are all the minimal cut sets, which are {C1},{C2},{C3},{C4},{C5},{C6},{C7},{...}. Minimal cut sets are independent of each other. Gear box fault Gear fault Bearing fault Shaft fault Box fault ... Shaft misalignment Shaft bending Shaft loose Unequal selection Improper placement Heating less Speed ​​too fast Load too large . Fig.3 Qualitative analysis of gear box fault 3.2. Quantitative analysis After obtaining all minimal cut sets, if there is enough data which can support inferring probability of each basic event, the further quantitative analysis can be made [14]. The main purpose of quantitative analysis is to find the characteristic quantities of the top event (such as the top event probability and system reliability) and the importance of the basic event, which can determine the weaknesses in the system. 1) Calculate the probability of top event The probability of occurrence of top event with "and" gate can be calculated by following formula: 11 ( ) ( ) ( ) n n i i ii p x p x p x == = =∏ (2) The probability of occurrence of top event with "or" gate can be calculated by following formula: 11 ( ) ( ) 1 [1 ( )] n n i i ii p x p x p x == = = − −∏ (3) 193YANG Zhi-Ling et al. / Systems Engineering Procedia 4 (2012) 189 – 195 YANG Zhi-Ling / Systems Engineering Procedia 00 (2011) 000–000 5 In which, p(x) is probability of the fault tree top event, p(xi) is probability of the i-th minimal cut set of fault tree, n is the total number of fault tree minimal cut set. 2) Importance analysis of basic event Importance analysis is to analyze the contribution of fault, which occurred in the minimal cut sets of a component, to the probability of top event. Probability importance degree of basic event describes the extent of probability change of top event caused by the probability change of i-th basic event fault, which is expressed by the formula as follow: ni xp xp xI i i ...1,)( )( )( = ∂ ∂ = (4) Ii(x) is probability importance degree of i-th basic event, p(x) is probability of the tree top event fault, p(xi) is probability of the i-th minimal cut set of fault tree, n is the amount of minimal cut set in the fault tree. 4. Expert System of Fault Diagnosis On the basis of fault tree analysis for gear box, a Web-oriented expert system for fault diagnosis of gear box is developed. The system is designed to have three grades, including symptoms page, fault causes page, solutions page. After entering the system, the default page is a symptoms page. As shown in Fig. 4. Fig.4 Qualitative analysis of gear box fault This system is developed by C# on the .NET platform. The overall framework of the system is as shown in Fig. 4. According to the structural of the gear box, the left side of page is made to accordion style, in which the six components fault of gear box are defined to the category name; the right side of the page defaults symptoms page, in which the previous column of symptoms column displays corresponding fault category and the next column links possible causes. 1) To query fault symptom In the symptoms page, the user can query page by page or do fuzzy query by writing keywords in the search box, and click the “search” button to get the results you want. 2) To query reason of fault 194 YANG Zhi-Ling et al. / Systems Engineering Procedia 4 (2012) 189 – 195 6 YANG Zhi-Ling / Systems Engineering Procedia 00 (2011) 000–000 The user can get corresponding possible causes of fault symptom by clicking on the details link of possible cause column in the above list. They can also directly click on the corresponding fault symptom in accordion structure to get a list of corresponding reasons. The query result is ordered by the weight of reason and the reason whose weight is large is sorted in the top of the list. 3) To query solution of fault After getting the list of fault reason, the user can click on the details link of solution column to get the recommended solution given by expert. Recommended solution will be ordered from top to bottom, if there are multiple. Maintenance staff can choose different solutions according to the circumstances. 5. Conclusion Gear box is one of the key components of wind turbine, whose fault will directly affects the safety and the operation of the overall wind turbine. This paper develops an expert system of fault diagnosis for gearbox in wind turbine based on fault tree analysis. With the gradual accumulation of fault and repair instance about wind turbine gear box, the expert system based on fault tree analysis will play a better role in the fault diagnosis of wind turbine gear box. Acknowledgements This study is supported by “the Fundamental Research Funds for the Central Universities” (09MG23). References [1]WANG Yi, FENT Xiaoyun. Electric locomotive fault diagnosis expert system based on fault tree. Electric Locomotives & Mass Transit Vehicles 2004; 27: 35-36. [2]Jian-Da Wu, Chiu-Hong Liu. An expert system for fault diagnosis in internal combustion engines using wavelet packet transform and neural network. Expert Systems with Applications 2009; 36: 4278-4286. [3]Yaguo Lei, Ming J. Zuo, Zhengjia He, Yanyang Zi. A multidimensional hybrid intelligent method for gear fault diagnosis. Expert Systems with Applications 2010; 37: 1419-1430. [4]N. Saravanan, V.N.S. Kumar Siddabattuni, K.I. Ramachandran. A comparative study on classification of features by SVM and PSVM extracted using Morlet wavelet for fault diagnosis of spur bevel gear box. Expert Systems with Applications 2008; 35: 1351-1366. [5]N. Saravanan, S. Cholairajan, K.I. Ramachandran. Vibration-based fault diagnosis of spur bevel gear box using fuzzy technique. Expert Systems with Applications 2009; 36: 3119-3135. [6]Tomasz Barszcz, RobertB Randall. Application of spectral kurtosis for detection of a tooth crack in the planetary gear of a wind turbine. Mechanical Systems and Signal Processing 2009; 23:1352–1365. [7]Baoping Tang, Wenyi Liu, Tao Song. Wind turbine fault diagnosis based on Morlet wavelet transformation and Wigner-Ville distribution. Renewable Energy 2010; 35: 2862-2866. [8]Yonghua Jiang, Baoping Tang, Yi Qin, Wenyi Liu. Feature extraction method of wind turbine based on adaptive Morlet wavelet and SVD. Renewable Energy 2011; 36: 2146-2153. [9]Sajid Hussain, HossamA.Gabbar. A novel method for real time gear fault detection based on pulse shape analysis. Mechanical Systems and Signal Processing 2011; 25: 1287–1298. [10]ZHU Xuejun,CHEN Yu. The research of intelligent fault diagnosing methods based on FTA. Microcomputer Information 2005; 21(6): 123-124. [11]YANG Bin,DING Hui,LUO Weimin,ZHANG Xiaofei.( 2002) Expert system of transformer fault diagnosis based on knowledge, Proceedings of the CSEE, 22(10): 121-124. 195YANG Zhi-Ling et al. / Systems Engineering Procedia 4 (2012) 189 – 195 YANG Zhi-Ling / Systems Engineering Procedia 00 (2011) 000–000 7 [12]YU Aihua. Fault tree analysis and application to hydraulic system maintenance. Hoisting and Conveying Machinery 2006; 6: 60-62. [13]CHEN Yongwang. Application of fault tree analysis method in fault diagnosis for 16V280 diesel engine fuel system. Mechanical Research & Application 2010;21(3): 85-86. [14]SHI Dinghua,WANG Songrui. Fault tree analysis methods and theories. Beijing Normal University Press 1993:. 
work_mruwvk4m2rafjnvrt6lc63k3fq	----	Visual Privacy Protection Methods: A Survey José Ramón Padilla-Lópeza,∗, Alexandros Andre Chaaraouia, Francisco Flórez-Revueltab aDepartment of Computer Technology, University of Alicante, P.O. Box 99, E-03080 Alicante, Spain bFaculty of Science, Engineering and Computing, Kingston University, Penrhyn Road, KT1 2EE, Kingston upon Thames, United Kingdom Abstract Recent advances in computer vision technologies have made possible the de- velopment of intelligent monitoring systems for video surveillance and ambient- assisted living. By using this technology, these systems are able to automatically interpret visual data from the environment and perform tasks that would have been unthinkable years ago. These achievements represent a radical improve- ment but they also suppose a new threat to individual’s privacy. The new capabilities of such systems give them the ability to collect and index a huge amount of private information about each individual. Next-generation systems have to solve this issue in order to obtain the users’ acceptance. Therefore, there is a need for mechanisms or tools to protect and preserve people’s privacy. This paper seeks to clarify how privacy can be protected in imagery data, so as a main contribution a comprehensive classification of the protection methods for visual privacy as well as an up-to-date review of them are provided. A sur- vey of the existing privacy-aware intelligent monitoring systems and a valuable discussion of important aspects of visual privacy are also provided. Keywords: visual privacy protection, video surveillance, ambient-assisted living, computer vision, image processing 1. Introduction It can be observed that world population is ageing. In fact, it is estimated that population over 50 will rise by 35% between 2005 and 2050, and those over 85 will triple by 2050 (EC, 2010). Furthermore, the number of people in long-term care living alone is expected to increase by 74% in Japan, 54% in France and 41% in the US (EC, 2008). Therefore, this situation will not be ∗Corresponding author Email addresses: jpadilla@dtic.ua.es (José Ramón Padilla-López), alexandros@dtic.ua.es (Alexandros Andre Chaaraoui), F.Florez@kingston.ac.uk (Francisco Flórez-Revuelta) Preprint submitted to Expert Systems with Applications February 3, 2015 sustainable in the near future, unless new solutions for the support of the older people, which take into account their needs, are developed. Ambient-assisted living (AAL) aims to provide a solution to this situation. AAL applications use information and communication technologies to provide support to people so as to increase their autonomy and well-being. Video cam- eras are being used more and more frequently in AAL applications because they allow to get rich visual information from the environment. The advances pro- duced in the last decades have contributed to this. The computational power has been increased while, at the same time, costs have been reduced. Fur- thermore, computer vision advances have given video cameras the ability of ‘seeing’, becoming smart cameras (Fleck & Strasser, 2008). This has enabled the development of vision-based intelligent monitoring systems that are able to automatically extract useful information from visual data to analyse actions, activities and behaviours (Chaaraoui et al., 2012b), both for individuals and crowds, monitoring, recording and indexing video bitstreams (Tian et al., 2008). By installing networks of cameras in people homes or care homes, novel vision- based telecare services are being developed in order to support the older and disabled people (Cardinaux et al., 2011). But these new technologies also sup- pose a new threat to individual’s privacy. Traditionally, cameras are used in public spaces for surveillance services in streets, parking lots, banks, airports, train stations, shopping centres, museums, sports installations and many others. It is estimated that there is an average of one camera for every 32 citizens in the UK, one of the most camera-covered countries in the world (Gerrard & Thompson, 2011). In short, video cameras are mainly used in outdoor environments and in public places, but they are not commonly used within private environments due to people’s concerns about privacy. Generally, the use of video cameras in public places has been tolerated or accepted by citizens, whereas their use in private spaces has been refused. There may be several reasons to explain this difference. On the one hand, the perceived public-safety benefits favor the usage of cameras in public places for crime prevention, fight against terrorism and others. On the other hand, there is a widespread belief that while staying in public environments, people’s sensitive information will not be exposed. Finally, there are some attitudes which have also contributed to accept their use, for example, to assume that anyone demanding privacy must have something to hide (Caloyannides, 2003). In traditional video surveillance systems cameras are managed by human operators that constantly monitor the screens searching for specific activities or incidents. As estimated by Dee & Velastin (2008), the ratio between human operators and screens is around 16 displays for every operator in four local authority installations within the UK. Although they can only really watch 1- 4 screens at once (Wallace & Diffley, 1988), this does not prevent abuses of these systems by their operators. Furthermore, the processing capacities of next-generation video surveillance systems and the increasing number of closed- circuit television cameras installed in public places are raising concerns about individual’s privacy in public spaces too. In the near future it is expected that cameras will surround us in both public 2 and private spaces. Intelligent monitoring systems threaten individual’s right to privacy because of automatic monitoring (Adams & Ferryman, 2013). These systems can retain a variety of information about people habits, visited places, relationships, and so on (Coudert, 2010). It is known that some systems already use facial recognition technology (Goessl, 2012). This way, these systems may build a profile for each citizen in which the people identity and related sensi- tive information is revealed. Therefore, this evolution of intelligent monitoring systems could be seen as approaching an Orwellian Big Brother, as people may have the feeling of being constantly monitored. In the light of the above, it is clear that the protection of the individual’s privacy is of special interest in telecare applications as well as in video surveil- lance, regardless whether they operate in private or public spaces. Therefore, privacy requirements must be considered in intelligent monitoring systems by design (Langheinrich, 2001; Schaar, 2010). As aforementioned, smart cameras become essential for AAL applications. Given that security and privacy protec- tion have become critical issues for the acceptance of video cameras, a privacy- aware smart camera would make it possible to use video cameras in realms where they have never been used before. If individual’s privacy can be guaran- teed through the use of this technology, public acceptance would be increased giving the opportunity of installing these cameras in private environments to replace simpler binary sensors or, most importantly, to develop new telecare services (Olivieri et al., 2012; Chen et al., 2012; Morris et al., 2013). This breakthrough could open the door to novel privacy-aware applications for am- bient intelligence (AmI) (Augusto et al., 2010), and more specifically in AAL systems for ageing in place (O’Brien & Mac Ruairi, 2009), being beneficial to improve the quality of life and to maintain the independence of people in need of long-term care. 1.1. Related studies The focus of this review is on the protection of visual privacy. There are valuable reviews about AmI and AAL that have already considered privacy in video and have also highlighted its importance for the adoption of video-based AAL applications (Cook et al., 2009; Cardinaux et al., 2011). But these works scarcely go into detail about how visual privacy protection can be achieved. Other works from the video surveillance field have also analysed this topic but from a different point of view (Senior et al., 2003, 2005; Cavoukian, 2013). The main threats and risks of surveillance technologies like closed-circuit television cameras, number plates recognition, geolocation and drones are discussed in depth. As a consequence, some guidelines to manage privacy are also proposed, but how to protect visual privacy is not considered. In the same line, Senior & Pankanti (2011) unify their previous works and extend the review of visual privacy not only to video surveillance but also to medical images, media spaces and institutional databases. They consider some technologies to protect visual privacy and provide a classification. As far as we know, this is the first attempt to provide such a classification of protection methods for visual privacy but it is not a comprehensive one. In a more recent work (Winkler & Rinner, 2014), 3 security and privacy in visual sensor networks are reviewed. Although they per- form a detailed analysis of the security from several points of view (data-centric, node-centric, network-centric and user-centric), they do not provide an in-depth analysis of privacy protection. In this survey, we focus on giving an answer to the question of how the vi- sual privacy can be protected, and how such a kind of protection is developed by some of the existing privacy-aware intelligent monitoring systems that have been found in the literature. Because of this, a comprehensive classification of visual privacy protection methods is provided as the main contribution of this work. The remainder of this paper is organised as follows: Section 2 gives an intuitive notion of what visual privacy protection is. A comprehensive review of visual privacy protection methods is presented in section 3. In section 4, relevant privacy-aware intelligent monitoring systems are introduced. A discus- sion of important privacy-related aspects is carried out in section 5. Finally, a summary of the present work as well as future research directions are given in section 6. 2. Visual privacy protection Privacy protection consists in preventing that the information that an in- dividual wants to keep private becomes available to the public domain. In the context of images and videos, we refer to it as visual privacy protection. In this paper, the terms visual privacy and privacy will be used indistinctly, except when indicated. First of all, it is worth to clarify when individual’s privacy needs to be protected. When protecting privacy, it can be differentiated between person’s identity and sensitive information which has to be kept in private. Video can convey an enormous amount of information that can be qualified as sensitive. Nevertheless, if sensitive information is present in a video but person’s identity is not, there is no privacy loss. The same is true whether person’s identity is in a video but without any sensitive information. In both cases, privacy is protected because there does not exist any association or mapping between sensitive information and person’s identity. Another important issue related to visual privacy is which is the sensitive information or region of interest to be protected. In many works only the face is obscured but that is not enough to protect visual privacy. Even when the person’s face is obscured, other elements could exist in the image through which person identification may be performed, for instance, using inference channels and previous knowledge (Saini et al., 2014). Visual cues like clothes, height, gait, and the like can be used to identify the person. For instance, in a pair- wise constraints identification (Chang et al., 2006; Chen et al., 2009) where faces had been masked, observers were able to identify whether a person in one image was the same one than in a different image. In that test, recognition hit rate was higher than 80%. By using this information and only detecting an image where a privacy breach exists, the person may be identified and tracked 4 in images where privacy was presumably preserved. These visual clues must be considered in order to protect privacy as they affect to the election of which regions of interest have to be protected. So, there is not actually only one region of interest but multiple. A region of interest should be extended to a wider area in some cases, while two or more regions of interest should be created in others. 3. Protection methods There are different ways to protect and preserve the privacy of individuals appearing in videos and images (see Table 1). Two approaches can be found if we consider the temporal aspect of when a protection method is used, i.e. before the image is acquired, or after it. On the one hand, it is possible to prevent others of successfully capturing images in which an individual appears. On the other hand, once an image exists sensitive or private information (e.g. faces, number plates, and so on) can be removed (Figure 1). According to how privacy is protected, protection methods can be classified in five large categories: intervention (section 3.1), blind vision (section 3.2), secure processing (section 3.3), redaction (section 3.4) and data hiding (sec- tion 3.5). Although the latter is often used along redaction methods, it has been added as it is a large field of research itself. In this section, a review of commonly used protection methods to protect visual privacy is presented. Table 1: Overview of reviewed papers according to the used visual privacy protection method. Category Count References Intervention 4 Wagstaff (2004), Patel et al. (2009), Harvey (2010), Mitskog & Ralston (2012) Blind vision 5 Avidan & Butman (2006), Erkin et al. (2009), Avidan et al. (2009), Sadeghi et al. (2010), Shashanka (2010) Secure processing 8 Erturk (2007), Shashank et al. (2008), Park & Kautz (2008), Ito & Kiya (2009), Upmanyu et al. (2009), Ng et al. (2010), Chaaraoui et al. (2012a), Zhang et al. (2012) Redaction: Image filter 10 Hudson & Smith (1996), Zhao & Stasko (1998), Boyle et al. (2000), Neustaedter & Greenberg (2003), Kitahara et al. (2004), Mart́ınez-Ponte et al. (2005), Neustaedter et al. (2006), Zhang et al. (2006), Frome et al. (2009), Agrawal & Narayanan (2011) Continued on next page 5 Table 1 – Continued from previous page Category Count References Redaction: Encryption 27 Spanos & Maples (1995), Macq & Quisquater (1995), Agi & Gong (1996), Tang (1996), Qiao et al. (1997), Zeng & Lei (2003), Yang et al. (2004), Boult (2005), Yabuta et al. (2005), Yabuta et al. (2006), Dufaux & Ebrahimi (2006), Dufaux et al. (2006), Chattopadhyay & Boult (2007), Baaziz et al. (2007), Raju et al. (2008), Dufaux & Ebrahimi (2008a), Dufaux & Ebrahimi (2008b), Xiangdong et al. (2008), Carrillo et al. (2010), Dufaux & Ebrahimi (2010), Tong et al. (2010), Tong et al. (2011), Dufaux (2011), Sohn et al. (2011), Pande & Zambreno (2013), Li et al. (2013), Ra et al. (2013) Redaction: K-same family 7 Newton et al. (2005), Gross et al. (2006a), Gross et al. (2006b), Bitouk et al. (2008), Gross et al. (2009), De la Hunty et al. (2010), Lin et al. (2012) Redaction: Object / people removal 22 Kokaram et al. (1995), Igehy & Pereira (1997), Masnou & Morel (1998), Morse & Schwartzwald (1998), Efros & Leung (1999), Bertalmio et al. (2000), Efros & Freeman (2001), Criminisi et al. (2003), Criminisi et al. (2004), Zhang et al. (2005b), Chan et al. (2006), Cheung et al. (2006), Shiratori et al. (2006), Wexler et al. (2007), Patwardhan et al. (2007), Whyte et al. (2009), Vijay Venkatesh et al. (2009), Koochari & Soryani (2010), He et al. (2011), Ma & Ma (2011), Ghanbari & Soryani (2011), Abraham et al. (2012) Redaction: Visual abstraction 14 Hodgins et al. (1998), Lyons et al. (1998), Tansuriyavong & Hanaki (2001), Fan et al. (2005), Williams et al. (2006), Chen et al. (2007), Baran & Popović (2007), Hogue et al. (2007), Chinomi et al. (2008), Chen et al. (2009), Qureshi (2009), Sadimon et al. (2010), Borosán et al. (2012), Chen et al. (2014) Data hiding 12 Petitcolas et al. (1999), Wu (2001), Cox et al. (2002), Yabuta et al. (2005), Zhang et al. (2005a), Ni et al. (2006), Yu & Babaguchi (2007), Paruchuri & Cheung (2008), Cheung et al. (2008), Cheung et al. (2009), Paruchuri et al. (2009), Guangzhen et al. (2010) 3.1. Intervention Intervention methods deal with the problem of preventing someone to cap- ture private visual data from the environment. They aim to create capture- resistant spaces. These methods physically intervene camera devices to prevent the acquisition of an image by means of a specialised device that interferes with the camera optical lens. For instance, Patel et al. (2009) developed the BlindSpot system. It locates any number of retro-reflective CCD or CMOS cam- era lenses around a protected area and, then, it directs a pulsing light at the detected lens, spoiling any images that cameras may record as Figure 2 shows. 6 Figure 1: Some privacy protected sequences of images: the first column shows the sensitive information or region of interest; the second column shows the region of interest replaced by the silhouette; the third column shows the sensitive information scrambled; and the last column shows the sensitive areas inpainted. Reprinted from Paruchuri et al. (2009). Similarly, Harvey (2010) proposed an anti-paparazzi device. It uses an array of high-power LEDs to produce a stream of light at over 12 000 lumen. Mitskog & Ralston (2012) have patented a camera blocker for a device with an integrated camera that uses a thin film organic polymer to neutralise camera lens. This camera blocker is reusable and adhesively sticks to any surface without leaving 7 Figure 2: Light beam neutralizing a camera lens. A light beam of single colour is used on the left picture, whereas a light beam generated using colour patterns is used on the right picture. Reprinted from Patel et al. (2009). a sticky residue. Aforementioned approaches are suitable when, once recorded, no control can be enforced over the use of images. However, when enforcement is possible, software-based methods can be used. For instance, the firmware of a camera or an application installed on a smartphone could be responsible of preventing the capture of certain environments, such as artworks in a museum, when the device comes into the range of a bluetooth transmitter (Wagstaff, 2004). Nev- ertheless, software-based intervention has some drawbacks. First of all, it can be easily thwarted by using a camera without any privacy software installed. Furthermore, users consent and their collaboration is required in order to work successfully. Because systems like these depend on third parties, they are likely doomed to failure. A new privacy legislation that enforces cameras to be in accordance with privacy protocols is needed. Anyway, given that no system is completely secure by nature, even under this assumption, it might be hacked. 3.2. Blind vision Blind vision has to do with image or video processing in an anonymous way. It addresses privacy related processing issues by means of secure multi-party computation (SMC) techniques that are applied to vision algorithms (Avidan & Butman, 2006). SMC is a subfield of cryptography that enables multiple parties to jointly compute a function in such a way that their inputs and the function itself are not revealed. When applied to vision this means that one party, Bob, could use a vision algorithm of another one, Alice, without enabling Bob to learn anything about Alice’s algorithm and, in the same extent, without enabling Alice to learn anything about Bob’s images. These methods are useful in order to use third-party algorithms to process data in a privacy-respectful way. For instance, using blind vision techniques, common tasks such as face detection, image matching, object tracking or image segmentation could be done in an anonymous way. 8 Concerning this, Avidan & Butman (2006) developed a secure classifier that was used in a blind face detector. This classifier is based on an oblivious transfer method and a secure dot-product protocol in order to compute the face detection algorithm. However, it is very time consuming due to the high computational load of the algorithm. Nevertheless, a proposal to accelerate the algorithm by using histograms of oriented gradients is considered at the risk of revealing some information about the original image. Similarly, Erkin et al. (2009) proposed an efficient algorithm that allows to jointly run the Eigenfaces (Turk & Pentland, 1991) recognition algorithm. This algorithm operates over homomorphically encrypted data. It allows matching an encrypted facial image with a database of facial templates in such a way that the biometric itself and the detection result are kept in secret from the server that performs the matching. Sadeghi et al. (2010) improved the communication and computation efficiency of the previous algorithm. Blind vision algorithms are also used for image matching. Avidan et al. (2009) presented a protocol for image matching related algorithms. Image matching is described as a generalisation of detection and recognition tasks. The described algorithm is built upon a secure fuzzy matching of SIFT at- tributes, which are treated as 16 character text strings, and it is used in their bag-of-features approach. A framework for privacy-preserving Gaussian Mixture Model (GMM) com- putations is proposed by Shashanka (2010). GMMs are commonly used in machine learning for clustering and classification. In computer vision they are mainly used in background subtraction (Zivkovic, 2004) for motion anal- ysis (Chan & Liu, 2009). 3.3. Secure processing There are other methods that are not based on SMC but that process visual information in a privacy respectful way. They are referred as secure processing in this work. For example, an image matching algorithm for private content- based image retrieval (PCBIR) is proposed by Shashank et al. (2008). PCBIR is related with similarity search. Conversely to blind vision, privacy is required in one direction because the whole database is usually public, while the query is kept private. Another possibility is to work with images in a domain that does not reveal visual information. Ito & Kiya (2009) presented an image matching algorithm using phase-only correlation (POC) in the frequency domain. This algorithm preserves the visual privacy of the images in a template database. In order to achieve this, all the images of the template database are converted to the frequency domain. Then, a phase scrambling using an one-time key is applied to the discrete Fourier transformation (DFT) coefficients. Afterwards, in order to match a query image with an image of the templates database, the query image is converted to the frequency domain and the matching is done with POC using the DFT coefficients as inputs. Algorithms that preserve privacy in an implicit manner, for instance, re- jecting visual information that is not necessary for the algorithm to work are 9 also considered under the umbrella of secure processing. Ng et al. (2010) pro- posed a privacy-preserving stereoscopic vision algorithm. It preserves privacy by calculating the disparity map using One-Bit Transform (Erturk, 2007). This way, each pixel in the input images is reduced to one bit. In this process, a huge amount of information (colour, texture and so on) is removed in the out- put images, complicating thus the identification of persons appearing on the images. A prototype of a privacy-preserving system for recognition of activities of daily living is proposed by Park & Kautz (2008). This prototype relies on the silhouette of detected foreground objects and the motion-map of the frame in order to analyse the activities. This way, if the silhouette and motion-map gen- eration is performed within the camera and it is not possible to access the RGB signal, the visual privacy of the persons inhabiting the environment would be ensured from the algorithm point of view. Similarly, Chaaraoui et al. (2012a) use silhouettes in their efficient approach for multi-view human action recogni- tion based on bag of key poses. By using silhouettes, identifiable information is removed, hence it can also be considered a privacy-aware method. Depth infor- mation from RGB-D cameras can be used as a way to preserve privacy (Zhang et al., 2012) as well. Depth data can be obtained from low-cost structured-light cameras like Microsoft Kinect or Asus Xtion, or time-of-flight cameras like Mi- crosoft Kinect 2. Given that no colour information is involved, a depth map visualisation does not enable face recognition and prevents direct extraction of visual clues for person identification. Upmanyu et al. (2009) presented an interesting approach where an efficient framework to carry out privacy-preserving surveillance is proposed. This so- lution is inspired in a secret sharing scheme adapted to image data. In this framework an image is split into a set of random images in such a way that the shares do not contain any meaningful information about the original image. De- spite of shares being distributed between a certain number of servers, computer vision algorithms can be applied securely and efficiently. 3.4. Redaction Image and video modification or redaction methods are the most common vi- sual privacy protection methods. They modify the sensitive regions of an image such as faces, bodies, number plates, etc. to conceal private information con- cerning the subjects appearing on it. In order to determine the privacy sensitive regions in which a redaction method operates, computer vision algorithms are used. However, this section focuses only on the application of privacy-preserving methods, therefore it is assumed that sensitive regions are correctly detected. According to the way in which an image is modified, redaction methods can be classified in several categories. To begin with, there are ad-hoc distor- tion/suppression methods (section 3.4.1). These methods modify the region of interest of an image, either completely removing sensitive information from the image, or modifying the information using common image filters like blurring or pixelating. By using these filters, sensitive regions are modified in order to make them unrecognisable. 10 More robust methods like image encryption (section 3.4.2) are also used to conceal the region of interest. Using image encryption techniques the privacy sensitive region of an image is ciphered by a key. Generally encryption tech- niques based on scrambling (permutation) were commonly used in analogue video, but they can also be used in digital video. Besides of that, it must be taken into account that in the literature scrambling is used as a synonym of encryption in some cases. In this paper, we consider scrambling techniques as a subfield of image encryption. Another approach to image redaction is face de-identification (section 3.4.3). These methods are focused on de-identifying the faces appearing in an image. Although the aforementioned methods can be used for face de-identification, we focus here on those based on the k-same family of algorithms that implements the k-anonymity protection model (Sweeney, 2002). These algorithms alter the face of a person in such a way that identity cannot be recognised but facial expressions are preserved. Finally, some approaches like object removal (section 3.4.4) use inpainting- based algorithms to completely remove sensitive regions of an image by filling the left gap with the corresponding background. Once the image has been inpainted, a visual abstraction of the removed sensitive information can be rendered, like a stick figure, a point, a silhouette, and the like (Chinomi et al., 2008). Regarding image modification, some questions have to be outlined. Redac- tion methods cannot modify an image in whatever way. As reported by Hudson & Smith (1996), there is a trade-off between providing privacy and intelligibil- ity in images. For instance, when an image is modified, information needed for image understanding may be also removed. So, a modified image could lack of utility and balancing privacy and intelligibility is needed. In other words, privacy protected images have to retain useful information needed by applica- tions built upon this information, such as telecare monitoring systems or video surveillance. It is also worth mentioning that whereas some of the redaction methods are irreversible, i.e. the modification cannot be undone, there are others methods that are reversible so the original version can be recovered if needed. Nevertheless, reversible methods can be developed by using irreversible redaction methods along with data hiding algorithms. An overview of redaction methods has been given so far. In next subsections, each one of the described categories will be dealt with in depth, focusing on the most representative works and summarising how the presented methods are used. 3.4.1. Image filtering Redaction methods based on image filtering use common image filters to apply various effects on images in order to modify privacy sensitive regions (Figure 3). Depending on the application, image filters can be used for ob- scuring human faces, human bodies, number plates or even background in video conferences (Kitahara et al., 2004; Mart́ınez-Ponte et al., 2005; Zhang 11 (a) Lena (b) Lena with Gaussian Blur (c) Lena Pixelation Figure 3: Several examples of image filtering methods: a) the original image without any modification, b) image modified by applying a Gaussian Blur filter, and c) image modified by applying a pixelating filter. et al., 2006; Frome et al., 2009). Among the most commonly used filters, blur- ring (Neustaedter & Greenberg, 2003; Zhang et al., 2006; Neustaedter et al., 2006; Frome et al., 2009) and pixelating (Boyle et al., 2000; Kitahara et al., 2004; Neustaedter et al., 2006) stand out. A blurring filter applies a Gaussian function over an image. This function modifies each pixel of an image using neighbouring pixels. As a result, a blurred image is obtained in which the details of sensitive regions have been removed. Blurring is widely used in Google Street View (Frome et al., 2009) to modify human faces and number plates. A pixelating filter divides an image into a grid of eight-pixel wide by eight-pixel high blocks. The average colour of the pixels of each block is computed and the resultant colour is assigned to all of the pixels belonging to that block. As a result, an image where the resolution of sensitive regions have been reduced is obtained. Pixelating is commonly used in television to preserve the anonymity of suspects, witnesses or bystanders. However, it is vulnerable to some kind of attacks such as integrating pixels along trajectories over time that may allow partly recovering of the obscured information. Lander et al. (2001) evaluated the effectiveness of pixelating and blurring in videos and static images. Results indicated that participants were still able to recognise some of the faces. Similarly, Newton et al. (2005) showed that pixelating and blurring alike filters do not thwart recognition software. Training a parrot recogniser using the same distortion as the probe on gallery images, high recognition rates are obtained (near 100%) despite looking somewhat de- identified to humans. 3.4.2. Encryption of videos and images Image and video encryption methods encode imagery data in such a way that the original data becomes unintelligible as can be observed in Figure 4. The main goal of these methods is reliable security in storage and secure transmis- sion of content over the network. Usually, security in cryptographic algorithms resides in the strength of the used key instead of in keeping the algorithm pri- 12 Figure 4: Two examples of an encrypted image where the face of the person is considered the sensitive region. Reprinted from Boult (2005). vate. By using encryption, a distorted video that unauthorised viewers cannot visualise is obtained. Only users who have the proper key for decryption can visualise it. Through the inverse operation and the used key, the ciphered data can be deciphered in order to retrieve the original images. This is named condi- tional access through encryption. Encryption methods operate over the whole frame or a delimited region of all the video frames. Although such methods do not provide a balance between privacy and intelligibility, they enable to perform data analysis over unprotected data once authorisation and required permissions have been granted. In such cases, privacy would be protected until access to raw data is eventually requested. Generally näıve video encryption algorithms have treated the compressed video bitstream as text data, therefore encrypting the entire video bitstream. Hence, commonly used encryption algorithms such as Data Encryption Standard (DES), Rivest’s Cipher (RC5), Advanced Encryption Standard (AES), Rivest, Shamir and Adleman (RSA) and so on, have been used. These algorithms guarantee the highest security level but, unfortunately, they are not suitable for real-time video encryption because they are very time consuming (Yang et al., 2004; Pande & Zambreno, 2013). Due to this, selective encryption algorithms have been proposed (Spanos & Maples, 1995). These algorithms keep using text- based encryption but encrypt only a selected part of the video bitstream so as to get real-time encryption. Other encryption algorithms have also been proposed for real-time encryption, namely light-weight encryption algorithms (Zeng & Lei, 2003). These algorithms are suitable for real-time applications because, when encrypting, they use a simple XOR cipher or only encrypt some bits of the video bitstream. Thereby, they are much faster than the first ones. Finally, there are methods based on scrambling (Tang, 1996). Traditional scrambling methods modify an analogue video signal like those found on closed-circuit television cameras to make it unintelligible. However, with the proliferation of digital video cameras, scrambling techniques are also applied to digital videos in the field of 13 video encryption. Mainly, scrambling algorithms are based on permutation only methods in which transformed coefficients are then permuted in order to distort the resulting image. Each one of these methods can operate in a specific domain, like the spatial domain, frequency domain (transform domain) or code-stream domain (com- pressed video). Furthermore, it is important to note that light-weight en- cryption and scrambling-based methods are less secure than näıve encryption. For instance, scrambling video in the spatial domain is subject to efficient at- tacks (Macq & Quisquater, 1995). Generally these algorithms trade-off security for encryption speed. However, compared to blurring and pixelating, scram- bling approaches as in Dufaux & Ebrahimi (2008a) are successful at hiding identity (Dufaux & Ebrahimi, 2010; Dufaux, 2011). Regarding when encryption is performed, there are several approaches: prior to encoding, after encoding or during encoding. Each approach has advantages and disadvantages. Prior-to-encoding encryption is a very simple method that works with the original image independently from the used encoding scheme. However, it significantly changes the statistics property of the video signal, resulting in a less efficient compression later. Regarding after-encoding encryp- tion, the compressed code-stream is encrypted after video encoding. The result- ing encrypted and compressed code-stream could hardly be reproducible in a standard player, and it could even cause the player to crash. However, it avoids to fully decode and re-encode the video. Finally, during-encoding encryption has the advantage of a fine-grained control over the encoding process but it is closely linked to the used video encoding scheme. Next, some of the selective and light-weight encryption as well as scrambling methods found in the literature will be analysed. Concerning selective encryption, AEGIS (Spanos & Maples, 1995) is an algo- rithm that uses DES to encrypt only the I-frames of the MPEG video bit- stream. By ciphering I-frames, the needed B-frames and P-frames cannot be reconstructed either. However, the algorithm is not completely secure due to partial information leakage from the I-blocks in P and B frames (Agi & Gong, 1996). Similarly, the video encryption algorithm proposed by Qiao et al. (1997) works with I-frames. It divides them in two halves that are XORed and stored in one half. One of the half is encrypted using DES algorithm. Although this algorithm is secure, it is not suitable for real-time applications. Raju et al. (2008) analyse the distribution of the DCT coefficients (DC and AC) of com- pressed MPEG videos in order to develop a computationally efficient and secure encryption scheme for real-time applications. DC and AC coefficients are man- aged differently regarding their visual influence, and electronic code block and cipher block chaining modes are interleaved to adapt the encryption process to the video data. The described scheme uses RC5 for encrypting DCT coefficients. Boult (2005) used DES and AES to encrypt faces in JPEG images during com- pression in their privacy approach through invertible cryptographic obfuscation. The information required for the decrypting process is stored inside the JPEG 14 file header. Although this information can be publicly read, it cannot be used without the private key. Chattopadhyay & Boult (2007) used this technique for real-time encryption, using uCLinux on the Blackfin DSP architecture. Simi- larly, an encryption scheme for JPEG images is used by Ra et al. (2013). The JPEG image is divided into two parts, one public and one private. The first one is unaltered, whereas in the second one the most significant DC coefficients are encrypted during the encoding process after the quantisation step. This ap- proach is designed to obtain JPEG-compliant images that can be sent to photo sharing services under storage and bandwidth constraints. As for light-weight video encryption, Zeng & Lei (2003) presented an efficient algorithm for H263 that operates in the frequency domain. Bit scrambling is used to transform coefficients and motion vectors during video encoding without affecting the compression efficiency. By using this method each frame of the re- sulting video is completely distorted. A cryptographic key is used to control the scrambling process, thereby authorised users will be able to undo the scrambling using the key. A similar video encryption algorithm for MPEG-4 is proposed by Dufaux & Ebrahimi (2006) where security is provided by pseudo-randomly inverting the sign of selected transform coefficients (frequency domain) corre- sponding to the regions of interest. The encryption process depends on a pri- vate key that is RSA encrypted and inserted in the stream as metadata. In this method the amount of distortion introduced can be adjusted from merely fuzzy to completely noisy. This method is deeply explained for the case of Motion JPEG 2000 in (Dufaux et al., 2006), and for the case of H264/AVC in (Dufaux & Ebrahimi, 2008a). Concerning the latter, Tong et al. (2010) made a proposal to correct the drift error produced in H264/AVC during video encryption. The described method is also used by Baaziz et al. (2007) in an automated video surveillance system. Dufaux & Ebrahimi (2008b) presented an extension of their previous work based on code-stream domain scrambling. The scrambling is per- formed after the MPEG-4 encoding process directly on the resulting code-stream output by the camera. This way, it avoids to decode and re-encode the video, saving computational complexity. An encryption scheme for JPEG XR (Srini- vasan et al., 2007) working in the frequency domain is proposed by Sohn et al. (2011). It uses subband-adaptive scrambling for protecting face regions. Con- cretely, different scrambling techniques are used for each subband: random level shift for DC subbands, random permutation for LP subbands, and random sign inversion for HP subbands. A different encryption scheme based on compressive sensing (CS) (Candes et al., 2006; Donoho, 2006) and chaos theory is proposed by Tong et al. (2011). This method scrambles sensitive regions of a video by using block-based CS sampling on their transform coefficients during encoding. The scrambling process is controlled by a chaotic sequence used to form the CS measurement matrix. In order to prevent drift error they also use their previous method. Finally, regarding scrambling, Tang (1996) proposed an encryption algo- rithm that works in the frequency domain and uses the permutation of the DCT coefficients in order to replace the zig-zag order. Specifically, instead of using the zig-zag order to map the 8x8 block to a 1x64 vector, a random per- 15 mutation list is used. A different approach is proposed by Yabuta et al. (2005) where scrambling is performed in the space domain before encoding. The scram- bling only affects the moving regions and is performed by randomly permuting pixels. Before being scrambled, the original moving regions are encrypted with AES, and embedded inside the masked image using digital watermarking (Pe- titcolas et al., 1999). Thereby, the scrambled regions can be recovered later if needed. An improved version of this method is presented in (Yabuta et al., 2006) being able of real-time encoding, and decoding only one specific object by exploiting object tracking information. Xiangdong et al. (2008) presented a novel encryption scheme that relies on chaos theory. It works in the space domain and prior to video encoding. It permutes the pixels of an image row by row by using a chaotic sequence of sorted real numbers as a key. This key can be used to de-scramble the image when necessary. A similar approach that sorts the chaotic sequence as Vigenère cipher is proposed by Li et al. (2013). Carrillo et al. (2010) proposed an encryption algorithm working in the spatial domain. Before encoding, pseudo-random permutations are applied to the pix- els. A secret pass phrase which controls the permutation process is used as key. This algorithm is independent of the used compression algorithm and robust to transcoding at the cost of an increase in bitrate depending on the percentage of encrypted blocks. 3.4.3. Face de-identification Face de-identification consists in the alteration of faces to conceal person identities. The goal is to alter a face region in such a way that it cannot be recognised using face recognition software. In order to achieve this, the faces appearing in images are modified by using some of the aforementioned methods such as image filtering previously discussed in section 3.4.1. Nevertheless, there are cases in which privacy must be preserved while images must still keep their capacity of being analysed. In such cases, it is then necessary to balance privacy and intelligibility. Concerning this, there are methods in the literature that consider this trade-off, and some of them will be reviewed in this section. The k-Same family of algorithms (Figure 5) is one of the most recent and commonly used algorithms for face de-identification. K-Same was first intro- duced by Newton et al. (2005). Intuitively, k-Same performs de-identification by computing the average of k face images in a face set. Then, all of the im- ages of the given cluster are replaced by the obtained average face. Using this algorithm, a de-identified face is representative of k members of the original used face set. This way, if the probability of a de-identified face of being cor- rectly recognised by a face recognition software is no more than 1 k , it is said that this algorithm provides k-anonymity privacy protection (Sweeney, 2002). However, despite the fact that this formal model can preserve privacy, there are no guarantees of the utility of the data. Gross et al. proposed some extensions to the k-Same. On the one hand, an algorithm named k-Same-Select that extends the k-Same algorithm is presented in (Gross et al., 2006a). It guarantees the utility of the data, for example, by preserving facial expressions or gender in face images. This algorithm di- 16 Figure 5: Several examples of de-identified face images: a) faces de-identified by using the k-Same algorithm where some ghosting artefacts appear due to misalignments in the face set. b) faces de-identified by using k-Same-M algorithm. Reprinted from Gross et al. (2009). vides the face image set into mutually exclusive subsets using a data utility function and, then, it uses the k-Same algorithm on the different subsets. On the other hand, the k-Same-M algorithm is introduced in (Gross et al., 2006b) to fix a shortcoming of the k-Same-Select algorithm. Appearance-based algo- rithms, like k-Same-Select, work directly in the pixel level producing sometimes alignment mismatch that leads to undesirable artefacts in the de-identified face images. In order to overcome this issue k-Same-M relies on active appearance model (Cootes et al., 1998, 2001), a statistical and photo-realistic model of the 17 shape and texture of faces. This model is also used by De la Hunty et al. (2010) for real-time facial expression transferring from one’s person face to another. Despite this method is not used for face de-identification, it could be useful to transfer expressions from an already de-identified face to another one. The k-Same-based algorithms have some constraints. For instance, if a sub- ject is represented more than once in the dataset then k-Same does not provide k-anonymity privacy protection. In order to address this shortcoming, Gross et al. (2009) proposed a multi-factor model which unifies linear, bilinear and quadratic models. By using a generative multi-factor model, a face image is fac- torised into identity and non-identity factors. Afterwards, the de-identification algorithm is applied and the de-identified face is reconstructed using the multi- factor model. A different approach is described by Bitouk et al. (2008), where faces are automatically replaced in photographs. A large library of 2D faces downloaded from the Internet is built according to appearance and pose. In order to replace a detected face in an image, a similar candidate to the input is selected from the library. Lin et al. (2012) presented a similar work for face swapping where a personalised 3D head model from a frontal face is built, thereby any pose can be rendered by directly matching with the face that wants to be substituted. 3.4.4. Object / people removal Object and people removal deals with concealing persons or objects appear- ing in an image or a video in such a way that there are not trails of them in the resulting modified version (Figure 6). For the sake of clarity, the term ‘object’ will be used to refer both person and object indistinctly. When removing an object from an image a gap is left. This gap has to be filled in order to create a seamless image. Inpainting methods are used to repair regions with errors or damages. Inpainting consists in reconstructing missing parts in such a way that the modification is undetectable. Information from surrounding area is used to fill in the missing areas. Therefore, these methods can be used for visual privacy to remove people that do not commit suspicious activities from video surveillance recordings, removing inactive participants from the video stream of a video conference, concealing people from images in online social networks, and so many other applications. However, due to computational restrictions, inpainting methods are rarely used in real-time applications running on com- modity hardware (Granados et al., 2012). Inpainting can be divided into two large groups: image inpainting (Bertalmio et al., 2000) and video inpainting (Kokaram et al., 1995; Abraham et al., 2012). This distinction is mainly due to the content nature. While in an image is only needed to ensure spatial consistencies, in video, temporal consistencies between all of the frames have to be ensured as well. Regarding image inpainting, there are several approaches: texture synthe- sis (Igehy & Pereira, 1997; Efros & Leung, 1999; Efros & Freeman, 2001), partial differential equations (PDE) inspired algorithms (Masnou & Morel, 1998; Morse & Schwartzwald, 1998; Bertalmio et al., 2000; Chan et al., 2006), and exemplar based (Criminisi et al., 2003, 2004; Whyte et al., 2009; Koochari & Soryani, 18 (a) Real image (b) Modified image Figure 6: An example of a people removal method where the person has been manually selected in the real image (a), and then automatically removed in the second image (b) by filling the region concerning the person using an exemplar-based image inpainting method. Reprinted from Criminisi et al. (2004). 2010; He et al., 2011; Ma & Ma, 2011). In texture synthesis, a synthetic texture derived from one portion of the image is used to fix another portion. This syn- thetic texture is a plausible patch that does not have visible seams nor repetitive features. Algorithms based on texture synthesis are able to fill in large regions but at the cost of not preserving linear structures. PDE inspired algorithms re- construct the gap using geometry information to interpolate the missing parts. A diffusion process propagates linear structures of equal gray value (isophotes) of the surrounding area into the gap region. Although these algorithms pre- serve well linear structures, the diffusion process introduces some blurring when filling in large regions. Finally, exemplar-based methods are of particular in- terest. Instead of generating synthetic textures, these methods operate under the assumption that the information that is necessary to complete the gap has to be fetched from nearby regions of the same image. They generate new tex- tures by searching for similar patches in the image with which the gap is filled. Moreover, some exemplar-based algorithms also search the needed information in databases of millions of images in order to complete the remaining infor- mation (Whyte et al., 2009). Furthermore, these algorithms often combine the advances of texture synthesis and isophote-driven inpainting by a priority-based mechanism which determines the region filling order, thereby reducing blurring caused by prior techniques. 19 (a) Real image (b) Blur (c) Pixelating (d) Emboss (e) Solid silhouette (f) Skeleton (g) 3D avatar (h) Invisibility Figure 7: Several examples of visual abstractions where the person in the real image (a) has been replaced by a visual model. Regarding video inpainting, some of the first straightforward approaches tried to apply image inpainting methods to individual images of the underlying video data. However, they did not take full advantage of the temporal correla- tion of video sequences. Due to this, the previous methods are often modified in order to be adapted to sequences of images, reconstructing a given frame by interpolating missing parts from adjacent frames. These methods can be classified into patch-based methods (Shiratori et al., 2006; Wexler et al., 2007; Patwardhan et al., 2007; Ghanbari & Soryani, 2011) and object-based meth- ods (Zhang et al., 2005b; Cheung et al., 2006; Vijay Venkatesh et al., 2009). As patch-based methods are unable to perform both spatial and temporal aspects simultaneously, object-based methods were introduced in order to overcome these constraints. 3.4.5. Visual abstraction / object replacement Object replacement involves the substitution of objects (or persons) appear- ing in an image or video by a visual abstraction (or visual model) that protects the privacy of an individual while enabling activity awareness. As far as we know, the term ‘visual abstraction’ was early coined by Chinomi et al. (2008) 20 to refer to a visual model that abstracts a replaced object. Common visual models could be a point, a bounding box, a stick figure, a silhouette, a polyg- onal model, and many others as seen in Figure 7. Visual abstraction may be obtained in a a variety of ways. Hence, there is an overlap with some of the previous reviewed methods, such as image filtering or face de-identification, that could also be used as visual models. Object replacement does not necessarily imply removing the object to be replaced, but quite often the aforementioned techniques in section 3.4.4 intervene. In such cases, an object removal method is applied, followed by the rendering of the visual model over the inpainted image. The abstract object is often located in the same relative position, pose, and orientation as the original object. Although it depends on the used visual model, by using a proper abstract representation of an object, the information of the scene remains useful so as visual analysis can still be carried out, for instance, to assess the underlying activity before an accident in a home risk detection service. Then, the activity can be analysed without violating the right to privacy of people appearing in the image because it is not possible to directly identify them. However, as reported by Hodgins et al. (1998), some works in human motion perception showed that viewers easily recognise friends by their gaits, as well as the gender of unfamiliar persons when using moving light displays (Johansson, 1973). Concerning this, it seems that motion is essential for identifying human figures. Furthermore, apparently the geometric model used for rendering human motion affects the viewer’s perception of motion. For example, in experiments carried out by Hod- gins et al., subjects were able to better discriminate motion variations using the polygonal model than they were with the stick figure model. Hence, a study to determine how visual abstraction techniques actually preserve privacy must be done. Different works employ visual abstraction and object replacement. For ex- ample, a silhouette representation can be used as a visual abstraction of a per- son (Tansuriyavong & Hanaki, 2001). This removes information about textures while maintaining the shape of the person, thereby complicating the identifi- cation. A silhouette representation is used, for instance, to preserve individ- ual’s privacy in a fall detector and object finder system (Williams et al., 2006). Another representation can be obtained by using an edge motion history im- age (Chen et al., 2007, 2009). By using this pseudo-geometric model, the whole human body is obscured and the person looks like a ghost in the final image. This method detects edges of the body appearance and motion, accumulating them through the time, in order to smooth the noise, and partially preserve body contours. Chinomi et al. (2008) proposed twelve operators for visual abstrac- tions: as-is, see-through, monotone, blur, mosaic, edge, border, silhouette, box, bar, dot and transparency. In order to choose among one of them, the relation- ship between the subject being monitored and the viewer is taken into account. In addition to the representation that have been seen so far, 3D avatars can be used too. Avatar creation (Sadimon et al., 2010) is a very interesting field with straightforward application in visual abstraction as well. Lyons et al. (1998) presented a method that uses automatic face recognition and Gabor wavelet 21 transform for creating avatars. It automatically extracts a face from an image, and create a personalised avatar by rendering the face into a generic avatar body. However, given that the avatar maintains recognisable aspects of the face, pri- vacy would not be protected. Hogue et al. (2007) proposed a method based on 3D reconstruction for whole body avatar creation. They use a portable stereo video camera to extract the geometry of the person and create a 3D model. A similar and more recent work is proposed by Chen et al. (2014), where two low- cost RGB-D cameras are used to scan the whole body of a person in order to create a mesh model. Although the resulting textured model of both methods do not protect privacy, they can constitute a building block for other methods where generic 3D avatars are used to replace persons, as proposed by Fan et al. (2005). Furthermore, automatic rigging (Baran & Popović, 2007; Borosán et al., 2012) could be used to animate 3D models, enabling the use of customised 3D models in real-time. Hence, opening the door to new protection methods that transform or deform mesh models. Finally, the use of object-video streams for preserving privacy is proposed by Qureshi (2009). A separated video bitstream is generated for each foreground object appearing in the raw video. Thereby, the original video can be reconstructed from object-video streams without any data loss. During reconstruction, each video bitstream can be rendered in several ways, for instance, obscuring people identities using a silhouette representation or just not showing an object-video stream at all. 3.5. Data hiding based methods In order to protect privacy, there are redaction methods that apart of mod- ifying the region of interest, they embed the original information inside of the modified version so as to be retrieved in the future if needed (Cheung et al., 2009). These redaction methods make use of data hiding techniques (Petitcolas et al., 1999) to develop reversible methods when the underlying redaction does not support it. In Figure 8, a classification of data hiding methods is presented. Concerning the terminology used in data hiding, the embedded data is the mes- sage that will be sent secretly. It is often hidden in another message referred to as cover message whose content can be text, audio, image or video. As a result of a hidden process, a marked message is obtained. Generally data hiding techniques are used for steganography, digital water- marking and fingerprinting. Steganography is a practice that consists in con- cealing a secret message as embedded data inside a cover message. The hiding process is often controlled by a key in order to allow the recovery of the secret message to parties that know it. Regarding digital watermarking and finger- printing, they both use steganography techniques but are focused on copyright protection. On the one hand, digital watermarking encodes information about the owner of an object by means of a visible pattern (e.g. a company logo) or hidden information. On the other hand, fingerprinting is used to embed hidden serial numbers that uniquely identify an object in such a way that the owner of the copyright can detect violations of licence agreements. Data hiding methods have to provide different kind of features in terms of capacity, perceptibility and robustness (Cox et al., 2002). The main difference between steganography 22 Information hiding Covert channels Steganography Linguistic steganography Technical steganography Anonymity Copyright marking Robust copyright marking Fingerprinting Watermarking Imperceptible watermarking Visible watermarking Fragile watermarking Figure 8: A classification of data hiding techniques according to Petitcolas et al. (1999). and digital watermarking (and fingerprinting) is their application. Whereas the former aims for imperceptibility to human vision, the latter is focused on the robustness. Invisible watermarking can be used for privacy protection purposes. Instead of embedding information about the owner, the original video is embedded in- side of the privacy protected version of it. Due to this, data hiding methods that provide large embedding capacity are required. Furthermore, if the reversibility of the hiding process is considered, irreversible and reversible data hiding tech- niques can be found (Ni et al., 2006). In the former, the cover video cannot be fully restored after the hidden data is extracted, but it usually produces higher hiding capacity. In the latter, the cover video can be fully restored, thereby maintaining the authenticity of the original video. Nevertheless, reversible data hiding is not required in privacy protection because the cover video is discarded once the embedded data has been extracted. Some of the privacy protection methods that make use of non-reversible and imperceptible watermarking are described next. For instance, Yabuta et al. (2005) extract moving objects of a motion JPEG video. These regions are scrambled in the original video, resulting in a privacy protected video that is used as the cover video. Then, the extracted objects are JPEG compressed and encrypted with AES, followed by a hiding process that embeds them into the least significant bits of middle frequency DCT coefficients of the cover video. The added perturbation is small and does not visually affect the reconstructed 23 image. A similar approach is proposed by Zhang et al. (2005a), where fore- ground objects are extracted from a H263 video, compressed and encrypted as a regular video bitstream. The gaps left by the foreground objects are inpainted in the original video using background replacement. Then, the encrypted fore- ground objects are embedded in the DCT coefficients of the inpainted video considering the perceptual quality of the marked video. However, one drawback is the increased bitrate of the resultant video. Yu & Babaguchi (2007) presented a method for hiding a face into a new generated face. It considers the face as the information to be embedded. Initially, an active appearance model is built and faces are characterised as parameters according to this model in order to reduce the payload of the embedded information. Then, the face appearing in the orig- inal MPEG2 video is masked. Finally, face parameters are embedded into DCT coefficients of the video following the quantised index modulation scheme (Cox et al., 2002). A novel data hiding algorithm for M-JPEG that minimises both the output perceptual distortion and the output bitrate is proposed by Paruchuri & Cheung (2008). The algorithm identifies the optimal locations to hide data by selecting the DCT coefficients that minimise a cost function that considers both distortion and bitrate. The person appearing in a video (privacy data) is extracted and removed from it. Then, the privacy data is embedded into the selected DCT coefficients of the inpainted video. The proposed algorithm is used by Cheung et al. (2008) but modifying the rate-distortion scheme to determine the optimal number of bits to be embedded in each DCT block. This last version is used for a reversible data hiding scheme in a video surveillance prototype (Cheung et al., 2009; Paruchuri et al., 2009). Guangzhen et al. (2010) presented a different method for JPEG images. It uses scaling and wavelet coef- ficients generated by DWT. The former is used to build a low-resolution image where privacy information appears pixelated, whereas the latter is used along scaling coefficients to jointly recover the original image. In this way, wavelet coefficients represent only the private information that needs to be embedded. They are embedded in the DCT coefficients of the low-resolution image after quantisation via amplitude module modulation (Wu, 2001). 4. Privacy-aware intelligent monitoring systems This section introduces some of the existing expert and intelligent systems in the video surveillance and AAL fields that take privacy into account and, thereby, are developed as a framework to protect it. Some of the reviewed systems focus only on the way in which private date is managed, whereas others use redaction methods to protect individual’s privacy before imagery data is stored, distributed, shared or visualised. Because of this, some privacy design questions which such systems should address are drawn (Senior et al., 2005; Yu et al., 2008): • What data is captured? • Has the subject given consent? 24 • How does the subject specify the privacy preferences? • What form does the data take? • Who sees the data? • How long is the data kept? • How is the data used? • How raw is the data? • Who and what should be protected? • How should privacy be protected in a video surveillance system? • How can privacy be protected without losing utility? The answers to these questions lead to privacy policies and technical aspects that such systems should cope with in one way or another. In the same line, it is also important to question when redaction is performed. As stated by Senior (2009), there are mainly several locations where redaction can be performed for a general video-based monitoring architecture made up of cameras, video processor, database and user interfaces. Concretely, these locations include all of the mentioned parts but cameras because of technical issues that may arise in legacy systems. Nevertheless, cameras have been also included here from a conceptual point of view, so there are a total of four locations. In such an architecture, data flows from video cameras to user interfaces, crossing through the video processor and the database. Furthermore, depending on privacy policies of subjects, viewers permission, and others, both redacted and unredacted data may have to be displayed. Next, we enumerate the several locations where redaction can be applied: 1. User Interface. This location is the most insecure. Data crosses the system without being protected until it reaches the user interface. Then, redac- tion is carried out by the user interface and the protected information is presented to the viewer. An advantage of this approach is that redacted data does not require to be stored. However, metadata information needs to be delivered with the raw data. 2. Database. When the user interface requests the database for some infor- mation to view, the latter can redact it. Thereby, redacted data does not require to be stored but it involves additional processing because the same data may be redacted multiple times. Besides of that, latency issues may arise due to extra processing. 3. Video Processor. Redaction can be performed by the video processor before storing. It analyses video bitstreams to detect activities and extract useful information. Nevertheless, bandwidth and storage requirements are increased because both redacted and unredacted data need to be sent to the database. 25 4. Video Camera. A privacy-aware smart camera can redact sequences of images itself. This is the earliest possible stage in which redaction can be applied. Similarly to the video processor, in this location bandwidth and storage requirements are increased too. Several redaction locations can be combined to enhance security as in double redaction also proposed by Senior (2009), in which privacy protection is applied at the earliest stage as well as in other locations. Video is decomposed in mul- tiple information streams containing the private data which are encrypted and flow through the system to the database. The information can be recombined later, when needed by an authorised viewer. By using double redaction more than one level of privacy protection can be provided, where each one could be suitable for different applications. This way, some visual information could not be protected but securely stored, whereas other information could be redacted according to several factors like subject, viewer, ongoing activity and so on. Next, some of the existing systems found in the literature are described. CoMedi (Coutaz et al., 1999) is a media space prototype that facilitates remote informal communication and group awareness while assuring privacy protection. A porthole display with fish-eye feature is used in order to provide awareness of the remote activities that are being carried out. It shows the personal in- formation that the corresponding remote user has previously accepted to reveal without losing awareness about peripheral activities. Regarding privacy, this prototype uses a face tracker and a privacy filter based on eigenspace coding in order to filter the captured faces not belonging to the image set of ‘socially correct’ faces. NeST (Fidaleo et al., 2004), the networked sensor tapestry, is a general ar- chitecture to manage the broad range of distributed surveillance applications. It is scalable and is composed of software modules and hardware components. The core component of a NeST server is the privacy buffer. It utilises programmable plug-in privacy filters which operate on data coming from sensors in order to prevent access to it or to remove personal identifiable information. These pri- vacy filters are specified using a privacy grammar that is able of connecting multiple low-level privacy filters to create arbitrary data-dependent privacy def- initions. Moreover, the privacy of the monitored subjects is also considered. In this sense, the NeST architecture integrates individuals’ privacy (according to behaviours) denying access to specific information to some or all of the modules or operators. NeST features secure sharing, capturing, distributed processing and archiving of surveillance data. Stealth Vision (Kitahara et al., 2004) is an anonymous video capturing sys- tem that protects the privacy of objects by pixelating or blurring their appear- ance. It determines private areas of captured videos from a mobile camera using 3D space models. The 3D position of the target object is estimated by a ho- mographic transformation using images coming from a static overhead camera. Afterwards, a calibrated mobile camera estimates the privacy area by project- 26 ing the 3D models onto the captured image plane. Finally, the privacy area is protected. RFID sensors are employed in order to indicate which persons or objects are allowed to be captured. PriSurv (Chinomi et al., 2008) is a system that uses visual abstract repre- sentations to protect individual’s privacy. It is composed of several modules: analyser, profile generator, profile base, access controller, abstractor and video database. In order to identify subjects, RFID tags and image processing is used. This system is focused on small communities where a certain number of members are registered and they monitor each other. Given that the sense of privacy is different for everyone, privacy policies are determined according to the closeness between subjects in the video and viewers monitoring. Hence, the appearance of subjects is opened to close viewers while it is hidden to the viewers that subjects feel distant from. Depending on this closeness, different abstract representations are chosen. Respectful Cameras (Schiff et al., 2009) is a prototype system focused on addressing privacy concerns. This system works by detecting coloured markers in real time that people wear such as hats or vests in order to indicate their privacy preferences. This are expressed as their will of remaining anonymous. The system obscures the face of a person when a marker has been detected. CareLog and BufferWare (Hayes & Truong, 2009) are two systems based on the concepts of the selective archiving model. Under this model, data is con- stantly buffered but an explicit input is required in order to archive it, otherwise data is lost. It represents a compromise whereby people can negotiate their own policies concerning control, privacy, information access and comfort. CareLog is a system aimed at education classrooms for the recording of diagnostic be- havioural data of children with severe behaviour disorders. Using such a system, a teacher takes control of data archiving to document an incident after it has occurred. Regarding BufferWare, it is a system aimed at semi-public spaces in which people can save recorded data, if desired, using a touch-screen. They are also able of viewing previous recordings. In this case, BufferWare was placed in a social area of an academic building. These two systems were proposed in order to conduct an experiment to assess issues of the selective archiving model, such as ownership of data, choice of which data should be saved or deleted, vis- ibility and awareness of recordings, and trust in the fact that policies are being followed and features were implemented as described. Altcare (Shoaib et al., 2010) is a monitoring system based on a static net- work of video cameras aimed for emergency detection at home. It focuses on fall detection, and works without any manual initialisation or interaction with older persons. When a fall is detected, Altcare first attempts to get the confirmation from the involved person. Afterwards, if the verification is positive, it auto- matically communicates the emergency to the responsible person. In addition to information concerning the involved person, the system transmits a patch of the video that shows the emergency. In order to protect the privacy, the person is replaced by the silhouette in the video. Furthermore, Altcare enables system administrators to check the state of the person at any time, if so desired by the monitored person. 27 In Sohn et al. (2010), a privacy-preserving watch list screening system for video-surveillance is proposed. Depending on the group of interest the watch list comprise either blacklisted or white-listed identities. Hence, the goal is to verify whether a person is enrolled in the watch list or not. In order to preserve privacy it relies on homomorphic encryption. This system is composed of a network of distributed cameras and a central management server (CMS) which owns the watch list database. This database is considered private and it contains face feature vectors corresponding to identities of interest. In this system, cameras analyse the recorded raw video in order to detect face regions. When a face is detected, it is encrypted and sent to the CMS where face feature vectors are retrieved from the watch list database. Afterwards, a comparison between the face coming from a camera and those from the watch list is performed in the encryption domain. Then a result confirming whether the identity is included in the watch list or not is encrypted and sent back to the camera. Finally, according to the obtained result the camera notifies it to the security service. TrustCAM (Winkler & Rinner, 2010b,a) is a smart camera that provides se- curity and privacy-protection based on trusted computing. By using this smart camera video bitstreams are digitally signed. Thereby, the manipulation of recorded images or videos can be detected, origin of images are provided, and multi-level access control is supported. Each TrustCAM node is operated by a central control station (CS) which runs a protected database where crypto- graphic keys generated during camera setup are securely stored. Camera nodes encrypt privacy sensitive image data such as faces or number plates using a spe- cial cryptographic key stored in each one. Moreover, an abstracted version of the sensitive regions as, for example, a silhouette can also be generated and en- crypted. These encrypted sensitive regions cannot be decrypted without having access to the CS. Indeed, each representation requires a different key in order to decrypt it, thereby providing multi-level access control. Finally, PAAS (Barhm et al., 2011) is a privacy-aware surveillance system that enables monitored subjects to specify their privacy preferences via gestures. It uses face recognition technology in order to map individuals to their privacy preferences and security credentials that are encoded using an extension of the P3P-APPEL framework (Cranor et al., 2002). Users have access to one of three privacy settings, namely no privacy, blurred face and blurred full body. 5. Discussion 5.1. Recognition of the region of interest As it has been seen throughout this survey, privacy protection methods mainly focus on preserving visual privacy in images and videos, either modifying completely the whole image or only a region of interest. Although the detection of the region of interest has not been specifically considered in this work, correct detection is fundamental in order to make protection methods work as desired. For instance, it could be imagined a hypothetical case in which a scrambling algorithm was used to obfuscate a person’s face in a video. Supposing that 28 the video sampling rate is 25 fps, there are 3 000 frames in 2 minutes of video where a face appears. It could be supposed also that the hit rate of the used face detector was around 98%. Under these assumptions, the subject face will be poorly detected in roughly 60 frames of video, thereby not enabling the obscuring algorithm to work properly in protecting privacy. If the face is fully observable in a single frame, subject identification can already be performed. Not only that, the face could also be inferred observing multiple frames in which the face detector failed. Hence, it is essential that the computer vision techniques (segmentation, object detection, tracking, recognition, etc.) involved before the obscuring process are reliable and robust enough in order to guarantee the effectiveness of the protection method. 5.2. Privacy and intelligibility trade-off Although we have mentioned the privacy-intelligibility trade-off in the intro- duction of the section 3.4, let us provide a further discussion here. Zhao & Stasko (1998) compared several image filters (blurring, pixelating, edge-detector, live- shadow and shadow-view) and found out that whereas intelligibility is provided almost in all of the tested filters, actors were also recognised in most of the cases. Boyle et al. (2000) performed an in depth evaluation of blurring and pixelating to analyse how the variation of the filtration level affects to this balance. The obtained results suggested that blurring and, to a lesser extent, pixelating may provide a balance, but it would be a precarious balance at best. Furthermore, Neustaedter et al. (2006) state that blurring by itself does not suffice for privacy protection when a balance is needed. Therefore, it seems that image filters can hardly provide a balance between privacy and information utility, and generally privacy loses in this balance. Nevertheless, when such a balance is not needed, image filters may be used. As reported by Agrawal & Narayanan (2011), blur- ring is enough to hide the identity if gait is not involved. But gait and other temporal characteristics are difficult to hide if there is some familiarity between the subject and the user. Anyway, it would be interesting to expand these stud- ies to also include visual abstraction methods and all those where such a balance could be analysed. 5.3. Real-time applications Regarding the techniques that could be used in real-time intelligent moni- toring systems, apart from image filtering, the remainder techniques would be impracticable because they are very time consuming. In such systems, all the involved computer vision algorithms along with the protection method must jointly work in real time. Like inpainting techniques, encryption ones have high computational requirements. Nevertheless, some light-weight encryption techniques may work properly. In turn, some image inpainting techniques are designed to work in real time too. The election of which protection method to use depends on the application requirements. 29 5.4. Privacy model and evaluation Relying on a formal model to guarantee privacy preservation is required in order to have a common framework in which privacy solutions can be devel- oped and evaluated. Common definitions about what visual privacy should be, when privacy is protected, which are the sensitive areas, and so on are needed. A model that currently fulfils some of these requirements is the face de- identification based on the k-same family of algorithms. Although it guarantees that a de-identified face cannot be recognised with a probability higher than 1 k , inference channels are not considered. In that sense, Saini et al. (2014) propose a model that takes this into account. This model measures the privacy loss of the individuals that usually inhabitant a surveillance area. This measure is given as a continuous variable in the range [0, 1]. As far as we know, this is the only approach that currently measures privacy loss in such a way. Regarding the remainder methods, despite they are not based on a formal model, it can be intuitively realised that they provide different protection levels. For instance, näıve image filtering techniques like blurring or pixelating are not sufficient to protect privacy when intelligibility is involved. But blanking out or encrypting sensitive regions could be enough. Also related to the previous point is how the evaluation of visual privacy pro- tection is performed. We have not found a common framework for it. Although the aforementioned model for measuring privacy loss is very valuable, it cannot be used to build such a framework that automatically performs the comparison because current technology is not robust enough. This would have been very useful in order to perform an exhaustive comparison covering all of the reviewed methods. Moreover, the lack of more manually-labelled privacy datasets also contributes to deteriorate visual privacy evaluation. The only datasets focused on privacy that have been found in the literature are PEViD-HD (Korshunov & Ebrahimi, 2013) and PEViD-UHD (Korshunov & Ebrahimi, 2014). Both datasets consider video surveillance scenarios and provide high definition video sequences and ultra high definition ones, respectively. A similar dataset for ambient-assisted living would be appreciated where, in addition, other kind of visual sensors were included (e.g. RGB-D sensors). In addition to assess the grade in which privacy is protected, to what extent useful information is retained should also be studied. In the absence of such evaluation mechanisms, several works have been found in which some of the protection methods have been tested by users as part of the experimentation. These experiments are mainly based on conducting interviews and questionnaires with users where their skills in extracting useful information from a privacy-protected image (subject identification, pose, activity, presence, etc.) are evaluated (Korshunov et al., 2012). Results obtained from an objective study are missed. 5.5. User studies about privacy requirements Concerning users, there is no agreement about privacy requirements be- cause privacy is highly subjective and which information is considered sensitive 30 depends on each individual. Despite of what has been previously said about blurring and pixelating not being effective in providing a balance between pri- vacy and intelligibility, some users that have participated in studies about this matter feel satisfied with these filters when they are configured correctly (Zhao & Stasko, 1998; Boyle et al., 2000; Lander et al., 2001; Neustaedter et al., 2006). Other participants felt that there is not such a balance because as they were able to know what subjects were doing, privacy was not being suitably protected. Indeed, while some of the participants showed interest in using video cameras at work and even at their homes imposing some constraints, others commented that they would not use them because video cameras are very intrusive. Any- way, what can be extracted about this lack of consensus is that privacy is a very subjective topic and user’s requirements vary. Considering which protection methods preserve privacy better for a fall de- tection system, a user survey was conducted by Edgcomb & Vahid (2012). They analysed several visual representations: blur, silhouette, oval, box, and trailing arrows. Results indicated that silhouettes and blur were perceived to provide insufficient privacy, whereas an oval provides sufficient perceived privacy while still supporting fall detection accuracy of 89%. Others have also researched about possible features of subjects’ sense of security and privacy for video surveillance systems (Koshimizu et al., 2006; Babaguchi et al., 2009). Results showed how subjects classify viewers that monitor them using cameras in: familiar persons, unfamiliar persons and per- sons in duty. Moreover, subjects expect a very familiar person to protect them in an emergency. Concerning the disclosure of private information, it is af- fected by the closeness between subjects and viewers. Familiarity or closeness between persons facilitates the recognition of each other. It is also curious to see how people in their 50s are less sensitive to privacy than those in their 20- 40s. Regarding older people, most of them agree with using silhouettes as a visual representation in order to protect their privacy. Furthermore, they show interest in customising the system operation. They demand control over the camera and who has access to the captured or stored information (e.g. turning a camera off, setting up the filtration level, seeing how they are being watched, and so on). However, there are others who consider that they do not require a monitoring system because they are independent enough (Demiris et al., 2009). Finally, it is necessary that users demand protection methods in order to get visual privacy protection techniques widely used in video surveillance systems. Without such a demand, main actors of the security market, more interested in making a profit, will not pay attention to privacy issues because they do not have enough motivation to kick-start self-regulation (Gutwirth et al., 2012). We consider that a strong demand of privacy-aware systems will increase research on this field also benefiting to future AAL systems. Either way, more discussion is needed in visual privacy, privacy requirements of users according to applications need to be collected, and more research on visual privacy protection is required. 31 6. Conclusion As we have seen throughout this work, the proliferation of networks of video cameras for surveillance in public spaces and the advances in computer vision have led to a situation in which the privacy of individuals is being compro- mised. Moreover, nowadays video cameras are being used more often in ambient- assisted living applications that operate in private spaces. Because of this, ex- pert and intelligent systems that handle these tasks should take privacy into account by means of new tools that restore the individual’s right to privacy. In this paper, an up-to-date review of visual privacy protection methods has been provided. As far as we know, this is the first review that focuses on the protection methods and tries to give a comprehensive classification of all of them. In our classification, we have considered the following categories: intervention, blind vision, secure processing, redaction and data hiding. As it has been seen, redaction methods cover the largest part of this work and they have also been classified according to the way in which imagery data is modified: image filtering, encryption, face de-identification, object removal and visual abstraction. Given that the previous methods are required for expert and intelligent sys- tems for video surveillance and ambient-assisted living, some of them have also been reviewed. Such systems are developed as a framework to provide some kind of privacy protection. However, only using some of the described meth- ods is not enough to build a privacy-aware system. As it has been seen, they have to support also other mechanisms to implement privacy policies. These policies specify what data is protected, how it is protected, who has access, and a variety of other fundamental aspects concerning involved subjects, ongoing activities, captured data, and so on. Despite of visual privacy protection being necessary in expert and intelligent systems for video surveillance, due to its in- creasing power of tracking, storing and indexing of our public daily activities, the use of privacy protection methods can also be justified for such systems but from an ambient-assisted living perspective. In that sense, privacy-aware smart cameras are essential in the construction of future low-cost systems for telecare applications aimed at older and disabled people. As another contribution of this work, we have also provided a discussion about important factors and current limitations of visual privacy protection. For instance, a key aspect, i.e. the correct detection of the region of interest, has been highlighted. It is a very relevant topic because it can be considered as the first building block of privacy protection and, therefore, it would deserve a self-standing review of the related research fields. We have also mentioned the problem faced by redaction methods, i.e. the privacy and intelligibility trade-off, and the necessity of expanding these studies to cover other protection methods. Furthermore, we have stood out the lack of a formal model and evaluation mechanisms that would enable to make fair comparisons between different protection methods and, more importantly, it would provide guarantees about their accuracy in protecting privacy. This work provides a comprehensive picture of how to protect visual privacy, 32 so it can be a good starting point for novel researchers in the field of expert and intelligent systems, and more specifically in intelligent monitoring systems for video surveillance and ambient-assisted living. For future research directions we consider that there are two main axis that need special attention. On the one hand, as aforementioned, a standard way to quantify and evaluate visual privacy protection is needed so as to fairly compare works in this matter. It would be better if this is included as a part of a formal model that considers more privacy related aspects. On the other hand, recognition accuracy of sensitive regions needs to be improved, so research is required in more robust computer vision algorithms for recognition, tracking and event detection. Regarding other future research directions, protection methods would also deserve some consideration. Although a lot of work has been done so far, novel redaction methods that provide a better balance between privacy and intelligi- bility are welcome. For instance, a textual representation of the scene, where the essential context (e.g. individuals, events, etc.) is captured, could be enough for some applications like telecare. Privacy preferences are also very important. Mechanisms to empower users and put them in control of their private data captured by intelligent systems should be researched. It would be appreciated to have a common way to let users specify their preferences so they can decide who, how, where and when they are watched in normal scenarios where no law infringement is involved. Finally, users should be able of knowing and track- ing which systems have collected data about them in order to enable them to take legal actions in this matter just in case they need it. Therefore, work in these areas may lead to obtain more effective and reliable visual privacy protec- tion methods and systems in a near future and also helps to increase the users’ acceptance of using video cameras in private spaces. Acknowledgement This work has been partially supported by the Spanish Ministry of Science and Innovation under project “Sistema de visión para la monitorización de la actividad de la vida diaria en el hogar” (TIN2010-20510-C04-02) and by the European Commission under project “caring4U - A study on people activity in private spaces: towards a multisensor network that meets privacy require- ments” (PIEF-GA-2010-274649). José Ramón Padilla López and Alexandros Andre Chaaraoui acknowledge financial support by the Conselleria d’Educació, Formació i Ocupació of the Generalitat Valenciana (fellowship ACIF/2012/064 and ACIF/2011/160 respectively). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. This article was originally published in José Ramón Padilla-López, Alexan- dros Andre Chaaraoui, Francisco Flórez-Revuelta, Visual Privacy Protection Methods: A Survey, Expert Systems with Applications, ISSN: 0957-4174, http: //dx.doi.org/10.1016/j.eswa.2015.01.041. 33 http://dx.doi.org/10.1016/j.eswa.2015.01.041 http://dx.doi.org/10.1016/j.eswa.2015.01.041 References Abraham, A. R., Prabhavathy, K. A., & Shree, D. J. (2012). A survey on video inpainting. International Journal of Computer Applications , 56 , 43–47. Adams, A. A., & Ferryman, J. M. (2013). The future of video analytics for surveillance and its ethical implications. Security Journal , . doi:10.1057/sj. 2012.48. Agi, I., & Gong, L. (1996). An empirical study of secure MPEG video trans- missions. In Network and Distributed System Security, Proceedings of the Symposium on (pp. 137–144). IEEE. Agrawal, P., & Narayanan, P. (2011). Person de-identification in videos. Circuits and Systems for Video Technology, IEEE Transactions on, 21 , 299–310. Augusto, J., Nakashima, H., & Aghajan, H. (2010). Ambient intelligence and smart environments: A state of the art. In H. Nakashima, H. Aghajan, & J. Augusto (Eds.), Handbook of Ambient Intelligence and Smart Environments (pp. 3–31). Springer US. Avidan, S., & Butman, M. (2006). Blind vision. In A. Leonardis, H. Bischof, & A. Pinz (Eds.), Computer Vision - ECCV 2006 (pp. 1–13). Springer Berlin / Heidelberg volume 3953 of Lecture Notes in Computer Science. Avidan, S., Elbaz, A., Malkin, T., & Moriarty, R. (2009). Oblivious image matching. In A. Senior (Ed.), Protecting Privacy in Video Surveillance (pp. 49–64). Springer London. Baaziz, N., Lolo, N., Padilla, O., & Petngang, F. (2007). Security and privacy protection for automated video surveillance. In Signal Processing and Infor- mation Technology, 2007 IEEE International Symposium on (pp. 17–22). Babaguchi, N., Koshimizu, T., Umata, I., & Toriyama, T. (2009). Psychological study for designing privacy protected video surveillance system: PriSurv. In A. Senior (Ed.), Protecting Privacy in Video Surveillance (pp. 147–164). Springer London. Baran, I., & Popović, J. (2007). Automatic rigging and animation of 3D charac- ters. Graphics, ACM Transactions on, 26 . doi:10.1145/1276377.1276467. Barhm, M., Qwasmi, N., Qureshi, F., & el Khatib, K. (2011). Negotiating pri- vacy preferences in video surveillance systems. In K. Mehrotra, C. Mohan, J. Oh, P. Varshney, & M. Ali (Eds.), Modern Approaches in Applied Intelli- gence (pp. 511–521). Springer Berlin Heidelberg volume 6704 of Lecture Notes in Computer Science. Bertalmio, M., Sapiro, G., Caselles, V., & Ballester, C. (2000). Image inpainting. In Computer Graphics and Interactive Techniques, Proceedings of the 27th Annual Conference on SIGGRAPH ’00 (pp. 417–424). New York, NY, USA: ACM Press/Addison-Wesley Publishing Co. 34 http://dx.doi.org/10.1057/sj.2012.48 http://dx.doi.org/10.1057/sj.2012.48 http://dx.doi.org/10.1145/1276377.1276467 Bitouk, D., Kumar, N., Dhillon, S., Belhumeur, P., & Nayar, S. K. (2008). Face swapping: automatically replacing faces in photographs. Graphics, ACM Transactions on, 27 , 39:1–39:8. Borosán, P., Jin, M., DeCarlo, D., Gingold, Y., & Nealen, A. (2012). RigMesh: Automatic rigging for part-based shape modeling and deformation. Graphics, ACM Transactions on, 31 , 198:1–198:9. Boult, T. (2005). PICO: Privacy through invertible cryptographic obscuration. In Computer Vision for Interactive and Intelligent Environment, 2005 (pp. 27–38). IEEE. Boyle, M., Edwards, C., & Greenberg, S. (2000). The effects of filtered video on awareness and privacy. In Computer supported cooperative work, Proceedings of the 2000 ACM conference on CSCW ’00 (pp. 1–10). New York, NY, USA: ACM. Caloyannides, M. A. (2003). Society cannot function without privacy. IEEE Security and Privacy , 1 , 84–86. Candes, E., Romberg, J., & Tao, T. (2006). Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information. Information Theory, IEEE Transactions on, 52 , 489–509. Cardinaux, F., Bhowmik, D., Abhayaratne, C., & Hawley, M. S. (2011). Video based technology for ambient assisted living: A review of the literature. Jour- nal of Ambient Intelligence and Smart Environments , 3 , 253–269. Carrillo, P., Kalva, H., & Magliveras, S. (2010). Compression independent reversible encryption for privacy in video surveillance. EURASIP Journal on Information Security , 2009 , 1–13. doi:10.1155/2009/429581. Article ID: 2009:429581. Cavoukian, A. (2013). Surveillance, Then and Now: Securing Privacy in Public Spaces . Technical Report Information and Privacy Commmissioner of On- tario. Chaaraoui, A. A., Climent-Pérez, P., & Flórez-Revuelta, F. (2012a). An efficient approach for multi-view human action recognition based on bag-of-key-poses. In A. A. Salah, J. Ruiz-del Solar, c. Meriçli, & P.-Y. Oudeyer (Eds.), Human Behavior Understanding (pp. 29–40). Springer Berlin Heidelberg volume 7559 of Lecture Notes in Computer Science. Chaaraoui, A. A., Climent-Pérez, P., & Flórez-Revuelta, F. (2012b). A review on vision techniques applied to human behaviour analysis for ambient-assisted living. Expert Systems with Applications , 39 , 10873–10888. Chan, C. S., & Liu, H. (2009). GMM-QNT hybrid framework for vision-based human motion analysis. In Fuzzy Systems, 2009. FUZZ-IEEE 2009. IEEE International Conference on (pp. 1820–1825). 35 http://dx.doi.org/10.1155/2009/429581 Chan, T., Shen, J., & Zhou, H.-M. (2006). Total variation wavelet inpainting. Journal of Mathematical Imaging and Vision, 25 , 107–125. Chang, Y., Yan, R., Chen, D., & Yang, J. (2006). People identification with limited labels in privacy-protected video. In Multimedia and Expo, 2006 IEEE International Conference on (pp. 1005–1008). Chattopadhyay, A., & Boult, T. (2007). PrivacyCam: a privacy preserving camera using uCLinux on the blackfin DSP. In Computer Vision and Pattern Recognition, 2007. CVPR ’07. IEEE Conference on (pp. 1–8). Chen, D., Chang, Y., Yan, R., & Yang, J. (2007). Tools for protecting the pri- vacy of specific individuals in video. EURASIP Journal on Advances in Signal Processing , 2007 , 1–9. doi:10.1155/2007/75427. Article ID: 2007:075427. Chen, D., Chang, Y., Yan, R., & Yang, J. (2009). Protecting personal identifi- cation in video. In A. Senior (Ed.), Protecting Privacy in Video Surveillance (pp. 115–128). Springer London. Chen, Y., Dang, G., Cheng, Z., & Xu, K. (2014). Fast capture of personalized avatar using two kinects. Journal of Manufacturing Systems , 33 , 233–240. Chen, Y.-Y., Cho, C.-W., Lin, S.-H., Lai, H.-Y., Lo, Y.-C., Chen, S.-Y., Chang, Y.-J., Huang, W.-T., Chen, C.-H., Jaw, F.-S., Tsang, S., & Tsai, S.-T. (2012). A vision-based regression model to evaluate parkinsonian gait from monocular image sequences. Expert Systems with Applications , 39 , 520–526. Cheung, S.-c., Paruchuri, J., & Nguyen, T. (2008). Managing privacy data in pervasive camera networks. In Image Processing, 2008. ICIP 2008. 15th IEEE International Conference on (pp. 1676–1679). Cheung, S.-C., Venkatesh, M., Paruchuri, J., Zhao, J., & Nguyen, T. (2009). Protecting and managing privacy information in video surveillance systems. In A. Senior (Ed.), Protecting Privacy in Video Surveillance (pp. 11–33). Springer London. Cheung, S.-C. S., Zhao, J., & Venkatesh, M. (2006). Efficient object-based video inpainting. In Image Processing, 2006 IEEE International Conference on (pp. 705–708). Chinomi, K., Nitta, N., Ito, Y., & Babaguchi, N. (2008). PriSurv: Privacy protected video surveillance system using adaptive visual abstraction. In S. Satoh, F. Nack, & M. Etoh (Eds.), Advances in Multimedia Modeling (pp. 144–154). Springer Berlin / Heidelberg volume 4903 of Lecture Notes in Computer Science. Cook, D. J., Augusto, J. C., & Jakkula, V. R. (2009). Ambient intelligence: Technologies, applications, and opportunities. Pervasive and Mobile Com- puting , 5 , 277–298. 36 http://dx.doi.org/10.1155/2007/75427 Cootes, T., Edwards, G., & Taylor, C. (1998). Active appearance models. In H. Burkhardt, & B. Neumann (Eds.), Computer Vision – ECCV’98 (pp. 484– 498). Springer Berlin Heidelberg volume 1407 of Lecture Notes in Computer Science. Cootes, T., Edwards, G., & Taylor, C. (2001). Active appearance models. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 23 , 681– 685. Coudert, F. (2010). When video cameras watch and screen: Privacy implications of pattern recognition technologies. Computer Law & Security Review , 26 , 377–384. Coutaz, J., Bérard, F., Carraux, E., Astier, W., & Crowley, J. L. (1999). CoMedi: Using computer vision to support awareness and privacy in me- diaspaces. In CHI ’99 Extended Abstracts on Human Factors in Computing Systems CHI EA ’99 (pp. 13–14). New York, NY, USA: ACM. Cox, I., Honsinger, C., Miller, M., & Bloom, J. (2002). Digital watermarking. Journal of Electronic Imaging , 11 , 414–414. Cranor, L., Langheinrich, M., & Marchiori, M. (2002). A P3P prefer- ence exchange language 1.0 (APPEL1.0). URL: http://www.w3.org/TR/ P3P-preferences/ last visited: 2015/01/08. Criminisi, A., Perez, P., & Toyama, K. (2003). Object removal by exemplar- based inpainting. In Computer Vision and Pattern Recognition, Proceedings of IEEE Computer Society Conference on (pp. II–721–II–728 vol.2). volume 2. Criminisi, A., Perez, P., & Toyama, K. (2004). Region filling and object removal by exemplar-based image inpainting. Image Processing, IEEE Transactions on, 13 , 1200 –1212. Dee, H., & Velastin, S. (2008). How close are we to solving the problem of automated visual surveillance? Machine Vision and Applications , 19 , 329– 343. Demiris, G., Oliver, D., Giger, J., Skubic, M., & Rantz, M. (2009). Older adults’ privacy considerations for vision based recognition methods of eldercare ap- plications. Technology and Health Care, 17 , 41–48. Donoho, D. (2006). Compressed sensing. Information Theory, IEEE Transac- tions on, 52 , 1289–1306. Dufaux, F. (2011). Video scrambling for privacy protection in video surveil- lance: recent results and validation framework. In Mobile Multimedia/Image Processing, Security, and Applications 2011, Proceedings of SPIE (pp. 1–12). volume 8063. doi:10.1117/12.883948 Article ID: 806302. 37 http://www.w3.org/TR/P3P-preferences/ http://www.w3.org/TR/P3P-preferences/ http://dx.doi.org/10.1117/12.883948 Dufaux, F., & Ebrahimi, T. (2006). Scrambling for video surveillance with pri- vacy. In Computer Vision and Pattern Recognition Workshop, 2006. CVPRW ’06. Conference on (p. 160). doi:10.1109/CVPRW.2006.184. Dufaux, F., & Ebrahimi, T. (2008a). H.264/AVC video scrambling for privacy protection. In Image Processing, 2008. ICIP 2008. 15th IEEE International Conference on (pp. 1688–1691). Dufaux, F., & Ebrahimi, T. (2008b). Scrambling for privacy protection in video surveillance systems. Circuits and Systems for Video Technology, IEEE Transactions on, 18 , 1168–1174. Dufaux, F., & Ebrahimi, T. (2010). A framework for the validation of privacy protection solutions in video surveillance. In Multimedia and Expo (ICME), 2010 IEEE International Conference on (pp. 66–71). Dufaux, F., Ouaret, M., Abdeljaoued, Y., Navarro, A., Vergnenègre, F., & Ebrahimi, T. (2006). Privacy enabling technology for video surveillance. In Mobile Multimedia/ Image Processing for Military and Security Applications, Proceedings of SPIE (pp. 1–12). volume 6250. doi:10.1117/12.664945 Article ID: 62500M. EC (2008). Seniorwatch 2. Assessment of the Senior Market for ICT Progress and Developments . Technical Report European Commission. EC (2010). Overview of the European strategy in ICT for Ageing Well . Technical Report European Commission. Edgcomb, A., & Vahid, F. (2012). Privacy perception and fall detection accuracy for in-home video assistive monitoring with privacy enhancements. ACM SIGHIT Record , 2 , 6–15. Efros, A., & Leung, T. (1999). Texture synthesis by non-parametric sampling. In Computer Vision, 1999. The Proceedings of the Seventh IEEE International Conference on (pp. 1033–1038). volume 2. Efros, A. A., & Freeman, W. T. (2001). Image quilting for texture synthesis and transfer. In Computer Graphics and Interactive Techniques, Proceedings of the 28th Annual Conference on SIGGRAPH ’01 (pp. 341–346). New York, NY, USA: ACM. Erkin, Z., Franz, M., Guajardo, J., Katzenbeisser, S., Lagendijk, I., & Toft, T. (2009). Privacy-preserving face recognition. In I. Goldberg, & M. Atallah (Eds.), Privacy Enhancing Technologies (pp. 235–253). Springer Berlin / Heidelberg volume 5672 of Lecture Notes in Computer Science. Erturk, S. (2007). Multiplication-free one-bit transform for low-complexity block-based motion estimation. Signal Processing Letters, IEEE , 14 , 109– 112. 38 http://dx.doi.org/10.1109/CVPRW.2006.184 http://dx.doi.org/10.1117/12.664945 Fan, J., Luo, H., Hacid, M.-S., & Bertino, E. (2005). A novel approach for privacy-preserving video sharing. In Information and knowledge management, Proceedings of the 14th ACM international conference on CIKM ’05 (pp. 609– 616). New York, NY, USA: ACM. Fidaleo, D. A., Nguyen, H.-A., & Trivedi, M. (2004). The networked sen- sor tapestry (NeST): a privacy enhanced software architecture for interactive analysis of data in video-sensor networks. In Video surveillance & sensor networks, Proceedings of the ACM 2nd international workshop on VSSN ’04 (pp. 46–53). New York, NY, USA: ACM. Fleck, S., & Strasser, W. (2008). Smart camera based monitoring system and its application to assisted living. Proceedings of the IEEE , 96 , 1698–1714. Frome, A., Cheung, G., Abdulkader, A., Zennaro, M., Wu, B., Bissacco, A., Adam, H., Neven, H., & Vincent, L. (2009). Large-scale privacy protection in google street view. In Computer Vision, 2009 IEEE 12th International Conference on (pp. 2373–2380). Gerrard, G., & Thompson, R. (2011). Two million cameras in the UK. CCTV Image, 42 , 10–12. Ghanbari, A., & Soryani, M. (2011). Contour-based video inpainting. In Ma- chine Vision and Image Processing (MVIP), 2011 7th Iranian (pp. 1–5). IEEE. Goessl, L. (2012). New japanese security camera scans 36 million faces per second. URL: http://digitaljournal.com/article/321848 last visited: 2014/05/08. Granados, M., Tompkin, J., Kim, K., Grau, O., Kautz, J., & Theobalt, C. (2012). How not to be seen – object removal from videos of crowded scenes. Computer Graphics Forum, 31 , 219–228. Gross, R., Airoldi, E., Malin, B., & Sweeney, L. (2006a). Integrating utility into face de-identification. In G. Danezis, & D. Martin (Eds.), Privacy Enhancing Technologies (pp. 227–242). Springer Berlin / Heidelberg volume 3856 of Lecture Notes in Computer Science. Gross, R., Sweeney, L., Cohn, J., Torre, F., & Baker, S. (2009). Face de- identification. In A. Senior (Ed.), Protecting Privacy in Video Surveillance (pp. 129–146). Springer London. Gross, R., Sweeney, L., de la Torre, F., & Baker, S. (2006b). Model-based face de-identification. In Computer Vision and Pattern Recognition Workshop, 2006. CVPRW ’06. Conference on (pp. 161–168). Guangzhen, L., Yoshimichi, I., Xiaoyi, Y., Naoko, N., & Noboru, B. (2010). Re- coverable privacy protection for video content distribution. EURASIP Journal on Information Security , 2009 , 1–11. doi:10.1155/2009/293031. Article ID: 2009:293031. 39 http://digitaljournal.com/article/321848 http://dx.doi.org/10.1155/2009/293031 Gutwirth, S., Leenes, R., De Hert, P., & Poullet, Y. (2012). European Data Protection: In Good Health? . Springer. Harvey, A. (2010). Camoflash - anti-paparazzi clutch. URL: http:// ahprojects.com/projects/camoflash/ last visited: 2014/05/08. Hayes, G., & Truong, K. (2009). Selective archiving: A model for privacy sen- sitive capture and access technologies. In A. Senior (Ed.), Protecting Privacy in Video Surveillance (pp. 165–184). Springer London. He, L., Bleyer, M., & Gelautz, M. (2011). Object removal by depth- guided inpainting. In The Austrian Association for Pattern Recognition (OAGM/AAPR) Workshop 2011 (OAGM2011) (pp. 1–8). Hodgins, J. K., O’Brien, J. F., & Tumblin, J. (1998). Perception of human mo- tion with different geometric models. Visualization and Computer Graphics, IEEE Transactions on, 4 , 307–316. Hogue, A., Gill, S., & Jenkin, M. (2007). Automated avatar creation for 3D games. In Future Play, Proceedings of the 2007 Conference on Future Play ’07 (pp. 174–180). New York, NY, USA: ACM. Hudson, S. E., & Smith, I. (1996). Techniques for addressing fundamental pri- vacy and disruption tradeoffs in awareness support systems. In Computer sup- ported cooperative work, Proceedings of the 1996 ACM conference on CSCW ’96 (pp. 248–257). New York, NY, USA: ACM. De la Hunty, M., Asthana, A., & Goecke, R. (2010). Linear facial expression transfer with active appearance models. In Pattern Recognition (ICPR), 2010 20th International Conference on (pp. 3789–3792). Igehy, H., & Pereira, L. (1997). Image replacement through texture synthesis. In Image Processing, Proceedings of the International Conference on (pp. 186–189). volume 3. Ito, I., & Kiya, H. (2009). One-time key based phase scrambling for phase- only correlation between visually protected images. EURASIP Journal on Information Security , 2009 , 1–11. doi:10.1155/2009/841045. Article ID: 2009:841045. Johansson, G. (1973). Visual perception of biological motion and a model for its analysis. Perception & Psychophysics , 14 , 201–211. Kitahara, I., Kogure, K., & Hagita, N. (2004). Stealth vision for protecting privacy. In Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference on (pp. 404–407). volume 4. Kokaram, A., Morris, R., Fitzgerald, W., & Rayner, P. (1995). Interpolation of missing data in image sequences. Image Processing, IEEE Transactions on, 4 , 1509–1519. 40 http://ahprojects.com/projects/camoflash/ http://ahprojects.com/projects/camoflash/ http://dx.doi.org/10.1155/2009/841045 Koochari, A., & Soryani, M. (2010). Exemplar-based video inpainting with large patches. Journal of Zhejiang University SCIENCE C , 11 , 270–277. Korshunov, P., Araimo, C., De Simone, F., Velardo, C., Dugelay, J., & Ebrahimi, T. (2012). Subjective study of privacy filters in video surveil- lance. In Multimedia Signal Processing (MMSP), 2012 IEEE 14th Interna- tional Workshop on (pp. 378–382). Korshunov, P., & Ebrahimi, T. (2013). PEViD: privacy evaluation video dataset. In Applications of Digital Image Processing XXXVI, Proceedings of SPIE (pp. 1–9). volume 8856. doi:10.1117/12.2030974 Article ID: 88561S. Korshunov, P., & Ebrahimi, T. (2014). UHD video dataset for evaluation of privacy. In Sixth International Workshop on Quality of Multimedia Experience (QoMEX 2014). Koshimizu, T., Toriyama, T., & Babaguchi, N. (2006). Factors on the sense of privacy in video surveillance. In Continuous archival and retrieval of personal experiences, Proceedings of the 3rd ACM workshop on CARPE ’06 (pp. 35– 44). New York, NY, USA: ACM. Lander, K., Bruce, V., & Hill, H. (2001). Evaluating the effectiveness of pixela- tion and blurring on masking the identity of familiar faces. Applied Cognitive Psychology , 15 , 101–116. Langheinrich, M. (2001). Privacy by design - principles of privacy-aware ubiqui- tous systems. In Ubiquitous Computing, Proceedings of the 3rd international conference on UbiComp ’01 (pp. 273–291). Springer-Verlag. Li, S., Zhao, Y., Qu, B., & Wang, J. (2013). Image scrambling based on chaotic sequences and veginère cipher. Multimedia Tools and Applications , 66 , 573– 588. Lin, Y., Wang, S., Lin, Q., & Tang, F. (2012). Face swapping under large pose variations: A 3D model based approach. In Multimedia and Expo (ICME), 2012 IEEE International Conference on (pp. 333–338). Lyons, M., Plante, A., Jehan, S., Inoue, S., & Akamatsu, S. (1998). Avatar creation using automatic face processing. In Multimedia, Proceedings of the sixth ACM international conference on (pp. 427–434). ACM. Ma, W., & Ma, Q. (2011). An effective method for removing large object. In Image and Signal Processing (CISP), 2011 4th International Congress on (pp. 80–84). volume 1. Macq, B., & Quisquater, J.-J. (1995). Cryptology for digital tv broadcasting. Proceedings of the IEEE , 83 , 944–957. 41 http://dx.doi.org/10.1117/12.2030974 Mart́ınez-Ponte, I., Desurmont, X., Meessen, J., & Delaigle, J.-F. (2005). Ro- bust human face hiding ensuring privacy. In Image Analysis for Multimedia Interactive Services (WIAMIS), Proceedings of the International Workshop on. volume 4. Masnou, S., & Morel, J.-M. (1998). Level lines based disocclusion. In Image Processing, 1998. ICIP 98. Proceedings. 1998 International Conference on (pp. 259–263). volume 3. Mitskog, T. F. R., & Ralston, R. A. (2012). Camera blocker for a device with an integrated camera that uses a thin film organic polymer. Morris, M. E., Adair, B., Miller, K., Ozanne, E., Hansen, R., Pearce, A. J., Santamaria, N., Viegas, L., Long, M., & Said, C. M. (2013). Smart-home technologies to assist older people to live well at home. Journal of Aging Science, 1 . doi:10.4172/2329-8847.1000101. Morse, B., & Schwartzwald, D. (1998). Isophote-based interpolation. In Image Processing, 1998. ICIP 98. Proceedings. 1998 International Conference on (pp. 227–231). volume 3. Neustaedter, C., & Greenberg, S. (2003). The design of a context-aware home media space for balancing privacy and awareness. In A. Dey, A. Schmidt, & J. McCarthy (Eds.), UbiComp 2003: Ubiquitous Computing (pp. 297–314). Springer Berlin Heidelberg volume 2864 of Lecture Notes in Computer Sci- ence. Neustaedter, C., Greenberg, S., & Boyle, M. (2006). Blur filtration fails to preserve privacy for home-based video conferencing. Computer-Human Inter- action, ACM Transactions on, 13 , 1–36. Newton, E., Sweeney, L., & Malin, B. (2005). Preserving privacy by de- identifying face images. Knowledge and Data Engineering, IEEE Transactions on, 17 , 232–243. Ng, P. L., minn Ang, L., & Seng, K. P. (2010). Privacy preserving stereoscopic vision with one-bit transform. In Computer Science and Information Tech- nology (ICCSIT), 2010 3rd IEEE International Conference on (pp. 70–74). volume 9. Ni, Z., Shi, Y.-Q., Ansari, N., & Su, W. (2006). Reversible data hiding. Circuits and Systems for Video Technology, IEEE Transactions on, 16 , 354–362. O’Brien, A., & Mac Ruairi, R. (2009). Survey of assistive technology devices and applications for aging in place. In Advances in Human-oriented and Personalized Mechanisms, Technologies, and Services, 2009. CENTRIC ’09. Second International Conference on (pp. 7–12). 42 http://dx.doi.org/10.4172/2329-8847.1000101 Olivieri, D. N., Conde, I. G., & Sobrino, X. A. V. (2012). Eigenspace-based fall detection and activity recognition from motion templates and machine learning. Expert Systems with Applications , 39 , 5935–5945. Pande, A., & Zambreno, J. (2013). Securing multimedia content using joint compression and encryption. In Embedded Multimedia Security Systems (pp. 23–30). Springer London. Park, S., & Kautz, H. (2008). Privacy-preserving recognition of activities in daily living from multi-view silhouettes and RFID-based training. In AI in Eldercare: New Solutions to Old Problems, AAAI Fall Symposium on. Paruchuri, J., & Cheung, S.-c. (2008). Joint optimization of data hiding and video compression. In Circuits and Systems, 2008. ISCAS 2008. IEEE Inter- national Symposium on (pp. 3021–3024). Paruchuri, J. K., Cheung, S.-c. S., & Hail, M. W. (2009). Video data hiding for managing privacy information in surveillance systems. EURASIP Journal on Information Security , 2009 , 1–18. doi:10.1155/2009/236139. Article ID: 2009:236139. Patel, S. N., Summet, J. W., & Truong, K. N. (2009). Blindspot: Creating capture-resistant spaces. In A. Senior (Ed.), Protecting Privacy in Video Surveillance (pp. 185–201). Springer London. Patwardhan, K., Sapiro, G., & Bertalmio, M. (2007). Video inpainting under constrained camera motion. Image Processing, IEEE Transactions on, 16 , 545–553. Petitcolas, F., Anderson, R., & Kuhn, M. (1999). Information hiding-a survey. Proceedings of the IEEE , 87 , 1062–1078. Qiao, L., Nahrstedt, K. et al. (1997). A new algorithm for MPEG video encryp- tion. In Imaging Science, Systems and Technology (CISST’97, Proceedings of First International Conference on (pp. 21–29). Qureshi, F. (2009). Object-video streams for preserving privacy in video surveil- lance. In Advanced Video and Signal Based Surveillance, 2009. AVSS ’09. Sixth IEEE International Conference on (pp. 442–447). Ra, M.-R., Govindan, R., & Ortega, A. (2013). P3: Toward privacy-preserving photo sharing. In Networked Systems Design and Implementation, Proceed- ings of the 10th USENIX conference on (pp. 515–528). Raju, C., Umadevi, G., Srinathan, K., & Jawahar, C. (2008). Fast and secure real-time video encryption. In Computer Vision, Graphics Image Processing, 2008. ICVGIP ’08. Sixth Indian Conference on (pp. 257–264). IEEE. 43 http://dx.doi.org/10.1155/2009/236139 Sadeghi, A.-R., Schneider, T., & Wehrenberg, I. (2010). Efficient privacy- preserving face recognition. In D. Lee, & S. Hong (Eds.), Information, Secu- rity and Cryptology - ICISC 2009 (pp. 229–244). Springer Berlin / Heidelberg volume 5984 of Lecture Notes in Computer Science. Sadimon, S., Sunar, M., Mohamad, D., & Haron, H. (2010). Computer gener- ated caricature: A survey. In Cyberworlds (CW), 2010 International Confer- ence on (pp. 383–390). Saini, M., Atrey, P., Mehrotra, S., & Kankanhalli, M. (2014). W3-privacy: understanding what, when, and where inference channels in multi-camera surveillance video. Multimedia Tools and Applications , 68 , 135–158. Schaar, P. (2010). Privacy by design. Identity in the Information Society , 3 , 267–274. Schiff, J., Meingast, M., Mulligan, D. K., Sastry, S., & Goldberg, K. (2009). Respectful cameras: Detecting visual markers in real-time to address privacy concerns. In A. Senior (Ed.), Protecting Privacy in Video Surveillance (pp. 65–89). Springer London. Senior, A. (2009). Privacy protection in a video surveillance system. In A. Senior (Ed.), Protecting Privacy in Video Surveillance (pp. 35–47). Springer London. Senior, A., & Pankanti, S. (2011). Privacy protection and face recognition. In S. Z. Li, & A. K. Jain (Eds.), Handbook of Face Recognition (pp. 671–691). Springer London. Senior, A., Pankanti, S., Hampapur, A., Brown, L., li Tian, Y., & Ekin, A. (2003). Blinkering Surveillance: Enabling Video Privacy through Computer Vision. Technical Report IBM Research Division. Senior, A., Pankanti, S., Hampapur, A., Brown, L., Tian, Y.-L., Ekin, A., Connell, J., Shu, C. F., & Lu, M. (2005). Enabling video privacy through computer vision. Security Privacy, IEEE , 3 , 50–57. Shashank, J., Kowshik, P., Srinathan, K., & Jawahar, C. (2008). Private content based image retrieval. In Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on (pp. 1–8). Shashanka, M. (2010). A privacy preserving framework for gaussian mixture models. In Data Mining Workshops (ICDMW), 2010 IEEE International Conference on (pp. 499–506). Shiratori, T., Matsushita, Y., Tang, X., & Kang, S. B. (2006). Video completion by motion field transfer. In Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on (pp. 411–418). volume 1. Shoaib, M., Elbrandt, T., Dragon, R., & Ostermann, J. (2010). Altcare: Safe living for elderly people. In Pervasive Computing Technologies for Healthcare (PervasiveHealth), 2010 4th International Conference on (pp. 1–4). 44 Sohn, H., De Neve, W., & Ro, Y. M. (2011). Privacy protection in video surveillance systems: Analysis of subband-adaptive scrambling in JPEG XR. Circuits and Systems for Video Technology, IEEE Transactions on, 21 , 170– 177. Sohn, H., Plataniotis, K., & Ro, Y. (2010). Privacy-preserving watch list screen- ing in video surveillance system. In G. Qiu, K. Lam, H. Kiya, X.-Y. Xue, C.-C. Kuo, & M. Lew (Eds.), Advances in Multimedia Information Process- ing - PCM 2010 (pp. 622–632). Springer Berlin / Heidelberg volume 6297 of Lecture Notes in Computer Science. Spanos, G. A., & Maples, T. B. (1995). Performance study of a selective en- cryption scheme for the security of networked, real-time video. In Computer Communications and Networks, International Conference on. Srinivasan, S., Tu, C., Regunathan, S., & Sullivan, G. (2007). HD photo: a new image coding technology for digital photography. In Applications of Digital Image Processing XXX, Proceedings of SPIE (pp. 1–19). volume 6696. doi:10.1117/12.767840 Article ID: 66960A. Sweeney, L. (2002). k-anonymity: a model for protecting privacy. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems , 10 , 557– 570. Tang, L. (1996). Methods for encrypting and decrypting MPEG video data efficiently. In Multimedia, Proceedings of the fourth ACM international con- ference on (pp. 219–229). ACM. Tansuriyavong, S., & Hanaki, S.-i. (2001). Privacy protection by concealing per- sons in circumstantial video image. In Perceptive user interfaces, Proceedings of the 2001 workshop on (pp. 1–4). ACM. Tian, Y.-l., Brown, L., Hampapur, A., Lu, M., Senior, A., & Shu, C.-f. (2008). IBM smart surveillance system (S3): event based video surveillance system with an open and extensible framework. Machine Vision and Applications , 19 , 315–327. Tong, L., Dai, F., Zhang, Y., & Li, J. (2010). Prediction restricted H.264/AVC video scrambling for privacy protection. Electronics Letters , 46 , 47–49. Tong, L., Dai, F., Zhang, Y., Li, J., & Zhang, D. (2011). Compressive sensing based video scrambling for privacy protection. In Visual Communications and Image Processing (VCIP), 2011 IEEE (pp. 1–4). Turk, M., & Pentland, A. (1991). Face recognition using eigenfaces. In Com- puter Vision and Pattern Recognition, 1991. Proceedings CVPR ’91., IEEE Computer Society Conference on (pp. 586–591). 45 http://dx.doi.org/10.1117/12.767840 Upmanyu, M., Namboodiri, A., Srinathan, K., & Jawahar, C. (2009). Efficient privacy preserving video surveillance. In Computer Vision, 2009 IEEE 12th International Conference on (pp. 1639–1646). Vijay Venkatesh, M., Cheung, S.-c. S., & Zhao, J. (2009). Efficient object-based video inpainting. Pattern Recognition Letters , 30 , 168–179. Wagstaff, J. (2004). Using bluetooth to disable camera phones. URL: http://www.loosewireblog.com/2004/09/using_bluetooth.html last vis- ited: 2014/05/08. Wallace, E., & Diffley, C. (1988). CCTV control rooms ergonomics . Technical Report 14/98 Police Scientific Development Branch (PSDB) UK Home Office. Wexler, Y., Shechtman, E., & Irani, M. (2007). Space-time completion of video. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 29 , 463– 476. Whyte, O., Sivic, J., & Zisserman, A. (2009). Get out of my picture! internet- based inpainting. In Proceedings of the 20th British Machine Vision Confer- ence, London. Williams, A., Xie, D., Ou, S., Grupen, R., Hanson, A., & Riseman, E. (2006). Distributed Smart Cameras for Aging in Place. Technical Report DTIC Doc- umen. Winkler, T., & Rinner, B. (2010a). Securing embedded smart cameras with trusted computing. EURASIP Journal on Wireless Communications and Net- working , 2011 , 1–20. doi:10.1155/2011/530354. Article ID: 2011:530354. Winkler, T., & Rinner, B. (2010b). TrustCAM: Security and privacy-protection for an embedded smart camera based on trusted computing. In Advanced Video and Signal Based Surveillance (AVSS), 2010 Seventh IEEE Interna- tional Conference on (pp. 593–600). Winkler, T., & Rinner, B. (2014). Security and privacy protection in visual sensor networks: A survey. ACM Computing Surveys , 47 , 1–42. Wu, M. (2001). Multimedia data hiding . Ph.D. thesis Princeton University. Xiangdong, L., Junxing, Z., Jinhai, Z., & Xiqin, H. (2008). Image scrambling algorithm based on chaos theory and sorting transformation. IJCSNS Inter- national Journal of Computer Science and Network Security , 8 , 64–68. Yabuta, K., Kitazawa, H., & Tanaka, T. (2005). A new concept of security camera monitoring with privacy protection by masking moving objects. In Y.-S. Ho, & H. Kim (Eds.), Advances in Multimedia Information Processing - PCM 2005 (pp. 831–842). Springer Berlin / Heidelberg volume 3767 of Lecture Notes in Computer Science. 46 http://www.loosewireblog.com/2004/09/using_bluetooth.html http://dx.doi.org/10.1155/2011/530354 Yabuta, K., Kitazawa, H., & Tanaka, T. (2006). A new concept of real-time se- curity camera monitoring with privacy protection by masking moving objects. In Proceedings of SPIE (pp. 188–199). volume 6063. Yang, M., Bourbakis, N., & Li, S. (2004). Data-image-video encryption. Poten- tials, IEEE , 23 , 28–34. Yu, X., & Babaguchi, N. (2007). Privacy preserving: Hiding a face in a face. In Y. Yagi, S. Kang, I. Kweon, & H. Zha (Eds.), Computer Vision - ACCV 2007 (pp. 651–661). Springer Berlin / Heidelberg volume 4844 of Lecture Notes in Computer Science. Yu, X., Chinomi, K., Koshimizu, T., Nitta, N., Ito, Y., & Babaguchi, N. (2008). Privacy protecting visual processing for secure video surveillance. In Image Processing, 2008. ICIP 2008. 15th IEEE International Conference on (pp. 1672–1675). Zeng, W., & Lei, S. (2003). Efficient frequency domain selective scrambling of digital video. Multimedia, IEEE Transactions on, 5 , 118–129. Zhang, C., Rui, Y., & wei He, L. (2006). Light weight background blurring for video conferencing applications. In Image Processing, 2006 IEEE Interna- tional Conference on (pp. 481–484). Zhang, C., Tian, Y., & Capezuti, E. (2012). Privacy preserving automatic fall detection for elderly using RGBD cameras. In K. Miesenberger, A. Karshmer, P. Penaz, & W. Zagler (Eds.), Computers Helping People with Special Needs (pp. 625–633). Springer Berlin Heidelberg volume 7382 of Lecture Notes in Computer Science. Zhang, W., Cheung, S., & Chen, M. (2005a). Hiding privacy information in video surveillance system. In Image Processing, 2005. ICIP 2005, IEEE In- ternational Conference on (pp. II–868–71). volume 3. Zhang, Y., Xiao, J., & Shah, M. (2005b). Motion layer based object removal in videos. In Application of Computer Vision, 2005. WACV/MOTIONS ’05 Volume 1. Seventh IEEE Workshops on (pp. 516–521). volume 1. Zhao, Q. A., & Stasko, J. T. (1998). The awareness-privacy tradeoff in video sup- ported informal awareness: A study of image-filtering based techniques . Tech- nical Report Georgia Institute of Technology. GVU Technical Report;GIT- GVU-98-16. Zivkovic, Z. (2004). Improved adaptive gaussian mixture model for background subtraction. In Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference on (pp. 28–31). volume 2. 47 Introduction Related studies Visual privacy protection Protection methods Intervention Blind vision Secure processing Redaction Image filtering Encryption of videos and images Face de-identification Object / people removal Visual abstraction / object replacement Data hiding based methods Privacy-aware intelligent monitoring systems Discussion Recognition of the region of interest Privacy and intelligibility trade-off Real-time applications Privacy model and evaluation User studies about privacy requirements Conclusion 
work_ms2ruztupvforj5fnbvzf7tofy	----	AN ABSTRACT OF THE THESIS OF Yimin Zenci for the degree of Master of Science in Industrial and Manufacturing Enciineerinci presented on MaY 5, 1993 Title: An Expert System For Softwood Lumber Grading Abstract approved: Dr. Sabah Randhawa The focus of this research is to develop a prototype expert system for softwood lumber grading. The grading rules used in the knowledge base of the system are based on Western Lumber Grading Rules 88 published by the Western Wood Products Association. The system includes 27 grades in Dimension, Select/Finish, and Boards categories. The system is designed to be interactive and menu- driven. The user input to the system consists of lumber size, grade category, and type, location and size of defects for each face. The system then infers the grade corresponding to each face, and an overall grade for the lumber. The system provides limited explanation capabilities. Evaluation of the system was performed using 85 samples of pre-graded Siberian larch 2x4x12s in Structural Light Framing category. The initial evaluation was performed using the two wide faces of boards. Results indicated a 60 percent match between the grade assigned by the human expert and the system. The largest cause of deviation was exclusion of defects on the two narrow faces. The knowledge base was expanded to include the two narrow faces; the match rate improved to 76.5 percent. Evaluations for other grading categories need to be conducted in the future to assess the adequacy of the knowledge base. The prototype development concentrates on selected defect characteristics for each grade. These characteristics are clearly defined and described in the rule book, and are usually the most frequently encountered defects on softwood lumber. The knowledge base needs to be refined and expanded if additional factors such as knot positions relative to each other, warp, manufacturing imperfections and clustering of defects are to be considered. (C) Copyright by Yimin Zeng May 5, 1993 All Rights Reserved An Expert System For Softwood Lumber Grading by Yimin Zeng A THESIS submitted to Oregon State University in partial fulfillment of the requirements for the degree of Master of Science Completed May 5, 1993 Commencement June 1993 APPROVED: Professor of Industrial and Manufacturing Engineering in charge of major Head of Department of Industrial and Manufacturing Engineering Dean of Gradua Sè'hdol Date thesis is presented May 5. 1993 Typed by Yimin Zeng for Yimin Zenci ACIOWLEDGEMENTB I wish to express my deepest gratitude and appreciation to my major professor, Dr. Sabah Randhawa, for his guidance, encouragement and support throughout the course of my graduate study. His tremendous patience and judicious suggestions made possible the successful completion of this thesis. I wish to thank Dr. Kenneth Funk and Dr. Eldon Olsen for contributing their time and expertise while serving on my graduate committee. My very special appreciations go to Dr. James Funk. He spent tremendous amount of time to teach me about lumber production and grading, and to help me to evaluate the system using real lumber. I would also like to thank Dr. Robert L. Ethington and Dr. Rakesh Gupta for providing me with lumber samples. I am grateful to my friend Mr. Johannes B. Forrer. The discussions with him always inspired me. I am deeply indebted to my wife, Yuan Zhong, for her love, understanding, encouragement and help in all aspects of my daily life and graduate studies. I am deeply indebted to my parents and sisters for their deep and eternal love, understanding, encouragement and support. TABLE OP CONTENTS CHAPTER1INTRODUCTION . . 1 CHAPTER 2 RESEARCH OBJECTIVES ............ 7 CHAPTER 3 LITERATURE REVIEW ....... . . . . . . 8 3.1 Programs Using Algorithmic Approach . . . 9 3.2 Programs Using Expert System Approaches . 12 3.2.1 Structure of Expert Systems . . 12 3.2.2 Expert Systems in Lumber Grading . 15 3.3 Systems Developed in Industry ....... 17 CHAPTER4BYSTEMDEVELOPMENT ............ 19 4.1 ProblemDomain .............. 19 4.1.1 Softwood Lumber Grading Practices 19 4.1.2 Grading Categories ....... 2]. 4.2 Knowledge Acquisition and Representation 31 4.2.1 Knowledge Acquisition ....... 31 4.2.2 Knowledge Representation ...... 32 4.3 Inference Mechanism ............ 38 4.4 Tool Selection .............. 43 4.5 Prototype Development ........... 48 4.5.1 Prototype Components ........ 48 4.5.2 Knowledge Base Implementation . 51 4.5.3 Grading Process, Input, and Output . 57 CHAPTER 5 EVALUATION AND ENHANCEMENTS ....... 67 5.1 System Evaluation ............ 67 5.2 System Improvement ............ 69 CHAPTER6CONCLUSIONS ................ 73 BIBLIOGRAPHY .................... 76 APPENDICES Appendix A. Source Code for Structural Light Framing Category ................ 7 8 Appendix B. Test Results Using Generated Data . . . 93 LIST OF FIGURES Figure Page 1.1. Flow chart of a typical sawmill ........ 2 1.2. Flow chart of a typical planer mill ...... 3 3.1. Basic structure of an expert system ...... 13 4.1. Grading categories, sub-categories, and grades included in this project ........ 22 4.2. Basic structure of the prototype ....... 49 4.3. Gradingprocess ............... 58 4.4. Opening window of the run-time version ofthesystein ................ 61 4.5. Dialog box for selecting input method ..... 62 4.6. Dialog box for entering lumber thickness . . . 62 4.7. Dialog box for confirming lumber dimension . . 62 4.8. Dialog box for selecting grade category . . . . 63 4.9. Dialog box for selecting dimension sub- categorr ................... 63 4.10. Dialog box for selecting lumber face ..... 64 4.11. Dialog box with partial defect list ...... 64 4.12. Dialog box with another partial defect list . . 65 4.13. Dialog box for entering knot type and location ................... 65 4.14. Dialog box for entering knot diameter ..... 66 4.15. System output display ............. 66 5.1. Evaluation results considering all factors .................... 68 5.2. Evaluation results when warp, manufacturing imperfections, and measurement bias were not considered................... 71 LIST OF TABLES Table Page 4.1. Grading categories .............. 21 4.2. Grade characteristics considered by the project for grades in Select/Finish's Selectsub-category .............. 24 4.3. Grade characteristics considered by the project for grades in Select/Finish's Finish sub-category .............. 25 4.4. Grade characteristics considered by the project for grades in Boards' Commons sub-category ................. 26 4.5. Grade characteristics considered by the project for grades in Boards' Alternate sub-category ................. 27 4.6. Grade characteristics considered by the project for grades in Dimensions' Light Framing sub-category ........... 28 4.7. Grade characteristics considered by the project for grades in Dimensions' Structural Light Framing sub-category ..... 29 4.8. Grade characteristics considered by the project for grades in Dimensions' Studsub-category ............... 30 4.9. Characteristics permitted and limiting provisions for CONSTRUCTION grade in Light Framing sub-category of dimension lumber . . . 35 4.10. Grading rules for the CONSTRUCTION grade represented using production rules ...... 36 4.11. Source files and their contents ........ 50 LIST OF APPENDICES TABLES Table Bi. Specifications of grades for a piece of Page lumber in the Dimension category (lumber size= 94 B2. Test results of grading the lumber in the Dimensioncategory ............... 95 B3. Specifications of grades for a piece of lumber in the Select/Finish category (lumber size = l"x8"x12') .......... 96 B4. Test results of grading the lumber in the Select/Finish category ........... 97 B5. Specifications of grades for a piece of lumber in the Boards category (lumber size = l"x8"x12') ........... 98 B6. Test results of grading the lumber in theBoardscategory .............. 99 AN EXPERT SYSTEM FOR SOFTWOOD LUMBER GRADING CHAPTER 2 INTRODUCTION Lumber sawiuilling is one of the oldest industries in America. In mills, logs are converted into lumber by sawing, edging and trimming operations (Williston, 1988). At the sawing station, logs are broken down into rough, unedged and untrimmed lumber. The rough lumber is then cut into green lumber in specified width and length at edging and trimming stations. Rough green lumber is then dried and surfaced into finished products. Figures 1.1 and 1.2 show the primary steps in sawmill and planer mills for lumber manufacturing. Configuration of different mills may vary depending upon the log supply and finished products. One of the most important issues concerning the lumber manufacturing industry is the efficiency of raw material usage. Lumber recovery, a term used to describe how much finished lumber is produced from a log or a given population of logs, is used as the primary measure for the performance of a sawmill. The trend is to use computer- based technology to optimize lumber recovery. Trimming and edging stations first used computer technology to obtain optimum cutting patterns of unedged and untrimmed lumber. Later, optimization techniques were used at log breakdown stations where the optimum log breakdown patterns are determined. 2 Raw Material Storage Debark! Buck Log Breakdown Edging Trimming Rough Green Lumber Grading/Sorting/Stacking Rough Green Lumber Storage Drying Dry Storage Figure 1.1. Flow chart of a typical sawmill 3 Unstacking Planing Grading and Trimming Sorting Stacking and Packing Storage Ship Figure 1.2. Flow chart of a typical planer mill The use of optimization techniques at the log breakdown stations has evolved from initially considering only one geometric dimension of a log to where three dimensional features of the log are now being considered. 4 However, currently available models make their "optimum" decisions based only on the log's or the board's geometrical dimensions. An important aspect ignored by the optimization models is the effect of log internal defect type, size, and location on final lumber grade (i.e., recovery and quality). The presence of defects needs to be considered in all manufacturing steps, especially in the log breakdown, edging and triimning operations. More value recovery is expected if an optimization system can combine both geometrical and quality information, and predict lumber grade when making cutting decisions. Natural log characteristics and manufacturing imperfections have a great impact on lumber quality. To provide specified quality levels for end-use, lumber must be graded according to its quality characteristics. Lumber grading is performed in a sawmill or a sawmill and planing mill, and is monitored by authorized agencies who publish grading rules which conform to the American Softwood Lumber Standards (ASLS, 1970). Traditionally, lumber grading is performed by trained human graders through visual inspection. Although computer aided manufacturing techniques have been widely used in modern sawmills, development of an automated lumber grading system was not initiated until recently because of the complexity of grading rules and the difficulties in the detection and recognition of defects using a machine vision system. 5 A lumber grading software package can be invaluable to the lumber manufacturing industry in the following aspects: 1. It can be linked to a log breakdown optimization program, such as SAW3D (Zeng, 1991), to improve lumber recovery. As mentioned before, lumber breakdown optimization programs have been used extensively in sawmills. After receiving log scanning data, the optimization system can determine optimum sawing, edging and trimming patterns to find the best lumber recovery from the given log. Nevertheless, such optimization systems do not consider external and internal defects of the log. As such, the resulting solution is not optimal. From a value stand point, an advanced log breakdown optimization system that considers log defects in addition to other characteristics will be able to provide better solutions. The lumber grading program can be embedded in a log breakdown program (such as SAW3D). 2. It can be coupled with a machine vision system to make an on-line automatic lumber grading system. An on- line automatic lumber grading system consists of two basic components, a machine vision system which performs as human eyes to capture the size and quality characteristics of lumber, and a grading software program which uses the information provided by the machine vision system to determine the grade of lumber. An integrated system should be able to provide more reliable and accurate grading [.1 results than a human grader, and it should be more economical than using .a human grader. 3. It can be used as an instruction tool to aid in training novices and upgrading knowledge of grading personnel. 7 CHAPTER 2 RESEARCH OBJECTIVES The objective of this research was to develop an expert system for softwood lumber grading. The primary objective of developing this system was to link it to a log breakdown optimization system, SAW3D (Zeng, 1991), so that the integrated system can consider both log geometric and quality information when making sawing, edging, and trimming decisions to optimize lumber recovery. Grading rules used in this system were based on "Western Lumber Grading Rules 88" published by the Western Wood Products Association. The research focused on feasibility analysis, and the conceptual and structure design of the system. Therefore, only selected grades and wood characteristics were included in the system. When SAW3D makes optimum log breakdown decisions, some defects such as manufacturing imperfections do not exist. Such defects were not included in the grading system. 8 CHAPTER 3 LITERATURE REVIEW Although the significance of computer lumber grading programs has been recognized and some researchers developed programs for lumber grading in the early 1970's (Hallock and Galiger, 1971), such programs have not been extensively used in industry. One possible reason for lack of use is that a prerequisite for a defect detection system is a machine vision system for lumber grading. Developing a reliable machine vision system is a difficult task. As research on machine vision has evolved, activity in the development of lumber grading programs has also increased. Some prototype automatic grading systems have been recently developed in the sawmill industry. Currently available computer grading programs have been developed for hardwood lumber grading, and use grading rules established by the National Hardwood Lumber Association (NHLA). There are two basic problems that exist while grading lumber: 1) the recognition of the type and seriousness of each defect, and 2) the calculation of usable areas. In grading hardwood lumber, all grades are based on the recognition of defects and the calculation of usable area. The focus of this research is softwood lumber, for which little has been developed in way of automatic grading systems. In contrast to hardwood lumber, softwood lumber grading is primarily based on the recognition of defects; only grades of Factory Lumber need to calculate usable area. Since grading problems for hardwood lumber and softwood lumber are very similar except for the differences in using different grading rules and grading procedures, it is worthwhile to review some of the research for hardwood lumber grading. 3.1 Programs Using Algorithmic Approach Most of the early programs were developed using an algorithmic programming approach. Although the expert system approach has an obvious advantage over the algorithmic approach, lack of good development tools and the requirement of expensive computing facilities prevented use of the expert system approach until recently. Hallock and Gauger (1971) developed a computer program for grading hardwood lumber. The program analyzes the board with its defects, finds the number of cuttings of the sizes permitted, and assigns a grade according to NHL rules. It classifies lumber according to six major classifications. The board and all its defects are described by rectangles in a Cartesian coordinate system; the X-Y coordinates of those rectangles are entered on a series of cards. Once the program receives the input data, 10 it begins the grading process by assigning the board a potential grade based on the board size. This is the highest possible grade for the board without considering defects. Then the defects are checked and the board is assigned a potentially lower grade because of the impact of defects. Next, sizes and number of usable areas are calculated, and the board is assigned the final grade. The program was developed in FORTRAN. The authors believed the program's accuracy to be well within NHLA tolerances. However, some limitations exist. For example, grading rules and procedures were all coded into the program algorithms. This programming practice makes modification and expansion of grading rules difficult. Additionally, the algorithms considered only one face of the board, whereas both faces of a board are evaluated in the actual grading process. Klinkhachorn et al. (1988) created a hardwood lumber grading program as part of an automated lumber process system. The program was implemented using FORTRAN 77 for an IBM PC or compatible. After reading in the board dimensions and defect types and their location on each face, the program assigns a potential grade to each face. Next the program calculates usable area to determine the final grade assigned to the board. This program improved upon the previous efforts by considering both faces of the board and applying a more flexible grading procedure which provides the option of admitting or excluding certain types 11 of defects when searching for cuttings. In addition, the program uses the BLOCK DATA feature of FORTRAN 77 to logically separate grading rules from the program logical code so that reconfiguration of the grading rules is easier. When grading rules need to be modified, only the data specified in the BLOCK DATA segment need to be changed. Klinkhachorn et al. (1989) presented a computer-aided instruction tool, named HaLT (Hardwood Lumber Training) for training hardwood lumber graders. The HaLT system incorporates a graphic user interface and explanation facility in the hardwood lumber grading program discussed above. The interface provides graphic views of the board including each side of the entire board with defects or the expanded view giving the detailed section of the board, asks questions, compares answers, and provides reasoning. The program runs on an IBM PC or compatible. The objective is to provide the lumber industry with an interactive learning tool to educate and train their graders on-site. Klinkhachorn et al. (1991) improved their previous program (Klinkhachorn et al, 1988) by converting the code from FORTRAN 77 to C, and restructuring it. The grading rules for various species, the mechanical dimensions of a board, the violations of the various rules, and the grade of the board are all mathematically represented in terms of program variables in data structure so that the program 12 data are logically separated from the program code. This enhances the program's capabilities, its adaptability and portability, and makes the restructuring of grading rules easier. 3.2 Programs Using Expert System Approaches 3.2.1 Structure of Expert Systems An expert system is a computer program that uses knowledge, facts, and reasoning techniques to solve problems that normally require the abilities of human experts. Expert system architectures include different components. However, certain components are common to most expert systems. These include a user interface, a knowledge base, and an inference engine. Figure 3.1 shows the basic structure of an expert system. The interface provides the communication between the user and the system. It allows the user to query the system, enter facts about a specific situation that is relevant to the system's subject domain, and receive advice, consultation and explanation. Many expert systems also accept new knowledge through the user interface. The knowledge base contains expert-level knowledge on the system's subject domain. The knowledge can be obtained from books and publications, and/or from one or more human experts, and is stored using a knowledge representation 13 scheme. The inference engine provides the reasoning capabilities of the system. It uses the domain specific knowledge in the knowledge base and facts provided by the user to infer new facts that may be conclusions to goals and subgoals as required by the system. The inference method is built into the inference engine, while the domain specific knowledge is contained in the knowledge base. The general problem solving knowledge and the domain specific User User Interface Inference Engine Knowledge Base Figure 3.1. Basic structure of an expert system 14 knowledge are separated, thus the same inference engine can be used to reason with different knowledge bases. In addition to those basic components, many expert systems have various supplemental components, including knowledge acquisition facilities, knowledge-base maintenance and update facilities, explanation facilities, self-training facilities, and validation tools. Expert system techniques are very appropriate to problems for which human expertise is required, which. need symbolic reasoning but cannot be easily handled numerically or algorithmically, or for which information required for a solution is incomplete or uncertain. The lumber grading problem is declarative rather than algorithmic in nature. The grading process involves symbolic rather than numeric reasoning. Maintaining and updating grading rules is required as criteria change over time, and from one mill to another. In some cases, data about a defect may be uncertain. If used as an instruction tool, explanation on the actions and conclusion would be necessary. All these facts indicate that while an algorithmic approach is possible, the expert system approach is more appropriate. Using the expert system approach for lumber grading has the following advantages over the algorithmic approach: 1. In knowledge-based systems, lumber data and grading rules are stored in a knowledge base which is decoupled 15 from the program logic. Hence, modification and expansion of grading rules can be made easily without affecting the program logic. 2. When used as a computer-aided instruction (CAl) tool, knowledge-based systems are more appropriate because they provide consultation and explanation capabilities required in training situations. 3. Rules-of-thumb used by grading experts can also be incorporated in the knowledge base to supplement the rules from publications, and to speed up the grading process as these may result in reducing the search space. Furthermore, due to the separation of the knowledge base from the program logic, customization of such rules can be easier. 3.2.2 Expert Systems in Lumber Grading Huang and sparrow (1989) developed a CAl tool for hardwood lumber grading using the expert system approach. The expert system consists of five main components: a knowledge base, an inference engine, a procedure for calculating the usable area, a user interface facility, and an explanation subsystem. The knowledge base contains facts and rules. Facts include lumber data, current data, and information derived from the rules. Rules are the conditions and restrictions for judging the grades of the boards. Knowledge is represented by rules and working memory elements. A forward-chaining inference mechanism is 16 used. The system was implemented using the LISP language and runs on the VAX 11/780 computer. The program facilitates the training of professional hardwood lumber graders by providing consultation and explanation of lumber grading procedure in an interactive question-answer process. Lumber data, including lumber dimensions, types and sizes of defects, need to be entered by the user. The program instructs the user in three steps: 1) to judge grades of the two faces based on a combination of defect types and sizes; 2) to judge grades of the two faces based on calculation of usable areas on both faces; and 3) to judge the overall grade of the board. At each step, it asks the user to determine a tentative grade, and then explains the user's mistakes, if any. Otherwise, it provides an explanation of the process of reaching the correct grade. The program reduces the training costs by decreasing the need to rely solely on an expensive human instructor, and increases the speed and retention rates of students because it offers a flexible time table that makes it more convenient for the student to work at his/her own pace. The knowledge base is limited in the system. Among over 100 types of hardwood lumber, only two types (walnut and poplar) were chosen for this initial work. Softwood lumber grading programs using the expert system approach have not been developed. 17 3.3 Systems Developed in Industry As mentioned before, lumber recovery optimization systems based on dimension scanning have been widely used in sawmills. While more complete geometric information is included in recent three-dimensional scanning and optimization systems, higher recovery is to be expected if the defect information along with the dimensional data is used in making cutting decisions. Though some corporations have been developing grade scanning and optimization systems which combine both dimension and quality information for logs and lumber in making optimum decisions, there is little information available in published literature. In addition to the difficulty of accessing information from industry, available papers focus on defect detection systems, as compared to implementing grading rules in their programs. This is not surprising since the development of real time defect detection systems has been a long time barrier in this area. VisionSmart Inc. of Canada developed a Grade-based Optimizer system which uses a camera to scan all features related to dimension and wane, and an X-ray sensor array to detect knots and other density related features such as compression wood and pitch pockets (Flatman and Bodell, 1989; Flatinan and Kenway, 1991; Kenway, 1990; Aune, 1991). Other features such as shakes, checks, holes, and 18 stains are detected by cameras operating under different lighting than the lighting for the camera used for dimension and wane. The system combines both profile and defect data and then determines optimum cutting solutions for maximum value return. The system can be used either for green lumber or for dry lumber. A Grade-Based Edger Optimizer was installed in MacMillan Bloedel's Alberni Pacific Division sawmill in March, 1989. The Edger Optimizer scans wet green lumber for grading features and based on the scanning data, uses a dynamic programming algorithm to determine an optimum edging solution. The Edger Optimizer showed a $160,000 return per month, as measured by the mill. A Grade-Based Trimmer Optimizer was also installed in CANFOR'S Fort St. John Mill in May 1990, where dry lumber is scanned after planing and the optimum trimming solutions calculated. The Coe Manufacturing Company developed the D*TEC GradeScan system that is able to detect knots, splits, pockets, stain and decay, as well as wane and dimensions1. The edger, trimmer and planer mill applications manufactured by the company have been equipped with the GradeScan system so that both defects and shape can be considered when optimizing for value recovery. The Coe Manufacturing Company. 1991. Company brochure. Portland, Oregon. 19 CHAPTER 4 SYSTEM DEVELOPMENT 4.1 Problem Domain 4.1.1 Softwood Lumber Grading Practices Lumber grading is performed in a sawmill or a planer mill by human graders. The graders evaluate each piece of lumber and make their decision according to grading rules as specified by regional association, proprietary, or national grading rules. The following çiescriptions are given in Western Lumber Grading Rules 88 (1988): The grading of lumber cannot be considered an exact science because it is based on either a visual inspection of each piece and the judgment of the grader or on such visual inspection and judgment combined with the results of method of mechanically determining the stiffness characteristics of structural lumber. The results are, however, sufficiently explicit to establish a maximum of 5 percent below grade as a reasonable variation between graders. In general, American Standard softwood lumber is classified according to use, manufacturing, and size 20 classifications. Grading of lumber is based upon a visual inspection and is primarily a judgment of end-use strength or appearance. Natural log characteristics which have an effect upon strength are, taken into account along with manufacturing imperfections. In most cases these include limiting provisions on characteristics such as stain, checks, grain, skip, cup, crook, wane, twist, knots, holes, pitch, pocket, split, steaks and patches, shakes, decay, and cutout. Definitions and detailed descriptions of these characteristics can be found in the Western Lumber Grading Rules 88 (1988). Grading rules applied to softwood lumber for domestic market are governed by National Bureau of Standards Voluntary Product Standard PS 20-70 (American Softwood Lumber Standard, 1970). All of the grading rule books which are published by grading organizations are based on this standard. The two largest grading organizations on the West Coast are the Western Wood Products Association (WWPA) and the West Coast Lumber Inspection Bureau (WCLIB). In 1986, nearly 35 percent of the nation's softwood lumber production was graded under WWPA supervision. This research project will use the grading rules published by the WWPA as described in the Western Lumber Grading Rules 88 (1988). 21 4.1.2 Grading Categories According to the WWPA, lumber is classified into one of the following five categories (Table 4.1) based on the size and end-use: Select and Finish lumber, Boards, Dimension, Timbers, and Factory lumber and related products. Table 4.1. Gradinci cateaories Name of category Nominal Size Judgment on Use EmphasisThickness Width Selects and Finish______________ 3/8" - 16/4" >= 2" appearance Boards 3/4" - 16/4" >= 2" appearance Dimension 2" - 4" >= 2" strength Timber >= 5" >= 5" strength Factory any any remanufac- turing Each major category is further grouped into many sub- categories and each sub-category includes many grades. Since the focus of this project is prototype development, it considers only grading rules for Select/Finish, Boards, and Dimension categories. Figure 4.1 show these categories and their sub-categories, and the grades selected for this project implementation. 22 Grading Category 8ub-categorv Grade B&BTR - l&2 CLEAR Selects C SELECT D SELECT Selects and Finish SUPERIOR FINISH h PRIME FINISH E FINISH Boards Dimension 1 COMMON 2 COMMON Commons 3 COMMON 4 COMMON 5 COMMON SELECI MERCHJNTABLE Alternate CONSTRUCTION Board STANDARD Grades UTILITY ECONOMY CONSTRUCTION Light STANDARD Framing UTILITY ECONOMY SELECT STRUCTURAL Structural NO. 1 Light NO. 2 Framing NO. 3 ECONOMY Stud STUD I ECONOMY Figure 4.1. Grading categories, sub-categories, and grades included in this project 23 Characteristics and limiting provisions for each grade include wood defects and manufacturing imperfections, and their type, size, location and distribution. Tables 4.2 to 4.8 give the characteristics for each grade as described by the rule book, and show the characteristics included in the scope of this project. There are two primary reasons for focusing on a subset of the grading rules. First, in many applications, not all characteristics listed in the rule book appear on lumber. At this initial development stage, only the most frequently occurring characteristics are included in the system. The focus was on conceptual design and testing the feasibility of the expert system approach. The system design is such that the knowledge base can be easily refined and extended. Second, the system is planned to be used as a component embedded into SAW3D (Zeng, 1991), a log breakdown optimization system. When SAW3D makes optimum breakdown decisions, manufacturing imperfections are not included. For example, warp, such as bow, crook, cup and twist, are not included. 24 Table 4.2. Grade characteristics considered by the project for grades in Select/Finish's Select sub- cateciory Character- istics Grade (Selects/Finish - Selects) B&BTR - l&2 CLEAR C SELECT D SELECT Stain Y Y Y Checks Y Y Y Splits N/A - Y N/A - Y Y Torn Grain N N N Skips N N N Cup N N N Crook N N N Wane Y Y Y Twist N N N Knots Y Y Y Holes N/A-Y Y Y Pitch Y Y Y Pockets Y Y Y Pitch Streak_____________ N/A N N Cutout N/A N/A N Y Listed in the rule book and considered by the project N : Listed in the rule book, but not considered by the project N/A : Not listed in the rule book and not considered by the project N/A - Y : Not listed in the rule book, but considered by the project 25 Table 4.3. Grade characteristics considered by the project for grades in Select/Finish's Finish sub- cateciory Character- istics Grade (Selects/Finish - Finish) SUPERIOR FINISH PRIME FINISH E FINISH Stain Y Y Y Checks Y Y Y Splits Y N/A - Y Y Torn Grain N N N Skips N N N Cup N N N Crook N N N Wane Y Y Y Twist N N N Knots Y Y Y Holes N/A-Y Y Y Pitch Y Y Y Pockets Y Y Y Pitch Streak_____________ N N N White Speck____________ N/A N N Cutout N N N Chip Marks N N N Y : Listed in the rule book and considered by the project N : Listed in the rule book, but not considered by the project N/A : Not listed in the rule book and not considered by the project N/A - Y : Not listed in the rule book, but considered by the project Table 4.4. Grade characteristics considered by the project for qrades in Boards' Commons sub-cateqory Character- istics Grade_(Boards - Commons) 1 COMMON 2 COMMON 3 COMMON 4 COMMON 5 COMMON Stain Y Y Y Y Y Checks Y y y y y Torn/Raised Grain_______ N N N N N Skips N N N N N Cup N N N N N Crook N N N N N Wane Y Y Y Y Y Twist N N N N N Splits y y y y y Streaks and Patches N/A N N N N Pitch Y Y Y Y y Pockets Y Y Y Y Y White Speck Honeycomb N/A N/A N N Y Unsound N/A N/A N N Y Shake N/A N N N N Roller Checks N/A N N/A N N Pith Y Y N/A N/A Y Knots Y Y Y Y Y Holes N/A-Y Y Y Y Y Y : Listed in the rule book and considered by the project N : Listed in the rule book, but not considered by the project N/A : Not listed in the rule book and not considered by the project N/A - Y : Not listed in the rule book, but considered by the project 27 Table 4.5. Grade characteristics considered by the project for grades in Boards' Alternate sub-category Character- istics Grade (Boards - Alternate) SEL- MERCH CONST STAND UTIL ECONOMY Stain Y Y Y Y Y Checks Y Y Y Y Y Torn/Raised Grain_______ N N N N N Skips N N N N N Cup N N N N N Crook N N N N N Wane Y Y Y Y Y Twist N N N N N Splits Y Y Y Y Y Streaks and Patches N N N N N Pitch Y N/A N/A N/A Y Pockets Y Y y y y White Speck Honeycomb________ N N/A N N Y Unsound Wood N N/A N/A N Y Shake N/A N/A N N N Roller Checks N N/A N/A N/A N Pith N/A N/A N/A N/A Y Knots Y Y Y Y Y Holes Y Y Y Y Y Burls N/A N/A N/A N N Y : Listed in the rule book and considered by the proj ect N : Listed in the the project N/A : Not listed in by the project rule book, but not considered by the rule book and not considered 28 Table 4.6. Grade characteristics considered by the project for grades in Dimensions' Light Framing sub- cateaorv Character- istics ______________ Grade (Dimension - Light Framing) CONST STAND UTIL ECONOMY Checks Y Y Y y Knots Y Y Y Y Holes Y Y Y Y Manufacture N N N N Pitch Y Y Y Y Pitch Streaks N N N N Pockets Y Y y y Shake N N N N Skips N N N Y Slopeof y y y y Splits Y Y Y Y Stain N N N N Wane Y Y Y Y Warp N N N N Unsound Wood N/A -Y Y I Y White Speck N/A N N N Y : Listed in the rule book and considered by the project N : Listed in the rule book, but not considered by the project N/A : Not listed in the rule book and not considered by the project N/A - Y : Not listed in the rule book, but considered by the project 29 Table 4.7. Grade characteristics considered by the project for grades in Dimensions' Structural Light Framing sub-category Character- istics Grade (Dimension - Structural Light Framing) SEL- STR NO.1 NO.2 NO.3 ECONOMY Checks Y Y Y Y Y Grain N N N N/A N Knots Y Y Y Y Y Holes Y Y Y Y Y Manufacture N N N N N Pitch Y Y Y Y Y Pitch Streaks N N N N N Pockets Y Y Y Y Y Shake N N N N N Skips N N N N N Slopeof Grain_______ Y Y Y Y Y Splits Y Y Y Y Y Stain N N N Y Y Wane Y Y Y Y Y Warp N N N N N Unsound Wood N/A-Y N/A-Y Y Y Y White Speck N/A N/A N N N Y : Listed in the rule book and considered by the project N : Listed in the rule book, but not considered by the project N/A : Not listed in the rule book and not considered by the project N/A - Y : Not listed in the rule book, but considered by the project 30 Table 4.8. Grade characteristics considered by the project for grades in Dimensions' Stud sub-category Character- istics Grade (Dimension - Stud) STUD ECONOMY Checks Y Y Grain N/A N/A Knots Y Y Holes Y Y Manufacture N N Pitch Y Y Pitch Streaks N N Pockets Y Y Shake N N Skips N N Slope of Grain Y Y Splits y y Stain N Y Wane Y Y Warp N N Unsound Wood Y y White Speck N N Y : Listed in the rule book and considered by the project N : Listed in the rule book, but not considered by the project N/A : Not listed in the rule book and not considered by the project N/A - Y : Not listed in the rule book, but considered by the project 31 4.2 Knowledge Acquisition and Representation 4.2.1 Knowledge Acquisition Knowledge acquisition deals with the issue of acquiring information about the problem domain. Knowledge used in problem solving may be based on experiences of one or more human experts, case histories, reference sources, etc. The key source of knowledge for this project is the grading rules. These are based upon the "Western Lumber Grading Rules 88" published by the Western Wood Products Association (1988). Compared with many other expert systems, the knowledge acquisition task for this project is relatively simple because most grading rules are well defined in the rule book. However, the rule book cannot give all the information required to perform the grading task accurately and efficiently. First, some of the rules are not explicitly defined. For example, certain grades allow "several" as the number of certain defects. However, "several" is not precisely defined. Second, the rule book just stipulates characteristics and limiting provisions, but not the grading procedures. In practice, grading procedures affect the speed with which lumber is graded. For instance, knots are the most frequently encountered defects. Looking for frequently encountered defects accelerates the grading process. Third, as mentioned 32 before, grading is more of an art than a science, and the results depend on the inspection and judgement of the grader. Five percent variation from standards is accepted as a reasonable difference between graders. Some mills and markets may use their local rules, and take advantage of the allowed variations. This may involve specific knowledge not documented in the rule book. Thus, the knowledge base using grading rules from the rule book can be refined based on actual grading practices. 4.2.2 Knowledge Representation The domain knowledge must be well organized and structured in the knowledge base, thus allowing the inference mechanism of the expert system to operate on it effectively and efficiently. There are many ways to represent domain knowledge, including production rules, frames, semantic networks, and first-order logic. Perhaps the most frequently used representation scheme is production rules. Production rules represent knowledge in the following form: IF (list of antecedents) AND OR NOT THEN (list of consequents or actions). 33 This format is almost identical to the description of grading rules stipulated in the rule book. Thus, the scheme of production rules is intuitively the best representation scheme for the domain knowledge for this project. The grading rules are of the following form: IF (list of characteristics permitted and limiting provisions) THEN Grade x. For example, consider the CONSTRUCTION grade in the Light Framing sub-category of dimension lumber. The characteristics permitted and limiting provisions for this grade are summarized in Table 4.9. These can be listed as the condition part of a production rule; the conclusion of the rule will be grade CONSTRUCTION if all the conditions are satisfied. However, such a single rule containing a description of all the characteristics will be inefficient because the inference process will be long. A better representation mechanism is to break down a single rule into several smaller rules where each component rule focuses on a distinct characteristic. Since the example in Table 4.9 has eleven components, the 11 rules shown in Table 4.10 may be used to represent the CONSTRUCTION grade. Finally, Rule 12 integrates the conclusions derived from 34 the individual sub-components. In this illustrative example, not all characteristics are included in the rules. For instance, spike knots and stains are not considered. Also, streaks and pockets are not listed because they have no impact on the grading result. Some rules, such as Rules 5, 7, and 11, need more details or other higher level rules to provide the antecedents. 35 Table 4.9. Characteristics permitted and limiting provisions for CONSTRUCTION grade in Light Framing sub-category of dimension lumber Character- Provisions istics Checks Surface seasoning checks not limited. Through checks at ends are limited as splits. Knots Sound, firm, encased and pith, must be tight and are permitted in the following sizes or their equivalent displacement: Unsound or Nominal Anywhere Loose Knots Width On Wide Face and Holes 2" 3/4" 5/8" 3" l&]./4" 3/4" 4" l&l/2" 1" Knots spiked entirely across the wide face are limited to a displacement of approximately 1/4 the cross section. Nanufacture Standard "E". Pitch and Not limited. pitch steaks Pockets, pitch Not limited. or bark Shake Several heart shakes up to 2' long, similar to seasoning checks, none through. Skips Hit and miss on 10% of the pieces. Slope of grain 1 in 6. Splits Equal in length to the width of the piece. Stain Stained sapwood. Firm heart stain or firm red heart. Wane 1/4 the thickness, 1/4 the width. 5% of the pieces may have wane up to 1/2 the thickness and 1/3 the width for 1/4 the length. Warp 1/2 of medium. Table 4.10. Grading rules for the CONSTRUCTION grade represented usinq production rules Rule 1: IF NO through checks at ends OR longest through check at end <= lumber width THEN Checks satisfy grade CONSTRUCTION Rule 2: IF NO firm knots on face OR (lumber width = 2" AND diameter of largest firm knot <= 0.75") OR (lumber width = 3" AND diameter of largest firm knot <= 1.25") OR (lumber width = 4" AND diameter of largest firm knot <= 1.5") THEN firm knots satisfy grade CONSTRUCTION Rule 3: IF NO loose knots on face OR (lumber width = 2" AND diameter of largest firm knot <= 5/8") OR (lumber width = 3" AND diameter of largest firm knot <= 3/4") OR (lumber width = 4" AND diameter of largest firm knot <= 1") THEN loose knots satisfy grade CONSTRUCTION Rule 4: IF NO holes on face OR (lumber width = 2" AND diameter of largest hole <= 5/8") OR (lumber width = 3" AND diameter of largest hole <= 3/4") OR (lumber width = 4" AND diameter of largest hole <= 1") THEN holes satisfy grade CONSTRUCTION Rule 5: IF manufacture standard is "E" THEN manufacture satisfies grade CONSTRUCTION Rule 6: IF NO shakes OR (NO through shakes AND length of largest shake <= 2') THEN shakes satisfy grade CONSTRUCTION 37 Table 4.10. (Continued) Rule 7: IF NO skips OR hit and miss on <= 10% pieces THEN skips satisfy grade CONSTRUCTION Rule 8: IF slope of grain <= 1/6" THEN slope of grain satisfies grade CONSTRUCTION RULE 9: IF NO splits OR length of longest split <= lumber width THEN splits satisfy grade CONSTRUCTION RULE 10: IF NO wane OR (thickness of wane <= 1/4 lumber thickness AND width of wane <= 1/4 lumber width) THEN wane satisfies grade CONSTRUCTION RULE 1.1: IF NO warp OR warp <= 1/2 medium THEN warp satisfies grade CONSTRUCTION RULE 12: IF checks satisfy grade CONSTRUCTION AND firm knots satisfy grade CONSTRUCTION AND loose knots satisfy grade CONSTRUCTION AND holes satisfy grade CONSTRUCTION AND manufacture satisfies grade CONSTRUCTION AND shakes satisfy grade CONSTRUCTION AND skips satisfy grade CONSTRUCTION AND slope of grain satisfies grade CONSTRUCTION AND splits satisfy grade CONSTRUCTION AND wane satisfies grade CONSTRUCTION AND warp satisfies grade CONSTRUCTION THEN grade is CONSTRUCTION 38 4e3 Inference Mechanism Given a knowledge base, an expert system performs inference or deduction to arrive at conclusions. For rule- base systems, the logic deduction is based on the following modus ponens: Given the rule: IF antecedent 1 AND antecedent 2 AND antecedent N THEN conclusion 1 and facts: antecedent 1 antecedent 2 antecedent N, it is possible to infer that conclusion 1 is true. Facts and rules are basic components of the knowledge base. The reasoning process involves looking for facts and using rules to arrive at conclusions. Inference strategies used in expert systems include backward chaining, forward chaining, breadth-first search, depth-first search, heuristic search, problem reduction, pattern matching, hierarchical control, unification, and event-driven control. Selection of a proper strategy for 39 an expert system depends on the knowledge representation scheme used in the system. Matching the task to the correct combination of knowledge representation and inference strategy is a complex task. For rule-based systems using production rules as the representation scheme, the major inference strategies used are backward chaining, forward chaining, or a combination of these two strategies. These strategies are primarily used to specify how rules contained in a knowledge-base are to be processed. Backward chaining is also called goal-driven inference. It begins with the selection of a goal using some selection criterion, then tries to prove that the goal is true. A goal is the conclusion of a rule. The process usually starts with matching existing facts with antecedents of that rule. If the required facts cannot be found, the inference engine then looks for rules that may have the required facts as their conclusions. This may direct the search process to sub-goals. If the required facts still cannot be found, the system may begin a query process if the facts can be obtained from the user. Otherwise the system will select another goal and repeat the process until a solution is found or there are no more goals to be checked. To illustrate how this strategy works, assume that there is a piece of lumber 2 inches in nominal thickness, 40 4 inches in width, and 10 feet long. There are no defects on its face except for a few well spaced firm knots. Assune that the largest knot has a diameter of 0.5 inches. The rules in Table 4.10 show that the grade of this piece of lumber is CONSTRUCTION in the light framing category of dimension lumber. The backward chaining process starts by selecting a grade from the 27 grades listed in Figure 4.1, and then tries to prove that this grade is the appropriate one. Assume that the grade selected is CONSTRUCTION. The system first looks for facts that are antecedents of Rule 12. Since these facts do not exist at this stage, it looks for rules 1 through 11 which have those facts as conclusions and tries to prove that rules 1 through 11 are true. To prove these rules, additional information is required. This can be obtained from the user, or by a search for more rules. Once rules 1 through 11 are proven to be true, all antecedents in Rule 12 are generated a.s new facts. This in turn proves that Rule 12 is true and the grading process is completed. If the selected initial grade is not CONSTRUCTION, the process may fail to find a solution; another grade may then be selected, and the process is repeated until a grade is found. This strategy is best suited for the problems with many initial facts but relatively few solutions. It has been successfully used for diagnostic problems. However, for the lumber grading problem, it is not the appropriate 41 method. One reason is that there are a large number of characteristics that need to be checked before a grade can be determined. It can be very inefficient if a wrong grade (a goal) is assumed. Another reason is that this inference mechanism is not analogous to the human grading process where conclusions are reached based on facts. The forward chaining inference mechanism, also called data-driven chaining or reasoning, starts with existing facts and works on deducing new facts that will eventually lead to the goal. The inference process works in the following cycle: 1) Matching rule(s): find all rules whose antecedents are satisfied by working storage (that is, where existing facts are stored). The collection of such rules is called the conflict set. 2) selecting rule(s): When more than one rule matches the facts, apply a conflict resolution strategy to determine which rule will be selected among those rules. Various conflict resolution strategies have been used for forward chaining systems, including: Do One: Select the first rule that matches the facts. Do All: Apply all the matching rules in one batch. 42 Do in Seauence: Apply the matching rules one at a time. Do the Most Specific: Select the more specific rule. Do the Most Recent: Select the one that matched a most recently derived fact. 3) Executing rule(s): the selected rule is proved, also called "fire the rule", and the working storage is updated with fact(s) of the newly fired rule. 4) The process is repeated until a solution is found or no more rules can be applied. For the lumber example used to illustrate backward chaining, the forward chaining process will start by collecting all the defect information, and storing it in the working storage (memory). It then tries to find all rules whose antecedents are matched with the existing facts. For the given data, it will conclude that antecedents of rules 1 through 11 exist in the working storage. Thus rules 1 through 11 are placed in the conflict set. Selection and firing of these 11 rules result in the facts that match the antecedents of Rule 12; these are then placed in the working storage. Now Rule 12 is also put into the conflict set because all of its antecedents are satisfied. Finally, firing Rule 12 results 43 in the CONSTRUCTION grade for the specified input data. The forward chaining procedure is best used to solve problems in which data is to be used as the starting point for problem solving. It is intuitively the most appropriate method for the lumber grading problem; it also matches the human grader's reasoning process. The inference used in this implementation is the forward chaining inference mechanism. 4.4 Tool Selection After identifying an appropriate knowledge representation scheme and inference mechanism, the next phase in the development process is to select the implementation software. Four options were considered for this project: conventional programming languages, knowledge engineering languages, knowledge-based programming tool kits, and knowledge-based programming shells. Conventional languages, such as C, C++ and Pascal, can be used to implement a knowledge-based system. They support a broad range of machine platforms. Compared with other options, they are very flexible and efficient in terms of execution speed and memory usage. The major disadvantage is that the development time is long requiring extremely large amount of coding. This option is obviously 44 not recommended for rapid prototyping and system design. Knowledge engineering languages, such as LISP and PROLOG, support symbolic manipulation and/or have built-in inference engines. They are very powerful and efficient when used on the right platforms such as LISP machines and work stations. However, the opposite is true on microcomputers. Since this project selected microcomputers as the implementation platform, this option also did not represent the best choice. Knowledge-based programming toolkits, such as KEE, OPS5, ART, Clips and Eclipse, are somewhat like high-level languages. While relieving the programmer from the burden of tedious coding by providing built-in knowledge representation schemes and inference engines, they do require some programming following their syntax so that maximum flexibility can be achieved. This option was considered appropriate for this project. Knowledge-based shells usually provide a complete development environments. The programmer does not need to do any real programming. However, they are usually expensive and have limited flexibility. This option was prohibitive due to the cost involved and the flexibility required by the project, particularly interaction with external systems. Several features were required of the development tool to be used. First, the tool should support rule-based 45 knowledge representation and a forward chaining inference mechanism. Second, the system should interact with C/C++ programs since the intended use of this product is to integrate it into SAW3D, a log breakdown optimization program written in C. Therefore, support of links between this project and any ANSI C/C++ compiler was considered to be an important feature of the development tool. Third, the tool should be run on an IBM-PC and support protected mode memory access. The knowledge base can potentially be so large that the 640k conventional memory limitation can become an obstacle. Fourth, the tool should be financially feasible. Other considerations included user interface and explanation facilities, ease of learning and use, documentation, and technical support. Considering the above factors, Eclipse, a commercial rule-based programming tool from The Haley Enterprise (The Haley Enterprise, 1991), was selected as the implementation tool for this project. Eclipse supports the features described above, including rule-based knowledge representation, a forward chaining inference engine, running under protected mode and integration with other applications. The syntax for Eclipse is derived from Inference Corporation' s ART (for Automatic Reasoning Tool), and is compatible with NASA'S Clips (for C Language Integrated Production System). The basic syntax uses LISP-like prefix 46 notation and parentheses. Eclipse uses the Rete Algorithm, which is one of the most efficient techniques for performing complex pattern matching of the type used in rule-based systems (Forgy, 1979, 1982). Although many other tools also use the Rete Algorithm, some benchmarks and applications have proven that Eclipse is superior to many other tools in terms of execution speed and program size2 (Brooke, 1992). For example, it runs four times as fast as Clips on 386 machines and requires roughly one half the memory of Clips. Eclipse supports more advanced functionality and better integration, especially on personal computers. It supports more powerful logical reasoning capabilities and provides more expressive pattern matching. One of Eclipse's best features is its integration capacity. It supports two-way integration, integrating other applications into Eclipse, and Eclipse into other applications. Other applications can be integrated into Eclipse by calling routines written in the C language, in some cases even Pascal or FORTRAN. Eclipse toolkits are provided with run-time libraries and source code necessary to construct applications which embed Eclipse. It allows declarative domain and procedural control knowledge to be explicitly separated and codified using rules or 2. The Haley Enterprise, Inc. 1992. Company brochure. Sewickley, PA. 47 procedures, retaining more of the advantages of a pure production system even when the requirement to procedurally control and embed a knowledge-base is implemented within any real application. The Microsoft Windows version of Eclipse has a real- time WYSIWYG (What You See Is What You Get) development interface. It provides knowledge base monitoring windows which can be used to simultaneously view the relations between facts and goals, and the match states of any number of rule sets, agenda, rules, joins, patterns, templates, relations, facts, matches, etc. Each of these views is supported in real time as commands or rules are executed. In effect, these modeless dialog windows allow full manipulation and monitoring of all knowledge base components, scheduling and pattern matching activities. The windows version also supports dynamic data exchange (DDE). This support provides more communication methods between Eclipse and other applications. Using this feature, Eclipse commands can be passed from other applications and Eclipse can issue commands to other applications. Eclipse can request and respond to requests for data from other applications. Dynamic link libraries (DLLs) are also provided in the Windows version. This research uses the Windows version of Eclipse. 48 4.5 Prototype Development Expert systems are best developed using an incremental prototyping methodology. A running prototype should be developed at an early stage. The prototype deals with a restricted portion of the problem domain; nonetheless, it will validate the feasibility and conceptual design of the system. After the prototype has been constructed, it can be extended into a fully functional system while its basic structure is preserved. 4.5.3. Prototype Components The prototype for this project has four basic components as shown in Figure 4.2. The user interface collects data from the user and provides the grading results. The inference engine is a built-in component of Eclipse. The knowledge base contains selected grading rules. The explanation facility consists of a set of rules for tracing back the facts upon which the grading results are based. At the beginning of system execution, the user loads the knowledge base, and then enters data describing lumber size and defect characteristics via the user interface. The inference engine tries to map the input data to one or more proper grades by performing the reasoning process on the input data and the knowledge base. Finally, the user 49 interface shows the final grades resulting from the matching process, along with explanations. The knowledge base is implemented in several source files. Each file contains grading rules for a sub-category of a category (see Table 4.11 for explanation). The user can load one or more of them, depending on the grading categories to be worked on. The user interface is implemented using rules and functions, and is contained in User Interface Inference Engine j Explanation Facility Knowledge Base Figure 4.2. Basic structure of the prototype 50 a separate file. Also, separate files contain output printing and some supplemental control strategies. Table 4.11 summarizes the file list used in the system. When using the Eclipse development toolkits, the development interface allows the system developer to watch how conclusions are reached and to monitor all facts in the knowledge base. Thus, an explanation facility is not necessary. However, the run-time version of the prototype does not include those knowledge-base debugging facilities. Table 4.11. Source files and their contents File Name Content Alt.clp Grading rules for ALTERMATE sub- category of BOARD Coinmons.clp Grading rules for COMMONS sub- category of BOARD Define.clp Definition of relations used to describe facts Finish.clp Grading rules for FINISH sub- category of SELECT/FINISH Lframe.clp Grading rules for LIGHT FRAMING sub-category of DIMENSION Query.clp Rules and functions for query interface. Select.clp Grading rules for SELECT sub- category of SELECT/FINISH Sframe.clp Grading rules for STRUCTURAL sub-category of DIMENSION Stud.clp Grading rules for STUD sub- category of DIMENSION System.clp Rules for system control, explanation, and output printing 51 Instead, a simple explanation facility is coded into a set of rules contained in the system.clp file to explain how the conclusions are reached. If used as an instruction tool, the prototype should be expanded by developing more detailed explanation features. 4.5.2 Knowledge Base Implementation The domain knowledge is described by objects, concepts, and relationships between them. The major objects in the problem domain are those defects considered by the system, including various types of stains, checks, splits, wane, knots, holes, pitch, and pockets. The major concepts are lumber grades. As discussed earlier, this system is a rule based system, in which the knowledge base consists of a collection of rules completely describing the relationships between the objects and concepts. In Eclipse syntax, the general format of a rule is as follows: (defrule rule_name "optional_comment" (pattern 1) (pattern 2) (pattern N) => (action 1) (action 2) (action 1.1)) 52 The first line is the rule header; "defrule" is a keyword indicating that this segment of code is a rule. Zero or more patterns are specified after the rule header. Each pattern consists of one or more fields. The patterns before the arrow symbol, "=>", are antecedents of an IF- THEN rule, and called the left-hand of the rule. The actions after "=>" are the THEN part of the IF-THEN rule, called the right-hand of the rule. Eclipse attempts to match the patterns of rules against existing facts. If all the patterns of a rule match facts, the rule is activated and put into an agenda (also called conflict set). The agenda is a collection of activated rules. Zero or more rules may be in the agenda. When more than one rule is in the agenda, Eclipse applies a conflict resolution procedure to select a rule to execute, called firing a rule. The actions following the arrow symbol will be executed when the rule in the agenda fires. Patterns, or antecedents of a rule are facts that are required to fire the rule. Facts are usually thought of as propositions, as in proposition calculus. A propositional fact is a fact that consists of a relation and one or more arguments following the relation. The relation is always in the first place in the fact with arguments following the relation. The relation describes how the argument relate to each other. For instance, a fact may be represented as (dia hole 0.5), where dia (for diameter) is the relation, hole and the 53 value of 0.5 are arguments. This pattern means that the hole has a diameter of 0.5 (inches). To illustrates how a rule is implemented in Eclipse, consider Rule 2 in Table 4.10. The rule is modified to accommodate knot description. As stated in Table 4.10, the rule only considers knot type (firm or loose), but not knot location (at edge or center line). However, some grades in the light framing sub-category need to consider both knot type and location. To keep the query interface consistent, Rule 2 is modified as follows: Modified Rule 2: IF NO firm knots at edge OR (lumber width = 2" AND diameter of largest firm knot at edge <= 0.75") OR (lumber width = 3" AND diameter of largest firm knot at edge <= 1.25") OR (lumber width = 4" AND diameter of largest firm knot at edge <= 1.5") THEN firm knots at edge satisfy grade CONSTRUCTION of Light Framing The rule is implemented in Eclipse as follows: (defrule CONST-LF-firin-edge-knot (declare (salience 1000)) (face ?) (category light-framing) (face_width ?width) (or (not (symptom firm_knot_at_edge)) (dia firm_knot_at_edge ?firiudia&: ( or (and (= ?width 2) (<= ?firmdia 0.75)) (and (= ?width 3) (<= ?firmdia 1.25)) (and (= ?width 4) (<= ?firmdia 1.5))))) => (assert (satisfy firm_edge_knot CONST-LF))) The name of this rule is CONST-LF-firiu-edge-knot. The patterns before "=>" look more complicated than the general 54 format described above because many Eclipse-defined functions are used. A function looks like a fact pattern. The entire function is surrounded by parentheses. The name of function precedes the arguments. The arguments for the function are patterns of facts or other functions. Here "declare", "or", "not", "and", and "assert" are names of functions which perform the functions implied by their names. Any field with the symbol ? means that value of this field is a variable. Comparing the modified Rule 2 and its Eclipse implementation, it becomes apparent that the first four patterns, or antecedents, within the implementation are not present in the modified Rule 2. These are supplemental patterns used for process control. The patterns after these process control patterns are just the Rule 2 expressed using Eclipse's syntax. An explanation of each component of the rule is given below. 1. (declare (salience 1000)): assigns a priority number to this rule. "Salience" is a keyword to indicate the priority of the rule, followed by a priority value. The system is driven by facts in the working memory. The system starts with matching the working memory facts with antecedents of rules. Once all antecedents of a rule are matched, that rule is put in the agenda. In many case more than one rule will be in the agenda waiting for firing. The system then applies a conflict resolution strategy to 55 fire those rules. However, sometimes priority ordering other than the Eclipse's predefined control strategy is necessary; "salience" orders the rule execution sequence. 2. (face ?): This has a special purpose. If both faces have the same defect characteristic described by a rule, without (face ?) once the first face has used in a rule, that rule will not be fired again for the second face, even if all the facts were present. This is due to the design of the Eclipse system. When grading different faces, ? in the pattern of (face ?) will have a different value. Therefore, even if no other facts in the rule change, the rule will still be activated because (face ?) pattern has changed. 3. (category light-framing): is used to prevent other grading categories from using this rule. This fact will be asserted into working memory during the query process when the user selects this category. Without using this pattern, Eclipse will try to match all antecedents of all grading rules, no matter what grading category the system is working on. That would be extremely inefficient. 4. (face_width ?width): is used to introduce variable ?width required in other patterns. 5. The patterns after (face_width ?width) and before "=>" are just statements of the modified Rule 2 in Eclipse syntax. They are written in compound logical "or", "not", and "and" format. The symbol "&:" has a special meaning. 56 The variable "?firmdia" before "&:" is tested in the calculations following "&:". For example, in the fact (dia firm_edge_knot 0.5), the ?firmdia is 0.5 If the lumber width is 2 inches, the ?firmdia will be tested to see whether the diameter of the knot is less than 0.75. The result will be true, and the system considers that the antecedent has been matched by the fact. If all the facts on the left-hand side of the rule are present in the working memory, the rule will be put on the agenda, and the new fact (satisfy firm_edge_knot CONST-LF) will be asserted into the working memory. 6. In the pattern of (satisfy firm_edge_knot CONST- LF), "satisfy" describes the relationship between the knot and the grade. Argument firm_edge_knot is an abstracted object, which is a combination of a knot object and its two attributes, type and location. Alternatively, the knot object can be represented using knot-type-value and knot- location-value triples, or using a knot template structure with type and location as its attributes. As a matter of fact, Eclipse has a "deftemplate" representation to support this kind of structure. Using deftemplate to represent a knot object was tried during prototyping, but was less preferable because of some inconvenience encountered in fact manipulation. All grading rules in the knowledge base have a similar structure. The query interface, system output, and system 57 controls are all implemented using such rules. To illustrate, Appendix A gives the Eclipse implementation of all grading rules for the Structural Light Framing category. 4.5.3 Grading Process, Input, and Output The system is menu-driven and interacts with the user through menus and a series of dialog boxes. The input process begins with loading the knowledge base and selecting the input method (interactively or loading a file containing facts). If the interactive input method is selected, a series of dialog boxes will guide the user to enter lumber size, select grade category and sub- categories, and specify the face where defects are located along with the defects' types, locations and sizes. The system then infers the grade corresponding to this face. Once the grade has been determined, the system allows the user to ask for explanation. The explanation provided by the system includes types of defects on the face, size of each defect, the grade each defect has satisfied, and the defects that determine the grade. The system determines the grade for each face separately and the overall grade. The flow chart in Figure 4.3 shows the grading process. Figures 4.4 to 4.15 explain the various menus and dialog boxes in details. Initially, the prototype only considered the two wide faces. After 58 Start Load knowledge base id fact file? Load fact file son andInput lumber dimension print output Select grade category End Select one or more subcategories Yes Work on facel? No Yes Work on face2 2 No No Find overall grade? End_D Reason Enter defect and characteristics print results chrocteristics Wont explona fect No from working Figure 4.3. Grading process (This prototype has been improved by including all four faces as described in Chapter 5) 59 prototype evaluation, the system was enhanced to consider all four faces (see Chapter 5). The system's initial window is shown in Figure 4.4. This window contains several pull-down menus for loading the knowledge base, cleaning memory space, and controlling system execution. The user can load a selected knowledge base into the system. Grading categories can be placed into one knowledge base, or several different knowledge bases, depending upon the user's preference. The development and maintenance of knowledge bases are done using Eclipse's development tool kits. Once the knowledge bases are error free, they can be saved as binary files. The run-time version of the system can only load knowledge bases in binary files. Binary files are efficient for fast loading and secure delivery. The system allows the user to input lumber characteristics in two ways: either by entering the facts during the query process, or by loading a fact file at the beginning of the session (Figure 4.5). Once the user selects to input facts interactively, a series of dialog boxes will pop up to guide the user to enter the facts. First, the user enters lumber thickness (Figure 4.6); this is followed by lumber width and length. Each input dialog box is followed by a confirmation dialog box (Figure 4.7) so that the user has a chance to modify the input data. 60 Next the system asks the user to select a grade category (Figure 4.8) and sub-categories in the category (Figure 4.9). Although the same piece of lumber may have more than one grade in different categories, the current version of the system does not consider grade overlap between different categories. However, the system allows the user to select one or all sub-categories in the category, that is, grade overlap is allowed between sub- categories. After selecting grade sub-categories, the user specifies the face of lumber to be graded (Figure 4.10) and then enters defects on this face. After a face has been graded, the dialog box in Figure 4.10 may again be used to specify another face. The overall grade is determined after both faces have been graded. Figures 4.1]. and 4.12 show the dialog boxes for entering defects. The defects shown in the figures include all defect types considered in the system. If the same defect type occurs more than once, only the worst defect is input. For example, if there are two holes, 0.1 inches and 0.5 inches in diameter, the larger hole (0.5 inches) is entered into the system. Once a defect type has been selected, a dialog box will pop up to ask for more detailed information about the defect type and location, as appropriate. Figure 4.13 shows a dialog box for entering information on knots. This 61 dialog box asks for knot type and location. The information in the dialog box may differ across grade categories as various categories consider knot types differently. The last piece of information required for a defect is its size, as shown in Figure 4.14. After all the defect data is entered, the user selects the no_more_defects item in Figure 4.12 to exit the input mode. The system maps the input data to appropriate grade(s). Figure 4.15 shows a an example output. The output includes each face grade based on the defect distribution on that face, the overall lumber grade, and the corresponding sub-category. Overlap between sub- categories is allowed; Figure 4.15 shows the grades for a piece of dimension lumber. {ile Lxecute stop help Figure 4.4. Opening window of the run-time version of the system Please select data input method load_tact_tile fact_tile_I oa d ed Figure 4.5. Dialog box for selecting input method enter lumber nominal thickness in inches Figure 4.6. Dialog box for entering lumber thickness Youve entered the following lumber size: 0 Nominal thIckness: 2 InNominal width: 4 InNomInal length: 8 ft Want to modify? I4. Figure 4.7. Dialog box for confirming lumber dimension 63 determine grade category sel cctlfinish jIiiIi4iE1[u1. boards Figure 4.8. Dialog box for selecting grade category determine dimension sub category light_framing structural_light_framing stud Figure 4.9. Dialog box for selecting Dimension sub- category 64 select face to work on Figure 4.10. Dialog box for selecting lumber face Select defect to enter holes wane splits checks grain pith Figure 4.11. Dialog box with partial defect list Select detect to enter grain pith decay stain pitch pockets Figure 4.12. Dialog box with another partial defect list select knot type and location fi rmkn ot_at_ce nte rI inc loose_knot_at_edge loose_knot_at_centerline Figure 4.13. Dialog box for entering knot type and location Figure 4.14. enter diameter of firm_knot_at_edge in inches o.1 Dialog box for entering knot diameter ace: Facel Grade: CONST Category: Light_framing ace: facel Grade: SEL-STR Category: Structural_light_Framing ace: Facel Crade: STUD Category: Stud ace: face2 Grade: NO.3 Category: Structural_light_framing ace: Face2 Grade: CONST Category: Light_Framing ace: Face2 Grade: STUD Category: Stud uerall Grade: CONST Category: Light_fraiinq uerall Grade: NO.3 Category: Structural_light_f raiiing uerall Grade: STUD Category: Stud Figure 4.15. System output display 67 CHAPTER 5 EVALUATION AND ENHANCEMENTS 5.1 System Evaluation The system was initially tested using generated input data. Simulated lumber and defect data were entered into the system. The resulting grades assigned by the system were compared with those assigned by using the rule book manually. All rules in the system were checked and 100 percent match was achieved for the simulated data. The test results are given in Appendix B. The system was then evaluated using real lumber. Eighty five samples of Siberian larch 2x4x12 lumber were used to evaluate the system. The samples had been graded by a human grader from the West Coast Lumber Inspection Bureau (WCLIB). All samples belong to the Structural Light Framing category. Among the five grades in this category, 20 sample pieces were randomly selected from 61 pieces of Select Structural lumber, 20 from 31 pieces of No.1, 20 from 53 No.2, 20 from 73 pieces of No.3, and 5 from 10 Economy. To test the performance of the system, defects on these pieces were measured by the author. These were then entered into the system. As shown in Figure 5.1, 60 percent of the samples (51 pieces) matched the original 0, 0 E .4- 0 I- 0) E z 68 Matched UM-1 UM-2 UM-3 UM-4 UM-5 UM-6 Before Modification After Modification Figure 5.1. Evaluation results considering all factorsgrades as judged by the human grader. A total of 31 pieces (36.5 percent) did not match the original grades because of limitations of the computer system, primarily not considering defects on the two narrow faces (edges), shake, warp, and manufacturing imperfections. An additional 3.5 percent samples (3 pieces) did not match the grades assigned because of measurement bias or the mismatch could not be explained. For example, a firm knot at edge on a piece was measured as being 1.1 inches in diameter, which made this piece No.2. However, for some reason the human grader assigned Select Structural grade to the piece. Since the human grader was not present, it was difficult to correlate some of the mismatched observations. If used as a sub-system of SAW3D, warp and manufacturing imperfections do not exist when SAW3D making decisions. If these defects are excluded, the matched rate between the grade assigned by the human grader and that assigned by the system is 72.9 percent. Although some marginal errors are introduced due to measurement bias, a difference of 40 percent requires modifications of the system to improve its performance. 5.2 System Improvement As shown in Figure 5.1, the largest cause of deviation between the grades assigned by the computer and those assigned by the human grader was exclusion of defects on the two narrow faces. Thus the focus of improvement was on including the narrow faces in the evaluation process. The system was improved to allow defects on narrow faces to be input and evaluated. The enhanced model allows rLi the user to input defects on all four faces, assigns a grade to each of the four faces, and then determines the overall grade. The eighty five samples tested earlier were again evaluated using the revised system. Figure 5.1 shows the grading results after the modification. The percentage of matched grades increased from 60 to 76.5 percent. The samples that did not match due to system limitations decreased from 36.5 to 17.6 percent. Samples which were assigned a grade different from that assigned by the human grader because of measurement error and unexplained differences increased from 3.5 to 59 percent. All 15 unmatched samples caused by system limitations are upgraded in the revised system because defects such as shakes, warp and manufacturing imperfections are not considered by the system. All five unmatched samples caused by measurement difference are downgraded because defect measurements are larger than those permitted by the original grades. Although warp and manufacturing imperfections also contributed to deviations in grade assignment, these are introduced as a result of the manufacturing process and do not exist before log breakdown. As such, they are not considered in a log breakdown model such as SAW3D. If the system is to be used for a log breakdown optimization system where warps and manufacturing imperfections are not 7]- present, the percentage of matched samples increases to 72.9 percent before the modification, and to 89.4 percent after the modification (Figure 5.2). However, if the system is to be used as an instruction tool or a component of an automatic grading system, warp and manufacturing imperfection are important factors that affect grading accuracy, and therefore need to be included in the system. The bias due to measurement error can be eliminated by having the individual who visually graded the samples also a, w a- E Ct 0) 1- 0 -a E z Matched UM-1 UM- UM-O Before Modification After Modification if no measure bias Figure 5.2. Evaluation results when warp, manufacturing imperfections, and measurement bias were not considered 72 measure the defects. In this case the matched samples represent 95.3 percent of the total boards (Figure 5.2). The nuiber of unmatched samples due to shake and not-well- spaced knots is very small. These may also depend upon wood species; inclusion of these factors will further improve the system accuracy. 73 CHAPTER 6 CONCLUSIONS A prototype expert system for softwood lumber grading was developed. The prototype can grade lumber for 27 grades in Dimension, Select/Finish, and Boards category. It can judge each of four faces of lumber individually and assign an overall grade based on the grades of the individual faces. The user interface helps to obtain facts about a piece of lumber, and provides grading results and explanations about how the conclusions are reached. The initial prototype considered only two faces of lumber. After an evaluation using real lumber samples, the prototype was modified to include all four faces resulting in significant performance improvement. The intended use of the system is in conjunction with SAW3D, a log breakdown optimization system, to determine optimum log breakdown decisions. SAW3D will provide this grading system with lumber size and defect characteristics information. The expert system determines the lumber grade and will then pass this information back to SAW3D where SAW3D will determine the lumber value based on the grade, and use it to arrive at optimum log breakdown decisions. This application represents a new direction for improving lumber production efficiency: optimization of cutting strategies using both lumber geometric and defect 74 information. The grading system can be used to build and maintain grading rules independently, and communicate with other components of the overall system in different ways. It can be embedded in the overall system, or it can exchange data with the system using Dynamic Data Exchange facility supported by MS-WINDOWS. The separation of the grading system from other parts of the overall system is very important. The knowledge base can be developed and maintained without interfering with any other program logic. This provides greater portability and flexibility of the grading system, and in turn simplifies the development and maintenance of the system. Areas for further research include: 1. The evaluation using real lumber was only performed for the Structural Light Framing sub-category of the Dimension. Such a evaluation should be done for all grading categories. 2. For rapid prototyping, the development concentrates on selected defect characteristics for each grade. The selected characteristics are clearly defined and described in the rule book, and are usually the most frequently encountered defects on softwood lumber. Once the specific use of the system is determined, the knowledge base should be refined and more rules should be added to the knowledge base with the help of human grading experts. For example, 75 knot positions relative to each other, shake, warp and manufacturing imperfections need to be considered. Clustering of defects is another important feature that may effect a grade, and thus may need to be included in the knowledge base. 3. As mentioned before, this system is to be used as a sub-system of SAW3D. The system may also be used as an instruction or training tool. When used as an instruction tool, a better user interface should be developed. A help facility should be added to guide the user at each step, and more detailed explanations should also be provided at each step. A graphic display of defect distribution and grade stamp bitmaps may be very helpful. The system may need to include some controls allowing the user to grade the lumber, and then judge the results using the knowledge base. 76 BIBLIOGRAPHY 1. American Lumber Standards Committee. 1970. American Softwood Lumber Standard. National Bureau of Standards, U.S. Department of Commerce, Washington, D.C. 2. Aune, Jan. 1991. X-Ray edger optimizer makes money at Macmillan Bloederl's Alberni Pacific Division. Proceedings of the 4th International Conference on Scanning Technoloav in the Wood Industry, pp. Aune-1 - Aune-2, San Francisco, CA. 3. Brooke, Tom. 1992. The art of production systems. Expert, 7(1):31-35. 4. Flatman, Carl, and Egan Bodell. 1989. Computer-aided lumber grader-optimizer system. Proceedings of the Third International Conference on Scanning Technology in Sawmilling, pp. XIX-1 to XIX-15, San Francisco, CA. 5. Flatman, Carl, and Daniel J. Kenway. 1991. Mill experience with high speed grade based optimizers. Proceedings of the 4th International Conference on Scanning Technolov in the Wood Industry, pp. kenway-1 to kenway-l4, San Francisco, CA. 6. Forgy, L. Charles. 1979. On the efficient implementation of production systems. Ph.D Thesis. Carnegie-Mellon University. 7. Forgy, L. Charles. 1982. Rete: a fast algorithm for the many pattern/many object match problem. Artificial Intelligence, 19(1) :17-37. 8. Haley Enterprise. 1992. Eclipse Reference Manual. The Haley Enterprise, Inc., Sewickley, PA. 9. Hallock, H., and L. Gauger. 1971. Grading hardwood lumber by computer. Res. Paper FPL-157. Forest Products Laboratory, Madison, WI. 10. Huang, S.S.L., and F.T. Sparrow. 1989. A computer- aided instruction tool for grading hardwood lumber. Forest Products Journal, 39(10) :39-42. 11. Kenway, J. Daniel. 1990. A supercomputer-based machine vision grading system and trimmer optimizer for dimension lumber. Proceedings of Process Control/Production Manaaement of Wood Products: Technoloav for the '90s, pp. 53-65, Athens, Georgia. 77 12. Klin]thachorn, P., J.P. Franklin, C.W. McMillin, R.W. Conners, and H.A. Huber. 1988. Automated computer grading of hardwood lumber. Forest Products Journal, 38(3) :67-69. 13. Klinkhachorn, P., C.J. Schwehiu, CW. McMillin, and H.A. Huber. 1989. HaLT: a computerized training program for hardwood lumber grading. Forest Products Journal, 39(2) :38-40. 14. Klinkhachorn, P., D. Yost, R. Kothari, C.W. McMillin, H.A. Huber. 1991. Computer lumber grading and HaRem for an automated lumber processing system. Proceedinas of the First International Conference on Automated Lumber Processing Systems and Laser Machining of Wood, pp. 173-181, Michigan State University, East Lansing, MI. 15. Western Wood Products Association. Western lumber grading rule 88. 1988. Portland, OR. 16. Williston, Edward M. 1988. Lumber manufacturing - The design and operation of sawmills and planer mills. Miller Freeman Publications, Inc. 17. Zeng, Y. 1991. Log breakdown using dynamic programming and 3-D log shape. M.S. thesis. Oregon State University, Corvallis, OR. APPENDICES 78 Appendix A. Source Code for Structural Light Framing Category , , , ------------------------------------------------------ ; This file contains rules for Structural Light Framing ; category of Dimension lumber grades. This category ; contains the following four grades: 1) SEL-STR (Select Structural, Structural Light Framing) ; 2) NO.1 (NO. 1, Structural Light Framing) 3) NO.2 (NO. 2, Structural Light Framing) 4) NO.3 (NO. 3, Structural Light Framing) 5) ECONOI1Y-SLF (Economy, Structural Light Framing) Validation of defects for each grade is checked by the ; rules separated from the grading rule. SE L-S TR GRADE ;----Rules to check validation of defects for SEL-STR ----- (defrule SEL-STR_firm_edge_knot "Firm knots at edge <= SEL-STR limit?" (declare (salience 1000)) (face ?) (category structural_light_framing) (face_width ?width) (or (not (symptom firm_knot_at_edge)) (dia firm_knot_at_edge ?firmdia&:( or (and (= ?width 2) (<= ?firmdia 0.375)) (and (= ?width 3) (<= ?firmdia 0.5)) (and (= ?width 4) (<= ?firmdia 0.75))))) => (assert (satisfy firm_knot_at_edge SEL-STR))) (defrule SEL-STR_firm_center_knot "Firm knots at center <= SEL-STR limit?" (declare (salience 1000)) (face ?) (category structural_i ight_f raming) (face_width ?width) (or (not (symptom firm_knot_at_centerline)) (dia firm_knot_at_centerline ?firmdia&:( or (and (= ?width 2) (<= ?firmdia 0.375)) (and (= ?width 3) (<= ?firmdia 0.5)) (and (= ?width 4) (<= ?firmdia 0.875))))) (assert (satisfy firm_knot_at_centerline SEL-STR))) 79 (defrule SEL-STR_loose_edge_knot "Loose edge knots <= SEL-STR grade limit?" (declare (salience 1000)) (face ?) (category structural_i ight_f raining) (face_width ?width) (or (not (symptom loose_knot_at_edge)) (dia loose_knot_at_edge ?loosedia&: ( or (and (= ?width 2) (<= ?loosedia 0.375)) (and (= ?width 3) (<= ?ioosedia 0.5)) (and (= ?width 4) (<= ?loosedia 0.75))))) => (assert (satisfy loose_knot_at_edge SEL-STR))) (defrule SEL-STR_loose_center_knot "Loose cnt knots <= SEL-STR grade limit?" (declare (salience 1000)) (face ?) (category structural_i ight_f raining) (face_width ?width) (or (not (symptom loose_knot_at_centerline)) (dia loose_knot_at_centerline ?loosedia&: ( or (and (= ?width 2) (<= ?loosedia 0.375)) (and (= ?width 3) (<= ?loosedia 0.5)) (and (= ?width 4) (<= ?loosedia 0.75))))) => (assert (satisfy loose_knot_at_centerline SEL-STR))) (defrule SEL-STR_hoie "Holes <= SEL-STR grade limit ?" (declare (salience 1000)) (face ?) (category structural_i ight_f raining) (face_width ?width) (or (not (symptom holes)) (dia hole ?holedia&:( or (and (= ?width 2) (<= ?holedia 0.375)) (and (= ?width 3) (<= ?holedia 0.5)) (and (= ?width 4) (<= ?holedia 0.75))))) => (assert (satisfy holes SEL-STR) )) (defrule SEL-STR_wane "Wanes <= SEL-STR grade limit ?" (declare (salience 1000)) (face ?) (category structural_light_framing) (lumber thickness ?thickness) (lumber_width ?width) (or (not (symptom wane)) (and (thickness wane ?thx&: ( <= 4.0))) (width wane ?wid&: ( <= ?wid => (assert (satisfy wane SEL-STR) )) ?thx (/ ?thickness (/ ?width 4.0))))) (defrule SEL-STR_split "Splits <= SEL-STR grade limit ?" (declare (salience 1000)) (face ?) (category structural_light framing) (lumber_width ?width) (or (not (symptom splits)) (length split ?splitlength&:(<= (- ?splitlength ?width) 0.000001))) => (assert (satisfy splits SEL-STR) )) I (defrule SEL-STR_check "Checks <= SEL-STR grade limit ?" (declare (salience 1000)) (face ?) (category structural_i ight_f raming) (lumber_width ?width) (or (not (symptom through_check_on_end)) (length through_check_on_end ?checklength&: (<= (- ?checklength ?width) 0.000001))) (assert (satisfy through_check_on_end SEL-STR) )) (defrule SEL-STR_slope "Slope within SEL-STR grade limit ?" (declare (salience 1000)) (face ?) (category structural_i ight_f raining) (or (not (symptom grain)) (slope_of_grain ?slope&:( <= ?slope (/ 1 12.0)))) => (assert (satisfy grain SEL-STR) )) Grading rule for grade of SEL-STR ----------- (defrule grade_SEL-STR "SEL-STR Grade" (declare (salience 900)) 81 (not (symptom decay)) (satisfy firm_knot_at_edge SEL-STR) (satisfy firm_knot_at_centerline SEL-STR) (satisfy loose_knot_at_edge SEL-STR) (satisfy loose_knot_at_centerline SEL-STR) (satisfy holes SEL-STR) (satisfy wane SEL-STR) (satisfy splits SEL-STR) (satisfy through_check_on_end SEL-STR) (satisfy grain SEL-STR) => (assert (grade Structural_light_framing SEL-STR 1))) NO. ]. GRADE Rules to check validation of defects for NO.1 ------ (defrule NO.1_firm_edge_knot "Firm knots at edge <= NO.1 grade limit ?" (declare (salience 800)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (face_width ?width) (or (not (symptom firm_knot_at_edge)) (dia firm_knot_at_edge ?knotdia&:( or (and (= ?width 2) (<= ?knotdia 0.5)) (and (= ?width 3) (<= ?knotdia 0.75)) (and (= ?width 4) (<= ?knotdia 1.0))))) => (assert (satisfy firm_knot_at_edge NO.1))) (defrule NO. 1_firm_center_knot "Firm knots at center <= NO.1 grade limit ?" (declare (salience 800)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (face_width ?width) (or (not (symptom firm_knot_at_centerline)) (dia firm_knot_at_centerline ?knotdia&: ( or (and (= ?width 2) (<= ?knotdia 0.5)) (and (= ?width 3) (<= ?knotdia 0.75)) (and (= ?width 4) (<= ?knotdia 1.5))))) => (assert (satisfy firm_knot_at_centerline NO.1))) (defrule NO.1_loose_edge_knot "Loose edge knots <= NO.1 82 grade limit ?" (declare (salience 800)) (face ?) (category structural 1 ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (face width ?width) (or (not (symptom loos (dia loose_knot_at (and (= ?width (and (= ?width (and (= ?width s_knot_at_edge)) edge ?knotdia&: ( or 2) (<= ?knotdia 0.5)) 3) (<= ?knotdia 0.75)) 4) (<= ?knotdia 1.0))))) => (assert (satisfy loose_knot_at_edge NO.1))) (defrule NO.1_loose_center_knot "Loose center knots <= NO.1 grade limit ?" (declare (salience 800)) (face ?) (category structural_light_framing) (not (grade Structural_light_framing SEL-STR 1)) (face_width ?width) (or (not (symptom loose_knot_at_centerline)) (dia loose_knot_at_centerline ?knotdia&: ( or (and (= ?width 2) (<= ?knotdia 0.5)) (and (= ?width 3) (<= ?knotdia 0.75)) (and (= ?width 4) (<= ?knotdia 1.0))))) => (assert (satisfy loose_knot_at_centerline NO.1))) (defrule NO. 1_hole "Holes <= NO.1 grade limit ?" (declare (salience 800)) (face 7) (category structural_light_framing) (not (grade Structural_light_framing SEL-STR 1)) (face_width ?width) (or (not (symptom holes)) (dia hole ?holedia&:( or (and (= 7width 2) (<= ?holedia 0.5)) (and (= ?width 3) (<= ?holedia 0.75)) (and (= ?width 4) (<= ?holedia 1.0))))) (assert (satisfy holes NO.1))) (defrule NO.1_wane "wanes <= NO.1 grade limit ?" 83 (declare (salience 800)) (face ?) (category structural_light_framing) (not (grade Structural_light_framing SEL-STR 1)) (lumber_thickness ?thickness) (lumber_width ?width) (or (not (symptom wane)) (and (thickness wane ?thx&: ( <= ?thx (/ ?thickness 4.0))) (width wane ?wid&:( <= ?wid (/ ?width 4.0))))) => (assert (satisfy wane NO.1))) (defrule NO.1_split "splits <= NO.1 grade limit ?" (declare (salience 800)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (lumber_width ?width) (or (not (symptom splits)) (length split ?splitlength&:(<= (- ?splitlength ?width) 0.000001))) => (assert (satisfy splits NO.1))) (defrule NO.1_check "checks <= NO.1 grade limit ?" (declare (salience 800)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (lumber_width ?width) (or (not (symptom through_check_on_end)) (length through_check_on_end ?checklength&: (<= (- ?checklength ?width) 0.000001))) => (assert (satisfy through_check_on_end NO.1))) (def rule NO.1_slope "slope is within NO.1 grade limit ?" (declare (salience 800)) (face ?) (category structural_light_framing) (not (grade Structural_light_framing SEL-STR 1)) (or (not (symptom grain)) 84 (Slope_of_grain ?slope&:(<= ?slope (1 1 10.0)))) => (assert (satisfy grain NO.1))) Grading rule for grade of NO.]. -------------- (defrule grade NO.1 "NO.1 Grade" (declare (salience 700)) (not (symptom decay)) (satisfy firm_knot_at_edge NO.].) (satisfy firm_knot_at_centerline NO.1) (satisfy loose_knot_at_edge NO.].) (satisfy loose_knot_at_centerline NO.].) (satisfy holes NO.].) (satisfy wane NO.1) (satisfy splits NO.].) (satisfy through_check_on_end NO.1) (satisfy grain NO.].) => (assert (grade Structural_light_framing NO.1 2))) NO . 2 GRADE ; ------ Rules to check validation of defects for NO.2 (defrule No.2_firm_edge_knot "Firm knots at edge <= NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (face_width ?width) (or (not (symptom firm (dia firm_knot_at_ (and (= ?width (and (= ?width (and (= ?width _knot_at_edge)) dge ?knotdia&:( or 2) (<= ?knotdia 0.625)) 3) (<= ?knotdia 0.875)) 4) (<= ?knotdia 1.25))))) => (assert (satisfy firm_knot_at_edge NO.2))) (defrule NO.2_firm_center_knot "Firm knots at center <= NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.]. 2)) (face_width ?width) (or (not (symptom firm_knot_at_centerline)) (dia firm_knot_at_ (and (= ?width (and (= ?width (and (= ?width => (assert (satisfy firm enter1ine ?knotdia&:( or 2) (<= ?knotdia 0.625)) 3) (<= ?knotdia 0.875)) 4) (<= ?knotdia 2.0))))) _knot_at_centerline NO.2))) 85 (defrule No.2_loose_edge_knot "Loose edge knots <= NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_i ight_framing) (not (grade Structural (not (grade Structural (face_width ?width) (or (not (symptom loos (dia loose_knot_at (and (= ?width (and (= ?width (and (= ?width light_framing SEL-STR 1)) _light_framing NO.1 2)) s_knot_at_edge)) _edge ?knotdia&:( or 2) (<= ?knotdia 0.625)) 3) (<= ?knotdia 0.875)) 4) (<= ?knotdia 1.25))))) (assert (satisfy loose_knot_at_edge NO. 2))) (defrule NO.2_loose_center_knot "Loose cnt knots <= NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_light_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (face_width ?width) (or (not (symptom loose_knot_at_centerline)) (dia loose_knot_at_centerline ?knotdia&: ( or (and (= ?width 2) (<= ?knotdia 0.625)) (and (= ?width 3) (<= ?knotdia 0.875)) (and (= ?width 4) (<= ?knotdia 2.0))))) => (assert (satisfy loose_knot_at_centerline NO. 2))) (defrule NO.2_hole "Holes <= NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (face_width ?width) (or (not (symptom holes)) (dia hole ?holedia&:( or (and (= ?width 2) (and (= ?width 3) (and (= ?width 4) (assert (satisfy holes NO.2))) I (defrule NO. 2_wane I 86 (<= ?holedia 0.625)) (<= ?holedia 0.875)) (<= ?holedia 1.25))))) "wanes <= NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_i ight_f raining) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (lumber_thickness ?thickness) (lumber_width ?width) (or (not (symptom wane)) (and (thickness wane ?thx&: ( <= ?thx (/ ?thickness 3.0))) (width wane ?wid&: ( <= ?wid (/ ?width 3.0))))) => (assert (satisfy wane NO.2))) (defrule NO.2_split "splits <= NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_i ight_f raining) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (lumber_width ?width) (or (not (symptom splits)) (length split ?splitlength&: (<= ?splitlength (* 1.5 ?width)))) => (assert (satisfy splits NO.2))) (defrule NO.2_check "checks <= NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_light_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) 87 (lumber_width ?width) (or (not (symptom through_check_on_end)) (length through_check_on_end ?checklength&: (<= ?checklength (* 1.5 ?width)))) (assert (satisfy through_check_on_end NO. 2))) (defrule NO.2_slope "slope is within NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_light_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (or (not (symptom grain)) (slope_of_grain ?slope&:(<= ?slope (/ 1 8.0)))) => (assert (satisfy grain NO.2))) (defrule NO.2_decay "decay <= NO.2 grade limit ?" (declare (salience 600)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (lumber_thickness ?thickness) (lumber_width ?width) (or (not (symptom decay)) (and (test (= ?thickness 2 )) (thickness decay ?thx&: ( <= ?thx (I ?thickness 3.0))) (width decay ?wid&: (<= ?wid (I ?width 3.0))))) (assert (satisfy decay NO.2))) (defrule gra (declare (satisfy (satisfy (satisfy (satisfy (satisfy (satisfy (satisfy Grading rule for grade of NO.2 ------------- e NO.2 "NO.2 Grade" (salience 500)) firm_knot_at_edge NO.2) firm_knot_at_centerline NO.2) loose_knot_at_edge NO.2) loose_knot_at_centerline NO.2) holes NO.2) wane NO.2) splits NO.2) (satisfy through_check_on_end NO.2) (satisfy grain NO.2) 88 (satisfy decay NO.2) => (assert (grade Structural_light_framing NO.2 3))) NO. 3 GRADE Rules to check validation of defects for NO.3 (defrule NO.3_firm_edge_knot "Firm knots at edge <= NO.3 grade limit ?" (declare (salience 400)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (face_width ?width) (or (not (symptom firm_knot_at_edge)) (dia firm_knot_at_edge ?knotdia&:( or (and (= ?width 2) (<= ?knotdia 0.75)) (and (= ?width 3) (<= ?knotdia 1.25)) (and (= ?width 4) (<= ?knotdia 1.75))))) => (assert (satisfy firm_knot_at_edge NO.3))) (defrule NO.3_firm_center_knot "Firm knots at center <= NO.3 grade limit ?" (declare (salience 400)) (face ?) (category structural_light_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (face_width ?width) (or (not (symptom firm_knot_at_centerline)) (dia firm_knot_at_centerline ?knotdia&: ( or (and (= ?width 2) (<= ?knotdia 0.75)) (and (= ?width 3) (<= ?knotdia 1.25)) (and (= ?width 4) (<= ?knotdia 2.5))))) => (assert (satisfy firm_knot_at_centerline NO.3))) (defrule NO.3_loose_edge_knot "Firm knots at edge <= NO.3 grade limit ?" (declare (salience 400)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (face_width ?width) (or (not (symptom loose_knot_at_edge)) (dia loose_knot_at_edge ?knotdia&: ( or (and (= ?width 2) (<= ?knotdia 0.75)) (and (= ?width 3) (<= ?knotdia 1.25)) (and (= ?width 4) (<= ?knotdia 1.75))))) (assert (satisfy loose_knot_at_edge NO.3))) 1 89 (defrule NO.3_loose_center_knot "Firm cntknots at edge <= NO.3 grade limit?" (declare (salience 400)) (face ?) (category structural_i ight_f raining) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (face_width ?width) (or (not (symptom loose_knot_at_centerline)) (dia loose_knot_at_centerline ?knotdia&: ( or (and (= ?width 2) (<= ?knotdia 0.75)) (and (= ?width 3) (<= ?knotdia 1.25)) (and (= ?width 4) (<= ?knotdia 2.5))))) (assert (satisfy loose_knot_at_centerline NO.3))) I (defrule NO.3_hole "Holes <= NO.3 grade limit ?" (declare (salience 400)) (face ?) (category structural_light_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (face_width ?width) (or (not (symptom holes)) (dia hole ?holedia&:( or (and (= ?width 2) (<= ?holedia 0.75)) (and (= ?width 3) (<= ?holedia 1.25)) (and (= ?width 4) (<= ?holedia 1.75))))) => (assert (satisfy holes NO.3))) (defrule NO.3_wane "wanes <= NO.3 grade limit ?" (declare (salience 400)) 90 (face ?) (category structural 1 ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (lumber_thickness ?thickness) (lumber width ?width) (or (not (symptom wane)) (and (thickness wane ?thx&: ( <= ?thx (/ ?thickness 2.0))) (width wane ?wid&: ( <= ?wid (/ ?width 2.0))))) => (assert (satisfy wane NO.3))) (defrule NO.3_split "splits <= NO.3 grade limit ?" (declare (salience 400)) (face ?) (category structural_light_framing) (lumber_length ?length) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (lumber_width ?width) (or (not (symptom splits)) (length split ?splitlength&:(<= ?splitlength (/ (* 12 ?iength) 6.0)))) => (assert (satisfy splits NO.3))) (defrule NO.3_check "checks <= NO.3 grade limit ?" (declare (salience 400)) (face ?) (category structural_i ight_framing) (lumber_length ?length) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (lumber_width ?width) (or (not (symptom through_check_on_end)) (length through_check_on_end ?checklength&: (<= ?checklength (/ (* 12 ?length) 6.0)))) => (assert (satisfy through_check_on_end NO.3))) (defrule NO.3_slope "slope is within NO.3 grade limit ?" (declare (salience 400)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (or (not (symptom grain)) (slope_of_grain ?siope&:(<= ?slope (/ 1 4.0)))) => (assert (satisfy grain NO.3))) (defrule NO.3_decay "decay <= NO.3 grade limit ?" (declare (salience 400)) (face ?) (category structural_i ight_framing) (not (grade Structural_light_framing SEL-STR 1)) (not (grade Structural_light_framing NO.1 2)) (not (grade Structural_light_framing NO.2 3)) (lumber_cross_section ?iumberCS) (or (not (symptom decay)) (cross_section decay ?decaySize&: (<= ?decaySize (I ?lumberCS 3.0)))) => (assert (satisfy decay NO.3))) Grading rule for grade of NO.3 ------------ (defrule gra (declare (satisfy (satisfy (satisfy (satisfy (satisfy (satisfy (satisfy le_NO.3 "NO.3 Grade" (salience 300)) firm_knot_at_edge NO.3) firm_knot_at_centerline NO.3) loose_knot_at_edge NO.3) loose_knot_at_centerline NO.3) holes NO.3) wane NO.3) splits NO.3) (satisfy through_check_on_end NO.3) (satisfy grain NO.3) (satisfy decay NO.3) => (assert (grade Structural_light_framing NO.3 4))) ECONOMY -SLF GRADE ================== Grading rule for grade of ECONOMY-SLF ---------- (defrule grade_ECONOMY-SLF "ECONONY-SLF Grade" (declare (salience 200)) (face ?) (category structural_i ight_f: (not (grade Structurai_iight (not (grade Structural_light (not (grade Structurai_iight (not (grade Structural_light (not (symptom broken ends )) => raming) _framing SEL-STR 1)) framing NO.1 2)) framing NO.2 3)) _framing NO.3 4)) 92 (assert (grade Structural_light_framing ECONOMY 5))) ;======== END OF STRUCTURAL LIGHT FRAMING GRADES 93 Appendix B. Test Results Using Generated Data Tables Bi, B3, and B5 show the grade specifications. Data in each column in the upper section of these tables represent the limiting provisions on defects; the grades are listed in the lower section. An X in the lower section means that data in the column are limiting provisions for the grade in the row. The grading rules for each grade in the knowledge base were tested using facts entered through the interactive query process. These facts represent realistic data values. Tables B2, B4, and B6 show the resulting grades. These tables use a format similar to the rule specification tables (Tables Bi, B3, and B5). Each column contains data for a face of a lumber sample. Data in the top half of the table is the generated defect measurements (facts). Empty cells in a column reflect defects that do not exist in a sample. As in the specification tables, column and row intersections mean that the data (facts) in the column satisfies the grade in the row. Multiple Xs in a column means that the input data satisfy more than one grade. For example, a lumber sample with defects indicated in the first column of Table B2 satisfies CONST, NO.3, and STUD grades. 94 Table Bi. Specifications of grades for a piece of lumber in the Dimension category (lumber size = 2x4txlO) )efect Fype/Location easure imiting Provisions <= inches) firm-edge a____ 1 . 0.75 T 1 not firm-center a_____ 1 .5_ 2 2.5 0.875 1 .5_ 2 2.5 y 2.5 6 loose-edge d a_____ 1 2 2.5 0.75 1.25 1.75 y 1.75 6 loose-center a_____ 1 2 2.5 0.75 _1 2 2.5 y 2.5 6 lole d a_____ 1.25 1.5 y 0.75 _1 1.25 1.75 1.5 Vane hx___ _1 0.5 2(3 1 0.5 _1 0.5 2/3 1 y 2/3 2 wid 1 4/3 2 1 1 4/3 2 4/3 3 plft ength 4 6 6 4 4 6 20 8 30 heck through-on-end ength 4 6 6 y 4 4 6 20 8 30 )ecay hx none 2/3 area 8/3 none none 2/3 area 8/3 area 8/3wid none 4/3 y none none 4/3 y 3rain slope 1/6 1/41/4 1/12 1/10 1/8 1/4 an_n 1/4 ub- ategory Grade GONST ight STAND X raming UTIL X ECONOMY X SEL-STR X itructural NO.1 X ight NO.2 X raming NO.3 X ECONOMY X itud STUD X ECONOMY x X - Data in the column is the specification for the grade in the row. none - Not allowed any - No limitations area - Width x length 95 Table B2. Test results of grading the lumber in the Dimension category )efect Type/Location Measure fi =inchL (not firm-edge die 0.75 ____ 1 T5 firm-center die 2 0.875 1.5 2.5 loose-edge dia 1 2.5 0.75 1.75 1.75 loose-center dia 2 0.75 2 2.5 lole dia 1 1.25 1.5 0.75 1 1.25 1.75 1.5 Vane thx 0.5 0.6 1 0.5 0.5 0.6 1 0.6 wid 1 1.3 2 1 1 1.3 2 2 pIft length 4 6 6 4 4 6 20 8 )heck through-on-end length 4 6 6 4 4 6 20 8 )ecay thx 0.5 1 0.66 1 1 wid 1 2.6 1.33 2.6 2.6 rain slope 0.16 0.25 0.25 0.08 0.1 0.13 0.25 0.25 pub- ategory Grade ight raming CONST x r STAND X X UTIL X ECONOMY X X X tructuraI ight raming SEL-STA X X NO.1 X NO2 X NO.3 X X X X ECONOMY X X tud STUD X X X X X XX ECONOMY X X X 1) X indicates that input data in the column satisfies the grade in the row. 2) Blank cell means that no such characteristic exists. 3) For the same set of facts, grade overlaps are allowed. For example, facts in the first column match grades CONST, NO.3, and STUD. 4) The results satisfy the specification, indicating the system works well. Table B3. Specifications of grades for a piece of lumber in the Select/Finish category (lumber size = l"x8"x12') pet ect Type/Location Measure Limiting Provisions (<= inches) not tirm number * 2 * 2 any * 2 any any dia 0.5 0.75 any 0.75 any any fixed number none none * 4 none 4 6 dia none none 0.75 none 1 1.25 loose number none none none none none none dia none none none none none none oIe dia none none none none 1/16 0.75 Vane on face thx none none none none 0.25 0.75 wid none none none none 1 2 ength none none none none 24 48 on back thx 0.25 0.5 0.75 0.5 any any wid 0.5 1 2 1 any any ength 24 36 36 36 any any plft number none none any any any any length none none 24 24 24 48 Theck surface number 3 any any any any any wid 1/32 1/32 1/32 1/32 1/32 any length 4 4 4 4 10 any 'itch light wid none *area 576 any *area 576 any any length none any any any medium wid none none *area 768 none *area 768 any length none none none any heavy wid none none none none none any length none none none none none any 'ocket number *1 *2 *4 *4 4 6 width 1/16 1/16 1/16 1/16 1/16 area 4length 3 3 6 6 8 taln type light medium medium medium any any wid area 115.2 area 384 any any any any length any any any any pub- ategory Grade elect &BTR-1&2CLEAR C SELECT X DSELECT X inish SUPERIOR X PRIME X EFINISH X X - Data in the column is the specification for the grade in the row. none - Not allowed any No limitations area - Width x length * - Only one of the characteristics with symbol * in the same column is allowed. For example, grade B&BTR-1&2CLEAR allows 2 firm knots or 1 pocket. 97 Table B4. Test results of grading the lumber in the Select/Finish category )efect Type/Location Measure fact (<= inches)_____ (not firm number 2 2 dia 0.5 0.75 fixed number 4 6 dia 1 1.25 loose number dia dole die 1/16 0.75 vane on face thx 0.25 0.75 wid 1 2 length 24 48 on back thx 0.25 0.5 0.75 0.5 wid 0.5 1 2 1 length 24 36 36 36 put number 2 1 1 1 length 24 24 24 48 heck surface number 3 1 2 1 2 wid 1/32 1/32 1/32 1/32 1/32 length 4 4 4 4 10 itch light wid length medium wid 20 20 length 30 30 heavy wid length 'ocket number 2 4 6 width 1/16 1/16 1 length 3 8 3 itain type light medium medium medium wid 8 15 20 20 length 10 20 30 30 pub- ategoIy Grade elect I&BTR-1&2CLEAR C SELECT X DSELECT X X inish SUPERIOR X X X PRIME X X EFINISH X 1) X indicates that input data in the column satisfies the grade in the row. 2) Blank cell means that no such characteristic exists. 3) For the same set of facts, grade overlaps are allowed. For example, facts in the first column match grades B&BTR-l&2CLEAR and SUPERIOR. 4) The results satisfy the specification, indicating the system works well. Table B5. Specifications of grades for a piece of lumber in the Boards category (lumber size = l"x8"x12') )efect rype/Location Measure Limiting Provisions (<.. in) rirm-red dla 2.25 3 3.5 1W5 2.5 T not firm-black dia 0.75 1.25 7/3 16/3 1.25 2.5 3.5 16/3 loose dia none none 1.5 2.5 none 1.25 1.75 2.5 -lole dia none none 1.5 25 1/16 1.25 1.75 2.5 dane on face thx none none none 0.5 none 0.5 0.5 <1 wid none none none 1 none 1 4/3 2 length none none none 24 none 24 on back thx 0.5 0.5 2/3 any wid 1 4/3 2 _y any any length 36 48 72 any y_ pUt number 1 1 any any length 4 8 24 48 8 ..!_ 8 24 36 heck surface number 4 2 any any wld 1/32 1132 1/32 1/32 length 4 10 10 10 ith firm-heart wid 0.25 0.5 any ..! ? ? .!L ? ? ? length 24 72 any ? ? 7 ? ? light wid area 144 any any y_ any !k streak streak length any any any streak streak streak pitch medium wid none area 576 any any streak streak streak length none any y any streak streak streak heavy wid none none area 576 area none streak streak streak ylength none none 576 none streak streak streak y 'ocket number 2 3 any ! y any Lwidth 1/16 1/16 1/16 area 4 1/16 1/16 .i. ..!!i. length 3 6 12 12 12 _L itain type light medium medium !' medium ywid area 384 any any any length any any _y any _!L )ecay wid length none none area none none none samli y none none 144 _!i an an none none none small anub- :ategory Grade 1 COMMON 2 COMMON x ommons 3 COMMON X 4COMMON X 5 COMMON X SEL-MERCH x CONST X lternate STAND X UTIL X -- ECONOMY x X - Data in the column is the specification for the grade in the row. none - Not allowed any - No limitations area - Width x length Table B6. Test results of grading the lumber in the Boards category ii-. FJF YA'fli fli: I__ ------ - -- ;I II k'I -- -- li rr --_---- --_ [.i.J -- _________Lull I - 1) X indicates that input data in the column satisfies the grade in the row. 2) Blank cell means that no such characteristic exists. 3) For the same set of facts, grade overlaps are allowed. For example, facts in the first column match grades 1 COMMON and SEL-MERCH. 4) The results satisfy the specification, indicating the system works well. 
work_msj4tbct5vgrlkb42f3o5lgeva	----	doi:10.1016/j.eswa.2007.08.056 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 35 (2008) 1444–1450 Expert Systems with Applications Constructing and application of multimedia TV-news archives q H.T. Pao a,*, Y.H. Chen b, P.S. Lai b, Y.Y. Xu b, Hsin-Chia Fu b a Department of Management Science, National Chiao Tung University, Hsinchu, Taiwan, ROC b Department of Computer Science, National Chiao Tung University, Hsinchu, Taiwan, ROC Abstract This paper addresses an integrated information mining techniques for broadcasting TV-news. This utilizes technique from the fields of acoustic, image, and video analysis, for information on news story title, newsman and scene identification. The goal is to construct a compact yet meaningful abstraction of broadcast TV-news, allowing users to browse through large amounts of data in a non-linear fash- ion with flexibility and efficiency. By adding acoustic analysis, a news program can be partitioned into news and commercial clips, with 90% accuracy on a data set of 400 h TV-news recorded off the air from July 2005 to August 2006. By applying speaker identification and/ or image detection techniques, each news stories can be segmented with a better accuracy of 95.92%. On-screen captions or subtitles are recognized by OCR techniques to produce the text title of each news stories. The extracted title words can be used to link or to navigate more related news contents on the WWW. In cooperation with facial and scene analysis and recognition techniques, OCR results can provide users with multimodal query on specific news stories. Some experimental results are presented and discussed for the system reli- ability, performance evaluation and comparison. � 2007 Published by Elsevier Ltd. Keywords: TV-news archives; Multimedia; Information mining; Multimodal query; Video OCR 1. Introduction Among the major sources of news program, TV has clearly had the dominant influence atleast since the 1960s. Yet it is easy to find the old newspaper in microfilm in any public library, but it is impossible to find the old foot- age of television news in the same library. TV-news archive has existed in the United States for 35 years. Paul C. Simp- son founded the Vanderbilt University Television News archive in 1968. In Huffman, Yang, Yan, and Sanders (1990), a team in the University of Missouri-Columbia decided to do a content analysis of the three US network coverage of the 1989 Tiananmen Massacre, they located these news items in the Vanderbilt Archive Index Vander- bilt television. The Vanderbilt archive promptly provided the 11 h-video clips all related to the Tiananmen Massacre. 0957-4174/$ - see front matter � 2007 Published by Elsevier Ltd. doi:10.1016/j.eswa.2007.08.056 q This research was supported in part by the National Science Council under Grant NSC 94-2213-E009-139. * Corresponding author. E-mail address: htpao@cc.nctu.edu.tw (H.T. Pao). At the same time, the Missourian team also planned to do a comparable study of Taiwanese reportage on Tiananmen Massacre. But the equivalent material of the Vanderbilt archive did not exist in Taiwan then. Therefore, that study only contained the US perspective of the Tiananmen Mas- sacre. This paper proposes an integrated methodology for the information mining on a multimedia TV-news archive in Taiwan. As described in Xu, Chen, Tseng, Lai, and Fu (2004) Lai, Lai, Tseng, Chen, and Fu (2004), a fully auto- mated Web-Based TV-news system were implemented to achieve the following goals: 1. Academic and applied aspects: This archive will greatly improve the quality of TV-news. As Dan Rather, the CBS anchorman, once mentioned that he lives with two burdens – the ratings and the Vanderbilt Television News Archive Therefore, once the archive is there, the researchers and the public will do some content analysis on the TV-news. And the journalists will be more careful in what they report. mailto:htpao@cc.nctu.edu.tw Fig. 1. The overall architecture and information processing flow of the proposed fully automated web-based TV-news system. Fig. 2. The flow diagram of TV news acquisition and content segmentation. H.T. Pao et al. / Expert Systems with Applications 35 (2008) 1444–1450 1445 2. Timing factor: Vanderbilt archive started its project with Betacam videotapes in 1968. There will be a problem of preservation because these tapes deteriorate along the years. Today, we can save all the TV-news in hard disc, VCD or DVD. Informedia (http://www.informedia.cs.cmu.edu/) is an integrated project launched in Carnegie Mellon university. Its overall goal is to use modern AI techniques to archive video and film media. VACE-II Informedia (http:// www.informedia.cs.cmu.edu/arda/vaceii.htm), a sub-pro- ject of Informedia, automatically detects, extracts, and edits highly interested people, patterns, and story evolves and trends in visual content from news video. This paper proposes a integrated information mining technique that can automatically generate semantic labels from news video Daniel and Daniel (2002), and statistical methods to discover hidden information. We intend to expect that the following significance will come to exist. • Although web-news provides another efficient way to access news, watching TV-news already became habit of many people. Beside this, most of web-news system can only provide text-based news. • There are so many channels providing TV-news. People need more information for searching like-minded channel. • Although almost every channel announced that they are dispassion, real dispassion is hard to archive with human editing. We need some evaluation to check if the channel is really dispassion. The rest of this paper is organized as follows. Section 2 introduces the overall concepts of the multimedia TV-news archive. In Section 3, methods of generating necessary semantic labels from the recording TV-news video are pre- sented. Section 4 focus on describing the information min- ing from these semantic labels. Finally, summary and concluding remarks are given in Section 5. 2. TV-news archive A fully automated Web-based TV-news system Lai, Lai, Tseng, Chen, and Fu (2004); Xu et al. (2004) consists of three modules: (1) TV-news video acquisition and content segmentation, (2) news content analysis, and (3) user inter- face for news query, search and retrieval. Fig. 1 depicts the overall architecture and interaction of these three modules. The major tasks of the acquisition module are to record TV-news programs in a proper video format, and to fetch related news text contents from Internet webs. Content analysis module segments the recorded news video into story based units, and extracts news title and keywords from each story unit. Providing a friendly querying and browsing environment for retrieving interested news stories is the most important task of the user interface module. An overall news video processing and content analyzing are depicted in Fig. 2. At the beginning, a TV-news pro- gram is captured and encoded into stream video format (Iain & Richardson, 2003; Wang, Ostermann, & Zhang, 2002). The recorded streaming video is named with dates first, and then stored in database. In the meantime, a shot detector is used to segment the streaming video into scene based shots for news unit generation and key-frame extrac- tion (Lee, Yoo, & Jang, 2006; Patel & Sethi, 1997). Within a scene shot, speaker identification techniques are then applied to detect anchor frames (Cheng, Wang, & Fu, 2004). The close-captions in the anchor frames are then extracted and recognized by using video OCR techniques (Lin, Liu, & Chen, 2001) as candidates for the news title and keywords of each news units. The extracted keywords can then be used to match with (1) Internet news stories to construct links between TV-news stories and Internet news, http://www.informedia.cs.cmu.edu/ http://www.informedia.cs.cmu.edu/arda/vaceii.htm http://www.informedia.cs.cmu.edu/arda/vaceii.htm Detect Commercial A A A A WA A A A WA A A ACC WA A A ACC Story1 Story2 Story3 A:anchor C:commercial W:weathercast Detect Anchor Video Clip Detect Weather Report Shot News Program Extract News Stories 1446 H.T. Pao et al. / Expert Systems with Applications 35 (2008) 1444–1450 and (2) the users’ query words for retrieving interested and/ or related news stories. In addition, the extracted charac- ters by video OCR from each news units can also be used as semantic labels of the news units. More description on the semantic labels will be presented in Section 3.3. 3. News information tree generation The most important things in news story writing are that journalists commonly refer to as the 5 W’s: who, what, when, where and why. These questions are crucial for catching a reader’s attention and introducing the essential facts of the story. Standing on this basic rules, the news archive system introduced in Section 2, are further improved to extract more information from a recorded news video. A news information tree (Feinstein & Morris, 1988) is suggested to structure the contents of recorded video clips for helping the user focus on specific news infor- mation, and information that is a little more general. 3.1. News information tree Fig. 3 illustrates a hierarchical structure of a news infor- mation tree. The hierarchical tree contains five types of video information records: (1) date (when), (2) channel (where), (3) title (what), (4) content (how), and (5) commer- cial. The title record contains the starting time, length, and brief description of the corresponding video clips. The con- tent record can also be further divided into the following sub-records: (a) on-site locations, (b) interview, and (c) tables or quoted word. 3.2. Analysis units Usually, a TV-news program contains the following items: news stories, commercials and weather reports. Fig. 3. The data structure of a news information tree. Complete description of shot detection and scene segmen- tation can be found in Huang, Lai, and Fu (2004). A flow chart of TV-news program segmentation is shown in Fig. 4, and brief introduction are described as follows. Among various scene shots, anchor video clips are detected first (Kim, Kim, Ra, & Choi, 1999). In general, anchor segments are the most appeared video clips (Saracoglu, Tutuncu, & Allahverdi, 2007), thus we propose to use BIC (Fraley & Raftery, 1998), an unsu- pervised method, to cluster anchor segments from the other clips. As shown in Fig. 5, this method contains the follow- ing procedures: 1. The MFCC audio feature sequence X is generated from input audio at first. 2. BIC segments X into segments X1,X2,. . .Xn. 3. These segments then are clustered as several clusters C1,C2,. . .,C3. 4. The cluster containing most clip segments is the set of anchor clip. Fig. 4. The flow diagram of the proposed news story analysis and information extraction processes. Fig. 5. The flow diagram of a BIC-based audio segmentation method. H.T. Pao et al. / Expert Systems with Applications 35 (2008) 1444–1450 1447 After locating each anchor shot, a SVM model based video classifier (Sun, Tseng, Chen, Chuang, & Fu, 2004) is used to detect weather report shots. Finally, commercials are detected and separated from on-site news stories. The following feature detecting techniques are integrated to achieve a high performance commercial detector: • The variation rate of zero crossing rate. • Short time energy. • Shot change rate. • Clip length. At this stage, anchor’s briefing, weather reports, commer- cial, and the background stories are all separated and iden- tified. For each on-site stories, as shown in Fig. 6, we further segment and classify each on-site scene into three categories: locations, interview and tables or quoted words (what). Fig. 7 shows how to partition the on-site news story into (1) location, interview, and tables or quoted words scenes. In general, on-site narration is not active during the inter- view scene, as shown in Fig. 6, the interview scene can be distinguished from location scene. By using its special char- Newshawk Scene Clip The Speaker Newshawk Newshawk B ac kg ro un d S to ry Locality Interview Locality Data Chart Locality Newshawk Interviewee Fig. 6. The general structure of a news story. On-site scene story contains three major news contents: locations, interview and tables or quoted words. :locality News Story Newshawk Voice Model Mark the Rest of Newshawk Periods as Locality Mark Data Chart from Newshawk’s Period Mark the Rest as Interview Mark Newshawk Speaking Periods N N N NI N NI C L I L C L L IN N C N I :newshawk :interview C L :data chart Fig. 7. On-site scene segmentation flow. The narration periods can be detected by using speaker identification techniques to distinguish narra- tor’s speech voice from the rest of scenes. Detecting the screen characters regions can find the tables or quoted words scenes. Then, the rest of scenes that are not belonging to interview or tables or quoted words scenes must be the location scenes. acteristics of the character regions, tables or quoted words scene can be distinguished from location scenes. 3.3. Semantic labels of units This section describes how to assign each segmented unit with semantic labels. Basically, the text words for each label are extracting from text streams in close-caption. Usually, a TV-news program often provides audience a quick overview of each news story in on-screen captions, such as names of location, people, and keywords of events, . . . etc. In general, these texts are quiet enough to give enough information for labeling each segmented units. Fig. 8 shows how to establish a news information tree. The information tree establishing process contains two phases: story and scene phases. In story phase, all on-screen characters are recognized by video OCR first. Then, the recognized characters or words are used to match with text-based news documents, which are usually retrieved from Internet. The title and contents of best matched text-news document not only fill out the story information record of news information tree, but also used to picking label candidates, including loca- tions, people names, event words, quoted phrases, and tab- ular data, up for scene phase processing. Fig. 8. Information flow of the generation of a news information tree. For each story, video OCR (optical character recognition) technique is applied to extract characters in the close-caption. The extracted characters will then be used to match with the retrieved news document over the Internet web sites. From the matched documents, key information, such as associated people, event location, reporters’ names etc., can be retrieved accordingly. Finally, video clips associated with location, interview, tables and quoted words scene can then be labeled according to the extracted keywords. 1448 H.T. Pao et al. / Expert Systems with Applications 35 (2008) 1444–1450 In scene phase, picking semantic labels up from label candidates for each scene is done in this phase. As shown as Fig. 9, the on-screen captions of locality scene provide location name and event descriptions. Therefore, in locality scene, the location and event words of label candidates are searched from on-screen captions to find which ones exactly appeared. Fig. 10 is an example of interview scene. In interview scene, the interviewee’s name and their points are always given by on-screen captions. Therefore, in interview scene, we search people name, and events word of label candi- dates instead. The example of data chart scene is shown as Fig. 11. In general, on-screen captions fill data chart scene. These cap- tions may present quoted sentence or tabular data. There- Fig. 10. An example of interview scene frame. An interview scene is used to present a reporter’s point of view. Normally, the reporter’s name and opinions can be extracted from subtitles or closed captioning. Fig. 11. An example of data chart scene frame. The data chart scene is used to present information in a organized manner. Additional informa- tion is also available from the on-screen characters. Fig. 9. An example of locality scene frame. The locality scene is used to show where and what news occurred. Thus, the location information and event description can be retrieved from the close-captions of a locality scene. fore, searching for quoted sentence or tabular data in data chart scene is the major task. 4. Data mining on the news information tree This section presents how and what to mine from a news information tree (NIT). In Section 4.1, we propose to mine the favored or preferred news contents of a TV-station. The news information tree can also be used to track the evolution of a series of news stories (see Section 4.2). In addition, the mining results from the NIT and the realtime ratings can be combined to provide TV-news commercial buyers a very useful guidance. 4.1. Mine the news preference of a TV-Station Generally speaking, a TV-station arranges the broad- casting sequence of each story in a news program according to their impact and attractiveness to audience. In fact, a preferred news story often gets more time on the air. By analyzing the sequence order and the length of stories, the preferred or the favored news stories of a TV-station can be roughly estimated or judged. Mining the NIT to extract favored or preferred types of news story from a TV-station will help audience to find the favored news channel. The proposed news mining method is described as fol- lows: Given N sets of keywords, K1,K2, . . ., Ki, . . ., KN, which correspond to N news topics (or subjects), let the fol- lowing delta function d(k,Ki) define the relations between a keyword k and a keyword set Ki: dðk; K iÞ¼ 1; if keyword k 2 K i 0; otherwise: � 1. Extract keywords {klsj ; l ¼ 1; � � � ; Lj} from a scene unit sj in a news program. 16 "elect" H.T. Pao et al. / Expert Systems with Applications 35 (2008) 1444–1450 1449 2. For each scene units sj, compute its association fre- quency F(Kijsj) with respect to a subject Ki, 0 500 1000 1500 2000 2500 3000 3500 4000 0 Fig. 12 appear 12 14 FðK ijsjÞ¼ PLj l¼1dðk l sj ; K iÞPN i¼1 PLj l¼1dðk l sj ; K iÞ 0 2 4 6 8 10 0 20 40 60 80 100 120 140 160 Fig. 13. The life-cycle of a specific news events along with a period of time. 3. Compute the the association frequency of news program Fd(tjKi) at time t and day d: F dðt; K iÞ¼ FðK i; s1Þ; for s1:start 6 t 6 s1:end FðK i; s2Þ; for s2:start 6 t 6 s2:end .. . .. . FðK i; sMÞ; for sM:start 6 t 6 sM:end; 8>>>>< >>>>: where sj.start and sj.end are the start and the end time of a scene unit sj in a news program at time t of the day d. The associated frequency distribution from one segment of news program is not enough to represent the overall pref- erence or trend of a news channel, thus long term statistics is needed. By accumulating a longer period (say one month) of associated frequency of news subjects, the preference of a channel can be discovered. As shown in Fig. 12, keywords that are related to social news, political news, and enter- tainment news are applied to associate with and to accumu- late frequency of news topics. As we can see in this example, the monitored news channel favors social and political news more than entertainment news. 4.2. The evolution of a series of news stories The evolution of a news story can also be mined from the news information tree. By associating the keywords of a specific event with recorded news scenes over a period of days, then the accumulated association frequency of matched scene units presents an overall developing and progressing of the specific news stories. Fig. 13 shows a sort of life-cycle of a particular news events. In addition, 500 1000 1500 2000 2500 3000 3500 4000 "politics" "society" "sport" . Three sets of (representative) keywords are used to associate the ing frequency of social, political and sport news in a news program. the spreading of the specific events to other areas, e.g. cit- ies, counties, countries, etc., can also be retrieved from the associated names of locations in the matched scene units. For example, one can query a news story by using a par- ticular people’s name, then the person’s daily schedule and/or whereabouts can be retrieved from the recorded NIT. 4.3. The mining on TV commercial Beside background stories, commercial records are also valuable information. Huang et al. (2004) proposed com- mercial detecting and identifying methods in TV video clips. When a commercial frame contains image keywords in a video frame, video OCR techniques can be used to extract keywords to label the corresponding video clips. Otherwise, keyblock-based image retrieval methods (Zhu, Rao, & Zhang, 2002) may be utilized to represent and to identify each commercial clips. However, manual annota- tion is needed to label the keyblock. By gathering statistical information of these labels and keywords in news pro- grams, cross relationship between TV commercials, real- time ratings, and news stories can be observed and analyzed to achieve a useful marketing database. Two example areas, customer modeling and cross-selling, in data- base marketing are discussed in the following. 4.3.1. Customer modeling The basic idea behind customer (i.e. the commercial buy- ers and news audience) modeling is to improve audience response rates by targeting prospects that are predicted as most likely to respond to a particular advertisement or pro- motion. This is achieved by building a model to predict the likelihood that groups of news audience will respond based on news type, viewing time and news channels as well as previous viewing behavior. In addition, by targeting more effectively to prospects and existing commercial buyers, TV-station operators can improve and strengthen customer relationships. The customer can perceive more value in TV- 1450 H.T. Pao et al. / Expert Systems with Applications 35 (2008) 1444–1450 news and commercials (i.e. both commercial buyers and news audience receive only products and/or services of interest to them). 4.3.2. Cross-selling The basic idea behind cross-selling is to leverage the existing customer base by selling them additional products (commercial time slots) and/or news services. By analyzing the groups of products or services that are commonly pur- chased together and predicting each customer’s affinity towards different products using historical data, a TV-sta- tion can maximize its selling potential to the existing cus- tomers. Cross-selling is one of the important areas in database marketing where predictive data mining tech- niques can be successfully applied. Using historical pur- chase data of different products from the customer database along with news type, viewing time and news channels, commercial buyers can identify their products that are most likely to be of interest to targeted news audi- ence. Similarly, for each type of product (i.e. commercial or groups of commercials), a ranked list of different types of news or groups of audience, that are most likely to be attracted to that product. Then, arrangement of commer- cials with matched types of news to achieve a high likeli- hood of audience response rate. 5. Conclusion This paper addresses techniques and possible applica- tions of fully automated information mining on a multime- dia TV-news archive. The proposed automated information mining contain the following processes: (1) segmenting a TV-news program video recording into scene clips, (2) using video OCR to extract and recognize close-caption and/or image characters into keywords for each scenes, (3) using keywords to generate semantic labels for each scenes, and (4) segmenting commercial video clips from news clips. Information associated with various labels and scenes (e.g. the starting and ending time of a scene) are stored in the proposed news information tree. Performing statistical analysis on the data items in the news information tree can reveal hidden information, like popular channels and evolution of some hot news stories. These information can help general multitude in finding their favored or desired news channel, searching focal point person, tracking hot news stories, . . ., and so on. References Cheng, S.-S., Wang, H. m., & Fu, H.-C. (2004). A model-selection-based self-splitting gaussian mixture learning with application to speaker identification. EURASIP Journal on Applied Signal Processing, 17, 2626–2639. Daniel, Gildea, & Daniel, Jurafsky (2002). Automatic labeling of semantic roles. Computational Linguististics, 28(3), 245–288. Feinstein, C. D., & Morris, P. A. (1988). Information tree: A model of information flow in complex organizations. Systems, Man and Cyber- netics, IEEE Transactions, 18(3), 390–401. Fraley, C., & Raftery, A. E. (1998). How many clusters? which clustering method? answers via model-based cluster analysis. Computer Journal, 41, 578–588. Huang,T. -Y., Lai, P.- S., & Fu, H.-C. (2004). A shot-based video clip search method. In Proceedings of CVGIP2004, Taipei, Hualien, ROC, August 2004. Huffman, S., Yang, T., Yan, L., & Sanders, K. (1990). Genie out of the bottle: Three US Networks report tiananmen square. In Proceedings of the annual meeting of association for education in journalism and mass communication, Minneapolis, Minnesota, USA. Iain, E. G., & Richardson, H. (2003). 264 and mpeg-4 video compression. Wiley Press. Kim, D.-W., Kim, J.-T., Ra, I.-H., & Choi, Y.-S. (1999). A new video interpolation technique based on motion-adaptive subsampling. IEEE Transactions on Consumer Electronics, 45(3), 782–787. Lai, P. S., Lai, L. Y., Tseng, T. C., Chen, Y. H., & Fu, H. C. (2004). A fully automated web-based TV-news system. In Proceedings of PCM 2004, Tokyo, Japan, Dec. 2004. Lee, M. H., Yoo, H. W., & Jang, D. S. (2006). Video scene change detection using neural network: Improved art2. Expert System with Applications, 31(1), 13–25. Lin, C.-J., Liu, C.-C., & Chen, H.-H. (2001). A simple method for chinese video ocr and its application to question answering. International Journal of Computational Linguistics and Chinese Language Processing, 6(2), 11–30. Patel, N. V., & Sethi, I. K. (1997). Video shot detection and character- ization for video databases. Pattern Recognition, 30(4), 583–592. Saracoglu, R., Tutuncu, K., & Allahverdi, N. (2007). A fuzzy clustering approach for finding similar documents using a novel similarity measure. Expert System with Applications, 33(3), 600–605. Sun, S.- Y., Tseng, C. L., Chen, Y. H., Chuang, S. C., & Fu, H. C. (2004). Cluster-based support vector machine in text-independent speaker identification. In Proceedings of international joint conference on neural networks IJCNN 2004, Budapest, Hungary; 2004. Vanderbilt television news archive, http://www.vanderbilt.edu/vtna. Wang, Y., Ostermann, J., & Zhang, Y.-Q. (2002). Video processing and communications. Prentice Hall Press. Xu, Y. Y., Chen, Y. H., Tseng, C. L., Lai, P. S., & Fu, H. C. (2004). Multimedia TV-news browsing system. In Proceedings of IEEE international conference on multimedia and expo (ICME), Taipei, Taiwan, ROC; June 2004. Zhu, L., Rao, A., & Zhang, A. (2002). Theory of keyblock-based image rerieval. ACM Transactions on Information Systems, 20(2), 224–257. http://www.vanderbilt.edu/vtna Constructing and application of multimedia TV-news archives Introduction TV-news archive News information tree generation News information tree Analysis units Semantic labels of units Data mining on the news information tree Mine the news preference of a TV-Station The evolution of a series of news stories The mining on TV commercial Customer modeling Cross-selling Conclusion References 
work_mvrzshdomjex5cflk6oubgvvvy	----	AD-A1IX 672 STANFORD UNIV CA.DEPT OF COMPUTER SCIENCE F/G 9/2 THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM: A FRAMEWORK FOR--ETC(Ul JAN 82 W J CLANCEY N00014-79..C-0302 UNCLASSIFIED STAN-CS-81-896 NL H . 1II.I5 Noiernbr 1981 Report. No. STAN-CS--R -- .lso numinered: The Epistemology of A Rule-Based Expert System: A Framework for Explanation .4 by William J. Clancey Department of Computer Science Stanrord University Stanrord, CA 94305 C MAR 3 1,4 C.2. L= $. is re._., =ma ';,ag b.e rrI]t . w, JA ERRATUM PAGE 20: THE NODE THAT IS LABELED "P M N S>100" SOULD READ: "P M N S + BNDS < 1000" '1 UN CT.AgRTFTpT SECURITY CLASSIFICATION OF THIS PAGE lWhen Oat& Entered) REPORT DOCUMENTATION PAGE READ INSTRUCTIONS BEFORE COMPLETING FORM 1 REPORT NUMBER 2. GOVT ACCESSION NO. 3 RECIPIENT'S CATALOG NUMBER STAN-CS-81-896, Tech. Report #4 im .i _ _______'___ 4. TITLE (and Subtitle) 5. TYPE OF REPORT & PERIOD COVERED The Epistemology of a Rule-Based Expert System: technical, January 1982 A Framework for Explanation 6. PERFORMING ORG. REPORT NUMBER 7. AUTHOR(s) STAN-CS-81-896 8. CONTRACT OR GRANT NUMBER(s) William John Clancey, Ph.D. N0014-79-C-0302 9 PERFORMING ORGANIZATION NAME AND ADDRESS 10. PROGRAM ELEMENT. PROJECT. TASK Department of Computer Science AREA & WORK UNIT NUMBERS Stanford University NR 154-436 Stanford, CA 94305 USA 12. REPORT DATE 13. 0O OF PAGES 11. CONTROLLING OFFICE NAME AND ADDRESS Personnel and Training Research Programs Jan SECURITY CLASS (of 9h82 repot' Office of Naval Research (Code 458) Arlington, VA 22217 Unclassified 14. MONITORING AGENCY NAME & ADDRESS (if diff. from Controlling Office) ONR Representative-Mr. Robin Simpson 15a. DECLASSIFICATION/DOWNGRADING Durand Aeronautics Building, Room 165 SCHEDULE §an ora U"veaor_, 16, DISTRIBUTION STATEMENT (of this report) Approved for public release; distribution unlimited. 17. DISTRIBUTION STATEMENT (of the abstract entered in Block 20, if different from report) 18. SUPPLEMENTARY NOTES To appear in the AI Journal 19. KEY WORDS IContinue on reverse side if necessary and identify by block number) 20.A RACT (Continue on rererie side if necessary and identify by block number) Production rules .are a popular representation for encoding heuristic know- ledge in programs for scientific and medical problem solving. However, experience with one of these programs, MYCIN. indicates that the representation has serious limitations: people other than the original rule authors find it difficult to modify the rule set, and the rules are unsuitable for use in other settings, such as for application to teaching. These problems are rooted in fundamental limitations in MYCIN's original rule representation: the view that expert knowledge can be encoded as a uniform, weakly-structured set of if/then associations is found to be wanting. (continued) D D O J 1473 UNCLASSIFIED EDITION OF I NOV 65 IS OBSOLETE SECURITY CLASSIFICATION OF THIS PAGE (When Data Enleredl UNCLASSIFIED SEC rITY CLASSIFICATION OF THIS PAGE (When Data Entered) 19. KEY WORDS (Continued) 20 ABSTRACT (Continued) To illustrate these problems, this paper examines MYCIN's rules from the perspective of a teacher trying to justify them and to convey a problem-solving approach. We discover that individual rules play different roles, have different kinds of justifications, and are constructed using different rationales for the ordering and choice of premise clauses. This design knowledge, consisting of structural and strategic concepts which lie outside the representation, is shown to be procedurally embedded in the rules. Moreover, because the data/hypothesis associations are themselves a proceduralized form of underlying disease models, they can only be supported by appealing to this deeper level of knowledge. Making explicit this structural, strategic and support knowledge enhances the ability to understand and modify the system. D0 FoAN7 3147 ( BACK) EOITION OF 1 NOV 66 IS O1SOLETE SECURITY CLASSIFICATION OF THIS PAGE (When Date Entered) THE EPISTEMOLOGY OF A RULE- BASED EXPERT SYSTEM: A FRAMEWORK FOR EXPLANATION William J. Clancey Department of Computer Science Stanford University, Stanford CA 94305 Contract No. N000C4-79-0302. effective March 15, 1979. Expiration Date: March 14, 1979 Total Amount of Contract -- $396,325 Principal Investigator, Bruce G. Buchanan (415) 497-0935 Associate Investigator, William J. Clancey (415) 497-1997 Sponsored by: Office of Naval Research, Personnel and Training Research Programs, Psychological Sciences Division. Contract Authority No. NR 154-436 Scientific Officers: Dr. Marshall Farr and Dr. Henry Halff The views and conclusions contained in this document are those of the authors and should not be interpret as necessrily representing the official policies, either expressed or implied, of the Office of Naval Research or the U.S. Government. Approved for public release: distribution unlimited. Reproduction in whole or in part is permitted for any purpose of the United States Government Aci, ', or ",'. ,- .A~ ~. E.t, : Abstract Production rules are a popular representation for encoding heuristic knowledge in programs for scientific and medical problem solving. However, experience with one of these programs, MYCIN, indicates that the representation has serious limitations: people other than the original rule authors find it difficult to modify the rule set, and the rules are unsuitable for use in other settings, such as for application to teaching. These problems are rooted in fundamental limitations in MYCIN's original rule representation: the view that expert knowledge can be encoded as a uniform, weakly-structured set of If/then associations is found to be wanting. To illustrate these problems, this paper examines MYCIN's rules from the perspective of a teacher trying to Justify them and to convey a problem-solving approach. We discover that individual rules play different roles, have different kinds of Just'fications, and are constructed using different rationales for the ordering and choice of premise clauses. This design knowledge, consisting of structural and strategic concepts which lie outside the representation, is shown to be procedurally embedded in the rules. Moreover, because the data/hypothesis associations are themselves a proceduralized form of underlying disease models, they can only be supported by appealing to this deeper level of knowledge. Making explicit this structural, strategic and support knowledge enhances the ability to understand and modify the system. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 2 1 Introduction Production rules are a popular representation for capturing heuristics, "rules of thumb," In expert systems. These artificial intelligence programs are designed to provide expert-level consultative advice in scientific (CRYSAIS--[1])(R1--[22]) and medical problem solving (EXPERT--[23]) (VM--(1g]). MYCIN (24] Is generally acknowledged to be one of the forerunners of this research. Shortliffe's and Buchanan's original design goal was to use a simple, uniform formalism, which was easy to modify, to encode heuristics. They cited the Importance of providing reasonable explanations, as well as good advice, for the program to be acceptable to users. This approach was in stark contrast to the common Bayesian programs which did not seek to capture an expert's reasoning steps, and so were not easily understood by clients [26]. MYCIN set the standards for much subsequent research. The success of MYCIN as a problem solver In infectious disease diagnosis, as shown in several formal evaluations [33] [34], suggested that the program's knowledge base might be a suitable source of subject material for teaching students. This use of MYCIN was consistent with the design goals that the program's explanations be educational to naive users, and that the representation be flexible enough to allow for use of the rules outside of the consultative setting. In theory, the rules acquired from human experts would be understandable and useful to students. The GUIDON program (6] [7] (8] was developed to push these assumptions by using the rules In a tutorial interaction with medical students. In attempting to "transfer back" the experts' knowledge through GUIDON, we find that the expert's approach and understanding of rules have not been represented. GUIDON cannot justify the rules because MYCIN does not have an encoding of how the concepts In 3 a rule fit together. GUIDON cannot fully articulate MYCIN's problem solving approach because the structure of the search space and the strategy for traversing it are Implicit In the ordering of rule concepts. Thus, the seemingly straightforward task of converting a knowledge based-system into a computer-aided instruction program has led to a detailed re-examination of the rule base and the foundations upon which rules are constructed, an epistemological study. In building MYCIN, rule authors did not recognize a need to record the structured way In which they were fitting rule parts together. The rules are more than simple associations between data and hypotheses. Sometimes clause order counts for everything (and the order can mean different things), and some rules are present for effect, to control the invocation of others. The uniformity of the representation obscures these various functions of clauses and rules. In looking beyond the surface of the rule representation to make explicit the intent of the rule authors, this paper has a purpose similar to Wood's "What's In a Link?" (32] and Brachman's "What's in a Concept?" [2]. We ask, "What's in a Rule?" The demands of tutoring provide a "forcing function" for articulating the structure of the rule base and the limitations of the program's explanation behavior. These Insights have implications for generating and explaining consultative advice, and modifying the system. For the moment, consider that many, if not most or all, of the issues of tutoring must be addressed In providing explanations to naive users. Consider how an expert can violate a rule In difficult, "non-standard" situations, because he can reason about the rule's Justification. Finally, consider the difficulties of changing a program whose design Is not fully documented. In building GUIDON, we thought that we were simply being "applications engineers" by THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 4 making use of MYCIN's explanation facility for a tutorial setting. it was completely surprising to find out how little the facility could accomplish for a student. Without a crisp characterization of what we expected an "explanation" to convey, the program was of questionable value. On the positive side, the study of knowledge that these shortcomings motivated led to a radical change in our conception of MYCIN's rules, and supplied a new epistemological framework for building expert systems, the subject of this paper. Subsequent sections provide a review of MYCIN's explanatory capability (Section 2), and an overview of an epistemological framework for enhancing that capability (Section 3). The following two sections examine In detail the problems of justifying a rule and explaining an approach, thereby elucidating the support and strategic aspects of the epistemological framework. In Section 6 Implications for performance of a consultation system and modifiability are considered briefly. Finally, In Section 7 the framework is used to analyze a variety of other expert systems. !6 2 MYCIN's Explanation Capability As already mentioned, MYCIN was designed so that It could explain Its reasoning to the user. This means that it can describe its reasoning steps: how a request for date Is related to a goal, how one goal leads to another, and how a goal is achieved. Figure 1 shows a consultation excerpt In which the program demonstrates a line of reasoning (chained goals) that lead backwards from the current question: Is the patient a compromised host (by rule 343 would imply) -> is the patient Immunosuppressed? (rule 395) -> is the patient receiving steroids? (rule 543) -> is e.coi causing the meningitis? Figure 2 illustrates how, In the questioning session after the consultation, one can inquire further about the program's itermediate reasoning steps, Including why It didn't ask about something. These are ibe explanation capabilities that we sought to exploit In a teaching program. MYCIN's explanations are entirely In terms of Its rules and goals. The question "WHY?" means, "How Is this Information useful?," translated internally as "in what rule does this goal appear, and what goal does the rule conclude about?" Davis, who developed the explanation facility (Figure 1), pointed out that MYCIN did not have the knowledge to respond to other Interpretations of a "WHY" question [11]. He mentioned specifically the lack of rule justifications and planning knowledge which we have studied and address In this paper. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 6 23) Is J.Smith a compromised host? *WHY [i. e. WHY is it important to determine whether J.Smith is a compromised host?] (3.8] This will aid in determining whether J.Smith is immunosuppressed. If ( 3. 1] J.Smith Is not a compromised host then It Is definite (7.0) that J.Smith Is not immunosuppressed (ROLE343] (back to question 23...] ~WHY (i. e. WHY is it important to determine whether J.Smith is immunosuppressed?] (4.8]... in order to determine whether J.Smith has received corticosteroids. If ( 4. 1 ] J.Smith Is not Immunosuppressed then (R135It Is definite (1.0) that J.Srnith has not received corticosteroids.I (back to question 23 ...] SWHY (i.e9. WHY is it important to determine whether J.Smith has received corticosteroids?] (5.8J]... in order to determine the organisms causing the infection. It has already been established that [(5. 1]) the Infection which requires therapy Is meningitis, and [(5.2] only circumstantial evidence Is available for this case, and ( 5. 3 ] the type of meningitis Is bacterial. Therefore, if (5.4 ] J.Smith Is receiving corticosteroids then there Is evidence that the organisms causing the Infection are: e.coll (.4), Alebsiela-pneumoniae (.2), or pseudomonas-aerugiroosa (.1) (RULE 543] (back to question 23...] ~NO Figure i. Consultation excerpt showing explanation capability (User input in BOLD ITALICS; all other text is printed by NYCIN.) 7 m DID YOU CONSIDER WHETHER THE PATIENT WAS RECEIVING STEROIDS? RULE543 could have used: whether J.Smith has received corticosterolds. Clause 4 of RULE543 (the patient has received corticosteroids) was already known to be false. SHOW DID YOU KNOW THAT THE PATIENT HAD NOT RECEIVED STEROIDS? RULE395 was used to conclude that J.Smith has not received corticosterolds. The last question asked before the conclusion was made was 23. SWHAT WAS QUESTION 23? Question 23 was asked in order to find out whether J.Smith is a compromised host in an effort to execute RULE343. Figure 2. Excerpt from question/answer session (User input appears in BOLD ITALICS.) In order to Illustrate other meanings for the question "Why?" In MYCIN, we Illustrate the rule set as a network of goals, rules and hypotheses (Figure 3). At the top level are all of the system's goals that It might want to pursue to solve a problem (diagnostic and therapeutic decisions). Examples of goals are: "What is the shape of the organism?" "What organism Is causing the meningitis?" At the second level are hypotheses or possible choices for each of the goals. Examples of hypotheses are: "The organism Is a rod." "E.coll Is causing the meningitis." At the third level are the rules that support each hypothesis (this Is a graph because a rule might support more than one hypothesis). At the fourth level appear the premises of these rules--specific hypotheses that must be believed for the rule to apply (multiple rules might predicate the same hypothesis). For example, for rule 643 to apply, It must be the case that the Infection Is meningitis, that the meningitis be caused by bacteria, that the patient be receiving steroids, and so on. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM S Level 1 Goal What cause of inf? What therapy? ~ more specific" 2 Hypothesis e.coli cryptococcus "concluded by" 3 Rule Rule543 Rule535 predicates" 4 Hypothesis meningitis bacterial steroids a3coholic "more general" 5 Goal What infection? What kind of meningitis? Steroids? Acohofic? Figure 3. Rule set shown as a network linking hypotheses and goals A key aspect of MYCIN's Interpreter Is that, when confronted with a hypothesis In a rule premise that It needs to confirm, it considers all related hypotheses by pursuing the more general goal. For example, attempting to apply rule 643, the program will consider all rules that conclude about the infection, rather than just those that conclude that the Infection Is meningitis. Similarly, It will consider all rules that conclude about the kind of meningitis (viral? fungal? tb? bacterial?), rather than just those that hypothesize that the meningitis Is bacterial1 . These new goals deriving from rules can now be seen conceptually This is not Inefficient, given the program's exhaustive search strategy and the fact that the other hypotheses will be referenced by other rules. Note also that some hypotheses, such as "the patient Is receiving steroids," are not generalized, but represented as goals directly. Whether or not a hypothesis is represented as a "yes/no parameter" or as a "value" of "multi-valued parameter" (such as "kind of meningitis") Is a rule-author decison? deriving from a pattern of hypotheses that he wishes to collapse for clarity into a more general goal. By this process of abstraction, a single "multi-valued as level 1 goals and the process recurs. The links in Figure 3 and their ordering are points of flexibility in the rule representation. For example, the rule author defines each goal and Its specific hypotheses (levels 1-2 and 4-5). Less trivially, it Is the rule author's choice to define rules that link hypotheses to one another (rules on level 3 link levels 2 and 4). We call the rationale behind this link the justification of the rule. GUIDON cannot teach rule justifications because they are not represented In MYCIN. Section 4 examines the nature of rule justifications and how a tutoring system can provide them. Next, by the rule author's ordering of hypotheses in a rule's premise, he will affect the order in which goals are pursued (level 5). The rationale for this choice again lies outside of the rule network. Thus, the program cannot explain why It pursues meningitis (goal 6.1 In Figure 1) before determining that the infection Is bacterial (goal 6.3). Section 5 examines how this ordering constitutes a strategy and how It can be made explicit. The order in which rules for a goal are tried (level 3) also affects the order In which hypotheses and hence subgoals are pursued (level 5). For example, rule 535 considers whether the patient Is an alcoholic, so if this rule is tried before rule 543, alcoholism will be considered before steroids. As these goals lead to asking questions of the user, it Is evident that the ordering of questions, when it derives from rule order as opposed to clause order, Is also determined by the ordering of rules. Here there Is no Implicit author rationale, for rule order lies outside of his choice; It Is parameter" dealing with kinds of surgery would replace individual "yes/no parameters" that specified "cardiac surgery," "neurosurgery," etc. These organizational decisions have no bearing on system performance, so the knowledge base Is somewhat Inconsistent In how these choices are made. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 10 fixed, but completely arbitrary. As pointed out above, MYCIN does not decide to pursue the hypothesis "bacterial meningitis" before "viral meningitis"--it simply picks up the bag of rules that make some conclusion about "kind of meningitis" and tries them In numeric order. Hence, rule order Is the answer to the question "Why is one hypothesis (for a given goal) considered before another?" And rule order is often the answer to "Why Is one question asked before another?" Focusing on a hypothesis and choosing a question to confirm a hypothesis are not necessarily arbitrary In human reasoning, raising serious questions about using MYCIN for interpreting a student's behavior and teaching him how to reason, also discussed In Section 5. To summarize, we have used a rule network as a device for illustrating aspects of MYCIN's behavior which it cannot explain. We are especially Interested In making explicit the knowledge that lies behind the behavior that is not arbitrary, but which cannot be explained because it Implicit in rule design, particularly the nature of the rule link between hypotheses ahd the ordering of hypotheses In rule premises. To do this, we will need some sort of framework for characterizing the knowledge Involved, since the rule link Itself is not sufficient. An epistemological framework for understanding MYCIN's rules is presented In the next section. 3 An Epistemological Framework for Rule-based Systems The framework presented In this section stems from an extensive study of MYCIN's rules. It is the basic framework that we have used for understanding physician's explanations of their reasoning, as well as being a foundation for re-representing the knowledge In MYCIN's rules. As an Illustration, we will consider In detail the "steroids rule" (Figure 4)2. RULE543 If: 1) The infection which requires therapy is meningitis, 2) Only circumstantial evidence is available for this case, 3) The type of the infection is bacterial, 4) The patient is receiving corticosteroids, Then: There is evidence that the organisms which might be causing the infection are e.coli (.4), klebsiella-pneumoniae (.2), or pseudomonas-aeruginosa (.1) Figure 4. The steroids rule. Figure 6 shows how this diagnostic heuristic is justified and Incorporated in a problem- solving approach by relating It to strategic, structural, and support knowledge. Recalling Figure 3, we use the term strategy to refer to a plan by which goals and hypotheses are ordered In problem solving. A decision to determine "cause of the Infection" before "therapy to administer" is a strategic decision. Similarly, a decision to pursue the hypothesis "e.coll Is causing meningitis" before "cryptococcus is causing meningitis" Is strategic. And recalling an earlier example, deliberately deciding to ask the user about steroids before alcoholism would be a strategic decision. These decisions all lie above the plane of goals and hypotheses, and as shown In Section 5, they can often be stated In 2 The English form of rules stated in this paper has been simplified for readability. Sometimes clauses are omitted. Medical examples are for purposes of Illustration only. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 12 domain-Independent terms. "Consider differential-broadening factors" Is an example of a domain-Independent statement of strategy. In order to make contact with the knowledge of the domain, a level of structural knowledge is necessary. Structural knowledge consists of abstractions that are used to Index the domain knowledge. For example, one can classify causes of disease Into common and unusual causes, as there are common and unusual causes of bacterial meningitis. These concepts provide a handle by which a strategy can be applied, a means of referencing the domain-specific knowledge. For example, a strategy might specify to consider common causes of a disease; the structural knowledge about bacterial meningitis allows this strategy to be Instantiated in that context. This conception of structural knowledge follows directly from Davis' [12) technique of content-directed invocation of knowledge sources. A handle Is a means of Indirect reference and Is the key to abstracting reasoning in domain-Independent terms. The discussion here (particularly Section 7) elaborates upon the nature of handles and their role In the explanation of reasoning. The structural knowledge we will be considering Is used to Index two kinds of hypotheses (as the term was used In Section 2): problem features which describe the problem at hand (for example, whether or not the patient Is receiving steroids Is a problem feature) and diagnoses which characterize the cause (disorder or disease) of the observed problem features (for example, acute meningitis Is a diagnosis). In general, problem features appear In the premise of diagnostic rules and diagnoses appear In the conclusion. Thus, organizations of problem features and diagnoses provide two ways of Indexing rule associations: one can use a strategy that brings certain diagnoses to mind, and consider rules that support those hypotheses; or one can use a strategy that brings certain problem 18 features to mind, gather that Information, and draw conclusions (apply rules) In a data- directed way. Figure 5 shows how a rule model or generalized rule3 , as a form of structural knowledge, enables both data-directed consideration of the steroids rule (via compromised host risk factors) or hypothesis-directed consiaeration (via unusual causes of meningitis). Illustrated are partial hierarchies of problem features (compromised host factors) and diagnoses (kinds of Infections, meningitis, etc.)--a typical form of structural knowledge. The specific organisms of the steroids rule are replaced by the set "gram-negative rods," a key hierarchical concept we use for understanding this rule. Finally, the justification of the steroids rule, a link between the problem feature hypothesis "patient Is receiving steroids" and diagnostic hypothesis "gram-negative rod organisms are causing acute bacterial infectious meningitis," is based on a causal argument about steroids Impairing the body's ability to control organisms that normally reside In the body. While this support knowledge is characteristically low-level or narrow In contrast with the strategical justification for considering compromised host risk factors, It still makes Interesting contact with structural terms, such as the mention of enterobacteriaceae (kinds of gram-negative rod organisms). In the next section, we will consider the nature of rule justifications in more detail, illustrating how structural knowledge enables us to make sense of a rule by tying It to the underlying causal process. 3 Davis' rule models (13], generated automatically, capture patterns, but they do not restate rules more abstractly as we Intend here. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 14 (STRATEGY) ESTABLISH HYPOTHESIS SPACE: CONSIDER DIFFERENTIAL-BROADENING FACTOR (RULE MODEL) IN BACTERIAL MENINGITIS, COMPROMISED HOST RISK FACTORS SUGGEST UNUSUAL ORGANISMS ANY-DISORDER INFECTION (STRUCTURE) MENINGITIS COMPROMISED HOST ACUTE CHRONIC CURRENT BACTERIAL VIRAL MEDICATIONS UNUSUAL-CAUSES SKIN ORGS (INFERENCE RULE) if STEROIDS then GRAM-NEGATIVE ROD ORGS (SUPPORT) STEROIDS IMPAIR IMMUNO-RESPONSE MAKING PATIENT SUSCEPTIBLE TO INFECTION BY ENTEROBACTERIACEAE, NORMALLY FOUND IN THE BODY. Figure 5. Knowledge for Indexing, justifying, and Invoking a MYCIN Rule 4 Justifying a Rule Here we consider the logical bases for rules: What kinds of arguments justify the rules and what is their relation to a mechanistic model of the domain? We use the terms "explain" and "justify" synonymously, though the sense of "making clear what Is not understood" (explain) is Intended more than "vindicating, showing to be right or lawful" (justify). 4.1 Different Kinds of Justifications There are four kinds of Justifications for MYCIN's rules: Identification, causal, world fact, and definition. In order to explain a rule, it is first necessary to know what kind of justification it Is based upon. Rules that use identifying properties of an object to classify it are called i(dentification rules. Most of MYCIN's rules that use laboratory observations of an unknown are like this: "if the organism Is a gram-negative, anaerobic rod, Its genus may be bacteroldes (.6)." Thus, an Identification rule is based on the properties of a class. Rules whose premise and action are related by a causal argument are called causal rules. The causality can go in either direction In MYCIN rules: "symptom caused by disease" or (more commonly) "prior problem causes disease". Szolovits suggests that It Is possible to subdivide causal rules according to the scientific understanding of the causal link: empirical association (a correlation for which the process Is not understood), complication (direction of causality is known, but the conditions of the process are not understood), and mechanistic (process Is well-modeled) (after Szolovits [28]). The CF's In MYCIN's causal rules generally represent a mixture of probabilistic and cost/benefit Judgment. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 16 Rules that are based on empirical, common sense knowledge about the world are called world fact rules. An example is, "if the patient is male, then the patient Is not pregnant or breast feeding." Other examples are based on social patterns of behavior, such as the fact that a young male might be a military recruit (and so living in a crowded environment where disease spreads readily). Domain fact rules link hypotheses on the basis of domain definitions. An example is "if a drug was administered orally and It is poorly absorbed In the GI tract, then the drug was not administered adequately." (By definition, to be administered adequately a drug must be present in the body at high enough dosage levels.) By using domain fact rules, the program can relate problem features to one another, reducing the amount of Information it has to request from the user. In summary, a rule link captures class properties, social and domain facts, and probabilstic and cost/benefit judgments. When a definition, property or world fact is Involved, simply saying this provides a reasonable explanation. But causal rules, with their connection to an underlying process of disease, require much more, so we will concentrate on them. 4.2 Levels of explanation -- what's not In a rule? In this section we will consider the problem of justifying a causal rule, the tetracycline rule, "If the patient is less than 8 years old, don't prescribe tetracycline." This rule simply states one of the things that MYCIN needs to know to properly prescribe drugs for youngsters. The rule does not mention the underlying Causal process (chelation- -drug deposition In developing bones) and the social ramifications (blackened permanent teeth) Levels of explanation -- what's not In a rule? 17 upon which It is based. From this example, It should be clear that the Justifications of MYCIN's rules lie outside of the rule base. In other words, the record of Inference steps that ties premise to action has been left out. A few issues need to be raised here: Did the expert really leave out steps of reasoning? What Is a justification for? And what Is a good Justification? Frequently we refer to rules like MYCIN's as "compiled knowledge." However, when we ask physicians to justify rules that they believe and follow, they very often can't explain why they are correct, or rationales are so long In coming and so tentative, it Is clear that we are not being told reasoning steps that are consciously followed. Leaps from data to conclusion are justified because the intermediate steps (like the process of chelation and the social ramifications) generally remain the same from problem to problem. There Is no need to step through this knowledge--to express It conditionally In rules. Thus, for the most part, MYCIN's rules are not compiled in the sense that they represent a deliberate composition of reasoning steps by the rule authors. They are compiled In the sense that they can be expanded (decomposed). If an expert does not think about the reasoning steps that justify a rule, why does a student need to be told about them? One simple reason why a student needs a justification for a rule Is so that he can remember the rule. A justification can even serve as memory aid (mnemonic) without being an accurate description of the underlying phenomena. For example,, medical students have long been told to think In terms of "bacteria eating glucose" from which they can remember that low CSF glucose Is a sign of a bacterial (as opposed to fungal or viral) meningitis. The interpretative rule Is learned by analogy to a familiar association (glucose Is a food and bacteria are analogous to larger organisms that THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 18 eat food). This explanation has been discredited by biological research, but It Is still a useful mnemonic. Given that an accurate causal argument is usually expected, how Is a satisfying explanation constructed? To see the difficulty here, observe that, In expanding a rule, there Is seemingly no limit to the details that might be included. Imagine expanding the tetracycline rule by Introducing three intermediate concepts: tetracycline in youngster 0> chelation of the drug In growing bones =0 teeth discoloration 0> undesirable body change :) don't administer tetracycline The choice of Intermediate concepts is arbitrary, of course. For example, there Is no mention of how the chelation occurs. What are the conditions? What molecules or Ions are Involved? There are levels of detail In a causal explanation. To explain a rule, we not only need to know the intermediate steps, we need to decide which steps in the reasoning need to be explained. Purpose (how deep an understanding is desirable) and prior knowledge are obviously Important. Conceptually, the support knowledge for a causal rule Is a tree of rules, where each node Is a reasoning step that can theoretically be justified In terms of finer-grained steps. The Important thing to remember is that MYCIN Is a flat system of rules. It can only state Its Immediate reasoning steps, and cannot explain them on any level of detail. 4.3 Problem features, the hypothesis taxonomy, and rule generalizations A tree of rules seems unwieldy. Surely most teachers cannot expand upon every Problem features, the hypothesis taxonomy, and rule generalizations 1i reasoning step down to the level of the most detailed physical knowledge known to man. (The. explanation tree for the tetracycline rule quickly gets into chemical bonding theory.) Explaining a rule (or understanding one) does not require that every detail of causality be considered. Instead, a relatively high level of explanation is generally satisfying--you probably felt satisfied by the explanation that tetracycline causes teeth discoloration. This level of "satisfaction" has something to do with the student's prior knowledge. For an explanation to be satisfying, It must make contact with already known concepts. We can characterize explanations by studying the kinds of Intermediate concepts they use. For example, It is significant that most contraindication rules, reasons for not giving antibiotics, will refer to "undesirable body changes." This pattern Is Illustrated hierarchically below. (The first level gives types of undesirable changes; the second level gives causes of these types.) undesirable body changes/ I tt,(<-'types' IZ'photosensitivity, diarrhea, ...teeth nausea discoloration tetracycline drug x Figure 7. Problem feature hierarchy for contraindication rules Notice that this figure contains the last step of the expanded tetracycline rule, and a leap from tetracycline to this step. The pattern connecting drugs to the Idea of undesirable body changes forms the basis of an expectation for explanations: we will be satisfied If a particular explanation connects to this pattern. In other words, given an effect that we can Interpret as an undesirable body change, we will understand why a drug causing that effect should not be given. We might want to know how the effect occurs, but here again, we will THE EPISTEMOLOGY Of A RULE-BASEO EXPERT SYSTEM 20 rest easy on Islands of familiarity, Just as we don't feel compelled to ask. why people don't want black teeth. To summarize, key concepts in rule explanations are abstractions that connect to a pattern of reasoning we have encountered before, premises that we readily accept. This suggests that one way to explain a rule, to make contact with a familiar reasoning pattern, Is to generalize the rule. We can see this more clearly from the viewpoint of diagnosis, which makes rich use of hierarchical abstractions. Consider the rule fragment, "if a complete blood count Is available and the white blood count Is less than 2.6 units, then the following bacteria might be causing Infection: e.coll (.75). pseudomonas-aeruginosa (.5), klebsielia-pneumoniae (.6)." How can we explain this rule? First, we generalize the rule (Figure 8). level 3 Compromised cc> bacteria normally found In host condition the body cause infection "causes" 0 "subtype* T "subset" pregnancy I 2 inmunosuppresslon ==> Gram negative rods and condition enterobacteriaceae "causes" "evidence" "i s a" steroids I I leukopenla ::> E.coli, Pseudomonas, \e eand Klebsiella"evidence= WBC 2.5 PNNS > 188 \ -'componentof CC data Figure 8. Generalizations of the leukopenia rule The premise concepts In the rules on levels 1 through 3 are problem features (recall discussion in Section 3), organized hierarchically by different kinds of relations. Generally, - ii Problem featw es, the hypothesis taxonomy, and rule generalizations 21 a physician speaks loosely about the connections--referrIng to "leukopenia" both as a cause of immunosuppresslon as well as a kind of Immunosuppresslon--probably because the various causes are thought of hierarchically. The similarity of this part of the diagram to Figure 7 should be evident. The relationships among CBC, WBC and Leukopenla reveal some Interesting facts about how MYCIN's rules are constructed. WBC is one component of a "complete blood count" (CBC). If the CBC Is not available, it makes no sense to ask for any of the components. Thus, the CBC clause In the leukopenia rule Is an example of a screening clause. Another example of a screening clause Is the age clause In the rule fragment, "if ... age Is greater than 1 7 and the patient is an alcoholic, then .... " Here the relation Is a social fact; If the patient Is not an adult, we assume that he Is not an alcoholic. The third relation we observe Is subtype, as in "if ... the patient has undergone surgery and the patient has undergone neurosurgery, then...." All screening relations can be expressed as rules, and some are, such as "If the patient has not undergone surgery, then he has not undergone cardiac surgery" (stated negatively, as Is procedurally useful). MYCIN's rule set Is inconsistent in this respect; to be economical and make the relationship between clauses explicit, all screening clauses should be expressed as rules. Indeed, the age/alcoholic relation suggests that som of the relations are uncertain and should be modified by certainty factors. But even when rules explicitly link problem features, (as a rule could state "if no CBC was taken, then WBC is not available"), the Aind of relation Is not represented (WBC Is a component of a CBC) because MYCIN's rule language does not allow the link to be labeled In this way. Finally, when one problem feature serves as a redefinition of another, such as the THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 22 relation between leukopenia and WBC, the more abstract problem feature tends to be left out altogether ("leukopenia" is nt a MYCIN parameter, the rule mentions WBC directly). For purposes of explanation, we argue that problem features, their relations, and the nature of the link should be explicit. Returning to Figure 8, the action concepts, diagnostic hypotheses, are part of a large hierarchy of causes that the problem solver will cite In the final diagnosis. The links In this diagnosis space generally specify refinement of cause, though in our example they strictly designate subclasses. Generally, problem features are abstractions of patient states Indicated by the observable symptoms, while the diagnosis space is made up of abstractions of causal processes that produce the symptoms. Paralleling our observations about rule problem features, we note that the relations among diagnostic hypotheses are not represented in MYCIN--nowhere In the knowledge base does It explicitly state that e.colU is a bacterium. Now suppose that the knowledge in Figure 8 were available, how would this help us to explain the leukopenla rule? The Idea is that we first restate the rule on a higher level. We point out that a low WBC Indicates leukopenia, which is a form of Immunosuppresslon, thus tying the rule to the familiar pattern that Implicates gram-negative rods and enterobacterlaceae. This is directly analogous to pointing out that tetracycline causes teeth discoloration, which Is a form of undesirable body change, suggesting that the drug should not be given. By rerepresenting Figure 8 linearly, we see that it Is an expansion of the original rule: WBC < 2.5 0) Leukopenia > Immunosuppression a> compromised host . Infection by organisms found in body 0 gram-negative rods & enterobacterlaceae 0) e.coii, paeudomonas, and klebsiella. Problem features, the hypothesis taxonomy, and rule generalizations 23 The expansion marches up the problem feature hierarchy and then back down the hierarchy of diagnoses. The links of this expansion Involvd causality composed with Identification, subtype and subset relations. By the hierarchical relationships, a rule on one level "explains" the rule below It. For example, the rule on level 3 provides the detail that links Immunosuppression to the gram-negative rods. By generalizing we have made a connection to familiar concepts. Tabular rules provide an interesting special case. The CSF protein rule (Figure 9) appears to be quite formidable. RULE508 If: 1) The infection which requires therapy is meningitis, 2) A lumbar puncture has been performed on the patient, and 3) The CSF protein is known Then: The type of the infection Is as follows: If the CSF protein is: a) less than 41 then: not bacterial (.5), viral (.7), not fungal (.6), not tb (.5); b) between 41 and 109 then: bacterial (.A), viral (.4), fungal (.1); c) between 198 and 298 then: bacterial (.3), fungal (.3), tb (.3); d) between 288 and 388 then: bacterial (.4), not viral (.5), fungal (.4), tb (.4); e) greater or equal to 388 then: bacterial (.4), not viral (.6), fungal (.4), tb (4); Figure 9. The CSF protein rule Graphing this rule (Figure 10), we find a relatively simple relation that an expert stated as, "If the protein value is less than 40, I think of viral; If it Is more than 100, I think of bacterial, fungal or TB." This is the first level of generalization, the principle that Is Implicit In the rule. The second level elicited from the expert is, "If the protein value Is low, I think of an acute process; if it Is high, I think of a severe or long term process." 4 Then, a 4 Bacterial meningitis Is a severe, acute (short term) problem, while fungal and TB meningitis are problems of long (chronic) duration. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 24 level higher, "an infection In the meninges stimulates protein production." So in moving up abstraction hierarchies on both the premise and action side of the rule (acute and chronic are subtypes of infection), we arrive at a mnemonic, Just like "bacteria eat glucose." Abstractions of both the observations and the conclusions are important for understanding the rule. Bel ief + CF TYPE of MENINGITIS given the CSF PROTEIN .8 OF8- -V-'- V '' , 58 le 156 266 250 306 CSF Protein Value -. 4 -BT -F -. 8 Key: B z BACTERIAL T = TUBERCULOSIS V = VIRAL F = FUNGAL Figure 10. Graph of the conclusions made by rule 500 (Figure 9) We might be surprised that explanations of rules provide levels of detail by referring to more general concepts. We are accustomed to the fact that principled theoretical explanations of, say, chemical phenomenon, refer to atomic properties, finer-grained levels of causality. Why should a rule explanation refer to concepts like "compromised host" or "organisms normally found In the body"? The reason is that In trying to understand a rule like the steroids rule, we are first trying to relate it to our understanding of what an Infection Is at a high, almost metaphorical level. In fact, there are lower level "molecular" details of the mechanism that could be explained, for example, how steroids actually Problem features, the hypothesis taxonomy, and rule generalizations 25 change the immunological system. But our first goal as understanders, our focus, Is at the top level--to link the problem feature (steroids) to the global process of meningitis infection. We ask, "What makes it happen? What role do steroids play In the Infectious meningitis process?" The concept of "compromised host" Is a label for a poorly understood causal pattern that has value because we can relate It to our understanding of the Infectious process. It enables us to relate the steroids or WBC evidence to the familiar metaphor in which Infection is a war that is fought by the body against Invading organisms. (If a patient Is compromised, his defenses are down; he is vulnerable to attack.) In general, causal rules argue that some kind of process has occurred. We expect a top-level explanation of a causal rule to relate the premise of the rule to our most general Idea of the process being explained. This provides a constraint for how the rule should be generalized, the subject of the next section. 4.4 Tying an Explanation to a Causal Model MYCIN's diagnostic rules are arguments that a process has occurred In a particular way, namely, that an Infectious process has transpired In the patient along certain lines. There are many kinds of Infections, which have different characteristics, but bacterial Infections tend to follow the same script: entry of an organism Into the body, passage of the organism to the site of infection, reproduction of the organism, and causing of observable symptoms. An explanation of a rule that concludes that an organism Is causing an Infection must demonstrate that this generic process has occurred. In short, this Is the level of abstraction that the explanation must connect to. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 26 A program was written to demonstrate this idea. The data parameters in MYCIN's 40 diagnostic rules for bacterial meningitis are restated as one or more of the steps of the Infectious process script. This restatement is then printed as the explanation of the rule. For example, the program's explanation of the alcoholic rule (alcoholism > diplococcus meningitis) is. The fact that the patient Is an alcoholic allows access of organisms from the throat and mouth to lungs (by reaspiration of secretions). The fact that the patient Is an alcoholic means that the patient is a compromised host, and so susceptible to Infection. Words in Italics (of the first sentence) constitute the pattern of "portal and passage." We find that the premise of a rule generally supplies evidence for only a single step of the causal process: the other steps must be inferred by default. For example, the alcoholic rule argues for passage of the diplococcus to the lungs. The person reading this explanation must know that diplococcus is normally found in the mouth and throat of. any person and it proceeds from the lungs to the meninges by the blood. The organism finds conditions favorable for growth because the patient is compromised, as stated in the explanation. In contrast, the leukopenla rule only argues for the patient being a compromised host, so the organisms are the default organisms, those already in the body which can proceed to the site of Infection. 5 These explanations say which steps are enabled by the data. They place the patient on the path of an infection, so to speak, and leave It to the understander to fill In the other 6 As might be expected, alcoholism also causes Infection by the gram negative rods and enterobacterlaceae. We have omitted these for simplicity. However, this example Illustrates that a MYCIN rule can have multiple conclusions reached by different causal paths. Tying an Explanation to a Causal Model 27 steps with knowledge of how the body normally works. This Is why the physician generally refers to the premise data as "predisposing factors." (MYCIN's heuristics are generally of the form "predisposing factor causes disease.") They argue that prior steps in a causal process have occurred. it is to the level of these prior steps, general concepts that explain many rules, that the rule should be related In order to be understood. The process of explanation Is a bit more complicated In that a causal relation may exist between clauses In the rule. We have already seen that one clause may screen another on the basis of world facts, multi-component test relations, and the subtype relation. The program described here knows these relations and "subtracts off" screening clauses from the rule. Moreover, as discussed in Section 6, some clauses describe the context in which the rule applies (the role of the first 3 clauses in the Steroids rule, Figure 4). These, too, are maide explicit for the explanation program, and subtracted off. In the vast majority of MYCIN rules, only one premise clause remains, and this is related to the process of infection in the way described above. When more than one clause remains after screening and contextual clauses have been removed, our study shows that a causal connection exists between the remaining clauses. We can always isolate one piece of evidence that the rule is about (for example, WBC in the Leukopenia rule); we call this the key factor of the rule. We call the remaining clauses restriction clauses and observe three kinds of relations between a restriction clause and a key factor: -- A confirmed diagnosis explains a symptom. (Example: a petechial rash would normally be evidence for neisseria, but If the patient has leukemia, It may be the disease THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 28 causing the rash. Therefore the rule states, "if the patient has a petechial rash (the key factor) and does not have leukemia (the restriction clause), then nelsserla may be causing the meningitis.") -- Two symptoms In combination suggest a different diagnosis than one taken alone. (Example: When both purpuric and petechial rashes occur, then a virus Is a more likely cause than neisseria. Therefore, the petechial rule also Includes the restriction clause "the patient does not have a purpuric rash.") -- Weak circumstantial evidence Is made Irrelevant by strong circumstantial evidence. (Example: a head Injury so strongly predisposes a patient to Infection by skin organisms that the age of the patient, a weak circumstantial factor, Is made irrelevant.) Restriction clauses are easy to detect when examining the rule set because they are usually stated negatively ("some problem feature or diagnosis does not apply"). These examples are suggestive of the causal reasoning about problem features that we find in diagnosis when the causality of the system is better understood, as in electronic troubleshooting. In summary, to explain a causal rule, a teacher must know the purposes of the clauses and connect the rule to abstractions in the relevant process script. 4.5 The Relation of Medical Heuristics to Principles It might be argued that we have to go to so much trouble to explain MYCIN's rules because they are written on the wrong level. Now that we have a "theory" for which The Relation of Medical Heuristics to Principles 29 intermediate parameters to Include ("portal," "pathway," etc.), why don't we simply rewrite the rules? The medical knowledge we are trying to codify is really on two levels of detail: 1) principles or generalizations, and 2) empiric details or specializations. MYCIN's rules are empiric. Cleaning them up by representing problem feature relationships explicitly would give us the same set of rules at a higher level. But what would happen If process concepts were Incorporated In completely new reasoning steps, for example, If the rule set related problem features to hypotheses about the pathway the organism took through the body? It turns out that reasoning backwards in terms of a causal model Is not always appropriate. As we discovered when explaining the rules, not all of the causal steps of the process can be directly confirmed; we can only assume that they have occurred. For example, rather than providing diagnostic clues, the concept of "portal of entry and passage" is very often deduced from the diagnosis Itself. According to this view, principles are good for summarizing arguments, and good to fall back on when you've lost grasp on the problem, but they don't drive the process of medical reasoning. Specifically, (1) If a symptom needs to be explained (is highly unusual). we ask what could cause it ("Strep-viridans? It is normally found In the mouth. How did it get to the heart? Has the patient had dental work recently?"); (2) to "prove" that the diagnosis is correct (after it has been constructed), we use a causal argument ("He has pneumonia; the bacteria obviously got into the blood from the lungs"). Thus, causal knowledge can be used to provide feedback that everything fits. It may be difficult or Impossible to expect a set of diagnostic rules to serve both as concise, "clincher" methods for efficiently getting to the right data, and still represent a THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 30 model of disease. Put another way, a student may need the model If he is to understand new associations between disease and manifestations, but he will be an Inefficient problem solver If he always attempts to directly convert that model to a subgoal structure for solving ordinary problems. Szolovits [28] points out that these "first principles," used by a student, are "compiled out" of an expert's reasoning. In meningitis diagnosis, the problem is to manage a broad, if not Incoherent, hypothesis set, rather than to pursue a single causal path. The underlying theory recedes to the background, and the expert tends to approach his problem simply In terms of weak associations between observed data and bottom-line conclusions. This may have promoted a rule-writing style that discouraged introducing intermediate concepts, even where they might have been appropriate, for example, the concept of "leukopenla" described above. 31 6 Teaching Problem-Solving Strategy A strategy is an approach for solving a problem, a plan for ordering methods so that a goal is reached. it is well-accepted that strategic knowledge must be conveyed In teaching diagnostic problem solving: Without explicit awareness of the largely tacit planning and strategic knowledge inherent In each domain, it is difficult for a person to "make sense of" many sequences of behavior as described by a story, a set of instructions, a problem solution, a complex system, etc.... The teacher should articulate for that domain the higher-order planning knowledge and strategic knowledge for formulating and revising hypotheses about what something means. [3] Strategic knowledge is general, much like the principles of mechanism we discussed earlier; both relate to processes that have structure. Thus, it Is not sufficient to merely show a student MYCIN's solution, the surface structure of the program, we must explain why the rules are Invoked in a particular order. Here it is clear how teaching how to do something differs from merely explaining how something was done: we want the student to be able to replicate what he observes, to go off and solve similar problems on his own. This Is why providing mnemonics Is useful when justifying a rule. Regarding strategy, we must again address human foibles and preference: we must teach a strategy that a human can follow. The main points of this section are: -- MYCIN's strategy is different from a physician's strategy THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 32 -- MYCIN's strategic knowledge is embedded in the rules, indistinguishable from screening and problem feature clauses. -- A domain-independent represt Ltation of strategy Is useful for teaching and for purposes of economy. 5.1 Surface and Deep Structure of MYCIN A person trying to understand a MYCIN consultation observes that pieces of data are requested by the program (Figure 11). Conceptually, these questions are terminals hanging below an AND (rule) node in a subgoal tree (one portion corresponding to Figure 1 1 Is shown In Figure 12). Observe that this example, considering the order of questions asked by the program, is a special case of the rule network shown earlier (Figure 3)--here we show only the hypotheses (Figure 3, level 4) that lead directly to a question. Following the terminology of Brown, et al. [3], a rule node Is a method for achieving a goal (e.g., "organisms that might be causing the infection") by asking questions or pursuing a set of subgoals to achieve (the premise of a particular rule). Therefore, the tree of rules and subgoals is part of a deep-structured trace that Brown, et al. postulate is constructed when the understander makes sense of the surface problem solution. Surface and Deep Structure of MYCIN 33 31) Has Pt538 ever undergone any type of' surgery? ** YES 32) Does Pt538 have a history of NEUROSURGERY? ** NO 33) Does Pt538 live in a crowded environment? ** No 34) Do you suspect recent alcoholic history in Pt538? ** No 35) Is meningitis a hospital-acquired infection? ** YES 36) Is Pt538's clinical history consistent with EPIGLOTTITIS? ** NO 37) Is Pt538-s clinical history consistent with OTITIS-fIEDIA? ** No 38) Has Pt538 ever undergone splenectony? ** NO 39) Is Pt538 a burn patient? a'YES Figure 11. MYCIN consultation excerpt GOAL HYPOTHESIS METHOD QUESTION E.COLI -(Rule5ll)- Q32 NEUROSURGERY -NEISSERIA -(Rule533)- Q33 CROWD -D.PNU r 1(11,e035)- Q34 ALCOHOLIC (Rule559)- Q38 SPLENECTOIY COVERFOR - H.NFL j (Rle55)-Q35 NOSOCOMIAL (Rul395) Q36 EPIGLOTTITIS Q37 OTITIS-MEDIA PSEUDOlO. -(Rule578)- Q39 BURN Figure 12. Portion of the AND/OR tree corresponding to the questions shown In Figure 11 (reorganized according to the hypothesis each rule supports). THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 34 It is not sufficient for a student to know all of the possible methods he can bring to bear on a problem. For example, a student who knows what kinds of algebraic transformqtlons can be used to solve for "x" in "x*'2 - 8 a1" could only proceed to apply the methods randomly if he didn't have a plan for solving the problem (that Is, have schemas for kinds of problems that can be tackled using different approaches or lines of reasoning.) A plan sets up a rational sequence of applications of methods that might get you closer to the solution (though this Is not guaranteed). The hypothetico-deductive strategy used in medical problem solving constitutes a plan for focusing on hypotheses and selecting confirmatory questions [1 7]. However, the methods selected In Figure 12 (rules 511 through 578) have been applied In a fixed, arbitrary order--not planned by the rule author. MYCIN has no "deep structure" plan at this level; the program is simply applying rules (methods) exhaustively. This lack of similarity to human reasoning severely limits the usefulness of the system for teaching problem-solving. However, MYCIN does have a problem solving strategy above the level of rule application, namely the control knowledge that causes it to pursue a goal at a certain point In the diagnosis. We can see this by examining how rules Interact In backward chaining. Figure 13 shows the goal rule and a rule that it Indirectly Invokes. Surface and Deep Structure of MYCIN 385 RuleS92 (The Goal Rule) If: 1) Gather information about cultures taken from the patient and therapy he is receiving, 2) Determine if the organisms growing on cultures require therapy 3) Consider circumstantial evidence for additional organisms that therapy should cover Then: Determine the best therapy recommendation RULE535 (The Alcoholic Rule) If: 1) The infection which requires therapy is meningitis, 2) Only circumstantial evidence is available for this case, 3) The type of meningitis is bacterial, 4) The age of the patient is greater than 17 years, and 5) The patient is an alcoholic. Then: There Is evidence that the organisms which might be causing the infection are diplococcus-pneumoniae (.3) or e.coli (.2) Figure 13. The goal rule and alcoholic rule. In order to evaluate the third clause of the goal rule, MYCIN tries each of the "cover for" rules; the alcoholic rule Is one of these (see also Figure 12). We call the goal rule a task rule to distinguish it from inference rules. Clause order counts here; this Is a procedure, not a logical conjunction. MYCIN has a few other task rules, but procedural knowledge appears In almost every rule. The first three clauses of the alcoholic rule, the context clauses, really control the order in which goals are pursued, just as In a task rule. We can represent this hidden structure of goals by a tree which we call the Inference structure of the rule base (produced by "hang,ng" the rule set from the goal rule). Figure 14 Illustrates part of MYCIN's Inference structure. THE EPISTEMOLOGY OF A RULEBASED EXPERT SYSTEM 36 REGIMEN = main goal rule2,c12 TREATFO COVERFOR a rule92, c13 (9)1 W"AT-INF? SIGNIJICANT? IDENTITY? MENINGIT\IS? ACTERIAL? (2)1 (4) (5) ()8 INFECION? CONTA INANT? INFECTON? (1) (3) (6) Figure 14. Portion of MYCIN's Inference structure (Numbers give the order in which non-place-holder goals are achieved by the depth-first interpreter.) The first three goals on the third level appear In a single task rule, while the last two goals (meningitis? and bacterial?) correspond to the first and third clauses of the 40 "cover for" rules similar to the alcoholic rule. The program's strategy comes to light when we list these goals in the order in which they are actually achieved. (By "achieved" we mean that all rules that conclude about that goal have been tried.) For example, under the depth-first control of the Interpreter, all "bacterial?" rules must be applied before any "cover for" rules are completed, so "bacterial?" Is achieved before "cover for." This gives us the ordering (numbers correspond to the numbering In Figure 14): 1. Is there an Infection? 2. Is It bacteremia, cystitis, or meningitis? 3. Any contaminated cultures? 4. Any good cultures with significant growth? 5. Is the organism Identity known? 6. Is there an infection? (already done In step 1) 7. Does the patient have meningitis? (already done in step 2) 8. Is It bacterial? . .. ..... , -I I II i4 Surface and Deep Structure of MYCIN 37 9. Are there specific bacteria to cover for? (We leave out the goals "regimen" and "treatfor" because they are just place holders for task rules, like subroutine names.) MYCIN's diagnostic plan Is in two parts, and both proceed by top-down refinement. This demonstrates that a combination of structural knowledge (the diagnosis space taxonomy--infection, meningitis, bacterial, diplococcus...) and strategic knowledge (traversing the taxonomy from the top down) is procedurally embedded in the rules. In other words, we could write a program that Interpreted an explicit, declarative representation of the diagnosis taxonomy and domain-independent form of the strategy to bring about the same effect. At this level, MYCIN's diagnostic strategy Is not a complete model of how physicians' think, but it could be useful to a student. As the quote from Brown et al. would Indicate, and has been confirmed In GUIDON research, teachers do articulate both the structure of the problem space and the nature of the search strategy to students. This means that we need to explicitly represent the fact that the diagnosis space Is hierarchical, and represent strategy in a domain-independent form. If the strategy Is not in domain-independent form, It can be taught by examples, but not explained. 5.2 Representing Strategic Knowledge in Meta-rules How might we represent domain-independent strategic knowledge In a rule-based system? In the context of the MYCIN system, Davis pursued the representation of strategic knowledge by using meta-rules to order and prune methods. These meta-rules are Invoked just before the object-level rules are applied to achieve a goal. An example of an Infectious disease meta-rule is shown in Figure 16. Observe that this Is a strategy for S..... ...... .... i THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 38 pursuing a goal. In particular, this meta-rule might be associated with the goal "identity of the organism." It will be Invoked to order the rules for every subgoal In the search tree below this goal; In this simple way, the rule sets up a line of reasoning. This mechanism causes some goals to be pursued before other, orders the questions asked by the system, and hence changes the surface structure of the consultation. META-RULEee I If: 1) the infection is pelvic-abscess, and 2) there are rules which mention in their premise enterobacteriaceae, and 3) there are rules which mention in their premise grampos-rods, Then: There is suggestive evidence (.4) that the former should be done before the latter. Figure 15. A MYCIN meta-rulez While meta-rules like this can capture and Implement strategic knowledge about a domain, they have their deficiencies. For ike the MYCIN performance rules we have examined, Davis's domain-dependent examples of meta-rules leave out knowledge Important for explanation. Not only do they, like the object-level rules, leave out the domain-specific support knowledge that justifies the rules, they leave out the domain- Independent strategic principles that GUIDON should teach. In short, meta-rules provide the mechanism for controlling the use of rules, but not the domain-independent language for making the strategy explicit. The implicit strategic principle that lies behind meta-ruleOO1 is that common (frequent) causes of a disorder should be considered first. The structural knowledge that ties this strategy to the object-level diagnostic rules is an explicit partitioning of the diagnosis space taxonomy, Indicating that the group of organisms called the Representing Stralegic Knowledge In Mete-rules 39 enterobacterlaceae are more likely than gram-positive rod organisms to cause pelvic infections. This is what we want to teach the student. One can Imagine, for different infection types, different common causes, requiring a different meta-rule for each Infectio. But if all mete-rules are as specific as meta-rule-O01, principles will be compiled into many rules redundantly and the teaching points will be lost. What does a domain-Independent mete-rule look like, and how Is it Interfaced with the object-level rules? To explore this question, we have reconfigured the MYCIN rule base into a new system, called NEOMYCIN [9]. Briefly, meta-rules are organized hierarchically (again!) into tasks, such as "group and refine the hypothesis space." These rules manage a changing hypothesis list by applying different kinds of knowledge sources, as appropriate. Knowledge sources are essentially the object level rules, Indexed In the diagnosis space taxonomy by a domain-independent structural language. For example, one meta-rule for achieving the task of pursuing an hypothesis Is "If there are unusual causes, then pursue them." 6 Suppose that the current hypothesis Is "bacterial meningitis." The program will use the structural label "unusual causes" to retrieve the nodes "gram-negative rods", "enterobacteriaceae", and "listerla", add them to the hypothesis list, and pursue them in turn. (When there are no "unusual causes" indicated, the meta-rule simply does not apply.) Pursuing gram-negative rods, the program will find that leukopenia is a relevant factor, but first ask if the patient Is a compromised host (Figure 8), modelling a physician's efficient casting of wider questions. 6 This rule appears after the rule for considering common causes, and the ordering Is marked as strategically significant. Domain-independent mete-rules have justifications, organization, and strategies for using them. Their justification refers to properties of the search space and the processor's capabilities. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 40 Other terms In the structural language used by NEOMYCIN's domain Independent meta- rules are: (disease) process features, such as extent and location; the enabling step of a causal process; subtype; cause; trigger association; problem feature screen; and structure. properties of the taxonomy, such as sibling. In effect, the layer of "structural knowledge" allows us to separate out what the heuristic Is from how It will be used. How domain- specific heuristics (like MYCIN's rules) should be properly Integrated with procedural, strategic knowledge Is an Issue at the heart of the "declarative/procedural controversy" (30]. We conclude here that, for purposes of teaching, the hierarchies of problem features and the diagnosis space should be represented explicitly, providing a useful means for Indexing the heuristics by both their premise and action (Figure 5). A structural language of cause, class, and process connects this domain-specific knowledge to domain-independent meta-rules, the strategy for problem solving. 5.3 Self-Referencing Rules Self-referencing rules provide an interesting special example of how problem solving strategies can be embedded in MYCIN's rules. A rule is self-referencing if the goal concluded by the action is also mentioned in the premise. An example Is the "aerobicity rule" (Figure 16).7 7 Aerobicity refers to whether an organism can grow in the presence of oxygen. A facultative organism can grow with or without it; a non-aerobic organism cannot grow with oxygen present; and an obligate-aerob Is aerobic only In a certain stage of growth. Self-Referencing Rules 41 Ru I e986 If: 1) The aerobicity of the organism is not known and 2) The culture was obtained more than 2 days ago, Then: There is evidence that the aerobicity of the organism is obligate-aerob (.5) or facultative (.5) Figure 16. The self-referencing aerobicity rule. This rule Is tried only after all of the non-self-referencing rules have been applied. The conclusion of these rules Is held fixed, then the self-referencing rules are tried. The effect Is to reconsider or reflect upon a tentative conclusion. When the original conclusion Is changed by the self-referencing rules, this is a form of non-monotonic reasoning [31]. We can restate MYC(NWs self-referencing rules in domain-independent terms: -- If nothing has been observed, consider situations that have no visible manifestations. (Example: The aerobicity rule--"if no organism Is growing in the culture, It may be an organism that takes a long time to grow (obligate-aerob and facultative organisms).") The self-referencing mechanism makes It possible to state this rule without requiring a long premise that is logically exclusive from the remainder of the rule set. -- If unable to make a deduction, assume the most probable situation. (Example: "if the gram stain is unknown and the organism is a coccus, then assume that It Is gram positive.") -- If there is evidence for two hypotheses, A and 8, that tend to be confused, then rule out B. (Example: "if there is evidence for Tb and Fungal, and you have hard data for Fungal, rule out Tb.") THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 42 Like Meta-ruleO01, self-referencing rules provide a useful mechanism for controlling the use of knowledge, but they leave out both the domain-dependent justification end the general, domain-independent reasoning strategy of which they are examples. These rules Illustrate that strategy Involves more than a search plan; It also takes In principles for reasoning about evidence, what Collins calls "plausible reasoning" [10]. It is not clear that a teacher needs to explicitly state these principles to a student. They tend to be either "common sense" or almost impossible to think about Independently of an example. Nevertheless, they are yet another example of strategic knowledge that Is Implicit In MYCIN's rules. / 43 8 Implications for Modifiability and Performance MYCIN's rule authors believed that the program achieved good problem-solving performance without the structural, strategic, and support knowledge we have been considering. However, It Is possible to Imagine situations In which knowledge of justification and strategy allows one to be a more flexible problem solver, to cope with novel situations. MYCIN cannot solve some kinds of difficult problems, the problems we say only an expert can solve. Knowing the basis of a rule allows you to know when not to apply It, or how to modify It for special circumstances. For example, knowing that tetracycline won't kill the patient, but the Infection might, you may have to dismiss social ramifications and prescribe the drug. You can deliberately break the rule because you understand it. There will also be problems which cannot be diagnosed using MYCIN's rules. For example, several years ago coccidioides meningitis strangely appeared In the San Francisco • ay Area. We would say that such a case "violates all the rules." To explain what was happening, one has to reason about the underlying mecharnisms. The organisms were travelling from the San Joaquin Valley to the Bay Area by "freak Southeastern winds," as the newspapers reported. The basic mechanism of disease was not violated, but this time the patients didn't have to travel to the Valley to come in contact with the disease. A human expert can understand this because he can fit the new situation to his model. Examples like these make us realize that Al systems like MYCIN can only perform some of the functions of an expert. Regarding modifiability, the process of implementing NEOMYCIN required many hours of THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 44 consultation with the original rule authors in order to unravel the rules. As shown In this paper, the lack of principles for using the representation makes It difficult to Interpret the purposes of clauses and rules. The strategy and overall design of the program have to be deduced by drawing diagrams like Figure 14. Imagine the difficulty any physician new to MYCIN would have modifying the CSF protein table (Figure 9); clearly he would first need the program to explain why it is correct. We also need a principled representation to avoid a problem we call concept broadening. When intermediate problem abstractions are omitted, use of goals becomes generalized and weakened. This happened In MYCIN as the meaning of "significance" grew to Include both "evidence of infection" and "non-contaminated cultures." (So rules were written In the form "If X then significant disease," rather than "if X then evidence of infection" and "if evidence of infection, then significant disease.") As long as the rule author makes an association between the data and some parameter he wants to influence, it doesn't matter for correct performance that the rule is vague But vague rules are difficult to understand and modify. A rule base is built upon and extended like any other program. Extensive documentation and a well-structured design are essential, as in any engineering endeavor. The framework of knowledge types and purposes that we have described would constitute a "typed" rule language that could make it easier for an expert to nrganize his thoughts. On the other hand, we must realize that this meta-level analysis may Impose an extra burden by turning the expert Into a taxonomist of his own knowledge--a task that may require considerable assistance. 45 7 Application of the Framework to Other Systems To Illustrate further the idea of the strategy, structure, support framework and to demonstrate its usefulness for explaining how a program reasons, several knowledge-based programs are described below in terms of the framework. For generality, we will call Inference associations like MYCIN's rules "knowledge sources" (KS). This analysis Is not concerned with the representational notation used In a program, whether It be frames, production rules, units, and so on. Instead we are trying to establish an understanding of the knowledge contained In the system: what kinds of Inferences are made at the KS level, how these KS's are structured explicitly In the system, and how this structure Is used by strategies for invoking KS's. THE EPISTEMOLOOY OF A RULEBASED EXPERT SYSTEM 46 META-LEVELS OF PROBLEM-SOLVING KNOWLEDGE SYSTEM STRATEGY J STRUCTURE KS examples J SUPPORT Domain = Chemistry, Mass spectrometry analysis DENDRAL Aggregation Family trees Molecular Buch- heuristics of functional identification chemistry anan, build superatms groups rules relating et al., & generate all (ketones, functional 1969 plausible ethers, etc.) groups to interstitial spectral peaks structures Domain= Speech Understanding HEARSAY It Policy KS's hierarchy of Example: KS's grammar & control hypoths interpretation hypothesizing identification Lesser, to generate & levels with words use properties of et al.. thresholds Jlinks to KS-s syllable level phonemes, 1975 (data-directed) for data syllables & words Domain = Concept formation, mathematical discovery AM Activity hierarchy of rules to create theory of Lenat heuristics heuristics concepts and interesting- 1976 propose tasks associated with fill-in facets, ness-chiefly (priority most general e.g., heuristics based on agenda & focus concept/context to fill-in the generalizing heuristic) to which they "examples" facet & specializing apply for ANY-SET Domain = Molecular Genetics, Experiment planning NOLGEN Determine hierarchy of specific lab Molecular Stefik differences, laboratory techniques: biology 1979 sketch.plan, operation types Input objects processes refine steps (used by -> molecular (message refinement changes & passing) design operator) byproducts Domain = Medical diagnosis, pulmonary function CENTAUR Hypothesis- I Hierarchy of disease I disease Aikins directed, top- disease component -> patterns; 1989 down refinement prototypes evidence for J biological (agenda) I prototype processes Domain = Medical diagnosis, diseases causing neurological symptoms NEO- MYCIN Grouping and Multiple data -> refining hierarchies evidence for same as abov Clancey hypothesis list of etiological disease process 1981 meta-rules processes or causall91 ocus £pursue) l state/category Figure 17. Relation of strategical, structural, and support meta-knowledge to knowledge sources of expert programs The Character of Structural Knowledge 47 7.1 The Character of Structural Knowledge One product of this study is a characterization of different ways of structuring KS'3 for different strategical purposes. In all cases, the effect of the structural knowledge Is to provide a handle for separating out what the KS Is from when It is to be applied s . The different ways of structuring KS's are summarized here according to the processing rationale: -- Organize KS's hierarchically by hypothesis for consistency in data-directed interpretation. In DENDRAL, if a functional group is ruled out, more specific members of the family are not considered during forward-directed, preliminary Interpretation of spectral peaks. Without this organization of KS's earlier versions of DENDRAL could generate a subgroup as a plausible interpretation, while ruling out a more general form of the subgroup, as if to say, "This is an ethyl ketone but not a ketone." [6] -- Organize KS's hierarchically by hypothesis to eliminate redundant effort In hypothesis-directed refinement. In DENDRAL, the family trees prevent the exhaustive structure generator from generating subgroups whose more general forms have been ruled out. The same principle is basic to most medical diagnosis systems that organize diagnoses In a taxonomy and use a top-down refinement strategy, such as CENTAUR and NEOMYCIN. -- Organize KS's by multiple hypothesis hierarchies for efficient grouping 8 In this section, the term "hypothesis" generally refers to a diagnostic or explanatory Interpretation made by a KS (In terms of some model), though It can also be a hypothesis that a particular problem feature is present, as In CRYSALIS. THE EPISTEMOLOGY Of A RULE-BASEO EXPERT SYSTEM 48 (hypothesis-space splitting). Besides using the hierarchy of generic disease processes (infectious, cancerous, toxic, traumatic, psychosomatic, etc.), NEOMYCIN groups the same diseases by multiple hierarchies according to disease process features (organ system involved, spread in the system, progression over time, etc.). When hypotheses are under consideration that do not fail Into one confirmed subtree of the primary etiological hierarchy, the "group and differentiate" strategy is invoked to find a process feature dimension along which two or more current hypotheses differ. A question will then be asked, or a hypothesis pursued, to differentiate among the hypotheses on this dimension. -- Organize KS's for each hypothesis on the basis of how KS data relates to the hypothesis, for focusing on problem features. In NEOMYCIN, additional relations make explicit special kinds of connections between data and hypotheses, such as "this problem feature Is the enabling causal step for this diagnostic process," and meta-rules order the selection of questions (invocation of KS's) by Indexing them indirectly through these relations ("if an enabling causal step is known for the hypothesis to be confirmed, try to confirm that problem feature"). The meta-rules that reference these different relations ("enabling step", "trigger", "most likely manifestation") are ordered arbitrarily. Meta-meta- rules don't order the meta-rules because we currently have no theoretical basis for relating the first order relations to one another. -- Organize KS's Into data/hypothesis levels for opportunistic triggering at multiple levels of Interpretation. HEARSAY's blackboard levels (sentence, word sequence, word, etc.) organize KS's by the level of analysis they use for data, each level supplying data for the hypothesis level above it. When new results are posted on a given level, KS's The Character of Structural Knowledge 49 that care about that level of analysis are polled to see If they should be given processing time. Policy KS's give coherence to this opportunistic Invocation by affecting which levels will be given preference. CRYSALIS takes the idea a step further by having multiple planes of blackboards; one abstracts problem features (the "density plane") and the other abstracts interpretations (the "model plane"). -- Organize KS's Into a task hierarchy for planning. In MOLGEN, laboratory operators are referenced indirectly through tasks that are steps in an abstract plan. For example, the planning level design decision to refine the abstract plan step, MERGE, is accomplished by indexing laboratory operators by the MERGE task (e.g., MERGE could be refined to using a ligase to connect DNA structures, mixing solutions, or causing a vector to be absorbed by an organism). Thus, tasks In planning are analogous to hypotheses in interpretation problems. -- Organize KS's into a context specialization hierarchy for determining task relevance. In AM, relevant heuristics for a task (typically filling In a mathematical concept slot) are Inherited from all contexts (concepts) that appear above It in the specialization hierarchy. Thus, AM goes a step beyond most other systems by showing that policy KS's must be selected on the basis of the kind uf problem being solved. Lenat's work suggests that this might be simply a hierarchical relationship among kinds of problems. The characterizations of structural knowledge above are a first step towards a vocabulary or language for talking about indirect reference of KS's. It Is clear that strategy and structure are Intimately related; to make this clearer, we return to our original Interest In explanation. rI THE EPISTEMOLOGY Or A RULE-BASED EXPERT SYSTEM 50 Teaching a strategy might boil down to saying, "Think In terms of such-and-such a structural vocabulary in order to get this strategical task done"--where the vocabulary Is the Indexing scheme for calling KS's to mind. So we might say, "Think in terms of families of functional subgroups In order to rule out interpretations of the spectral peaks." Or. "Consider process features when diseases of different etiologies are possible." That is, teaching a strategy involves in part the teaching of a perspective for relating KS's hierarchically (e.g., "families of functional subgroups" or "disease process features"), and then showing how these relations provide leverage for managing a large amount of data or number of hypotheses. The explanation of the leverage or structuring effect must be in terms of some task for carrying the problem forward, thus tying the indexing scheme to the overall process of what the problem solver Is trying to do (thus we say, "to rule out Interpretations" or "to narrow down the problem to one etiological process" or (recalling Figure 6) "to broaden the spectrum of possibilities"). In this way, we give the student a meta-rule that specifies what kind of vocabulary or "cut on the problem" to consider for a given strategical task. Davis's study of meta-rules [14] suggested a need for a vocabulary of meta-rule knowledge. His examples suggested just a few conceptual primitives for describing refinement (ordering and utility of KS's) and a few primitives for describing object-level knowledge (KS input and output). All of the strategies In our examples deal with ordering and utility criteria for KS's; so we have nothing to add there. All of the examples given here reference KS's by the data they act upon, the hypotheses they support or the tasks they accomplish, except for AM, which references KS's by their scope or domain of applicability. What is novel about the analysis here Is the focus on relations among hypotheses and among data. The Character of Structural Knowledge 51 From our domain-independent perspective, strategical knowledge selects KS's on the basis of the causal, subtype, process, or scoping relation they bear to hypotheses or data currently thought to be relevant to the problem at hand. Thus, our mete-rules make statements like: "Consider KS's that would demonstrate a prior cause for the best hypothesis." "Don't consider KS's that are subtypes of ruled-out hypotheses." "Consider KS's that abstract known data." "Consider KS's that distinguish between two competing kinds of processes." "Consider KS's relevant to the current problem domain." To summarize, the structural knowledge we have been studying consists of relations that hierarchically abstract data and hypotheses. These relations constitute the vocabulary by which domain-independent meta-rules invoke KS's. The key to our analysis Is our insistence on domain-independent statement of meta-rules--a motivation deriving from our Interest in explanation and teaching. 7.2 Explicitness of Strategical Knowledge Another consideration for explanation is whether or not the strategy for Invoking KS's is explicit or encoded Indirectly. As we proceed to higher strategical levels, It becomes difficult to represent a strategy declaratively. For example, top-down refinement Is "compiled into" CENTAUR's hierarchy itself by the control steps that specify on each level what to do next (e.g., "after confirming Obstructive Airways disease, determine the subtype of Obstructive Airways Disease"). By separating control steps from disease inferences, Alkins's Improved the explanation facility, one of the goals of CENTAUR. However, the rationale for these control steps is not represented--it is just as Implicit as it was in PUFF'S contextual clauses. In contrast, NEOMYCIN's "explore and refine" task clearly implements THE EPISTEMOLOGY Of A RULE-BASED EXPERT SYSTEM 52 top-down refinement through domain-independent meta-rules. However, these mete-rules are ordered to give preference to siblings before descendents--an example of an implicit strategy. One common way of selecting KS's Is on the basis of numerical measures of priority, utility, Interestingness, etc. For example, CENTAUR, like many medical programs, will first request the data that gives the most weight for the disease under consideration. Thus, the weight given to a KS Is another form of indexing by which a strategy can be applied. If we wish to explain these weights, we should ideally replace them by descriptors that "generate" them, and then have the strategy give preference to KS's having certain descriptors. NEOMYCIN's mete-rules for requesting data (described above) are a step In this direction. MOLGEN's "least-commitment" meta-strategy is a good example of implicit encoding by priority assignment. The ordering of tasks specified by least commitment Is: "look first for differences, then use them to sketch out an abstract plan, and finally refine that plan .... " This ordering of tasks is implicit in the numerical priorities that Stefik has assigned to design operators (e.g., propose-goal, refine-object, find-features). Therefore, an explanation system for MOLGEN could not explain the least-commitment strategy, but only say that the program performed one task before another because Its priority was higher. 7.3 Absence of Support Knowledge We have little to say about support knowledge In these systems because none of them represent it. That Is, the causal or mathematical models, statistical studies, or world knowledge that justifies the KS's Is not used during reasoning. As discussed In Section 8, Absence of Support Knowledge 83 this limitation calls Into question the problem-solving flexibility or "creativeness" of these programs. In any case, the knowledge is not available for explanation. 7.4 Summary The strategy/structure/support framework can be applied to any knowledge-based system by asking the questions: What are the KS's in the system (what kinds of recognition or construction operations are performed)? How are the KS's labeled or organized (by data/constraint or hypothesis/operation)? Is this Indexing used by the Interpreter or by explicit strategical KS's, or is It just an aid for the knowledge engineer? What theoretical considerations justify the KS's? Is this knowledge represented? With this kind of analysis, it should be clear how the knowledge represented needs to be augmented or decomposed, If an explanation facility is to be built for the system. Quite possibly, as In MYCIN, the representational notation will need to be modified as well. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 54 8 Conclusions The production rule formalism is often chosen by expert system designers because It Is thought to provide a perspicuous, modular representation. But we have discovered that there are points of flexibility in the representation that can be easily exploited to embed structural and strategic knowledge in task rules, context clauses and screening clauses. Arguing from a teacher's perspective, we showed that hierarchies of problem features and diagnoses (allowing rules to be generalized), in addition to domain-independent statement of strategy are useful to justify a rule and teach an approach for using It. Also, when the rule Is causal, satisfying explanations generalize the rule In terms of an underlying process model. This same knowledge should be made explicit for purposes of explanation after a consultation, ease of modification, and potential Improvement of problem solving ability. Characterizing knowledge In three categories, we concluded that MYCIN's rules were used like a programming language to embed strategic and structural principles. However, while context and screening clauses are devices that don't precisely capture the paths of expert reasoning, the basic connection between data and hypothesis Is a psychologically valid association. As such, the "core rules" represent the experts' knowledge of causal processes In proceduralized form (again, not necessarily compiled Into this form, but compiled with respect to causal models which may be incomplete or never even learned). For this reason, support knowledge needs to be represented In a form that Is somewhat redundant to the diagnostic associations, while structure and strategy can be directly factored out and represented declaratively. The lessons of this study apply to other knowledge-based programs, Including those which don't use the production rule representation. The first moral Is that one cannot simply 55 slap an Interactive front end onto a good Al program and expect to have an adequate teaching system. Similarly, an explanation system may have to do more than just read back reasoning steps and recognize questions: it may be useful to abstract the reasoning steps, relating them to domain models and problem-solving strategies. Other knowledge bases could be studied as artifacts to evaluate the expressiveness of their representation. Is the design of the inference structure explicit? Can It be reasoned about and used for explanation? You must ask: where are the choice points In the representation and what principles for their use have not been represented explicitly? For production systems one should ask: What Is the purpose of each clause in the rule and why are clauses ordered this way? Why Is this link between premise and conclusion justified? Under what circumstances does this association come to mind (structure and strategy)? Finally, future "knowledge engineering" efforts in which human experts are interviewed and their knowledge codified could benefit from first constructing an epistemology along the lines of the strategy-structure-KS-support distinction, and then, relative to that framework, representing knowledge using their chosen notation (rules, units, etc.). Then, when the system fails to behave properly (whether the purpose Is teaching or problem solving), changes to either the epistemology or the rules should be entertained. In fact, this is a cyclic process where changes are made to the rules that subtly tear at the framework, and after Incorporating a series of changes, a new, better epistemology and revised notation can be arrived at. (So a single MYCIN rule might seem awkward, but a pattern such as 40 rules with the same first 3 clauses suggests the underlying nature of the knowledge). Thus, a methodology for converging on an adequate epistemology comes In part from constant cycling and re-examinlng of the entire system of rules. THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 56 The epistemology that evolved from attempts to reconfigure MYCIN's rules Is NEOMYCIN's etiological taxonomy, multiple disease process hierarchies, data that trigger hypotheses, etc., plus the domain-independent task hierarchy of meta-rules. In our use of terms like "problem feature," we have moved very far from the too abstract "clinical parameter" which did not distinguish between data and hypotheses. Our epistemology provides an Improved basis for interpreting expert reasoning, a valuable foundation for knowledge engineering, as echoed by Swanson: Three aspects of the expert's adaptation are especially important to the design of decision support systems: the generative role of basic principles of pathophysiology, the hierarchical structure of disease knowledge, and the heuristics used in coping with information processing demands. [27] These categories of knowledge provide a framework for understanding an expert. We ask, "What kind of knowledge Is he describing?" This framework enables us to focus our questions so that we 'can separate out detailed descriptions of the expert's causal model from his associations that link symptom to disorder, and his strategies for using this knowledge. 9 Postscript: How the rule formalism helped Despite the now apparent shortcomings of MYCIN's rule formalism, we must remember that the program was influential because it worked well. The uniformity of representation, so much the cause of the "missing knowledge" and "disguised reasoning" described here, was perhaps an Important asset. With knowledge so easy to encode, It was perhaps the simple parameterlzation of the problem that made MYCIN such a success. The program could 57 be built and tested quickly at a time when little was known about the structure and methods of human reasoning. Again demonstrating the importance of the original research, we can treat the knowledge base as a reservoir of expertise, something never before captured In quite this way, and use It to suggest better representations. 10 Acknowledgments This paper is based on a chapter of my thesis. Early encouragement and comments were provided by Bruce Buchanan, John Seely Brown and Adele Goldberg. An early draft of this version was read and critiqued by Ted Shortliffe and Jim Bennett. The Al Journal reviewers were particularly helpful during the long gestation period of these Ideas. I have quoted a number of John Brown's suggestions directly. Without the careful, patient explanations of Tim Beckett, my meningitis tutor for a year. this paper would not have been possible. This research was supported In part by a grant from a joint ONR/ARPA contract (ONR 14-79C-0302). Computing resources provided by the SUMEX-AIM national resource (NIH grant RR 00786-07). THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM 58 References 1. Alkins, J. S. Prototypes and production rules: a knowledge representation for computer consultations. PhD thesis, Computer Science Department, Stanford University, (HPP-80-1 7, STAN-CS-80-814), August 1980. 2. Brachman, R. J. What's in a Concept: Structural Foundations for Semantic Networks. BBN Report No. 3433, October 1976. 3. Brown, J. S., Collins, A., & Harris, G. Artificial Intelligence and learning strategies. In O'Neill (Ed.), Learning Strategies. New York: Academic Press, 1977. 4. Buchanan, B. G., Sutherland, G., & Feigenbaum, E. A. HEURISTIC DENDRAL: a program for generating explanatory hypotheses in organic chemistry. In B. Meltzer and D. Michie (Eds.), Machine Intelligence 4. Edinburgh University Press, 1969, 209-254. 6. Buchanan, B. G., Sutherland, G., & Feigenbaum, E. A. Rediscovering some problems of artificial Intelligence In the context of organic chemistry. In B. Meltzer and D. Michie (Eds.), Machine Intelligence 5. Edinburgh University Press, 1970, 263-280. 6. Clancey, W. J. Tutoring rules for guiding a case method dialogue. International Journal of Man-Machine Studies, 1979, 11, 25-49. 7. Clancey, W. J. Transfer of Rule-Based Expertise through a Tutorial Dialogue. Computer Science Doctoral Disserta.tion, Stanford University, STAN-CS-769, August, 1979. 8. Clancey, W. J., Shortliffe, E. H., and Buchanan, B. G. Intelligent computer-aided instruction for medical diagnosis. Proceedings of the Third Annual Symposium on Computer Applications in Medical Care, Silver Spring, Maryland, October 1979. 9. Clancey, W. J. and Letsinger, R. NEOMYCIN: Reconfiguring a rule-based expert system for application to teaching. Seventh International Joint Conference on Artificial Intelligence, 1981, 829-836. 10. Collins, A. Fragments of a theory of human plausible reasoning. TINLAP-2, 1978, 194- 201. 11. Davis, R. Applications of meta-level knowledge to the construction, maintenance and use of large knowledge bases (STAN-CS-76-552, HPP-76-7). Stanford University, July 1976. 12. Davis, R. Generalized procedure calling and content-directed Invocation. Proceedings of the Symposium on Artif icial Intelligence and Programming Languages. SIGPLAN/SIGART Combined Newsletter, August 1977, 46-64. 13. Davis, R. Interactive transfer of expertise: acquisition of new Inference rules. Artificial Intelligence, 1979, 12, 121-167. 59 14. Davis, R. Meta-rules: reasoning about control. Artificial Intelligence, 1980, 15, 179- 222. 16. Davis, R., Buchanan, B., & Shortliffe, E. H. Production rules as a representation for a knowledge-base consultation program. Artificial Intelligence, 1977, 8, 16-46. 16. Davis, R., & King, J. J. An overview of production systems. In E. W. Elcock & D. Michle (Eds.), Machine intelligence 8. New York: Wylie & Sons, 1977,.300-332. 17. Elstein, A. S., Shulman, L. S., & Sprafka, S. A. Medical problem-solving: An analysis of clinical reasoning. Cambridge: Harvard University Press, 1978. 18. Engelmore, R. and Terry, A. Structure and function of the CRYSALIS system. Sixth International Joint Conference on Artificial Intelligence, 1979, 260-266. 19. Fagan, L. M., Kunz, J. C., Felgenbaum, E. A., & Osborn, J. J. Representation of dynamic clinical knowledge: measurement interpretation in the Intensive care unit. Sixth International Joint Conference on Artificial Intelligence, 1979, 260-262. 20. Lenat, D. B. AM: An artificial intelligence approach to discovery In mathematics as heuristic search. PhD thesis, Computer Science Department, Stanford University, (CS-STAN-76-670, AIM-286), July 1976. 21. Lesser, V. R., Fennell, R. D., Erman, L. D., Reddy, 0. R. Organization of the HEARSAY II speech understanding system. IEEE Transactions on Acoustics, Speech, and Signal Processing, ASSP-23, February 1975, 11-24. 22. McDermott, J. RI: A rule-based configurer of computer systems. Department of Computer Science, Carnegie-Mellon University, CMU-CS-80-1 19, April, 1980. 23. Mizoguchi, F., Maruyama, K., Yamada, T., Kitazawa, K., Saito, M., Kulikowski, C. A. A case study of EXPERT formalism--an approach to a design of medical consultation system through EXPERT formalism. Sixth International Joint Conference on Artificial Intelligence, 1979, 583-685. 24. Shortliffe, E. H. MYCIN: A rule-based computer program for advising physicians regarding antimicrobial therapy selection. Ph.D. dissertation in Medical Information Sciences, Stanford University, 1974. (Also, Computer-based medical consultations: MYCIN. New York: Elsevier, 1976.) 26. Shortliffe, E. H., Buchanan, B. G., and Feigenbaum, E. A. Knowledge engineering for medical decision making: a review of computer-based clinical decision aids. Proceedings of the IEEE, 1979, 67:1207-1224. 26. Stefik, M. J. Planning with constraints. PhD thesis, Computer Science Department, Stanford University, (HPP-80-2, STAN-CS-80-784), January 1980. 27. Swanson, D. B., Feltovich, P. J., & Johnson, P. E. Psychological analysis of physician expertise: implications for design of decision support systems. MEDINFOiF7, 1977, 161-164. k- s ]LZNG PAUR BLANK-NOT n J THE EPISTEMOLOGY OF A RULE-BASED EXPERT SYSTEM s0 28. Szolovits, P. , & Pauker, S. G. Categorical and probabilistic reasoning In medical diagnosis. Artificial Intelligence, 1978, 11(1), 115-144. 29. van Melle, W. A domain-independent production-rule system for consultation programs. Computer Science Doctoral Dissertation, Stanford University, In press, 1980. 30. Winograd, T. Frame Representations and the Declarative/Procedural Controversy. In D. G. Bobrow & A. Collins (Eds.), Representation and Understanding. New York: Academic Press, 1975, 185-210. 31. Winograd, T. Extended inference modes in reasoning by computer systems. Artificial Intelligence, 1980, 13, 6-26. 32. Woods, W. A. What's in a Link: Foundations for Semantic Networks. In D. G. Bobrow & A. Collins (Eds.). Representation and Understanding. New York: Academic Press, 1975, 35-82. 33. Yu, V. L., Buchanan, B. G., Shortliffe, E. H., Wraith, S. M., Davis, R., Scott, A. C., & Cohen, S. N. Evaluating the per'ormance of a computer-based consultant. Computer Programs in Biomedicine, 1979, 9, 95-102. (a) 34. Yu, V. L., Fagan, L. M., Wraith, S. M., Clancey, W. J., Scott, A. C., Hannigan, J. F., Blum, R. L., Buchanan, B. G. ahn,S, S.N. Antimcrobial selection by a computer -- a blinded evaluation by infectious disease experts. Journal of the American Medical Association, 1979, 242, 1279-1282. (b) STANFORD/CLANCEY December 30, 1981 Page 1 Navy Navy Dr. Ed Aiken 1 CAPT Richard L. Martin, USN Navy Personnel R&D Center Prospective Commanding Officer San Diego. CA 92152 USS Carl Vinson (CVN-70) Newport News Shipbuilding and Drydock Co Meryl S. Baker Newport News, VA 23607 NPRDC Code P3,09 1 Dr. James cBride San Diego, CA 92152 Navy Personnel R&D Center San Diego, CA 92152 Dr. Robert Ereaux Code N-711 Dr Wiiliam Montague I:AVTRAEQU IPCEN Navy Personnel R&D Center Orlando, FL 32813 Sar Diego, CA 92152 CDR Mike Curran Tet . . Yeller, Office of Naval Research Technical information Office, Code 2C1 FCC N. Quincy St. NAVY PEESON" EL F$&D CENTER Code 270 SAN !IEGO, CA 921F2 Prington, VA 22217 1 Library, Cc, P2CIL DR. PAT FEDERICC Navy Personnel &D Center NAVY PERSONNEL R&D CEUNTER an Diego, CA 92152 SA" DIECO, CP c2152 1 Technic rI Director Dr. John Ford Npvy Personnel r, D Center Navy Pprsonnel R&D Center San Dicgo, CI 92152 San, Diego, CA 92152 6 Commanding Officer LT Steven D. Harris, TISC, USN Naval PesearcY Laboretory Code 6021 Code 2(27 Nav . Air Devclopment Center VaE!"Lngton, r.C 22 e.. 1;prminster, Pennsylvania 197L 1 Psychologist Dr. Jim Hollpn ONR Ernch Office Code 3"C Elg 11L, Section D Navy Personnel P & D Center 666 Sure.-r Street. San Diego, Ct 92152 Foston, MA C22 r Dr. I'orman J. Kerr 1 Psychologist Chief of lnval Technical Training O1:R Frz ncl Offie N val Pir Ftt.tior Vemphis (75) 56 S. Cl:,rk Ftr,-ct , ' cngton, TN 3P054 Chicac, -L Dr. Willi m L. Xaloy 1 C)ffice of !'ava] Fe.(.arc:r Principal Civilian Advisor for Code N'.- Educrtion and Training PDC . uincy SFtreet --vtl Tr-ining Comr'nd, Code 0OA trlirgtor, V4 22217 Pensacol., FL 250 STA ',F 0RD / CLA N'C FY Dt'ccrnber 3C, 1CS1 Page 2 r; Personnel ?. Tr,-ining Research Programs 1 Dr. Robert G. Smith (Code 4~5P) ODffice of Chief of Naz Oper~tic Offirnc of Navi Research OP-( 0 PW! Arlington. VA 22217 Washington, DC ?C31; Psycholog-ist 1 Dr. Plfred F. Sriode 2,,I Frz-nch Cffice Training AnalIysis P.EVF.ILI?tLor ' lr'.e East Grccn Street (7AECQ) Pasadena, CA 011C1 Dept. of the I!avy Orlando, FL -2R I Spec-la1. Asst. for Fducation and Trzirdng (OP-01E) 1 Dr. Fichprd Sorensen Rni. ?"Cr Pr'hingtor Annex Navy Personnel R&2 Center LVss!%Ington, DC 7 7 r Stnn Diego, Cf. p2152 i Iffice of the Chief of !.aval r0 per~tiens I Roger Wpissirrer-P-ylor Pese, rch Tevplopment t C;tudies Erarch Department of f. miristrrtfve Fc.-' (C-P-1 lE.) ha;val! Pcstgrciu.,te Sch-oo2 ~~V~.~o.PC 20er,( !-Ioterey, CAI Cq% 4P I LT Fr.-nk C. Patlio, !"'C, US" (Ph.D) 1 Dr. Pobcrt Uis*,r 5>2 octicr rd Trining lcs-carch Division Code c I Huin.n Pe'-fcrntnence s Dept. !:z3%y Pcrsonne2 F',D Center !mrl rec 1,ec ic~l Fes, F-rch Labor: t S, .n Dicgo, CP 2~ Penc'clF'L 2- 1 M'r John E. IMo Cc I Dr. \(Thry Poch-! Code P*:Ir -~rtin PsrhDprtmen U. S. !avy Pers onnP2 ser' Code 5 ppK Devclopmnr C-nte-r 'Sa'.'] Post~rrmdur-t~c 5chcol nPicCE '% 1.,nt'.r'y, C,. c 0.4C Pagrer W. nrem'ingtcn. Ph.P CodE Lr,3 Pensrano' r, FL ?2 ,OF Dr. F~rnlrd Firnrd (C'.D) M'nvy Personr<l P' P Center 7',P c CA 21r 'r, Unrt Sc"fn irrctor Fi~~.rP-, vvlcp.ent, Tcvt & Ev .upat inn !, r !:Pv; Ed~uc;aion ;nd Tr,-ining Cornanci ?NAF, P-rs;nscol; , FL -a25OF STANFORD/CLANCEY December 30O, 1981 Page~ A rry Army 1 Techriccl Director 1 Dr. Frederick ' Steinh-iser U. S. trmy Research Institute for the Dept. of Navy Behavior .I Fnd Social Sciences Chief of Naval Operatior!7s r01 Eisenhower Avenue QP-113 Alexandria, VA 223' Washinstor, DC 2C750 1 Mr. James Pzker 1 ZDr. JcsepK Ward Systemns Mzbnning Technical Area U.S. Army Fese~rch irnstitute Army Reseprch institute 5 71 Eisenh;owr A~venue 5001 Eisenhower Pve. Alcxzr.riiv, VA 22---- L lexandriz., VA 2 23 1 Dr. Fe?tr-ice J. Farr U. S. Army Feseprch institute 5rrl Ei--nhower Avenup Ai.Cxarcrin, VA 22"'7 1 17 FPt'V J. HPP7RS U.S. AP'Y RESEt PCH i!STIT-UTE 5 21 ETSENHPWEP AVENUE ALEXANDTRTA, VA 22?-33. Dr . 'i c e 1 K :pl.n U.S. IF'Y RESEA- PCH T NSTTMUTE 5" F7PE:HOWEP A.VENUE ALEYA!,DFT;., V.- 22.? Dr. -I'1 on F. Katz Trz&inir17 Ychnic~zl Area U.F. Prmy Reserrch Tn titute C Riscrnhower Avcnue A-ex~rr!riz., VA 22: '7 1 Dr. VWroid F. C'Neil, Jr. tttn: PETIT-OK Army Peseirch -nst itut e 5031l Fis'-nhowpr Avenue Alexandri _, VA P2T33 Dr. Robert 5Y-sror P'. F. irmy F'%-,rOh Tnztitutc for '!hc Pnh.vior;.! -.d Sori'.1 Fcicrw'-s 5'C Eifcrj1howr t.vcnufe ."'cxandria, VA?23 qT! NFOPD/CLAN*CEY December '-C, 19 1 Page J Air Force Marines U.S. Air Force Office of Scientific 1 H. William Greenup Research Education Advisor (E031) Life Sciences Directorate, NL Education Center, V'CDEC Bolling Air Force Base Quantico, VA 22134 Washington, DC 2032 1 Special Assistant for Marine Dr. Fvrl A. Alluisi Corps Matters HO, AFHRL (AFSC) Code 10OM Erooks fFB, TX 75235 Office of Naval Research 800 N. Quincy St. Dr. Genevieve Haddad Arlington, VA 22217 Prograr !!anager Life Sciences Pirectorpte 1 DR. A.L. SLAFKOSKY AFCSR SCNENTIFIC ADVISOP (CODE RD-i) Bolling AFP, DC 20332 !1Q, U.S. IPM:NE CORPS WtSHIl'GTON, DC 207PO 2 700 TCHP2,,/TTGH Stop 32 Sheppard .Fr, TX 76311 m TANFOED/CLANCEY December 30, 1981 Page 5 CoastGuard Other DoD Chief, Psychological Reserch Branch 12 Defense Technicel Information Center U. S. Coast Guard (G-P-1/2/TP42) Cameron Station, Bldg 5 Washington, DC 20593 Alexandria, VA 22314 Attn: TC 1 Military Assistant for Training and Personnel Technology Office of the Under Secretary of Defense for Research & Engineering Room 3D129, The Pentagon Washfhgton, DC 20301 1 DARPA 1OO Wilson [Ilvd. Arlington, VA 22209 STA:FOPr/CLA ICEY December 10, 1981 Page . Civil Govt Non Govt Dr. Susan Chipman 1 Dr. John R. Anderson Learning and Development Department of Psychology National Institute of Education Carnegie 1'ellon University 1200 1th Street NW Pittsburgh, PA 15213 Washington, DC 20208 1 Anderson, Thomas H., Ph.D. Dr. John Mays Center for the Study of Peading National Institute of Education 174 Children's Research Center 1200 19th Street N 51 Gerty Drive Washington, DC 2020? Champizgn, IL 61820 lillirn J. f'cLp.urin 1 Dr. John Annett ~6F10 Howic Court Department cf Psychology Camp Springs, I'D 20031 University of Warwick Coventry CV4 7ALDr. Ar4 hur "elmed EIGLA HD Nationsl intitute of Fducatjon 120P lCth Street KW 1 psychologicel research unit V.shington, DC 20_20 Dept. of Deffnse (Army Office) Campbe . Park Offices Dr. Andrei: R. Volnpr Canberra ACT 2600, Australin Science Education Dav. cnd Pesezrch 1 Dr. Plan Eaddeley zticnz.' Sci ence Foundation edical Peserrch Council !P'shington, DC 20rjcn . Applied Psychology Unit 15 Chauccr Roid Dr. Joseph Psotkai Cambridge CE2 2EF Nvtion&A] nstitute of Education ENGLAND 1200 10t! St. M Washington,DC 2020 1 Dr. Patricia Paggett Department of Psychology Dr. Frrank Uithrow University of Colorado U. I!. Office of Education Eoulde-r, CC ?03 '9 tn(, ?'=ryland Ave. SW Wcshi gton, DC 20212 1 Mr Pvron Parr Department of Computer Science Dr. Joseph L. Young, Pirector St.nford University Memory & Cognitive Processes Stanford, CA 9)1305 National Scipnee Foundtion I:hir ton, DC 2055(' 1 Liaison Sciert.ists Cff (ee of Iav'l Peseirc., Prrnc.h Office , London Pox .c FPO rcw YorlN 0?9I0 1 Dr. Lyle Pourne rp.rt.-.nt of Psychology Univerrity of Colorado rouldcr, CO S0'0c STANFCRD/CLANCEY December 30, 1981 Page 7 Non Govt Non Govt 1 Dr. John S. Brown 1 Dr. Meredith P. Crawford XEROX Palo Alto Research Center American Psychological Association 3333 Coyote Road 1200 17th Street, N.W. Palo Alto, CA 94304 Washington, DC 20036 Dr. Bruce Buchanan 1 Dr. Kenneth B. Cross Department of Computer Science Anpcapa Sciences, Inc. Stanford University P.O. Drawer Q Stanford, CA 9J305 Santa Barbara, CA 93102 DR. C. VICTOR BUNDERSON 1 LCOL U. C. Eggenberger WICAT INC. DIRECTCRATE OF PERSONNEL APPLIED RESEARC UNIVERSTTY PLAZA, SUITE 10 NATIONAL DEFENCE HQ 1160 SO. STATE ST. 101 COLONEL BY DPIVE OREM, UT 8L057 OTTAWA, CANADA KIA OK2 Dr. Pat Carpenter 1 Dr. Ed Feigenbp.um Department of Psychology Department of Computer Science Carnegie-Mellon University Stanford University Pittsburgh, PA 15213 Stanford, CA 94305 Dr. John B. Carroll 1 Dr. Richard L. Ferguson Psychometric Lab The Arericn College Testing Program Univ. of No. Carolina P.C. Box 168 Davie H311 017A Towz City, TA 52240 Chapel Hill, NC 27514 1 Mr. Wallace Feurzeig Dr. William Chase Bolt B5rmnek & Netnen, Inc. Department of Psychology 50 Voulton St. Carnegie Vellon University Cambridge, M1A C213 Pittsburgh, PA 15213 1 Dr. Victor Fields Dr. Micheline Chi Dept. of Psychology Learning R & D Center l'ontgomery College University of Pittsburgh Pockville, MD 2050 393 O'Harr Ftreet Pittsburgh, PA 15213 1 Univ. Prof. Dr. Ger1-ard Fischer Liebiggasse 5/? Dr. Allan M. Collins A 1010 Vienn,' Polt Fer-.nek & ?,ewman, Inc. fAISTRIA 5r 11ou]ton Street Crmbridge, Na C2138 1 DR. JOHN D. FCLLEY JR. APPliED SCTENCE AFSOCI!TES 7NC Dr. Lynn A. Cooper VALENCIA, PA 16059 LRDC University of Pittsburgh 1 Dr. John R. Frederiksen 3923- O'Harr Street Bolt Eernnek & Newman Pittsburgh, PA 15213 50 !oulton Street Cambridge, MA C21-r STAMFORD/CLANCEY December 30, 199I Page 8 Non Govt Non Govt Dr. Plinda Friedmpn 1 Dr. Frederick Hayes-Roth Department of Psychology The Rand Corporation University of Alberta 1700 Main Street Edmonton, Alberta Santa Monica, CA 90406 CANADA T6G 2E9 I Dr. James R. Hoffman Dr. P. Edward Geiselman Department of Psychology Department of Psychology University of Delaware University of California Newark, DE 19711 Los Angeles, CA 90024 1 Dr. Kristina Hooper DR. RCBERT GLASER Clark Kerr Hall LRDC University of California UNIVERSITY OF PITTSBURGH Santa Cruz, CA 95060 10o C'HARA STREET PITTSBURGH, PA 1521 - 1 Glenda Greenwald, Ed. "Human Intelligence Newsletter" Dr. 11arvin D. Clock P. 0. Box 1163 217 Stone Hal Birmingham 11I 421012 Cornell University Ithaca, NY 1465'? 1 Dr. Earl Hunt Dept. of Psychology Dr. Daniel Gopher University of Washington Industripl & frenagement Engineering Seattle, W. 9,c15 Technion-Israf- Institute of Technology Haifa 1 Dr. Ed Hutchins ISRAEL Navy Personnel R&D Center San Diego, CA 92152 DR. JAMES G. GREENO LRDC 1 DR. KAY INABA UNIVFRSITY OF PITTSBURGH 21116 VA!:O'.EN ST 3.o9 O'HAFA STREET CANOGA P.RK, CA 91 '03 PITTSBURGH, PA 15213 1 Dr. Stevpn W. Keele Dr. Ron Hambleton Dept. of Psycho2ogy School of Education University of Cregon University of fassechusetts Eugene, OR 97140? Amherst, lIA 01002 1 Dr. Walter Kintsch Dr. Harold HPwkins Department, of Psychology Depirtment of Psychology University of Colorudo University of Orcgon Boulder, CC P0C02 Eugene OB 9714C7 1 Dr. David Kieras Dr. Bprbara Hpyes-Roth Department of Psychology The Rand Corporption University of Arizone 17CC Mrin Street Tuscon, AZ e5721 Santa l'onicr, CA 90406 STANFORD/CLANCEY December 30, 1981 Page 9 Non Govt Non Govt Dr. Stephen Kosslyn 1 Committee on Human Factors Harvard University JH 811 Department of Psychology 21C1 Constitution Ave. NW 33 Kirkland Street Washington, DC 20418 Cambridge, MA 02138 1 Dr. Jesse Orlansky Dr. Marcy Lansman Institute for Defense Analyses Department of Psychology, NI 25 400 Army Navy Drive University of Washington Arlington, VA 22202 Seattle, WA 98195 1 Dr. Seymour A. Papert Dr. Jill Larkin Massachusetts Institute of Technology Department of Psychology Artificip! Intelligence Lab Carnegie Vellon University 514r Technology Square Pittsburgh, PA 15213 Cambridge, MA C2139 Dr. Alan Lesgcld I Dr. James A. Poulson Learning R&D Center Portland State University University of Pittsburgh P.O. Box 751 Pittsburgh, PA 152"0 Port!and, OR c7207 Dr. Michael Levine 1 Dr. James L'. Pellegrino Department of Educational Psychology University of Cclifornia, 210 Education Rldg. Santp Parhara University of Illinois Dept. of Psychology ChampaiLr, 1L 6!C01 SantF Earsbarz, CA 93106 Dr. Erik McWilli-ris 1 MR. LUIGI PETRULLC Science Education Dev. and Pesearch 2431 N. ErSEWClD STFrFT National Science Foundation ARLIGTON, VA 222C7 Washington, DC 20550 1 Dr. VFrtha Poison Dr. Mark Miller Department of Psynhology TI Computer Science Lab Campus Box 746 f/O 2?24 Winterplace Circle University of Colorado Pl;:no, TY 75'75 Boulder, CO ^0709 Dr. Allen Vunro 1 DR. PETF POLSON Behavioral Technology Laboratories DEPT. OF PSYCHnLOGY 1?145 El!na Ave., Fourth Floor UN.VERC'TTY CF CCLOPADO Fedondo Pepch, Ct 90277 POULEP, CO FO] c Dr. Don;1d A iorman 1 Dr. Steven E. Poltrock herr. of Psychology C-009 Department of Psychology Univ. of Callfornia, ran Diepo Univ(rsity of Denver La Jolla, CA 92 c Denver,CO a0?C' FTANFOFD/CLANCEY December 30, 1981 Page iC ;"on Govt. Non Govt 1 MNRAT M. L. RAUCH 1 DR. ROBERT J. SEIDEL P II 4 INSTRUCTIONAL TECHNOLOGY GROUP BUNDESMINISTERIUM DER VERTEIDIGUNG HUMRPRO POSTFACH 1328 30C N. WASHINGTON ST. D-53 BONN 1, GERMANY ALEXANDRIA, VA 22314 Dr. Fred Reif 1 Committee on Cognitive Research SESPME % Dr. Lonnie P. Sherrod c/o Physics Department Social Science Pesearch Council University of California 605 Third Avenue Ferkely, CA 94720 New York, NY 10016 Dr. Lauren Resnick 1 Robert S. Siegler LRDC Pssocipte Professor University of Pittsburgh Carncgie-Mellon University z2c9 O'Hara Street Department of Psychology Pittsburg., PA 15213 Schenley Park Pittsburgh, PA 15213 1 ary Riley LFDC 1 Dr. Edwird E. Smith University of Pittsburgh Polt Pernnek & 1,.e;,inan, Inc. '10c O'Hsr& Street 50 ;!oulton Street Pittsburgh, PA 1511? Cambridge, .,A 0213P 'Dr. Andrew M.. Pose 1 Dr. Pobert Smith Am,,eric.n Institutes for Research Department of Computer Science 1P55 Thorias Jefferson St. N'J Rutgers University .:shington, DC 200r7 Hew Brunswick, J 00,03 Dr. Ernst Z. Rothkopf 1 Dr. Fichard Snow Pel! Labcrctoripr School of Educaticn 60. Mountain Avenue Stanford University Murray Hill, NJ 07974 Stanford, CA 9430r Dr. David Rumelhart 1 Dr. Rob~rt Sternberg Cpnter for Human Information Processing Dept. of Psychology Univ. of California, San Diego Ynle University L= loll&, CA 92093 Box 11A, Yale Sta-tion New Haven, CT C652C DF. :UALTER FCI'EIDER DEPT. CF PSYCNCLOGY 1 DR. ALRERT FTEVENS U,'TV7-RSITY OF ILLINOIS POLT EFRAVEK & NEI-AN,11, 1"!C. CHAPATG!;, 1L 61P20 5C I'CULTC0; STREET CA?!PIDGE, Mk C217. Dr. Al-r .echoenfeld Department of Vpthematics 1 David E. Stone, Ph.D. Hrmilton College Hazeltine Corporptior Clintor, NY 1--2: " ,C Ol.d pringhouse Road 1'cLean, VA 22202 Page 11 ZTANFORD/CLANCEY December 30, 19Va Non Govt Non Govt DR. PATRICK SUPPES DR. GERSHON WELTMAN INSTITUTE FOR MATHEMATICAL STUDIES IN PERCEPTRONICS INC. THE SOCIAL SCIENCES 6271 VARIEL AVE.HE SILSCNSwOODLkND HILLS,.A916 STANFORD UNIVERSITY STANFORD, CA 94305 1 Dr. Keith T. Wescourt Information Sciences Dept. 1 Dr. Yikumi Tatsuoka The Rand Corporation Computer Based Education Research 1700 Main Spt Laboratory 0y Santa Snc, CA 9006 252 Engineering Research Laboratory S ,A University of Illinois Urbana, TL 61PO1 Dr. John ThomasIBM Thomas j. W4atson ReseparcCne P.O. Box 215 Yorktown Heights, NY 10598 DR. PERRY THORNDYKE THE RAND CORPORATION 1700 MATV] STREET S411TA V'ONTC, CA 90406 Dr. Douglas Towne Univ. of So. Celifornia FBehavioral Technology Labs 1e45 S. Elene Ave. Redondo Peach, CA 90277 Dr. J. Uhlener Perceptronics, Inc. 6271 Voriel Avenue Woodlpnd Hills, CA 91364' Dr. BentOn J. Underwood Dept. of PsychOlogy Northwestern University Evanston. IL FC201 Dr. :illard S. Vaughan, Jr. OceFlnutics, Inc. L22 Sixth street AnnFpolis, !.,D 21'3 Dr. David J. VJeiss N66C Elliott Hall University of Minnesota 7r, F. River Road inn~apo.i~s, F'N 55455 &= . a II 
work_n25ntqmt35gq5ho2os4zbwvify	----	Expert Systems With Applications 105 (2018) 23–33 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Robust Semi-Supervised Growing Self-Organizing Map Ali Mehrizi, Hadi Sadoghi Yazdi ∗, Amir Hossein Taherinia Department of Computer Engineering, Ferdowsi University of Mashhad, Mashhad, Iran a r t i c l e i n f o Article history: Received 14 October 2017 Revised 22 March 2018 Accepted 23 March 2018 Available online 26 March 2018 Keywords: Semi-supervised learning Online learning Dynamic self-organization network Adaptive learning Half quadratic a b s t r a c t Semi-Supervised Growing Self Organizing Map (SSGSOM) is one of the best methods for online classifica- tion with partial labeled data. Many parameters can affect the performance of this method. The structure of GSOM network, activation degree and learning approach are the most important factors in SSGSOM. In this paper, a comprehensive robust mathematical formulation of the problem is proposed and then half quadratic (HQ) is used to solve it. Furthermore, an adaptive method is proposed to adjust activation de- gree optimally to improve the performance of SSGSOM. The results are reported on a variety of synthetic and UCI datasets and in the noisy conditions, which show superiority and robustness of the proposed method compared with the state of the art approaches. © 2018 Elsevier Ltd. All rights reserved. 1 p b a u f l t t l ( p Y & d b o i l t c t l Y i b o t i w n n t p m fl a o t s o d w f t b ( t h 0 . Introduction Extracting knowledge from unlabeled data, also known as unsu- ervised learning, improves the performance of systems. However, y increasing the volume of imported data, they require a large mount of memory and computation time and therefore become nusable or very inefficient. If a small portion of data has side in- ormation such as labels, it can increase the accuracy and speed of earning. The use of labeled data for learning together with clus- ering methods is called semi-supervised learning. In other words, he learning algorithms that combine unsupervised and supervised earning algorithms, are called semi-supervised learning algorithms Zhu & Goldberg, 2009 ). Semi-supervised learning has been ap- lied for many problems such as those reported in ( Cong, Liu, uan, & Luo, 2013; Lu, Wang, Xue, & Pan, 2014; Chen, Li, Su, Cao, Ji, 2014 ). Considering the data importing method, learning algorithms are ivided into two categories: online learning and batch learning. In atch learning, all the learning data are available at the beginning f the process. In contrast, in online learning, none of the data ex- sts at the beginning and data are logged gradually. Although, on- ine learning is adapted to the varying environment better, it lacks he accuracy of batch learning. One of the major reasons for de- reasing accuracy of the online algorithms is their dependency on he sequence of data entry. In recent years, attention to online earning is growing, because by increasing volume of data, learn- ∗ Corresponding author. E-mail addresses: mehrizi@um.ac.ir (A. Mehrizi), h-sadoghi@um.ac.ir (H. Sadoghi azdi), taherinia@um.ac.ir (A.H. Taherinia). s a D i ttps://doi.org/10.1016/j.eswa.2018.03.046 957-4174/© 2018 Elsevier Ltd. All rights reserved. ng methods must have online capabilities and rapid responses ecomes inevitable. Accordingly, many researchers have focused n eliminating defects of online learning methods to achieve bet- er accuracy in a reasonable time. Some promising online cluster- ng methods are online k-mean ( Alpaydin, 2004 ), neural gas net- ork ( Martinetz, Berkovich, & Schulten, 1993 ), and self-organizing etwork ( Kohonen, 1990 ). Among these methods, self-organization etworks and neural gas networks have been more useful because hey are inherently online and have properties such as topology reservation ( Alahakoon, Halgamuge, & Srinivasan, 20 0 0 ). Since the ajority of studies on semi-supervised learning have been on of- ine method, in the following we first review the learning methods nd then discuss studies on online semi-supervised learning. Farid et al. (2013) presented a semi-supervised algorithm based n the ensemble learning method, which is an offline approach hat determines the label of the data via voting among several clas- ifications. They used decision trees for data classification. Then, an nline algorithm was proposed by modifying the structure of the ecision tree. For this purpose, in the learning phase, a random eight is allocated to every datum and the decision tree is built or them. Afterward, the allocated weights are modified based on he training data. Leite and colleagues reported that features and main distri- ution of data change by time in non-stationary environments Leite, Costa, & Gomide, 2010 ). In machine learning terminology, his change is called as a “concept drift”. For example, in clas- ification, a concept can be either a class of data or the bound- ry of data and the drift is data boundary that changes over time. rifts can be gradual or sudden, contracting or expanding, and def- nite/random or periodic change of concept. The main requirement 24 A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 t a t m l m ( s g t u t K t l a o c s 2 m d c A fi a ( p G A c 2 i t c & t t f fi t t o S of online learning is to identify and deal with the drift. They pro- posed a semi-supervised method based on a granular neural net- work to deal with data streams. This model has five layers; the first and fifth layers of the neural network are the input and output layer, respectively. The second layer performs the data-clustering task. The third layer performs the merge operation and the fourth layer makes the decisions about data class. Macario and de Carvalho (2012) proposed a semi-supervised FCM algorithm that adds a supervised ability to the original FCM algorithm. The main idea of this paper is that data of the same class have similar degrees of membership. Therefore, if data be- longs to a certain class, differences between their cluster member- ship degrees are low and the data have stable degrees of mem- bership in all clusters. They also proposed an adaptive method to prevent uncontrolled growth of a cluster. For the first time, an online constrained clustering algorithm was presented by Halkidi, Spiliopoulou, and Pavlou (2012) in order to find clusters that are centralized and have constraints entering as a stream. The Halkidi algorithm was based on the MPCK 1 -Means clustering algorithm. The MPCK-Means algorithm was an offline algorithm and cannot be used for online data. Therefore, Halkidi considered data as a chunk that enters the algorithm over time. The MPCK-Means clustering algorithm performs offline on small chunk of data. This method can be used to reduce the computa- tional memory. Cong et al. (2013) introduced a self-supervised online met- ric learning with the low-rank constraint for scene categorization. Their work used a two-sided graph to evaluate the similarity be- tween data and metric learning. The labeled and unlabeled data were used to update metric learning by a high score. The proposed algorithm of Kung was an online algorithm. Li and colleagues introduced a semi-supervised learning algo- rithm based on the extreme learning combined with cooperative learning ( Li, Zhang, Xu, Luo, & Li, 2013 ). The main idea of cooper- ative learning is using data from a learner’s testing phase as train- ing data for other learners. Looping in the learning process is the disadvantage of cooperative learning, which leads to a decrease in learning speed and propagation of the learning process errors to others. In their paper, using extreme learning was proposed to deal with these problems because it is very fast and can cover the prob- lem of cooperative learning ( Huang, Zhu, & Siew, 2004 ) ( Huang, Li, Chen, & Siew, 2008 ). Zhang and colleagues proposed a semi-supervised algorithm based on co-training ( Zhang, Wen, Wang, & Jiang, 2014 ). This al- gorithm applies co-training to select the most reliable instances according to two criteria of high confidence and nearest neighbor for boosting the classifier and also exploits the most informative instances with human annotation for improving the classification performance ( Zhang et al., 2014 ). Dornaika and El Traboulsi (2016) proposed a Graph-Based Semi- Supervised method and its kernelized version for generic classifi- cation. Although this method is flexible, it is not online. Beyer and Cimiano (2012) proposed a semi-supervised method based on neural gas. This paper extends the offline Growing Neu- ral Gas classifier to an online method that predicts labels for un- labeled example and incorporates these labeled examples into the model on-the-fly ( Beyer & Cimiano, 2012 ). Maximo, Quiles, & Nascimento (2014) proposed a new semi- supervised growing neural gas (GNG) model, wherein instead of assigning a single scalar label value to each neuron, a vector con- taining the representativeness level of every class is associated with each neuron. Also, to propagate the labels among the neurons 1 Metric pairwise constrained K-means. e S d f he Consensus-Based Semi-Supervised GNG employs a consensus pproach ( Maximo et al., 2014 ). Fritzke (1994) proposed growing self-organizing map (GSOM) hat is able to find a suitable network structure and size of the odel automatically. Next, this researcher proposed a supervised earning method that results from the combination of the above- entioned self-organizing network with the radial basis function RBF) approach. Hsu and Halgamuge (2008) improved Fritzke method and pre- ented a semi supervised algorithm based on GSOM. In their al- orithm, a two-layer model for online data classification with par- ial labels was provided. In the first layer of this model, GSOM is sed for data clustering and when the labeled data were logged, he label of this data is used for classification. In this method, like -mean and fuzzy C-means, there are repetitive operations with he difference, where in each step only one datum is randomly se- ected and entered into the GSOM. Shen, Yu, Kamiya, and Hasegawa (2010) and colleagues offered three-layer architecture that represents the topological structure f training data (with SOM), learned node labeling, and classifying onstruction. In our previous work, we proposed an online constraint semi- upervised learning based on GSOM ( Allahyar, Yazdi, & Harati, 015 ), based on a two-layer model that came from the develop- ent of the HSU model ( Hsu & Halgamuge, 2008 ). Self-Organizing Maps (SOM) is a suitable learning method for ata clustering ( Kohonen, 1990 ). This method is based on artifi- ial neural networks (ANNs) and its structure is formed online. ccordingly, the network is suitable for online learning. SOM was rst proposed in 1990 by Kohonen (1990) . Dimensions of SOM are parameter that strongly influences the final result of clustering Alahakoon et al., 20 0 0 ). In the Kohonen’s model, the user sets this arameter manually. To resolve this problem, ALkahon presented rowing Self-organizing Map (GSOM) ( Alahakoon et al., 20 0 0 ). In Lkahon models, expansion dimensions are performed automati- ally depending on type and structure of data ( Alahakoon et al., 0 0 0 ). Hsu and Halgamuge (2008) presented Semi-supervised Grow- ng Self-Organizing Map (SSGSOM) by adding a supervised layer o GSOM. Hsu model is an online semi-supervised model that an perform clustering and classification of data very well ( Hsu Halgamuge, 2008 ). Their model uses labeled data to control he Expansion nodes in the SOM. This model employs an innova- ive method to determine the effective parameters in the objective unction ( Hsu & Halgamuge, 2008 ). In the present paper, we rede- ne the objective function and then discuss the determination of he optimum effective parameters. Table 1 summarizes the significant semi-supervised methods. In his table, type of algorithm, advantages, and unresolved problems f each algorithm are presented. The main contributions of the proposed method are: • Redefining robust cost function of online Semi-Supervised GSOM by half quadratic, • An adaptive online semi-supervised GSOM algorithm is pro- posed. The remainder of this paper is organized as follows: In ection 2 , the primary concepts for the proposed algorithm are xplained. In Section 3 , the proposed algorithm is presented. In ection 4 , the results of the performed experiments on different ata are shown. In the final section, the conclusion and guidelines or future works are expressed. A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 25 Table 1 Summary of Semi-supervised Algorithms. Proposed algorithm Advantage Unsolved problems Farid et al. √ Semi Supervised learning ❖ Slow execution speed because of creating the decision tree Leite et al. √ Fuzzy ❖ Slow execution speed √ Resistant data types ❖ Not improve clustering by using the feedback of labeled data Macario et al. √ Fuzzy clustering ❖ Offline method √ Adaptive distance Shen et al. √ online ❖ Not adaptive √ Based on SOM ❖ Not robust Halkidi et al. √ Data binding ❖ Receiving data in chunk format Cong et al. √ Metric learning ❖ Receiving in chunk format ❖ Slow execution speed Li et al. √ Cooperative extreme learning ❖ Not suitable for online data ❖ Slow execution speed because of cooperative learning Zhang et al. √ Co-training ❖ Offline ❖ Slow (because used active learning to determine appropriate data for labeling) Dornaika et al. √ Flexible method ❖ Offline √ Graph-based ❖ Adjust many parameters Beyer et al. √ Based on neural gas ❖ Not adaptive √ Online ❖ Not robust Maximo et al. √ Based on neural gas ❖ Not adaptive √ Online ❖ Not robust Allahyar et al. √ Online ❖ Not adaptive √ Based on GSOM ❖ Not robust Hsu et al. √ online ❖ Determine the parameters of the GSOM innovatively √ Based on GSOM √ Improve clustering by using feedback of labeled data ❖ Not robust Proposed method of the paper √ Online √ Based on GSOM √ Adaptive √ Robust Fig. 1. Semi-supervised GSOM Structure ( Hsu & Halgamuge, 2008 ). 2 G s s n t i n c t w t t t d m � t d t ζ P t d b m w s b δ . Preliminary concept In this section, the preliminary concepts of semi-supervised SOM are expressed. GSOM is a method that is able to show the tructure of data ( Fritzke, 1994 ). Hsu and colleagues added the emi-supervised ability by adding a supervised layer on top of this etwork ( Hsu & Halgamuge, 2008 ). Fig. 1 shows the structure of his model. In the presented semi-supervised model, the number of nodes n the upper layer is equal to the number of classes (M) and the umber of nodes in the bottom layer represents the number of lusters (C). There is a full connection of weights between the wo layers that is indicated as w c,m . Its purpose is to train the eights in such a way that by entering data into the cluster layer, he nearest node is activated and then the suitable class is de- ermined for it. Based on the weights between the two layers, raining of weights between the two layers is done by the labeled ata and according to Eq. (1) for each data sample ( Hsu & Halga- uge, 2008 ). w c,m k = β × ( ζ m k − P S m k ) × o c k (1) In Eq. (1) , ζ m k is actual label for datum k that is in class m and he P S m k is the estimated value of the semi-supervised GSOM for atum k that is in class m ; where ζ m k and P S m k are defined respec- ively by Eqs. (2) and (3) ( Hsu & Halgamuge, 2008 ). m k = { 0 l k � = ˆ m k or missing label 1 l k = ˆ m k (2) S m k = C ∑ c=1 w c,m k × o c k m = { 1 , 2 , . . . , M } (3) In Eqs. (1) and (3) , o c k the function determines the similarity of he entered data with cluster nodes and is known as the activation egree. The value of activation degree for each c node is calculated y using a Gaussian function that is given in Eq. (4) ( Hsu & Halga- uge, 2008 ). o c k = exp ( − ‖ x k −v c ‖ 2 ( δc k ) 2 ) c ∈ { 1 , . . . , C } (4) here δc k is average distance of all neighbor’s nodes of c for k th ample and is calculated by (5) and | Ne ( c )| in (5) represents num- er of neighbor’s node c ( Hsu & Halgamuge, 2008 ). c k = ∑ k ∈ Ne ( c ) ‖ v k − v c ‖ | Ne ( c ) | (5) 26 A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 w E w fi l c r t v δ s J ( m fi δ w a 3 r ∇ o More information about how growing GSOM in ( Hsu & Halga- muge, 2008 ). 3. Proposed method Since the number of labeled data may be few, optimal opera- tion of semi-supervised algorithms is very important. Although the method proposed by Hsu is semi-supervised and online, it faces the following problems: • The objective function is optimized only based on the weight of classifier layer, but in fact, the objective function also depends on the parameters of the shape and location of the nodes in the clustering layer and the degree of activation. • Growing of GSOM is not optimized, because the activation func- tion plays an important role in amount of similarity of data for each node, but Hsu algorithm adjusts this parameter by the av- erage distance to neighbor nodes. This innovative solution is a local solution and this method is not optimal by changing data. • It suffers the problem of instability versus noisy data. Accordingly, to overcome the mentioned problems, the objec- tive function was redefined and resolved in the present work. 3.1. Definition of the problem According to Section 2 , in the semi-supervised GSOM error for each data k is defined as: e k = M ∑ m =1 ( l m k − ps m k ) (6) And cost function based on the expectation of error function is defined as: g ( e k ) = N ∑ k =1 exp ( −ηe 2 k ) (7) J = max w, δ, v N ∑ k =1 exp ( −ηe 2 k ) (8) The half-quadratic technique Rockafellar, 1970 ; He, Hu, Zheng, & Guo, 2010 ; Zhang et al., 2014 ) is often used to solve the nonlin- ear optimization problem ( He, Hu, Zheng, & Kong, 2011 ). Therefore, we derived an algorithm to solve ( (8) based on the half quadratic. Based on the theory of convex conjugated functions, we can solve (8) as following proposition. Proposition 1. There exists a convex conjugated function φ of g ( e ) such that g ( e ) = sup p ( ηe 2 p − ϕ ( p ) ) (9) where convex function φ( p ) is defined as φ( p ) = −p log ( − p ) + p . Therefore, J is rewritten as follows: J = max w,δ, v N ∑ k =1 ( ηe 2 k p k − ϕ ( p k ) ) (10) According to Proposition 1 , we notice that when [ w , δ, v ] is fixed, the following equation holds: max w,δ, v J ( w, δ, v ) = max w,δ, v ,p ˆ J ( w, δ, v , p ) (11) Then, the alternative solution is used: ste p 1 : max p ˆ J ( w, δ, v , p ) ⇒ p = −exp ( −ηe 2 ) . ste p 2 : max w,δ, v ˆ J ( w, δ, v , p ) ⇒ max w, δ, v J = max w,δ, v N ∑ k =1 ηe 2 k p k . By defining q k = −p k = exp ( −ηe 2 k ) : min , δ, v J = min w,δ, v N ∑ k =1 ηe 2 k q k (12) Let assume η is constant and expand (12) as: q. ( 8 ) ⇒ J = ∑ n k =1 ⎛ ⎝ ( M ∑ m =1 ( l m k − ps k m )) 2 × q k ⎞ ⎠ Eq. 3 ⇒ J = ∑ n k =1 ((∑ M m =1 ( l m k − (∑ C c=1 w c,m k × o c k )))2 × q k ) = ∑ n k =1 ((∑ M m =1 ∑ C c=1 ( l m k − w c,m k × o c k ))2 × q k ) Eq. 4 ⇒ J = ∑ n k =1 ((∑ M m =1 ∑ C c=1 ( l m k − w c,m k × exp ( − ‖ x k − v c ‖ 2 σ 2 c )))2 × q k ) (13) Therefore: min ,δ, v ,p J ( w, σ, v ) = min w,δ, v ,p ∑ n k =1 ((∑ M m =1 ∑ C c=1 ( l m k − w c,m k × exp ( − ‖ x k − v c ‖ 2 σ 2 c )))2 × q k ) (14) Based on (14) , the effective parameters include weight classi- er layer ( w ), the shape and location of the nodes in the clustering ayer ( v ) and the width of activation degree function ( δ). It is so lear that the parameter δ is important because it affects w di- ectly. Moreover, the calculation of the optimal value of parame- er v in the objective function practically is impossible and will be ery time consuming. Therefore, if the optimal value of parameter can be determined, other parameters will be optimal. In conclu- ion, it can be derived: ( w, δ, v ) ∼= J ( w, δ) (15) Next, we use the gradient search method in the following form: w d ) k +1 = ( w d ) k − μ ˆ ∇ ( w d ) k (16) Now, Eq. (16) is simplified in two steps: one step is an adjust- ent of clustering layer and the other is the adjustment of classi- cation layer. For clustering layer, it is: k +1 = δk − μδ ˆ ∇ δk (17) And in classification layer, k +1 = w k − μw ̂ ∇ w k (18) By derivation of each variable, the minimal cost function is chieved by the corresponding variable ( μδ and μw ). .2. Parameters adaption To obtain the adaptive value of δ, we use a least mean algo- ithm for Eq. (17) as: ˆ δk = ∂ J k ∂ δk = 2 e k ∂ e k ∂ δk (19) The optimal value of δ is calculated through a derivation of the bjective function (14) based on δ as follows: ∂ e k ∂ δk = ∂ (∑ n k =1 ∑ M m =1 ∑ C c=1 ( l m k − w c,m k ex p ( − ‖ x k −v c ‖ 2 σ 2 c )) × q k )) ∂ δk A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 27 n δ ∇ s s b ∇ w n w 3 b e e ⇒ 3 fi f m w l t η 3 u t the width degree of activation of each node. = 0 − ∂ (∑ n t=1 (∑ M m =1 ∑ C c=1 ( w c,m k ex p ( − ‖ x t −v c ‖ 2 σ 2 c ))) × q k ) ∂ δk (20) If c is replaced with Best Matching Node (BMN of winning ode) then (20) is rewritten as follows, = − ∂ ∂ δk (∑ n k =1 (∑ M m =1 ( w BMN,m k × exp ( − ‖ x k − v BMN ‖ 2 σ 2 BMN )) × q k )) = − ∑ n k =1 ∑ M m =1 ( 2 × w BMN,m k × exp ( − ‖ x k −v BMN ‖ 2 σ 2 BMN ) × ( ‖ x k − v BMN ‖ 2 )) × q k δ3 BMN (21) So, according to Eq. (19) , δ is updated as follows: k +1 = δk + 4 μδ × ∑ n k =1 ∑ M m =1 ( l m k − P S m ) × ( w BMN,m k × exp ( − ‖ x k −v BMN ‖ 2 σ 2 BMN ) × ( ‖ x k − v BMN ‖ 2 )) × q k δ3 BMN δ = 4 μδ × ∑ n k =1 ∑ M m =1 ( w BMN,m k × ( l m k − P S m ) ×exp ( −‖ x k −v BMN ‖ 2 σ 2 BMN ) × ( ‖ x k −v BMN ‖ 2 )) q k δ3 BMN (22) The coefficient μδ affects the convergence speed of δ. For the ake of simplicity, the learning rate can be assumed as a con- tant and positive value that is reduced by time and exceeding the oundary confidence. To obtain adaptive w , likewise the δ, we have: ˆ w t ( e 2 k ) = ∂ J k ∂ w k = 2 e k ∂ e k ∂ w k (23) here ∂ e k ∂ w k = ∂ ( ∑ n k =1 ∑ M m =1 ∑ C c=1 ( l m k − w c,m k ex p ( − ‖ x k −v c ‖ 2 σ 2 c )) × q k )) ∂ w k = 0 − ∂ ( ∑ n k =1 ∑ M m =1 ∑ C c=1 ( w c,m k ×ex p ( −‖ x k −v c ‖ 2 σ 2 c )) × q k )) ∂ w k (24) If c is replaced with Best Matching Node (BMN of winning ode) then (24) is rewritten as follows: = − ∂ ((∑ n k =1 (∑ M m =1 ( w BMN,m k × exp ( − ‖ x k −v BMN ‖ 2 σ 2 BMN )) × q k ))) ∂ w k = − ∑ n k =1 (( exp ( − ‖ x k − v BMN ‖ 2 σ 2 BMN )) × q k ) (25) Therefore, w is updated as follows: k +1 = w k + 2 μw × n ∑ k =1 M ∑ m =1 ( l m k − ps m k ) × exp ( − ‖ x k − v BMN ‖ 2 σ 2 BMN ) × q k ∇w = 2 μw × ∑ n k =1 ∑ M m =1 ( l m k − ps m k ) × exp ( − ‖ x k −v BMN ‖ 2 σ 2 BMN ) × q k (26) .3. Parameters adaption through back propagation method In addition to Section 3.1 , the back propagation method can also e used for error estimation. In the back propagation method, the rror is defined as Eq. (27) . k = M ∑ m =1 ( l m k − ps m k ) × w m k (27) Hence, (21) id rewritten as: ∂ e k ∂ δk = ∂ ( ∑ n k =1 ∑ M m =1 ∑ C c=1 ( l m k − w c,m k × exp ( − ‖ x k −v c ‖ 2 σ 2 c )) × w c,m k × q k )) ∂ δk After simplification: ∂ e k ∂ δk = − n ∑ k =1 M ∑ m =1 ( 2 × ( w BMN,m k )2 × exp ( − ‖ x k − v c ‖ 2 σ 2 BMN ) × ( ‖ x k − v c ‖ 2 σ 3 BMN )) × q k ) ⇒ ∇δ = 4 μδ × ∑ n k =1 ∑ M m =1 (( w BMN,m k )2 × (l m k − ps m k ) × exp ( − ‖ x k −v BMN ‖ 2 σ 2 BMN ) × ( ‖ x k − v BMN ‖ 2 )) q k δ3 BMN (28) Also, for w, it is as follows: ∂ e k ∂ w k = ∂ ( ∑ n k =1 ∑ M m =1 ∑ C c=1 ( l m k − P S m ) × w k m × q k )) ∂ w k (29) After simplification, we have: ∂ e k ∂ w k = −2 × ∑ n k =1 ∑ M m =1 ∑ C c=1 w BMN,m k × exp ( − ‖ x k − v BMN ‖ 2 σ 2 BMN ) × q k ∇w = 4 μw × ∑ n k =1 ∑ M m =1 ( l m k − ps m k ) × ( w BMN,m k )2 × exp ( − ‖ x k − v BMN ‖ 2 σ 2 BMN ) × q k (30) .4. ƞ tuning In this section, we discuss adaption of the ƞ parameter. We de- ne a constraint for ƞ parameter over the minimization problem as ollows: in ηi ∑ N i =1 η2 i e 2 i q i , ∑ ηi = p, ηa i ∈ [ 0 , 1 ] (31) Producing the Lagrange equation by adding the penalty term, e have: = N ∑ i =1 η2 i e 2 i q i − λ (∑ ηi − p ) (32) Then, we obtain ƞi as follows: ∂l ∂ ηi = 0 ⇒ ηi = λ 2 e 2 i q i ∑ ηi = p ⇒ ηi = p ∑ n j=1 e 2 i q i e 2 j q j (33) To deal with division by zero problems, we add ξ parameter to he denominator of the equation as follows: i = p ∑ n j=1 e 2 i q i e 2 j q j + ξ (34) .5. Structure of the proposed method As shown in Fig. 2 , the proposed algorithm receives labeled and nlabeled data continuously. Depending on the type of the data, wo events may happen: • Unlabeled data are received: GSOM performs clustering and up- dates weight of nodes in the clustering layer. In this study, we assumed that most of data sizes are unlabeled. • Data with known labeled is received: First, the winner node in the clustering layer is identified and then the label of data is estimated by the system based on this node. The system error is calculated from the difference between actual Label and esti- mated Label system. This error, which updates weight between layers of clustering and classification, is applied to determine 28 A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 Fig. 2. The proposed model for creation adaptive Semi-supervised GSOM. Fig. 3. Synthetic evaluated datasets: (a) four cluster (b) two moon (c) spiral (d) two ring. Table 2 Abbreviation and title of proposed and state of the art methods. Abbreviated name Title of method ASSGSOM Adaptive Semi Supervised GSOM BASSGSOM Back propagation Adaptive Semi Supervised GSOM SSGSOM Hsu’s Semi-supervised GSOM{Hsu, 2008 #20} CS2GS Constrained Semi-Supervised GSOM{Allahyar, 2015 #2} HSS Halkidi’s Semi Stream algorithm{Halkidi, 2012 #18} P 4. Performance evaluation and tests In this section, the efficiency of the proposed algorithm is eval- uated. The datasets used for the evaluation include a synthetic and UCI repository, 2 which are detailed in the next sections. Three online algorithms were selected among the state of the art methods mentioned in Section 1 . Table 2 shows title and ab- breviation for the two proposed method and state of the art meth- ods. The measures considered for assessing the proposed method and those from the literature include Accuracy and F-measure. The Accuracy measure calculates based on (35) . In (35) , TP, TN, FP, and FN are true positive, true negative, false positive, and false nega- tive, respectively. Ac c q = T P T P + T N + F P + F N (35) 2 http://archive.ics.uci.edu/ml/ R Also, Precision and Recall measures are defined as follows: R = T P T P + F P (36) E = T P T P + F N (37) A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 29 Fig. 4. Result on Two Moon dataset. d F a a c G c t r t v s o t t T a i t l r d 4 s s d f r d t t l o s i w Fig. 5. Result on Four Cluster dataset. Fig. 6. Result on Two Ring dataset. Fig. 7. Result on Spiral dataset. T S s A A 4 U ( m d S And then F -measure, which is based on precision and recall, is efined as follows: M = 2 × P R q × R E q P R q + R E q (38) For all algorithms that are based on GSOM (i.e., SSGSOM, CS2GS, nd both proposed methods), the scale factor (SF) is considered as constant value and based on growing threshold of GSOM is cal- ulated as (39) . Parameter d in (39) represents data dimension. T = −d ∗ log ( SF ) (39) Furthermore, the regularization parameter for GSOM layer and lassification layer is very influential on the results. So, in all of he experiments, constant values of 0.4 and 0.6 were used for the egularization parameter for GSOM and Classification Layer, respec- ively. Besides, the initial value of w (weight of classification layer), (weight of GSOM node) and other settings related to GSOM, are elected randomly and are the same for all algorithms. The order f data entry, which affects the performance of the algorithms is he same for all of them. The test scenario is that at first each data set is divided into he train and test subsets considering a 10-folds cross-validation. hen, 5% of training data is labeled randomly and all the methods re trained using them. Next, the performance of each algorithm s evaluated using test data. This process is replicated 5 times and he average of results is reported. In the following, the number of abeled training data is increased by 5% and the test scenario is epeated. This procedure will continue until the number of labeled ata is 30%. .1. The evaluation results on synthetic datasets In this section, the performances of the proposed methods and tate of the art methods are shown on synthetic datasets. The tructures of the synthetic datasets are shown in Fig. 3 . All these ata sets have two classes. Also, the number of data in two moons, our clusters, two ring and spiral sets are 258, 160, 237, and 30 0 0, espectively. Figs. 4–7 present the obtained results using the synthetic atasets based on F-measure. As shown in Figs. 4–7 , ASSGSOM method is superior compared o other methods. The closest results to this method are those ob- ained by SSGSOM method. Moreover, a statistical t -test was estab- ished to determine whether the difference between these meth- ds is statistically significant. Table 3 , presents the t -test compari- on between ASSGSOM and SSGSOM. In this test, 15% of the data s labeled and a number of iteration is 35. In Table 3 , if the value of t -test is 1, it means that the method ith a higher average is statistically superior. For example, in the wo Ring dataset, the value of Accuracy t -test column for ASSG- OM is 1. It means that Accuracy result of ASSGSOM is statistically uperior to the SSGSOM because the average of the accuracy of the SSGSOM algorithm (98%) from SSGSOM (93%) is higher, then the SSGSOM algorithm is superior. .2. The evaluation results on UCI datasets In this section, the performance of the proposed methods on CI datasets is evaluated. Several datasets from UCI repository Blake & Merz, 1998 ) were selected to evaluate the proposed ethod. Table 4 shows the details of each dataset. In Figs. 8–13 , results on a part of UCI datasets are shown and etailed results are reported in Table 5 . Results in Figs. 8–13 indicate the relative superiority of ASSG- OM compared to state of art methods. As can be seen, when the 30 A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 Table 3 The t -test results between ASSGSOM and SSGSOM. Dataset method Accuracy (%) F -Measure (%) Accuracy t -test F -Measure t -test Two Moon ASSGSOM 96 95 1 1 SSGSOM 88 87 Four Cluster ASSGSOM 94 94 1 1 SSGSOM 86 85 Two Ring ASSGSOM 98 98 1 1 SSGSOM 93 92 Spiral ASSGSOM 88 87 1 1 SSGSOM 86 85 Fig. 8. Result on Spiral dataset. Fig. 9. Result on Iris dataset. Fig. 10. Result on Wine dataset. Fig. 11. Result on E.coli dataset. A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 31 Fig. 12. Result on Vehicle dataset. Fig. 13. Result on Vowel dataset. Table 4 Detail of used UCI repository. Dataset name Number of class Number of features Number of data Soybean 4 35 47 Iris 3 4 150 Wine 3 13 178 Sonar 2 60 208 E.coli 6 6 332 Ionosphere 2 34 351 WDBC 2 30 569 Scale 3 4 625 Vehicle 4 18 846 Vowel 11 10 990 n l u 3 i f S s S Table 5 The results of the proposed methods and state of art me Dataset/Method Measure ASSGSOM (%) BAS Soybean F -Measure 94 96% Accuracy 87 89% Iris F -Measure 94 90% Accuracy 94 90 Wine F -Measure 70 63% Accuracy 71 63% Sonar F -Measure 60 67% Accuracy 61 67% E.coli F -Measure 83 73% Accuracy 84 75% Ionosphere F -Measure 90 85% Accuracy 91 87% WDBC F -Measure 90 87% Accuracy 90 87% Scale F -Measure 75 70% Accuracy 84 76% Vehicle F -Measure 51 50% Accuracy 52 50% Vowel F -Measure 88 71% Accuracy 88 72% umber of labeled data is increased the proposed method shows ess fluctuate and it becomes more stable. Results in different UCI dataset are summarized in Table 5 . Eval- ation is done in the condition that the number of labeled data is 0% of the total data in the dataset. The results in Table 5 show that when the number of Attributes s high (such as Soybeans and Sonar), ABSSGSOM is better. Except or the results on Iris dataset, in the other datasets ASSGSOM and SGSOM methods are superior compared to other methods. The re- ults of these two methods are very close to each other but ASSG- OM is superior totally. thods on UCI datasets. SGSOM SSGSOM (%) CS2GS (%) HSS (%) 93 85 82 85 70 80 97 90 90 97 90 89 68 70 67 70 71 67 57 57 51 58 58 51 81 70 64 82 72 67 88 86 80 89 87 81 88 87 80 88 87 80 71 68 55 84 69 56 52 45 42 51 45 42 88 64 60 87 65 61 32 A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 Fig. 14. Effect of the noise on the four datasets selected. Fig. 15. Resistance analysis of the accuracy of the all evaluated datasets. 5 t q m a a t s o b c r a R A 4.3. The effect of noise In this section, noise effect on the proposed methods is eval- uated. Fig. 14 presents the results for the datasets contaminated with noise. Eq. (40) is used for calculating the SNR. 3 In this for- mula, D is the domain of signal. SN R dB = 20 lo g 10 ( D signal D noise ) (40) 4.4. Resistance analysis Resistance analysis shows the stability of the algorithm in dif- ferent condition ( Geng, Zhan, & Zhou, 2005 ). Eq. (41) presents how to calculate resistance analysis. Fig. 15 exhibits resistance analysis for the discussed dataset. R A q = Ac c q ∀ j ∈ { 1 , 2 , . . . , Q } (41) max Ac c j 3 Signal to noise ratio. A A . Conclusion and future works In this paper, a comprehensive robust mathematical formula- ion of the semi-supervised GSOM was proposed and then half uadratic (HQ) was used to solve it. Furthermore, an adaptive ethod was proposed to adjust activation degree optimally for chieving a better performance of semi-supervised GSOM. Evalu- tion in noisy conditions and on different datasets suggests that he proposed method has superiority and stability compared to the tate of the art methods. In the future, we add a regular term to cost function to avoid ver-fitting the complex dataset. Also, we plan to develop a ro- ust active semi-supervised SSGSOM. Using this active learning, we an determine which data are important for increasing the accu- acy. Therefore, a higher degree of accuracy can be achieved with smaller number of labeled data. eferences lahakoon, D. , Halgamuge, S. K. , & Srinivasan, B. (20 0 0). Dynamic self-organizing maps with controlled growth for knowledge discovery. IEEE Transactions on Neu- ral Networks, 11 (3), 601–614 . llahyar, A. , Yazdi, H. S. , & Harati, A. (2015). Constrained semi-supervised growing self-organizing map. Neurocomputing, 147 , 456–471 . lpaydin, E. (2004). Introduction to machine learning . MIT Press . A. Mehrizi et al. / Expert Systems With Applications 105 (2018) 23–33 33 B B C C D F F G H H H H H H K L L L M M M R S Z Z eyer, O. , & Cimiano, P. (2012). Online semi-supervised growing neural gas. Interna- tional Journal of Neural Systems, 22 (05), 1250023 . lake, C. L., & Merz, C. J. (1998). UCI repository of machine learning databases, Irvine, CA: Department of Information and Computer Science, University of Cal- ifornia , 55. [http://www.ics.uci.edu/ ∼mlearn/MLRepository.html] . hen, S. , Li, S. , Su, S. , Cao, D. , & Ji, R. (2014). Online semi-supervised compressive coding for robust visual tracking. Journal of Visual Communication and Image Representation, 25 (5), 793–804 . ong, Y. , Liu, J. , Yuan, J. , & Luo, J. (2013). Self-supervised online metric learning with low rank constraint for scene categorization. IEEE Transactions on Image Process- ing, 22 (8), 3179–3191 . ornaika, F. , & El Traboulsi, Y. (2016). Learning flexible graph-based semi-supervised embedding. IEEE Transactions on Cybernetics, 46 (1), 206–218 . arid, D. M. , Zhang, L. , Hossain, M. A. , Rahman, C. M. , Strachan, R. , Sexton, G. , et al. (2013). An adaptive ensemble classifier for mining concept drifting data streams. Expert Systems with Applications, 40 (15), 5895–5906 . ritzke, B. (1994). Growing cell structures—a self-organizing network for unsuper- vised and supervised learning. Neural Networks, 7 (9), 1441–1460 . eng, X. , Zhan, D.-C. , & Zhou, Z.-H. (2005). Supervised nonlinear dimensionality re- duction for visualization and classification. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, 35 (6), 1098–1107 . alkidi, M. , Spiliopoulou, M. , & Pavlou, A. (2012). A semi-supervised incremental clus- tering algorithm for streaming data advances in knowledge discovery and data min- ing (pp. 578–590). Springer . e, R. , Hu, B. G. , Zheng, W. S. , & Guo, Y. (2010). Two-Stage Sparse Representation for Robust Recognition on Large-Scale Database. In AAAI: Vol. 10 (p. 1) . e, R. , Hu, B.-G. , Zheng, W.-S. , & Kong, X.-W. (2011). Robust principal component analysis based on maximum correntropy criterion. IEEE Transactions on Image Processing, 20 (6), 1485–1494 . su, A. , & Halgamuge, S. K. (2008). Class structure visualization with semi-super- vised growing self-organizing maps. Neurocomputing, 71 (16), 3124–3130 . uang, G.-B. , Li, M.-B. , Chen, L. , & Siew, C.-K. (2008). Incremental extreme learning machine with fully complex hidden nodes. Neurocomputing, 71 (4), 576–583 . uang, G.-B. , Zhu, Q.-Y. , & Siew, C.-K. (2004). Extreme learning machine: A new learning scheme of feedforward neural networks. In Proceedings of the IEEE in- ternational joint conference on neural networks . ohonen, T. (1990). The self-organizing map. Proceedings of the IEEE, 78 (9), 1464–1480 . eite, D. , Costa, P. , & Gomide, F. (2010). Evolving granular neural network for semi– supervised data stream classification. In Proceedings of the international joint conference on neural networks (IJCNN) . i, K. , Zhang, J. , Xu, H. , Luo, S. , & Li, H. (2013). A semi-supervised extreme learn- ing machine method based on co-training . Journal of Computational Information Systems, 9 (1), 207–214 . u, K. , Wang, Q. , Xue, J. , & Pan, W. (2014). 3D model retrieval and classification by semi-supervised learning with content-based similarity. Information Sciences, 281 , 703–713 . acario, V. , & de Carvalho, F. d. A. (2012). An adaptive semi-supervised fuzzy clus- tering algorithm based on objective function optimization. In Proceedings of the IEEE international conference on fuzzy systems (FUZZ-IEEE) . artinetz, T. M. , Berkovich, S. G. , & Schulten, K. J. (1993). Neural-gas’ network for vector quantization and its application to time-series prediction. IEEE Transac- tions on Neural Networks, 4 (4), 558–569 . aximo, V. R. , Quiles, M. G. , & Nascimento, M. C. (2014). A consensus-based semi– supervised growing neural gas. In Proceedings of the international joint conference on neural networks (IJCNN) . ockafellar, R. (1970). Convex analysis . Princeton, NJ: Princeton University Press Google Scholar . hen, F. , Yu, H. , Kamiya, Y. , & Hasegawa, O. (2010). An online incremental semi– supervised learning method. Journal of Advanced Computational Intelligence and Intelligent Informatics, 14 (6), 593–605 . hang, Y. , Wen, J. , Wang, X. , & Jiang, Z. (2014). Semi-supervised learning com- bining co-training with active learning. Expert Systems with Applications, 41 (5), 2372–2378 . hu, X. , & Goldberg, A. B. (2009). Introduction to semi-supervised learning. Synthesis Lectures on Artificial Intelligence and Machine Learning, 3 (1), 1–130 . 
work_n2nvxlrrxzgkxn4tlkfxom5ao4	----	Multi-BP expert system for fault diagnosis of powersystem Engineering Applications of Artificial Intelligence 26 (2013) 937–944 Contents lists available at SciVerse ScienceDirect Engineering Applications of Artificial Intelligence 0952-19 http://d n Corr E-m journal homepage: www.elsevier.com/locate/engappai Multi-BP expert system for fault diagnosis of power system Deyin Ma, Yanchun Liang, Xiaoshe Zhao, Renchu Guan, Xiaohu Shi n College of Computer Science and Technology, Key Laboratory of Symbol Computation and Knowledge Engineering of the Ministry of Education, Jilin University, Changchun, China a r t i c l e i n f o Article history: Received 7 March 2012 Accepted 28 March 2012 Available online 18 April 2012 Keywords: Power system Expert system Multi-BP networks 76/$ - see front matter & 2012 Elsevier Ltd. A x.doi.org/10.1016/j.engappai.2012.03.017 esponding author. ail address: shixh@jlu.edu.cn (X. Shi). a b s t r a c t Fault diagnosis and assessment is a crucial and difficult problem for power system. Back propagation neural network expert system (BPES) is an often used method in fault diagnosis. However, with the layer numbers increasing, BPES becomes time consuming and even hard to converge. To solve this problem, we divide the whole networks into many sub-BP groups within a short depth and then propose a novel Multi-BP expert system (MBPES) based method for power system fault diagnosis. We use two real power system data sets to test the effectiveness of MBPES. Experimental results show that MBPES obtains higher accuracy than two commonly used methods. & 2012 Elsevier Ltd. All rights reserved. 1. Introduction In modern times, power system becomes larger and more complex than before. With its fast development, higher demand for the sustainability and stability of power system is of great requirement. However, some common faults in power system have never been resolved very well and are still hindering the stability of power system, such as transmission fault, network distribution fault, power variable fault (Mizutani et al., 2007). Sometimes, even only one fault could destroy the equipments in power system, and might affect the whole power system. An even worse damage could cause conflagration and casualties, and leading to a huge pecuniary loss. Therefore, it is of great significance to do researches for preventing those faults from power system. And fault diagnosis is a powerful tool to guarantee the safety and reliability of power system. In power system, transformer is a kind of major equipment and plays an important role in power transmission. It can raise voltage so that power can be transported to the user with less loss. On the other hand, the transformer can reduce power into different voltage levels, which can satisfy variety needs of users. Because of its complex structure and function, transformer is tending to cause fault. Unfortunately, transformer fault is very difficult to predict. Moreover, if some accidents take place in transformer, the whole system has no choice but to stop to check and maintain the equipments. Therefore, it is believed that keeping the transformer running in perfect situation plays a key role in power system diagnosis (Lin and Zeng, 2009). High voltage circuit breaker (over 3 kV) is another important element in power system. It has two main functions, namely controlling and ll rights reserved. protecting. Firstly, it decides when and which parts of the power system should be started or stopped according to the require- ments; secondly, when some errors occur in power networks or equipments, the high voltage circuit breaker will quickly break off the error parts from power system, so that other parts can work without influence. In other words, high voltage circuit breaker is able to control the normal current in power lines, and to deal with the overload current, short-circuit current and other abnormal current within a limited time. When a mistake happens in high voltage circuit breaker, it will usually expand to other parts of the power system and finally lead to a worse accident. There are some classical artificial intelligence technologies have been used in power system fault diagnosis, for example: the expert system (Ma et al., 2010), artificial neural networks (El-madany et al., 2011; Zhu et al., 2006; Huang et al., 2002; Karthikeyan et al., 2005), decision tree theory (Qu and Gao, 2008) etc. In recent years, some new theories have been applied in this field, such as data mining (Athanasopoulou and Chatziathanasiou, 2009), fuzzy set theory (Lee et al., 2000; Zhang et al., 2010), rough set theory (Li and Wang, 2010, Li et al., 2011), petri-network (Yang et al., 2004), support vector machine (Eristi and Demir, 2010), multi-agent systems (Zaki et al., 2007), and so on. Li and Liu had performed a comprehensive review of the above-men- tioned methods (Li and Liu, 2010). They pointed out that there are some problems in the existing intelligent fault diagnosis expert system theology, such as the difficulty for knowledge gaining and managing, low on-line usage of fault diagnosis, high error rate, poor efficiency of inference process, and so on. Back propagation neural network (BPNN) expert system is an often used method in fault diagnosis. In real applications, BPNN usually has many layers. However, the training time of BPNN will grow exponen- tially with the layer number increasing. While more serious problem is that it is difficult to converge when BPNN has a large number of layers. Another problem is that the diagnostic accuracy www.elsevier.com/locate/engappai www.elsevier.com/locate/engappai dx.doi.org/10.1016/j.engappai.2012.03.017 dx.doi.org/10.1016/j.engappai.2012.03.017 dx.doi.org/10.1016/j.engappai.2012.03.017 mailto:shixh@jlu.edu.cn dx.doi.org/10.1016/j.engappai.2012.03.017 D. Ma et al. / Engineering Applications of Artificial Intelligence 26 (2013) 937–944938 of BPNN is still not satisfied. To solve those problems, we propose a so called multi-BP expert system (MBPES) method. In MBPES, the whole BPNN networks are divided into many sub-BP groups within a short depth, saying about 5 layers. In this manner, the consumed training time is greatly reduced and it is easy to achieve the convergence of the training process. In the experiments, we firstly compare the performance of BPNN with different number layers according to a XOR problem. Numerical results show that when the layer number is more than 6, BPNN is very time consumed and even hard to converge. To test the effectiveness of the proposed MBPES method, it is applied to two real power system data sets, namely that the transformer data set and high voltage circuit breaker data set. Experimental results show that MBPES is very efficient, and it is more accurate than two other compared methods. 2. Back propagation expert system Back propagation (BP) algorithm is one of the most classical and successful learning methods of feed forward artificial neural network, which is based on gradient descent algorithm. For its success, those feed forward neural networks using BP algorithm are always called BPNNs. Fig. 1 shows an architecture of BPNN model with K hidden layers. There are n nodes in the input layer, which is corresponding to the sample vector’s dimension. And the inputs of the input layer are the components of the sample’s vector. Denote the node’s number of the jth hidden layer as Nj, then the outputs of first hidden layer are Y 1j ¼ 1 1�ð Pn i ¼ 1 XiUw 1 ji Þ j¼1,. . .,N1 ð1Þ And the outputs of the other hidden layers are Y kj ¼ 1 1�ð PNk�1 i Y k�1i Uw k ji Þ k¼2,. . .,K; j¼ 1,. . .,Nk ð2Þ Suppose the related problem has m expected outputs, then the output layer should have m nodes, and their outputs are Z1j ¼ 1 1�ð PNK i ¼ 1 Y Ki Uw O ji Þ j¼1,. . .,m ð3Þ Here wji is the weight between the jth node of one layer and the ith node of its former layer. The weights should be trained before application. The training method could be referred to Rumelhart et al. (1986). BPNN also can be used as an expert system, which is called back propagation expert system (BPES). In BPES, the rules with the form X1 X2 Input layer Y11 Hidden la Xm YN11 Fig. 1. Architectu of IF-THEN in knowledge database are represented by the weights of networks. An example of BPNN expert system is shown as Fig. 2. In the networks, the nodes represented by rectangles are the ante- cedents or consequents of the rules, while the nodes represented by circles are corresponding to the rules. The initial structure of the networks is generated by rules in knowledge database, as well as the values of the weights. If we could obtain enough known sample data, the weights can be improved by training process, even the topology of the networks might be amended. The training algorithm is the same as that of BPNN. In BPES networks, each rule creates a 3-layer sub-network. The first layer represents the antecedents of the rule, the second layer represents the transform relationship including only one node, the third layer represents consequent of the rule, also including one node. Therefore, the nesting of different rules may extend the length of the network structure. Unfortunately, it is often the case in the knowledge database. However, when the BPNN networks have more than 5 layers, the training time and iteration number will grow exponentially and even hard to converge. To solve such problem, a called Multi-BP Expert System (MBPES) is proposed in this paper, in which the networks are divided into sub-networks with small scale. 3. Multi-BP expert system (MBPES) 3.1. The structure of MBPES In a BPNN, a rule is corresponding to a substructure of the networks with 3 layers. Accordingly, 2 and 3 nested rules are corresponding to a substructure with 5 and 7 layers, respectively. Because those BPNN networks with more than 5 layers are difficult to train and even hard to converge, we divide the whole networks into sub-networks with 5 or less layers. Fig. 3 shows the division of a BPNN in serial structure. However, in real applications, rules are not always nested in serial. For example, the inputs of a rule might be the outputs from other 2 or more rules. For 2 independent rules, the accordingly sub-networks are in parallel. Moreover, the inputs of a rule could come from different rules in different layers. For such a compli- cated BPNN, we also need to divide it into sub-networks with 5 or less layers, namely that Multi-BP expert system (MBPES). A classical structure of MBPES is depicted as Fig. 4. 3.2. The construction algorithm of MBPES Given a knowledge base, the structure of the MBPES should be determined firstly. To begin with the construction algorithm of Output layeryers Y1K YNKK Z1 Z2 Zn re of BPNN. R1 1 a1 a2 b1 1.0 1.0 cf1 R k 1 c1 cf1 a3 a4 an R1 2 R1 3 R1 4 R 1i b2 bm R k 2 R kj cp 1.0 1.0 1.0 1.0 1.0 1.0 cf2 cf3 cf4 cf i cf2 cfj 1.0 1.0 1.0 1.0 Input layer Hidden 1 ... ... 1.0 ... ... ... ... ... Hidden 2 Hidden 3 Hidden 4 Fig. 2. Architecture of BPES. PBPB X1 X4 Xn X5 X3 X2 Y1 Z1 Y3 Y2 Ym Z2 Z3 Z4 Zk... ... ... layers5≤ layers5≤ Fig. 3. Division of a BPNN in serial structure. X1 1 Y1 Y2 Ym ... ... ... Z11 Z2 1 ZK11 ... PBPB ... ... ... ...... X2 1 XN1 1 X1 G X 2 G XNG G Z1 G Z2 G ZKGG layers5 layers5< < Fig. 4. A classical structure of MBPES. R1 a1 a2 b1 1.0 1.0 0.8 Fig. 5. The digraph generated from R1. D. Ma et al. / Engineering Applications of Artificial Intelligence 26 (2013) 937–944 939 MBPES, we look at how to build up a digraph from only one rule (Shi et al., 2006). Denote an empty digraph as F, and assume we have a rule R1 as R1: IF a1 AND a2 THEN b1 (0.8) Then, R1 could be input into F, and a new digraph D is created which is shown as Fig. 5. Next, if we have rules which have antecedents or consequents associated with D, they could be added into the digraph. For instance, in Fig. 6 there are seven rules (from R2 to R8) having related antecedents or consequents to digraph D. Then D can be updated as Fig. 7 shown, which has 5 layers. However, when more rules are added into digraph, it might exceed 6 layers. Taking aforementioned rules as an example, the antecedents of the following rule R9 are associated with D: R 9: IF c1 AND c2 THEN c3 (0.8). But if R9 is added into D directly, it will be changed into 7 layers. Therefore it is divided into 2 sub-digraphs according to the 5th layer. It is shown in Fig. 8. Now considering a knowledge database KD which includes N rules, a MBPES could be built up similarly. Denote Ri is the ith rule in KD (i¼1,2,y,N), D is the digraph, Dk is the kth divided sub-digraph (k¼1,2,y,j), j is the number of sub-digraphs, LDk is the layer number of Dk. Then the main steps of the construction algorithm are described as follows: (1) Initialization: Set i¼1, j¼1, D1¼F, D¼{ D1}; (2) Determination: If all the rules in KD have been read, namely i4N, output MBPES, END the algorithm, else go to step (3); D. Ma et al. / Engineering Applications of Artificial Intelligence 26 (2013) 937–944940 (3) Input of rule Ri: (3.1) Search Ri in sub-digraphs. If (k, s.t. the antecedents or consequents of Ri are related to sub-digraph Dk, then check if Dk has more than 5 layers. If yes, jþþ, create a new sub-digraph Dj, input Ri into Dj, else, input Ri into Dk. go to step (3.3). (3.2) Create a new sub-digraph for Ri. If 8k, the antecedents or consequents of Ri are not related to sub-digraph Dk, jþþ, create a new sub-digraph Dj, input Ri into Dj, go to step (3.3). (3.3) iþþ, go to step (2). The flowchart of the algorithm is depicted in Fig. 9. 3.3. The training algorithm of MBPES The constructed MBPES could be considered as a weighted digraph. However, there are lots of noises in the rules saved in KD. So a training process is necessary to improve the accuracy of the system. By the training process, the weights in digraph are updated according to a real data set, which are more reliable compared to the rules in KD. Even some weights can be deleted and some new ones can be added. Fig. 6. Rules from R2 to R8. R1 a1 a2 b1 1.0 0.8 a3 a4 a5 R2 R3 R4 R5 b2 b3 1.0 1.0 1.0 1.0 1.0 1.0 0.2 -0.1 Input layer Hidden 1 Hidd 0.5 Fig. 7. The Updated dig Assume there are J sub-digraphs in MBPES D. Denote Dk as the kth of sub-digraphs, Nk the output number of Dk, T k j and Y k j the jth expected output and the jth real output of Dk, respectively. Then the main steps of the training algorithm are as follows: (1) en 2 raph For each sub-digraph Dk (k¼1,2,y,J), define an error function as Ek ¼ 1 2 XNk j ¼ 1 ðT kj �Y k j Þ 2 ð4Þ (2) According to the basic BP training algorithm, train each sub-digraph Dk. (2.1.) If Ek reaches a value less than the error threshold Ethr, then terminate the training process of Dk, and remove those weights with 0 values. (2.2.) Else, if Ek could not decrease below to Ethr within the max iteration number Nmax loops, then add new links between unlinked nodes with 0 weights, and retrain Dk. 1.0 1.0 1.0 1.0 H 1.0 D by R2 (3) When all the sub-digraphs finish the training procedures, update the whole structure of MBPES. 4. Results and analysis 4.1. Comparison for different network layer numbers To test the performance of BPNNs with different layer number, we take XOR problem as an example. In the experiment, the main parameters are set as follows: the inertial weight a is 0.5, the momentum coefficient Z is 0.9, the threshold b is �0.5, and the error threshold Ethr is 0.001, respectively. To get the statistical results, each experiment is repeated for 100 times. Numerical results are shown in Table 1 and Fig. 10. From Table 1 and Fig. 8, we can see that the training time is raising with the increasing of BPNN layer’s number. On the contrary, the convergence rate is falling in a converse manner. It is noteworthy that there is a steep slope between the layer numbers 5 and 6 for both training time and convergence rate. When the layer number is more than 5, training time for BPNN is increasing exponentially and the convergence rate is decreasing sharply. In fact, that is just the reason we want to divide the BP subgroup within 5 layers. R6 c1 0.7 R7 R8 c2 0.8 idden 3 Output layer to R8. R1 a1 a2 b1 1.0 0.8 1.0 R6 c1 0.7 a3 a4 a5 R2 R3 R4 R5 b2 b3 R7 R8 c2 1.0 1.0 1.0 1.0 1.0 1.0 0.2 -0.1 0.5 0.8 1.0 1.0 1.0 Input layer Hidden 1 Hidden 2 Hidden 3 Output layer 1.0 R9 c30.8 1.0 1.0 Hidden 4 Hidden 5 Fig. 8. The Updated digraph D by R9. i k D i m k k j Ri Dk Ri Dk i i j j Dj Ri Dj k k Fig. 9. The flowchart of construction algorithm of MBPES. 2 0 2000 4000 6000 8000 10000 12000 C on ve rg en ce ( % ) T ra in in g T im e (S ec ) Layer number Training Time 0 20 40 60 80 100 Convergence 10864 Fig. 10. training time and convergence. D. Ma et al. / Engineering Applications of Artificial Intelligence 26 (2013) 937–944 941 4.2. Experiment on transformer data Our experimental data used in this section are the real data from the on-site transformer in an electric power design institute. The rules in knowledge base are provided by the engineers in the institute and arranged by authors. There are more than 1200 knowledge rules, of which more than 700 rules are about transformer fault diagnose. For comparison, BPNN, BPES and our proposed MBPES methods are all applied as the inference engines, respectively. There are 150 sample data in all, 100 of which are set as the training data and the other 50 ones are set as the testing data. The first 15 sample data are shown in Table 2. Each sample contains 8 gas acquisition data (in the unit of mL/L), namely H2, CH4, C2H6, C2H4, C2H2, CO, CO2, and HC, together with ‘‘Fault type’’ which indicates the real fault type of the sample. In our experi- ment, there are five real fault types: final superheater, medium superheater, primary superheater, spark discharge and arc dis- charge (Wang et al., 2006) etc. For the preprocessing procedure, the values of the eight gas acquisition data are normalized into interval [0,1] using the following equation: yi ¼ yi�ymin ymax�ymin ð5Þ where yi, ymin, ymax represent the original i-th sample data, the minimum and maximum values of all of the samples with the same input attribute respectively, and yi indicates the normalized i-th sample data. Firstly, according to the construction algorithm proposed in Section 3.2, an MBPES structure has been built using the 700 rules which are related to transformer fault diagnose. And then it is updated according to the learning algorithm in Section 3.3. The final topology of MBPES is shown as Fig. 11. It is interesting that our groupment result is highly consistence with the theoretical results in Li (2010) about the relationship between fault type and the partition of gas inputs. To test the effectiveness of MBPES, it is compared with BPNN and BPES on the prediction results of the 50 test sample data. The comparison results are shown in Table 3. From the table, it could be found that the correct predicted diagnosis record numbers of BPNN, BPES and MBPES are 39, 40 and 43, with the accuracy rates Table 2 The first 15 data in transformer (mL/L). ID H2 CH4 C2H6 C2H4 C2H2 CO CO2 HC Fault type 1 0.00 16.18 10.26 63.76 5.27 141.69 2096.60 95.74 Final superheater 2 0.00 15.65 11.99 63.21 6.05 153.05 2284.45 96.90 Final superheater 3 0.00 200.00 222.00 826.00 0.00 84.00 1114.00 1248.00 Final superheater 4 0.00 179.00 189.00 137.00 0.00 108.00 1075.00 1405.00 Primary superheater 5 0.00 59.00 15.90 111.10 0.00 262.20 764.90 186.50 Final superheater 6 3.00 198.00 191.00 701.00 0.00 396.00 2256.00 1093.00 Final superheater 7 4.00 79.00 112.00 312.00 0.00 45.00 4155.00 503.00 Primary superheater 8 5.49 35.89 28.08 156.00 0.00 12.79 304.56 219.97 Final superheater 9 6.50 38.30 15.50 73.60 0.00 120.60 67.40 127.40 Final superheater 10 54.60 7.20 5.00 37.20 7.00 63.00 571.00 116.80 Spark discharge 11 86.13 31.82 8.58 8.52 21.05 564.48 4837.00 678.00 Arc discharge 12 11.70 55.00 111.40 23.60 0.90 308.90 669.30 190.90 primary superheater 13 14.00 221.00 199.00 804.00 0.00 138.00 1193.00 1227.00 Final superheater 14 29.00 26.30 1.80 27.00 82.40 522.90 398.30 344.90 Arc discharge 15 14.6 98 42.2 204.7 0 715.1 1611 344.9 Final superheat Table 1 XOR BPNN data of experiments. Layer Number 3 4 5 6 7 8 9 Time (Sec) 0.043 0.272 1.85 8.96 190.65 2000.72 11,935.84 Iteration 5642.25 7914.56 20,143.84 1,340,879.96 3,095,324.14 7,971,254.69 30,326,842.44 Convergence (%) 100 100 95 12 5 3 1 Structure of BPNNs with different layer number: 3 layers:2-4-1; 4 layers: 2-4-3-1; 5 layers: 2-4-3-2-1; 6 layers: 2-4-3-2-2-1; 7 layers: 2-4-3-3-2-2-1; 8 layers: 2-4-3-3-2- 2-2-1; 9 layers: 2-4-3-3-3-2-2-2-1. Y1 Y2 Y3 Y4 Y5 X2 X4 X6 X7 C2 C3 C1 A1 A2 A3 B1 B2 B3 B4 Z2 Z3 Z1 E1 E2 E3 E4 E5 Y10 Y11 Y12 X1 X2 X5 X6 Z7 B8 B9 B10 B11 Y6 Y7 Y8 Y9 X1 X3 Z6 Z5 A4 A5 A6 B5 B6 B7 Z4 X5 X8 X3 X4 X5 X6 X1 X2 X7 X8 Z2 Z3 Z1 Z7 Z4 Z6 Z5 C2 C3 C1 Y13 Y14 Y15 A7 A8 A9 Fig. 11. MBPES topology graph for transformer fault diagnose. D. Ma et al. / Engineering Applications of Artificial Intelligence 26 (2013) 937–944942 Table 3 Comparison results (I). ID BPNN BPES Mutli-BP Accuracy rate 78% 80% 86% BPNN: training error is 0.001; main parameters: a¼0.5, b¼�0.5; structure: 8-16- 10-6 BPES: rules available in database: 31; training error 0.001; main parameters a¼0.5, b¼�0.5. Table 4 The first 15 data in high voltage vacuum circuit breaker. ID I1/A I2/A I3/A t1/ms t2/ms t3/ms t4/ms t5/ms T 1 1.61 1.14 2.21 24.5 37.8 43.6 46.7 50.3 ZC 2 1.64 1.16 2.2 24.59 37.78 43.57 46.54 50.35 ZC 3 1.23 0.91 1.81 23.8 36.7 42.3 45.8 50.1 GD 4 1.26 0.95 1.82 23.82 36.84 42.29 45.81 49.97 GD 5 1.61 1.11 2.25 30.1 43.5 49.1 52.4 56.1 HKS 6 1.62 1.08 2.26 30.17 43.47 49.11 52.48 55.99 HKS 7 1.63 1.17 2.21 24.21 37.57 43.25 50.02 54.09 CKS 8 1.62 1.15 2.31 24.2 37.5 43.3 49.9 54.1 CKS 9 1.61 1.13 2.26 24.09 39.75 42.86 45.89 49.79 TD 10 1.64 1.14 2.22 24.15 39.73 42.83 45.94 49.82 TD 11 1.65 1.13 2.23 23.95 37.53 43.35 48.17 52.21 FK 12 1.63 1.16 2.22 23.93 37.47 43.3 48.09 52.09 FK 13 1.62 1.19 2.18 24.36 37.88 43.39 46.59 50.21 ZC 14 1.61 1.15 2.19 23.9 37.49 43.28 48.09 52.19 FK 15 1.22 0.92 1.81 23.76 36.8 42.27 45.77 50.1 GD Y1 Y2 Y3 Y4 X1 X2 Z3 Z2 A1 A2 A3 B1 B2 B3 X3 X4 X5 Fig. 12. A sub-digraph of MBPES for high voltage breaker fault data. Table 5 Comparison results (II). ID BPNN BPES Mutli-BP Accuracy rate 80% 82.5% 90% BPNN: training error is 0.001; main parameters: a¼0.5, b¼�0.5; structure: 8-16- 10-6 BPES: rules available in database: 19; training error 0.001; main parameters a¼0.5, b¼�0.5. D. Ma et al. / Engineering Applications of Artificial Intelligence 26 (2013) 937–944 943 of 78%, 80% and 86%, respectively. Obviously, MBPES outperforms the other two compared algorithms. Taking the first record in Table 2 as an example to describe the output of MBPES, the predicted diagnosis result is ‘‘final superheater’’ which is just met the real fault type of the sample. Furthermore, MBPES gives an advice for the fault type as (1) Please check the wire connector to see whether it is badly contacted and many spots of the iron core touch floor. (2) Please check the current of iron core touching floor. (3) Please check the iron core insulation resistance. (4) Please check DC Resistance. 4.3. Analysis of vacuum circuit breaker experimental data Similarly as stated in Section 4.2, MBPES is compared with BPNN and BPES in another real data set in this section. The experimental data are from a power plant, which are related to high voltage vacuum circuit breaker. There are 300 rules and 120 sample data in all. 80 of the sample data are used for learning and the other 40 data are used for testing. Table 4 lists the first 15 sample data, each of them contains 3 current data (I1, I2 and I3), 5 time data (t1, t2, t3, t4 and t5) and the real fault type (T). Here t1, t2, t3, t4 and t5 are five time points, and the current variation within each time range could reflect mechanics transmission system’s working performance. I1, I2 and I3 are the current values corre- sponding to time points t1, t2 and t4, respectively. In this experi- ment, there are 6 real fault types: running properly (ZC), power in low (GD), Jam in the beginning (HKS), Jam in operation mechanism (CHK), Iron core idle running too long (TD), Auxiliary switch contacts badly (FK) (Lei and Liu, 2010). The same as the above experiment, the values are also normalized according to Eq. (5). According to the construction algorithm and learning algo- rithm, we obtain the final MBPES structure. A sub-digraph is shown as Fig. 12. Wang had described that the Jam related faults (HKS and CHK) are relevant to the current values at time points t1 and t2 (Wang, 2008). Analyzed the inputs and the outputs of the sub-digraph in Fig. 12, we could find that it is highly concordant with Wang’s conclusion. The comparison results of MBPES, BPNN and BPES are shown in Table 5. The accuracy rates of BPNN, BPES and MBPES are 80%, 82.5% and 90%, respectively. MBPES is also better than the other two compared algorithms. Taking the first record in Table 4 as an example, the predicted diagnosis result is ‘‘Auxiliary switch contacts badly (FK)’’, the same with the real fault type of the sample. And an advice for the fault type is given as (1) Please check all the possible leaking points and block them. (2) Please install sealing sleeve in the output link. (3) Please open the heating and drive wave device in the box. 5. Conclusions In this paper, facing the real data in transformer and high voltage breaker, with serial, parallel and hybrid grouping algo- rithm, we proposed two typical new Multi-BP networks. More- over, they successfully solved the problem of low convergence speed caused by the large number of network layers and also greatly increased the speed of diagnosis. Furthermore, by the comparison with BPNN and BPES, we can see that the results from Multi-BP are more accurate. Using multi-BP network to diagnosis power system, not only can we diagnose the transformer online and offline, but also the system will give us the diagnosis advice quickly. Acknowledgment This work was supported in part by the National Natural Science Foundation of China (61073075, 61103092), China Post- doctoral Science Foundation (2011M500613), the Science-Tech- nology Development Project from Jilin Province (20120730, 201215022), Special Fund for Basic Scientific Research of Central Colleges, Jilin University (no. 201103036). References Athanasopoulou, C., Chatziathanasiou, V., 2009. Intelligent system for identifica- tion and replacement of faulty sensor measurements in thermal power plants. Expert Syst. Appl. 36 (5), 8750–8757. D. Ma et al. / Engineering Applications of Artificial Intelligence 26 (2013) 937–944944 El-madany, H.T., Fahmy, F.H., El-Rahman, N.M.A., Dorrah, H.T., 2011. Spacecraft power system controller based on neural network. Acta Astronaut. 69 (7–8), 650–657. Eristi, H., Demir, Y., 2010. A new algorithm for automatic classification of power quality events based on wavelet transform and SVM. Expert Syst. Appl. 37 (6), 4094–4102. Huang, Y.C., Yang, H.T., Huang, K.Y., 2002. Abductive network model-based diagnosis system for power transformer incipient fault detection. IEE Proc. Gener. Transm. Distrib. 149 (3), 326–330. Karthikeyan, B., Gopal, S., Vimala, M., 2005. Conception of complex probabilistic neural network system for classification of partial discharge patterns using multifarious inputs. Expert Syst. Appl. 29, 953–963. Lee, H.J., Park, D.Y., Ahn, B.S., Park, Y.M., Park, J.K., Venkata, S.S., 2000. A fuzzy expert system for the integrated fault diagnosis. IEEE Trans. Power Del. 15 (2), 833–838. Lei, H.Y., Liu, J.G., 2010. Common fault and disposal methods of vacuum circuit breaker. Econ. Res. Guide 12, 205–254. Li, C., 2010. Research of Power Transformer Fault Diagnosis System. Shanghai University of Electric Power 6–8. Li, J.R., Wang, Q.H., 2010. A rough set based data mining approach for house of quality analysis. Int. J. Prod. Res. 48 (7), 2095–2107. Li, Q., Li, Z.B., Zhang, Q., 2011. Research of power transformer fault diagnosis system based on rough sets and Bayesian networks. Adv. Mater. Res. 320, 524–529. Li, Z.H., Liu, M.K., 2010. Review of intelligence fault diagnosis in power system. Electr. Eng. 8, 21–24. Lin, Y.X., Zeng, Q.J., 2009. Development of on-line monitoring system for high voltage vacuum circuit breaker. Electron. Des. Eng. 17 (12), 21–23. Ma, D.Y., Sun, C., Li, Z.H.X., Guan, R.C., Liang, Y.C., 2010. Uncertain management in rule-based power system using object-oriented design. In: Proceedings of the 3rd International Conference on Power Electronics and Intelligent Transporta- tion System (PEITS 2010), 363–366. Mizutani, Y., Takahashi, T., Ito, T., 2007. Lifetime evaluation method for pole transformer based on transient temperature analysis. IEEE Trans. Power Energy 127 (5), 653–658. Qu, Z.Y., Gao, Y.F., Nie, X., 2008. Network fault diagnoses expert system model based on decision tree. Comput. Eng. 34 (22), 215–217. Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1986. Learning representations by back-propagating errors. Nature 323 (6088), 533–536. Shi, Y., Qiu, T., Chen, B.Z., 2006. Fault analysis using process signed directed graph model. Chem. Ind. Eng. Prog. 25 (12), 1484–1488. Wang, J., Liu, J.X., Wu, L.Z., 2006. Fault forecast of the transformer using new theory of grey system. In: Proceedings of the International Conference on Electricity Distribution (CICED 2006), 20, 340–345. Wang, N., 2008. Study on Fault Diagnosis of High Voltage Circuit Breaker Based on Neural Network Expert System. Henan University 37–40. Yang, B.S., Jeong, S.K., Oh, Y.M., Tan, A.C., 2004. Case-based reasoning system with Petri nets for induction motor fault diagnosis. Expert Syst. Appl. 27 (2), 301–311. Zaki, O., Brown, K., Fletcher, J., Lane, D., 2007. Detecting faults in heterogeneous and dynamic systems using DSP and an agent-based architecture. Eng. Appl. Artif. Intell. 20 (8), 1112–1124. Zhang, L.L., Li, P., Wang, T.T., Wang, H., 2010. Research of fault diagnosis based on fuzzy neutral network to power transformer. Ind. Instrum. Autom. 5, 3–5. Zhu, Y.L., Huo, L.M., Lu, J.L., 2006. Bayesian networks-based approach for power systems fault diagnosis. IEEE Trans. Power Del. 21 (2), 634–639. Multi-BP expert system for fault diagnosis of power system Introduction Back propagation expert system Multi-BP expert system (MBPES) The structure of MBPES The construction algorithm of MBPES The training algorithm of MBPES Results and analysis Comparison for different network layer numbers Experiment on transformer data Analysis of vacuum circuit breaker experimental data Conclusions Acknowledgment References 
work_n743hftqorg4nglq67wrje4dre	----	The role of an extensionist in expert system applications: 1 Understanding Expert Systems Applications from a Knowledge Transfer Perspective Weizhe Feng ab, Yanqing Duan 1c , Zetian Fu ab , Brian Mathews c a International College at Beijing, and College of Information and Electrical Engineering China Agricultural University, Beijing, 100083, People‟s Republic of China b Key Laboratory of Modern Precision Agriculture System Integration, China Agricultural University, Beijing, 100083, People‟s Republic of China c University of Bedfordshire Business School, Luton, LU1 3JU, UK Abstract Expert systems were introduced more than two decades ago but their effectiveness and success are still in debate. This paper attempts to make a contribution to the better understanding of expert system applications from a knowledge transfer perspective. The paper argues that an expert system application is knowledge transfer using Information and Communication Technologies (ICT). Underpinned by knowledge transfer theories and through a series of empirical investigations of agriculture expert system projects, the study recognises the importance of human interactions in the expert systems implementation process. Based on the evidence collected, a number of key players are examined. These are; knowledge provider/sender, knowledge engineer, knowledge extensionist and knowledge recipient. The paper represents a first attempt to introduce and acknowledge the role of a knowledge extensionist in the ICT-based knowledge transfer process. The name “extensionist” is borrowed from previous literature and describes an actor whose role is an intermediary in supporting transferring knowledge towards the knowledge user. Findings demonstrate the significant contributions made by extensionists to the success of expert systems applications. It is argued that the rigidity and limitation of ICT based knowledge transfer can be significantly reduced with the involvement of close human interactions towards the knowledge recipient. Keywords: knowledge extensionist; expert system; knowledge transfer; agriculture; China. 1. Introduction 1 Corresponding author – Prof Yanqing Duan, University of Bedfordshire Business School, Luton LU1 3JU, UK, tel: (+)44 (0) 1582 743134, fax: (+)44 (0) 1582 743924, email: yanqing.duan@beds.ac.uk 2 The main function of an expert system is to mimic expertise and distribute expert knowledge to non-experts. The rapid development of Internet technology has changed the way that an expert system can be developed and distributed, but the distribution of expert systems to a large scale of end users can be challenging in terms of both effectiveness and efficiency. From a knowledge transfer perspective, it is argued in this paper that an expert system application is a knowledge transfer process in which expert knowledge is captured by a computer system (ICT) and delivered to non-expert recipient. It is believed that looking at expert system applications from a knowledge management point of view could benefit researchers and practitioners working in either the expert system or knowledge management domains. The traditional knowledge transfer approach is frequently criticised for its rigidity and bureaucracy because all knowledge transfer activities solely rely on the face-to-face communications from a centralised extension agency down to local recipients. After a few layers of people-to-people contact, knowledge can be easily lost or distorted. At the same time, knowledge transfer cannot achieve a great deal of efficiency by limited source when dealing with many users. “Sharing knowledge is power” (Liebowitz, 2001). This is especially true when agricultural expert systems are introduced to farmers in the developing world. In many developing countries the agriculture sector remains the largest employer in the country and agricultural productivity is one of the major concerns for the country‟s economy. Agricultural knowledge extension is seen as an effective solution for improving agricultural productivity. The word “extension,” in this context, derives from an education development in England during the 19th century when Oxford University and Cambridge University attempted to serve the rapid expansion of educational needs from society. It was called “university extension”. In the early 20th century, the word extension was applied to describe the transfer of knowledge and technology to serve the needs of rural development by American land-grant universities (Jones & Garforth, 1997). The rapid development of ICT, such as agricultural expert systems, brings new opportunities to the agricultural extension methodology (Rees, Momanyi & Wekundah, 2000). From a knowledge transfer perspective an agricultural expert system is a knowledge transfer medium, through which advanced knowledge is encoded and transferred to a recipient, who can learn and benefit from the knowledge transferred. 3 Expert system applications may involve a large number of users with diversified application scenarios. For example, a web-based expert system project in China can have potential impact on millions of farmers. Because of the large scale of the application, the application scenarios may vary in accordance with specific local farming conditions. The success of large scale expert system projects in the agriculture sector faces a number of challenges: 1. the large knowledge gap between the knowledge provider and the farmers, as sometimes, poorly educated farmers are not able to absorb the knowledge delivered to them nor follow the advice provided by the system; 2. the sheer number of users involved in using the expert systems; 3. physical distances between the knowledge provider/knowledge engineer and end users; 4. complex and diversified application contexts. Through empirical investigation into current ES applications in Chinese agriculture sector, it is evident that those challenges can be significantly reduced through the use of knowledge extensionists. This paper aims to analyse expert systems application cases, identify the roles of knowledge extensionists and report their contributions to the success of expert systems from a knowledge transfer perspective. 2. Literature Review An expert system is “a system that uses human knowledge captured in a computer to solve problems that ordinarily require human expertise” (Turban and Aronson, 2001). Expert systems are considered as a branch of artificial intelligence (AI) because the method of problem solving is predominantly based on heuristics (Darlington 2000). Durkin (1996) reports that many organisations have leveraged the technology to increase productivity and profits through better business decisions. Although there have been reports of ES failures (O‟Keefe & Rebne, 1993; Wong, 1996), research (Yoon et al, 1995; Kunnathur et al, 1996) shows that many companies have remained enthusiastic proponents of the technology and continue to develop important ES applications. The early applications of expert systems were standalone applications based on mainframe, AI workstation or PC platforms. The Internet offers an ever-expanding set of capabilities and a web based ES is capable of offering much more than traditional ES. Perhaps the most successful example is web-based legal ES reported by Bodine (2001), which enable law firms to collect hundreds of thousands of dollars in subscription fees from clients who use their advisory services. PT Consulting Partners in USA (2005) reported that they have helped its clients to build a number of successful web based ES which have brought significant benefits to the client 4 company. Grupe (2002) reported a web based ES called Student Advisor, which is an online ES helping students to select an academic major. From knowledge management point of view, an expert system application is a knowledge transfer activity using ICT. The objective of any knowledge transfer is to transfer the source knowledge successfully to the recipient (Cummings and Teng 2003). Brown and Duguid (1998) point out that knowledge transfer to a recipient who has a limited knowledge base will be difficult. Lin, Geng & Whinston (2005) develop a knowledge sender and recipient framework to classify knowledge transfer structures. They indicate that most literature implicitly assumes that knowledge transfer occurs under a symmetric complete structure. Within this structure both sender and recipient are underpinned by the same expertise and have the same knowledge capability. However, it is argued by the authors that, in practice, a sender-advantage asymmetric structure occurs commonly, within which the sender has a much better expertise and stronger knowledge capability, and so has an knowledge advantage over the receiver. The mainstream ICT based knowledge transfer literature is focused on the business sector and examines inter- and intra-organisational knowledge transfer facilitated by ICT (see e.g. Pawlowski & Robey 2004). With the knowledge base and knowledge capability at similar levels, inter- and intra-organisation knowledge transfer can be defined as a symmetric complete structure. In expert systems applications, the end users are normally non- experts who are seeking advice and help from domain experts. It is evident that end users expertise and knowledge capability are much lower than domain experts. With the classification given by Lin, Geng, & Whinston (2005), it would be appropriate to regard expert system applications as a sender-advantage asymmetric structure. Communication theory has been one of the fundamental bases in knowledge transfer research. It emphasises the critical players and their interactions during the knowledge transfer process (Szulanski, 1996). Based on communication theory, knowledge transfer is influenced by five basic elements: source, channel, message, recipient, and context (Szulanski, 2000). The process of knowledge transfer has been described as a two-stage process from the knowledge transmission by a sender to the absorption by a receiver by Garavelli, Gorgoglione, & Scozzi (2002), a multi-stage process by Santoro & Gopalakrishnan (2000), and multi-stage knowledge flow by Kim, Hwang, & Suh (2003). 5 The knowledge transfer intermediary has also drawn certain research attention and is described as an irreplaceable middleman of a transfer process for recombining past experience and leading towards new knowledge (Hargadon, 2002), or helping an organisation in procuring new knowledge (Hinloopen, 2004). Lack of intermediaries may be the one of the reasons leading to the failure of some ICT-based knowledge transfer projects (Matson et al. 2003). Pawlowski & Robey (2004) identify that IT professionals are committed to facilitating the knowledge flow across business boundaries for ICT based knowledge transfer. Meera, Jhamtani, & Rao (2004) also point out that local agricultural extensionists are important facilitators in farmers‟ access to computer applications for knowledge. There has been a different understanding in knowledge transfer texts concerning the capability of ICT in transferring knowledge. There seem few doubts about the ability and efficiency of ICT in transferring explicit knowledge (e.g. Cummings and Teng, 2003), but considerable Debates have been raised about the transferability of tacit knowledge via ICT because tacit knowledge cannot be articulated, codified and be easily understood (Nonaka, Reinmoeller & Senoo, 1998). A simple dichotomy of knowledge into the explicit and tacit can be somewhat misleading for the study of ICT based knowledge transfer for two reasons. Firstly, the boundary between explicit knowledge and tacit knowledge is porous, flexible (Spender 1996), and confusing (Clark, Carter & Szmigin, 2000). A vague distinction of explicit and tacit knowledge leaves much room for researchers to find themselves with contradictory viewpoints. Some literature attempts to avoid a distinction between the two in their arguments. For instance, Blumentritt & Johnston (1999) insist that the human intelligent system is the only repository of knowledge and only information can be transferred with ICT. Hislop (2002) points out that ICT is effective in facilitating explicit knowledge transfer but tacit knowledge can only be transferred by face-to-face communication. Bolisani & Scarso, (1999) argue that ICT can transfer both explicit and tacit knowledge. Secondly, being commonly defined by the school of management studies, explicit knowledge is challenged conceptually. Knowledge transfer does not mean that knowledge is moved as goods, instead, the recipient absorbs it by reconstructing their version of the new knowledge based on that transmitted (Sveiby, 1996). In a symmetric knowledge transfer structure, it is implicitly assume that electronically codified and transferred knowledge is explicit knowledge, as it will be understood and absorbed by a recipient. However, such implicit 6 assumption may be unable to explain knowledge transfer process in an asymmetric structure. Cowan, David, & Forey (2000) explains that the knowledge being “codified for one person or group may be tacit for another and an utterly impenetrable mystery for a third”. To avoid the conceptual confusion, this paper attempts to divide the knowledge into ICT transferable and ICT non-transferable knowledge, which is broadly in line with the arguments of Boutellier, Grassmann, & Macho (1998), Cowan, David & Forey (2000), and Cowan (2001). ICT transferable knowledge is the knowledge that can be electronically codified and transmitted by ICT. ICT non-transferable knowledge is the knowledge that can only be transferred by face-to-face communications. 3. Research Method To better understand the expert systems applications from a knowledge transfer process through expert systems, the key players involved, and the specific role of knowledge extensionists in the process, a case study approach was adopted. Field investigations and data collection were based on five expert system application projects in China‟s agriculture sector. The expert systems offer technological knowledge on advanced farming methods and diseases prevention and treatment. With state key research institutes or universities as the project initiators, five ES projects are delivered at either regional or national level. The profile of five projects are summarised in table 1. (table 1 is about here) The case study method was employed in this research because it allows researchers to retain the “holistic and meaningful characteristics and real-life events” (Yin, 2003). In-depth interviews were a major data collection method. A total of 65 individuals were from interviewed with between 10 and 15 participants from each of the five projects. The interviewees were selected based on the nature of their roles in the project. They include project manager, domain expert, knowledge engineer, computing system programmer, expert system applications facilitator (extensionists), and users (farmers). Some participants play dual roles in the project. To enhance the reliability of qualitative data, all in-depth interviews were either tape recorded or field notes taken immediately after the interview. 7 The interviews were not structured in nature and covered the following main questions: what is your role in the ES project? What is the knowledge source? Who are the recipients? What does a typical recipient like? Can you explain the ES application (knowledge transfer) process? What is your understanding of the role of an extensionist? What difficulties and problems you have experienced? What is your opinions on the effectiveness of the ES, why? etc. All interview data was finally transcribed into MS-Word. Triangulation of evidence was achieved by collecting project documents and conducting direct observations. In total 32 project briefings and reports were collected and examined. To analyse the qualitative data collected from interviews, NVivo, computer-assisted qualitative data analysis software, was used. With the assistance of NVivo, an analyst‟s physical tasks can be taken over by the software for marking sequences of text by coding, and retrieving together all sequences of coded text (Bryman, 2004). More conveniently, in qualitative analysis it can be relatively easier to visualise the data by using the Model Explorer facility in NVivo. Different icons and shapes are used as the representation of themes with the links to represent the relationships. Diagrams produced in the Model Explorer are a powerful way of exploring and analysing the data and auditing the development of ideas (Gibbs, 2002). 4. Findings and discussions The main purpose of case studies is to analyse the knowledge transfer process, the players involved and their roles in the transfer process. The following sections discuss the identified actors in knowledge transfer, the transfer process and the interaction among involving actors and the roles of the extensionist. 4.1. Expert system based knowledge transfer process Inspired by the communication theory and its emphasis on the critical players and their interactions during the knowledge transfer process, this paper focuses on examining the key actors involved in the expert systems applications and their interaction and contribution to the knowledge transfer. 8 The expert system based knowledge transfer process was examined with the empirical evidences collected from the case studies. Figure 1 illustrates the knowledge transfer process from senders to recipients with the intermediary actors between them, such as knowledge engineers and extensionists. Two types of knowledge flow are identified, namely as ICT-transferable knowledge and ICT non-transferrable knowledge. The interactions between actors, and an actor with ICT systems are also clearly visualised. This framework is partly in line with a knowledge transfer framework proposed by Argote & Ingram (2000). They suggest that knowledge is embedded and transferred within a network that consists of three repositories: people, tools and tasks. In this study, people refers to human expert, knowledge engineer, extensionist and farmers; tools refers to expert systems; and tasks refer to activities undertaken along the transferring process by various people, such as knowledge acquisition, codification, absorption, internalisation, etc.. In all of the cases investigated, it is evident that knowledge extensionists play very important roles in facilitating the knowledge transfer. The long distances and widely scattered project sites do not permit knowledge engineers to contact farmers face-to-face. The complex local agriculture environment also does not permit the domain expert and knowledge engineer to cover all possible application problems. Local extensionists are trained jointly by agriculture experts and knowledge engineers. During the training process, the different versions of expert systems are finalised with the joint efforts of local extensionist, knowledge engineer and agriculture experts in response to different local needs. In the meantime, the training programme provides opportunities for extensionists from different regions to share and exchange knowledge that may be beyond the competence of the knowledge base embedded in an expert system. (figure 1 is about here) 4.2. Actors involved in ICT based knowledge transfer 4.2.1. Knowledge provider (sender) The knowledge provider for the five expert systems is mainly human domain experts. Depending on the context of different projects, knowledge acquisition approaches vary. For example, in ES-a and ES-b which are the development of a fruit farming and disease prevention system, the research institutions has no first-class experts available within the organisation, therefore, external experts and text based publications are the main knowledge source. Published sources help the knowledge engineer not only to develop his/her personal knowledge in the 9 domain, but also to reduce reliance on the domain expert. In project ES-c, husbandry experts‟ contribution to knowledge is in an organisational setting. The project is a joint collaboration between an artificial intelligence research institute and an agriculture research institute. In project ES-d and ES-e, almost all knowledge was acquired from a group of aquiculture experts who were led by a chief expert. Thus, contrary to expectations, the field study finds that prior to the initiation of some specific projects a knowledge provider per se did not readily exist and had to be created from both individuals and published resources. 4.2.2. Knowledge receiver A knowledge recipient is the system user or the knowledge recipient at the end of the knowledge transfer chain. In this study, rural farmers are identified as the knowledge user. The majority of Chinese farmers can be characterised as poorly educated and computer illiterate hence a large gap exists between knowledge senders and recipients. Therefore, knowledge transfer effectiveness through ICT poses a significant challenge. Field investigations found that a large number of expert systems are installed in places accessible to farmers in each project location. Farmers report that an expert system enables them to access expertise and help them to improve agricultural production. In their words, computers bring them “experts at the farm gate”. 4.2.3. Knowledge Engineer A knowledge engineer plays critical roles in expert systems applications. A knowledge engineer is normally responsible for building the knowledge based system (Giarratano & Riley, 2002). A knowledge engineer carries out knowledge acquisition, knowledge representation and codification, and system programming. In the expert system projects investigated, the knowledge engineer actively worked with various domain experts for the knowledge acquisition and codification tasks. None of the knowledge engineering work was carried out by an agricultural expert. It is interesting to see that neither the agricultural expert nor the engineer are interested in playing the roles which they feel that they are not good at. 4.2.4. Knowledge Extensionist 10 This study represents a first attempt to introduce and acknowledge the unique role of an extensionist in the knowledge transfer in expert systems applications. In the early literature, the role of extensionist was reported to transfer the knowledge and skills in a social network originally, and more recently in the agriculture sector. As the demand for the intensive involvement of domain experts increases, such an institutional setting is mainly beneficial to developed countries that have sufficient experts in the agricultural extension system. With insufficient agriculture experts and scarce investment, the practices of agricultural extension in many developing countries experience a high rate of failure. The analysis of the expert systems applications shows that an extensionist is needed to facilitate the knowledge transfer activities and enhance the effectiveness of the applications. Various local organisations are found to have been directly involved in facilitating the application of agricultural expert systems in the local region. For example, the local applications of expert system in project ES-b, ES-c and ES-e indicate that expert system facilitators include various intermediary organisations at both private and public sector. In project ES-b, ES-c, and ES-e, it is found that knowledge engineers only have contacts with local extension agencies or other intermediary organisations. In the projects at the local level there is a co-location of knowledge engineer and user, e.g. project ES-a and ES- d, and relatively small number of potential users compared with those at the national level. In these cases the knowledge engineer sometime has to play a dual role as a knowledge engineer and extensionist. In both ES-a and ES-d, knowledge engineers travel frequently between the systems development institution and the farmer‟s village to help farmers in system installation and application. The experience gained and the information collected in our investigation demonstrates the need as well as the importance of the extensionist in facilitating knowledge transfer in expert systems applications. 4.3. Analysis of an extensionists‟ role 11 The investigation of the five projects finds that expert systems are being facilitated by extensionists from various organizations, such as:  Public extension agency, e.g. local agricultural extension station, agricultural science park;  Public science and education institution, e.g. agricultural polytechnic, agriculture college, agriculture research institute, rural evening schools;  Local farmer community, e.g. farmer‟s agriculture association;  Commercial organisations, e.g. agro-business companies; rural internet bars. Analysis of the case study materials and the interviews clear suggest that extentionists play four different roles. These are: a knowledge transfer intermediary; a knowledge sharing facilitator; a new knowledge creator; and a continuous expert system developer. 4.3.1. A knowledge transfer intermediary Extensionists help to transfer the knowledge and sometime provide support beyond the capacity of an expert system. At the early stage of the expert system project, it was assumed a tailor made expert system will fully satisfy farmer‟s knowledge needs. The preliminary assumption was drawn from many farmers‟ expectation that: “The computer [expert system] can transfer knowledge without a discount (quality reduction)” At a later stage, it was realised that a complete knowledge transfer sometimes can only be achieved when both ICT transferable and non-transferrable knowledge are delivered to knowledge users. In a sender-receiver asymmetric knowledge transfer structure, it is not a problem for ICT transferable knowledge being articulated, codified and transferred using ICT, but the local absorption of electronically transmitted knowledge sometimes may be difficult. Farmers may not be familiar with knowledge transfer process of expert system as pointed out by farmer Qin in ES-a: “The computer [expert system] does not speak in a way as a real expert does. Questions it asks cannot be answered by me, and vice versa. If I talk to a real expert, I can easily explain my questions to the person”. Farmers may also find difficult to understand the explanation provided in a scientifically rigorous approach by an ES. For example, Farmer Zheng in project ES-c said: “I am always puzzled with some suggests given by computer [expert system], for example, it says that my duck may have 75% percent possibility of having enteritis. So what shall I do with the 75% possibility? I really do not know”. 12 The investigation found that a small number of expert systems are installed in an individual farmers‟ site and used by them without the assistance of the extensionist, but the majority of expert systems are installed in a village office or a community centre and used by farmers with the assistance of an extensionist. Various local organisations assign such tasks to staff who possess certain levels of both domain and IT knowledge. Although the expert systems have been carefully adapted by extensionist to the local needs, many low educated farmers still find that it is difficult for them to understand the advice and instructions provided by an ES. Such difficulties can be better addressed by an extensionist with face-to-face communication. For example, Mr Li said: “Farmers collect their information through their eyes and ears. To farmers, a picture is more meaningful than one hundred words, and a story is more meaningful than ten theories. Believe it or not, a farmer may get lost in a series of IF-THEN instructions.” Another example is in ES-b, a series learning references paper, which is called locally “easy knowledge card”, is prepared by an extensionist for farmers. With the support of an expert system, selected knowledge items drawn from the expert system are reedited with the local language that farmers are more familiar with. In addition to textual explanation of those disease symbols, colourful pictures are added to illustrate the concept and descriptions. The case studies also found that even when farmers learned how to deal with the disease from an ES, some of them still liked to confirm what they have learned from a computer with a local extensionist. It seems that some farmers would only trust the knowledge provided by an ES if it is backed up with an extensionist. Knowledge transfer with ES can also be regarded as a learning process for farmers. As a result, learning is not a one off event but ongoing process. An extensionist is found important in facilitating a farmer‟s new knowledge application process. For example, farmer Liu in project ES-e said: “I am not a slow learner, but you know we have to learn by doing. I assume the artificial fertilisation process for trout is easy to learn with the help of an ES, but it is not. You always find something is not covered by the computer (ES) when you try to do it. Thanks to Lao Xie [a local extensionist] for his 24 hours cell phone consulting service and his visit to my pond. Now I have mastered the process successfully with his timely help.” 13 Regarding the ICT non-transferable knowledge, it is contextually related and difficult to be transferred using ICT. For example, a human veterinarian expert Guo in project ES-c stated: “When you diagnose a sick pig, you may comprehensively scan the circumstances it lives, its physical appearances, and its body movement. But to some extent, some facts only can be known when you are on the sites, e.g. smell, touching behaviour, etc.. To ensure the farmer to learn it accurately, we have to teach their local extensionist to learn it by direct observation and practice with my supervision. When they return to their villages, they may find appropriate cases and demonstrate such knowledge face to face to farmers” An extensionist Mr Gong in project ES-c told the researcher: “In our cow expert system, traditional animal acupuncture is demonstrated with some diagrams and pictures. But it is not easy as it is very experience based. Each time when farmers would like to try, I will teach them in person. In terms of finding the accurate acupoint, controlling the needle, perceiving the cow‟s reaction…it is learned only by doing with your close watch.” For more effective knowledge transfer, it can be argued that the use of an extensionist as a knowledge intermediary can enhance knowledge absorption and transfer to ensure completeness and effectiveness. 4.3.2. A knowledge sharing facilitator The second finding of this research is the acknowledgement and identification of the extensionist as a knowledge sharing facilitator for fostering local farmer‟s knowledge exchange activities in expert system applications. It is argued that knowledge on agriculture production is “locally and historically embedded and socially constructed” (Osbahr & Allan, 2003). With a dense population in the countryside, farmers are used to sharing knowledge via socially constructed knowledge sharing network. In project application sites, the expert system is introduced into such networks. Farmers are organised together with kinship members and friends for collective learning with the help of the local extensionist. In this case, an extensionist is identified with a role of knowledge sharing organiser and facilitator. For example, extensionist Xiao said: “An efficient way to teach farmers in a village to use the expert system is to teach some smart and capable farmers to use it firstly. Then few guys of the first group are invited by me to teach the second group. As being familiar with each other, farmers know the best way to get the message crossed and share their experience 14 with their kinships and friends and help each other to use the system.” This process is regarded as very useful as project manager Yang explained: “Some less educated farmers are incapable users in determine what knowledge is useful when mass electronic knowledge suddenly available in front of them. We must send them a guide [extensionist] for an easy navigation in their knowledge seeking. When teaching swimming you can‟t throw a beginner into an ocean alone in his beginning class, right? You know, our local extension station is like a safe swimming pool which has coaches (extensionists) who also organize them to learn from each other.” 4.3.3. A new knowledge creator The third finding of this research is the acknowledgement and identification of the extensionist as a new knowledge creator to update the knowledge base with new knowledge. In the study of knowledge management, knowledge transfer is commonly associated with knowledge creation (Nonaka, Toyama, & Konno, 2000). This study offers evidence to support such an argument. It is found that extensionists are able to enhance the knowledge base with new knowledge or best of practice discovered by either themselves or farmers. For example, in project ES-d and ES-e, newly introduced chemical medicines are suggested to farmers by the ES with clear instruction on the standard doses for the treatment. However, farmers sometime seem to use a higher dose than suggested for more significant effect. Extensionist Chen identified the risk of the application of the overdose and incorporated new knowledge he has collected for a better result of disease treatment. The new knowledge is about the use of a mixture of chemical medicines and traditional herbs. Green herbs are cheaply available and have less risk for product pollution. Similar examples are also demonstrated in other three projects. This kind of the knowledge creation by extensionist benefits the rural expert system users. Extensionists also facilitate new knowledge creation by reporting new problems and cases back to knowledge engineers, so they can consult knowledge providers and develop solutions to the new problems. For instance, extensionist Zheng explained the interaction between knowledge engineer and extensionist for fostering new knowledge creation: “Since some new farming or treatment techniques suggested by an ES can be expensive, farmers may be hesitated to adopt it. But some farmers applied a new method with a mixture of the newly introduced knowledge and the local indigenous one. Later, professors [knowledge engineers] became so interested in 15 such creativity in new methods that they asked me to collect it for them. When the system was upgraded he added this new knowledge item to the knowledge base. You see, professor and farmer jointly change a „dead‟ technology into a „live‟ one.” 4.3.4. A continuous system developer The fourth finding of this research is the acknowledgement and identification of the extensionist as a continuous expert system developer to update, modify and adapt the expert system to the specific local conditions and user needs. Field investigations showed that an extensionist plays an important role in helping the knowledge engineer to complete and validate the expert system for local applications. Local heterogeneity of agricultural systems and complexity of the biological systems are such that scientifically derived technologies cannot cope alone with the scale of the problem (Hall & Clark, 1995). What is of interest to the study is the approach of tailoring the expert systems toward an applicable one to meet the needs of diversified agricultural production. In field interviews, the words „second time system development‟ appear frequently and interchangeably with the words „system localisation‟, which refers to the continuous development of the expert system. Although modern technological knowledge has brought advanced farming methods, continuous improvement on the technological approaches and their applications may ensure their fulfilment of the requirement of local farming communities (Weiss, Crowder, & Bernardi, 2000). All of the expert systems application projects investigated demonstrate the necessity of localising the expert system for local needs. They also all undertake localisation tasks either by involving knowledge engineers or local agricultural experts before the final systems delivery, or by providing the training to the regional extensionists on how to modify the expert system toward a successful local application. For example, in ES-b, the knowledge engineer develops the standard expert systems for vegetables, fruits, crops, and husbandry production. Local extensionists are called in from various expert system facilitating organisations to attend the training programme delivered jointly by agricultural experts and project knowledge engineers. One of the major goals of the training programme is to teach all extensionists the skills of using the expert system development shell to modify the system for the specific local conditions. With the flexibility of the expert system shell, not only knowledge items, communication language and knowledge searching process can be modified, but also the knowledge base and system interface can be adjusted. Extensionists with stronger 16 computer skills and greater expertise can even be trained in developing a new expert system for a specific local application. 5. Conclusions Expert systems provide an opportunity for developing countries to transfer agricultural knowledge to rural farmers by the mass duplication of the valuable knowledge of domain experts. However, effective implementation of expert systems cannot by itself override the disadvantage of the rigidity and narrow focus in transferring knowledge. This paper makes a number of important contributions to a better understanding of expert systems applications: Firstly, from a knowledge management point of view, it points out that an expert systems application is a knowledge transfer process and attempts to look at ES applications using knowledge transfer theories and analysis. It argues that researchers and practitioners working in the ES domain can benefit from knowledge management theories and studies. Secondly, the paper reviews the characteristics of expert systems using the Lin et al (2005) sender-receiver framework for knowledge transfer and recognises that ES-based knowledge transfer has the sender-advantage asymmetric transfer structure. In a sender-advantage asymmetric structure there is normally a significant gap of expertise and knowledge capability between the knowledge provider and a recipient. Thirdly, through case studies, the paper reveals that when there is a huge knowledge gap between knowledge sender, i.e. experts, and knowledge receiver, i.e. farmers in this case, the effectiveness of knowledge transfer could be seriously jeopardised without adequate human intermediaries and social interactions. Findings confirm that in regard to transferring the remaining ICT non-transferable knowledge beyond the capacity of an ES, a face-to-face approach or rich contextual communication support is one way to reduce the gap. The paper argues that in ES applications, as a sender-advantage asymmetric knowledge transfer process, human actors and especially knowledge extensionists play an important role in facilitating both ICT-transferable and ICT-non transferable knowledge. 17 Through a series of case studies involving interviews with the main players in expert systems application projects, the research develops an expert system based knowledge transfer framework. In this framework, all key actors, their interactions with other human actors and the ICT system, and knowledge flows are clearly located and visualised. Extensionists make significant contributions to expert system effectiveness. According to the findings, a knowledge extensionist can act as:  A knowledge transfer intermediary to help reduce the knowledge gap and shorten the knowledge distance between the sender and the receiver.  A knowledge sharing facilitator to help knowledge receivers share and learn knowledge through their local social networks.  A new knowledge creator to enhance or extend the knowledge base with new knowledge.  A continuous expert system developer to help adapt the system to the local needs. Having formally recognised the role and the important contributions of the extensionists in ICT based knowledge transfer, issues related to the education and the training of the extensionists need to be addressed. Also, further research on the success factors affecting the expert system applications from knowledge users and extensionists‟ point of view should be examined. Regarding knowledge theories, this study finds limitation in current taxonomy of knowledge properties. By simply dividing knowledge into the ICT transferable and non-transferable knowledge, there can be conceptual conflicts or confusion caused with respect to another common knowledge dichotomy, i.e. explicit and tacit knowledge. Explicit knowledge, by its definition, is easily transferred by using ICT, but the ICT‟s transferability of tacit knowledge remains debatable. Therefore, further research on the theoretical development of knowledge classifications, which are suitable for ICT based knowledge transfer systems, should be carried out. References Argote, L., & Ingram P. (2000). Knowledge Transfer: A Basis for Competitive Advantage in Firms. Organizational Behavior and Human Decision Processes, 82(1), 150–169. Blumentritt, R., & Johnston, R. (1999). Towards a Strategy for Knowledge Management. Technology Analysis & Strategic Management, 11 (3), 287-300. 18 Bodine, L. (2001). Delivering Legal Services via Web-Based Expert Systems: Law Firms Find New Paths to Profit on the Web. The Sugarcrest Report, June 12 2001, issue 4, http://www.sugarcrest.com/newsletters/v1_issue4.htm, 30 January 2007. Bolisani E., & Scarso, E. (1999). Information technology management: a knowledge-based perspective. Technovation, 19, 209-217. Boutellier, R., Grassmann, O., & Macho, H. (1998). Management of dispersed product development teams: the role of information technologies. R&D Management, 28, 13–25. Brown, J., & Duguid, P. (1998). Organizing knowledge. California Management Review, 40 (3), 90–113. Bryman, A. (2004). Social research methods, 2nd ed. New York: Oxford University Press. Clark, P., Carter C., & Szmigin I. (2000). The spectrum of (Explicit) knowledge in firms and nations. Prometheus,18(4), 453-460. Cowan, R. (2001). Expert systems: aspects of and limitations to the codifiability of knowledge. Research Policy, 30, 1355–1372. Cowan, R., David, P., & Forey D. (2000). The explicit economics of knowledge codification and tacitness. Industrial and corporate change, 9 (2), 211-253. Cummings, J., &. Teng B-S. (2003). Transferring R&D knowledge: the key factors affecting knowledge transfer success. Journal of Engineering and Technology Management, 20, 39–68. Garavelli, C., Gorgoglione M., & Scozzi B. (2002). Managing knowledge transfer by knowledge technologies. Technovation, 22, 269–279. Giarratano, J., & Riley G. (2002). Expert systems: principles and programming (3rd ed.), Beijing: Thomson Asia and China Machine Press Gibbs, G. R. (2002). Qualitative data analysis: explorations with NVivo. Buckingham, Open University Press. Grupe, F. H. (2002). Student advisement: Applying a web-based expert system to the selection of an academic major. College Student Journal, Dec 2002. http://www.findarticles.com/p/articles/mi_m0FCR/is_4_36/ai_96619963, 30 January 2007. Hall, A., & Clark N. (1995). Coping with change, complexity and diversity in agriculture - The case of Rhizobium Inoculants in Thailand. World Development, 23(9), 1601-1614. Hargadon, A. (2002). Brokering knowledge, linking learning and innovation. Research in Organisational Behavior, 24, 41-85. Hinloopen, J. (2004) The market for knowledge brokers. Small Business Economics, 22, 407–415. Hislop, D. (2002). Mission impossible? Communicating and sharing knowledge via information technology. Journal of Information Technology, 17, 165–177. Jones, G. E., & Garforth C. (1997). Chapter 1. The history, development, and future of agricultural extension, in Swanson, B. E., Bentz, R. P., & Sofranko (Ed.) Improving Agricultural Extension. A Reference Manual. Rome: FAO. Kim, S., Hwang, H., & Suh, E. (2003). A Process-based Approach to Knowledge-Flow Analysis: A Case Study of a Manufacturing Firm. Knowledge and Process Management,10(4), 260-276. Kunnathur, A,S,, ahmed, M,U, & Charles, R,J,S, (1996). Expert systems adoption: an analytical study of managerial issues and concerns. Information & Management, 30(1), 15-25. Liebowitz, J. (2001). Knowledge management and its link to artificial intelligence. Expert Systems with Applications, 20, 1-6. Liebowitz, J. (1995). Expert systems: a short introduction. Expert Systems with Applications, 50 (5/6), 601-607. Lin L., Geng X., & Whinston A.B., (2005) A sender-receiver framework for knowledge transfer, MIS Quarterly, 29(2), 197-219. Matson, E., Patiath, P., & Shavers, T. (2003). Stimulating Knowledge Sharing: Strengthening Your Organization‟s Internal Knowledge Market. Organizational Dynamics, 32(3), 275–285. Meera, S. N., Jhamtani, A., & Rao D. (2004). Information and communication technology in agricultural development: A comparative analysis of three projects from India. ODI: Agricultural Research & Extension Network, Network Paper No. 135. http://www.odi.org.uk/agren/papers/agrenpaper_135.pdf, retrieved on Oct.25, 2004. Nonaka, I., Toyama, R., & Konno, N. (2000). SECI, Ba and Leadership: a Unified Model of Dynamic Knowledge Creation. Long Range Planning, 33, 5-34. Nonaka, I., Reinmoeller, P., & Senoo, D. (1998). The „ART‟ of Knowledge: Systems to Capitalize on Market Knowledge. European Management Journal, 16, 673-684. O‟Keefe, R.M. & Rebne, D. (1993) Understanding the applicability of expert systems. International Journal of Applied Expert Systems. 1(1), 3-24. Osbahr, H., & Allan, C. (2003). Indigenous knowledge of soil fertility management in southwest Niger. Geoderma, 111, 457–479. http://www.sugarcrest.com/newsletters/v1_issue4.htm http://www.findarticles.com/p/articles/mi_m0FCR/is_4_36/ai_96619963 http://www.odi.org.uk/agren/papers/agrenpaper_135.pdf 19 Pawlowski, S., & Robey D. (2004). Bridging user organizations: Knowledge brokering and the work of information technology professionals. MIS Quarterly, 28(4), 645-672. Rees, D., Momanyi, M., & Wekundah J. (2000). Agricultural knowledge and information system in Kenya- implications for technology dissemination and development. ODI: Agricultural Research & Extension Network, Network Paper No.107, http://www.odi.org.uk/agren/papers/agrepaper_107.pdf. Retrieved on 16, Jan., 2004. Santoro, M. D., & Gopalakrishnan, S. (2000). The institutionalization of knowledge transfer activities within industry–university collaborative ventures. Journal of Engineering Technology Management, 17, 299–319. Spender, J. C. (1996). Organizational knowledge learning and memory: three concepts in search of a theory. Journal of Organizational Change Management, 9 (1), 63-78. Sveiby K.E. (1996). Transfer of Knowledge and the Information Processing Professions. European Management Journal, 14(4), 379--388. Szulanski, G. (1996). Exploring internal stickiness: Impediments to the transfer of best practice within the firm. Strategic Management Journal, 17 (Winter special issue), 27-43. Szulanski, G. (2000). The process of knowledge transfer: A diachronic analysis of stickiness. Organizational Behavior and Human Decision Processes, 82(1), May, 9–27. Weiss, A., Crowder, L. V., & Bernardi, M. (2000). Communicating agrometeorological information to farming communities. Agricultural and Forest Meteorology, 103, 185–196. Wong, B. K. (1996). The role of top management in the development of expert systems. Journal of Systems Management 47(4), 36-40. Yin, R. (2003). Case study research design and method, 3rd ed. Thousand Oaks: Sage Publications. Yoon, Y, Guimaraes, T. and O‟Neal, Q. (1995). Exploring the factors associated with expert systems success. MIS Quarterly, 19(1), 83-106. http://www.odi.org.uk/agren/papers/agrepaper_107.pdf 20 Table 1 The profile of five agricultural expert system applications projects investigated in the study Project Application domain Development approach Project initiator and system developer Local facilitator Target user ES-a Web based crop and fruit farming expert systems for rural Beijing farmers at the local level. Knowledge engineer develops an expert system using an expert system shell-PAID 4.0. Domain knowledge are acquired and encoded into the system by the knowledge engineer. State Agricultural Information Engineering Research Centre in Beijing Agriculture and Forestry Research Institute. Expert system team members. Farmers in the rural areas of Beijing. Potential users are 800. ES-b A group of web based crop and fruit farming expert systems at the national level. Using the same expert system shell in ES-a. Domain knowledge are acquired and encoded by various local intermediary organisations for applications in different regions. State Agricultural Information Engineering Research Centre in Beijing Agriculture and Forestry Research Institute. Local agricultural extension stations, agricultural science parks, agricultural polytechnics, agro-business companies, local agricultural research institute, local farmer‟s associations, rural evening schools, rural internet bars. Farming workers of modern farming demonstration farms in 20 provinces in north, northeast, northwest and southwest. Potential users are about 7 million. ES-c A group of web based poultry and husbandry farming expert systems at the national level. Knowledge engineer develops an expert system using an expert system shell-DET. Domain knowledge are acquired and encoded into the system by local intermediary organisations for applications in different regions. A collaboration between a State Artificial Intelligent Machine Research Institute and Anhui Provincial Academy of Agricultural Science Local agricultural extension stations, and local agro-business companies. Farmers in a remote southwest province of Yunnan. Potential users are about 1 million. ES-d A comprehensive aquaculture disease diagnosis expert system called Fish-Expert at the local level. Knowledge engineer develops the complete expert systems without using expert system shell. Ministry‟s Modern Precision Agriculture System Integration in China Agricultural University in Beijing Expert system team members Large farms in rural counties of Beijing. Potential users are about 700. ES-e A comprehensive aquaculture disease diagnosis expert system further locally modified based on Fish-Expert at the local level. Using the expert system architecture of ES-d. Knowledge based is modified by local intermediary organisations for local applications. Ministry‟s Modern Precision Agriculture System Integration in China Agricultural University in Beijing A local agricultural college Large fishing farms in rural counties in Tianjin city. Potential users are about 100 thousand. 21 ICT-Base Knowledge Transfer Systems (e.g. expert system) Knowledge Engineer Knowledge Extensionist Knowledge Receiver Knowledge Provider Figure 1. ICT based knowledge transfer process ICT non-transferable knowledge flow ICT transferable knowledge flow 
work_n7qhzgftafabvhusi7cxovfv6y	----	PII: S0003-2670(00)83876-3 Analytica Chimica Acta, 210 (1988) 51-62 Elsevier Science Publishers B.V., Amsterdam - Printed in The Netherlands 51 EXPERT SYSTEM FOR INTERPRETATION OF THE INFRARED SPECTRA OF ENVIRONMENTAL MIXTURES LI-SHI YING and STEVEN P. LEVINE* School of Public Health, The University of Michigan, Ann Arbor, MI48109 (U.S.A.) STERLING A. TOMELLINI Department of Chemistry, University of New Hampshire, Durham, NH 03824 (U.S.A.) STEPHEN R. LOWRY Nicokt Instrument Corporation, 5225 Verona Road, Madison, WI 53711 (U.S.A.) (Received 1st September 1987) SUMMARY A program for the identification of the principal components of mixtures through interpretation of the infrared mixture spectrum (IntIRpret ) was developed. This program, which was developed as a preliminary screening tool for unknown organic mixtures, has five main subroutines: the interferogram processing and peak-selection subroutine (PUSHSUB), the automated knowledge- acquisition subroutine ( AUTOGEN ) , the system optimization subroutine (STO) , the interpre- tation subroutine (PAIRS), and fii processing subroutine to subtract spectral similarity (PAIRSPLUS) . Principal advantages of this system compared to earlier systems are speed, fIex- ibility and accuracy. In order to satisfy the requirements of hazardous waste analysis [l-6], a program for automated waste mixture identification (PAWMI) through the interpretation of the infrared (IR) spectrum of the waste mixture was devel- oped [ 7,8] and tested on samples from hazardous waste drums [ 9 1. Two lim- itations of PAWMI were that once a training set, consisting of a library of reference spectra, was defined, the rules for the inference engine (PAIRS) [lo-161 had to be generated manually. The second limitation was that the PAWMI software for compound identification only uses information on peak location. An approach to the automated generation of functional group interpretation rules for PAIRS was previously developed [ 161. This system defines an “oc- currence” value, which was used to weight information on peak position for the generation of expectation values for the presence of certain functional groups. 0003-2670/88/$03.50 0 1988 Elsevier Science Publishers B.V. 52 Efforts by other investigators have included the fuzzy data set [ 171, as well as hierarchical tree [ 181 and table-driven [ 191 programs developed by Munk and co-workers, and the pattern recognition approach of Frankel [ 201. Most of these systems were primarily aimed at identifying functional groups in com- pounds, as was the original PAIRS program. A related work aimed primarily at identifying compounds in mixtures was that of Lowry and Huppler [ 211, which used a search system based on Boolean logic. Many of these approaches owe their origins to earlier efforts by Jurs, Isenhour and co-workers [ 22-241. Recent publications have included improvements in the PAIRS and hierar- chical tree approaches, multi-spectroscopy expert systems, and various com- puter-aided spectral interpretation systems [ 25-271. In this paper, a program is described for the identification of the principal components of mixtures based on computer-assisted interpretation of the in- frared spectrum of the mixture. This program (IntIRpret) has five main sub- routines: the interferogram processing and peak-selection subroutine (PUSHSUB ) [ 81, the automated knowledge-acquisition subroutine [ 161 (AUTOGEN ), the system training and optimization subroutine (ST0 ) , the interpretation subroutine (PAIRS) [ 7, 10-161, and final processing subrou- tine to subtract spectral similarity (PAIRSPLUS) [8]. The principal advantages of this system compared to the previously reported PAWMI system are speed (all spectral information is encoded automatically), flexibility (changes in the data base and in interpretation rules are readily accommodated) and accuracy (interpretation is based on peak position, fre- quency of occurrence and peak size, each of which is weighted in an optimal fashion). The method was evaluated for the 62 most commonly identified organic compounds on hazardous waste sites [ 8,9]. IntIRpret was designed to be au- tomatic, self-training, and self-optimizing. EXPERIMENTAL All solvents were Aldrich Spectrophotometric Grade or equivalent. Mixtures were prepared on a weight basis. Film transmission spectra were acquired by placing a drop of sample between two KBr crystals. Spectra were acquired on a Nicolet 20-SX optical bench. Each spectrum was generated with back- ground- and sample-signal averaging over 128 scans. The number of data points collected was 16 384 resulting in a nominal spectral resolution of 2 cm-‘. All programming and spectral analysis, including rule writing, compiling and spec- tral interpretation were done with a Nicolet 1280 computer. RESULTS AND DISCUSSION IntIRpret has five main subroutines: the interferogram processing and peak- selection subroutine (PUSHSUB ) [ 81, the automated knowledge acquisition 53 subroutine [ 161 ( AUTOGEN), the system optimization subroutine (ST0 ), the inference engine (PAIRS) [ 7,10-16 J, and the final processing subroutine which subtracts spectral similarity (PAIRSPLUS) [ 8 1. Figure 1 is a flow chart of the IntIRpret process, where the logic of each of the five major subroutines is outlined. TRAINING SET OF PURE COMPOUNDS PUSnSUB 1 AUTOGEN [ PAIRS [ PAIRSPLUS [ lnterferogram Subtract low resolution from high resolution PAlRSdormat file Read the peak factors ,n,o PAIRS SPECTRUM OF MIXTURE OF UNKNOWN COMPOUNDS ldentlfy GOtTQOnentS Fig. 1. Flow chart of the IntIRpret process, showing the logic of each of the five major subroutines. 54 Here, emphasis is placed on describing STO, which is central to the opera- tion of the self-training, self-optimizing mode of operation of IntIRpret. PUSHSUB In order to automate PAWMI, a peak-selection subroutine PUSHSUB, was developed that does not require the operator to set a peak-selection threshold, and successfully follows nonlinear baselines [ 81. PUSHSUB selects peaks by transforming the first 256 data points right of the center-burst from the orig- inal 16 384-point sample interferogram into a threshold curve. PUSHSUB au- tomatically calculates the threshold value from this file. This has been described in detail [8]. PUSHSUB stores the peak file in a format that can be used by AUTOGEN and STO. ST0 This subroutine, the flow diagram of which is given in Fig. 2, accesses the peak tables generated by PUSHSUB. The peaks in a spectrum that are chosen for the purposes of decision-making are called rule peaks. Not all spectral peaks are rule peaks. Each rule peak is assigned a “goodness value” that indicates the probable presence or absence of each compound in the training set. The question of “goodness” has been discussed [ 7,8]. It is the purpose of the ST0 program to utilize the maximal amount of spectral information in an effort to enhance the predictive power of the goodness value. Three factors are used to weight the goodness values assigned to each rule peak listed by AUTOGEN: Kl (frequency of occurrence), K2 (intensity), and K3 (frequency of occurrence x intensity). These three factors are designed to follow the logic used by an expert during the interpretation of the infrared spectra of mixtures. In this respect, the underlying intellectual framework is similar to that described earlier [ 28, 291, in which match factors were auto- matically calculated for the interpretation of mass spectra. ST0 is structured around five subroutines, plus a “main”, or driver, pro- gram. SUB 1 reads the peak table for each compound that was generated by PUSHSUB. The peak table is compared to the operator-defined window-widths. If there is more than one peak in any given window, only the largest is retained. This results in the loss of potentially useful information, but it greatly simpli- fies later steps in the program with no apparent degradation of results. SUB 2 reads the peak tables of all spectra in the training set and creates an array consisting of peak position and intensity information. This is used for the cal- culations done in SUB 4 and SUB 5. SUB 3 decides which peaks in the peak table of each compound will be used for rule peaks for the PAIRS inference engine. This subroutine is designed to pick the largest peaks in the spectrum, up to a maximum of 20 peaks. If less than 20 peaks are present when the highest threshold value is used, the thresh- old is automatically lowered incrementally until 20 peaks are chosen. In some 55 SUB 1 I SUB 3 1 SUB 5 I MAIN choose Create largest c peak array peak In for each wlndow tralnlng set and pick peaks Count the number of peaks present for each rule peak in each window t Asaign welghtlng factors to all goodness values t Determine the peak lntenslty for each rule peak I SUB 2 SUB 4 Fig. 2. Flow chart of the system training and optimization (STO) process, showing the relation- ship between each of the five subroutines, and the main driver subroutine. cases, 20 peaks will not be present even at a low threshold, so the number of peaks necessary to satisfy this step of the program is lowered along with the threshold. If at least three peaks are not present at a threshold of 2 3% of the largest peak in the spectrum, then an error message is printed, and the spec- trum of that compound in the training set must be re-examined by the opera- 56 tor. If the criteria for numbers of rule peaks and threshold are satisfied, the rule peak array is created from the information in SUB 2. SUB 4 analyzes the frequency of occurrence and intensity of data in the spectral array created by SUB 2 and SUB 3. This procedure counts the number of peaks within the window width surrounding each rule peak. The default value of the window widths was set at If: 3,5, and 10 cm-‘, which compensates for peak shifts expected in condensed phase mixtures [ 7, 8, 10-171. For ex- ample, for a peak at 1036 cm-l in the spectrum of benzene, there are eleven other peaks for spectra in the training set within the tightest window of +3 cm-’ ,19 other peaks present within the 2 5 cm-l window, and 26 other peaks within the ? 10 cm-’ window. This information is utilized to assign weighted goodness values in SUB 5. A similar calculation is done for the peak-intensity parameter. All peaks within the preset windows are not only counted, but their intensities are summed. This information is also used in SUB 5. SUB 5 divides the total goodness between peaks and peak windows (Fig. 3 ) . The first division of goodness is between windows, with the default value set at 50% for the tightest window, and 30% and 20% for the remaining two in- creasingly wide windows. These default values can be changed by the operator, if so desired. Secondly, the factors Kl, K2 and K3 are defined by the program. The goodness available to each peak window is divided between Kl, K2 and K3, with the default value for the constants set equal. These default values can be changed by the operator. Data generated by SUB 5 is accessed by the MAIN or driver program, which both calls SUB l-5 in sequence, and then creates a file for storing the goodness value for each window, peak, and compound in the training set. This data is stored in a form that is usable by AUTOGEN. An example is the generation of the optimized goodness value for the rule peaks of benzene (Table 1) . The values of Kl, K2 and K3 are given for each of the peaks. For each compound, a total of 100 000 goodness units are allocated by STO. This is a change from the original PAIRS program in which 100 good- ness units were allocated to the spectrum of each pure compound. For benzene, the allocation is made by apportioning the goodness between six rule peaks. The peak at 674 cm-l is illustrative of the manner in which the system works. This peak is the largest in the spectrum of benzene, therefore the K2 value is the highest, with a value of 9821, 5892, and 3928, totalling 19 641 goodness units. The value 19 641 can be found in Table 1. These three values are for the + 3,5, and 10 cm-l windows, and represent an allocation of 50,30 and 20% of the total K2 goodness. The peak at 674 cm-l has 4,8, and 13 peaks in all of the other spectra of the compounds in the training set within 2 3,5, and 10 cm-’ windows. Thus, the peak is in a window in which the frequency of occurrence of potentially inter- fering peaks is low, and the Kl values are correspondingly high. These are set 57 I I I t- K1 1 INumber of Peaks in the rule peak wndow) ~’ x fNumbar ol Praks for the I 1 spectra Ir<~~n~r,q st’l w~lh!n each w~ndowl -’ , -- I KP l-4 lntenstty of r;~le peak, x lntenslty of that compounds rule .oea:is ) .- lntenslty of rule peak 1 lntenslty of all rule peaks tin that wlndow 1 AUTOGEN 1 m----_-L MAIN ] - Fig. 3. Flow chart of ST0 SUB 5, showing the relationship between Kl, K2 and K3. at 3450, 1991, and 1218, respectively, totalling 6659, which is the value given in Table 1. The total intensity, on a scale where the largest peak in a spectrum has an intensity of 99, of all other peaks in the spectra of the compounds in the train- ing set, is 177,335, and 502 for the three windows surrounding the 674 cm-’ peak. Thus, not only does this peak occur at a location where there are few other peaks in the spectra of other compounds in the training set, but those other peaks are relatively small. Therefore, the K3 values for this peak are set at the relatively high values of 4696, 3439, and 2547 for the three windows, totalling 10 682, which is the value found in Table 1. The total goodness assigned to the peak at 674 cm-’ is 36 986, or 37% of the 58 TABLE 1 Comparison of Kl, K2, K3 and goodness values for six peaks in the spectrum of benzene. Values associated with the peak at 674 cm-’ are discussed in the text Peak position (cm-‘) Relative intensity (O-99 ) Number of peaks in all spectra in +3 cm-’ F5 cm-’ f 10 cm-’ Kl (total for all 3 windows)” K2 (total for all 3 windows) b Total intensity of peaks in all spectra in f 3 cm-’ +5 cm-’ f 10 cm-’ K3 (total for all 3 windows)’ Kl+K2+K3for f 3 cm-’ f 5 cm-’ +lOcm-’ Totald 674 1036 1479 3036 3071 3091 99 9 28 18 5 9 4 11 6 5 6 3 8 19 10 8 10 8 13 26 14 14 16 10 6659 2701 5024 5882 4883 8175 19 641 1784 5554 3570 991 1784 177 203 228 40 24 15 335 254 368 80 39 91 502 515 446 170 110 103 10 682 1009 2726 7763 3828 7317 17 968 2519 6109 8324 4545 10 537 11324 1786 4145 5681 3383 3678 7694 1192 3053 3213 1775 3070 36 986 5497 13 307 17 218 9703 17 279 “Apportioned baaed on the reciprocal of the number of peaks in window in all .spectraX0.5,0.3, and 0.2 for the three window widths. b Apportioned based on 0.5,0.3, and 0.2 for the three window widths. “Apportioned based on the reciprocal of the total intensity of peaks in window in all spec- tra x 0.5,0.3 and 0.2 for the three window widths. d Divide by 1000 for % contribution. goodness for all of the peaks in the entire spectrum of six rule peaks. Goodness is divided into 18% for Kl, 53% for K2, and 29% for K3. As stated earlier, the default values chosen for this study were: window widths of ? 3,5, and 10 cm-‘; goodness values divided between these windows of 50%, 30%, and 20%, respectively; and Kl =K2 = K3. It is not known if these are the optimal values for this training set, for all possible mixtures that can be pre- pared for compounds in this training set, or for other training sets. 59 Using the ST0 portion of IntIRpret allows the optimization of goodness values for each rule peak in each training set. AUTOGEN The automated generation of rules for a defined training set is essential to the success of this approach. Without AUTOGEN, PAIRS and PAWMI are hampered by the potential for errors that always occurs when data is manually encoded, and by the constraints imposed by the length of time it takes to enter data for new or modified training sets. Because of these problems, such a sys- tem is inherently inflexible. AUTOGEN solves these problems. This subroutine has been modified from the program first reported [16]. The earlier version generated single level rules, plus a value for each functional group called “occurrence”. Occurrence was used to generate a “maximum ex- pectation value” which related the probability of the presence of a peak that was associated with a given functional group in a given wavenumber range. The present version generates a three-level filter algorithm. The intensity al- gorithm was also modified to generate information based on a scale of O-99, rather than the previously utilized O-9 scale. Completion of the running of AUTOGEN for a given training set generates a complete set of three-level “if-then” rules for the PAIRS inference engine. If ST0 had not been used, goodness values, which are a measure of the probabil- ity of the presence of an unknown compound in a mixture, would be assigned on an equal basis to each peak in each spectrum of the training set. The use of ST0 allows the optimized goodness values to be entered in the rules by AU- TOGEN for use by PAIRS. PAIRS As previously reported, PAIRS [lo-161 was modified in the PAWMI pro- gram [ 7,8]. The goodness scale ranges from 0.001 for a complete mismatch to 0.999 for a complete match. In the IntIRpret program, peaks in the library spectra are picked by PUSH- SUB, the goodness values are weighted by STO, and the three-level rules are written by AUTOGEN. A peak table is then created for the unknown mixture by PUSHSUB. PAIRS accesses that table and generates goodness values that indicate the probable presence of compounds in the mixture of unknowns. PAIRSPLUS PAIRSPLUS was developed to limit the effect of spectral similarity and has been described in detail [8]. Statistical studies have been conducted [ 71 to evaluate the quality of the final goodness value reported by PAIRS for actual compound assignment. Based on these results, a goodness value greater than 0.60 out of a possible 0.99 indicated the likely presence of the compound in the unknown spectrum. However, if many compounds in the training set are spec- 60 trally similar, then goodness values greater than 0.60 may be returned for these compounds as well. Because these compounds are not actually in the sample, but are predicted to be there, they are considered false positives. PAIRSPLUS accesses both the complete array of known spectra and the PAIRS interpretation results and subtracts the percentage of spectral similar- ity corresponding to the compound with the largest goodness value from all the goodness values of the remaining compounds. Note that this is a subtrac- tion of goodness values, not of actual spectra. A statistical check is then con- ducted on the remaining compounds to establish if another compound should be reported as present in the unknown sample. This is accomplished by cal- culating the mean and standard deviation of the remaining goodness values. If the next largest goodness value in the remaining spectra is greater than the 95% confidence interval, it is reported, the array is accessed, and the percent- age of spectral similarity corresponding to its goodness value is subtracted from the goodness values of all the remaining compounds. This is repeated until the statistical check establishes that there are no goodness values greater than the 95% confidence interval. At that point, the program terminates. In conclusion, results obtained through the use of PAWMI and IntIRpret are shown in Table 2 for the training set of the spectra of 62 compounds fre- quently found at hazardous waste sites [3] and 67 four-component mixtures of those compounds. As stated previously, the differences between PAWMI and IntIRpret are the subroutines ST0 and AUTOGEN, and a minor improve- ment in PAIRSPLUS. Thus, in PAWMI, rule peaks are operator chosen and TABLE 2 Results obtained with the PAIRS and PAIRSPLUS subroutines of PAWMI and IntIRpret. The training set consisted of 62 compounds frequently found at hazardous waste sites [ 3 1. The test mixtures consisted of 67 four-component mixtures of chlorobenzene, l,I,l-trichloroethane, tol- uene, and benzene PAIRSPLUS results With PAWMI” With IntIRpret Positives True False Improvement Negatives True False Improvement Total decisions 200 216 71 46 - 40% 3809 3840 68 52 - 24% 4154 4154 ‘These data do not match those previously reported [8] because the data set has been altered. 61 entered by hand into PAIRS using the subroutine, CONCISE. All peaks are weighted equally, and a three-level logic structure is used to compensate for shifts of peak positions from the spectrum of the pure compound to the spec- trum of the mixture. In IntIRpret, rule peaks are chosen by STO, and weighted for frequency of occurrence (Kl ) , intensity (K2 ) , and for the cross-term (K3). Rules are en- tered automatically by AUTOGEN and compiled into PAIRS. The software system is several orders of magnitude faster than when peaks are entered man- ually, is immune from mistakes made when complex data is entered manually, and is based on results that are consistently applied regardless of the operator or data set. These data show a 40% decrease in false positive results and a 24% decrease in false negative results when IntIRpret is compared to PAWMI. Some addi- tional improvements in results can be expected after completion of a study of the optimal values of window widths, window weighting factors and the rela- tive weights of Kl, K2 and K3. However, a certain degree of uncertainty will remain in the direct interpretation of the infrared spectra of mixtures because of peak shifts in solution, the similarity of the spectra of structurally similar compounds, and the inability of the peak-picking routines to recognize the presence of peaks that appear as unresolved shoulders or in poorly resolved envelopes. The authors thank Greg Kinnes for his help in preparing the mixtures and acquiring the IR spectra, and Mary Weed for preparation of manuscript fig- ures. This work was supported by grants l-ROl-OH02066-01 and OH02404- 01 from the National Institute for Occupational Safety and Health of Centers for Disease Control. REFERENCES M.A. Puskar, S.P. Levine and R. Turpin, in S.P. Levine and W.F. Martin (Eds.), Protecting Personnel at Hazardous Waste Sites, Butterworths/Ann Arbor, Woburn, MA, 1985, Chap. 6. D.F. Gurka, Project Summary: Interlaboratory Comparison Study: Methods for Volatile and Semivolatile Compounds, Environmental Monitoring Systems Laboratory, EPA-600/S4-84- 027, Las Vegas, NV, June, 1984. P .A. Hahstedt, M.A. Puskar and S.P. Levine, J. Hax. Waste Haz. Mater., 3 (2 ) (1986) 221. W.P. Eckel, D.P. Trees and S.P. Kovell, Distribution and Concentration of Chemicals and Toxic Materials Found at Hazardous Waste Dump Sites, Proc. Natl. Conf. Haz. Waste En- viron. Emergencies, Washington, DC, May, 1985. J.D. Mayhew, G.M. Sodaro and D.W. Carroll, A Hazardous Waste Site Management Plan, Chemical Manufacturers Association, Washington, DC, 1982. The Hazardous and Solid Waste Amendments of 1984, Congressional Record, H11103, Washington, DC, Oct., 1984. M.A. Puskar, S.P. Levine and S.R. Lowry, Anal. Chem., 58 (1986) 1156. M.A. Puskar, S.P. Levine and S.R. Lowry, Anal. Chem., 58 (1986) 1981. M.A. Puskar, S.P. Levine and S.R. Lowry, Environ. Sci. Technol., 21 (1987) 90. 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 H.B. Woodruff and M.E. Munk, J. Org. Chem., 42(10) (1977) 1761. H.B. Woodruff end M.E. Munk, Anal. Chim. Acta, 95 (1977) 13. H.B. Woodruff and G.M. Smith, Anal. Chem., 52 (1980) 2321. H.B. Woodruff and G.M. Smith, Anal. Chim. Acta, 133 (1981) 545. S.A. Tomellini, D.D. Saperstein, J.M. Stevenson, G.M. Smith and H.B. Woodruff, Anal. Chem., 53 (1981) 2367. S.A. Tomellini, J.M. Stevenson and H.B. Woodruff, Anal. Chem., 56 (1984) 67. S.A. Tomellini, R.A. Hartwick, J.M. Stevenson and H.B. Woodruff, Anal. Chim. Acta, 162 (1984) 227. T. Blaffert, Anal. Chim. Acta, 161 (1984) 135. J. Zupan and M.E. Munk, Anal. Chem., 57 (1985) 1609. M.O. Trulson and M.E. Munk, Anal. Chem., 55 (1983) 2137. D.S. Frankel, Anal. Chem., 56 (1984) 1011. S.R. Lowry and D.A. Huppler, Anal. Chem., 55 (1983) 1288. P.C. Jure and T.L. Isenhour, Applications of Pattern Recognition, Wiley, New York, 1975. G.T. Rasmussen, T.L. Isenhour, S.R. Lowry and G.L. Ritter, Anal. Chim. Acta, 103 (1978) 213. J.A. de Haseth, H.B. Woodruff, S.R. Lowry and T.L. Isenhour, Anal. Chim. Acta, 103 (1978) 109. D.D. Saperstein, Appi. Spectroec., 40(3) (1986) 344. J. Zupan (Ed. ) , Computer Supported Data Bases, Howard/Wiley, New York, 1986. P.C. Jurs, in P.C. Jurs (Ed.), Computer Software Applications in Chemistry, Wiley, New York, 1986, Chap. 16. K.-S. Kwok, R. Venkataragahaven and F.W. McLafferty, J. Am. Chem. Sot., 95 (1983) 4185. B.L. Atwater, D.B. Stauffer, F.W. McLafferty and D.W. Peterson, Anal. Chem., 57 (1985) 899. 
work_nacpmdwfr5a47ldi3g65kmssgu	----	00190.dvi April 12, 2008 15:19 00190 Journal of Information & Knowledge Management, Vol. 7, No. 1 (2008) 37–46 c© iKMS & World Scientific Publishing Co. Knowledge-Based Expert System Development and Validation with Petri Nets Madjid Tavana Management Information Systems Lindback Distinguished Chair of Information Systems La Salle University, Philadelphia, PA 19141, USA tavana@lasalle.edu http://lasalle.edu/∼tavana Abstract. Expert systems (ESs) are complex information systems that are expensive to build and difficult to validate. Numerous knowledge representation strategies such as rules, semantic networks, frames, objects and logical expressions are developed to provide high-level abstraction of a system. Rules are the most commonly used form of knowledge representation and they are derived from popular techniques such as deci- sion trees and decision tables. Despite their huge popularity, decision trees and decision tables are static and cannot model the dynamic requirements of a system. In this study, we pro- pose Petri Nets (PNs) for dynamic system representation and rule derivation. PNs with their graphical and precise nature and their firm mathematical foundation are especially useful for building ESs that exhibit a variety of situations, includ- ing: sequential execution, conflict, concurrency, synchronisa- tion, merging, confusion, or prioritisation. We demonstrate the application of our methodology in the design and development of a medical diagnostic expert system. Keywords: Petri-Net; expert system; knowledge representa- tion; rules; decision tree; decision table; knowledge map; pro- cess map. 1. Introduction Knowledge-based information systems are complex arte- facts that are expensive to build and difficult to validate, especially when the components of the system exhibit a variety of situations, including: sequential execution, con- flict, concurrency, synchronisation, merging, confusion, or prioritisation (Balduzzi et al., 2000; Mehrez et al., 1995). Sequential execution refers to the processing of precedence constraints; conflict refers to mutually exclusive activities or results; concurrency refers to simultaneous task opera- tion; synchronisation refers to multiple resource usage in a single operation; merging refers to multiple precedence constraints; confusion refers to the combination of conflict and concurrency; and prioritisation refers to the determi- nation of the priorities of activities. Lee et al. (2001) have argued the need for verifying and validating the reliability and quality of the data and processes for such complex systems. Numerous knowledge representation strategies such as rules, semantic networks, frames, objects and logical expressions are developed to provide high-level abstrac- tion of a system. Rules are the most commonly used form of knowledge representation and they are derived from popular techniques such as decision trees and deci- sion tables. The decision trees are widely used for build- ing knowledge-based systems by the inductive inference (Chen et al., 2003; Lee et al., 2007; Yang et al., 2005). Decision trees provide a useful solution for many prob- lems of classification where large datasets are used and the information contained is complex (Aitkenhead, 2008). Decision trees are easily converted into rules by deriving a rule for each path in the tree that starts at the root and ends at the leaf node (Janssensa et al., 2006). Similarly, decision tables represent an exhaustive set of mutually exclusive expressions that link conditions to particular actions. Decision tables and decision trees are practical tools that enhance clarity, conciseness and comprehensi- bility (Baesens et al., 2003; Yang et al., 2005). Knowl- edge and process maps are also used to analyse business problems in order to transfer certain aspects of knowledge into a transparent graphical form (Becerra-Fernandez and Sabherwal, 2001; Eppler, 2004; Handzic, 2003). Eppler (2001) has classified knowledge maps into five different categories, including knowledge source map, knowledge asset map, knowledge structure map, knowledge applica- tion map and knowledge development map. A knowledge map generally consists of a ground layer which represents the context for the mapping and the individual elements that are mapped within this specific context. Despite their huge popularity, decision trees, decision tables, knowl- edge and process maps are static and cannot model the dynamic requirements of a system. 37 April 12, 2008 15:19 00190 38 M. Tavana Petri Nets (PNs) are well-suited for the design, spec- ification and formal verification of complex information systems (Sakthivel and Tanniru, 1988–1989). PNs with their graphical and precise nature and their firm math- ematical foundation are commonly used to model many complex systems. Their graphical nature allows for mod- els that are easy to understand while their formal seman- tics allow for precise and unambiguous descriptions. PNs are particularly considered a richer, more versatile, and dynamic graphical tool in the development and vali- dation of expert systems (Mehrez et al., 1995; Wong, 2001). PNs were initially defined by Carl Adam Petri (1962) and later refined and named after him by Holt (1971). Peterson (1981) elegantly discusses the dynamic behaviour of PNs, while Murata’s tutorial review paper provides a thorough review of PN’s history and appli- cations (Murata, 1989). PNs are abstract, formal mod- els used to describe and analyse the flow of infor- mation and control in systems, particularly systems that exhibit asynchronous and concurrent activities, with conditions for the performance of events within the system (Bullers, 1991). They have been proven to be useful for the modelling and analysis of several classes of systems, including web-based systems (Huang et al., 2008; Zhovtobryukh, 2007), communication sys- tems (Berthelot and Terrat, 1982), knowledge-based sys- tems (Jantzen, 1980), simulation systems (Piera et al., 2004) and process control systems (Bruno and Marchetto, 1986). In addition to ordinary PNs, timed PNs, stochas- tic PNs, coloured PNs and fuzzy PNs are widely used to model engineering and business systems. Timed PNs are those with places or transitions that have time dura- tions in their activities (Liu et al., 2007). Stochastic PNs include the ability to model randomness in a sit- uation, and also allow for time as an element in the PN (Murata, 1989; Lee et al., 2001). Coloured PNs allow the user and developer to witness the changes in places and transitions through the application of colour-specific tokens, and movement through the sys- tem can be represented through the changes in colours (Chen et al., 2001). Fuzzy PNs are used to model fuzzy rule-based reasoning to handle uncertain and imprecise information (Huang et al., 2008; Li and Lara-Rosano, 2000). In the next section, we present PN principles and formalism followed by a discussion of ES modelling with PNs in Sec. 3. In Sec. 4, we use PN process modelling to develop a medical diagnostic expert system, and in Sec. 5, we present our concluding remarks and future research directions. 2. PN Principles and Formalism Classical PNs as defined by Petri (1962) and further discussed by Peterson (1981). They are identified by 5-tuples (P, T, I, O, Eµ), where P = {p1, p2, . . . , pm} is a set of places; T = {t1, t2, . . . , tn} is a set of transitions; [I ⊆ P xT ] is the input function from P to T ; [O ⊆ T xP ] is the output function from T to P ; and Eµ, called mark- ing, is a function that defines a mapping from a set of places P to Z (here Z denotes the set of all nonnegative integers). µ : P → Z where µi = µ(pi) ∈ Z, pi ∈ P i = {0, 1, . . . , m}. Mathematically, a PN is a directed bipartite graph with two different types of node called places and transi- tions. A place p is presented with a circle and a transition t is presented by a rectangle. The nodes are connected through directed arcs. Directed arcs from p to t create input places while directed arcs from t to p create out- put places. Input places are a set of places that can fire a transition, while output places are a set of places that are associated with the results (outputs) from a transition. Only the static properties of a system are presented by a PN structure; however, dynamic properties result from PN execution. Execution requires the use of tokens or markings (denoted by dots) associated with places. Each place contains zero or many tokens drawn as black dots. The execution of a PN may affect the number of tokens in a place. A transition is called enabled when each of its input places has enough tokens. A transition can be fired only if it is enabled. When a transition is fired, tokens from input places are used to produce tokens in output places. We will use the PN shown in Fig. 1 to illustrate the classical PN. With one token placed in p1 and two tokens placed in p2, transitions t2 and t3 are enabled. Firing t2 and t3 consumes three tokens (one from p1 and two from p2) and produces two tokens in p3. Now transition t1 is enabled. Firing t1 consumes one of the two tokens in p3 (leav- ing p3 with one token) and produces two tokens, one in t1 p1 t3 t2 p2 p3 Fig. 1. Simple Petri Net example. April 12, 2008 15:19 00190 Knowledge-Based Expert System Development and Validation with Petri Nets 39 p1 and one in p2. Mathematical properties of the PN shown in Fig. 1 are presented next. The PN in Fig. 1 is constructed with three places (P = {p1, p2, p3}) and three transitions (T = {t1, t2, t3}). The input and output mapping of this PN is given as: I(t1) = {p3} O(t1) = {p1, p2} I(t2) = {p1, p2} O(t2) = {p3} I(t3) = {p2} O(t3) = {p3}. Any PN can be specified in matrix form as a D− Matrix with m rows and n columns, where m is the num- ber of transitions and n is the number of places in the PN. For each position [i, j] in the matrix, a 1 is placed in the position if transition i receives input from place j. A 0 is placed if transition i does not receive input from place j: D− =   0 0 1 1 1 0 0 1 0   . Similarly a D+ Matrix with m rows and n columns can be constructed, where m is the number of transitions and n is the number of places in the PN. For each posi- tion [i, j] in the matrix, a 1 is placed in the position if transition i produces output to place j. A 0 is placed if transition i does produce output to place j: D+ =   1 1 0 0 0 1 0 0 1   . The Composite Change Matrix (Matrix D) can be computed by subtracting D− form D+: [D+] − [D−] = [D]   1 1 0 0 0 1 0 0 1   −   0 0 1 1 1 0 0 1 0   =   1 1 −1 −1 −1 1 0 −1 1   . In addition, a 1 × m matrix, representing the firing of the PN is constructed. In each position [1, j] place the number of times transition j is to fire. The Transition Matrix for our PN is Transition Matrix = [0 1 1] (t2 and t3 firing because of the tokens in p1 and p2). Finally, a 1 × n matrix is constructed showing the current marking of the PN. In each position [1, j], the identified number of tokens in position j are placed. The Marking Matrix for our PN is Marking Matrix = [1 2 0] (given one token in p1 and two tokens in p2). The marking of the PN after the transition specified in the Transition Matrix (next marking) can be found as: [Transition Matrix][D] + [Marking Matrix] = [Next Marking][0 1 1]   1 1 −1 −1 −1 1 0 −1 1   + [1 2 0] = [0 0 2]. From a modelling perspective, PNs can be char- acterised as a “conceptual model with analytical quali- ties”, where a “conceptual model” generally represents a graphical approach, while an “analytical model” expresses functional and mathematical relationships (Mehrez et al., 1995). As a hybrid modelling approach, PNs can dis- play several important properties, including the ability to model situations for simulation analysis and to describe system behaviour (Kim et al., 2001). These abilities are useful in developing and validating expert systems that are used to arrive at an “expert advice”. The user supplies a set of facts and the expert sys- tem returns advice. Expert systems consist of a knowledge base and an inference engine. The inference engine con- sults the knowledge base with the facts provided by the user to return expert advice. The building blocks of expert systems are rules. When the facts supplied by the user, comply with a particular set of rules, an expert advice is produced. The rules of an expert system are often used to form a tree structure, called a decision tree. The nodes in a decision tree describe states and branches which repre- sent the relationships between the states. 3. Expert Systems Modelling with PNs There are many frameworks, methods and tools to build and validate information systems. Some are more suited to transaction processing systems, while others are more effective when assisting in the design and validation of decision support and ES (Whitten et al., 2001). According to Van Hee et al. (1991), designing a classical information system means automating an existing, well-defined man- ual system (or set of processes). However, when designing a decision support or ES, it is unwise to automate the existing processes completely, since many of these pro- cesses rely on heuristics and are not easily replicated by a machine. Many knowledge-based systems have failed as a result of the designers’ inability to differentiate between those processes that should be automated and those that should remain manual (Van Hee et al., 1991). Problems can arise in both the design and validation phases of the development of a knowledge-management system (Whitten et al., 2001), and care must be taken to ensure that the systems’ processes and the knowledge April 12, 2008 15:19 00190 40 M. Tavana base data are correctly designed, accessed and manipu- lated. With these complex systems, prototypes are not always sufficient, since it can be difficult to simulate all the permutations of a decision process for all possible sit- uations under which an ES is expected to function (Van Hee et al., 1991). Therefore, Van Hee et al. (1991) and others (Wu and Lee, 1997; Bullers, 1991) advocate the development of a model of the target system that can reflect the performance characteristics (i.e., the effects of the decisions made by or through the ES) as well as the functionality of the system itself. PNs have been used to perform both the knowledge validation and functional- ity testing of many complex systems, and can be used to assist in designing and validating rule-based ESs, since they have been used to analyse conditional relationships (Lee and Lai, 2002; Liu et al., 1999; Liu et al., 2000; Zhu and Xudong, 2002). As detailed in Wu and Lee (1997) as well as in Bullers (1991), PNs can be used successfully to detect flaws in pro- cess as well as in the information passing from rule-based applications. PNs can address problems such as conser- vation of known and unknown facts and refraction that arise in the design and/or validation of an ES. The use of a PN model can serve as an inference model for rule- based ESs and as a platform for knowledge management and verification (Wu and Lee, 1997). According to Van der Aalst (1999), PNs are also use- ful for representing workflow processes because the PN model can be used to study the flow of control of and its relevant data. This representation enables designers to verify the correctness of the process and its manage- ment of information before the system becomes opera- tional (Verbeek et al., 2001). This benefit is an important consideration in choosing a model for the creation and validation of expert systems, since these systems generally involve complex processes and the management of consid- erable amounts of rule-based knowledge (Bullers, 1991). According to Lee et al. (1999), there are several rationales for basing the design and validation of an expert system on a PN model: • PNs achieve the structuring of knowledge within rule bases, which can express relationships among rules and help experts construct and modify rule bases, • The PN’s graphic nature provides visualisation of the dynamic behaviour of rule-based reasoning, • The PN’s analytic capability provides a basis for devel- oping a knowledge verification technique, and • Concurrency among rules can be modelled by PNs, which is an important aspect in real-time performance validation. 4. Application Problem In this section, we use PN process modelling to develop a medical diagnostic expert system at Kennedy Memo- rial Hospital.∗ The number of incorrectly diagnosed acute appendicitis cases at the emergency room (ER) at Kennedy Memorial Hospital had increased steadily over a three-year period. Several patients were misdiagnosed and sent home to develop ruptured appendicitis while some patients were incorrectly sent to the operating room (OR) for emergency appendectomy. The ER physicians and staff agreed to implement a strict system of diagnos- tic protocols for patients with right lower quadrant (RLQ) pain. After several meetings, the ER physicians and staff were able to reach consensus on the following protocols for patients with RLQ pain. If a female patient complains about abdominal pain, her temperature should be taken. Next, the patient should be examined for RLQ pain. Those with no fever and no RLQ pain, should receive a blood test to determine if their white blood cell (WBC) count is elevated or not. An elevated WBC count is suggestive of acute appendici- tis. If the patient’s WBC count is normal, she will be discharged home with instruction to return to the ER if any fever, worsening abdominal pain, or vomiting devel- ops. If the WBC count is elevated, the patient will be admitted to the hospital and receive an emergency CAT (Computerised Axial Tomography) scan of the abdomen and pelvis. If the CAT scan is positive for appendicitis, she will be taken to the OR for appendectomy. If the CAT scan is negative, she will be observed in the hospital for the next 24 h to determine the course of her disease. Patients with no fever and RLQ pain must be sent for WBC count. If the WBC count is normal, the patient will be scheduled for an outpatient ultrasound within 24 hours to rule out any other significant female pathology. If the WBC count is high, the patient will be scheduled for emergency CAT scan. If the CAT scan is positive for appendicitis, she will be taken to the OR for appendectomy. If the CAT is neg- ative, she will be observed in the hospital for the next 24 hours to determine the course of her disease. Female patients with fever are examined for RLQ pain followed by a blood test for WBC count. Those with fever, RLQ pain, and a high WBC count, will be taken to the OR for emergency appendectomy. Those with a nor- mal WBC count are sent for a CAT scan. If the CAT scan is positive for appendicitis, the patient will be taken to the OR for an appendectomy. If the CAT scan is negative, the patient will be observed in the hospital for the next 24 hours to determine the course of her disease. If the patient has fever, with no RLQ pain and a normal WBC count, ∗The hospital name has been changed to protect the anonymity of all concerned. April 12, 2008 15:19 00190 Knowledge-Based Expert System Development and Validation with Petri Nets 41 she will be discharged home with instructions to return to the ER if any fever, worsening abdominal pain, or vomit- ing develops. Those with a high WBC count are sent for a CAT scan. If the CAT scan is positive for appendicitis, the patient will be taken to the OR for appendectomy. If the CAT is negative, the patient is observed in the hospital for the next 24 hours to determine the course of her disease. For male patients, the temperature should be taken to determine if the patient has a fever or not. The abdomen should be examined in all patients for RLQ pain. Male patients with no fever and no RLQ pain are dis- charged with instructions to return to the ER if any fever, worsening abdominal pain, or vomiting develops. Those with no fever and RLQ pain are sent for an emergency CAT scan. If the CAT scan is positive for acute appen- dicitis, the patient will be taken to the OR for emergency appendectomy. If the CAT scan is negative, a WBC count will be ordered. If the count is high, the patient will be observed in the hospital for the next 24 hours to deter- mine the course of his disease. If the count is normal, he should be discharged with instructions to return to the ER if any fever, worsening abdominal pain, or vom- iting develops. Male patients with fever and RLQ pain are taken to the OR for emergency appendectomy. Those with fever and no RLQ pain are observed in the hospital for the next 24 hours to determine the course of their dis- ease. Figures 2(a) and 2(b) present the ER screening PN diagram and notations. The PN presented in Fig. 2(a) is a complex network with 31 places and 38 transitions. In this problem, it is very common for the ER staff to work on more than one patient or more than one task at a time. Concur- rent process modelling capabilities of PNs are useful in modelling this requirement. Furthermore, the arrival of patients to the system is unplanned and their condition is unknown. This requirement can be captured with the t1 p1 t3 t2 t4 t8 t5 t7 t9 t6 t11 t12 p2 p3 p4 p5 p8 p6p7 p9 p11 p12 p13 p10 p14 p15 p16 t13 t10 t14 t15 t16 t17 t18 t19 t20 t21 t22 p23 p22 p21 p20 p19 p18 p17 t27 t26 t25 t24 t23 p25 p24 p27 p26 t34 t33 t32 t31 t30 t29 t28 p31 p30 p29 p28 t35 t37 t38 t36 (a) Fig. 2. Emergency room screening (a) Petri Net diagram. April 12, 2008 15:19 00190 42 M. Tavana PLACES p1: Idle Screener p2: Male Signal p3: Female Signal p4: Male Patient p5: Female Patient p6: Fever Signal p7: No Fever Signal p8: Male with Fever p9: Female with Fever p10: Male with No Fever p11: Female with No Fever p12: Right Lower Quadrant (RLQ) Pain p13: No Right Lower Quadrant (RLQ) Pain p14: Male with No Fever and RLQ Pain p15: Female with Fever and RLQ Pain p16: Female with Fever and No RLQ Pain p17: Female with No Fever and RLQ Pain p18: Female with No Fever and No RLQ Pain p19: Elevated White Blood Cell (WBC) p20: Normal White Blood Cell (WBC) p21: Emergency Appendectomy p22: Discharge p23: Admit p24: Ultrasound p25: Positive CAT Scan Signal p26: Negative CAT Scan Signal p27: Male with No Fever and RLQ Pain and Negative CAT Scan p28: Female with No Fever and No RLQ Pain and Elevated WBC p29: Female with No Fever and RLQ Pain and Elevated WBC p30: Female with Fever and No RLQ Pain and Elevated WBC P31: Female with Fever and RLQ Pain and Normal WBC TRANSITIONS t1: Male Sex Identification t2: Female Sex Identification t3: Male Temperature Measurement t4: Female Temperature Measurement t5: Male No Fever Determination t6: Female No Fever Determination t7: Male with No Fever and No RLQ Pain Determination t8: Male with No Fever and RLQ Pain Determination t9: Male with Fever and No RLQ Pain Determination t10: Male with Fever and RLQ Pain Determination t11: Female with Fever and RLQ Pain Determination t12: Female with Fever and No RLQ Pain Determination t13: Female with No Fever and RLQ Pain Determination t14: Female with No Fever and No RLQ Pain Determination t15: Male with No Fever and No RLQ Pain Positive CAT Scan Determination t16: Male with No Fever and RLQ Pain Negative CAT Scan Determination t17: Female with Fever and RLQ Pain Elevated WBC Determination t18: Female with Fever and RLQ Pain Normal WBC Determination t19: Female with Fever and No RLQ Pain Elevated WBC Determination t20: Female with Fever and No RLQ Pain Normal WBC Determination t21: Female with No Fever and RLQ Pain Elevated WBC Determination t22: Female with No Fever and RLQ Pain Normal WBC Determination 23: Female with No Fever and No RLQ Pain Elevated WBC Determination t24: Female with No Fever and No RLQ Pain Normal WBC Determination t25: Male with No Fever and RLQ Pain and Negative CAT Scan Elevated WBC Determination t26: Male with No Fever and RLQ Pain and Negative CAT Scan Normal WBC Determination t27: Female with No Fever and No RLQ Pain and Elevated WBC Positive CAT Scan Determination t28: Female with No Fever and No RLQ Pain and Elevated WBC Negative CAT Scan Determination t29: Female with No Fever and RLQ Pain and Elevated WBC Positive CAT Scan Determination t30: Female with No Fever and RLQ Pain and Elevated WBC Negative CAT Scan Determination t31: Female with Fever and No RLQ Pain and Elevated WBC Positive CAT Scan Determination t32: Female with Fever and No RLQ Pain and Elevated WBC Negative CAT Scan Determination t33: Female with Fever and RLQ Pain and Normal WBC Positive CAT Scan Determination t34: Female with Fever and RLQ Pain and Normal WBC Negative CAT Scan Determination t35: Return from Emergency Appendectomy t36: Return from Discharge t37: Return from Admit t38: Return from Ultrasound (b) Fig. 2. (Continued ) stochastic modelling capabilities of PNs. Finally, some patients might be in a more critical state than others and require immediate treatment. This requirement can be captured by establishing priorities in the PN. Figure 3, shows the rules derived from the ER PN. In order to validate our expert system, a pilot study was designed. Seven doctors, including the chief of the ER, two staff specialists and four ER doctors partici- pated in the pilot study. Multiple teams with at least five ER and specialist doctors in each team were formed to validate the knowledge base. The dynamic properties of this PN was examined and verified with the partic- ipants in this study through execution which required the use of tokens associated with places. Tokens in dif- ferent places enable transitions. The firing of the transi- tions produces tokens which again enable transitions for firing. Repeated examinations of this dynamic behaviour showed that the PN properly exhibits the current diag- nosis protocols in the ER at the Kennedy Memorial Hospital. A total of 814 cases were used in a pre- and post-ES implementation study to develop a prototype for patients arriving at the Kennedy Memorial Hospital with severe abdominal pain. The results are presented in Table 1. As is shown in the pre-ES statistics portion of Table 1, a total of 543 patients (45.25 patients on average per month) arrived into the ER with severe abdomi- nal pain over a 12-month period. 245 patients (20.42 patients on average per month) were referred to a special- ist after the initial screening in the ER and the remain- ing 298 patients were either discharged or treated for a non-appendicitis related problem. 137 patients (11.42 patients on average per month or 55.92%) were diagnosed with appendicitis by the specialists and the remaining 108 patients (9.00 patients on average per month or 44.08%) were diagnosed with no appendicitis and were either dis- charged or treated for a non-appendicitis related prob- lem. Further follow-up observations showed that 28 (2.33 patients on average per month) of the 108 patients who were either discharged or treated for a non-appendicitis related problem, were treated for appendicitis within seven days of their initial diagnosis. As is shown in the post-ES statistics portion of Table 1, a total of 271 patients (45.17 patients on average April 12, 2008 15:19 00190 Knowledge-Based Expert System Development and Validation with Petri Nets 43 Q U A L IF IE R S : 1 S e x Id e n ti fi c a ti o n : (F e m a le /M a le ) 2 T e m p e ra tu re M e a s u re m e n t: ( F e v e r/ N o F e v e r) 3 R L Q P a in D e te rm in a ti o n : (R L Q P a in /N o R L Q P a in ) 4 W B C L e v e l D e te rm in a ti o n : (E le v a te d /N o rm a l) 5 C A T S c a n D e te rm in a ti o n : (P o s it iv e /N e g a ti v e ) C H O IC E S : 1 E m e rg e n c y A p p e n d e c to m y 2 D is c h a rg e 3 A d m it 4 U lt ra s o u n d R U L E S : R U L E N U M B E R : 1 ( T 1 :1 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : E le v a te d T H E N : E m e rg e n c y A p p e n d e c to m y - C o n fi d e n c e = 1 R U L E N U M B E R : 2 ( T 1 :2 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : N o rm a l a n d C A T S c a n D e te rm in a ti o n : P o s it iv e T H E N : E m e rg e n c y A p p e n d e c to m y - C o n fi d e n c e = 1 R U L E N U M B E R : 3 ( T 1 :3 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : N o rm a l a n d C A T S c a n D e te rm in a ti o n : N e g a ti v e T H E N : A d m it - C o n fi d e n c e = 1 R U L E N U M B E R : 4 ( T 1 :4 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: F e v e r a n d R L Q P a in D e te rm in a ti o n : N o R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : E le v a te d a n d C A T S c a n D e te rm in a ti o n : P o s it iv e T H E N : E m e rg e n c y A p p e n d e c to m y - C o n fi d e n c e = 1 R U L E N U M B E R : 5 ( T 1 :5 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: F e v e r a n d R L Q P a in D e te rm in a ti o n : N o R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : E le v a te d a n d C A T S c a n D e te rm in a ti o n : N e g a ti v e T H E N : A d m it - C o n fi d e n c e = 1 R U L E N U M B E R : 6 ( T 1 :6 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: F e v e r a n d R L Q P a in D e te rm in a ti o n : N o R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : N o rm a l T H E N : D is c h a rg e - C o n fi d e n c e = 1 R U L E N U M B E R : 7 ( T 1 :7 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : E le v a te d a n d C A T S c a n D e te rm in a ti o n : P o s it iv e T H E N : E m e rg e n c y A p p e n d e c to m y - C o n fi d e n c e = 1 R U L E N U M B E R : 8 ( T 1 :8 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : E le v a te d a n d C A T S c a n D e te rm in a ti o n : N e g a ti v e T H E N : A d m it - C o n fi d e n c e = 1 R U L E N U M B E R : 9 ( T 1 :9 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : N o rm a l T H E N : U lt ra s o u n d - C o n fi d e n c e = 1 R U L E N U M B E R : 1 0 ( T 1 :1 0 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : N o R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : E le v a te d a n d C A T S c a n D e te rm in a ti o n : P o s it iv e T H E N : E m e rg e n c y A p p e n d e c to m y - C o n fi d e n c e = 1 R U L E N U M B E R : 1 1 ( T 1 :1 1 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : N o R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : E le v a te d a n d C A T S c a n D e te rm in a ti o n : N e g a ti v e T H E N : A d m it - C o n fi d e n c e = 1 R U L E N U M B E R : 1 2 ( T 1 :1 2 ) IF : S e x Id e n ti fi c a ti o n : F e m a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : N o R L Q P a in a n d W B C L e v e l D e te rm in a ti o n : N o rm a l T H E N : D is c h a rg e - C o n fi d e n c e = 1 R U L E N U M B E R : 1 3 ( T 1 :1 3 ) IF : S e x Id e n ti fi c a ti o n : M a le a n d T e m p e ra tu re M e a s u re m e n t: F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in T H E N : E m e rg e n c y A p p e n d e c to m y - C o n fi d e n c e = 1 R U L E N U M B E R : 1 4 ( T 1 :1 4 ) IF : S e x Id e n ti fi c a ti o n : M a le a n d T e m p e ra tu re M e a s u re m e n t: F e v e r a n d R L Q P a in D e te rm in a ti o n : N o R L Q P a in T H E N : A d m it - C o n fi d e n c e = 1 R U L E N U M B E R : 1 5 ( T 1 :1 5 ) IF : S e x Id e n ti fi c a ti o n : M a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in a n d C A T S c a n D e te rm in a ti o n : P o s it iv e T H E N : E m e rg e n c y A p p e n d e c to m y - C o n fi d e n c e = 1 R U L E N U M B E R : 1 6 ( T 1 :1 6 ) IF : S e x Id e n ti fi c a ti o n : M a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in a n d C A T S c a n D e te rm in a ti o n : N e g a ti v e a n d W B C L e v e l D e te rm in a ti o n : E le v a te d T H E N : A d m it - C o n fi d e n c e = 1 R U L E N U M B E R : 1 7 ( T 1 :1 7 ) IF : S e x Id e n ti fi c a ti o n : M a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : R L Q P a in a n d C A T S c a n D e te rm in a ti o n : N e g a ti v e a n d W B C L e v e l D e te rm in a ti o n : N o rm a l T H E N : D is c h a rg e - C o n fi d e n c e = 1 R U L E N U M B E R : 1 8 ( T 1 :1 8 ) IF : S e x Id e n ti fi c a ti o n : M a le a n d T e m p e ra tu re M e a s u re m e n t: N o F e v e r a n d R L Q P a in D e te rm in a ti o n : N o R L Q P a in T H E N : D is c h a rg e - C o n fi d e n c e = 1 F ig . 3 . R u le d e ri v e d fr o m th e e m e rg e n c y ro o m P e tr i N e t. April 12, 2008 15:19 00190 44 M. Tavana Table 1. Pre- and post-pilot study statistics. Month Total patients ER diagnosis: Specialist initial Specialist initial Specialist secondary with severe referral to a diagnosis: diagnosis: diagnosis: abdominal pain specialist appendicitis no appendicitis appendicitis Pre expert system statistics 1 48 21 12 9 3 2 42 19 13 6 2 3 52 23 11 12 3 4 45 18 10 8 3 5 39 18 9 9 2 6 47 20 11 9 2 7 51 23 13 10 2 8 45 22 12 10 3 9 44 20 11 9 2 10 39 18 10 8 2 11 43 21 11 10 3 12 48 22 14 8 1 Total 543 245 137 108 28 Pre-mean 45.25 20.42 (100%) 11.42 (55.92%) 9.00 (44.08%) 2.33 Post expert system statistics 13 52 16 15 1 0 14 38 14 12 2 1 15 48 15 14 1 0 16 49 14 13 1 1 17 39 14 14 0 0 18 45 15 14 1 0 Total 271 88 82 6 2 Post-mean 45.17 14.67 (100%) 13.67 (93.18%) 1.00 (6.82%) 0.33 Pre-post expert system statistics Mean difference −0.08 −5.75 2.25 −8.00 −2.00 t statistics 0.0356 7.0769 −3.3866 12.5514 6.5320 Significant (α = 0.05) No Yes No Yes Yes per month) arrived into the ER with severe abdominal pain over a 6 month period. The ES developed in this study was used for screening patients with severe abdomi- nal pain. 88 patients (14.67 patients on overage per month) were referred to a specialist after the initial screening in the ER and the remaining 183 patients were either discharged or treated for a non-appendicitis related problem. 82 patients (13.67 patients on average per month or 93.18%) were diagnosed with appendicitis by the specialists and the remaining 6 patients (1.00 patients on average per month or 6.82%) were diagnosed with no appendicitis and were either discharged or treated for a non-appendicitis related problem. Further follow-up observations showed that 2 (0.33 patients on average per month) of the 6 patients who were either discharged or treated for a non-appendicitis related problem, were treated for appendicitis within seven days of their initial diagnosis. The average number of patients arrived into the ER room pre ES implementation was 45.25 patients versus 45.17 patients for post-ES implementation. Test of means between pre- and post-ES implementation shows that the 0.08 difference in the average number of patients arrived into the ER with severe abdominal pain is not signifi- cant (α = 0.05). In addition, the 2.25 difference between the number of patients diagnosed with appendicitis in pre- and post-ES implementation studies is not signifi- cant either. However, further test of means shows that the average number of patients referred to a specialist was dropped from 20.42 to 14.67. This 5.75 patient or a 28% reduction in the number of patients referred to a spe- cialist after the implementation of our ES is statistically significant. Furthermore, the average number of patients initially diagnosed with no appendicitis by the special- ist dropped from 9.0 patients per month to 1.0 patient per month after the implementation of our ES. This 8.00 patient or an 89% reduction in the number of patients initially diagnosed with no appendicitis by the special- ist is statistically significant. Finally, the difference in the average number of patients who were either discharged or treated for a non-appendicitis related problem but upon April 12, 2008 15:19 00190 Knowledge-Based Expert System Development and Validation with Petri Nets 45 return within seven days of their initial diagnosis were treated for appendicitis dropped by 2 patients per month. This 86% reduction in the number of patients is statisti- cally significant. In summary, this study revealed two major findings. First, the implementation of our ES resulted in significant drop in “wrong” referrals as evidenced by significant decrease in average number of referrals by 5.75, as well as increase and decrease in specialists’ percentage of initial diagnoses to 93.18% and 6.82%. Second, there was no sig- nificant increase in specialists’ diagnosis accuracy as evi- denced by similar rations (28/108 = 26% and 2/6 = 33%). 5. Conclusion and Future Research Directions ESs are complex artefacts that are difficult to design develop and validate. Various knowledge representation techniques such as decision trees and decision tables are commonly used to develop rule-based ESs. However, these static approaches cannot replicate the dynamic behaviour of systems with sequential execution, conflict, concur- rency, synchronisation, merging, confusion or prioritisa- tion. In this study, we showed that PNs have great potential for providing high-level abstraction in the ES development life cycle. The PN’s graphic nature can provide visualisation of the dynamic behaviour of rule-based reasoning to design- ers for development and to users for validation. PNs help ES designers construct, validate and modify rule bases by observing the behaviour of the rules within the knowledge base. The PN’s with their analytic capa- bilities support knowledge verification techniques and PNs can effectively model sequential execution, conflict, concurrency, synchronisation, merging, confusion or pri- oritisation among rules which are crucial in real-time performance validation. PNs are traditionally used in electrical engineer- ing for process control. Over the past decade, several researchers have focussed on the application of ordinary PNs in business. In addition to ordinary PNs, timed PNs, stochastic PNs, and coloured PNs and fuzzy PNs are pow- erful tools for modelling complex business systems. More research is necessary to expand PN research and devel- opment into other areas of knowledge-based management and artificial intelligence. References Aitkenhead, MJ (2008). A co-evolving decision tree clas- sification method. Expert Systems with Applications, 34(1), 18–25. Baesens, B, R Setiono, C Mues and J Vanthienen (2003). Using neural network rule extraction and decision tables for credit-risk evaluation. Management Science, 49(3), 312–329. Balduzzi, F, A Giua and G Menga (2000). First-order hybrid Petri Nets: A model for optimization and con- trol. IEEE Transactions on Robotics and Automation, 16(4), 382–399. Becerra-Fernandez, V and R Sabherwal (2001). Organiza- tional knowledge management: A contingency perspec- tive. Journal of Management Information Systems, 18, 23–56. Berthelot, G and R Terrat (1982). Petri-Net theory for correctness of protocols. IEEE Transactions on Com- munication, 30(12), 2497–2505. Bruno, G and G Marchetto (1986). Process-translatable Petri Nets for the rapid prototyping of process control systems. IEEE Transactions on Software Engineering, 12(2), 346–357. Bullers, WI Jr. (1991). A tripartite approach to infor- mation systems development. Decision Sciences, 22(1), 120–135. Chen, SC, YL Ke and JS Wu (2001). Coloured Petri-Nets approach for solving distribution system contingency. Proceedings of the IEEE, 148(5), 463–470. Chen, YL, CL Hsu and SC Chou (2003). Constructing a multi-valued and multi-labeled decision tree. Expert Systems with Applications, 25(2), 199–209. Eppler, M (2001). Making knowledge visible through intranet knowledge maps: Concepts, elements, cases. Proceedings of 34th Hawaii International Conference on System Sciences, pp. 1530–1539, Piscataway, NJ: IEEE. Eppler, M (2004). Facilitating knowledge communication through joint interactive visualization. Journal of Uni- versal Computer Science, 10, 683–690. Handzic, M (2003). An integrated framework of KM. In Australian Studies in Knowledge Management, Hasan, H and M Handzic (eds.), pp. 3–34, Wollongong: University of Wollongong Press. Holt, AW (1971). Introduction to occurrence systems. In Associate Information Techniques, Jacks, L (ed.), pp. 175–203, New York: American Elsevier. Huang, Y-M, J-N Chen, T-C Huang, Y-L Jeng and Y-H Kuo (2008). Standardized course generation process using Dynamic Fuzzy Petri Nets. Expert Systems with Applications, 34, 72–86. Janssensa, D, G Wetsa, T Brijsa, K Vanhoofa, T Arentzeb and H Timmermansb (2006). Integrating Bayesian net- works and decision trees in a sequential rule-based transportation model. European Journal of Operational Research, 175(1), 16–34. Jantzen, M (1980). Structures representation of knowl- edge by petri nets as an aid for teaching and research. Lecture Notes in Computer Science. Berlin: Springer Verlag. April 12, 2008 15:19 00190 46 M. Tavana Kim, CH, DS Yim and RH Weston (2001). An inte- grated use of IDEF0, IDEF3 and Petri-Net meth- ods in support of business process modeling. Proceed- ings of the Institution of Mechanical Engineers, 215(4), 317–329. Lee, J, KFR Liu and W Chiang (1999). A fuzzy Petri Net-based expert system and its application to dam- age assessment of bridges. IEEE Transactions on Sys- tems, Man, and Cybernetics — Part B: Cybernetics, 29, 350–369. Lee, J and LF Lai (2002). A high-level Petri-Nets-based approach to verifying task structures. IEEE Trans- actions on Knowledge and Data Engineering, 14(2), 316–335. Lee, J, JI Pan and JY Kuo (2001). Verifying scenarios with time Petri-Nets. Information and Software Tech- nology, 43(13), 769–781. Lee, S, S Lee and Y Park (2007). A prediction model for success of services in e-commerce using decision tree: E-customer’s attitude towards online service. Expert Systems with Applications, 33(3), 572–581. Li, X and FL Rosano (2000). Adaptive fuzzy Petri nets for dynamic knowledge representation and inference. Expert Systems with Applications, 19(3), 235–241. Liu, R, A Kumar and W van der Aalst (2007). A for- mal modeling approach for supply chain event man- agement. Decision Support Systems, 43, 761–778. Liu, KFR, J Lee and W Chiang (1999). A fuzzy Petri- Net-based expert system and its application to damage assessment of bridges. IEEE Transactions on Systems, Man and Cybernetics, 29(3), 350–369. Liu, KFR, J Lee and W Chiang (2000). High-level fuzzy Petri-Nets as a basis for managing symbolic and numer- ical information. International Journal on Artificial Intelligence Tools, 9(4), 569–588. Mehrez, A, M Muzumdar, W Acar and G Weinroth (1995). A Petri-Net model view of decision making: An operational analysis. Omega — The International Journal of Management Science, 23(1), 63–78. Murata, T (1989). Petri Nets: Properties, analysis and applications. Proceedings of the IEEE, 77(4), 541–580. Peterson, JL (1981). Petri Net Theory and the Modeling of Systems. Morristown, NJ: Prentice-Hall, Inc. Petri, CA (1962). Kommunikation mit Automaten. Uni- versity of Bonn. English Translation by C.F. Greene (1965) Communication with automata, Supplement to Technical Report RADC-TR-65-377, Vol. 1, Rome Air Development Center, Griffith Air Force Base, Rome, N.Y. Piera, MA, M Narciso, A Guasch and D Riera (2004). Optimization of logistic and manufacturing systems through simulation: A colored Petri Net-based method- ology. Simulation, 80(3), 121–129. Sakthivel, S, MR Tanniru (1988–1989). Information verifi- cation and validation during requirement analysis using Petri nets. Journal of Management Information Sys- tems, 5(3), 33–52. Van Der Aalst, WMP (1999). Woflan: A Petri-net-based workflow analyzer. Systems Analysis, Modeling, Simu- lation, 35(3), 345–357. Van Hee, KM, LJ Somers and M Voorhoeve (1991). A modeling environment for decision support systems. Decision Support Systems, 7, 241–251. Verbeek, HMW, T Basten and WMP Van Der Aalst (2001). Diagnosing workflow processes using Woflan. The Computer Journal, 44(4), 246–279. Whitten, J, L Bentley and K Dittman (2001). Sys- tems Analysis and Design Methods. Boston, MA: Irwin McGraw-Hill. Wong, ML (2001). A flexible knowledge discovery system using genetic programming and login grammars. Deci- sion Support Systems, 31, 405–428. Wu, CH and SJ Lee (1997). Knowledge verification with an enhanced high-level Petri-Net model. IEEE Expert, October, 73–80. Yang, B-S, D-S Lim and ACC Tan (2005). VIBEX: An expert system for vibration fault diagnosis of rotating machinery using decision tree and deci- sion table. Expert Systems with Applications, 28(4), 735–742. Zhovtobryukh, D (2007). A Petri Net-based approach for automated goal-driven web service composition. Simu- lation, 83(1), 33–63. Zhu, H and H Xudong (2002). A methodology of testing high-level Petri-Nets. Information and Science Tech- nology, 44, 473–489. Madjid Tavana is a Professor of Management Informa- tion Systems and the Lindback Distinguished Chair of Information Systems at the La Salle University where he served as the Chairman of the Management Depart- ment, Director of the Center for Technology and Man- agement, and the Executive Director of the E-Commerce Institute. He has been a Faculty Fellow in Aeronautics and Space Research at the NASA’s Kennedy Space Cen- ter, developing Group Decision Support Systems for the Space Shuttle Program and Johnson Space Center, devel- oping Intelligent Decision Support Systems for Human Exploration of Mars at the Mission Control Center. In 2005, he was awarded the prestigious Space Act Award by NASA. He holds an MBA, a PMIS, and a PhD in Manage- ment Information Systems and received his Post-Doctoral Diploma in Strategic Information Systems from the Whar- ton School of the University of Pennsylvania. He is the Editor-in-Chief of the International Journal of Applied Decision Sciences and has published in journals such as Decision Sciences, Information Systems, Interfaces, Infor- mation and Management, Computers and Operations Research, among others. 
work_ncrbjlito5eftbiht5vr5svawq	----	PII: 0898-1221(85)90050-1 o(w7-4943185 $3.00 + .oo 0 IYX5 Pergamon Preu Ltd CSRL: A LANGUAGE FOR EXPERT SYSTEMS FOR DIAGNOSISt TOM BYLANDER, SANJAY MITTAL* and B. CHANDRASEKARAN Artificial Intelligence Group, Department of Computer and Information Science, The Ohio State University, Columbus, OH 43210, U.S.A. (Received September 1984) Abstract-We present CSRL (Conceptual Structures Representation Language) as a language to facilitate the development of expert diagnosis systems based on a paradigm of “cooperating diagnostic specialists.” In our approach diagnostic reasoning is one of several generic tasks, each of which calls for a particular organizational and problem-solving structure. A diagnostic structure is composed of a collection of specialists, each of which corresponds to a potential hypothesis about the current case. They are organized as a classification or diagnostic hierarchy, e.g. a classification of diseases. A top-down strategy called establish-refine is used, in which either a specialist establishes and then refines itself, or the specialist rejects itself, pruning the hierarchy that it heads. CSRL is a language for representing the specialists of a diagnostic hierarchy and the diagnostic knowledge within them. The diagnostic knowledge is encoded at various levels of abstractions: message procedures, which describe the specialist’s behavior in response to messages from other specialists; knowledge groups, which determine how data relate to features of the hypothesis; and rule-like knowledge, which is contained within knowledge groups. I. INTRODUCTION Many kinds of problem-solving for expert systems have been proposed within the AI community. Whatever the approach, there is a need to acquire the knowledge in a given domain and implement it in the spirit of the problem-solving paradigm. Reducing the time to implement a system usually involves the creation of a high-level language that reflects the intended method of problem-solving. For example, EMYCIN[ 11 was created for building systems based on MYCIN- like problem-solving[2]. Such languages are also intended to speed up the knowledge acquisition process by allowing domain experts to input knowledge in a form close to their conceptual level. Another goal is to make it easier to enforce consistency between the expert’s knowledge and its implementation. CSRL (Conceptual Structures Representation Language) is a language for implementing expert diagnostic systems that are based on our approach to diagnostic problem-solving. This approach is an outgrowth of our group’s experience with MDX, a medical diagnostic program[3], and with applying MDX-like problem-solving to other medical and nonmedical domains. CSRL facilitates the development of diagnostic systems by supporting constructs that represent di- agnostic knowledge at appropriate levels of abstraction. First, we will overview the relationship of CSRL to our overall theory of problem-solving types and the diagnostic problem-solving that underlies CSRL. We then present CSRL, illus- trating how its constructs are used to encode diagnostic knowledge. Two expert systems under development in our laboratory, which use CSRL, are then briefly described. Based on our experience with these systems, we point out where improvements in CSRL are needed. 2. CLASSIFICATORY DIAGNOSIS The central problem-solving of diagnosis, in our view, is classificatory activity. This is a specific type of problem-solving in our approach, m;aning that a special kind of organization and special strategies are strongly associated with performing expert diagnosis. In this section we will briefly review the theory of problem-solving types, as presented by Chandrasekaran[4], and the structure and strategies of the diagnostic task[5]. tThis an expanded version of a paper of the same title presented at the 1983 International Joint Conference on Artificial Intelligence. Wurrently at Knowledge Systems Area, Xerox PARC, 3333 Coyote Hill Rd., Palo Alto, CA 94304, U.S.A. 449 450 2.1 Types of problem-solving T. BYLANDER et al. We propose that expert problem-solving is composed of a collection of different problem- solving abilities. The AI group at Ohio State has been working at identifying well-defined types of problem-solving (called generic tasks), one of which is classificatory diagnosis. (For the purposes of this discussion we will use “diagnosis ” in place of “classificatory diagnosis” with the understanding that the complete diagnostic process includes other elements as well.) Other examples include knowledge-directed data retrieval, consequence finding, and a restricted form of design. Each generic task calls for a particular organizational and problem-solving structure. Given a specific kind of task to perform, the idea is that specific ways to organize and use knowledge are ideally suited for that task. Even when the specification of a problem is reduced to a given task within a given domain, the amount of knowledge that is needed can still be enormous (e.g. diagnosis in medicine). In our approach the knowledge structure for a given task and domain is composed of specialists, each of which specialize in different concepts of the domain. Domain knowledge is distributed across the specialists, dividing the problem into more manageable parts, and organizing the knowledge into chunks that become relevant when the corresponding concepts become relevant during the problem-solving. Decomposing a domain into specialists raises the problem of how they will coordinate during the problem-solving process. First, the specialists as a whole are organized primarily around the “subspecialist-of” relationship. Each task may specify additional relationships that may hold between specialists. Second, each task is associated with a set of strategies that take advantage of these relationships and the problem-solving capabilities of the individual specialists. The choice of what strategy to follow is not a global decision, but is chosen by the specialists during problem-solving. 2.2 The diagnostic task The diagnostic task is the identification of a case description with a specific node in a predetermined diagnostic hierarchy. Each node in the hierarchy corresponds to a hypothesis about the current case. Nodes higher in the hierarchy represent more general hypotheses, and lower nodes are more specific. Typically, a diagnostic hierarchy is a classification of malfunc- tions of some object, and the case description contains the manifestations and background information about the object. For example, the Auto-Mech expert system[6] attempts to classify data concerning an automobile into a diagnostic hierarchy of fuel-system malfunctions. Figure 1 illustrates a fragment of Auto-Mech’s hierarchy. The most general node, the fuel system in this example, is the head node of hierarchy. More specific fuel-system malfunctions, such as fuel-delivery problems, are classified within the hierarchy. Each node in the hierarchy is associated with a specialist that contains the diagnostic knowledge to evaluate the plausibility of the hypothesis from the case description. From this knowledge the specialist determines a confidence value representing the amount of belief in the hypothesis. If this value is high enough, the specialist is said to be established. The basic strategy of the diagnostic task is a process of hypothesis refinement, which we call establish-refine. In this strategy, if a specialist establishes itself, then it refines the hypothesis by invoking its subspecialists, which also perform the establish-refine strategy. If its confidence value is low, the specialist rejects the hypothesis and performs no further actions. Note that when this happens, the whole hierarchy below the specialist is eliminated from consideration. Otherwise the specialist suspends itself and may later refine itself if its superior requests it. Bad Fuel Problems n--._ Fuel Mixture Problems Low Octane Water in Fuel Dirt In Fuel Fig. I Fragment of a diagnostic hierarchy CSRL 451 With regard to Fig. 1, the following scenario might occur. First, the fuel-system specialist is invoked, since it is the top specialist in the hierarchy. This specialist is then established, and the two specialists below it are invoked. Bad fuel problems are rejected, eliminating the three subspecialists of bad fuel from consideration. Finally, the fuel-mixture specialist is established, and its subspecialists (not shown) are invoked. An important companion to the diagnostic hierarchy is an intelligent data-base assistant that organizes the case description, answers queries from the diagnostic specialists, and makes simple inferences from the data[7]. For example, the data base should be able to infer that the fuel tank is not empty if the car can be started. The diagnostic specialists are then relieved from knowing all the ways that a particular datum could be inferred from other data. There are several issues relevant to diagnostic problem-solving that we will not address here. The simple description above does not employ strategies for bypassing the hierarchical structure for common malfunctions, for handling multiple interacting hypothesis, or for ac- counting for the manifestations. Also, additional control strategies are required when many nodes are in a suspended state. For discussion on some of these topics, see Gomez and Chan- drasekaran[5]. Test ordering, causal explanation of findings, and therapeutic action do not directly fall within the auspices of the classificatory diagnosis as defined here, but expertise in any of these areas would certainly enhance a diagnostic system. Fully resolving all of these issues and integrating their solutions into the diagnostic framework are problems for future research. 2.3 Differences from other approaches The usual approach to building knowledge-based systems is to emphasize a general knowl- edge representation structure and different problem-solvers that use that knowledge. One dif- ference in this approach is that the organization of knowledge is not intended as a general representation for all problems. Rather it is tuned specifically for diagnosis. By limiting the type of problem to be solved, a specific organizational technique (classification hierarchy) and problem-solving strategy (establish-refine) can be used to provide focus and control in the problem-solving process. Another difference is that the specialists in the hierarchy are not a static collection of knowledge. The knowledge of how to establish or reject is embedded within the specialists. Each specialist can then be viewed as an individual problem-solver with its own knowledge base. The entire collection of specialists engages in distributed problem-solving. 3. CSRL CSRL is a language for representing the specialists of a diagnostic hierarchy and the diagnostic knowledge within them. The diagnostic knowledge is encoded at various levels of abstractions. Message procedures describe the specialist’s behavior in response to messages from other specialists. These contain the knowledge about how to establish or refine a specialist. Knowledge groups determine how selected data relate to various features or intermediate hy- potheses that are related to the specialist. The selected data may be the values of other knowledge groups, so that a single knowledge group can “summarize” the results of several others. Knowledge groups are composed of rule-like knowledge that matches the data against specific patterns and, when successful, provides values to be processed by the knowledge group. 3.1 Specialists In CSRL a diagnostic expert system is implemented by individually defining each specialist. The super- and subspecialists of the specialist are declared within the definition. Figure 2 is a (Specialist BadFuel (declare (superspecialist FuelSystem) (subspecialists LowOctane WaterInFuel DirtInFuel)) (kgs . ..I (messages . ..)I Fig. 2. Skeleton specialist for BadFuel 452 T. BYLANDER et al skeleton of a specialist definition for the bad fuel node from Fig. 1. The declare section specifies its relationships to other specialists. The other sections of the specialist are examined below. Since CSRL is designed to use only a simple classification tree, many choices concerning the composition of the hierarchy must be made. This is a pragmatic decision, rather than a search for the “perfect” classification tree. The main criteria for evaluating a classification is whether enough evidence is normally available to make confident decisions. To decompose a specialist into its subspecialists, the simplest method is to ask the domain expert what subhy- potheses should be considered next. Usually the subspecialists will differ from one another based on a single attribute (e.g. location, cause). For further discussion on this and other design decisions in CSRL, see Bylander and Smith[8]. 3.2 Messuge procedures The messages section of a specialist contains a list of message procedures that specify how the specialist will respond to different messages from its superspecia1ist.t “Establish,” “Re- fine, ” “Establish-Refine” (combines Establish and Refine), and “Suggest” are predefined messages in CSRL; additional messages may be defined by the user. Below, we will examine how Establish and Refine procedures are typically constructed. Message procedures are the highest level of abstraction for diagnostic knowledge within specialists. Just as in general message-passing languages, messages provide a way to invoke a particular kind of response without having to know what procedure to invoke. Strategies for diagnosis, such as Establish-Refine, are usually easy to translate into a message protocol. However. CSRL does not provide any way to specify and enforce message protocols. Figure 3 illustrates the Establish message procedure of the BadFuel specialist. “relevant” and “summary” are names of knowledge groups of BadFuel. “self” is a keyword that refers to the name of the specialist. This procedure first tests the value of the relevant knowledge group. (If this knowledge group has not already been executed, it is automatically executed at this point.) If it is greater than or equal to 0, then BadFuel’s confidence value is set to the value of the summary knowledge group, or else it is set to the value of the relevant knowledge group. In CSRL a confidence value scale of - 3 to + 3 is used (integers only). A value of + 2 or + 3 indicates that the specialist is established. In this case the procedure corresponds to the following diagnostic knowledge. First perform a preliminary check to make sure that BadFuel is a relevant hypothesis to hold. If it is not (the relevant knowledge group is less than 0). then set BadFuel’s confidence value to the degree of rclcvancy. Otherwise, perform more complicated reasoning (the sum- mary knowledge group combines the values of other knowledge groups) to determine BadFuel’s confidence value. Figure 4 shows a Refine procedure that is a simplified version of the one that BadFuel uses. “subspecialists” is a keyword that refers to the subspecialists of the current specialist. The procedure calls each subspecialist with an English message. $ If the subspecialist establishes itself (t ? tests if the confidence value is +2 or +3), then send it a Refine message. CSRL has a variety of other kinds of statements and expressions so that more complicated (Establish (if (GE relevant 0) then (SetConfidence self sunrmary) else (SetConfidence self relevant))) Fig. 3. Establish procedure of BadFuel. +A specialist I\ not allowed to send messages to its superspecialist. However. other message-passing routes are allowed. Specifically. a specialist may send a message to itself. across the hierarchy, and to indirect subspecialists. In the latter case each interconnecting specialist is sent a “suggest” message and decides within its suggest message procedure whether or not to pass the original message downwards. $‘For convenience many of CSRL’s control constructs mimic those of INTERLISP; however, these constructs are executed by the CSRL interpreter, not by using LISP EVAL. LISP code is allowed within message procedures, but only within a construct called “DoLisp. ” This is not intended to let specialists have arbitrary code, but to allow interaction with other LISP-implemented systems. CSRL (Refine (for specialist in subspecialists do (Call specialist with Establish) (if (+? specialist) then (Call specialist with Refine)))) Fig. 4. Refine procedure. 453 strategies can be implemented. For example, a “Reset” statement deletes the confidence value and the knowledge group values of a specialist. This might be used when additional tests are performed, making it necessary to recalculate the confidence value. Also, messages can be parameterized, and message procedures can declare local variables. 3.3 Knowledge groups The kgs section of a specialist definition contains a list of knowledge groups that are used to evaluate how selected data indicate various features or intermediate hypotheses that relate to specialist’s hypothesis. A knowledge group can be thought of as a cluster of production rules that map the values of a list of expressions (boolean and arithmetic operations on data) to some conclusion on a discrete, symbolic scale. Different types of knowledge groups perform this mapping differently: e.g. directly mapping values to conclusions, or ha;ling each rule add or subtract a set number of “confidence” units. Knowledge groups are intended for encoding the heuristics that a domain expert uses for inferring features of an hypothesis from the case description. The main problem is that this inference is uncertain--there is rarely a one-to-one mapping from data to the features of the hypothesis. The way that this is handled in CSRL is borrowed from the uncertainty handling techniques used in MDX[9]. Each feature or intermediate hypothesis is associated with a knowledge group. The data that the domain expert uses to evaluate the feature are encoded as expressions in the knowledge group. These are usually queries to a separate data-base system. Each combination of values of the expressions is then mapped to a level of confidence as determined by the domain expert. This set of knowledge groups becomes the data for another knowledge group, which determines the confidence value of the specialist from the confidence values of the features.? By examining the results of test cases, we see that the knowledge groups are relatively easy to debug, since the attention of the domain expert can be directed to the specific area of knowledge that derived the incorrect result. As an example, Fig. 5 is the relevant knowledge group of the BadFuel specialist mentioned above. It determines whether the symptoms of the automobile are consistent with bad fuel problems. The expressions query the user (who is the data base for Auto-Mech) for whether the car is slow to respond, starts hard, has knocking or pinging sounds, or has the problem when accelerating. ‘ ‘AskYNU?” is a LISP function that asks the user for a Y, N, or U (unknown) answer from the user, and translates the answer into T, F, or U, the values of CSRL’s three- (relevant Table (match (AskYNU? “Is the car slow to respond”) (AskYNU? “Does the car start hard”) (And (AskYNU? “Do you hear knocking or pinging sounds”) (AskYNU? “Does the problem occur while accelerating”)) with (if T ? ? then -3 elseif ? T ? then -3 elseif ? ? T then 3 else 1))) Fig. 5. Relevant knowledge group of BadFuel tActually, any number of knowledge group levels can be implemented 454 T. BYLANDER ef al. valued logic. Each set of tests in the if-then part of the knowledge group is evaluated until one matches. The value corresponding to this “rule” becomes the value of the knowledge group. For example, the first rule tests whether the first expression is true (the “?” means doesn’t matter). If so, then - 3 becomes the value of the knowledge group. Otherwise, other rules are evaluated. The value of the knowledge group will be 1 if no rule matches. This knowledge group encodes the following diagnostic knowledge: If the car is slow to respond or if the car starts hard, then BadFuel is not relevant in this case. Otherwise, if there are knocking or pinging sounds and if the problem occurs while accelerating, then BadFuel is highly relevant. In all other cases BadFuel is only mildly relevant. Figure 6 is the summary knowledge group of BadFuel. Its expressions are the values of the relevant and gas knowledge groups (the latter queries the user about the temporal relationship between the onset of the problem and when gas was last bought). In this case, if the value of the relevant knowledge group is 3 and the value of the gas knowledge group is greater than or equal to 0, then the value of the summary knowledge group (and consequently the confidence value of BadFuel) is 3, indicating that a bad-fuel problem is very likely. 3.4 Comparison with rule-bused languages There is nothing in CSRL that is not programmable within rule-based languages such as OPSS[lO] or EMYClN[l]. The difference between CSRL and these languages is that CSRL makes a commitment to a particular organizational and programming style. CSRL is not intended to be a general-purpose representation language, but is built specifically for the classificatory diagnosis problem. It is possible to program in a rule-based language, so that there is an implicit relationship between rules so that they correspond to knowledge groups and specialists. Rl, although not a diagnostic expert system, is an excellent example of how one creates implicit grouping of rules in such a system[ 111. The central idea underlying CSRL is to make these relationships explicit. The expert system implementor is then relieved from trying to impose an organization on a organizationless system and is free to concentrate on the conceptual structure of the domain. Also, there is a greater potential to embed explanation and debugging facilities that can take advantage of the expert system organization. 3.5 The CSRL environment The current version of CSRL is implemented in INTERLISP-D and LOOPS, an object- oriented programming tool. Each specialist is implemented as a LOOPS class, which is in- stantiated for each case that is run. The LOOPS class hierarchy is used to specify default message procedures and shared knowledge groups, making it easy to encode a default establish- refine strategy, and letting the user incrementally modify this strategy and add strategies as desired. A graphical interface displays the specialist hierarchy and, through the use of a mouse, allows the user to easily access and modify any part of the hierarchy. Additional facilities for debugging and explanation are being implemented. 4. EXPERT SYSTEMS THAT USE CSRL 4.1 Auto-Mech Auto-Mech is an expert system that diagnoses fuel problems in automobile engines[6]. This domain was chosen to demonstrate the viability of our approach to nonmedical domains, as well as to gain experience and feedback on CSRL.t The purpose of the fuel system is to deliver a mixture of fuel and air to the air cylinders of the engine. It can be divided into major subsystems (fuel delivery, air intake, carburetor, vacuum manifold) that correspond to initial hypotheses about fuel-system faults. tAuto-Mech was developed using an early version of the language CSRL (sunmrary Table (match relevant gas with (if 3 (GE 0) then 3 elseif 1 (GE 0) then 2 elseif ? (LT 0) then -3))) Fig. 6. Summary knowledge group of BadFuel 455 Auto-Mech consists of 34 CSRL specialists in a hierarchy that varies from four to six levels deep. Its problem-solving closely follows the establish-refine strategy. Before this strategy is invoked, Auto-Mech collects some initial data from the user. This includes the major symptom that the user notices (such as stalling) and the situation when this occurs (e.g. accelerating and cold engine temperature). Any additional questions are asked while Auto-Mech’s specialists are running. The diagnosis then starts and continues until the user is satisfied that the diagnosis is complete. The user must make this decision because the data that Auto-Mech uses are very weak at indicating specific problems, and, more importantly, Auto-Mech is unable to make the repair and determine whether the problem has been fixed. A major part of Auto-Mech’s development was determining the assumptions that would be made about the design of the automobile engine and the data that the program would be using. Different automobile engine designs have a significant effect on the hypotheses that are considered. A carbureted engine, for example, will have a different set of problems than a fuel- injected engine (the former can have a broken carburetor). The data was assumed to come from commonly available resources. The variety of computer analysis information that is available to mechanics today was not considered, in order to simplify building Auto-Mech. 4.2 Red Red is an expert system whose domain is red-blood-cell antibody identification[l2]. An everyday problem that a blood bank contends with is the selection of units of blood for transfusion during major surgery. The primary difficulty is that antibodies in the patient’s blood may attack the foreign blood, rendering the new blood useless as well as presenting additional danger to the patient. Thus, identifying the patient’s antibodies and selecting blood that will not react with them is a critical task for nearly all red-blood transfusions. The Red expert system is composed of three major subsystems, one of which is implemented in CSRL. The non-CSRL subsystems are a data base, which maintains and answers questions about reaction records (reactions of the patient’s blood in selected blood samples under a variety of conditions), and an overview system, which assembles a composite hypothesis of the anti- bodies that would best explain the reaction record[ 131. CSRL is used to implement specialists corresponding to each antibody that Red knows about (about 30 of the most common ones) and to each antibody subtype (different ways that the antibody can react). The major function of the specialists is to rule out antibodies and their subtypes whenever possible, thus simplifying the job of the overview subsystem, and to assign confidence values, informing overview of which antibodies appear to be more plausible. The specialists query the data base for information about the test reactions and other patient information, and also tell the data base to perform certain operations on reaction records. An interesting feature of Red is the way it handles the problem of interacting hypotheses. It is possible for the patient’s blood to have practically any number or combination of antibodies, which makes it very hard for a single specialist to determine how well it will fit with other specialists in a composite hypothesis. In Red each specialist is encoded to assume that it is independent-it looks at the data as if no other specialist can account for the same data. The knowledge of how the specialists can interact is left to the overview subsystem. This would be problematic if few specialists could rule themselves out, but it so happens that in this domain it is rare to have more than a few antibodies that cannot be independently ruled out. Thus Red’s CSRL subsystem makes overview’s problem-solving computationally feasible since it consid- erably reduces the amount of search that would otherwise be necessary. 456 T. BYLANDER et al. 5. NEEDED IMPROVEMENTS IN CSRL The largest flaw in CSRL is that there is no strategy that determines when diagnosis should stop. Currently, the default procedures simply ask the user if the current diagnosis is satisfactory. Some notion of what it means to account for the data needs to be added to the language. The work on Red’s overview system is a step in this direction, but there needs to be more integration of overview and CSRL (currently overview starts after the specialists are finished) and a better understanding of what kinds of interactions can occur between two hypotheses. Progress in this area would also help increase the focus of the diagnosis; i.e. the diagnosis could concentrate on accounting for the most important manifestation(s). Another problem is the meaning of the confidence value of a specialist. In MDX this value was directly associated with the amount of belief in the specialist. However, in both Auto- Mech and Red, this meaning had to be slightly altered to fit the purposes of the expert system. In Auto-Mech the confidence value is used to indicate whether the hypothesis was worth pursuing. In Red it is used to indicate the specialist’s plausibility, given the independence assumption mentioned earlier. It is not possible in either expert system to confirm a specialist without outside help. In Auto-Mech a repair or highly specific test must be performed, but in Red all the specialists must be considered together. This does not create a problem for the process of establish-refine problem-solving, but makes it difficult to explain what the confidence value means. Any explanation facility must understand the assumptions that are being made in order to make coherent explanations. 6. CONCLUSION We believe that the development of complex expert systems will depend on the availability of special-purpose languages with organizational and problem-solving tools that match the conceptual structure of the domain. CSRL represents an initial step in this direction. It provides facilities to organize diagnostic knowledge in accordance with the structure of the domain. In particular, CSRL’s constructs facilitate the encoding of rule-like and strategic knowledge into appropriate abstractions: knowledge groups, message procedures, and specialists. Acknow~/edgments~We would like to acknowledge Jack Smith and Jon Sticklen for many fruitful discussions concerning CSRL’s design. Many improvements in the language are due to Mike Tanner and John Josephson, who implemented the CSRL specialists in Auto-Mech and Red., The language development is funded by a grant from the Battelle Memorial Laboratories University Distribution Program, and experimentation and application in different domains is supported by AFOSR grant 82-0255, and NSF grant MCS-8103480. I 8 9. IO. II. 12. 13. REFERENCES W. van Melle, A domain independent production-rule system for consultation programs. Proc. Sixth Int. Cot$ on ArtijZcial Intelligence, pp. 923-925. Tokyo (1979). E. H. Shortliffe, Computer-Based Medical Consultations: MYCIN. Elsevier. New York (1976). B. Chandrasekaran and S. Mittal, Conceptual representation of medical knowledge for diagnosis by computer: MDX and related systems, in Advancrs in Computers, pp. 217-293. Academic Press. New York, (1983). B. Chandrasekaran, Towards a taxonomy of problem solvin, ~7 types. AI Mq. 4. 9917 (1983). F. Gomez and B. Chandrasekaran, Knowledge org,;nization and distribution for medical diagnosis. IEEE Trans. Svst.. Man Cybentetics SMC-11. 34-42 (1981). M. C. Tanner and ‘f. Bylander, Application of the CSRL language to the design of expert diagnosis systems: the auto-mech experience. Proc. Joint Swvices Workshop on Artificial /nte//igence in Maintenunce. Department of Defense, pp. 13lll52, Denver (1984). S. Mittal and B. Chandrasekaran, Conceptual representation of patient data bases. J. Medical Swt. 4, 169-185 (1980). T. Bylander and J. W. Smith, Using CSRL for medical diagnosis. Pmt. .%wmd Int. Cmfl IVI Medical Computer Science and Compututimal Medicine. IEEE Computer Sot.. Glouster. Ohio C 1983). B. Chandrasekaran, S. Mittal and J. W. Smith, Reasoning with uncertain knowledge: the MDX approach. Proc. Con,qrcts Americtm Medical Informtics A.ssociation. San Francisco (19X2). C. L. Forpy, OPS5 Users Manuul. Tech. Rept. CMU-CS-8 I 135, Carnegie-Mellon University ( I98 I). J. McDermott. RI: a rule-based configurer of computer systems. Artificial Irttrll. IY. 39-88 (1982). J. W. Smith. J. Josephson, C. Evans, P. Straum and J. Noga, Design for a red-cell antibody identification expert. Proc. Srcond lat. Co@ on Medical Computer Science and Computational Mrdir,inr. IEEE Computer Sot., Glouster, Ohio (1983). J. Josephson, B. Chandrasekaran and J. W. Smith, The Ovrrvte~~ Fmctim in Diagnostic Problem Solving. Technical Paper, Al Group, Dept. of Computer and Information Science, The Ohio State University (1984). 
work_nepu4n3jovbb5mrn3zv7tcstja	----	Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 A SCADA Oriented Middleware for RFID Technology Ismael Abad Cardiel ∗1, Ruben Heradio†1, Carlos Cerrada Somolinos‡1, and Jose Cerrada Somolinos§1 1ETS de Ingenieria Informatica, Universidad Nacional de Educacion a Distancia, Madrid, Spain Abstract Radio Frequency IDentification (RFID) has emerged as the new technology paradigm for acquisition and information management. RFID can be used to improve significantly the efficiency of business processes by providing the capability of automatic identification and data capture. This technology introduces new challenges on data and process information management in current systems. RFID data are time- dependent and dynamically changing. In addition, data carry implicit semantics. The homogeneous data processing of such implicit semantics allows us to propose RFID middleware as a WHO–WHEN–WHERE data problem. This paper presents DEPCAS, a new middleware for RFID information based on the SCADA architecture for control systems. An application of DEPCAS is the resolution of heterogeneous situations, which solves the WHAT or context–aware to apply the auto identification data received from RFID systems in business applications. 1 Introduction Due to the ubiquity of Radio Frequency IDentification (RFID) deployments and low cost tagging [1], a growing amount of data are being generated from multiple sources that need to be processed and managed. So, new questions are arising, such as how to encompass current and future identification technology at the same time and how this information can be combined with existing applications [2, 3, 4]. The large amount of information generated in the RFID startup projects requires to know how and when the information is going to be summarized [5, 6]. Many of the non–finished RFID pilot projects have been focused in the RFID data instead of being fo- cused on the business related events [7]. Raw RFID data have little value until they are is transformed into a form suitable for application-level interaction [8]. Moreover, data have implicit meanings and associated relationships with other RFID data (e.g., data about containment, position, aggregation...). So, RFID appli- cations must make appropriate inferences. For instance, observations of RFID tags associated to several items and a RFID tag associated to a container in a certain time interval can imply that the items are packed in the container. In addition, the wide spread use of RFID technology involves a huge migration problem. There are thousands of applications using identification through codes, bar codes, plates, badges, 2D codes, operator identifications, etc [9]. that in a mid-time will change to RFID auto identification [10]. From the application ∗iabad@issi.uned.es †rheradio@issi.uned.es ‡ccerrada@issi.uned.es §jcerrada@issi.uned.es 1 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 that can be found in any little market, to the large and more complex information system of a big and global company, the change from identification to auto identification will arrive [11, 12]. The state of the art in RFID integration in existing and real systems provide a research prospect to propose a general middleware architecture that processes and consolidates RFID information solving the former issues and helping efficiently to migrate from identification to auto identification systems [13]. This paper proposes a new middleware architecture for RFID called DEPCAS (Data EPC1 Acquisition System), which integrates existing proposals for RFID middleware and it is based on the Supervisory Control And Data Acquisition (SCADA) architecture for control systems [14]. Key concepts supported by DEPCAS are: 1. Converting RFID information to business information: RFID applications receive basic and simple data source inputs as time–stamp, tag code or reader dock (reader plus antenna) [15]. This raw information should be translated to business relevant information using abstract process that can be applied in heterogeneous RFID installations. 2. Distributing RFID information to business processes: the RFID processed information need to be de- fined and accessible for external applications in a general way. The format and the exchange mech- anisms must be transparent and general for any kind of applications. The exchange flow between existing running programs and RFID acquisition system is bidirectional. It is necessary to solve how to send the necessary business information data to the RFID acquisition system and the opposite flow from RFID middleware to external entities. 3. Filtering information: the computer systems that manage a big quantity of input data must solve a filter policy. Filter process could be as simple as eliminating all unknown information or as complex as operations that manages every time–stamp and any network structure to ensure the data quality. In the RFID acquisition systems the raw tag readings are filtered according to a previous algorithm that helps the back end application to receive only relevant information. 4. Triggering business process from RFID process and vice versa: it is a common situation to solve that RFID acquisition system can receive any kind of input that triggers external activities. And in the opposite way, external input can modify RFID acquisition process. In any direction, RFID acquisition must solve the event trigger process. 5. Managing RFID devices: this requirement deals with an abstraction of physical communication and protocol exchange between reader devices and the acquisition platform. It is necessary to manage also the topological reader network, i.e., how readers are distributed or centralized. 6. Consolidating RFID information: the general RFID process generates relevant information that has to be consolidated. This permanent information is the result of specific process according to the homogeneous process. The homogeneous process creates detailed information about how, where, and why RFID information is being used in a productive process. The remainder of this paper is structured as follows. Section 2 summarizes related work. Section 3 describes DEPCAS. Section 4 outlines our current implementation of DEPCAS. Section 5 presents a scenario where DEPCAS is used. Finally, Section 6 presents the conclusions of our work. 2 Related work The need of RFID middleware is gradually becoming a basic question for non–trivial RFID installations. This is particularly true in heterogeneous environments, including multiple readers, applications instances, supply chain management, complex process and sophisticated business semantics [16, 17, 18, 19]. 1EPC stands for Electronic Product Code. 2 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 RFID middleware is indispensable for these main reasons: 1. The need to filter RFID data in order to avoid redundant or erroneous information that is not needed to business applications, while at the same time optimizing resources. 2. The need to process and interface RFID data from origin (tag reads) in heterogeneous deployments (multi-tagging and multi-readers systems) without resorting to business integration logic. 3. The flexibility to integrate RFID systems to support auto–identification in different applications and process models. There are several surveys in the literature that propose system taxonomies and compilations about RFID middleware [20, 21]. The RFID middleware market is dominated by three tendencies: 1. Specialized companies which offer RFID middleware. These companies can be classified into product oriented and specific RFID middleware producer. If they are reader/antenna, or general hardware producer, then they include specific RFID middleware to support the hardware. If we speak about a general middleware producer it is quite common they have included an specific plugin to acquire RFID data. There are new companies that have born with the specific commercial target to create, distribute and install RFID middleware. We can mention in this group: (a) WinRFID from UCLA-WINMEC Research Lab [22] is a research result to solve RFID features like scalability and administration, event and data intelligent process and dispatching. (b) OAT Foundation Suite2 is a RFID middleware platform that includes two-way process integration between enterprise applications and RFID acquisition systems. (c) DETEGO/You-R OPEN3 is the RFID middleware from RF-iT Solutions based on three main func- tions: business integration, data management and device management. (d) TrueVUE RFID Platform4 from VueTechnology is a complete software infrastructure to manager RFID deployment. It contains several modules to support different points of view in item level management. 2. The second group is the one of giants’ software vendors such as IBM, Microsoft, ORACLE/Sun Mi- crosystems, or SAP. Every of these huge companies have dedicated in the last decade important resources to include RFID acquisition inside their own products. Between the RFID middleware we can mention: (a) RFID Java Middleware from ORACLE/Sun Microsystems5. With ORACLE and Sun Microsystems company fusion, the RFID middleware from both companies was renamed using the ORACLE reference. Now the product name is ORACLE Sensor Edge Server based on ORACLE Fusion Mid- dleware. (b) SAP Solutions for Auto-ID and Item serialization6 is a set of components to include RFID tech- nologies in SAP management tool. The solution offers three major components: SAP Auto-ID infrastructure, SAP object event repository and SAP event manager. 2http://www.oatsystems.com 3see “Plug and Trace. Excellent RFID Solutions” available at http://www.enso-detego.com 4http://tycoretailsolutions.com 5see “Oracle Application Server Wireless Developer’s Guide” available at http://docs.oracle.com 6see “SAP Research Report 2009/2010. From creativity to value” available at http://research-report-20092010.research-events.com/ 3 http://www.oatsystems.com http://www.enso-detego.com http://tycoretailsolutions.com http://docs.oracle.com http://research-report-20092010.research-events.com/ Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 (c) IBM WebSphere RFID7 is the IBM middleware to implement RFID acquisition. It consists of Edge- Controllers and the WebSphere RFID Premises Servers which includes WebSphere Application Server and DB2 Workgroup server. (d) RFID Anywhere8 from Sybase iAnywhere is a software infrastructure that simplifies the develop- ment, deployment, configuration and management task for RFID networks. 3. The last group of RFID middleware producers is the open source initiatives. These proposals are research results that take origin in university groups or consortiums. There are several Open Source Initiatives to solve an RFID middleware: (a) CUHK RFID Middleware9 from MobiTEC Technologies Centre of the Chinese University of Hong Kong. (b) ASPIRE RFID Middleware10 is an EU funded project to deliver an Advanced Sensors and lightweight Programmable middleware for Innovative RFID Enterprise applications as an open source mid- dleware. (c) RIFIDI Edge Server11 is a software architecture proposed by the RFID Research Center in the University of Arkansas. (d) FOSSTRAK [3] (called ACCADA) is a middleware implementation sponsored by Auto-ID Labs oriented to produce a free and open source software for track and trace. SCADA and RFID middleware systems share a set of commonalities about structure, functions and characteristics. From an acquisition point of view, the physical low level is a set of devices acquiring data: we have called sensors and RTU’s in SCADA world and we call antenna/edge/reader in RFID world. The communication networks to integrate the data capture system with the master processing system. This network communication could be defined with a variety of technologies and topologies. And the last element: the server data processor. The functions in SCADA and RFID middleware systems are similar. Main question of both of them is to acquire data: SCADA data acquisition could be expressed in terms of discrete inputs (equipment is on or off, tripwire alarms, power failures, etc.) and analog values (level in tanks, voltage levels, temperature, or any other factor, etc.). RFID middleware data acquisition system is a three-tuple: the RFID code or the WHO, the time–stamp or the WHEN, and the reading point or the WHERE. If we analyze the characteristics in the two kind of acquisition systems, there are a set of common points between SCADA software systems and RFID acquisition system: 1. Hiding hardware: SCADA software system hides the specific Remote Terminal Units (RTUs) hard- ware. It is a common situation that SCADAs manage a set of RTUs from different companies or with different structure. This same situation could be found in RFID middleware with a deployment of an RFID infrastructure. At our days we can find a set of various RFID readers’ trades like Alien, Intermec, Feig, Siemens, Symbol, Sick, etc. 2. Hiding communication between devices and master software: data transport from acquiring devices to master software can be made using diverse communication infrastructures and different protocols. There are all kind of SCADA communication infrastructures, from the simplest serial wired structure to the most complex satellite structure. Additionally, any advance in telecommunication field will be used if there was any opportunity. The same situation may be repeated in RFID systems. The present 7see “IBM WebSphere Premises Server” available at http://www.ibm.com 8see “Getting The Most From Mobile RFID With Middleware” available at http://www.sybase.com 9see “CUHK RFID Middleware - System Design Document” available at http://mobitec.ie.cuhk.edu.hk 10http://wiki.aspire.ow2.org 11http://www.rifidi.org 4 http://www.ibm.com http://www.sybase.com http://mobitec.ie.cuhk.edu.hk http://wiki.aspire.ow2.org http://www.rifidi.org Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 readers have limited options to communicate to master software using Ethernet network or serial communications. But there are prototypes using optical fiber or wireless communications. 3. Quantity Information: one of the more significant common points between SCADA and RFID acqui- sition software is that they receive a relevant quantity of dynamic information. SCADAs can receive data with milliseconds of resolution from a big quantity of data signal to monitor a complete instal- lation. The same situation may be produced inside a RFID monitoring installation. This data should be received and processed to obtain relevant information. 4. From fine-grain information to relevant information: both systems acquire data characterized from a very fine grain. SCADA basic data are status signal, analog value, and may be rate counters. RFID acquire system receive data about auto identification with a code. The two type of systems have the possibility to manage additional basic information like time–stamp, read quality, read origin, etc. All this individual information has very little meanings and it has a very low meaning relevance. The relevant information is obtained when it is processed with others data to obtain significant information. 5. Real time process: SCADA systems are typical real time systems. Data is acquired and processed to obtain the relevant alarms or calculation that give information to the operators. Of course, there are differences about the meaning of real time depending on specific situations. It is not the same to speak about real time in electrical SCADA systems that in water SCADA. Time scale it is not comparable between electrical processes with water processes systems. The same situation it is reproduced in RFID acquisition system. Auto identification could be time critical as the process where it has been using. 6. Consolidating information: the SCADA and RFID middleware process data is defined to obtain sig- nificant information not only real time alarms or events. This processed data is structured, stored, and summarized to be available for other systems. All this operations allow filtering managing the information that is received in the system. 7. External interfaces: the consolidated information should be offered to others systems. Modern SCADA software systems manage external structured information to define and configure the own process. This information has in-out components. SCADA receives information about configuration, Geographic Information System (GISs), maintaining systems or any internal warehouse. SCADA offers information about the process results to any back-end company software. This situation is repeated in RFID middleware environments. 3 DEPCAS Figure 1 depicts the overall DEPCAS architecture. The scheme proposed here is based on the architecture of modern SCADA (Supervisory Control and Data Acquisition) systems. The basic issues in DEPCAS are: 1. Producing RFID processed data. 2. Hiding heterogeneous RFID deployment systems with an homogeneous layer approach. 3. Translating business (async or sync) needs to RFID systems. 4. Providing management capabilities in RFID middleware environment. To solve these main targets the structure of the acquisition system that we propose is organized in four great layers or subsystems: 5 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 Figure 1: Architecture of DEPCAS 6 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 1. Middleware Device Manager (MDM). It has the following basic functions. First, the communication management with one or more devices: RFID readers. Second, implementing communication pro- tocol with the readers and event processor implementation (ALE: Application Level Event used in the EPCglobal standard12). And third, topological configurations support defined by the acquisition equipment: antennas, edge readers and readers. 2. Middleware Logic Manager (MLM). It supports the process to obtain RFID summarized data from auto-identification raw acquired data. The logical process to resolve will depend on each particular scenario [23]. This process will generate information of a permanent nature which is the result of the process in every scenario instance. The basic scenarios on the proposed work are related with: monitoring and tracking operations, aggregation of information items to generate new information, sorting of items, format raw RFID data conversion, and state machine delivery information. The concept of scenario is equivalent to a logical process that generally can be applied in all situations where information is used for self-identification. At each stage information is received, processed, summarized and accumulates, and generate and consolidate the data according to an specific logic. 3. Graphical User Viewer (GUV). It is the mechanism that DEPCAS supplies to help in the installation and management of a RFID system from the engineering point of view. This subsystem supports two basic operations: the supervisory capability to the RFID system and certain data manageability (configuration and acquisition data). 4. Information system exchange (EPCIS: EPC Information Services). EPCIS services should be the gateway of information between the data provided by the MDM and different MLMs to external business. These services will enable both receiving and sending information from the DEPCAS system to other systems. Following subsections describe in detail MDM, MLM, GUV and EPCIS. 3.1 MDM One of the main purposes of RFID middleware is to manage and hide the broad range of device readers existing in acquisition RFID networks. There are two basic solutions to override this question. The first one is to define a specification that should be accomplished for every RFID reader inside a network. The second solution is to define an abstraction layer that can translate proprietary protocols and specific characteristics from a vendor to a general solution. MDM is the specific DEPCAS layer to support all these management operations. The study of different RFID Device Management components allows obtaining a set of features that are included to solve it in this kind of systems. Among these features we include: 1. Device Agnostic: the device management should include a specific function to hide the alternatives of existing devices related with: physical parametrization, communication protocol, communication options, etc. 2. Configuration and parametrization of RFID devices: an agnostic approach to existing devices sets the options to manage from an homogeneous point of view the configuration and parametrization of every equipment in the RFID acquisition network. 3. Topological RFID devices: the RFID device management can include some network organization that give additional capabilities to the acquisition network as: redundant point of reading, input charac- teristics, out characteristic, parallel reading devices, etc. 12see “Application Level Events Standard v. 1.1.1” at http://www.gs1.org/gsmp/kc/epcglobal/ale 7 http://www.gs1.org/gsmp/kc/epcglobal/ale Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 4. Start-up devices: the set of operations to startup and configure an RFID equipment inside a network are known as starting up protocols. Some of the existing RFID middleware support a starting up protocol giving instructions about how to include RFID reader in the network. 5. Statistical and Report RFID devices data: the deployments of RFID reader require to introduce mainte- nance operations in the network. The statistical and reporting information are essential to optimize these operations. 6. Monitoring devices in live operation: some existing RFID networks can manage critical operations thus the RFID middleware supports a supervisory management about working devices. 7. Diagnosis of devices in live operation: the RFID middleware includes operations like device signals, examining logs, updating software devices, etc. 8. LLRP implementation: the Low Level Reader Protocol13 is the standard proposal from EPCGlobal Network to specify the interface between RFID readers and any software client. 9. OSGi Technology: the increase of software complexity in heterogeneous hardware environment has promote the development of frameworks to support the continuous changes. OSGi framework is a module and service platform that implements a complete and dynamic model. Any framework that implements OSGi provides an environment for modularization of application in small pieces tightly-coupled, dynamic loadable and configuration files to declare external dependencies. The OSGi Architecture is one of the most common specification framework used in our days. The OSGi specifi- cation enable components to hide their implementation from others components while communicat- ing through services. These features and functions define the operation of the middleware device manager in RFID mid- dleware. For instance, Device Agnostic Feature is provided when middleware infrastructure acts as an integration point for RFID readers, printers, and other infrastructure devices, such as any sensors and actuators. RFID reader and printer manufacturers typically provide complex configuration options, pro- prietary interface languages, and communication protocols. If RFID device infrastructure eliminates the need for RFID integrators to be aware of these complexities and provides a device agnostic interface with which supported devices can be configured hiding the complexity. Other features are related to scope and technological implementation of middleware device manager. Table 1 compares the features supported by DEPCAS MDM to the related work summarized in Section 2. The main goals of MDM in DEPCAS are: 1. Managing the topology RFID reader’s network. 2. Agnostic management of different RFID vendors. 3. Homogeneous process of different RFID tag reads. 4. Supporting LLRP and ALE specification process. 5. Providing configurable RFID network. 6. Hiding operational environment. 7. Minimum tag management operations. 8. Monitoring and controlling RFID readers. 13see “LLRP Standard v. 1.1” available at http://www.gs1.org/epcglobal 8 http://www.gs1.org/epcglobal Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 To support these objectives, the MDM definition includes the concept of Minimum Abstract Reader Commands (MARC) and the RFID Reader Topology Language (RRTL). The MARC allows hiding the particu- larities of every RFID reader vendor through a set of operations that are mapped to specific vendor API. The RRTL includes the description of RFID topologies including reader control and management. The MARC is the minimum set of instructions that an RFID reader must supply to support the operational instructions inside a DEPCAS network. The set of MARC is closed and give an independent layer of RFID management. The implementation of MARC is related to RFID reader API and gives an independent hardware layer. The RRTL is a language that includes primitives to define DEPCAS network. The sentences of this language include the necessary semantic to explain how the readers are organized inside the acquisition network. F O S S T R A C K R IF ID I A S P IR E O R A C L E IB M D E T E G O O A T R F ID P la t. D E P C A S Device Agnostic Yes/No No Yes Yes Yes Yes Yes Yes 1 Device Configuration Yes Yes Yes No No Yes Yes Yes Topological Configuration No No Yes Yes No Yes Yes Yes2 Needs Programming Starting-up Device Yes Yes Yes No No No No Yes/No3 Statistical Data from Device Yes No Yes No No Yes Yes Yes4 Monitoring Data from Device No No No Yes Yes Yes Yes Yes5 Diagnosis capabilities Yes No Yes No Yes No Yes Yes6 LLRP Implementation Yes Yes Yes No No No No No OSGi Technologies No Yes Yes No No No No Yes7 1 using MARC sentences to configure and connect RFID devices. 2 using RRTL. 3 depending on the API of the chosen reader. 4 statistical operations are included in MARC sentences. 5 MARC sentences include data quality evaluation if it is possible to obtain information from readers. 6 there are MARC sentences to produce device diagnosis and there are operations included in RRTL about topological diagnosis. 7 to process the parallel RFID readers. Table 1: DEPCAS MDM support compared to related word 3.2 MLM Current RFID middleware systems capture raw readings from readers, filter and aggregate the data, and finally yield cleansed RFID raw events. However, RFID raw events only provide low-level information such as EPC [24], reader identification and time–stamp, which are not directly related to business processes. Therefore, the ability to quickly transform these raw events into business events useful for decision-making becomes a central issue for realizing the maximum value from RFID technology in business. The key aspects to take in count in processing RFID information are: 1. Process data: the RFID raw process is more than a basic filtering, it is a cleansing, consolidation and summarization set of operations that have as objective to create and consolidate useful RFID data for business. 2. Alarm: one of the purposes of RFID middleware is to make difference between relevant and not relevant data in acquisition. The alarms are the first processed data that are generated inside the middleware. RFID middleware’s can generate two types of alarms: infrastructure alarming and pro- cess alarming. 9 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 3. Event: the RFID acquisition system should digest the meaningful events. The alarms are also event from this point of view but not only. The RFID middleware can define other relevant situations that are interesting for external application or to support the system exploitation. 4. Filter: the filter RFID process refers to the removal of certain tag read events based on any dis- criminator parameter as tag code repetition in a short period or the same tag code in two different readers. 5. Consolidate: the RFID middleware process will create a new RFID data that should be consolidated and maintained inside any database systems. This consolidation opens the way to share and to transmit to external applications the cooked RFID information. Table 2 compares the features supported by DEPCAS MLM to the related work summarized in Section 2. Existing RFID middleware includes a basic RFID data process. This basic process is related with event pro- cess and filter reads in the RFID middleware. Some of the existing middleware also propose an additional operation, called aggregation or summary reads. The basic idea under this operation is to reduce the tag reads using any criteria. For instance, aggregating operation can detect input/output read, selecting just the first and the last read when a tag appear under read range. Other approach is to count the tag reads and accumulate it. Other RFID include specific business modules. These modules perform business explicit process from the raw acquisition systems and offers direct connectors with the back end application. O R A C L E IB M F O S S T R A C K R IF ID I T ru e V U E D E T E G O D E P C A S Alarm No Yes No Yes Yes Yes Yes Event No Yes Yes Yes Yes Yes Yes Filter Process Yes Yes Yes Yes Yes Yes Yes Aggregation Process No Yes No Yes No Yes Yes/No1 Configurable Data Process No No No No Yes Yes Yes2 Consolidate Data Yes Yes No Yes Yes Yes Yes3 Specific Business Modules No No No Yes No Yes No4 Supply Chain Process Yes Yes Yes Yes Yes Yes Yes5 EPC Network Support Yes Yes Yes Yes Yes Yes Yes6 1 depending on every scenario process. 2 scenario configuration. 3 scenario process results. 4 scenario configuration applies to every business case. 5 using the scenarios according to the supply chain process definition. 6 using the scenario process. Table 2: DEPCAS MLM support compared to related word The main target in the conception of DEPCAS is to think about the use of homogenous RFID middleware that can be introduced in working systems that already use identification (and want to change to use auto identification). For instance, we are habituated to use the bar code in many circumstances in daily life, like shopping, luggage, personal cards, etc. All these heterogeneous systems that use identification in a half term will transform into systems that use the auto identification. The mechanism that solves this homogenization in DEPCAS is the scenario concept, which is equivalent to a logical process with a general application that it is possible to be used in a set of situations where the information takes advantage of auto identification. In each scenario there are RFID data that is received, processed, summarized and accumulated, according to their own logic. The basic defined scenarios in DEPCAS are: 10 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 1. Tracking of items: this scenario corresponds to an application in which it is wanted to make the tracking of some object in a chain that is supervised with RFID readers. The objects when happening through each point generate the identification events that are registered from an origin to a destiny. The different events will depend on the parameters that are desired to check: time, distance, speed, step/no step, complete route, etc. Section 5 details this scenario. 2. Items classification: this scenario consists of a classification of elements after an identification of an object. The read event from an object generates the classification according to certain criteria. The application of this scenario would be applicable to an automatic robotically classification system. or to a system of shipments classification. 3. Aggregation of items: this scenario would allow adding auto identified objects to generate a new information registry. This scenario can be applied in those cases in which a product is develop by parts that are added, like for example the creation of baskets of gifts, pallets, etc. 4. Format conversion: this scenario would allow generating information from auto identification codes. For instance, it could be applied for systems in which hybrid identifications are received and wanted to make uniform. This scenario allows to turn formats EPC to other formats of particular objects like the ISBN, ISSN, etc. 5. Execution of machine of states: the scenarios that represent state machines would allow associating information of auto identification to the identified object states. For instance, it could be used to make the track of a rent of vehicles or the management of loans in a library. 3.3 GUV GUV is a combination of configurable components, that graphically reflect the dynamic state of the simple components (tags, readers...) or complex ones (routes, adders, state machines...) obtaining different types of information from data servers. The creation of a specific application will be carried out by the grouping of these components, properly configured like a client data or on a web node for an Internet browser. The study of different RFID GUVs allows obtaining a set of characteristics that are implemented to solve the user interface in the RFID middleware. Among these features we include: 1. Web Client: the graphical component can be supported as lightweight software using web client applications. 2. Collection of dynamic information: GUV should include dynamic graphical components to show the status and quality of represented data. 3. Managing the database infrastructure of RFID: the database management tool is a graphical com- ponent that simplifies the database operations (i.e., search, create, delete or update) about RFID infrastructure (e.g., readers, tags...) 4. Managing the process database: the database management tool is a graphical component that simpli- fies the database operations (i.e., search, create, delete or update) about RFID process configuration. 5. Lists of information: the list graphical resource presents the database information like a summary where the user can choose the rows and the columns shown for an specific table. 6. Editing and management of graphical elements: the option to edit and manage the graphical resources allows building specific user environment depending on installation or user authorization policy. 11 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 7. Management and configuration of alarms: alarms are one of the most important resources on RFID middleware. GUV includes several mechanisms to present the alarming information: tables, sum- maries, flashing and sound signals, etc. 8. User management and privileges: these kinds of options allow defining graphical user organization and policy to show/not to show data to the users. Table 3 compares the features supported by DEPCAS GUV to the related work summarized in Section 2. The main objectives of GUV are: 1. Including a single common environment and information management in DEPCAS middleware. This environment is common to all management information system configuration data regardless of being related to infrastructure, the processing logic sets, or the EPCIS modules. 2. Managing a common graphical object to support information lists. These lists can be used like dynamic queries with real time information or static queries to show database information. 3. Supporting graphical displaying information about a particular object in real time. You may include in new graphics another graphical object with real time attributes. 4. Representing alarm system that can be configured. This treatment includes alarm management levels and the realization of automated operations 5. Defining the process and system events for the acquisition of auto identification data. 6. Managing graphical objects that are used and which allows their reuse in different graphical inter- faces. O R A C L E IB M S y b a se T ru e V u e F O S S T R A K W in R F ID R IF ID I D E P C A S Web Client Yes Yes No No Yes No No No Dynamic Information Yes Yes Yes Yes Yes Yes Yes Yes1 Configurable Dynamic Information No Yes Yes Yes No Yes Yes Yes2 RFID Database Management Yes Yes Yes Yes Yes Yes No Yes3 RFID Process Configuration No Yes Yes Yes No Yes No Yes3 Lists Yes Yes Yes Yes Yes Yes No Yes4 Graphical Edition No No Yes Yes No No Yes Yes2 Alarms No Yes Yes Yes No Yes No Yes5 Events Yes Yes Yes Yes Yes Yes No Yes6 User Management Yes Yes No No Yes No No Yes7 1 using graphical connection to real time DEPCAS database and DEPCAS repository. 2 using ediGUV to create and compound graphics. 3 using a uniform Database Management Tool (DMT). 4 dynamic and static graphical representation. 5 like an special dynamic list and connected to the real time database. 6 like an special static list and connected to the repository database. 7 using user and user groups to define operational capabilities. Table 3: DEPCAS GUV support compared to related word 12 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 3.4 EPCIS The EPC Information Service provides a specification framework to solve the external RFID middleware applications to capture, secure and access RFID data. The original EPC network architecture was defined in [25] like an “omniscient entity” to provide relevant information required from business application to RFID middleware. The main four functions included in EPCS in this initial proposal were: providing a long term repository of EPC event data, supplying an access point to higher-level information obtained from EPC process acquisition, solving the storage and access to detailed business information associated to EPC process acquisition, and providing transparent on-demand access to data attribute held in others systems. EPC data was interchanged using the Physical Markup Language (PML) [26], an XML schema for representing EPC–related data including auto identification information (e.g., codes, readers, antennas, time–stamp...) and business information (e.g., size, quantities, width...). These functions were so broad that a general confusion about how to implement them happened. These pending questions are got pass in the definite EPCIS specification definition [27] including an standard in- terface to allow EPC data interchange. This standard interface is characterized by: managing only historical data (real-time EPC data is decoupled from EPCIS), operating with cooked EPC related information (raw EPC data is processed in low EPC Network layers), and dealing only with enterprise systems environment. All this specific functions are solved using a three layer interface: the abstract data model layer, the data def- inition layer and the service layer. The abstract data model is a closed and defined interface that proposes a specification of general requirements for creating data definitions in the data definition layer. This data definition is called core event type. This structural organization in the interface allow to use standard vo- cabulary elements used by inter organizations and user specific vocabulary elements defined and meaning inside intra organizations. Despite the objective of all these implementations is the same: to connect business with RFID middle- ware, there are a set of characteristics that differ between specific solutions. This analysis introduces the following features: 1. Standard compliance. What is the compliance degree in the existing implementation according to EPC Network standard? 2. EPCIS permanent data model implementation. What is the existing repository implementation sup- ported by EPCIS solutions. 3. Security EPCIS control. There is a large range of security control levels in EPCIS data, from general user access to low granular data access. 4. Metadata management. The business process defines process data and specific attributes that are introduced in the EPCIS repository. 5. Supply chain support. The EPCIS implementation has specific elements to solve supply chain ques- tions. 6. Reports. The EPCIS includes report facility to data access using XML exporters. Table 4 compares the features supported by DEPCAS EPCIS o the related work summarized in Section 2. The main functionalities of EPCIS in DEPCAS are: 1. Publish/subscription manager. 2. Permanent historical data information. 13 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 O R A C L E IB M W in R F ID F O S S T R A C K D E P C A S Standard Compliance Partial Total Partial Partial No Data model Oracle Oracle Oracle MySQL Yes1 PostgreSQL DB2 SQLServer Web Server Jini WebSphere Unknown Apache Yes2 Security WebServer Propietary Remote Authentication No security security objects and authorization for query operations Metadata Using Specific XML No support Yes3 data in metadata framework for new repository model vocabulary type Supply chain support Yes Yes Specific Yes Yes4 adaptor Reports Not Business Not Not Not Reports included Intelligence included included included Reporting Tool EPCIS events Yes Events Rule Yes Yes5 and manager alerts exceptions 1 to solve the relation between business and MLM scenario. 2 as a publish/subscription server to the DEPCAS repository. 3 metadata captures the relationships between the RFID product-process data. 4 as a middleware connector. 5 to manage the external connection and to supervise the way connections are working. Table 4: DEPCAS EPCIS support compared to related word 14 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 3. Transfer in/out data between DEPCAS and business applications. 4. Security distribution of DEPCAS information. 5. XML data mapping. 6. Specific application connector with DEPCAS 4 A DEPCAS prototype DEPCAS basic implementation includes a prototype module for all its components. 4.1 MDM implementation The DEPCAS MDM core architecture prototype has been developed using TCP and HTTP for transporting reader protocol messages, while the message content can be either XML or Text. In addition, it supports synchronous and asynchronous messaging (through the reader’s protocol Notification Channels mecha- nisms). The MDM MARC sentences implementation includes the following operations: 1. setAntennaSequence: Set an order sequence to read from multiple antennas installed in one reader 2. setAutoFalseOutput: Order to autonomous function mode to ignore tag reads 3. setAutoFalsePause: Order to autonomous function mode to pause working 4. setAutoMode: Start/Stop autonomous reader functions 5. setAutoStartTrigger: Activate the trigger sending functions in reader autonomous function 6. setAutoStartTimer: Activate periodic timer reader function in reader autonomous function 7. setAutoStopTimer: Reset periodic timer reader function in reader autonomous function 8. setAutoStopTrigger: Reset the trigger sending functions in reader autonomous function 9. setAutoTrueOutput: Order to autonomous function mode to process tag reads 10. setAutoTruePause: Order to autonomous function mode to start reading process 11. setConnectTCP: Establish a TCP connection 12. setNotifyAddress: Init a listening TCP port to instruct the reader to send notification messages 13. setNotifyFormat: Configure the format message to be used by the reader 14. setNotifyTime: Establish the period that the reader notifier should use 15. setPassword: Identify the password reader access 16. setRun: Start to execute a configuration reader mode 17. setTagType: Establish the expected tag type property 18. setTime: Synchronize reader time 15 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 19. setUsername: Identify the user reader access The MDM prototype includes MARC implementation for RFID reader ALIEN8780, FEIG2000, SKYWRAEM2 and we are working in MARC implementation for ThingMagic Mercury 4, Intermec IF61 and Impinj Speed- way R420. 4.2 MLM implementation The general process in any of MLM configuration follows the general flow presented in Figure 2. MLM uses two main groups of configuration tables: the scenario configuration tables and the static configuration ta- bles. The scenario configuration tables include data structure about the specific scenario that is processed. For instance, in the tracking scenario (see Section 5 the most relevant table is the “routes” table that in- cludes information about the itinerary that items are going to be monitoring. The static configuration tables for MLM includes data about the general elements processed. For instance, what codes are going to be supervised (the who’s that DEPCAS is going to process) or the reader points in the systems (the where’s that DEPCAS logic manages). The MLM initial trigger is always a RR (relevant read) that always include a who (RFID code), a where (RFID reader point) or a when (RFID read time–stamp). RRs are processed one by one or can be processed by set of data by the RR Process. These processes generate data logs information from the processing result of incoming new RR and static configuration. For instance, if there is a RR with a who (RFID code) that is not inside the static configuration table the process generate a NOT_OK log and create the necessary data inside alarm, event or statistical tables. Figure 2: General process of DEPCAS MLM The scenario loader initializes the configuration parameter of each scenario. For instance, in the state machine scenario process the initial operation starts every state machine defined to the initial state, pro- ducing the ignorance of every read associated to other states. The Dynamic Scenario Tables define pro- prietary data structures for every particular scenario. These tables are managed by the dynamic scenario 16 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 process and they are scenario dependent. In the aggregation scenario, there are several scenario depen- dence process, for example the compositions table receives the items that are made up of the complete aggregation. When the items are processed they area accumulated to an specific aggregation. The notifier process communicates with the Business Event Process (BEP) to trigger it the events that are configured in every scenario. 4.3 GUV implementation Based on the concepts of high-level design have been shown we have developed a GUV prototype envi- ronment that allows building graphical environment for a specific domain. The prototype has a graphic editing system that allows creation, modification and deletion of graphical elements. This environment, called ediGUV defines a core container referred as DEPCASGUV in which may include other edited elements. The prototype operates two types of graphic elements: the basic elements that are provided directly by the window SDKs and the compound DEPCAS elements that are the result of basic elements compositions for a specific function in DEPCAS middleware. The basic graphic elements in the actual DEPCAS version are: 1. Windows 2. Buttons 3. Edit Fields 4. Text Fields 5. Check boxes 6. Combo boxes DEPCAS compound graphic elements are: 1. Panel database (Add-Delete-Modify Data) 2. List of dynamic database 3. List of static database 4. RFID / EPC Label 5. RFID Reader Point 6. Dynamic graph (Route / Forecast) 4.4 EPCIS implementation EPCIS is implemented using the following four subsystems: 1. Publish/Subscription Manager supports two kinds of relationships: relationships that are simple data representations of some actual relationship that exists between business atomic data and DEPCAS data, and relationships between tables that have their own properties. Data representations of rela- tionships can contain a Web service reference to access the relationship Web service. The EPCIS imple- mentation defines interfaces to query a manageable resource about the relationships it participates in directly, such as a containment relationship. The Publish/Subscription manager also defines an 17 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 interface for a service that offers relationship information for many manageable resources enabling a relationship registry. Manageable resources do not have to be aware that they are participating in relationships. 2. Converter Manager implements a global repository that captures the relationships between the RFID product-process data, that we call scenario in DEPCAS terminology, and existing business data. The converter manager layer offers the possibility to control the level of aggregation by customized com- bination of several built-in filters as well as those developed by specific environments. Hence a wide range of variation is at hand. Attaching meta-data from backend-systems to DEPCAS scenarios re- quires connecting specific existing external data to DEPCAS scenarios tables. 3. Feeder EPCIS interface works using a Java Message Service (JMS) to exchange data from DEPCAS real time processing scenarios to EPCIS process. This capturing interface distributes the information to the consolidate database and to synchronous ECPIS connectors. The message structure is defined using an XML schema. This schema structures the general data field of any message according to a specific scenario. For example, the JMSType for a tracking scenario includes: general position, specific place, building, area, reader used, department, auto id read. This message is instantiated as showed in Figure 3. 4. Query EPCIS server supports the DEPCAS servers/clients connections. It solves the different data interchange (HTTP, web services or XML) between DEPCAS and external business applications. In the same way, EPCIS server is designed as a platform with a uniform query and update interface to ap- plications, while the actual implementation solves the details and data binding to existing databases and information systems. EPCIS server supports simultaneous binding to multiple databases and information systems from multiple vendors, as well as EPCIS as a managed application service. In order to accommodate various types of data relationships such as routes, predictions, reader points, associations with particular transactions, EPCIS server is implemented as a modular framework, with a very lightweight core functionality, which merely described which “profiles” or relationship types are implemented on any particular instantiation providing an EPCIS interface. Startup time: 21.jan.2008 11:30:28:343 21.jan.2008 11:30:28:343 Queue name: jms/Queue Consumer Queue Selector: JMSCorrelationID = ’UNED’ AND JMSType LIKE ’ISSI_2’ To end program, type Q or q, then <return> Recv message: 21.jan.2008 11.38.01:531 Message header: 21.jan.2008 11:38:01:437 Producer-63, message order: 3 Data message: UNED, Campus 1, Edificio A, Zona 3 Reader_A 00 ISSI_2 0000 1111 2222 3333 4444 8391 1981 Recv message: 21.jan.2008 11:38:01:562 Message header: 21.jan.2008 11:38:01:531 Producer-63, message order: 9 Data message: UNED, Campus 1, Edificio A, Zona 3 Reader_A 00 ISSI_2 0000 1111 2222 3333 4444 7134 2187 Figure 3: Example of EPCIS message 5 An example of DEPCAS scenario The tracking of people or goods with identification mechanisms is repeated in heterogeneous form in diverse situations. The tracking operation abstraction can be considered in terms of a route to monitor 18 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 and the parameters to process in this route, that we call predictions. The fundamental software function is tracking objects that travel by the defined routes. The process for each case depends on the type of parameter assigned to the route. The different types of routes make the corresponding verifications in each case: 1. BDR (Basic Data Routes): they are fixed routes in which the passage in a certain absolute time is verified. 2. BDTR (Basic Data Time Routes): they are fixed routes in which the passage in a certain relative time relative to the origin is verified. 3. ADR (Alternative Data Routes): they are alternative routes in which one verifies the origin and the destiny in a certain absolute time. 4. ADITR (Alternative Data Intermediates Time Routes): they are alternative routes in which one verifies the alternative origin, destiny and intermediate steps in an absolute time. 5. ADRTR (Alternative Data Relative Time Route): they are alternative routes in which one verifies the origin and the destiny according to a time relative to the origin. 6. ADRTITR (Alternative Data Relative Time Intermediate Time Routes): they are alternative routes in which one verifies the alternative origin, destiny and intermediate steps in a time relative to each made step. 7. UPUR (Undefined Predictions Undefined Routes): routes of any type for which any type of prediction algorithm is not defined. The different processes generate the warnings that determine if the readings correspond to the pre- dicted thing or they do not correspond and why. Between the system warnings we can distinguish three groups. The first group are the warnings generated in correct tracking, we denominated LOG_OK. A second group is related with a correct tracking in which some type of incidence takes place, like for example not fulfilling a certain part of a route or for those cases in which the route is fulfilled with some time delay error. The third group of warnings are those generated in case of breaks of established routes, an example of this type are the LOG_LR_NOK warning that is generated in the case of appearing labels in routes in which they would not have to appear. In order to verify the operation of this type of scenarios, a monitor has been developed that shows all information about the labels tracking. 6 Conclusions Research on RFID middleware is oriented to find new alternatives to efficiently include auto–identification into productive processes. New topological networks of tag readers and heterogeneous business applica- tions need to be connected through RFID middleware, solving all the established requirements. In this paper, we have summarized the features that must be supported by any “real” RFID middleware. As a result, we propose the DEPCAS middleware, which starting point is the well known SCADA architec- ture. After describing the theoretical basis of our approach, i.e., the architecture of DEPCAS, its current implementation and an application scenario have been reported. 19 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 Acknowledgements This work has been supported by (i) the Spanish Government under the CICYT projects DPI-2008-05444, DPI-2011-26094 and DPI-2011-24588, (ii) the Comunidad de Madrid under the RoboCity2030-II excellence research network S2009/DPI-1559, and (iii) the Universidad Nacional de Educacion a Distancia under grant 2010V/PUNED/004 References [1] J. J. Jung, Ubiquitous conference management system for mobile recommendation services based on mobilizing social networks: A case study of u-conference, Expert Systems with Applications 38 (10) (2011) 12786 – 12790. [2] X. Su, C.-C. Chu, B. Prabhu, R. Gadh, The Internet of Things: From RFID to the Next-Generation Per- vasive Networked Systems (Wireless Networks and Mobile Communications), Auerbach Publications, 2008, Ch. On the creation of Automatic Identification and Data Capture infrastructure via RFID and other technologies. [3] C. Floerkemeier, C. Roduner, M. Lampe, Rfid application development with the accada middleware platform, IEEE Systems Journal 1 (2) (2007) 82–94. [4] Q. Z. Sheng, X. Li, S. Zeadally, Enabling Next-Generation RFID Applications: Solutions and Challenges, Computer 41 (9) (2008) 21–28. [5] A. Ustundag, M. S. Kilinç, E. Cevikcan, Fuzzy rule-based system for the economic analysis of rfid investments, Expert Systems with Applications 37 (7) (2010) 5300 – 5306. [6] C. Lee, T. Chan, Development of rfid-based reverse logistics system, Expert Systems with Applications 36 (5) (2009) 9299 – 9307. [7] C. Bornhövd, T. Lin, S. Haller, J. Schaper, Integrating automatic data acquisition with business pro- cesses experiences with sap’s auto-id infrastructure, in: Proceedings of the Thirtieth international conference on Very large data bases - Volume 30, VLDB ’04, VLDB Endowment, 2004, pp. 1182–1188. [8] J. Curtin, R. Kauffman, F. Riggins, Making the ŚmostŠ out of rfid technology: a research agenda for the study of the adoption, usage and impact of rfid, Information Technology and Management 8 (2007) 87–110. [9] M. He, S.-J. Horng, P. Fan, M. K. Khan, R.-S. Run, J.-L. Lai, R.-J. Chen, A fast rfid tag identification algorithm based on counter and stack, Expert Systems with Applications 38 (6) (2011) 6829 – 6838. [10] A. Brewer, N. Sloan, T. L. Landers, Intelligent tracking in manufacturing, Journal of Intelligent Manu- facturing (1999) 245–250. [11] C.-S. Lin, L.-S. Chen, C.-C. Hsu, An innovative approach for rfid product functions development, Expert Systems with Applications 38 (12) (2011) 15523 – 15533. [12] Y.-C. Lee, S.-S. Lee, The valuation of rfid investment using fuzzy real option, Expert Systems with Applications 38 (10) (2011) 12195 – 12201. [13] N. Kefalakis, N. Leontiadis, J. Soldatos, D. Donsez, Middleware building blocks for architecting rfid systems, in: MOBILIGHT, 2009, pp. 325–336. 20 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.03.045 [14] Zafer, Aydogmus, Implementation of a fuzzy-based level control using scada, Expert Systems with Applications 36 (3, Part 2) (2009) 6593 – 6597. [15] C.-C. Hsu, P.-C. Yuan, The design and implementation of an intelligent deployment system for rfid readers, Expert Systems with Applications 38 (8) (2011) 10506 – 10517. [16] C. Cheung, C. Cheung, S. Kwok, A knowledge-based customization system for supply chain integration, Expert Systems with Applications 39 (4) (2012) 3906 – 3924. [17] C. Lee, W. Ho, G. Ho, H. Lau, Design and development of logistics workflow systems for demand management with rfid, Expert Systems with Applications 38 (5) (2011) 5428 – 5437. [18] J. M. Ko, C. Kwak, Y. Cho, C. O. Kim, Adaptive product tracking in rfid-enabled large-scale supply chain, Expert Systems with Applications 38 (3) (2011) 1583 – 1590. [19] M.-C. Kim, C. O. Kim, S. R. Hong, I.-H. Kwon, ForwardŰbackward analysis of rfid-enabled supply chain using fuzzy cognitive map and genetic algorithm, Expert Systems with Applications 35 (3) (2008) 1166 – 1176. [20] X. Huang, S. Le, D. Sharma, A taxonomy for rfid systems, in: IEEE International Conference on Signal Processing and Communication Systems, Gold Coast, Australia, 2007. [21] F. Bo, L. Jin-Ta, Z. Ping, G. Jun-Bo, D. Zhen-Hua, Study of rfid middleware for distributed large-scale systems, in: Information and Communication Technologies, 2006. ICTTA ’06. 2nd, Vol. 2, 2006, pp. 2754 –2759. [22] B. S., X. Su, H. Ramamurthy, C. cheng Chu, R. Gadh, Winrfid Ű a middleware for the enablement of radio frequency identification (rfid) based applications, in: in Mobile, Wireless and Sensor Networks: Technology, Applications and Future, John Wiley & Sons, Inc, 2005, pp. 331–336. [23] H. K. Chow, K. L. Choy, W. Lee, K. Lau, Design of a rfid case-based resource management system for warehouse operations, Expert Systems with Applications 30 (4) (2006) 561 – 576. [24] Eun-Jun, Yoon, Improvement of the securing rfid systems conforming to epc class 1 generation 2 standard, Expert Systems with Applications 39 (1) (2012) 1589 – 1594. [25] M. Harrison, Epc information service–data model and queries, Tech. rep., Auto–Id Centre Institute For Manufacturing, University Of Cambridge (2003). [26] C. Floerkemeier, D. Anarkat, T. Osinski, M. Harrison, Pml core specification 1.0, Tech. rep., Auto–Id Centre Institute For Manufacturing, University Of Cambridge, (2003). [27] EPCglobal, Epc information services (epcis) version 1.0 specification, Tech. rep. (2007). 21 Introduction Related work DEPCAS MDM MLM GUV EPCIS A DEPCAS prototype MDM implementation MLM implementation GUV implementation EPCIS implementation An example of DEPCAS scenario Conclusions 
work_nfibvkgxobfxfk4btcxuz6hr3u	----	Dipmeter Advisor Expert System: ABSTRACT Association Round Table 1703 Recognition of the turbidite nature of many petroleum- producing sands worldwide makes it imperative to understand both the facies types encountered and the trap types to be antici- pated in future exploration and development work. Several facies models have been proposed, most of which are not read- ily usable with the data generally available in subsurface work. The facies analysis presented examines large-scale controls, such as climate, provenance, basin geometry, and tectonics, and then considers the various large to small-scale facies and sand-body geometries that result. Use of wire-line log pattern, dipmeter, and core descriptive data as criteria is emphasized. The newly recognized meander channel facies is shown to be important in prograded muddy slope areas. The concept of a "facies" as referring to a mappable assemblage of beds of vary- ing natures and origins is favored, in contrast to the practice of assigning a separate facies designation to each single successive bed, with resulting uncertainty as to overall significance. The author's trap-type classification, which includes canyon- dependent, fan-dependent, anticline-dependent, fault- dependent, and uplift topography-dependent traps, is used in conjunction with the facies analysis to predict the trap types most likely to be encountered in the various facies and basin set- tings. The trap classification itself is developed as a predictive tool rather than as a pigeonholing exercise. iiacc lornpositing, and two-dimensional, area! stacking of the data in order lo identify anomalies. An anomaly was clearly dis- cernible resulting from the rubble-collapse void in the burn /.one. The anomaly was studied in detail and compared to syn- thetic models. The fault system was found to be a graben com- plex with antithetic faults. The fault system contains folded beds. A series of anomalies was discovered on the northeast end of one of the seismic lines. These reflections have been identi- fied as underground mine adits from the old Hanna No. 1 coal mine. AAPG RESEARCH CONFERENCE Interactive Computer Applications in Petroleum Exploration Keystone, Colorado May 9-12,1982 Abstracts BANKS, RICHARD B., and JOSEPH K. SUKKAR, Scientific Computer Applications, Inc., Tulsa, OK WHITTAKER, ALUN, and RAFAEL GURVIS, Exploration Logging Inc. Wellsite Geochemical Analysis in Frontier E x p l o r a t i o n - Logistics, Benefits, and Examples In the past 10 years, organic geochemistry has become an increasingly significant factor in petroleum exploration. In frontier exploration areas, the need for geochemical data dur- ing drilling operations, rather than months after a well is fin- ished, is now well appreciated. Real-time geochemical data have proven to be an important additional parameter in explo- ration and well completion decisions. Wellsite geochemistry mandates unique operational, techni- cal, and data interpretation requirements. However, the timely nature of data availability during drilling operations as well as the ability for field geochemists to work closely with other well- site personnel greatly enhances exploration efficiency. Sample procurement, preparation, and selection can be much more accurately realized by personnel familiar with local geology and a specific drilling operation. Furthermore, sample contamina- tion by drilling fluid additives is more easily prevented, detected, and isolated from fresh samples at the wellsite. Complex Subsurface Analysis Using Interactive Modeling and Simulation Techniques Many offshore producing regions (and some onshore regions, such as Prudhoe Bay) are complicated by dual neme- ses: intersecting non-vertical faults and directional well bores. Fault occurrences are observed in well bores or on seismic lines. When many fault occurrences are present, it often is difficult to determine which fault cuts are associated with which faults. An interactive contouring system allows the user to try various ways of connecting fault occurrences to form (intersecting) fault surfaces in (subsurface) space. Once the faults have been modeled, the next task is to model (reconstruct, contour) subsurface formations as they interact with the various faults. This task is difficult enough when there are many formations and many faults. The task is even more difficult when directional wells are involved, since isopachs needed for reconstruction (i.e., stacking) are hard to deter- mine. A multi-surface multi-fault contouring system is being used to perform these complex subsurface analyses. A case study from offshore Gulf Coast illustrates its use. CARLBOM, INGRID, Schlumberger-Doll Research, Ridge- field, CT YOUNGBERG, A. D., Laramie Energy Technology Center, Laramie, WY, and E. BERKMAN and A. ORANGE, Emerald Exploration Consultants, Inc., Austin, TX Location of Burns and FauUs at Hanna Underground Coal Gasification Area by Use of High-Resolution Seismic Survey In November 1980, a high-resolution seismic survey was con- ducted at the Hanna underground coal gasification area. The objectives of the survey were to locate and characterize the burn cavity at the Hanna II, phases 2 and 3 gasification site, and to locate shallow geologic faults. Seismic data acquisition and processing parameters were specifically designed to emphasize reflections at the shallow, 200 to 500 ft (60 to 160 m) depths. A three-dimensional grid of data was obtained over the burn zone. Processing included time-varying filters, deconvolution. Dipmeter Advisor Expert System The dipmeter services offered by Schlumberger consist of measurements made in a well bore to determine the inclina- tions, or dip, of the bedding layers penetrated by the well. The measurements are represented on a log in the form of arrow plots, which indicate the magnitude and direction of the dip as a function of depth. Information about the underground struc- ture and stratigraphy can be derived from arrow plot patterns on the dipmeter log in conjunction with other types of data from the borehole as well as some general knowledge of the geo- logical area. Traditionally, interpretation of dipmeter data has been made by a human expert who identifies and decodes the arrow plot patterns. The Dipmeter Advisor system is an application of artificial intelligence and expert system techniques to dipmeter log inter- 1 7 0 4 Association Round Table pretation. It incorporates the inference and knowledge of human dipmeter experts. The first prototype system, which was completed in December 1980, consisted of interpretation in marine environments, but recently the system has been extended to continental and transitional environments as well. The Dipmeter Advisor system has a highly interactive graph- ics user interface. During the interpretation, the user has access to the arrow plot data and other related well data through a graphics display screen. The interpretation is made in a sequence of passes over the data, each pass arriving at some conclusions based on user input, combined with applications of the rules of dipmeter interpretation and of pattern recognition algorithms in previous passes. The partial results are displayed on the graphics screen for user verification and the user can, with graphics interaction devices, make any additions, dele- tions, and modifications of the results in each pass. The final output of the system consists of a log annotated much in the same way as dipmeter logs are currently annotated by human experts. Several aspects of the Dipmeter Advisor system will be described: the system organization, the graphics interface, and the general form of the rules and the inference structure. The operation of the Dipmeter Advisor system will be demonstrated with an example of dipmeter interpretation. CRANE, DAVID C , Consultant, Dallas, TX Microcomputers in Exploration—A Survey The author has contacted about 50 geoscientists who are using "personal" microcomputers (micros) in their profession. They include respondents to notes in the AAPG Explorer, peo- ple referred by members of the AAPG Committee on Com- puter Applications to Geology, and numerous other members of the AAPG and SEG whose assistance is gratefully acknowl- edged. Microcomputers have been widely available for only about five years. Software sophisticated enough to realize their poten- tial has been slow to appear, but a surprising number of geosci- entists, including independents and consuhants, now wonder how they ever survived without computers. Reported applica- tions areas range from word-processing, production account- ing, and financial analysis of prospects to log analysis and creation of synthetic seismograms. System costs range from under $2,000 to over $30,000plus software. Explorationists involved in the rapidly growing use of micros are invited to submit their names, addresses, applications, sys- tem descriptions, and interests to the Committee on Computer Applications to Geology for inclusion in the next survey mail- ing. The first survey and mailing list are being made available as part of this paper. DOWNING, JAMES A., ZYCOR, Inc., Austin, TX interactive control for manual interpretation of data hold the potential for dramatically improving results and expanding acceptance of the technology. Several conventional algorithms are reviewed and two new gridding techniques are introduced. Procedures for interactively controlling gridding techniques and adapting the techniques to respond to manual interpreta- tions of the data will be discussed. Also, procedures for intrac- tively adjusting gridded surfaces to conform to manually input contour curves will be described and demonstrated. DUDA, RICHARD, Fairchild Laboratory for Artificial Intel- ligence, Palo Alto, CA An Overview of Rule-Based Expert Systems One of the more successful apphcations of artificial intelli- gence techniques has been the development of programs that have come to be known as rule-based expert systems. These programs encode knowledge about specialized problem areas in the form of sets of if-then rules, typically obtained by inter- viewing people who are specialists or experts in those problems. The rules can then be used by the program to solve similar prob- lems. The modular structure that results allows incremental development, leading to performance that continually improves as the rule base is expanded. Most of the rule-based expert systems that have been devel- oped to date have been designed for problems in medical diag- noses. However, among several efforts that are relevant to the petroleum industry, a program called Prospector has been developed for the U.S. Geological Survey for certain problems in hard-rock mineral exploration, and a program called the Dipmeter Adviser has been developed by Schlumberger for the geological interpretation of dipmeter data. This presentation will describe the basic principles behind all such systems and will summarize the current state of the art. DUPREE, R. L., Amoco Production Co., Tulsa, OK Highly Interactive Contouring Systems Computer contour systems are intended to aid in the inter- pretation of geologic data as well as to prepare drafted quality displays. To accomplish this, completely automated (batch) contour systems require complex algorithms and a significant amount of computer resources. Multiple submissions are also usually required to obtain a finished display. A highly interactive contour system, however, rehes much more heavily on the interpreters and the graphical functions of an automated drafting system to prepare a finished display. This approach uses simpler algorithms, less computer resource, and more interpretative interaction from the end user. Exam- ples of the two approaches include the use of color schemes for displaying results. Interactive Gridding Gridding of geologic or geophysical data is at the foundation of most computerized mapping and modeling operations. Sur- face display techniques such as contouring, fish-net isometrics, and cross sections are usually derived from gridded surface models. Processing and analysis techniques such as surface fil- tering, trend analysis, Fourier transforms, and simple algebraic combinations of surfaces use gridded models in intermediate steps. A combination of conventional gridding techniques with HILDEBRAND, H. A., CYBERAN Corp., Houston, TX A Microcomputer Workstation for Interactive Geology and Geophysics The capabilities of a new microcomputer system are reviewed. This system allows easy data management for both geologists and processing geophysicists. Currently available peripherals include a digitizer, pen plotter, raster graphics printer, and graphics video. Communications allow hardwired or modern access to other computers. 
work_ngt5vnd2krhozmk545tn6pzlqa	----	Attenzione: i dati modificati non sono ancora stati salvati. Per confermare inserimenti o cancellazioni di voci è necessario confermare con il tasto SALVA/INSERISCI in fondo alla pagina nascondi/visualizza icone a destra nascondi/visualizza menu in alto Aiuto Sfoglia Scorri i prodotti per: Autore Titolo Riviste Serie  Login PORTO @ Archivio Istituzionale della Ricerca IRIS è la soluzione IT che facilita la raccolta e la gestione dei dati relativi alle attività e ai prodotti della ricerca. Fornisce a ricercatori, amministratori e valutatori gli strumenti per monitorare i risultati della ricerca, aumentarne la visibilità e allocare in modo efficace le risorse disponibili. IRIS Pol. Torino Catalogo POLITO 1 Contributo su Rivista 1.1 Articolo in rivista Misleading Generalized Itemset discovery Italiano Italiano English Frequent generalized itemset mining is a data mining technique utilized to discover a high-level view of interesting knowledge hidden in the analyzed data. By exploiting a taxonomy, patterns are usually extracted at any level of abstraction. However, some misleading high-level patterns could be included in the mined set. This paper proposes a novel generalized itemset type, namely the Misleading Generalized Itemset (MGI). Each MGI represents a frequent generalized itemset X and its set E of low-level frequent descendants for which the correlation type is in contrast to the one of X. To allow experts to analyze the misleading high-level data correlations separately and exploit such knowledge by making different decisions, MGIs are extracted only if the low-level descendant itemsets that represent contrasting correlations cover almost the same portion of data as the high-level (misleading) ancestor. An algorithm to mine MGIs at the top of traditional generalized itemsets is also proposed. The experiments performed on both real and synthetic datasets demonstrate the effectiveness and efficiency of the proposed approach. Misleading Generalized Itemset discovery / Cagliero L.;Cerquitelli T.;Garza P.; Grimaudo L.. - In: EXPERT SYSTEMS WITH APPLICATIONS. - ISSN 0957-4174. - 41:4(2014), pp. 1400-1410. fake_placeholder_label_hidden fake_placeholder_label_hidden La pubblicazione è stata scelta per una campagna VQR Scheda breve Scheda completa Titolo:  Misleading Generalized Itemset discovery Autori:  CAGLIERO, LUCA CERQUITELLI, TANIA GARZA, PAOLO GRIMAUDO, LUIGI Data di pubblicazione:  2014 Rivista:  EXPERT SYSTEMS WITH APPLICATIONS   Digital Object Identifier (DOI):  http://dx.doi.org/10.1016/j.eswa.2013.08.039 Appare nelle tipologie: 1.1 Articolo in rivista File in questo prodotto: File Descrizione Tipologia Licenza   2515905_draft.pdf  1. Preprint / Submitted Version PUBBLICO - Tutti i diritti riservati Visibile a tuttiVisualizza/Apri Utilizza questo identificativo per citare o creare un link a questo documento: http://hdl.handle.net/11583/2515905 I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione. × informazioni i dati relativi al numero di citazioni sono recuperati in tempo reale dai servizi offerti da Scival di Elsevier, Pubmed-central e da WOS. Il grafico mostra l'andamento del dato citazionale di SCOPUS/WOS. I dati possono differire da quelli visualizzati in reportistica. Chiudi × informazioni i dati relativi ai percentili sono recuperati in tempo reale dai servizi offerti da Scival di Elsevier e da WOS. Si segnala che il miglior percentile visualizzato da IRIS è il primo (il più basso, secondo quanto implementato da Scival/WOS). E' possibile visualizzare l'elenco di tutte le categorie/percentili muovendo il mouse sopra al numero di percentile visualizzato. Chiudi       Powered by IRIS-about IRIS-Utilizzo dei cookie  Copyright © 2021  × Errore Errore × simulazione ASN Il report seguente simula gli indicatori relativi alla propria produzione scientifica in relazione alle soglie ASN 2018-2020 del proprio SC/SSD. Si ricorda che il superamento dei valori soglia (almeno 2 su 3) è requisito necessario ma non sufficiente al conseguimento dell'abilitazione. La simulazione si basa sui dati IRIS e sugli indicatori bibliometrici alla data indicata e non tiene conto di eventuali periodi di congedo obbligatorio, che in sede di domanda ASN danno diritto a incrementi percentuali dei valori. La simulazione può differire dall'esito di un’eventuale domanda ASN sia per errori di catalogazione e/o dati mancanti in IRIS, sia per la variabilità dei dati bibliometrici nel tempo. Si consideri che Anvur calcola i valori degli indicatori all'ultima data utile per la presentazione delle domande. La presente simulazione è stata realizzata sulla base delle specifiche raccolte sul tavolo ER del Focus Group IRIS coordinato dall’Università di Modena e Reggio Emilia e delle regole riportate nel DM 589/2018 e allegata Tabella A. Cineca, l’Università di Modena e Reggio Emilia e il Focus Group IRIS non si assumono alcuna responsabilità in merito all’uso che il diretto interessato o terzi faranno della simulazione. Si specifica inoltre che la simulazione contiene calcoli effettuati con dati e algoritmi di pubblico dominio e deve quindi essere considerata come un mero ausilio al calcolo svolgibile manualmente o con strumenti equivalenti. Annulla procedi × Modifica foto x Informazioni Si consiglia il caricamento di immagini con una proporzione 1-1 tra larghezza e altezza. La dimensione ottimale è 160x160 pixel 
work_nhd4rutu5rb53mugyx2ssa5egu	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220977252 Params is empty 220977252 exception Params is empty 2021/04/06-03:05:18 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220977252 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:18 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_nig5aub7bvgs7j6n5ekcacnbhu	----	University Homepage School Homepage School Contacts School Search Jump to the Web Login instructions Jump to the error message (if present) Jump to the ID and Password section Authentication Required You are connecting to a School of Informatics website which requires authentication. Please enter your username (either DICE username, or iFriend account name) and password to continue. This page displays best when JavaScript is enabled in your web browser. Hide this Information Back to Top Need an Account? If are a member of staff, or a student, contact Support for details of how to access your DICE account Other users may obtain an iFriend guest account which allows access to certain services. Forgot Your Password? Users with iFriend guest accounts may reset their password. All other users should contact Support  ID & Password Authentication Complete Login Password   Caps Lock is on Need an account or password help?     Important Information The School of Informatics uses the cosign single sign-on web authentication system. This will set a small number of session cookies in your browser as an essential part of the login mechanism. If you choose not to accept these you will not be able to view our authenticated web pages. If you do not wish to accept these cookies, do not complete the login process. Information on the cosign system is available here. By using this service you agree to adhere to the University of Edinburgh Computing Regulations.   Informatics Forum, 10 Crichton Street, Edinburgh, EH8 9AB, Scotland, UK Tel: +44 131 650 2690, Fax: +44 131 651 1426, E-mail: hod@inf.ed.ac.uk Please contact our webadmin with any comments or corrections. Unless explicitly stated otherwise, all material is copyright © The University of Edinburgh 
work_nj5zqb6v3ndx3jkckvkqfndi7q	----	Comparing Rankings from Using TODIM and a Fuzzy Expert System Procedia Computer Science 55 ( 2015 ) 126 – 138 1877-0509 © 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Organizing Committee of ITQM 2015 doi: 10.1016/j.procs.2015.07.019 ScienceDirect Available online at www.sciencedirect.com Information Technology and Quantitative Management (ITQM 2015) Comparing rankings from using TODIM and a fuzzy expert system Valério Antonio Pamplona Salomonb*, Luís Alberto Duncan Rangela aUFF, Av. dos Trabalhadores 420, 27255-125, Volta Redonda, RJ, Brazil bUNESP, Av. Ariberto Pereira da Cunha, 333, 12.516-410, Guaratingueta, SP, Brazil Abstract TODIM is, in its original formulation, an MCDA method developed to solve ranking problems. As an MCDA method TODIM combines the use of a multi-attribute value function as well as elements of the Outranking Approach, being founded on Prospect Theory. Recent advances in TODIM incorporate concepts from Fuzzy Sets. Although modelling multi- criteria decision problems with Fuzzy Sets has been utilized when the available data are imprecise, their use in MCDA is slightly controversial, because the data fuzzification can invalidate the outcome. Following a mixed qualitative-quantitative research strategy, our aim is to prove that for the ranking problems, TODIM can provide better solutions than Fuzzy Sets. Ranks from TODIM are linear, or strong, in a sense that it has no ties between the alternative solutions. The rank obtained with a Fuzzy Expert System can be weaker, that is, it may be a number of ties. The research strategy extends this result to ranking problems with the occurrence of crisp criteria. © 2015 The Authors. Published by Elsevier B.V. Selection and/or peer-review under responsibility of the organizers of ITQM 2015 Keywords: Fuzzy Expert Systems; Ranking Problem; TODIM 1. Introduction “On any one day people face a plethora of different decisions” [1]. This way, Multi-Criteria Decision Analysis (MCDA) methods have been developed to support decision makers in their decision problems. One reason for different MCDA methods is that there are different decision problems. First classifications of decision problems are Discrete Problem versus Continuous Problem. A Discrete Problem involves a discrete set of alternative solutions. A Continuous Problem involves a case where the number of possible alternatives are infinite [2]. There are four types of Discrete Problem: selecting an alternative solution (Choice Problem), * Corresponding author. Tel.: +55-12-3123-2232; fax: +55-12-3123-2468 E-mail address: salomon@feg.unesp.br © 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Organizing Committee of ITQM 2015 http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.07.019&domain=pdf 127 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 grouping alternatives (Sorting Problem), ordering alternatives from best to worst (Ranking Problem), or better descripting alternatives (Description Problem) [3]. This work focuses the Ranking Problem. The large number of MCDA methods engendered classifications for the methods. American School and European School [4] are, perhaps, the most well-known. These classifications are criticized, not only for xenophobia, but also to difficult developments by international teams [5]. Aggregation Approach and Outranking Approach are better classifications. However, both sets of classification shave often the same result. For instance, Analytic Hierarchy Process (AHP) and Multi-Attribute Utility Theory (MAUT) are MCDA methods for the Aggregation Approach, and they are from American School [6-7]; Elimination et Choix Traduisant la Realité (ELECTRE) and Preference Ranking Organization Method for Enriched Evaluation (PROMETHEE) are MCDA methods for Outranking Approach, and they are from European School [8-9]. An exception is Measuring Attractiveness by a Categorical Based Evaluation Technique (MACBETH), which is a MCDA method for Aggregation Approach, and is from European School [10]. Matter of fact, the choice of an MCDA method shall be based on characteristics from the decision problem, including necessary data and expected results. Nevertheless, the choice for an MCDA method have been a matter of opinion. This way, decision maker chooses to apply a familiar MCDA method to solve a new decision problem. Since decision maker has already applied this method, the new application gains in feasibility. On another way, the decision maker may choose a not familiar method. Or else, a method never applied before, just to expand decision maker’s knowledge in MCDA practice. Different MCDA methods may yield different results when applied to the same problem [11]. Still, a single method application can lead to different ranks. This can be a result from different individuals providing data, or it can be resulted from time-lapses in data collection. This work addresses the divergence between ranks from different MCDA applications with the concepts of rank correlation [12]. Multi-Criteria Interactive Decision-Making (shorted as TODIM, from Portuguese) is an MCDA method [13] developed to solve the Ranking Problem, TODIM combines elements from both Aggregation Approach and Outranking Approach. It first application was to rank projects with environmental impacts. Later, TODIM incorporated elements from Prospect Theory [14-15]. The previous case on environmental impacts was analyzed. The ranks with Prospect Theory diverge from the original rank. However, the ranks have some degree of correlation. The most well-know TODIM application was to rank residential properties [16]. It was recently revisited, considering criteria interactions [17]. The ranks with criteria interactions and with TODIM also diverge each other. However, the top two alternatives are the same from both applications. That is, the ranks are correlated each other. Recent advances in TODIM incorporate concepts from Fuzzy Sets [18-19]. Fuzzy Sets Theory (FST) was firstly proposed for the Classification Problem [20]. “A fuzzy set is a class of object with a continuum of grades of memberships. Such a set is characterized by a membership function which assigns to each object a grade of membership ranging between zero and one” [21]. In Classical Sets Theory (CST), sets are crisp. That is, an element belongs to a set or not. Then, when the available data are imprecise, FST is expected to better solve Classification Problem than CST. Therefore, new versions of classical MCDA methods were developed to incorporate FST [22-24]. FST have also been successfully applied to the Ranking Problem [25-26]. However, the use of Fuzzy Sets in MCDA is slightly controversial [27]. When these data are allowed to vary in choice over the values of a scale, as in AHP, these data are themselves already fuzzy [28]. Then, data fuzzification can invalidate the outcome. On the same issue, in real life crisp sets do exist [29]. Therefore, the use of Fuzzy Sets may result in loss of information, when transforming precise data in imprecise information. This work takes place on the side of questioning the indiscriminate use of FST in MCDA. Our aim is to prove that, for the Ranking Problem, TODIM can provide a better solution than FST. A mixed qualitative- quantitative research strategy [30] was adopted. That is, the goal is not to consider exhaustive cases. 128 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 Next section presents concepts on rank correlation, TODIM, and FST. Section 3 presents a case of Ranking Problem from a Brazilian real estate market. This problem involves fifteen alternatives and eight criteria, with a crisp criterion. The rank from TODIM is linear, or strong, in a sense that it has no ties between the alternative solutions. The rank obtained with a FST is weaker, that is, it has a number of ties. The research strategy extends this result to ranking problems with the occurrence of crisp criteria. Section 4 presents some conclusions and proposal for future researches. 2. Theory Background 2.1. Correlation between ranks As observed in Section 1, different MCDA methods may yield different ranks to the same problem [11]. The rank correlation coefficient is a measure of ranks agreement [12]. For two ranks of n elements, A and B, Kendall coefficient, b, is obtained by Equation 1, when aij and bij are score matrices obtained by Equation 2. 1 1 1 nn ba n i n j ijij b (1) tiedare and objects0 object of behind ranked is object 1 object of ahead ranked is object 1 ji ji ji aij (2) Two identical ranks will have agreements in all positions, aij = bij, then b = 1. For opposite ranks, that is two ranks with no agreement on any position, b = –1. For instance, if A = (1, 2, 3, 4), B = (1, 3, 2, 4), and C = (4, 3, 2, 1), we have b (A, B) ≈ 0.67, b (A, C) = –1, and b (B, C) ≈ – 0.67. Ranging from –1 to 1, Kendall coefficient measures the closeness of correspondence between two given ranks. In other words, it measures the compatibility between two ranks [31]. Kendall coefficient performs satisfactory to linear ranks. However, it presents some difficulties with partial and weak ranks. Linear ranks has no ties; weak ranks permits ties; two or more ranks are partial when at least one of them is incomplete, in the sense that each ranker may not necessarily rank all of the objects [32]. Emond-Mason coefficient, x, differs from Kendall coefficient by using a value of one for aij and bij to represent ties instead of the value of zero used by Kendall’s. By extending this interchange to acco mmodate ties, x is not flawed as b for weak ranks or for partial ranks [12]. An MCDA application may result weak ranks; a Fuzzy System may result weak ranks, too. Then, x will be adopted in this work, instead of b. 2.2. TODIM method TODIM has similarities with other MCDA methods as ELECTRE [33] and PROMETHEE [34]. However, while practically all other MCDA methods start from the premise that the decision maker always looks for some maximum overall value, TODIM method makes use of a measurement of overall value calculable according to Prospect Theory. TODIM application requires numerical values for the evaluation of the alternatives regarding the criteria. For qualitative criteria, alternatives can be evaluated in a verbal scale, but it must be then transformed into a cardinal scale. The numerical evaluation for the alternatives regarding to all the criteria composes the matrix of 129 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 evaluation. This matrix must be normalized, for each criterion: the value for one alternative must be divided by the sum of values for all the alternatives. This way, a stochastic matrix is obtained, that is, a matrix where all the components are in-between zero to one, and every column sums equal to one. This is the matrix of normalized alternatives’ scores against criteria, P = pnm, with n indicating the number of alternatives and m the number of criteria. The next step is the attribution of weights for the criteria. Usually, weights are attributed by DM using a linear 1 to 5 scale, similar to the Likert scale (Likert, 1932) [35] The decision makers must indicate a criterion r as the reference criterion. The criterion with the highest weight is usually chosen. The vector of weights, wr = wrc, is composed by the weight of the criterion c divided by the weight of the reference criterion r. The measurement of dominance (Ai, Aj) of each alternative Ai over each alternative Aj, incorporate concepts of Prospect Theory, according to Equation 3. m c jicji jiAAAA 1 ),(),,(),( (3) Where: 0)(if )()( 1 0)(if0 0)(if )( ),( 1 1 jcic rc icjc m c rc jcic jcicm c rc jcicrc jic PP w PPw PP PP w PPw AA The expression c(Ai, Aj) is the contribution of criterion c to the dominance of alternative Ai over alternative Aj. If pic was greater than pjc, it will represent a gain for (Ai, Aj); if pic and pjc were equal, then a zero will assigned to (Ai, Aj); if pic was less than pjc, then c(Ai, Aj) will be a loss to (Ai, Aj). The function c(Ai, Aj) allows the adjustment of problem data to the Prospect Theory, that is, considering the aversion to risk and the propensity to risk. This function has the shape of an ‘‘S’’, as presented in Figure 1. 130 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 Figure 1. Value function of the TODIM method [15] Above the horizontal axis, that is for value equal to zero, there is a concave curve representing the gains; below the horizontal axis, there is a convex curve representing the losses. The concave part reflects the aversion to risk in the face of gains and the convex part, in turn, symbolizes the propensity to risk when dealing with losses. is the attenuation factor of the losses. Different choices of lead to different shapes of the prospect theoretical value function in the negative quadrant. The overall value for alternative Ai, i, is obtained with Equation 4. n j ji n j ji n j ji n j ji i AAAA AAAA 11 11 ),(min),(max ),(min),( (4) 2.3. Fuzzy expert systems Expert system is an information system that emulates the decision-making ability of a human expert [36]. An expert system is typically made of three parts: a Knowledge Base, an Inference Engine and a Working Memory [37]. The Working Memory is the stored information gained by the user of the system. The Inference Engine uses the Knowledge Base together with information from the problem to provide an expert solution. If–Then rules are popular schemes for knowledge representation as “If premise then conclusion”. In a Fuzzy Expert System, premises and conclusion are fuzzy propositions, as in “If X is small then Y is large with a certainty factor equal to 0.8” [28], for instance, because Small and Large are fuzzy sets. Several membership functions can be used in the definition of a fuzzy set. One of the most used is the triangular function [38].As presented in Figure 2, a triangular fuzzy set has a triangular membership function. A triangular fuzzy set is usually represented as a vector, (x1, x2, x3), where A (x1) = A (x1) = 0, and A (x1) = 1. Figure 2. Triangular fuzzy set A = (x1, x2, x3) One of the most popular Fuzzy Expert System is the Mamdani Model [39]. In a Mamdani Model, If–Then rules may have several clauses as If A and B and C… Then Z, where all A, B, C… and Z are fuzzy propositions. For every clause in the rule, the matching degree between the current value for the variable and a linguistic label must be computed. The clauses are aggregated, using the minimum fuzzy operator. If more than one rule implies in the same result, the rules must be aggregated, using the maximum fuzzy operator. The overall matching degree can be obtained, also using the minimum fuzzy operator. This degree is referred as alpha-cut, 131 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 or -cut [40]. The -cut level will generate a new fuzzy set, with a trapezoidal membership function, as presented in ! . A real number may be obtained from the centroid of gravity (COG) of the resulting fuzzy set, within a process referred as defuzzification [41]. Figure 3. Defuzzification by the centroid of gravity 3. Illustrative Case 3.1. Data collection Volta Redonda is a Brazilian city situated in the south of the State of Rio de Janeiro. The city has approximately 260,000 inhabitants [42]. There is a large number of properties, residential and commercial, rented or available for rent. The major steel plant installed in the city in the 1940’s is a landmark of Brazilian industrialization. Because of this industrial vocation, Volta Redonda was nicknamed Steel City. However, its economy is quite diverse on services as education and transportation, to name a few. Local real estate agents mentioned eight criteria for the selection of a residential property: Location (C1), Constructed Area (C2), Construction Quality (C3), State of Conservation (C4), Garage Spaces (C5), Rooms (C6), Attractions (C7), and Security (C8). Criteria C2, C5, C6, and C8 are quantitative; C2 is measured in m 2; C5 and C6 are measured in unities of rooms or space garages; and C8is a crisp criterion, since a residential property has security or has not. Tables 1 to 4 present the possible scores to evaluate alternatives according to qualitative criteria. Table 1. Possible scores to C1 Location Score Periphery 1 Between periphery and an average location 2 Average location 3 Good location 4 Excellent location 5 Table 2. Possible scores to C2 Construction Quality Score Low standard 1 Average standard 2 132 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 High standard 3 Table 3. Possible scores to C3 State of Conservation Score Bad 1 Average 2 Good 3 Very good 4 Table 4. Possible scores to C4 Attractions Score Without attractions 0 Backyard or terrace 1 Barbecue 2 Swimming pool 3 Swimming pool, barbecue and others 4 Weights from 1 to 5 must be assigned to the criteria, where 1 goes to the lowest important criterion and 5 to the highest important criterion. Location (C1) was indicated as the reference criterion. Table 5 presents the assigned weighted and the normalized weights, that is, summing equal to one. Table 5. Weights of criteria Criterion Assigned weight Normalized weight Localization (C1) 5 0.25 Constructed Area (C2) 3 0.15 Construction Quality(C3) 2 0.10 State of Conservation (C4) 4 0.20 Garage Spaces (C5) 1 0.05 Rooms (C6) 2 0.10 Attractions (C7) 1 0.05 Security (C8) 2 0.10 Fifteen residential properties in different neighborhoods of Volta Redonda were evaluated. These alternatives were simply named as A1 to A15. Table 6 presents the scores assigned to the alternatives according to the qualitative criteria (C1, C3, C4, and C7) and real data for the quantitative criteria (C2, C5, C6, and C8). Table 6. Data and assigned scores of residential properties Residential properties C1 C2 C3 C4 C5 C6 C7 C8 A1 3 290 3 3 1 6 4 0 A2 4 180 2 2 1 4 2 0 A3 3 347 1 2 2 5 1 0 A4 3 124 2 3 2 5 4 0 A5 5 360 3 4 4 9 1 1 A6 2 89 2 3 1 5 1 0 A7 1 85 1 1 1 4 0 1 A8 5 80 2 3 1 6 0 1 A9 2 121 2 3 0 6 0 0 A10 2 120 1 3 1 5 1 0 A11 4 280 2 2 2 7 3 1 A12 1 90 1 1 1 5 2 0 133 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 A13 2 160 3 3 2 6 1 1 A14 3 320 3 3 2 8 2 1 A15 4 180 2 4 1 6 1 1 Table 7 presents the normalized score for the residential properties against the criteria. Table 7. Normalized scores for the residential properties against the criteria Residential properties C1 C2 C3 C4 C5 C6 C7 C8 A1 0.068 0.103 0.100 0.075 0.045 0.069 0.174 0 A2 0.091 0.064 0.067 0.050 0.045 0.046 0.087 0 A3 0.068 0.123 0.033 0.050 0.091 0.057 0.043 0 A4 0.068 0.044 0.067 0.075 0.091 0.057 0.174 0 A5 0.114 0.127 0.100 0.100 0.182 0.103 0.043 0.143 A6 0.045 0.031 0.067 0.075 0.045 0.057 0.043 0 A7 0.023 0.030 0.033 0.025 0.045 0.046 0 0.143 A8 0.114 0.028 0.067 0.075 0.045 0.069 0 0.143 A9 0.045 0.043 0.067 0.075 0 0.069 0 0 A10 0.045 0.042 0.033 0.075 0.045 0.057 0.043 0 A11 0.091 0.099 0.067 0.050 0.091 0.080 0.130 0.143 A12 0.023 0.032 0.033 0.025 0.045 0.057 0.087 0 A13 0.045 0.057 0.100 0.075 0.091 0.069 0.043 0.143 A14 0.068 0.113 0.100 0.075 0.091 0.092 0.087 0.143 A15 0.091 0.064 0.067 0.100 0.045 0.069 0.043 0.143 The overall values presented in Table 8 were obtained simply by aggregating the normalized scores for the residential properties (Table 7), weighted by the normalized vector (Table 5). The bolded A5, A11, and A14 have the highest overall values; the stricken throughA7, A9, A10, and A12 have the lowest values. Table 8. Overall values for the residential properties Residential properties Overall value Rank A1 0.301 6 A2 0.241 10 A3 0.245 9 A4 0.257 8 A5 0.454 1 A6 0.192 11 A7 0.159 14 A8 0.311 5 A9 0.185 12 A10 0.185 12 A11 0.351 3 A12 0.125 15 A13 0.291 7 A14 0.366 2 A15 0.338 4 3.2. TODIM application As in many other TODIM applications [43], = 1 is adopted. To illustrate TODIM computation, from Table 2, let us consider the pair A2 and A4: For C1, p21>p41, then 0.075 For C2, p22>p42, then = 0.054 134 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 For C3, p23 =p43, then = 0 For C4, p24<p44, then – 0.353 For C5, p25<p45, then – 0.959 For C6, p26<p46, then – 0.331 For C7, p27<p47, then – 1.319 For C8, p28 = p48, then 0 Then, substituting values in Equation 1, (A2, A4) – 2.833. In analogy, all other (A2,Aj) can be found, and adding then, = – 27.02. The minimum and maximum sums are = – 44.23 and = 0.343. By substituting in Equation 4, one gets 0.386. Table 9 presents he overall values for the residential properties with TODIM. It can be noted in Tables 8 and 9 that the incorporation of Prospect Theory implies in a different rank. However, the top three (the bolded A5, A11 and A14) and the bottom four (A7, A9, A10, and A12) will be the same. The Emond-Mason coefficient computed for these ranks is x ≈ 0.733, which indicates a positive correlation. Table 9. Overall values for the residential properties with TODIM Residential properties Overall value Rank A1 0.692 5 A2 0.386 10 A3 0.399 9 A4 0.621 7 A5 1 1 A6 0.286 11 A7 0 15 A8 0.441 8 A9 0.020 14 A10 0.213 12 A11 0.858 3 A12 0.107 13 A13 0.719 4 A14 0.937 2 A15 0.673 6 3.3. Fuzzy expert system application A fuzzy expert system was developed, as presented in Figure 4. To facilitate the implementation of the expert system, the Software FuzzyTECH [44] was selected. The choice for FuzzyTECH was mainly because of it has been successfully applied in MCDA practice [25]. 135 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 Figure 4. Fuzzy expert system to evaluate residential properties Three triangular fuzzy sets were defined for each qualitative criteria (C1, C3, C4 and C7): Bad (2, 2, 3), Average (2, 3, 4), and Good (3, 4, 4). For the quantitative criteria (C2, C5, C6 and C8), only one set was defined: Good. For Evaluation, two sets were defined: Bad (0.25, 0.75, 0.75) and Good (0.25, 0.25, 0.75). Figures 5 to 7 present the triangular fuzzy sets for Location, the fuzzy set for Constructed Area, and the triangular fuzzy sets for Evaluation. Figure 5. Triangular fuzzy sets for C1 Figure 6. Triangular fuzzy set for C2 136 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 Figure 7. Triangular fuzzy sets for Evaluation The fuzzy expert system is accomplished in 34 = 81 rules. In FuzzyTECH, these rules were created in groups of three rules, inserted in blocks of three groups. Table 10 presents the first three and the latest three groups. When there is, at least, one Bad set, the rule output will be Evaluation equals to Bad. Otherwise, the output will be Evaluation equals to Good. Table 10. Fuzzy rules Input Output Rule Location Construction quality State of conservation Attractions Evaluation 1 Bad Bad Bad Bad Bad 2 Bad Bad Bad Average Bad 3 Bad Bad Bad Good Bad … … … … … … 79 Good Good Good Bad Bad 80 Good Good Good Average Good 81 Good Good Good Good Good Table 11 presents the overall value obtained with defuzzification by COG. The first observation from Table 11 is that for 10 of 15 residential properties, Evaluation resulted overall values equal to zero. This is due to the fuzzy system design. Residential properties with the lowest value according to a quantitative criterion received a -cut equal to zero. Table 11. Overall values for the residential properties with fuzzy expert system Residential properties Overall value Rank A1 0 6 A2 0 6 A3 0 6 A4 0 6 A5 0.259 3 A6 0 6 A7 0 6 A8 0 6 A9 0 6 A10 0 6 A11 0.452 2 A12 0 6 A13 0.259 3 137 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 A14 0.741 1 A15 0.259 3 This situation was mainly originated because C8, is a crisp criterion. This way, A1, A2, A3, A4, A6, A9, A10, and A12 have overall values equal to zero. Nevertheless, it seems to be plausible, since they are residential properties without security system, and, Security is a major issue in Brazil, particularly in Rio de Janeiro. 4. Discussion and Conclusions Comparing the rank from the fuzzy expert system to the one resulted with TODIM application, both resulted A5, A11, and A14 as top residential properties; and A7, A9, A10, and A12 in the bottom. Therefore, besides the differences in the overall values, the ranks from both applications can be considered as compatible each other, in qualitative terms. However, the Emond-Mason coefficient computed for these ranks is x ≈ 0.408, which indicates a fragile correlation. From ordering theory, the rank obtained with TODIM is classified as a linear rank, since it has no ties. Now, the rank from fuzzy expert system is a weak rank, since there are ten alternatives tied in the last 6th position. Surprisingly, the fuzzy expert system’s rank is crisp and TODIM’s rank is fuzzy. That is, in this presented case the use of an MCDA method was superior to fuzzy systems. The fuzzy expert system’s fragility exposed in this case was due the necessity of a crisp evaluation on Security. This work satisfied its objective presenting a comparison between TODIM’s rank with the rank from a fuzzy expert system. The comparison follows a mixed qualitative-quantitative research strategy, that is, our findings are based in a single case. Then, new researches must be conducted comparing ranks from TODIM with ranks from other MCDA methods and mainly comparing fuzzy system ranks with ranks from other decision techniques. With multiple examples, that is with quantitative researches, the fuzzy system’s fragility observed in this work can be demonstrated. Another interesting subject for future investigation is the use of Emond-Mason coefficient to compare ranks previously obtained with MCDA methods in an ex-post facto research approach. Acknowledgments Authors need to thank Prof. Dr. Luiz Flavio Autran Monteiro Gomes for valuable advises, comments, and suggestions. This research has financial support from Brazilian Council for Scientific and Technological Development (Grant No. CNPQ 302692/2011-8) and Sao Paulo State Research Foundation (Grant No. FAPESP 2013/03525-7). References [1] Ishizaka A, Nemery P. Multi-criteria decision analysis. Chichester: Wiley; 2013. [2] Doumpos M, Zopounidis C. Multicriteria decision aid classification methods. Dordrecht: Kluwer; 2002. [3] Roy B. The optimisation problem formulation: criticism and overstepping. J Oper Res Soc 1981;32:427–36. [4] Vincke P. Multicriteria decision aid. Chichester: Wiley; 1992. [5] Olson D. Decision aids for selection problems. : Berlin: Springer; 1996. [6] Saaty TL. Analytic hierarchy process. New York: McGraw-Hill; 1980. [7] Keeney R, Raiffa H. Decisions with multiple objectives. Cambridge: Cambridge University Press; 1976. [8] Roy B. Classement et choix en pr´esence de points de vue multiples (la méthode ELECTRE). Revue d’Informatique et de Recherche Opérationnelle 1968; 2:57–75. [9] Brans JP, Vincke P. A preference ranking organisation method: (the PROMETHEE method for multiple criteria decision-making). Manage Sci 1985;31:647–56. [10] Bana e Costa CA, De Corte J-M, Vansnick J-C. On the mathematical foundation of MACBETH. In: Figueira F, Greco S, Ehrogott M, editors. Multiple criteria decision analysis, New York: Springer; 2005, p.409–37. 138 Valério Antonio Pamplona Salomon and Luís Alberto Duncan Rangel / Procedia Computer Science 55 ( 2015 ) 126 – 138 [11] Zanakis SH, Solomon A, Wishart N, Dublish S. Multi-attribute decision making: a simulation comparison of select methods. Eur J Oper Res 1998;107:507–29. [12] Emond EJ, Mason DW. A new rank correlation coefficient with application to the consensus ranking problem. J Multi-Crit Decis Anal 2002;11:17–28. [13] Gomes LFAM, Lima MMPP. TODIM: basic and application to multicriteria ranking of projects with environmetal impacts. Fund Computing Decis Sci 1991;16:113–27. [14] Kahneman D, Tversky A. Prospect theory: An analysis of decision under risk. Econometrica 1979;47:263–92. [15] Gomes LFAM, Lima MMPP. From modelling individual preferences to multicriteria ranking of discrete alternatives: a look at Prospect Theory and the additive difference model. Fund Computing Decis Sci 1992;17:171–84. [16] Gomes LFAM, Rangel LAD. An application of the TODIM method to the multicriteria rental evaluation of residential properties. Eur J Oper Res 2009;193:204–11. [17] Gomes LFAM, Machado MAS, Costa FF, Rangel LAD. Criteria interactions in multiple criteria decision aiding: A Choquet formulation for the TODIM method. Procedia Computer Science 2013;17:324–31. [18] Krohling RA, De Souza TT. Combining prospect theory and fuzzy numbers to multi-criteria decision making. Expert Syst Appl 2012;39:11487-93. [19] Krohling RA, Pacheco AG, Siviero AL. 2013. IF-TODIM: An intuitionistic fuzzy TODIM to multi-criteria decision making. Knowl- Based Syst 2013;53:142–6. [20] Bellmann R, Kalaba R, Zadeh LA. Abstraction and pattern classification, Santa Monica: RAND; 1964. [21] Zadeh LA. Fuzzy sets. Inform Control 1965;8:338–53. [22] Van Laarhoven PJ, Pedrycz W. A fuzzy extension of Saaty’s priority theory. Fuzzy Set Syst 1983;11:229–41. [23] Kahraman C. Fuzzy multi-criteria decision making. New York: Springer; 2012. [24] Giannopoulos D, Founti M. A fuzzy approach to incorporate uncertainty in the PROMETHEE multicriteria method. Int J Multicriteria Decision Making 2010;1:80–102. [25] Prodanovic P, Simonovic SP. Comparison of fuzzy set ranking methods for implementation in water resources decision-making. Can J Civil Eng 2002;29:692–701. [26] Valls A, Pijuan J, Schuhmacher M, Passuello A, Nadal M, Sierra, J. Preference assessment for the management of sewage sludge application agricultural soils. Int J Multicriteria Decision Making 2010;1:4–24. [27] Zhu K. The invalidity of triangular fuzzy AHP: a mathematical justification. Social Science Research Network 2012; doi:/10.2139/ssrn.2011922 [28] Saaty TL, Tran LT. On the invalidity of fuzzifying numerical judgments in the analytic hierarchy process. Math Comput Model 2007;46:962–75. [29] Zadeh LA. The role of fuzzy logic in the management of uncertainty in expert systems. Fuzzy Sets Syst 1983;11:199–227. [30] Bryman A, Bell E. Business research methods. Oxford: Oxford University Press; 2007. [31] Kendall MG. A new measure of rank correlation. Biometrika 1938;30:81–93. [32] Cook WD, Kress M, Seiford LM. An axiomatic approach to distance on partial orderings. Recherche operationnelle 1986;20:115–22. [33] Roy B, Bouyssou D. Aide multicritère à la décision. Paris: Economica; 1993. [34] Pomerol J-C, Barba-Romero S. Multicriterion decision in management. Dordrecht: Kluwer; 2000. [35] Likert R. A Technique for the Measurement of Attitudes. Archives of Psychology 1932:140:1–55. [36] Jackson P. Introduction to expert systems. Harlow: Addison-Wesley; 1999. [37] Kandel A. Fuzzy expert systems. Boca Raton: CRC; 1992. [38] Bobillo F, Delgado M, Gómez-Romero J. A semantic fuzzy expert system for a fuzzy balanced scorecard. Expert Syst Appl 2009;36:423–33. [39] Mamdani EH, Assilian S. An experiment in linguistic synthesis with a fuzzy logic controller. Int J Man Mach Stud 1975;7: 1–13. [40] Liang G-S, Ding, J-F. Fuzzy MCDM based on the concept of alpha-cut. J Multi-Crit Decis Anal 2003;12: 299310. [41] Chen SH. Ranking fuzzy number. Fuzzy Set Syst 1985;17:113–129. [42] Brazilian Institute of Geography and Statistics. IBGE | Cidades | Rio de Janeiro | Volta Redonda. http://cidades.ibge.gov.br/xtras/perfil.php?lang=&codmun=330630; May 29, 2015. [43] Ribeiro LS, Passos AC, Teixeira MG.. Selection of communication technologies in the Brazilian army using AHP, TODIM and Sapiens software. Prod 2012;22:132–41. [44] INFORM GmbH. FuzzyTech 6.02 Professional. http://www.fuzzytech.com; May 29, 2015. 
work_nj65ubti25eqthx75zbpbtwhnq	----	PII: 0888-613X(87)90013-2 Extended Fuzzy Relations: Application to Fuzzy Expert Systems J. J. Buckley M a t h e m a t i c s D e p a r t m e n t , University o f A l a b a m a at B i r m i n g h a m Douglas Tucker Carraway M e d i c a l Center, B i r m i n g h a m , A l a b a m a A B S T R A C T This paper introduces the concept o f an extended f u z z y relation, which is a relation whose values are vectors o f f u z z y relations, some o f which may also be extended f u z z y relations. Our motivation is to use extended fuzzy relations to replace blocks o f rules in a f u z z y expert system with one rule. The extended f u z z y relation method is shown to contain the generalized modus ponens as a special case. The construction o f extended f u z z y relations is illustrated in two examples taken f r o m diagnosing mental disorders and image processing. We argue that the existence o f an extended f u z z y relation f o r a block o f rules may be a criterion f o r parallel execution o f this block instead o f sequential firing o f the rules. K E Y W O R D S : e x p e r t systems, f u z z y relations, p a t t e r n recognition " l 1. I N T R O D U C T I O N The purposes o f this article are to introduce extended fuzzy relations (EFR), explain their application to fuzzy expert systems, and show how they m a y be employed to decide between parallel and sequential firing o f production rules in a fuzzy expert system. W e begin with defining extended fuzzy r e l a t i o n s . Let S and T be any two nonempty sets. A regular fuzzy relation R is a function on a subset o f S x T w i t h values in [0, 1]. The strength o f the relation between s E S and t E T is given b y s R t , the value o f R at (s, t), a number between zero and one. An E F R (R w i l l be defined in a sequence o f steps. The domain o f (R will be a Address correspondence to J. J. Buckley, Mathematics Department, University of Alabama at Birmingham, Birmingham, Alabama 35294. International Journal of Approximate Reasoning 1987; 1:177-195 © 1987 Elsevier Science Publishing Co., Inc. 52 Vanderbilt Ave., New York, NY 10017 0888-613X/87/$3.50 177 178 J . J . Buckley et al. subset o f U x V, where U = {Ul, " ' ' , Urn} and V = {Vl, " " , on} are two finite sets. The value o f 61 at (ui, vj) will be denoted by ui61v-i. Level Zero We have ui61v-i = 7i-i (1) where 70 E [0, 1]. A level zero E F R is a regular fuzzy relation. A level zero EFR will be represented by 61o. Level One For all levels e, f _ 1, the elements o f U m a y be vectors; so let ui = ( / d i l , " ° " , uiki), a vector o f length Ki. Assume that the possible values o f uik all belong to some universal set Uik. Then ui61vj= R # = (Rijl, "" ", Rok i) (2) where each Rijk is a regular fuzzy relation 610. RO is a vector o f length Ki o f regular fuzzy relations, and the length o f R 0 is the same as the length o f u~. A level one E F R is written as 611. Level T w o We have ui61vj = R 0 (3) where Rij is a vector o f length Ki whose components are 61~ or 61o extended fuzzy relations. A level two E F R is denoted by 612. Level L The value o f 61 is ui(Rv-i = Ri-i (4) where the components o f R t / a r e (Re, 0 __< g _< L - 1. Let 61L be a level L EFR. The definition o f an E F R will be completed when we specify the domains o f all the relations mentioned in the foregoing statements. The formal, recursive definition o f extended fuzzy relations is rather involved, and the reader may wish to turn to an example near the end o f this section which illustrates the notation. Because U and V are finite, we may use matrix notation to describe an EFR. Extended Fuzzy Relations: Application 179 Let the rows o f a m x n matrix be labeled by Ug and the columns by vj. W e then place Rij, or "Yij for a 6/0, in the (ij)th cell. I f some ui is not related to some vj, then we leave the corresponding ( i j ) cell empty. When some cells are empty, the domain o f the E F R will be a proper subset o f U x V. Obviously, extended fuzzy relations are generalizations o f regular fuzzy relations, because for levels e _ 1 the cells contain vectors o f relations. W e are primarily interested in evaluating an E F R given some data D . W e will represent D as ,u l , , , where di = (dil, " " , diki) is a vector o f length Ki if ui is a vector o f length Ki. Suppose each dgk takes its values in some universal set ~3ik. The structure o f the data will vary with the levels o f the extended fuzzy relations. The data for some (Re will b e c o m e (generate) all the data for all the extended fuzzy relations contained in (Re. Any E F R (Re under discussion will have its domain a subset o f U × V w i t h data D , but any E F R within (Re will be assumed to be defined on a subset o f O × I7 with d a t a / ) . There may be many 61y, 0 _< j _< g - 1, contained in (Re, but we will use the same notation O × f " a n d / 5 for all these 61j extended fuzzy relations. W e will now specify the domains o f all the relations contained in a given EFR. For L e v e l Zero 61o is defined on a subset o f U x V with values in [0, 1]. For L e v e l One each Rijk is defined on a subset o f ~ k X Uik with values in [0, 1]. For L e v e l T w o each Rijk is defined on a subset o f ~ i k X Uik with values in [0, 1]. Suppose some Rijk is a 611. This 611 is also defined on a subset o f O × I7 with d a t a / ) . W e require Uik C lTand 33ik contain the possible data v a l u e s / ) for 611. Finally, for L e v e l L each RUk is defined on a subset o f ~Da x Ua with values i n [ 0 , 1 ] . I f R i j k = (Re, 0--< e_< L - 1, then Ugk C l " a n d D C_ ~ k . W e now introduce some general notation for the evaluation o f an EFR. The evaluation o f 61, given D , produces output O given by O = D o 61 (6) where ,,o,, is some type o f composition o f D and 61. The output O will be a fuzzy subset o f V given by ,o 1 The composition o f D and 61 will be determined from the inner product di * Rij o f di and Rij and a generalized matrix product. Let CVk = dik RijkUik (8) 180 J . J . Bucldey et al. and c i j = d i * R u = F i j ( c o ' l , " " , ci.iki) (9) where F~/is some function that aggregates the cuk values into one value c U. Then we define oj in the output O to be o j = f ( c l j , c 2 j , " ' ' , C m j ) (10) where f is some other function that combines the c U values into one number o j . The entire fuzzy output O is represented as D o (R, so let D ( R v i = o 1 denote a single membership value. The evaluation procedure is easily visualized using the matrix representation for 61. Given the data ( d i , • • ", din), we first take the product o f this vector with the column vector under v 1. The individual products o f di and Riy are defined by the inner product d~ * R i i . We then " s u m " or aggregate these inner products o v e r an entire column into the output value oj. Therefore, the composition D o 61 may be interpreted as a generalized matrix product. Now we may present a more precise definition o f an EFR. LEVEL ZERO Let cij = g ( d i , "Yij), where g is some function that combines d; and 3'ij into one number c o in [0, 1]. No F,j function is required, and oj = f ( c U, • • . ~ C m j ) . LEVEL ONE The ciik values are obtained from dikRi.ikUik; t h e Fii functions produce the c U, and then f gives the output oy for all j . LEVEL TWO When R i j k is a 610, the cuk values are computed directly from dikRijkUik. S o a s s u m e Rijkis a (RI. We know Uik C I7", so let uik = 0j E 17". W e also know dik will be a data value D for (Ri. Then Cijk = d i k ( R l Uik :-- D ( R l ()j = O j (11) where 6j is an output o f (R i. When Riik is a level one EFR, c,)-k is the correct level one output. LEVELL SupposeRijk = (Re, 0 - < e_< L - 1. Then cijk = dik (ReUik = D (ReO i = Oj (12) where D , 0j, and 6 / a r e the correct level e values for (Re. Our intended use o f extended fuzzy relations is to model and evaluate simultaneously blocks o f rules in a fuzzy expert system. Let us consider constructing a level one E F R for the following block o f rules. If xilRijlAil and/or " " X i k i R i j k i A i k i (13) t h e n y is B j f o r 1 <_ i <_ m , 1 < j <_ n . L e t vj = By and R i j = ( R i j l , " " " , RiJki) Extended Fuzzy Relations: Application 181 be the vector o f regular fuzzy relations. Also let ui = ( A / l , • • " , Aiki), I <_ i < m , be a listing o f the attributes on the left side. The data D contains di = (xil, • " , xiki), 1 <_ i <_ m . In the matrix representation o f the EFR, each ( i j ) cell containing an Ri.i corresponds to a rule in equation (13). Cells are left empty when a uj and vy are not related through a rule. The entrees in a v~ column represent all rules with the same conclusion oj. The items in a ui row are all the rules with the same left-side attributes. The function F u is chosen to reflect the structure (ands, ors, nots) in the left side o f the rule. I f only " a n d " is used in a rule, then we would consider using minimum for F u. W e would also consider employing m a x i m u m f o r f b e c a u s e we are " o r i n g " over all rules with the same conclusion to obtain our confidence oj in vj. The block o f rules in equation (13) becomes, using the level one EFR, if x is D , then y is 0. (14) Example Let us now consider an example that requires a level three EFR. This is a fuzzy expert system for diagnosing mental disorders (Siler and Tucker [8]). The information flow in this system is shown in Figure 1. The user is first asked a number o f questions that, together with their answers, create a set o f facts 5:. The system asks i f certain behavioral manifestations exists in the patient, and the user is to enter his/her confidence, f r o m zero to one, that they are present in the patient. A subset o f these facts 5:0 are input to a level one E F R (Rs that produces a fuzzy set o f symptoms. Let ITs = {depressive ( D i ) , manic ( M ) , schizophrenic (SI)} be the set o f symptoms. The output from (Rs will be a fuzzy subset o f Vs. Now, (Rs represents a block o f rules. An example o f a rule from this block is as follows: I f [fact = excess energy] A N D [fact = many big plans] A N D [fact = hallucinations], then [symptom -- M ] (15) Let us code " e x c e s s e n e r g y " as E E , " m a n y big p l a n s " as M B P , and " h a l l u c i n a t i o n s " as H . I f the number o f this rule, within this block, is seven, then u7 = ( E E , M B P , H ) . Next assume that the u s e r ' s response to questions about these conditions produced confidences c f ( E E ) , c f ( M B P ) , and c f ( H ) , respectively. The data input Ds for 6Is would then contain ( c f ( E E ) , c f ( M B P ) , c f ( H ) ) U 7 (16) I f this rule has prior confidence 0.80 (discussed further in the section on the generalized m o d u s p o n e n s ) , then to evaluate (Rs we must first find all the c O 182 J . J . Buckley et al. Q L v c3 ~... O .1.4 r - I 0 J P.4 o~ o~ o ~ o -r-/ r-4 ¢0 ° ~ ~J ..= _o L L o Extended Fuzzy Relations: Application 183 and, in particular, c72 = F72(cf(EE)R721EE, cf(MBP)R722MBP, cf(H)R723H) = min (cf(EE), cf(MBP), cf(H), 0.80) (17) We have assumed the columns of (Rs are labeled D l , M , and $1, so the second column is for manic. The relations in 6Is are very simple in that they pick off the confidence in the given condition. (More complicated relations are presented in the example in the section on application.) In this way all the c U are determined, and after taking column maximums we obtain the output O ~ = [ ° ~ l ' M ' S 1 0 2 03 1 (18) We next input O~, together with more facts 5:1, into a level two ERF (Rp to determine a fuzzy set of preliminary diagnoses. Let Vp = {depression (/)2), schizophrenia ($2), paranoid (P)} be the set of preliminary diagnoses. Here (Rp also represents a block of rules, and a rule from this block is If [symptom = M ] AND [symptom = D d AND [fact = persecutory or jealous delusions] AND [fact = H I , then [preliminary diagnosis=DE] (19) We code "persecutory or jealous delusions" as PJD. I f the number of this rule is five, then u5 = (Mr, Dl, PJD, H ) , and the data Dp for (Rp will contain (5:0, 5:0, cf(PJD), cf(H)) (20) U5 The columns in 6~p are labeled DE, $2, P, so csl = Fsl(5:o6isM, 5:o(RsDi, cf(PJD)RsI3PJD, cf(H)RsI4H) = rain (o2, ol, cf(PJD), cf(H), 0.90) (21) where 0.90 is the prior confidence in this rule. The output from CRp is OP= [0~22 ' $2 'o2 ~I (22) Lastly, Op and other facts 5:2 are put into a level three EFR (R: to obtain a fuzzy set of final diagnoses. Let V: = {D31, D3E, D33, S3I, S32, $33, $34, Pl, PE} be the collection of final diagnoses where, for example, D32 = manic- depressive, $33 = paranoid-schizophrenic, and so on. A rule from fRf is If [preliminary diagnosis = $2] AND [fact = PJD] AND [fact = H], then [final diagnosis = $33] (23) 184 J . J . Buckley et al. This rule has number three, so u3 = ( $ 2 , PJD, H) and ((5:0 U 5:1), c f ( P J D ) , c f ( H ) ) U3 is in Df. The column number for $33 is six, so C36=F36[(5:o O 5 : l ) ( ~ p S 2 , cf(PJD)R362PJD, cf(H)R363H] (24) = min [02, c f ( P J D ) , c f ( H ) , 1.01 if the rule confidence is one. The final output is , ° 21 (25) (26) Using (~f all the rules become I f the facts are 5:, then the (final) diagnosis is Of (27) Notice that multiple copies o f columns from (Rs can be in different cells in CRp, and multiple copies o f columns from ~ p may be in many different cells in (Ry. It may happen that we have two rules with the same attributes ui, same conclusion oj = Bj, but different relations Rij and R~. We would place both Rij t and Rij in the same (ij) cell but use possibly different functions F U and F:. to U obtain c U and c~, respectively, given data D. Then both cij and c~. would be input into f to obtain oj. The model can be easily extended to incorporate compound right sides. I f rules have multiple conclusions, then we would add more columns to the EFR; each vj would thus be a vector o f conclusions. We may also consider generalizing to process multidata simultaneously. The data D were assumed to be all the information necessary for one problem (run). To process many problems at once, the data are a vector ( D I , / ) 2 , " " ) and the output is separated into (Oi, 02, "" "). In the example in the section o f application, we process multidata simultaneously. In the next section we compare the generalized modus ponens approach to the extended fuzzy relation technique for modeling fuzzy production rules. We illustrate the use o f level one and level two extended fuzzy relations in an image processing problem in the section on application. We then argue, in the section o f parallel versus sequential processing, that the existence o f an EF R is a key to knowing when a block o f rules may be executed in parallel instead o f sequentially in a fuzzy expert system. The last section contains a brief summary and conclusions. Extended Fuzzy Relations: Application 185 T H E G E N E R A L I Z E D M O D U S P O N E N S The structure o f the generalized m o d u s p o n e n s (GMP) is (Zadeh [115, 16]) if x is A , then y is B x is D , then y is O (28) where A and D are fuzzy subsets o f U = {u~, . . . , urn} and B and O are fuzzy subsets o f V = {vl, • • ", v,}. One constructs, using A and B, a regular fuzzy relation R on U x V with values uiRvj = ~/ij (Mizumoto and Z i m m e r m a n [5], Mizumoto [6], Whalen and Schott [13], Yager [14]). Then O = D o R for some composition , ' o " . This is exactly our level zero evaluation o f an EFR. We m a y also model the G M P as a level one EFR. Consider the following block o f rules: I f xiRijui, then y is vj (29) for 1 _ i _ n, 1 < j _ m . The definition o f Rij is cij = diRijui= g ( di, "Yij) (30) and we therefore obtain the same conclusion that y is O as in the G M P . Both the G M P and the E F R method are used to construct fuzzy sets; however, there are basic differences in the two approaches. The G M P employs a fuzzy set B, which is actually part o f the rule, to obtain the fuzzy relation on U x V, whereas the E F R procedure uses the structure o f the left side o f a block o f rules to build the relation on U x V. In general, an EFR cannot be modeled as a single GMP. There are three main reasons for this conclusion: (1) the EFR method allows for general input; (2) a high-level EFR is evaluated differently than a G M P ; and (3) the E F R approach can incorporate prior rule confidence, thresholding, and consideration o f preexisting data stored in working m e m o r y . In the G M P the di belong to [0, 1], as D is a fuzzy subset o f U; in an EFR, however, the components o f a di can be numbers, strings, fuzzy numbers, or discrete fuzzy sets. The only restriction on the di is the data types and relations allowed in the system. The example in the next section has the di vectors o f fuzzy numbers and strings. Therefore, the E F R technique allows for general input, whereas the G M P accepts only discrete fuzzy set input. T o understand the second reason given above, consider a level two E F R 612. N o w 612 represents blocks o f rules, and each block could possibly be modeled as a GMP. W e m a y even be able to construct a larger G M P , say &, to model all the blocks and 612. However, the evaluation o f 612 and (P would be different. Assume we are employing the traditional m a x and min for finding final confidences. Then the output f r o m 612 is a fuzzy set O where oj is the column m a x o f a group o f numbers each the min o f other numbers c~/k, and each cuk could be the output o f a level one EFR, hence is the column max o f a set o f numbers each the min o f 186 J . J . Buckley et al. other numbers. None o f the suggested methods o f evaluating (P will produce the oj numbers. The third reason will be discussed below. Modeling the G M P as a level one E F R was artificial because we never use 7o = uiRvj tO evaluate an EFR. That is, our fuzzy production system (Buckley, Siler, and Tucker [3]; Buckley, Siler, and Tucker [4]; Siler, Buckley, and Tucker [10], Siler, Tucker [8]) does not require a fuzzy set B to obtain O. However, we do take into consideration the preexistence o f some fuzzy set for y . Suppose, through other blocks o f rules or from the initial data base, we already have in working m e m o r y y is E , where E is a fuzzy subset o f V. W e would not use E to evaluate the EFR, given x is D , but instead we would combine O and E into one final fuzzy subset for y . The G M P method does not allow for the existence o f this fuzzy set E . We will now show how the E F R procedure can handle E together with prior rule confidence and thresholding b y simply changing the Fi/functions. Consider the block o f rules given in equation (13) modeled by a level one E F R (Rj. Let r U, a number between zero and one, be the prior confidence in the rule given in the (ij)th cell in (R i. Thresholding is a procedure by which time is saved by not completely executing a rule whose left side has low confidence. Let T, a number between zero and one, be the threshold value. Suppose there already exists in working m e m o r y y is E where E = [ e ~ e ~ l , . . . , ( 3 1 ) Given that x is D , we first compute cok = d~k Ri/kUik (32) and ~ij=di * Riy=hu(col, "" ", cok i) (33) where now hi~ is the function that aggregates the ci/k into c,7- Then I e j if t?/7 < T c U m a x [min (t?i/, ri/), e/I, if G0_> T (34) The function F U is the method o f obtaining the c 0 from the cok, so now F 0 contains hi~ and equation (34). Given the c#, we find oj, our final confidence in o/, as before, using the f function. We use the term " w e a k l y m o n o t o n i c " for the procedure given in equation (34) because it never allows a confidence to decrease. That is, we never replace a preexisting confidence e / w i t h a lower value. The weakly monotonic method may be summarized as follows. First, the confidence in the left side o f a rule t~t/ is tested for threshold. Second, if t?i/ passes threshold, then the new rule confidence is the minimum o f t? U and the prior rule confidence r/j. Last, i f the new rule confidence exceeds e/, then the new rule confidence becomes our new Extended Fuzzy Relations: Application 187 confidence in oj; otherwise, there is no change in our confidence in oj. The final confidence in vj is still given as oj, a function o f the c U. O f course, one may consider using other appropriate functions for max and min in equation (34). The main difference between the GMP method and the level one EFR approach is their use o f preexisting fuzzy sets for the right sides o f rules. The GMP technique employs a fuzzy set, as part o f the rule, for the right side o f a rule in order to evaluate the rule. The EFR method, in contrast, is used to construct the fuzzy set for the right sides o f a block o f rules. Also, the EFR approach allows for preexisting fuzzy sets in working memory for the right sides o f rules. It is our experience that the G M P is not generally applicable to rule- based fuzzy expert systems because (1) we usually do not have fuzzy sets for right sides o f rules but instead wish to construct these fuzzy sets; and (2) we quite often have to deal with preexisting fuzzy sets, in working memory, for the right sides o f rules. A P P L I C A T I O N Our problem was to design an expert system to medically classify regions that were previously identified in an echocardiogram. All these details have been reported elsewhere (Buckley and Siler [2]; Siler, Tucker, Buckley, Hess and Powell [7]; Tucker, Siler, Powell, and Stanley [11]); therefore, we will be concerned here only with the facts necessary to construct fuzzy relations. Numerical feature extraction was first carried out for each region. The features o f each region used are area (a fuzzy number a); x-coordinate o f the centroid (a fuzzy number ~); y-coordinate o f the centroid (a fuzzy number .P); type; and border. Type equals 1(0) if the pixels in the region are turned on (off), and border equals 1(0) if the region touches (does not touch) the border o f the picture. Using the fuzzy numbers a, .~, and.~, we first constructed fuzzy subsets o f sets SIZE, XPOS, and YPOS, respectively. We will now discuss in detail the type one EFR that goes with SIZE. The set SIZE equals {teeny, small, medium, large, huge}. N(r) denotes an appropriate fuzzy number centered at r, and L T E (GTE) are regular fuzzy relations less than or equal to (greater than o f equal to) defined on fuzzy numbers. The block o f rules used to build the fuzzy subset o f SIZE are as follows: I f a L T E N(100), then SIZE teeny I f a L T E N(200) and G T E N(100), then SIZE small I f a L T E N(500) and G T E N(200), then SIZE medium I f a L T E N(1000) and G T E N(500), then SIZE large I f a G T E N(1000), then SIZE huge 188 J . J . Buckley et al. We now define a level one EFR 6t a to take the place o f this block o f rules. Let V = SIZE and ul = N(100), u2 = (N(100), N(200)), u3 = (N(200), N(500)), u4 = (N(500), N(1000)), and u5 = N(1000) with U = {Ul, " " , us}. The values o f the R U are Rll = LTE, R22 = R33 = R44 = (GTE, LTE), R55 = GTE, and all the ( i j ) cells are empty for i g: j . The data D values are d l = d5 = a and d2 = d3 = d4 = (a, a). This is a very simple EFR because all the cells o ff the main diagonal are empty. The output O = D o 61a, a fuzzy subset o f V = SIZE, is given by O = I °te--~ny " " " h~gge05 1 (35) where we use minimum for h22, h33, and h44, and n o f i s required because 61~ is diagonal. For example, c22 = min (a G T E N ( 1 0 0 ) , a LT E N(200)). (36) and 0 2 =- min (£~22, r 2 2 ) (37) where r22 is the prior confidence in this rule. The confidence o2 will be the same confidence in " s m a l l " as obtained in the second rule in the foregoing block o f rules. In constructing the fuzzy set SIZE we do not employ thresholding and the weakly monotonic procedure discussed in the preceding section. In this situation we would not have a preexisting fuzzy set for SIZE in working m e m o r y . Using 61~, this block o f rules is replaced by one rule: if the area is a, then SIZE is 0 (38) In a similar manner two other fuzzy subsets o f XPOS = {far left, left, center, right, far right} and YPOS = {very high, high, middle, low, very low} are constructed using ~? and y , respectively. The blocks o f rules used to accomplish this may be represented by level one extended fuzzy relations 61x and 61y, respectively. Sometimes blocks o f rules, each represented by an EFR, will form a network, with the output o f some EFR being the input into another EFR. This is what occurred in the image processing problem, as shown in Figure 2. The four EFRs in Figure 2 will now be combined into one level two EFR 612. The block of rules forming the primary classification may also be represented by an EFR 61c. Instead o f discussing how to define 61c, we will concentrate on combining the EFRs into one level two EFR. A sample o f the primary classification rules is as follows: I f SIZE teeny and Border = 1, then Artifact I f SIZE small and XPOS left and YPOS low, then RA I f SIZE large and YPOS low and Type = 0 and Border = 1, then Dead-Zone Extended Fuzzy Relations: Application i e~ .o ~.~ 0 189 .o 3 ~ N N N I,< Z d 190 J . J . Buckley et al. I f S I Z E small and X P O S right and Y P O S low and T y p e = 0, then L A I f X P O S c e n t e r and Y P O S v e r y high and T y p e = 0 and B o r d e r = 1, then D e a d - Z o n e I f S I Z E m e d i u m and X P O S left and Y P O S high, then R V I f S I Z E h u g e and V P O S l o w and T y p e = 0 and B o r d e r = 1, then D e a d Z o n e T h e r e a r e 11 p o s s i b l e classifications, which w e will call C l , • • ", Cll and then let V = {C1, " " , C11}. E a c h ui is a v e c t o r , o f length at m o s t 5, w h o s e c o m p o n e n t s a r e m e m b e r s o f S I Z E , X P O S , o r Y P O S , o r uik = 1 (0) f o r T y p e , o r uik = 1 (0) f o r B o r d e r . Also, e a c h R 0 is a v e c t o r w h o s e c o m p o n e n t s a r e ffi~, 61,~, ffty, R r , o r RB w h e r e R r (RB) is a b i n a r y relation f o r T y p e (Border). F o r the classification rules p r e s e n t e d a b o v e , let C1 = Artifact, (?2 = R A , C3 = D e a d - Z o n e , C4 = L A , C5 = R V, and ul = (teeny, 1) u2 = (small, left, low) u3 = (large, l o w , 0, 1) u4 = (small, right, l o w , 0) us = (center, v e r y high, 0, I) u6 = ( m e d i u m , left, high) u7 = (huge, l o w , 0, 1) T h e n w e see that R l l = ((Ra, RB) R22=((Pta, (R.¢, (Pry) R 3 3 = ( ( R a , (Ry, R r , RB) R44= (ffto, 61~, 6ty, R r ) R53 = ((R~, 61y, R T , R n ) R65 = (6ta, fft,~, 6~y) R73 = ((Ra, (Ry, R T , RB) T h i s is a level t w o E F R b e c a u s e the R o contain level o n e e x t e n d e d f u z z y relations. T h e structure o f the data e l e m e n t s di m a t c h e s the structure o f the ui. F o r the rules p r e s e n t e d , w e see that f o r a g i v e n r e g i o n dl = (a, B o r d e r ) d2 = ((a, a ) , (.~, .¢), (.~, y ) ) Extended Fuzzy Relations: d3 d4 d5 d6 d7 Application = ((a, a), ( y , y ) , Type, Border) = ((a, a), (.¢, .¢), ( y , y ) , Type) = ((.~, ~¢), fl, Type, Border) = ((a, a), (~, ~), ( y , y ) ) = (a, (.P, y ) , Type, Border) 191 The fuzzy set o f primary classifications O is determined by D o CR2. We use min for the hij functions, max for f , prior rule confidence ru, but no thresholding or weakly monotonic methods in equation (34). Let us illustrate the method by finding o3 for Dead-Zone. We have ~33=min ((a, a)(Ra large, ( y , y)¢Ry low, Type RrO, Border RB1) (39) and C33 = min (e33, 7"33) (40) The value o f (a, a)(Ra large is the 64 output from the CRa EFR. The relation Type RrO is one if Type = 0 and zero otherwise, and Border R s 1 is one if Border = 1 and zero otherwise. Similarly, we compute c53, c73, ct3, and so forth and obtain 03 = f ( ¢ 3 3 , C53, C73, "" ") (41) Using (R2, the entire primary classification becomes one rule: if the data is D, then the classification is O (42) In a fuzzy expert system not all rules or blocks o f rules may be modeled by extended fuzzy relations. One obvious situation is where the right side o f a rule calls for user interaction. Any action in the right side o f a rule that cannot be interpreted as making or changing a fuzzy set will not come under the domain o f extend fuzzy relations. PARALLEL VERSUS SEQUENTIAL An important question being asked by many computer people (both algorithm and hardware oriented) is the question o f when a problem is suitable for parallel processing. In his book Uhr [12] points out that parallel processing will be an important part o f future computing systems but that many obstacles must be overcome prior to widespread use o f these techniques. Through our work with a parallel rule-firing expert system (Siler, Tucker, and Buckley [9]), we have also become interested in this question. We would like to be able to quantify those characteristics which are possessed by problems suitable for parallel processing and use this as a means o f discriminating between 192 J . J . Bucldey et al. those problems which are suitable for parallel processing and those which are not. The motivation for such work is to us obvious. It has been argued that parallel processing will result in an increase in system performance. We have shown [9] that for a particular problem, that of echocardiogram image analysis, there is an overall increase in run-time performance by a factor o f 6 using a software emulation o f a parallel machine. Unfortunately, not all problems are ones that stand to benefit from use o f parallel computing techniques. In a broad sense we have grouped those problems which yield to inductive reasoning as being suitable for parallel processing, and those which yield to deductive reasoning as being suitable for sequential processing. This measure is a qualitative one and gives no indication o f the specific nature o f the parallelism within the problem. A problem may be one that does not fit this classification scheme well, having subproblems that are both inductive and deductive in nature. What is needed is a quantitative method for expressing the " a m o u n t " o f parallelism that a problem possesses and where that parallelism lies. Once this has been established, we may better apply existing techniques to solve the problem at hand. To this end we propose that the development o f an EFR or a set o f related EFRs is a technique whereby the parallel nature o f a problem can be demonstrated. In the preceding discussion we have shown that the evaluation o f an EFR is (minimally) equivalent to the processing o f several rules simultane- ously. On a suitably designed machine, we think that an EFR can be effectively evaluated in one step. What we must do is to describe such a machine and show where the parallelism lies. For such a machine, the input would be as described above, a vector o f data. In a preprocessing step, copies o f the data, one for each nonempty (ij) position in the EFR, would be produced and distributed to the processing elements, one processing element for each nonempty (ij) position. After this preprocessing, the evaluation o f each o f the inner products di * R 0 and the cij would take place. In a final post-processing step, the outputs o f the rn processors in each column can be aggregated to produce the resulting fuzzy output O. According to our definition o f a level g >_ 1 EFR, each Rij may be a vector o f Ki extended fuzzy relations. For the EFR to be evaluated as efficiently as possible, each o f the Ki relations Rijk must be processed simultaneously. This requires that there be Ki such machines at each o f the (ij) processing elements in (R. This results in a hierarchical view o f the parallelism that may exist in a problem. If we look at the echocardiogram analysis example presented earlier, we see that this is an EFR in which each of the relational components Rij can contain three EFRs and two binary relations. Clearly this could be represented as a problem with two levels o f parallelism, the first (lower level) being the processing o f area and centroid data to generate suitable fuzzy sets for the Extended Fuzzy Relations: Application 193 higher-level process, the generation o f the preliminary classifications (see Figure 2). In our production system, operating in parallel mode, we emulate the machine just described. The production s y s t e m ' s working m e m o r y acts as a central store for data that are accessible to all the R i f t c relations, and all those R i j k relations which have suitable data are made active at once. The m e m o r y management techniques (weakly monotonic) described in the section on the generalized m o d u s p o n e n s function as the post-processing step and produce a column output for the EFR. S U M M A R Y A N D C O N C L U S I O N S This article first introduced the idea o f an extended fuzzy relation 6t. The relation between two elements, say u and v, is u61v = (R1, " " , R x ) , where each Rk is a regular fuzzy relation or an extended fuzzy relation. Another type o f extended fuzzy relation has been considered by Bezdek, Pettus, Stephens, and Zhang [1] and Zhang, Bezdek, Pettus, and Stephens [17], where u61v is a fuzzy set representing a linguistic variable. Their application is knowledge representa- tion or information retrieval in an expert system. Our application is the use o f extended fuzzy relations to replace blocks o f rules in a fuzzy expert system with one rule. We have shown that our procedure contains the generalized m o d u s p o n e n s and have also shown how to construct extended fuzzy relations in two examples. W e have then argued that knowing that an extended fuzzy relation exists for a block o f rules may be the key to knowing when to process these rules in parallel instead o f sequentially. Firing rules in parallel, as opposed to sequentially, has two main advantages. First, a rule conflict algorithm, to determine which rule to fire when a group o f rules become fireable, is not needed. Second, there is no stacking o f unfired rules with subsequent backtracking. Parallel execution can result in a substantial reduction in system overhead and an increase in computational efficiency [9]. However, when operating in a parallel mode we need a m e m o r y conflict algorithm. When the system attempts to execute several rules all having the same conclusion, we need to decide on the final confidence we will place in this conclusion. Our system employs weakly monotonic logic for m e m o r y conflict resolution. The incorporation o f weakly monotonic logic into the extended fuzzy relation technique has been discussed in the section on the generalized m o d u s p o n e n s . Future research is needed to determine existence theorems for extended fuzzy relations. Results such as the following are needed for expert systems: I f your problem has characteristics ~,/3, .y, • • . , then theorem 3 says that there exists an extended fuzzy relation for this problem. Therefore, characteristics c~,/3, 3', " " " are sufficient for putting the problem on a parallel machine. 194 J . J . Buckley et al. References 1. Bezdek, J. C., Pettus, R. O., Stephens, L. M., and W.-R. Zhang, Knowledge representation using linguistic fuzzy similarity relations, Int. J. Man-Machine Stud. (to appear). 2. Buckley, J. J., and Siler, W., Echocardiogram analysis using fuzzy numbers and relations, Fuzzy Sets & Syst. (to appear). 3. Buckley, J. J., Siler, W., and Tucker, D., A fuzzy expert system, Fuzzy Sets & Syst. 20, 1-16, 1986. 4. Buckley, J. J., Siler, W., and Tucker, D., FLOPS, A fuzzy expert system: Applications and perspectives, in Fuzzy Logics in Knowledge Engineering (C. V. Negoita, and H. Prade, Eds.), Rheinland, Germany, Verlag TUV, 256-274, 1986. 5. Mizumoto, M., and Zimmermann, H.-J., Comparison of fuzzy reasoning methods, Fuzzy Sets & Syst. 8, 253-283, 1982. 6. Mizumoto, M., Fuzzy reasoning under new compositional rules of inference, Kybernetics 12, 107-117, 1985. 7. Siler, W., Tucker, D., Buckley, J. J., Hess, R. G., and Powell, V. G., Artificial intelligence in processing a sequence of time-varying images, Proceedings o f the International Society f o r Optical Engineering 548, 194-199, 1985. 8. Siler, W., and Tucker, D., FLOPS, A Fuzzy Logic Production System User's Manual, Kemp-Carraway Heart Institute, Birmingham, Ala. 1986. 9. Siler, W., Tucker, D., and Buckley, J. J., A parallel rule firing fuzzy production system with resolution of memory conflicts by weak fuzzy monotonicity, applied to the classification of multiple objects characterized by multiple uncertain features, Int. J. Man-Machine Studies (to appear). 10. Siler, W., Buckley, J. J., and Tucker, D., Functional requirements for a fuzzy expert system shell, in Artificial Intelligence: Applications o f Quantitative Reasoning (L. Zadeh and E. Sanchez, Eds.), Pergamon Press, New York, 21-31, 1987. 11. Tucker, D., Siler, W., Powell, V. G., and Stanley, A. W. H., FLOPS: A fuzzy expert system used in unsupervised enchocardiogram analysis, Computers in Cardiology, 341-344, 1985. 12. Uhr, L., Algorithm-Structured Computer Arrays and Networks, Academic Press, New York, 1984. 13. Whalen, T., and Schott, B., Alternative logics for approximate reasoning in expert systems: A comparative study, Int. J. Man-Machine Stud. 22, 327-346, 1985. 14. Yager, R. R., On the implication operator in fuzzy logic, Info, Sci. 31, 141-164, 1983. Extended Fuzzy Relations: Application 195 15. Zadeh, L. A., A theory of approximate reasoning, in Machine Intelligence 9 (J. E. Hayes, D. Michie, and L. I. Kulich, Eds.), John Wiley, New York, 149-194, 1979. 16. Zadeh, L. A., The role of fuzzy logic in the management of uncertainty in expert systems, F u z z y Sets & Syst. 11, 199-227, 1983. 17. Zhang, W. R., Bezdek, J. C., Pettus, R. O., and Stephens, L. M., Linguistic fuzzy relations in expert systems, presented at the Third Annual Computer Science Symposium on Knowledge Based Systems: Theory and Applications, Univ. South Carolina, Columbia, So. Carolina, 1986. 
work_njzbswnor5hvngpfyf2j222agy	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy An expert system based on least square support vector machines for diagnosis of the valvular heart disease Davut Hanbay * Firat University, Technical Education Faculty, Department of Electronics and Computer Science, 23119 Elazig, Turkey a r t i c l e i n f o Keywords: Valvular heart disease Wavelet packet decomposition Entropy Fast-Fourier transform Least squares support vector machine a b s t r a c t There has been a growing research interest in the use of intelligent methods in biomedical studies. This is the result of developments in the area of data analysis and classifying techniques. In this paper, an expert system based on least squares support vector machines (LS-SVM) for diagnosis of valvular heart disease (VHD) is presented. Wavelet packet decomposition (WPD) and fast-Fourier transform (FFT) methods are used for feature extracting from Doppler signals. LS-SVM is used in the classification stage. Threefold cross-validation method is used to evaluate the proposed expert system performance. The performances of the developed systems were evaluated in 105 samples that contain 39 normal and 66 abnormal subjects for mitral valve disease. The results showed that this system is effective to detect Doppler heart sounds. The average correct classification rate was about 96.13% for normal subjects and abnormal subjects. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction In recent years, realized medical studies showed that important causes of human deaths in the world are heart diseases. The heart valve disorders are of importance among the heart diseases. Among them, mitral and aortic valve disorders are the most com- mon ones. For this reason, the early detection of valvular heart dis- eases is one of the most important medical research areas (Akay, Akay, & Welkowitz, 1992; Turkoglu, Arslan, & Ilkay, 2002). Nowa- days, the used methods for diagnosis of the valvular heart diseases are non-invasive techniques (electrocardiograms, chest X-rays, heart sounds and murmur from stethoscope, ultrasound imaging and Doppler techniques) and invasive techniques (angiography and transozefagial echocardiograph) (Nanda, 1993). However, each method is limited in its ability to offer efficient and thorough detection and characterization. All these methods are based on experience and information of physician. Therefore, developing Human–Machine interfaces with the existing methods of studies has become popular in these areas. By using these interfaces, the cardiologist can understand the output of the examination systems more easily and diagnose the problem more accurately (Philpot, Yoganathan, & Nanda, 1993; Turkoglu et al., 2002). Doppler techniques are the most preferred methods because they are completely non-invasive and without a risk in the serial studies (Keeton & Schlindwein, 1997; Turkoglu et al., 2002). In the recent years, Doppler technique has found increasing use in the medical area (Wright, Gough, Rakebrandt, Wahab, & Wood- cock, 1997). Doppler heart sounds (DHS) are one of the most important sounds produced by blood flow, valves motion and vibration of the other cardiovascular components (Jing, Xuemin, Mingshi, & Wie, 1997). However, the factors such as calcified dis- ease or obesity often result in a diagnostically unsatisfactory Dopp- ler techniques assessment, therefore, it is sometimes necessary to assess the spectrogram of the Doppler shift signals to elucidate the degree of the disease (Wright et al., 1997). The main aim of our work is to aid the diagnosis in such cases. Among Doppler tech- niques, the most ubiquitous and straightforward are the waveform profile indices such as the pulsatility index (PI), Pourcelot or resis- tance index (RI) and A/B Systolic Diastolic ratio, which are highly correlated and led to highly erroneous diagnostic results (Izzeto- glu, Erkmen, & Beksac, 1995; Turkoglu et al., 2002). These indices rely on the peak systolic and end-diastolic velocities, with only the PI making use of the mean velocity over the cardiac cycle. More sophisticated methods have also been developed such as the La- place transform and principal components analysis. However, none of the simple or more complex analytical techniques has yielded an acceptable diagnostic accuracy so as to be commonplace in the vascular clinic (Wright et al., 1997). In this study, the developed method is a decision support system and will cause more effective usage of the Doppler technique. Until now, many studies have been realized to automatically classify Doppler signals using pattern rec- ognition techniques (Chan, Chan, Lam, Lui, & Poon, 1997; Guler & Kara, 1995; Turkoglu et al., 2002). Nevertheless, the effective stud- ies on the Doppler heart sounds are also limited (Turkoglu, Arslan, & Ilkay, 2002a). The Doppler heart sounds can be obtained simply by placing the Doppler ultrasonic flow transducer over the chest of the patient. 0957-4174/$ - see front matter � 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2008.04.010 * Tel.: +90 424 2370000 4257; fax: +90 424 2367064. E-mail address: dhanbay@firat.edu.tr Expert Systems with Applications 36 (2009) 4232–4238 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy But the disadvantage of the Doppler method is that it requires the constant attention of the doctor to detect subtle changes in the DHS (Chan et al., 1997). The presented method prevents subtle changes in the DHS from escaping the physician’s attention by per- ceiving them, even if the physician does not pay a continuous attention (Turkoglu, Arslan, & Ilkay, 2002b). Up to now, several papers have been proposed for classifying Doppler signals by using pattern recognition techniques (Chan et al., 1997; Çomak, Arslan, & Türkoğlu, 2007; Turkoglu et al., 2002b; Uguz, Arslan, & Türkoğlu, 2007; Wright et al., 1997). The data set which was obtained by Turkoglu et al. (2002b) was used by Çomak et al. (2007) and Uguz et al. (2007) too. Çomak et al. pro- posed to develop a decision support system based on wavelet decomposition and short-time Fourier transform to develop the performance of Turkoglu et al. (2002b). Uguz et al. proposed a bio- medical system based on continuous hidden Markov model classi- fier for the diagnosis of valvular heart diseases. The proposed methodology was composed of two stages. At the first stage, the initial values of average and standard deviation were calculated by separating observation symbols into equal segments according to the state number and using observation symbols appropriate to each segment. At the second stage, the initial values of average and standard deviation were calculated by separating observation sym- bols into the clusters (FCM or K-means algorithms) that have equal number of states and using observation symbols appropriate to the separated clusters. Least square SVM uses equality constraints and solves a set of linear equations in the dual space instead of solving a quadratic programming problem as for the standard SVM. This simplifies the computations and enhances the speed considerably. There ex- ists a link between the LS-SVM classifier formulations with the well-known Fisher discriminant analysis, namely by extending it to a high-dimensional feature space. Some parameters have to be tuned to achieve a high level performance of the LS-SVM, including the regularization parameter and the kernel parameter corre- sponding to the kernel type (Lukas, Devos, Suykens, et al., 2004; Suykens & Vandewalle, 1999). This study will introduce the technique that will aid clinical diagnosis, enable the further research of VHD, and provide a deci- sion support system for recognition of VHD. This study uses the powerful mathematics of wavelet signal processing and entropy, FFT to efficiently extract the features from pre-processed Doppler signals for the purpose of recognizing between abnormal and nor- mal subjects of the VHD. Thus, the doctor can make a comparison between the diagnoses by the developed method and the diagno- ses by the existing methods. If the results are different, the exam- inations can be repeated or performed more carefully. In this way, the physician can decide more realistically. This paper is organized as follows. In Section 2, we review some basic properties of the pattern recognition, the Doppler heart sig- nals, WPD, FFT, wavelet entropy and LS-SVM. In Section 3, the implementation stage is described and the effectiveness of the pro- posed method for the classification of Doppler signals in the diag- nosis of VHD is demonstrated. Finally, conclusion is presented in Section 4. 2. Preliminaries In this section, the theoretical foundations of the presented study are given in the following subsections. 2.1. Pattern recognition Pattern recognition consists of some sequential stages, the first stage is feature extraction from the patterns, which is the conver- sion of patterns to features that are regarded as a condensed rep- resentation, ideally containing all-important information. The second stage is feature selection. At which a smaller number of meaningful features that best represents the given pattern without redundancy is identified. The next stage is classification. At which a specific pattern is assigned to a specific class according to the char- acteristic features selected for it. This general abstract model is shown in Fig. 1, allows a broad variety of different realizations and implementations. Applying this terminology to the medical diagnostic process, the patterns can be identified, for example, as particular, formalized symptoms, recorded signals, or as a set of images of a patient. The classes obtained represent the variety of different possible diagnoses or diagnostic statements (Dickhous & Heinrich, 1996; Turkoglu et al., 2002). The techniques applied to pattern recognition use artificial intelligence approaches (Bishop, 1996; Turkoglu et al., 2002). 2.2. DHS signals The audio DHS can be obtained by simply placing the Doppler ultrasonic flow transducer over the chest of the patient (Chan et al., 1997; Turkoglu et al., 2002). (Fig. 2) shows a DHS signal. The DHS produced from echoes backscattered by moving blood cells is generally in the range of 0.5–10 kHz (Saini, Nanda, & Maulik, 1993). DHS signal spectral estimation is now commonly used to evaluate blood flow parameters in order to diagnose cardiovascular diseases. Spectral estimation methods are particularly used in Doppler ultrasound cardiovascular disease detection. Clinical diag- nosis procedures generally include analysis of a graphical display and parameter measurements, produced by blood flow spectral evaluation. Ultrasonic instrumentation typically employs Fourier- based methods to obtain the blood flow spectra and blood flow measurements (Madeira, Tokhi, & Ruano, 2000; Turkoglu et al., 2002). A Doppler signal is not a simple signal. It includes random char- acteristics due to the random phases of scattering particles present in the sample volume. Other effects such as geometric broadening feature extraction / selection classification training learning patterns classes Fig. 1. The general abstract model of pattern recognition approach. 0 1 2 3 4 5 -1 -0.5 0 0.5 1 Time (sec) A m pl it ud e (v ol t) Fig. 2. The waveform pattern of the Doppler heart sound. D. Hanbay / Expert Systems with Applications 36 (2009) 4232–4238 4233 Author's personal copy and spatially varying velocity also affect the signal (Turkoglu et al., 2002). The following is the Doppler equation: Df ¼ 2vf cos h c ð1Þ where v equals the velocity of the blood flow, f equals the frequency of the emitted ultrasonic signal, c equals the velocity of sound in tis- sue (approximately 1540 m/s), Df equals the measured Doppler fre- quency shift, and h equals the angle of incidence between the direction of blood flow and the direction of the emitted ultrasonic beam (Saini et al., 1993; Turkoglu et al., 2002). 2.3. Wavelet packet decomposition Wavelet transforms are finding inversed use in fields as diverse as telecommunications and biology. Because of their suitability for analyzing non-stationary signals, they have become a powerful alternative to Fourier methods in many applications, where such signals abound (Akay, 1997; Burrus, Gopinath, & Guo, 1998; Coif- man & Wickerhauser, 1992; Devasahayam, 2000; Karabetsos, Papaodysseus, & Koutsouris, 1998; Liang & Nartimo, 1998; Quirog- a, 1998; Quiroga, Roso, & Basar, 1999). The main advantage of wavelets is that they have a varying win- dow size, being wide for slow frequencies and narrow for the fast ones, thus leading to an optimal time-frequency resolution in all the frequency ranges. Furthermore, owing to the fact that windows are adapted to the transients of each scale, wavelets lack the requirement for stationary (Quiroga, 1998). Wavelet decomposi- tion uses the fact that it is possible to resolve high-frequency com- ponents within a small time window, while only low-frequency components need large time windows. This is because a low-fre- quency component completes a cycle in a large time interval, whereas a high-frequency component completes a cycle in a much shorter interval. Therefore, slow varying components can only be identified over long time intervals but fast varying components can be identified over short time intervals. Wavelet decomposition can be regarded as a continuous time wavelet decomposition sam- pled at different frequencies at every level or scale. The wavelet decomposition functions at level m and time location tm can be ex- pressed as the following equation: dmðtmÞ¼ xðtÞ � Wm t � tm 2m � � ð2Þ where Wm is the decomposition filter at the frequency level m. The effect of the decomposition filter is scaled by the factor 2m at stage m, but otherwise the shape is the same at all scales. Wavelet packet analysis is an extension of the discrete wavelet transform (DWT) (Burrus et al., 1998) and it turns out that the DWT is only one of the much possible decomposition that could be performed on the signal. Instead of just decomposing the low-frequency component, it is therefore possible to subdivide the whole time-frequency plane into different time-frequency pieces as can be seen from Fig. 3. The advantage of wavelet packet analysis is that it is possible to com- bine the different levels of decomposition in order to achieve the optimum time-frequency representation of the original (Keeton & Schlindwein, 1997). 2.4. Fast-Fourier transform The discrete Fourier transform (DFT) of a sequence can be eval- uated directly by the following equation, which involves the order of N2 complex multiplications and additions, so these processes need more computation time. Because of this problem, a quicker method must be used to evaluate this equation. The fast-Fourier transform (FFT) provides such methods FðkXÞ¼ XN�1 n¼0 fðnTÞe�jXTnk 0 6 k 6 N � 1 ð3Þ where f is the impulse response of a finite impulse response filter, N is finite number, the Eq. (3) can be written in the form of FðkXÞ¼ XN�1 n¼0 fðnTÞW nkN ð4Þ where W N ¼ e�jð2p=NÞ. If N is even, then we can split the sequence f into even and odd terms. If N = 1024, then the FFT requires 1% of the time required for the discrete calculation of DFT (Banks, 1991). 2.5. Least square SVM LS-SVM was proposed by Suykens and Vandewalle (1999). It is the reformulation of the standard SVM, which was intended by Vap- nik in 1995 (Lukas et al., 2004). LS-SVMs generalization performance was compared with the standard SVM by Van Gestel et al. We review the LS-SVM formulation as follows: we are given a training data set of n data points fxi; yig n i¼1, where xi 2 R d is the ith input vector and yi 2 R is the label of ith class. For binary classification, yi takes only two values {-1, +1}, in regression stage it can take any real value. In kernel designs, the idea was to transform the input patterns into Reproducing Kernel Hilbert Space (RKHS) by a set of mapping func- tions /ðxÞ If reproducing kernel denoted in RKHS as K ¼ðx; x’Þ; which is defined as (Lukas et al., 2004; Suykens & Vandewalle, 1999) Kðx; x0Þ ¼ /ðxÞ � /ðxÞ0 ð5Þ In the RKHS, a linear classification is performed. The discriminant function takes the form yðxÞ¼ Xn i¼1 w � /ðxÞþ b ð6Þ where w is the weight vector in the RKHS, and b 2 R is bias term. The discriminant function of LS-SVM classifiers is constructed by minimizing the following primal problem: min w;b;n Pðw; b; nÞ¼ 1 2 kwk2 þ C 2 Xn i¼1 n2i ð7Þ subject to the equality constraints yi �ðw � /ðxiÞþ bÞ¼ ni8i, where the regularization parameter, C > 0. Note that the formulation for classifier design is same as that for regression. Traditionally, an inequality constraint is exploited on the slack variables ni to pun- ish the misclassified patterns only. Nevertheless, the formulation of LS-SVM produces penalty to all the patterns if their discrimi- nant function is not equal to the corresponding target value, as the traditional regression formulation does. Standard Lagrangian techniques are used to derive the dual problem (Lukas et al., 2004; Suykens & Vandewalle, 1999). The Lagrangian for the pri- mal problem is Lðw; b; n; aÞ¼ 1 2 kwk2 þ C 2 Xn i¼1 n2i þ Xn i¼1 aifyi �ðw:/ðxÞþ bÞ� nig ð8Þ Signal C D DWThigh-passlow-pass Terminal nodes Fig. 3. Wavelet packet decomposition. 4234 D. Hanbay / Expert Systems with Applications 36 (2009) 4232–4238 Author's personal copy where ai are Lagrangian multipliers. The Karush–Kuhn–Tucker (KKT) conditions for the primal problem are oL ow ¼ 0 ) w ¼ Xn i¼1 ai/ðxiÞ ð9Þ oL ob ¼ 0 ) Xn i¼1 ai ¼ 0 ð10Þ oL oni ¼ 0 ) ai ¼ Cni 8i ð11Þ Together with the KKT conditions for Lagrangian multipliers ai, Suy- kens and Vandewalle (1999) write them as a linear system, and sug- gest to solve two linear systems and then combine the results to get the final solution to the KKT linear system. 3. Methodology Fig. 4 shows the used decision support system block diagram. It consists of three parts: (a) data acquisition and pre-processing, (b) feature extraction, and (c) classification by LS-SVM. 3.1. Data acquisition and pre-processing All the original audio DHS signals were acquired from the Acu- son Sequoia 512 Model Doppler Ultrasound system in the Cardiol- ogy Department of the Firat Medical Center. DHS signals were sampled at 20 kHz for 5 s and signal-to-noise ratio of 0 dB by using a sound card which has a 16-bit A/D conversion resolution and a computer software prepared by us in the MATLAB (version 5.3) (The MathWorks Inc. Natick, MA, USA). The Doppler ultrasonic flow transducer used (Model 3V2c) was run at an operating mode of 2 MHz continuous wave. The Doppler signals of the heart valves were obtained by placing the transducer over the chest of the pa- tient with the aid of an ultrasonic image. The digitised data, which have 39 normal and 66 abnormal subjects, were stored on hard disk of the PC. The subject group consisted of 58 males and 47 fe- males with the ages ranging from 19 to 78 years. The average age of the subjects was 47.5 years. Pre-processing to obtain the feature vector was performed on the digitized signal in the following order. 3.1.1. Filtering The reserved DHS signals were high-pass filtered to remove un- wanted low-frequency components, because the DHS signals are generally in the range of 0.5–10 kHz. The filter is a digital FIR, which is a fiftieth-order filter with a cut-off frequency equal to 500 Hz and window type is the 51-point symmetric Hamming window. 3.1.2. White de-noising White noise is a random signal that contains equal amounts of every possible frequency, i.e., its FFT has a flat spectrum (Devasa- hayam, 2000). The DHS signals were filtered by removing the white noise by using a wavelet packet. The white de-noising procedure contains three steps (Bakhtazad, Palazoglu, & Romagnoli, 1999): (1) Decomposition: computing the wavelet packet decomposition of the DHS signal at level 4 and using the Daubechies wavelet of order 4; (2) Detail coefficient thresholding: for each level from 1 to 4, soft thresholding is applied to the detail coefficients; (3) Reconstruction: computing wavelet packet reconstruction based on the original approximation coefficients of level 4 and the mod- ified detail coefficients of levels from 1 to 4. 3.1.3. Normalization The DHS signals in this study were normalized using the follow- ing equation so that the expected amplitude of the signal is not af- fected by the rib cage structure of the patient, DHSsignal ¼ DHSsignal jðDHSsignalÞmaxj ð12Þ 3.2. Feature extraction Feature extraction is one of the important steps of pattern rec- ognition so that it is the most important step of designing the deci- sion support system based on pattern recognition because the best classifier performance will decrease if the features are not chosen well. The feature extraction stage must reduce the pattern vector (i.e., the original waveform) to a lower dimension, which contains most of the useful information from the original vector. The DHS waveform patterns from heart valves are rich in detail and highly non-stationary. In this study, the goal of the feature extraction is to extract features from these patterns for intelligent classification. After the data pre-processing has been realized, three steps are proposed in this paper to extract the characteristics of these wave- forms using MATLAB with the wavelet toolbox and the signal pro- cessing toolbox. 3.2.1. Wavelet packet decomposition: to decomposition To evaluate the wavelet packet decomposition, the tree struc- ture was used. Wavelet packet decomposition was applied to the DHS signal with the Daubechies-1 wavelet packets using the Shan- non entropy as defined in the following equation. In this equation, s is the DHS signal and si is ith coefficients of wavelet packet decomposition of s. A representative example of the wavelet pack- et decomposition of the Doppler sound signal of the heart mitral valve is shown in Fig. 5, EðsÞ¼� X i s2i � logðs 2 i Þ ð13Þ PREPROCESSING Data acquisition Filtering White de-noising Normalization Doppler Ultrasound FEATURE EXTRACTION Wavelet Packet FFT Wavelet entropy CLASSIFICATION LS-SVM Cleaned DHS Signal (The heart valves) Feature Space Decision Space CLASSIFICATION RESULTS Abnormal valve Normal valve Fig. 4. The algorithm of the expert system. D. Hanbay / Expert Systems with Applications 36 (2009) 4232–4238 4235 Author's personal copy 3.2.2. Fast-Fourier transform The FFT is the most robust and understood one of the various time-frequency representations (Keeton & Schlindwein, 1997; Turkoglu et al., 2002). Convert to frequency domain of waveforms of terminal nodes was computed using a 512-point FFT. Because the length of the terminal node signal was less than 512, FFT pad of the terminal node signal was constituted with trailing zeros to length 512. A representative example of the FFT spectrum of wave- form of a terminal node is indicated in Fig. 6. 3.2.3. Wavelet entropy We next calculated the norm entropy as defined in the follow- ing equation of waveforms of the FFT spectrum: EðsÞ¼ X i jsij 3=2 ð14Þ where s is the FFT spectrum and si is ith coefficients of s Quiroga et al. (1999). The resultant entropy data, which were normalized with 1/1000, are plotted in Fig. 7. The plot of the entropy data includes 256 features obtained from 256 terminal nodes, where each one contains waveform of one frequency spectrum per DHS signal. Thus, the feature vector was obtained by computing the wavelet packet entropy values for each DHS signal. 3.3. Classification using LS-SVM The objective of the classification is to demonstrate the effec- tiveness of the proposed feature extraction method from the DHS signals. For this purpose, the feature vectors were applied as the in- put to an LS-SVM classifier. The KULeuven’s LS-SVMlab MATLAB/C Toolbox was used for the purpose of training and testing. RBF 0 1 2 3 4 5 -1 0 1 DHS signal -0.02 0 0.02 0 -5 0 5 -5x 10-3 -0.01 0 0.01 -0.05 0 0.05 -0.02 0 0.02 -0.05 0 0.05 0 100 200 300 400 -0.02 0 0.02 5x 10-3 A m pl it ud e (V ) A m pl it ud e Time (sec) Samples Fig. 5. The waveforms of terminal nodes (i = 1–256) of wavelet packet decomposition at eight-level of the DHS signal. 0 1 2 3 4 5 0 2000 4000 6000 8000 10000 0 0.1 0.2 Ti m e ( se c) Freq.(Hz) A m pl it ud e (V ) Fig. 6. The WPD-FFT based time-frequency representation. 4236 D. Hanbay / Expert Systems with Applications 36 (2009) 4232–4238 Author's personal copy kernel is used. Grid search algorithm is used to tune the c regular- ization constant and r width of RBF kernel parameters. The deter- mined optimal c value is 2.8439 and optimal r value is 29.316 for predicting mitral valve diseases. Threefold cross-validation meth- od was applied to the 105 experimental data sets for computing the validation of LS-SVM model. In k-fold cross-validation method, the data set is divided into k subsets, and the holdout method is re- peated k times. At each time, k � 1 subsets are used for training and kth subset is used for testing. Then the average error across all k trials is computed. Therefore, every data point gets to be in a test set exactly once, and gets to be in a training set k-1 times. Different evaluation methods were used for calculating the perfor- mance of the proposed expert system. The best test performance of LS-SVM model is graphically presented in Fig. 8. As seen in Fig. 8, LS-SVM model predicts the measured values at a high accuracy rate. Threefold test performance of LS-SVM model is shown in Ta- ble 1. The average correct classification rate is 96.13%. 4. Conclusions In this study, a decision support system was developed for the interpretation of the DHS signals using pattern recognition, and demonstrated the diagnosis performance of this method on the heart mitral valves. The task of feature extraction was performed using the wavelet packet decomposition for multi-scale analysis, FFT for time-frequency representations, and the wavelet entropy, while classification was carried out by LS-SVM. The stated results show that the proposed method can make an efficient interpretation. The feature choice was motivated by a realization that wavelet decomposition essentially is a representation of a signal at a vari- ety of resolutions. In brief, the wavelet packet decomposition has been demonstrated to be an effective tool for extracting informa- tion from the DHS signals. Moreover, the proposed feature extrac- tion method is robust against the noise in the DHS signals. In this paper, the application of the wavelet entropy to the fea- ture extraction from DHS signals was shown. Wavelet entropy proved to be a very useful tool for characterizing the DHS signal, furthermore the information obtained with the wavelet entropy proved not to be trivially related with the energy and consequently with the amplitude of the signal. This means that with this method, new information can be accessed with an approach different from the traditional analysis of amplitude of DHS signal. This system is of better clinical application over others, espe- cially for earlier survey of a population. However, the position of the ultrasound probe, which is used for data acquisition from the heart valves, must be taken into consideration by a physician. Although this decision support system was carried out on the heart mitral valves, similar results for the other valves (tricuspid and pulmonary) and other Doppler studies can be expected. Besides the feasibility of a real-time implementation of the decision sup- port system, by increasing the variety and number of DHS signals additional information (i.e., quantification of the heart valve regur- gitation and stenosis) can be provided for diagnosis. Acknowledgments The author thanks Dr. Ibrahim Turkoglu for his valuable support and suggestion in improving the technical presentation of this pa- per. The author thanks Dr. Erdogan ILKAY and the Cardiology Department of the Firat Medicine Center, Elazig/Turkey for provid- ing the DHS signals too. References Akay, M. (1997). Wavelet applications in medicine. IEEE Spectrum, 34(May), 50–56. Akay, M., Akay, Y. M., & Welkowitz, W. (1992). Neural networks for the diagnosis of coronary artery disease. International Joint Conference on Neural Networks, IJCNN, 2, 419–424. Bakhtazad, A., Palazoglu, A., & Romagnoli, J. A. (1999). Process data de-noising using wavelet transform. Intelligent Data Analysis, 3. Banks, S. P. (1991). Signal processing, image processing and pattern recognition. NJ, USA: Prentice Hall. Bishop, C. M. (1996). Neural networks for pattern recognition. Oxford: Clarendon Press. Burrus, C. S., Gopinath, R. A., & Guo, H. (1998). Introduction to wavelet and wavelet transforms. NJ, USA: Prentice Hall. Chan, B. C. B., Chan, F. H. Y., Lam, F. K., Lui, P. W., & Poon, P. W. F. (1997). Fast detection of venous air embolism is Doppler heart sound using the wavelet transform. IEEE Transactions on Biomedical Engineering, 44(4), 237–245. Coifman, R. R., & Wickerhauser, M. V. (1992). Entropy-based Algorithms for best basis selection. IEEE Transactions on Information Theory, 38(2), 713–718. Çomak, E., Arslan, A., & Türkoğlu, I. (2007). A decision support system based on support vector machines for diagnosis of the heart valve diseases. Computers in Biology and Medicine, 37(1), 21–27. Devasahayam, S. R. (2000). Signals and systems in biomedical engineering. Kluwer Academic Publishers. Dickhous, H., & Heinrich, H. (1996). Classifying biosignals with wavelet networks. IEEE Engineering in Medicine and Biology, 103–111. Guler, I., & Kara, S. (1995). Detection of mitral stenosis by a pulsed Doppler flow meter and autoregressive spectral analysis method. 14th Conference of the biomedical engineering society of India. An international meeting, proceedings of the first regional conference, IEEE (pp. 2/91–2/92). Izzetoglu, K., Erkmen, A. M., & Beksac, S. (1995). Intelligent classification of fetal Doppler blood velocity waveform abnormalities using wavelet transform and 0 200 400 600 0 2 4 6 8 10 E nt ro py Feature index Fig. 7. The WP entropy of DHS signal. 0 5 10 15 20 25 30 35 -1 -0.5 0 0.5 1 1.5 2 Sample O ut pu t Desired Predicted Fig. 8. The best test performance of LS-SVM model for mitral valve diseases. Table 1 Threefold test performance of LS-SVM model Data set (105) Correct classified Wrong classified Performance (%) Training sets Test sets Set-1, set-2 (70) Set-3 (35) 35 0 100 Set-1, set-3 (70) Set-2 (35) 33 2 94.2 Set-2, set-3 (70) Set-1 (35) 33 2 94.2 Average performance 96.13 D. Hanbay / Expert Systems with Applications 36 (2009) 4232–4238 4237 Author's personal copy vector quantization algorithm. IEEE International Conference on Systems, Man and Cybernetics, Intelligent Systems for the 21st Century, 1, 724–729. Jing, F., Xuemin, W., Mingshi, W., & Wie, L. (1997). Noninvasive acoustical analysis system of coronary heart disease. In Biomedical engineering conference, proceedings of the 1997 sixteenth southern (pp. 239–241). Karabetsos, E., Papaodysseus, C., & Koutsouris, D. (1998). Design and development of a new ultrasonic doppler technique for estimation of the aggregation of red blood cells. Measurement, 24, 207–215. Madeira, M. M., Tokhi, M. O., & Ruano, M. G. (2000). Real-time implementation of a Doppler signal spectral estimator using sequential and parallel processing techniques. Microprocessors and Microsystems, 24, 153–167. Nanda, N. C. (1993). Doppler echocardiography (2nd ed.). London: Lea & Febiger. Philpot, E. F., Yoganathan, A. P., & Nanda, N. C. (1993). Future directions in Doppler echocardiography. Doppler echocardiography. London: Lea & Febiger Philadelphia. Quiroga, R. Q. (1998). Quantitative analysis of EEG signals: Time–frequency methods and Chaos theory. Intitute of Physiology – Medical University Lübeck. Quiroga, R. Q., Roso, O. A., & Basar, E. (1999). Wavelet entropy: A measure of order in evoked potentials. Evoked potentials and magnetic fields (Vol. 49). (pp. 298–302). Saini, V. D., Nanda, N. C., & Maulik, D. (1993). Basic principles of ultrasound and Doppler effect. Doppler echocardiography. London: Lea & Febiger Philadelphia. Suykens, J. A. K., & Vandewalle, J. (1999). Least squares support vector machine classifiers. Neural Processing Letters, 9(3), 293–300. Turkoglu, I., Arslan, A., & Ilkay, E. (2002). A pattern recognition system for diagnose of the heart mitral valve diseases based on wavelet packet and neural network. Fırat University, Journal of Science and Technology, 14(1), 1–10. Turkoglu, I., Arslan, A., & Ilkay, E. (2002a). A decision support system for diagnose of the heart valve diseases. Journal of the Institute of Science and Technology of Sakarya University, 6(2), 57–64. Turkoglu, I., Arslan, A., & Ilkay, E. (2002b). An expert system for diagnosis of the heart valve diseases. Expert Systems with Applications, 23(3), 229–236. Uguz, H., Arslan, A., & Türkoğlu, I. (2007). A biomedical system based on hidden Markov model for diagnosis of the heart valve diseases. Pattern Recognition Letters. Wright, I. A., Gough, N. A. J., Rakebrandt, F., Wahab, M., & Woodcock, J. P. (1997). Neural network analysis of Doppler ultrasound blood flow signals: A pilot study. Ultrasound in Medicine and Biology, 23(5), 683–690. 4238 D. Hanbay / Expert Systems with Applications 36 (2009) 4232–4238 EXPERT SYSTEMS WITH APPLICATIONS An International Journal EDITOR-IN-CHIEF Jay Liebowitz, D.Sc. 966 Farm Haven Drive, Rockville, Maryland 20852, USA EDITORIAL ADVISORY BOARD Edward Feigenbaum Stanford University, USA Lance B. Eliot University of Southern California, USA Dianne C. Berry University of Reading, UK Walter Reitman Rensselaer Polytechnic Institute, USA Shri K. Goyal GTE Laboratories, USA Daniel E. O’Leary University of Southern California, USA Richard G. Vedder University of North Texas, USA James L. Rash NASA Goddard Space Flight Center, USA Randall Shumaker Naval Research Laboratory, USA David Bendel Hertz University of Miami, USA A. Desai Narasimhalu Singapore Management University Patricia Lightfoot NASA Goddard Space Flight Center, USA Mark Fox Toronto, Ontario, Canada Jae Kyu Lee Korea Advanced Institute of Science and Technology I. Burhan Turksen University of Toronto, Canada Ching Suen Concordia University, Canada Ernest H. Forman George Washington University, USA Roar Fjellheim Computas A.S., Norway Jan Vanthienen Catholic University, Leuven, Belgium Jon R. Wright AT&T Bell Laboratories, USA Roy Rada University of Maryland, USA John P. Coyne George Washington University, USA John H. Carson George Washington University, USA Daniel Schutzer Citicorp Investment Bank, USA Paul Harmon Harmon Associates, USA Francisco J. Cantu Instituto Tecnologico y de Estudios Superiores de Monterrey, Mexico Brian Gaines University of Calgary, Canada Michel Clerget Cognitech, France Kenneth J. Fordyce IBM, USA James R. Slagle University of Minnesota, USA Eliot Weinman Software Productivity, USA Randall Davis Massachusetts Institute of Technology, USA E. Balagurusamy Mahareer Academy of Technology and Sciences, India Dieter Specht Technische Universitat Cottbus, Germany Vladimir Milacic University of Belgrade, Yugoslavia Hans Bergkvist ATTEXOR S. A., Switzerland Laura C. Davis U.S. Navy Center for Applied Research in Artificial Intelligence, USA Jerald Feinstein MITRE, USA Efraim Turban University of Hawaii, 3435 Kehala Dr., Kiehi HI 96753, USA Bernt Bremdal GeoKnowledge, Norway Milton White Datanamics Inc., USA Editorial Office: Jay Liebowitz, 966 Farm Haven Drive, Rockville, Maryland 20852, USA. Tel./Fax: (+1) 301-770-2978; e-mail: jliebow1@jhu.edu Author enquiries: For enquiries relating to the submission of articles (including electronic submission where available) please visit this journal’s homepage at http://www.elsevier.com/ locate/eswa. You can track accepted articles at http://www.elsevier.com/trackarticle and set up e-mail alerts to inform you of when an article’s status has changed. Also accessible from here is information on copyright, frequently asked questions and more. Contact details for questions arising after acceptance of an article, especially those relating to proofs, are provided after registration of an article for publication. Publication information: Expert Systems with Applications (ISSN 0957-4174). For 2009, volumes 36, 37 are scheduled for publication. Subscription prices are available upon request from the Publisher or from the Regional Sales Office nearest you or from this journal’s website (http://www.elsevier.com/locate/eswa). Further information is available on this journal and other Elsevier products through Elsevier’s website: (http://www.elsevier.com). Subscriptions are accepted on a prepaid basis only and are entered on a calendar year basis. Issues are sent by standard mail (surface within Europe, air delivery outside Europe). Priority rates are available upon request. Claims for missing issues should be made within six months of the date of dispatch. Advertising information: Advertising orders and enquiries can be sent to: James Kenney, Elsevier Ltd., 32 Jamestown Road, London NW1 7BY, UK; phone: (+44) 207 424 4216; fax: (+44) 1865 853 136; e-mail: j.kenney@elsevier.com. Customers in the US and Canada can also contact: Mr Tino DeCarlo, Advertising Department, Elsevier Inc., 360 Park Avenue South, New York, NY 10010-1710, USA; phone: (+1) (212) 633 3815; fax: (+1) (212) 633 3820; e-mail: t.decarlo@elsevier.com. Orders, claims, and journal enquiries: please contact the Regional Sales Office nearest you: St. Louis: Elsevier, Customer Service Department, 11830 Westline Industrial Drive, St. Louis, MO 63146, USA; phone: (877) 8397126 [toll free within the USA]; (+1) (314) 4537076 [outside the USA]; fax: (+1) (314) 5235153; e-mail: JournalCustomerService-usa@elsevier.com Amsterdam: Elsevier, Customer Service Department, PO Box 211, 1000 AE Amsterdam, The Netherlands; phone: (+31) (20) 4853757; fax: (+31) (20) 4853432; e-mail: JournalsCustomer- ServiceEMEA@elsevier.com Tokyo: Elsevier, Customer Service Department, 4F Higashi-Azabu, 1-Chome Bldg, 1-9-15 Higashi-Azabu, Minato-ku, Tokyo 106-0044, Japan; phone: (+81) (3) 5561 5037; fax: (+81) (3) 5561 5047; e-mail: JournalsCustomerServiceJapan@elsevier.com Singapore: Elsevier, Customer Service Department, 3 Killiney Road, #08-01 Winsland House I, Singapore 239519; phone: (+65) 63490222; fax: (+65) 67331510; e-mail: JournalsCustomer- ServiceAPAC@elsevier.com USA mailing notice: Expert Systems with Applications (ISSN 0957-4174) is published 10 issues per annum in January, March, April, May, July, August, September, October, November, December by Elsevier Ltd. (The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, UK). Periodical postage paid at Rahway NJ and additional mailing offices. USA POSTMASTER: Send change of address: Expert Systems with Applications, Elsevier, Customer Service Department, 11830 Westline Industrial Drive, St. Louis, MO 63146, USA. Periodical postage paid at Rahway NJ and additional mailing offices. AIRFREIGHT AND MAILING in USA by Mercury International Limited, 365, Blair Road, Avenel, NJ 07001. 
work_nlcqi2h725a3vh67xurzf5wy7a	----	Text Summarization Using Unsupervised Deep Learning Mahmood Yousefi-Azar and Len Hamey Department of Computing Faculty of Science and Engineering Macquarie University, Sydney, NSW, Australia Email: mahmood.yousefiazar@students.mq.edu.au; len.hamey@mq.edu.au May 19, 2016 Abstract We present methods of extractive query-oriented single-document summarization using a deep auto-encoder (AE) to compute a feature space from the term-frequency (tf ) input. Our experiments explore both local and global vocabularies. We investigate the effect of adding small random noise to local tf as the input representation of AE, and propose an ensemble of such noisy AEs which we call the Ensemble Noisy Auto-Encoder (ENAE). ENAE is a stochastic version of an AE that adds noise to the input text and selects the top sentences from an ensemble of noisy runs. In each individual experiment of the ensemble, a different randomly generated noise is added to the input representation. This architecture changes the application of the AE from a deterministic feed-forward network to a stochastic runtime model. Experiments show that the AE using local vocabularies clearly provide a more discriminative feature space and improves the recall on average 11.2%. The ENAE can make further improvements, particularly in selecting informative sentences. To cover a wide range of topics and structures, we perform experiments on two different publicly available email corpora that are specifically designed for text summarization. We used ROUGE as a fully automatic metric in text summarization and we presented the average ROUGE-2 recall for all experiments. I. INTRODUCTION Text summarization is an automatic technique to generate a condensed version of the original documents. Manual summa- rization requires a considerable number of qualified unbiased experts, considerable time and budget and the application of the automatic techniques is inevitable with the increase of digital data available world-wide. The early technique to address with manual text summarization dates back as early as 1958 (Luhn, 1958) and the proposed techniques were reviewed extensively (Lloret and Palomar, 2012; Nenkova and McKeown, 2012). Text summarization can be categorized into two distinct classes: abstractive and extractive. In the abstractive summariza- tion, the summarizer has to re-generate either the extracted content or the text; however, in extractive category, the sentences have to be ranked based on the most salient information. In many research studies extractive summarization is equally known as sentence ranking (Edmundson, 1969; Mani and Maybury, 1999). In practice, specific text summarization algorithm is needed for different tasks. In particular, a summarization technique can be designed to work on a single document, or on a multi-document. Similarly, the purpose of summarization can be to produce a generic summary of the document, or to summarize the content that is most relevant to a user query. The focus of this paper is to propose an extractive query-oriented single-document summarization technique. Deep learning showed strong promise in various areas, specifically in natural language processing (NLP) tasks (Collobert et al., 2011; Srivastava and Salakhutdinov, 2012). The pivot of our model is a deep auto-encoder (AE) (Hinton and Salakhut- dinov, 2006a) as an unsupervised model. The AE learns the latent representations for both the query and the sentences in the document. In this paper, the word ”automatic” implies the feature learning process, that is, completely independent from human- crafted features. More clearly, the concept of deep learning (i.e. neural networks with more than one hidden layer) can be considered as a wide class of machine learning approaches and architectures in which the main characteristic is hierarchically using many layers of nonlinear information processing. The aim of the techniques is learning feature hierarchies with higher level features of the hierarchy extracted from lower level features (Bengio, 2009). In fact, with automatically learned features at multiple levels of abstraction a system may execute complex functions to directly transfer the input to the output. The key factor of our model is the word representation. Typically, automatic text summarization systems use sparse input representations. Sparse representations can cause two problems for the model. First, not observing (enough) data in training 1Preprint of manuscript accepted 10 October 2016 Please refer to DOI: 10.10.16/j.eswa.2016.10.017 process. This problem is intensified when the selected vocabulary consists of only a subpart of the total presented words in the training data. Second, too much zero in the input and output of AE. To address these problems, we propose two techniques to reduce sparsity. First, we develop a local word representation in which each vocabulary is designed to construct the input representations of sentences in that document. Second, we add random noise to word representation vector, affecting both the inputs and outputs of the AE. In our case, after ranking the sentences of a document based on the cosine similarity, they must be selected to generate the summary. Picking the top ranked sentences up is a straightforward selection strategy that is used for AE networks trained without added noise. However, we propose to use an ensemble approach that aggregates the rankings of different experiments, each the result of adding randomly generated noise. Each application uses different random noise added to the input word representation, producing a possibly different extractive summary. The final summary is obtained by selecting the sentences that occur most frequently in the individual summaries. This ensemble approach can make the summarization process robust against small differences between training methods and analogous with manual summarization where annotator(s) would produce different summaries for a document in each review of the document. The specific contributions of the paper are as follow: We introduce an unsupervised approach for extractive summarization using AEs. Although AEs have previously applied for summarization as a word filter (Zhong et al., 2015), to the best of our knowledge we are the first to use the AE for summarization by sentence ranking. We will evaluate how AEs handle a sparse word representation such as tf-idf and a less sparse word representation based on a document-specific vocabulary, and also the impact of adding random noise to the local word representation vector. The addition of noise changes the AE from a deterministic feed-forward network to a stochastic model. To our best knowledge, this representation technique is not previously explored in the application of auto-encoders. We introduce the Ensemble Noisy Auto-encoder (ENAE) in which the model is trained once and used multiple times on the same input, each time with different added noise. We show adding stochastic noise to the input and running an ensemble on the same input can make improvements. We expand the email summarization features beyond a set of features and develop to automatically extracted features of emails. The next section describes recent related work. Section three is dedicated to our model in terms of topology and training algorithms, in particular, Restricted Boltzmann Machine (RBM) and AEs. The word representation is presented in section four. Section five presents sentence ranking measurement detail. The ENAE scheme is described in the six section. After providing experimental setup in the section seven, the results and discussion on the two different email datasets is presented in the section eight. The conclusion and future work is explained in the section nine. Bibliography is placed at the last section. II. RELATED WORK Machine learning techniques have widely used for text summarization (Li et al., 2009; Ouyang et al., 2011; Conroy and O’leary, 2001). They are trainable systems in which models learns how to generalize its parameters to extract salient points. Most of machine learning approaches in text summarization are inspired from information retrieval and adapted into the task such as Bayesian models, Hidden Markov Models (HMM), Support Vector Machines (SVM) and Support Vector Regression (SVR). In general, many supervised and unsupervised approaches have been proposed and they can be categorized into the following groups: Latent topic models as unsupervised techniques, classification and regression as the supervised techniques. Kaikhah (2004) successfully introduced a shallow neural network for automatic text summarization. Also, Svore et al. (2007) proposed a neural network-based system, called NetSum. This technique was inspired from (Burges et al., 2005). Shardan and Kulkarni (2010) proposed a combination of the multilayer perceptron (MLP) with fuzzy logic and they reported that this combination improves the result. In addition to feed-forward neural networks, recurrent neural networks (RNNs) have been applied for text summarization (Prasad et al., 2009). In spite of different topology, the approach was generally inspired from (Kaikhah, 2004). Zhong et al. (2015) used a deep architecture that is similar to an AE. They used the learned representations for filtering out not important words of a document in the early layer and discovering key words in later layer. The concept space is to extract the candidate sentences for the summary. However, in our model both the term and concept space are developed based on a sentence of a document. We use the learned representations directly to represent the semantic similar of sentences and in the ranking function. In supervised models, Denil et al. (2014) proposed a model based on a convolutional neural network (CNN) to extract candidate sentences to be included into the summary. A key contribution of the paper is that CNN is trained to classify sentiment labels that are either positive or negative (i.e. a binary label) and sentence extraction can affect the results of predicted a sentiment. Cao et al. (2015) used a recursive neural network for text summarization using hand-crafted word features as inputs. With the capability of dealing with a variable-length input in a RNN, the proposed system formulates the sentence ranking task in a hierarchical regression fashion. Not using any hand-crafted word representations and labeling data are significant in our proposed model. The proposed method of this paper in adding noise to the input can be analogous with the denoising auto-encoders (Vincent et al., 2008). To stop overfitting, denoising auto-encoders use noise in the input of AE. In a denoising AE, in order to prevent 2 Figure 1: Our AE topology for dimensionality reduction. The weights of decoder are obtained from unrolled weights of encoder. x and x̂ denote the input and reconstructed inputs respectively. hi are the hidden layers and wi are the weights. Concept space in which features/codes are extracted C(x) are in this scheme the output of the hidden layer h4. learning identity function, thereby capturing important information about the input in hidden layers, the input is corrupted and the network tries to undo the effect of this corruption. The intuition is that this rectification can occurs if the network can capture the dependences between the inputs. However, first, in most denoising AEs the input is corrupted using a random zero mask while in our model a small noise is added to the input. Second, in our model, the training algorithm adds very small noise to the input but the output is still the same as input. Third, the denoising AE adds noise to the inputs only during training, while we also add noise to input during test time. III. THE METHOD DESCRIPTION An AE (Figure 1) is a feed-forward network with the function of learning to reconstruct the input x (Hinton and Zemel, 1997). The AE is indeed trained to encode the input x using a set of recognition weights into a concept space C(x). Then, the features (a.k.a latent representations, codes) from concept space C(x) are converted into an approximate reconstruction of x̂ using a set of generative weights. In our model, the generative weights are obtained firstly from unrolled weights of encoder and then from fine-tune phase. Through the encoding process, (i.e. mapping to a hidden representation) the dimensionality of x is reduced to the give number of codes. The codes are then mapped to x̂ through the decoder. Both non-linear mappings are deterministic. A deep AE transforms the original input to a more discriminative representation in coding layer (a.k.a discriminative layer or bottleneck) over a hierarchical feature detection in which each level corresponds a different level of abstraction. Our model has two training stages: pre-training and fine-tune. Pre-training provides an appropriate starting point for the fine-tune phase. That is, obtained weights in pre-training phase will be the initial weights of the fine-tune phase. This initialization could significantly improve the performance of an AE in a wide variety of applications compared with random initialization (Hinton and Salakhutdinov, 2006a; Hinton et al., 2006b). AEs with pre-training have been successfully applied primarily to documents retrieval (Salakhutdinov and Hinton, 2009; Genest et al., 2011). We apply this technique to shorter texts. I. Pre-training Stage Assuming observations are produced from a stochastic generative model in which parameters control the model’s behavior (Hinton and Zemel, 1997; Hinton, 2010), Restricted Boltzmann Machine (RBM) as a generative model can be an appropriate 3 Figure 2: The schematic representation of the restricted Boltzmann machine (RBM) as a generative model. prototype to discover the parameters. The RBM (Figure 2) is the undirected graphical model corresponding a two-layer neural network with one layer of visible units (i.e. input data) and one layer of hidden units (i.e. feature detectors). The symmetric connections of units are restricted between hidden units and visible units, and no visible/hidden unit of the same layer is connected to any other visible/hidden unit. In our layer-wise pre-training, the first RBM is Gaussian-Bernoulli and the other RBMs, except for bottleneck hidden units, are Bernoulli-Bernoulli. In the last RBM, developing to have bottleneck, we replaced binary hidden units with a Noisy Rectified Linear Unit (NReLU) because of presented empirical successfulness in (Nair and Hinton, 2010). For both binary states of visible units and feature detectors (i.e. Bernoulli-Bernoulli units) and as an energy-based network, the energy function (Hopfield, 1982) is bilinear: E(x,h; θ) = − ∑ i∈visible bixi − ∑ j∈hidden ajhj − ∑ i,j xihjwij (1) Where wij is the weights between visible units xi and hidden units hj, bi and aj are their biases and θ = {W,a,b} represents the network parameters. In terms of the energy function, the joint distribution p(x,h; θ) has the following form: p(x,h; θ) = exp (−E(x,h; θ)) Z (2) Where Z = ∑ x,h exp (−E(x,h; θ)) is a partition function as a normalization constant. The marginal probability as- signed to a visible vector is the sum over all hidden units: p(x; θ) = ∑ h exp (−E(x,h; θ)) Z (3) The symmetric structure of the network leads the conditional probabilities for Bernoulli-Bernoulli RBM to be: p(hj = 1|x; θ) = exp ( ∑ i wijxi + aj) 1 + exp ( ∑ i wijxi + aj) = f( ∑ i wijxi + aj) (4) p(xi = 1|h; θ) = exp ( ∑ j wijhj + bi) 1 + exp ( ∑ j wijhj + bi) = f( ∑ j wijhj + bi) (5) For visible units with real values, the energy function is: E(x,h; θ) = ∑ i∈visible (xi − bi)2 2σ2i − ∑ j∈hidden ajhj − ∑ i,j xi σi hjwij, (6) Where σi is the standard deviation of the ith visible unit. For Gaussian-Bernoulli conditional probabilities becomes: p(hj = 1|x; θ) = exp ( ∑ i wijxi + aj) 1 + exp ( ∑ i wijxi + aj) = f( ∑ i wijxi + aj) (7) p(xi|h; θ) = 1 √ 2π exp ( − (x− bi − ∑ j wijhj) 2 2 ) = N  ∑ j wijhj + bi, 1   (8) Maximum likelihood estimation can be applied to estimate the parameters of the model. To do so, taking the derivative of the negative log-probability of the inputs with respect to the parameters is: ∂ − log p(x) ∂θij = ∂ ∂θ (− log ∑ h p(x,h)) (9) 4 Figure 3: Multiple RBMs stacked on top of each other. ∂ − log p(x) ∂θij = E[ ∂E(x,h) ∂θ |x] −E[ ∂E(x,h) ∂θ ] (10) ∂ − log p(x) ∂θij = 〈xi,hj〉data −〈xi,hj〉recon (11) Where the angle brackets are used to denote the expectation of the distribution of the subscript that follows. This leads to a simple learning algorithm by which the parameters update rule is given: ∆wij = �(〈xi,hj〉data −〈xi,hj〉recon) (12) Where � is a learning rate, 〈xi,hj〉data is the so-called positive phase contribution and 〈xi,hj〉recon is the so-called negative phase contribution. The positive phase decreases the energy of observation and negative phase increases the model’s energy. However, the computation of the expectation defined by the model is not easily tractable. Hinton (2002) presented maximizing the likelihood or log-likelihood of the data is equivalent to minimize Kullback–Leibler (KL) divergence between data distribution and the equilibrium distribution over the visible variables. In practice, the k-step contrastive divergence (CD-k) approximation provides surprising results (Hinton, 2002). Carreira-Perpinan and Hinton (2005) numerically compared the update rule with exact log-likelihood gradient with CD-k (Hinton, 2002). Based on this comparison and to have a very good approximation we used CD-1, with running one step Gibbs sampler that is: x = x0 p(h|x0) −−−−−→ h0 p(x|h0) −−−−−→ x1 p(h|x1) −−−−−→ h1 (13) The above mentioned algorithm can be used to have more than one RBMs blocks. To draw a conclusion for the pre-training process, a greedy layer-wise fashion using individual RBMs is employed in which the output of one training RBM is used as an input to the next upper RBM block (Figure 3). In terms of topology, individual RBM blocks would be stacked on top of each other and having the RBM blocks stacked, the topology of the AE with desire deepness meets. II. Global Adjustment After extracting the weights of stacked RBMs from the greedy layer-wise unsupervised algorithm (i.e. pre-training), the weights are used as an initialization point of AE with a corresponding configuration. To globally adjust parameters and develop the concept space of our AE, first the stacked RBMs are unrolled (i.e. tied weights). This ensures that the encoder is not operating in the linear regime of its non-linearity (Vincent et al., 2010). Then a fine-tune phase using back-propagation is applied. The fine-tune phase is for the whole network (i.e. all the layers together) and uses a gradient-based optimization algorithm. This approach of using RBM modelling for pre-training followed by back-propagation algorithm for fine-tune is performing two different search methods in the parameter space of the AE. The greedy pre-training learning algorithm performs a search that extracts a good region in the parameter space. In information retrieval applications, it has been observed that deep generative models with only a pre-training phase are comparable with a model that has only one hidden layer (Hinton and Salakhutdinov, 2011). Applying back-propagation from a point in this good region by using a gradient-based local search to fine-tune the model parameter helps take advantage of the multiple hidden layers. 5 In this unsupervised fine-tuning phase for optimal reconstruction, given the encoding C(x), the whole network tries to minimize the cross-entropy error as the loss function: − log p(x|c(x)) = − 1 N N∑ i=1 xi log x̂i − N∑ i=1 (1 −xi) log(1 − x̂i) (14) Where n is the total number of items of training data, x is the input of the AE, x̂ = fθ(c(x)) is the reconstructed values. Although stochastic gradient descent (SGD) is a common methodology to fine-tune a neural network, Ngiam et al. (2011) observed that mini-batch Conjugate Gradient (CG) with line search can simplify and speed up training of AEs compared to SGD; thus, we used mini-batch CG with line search and the Polak-Ribiere rule to compute search directions. IV. WORD REPRESENTATION A word representation is a type of math object regarding each word. We used a bag-of-words (BOW) representation. This representation does not contain task-specific features and provides a fixed size vector for documents of different sizes. In the BOW representation, a sentence or a document is presented as the collection of its words ignoring the exact word ordering and only retaining the number of occurrences of each word. Intuitively, sentences or documents with similar BOW representations are likely to be similar in content. Although BOW does not embed any semantic information about the words, a neural network with this representation as the input can learn a distributed semantic representation in which semantically related words are relatively close to each other in the generated vector space (Turian et al., 2010). Term frequency - inverse document frequency (tf-idf ) is the most commonly used BOW representation for information retrieval and text summarization. It represents each word as an element of a vector comprising of its term frequency in the document, as well as over all documents (idf ) (Wu et al., 2008). In our model, the tf-idf representation is constructed for each sentence as a basis for sentence ranking. Therefore, the input vectors are very sparse since each sentence only contains a small number of words. Sparse representations typically can cause two problems for the leanring algorithm. First, some words may only appear in the test set and never be seen in training. Second, extensive zeros in the output of the AE has an effect on the derivative; i.e. small errors on the zeroes can dominate larger errors on the non-zero outputs when back-propagation modifies the hidden nodes. Genest et al. (2011) and Glorot et al. (2011) presented a re-sampling algorithm to solve this problem for AE. However, to produce a less sparse representation, we propose using a local vocabulary to construct a locally normalised term-frequency (tf ) representation. In this representation, the vocabulary for each document is separately constructed from the most frequent terms occurring in that document. The vocabulary size is the same for all documents. This local representation logically has to be less sparse compared to tf-idf in which the vocabulary is built from all the documents because the input dimensions correspond only to terms occurring in the current document. A side effect is that the AE input dimensions now correspond to different words in different documents and the encoder maps the queries and sentences of different documents into different semantic subspaces. However, this causes no adverse effects on the task because our model ranks sentences in each document against that document’s individual query. Also, we added small random values (Gaussian or Uniform distribution) to the tf representation to eliminate zeroes in the input. The idea is that when the noise is small, the information in the noisy inputs is essentially the same as in the input vectors without noise, however, the noisy inputs do not contain absolute zeros any more. V. SENTENCE RANKING Extractive text summarization is also known as sentence ranking. Our model ranks the sentences based on the most relevant and salient characteristics of the documents that are semantically similar to either the subject of an email or previously selected key phrases of a text; thus, this ranking process is an example of query-oriented summarization. The email subject or key phrase is the query text. Having the AE globally adjusted, the output of the encoder provides the concept space where the codes capture the concepts of a sentence. Because the term space was based on the content of a document and its query, the encoding function maps each sentence into an abstract concept. In the concept space, we use cosine similarity as a common metric. More clearly, the AE learns a low-dimensional concept space. In this space, cQ = C(x|Q), cs = C(x|S1, . . . ,sn) where C(x) is the mapping, Q is the query and S is the sentences. The relevance score, driven from concept codes, between the query cQ and the sentences cS is calculated by cosine similarity as following: R(Q,D) = cos(cQ,cS) = cQ.cS ||cQ|| ||cS|| (15) Sentences can be ranked according to the relevance score. 6 Figure 4: The Ensemble Noisy Auto-Encoder for query-oriented summarization. VI. ITERATIVE DATA FUSION AND VOTING Ensemble methods use multiple models, mostly classifiers, that are combined to solve a particular problem. In particular, in classifier fusion algorithms, all classifiers are trained in parallel over the entire feature space (Hansen and Salamon, 1990). However, in particular, Bucilua et al. (2006) have presented a neural network model by which a large and complex ensemble can be compressed into a smaller and faster model. Hinton et al. (2015) developed this model compressing method further using a different compressing technique and showed that an ensemble of deep neural networks can be distilled into a single deep network while the quality of trained ensembles is retained. Although our model is a single deep network, the proposed technique is not distilling an ensemble of models into a single model but developing a single model from scratch using an ensemble of runs on the same network. Figure 4 presents the proposed model. It is an integrated system in which semantically similar sentences to a query can be ranked. Although after training, an AE applies a deterministic non-linear transformation to compute a new feature space for the data, the model changes AEs from a deterministic feed-forward network to a stochastic model. In detail, after ranking the sentences of a document based on the cosine distance to the given query, a number of them must be selected to be included into the summary. A straightforward selection strategy is just to select the top ranked sentences. We use this selection strategy for the AE without additive noise to the input. However, with the noisy input representations introduced in the word representation section, an experiment can be repeated using the same input vector but with differently generated noise several times. This iterative ranking (i.e. each experiment) provides a variations in ranking. Then, an ensemble approach that aggregates the rankings of different experiments on the same input document is used to fuse the ranking. Specifically, we use a majority voting scheme to fuse the ranks. Since the raw rankings have been obtained using additive noise, the voting algorithm ignores the rank and only considers the number of times that each sentence appears in the top group. Thus, in the final sentence selection phase we have a set of items D (the total number of sentences produced by t experiments) and the algorithm picks a subset Sn ⊂ D (Where n is the number of the sentences of the final summary). In a greedy approach ŝk is picked given: Sk−1 = {ŝ1, ..., ŝk−1}. Where Sk = Sk−1 ∪{ŝk} with the following formula: ŝk = arg maxsk∈D|Sk−1 [score(sk)] (where score(sk) is the number of times sk is selected by each component). VII. EXPERIMENTAL SETUP To have a extensive exploration of the model, we conducted different experiments on the two different publicly available email datasets that are specifically developed for summarization: Summarization and Keyword Extraction from Emails (SKE) (Loza et al., 2014), BC3 from British Columbia University (Ulrich et al., 2008). SKE consists of 349 emails, of which 319 are from the Enron mailboxes and 30 were provided by volunteers. The email sets contain both private and corporate emails and are divided into two parts: single emails of at least 10 sentences and threads containing at least 3 emails. In total, SKE consists of more than 100,000 words. Stemming using Porter (1980) and stop-word removal 1 reduces this to 46,603 words comprising 7478 unique words. There are 6801 sentences — an average of 19.5 sentences per email. The average email is 303 words and the average sentence is 15.5 words. Two different annotators provided an abstractive summary, an extractive summary and a set of key phrases for each email. For extractive summaries, the annotators identified 5 different sentences in a ranking fashion; therefore, in our experiments with this dataset, generated summaries have 5 sentences. The abstractive summaries are between 33 and 96 words long, averaging 73 words per email. Key phrases were restricted to 4 words, averaging 2.1 words per phrase. We conducted two different query-oriented summarization experiments on the SKE corpus. The first experiment used the email subject as the query text — this experiment could only be performed on the 289 emails that have non-empty subjects. 1obtained from http://xpo6.com/list-of-english-stop-words 7 http://xpo6.com/list-of-english-stop-words The second experiment used keyword phrases, which are available for the whole dataset, as the queries. Subject: Nissho NDA 01 Kobayashi-san, Sorry for the delay. 02 I am working through a few issues here and I expect to get back to you shortly. 03 Diamond-san, As I wrote in the past, Nissho Iwai’s LNG related department has been transferred into a new joint venture company between Nissho and Sumitomo Corp. as of October 1, 2001, namely, ”LNG Japan Corp.”. 04 In this connection, we would like to conclude another NDA with LNG Japan Corp, as per attached. 05 Please review it and it would be great if you could make original again and send us for Mr.Miyazawa’s signature. 06 Also, please advise us how we should treat Nissho’s NDA in these circumstances. 07 BTW, I talked with Mr.Matsubara, Manager of Enron Global Finance in Tokyo. 08 We are internally discussing when we start our official meeting. 09 Please let us know when you are ready. 10 Please approve or make changes to their new NDA. 11 They need to change the counterparty name due to a joint venture. 12 I wanted to let you know this was coming in as soon as Mark approves the changes. 13 Dealing with Japan is somewhat time challenging. Subject: Next face to face meeting 01 Dear all, Hoping that we will have a last call review shortly (see my next email) we are planning a face to face meeting at which we can resolve issues raised, and generally tidy the document up (write up lots more good techniques). 02 At the moment we are talking to a potential host in the Boston area about a meeting on thursday and friday 7-8 October. 03 More details as they come to hand. 04 Folks, Please note that this meeting would be on at the same time as the ATIA meeting in Orlando, Florida. 05 If the time is impossible for anyone please let us know as soon as possible at the moment the next closest possible times are a week earlier (which will cut into the time available for last call review) or ten days later (which will possibly push our schedule back further). 06 Hello, I have other meetings tentatively scheduled for 6 and 7 October in Boston. 07 I would probably be able to attend starting the 7th. 08 I believe that I will be in Rome on the 6th and 7th. 09 Dear working group participants, As discussed, the meeting will indeed be held on thursday 7 and friday 8 October. 10 I am pleased to announce that Allaire Corp will be our hosts, and look forward to a productive two days, and some time to say hi. The nitty-gritty. 11 Meeting registration is open from now until 1 October. 12 register via the meeting page at http://www.w3.org/WAI/AU/f2f-oct99 which also provides details on location, accommodation, etc. 13 There are a limited number of rooms available in the Royal Sonesta Hotel (which is the meeting venue) at a discounted rate in order to qualify for this rate reservations and meeting registration must be completed by September 10. 14 Since October is the peak season for Boston hotels, rooms are expensive and hotels book out early, so you are advised to make your plans and reservations as soon as possible. 15 If you have any questions or special requirements please contact me as early as possible. 16 Most important if you haven’t registered yet, don’t forget. 17 http://www.w3.org/WAI/AU/f2f-oct99 for the details. 18 Given that there are a small number of us, I thought I would ask what people want to eat for lunch. 19 The default choices are roast turkey breast with vegetables (without turkey if you are a vegan), and pasta with vegetable rataouille. 20 Possible substitutions: baked Scrod (fish) chicken caesar salad hommus roll-up sandwich with salad salad with crab cakes churrasco flank steak with chimichurri and onions. 21 I will do this in the following way: If you cannot eat one of these things please make that very clear. 22 It will take at least two votes to change something, and will have to be compatible with what people can’t do. 23 The deadline is 3pm (boston time) Monday. 24 No response will not be considered a vote for the default it will be considered a vote for ”whatever gets chosen” Sorry to bore you with administrivia, but I figure that having food that people like helps the meeting go well. Figure 5: A randomly selected sample email thread from SKE, corporate section (ECT020.xml) and from BC3 (list number: 058-13714734). The subject of SKE email is ”Nissho NDA”. The subject of BC3 email is ”Next face to face meeting”. The BC3 data was collected from mailing lists that are mostly not as informal as ordinary emails. The dataset contains 40 email threads each of which has less than 11 emails and less than 6 participants, on average. The dataset contains 3222 sentences including repetitions due to quoting. Removing repetition, signatures and sender’s name leaves 736 sentences, an average of 18.4 sentence per email thread. Stemming and stop-word removal further reduce the data set from 15,000 to 7101 words with 2306 unique words. The average number of words per email thread is 380 and per sentence is 21. BC3 has been annotated by 3 different annotators for both extractive and abstractive summarization. However, the ex- tractive summaries are not restricted to a limited number of sentences and do not involve sentence ranking so they are not suitable as goals for our approach. The shortest extractive summaries were 4 sentences, so our experiments produce 4 sen- tences as extractive summaries for this dataset. These summaries are compared with the annotators’ abstractive summaries using ROUGE-2 as explained below. The abstractive summaries were limited to 250 words and are on average 122 words for each email thread. All 40 email threads of BC3 have an informative subject which we use as the query text for summarization. 8 The average subject length is 5.2 words. Our baseline is tf-idf (with the inverse document frequency as the idf term) and over experiments we use several different vocabulary sizes. tf-idf is a term based technique and a case study helps better undestanding of our model regarding meaning extraction. Figure 5 shows randomly selected emails from SKE and BC3 for the case study. Because there is no general agreement on a vocabulary size in the literature we decided to choose vocabulary sizes by which two datasets can be compa- rable and also providing different test settings to evaluate the proposed model. The smallest vocabulary, 60 words, is selected because Genest et al. (2011) used less than 1.5% of the unique words of a large dataset like Text Analysis Conference (TAC) corpus for the input of an AE and we want to provide at least 2% of the unique words. The largest vocabulary of 1000 words was selected because assuming words with frequency 1 as idiosyncratic words, in BC3, only 1000 unique words have the frequencies of at least 2. In short, we constructed vocabularies based on the 1000, 374/120 (about 5% of the whole vocabulary of SKE/BC3), 150 (about 2% of the whole vocabulary of SKE), and 60 most frequently occurring terms (around 2% of the whole vocabulary of BC3). The selection of a metric for comparison of the generated summary with the human summary is still an open question (Louis and Nenkova, 2013). We used ROUGE (Lin, 2004) package as a fully automatic metric for evaluating the performance of our model. The pyramid metric is an indirect manual evaluation method (Nenkova et al., 2007) that has previously been used for BC3 (Ulrich et al., 2009). However, due to lack of resources we used the fully automatic metric. Of the different forms of ROUGE, we use ROUGE-2 Recall because Dang and Owczarzak (2008) observed that it has the highest correlation with manual scores. In all our experiments, the confidence interval was 95% and a jackknifing procedure was used for multi- annotation evaluation Lin (2004). We used 10-fold cross-validation with 90% training data. The AE has 140 units in the first hidden layer, then 40 units, 30 units and 10 units as the bottleneck (i.e. codes). The decoder portion uses the reverse pattern. We used the same architecture for all the different vocabulary sizes. We used G Hinton’s training algorithms 2. The size of the first hidden layer (140 units) was chosen to be at least double the smallest vocabulary size of 60 inputs, because this may help RBM to obtain a better representation (Bengio, 2009; Erhan et al., 2009; Le Roux and Bengio, 2008). We chose to include two additional hidden layers with 40 and 30 units because (Hinton and Salakhutdinov, 2006a) showed in the supplementary material paper 3 that a pre-trained deep AE can have better performance than a pre-trained shallow AE with the same number of parameters. The ten code units provide a ten-dimensional concept space to represent the semantics of each sentence. Theoretically, learning a desired concept in a large concept space can be as difficult as finding a needle in the haystack and experimental results show having more units may negatively impact on the AE for summarization (Zhong et al., 2015). We used the same hidden units structure throughout our experiments for two reasons. First, this makes the results more comparable. Second, an AE with more parameters to be tuned would require more manually annotated data that was not available to us. Also, we used the same hyper-parameters, obtained on the basis of presented practical guides in (Bengio, 2012; Hinton, 2010), over the all experiments. We explored several input representations: tf-idf, tf using local vocabularies (Ltf ), tf-idf and Ltf as the input of AE, Ltf with added Gaussian/Uniform noise (Ltf-NAE) and in the noisy ensemble Ltf (Ltf-ENAE) is used for the input of the AE. Both additive Gaussian and Uniform noise have the same variance over the all experiments. VIII. RESULTS AND DISCUSSION I. Subject-oriented Summarization with SKE The ROUGE-2 recall of the tf-idf baselines and proposed models is presented in table 1. The table presents the performance of models in extracted summaries with different length. This is hardly surprising that recall improves with more longer summaries and larger vocabulary size for tf-idf. This improvement is relatively constant for our baselines. Using tf-idf as the input of the AE, the recall is improved by increasing the number of sentences in the summary. However, increasing the vocabulary size provides less effective AE representation and reduces the recall, possibly because the larger vocabulary increases the proportion of zero inputs. Comparing tf-idf, Ltf, AE (tf-idf ) and AE (Ltf ) shows three important points. First, AE (Ltf ) is always much better than Ltf. The relative difference is greater for longer summaries, demonstrating that AE enriches the representation and provides a more discriminative space. Second, AE enhances the discriminative power of both tf-idf and Ltf, but the recall of AE (Ltf ) is higher. As above, Ltf provides less sparse inputs than tf-idf, and the AE learns more effectively. Finally we note that AE (Ltf ) outperforms the tf-idf baselines even when tf-idf uses much larger vocabulary sizes. The table shows that in general, the ROUGE score of AE (Ltf ), Ltf-NAE and Ltf-ENAE, both perturbed with either Gaussian or Uniform noise, is itself noisy (i.e. having no deterministic order) for summaries with only one sentence. However, 2http://www.cs.toronto.edu/∼hinton/MatlabForSciencePaper.html 3http://www.cs.toronto.edu/∼hinton/absps/science som.pdf 9 http://www.cs.toronto.edu/~hinton/MatlabForSciencePaper.html http://www.cs.toronto.edu/~hinton/absps/science_som.pdf Model n = 1 n = 2 n = 3 n = 4 n = 5 tf-idf V=1000 0.1936 0.2002 0.2062 0.2217 0.2312 tf-idf V=5% (374 words) 0.1453 0.1473 0.1509 0.1688 0.1838 tf-idf V=2% (150 words) 0.0985 0.1048 0.1074 0.1291 0.1435 tf-idf V=60 0.0859 0.0860 0.0823 0.0935 0.1068 AE (tf-idf V=2%, 150 words) 0.1277 0.1951 0.2572 0.3227 0.3580 AE (tf-idf V=60) 0.1805 0.2310 0.2871 0.3422 0.3913 Ltf V=60 0.1165 0.1682 0.2452 0.3032 0.3349 AE (Ltf V=60) 0.1205 0.2272 0.3360 0.4251 0.4948 Ltf-NAE (Gaussian) 0.1067 0.2219 0.3289 0.4033 0.4664 Ltf-ENAE (Gaussian) 0.1370 0.2471 0.3510 0.4325 0.5031 Ltf-NAE (Uniform) 0.0841 0.1583 0.2618 0.3592 0.4428 Ltf-ENAE (Uniform) 0.1140 0.1843 0.2711 0.3659 0.4377 Table 1: ROUGE-2 recall of subject-oriented summarization for tf-idf, tf and models versus the number of selected sentences. n is the number of sentences for a summary. with the increase of the number of extracted sentences, the scores of all variants of AEs are relatively closer to each other, in particular for summaries with 5 sentences. Among AE-based techniques, more relevant sentences can be extracted using Ltf-ENAE (Gaussian) and AE (Ltf ) compared to Ltf-NAE (with either Gaussian or Uniform) and Ltf-ENAE with voted from Uniform noise runs. In extracted summaries with one sentence, the highest recall belongs to tf-idf whereas for summaries with more sentences AE-based models preform much better. This can be interpreted that in subject-based summarization mapping term space to concept space is not quite accurate for every single vector, but the mapping is much accurate on average (i.e. the summaries with more sentences). To further evaluate our model, we analyse the results of Ltf-NAE and Ltf-ENAE generated from the sample email thread (Figure 5). The first annotator selected and ranked sentences 3, 4, 10, 11 and 5 while the second annotator selected and ranked sentences 3, 4, 11, 6 and 5. For comparison, Ltf-NAE and Ltf-ENAE ranked sentences number 4, 10, 3, 6, 12 and 3, 8, 4, 6, 11, respectively, as the top 5 sentences. Both generated summaries contain sentences 03, 04 with high rank similar to annotators. Sentence 11 is selected by both annotators at a middle rank and is selected by Ltf-ENAE but not by Ltf-NAE in this particular example. Sentences 10 and 6 are each selected by only one annotator and are ranked lower by both algorithms. The only annotator selected sentence that is not ranked by the algorithms is sentence 5, which was ranked last by both annotators. II. Key-phrase-oriented Summarization with SKE Key phrase extraction has been used for text summarization (Qazvinian et al., 2010; Zhao et al., 2011). Where the annotators have provided key phrases, we can use them as the query. Table 2 shows the ROUGE-2 recall of the baselines and AE-based models with various input representations. Since the summarization uses carefully extracted key phrases, recall increased as expected. Although this improvement is for all techniques, the amount of improvement varies for each technique. AE (tf-idf ) provides better performance than plain tf-idf with the same size vocabulary. Considering n=5 which has the highest performance, tf-idf with vocabulary sizes 150 and 60 yields recalls of 0.3166 and 0.2224 respectively which are improved by AE (tf-idf ) to 0.4795 and 0.4220. It is worth noting that the larger vocabulary size here yields better recall, in contrast with the subject-oriented summarization results (Table 1). We hypothesize that the enriched term space of the key phrase queries enables AE, as an unsupervised model, to learn a richer concept space, yielding richer summaries. The AE (Ltf ) performance is better than Ltf. There is a fairly important difference between subject-oriented and key- phrase-oriented summarization for SKE corpus when n=1. The ROUGE score of the AE with various Ltf -based input rep- resentations is higher than tf-idf and AE (tf-idf ). Ltf uses a local vocabulary for each e-mail, meaning that there is a greater likelihood of the words in the key phrases being present in the vocabulary. Words that are not in the vocabulary cannot contribute to the concept space, so Ltf supports a richer concept space for small vocabulary sizes. Table 2 shows that AE (Ltf ) and ENAE (Uniform) provide better results in comparison with other techniques for values of n between 3 and 5. Comparing the performance of ENAEs with the corresponding NAEs, using either subject or key phrase, shows a small improvement in most cases. We hypothesize that the ensemble regularises the concept space, producing a more effective concept mapping irrespective of the information content. Our case study (Figure 5) also shows that summarization using key phrases with Ltf-NAE and Ltf-ENAE provide better results compared with their subject-oriented counterparts. Ltf-NAE provides a good ranking in which 3 sentences are ranked 10 Model n = 1 n = 2 n = 3 n = 4 n = 5 tf-idf V=1000 0.2217 0.3432 0.4059 0.4607 0.4845 tf-idf V=5% (374 words) 0.2105 0.2874 0.3437 0.4118 0.4217 tf-idf V=2% (150 words) 0.1482 0.2176 0.2655 0.3060 0.3166 tf-idf V=60 0.1215 0.1661 0.1988 0.2103 0.2224 AE (tf-idf V=2%, 150 words) 0.1321 0.2251 0.3179 0.4077 0.4795 AE (tf-idf V=60) 0.1291 0.2226 0.2879 0.3678 0.4220 Ltf V=60 0.2034 0.3044 0.3708 0.4395 0.4972 AE (Ltf V=60) 0.2397 0.3410 0.4365 0.5035 0.5657 Ltf-NAE (Gaussian) 0.2200 0.3107 0.3887 0.4581 0.5179 Ltf-ENAE (Gaussian) 0.1870 0.3168 0.4017 0.4809 0.5370 Ltf-NAE (Uniform) 0.2547 0.3416 0.4277 0.4872 0.5380 Ltf-ENAE (Uniform) 0.2179 0.3450 0.4334 0.4938 0.5504 Table 2: ROUGE-2 recall of key-phrase-oriented summarization for tf-idf, tf and models versus the number of selected sentences. n is the number of sentences for a summary. in the top by both annotators and the other 2 by one annotator. Ltf-ENAE also has 3 sentences in common with both annotators while one other sentence has also been selected by one of annotators. The only sentence that is unranked by the annotators is sentence number 12 — we will see in the next subsection that this sentence is also informative. III. Key-phrase-oriented Summarization Versus Abstractive Summary with SKE SKE also provides annotators’ abstractive summaries. These can be used as a ’gold standard’ to compare our model’s sum- maries with human abstracts. Table 3 present the ROUGE-2 scores of the extracted sentences versus abstractive summaries. It is to be expected that the recall will decrease to small values because the abstractive summaries are written by humans and only contain the abstract meaning of the emails, not the exact sentences. Despite this, the trend is very similar to the previous subsection. This experiment reinforces that AE can provide a more discriminative concept space and that Ltf-ENAE (Gaussian/Uniform) mostly yield better results compared other baselines. Model Recall tf-idf V=1000 0.2120 tf-idf V=5% (374 words) 0.1848 tf-idf V=2% (150 words) 0.1424 tf-idf V=60 0.0942 AE (tf-idf V=2%, 150) 0.2115 AE (tf-idf V=60) 0.1875 Ltf V=60 0.2137 AE (Ltf V=60) 0.2319 Ltf-NAE (Gaussian) 0.2110 Ltf-ENAE (Gaussian) 0.2255 Ltf-NAE (Uniform) 0.2210 Ltf-ENAE (Uniform) 0.2299 Table 3: ROUGE-2 recall for tf-idf, tf and models, abstractive summaries generated by human versus extractive summaries 11 First Annotator - LNG Japan Corp. is a new joint venture between Nissho and Sumitomo Corp. Given this situation a new NDA is needed and sent for signature to Daniel Diamond. Daniel forward the NDA to Mark for revision. Second Annotator - An Enron employee is informed by an employee of Nissho Iwai that the Nissho Iwai’s LNG related department has been transferred into a new joint venture company, namely, ’LNG Japan Corp.’. As a result, there is a need to change the counterparty name in the new NDA. The new change has to be approved and then applied to the new NDA with LNG Japan Corporation. Figure 6: the generated abstractive summaries of sample email (Figure 5) by two annotators. Figure 6 presents the generated abstractive summaries by annotators. Taking the sentence number 12 into consideration, although the sentence is not among extracted sentences by Ltf-ENAE, it contains important information that both annotators have pointed out. IV. Subject-oriented Summarization with BC3 As discussed previously, only the abstractive summaries of BC3 are appropriate for our study. Table 4 presents the ROUGE-2 recall of extracted summaries compared with the three annotators’ abstractive summaries; thus, it is expected that recall scores are small values. While each extracted summary contains 4 sentences, for completeness we also evaluate summaries of 1, 2 and 3 sentences. In table 4, AE (tf-idf ) performs worse than the other approaches, in contrast with the previous results in tables 1, 2 and 3. This may come from the relative difference between SKE and BC3. More clearly, the number of emails in BC3 is about 12% of SKE while the number of unique words in BC3 is about 30% of SKE. This results in a more sparse tf-idf representation in BC3 than SKE. Similar to the subject-oriented summarisation of SKE, the performance of AE (tf-idf ) is better with the smaller vocabulary (n=2%). For small vocabularies, the best results are obtained using variants of AE with Ltf. As in the above subsections, this performance improvement is related to the local vocabulary providing a richer representation of both the query and text. For the case study presented in figure 5, we again consider the output of Ltf-NAE and Ltf-ENAE. Although in this randomly selected email, there is no intersection between the top ranked sentences of Ltf-NAE and Ltf-ENAE, both models do extract important information from this long email thread. Ltf-NAE extracted information such as possible menu substitutions (sentence 20), flexibility of schedule changes (sentence 5) and avoiding certain foods (sentence 19); the non-informative sentence 15 was also included in this extract. Ltf-ENAE selected very important information, that is highly correlated with annotator’s summaries, such as the hosting organisation and duration (sentence 10), place and time of meeting (sentence 2), some people may not attend and lunch planning (sentence 18) and a competing meeting (sentence 4). Model n = 1 n = 2 n = 3 n = 4 tf-idf V=1000 0.0380 0.0636 0.0856 0.1002 tf-idf V=5% 0.0359 0.0517 0.0612 0.0711 tf-idf V=2% 0.0294 0.0420 0.0528 0.0583 AE (tf-idf V=5%) 0.0145 0.0300 0.0429 0.0594 AE (tf-idf V=2%) 0.0166 0.0315 0.0522 0.0697 Ltf V=60 0.0317 0.0566 0.0758 0.0967 AE (Ltf V=60) 0.0362 0.0659 0.0853 0.1084 Ltf-NAE (Gaussian) 0.0406 0.0650 0.0825 0.1001 Ltf-ENAE (Gaussian) 0.0264 0.0606 0.0839 0.1017 Ltf-NAE (Uniform) 0.0354 0.0566 0.0832 0.1010 Ltf-ENAE (Uniform) 0.0334 0.0581 0.0827 0.1028 Table 4: ROUGE-2 recall of subject-oriented summarization for tf-idf, Ltf and models versus the number of selected sentences. n is the number of sentences for a summary. 12 V. Generalization and Data Comparison The generalization capability for deep structures with millions of parameters is quite crucial. Poor generalization arises from over-fitting of the training data, where the error on the training set is significantly small while this error is still large for test set (i.e. new data). To address this problem, AEs employ training algorithms such as a sparsity constraint, contractive mapping, dropout, de-noising, early stopping, L2 regularization (i.e. weight-decay) and unsupervised pre-training (Srivastava et al., 2014; Rifai et al., 2011; Vincent et al., 2008; Erhan et al., 2010). While a gradient-based algorithm gradually increases weights, corresponding to a regularized learning process (Collobert and Bengio, 2004), early stopping alone is not sufficient enough to prevent over-fitting. Our experiments use a combination of early stopping, L2 regularisation and unsupervised pre- training. Figure 7 demonstrates that both the training and test errors are decreasing throughout training, and that the trained AE has good generalisation. The two reasons for adding small random noise to the input representation were to eliminate zeroes and to support ensem- bles. The results show that eliminating zeroes alone is not beneficial — the performance of AE (Ltf ) is typically better than Ltf-NAE. The comparison is based on the SKE data because it is larger and more diverse. Figure 8 illustrates the improvement of summarization provided by AE(Ltf ), particularly for summaries of 5 sentences. Comparing AE(Ltf ) with tf-idf in this fig- ure, we note that AE(Ltf ) improves linearly with increasing the number of sentences in a summary, whereas tf-idf gains less from larger summaries. To evaluate ensembles, we compare Ltf-NAE with Ltf-ENAE. Figure 9 shows that the Ltf-ENAE technique usually out- performs Ltf-NAE. Ltf-ENAE scheme may even improve on AE(Ltf ). The results particularly confirm the value of AEs. The AE-based models generally perform much better than the tf-idf baselines in both the SKE and BC3 data sets — only tf-idf with the vocabulary of size of 1000 can compete with the AE models. This demonstrates that the AE models are more capable of encoding the information content of a small vocabulary. Figure 7: the average squared reconstruction error of the train and test set versus the number of epochs (iterations) for AE(Ltf) (a) and Ltf-NAE (Uniform) (b) for the SKE corpus. Figure 8: the ROUGE-2 recall of the best baseline td-idf, tf, the AE (tf-idf) and AE(Ltf) using keyword phrase as query. 13 Figure 9: ROUGE-2 recall using the keyword phrases as queries. (a) shows the results based on Gaussian noise. (b) presents the results based on Uniform noise. IX. CONCLUSION AND FUTURE WORK In this paper, we presented a query-based single-document summarization scheme using an unsupervised deep neural network. We used the deep auto-encoder (AE) to learn features rather than manually engineering them. A series of experiments were conducted on the SKE and BC3 email datasets, using the subject and keyword phrases as queries. We compared using global and local vocabularies to construct word representations as the input of the AE. Comparing the results of local term frequency (Ltf ) with and without AE demonstrates that the AE provides a more discriminative feature space in which semantically similar sentences and queries are closer to each other. More precisely, AE improves the ROUGE-2 recall of Ltf by on average 11.2%. Also, AE (Ltf ) in most cases significantly outperforms the tf-idf state-of-the-art term matching baselines. To investigate the semantics relevance of the extracted summary, we examined a randomly selected case study from each data set and found that the extracted sentences are mostly highly informative. We also presented noisy AE approaches, where random noise is added to the input and output of the AE during training and testing. Using a single AE with noisy input generally reduces test performance, but using an ensemble of 5 noisy AEs restores the recall performance and is generally competitive with AE Ltf. Further, the case study shows that the ensemble can extract a summary that is more closely aligned to human summaries. In future, we intend to extend our proposed model to generic summarization by clustering the AE features. A semi- supervised learning technique may be useful because of the limited availability of manually annotated data, or noisy labelling could be used with the unlabelled data. Other ensemble approaches should also be investigated to improve performance including accuracy. REFERENCES Yoshua Bengio. Learning deep architectures for ai. Foundations and trends R© in Machine Learning, 2(1):1–127, 2009. ISSN 1935-8237. Yoshua Bengio. Practical recommendations for gradient-based training of deep architectures. In Neural Networks: Tricks of the Trade, pages 437–478. Springer, 2012. Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535–541. ACM, 2006. Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning, pages 89–96. ACM, 2005. Ziqiang Cao, Furu Wei, Li Dong, Sujian Li, and Ming Zhou. Ranking with recursive neural networks and its application to multi-document summarization. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015. Miguel A Carreira-Perpinan and Geoffrey E Hinton. On contrastive divergence learning. In Proceedings of the tenth international workshop on artificial intelligence and statistics, pages 33–40. Citeseer, 2005. Ronan Collobert and Samy Bengio. Links between perceptrons, mlps and svms. In Proceedings of the twenty-first international conference on Machine learning, page 23. ACM, 2004. Ronan Collobert, Jason Weston, Léon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. Natural language processing (almost) from scratch. The Journal of Machine Learning Research, 12:2493–2537, 2011. ISSN 1532-4435. John M Conroy and Dianne P O’leary. Text summarization via hidden markov models. In Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval, pages 406–407. ACM, 2001. Hoa Trang Dang and Karolina Owczarzak. Overview of the tac 2008 update summarization task. In Proceedings of text analysis conference, pages 1–16, 2008. Misha Denil, Alban Demiraj, and Nando de Freitas. Extraction of salient sentences from labelled documents. arXiv preprint arXiv:1412.6815, 2014. Harold P Edmundson. New methods in automatic extracting. Journal of the ACM (JACM), 16(2):264–285, 1969. Dumitru Erhan, Pierre-Antoine Manzagol, Yoshua Bengio, Samy Bengio, and Pascal Vincent. The difficulty of training deep architectures and the effect of unsupervised pre- training. pages 153–160, 2009. Dumitru Erhan, Yoshua Bengio, Aaron Courville, Pierre-Antoine Manzagol, Pascal Vincent, and Samy Bengio. Why does unsupervised pre-training help deep learning? The Journal of Machine Learning Research, 11:625–660, 2010. Pierre-Etienne Genest, Fabrizio Gotti, and Yoshua Bengio. Deep learning for automatic summary scoring. In Proceedings of the Workshop on Automatic Text Summarization, pages 17–28, 2011. 14 Xavier Glorot, Yoshua Bengio, and Yann N Dauphin. Large-scale learning of embeddings with reconstruction sampling. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 945–952, 2011. Lars Kai Hansen and Peter Salamon. Neural network ensembles. IEEE Transactions on Pattern Analysis & Machine Intelligence, pages 993–1001, 1990. Geoffrey Hinton. A practical guide to training restricted boltzmann machines. Momentum, 9(1):926, 2010. Geoffrey Hinton and Ruslan Salakhutdinov. Discovering binary codes for documents by learning deep generative models. Topics in Cognitive Science, 3(1):74–91, 2011. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8):1771–1800, 2002. Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. Science, 313(5786):504–507, 2006a. Geoffrey E Hinton and Richard S Zemel. Minimizing description length in an unsupervised neural network. Preprint, 1997. Geoffrey E Hinton, Simon Osindero, and Yee-Whye Teh. A fast learning algorithm for deep belief nets. Neural computation, 18(7):1527–1554, 2006b. John J Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences, 79(8):2554–2558, 1982. ISSN 0027-8424. Khosrow Kaikhah. Text summarization using neural networks. Faculty Publications, Texas State Universit, 2004. Nicolas Le Roux and Yoshua Bengio. Representational power of restricted boltzmann machines and deep belief networks. Neural computation, 20(6):1631–1649, 2008. Liangda Li, Ke Zhou, Gui-Rong Xue, Hongyuan Zha, and Yong Yu. Enhancing diversity, coverage and balance for summarization through structure learning. In Proceedings of the 18th international conference on World wide web, pages 71–80. ACM, 2009. Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out: Proceedings of the ACL-04 workshop, volume 8, 2004. Elena Lloret and Manuel Palomar. Text summarisation in progress: a literature review. Artificial Intelligence Review, 37(1):1–41, 2012. Annie Louis and Ani Nenkova. Automatically assessing machine summary content without a gold standard. Computational Linguistics, 39(2):267–300, 2013. Vanessa Loza, Shibamouli Lahiri, Rada Mihalcea, and Po-Hsiang Lai. Building a dataset for summarization and keyword extraction from emails. Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), 2014. H. P. Luhn. The automatic creation of literature abstracts. IBM Journal of Research Development, 2(2):159, 1958. Inderjeet Mani and Mark T Maybury. Advances in automatic text summarization. MIT press, 1999. Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pages 807–814, 2010. Ani Nenkova and Kathleen McKeown. A survey of text summarization techniques. In Mining Text Data, pages 43–76. Springer, 2012. Ani Nenkova, Rebecca Passonneau, and Kathleen McKeown. The pyramid method: Incorporating human content selection variation in summarization evaluation. ACM Transac- tions on Speech and Language Processing (TSLP), 4(2):4, 2007. Jiquan Ngiam, Adam Coates, Ahbik Lahiri, Bobby Prochnow, Quoc V Le, and Andrew Y Ng. On optimization methods for deep learning. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 265–272, 2011. You Ouyang, Wenjie Li, Sujian Li, and Qin Lu. Applying regression models to query-focused multi-document summarization. Information Processing & Management, 47(2): 227–237, 2011. Martin F Porter. An algorithm for suffix stripping. Program, 14(3):130–137, 1980. Rajesh S Prasad, UV Kulkarni, and Jayashree R Prasad. Connectionist approach to generic text summarization. World Academy of Science, Engineering and Technology, 55, 2009. Vahed Qazvinian, Dragomir R Radev, and Arzucan Özgür. Citation summarization through keyphrase extraction. In Proceedings of the 23rd International Conference on Computational Linguistics, pages 895–903. Association for Computational Linguistics, 2010. Salah Rifai, Pascal Vincent, Xavier Muller, Xavier Glorot, and Yoshua Bengio. Contractive auto-encoders: Explicit invariance during feature extraction. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 833–840, 2011. Ruslan Salakhutdinov and Geoffrey Hinton. Semantic hashing. International Journal of Approximate Reasoning, 50(7):969–978, 2009. Rajesh Shardan and Uday Kulkarni. Implementation and evaluation of evolutionary connectionist approaches to automated text summarization. Journal of Computer Science, 2010. Nitish Srivastava and Ruslan R Salakhutdinov. Multimodal learning with deep boltzmann machines. In Advances in neural information processing systems, pages 2222–2230, 2012. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1):1929–1958, 2014. Krysta Marie Svore, Lucy Vanderwende, and Christopher JC Burges. Enhancing single-document summarization by combining ranknet and third-party sources. In EMNLP-CoNLL, pages 448–457, 2007. Joseph Turian, Lev Ratinov, and Yoshua Bengio. Word representations: a simple and general method for semi-supervised learning. In Proceedings of the 48th annual meeting of the association for computational linguistics, pages 384–394. Association for Computational Linguistics, 2010. Jan Ulrich, Gabriel Murray, and Giuseppe Carenini. A publicly available annotated corpus for supervised email summarization. 2008. Jan Ulrich, Giuseppe Carenini, Gabriel Murray, and Raymond T Ng. Regression-based summarization of email conversations. In ICWSM, 2009. Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096–1103. ACM, 2008. Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. The Journal of Machine Learning Research, 11:3371–3408, 2010. Ho Chung Wu, Robert Wing Pong Luk, Kam Fai Wong, and Kui Lam Kwok. Interpreting tf-idf term weights as making relevance decisions. ACM Transactions on Information Systems (TOIS), 26(3):13, 2008. Wayne Xin Zhao, Jing Jiang, Jing He, Yang Song, Palakorn Achananuparp, Ee-Peng Lim, and Xiaoming Li. Topical keyphrase extraction from twitter. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1, pages 379–388. Association for Computational Linguistics, 2011. Sheng-hua Zhong, Yan Liu, Bin Li, and Jing Long. Query-oriented unsupervised multi-document summarization via deep learning model. Expert Systems with Applications, 42 (21):8146–8155, 2015. 15 Introduction Related Work The Method Description Pre-training Stage Global Adjustment Word Representation Sentence Ranking Iterative Data Fusion and Voting Experimental Setup Results and Discussion Subject-oriented Summarization with SKE Key-phrase-oriented Summarization with SKE Key-phrase-oriented Summarization Versus Abstractive Summary with SKE Subject-oriented Summarization with BC3 Generalization and Data Comparison Conclusion and Future Work 
work_nll7n7rkwfcffow4gne4cohqqu	----	1 Impact of a spirometry expert system on general practitioners� decision-making Patrick JP Poels, Tjard RJ Schermer, Daan PA Schellekens, Reinier P Akkermans, Pieter F de Vries Robbé, Alan Kaplan, Ben JAM Bottema, Chris van Weel Short Title: Spirometry expert support in general practice Patrick JP Poels General Practitioner Radboud University Nijmegen Medical Centre, Department of General Practice (117), PO box 9101, 6500 HB, Nijmegen, The Netherlands, Phone ++31 24 361 33 15 Fax ++31 24 354 18 62, p.j.p.poels@hag.umcn.nl Tjard RJ Schermer Senior researcher Radboud University Nijmegen Medical Centre, Department of General Practice (117), PO box 9101, 6500 HB, Nijmegen, The Netherlands, Phone ++31 24 361 33 15 Fax ++31 24 354 18 62, t.schermer@hag.umcn.nl Daan PA Schellekens MD Radboud University Nijmegen Medical Centre, Department of General Practice (117), PO box 9101, 6500 HB, Nijmegen, The Netherlands, Phone ++31 24 361 33 15 Fax ++31 24 354 18 62, d.schellekens@hag.umcn.nl Reinier P Akkermans Statistician Radboud University Nijmegen Medical Centre, Department of General Practice (117), PO box 9101, 6500 HB, Nijmegen, The Netherlands, Phone ++31 24 361 33 15 Fax ++31 24 354 18 62, r.akkermans@ives.umcn.nl . Published on June 27, 2007 as doi: 10.1183/09031936.00012007ERJ Express Copyright 2007 by the European Respiratory Society. 2 Pieter F de Vries Robbé Professor of Medical Informatics Radboud University Nijmegen Medical Centre, Department of Medical Informatics (152), PO box 9101, 6500 HB, Nijmegen, The Netherlands, Phone ++31 24 361 32 51 Fax ++31 24 361 3504, p.devriesrobbe@mi.umcn.nl Alan Kaplan General Practitioner 17 Bedford Park Avenue, Richmond Hill , ON L4C 2N9, Ontario, Canada Phone: 905-883-1100, Fax: 905-884-1195, FOR4KIDS@sympatico.ca Ben JAM Bottema General Practitioner Radboud University Nijmegen Medical Centre, Department Postgraduate Training (166), PO box 9101, 6500 HB, Nijmegen, The Netherlands, Phone ++31 24 361 53 00 Fax++31 024 361 95 53 b.bottema@voha.umcn.nl Chris van Weel Professor of General Practice Radboud University Nijmegen Medical Centre, Department of General Practice (117), PO box 9101, 6500 HB, Nijmegen, The Netherlands, Phone +31 24 361 33 15 Fax ++31 24 354 18 62, c.vanweel@hag.umcn.nl Correspondence to: p.j.p.poels@hag.umcn.nl Word count: 3368 Number of tables: 4 (plus 1 online depository table) Number of figures: 2 Number of references: 25 3 Abstract This study assessed the impact of computerised spirometry interpretation expert support on the diagnostic achievements of general practitioners (GPs), and on GPs� decision-making in diagnosing chronic respiratory disease. We performed a cluster-randomised controlled trial in 78 GPs who completed 10 standardised paper case descriptions each. Intervention consisted of support for GPs� spirometry interpretation either by an expert system (expert support group) or by sham information (control group). Agreement of GPs� diagnoses were compared with an expert panel judgement, which served as the primary outcome. Secondary outcomes were additional diagnostic test rates, width of differential diagnosis, certainty of diagnosis, estimated severity of disease, referral rate, and medication or non-medication changes. Effects were expressed as odds ratios (OR) with 95% confidence intervals. There were no differences between the expert support and the control group in the agreement between GP and expert panel diagnosis of COPD (OR=1.08 [95% CI 0.70-1.66]), asthma (OR=1.13 [95% CI 0.70-1.80]), and absence of respiratory disease (OR=1.32 [95% CI 0.61-2.86]). A higher rate of additional diagnostic tests was observed in the expert support group (OR= 2.5 [95% CI 1.17-5.35]). Computerised spirometry expert support had no detectable benefit on GPs� diagnostic achievements and decision-making process when diagnosing chronic respiratory disease. Trial Number: http://www.clinicaltrials.gov/ct/show/NCT00131157?order=1 Keywords: Diagnosis Computer-Assisted, Expert Systems, Family Practice, Spirometry 4 Introduction Although all major COPD guidelines stress the central role of spirometry in diagnosing and managing chronic respiratory disease,1;2 this does not guarantee that general practitioners (GPs) will use spirometry consequently in care of their patients with respiratory symptoms.3;4 Most common barriers that impede utilisation of spirometry in general practice are the absence of properly trained staff,5 lack of time and practice support to fit spirometry into the daily practice routine,6 absence of a spirometer in the practice,7 and GP�s lack of confidence in the ability to interpret the test results.8;9 A recent survey showed that a third of Australian GPs interpreted less than one spirometry test per week.4 Due to this low prevalence of test interpretations, it seems difficult for GPs to become an expert in this. We have previously demonstrated the influence of spirometry on GPs� diagnostic achievements and management decisions in a non-randomised simulation study.10 Other recent non-randomised studies confirm that spirometry increases diagnostic rates of chronic respiratory disease and may lead to management changes in a general practice population.11;12 However, an absolute prerequisite for the use of spirometry is the validity (or �reliability�) of spirometric tests. In a previous study with patients with COPD we observed that the most relevant indices as measured by trained general practice staff were comparable with those measured in pulmonary function laboratories.13 Therefore, once GPs have had initial spirometry training and spirometry equipment and test validity is adequate, the next step to improve implementation of spirometry in general practice is to arrange for a possibility to receive continuous advice and support for the test interpretation.14 This could be done by means of a diagnostic computerised clinical decision support system.15;16 While there is already such an expert support system available on the market17 and GPs welcome such kind of support,18 empirical studies on the effects of ongoing expert support on the interpretative capacity and self confidence of GPs are warranted. 5 The objective of the present study was to assess the impact of expert support for the interpretation of spirometry tests on the diagnostic achievements of GPs, and on GPs� decision-making when diagnosing chronic respiratory disease. Methods Study design The study was a simulated cluster-randomised trial of GP�s diagnostic acuity of chronic respiratory disease in a process of diagnostic assessment of 10 standardised cases, with expert system support. An expert panel�s diagnosis of the cases served as the gold standard. Differences in GPs� diagnostic achievements and in GPs� decision-making process were compared between the study groups and within groups. Participants GPs from the catchment�s area of our academic hospital and from a specific general practice network of our department19 were invited by postal mailing to participate. Intervention GPs were randomly allocated to the (i) computerised spirometry expert interpretation support group, or (ii) the control group. GPs in the expert support group received the spirometry test results, flow-volume curve, plus the graphical interpretation and the textual interpretative notes. GPs in the control group received the spirometry test results, flow-volume curve, plus the volume-time curve (figure 1). The spirometry expert system (SpidaXpert®, Micro Medical Ltd, Rochester, UK),17 contains a diagnostic algorithm based on pre- and postbronchodilator FEV1/FVC and FEV1 values and the accompanying age, sex, and origin specific predicted values. The expert interpretation module in SpidaXpert® had been developed with funding of the Dutch Asthma Foundation by a group of independent experts (http://www.spirxpert.com/spirxpertgroup.htm). The spirometry interpretation is presented as coloured bars that indicate levels of FEV1/FVC and 6 FEV1, and compares the values before and after bronchodilatation. The graphical representation is further elucidated by a textual interpretation, that provides information on and suggestions for additional diagnostic testing and treatment options. GPs in the control group received the volume-time curve as sham information. We introduced sham information in the control group to be able to compare GPs re-assessment of a diagnosis in the control group in the same way as in the expert support group. Sham information is in fact a �placebo-effect� as we presented no new data to these GPs; we presented earlier data (i.e. the flow-volume curve) in each case again in another way (i.e. the volume-time curve). Although clearly important to evaluate the quality of forced expiratory manoeuvres (i.e., end of test criteria),20 from a diagnostic point of view the volume-time curve does not add relevant new information to the information provided by the flow-volume curve and the numeric test results. Prior to the study, participants were informed that they would receive additional information on the spirometry and asked to reconsider their diagnosis. No further specification was given of the nature or the background of that information. INSERT FIGURE 1 HERE Standardised case descriptions and gold standard Based on our experiences in a previous study,10 we knew beforehand that GPs are quite able to diagnose common respiratory disease patterns, whereas rare pathology and inadequate test results are more difficult for them to recognise. Furthermore, the challenge to differentiate COPD from other conditions that result in respiratory symptoms (e.g., heart failure, asthma) growths with age. That is the reason for including case descriptions of adult patients only, with a special focus on the 50-60 year age category. This category reflects daily practice patterns in primary care. The case descriptions - in which a GP would use spirometry as a diagnostic test � were: COPD (classified as GOLD stage 1 (n=1); GOLD stage 2 (n=1); GOLD stage 3 (n=2))2; asthma (n=2); allergic asthma (n=1); lung fibrosis 7 (n=1); no respiratory disease (n=1); incorrect test manoeuvre (n=1); exercise-induced asthma (n=1) [example case] (see online depository). At inclusion, a research assistant visited the participating GPs in their practice. During a 90 minutes audio taped session, an example case and 10 standardised cases were presented on a laptop computer using PowerPoint® slides. GPs worked through the cases in a random order. GPs first practised on one separate example case to become familiar with case structure. For each case, a concise medical history, the results of physical examination, and the medication were presented to the GP first. Subsequently, absolute predicted pre- and post bronchodilator spirometry test results (including FEV1, FVC, FEV1/FVC and flow-volume curves) were provided. GPs were asked to consider their diagnosis and management before the upcoming intervention. Next, GPs received additional information next to the spirometry test results: either the graphical representation of FEV1, FEV1/FVC together with interpretative notes (expert support group) or the volume-time curve (control group). Again GPs were asked to re-consider their diagnosis and management after the intervention. An example of the case structure is depicted in Figure 2. Due to time limitation we asked only for specific medication and non-medication changes after the intervention in cases with already diagnosed respiratory disease (6 out of 10 cases). INSERT FIGURE 2 HERE. Before their use in the study, the cases were judged by an expert panel consisting of two chest physicians, a GP (PP) with specific expertise in spirometry and a health scientist (TS). The panel consensus diagnoses served as �the gold standard� in the subsequent evaluation of GPs� diagnostic achievements. The whole approach was piloted in 4 GPs. Shortly after the first six visits, we adjusted the case set by switching the example case with a case out of the actual set. As a result, no data of the new introduced case were available for those first 6 GPs (equally divided over the 2 groups). 8 Primary and secondary outcome measures Difference in proportion of agreement of the cases� diagnosis between GP�s and the expert panel judgement before and after interpretation of spirometry served as the primary outcome and were directed at five outcome categories: (1) COPD; (2) asthma; (3) rare respiratory pathology (lung fibrosis); (4) absence of respiratory disease and (5) incorrect test manoeuvre. Six predefined secondary outcome measures were assessed using indicators that show the impact of the expert system intervention on GP�s decision-making process: (1) probability of ordering additional diagnostic tests (yes/no); (2) �width� of the differential diagnoses (i.e. the working diagnosis plus the number of alternative diagnoses considered by the GP); (3) GP�s certainty of the working diagnosis (self-scored between 0-10, [0 = uncertain, 10=certain]); (4) GP�s perception of severity of the working diagnosis (self-scored between 0-10, [0=no severe disease, 10=severe disease]); (5) probability of referral to secondary care (yes/no); (6) probability of medication and non-medication changes. Medication change included either to stop or lower of inhaled corticosteroids or bronchodilators or to start of bronchodilators, inhaled corticosteroids, oral corticosteroids, and combination drugs. Non-medication included giving smoking cessation advice. Sample size Calculation of the sample size was based on an estimated relevant proportion correctly interpreted cases after spirometry expert support of 25% compared with no expert support. Assuming a correctly interpreted proportion of cases without support of 50%5, α=0.05, 1- β=0.80, and an intra-cluster correlation r=0.18, 31 GPs were required in each randomisation group. To allow for drop-outs and subgroup analyses we aimed at including a minimum of 70 GPs. 9 Randomisation The research assistant used restricted randomisation (minimisation) with a computer program on a laptop computer using three stratification factors: GP�s prior experience with the specific computerised spirometry interpretation support package (yes/no), the average number of spirometry tests a GP reported to interpret per week, and GP�s experience (years) with spirometry. The researchers and the statistician (RA) were blinded while assessing and reporting all outcomes. Statistical analysis Agreement between GP�s and the expert panel judgement was expressed as percentages with 95% confidence intervals (95%CI). Multilevel regression logistic modelling was used to account for the intra-cluster correlation induced by the fact that each GP assessed more than one case, and the fact that the same cases were applied repeatedly in different GPs. We performed multilevel logic analyses for dichotomous variables and multilevel regression analyses for continuous variables in SAS V8.2 for Windows (SAS Institute Inc, Cary USA 1999-2001). Odds ratios (OR) with 95%CI were calculated to evaluate differences in percentages of agreement before and after the intervention with the expert judgement between the study groups. Sensitivity, specificity, positive and negative predictive values (further referred to as PPV and NPV, respectively), and the diagnostic odds ratio (DOR)21 with 95% confidence intervals were calculated for GP judgements of COPD, asthma, rare respiratory pathology and no respiratory disease after the intervention. Odds ratios with 95% confidence intervals were also used to evaluate differences in indicators GPs� decision- making process. To detect possible effect modifications before intervention, subgroup analyses were performed for GP�s prior experience with spirometry, GP�s prior experience with expert support and GP�s number of interpreted spirometry tests per week. 10 Results Baseline characteristics of GPs Between January and October 2006 we enrolled 78 GPs in our study; 36 were allocated to the expert support group and 42 to the control group (Figure 1). All GPs completed the study. Relevant characteristics at baseline were similar between the two groups (table 1). Table 1 Baseline characteristics of all randomised GPs. Expert support group (n=36 GPs) Control group (n=42 GPs) Type of practice, n (%) -single handed -duo -group (> 3 GPs) -multidisciplinary health care centre 2 (5) 9 (25) 15 (42) 10 (28) 4 (10) 9 (21) 19 (45) 10 (24) Gender, % male 64 57 GP�s experience with spirometry in years, mean (SD) 5.5 (4.3) 5.3 (3.3) No. of spirometry results interpreted per week, mean (SD) 1.4 (0.8) 1.4 (0.7) Prior experience with expert support, % yes 47 36 Primary outcome: diagnostic achievements by GPs GPs assessed in total 774 cases; 357 cases in the expert support group and 417 cases in the control group. There was no difference between the expert support and control group in agreement on judgement between GPs and the expert panel for presence of COPD, asthma, absence of respiratory disease, and incorrect test manoeuvre after intervention (table 2). GPs� agreement with the expert panel for all cases - except the incorrect test manoeuvre case - was before intervention 66.0% (expert support) versus 65.9% (control) and after intervention 68.5% (expert support) versus 63.5% (control). 11 Table 2 Agreement on case diagnoses between GPs� and expert panel judgement before and after intervention. The Odds Ratio expresses the difference in GP�s judgement before and after intervention, expert support compared to control. Expert support group (n=357 cases) Control group (n=417 cases) Expert panel Odds ratio (95% CI) intervention intervention judgment GP judgement before after before after presence of: COPD 32.5% 32.5% 32.4% 30.7% 40% 1.08 (0.70 � 1.66) Asthma 23.5% 25.2% 23.5% 23.0% 30% 1.13 (0.70 � 1.80) Rare respiratory pathology 0.6% 0.3% 0.0% 0.0% 10% NA Absence of respiratory disease 2.8% 3.6% 3.4% 4.4% 10% 1.32 (0.61 � 2.86) Incorrect test manoeuvre NA 5.9% NA 6.71% 10% 0.87 (0.48 � 1.56) NA = not available, 95%-CI = 95% confidence interval Although the DORs in the expert support group were consistently higher than in the control group, we found no significant differences between the groups (table 3). GPs did not recognise an incorrect test manoeuvre in 28.6% (expert support) and 28.6% (control) of cases. For cases with the conditions asthma and absence of respiratory disease we found the highest NPVs. 12 Table 3 Sensitivity, specificity, predictive values, and diagnostic odds ratios for GPs� judgment after intervention. NA = not available, 95%-CI = 95% confidence interval Secondary outcomes: indicators of GPs� decision-making process GPs in the expert support group ordered slightly more additional diagnostic tests compared with the control group (OR 2.5 [95%CI 1.2 � 5.4]) (table 4). There were no significant differences between the two groups for other secondary outcome measures. There were also no specific changes (start, stop or lower) in medication (bronchodilators, inhaled steroids or non-pulmonary drugs) between the study groups. COPD Asthma Rare respiratory pathology Absence of respiratory disease Expert support group Sensitivity,% Specificity,% Positive predictive value,% Negative predictive value,% Diagnostic odds ratio 95%-CI 80.6 77.5 70.7 85.5 14.2 8.46-23.98 83.3 90.4 78.9 92.6 46.9 24.35-90.23 2.8 99.4 33.3 90.1 4.6 0.59-35.90 39.4 97.5 61.9 94.0 25.7 9.74-67.71 Control group Sensitivity,% Specificity,% Positive predictive value,% Negative predictive value,% Diagnostic odds ratio 95%-CI 76.2 79.1 71.1 83.1 12.1 7.60-19.34 76.2 90.7 78.0 89.8 31.3 17.73-55.22 0.0 99.7 0.0 89.9 NA 0.0-34.79 35.9 97.6 60.9 93.7 23.0 9.22-57.17 p-value 0.65 0.36 NA 0.87 13 Table 4 Impact of the intervention on six indicators of GPs� decision-making process. The Odds Ratio expresses the difference in an indicator of GPs� decision-making process before and after the intervention, expert support compared to control. Statistically significant differences (p<0.05) are printed in bold. Expert support group (N=357) Control group (N=417) Odds Ratio 95% CI intervention intervention Indicators before after before after (1) Additional diagnostic tests,% - X ray imaging - blood tests - lung function - prednisolon course - electrocardiography - othera 70.2 48.7 29.2 13.2 11.2 3.6 1.1 5.2 0.9 0.9 2.3 2.6 0.3 0.0 75.3 51.3 39.1 15.2 8.9 6.0 0.7 3.0 0.8 1.3 0.3 1.0 0.2 0.0 2.5 1.27 1.0 1.0 1.0 1.0 NA (1.17 � 5.35) (0.25 � 6.41) (0.65 � 1.55) (0.59 � 1.70) (0.61 � 1.63) (0.55 � 1.80) NA (2) Width of differential diagnoses, mean (SD) 2.2 (1.0) 1.7 (0.9) 2.3 (1.0) 1.7 (1.0) 1.0 (0.81 � 1.19) (3) Certainty of diagnosis, mean (SD) 6.8 (2.0) 7.1 (1.9) 7.3 (1.9) 7.3 (1.8) 1.0 (0.57 � 1.43) (4) Perception of severity of diagnosis, mean (SD) 6.0 (2.2) 5.9 (2.3) 6.3 (2.1) 6.3 (2.1) 1.0 (0.57 � 1.42) (5) Referral rate,% 18.6 1.7 17.8 2.5 1.0 (0.60 � 1.66) (6) Medication & non-medication changesb,% yes - 76.2 - 69.1 1.44 (0.80 � 2.59) - Stop or lower medication - inhaled corticosteroids - bronchodilators NA NA 10.2 0.9 NA NA 6.3 0.8 1.67 1.17 (0.85 � 3.28) (0.16 � 8.40) - Start medication - short-acting bronchodilators - long-acting bronchodilators - inhaled corticosteroids - oral corticosteroids - combinational drug NA NA NA NA NA 26.9 15.3 31.0 6.5 3.7 NA NA NA NA NA 22.2 15.5 31.0 5.6 2.8 1.27 1.01 1.00 1.34 1.27 (0.73 � 2.20) (0.46 � 2.18) (0.68 � 1.49) (0.44 � 4.10) (0.26 � 6.19) - Non-medication changes - smoking cessation advice NA 30.1 NA 33.7 0.85 (0.57 � 1.26) NA = not available, 95%-CI = 95% confidence interval a includes urine test, gastroscopy, ergometry, blood pressure, temperature and oxygen saturation b this information was available for 6 out of 10 cases (expert support n=216, control group n=252 cases) Subgroup analyses Neither GP�s experience with spirometry (OR 1.02 [95%CI 0.97 - 1.06]), nor GP�s prior experience with expert support (OR 0.97 [95%CI 0.72 - 1.31]), nor GP�s number of interpreted spirometry test per week (OR 1.02 [95%CI 0.84 - 1.23]) were associated with the effectiveness of expert support, as their agreements with the expert panel was not different before intervention. If GPs interpreted more spirometry test per week and had prior experience with expert support the probability of agreement with the expert panel before intervention increased. It decreased if GPs had no prior experience with expert support (interaction effect p=0.02). 14 Discussion Statement of principal findings Computerised spirometry expert support for the interpretation of spirometry tests by GPs had no detectable benefit over sham information on GPs� diagnostic achievements of chronic respiratory disease. Overall, expert support did not influence GPs� decision-making process. Strengths of the study This is the first diagnostic study that assesses the impact of a commercially available computerised expert support for spirometry in a randomised simulation study in primary care. The study used standardized patients, which means that all participants were faced with the same diagnostic challenges. This can only be achieved in an in-vitro design as in real life practice patient mix would cause difficult to capture variation in diagnostic challenges. The standardised complex and original method that we used to assess this impact had been used before in a non-randomised design.10 Based on that information we were able to create a balanced mixture of cases relevant for GPs. We focussed stronger on the confirmative role of spirometry than the exclusive role of spirometry in primary care. To avoid bias, cases were presented in a (computer generated) random order, and analyses were performed blinded for both the investigators and the statistician (RA). Subgroup analyses did not show any difference in baseline diagnostic achievements of GPs on prior experience with spirometry, on GP�s prior experience with this expert support system, and on the number of spirometry test a GP interpreted per week. Therefore, the external validity seems quite good given the fact that the participants were not specifically interested in spirometry. 15 Possible limitations Our trial has some limitations. In a diagnostic assessment of chronic respiratory disease, GPs� consideration to perform spirometry in case of a intermediate prior probability of disease is a great diagnostic step.21 This step was already foreseen in our study design. The next step of diagnostic refinement does not seem to influence the posterior probability extensively. In our study the diagnostic achievements of GPs in both groups were high (prior probability of a correct diagnosis was ~66%). Overall only 4.3% of initial diagnoses changed after intervention. As the posterior probability in both groups was nearly the same as the prior probability, the role for expert support to change diagnosis and management was very small. Furthermore, the diagnostic achievements of the GPs exceeded our assumptions in the power calculation (50% correct diagnoses without expert support). Probably instruction and support for these GPs had not been effective, as these GPs could already be considered as �experts� due to prior participation in other studies or postgraduate spirometry training programs from our department. Therefore, the expert system had hardly additional value and could be considered as a sort of luxury appendix for these GPs. We found a large within group difference for ordering additional diagnostic tests, which may be an effect of the study design: because GPs expected from us the results of their diagnostics already discounted before intervention they hardly reassessed their diagnostics after intervention. However, the objective was to re-assess GP�s opinion when new information (i.e. expert support) was available, regardless of their earlier assessment in the same case. From a methodological point of view, one could question the use of the volume-time curve as sham information. Theoretically these curves do not add new information to GPs after presentation of the flow-volume curves. On the one hand this is not really �usual care� as most GPs in our country are trained to look at flow-volume curves rather than volume-time curves. On the other hand, the volume-time curve is much more intuitive and may have 16 improved unconscious performance of spirometry interpretation in the control group. Furthermore, providing the expert panel and GPs in our study with the fixed cut-off <0.7 in stead of the lower limit of normal (LLN) for the FEV1/FVC ratio in the standardised cases may have led to an overestimation of diagnosed airflow obstruction.22Further discussion about the pros and cons of using the fixed cut-off versus the LLN for FEV1/FVC23 is beyond the scope of this paper. Finally, a possible reason why we could not demonstrate differences in diagnostic achievements should be sought in the used expert support system. Although the expert support system we used meets the criteria of a good system15 � involvement of authors by development, integrated in computer, displaying specific recommendations at the right place and time � it was not actually tested in the target group (i.e. GPs) before the study. Therefore, it may not optimally comply with the decision-making process of GPs. The information presented by the system to the GP possibly lacked explanation of what the output means exactly. These are known barriers to the adoption of expert support in primary care.24 Relation to other studies A recent systematic review demonstrated two relevant issues with respect to expert support systems: (1) the effects of diagnostic expert support systems on practitioners performance was low, and (2) trials evaluating diagnostic systems were scarce.16 Currently, there are no similar expert support studies available to compare our results with directly. It is important to realise that - like electrocardiography (ECG) - spirometry is a highly complex diagnostic tool in the perception of many GPs. Although a promising idea to evaluate the ECG interpretations skills of GPs and the value of automatic ECG recorded interpretations, a recent study lacked the right design to compare our results with.25 Like the results from the study from Jensen et al.25 we found in our study that the positive predictive values were lower than the negative predictive values. For the cases with the conditions asthma and 17 absence of respiratory disease we found the highest NPVs. This probably reflects the fact it is more difficult for a GP to confirm the presence of a disease than to exclude its presence. The acuity of GPs� interpretation of test results had been evaluated by others. In 1999, Eaton et al. found already that 53% of GPs� interpretation of spirometry test results was judged correct according to an expert panel.5 Recently, Raghununath et al. found that the agreement in interpretation of spirometry and peak flow results between nurses, GPs and an expert panel was only 20%.9 The lower agreement in this latter study could probably be explained by the fact that GPs as well as nurses (i.e. less trained professionals) assessed a common diagnosis. Furthermore, contrary to the nurses and GPs, the expert panel did not have detailed clinical history information to assess their final diagnosis on. Due to design artefact, interpretation of their study results is difficult. Results of our study confess Eaton�s5 results and show that generally GPs have made progress in the interpretation of test results relevant for respiratory diseases in primary care. The current acuity of GP�s interpretation of the test results should weaken earlier reported lack of confidence in the ability to interpret the test results.8;9 Unanswered questions and future research Generally, the first question to answer is how to achieve optimal quality spirometry results in primary care outside of research settings. The second question to answer is what is the effective way to give continuous expert support for the interpretation of spirometry test results given a situation of optimal quality results.14 Continuous expert support could be provided by means of consultation of feedback from a chest physician or by means of an expert support system. Results of the current study add to knowledge that computerised spirometry expert support had no detectable benefit over sham information on GPs� diagnostic achievements and decision-making process when diagnosing chronic respiratory disease. Comparison of support from a chest physician versus computerised expert support for spirometry test results calls for further studies. 18 Acknowledgements We thank all participating GPs; we thank Yvonne Heijdra (PhD) and Johan Molema (PhD), chest physicians, for their role in the expert panel; Margreet van den Bogart-Jansen (MSc), Antoinette Marks (RN) who helped with the interviews of the GPs; Joke Grootens- Stekelenburg for data management. Executive steering committee: Henk van den Hoogen, MSc; Annelies Jacobs, MSc PhD; Philip Quanjer, MD PhD; Bart Thoonen, MD PhD. Funding: This study was supported by grant 3.4.02.18 of the Netherlands Asthma Foundation and grant 920.03.265 of the Netherlands Organisation for Health Research and Development. The authors� work was independent of the funders (the funding source had no involvement). Ethical approval: This study was approved by the medical ethics review board of the Radboud University Nijmegen Medical Centre. Competing interests: All authors declare have nothing to declare. Legends Figure 1 Participants and intervention Figure 2 Schematic representation of a case structure Reference List (1) National Institute for Clinical Excellence (NICE). Chronic obstructive pulmonary disease: national clinical guideline for management of chronic obstructive pulmonary disease in adults in primary and secondary care. Thorax 2004; 59 (Suppl I):i1-i232. (2) Global Initiative for Chronic Obstructive Lung Disease; www.goldcopd.com. Date accessed: November 15 2006. (3) Caramori G, Bettoncelli G, Tosatto R, Arpinelli F, Visona G, Invernizzi G et al. Underuse of spirometry by general practitioners for the diagnosis of COPD in Italy. Monaldi Arch Chest Dis 2005; 63(1):6-12. 19 (4) Johns DP, Burton D, Walters JA, Wood-Baker R. National survey of spirometer ownership and usage in general practice in Australia. Respirology 2006; 11(3):292-298. (5) Eaton T, Withy S, Garrett JE, Mercer J, Whitlock RM, Rea HH. Spirometry in primary care practice: the importance of quality assurance and the impact of spirometry workshops. Chest 1999; 116(2):416-423. (6) Poels PJ, Schermer TR, Jacobs A, Akkermans RP, Hartman J, Bottema BJ et al. Variation in spirometry utilisation between trained general practitioners in practices equipped with a spirometer. Scan J Prim Health Care 2006; 24(2):81-87. (7) O'Dowd LC, Fife D, Tenhave T, Panettieri RA, Jr. Attitudes of physicians toward objective measures of airway function in asthma. Am J Med 2003; 114(5):391-396. (8) Bolton CE, Ionescu AA, Edwards PH, Faulkner TA, Edwards SM, Shale DJ. Attaining a correct diagnosis of COPD in general practice. Respir Med 2005; 99(4):493-500. (9) Raghunath AS, Innes A, Norfolk L, Hannant M, Greene T, Greenstone M et al. Difficulties in the interpretation of lung function tests in the diagnosis of asthma and chronic obstructive pulmonary disease. J Asthma 2006; 43(9):657-660. (10) Chavannes N, Schermer T, Akkermans R, Jacobs JE, van de GG, Bollen R et al. Impact of spirometry on GPs' diagnostic differentiation and decision-making. Respir Med 2004; 98(11):1124-1130. (11) Dales RE, Vandemheen KL, Clinch J, Aaron SD. Spirometry in the primary care setting: influence on clinical diagnosis and management of airflow obstruction. Chest 2005; 128(4):2443-2447. (12) Walker PP, Mitchell P, Diamantea F, Warburton CJ, Davies L. Effect of primary care spirometry on the diagnosis and management of COPD. Eur Respir J 2006; 28(5):945-952. (13) Schermer TR, Jacobs JE, Chavannes NH, Hartman J, Folgering HT, Bottema BJ et al. Validity of spirometric testing in a general practice population of patients with chronic obstructive pulmonary disease (COPD). Thorax 2003; 58(10):861-866. (14) Poels PJ, Schermer TR, van Weel C, Calverley PM. Spirometry in chronic obstructive pulmonary disease. BMJ 2006; 333(7574):870-871. (15) Purcell GP. What makes a good clinical decision support system. BMJ 2005; 330(7494):740- 741. (16) Garg AX, Adhikari NKJ, McDonald H, Rosas-Arellano MP, Devereaux PJ, Beyene J et al. Effects of Computerized Clinical Decision Support Systems on Practitioner Performance and Patient Outcomes: A Systematic Review. JAMA 2005; 293(10):1223-1238. (17) SpidaXpert, software for interpreting pulmonary function. www.spirxpert.com. Date accessed: January 15 2007. (18) Poels PJP, Schermer TRJ, Akkermans RP, Jacobs A, van den Bogart-jansen M, Bottema BJAM et al. General practitioners' needs for ongoing support for the interpretation of spirometry tests. Eur J Gen Pract 2007; 2007(13):16-19. (19) van Weel C, de Grauw WJC. Family practices registration networks contributed to primary care research. J Clin Epidemiol 2006; 59(8):779-783. (20) Miller MR, Hankinson J, Brusasco V, Burgos F, Casaburi R, Coates A et al. Standardisation of spirometry. Eur Respir J 2005; 26(2):319-338. 20 (21) Knottnerus JA, van Weel C, Muris JW. Evidence base of clinical diagnosis: Evaluation of diagnostic procedures. BMJ 2002; 324:477-480. (22) Hardie JA, Buist AS, Vollmer WM, Ellingsen I, Bakke PS, Morkve O. Risk of over-diagnosis of COPD in asymptomatic elderly never-smokers. Eur Respir J 2002; 20(5):1117-1122. (23) Schermer TR, Quanjer PH. COPD screening in primary care: who is sick? Prim Care Respir J 2007; 16(1):49-53. (24) Short D, Frischer M, Bashford J. Barriers to the adoption of computerised decision support systems in general practice consultations: a qualitative study of GPs' perspectives. Int J Med Inform 2004; 73(4):357-362. (25) Jensen MSA, Thomsen JL, Jensen SE, Lauritzen T, Engberg M. Electrocardiogram interpretation in general practice. Fam Pract 2005; 22(1):109-113. 21 22 
work_nmel43xsgvdbnamkhipmhqvbq4	----	Minimum Cost Consensus Models based on Random Opinions Ning Zhanga, Zaiwu Gonga,∗, Francisco Chiclanab aSchool of Economics and Management, Nanjing University of Information Science and Technology, Nanjing 210044, China bCentre for Computational Intelligence, Faculty of Technology, De Montfort University, Leicester, UK Abstract In some complex group decision making cases, the opinions of decision makers (DMs) present random characteristic. However, it is difficult to determine the range of opinions by knowing only their probability distributions. In this paper, we construct cost consensus models with random opinions. The objective function is obtaining the minimum consensus budget under a certain confidence level. Nonetheless, the constraints restrict the upper limit of the consensus cost, the lower limit of DMs’ compensations, and the opinions deviation between DMs and the moderator. As such, probabilistic planning based on a genetic algorithm is designed to resolve the minimum cost consensus models based on China’s urban demolition negotiation, which can better simulate the consensus decision- making process and obtain a satisfactory solution for the random optimization consensus models. The proposed models generalize both Ben-Arieh’s minimum cost consensus model and Gong’s consensus model with uncertain opinions. Considering that the opinions of DMs and the moderator obey various distributions, the models simulate the opinion characteristics more effectively. In the case analysis, a sensitivity analysis method is adopted to obtain the minimum budget, and probabilistic planning based on genetic algorithm to obtain a satisfactory solution that is closer to reality. Keywords: Group decision making, Consensus, Probability distribution, Probabilistic planning, Genetic algorithm. 1. Introduction Group consensus is mainly used to analyze the consistency of group opinions. Four aspects of consensus research can be summarized as follows: the degree of consensus is related to the preferences of decision makers (DMs); to the level of DMs’ logical thinking, that is, the level of DMs’ consistency; consensus is also reflected in the dynamic adjustment nature of the process; and consensus need to aggregate individual opinions in a group decision making (GDM) process. The first category focuses on the different representation formats of preferences in consensus prob- lems. In GDM processes, different DMs will have and provide opinions in different ways, based on ∗Corresponding author Email addresses: yguangzhiying@163.com (Ning Zhang), yzwgong26@163.com (Zaiwu Gong), chiclana@dmu.ac.uk (Francisco Chiclana) Preprint submitted to Expert Systems with Applications. Revised version of ESWA-D-17-00847 July 23, 2017 their knowledge background, interests, habits and other factors, among others in the form of crisp pref- erences (Fishburn, 1979), fuzzy preferences (Nurmi, 1981; Fodor and Roubens, 1994; Cabrerizo et al., 2015), interval-valued preferences (Dong et al., 2016), triangular fuzzy preferences (Wu and Chiclana, 2014a), intuitionistic fuzzy numbers (Wu and Chiclana, 2014b; Xu and Liao, 2015), linguistic assess- ments (Zadeh, 1975; Wu et al., 2015; Cid-Lopez et al., 2017). Moreover, in the selection and evaluation of various alternatives, utility function, preference relation, and preference orderings (Chiclana et al., 1998) are the manifestation of DMs’ preference structures. For instance, Chiclana et al. (2001) re- garded multiplicative preference relations as a preference representation structure, and constructed a multipurpose decision-making model based on fuzzy preference relations. Jiang et al. (2013) con- structed two different consensus models with intuitive multiplicative preference relations, and tested, achieved, and improved the group consensus level. To solve the multi-person decision-making problem with heterogeneous preference representation structures, Zhang et al. (2015) composed a new consen- sus model, while Dong et al. (2015) constructed a new framework based on the prospect theory and DMs’ psychological behaviors. Moreover, Wang and Li (2016) used quadratic programming models to solve GDM with incomplete preference relations. Through a specialized model in the R software environment, Ureña et al. (2016) applied fuzzy preference relations to support fuzzy GDM processes. The second category deals with individual consistency in consensus problems. It is beneficial to obtain more rational and reliable results by considering individual consistency and group consensus. In this field, Wu and Xu (2012) provided a decision support model to promote group consensus while keeping individual consistency for DMs; later, they developed a consensus reaching process for man- aging individual and group rationality of hesitant fuzzy linguistic preference relations (Wu and Xu, 2016). In the process of consensus building with missing values, Wu and Chiclana (2014c) mainly utilized multiplicative consistency of intuitionistic reciprocal preference relations. Driven by the con- sistency of group decision, Ureña et al. (2015) proposed the method to solve incomplete reciprocal intuitionistic preference relations. In addition, Chiclana et al. (2013) applied different distance func- tions to individual consistency, and analyzed those different convergence to consensus speeds. Based on a multiplicative analytic hierarchy process (AHP) model with log normal errors, Lin and Kou (2015) put forward a Bayesian revision method for improving the individual pairwise comparison ma- trices’ consistency. Additionally, Li et al. (2017) developed a consistency-driven model based on the two-tuple linguistic model, which benefits reaching consensus with linguistic preference. The third category addresses dynamic consensus problems. Group consensus itself is a dynamic adjustment process, which reflects active consensus (stigmergy) and non-active consensus. For ac- tive consensus within large group emergency decision-making, Xu et al. (2015) presented a dynamic consensus method based on an exit-delegation mechanism, which helped create a solution to satisfy the consistency and consensus criteria simultaneously. By exploring a Markov chain method, Dong and Cooper (2016) structured a peer-to-peer dynamic adaptive consensus reaching model. Moreover, 2 Zhao et al. (2017) explored an H∞ sliding mode control model to solve the scaled consensus control problem. For non-active consensus, a moderator of extraordinary leadership participates in a consen- sus process, which is beneficial to accelerate consensus and improve the efficiency. To minimize the deviation between DMs’ and the moderator’s opinions, Ben-Arieh and Easton (2007) and Ben-Arieh et al. (2009) adopted the distance-measured method to construct the minimum and minimum quadratic cost consensus models. Generalizing Ben-Arieh’s models, Gong et al. (2015a) explored consensus models based on minimum cost and maximum return, and developed consensus models with interval preference opinions. For investigating network-based leader-following consensus problems, He et al. (2017) designed nonlinear multi-agent systems under distributed impulsive control. Dong et al. (2017) exploited a convergent group AHP consensus reaching model with a twofold feedback mechanism, which fully considered the leader or customer opinion in the process of dynamic interaction. The fourth category relates to information aggregation in consensus problems. To obtain an ac- ceptable consensus result for DMs in GDM, it is feasible to aggregate individual opinion through model construction and graph analysis (Fu and Yang, 2010; Von Winterfeldt and Edwards, 1986). Analytic Hierarchy Process (AHP) (Saaty, 1980) is a systematic and hierarchical analysis method combining qualitative with quantitative information, which is applied to information aggregation problems with significant impact and contribution to group consensus. Indeed, Forman and Peniwati (1998) aggre- gated DMs’ judgements and AHP’s priorities to deal with group decision problems. To study the consensus process, Honert (1998) constructed stochastic group preference models based on AHP mul- tiplicative variant. Regan et al. (2006) constructed a mathematical consensus convergent model and demonstrated its utility in an environmental management framework. Besides, Altuzarra et al. (2007) put forward a new priorization procedure based on AHP-group decision making. Srdjevic and Srdjevic (2013) adopted AHP to combine DMs’ local priorities, and reduced the group inconsistency; while Srdjevic et al. (2013) designed a two-phase algorithm based on optimal aggregation for consensus. Nevertheless, previous aggregation models focused on crisp, fuzzy, interval, intuitionistic fuzzy, lin- guistic and other preference representation forms, but rarely referred to opinions/preferences obeying random distributions. Stochastic programming is a type of mathematical programming that deals with random data, being also an effective tool for modeling optimization problems with random variables. For decades, the stochastic programming problem has been widely used in decision science, investment portfolio, system optimization, and other fields. For instance, Sakawa and Matsui (2013) constructed two-level linear programming problems under fuzzy random environments by considering the cooperative re- lation between DMs, and derived a satisfactory solution for the upper-level DM. In multi-attribute decision-making problems with interval values and random data, Wu et al. (2016) defined interval number with probability distribution and also identified a new stochastic dominance degree definition. For investment portfolio optimization, Sadati and Nematian (2013) developed a portfolio optimization 3 model involving fuzzy random variables, and maximized the possibility and necessity of investment portfolio. By employing random fuzzy variable to describe the stochastic return of individual se- curity with fuzzy information, Qin (2016) exploited mean-absolute deviation portfolio optimization models, while for reducing aircraft noise and optimizing the structural-acoustic system, Wang et al. (2017) proposed a robust optimization method of structural-acoustic coupled systems with random parameters. Chance constrained programming (Charnes and Cooper, 1959), an important branch of stochastic programming, mainly deals with situations with constraints containing random variables, and the decision must be made before the implementation of these random variables. It is considered that the decision may not satisfy the constraints when the adverse situation occurs, i.e., the decision cannot satisfy the constraint to a certain extent when a principle is adopted, while the probability the constraint should be established is not below a certain confidence level 1 −α. Hong et al. (2011) employed a gradient-based Monte Carlo method to realize convex approximations, which was used for joint chance-constrained programs. To obtain the optimal strategies for carbon capture, utilization, and storage, Dai et al. (2014) exploited an inexact mλ-measure fuzzy chance-constrained programming. Moreover, for the improvement of wheat crop quality, Bai et al. (2015) instructed a multi-step rolled forward chance-constrained model to deal with uncertain climate in the harvest season. Aiming at the New York center location problem, Gao and Qin (2016) introduced uncertain variables of travel times and established a chance-constrained programming model. However, opinions are almost always uncertain in GDM, being reflected by interval-valued and intuitionistic fuzzy numbers and other characteristics. Moreover, it is difficult to express DMs’ opinions only with probability distribution (e.g., the opinions of DMs can obey a normal distribution with mean x and variance σ2). This article studies the consensus problem when the opinions of DMs or the moderator obey a certain distribution or are intervals. In cost consensus decision making problems (Ben-Arieh and Easton, 2007; Ben-Arieh et al., 2009; Gong et al., 2015a) it is difficult to obtain a cost paid to all DMs for consensus below budget. In this article, it s assumed that the cost paid to all DMs for consensus is lower than the budget under a certain confidence level, i.e., the cost is less than a chosen probability. To obtain the minimum consensus budget under a certain confidence level and to place restrictions on the probability of the cost paid for consensus no more than budget that is not less than a certain confidence level, we construct the cost consensus model based on a random distribution considering the probabilistic budget constraint. The advantages of the new models proposed in the paper are as follows: • Probability constraints are more realistic than complete constraints; • Apart from interval values, adding different random distributions can increase the available of uncertain values; 4 • With the generalization of the minimum cost consensus models by Ben-Arieh and Easton (2007) and Gong et al. (2015b), our models consider three constraints, including the cost upper limit, the lower DM compensations limit, and the opinion deviation between DMs and the moderator. The rest of this article is structured as follows. Section 2 introduces the principle and basic assumptions of the stochastic consensus model. Section 3 constructs the cost consensus model based on random distribution, only considering the probabilistic constraint of the moderator’s maximum budget. By improving the model, we create a probabilistic constraint of the DM’s minimum compensation and the constraint of distance deviation between the moderator’s and DM’s opinions, and develop a multi-constraint cost consensus model. In Section 4, we use a sensitivity analysis and probabilistic programming based on genetic algorithms (GA), combine them with an application to China’s urban demolition negotiation, and simulate the minimum budget in a certain confidence level under different constraints. The article is closed with Section 5 where conclusions are drawn. 2. Model principle and basic assumptions Based on consensus research, by introducing an ε general consensus concept, Ben-Arieh and Easton (2007) employed a distance-measured method to construct the minimum cost consensus model with the threshold ε constraint of opinion deviation: Min ϕ = m∑ i=1 ci|o′i −oi| s.t. |o′i −oi| ≤ ε,i ∈ M = {1, 2, . . . ,m} (1 − 1) (1) where i ∈ M = {1, 2, . . . ,m} denotes the m DMs participating in GDM, oi is the original DM i’s opinion, o′i denotes i’s opinion after adjustment (consensus opinion), and |o ′ i − oi| is the deviation between the original opinion oi and the adjusted opinion o ′ i. We define ci to be the cost of changing a unit opinion, and the objective function in Model (1) can be interpreted as the minimum cost for obtaining final consensus. In constraint (1-1), ε is the maximum units of the deviation for reaching common consensus, and so |o′i − oi| ≤ ε denotes the threshold limits of deviation between the DM i’s original and consensus opinion. Subsequently, due to the construction of the minimum quadratic cost consensus model, Ben-Arieh et al. (2009) discussed the problem of single-objective quadratic cost decision, i.e., changed the objective function to m∑ i=1 ci(o ′ i −oi) 2, and proposed a linear time algorithm for the minimum consensus cost. By introducing aggregation operators, Zhang et al. (2013) generalized the minimum cost con- sensus model, and put forward minimum-cost and maximum expert consensus models. Based on the primal-dual linear programming theory, Gong et al. (2015a) constructed two consensus models based on minimum cost and maximum return regarding the DMs or moderators, and investigated the relationship and economic significance between models and variables. In realistic decision making, the DMs’ opinions often present the following uncertain characteristics: 5 • Employing interval values to represent the opinions of DMs and the moderator, and determining the explicit values (i.e., optimal values) of DMs and the moderator using the optimization method, Gong et al. (2015b) constructed the minimum cost consensus model based on interval opinions: Min ϕ = m∑ i=1 ωi|o′ − [oli + αi(ori −oli)]| s.t.   o′ ∈ O oi ∈ [oli,ori], i ∈ M 0 ≤ αi ≤ 1, i ∈ M (2) where i ∈ M denotes the m DMs participating in GDM, oi is the DM i’s interval valued opinion with lower and upper bounds oli and ori, respectively. From interval [oli,ori] the following crisp value oli + αi(ori − oli) is derived, where 0 ≤ αi ≤ 1. o′ is the moderator’s opinion, and O = {o′ ∈ R|o′ > 0} is the set of all moderators’ opinions. To obtain the minimum cost for consensus, the optimal weight ωi are obtained by Model (2). • When the opinions of DMs and the moderator present random characteristics, it is difficult to determine the range of opinion by knowing only the opinion distribution. To determine the mini- mum cost considering random opinions, this paper further generalizes Model (2), and constructs a minimum cost consensus model with random distributed opinions. The basic assumptions of the model are as follows. Hypothesis 1: The opinions of DMs and the moderator are uncertain, and obey a certain distribution; Hypothesis 2: Consensus budget is limited. In a budget range, group consensus reaches under a certain confidence level (i.e. a certain probability); in other words, the probability of group consensus is not below the confidence level; Hypothesis 3: Under the stated certain confidence level, the budget for the reached group consensus has a minimum value. 3. Model construction 3.1. Cost consensus model with single-constraint: cost upper limit Assume that m DMs participate in the GDM process, where oi denotes the DM i (i ∈ M)’s opinion, and o′ is the moderator’s opinion (i.e. o′ can be interpreted as the group consensus opinion). Here ωi denotes the unit cost that the moderator is willing to pay DM i to achieve the optimal consensus opinion. Moreover, if ωi|o′−oi| is used to measure the deviation between the moderator’s opinion and DM i’s opinion, then ωi|o′ − oi| is the cost that the moderator pays DM i. Therefore, m∑ i=1 ωi|o′ − oi| denotes the total cost that the moderator pays all DMs. 6 In realistic decision making (Hypothesis 1), the opinions of DMs and the moderator are uncertain, and present a certain distribution. Since these opinions are difficult to be expressed by a certain value, GDM reaches consensus under a certain confidence level α (Hypothesis 2). Next, on the assumption of a certain confidence level, we study the minimum cost for consensus (Hypothesis 3). Let the probability of the cost paid for consensus no more than budget β be not less than a certain confidence level 1−α (α is called as significance level): Pr{ m∑ i=1 ωi|oi −o′| ≤ β} ≥ 1 −α. To measure the minimum budget for the group consensus reached, we construct a consensus model based on the random distribution under the confidence level: W(α) = Min β s.t. Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α (3) The following result proves that Model (2) is a special case of Model (3). Proposition 1. When α = 0, Model (3) is equivalent to Model (2). Proof. When α = 0, Model (3) can be transformed to a general nonlinear programming model: W(α) = Min β s.t. m∑ i=1 ωi|oi −o′| ≤ β (4) According to Proposition 1, Model (3) (and in turn Model (4)) generalizes Model (2). The following result proves the monotonicity property of W(α). Theorem 1. W(α) is a monotonic non-increasing function of significance level α. Proof. Let ϕ(α1), ϕ(α2) be the feasible region of Model (3) under significance level α1, α2. When α1 ≤ α2, 1 −α1 ≥ 1 −α2, then ϕ(α1) ⊆ ϕ(α2). Thus, W(α1) ≥ W(α2). Since the opinions of DM i and the moderator o′ are uncertain, they are divided into two categories: interval value and probability distribution. Case 1. DMs’ opinions are uncertain, their probability distribution unknown, and the moderator’s opinion obeys a certain distribution. Min β s.t.   Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α oi ∈ [oLi ,o U i ], i ∈ M (5) where the moderator’s opinion o′ obeys a certain distribution, and oi is DM i’s opinion in interval [oLi ,o U i ]. o L i , o U i are the lower and upper bounds of DM i’s opinion, respectively. Replacing 7 [oLi ,o U i ] by its equivalent crisp number o L i + µi(o U i −o L i ), Model (5) can be transformed into the following Model (6): Min β s.t.   Pr{ m∑ i=1 ωi|oLi + µi(o U i −o L i ) −o ′| ≤ β}≥ 1 −α 0 ≤ µi ≤ 1, i ∈ M (6) where o′ obeys a certain distribution. Case 2. DMs’ opinions obey a certain distribution, the moderator’s opinion is uncertain and its distribution is unknown. Min β s.t.   Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α o′ ∈ [o′L,o ′ U ] (7) where DM i’s opinion obeys a certain distribution, and o′ is the moderator’s opinion in interval [o′L,o ′ U ]. o ′ L,o ′ U are the lower and upper bounds of the moderator’s opinion respectively. Replac- ing [o′L,o ′ U ] by its equivalent crisp number o ′ L + µ(o ′ U −o ′ L), Model (7) can be transformed into the following Model (8): Min β s.t.   Pr{ m∑ i=1 ωi|oi − [o′L + µ(o ′ U −o ′ L)]| ≤ β}≥ 1 −α 0 ≤ µ ≤ 1 (8) where oi, i ∈ M obeys a certain distribution. The significance of the objective function in Models (5)–(8) is that whatever β is, consensus can be realized/reached with confidence level of 1 −α. 3.2. Cost consensus model with two-constraints: upper cost limit and lower DM compensation limit From the moderator point of view, the lower the group consensus budget the better, as he/she aims at the budget to not exceed a certain upper limit. However, for the DMs, it is unfavorable to reach group consensus with a low budget, and they hope the compensation is as large as possible because they sacrifice individual interests for the benefit of the group. Similarly, from the perspective of the moderator, the probability of the cost paid for consensus being no more than budget β is set to be not less than a certain confidence level 1 −α1, i.e. Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α1. However, the DMs must be able to obtain compensation for reaching consensus, so let the cost paid for consensus be used to guarantee the lowest reward γ, and the probability of exceeding the lowest reward γ be not less than another confidence level 1 −α2, i.e. Pr{ m∑ i=1 ωi|oi −o′| ≥ γ}≥ 1 −α2. Under the constraints 8 of the above two conditions, to discuss the minimum of β −γ, we construct the following Model (9): Min β −γ s.t.   Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oi −o′| ≥ γ}≥ 1 −α2 (9) As before, two cases are possible: Case 1. DMs’ opinions are uncertain wth unknown probability distribution, and the moderator’s opinion obeys a certain distribution. Min β −γ s.t.   Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oi −o′| ≥ γ}≥ 1 −α2 oi ∈ [oLi ,o U i ], i ∈ M (10) Model (10) can be transformed into the following Model (11): Min β −γ s.t.   Pr{ m∑ i=1 ωi|oLi + µi(o U i −o L i ) −o ′| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oLi + µi(o U i −o L i ) −o ′| ≥ γ}≥ 1 −α2 0 ≤ µi ≤ 1, i ∈ M (11) Case 2. DMs’ opinions obey a certain distribution, the moderator’s opinion is uncertain and its distribution is unknown. Min β −γ s.t.   Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oi −o′| ≥ γ}≥ 1 −α2 o′ ∈ [o′L,o ′ U ] (12) Model (12) can be transformed into the following Model (13): Min β −γ s.t.   Pr{ m∑ i=1 ωi|oi − [o′L + µ(o ′ U −o ′ L)]| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oi − [o′L + µ(o ′ U −o ′ L)]| ≥ γ}≥ 1 −α2 0 ≤ µ ≤ 1 (13) Thus, we can similarly discuss the minimum of β −γ, and the change in β −γ with the change in α. 9 3.3. Cost consensus model with three-constraints: cost upper limit, lower DM compensations limit, and opinion deviation between DMs and the moderator In GDM, the role of the moderator is similar to a leader’s role in an organization, where he/she has good communication and negotiation skills, and leads the development trend of the entire group decision to approach the moderator’s opinion. In some cases, the moderator needs to guide (control) some DMs, and to guarantee that the opinion deviation between these DMs and himself/herself is not too large, i.e. the opinions deviation is within a certain range (|os −o′| ≤ ζ,s ∈ S ⊆ M). Based on Subsection 3.2., we construct the following consensus model (Model (14)) based on the random distribution subject to the above opinion constraint: Min β −γ s.t.   Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oi −o′| ≥ γ}≥ 1 −α2 |os −o′| ≤ ζ,s ∈ S ⊆ M (14) Again, two cases are possible: Case 1. DMs’ opinions are uncertain, probability distribution is unknown, and the moderator’s opin- ion obeys a certain distribution. Min β −γ s.t.   Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oi −o′| ≥ γ}≥ 1 −α2 |os −o′| ≤ ζ,s ∈ S ⊆ M oi ∈ [oLi ,o U i ], i ∈ M (15) Model (15) can be transformed into Model (16): Min β −γ s.t.   Pr{ m∑ i=1 ωi|oLi + µi(o U i −o L i ) −o ′| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oLi + µi(o U i −o L i ) −o ′| ≥ γ}≥ 1 −α2 |os −o′| ≤ ζ,s ∈ S ⊆ M 0 ≤ µi ≤ 1, i ∈ M (16) Case 2. DMs’ opinions obey a certain distribution, the moderator’s opinion is uncertain and its 10 distribution is unknown. Min β −γ s.t.   Pr{ m∑ i=1 ωi|oi −o′| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oi −o′| ≥ γ}≥ 1 −α2 |os −o′| ≤ ζ,s ∈ S ⊆ M o′ ∈ [o′L,o ′ U ] (17) Model (17) can be transformed into Model (18): Min β −γ s.t.   Pr{ m∑ i=1 ωi|oi − [o′L + µ(o ′ U −o ′ L)]| ≤ β}≥ 1 −α1 Pr{ m∑ i=1 ωi|oi − [o′L + µ(o ′ U −o ′ L)]| ≥ γ}≥ 1 −α2 |os −o′| ≤ ζ,s ∈ S ⊆ M 0 ≤ µ ≤ 1 (18) Again, we can discuss the minimum of β −γ, and the change in β −γ with the change in α. 4. Application to the minimum cost consensus models based on China’s urban demolition negotiation In urban relocation negotiations, the government plays the role of moderator while urban residents are DMs. Naturally, the government needs to negotiate with residents regarding the compensation price. Consensus is reached only when all residents’ expected and the government’s compensation price reach an agreement, i.e. the expected utilities of all residents’ expected compensation are largely satisfied. Taking the background of China’s urban demolition negotiation (Gong et al., 2017), we assume the demolition negotiation involves four residents d1, d2, d3, d4, and their expected compen- sation price intervals are o1 ∈ [48, 52], o2 ∈ [50, 55], o3 ∈ [60, 65], and o4 ∈ [62, 67]. Meanwhile, the government’s compensation price is o′ ∈ [50, 65] (unit: ten thousands RMB). To reach the final relocation consensus, the unit negotiation costs that the government pays residents are ω1 = 0.8, ω2 = 1.5, ω3 = 1.2, and ω4 = 0.9 (unit: thousand RMB). In the following, we discuss situations for demolition consensus considering interval-valued and random prices. 4.1. Cost consensus model with single-constraint: cost upper limit Case 1. oi ∈ [oLi ,o U i ] with unknown distribution, and o ′ obeys a certain distribution. According to the 3σ principle of normally distributed variables (Chen et al., 2016), it is known that a random normally distributed variable X ∼ N(µ,σ2) verifies P(µ−3σ ≤ X ≤ µ+ 3σ) ≈ 1. Thus, interval values can be linked to a random normally distributed variable using the relation [a−,a+] = [µ− 3σ,µ + 3σ],µ = a− + a+ 2 ,σ = a+ −a− 6 (19) 11 Using (19), we obtain the distribution of the government’s compensation price o′ ∼ N(57.5, 2.52), i.e. the government’s expected compensation price is 57.5 with variance 2.52. Setting the proba- bility of the cost paid for consensus no more than the budget not less than 0.9, the optimization consensus model (Model (6)) based on the minimum cost becomes: Min β s.t.   Pr{0.8 ∗ |(48 + 4µ1) −o′| + 1.5 ∗ |(50 + 5µ2) −o′|+ 1.2 ∗ |(60 + 5µ3) −o′| + 0.9 ∗ |(62 + 5µ4) −o′| ≤ β}≥ 0.9 o′ ∼ N(57.5, 2.52) 0 ≤ µi ≤ 1, i = 1, 2, 3, 4 Case 2. oi obeys a certain distribution; o ′ ∈ [o′L,o ′ U ] with unknown distribution. Applying again (19), the four residents’ expected compensation price distributions are o1 ∼ N(50, 0.66672), o2 ∼ N(52.5, 0.83332), o3 ∼ N(62.5, 0.83332), and o4 ∼ N(64.5, 0.83332). Similarly, the optimization consensus model (Model (8)) based on the minimum cost becomes: Min β s.t.   Pr{0.8 ∗ |(50 + 15µ) −o1| + 1.5 ∗ |(50 + 15µ) −o2|+ 1.2 ∗ |(50 + 15µ) −o3| + 0.9 ∗ |(50 + 15µ) −o4| ≤ β}≥ 0.9 o1 ∼ N(50, 0.66672) o2 ∼ N(52.5, 0.83332) o3 ∼ N(62.5, 0.83332) o4 ∼ N(64.5, 0.83332) 0 ≤ µ ≤ 1 The solutions of these optimization consensus models are shown in Table 1. The final relocation consensus has a probability of 90% = 1−10% with the conditions of Case 1 (when Min β = 19.0451) and Case 2 (when Min β = 27.1073). Table 1: Results of the cost consensus models with single-constraint o1 o2 o3 o4 o ′ β Case 1 51.9664 54.3240 61.4760 62.1320 - 19.0451 Case 2 - - - - 53.2295 27.1073 Moreover to discuss the influence of the budget with different distributions, we consider the normal distribution, and the uniform distribution. The results are shown in Fig. 1. Case 1 is represented in the top figure: on the x-axis 1 represents the moderator’s opinion obeys the normal distribution, 12 while 2 represents the moderator’s opinion obeys the uniform distribution. Case 2 is represented in the botton figure: 1 on the x-axis 1 represents the DMs’ opinions obey the normal distribution, while 2 represents the DMs’ opinions obey the uniform distribution. 0 0.5 1 1.5 2 2.5 3 15 20 25 30 1−1 Different distributions of the moderator’s opinion 1−N(57.5,2.5 2 ); 2−U(50,65) M in β 0 0.5 1 1.5 2 2.5 3 25 26 27 28 29 30 1−2 Different distributions of the DMs’ opinion 1−N(50,0.6667 2 ),N(52.5,0.8333 2 ),N(62.5,0.8333 2 ),N(64.5,0.8333 2 ); 2−U(48,52),U(50,55),U(60,65),U(62,67) M in β Figure 1: Different opinion distributions with single-constraint. Taking Case 1 as an example, the change in budget β is shown in Fig. 2 with the change in the confidence level, 1 −α, and the distributions of the moderator’s opinion. The confidence level 1 −α values shown are 1, 0.9, 0.8, 0.7, 0.6, and 0.5; and the distributions the moderator’s opinion being N(57.5, 2.52) and U(50, 65). It is clear that the minimum budget decreases as the significance level α increases, i.e. the confidence level decreases. 0 0.1 0.2 0.3 0.4 0.5 10 15 20 25 30 35 The change in Min β with the change in α for two distributions of the moderator’s opinion α M in β Min N(57.5,2.5 2 ) Min U(50,65) Figure 2: The change in Min β with the change in α for three distributions of the moderator’s opinion 13 4.2. The cost consensus model with two-constraints: the upper limit of the cost and the lower limit of DMs’ compensations Case 1. oi ∈ [oLi ,o U i ], the distribution of oi is unknown; and o ′ obeys a certain distribution. In urban relocation negotiations, we assume the conditions are the same as in Case 1 in Section 4.1. We not only let the probability of the cost paid for consensus no more than budget be not less than 0.9, but also let the probability of the cost paid for consensus exceeding the lowest reward be not less than 0.9. The optimization consensus model (Model (11)) based on the minimum cost is: Min β −γ s.t.   Pr{0.8 ∗ |(48 + 4µ1) −o′| + 1.5 ∗ |(50 + 5µ2) −o′|+ 1.2 ∗ |(60 + 5µ3) −o′| + 0.9 ∗ |(62 + 5µ4) −o′| ≤ β}≥ 0.9 Pr{0.8 ∗ |(48 + 4µ1) −o′| + 1.5 ∗ |(50 + 5µ2) −o′|+ 1.2 ∗ |(60 + 5µ3) −o′| + 0.9 ∗ |(62 + 5µ4) −o′| ≥ γ}≥ 0.9 o′ ∼ N(57.5, 2.52) 0 ≤ µi ≤ 1, i = 1, 2, 3, 4 Case 2. oi obeys a certain distribution; o ′ ∈ [o′L,o ′ U ], and the distribution of o ′ is unknown. We assume the conditions are the same as in Case 2 in Section 4.1. We not only let the probability of the cost paid for consensus no more than budget be not less than 0.9, but also let the probability of the cost paid for consensus exceeding the lowest reward be not less than 0.9. The optimization consensus model (Model (13)) based on the minimum cost is: Min β −γ s.t.   Pr{0.8 ∗ |(50 + 15µ) −o1| + 1.5 ∗ |(50 + 15µ) −o2|+ 1.2 ∗ |(50 + 15µ) −o3| + 0.9 ∗ |(50 + 15µ) −o4| ≤ β}≥ 0.9 Pr{0.8 ∗ |(50 + 15µ) −o1| + 1.5 ∗ |(50 + 15µ) −o2|+ 1.2 ∗ |(50 + 15µ) −o3| + 0.9 ∗ |(50 + 15µ) −o4| ≥ γ}≥ 0.9 o1 ∼ N(50, 0.66672) o2 ∼ N(52.5, 0.83332) o3 ∼ N(62.5, 0.83332) o4 ∼ N(64.5, 0.83332) 0 ≤ µ ≤ 1 The solutions of these optimization consensus models are shown in Table 2. The final relocation consensus has a probability of 90% = 1−10% with the conditions of Case 1 (when Min β−γ = 1.2002) and Case 2 (when Min β−γ = 3.6950). Similarly, Fig. 3 shows results that help to discuss the influence of the minimum budget with different distributions, and it is to read similarly as Fig. 1 above. 14 Table 2: Results of the cost consensus models with two-constraints o1 o2 o3 o4 o ′ β −γ Case 1 51.9972 52.9920 64.3750 66.1140 - 1.2002 Case 2 - - - - 52.4690 3.6950 0 0.5 1 1.5 2 2.5 3 0 1 2 3 3−1 Different distributions of the moderator’s opinion 1−N(57.5,2.5 2 ); 2−U(50,65) M in β − γ 0 0.5 1 1.5 2 2.5 3 0 5 10 3−2 Different distributions of the DMs’ opinion 1−N(50,0.6667 2 ),N(52.5,0.8333 2 ),N(62.5,0.8333 2 ),N(64.5,0.8333 2 ); 2−U(48,52),U(50,55),U(60,65),U(62,67) M in β − γ Figure 3: Different opinion distributions with two-constraints 4.3. Cost consensus model with three-constraints: cost upper limit, lower DM compensations limit, and opinion deviation between DMs and the moderator Case 1. oi ∈ [oLi ,o U i ] with unknown distribution, and o ′ obeys a certain distribution. In urban relocation negotiations, we assume the same conditions as in Case 1 in Subsection 4.1. We add the deviation constraints between four residents’ expected compensation prices and the government’s compensation price, and the optimization consensus model (Model (16)) based on the minimum cost is: Min β −γ s.t.   Pr{0.8 ∗ |(48 + 4µ1) −o′| + 1.5 ∗ |(50 + 5µ2) −o′|+ 1.2 ∗ |(60 + 5µ3) −o′| + 0.9 ∗ |(62 + 5µ4) −o′| ≤ β}≥ 0.9 Pr{0.8 ∗ |(48 + 4µ1) −o′| + 1.5 ∗ |(50 + 5µ2) −o′|+ 1.2 ∗ |(60 + 5µ3) −o′| + 0.9 ∗ |(62 + 5µ4) −o′| ≥ γ}≥ 0.9 Min{|48 + 4µ1 −o′|, |50 + 5µ2 −o′|, |60 + 5µ3 −o′|, |62 + 5µ4 −o′|}≤ 10 o′ ∼ N(57.5, 2.52) 0 ≤ µi ≤ 1, i = 1, 2, 3, 4 Case 2. oi obeys a certain distribution; o ′ ∈ [o′L,o ′ U ] with unknown distribution. With the same conditions of Case 2 in Subsection 4.1, and the addition of the deviation constraints between 15 four residents’ expected compensation prices and the government’s compensation price, the optimization consensus model (Model (18)) based on the minimum cost is: Min β −γ s.t.   Pr{0.8 ∗ |(50 + 15µ) −o1| + 1.5 ∗ |(50 + 15µ) −o2|+ 1.2 ∗ |(50 + 15µ) −o3| + 0.9 ∗ |(50 + 15µ) −o4| ≤ β}≥ 0.9 Pr{0.8 ∗ |(50 + 15µ) −o1| + 1.5 ∗ |(50 + 15µ) −o2|+ 1.2 ∗ |(50 + 15µ) −o3| + 0.9 ∗ |(50 + 15µ) −o4| ≥ γ}≥ 0.9 Min{|(50 + 15µ) −o1|, |50 + 15µ) −o2|, |(50 + 15µ) −o3|, |(50 + 15µ) −o4|}≤ 15 o1 ∼ N(50, 0.66672) o2 ∼ N(52.5, 0.83332) o3 ∼ N(62.5, 0.83332) o4 ∼ N(64.5, 0.83332) 0 ≤ µ ≤ 1 The solutions of these optimization consensus models are shown in Table 3. The final relocation consensus has a probability of 90% = 1−10% with the conditions of Case 1 (when Min β−γ = 1.5111) and Case 2 (when Min β − γ = 3.6978). Similarly, Fig. 4 helps in discussing the influence of the minimum budget with different distributions. Table 3: Results of the cost consensus models with three-constraints o1 o2 o3 o4 o ′ β −γ Case 1 50.9672 52.3935 64.9995 65.7540 - 1.5111 Case 2 - - - - 52.4420 3.6978 4.4. The effectiveness of the proposed probabilistic programming method based on GA To demonstrate the effectiveness of the proposed probabilistic programming method based on genetic algorithms (GA) (details in Appendix I), we compute Gong et al. (2017) model by our algorithm as shown below in (20), with the iteration results shown in Fig. 5. Min 0.8 ∗ |(48 + 4µ1) − (50 + 15µ5)| + 1.5 ∗ |(50 + 5µ2) − (50 + 15µ5)|+ 1.2 ∗ |(60 + 5µ3) − (50 + 15µ5)| + 0.9 ∗ |(62 + 5µ4) − (50 + 15µ5)| s.t. 0 ≤ µi ≤ 1, i = 1, 2, 3, 4, 5 (20) Compared with the optimal minimum negotiation budget of 14.7 obtained in (Gong et al., 2017), it is clear that our iteration results produce an optimal result close to the above one after 200 iter- ations. Apart from interval values, our algorithm can also deal with fuzzy numbers obeyed random 16 0 0.5 1 1.5 2 2.5 3 0 1 2 3 4−1 Different distributions of the moderator’s opinion 1−N(57.5,2.5 2 ); 2−U(50,65) M in β − γ 0 0.5 1 1.5 2 2.5 3 0 5 10 4−2 Different distributions of the DMs’ opinion 1−N(50,0.6667 2 ),N(52.5,0.8333 2 ),N(62.5,0.8333 2 ),N(64.5,0.8333 2 ); 2−U(48,52),U(50,55),U(60,65),U(62,67) M in β − γ Figure 4: Different opinion distributions with three-constraints 0 100 200 300 400 14 15 16 17 18 the number of generations Figure 5: The iteration results of Gong’s model computed with proposed algorithm 17 distributions. The detail steps of the proposed probabilistic programming method based on GA are as follows: Step 1: randomly generate initial population between all variables’ stated ranges based on GA; Step 2,: randomly generate some variables obeyed a certain distribution; Step 3: combined initial population with random variables, evaluate whether all generated variables satisfy the constraints. If yes, then go Step 5; if no, then go Step 4; Step 4: through selective reproduction, crossing over and mutation, go into generation cycle and go Step 2; Step 5: collect all feasible populations and obtain the optimal solution. 5. Conclusions Uncertain opinions often exist in GDM and obstruct the consensus process. Except for interval values, we assume that the opinions of DMs and the moderator obey a certain distribution, and con- struct the cost consensus model based on random distribution with a probability of budget constraints. The objective function of the stochastic optimization model is to obtain the minimum of the total budget under a certain confidence level. The constraints of the model include: the probability of the cost paid for consensus no more than the budget is not less than a certain confidence level; the probability of the cost paid for consensus, which exceeds the lowest reward, is not less than another confidence level; and the qualified opinion deviations between the moderator and some DMs do not exceed a certain threshold. Additionally, for the China’s urban demolition negotiation application, we design the minimum cost consensus models and introduce GA to solve the corresponding stochastic optimization models. The algorithm not only has better convergence, but also can better simulate the consensus GDM process, providing a satisfactory solution to the stochastic optimization consensus model. The main innovative contributions of this article can be summarized as follows: • The cost consensus model based on random opinion distribution is not only the generalization of the minimum cost consensus model by Ben-Arieh and Easton (2007), but also the further improvement of the consensus model with uncertain opinions by Gong et al. (2015b). • The opinions of DMs and the moderator cab be assumed to obey different distributions, such as normal distribution, uniform distribution, Chi-square distribution, etc, which improves the expression forms of uncertain opinions and are more relevant to reality. • Probabilistic planning based on GA is used to resolve the model, which has the advantages of good convergence, high precision, and saving computation time. 18 • In the case analysis, the sensitivity analysis method is adopted to obtain the lower bounds of the budget, and the minimum budget is obtained under different constraints or various opinions distributions. Appendix I Probabilistic programming (Kushmerick et al., 1995) is used to deal with goal programming prob- lems with probabilistic constraints. In this article, we explore a new probabilistic programming method based on GA, which combines probabilistic programming with goal programming problems. GA is a random search method referenced in biological evolution with regularity (survival of the fittest) and first proposed by Holland (1992). Its main characteristics are as follows: it operates directly on the structure objects, and there is no limit of derivation and function continuity; it has conceal concur- rence and splendid capability with the better situation as a whole; it uses probability optimization methods to automatically obtain and optimize the search space, adaptively adjust the search direc- tion, but does not need to determine the rules. With these properties, GA has been widely used in combinatorial optimization, machine learning, signal processing, adaptive control, and artificial life. It is a key technique in modern intelligent computing (the basic operation of GA is shown in Fig. A1). Figure A1: GA To obtain the minimum of the objective function, it is feasible to adopt goal programming. As for the probability distribution existing in constraints or the objective function, it can be calculated after a certain amount of processing. The article designs a probabilistic planning method based on GA, which effectively resolves the proposed model (details in Fig. A2) 19 Figure A2: Probabilistic planning based on GA References Altuzarra, A., Moreno-Jiménez, J. M., and Sandalvador, M. (2007). A Bayesian priorization procedure for AHP-group decision making. European Journal of Operational Research, 182(1), 367–382. Bai, J., Yang, J., Zhou, Y., and Yang, Q. (2015). A multi-step rolled forward chance-constrained model and a proactive dynamic approach for the wheat crop quality control problem. European Journal of Operational Research, 246(2), 631–640. Ben-Arieh, D., and Easton, T. (2007). Multi-criteria group consensus under linear cost opinion elas- ticity. Decision Support Systems, 43(3), 713–721. Ben-Arieh, D., Easton, T., and Evans, B. (2009). Minimum cost consensus with quadratic cost functions. IEEE Transactions on Systems Man and Cybernetics Part A: Systems and and Humans, 39(1), 210–217. Cabrerizo, F. J., Chiclana, F., Al-Hmouz, R., Morfeq, A., Balamash, A. S. and Herrera-Viedma, E. (2015) Fuzzy decision making and consensus: challenges. Journal of Intelligent & Fuzzy Systems, 29(3): 1109–1118. Cid-Lopez, A., Hornos, M. J., Carrasco, R. A., Herrera-Viedma, E. and Chiclana, F. (2017) Linguistic multi-criteria decision-making model with output variable expressive richness. Expert Systems with Applications, 83, 350–362. Charnes, A., and Cooper, W. W. (1959). Chance-constrained programming. Management Science, 6(1), 73–79. 20 Chen, M., Wang, S. G., Wang, P. P., and Ye, X. (2016). A new equivalent transformation for interval inequality constraints of interval linear programming. Fuzzy Optimization and Decision Making, 15(2), 155–175. Chiclana, F., Herrera, F., and Herrera-Viedma, E. (1998). Integrating three representation models in fuzzy multipurpose decision making based on fuzzy preference relations. Fuzzy Sets and Systems, 97, (1) 3–48. Chiclana, F., Herrera, F., and Herrera-Viedma, E. (2001). Integrating multiplicative preference rela- tions in a multipurpose decision-making model based on fuzzy preference relations. Fuzzy Sets and Systems, 122(2), 277–291. Chiclana, F., Tapia Garćıa, J. M., del Moral, M. J., and Herrera-Viedma, E. (2013). A Statisti- cal Comparative Study of Different Similarity Measures of Consensus in Group Decision Making. Information Sciences, 221, 110–123. Dai, C., Cai, Y. P., Li, Y. P., Sun, W., Wang, X. W., and Guo, H. C. (2014). Optimal strategies for carbon capture, utilization and storage based on an inexact mλ-measure fuzzy chance-constrained programming. Energy, 78, 465–478. Dong, Q., Zhü, K., and Cooper, O. (2017). Gaining consensus in a moderated group: a model with a twofold feedback mechanism. Expert Systems with Applications, 71, 87–97. Dong, Q. X., and Cooper, O. (2016). A peer-to-peer dynamic adaptive consensus reaching model for the group AHP decision making. Ssrn Electronic Journal, 250(2), 521-530. Dong, Y. C., Luo, N., and Liang, H. (2015). Consensus building in multiperson decision making with heterogeneous preference representation structures: A perspective based on prospect theory. Applied Soft Computing, 35, 898–910. Dong, Y., Li, C., Chiclana, F., and Herrera-Viedma, E. (2016). Average-case consistency measurement and analysis of interval-valued reciprocal preference relations. Knowledge-Based Systems, 114, 108– 117. Fishburn, P.C. (1979) Utility Theory for Decision Making, Robert E. Krieger Publishing Company. Fodor, J.C. and Roubens, M. (1994) Fuzzy Preference Modelling and Multicriteria Decision Support, Kluwer, Dordrecht. Forman, E., and Peniwati, K. (1998). Aggregating individual judgments and priorities with the analytic hierarchy process. European Journal of Operational Research, 108(1), 165–169. 21 Fu, C., and Yang, S. L. (2010). The group consensus based evidential reasoning approach for multiple attributive group decision analysis. European Journal of Operational Research, 206(3), 601–608. Gao, Y., and Qin, Z. (2016). A chance constrained programming approach for uncertain p-hub center location problem. Computers and Industrial Engineering, 102, 10–20. Gong, Z. W., Zhang, H. H., Jeffrey, F., Li, L. S., and Xu, X. X. (2015a). Two consensus models based on the minimum cost and maximum return regarding either all individuals or one individual. European Journal of Operational Research, 240(1), 183–192. Gong, Z. W., Xu, X. X., Zhang, H. H., Ozturk, U. A., Herrera-Viedma, E., and Xu, C. (2015b). The consensus models with interval preference opinions and their economic interpretation. Omega, 55, 81–90. Gong, Z. W., Xu, C., Chiclana, F., and Xu, X. X. (2017). Consensus measure with multi-stage fluctuation utility based on China’s urban demolition negotiation. Group Decision and Negotiation, 26(2), 379–407. He, W., Chen, G., Han, Q. L., and Qian, F. (2017). Network-based leader-following consensus of nonlinear multi-agent systems via distributed impulsive control. Information Sciences, 380, 145– 158. Holland, J. H. (1992). Adaptation in natural and artificial systems. MIT Press. Honert, R. C. V. D. (1998). Stochastic group preference modelling in the multiplicative AHP: a model of group consensus. European Journal of Operational Research, 110(1), 99–111. Hong, L. J., Yang, Y., and Zhang, L. (2011). Sequential convex approximations to joint chance constrained programs: a Monte Carlo approach. Operations Research, 59(3), 617–630. Jiang, Y., Xu, Z., and Yu, X. (2013). Compatibility measures and consensus models for group decision making with intuitionistic multiplicative preference relations. Applied Soft Computing, 13(4), 2075– 2086. Kushmerick, N., Hanks, S., and Weld, D. S. (1995). An algorithm for probabilistic planning. Artificial Intelligence, 76(1–2), 239–286. Li, C. C., Dong, Y., Herrera, F., Herrera-Viedma, E., and Mart́ınez, L. (2017). Personalized individual semantics in computing with words for supporting linguistic group decision making. An application on consensus reaching. Information Fusion, 33(C), 29-40. Lin, C. S., and Kou, G. (2015). Bayesian revision of the individual pair-wise comparison matrices under consensus in AHP–GDM. Applied Soft Computing, 35(C), 802-811. 22 Nurmi, H. (1981) Approaches to collective decision making with fuzzy preference relations, Fuzzy Sets Syst., vol. 6, pp. 249–259. Qin, Z. (2016). Random fuzzy mean-absolute deviation models for portfolio optimization problem with hybrid uncertainty. Applied Soft Computing, in press. Regan, H. M., Colyvan, M., and Markovchicknicholls, L. (2006). A formal model for consensus and negotiation in environmental management. Journal of Environmental Management, 80(2), 167–176. Saaty, T. L. (1980). The Analytic Hierarchy Process, McGraw-Hill, New York. Sadati, M. E. H., and Nematian, J. (2013). Two-level linear programming for fuzzy random portfolio optimization through possibility and necessity-based model. Procedia Economics and Finance, 5(13), 657–666. Sakawa, M., and Matsui, T. (2013). Interactive fuzzy random two-level linear programming based on level sets and fractile criterion optimization. Information Sciences, 238(7), 163–175. Srdjevic, B., and Srdjevic, Z. (2013). Synthesis of individual best local priority vectors in AHP-group decision making. Applied Soft Computing, 13(4), 2045–2056. Srdjevic, B., Srdjevic, Z., Blagojevic, B., and Suvocarev, K. (2013). A two-phase algorithm for consensus building in AHP-group decision making. Applied Mathematical Modelling, 37(10–11), 6670–6682. Ureña, R., Chiclana, F., Fujita, H., and Herrera-Viedma, E. (2015). Confidence-consistency driven group decision making approach with incomplete reciprocal intuitionistic preference relations. Knowledge-Based Systems, 89, 86–96. Ureña, R., Cabrerizo, F. J., Morente-Molinera, J. A., and Herrera-Viedma, E. (2016). GDM-R: a new framework in R to support fuzzy group decision making processes. Information Sciences, 357, 161–181. Von Winterfeldt, D., and Edwards, W. (1986). Decision analysis and behavioral research. Journal of the Mount Sinai Hospital New York, 18(1), 4–14. Wang, X., Li, Y., Ma, Z., Fan, W., Wang, L., and Xu, M. (2017). Robust optimization of structural- acoustic coupled system with random parameters. Aerospace Science and Technology, 60, 48–57. Wang, Z. J., and Li, K. W. (2016). Group decision making with incomplete intuitionistic preference relations based on quadratic programming models. Computers and Industrial Engineering, 93(C), 162–170. 23 J. Wu and F. Chiclana. (2014a) Visual information feedback mechanism and attitudinal prioritisa- tion method for group decision making with triangular fuzzy complementary preference relations. Information Sciences, 279, 716–734. J. Wu and F. Chiclana. (2014b) A risk attitudinal ranking method for interval-valued intuitionistic fuzzy numbers based on novel score and accuracy expected functions. Applied Soft Computing, 22, 772–786. Wu, J., and Chiclana, F. (2014c). Multiplicative consistency of intuitionistic reciprocal preference relations and its application to missing values estimation and consensus building. Knowledge-Based Systems, 71, 187–200. Wu, J., Chiclana, F. and Herrera-Viedma, E. (2015) Trust Based Consensus Model for Social Network in an Incomplete Linguistic Information Context. Applied Soft Computing, 35, 827–839. Wu, Z., and Xu, J. (2012). A consistency and consensus based decision support model for group decision making with multiplicative preference relations. Decision Support Systems, 52(3), 757–767. Wu, Z. B., and Xu, J. (2016). Managing consistency and consensus in group decision making with hesitant fuzzy linguistic preference relations. Omega, 65(3), 28–40. Wu, Y. N., Xu, H., Xu, C., and Chen, K. (2016). Uncertain multi-attributes decision making method based on interval number with probability distribution weighted operators and stochastic dominance degree. Knowledge-Based Systems, 113, 199–209. Xu, Z. S., and Liao, H. (2015). A survey of approaches to decision making with intuitionistic fuzzy preference relations. Knowledge-Based Systems, 80(C), 131–142. Xu, X. H., Zhong, X. Y., Chen, X. H., and Zhou, Y. J. (2015). A dynamical consensus method based on exit-delegation mechanism for large group emergency decision making. Knowledge-Based Systems, 86(C), 237–249. Zadeh,L. A. (1975) The concept of a linguistic variable and its applications to approximate reasoning. Part I. Information Sciences, 8:199–249. Zhang, B., Dong, Y., and Xu, Y. (2013). Maximum expert consensus models with linear cost function and aggregation operators. Computers and Industrial Engineering, 66(1), 147–157. Zhang, F., Ignatius, J., Zhao, Y., Lim, C. P., Ghasemi, M., and Ng, P. S. (2015). An improved consensus-based group decision making model with heterogeneous information. Applied Soft Com- puting, 35, 850–863. Zhao, L., Jia, Y., Yu, J., and Du, J. (2017). H ∞ sliding mode based scaled consensus control for linear multi-agent systems with disturbances. Applied Mathematics and Computation, 292, 375–389. 24 
work_nn2jkt2gw5cp5bqwgxsf7va2gu	----	An expert system for real-time traffic management in wireless local area networks An expert system for real-time traffic management in wireless local area networks T. Frantti a,⇑, M. Majanen b a Renesas Mobile Europe Ltd., Elektroniikkatie 10, FI-90570 Oulu, Finland b VTT Technical Research Centre of Finland, Finland a r t i c l e i n f o Keywords: Expert system Fuzzy control Flow control Offloading Real-time traffic WLAN a b s t r a c t The paper explores delay-based congestion and flow control and the offloading of real-time traffic from wireless local area networks (WLANs) to mobile cellular networks (MCNs) in multihomed devices. The control system developed is based on an embedded hierarchical expert system. It adjusts transceivers’ traffic flow(s) for prevailing network conditions to achieve application-dependent delay and throughput limits. In wireless networks, delay and throughput depend on the packet size, packet transmission inter- val, and node connection density. Therefore, the controller on the destination node monitors average one-way delay and the change of one-way delay of the incoming traffic. On this basis, it adjusts the packet size and transmission interval of the source node by transmitting a control command to the source. If the prevailing level of traffic in the network exceeds its capacity despite of the control actions taken, devices prepare for developed asynchronous offloading of traffic to another access network. The control model was validated via simulation of Voice over Internet Protocol (VoIP) traffic in the OMNeT++ network simulator. The results demonstrate that the expert system developed is able to reg- ulate packet sizes to match the prevailing application-dependent optimum and transfer traffic to another network if the network exceed its capacity no matter the control actions taken. Although this work is motivated mainly by issues of congestion and flow control of WLAN systems and the simulations and results were prepared for the IEEE 802.11b system, the approach and techniques are not limited to these systems, but they are applicable for other packet switched access networks (PSANs), too. � 2014 Elsevier Ltd. All rights reserved. 1. Introduction With the everyday use of cellular networks, extensive mobility during communication became self-evident reality for end users. A mobility requirement soon appeared also for Internet-based com- munication, and currently mobile phone networks are moving with Long Term Evolution (LTE) toward an Internet Protocol (IP) infrastructure with the ultimate aim of offering almost all the ser- vices provided by current second- and third-generation (2G/3G) cellular networks through an all-IP network (see ITU-T (2004a) and ITU-T (2004b) for more details), although Transmission Con- trol Protocol/Internet Protocol (TCP/IP) networks were not origi- nally designed for mobile use. For example, devices on the Internet are identified by an IP address, which has a dual role: it serves as an identifier and as a locator of the networked device. If a device changes its point of attachment, such as wireless access network, its IP address too may change, which means that other devices in the network need means to access the new address if they are to reach the device at its new location. The change in IP address causes the upper-layer protocol (at the layer above the network layer) connection to break, which is problematic for appli- cations with more persistent connections or applications requiring registration of an IP address. Today, mobility in IP networks relies on wireless access technol- ogies, and the trend is toward equipping mobile devices, such as smartphones, with multiple network interfaces. The various inter- faces are, in the general case, operated by different Internet service providers (ISPs), for devices’ improved resilience via an opportu- nity to connect to the Internet through at least one of the access technologies. For end users, these technologies vary mainly in their coverage area and performance. A networked mobile device can use the available interfaces either one at a time or several simultaneously. The device using its multiple network interfaces simultaneously is a special case of node multihoming. A multihomed device has multiple IP ad- dresses, assigned to the same network interface or different ones. http://dx.doi.org/10.1016/j.eswa.2014.01.024 0957-4174/� 2014 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. Tel.: +358 40 547 0819. E-mail addresses: tfrantti@ee.oulu.fi (T. Frantti), mikko.majanen@vtt.fi (M. Majanen). Expert Systems with Applications 41 (2014) 4996–5008 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.01.024&domain=pdf http://dx.doi.org/10.1016/j.eswa.2014.01.024 mailto:tfrantti@ee.oulu.fi mailto:mikko.majanen@vtt.fi http://dx.doi.org/10.1016/j.eswa.2014.01.024 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Node multihoming differs from site multihoming, which can be de- fined as an edge network configuration that has more than one ser- vice provider but does not provide transit communication between them; see the work of Baker (2011). In site multihoming, end users do not need to manage their devices in any way. When a net- worked device is multihomed, there are multiple paths between the source and destination devices. The paths may differ in price, data rate, latency, jitter, and packet loss. Node multihoming, with parallel connections, enables new kinds of traffic engineering tech- niques, whilst site multihoming is intended mainly to increase the reliability of the Internet connection. In this paper, we explore delay-based congestion and flow con- trol and offloading of real-time traffic from wireless local area net- works (WLANs) to cellular networks in multihomed mobile devices. In WLANs, delay and throughput depend on the packet size, packet transmission interval, and connection density. There- fore, we developed and applied control systems to adjust trans- ceivers’ packet sizes for prevailing network conditions to achieve application-dependent delay and throughput limits for real-time traffic and to avoid unnecessary offloading. If the prevailing level of traffic in networks exceeds capacity regardless of the control ac- tions, devices prepare to perform asynchronous offloading of traffic to another access network. In our work, the research assumptions can now be presented as: ‘‘Multihomed and multiradio (WLAN and cellular) devices operating in WLAN and running real-time applica- tions try to maximize network capacity by packet size optimization. When the performance threshold is reached, connections are asyn- chronously offloaded to cellular networks.’’ By real-time traffic we mean Voice over Internet Protocol (VoIP), video calls, and interac- tive games. With the assumptions above, the research question for the work can be formulated as follows: ‘‘When should we begin to transfer traffic of real-time applications to another (cellular) network as the WLAN’s traffic and number of users increase, and how can we do that to avoid unnecessary offloading?’’ The control model was validated through simulation of VoIP traffic in the OMNeT++ network simulator. Even if this work is ad- dressed mainly to congestion and flow control of WLAN systems, the approach and the techniques are not limited to these systems. They are applicable also for other packet switched access networks (PSANs). The rest of the paper is organized as follows. Section 2 presents a review of the literature on packet size control in WLANs, multih- omed path management, and network selection. Section 3 focuses on multihome- and offloading-related standardization activities. Section 4 summarizes the hierarchical decision making system developed, which features a packet size control unit and real-time offloadig of traffic from WLANs to mobile cellular networks unit. Section 5 describes the simulation model, and Section 6 presents the simulation results. Section 7 delineates our future research. Fi- nally conclusions are drawn in Section 8. 2. Literature review 2.1. Packet size optimization in WLANs Korhonen and Wang (2005) have studied the effect of packet size on loss rate and delay in an IEEE 802.11-based WLAN. The analysis shows that there is a straightforward connection between packet size, bit error characteristics, and observed delay character- istics. In general, it is evident throughout the literature that the performance of wireless networking is sensitive to packet size and that significant performance improvements are obtained if a ‘‘good’’ packet size is used. For example, Bakshi, Krishna, Vaidya, and Pradhan (1997) show this for TCP traffic over a wireless net- work. Chee and David (1989), Lettieri and Srivastava (1998), and Chien et al. (1999) all have studied the relationship between frame length and throughput, but they do not propose any precise meth- od for dynamic control of frame length to maximize throughput. Smadi and Szabados (2006) focus on optimization of packet size in error recovery but do not consider the optimal packet size for performance optimization. Sheu, Lee, Chen, Yu, and Huang (2000) present a fuzzy packet length controller (PLFC) for improving the performance of WLANs suffering from interference from a micro- wave oven. It was demonstrated that the PLFC improves the throughput of User Datagram Protocol (UDP) traffic from that with fixed-length packets, but the authors did not consider performance improvements when the number of users and the amount of traffic are increased. Sankarasubramaniam, Akyildiz, and McLaughlin (2003) have studied packet size optimization for energy efficiency, and Younis, Farrag, and D’Amico (2009) consider it for security and throughput, but their solutions are statistical in nature, meaning that the packet size is optimized beforehand. In the most recent of our publications (Frantti & Majanen, 2011), we presented and compared proportional-integral-derivative (PID) and fuzzy control systems to achieve maximum throughput and minimal delay by adjusting the packet sizes of UDP-based uni- or bi-directional real-time traffic in WLANs in line with prevailing channel condi- tions. The aim with the hierarchical expert system developed for this paper is to reach application-dependent delay and throughput limits for the maximum number of such real-time connections as VoIP calls and, if necessary, offload real-time connections to an- other network interface in a controlled manner. 2.2. Multihomed path management In node multihoming, the end nodes are responsible for paths’ management. Fekete (2010) notes that multihoming can be used to enable more advanced traffic engineering or load sharing tech- niques, such as load balancing and load spreading, by distributing traffic over multiple interfaces and addresses. In load balancing, different traffic flows are sent on different paths. In load spreading, packets belonging to the same flow are sent through different paths. Kandula, Lin, Badirkhanli, and Katab (2008), Singh, Alpcan, Agrawal, and Sharma (2010), and Yao, Kanhere, and Hassan (2009) use quality of service (QoS) parameters for balancing the traffic load over the available network interfaces. In host-centric traffic engineering, this is done by selecting the proper source IP address for outgoing packets and notifying peer nodes about the IP address for incoming packets. In network-centric traffic engi- neering, this is done within the network through tuning of the routing protocols (see Schiller (2005)). 2.3. Multihomed interface selection Common approaches to network selection are based on esti- mated QoS parameters for each available network. Adamopoulou, Demestichas, Koutsorodi, and Theologou (2005), Bari and Leung (2007), Psaras and Mamatas (2011), Song and Jamalipour (2005), Wang, Katz, and Giese (1999), Wilson, Lenaghan, and Malyan (2005), Xing and Venkatasubramanian (2005) and Yahiya and Cha- ouchi (2009) consider network QoS parameters such as delay and capacity. In the literature, researchers have also proposed more user-centric criteria, among them the power consumption of de- vices (see Yahiya & Chaouchi (2009) and Petander (2009)) or the cost of network use (see Adamopoulou et al. (2005), Bari & Leung (2007), Wilson et al. (2005) and Alkhawlani & Ayesh (2008)). In some of the aforementioned references, including those of Adamo- poulou et al. (2005), Song and Jamalipour (2005), Wang et al. (1999), and Alkhawlani and Ayesh (2008), a user is expected to supply policies and preferences. The approach of Alkhawlani and Ayesh (2008) proposes a combination of fuzzy logic and genetic T. Frantti, M. Majanen / Expert Systems with Applications 41 (2014) 4996–5008 4997 https://isiarticles.com/article/52584 
work_noodcfgro5bcdknzut2al2wpc4	----	Sddhand, Vol . 17, Part 2, June 1992, pp. 275-298. © Printed in India. Expert systems and finite element structural analysis - a review B P NAGANARAYANA and G PRATHAP Structural Sciences Division, National Aeronautical Laboratory, Bangalore 560017, India MS received 16 April 1991 ; revised 7 December 1991 ,Abstract . Finite element analysis of many engineering systems is practised more as an art than as a science . It involves high level expertise (analytical as well as heuristic) regarding problem modelling (e .g. problem specification, choosing the appropriate type of elements etc .), optical mesh design for achieving the specified accuracy (e .g . initial mesh selection, adaptive mesh refinement), selection of the appropriate type of analysis and solution routines and, finally, diagnosis of the finite element solutions . Very often such expertise is highly dispersed and is not available at a single place with a single expert. The design of an expert system, such that the necessary expertise is available to a novice to perform the same job even in the absence of trained experts, becomes an attractive proposition . In this paper, the areas of finite element structural analysis which require experience and decision-making capabilities are explored . A simple expert system, with a feasible knowledge base for problem modelling, optimal mesh design, type of analysis and solution routines, and diagnosis, is outlined . Several efforts in these directions, reported in the open literature, are also reviewed in this paper. Keywords. Finite element method ; optimal mesh design; structural analysis; expert system; knowledge base. The recent trend in the field of software development for scientific and engineering applications is to relieve the lay user of the need for expert knowledge necessary for the optimal use of the software as well as the theoretical and mathematical background for Using the code . For this purpose, a combination of knowledge-based expert system (KBEs) control and the automation of data generation and problem optimisation (e .g . adaptive mesh generation), with the computational method of solving the problem (e .g . finite element method), can make the entire system a powerful, efficient and cost- eRective tool for design, analysis and diagnosis of complex problems . Today, the finite element method has become an essential (and rather inevitable) tool for modelling and evaluating the physical performance of structural, mechanical and thennal systems in several engineering disciplines -aerospace, mechanical, civil, 275 276 B P Naganarayana and G Prathap nuclear, marine etc . It has considerable impact on other engineering fields like electromagnetism and fluid mechanics as well . It has become an integral part of the present day CAD/CAM cycles, and has even been used to simulate manufacturing processes like casting, forging etc . The scope of the present paper, however, is confined to the role expert systems can play to simplify finite element structural analysis. Structural analysis is the process of determining the response of a fully specified structure to its environment . Full specification of a structure needs a priori description of the topology of the structural domain, the material properties and the boundary conditions . The loads acting on the structure - including the equivalent loads produced by thermal effects, material changes like shrinkage and creep, initial strain/stress etc . - form the environment. The structural deformation and stresses (or stress resultants) are the standard response quantities of interest . The finite element analysis of a structural system has attained a high level of sophistication and involves a considerable amount of experience and judgement in taking a priori and/or a posteriori decisions regarding modelling of the system with proper elements, data management for optimum mesh size for achieving the desired accuracy, choice of the type of analysis required and choice of the correct combination of the solution tools and finally the diagnosis of the results obtained . A great deal of expertise and a considerable amount of heuristic knowledge are essential for such an exercise . Today, with the availability of numerous general purpose finite element programs (e .g . NASTRAN, ASKA, SAP, NISA, ANSYS, COSMOS, MARC etc .), coupled with their broad range of capabilities, the need for expertise in finite element analysis has been multipled . Often an experienced analyst might have attained expertise in one field of application using only one such general purpose software packages . Thus, apart from the novice user, even the experienced user often confronts difficulties while modelling a new class of problems with a new package . Thus, one of the major requirements is to accumulate, organise and thereby transfer the expertise of such experienced users to the less experienced ones, for the same class of problems and the same finite element programs, and, eventually aggregate such expertise over many different classes of problems and many different finite element packages . Knowledge-based expert systems hold the promise of providing a methodology for such finite element modelling aids . Often these are called intelligent pre- and post , processors (Adeli 1988), in the sense that they act as knowledge-based interfaces to the algorithmic programs . Architecturally, such expert systems are closely related to the so-called classification or diagnostic systems (Harmon & King 1985) since they are data-driven (either input or generated data). Several expert systems have been developed in practice with reference to modelling structural problems using some specific finite element programs (Bennett et al 1978; Bennett & Engelmore 1979, pp . 47-49 ; Rivlin et al 1980; Corlett et al 1984 ; Fenves 1985, 1986; Holt & Narayana 1986; Rehak 1986; Taig 1986; Wilson & Itoh 1986 ; Logo& Genberg 1987 ; Chen & Hajela 1988). SACON (Bennett et al 1978 ; Bennett & Eng 1979, pp. 47-49) is one of the earliest . It used the popular diagnostic expert system shell EMYCIN (Zumsteg & Flaggs 1985) and interacted with the user for proper application °f the MARC finite element program for structural analysis (Bennett et al 1978, Bennett & Engelmore 1979, pp . 47-49 ; Rivlin et al 1980; Fjellheim & Syversen 1983) . Most are knowledge-based consultation systems and are developed for experi mental purposes . Their applications are often confined to initial modelling of the problem for particular finite element program . Expert systems and finite element analysis Another area of the finite element method which demands experience, judgement and heuristic knowledge is that of identifying the singularities present, the locations of stress concentrations and their relative strengths and the designing of an optimal mesh for achieving the specified accuracy. Several strategies of adaptive refinement for this purpose (e.g. increasing the number of elements of the same order, h-type, or increasing the order of the element interpolations, p-type, or the combination of the two, hp-type, refinement) are developed in the literature (Babuska & Rheinboldt 1978 ; Atluri et al 1986; Babuska et al 1986 ; Zienkiewicz & Zu 1987). Recently, a few attempts have also been made towards developing expert-like systems to identify the singularities in the structure and arrive at optimal refined mesh configurations (Babuska & Rheinboldt 1978 ; Babuska & Rank 1987; Rank & Babuska 1987; Szabo, PROBE) . These systems mainly deal with the rules for adaptive refinement based on analytical error norms (e .g. strain energy norms, stress gradient norms etc .) and heuristic knowledge about the strength of the singularities and/or stress-concentrations . However, expert systems for such tasks are not yet well-established in practice. In this paper, the role that expert systems can play in general finite element structural analysis is discussed . An overview of several aspects of finite element analysis which need both theoretical/analytical and heuristic expertise is given . Finally a proposal for a simple expert-system architecture for such a task is presented. This architecture was the basis for building an experimental expert-system module for problem modelling and optimal mesh design for a special purpose 3-D finite element package at the National Aeronautical Laboratory in Bangalore (Prathap & Naganarayana 1991) . This is discussed briefly as a typical case study . 2. Finite element analysis - scope for expert systems The main areas of finite element analysis which demand a high level of experience and judgement capabilities (figure 1) are problem modelling (e.g. problem specification, choosing the appropriate type of elements etc .), optimal mesh design for achieving domain topology analysis pro bI em mode I I i n g optimisation diagnostics do porn specification environment specific of Ion error norms p r a b I e m adoptive refinement singularities material properties boundoy conditions ~---{ type of analysis type of solution Figure 1 . Scope of expert systems in finite element analysis . loading conditions . 277 278 B P Naganarayana and G Prathap specified accuracy, selection of appropriate type of analysis and solution routines and, finally, diagnosis of the finite element solutions . 2 .1 Problem modelling Modelling any problem requires expertise in selecting and/or development of the optimal theoretical basis (e .g. plate theories, shell theories etc.) and hence the type of elements to be used . Considerable knowledge regarding topological configuration and material constitution of the structure, boundary conditions and loading environment is required for this purpose . A decision on the theoretical basis is crucial for making the analysis meaningful, optimally accurate as well as cost-effective. It should be noted that the cost of finite element analysis is directly related to the continuity requirements and the total degrees of freedom used for modelling the problem (table 1) . An intelligent use of the available and/or generated knowledge is necessary to decide whether 3-dimensional, 2- dimensional or 1-dimensional modelling is appropriate for the problem, depending on the geometrical configuration and the material constitution of the structure, the boundary conditions and the loading environment . If 2-D or 1-D analysis seems sufficient with some loss of accuracy and/or compatibility in the displacement and the stress recovery, one has to decide whether a refined higher order shell, plate or beam theory is necessary for achieving the required accuracy . In addition, the degree of importance and risk associated with the type of application can also dictate the type of theoretical support required . It should be kept in mind that the time and cost of analysis increases sharply with an increase in the total degrees of freedom and continuity requirements for modelling the structure (table 1). Thus, 3-D analysis may be unwarranted, for example, when one of the dimensions of the domain is `small' as compared with the others . If this is not done, the cost of computation increases dramatically requiring much more computer time and larger storage requirements . Other deciding factors for choosing 1-D, 2-D or 3-D elements are the surface boundary conditions and loading. For example if 3 adjacent surfaces of a prismatic slab are completely or partially supported or loaded, 3-D analysis may become mandatory. Often, depending on the geometrical and material properties of the Table 1 . Comparison of different theories regarding the total d.o .f. and continuity requirement . Total d .o .f. (20 nodes per edge of hexa- d .o.f. hedral domain) per node 20 1 40 2 60 3 400 1 1,200 3 2,000 5 2,400 6 24,000 3 Expert systems and finite element analysis 279 structure, the boundary conditions and the loading environment, one can reduce a 3-D problem to a 2-D problem in a plane-stress or plane-strain sense (Timoshenko & Goodier 1970) . Accordingly, a 2-dimensional plane-stress or a plane-strain element may suffice for the modelling requirements . Depending on the loading and the type of element choosen, only the active components of strain energy need to be computed e.g . for purely in-plane loading on a plane-stress model of a plate, one requires to compute only the membrane component of the strain energy and the in-active degrees of freedom (transverse deflection and all rotations) should be suppressed . Such decisions reduce the computer time and cost effectively, in particular, when one deals with large problems . Very often, prescribing the boundary conditions in idealised terms, e .g . simple support or clamped, can be unrealistic and unreliable . Boundary conditions become much more complex when they become nonlinear. On the other hand, for some problems, one may go for a combination of simple boundary conditions without losing accuracy of solution. Similarly, modelling of loads acting on a structure can also be non-unique. Normally, the general loads acting upon a structure (e.g. airplane wing, landing gears etc .) do not fit into any standard category . In such cases, an expert may arrive at a decision regarding modelling of the boundary conditions and the loads based on the structure, its environment, his experience and intuition, and his knowledge of the information available in the literature . Hence, in such cases, it may be preferable to go according to the rule of thumb followed by experts in the relevant fields of application . Recently, Daniel (1991) has proposed an expert system for choosing proper boundary conditions for a given structure. Sometimes, the global geometry of the structure and the possible local singularities may lead to conflicting opinions about using any particular theory for modelling and/or the method of solution . For example, a 2-D structural domain with a small, but critical, crack embedded deep in the domain may just require a global 2-D analysis and a local 3-D analysis around the crack . Sometimes, for a small crack in a relatively large 2-D or 3-D structural domain for which only the global behaviour of the structure is of interest, it may be desirable to have a combination of the boundary element and the finite element methods in modelling the problem . This achieves a cheap but sufficiently accurate solution . On the other hand, if the strength of the singularity is low enough, a refined mesh may be suf . Tent for achieving the specified accuracy. Problem modelling cannot be complete without defining proper correlation (geometrical as well as material) between any two points in the domain . Experience is required in choosing the interpolation functions for generating the geometrical configuration of the structure (e .g. automatic volume, surface and line generation) from the minimum data input . This is turn can play a decisive role in choosing the type and order of the elements (as is apparent above) and the type of the basic mesh required. Material modelling requires special attention, in particular when the structure is urinated . Material correlation in the thickness direction of the laminate should be evolved from the lay-up data (e .g . thickness, orientation and material properties of each layer for each of the orthotropic layers and the global constitutive relationships for each of the anisotropic layers) from which one has to deduce knowledge on the type of the laminate structure e .g . symmetric. antisymmetric, cross-ply etc . Accordingly computations can be optimised later by generating only the required data and following only the required steps of analysis . Dimension Theoretical basis Continuity requirement 1-D Elementary C' First order C° Third order C° 2-D Elementary C' First order C° Third order-I C° 'Third order-I] C° 3-D C° 280 B P Naganarayana and G Prathap 2.2 Mesh optimisation Mesh optimisation gains importance, in particular, when singularities are embedded in the problem domain. In general, the singularities create error-prone zones in their proximity. This demands data refinement, localised to such zones, for achieving the specified accuracy of solution in an optimal sense. The domain singularities, with reference to solid and structural mechanics, fall in one of the four general categories: Material singularities - Hard grains, material voids etc . Geometric singularities - Cracks, notches, voids etc . Boundary conditions - Point supports, discontinuous supports etc. Loading environment - Concentrated forces, discontinuously distributed loads etc . Often, an expert uses his experience, and the analytical as well as heuristic knowledge available in the literature with reference to both the method of analysis and the problem under consideration so as to optimise the data refinement in the error-prone zones to achieve accuracy of solution in an efficient and cost-effective manner . Sometimes it may not be convenient to determine the strength of the singularities analytically . In such cases, the rule of thumb followed by the acknowledged experts in the corresponding field may have to be sought for mesh optimisation . In general, it is impossible to construct an a priori error norm for finite element modelling of a structure of complex geometry . In such cases one has to use a posteriori information from cheap coarse mesh computations of the model, construct the error norms, identify the error-prone zones and optimise the mesh appropriately to achieve the specified accuracy in the most efficient and cost-effective manner. A considerable amount of research and development has taken place in literature with reference to the error norms for finite element analysis of structure with singularities and several adaptive mesh refinement techniques (e .g. h-type, p-type and hp-type) have been proposed to achieve the specified accuracy (Babuska & Rheinboldt 1978 ; Atluri et al 1986; Babuska et al 1986; Zienkiewicz & Zu 1987) . Recently some efforts have been made toward developing some expert-like systems for co-ordinating such activities in finite element analysis (Babuska & Rank 1987; Rank & Babuska 1987 ; Szabo (year not given) . However, the resulting expert systems are primitive in the sense that they only consider the analytical knowledge with reference only to the method of analysis (i .e . adaptive finite element methods). 2 .3 Type of analysis and solution Another area in which a great deal of expert knowledge is essential is the understanding of how a structure with certain geometric and material properties responds to a given load under given boundary conditions . The major parameters involved in taking such decisions are the material properties, the domain geometry, the boundary conditions and the loads. One has to have a clear understanding of the independent as well as the collective effects of these parameters on structural behaviour before deciding on the type of analysis required . Knowledge of the microstructural as well as the macrostructural properties of th e material is required in deciding whether the material behaviour is linear or nonlinear , elastic or plastic, compressible or incompressible etc . One has to have an a priori knowledge of whether there are certain zones in the domain which require nonlinea r elastic or plastic analysis though the global structural behaviour is linearly elastic For example, the zone in the proximity of the crack tips in an elastic domain ma y Expert systems and finite element analysis 281 require plastic analysis in a local sense, in particular, under fatigue load . Such knowledge becomes important, for the design/analysis of failure of a structural component. Sometimes the loads can also lead to such requirements, though the material is linearly elastic, depending on their magnitude . Normally a material database is required for such purpose. The next important aspect is about the geometric nonlinear behaviour of a structure . Based on the structural geometry, boundary supports and the type of loads on the structure, one has to decide whether the structure is undergoing relatively small or large deflections and accordingly choose the theory for the analysis . The characteristic properties that influence geometric nonlinearity of a structure are mainly the structural geometry and magnitude and direction of the load . A good example of geometric nonlinearity is seen in the buckling of structures . Often such problems are appro- ximately solved as an eigenvalue problem for certain simple cases (e.g. buckling of beams) with reasonably accurate solutions . On the other hand, the true geometric nonlinear considerations often become mandatory for analysing buckling of complex structures like shells . Normally one needs a posteriori information about the deflections and stresses for taking a decision on nonlinear analysis for a structure, in particular if deformation or strain under load is large . If the deflection at any point of the structure is of the same order as the minimum characteristic structural dimension (e.g . thickness of a plate) one may need geometric nonlinear analysis . On the other hand, if any of the stress components is close to the corresponding elastic design limits for the material (e.g. corresponding yield stress) material nonlinear analysis may be required . Again this material non linearity can fall into any category like elasto-plastic, visco-elastic, plastic etc. depending on the structure and its environment . Often, an expert uses his past experience and depends on several heuristic rules for taking such decisions . Apart from the magnitude, location and direction of the loads which normally influence the geometric (sometimes material) nonlinearity, the type of the loads with respect to time also influences the type of analysis required . Expert decisions have to be taken about whether a static analysis is enough or a dynamic analysis is required depending on the type of loads . If dynamic analysis is required, the type of analysis again influenced by the type of application as well as the duration of application the load (e .g . eigenvalue analysis, frequency response analysis, impact load analysis, tier analysis, fatigue damage analysis etc .). The presence of certain singularities like a crack may require special knowledge h as fracture mechanics - which may again be linear or nonlinear - to be rporated into the analysis system . The analysis may differ for different types of ks . The analysis of crack initiation and propagation becomes very important, in cular, for enduring dynamic loads causing fatigue in the material structure . Depending on the type of numerical method and the type of analysis one has to an optimum combination of solution routines . Exact or direct solution hniques may be desirable in conventional finite element modelling of linear beems, probably enhanced with an `out-of-core' algorithm for very large problems . the other hand, iterative solution routines may be preferable for conventional e element modelling of nonlinear problems (e .g . Newton-Raphson method) and adaptive unite element modelling (e .g. element-by-element conjugate gradienth with hierarchical preconditioning of the stiffness matrix) . For dynamic analysis may need additional solution routines like time-step integration, eigenvalue tion etc . depending on the type of analysis required . 'tall Y, expertise is needed in determining the tools to be used for tackling the 282 B P Naganarayana and G Prathap problem in numerical and logical manner, for example, what combination of computer languages, what sort of data management, how the already available software can be optimally used etc . Today, most of the expert systems are written in C-language owing to its flexibility and efficiency in handling both numerical as well as logical operations . Blackboard architecture (Nilssion 1980 ; Nii 1986) is popular for the management of data and computational as well as logical software . 2 .4 Diagnostics The usual practice of analysis, especially for a complex problem, is to start with a . simplified approximate model and then choosing optimally efficient combinations of more sophisticated tools of analysis after an intelligent diagnosis of the initial results. Interpretation of the final results also demands high levels of intelligence and expertise. Thus, it is often the diagnosis and interpretation of the finite element results for a complex problem that calls for high levels of intelligence, expertise and judgement capabilities . On the other hand, this is the most difficult part to code, since such knowledge is normally heuristic and changes from person to person, context to context and finally organisation to organisation (depending on the policies) . In addition, most of the expertise on diagnosis is highly problem-dependent and cannot be generalised for all finite element applications as such. Therefore, developing a general purpose expert system for diagnosis of finite element results may not be a practical proposition and in fact hardly any effort has been documented in this area in the open literature . Thus one may have to confine the system to tackling specific problems like analysing rocket tanks (Desaleux & Fouet 1966) etc . using an appropriate finite element package. 3. Expert systems As discussed earlier, if the expertise has to be transparent to a novice user of an analysis package, it becomes mandatory to coordinate the knowledge regarding the package and the field of interest and make it readily available . Again, as explained earlier, an expert system is a very useful tool for such an exercise. problem In addition, a knowledge-based expert system combined with automated p modelling and problem optimisation can make the analysis process very efficient and cost-effective when the process is of repetitive nature (figure 2) . Such a system can have many advantages over conventional numerical analysis . Once the expert system is calibrated by the expert, he can release the software for use to less qualified assistants and coordinate several different tasks simultaneously. Since the algorithmic comp tions are also optimised, the total time and cost of the whole analysis process is drastically reduced . Finally, such systems can be effectively used to train novice users in specific fields . Various interpretations and definitions of expert systems can be found in the literature . For example : "An interactive computer program incorporating judgement, experience, rules of thumb, intuition and other expertise to provide knowledgeable advice about "variety of tasks" (Feigenbaum 1981). r ocedure to "An intelligent computer program that uses knowledge and inference P ertisef or solve problems that are difficult enough to require significant human exp their solution" (Gasching et at 1981). expert final analysis report Expert systems and finite element analysis ( problem assignment ) / ( results and diagnostics ) calibration establish ES no ledge analysis module I automated data generation / refinement I expert system use Figure 2. Role of expert systems in finite element analysis. "An expert system solves real-world, complex problems using a complex model of expert human reasoning, reaching the same conclusions that the human expert would reach if faced with a comparable problem" (Weiss & Kulikowski 1984). Keeping the dominantly computational nature of the finite element analysis in mind, the last interpretation for the expert system appears to be most appropriate in the present context . In this section, the aspects involved in building an expert system are discussed with particular reference to finite element analysis of structural problems . It is convenient to handle an expert system if its frame is divided into a domain- dependent part, the knowledge base and a domain-independent part, the Inference Engine (Adeli 1988; Levine et al 1988). The inference engine forms the context, working memory and decision paths for the system while the knowledge-base provides the technical informations and rules for reaching the specified goals in a particular domain of application, as an expert would do. Essentially, the inference engine coordinates different tasks (both computational and logical) and the knowledge base, normally constituted by database and rulebase, forms the body of knowledge for executing the tasks identified by the inference engine . The general configuration for an expert system for finite element analysis is shown )n figure 3 . It consists of an inference engine having two parts . The logical modulec oordinates the information provided by the knowledge-base, database, rule-base and User interaction enhanced with an expert advice facility to select the appropriate finitee lements, methods/types of analysis, a coarse initial mesh etc . Essentially, this module produces an optimum initial model for the given problem . On the other hand, the COnPutational module coordinates the computational, algorithmic and graphical toftware through a recurring cycle of finite element analysis, error estimations and mesh refinement resulting in an optimally accurate solution for the problem . This iiodule essentially involves problem optimisation and solution . Finally the system 9n be enhanced with the assistance of post-processing and diagnosis facilities . Other major modules in such an expert system are the Advise/Help, Explanation and subordinates know ledge acquisition module novice or experts 283 284 database B P Naganarayana and G Prathap rule-base L diagnostics inference mechanism knowledge base advice / help facility 0gICaI module use r procedural software computational module expi anation facility knowledge acquisition T explanation facility learning facility post - processor diagnostics advice/help facility Figure 3. A typical expert system for finite element analysis . Learning facilities. Figure 4 shows the details of the design of such an expert adaptive finite element structural analysis system. 3 .1 Knowledge base Knowledge can be categorised into two types. Surface Knowledge consists of rules regarding the necessary tasks, data input and the sequence of the tasks for the given particular problem (e .g. determining relative dimensions of the structure) . This can be normally handled by a rule-based expert system through a user-interactive logical questionnaire. On the other hand, Deep Knowledge involves knowledge about the use r procedural software line, surface and -- volume generator mesh generator I . element conne t"' 22 boundary and load conditions F .E . package error estimators solution routine graphics pack ag e post-proces sor Figure 4. Blackboard architecture for expert system in adaptive finite element analysis . Expert systems and finite element analysis 285 underlying problem structure (e.g . error estimation and recognition of critical zones) . Each can be Declarative (facts about objects, events, e .g . whether strain energy density at a point is high or low) or Procedural (information about courses of action e .g. the mesh refinement strategy depending on the error norms) . Surface knowledge is often heuristic while deep knowledge is analytical . The process of building a knowledge-base involves acquisition of relevant and reliable knowledge from the experts and other sources like the literature etc . and its representation in a form that the inference engine can process in an efficient manner. 3 .1a Knowledge acquisition : The most difficult challenge in building a knowledge- base is to acquire reliable knowledge relevant to the domain under consideration . This requires careful survey of different reliable sources of knowledge such as (Adeli 1987): (1) Technical literature (e .g . books, manuals, journal articles etc .); (2) domain experts (e .g. interviews, circulating questionnaires, example problem-solving sessions etc .) (3) experimental data (e .g. numerical or laboratory experiments) ; (4) input data for the particular problem . Again the knowledge acquired can be domain knowledge (i .e. the knowledge related to the specific area of application e .g . finite element structural analysis) or problem knowledge (i .e . knowledge related to a particular problem in the domain e.g. the dimensions of a structure) . However, they can be either deep or surface and declarative or procedural . Domain knowledge is normally independent of the particular ptoblems of application in the domain. Since such knowledge does not vary from problem to problem it is often called static knowledge . For the same reason, this can be in-built in the system permanently in the form of rules, nets or frames so that the end-user need not have any control on it. In general, the first three sources mentioned above are exploited to acquire such knowledge. Some of the examples relevant to the finite element analysis are rules for selection of appropriate theoretical bases for structural modelling, proper finite element mesh, type of analysis and solution etc . On the other hand, problem knowledge varies from problem to problem (hence often called dynamic knowledge) and it is generated from the input data for a particular problem in the domain in the user-interactive mode . This class of knowledge gains Particular importance in the present context . Some of the examples are whether the structure is 'thin' or 'thick', flat or curved, isotropic or anisotropic, whether a plane stress or a plane strain model is required etc. It can be observed that input data regarding the structural topology, material constitution and the loading environment are pre-requisites for generating such knowledge . This can be achieved through an efficient user interface . A typical flow of logic in modelling a complex structure with simple hexahedral substructures and determining the relative characteristic dimensions of each of the substructures is shown in figure 5 . Such knowledge is essential while i nterpreting domain knowledge (e .g. selecting the appropriate theoretical basis and type of elements) later . 3.1b Knowledge representation : An equally important task is to represent the k nowledge that is acquired in an efficient form that can easily be addressed by the inference mechanism. Knowledge can be represented in a program in two ways - inference mechanism knowledge or data generator problem modelling oblemoptimisation geometry _ dimension determination coarse mesh (initial ) ---- material boundary _ condition _ I ooding conditions ------ _ laminate properties mesh generator (automated) theoretical support finite element analysis method of analysis error estimation computational -- type of error analysis analysis knowledge base computational - tools mesh refinement database or _ rule base coarse mesh rules optimal solution 286 B P Naganarayana and G Prathap reduce your structure to parts of simpler geometry, preferably of standard shapes1if nonstandard, the part should be bound by 4 to 6 surfaces is feed the standard dimensions locate the part in the 3-D Cartesian space para metric generation of the rest of the surfaces flat faces create substructure volume determine characteristic dimensions the port of tandard shape y yes yes are all the surfaces I f standard geometry) ye~ Ino o do at least 2 opposite faces have standard shape ) dge mid points defined no e serendipity quadratic surfaces I ye no 0 no a It o many surfaces bind the port ? 1 t parametric generation all the surfaces I I surface centroid def ye Figure 5 . Flow of logic for geometric data generation . procedural and declarative . Procedural representation is normally used in conventional algorithmic programming and is very efficient . But the knowledge involved is highly context-dependent and embedded in the code, thus making the knowledge opaque - unintelligible and difficult to modify . On the other hand, declarative representation encodes knowledge as data so that it is context-independent and is therefore easier to comprehend and modify. While semantics are distributed over the code in procedural representation, they are collected in one place in declarative representation. Several approaches of declarative knowledge representation are available in the Al literature (Hayes-Roth et al 1983; O'Shea & Eisenstadt 1984; Brachman & Levesque 1985) . The major ones are briefly described below . Rule-based production system - This has been the most favoured representational approach for building expert systems . A production system (Brownston et al 1985 ; O'Shea & Eisenstadt 1984; Van Melle 1979, pp . 923-925) is a collection of several rules each of which consists of an ordered pair of symbols (IF and THEN). The first constitutes the premises or the conditions while the second constitutes the actions . This is the most popular and simplest method . The problem reduction method (Barr & Feigenbaum 1981-82 ; Cohen & Feigenbaum 1982 ; Ishizuka et al 1983 ; Rich 1983; Winston 1984; Charniak & McDermott 1985) is often useful for such knowledge representation so that a complex problem can be subdivided into simpler problems , described hierarchically, each having a subgoal . Logic-predicate calculus - This is basically an extension of the so-called propositio n logic where a proposition can be either true or false. A first order predicate calculus (Ishizuka et al 1983 ; Ogawa et al 1984) consists of a number of components or propositions (like predicate symbols, variable symbols, function symbols and constant symbols) which are interconnected through the logical connectives (like AND, OR, NOT, EQUIVALENT, IMPLIES etc.). This type of knowledge representation is very c onvenient, in particular, with declarative languages like PROLOG . red of Log rang ion quadratic surfaces Expert systems and finite element analysis 287 Semantic networks and triplets - Semantic network, normally, involves nodes and arcs (links). Nodes represent objects, concepts or situations, their attributes and values . The arcs link these nodes and define the nodal heirarchy . If the net consists of only one object, and its attribute and value, the three nodes form the so-called semantic triplet or OAV triplet (Adeli 1988). Frames - Sometimes frame-like structures are found to be convenient for knowledge representation, in particular for representing sequences of events in expert systems (Minsky 1975; Aikins 1983 ; Yao & Fu 1985). A frame is composed of its name and several slots. Each slot may have a number of facets. A slot in a frame may be an attribute of the object or it may be a relation. A relation slot is used to link different frames together . In the present context, for such a dominatingly analytical exercise, a rule-based production system appears to be most promising . 3.1c Representing uncertain knowledge -interval modelling: Often the available knowledge about the domain and the problems of interest in the domain are uncertain and incomplete . Usually, such knowledge is qualitative e .g. strength of a singularity is high or plate thickness ratio is low etc . In fact, the intelligence of an expert (and hence of an expert system) is reflected in his ability to handle such knowledge . In such cases the so-called approximate reasoning or inexact reasoning or reasoning under uncertainty(Zadeh 1974, pp. 591-594 ; Blockley 1980; Ruspini 1982 ; Ishizuka et al 1983; Negoita 1985 ; Lecot & Parker 1986) should be employed for representing and processing the available uncertain knowledge . Interval modelling (Wong 1986) is one of the most popular mathematical tools used for handling uncertainty . Different mathematical measures which use interval modelling concept are illustrated in Wong (1986) . Some of them, which are useful for the present exercise, are briefly discussed here . (a) Simple intervals -This needs minimum information as seen in figure 6a . The value of a parameter is estimated to be within aft interval [a, b] and nothing more is known . 1 0 N (X) (d) 0 (e ) 1 1 µ(x) x o a o Figure 6. Different types of fuzzy knowledge representation : (a) simple interval, (b) triangular set, (c) rectangular set, (d) trapezcidal set, (e) normal set . X x 288 B P Naganarayana and G Prathap (b) Probability measures - Probability measure implies exact knowledge of the uncertainty . Although the value of the parameter is uncertain, there is certainty about the character of its randomness (e .g . normal curve) . The probability measure is built on a simple interval and the probability of an event A, say p(A) is known at any point in the interval . It should be noted that p(A) implies that probability of the event (not A) is also certain and p(not A) = I - p(A). (c) Possibility measures and fuzzy sets - Possibility measures are again built upon a combination of simple intervals as seen in figures 6b to 6e . The possibility of an event u(A) is known at any point in these interval and the resulting set is generally called fuzzy set . Here, the stringent presumption of probability theory is relaxed by introducing the concept of ignorance (Shafer 1976) that the probability of event A should not say anything about the probability of (not A) unless there is explicit evidence supporting the latter i .e . µ(not A) is independent of µ(A). This measure is often identified with the so-called evidence measure . Depending on the variation of the possibility factor µ(A) associated with the event/object A the fuzzy set can be of any shape . For example, a triangular set (figure 6b) can be used for a pessimistic representation of the event A that at only one point p(A) = 1 . On the other hand, a rectangular set (figure 6c) can be used for optimistic representation such that µ(A)=1 at all points except at the two end points . Again the possibility function can be a continuous function (figure 6e) . But a trapezoidal set (figure 6d) is seen to be the most practical choice for representing fuzzy knowledge . There are several instances in finite element structural analysis where the knowledge representation becomes fuzzy . The most common one is of determining the dimension of analysis for the structure, in particular, if the structural geometry is irregular and distorted. Figure 7 shows how the thickness ratio for a general hexahedral structural domain can be represented using the fuzzy set theory (also see §41). Some other instances where fuzzy knowledge representation is required in finite element structural analysis are - the intensity of mesh refinement required in an adaptive finite element approach, the strength of a structural singularity, strength of the inter-laminar bond in a composite structure etc., where the knowledge is often heuristic and verbal . 3 .2 Inference mechanism and expert system architecture An inference mechanism is the part of an expert system that deduces new facts from known facts of the knowledge-base. This mainly constitutes the problem-solving strategy for the required task . Problem solving can be considered as a search for solution through a state space by the application of operators, where the space (the possible states in a problem solution) consists of an initial state, a goal state and an intermediate state . The solution path consists of all states that lead from an initial state to a goal state. A number of different problem-solving paradigms are available in the Al literature - the describe-and-match paradigm, the goal-reduction paradigm generate-and-test systems, means-ends analysis, rule-based (production) systems etc . (Barr & Feigenbaum 1981-82: Cohen & Feigenbaum 1982; Ishizuka et al 1983 ; O'Shea & Eisensfadt 1984; Winston 1984; Brownston et al 1985; Charniak & McDermot t 1985). Many of these approaches overlap or can be used, in conjunction with each other. However, it is the production system which is popularly used for developing most of the knowledge-based expert systems so far . This has additional advantages oper other strategies, being simple to code, with improved clarity, and being suitable for Expert systems and finite element analysis H 2 0 -- h/I ratio as triangular set h/I ratio as rectangular set min . 0/0 = 0067 mox .(t/I)= 0 . 173 mean 0/0 = 0 . 108 min (b/1)= 0105 mox(b/1)= 0 576 mean (b/I)= 0 307 min (t/b) = 0 . 128 max (t/b) = 1498 mean It/b) = 0594 A - classical B - 1 order C - III order (A) D - III order (8) E -3DeIasticity (approximate ranges) template Figure 7. Fuzzy sets for the dimension ratios : (a) generation of dimension ratios ; (b) comparison of fuzzy sets. Parallel computations and for designing an explanation facility for the system . In particular a rule-based system combined with the goal-reduction paradigm can form a very efficient and powerful model of the human information processing and problem solving abilities . A production system normally consists of three main elements - aset of IF-THEN rules (or knowledge base), a global database or working memory or the Context and an inference mechanism . 289 I point X Y A 0 0 B 21 3 0 4 C 20 5 0 6 D 1 9 0 E 1 0 2 2 F 20 4 09 G 19 6 0 8 H -I II 18 - - mean dlmn S(1,m) 10 863 13 928 10 198 7 071 10 515 1 S(j,m) 5 0 8 . 544 9 434 13 675 9 163 b S( k,m) 5 0 8 544 10 630 9 950 8531 t 290 B P Naganarayana and G Prathap 3 .2a Inference mechanism - problem solving strategies : The next question is, how are the rules determined from the knowledge base and which ones should be used at a certain state of the solution path . For this one has to choose an inference mechanism or a control strategy which constitutes the heart of a knowledge-based expert system. Inference mechanism fires the rules in accordance with its built-in reasoning process . The two major types of reasoning are antecedent reasoning and consequent reasoning. (a) Antecedent reasoning (forward-chaining, data-driven control strategy) - In this type of inference mechanism, the rules are scanned until one is found whose antecedents (left-hand sides) match the information for the problem entered in the working memory. Then the rule is fired and the working memory is appropriately updated . This process is repeated until the goal state is achieved or no useable rule is found . This strategy is particularly useful if the goal state is not known or if there are too many possible outcomes. Complex planning problems can be tackled through this approach . The reasoning involved in the finite element structural analysis predomin- antly falls into this category e .g. determination of the dimension of analysis and the theoretical support required, given the structural topology, material constitution and loading environment, error prediction from finite element analysis and heuristic expertise etc . (b) Consequent reasoning (backward-chaining, goal-driven control strategy) -In this type of inference mechanism, rules are scanned and those whose consequent actions can lead to the desired goal are determined . The associated antecedents are tested to see whether they match the working memory information . If they all match, the rules are fired and the problem is solved . If the goal states are known and their number is small then this reasoning process seems to be advantageous . This is popular in the diagnostic type of expert system . Sometimes it may be convenient to use backward chaining in finite element structural analysis e .g . selection of appropriate elements and design of the required finite element mesh based on the knowledge of the dimension of analysis and the theoretical supports that are already established, adaptive refinement of the finite element mesh based on error predictions etc . (e) Opportunistic reasoning (hybrid or mixed chaining) - This is a combination of both forward and backward chaining . It follows that this becomes mandatory, in particular, for an expert system for finite element structural analysis . This approach can be efficiently utilised through the use of the so-called 'Blackboard Environment' (Nil 1986). The blackboard will keep track of the simultaneous application of backward and forward reasoning chains which are activated at the most 'opportune' time . 3 .2b Blackboard architecture: The concept of blackboard architecture was first implemented in HEARSAY-Il, a speech understanding system (Reddy et al 1973, pp. 185-193) . The control of the program can be made flexible and efficient and the restricted accessibility of routines could be eliminated by isolating the routines (or modules), but allowing all the subroutines to make use of the common data structura l The blackboard model of problem solving is highly structured and essentially a special case of opportunistic reasoning . The knowledge necessary for solving the problem is divided into independent groups of rules called knowledge sources . A blackboar d plays the part of a communication vehicle among knowledge sources and keeps track of the incremental changes made in the problem state until a solution is found . In the present context of expert systems for finite element structural analysis such the an architecture appears to be the most suitable since all the information and are associated rules necessary to model the structural domain geometry, boun Expert systems and finite element analysis 291 conditions, material constitution and loading conditions (acquired/generated as mentioned in the previous section) can be stored in the respective independent knowledge sources. Figure 4 shows a schematic diagram of the blackboard model for the system which acts as a global database in controlling the above-mentioned logical knowledge sources as well as the computational knowledge sources. 3.2c User interface: The primary motivation for expert-system development is to make the expert knowledge, in a specific area of interest, available to the amateur user of the system. Thus inclusion of an exhaustive user interface in the system becomes mandatory. This module generally includes Advice/Help, Explanation and Learning facilities. This monitors functions like modelling the problem (e .g. structural topology and material constitution and the loading environment), selection of appropriate tools (e.g . finite elements, mesh size, type of analysis and solution routines etc.) through a live conversation with the user . At any state in the problem space the system should be able to advise the user with proper verbal/graphic support, if asked for, about taking decisions appropriately depending upon the problem environment. Also capability of self-justification is desirable so that the user should finally know why and how the system arrived at the suggested goal/advice. In fact the main purpose of an expert system, that transfer of the experts' knowledge to the novice users of the system be simple and flexible, cannot be fulfilled without such an explanation facility . For example, if the system decides that 1-D analysis is enough for a structure, it should be able to educate the user about why such a decision was taken and what are the preceding factors that lead to such a decision, thus enabling him to examine the reasoning process . Finally, the system should update its knowledge, as its human counterpart would normally do, from time to time . For example, if the user creates a new geometrical shape which is unknown to the system and if he feels that it invites frequent usage, the system should be able to 'learn' the necessary details and store them in its library with the help of the user . A more powerful example is of self learning . For example, if the system reaches the same goal (e .g. nonlinear analysis for a structure) during several runs, appropriate chunks of knowledge (e .g. geometry, material, boundary conditions and loading) involved in all such runs may be compared to generate, Probably, a new heuristic rule which can be used in future under similar conditions without doing the actual computations required . Building such automated learning systems is quite a difficult challenge and the expected efficiency and effectiveness are Yet to be achieved . 4. A case study - an expert system for finite element analysis of structural problems Using 3-dimensional elements A Project on developing an expert system for control of modelling of the structural Problems using finite elements (only 3-D elements at present) is under progress at the National Aeronautical Laboratory in Bangalore (Prathap & Naganarayana 1991) . Mesh refinement is automated using an h-type strategy based on strain energy density error norms. In this section, we take this up as a case study to demonstrate some of the issues involved in such developmental work . The inexact representation of certain Problem knowledge using fuzzy sets is first discussed . Later some rules related to the Present system are briefly listed below (the knowledge-base, however, is built for using 1 'B, 2-D as well as 3-D elements) . 292 B P Naganarayana and G Prathap 4.1 Inexact representation of knowledge The major parameters influencing the selection of the theoretical bases to be used, as we already discussed, are the domain dimensions, material constitution and presence of singularities. Usually it is not possible to represent the knowledge about these conditions in an exact sense . The fuzzy set representation of these can streamline the problem-solving process for elegance and efficiency . 4.1 a Fuzzy sets for ratios of characteristic structural dimensions -. <Dh,) and (Dkb ): First, the major sets of dimensions are identified and their mean values are calculated . Then the ratios of the smallest mean dimension with the rest are calculated. For example, consider a hexahedral domain (figure 7) which has 6 surfaces, and 3 sets of dimensions, each set having 4 values . Let S,(m), S .(m) and Sk (m) (for m = 1,4) represent the three sets and let L., L, ana Lk be their corresponding means . Let 1 and h represent the length and thickness dimensions corresponding to the maximum and minimum mean dimensions and let b represent the remaining (breadth) dimension. Let (Dh ,), <Dhb) and (D b ,) be the fuzzy sets representing the ratios h to l, h to b and b to I . These sets can be constructed as follows (also refer to figure 7) : Let, Xh, = (Sh/SI)min ; Yhl = (Sh/S,)m,= and Z,,, = (Sh/SI)mcan . Using the values X and Y for the two points (A and B) of zero possibility factor and Z for the point (C) of unit possibility factor a triangular fuzzy set can be formed (figure 7b), Trapezoidal sets can be constructed as follows . Let, Hh,l = Abs(Zhl - Xhl) and H h1s = Abs(Zh, - Yhl) and Hh , = Min(Hhn, Hh,2) x fhl, where fh, can be a factor of ignorance (normally ranging from 0-9 to 1 . 0) . Similarly Hhb and Hb , can be calculated . The two points (C and D) of the unit possibility factor for the trapezoidal fuzzy sets <D h,), (Dhb ) and <D b,) are then given by (Zi,l ± Hhl) ; (Zhb ± Hhb) and (Zbl ± Hbl), respectively. (figure 7b). From a comparison of these fuzzy sets with a template of standard sets (figure 7b) one can decide the dimension and theoretical requirements for the analysis of a structure (as will be seen in the next subsection) . Note: If fuzzy sets are not used, then, the scalars, D h , = Lh /L, and Dh,, = Lh/Lb , calculated based on the mean dimensions, can be used for comparison as practised conventionally . 4.1b Fuzzy sets for singularities-<Si ) : Let there be n singularities in the problem domain. Associated with each, there are two parameters that can influence theoretica l support-stress concentration factor or criticality of the singularities <SC,) (i =I,n) and the accuracy of stress recovery demanded by the user (S,4 1 ) . SA can be used as a weighting factor for deriving the resultant fuzzy set <S i ) for ith singularity . Expert systems and finite element analysis 293 4.Ic Fuzzy set for material structure-<C) : As discussed earlier, the laminated structure may require the use of higher order theories, especially if high accuracy is demanded for inter-laminar stresses . If the inter-laminar bonds are weak, failure normally occurs across the laminae. In this case, higher order theories are essential for meaningful stress calculations . Let <SL> and <B W> represent the fuzzy sets for the accuracy of laminar stress recovery demanded by the user and the interlaminar bond weakness, respectively. Using SL as a weighting factor, the resultant fuzzy set for lamination effects (C) can be constructed . 4.2 Some typical rules Some of the typical rules that can be used for modelling a structure using finite-elements and for remeshing the mesh adaptively, are listed here. 4 .2a Rules for 3-D analysis: (I) Global 3-D analysis if, {Number of the binding surfaces _< 4) or {3 or more adjacent surfaces are fully or partially supported/loaded) or {Both (D,,) and <Dhb ) are very high) or {Both <D>'s are high or very high and <S) and/or <C> are very high) . (2) Local 3-D analysis around ith singularity if {When <D,,> and/or <D hb ) are small} and {singularity <S1 > is very high) . 4 .2b Rules for 2-D analysis (3) 2-D analysis if, {any of <D>'s, <S) and (C) is not very high, . Theoretical basis: (4) Third-order plate/shell theory-II if, {all (D)'s, <S) and/or (C) are high) . (5) Third-order plate/shell theory - I if, {both <D)'s are high or moderate and <S) and/or <C) are moderate (or vice versa), . (6) First-order plate/shell theory if, {both (D)s, <S) and/or <C> are moderate or low or very low but not (D)'s as well as <S) and/or <C> are very low) . (7) Elementary plate/shell theory if, {all <D)'s and <S) and/or <C> are very low). Geometric basis: Now construct the mid-surface in the 1-b curvilinear 'plane' by considering the midpoints of the edges in the h-direction and find out the curvatures kl and k2 in the 1-h and b-h planes respectively . This is done if standard domain shapes are not chosen . Find the Gaussian curvature K = kl x k2 . 294 B P Naganarayana and G Prathap (8) Plates if, {kl =k2=0} . (9) Plane stress model if { (8) and (6) or (7) are satisfied) . (10) Plane strain model if, {both <D>'s are high or very high and the h-, b- dimensions, loads and boundary conditions do not vary with the I-dimension) . (11) Zero Gaussian curvature shells if, {either of kl and k2 are non-zero but K =0} . (12) Positive Gaussian curvature shells if, {K > 0} . (13) Negative Gaussian curvature shells if, {K < 0} . (14) Shallow shells if, kl and k2 are very small. 4.2c Rules for 1-D analysis : (15) 1-D analysis if, { (Dhb ) is very high but <Dh,> is not high or very high) . (16) Higher order beam theory if, { <Dh ,) is moderate) . (17) First-order beam theory if, { <D,,,) is low} . (18) Elementary beam theory if, { (D,,,) is very low} . Since both b- and h- dimensions are comparable, the line passing through the centroids of the b - h planes forms the 1-D arc for which the curvature k is calculated if non-standard domain shape was selected . If load is acting in the b-direction Db , should be used in place of Db , . (19) Straight beam if {k = 0) . (20) Curved beam if. {k is not zero) . (21) Warping corrections are necessary if, {b - h planes are non-circular) . 4.2d Rules for computation : (22) If the structure is a plate/straight-beam or shallow shell/curved-beam without singularies, {use linear elements in the coarse mesh} . (23) If the structure is a deep shell/curved-beam and if K = 0, (use linear elements), if K ~ 0, (use quadratic elements) . (24) If the structure has singularities, {use quadratic elements around them) . Expert systems and finite element analysis 295 (25) If the structure is a plate and load is acting only in the I - b plane, {calculate energy and the stiffness corresponding to inplane normal strain components only) and {suppress all the degrees of freedom other than the inplane displacements before assembling the stiffness} . (26) If the structure is a plate and load is acting only normal to the plate (calculate only transverse shear and flexural strain energies and the corresponding stiffness matrices) and (suppress all the degrees of freedom corresponding to inplane displacements before assembling the stiffness) (27) Calculate the laminate elastic matrices and carry out the related analysis only if {the structure is laminated) . (28) If the laminate is symmetric and the structure is a plate {calculate only [A] and/or [D] matrices and carry out analysis as applicable according to rules (25) and (26)) . (29) If the material is nonlinearly elastic or visco-elastic {use the material nonlinear capabilities of the code} . (30) If a surface is quadrilateral, {use initial mesh strategy as shown in figure 8a) . (31) If a surface is triangular, {use initial mesh strategy as shown in figure 8b} . (32) If a continuous surface of a domain has discontinous boundary conditions (e.g. cracks) leading to very critical singularities {appropriately bias the initial mesh) . 42f Rules for mesh refinement : (33) If the <discretisation error) is high for a zone {use hp-refinement) . (34) If the <discretisation error) is moderate for a zone {use h-refinement} . (35) If the <discretisation error) is low for a zone {use p-refnement} . (a) (b) Figure 8. Initial mesh patterns : (a) quadrilateral surface, (b) triangular surface . 296 B P Naganarayana and G Prathap (36) If the structure is 3-dimensional (refine the element (3-D elements) mesh in all the three principal directions (I-, b- and t- directions)} . (37) If the structure is 2-dimensional {refine the element mesh only in I- and b- directions} . (38) If the structure is 1-dimensional {refine the element mesh only in the I- directions} . 4 .2g Some diagnostics rules based on a posteriori knowledge : (39) If the deflection at any point of the structure is comparable with its minimum characteristic dimension (e .g . thickness of a plate/shell) {repeat the exercise with the geometric nonlinear capabilities of the code} . (40) If the stress at any point of the structure is comparable with the corresponding design stress (design stress is normally taken as equal to yield-strength x factor-of- safety) {repeat the exercise with the material nonlinear capabilities of the code} . It can be noted that both backward-chaining (e.g. rules for selection of dimension and theoretical basis) and forward-chaining (e .g . rules for computation and mesh generation/refinement) are used . Thus an expert system for the present task requires preferably an opportunistic strategy of problem solving. 5 . Conclusions In this paper, several aspects of the finite element analysis of problems in structural mechanics, which require high levels of expertise, are discussed . It is obvious that there is a need for knowledge-based expert system control in modelling complex structures with appropriate finite elements and proper mesh for optimal accuracy at minimum computational cost. A brief review of the ideas and concepts that are emerging in the field and of several efforts in this direction are given. A comprehensive set of rules which can constitute the skeletion of a full-fledged expert system for finite element analysis of a large scale structure is illustrated . References Adeli H 1987 Knowledge acquisition in expert systems for structural design . Proc. In t. SYMP• on Fuzzy Systems and Knowledge Engineering, Guangzhou, Guiyang, China Adeli H (ed .) 1988 Expert systems in construction and structural engineering (London : Chapman and Hall) Aikins J S 1983 Prototypical knowledge for expert systems . Artif. Intel(. 20 : 163-210 Atluri S N, Belytschko T, Oden J T, Reddy J N (eds) 1986 Comput . Methods Appl . Mech . Eng. 55: (Spec . issue on adaptive mesh refinement) Babuska I, Rank E 1987 An expert-system-like feedback approach in the h-p version of the finite element method . J . Finite Elements Anal. Design Babuska 1, Rheinboldt W C 1978 A posteriori error estimates for the finite element method • Int . J . Numer. Methods Eng . 12: 1597-1615 Babuska 1, Zienkiewicz 0 C, Gago J, de Arantes 0 E R 1986 Accuracy estimates and adaPttne refinements in -finite element computations (New York: Wiley) Expert systems and finite element analysis 297 Barr A, Feigenbaum E A 1981-82 The handbook of artificial intelligence (Los Altos, CA :William Kaufman) vol . 1 and 2 Bennett J S, Engelmore R S 1979 SALON : A knowledge-based consultant for structural analysis .Proc . 6th Int . Joint Conf. on Al, Tokyo, pp . 47-49 Bennett J, Creary L, Engelmore R, Melosh R 1978 SALON : A knowledge based consultant for structural analysis, Stanford Report HPP-78-23 (STAN-Cs-78-669), Stanford University Blockley D 1980 The nature of structural design and safety (Chichester: Ellis Horwood) Brachman R J, Levesque H J 1985 Readings in knowledge representation (Los Altos, CA :Morgan Kaufman) Brownston L, Farell R, Kant E, Martin N 1985 Programming expert systems in OPS5 - Anintroduction to rule-based programming (Reading, MA: Addison-Wesley) Charniak E, McDermott D 1985 Introduction to artificial intelligence (Reading, MA: Addison-Wesley) Chen J L, Hajela P 1988 FEMOD : A consultative expert system for finite element modeling .Comput. Struct. 29 : 99-109 Cohen P R, Feigenbaum E A (eds) 1982 The handbook of artificial intelligence (Los Atlos, CA: William Kaufman) vol. 3 Corlett R A, Humphrey A T, Trumpeter J N 1984 NASCON : A knowledge-based consultantfor NASTRAN Marconi Rh C 84/7,esearcentre, tTM4 Daniel W J T 1991 An expert system for validation of displacement and rotation boundary conditions on finite element models . Comput . Struct. 39: 105-119 Desaleux T, Fouet J M 1986 Expert systems for automatic meshing . Reliability methods forengineering analysis (eds) K J Bathe D R J O (S Pii,wenwansea:nerdge) Feigenbaum E 1981 Expert systems in the 80's . Machine intelligence (ed.) A Bond Fenves S 1985 A framework for a knowledge-based finite element analysis assistant . WAM/ASME/Symp. on Appl. of Knowledge-Based Systems to Eng. Analysis and Design, MiamiBeach, Florida Fenves S J 1986 A framework for cooperative development of a finite element modelling assistant . Reliability methods for engineering analysis (eds) K J Bathe, D R J Owen (Swansea :Pineridge) fjellheim R, Syversen P 1983 An expert system for Sesam-69 structural analysis program selection, TR CP-83-6010, Division for Data Technology, Computas, Norway Gasching J, Reboh R, Reiter J 1981 Development of a knowledge-based system for water resources problems. sRi Project 1619, Stanford Research Institute, Menlo Park, CA Harmon P, King D 1985 Artificial intelligence in business (New York : John Wiley) Hayes-Roth F, Waterman D A, Lenat D (eds) 1983 Building expert systems (Reading, MA :Addison-Wesley) Bolt R H, Narayana U V L 1986 Adding intelligence to finite element modelling . Expert SystemGovt. Symp (eds) K N Karna et al. hhizuka M, Fu K S, Yao J T P 1983 Inference procedure under certainty for the problem reduced method Inf cc; 28. . .1401 K, Parker D S 1986 Control over inexact reasoning . Al Expert (Premier Issue): 32-43 (evine R I Diane E D Edl B 1988 Ah,eson compreensive guide to Al and expert systems usingTURBO/PAS CAL(Singapore: McGHill)raw- LOgan J, Genberg V'1987 PLASHTRAN : An expert consultant on two dimensional finite element modelling techni E C 2 19920ques.ng. omput . : -8 wnsky M L 1975 A framework for representing knowledge. The psychology of computer visionn,(-0'. P H Winston (New York : McGraw-Hill) pp. 4510--4519 a C V 1985 Experr systems and fuzzy systems (Menlo Park, CA: Benjamin Cummings)Mn ofbRf P 1986a Blackboard systems: The blackboard model of problem solving and the evolution lackboard architect AJ M 7(2) 3853ure a. g . : - a H P 19866 Blackboard application systems : Blackboard systems from a knowledge rn(ineering perspective Al May 7(3) : 82106. .- DShton N J 1980 Principles of artificial intelligence (Palo Alto, CA : Tioga) pp. 163-190ea T Eisenstadt M 1984 Atifil illi, ricantegence - Tools, techniques and applications New York : Harper and Row) wa H F, u K S Yao J T P 1984 -PERIL-l1:, An ES for structural damage assessment ofexisting structures. Technical Report CE-sTR-84-1 ],School of Civil Eng ., Purdue University 298 B P Naganarayana and G Prathap Prathap G, Naganarayana B P 1991 A 3-D finite element package with expert system and adaptive mesh control : Theoretical manual, NAL Project Document PD-sT-9101, National Aeronautical Laboratory, Bangalore Rank E, Babuska 1 1987 An expert system for the optimal mesh design in the hp-version of the finite element method . Int . J . Numer . Method Eng. 24 : 2087-2106 Reddy D R, Erman L D, Neely R B 1973 The HEARSAY speech understandingg1 system: An example of recognition process. Proc. Int. Joint Conf. on AI, Tokyo, pp. Rehak D R 1986 Artificial intelligence based techniques for finite element program development . Reliability methods for engineering analysis (eds) K J Bathe, D R J Owen (Swansea : Pineridge) Rich E 1983 Artificial intelligence (New York : McGraw-Hill) Rivlin J M, Hsu M B, Marcal P V 1980 Knowledge-based consultation for finite element structural analysis, Report AFWAL-TR-80-3069, US Air Force Flight Dynamics Lab ., W right- Patterson Air Force Base, Ohio Ruspini E 1982 Possibility theory approaches for advanced information systems . Computer 15: 83-91 Shafer G 1976 A mathematical theory of evidence (Princeton, NJ : Princeton University Press Szabo B A Probe : Theoretical manual, NOETtc Technology Corprn ., St . Louis, MO 63117 Taig I C 1986 Expert aids to reliable use of finite element analysis . Reliability methods for engineering analysis r J sN 1970 Theory of elasticity 3rd edn . (N w rYork: McGraw-Hill) Van Me S, Van Melle W 1979 A domain dependent production rule system for consultation programs . Proc . 6th Int . Joint Conf. on Al, Tokyo, pp . 923-925 Weiss S M, Kulikowsi C A 1984 A practical guide to designing expert systems (Totowa, NJ: Rowman and Allanheld)Wilson E L, Itoh T 1986 EXPERT SAP-A computer program for adaptive mesh refinement in finite element analysis . Reliability methods in engineering analysis (eds) K 7 Bathe, D R J Owen (Swansea: Pineridge) Winston P H 1984 Artificial intelligence 2nd edn (Reading, MA : Addison-Wesley) Wong F S 1986 Outline of an approach to inexact reasoning for rule-based survivability ad vulnerability assessment systems . Quarterly report F29601-86-C-0213 (Albuquerque , New Mexico: Air Force Weapons Lab .) Yao J T P, Fu K S 1985 Civil engineering applications of expert systems . Proc . 4th Int . offshore Mechanical and Arctic Engineering Symp . (New York : ASME) vol. 2 Zadeh L A 1974 Fuzzy logic and its applications to approximate reasoning . Inf. Process .- 7 4, Vol. 3. Proc . I FI P Congress (Amsterdam : North Holland) Zienkiewicz 0 C, Zu J Z 1987 A simple error estimator and adaptive procedure for practical engineering analysis, Int . J . Numer . Methods Eng . 24 : 337-357 Zumsteg J R, Flaggs D L 1985 Knowledge-based analysis and design systems for aerospace structures. Applications of knowledge-based systems to engineering analysis and design C L Dym (New York : ASME) AD-10 Sadhana, Vol . 17, Part 2, June 1992, pp. 299-313 . © Printed in India . Approximate factorization scheme for transonic potential flow computations* SUNIL KUMAR CHAKRABARTTY Computational and Theoretical Fluid Dynamics Division, National Aeronautical Laboratory, Bangalore 560017, India MS received 15 March 1991 ; revised 6 April 1992 Abstract. The implicit approximate factorization scheme known as AF2 is investigated here for the purpose of application to the solution of two- and three-dimensional transonic full potential equations in conservative form. The artificial viscosity used by different authors has been deduced, and is discussed in detail . A second-order correction to the implicit artificial viscosity is tested for transonic flow past a Korn aerofoil at both design and off-design conditions . The inviscid transonic flow past different aerofoils, wings and wing-body configurations has been computed using the AF2 scheme and the solutions are compared with experimental and other numerical results . It is shown that the AF2 scheme is fast, and is not sensitive to grid stretching . Keywords. Implicit schemes; approximate factorization schemes; inviscid flows; transonic flows ; computational fluid dynamics . The computation of transonic potential flow past aerofoils, wings and bodies has reached a state of relative maturity due to the development of fast algorithms, such as the Alternating Direction Implicit (AM) and Approximate Factorization (AF) schemes and multigrid techniques, complemented by the availability of high speed computers with vast memories . The inaccuracy introduced by the assumption of u rotationality (which is justified only for weak shocks) in potential flow computation has been accepted in order to save on the computer storage and time that would be required to obtain solutions of the more accurate Euler and Navier-Stokes equations . For many years, Successive Line Over-Relaxation (sLOR) has proved to be a reliable but slow method for solving the compressible potential equation . The use of successive ash refinement does hasten convergence but demands a considerable number of ' Modified version of the paper presented at CAARC (Commonwealth Advisory Aeronautical a arch Council) Specialists Meeting on Computational Fluid Dynamics, held during December, 1988, at National Aeronautical Laboratory, Bangalore . 299 page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 
work_noservp5vng3li2jtosvm6x3ja	----	Karen Levy Development of an Expert System for Classification of Medical Errors D. KOPEC a, K. LEVY a, M. KABIR b, D. REINHARTH c, G. SHAGAS a a Department of Computer and Information Science, Brooklyn College, New York, USA b Department of Health Informatics, Dalhousie University, Halifax, Ontario, Canada c Long Island Jewish Medical Center, New Hyde Park, New York, USA Abstract. The 1999 report published by the Institute of Medicine (IOM) indicated that between 44,000 and 98,000 unnecessary deaths per year occurred in hospitals alone, as a result of errors committed by medical professionals in the United States. There has been considerable speculation that these figures are either overestimated or underestimated. For example, the possibility that they focus on isolated injuries rather than error, or the majority of surveyed respondents did not know what constitutes a (medical) error. These disagreements have led experts to challenge the estimates of patient harm attributable to error, as well as the methodologies used to enumerate them. Of particular concern is the process used in the identification, classification and prevention of medical errors. There have been numerous attempts to develop classifications of medical errors, and currently an abundance of taxonomies exist to describe their mechanism. In previous research, (Kopec, Kabir, Reinharth, Rothschild & Castiglione, 2003) a new taxonomy of Medical Errors was designed by expanding the IOM classification. This model and its extension can be used as a blueprint for future design, development and implementation of an expert system for classification of medical errors. Effective classification can facilitate pattern recognition, and pattern recognition will help in understanding the nature, background and abatement of medical errors. Such a system’s goal will be to perform convincingly as an advisory consultant, exhibiting expertise on a par with and beyond human experts in specified domains. Despite substantial disagreement on the validity of the published figures for fatalities in hospitals in the IOM report, what is of importance is that the number of deaths caused by such errors is nonetheless alarming. The identification and classification of errors in medical care delivery is a very complex process, and this process can be facilitated and simplified by the implementation of an effective classification system. Keywords. Medical errors, expert systems, error theory, CLIPS, taxonomy Introduction Since the publication of the famous 1999 report by the Institute of Medicine (IOM) [1] there has been continued concern about the effective identification and classification of human medical errors. A number of papers have been published in the field, addressing the validity of the manner in which the estimated range of the number of deaths per annum due to medical errors was obtained [2, 3, 4, 5, 6]. Despite disagreements concerning the accuracy of the quoted figures, it is evident that it is impossible to quantify the full magnitude of the challenges to safety with certainty, as the health care sector does not routinely identify and collect information on errors [7]. It is clear that even with discrepancies between the estimates, the mortality rates strongly suggest that effective strategies need to be employed to reliably identify and classify errors. The development of an effective classification system will aid in reducing the occurrence of errors and thereby assist in improving the quality of patient care for the American Health Care System. 1. Research Goal In previous research, the original IOM taxonomy was extended and a new approach to the classification, distribution and updating of medical information was recommended [7]. The goal of this study is to design an Expert System, utilizing this extended taxonomy, which will effectively classify medical errors and serve as a testbed for healthcare practitioners. 2. Taxonomy of medical errors There have been numerous attempts to develop classifications of medical errors, and currently an abundance of taxonomies exist to describe the mechanism behind the types of medical errors. The following etiology of categories of medication errors was given by the American Hospital Association [7]: a) Incomplete patient information b) Unavailable drug information c) Miscommunication of drug orders d) Lack of appropriate labeling e) Environmental factors The five error types most often observed and reported by U.S. family physicians were : [8] a) Errors in prescribing medications b) Errors in getting the right laboratory test done for the right patient at the right time c) Filing system errors d) Errors in dispensing medications e) Errors in responding to abnormal laboratory test results “Errors in prescribing medications” was the only one of these five error types that was also commonly reported by family physicians in other countries [9]. In an influential 1993 report, Kohn et. al. [2] developed a classification of medical errors. This report classifies medication-related errors under “Treatment Error, etc.”, but in our analysis we found that medication-related errors and errors related to clerical procedure are abundant in medical practice. Therefore, we altered the original IOM classification by specifically identifying and addressing these two types of errors, by including the NCC MERP [10] taxonomy. We also classified ‘Errors Related to Diagnosis’ into three clinical subgroups (“delayed”, “missed”, and “wrong”, as opposed to the four subgroups of the IOM). Specifically, we have divided each type into subtypes, numbered according to the NCC MERP [10]. Human error [11] can be subdivided as follows: “A knowledge error referred to as a mistake occurs from inadequate or incorrect information. If the information is correct, but the wrong method of application is chosen, a rule error occurs, termed a lapse. An example of such an error would be an incorrect diagnosis. Finally, the plan may be good, but the performance is faulty, often from distraction or inattention. This is a skill-based error, termed a slip [11]. Reason [11] distinguishes between two kinds of human error – active and latent. Active errors are the ones which are immediately discernible, whereas latent errors are much harder to detect and may require considerable analysis to discover or understand.” Based on our previous research [7], following is the classification of human errors in Medical practice: 1) Errors in prescribing medication: Misuse by incorrect: medication, route, dose, administration Overuse – using too much of a drug or prescribing a drug when not indicated Underuse – failure to provide medication 2) Treatment Procedure(s) – other than by medication 3) Errors related to Clerical Procedures 4) Errors related to Diagnosis Delayed Diagnosis Missed Diagnosis Wrong Diagnosis 5) Preventative errors – delayed or no follow-up treatment are examples 6) Other Communication Failure Administration Problems The above approach for distinguishing between errors will be used in the design of the Expert System for the Classification of Medical Errors. 3. Cases of Errors The study of more than 235,000 error reports submitted in 2003 by 570 health care facilities was the largest ever performed by the U.S. Pharmacopeia [12]. In their findings, as the number of reported errors goes up, the percentage that causes patient harm has gone down. However, the findings that were remarkable are those indicating that electronic prescribing is creating new types of errors. “Computer entry” was the 4th leading cause of errors accounting for 13% (27,711) of the medication errors reported in 2003[12]. In contrast, illegible or unclear handwriting was the 15th leading cause and accounted for 2.9% (6,134) of reported errors[12]. It might be expected that handwriting would move down the list as computerization becomes more widely implemented, however what was occurring was a new type of error. So we can see that information systems and the medical field present a double- edged sword. On one hand they offer the possibility of tremendous improvements in terms of memory capacity, speed and general processing power, but if coupled with arcane data entry systems, serious new problems are created. Can we return to manual data entry? Not at all. More important, is that data entry systems develop “anti-debugging” components to reduce human error. Another example that emphasizes the enormity of the medical error scenario is seen in the article by, Myhre and D. McRure [7, 13], which compares studies of errors in blood transfusions presented as Table 1 below. Table 1. Summary and Comparison of Various Reports of Fatal Errors in Blood Transfusions Figure 1. Pie Chart representation of the information in Table 1. Table 1 illustrates that errors occurred due to drawing specimens from the wrong patient (error by medical technologist), change of specimen in the laboratory (error by laboratory staff or medical technologist), and transfusion of blood into the wrong patient (errors by physician/nurse). Figure 2 below Table 1 illustrates the same data in a pie chart format. The numbers in the columns of the Table should add up to 100% since they are percentages of fatal error in Blood Transfusions, but they did not in the original published table [13]. Hence we have determined that the second row labeled “Other” should have the figures shifted one column to the left, for example, under “Mythre”, the second row labeled “Other” should read 8(10%), under Honig & Bove it should read 4(10%) etc. This type of error can be committed by anyone including a variety of medical practitioners, in addition to physicians. Another incident involved the assessment of the impact on ordering errors when physicians stopped writing patient identifiers on requests for blood transfusions by hand [14]. Physicians, frustrated by the amount of time required to complete forms to order blood, wanted to eliminate the need for their handwritten patient identifiers, which were in addition to such information “stamped” on blood requests. This change was implemented, the blood ordering forms were modified accordingly, and after elimination of the handwritten identifiers in 1997, ordering errors increased from an annual rate of 1 in 10,000 to 6 in 10,000 blood requests by late 1999. Subsequently, clinicians were alerted by newsletter and the rate decreased to 3 in 10,000. However, the error rate did not decrease to its previous level of 1 in 10, 000 requests until mid-2001, approximately 2.5 years after reinstitution of the requirement for handwritten patient identifiers. The conclusion of this study [14] was that an obligatory second entry of demographic identifiers on a blood order requires ordering physicians to be given careful consideration to the identity of the patient receiving the blood transfusion, thereby reducing the likelihood of transfusion of an unintended recipient. 4. The Design, Development and Implementation of an Expert System for Classification of Medical Errors As seen from examination of the above scenarios, errors in the medical field can occur in many different ways, with potentially diverse, wide-ranging and hazardous effects. Just review of the varied errors related to fundamental areas such as medication, diagnosis, treatment procedures and clerical procedures in terms of their number, etiology and possible ramifications, is a complex domain. Consideration of the number of possible permutations of specific elements or actions in a health-related setting that might be classified as error(s), coupled with the large variety of “system” type errors, leads us to conclude that we are dealing with an enormous range of possibilities. Which expert in the medical field will be able to hold in his or her head all the possible combinations of signs, symptoms and treatments that have occurred for all possible medical conditions? Expert systems can be used for effective classification of a diverse range of possibilities, and many have been built to solve different types of problems. These systems are unique in that they can draw conclusions from a store of task-specific knowledge principally through logic or plausible inference [15, 16]. They are also called knowledge-based systems because they contain the same kind of rules used by human experts when they make decisions in their field of expertise [15]. The heart of an expert system is the powerful corpus of knowledge that accumulates during system building. The knowledge is explicit and organized to simplify decision-making; and the accumulation and codification of knowledge is one of the most important aspects [15]. These systems are not locked into any specific decision path and as a result can select from alternative paths in their search for a conclusion [16]. When there are domain experts and a substantial number of rules, more than the human mind can effectively recall with speed and accuracy, such a situation can be remedied by building an expert system. We intend to use the Expert System shell CLIPS to design this system. The system will classify errors based on a set of production or decision rules. Rule-based programming is one of the most commonly used techniques for developing expert systems. In the programming paradigm, rules are used as heuristics or rules-of-thumb, which specify a set of actions to be performed for a given situation. In the event that two rules match a given problem situation, the system will utilize a conflict resolution strategy to best resolve the tie based on the specified decision rules. For example, it could break a tie based on which rule is more specific, or which rule is shorter or based on “refreshing”, that is, rules, which had recently been done after conflict resolution, might not be used again for some time in favor of new rules. Figure 2 illustrates how this Expert System could assist in recognizing and classifying what is considered a typical problem in hospitals, that of not implementing a proper hand-cleansing protocol for medical personnel: Assume a patient had been admitted to a hospital due to complications with influenza, however, after what was considered an acceptable amount of recovery time, the patient showed no signs of improvement, and after assessment was found to be physically worse than he had been upon admission. Some of the questions that the system might ask would be as follows: Q Is the patient male or female? A male Q What is the patient’s age? A 43 Q Did the patient stay overnight in the hospital? A yes Q How many days was the patient in hospital? A 7 Q Give the number of medical staff that were exposed to the patient during his stay? A 22 Q Did the patient come into contact with staff through: medical devices, food trays, medicine dispensation? A yes Q Does the hospital have a hand-cleansing protocol for staff? A no The system, based on the rule constraints that it is designed with might continue with more questions to get more information from the user, and might respond with the following summary: The patient is likely to have contracted a hospital-acquired infection due to the absence of a protocol for the effective practice of regular hand washing by staff. This determination is based on a 75% degree of certainty based on the following elements: Patient was in the hospital overnight Patient was exposed to more than five members of staff Patient was exposed to staff through medical devices, food trays, etc. No hand-washing protocol for staff exists Figure 2. Medical error scenario and possible outcome on running through our Medical Errors Classification System Conclusion and Further Research The identification and classification of errors in medical care delivery is a very complex process, and this process may be simplified by the implementation of an effective classification system.[7] In order to reach a consensus on the classification of medical errors, it is necessary to develop a generally accepted international medical error classification system. Findings have indicated that errors are likely to affect patients in similar ways in countries with similar healthcare systems [17]. With this taxonomy, the major types of errors can be categorized and each major type can then be associated with a specific underlying mechanism. This can explain why and even predict when and where an error will occur [18], which in turn will assist in the generation of intervention strategies for each type of error, and also assist in the reduction and abatement of medical errors. References: [1] Brennan T.A., The Institute of Medicine Report on Medical Errors: Could It Do Harm? Tex Med. 96(6), 13-15, 2000. [2] Kohn, L.T., Corrigan, J.M., and Donaldson, M.S. (eds.), To Err is Human: Building a Safer Health System, National Academy Press, Washington, DC, 2000. [3] Spencer, F.C., Human Error in Hospitals and Industrial Accidents: Current Concepts, Amer. Col. Of Surgeons 191:410-418, 2000. [4] Ballas, E.A., Information systems can prevent errors and quality. J. Am. Med. Inf. Assoc. 8:398-399, 2001. [5] Bates, D.W., Reducing the frequency of errors in medicine using information technology, J. Amer. Med. Inf. Assoc., 8: 299- 308, 2001. [6] Richardson, W.C., The Institute of Medicine report on medical errors, N. Eng. J. Med. 343:663-665, 2000 [7] Kopec, D., Kabir, M.H., Reinharth, D., Rothschild, O. & Castiglione, J.A., Human Errors in Medical Practice: Systematic Classification and Reduction with Automated Information Systems, Journal of Medical Systems, 27(4), 297-313, 2003. [8] Kopec, D., Shagas, G., Kabir, M.H., Reinharth, D., Castiglione, J.A. & Tamang, S., Errors in Medical Practice?: Identification, Classification and Steps toward Reduction, Medical and Care Compunetics 1 (Eds. Lodewijk Bos, Swamy Laxminarayan and Andy Marsh). Proceedings of the 1st ICMCC (International Congress on Medical Care Compunetics), The Hague, The Netherlands, June 2–4, 2004, pp. 126-134. [9] Dovey, S.M., Meyers D.S., Phillips R.L., Green L.A., et al, A preliminary taxonomy of medical errors in family practice. Qual. Saf. Health Care 11:233-238, 2002. [10] National Coordinating Council for Medication Error Reporting and Prevention. NCC MERP taxonomy of medication error. http://www.nccmerp.org/pdf/taxo1999-05-14.pdf [11] Reason, J. Human error, Cambridge: Cambridge University Press; 1990. [12] Robeznieks, A., Data Entry is a top cause of Rx Errors, American Medical News, pp. 14-15, January 24, 2005. [13] Myhre, B.A., and McRure, D. “Human error- A significant cause of transfusion mortality”. Transfusion, 40:879-885, 2000. [14] Moore, SB., Floss, ML., Ordering Blood for the Wrong Patient – Getting Inside the Minds of Ordering Physicians, Mayo Clin Proc. 78:1337-1339, 2003. [15] Waterman, D., A Guide to Expert Systems. [16] Webster, R., Miner, L., ExpertSystems – Programming Problem Solving, 1982. [17] Makeham, M.A.B., Dovey, S.M., County, M., Kidd, M.R., An international taxonomy of errors in general practice: A pilot study, Med. J. Australia 177: 68-72, 2002 [18] Zhang, J., Patel, V., Johnson, T. and Shortliffe, E., A cognitive taxonomy of medical errors, Journal of Biomedical Informatics 37: 193-204, 2004. http://www.nccmerp.org/pdf/taxo1999-05-14.pdf Development of an Expert System for Classification of Medica a Department of Computer and Information Science, Brooklyn C 
work_npfqwig6gfb2deid2yp77s5kpm	----	Table-driven Rules in Expert Systems CUC5-69-83 Alexander Pasik, \larshall Schor A.ugust 17th, 1 983 Departn1ent of Con1puter Science Columbia C niversity .\"ew York City. ~ew York 10027 Department of \lathematical Sciences IB\1 Thomas .J. 'Vatson Research Center Yorkto\vn Heights. ~ew York 10598 Abstract The structure and organization of expert systems can be usefully mod- eled after corresponding human experts. Often this modeling degrades because of insufficient expressive power in production system lan- guages. Relational table techniques provide additional abstraction ca- pabilities and are useful in extending the expressiveness of production system rules; the resulting systems can be easier to build. understand and debug because they can reflect more accurately human methods of reasoning. The number of superfluous rules is reduced by organizing much of the problem domain knowledge in relations in working mem- ory. The relational table methods also provide a tool for the interfacing of knowledge bases and databases. 1. Introduction Production Rule SyJtmu The knowledge base of many expert systems is composed of a set of production rules together with a set of working m~mory ~/ements. The inference engin~ matches the left hand side of each rule against working memory to determine which rules are satisfied. One of the satisfied rules is selected and the actions on its right hand side are performed. These actions in general alter the working memory so that on the next inference cycle, a different set of productions is satisfied. Using table-driven rules can ease the rule writing process in building expert systems. OPSS [Forgy 81 J, a high speed production system in- terpreter, lends itself well to the implementation of the relational table techniques which will be described. Elegance (and lAck of it) in Rule Writing While building an expert system. knowledge engineers attempt to encapsulate discrete pieces of knowledge into single rules. This has se- veral benefits. Human experts in a particular domain often can explain their expertise in small units. each of which is translated into a pro- duction rule. This strengthens the correspondence between human performance and the expert system. making the system easier to un- derstand and debug. Unfortunately, it is often difficult to encode a unit of knowledge into one rule. Sometimes this is inherent in the type of knowledge, but oc- casionally it is the result of the lack of certain structures in production system languages. In many procedural languages. the Case Statement is used to unify a collection of related but conditionally exclusive actions. Intuitively, production rules can be written for each condition in a desired case statement but again this does not reflect the singularity of the unit of knowledge that must be encoded. For example, if the elements of the class person were to be classified into the groups SHORT, MEOIl.JM and TALL. a case statement would express this as demonstrated in Figure 1. case person.he~qh~ of ;< 5.0) person.cateqory SHORT i> 6.0) person.cateqory:a TALL ielsel : person.cateqory :a MEDtUM endcase Figure 1. Case statement to categorize persons by height Since OPS5 does not allow conditional statements in the right hand side of rules, the case statement construct in not available. Instead. three OPS5 rules can be written to encode this prOCess. one corresponding to each clause in the case statement (see Figure 2).1 For all rules shown in this p:lper. English equivalents WIll be supplied. Thus it is not absolutely necesS:lry to understand OPSS synt:lx. .3hor:. -pe rso:,. ~ :r ~ ~ers~n ~as ~~~~n~ < and ~s ~~t ~~:~?~~l=ed, :HE~ modl~Y che ~~=s~n'5 :~~~qcry ~o :e 3H0RT. 'p shor:.-p~rsor. I person .hel~ht < 5.0 .c~t ?) <the-persc~> f --> ~Odl:¥ <:~e-~e~sor.> *~at 5HOR"' I :'a:'':'-rers.jn: !F ~ person has ~e~;~: > a ~nd ~s ~o: Cdteqor~=ed, :H::~! ~od::·:, ::1e ~erson' S ':.3cegory ':,= ::'e 7AL.L . ., :a"l-;:erson I ~erson .he!;nc > ~.': .cae ?) <,:~e-;er3cn> I --> ~Odl~Y <t~e-person> ·ca: !AL:I I :"'ed!:J~-pe=sor.: •. a ~erso~ hdS he~=n: ~ecween ~nd ~ and ~s noe ~3:eqorlzed, :HE~: ::'loJ~':i· o::'e ~I!rson's =at~qory:o te ~E=:·_·~. '., "'edl'.l",-pe~scn I ,person *hel:;ht I<~ ';.0 >= 3.01 teat ,I <:~e-.,erson> I --> Figure 2. Three OPS5 rules to categorize persons by height 2. Using Relational Tables in Rule Writing The Relational Table Technique Relational tables structure information into relations. Each relation is made up of a collection of tuples each of which represents a group of objects that are related to each other. For example. the relation party-()f can be defined over 2-tuples (i.e. doubles) in which the domain of one element is the set of all presidents and that of the other is the set of all political parties. One tuple in this relation can be expressed as party-()f( president = Kennedy ,party= Democrat). The complete re- lation can be expressed in a table. A small portion of the relation is shown below. President Party Reagan Republican Carter Democrat Ford Republican Nixon Republican Johnson Democrat Kennedy Democrat Eisenhower Republican Truman Democrat Roosevelt Democrat There is a natural analogy between relational table constructs and OPS5 working memory. A relation and its domains can be represented by a class and its attributes respectively, and the values of a tuple by values in a working memory element. The tuple party-()f(president= Truman, party:cDemocrat) could appear in working memory as (party-of t president Truman t party Democrat). Working Memory as Long Tum Memory Production rules are often referred to as long term memory. and work- ing memory elements are considered short term. In other words. knowledge of the problem domain resides in the rule base. whereas the state of the particular problem being solved is described in worldng memory. The relational table technique uses working memory as long term memory by encoding much of the problem domain knowledge in re- lations. Much of the inference process is then table driven according to the relations in working memory. This has the effect of decreasing the number of rules which then can more accurately reflect the proc- essing of the human expert. Rule:J Drivm by Relations Using the example of the categorization of persons by their height. the relational table technique improves the clarity of that portion of the expert system (see Figure 3). The relation beight-at would be com- posed of the the three tuples in worldng memory. The knowledge in- volved in this categorization is expressed in one simple rule. driven by the relation between heights and categories. This structure supports additional levels of abstraction allowing greater expressive power. IhelQht-cat 'low 0.0 .hlqh 5.0 .cat SHGRTI IhelQhr.-cat 'low 5.0 'hlQh 0.0 'cat MECIJMl \he~Qht-car 010· ... 6.0 .hlqh 9.0 'cat TALL) cate~crlze-helqht5: !F a person has he1~ht CH> and 15 nOt ~ateqcrlzed. ';ND <H> 1S w1thln he1qht-c3t <C>. THEN ",od~!~,. :~.e per~on' s ca::eqory ::0 be <C>. lP cateqor~z~-he!;ht5 I Iperson .helqht <H> 'ea:; 71 <:he-person> I Ihel~n:;-cat 'lo~ <a <H> .hlqh > <H> .cat <C>l --> i",odl~i' <:ne-po:rson> .cat <C>II Figure 3. The HEIGHT-CAT relation improves the system Table-4riVDI Anribuu Selection In OPS5. relations can be used to encode information about attributes also. For example. if one working memory element is to keep track of how many persons of each height exist. the relation and rule shown in Figure 4 could be used for the categorization. The OPS5 function substl' is used to reference the old value of the attribute <C> of the counter. 2 At a given time, the counter element might be (counter +SHORT 27 +MEDIUM 152 +TALL 112). Each time the rule fires, a person gets categorized and one of the counter attributes gets incre- mented. 2 OPSS's substt has the rollo"'ing rorm:1t and runction. (substr WME ATII ATI2) returns the values or working memory element WME rrom the position designated by attribute A TI 1 through the position designated by allribute A TI2. Thus (substr <the<ounter> <C> <C» returns the: value or the counter at the: alln- bute <C>. :'-:l..;:l:-C:lt . ::..: " ... -" ""c:a: ;-oe :..;::: --:.1:' . -- .:.1 :,.:...:~ -0::1: he:'1r.~-c"t :;'0": .J ·n::';:1 'cat ~~=eQorlze-hel~hcs: =, d ?erson has height <E> lnd :'3 no~ c3te~or~=ed, .;NC o;~~ .2;:~F_':"' :1E: :'_'!,\l TML:', CH> 15 ~:.thLn h~!;n:-cat <C>. :~:::!J ~cd!.::· :he person's ::a:~qcry co be <::>. ';:10 :.ncre~e;.~ ~he <e> ~::r!=u:e 0: :he counter oy ~. ~ ~~;eqo~:=e-he~qht~ I I?erso~ .helqn~ cs> ·c~t ;) <:he-~erson> , I Icoun~erJ <:he-co~n=er> t :'el;:-.":.-ca: - ~.::,.; <~ <H> ·h~;;h > <H> ·c:ac <':>1 --> :::Od1:',' <;he-pe:-son> 'c"t <C>I oo:.nc! COJ> ~suostr <:he-count~r> <0 CC>iJ ',"Odl:';- <;he-counter> '<C> 1 com~u-:e <:1> - 11 1 1 Figure 4. Values in tuples can be attribute names Table-driven State Transitions Another important use of this technique is the management of problem solving stages within the system. Expert systems using the Match problem-solving strategy [Newell 69] are composed of groups of rules. each group designed to solve a sub-problem within the system. When one sub-problem is solved, a transition rule fires which results in making the next set of rules active. This transition may need to be accompanied by the addition or reorganization of relevant information in working memory before the next stage begins. These transition values can be organized in a relation in working memory. An expert system designed to prepare gourmet meals would be built in terms of solving various sub-problems. The stages would include cooking the entree. appetizers and dessert as well as buying the ingre- dients and serving the meal. A relation for the organization of the transition values can reside in working memory as in Figure 5. By using this relation. transition rules between each stage all collapse to the one rule shown. i:.:rans -from s':d.rt .:':) 0'.1,/ ·:1e~d :::oney) t~:r.lns ·from cu,/ ·:0 d;pet!Ze!'s t!'leed .a-l.::grl t:r3ns ·!rom a;pt!t.lzers . ':',:) dessero; ~::eed d-l.nqrl (~~ans '(:-001 dessert ·:'0 en:ree ·:-.eec ~-l.nqr) It=ans +!rom en cree ·:'0 se~'/e -need i'!.acesl It:-ans +from se:-':e -':0 jone ·r..eed :"lone) \:~anqe-st.l<;es :F the ~o~l to solve problem <A> ,5 ~,,~~s::ed. AND there 15 " tr"nS'Clon f:-om <A> :0 <9> 'oIhlCh needs ",,,~e!'"!,,: <C>. ~H::~~ re~ove :.!1e goal :'0 so ~ ve <A> AND create 3 new qoal :0 sO!'le <9> wl:.h d pre~eq~,slte :0 ooc"ln materla1 <C>. 'p ~h"nqe-5c3qe5 I 1;0"1 .~ame <A> 'St3t~5 SATISFIEOI <old> ':rans +!ro~ <A> .co <9> +need <C» --> (re:nove <old>1 mai(e ~Od: • na:ne <8> • ~rereq'': loS 1. ':e <C>, ) Figure 5. The transition states in a gourmet expert system 3. Applications of the Technique In the examples presented. the relational tables were of only a few tuples. The relational table technique has a greater impact on the clarity of a system when the relations are large. An expen system to suggest travel routes using the Manhattan subway and bus systems was built at Columbia University as a project for a course in expen systems [Lerner and Cheng 83]. The system used several relational tables in- cluding one of over 500 tuples to determine intersections of the rapid transit system. Use of this techruque dramatically reduced the number of rules and succeeded in separating the true procedural knowledge from the database which described the transit map of the city. Besides the structural benefits of this method. the relational table tech- nique provides a tool that knowledge engineers can use in another emerging application: the connection of expen systems and databases [Walker 83]. As databases grow and require maintenance. analysts are hired whose responsibility it is to examine large quantities of data and make management decisions based on the information gathered and computed daily by the machines. This data analysis is performed by trained human expens. With the current expert system technology, in- telligent systems can be (and have been) built to analyze these large databases [Stolfo and Vesonder 82]. The relational table techniques can provide knowledge engineers with a tool for the manipulation of ponions of the databases in working memory. Information structured in a relational database can be' mapped into working memory elements in a natural way (one tuple results in one working memory element). The working memory elements are then used by the rule base. The relational table technique described can be used in almost any ex- pen system. Knowledge in a panicular domain is characterized by rules that solve problems. Expen systems have been constructed with a multitude of additional rules which. though syntactically separate. en- code the same knowledge as the original group of rules, varying only in the values of cenain parameters [McDermott 1982]. By building rela- tional tables in working memory consisting of related values and attri- butes. the superfluous rules can be eliminated. Benefits of these techniques include increased ease of preparation and improved read- ability. Systems are easier to maintain because the procedural know- ledge will be condensed. An erroneous result due to a portion of the system can be repaired by the inspection of one rule and a uniform table rather than relying on the complete understanding of a large number of rules that are only slightly different from each other. It has already been demonstrated that relational tables can improve expert systems in gen- eral. With the upcoming demand for expert data analysis. the impor- tance of these methods will continue to increase. References Forgy. C. L. "OPS5 User's Manual" Technical Repon. Carnegie-Mellon University, Department of Computer Science, 1981. Lerner, M. and Cheng. J. "Th~ Manhattan Mapper" Expert Systems Course Project, Columbia Uruversity. Department of Computer Sci- ence, 1983. (unpublished) McDermott. 1. "Rl: A Rule Based Configurer of Computer Systems" Artificial Intelligence, Volume 19(1), pp. 39-88, September 1982. ~ewell. A. "Heuristic Programming: Ill-structured Problems" In J. S. Aronofsky (Ed.). Progress in Operations Research. John Wiley and Sons. pp. 361-414. 1969. Stolfo. S. J. and Vesonder. G. "ACE: An Expert System Supporting Analysis and Management Decision Making" Technical Report. Columbia University. Department of Computer Science. 1982. Walker. A. "Data Bases. Expert Systems. and Prolog" Technical Report RJ 3870. IBM Research Laboratory San Jose. 1983. 
work_nqoovjor2rg7nmedtve6uw7bzy	----	PII: S0004-3702(99)00013-2 Artificial Intelligence 109 (1999) 187–209 Representation of propositional expert systems as partial functions Robert M. Colomb 1 Department of Computer Science and Electrical Engineering, The University of Queensland, Queensland 4072, Australia Received 10 February 1998; received in revised form 6 December 1998 Abstract Propositional expert systems classify cases, and can be built in several different forms, including production rules, decision tables and decision trees. These forms are inter-translatable, but the translations are much larger than the originals, often unmanageably large. In this paper a method of controlling the size problem is demonstrated, based on induced partial functional dependencies, which makes the translations practical in a principled way. The set of dependencies can also be used to filter cases to be classified, eliminating spurious cases, and cases for which the classification is likely to be of doubtful validity.  1999 Elsevier Science B.V. All rights reserved. Keywords: Propositional systems; Machine learning; Decision tables; Decision trees; Knowledge filtering 1. Introduction There is a large class of expert systems whose purpose is essentially to classify cases, for example, to diagnose disease from symptoms. Several different methods are used to represent the knowledge in such systems, and to form the basis of computer implementations, including propositional production rules, decision tables and decision trees. It has been known for some time that these representations are equivalent, in that a system represented in one can be automatically translated into another, exactly preserving the input–output behaviour of the original representation. This issue is addressed in detail by Colomb and Chung [5]. Translatability means that a body of knowledge can be represented in different ways for different purposes. For example, an expert system may be 1 Email: colomb@csee.uq.edu.au. 0004-3702/99/$ – see front matter  1999 Elsevier Science B.V. All rights reserved. PII: S 0 0 0 4 - 3 7 0 2 ( 9 9 ) 0 0 0 1 3 - 2 188 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 built by experts as a set of rules, which might be a natural way for humans to understand the knowledge; but a decision table representation can make it much easier to perform an analysis of the vulnerability of the knowledge base to measurement errors [3]; and a decision tree representation may give a fast bounded time implementation [19,20]. On the other hand, an expert system may be built originally as a decision tree, say by induction from a set of cases [15,16], but be translated into one of the other forms for analysis or understandability. Since these various representations of knowledge are interchangeable, we will call them all decision objects. However, even though the translations can be automatically performed by simple algorithms, there is a severe practical difficulty, namely that the translated representation can be very much larger than the original. Furthermore, most of the constituent parts of the translated object may not be used in practice, even if all the parts of the original are. The main result reported in this paper is to show why this problem of greatly increased size occurs and to show a simple method of representing decision objects which eliminates it. To do this, we first exhibit a unified representation of the various forms of decision objects and present in this representation the translatability results from the literature. We then show how the translated representation becomes inflated, and that this inflation is the result of representing inherently partial functions on the attribute space as total functions. To solve the problem, we show a simple way of representing and estimating the domain of the partial classification function represented by the decision object, and show that by use of the knowledge of the domain, we obtain practical translation among forms of decision objects with little inflation. The paper concludes with a discussion and implications of the results. 2. Forms of decision object A classification system typically processes a case consisting of measurements of a number of variables by assigning a classification to the case. This section is a collection and systematization of results from the literature. The proofs are therefore given somewhat informally since the full detail is available elsewhere. 2.1. Propositions We interpret a variable as a measuring instrument, used by a computer system to monitor a real-world process of some kind. A measurement by a specific variable is the assignment of a specific value to that variable, notionally by the real-world process. The value set belonging to a variable is a discrete set of names, usually describing qualitative properties. A value set must have at least two members. The prototypical case is a Boolean variable with values {true,false}, but other value sets are possible: for example, the variable sex has the value set {male,female}. If a variable refers to a continuous measurement, its value set frequently names the results of a series of relational tests on the measurement: for example, the variable tsh might have the values high,borderline_high,normal,borderline_low,low,missing. There is no limit in principle R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 189 to the cardinality of a value set. This notion of variable with finite value set is widely applicable: in particular, it is the basis for most expert systems. In an application, there are generally a number of variables. More formally, there is a set X={xi}, consisting of at least one variable. This set of variables is interpreted as a view of the real-world process. Each variable xi has a value set Vi. A state of the real-world process as measured by the system is an assignment xi =vj for vj in Vi for all variables. This measurement of state is referred to as a case. The set X will be called the variable set associated with the system under consideration. An elementary proposition consists of a variable and a value from the value set belonging to that variable, designated xi = vij for vij in Vi. We will designate by pi the set of elementary propositions pij associated with variable xi. The elementary propositions associated with the variables {xi} will be referred to as case elementary propositions. A case is represented as a proposition by a choice from {pi}, that is exactly one member of each pi. The ith choice will be designated ci and a case c will be represented by the set {ci}. Each ci is a case elementary proposition associated with variable xi. Classification is a function whose domain is the set of possible choices from {pi}, and whose range is a set of elementary propositions whose value set is a (typically small) finite set D={dj}. The classification assigned to a case will be represented as a proposition of the form class =dj for dj in D, where class is a reserved variable name. Elementary propositions associated with the variable class will be referred to as classification elementary propositions. Example 1. Consider a system with two variables x1 (sex) with value set {female,male} and x2 (pregnant) with value set {true,false}. The case elementary propositions asso- ciated with each variable are p1 = {sex = female,sex = male} and p2 = {pregnant = true,pregnant = false}. There are four possible choices from {pi}, two of which are {sex = male,pregnant = false} and {sex = male,pregnant = true}. D is the set {normal,plausible,extraordinary}. Assume that the classification assigned to the two cases exhibited are, respectively, the classification elementary propositions class = normal and class = extraordinary, while the classification assigned to the two other choices are both class= plausible. Definition 2. An elementary proposition is binary if its value set has cardinality 2. Proposition 3. The negation of an elementary proposition is a conjunction of elementary propositions. Proof. An elementary proposition is a formula of the form xi =vij . If it is not true that xi =vij , then it must be true that xi =v, where v∈Vi\vij , so that ∼(xi =vij)≡ ∨ k 6=j xi =vik, where k ranges over the indexes of the members of the value set Vi. 2 Corollary 4. The negation of a binary elementary proposition is a binary elementary proposition. 190 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 2.2. Rules A rule system form of decision object is a collection of production rules. Production rules are constructed from elementary propositions, classifications, and a distinguished set of propositions A={a} designated intermediate elementary propositions. Intermediate elementary propositions are distinguished by having variables distinct from the variables which can occur in cases, and not including the reserved variable class. The antecedent of a rule is a conjunction of case or intermediate elementary propositions, or the negations of either, while the consequent is either an intermediate or classification elementary proposition. There is a single consequent in each production rule, so that a rule system is a collection of propositional Horn clauses. A production rule system will be called well-formed if its dependency graph [2] is acyclic. The dependency graph for a system of production rules has one node for each rule. An arc is defined with source N1 and target N2 if the consequent of the rule at N1 appears in the antecedent of the rule at N2. Note that this requirement is stronger than stratification as defined in the cited work. That work labels an arc “positive” or “negative” according to whether the consequent of N1 appears as a positive or, respectively, negative literal in the antecedent of N2. A stratified system has no cycles including a negative arc, but allows cycles all of whose arcs are positive. The semantics of a propositional production rule based decision object is based on datalog under negation as failure. A set of propositional Horn clauses is trivially a datalog intensional database (IDB), since there are no variables. The extensional database (EDB) consists of a set of predicates each of which corresponds to one of the elementary case propositions. Each of the EDB predicates either contains one tuple, which consists of the token “true”, or is empty. The standard stratified naive bottom-up evaluation procedure populates the IDB predicates by propagating the “true” token. The classifications assigned to a case are the classification elementary propositions which are not empty. All this is a straightforward application of the principles of datalog as described, for example, by Ullman [21,22]. Example 5. We will re-cast Example 1 as a datalog system. The IDB consists of the three Horn clauses: (class = normal) :- (sex = male),(pregnant= false) (class = extraordinary) :- (sex = male),(pregnant= true) (class = plausible) :- (sex = female), and assume the EDB represents the case of a pregnant male: (sex = male)={true} (sex = female)={} (pregnant= true)={true} (pregnant= false)={}. The datalog evaluation produces the following population of IDB predicates (class = normal)={} (class = extraordinary)={true} (class = plausible)={}. R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 191 The datalog semantics of rule systems shows that the well-formedness criterion which excludes recursion entirely involves no loss of generality. Theorem 6. For every stratified propositional IDB P there is an IDB with an acyclic dependency graph which has the same perfect model. Proof. Based on the datalog naive evaluation procedure. This procedure starts with the given EDB and no population for any IDB predicate, and proceeds from stratum to stratum (making sure that clauses containing negated predicates in their antecedents are not evaluated until all the clauses with those predicates as consequents are fully evaluated). Each step evaluates all clause bodies and possibly generates tuples for some of the IDB predicates. This evaluation is called applying the T operator. The initial population is named T(0), and the population after n steps is called T(n). It terminates when a fixed point is reached: T(n+1)=T(n). T(n) is the perfect model. 2 Lemma 6.1. In a propositional IDB, each predicate generates tuples only once. Proof. Each predicate in the IDB can have at most one tuple. Therefore once a clause puts a tuple in a predicate, it is fully evaluated. 2 Even though a propositional predicate may be defined in several clauses, at most one clause firing fully evaluates the predicate. Lemma 6.2. For a given EDB, the evaluation procedure selects an acyclic subset of IDB clauses, which are the only clauses to generate tuples. Proof. A clause can generate tuples only if all of its positive antecedent literals have populations, and none of its negative antecedent literals can possibly have populations. At each step n of the evaluation, include the clauses which generate tuples in T(n) in the required subset. Designate the clauses added at T(n) by C(n), and the union up to n of the C(n) by S(n). Lemma 6.1 insures that a clause can be in at most one C(n). That the required subset at the fixed point is acyclic can be seen by induction. Clearly S(1) is acyclic, since the only clauses which can generate tuples in T(1) must have antecedents entirely in the EDB. If S(n) is acyclic, then C(n+ 1) includes only clauses all of whose antecedents contain only positive literals defined in S(n) and negative literals which are empty and which have been fully evaluated, so S(n+1) can have no cycles. 2 Each EDB e generates an acyclic subset of the IDB by Lemma 6.2. Designate that subset by S(e). Lemma 6.3. If e is an EDB and S′ is an IDB which generates no tuples when evaluated with EDB e, there is an IDB with an acyclic dependency graph which has the same perfect model as S′ when evaluated with EDB other than e and the same perfect model as S(e) when evaluated with EDB e. 192 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 Proof. By construction. If S(e) generated no tuples when evaluated with EDB other than e, and S(e) had no intermediate elementary propositions in common with S′, then S′∪S(e) would be the desired IDB. In general, this is not the case. We need to isolate S(e) from other EDBs and to make sure all intermediate elementary propositions in S(e) do not occur in any other clause in the IDB. It is easy to specify an additional set of rules which will serve to identify EDBs. A propositional EDB is a pattern of presence or absence of a token in a fixed collection of predicates. For EDB e, we can create a rule with consequent i(e) which is true if the EDB is the pattern e and empty otherwise. The antecedent of identification predicate i(e) has a positive literal for every proposition having a token in EDB e, and a negative literal for every proposition lacking a token. Also, since all elementary propositions assigned values in S(e) are assigned values by clauses in S(e) together with EDB e, we can replace the names of the variables associated with all intermediate propositions in S(e) with names tagged by e. We will call this process isolation of an IDB. For S(e), construct an IDB S′(e) by including i(e) as a conjunct in the antecedent of all clauses in S(e). The only IDB which can generate tuples when evaluated with EDB e is S′(e). Further, S′(e) has the same perfect model as S(e) (excluding the introduced identification predicates). If S′(e) is also the isolation of S(e), then S′ ∪S′(e) is IDB required for the lemma. 2 The identification predicate is what really does the job. The only reason for worrying about isolation is that the dependency graph is defined in terms of the propositions which are the consequents of clauses, not clauses in isolation. Even though the introduction of the identification predicates is enough to keep S′(e) from any influence in the evaluation of any other EDB, the isolation is required to separate the dependency graph. Proof of Theorem 6. Construct the finite set (cardinality N) of all possible choices from {pi} as the set of possible cases C. Impose an ordering on C, yielding a sequence of cases with the kth member identified by ck, k6N. For each ck construct S(ck). Construct S′′(0) as the empty IDB. For each k > 0 construct S′′(k) by the application of Lemma 6.3 to S′′(k−1) and ck. S′′(N) is the required IDB, which has an acyclic dependency graph and has the same perfect model as P for every possible case. 2 Theorem 6 is of course not a practicable procedure to convert recursively defined propositional production rule sets to nonrecursive production rule sets. However, it shows that the use of recursion adds no expressive power. Therefore it is reasonable to suggest to the programmer of a recursively defined propositional production rule set that the set be re-implemented to avoid recursion. Definition 7. Two decision objects are equivalent if they have the same perfect model for all EDBs. R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 193 2.3. Decision table A decision table consists of a two-dimensional array of cells. Associated with each row in the array is a classification. The ith cell in a row is a nonempty disjunction of elementary propositions from the set of elementary propositions associated with the ith variable. A row in a decision table can be viewed as a rule whose antecedent is a conjunction of cells, and whose consequent is a classification. Therefore, a decision table can be viewed as a conjunction of row rules. Example 8. The following is a decision table equivalent to the system in Example 5. Row Sex Pregnant Classification 1 sex= male pregnant = true class = extraordinary 2 sex= male pregnant = false class = normal 3 sex= female pregnant = true ∨ class = plausible pregnant = false Definition 9. A cell is a don’t care cell if it is the disjunction of all the elementary propositions in the set associated with its column. The cell in Example 8 row 3 associated with the variable pregnant is a don’t care cell. Theorem 10. An acyclic rule system is equivalent to a decision table. Proof. (Detailed in [5].) Essentially done by successively unfolding (partially evaluating) the production rules, replacing the intermediate propositions by formulas consisting entirely of case elementary propositions or their negations. The result is a conjunction of rules whose consequents are classification elementary propositions. The result is expressed in clausal form. These clauses are rows in a decision table, with the exception that a given row may contain no elementary proposition from the set associated with a given variable. If this is the case, the antecedent of the clause is augmented by a conjunct consisting of the don’t care cell associated with the variable, which evaluates to true. 2 Definition 11. A decision table is unambiguous if no assignment of truth values to elementary propositions can assign true to more than one classification elementary proposition. An ambiguous decision table can arise from several sources. Decision tables arising from Theorem 10 can be ambiguous due to errors in the rule set. One motivation for using Theorem 10 is to check the set of rules for ambiguity, which is very difficult to see in the rule representation. It can also be ambiguous only for impossible cases. Insulating a system from impossible cases is one of the primary motivations of this paper. Sometimes 194 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 decision tables are used to identify situations in which there is normally one classification, but sometimes can have more-multiple disease states for instance. In this situation, we can without loss of generality replace the variables in the classification elementary propositions with Cartesian products of variables, and the values by Cartesian products of values, so that the form of the elementary proposition is preserved and the table becomes unambiguous, preserving the remainder of the present theory. Definition 12. A decision table is complete if every assignment of truth values to elementary propositions assigns true to at least one classification elementary proposition. Observation 13. A rule system can be equivalent to an ambiguous and/or incomplete decision table. See [5]. 2.4. Decision tree A decision tree is a tree consisting of nodes of two classes: deciding nodes and leaf nodes. Each deciding node is associated with a variable and each leaf node is associated with a classification elementary proposition. Each arc has as its source a deciding node, and is associated with a case elementary proposition from the set associated with its source (associated arc proposition). Our formulation is somewhat unusual in that each leaf node is associated not only with a classification elementary proposition but with a proposition composed from case elementary propositions. The operational semantics is that a leaf node fires on a case if all the arc propositions in a path to the root are consistent with the case, and so is the leaf proposition. A decision tree may therefore fail to classify a case. Example 14. The following is a decision tree equivalent to the table in Example 8. Node 1 sex—female—Leaf 2 (pregnant= true)∨ (pregnant= false)→ (class = plausible) | male Node 3 pregnant—true—Leaf 4 true → (class = extraordinary) | false Leaf 5 true → (class = normal). Definition 15. A decision tree is complete if each deciding node is the source of arcs associated with all of the elementary propositions associated with its variable, and all of its leaf propositions are equivalent to true. Definition 16. A path proposition is associated with a path from the root to a leaf, and is the conjunction of all the associated arc propositions conjoined with the leaf proposition. Proposition 17. A conjunction of elementary propositions is either inconsistent or a choice from the sets of elementary propositions associated with a subset of the variables. Proof. Two distinct elementary propositions associated with the same variable are inconsistent. Therefore, any consistent conjunction contains at most one elementary proposition associated with each variable. 2 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 195 Proposition 18. A path proposition is equivalent to a decision table. Proof. The leaf proposition can be expressed in disjunctive normal form. The path proposition is therefore the conjunction of each of the leaf proposition disjuncts with all the arc propositions. By Proposition 17, each of the resulting disjuncts contains at most one elementary proposition associated with each of the variables associated with the deciding nodes in the path and the variables in the path propositions, or is inconsistent. If inconsistent, the disjunct can be removed. Finally, each disjunct can be augmented with don’t care cell propositions associated with any variable lacking any associated elementary proposition in the disjunct. 2 Theorem 19. A decision tree is equivalent to an unambiguous decision table, which is complete if the tree is. Proof. Each distinct path through the decision tree is equivalent to a decision table each row of which is associated with the path’s leaf classification elementary proposition. Each path proposition is inconsistent with every other path proposition, since every pair of paths must share at least one deciding node and the arcs in each path whose source is that deciding node have different and therefore, inconsistent associated arc propositions (which are are case elementary propositions with the same variable and a different value for each arc). The union of the decision tables from all paths is a decision table, which is unambiguous, since each leaf and therefore, path has a single classification. If the decision tree is complete, then each deciding node partitions the set of possible cases, so there is a partition of the set of possible cases where each partition is associated with a leaf. The leaf proposition is true, so that every case in that partition is consistent with the decision table equivalent to the path proposition associated with the leaf. Each case in the population of possible cases is therefore, contained in one of the parts of the partition, and is therefore, consistent with at least one row in the decision table. 2 Theorem 20. An unambiguous decision table is equivalent to a decision tree, which is complete if the table is. Proof. By induction on the number of attributes associated with the table. We adopt the convention that a decision table with no attributes has a single row with the propositional value true, and therefore a single classification. Basis step: If any of the following is the case, then the induction terminates with the indicated action: (1) The number of attributes is zero. Create a leaf node whose classification is the table’s single classification elementary proposition, and whose leaf proposition is true. (Depends on the table being unambiguous.) (2) The number of rows is zero. Create a leaf node whose classification elementary proposition is arbitrary and whose leaf proposition is false. (3) The number of distinct classifications is one. Create a leaf node whose classification elementary proposition is the table’s unique classification, and whose leaf proposi- tion is the disjunction of propositions created from each row by the conjunction of the cell propositions. 196 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 Induction step: There is at least one attribute, at least one row, and at least two distinct classifications in the table T . (1) Select an attribute a by some method (this point is discussed below). (2) Create a deciding node associated with this attribute. (3) For each v in the value set of a, create: • An arc whose source is the deciding node created and whose arc proposition is a=v. • A decision table T(a,v) derived by removing the cell associated with a from each row of T in which the cell associated with a is consistent with a=v. T(a,v) has one fewer attribute in its associated attribute set than T , so the induction proceeds. The tree is not complete unless every deciding node is the source of an arc associated with each elementary proposition associated with its variable. The arcs are created by part (3) of the induction step. If one of the elementary propositions is missing, then the table has no row consistent with that elementary proposition and no case with that elementary proposition as a conjunct will be classified. Similarly, if the leaf proposition does not evaluate to true, then no case consistent with the path proposition conjoined with the negation of the leaf proposition will be classified by the table, so the table must be incomplete. The tree is therefore, complete if the table is. 2 From a practical point of view, the key to the algorithm developed from this theorem is in the method of choosing an attribute in part (1) of the induction step. Shwayder [19,20] uses a heuristic based on equalizing the number of rows in the T(a,v) (maximum dispersion), giving one of the standard algorithms for translating from a decision table to a decision tree. Quinlan [16] uses a heuristic based on minimizing the maximum number of distinct classifications in the T(a,v) (maximum entropy gain). Theorem 20 with the maximum entropy gain selection procedure gives a variant of the ID3 algorithm. Theorems 19 and 20 show that decision trees and unambiguous decision tables are exactly equivalent. 2.5. Cases Quinlan’s 1986 paper [16] is based on the concept of a training set of cases, and is intended to construct a decision tree from a training set. Definition 21. A training set is a set of cases. Associated with each case is a classification. Proposition 22. A training set is a decision table. Proof. By inspection. 2 Definition 23. A training set is noise-free if it is unambiguous. Observation 24. Theorem 14 applied to a noise-free training set using the maximum entropy gain selection procedure is the ID3 algorithm. R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 197 2.6. Ripple-down rule trees Another form of propositional knowledge representation is the ripple-down rule tree [7]. A ripple-down rule tree is a binary tree consisting of a set of decision nodes and a set of leaf nodes. A decision node is associated with a conjunction of elementary propositions. A leaf node is associated with a classification elementary proposition. There are two arcs whose source is a decision node. The true arc is taken if the node proposition is consistent with a case, the false arc otherwise. Example 25. The following is an example of a ripple-down rule tree equivalent to the decision tree in Example 14: Node 1 if (sex =male)&(pregnant= true)—true—Leaf 2 (class= extraordinary) | false Node 3 if (sex = female)—true—Leaf 4 (class= plausible) | false Leaf 5 (class = normal). Definition 26. An rdr path proposition associated with a path through a ripple-down rule tree is constructed from a path from the root to a leaf. If the path takes a true arc, then the proposition associated with the decision node is a conjunct. If the path takes a false arc, then the negation of one of the conjuncts in the associated proposition is a conjunct. Proposition 27. Without loss of generality, an rdr path proposition is a decision table row. Proof. An rdr path is a conjunction of disjuncts. Each disjunct is a disjunction of elementary propositions associated with the same variable, since disjunctions are either single elementary propositions or are negations of elementary propositions (Proposition 3). By Boolean algebra, the rdr path can be represented as a disjunction of conjunctions of elementary propositions. By Proposition 17, each disjunct is either a choice from a set of variables or inconsistent. An rdr path is therefore, either inconsistent or a conjunction of disjuncts of elementary propositions associated with distinct variables. If the rdr path lacks elementary propositions associated with a given variable, the corresponding don’t care proposition can be conjoined. The classification is the classification associated with the leaf node. 2 Theorem 28. A ripple-down rule tree is equivalent to an unambiguous decision table. Proof. The decision table is the union of all rdr path propositions. Note that there can be many rdr path propositions associated with a single leaf, if the path from the root to the leaf takes any false paths whose source decision node is a nontrivial conjunction of elementary propositions, so there may be many rows in the decision table with the classification associated with a given leaf. However, rows with classifications from distinct leaves are inconsistent, since there must be a decision node in common, with one path taking the true arc and one the false arc. The table is therefore unambiguous. 2 198 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 Section 2 has presented a framework in which the equivalences among the various forms of decision object (rules, decision tables, decision trees, cases and ripple-down rule trees) can be easily seen. Using these equivalences, one can develop a classification system using a form of decision object convenient for the developers, then translate it into other forms for purposes of analysis and execution. 3. What is wrong with this picture? The idyllic theory described in the previous section encounters problems when put into practice. Section 3 describes the problems encountered and analyses their cause. Some terminology is needed. We will say that a decision object has a number of parts. What a part is depends on the form of decision object. For a set of rules, a part is a rule; for a decision table, a row; for a decision tree or ripple-down rule tree, a leaf. The size of an object is the number of parts it has. A decision object is designed to solve a practical classification problem. We would expect that when we build a decision object, no matter what its size, that all of its parts will be used in practice. A part that is not used represents a cost giving no benefit. We will say that a decision object is compact to the extent that all of its parts are used in practice (measured, for example, by monitoring the decision object over a long period of use). Generally speaking a decision object in its constructed form will be compact. The problem we encounter is that when an object is transformed, the new object can be very much larger than the original, and much less compact. We designate this phenomenon as inflation. For example, the transformations from Theorems 10, 20 and 28 were applied to a well-known expert system Garvan ES1 [11], which constructed clinical interpretations of thyroid hormone assays in a hospital pathology laboratory. It was in clinical use from about 1984–1990, and was applied to many thousands of cases per year. The compactness of the various forms was tested using a sample of 9805 cases, representing more than one year’s use towards the end of the system’s life when its accuracy was more than 99%. Garvan ES1 was constructed as a set of production rules, of which at the end of its life there were 661, all of which were exercised by the set of cases. The system was, as one might expect, very compact. Applying Theorem 10, the production rules were transformed into an equivalent decision table which had 5286 rows. Only 653 of these rows were fired by any of the year’s cases. The transformed object is much larger than the original, and much less compact. Theorem 20 was then applied to the 653-row subset of the transformed decision table consisting of the rows which were fired by any of the year’s cases, using the maximum dispersion selection procedure, resulting in a decision tree with 30,693 leaf nodes, of which only 807 were fired by any of the 9805 cases. The inflation is much worse than in the previous situation. Finally, Compton and Jansen [7] produced a ripple-down rule version of the knowledge base (called Garvan RDR) which had 557 leaves in its ripple-down rule tree, all of which were necessarily exercised in practice since the tree was constructed in a process based on the sample of cases. Application of the algorithm derived from Theorem 28 would result R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 199 in a decision table with 481,803,735 rows. One would expect, of course, that some of these rows would be inconsistent or would be subsumed by other rows. Nevertheless, only 680 of the rows fire on any of the cases, giving an extreme inflation. Inspection of the proofs of the theorems shows how the transformed object gets bigger than the original. Theorem 10 translates a set of production rules to a decision table by progressively substituting the antecedents of intermediate propositions for their occurrences in other rules. If the antecedent is a disjunction, or appears negated in a rule, or both, a rule absorbing the intermediate proposition gets broken into several rules whose antecedents are conjunctions of elementary propositions. If the absorbing rule has itself an intermediate proposition as a consequent, the rules multiply. The algorithm for transforming rules to decision tables derived from Theorem 10 therefore produces a result whose size is exponential in the length of a chain of intermediate propositions. (Garvan ES1 had an average of three intermediate propositions between the measurements and the classifications.) Translating from a decision table to a decision tree, part (3) of the induction step of Theorem 20 duplicates rows of the decision table where the cell associated with the selected variable contains a disjunction of elementary propositions. The algorithm derived from the theorem therefore, produces a result whose size is exponential in the number of variables. (Garvan ES1 had 34 variables, and the decision tree had an average path length from the root to a leaf of about 14.) Finally, the number of rows in the decision table equivalent of a ripple-down rule tree is more than the product of the number of conjuncts in the propositions associated with the decision nodes in the longest chain of false arcs. So we can see how inflation happens, but it remains to see why it happens. After all, the original decision object is created compact. We can see how it gets large, but where do all the unused parts come from? To see this, we step back to the black box view of a classification expert system. The system can be seen as monitoring some real-world process which generates cases. The system observes the cases through its variables. Each distinct case is represented by a choice of values for the variables. The variables can therefore be seen to determine a space defined by the distinct possible choices. We call this space the attribute space for the process. It depends solely on the variables and their value sets. The attribute space for a process can be quite large. A convenient measure of its size is the number of bits necessary to represent a case, disregarding the statistics of the process. If all the variables are Boolean, the size of the space is the number of variables. Otherwise, the size of the space is the total of the log to the base 2 of the cardinalities of the value sets. For example, Garvan ES1 has 34 variables, whose value sets range in cardinality from 2 to 6. Its size is 46 bits. If Garvan ES1 has an attribute space of 46 bits, there are 246 ≈ 1014 possible distinct cases. At 105 cases per year and if all cases are different, it would take 109 years before all the cases were encountered. In practice, as one might expect, there are fewer than 4000 distinct cases in the year’s history, some of which occur more than 100 times. Where the attribute space is large, one would expect that the process has many variables which are contingent on others taking specific values (pregnancy is a relevant variable only if the sex of the patient is female, for example). Also, values of some variables occur in patterns 200 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 depending on the values of others (the hormone profile of a pregnant woman or a child is different from a male adult, for example, and there are different characteristic disease states). In general, where the attribute space is large one would expect that almost every combination of values would produce a case the domain experts would regard as absurd. Even if there were 106 possible valid cases for the Garvan ES1 process, one would have to generate 108 cases at random before the expert would find one that makes sense. This observation suggests that in systems with large attribute spaces the real-world process is confined to a very small region. The knowledge of the experts is confined to this small region, which we might call the region of experience. The expert system classifies cases, so can be seen as a function taking the region of experience to a space of classifications. This accounts for the compactness of the decision objects constructed. These decision objects are, however, constructed as total functions on the attribute space, not as partial functions. The strategy works essentially because by definition the process being monitored never generates an impossible case. However, this strategy leaves the expert system vulnerable to errors or maliciousness. The famous experiment in which Lenat answered Mycin’s questions as for a dead person, and Mycin happily diagnosed a specific course of treatment, is almost certainly a case where the input lay outside Mycin’s region of experience. Decision objects get their compactness essentially by liberal use of don’t care propositions, or by conjunctions of elementary propositions which fail in limited ways. These elements are found and expanded by the transformation algorithms. Inflation is a side effect of the practice of constructing classification systems as total functions. 4. A cure for inflation If inflation is a side effect of the construction of classification systems as total functions on large attribute spaces, then it would seem reasonable to look for a cure by defining the systems as partial functions. If one could get a characterisation K of the region of experience associated with a process generating cases to be classified, then we could use K to prune the transformed decision objects as they are being created. In Theorem 10 (rule → table), we would keep only those disjuncts of the definition of an intermediate proposition which are consistent with K. In Theorem 20 (table → tree), part (3) of the induction step would only make copies of rows which are consistent with K. In Theorem 28 (rdr tree → table), we would keep only those rdr paths which are consistent with K. A good estimate of K would not only control inflation, but would act as a filter detecting cases which are either spurious or perhaps legitimate but outside the experience of the domain experts. A natural way to represent K is by a set of constraints stating that the values of certain variables are determined by the values of others. These sort of constraints are partial functional dependencies (PFDs). The variables are in general independent, but if the set of variables in the domain of the dependency take on a particular combination of values, then the value of the variable in the range of the dependency is fixed. For example, in general the variables sex and pregnant are independent. However, if sex takes the value male, then R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 201 pregnant takes the value false. Similarly, if pregnant takes the value true, then sex takes the value female (at least in the experience of Garvan ES1). Partial functional dependencies could be constructed by the domain experts as part of the process of building a decision object. Certainly, the sex/pregnant situation in the previous paragraph is pretty straightforward. However, many expert systems are constructed by induction using various methods to produce various forms of decision object. If a body of cases is available, then it is plausible to induce a set of partial functional dependencies. In fact, one can see a priori that it must be possible to induce a set of partial functional dependencies. If the set of cases available is much smaller than the attribute space, then by assumption there are very many combinations of variable values which do not occur. The problem is therefore, not whether we can induce a set of PFDs, but whether we can induce a set of practical size which gives a sufficiently tight upper bound on the region of experience, whether the induction can be done at an acceptable cost, and whether the set induced is statistically reliable. There are a number of algorithms available for estimating sets of PFDs. One is from the method of rough sets [13] which looks for rules resulting from value reducts, taking each attribute in turn as a decision attribute (set of classifications). Another comes from the data mining literature [1]. In data mining terminology, a PFD is an association with 100% confidence (no counterexamples). We want PFDs with a small number of elementary propositions in their antecedents. There are two reasons for this. The first is that a single PFD with m elementary propositions in the antecedent and consequent taken together constrains the attribute space more the smaller m is. If all the variables are Boolean and there are n variables, then a single constraint reduces the attribute space by a factor of 2n−m+1. The second reason for wanting PFDs with few propositions in their antecedent is that the algorithms for computing PFDs are exponential in the number of propositions in the antecedent. Small PFDs are therefore, both stronger and practical to compute. Finally, we want PFDs which are statistically reliable. In the absence of a good theory of the statistics of PFDs, it seems reasonable to look for PFDs which have a large number of positive examples (in data mining terminology, associations with high support). This prevents rare cases from contributing possibly spurious PFDs. The existence of a good set of PFDs is a question which can only be settled experimentally in particular situations. To gain experience, we estimated PFDs for Garvan ES1 from the set of 9805 cases (of which 3856 are distinct), using the rough sets library. 2 We applied the algorithm to the set of 3856 distinct cases, so that duplicated cases did not contribute to support. This process produced 382 PFDs with a single elementary proposition in their antecedent at a support level of 1 and 332 such PFDs with a support level of 10. Further, there were 13384 PFDs with two elementary propositions in their antecedent at a support level of 1 and 5858 such PFDs at a support level of 10. Since we are here concerned with the theoretical basis and computational feasibility of working with sets of PFDs, but want statistically reliable constraint sets, we chose the upper support level (10 examples with no counterexamples). We will call these two sets of constraints Garvan K1 2 Institute of Computer Science, Warsaw University of Technology, Warsaw, Poland. 202 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 and Garvan K2, respectively, and in general a set of constraints with one antecedent K1, and with two K2. Further experience might suggest a different support level, but at least in this case the number of rules computed does not vary greatly with differences in support level so it would make little difference in the following. This shows that it is possible, in at least one application, to induce a reasonable set of small PFDs. The remaining question is how strong they are—the size of the region of experience defined by them, and what difference they make to the inflation problem. This requires some more theory. Definition 29. A proposition covers a case if the case is consistent with the proposition. Definition 30. A variable is missing from a proposition if the proposition contains no elementary proposition associated with it. One way to compute the strength is to use a method of Cragun and Stuedel [8]. In our terminology, they show that if a proposition is in disjoint disjunctive normal form (each disjunct is inconsistent with every other) the number of cases covered by the proposition is the sum of the number of cases covered by each disjunct. The number of cases covered by each disjunct is one if the disjunct is the conjunction of propositions associated with all variables, and the product of the cardinality of the value sets of the missing attributes otherwise. Unfortunately, this method is of limited utility, since K is constructed in conjunctive normal form, and the process of conversion from conjunctive to disjunctive normal forms is exponential in the number of conjuncts. Garvan K1 requires the computation of 2332 ≈ 10100 conjuncts, while Garvan K2 requires 35858 ≈ 102610 conjuncts, clearly beyond the bound of computational feasibility, even for Garvan K1 only. One might think that there might be considerable redundancy in K, in that some of the constraints might be deduced from others. This is particularly easy to test in the case of K1, since the redundancy can be viewed as a case of transitive closure. If a→b and b→c are both in K1, then the algorithms used will ensure that the redundant a→c is also. Applying this method to the 332-conjunct Garvan K1 reduces it to 296 conjuncts, taking the cost of conversion to disjunctive normal form from 10100 to 1089, still well beyond computational feasibility. Although there may be problems where the Cragun and Steudel approach is feasible, it is clear that there are some in which it is not. A Monte Carlo method based on constraint propagation is more generally applicable. The basis of the Monte Carlo method is the generation of random cases. A naive approach would generate a population of random cases and count the cases which satisfy K. This is not generally practical, since we are expecting the region of experience to be a tiny fraction of the entire attribute space—in the case of Garvan ES1, the region of experience might be as small as 10−10 of the attribute space. If we augment the generation of random cases with propagation of constraints, we obtain a feasible method of generating an estimate of the size of the region of experience. R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 203 Algorithm 31. Estimate the strength of a set of constraints K. The algorithm is based on generating cases by successively choosing random values for variables and propagating these values, until all variables are assigned values. The average number of random bits needed to generate values is an estimate of the size of the space covered by K. A variable which is not assigned a definite value will be referred to as an open variable. Note that the method may restrict the possible values a variable may be assigned, but the variable remains open until a definite value is assigned (this last applies only to variables the cardinality of whose value set is greater than 2). (1) When sufficient cases have been generated terminate, returning the average number of bits needed to generate a case. (2) Generate a single case: • If all variables have values assigned, terminate, returning the number of bits needed to select values. Otherwise: • Select an open variable at random. Generate sufficient random bits to give this variable a definite value, and assign that value. (Note that the cardinality of the value set may be reduced due to earlier constraint propagation.) • Propagate the new value to the other open variables using K. Algorithm 31 is a special case of the more general problem of propagation of finite domain constraints. It would be possible to build a constraint size estimating procedure using a finite domain constraint logic programming language [12]. Using an implementation of Algorithm 31 specifically designed for K1 constraint sets, Garvan K1 was found to be of size 25.4 bits [6]. Since the attribute space for Garvan ES1 is of size 46 bits, the region of experience for Garvan K1 is 2−20 ≈ 10−6 of the total attribute space. This gain of 20 bits is more than half of the maximum possible gain for Garvan ES1. There are nearly 4000 distinct cases, so the region of experience is at least 12 bits large, so the maximum gain is 46−12= 34 bits. This result suggests that a plausible strategy for inducing K is to first induce K1, proceeding further only if K1 is not strong enough. A very simple algorithm suffices to compute K1. Algorithm 32. Compute K1 from a set of cases C. Construct an above the diagonal triangular array A of integers each of whose rows and columns is indexed by a member of one of the value sets of one of the variables underlying C. Initialise this matrix to 0. For each member c of C, increment all the cells in A which correspond to pairs of values of variables in c. Each cell of A now contains the number of instances the associated variable values co- occur in C. Each variable value x =v selects a set of cells of A corresponding to all other variable values. If all but one of the cells associated with a given value is zero, with x′ =v′ indexing the nonzero cell, then x=v→x′ =v′ is a partial functional dependency. Algorithm 32 requires an array which is square in the aggregate cardinality of the value sets (which is about 100 for Garvan ES1). It would be practicible to extend the algorithm 204 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 to three dimensions to compute K2, since the size of the corresponding matrix would be cubic in the aggregate cardinality of the value sets. We now know that it is possible, in at least one case, to economically induce a small, strong set of PFDs. It remains to see what effect K has on the inflation problem. An experiment was run using the algorithm resulting from Theorem 20 (table → tree), using a slightly different attribute selection procedure, applied to the 653-rule decision table resulting from applying the algorithm derived from Theorem 10 to the Garvan ES1 rule set, then selecting only those rows which were consistent with all of the 9805 cases [4]. In this experiment, the algorithm generated a tree with 46,378 leaves, of which 4562 were consistent with Garvan K1, and 985 were consistent with at least one case. We can conclude from this section that the method of representing K as a set of partial functional dependencies is plausibly a practical approach to the inflation problem. The practicality issue is addressed in Section 6 below. 5. Further properties of K The set K of constraints was introduced as a heuristic designed to control the inflation problem when translating propositional classification systems from one form to another. We observed in passing that another use for K was as a filter to detect cases that were either spurious or, if valid, at least outside the experience on which the classification expert system is based. If we take the use of K as a filter seriously, and stipulate that the expert system is a partial function defined only on the subset of the attribute space covered by K, we can incorporate K into the translation theorems obtaining a corresponding set of translation procedures which is likely to be much less susceptible to inflation. Proposition 33. Given a conjunction A of elementary propositions and a case C associated with the same variable set as A; either A is inconsistent, A and C are inconsistent, or A & C=C. Proof. By Proposition 17, a conjunction of elementary propositions is either inconsistent or a choice from the sets of elementary propositions associated with a subset of the variables. A case is a choice from the complete set of variables. Therefore in the conjunction of an arbitrary consistent conjunction A of elementary propositions, for each elementary proposition in A there is an elementary proposition in C associated with the same variable. If they are distinct, A and C are inconsistent with each other. Otherwise A & C=C by the Boolean algebra identity a & a=a applied to the elementary propositions in the conjunction. 2 Proposition 34. For any consistent set K of partial functional dependencies and any case C covered by K, K & C=C. Proof. K can in principle be represented in disjunctive normal form in elementary propositions. For each disjunct D, Proposition 33 gives either C&D = C or C is R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 205 inconsistent with D. Since C is consistent with K, at least one of the disjuncts must be consistent with C. The result follows from the Boolean algebra identity A∨A=A. 2 We now modify Theorems 10, 20 and 28. Theorem 35 (Rules → table K). An acyclic rule system defined on the region of the attribute space covered by a set K of partial functional dependencies is equivalent to a decision table defined on the same region of the attribute space. Proof. Theorem 10 proceeds by unfolding the set of rules, removing intermediate propositions, starting from intermediate propositions which are the consequents of rules having only elementary propositions in their antecedents. If a particular rule is of the form p→q, then the application of the rule to a case c follows the derivation process: (1) c, p→q. (2) c, c & p→q. (3) If c is consistent with p: c,c→q (by Proposition 33). (4) c, c→q, q. If q is an intermediate proposition, the process of unfolding in Theorem 10 creates a single proposition which is the only proposition whose consequent is the intermediate proposition, and similarly for its negation. By Proposition 34, step (2) in the derivation can be replaced by (2′) c,c & K & p→q. If c is consistent with p, then K is consistent with p, for any c, by the definition of K. If the antecedent p is expressed in disjunctive normal form, then any disjunct inconsistent with K can be removed, as step (3) of the derivation will never succeed. With this modification, the process of unfolding ultimately results in a decision table all of whose rows are consistent with K. 2 An algorithm derived from Theorem 35 will prune the rows of the decision table as they are constructed, so will arrive smoothly at the final decision table consistent with K. Theorems 19 and 20 were concerned with a complete propositional system, which was able to classify any possible case in the attribute space. Completeness as defined in Definitions 8 and 15 is no longer relevant, since we are looking at decision objects which are deliberately incomplete. This leads to the following definition of completeness with respect to a set of constraints. Definition 36. A decision object P is complete with respect to a set of partial functional dependencies K if P classifies every case covered by K. Proposition 37. If N(P) is the disjunction of cases not classified by P , then P is complete with respect to a set of partial functional dependencies K if N(P) is inconsistent with K. 206 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 Proof. By inspection. 2 Proposition 37 may not lead to computationally feasible algorithms in all circumstances. However, if N(P) can be represented in disjunctive normal form with m conjuncts, then the test for competeness with respect to K is no worse than testing whether m cases are in the region of experience defined by K. One would expect that the leaf propositions in a decision tree would tend to be relatively small (note that don’t care cells can be eliminated by the Boolean algebra identity a & true = a). A leaf proposition in disjunctive normal form with say 10 conjuncts per disjunct, having say 5 disjuncts, requires computation of 105 conjuncts to convert its negation to disjunctive normal form, which is tolerable. Most of these conjuncts would probably either be inconsistent or subsumed by others, so that the resulting expression would likely not have an excessive number of disjuncts. One would expect that there would be problems where the concept could be applied. Definition 36 leads to modifications of Theorems 19 and 20. Theorem 38 (Tree → table K). A decision tree defined on the region of the attribute space covered by a set K of partial functional dependencies is equivalent to an unambiguous decision table defined on the same region of the attribute space, which is complete with respect to K if the tree is. Proof. Follows Theorem 19, except that only paths through the decision tree consistent with K are included in the decision table, since if a path is inconsistent with K, no valid case will be consistent with it. If the tree is complete with respect to K, then for every case c covered by K, there is one path through the tree consistent with c. This path P(c) is consistent with K by definition of K, so is included in the table. Thus every case c covered by K is consistent with at least one row of the table. 2 Theorem 39 (Table → tree K). An unambiguous decision table system defined on the region of the attribute space covered by a set K of partial functional dependencies is equivalent to a decision tree defined on the same region of the attribute space, which is complete with respect to K if the table is. Proof. Parallels that of Theorem 20. Replace part (3) in the induction step with (3′) For each v in the value set of a such that the path proposition in the tree down to a P(a,v) is consistent with K, create: • An arc whose source is the deciding node created and whose arc proposition is a=v. • A decision table T(a,v) with one row derived from each row r of T in which the cell associated with a is consistent with a = v, and such that P(a,v)& r consistent with K, by removing the cell associated with a. This proves that the tree is equivalent to the table on K, since no branch pruned is consistent with K, nor is any row excluded from the T(a,v). The tree is therefore complete with respect to K if the table is. 2 Finally, we modify Theorem 28 (rdr-tree → table), giving R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 207 Theorem 40 (rdr-tree → table K). A ripple-down rule tree is equivalent with respect to K to an unambiguous decision table. Proof. The decision table is the union of all rdr path propositions which are consistent with K. The remaining observations from the proof of Theorem 28 apply. 2 We have now achieved a practical set of translation procedures for decision objects taking into account that they are partial functions on their attribute space. The issue of practicality is canvassed below. 6. Discussion The main result of this paper has been the solution of the problem of inflation in the translation of decision objects from one form to another. The key idea has been the characterisation of the region of experience using a set of constraints expressed as partial functional dependencies. These problems are conceptually quite simple. Their difficulty arises from their origin in exponential data structures and exponential algorithms. It is easy to construct worst-case examples where the computations are very long, so they depend for their utility on the typical case turning out to be tractable, in much the same way as the simplex algorithm in linear programming. The procedures have been tested on a large problem and shown to be practicable, but one would not be able to routinely recommend them until they had been used by a wide variety of people on a wide variety of problems and intractable cases did not arise in practice—again like the simplex algorithm—which is well beyond the scope of a single project. We can, however, consider the characteristics of situations where the procedures would fail. There are two things which could go wrong: • the set of PFDs comprising K could have few rules with a small number of antecedents, instead very many rules with a large number of antecedents; or • the use of K could fail to prune the decision objects in the translation process. The former would be a problem both because the cost of computing K increases rapidly with the number of antecedents in the rules, and partly because of the cost of using a very large set of rules with a large number of antecedents. A system with an impracticable K would be extremely complex: it would be large, else the exponential algorithms would not reach their practical limits (a six-attribute decision table can be exhaustively analysed); and it would lack simple relationships so would be difficult for humans to either understand or to gather enough data for a statistically reliable decision object induction. The latter mode of failure requires that the inflation of the decision objects be highly anti- correlated with the set of constraints. This would be surprising, because the decision object is built from the relationships of the case elementary propositions to the classification elementary propositions and the set of constraints is built from the relationships of the case elementary propositions among themselves. One would think that a system where the two were strongly correlated would be noticeably peculiar. This highly informal argument indicates that the results of this paper are likely to be of wide applicability. 208 R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 Solving the inflation problem with a set of constraints suggests a number of problems which will be the subject of further research. First, the study of the set of constraints itself: estimation of its size, simplifying it, etc., which can likely make use of the techniques of constraint logic programming (e.g., [14]). Second, results in the theory of propositional expert systems can be revisited and improved, for example, the computational stability of expert systems as reported in [3]. In that work, the chief problem was to analyse a decision object in the form of a decision table to identify situations, where small changes in attribute space make large changes in classification space, using essentially a Hamming distance measure on attribute space. The results could be improved by considering only errors which remain within the region of experience. Of course, use of the set of constraints as a filter can improve the brittleness of a wide range of expert systems, improving results such as that of Webb and Wells [23], which used method based on classification. In particular, there are problems such as described by Davidsson [9] in which a good characterisation of the region of experience is crucial. In that work, problems such as design of a coin-sorting machine are considered. In Europe, one often finds coins from other countries in a set of coins to be sorted. It becomes essential to recognise that a coin is foreign and to reject it while sorting the coins of a particular country. Use of partial functional dependencies should be able to improve Davidsson’ results. Further, in the collection of data describing the region of experience one could also obtain data describing a range of foreign coins, and so also have a set of cases which is explicitly outside the desired region of experience, which could improve the estimates of K. Edwards et al. [10] has investigated the problem of prudence in an expert system. The type of system concerned is an automated medical pathology laboratory system, which produces clinical interpretations of samples using the analysis machine results and a small amount of descriptive information about the patient. (The system, called PIERS, is a descendant of Garvan ES1.) In the application, it is a legal requirement that all clinical interpretations be signed by an appropriate specialist. This requirement represents a significant residual human involvement. The proposal of Edwards et al. is for the expert system to maintain records of all the cases it processes and to append to an interpretation how different the present case is from other cases previously processed. Cases very similar to previous cases can be assumed to have very reliable interpretations, while the more different a case is from previous cases, the closer checking the interpretation should have. Edwards et al.’s approach is to use the range of values for particular variables as a measure of similarity. We speculate that an incremental computation of a set of partial functional dependencies might give a more general and more reliable measure. In particular, we would investigate including all partial functional dependencies, no matter how weak their support, so long as they have 100% confidence. A new case can either increase the support of a dependency or break it, so that earlier estimates would tend to overconstrain the region of experience, which would be gradually widened as the system was used. Each case violating a constraint would be subject to special scrutiny. Finally, note that the characterisation of the region of experience by a set of constraints differs fundamentally from clustering methods: the set of constraints identifies the R.M.Colomb / Artificial Intelligence 109 (1999) 187–209 209 boundary of the region, while clustering methods generally identify the centres of regions of interest. It might be profitable to investigate the interaction of the two kinds of method. References [1] R. Agrawal, R. Srikant, Mining Generalized Association Rules VLDB’95, Morgan Kaufmann, Los Altos, CA, 1995, pp. 407–419. [2] K.R. Apt, H.A. Blair, A. Walker, Towards a theory of declarative knowledge, in: J. Minker (Ed.), Foundations of Deductive Database and Logic Programming, Morgan Kaufmann, Los Altos, CA, 1988, pp. 89–148. [3] R.M. Colomb, Computational stability of expert systems, Expert Systems with Applications 5 (1992) 411– 419. [4] R.M. Colomb, Y.-P. Chen, Use of partial functional dependencies to make practical approximate translations among forms of propositional expert systems, in: A. Sattar (Ed.), Proc. 5th Australian Joint Conference on Artificial Intelligence, Lecture Notes in Artificial Intelligence, Vol. 1342, Springer, Berlin, 1997, pp. 167– 176. [5] R.M. Colomb, C.Y. Chung, Strategies for building propositional expert systems, Internat. J. Intelligent Systems 10 (1995) 295–328. [6] R.M. Colomb, J. Sienkiewicz, Analysis of redundancy in expert systems case data, in: Proc. 8th Australian Joint Conference on Artificial Intelligence, World Scientific, Singapore, 1995, pp. 395–402. [7] P. Compton, R. Jansen, A philosophical basis for knowledge acquisition, Knowledge Acquisition 2 (1990) 241–257. [8] B.J. Cragun, H.J. Steudel, A decision-table-based processor for checking completeness and consistency in rule-based expert systems, Internat. J. Man-Machine Studies 26 (1987) 633–648. [9] P. Davidsson, Learning characteristic decision trees, in: Proc. 8th Australian Joint Conference on Artificial Intelligence, World Scientific, Singapore, 1995, p. 579. [10] G. Edwards, B.H. Kang, P. Preston, P. Compton, Prudent expert systems with credentials: managing the expertise of decision support systems, Internat. J. Biomedical Computing 40 (1995) 125–132. [11] K.A. Horn, P. Compton, L. Lazarus, J.R. Quinlan, An expert computer system for the interpretation of thyroid assays in a clinical laboratory, Australian Computer J. 17 (1985) 7–11. [12] J. Jaffar, M.A. Maher, Constraint logic programming: A survey, J. Logic Programming 19/20 (1994) 503– 581. [13] Z. Pawlak, Rough Sets: Theoretical Aspects of Reasoning about Knowledge, Kluwer Academic, Dordrecht, 1991. [14] M.J. Maher, Constrained dependencies, Theoret. Comput. Sci. 173 (1997) 113–149. [15] J.R. Quinlan, Semi-autonomous acquisition of pattern based knowledge, in: J.E. Hayes, D. Michie, Y.-H. Pao (Eds.), Machine Intelligence 10, Ellis Horwood, London, 1982, pp. 159–172. [16] J.R. Quinlan, Induction of decision trees, Machine Learning 1 (1986) 81–106. [17] T. Sato, Equivalence-preserving first order unfold/fold transformation systems, Theoret. Comput. Sci. 105 (1992) 57–84. [18] H. Seki, Unfold/fold transformations of stratified programs, in: G. Levi, M. Martelli (Eds.), Logic Programming: Proc. 6th International Conference (Lisbon), MIT Press, Cambridge, MA, 1989, pp. 554– 568. [19] K. Shwayder, Conversion of limited-entry decision tables to computer programs—A proposed modification to Pollack’s algorithm, Comm. ACM 14 (1971) 69–73. [20] K. Shwayder, Extending the information theory approach to converting limited-entry decision tables to computer programs, Comm. ACM 17 (1974) 532–537. [21] J.D. Ullman, Principles of Database and Knowledge-Base Systems, Vol. 1, Computer Science Press, Rockville, MD, 1988. [22] J.D. Ullman, Principles of Database and Knowledge-Base Systems, Vol. 2, Computer Science Press, Rockville, MD, 1989. [23] G.I. Webb, J. Wells, Recent progress in machine-expert collaboration for knowledge acquisition, in: Proc. 8th Australian Joint Conference on Artificial Intelligence, World Scientific, Singapore, 1995, pp. 291–298. 
work_nrnit4emvfbfhaqenwnkv7lyla	----	BarreroMorenoCamacho-eswa2010.dvi Repositorio Institucional de la Universidad Autónoma de Madrid https://repositorio.uam.es Esta es la versión de autor de la comunicación de congreso publicada en: This is an author produced version of a paper published in: Expert Systems with Applications: An International Journal 39.3 (2012): 3061- 3070 DOI: http://dx.doi.org/10.1016/j.eswa.2011.08.168 Copyright: © 2012 Elsevier B.V. All rights reserved El acceso a la versión del editor puede requerir la suscripción del recurso Access to the published version may require subscription https://repositorio.uam.es/ http://dx.doi.org/10.1016/j.eswa.2011.08.168 Adapting Searchy to Extract Data using Evolved Wrappers David F. Barreroa, Marı́a D. R-Morenoa, David Camachob aUniversidad de Alcalá. Computer Engineering Department. Escuela Politécnica. Ctra. Madrid-Barcelona km 31,600 28871 Alcal de Henares, Madrid, Spain Phone: (+34) 91-885-69-20, fax: (+34) 885-66-41 bUniversidad Autónma de Madrid. Escuela Politécnica Superior. C/ Francisco Tomás y Valiente 11. Ciudad Universitaria de Cantoblanco. 28049 Madrid, Spain Phone: (+34) 91-497-22-88, fax: (+34) 91-497-22-35 Abstract Organisations need diverse information systems to deal with the increasing requirements in information storage and processing, yielding the creation of information islands and therefore an intrinsic difficulty to obtain a global view. Being able to provide such an unified view of the -likely heterogeneous- information available in an organisation is a goal that provides added-value to the information systems and has been subject of intense research. In this paper we present an extension of a solution named Searchy, an agent-based mediator system specialized in data extraction and integration. Through the use of a set of wrappers, it integrates information from arbitrary sources and semantically translates them according to a mediated scheme. Searchy is actually a domain-independent wrapper container that ease wrapper development, providing, for example, semantic mapping. The extension of Searchy proposed in this paper introduces an evolutionary wrapper that is able to evolve wrappers using regular expressions. To achieve this, a Genetic Algorithm (GA) is used to learn a regex able to extract a set of positive samples while rejects a set of negative samples. Keywords: Wrappers, Genetic Algorithms, Information Extraction 1. Introduction Organisations have to deal with increasing needs of process automation, yielding a grown of the number and size of soft- ware applications. As a result there is a fragmentation of the in- formation: it is placed in different databases, documents of dif- ferent formats or applications that hide valuable data. Thus, it originates the creation of information islands within the organi- sation. Then, it has a negative impact when users need a global view of the information, increasing the complexity and devel- opment costs of applications. Usually ad-hoc applications are developed despite its lack of generality and maintenance costs. Information Integration [1] is a research area that addresses the several problems that emerge when dealing with such scenario. When a bunch of organizations are involved in an integra- tion process, the problems associated in the integration are in- creased. Some traditional integration problems, such as infor- mation heterogeneity, are amplified and new problems such as the lack of centralized control over the information systems arises. One of the most interesting problems in such context is how to ensure administrative autonomy, i.e., limit as much as possible the constrains that the integration might impose to data Email addresses: david@aut.uah.es (David F. Barrero), mdolores@aut.uah.es (Marı́a D. R-Moreno), david.camacho@uam.es (David Camacho) sources. We have developed a data integration solution called Searchy with the intention of addressing those constrains. Searchy [2] is a distributed mediator system that provides a virtual unified view of heterogeneous sources. It receives a query and maps it into one or more local queries, then trans- lates the responses from the local schema to a mediated one defined by an ontology and integrates them. It separates the in- tegration issues from the data extraction mechanism, and thus it can be seen as a wrapper container that eases wrapper de- velopment. It is based on Web Standards like RDF (Resource Description Framework) or OWL (Web Ontology Language). Then, Searchy can be easily integrated in other platforms and systems based on the Semantic Web or SOA (Service Oriented Architecture). Experience using Searchy in production environments has shown issues to be enhanced. One of the most successful wrap- pers in Searchy was the regex wrapper, a wrapper that extracts data from unstructured documents using a regular expression (or simply regex). Regex is a powerful tool able to extract strings that match a given pattern. Two problems were found related to wrapper-based regex utilization: the need of an engi- neer (or a specialized user, which we usually denoted as wrap- per engineer) with specific skills in regex programming, and the lack of automatic way to handle errors in the extraction pro- cess. These problems lead us to adapt the Searchy architecture to support evolved wrappers. That is, wrappers based on regex Preprint submitted to Expert Systems with Applications September 12, 2011 that have been previously generated using Genetic Algorithms (GAs). This wrapper uses supervised learning to generate a regex able to extract records automatically from a set of posi- tive and negative samples. Wrappers in Searchy may implement arbitrary complex ex- traction algorithms. The wrapper may, for instance, rely in a Multiagent System (MAS) to generate a composed regular ex- pression able to extract records that match with a training set, as it is described in detail in [3].This paper describes a wrapper able to evolve a simple regex thought a VLGA with an alpha- bet automatically generated and extract records matching the regex. It also provides an empirical evaluation of these evolved wrappers. A second contribution of this paper is the description of a complex wrapper able to extract data by means of evolu- tionary regular expressions. That is, regular expressions (or simply regex) that have been generated using Genetic Algo- rithms (GA). This wrapper uses supervised learning to generate a regex able to extract records automatically, without human supervision, from a set of positive and negative samples. This article is structured as follows. Section 2 provides a general overview the system architecture. The information re- trieval and information integration mechanism used in Searchy is briefly described in section 3. The evolved regex wrapper is presented in section 4 followed by a description of the alpha- bet construction algorithm. Some experiments carried out by the regex wrapper are shown in section 6. Section 7 describes related work. Finally, some future steps are outlined and con- clusions are summarized. 2. Searchy Architecture Many properties of Searchy are consequences of two design decisions: the MAS approach [4, 5] and the Web standards compliance. Using MAS gives Searchy a distributed and de- centralized nature well suited for the integration scenario de- scribed in the introduction. Web Services are used by Searchy agents as an interface to access their functionalities meanwhile the Semantic Web standards are used to provide an informa- tion model for semantic and structural integration [6]. From an architectural point of view agents were designed to maximize modularity decoupling integration from extraction issues, eas- ing the implementation of extraction algorithms. In our architecture each agent is composed by four compo- nents, as can be seen in Figure 1. Some of the key properties of Searchy are directly derived from this architecture. These ele- ments are the communication layer, the core, the wrappers and the information source. The next lines describe these compo- nents related to the FIPA Agent Management Reference Model. Communication layer It provides features related to the com- munications such as SOAP message processing, access control and message transport. The Communication layer is equivalent to the Message Transport System (MTS) in the FIPA model. Core It contains the basic skills used by all the agents, includ- ing configuration management, mapping facilities or agent Figure 1: Searchy platform architecture. identification. Any feature shared by all the agents is con- tained in the core. It presents some of the features defined by FIPA for the Agent Management System (AMS), how- ever they are not equivalents. AMS are supposed to con- trol the access of the agents to the Agent Platform (AP) and their life cycle. Meanwhile the agent core supports the operation of the wrappers. Wrapper A wrapper is the interface between the core agent and a data source, extracting information from the me- diated data source. Wrappers are a key point in order to achieve generality and extensibility. Agents in the FIPA model have some similarities with Searchy wrappers from an architectural point of view. An AP in the FIPA model may contain several agents meanwhile each agent in Searchy may contain several wrappers. Both of them are containers for some software asset, agents in case of FIPA or wrappers in case of Searchy. Data source It is where information that is the object of the integration process is stored. Almost any digital informa- tion source might be used as data source. Due to the nature of Searchy, data sources are usually some kind of informa- tion system such as a web server or an index. However any source of digital information is a potential Searchy data source. There is no equivalent in the FIPA model to data sources. Figure 1 shows the architecture of a Searchy agent with its four components. Agents interfaces are published thought the HTTP server, one of the subsystems of the communication layer. It receives the HTTP request that has been sent by the Searchy client and extracts the SOAP message. In order to pro- vide a first layer of security, the HTTP subsystem filters the 2 request using the Access Control Module. This module is an IP based filter that enables basic access control. The HTTP server has responsibilities with the SOAP messages transport, but the processing of these messages is done by its own module, the SOAP Processing Module. It processes the SOAP messages and then it transfers the operation to the Control Module or re- turns an error message. Once the message has been successfully processed, the Control Module begins to operate. The Control Module sets the flow of operations that the dif- ferent elements involved in the integration must perform, in- cluding the wrappers, the Mapping Module, and the Integra- tion Module. The Mapping Module is composed of three sub- systems, with different responsibilities in the mapping process. The Query Mapping subsystem performs the query rewriting translating the query from the mediated schema into the local schema, for example, SQL. Meanwhile, the Response Map- ping subsystem translates the response from a local schema like SQL, into RDF following a mediated schema defined by an on- tology. Both, Query and Response Mapping subsystems use the Mapping subsystem, that provides common services related to mappings and rules management to Query and Response Map- ping subsystems. The way in which the integration and map- ping processes operate is described in the section 3. Respon- sibility for Information extraction as well as communication among the agents falls in the wrappers. In our architecture the coordination among agents is based on an organizational structuring model with two different discov- ery mechanisms. In the first mechanism each agent has a static knowledge about which agents it must query, where it can find them and how access. The result is a static hierarchical struc- ture. It is useful in order to adapt a Searchy deployment to the hierarchy of a organisation, however it cannot take full benefice of a MAS such as parallelism, the reliability of the whole sys- tem is reduced and it is difficult to integrate in dynamic envi- ronments. To overcome some of these disadvantages a second coordina- tion mechanism has been implemented. Using our previous or- ganizational structuring model, relationships among the agents are not stored within the agents, but externally in a WSDL doc- ument that can be fetched by any agent from a HTTP or FTP server. This agent discovery mechanism is simpler than using an UDDI (Universal Description, Discovery, and Integration) directory or a Directory Facilitator (DF) in a FIPA platform. Agents are accessed as another data source, and thus it is done by a set of wrappers responsible of the discovery and communi- cation between Searchy agents: the Searchy and WSDL wrap- pers. These wrappers implement the coordination mechanism in Searchy, however wrappers’ main purpose is to extract data from data sources. At present Searchy includes four ordinary wrappers: SQL, LDAP, Harvest and regex. By means of SQL and LDAP wrap- pers, structured data in databases and LDAP directories may be accessed. Using the Harvest wrapper Searchy can integrate re- sources available in an intranet like HTML, LATEX, Word, PDF documents and other formats. The support of new data sources is done by the development of new wrappers. In section 4 we explain the regex wrapper. There is no restriction on the al- gorithm and data source that the wrapper might implement, it may be a direct access to a database, a data mining algorithm or data obtained from a sensor. Mapping and integration issues are managed by the agent’s core, and thus the wrapper has not to be concerned by these issues. Next section describes how these tasks are performed. 3. Mapping and Integration in Searchy Integrating information means dealing with heterogeneity in several dimensions [6]. Technical heterogeneity can be over- came by selecting the proper implementation technology. In our work it has been done using Web Services (WS) as an in- terface to access to the service. Addressing information hetero- geneity requires the definition of a global information model, the mediated schema, among all the entities involved in the in- tegration process, as well as a mapping mechanism to perform a mapping between the different local information models and the global information model. Defining this model is a critical subject in an information integration system. Searchy uses semantic technologies standardized by the WWW -RDF, RDFS and OWL- to represent the integrated in- formation. RDF is basically an abstract data model that can be represented using several syntaxes. Searchy uses RDF serial- ized with XML to represent the information. This combination of RDF and XML grants interoperability in a structural level. Semantic integration requires an agreement about the meaning of the information to deal with semantic heterogeneity. This agreement is performed by using shared ontologies expressed in RDFS or OWL. Then, there must be an explicit agreement among all the actors involved in a Searchy deployment to es- tablish at least one global ontology. A set of mapping rules are needed in order to map entities according to a local schema into the global schema. Rules are used to map queries to a local schema and responses to the mediated schema. Query format is a tuple <attribute, query> of strings, the Query Mapping subsystem rewrites the query to obtain a query valid for the local data source. The first element in the tuple is an URI that represents the concept to which the query is re- ferred, meanwhile the query is a string with the content of the concept that is being queried. The query model is simple but enough to fulfill the requirements of the application. The trans- lation of the query to the local schema is performed using the Mapping Module (see Figure 1). Mappings are done by means of a string substitution mechanism very similar to the traditional printf() function in C. This mechanism is enough to satisfy the needs in almost all cases. Once a query has been translated the response of the local information source must be extracted, mapped to a shared ontology and integrated, respectively, by the Response Mapping and Integration subsystems. Response mappings are done in two stages: 1. The response is mapped semantically conforming to a shared ontology. It is done using the same mechanism than the Query Mapping subsystem. A critical aspect is to provide a URI identifier for each resource, just like RDF requires to identify any resource. There is no unified way 3 to do this task: each type of wrapper and user policy define a way to name resources. 2. Each response of each wrapper is integrated in the Inte- gration Module. Integration is based on the URI of the resource returned by the wrappers. When two wrappers return two resources identified by the same URI, the agent interprets that they are referred to the same object and thus they are merged. Figure 2 shows a simple example of an integration pro- cess within Searchy. There are two data sources: a relational database and a LDAP directory service. In a first stage the wrappers retrieve the information from the local data source and this is mapped into a RDF model. The mapping is done by using the terms defined by an ontology and according to some rules given by the system administrator. The ontologies used within the integration process must be shared among all the agents. In general, a one to one correspondence between a data field and an ontology term will be defined. Several local fields or fixed texts may compose one value in RDF, this feature aids the administrator to define more accurate mappings. The mapping rules defined in the example shown in Figure 2 for the database wrapper are depicted in Example 1. Example 1 Query mapping rules example rdf:about IS "http://www.example.org/" + name dc:title IS name + " " + surname foaf:family_name IS surname The first rule defines that the RDF attribute rdf:about is built with the concatenation of the string ”http://www.example.org/” and the attribute Person as it is defined in the local schema. The rest of rules are defined in a similar way. Meanwhile the map- ping rules for the directory wrapper can be seen in Example 2. Example 2 Response mapping rules example rdf:about IS "http://www.example.org/" + uid rdf:type IS foaf:Person foaf:mbox IS email foaf:homepage IS web The wrappers in the example use two vocabularies: Dublin Core and FOAF. Each object retrieved from the data source must be identified by an URI, that in this case is built using local data with a fixed text. The second stage integrates the enti- ties returned by the wrappers. The agent core identifies the two objects as the same object by comparing their URI and merges the attributes, providing a RDF object with attributes retrieved from two different sources. Mapping and Integration Modules decouple data integration and mapping from the extraction, and thus it is possible to de- velop wrappers in Searchy without any concern about these issues. Next section shows an example of how a complex wrapper may be developed using the infrastructure provided by Searchy. The original architecture of Searchy [3] provided an easy to use extraction and integration platform. However, it required human supervision in some parts of the process. One of the most useful wrappers supported by Searchy is the regex wrap- per, which is able to extract data from unstructured documents. One problem associated with this wrapper is the need of a wrap- per engineer skilled in regex programming. Another problem is the error detection, that is, detect when the wrapper is not ex- tracting data correctly and correct it. It lead us to extend the original Searchy regex wrapper able to extract data using a regex created by the wrapper engineer with an evolved regex agent able to generate a regex from a set of positive and negative examples using a GA. Figure 3 depicts the extended architecture, where the original architecture is ex- tended with control and evolutive agents. The MAS contains three kind of agents: control, extractor and evolutive agents. The three types of agents share the same agent architecture de- picted in Figure 1, they differ from an architectural point of view in the wrappers they use. Figure 3 uses solid lines to rep- resent the iteration among the agents and resources with the exception of iterations that involve regex, which is represented with dotted lines. There must be one control agent that receives queries from the user and forwards it to the extractors, which are agents with a regex wrapper. Regex wrappers in the original Searchy ar- chitecture obtained the regex from the wrapper engineer, who generated manually the regex. When the wrapper detected a fail in the data extraction, i.e., when it was unable to extract data from a source, the wrapper notified it to the wrapper en- gineer who had to identify the problem and in case the regex was incorrectly constructed, he generates a new one. When the wrapper detected a fail in the data extraction, i.e., it is unable to extract data from a source, the wrapper notifies the wrapper en- gineer and he had to identify the problem and in case the regex was incorrectly constructed he had to generate a new one. The new architecture aims to automate this approach, using an evolutive agent that fulfills some roles of the wrapper en- gineer. Extraction agents obtain the regex from the evolutive agents on start-up time, but also when they identify an extrac- tion error. In this case, instead of requesting a new regex to the wrapper engineer, it would request it to the evolutive agent. When an evolutive agent is required to generate a new regex, it executes a GA that is described in the next section. 4. Wrapper based on evolved regular expressions The implementation of the evolved regex was done as a Searchy wrapper using the Searchy wrapper API. When an agent with the evolved regex wrapper is run, the wrapper gener- ates a valid regex executing the described VLGA with a given training set. Once a suitable regex is generated, the wrapper can begin to extract records from any text file accessible thought HTTP or FTP. It does not have to manage any issue related to mapping since the Mapping Module performs this task. 4 Figure 2: Example of the integration process in Searchy, with two data sources, one relational database and a directory. Figure 3: Searchy evolutive agents. 4.1. Codification Any GA has to set a way to codify the solution into a chro- mosome. The VLGA implemented in the wrapper uses a bi- nary genome divided in several genes of fixed length. Each gene codes a symbol σ from an alphabet Σ composed by a set of valid regular expressions constructions, as described in sec- tion 5. Some words should be dedicated to how genes codes regex. The alphabet is not composed by single characters, but by any valid regex, in this way the search space is restricted leading to a easier search. These simple regular expressions are the building blocks of all the evolved regex and cannot be divided, thus, we will call them atomic regex. The position (or locus) of a gene determines the position of the atomic regex. Gen in position i is mapped in the chromosome to regex transformation as an atomic regex in the position i. Figure 4 represents a simple example of how the regex ca[tr] could be coded in the GA. Figure 4: Example of chromosome encoding. 5 4.2. Evolution strategy Genetic operators used in the evolution of regular expres- sions are the mutation and crossover. Since the codifications rely in a binary representation, the mutation operator is the common inverse operation meanwhile the recombination is per- formed with a cut and splice crossover. Given two chromo- somes, this operator selects a random point in each chromo- some and use it to divide it in two parts, then the parts are in- terchanged. Obviously, the resulting chromosomes will likely be of different lengths. Selective pressure is introduced by a tournament selection where n individuals are randomly taken from the population and the one that scores a higher fitness is selected for reproduction. An elitist strategy has also be used, where some of the best individuals in the population are trans- ferred without any modification to the new generation. In this way it is assured that the best genetic information is not lost. 4.3. Fitness How goodness of any solution is measured is a key subject in the construction of a GA. In our case, for each positive exam- ple, the proportion of extracted characters is calculated. Then the fitness is calculated subtracting the average proportion of false positives in the negative example set to the average of characters correctly extracted. In this way, the maximum fit- ness that a chromosome can achieve is one. This happens when the evolved regex has correctly extracted all the elements of positive examples while any element of the negative examples has been matched. An individual with a fitness value of one is called ideal individual. From a formal point of view, the fitness function that has been adopted in the wrapper uses a training set composed by a positive and a negative subset of examples. Let P be the set of positive samples and Q the set of negative samples, such as P = {p1, p2, ..., pM} and Q = {q1,q2, ...,qN}. Both, P and Q are subsets of the set of all strings G and they have no common elements, so P∩Q=⊘. Chromosomes are evaluated as follows. Given a chromo- some it is transformed into the corresponding regex r ∈ R, then tries to match against the elements of P and Q. The set of strings that r extracts from a string p is given by the function ϕ( p,r) : (S×R) −→ R while the number of characters retrieved is represented by |ϕ( p,r)|. The percentage of extracted charac- ters of pi such as i = 0, ..., M is averaged and finally the fitness is calculated subtracting the average proportion of false posi- tives in the negative example set to the average of characters correctly extracted, as is expressed by equation (1). F(r) = 1 |P| ∑ pi∈P |ϕ( pi,r)| |pi| − 1 |Q| ∑ qi∈Q Mr (qi) (1) where |pi| is the number of characters of pi, |P| the number of elements of P, |Q| the number of elements of Q and Mr (qi) is defined as Mr (qi) = { 1 i f |ϕ(qi,r)|> 0 0 i f |ϕ(qi,r)|= 0 (2) 5. Zipf’s law based alphabet construction 5.1. Preliminary considerations Section 4.1 has shown how a classical binary codification is used to select one symbol σ from a predefined set Σ of symbols or atomic regex. The construction of Σ is a critical task since it determines the search space, its size and its capacity to express a correct solution. Of course, the simplest approach is to man- ually select the alphabet, however this approach may devaluate the added value of evolved regex: the automatic generation of regex. We can state that construction of Σ must satisfy three con- strains. 1. Σ must be sufficient, i.e., it must exist at least an element r ∈ Σ∗ such as r is an ideal individual. In other words, it must be possible to construct at least one valid solution using the elements of Σ. 2. |Σ| must contain the minimum number of elements able to satisfy the sufficiency constrain. Of course, being able to satisfy this condition is a challenging task with deep the- oretical implications. From a practical point of view, this constrain can be reformuled to keep |Σ| as low as possible. 3. Symbol selection must be automatic, with minimal human interaction and number of parameters. 5.2. Alphabet construction algorithm To reduce the number of elements of Σ, and keep the search space as small as possible, we aim to identify patterns in the positive samples and use them as building blocks. In order to satisfy the previous constrains we propose the following algo- rithm. Σ is built as the union of F , D and T , where F , is the set of fixed symbols, D the set of delimiters and T the set of tokens. Σ= {σi}\σi ∈F ∪D∪T (3) F contains hand-created reusable symbols that are meant to be common cross-domain regex, and thus, once they have been defined they can be used to evolve different regex. It should be noticed that F may contain any valid regex, nevertheless it is supposed to contain generic use regex such as \d+ or [1-9]+. Since F is supposed to include common used complex regex, it contributes to reduce the search space and increase individual fitness by introducing high fitness building blocks. The sets D and T are constructed using a more complex mechanism based on Zipf’s Law [7]. It states that occurrences of words in a text are not uniformally distributed, rather only a very limited number of words concentrates a high number of occurrences. This fact can be used to identify patterns in P and use them to construct a part of Σ. Since the tokens do not contain delimiters, the sufficiency constrain cannot be satisfied, so, each delimiter that appear in the examples is included in the set D. The overall process is described in Algorithm 1. Of course, |Σ| must be equal to the number of elements of the union of F , D and T , as is expressed in equation (4). 6 Algorithm 1 Selection of alphabet tokens. 1 .- P := Set of positive examples 2 .- S := Set of candidate delimiters 3 .- D := T := { } 4 .- 5 .- for each p in P 6 .- for each s in S 7 .- tokens := split p using s 8 .- numberTokens := number of tokens 9 .- 10.- for each token in tokens 11.- occurrence(token) := occurrence(token) + 1 12.- endfor 13.- 14.- if (numberTokens > 0) add s to D 15.- endfor 16.- endfor 17.- 18.- sort occurrence 19.- add n first elements of occurrence to T |Σ|= |F ∪D∪T| (4) given that |F ∩D∩T|= |F ∩D|= |F ∩T|= |D∩T|=⊘ (5) 5.3. Complexity analysis A better understanding of the algorithm can be achieved by a time complexity analysis. As can be seen in Algorithm 1, there are two main loops (see Algorithm 1, lines 5 and 6) that depends on the number of examples |P| and the number of po- tential delimiters |S |. The complexity of the algorithm is given by these loops and the operations that are performed inside. Splitting a string pi ∈P (line 7) is proportional to the length of the string |pi|, so the mean time required to perform this op- eration is proportional to the mean string length |p|. Lines 19 to 21 include a loop that is repeated as many times as the tokens are in the string. A hash table is accessed inside the loop (line 20), so it makes sense to suppose that its complexity is given by the calculus of the key, a string, therefore its time complex- ity is n|p|, where n is the number of tokens. Finally sorting occurrence can be performed in ntot log(ntot) where ntot is the number of tokens stored in occurence. The rest of the opera- tions in the algorithm can be performed in a negligible time. We can express these considerations in equation (6). t ∝ |P| · |S | · [|p|+ n|p|] + ntot log(ntot) (6) Both n and ntot are unknown and we have to estimate it for the average case. A string p ∈ P of length |p| can contain approx- imately a maximum of |p|2 tokens. We have supposed there is one delimiter for each token. The maximum number of tokens that can be stored in occurrences are |P|·|S |·|p|2 . Then n = |p| 2 (7) ntot = |P| · |S | · |p| 2 (8) and 6 can be expressed as t ∝ |P| · |S | · |p|[1 + |p| 2 ] + |P| · |S | · |p| 2 log( |P| · |S | · |p| 2 ) (9) Some terms can be removed t ∝ |P||S ||p| 2 log( |P| · |S | · |p| 2 ) (10) Using Big Oh notation it yields that the time complexity is given by O(k log(k)) (11) where k = |P||S ||p| and therefore we can conclude that the time complexity is linearithmic. 6. Evaluation Two phases have been used in the evaluation, a first phase where the basic behaviour of the GA is analyzed, and a second phase that uses the knowledge acquired along the first phase to measure extraction capabilities of the evolved regex wrap- per. Measures that have been used are the well known preci- sion, recall and F-measure. The sets of experiments described in this section are focused in the extraction of three types of data: URLs, phone numbers and email addresses. 6.1. Parameter tuning Some initial experiments were carried out to acquire knowl- edge about the behaviour of the regex evolution and select the GA parameters to use within the wrapper. Experiments showed that despite the differences between phone, URL and emails all the case studies have similar behaviors. In this way it is possible to extrapolate the experimental results and thus to use the same GA parameters. Setup experiments showed that best performance is achieved with a mutation probability of 0.003 and a tournament size of 2 individuals. A population composed by 50 individuals is a good trade-off between computational resources and convergence speed. Initial population has been randomly generated with chromosome lengths that range from 4 to 40 bits and elitism of size one has been applied. Table 1 summarizes the parameter values used in the experiments. 6.2. Regex evolution Once the main GA parameters have been set, the wrapper can evolve the regex. Experiments have used three datasets to evolve regex able to extract records in the three case studies un- der study. Figure 5 depicts the mean best fitness (MBF) and mean average fitness (MAF) of 100 runs. The fitness evolution of the case studies follows a similar path. The best MBF and 7 Table 1: GA parameters summary. Parameter Value Population 50 Mutation probability 0.003 Crossover probability 1 Tournament size 2 Elitism 1 Initial chromosome length 4 - 40 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 10 20 30 40 50 60 70 F itn e ss Generations Email Phone URL Figure 5: Best and average fitness of phone, URL and email regex. MAF are achieved by the email regex, while the poorest perfor- mance is given by the URL regex, with lower fitness values. The dynamics of the chromosome length can be observed in the Figure 6. It is clear that there is a convergence of the chro- mosome length and thus chromosome bloating does not appear. It can be explained by the lack of non-coding and overlapping regions in the chromosome, i.e, if the chromosome has achieved a maximum it hardly can increase its size without a penalty in its fitness. The longer is the chromosome, the more restrictive is the phenotype and it is closely related to the associated fit- ness. URL regex has a stronger tendency to local maximum, this fact is reflected in Figure 6, where a lower MBF and MAF are achieved. This fact also explains why URL chromosome length depicted in Figure 6 is shorter than phone regex: the lo- cal maximum of URL regex tends to generate populations with an insufficient chromosome length. Those results are not sur- prising since URLs follow a far more complex pattern than phone numbers or emails. The same can be affirmed about emails in comparison to phone numbers. Figure 6 shows another interesting behaviour. As the GA begins to run, the average chromosome length is reduced un- til a point where it begins to increase, then the chromosome length converges into a fixed value. In early generations indi- viduals have not suffered evolution and thus its genetic code has a strong random nature. Individuals with longer genotype have longer phenotypes and thus more restrictive regex that will likely have smaller fitness values. So long chromosomes are discarded in early stages of the evolutive process until the pop- ulation is composed by individuals representing basic pheno- types, then recombination leads to increase complexity of in- 6 8 10 12 14 16 18 20 22 24 0 10 20 30 40 50 60 70 A vg . ch ro m o so m e le n g th Generations URL Email Phone Figure 6: Evolution of regular expressions. dividuals until they reach a length associated with a local or global maximum. 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 S u cc e ss r a te ( % ) Generations Email Phone URL Figure 7: Probability of finding an ideal regex able to accept all the positive examples while rejecting the negative ones. Some of the facts found previously are confirmed by Fig- ure 7, where the success rate (SR) [8] is depicted versus the generation. SR is defined as the probability of finding an ideal individual in a given generation. It should be noted that Figure 7 depicts the average success rate of 100 runs of the experiment. It can be seen that email achieves a SR of 91%, phone numbers 54% and URLs 63% in generation 70. These results are consis- tent with those in Figure 6 and show that the hardest study cases are URLs, phone numbers, and emails, in that order. Here the term ”hard” should not be understood in a strict absolute way since the hardness of the search space is influenced from sev- eral factors, such as the training set, the selection of negative samples or the alphabet chosen. 6.3. Data extraction Three regex with an ideal fitness of one have been selected by the wrapper and its extraction capabilities have been evalu- ated by means of the precision, recall and F-measure. The ex- periments used a dataset composed by eight sets of documents from different origins containing URLs, emails and/or phone 8 Table 2: Extraction capacity of the evolved regex. The table shows the F-measure (F), precision (P) and recall (R) achieved in the three datasets (phone, URL and email addresses). Phone regex URL regex Email regex Phone URL Email F P R F P R F P R Set 1 99 0 0 1 1 1 - - - - - - Set 2 0 51 0 - - - 0.24 0.14 0.84 - - - Set 3 0 0 862 - - - - - - 0.79 0.51 0.62 Set 4 20 77 0 1 1 1 0.27 0.16 1 - - - Set 5 37 686 0 1 1 1 0.20 0.11 0.97 - - - Set 6 24 241 0 1 1 1 0.02 0.01 0.37 - - - Set 7 83 0 88 0.92 1 0.96 - - - 0.92 1 0.96 Set 8 0 51 0 - - - 0.63 0.47 0.96 - - - Avg. - - - 0.98 1 0.99 0.27 0.18 0.83 0.85 0.79 0.79 numbers. Table 2 shows basic information about the datasets and their average records and Table 3 contains some evolved regex with their fitness value. Sets one, two and three are com- posed by examples extracted from the training set. The rest of the sets are web pages retrieved from the Web classified by their contents. An extracted string has been evaluated as cor- rectly extracted if and only if it matches exactly the records, otherwise it has been computed as a false positive. The results, as can be seen in Table 2, are quite satisfactory for phone numbers and testing sets, but measures get worse for real raw documents, specially the ones containing URL records. Phone regex has a perfect extraction with a F-measure value close to 1. The training set used to evolve regex contains phone numbers in a simple format (000)000−0000, the same that can be found in the testing set, the reduction of recall in set 7 is due to the presence of phone extensions that are not extracted. On the contrary, measures achieved for URL extraction from raw documents are much lower. It can be explained looking at the regex used in the extraction, http://\w+\.\w+\.com. Docu- ments used in the test contain many URLs with paths, so the regex is able to partially extract them, increasing the count of false positives. The result is a poor precision. An explanation of the poor recall measures in URLs extraction is found in the fact that the evolved regex only is able to extract URL whose first level domain is .com, so its recall in documents with a high presence of first level domains in other forms is worse. Finally, email regex achieves an average F-measure of 0.85. Some of the factors that limits the URL regex extraction capa- bilities are also limiting email regex. However in this case the effects are not so severe for some reasons, for instance the lower percentage of addresses with more than two levels. 7. Related Work The use of ontologies [9] has attracted the attention of data integration community in the last years. It has provided a tool to define mediated schemas focused on knowledge sharing and interoperability, in contrast with traditional database cen- tric schemas, whose goal is to query single databases [10]. The adoption of ontologies has lead to reuse results achieved by two communities such as the database and the AI communities to Table 3: Some examples of evolved regular expressions with their fitness val- ues. Evolved regex (Phone) Fitness \w+ 0 \(\d+\) 0.33 \(\d+\)\d+ 0.58 \(\d+\)\d+-\d+ 1 Evolved regex (URL) Fitness http://-http://http:// 0 /\w+\. 0.55 http://\w+\.\w+\ 0.8 http://\w+\.\w+\.com 1 Evolved regex (Email) Fitness \w+\. 0.31 \w+\.\w+ 0.49 \w+@\w+\.\com 1 solve similar problems like schema mapping or entity resolu- tion. A deep discussion about the role of the ontologies in the data integration can be found in [11]. We can define a collection of semantic solutions based on on- tology technologies prior to the development of the SW. A in- troduction to this group of solutions can be found in [6]. We can remark classical literature examples such as InfoSleuth [12] or SIMS [13]. From these systems, we have to stress InfoSleuth, a solution that uses a MAS. Semantic integration tools in the last years have adopted WS standards and technologies. One of the first ones can be found in [14]. Vdovjak proposes a semantic mediator for querying heterogeneous information sources, but limited to XML docu- ments, furthermore, this solution relies on a wrapper layer that translates the local entities into XML and then the RDF is gen- erated. A step forward is done by Michalowski with Build- ing Finder [15], a domain specific mediator system aimed to retrieve and integrate information about streets and buildings from heterogeneous sources, presented to the user within satel- lite images. [16] describes an information integration tool that covers all the phases of integration, such as assisted mapping 9 definition and query rewrite. Another newcomer into the IT toolbox is the Web Services technology. WS provide a means to access services in a loose coupling way. Despite WS and the SW face different problems -one models and represents knowledge meanwhile the another one is concerned with service provision-, they are related by means of semantic descriptions of WS thought Semantic Web Services. In this way WS are enhanced with semantic descrip- tions, enabling dynamic service composition and data integra- tion [17]. A semantic integration solution based on SOA is SODIA (Service-Oriented Data Integration Architecture) [18]. It sup- ports some integration approaches such as federated searches and datawarehouse. By using a SOA approach SODIA has many of the benefits of using an agent technology. However, this is a process centric solution and has limited semantic sup- port. The most aligned solution to the one described in this paper is Knowledge Sifter [19]. It is an agent based approach that uses OWL to describe ontologies and WS as interface to the agents’ services. Despite the lack of semantic support, WS integration or distributed nature, we have to mention the system proposed by [20], a system able to automate the full integration process by creating the mediated schema and schema mapping on-the-fly. Another interesting integration suite related to bioin- formatics domain that that could be mentioned is INDUS [21]. Table 4 compares some representative federated ontology- driven search solutions. The scope of table 4 is limited, however some relevant facts are show. It depict if the integration system is supported by agents, it uses any WS or SW technology as well as the degree of specialization of the tool. 8. Future work and conclusions Some issues are still open. One of them is the limited num- ber of information sources that Searchy is able to integrate. WebMantic [22] is a wrapper web-centric information extrac- tion tool that once integrated in Searchy will enhance its Web information extraction and integration capabilities. One method to extract data from unstructured documents is the use of evolved regex. Genetic Algorithms provide a stochas- tic search method well suited for the complex search space con- formed by regular expressions. Despite the fact that VLGAs are much simpler than the fixed-length approach described in [3], it still presents some important drawbacks, such as the intrinsic difficulty to evaluate parts of the evolved regex or the linear na- ture of the codification. In order to avoid this type of problems it is expected to use Genetic Programming and Grammatical Evaluation to generate regex. In particular, Genetic Program- ming has shown well performance in related areas [23] and is a promising approach. From the point of view of the Searchy platform, The creation of a wide network of agents implies the management of huge amount of information. Then, the physical scalability of the system should be done in parallel with the intelligence system improvement. Some techniques such as ranking or information filtering with a case based reasoning or collaborative filtering are considered to provide some intelligence to the system that will produce better user satisfaction. Along this paper we have briefly presented the problem of information extraction and integration and we have proposed an extension of a partial solution called Searchy. This exten- sion aims to automatice some of the tasks asigned originally to the wrapper engineer thought some agents able to use Machine Learning to generate regex. When an extractor agent requires a regex (it can be in initialization time or because it cannot ex- tract data with a given regex) it request one to a evolutive agent that using a set of positive and negative examples and a Genetic Algorithm is able to generate the regex. 9. Acknowledgements The authors gratefully acknowledge Martı́n Knoblauch for his useful suggestions and valuable comments. This work has been partially supported by the Spanish Ministry of Science and Innovation under the projects ABANT (TIN 2010-19872), COMPUBIODIVE (TIN2007-65989) and by Castilla-La Man- cha project PEII09-0266-6640. References [1] L. Haas, Beauty and the beast: The theory and practice of information integration, ICDT 2007 (2007) 28–43. [2] D. F. Barrero, M. D. R-Moreno, D. R. López, Information integration in searchy: an ontology and web services approach, International Journal of Computer Science and Applications (IJCSA) 7 (2010) 14–29. [3] D. F. Barrero, D. Camacho, M. D. R-Moreno, In Data Mining and Mul- tiagent Integration. Chapter 9: Automatic Web Data Extraction Based on Genetic Algorithms and Regular Expressions (2009), Springer, pp. 143– 154. [4] F. Wan, M. P. Singh, Commitments and causality for multiagent design, in: A. Pres (Ed.), 2nd International Joint Conference on Autonomous Agents and multiagent Systems (AAMAS), Melbourne, Australia. [5] R. Aler, J. M. Valls, D. Camacho, A. Lopez, Programming robosoccer agents by modeling human behavior, Expert Systems with Applications 36 (2009) 1850–1859. [6] H. Wache, T. Vögele, U. Visser, H. Stuckenschmidt, G. Schuster, H. Neu- mann, S. Hübner, Ontology-based integration of information — a survey of existing approaches, in: IJCAI–01 Workshop: Ontologies and Infor- mation Sharing, Seattle, Washington, USA, pp. 108–117. [7] G. Zipf, The Psycho-Biology of Language, Houghton Mifflin, Boston, MA, 1935. [8] D. F. Barrero, D. Camacho, M. D. R-Moreno, Confidence intervals of success rates in evolutionary computation, in: Proceedings of Genetic and Evolutionary Computation Conference (GECCO-2010)., ACM, Portland, Oregon,USA, 2010, pp. 975–976. [9] T. Grubber, A translation approach to portable ontology specifications, Knowledge Acquisition 5 (1993) 199–220. [10] M. Uschold, M. Gruninger, Ontologies and semantics for seamless con- nectivity, SIGMOD Rec. 33 (2004) 58–64. [11] N. F. Noy, Semantic integration: a survey of ontology-based approaches, SIGMOD Rec. 33 (2004) 65–70. [12] M. H. Nodine, J. Fowler, T. Ksiezyk, T. Perry, M. Taylor, A. Unruh, Ac- tive information gathering in infosleuth, International Journal of Cooper- ative Information Systems 9 (2000) 3–28. [13] C. A. Knoblock, J.-L. Ambite, Agents for information gathering, in: J. M. Bradshaw (Ed.), Software Agents, AAAI Press / The MIT Press, 1997, pp. 347–374. [14] R. Vdovjak, G. Houben, Rdf based architecture for semantic integration of heterogeneous information sources, in: E. Simon, A. Tanaka (Eds.), Proceedings of the International Workshop on Information Integration on the Web, Rio de Janeiro, Brazil, pp. 51–57. 10 Table 4: Semantic information integration tools comparison. Platform Agent support Semantic Web Web Services Interdomain support InfoSleuth Yes No No Yes SIMS Yes No No Yes Building Finder No Yes No No SODIA No Yes Yes Yes Knowledge Sifter Yes Yes Yes Limited Searchy Yes Yes Yes Yes [15] M. Michalowski, J. Ambite, S. Thakkar, R. Tuchinda, C. Knoblock, S. Minton, Retrieving and semantically integrating heterogeneous data from the web, IEEE Intelligent Systems 19 (2004) 72–79. [16] J. Yuan, A. Bahrami, C. Wang, M. Murray, A. Hunt, A semantic infor- mation integration tool suite, in: VLDB ’06: Proceedings of the 32nd international conference on Very large data bases, VLDB Endowment, 2006, pp. 1171–1174. [17] P. Deepti, B. Majumdar, Semantic web services in action - enterprise in- formation integration., in: B. J. Kramer, K.-J. Lin, P. Narasimhan (Eds.), ICSOC, volume 4749 of Lecture Notes in Computer Science, Springer, 2007, pp. 485–496. [18] F. Zhu, M. Turner, I. A. Kotsiopoulos, K. H. Bennett, M. Russell, D. Bud- gen, P. Brereton, J. Keane, M. Rigby, J. Xu, Dynamic data integration using web services., in: IEEE International Conference on Web Services (ICWS’04), San Diego, California, USA, pp. 262–269. [19] L. Kerschberg, M. Chowdhury, A. Damiano, H. Jeong, S. Mitchell, J. Si, S. Smith, Knowledge sifter: Ontology-driven search over heterogeneous databases., in: 16th International Conference on Scientific and Statistical Database Management (SSDBM 2004), IEEE Computer Society, San- torini Island, Greece, 2004, pp. 431–432. [20] A. D. Sarma, X. Dong, A. Halevy, Bootstrapping pay-as-you-go data integration systems, in: SIGMOD ’08: Proceedings of the 2008 ACM SIGMOD international conference on Management of data, ACM, New York, NY, USA, 2008, pp. 861–874. [21] D. Caragea, J. Pathak, J. Bao, A. Silvescu, C. Andorf, D. Dobbs, V. Honavar, Information integration and knowledge acquisition from se- mantically heterogeneous biological data sources, in: Proceedings of the 16th International Workshop on Database and Expert Systems Applica- tions, Springer-Verlag, 2005, pp. 175–190. [22] D. Camacho, M. D. R-Moreno, D. F. Barrero, R. Akerkar, Semantic wrappers for semi-structured data extraction, Computing Letters (COLE) 4 (2008) 1–14. [23] You-Heng, L. Ge, Learning ranking functions for geographic information retrieval using genetic programming, Journal of Research and Practice in Information Technology 41 (2009) 39–52. 11 
work_nvdskcxcwfebbhpl6q2kliwoai	----	Recognizing white blood cells with local image descriptors Dan López-Puigdollersa,1, V. Javier Travera, Filiberto Plaa aInstitute of New Imaging Technologies, Jaume-I University, Av. de Vicent Sos Baynat, s/n 12071 Castelló de la Plana, Spain Phone: +34 964 38 7767, Fax: +34 964 38 7678 Abstract Automatic and reliable classification of images of white blood cells is desirable for inexpensive, quick and accurate health di- agnosis worldwide. In contrast to previous approaches which tend to rely on image segmentation and a careful choice of ad hoc (geometric) features, we explore the possibilities of local image descriptors, since they are a simple approachthey require no explicit segmentation, and yet they have been shown to be quite robust against background distraction in a number of visual tasks. Despite its potential, this methodology remains unexplored for this problem. In this work, images are therefore characterized with the well-known visual bag-of-words approach. Three keypoint detectors and five regular sampling strategies are studied and compared. The results indicate that the approach is encouraging, and that both the sparse keypoint detectors and the dense regular sampling strategies can perform reasonably well (mean accuracies of about 80% are obtained), and are competitive to segmentation-based approaches. Two of the main findings are as follows. First, for sparse points, the detector which localizes keypoints on the cell contour (oFAST) performs somehow better than the other two (SIFT and CenSurE). Second, interestingly, and partly contrary to our expectations, the regular sampling strategies including hierarchical spatial information, multi-resolution encoding, or foveal-like sampling, clearly outperform the two simpler uniform-sampling strategies considered. From the broader perspective of expert and intelligent systems, the relevance of the proposed approach is that, since it is very general and problem-agnostic, it makes unnecesary human expertise to be elicited in the form of explicit visual cues; only the labels of the cell type are required from human domain experts. Key words: White blood cells recognition, local image descriptors, SIFT, interest point detectors, visual vocabulary 1. Introduction In blood tests, the count of the different types of white blood cells provides a good quantitative account of a per- son’s health state, which is crucial for prevention or cure pur- poses. Since huge number of blood tests are performed on a daily basis world-wide, automatic, fast and accurate proce- dures are really called for. Methods based on computer vision and machine learning are still under investigation but repre- sent promising inexpensive tools compared to manual and laser- flow approaches (Kamentsky, 1973; Saphiro, 2003), which are massive-counting procedures and are not very precise in dis- criminating particular cell types. There are two main types of white blood cells, Granulocytes and Agranulocytes. In turn, Granulocytes include Neutrophil, Eosinophil, and Basophil, whereas Agranulocytes can be of types Lymphocyte and Monocyte. In addition, during the sam- ple preparation process any cell of these types can be broken, which greatly distorts its appearance and can therefore be re- garded as a class itself. By looking at examples of these classes Email addresses: vtraver@uji.es (V. Javier Traver) 1Present address: Image Processing Laboratory, Universitat de València, C/ Catedrático José Beltran, 2, 46980 Paterna (València), Spain. Phone: +34 96 354 32 29, Fax: +34 96 354 32 61 (Fig. 1), it is hard to think of any simple rule of thumb or heuris- tics that can guide the design of discriminative image descrip- tors, since there are few clear distinctive features and significant overlap. This situation seems to call for a feature learning scheme such as deep learning (DL). Despite the advantages of auto- matically learning features directly from data, DL is generally known to pay off when a great deal of training instances are available, often in the order of tens of thousands. An alternative is segmenting the images so that the region(s) of interest corresponding mainly to the cell type are isolated from the background. After the segmentation, one can char- acterize the region(s) by using a combination of shape, color and texture features. The main drawback of this approach is that segmentation itself is a non-trivial problem, and it is gen- erally hard to produce robust and reliable segmentations, be- ing a critical preliminary step for subsequent feature computa- tions (Bikhet, Darwish, Tolba & Shaheen, 2000; Sarrafzadeh, Dehnavi, Rabbani & Talebi, 2015). Back in the 1970s, early efforts were performed to automate the task of classifying 5 types of white blood cells using 74 instances and four very specific features (size and color of nu- cleus and cytoplasm) (Young, 1972), and reported that granular- ity, texture and shape do not provide additional discriminative ability. Preprint submitted to Expert Systems with Applications August 18, 2018 Some authors (Gómez-Gil, Ramı́rez-Cortés, González- Bernal, Pedrero, Prieto-Castro, Valencia, Lobato & Alonso, 2008) use specific shape and area descriptors to classify six types of leukocytes, with encouraging results in a limited dataset of 54 instances. However, this approach requires seg- mentation, which is accomplished manually, and makes use of ad hoc features; additionally, 7 out of 9 segmented images per class are synthetically generated by a human artist. Sim- ilarly, geometric features computed on segmented images by simple global thresholding are considered in a simplified 3- class Leukocytes discrimination problem (Hiremath, Bannigi- dad & Geeta, 2010). There exist other works with similar approaches (Piuri & Scotti, 2004; Gautam, Singh, Raman & Bhadauria, 2016). One interesting idea is trying to increase the classification rate by taking into account the class predicted for co-occurring cells within a given specimen instead of clas- sifying single cells individually (Song, Abu-Mostafa, Sill & Kasdan, 1997). In (Rezatofighi & Soltanian-Zadeh, 2011), af- ter nucleus-cytoplasm automatic segmentation, morphological, color and texture features are extracted and selected, resulting in a good-performing approach. A similar framework includ- ing segmentation and texture features is described in a sys- tem requiring user interaction for some tasks such as super- vising the segmentation results (Sabino, da Fontoura Costa, Rizzatti & Zago, 2004). Features derived from intensity his- tograms, and their projection with Kernel Principal Component Analysis (Habibzadeh, Krzyżak & Fevens, 2013) are tested for white blood cell discrimination in a small dataset of 140 low- resolution instances. They compared these features and linear Support Vector Machine (SVM) with a simple Convolutional Neural Network (CNN) model (Le-Net5). Despite their small dataset, the CNN is reported to perform on a par with or better than the SVM. An interesting alternative are local image descriptors com- puted at interest points, and the use of a pooling strategy for characterizing each individual image with a fixed-length repre- sentation. This approach has proven to be very effective in a number of computer vision tasks (Laptev & Lindeberg, 2003; Mikolajczyk & Schmid, 2005; Laptev, 2005; Wang & Mori, 2009), because it tends to be robust against background clut- ter and requires no explicit segmentation of the relevant regions of the image. However, in spite of its potential, to the best of our knowledge, this approach has not been tested before on this problem. We therefore explored the possibility of characteriz- ing white blood cells with local descriptors and the well-known bag of visual words as a pooling mechanism. More concretely, this work focuses on exploring and com- paring several different interest point sampling strategies within two big broad categories: interest point detection and regular dense-like sampling. Although interest point detection is aimed at selecting particularly good distinctive image locations and is therefore a promising procedure, many studies have revealed that dense-like sampling, i.e. ignoring the image contents for point selection, can perform similarly or better than point de- tectors. Since our main interest is in comparing these two main “detection” approaches, a common local descriptor, the well- known Scale Invariant Feature Transform (SIFT) (Lowe, 2004), will be used in all cases. From the broader scope of expert and intelligent systems, dif- ferent methodologies can differ in their potential for addressing practical, real-world issues. On the one hand, due to legal and other reasons, there is a growing interest towards making ar- tificial systems explainable (Gunning, 2016; Ribeiro, Singh & Guestrin, 2016), which is understandably particularly impor- tant in the medical domain (Holzinger, Biemann, Pattichis & Kell, 2017). In this sense, and in the context of the problem addressed in this paper, methods that use visual cues that are easier to interpret by humans, such as segmentation-based and geometric features, might be preferable. On the other hand, problem-specific ad hoc image features may require some form of expert knowledge to be elicited and implemented, a process that can be vague and costly. In this respect, general, problem- agnostic methodologies, such as the one proposed in this work, based on general-purpose local descriptors, might be desirable. This situation represents a challenging dichotomy between con- flicting properties of alternative methodologies, which is revis- ited later (Sect. 4.4). In the rest of this manuscript, the sparse interest-point de- tectors (Sect. 2.1), the regular sampling strategies (Sect. 2.2) and the bag-of-words approach (Sect. 2.4) are described, and a summary of the complete process is provided (Sect. 2.5). Then, the experimental work, results and discussions are de- tailed (Sect. 3). Finally, some discussion (Sect. 4) and the main conclusions of the work are given (Sect. 5). 2. Methodology 2.1. Interest point detectors Interest points, also called keypoints, are locally salient im- age points. Mainly meant for image matching, interest point de- tectors have been devised for robustness against view changes. Although for classification purposes the requirements may not be exactly the same as in matching, the properties of these de- tectors turn out to be generally useful also for image charac- terization. Among the many existing detectors (Tuytelaars & Mikolajczyk, 2008; Krig, 2014), we select and explore these three detectors: SIFT, oFAST, and CenSurE. They are repre- sentative of three broad kinds of keypoint detectors, as follows: • SIFT relies on the concept of scale-space, which can be scale-invariant, but may lose some location precision due to the involvement of coarser-resolution levels in the pyra- mid; • oFAST detects corners, which can be fast to compute and are stable image locations, robust to changes of view, but not to scale changes; • CenSurE aims at achieving both the stability of scale- space methods and accuracy of corners, by finding extrema in the responses of centre-surround filters. 2 SIFT detector The Scale-Invariant Feature Transform (SIFT) (Lowe, 2004) includes both a detector and a descriptor. We focus now on its detection step. Its descriptor (Section 2.3) is applied for every detector considered in this work. Unlike previous detectors at that time, the SIFT detector was designed to be scale and ro- tation invariant which provides it with robustness against affine image transformations and, hence, repeatability. SIFT uses a multi-scale scheme so that interest points at dif- ferent scales can be found. To that end, a Gaussian pyramid is built, and then Differences of Gaussians (DoGs) are computed between the consecutive levels of the pyramid. Next, local max- ima in the DoGs are identified not only within the 8 neighbors of a pixel in the DoG at a given scale but also within its 9 neigh- bors of the subsequent-level DoG. To remove the responses to edge-like structures which are not precise locations of interest points, a second-derivative fil- tering mechanism using the Hessian matrix is applied. An in- tensity threshold of the responses is finally applied so that low- response points are filtered out. This thresholding allows the procedure to be more selective in the number of the result- ing points. The adequate threshold to use is highly problem- dependent. In our case, in order to have an acceptable number of key points for adequately representing the images, a rela- tively low-intensity threshold was used (Table 1). oFAST (Oriented FAST) The FAST (Features from Accelerated Segment Test) detec- tor (Rosten & Drummond, 2006) is computationally very light, and builds on the conceptually simple idea (Rosten & Drum- mond, 2005) of comparing a pixel with its 16 neighbouring pixels lying on a circle centred on the test pixel. A pixel is decided to be a corner if a number n of contiguous pixels on that circle are brighter or darker than the test pixel, for a given threshold for the intensity difference. With n = 12, a high-speed version of the test is possible by checking only 4 pixels on the circle, and classifying the pixel as a corner if 3 out of these 4 pixels are all brighter or darker than the test pixel. However, this quick test was found to have some drawbacks such as poor generalization for n < 12. To address this and other shortcom- ings, a decision tree is built that learns to classify corners from a properly generated training set which uses a slow but accu- rate corner detector. Then, the resulting tree is converted into nested if-then-else rules, which can subsequently be used as the fast and more accurate detector. The oFAST (Rublee, Rabaud, Konolige & Bradski, 2011) in- cludes the orientation component which is missing in FAST. This orientation is estimated by using the angle of the vector joining the centre of the detected corner and the centroid of the image patch around the corner’s centre. Although the original FAST itself does not detect features at different scales, it can be applied at the different levels of a pyramid scale. This multi- scale FAST and a circular neighbourhood of radius 9 (FAST-9) are used in this work. CenSurE Like SIFT, the CENter SURround Extrema (CenSurE) de- tector (Agrawal, Konolige & Blas, 2008) uses a Laplace scale- space operator, but in contrast to SIFT, which uses a DoG approximation, CenSurE uses a a centre-surround one. Un- like SIFT, which subsamples images at larger scales, CenSurE uses the full-resolution images at any scale, which is an im- portant difference. Then, since CenSurE aims at computing all features at all scales, fast computation is a must. Therefore, simplified bi-level kernels (values in the kernels are only −1 and +1) are used as centre-surround filters. These kernels are computed in constant time regardless of their size, which may render this descriptor real-time-amenable. CenSurE considers a suite of bi-level kernels with increasing symmetry degrees (boxes, hexagons, octagons and circles), which also represent approximations to the Laplacian operator with increasing com- putational cost. The circular filter (STAR) is the slowest but the best approximation to the Laplacian; the other approxima- tions are computed efficiently by means of integral images (Vi- ola & Jones, 2004). Since in our study accuracy is more im- portant than efficiency, we use the circular filter. However, to gain some insight into the computation-recognition trade-off, the fastest and least accurate difference-of-boxes kernel (a ba- sic Haar wavelet) is also briefly tested. Additional configuration details of these three detectors are provided in Table 1. Most of these parameters take their corre- sponding default values in the libraries used. Some thresholds have been modified to increase the number of detected points, since it was too small in some images, particular in a few defo- cused ones. In addition, in oFAST the sensibility Harris factor was reduced to 0 to detect only pronounced corners. Table 1: Configuration of the keypoint detectors. Default values given in paren- theses if different to the values actually used SIFT Edge threshold 10 Intensity threshold 0.5 (5) Window size (σ) 2 Num. of octaves 9 Num. of levels per octave 3 oFAST Target number of features 200 Threshold 0.04 Sensibility Harris factor 0 (0.04) Number of scales 8 Scale factor 1.2 CenSurE Bi-level filter STAR Non-maximum suppression threshold 0.01 (0.15) Line threshold 50 (10) Minimum and maximum scale 1 and 7 Examples of keypoints are given in Fig. 1, where it is no- ticeable how oFAST detects corner-like points around contours whereas SIFT and CenSurE detect mostly points inside the 3 SIFT oFAST CenSurE B as op hi l M on oc yt e E os in op hi l N eu tr op hi l L ym ph oc yt e B ro ke n Figure 1: Keypoints detected with different detectors on example images of different white blood cell types 4 Table 2: Number of keypoints. For the regular sampling strategies this number is constant; for the sparse keypoint detectors, the average and standard deviation for a total of 1,315 images are given Detector Sampling strategy SIFT CenSurE oFAST GUS LUS FS mLUS hLUS Average 427 161 192 2,601 961 436 3,423 961 Std. dev. 209 115 31 - - - - - B as op hi l M on oc yt e E os in op hi l N eu tr op hi l L ym ph oc yt e B ro ke n Figure 2: Gray-level images corresponding to images in Fig. 1. Keypoints detectors are actually computed on the gray-level images. cells. Since keypoints are actually computed on gray-level im- ages, the gray-level versions of these images are given in Fig. 2 as a reference. The number of keypoints that each detector finds are given in Table 2. On average, more than twice as many key- points are detected with SIFT than with oFAST or CenSurE (Ta- ble 2). In oFAST, the mean number of keypoints per image is similar to the desired maximum number. 2.2. Regular sampling strategies As an alternative to keypoint detectors, the location of in- terest points can be defined with some kind of regular grid. In this case, the points are not “interest” or “key” points any- more, since their geometric distribution is independent of the actual image contents. Although intuition may suggest that this “dense” pattern is less discriminative than locally salient points, previous works have shown that dense sampling can be effec- tive in many visual tasks (Nowak, Jurie & Triggs, 2006; Wang, Ullah, Klaser, Laptev & Schmid, 2009). Since this might also be the case for this problem, we explore several sampling strate- gies, as discussed below. Global uniform sampling (GUS) The whole image is sampled at the same density, selecting one pixel every s pixels along both the horizontal and verti- cal axes. This amount s is commonly referred to as the stride. This sampling strategy makes sense when no clear insight is available on the possibly different relevance of different image regions. It is therefore implicitly assuming that all image parts may contribute equally to characterize the entire image con- tents. Local uniform sampling (LUS) As in GUS, a regular rectangular grid is defined, but instead of sampling the entire image, only a local region of interest (ROI) is covered. The ROI is expected to contain most of the discriminative information. In our problem, since the blood cell is generally roughly centred on the image and has a similar size across images in the dataset, a centred rectangular ROI of size 312 × 312 is consequently defined, so as to roughly cover the entire or most of the blood cell and not too much of the back- ground. This ROI represents a 37% of the total area of the im- age. Foveal-like sampling (FS) This is a space-variant sampling technique, and represents a trade-off solution between GUS and LUS. Rather than sampling the whole image evenly in one case, or just focusing on one re- gion and completely ignoring the rest of the image in the other, 5 Table 3: Strides and points per region at increasing distances to the centre for the foveal-like sampling (FS) # regions stride (s) # points/region # total points 4 8 49 196 8 16 16 128 20 32 4 80 32 64 1 32 we can sample more densely over the ROI, but still sample, even though less densely, the rest of the image. In particular, a foveal-like sampling is used with sampling density being in- versely proportional to the distance to the centre. The rationale for this strategy is two-fold: • On the one hand, since the extent of the blood cell varies across images, if we sample areas outside the ROI we may still capture some parts of the relevant blood cell which lie outside the ROI. • On the other hand, contextual information is included in the representation. In some problems the context has been shown to play an important role in recognition. Therefore, we can evaluate how important is the context in this prob- lem. The space-variant sampling is implemented as follows. Non- overlapping square image regions, each of size 64×64, are con- sidered, and they are sampled inversely to their distance to the centre. The number of regions, their sampling stride s, and the resulting number of points are given in Table 3, at increasing distance to the centre. Sampling points on an example image for these three sam- pling strategies are given in Fig. 3. Hierarchical uniform sampling (hLUS) As in LUS, the image is sampled uniformly, but after the vocabulary of visual words has been built (Section 2.4), sev- eral histograms are computed, one at each cell in a quad- tree, hence the name of this idea in the literature (Bosch, Zis- serman & Muñoz, 2007), Pyramid of Histograms Of visual Words (PHOW). The rationale for this representation is to cap- ture some spatial information that is otherwise lost with a sin- gle global histogram (Lazebnik, Schmid & Ponce, 2006). As in LUS, only descriptors in a central ROI of size 312 × 312 are considered. A 3-level pyramid is used, as in (Bosch, Zisserman & Muñoz, 2007), resulting in 1+4+16 = 21 histograms per im- age. Consequently, the feature vector describing each image turns out to be relatively high (21 · k), with k being the vocabu- lary size. For computational and comparison reasons, we opted to reduce this dimensionality by applying Principal Component Analysis (PCA) (Alpaydın, 2004) and choosing k as the target dimensionality, to match that of the rest of the compared meth- ods. Multiresolution uniform sampling (mLUS) This is also a dense sampling, but computed at several image resolutions. Since it also uses a local ROI, it can be seen as a multiresolution LUS (mLUS).2 For all the regular sampling but for FS, a common stride s = 10 is used. For GUS, and mLUS this stride results in a large number of descriptors per image (specifically, 3, 423 for mLUS), which raises memory issues in the clustering procedure (the overall requirement is of 1,800 MB). To circumvent this is- sue, and obtain a feasible reduced number of points, we tried two approaches, both seeking a fair comparison with the other strategies: (1) doubling the sampling stride s (i.e. halving the number of resulting points per image); and (2) selecting ran- domly a subset of m points per image for the clustering, with m = 961 corresponding to the number of points used in LUS. We tried both approaches, but found empirically the second one to be preferable andthus it is the one whose results are reported here. The number of keypoints for each of these sampling strate- gies (Table 2), except for FS, is an order of magnitude larger than the number of keypoints selected by the sparse detectors. 2.3. SIFT descriptor As said before, every keypoint in all the considered strategies is described locally using the SIFT descriptor (Lowe, 2004). This descriptor consists of the concatenation of 16 8-bin gradi- ent orientation histograms computed over a neighbourhood of size 16 × 16 around the keypoint divided into 4 × 4 cells. For the detectors that compute an orientation (SIFT and oFAST), this is plugged into the SIFT descriptor, to endow the descriptor with rotation invariance. The result is a L1-normalized vector of length 128 (= 4 ·4 ·8). Although the dataset used in the experi- ments consists of color images, in this work the SIFT descriptor is applied to their gray-level versions. 2.4. Vocabulary and bag of words The set of SIFT descriptors (or a subset of them in the case of GUS and mLUS, as discussed above) of the images in the training set are clustered using k-means (Jain, 2010), for a user- provided value of k, and the centroids of each cluster represent the visual words of our vocabulary. To compute these clus- ters and their corresponding centroids {ci}k1, k-means starts by some (random) initial centroids and then proceeds with these two steps, which are repeated until convergence, with a maxi- mum number of iterations: 1. Given the centroids, reassign each point p ∈ {pi}n1 to the cluster corresponding to its closest centroid. 2. Given the clusters, recompute the centroid of each cluster. Then, once this vocabulary is built, a given new local descrip- tor is assigned to the word corresponding to its closest centroid. Therefore, an image in both the training and test sets is coded as a k-bin histogram (the bag of words) with the counts of the words of all its descriptors. Formally, given m local (SIFT) descriptors in an image, {q j}m1 , the k-bin histogram {hb} k 1 corre- sponding to this given image is computed as follows: 2The function which extracts the data for our mLUS within the VLfeat pack- age is called, and refers to, the PHOW method, but in our understanding PHOW actually refers to the hierarchical approach followed in our hLUS strategy. 6 GUS LUS FS Figure 3: Sampling layouts for GUS, LUS, and FS. The other two approaches (mLUS and hLUS) are based on the same sampling as LUS 1. Find the closest centroid of each point q ∈ {q j}m1 , and assign it to the corresponding cluster. Let c j ∈ {1, 2, . . . , k}, j ∈ {1, . . . m} be these assignments of the m points to the k clusters. 2. Count the number of points assigned to each cluster, i.e. hb = ∑m j=1 δ(c j, b), where δ(x, y) = 1 if x = y, and δ(x, y) = 0 if x , y. Finally, each of these histograms is L1-normalized individu- ally, per image. 2.5. Summarizing the complete procedure To summarize (see the flowchart of the whole process in Fig. 4), at both the training and the prediction stages, key- points are detected on the input images (Sect. 2.1) or, alter- natively, selected with a regular sampling strategy (Sect. 2.2). In both cases, the points are described with the SIFT descrip- tor (Sect. 2.3). At training time, from a set of SIFT descriptors from the training images, a vocabulary is built using vector quantization with k-means (Sect. 2.4). Through the vocabulary, one bag- of-words histogram (Sect. 2.4) is computed per image, both at training and prediction stages. Then, only at training, a set of histograms is used to train the SVM classifier (Sect. 3). Only at the prediction stage, the trained classifier uses the BoW histogram of a novel image to predict its type of WBC. Therefore, the most costly procedures (BoW and SVM training) are performed only once, at the training stage. The input to the training stage is a set of training images with their correspond- ing ground-truth labels, and its output is a trained SVM. The ground-truth labels are only used during SVM training, since the k-means is completely unsupervised. As for the prediction stage, the input is a new, unlabelled image, and its output is the predicted label for the input image. The prediction stage uses the vocabulary and the trained SVM obtained at the training stage. 3. Experimental work 3.1. Setup Dataset. The dataset provided by the hematologists at Hospital General, Castellón, Spain, consists of 1,315 512×512 color im- ages of 6 types of blood cells (Table 4). The samples were ran- domly chosen from a set of alarms coming from the automatic differential flow system. These blood samples were considered as potentially abnormal so they had to be visually inspected by expert haematologists. The samples were stained with the common May-Grumwald- Giemsa procedure. Hematologists labelled each individual im- age as one of the 6 classes, through a graphical user interface with the possibility of correcting their mistakes and navigating back and forth along the images. These labels are used as the ground-truth labels for training and testing purposes. Table 4: Distribution of the classes of white blood cells in the dataset Class # instances Ratio (%) Lymphocyte 511 38.86 Neutrophil 476 36.20 Broken 185 14.07 Monocyte 99 7.53 Eosinophil 38 2.89 Basophil 6 0.46 Total 1,315 100 WBC localisation and counting. The images in this dataset were previously extracted in a pre-processing stage by an au- tomatic machine vision system consisting of a microscope, a color camera, a robotic platform to move the blood smears un- der the microscope objective, and a personal computer to con- trol and perform the image processing tasks. This system is able to scan the whole surface of a blood smear and it can eas- ily differentiate at a given scale between white, red, and other main types of cells, because of their colour and sizes. Once a white cell is identified, the system zooms in and takes an image 7 (a) Training (b) Prediction Figure 4: Flowchart of the process at the (a) training and (b) prediction stages. See text for details at a larger scale of the detected white cell. It is thus the out- put images of this system that are the input to our recognition approach, and the whole system would therefore be capable of counting the different types of WBC, as required for health re- porting. Vocabulary. To select the size of the vocabulary for the BoW model, tests in an initial data exploration phase were performed, and sizes k = 150 and k = 500 were found to give the overall best performance among a large set of sizes ranging from 10 to 1,000. Subsequent tests with these two sizes revealed that results with k = 500 were slightly better and thus used for the rest of the experimentation. Classifier. A support vector machine (SVM) is used, with ei- ther linear or radial-basis function (RBF) kernels, selected by cross-validation. Given the distribution of instances per class (Table 4), there is a significant class imbalance, a situation known to require particular strategies to avoid a bias towards the majority classes. As a preliminary effort to deal with this issue, the Synthetic Minority Over-sampling Technique (SMOTE) was explored, both the SVM-SMOTE and the bor- derline SMOTE (Nguyen, Cooper & Kamei, 2011; Han, Wang & Mao, 2005). When using these approaches we appreciated a slight improvement in the recognition for the minority classes, but at the expense of a slight degradation of the recognition of the majority classes. Besides the class imbalance, too few instances of the minority classes are available, and therefore this issue is left for addressing it in future work with a bigger dataset. Validation protocol and performance assessment. A hold-out `-fold procedure is followed, which is adequate when datasets are not large, which is the case. We used ` = 5 and, for ev- ery split, the whole procedure (clustering + histogram coding + training + testing) is repeated. Additionally, as a requirement of one statistical test (introduced below), a 2-fold hold-out was performed 5 times. The common accuracy performance is complemented with precision and recall as well as with two more suitable measures under class imbalance, the F1 measure and the Matthews Corre- lation Coefficient (MCC) (Matthews, 1975; Boughorbel, Jarray & El-Anbari, 2017). Although MCC was originally thought for a binary case, its generalization for multi-class case has also been derived (Gorodkin, 2004): MCC = ∑ k ∑ l ∑ m(CkkClm − CklCmk)√√√√∑ k (∑ l Ckl )  ∑l′ k′,k Ck′l′  √√√√∑ k (∑ l Clk )  ∑l′ k′,k Cl′k′  , where Ci j are the entries of the confusion matrix, i.e. the num- ber of instances belonging to class i which are classified as 8 the class j. The possible values for this coefficient are in the range [−1, 1], the higher the better. Box-and-whisker plots (Sahay, 2016) are provided for visu- ally depicting the average performance and its variability across the 5 folds. Hyperparameter selection. For each fold, a grid search of the SVM hyperparameters is performed to select those resulting in the highest precision. The hyperparameters included the ker- nel (linear or RBF), the regularization parameter (in the range C ∈ {1, 2, . . . , 11}), and the scale parameter in the case of the RBF (in the range γ ∈ {10 j : j ∈ {−6,−5, . . . , 3, 4}}). Note that ranges covering other more extreme values for C and γ were initially tested, but after observing the trend of the values be- ing chosen, they were adjusted to the reported ones. The RBF kernel was chosen almost always for all the methods. Statistical tests. To find out the statistically significant differ- ences in performance among the different methods, we applied two significance tests: the corrected resampled t-test (Nadeau & Bengio, 2003) (CTT), and the 5×2cv paired t-test (5×2cv) (Di- etterich, 1998), which are known to be suitable for comparing two methods on a single dataset. These test statistics follow a Student-t distribution t f , with f degrees of freedom, and are computed as follows (Bouckaert & Frank, 2004): CTT ≡ µ√( 1 ` + nte ntr ) ·σ2 ∼ t`−1, 5×2cv ≡ d11√ 1 5 ∑5 j=1 σ 2 j ∼ t5, where • for CTT, ` is the total number of runs of the method; ntr and nte are, respectively, the number of training and test- ing instances in each run; and µ ad σ are the estimates of the mean and standard deviation of the ` differences of the chosen performance measure for the two compared meth- ods; and • for 5×2cv, let ai j, bi j, i ∈ {1, . . . ,`}, j ∈ {1, . . . , r} be the performance measures of two compared methods (say A and B) for the r-times `-fold cross-validation (here, ` = 2, r = 5), using exactly the same training and testing sets for both methods in every run; then di j = ai j − bi j (the in- dividual differences in performance), and σ j = ∑2 i=1(xi j − d j)2, with d j = x1 j +x2 j 2 (the mean difference for a single run j). For the sake of convenience, the outcome of these tests will be presented graphically following the symbols in Table 5. Software. As supporting software, we used the VLfeat li- brary (Vedaldi & Fulkerson, 2008) for the SIFT detection, OpenCV (Howse, Joshi & Beyeler, 2016) for the SIFT de- scription, and Python packages for sparse detectors (scikit- image (van der Walt, Schönberger, Nunez-Iglesias, Boulogne, Table 5: Visual representations of significance degree found by the statistical test given a p-value. The lower the p-value, the higher the significance p < 0.01 p ≤ 0.1 p ≤ 0.05 p > 0.1 (no significance) Warner, Yager, Gouillart, Yu & the scikit-image contributors, 2014)), SVM, grid search and PCA (scikit-learn (Pedregosa, Varoquaux, Gramfort, Michel, Thirion, Grisel, Blondel, Pret- tenhofer, Weiss, Dubourg, Vanderplas, Passos, Cournapeau, Brucher, Perrot & Duchesnay, 2011)), and clustering, vector quantization, and t-test lookups (scipy (Jones, Oliphant, Peter- son et al., 2001–)). 3.2. Results Experimental results are discussed subsequently through the analysis of the boxplots, the confusion matrices, the statistical tests and the computation times. Performance comparison with other approaches are also provided as a general reference. Boxplots. By observing the boxplots (Fig. 5), it can be found that purely uniform sampling strategies (GUS and LUS) per- form the worst. Among them, it seems a better idea to focus on a central ROI (LUS), otherwise there seems to be too much noisy and irrelevant background included and this may end up being harmful. Interestingly, the foveal-like strategy (FS), by gradually decreasing the sampling density away from the image centre, provides a nice and better solution: the fact of captur- ing a bit of background with a coarser sampling seems to pro- vide useful contextual information. There exists not much dif- ference between the multi-resolution scheme (mLUS) and the hierarchical one (hLUS), and both tend to outperform the uni- form sampling strategies either global (GUS) or local (LUS). A bit against our intuition, the multi-resolution (mLUS) and hier- archical (hLUS) strategies performed very well compared with the other regular sampling strategies and the keypoint detectors. This might be explained by the fact that some form of discrim- inative scale or texture-like information is better captured. The three sparse detectors provide generally a good recogni- tion rate, with oFAST possibly resulting in (slightly) better per- formance. Compared to the regular sampling strategies, oFAST has a comparable performance to mLUS or hLUS. The statisti- cal significance analysis below provides some insight into the actual relevance of the similar or different performances among the methods. Confusion matrices. Confusion matrices averaged over the ` folds (Fig. 6) provide insights into how the different types of white blood cells are recognized by the different strate- gies. Neutrophil and Lymphocyte are generally the classes with higher recognition rates, which might be related to the fact that these are the classes with the most instances. A further reason is that Lymphocyte is a relatively distinct class with respect to any other class. The strong visual similarity, even for a human ob- server, between Neutrophil and Eosinophil, possibly explains why all the methods tend to confuse them, and calls for very 9 Figure 5: Recognition performance for the different detectors specialized knowledge in terms of human expertise and/or auto- matic feature learning. Basophils are often confused with a va- riety of other blood types, with oFAST being the detector which leads to better performance. No blood cell instance is classified as Basophil, most likely due to the very few instances of this under-represented class. In particular, the frequent confusion 10 of Basophils with Lymphocytes has also been reported by other authors (Habibzadeh, Krzyżak & Fevens, 2013). Monocytes are often confused with either Neutrophils or Lymphocytes. In- terestingly, some strategies (LUS, GUS) misclassify Monocyte more as Neutrophil, and others (CenSurE, oFAST, FS) more as Lymphocyte, which indicates the different information captured by sparse detectors and foveal-like sampling on the one hand, and the uniform sampling on the other. Significance analysis. The results of the significance tests roughly correspond to the intuitive visual interpretation one can get by comparing the box-plots. Furthermore, as it hap- pens with these plots, the tests applied to different metrics are very similar; therefore, only results for a single metric (accu- racy) are shown. For easier visual comparison, the visual rep- resentation of the significance level is given, instead of the the p-values (Table 6). Although both tests reveal many common differences among the methods, there are also some discrepan- cies. Overall, both tests reveal that the dense method LUS per- forms differently (worse) than most of the other methods (both dense and sparse). Both tests also agree to find CenSurE and oFAST different. And neither of the tests find any significant difference between the top performing regular-sampling meth- ods (FS, mLUS, hLUS). Comparison with other works. Since no standard public dataset of white blood cells is available for benchmarking, it is not possible to compare directly the different approaches. Furthermore, since the datasets used in these works are rela- tively small (often fewer than 100 instances), and some of them even use fewer classes (in one case, 3 instead of 6), the re- sults are also less representative and/or the classification tasks easier. However, and with these caveats in mind, it can still be illustrative to compare ours with the performance reported in other studies with a similar problem but different datasets. Some segmentation-based approaches are reported to have ac- curacies about 90% or higher. With a maximum mean accu- racy of 79% and a peak accuracy of 85%, the proposed strategy has comparable performance while being simpler and more ro- bust. The particularly good performance of some works such as (Rezatofighi & Soltanian-Zadeh, 2011), with an accuracy of 96% (on a dataset of 251 instances), may suggest that care- fully crafted and complex and possibly costly segmentation procedures such as snakes, the use of color and texture infor- mation, and many ad hoc design decisions and pre-processing steps (histogram equalization, color space transformation, im- age smoothing, etc.) may lead to good performance. In con- trast, the approaches tested in our work are much more general, work on gray-level images, perform no image pre-processing at all, and have no explicit segmentation (besides a rough ROI central selection in some of the cases). The promising results of the studied strategies and the ob- servation of the reported results of the alternative approaches suggest possibilities for recognition improvement within the framework of local descriptors and BoW, thus getting closer to what experts in this domain consider acceptable (e.g. accu- racies around 90%). Computation times. Given the similar performance of several of the tested methods, it is important to compare them also in computational terms. The times in our computer (Intel c©CoreTM i5 650, 8-GB RAM, ATI Radeon HD4650 GPU) and (unop- timized) Python implementation are given in Fig. 7. The de- tection time has been computed on a subset of about 20% of the total dataset, and scaled proportionally to represent the full dataset. It can be seen that detecting or extracting the points is gener- ally the most costly procedure, and apart from the mLUS case, it is naturally bigger when using actual detectors than in sam- pling strategies. The time for CenSurE is particularly large due to the bi-level filter chosen (STAR); other faster but less accu- rate filters are possible (see below). In contrast, the time for building the vocabulary, which is proportional to the number of points used in the clustering, is generally bigger with the dense methods. Similarly, computing the histograms depends on the number of points per image, and generally represents the small- est percentage of the total time. Finally, the training time of the classifier is essentially the same for all methods because the training data are the already-computed histograms of the same number of bins in all cases. Taken together, and excluding the most costly CenSurE and mLUS, computation times for the sparse and dense methods are comparable, although the sparse keypoint detectors might be slightly advantageous. It is important to bear in mind that build- ing the vocabulary and learning a classifier are off-line proce- dures, which usually need to be computed only once. At predic- tion time, the relevant costs include detecting and describing the keypoints, computing the histograms from the pre-computed vocabulary, and predicting the label from the already learned classifier. These steps can be computed fairly quickly and may be acceptable for real settings unless hard real-time constraints are demanded (e.g. for a responsive user interface). Accuracy-time. When assessing the methods by consider- ing jointly their recognition and computational performances (Fig. 8), the oFAST and the hLUS seem the most advantageous methods. Depending on the required classification-time trade- off, the SIFT and FS methods can also be competitive alterna- tives. Although the mean accuracy is only a partial picture of the recognition ability, this scatter plot provides an interesting insight into the overall recognition and computational perfor- mances. Another fact worth observing is how a faster kernel (Difference of Box, DoB) in CenSurE results in lower discrim- inative power than the more costly STAR filter. 4. Discussion This section is aimed at clarifying some aspects regarding the context, strengths and weaknesses of the proposed approach, as well as the limitations of the study, possible future work and the relevance of the contribution in the context of intelligent systems. 11 (a) SIFT (b) CenSurE (c) oFAST (d) FS (e) LUS (f) GUS (g) mLUS (h) hLUS Figure 6: Confusion matrices for the different detectors (a–c) and sampling strategies (d-h). True classes as rows, predicted ones as columns 12 Table 6: Results of significance tests on the accuracy CTT test CenSurE oFAST GUS LUS FS mLUS hLUS SIFT CenSurE oFAST GUS LUS FS mLUS 5×2cv test CenSurE oFAST GUS LUS FS mLUS hLUS SIFT CenSurE oFAST GUS LUS FS mLUS Figure 7: Time breakdown for the different steps of the BoW for each method 4.1. Image segmentation It is important to distinguish two types of image segmenta- tions that may arise in the context of the real-world problem: on the one hand, there are segmentation-for-localisation methods to roughly localise the cells of interest within a larger image, and whose solution in our work is taken for granted, since it is given by the robotized system, as described in Sect. 3, and it is out of the scope of this work. This segmentation is akin to the nucleus segmentation step in works such as (Rezatofighi & Soltanian-Zadeh, 2011). On the other hand, there are segmentation-for-characterisation methods, whose purpose is to extract features within the image region of the cell or within the nucleus and cytoplasm regions, separately (Rezatofighi & Soltanian-Zadeh, 2011). Figure 8: Accuracy-time scatter plot of the different strategies. The size of the circles is proportional to the variance in the accuracy In this respect, the proposed approach relies on some kind of segmentation-for-localisation procedure, and it is therefore sensitive to the potential failures of this system. However, it does not need any segmentation-for-characterisation method, and this fact is a strength of the approach (as discussed below in Sect. 4.3). 4.2. Computational and robustness issues Regarding the computational cost, although a complete and rigorous comparison with alternative approaches is out of the scope of this paper, our estimated computational time and those reported in papers of related work are provided (Table 7), just as a rough guideline. Our reported times correspond to the most and the least costly strategies, which differ in one order of mag- nitude. The times for the remaining strategies lie in between these two times. The caveat for the times in the table is obvi- ously that they correspond to very different settings, in terms 13 of image sizes, implementation choices, and computing facili- ties, and are therefore not directly comparable, but they can be somehow illustrative of the computational efficiency of local- descriptor-based methods. As for the robustness issue, general segmentation algo- rithms are known to easily over-segment or under-segment im- ages (Estrada & Jepson, 2009). Because of their nature (low contrast, noise, etc.), medical images can be particularly chal- lenging to segment and, despite the progress made, accurate segmentations are elusive (Elnakib, Gimel’farb, Suri & El- Baz, 2011), and robust segmentation remains an unsolved prob- lem (Zhang & Metaxas, 2016). In contrast to segmentation- based methods, local descriptors do not rely on particular im- age regions being defined. The holistic and general purpose of this methodology makes it inherently robust to image contents, being tolerant to the presence of significant background clut- ter (Zhang, Marszałek, Lazebnik & Schmid, 2007). 4.3. Strengths and weaknesses When comparing the proposed approach to segmentation- based approaches (Table 8), it can be observed that the same characteristic can be seen as an asset or become a weak point of the methodology. For instance, segmentation-for- characterisation (either segmenting the cell from the back- ground or even separating the nucleus from cytoplasm) is a good asset since very specific features can be computed, but since robust segmentations tends to be elusive, the reliability of these features can be compromised. Likewise, the use of lo- cal descriptors makes only mild assumptions of image contents, and this is a great advantage. However, the approach might fail under strong deviations from these assumptions, which might happen if the cell is too small with respect to the background area, or there are too many distractive background features, or more than a single type of cell is present in the image. If the relevant cell region was highly off-centred, regular-sampling strategies may be affected significantly more than keypoint de- tectors. Both approaches share the limitation of being dependent on the quality of the output of the segmentation-for-localisation step. 4.4. Future work Exploring the discriminative role of color. This work used the SIFT descriptor on gray-level images since single-channel de- scriptors are most known and explored. However, since color information may provide discriminative information for WBC recognition, the role of color is an interesting open topic to ex- plore. Possibilities range from early fusion (e.g. concatenat- ing SIFT descriptors in different color bands) and late fusion (e.g. combining classifiers) to color extensions of the SIFT descriptor and alternative descriptors (van de Sande, Gevers & Snoek, 2010; Guo, Huang & Qiao, 2017). Depending on the problem, the selection of the descriptors with the adequate amount of invariance to illumination and color changes is rec- ommended (van de Sande, Gevers & Snoek, 2010). Combining segmentation-based and segmentation-free ap- proaches. The former comparison (Sect. 4.3, Table 8), includ- ing the dichotomy between explainability and problem agnosti- cism discussed in the introduction, reveals that the strengths and weaknesses of local descriptors (segmentation-free) approach and segmentation-based approaches are essentially comple- mentary. Consequently, the design of a system combining the strengths of both worlds may be considered. More specifically, an open issue would be how to endow approaches based on local descriptors with the problem-domain insights that can ar- guably be obtained more naturally with segmentation-based ap- proaches. Active learning. A common issue that the recognition prob- lem tackled in this work and actually of many classifier-based learning systems is that they stop learning after a single (ex- tensive) training stage. This poses a serious practical limita- tion since even though the experts may detect the system’s mis- classification during its predition stage, the system is unable to learn from its own mistakes. In this sense, less explored learn- ing paradigms, such as on-line and incremental learning can be very valuable. Of particular interest in the scope of expert and intelligent systems is the combination of incremental and active learning (Settles, 2012), so that the system not only can keep learning incrementally from an initial extensive learning stage, but should also be able to, very selectively, ask the experts for the true label of cells and exploit this knowledge to update its model and perform better in the future. The human experts are thus not overloaded with many images to label explicitly. Their intervention is reduced to instances whose ground-truth label the system considers most valuable, thus making the most of human-expert skills and effort. As a longer-term, more ambi- tious endeavour, the problem can be studied with the perspec- tive of never-ending learning (Mitchell, Cohen, R., Talukdar, Yang, Betteridge, Carlson, Mishra, Gardner, Kisiel, Krishna- murthy, Lao, Mazaitis, Mohamed, Nakashole, Platanios, Ritter, Samadi, Settles, Wang, Wijaya, Gupta, Chen, Saparov, Greaves & Welling, 2018), whose design can be more challenging. 5. Conclusions This work has shown that white blood cells types can be rec- ognized with reasonable performance by means of local im- age descriptors and a conventional bag-of-words pipeline, us- ing only gray-level images, without the need of complex and error-prone explicit image segmentation, and without carefully crafted features, which makes this approach robust and much more general. A variety of sampling strategies and sparse de- tectors have been shown to produce generally satisfactory and very promising performances. Among the three sparse detectors, oFAST tends to outper- form the other two (SIFT and CenSurE). This suggests that vi- sual structures nearby the border of the cell might be slightly more discriminative than its interior (CenSurE) or other sparse points distributed throughout the cell image (SIFT). As for the sampling strategies, foveal, multi-resolution and spatial pyramids yields similar performance, without a clear 14 Table 7: Average times to process a single image Approach Image size Time (s) Gautam et al. (2016) 480 × 640 22 Rezatofighi & Soltanian-Zadeh (2011) 141 × 141 10 Our approach CenSurE (STAR) 512 × 512 2.7 LUS 0.3 Table 8: Comparison of segmentation-based and local descriptors-based approaches Approach Strengths (:) and weaknesses (�) Segmentation-based : Specific features for the cell or nucleus and cytoplasm regions can be obtained : Meaningful expert-based ad-hoc features can be explored (shapes, geometric ratios, etc.) : In principle, more directly suitable for explainable solutions � Problem-specific features have to be designed and implemented � Correct segmentations are generally hard to obtain and may be initialisation-dependent � Subsequent steps (and in turn, results) rely on robust segmentations Local descriptors : Generic, problem-agnostic methodology; no specific features have to be devised and implemented : Segmentation-free approach; only mild assumptions on cell-background contents : Robust against background contents; useful contextual information may be exploited � Potentially useful features such as shape or size cues are not explictly captured � Sensitive to strong deviations from the mild assumptions � In principle, less suitable for explainable solutions winner, and outperform the pure uniform-sampling strategies either covering the whole image or a central region of interest. The result for the foveal-like sampling is interesting in that it suggests the importance of taking into account a moderate form of contextual information, i.e. a little of background informa- tion might help to discriminate, but too much of it may turn out to be harmful. Furthermore, the good performance of the multi-resolution and hierarchical strategies might relate to their ability to encode useful scale and texture information. Given the different nature of some detectors and their (some- how) complementary performance, further work might explore some fusion strategies. It might also be investigated whether descriptors which capture color information and (more explic- itly) shape and texture cues can provide higher discrimina- tion ability. Another research avenue is the comparison of segmentation-based and segmentation-free approaches as well as smartly combining them to get the benefits of both. Studying incremental and active learning paradigms for this type of in- telligent system would increase its utility in real-world medical settings. Acknowledgments. This work has been partially supported by project P1·1B2014-09 of the Pla de Promoció de la Investigació from Jaume-I University. It has been carried out with the col- laboration of the Departament of Hematology of the Hospital General Universitario de Castellón. Sharing statement. The labelled image dataset will be made publicly available or upon request (publication permission and conditions are pending). References Agrawal, M., Konolige, K., & Blas, M. R. (2008). CenSurE: Center surround extremas for realtime feature detection and matching. In European Confer- ence on Computer Vision (pp. 102–115). Alpaydın, E. (2004). Introduction to Machine Learning. The MIT Press. Bikhet, S. F., Darwish, A. M., Tolba, H. A., & Shaheen, S. I. (2000). Segmen- tation and classification of white blood cells. In International Conference on Acoustics, Speech, and Signal Processing (pp. 2259–2261). Bosch, A., Zisserman, A., & Muñoz, X. (2007). Image classification using random forests and ferns. In International Conference on Computer Vision (pp. 1–8). Bouckaert, R. R., & Frank, E. (2004). Evaluating the replicability of signifi- cance tests for comparing learning algorithms. In Pacific-Asia Conference on Advances in Knowledge Discovery and Data Mining (pp. 3–12). Boughorbel, S., Jarray, F., & El-Anbari, M. (2017). Optimal classifier for im- balanced data using Matthews correlation coefficient metric. PLoS ONE, 12. Dietterich, T. G. (1998). Approximate statistical tests for comparing supervised classification learning algorithms. Neural Computation, 10, 1895–1923. Elnakib, A., Gimel’farb, G., Suri, J. S., & El-Baz, A. (2011). Medical image segmentation: A brief survey. In A. S. El-Baz, R. Acharya U, A. F. Laine, & J. S. Suri (Eds.), Multi Modality State-of-the-Art Medical Image Segmen- tation and Registration Methodologies: Volume II (pp. 1–39). Estrada, F. J., & Jepson, A. D. (2009). Benchmarking image segmentation algorithms. International Journal of Computer Vision, 85, 167–181. Gautam, A., Singh, P., Raman, B., & Bhadauria, H. (2016). Automatic classi- fication of leukocytes using morphological features and Naı̈ve Bayes classi- fier. In Region 10 Conference (pp. 1023–1027). Gómez-Gil, P., Ramı́rez-Cortés, M., González-Bernal, J., Pedrero, Á. G., Prieto-Castro, C. I., Valencia, D., Lobato, R., & Alonso, J. E. (2008). A feature extraction method based on morphological operators for automatic 15 classification of leukocytes. In Mexican International Conference on Artifi- cal Intelligence (pp. 227–232). IEEE. Gorodkin, J. (2004). Comparing two k-category assignments by a k-category correlation coefficient. Computational biology and chemistry, 28, 367–374. Gunning, D. (2016). Explainable artificial intelligence (XAI). Technical Re- port DARPA-BAA-16-53 Defense Advanced Research Projects Agency, (DARPA). Guo, S., Huang, W., & Qiao, Y. (2017). Improving scale invariant feature transform with local color contrastive descriptor for image classification. Journal of Electronic Imaging, 26. Habibzadeh, M., Krzyżak, A., & Fevens, T. (2013). White blood cell differen- tial counts using convolutional neural networks for low resolution images. In International Conference on Artificial Intelligence and Soft Computing (pp. 263–274). Han, H., Wang, W.-Y., & Mao, B.-H. (2005). Borderline-SMOTE: A new over- sampling method in imbalanced data sets learning. In D.-S. Huang, X.- P. Zhang, & G.-B. Huang (Eds.), International Conference on Intelligent Computing. Hiremath, P., Bannigidad, P., & Geeta, S. (2010). Automated identification and classification of white blood cells (leukocytes) in digital microscopic images. International Journal of Computer Applications, (pp. 59–63). Holzinger, A., Biemann, C., Pattichis, C. S., & Kell, D. B. (2017). What do we need to build explainable AI systems for the medical domain? CoRR, abs/1712.09923. Howse, J., Joshi, P., & Beyeler, M. (2016). OpenCV: Computer Vision Projects with Python. Packt. Jain, A. K. (2010). Data clustering: 50 years beyond k-means. Pattern recog- nition letters, 31, 651–666. Jones, E., Oliphant, T., Peterson, P. et al. (2001–). SciPy: Open source scientific tools for Python. http://www.scipy.org. Kamentsky, L. A. (1973). Cytology automation. Advances in Biological and Medical Physics, 14, 93–161. Krig, S. (2014). Interest point detector and feature descriptor survey. In Computer Vision Metrics: Survey, Taxonomy, and Analysis (pp. 217–282). Berkeley, CA: Apress. Laptev, I. (2005). On space-time interest points. International Journal of Com- puter Vision, 64, 107–123. Laptev, I., & Lindeberg, T. (2003). Space-time interest points. In International Conference on Computer Vision (pp. 432–439). Lazebnik, S., Schmid, C., & Ponce, J. (2006). Beyond bags of features: Spa- tial pyramid matching for recognizing natural scene categories. In IEEE Conference on Computer Vision and Pattern Recognition (pp. 2169–2178). Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60, 91–110. Matthews, B. W. (1975). Comparison of the predicted and observed secondary structure of T4 phage lysozyme. Biochimica et Biophysica Acta – Protein Structure, 405, 442–451. Mikolajczyk, K., & Schmid, C. (2005). A performance evaluation of local de- scriptors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27, 1615–1630. Mitchell, T. M., Cohen, W. W., R., H. J. E., Talukdar, P. P., Yang, B., Betteridge, J., Carlson, A., Mishra, B. D., Gardner, M., Kisiel, B., Krishnamurthy, J., Lao, N., Mazaitis, K., Mohamed, T., Nakashole, N., Platanios, E. A., Ritter, A., Samadi, M., Settles, B., Wang, R. C., Wijaya, D., Gupta, A., Chen, X., Saparov, A., Greaves, M., & Welling, J. (2018). Never-ending learning. Communications of the ACM, 61, 103–115. Nadeau, C., & Bengio, Y. (2003). Inference for the generalization error. Ma- chine Learning, 52, 239–281. Nguyen, H. M., Cooper, E. W., & Kamei, K. (2011). Borderline over-sampling for imbalanced data classification. International Journal of Knowledge En- gineering and Soft Data Paradigms, 3, 4–21. Nowak, E., Jurie, F., & Triggs, B. (2006). Sampling strategies for bag-of- features image classification. In A. Leonardis, H. Bischof, & A. Pinz (Eds.), European Conference on Computer Vision (pp. 490–503). Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Pas- sos, A., Cournapeau, D., Brucher, M., Perrot, M., & Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830. Piuri, V., & Scotti, F. (2004). Morphological classification of blood leucocytes by microscope images. In International Conference on Computational In- telligence for Measurement Systems and Applications (pp. 103–108). IEEE. Rezatofighi, S. H., & Soltanian-Zadeh, H. (2011). Automatic recognition of five types of white blood cells in peripheral blood. Computerized Medical Imaging and Graphics, 35, 333–343. Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016 (pp. 1135–1144). Rosten, E., & Drummond, T. (2005). Fusing points and lines for high per- formance tracking. In International Conference on Computer Vision (pp. 1508–1515). IEEE volume 2. Rosten, E., & Drummond, T. (2006). Machine learning for high-speed corner detection. In A. Leonardis, H. Bischof, & A. Pinz (Eds.), European Confer- ence on Computer Vision (pp. 430–443). Springer. Rublee, E., Rabaud, V., Konolige, K., & Bradski, G. (2011). ORB: An efficient alternative to SIFT or SURF. In International Conference on Computer Vision (pp. 2564–2571). IEEE. Sabino, D. M. U., da Fontoura Costa, L., Rizzatti, E. G., & Zago, M. A. (2004). A texture approach to leukocyte recognition. Real-Time Imaging, 10, 205– 216. Sahay, A. (2016). Data Visualization, Volume I. Business Expert Press. van de Sande, K., Gevers, T., & Snoek, C. (2010). Evaluating Color Descriptors for Object and Scene Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32, 1582–1596. Saphiro, H. M. (2003). Practical flow cytometry. Wiley-Liss. Sarrafzadeh, O., Dehnavi, A. M., Rabbani, H., & Talebi, A. (2015). A sim- ple and accurate method for white blood cells segmentation using k-means algorithm. In Workshop on Signal Processing Systems (pp. 1–6). Settles, B. (2012). Active Learning. Morgan & Claypool. Song, X. B., Abu-Mostafa, Y. S., Sill, J., & Kasdan, H. (1997). Incorporating contextual information in white blood cell identification. In Advances in Neural Information Processing Systems (pp. 950–956). Tuytelaars, T., & Mikolajczyk, K. (2008). Local invariant feature detectors: A survey. Foundations and Trends R© in Computer Graphics and Vision, 3, 177–280. Vedaldi, A., & Fulkerson, B. (2008). VLFeat: An open and portable library of computer vision algorithms. http://www.vlfeat.org/. Viola, P. A., & Jones, M. J. (2004). Robust real-time face detection. Interna- tional Journal of Computer Vision, 57, 137–154. van der Walt, S., Schönberger, J. L., Nunez-Iglesias, J., Boulogne, F., Warner, J. D., Yager, N., Gouillart, E., Yu, T., & the scikit-image contributors (2014). scikit-image: image processing in Python. PeerJ, 2, e453. Wang, H., Ullah, M. M., Klaser, A., Laptev, I., & Schmid, C. (2009). Evaluation of local spatio-temporal features for action recognition. In British Machine Vision Conference (pp. 124.1–124.11). BMVA Press. Wang, Y., & Mori, G. (2009). Human action recognition by semilatent topic models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31, 1762–1774. Young, I. T. (1972). The classification of white blood cells. IEEE Transactions on Biomedical Engineering, BME-19, 291–298. Zhang, J., Marszałek, M., Lazebnik, S., & Schmid, C. (2007). Local features and kernels for classification of texture and object categories: A comprehen- sive study. International Journal of Computer Vision, 73, 213–238. Zhang, S., & Metaxas, D. (2016). Large-scale medical image analytics: Recent methodologies, applications and future directions. Medical Image Analysis, 33, 98–101. 16 
work_nwb7pklhefdaxln3iskkriitom	----	[PDF] Modeling and identification of microbial batch fermentation using fuzzy expert system | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1016/J.APM.2013.02.042 Corpus ID: 16416262Modeling and identification of microbial batch fermentation using fuzzy expert system @article{Gao2013ModelingAI, title={Modeling and identification of microbial batch fermentation using fuzzy expert system}, author={Jinggui Gao and L. Wang and E. Feng and Z. Xiu}, journal={Applied Mathematical Modelling}, year={2013}, volume={37}, pages={8079-8090} } Jinggui Gao, L. Wang, +1 author Z. Xiu Published 2013 Engineering Applied Mathematical Modelling The bioconversion of glycerol to 1,3-Propanediol by Klebsiella pneumoniae (K. pneumoniae) is a complex bioprocess. Because the transport mechanism of glycerol across cell membrane of K. pneumoniae was still not adequately understood, it is difficult to formulate reliable mathematical models. In this study, we aim to explore a novel model for describing the process of glycerol batch fermentation. This is achieved by a hybrid of a fuzzy expert system with a physical model framework. Some… Expand View via Publisher doi.org Save to Library Create Alert Cite Launch Research Feed Share This Paper 10 CitationsBackground Citations 1 Methods Citations 1 View All Figures and Tables from this paper figure 1 table 1 figure 2 figure 3 figure 4 figure 5 View All 6 Figures & Tables 10 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency Modeling and identification of hybrid dynamic system in microbial continuous fermentation Yanan Mao, C. Gao, Ruidong Yan, Aruna Bai Computer Science 2015 1 Save Alert Research Feed MICROBIAL CONTINUOUS FERMENTATION Yanan Mao, C. Gao, Ruidong Yan, Aruna Bai 2015 Save Alert Research Feed Modeling the interaction between the central carbon metabolism of Escherichia coli and bioreactor culture media Fabián A Ortega-Quintana, M. Trujillo-Roldán, Héctor Botero-Castro, H. Álvarez Mathematics 2020 1 Save Alert Research Feed Optimal Control of a Nonlinear Time-Delay System in Batch Fermentation Process Y. Yu Mathematics 2014 6 PDF Save Alert Research Feed Analysis of hydrogen production by anaerobic fermentation from urban organic waste Edilson León Moreno Cárdenas, A. D. Zapata, F. Mejía Chemistry 2015 PDF Save Alert Research Feed Advances in bioconversion of glycerol to 1,3-propanediol: Prospects and challenges Yaqin Sun, J. Shen, +6 authors Z. Xiu Environmental Science 2018 34 Save Alert Research Feed Analysis of hydrogen production by anaerobic fermentation from urban organic waste Edilson León Moreno-Cárdenas, Arley David Zapata-Zapata, Fernando Álvarez-Mejía Mathematics 2015 Save Alert Research Feed A hybrid approach to optimize fuzzy if-then rules L. Aliahmadipour, E. Eslami Mathematics 2014 Iranian Conference on Intelligent Systems (ICIS) 2014 View 1 excerpt, cites methods Save Alert Research Feed Between the Poles of Data‐Driven and Mechanistic Modeling for Process Operation D. Solle, B. Hitzmann, +5 authors T. Steckenreiter Engineering 2017 20 Save Alert Research Feed Enfrentando el modelado de bioprocesos: una revisión de las metodologías de modelado F. O. Quintana, H. Álvarez, Héctor Antonio Botero Castro Philosophy 2017 4 PDF View 3 excerpts, cites background Save Alert Research Feed References SHOWING 1-10 OF 42 REFERENCES SORT BYRelevance Most Influenced Papers Recency Modeling and parameter identification of microbial bioconversion in fed-batch cultures Gang Wang, E. Feng, Z. Xiu Engineering 2008 43 Save Alert Research Feed Vector measure for explicit nonlinear impulsive system of glycerol bioconversion in fed-batch cultures and its parameter identification G. Wang, E. Feng, Z. Xiu Mathematics, Computer Science Appl. Math. Comput. 2007 21 Save Alert Research Feed Nonlinear hybrid kinetic system of microbial bioconversion in fed-batch culture Gang Wang, E. Feng, Z. Xiu Mathematics 2008 15 View 1 excerpt, references methods Save Alert Research Feed Complex metabolic network of glycerol fermentation by Klebsiella pneumoniae and its system identification via biological robustness J. Wang, Jianxiong Ye, E. Feng, H. Yin, B. Tan Chemistry 2011 20 View 4 excerpts, references background Save Alert Research Feed Pathway identification using parallel optimization for a complex metabolic system in microbial continuous culture Jingang Zhai, Jianxiong Ye, L. Wang, Enmin Feng, H. Yin, Z. Xiu Mathematics 2011 26 Highly Influential View 5 excerpts, references background and methods Save Alert Research Feed Nonlinear dynamical systems of bio-dissimilation of glycerol to 1,3-propanediol and their optimal controls C. Gao, E. Feng, Zongtao Wang, Z. Xiu Mathematics 2005 37 Save Alert Research Feed Mathematical modeling of kinetics and research on multiplicity of glycerol bioconversion to 1,3-propanediol A. Li Chemistry 2000 63 Save Alert Research Feed Bioprocess optimization and control: Application of hybrid modelling Jörg Schubert, R. Simutis, M. Dors, I. Havlik, A. Lübbert Computer Science 1994 140 View 1 excerpt, references background Save Alert Research Feed Modeling and Robustness Analysis of Biochemical Networks of Glycerol Metabolism by Klebsiella Pneumoniae Jianxiong Ye, E. Feng, L. Wang, Z. Xiu, Yaqin Sun Biology, Computer Science Complex 2009 7 Save Alert Research Feed Hybrid modelling of biochemical processes: A comparison with the conventional approach S. Azevedo, B. Dahm, F. Oliveira Engineering 1997 90 PDF View 1 excerpt, references background Save Alert Research Feed ... 1 2 3 4 5 ... Related Papers Abstract Figures and Tables 10 Citations 42 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_nx3k3cao7zhb3jn6vvxohbfx6m	----	doi:10.1016/j.eswa.2004.08.011 Web enabled expert systems using hyperlink-based inference Wooju Kim a,*, Yong U. Song b,1 , June S. Hong c,2 a Department of Computer and Industrial Systems Engineering, Yonsei University, 134 Shinchon, Seoul 120-749, South Korea b Department of Management Information Systems, Yonsei University Wonju Campus, 234 Maji, Wonju, Gangwon 220-710, South Korea c Division of Business Administration, Kyonggi University, San 94-6 Yiui-Dong, Paldal-Gu, Suwon, Kyonggi 442-760, South Korea Abstract With the proliferation of the WWW, providing more intelligent Websites has become a major concern in the e-business industry. Recently, this trend has been even more accelerated by the success of Customer Relationship Management (CRM) in terms of product recommendation, and self after service, etc. As a result, many e-companies are eager to embed Web-enabled, rule-based systems, i.e. that is, expert systems, into their Websites, and several well-known commercial tools to facilitate this are already available in the market. So far, most of those tools are based on CGI, but CGI-based systems inherently suffer from problems related to overburdening, when there are too many service demands at the same time. To overcome the limitations of the existing CGI-based expert systems, we propose a new form of Web-enabled expert system that uses a hyperlink-based inference mechanism. In terms of burden to the Web server, our approach has proven to outperform the CGI-based approach theoretically as well as empirically. For practical purposes, our approach is implemented in a software system, WeBIS, and uses a graphic rule editing methodology, Expert’s Diagram, that is incorporated into the system to facilitate rule generation and maintenance. WeBIS is now successfully operating for financial consultation on the Website of a leading financial consulting company in Korea. q 2004 Published by Elsevier Ltd. Keywords: Expert Systems; HTML; Inference; JavaScript; Rule; WWW 1. Introduction With the advent of the customer-centric paradigm in business strategy, e-companies go beyond merely making the sale to acquire relative competitiveness in terms of customer satisfaction, thereby maximizing the lifetime value of a business relationship (McCarthy, 1999). Giving more intelligence to e-commerce sites is popularly recog- nized as one of the effective strategies that increase customer satisfaction because they react intelligently and can give a personalized response to each customer. Current Web-enabled, rule-based system tools or expert system tools such as Blaze Advisor (Fair Isaac Corporation) and ILOG JRules are playing a major role in more intelligent Websites. This Web-enabled, rule-based inference technol- ogy is applied to various areas of application such as product 0957-4174/$ - see front matter q 2004 Published by Elsevier Ltd. doi:10.1016/j.eswa.2004.08.011 * Corresponding author. Tel.: C82 2 2123 5716; fax: C82 2 364 7807. E-mail address: wkim@yonsei.ac.kr (W. Kim). 1 Tel.: C82 33 760 2340; fax: C82 33 763 4324. 2 Tel.: C82 31 249 9459; fax: C82 31 249 9459. recommendations, distance learning and training, and help desks for technical support, etc. Most of the Web-enabled, rule-based systems have been developed using CGI technology, but these CGI-based systems usually cause a relatively high burden to Web servers in terms of both required memory and response time. In the worst case, the system may crash causing a very critical problem to most commercial Websites. One major reason for such crash is that each individual request to a CGI-based system requires higher resources than requests for HTML documents. Therefore, if we can build a rule- based inference system based only on HTML documents, we can be sure that the same inference task can be performed much more efficiently than the CGI-based approach within a given computer resource. If the HTML-based approach is better, the first problem that presents itself is the building of a system that can perform the same rule-based inference tasks using only one set of HTML documents. Song and Lee (2000) have already shown that a rule can be represented by a set of HTML documents hyperlinked to each other, and the rule can be inferred by following the hyperlinks. To solve this problem, Expert Systems with Applications 28 (2005) 79–91 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa Table 1 Categories of Web-enabled, rule-based systems Location of inference engine Type of inference engine Technical features Server-side CGI program Uses CGI (Common Gateway Interface) standard. Web server invokes a CGI program while passing required parameters according to the CGI standard. Server-side script Inference engines are developed under environments such as JSP, ASP, and PHP. Transaction processing and multi- threading functions are provided by default. Web server embedded module Inference engines are embedded into the Web server as a sub-module using API such as NSAPI. Client-side External viewer Is developed as an independent program. Is invoked by the Web browser accord- ing to predefined MIME type. Java applet Bytecodes for inference engine is located on the server-side but is trans- mitted to a Web browser. JVM on the client-side interprets the bytecodes and executes them. W. Kim et al. / Expert Systems with Applications 28 (2005) 79–9180 we applied a generalized transformation algorithm deriving from a set of rules a set of hyperlinked HTML documents equivalent to the set of rules in terms of inference. The second potential problem arising from adopting an HTML document-based rule inference is that the mainten- ance of rules will be somewhat different from the CGI-based approach. While the CGI-based approach can use a classical rule maintenance approach applied directly to rules, our approach might eventually need to maintain directly a set of HTML documents in order to properly manage the rule base. Since we provide an automatic transformation mechanism, the FES algorithm, deriving a set of HTML documents from the rules, the classical rule maintenance approach can still be applied to our HTML document-based rule inference without any additional efforts. To address this issue, we developed a framework that can systematically maintain a set of HTML documents. This framework adopts a graphical rule representation mechanism, Expert’s Dia- gram (Lee, Lee, & Choi, 1990), by which a rule base manager, that is, a knowledge engineer, can view and edit a set of HTML documents through an easy-to-understand and graphical rule representation model. Finally, we have designed and implemented our proposed rule transformation algorithm and HTML-based rule inference system maintenance framework as an automated system called WeBIS. For the empirical validation of our approach, we also present a performance evaluation of our approach through a comparative study with CGI-based approaches. The remainder of the paper is organized as follows. Section 2 presents reviews of the current rule-based inference systems and addresses their practical limitations. In Section 3, we first present our basic idea of using the hyperlinked HTML document for rule inference and then we propose the application of a generalized rule base transformation algorithm to a set of hyperlinked HTML documents. Section 4 presents an HTML document-based rule inference system maintenance framework. System architecture and implementation issues of WeBIS are also presented. We present a performance evaluation of our proposed inference approach in comparison with the currently popular rule-inference approach, that is, the CGI-based approach, in Section 5. Finally, our conclusions are addressed in Section 6. 2. Related work We can classify existing methodologies to build Web- enabled, rule-based systems into five categories in accord- ance with applied technologies. Table 1 shows those five categories and their summarized technical features as developed in Web-enabled, rule-based systems. As shown in Table 1, we can first categorize existing methodologies into two broad categories, the server-side and the client-side depending on the location of the inference engine of a Web-enabled, rule-based system. The server-side category can be further divided into three more detailed categories, the CGI program, the server-side script, and the Web server embedded module depending on the types of inference engine implemented. On the other hand, the client-side category is further classified into two sub-categories, the external viewer and the Java applet. The technical features column of Table 1 describes the characteristic features of each category that differentiate them from each other. Concerning practicality and efficiency in developing and operating a Web-enabled, rule-based system, let us examine the pros and cons of each category mentioned above. The most popular approach to the Web-enabled, rule-based system is the CGI program. Several well-known rule-based systems such as EXSYS, Blaze Advisor, and ILOG are based on the CGI program. One of the major reasons for CGI’s popularity is its generally easy maintenance and its ability to extend rule-based systems compared with the Web-server embedded module approach or the client-side approaches. The CGI program, however, has revealed a serious weakness in its ability to respond to requests from clients because the number of clients is growing faster than most computing resources. Currently, server-side script approaches are recognized as an alternative that can partially complement this weakness of the CGI program since they allow multi-threading instead of forking and provide an automated transaction processing function to reduce the burdens on Web servers more efficiently than the less savy CGI program. Even though the server-side script approach is better than the simple CGI program approach, Fig. 1. An AND/OR graph. W. Kim et al. / Expert Systems with Applications 28 (2005) 79–91 81 the Web server burden incurred by the server-side script approach is still much bigger than the HTML document- based approach, and causes a problem of overburdening with limited computer resources. The third method belonging to the server-side approach is the Web server embedded module approach. In this case, since the Web server contains a rule-based system as a sub- module, it is easily expected that this approach will cause more difficulties and cumbersome tasks in the maintenance and extension of the rule-based system in comparison with the CGI program or server-side script approaches. We speculate that this is a major cause of the low number of commercial Web-enabled, rule-based systems that can be developed by using the Web server embedded module approach. As a sub-category of the client-side approach, the external viewer approach uses an independent rule-based inference system which is contained in the client computer. Therefore, the computational burden to Web server may be reduced on some limited level, but maintenance of the system is more difficult than the server-side approaches since each program must be installed separately from the Web server and this requires a reinstallation of the system every time it is updated. These problems prevent it from being applied to the development of a Web-enabled, rule- based system for commercial use. Another sub-category in the client-side approaches is Java applet, which solves the major problems of the external viewer approach, such as version management and consistent maintenance, with its platform independent framework. Although ILOG and e2gLight (eXpertise2Go) also provide this Java applet based approach because of these advantages, it is never- theless true that Java applet is not widely accepted yet as a practical, commercial approach to develop the Web- enabled, rule-based system in a real world industry. The Java applet approach needs bytecodes, that is, executable programs, to be downloaded only once from the Web server to the client at the time of consultation to a rule-based system, and does not create any further burdens to the Web server. Instead, it creates additional data traffic such as rule- base or parts of databases (in some cases, it may be whole databases) between server and client. But the size of the rule base and required parts of databases for inference are usually much bigger than the size of inference engine in most commercial systems; this seems to be the main reason for the lack of popularity of the Java applet approach for commercial purposes. Reducing latency time is most critical to the success of a Website (Zari, Saiedian, & Naeem, 2001), and is also significant to Web-enabled, rule-based systems. In this review of related approaches to Web-enabled, rule-based systems, we have identified reducing latency time, that is, reducing burdens caused by rule-based systems to the Web server, as a most important, but not fully resolved, issue. This means it is still worthwhile to keep making an effort to improve the efficiency of Web-enabled, rule-based systems in terms of latency time. Based on the premise that we can make a rule-based system only with hyperlinked HTML documents, and no functional differences between hyperlink-based inference system and conventional Web-enabled, rule-based systems exist, the hyperlink-based inference system is definitely much more efficient than the conventional methodologies that use the CGI program or server-side scripts in terms of latency time. If, however, we want to make use of this advantage of the HTML-based system, we must first establish the premise of our proposal by showing how a rule-based system can be implemented with a set of hyperlinked HTML documents. In addition to this, we have to secure a general-purpose mechanism to transform a set of rules into a set of functionally equivalent hyperlinked HTML documents that work in real world, practical application. If we can address both of the above prerequisites, we can take full advantage of the hyperlink-based inference systems in Web-enabled, rule-based inference tasks. Section 3 discusses our proposed approach to satisfy both of these prerequisites. 3. Hyperlink-based inference systems 3.1. Hyperlink-based Inference If we want to use a hyperlink-based inference system instead of a conventional rule-based system, first we have to prove the hyperlink-based inference system is functionally equivalent to the conventional rule-based system with any given rule base. In this paper, the definition of the rule base that we will deal with is restricted only to the backward- chaining style rule. That is, this set of rules is devised to deduct one or a group of goals. Now, let us assume, we have the following two rules to prove proposition c: Rule #1: If a And b Then c Rule #2: If d And e Then a Fig. 1 depicts these two rules by using a popular graphic rule representation scheme, AND/OR graph (Giarratano & Riley, 1994; Nilson, 1980; Rich, 1983). If we apply these two rules to a typical rule-based system with the purpose of proving the proposition c, it will ask a user Fig. 2. An example of a functionally equivalent set of hyperlinked HTML documents. Fig. 3. An example FES for a traffic control rule set. W. Kim et al. / Expert Systems with Applications 28 (2005) 79–9182 about d, e, and b consecutively under the assumption that the applied backtracking method follows the left-to-right prin- ciple. If the results from these three questions are all true, then c is proved to be true. Otherwise, if any one of the answers is false, the rule-based system cannot accurately conclude anything. Related to the two above rules, we can posit the same series of questions and conclusions as typical rule-based systems do, by using a set of hyperlinked HTML documents. The set of hyperlinked HTML documents for our example is illustrated in Fig. 2, and the readers can easily identify that these hyperlinked HTML documents can do exactly the same inference tasks as the rule-based systems. Let us define such a functionally equivalent set of hyperlinked HTML documents for a given set of rules in terms of inference as Functionally Equivalent hyperlinked HTML document Set (FES) denoted by FES(x) for a given set of rules x. So far, we have only addressed the rules having propositional symbols in their rule sentences. This type of rule is usually called a ‘Fact Type’ rule. However, there is another type of rule which is frequently and generally used in rule-based systems that is different from the fact type rule in that it utilizes variables in representing sentences. If we introduce variables into rule representation, we have to consider a set of available values, that is, domains of the variables. The domain for the variable can usually be divided into two types, the symbolic type and the numeric type. For the cases of the symbolic type domain, the number of available values is usually finite, and so we are supposed to ask users to select one or several of the available values in the domain. Let us assume, we have the following three rules related to traffic control: Rule #3: If the color of the traffic signalZgreen Then go Rule #4: If the color of the traffic signalZyellow Then slow down and prepare to stop Rule #5: If the color of the traffic signalZred Then stop In this case, the variable is ‘the color of the traffic signal’ and its domain is {green, yellow, red}. We can also easily construct FES({Rule #3, Rule #4, Rule #5}) for these three rules as depicted in Fig. 3. In the case of the variable that has a numeric type domain, however, the construction of a corresponding FES( ) is not so simple because we must first obtain the exact numerical value via hypertext from a user, and this is practically impossible to obtain. Of course, in the case of only one set of a finite number of ranges in value exclusive to each other, we may construct FES( ) in a way similar to the symbolic type domain by regarding each range as a separate hyperlink. But, in a general sense, since we may need to compute an expression based on the value of the variable and to compare it with a constant or other variables, we can say that the hyperlink on its own is not complex enough to represent the rules containing numeric type of variables. One simple way to cope with this limitation of the hyperlink is to use a client-side script language such as JavaScript and VBScript to obtain the values of the variables and to compute and compare them if needed. Suppose the following two rules related to a tax consultation: Rule #6: If total incomeR0.2 * threshold Then Have to pay tax. Rule #7: If total income !0.2 * threshold Then Do not need to pay tax. From these two rules, we can identify two numeric type variables of total income and threshold. We also find that numerical computations and comparisons must be performed for inference. Using JavaScript, we can construct FES({Rule #6, Rule #7}) with three hyperlinked HTML documents: one JavaScript-embedded document for the premise part and two simple HTML documents for the conclusions. Fig. 4 illustrates the required source of the first JavaScript embedded document. This source is then divided into two parts, the Form and the Script as shown. The Form obtains the values of the numerical variables in rules from users. The Script performs computations and comparisons of the obtained values Fig. 4. Source of a JavaScript embedded HTML page for numeric type variables. Fig. 5. Example screen shots generated W. Kim et al. / Expert Systems with Applications 28 (2005) 79–91 83 required in the premise of the rule and then hyperlinks to the corresponding page for a conclusion depending on the results from the comparisons. The Script is written in boldface to distinguish it from the Form in Fig. 4. Fig. 5 illustrates example screen shots generated by FES({Rule #6, Rule #7}). Of course, which of the two screens will be activated appears on the lower side of Fig. 5 and depends on the computation of the user’s input performed in the upper part of the screen. So far, we have shown with several examples that typical but various types of rules can be transformed into functionally equivalent hyperlinked documents, that is, FES( )s. But, as of yet in spite of such examples, we are not sure if we can find FES( ) functionally equivalent to any arbitrary set of rules. To make sure that FES( ) exists for any arbitrary set of rules and can also be found, we propose an FES generation algorithm for an arbitrary set of rules. This algorithm is addressed in Section 3.2. 3.2. FES generation algorithm Basically, our FES generation algorithm takes an arbitrary set of rules and transforms it into a functionally equivalent set of hyperlinked documents. To achieve this, the FES generation algorithm works in two major phases. In the first phase, we convert each rule of an arbitrary rule set into its corresponding canonical form. Based on from FES ({Rule #6, Rule #7}). W. Kim et al. / Expert Systems with Applications 28 (2005) 79–9184 the converted rules in the canonical form, the second phase generates, in terms of inference, a functionally equivalent set of hyperlinked documents (FES). In the first phase, we define the required canonical form of rules in our context and name this required canonical form the simplified normal form (SNF). If a rule is in SNF, it must satisfy three conditions. First, the rule should not have any chaining relationship with any other rules in the rule set. In this case, the existence of a chaining relationship of a rule to other rules implies the rule has at least one element of the premise part which also appears in the consequent part of the other rules and vice versa. The second condition for SNF is that a rule can only have conjunctions in the premise part. In the third condition, each rule can have only one consequent proposition. To convert arbitrary rules into SNF, we provide a simple three step converting procedure in the figure below with some examples of each step. 3.2.1. Conversion procedure into SNF Step (1) Eliminate all chaining variables by merging related rules. Example: {QnR0C,PoC0S}/{Po (QnR)0S} Step (2) Convert the premise part of each rule into a disjunctive normal form by the associative law. Example: {Po(QnR)0S}/{(PoQ)n (PoR)0S} Step (3) Eliminate disjunctions of the premise part of each rule by splitting it into multiple rules with same consequent. Example: {(PoQ)n(PoR)0S}/{PoQ0 S,PoR0S} After converting the rules into SNF, we can proceed to the second phase where FES( ) will be generated based on the rules in the SNF. For this second phase, we developed an algorithm which transforms a set of SNF rules into FES( ) and we call it FES generation algorithm. Before going into the details of FES generation algorithm, let us look into the required definitions and notations first. Let R be a set of rules and r be a rule, which belongs to R. And each r has a premise part, pr and a consequent part, cs. pr consists of a set of atomic propositions while cs has a single consequent. p and c denote atomic propositions contained in pr and cs, respectively. Related to describing hyperlinked documents formally, let us adopt conventional notations from graph theory. T means a tree of hyperlinked documents and consists of two sets, V and E. V is a finite set of vertices, that is, HTML documents in our context, and E is a set of edges between vertices, that is, hyperlinks in our context. v and e denote, respectively, a vertex in V and an edge in E. Especially in our case, v corresponds to an HTML document and we can further represent the corresponding HTML document content of vk with content(vk). An edge e can be naturally represented as a pair of vertices, (vk, vl) where vk and vl are kth and lth arbitrary vertices in V, and this means that the edge connects the two vertices. Furthermore, each edge has a direction and the left-hand vertex in the pair is a parent vertex, while the right-hand vertex is a child vertex because the generated graph is a kind of tree. Therefore, in the above case, vk is a parent and vl is a child. In terms of the edge e, Let us denote the parent and child vertices by pv(e) and cv(e), respectively. In addition to this, since each vertex will eventually be matched with an atomic proposition in a rule, each edge also represents one of the Boolean constants, a true or false value of the corresponding atomic proposition. The represented Boolean value of an edge e is denoted by value(e). Up to this point, we have addressed the notations about rules and hyperlinked documents themselves, but more definition is needed in order to successfully transform rules into FES( ). We formally define a term as fact when an atomic proposition has an arbitrary Boolean value and it is denoted by f. That is, f is represented by (p, value(p)), which means the atomic proposition p has value(p) regardless of whether its value is true or false. To access p and value(p) in f, we define two functions, prop(f) and propval(f) where they designate p and value(p), respectively. Let us make one more definition, preceding fact set (PFS). PFS is defined from the point of view of the vertex, and it designates facts accumulated along the path from the root to a specific vertex, vk. As is commonly understood, a path consists of a set of edges; and each edge implies whether the related parent vertex, that is the proposition of that vertex, is true or false. Therefore, for a given vertex, we can have a series of related facts along the path to that vertex. The set of those facts to a given vertex vk is denoted by PFS(vk). To define PFS in a more formal manner, let path(vk) be a set of vertices laid along the path from the root to the vertex vk, {v1,vi1,vi2,.,vim,vk}. Then, PFS ðvkÞ Z fa set of f Z ðcontent ðvlÞ; value ðeÞÞ s:t: vl 2path ðvkÞ; pv ðeÞ Z vl; cv ðeÞ 2path ðvkÞ; and l skg: Based on the concepts and definitions addressed so far, we will now develop the second phase algorithm mentioned earlier in this subsection. The algorithm ‘FES generation algorithm’ consists of four routines (functions) including the main routine. Fig. 6 shows the main routine, CONVERTRU- LEBASEINTOFES and the remaining three subroutines, CONVERTRULEINTOTREE, MATCH, and EXIST, which are described in Figs. 7–9, respectively. It is easy to understand this FES generation algorithm with an example which shows how the algorithm proceeds. Let the following two rules in SNF be a rule set needed to be transformed into FES. Fig. 6. CONVERTRULEBASEINTOFES algorithm. Fig. 7. CONVERTRULEINT W. Kim et al. / Expert Systems with Applications 28 (2005) 79–91 85 R1 : Ao Bo C 0 X R2 : Bo D0 Y Fig. 10 illustrates how FES generation proceeds to generate hyperlinked HTML documents for the example rules at a step by step level. The following seven steps describe each step shown in Fig. 10. (1) OTRE In the first step, the main routine CONVERTRULEBASEIN- TOFES is invoked with a parameter RZ{R1, R2}. Then the main routine invokes the subroutine CONVERTRU- LEINTOTREE for each rule in R (ZR1 and R2) while passing the IncompleteVertice, a rule r, and a tree T. E algorithm. Fig. 8. MATCH algorithm. Fig. 10. Progress of FES generation algorithm with the example rules. W. Kim et al. / Expert Systems with Applications 28 (2005) 79–9186 At this moment, IncompleteVertice and T are empty and r is R1 in our example. Then the subroutine CONVER- TRULEINTOTREE transforms a given rule into a tree of HTML documents by creating and appending vertices and edges. This subroutine can be divided into two parts, the initialization part and the tree growing part. In the case of R1, it performs the initialization part which will process every atomic proposition in the premise part, and then it finishes by updating the content of the leaf node with a consequent proposition. During processing each proposition of the premise part, the tree, T will grow. As a result of the first iteration of this processing, Step 1 in Fig. 10 shows the generated tree based on the first proposition, A of R1 in our example. (2) By processing the second proposition, B of R1, we have the tree as depicted in Step 2 in Fig. 10. The white box node means it is expected to become a member of the IncompleteVertice in the next stage in order to process the remaining rules. (3) In the third step of Fig. 10, the algorithm processes the final proposition, C of R1 and updates the true side node with the consequent of the rule, X. The conclusion nodes are denoted by circle nodes here. t and f attached on the arcs means that the proposition of the parent node of that arc is either true or false. In this step, the transformation of rule R1 into the tree is finished and so, the execution of the subroutine Convertruleintotree for the rule R1 is also ended by returning to the main routine. (4) In the fourth step, the algorithm invokes the subroutine CONVERTRULEINTOTREE again for the rule, R2 together with the IncompleteVertice, which consists of the Fig. 9. EXIST algorithm. vertices marked with a white box and the generated tree, T that was the result in (3). In this case, CONVERTRU- LEINTOTREE applies the tree growing part procedure which is located at the outermost else part to the rule R2. The first vertex of the IncompleteVertice is the false side child node of A and the algorithm first applies MATCH with {B, D} as pr to this vertex. Since there is no evidence that B or D is false along the path, MATCH is evaluated as true and this leads to the else part where the tree is growing while considering the existence of propositions in the path. In our example, B is first considered and is appended to the tree and its result is shown in Step 4 of Fig. 10. (5) Proposition D of rule R2 is checked against MATCH and EXIST and then is added into the tree as shown in Step 5 of Fig. 10. At the same time, since all propositions in the premise part of R2 are processed, the consequent Y is updated on the true side leaf node. (6) In the sixth step, the false side vertex of B in the upper side of the tree is examined, but it does not pass the MATCH test, so the vertex still remains as the IncompleteVertice. Now, the false side vertex of C in Step 3 is examined and the rule satisfies the MATCH condition. Therefore, B and D are examined against EXIST and only D satisfies the false condition. As a result, D is added to the tree and finally, its consequent part Y is also updated on the true side of D. (7) In the final step, the algorithm updates all remaining vertices in the IncompleteVertice by ‘No Conclusion’ which is denoted by a question mark, ‘?’, because all the rules were processed in R. The completed result of the hyperlinked HTML documents, that is, the FES correspondence to the example rule set is depicted in Step 7 of Fig. 10. W. Kim et al. / Expert Systems with Applications 28 (2005) 79–91 87 So far, we have discussed our proposed FES generation algorithm. In Section 3.3, we will address the minor extension of the algorithm and show a brief real world example based on a bank loan approval. 3.3. Algorithm extension and its real word example The algorithm proposed in Section 3.2 can only cover the so called ‘Fact Type’ rule mentioned earlier. We already explained that there are two additional popular rule types, rules with symbolic variables and numerical variables. But with minor revision on the FES generation algorithm, we can easily take into account those types of rules. We are not going into the details of revision here since they are so trivial. Fig. 11 shows a part of the real world example rules for a bank loan approval process. These example rules include all three types of rules, so this real world example will show how the example rules are transformed into a set of hyperlinked HTML documents by using our extended algorithm. We applied the extended FES generation algorithm to the example rules in Fig. 11 that creates a set of hyperlinked HTML documents in Fig. 12. As you can see in this figure, the condition which has a symbolic variable is reflected in the root HTML document and the condition having a numerical variable is reflected in the HTML document Fig. 11. Example rules for a titled, “Current Amount of Monthly Debt”. Especially in the latter document, users are required to enter the exact value of their current amount of monthly debt and the corresponding JavaScript code will decide which condition is satisfied by the user’s input and then lead to the corresponding HTML document. 4. WeBIS approach 4.1. Architecture of WeBIS So far, we have proposed the idea that the hyperlinked HTML document can perform inference tasks, and we have shown how a rule base can be automatically converted into a FES. To facilitate our approach, we provide a system architecture, Web Based Inference System (WeBIS), which can support users to develop a set of hyperlinked HTML documents to perform an inference task systematically. Fig. 13 shows the overall architecture of WeBIS and its three major components. Major information flows are also depicted. As shown in Fig. 13, the knowledge engineer can input rules to the system via both the general text editor and the graphical rule editor provided by WeBIS. Then, WeBIS transforms rules into functionally equivalent set of hyperlinked bank loan approval. Fig. 12. Generated hyperlinked HTML documents for bank loan example rules. W. Kim et al. / Expert Systems with Applications 28 (2005) 79–9188 HTML documents. Those documents are registered to Web servers and finally users can access them and get inference services via the Internet. Now Let us briefly introduce each of the components, their roles, and related architectural issues. Fig. 13. Architectur (1) e of Graphical Rule Editor (Expert’s Diagram Approach). The graphical rule editor supports complete rule editing process by specifying Expert’s Diagram in a GUI environment. Lee et al. (1990) claimed that expressing rules in a graphical manner is more efficient in most cases WeBIS. W. Kim et al. / Expert Systems with Applications 28 (2005) 79–91 89 especially when a knowledge engineer is not skilled in knowledge engineering, and they proposed a graphical rule representation scheme, Expert’s Diagram. We adopt this rule representation scheme as default in our system and develop a rule editor which supports user to express rules in Expert’s Diagram format. But skilled knowledge engineer can still input rules directly by using any general text editors. Fig. 13 also depicts such direct rule input process. Whether the rules are edited by a graphical rule editor or general text editor, they are fed into an extended FES generation module. (2) Extended FES Generation Module. The extended FES generation module implements the algorithm addressed in Section 3. It is responsible for transforming rules into a functionally equivalent set of hyperlinked HTML documents. After finishing the transformation, it invokes the page manager by passing the generated HTML document set. (3) Page Manager. The page manager has two major roles: (1) to finalize completion of each generated HTML document by applying a corresponding template to it, and (2) to manage consistency between Expert’s Diagram generated by a graphical rule editor and a generated set of hyperlinked HTML documents. The additional benefit of using Expert’s Diagram is that the user can easily manage consistency by cooperation between the page manager and the graphical rule editor when there are minor changes in rule content. 4.2. Implementation We have incorporated the framework of our Web-based inference approach into a working prototype written in Fig. 14. An illustrative scr Visual CCC on the Windows environment. Fig. 14 shows an illustrative screen where the user creates a rule base and hyperlinked HTML documents, transforms the rule base into equivalent hyperlinked HTML documents, FES back and forth, edits FES using a graphical interface, and generates styled HTML documents using templates. As shown in Fig. 14, the user can choose a rule base file on the left frame and then convert it into FES as displayed graphically on the right upper frame. The user can also create directly hyperlinked HTML documents based on Expert’s Diagram technology. Finally the user can create the templates to be applied to the HTML documents, then the system automatically generates and displays completed HTML documents after applying the templates. The right lower frame in Fig. 14 shows a template-applied HTML document for the selected node (marked with a red border) on the right upper frame. At the time of finishing all these jobs, you can publish the HTML documents to the Web server as a hyperlink-based inference service. 5. Performance evaluation To prove the efficiency of our approach, we compare the throughput of a hyperlink-based inference system with that of inference systems which were developed by using conventional commercial rule-based systems. To evaluate the throughput of each system, we count the number of HTTP (Fielding et al., 1999) responses per 3 min from the Web server when 10 client computers send HTTP requests simultaneously via the Web. The hardware specification of the server computer is Pentium 4 CPU of 1.5 GHz with 256 MB of main memory. The OS of the server is the Microsoft Windows 2000 Server. een shot of WeBIS. Table 4 Performance evaluation results (10 clients) Client Hyper- link A a Ratio b B a Ratio c PC1 7471 34 219.74 932 8.02 PC2 7568 34 222.59 931 8.13 PC3 7496 33 227.15 900 8.33 PC4 7563 33 229.17 921 8.21 PC5 7534 32 235.44 895 8.42 PC6 7435 35 212.43 893 8.33 PC7 7486 34 220.16 913 8.20 PC8 7189 33 217.83 930 7.73 PC9 7681 34 225.90 947 8.11 PC10 7755 33 235.00 919 8.44 Sum 75,176 335 224.41 9181 8.19 Average 7518 34 918 Unit: #responses per 3 min. a Conventional rule-based systems A and B. b Ratio between Hyperlink and A. c Ratio between Hyperlink and B. Table 2 Performance evaluation results (one client) Client Hyper- link A a Ratio b B a Ratio c PC1 30,411 307 99.06 5832 5.21 Sum 30,411 307 99.06 5832 5.21 Average 30,411 307 5832 Unit: #responses per 3 min. a Conventional rule-based systems A and B. b Ratio between Hyperlink and A. c Ratio between Hyperlink and B. W. Kim et al. / Expert Systems with Applications 28 (2005) 79–9190 The hardware specification of all of the client computers is Celeron CPU of 633 MHz with 128 MB of main memory. The OS of the client is Microsoft Windows 98. Because the commercial rule-based systems deploy their application systems using Java Server Pages (JSP), we use Tomcat as the Web server. The results appear in Tables 2–4, and Fig. 15. Table 2 shows the results of performance evaluation when a single PC is used as the client. The second column denoted as Hyperlink shows the throughput of our hyperlink-based inference approach. The third column denoted as A shows the throughput of a commercial rule- based system, and the fourth column shows the ratio between the throughput of our approach (Hyperlink) and the throughput of the rule-based system A. The fifth column denoted as B shows the throughput of another commercial rule-based system, and the sixth column shows the ratio between the throughput of Hyperlink and the throughput of the rule-based system B. When the number of clients is one, the total numbers of HTTP responses per 3 min for Hyperlink, A, and B are 30,411, 307, and 5832, respectively. The ratio between Hyperlink and A is 99.06 and the ratio between Hyperlink and B is 5.21. The result means that even in the case of serving a single client our approach is much faster than the two competitive commercial tools. Table 3 shows the results produced when five PCs are used as the clients. In this case, the total numbers of HTTP Table 3 Performance evaluation results (five clients) Client Hyper- link A a Ratio b B a Ratio c PC1 14,829 69 214.91 1997 7.43 PC2 15,092 70 215.59 1995 7.56 PC3 14,886 70 212.65 1845 8.07 PC4 14,950 72 207.64 1897 7.88 PC5 15,030 71 211.68 1849 8.13 Sum 74,785 352 212.46 9583 7.80 Average 14,957 70 1917 Unit: #responses per 3 min. a Conventional rule-based systems A and B. b Ratio between Hyperlink and A. c Ratio between Hyperlink and B. responses are 74,785, 352, and 9583, respectively, and the average numbers of HTTP responses are 14,957, 70, and 1917, respectively. The average performance ratios of Hyperlink to A and B are 212.46 and 7.80, respectively. In terms of ratio, the readers can easily see that the out- performing level is higher than in the case of serving a single client only. The results produced when there are ten PC clients are shown in Table 4. In this case, the total numbers of HTTP responses are 75,176, 335, and 9181, respectively, and the average numbers of HTTP responses are 7,518, 34, and 918, respectively. The average ratios are 224.41 and 8.19, respectively, and the performance gaps between our approach and commercial systems keep growing. Fig. 15 summarizes the transition of the total numbers of HTTP responses per 3 min for 1-client, 5-client, and 10-client cases. We can expect here that the total number of HTTP responses will be stabilized as the number of clients is increased, and at the same time, the high performance of our approach over two competitive commercial rule-based systems will persist as well. In summary, we can conclude that there are some fixed Fig. 15. Performance evaluation results. W. Kim et al. / Expert Systems with Applications 28 (2005) 79–91 91 and significant gaps between the performance of Hyper- link and those of A and B, and the results of our performance evaluation predicts that the ratio will be approximately 224.41 and 8.19. 6. Conclusions We have proposed a hyperlink-based inference approach to alleviate the over-burdening problem that troubles most of the conventional Web-enabled, rule-based systems including commercial tools. Since this problem is nearly caused by the CGI-based approach, we have found a way to perform equivalent inference tasks based on hyperlinks between HTML documents as an alternative to the CGI- based approach. To demonstrate our idea, we first show that a hyperlink- based inference can perform an equivalent inference task to conventional Web-enabled, rule-based systems. Secondly, to facilitate our approach, we proposed an FES generation algorithm which automatically transforms the rule base in the conventional text format into the hyperlinked HTML documents, FES, and proved its completeness. Furthermore, to support the user to create directly hyperlinked HTML documents for inference, we also develop an ease-to-use graphical knowledge acquisition function to build hyper- linked HTML documents systematically based on the Expert’s Diagram. Finally, we design and implement a WeBIS which incorporates all of the functions and mechanisms mentioned above. Using this WeBIS, we have performed an empirical experiment and performance comparison between our approach and two major commercial Web-enabled, rule- based systems. Through this experiment, our approach has proven to outperform both of the commercial rule-based systems in terms of throughput. In addition to this, WeBIS was successfully applied to a real world example, the financial consultation done via the Website of a leading financial consulting company in Korea and it is still running at a very satisfactory level. We expect the best performance of our approach when the number of service requests for rule inference is very large. References eXpertise2Go.com, [Online]. Available: http://www.expertise2go.com/ EXSYS Inc., [Online]. Available: http://www.exsys.com/ Fair Isaac Corporation, [Online]. Available: http://www.blazesoft.com/ Fielding, R., Gettys J., Mogul J., Frystyk H., Masinter L., Leach P., Berners-Lee T. Hypertext Transfer Protocol—HTTP/1.1, RFC2616, IETF, 1999. [Online]. Available: ftp://ftp.isi.edu/in-notes/rfc2616.txt. Giarratano, J., & Riley, G. (1994). Expert systems: principles and programming (2nd ed.). PWS Publishing Company. ILOG Inc., [Online]. Available: http://www.ilog.com/ Lee, J. K., Lee, I. K., & Choi, H. R. (1990). Automatic rule generation by the transformation of Expert’s Diagram: LIFT. International Journal of Man–Machine Studies, 32, 275–292. McCarthy, J. C. (1999). The social impact of electronic commerce. IEEE Communications Magazine, 37(9), 53–57. Nilson, N. J. (1980). Principles of artificial intelligence. Berlin: Springer. Rich, E. (1983). Artificial intelligence. New York: McGraw-Hill. Song, Y. U., & Lee, J. K. (2000). Automatic generation of web-based expert systems. Journal of Intelligent Information Systems, 6(1), 1–16 [in Korean]. Zari, M., Saiedian, H., & Naeem, M. (2001). Understanding and reducing web delay. Computer, 34(12), 30–37. http://www.expertise2go.com/ http://www.exsys.com/ http://www.blazesoft.com/ http://ftp://ftp.isi.edu/in-notes/rfc2616.txt http://www.ilog.com/ Web enabled expert systems using hyperlink-based inference Introduction Related work Hyperlink-based inference systems Hyperlink-based Inference FES generation algorithm Algorithm extension and its real word example WeBIS approach Architecture of WeBIS Implementation Performance evaluation Conclusions References 
work_nxdy636kujcvvj3ogvkz3iw7ca	----	Microsoft Word - Zhang et al.pdf African Journal of Biotechnology Vol. 10(32), pp. 6162-6171, 4 July, 2011 Available online at http://www.academicjournals.org/AJB DOI: 10.5897/AJB10.2086 ISSN 1684–5315 © 2011 Academic Journals Full Length Research Paper Comparison of ν-support vector regression and logistic equation for descriptive modeling of Lactobacillus plantarum growth Lijing Zhang 1,3 , Zhijian Song 2,3 , Xiaodong Pan 3 , Mingguang Feng 1 and Zhihua Jin 3 * 1 Institute of Microbiology, Zhejiang University, Hangzhou, 310058, China. 2 Institute of Biomedical Informatics/Zhejiang Provincial Key Laboratory of Medical Genetics, Wenzhou Medical College, Wenzhou 325000, China. 3 Ningbo Institute of Technology, Zhejiang University, Ningbo 315104, China. Accepted 15 April, 2011 Due to the complexity and high non-linearity of bioprocess, most simple mathematical models fail to describe the exact behavior of biochemistry systems. As a novel type of learning method, support vector regression (SVR) owns the powerful capability to characterize problems via small sample, non- linearity, high dimension and local minima. In this paper, we developed a ν-SVR model with genetic algorithms (GA) in the pre-estimate in Lactobacillus plantarum fermentation by comparing the predicting capability of logistic model and SVR model. 5-fold cross validation technique was applied in the SVR train to avoid over-fitting. The information of SVR parameters were obtained in the generation of 150 and the optimal parameters were C= 235.8935, σ= 8.3608, ν=0.7587. Correspondingly, the logistic model parameters maxµ and maxx were estimated as 0.4791 and 0.3498, respectively. The experimental results demonstrated that, SVR model excelled the logistic model based on the normalized mean square error (NMSE), mean absolute percentage error (MAPE) and the Pearson correlation coefficient R. We found that the ν-SVR model optimized by genetic algorithms could be a potential monitoring method for prediction of biomass. Key words: Support vector regression, genetic algorithm, logistic model, prediction of biomass. INTRODUCTION Because microorganism fermentation process is complex and non-linear, many parameters such as the concentration of bacteria, substrates and products are difficult to measure on-line (Wang et al., 2006). In order to optimize bioprocess and put the advanced algorithms of control into practice, it is necessary to monitor and diagnose the bioprocess parameters. Kinetic models are generally experimentally derived mathematical formulas that can agree well with the cultivation data (Bessaou and Siarry, 2001). Kinetic models can be linear or nonlinear, such as ordinary (or * Corresponding author. E-mail: zlj@nit.zju.edu.cn. Tel: +86 574 88229516. Fax: +86 574 88229516. partial) differential equations, cellular automata (Xiao et al., 2005; Xiao et al., 2006) and logistic model. The logistic model firstly, proposed by Verhulst in the eighteenth century(Liu et al., 2002), has been demonstrated to be the most illustrative model of microbial growth dynamics in a habitat of finite resources (Fujikawa et al., 2002; Peleg et al., 2007; Schepers et al., 2000). Logistic equation is based on the notion that the momentary growth rate of a given population inoculated into a virgin environment is proportional to the momentary population’s size. Recently, a statistical learning theory based on formalism known as the support vector machines (SVMs) emerged as a novel powerful tool for data classification and regression, which are called support vector classification (SVC) and support vector regression (SVR). SVMs could solve the problems resulted from small samples, nonlinearity, high dimension and local minimum (Vapnik, 1995; Borin et al., 2006; Cai et al., 2003; Moreira et al., 2007). Therefore, SVR is gradually accepted in the data-driven nonlinear modeling applications (Desai et al., 2006). Unfortunately, for the generalization performance of SVR which depends on SVR parameters, how to set these parameters by a given set of data is the main issue for the application of SVR. Cherkassky (2004) described ε-support vector regression (ε-SVR) for setting meta- parameters. However, as the parameter ε influences the solution indirectly and lacks intuitionistic means, it is difficult to give appropriate value for ε. To overcome this problem, a new parameter ν was introduced to control the fitness and predication accuracy instead of parameter C in SVR (Schepers et al., 2000). Chalimourda et al. (2004) studied optimal ν in SVR for different noise models using the data from sine function and Boston housing problem. Recently, the performance of SVR in bioprocess was also researched (Li et al., 2006a; Li and Yuan, 2006b). However, ν-SVR coupled with genetic algorithm has not been investigated in fermentation process. In this study, ν-SVR was applied to determine the concentration of cell via the data from fermentation process. We aimed to find a suitable method for the simulation of microbial growth by comparing different modeling strategies in process simulation and fermentation control. The organism used in the model was Lactobacillus plantarum. MODEL STRUCTURE Logistic model As unstructured kinetic model, logistic model is the most frequently employed for modeling microbial systems, because it is simple and good enough for technical purposes ((Kiviharju et al., 2006; Mu et al., 2006). The logistic model can describe the kinetics of the microbial growth as follows: m 1 max ax dx x x dt x µ   = −    (1) Where, x (g l −1 ) is the cell concentration, t (h) is the fermentation time, maxµ (h -1 ) is the maximum specific growth rate and m axx (g l −1 ) is the maximum cell concentration. Using x = 0 x (t = 0), integration of equation 1 gives the following equation for microbial biomass: Zhang et al. 6163 m 0 m m 0 0 axm ax ax ax t t x x e x x x x e µ µ = − + (2) Support vector regression Support vector regression The ν-support vector regression (ν-SVR) is a promising method for solving non-linear regression problem, which depends on the theory of support vector machines (SVMs). In this section, a brief introduction of SVR is given. Given a set of data points: l i i i=1 T={ (x , y )} N R R∈ × , where m i x R∈ is the input vector, iy R∈ is the desired value and l is the total number of data points. SVR estimation aims to seek for a function using the following formula: ( ) T i y w x b= ⋅ Φ + (3) Where, w is the weight, b is the coefficient, and ( )ixΦ denotes the data in high dimensional feature space mapped by applying nonlinear kernel function from the input space i x . The coefficients (both w and b) could be optimized only by minimizing the regularized risk function: 2 1 1 1 ( ) 2 l i ik i w C l ζ ζ ∗ = + +∑ (4) Where, 21 2 k w is called the regularized term. 1 1 ( ) l i i il ζ ζ ∗ = +∑ is the empirical risk measured by the ε − insensitive loss function. The slack variable i ζ , is called the upper training error and iζ ∗ represents the lower training error subject to the " ε -radius-tube” where (( ( ) ) i i y x bω ε− ⋅ Φ + ≤ . C is the regularization factor which has the ability of determining the trade-off between the empirical risk and the regularized term. As a consequence, the ε-SVR problem can be 6164 Afr. J. Biotechnol. described as the following quadratic optimization problem: 2 , , 1 1 1 min ( ) 2 l i ikb i w C lω ζ ζ ζ ∗ =   + ⋅ +    ∑ (5) (( ( ) ) , (( ( ) ) , , 0, 1, 2, 3 i i i i i i i i subject to x b y y x b i l ω ε ζ ω ε ζ ζ ζ ∗ ∗ ⋅ Φ + − ≤ + − ⋅ Φ + ≤ + ≥ = L However, the varied value of the ε could be selected from all values in real number and would deeply influence the solution of quadratic optimization problem. Fortunately, Schölkopf et al. (2000) cited a weight parameter (ν) multiplied by ε in the ε-SVR problem to control the tube size, which avoids exhausting search of the ε parameter in the real number range. After this transition, the optimization problem is as follows: 2 , , 1 1 1 m in ( ) 2 l i ikb i w C lω ζ υ ε ζ ζ ∗ =    + ⋅ + +      ∑ (6) (( ( ) ) , (( ( ) ) , , 0, 1, 2, 3 i i i i i i i i su b jec t to x b y y x b i l ω ε ζ ω ε ζ ζ ζ ∗ ∗ ⋅ Φ + − ≤ + − ⋅ Φ + ≤ + ≥ = L Usually, the problem stated earlier could be successfully solved in its dual form by exploiting the Lagrange multipliers approach (Vapnik, 1995; Belousov et al., 2004). Through the Karush-Kuhn-Tucker (KKT) conditions of solving quadratic programming problem, the dual Equation 5 leads to the solution by maximizing the , i i α α∗ , which are Lagrange multipliers . * , , 1 1 1 max ( )( ) ( , ) ( ) 2 l l i i j j i j i i i i j i K x x y α α α α α α α α∗ ∗ ∗ = = − − + −∑ ∑ (7) 1 1 ( ) 0 0 , , 1, 2, 3 ; ( ) l i i i i i l i i i su bject to C i l l C α α α α α α υ ∗ = ∗ ∗ = − = ≤ ≤ ∀ = + ≤ ⋅ ∑ ∑ L They satisfy the condition of 0, 0, 0 i i i i xα α α∗ ∗× = ≥ ≥ . Where 1, 2...i l= . Parameters determination of the ν-SVR model To construct an efficient ν-SVR model, SVR’s parameters are crucial and should be set carefully (Fujikawa et al., 2004). These parameters include: Kernel function: The term ( , )i jk x x in Equation 7 is defined as kernel function. Generally speaking, as the key in the SVM theory, kernel function is used to map the input feature data ix into a higher-dimensional feature space via non-linear mapping (Bessaou and Siarry, 2001). Three typical examples of kernel functions are polynomial, radial basis function (RBF) and linear kernel. As a kind of RBF kernel functions, Gaussian function yielding better prediction performance was used in this study (Samanta et al., 2003). Penalty parameter C: As mentioned earlier, C determines the trade off between margin maximization and error minimization. Bandwidth of the kernel function ( 2σ ): In this paper, sigma square ( 2σ ) denotes the variance of the Gaussian kernel function. Parameter ν: ν represents an upper bound on the fraction of errors and a lower bound on the fraction of parameters to build the regression function and the range is from 0 to 1. In general, the accurate estimation of ν-SVR model is based on the hyper-parameters (C, σ , ν), these parameters need to be selected in ν-SVR. But now it is still a complex problem and hard to obtain the optimized parameters value in an effective algorithm. Heuristic search techniques are usually used to find the optimal SVR parameter settings, such as genetic algorithm (GA) or simulation annealing algorithm (SA). Thereinto, GA is better for robust optimization in a complex search space (Potocnik and Grabec, 1999; Ustun et al., 2005). In this study, the coupled-model of GA-SVR was proposed to find the optimal values of ν-SVR parameters. GA-SVR optimization procedure Genetic algorithm (GA), which was formerly introduced by Holland (1975), employs the biological techniques of Figure 1. GA-ν-SVR model. mutation and crossover to search the local optimal solutions. Figure 1 illustrates the algorithm of the GA- SVR model. Our proposed GA- SVR is described as follows in detail: Step 1. Population initialization: In this process, the three ν-SVR parameters (C, σ , ν) were converted decimal code into a binary format. The chromosome X was represented as X = p1, p2, p3, where p1, p2 and p3 denote the regularization parameters C, σ and ν, respectively. One population contains n sizes of such kind chromosomes and is randomly generated. In this study, we used 100 sizes of such chromosomes and when the chromosomes were put into SVR to train, the binary format was translated into decimal once again. Step 2. Evaluate the fitness function: The fitness of training data set is easy to be calculated, but also, it is prone to be over-fitting. The 5-fold cross validation technique was applied in the SVR train to avoid over- fitting and to make the result generalized and reasonable (Holland, 1975; Browne, 2000). The fitness function of the parameter set is measured as the criteria of minimizing the MAPE (mean absolute percentage error): Zhang et al. 6165 MAPE= 1 100% / × n i i ii a p a n = −∑ Where, ip denotes the predicted value in SVR model, i a represents the actual experimental observation and n is the number of experimental observations. The smaller the values of MAPE in the regression, the better the prediction performance would be. Step 3. Genetic operation: The genetic operators (Figure 2) which contain selection, crossover and mutation are mainly illustrated as follows: Selection: A spin of a weighted roulette wheel is applied to choose chromosomes in the current population with the possibility of selection frequency rate. Crossover: Crossover is a typical genetic operation in genetic algorithm which allows new population to be created in the search space. This mechanism operation aims to create a crossover point randomly in a pair of parental chromosomes and then, exchange genes between these two chromosomes. This situation occurs between two chromosomes and the crossover rate could be described with the frequency of crossover. Mutation: Mutation process is the same genetic operator of crossover which also alters the solution of parental chromosome. However, it happened in one chromosome and the genes may occasionally be altered. In binary code, one mutation means genes changing code from 0 to 1 at one random point. After these operators, if the minimum fitness value of the new population is smaller than that of the old population, the offspring chromosome could replace the old chromosome and create a new population. Step 4. Termination criteria: The process was repeated from the genetic operation and fitness function evaluation until the number of generations was up to termination. After that, the best chromosomes would be presented as a solution. Of course, the information of parameters in the chromosomes could be acquired and the best MAPE could be represented. In this study, we defined 150 generation. MATERIALS AND METHODS Strain L. plantarum C263, originally isolated from diary food were 6166 Afr. J. Biotechnol. Figure 2. Genetic operation. rendered by China Center of Industrial Culture Collection (CCICC) and cultured in MRS broth (Difco). Culture stocks were inoculated into MRS medium at 5% and incubated for 24 h at 37°C. This culture was propagated twice under the same conditions. Fermentation experiment The optimized fermentation medium consisted of (g l -1 ): glucose, 2; peptone, 1; beef extract, 1; yeast extract, 1; Na3C6H5O7 ·2H2O, 0.1; (NH4)2HPO4, 1; K2SO4, 0.2. The seed culture was inoculated at 5% into 250 ml Erlenmeyer flasks containing 150 ml MRS medium cultured at 37°C. For batch cultures in a bioreactor, 150 ml seed culture was added to a KF-5l fermentor (KBT, Korea) with 3 L production medium at 37°C anaerobically. The samples were fetched at intervals and the pH was recorded. The microbial growth was measured by recording the optical density (OD) at 660 nm. The glucose concentration in the fermentation broth was determined by the DNS method (Zhang et al., 1997). Data scaling Because original data may disturb each other during the training process, which could mislead the predicting result, data re-scaling plays an important role to improve the predicting accuracy. Another advantage for data scaling is to avoid numerical difficulties during the calculation. All the data sets are scaled within the range of 0 and 1 in this formation shown as follows: min , max min i N i x x x x x − = − Where, i x is the feature vector which influences the L. plantarum growth in the batch fermentation, ,N i x is the normalized value of the ith process variable after the process of scaling, maxx is the maximized value of the feature vector i x , minx is the minimized value of the feature vector ix and N is the number of experimental observations. RESULTS AND DISCUSSION Variables selection Variables selection is necessary for the v-SVR model with a good prediction. In this paper, we selected the most important features to describe the fermentation process, which included temperature, pH value, glucose concentration and fermentation time. We set the four vectors as the input variables and the cell concentration as output variables. Performance criteria In the statistical prediction, the following three cross- validation methods are often used to examine a prediction model for its effectiveness in practical application; independent dataset test, subsampling (K- fold cross-validation) test and jackknife test. However, among the three cross-validation methods, the jackknife test is deemed the most objective which can always yield a unique result for a given benchmark dataset and thus, has been increasingly used and widely recognized by investigators to examine the performance quality of various predictors (Chou, 2009; Lin et al., 2009; Wu et al., 2010; Xiao et al., 2008a; Xiao et al., 2008b; Xiao et al., 2009a; Xiao et al., 2009b; Xiao et al., 2009c; Xiao et al., 2010). In this work, we used 5-fold cross-validation to examine the performance quality of the two models and chose some statistical metrics such as normalized root mean square error (NRMSE), MAPE and R. Table 1 shows these performance metrics and their calculations. NRMSE, MAPE were used to measure the deviation between the measured and predicted values. The smaller Zhang et al. 6167 Table 1. Performance metrics and their calculations. Metric Calculation MAPE MAPE= 1 100% / × n i i ii a p a n = −∑ NRMSE 2 1 2 1 ( ) NRMSE= n i ii n ii a p a = = −∑ ∑ R 1 2 2 1 1 ( )( ) ( ) ( ) n i ii n n i ii i a a p a R a a p a − − = − − = = − − = − − ∑ ∑ ∑ the values of NRMSE and MAPE were the closer the predicted values were to the measured values. The Pearson correlation coefficient R was adopted to measure the correlation of the experimental and predicted values. Growth models obtained by v-SVR The GA was used for searching the optimal parameter sets when the MAPE was at its minimum. The searching process of optimal parameters was operated with 150 generations in total. The genetic process was recorded. Figure 3 shows that the tendency of the optimal and average fitness of population varied during genetic process. We found that when the generation of population rose to 60, the optimal fitness value (MAPE) became steady. The information of the best parameters were acquired in the generation of 150 and the corresponding optimal ones were C = 235.8935, σ = 8.3608 and ν = 0.7587. The mean error applied to describe the average MAPE in the total chromosome population was 0.068969. After the GA method was applied to search for the optimal parameter of ν-SVR, the predicting model for the batch fermentation with L. plantarum was constructed. The graphical comparison of the measured and predicted value modeling by GA-ν-SVR is shown in Figure 4. Growth models obtained by logistic model The logistic model is an approximation of the microbial growth curve for the batch experiments in this study. Figure 5 showed the relationship between the cell concentration and cultivation time, including both experimental data and predicted values obtained based on the logistic model. Correspondingly, the logistic model parameters maxµ and maxx were estimated as 0.4791 and 0.3498. Performance comparison of SVR and logistic model As shown in Table 2, the values of MAPE and NRMSE generated from GA-ν-SVR model were smaller than those from logistic model, which indicated smaller deviations between the measured and predicted values. Moreover, the Pearson correlation coefficient R was higher. To illustrate the better performance of the SVR model, we compared the measured value with the predicted one. The plots between the measured value and the predicted value are shown in Figures 6 and 7. The predicting simulation was performed with the testing data. From Figures 6 and 7, the proposed GA-ν-SVR model qualified this particular data set very well (Pearson correlation coefficient was 0.9966), while two particular points lacked accuracy in the logistic model (Pearson correlation coefficient was 0.9828). Therefore, the performance of GA-ν-SVR model was better than that of the logistic equation. Conclusions Fermentation process is very complex and it is very difficult to obtain a complete picture of what is actually going on in a particular fermentation. In order to estimate the biomass online, the models presented in this work can be use to predict fermentation process of microbes. In this paper, a relatively easy strategy was given based on the investigation of ν-SVR parameters, which determined the value of ν beforehand and then, selected other two parameters(C and σ). From the results of MAPE, NRMSE and Pearson correlation coefficient, we concluded that, the approximation ability and generalization of v-SVR model were better than the logistic model for the on-line monitoring of biomass. SVR models could be considered as intracellular metabolic pathways, which are described by four variables (temperature, pH value, glucose concentration and fermentation time), while the logistic model has only one variable to describe the biomass. In conclusion, GA v- 6168 Afr. J. Biotechnol. Figure 3. The fitness alternation during the evolutionary process. Table 2. Comparison of the predicted results from GA-ν-SVR and logistic models. Model R MAPE NRMSE GA-ν-SVR 0.9966 0.0185 0.0209 Logistic 0.9828 0.0375 0.0465 Figure 4. Measured and estimated cell concentration by ν-SVR (○-measured value ▲-predicted value). 0 10 20 30 40 50 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 T he O D 66 0 v al ue o f L . p la nt ar um Fermentation time (h) Zhang et al. 6169 Figure 5. Measured and estimated cell concentration by logistic model (○-measured value ▲- predicted value). Figure 6. The plots of measured cell concentration vs. predicted value by the ν-SVR model. 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 SV R p re di ct ed v al ue ( O D 66 0) The measured value (OD 660 ) 6170 Afr. J. Biotechnol. Figure 7. The plots of the measured cell concentration vs. predicted value by logistic model. SVR has better predicting capacity on simulating bacteria biomass than logistic model and is a promising method for estimating the microbial growth. Acknowledgements This work was supported by the National Natural Science Foundation of China (Grant No. 20976161 and Natural Science Foundation of Ningbo City, China (Grant No. 2009A610149). REFERENCES Belousov AI, Verzakov SA, Frese JV (2004). A flexible classification approach with optimal generalisation performance: support vector machines. Chemom. Intell. Lab. Syst. 64: 15–25. Bessaou M, Siarry P (2001). A genetic algorithm with real-value coding to optimize multimodal continuous functions. Struct. Multidisc. Optim. 23: 63–74. Borin A, Ferrao MF, Mello C, Maretto DA, Poppi RJ (2006). Least- squares support vector machines and near infrared spectroscopy for quantification of common adulterants in powdered milk. Anal. Chim. Acta. 579: 25–32. Browne MW (2000). Cross-validation methods. J. Math. Psychol. 44: 108-132. Cai CZ, W ang W L, Sun LZ, Chen YZ (2003). Protein function classification via support vector machine approach. Math. Biosci. 185: 111-122. Chalimourda A, Schölkopf B, Smola AJ (2004). Experimentally optimal ν in support vector regression for different noise models and parameter settings. Neural Netw. 17: 127-141. Cherkassky V, Ma YQ (2004). Practical selection of SVM parameters and noise estimation for SVM regression. Neural Netw. 17: 113-126. Cho KC (2009). Pseudo amino acid composition and its applications in bioinformatics, proteomics and system biology. Current Proteomics 6, 262-274. Desai K, Badhe Y, Tambe SS, Kulkarni BD (2006). Soft-sensor development for fed-batch bioreactors using support vector regression. Biochem. Eng. J. 27: 225–239. Fujikawa H, Kai A, Morozumi SA (2004). New logistic model for Escherichia coli growth at constant and dynamic temperatures. Food Microbiol. 21: 501–509. Holland JH (1975). Adaptation in Natural and Artificial systems, University of Michigan Press. Kiviharju K, Salonen K, Leisola M, Eerikainen T (2006). Modeling and simulation of Streptomyces peucetius var. caesius N47 cultivation and ε-rhodomycinone production with kinetic equations and neural networks. J. Biotechnol. 126: 365–373 Li YF, W ang ZF, Yuan JQ (2006a). On-line fault detection using SVM- based dynamic MPLS for batch process. Chinese J. Chem. Eng. 14: 754-758. Li YF, Yuan JQ (2006b). Prediction of key state variables using support vector machines in bioprocess. Chem. Eng. Technol. 29: 313-319. Lin, W .Z., Xiao, X., and Chou, K.C., 2009. GPCR-GIA: a web-server for identifying G-protein coupled receptors and their families with grey incidence analysis. Protein Eng Des Sel 22, 699-705. Liu JZ, W eng LP, Zhang QL, Xu H, Ji LN (2002). A mathematical model for gluconic acid fermentation by Aspergillus niger. Biochem. Eng. J. 14: 137-141. Moreira GA, Micheloud GA, Beccaria AJ, Goicoechea HC (2007). Optimization of the Bacillus thuringiensis var. kurstaki HD-1 δ- endotoxins production by using experimental mixture design and artificial neural networks. Biochem. Eng. J. 35: 48–55. Mu Y, W ang G, Yu HQ (2006). Kinetic modeling of batch hydrogen production process by mixed anaerobic cultures. Bioresource. Technol. 97: 1302–1307. Peleg M, Corradini MG, Normand MD (2007). The logistic (Verhulst) model for sigmoid microbial growth curves revisited. Food Res. Int. 40: 808–818. Potocnik P, Grabec I (1999). Empirical modeling of antibiotic fermentation process using neural networks and genetic algorithms. Math. Comput. Simul. 49: 363-379. Samanta B, Al-Balushi KR, Al-Araimi SA (2003). Artificial neural networks and support vector machines with genetic algorithm for bearing fault detection. Eng. Appl. Artif. Intell. 16: 657–665. Schepers AW , Thibault J, Lacroix C (2000). Comparison of simple neural networks and nonlinear regression models for descriptive modeling of Lactobacillus helveticus growth in pH-controlled batch cultures. Enzyme. Microb. Tech. 26: 431-445. Schölkopf B, Smola AJ, W illiamson PC, Bartlett PL (2000). New Support Vector Algorithms. Neural Comput. 12: 1207-1245. Ustun B, Melssen W J, Oudenhuijzen M, Buydens LMC (2005). Determination of optimal support vector regression parameters by genetic algorithms and simplex optimization. Anal. Chim. Acta. 544: 292–305. Vapnik V (1995). The Nature of Statistical Learning Theory, Springer, New York. W ang JL, Yu T, Jin CY (2006). On-line estimation of biomass in fermentation process using support vector machine. Chinese J. Chem. Eng. 14: 383-388. Wu ZC, Xiao X, Chou KC (2010). 2D-MH: A web-server for generating graphic representation of protein sequences based on the physicochemical properties of their constituent amino acids. J. Theoret. Bio. 267(1): 29-34. Xiao X, Lin W Z (2009c). Application of protein grey incidence degree measure to predict protein quaternary structural types. Amino Acids. 37(4) ：741-749. Xiao X, Lin W Z, Chou, KC (2008a). Using grey dynamic modeling and pseudo amino acid composition to predict protein structural classes. J. Comput. Chem. 29, 2018-2024. Zhang et al. 6171 Xiao X, W ang P, Chou KC ( 2008b). Predicting protein structural classes with pseudo amino acid composition: an approach using geometric moments of cellular automaton image. J. Theor. Biol. 254, 691-696. Xiao X, W ang P, Chou KC (2009a). GPCR-CA: A cellular automaton image approach for predicting G-protein-coupled receptor functional classes. J. Comput. Chem. 30, 1414-1423. Xiao X, W ang P, Chou, KC (2010). Quat-2L: a web-server for predicting protein quaternary structural attributes. Molecular Diversity, DOI 10.1007/s11030-010-9227-8. Xiao X, W ang P, Chou KC (2009b). Predicting the quaternary structure attribute of a protein by hybridizing functional domain composition and pseudo amino acid composition. J. Applied Crystallography, 42: 169-173. Zhang LX, Zhang TF, Li LY (1997). Biochemical experimental method and technology. Higher Education Press, Beijing. 
work_nxwzu7fskrg2vol2ztglvz5hpq	----	CORELAŢIA Annals of the University of Petroşani, Economics, 11(1), 2011, 151-160 151 DECISIONS IN NEGOTIATIONS USING EXPERT SYSTEMS AND MATHEMATICAL METHODS CORNELIA MUNTEAN, ILEANA HAUER, ANTOANETA BUTUZA * ABSTRACT: The goal of this paper is at the very instant to highlight the role and the importance of mathematical multi-attribute methods, implemented in expert systems, in the preparing of international negotiation processes. We used Exsys Corvid as an expert system generator to implement a prototype of an expert system with three mathematical methods: the method of simple additive weight, the diameter method and the TOPSIS method. The applying of these methods in the exposed case study permitted the preparation and deployment of the negotiation in such a manner that the goals expected by the negotiator should be fulfilled. KEY WORDS: business decision making; negotiation preparing process; expert systems; mathematical methods; simple additive weight method; diameter method; TOPSIS method. JEL CLASSIFICATION: C02; M19. 1. INTRODUCTION Some negotiations have far too high stakes to be lost. That's the reason why it is so important to choose and apply the most effective strategy, which should warrant the winning of the negotiation. In this respect one can invoke mathematical models for obtaining a better vision upon the negotiation process and also for identifying the most efficient variant to overcome a deadlock through anticipation of the partner's movements (Vasiliu, 2003). A basic tool of mathematical models is the game theory (Neamtu & Opris, 2008), which could be very useful in identifying settlements of negotiations due to the * Assist. Prof., Ph.D., West University of Timisoara, Romania, cornelia.muntean@feaa.uvt.ro Assist. Prof., Ph.D., West University of Timisoara, Romania, ileana.hauer@feaa.uvt.ro Lecturer, Ph.D., West University of Timisoara, Romania, antoaneta.butuza@feaa.uvt.ro mailto:cornelia.muntean@feaa.uvt.ro mailto:ileana.hauer@feaa.uvt.ro mailto:antoaneta.butuza@feaa.uvt.ro 152 Muntean, C.; Hauer, I.; Butuza, A. fact that it allows designing a model of analyzing the situations where the decisions of a negotiator could affect the benefits of the other/s. But there are also situations when the negotiators must select from a multitude of variants, make a hierarchy and choose the optimal one. In these situations the most suitable mathematical tools are multi-criteria methods. The purpose of this paper is just to relieve the usefulness and importance of multicriterial methods. This is highlighted by a case study on a Romanian company by programming a prototype of an expert system used for the preparation stage of an international business negotiation process. While applying three different mathematical methods, the results were nearly the same, so the decision-maker could select the most profitable offer among many in the pre-negotiation stage, in order to organize the negotiation processes accordingly. 2. MATHEMATICAL MULTI-CRITERIA METHODS USED IN PREPARING THE NEGOTIATION PROCESS 2.1. Multi-criteria decisions Multi-criteria decision problems are those which arise when the selection of an alternative or of an action plan go through, given that the decider must consider withal several goals. These goals are frequently of different nature and could be in contradiction with each other. A decision is the compendium of some activities with open eyes of selecting the direction of an action and engagement in that, which usually involves the allocation of some resources. This decision appears following some information and knowledge and belongs to a person or to a group of people who has the necessary authority and who answer for the efficient using of the resources in some given situations. The most important characteristic of a decision is the possibility of choosing between alternatives. Most situations in real life presume the existence of some multi-criteria decisions. The decision itself consists in a selection, “a good choice” from a number of available solutions. Every solution is an alternative. In context of multi-criteria decisions, the selection goes through the evaluation of each possible variant. The criterions will be compulsory quantifiable, even if comes through only by a nominal scale (of type yes/no), and their result must be calculated for each alternative, individually. The results of these criterions are the minimal necessary information for comparing de available variants and, accordingly, they facilitate the selection of one of these, of the most suitable one. The complexity of problems which require to make a decision taking into account several criterions which are to be accomplished, derives from the fact that, whether the state of certainty or the state of uncertainty is involved, the results of a decisional variant must include more sometimes incomparable attributes, and to measure these for comparison is difficult. Practically it is impossible to achieve the maximum levels desired separately, for each of the criterions, at the same time. Decisions in Negotiations Using Expert Systems and … 153 The decisions must be the result of a thinking process, preceded by an information and a thoroughgoing analysis of all the problem presumptions, of the influence elements, considering the concrete conditions in which each enterprise conducts its activities. For choosing the best decision there must exist the following elements: - an economic objective or a well established goal which could be quantified; - a great amount of information which should reflect as well as possible the economical phenomena and processes that actually occur, with influences on adopting a decision; - a well done investigation and data processing machine, to allow the achievement of a reasonable selection process. Commonly, the results of a criterion in the case of a decision process appears in a table (called decisional matrix or decisional table), delimited by several columns and lines. The lines of a table represent the alternatives and the columns the criterions. A value that resides at the intersection of a line and a column in a table represents the result of a criterion – a calculated or evaluated characteristic of an alternative regarding a criterion. The decisional matrix is the central structure for MCDM (Multiple criteria decision making), because it contains the necessary data for comparing the alternatives of the decision. The process of decision-making is defined by following elements (adapted from Gheorghiţă, 2001): the decision-maker, the assemblage of decision alternatives, the assemblage of decision criterions, the assemblage of goals. The decision-maker is the person who must select the most advantageous variant from a lot of possible ones, variant called the optimum choice. The assemblage of decision alternatives, V, is the assemblage of action possibilities at a given moment. The assemblage of decision criterions, C, is the assemblage of parameters which defines the process and in respect of which we have in view the comparison of alternatives. The decision criterions are characterized by a number of levels according to the different alternatives and/or status of impartial conditions. Decision models with an assemblage of criterions, called also multi-criteria decision models, could be multi-attribute decision models, which are presented below, or multi-objective decision models, which are subject of linear programming. 2.2. Mathematical multi-attribute decision models Multi-attribute decision models subsist in the determination of the optimum variant from a finite variant assemblage V={V1, V2, ...,Vm}, variants that are compared one with another in respect with numerical or non-numerical criterions belonging to a finite assemblage C={ C1, C2, ...,Cn}. Each criterion has a minimum or maximum goal. For some multi-attribute decision problems, in which the matrix of consequences contains heterogeneous data, numerical or non-numerical, the homogenization of these 154 Muntean, C.; Hauer, I.; Butuza, A. data is done by the normalization procedure [6], which transforms the matrix of consequences in a matrix R=(rij)i=1,m; j=1,n} with elements in the interval [0,1]: rij = ⎪ ⎪ ⎩ ⎪ ⎪ ⎨ ⎧ ≤≤ ≤≤ criterionsfor a a criterionsfor a a ij ijmi ij mi ij min, min max, max 1 1 (1) In almost all multi-attribute decision problems there is information regarding the importance of each criterion. This is generally expressed by the vector P={p1, p2, ..., pn} and indicates the level of importance given by the decision-maker to each criterion. Every multi-attribute decision problem could be expressed by a matrix A, called the matrix of consequences, with elements aij indicating the evaluation (consequence) of variant i, i=1, 2, ..., m ((Vi)), by criterion j, j=1, 2, ..., n, (Cj). The data could be stored in table 1, where P={p1, p2, ..., pn} indicates the level of importance given by the decision-maker to each criterion. Table 1. The matrix of consequences Ci Vj C1 ... Cn V1 a11 ... a1n ... ... ... ... Vm am1 ... amn P p1 ... pn Source: Andrasiu, et. al., 1986 Multi-attribute decision problems could be classified into three categories: direct methods, indirect methods and methods which use a certain distance for the construction of hierarchies (Neamţu, 2008) In our case study we will use two direct methods (the method of simple additive weight and the diameter method) and a method which uses the distance (TOPSIS), which are presented below. 2.3. The method of simple additive weight The method consists in defining the function f : V→R, given by: Decisions in Negotiations Using Expert Systems and … 155 f(Vi) = mi p rp n j j n j ijj ,1, 1 1 = ∑ ∑ = = (2) The optimum variant will be that for which f(Vi) takes the maximum value. The method uses the normalized matrix. 2.4. The diameter method The diameter method has the advantage that in the hierarchy of variants it takes into consideration the homogeneity or heterogeneity of the data in respect to the different criteria. A variant will be homogeneous if it takes close values for all criteria and will be heterogeneous if it takes very big values for some criteria and very small values for others, with the presumption that all criteria are alike (minimum or maximum). We can build the matrix: L= (3) ⎟⎟ ⎟ ⎟ ⎟ ⎠ ⎞ ⎜⎜ ⎜ ⎜ ⎜ ⎝ ⎛ mnmm n n LLLm LLL CCC ... ............... ...1 ... 21 11211 21 in which the column j, j=1,n contains the variants corresponding to the ordering of the elements of the assemblage {a1j, a2j, ..., amj} in ascending (descending) order if the criterion Cj is a maximum (minimum). If ai1j, ai2j, ..., aimj, i=1,m, j=1,n is the ordering of the assemblage {a1j, a2j, ..., amj}, j=1,n, then L1j=Vi1, L2j=Vi2, ..., Lmj=Vim. For i=1..m and j=1..n are defined the following functions: • The estimate function: A:V→R, A(Vi) = (4) ( )[ ] ∑∑ == − n j jj n j ji ppCVlocm 11 /, • The diameter function: d:V→N d(Vi) = )],([min)],([max jijjij CVlocCVloc − (5) where loc : V x C →{1, 2, ..., m} loc(Vi,Ci)=k, k = m,1 ⇔ Vi = Lkj (6) • The aggregation function: Aggr : V→ R , Aggr(Vi) = ( ) ( ) 2 VidmVA i −+ (7) 156 Muntean, C.; Hauer, I.; Butuza, A. The hierarchy of variants is given by the descending values of the aggregation function (Aggr). 2.5. The TOPSIS method The TOPSIS method (Technique for Order Preference by Similarity to Ideal Solution) is based on the idea that the optimum variant must have the minimum distance to the ideal solution. The steps of the TOPSIS method are: - Step 1. We build the normalized matrix R=(rij), i=1,...,m, j=1,...,n; - Step 2. We build the weighted normalized matrix V=(vij), i=1,...,m, j=1,...,n, where vij = ∑ = n j j ijj p rp 1 (8) Step 3. We calculate the ideal solution A and the ideal negative solution B, defined as: A= (a1, a2, ..., an), B= (b1, b2, ..., bn) where: aj = (9) ⎪⎩ ⎪ ⎨ ⎧ ≤≤ ≤≤ min,min max,max 1 1 isCcriteriontheifv isCcriteriontheifv jij mi jij mi bj = (10) ⎪⎩ ⎪ ⎨ ⎧ ≤≤ ≤≤ max,min min,max 1 1 isCcriteriontheifv isCcriteriontheifv jij mi jij mi Step 4. We calculate the distance between the solutions: Si = ( )∑ = − n j ij ajv 1 2 , i = 1, 2, … ,m; (11) Ti = ( )∑ = − n j jij bv 1 2 , i = 1, 2, ... , m; (12) Step 5. We calculate the relative nearness from the ideal solution: Ci = ii i TS T + (13) Step 6. We make a classification on the assemblage V according to the descending values of Ci obtained in step 5. Decisions in Negotiations Using Expert Systems and … 157 3. CASE STUDY The case study in this paper wants to mark out the role of mathematical methods implemented in an expert system for the stage of preparation in an international negotiation process. An Expert System (Andone I., Mockler R., 2001) is a knowledge-based computer program containing expert domain knowledge about objects, events, situations and courses of action, which emulates the process of human experts in the particular domain. For long term use, a knowledge base stores rules, facts and other knowledge structures, much as a database stores data. When the ES is used, an inference engine processes the knowledge structures, bringing problem specific information into the system, and makes recommendations to the user based on the information and knowledge structures available. Exsys Corvid (www.exsys.com) provides an object-oriented structure that makes it easy to build expert systems using methods and properties of variables, while not requiring the developer to change the way they think and describe their decision- making steps and logic (Muntean, Butuza, Dobrican 2002). The result is a very flexible and powerful development environment that can easily be learned. We used Corvid for implementing our application. We considered the Romanian textile company IASITEX S.A. and used the three multi-attribute methods presented in paragraph 2 (the method of simple additive weight, the diameter method, the TOPSIS method) for selecting, in the stage of pre- negotiation, of the best offer and for organizing the negotiation processes thereafter. After an initial stage of commercial tender, Iasitex S.A. will have to decide between two offers of two foreign companies, taking into account six selection criterions: • C1 : the account of the good that has to be purchased (million Euro); • C2 : requested advance money (%); • C3 : time period allowed for the payments (years); • C4 : payment staggering (month); • C5 : guarantee period (years); • C6 : offer validity (month). The following two offers of two foreign companies will be approached further as potential variants of the Romanian company Iasitex S.A: • the offer of VAKONA GmbH, from Germany – variant 1 (V1); • the offer of Hashima Co., Ltd., from Japan – variant 2 (V2) In following table are presented the offers of the two companies according to the criteria invoked by Iasitex S.A. Table 2. The characteristics of the variants according to the criteria C1Mil.Euro C2 % C3 years C4 month C5 years C6 month V1 10 10 3 12 2 1 V2 8 15 2 6 3 2 158 Muntean, C.; Hauer, I.; Butuza, A. The Romanian company Iasitex S.A. confers to each invoked criterion a specific rate of importance on a scale from 0 to 1: For C1 :0.3; for C2 : 0.2; for C3 : 0.1; for C4 : 0.1; for C5 :0.2; for C6 : 0.1. For doing all calculations more quickly, we used the expert system generator Corvid (Muntean & Muntean 2010, pp.199-205), and put all entrance data into two input files, the first one used as a metablock, containing the characteristics of each variant in respect to each criterion (figure 1), and the other as a simple data file, containing the importance rates of the six criteria (figure 2). Figure 1. The content of the input text file, used further as a metablock Figure 2. The content of the simple input file, containing the importance rates Further we will apply, using Corvid variables, the three mathematical methods described in paragraph 2 (the method of simple additive weight, the diameter method and the TOPSIS method) for deriving the best variant between the two remaining offers for the Romanian company. Finally we will compare the results obtained with each of the three methods. Considering that the matrix in figure 1 contains heterogeneous data, there will be necessary a normalization procedure and we will obtain a normalized matrix R=(rij) with i= 2,1 j= 6,1 . For the Romanian company every criterion must be optimized, and this occurs by minimization for C1, C2, and by maximization for C3, C4, C5, C6. For the maximum and minimum criteria we used relation (1) and calculated de elements of the normalized matrix R. The obtained normalized matrix is: R= (14) ⎟ ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎜ ⎝ ⎛ 115.067.067.01 5.067.01118.0 654321 CCCCCC 4. RESULTS, DISCUSSIONS AND CONCLUSIONS Applying the first method, the simple additive weight method, we obtain following values for the functions f(Vi) using relation (2): f(V1) = 0,82 and f(V2) = 0,85. According to this method, the order of the variants is: V2 → V1. (figure 3). Next we apply the diameter method. We calculate with Corvid variables the aggregation functions Aggr(Vi), with i= 3,1 using (7). We obtain: Aggr(V1) = 0,7 and Decisions in Negotiations Using Expert Systems and … 159 Aggr(V2) = 0,8. So, applying this method the variants order is the same: V2 → V1. (figure 4) Figure 3. Results calculated by the Expert System with the simple additive weight method Figure 4. Results calculated by the Expert System with the diameter method Finally, we apply the TOPSIS method. After calculating the relative nearness from the ideal solution Ci, with i= 3,1 using relation (13), we obtain: →C1= 0,44 →C2= 0,56 We make a classification on the assemblage V according to the descending values of Ci, and we obtain following order of variants: V2 → V1 (fig.5), the same as with the other two methods. Figure 5. Results calculated by the Expert System with the TOPSIS method As a conclusion, this paper presented a short overview of three mathematical multi-criteria methods that were implemented in an expert system for deciding in a negotiation problem. The expert system used in a quick and easy way the input data as text files and calculated, using confidence variables, the functions that indicate the proposed order of variants, accordingly to each of the three different methods. 160 Muntean, C.; Hauer, I.; Butuza, A. In negotiation, as well as in most other circumstances, people must make a decision from among a lot of possible decisions, in order to achieve a certain goal. It is perfectly normal for human reasoning to analyse and to compare the possibilities, in order to adopt that decision which permits the best fulfilment of the desired goal. Although we frequently use the term "optimal decision", in most situations this "optimality" is a very complex concept which can't be defined but by mean of a mathematical model. Mathematical models appeared and were used in the process of decision making in business, particularly in negotiation, quite from the necessity to sustain the logical reasoning in negotiation and to manage a great number of factors simultaneously. Furthermore, applying these mathematical models permits the approach of some new qualitative problems, so that it’s not surprising at all the fact that in negotiation there are used more and more mathematical tools, methods and techniques. REFERENCES: [1]. Andone, I.; Mockler, R.; Tugui, A. (2001) Developing expert systems in economy, Editura Economică, București [2]. Andraşiu, M.; Baciu, A.; Pascu, A.; Puşcaş, E.; Taşnadi, A. (1986) Metode de decizii multicriteriale, Editura Tehnică, Bucureşti [3], Gheorghiţă, M. (2001) Modelarea şi simularea proceselor economice, Editura ASE, Bucureşti [4]. Ionescu, G.; Cazan, E.; Negruţ, A.L. (2000) Modelarea şi optimizarea deciziilor manageriale, Editura Dacia, Cluj – Napoca [5]. Muntean, C.; Butuza, A.; Dobrican, O. (2007) Sisteme expert. Elemente de teorie şi aplicaţii. Editura Mirton, Timisoara [6]. Muntean, C.; Hauer, I. (2010) Improving the management in organizations by using Expert Systems, Simpozion Ştiinţific Internaţional, „Managementul dezvoltării rurale durabile”, Timişoara [7]. Muntean, C.; Muntean, M. (2010) Multicriterial Methods Used in Expert Systems for Business Decision Making, Revista Informatica Economică, Vol.14, No.3/2010, pp.199- 205 [8]. Neamţu, M.; Opriş, D. (2008) Jocuri economice. Dinamică economică discretă. Aplicaţii, Editura Mirton, Timişoara [9]. Vasiliu, C. (2003) Tehnici de negociere şi comunicare în afaceri, Editura ASE, Bucureşti [10]. Zimmermann, H.J. (2001) Fuzzy set theory and its applications, 4th Edition, Springer [11]. http://www.exsys.com and http://www.exsys.com/CorvidTutorials.html http://www.biblioteca.ase.ro/catalog/detalii.php?c=2&q=curry&st=s&dela=0&ct=108788 http://www.biblioteca.ase.ro/catalog/detalii.php?c=2&q=curry&st=s&dela=0&ct=108788 http://www.exsys.com/ http://www.exsys.com/CorvidTutorials.html 2. MATHEMATICAL MULTI-CRITERIA METHODS USED IN PREPARING THE NEGOTIATION PROCESS 
work_o226xx3labcrbpre5mmuxnyeuu	----	Document downloaded from: This paper must be cited as: The final publication is available at Copyright http://dx.doi.org/10.1016/j.eswa.2011.04.092 http://hdl.handle.net/10251/38498 Elsevier Braysy, O.; Martínez Molada, E.; Nagata, Y.; Soler Fernández, D. (2011). The mixed capacitated general routing problem with turn penalties. Expert Systems with Applications. 38(10):12954-12966. doi:10.1016/j.eswa.2011.04.092. The mixed capacitated general routing problem with turn penalties Olli Bräysy1, Eulalia Mart́ınez2, Yuichi Nagata3, David Soler2∗ (1) Agora Innoroad Laboratory, Agora Center, P.O.Box 35, FI-40014 University of Jyväskylä, Finland (2) Instituto Universitario de Matemática Pura y Aplicada. Universidad Politécnica de Valencia Camino de Vera s/n, 46022, Valencia, Spain (3) Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, 4259 Nagatsuta Midori-ku Yokohama, Kanagawa 226-8502, Japan Abstract In this paper we deal with the Mixed Capacitated General Routing Problem with Turn Penalties. This problem generalizes many important arc and node routing problems, and it takes into account turn penalties and forbidden turns, which are crucial in many real-life applications, such as mail delivery, waste collection and street maintenance operations. Through a polynomial transfor- mation of the considered problem into a Generalized Vehicle Routing Problem, we suggest a new approach for solving this new problem by transforming it into an Asymmetric Capacitated Vehicle Routing Problem. In this way, we can solve the new problem both optimally and heuristically using existing algorithms. A powerful memetic algorithm and a set of 336 new benchmark instances are also ∗Corresponding author: Address: Instituto Universitario de Matemática Pura y Aplicada. Univer- sidad Politécnica de Valencia. Camino de Vera s/n, 46022, Valencia, Spain. Tel:+34 963877007-76667; fax:+34 963877669. E-mail addresses: olli.m.p.braysy@jyu.fi (O. Bräysy), eumarti@mat.upv.es (E. Mart́ınez), nagata@fe.dis.titech.ac.jp (Y. Nagata), dsoler@mat.upv.es (D. Soler). 1 suggested. The experimental results show that the average deviation of the sug- gested solution method is less than 0.05 per cent with respect to optimum. Keywords: Vehicle routing, capacitated general routing problem, turn penalties, transformation. 1 Introduction In many real-life vehicle routing problems it is important to consider the risk and time associated with turns, i.e., the cost of the turn. Moreover, some turns, especially U- turns, can be forbidden. This last implies that a vehicle route made with a classical graph route generator may be illegal if it does not respect the traffic signals. For example, Figure 1 shows a traffic signal located in an arterial avenue from Valencia (Spain). It indicates that the next two left turns are forbidden and the way to dodge them. These forbidden turns cannot be taken into account if we model the city map with a classical graph. Considering turn penalties and forbidden turns is particularly important in down- town areas and for large-size vehicles. But also turn penalties are important in order to save time in tours on foot. From Figure 2 it is easy to see that going on foot, right turn a → b can be considered with zero cost, while turn a → c is much more time-consuming, because it implies to cross two streets, with up to two traffic lights. Usually the real-life vehicle routing software are based on separate modules for shortest path calculation and vehicle route optimization. The latter is based on given time and distance matrices, calculated typically in the beginning with the shortest path procedure. In this context, the distance and time calculation are based on fixed and known stopping points (usually addresses) for the vehicles and the possible turn penalties and forbidden turns are taken into account during the shortest path calcu- lation. However, in practice, there are several applications where the exact stopping points are not known a priori and are part of the optimization problem, such as mail collection and delivery, waste collection and street maintenance operations. In these cases, including turn penalties and forbidden turns in the vehicle routing model is very important. 2 Figure 1. Traffic signal indicating two forbidden left turns. Figure 2. A street crossing. So far research on extended vehicle routing models with turn penalties and forbidden turns has been scarce. For previous research, see Benavent and Soler (1999), Clossey et al (2001), Corberán et al (2002) and Soler et al (2008), that generalize several well- known single vehicle routing problems to the existence of turn penalties. They provide theoretical results about complexity and resolution and/or computational results on these extensions. The last cited work studies the most general problem, the Mixed General Routing Problem (MGRP) with turn penalties, that includes all the cases studied in the previous cited papers. MGRP consists of finding a minimal cost closed walk on the links of a mixed graph G which traverses a given subset of “required” links and a given subset of “required” vertices. With respect to multivehicle routing problems, early papers that considered turn penalties focused on real-life applications and heuristic solution methods. See for exam- ple Bodin et al (1989) and Roy and Rousseau (1989). Later, turn penalties have been considered in the context of the Mixed Capacitated Arc Routing Problem (MCARP), see for example Bautista and Pereira (2004) and Belenguer et al (2006). The MCARP is an arc routing problem, in which a fleet of vehicles (with a known capacity) is based 3 on a specified vertex (the depot) and must service a subset of the links of a mixed graph, with minimum total cost and such that the load assigned to each vehicle does not exceed its capacity and each link is serviced by exactly one vehicle. Bautista et al (2008) present two ant colony metaheuristics for a real urban waste collection problem. This real problem is modeled as a particular case of the MCARP with turn penalties, in which they only consider two kind of turns: forbidden or allowed with zero cost. Finally, Perrier et al (2008) heuristically solve a real vehicle routing problem in the context of snow plowing operations that also takes into account the existence of forbidden turns. The multivehicle extension (with capacity constraints) of the MGRP is called the Mixed Capacitated General Routing Problem (MCGRP). Due to its complexity, there are only few works on MCGRP, see e.g. Jansen (1993) who studied the undirected case and Pandit and Muralidharan (1995). However, particular cases of the MCGRP, such as the capacitated arc routing problem (CARP) and the capacitated vehicle routing problem (CVRP) have attracted a huge amount of research. In this paper we present a generalization of the MCGRP that considers turn penal- ties and forbidden turns. The objective is to minimize the sum of the costs of the traversed arcs and edges together with the penalties associated with the turns made. We call the new problem the Mixed Capacitated General Routing Problem with Turn Penalties (MCGRPTP). As far as we know, this is the first time that MCGRPTP is presented in the liter- ature. Moreover, there is no previous research on multivehicle node routing problems with turn penalties. In this paper we also present for the first time an approach for solving capacitated routing problems with turn penalties, through suggesting a new polynomial transformation from the MCGRPTP to an asymmetric CVRP (ACVRP). To be more precise, the transformation is done in two steps: we first transform the MCGRPTP into a generalized VRP (GVRP), using a new approach suggested in this paper. The key idea of GVRP, compared to CVRP is that in GVRP each customer has several alternative service locations, and only one of them has to be selected for service. For more details on GVRP, see e.g. Ghiani and Improta (2000). In the second step we transform the GVRP into an ACVRP, using a model presented in Soler et al (2009). 4 Finally, we present a set of new benchmark problems and a very powerful memetic algorithm for the ACVRP, based on a previous study by Nagata and Bräysy (2009). The experimental results show an average deviation equal to 0.05% for instances with known optimal solution and that large-size problems can be solved with the suggested MA. The suggested transformation makes it possible to use also any other powerful algorithm developed for ACVRP, see e.g. Fischetti et al (1994), Vigo (1996) or the more recent heuristic by De Franceschi et al (2006). The remainder of this paper is organized as follows. Section 2 introduces some def- initions and notations in order to formally define and solve the MCGRPTP. In Section 3, through two transformations, we prove that the MCGRPTP can be transformed in polynomial time into a Generalized Vehicle Routing Problem (GVRP). It is known (Soler et al (2009)) that the GVRP can be transformed into an ACVRP, so in Section 4 we show computational results for several sets of ACVRP benchmarks. Finally, in Section 5 we present our conclusions. 2 Definitions and notations First, to our aim, we formally define two known problems cited before: the Asymmetric Capacitated Vehicle Routing Problem (ACVRP) and the Generalized Vehicle Routing Problem (GVRP). The second one is an extension of the ACVRP, introduced by Ghiani and Improta (2000), that can model several real-world situations and that will be the “cornerstone” to solve our problem: The ACVRP is defined as follows: Let G = (V, A) be a complete digraph, V = {vi}ni=0 being its set of vertices, where v0 is the depot vertex. Each vertex vi with i > 0 has an associated demand di > 0 and each arc (vi, vj) ∈ A has an associated cost ci,j ≥ 0. Moreover, a fleet of k vehicles with the same capacity W is available at the depot. Find a set of k shortest routes, each starting and ending at the depot, such that each vertex vi (∀ i ∈ {1, ..., n}) must be visited by one and only one vehicle and the sum of the demands of the vertices visited by each vehicle does not exceed its capacity W . 5 As in other papers cited in the introduction, in all the capacitated routing problems discussed here, we will consider k to be equal to the minimum number of vehicles needed to serve all demands. The GVRP is defined as follows: Let G = (V, A) be a directed graph where the vertex set V is partitioned into m + 1 nonempty subsets S0, S1, ..., Sm such that S0 has only one vertex v0 (the depot), Sh (h = 1, ..., m) represents l(h) possible locations of the same vertex which has associated a positive demand dh, and each arc (vi, vj) ∈ A has associated a cost ci,j ≥ 0. Moreover, a fleet of k homogeneous vehicles having the same capacity W is available at the depot. Find a set of k shortest routes, each starting and ending at the depot, such that each subset Sh (h = 1, ..., m) is visited exactly once and the sum of the demands of every route does not exceed the capacity W of the vehicle. Next, we need to show some concepts and notations that have been used in previous works on turn penalties: Given a mixed graph G = (V, E, A), each pair of links a = (u, v), b = (v, w) ∈ E ∪A has an associated turn at v, based on going from a to b, denoted as [ab]. Moreover, if a, b ∈ E, the same pair has another associated turn at v, based on going from b to a, denoted as [ba]. Each edge e incident with v has an associated U-turn at v that, if necessary, will be denoted by [eve]. Each link in G has associated a nonnegative cost and each allowed turn in G has associated a nonnegative penalty. Given a = (u, v), b = (s, t) ∈ E ∪A, a v-s feasible chain from a to b is an alternating sequence of links and allowed turns {a1, [a1a2], a2, . . . , [ar−1ar], ar, [arb]}, where a1 = a. The cost of a feasible chain is defined as the sum of the costs of the arcs it traverses plus the sum of the penalties of the turns it makes. A v-s feasible chain from a to b is closed if a = b and s 6= v. Given a = (u, v), b = (s, t) ∈ E ∪ A, a shortest (minimum cost) v-s feasible chain from a to b will be denoted by s.f.c.(va, sb). Note that a feasible chain is defined such that it begins at a link and ends at a turn. This is very important in the context of forbidden or penalized turns. In classical routing problems, if we have to go from a vertex u to a vertex v and then to a vertex w, we only have to connect the shortest path from u to v with the shortest path from v 6 to w. But even if these shortest paths have been constructed taking into account turn conditions, the connection of both paths at v can give rise to an unavoidable forbidden turn (U-turn for example). In our case, the connection between two feasible chains at a vertex v is possible only if the first one ends at a turn [(t, v)(v, s)] and the second one begins at the link (v, s), which avoids the existence of forbidden turns. Therefore, we cannot use paths between vertices as in the classical way, and this increases the difficulty of modeling how to serve demands at vertices, specially if these vertices have undesirable turns. With these previous concepts we can formally define the problem that we study in this paper. The Mixed Capacitated General Routing Problem with Turn Penalties (MCGRPTP) is defined as follows: Let G = (V, E, A) be a mixed graph where each link (i, j) ∈ E ∪A has an associated cost cij ≥ 0 and each turn [ab] has an associated penalty p[ab] ≥ 0 ( p[ab] = +∞ if turn [ab] is forbidden). One of the vertices, say v0, represents the depot where there are k vehicles of an identical capacity W > 0 and in it all the turns are allowed with zero penalty. Let R ⊆ E ∪ A be a set of required links such that each (i, j) ∈ R has an associated positive demand qij ≤ W , and let VR ⊆ V −{v0} be a set of required vertices such that each v ∈ VR has an associated positive demand qv ≤ W . Find k closed feasible chains in G, one for each vehicle, that minimize the total cost and such that each chain passes through the depot, each demand is served by only one vehicle and the total demand served by each vehicle does not exceed its capacity W . Note that allowing all the turns with zero penalties at the depot is due to the fact that in real-world situations, the depot normally represents a warehouse from which the vehicles begin their journey and to which they return. It makes no sense considering forbidden/penalty turns in the warehouse as the truck leaves from depot independently of the route the truck made before. Moreover, these warehouses are usually placed outside the cities with good road communications and of easy access. Hereinafter and as in similar papers, each non-required edge will be replaced by two arcs of the same cost and opposite direction; then we assume that all edges in the graph are required (ER = E, with R = ER ∪ AR). Finally, for simplicity, we will not write 7 the middle turns of a feasible chain. For example, the feasible chain {a, [ab], b, [b, c]} will be written as {a, b, [b, c]}. In the particular case of the MCGRPTP in which VR = ∅, we have the MCARP with turn penalties, and in the particular case of the MCGRPTP in which k = 1 we have the MGRP with turn penalties. Therefore, the problem presented here generalizes both the single vehicle and the multivehicle routing problems with turn penalties studied in the literature. 3 Solving the MCGRPTP To solve the MCGRPTP, we will first transform it into a GVRP, which in turn can be transformed into an ACVRP. 3.1 Transformation of the MCGRPTP into a GVRP Let G = (V, E, A) be a mixed graph where a MCGRPTP is defined, E ∪ AR being the required link set and VR the required vertex set. Due to the fact that in the GVRP the demand is at the vertices, we will transform graph G in which we have defined the MCGRPTP into a directed graph G∗ = (V ∗, A∗) such that the vertices in G∗ are related with the required links and required vertices in G. To this aim, we first construct an intermediate directed graph G′ = (V ′, A′) from G as follows: First, each required edge is replaced in G by two opposite required arcs, both with the edge cost and the edge demand. Second, subset VR is partitioned into two subsets, VR1 and VR2 , such that ver- tices containing all allowed zero-penalty turns belong to VR1 , and vertices containing forbidden or positive-penalty turns belong to VR2 . For each v ∈ VR1 ∪ {v0}, replace vertex v in G by two vertices ve and vl, so that ve has only entering arcs (the arcs entering at v) and vl has only leaving arcs (the arcs leaving from v). Add a required arc av = (v e, vl) to G such that all turns at ve and vl are allowed with zero penalty and the demand of this arc in G′ is the one of the required vertex v in G (the demand corresponding to the arc from the depot vertex is 8 obviously zero). Note that traversing arc av in G ′ is then equivalent to passing through vertex v in the original graph G. For each v ∈ VR2 do: - Replace vertex v in G by the same number of vertices vij as those of allowed turns [aibj] at v, so that each of these copies has only one entering arc (ai) and one leaving arc (bj), with its corresponding allowed turn. Note that if ai is an entering arc at v, G′ will contain at least as many copies of the entering arc ai as there are allowed turns involving ai at v. The same applies to the leaving arc bj from v. - Then, replace each vertex vij by two vertices, v e ij and v l ij, and add a “required” arc avij = (v e ij, v l ij) between them with cost zero, such that p[aiavij ] = p[aibj ] and p[avij bj ] = 0, i.e. the penalty that was in the turn at vij is moved to vertex v e ij, and all these arcs will have the same demand, the one of the required vertex v. Note that traversing only one of these required arcs avij in G ′ involves passing through vertex v in G. Note that in the last paragraph we have written required in quotes because for each v ∈ VR2 , only one of the generated arcs must be served. After this transformation we have a directed graph G′ = (V ′, A′) such that the subset A′R comes from the required arcs, required edges and required vertices in G. A′R will give rise to a partitioned set of vertices V ∗ in the graph G∗ in which we will define the GVRP. Each arc between two of these vertices that do not form part of the same subset, will have associated the cost of the shortest feasible chain between the two corresponding links in G′. From G′ we then construct the graph G∗ = (V ∗, A∗) as follows: - For each arc av ∈ A′R with v ∈ VR1 ∪ {v0}, associate a vertex set Sv = {xav} in G∗, with its corresponding demand in xav . - For each v ∈ VR2 , associate a vertex set Sv in G∗ with as many vertices xavij as arcs avij are in A ′ R, all of them with the same corresponding demand. 9 - For each arc a ∈ A′R that comes from a required arc in G, associate a vertex set Sa in G ∗ with as many copies of a vertex xa as copies of arc a appear in G ′, all of them with the same corresponding demand. - For each pair of opposite required arcs e1, e2 ∈ A′R that come from a required edge e in G, associate a vertex set Se in G ∗ with as many copies of vertices xe1 and xe2 as copies of arcs e1 and e2 respectively appear in G ′, all of them with the same corresponding demand. - For each pair of vertices xa, xb ∈ V ∗ with xa ∈ Si, xb ∈ Sj, i 6= j being a = (u, v) and b = (s, t), add arcs (xa, xb) and (xb, xa) to A ∗, with the cost of the s.f.c.(va, sb) and of the s.f.c.(tb, ua) respectively in G′ (if they exist). - There is no arc between vertices belonging to the same Si. Given an MCGRPTP in G, we define a GVRP in G∗ where the vertex set V ∗ is partitioned into the following subsets: Sv for all v ∈ VR1 ∪ {v0} (hereinafter we will denote the subset corresponding to the depot by S0 = {vD}), Sv for all v ∈ VR2 , Sa for all a ∈ AR and Se for all e ∈ ER. Let us see an example of the construction of the auxiliary graph G∗. Consider the mixed graph given in Figure 3 corresponding to an MCGRPTP in which vertex 1 represents the depot, there are two vehicles both with capacity 55, and there are three required arcs a, b and c, with demands 25, 25 and 20 respectively, a required edge e with demand 20 and a required vertex 5 with demand 10. Then the total demand in the graph is 100 units. Link costs and demands appear in Figure 3 with different size (demands appear in bold and small size). We will suppose in this graph that all U-turns are forbidden except at vertex 1 (the depot) at which all turns are allowed with penalty zero, and in the rest of the vertices, right turns (according to the drawing of the graph) have penalty 1, left turns have penalty 3 except for the turn from arc b to arc (2, 1) that is considered forbidden, and going straight ahead, as it occurs in the turn from arc a to edge e, has penalty zero. 10 Figure 3. Graph G with AR = {a, b, c}, ER = {e} and VR = {5}. Starting from the information given by the initial graph G = (V, E, A), where AR = {(3, 4), (4, 2), ((5, 6)}, ER = {(4, 6)}, VR = {5} and the depot is vertex 1, we construct the intermediate graph G′ (see Figure 4): First, we replace the required edge e = (4, 6) with demand qe = 20 by two opposite arcs e1, e2 with the same cost 5 and the same demand 20. Second, we replace vertex 1 by the sequence {1e, a1, 1l}, a1 with cost zero. Finally, we transform vertex 5, which belongs to VR2 (here VR1 = ∅), into four arcs, as many as allowed turns in it, all of them with cost zero and demand 10. Figure 4. Intermediate graph G′. From the intermediate graph G′ = (V ′, A′) we can already construct the directed graph G∗ = (V ∗, A∗). The vertex set V ∗ is given by the following partition: 11 - A subset S0 with only one vertex xD representing the depot and corresponding to the required arc a1. - Subsets Sa and Sb, each one with only one vertex xa and xb respectively, for the required arcs a = (3, 4) and b = (4, 2), both with demand 25, and subset Sc with two vertices xc1 and xc2 , both with demands 20, for the required arc c = (5, 6) which have two copies in G′. - Subset Se with two vertices xe1 and xe2 both with demand 20, for the required edge e = (4, 6). - Subset S5 with four vertices x5f h , x5f c , x5dc and x5gh all of them with demand 10, for the required vertex 5 in VR2 . Figure 5 shows graph G∗ in which, for simplicity, each pair of arcs (xr, xt) and (xt, xr) with xr ∈ Si, xt ∈ Sj and i 6= j has been drawn as a line with two arrow heads, one at each end, and the arc costs (normally different for each direction) have been omitted. Figure 5. Directed graph G∗ associated with G. Theorem 1 An MCGRPTP defined in G can be transformed in polynomial time into the corresponding GVRP defined in G∗. 12 Proof. By the construction of G′, given B = {Ti}ki=1 a set of k feasible closed chains in G corresponding to a solution to the MCGRPTP, we can associate with B a set B′ = {T ′i}ki=1 of k feasible closed chains in G′ such that for all i ∈ {1, . . . , k} we have: - T ′i traverses arc a1 (corresponding to the depot node 1 in G). - Ti passes through a vertex v ∈ VR1 iff T ′i traverses arc av in G′. - Ti passes through a vertex v ∈ VR2 iff T ′i traverses an arc avij . - Ti traverses arc a ∈ AR iff T ′i traverses a copy of arc a in G′. - Ti traverses edge e ∈ E iff T ′i traverses a copy of arc e1 or a copy of arc e2 in G′. - T ′i has the same cost as Ti. - Moreover, we will suppose that if Ti satisfies demand at v ∈ V1 (v ∈ V2) (a ∈ AR) (e ∈ E), then T ′i satisfies the demand located at av (one and only one arc avij ) (one and only one copy of arc a in G′) (one and only one copy of e1 or e2 in G ′). Summarizing, for all i ∈ {1, . . . , k} T ′i is a feasible closed chain in G′ that traverses a1, satisfies the same demands as Ti and has the same cost as Ti. Once we have constructed B′ from B, for each i ∈ {1, . . . , k}, from T ′i we construct a cycle CBi in G ∗ as follows: Suppose that T ′i satisfies, in this chronological order, the demands qj1 , qj2 , . . . , qjmi . Then CBi is a cycle in G ∗ that visits, in this order, the sets of vertices S0, Sj1 , Sj2 , . . . , Sjmi , S0, and for all t ∈ {1, . . . , mi}, Ci visits only the vertex at St coming from the arc in G′ at which demand has been satisfied by T ′i . It is evident that the set of cycles LB = {CBi }ki=1 is a solution to the GVRP in G∗ its cost c∗(LB) being less than or equal to c(B), this last due to the fact that the cost of the route segment of one feasible closed chain T ′i in G ′ between two consecutive serviced arcs (including the depot arc a1) is greater than or equal to the cost of the shortest feasible chain in G′ between them (from the first serviced one to the second serviced one), while the cost of the arc in CBi in G ∗ from the vertex coming from a 13 serviced arc in G′ to the vertex coming from the next serviced arc in G′ is equal to the one of the shortest feasible chain in G′ between them. On the other hand, let L = {Ci}ki=1 be a set of k cycles corresponding to a solution to the GVRP in G∗. For each i ∈ {1, . . . , k}, from Ci we construct a feasible closed chain T ′Li in G ′ as follows: Let (xa, xb) be a generic arc of Ci, a = (u, v) and b = (w, r) being the arcs in A ′ R from which vertices xa and xb come, respectively. Arc (xa, xb) will give rise in T ′L i to the s.f.c.(va, wb), and such that T ′Li assumes the demand at a (the same as the one in xa) and b (the same as the one in xb). Note that in this way, T ′L i has the same cost as Ci and satisfies the same demands as Ci. From the set B′L = {T ′Li }ki=1 of k feasible closed chains in G′, we construct now a set BL = {T Li }ki=1 of k feasible closed chains in G as follows: - “Contract” each sequence in T ′Li of the form (u, v e)(ve, vl)(vl, w) with av = (v e, vl) if v ∈ VR1 ∪ {1} by (u, v)(v, w) in T Li . - “Contract” each sequence in T ′Li of the form (u, v e ij)(v e ij, v l ij)(v l ij, w) with avij = (veij, v l ij) if v ∈ VR2 by (u, v)(v, w) in T Li . - If T ′Li traverses a copy of arc e1 or a copy of arc e2 in G ′, with e ∈ E, replace this copy in T Li by edge e. - Any other link or turn in T ′Li is replaced by itself in T L i . - Demand at v ∈ V1 (v ∈ V2) (a ∈ AR) (e ∈ E) is assigned to route T Li iff T ′Li satisfies the demand located at av (one and only one arc avij ) (one and only one copy of arc a in G′) (one and only one copy of e1 or e2 in G ′). It is evident that BL = {T Li }ki=1 is a solution to the MCGRPTP in G with c(BL) = c∗(L), this last due to the fact that for all i, c(T Li ) = c ′(T ′Li ). Let then Lopt be an optimal GVRP solution in G∗ and let BL opt be the MCGRPTP solution in G obtained from Lopt as described above. For each MCGRPTP solution B in G we have: c(B) ≥ c∗(LB) ≥ c∗(Lopt) = c(BLopt ) (1) Therefore, BL opt is an optimal MCGRPTP in G. 14 3.2 Solving the GVRP in G∗ through an ACVRP Once we have transformed our MCGRPTP into a GVRP defined in G∗ = (V ∗, A∗), from G∗ we construct a digraph Ĝ = (V̂ , Â) as follows: - V̂ = V ∗. - For each Si with i ∈ {0, 1, . . . , m} and |Si| > 1, order its vertices consecutively in an arbitrary way {vi1, . . . , vil(i)}; then, for j = 1, . . . , l(i) − 1, define the cost of arc (vij, v i j+1) ∈  as zero; also define the cost of arc (vil(i), vi1) as zero. - For every vij ∈ Si and every w /∈ Si define the cost of arc (vij, w) in  equal to the cost in G∗ of the arc (vij+1, w) ((v i 1, w) if j = l(i)) plus a fixed positive large quantity M if |Si| > 1, and equal to the cost in G∗ of arc (vij, w) plus M if |Si| = 1. - Any other arc in  has infinite cost. - Assign positive demands having sum equal to di to the vertices in Si ∀ i, except for the depot subset. As it was proved by Soler et al (2009), the GVRP can be transformed into an ACVRP. Following their proof, we can see that to solve the GVRP in G∗ we can solve the ACVRP in the digraph Ĝ, and to obtain a GVRP solution from an ACVRP one, we just have to identify each path in the ACVRP solution (vij, v i j+1, . . . , v i l(i), v i 1, . . . , v i j−1, w) w /∈ Si (w can be the depot) if j 6= 1 and |Si| > 1, or (vi1, . . . , vil(i), w) if j = 1 and |Si| > 1, or (vij, w) if |Si| = 1 (in this last case vij can be the depot), with the arc (vij, w) in G ∗. An optimal ACVRP solution Hopt in Ĝ, will give rise to an optimal GVRP solution Lopt in G ∗ with cost c∗(Lopt) = ĉ(Hopt) − M (m + k). Going on with our example, from the graph G∗ where the GVRP is defined (Figure 5) we define the ACVRP in the digraph Ĝ (see Figure 6) where, for simplicity again, the pairs of arcs (xr, xt) and (xt, xr) with xr ∈ Si, xt ∈ Sj and i 6= j have been drawn as lines with two arrow heads, one at each end, and the arc costs (normally different for each direction) have been omitted. Figure 6 shows the cost zero “intraset” arcs and the demand assigned to each vertex. For example, vertex 5, belonging to VR2 and 15 with demand 10 in G, has associated the set S5 in G ∗ with four vertices that in Ĝ have demands 2, 2, 3 and 3, respectively. Figure 6. Directed graph Ĝ associated with G∗. Figure 7 shows the optimal solution H to the ACVRP in Ĝ given in Figure 6 corresponding to our example; it consists of two cycles: - H1 = (xD, xa, xb, xD) with associated cost 7 + 7 + 22 + 3M = 36 + 3M (these arc costs will be deduced above when explaining the solution to the MCGRPTP in G) and the demand served by this cycle is qa + qb = 25 + 25 = 50. - H2 = (xD, x5dc, x5gh, x5f h, x5f c, xc2 , xc1 , xe2 , xe1 , xD) with associated cost: 15 + 0 + 0 + 0 + 0 + 0 + 4 + 0 + 12 + 4M = 31 + 4M and the demand serviced by this cycle is qc + qe + qv5 = 20 + 20 + 2 + 2 + 3 + 3 = 50. Note that the total cost ĉ(H) = 67 + 7M = 36 + 31 + (5 + 2)M since m = 5 is the number of vertex subsets in G∗ (the depot not included) and k = 2 is the number of cycles in the solution. This optimal solution H in Ĝ gives rise to the optimal solution L in G∗ shown in Figure 8, with total cost c∗(L) = ĉ(H) − 7M = 67, and that consists of two cycles: C1 = (xD, xa, xb, xD) with cost 36 and C2 = (xD, x5dc, xc2 , xe2 , xD) with cost 31. 16 Figure 7. Optimal solution to the ACVRP in Ĝ. Figure 8. Optimal solution to the GVRP in G∗. Let us find the optimal solution to the MCGRPTP in G given by these cycles, according to the proof of Theorem 1. From C1 = (xD, xa, xb, xD) we have: - (xD, xa) corresponds to the s.f.c.(1 la1 , 3a) in G′: (1e, 1l)(1l, 3)[(1l, 3), (3, 4)] with cost 0+0+4+3=7 (in the addition, normal-font numbers indicate arc costs while bold numbers indicate turn penalties). See Figure 4 to follow the chains in G′. - (xa, xb) corresponds to the s.f.c.(4 a, 4b) in G′: (3, 4)[(3, 4), (4, 2)] with cost 4+3=7. 17 - (xb, xD) corresponds to the s.f.c.(2 b, 1e a1 ) in G′: (4, 2)(2, 5ef h)(5 e f h5 l f h)(5 l f h, 4)(4, 1 e) [(4, 1e), (1e, 1l)] with cost 4+1+4+1+0+0+5+1+6+0=22. Therefore, joining the chains we have the following feasible closed chain in G′ T ′1 = { (1e, 1l), (1l, 3), (3, 4), (4, 2), (2, 5ef h), (5 e f h, 5 l f h), (5 l f h, 4), (4, 1 e), [(4, 1e), (1e, 1l)] } with cost 36, and assuming the demands of arcs (3, 4) and (4, 2) (a and b), with a total demand of 50. From T ′1, following again with the proof of Theorem 1, we construct the feasible closed chain with the same cost as T ′1 in the original graph G, T1 = {(1, 3), (3, 4), (4, 2), (2, 5)(5, 4), (4, 1), [(4, 1), (1, 3)]} that also assumes the de- mands of arcs a and b (25+25). Similarly, from C2 = (xD, x5dc, xc2 , xe2 , xD) we have: - (xD, x5dc) corresponds to the s.f.c.(1 la1 , 5 ea5dc dc ) in G ′: (1e, 1l)(1l, 4)(4, 5edc)[(4, 5 e dc), (5 e dc, 5 l dc)] with cost 0+0+6+3+5+1=15. - (x5dc, xc2 ) corresponds to the s.f.c.(5 la5dc dc , 5 lc2 dc ) in G ′: (5edc, 5 l dc)[(5 e dc, 5 l dc), (5 l dc, 6)] with cost 0+0=0. - (xc2 , xe2 ) corresponds to the s.f.c.(6 c2 , 6e2 ) in G′: (5ldc, 6)[(5 l dc, 6), (6, 4)] with cost 3+1=4. - (xe2 , xD) corresponds to the s.f.c.(4 e2 , 1e a1 ) in G′: (6, 4)(4, 1e)[(4, 1e), (1e, 1l)] with cost 5+1+6+0=12. Therefore, joining the chains we have the following feasible tour in G′, T ′2 = { (1e, 1l), (1l, 4), (4, 5edc), (5 e dc, 5 l dc), (5 l dc, 6), (6, 4), (4, 1 e), [(4, 1e), (1e, 1l)] } with cost 31, and assuming the demands of arcs (5edc, 5 l dc), (5 l dc, 6) and (6, 4), with a total demand of 50. From T ′2, we construct the feasible closed chain with the same cost as T ′ 2 in G, T2 = {(1, 4)(4, 5)(5, 6)(6, 4)(4, 1)[(4, 1), (1, 4)]}, that assumes the demands of vertex 5 belonging to VR2 , of arc (5, 6) and of edge (4, 6) (10+20+20). Figure 9 shows the optimal solution to the MCGRPTP in the original graph G: T1 with solid bold line and T2 with broken bold line. 18 Figure 9. Optimal solution to the MCGRPTP in G. 4 Computational experiments The aim of this section is to show that the transformation presented here can be considered as a good tool to solve MCGRPTP instances, at least heuristically due to the complexity of the problem. That is, if there exists any competitive procedure to solve the ACVRP, we can solve MCGRPTP instances within a reasonable running time. To do this, we first present a powerful heuristic algorithm for the ACVRP based on the memetic algorithm (MA) for the (symmetric) CVRP proposed by Nagata and Bräysy (2009). MA is a population-based heuristic search approach that combines evolutionary algorithm with local search algorithm. Although there are other heuristic approaches for the ACVRP, as those cited in the introduction, we selected the above mentioned MA because it is shown to be currently the most powerful heuristic method for the CVRP, and it can be applied to the ACVRP by a straightforward extension. Here the suggested MA has been tested on a set of 32 ACVRP instances by Pessoa et al (2007). We have also applied the MA to a set of 126 single vehicle instances by Soler et al (2008) because the optimal solutions are known to these instances. Finally, the MA has been applied to a set of 336 ACVRP instances with up to 623 vertices that come from the transformation of MCGRPTP instances. 4.1 The MA for the ACVRP The main feature of the MA by Nagata and Bräysy (2009) is that the edge assem- bly crossover (EAX) operator generates offspring solutions by combining edges of two 19 solutions selected as parents from the population. The generated offspring solutions may violate the capacity constraint. In this case, a subsequent local search-based re- pair procedure is used to restore the feasibility of the temporarily infeasible solutions. Moreover, a simple local search is applied to the obtained feasible solutions according to a standard MA procedure. Note that the EAX was adapted to the ACVRP by defining it on the directed graph whereas the original EAX for the symmetric CVRP was defined on the undirected graph. The MA by Nagata and Bräysy minimized the total travel distance without putting any constraint on the number of vehicles and the number of vehicle was also a decision variable. However, according to the literature and therefore to the definition of the ACVRP given here, in the ACVRP the travel distance must be minimized with a given number of vehicles. So we have made a new version of the MA that minimizes the total travel distance for a fixed number of vehicles. The suggested MA has been implemented in C++ and has been executed on a ADM Opteron 2.4 GHz computer. For each instance, the MA has been executed five times. 4.2 Results for the ACVRP instances We first analyze the efficiency of the two MA versions (fixed or variable number of vehicles) on a set of 32 ACVRP instances with known upper bounds that appear in the work by Pessoa et al (2007), which are variants of the 8 benchmark ACVRP instances given in http://or.ingce.unibo.it/research/cvrp-and-dcvrp. As far as we know, the work by Pessoa et al is the most recent paper with computational results on the ACVRP. The results for the 32 ACVRP instances without constraint on the number of vehi- cles are presented in Table 1. The columns in the table list instance names (Instance), the capacity of the vehicles (C), the number of vehicles and the total travel distance of the best-known upper bound solutions (k and U B), the number of vehicles and the total travel distance of the best result in five runs (best-k and best-d.) with our MA, the average number of vehicles and the average total travel distance in these five runs (ave-k and ave-d.), and the average computation time in seconds for a run (Time). 20 Table 1. Computational results on known ACVRP instances. Instance C k Best U B Best − k Best − d Ave − k Ave − d T ime a034-14f 150 14 4046 14 4046 14.0 4046.0 1.37 a036-18f 150 18 5296 19 5224 19.0 5224.0 1.21 a039-20f 150 20 5903 20 5903 20.0 5903.0 1.47 a045-18f 150 18 6399 19 5564 19.0 5568.0 4.59 a048-16f 150 16 4955 16 4955 16.6 4960.0 6.35 a056-17f 150 17 4998 17 4998 17.6 5020.0 8.82 a065-19f 150 19 6014 20 5862 20.0 5862.0 12.52 a071-17f 150 17 5006 17 5006 17.0 5014.0 13.86 a034-08f 250 8 2672 8 2672 8.0 2672.0 1.94 a036-10f 250 10 3338 11 3294 11.0 3294.0 1.67 a039-12f 250 12 3705 12 3705 12.0 3705.0 2.63 a045-11f 250 11 3544 11 3544 11.2 3555.6 5.41 a048-10f 250 10 3325 10 3325 10.0 3325.6 4.32 a056-10f 250 10 3263 10 3263 10.0 3263.0 7.46 a065-12f 250 12 3902 12 3902 12.0 3902.0 10.64 a071-10f 250 10 3486 10 3486 10.0 3486.6 12.73 a034-04f 500 4 1773 4 1773 4.0 1773.0 1.76 a036-05f 500 5 2110 5 2110 5.0 2110.0 1.95 a039-06f 500 6 2289 6 2289 6.0 2289.0 2.00 a045-06f 500 6 2303 6 2303 6.0 2303.0 3.41 a048-05f 500 5 2283 5 2283 5.0 2283.0 3.26 a056-05f 500 5 2165 5 2165 5.0 2165.0 4.70 a065-06f 500 6 2567 6 2567 6.0 2568.0 7.82 a071-05f 500 5 2475 5 2457 5.0 2457.0 7.62 a034-02f 1000 2 1406 2 1406 2.0 1406.0 1.31 a036-03f 1000 3 1644 3 1644 3.0 1644.0 1.13 a039-03f 1000 3 1654 3 1654 3.0 1654.0 1.17 a045-03f 1000 3 1740 3 1740 3.0 1740.0 1.97 a048-03f 1000 3 1891 3 1891 3.0 1891.0 1.69 a056-03f 1000 3 1739 3 1739 3.0 1739.0 3.23 a065-03f 1000 3 1974 3 1974 3.0 1974.0 4.17 a071-03f 1000 3 2054 3 2054 3.0 2054.0 4.24 21 We can see that in four instances out of the 32 instances the MA has improved the best known upper bound by using one more vehicle, and in the other 28 instances the solution given by the MA coincides with the best known solution both in the total travel distance and in the number of vehicles. We also have run the MA with fixing a priori the number of vehicles equal to the one given in the best known solution (see third column in Table 1). In this case, in all of the 32 instances the total travel distance obtained with the MA coincides with the best known upper bound, with running time similar to the one given in Table 1. Therefore, the table corresponding to these results has been omitted. In Table 2 we show the results to the set of 126 single vehicle instances given by Soler et al (2008) for which the optimal solution is known. In this table, results are averaged over selected groups of the instances where the 126 instances are partitioned into 24 groups according to features of the instances. The columns in the table lists the number of instances in each groups (Ins), the average number of required arcs in the subset (ARA), the number of (required) edges in all the instances in the subset (|E|), the average number of required vertices in the subset (ARV ), the average number of vertices in the asymmetric TSP instances obtained from the original instances (AVAT SP ), the average time in seconds to obtain the optimal solution with the exact procedure (AT O), the average time in seconds to obtain the heuristic solution with the MA (AT M A), the number of instances optimally solved with the MA in the subset (OP T ), and the average deviation of the MA solutions in the subset (ADEV ). By deviation we mean (U B − LB)/LB. 22 Table 2. Computational results on known single vehicle instances. Group Ins ARA |E| ARV AVAT SP AT O AT M A OP T ADEV 1 6 51 0 6 82 14.48 5,76 6 0 2 4 84 0 8 127 65.04 11,04 4 0 3 6 108 0 8 149 392,81 16,20 6 0 4 5 125 0 12 195 348,34 27,89 4 0,00010 5 7 145 0 14 212 932,43 26,87 6 0,00050 6 6 164 0 16 248 1269,71 41,29 5 0,00025 7 5 54 0 3 86 42.65 7,53 4 0,00078 8 4 84 10 8 137 478,8 12,73 4 0 9 7 107 10 9 175 490,38 21,33 6 0,00071 10 6 126 10 11 201 819,55 29,76 1 0,00059 11 5 141 10 13 231 1273,89 41,44 4 0,00052 12 6 156 10 14 245 1037,18 50,19 5 0,00007 13 6 172 10 15 265 3211,59 61,19 2 0,00093 14 5 61 10 3 119 139,92 7,18 5 0 15 6 105 20 5 167 1464,62 21,52 3 0,00049 16 4 126 20 9 208 516,69 40,62 1 0,00141 17 6 142 20 11 240 3095,38 43,53 2 0,00142 18 4 156 20 9 236 3539,15 42,90 3 0,00089 19 3 168 20 12 236 3646,25 66,26 1 0,00130 20 4 58 20 3 129 846,88 16,77 2 0,00088 21 6 100 20 5 187 1218,35 22,66 4 0,00029 22 6 65 30 3 156 2437,78 22,63 3 0,00034 23 6 113 30 6 218 2931 37,87 1 0,00087 24 3 140 30 6 191 > 2h. 64,47 3 0 We can see from Table 2 that the results obtained with the MA are very good. The MA was able to optimally solve 86 instances out of the 126 instances, with 0.05% average deviation. 4.3 Solving MCGRPTP instances with the MA We applied the MA to solve also three sets of random MCGRPTP instances with up to 700 arcs, 160 (required) edges, 160 vertices, 225 required arcs, and 28 required 23 vertices, which are transformed into ACVRP instances with up to 623 vertices. As far as we know, the largest ACVRP instance described in the literature until now has 300 vertices. Next we describe the data generation procedure for each set. The first set was generated from the 128 single vehicle instances given by Soler et al (2008), and we obtain two MCGRPTP instance sets (one with vehicle capacity 1000 and the other with vehicle capacity 2000) as follows: We choose the depot node as the first required vertex in the numerical order. In this node all turns are changed to be allowed with zero cost. Each required arc, required edge and required vertex will have a randomly generated integer demand in range [13,120], [7,60], and [50,120] respectively. Note that when we transform one of these MCGRPTP instances into an ACVRP instance, if an original required vertex v has demand dv and it gives rise to a vertex subset in the ACVRP with t vertices, let r be the largest integer positive number such that dv = rt + c, with c ≥ 0, then t − 1 vertices in this subset will have demand r, and the last one will have demand r + c. These 256 generated MCGRPTP instances act as a basis for ACVRP instances with number of vertices in the interval [61,290]. The second set contains 32 MCGRPTP instances generated from 16 of the biggest single vehicle instances by Corberán et al (2002). Each original instance is first trans- formed into an MGRP with turn penalties instance with the procedure explained in Soler et al (2008) and then transformed into two MCGRPTP instances (one with ca- pacity 1000 and the other with capacity 2000) with the procedure given above for the first set. These 32 instances form the ACVRP instances with number of vertices in the interval [235,406]. Finally, the third set contains 48 instances that have been obtained from 24 new large single vehicle instances randomly generated with the same instance generator used by Corberán et al (2002). These 48 MCGRPTP instances are transformed into ACVRP instances with the number of vertices in the interval [382,623]. In these three sets we have used the MA version with variable number of vehicles 24 in order to obtain best upper bounds with respect to the total travel distance. The appendix shows a table containing all data and results corresponding to each individual MCGRPTP instances. In this table, each instance is named as Ixxy, where xx indicates the subset to which the instance belongs and y indicates the number of the instance inside that subset. The first 24 subsets correspond to the first set of (128) instances, subsets 25 to 27 correspond to the second set of (16) instances and subsets 28 to 31 correspond to the third set of (24) instances. The columns in the table list the following data: the number of vertices in the ACVRP instance obtained from the corresponding MCGRPTP instance (|VACV RP |), the total demand in the ACVRP instance (T.D.), and for i ∈ {1, 2}, the total travel distance of the best result in five runs obtained for a vehicle capacity of Ci = i·1000 units (di), the number of vehicles corresponding to this best result (Ki) and the average computing time in seconds for a run corresponding to the capacity Ci (Timei). Note that di is the total distance in the original MCGRPTP instance, not in the auxiliary ACVRP instance. The MA was able to find feasible solutions for all 336 instances including the large- size instances with up to 600 vertices, within a reasonable computation time, as re- ported in the appendix. The computation times vary from a few seconds to more than one hour depending on the problem size. We consider the reported computing times reasonable, given the size and complexity of the considered ACVRP instances. Based on these results, one can conclude that at least medium-size real-world MC- GRPTPs can be solved by a state-of-the-art heuristic method for the ACVRP through their transformation into an ACVRP as explained here. By the way, we have gener- ated a large number of instances to the, until now, limited set of ACVRP benchmarck instances. Of course these new instances will be available to any researcher interested on them. 5 Conclusions In this paper we have studied a generalization of the MCGRP including turn penalties and forbidden turns. Through an intermediate transformation into a GVRP, we have provided a procedure to transform it into an ACVRP. Then, at least from a theoretical 25 point of view, this generalization can be solved both optimally and heuristically with existing algorithms. We have also introduced a set of new benchmark problems and adapted a recent and powerful memetic algorithm to ACVRP. The experimental results show an average deviation equal to 0.05% for instances with known optimal solution and that large-size problems can be solved with the memetic algorithm. We are convinced that research on turn penalties will increase and be of important value in the future to reduce the gap between theoretical research and real-life appli- cations. We hope that the theoretical and experimental results presented here can be used in the future as ideas or tools to test the efficiency of specific procedures to solve capacitated routing problems with turn penalties. Acknowledgements This work has been partially supported by the Ministerio de Educación y Ciencia of Spain (project TIN2008-06441-C02-01). 26 Appendix Inst. |V |-|A|-|E| |AR| |VR| |VACV RP | T.D. d1 K1 Time1 d2 K2 Time2 I11 40-110-0 44 3 61 3819 4255 4 13.11 3966 2 4.42 I12 60-140-0 56 5 75 6462 6848 5 25.63 6145 3 17.9 I13 40-90-0 36 3 46 3009 4103 4 11.04 3547 2 19.67 I14 80-170-0 68 6 92 5415 7578 6 25.15 6806 3 20.9 I15 40-90-0 36 6 54 3142 4126 4 7.56 3625 2 8.38 I16 80-170-0 68 11 118 5772 7651 6 65.18 6798 3 37.18 I21 80-200-0 80 7 107 6936 7464 7 53.7 6962 3 35.13 I22 100-220-0 88 7 121 7067 10152 8 41.1 8999 4 28.14 I23 80-200-0 80 4 95 5815 7300 6 53.97 7026 4 28.3 I24 100-220-0 88 14 158 8907 10643 9 79.62 9439 5 54.04 I31 100-260-0 104 3 112 6704 11831 7 63.81 11272 4 24.74 I32 120-260-0 104 6 127 8055 13666 11 256.67 12364 6 112.48 I33 120-290-0 116 7 147 8946 13493 9 228.25 12332 5 59.84 I34 100-260-0 104 6 136 8154 12352 9 62.52 11466 5 47.86 I35 120-260-0 104 11 147 8387 12264 9 43 11130 5 41.33 I36 120-290-0 116 14 195 10603 14253 11 248.94 12728 6 102.05 I41 140-300-0 120 8 151 9108 15107 10 93.41 13086 5 69.21 I42 140-330-0 132 12 216 10775 16681 11 320.16 15176 6 114.81 I43 140-330-0 132 6 164 10232 16542 11 139.86 15126 6 65.39 I44 140-300-0 120 20 220 11314 16130 12 214.2 13732 6 156.12 I45 140-300-0 120 14 185 9746 15101 10 204.5 13184 5 122.74 I51 160-340-0 136 8 166 10460 14698 11 156.37 13806 7 224.51 I52 160-340-0 136 19 219 8387 15947 14 284.86 13405 6 69 I53 160-340-0 136 13 191 11192 14987 12 183.21 13413 6 117.01 I54 180-380-0 152 8 183 10750 18322 11 219.97 16710 6 92.44 I55 160-370-0 148 14 218 12729 17818 13 387.24 15959 7 151.67 I56 180-380-0 152 21 250 13252 19431 14 422.96 17000 7 236.26 I57 180-380-0 152 15 217 12690 18806 13 369.88 16687 7 117.98 I61 200-420-0 168 10 210 12868 21039 13 431.41 18289 7 131.69 I62 180-400-0 160 8 193 12364 19634 13 227.68 17122 7 91.19 I63 200-420-0 168 17 289 15961 25569 16 1608.23 19197 9 348.06 I64 180-400-0 160 20 255 14227 20806 15 522.98 17808 8 214.22 I65 180-400-0 160 14 217 13445 20295 14 392.63 17340 7 180.87 I66 200-420-0 168 18 254 14903 22681 15 669.9 18661 8 235.08 I71 40-110-10 44 1 65 3612 4830 4 12.93 4528 2 11.29 I72 40-90-10 36 3 67 3501 4353 4 14.94 4151 2 16.83 I73 60-140-10 56 2 81 4290 6101 5 16.09 5808 3 16.85 I74 60-160-10 64 2 89 5011 7201 6 28.32 6859 3 30.66 I75 80-170-10 68 4 107 6462 7924 7 45.76 7248 4 31.74 I81 80-200-10 80 4 113 6703 7403 7 62.7 6993 4 30.44 I82 100-220-10 88 5 113 6824 7409 7 86.16 6993 4 34.07 I83 80-200-10 80 7 133 7279 7654 8 51.37 7034 4 47.82 I84 100-220-10 88 10 184 9826 9568 9 109.2 8820 5 47.64 27 Inst. |V |-|A|-|E| |AR| |VR| |VACV RP | T.D. d1 K1 Time1 d2 K2 Time2 I91 120-260-10 104 6 149 8587 13401 11 157.11 12181 6 78.18 I92 100-260-10 104 5 157 8386 12508 9 209.1 11584 5 81 I93 120-290-10 116 5 158 8700 12457 9 145.08 11757 5 47.07 I94 120-260-10 104 16 204 10556 12734 9 111.9 11894 5 60.81 I95 100-260-10 104 10 184 9826 13780 10 248.3 12055 5 129.32 I96 120-290-10 116 10 183 9539 12771 10 182.84 11934 5 106.03 I97 120-260-10 104 11 177 10132 12703 9 95.41 11877 5 61.5 I101 140-300-10 120 6 169 9675 14174 10 190.19 12609 5 121.77 I102 140-330-10 132 6 179 10520 16205 11 195.82 15124 6 64.7 I103 140-330-10 132 11 203 11670 16628 12 330.09 15468 6 172.42 I104 140-300-10 120 16 215 11522 15189 12 314.17 13089 6 169.71 I105 140-330-10 132 15 243 13154 17056 14 389.77 15618 7 177.68 I106 140-300-10 120 11 193 10832 14880 11 274.03 12822 6 102.74 I111 160-340-10 136 9 193 11457 16445 12 273.48 14640 6 172.66 I112 160-340-10 136 23 275 13274 17141 14 484.04 14918 7 330.28 I113 160-340-10 136 16 230 12651 16616 13 403.95 14690 7 152.03 I114 160-370-10 148 6 195 12448 18102 13 247.19 16335 7 113.22 I115 160-370-10 148 12 230 13856 19240 14 586 16446 7 323.42 I121 180-380-10 152 8 203 11981 20136 12 782.32 17250 8 397.97 I122 180-380-10 152 20 277 15076 20133 16 530.39 16547 6 406.85 I123 180-380-10 152 14 234 12662 18966 13 500.68 16671 7 211.59 I124 180-400-10 160 13 241 13650 18237 14 411.86 16348 7 265.88 I125 180-400-10 160 8 219 12867 18404 13 431.07 16303 7 169.24 I126 180-400-10 160 19 276 15023 18594 16 415.15 16647 8 308.28 I131 200-420-10 168 8 234 14073 20264 13 454.48 18009 7 153.91 I132 200-440-10 176 8 223 12780 21636 15 308.71 19289 8 140.3 I133 200-420-10 168 22 288 15931 23868 16 1173.59 19076 8 661.17 I134 200-440-10 176 21 290 15052 21369 16 557.87 18686 8 343.22 I135 200-420-10 168 15 257 14872 21877 15 689.08 18479 8 278.81 I136 200-440-10 176 15 275 15877 24410 16 790.1 20049 8 522.07 I141 60-90-20 36 1 77 3699 4084 4 20.51 3833 2 21.02 I142 60-160-20 64 5 115 5659 8403 7 52.74 7629 4 30.23 I143 60-140-20 56 5 129 6107 7284 6 64.19 6522 3 42.57 I144 80-170-20 68 5 125 6565 8181 7 82.09 7743 4 52.77 I145 80-200-20 80 1 121 6542 7986 7 58.58 7620 4 49.68 I151 100-220-20 88 4 141 8387 10444 8 195.27 9800 4 106.95 I152 120-260-20 104 4 157 8391 12665 9 102.45 11959 5 76.46 I153 120-290-20 116 3 169 10033 13550 11 156.71 12303 6 76.88 I154 100-260-20 104 3 153 8387 13614 11 189.79 12502 6 98.08 I155 120-260-20 104 7 171 9857 13125 10 299.63 12032 5 159.41 I156 120-290-20 116 6 183 9777 13596 10 257.16 12317 5 172.63 I161 140-300-20 120 6 184 10131 15037 11 133.85 13473 6 77.48 I162 140-330-20 132 6 196 10861 16840 11 304.8 14957 6 137.01 I163 140-300-20 120 11 208 11277 15554 12 239.37 13585 6 136.5 I164 140-330-20 132 11 235 12807 17620 13 497.14 15329 7 175.13 28 Inst. |V |-|A|-|E| |AR| |VR| |VACV RP | T.D. d1 K1 Time1 d2 K2 Time2 I171 160-340-20 136 16 249 13059 16078 14 418.03 14551 7 261.08 I172 160-340-20 136 6 197 11250 15202 12 251 14189 6 154.1 I173 160-370-20 148 6 214 12080 18298 13 237.53 16272 7 144.29 I174 160-340-20 136 11 224 11824 15763 12 388.2 14244 6 325.71 I175 160-370-20 148 11 245 13134 18765 14 355.11 16396 7 248.01 I176 160-370-20 148 16 288 14216 19257 15 500.41 16827 8 247.63 I181 180-380-20 152 5 212 11643 18628 12 332.17 16663 6 246.78 I182 180-380-20 152 9 230 12888 19989 13 756.61 17339 7 231.97 I183 180-400-20 160 7 229 13028 18463 14 303.72 16648 7 190.11 I184 180-400-20 160 14 267 13991 18851 15 1343.62 18193 7 914.38 I191 200-420-20 168 7 237 13896 22059 14 703.29 19778 8 802.64 I192 200-420-20 168 17 289 15961 25569.00 16 1608.23 18782 7 406.29 I193 200-420-20 168 12 267 14877 23137 15 907.11 19184 8 251.88 I201 40-110-30 44 1 105 5266 6069 6 40.75 5444 3 38.24 I202 60-140-30 56 2 123 5777 6705 6 96.64 6359 3 54.63 I203 60-160-30 64 2 129 6846 8571 7 154.47 7915 4 80.08 I204 80-170-30 68 5 144 7021 8487 8 66.81 7896 4 74.44 I211 80-200-30 80 3 149 7597 8704 8 131.29 7943 4 88.27 I212 80-200-30 80 6 163 7744 8818 8 198.14 8056 4 103.63 I213 100-260-30 104 1 192 10061 12822 11 284.98 11984 6 143.23 I214 120-260-30 104 5 183 8387 12670 10 161.83 11516 5 122.43 I215 120-290-30 116 6 199 10056 13642 11 166.43 12382 6 116.95 I216 120-290-30 116 11 233 12343 14384 13 356.58 12711 7 152.71 I221 40-90-40 36 1 117 5416 5321 6 59.06 4962 3 56.52 I222 60-140-40 56 3 147 7012 7742 8 88.71 7093 4 81.97 I223 60-160-40 64 2 149 7524 8892 8 150.32 8225 4 90.57 I224 80-170-40 68 1 149 7651 8818 8 129.22 8134 4 82.15 I225 100-220-40 88 5 214 10793 11285 11 368.81 10027 6 160.24 I226 80-200-40 80 2 165 7986 9410 8 509.77 8253 4 275.54 I231 120-260-40 104 3 207 10567 14740 11 403.89 13088 6 145.29 I232 120-290-40 116 3 207 10678 14281 11 430.29 13110 6 150.35 I233 120-290-40 116 6 221 11151 14642 12 369.99 13282 6 178.76 I234 100-260-40 104 5 208 10307 12485 9 88.39 11297 5 50.78 I235 140-300-40 120 5 216 11258 15690 12 327.21 14373 6 202.45 I236 140-300-40 120 10 242 7021 16012 13 460.42 14576 7 207.78 I241 140-330-40 132 5 231 11865 17444 12 606.84 15789 6 372.98 I242 160-340-40 136 6 241 12797 17189 13 687.33 15339 7 207.94 I243 160-370-40 148 5 249 13426 18010 14 552.3 16245 7 365.67 I244 160-340-40 136 12 271 13973 20563 14 1586.52 16225 7 904.7 I245 160-370-40 148 10 275 14247 18365 15 566.35 16602 8 239.05 I251 120-260-60 104 3 235 10952 16098 11 836.73 13467 6 287.68 I252 140-330-60 132 7 283 13898 18238 14 1108.62 15406 7 799.45 I253 160-370-60 148 6 293 14768 20488 15 1189.57 18175 8 476.2 I254 180-400-60 160 8 315 16339 21517 17 1184.31 19329 9 522.9 I256 200-420-60 168 17 361 19015 24841 20 1212.79 21615 10 1023.6 29 Inst. |V |-|A|-|E| |AR| |VR| |VACV RP | T.D. d1 K1 Time1 d2 K2 Time2 I261 100-220-80 88 2 253 10764 12723 11 643.9 11708 6 424.9 I262 120-260-80 104 3 273 12387 15312 13 529.4 14127 7 285.35 I263 140-330-80 132 2 299 14801 18469 15 1370.78 16110 8 552.23 I264 160-370-80 148 11 359 16695 22098 17 1434.42 18966 9 651.6 I265 180-400-80 160 10 361 18031 22641 19 1089.58 19829 10 793.31 I266 200-420-80 168 9 363 18409 23206 19 1685.57 20366 10 1026.85 I271 100-220-100 88 2 293 12705 12708 13 912.82 11754 7 461.76 I272 120-260-100 104 3 315 14161 15851 15 784.83 14099 8 490.66 I273 140-330-100 132 5 353 16462 18388 17 1280.67 16269 9 1086.19 I274 180-400-100 160 9 406 19041 23053 20 1927.5 20180 10 1489.17 I275 200-420-100 168 8 399 19914 29504 20 4160.04 23017 10 2203.86 I281 200-600-60 180 9 382 23593 25022 18 4449.39 20696 9 1819.96 I282 200-600-80 180 13 452 19844 26271 20 5636.98 223670 10 3160.30 I283 200-600-100 180 6 440 20324 24640 21 2681.01 21966 11 1802.93 I284 200-600-120 180 8 463 20555 26060 21 3841.40 23267 11 2136.15 I285 200-600-140 180 7 497 21577 27645 22 5756.68 24330 11 3655.87 I286 200-600-160 180 7 548 25148 25999 26 5966.49 23482 13 3879.16 I291 225-650-60 195 19 439 21355 27266 22 2306.38 23901 11 2337.94 I292 225-650-80 195 15 470 22727 29510 23 4450.06 248090 12 2900.06 I293 225-650-100 195 12 494 23987 28549 25 3187.86 25247 12 5064.86 I294 225-650-120 195 11 507 23593 31508 24 4747.81 27241 12 4546.69 I295 225-650-140 195 6 502 23593 30096 24 4434.57 26147 12 5327.40 I296 225-650-160 195 5 538 24429 30333 25 5490.39 267850 13 5215.72 I301 250-700-60 210 24 484 22346 28475 23 3718.08 247730 12 1945.99 I302 250-700-80 210 19 484 22746 32558 23 4312.41 27064 12 2631.71 I303 250-700-100 210 18 521 24925 29632 26 5427.05 25985 13 3103.63 I304 250-700-120 210 7 490 23071 30786 24 3635.21 27128 12 2674.20 I305 250-700-140 210 11 563 24875 374830 25 10734.52 297940 13 4286.82 I306 250-700-160 210 11 604 27643 34824 28 9304.83 29962 14 6554.05 I311 270-750-60 225 25 503 23948 30713 25 3003.68 29216 12 5038.99 I312 270-750-80 225 28 599 27111 32940 28 4606.29 28394 14 5215.76 I313 270-750-100 225 16 543 26907 33859 28 5026.30 29348 14 3865.43 I314 270-750-120 225 13 548 25044 30692 26 4782.07 27320 13 4545.20 I315 270-750-140 225 14 602 27722 36708 28 9063.87 306540 14 7840.02 I316 270-750-160 225 9 623 27771 39925 28 11808.50 32859 14 9424.87 References Bautista, J., & Pereira, J. (2004). Ant Algorithms for Urban Waste Collection Routing. Lecture Notes in Computer Science, 3172, 3020-3033. Bautista, J., Fernández, E., & Pereira, J. (2008). Solving an Urban Waste Collec- tion Problem Using Ants Heuristics. Computers & Operations Research, 35, 302-309. 30 Belenguer, J.M., Benavent, E., Lacomme, P., & Prins, C. (2006). Lower and Upper Bounds for the Mixed Capacitated Arc Routing Problem. Computers & Operations Research, 33, 3363-3383. Benavent, E., & Soler, D. (1999). The Directed Rural Postman Problem with Turn Penalties. Transportation Science 33, 408-418. Bodin, L., Fagin, G., Welebny, R., & Greenberg, J. (1989). The Design of a Com- puterized Sanitation Vehicle Routing and Scheduling for the Town of Oyster Bay, New York. Computers & Operations Research, 16, 45-54. Clossey, J., Laporte, G., & Soriano, P. (2001). Solving Arc Routing Problems with Turn Penalties. Journal of the Operational Research Society, 52, 433-439. Corberán, A., Mart́ı, R., Mart́ınez, E., & Soler, D. (2002). The Rural Postman Problem on Mixed Graphs with Turn Penalties. Computers & Operations Research, 29, 887-903. De Franceschi, R., Fischetti, M., & Toth, P. (2006). A New ILP-Based Refinement Heuristic for Vehicle Routing Problems. Mathematical Programming Series B, 105, 471-499. Fischetti, M., Toth, P., & Vigo, D. (1994). A Branch-and-Bound Algorithm for the Capacitated Vehicle Routing Problem on Directed Graphs. Operations Research, 42, 846-859. Ghiani, G., & Improta, G. (2000). An Efficient Transformation of the Generalized Vehicle Routing Problem. European Journal of Operational Research, 122, 11-17. Jansen, K. (1993). Bounds for the General Capacitated Routing Problem. Net- works, 23, 165-173. Nagata, Y., & Bräysy, O. (2009). Edge Assembly Based Memetic Algorithm for the Capacitated Vehicle Routing Problem. Networks, 54, 205-215. Pandit, R., & Muralidharan, B. (1995). A Capacitated General Routing Problem on Mixed Networks. Computers & Operations Research, 22, 465-478. Perrier, N., Langevin, A., & Amaya, C.V. (2008). Vehicle Routing for Urban Snow Plowing Operations. Transportation Science, 42, 44-56. Pessoa, A., Poggi, M., & Uchoa,E. (2007). Robust Branch-Cut-and-Price Algo- rithms for Vehicle Routing Problems. Presented at Route 2007 (International Work- 31 shop on Vehicle Routing and Transportation), Jekyll Island, Georgia. Roy, S., & Rousseau, J.M. (1989). The Capacitated Canadian Postman Problem. INFOR, 27, 58-73. Soler, D., Mart́ınez, E., & Micó, J.C. (2008). A Transformation for the Mixed General Routing Problem with Turn Penalties. Journal of the Operational Research Society, 59, 540-547. Soler, D., Albiach, J., & Mart́ınez, E. (2009). A Way to Optimally Solve a Time- Dependent Vehicle Routing Problem with Time Windows. Operations Research Letters, 37, 37-42. Vigo, D. (1996). A Heuristic Algorithm for the Asymmetric Capacitated Vehicle Routing Problem. European Journal of Operational Research, 89, 108-126. 32 
work_o22yfbkr7bgoxaqh7pzz2assqa	----	Regional Appraisal of Hydrocarbon Potential of Trans-Pecos Texas: Methodology and Conclusions: ABSTRACT; Expert Systems in Seismic Exploration: ABSTRACT; Geotectonic Evolution of Bering Sea Area, Alaska: ABSTRACT; Geology and Petroleum Potential of ... Association Round Table 249 eral is monocrystalline quartz and the matrix components are ferroan cal- cite, siderite, and kaolinite. Porosity appears primarily secondary in nature. Core analysis has shown 23% average porosity and 210 md per- meability. Oil-base cores indicate an irreducible water saturation of 44% and residual oil saturation of 22%. In the area of study, estimated reserves are 22.27 million m ' of oil in place. The produced oil has a density o f 8 4 4 k g / m , a viscosity of3,3 cp,anda gravity of 37° API. The produced oil is undersaturated in relation to gas and the initial reservoir pressure was 3,960 kPa. The absence of a gas cap and an active aquifer has resulted in implementation of a secondary- recovery waterflood mechanism based on a 5-spot injection pattern. Pri- mary recovery is estimated to be 10%, with an additional 31% from waterflood. DEJONG, H. W., SUNIT K. ADDY, GEORGE W. WHITNEY, and RALPH E. WORTHINGTON, ARCO Exploration Co., Denver, CO Regional Appraisal of Hydrocarbon Potential of Trans-Pecos Texas: Methodology and Conclusions Evaluation of large areas requires a different approach from that used to develop prospects. Regional cross-section networks generally ignore minor correlation problems. We examined and related all potentially sig- nificant parameters, and used reduced scales to subdue nonessentials and to cover a large area. Organization and planning must permit free inter- change of ideas and close cooperation among those working on the pro- ject. Regional structural and stratigraphic analysis of Trans-Pecos Texas strongly suggests that widespread trap-destruction by faulting and ero- sion led to hydrocarbon leakage and induction of fresh water into pro- spective zones, which reduced the likelihood of accumulation and preservation of economic reserves. At least 4 major periods of tectonism, pervasive fresh water, mineral- ization, basic igneous intrusives and extrusives, high heat-flow regime, increasing percentage of carbon dioxide southwestward from the Dela- ware basin, and contemporaneous vertical movement suggest that the opportunity for major reserves is minimal. A review of 1,000 mi of CDP seismic data from 4 areas disclosed 65 structural leads or traps, of which 45 had significant tests. All these tests were failures; many gave indications of sizable amounts of fresh water. None of the areas have been completely condemned. However, only relatively small reserves can be anticipated for nearshore to continental Jurassic and Cretaceous sections in the Chihuahua Trough, Pennsylva- nian and Permian reefs formed on a shelf environment or on the flanks of major uplifts, or in areas such as the Marfa basin. DENHAM, L. R., Seiscom Delta United Corp., Houston, TX Expert Systems in Seismic Exploration Artificial intelligence research has produced few practical results in most of its branches. However, "expert systems" in limited fields of expertise are potentially practical and cost-effective tools in many fields of exploration geophysics. Recent breakthroughs, such as writing expert systems in languages less exotic than Lisp, have made it possible to install a practical expert system on even the smallest computer. A recently pub- lished expert system written in Forth compiles a rule base into very com- pact code, and then uses it to reach decisions based on data supplied by the user. Such a system makes it possible for a small computer to be the geophysicist's advisor on many different subjects, because one expert sys- tem can use any number of rule bases. The expert system then becomes a practical tool for standardizing the decision-making process, even in comparatively trivial areas. DESAUTELS, DAVID A., Elf Aquitaine Petroleum, Houston, TX Geotectonic Evolution of Bering Sea Area, Alaska The geologic, structural, and tectonic history of the Bering Sea area since Paleozoic time is best viewed in terms of major plate-tectonic inter- actions. The geotectonic style of disparate areas is apparently related to the nature of plate motion at the time of tectonic imprint. Three major structural belts that have existed since the Mesozoic can be traced from the Siberian .sector across the Bering Sea and into Alaska. The northern belt, the Verkhoyansk-Chukotsk-Seward-Brooks, consists of miogeosyn- clinal sediments that were deposited beginning in earliest Mesozoic time. The middle belt, the Okhotsk-Chukotsk-Yukon-Koyukuk, consists of a Mesozoic magmatic arc and numerous allochthonous terranes, formed due to the convergence-subduction of a southern oceanic plate. The southern belt, the Koryak-Anadyr-Peninsular, consists of terranes accreted during Cretaceous time and forms the southern hmit of Meso- zoic subduction. During Late Cretaceous to early Tertiary time, these belts were oro- clinally bent southward by an east-west compressional event, causing the subduction zone to shift to a more southerly location, thus forming the current Aleutian Island arc system, behind which the fragments of 2 Cre- taceous oceanic plates were "trapped." These oceanic plate fragments may consist of an Early Cretaceous plate and a portion of the Kula plate(?), which carried a northward-migrating arc system. The hypothe- sized Early Cretaceous plate may have had a counterpart separated by a spreading ridge, both of which have been subducted beneath the Berin- gian margin. DESPRETZ, JEAN-MARIE, Petrofina, Brussels, Belgium, THOMAS E. DALY, REMI, Inc., Englewood, CO, and EDWARD ROBINSON*, Consultant, Kingwood, TX Geology and Petroleum Potential of Saba Bank Area, Northeastern Car- ibbean Recent exploratory activity on Saba Bank in the northeastern Carib- bean has provided geologic information showing that this frontier area possesses all of the attributes necessary for the commercial accumulation of hydrocarbons. The first well drilled in the area penetrated 9,370 ft (2,856 m) of sediments including 3,021 ft (921 m) of Eocene carbonates containing zones of good to excellent porosity. Geochemical studies show the presence of good but immature source rocks with the extractable hydrocarbons being migrated rather than indigenous, The geothermal gradient and vitrinite reflectance data indicated the threshold of the oil window would be reached around 10,000 ft (3,048 m). The second well was drilled to test a postulated reef on a basement high at a sufficient depth to fall within the oil window. The well bottomed in Eocene andesite at 13,881 ft (4,231 m). Reef carbonate was not encountered; the well pen- etrated turbidite sandstones and siltstones with low to moderate porosity and permeability. A test of gas shows recovered small amounts of C1-C5, but the formation is believed to have been badly damaged by severe mud loss during drilling. Geochemical studies confirm the presence of good source rocks. Reworked unmetamorphosed organic matter of probable early Eocene to Cretaceous age suggests that the Cretaceous cannot be considered economic basement in this area. Reinterpretation of the seis- mic data explains why the two wells were dry and indicates the presence of a submarine fan area, reefs within the oil window, and large structures in an area of thick sediments of probable Cretaceous age. DICKEY, RODGER L., Independent, Waco, TX Depositional and Diagenetic History of Buda Limestone in Central Texas and Its Relationship to Petroleum Potential The upper Comanchean Buda Limestone (Cretaceous) is a known res- ervoir for hydrocarbons in central Texas, producing from depths as shal- low as 700 ft. Understanding the character of the Buda Limestone and its complex depositional and diagenetic history is essential to developing a sound exploration strategy and to insure maximum production. In central Texas, the Buda Limestone may be divided into a lower, micritic facies, and a dense, in places dolomitized, intrasparite upper fades. The upper Buda is more porous and contains most of the produc- ible hydrocarbons in the formation. The upper contact is an undulating erosional surface, unconformably overlain by impermeable Woodbine shales. Porosity enhancement appears greater in areas of faulting and fractur- ing, especially where occurring along erosional drainage divides. Because of the apparent correlations between favorable structure and marketable oil production, economic prospecting methods should seek to delineate zones of faulting and fracturing along areas where the upper Buda was exposed to weathering. 
work_o3c5ngi7nrherncep3tg7nhv4u	----	doi:10.1016/j.eswa.2005.07.034 XKey: A tool for the generation of identification keys M. Delgado Calvo-Flores a , W. Fajardo Contreras a , E.L. Gibaja Galindo b,*, R. Pérez-Pérez a a E.T.S. de Ingenierı́a Informática, Departamento de Ciencias de la Computación e Inteligencia Artificial, Universidad de Granada, C/Periodista Daniel Saucedo Aranda, 18071 Granada, Spain b Escuela Politécnica Superior, Departamento de Informática y Análisis Numérico, Universidad de Córdoba, Campus de Rabanales, Edificio Albert Einstein, 3a planta, 14071 Córdoba, Spain Abstract This paper presents the development of XKey, a tool for generating taxonomical identification keys by means of decision tree construction. The tool is based on an XML standard for the representation of general taxonomical information, which makes it ideal for different fields of application. The article analyses the problem by examining the adaptation of machine learning techniques to the sphere of biology so as to incorporate the viewpoints of biologists and computer science experts. It also analyses the effect of using various division criteria on a set of real data: the Gymnosperm plant groups present in the Iberian peninsula. q 2005 Elsevier Ltd. All rights reserved. Keywords: Dichotomous keys; Decision trees; XML; Division criteria; Interactivity 1. Introduction Identification keys are fundamental in the study of biodiversity. This term, which arises from concern for the environment, and the need for conservation and the sustainable use of natural resources, is defined as the variability of living organisms on earth and includes all organizational levels: from the simplest of individuals to ecosystems (even the entire planet). The first step in its study is to identify the types of organisms present in the biotype being analyzed (in the field, the researcher finds an organism, but in a database register, its identification is stored). This identifi- cation (or determination) consists in recognizing characters in a sample, so as to give it a name which was previously given to a similar organism. Given the age of taxonomy as a science, the most used identification tool in the last 200 years is the printed key; a tool which must also be included in field guides and floras. In general, experts design their keys manually, without the use of any computer support tool. The development of a key is 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.07.034 * Corresponding author. E-mail addresses: mdelgado@decsai.ugr.es (M. Delgado Calvo- Flores), egibaja@uco.es (W. Fajardo Contreras), egibaja@uco.es (E.L. Gibaja Galindo). therefore time-consuming, and also extremely costly when an error is detected: the later the error is detected, the more it will cost to rectify it. Since the computer tools developed to support this process have been suggested by specialist biologists, we find that considerable improvements could be made. This article presents XKey, a tool for generating identification keys, and is organized in the following way: Section 2 details the background of the problem, focusing on the aspects related to the representation of taxonomical knowledge and the tools for the generation of previously designed keys; Section 3 analyses the problem and describes the division criteria used by the tool; Section 4 examines the characteristics of XKey and its contribution to its field of application; Section 5 outlines the test procedure and the results obtained; and finally, Section 6 presents a series of conclusions and the references consulted to produce this work. 2. Background 2.1. Identification keys An identification key is a tree-shaped structure which presents the user with a series of choices at each step. We can distinguish between: † Printed keys. Most printed keys have a treelike structure (see Fig. 1). They are very powerful tools when dealing Expert Systems with Applications 30 (2006) 337–351 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa Single leaves Leaves which are green on both sizes Leaves which are tomentous on the back Myrtus communis (myrtle) Olea europaea (wild olive) Composite leaves Trifoliate leaves Pinnate-composite leaves Jasminum fructicans (jasmine) Pistacia lentiscus (lentisk) Fig. 1. Example of an identification key (Morales, Quesada, & Baena, 2001). M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351338 with primary identification characters (Fortuner, 1989), or rather characters which are capable of differentiating between species or groups of species with a very small risk of error. They proceed by elimination, and, for this reason, identification is prevented as they do not have the necessary information to go on to the next level. In order to avoid this, multi-entry or tabular keys are used whereby the elimination of species or groups depends on several characters. The problem of this type of key is that it is difficult to use with large genera. † Computerized keys. This type of key is characterized by its interactivity: a computer enables a multi-entry key to be easily used. Within this group, we should also include the graphic key, in which the choices are presented in image format to represent the options. It is not possible to say that one type of key is better than another; this aspect will depend on the sphere of its use. We shall focus our study on the development of tree-like printed keys, which are widely used and whose development is not particularly automated. Fig. 2. General structure of an SDD document 1.0. In TDWG-SDD (2005). 2.2. Digital representation of taxonomical knowledge The automatic generation of identification keys requires highly structured descriptive information to be used. In spite of having described certain models such as Lif (UBio, 2003), Delta (Dallwitz, 1974; Dallwitz, Paine, & Zurcher, 2000) and Nexus (Maddison, Swofford, & Maddison, 1997), there is no widely accepted and used standard model, and this prevents a greater use of digital taxonomical infor- mation and leads to substantial inefficiency for taxonomy as a whole. At the present moment in time, information globalization is being pursued so that it may be used by different authors and with different ends (identification, key production, floras, field guides, etc.). Delta has been standard of the IUBIS-TDWG since 1991, and in September 1998, after analyzing the new challenges for the scientific community, the subgroup SDD was set up to develop an XML-based standard for representing and handling descrip- tive information of organisms. The aim is to design a sufficiently standardized knowledge representation model which is accepted by the entire community of biologists, whatever their discipline of origin (zoology, botany, etc.). SDD shall attempt to provide a flexible and independent platform to facilitate the exchange of data sets without loss of information between applications, in addition to using a same description for different purposes. Fig. 2 shows the general structure of an SDD document. 2.3. Tools for the generation of identification keys There are certain tools for generating interactive keys manually. This is the case of LucID (CBIT, 1994), Linnaeus II (Schalk & Heijman, 1996; Schalk & Troost, 1999), Xid (Intelsys, Inc., 2000***–2001), Meka (Duncan & Meacham, 1986; Meacham, 1986–1996), NaviKey (Bartley & Cross, 2000; University of Toronto, Department of Botany, 1996). This type of tool does not give any information about the suitability of the selected characters for branching nor does it check the existence of inconsistencies in the key (unclassified objectives, dead ends, etc.). Pankey (Pankhurst, 1991; Pankhurst & Pullan, 1994) is a commercial system based on Delta and available for MS-Dos which incorporates two tools: Key3m3 (to generate keys totally automatically), and Kconi (to generate keys under the user’s supervision). It is capable of showing the partial key resulting from the choice of characters made by the user and to suggest the best option at each step. We should also mention Delta which is a general system for the field of systematics based on the model with the same name which includes Key, a printed key generation program M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351 339 which uses text files to register the user’s preferences and thereby offer some interactivity. This review reveals the shortage of tools, and also the fact that until recently, they were developed by biologists and have become obsolete. In addition, they incorporate few facilities for interactivity, and can be improved from the computing point of view. Consequently, we have developed XKey, an SDD-based tool which combines proven artificial intelligence techniques with the specific functional needs of taxonomy. It has been designed and implemented within the framework of a multidisciplinary team and from the study of previously developed tools, and therefore represents an integratory vision from the viewpoints of biologists and computer specialists. 3. Analysis of the problem 3.1. Construction of identification keys An identification key is a tree-like structure in which each node represents a character, and the arcs its possible values. Because of its structure, it is constructed by performing recursive partitions of the set of training cases in order to finally obtain a decision tree. In short, when a node is constructed or expanded, the subset of training cases belonging to each class is considered. If all the examples belong to one class, or if some stopping rule is verified, the node is a leaf of the tree. Otherwise, the training set is divided into subsets (using a rule of heuristic division), and the same procedure is applied to each subset of the training set. Due to the special characteristics of biology, an identification key does not totally fit the definition of a decision tree in the classic sense. Below, we shall describe the characteristics of these keys and how they are adapted by our algorithm to generate identification keys, as outlined in Fig. 3. 3.2. Objectives of the division criterion The general aim is to obtain a compact decision tree. In accordance with the principle of Occam’s razor, a small decision tree enables a better understanding of the classification model obtained. This principle, although allowing for the construction of easily understandable models, does not guarantee that these are better than other apparently more complex ones (Domingos, 1998, 1999). From the point of view of taxonomy, a good division rule (Dallwitz et al., 2000): 1. divides the taxa into uniform subgroups; 2. selects characters with high reliability and low intra- taxon variability (we understand intra-taxon variability to be the set of characters which can distinguish between individuals or populations belonging to a same taxonomical category); 3. discriminates the individuals which are going to be identified more frequently in the first steps (this frequency is also called abundance). Although many division criteria have been proposed, we shall focus our study on four of these: the first three because they are classic and widely used, and the fourth because it has been proposed from the area of taxonomy. 3.3. Classic division criteria Within this first group of criteria, we can distinguish: † Entropy (Quinlan, 1986). This measures the impurity of a node of the tree and reaches its minimum when all the elements of the node are the same. It satisfies the requirements established by Dallwitz, i.e. it divides the taxa into more uniform groups and minimizes the intra- taxon variability. It also considers the abundance of an individual, so that the most frequent will appear in the first steps of the key. The entropy normally constructs decision trees with a high degree of branching, which favors questions with the most possible results. This aspect does not always reflect the work methodology of the expert biologist, who, depending on the objective with which the key is developed, may have other criteria which are different from the minimization of the length. † The gain ratio criterion (Quinlan, 1993). This measure- ment is based on the entropy which normalizes the information gain obtained in order to avoid the construc- tion of decision trees which classify the cases using their keys. It has been observed that, like the entropy, it favors partitions of the training set which are very uneven in size when any of them is very pure (all the cases which it includes correspond to the same class) even if it is not particularly significant (covering very few training cases). † The Gini diversity index (Breiman, Friedman, Olshen, & Stone, 1984). This is a measurement for class diversity in a tree node which attempts to minimize the impurity existing in the subsets of training cases generated when the decision tree is branched. 3.4. Division criteria in taxonomy In the area of taxonomy, very few criteria have been described for generating identification keys, and one such example is that used by Dallwitz et al. (2000). This criterion calculates the cost of using a certain character i according to the expression in Formula 1. Since it concerns a cost measurement, the aim is to minimize it, and consequently characters with a low cost will be preferred. where: † s is the number of values for the attribute; † nj is the number of taxa in the jth subgroup; START Are there cases in the training set? Can tree branching continue? Calculate division criterion and select best attribute Do all cases belong to the same class c Generate leaf node labeled class c END No No No Yes Yes Yes For each subset resturn to START Generate intermediate node labeled with th selected attribute Add the confirmatory characters to the node Subdivide the training set into as many subsets as the number of values in the attribute If an example presents null values and the interpretation is: “unknown”, add the example to all subsets “inapplicable”, do not add the example to any subset If several attributes ar candidates, determine their priority: from the information in the data set by asking the user Interactive mode? Ask the user for the attribute and the branching node No .. .. Fig. 3. Algorithm for the generation of identification keys. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351340 † c is the cost of a character. This is related to the reliability using the expression cZRbase5Kr, where: B r is the value ascribed to the reliability of the character (a value fixed by the expert) and takes values in [0–10]. B Rbase is a constant which takes values in the interval [1,5]. † cmin is the lowest cost for the characters being considered; † fj is the total frequency of the items in the jth group. (0) Flexible leaves; Discolorous leaves Yes; Exserted bracts; Color of the branches ash-gray; Pubescent branches .............Species Abies alba (0) Rigid leaves; Discolorous leaves No; Included bracts; Color of the branches reddish-brown; Glabrous branches .............Species Abies pinsapo main character confirmatory characters Fig. 4. Key with confirmatory characters. Author: Eva Gibaja Galindo. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351 341 The frequency f of an item is given by the expression where: B a is the abundance of the item (a value fixed by the expert) and takes values in [0–10] B Abase is a constant which takes values in the interval [1,5]. † V controls the effect of the intra-taxon variability. In its expression: B Varywt is a constant value. If its value is 0, those characters which have any intra-taxon variability are excluded from the key. If, on the other hand, the value of this parameter is 1, this aspect is not penalized. It takes values in the interval [0–1]. B n is the number of taxa. 3.5. Treatment of null values When a key is generated, it is common for null values to appear in the data set. The final result of the process depends on the interpretation made by the algorithm of these. In taxonomy, there could be three different interpretations for the appearance of null values: 1. The value does not appear because the attribute is inapplicable for a certain taxon. In this case, the taxon is not added to any branch of the tree. It should be observed that when this interpretation of the null values is used, the taxon to which the null value corresponds could remain unclassified. 2. The value does not appear because it is indifferent. This means that, for a certain taxon, the attribute can take any valid value. If this character is used for branching the decision tree, it is necessary to add the taxon with the null value to each branch of the tree. 3. The value does not appear because it is unknown. In this case, and in order to avoid loss of information, the same procedure is followed as in the previous case. 3.6. Reliability of a character In taxonomy, not all the characters have the same relevance: some are more important than others, for example, owing to the ease with which they are observed, because they are distinguishing characters, etc. This aspect is called the reliability of a character (Dallwitz, 1974). If the expert provides information about the most reliable character for branching, it is possible to combine the measurement given by a division criterion and expert knowledge about the domain, in order to decide in those cases in which several characters present the same value of the division criterion. In this way, the key will adapt better to the characteristics of each taxonomical group and will be more satisfactory. 3.7. Confirmatory characters It is normal in a key to include more than one character in the same node of the classification tree (see Fig. 4). This means that the node contains a main character and set of characters called confirmatory characters (Dallwitz et al., 2000). This is a fundamental aspect since it enables: † identification to be continued when the main character is not available; † confirmation of the decision to continue identification by a certain node. The main character and the confirmatory characters are equivalent, and any of these can be used to branch the tree. This means that, within the taxa group being considered, the two characters: 1. have the same value for the division criterion; 2. have the same number of possible values; 3. generate identical taxa groups since they are used as the branching character. 3.8. Construction of knowledge bases By taking advantage of the construction of the decision tree, it is possible to generate a knowledge base comprising rules from this. In botany, the classification problem and subsequent identification has singular characteristics: a fold does not, in every case, have the complete individual for its identification, and the samples gathered frequently depend on seasonality, age of the individual, etc. This can lead to a very well-informed consultation not offering conclusions if there is no value for any character. In order to tackle this aspect, various decision trees are constructed (which shall subsequently be converted into rules), taking a different generator node each time. For this reason, it is necessary to include a consistency reinforcer, and since the syntax of a rule is sufficiently restrictive, it systematically analyses each rule in the knowledge base to detect and correct unnecessary if conditions, redundant rules, conflictive rules, and subsumed rules (see Fig. 5). 3.9. Treatment of uncertainty Given the characteristics of taxonomical identification, the application of some method for the treatment of Fig. 5. Algorithm for the XKey consistency reinforcer. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351342 uncertainty is extremely useful in order to enrich knowledge base consultation. There are various theories for treating uncertainty such as the theory of evidence (Dempster, 1967; Shafer, 1976), or the probability or fuzzy set theory (Zadeh, 1965). Due to the absence of historical data and the difficulties of defining fuzzy variables and probability mass functions, we have chosen the theory of certainty factors (Shortliffe & Buchanan, 1975). This model has been used successfully in many other recently published systems (Cabrero-Carnosa, Castro-Pereiro, Graña-Ramos, Hernández-Pereira, Moret-Bonillo and Martı́n-Egaña, 2003; Mahaman, Harizanis, Filis, Antonopoulou, Yialouris and Sideridis, 2002). Furthermore, the theory of certainty factors is a computationally simple and more natural model for a non-mathematical user than other approaches (Harrison & Kovalchic, 1998). Taking all this into consideration, each rule in the knowledge base has an associated certainty factor (CF). This value, which is determined during generation and remains unchanged unless redefined by an expert, is ascribed according to the following criteria: † if a pure leaf node is generated (all the elements belong to the same class), the associated rule has an associated certainty factor with a value of C1; † if the leaf node is not pure, the result is that of a rule with disjunction in the consequent, which is fractionized into rules with simple consequents. The certainty factor of the original rule is also fractionized by the following Formula 2 where † p(x) is the probability of the objective in question occurring in the initial table; † p(x/a) is the probability of objective x occurring knowing that antecedent a occurs, regardless of whether this is simple or compound. 4. Characteristics of XKey 4.1. Use of different division criteria XKey enables the division criterion to be selected, and includes the four criteria presented in Sections 3.3 and 3.4: entropy, the gain ratio criterion, the Gini diversity index, and the division rule proposed by Dallwitz. This allows generated keys to be compared from a same set of data but using different criteria and aids a better adaptation to each specific problem. The criterion is selected from the user options menu (Fig. 6). 4.2. Treatment of null values In its operation mode, XKey reflects the three different interpretations which we can give to the appearance of null values in taxonomy (see Fig. 5) and enables the distinction to be made between: 1. not applicable (the taxon with the null value is not added to any tree branch); 2. indifferent (XKey adds the taxon with the null value to each of the tree branches); 3. unknown (XKey acts as in Case 2). SDD also enables explicit representation of null values. For this, it uses the global states ‘Unknown’ and ‘NotApplicable’. An attribute with the special value ‘Unknown’ will be treated by XKey as unknown (Case 3), whereas an attribute with the special value ‘NotApplicable’ will be treated as inapplicable (Case 1) independently of the general interpretation with which XKey has been configured. 4.3. Control of the reliability of a character The SDD data sets do not include information about the reliability of a character, so when two attributes have Fig. 6. Execution options of XKey. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351 343 the same value for the division criterion, the construction of the tree depends on the order in which the data are represented. In order to avoid this situation, XKey shows the user the different alternatives, the current state of the tree, and the objectives which must still be classified so that he/she may select the attribute which they consider to be the most suitable. It is also possible to ascribe a priority value in Fig. 7. Attribute selection execution time to a specific attribute. This weight will be remembered by the system throughout execution and will be used to decide between several branching attributes with the same value for the division criterion (see Fig. 7). At certain times, the data set can be too large to apply this strategy, or it could be that the expert is not interested in making this type of choice. For this reason, users can tell in execution time. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351344 XKey at any time that they no longer want to be asked about this aspect. 4.4. Inclusion of confirmatory characters In order to adjust to this feature, XKey includes an option in the execution options menu so that users can indicate whether they want to include this type of character and the maximum number of confirmatory characters per tree node. The choice of the value will depend on the user’s criterion and on the final use of the key; the final quantity included will depend on the characteristics of each data set. It is therefore possible to indicate that the inclusion of confirmatory characters is not required. 4.5. Operation modes One difference between the learning of traditional classification models and identification keys is interactivity in character selection. The best character from the point of view of the division criterion may not be the best character from the expert’s point of view: it might be difficult to visualize, at times it is preferable to eliminate the exceptions in the first steps of the key, etc. For these reasons, XKey offers various operation modes: † Automatic mode. This applies the algorithm for the generation of decision trees without the user’s interven- tion and uses the branching criterion selected in the options menu. The algorithm for the automatic gener- ation of XKey matches the general guidelines of a TDIDT algorithm (Top-Down Induction On Decision Trees). What XKey offers is its capacity to add confirmatory characters, to treat null values according to the interpretation established by the user, and to select, in execution time, the best branching attribute (if several attributes have the same value for the division criterion). † Semi-automatic mode. This applies the algorithm for the generation of decision trees automatically, but it consults the user in those cases where there are ties in the selection of the branching attribute. XKey shows the list of all the characters which can be used for branching, ordered according to the division criterion. In this way, it provides very useful statistical information which the expert does not have access to when generating keys manually with other tools. † Interactive mode (supervised by the user). This follows the schema for automatic generation, to which inter- action facilities are added. The operations permitted during execution in interactive mode are: † Add node. The user selects the node to branch and the branching attribute, and generates the subtree corresponding to this node. XKey again shows the list of all the characters which can be used for branching, ordered according to the division criterion. † Delete node. In this case, the node and all its descendants are eliminated. For this, it is necessary to maintain an execution memory for each node so that it may return to previous states. † End. This option enables the user to establish where certain characters will appear so that the system can subsequently finish the execution automatically. When this operation is selected, XKey analyzes the classification model in order to detect the leaf nodes which do not contain classes of the example (intermediary nodes) so as to finish construction of these nodes automatically. 4.6. Diversity of output formats Once the identification key has been generated, it is presented to the user in tree form (see Fig. 8). It is possible to save this key in the formats described below: † Text format. The key is saved in a plain text file. This format facilitates the modification of the key with other more sophisticated text editors, and in turn, does not limit the use of any of these. † XML format. The key is saved in an XML format file, with a similar structure to the expected format of the SDD keys. This format enables the keys to be edited once they have been generated and for information to be easily exchanged. 4.7. Construction of a knowledge base XKey incorporates the generation of a knowledge base into its functionality from the data set as described in Sections 3.8 and 3.9. If generation of a knowledge base is chosen, this can be saved in: † CLIPS format (Giarratano, 1998). The key is saved in the form of rules for this well-known expert system shell, with which we can intercommunicate the standard XML model for taxonomical description with one of the most popular shells within the field of expert systems. † GREEN format (Fajardo, Gibaja, Bailon, & Moral, 2003). We have mentioned that as well as generating keys, XKey can generate sets of rules which can be used directly by the GREEN system. 4.8. Execution results In addition to the keys, the execution of XKey returns a set of measurements which facilitate analysis of the results obtained. This information includes: † average length of the key; † typical deviation of the length of the key (in order to compare some keys with others, the system returns Pearson’s skewness coefficient); Fig. 8. Interactive selection of attributes with XKey. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351 345 † maximum and minimum length of the key; † number of terminal and non-terminal nodes; † number of OR terminal and exclusive nodes (the appearance of OR nodes indicates that it is not possible to clearly identify a particular node, and, therefore, that the data set is incomplete); † number of total attributes in the data set and number of attributes used in the key; † number of confirmatory attributes included in the key. Table 1 Subgroups of Gymnospermae being studied Division Gymnospermae Family Araucariaceae Family Cephalotaxaceae Family Cupressaceae Family Cycadaceae Family Ephedraceae Family Ginkgoaceae Family Pinaceae Family Taxaceae Family Taxodiaceae Genus Abies Genus Calocedrus Genus Cedrus Genus Chamaecyparis Genus Cryptomeria Genus Cupressus Genus Cycas Genus Ephedra Genus Juniperus Genus Larix Genus Picea Genus Pinus Genus Platycladus Genus Pseudotsuga Genus Sequoia Genus Sequoiadendron Genus Taxodium Genus Tetraclinis 5. Test of the XKey tool 5.1. Execution data for the XKey tool and methodology followed The information provided by XKey enables the keys obtained for the same data set to be compared by using, for example, different division criteria. In our case, experiments have been carried out with a real group: Gymnosperms present in the Iberian peninsula. The evaluation of the keys generated with XKey for this group offers us a sufficiently wide set of tests for determining what criteria, in what cases or conditions, and why they enable the most ideal key to be generated. It should be noted that in order to increase the diversity of casuistries which can be presented, not only have wild or naturalized taxa been included, but also ones which have been cultivated and which are widely used in gardening. The subgroups of Gymnosperms used are described in Table 1. In order to carry out this experiment, a knowledge acquisition process has been performed with the partici- pation of two expert botanists. Bibliographical sources have also been consulted, and in particular (López Gonźlez, 2001; López Gonźlez & Do Amaral, 1986), which are essential reference books for the taxonomical group being studied. In our study, we have analyzed a set of qualitative aspects which are directly related to the reliability of the generated key: 0 5 10 15 20 25 30 35 40 G im n o sp e rm a s P in u s Ju n ip e ru s C u p re ss a ce a e P in a ce a e T a xo d ia ce a e C u p re ss u s E p h e d ra A b ie s C yc a s C e d ru s Entropy Gain Gini Dallwitz Expert Average 0 10 20 30 40 50 60 70 p e rm a s P in u s n ip e ru s ss a ce a e in a ce a e d ia ce a e p re ss u s p h e d ra A b ie s C yc a s C e d ru s Entropy Gain Gini Dallwitz Expert Average A B M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351346 † Average length of the key. This is an indicator of the average number of steps for identification. This is an important aspect since the expert often searches from the key which leads to the identification in the least number of steps. † Typical deviation of the average length. This is an indicator of the equilibrium of the keys: if the typical deviation is large, the different paths of the tree have very different lengths. † Terminal nodes/number of objectives ratio. This is an indicator of the degree of branching of the keys: if there are many more leaf nodes than objectives to be classified, the tree is extremely branched. † Internal nodes/confirmatory characters ratio. This is an indicator of the power of generating confirmatory characters of a certain division criterion. Those criteria which identify a greater quantity of confirmatory characters are best. † Characters used/total number of characters ratio. With this, it is possible to determine the ratio of attributes in the data set which have been used to generate the key and what these attributes are. † In addition to the quantitative interpretation of the results, another important aspect has been the analysis of the adaptation to the reality of the keys obtained. G im n o s Ju C u p re P T a xo C u E Fig. 10. Number of external (A) and internal nodes (B). 5.2. Observations about the average length of the keys generated From the expert botanist’s point of view, the common factor is the inadequacy of the character selection of the keys when the Gini diversity index is used as the division criterion. This criterion always produces longer keys (see Fig. 9) and trees with a greater number of internal and external nodes (see Fig. 10). If we consider the sum of differences in relation to the minimum average length (Fig. 11), the keys generated with the entropy criterion are practically in all the cases of minimum length. In second place is the profit gain ratio, followed by Dallwitz’s criterion, which does not produce as good results due to the fact that the measurement it uses for 0 1 2 3 4 5 6 G ym n o sp e rm a e P in u s Ju n ip e ru s C u p re ss a ce a e P in a ce a e T a xo d ia ce a e C u p re ss u s E p h e d ra A b ie s C yc a s C e d ru s Entropy Gain Gini Dallwitz Expert Minimum Fig. 9. Average length of the keys. counting the intra-taxon variability is cruder than that used by the two previous criteria: Dallwitz only considers the number of different taxa in each partition, whereas the other two also count the frequency of each taxon. 5.3. Observations about the equilibrium of the keys The deviation of the length of the branches in relation to the average is always greater for the Gini measure, which indicates that the length of the tree branches generated with 0 1 2 3 4 5 6 7 8 Entropy Gain Gini Dallwitz Fig. 11. Sum of the differences in relation to the minimum average length. 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 G ym n o sp e rm a e P in u s Ju n ip e ru s C u p re ss a ce a e P in a ce a e T a xo d ia ce a e C u p re ss u s E p h e d ra A b ie s C yc a s C e d ru s Entropy Gain Gini Dallwitz Expert Minimum Fig. 12. Typical deviation of the average length. 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 G ym n o sp e rm a e P in u s Ju n ip e ru s C u p re ss a ce a e P in a ce a e T a xo d ia ce a e C u p re ss u s E p h e d ra A b ie s C yc a s C e d ru s Entropy Gain Gini Dallwitz Minimum Fig. 14. Terminal/objective ratio. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351 347 this criterion is more variable. The most balanced keys are those generated with the gain ratio criterion, followed by those generated with entropy and Dallwitz’s division criterion (Fig. 12). The sum of the differences in relation to the minimum typical deviation (Fig. 13) is less for the gain ratio, followed by entropy. 5.4. Observations regarding the degree of branching The gain ratio and entropy criteria generate the least branched key, i.e. they tend to detect the smallest number of possible paths for a certain objective and, therefore, to generate more compact models than Gini and Dallwitz’s criteria (Figs. 14 and 15). 5.5. Observations about the number of confirmatory characters included The entropy criterion produces a better number of confirmatory attributes/number of internal nodes ratio, followed by the profit gain criterion (see Fig. 16). If we observe the differences regarding the maximum of the confirmatory/internal nodes ratio in each case, we can see 0 0.5 1 1.5 2 Entropy Gain Gini Dallwitz Fig. 13. Sum of the differences in relation to the minimum typical deviation. how the entropy is the one which presents a smaller difference, followed by the profit gain criterion and, some way behind, by the division criteria of Dallwitz and Gini (see Fig. 17). 5.6. Observations about the number of characters used The entropy uses the least number of characters in the generation of keys, i.e. it locates the smallest amount of information in order to carry out identification. In second place is the gain ratio criterion, followed by Dallwitz’s criterion. The Gini diversity index, in addition to producing longer keys, includes a greater number of attributes in the generation of keys (Figs. 18 and 19). 5.7. Qualitative comparison of the criteria The tests carried out show that entropy and gain ratio offer the best results for all the measures proposed. We can see a summary in Table 2: 4 is the score given to the best criterion, 1 is the score for the worst criterion, and 2 and 3 are the scores for intermediary cases. Table 2 shows that it is precisely entropy and profit gain which are the best criteria. The measure used by Dallwitz also produces good results, but in some unwanted cases since it measures the intra-taxon variability more crudely. The difference with Dallwitz’s division criterion can be clearly seen in the case 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 Entropy Gain Gini Dallwitz Fig. 15. Sum of the differences in relation to the minimum terminal/objec- tive ratio. 0 0.5 1 1.5 2 2.5 G ym n o sp e rm a e P in u s Ju n ip e ru s C u p re ss a ce a e P in a ce a e T a xo d ia ce a e C u p re ss u s E p h e d ra A b ie s C yc a s C e d ru s Entropy Gain Gini Dallwitz Expert Maximum Fig. 16. Number of confirmatory characters/internal node. 0 10 20 30 40 50 60 70 80 90 100 G ym n o sp e rm a e P in u s Ju n ip e ru s C u p re ss a ce a e P in a ce a e T a xo d ia ce a e C u p re ss u s E p h e d ra A b ie s C yc a s C e d ru s Entropy Gain Gini Dallwitz Minimum Fig. 18. Ratio of total number of characters/characters used. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351348 of the keys of the genus Cedrus: in order to separate two taxa, Dallwitz gives five paths for C. atlantica and three for C. deodara (Table 3). With the entropy, we have two classes of Cedrus and one path to arrive at each species (Table 4). 5.8. Discussion and biological interpretation of the results obtained/effect of interactivity The entropy produces more optimized keys in terms of the number of questions, but in the field of biology, this aspect must be qualified. At times, there is no reason why the distinguishing character from the point of view of information theory need be the best character from the biological point of view, due, for example, to the difficulty in its observation (need for microscopic observation, temporality, size, layout, etc.). This is a fundamental aspect as the identification process is interrupted if no character is available. It is therefore necessary for those characters which as well as having a high distinguishing power can be easily observed to be included in the first steps of the key. All this entails considering not only the division criteria but also the characteristics of the taxonomical group being studied and the future receptors of the keys. The possibility of choosing between the four criteria set out in our article enables different types of keys to be 0 0.2 0.4 0.6 0.8 1 1.2 Entropy Gain Gini Dallwitz Fig. 17. Sum of the differences in relation to the maximum number of confirmatory characters/internal node. produced according to the type of work, research, and end user: † If the key is going to be all the information about the taxa which shall be available in the end, we shall need long keys which gather the greatest number of possible characters. † If the keys are the tool to access information and, once the taxon has been identified, this is accompanied by an exhaustive description, we can use more precise keys with a high level of discrimination. † The generation of keys for experts, which assumes an advanced knowledge of the group and of the plant morphology and characteristics, prefers criteria such as that of the entropy. † The case of considering keys as a tool for an intermediary user would probably entail using the ideal criterion according to the information theory, but combined with interactivity. It might be that the attributes on which the entropy has generated its key are not the ones which an expert in the field would consider to be the most suitable. In these cases, the combination of interactivity with a division criterion 0 20 40 60 80 100 120 Entropy Gain Gini Dallwitz Fig. 19. Sum of the differences in relation to the minimum of the total number of characters/characters used ratio. Table 2 Comparison of different division criteria on the Gymnospermae group Length Balance Branching Confirmatory No. of characters Total Entropy 4 3 3 4 4 18 Profit 3 4 4 3 3 17 Gini 1 1 1 1 1 5 Dallwitz 2 2 2 2 2 10 M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351 349 such as the entropy produces very satisfactory results: the entropy suggests and orders the characters, but it is the expert who ultimately decides which character to select by using his/her knowledge about the specific taxonomical group. In the case of the keys for Gymnosperm families, we have managed to reduce the average length of the key in relation to the entropy by means of the use of expert knowledge (see Fig. 20). A similar case occurs with the keys of the genus Pinus. In this case, the character in which the key is based is ‘characteristics of the apophysis’, which rapidly separates the different species of pine trees. It is not easy, however, to see this character as it refers to the pine cone, a character which is not always present and which supposes knowledge of the plant morphology which is not within the reach of any user. According to the experts’ opinion, a good attribute is ‘leaf shape’. The interactivity can be used on other levels. Continuing with the example of Pinus, the second attribute was made to be the ‘presence/absence of pine nuts’. The final result is shown in Table 5, and this became the key which best responded to the expert’s expectations (Fig. 21). 6. Conclusions and final notes Throughout this article, we have observed that the results can be better adapted to the biological reality by adding Table 3 Key for the genus Cedrus generated according to Dallwitz’s criterion Genus Cedrus /Dallwitz (0) Size of the leaves (long) between 2 and 2.5 cm 1 (0) Size of the pine cone (long) between 8 and 12 cm Species Cedrus deodara (0) Size of the pine cone (long) between 4 and 6 cm Species Cedrus atlantica (1) Size of the pine cone (long) between 6 and 8 cm Species Cedrus atlantica (0) Size of the leaves (long) between 2.5 and 3 cm 2 (2) Size of the pine cone (long) between 8 and 12 cm Species Cedrus deodara (2) Size of the pine cone (long) between 4 and 6 cm Species Cedrus atlantica (2) Size of the pine cone (long) between 6 and 8 cm Species Cedrus atlantica (0) Size of the leaves (long) between 3 and 5 cm Species Cedrus deodara (0) Size of the leaves (long) between 1 and 2 cm Species Cedrus atlantica meta-knowledge. For example, by ascribing a reliability value for the characters, it is possible to decide in those cases in which the system does not have the necessary information to make a decision. This alternative has its disadvantages: firstly, all this information must be gathered in the knowledge representation model and this makes design more complicated; secondly, the large number of characters normally being dealt with; and thirdly, all the additional meta-knowledge must be entered by the expert developing the data set. It is necessary to find a balance between the quantity of the additional information and the functionality of the tools. We should also highlight that the adaptation of the results also depends on the characteristics of the data set and on the specific problem being tackled. A good selection of identification characters will result in keys which are much more satisfactory and better adapted to reality (Fig. 22). We shall end this description of our work with a set of conclusions: 1. We have developed XKey, a tool for generating identification keys which operate directly from data sets developed with SDD. XKey can also operate with descriptions in DeltaAccess and Delta format (by means of the use of translation utilities). 2. We have shown the adaptation of artificial intelligence techniques for automatic learning and treatment of uncertainty to the area of taxonomical identification. The expert’s satisfaction with the keys generated by XKey endorse the suitability of the techniques used. 3. The output of the tool is presented in various formats: text format, XML format, and CLIPS format, and is supplemented with statistical information which facili- tates the study and comparison of the results obtained. 4. XKey generates identification keys rapidly and con- veniently, which means a considerable saving in time for the expert. It is also versatile, since it enables different division criteria to be selected, the meaning of null Table 4 Key for the genus Cedrus generated according to minimum entropy criterion Genus Cedrus/entropy (0) Size 6000 cm; Tree guide: recurved; Hanging branches: Yes Species Cedrus deodara (0) Size 5000 cm; Tree guide: not recurved; Hanging branches: Yes Species Cedrus atlantica 0 1 2 3 4 5 6 Gymnospermae Pinus Entropy Gain Gini Dallwitz Expert 1 Expert 2 Average Fig. 20. Average length of the keys for the Gymnospermae division and the genus Pinus. Fig. 21. Division criterion proposed by Dallwitz. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351350 values to be configured, distinguishing characters to be included, and weights to be ascribed to the characters in execution time. This added functionality produces keys with a more suitable biological meaning than those generated with classic decision trees. 5. In addition to generating identification keys, Xkey also generates complete, consistent knowledge bases which are compatible with the CLIPS system for subsequent use in expert systems. Table 5 Key for the genus Pinus generated according to the expert’s criterion Genus Pinus/expert criterion (0) Number of leaves per fascicle 2 1 (1) With pine nut: No; Crown shape: pyramidal; Persistent, winged seed: Yes 2 (2) Characteristics of the apophysis: prominent and sharp Species Pinus pinaster (2) Characteristics of the apophysis: not very prominent 3 (3) Colour of the bark (rhytidome): Ash-grey; Shiny pine cones: Yes; Leaf colour: intense green; Leaves: flexible Species Pinus nigra subsp. Salzmannii (3) Colour of the bark (rhytidome): Reddish-brown; Shiny pine cones: No; Leaf colour: light green; Leaves: rigid Species Pinus sylvestris (2) Characteristics of the apophysis: very prominent, hooked Species Pinus uncinata (2) Characteristics of the apophysis: not very convex Species Pinus halepensis (2) Characteristics of the apophysis: very prominent and sharp Species Pinus radiata (1) With pine nut: Yes; Crown shape: umbrella-shaped; Persistent, winged seed: No Species Pinus pinea (0) Number of leaves per fascicle 3 4 (4) Size 2500 cm; Characteristics of the apophysis: very prominent, hooked Species Pinus uncinata (4) Size 4000 cm; Characteristics of the apophysis: very prominent and sharp Species Pinus radiata (4) Size 6000 cm; Characteristics of the apophysis: prominent Species Pinus canariensis 6. The adaptation of the key to reality has been shown to depend to a large extent on the attributes selected for its branching. In view of various equivalent branching alternatives, XKey stops its execution and resorts to the user’s criterion. This semi-automatic mode of execution can be deactivated in order to operate totally automati- cally. 7. In addition to the automatic and semi-automatic execution modes, XKey has been provided with the capability of generating keys interactively. In this way, the user can select what node to branch and what attribute to use at any given moment; it is also possible to eliminate nodes from the tree, and to change from interactive to automatic mode at any time so as to finish the key generation. 8. In order to help the user, XKey presents the available characters in each step, ordered according to the division criterion. This aspect is particularly important as it enables the discrimination capacity of division rules (such as the entropy) to be combined with the human expert’s criterion and for the keys obtained to have a biological content which is more satisfactory to the expert. 9. Finally, a comparative study has been carried out of the effect of several division criteria on the generation of dichotomic keys with a sufficiently complex, real taxonomical group: Iberian Gymnosperms. This study reveals that the division criterion of the entropy is the one which produces the shortest keys and which favors the inclusion of a greater number of distinguishing characters. The gain ratio criterion generates somewhat longer keys, but in return, they are more balanced (the length of the paths is less variable). These two criteria detect the most important information for the classifi- cation of a set of taxa. Dallwitz’s criterion does not offer as good results in terms of the average length of the key, power of generating confirmatory characters, etc. Fig. 22. Calculation of the certainty factor of a rule with OR in the consequent. M. Delgado Calvo-Flores et al. / Expert Systems with Applications 30 (2006) 337–351 351 In spite of this, it can be useful in cases where the objective is not to minimize the length of the key. The same happens with the Gini diversity index, with the observation that this last criterion is not advisable with large data sets since it produces excessively complicated keys. References Bartley, M., & Cross, N. (2000). Navikey 2.0 (En linea, Consulta: 03/06/2003, http://www.huh.harvard.edu/databases/legacy/navikey/ index.html). Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and regression trees. California: Wadsworth. Cabrero-Carnosa, M., Castro-Pereiro, M., Graña-Ramos, M., Hernández- Pereira, E., Moret-Bonillo, V., Martı́n-Egaña, M., et al. (2003). An intelligent system for the detection and interpretation of sleep apneas. Expert Systems with Applications, 24, 335–349. CBIT (Centre for Biological Information Technology, University of Queensland) (1994). LucID version 2.1 (En lı́nea. Consulta: 25/05/2003. http://www.lucidcentral.com/default.htm). Dempster, A. (1967). Upper and lower probabilities induced by a multi- valued mapping. Annals of Mathematical Statistics, 38(2), 325–399. Dallwitz, M. J. (1974). A flexible computer program for generating identification keys. Systematic Zoology, 1, 50–57. Dallwitz, M. J., Paine, T. A., & Zurcher, E. J. (2000). User’s guide to the DELTA system: A general system for processing taxonomic descrip- tions, 4.12 edition. Canberra: CSIRO Division of Entomology (En lı́nea, Consulta: 25/05/2003, http://biodiversity.uno.edu/delta/). Domingos, P. (1998). Occams’s two razors: The sharp and the blunt Proceedings of the fourth international conference of knowledge discovery and data mining (KDD-98), August 27–31, New York City, USA pp. 34–37. Domingos, P. (1999). The role of occams’s raxor in knowledge disvorvery. Data Mining and Knowledge Discovery Volume, 3, 409–425. Duncan, T., & Meacham, C. A. (1986). Multiple-entry-keys for the identification of angiosperm families using a microcomputer. Taxon, 35, 492–494. Fajardo, W., Gibaja, E., Bailon, A., & Moral, P. (2003). An application of expert systems to botanical taxonomy. Expert Systems with Appli- cations, 25/3, 425–430 (Elsevier/2003). Fortuner, R. (1989). Nematode identification and expert system technology. New York: Plenum Publishing Corp.. Giarratano, J. E. (1998). CLIPS user’s guide. NASA, Artificial Intelligence Section, L.B. Johnson Space Center. Harrison, P. R., & Kovalchik, J. G. (1998). Expert systems and uncertainty. In J. Liebowitz (Ed.), Handbook of applied expert systems (pp. 1–11). West Palm Beach, FL: CRC Press. Intelsys Inc. (2000–2001). XID authoring system 3.0 (Demo Version) (En lı́nea, Consulta: 3/06/2003, http://www.xidservices.com). López González, G. (2001). Los árboles y arbustos de la Penı́nsula Ibérica e islas Baleares Tomo I. Madrid: Ediciones Multiprensa. 861 pp. López González, G., & Do Amaral Franco, J. (1986). In S. Castroviejo, et al., Gymnospermae. Flora Ibérica (Vol. I) (pp. 161–195). Madrid: Real Jardı́n Botánico, CSIC. Maddison, D. R., Swofford, D. L., & Maddison, W. P. (1997). NEXUS: An extensible file format for systematic information. Systematic Biology, 46. Mahaman, B. D., Harizanis, P., Filis, I., Antonopoulou, E., Yialouris, C. P., & Sideridis, A. B. (2002). A diagnostic expert system for honeybee pests. Computers and electronics in agriculture, 36, 17–31. Meacham, C. A. (1986–1996). Meka version 3.0 (En lı́nea, Consulta: 3/06/2003, http://ucjeps.berkeley.edu/meacham/meka/). Morales, C., Quesada, C., & Baena, L. (2001) Guı́as de la Naturaleza. Árboles y Arbustos. Diputación de Granada, Granada. Pankhurst, R. J. (1991). Practical taxonomic computing. Cambridge: Cambridge University Press. Pankhurst, R. J., & Pullan, M. (1994). Pandora user guide version 3.1 (En lı́nea, Consulta:8/06/2003, http://www.ibiblio.org/pub/academic/ biology/ecologyCevolution/software/pandora). Quinlan, J. R. (1986). Induction on decision trees. Machine Learning, 1, 81–106. Quinlan, J. R. (1993). C4.5: Programs for machine learning. Los Altos, CA: Morgan Kaufmann1-55860-238-0. Shafer, G. (1976). Mathematical theory of evidence. Princeton, NJ: Princeton University Press. Schalk, P. H., & Heijman, R. P. (1996). ETI’s linnaeus II taxonomic software: A new tool for interactive education, Vol. 3. UniSer- ve.Science News, University of Sydney (En lı́nea, Consulta 7/5/2003, http://www.eti.uva.nl/Home/Articles.html). Schalk, P. H., & Troost, D. G. (1999). Computer tools for accessing biodiversity information. Nature and Resources, 35(3), 31–38. Shortlife, E., & Buchanan, B. (1975). A model of inexact reasoning in medicine. Mathematical Biosciences, 23, 351–379. TDWG-SDD (2005). SDD part 0: Introduction and primer to the SDD standard (En lı́nea, Consulta 7/02/2005, http://160.45.63.11/Projects/ TDWG-SDD/Primer/index.htm). UBio (2003). X:ID, version (En lı́nea, Consulta 30/7/2003, http://ubio.org/ offerings/applications/key/index.html). University of Toronto. Department of Botany; the University of Toronto Libraries; the Royal Ontario Museum (1996). Pollyclave a multi-entry identification key version 1.6 (En lı́nea, Consulta 6/06/2003, http://prod. library.utoronto.ca/pollyclave/index.html). Zadeh, L. A. (1965). Fuzzy Sets. Information and Control, 8(3), 338–353. http://www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa http://www.lucidcentral.com/default.htm http://www.biodiversity.uno.edu/delta/ http://www.xidservices.com http://www.ucjeps.berkeley.edu/meacham/meka/ http://www.ibiblio.org/pub/academic/biology/ecology&plus;evolution/software/pandora http://www.ibiblio.org/pub/academic/biology/ecology&plus;evolution/software/pandora http://www.eti.uva.nl/Home/Articles.html http://www.160.45.63.11/Projects/TDWG-SDD/Primer/index.htm http://www.160.45.63.11/Projects/TDWG-SDD/Primer/index.htm http://www.ubio.org/offerings/applications/key/index.html http://www.ubio.org/offerings/applications/key/index.html http://www.prod.library.utoronto.ca/pollyclave/index.html http://www.prod.library.utoronto.ca/pollyclave/index.html XKey: A tool for the generation of identification keys Introduction Background Identification keys Digital representation of taxonomical knowledge Tools for the generation of identification keys Analysis of the problem Construction of identification keys Objectives of the division criterion Classic division criteria Division criteria in taxonomy Treatment of null values Reliability of a character Confirmatory characters Construction of knowledge bases Treatment of uncertainty Characteristics of XKey Use of different division criteria Treatment of null values Control of the reliability of a character Inclusion of confirmatory characters Operation modes Diversity of output formats Construction of a knowledge base Execution results Test of the XKey tool Execution data for the XKey tool and methodology followed Observations about the average length of the keys generated Observations about the equilibrium of the keys Observations regarding the degree of branching Observations about the number of confirmatory characters included Observations about the number of characters used Qualitative comparison of the criteria Discussion and biological interpretation of the results obtained/effect of interactivity Conclusions and final notes References 
work_o4xy7zmaafdrvgjqdwbncbbxg4	----	Tajweed: An Expert System for Holy Qur’an Recitation Proficiency Procedia Computer Science 65 ( 2015 ) 807 – 812 Available online at www.sciencedirect.com 1877-0509 © 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of Universal Society for Applied Research doi: 10.1016/j.procs.2015.09.029 ScienceDirect International Conference on Communication, Management and Information Technology (ICCMIT 2015) Tajweed: An Expert System for Holy Qur’an Recitation Proficiency Musbah J. Aqela, Nida M. Zaitounb * aDepartment of Computer Science, Faculty of Information Technology, Applied Science University, Amman, Jordan bDepartment of Computer Science, Faculty of Information Technology, Applied Science University, Amman, Jordan Abstract An expert system for the holy Qur'an recitation proficiency is developed to help non Arabic Muslims to recite the holy Qur'an according to the Islam's rule. The system was developed as a rule-based system and implemented using Prolog language. The system was tested by experts in Tajweed and the results were very excellent. © 2015 The Authors. Published by Elsevier B.V. Peer-review under responsibility of Universal Society for Applied Research. Keywords: Type your keywords here, separated by semicolons ; 1. Introduction Expert systems are computer applications [4] or a programming approach [5] which embodies some non- algorithmic expertise for solving certain types of problems [4] by providing answers for complicated problem [5]. The applications of expert systems are rapidly increasing. Such applications are very effective in situations when the domain expert is not available. Expert systems found many applications in many fields, like science, business, medicine, and social area [12]. Expert systems tools are valuable because they provide rich software development environments, and the knowledge representation and the inference engine are already built into them. KHABEER is an Arabic CLIPS- based expert system tool where all the commands and syntax are written in Arabic [6]. KHABEER was developed using the conventional language C. KHABEER uses rules as its primary knowledge representation approach and supports a rich pattern-matching language for specifying rule conditions. The system has interface that supports pull-down menus [6]. Tajweed of the Holy Qur'an is the knowledge and application of the rules of recitation so the reading of the * amusbahaqel@yahoo.com, bnida_zaitoun@yahoo.com © 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of Universal Society for Applied Research http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.09.029&domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.09.029&domain=pdf 808 Musbah J. Aqel and Nida M. Zaitoun / Procedia Computer Science 65 ( 2015 ) 807 – 812 Qur'an is as the Prophet Mohammed peace and blessings are upon him, recited [1], [2], [3]. The word "Tajweed" means to improve, make better [2], [3], betterment [1]. It is one of the most honoured of sciences and one of the best of them due to its relation to Allah’s words [1]. Applied definition of Tajweed is articulating every letter from its articulation point and giving the letter its rights and dues of characteristics [1]. It is preserving the tongue from mistakes in pronunciation of the Glorious Qur’an during reading [1]. One who wishes to learn a Tajweed needs to understand and well practice its several principles [1], [2]. The rule setter from the practical point of view is the Messenger of Allah the Prophet Mohammed peace and blessings be upon him, because the Qur’an was revealed to him from Allah, The most high [1]. The knowledge of Tajweed is contingent on four matters: a. Knowledge of the articulation points of the letters. b. Knowledge of the characteristics of the letters. c. Knowledge of what rules change in the letters due to the order of letters. d. Exercising the tongue and a lot of repetition. [1] 2. Tajweed Rules The rules of tajweed are classified as follows: I. Noon and Meem Mushaddad is that noon or meem which has a shaddah with Ghunnah of 2 beats. II. Al –Qalqalah is vibration the sound at the end of the pronunciation of any letter of Qaaf, Ttaa, Baa, Jiim or Daal, when it is Saakin; with sukoon or shaddah. III. Noon saakinah and Tanween Rules Noon Saakinah is the noon with no Harakah or with a Sukoon. Tanween is a noon Saakinah at the end of the nouns is pronounced as Noon Saakinah without writing noon. IV. Meem Saakinah Rules Meem Saakin is the meem with no Harakah or with a Sukoon sign on it. Rules of Meem Saakinah a. Ikhfaa Shafawi: Hiding meem by the Baa with the two lips are not completely contact with 2 beats Ghunnah. b. Idghaam Shafawi: Mixing of a saakin Meem into a Mutaharrik Meem following it with 2 beats Ghunnah. c. Izhaar Shafawi: Clear Meem Saakinah with a complete contact of the two lips when it is followed by any letter other than Baa and Meem. I. Al-Madd Rules Al-Madd means long, to make the Madd letters long under some conditions from two to six beats depending upon its kind. Madd letters: Leen letters both are preceded by a letter with a Fathah i. Yaa Saakinah. ii. Waaw Saakinah. b. Huroof Maddiyyah each is preceded by a letter with a likely haraka i. Alif saakinah preceded by a Fathah. ii. Waaw Saakinah preceded by a Dhammah. iii. Yaa Saakinah preceded by a Kasrah. The complete model for the Qur'an tajweed rules are shown in fig. 1. 809 Musbah J. Aqel and Nida M. Zaitoun / Procedia Computer Science 65 ( 2015 ) 807 – 812 Figure (1): A model for Tajweed rules 810 Musbah J. Aqel and Nida M. Zaitoun / Procedia Computer Science 65 ( 2015 ) 807 – 812 3. Tajweed Expert System Design The expert system consists of four components and the process of development as shown in Figure 2. These components are as follows: Inference Engine: The core of the system, which obtains the solution for a particular problem from the knowledge Base and data in working storage. User Interface: The interface is designed as a set of windows and menus to facilitate the interaction between the user and the system in order to make the system easy-to-use. Explanations: the ability of the system to explain the reasoning process that it used to reach a conclusion and a recommendation. Knowledge Base: consists of rules and facts that are acquired from domain experts. This knowledge is represented as a set of rules in the form of IF THEN. For example: IF (Noon or Meem has a shaddah) then The rule is Mushaddad IF (Q aaf, Ttaa, Baa, Jiim or Daal is saakin) then The rule is Qalqalah IF (Noon Saakinah is followed by Hamza, Haa, hhaa, Aiin, Khaa or Ghaiin) then Then the rule is Izhaar IF (Noon Saakinah is followed by Baa) then Then the rule is Iqlaab IF (Noon Saakinah is followed by Yaa or Waaw) then Then the rule is In Complete Idghaam with Ghunnah Figure (2) Proposed expert system components IF (Noon Saakinah is followed by Noon or Meem) then Then the rule is Complete Idghaamwith Ghunnah IF (Noon Saakinah is followed by Raa or Laam) then Then the rule is Complete Idghaam without Ghunnah IF (Noon Saakinah is followed by Baa) then Then the rule is Iqlaab 4. Expert System Results Assume a user would like to consult the expert system for the proper recitation of a phrase of a holy Quran (Aya). The system starts by asking some questions regarding to word of the phrase, and each two successive letters of this word and the user will answer as follows: Is the letter Noon or Miim Mushashaddad (y/n) ? no Is the letter non saakinah or tanween (y/n) ? n Is the letter miim saakinah (y/n) y Is the letter follows the mii saakinah miim (y/n) y Then the system will conclude that the correct recitation for this word is by making IDGHAAM Shafawi for the two Miim letters as shown in Figure 3. 811 Musbah J. Aqel and Nida M. Zaitoun / Procedia Computer Science 65 ( 2015 ) 807 – 812 5. Conclusion An expert system was developed for proper recitation of holy Qur'an which is called Tajweed rules. The expert system was designed and implemented as a rule-based system. The system was tested and evaluated by experts in the field of Tajweed and they results were very excellent since the system could provide the proper recitation for any word in holy Qur'an. This is very useful for non-Arab Muslims and for the Arab students who are studying at the universities the holy Qur'an Tajweed rules. 6. References [1] Kareema Carol Czerepinski, Ash-sheikh Dr. Ayman Rushdi Swayd. (2000, 6, 25) “Tajweed Rules of the Qur’an”, Part One. [Online]. Available (2014, January, 30): http://d1.islamhouse.com/data/en/ih_books/single2/en_Tajweed_Rules_of_the_Quran_Part_01.pdf. [2] About Tajweed. (2014, January, 30), [Online]. Available: http://www.abouttajweed.com/index2.php?option=com_content&do_pdf=1&id=15 [3] Tajweed Study. (2014, January, 30), [Online]. Available: http://tajweedstudy.com/index.php?option=com_content&view=article&id=126&Itemid=124 [4] Dennis Merritt. (2001, August), “Building Expert System in Prolog”. [Online]. Available (2014, January,30): http://ce.yazd.ac.ir/ghasemzadeh/eLecture-ExpertSystem/Ch00-Expert%20systems%20in%20Prolog.pdf [5] Venus Samawi, Akram Mustafa, Abeer Ahmad. (2013, January), “Arabic Expert System Shell”. [Online]. Available (2014, January,30): http://www.ccis2k.org/iajit/PDF/vol.10,no.1/4094-6.pdf [6] Mostafa M. Aref, Husni A. Al-Muhtaseb. (2013, January), “KHABEER (خبير): An Object-Oriented Arabic Expert System Shell”. Presented at Information and Computer Science Department, Information and Computer Science Department, Saudi Arabia. [Online]. Available (2014, January,30): http://faculty.kfupm.edu.sa/ics/muhtaseb/Research/KH-JNL1F1.pdf [7] Tajweed rules of the QUR'AN, Mohamed A. Abdelhadi, Ghassan Kanann. (2013, Noveber), “A New Developed Model for Arabic Information Retrieval System based on Knowledge Base System”. “International Journal of Emerging Research in Management &Technology”. [Online]. Volume-2, (Issue-11), ISSN: 2278- 9359, Available (2014, January,30): http://www.ermt.net/docs/papers/Volume_2/issue11_November2013/V2N11-110.pdf [8] http://d1.islamhouse.com/data/en/ih_books/single2/en_Tajweed_Rules_of_the_Quran_Part_01.pdf. Figure 3: A consultation session with Tajweed expert system 812 Musbah J. Aqel and Nida M. Zaitoun / Procedia Computer Science 65 ( 2015 ) 807 – 812 [9] http://www.abouttajweed.com/ [10] Tajweed Rules. http://understandquran.com/fileadmin/user_upload/extras/tajweed/Tajweed_Rules.pdf [11] A. Shukri, M. AL-Majali, A. AL-Qudah, A. Abu-Ghalion, A. Jaiousi, O. Yusef, " Al-Muneer: Tajweed recitation", Amman, Jordan, 2012. [12] Rajdeep Borgohain, Suguta Sanyal, (2012). "Rule based expert system for diagnosis of neuromuscular disorders". 
work_o5akemxz5naoloykghbfuhkji4	----	Hindawi Publishing Corporation Advances in Fuzzy Systems Volume 2012, Article ID 823052, 11 pages doi:10.1155/2012/823052 Research Article A Fuzzy Rule-Based Expert System for Evaluating Intellectual Capital Mohammad Hossein Fazel Zarandi,1 Neda Mohammadhasan,1 and Susan Bastani2 1 Department of Industrial Engineering, Amirkabir University of Technology, P.O. Box 15875-4413, Tehran, Iran 2 Department of Social Sciences, Alzahra University, P.O. Box 1993891176, Tehran, Iran Correspondence should be addressed to Mohammad Hossein Fazel Zarandi, zarandi@aut.ac.ir Received 17 April 2012; Accepted 30 August 2012 Academic Editor: Zeng-Guang Hou Copyright © 2012 Mohammad Hossein Fazel Zarandi et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. A fuzzy rule-based expert system is developed for evaluating intellectual capital. A fuzzy linguistic approach assists managers to understand and evaluate the level of each intellectual capital item. The proposed fuzzy rule-based expert system applies fuzzy linguistic variables to express the level of qualitative evaluation and criteria of experts. Feasibility of the proposed model is demonstrated by the result of intellectual capital performance evaluation for a sample company. 1. Introduction Today’s “new economy” is actually the “knowledge econ- omy” or the “knowledge society”. This development is characterized by increased digitalization, globalization, and application of knowledge [1]. Intellectual capital consists of assets created through intellectual activities ranging from acquiring new knowledge (learning) and inventions to creating valuable relationships [2]. In strategy decision- making, an intellectual capital evaluation model should be a feasible tool for managers to handle and improve existing intellectual capital effectively and efficiently in accordance with different performance levels of each item and criterion [1]. Tai and Chen [3] proposed a feasible enterprise intel- lectual capital evaluation model by means of 2-tuple fuzzy linguistic approach to assist managers understand the per- formance of intellectual capital. In this model, experts apply linguistic variables to express the level of intellectual items. Fasanghari et al. [4] proposed a fuzzy expert system which can measure intellectual capital based on three categories: knowledge capital, management capital, and market capital. This expert system uses 27 rules to assess level of intellectual capital based on these three components. However, these measurement models help enterprise managers to under- stand the status of intellectual capital, but they frequently use the term “intellectual capital” and its items in an all- encompassing fashion with the risk that in time the identity of object will become unclear [5]. Managers are faced with many questions such as: “what is intellectual capital?”, “what does employee’s knowledge mean?”, “how can I found out the level of employee’s knowledge in my company?”, or “is employee’s knowledge important for my company?” a set of rules which determine the level of intellectual capital items by asking some questions from managers is useful. Manager’s answer to these questions is expressed with linguistic labels. Here, we deal with a qualitative model as a system model based on linguistic description. Sugeno and Yasukawa [6] defined fuzzy modeling as a qualitative modeling scheme by which we quantitatively describe system behavior using a natural language. Thus, a suitable fuzzy expert system, which could identifies the performance degree of intellectual capital, is proposed. The proposed fuzzy rule-based expert system evaluates the performance degree of intellectual capital based on the model proposed by Tai and Chen [3] using performance level and importance level of each intellectual item. In this expert system, the importance and performance level of each 2 Advances in Fuzzy Systems intellectual item are determined indirectly through asking some questions from managers. The main contributions of this paper are as follows. (i) Instead of asking managers directly about the status of enterprise’s intellectual items—so as previous works—the importance and performance level of each intellectual capital is determined indirectly through asking some question. (ii) Developing two sets of rules (totally about 80 rules) for each intellectual capital items to assess the overall status of intellectual capital more efficiently (one rule set for importance level and one for performance level). This paper is organized as follows. Section 2 introduces the concept of intellectual capital and fuzzy control rules. In Section 3, proposed fuzzy expert system to evaluate intellectual capital is presented. In Section 4, the proposed system is implemented in a real company. Finally, some conclusions and future works are presented at the last section of the paper. 2. Background 2.1. Intellectual Capital (IC). Edvinsson and Malone [7] define the difference between a firm’s market value and book value as the value of intellectual capital. Stewart [8] defines intellectual capital as “intellectual material—knowledge, information, intellectual property, experience—that can be put to use to create wealth”. He recognizes human capital, structure capital, and customer capital. Sveiby [9] proposed that intellectual capital includes employee competence, internal structure, and external structure. Edvinsson [10] divides structure capital into organization capital and cus- tomer capital. Liebowitz and Wright [11] divide intellectual capital into four unique categories such as human capital, customer capital, process capital, and innovation capital. Organization for Economic Cooperation and Development (OECD) [12] describes intellectual capital as “the economic value of two categories of intangible assets of a company: (1) organizational (“structural”) capital and (2) human capital“. Bukh et al. [13] identify that in most models of intellectual capital, human capital, customer capital, and organization capital are the components of intellectual capital. Human capital is the foundation of IC, a primary element to perform IC’s functions. It refers to such factors as employee’s knowledge, skill, capability, and attitudes in relation to fostering performances which customers are willing to pay for and the company’s profit comes from. Customer capital, which acts as a bridge and a catalyst on the operations of IC, is the main requirement and determinant in converting IC into market value and thereupon organization business performance [14]. Structural capital is the use of the part of the intellectual capital that is owned by the company. Edvinsson [6] defines the term as: “Hardware, software, databases, organizational structure, patents, trademarks, and everything else of organizational capability that supports that employee’s productivity. Everything left at the office, when the employees go home. Unlike human capital, structural capital can be owned and thereby traded.” To evaluate the overall performance of intellectual ca- pital, we should define the performance of intellectual capital components. So, by reviewing the literature, the following classification is used in this paper: human capital can be defined as a combination of employee’s knowl- edge, employee’s innovativeness, satisfaction degree, and employee’s turnover rate. Customer capital can be classified into market share rate, customer loyalty, customer satis- faction, and customer relationship. Structural capital can be classified into trademark, operation process, information system, and corporate culture. By defining the performance degree of each item, we can evaluate the performance degree of intellectual capital. The importance and performance degree of each intellectual item is determined with linguistic variables. Fuzzy set theory is suitable in qualitative assess- ment. So for each intellectual item, we define a fuzzy rule set to evaluate the item through asking question from managers. 2.2. Fuzzy Control Rules. Conventional rule-based expert systems use human expert knowledge to solve real-world problems that normally would require human intelligence. Expert knowledge is often represented in the form of rulesor as datawithin the computer. Depending upon the problem requirement, these rules and data can be recalled to solve problems [15]. Some of the important advantages of expert systems are as follows [16]: (i) ability to capture and preserve irreplaceable human experience; (ii) ability to develop a system more consistent than human experts; (iii) minimize human expertise needed at a number of locations at the same time (especially in a hostile environment that is dangerous to human health); (iv) solutions can be developed faster than human experts. The basic components of an expert system are illustrated in Figure 1. The knowledge base stores all relevant informa- tion, data, rules, cases, and relationships used by the expert system. A knowledge base can combine the knowledge of multiple human experts. A rule is a conditional statement that links given conditions to actions or outcomes [16]. Zadeh [17] states that precise qualitative analyses of the behavior of humanistic systems are not likely to have much relevance to the real-world societal, political, economic, and other types of problems which involves humans either as individuals or in groups. He suggested linguistic analysis in place of quantitative analysis. Mamdani was the first to apply fuzzy logic to control (Figure 2). This topic became known as fuzzy algorithmic control or linguistic control. Fuzzy control can be viewed as a result of qualitative modeling of a human operator. Advances in Fuzzy Systems 3 Knowledge base Knowledge base acquisition facility User interface Inference engine Explanation facility Expert knowledge Users Figure 1: Architecture of a simple expert system [16]. Input MF Output MFMin Max YX YX Output Z Input (X, Y) µµµ µ µ µ µ A1 B1 B2 C1 C  C 1 C2 Z(COA) C2 Z2 Z1 A2 Figure 2: Mamdani fuzzy inference system using min and max for t-norm and t-conorm operators [16]. 3. Proposed Fuzzy Expert System to Evaluate Intellectual Capital Performance 3.1. Overview of the Problem Domain. Intellectual capital is intangible, and assessing its components needs knowl- edge of intellectual capital experts. In knowledge econ- omy, enterprises should manage their intellectual capital to maintain their competitive advantage in capital market. To manage intellectual capital, enterprises should measure it. In this paper, we have developed a fuzzy rule base expert system for this purpose. Using knowledge of intellectual capital experts, this expert system helps enterprise decision makers to determine organization’s performance degree of intellectual capital. Fuzzy expert system models are similar to the classical expert system models, where input spaces are mapped to an output space. A fuzzy expert system uses a collection of fuzzy membership functions and rules for knowledge representation and reasoning with observed system state data. An approximate reasoning process in a fuzzy expert system is executed as follows [18]. Fuzzification and forming a fuzzy set is a sub process in a fuzzy expert system: 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Very low Low Medium High Very high Figure 3: Fuzzy sets on proportion of staff with a university degree. (a1) Fuzzification. The degree of the memberships for each rule’s premise is determined by matching the linguistic terms with the actual values of the input variables [18]. For example, in the proposed evaluation model, “Proportion of staff with a university degree” is the linguistic variable, and “high” is one of its fuzzy sets. A fuzzy set can be formed by assigning a membership value to each object in the interval of [0, 1]. Membership values represent the degree to which an object belongs to a fuzzy set. Let X denotes the universe of discourse, where x represents an element of the universe, X and A denote a fuzzy set. A fuzzy set is hence characterized by its membership function, µA(x) as [19] µA(x) : X → [0, 1]. (1) Membership function states that values assigned to the elements of the universal set, X, fall within a specified range. In the meantime, it indicates the membership grade of these elements in fuzzy set A. A fuzzy set A, on universe of discourse of X, can also be defined as a set of ordered pairs as [19] A = {(x, µA(x) ) | x ∈ X}. (2) Figure 3 shows membership functions of fuzzy sets “very low”, “low”, “medium”, “high”, and “very high”. Horizontal 4 Advances in Fuzzy Systems axis of graph in Figure 3 represents “proportion of staff with university degree” in percentage (the universal set X) and vertical axis represents the degree to which this proportion for a company can be labeled “very low”, “low”, “medium”, “high”, and “very high”. (a2) Forming Fuzzy Sets. To represent a fuzzy set in a computer, we need to define its membership function. One approach is to poll a group of people for their understanding of the term that we are attempting to represent by the fuzzy set [20]. For example, consider the concept of high proportion of staff with university degree. We could ask intellectual capital experts to what degree they believe a proportion of staff with university degree is high. After acquiring answers for a range of proportions, we could perform simple averaging to produce a fuzzy set of high proportion of staff with university degree. We can now use this function to ascribe a belief (or membership value) to a given proportion belonging to the fuzzy set of high. We could continue this polling to account for other proportion descriptions such as very low, low, medium, and very high (as shown in Figure 3). Using linguistic variables in fuzzy expert systems make all modules of a classical expert system fuzzy. (b1) Fuzzy Knowledge Acquisition. This module involves the adaption of the expert system to a specific domain. Zimmerman [18] state that the goal of this module is extracting and organizing the knowledge and expertise embedded in the problem domain. Knowledge acquisition is described in literature as the most difficult process, one of the greatest difficulties, the biggest bottleneck, or the most time consuming task. There is no all-encompassing, unified theory of how to acquire knowledge, and probably never will be [21]. In this paper, we use direct approach for knowledge acquisition process. Direct approach or interviewing is good for obtaining a sense of the knowledge domain [21]. Interviewing consists of asking the domain experts questions about the domain of interest and how they perform their tasks [22]. To obtain the knowledge of intellectual capital from the experts, we provided them with a set of indicators which may be included in an intellectual capital statement. We used a structured interview in which the experts were asked to determine which of these indicators have impact on importance and performance of which intellectual capital components. Next, the experts were asked to determine to what degree they are effective on outputs. Finally, they were asked to connect indicators to one another, outputs to one another, and indicators to outputs through rules. For example, intellectual capital experts believe that proportion of staff with a university degree is positively related with employee’s knowledge capital, so, for instance, one of the extracted rules is as follows: IF proportion of staff with a university degree is very low, THEN employee’s knowledge degree is very poor. Table 1: Fuzzy rule set for employee’s knowledge performance degree. K-1-1 IF Proportion of staff with a university degree is very low THEN Employee’s knowledge degree is (VP). K-1-2 IF Proportion of staff with a university degree is low THEN Employee’s knowledge degree is (P). K-1-3 IF Proportion of staff with a university degree is medium THEN Employee’s knowledge degree is (F). K-1-4 IF Proportion of staff with a university degree is high AND Proportion of 30–50 age employees is low THEN Employee’s knowledge degree is (F). K-1-5 IF Proportion of staff with a university degree is high AND Proportion of 30–50 age employees is medium THEN Employee’s knowledge degree is (G). K-1-6 IF Proportion of staff with a university degree is high AND Proportion of 30–50 age employees is high THEN Employee’s knowledge degree is (VG). K-1-7 IF Proportion of staff with a university degree is very high AND Proportion of 30–50 age employees is low THEN employee’s knowledge degree is (F). K-1-8 IF Proportion of staff with a university degree is very high AND Proportion of 30–50 age employees is medium THEN Employee’s knowledge degree is (G). K-1-9 IF Proportion of staff with a university degree is very high AND Proportion of 30–50 age employees is high THEN Employee’s knowledge degree is (VG). Most of the indicators are obtained from “A guideline for intellectual capital statements” [1] and others are determined by experts. Like the evaluation model for intellectual capital proposed by Tai and Chen [3], we use a classification for intellectual capital components (as shown in Figure 4), and we evaluate each components through its importance degree and performance degree. Then, by combining all these degrees, the overall performance of intellectual capital will be determined (Tables 1 and 2). (b2) Fuzzy Knowledge Base. The knowledge base represents the knowledge related to the problem domain. A subset of knowledge base is the fact base, which contains both sym- bolic information and numeric information [18]. Factual Advances in Fuzzy Systems 5 Intellectual capital Structural capital Trademarks Operation process Information system Corporate culture Customer capital Market share rate Customer loyalty Customer satisfaction Customer relationship Human capital Employee’s knowledge Innovativeness Satisfaction degree Employee’s turnover rate Figure 4: Components of intellectual capital (revised from [3]). knowledge or observed system state or facts in a knowledge- based system are expressed as fact base. A notation of observed system state is X1 isr A ∗ 1 AND X2 A ∗ 2 AND ··· AND Xn isr A∗n , (3) where “isr” is a short form for “is contained in”, “is similar to”, “is compatible to”, and so forth [18]. In short, it is usually expressed as A∗ = A∗1 AND A∗2 AND . . . AND A∗n (4) The user of an expert system inputs the factual knowl- edge and chooses the linguistic terms for observed states of the system [18]. For example, in proposed expert system in this paper, an observed system state is as follows: Proportion of staff with a university degree isr high and proportion of 30–50 age employees isr medium or in short notation: High AND Medium (5) Another part of the knowledge base is rule bases which contain heuristics. They are used by domain experts to diagnose, predict, control, and so forth a system behavior [18]. Experts of intellectual capital domain use heuristics to evaluate components of intellectual capital. A notation of rule-base is IF X1 isr A ∗ 1 AND X2 isr A ∗ 2 AND . . . AND Xn isr A ∗ n , THEN Y isr B (6) For example, in proposed expert system in this paper, an observed system state is as follows: IF Proportion of staff with a university degree isr high AND Proportion of 30–50 age employees isr medium, THEN employee’s knowledge degree isr good During the knowledge acquisition phase, the domain experts provide these rules. The linguistic terms and their membership values and/or functions are also specified in such rules by experts [18]. (b3) Fuzzy Inference Mechanism. The inference mechanism uses the knowledge base and the rule base for inference and derivation of decisions for given observed facts [18]. The inference mechanism which we use in this paper is data driven or forward chaining. Forward chaining is an inference strategy that begins with a set of known facts, derives new facts using rules whose premises match the known facts and continues this process until a goal state is reached or until no further rules have premises that match the known or derived facts (Figure 5) [20]. For example, in this paper to model evaluation procedure of employee’s knowledge capital in a forward chaining rule-based expert system (Figure 5), system’s rules are as follows. We assert the following facts into the working memory as supplied by sample company’s experts. (i) Proportion of staff with a university degree is high. (ii) Proportion of 30–50 age employees is medium. (iii) Total training costs/total payroll expenses are high. 6 Advances in Fuzzy Systems Table 2: Fuzzy rule set for employee’s knowledge importance degree. K-2-1 IF Total training costs/total payroll expenses are very low THEN Employee’s knowledge are very unimportant for company. K-2-2 IF Total training costs/total payroll expenses are low THEN Employee’s knowledge is unimportant for company. K-2-3 IF Total training costs/total payroll expenses are medium AND Number of employees with a competency development plan/total employees is low AND Number of employees being trained as project managers/total employees is low THEN Employee’s knowledge is very unimportant for company. K-2-4 IF Total training costs/total payroll expenses are medium AND Number of employees with a competency development plan/total employees is medium AND Number of employees being trained as project managers/total employees is low THEN Employee’s knowledge is unimportant for company. ··· K-2-5 IF Total training costs/total payroll expenses are medium AND Number of employees with a competency development plan/total employees is low AND Number of employees being trained as project managers/total employees is medium THEN Employee’s knowledge is unimportant for company. K-2-6 IF Total training costs/total payroll expenses are medium AND Number of employees with a competency development plan/total employees is medium AND Number of employees being trained as project managers/total employees is medium THEN Employee’s knowledge is (F) for company. K-2-7 IF Total training costs/total payroll expenses are high AND Number of employees with a competency development plan/total employees is medium AND Number of employees being trained as project managers/total employees is medium THEN Employee’s knowledge is important for company. K-2-8 IF Total training costs/total payroll expenses are high AND Number of employees with a competency development plan/total employees is high AND Number of employees being trained as project managers/total employees is medium THEN Employee’s knowledge is important for company. Table 2: Continued. K-2-9 IF Total training costs/total payroll expenses are high AND Number of employees with a competency development plan/total employees is medium AND Number of employees being trained as project managers/total employees is high THEN Employee’s knowledge is important for company. K-2-10 IF Total training costs/total payroll expenses are high AND Number of employees with a competency development plan/total employees is high AND Number of employees being trained as project managers/total employees is high THEN Employee’s knowledge is very important for company. K-2-11 IF Total training costs/total payroll expenses are very high AND Number of employees with a competency development plan/total employees is medium AND Number of employees being trained as project managers/total employees is medium THEN Employee’s knowledge is important for company. K-2-12 IF Total training costs/total payroll expenses are very high AND Number of employees with a competency development plan/total employees is high AND Number of employees being trained as project managers/total employees is medium THEN Employee’s knowledge is very important for company. K-2-13 IF Total training costs/total payroll expenses are very high AND Number of employees with a competency development plan/total employees is medium AND Number of employees being trained as project managers/total employees is high THEN Employee’s knowledge is very important for company. K-2-14 IF Total training costs/total payroll expenses are very high AND Number of employees with a competency development plan/total employees is high AND Number of employees being trained as project managers/total employees is high THEN Employee’s knowledge is very important for company. (iv) Number of employees with a competency development plan/total employees is medium. (v) Number of employees being trained as project man- agers/total employees is medium. The system takes each rule in turn and checks to see if its premises are listed in the working memory. When the system finds matches for all the premises, it places the rule’s conclusion in the working memory. Advances in Fuzzy Systems 7 Table 3: Inputs of fuzzy rule-based expert system for sample company. Criteria Value Item Rating Performance degree Proportion of staff with a university degree is high Employee’s knowledge Proportion of 30–50 age employees is medium Importance degree Total training costs/total payroll expenses are medium Number of employees with a competency development plan/total employees is medium Importance degree Employee’s knowledge Number of employees being trained as project managers/total employees is medium Performance degree Number of patent claims is low Human capital Employee’s innovativeness Number of new products introduced during the past 12 months/total products is medium Importance degree Cost of research and development/total turnover is low Performance degree Proportion of generally satisfied employees is high Satisfaction Number of day absence/total working days is low Importance degree The response rate in staff satisfaction survey is high Performance degree Employee’s turnover rate Number of employees leaving the company/total employees is low Importance degree Employees’ total number of seniority/total employees is fair Performance degree Market share rate Company’s sales revenue from market/total sales revenue available in market is high Importance degree Marketing initiative is of a considerable amount of importance for the company Performance degree Customer loyalty longevity of customer relationship is medium Importance degree Number of repeat customers buying the brand /total customers is of a considerable amount of importance for the company Customer capital Performance degree Customer satisfaction Proportion of customers who are satisfied is high Importance degree The five or ten customer’s proportion of total turnover is fair Performance degree Company considerably have close contact with customer during specification of requirements Customer relationship Company advising the customer considerably during the entire process Importance degree Guests sometimes participate in company lunch buffet Customer representatives sometimes participate at company’s conferences Performance and Importance degree Trademark Company has a distinctive sign or indicator to distinguish its products or services from those of other entities Performance degree Operation process The proportion of procedures described or proportion of instruction described is fair Importance degree Standard operating procedure is of considerable importance for the company Structural capital Performance degree Information system Proportion of updated knowledge documents on the intranet is fair Importance degree Its costs are high Performance degree Corporate culture Employees are averagely identified with company’s perspective Importance degree Having values, faith, and behavior criterions which are approved and shared by all the staff is of considerable importance for the company 8 Advances in Fuzzy Systems Insert information into working memory Check first rule Check next rule Add conclusion to working memory Stop Rules remainPremises match working memory F T T F Figure 5: Forward chaining inference process [20]. From the initial information entered into the working memory, the system concludes new information. (1) Employee’s knowledge degree is good. (2) Employee’s knowledge is important for company. Since the proposed expert system is a fuzzy expert system which uses linguistic variables, a fuzzy inference should be used to evaluate employee’s knowledge capital. Fuzzy Inference. The conjunctive combination of antecedent memberships is executed and projected to the conclusion part of each rule to determine the degree of membership on its consequent [18]. In this paper, we used Mamdani type of linguistic model inference as follows. Step 1. The antecedent of each rule is evaluated. When the antecedent of rule has more than one part, fuzzy intersec- tions (t-norms) and fuzzy unions (t-conorms) are applied to obtain a single membership value. Fuzzy intersection refers to AND, that we interpret it by Yager t-norm in this paper. Hence, intersection of two fuzzy sets A and B defined on X is given as µA∧B(X) = 1 − min ( 1, (( 1 − µA(x) )p + ( 1 − µB(x) )P )1/P ) . (7) Fuzzy union refers to OR, that we interpret it by Yager t- conorm in this paper. So, union of two fuzzy sets A and B defined on X is given as µA∨B(X) = min ( 1, ( µA(x) p + µB(x) P )1/P ) . (8) Step 2. The consequent of each rule is given as a fuzzy set, and the antecedent value has obtained in Step 1. To obtain a new fuzzy set, we apply a fuzzy implication operator. We use minimum operator which cuts the consequent’s membership function (Figure 6). 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Figure 6: Obtaining each rule conclusion. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Figure 7: Output result for employee’s knowledge performance. Step 3. The outputs obtained for each rule are combined to a single fuzzy set, using Yager s-norm (8). The last subprocess in a fuzzy expert system is defuzzifi- cation. The fuzzy value needs to be converted to a crisp value. Defuzzification. The overall output fuzzy set is converted into a crisp number on the base variable domain [18]. In this model we used Yager defuzzification method (9). Z = ∫ z [ µA(z) ]w zdz ∫ z [ µA(z) ]w dz . (9) For example, to define the performance degree of employee’s knowledge, the fuzzy inference mechanism is as follows. System 1. proportion of staff with a university degree is (1) very low, (2) low, (3) fair, (4) high, and (5) very high. Assume that a user chooses high, then system goes to rules k-1-4 and k-1-5, for second premise of rules k-1-4 and k-1-5. System 2. Proportion of 30–50 age employees is (1) low, (2) medium, and (3) high. Assume that a user chooses medium. Using Mamdani inference with Yager s-norm and t-norm, and the output is as shown in Figure 7. Advances in Fuzzy Systems 9 Total training costs/total payroll expenses are medium Number of employees with a competency development plan/total employees is medium Number of employees being trained as project managers /total employees is medium Fuzzy rule set for employee’s knowledge importance degree Crisp inputs Then Fuzzy inference engine K-2-1: K-2-2: K-2-3: K-2-4: K-2-5: K-2-6: K-2-7: Deffuzification AndAnd User interface Fuzzy inputs If Yager t-norm Yager t-norm Yager t-norm Ya ge r s- n o rm Yager t-norm Yager t-norm Figure 8: Sample procedure for measuring employee’s knowledge with proposed fuzzy expert system. By defuzzifying employee’s knowledge, performance degree is 0.7038. Sample procedure for measuring employee’s knowledge performance in proposed fuzzy expert system is as shown in Figure 8. (b4) User Interface. In many cases, the disappointing history, that reasoning systems have had can be traced to a lack of respect for the user. The successful systems strive to use oper- ational techniques and decision processes with which users are comfortable. Otherwise, frustrated users will shut the system and quickly return to their old way of doing things. The user interface must be at least able to display and update system message or problem results comprehensibly [21]. In the proposed expert system for evaluating intellectual capital, a user is provided an interactive interface that acquires the information about company through asking questions from them. In Figure 9 a view of evaluation panel for employee’s knowledge capital is shown. 10 Advances in Fuzzy Systems Figure 9: A view of the proposed model interface. 4. Case Study To evaluate applicability of the proposed fuzzy expert system, it is implemented in an industrial company which its activity is to design, source, assemble, install, and commission control systems, precision instrumentation, and electrical systems for power plants as well as mining and process industries. This company has a large number of experts in different fields, and intellectual capital plays an important role in it. Thus, using an expert system which helps managers to handle and understand the status of intellectual capital efficiently and effectively is of great importance in this company. Initially managers answer the questions asked by the system to determine performance degree and importance degree of each item using linguistic terms (shown in Table 3). The Total performance of every element is produced by multiplying performance degree by importance degree of each element. Hence, the overall performance of intellectual capital for a company is equal to the summation of all elements’ total performance degree. The overall degree of employee’s knowledge is 0.3593 which results by multiplication of performance degree and importance degree of employee’s knowledge. According to fuzzy sets on employee’s knowledge (similar to Figure 3), 0.3593 is a member of “Low” fuzzy set with 0.2035 mem- bership value and is a member of “Medium” fuzzy set with 0.7965 membership value. Thus, employee’s knowledge capital of company is medium (0.7965 > 0.2035). Employee’s performance or human capital performance is evaluated through employee’s knowledge, employee’s innovativeness, employee’s satisfaction, and employee’s turnover rate. Uni- versity degree, age, training costs, and competency devel- opment plan are used for evaluating employee’s knowledge. Number of patent claims, number of new products intro- duced during past 12 months, and cost of research and development are used for evaluating employee’s innovative- ness. By adding more factors into each item, employee’s performance would be evaluated more precisely. The overall performance degree of intellectual capital which is summation of the crisp degree for each element is equal to 3.7989 (scale 0–11. If all the intellectual items of a company are in best degree, the overall performance of intellectual capital will be 11). All the previous work in assessing intellectual capital was based on expert judgment of every intellectual item. Experts evaluate every intellectual item using linguistic variables. This kind of evaluation largely depends on expert’s concept about intellectual items, which can lead to misevaluation in many cases and finally unreliable results. So, it is better to determine indicators which affect every intellectual item. In this way, a set of rule sets is defined which links indictors to each item, and by evaluating the indictors by experts, each intellectual item is evaluated. The proposed expert system in this paper uses a set of rules to evaluate company’s intel- lectual capital items performance and provides companies more reliable results which are independent of individual’s judgment. 5. Conclusions and Future Works In this paper, a fuzzy rule-based expert system is proposed to reveal and evaluate overall performance degree of intel- lectual capital. Through using the proposed expert system, managers can understand the enterprise situation in every item of intellectual capital even if they do not know exactly what every item means, and they also understand whether each item is important for them in decision making or not. Moreover, using linguistic terms for items evaluation leads to more accurate judgment. Recent proposed model [3] asked expert the performance degree and importance degree of each intellectual capital item. In this way, there is always the risk that experts do not have a clear concept for every item in mind and his/her judgments will be done with some inaccuracy. Thus, using a fuzzy expert system which can help experts to improve his/her judgments is useful. A suggestion for future studies is concerned is with different kinds of membership functions for linguistic terms to find the best one. Finally, the model is seen as open for future extension and development, especially extending the expert system by adding more precise and detailed rules to evaluate intellectual capital performance more accurately. References [1] Danish Agency for Trade and Industry, A Guideline for Intel- lectual Capital Statements—A Key to Knowledge Management, 2000. [2] K. M. Wiig, “Integrating intellectual capital and knowledge management,” Long Range Planning, vol. 30, no. 3, pp. 399– 405, 1997. [3] W. S. Tai and C. T. Chen, “A new evaluation model for intel- lectual capital based on computing with linguistic variable,” Expert Systems with Applications, vol. 36, no. 2, pp. 3483–3488, 2009. [4] M. Fasanghari, M. Mohamedpour, and A. Abdollahi, “Assess- ing intellectual capital with fuzzy expert system,” in Pro- ceedings of the 4th International Conference on Networked Computing and Advanced Information Management (NCM ’08), pp. 278–283, September 2008. [5] R. Petty and J. Guthrie, “Intellectual capital: literature review,” Journal of Intellectual Capital, vol. 7, no. 2, pp. 254–271, 2006. Advances in Fuzzy Systems 11 [6] M. Sugeno and T. Yasukawa, “Fuzzy-logic-based approach to qualitative modeling,” IEEE Transactions on Fuzzy Systems, vol. 1, no. 1, pp. 7–31, 1993. [7] L. Edvinsson and M. S. Malone, Intellectual Capital: Realizing Your Company’s True Value by Finding Its Hidden Brainpower, Harper Business, New York, NY, USA, 1997. [8] T. A. Stewart, Intellectual Capital: The New Wealth of Organi- zations, Currency, New York, NY, USA, 1998. [9] K. E. Sveiby, The New Organizational Wealth: Managing and Measuring Knowledge-Based Assets, Berrett-Koehler, San Fran- cisco, Calif, USA, 1997. [10] L. Edvinsson, “Developing intellectual capital at Skandia,” Long Range Planning, vol. 30, no. 3, pp. 366–321, 1997. [11] J. Liebowitz and K. Wright, “Does measuring knowledge make cents?” Expert Systems with Applications, vol. 17, no. 2, pp. 99– 103, 1999. [12] Organization for Economic Co-operation and Development (OECD), Guidelines and instructions for OECD Symposium, Intellectual Symposium Measuring Reporting Intellectual Capital: Experiences, Issues, and Prospects OECD, Amster- dam, The Netherlands, Paris, France, 1999. [13] P. N. Bukh, H. T. Larsen, and J. Mouritsen, “Constructing intellectual capital statements,” Scandinavian Journal of Man- agement, vol. 17, no. 1, pp. 87–108, 2001. [14] J. Chen, Z. Zhu, and H. Xie, “Measuring intellectual capital: a new model and empirical study,” Journal of Intellectual Capital, vol. 5, no. 1, pp. 195–212, 2004. [15] J. P. Ignizio, Introduction to Expert Systems: The Development and Implementation of Rule-Based Expert Systems, McGraw- Hill, 1991. [16] A. Abraham, Rule-Based Expert Systems, Oklahoma State University, Stillwater, Okla, USA, 2005. [17] L. A. Zadeh, “Outline of a new approach to the analysis of complex systems and decision processes,” IEEE Transactions on Systems, Man, and Cybernetics, vol. 3, p. 2844, 1973. [18] H. J. Zimmerman, Practical Applications of Fuzzy Technologies, Kluwer Academic Publishers, 1999. [19] A. Celikyilmaz and I. B. Türkşen, Modeling Uncertainty with Fuzzy Logic: With Recent Theory and Applications, Springer, 2009. [20] J. Durkin, Expert Systems: Design and Development, Macmil- lan, 1994. [21] C. Mount and T. W. Liao, “Prototype of an intelligent failure analysis system,” in Proceedings of the 4th International Conference on Case-Based Reasoning (ICCBR ’01), pp. 716– 731, Vancouver, BC, Canada, 2001. [22] J. E. Burge, “Knowledge Elicitation Tool Classification,” 2011, http://web.cs.wpi.edu/∼jburge/thesis/kematrix.html# Toc41- 7957386. Submit your manuscripts at http://www.hindawi.com Computer Games Technology International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Distributed Sensor Networks International Journal of Advances in Fuzzy Systems Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Reconfigurable Computing Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Applied Computational Intelligence and Soft Computing Advances in Artificial Intelligence Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Advances in Software Engineering Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Electrical and Computer Engineering Journal of Journal of Computer Networks and Communications Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Advances in Multimedia International Journal of Biomedical Imaging Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Artificial Neural Systems Advances in Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Robotics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Computational Intelligence and Neuroscience Industrial Engineering Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Modelling & Simulation in Engineering Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Human-Computer Interaction Advances in Computer Engineering Advances in Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 
work_o5pckqbnsfdtndk4vlsk5e43ge	----	PII: S0957-4174(03)00073-3 Goal-oriented sequential pattern for network banking churn analysis Ding-An Chiang a , Yi-Fan Wang b,*, Shao-Lun Lee a , Cheng-Jung Lin a a Department of Information Engineering, Tamkang University, Tanshui, Taipei, Taiwan, ROC b Department of Information Management, Chang Gung Institute of Technology, Kwei-Shan, Tao-Yuan, Taiwan, ROC Abstract Discovering sequential patterns is one of the most important task in data mining. In this paper we propose an efficient algorithm, called Goal-oriented sequential pattern. It can provide enterprises warning signs soon before they are losing valuable customers and give them reference for decision making. Experiments comparing Apriori showed that Goal-oriented is more efficient, and performs reasonably well for the rules. q 2003 Elsevier Ltd. All rights reserved. Keywords: Data mining; Association rule; Sequential pattern; Goal-oriented; Retention analysis 1. Introduction The Pareto Principle or the 80:20 Rule is commonly employed in customer relationship management in which 80% of the company income comes from 20% of the major customers while 90% of the company income is generated from the basic customers. According to Don Peppers and Martha Rogers, the marketing experts, most enterprises average lost 25% customers annually. How- ever, the cost of obtaining a new customer is five times higher than maintaining an existing customer (Gronroos, 1984; Reich & Benbasat). Nevertheless, the major concerns of the enterprises today are to maintain customers’ loyalty and to find out the possible reasons for the failures of keeping customer loyalty (Reichheld, 1996). Further, to establish a system that can get alert before losing customers in order to take further actions to keep the customer loyalty is also the main consideration. It is very important for every company to fully understand the changes of numbers about their customers and how the changes affect the company. The prime concern for this research is to help the company understand the customer behaviors and list out the possible losing customers in order to retain customers. In this research, we specify how to judge whether a customer is leaving and the retaining strategies. The Sequential Pattern cannot conduct a research function by a specific item; meanwhile, the regulations it figures are relatively irrelevant. This does not only increase the difficulties for interpreting the regulations, but also lack of efficiency. Therefore, to alter the flaw of being lack of efficiency, and the incapacity of searching a specific item, Goal-oriented pattern is adapted in this research to implement the searches for some particular items. We use the association analysis, comprising sequential patterns (Agrawal & Srikant, 1995; Han et al., 2000; Han, Pei, & Yin, 2000; Masseglia, Cathala, & Poncelet, 2001; Pei et al., 2001; Tseng & Hsu, 2001) and association rules (Agrawal, Imielinski, & Swami, 1993; Agrawal & Srikant, 1994; Park, Chen, & Yu, 1995; Srikant & Agrawal, 1995). The cause of leaving customers can be generated by sequential patterns. Consequently, the strategy planners can take proper actions to retain customers. We list the possible reasons of customer leaving by means of reverse sequence in which the original sequence is being reversed so that the main concerns of the item are arranged in front of the array. The contrasts between the goal items are then used to exclude the irrelevant items. Finally, the reversed sequence is re-reverse again to the original sequence in order to enhance the efficiency for searching the goal item rules. 0957-4174/03/$ - see front matter q 2003 Elsevier Ltd. All rights reserved. doi:10.1016/S0957-4174(03)00073-3 Expert Systems with Applications 25 (2003) 293–302 www.elsevier.com/locate/eswa * Corresponding author. E-mail addresses: yfwang@mail.cgit.edu.tw (Y.-F. Wang), chiang@cs. tku.edu.tw (D.-A. Chiang). http://www.elsevier.com/locate/eswa 2. Relative research 2.1. Association rules Association rules are mainly used to find out the relations between items or features that happened synchronously in the database, such as generating the data about the groceries that are bought at the same time when people shop in the mall. For instance, 80% of people who merchandise milk buy bread as well. As soon as the decision maker fetches this information, strategies, such as setting the relevant counters nearby or held a promotion, can be generated from this aspect. Therefore, the main purpose of implementing association rules is to find out the synchronous relationship by analyzing the random data for the reference of making decisions. The definition of association rule is as following: Make I ¼ {i1; i2; …; im} as the itemset, in which each item represents a specific commodity. D stands for a trading database in which each transaction T represents a itemset. That is T # I: Each itemset is a non-empty sub-itemset of I; and the only identify code is TID: Each itemset X , I has a measure standard—Support, to evaluate the statistical importance in D: SupportðX; DÞ denotes the rate of merchandising X in transaction D: The format of the association rule is X ! Y; in which X; Y , I; and X > Y ¼ B: The interpretation of this association rule is that if X is purchased, Y can be bought at the same time. Each rule has a measuring standard called Confidence; i.e. ConfidenceðX ! YÞ ¼ SupportðX < Y; DÞ= SupportðX; DÞ: In this case, ConfidenceðX ! YÞ denotes if the merchandise including X; the chance of buying Y is relatively high. There are two steps to find out the association rules. First, detecting the large itemset. Second, generating the association rules by utilizing the large itemset. Therefore, to explore the association rules also means to find out all the association rules of X ! Y formats and meet the following conditions: 1. SupportðX < Y; DÞ $ Minsup 2. ConfidenceðX ! YÞ $ Minconf The Minsup and Minconf are both set by the users. In general, the numbers of the transactions that comprising X is called the support of X; denoted by sx: Make Minsup the minimum value of support. If the support of X meets the condition, sx $ Minsup; X is the large itemset. As for the exploration of the association rules, many researchers take the Apriori algorithm (Agrawal & Srikant, 1994) supported by Agrawal et al. as the basic formulation. The Apriori algorithm forms the rules by way of simple and sequential progresses to lay out the relationship in the database. It is widely adopted to find out the large itemset, as demonstrated in Fig. 1. By means of this algorithm, each support can be calculated followed by scanning the transaction database. Further, judge if the support exceeds the minimum support so that the large itemset can be identified. In the following rounds, we use the itemset selected previously to be the seed itemset, which can be utilized to generate a new potential large itemset, called Candidate itemsets. Calculate the support of each candidate itemset to decide whether the itemsets can be the genuine large itemset and be the seed itemset for the next round. This algorithm can be ceased when no further candidate itemset can be generated. For example, in Fig. 2, database D is the original transaction database. Assume the minimum support is 2. First, calculate the number of each item that appears in the transaction database, which is to calculate the support and to evaluate whether the number is bigger than or equal to the minimal support and determine the Large 1-itemsets, L1: Next, generate candidate 2-itemsets, C2 from L1 £ L1: Further, calculate the support of C2 to create L2: From L2 £ L2 brings about Candidate 3-itemsets, C3: In the phase of Fig. 1. Apriori algorithm. Fig. 2. Original transaction database. D.-A. Chiang et al. / Expert Systems with Applications 25 (2003) 293–302294 join, we have {40; 70; 80}: However, in the phase of pruning, the sub-itemsets {40; 70}; {40; 80} of {40; 70; 80} are both comprised in L2; but {70; 80} is not enclosed in L2: Thus, it does not meet Candidate 3-itemset C3: The algorithm, therefore, ceased. In consequence, the large itemsets generated are L1 ¼ {10}; {20}; {30}; {40}; {70}; {80}; L2 ¼ {40; 70}; {40; 80} as demonstrated in Fig. 3. In step 2, the association rules are created by use of the large itemsets. As the large itemsets figured in the previous step are L1 ¼ {10}; {20}; {30}; {40}; {70}; {80}; L2 ¼ {40; 70}; {40; 80}: Therefore, the association rules we intend to find out are {40} ! {70}; {70} ! {40}; {40} ! {80} and {80} ! {40}: 2.2. Sequential patterns The sequential patterns originate the association rules that occur frequently. It emphasizes the factor of time. What being concerned is the data that related on the base of time. We use this pattern to analyze the association of each individual event chronologically. The major method of mining sequential patterns are Apriori algorithm (Agrawal & Srikant, 1994), DHP algorithm (Park et al., 1995), FP-tree algorithm (Han et al., 2000), and FPL algorithm (Tseng & Hsu, 2001). The sequential pattern can be divided into Sequential Procurement (Agrawal & Srikant, 1995) and Cyclic Procurement (Ozden, Ramaswamy, & Silberschatz, 1998) by the sequence and the section of time. What sequential procurement concerns is the chronological pattern. It only analyses the association of data by the sequence of time. On the other hand, what cyclic procurement concerns are the events that happen in the same time section and whether they reoccur in other time section as well. It emphasizes the variation in different time section. The main concern in this research is on sequential procurement. In general, there are two phases to create the sequential pattern: Setting large itemset and establishing sequential pattern rules by large itemset. As the association rule, the meaning of sequential patterns is similar to the large itemset, which is to find out the association between items. The only difference is that in sequential patterns, the chronological pattern is one of the major concerns. In phase I, Apriori algorithm introduced in Fig. 1 is applied to find the large itemset. Whereas in phase II, there are five phases to generate sequential pattern rule from the large itemset. Sort Phase. Customer series is the major figure while transaction time is the sub-major figure to arrange the raw data from the original transaction database. The raw transaction database is demonstrated in Fig. 2. The new database after sorting incrementally by time is showed in Fig. 4. Large Itemset Phase. Find out all the large itemsets and denote each large itemset with a specific symbol. Assume the minimum support is 2, from Section 2.2, we get the large itemset as L1 ¼ {10}; {20}; {30}; {40}; {70}; {80}; L2 ¼ {40; 70}; {40; 80}: Then code them with an integer as showed in Fig. 5. Transformation Phase. Delete the non-large itemsets from the raw transaction database. The items not in the large itemset from the raw transaction database are deleted and rearranged data by customers as shown in Fig. 6. Sequential Phase. Delete the sequences that each contains two items from the data. The sequences can be generated by algorithm similar to Apriori. The difference from finding the large itemset is in Fig. 4. The customer sequence is ordered by increasing transaction time. Fig. 3. Generation of candidate itemsets and large itemsets. D.-A. Chiang et al. / Expert Systems with Applications 25 (2003) 293–302 295 this phases, the factor of sequence is being concerned. Hence, the combination of candidate sequence of two items is square of these of single item when we want to find the combination sequence from one item to two items. Similarly, we also can use Apriori-gen Algorithm to find out the candidate sequence if the items are above three. The LS1 conducted by all large itemsets is shown in Fig. 7. In order to produce CS2, we proceed L1 £ L1; which needs 64 combinations (8 £ 8). We can get LS2 by calculating supported level of CS2. Similarly, we can get CS3 via LS2 £ LS2. Although we generate k3,4,6l, k3,4,8l and k3,6,8l in the merging phase, k4,6l, k4,8l, k6,8l (the subset of k3,4,6l, k3,4,8l and k3,6,8l, respectively) are not included within L2 in pruning phase. This cannot satisfy the 3-items’s candidate sequence (CS3) and then algorithm will stop. Maximal Phase. Find out the maximal sequences within the large sequences. To link up sequence phase to reduce the time of counting non-maximal sequences. The sequence generated in the previous step is not contained in others is the largest sequence (Fig. 7). Sequence k3,4l,k3,6l and k3,8l are the largest sequence. Transforming the largest sequence to the original items, we can get our mining sequence pattern (k(30)(40)l, k(30)(80)l, k(30)(40,80)l). 3. Problem statement At the present business field, creating a customer retention analysis model is the first step to start customer relation managements. Many businesses, therefore, hope to figure out the real cause of the loss of a customer, or even to be told that they are about to lose the customer by some clues before it occurs, and then they can propose or make some new sales strategies against the loss in advance. Although many businesses eager to create a customer-losing model to solve the problem of the loss of customers, they cannot find an effective way to solve the problem that has been confused them for a long time. This research, therefore, proposes a new method to solve such a problem, and according to the method, we can find out behavior patterns of losing customers or clues before they stop using some products through mining Sequential Patterns. In this section, we will take network banking services as an example to elaborate this new method, and moreover, to identify the real causes of the loss of customers in different trades. 3.1. Definition of loss At first, we will make a simple definition for the analysis of the loss of customers in network banking services. We focus on each user who applied for network banking services to find out the periodicity of transaction time of the user. Generally, there are two ways for us to judge whether or not a customer is lost: One is judging by the average periodic transaction days of the customer in the past. If the interval between the present day and the last transaction day is longer than the average interval of transaction days in the past, we can tell that the customer has been lost. The other is judging by the longest interval from one transaction day to another in the past. We can also tell the loss of a customer, if the total of no transaction days since the last transaction day is longer than the longest interval between two transaction days. 3.2. Methodology In this section, we offer a method to solve the constant loss of customers in network banking services, that is, to find out main reasons why customers stop using the network banking services through Sequential Patterns Algorithm and the definition of loss mentioned in Section 3.1. To identify the real cause of the loss of the customer, common sequential patterns cannot entirely be applied to it. Because if we put all of the transaction data into the program of sequential patterns, it will cost us lots of time to calculate to find out a rule, and even if we have found one, the chances are that it is useless rule. Fig. 5. Large itemsets. Fig. 6. Transformed database. D.-A. Chiang et al. / Expert Systems with Applications 25 (2003) 293–302296 3.2.1. Time windows and normalization For solving the problem described above, we use the concept of Time windows and Normalization to select the data we want. The size of the time window will affect many factors, in other words, the larger the window is, the more the transaction data we have to deal with, the more time we have to spend on calculation, and the more related rules we can find out. The supportability rating, therefore, may be raised. Conversely, the smaller the window is, the less or even none related rules would be found. Firstly, we have to decide the size of the time window. For example, one year, six months, one-quarter, one month, or one week, etc. Next, in accordance of the characteristics of the problem, we use a way to set time windows that is different from other sequential patterns. For instance, supposing that we set the time window size as one month, it means that we only select the transaction record of each customer in the last month from the database. It helps us a lot to differentiate the real cause of the loss of the customer, for the transaction activity of each customer before loss is what we are concerned, and these activities will reflect on the transaction data in the last month. Thus, by the concept of Time Windows and Normalization, we can get a better effect in finding a sequential pattern rule. After the time window was selected (e.g. one month), it is very common to select the transaction data for one month from the entire data. However, in order to find out the real factor of the loss of the customer, we use the idea of Normalization. We regard the last transaction date of each customer as a datum point, and so different customers have different datum points. We select the data that is traced back for one month from the datum point of each customer. As shown in Fig. 8, we take the vertical axis as the customer, the horizontal axis as the transaction date, dark color as the selected data, and light color as the unselected data. The numbers in dark color represent the number of transaction times. When the time window is set as one month, the last transaction date of customer ID0001 is on 15/11/2001, then the data we selected are four transactions from 16/10/2001 to 15/11/2001; the last transaction date of customer ID0002 is on 16/08/2001, then the data we selected are the five transaction data from 17/07/2001 to 16/ 08/2001; the last transaction date of customer ID003 is on 20/04/2001, then the data we selected are the three transactions between 21/03/2001 and 20/04/2001, and so are the rests on analogy of this. 3.2.2. Data preprocessing and virtual labels As among the four transactions of customer ID0001, ways of practicing general sequential patterns is regard Fig. 8. Use time window to extract data. Fig. 7. Generation of candidate sequences and large sequences. D.-A. Chiang et al. / Expert Systems with Applications 25 (2003) 293–302 297 these four descriptions of the transactions as a single record, and then we induce rules by using sequential pattern algorithm. However, due to the characteristics of the problem, we add the concept of virtual labels before we delete duplicated descriptions of transactions. For general sequential patterns cannot find out number of times of every transaction activity when dealing with these four transaction data, it is surely that some information will be omitted by means of such calculation. Thus, in order to retain key information, we put some key descriptions of transaction different virtual labels. Take the customer as an example, if there are two repeated transactions described as ‘login in failure’, at the same period, we will combine these two consecutive data and label it a new description ‘login in failure 2’, which means two repeated attempts are made to sign on with incorrect PIN. Corresponding, if there are three repeated transaction descriptions of the customer, which are described as login in failure, at the same period, we will combine these three consecutive data and label it a new description ‘login in failure 3’, which means PIN is entered incorrectly three consecutive times. These descriptions like Login in failure 2 and Login in failure 3 are virtual labels that separately represent some important information such as the failures may be caused by the user who forgot the PIN or hackers who attempt to enter into the system. 3.3. Goal-oriented sequential patterns Traditional sequential patterns cannot excavate specific goal items; however, for most of the decision makers, they have a need for finding out the happening order of some specific items. Therefore, we propose a method to deal with the problem that happens when we search the specific goal items. Generally, if the goal items appear at the top of the sequence, we only have to delete the large itemset that is not related to the goal items and then make the rules. Relatively, if the goal items appear at the bottom of the sequence, then we have to use the concept of ‘the reversed sequence’ to reverse the original sequence and make the items we concern at top of the whole sequence; after that, comparing the goal times and sifting the rules irrelevant to the goal items to increase the rule-finding efficiency. In addition, if the goal items appear at the middle of the sequence, we firstly delete the sequences after the goal items, and then follow the steps as mentioned in above. As mentioned in Section 2.2, a general way of finding sequential patterns can be resolved into two steps. The first step is to find out large itemset. The second step is use the large itemset to conclude sequential patterns rules. At the first step, we use Apriori calculation to find out the large itemset, while at the second step as to use the large itemset to induce rule, we propose a new algorithm to coincide with the characteristics of the problem that is to find out the rules of loss. The algorithm can delete some rules that are irrelevant to the goal item in advance, and the efficiency of finding rules is increased by a big margin herein. Although the five steps of the traditional sequential patterns mentioned above can also find out the rule of loss, it is a very inefficient way because we will find many rules, which are mostly useless ones that irrelevant to the loss. Because the purpose of this research is to find out the rule of loss, in other words, to find out the transaction activity of the customer before losing, to deal with the problem like this, we propose a new sequential patterns called Goal-oriented algorithm to solve such a problem towards a specific goal. The problem of mining Goal-oriented sequential patterns is split into six phases. Descend Sort Phase. The sort phase converts the original transaction database into a database of customer sequences. The customer sequence is a list of transactions ordered by decreasing transaction time. Delete the sequences when it was not beginning with goal item. We can get a new database (Fig. 9) from decreasing sorting raw one (Fig. 2) by time, if we set the target as 80 and omit the sequence that does not start from the target item. Large Itemset Phase. Find out all the large itemsets and denote each large itemset with a specific symbol. Assume the minimum support is 2, we get the large itemset as L1 ¼ {30}; {40}; {80}; L2 ¼ {40; 80}: Then code them with an integer as shown in Fig. 10. Fig. 9. The customer sequence is ordered by decreasing transaction time. Fig. 10. Large itemsets. D.-A. Chiang et al. / Expert Systems with Applications 25 (2003) 293–302298 Transformation Phase. Delete the non-large itemsets from the raw transaction database. The items not in the large itemset from the raw transaction database are deleted and rearranged data by customers as shown in Fig. 11. Sequential Phase. Delete the sequences that each contains two items from the data. Whereas in phase II, the numbers of large itemsets, which contains target item (80) are 2. One is large itemset (80), the mapping value is 3, the others is (40,80), which mapping value is 4. The LS1 conducted by all large itemsets. In order to produce CS2, we proceed L1 £ L1: Because of the numbers of large itemsets, which contain target item (80) are 2, so it only needs eight combinations ð4 £ 2 ¼ 8Þ: The result is shown in Fig. 12. Comparison with traditional method (Section 2.2), it needs compute 64 times, there is a major difference between them. And then we can get LS2 by calculating supported of CS2. Similarly, we can get CS3 via LS2 £ LS2, because we cannot generate CS3 in the merging phase, the algorithm will stop. Maximal Phase. Find out the maximal sequences within the large sequences. To link up sequence phase to reduce the time of counting non-maximal sequences. The sequence of k3,1l and k4,1l is the large sequences (Fig. 12). Transforming the largest sequence to the original items, we can get our mining sequence pattern k(80)(30)l and k(40,80)(30)l. Reverse Phase. Reverse the maximal sequences. Reverse the maximal sequences k(80)(30)l and k(40,80)(30)l, and we get the sequences k(30)(80)l and k(30)(40,80)l. Therefore, the goal oriented sequential pattern rules we intend to find out are k(30)(80)l and k(30)(40,80)l. 4. The experiment results To examine the executing efficiency of the Goal- oriented sequential pattern algorithm, we separately design some experiments to examine our method, and compare it with Apriori algorithm. At last, we will explain the result of our experiment. In Section 4.1, we will introduce experimental environments, including the exper- iment desktop, the practicing program language and information resources. In Section 4.2, we will introduce the experiment results and compare the executing efficiency of the Goal-oriented sequential patterns algor- ithm. In Section 4.3, we will introduce the customer loss analysis result. 4.1. Experimental environments 1. Experimental platform * CPU: Pentium III 600 MHz * Memory: 392 MB RAM’13.9 GB HD * Operating System: Window 2000 * Database: SQL Server 2000 2. Implementation Language * Visual Basic 6.0 3. Experimental Data Source * We analyze the customers’ data from a famous network banking in Taiwan, total more than one million data. These data are banks’ customers more than 6 months transaction data. 4.2. The executing efficiency analysis Fig. 13 shows the executing frame of Apriori algorithm, and we donot have to restrict its items of producing rules, while the executing frame of Goal-oriented algorithm as shown in Fig. 14, we have to restrict its items of producing rules. The program offers two choices: One is to begin at one target item; the other is to end at one target item. The customer loss model offered in the thesis utilizes the latter choice. We focus on the target item ‘Termination’ as the ending rule, so that we can know what activities happened before the termination. In order to compare the executing efficiency between these two algorithms, we utilize four different supports 0:1; 0:2; 0:3; 0:4; and examine them under the condition of Fig. 11. Transformed database. Fig. 12. Generation of candidate sequences and large sequences. D.-A. Chiang et al. / Expert Systems with Applications 25 (2003) 293–302 299 allowing 5 as the max ruling length. The result was shown in Fig. 15. Compare with these two algorithms, in the same minimum support 0.1 and different data set, the result was shown in Fig. 16. It indicates that the executing efficiency of the Goal-oriented algorithm is evidently better than that of the Apriori algorithm, and it is more apparently as there are more activities and numbers of item and the minimum support comes to less. The numbers of executing efficiency between these two algorithms can up to multi-ten times difference. Fig. 14. The executing frame of goal-oriented algorithm. Fig. 13. The executing frame of apriori algorithm. D.-A. Chiang et al. / Expert Systems with Applications 25 (2003) 293–302300 As in Fig. 17, which shows the numbers of rules these two algorithms generated, the Apriori algorithm produces more rules though, it makes users not so easy to read and check. Compare with these two algorithms, in the same minimum support 0.1 and different data set, the result was shown in Fig. 18. As in Fig. 13, if we do not restrict an item, it contains all the rules, including the rules we concern and those we do not concern. It is difficult for us to read. Contrarily, the numbers of rules that are produced by the Goal-oriented sequential patterns focus on the rules relate to the target item, and therefore have a better readability. In Fig. 14, we restrict each rule to end up with the description ‘transaction inquiry’, and therefore, the last description of each rule is ‘transaction inquiry’. It gets us easy to read. 4.3. The customer loss analysis result After we use Goal-oriented sequential patterns algor- ithm to process the customer loss analysis, the rule of loss will be produced as shown in Fig. 19; take the model login in failure 2 as an example, 77.894% of the customers who accord with the situation do not enter into the system over 30 days, that is to say that it is very possible for them to stop using the service. Login in failure 2 indicates the meaning of entering incorrect PIN two times, and due to the agreements of the banking services, the bank will disable the service if three consecutive attempts are made to sign on with incorrect PIN, such a regulation causes the customers choose to terminate the services. According to the loss analysis data, we can find out the customers of this type, if the bank agent who is responsible for the customer service can take the initiative in contacting the customer and solving the problem in time, then the efficiency of constant use of the services is increased by a big margin, it also applies to other customers in other models. While the bank is blindly promoting its sales to get more new users, it does not realize that the existing customers are losing gradually. For the bank, such a gradual loss is larger than the benefits of getting new customers. Fig. 19 shows the main reasons for the customer loss after the analysis. We list the first 20 rules in proportion to the termination rate, and we give the customer center the name lists of the customers who are classified by the causes. The center then can trace the losing customers, and because the customer activity is subject to change, we can use the system to add new information and renew the rules of loss as we find out the rate of the causes of these rules is decreased by a big margin. Fig. 15. Execution time (same data set and different minimum support). Fig. 16. Execution time (same minimum support and different data set). Fig. 18. Number of rules (same minimum support and different data set). Fig. 17. Number of rules (same data set and different minimum support). D.-A. Chiang et al. / Expert Systems with Applications 25 (2003) 293–302 301 5. Conclusion In this paper, in order to set up the customer-losing pattern, we present a novel method to find out the activities of the losing customers by the excavating sequential pattern algorithm. Moreover, we propose another new algorithm called the Global-oriented algorithm in search of specific goal items. The most distinct difference between this algorithm and the traditional one is that we use the concept of reversed sequence. We firstly reverse the original sequence, and then sort the items we concerns and put them to the top of the whole sequence, next, comparing goal items to sift the rules which are irrelevant to the goal items. The algorithm makes the rule-finding efficiency increased by a big margin. Besides, the rules concluded by the algorithm are simple, which result from the goal items we concern; and the rules, therefore, are easier to read. Finally, we attempt to find out the main reasons why the customers terminate the services through the Goal-oriented algorithm. The results show that up to 80% of the terminated users, the cause of the termination is just because they repeated entered incorrect PIN for two times. The customers stop using the network banking services or become a static account in consequence. After obtaining such information, with some timely solutions, businesses can effectively prevent the loss from happening. References Agrawal, R., Imielinski, T., & Swami, A. (1993). Mining association rules between sets of items in large database. Proceedings of the ACM SIGMOD international conference on management of data, 207 – 216. Agrawal, R., & Srikant, R. (1994). Fast algorithms for mining association rules. Proceedings of the 20th international conference on very large data bases, 478 – 499. Agrawal, R., & Srikant, R. (1995). Mining sequential patterns. Proceedings of the 11th international conference on data engineering, 3 – 14. Gronroos, C. (1984). A service quality model and its marketing implication. European Journal of Marketing, 18(4), 36 – 44. Han, J., Pei, J., Mortazavi-Asl, B., Chen, Q., Dayal, U., & Hsu, M.-C. (2000a). Freespan: frequent pattern-projected sequential pattern mining. Proceedings of the international conference of data mining and knowledge discovery, 355 – 359. Han, J., Pei, J., & Yin, Y. (2000b). Mining frequent patterns without candidate generation. Proceedings of the ACM SIGMOD conference on management of data, 241 – 250. Masseglia, F., Cathala, F., & Poncelet, P. (1998). The PSP approach for mining sequential patterns. Proceedings of the second european symposium on principles of data mining and knowledge discovery, 1510, 176 – 184. Ozden, B., Ramaswamy, S., & Silberschatz, A. (1998). Cyclic association rules. In Proceedings of the 14th international conference on data engineering, 412 – 421. Park, J. S., Chen, M. S., & Yu, P. S. (1995). An effective hash based algorithm for mining association rules. Proceedings of the ACM SIGMOD conference on management of data, 175 – 186. Pei, J., Mortazavi-Asl, B., Pinto, H., Chen, Q., Dayal, U., & Hsu, M.-C. (2001). Prefixspan: mining sequential patterns efficiently by prefix projected pattern growth. Proceedings of the international conference of data engineering, 215 – 224. Reich, B. H., & Benbasat, I (2003). An empirical investigation of factors influencing the success of customer-oriented strategic system. Infor- mation Systems Research, 1(3) 325 – 347 Reichheld, F. (1996). The loyalty effect: The hidden force behind growth, profit and lasting value. Cambridge: Harvard Business School Press. Srikant, R., & Agrawal, R. (1995). Mining generalized association rules. Proceedings of the 21th international conference on very large data bases, 407 – 419. Tseng, F. C., & Hsu, C. C. (2001). Generating frequent patterns with the frequent pattern list. Proceedings of the asia pacific conference of data mining and knowledge discovery, 376 – 386. Fig. 19. Main losing mode. D.-A. Chiang et al. / Expert Systems with Applications 25 (2003) 293–302302 Goal-oriented sequential pattern for network banking churn analysis Introduction Relative research Association rules Sequential patterns Problem statement Definition of loss Methodology Goal-oriented sequential patterns The experiment results Experimental environments The executing efficiency analysis The customer loss analysis result Conclusion References 
work_o632qjptwnb6pdz5sfk3z3jm6a	----	This article was published in an Elsevier journal. The attached copy is furnished to the author for non-commercial research and education use, including for instruction at the author’s institution, sharing with colleagues and providing to institution administration. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy Adaptive wavelet network for multiple cardiac arrhythmias recognition Chia-Hung Lin a,b,*, Yi-Chun Du b, Tainsong Chen b a Department of Electrical Engineering, Kao-Yuan University, Kaohsiung, Taiwan b Institute of Biomedical Engineering, National Cheng-Kung University, Tainan, Taiwan Abstract This paper proposes a method for electrocardiogram (ECG) heartbeat detection and recognition using adaptive wavelet network (AWN). The ECG beat recognition can be divided into a sequence of stages, starting with feature extraction from QRS complexes, and then according to characteristic features to identify the cardiac arrhythmias including the supraventricular ectopic beat, bundle branch ectopic beat, and ventricular ectopic beat. The method of ECG beats is a two-subnetwork architecture, Morlet wavelets are used to enhance the features from each heartbeat, and probabilistic neural network (PNN) performs the recognition tasks. The AWN method is used for application in a dynamic environment, with add-in and delete-off features using automatic target adjustment and parameter tuning. The experimental results used from the MIT-BIH arrhythmia database demonstrate the efficiency of the proposed non-invasive method. Compared with conventional multi-layer neural networks, the test results also show accurate discrimination, fast learning, good adaptability, and faster processing time for detection. � 2007 Elsevier Ltd. All rights reserved. Keywords: Electrocardiogram (ECG); Adaptive wavelet network (AWN); Cardiac arrhythmia; Morlet wavelet; Probabilistic neural network (PNN) 1. Introduction Bio-signals on the surface of the human reflect the inter- nal status and electric activity, thus providing information on internal organs with non-invasive measurement such as ECG, echocardiogram, or scintigraphy (Chen, Chen, & Tsai, 1997; Silipo & Marchesi, 1998). ECG has commonly used to collect large amounts of measurements that contain particular information in the signals. The typical diagnostic method is off-line analysis from the recorded data, and using a cardiogram to identify arrhythmic types of the patients. Designing non-invasive tools, abnormal monitor- ing techniques, signal processing, and classification capa- bility for stationary/portable instruments has become an essential task. In addition, further integrating several tech- niques such as data analysis, pattern detection and discri- mination, decision support, and human computer interface is also necessary (Dickhaus & Heinrich, 1996; Qin et al., 2003). To ensure accurate detection, the algorithm requires real-time automatic classification, non-invasive, high- performance computing technique, reliable solutions, and simple for diagnosing disturbances of cardiac rhythm. Diagnostic approaches have been applied to detection with frequency-domain and time-domain techniques. The QRS complex in ECG signals varies with the origination and the conduction path of the activation pulse in the heart- beat. When the activation pulse does not travel through the normal conduction path, the QRS complex becomes wide, and high-frequency components are attenuated. Power spec- tra of individual QRS complex are found at frequencies between 4 Hz and 20 Hz (Minami, Nakajima, & Toyoshima, 1999). With the time-domain technique, various features from each heartbeat are extracted to detect arrhythmia waveforms, such as width, height, area of QRS complex, and QRS morphology, etc. (Osowski & Linh, 2001). Apply- ing these particular features, artificial intelligence (AI) approaches are used to perform ECG beat recognition. Var- ious architectures for artificial neural networks (ANN) are 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.05.008 * Corresponding author. Address: Department of Electrical Engineer- ing, Kao-Yuan University, Kaohsiung, Taiwan. Tel.: +886 7 6077014. E-mail address: eechl53@educities.edu.tw (C.-H. Lin). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 34 (2008) 2601–2611 Expert Systems with Applications Author's personal copy used in this research, such as wavelet neural networks (Dick- haus & Heinrich, 1996), back-propagation networks (Ach- arya, Kumar, Bhat, Lim, & Lyengar, 2004; Minami et al., 1999; Silipo & Marchesi, 1998), Fuzzy hybrid neural net- works (Osowski & Linh, 2001; Wang, Zhu, Thakor, & Xu, 2001), self-organizing map network (Hu, Palreddy, & Tompkins, 1997). To develop a diagnostic method for arrhythmia classifi- cation, signal analysis based on wavelet transform (WT) has been presented to extract the characteristics of ECG signals (Dickhaus & Heinrich, 1996; Qin et al., 2003). The wavelet coefficients represent measures of similarity of the local shape of the signal to the mother wavelet under different dilation and translation parameters. This analysis is robust to time-varying signal analysis and it can point out occurrence time, but it is not capable of recognition. With multiresolution and localization of the wavelets (Lin & Wang, 2006) and pattern recognition capability of the ANN, wavelet network (WN) has become important for signal analysis and pattern recognition. When WN is applied in the real word, for example, the morphology vari- ations of ECG waveforms are different for different patients, and even for the same patient or for the same type (Osowski & Linh, 2001), traditional networks can become a bottleneck requiring retraining with new features added into the current database. PNN (Specht et al., 1988) and general regression neural networks (GRNN) (Masters & Land, 1997; Seng et al., 2002) have been presented, and are recognized as having expandable or reducible network structure, fast learning speed, and promising results. In these adaptation methods, the choice of smoothing para- meter has significant effects on the network outcome, and the choice of parameter is usually based on the overall sta- tistical calculation from pre-collected training data. A dynamic model needs a non-statistical method as well as automatic adjustment of the targets and smoothing para- meters for dynamic process technique (Masters & Land, 1997; Seng et al., 2002). In this paper, AWN is proposed to recognize normal beat and six cardiac arrhythmias. The WN consists of two subnetworks connected in cascade. In the wavelet layer, the activation functions take the Morlet wavelets and are responsible for extracting features from each ECG signal. Subsequently, the PNN is an adaptive net- work with automatic tuning parameters and is used to clas- sify cardiac arrhythmias. The results show computational efficiency and accurate recognition. 2. Adaptive wavelet network (AWN) 2.1. Morlet wavelet In applications of signal analysis, it is necessary to extract signal features. Fourier analysis consists of break- ing up a signal into sinusoidal waves of various frequencies, but it is only a time-domain transform, which has no time– frequency localization features. Similarly, wavelet analysis is the breaking up of a signal into dilations and translation versions of the original wavelet, referred to as the mother wavelet. The wavelets must be oscillatory, have amplitudes that quickly decay to zero, and have at least one vanishing moment. The Morlet wavelet is the modulated Gaussian function, the family function is built starting from the fol- lowing complex Gaussian function (Lin & Wang, 2006) wðxÞ¼ ejx0 x � e� x2 0 2 � � e� x2 2r; r > 0: ð1Þ The Fourier transform of w(x) is ŵðxÞ¼ ffiffiffiffiffiffiffiffi 2pr p e �ðx�x0Þ 2 2r � e� x2 0 2 e� x2 2r � � : ð2Þ Let x = 0; then ŵð0Þ¼0, that is R R wðxÞdx ¼ 0, which rep- resents the collection of all measurable functions in the real space, and w(x) satisfies the admissibility condition. When x0 P 5, the appropriate Morlet wavelet becomes uðxÞ¼ ejx0 xe� x2 2r ¼ ½cosðx0xÞþ j sinðx0xÞ�e� x2 2r ¼ uRðxÞþ juIðxÞ; x0 P 5; r > 0; ð3Þ where uR(x) and uI(x) are the real and imaginary part respectively. Morlet wavelet is modulated Gaussian func- tion by cosine. It has a better time–frequency localization feature and smoothes noise interference. When u(x) 2 L 2(R), then mother wavelet becomes the daughter wavelet ud,t(x) with dilation parameter d and translation parameter t, as Eq. (4) ud;tðxÞ¼ uRd;t x � t d � � þ juId;t x � t d � � ; d 2 R nf0g; t 2 R; uRd;tðxÞ¼ cos 5ðx�tÞ d h i e �ðx�tÞ 2 2rd 2 ; uId;tðxÞ¼ sin 5ðx�tÞ d h i e �ðx�tÞ 2 2rd2 : 8>< >: ð4Þ Fig. 1 shows the wavelets with various dilation parameters ðd ¼ 1; 2; 3Þ and translation parameters (t = �1, 0, 1). In this study, both real and imaginary parts of the ud,t(x) can use to extract features from the ECG signals. Real part wavelets are applied to extract the features of normal beat (•) and cardiac arrhythmia disturbances including prema- ture ventricular contraction (V), atrial premature beat (A), right bundle branch block beat (R), left bundle branch block beat (L), paced beat (P), and fusion of paced and normal beat (F). The activation functions of the wavelet nodes are derived from the mother wavelet udi,ti(xi) for i = 1, 2, 3, . . . , n, where n is the number of the wavelet nodes. The input vector X ¼ ½x1; x2; x3; . . . ; xi; . . . ; xn� is con- nected to the WN, and inputs are the sample data from the QRS complexes as shown in Fig. 2. 2.2. Adaptive probabilistic neural network Wavelets hybrid ANN is proposed to detect arrhythmia disturbances, and WN combines the properties of the 2602 C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 Author's personal copy Morlet wavelets with the advantages of PNN. The second subnetwork with hidden, summation, and output layer is shown in Fig. 2. The number of hidden nodes Hk ðk ¼ 1; 2; 3; . . . ; KÞ is equal to the number of training exam- ples, while the number of summation nodes Sj and output nodes Oj ðj ¼ 1; 2; 3; . . . ; mÞ equals to the types of distur- bances. The weights wWHki (connecting the kth hidden node and the ith wavelet node) and wHSjk (connecting the jth sum- mation node and the kth hidden node) are determined by K input–output training pairs. The final output of node Oj is (Masters & Land, 1997; Seng et al., 2002) H k ¼ exp � XK k¼1 ðuiðxiÞ� wWHki Þ 2 2r2k " # ; ð5Þ OjðuÞ¼ PK k¼1w HS jk H kPK k¼1H k ¼ sðuÞ hðuÞ ; ð6Þ where r1 ¼ r2 ¼ �� � ¼ rk ¼ �� � ¼ rK . However, there are no means of generating an optimal smoothing parameter rk, and adjusting the rk would refine the accuracy in the dynamic environment. The optimal parameter rk is in- tended to minimize the object function, which is defined as squared error function ej (Masters & Land, 1997) ejðu; TÞ¼ ½T j � OjðuÞ� 2 ; ð7Þ where Tj is the desired output for input vector X, and the implicit constraint is rk 5 0. The first partial derivatives of error are shown in Eq. (8) �oejðu; TÞ ork ¼ 2 T j � OjðuÞ oOjðuÞ ork ¼ 2 T j � OjðuÞ osðuÞ ork � OjðuÞ ohðuÞork hðuÞ " # ; ð8Þ osðuÞ ork ¼ 2 XK k¼1 wHSjk H k " # ðuiðxiÞ� wWHki Þ 2 2r3k " # ; ð9Þ ohðuÞ ork ¼ 2 XK k¼1 H k " # ðuiðxiÞ� wWHki Þ 2 2r3k " # : ð10Þ Fig. 1. The wavelets with various dilations and translations. Wavelet Nodes x 1 x2 x n-1 x n . . . . . . x i . . . . . . . . . 1ϕ ϕ 2 1nϕ nϕ H 1 H k H K S1 S 2 S m . . . iϕ O1 O 2 O m kH . . . . . . weight=1Hidden Nodes Summation Nodes Output Nodes Tune Smoothing Parameters Desired Output Error Signal ECG Signal ∑ ∑ Fig. 2. Architecture of the adaptive wavelet network. C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 2603 Author's personal copy The gradient method is used to update parameter rk with iteration process, as in Eq. (11) rkðp þ 1Þ¼ rkðpÞþ g oejðu; TÞ ork ; ð11Þ where g is the learning rate, and p is the iteration number. In this study, AWN based algorithm contains two stages: the learning stage and recalling stage, as detailed below (Lin & Wang, 2006). 2.2.1. Learning stage Step (1) For each training example XðkÞ¼ ½x1ðkÞ; x2ðkÞ; . . . ; xiðkÞ; . . . ; xnðkÞ� for k ¼ 1; 2; 3; . . . ; K, and i ¼ 1; 2; 3; . . . ; n, create weights wWHki between wavelet node ui and hidden node Hk by uiðkÞ¼ cos 5ðxiðkÞ�tiÞ d i h i e �ðxiðkÞ�tiÞ 2 2rd2 i xi ¼ V ðiÞ; ti ¼ V norðiÞ 8>< >: ; ð12Þ wWHki ¼ uiðkÞ; ð13Þ where di is the dilation parameters, di 2 Z; xi is a se- quence of samples obtained from the QRS complex of unknown signal V; ti is the translation parameters, a se- quence of samples obtained from the QRS complex of normal beat Vnor; n is the number of sampling points; and WWH = [wWHki ] is a k by n matrix. Step (2) Create weights wHSjk between hidden node Hk and summation node Oj by wHSjk ¼ 1 0 � ðj ¼ 1; 2; 3; . . . ; mÞ; ð14Þ where the values of wHSjk are the predicted outputs asso- ciated with each stored pattern wWHki , and W HS = [wHSjk ] is k by m matrix. Connection weights from hidden nodes Hk to summation node P are set 1. 2.2.2. Recalling stage Step (1) Get network weights wWHki and w HS jk . Step (2) Apply test vector X ¼ ½x1; x2; x3; . . . ; xi; . . . ; xn� to the AWN. Compute the output of wavelet node ui uiðxiÞ¼ cos 5ðxi � tiÞ d i � � e �ðxi�tiÞ 2 2rd2 i ð15Þ Step (3) Compute the output of hidden node Hk by using Eq. (5). The optimal value rk can be obtained by using Eqs. (8)–(11) based on minimum misclassifi- cation error (Convergent Condition: Infinity Norm). Step (4) Compute the outputs of node Oj by using the Eq. (6). 3. Cardiac arrhythmias detection procedure 3.1. ECG characteristic and feature extraction An ECG signal represents the changes in electrical potential during the heartbeat as recorded with non-inva- sive electrodes on the limbs and chest; a typical ECG signal consists of the P-wave, QRS complex, and T-waves. The P- wave is the result of slow-moving depolarization of the atria. The rapid depolarization of the ventricles results in the QRS complex of the ECG, which is a sharp wave about 1 mV amplitude and 80–100 ms duration. The plateau part of the action potential after QRS is called the ST segment (Silipo & Marchesi, 1998). The ECG waveform is the result of the potential change that propagates within the heart (electrical activity) and causes the cardiac muscle contrac- tion, even varying with rhythm origin and conduction path. For example, when the activation pulse originates in the atrium and travels through the normal conduction path, the QRS complex has a sharp and narrow deflection, or else the QRS complex becomes wide and distorted (Minami et al., 1999). In the time domain, the normal beat and typ- ical arrhythmia heartbeats are normalized as shown in Fig. 3. Each ECG signal has various morphological infor- mation and features, which can be used to classify seven categories. The QRS complex of the ECG is important information in heart-rate monitoring and cardiac diseases diagnosis. The R-waves are detected by a peak detection algorithm, which begins by scanning for local maxima in the absolute value of ECG data. For certain window duration, the searching continues to look for a larger value. If this search finishes without finding a larger maximum, the current maximum is assigned as the R peak (Minami et al., 1999). Centered on the detected R peak, the QRS complex portion is extracted by applying a window of 280 ms, and P-wave and T-wave are excluded by this window duration. Based on 360 sampling rate, 100 samples can be acquired around the R peak (Sampling point n = 100, 50 points before and 50 points after). After sampling and performing analog-to-digital conversion, individual QRS complexes are extracted. The real part of wavelets uRd,t(xi), d = 3, i = 1, 2, 3, . . ., 100, as Eq. (12), are responsible for extract- ing features under low frequency analysis, and these fea- tures are reconstructed by 100 wavelet nodes to form the symptomatic patterns, as shown in Fig. 4. For example, the symptomatic pattern of a normal beat has a near rectangular-impulse-sequence graph with amplitude one. The morphology of symptomatic patterns will reveal the serious dip and bumpy shapes for abnormal heartbeats. Symptomatic patterns of the same categories have similar morphology or multiform. The amplitudes having a spe- cific dip range can be observed between zero and one in the rectangular section. According to the morphology, various patterns indicate different cardiac diseases. These symptomatic patterns are considered for training the AWN. 2604 C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 Author's personal copy 3.2. Training patterns creation In this study, the dataset of QRS complexes typically for seven categories are taken from the MIT-BIH arrhythmias database (from Record 100 to Record 234) (Goldberger et al., 2000). The database contains 48 records, and each record is slightly over 30 min long. In most records, the upper signal is a modified limb lead II (ML II) and the lower signal is a modified lead V1 (VI). Seven heartbeat classes have been included in the investigations, involving normal beat, supraventricular ectopic beat, bundle branch ectopic beat, and ventricular ectopic beat as shown in Table 1 (Chazal, Dwyer, & Reilly, 2004). A total of 43 QRS complexes (ML II Signal) are selected including patient numbers 107, 109, 111, 118, 119, 124, 200, 209, 212, 214, 217, 221, 231, 232, and 233, and the templates of seven classes are produced by Eqs. (12) and (13). The numbers of symptomatic patterns from the same class are 7-, 11-, 2-, 7-, 8-, 6-, and 2-set data (Knor = 7, KV = 11, KA = 2, KL = 7, KR = 8, KP = 6, KF = 2) respectively. Because the characteristics are enhanced, the number of training data requirements can be reduced. We can system- atically create weights between wavelet nodes and hidden nodes with 43-set training data, k ¼ 1; 2; 3; . . . ; 43. The weights between hidden nodes and summation nodes are encoded as binary values by Eq. (14) with signal ‘‘1’’ for belonging to Class j, j ¼ 1; 2; 3; . . . ; 7. The AWN contains 100 wavelet nodes, 43 hidden nodes, eight summation nodes, and seven output nodes. The num- ber of wavelet nodes is equal to the number of the sampling points, the number of hidden nodes is equal to the number of training data, and each output represents one normal beat and six types of arrhythmias as defining output vector O = [O1, O2, O3, O4, O5, O6, O7] = [ONor, OV, OA, OL, OR, OP, OF]. The selection sort is applied to find the maximum value that indicates the arrhythmic type. The output values are between 0 and 1, where a value close to 1 means ‘‘Nor- mal’’, and close to 0 means ‘‘Abnormal’’. If clinicians pro- vide some suggestion or more patterns are generated in clinical investigation, training data can be continually added to the current database. The database can be enhanced at any time with new training data. The corre- sponding hidden nodes will continue to grow, and will update the network weights without re-iteration to corrupt Fig. 3. Typical arrhythmia heartbeats in time domain (Lead II signal). C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 2605 Author's personal copy the previous database or weights. This process results in very fast training, and the network is adaptive to data changes by tuning smoothing parameters. 4. Experiment result The proposed detection algorithm was developed on a PC Pentium-IV 2.4 GHz with 480 MB RAM and Matlab workspace, based on the MIT-BIH arrhythmias records. The performance of the proposed model was tested with learning performance for the training data and detection accuracy for the untrained data, as detailed below. 4.1. Single cardiac arrhythmia with noise influence In ECG measurement, signals may be disturbed by noise such as power line interference (Minami et al., 1999) or quantification error (Gaussian noises). The ECG signal sometimes is disturbed by 50 Hz or 60 Hz interference whose amplitude is approximately 5–6 times less than the R-peak, as Fig. 5a shows the ECG signals of normal heart- beat and V in the time domain, and Fig. 5b shows the ECG signals with 60 Hz noise influences. Using 100 heartbeats (about 1.5 min long) of the patient numbers 119, 200, and 212 (Goldberger et al., 2000) containing normal beats, pat- tern Vs, and pattern Rs, the results show that high accura- cies of the proposed algorithm, as shown in Table 2. Test 1 shows the test results without any noise, and Test 2 shows the results with presenting ECG signals involving 60 Hz interference. The proposed method is robust enough to han- dle noisy environments. Overall accuracies are greater than 90%. The positive predictivity of more than 80% is obtained to quantify the performance of proposed method with or without a noisy background. It can be seen that the features of heartbeats are still strong enough to recognize, and the symptomatic patterns are not corrupted by noise influences as shown in Fig. 6. When the symptomatic patterns have Fig. 4. Various symptomatic patterns in time–frequency domain. 2606 C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 Author's personal copy morphological variation or sag shapes, the critical times for starting and ending of occurrences are clearly noted, and the number of abnormal beats is easy to count. This study case confirms that the proposed method can work in an uncertain environment with a noisy background. 4.2. Multiple cardiac arrhythmias Some of the clinical cases include multiple cardiac arrhythmias, for example ventricular ectopic beat, bundle branch ectopic beat, fusion, and paced beats. Using 100 heartbeats of the patient numbers 217, 214, and 118 includ- ing multiple cardiac arrhythmias (Masters & Land, 1997), Table 1 Heartbeat types of human ECG data Class Beat Symbol Record Patient Normal Normal beat • MIT-100 Male, age 69 MIT-103 Male, age * MIT-119 Female, age 51 MIT-200 Male, age 64 MIT-209 Male, age 62 MIT-212 Female, age 32 MIT-221 Male, age 83 Ventricular ectopic beat Ventricular premature contraction V MIT-119 Female, age 51 MIT-200 Male, age 64 MIT-221 Male, age 83 MIT-233 Male, age 57 Supraventricular ectopic beat Atrium premature beat A MIT-202 Male, age 68 MIT-232 Female, age 76 Bundle branch ectopic beat Left bundle branch block beat L MIT-109 Female, age 64 MIT-111 Female, age 47 MIT-207 Female, age 89 MIT-214 Male, age 53 Right bundle branch block beat R MIT-118 Male, age 69 MIT-124 Male, age 77 MIT-212 Female, age 32 MIT-231 Female, age 72 Unknown Paced beat P MIT-107 Male, age 63 MIT-217 Male, age 65 Fusion beat Fusion of paced and normal beat F MIT-217 Male, age 65 Note: *: Not recorded. Fig. 5. ECG signals in time domain (Record 119) (a) ECG signals of normal beat and V; (b) ECG signals of normal beat and V with 60 Hz power line interference. Table 2 The test results of single cardiac arrhythmias Record Number of arrhythmias CPU time (s) Accuracya (%)• V A L R P F 119 Actual 75 25 0 0 0 0 0 – – Test1 75 25 0 0 0 0 0 0.096 100 Test2 75 25 0 0 0 0 0 0.105 100 200 Actual 62 38 0 0 0 0 0 – – Test1 62 35 0 1 2 0 0 0.109 97 Test2 62 35 0 1 2 0 0 0.156 97 212 Actual 5 0 0 0 95 0 0 – – Test1 9 0 0 0 91 0 0 0.109 96 Test2 9 0 0 0 91 0 0 0.110 96 a Accuracy (%) = (Nr/Nt) · 100%; Nr: the number of correctly discri- minated beats; Nt: total number of heartbeats. C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 2607 Author's personal copy Test 1 and Test 2 show that the accuracies are greater than 90%, as shown in Table 3. The results confirm that the major types are paced beat (P), left bundle branch beat (L), and right bundle branch beat (R). Fig. 7 shows the traced detection results of patient number 118 and appears six misclassification errors (�) of class R in the output OR. The processes recognized 94 paced beats with six fail- ures, 95 left bundle branch ectopic beats with four failures, and 99 right bundle branch ectopic beats with six failures, respectively. The positive predictivity of more than 80% is also obtained to quantify the performance with or without a noisy background. Second study case confirms that the proposed method can detect multiple cardiac arrhythmias with high confidence. Fig. 6. Symptomatic patterns in time–frequency domain (a) symptomatic patterns of normal beat and V; (b) symptomatic patterns of normal beat and V with 60 Hz power line interference. Table 3 The test results of multiple cardiac arrhythmias Record Number of arrhythmias CPU time (s) Accuracya (%)• V A L R P F 217 Actual 0 3 0 0 0 94 3 – 94 Test 0 4 0 0 1 88 7 0.113 214 Actual 0 5 0 95 0 0 0 – 96 Test 1 5 3 91 0 0 0 0.108 118 Actual 0 1 0 0 99 0 0 – 94 Test 4 1 0 2 93 0 0 0.111 a Accuracy (%) = (Nr/Nt) · 100%; Nr: the number of correctly discri- minated beats; Nt: total number of heartbeats. Fig. 7. Detection results of patient number 118. Note: �: Error. 2608 C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 Author's personal copy 4.3. Learning performance tests Figs. 8a and b show the squared errors and smoothing parameters versus learning cycles, respectively. The perfor- mance of AWN is affected by the width of the Gaussian activation function. As the width of Gaussian function decreases, decision boundaries can become increasingly nonlinear. For a very narrow Gaussian function, the net- work approaches a nearest-neighbor classifier (Lin & Tsao, 2005; Lin & Wang, 2006). Eqs. (8)–(11) are used to find near-optimum smoothing parameters that minimize the training-data error of the AWN as the number of training data increases from K1=18 to K6=43. The corresponding hidden nodes will continue to grow from 18 to 43, and con- struct the network weights without any iteration process. This process results in very fast training, and the network is adaptive to match desired output with add-in/delete-off training patterns by tuning smoothing parameters for six learning stages. Learning rates g = 0.1–0.3 are selected for training AWN as shown in Table 4. Fig. 8 shows the final squared errors after the training AWN has been responsible for the determining smoothing parameters. For the convergent condition e 6 10�3, AWN rapidly con- verges to the nearest local minimum for less than 40 learn- ing cycles in a shorter processing time. It takes 0.782 s (Average CPU Time) to classify the 43 training data into seven categories. The output values of AWN are shown in Fig. 9a. AWN automatically adjusts outputs to approach the targets with tuning smoothing parameters in the learning stages. Threshold value 0.5 is used to separate ‘‘Normal’’ from ‘‘Abnormal’’, and can correctly discriminate. The proposed method has the advantages of fast learning process, learn- ing stage with slight iteration for updating weights, and adaptation capability. However, the performance is affected by the parameter of Gaussian activation function. For 43 trained data and 1200 untrained data including sin- gle and multiple cardiac arrhythmias, the ranges of param- eters could be determined by experience and trial-and-error procedure with tuning parameters. The detection accuracy decreases as smoothing parameters increase. The choice of parameters will affect the estimation error, and the suitable range is from 0.05 to 0.30 for reducing misclassification errors, and the accuracies are greater than 90% as shown in Fig. 9b. Refining the parameters can enhance the detec- tion accuracy by using the proposed optimum method. The near-optimal parameter r = 0.078 can minimize the classi- fication error, and the accuracies are the maximums for single and multiple cardiac arrhythmias, as shown in Fig. 9b. For comparison purposes, we have also applied the WBPN composed of 100 wavelet nodes in the wavelet layer and multi-layer neural network (MLNN). For the second subnetwork, a MLNN is used for training with the back- propagation learning algorithm. Only one hidden layer is used, and the number of hidden nodes is determined by the experience formulas as shown in Table 4. Traditional MLNN has some limitations including very slow learning process, needs iteration for determining weights and learn- ing rates (g = 0.2–0.8), and needs to determine the network architecture such as the number of hidden layers and hid- den nodes, which is difficult to retrain with new training data. With various tests, we can see that the training time of AWN outperformed WBPN. AWN has a fast learning process needing no iteration for updating weights, a flexible hidden nodes mechanism with add-in or delete-off, and automatic adjustment of the targets and parameter r. With the same training data, the proposed AWN shows better performance than WBPN as shown in Table 4. Table 4 Comparison of AWN with WBPN Method Network topology Training patterns Learning rate g Learning cycles Average CPU time (s) *AWN 100-43-8-7 43 0.1–03 640 0.782 WBPN 100-100-27-7 43 0.2–0.8 <10000 <800 100-100-54-7 100-100-107-7 Note: (1) NH = (NI + NO) 1/2; (2) NH = (NI + NO)/2; (3) NH = (NI + NO). NH: the number of hidden node; NI: the number of input node; NO: the number of output node. 0 1 2 3 4 0 10 20 30 40 K1 K2 K3 K4 K5 K6 Learning Cycle S qu ar ed E rr or 0 10 20 30 40 0 0.2 0.4 0.6 0.8 1 K1 K2 K3 K4 K5 K6 Learning Cycles S m oo th in g P ar am et er Fig. 8. (a) Squared errors versus learning cycles, (b) smoothing param- eters versus learning cycles. Note: K1 ¼ K nor þ K V ¼ 18; K2 ¼ K1 þ K A ¼ 20; K3 ¼ K2þK L ¼ 27; K4 ¼ K3þK R ¼ 35; K5 ¼ K4þK P ¼ 41; K6 ¼ K5þ K F ¼ 43. C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 2609 Author's personal copy 5. Conclusion The diagnostic procedure based on WT and PNN has been presented to recognize cardiac arrhythmias. For ECG signals, classifier based on AWN is proposed to rec- ognize normal beat, premature ventricular contraction, atrial premature beat, right/left bundle branch block beat, paced beat, and fusion of paced and normal beat. The wavelets act to extract and enhance the features from QRS complexes in the time domain. These features are slightly affected by noise influences. Subsequently, PNN can classify the applied input pattern with/without a noisy background. The AWN model can also work in a dynamic environment with continuity add-in/delete-off features by automatically tuning the targets and smoothing parameters of hidden nodes. For both trained and untrained data, the results demonstrate the efficiency of the proposed method. Compared with the MLNNs, AWN can be built using adaptive training algorithms, and can avoid the determina- tion of network weights by the trial-and-error procedure. Test results show accurate diagnosis, fast learning, good adaptability, and faster processing time for detection. With this model, special features can be further added to the cur- rent database such as ventricular bigeminy (B), ventricular trigeminy (V), ventricular tachycardia (VT), normal sinus rhythm (N), etc. Thus, the database is always enhancible with new symptomatic patterns. The AWN can adapt itself in a new environment by adding features, and also promise the high confidence value of detection results with special features considering. The proposed method can be used as an aided tool for heartbeat recognition, and be inte- grated in the monitoring device. Acknowledgement This work is supported in part by the National Science Council of Taiwan under Contract Number: NSC 93- 2614-E-244-001 (December 1 2004–July 31 2005). References Acharya, R., Kumar, U. A., Bhat, P. S., Lim, C. M., & Lyengar, S. S. (2004). Classification of cardiac abnormalities using heart rate signals. Medical and Biological Engineering and Computing, 42(3), 288–293. Chazal, Philip de, Dwyer, Maraia O’, & Reilly, Richard B. (2004). Automatic classification of heartbeats using ECG morphology and heartbeat interval features. IEEE Transactions on Biomedical Engi- neering, 51(7), 1196–1206. Chen, Tain-Song, Chen, Tzu-Pei, & Tsai, Liang-Min (1997). Computer- ized quantification analysis of left ventricular wall motion from echocardiograms. Ultrasonic Imaging, 19, 1–9. Dickhaus, Hartmut, & Heinrich, Hartmut (1996). Classifying biosignals with wavelet networks – A method for noninvasive diagnosis. IEEE Engineering in Medicine and Biology(September/October), 103–111. Goldberger, A.L., Amaral, L.A.N., Glass, L., Hausdorff, J.M., lvanov, P.Ch., Mark, R.G., et al. (2000) PhysioBank, Physio Toolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic signals. Circulation 101 (23), e215–e220 [Circulation electronic Pages; http://circ/cgi/content/full/101/23/e215]; (June 13). 0 0.2 0.4 0.6 0.8 1 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 Nor V A L R P F Threshold value = 0.5 0% 20% 40% 60% 80% 100% 0.01 0.05 0.078 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Single Cardiac Arrhythmia Multiple Cardiac Arrhythmias Smoothing parametersSmoothing parameters Optimal ParameterOptimal Parameter A cc ur ac y A cc ur ac y Accuracy The number of training data O ut pu t V al ue Fig. 9. (a) Output target values of the AWN, (b) detection accuracies versus smoothing parameters. 2610 C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 Author's personal copy Hu, Y.-H., Palreddy, S., & Tompkins, W. (1997). A patient adaptable ECG beat classifier using a maxture of experts approach. IEEE Transactions on Biomedical Engineering, 44(September), 891–900. Lin, Chia-Hung, & Tsao, Ming-Chieh (2005). Power quality detection with classification enhancible wavelet-probabilistic network in a power system. IEE Proceedings-Generation, Transmission, and Distribution, 152(6), 969–976. Lin, Chia-Hung, & Wang, Chia-Hao (2006). Adaptive wavelet networks for power quality detection and discrimination in a power system. IEEE Transactions on Power Delivery, 21(3), 1106–1113. Masters, Timothy, & Land, Walker (1997). A new training algorithm for the general regression neural network. 1997 IEEE international conference on computational cybernetics and simulation, 3, 1990–1994, 12–15 October. Minami, Kei-ichiro, Nakajima, Hiroshi, & Toyoshima, Takesshi (1999). Real-time discrimination of ventricular tachyarrhythmia with fourier- transform neural network. IEEE Transactions on Biomedical Engi- neering, 46(2), 179–185. Osowski, Stanislaw, & Linh, Tran Hoai (2001). ECG beat recognition using fuzzy hybrid neural network. IEEE Transactions on Biomedical Engineering, 48(11), 1265–1271. Qin, Shuren, Ji, Zhong, Zhu, Hongjun (2003). The ECG recording analysis instrumentation based on virtual instrument technology and continuous wavelet transform. In Proceedings of the 25th annual international conference of the IEEE EMBS Cancun, Mexico, Septem- ber 17–21 (pp. 3176–3179). Seng, Teo Lian, Khalid, Marzuki, Tusof, Rubiyah (2002) Adaptive GRNN for the Modeling of Dynamic Plants. In Proc. of the 2002 IEEE International Symposium on Intelligent Control Vancouver, Canada, October 27–30 (pp. 217–222). Silipo, Rosaria, & Marchesi, Carlo (1998). Artificial neural networks for automatic ECG analysis. IEEE Transactions on Signal Processing, 46(5), 1417–1425. Specht, Donald F. (1988). Probabilistic neural network for classification mapping, or associative memory. In Proc. IEEE Int. Conf. Neural network, San Diego, CA, Vol. 1 (pp. 525–532), July. Wang, Yang, Zhu, Yi-Sheng, Thakor, Nitish V., & Xu, Yu-Hong (2001). A short-time multifractal approach for arrhythmias detection based on fuzzy neural network. IEEE Transactions on Biomedical Engineering, 48(9), 989–995. C.-H. Lin et al. / Expert Systems with Applications 34 (2008) 2601–2611 2611 
work_o6apwsmoubhl7pn66lgybiegnu	----	A Two Fold Expert System for Yawning Detection Procedia Computer Science 92 ( 2016 ) 63 – 71 Available online at www.sciencedirect.com 1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Organizing Committee of ICCC 2016 doi: 10.1016/j.procs.2016.07.324 ScienceDirect 2nd International Conference on Intelligent Computing, Communication & Convergence (ICCC-2016) Srikanta Patnaik, Editor in Chief Conference Organized by Interscience Institute of Management and Technology Bhubaneswar, Odisha, India A Two Fold Expert System for Yawning Detection Anitha Ca*, M K Venkateshab, B Suryanarayana Adigac aResearch Scholar – BMS College of Engineering/ Asst. Prof. – NIE, Mysuru- 5700 08, INDIA *is the corresponding author bPrincipal, RN Shetty Institute of Technology, Bengaluru – 560 098, INDIA, cFormer Chief Consultant, TCS Ltd., Bengaluru – 560 066, INDIA Abstract One of the prominent indicators of drowsiness is yawning. The main intention for a real-time application such as detecting the driver’s yawning is that the response of the detector must be as quick as possible. A novel yawning detection system is proposed which is based on a two agent expert system. The features of the face have to be extracted to detect yawning in the driver’s face. In the proposed system, as the first part of detection we use the face detection algorithm’s skin detection part. The skin region is extracted. For all the skin region blocks detected, their boundaries are defined. Then segmented face is divided into two halves. The lower half of the face is considered for the mouth region extraction. The presence of yawning would be indicated by a black blob in the mouth region of the binary image. But, there may be multiple blobs present in the image which may be due to the presence non-skin like regions around the driver’s face. So, identifying the exact position of the mouth and checking for its containment inside the face is necessary. The features extracted for yawning detection are the histogram values taken from the vertical projection of the lower part of the face. © 2014 The Authors. Published by Elsevier B.V. Selection and peer-review under responsibility of scientific committee of Missouri University of Science and Technology. Keywords: Yawning detection; two-fold expert system; mouth region extraction; histogram; containment; vertical projection; non-frontal face images; blob detection © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Organizing Committee of ICCC 2016 http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2016.07.324&domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2016.07.324&domain=pdf 64 C. Anitha et al. / Procedia Computer Science 92 ( 2016 ) 63 – 71 1. Introduction Driver’s behavior monitoring system finds its application in many real-time situations that include drowsiness detection and surveillance. One of the applications that are considered is the yawning detection while driving. Yawning is a vital indicator for drowsiness or sleepiness. It is crucial to identify the sleepiness at an early stage. Repeated yawning most of the time, leads to lack of alertness. When the driver is not alert, mishaps tend to occur. In order to avoid mishaps occurring through non-alert drivers, the vital indicators have to be detected at the earliest. As per the statistics provided by the National Highway Traffic Safety Administration1 (NHTSA), about 72,000 crashes, 800 fatalities and 44,000 injuries have occurred in 2013 due to drowsy driving. Hence, there is a need for automatically alerting the driver in advance. In this paper, we propose a two-fold expert system for yawning detection. The proposed system is computationally less complex, robust and has less reaction time. In the next section, the state of the art yawning detection systems are described. Section 3 gives a detailed description of the proposed system. The results are discussed in section 4. The last section provides the conclusion for the proposed work. 2. Related Work In4, three different approaches for yawning detection are described. The color segmentation technique, the Snake Contour method and the third method was usage of Viola-Jones theory9 for face and mouth detection for detecting and locating the mouth region. The yawning detection was based on the openness of the mouth and the number of frames it was open. A comparative histogram was used between the closed mouth condition and the yawning condition. In5, an SVM classifier is used to detect yawning condition. The classifier is given the width to height ratios of eyes and mouth. Both the eyes and the mouth features are considered for concluding the yawning condition. In6, the location of the chin and nostrils are used to detect yawning. The distance between the chin and the nostrils increases as a person yawns, according to the authors’ of6. The localization of the chin and the nostrils is done using directional integral projection method. The Viola-Jones technique9 for face and mouth detection is used in paper7. An SVM classifier is trained with mouth and yawning images. The mouth region is detected from the face using a cascade of classifiers during fatigue. SVM is used to classify the mouth regions as yawning or alert. In8, mouth corners are detected by grey projections and extracted using Gabor wavelets. Least Discriminant Analysis (LDA) is applied to classify these Gabor wavelets into yawning and not yawning condition. In8, the authors also detect yawning based on the mouth’s height to width ratio. Most of the techniques discussed for yawning detection make use of classifiers. The use of classifiers adds to the complexity of the system. In this paper we propose a system that does not use any classifiers for yawning detection; this reduces the computational time and complexity. Another feature of the proposed system is that the yawning detection is performed in two folds; this improves the accuracy. 3. Proposed System A two-fold yawning detection system is proposed in this paper. The word two-fold is used as the yawning is detected in two ways. Firstly, yawning detection is based on the skin tone detection and secondly based on blob dimensions and face containment. As the first step towards yawning detection, we separate the face from the background. 3.1. Feature Extraction The features of the driver’s face have to be extracted to detect yawning. In the proposed system, as the first part of detection we use the face detection algorithm’s skin detection part. The skin region is extracted. For all the skin region blocks detected, their boundaries are defined. Then extracted face is divided into two halves. The lower half of the face is considered for the mouth region extraction. 65 C. Anitha et al. / Procedia Computer Science 92 ( 2016 ) 63 – 71 The presence of yawning would be indicated by a black blob in the mouth region of the binary image. But, there may be multiple blobs present in the image which may be due to the presence non-skin like regions around the driver’s face. So, identifying the exact position of the mouth and checking for its containment inside the face is necessary. The features extracted for yawning detection are the histogram values taken from the vertical projection of the lower part of the face. 3.2. Design The most important assumptions made for yawning detection: Driver’s face is expected to be frontal to the camera (located on the dashboard of the car). Head rotation maximum up to 45° permitted, not beyond that (practically not feasible to drive with head turned beyond 45°) Figure 1 depicts the steps involved in the proposed yawning detection system. Fig 1: Block diagram of the yawning detection approach. 3.3. The Approach The method adopted is simple computationally, yet efficient and accurate. Simple as there is no use of classifiers for yawning detection. The proposed system is a two-agent expert system to detect yawning. The first agent detects yawning based on skin tone detection. The second agent detects yawning based on the blob dimensions and containment within the face region. The design steps are elaborated as below: Detect skin tone The technique discussed in3 is used for skin detection. The input is resized to standard dimensions. This algorithm displays a blob at regions containing non-skin values. All the detected skin regions are bounded in a rectangle and indexed. The minimum and maximum index values are extracted. The maximum index 66 C. Anitha et al. / Procedia Computer Science 92 ( 2016 ) 63 – 71 value is used to define the region of interest (ROI) rectangle, which is the candidate face. Segment the detected face To minimize the noise and to reduce the size of computation, only the lower part of the face is considered. This segment contains the mouth region, which is further analysed. Verify the presence of blob The non-skin regions in the segmented part of the face will be represented by black blobs. Check if the blobs appearing are due to features inside the face or not. This is because; the noise from the surroundings should not be mistaken for inside face non-skin regions. Bound the blobs The blobs confirmed to be inside the face are bounded in rectangles as per their dimensions. Take the histogram from the vertical projection of the bounded blobs. Verify the yawning through histogram The histogram of blobs is obtained through vertical projection of the segmented face. Histograms are considered for blobs present only at the centre and near-centre regions of the segmented face. This is in accordance with the assumptions stated earlier in this section. The histogram’s length and sum values are verified with the threshold values. If the values satisfy the thresholds specified, yawning is confirmed. 3.4. Dealing with Occlusions The occlusions on the driver’s face may be many, such as mouth covered by hand while yawning; wearing sunglasses while driving, hair covering the face while driving and so on. Among these, the mouth covered by hand while yawning is one which cannot be handled. The frames with the hand covering the mouth, even though yawning, would indicate not yawning. This can be overcome by monitoring the frames’ output before and after the occluded frames. The wearing of sunglasses or any other type of glasses does not bother the yawning detection as we have segmented the face, and only the lower part of the face is considered for yawning detection. The covering of face with the hair also causes occlusion to some extent as the skin region covered cannot be detected. But this is not as serious as it does not affect the yawning decision made finally. 4. Results and Discussion The methodology adopted for yawning detection was tested on the YAWDD data set2. Different cases were considered for evaluation. These include image sequences with frontal faces, sequences with sunglasses, sequences with prescribed spectacles, sequences with mouth occlusion, and sequences with non-frontal faces. This section discusses the results into the following categories: • Non-occlusive, frontal faces with/without glasses • Occlusion with hand, frontal with/without glasses • Non-frontal faces with/without glasses 4.1. Non-occlusive, frontal faces with/without glasses 67 C. Anitha et al. / Procedia Computer Science 92 ( 2016 ) 63 – 71 The detection is accurate for non-occlusive, frontal faces. The driver’s sunglasses do not affect the response of the yawn detector, since only the lower part of the face is considered. Fig. 2 True Positives; (a) Input image, (b) Segmented region, (c) Yawn detector’s output, (d) Histogram The histogram values indicated in the rightmost column of figure 2 indicates the blob measurements in the lower segmented region of the face. The presence of multiple bins indicates the presence of other blob regions. But mouth blob is identified based on the location and the containment inside the face. 4.2. Occlusion with Hand The detection in this case was not possible in the frames where the hand completely covered the mouth region. But, due to the detection of yawning in the previous and continuing frames, this occlusion could be easily handled. Figure 3 shows the input and output resulting in this case. When the driver occludes the mouth region during yawning, the blob in the mouth region is covered and the vertical projection does not reflect any value. Hence the histogram does not reflect any variations. During these frames, the yawning condition goes undetected. As our system increments a count value for every frame when yawning is detected based on the histogram values. The yawning is detected the moment the hand is removed away from the mouth region. The proposed system could not overcome this occlusion as the feature used for yawning detection was the based on the mouth region. 68 C. Anitha et al. / Procedia Computer Science 92 ( 2016 ) 63 – 71 69 C. Anitha et al. / Procedia Computer Science 92 ( 2016 ) 63 – 71 Fig. 3 Occlusion with hand, Yawn Undetected (a) Input image, (b) Segmented region, (c) Yawn detector output, (d) Histogram 4.3. Non – frontal faces The frames with subject completely frontal to the camera are the actual and typical way a driver drives a car. But some cases where the face was not completely in front of the camera during yawning were considered. The success rate was good for faces up to 45° inclination, but went bad for faces inclined beyond that. Figure 4 shows some of the non-frontal frames from the database. As per the assumptions made in section 3.2, the proposed system detects yawning when the driver’s face is completely parallel to the camera (completely facing the camera) or slightly rotated. The extent to which yawning is detected with head rotation is 45°. Beyond this the blob’s containment in the face becomes false, leading to undetected yawning condition. Moreover, the driver cannot afford to keep his head away from the windshield all the time. 70 C. Anitha et al. / Procedia Computer Science 92 ( 2016 ) 63 – 71 Fig. 4 Non – frontal frames, less than 45° detectable (a) Input image, (b) Segmented region, (c) Yawn detector output, (d) Histogram 5. Conclusion A robust, computationally less complex yawning detection system is proposed. The two fold expert system was successful in detecting yawning with an accuracy of 94% when evaluated on the YAWDD – Yawning Detection Database2. The first clue for the yawning condition is at the skin tone detection phase. The presence of blob is an indication of yawning condition. The confirmation of the yawning condition is when the blob is tested for its containment within the face and the histogram values of the vertical projection of the blobs in the lower segmented face region. The proposed system indicates yawning condition only when the histogram value is above the threshold value. The system successfully detects yawning when the driver’s face is completely frontal or slightly frontal. The system is unable to detect in the driver’s head is rotated beyond 45°. Yawning with hand occluding the mouth cannot be detected as the entire mouth region is covered by the hand. The previous and next frames can be used to detect yawning in such situations. The proposed system deals with head movement up to 45°. The yawning detection systems discussed in5, 6, 7, 8 consider only frontal images. The proposed system is evaluated on the publicly available yawning detection database YAWDD2, which consists of 322 video sequences shot in the car with varying light conditions, when compared to5, 6, 7, 8 which use a small private data set under controlled conditions for evaluation. 71 C. Anitha et al. / Procedia Computer Science 92 ( 2016 ) 63 – 71 Acknowledgement The authors express their deepest gratitude to BMS College of Engineering, Bengaluru for providing the required help in conducting this research work and The National Institute of Engineering, Mysuru for their continuous encouragement and support. References 1. NHTSA – National Highway Traffic Safety Administration, Washington DC, Online: http://www.nhtsa.gov/Driving+Safety/Drowsy+Driving 2. S Abtahi, M Omidyeganeh, S Shirmohammadi, B Hariri, ‘YawDD: A Yawning Detection Dataset’, in proceedings of ACM Multimedia Systems, Singapore, pp. 24-28, March 2014. 3. Anitha C, M K Venkatesha, B Suryanarayana Adiga, “Real Time Detection and Tracking of Mouth Region of Single human Face”, in Proceedings of IEEE Third International Conference on Artificial Intelligence, Modeling and Simulation (AIMS2015), Malaysia, pp. 297 – 304, December 2015. 4. S Abtahi, B Hariri, S Shiromohammadi, “Driver Drowsiness Monitoring Based on Yawning Detection”, in Proceedings of IEEE International Instrumentation and Measurement Technology, Binjiang (Hangzhou), China, May 10-12, Pages 1 – 4, 2011. 5. T Azim, M Jaffar, A Mirza, “Automatic Fatigue Detection of Drivers through Pupil Detection and Yawning Analysis”, in Proceeding of Fourth International Conference on Innovative Computing, Information and Control, pp. 441 – 445, 2009. 6. L Yufeng, Z Wang, “Detecting Driver Yawning in Successive Images”, in Proceedings of First International Conference on Bioinformatics and Biomedical Engineering, 2007. 7. M Saradadevi, P Bajaj, “Driver Fatigue Detection Using Mouth and Yawning Analysis”, IJCSNS International Journal of Computer Science and Network Security, Vol. 8, No. 6, pg. 183-188, June 2008. 8. X Fan, B Yin, Y Sun, “Yawning Detection for Monitoring Fatigue”, in International Conference on Machine Learning and Cybernetics, 2007. 9. P Viola, M Jones, “Rapid Object Detection Using a Boosted Cascade of Simple Features”, in Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2001. 
work_o6ast2mhqnhb7phhyj4jf6zvx4	----	Trend following algorithms in automated derivatives market trading Expert Systems with Applications 39 (2012) 11378–11390 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Trend following algorithms in automated derivatives market trading Simon Fong ⇑, Yain-Whar Si, Jackie Tai Department of Computer and Information Science, University of Macau, Macau a r t i c l e i n f o a b s t r a c t Keywords: Trend following Automated trading system Futures contracts Mechanical trading 0957-4174/$ - see front matter � 2012 Elsevier Ltd. A http://dx.doi.org/10.1016/j.eswa.2012.03.048 ⇑ Corresponding author. Tel.: +853 83974460; fax: E-mail address: ccfong@umac.mo (S. Fong). Trend following (TF) is trading philosophy by which buying/selling decisions are made solely according to the observed market trend. For many years, many manifestations of TF such as a software program called Turtle Trader, for example, emerged in the industry. Surprisingly little has been studied in academic research about its algorithms and applications. Unlike financial forecasting, TF does not predict any mar- ket movement; instead it identifies a trend at early time of the day, and trades automatically afterwards by a pre-defined strategy regardless of the moving market directions during run time. Trend following trading has been popular among speculators. However it remains as a trading method where human judgment is applied in setting the rules (aka the strategy) manually. Subsequently the TF strategy is exe- cuted in pure objective operational manner. Finding the correct strategy at the beginning is crucial in TF. This usually involves human intervention in first identifying a trend, and configuring when to place an order and close it out, when certain conditions are met. In this paper, we evaluated and compared a col- lection of TF algorithms that can be programmed in a computer system for automated trading. In partic- ular, a new version of TF called trend recalling model is presented. It works by partially matching the current market trend with one of the proven successful patterns from the past. Our experiments based on real stock market data show that this method has an edge over the other trend following methods in profitability. The results show that TF however is still limited by market fluctuation (volatility), and the ability to identify trend signal. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction Trend following (TF) (Fong & Tai, 2009) is a reactive trading method in response to the real-time market situation; it does nei- ther price forecasting nor predicting any market movement. Once a trend is identified, it activates the trading rules and adheres rigidly to the rules until the next prominent trend is identified. Trend fol- lowing does not guarantee profit every time, but nonetheless in a long term period it may probably profit by obtaining more gains than loses. The nature of TF makes it as an ideal ingredient in implementing a decision-making component in automated trading software where human intervention is not required in automatic buying and selling. Automated trading here is referred to an ‘auto-pilot’ mode by which the software system decides when to buy or sell a stock (in a form of option, future contract, warrant, etc.) from the derivative market. Though the philosophy of trend following is simple – identify a current market trend and trade strictly according to the pre-defined rules (sometimes known as strategies), how the rules or strategies should be formulated has become an important research problem that deserves attention from researchers. Rules such as when to buy and sell or when to ll rights reserved. +853 2883 8314. open and close a position, as signaled from the market trend have a direct impact into the profitability of automated market trading. TF algorithms in this context are termed as automated trading methods that are guided by the current market trend (signals from the trend) and specified by the rules (reactions to the trend). For an example of a most basic TF algorithm, buying and selling are cued by the conditions when the market trend which is represented by its moving average rises over an up-threshold, and falls below a down-threshold respectively. The values of the thresholds are pre- defined as a part of the trading rules. To the best of the authors’ knowledge, surprisingly, TF algo- rithms have not been investigated in the computer science com- munity. In contrast a lot of research papers on stock market forecast and prediction by soft computing can be found in litera- ture. Fig. 1 classifies clearly that trading algorithms can be pow- ered by two different domains of techniques, namely Predictive and Reactive. Predictive types of forecasting and trading in a stock market have a long history in academic research, which generally covers non-linear prediction by artificial neural works, decision trees and other regression models. These tools usually aim at pre- dicting the future market movement ahead by analyzing over the historical data. TF algorithms however, belong to the latter cate- gory, which conduct trading decision solely by reacting to the cur- rent market trend. There are several variants of TF algorithms, depending on how the rules and the predefined thresholds are http://dx.doi.org/10.1016/j.eswa.2012.03.048 mailto:ccfong@umac.mo http://dx.doi.org/10.1016/j.eswa.2012.03.048 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. Classification of trading algorithms. P Q T Fig. 2. Illustration of P and Q applied on trend T. S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 11379 specified. Without the need of training up a decision model, and subsequently updating it, like those under the predictive type, TF is readily implemented as a rule-based system in automated trad- ing software. The objective of this paper is to present a performance compar- ison of the five TF algorithms which have been recently proposed by the authors. The contribution of this paper is twofold: Theoret- ically insights are gained in terms of performance strength, weak- ness and characteristics of each TF algorithm under test, so that prospect of future research in improving from the existing TF algo- rithms is warranted. Secondly the TF algorithms presented in this paper serve as useful designs of the decision-making core for implementing an automated stock market trading system. The structure of this paper is as follow: Section 2 describes four possi- ble implementations of TF algorithms, highlighting especially their shortcomings and possible improvement. Section 3 presents in de- tails the latest addition to the family of TF algorithms called trend recalling algorithm. A simulation prototype of automated trading system is developed, for the purpose of evaluating their perfor- mances comparatively. The experiments carried out by the simula- tor are reported in Section 4. Section 5 concludes. 2. Trend following algorithms In this paper, we discuss on the design of an automated trading system, which incorporates these trend following algorithms, which is a core element of the system, and the application of this type of system in financial market. We also investigate how the market fluctuation can affect the overall performance, and bring new perspective to handling the financial cycles. One of our goals is to build up this automated trading system, which primarily operates on financial derivative market (Mayo, 2003), such as the Hang Sang Index Futures Contract that trade on HKFE (Hong Kong Future Exchange) or the Dow Jones Industrial Average Index Futures Contract that trade on CBOT (Chicago Board of Trade). In this thesis ‘‘Hang Sang Index Futures’’ is selected as the primary simulation market, although the system can be ap- plied on any market theoretically. By using only historical market data, the system is able to react according to real-time market state, and make trade decision on its own. To reduce the overnight risk and cost-of-carry, trade will be performed on a daily basic only, which means no contracts will be carried overnight. The P&L (profit and loss) on each trade will be recorded and accumulated as the total of ROI (return on invest- ment), which is a common indication for the performance of an investment in financial world. 2.1. Static TF algorithm The basic practice of trend following is to find the trend, identify trade signals and trade along with them. Based on a phenomenon pointed out by Murphy (1999), which says ‘‘A trend in motion is more likely to continue than to reverse’’, TF algorithm can possibly reap benefit by recognizing the direction of the trend at the start and bet on that direction for the remaining length of the trend, as long as the trend does not reverse. For confirming whether a trend is now in effect of development, it is necessary to measure how much it has progressed since the last reverse, and then ensure the current progress is greater than the minimum threshold that it has risen or fallen. This is to test by comparing the trend momen- tum to a specific level. If the amount of increase or decrease of a trend has already exceeded a certain level, a significant trend is as- sumed to emerge and trade signals can be safely generated accord- ingly. These marks or minimum levels will have to be chosen carefully in order to avoid false signals as a result of temporary fluctuation or short-term velocity of the trend. If they are too long, few positions will be opened or closed because it is not often that a large scale charge on a trend would occur. If the marks are chosen too short, the trading is prone to be buying very frequently and selling too early, leaving little or no profit in between the trades. In this Static TF algorithm, two constants are used to function as the two comparison marks, only over which a substantial change in the trend would trigger the trading system to open or close a posi- tion accordingly. These constants are defined as P and Q, where P is the amount of up-trend required for opening a position, and Q is the amount of opposite trend required to close this position. Fig. 2 illustrates the basic concept of TF. T denotes as the market price trend. For example on the diagram, it will open a long posi- tion when the trend T advances over P, and will close out this posi- tion when the trend T declines for more than Q. In reality market price does not move in a straight line. It is therefore impractical to apply the P and Q rules directly on the trend T, because the frequent fluctuation will alarm off too many signals of trading actions. An Exponential Moving Average (EMA) algorithm is used to smooth out this fluctuation, which equation is shown as follow: EMAðtÞ ¼ priceðtÞ � EMAðt�1Þ � 2 n þ 1 � � þ EMAðt�1Þ ð1Þ Fig. 3 demonstrates the effect of the smoothing. This trading algorithm is represented by the following pseudo- code (Algorithm 1), in which P and Q are static constants. Their val- ues are obtained by studying the historical data and deriving the optimal combination of P and Q. Different combinations of P and Q are tried by brute-force, testing out for the best values that generate the maximum profit in the simulation. EMA(T) is the P Q T EMA(T) Fig. 3. P and Q applied on smoothed trend EMA(T). 11380 S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 Exponential Moving Average of the real time market price trend, and EMAð:TÞ is the reversion (opposite direction) of the trend counting from the highest(or lowest) point of this trend. Algorithm 1. Pseudo-codes of the Static TF algorithm. Repeat until end of market Compute EMA(T) If no position opened If EMA(T) >= P If trend is going up Open a long position Else if trend is going down Open a short position Else if any position is opened If EMAð:TÞ >= Q Close position If end of market Close all opened position Fig. 4. Example of relat The basic premise is that the most profit is gained when a trade is synchronized to an enduring trend. The rules of this algorithm are statically formulated as the trade parameters remain un- changed once they have been defined throughout the whole trad- ing process. The experimental results from simulation with Static TF algorithm are detailed in Section 4. 2.2. Dynamic TF algorithm In this dynamic TF algorithm, P and Q are variables instead of static constants, and their values change adaptively to the current market trends. Based upon this initial concept of trading algorithm, dynamic TF algorithm is introduced with incorporation of technical analysis concept. Technical analysis makes trade decision through technical indicators such as RSI (relative strength index), STC (sto- chastic oscillator), and EMA and these indicators are changing dynamically according to the market situation. By adopting one or more of these indicators and by studying how they react to the market, rules can be formed that are able to inherit this dy- namic nature. By following these rules during trade session, we up- date the trade parameters P and Q with the latest dynamic values attribute. There are hundreds of indicators in use today, but not all are tested to be reliable. Experiments have been conducted to try out many popular ones and RSI is found to be the best for TF. RSI com- pares the magnitude of underlying recent gains of an asset to the magnitude of its recent losses, and normalized to a number that ranges from 0 to 100 RSIðtÞ ¼ 100 � 100 1 þ RS ð2Þ RS ¼ AUðtÞ ADðtÞ ð3Þ AUðtÞ¼ UpðtÞþ Upðt � 1Þþ �� �þ Upðt � n þ 1Þ n ð4Þ ADðtÞ¼ DownðtÞþ Downðt � 1Þþ �� �þ Downðt � n þ 1Þ n ð5Þ ive strength index. Fig. 5. Example of stochastic oscillator index. S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 11381 where AU is average price upward in n periods, AD is average price downward in n periods, t is the time, n is the number of RSI periods usually 14. Fig. 4 shows an example of this indicator, which is below the primary price trend. Stochastic Oscillator (STC) is a momentum indicator that shows the location of the current close relative to the high/low range over a set number of periods. Formula: %KðtÞ¼ 100 � CloseðtÞ� LLðnÞ HHðnÞ� LLðnÞ ð6Þ %DðtÞ¼ EMAð%KðtÞÞðmÞ ð7Þ where HH is highest high in n periods, LL is lowest low in n periods, n is number of STC periods, m is number of periods of EMA that ap- plied on %K. Fig. 5 shows an example of this indicator, which is below the primary price trend. By studying how this indicator reacts to the market, we can de- rive rules and criteria that describe the current market situation systematically. And this information could eventually lead to trade execution. The logic of rules is given below: As for long position, the following criteria should be satisfied. 1. Price is advancing. 2. RSI(t) is greater than RSI(EMA(t)). 3. RSI(EMA(t)) is less than 40 or greater then 60. And for short position, the following criteria should be satisfied. 1. Price is declining. 2. RSI(t) is less than RSI(EMA(t)). 3. RSI(EMA(t)) is less than 40 or greater then 60. By following those rules and criteria, we can see the values of P and Q are now changing along the trend with reference to RSI. Since RSI is changing according to market condition dynamically, rules that are derived from it will also adopt to this dynamic situ- ation. By applying these rules during trading session, we therefore obtain the values of P and Q that give good results. When the cri- terion that is described above is satisfied, it triggers the amount of up (or down) trend and causes a position to be open (or close). Now we summarize this trading algorithm as the following pseu- do-code (Algorithm 2) Algorithm 2. Pseudo-codes of the dynamic TF algorithm. Repeat until end of market Compute RSI(t) and RSI(EMA(t)) If price is advancing: If RSI(t) > EMA(t) and 40 < EMA(t) > 60 If no position has been opened Open a long position Else if short position has been opened Close out short position Else if price is declining: If RSI(t) < EMA(t) and 40 < EMA(t) > 60 If no position has been opened Open a short position Else if long position has been opened Close out long position If end of market Close all opened positions The experimental results from simulation with dynamic TF algorithm are detailed in Section 4. 2.3. Fuzzy TF algorithm The static and dynamic trend following algorithms described in previous section are designed to make trading decisions based on criteria which are formulated as crisp membership of classical binary logic. In this section, we extend the trending following algo- rithms based on Fuzzy logic (Zadeh, 1973). jFuzzyLogic (jFuzzyLog- ic, 2012) was used to develop a fuzzy inference system. Based on our experience with previous trend following algorithms, we de- fine three membership functions for input and output variables. Fig. 7. Input membership function for relative strength index (RSI). Fig. 6. Example of momentum index. Fig. 8. Input membership function for momentum indicator (MTM). 11382 S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 These functions are depicted in Figs. 7–9. The fuzzy inference en- gine accepts relative strength index (RSI) and momentum indicator (MTM) as input and produces recommendations on whether or not to take a position (POS) as output. Momentum (MTM) is an oscillator type indicator used to detect overbought and oversold conditions and to perform as a gauge indicating the strength of the current trend. Momentum calcula- tions are either positive or negative and fluctuate around a zero- line MTMðtÞ¼ CðtÞ� Cðt � nÞ ð8Þ where C(t) is the closing price, n is the number of MTM periods. Fig. 6 shows an example of this indicator, which is below the primary price trend. From these membership functions, 27 inference rules can be implemented. Among these rules, five of them are selected for the experiments. 1. IF RSI IS whipsaw OR MTM IS whipsaw THEN POS IS doNothing. 2. IF RSI IS overSold AND MTM IS long THEN POS IS goLong. 3. IF RSI IS overBought AND MTM IS short THEN POS IS goShort. 4. IF RSI IS overSold AND MTM IS short THEN POS IS goShort. 5. IF RSI IS overBought AND MTM IS long THEN POS IS goLong. Whipsaw is a condition where a security’s price heads in one direction, but then is followed quickly by a movement in the opposite direction (Investopedia, 2012). Whipsaw pattern for rel- ative strength index (RSI) can be considered as a neural signal in terms of the velocity and magnitude of directional price move- ments. The security is considered to be in overbought territory when RSI is above 70 and considered to be over sold when RSI is below 30. Momentum shows the difference between today’s closing price and the closing price of n days ago (Wikipedia, Momentum (technical analysis), 2012). Momentum can be calcu- lated as follows: mometum ¼ closetoday � closen days ago ð9Þ The first fuzzy inference rule directs the trend following system to hold the current position if relative strength index and momentum are neutral. The second fuzzy inference rule directs the system to buy when relative strength index is oversold and momentum is high. Remaining inference rules can be interpreted in a similar way. The experimental results from simulation with Fuzzy TF algorithm are detailed in Section 4. In the next section, Fig. 9. Output membership function on taking a position (POS). S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 11383 we further extend the algorithm to take into account the volatil- ity in price. 2.4. Fuzzy TF with Volatility algorithm Fluctuation or volatility is one of the important aspects consid- ered by professional traders before making any trading decision. Volatility can be described as the intensity of the ups and downs in price. Volatility is often driven by factors such as emotion, greed, fear, and political situation. Optimism and greed can also drive the prices up when the majority of investors are making profit in up- trend market. During downtrend, fear and panic among investors can drive the prices into lower levels. There are two type of volatil- ity, Historical and Implied (Fontanills & Gentile, 2002). Historical volatility is a measure of price fluctuation over a time period in the past. Historical volatility is usually derived from standard devi- ation or from the average of price difference. Implied Volatility is often used on option trading. Implied Volatility is a measure of market expectations to an asset’s future volatility, which is implied Fig. 10. Example of simulated marke by the underlying asset’s option price. It is based on an option pric- ing model such as the famous Black-Scholes (Hull, 2000). Implied Volatility is often used on option trading. To find out how market fluctuation can affect our algorithm, we attempt to simulate the market with a controllable fluctuation. We then apply our trading algorithms on it, and observe how it per- forms under certain degree of fluctuation. Although it is almost impossible to model a real-world market fluctuation exactly in lab, we therefore turn to simulate the fluctuation scenario by nu- meric patterns. To do that we need to generate a series of simu- lated market data; these data must contain some degree of randomness, so it will behave similar to the market movement. Meanwhile we still need to control it, so it will move and fluctuate as expected. The following formula is designed to accomplish these: MarketDataðtÞ¼ ðCOSðeÞ� C � RðtÞÞþ B ð10Þ where COS is the geometry cosine which creates a basic oscillating wave structure, e is the angle that controls the fluctuation frequency, C is a constant that controls the fluctuation depth, R is a random number that creates the Whipsaw sharp, and B is a fixed base price that the generated prices will fluctuate along with. By controlling the parameter in the formula (10), we can gener- ate the market data that gradually increases the level of fluctuation, with a degree starting from 0 that indicates zero fluctuation, up to 100 that indicates maximum fluctuation. Fig. 10 shows some of the simulated market data sets in an intra-day scale chart. As the fluctuating data are generated, they are fed into the trad- ing simulation engine. The result is summarized in Fig. 11. It shows when the fluctuation is over 50%, it starts to make loss. On each time when the simulation data set is regenerated, though there might be some differences on the P&L numbers, but in general the observation of making loss gradually stay true, no matter how many times the simulated data set regenerate. Table 1 summarizes all fluctuation test result data. t trend with fluctuation control. Market Simulation P&L -3000 -2500 -2000 -1500 -1000 -500 0 500 1000 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 Fluctuation(%) M ea n (In de x P oi nt ) Fig. 11. Fluctuation test result. Table 1 Fluctuation test result data. Fluctuation (%) Mean Standard deviation Average trade 0 0 0 0 10 245 23 1 20 471 189 2 30 296 405 3 40 150 433 4 50 41 560 4 60 �297 642 5 70 �606 662 5 80 �729 707 5 90 �1953 474 6 95 �2515 446 7 11384 S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 In previous section, we have discussed how the market fluctua- tion (or volatility) can affect our trading algorithms. Given this vol- atility information, we can formalize and integrate it into our trading algorithm and it should help to improve our trading algo- rithm profitably. Based on this assumption, market fluctuation should be taken into account during decision making. There are many ways to calculate volatility. The most common one is stan- dard deviation of asset closing prices over a year. However, since we only concentrate on intra-day trades, it may not be very suit- able here. The central concept of volatility is finding the amount of change in the price of the asset over a period of time. Therefore we can measure it simply by the following formula: VolatilityðtÞ¼ SMAnððlnðpriceðtÞÞ� lnðpriceðt � 1ÞÞÞ� CÞ ð11Þ where ln is the natural logarithm, n is the number of periods, t is current time, C is a constant that enlarge the digit to significant figure. SMA (Simple Moving Average) is formed by computing the average (mean) price of a security over a specified number of peri- ods. It is possible to create moving averages from the Open, the High, and the Low data points, most moving averages are created using the closing price. For example, a 5-day simple moving aver- age is calculated by adding the closing prices for the last 5 days and dividing the total by 5 SMAðtÞ¼ CloseðtÞþ Closeðt � 1Þþ �� �þ Closeðt � n þ 1Þ n ð12Þ where n is the number of periods, t is the time. From these formulas, we can see that the volatility is actually a period of price difference in average, and there is no upper or lower boundary. When a price is advancing while increasing volatility, the formula result in positive, in contrary when price is falling while increasing in volatility, the formula results in negative; and when there is no volatility it tends to be zero. By observation, we find the highest reading of the equation is around ‘‘±18’’. Now if we follow on previous fluctuation results, when it is over 50% we should consider taking action, such as getting out of a position; and when it is under 10% we may not want to open any position, as there is a chance that we will make a loss. To include volatility measurement during decision-making, we add an additional factor into our fuzzy inference system. Fig. 12 shows the fuzzy membership function for volatility. Based on this additional input factor, we revise the fuzzy mem- bership rules as follows: 1. IF VIX IS nagHigh OR VIX IS posHigh THEN POS IS doNothing. 2. IF RSI IS whipsaw OR MTM IS whipsaw or VIX IS low THEN POS IS doNothing. 3. IF RSI IS overSold AND MTM IS long AND VIX IS pMiddle THEN POS IS goLong. 4. IF RSI IS overSold AND MTM IS short AND VIX IS nMiddle THEN POS IS goShort. 5. IF RSI IS overBought AND MTM IS short AND VIX IS nMiddle THEN POS IS goShort. 6. IF RSI IS overBought AND MTM IS long AND VIX IS pMiddle THEN POS IS goLong. The experimental results from simulation with Fuzzy TF with Volatility Algorithm are detailed in Section 4. 3. Trend recalling algorithm Since TF is an objective mechanism that is totally free from hu- man judgment and technical forecasting, the trends and patterns of the underlying data play an absolutely influential role in deciding its ultimate performance. It was already shown in Fong, Tai, and Si (2011) that market fluctuation adversely affects the performance of TF. Financial cycles are inevitable phenomena and it is a contro- versy whether cycles can be predicted or past values cannot forecast future values because they are random in nature. None- theless, we observed that cycles could not be easily predicted, but the abstract patterns of such cycles can be practically recalled by simple pattern matching. The formal interpretation of financial cycle (or better known as economic cycle) refers to economy-wide fluctuations in production or economic activity over several months or years. Here we consider it as the cycle that run contin- uously between bull market and bear market, some people refer this as market cycle (although they are highly correlated), In gen- Fig. 12. Volatility membership function. S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 11385 eral a cycle is made of four stages, and these four stages are: ‘‘(1) consolidation (2) upward advancement (3) culmination (4) decline’’ (Weistein, 1988). Despite being termed as cycles, they do not follow a mechanical or predictable periodic pattern. But similar patterns are observed to always repeat themselves in the future, just as a question of when, though in approximate shapes. We can anticipate that some exceptional peak (or other particular pattern) of the market trend that happen today, will one day hap- pen again, just like how it did happen in history. For instance, in the ‘‘1997 Asian Financial Crisis’’ (Wikipedia, 1997 Asian Financial Crisis, the Hang Seng Index, 2012) plunged from the top to bottom (in stages 3 to 4); then about 10 years later, the scenario repeats itself in the ‘‘2008 Financial Crisis’’ (Wikipedia, Financial Crisis of 2007–2010, 2012). Dow theory (Schannep, 2008) describes the market trend (part of the cycle) as three types of movement. (1) The ‘‘primary move- ment’’, main movement or primary trend, which can last from a few months to several years. It can be either a bullish or bearish market trend. (2) The ‘‘secondary movement’’, medium trend or intermediate reaction may last from 10 days to three months. (3) The ‘‘minor movement’’ or daily swing varies from hours to a Fig. 13. Intra-day of 2009-12-07 and day. Primary trend is a part of the cycle, which consist of one or several intermediate reaction and the daily swings are the minor movements that consist of all the detailed movements. Now if we project the previous assumption that the cycle is ever continu- ously rolling, into the minor daily movement, can we assume the trend that happens today, may also appear some days later in the future? Here is an example for this assumption; Fig. 13 shows two in- tra-day 2009-12-07 (top) and 2008-01-31 (bottom) trend graphs of Hang Seng Index Futures. Although they are not exactly the same, in terms of major upwards and downwards trends the two graphs do look alike. This observation provides a foundation for recalling trading strategies that are based on searching for similar patterns from the past. This concept is valid for TF because TF works by smooth- ing out the averages of the time series. Minor fluctuations or jitters along the trend are averaged out. This is important because TF is known to work well on major trending cycles aka major outlines of the market trend. Based on this underpinning, an improved version of Trend Fol- lowing algorithm is proposed in this section, which looks back to the past for reference of selecting the best trading strategy. The de- sign of a TF system is grounded on the rules that are summarized by Covel, into the following five questions (Covel, 2004): 1. How does the system determine what market to buy or sell at any time? 2. How does the system determine how much of a market to buy or sell at any time? 3. How does the system determine when you buy or sell a market? 4. How does the system determine when you get out of a losing position? 5. How does the system determine when you get out of a winning position? The first and second questions are already answered in our pre- vious work (Fong & Tai, 2009). The third question is rather chal- 2008-01-31 day trend graphs. Fig. 14. Improved TF process with recalling function. 11386 S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 lenging, that is actually the core decision maker in the TF system and where the key factor in making profit is; questions 4 and 5 are related to it. Suppose that we have found a way to identify trend signal (to buy or sell), and we have a position opened. Now if the system along the way identifies another trend signal, which complies with the current opened position direction, then we should keep it open, since it suggested that the trend is not yet over. However, if it is counter to the current position, we should probably get a close out, regardless whether you are currently win- ning or losing, as it indicates a trend reversion. Our improved TF algorithm is designed to answer this question: when to buy or sell. It is a fact that financial cycles do exist, and it is hypothesized that a trend on a particular day from the past could happen again some days later. This assumption supports the recall- ing trading mechanism, which is the basic driving force that our improved trend following algorithm relies on. The idea is expressed in Fig. 14. As it can be seen in the diagram there’s four major processes for the decision making. Namely they are Pre- processing, Selection, Verification and Analysis. Fig. 14 also shows the process of which our improved TF model works by recalling a trading strategy that used to perform well in the past by matching the current shape of the pattern to that of the old time. A handful of such patterns and corresponding trading strategies are short- listed; one strategy is picked from the list after thorough verifica- tion and analysis. Fig. 15. Example of EDM and prep 3.1. Pre-processing In this step, raw historical data that are collected from the past market are archived into a pool of samples. A sample is a day trend from the past with the corresponding trading strategy attached. The trend is like an index pattern for locating the winning trading strategy that is in the format of a sequence of buy and sell deci- sions. Good trading strategy is one that used to maximize profit in the past given the specific market trend pattern. This past pat- tern which is deemed to be similar to the current market trend, is now serving as a guidance to locate the strategy to be applied for decision making during the current market trade session. Since the past day trend that yielded a great profit before, reus- ing it almost can guarantee a perfect trading strategy that is supe- rior to human judgment or a complex time series forecasting algorithm. The past samples are referenced by best trading strate- gies on an indicator that we name it as ‘‘EDM’’ (exponential diver- gence in movement). EDM is a crisp value indicator that is based on two moving average differences EDMðtÞ¼ fðEMAsðtÞ� EMAlðtÞÞ ð13Þ EMAðtÞ¼ priceðtÞ� EMAðt � 1Þ� 2 n þ 1 � � þ EMAðt � 1Þ ð14Þ rocessed trend of 2009-12-17. S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 11387 where price(t) is the current price at any given time t, n is the num- ber of periods, s denotes a shorter period of EMA(t) at time t, l rep- resents a longer period EMA(t), f(�) is a function for generating the crisp result. The indicator sculpts the trend; and based on this infor- mation, a TF program finds a list of best trading strategies, which can potentially generate high profit. An example of pre-processing a 2009-12-07 day trend that shows the EDM is depicted in Fig. 15. As indicated from the diagram, the program first found a long posi- tion at 10:00 followed by a short position at around 10:25, then a long position at 11:25, finally a short position around 13:51 and closes it out at the end of the day, which reaps a total of 634 index points. In Hong Kong stock market, there is a two hours break be- tween morning and afternoon sessions. To avoid this discontinua- tion on the chart, we shift the time backward, and joined these two sessions into one, so 13:15 is equivalent to 15:15. Fig. 16. Example of 2009-12-07 sample with best 3.2. Selection Once a pool of samples reached a substantial size, the improved TF with recalling mechanism is ready to use. The stored past sam- ples are searched and the matching ones are selected. The goal of this selection process is to find the most similar samples from the pool, which will be used as a guideline in the forthcoming trad- ing session. A foremost technical challenge is that no two trends are exactly the same, as they do differ day by day as the market fluctuates in all different manners. Secondly, even two sample day trends look similar but their price ranges can usually be quite different. With consideration of these challenges, it implies that the sample cannot be compared directly value to value for a simple match. Some normalization is necessary for enabling some rough matches. Furthermore the comparison should allow certain level fitness (2008-01-31) day trend and RSI graph. Table 2 Fitness test applied on 2009-12-07 at the time 14:47. Table 3 Results of the overall simulation runs. Simulation Static Dynamic Fuzzy FyzzyVix Recalling Total Index Point 1115 1142 1476 2981 7201 Net worth (HKD) 55,750 57,100 73,800 149,050 360,050 Total trade 1088 987 1578 1149 1186 Cost 41,997 38,098 60,911 44,351 45,780 P&L 13,753 19,002 12,889 104,699 314,270 Total ROI (%) 19 26 17 141 425 11388 S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 of fuzziness. Hence each sample trend should be converted into a normalized graph, and by comparing their rough edges and mea- sure the difference, it is possible to quantitatively derive a numeric list of similarities. In pattern recognition, the shape of an image can be converted to an outline like a wire-frame by using some image processing algorithm. The same type of algorithm is used here for extracting features from the trend line samples for quick compar- ison during a TF trading process. In our methodology, each sample is first converted into a nor- malized graph, by calculating their technical indicators data. Indi- cators such as RSI and STC have a limited value range (from 1 to 100) which is suitable for fast comparison, and they are sufficient to reflect the shape of a trend. In other words, these indicators help to normalize each trend sample into a simple 2D line graph. We can then simply compare each of their graphical difference by superimposing these line graphs on top of each other for estimat- Fig. 17. Volatility membership d ing the differences. This approach produces a hierarchical similar- ity list, such that we can get around with the inexact matching problem and allows a certain level of fuzziness without losing their similarity attributes. Fig. 16 shows an example of two similar sam- ples trend graphs with the RSI displayed. efinition for trend recalling. -1000 0 1000 2000 3000 4000 5000 6000 7000 8000 20100104 20100201 20100303 20100331 20100503 20100601 20100630 20100729 20100826 20100924 20101025 20101122 20101220 Date In de x P oi nt Static Dynamic Fuzzy Fuzzy Vix Recalling Fig. 18. Performance comparison of five TF algorithms. Daily Profit and Loss -400 -300 -200 -100 0 100 200 300 400 500 600 Date In de x P oi nt Fig. 19. Daily profit and loss diagram of TF algorithm with recalling function. S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 11389 3.3. Verification During this process, each candidate from the list is tested against the current market state. Each trend sample corresponds to a specific trading strategy (that was already established in the pre-processing step). Each trading strategy is extracted and evalu- ated against historical to determine how well it performed as a trial. Each of their performances is then recorded and the trial per- formance is used as a criterion to rearrange the list. An example of the sample list before and after the fitness test is shown in Table 2. The fitness test was run on the 2009-12-07 during the middle of simulated trade session. Verification is needed because the selection of these candidates is by a best effort approach. That is because the market situations now and the past may still differ to certain extent. 3.4. Confirmation After the verification process is finished, the candidate list is re- sorted according to the fitness test results. The fittest one will be used as the reference of subsequent trading strategy during the TF decision making. In order to further improve the performance on top of the referencing to the past best strategy, some technical analysis is also referenced as well. By the advice of Weissman from his book (Weissman, 2004), the two-moving average crossover sys- tem is used as a signal confirmation. (The two-moving average crossover system entails the rise of a second, shorter-term moving average.) But instead of using simple moving average, EMA – expo- nential moving average with RSI should be used, that is a short- term RSI EMA and a long-term RSI EMA crossover system. When a trend is identified and it appears as a good trading signal, the crossover system must also be referenced and check if it gives a consistent signal. Otherwise the potential trend is considered as a false signal or noise. For example in our case the trading strategy from the recalled sample hints a long position trade. We check if RSI crossover system shows a short-term EMA crossing over its long-term EMA or not. In addition to validating the hinted trading signals from past strategy, market volatility should be considered during decision making. By observing how the equation responds to historical data, we can find the maximum volatility as ±15. Based on the previous fluctuation test result, we define the volatility fuzzy membership function for trend recalling in Fig. 17. During the trading session, volatility will be constantly refer- enced while the following rules apply at the TF system: 1. IF volatility is too positive high and long position is opened THEN close it out. 2. IF volatility is too positive high and no position is opened THEN open short position. 3. IF volatility is too low THEN do nothing. 4. IF volatility is too negative high and short position is opened THEN close it out. 5. IF volatility is too negative high and no position is opened THEN open long position. These rules have a higher priority over the trade strategies, such that when the condition has met any of these rules, it will take over 11390 S. Fong et al. / Expert Systems with Applications 39 (2012) 11378–11390 the control regardless of what decision that the trade strategies has made. The improved TF algorithm with recalling function is pro- grammed as an automated trading simulator, written in JAVA. 4. Experimental results In this section, we perform simulation experiments on the five trading algorithms previously proposed based on The Hong Kong Index Futures Contracts. All trials of simulations are run on histor- ical market data within 2.5 years prior to the year of 2010. A price that spread between bid and ask price is considered on each trade. In the simulation, each trade is calculated in the unit of index point. Each index point is equivalent to 50 HKD, which is subject to overhead cost as defined by Interactive Broker unbundled com- mission scheme at 19.3 HKD per trade. Six attributes are used to compare the performance of the algorithms including net worth, cost, and profit & lost. Net Worth ¼ Index Point � CM ð15Þ Cost ¼ Trade � C ð16Þ P&L ¼ Net Worth � Cost ð17Þ CM is the contract multiplier that is defined by the Hong Kong Fu- tures Exchange, which currently set as HKD 50 per index point. C is the commission that the broker charges. According to the IB (Group, 2012), bundled commission scheme during the year of 2007–2009, each trade cost 60 HKD commission. ROI is the prime performance index that is based on Hong Kong Exchange current Initial margin requirement (each contract 7400 HKD in year of 2010). From the simulation result from Table 3, we find that trend recalling algorithm achieves the best performance at the end of the experimental run. This is a significant result as it implies the proposed TF algorithm can reap more than three folds of whatever the initial investment is over just 2.5 years of trading. In Fig. 18, we summarize the performance comparison of five TF algorithms. Fig. 19 shows a horizon of trading results and we can observe that in overall, there are more profits than losses. 5. Conclusion and future work Trend following has been known as a rational stock trading technique that just rides on the market trends with some preset rules for deciding when to buy or sell. TF has been widely used in industries, but none of it was studied academically in computer science communities. In this paper, we pioneer in formulating TF into algorithms and evaluating their performance. First, we describe statistic and dynamic TF algorithms. Next, to improve the performance, we introduce fuzzy logic into our trade system, which forms our third and fourth versions of trading algo- rithms. The properties of these trading algorithms are generally built upon the experiences of previous trading algorithms, such as the membership definition and fuzzy sets generation. All these trading algorithms are later verified on Hang Sang Index futures market in a simulated environment, and the result is encouraging. These trading algorithms show an outstanding performance in the wild bullish and bearish markets between the years of 2007– 2009. However, during the year of 2010, they seem to be under performing, as they are no longer generating great profits. This is due to the large flip-flop changes of the market. We observe that frequent market trend fluctuations in large extents deter trend fol- lowing algorithms and cause the degradation in performance. To alleviate this problem, we extend the original TF algorithm by adding a market trend recalling function in our last trading algorithm. Trading strategy that used to make profit from the past is recalled for serving as a reference for the current trading. The trading strategy is recalled by matching the current market trend with the most similar past market trend which is known to pro- duce good profit. Matching market trend patterns is a complex task since patterns can be quite different in details. The concept of financial cycle was also taken into account of in the trend recalling algorithm so that the algorithm is adaptable to market behavioral changes. This trading algorithm is also verified on Hang Sang Index futures contracts in simulated environment. Our simulation results show that the improved TF model with trend recalling is able to generate profit from stock market at more than three times of ROI. In summary, contributions of this research are summarized as follows: � Developed an automated trading system prototype, which pro- vides a cornerstone for future development, and experimental platform for evaluating trading algorithms and programmed. � Provided an alternative view and comparative of two kinds of trading algorithms (predictive model vs. reactive model). � Proposed five innovative trading algorithms based on trend following concepts. As for the future work, we are planning to test the algorithms on other stock market data. We are also planning to further improve the calculation of P and Q in dynamic TF algorithms by applying time series segmentation methods. Acknowledgement Credits go to Tai Kam Fong, Jackie, who is the Master student of the University of Macau in year 2010, for his theoretic contribu- tions and for programming the simulator. References Covel, M. W. (2004). Trend following: How great traders make millions in up or down markets. Financial Times Prentice Hall. Fong, S., & Tai, J. (2009). The application of trend following strategies in stock market trading (pp. 1971–1976). <http://ieeexplore.ieee.org/xpl/articleDetails.jsp? arnumber=5331484>. Fong, S., Tai, J., & Si, Y. W. (2011). Trend following algorithms for technical trading in stock market. Journal of Emerging Technologies in Web Intelligence (JETWI), 3(2), 136–145. Fontanills, G. A., & Gentile, T. (2002). The volatility course. Wiley. Group, I. B. (2012). About the interactive brokers group. <http://www. interactivebrokers.com/en/general/about/about.php> Retrieved 18.02.12. Hull, J. C. (2000). Options, futures, and other derivatives (4th ed.). Prentice Hall. Investopedia (2012). Whipsaw. <http://www.investopedia.com/terms/w/whipsaw. asp> Retrieved 18.02.12. jFuzzyLogic (2012). An open source fuzzy logic library and FCL language implementation. <http://jfuzzylogic.sourceforge.net/html/index.html> Retrieved 18.02.12. Mayo, H. B. (2003). Financial institutions, investments, and management: An introduction (8th ed.). South-Western College Pub.. Murphy, J. J. (1999). Technical analysis of the financial markets: A comprehensive guide to trading methods and applications. New York Institute of Finance. Schannep, J. (2008). Dow theory for the 21st century: Technical indicators for improving your investment results. Wiley. Weissman, R. L. (2004). Mechanical trading systems: Pairing trader psychology with technical analysis (1st ed.). Wiley. Weistein, S. (1988). Stan Weinstein’s secrets for profiting in bull and bear markets (1st ed.). McGraw-Hill. Wikipedia (2012a). 1997 Asian Financial Crisis. <http://en.wikipedia.org/wiki/ 1997_Asian_Financial_Crisis> Retrieved 18.02.12. Wikipedia (2012b). Financial Crisis of 2007–2010. <http://en.wikipedia.org/wiki/ Financial_crisis_of_2007> Retrieved 18.02.12. Wikipedia (2012c). Momentum (technical analysis). <http://en.wikipedia.org/wiki/ Momentum_%28technical_analysis%29> Retrieved 18.02.12. Zadeh, L. (1973). Outline of a new approach to the analysis of complex system and decision process. IEEE Transactions on Systems, Man, and Cybernetics, 3(1), 28–44. http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=5331484 http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=5331484 http://www.interactivebrokers.com/en/general/about/about.php http://www.interactivebrokers.com/en/general/about/about.php http://www.investopedia.com/terms/w/whipsaw.asp http://www.investopedia.com/terms/w/whipsaw.asp http://www.jfuzzylogic.sourceforge.net/html/index.html http://www.en.wikipedia.org/wiki/1997_Asian_Financial_Crisis http://www.en.wikipedia.org/wiki/1997_Asian_Financial_Crisis http://www.en.wikipedia.org/wiki/Financial_crisis_of_2007 http://www.en.wikipedia.org/wiki/Financial_crisis_of_2007 http://www.en.wikipedia.org/wiki/Momentum_%28technical_analysis%29 http://www.en.wikipedia.org/wiki/Momentum_%28technical_analysis%29 Trend following algorithms in automated derivatives market trading 1 Introduction 2 Trend following algorithms 2.1 Static TF algorithm 2.2 Dynamic TF algorithm 2.3 Fuzzy TF algorithm 2.4 Fuzzy TF with Volatility algorithm 3 Trend recalling algorithm 3.1 Pre-processing 3.2 Selection 3.3 Verification 3.4 Confirmation 4 Experimental results 5 Conclusion and future work Acknowledgement References 
work_o7m65kde35b6je4cjstmqacd7y	----	Soft ontologies, spatial representations and multi-perspective explorability Soft ontologies, spatial representations and multi-perspective explorability Mauri Kaipainen, Peeter Normak, Katrin Niglas, Jaagup Kippar and Mart Laanpere KERG Knowledge Environments Research Group, Tallinn University, Estonia E-mail: mauri.kaipainen@tlu.ee Abstract: It is against the dynamically evolving nature of many contemporary media applications to be analysed in terms of conventional rigid ontologies that rely on expertise-based fixed categories and hierarchical structure. Many of these rely on sharing ‘folksonomies’, personal descriptions of information and objects for one’s own retrieval. Such applications involve many feedback mechanisms via the community, and have been shown to have emergent properties of complex dynamic systems. We propose that such dynamically evolving information domains can be more usefully described by means of a soft ontology, a dynamically flexible and inherently spatial metadata approach for ill-defined domains. Our contribution is (1) the elaboration of the so far intuitive concept of soft ontology in a way that supports conceptualizing dynamically evolving domains. Further, our approach proposes (2) a whole new mode of interaction with information domains by means of recurring exploration of an information domain from multiple perspectives in search of more comprehensive understanding of it, i.e. multi-perspective exploration. We demonstrate this concept with an example of collaborative tagging in an educational context. Keywords: folksonomies, tagging, exploration, multidimensional scaling, soft ontologies, ontospaces, interactive media, multi-perspective exploration 1. Introduction It is against the dynamically evolving nature of many contemporary media applications to be analysed in terms of conventional rigid ontolo- gies that rely on expertise-based fixed categories and hierarchical structure. In particular, appli- cations of the second-generation Internet, popu- larly referred to as Web 2.0, are typically characterized by user-contributed content and metadata. Many of these, e.g. user content sites like Flickr and YouTube and shared bookmark- ing applications like Del.icio.us, rely on ‘folkso- nomies’, i.e. the result of personal free tagging of information and objects for one’s own retrieval in a social environment (Vander Wal, 2004). Such applications involve many feedback me- chanisms via the community, and have been shown to have emergent properties of complex dynamic systems (e.g. Golder & Huberman, 2005; Halpin & Shepard, 2006). We propose that such dynamically evolving information domains can be more usefully de- scribed by means of a soft ontology (SO) (Aviles Collao et al., 2003), a dynamically flexible and inherently spatial metadata approach for ill- defined domains. Our contribution to the SO discussion is (1) the elaboration of the so far intuitive concept in a way that supports concep- tualizing dynamically evolving domains. Further, our approach proposes (2) a whole new mode of interaction with information do- mains by means of recurring exploration of an information domain from multiple perspectives DOI: 10.1111/j.1468-0394.2008.00470.x Article _____________________________ 474 Expert Systems, November 2008, Vol. 25, No. 5 c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd (multi-perspective exploration, MPX) in search of more comprehensive understanding of it, and thereby suggests new potential for interactive media and online communities. We demonstrate this concept with an example of collaborative tagging in an educational context. 2. Soft ontologies Soft ontologies (SOs) are flexible sets of meta- data that describe a domain of information by means of spatially conceptualized properties, ontodimensions, that jointly define the ontologi- cal space (ontospaces) in which an information domain ‘is’ or exists. We interpret that an SO is commensurate on a general level with Gruber’s (1993) definition of an ontology as ‘a specifica- tion of conceptualization’. Individual items of a domain are character- ized by values representing the degree of salience of each ontodimension (Aviles Collao et al., 2003). SOs are open-ended in the sense that they allow the creation of new ontodimensions, as well as the deletion of existing ones. Further, they are flat, i.e. not structured a priori by multi- level hierarchies. Instead, such a specification of an information domain can be interpreted as a priority order of organizing criteria, which in this sense corresponds to a hierarchical concep- tualization of a conventional ontology. The difference is that the implied hierarchy is malle- able and interactively explorable instead of being rigidly fixed a priori. Because of these characteristics, we suggest that SOs are better suited to modelling dynami- cally evolving information domains, such as those of collaborative tagging practices de- scribed by, for example, Vander Wal (2004) and Mathes (2004), than conventional rigid hierarchical ontologies. Relying on this concep- tualization, we propose (1) a formal definition of SO, which in turn supports (2) exploration of multiple perspectives to such domains by allow- ing each property to be taken into account to a degree chosen by the user. Formally, an SO is an open-ended coordinate system O¼ x1;x2; . . . ; xm½ � that defines shared m-dimensional ontospace A, i.e. the shared and expanding vocabulary of describing a domain D. Each item i of domain D can be represented by an m-tuple Ai ¼ ai1;ai2; :::; aim½ �, were aij stands for the salience of property xj with respect to item i, spatially interpreted as the position of item i with respect to ontodimension xj. In collaborative tag- ging applications, aij may represent the strength of tag j for item i calculated, say, by scaling the frequency of tag j for item i between 1 and 0. From a more general semantic point of view, aij allows a range of reading options depending on the nature of that property, e.g. presence, proximity, probability, strength-of-relation or agreement of item i with xj. As another inter- pretation, following Zadeh (1965), aij can be seen to stand for the degree of membership of item i in one-dimensional fuzzy set xj. 3. Reduction of dimensionality as a model of sense-making Several fields of research suggest quite consen- sually that organizing items to spatially laid-out clusters by their mutual similarity relations is the most natural strategy for making sense of the environment’s complexity. On the neural level, adaptive cortical maps, such as tonotopies (e.g. Hood, 1977; Wessinger et al., 1996), soma- totopies (e.g. Merzenich et al., 1988; Wall, 1988) and spatial representations (Olton et al., 1977), altogether suggest that mapping from multidi- mensional experiential space onto the cortical surface, i.e. dimensionality-reducing neural pro- cesses, is the physiological means of managing sense of the environment. As to language, Lakoff and Johnson have elaborated a theory according to which the very core elements of language and cognition are spatial metaphors (e.g. Lakoff & Johnson, 1980, 1999; Lakoff, 1986) originating from bodily– motor–spatial experiences, such as the expres- sions ‘under-stand’, ‘get-around’ or ‘up-load’. Further, in his geometrical approach to thought, Gärdenfors (2000, p. 258) proposes that spatial representations can model both dimensionality- reducing neural processes, such as discussed above, and a range of symbolic conceptualiza- tions, and can thereby serve as an explanatory c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd Expert Systems, November 2008, Vol. 25, No. 5 475 bridge between the neural and symbolic domains. Kohonen’s self-organizing map algorithm (1982), in turn, must be credited for the particular episte- mological value of demonstrating how two-di- mensional or three-dimensional representations, which are reminiscent of cortical projections on one hand and the result of order-preserving statistical algorithms on the other, can be ap- proximated with an extremely simple computa- tional abstraction of adaptive neuronal activity. Based on the above, our approach relies on the general assumption that a representation of com- plex information on a low-dimensional space can be considered a rather universal model of sense- making, to be referred to as the assumption of dimensionality reduction. Various algorithms exist for the purpose of revealing the structure of a multidimensional data set by producing approx- imating similarity-preserving representations of lower dimensionality. The most generic and well- known algorithms of these fall into the family of multidimensional scaling (MDS) (e.g. Kruskal & Wish, 1978; Kotz & Johnson, 1985). MDS algo- rithms represent items characterized by points on a low-dimensional (usually two-dimensional) Eucli- dean space so that the proximity of points reflects their mutual similarity. In our treatment we gen- erally refer to MDS even though other algorithms can also be accommodated with this formalization. Spatial representation of a finite set of items s¼fi1, i2, . . ., ing of a domain D by points in a lower-dimensional space B can be considered as a mapping Rs: fAi1; Ai2; . . . ; Aing! B. In or- der to satisfy the conceptualization of an SO it is necessary to introduce means that allow each property to be taken into account to the degree chosen by the user. For this purpose we define weights P¼ p1;p2; . . . ; pm½ �, 0rpjr1, for corre- sponding ontological dimensions xj of A. P is conceived of as an ontological perspective, defin- ing transformation P of A as P[x1, x2, . . ., xm]¼ [p1x1, p2x2, . . ., pmxm]. It should be emphasized that in our concep- tualization A itself is always a result from some a priori choice of perspective, i.e. a choice of a particular set of salience weights out of the multitude of possible sets. We argue that some ontological perspective P is always present, at least implicitly, typically due to the choice of the dimensions, or by means of statistical prepro- cessing, scaling, standardization or weighting. Therefore P should not be considered as an additional factor but rather as an intrinsic term. In our approach it is made explicit, accessible and negotiable by the user, instead of accepting it as given by the expert (author, designer, editor, owner etc.) of the domain. We claim that in this way our formalization reflects new own- ership relations of Web 2.0 media. Thus, a spatial representation RP,s of a finite set of items s¼{i1, i2, . . ., in} of a domain D consists of the transformation P: A ! A followed by application of an algorithm that preserves similarity patterns from an m-dimen- sional domain A to some q-dimensional domain B, qrm. In our application, P is used as the means of determining the desired degree to which each ontological dimension should be prioritized by a spatial representation. For the user-chosen values 0rpjr1, the extremes can be interpreted as follows: pj¼1 reflects the desire to maximize the preservation of the variance along dimen- sion xj and thereby prioritize the dimension over dimensions with lower values of p, while pj¼0 reflects the decision to totally ignore the variance along the dimension. Thus, it should be noted that transformation P will not preserve the distributions of all dimen- sions equally, but distorts some more than others. The potential of expressing the priority order of ontological dimensions in terms of ontological perspective P is an additional benefit with regard to searches from typical ‘folksonomies’, in which one either takes any particular tag (ontological dimension) into account or does not. 4. Multi-perspective exploration To distinguish our approach from standard applications in statistics (e.g. from MDS), we hold that a single representation corresponding to a particular perspective should not be re- 476 Expert Systems, November 2008, Vol. 25, No. 5 c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd garded as more than a transient and partially revealing view of multidimensionally complex information. We argue that a more profound comprehension emerges in the course of an iterative process of exploring the data from alternative ontological perspectives. As a meta- phor that helps to clarify this idea, one can consider a typical architectural design that can- not be fully comprehended just with a two- dimensional visualization, but for any better understanding it is instrumental to see the object from different perspectives using three-dimen- sional miniatures, computer-aided design visua- lizations or virtual reality models. This implies the assumption of some kind of cognitive system that binds together subsequent perceptions. An explanatory framework for what binds subsequent mappings together in the mind is the recursively iterative perceptual cycle of Neisser (1976, pp. 112–113) (Figure 1), in which perceptual exploration samples avail- able information in an object, of which the perception modifies the orienting schema, which again directs exploration, then feeding back to exploration, ad infinitum. Reflecting our approach against this concep- tualization we propose that the accumulating outcome of the explorative activity is like an orienting schema that keeps integrating subse- quent representations into increasingly encom- passing syntheses of a domain. If this kind of conceptualization is accepted, it is not mean- ingful in this framework to discuss MPX of SO spaces as an operation or algorithm with a predefined or fixed end condition. It is assumed that the activity of exploration is driven by the purpose of discovering new insights and qualities of the domain, or interdependences of the onto- logical dimensions. Given that aim, it is then the point of interpretive saturation, i.e. the point when new perspectives will not add anything of substantial importance to the understanding of the domain, when the user may decide to end the process of exploration. However, this end point should be taken only as ‘local’. 5. Demonstration In order to demonstrate the idea of MPX we have developed a prototype application, 1 in which the interface provides a set of sliders for the user to adjust and change perspectives P. The MDS representation RP;s: fAi1;Ai2; . . . ; Aing! B is performed and visualized as a real-time response to every interaction the user has with the inter- face. In the activity of exploration, a dimension to be taken fully into account, i.e. to be given priority as an organization criterion, is assigned the weight 1 (slider to the right) while a dimension the user wants to ignore totally remains with the weight 0 (slider on the left). In our example we use data 2 from a junior high school students’ class work assignment in which they were asked to tag learning materials of a project management course with their own de- scriptive keywords, creating a kind of ‘folkson- omy’, with the purpose of facilitating knowledge building of the domain of information and to help to identify right materials for reference. For the purposes of the present demonstration, every tag is represented as an ontological dimension, and a learning material document as an item. Corre- spondingly, the domain is described as a matrix of tag frequencies for every learning material. We consider two main strategies of explora- tion that contribute to more profound compre- Figure 1: Neisser’s perceptual cycle (adapted from Neisser, 1976). 1 Online demonstration at http://kerg.tlu.ee/demos/multi- perspective-exploration. 2 Data accessible at http://www.tlu.ee/imke/data/ProJuht_ EN.txt. c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd Expert Systems, November 2008, Vol. 25, No. 5 477 hension of the domain, i.e. reductive and induc- tive strategies. Within both, the observer may either pay attention to the overall clustering of the domain, or use the representation as a visualization of a search result where the per- spective is regarded as a multi-term search key allowing the user to adjust the weight for every particular search key according to its relative importance. In all cases, the user can promote or demote dimensions at will in the course of iterative exploration, i.e. either increase or de- crease their weights relative to other dimensions. 6. Reductive strategies Reductive strategies start with some given high- dimensional ontospace (Figure 2) that will be scrutinized in order to identify perspectives that are defined by fewer dimensions than the total dimensionality of the ontospace, and from which the domain appears ordered in a way that matches with the orienting schemas, i.e. existing knowledge. This, according to our dimensional- ity reduction assumption discussed above, con- tributes to better comprehension of the domain. Involved in reductive exploration, the obser- ver explores ontological dimensions one by one, looking for those whose weights can be demoted without causing additional ambiguity, i.e. overlapping item labels in the overall spatial representation. As a result of every change, a two-dimensional visualization of the similarity cluster representation computed by means of an MDS algorithm appears. The point of satura- tion is reached when a low-dimensional perspec- tive is identified that results in a sufficiently disambiguated representation. In the case of the search interpretation, the chosen perspective can be regarded as a search key representing the priority order of search Figure 2: A high-dimensional representation of the ontospace in the initial stage of a reductive exploration. The values (relative tagging frequencies) of the item pointed to by the cursor are circled. Items that are similar with respect to the chosen perspective are positioned near each other to form similarity clusters. For the search interpretation, the best match is pointed to by the cursor, and near- best hits to be explored are indicated with question marks. 478 Expert Systems, November 2008, Vol. 25, No. 5 c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd terms, leaving the overall spatial order as the secondary concern. For this interpretation, the best matching item is displayed at the top-right corner, a design feature of the demonstration software to comply with the convention of read- ing two-dimensional plottings in statistics. Search results visualized in this way are essen- tially more informative than standard one- dimensional lists of search outcomes, because the spatial layout supports immediate consid- eration of the position of the best hit with respect to the next-best ones (indicated by ques- tion marks in Figure 2), and allows its super- iority over the next-best hits to be estimated as perspectives change in the course of exploration. 7. Inductive strategies Inductive strategies of exploration are suited for making sense of ill-defined domains with weak or non-existent ontologies, i.e. for building ontologies from scratch. The goal is to identify sets of dimensions, i.e. perspectives, that cluster the domain items in some meaningful and co- herent manner with respect to existing knowl- edge of the domain (schema). The saturation point of this activity is when a compact set of perspectives is recognized that allows making sense of the domain as a whole, while keeping the dimensionality spatially comprehensible. Starting from no order at all, the activity proceeds by experimenting with dimensions one at a time, and observing the effect of each on the spatial representation using the slider controls to promote or demote dimensions. In them, the data are first clustered with respect to one dimension (Figure 3), and gradually to more dimensions. Initially the visual representation displays clusters of more or less superimposed labels. Additional dimensions may or may not contribute to revealing items obscured by others or by breaking tight clusters into sub-clusters. Figure 3: One-dimensional search as the initial setting of inductive exploration. Full weight is given to the dimension ‘project’ using the slider interface, and the corresponding value of the item under the cursor is indicated by the horizontal bars. Items with highest values on the dimension ‘project’ (highest tagging frequency) are piled on top of each other at the top-right corner. c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd Expert Systems, November 2008, Vol. 25, No. 5 479 The visual effect of each added dimension is perceptually self-explanatory, reminiscent of perspective change in the physical world. As with the reductive exploration strategy, in addition to being an overall visualization of the similarity clustering of the domain, a perspective can be interpreted as a search for the best-match- ing item (Figure 4, item pointed at by the cursor). 8. Adding ontological dimensions Adding new tags, treated in this framework as ontodimensions, at will is the defining charac- teristic of collaborative tagging practices. Add- ing dimensions is to be seen as another inductive strategy of making sense of complex informa- tion domains in the dimensionality-increasing direction. In this case, an added ontodimension constitutes a request for the community to apply the new tag, or in other words to evaluate content items with respect to the added onto- logical dimension and thereby provide addi- tional data, which allow the domain to be explored from new perspectives. In collaborative tagging practices it is a well- recognized problem that contributors often add new tags without checking whether there already exist tags with nearly the same meaning, which eventually leads to the accumulation of synonymous tags. This problem occurs because at present the environments and interfaces typi- cally do not allow easy exploration of the onto- space (tag space). Our approach can be regarded as a suggestion of how the quality of collabora- tive tagging applications can be improved by supporting the awareness of the community members of the ontospace. 9. Conclusions and implications The present approach implies new kinds of inter- active media, i.e. in which the user interacts directly with the ontology instead of taking for granted an implicit ontology predetermined by Figure 4: Additional ontodimension ‘planning’ is taken into account in addition to ‘project’, splitting the clusters of Figure 3 (double-ended arrow). The arrow points to the best matching item for the perspective, and the circled bar indicator indicates the respective data, as in Figure 3. 480 Expert Systems, November 2008, Vol. 25, No. 5 c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd somebody else, e.g. the designer or owner of the medium. We have argued that conventional fixed ontologies with hierarchical structure are not optimally suited for the analysis of dynamically evolving domains of information, e.g. those which involve non-constrained collaborative tagging, and we have claimed SOs to be more appropriate for that purpose. In addition to the flexibility of SOs with respect to the dynamically evolving dimensionality, we propose that the inherent spa- tial representation is their major advantage. On one hand, the latter allows dimensionality-redu- cing representations, such as those suggested by a range of research results across disciplinary boundaries, which we propose to serve as a universal model of sense-making. On the other hand, it allows dimensionality induction necessary in the case of ill-defined domains. We have demonstrated two main strategies of exploiting such explorability in a purpose-driv- en manner: (1) the reductive approach, i.e. identifying a range of perspectives that make sense of the ontospace as selective subsets of its dimensions, and (2) the inductive approach, i.e. iteratively promoting ontological dimensions to construct a comprehensive ontospace, yet with enough dimensions to distinguish domain items in a meaningful way. We argue that it is reason- able to compare the resulting knowledge to the comprehension of concrete artifacts, such as buildings or statues, which are hardly collapsi- ble to any single perspective. With respect to both strategies, the final out- come is not necessarily a single perspective but rather a range of perspectives that together contribute to overall comprehension of the do- main. Beyond the present treatment, different exploration strategies under the two main direc- tions are conceivable. The concept of MPX is more than just apply- ing one-time spatial visualization to complex content. We claim that the MPX concept sup- ports a new kind of self-directed interactive relationship between the user and abstract con- tent, comparable to navigation in computer- aided design or virtual reality environments. We suggest that the explorative activities ad- dressed above involve construction of mental schemas, i.e. understanding not only the data prima facie but also their underlying conceptual ontology via recursive action, in the manner of the Neisserian perceptual cycle, so that the immediate visualization of the exploration facil- itates and encourages this dynamic activity. There is a good reason to assume that making ontologies of abstract information domains inter- actively explorable allows going ‘beyond the in- formation given’ in the sense of Bruner (1973). We argue that this type of exploration is comparable to hands-on learning, say in science education, generally considered superior to top-down-dic- tated teaching, and we propose that the suggested kind of activity will contribute to knowledge building in line with constructivist and socio- cultural learning theories, where social negotiation of meaning is considered one of the key mechan- isms in human learning. Furthermore, when the ontospace is shared by the members of an online community, as in the case of shared tagging applications, then it is conceivable that MPX will contribute to joint sense-making (Golder & Hu- berman, 2005, p. 3) and social knowledge building. Bereiter and Scardamalia (2003) have defined knowledge building as a collaborative activity aimed at creation or modification of public knowl- edge. We argue that combining MPX with the use of discursive knowledge building tools will en- hance the quality and efficiency of the collabora- tive knowledge construction process thanks to spatial visualization of the SO, and also because it facilitates scaffolding (Bruner, 1975) both reduc- tive and inductive strategies of exploration. The potential application field of MPX be- yond collaborative tagging is as broad as the need to facilitate understanding of complex information domains in general. Elsewhere we have introduced MPX as a research tool in the context of mixed methods, i.e. hybrids of quan- titative and qualitative research (Niglas et al., 2008). As another type of potential application domain, one may consider that any corpus of text documents has an enormous number of hidden ontological dimensions, with respect to which each document is positioned in terms of frequencies of words. Such dimensions can be revealed at request, assuming search functional- c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd Expert Systems, November 2008, Vol. 25, No. 5 481 ities that return counts of word frequencies from the corpus, scaled and normalized in some reasonable way. Given this potential, the user could define his=her private and shareable do- main ontologies ‘softly’ at will and explore them with the method suggested in this paper. Furthermore, more generally multi-perspec- tive explorability of ontospaces relates to new ownership and intellectual property relations of the ‘democratized’ web media in terms of explicating the omnipresence of an ontological perspective and making it interactively negoti- able rather than taking the media owner’s per- spective for granted. Acknowledgements The work has been funded by European Social Funds, priority 1.1, and by Estonian Science Foundation grants 6148 and 7663. We thank Martin Sillaots for example data and Kaido Kikkas for comments. References AVILES COLLAO, J., L. DIAZ-KOMMONEN, M. KAIPAI- NEN and J. PIETARILA (2003) Soft ontologies and similarity cluster tools to facilitate exploration and discovery of cultural heritage resources, in 14th International Workshop on Database and Expert Systems Applications (DEXA ’03), available at http:// www2.computer.org/portal/web/csdl/doi/10.1109/ DEXA.2003.1232001. BEREITER, C. and M. SCARDAMALIA (2003) Learning to work creatively with knowledge, in Powerful Learning Environments: Unravelling Basic Compo- nents and Dimensions, E.D. Corte, L. Verschaffel, N. Entwistle and J. van Merrinboer (eds), Oxford: Elsevier Science. BRUNER, J.S. (1973) Going Beyond the Information Given, New York: Norton. BRUNER, J.S. (1975) The ontogenesis of speech acts, Journal of Child Language, 2, 1–40. GÄRDENFORS, P. (2000) Conceptual Spaces: On the Geometry of Thought, Cambridge, MA: MIT Press. GOLDER, S.A. and B.A. HUBERMAN (2005) The struc- ture of collaborative tagging systems, Information Dynamics, at http://www.hpl.hp.com/research/idl/ papers/tags/tags.pdf, accessed October 2007. GRUBER, T. (1993) A translation approach to portable ontologies, Knowledge Acquisition, 5 (2), 199–220. HALPIN, H. and H. SHEPARD (2006) Evolving ontolo- gies from folksonomies: tagging as a complex sys- tem, at http://www.ibiblio.org/hhalpin/homepage/ notes/taggingcss.html, accessed October 2007. HOOD, J. (1977) Psychological and psychological aspects of hearing, in Music and the Brain, M. Critchley and R. Henson (eds), London: Heinemann. KOHONEN, T. (1982) Self-organized formation of topo- logically correct feature maps, Biological Cyber- netics, 43, 59–69. KOTZ, S. and N.L. JOHNSON (1985) Encyclopedia of Statistical Sciences, Vol. 5, New York: Wiley. KRUSKAL, J.B. and M. WISH (1978) Multidimensional Scaling, Beverly Hills, CA: Sage. LAKOFF, G. (1986) Women, Fire, and Dangerous Things: What Categories Reveal about the Mind, Chicago, IL: University of Chicago Press. LAKOFF, G. and M. JOHNSON (1980) Metaphors We Live By, Chigago, IL: University of Chicago Press. LAKOFF, G. and M. JOHNSON (1999) Philosophy in the Flesh: the Embodied Mind and its Challenge to Western Thought, New York: Basic Books. MATHES, A. (2004) Folksonomies – cooperative classi- fication and communication through shared meta- data, at http://www.adammathes.com/academic/ computer-mediated-communication/folksonomies.html, accessed October 2007. MERZENICH, M.M., T. ALLARD, W.M. JENKINS and G. RECANZONE (1988) Self-organizing processes in adult neo-cortex, in Organisation of Neural Net- works: Structures and Models, W. von Seelen, G. Shaw and U. Leinhos (eds), Weinheim: VCH Verlagsgesellschaft. NEISSER, U. (1976) Cognition and Reality. Principles and Implications of Cognitive Psychology, San Fran- cisco, CA: W.H. Freeman. NIGLAS, K., M. KAIPAINEN and J. KIPPAR (2008) Multi-perspective exploration as a tool for mixed methods research, in Advances in Mixed Methods Reseach. Theories and Applications, M. Bergman (ed.), Beverly Hills, CA: Sage. OLTON, D.S., C. COLLISON and M.A. WERZ (1977) Spatial memory and radial arm maze performance of rats, Learning and Motivation, 8, 289. VANDER WAL, T. (2004) Folksonomy, at http:// vanderwal.net/folksonomy.html, accessed October 2007. WALL, J.T. (1988) Variable organization in cortical maps of the skin as an indication of the life- long adaptive capabilities of circuits in the mammalian brain, Trends in Neuroscience, 11 (12), 549–557. WESSINGER, C.M., M.H. BUONOCORE, C.L. KUSS- MAUL and G.R. MANGUN (1996) Tonotopy in hu- man auditory cortex examined with functional magnetic resonance imaging, Human Brain Map- ping, 5 (1), 18–25. ZADEH, L.A. (1965) Fuzzy sets, Information and Con- trol, 8, 338–353. 482 Expert Systems, November 2008, Vol. 25, No. 5 c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd The authors Mauri Kaipainen Mauri Kaipainen is Professor of New Media at Tallinn University and Guest Professor of Med- ia Technology at Södertörn University College. He studied education, musicology and cognitive science at the University of Helsinki and earned his PhD in 1994. He has worked with a number of media projects involving topics of learning and cognition, narrative spaces and logics, e- participation and cultural heritage. He has de- veloped the approach of modelling interactive media in terms of knowledge ecologies that involve continuous exchange of conceptual ar- tefacts with dynamically flexible spatially de- fined ontologies. Following this line, his present research activity is focused on defining the con- cept of interactively explorable multi-perspec- tive media, particularly suited for the analysis and design of applications for online commu- nities, collaboration and content sharing. Peeter Normak Peeter Normak is Professor of Informatics at Tallinn University. He studied mathematics at the University of Tartu and earned his PhD in 1982 from the M.V. Lomonossov Moscow State University. He is studying representations of semi-groups (algebraic automata), problems that belong both to algebra and to theoretical computer science. He has also coordinated a number of research and development projects in mathematical modelling, e-learning and curricu- lum development. Katrin Niglas Katrin Niglas is associate professor of data analysis at Tallinn University. She has taken part in various research projects in the fields of education, social sciences and humanities as an expert in methodology and data analysis and has successful experience in leading several re- search and development projects. Her main field of scientific interest as well as teaching is re- search methods and data analysis. After receiv- ing a teacher training diploma and MA degree in Tallinn, Katrin Niglas studied at the Univer- sity of Cambridge and obtained an MPhil de- gree in educational research. In her PhD dissertation, defended in 2004 at Tallinn Uni- versity, she focused on the combined use of qualitative and quantitative methods in educa- tional research. Her continuous work on elabor- ating mixed methods theory and practice is internationally recognized. She is a member of the editorial board of the International Journal of Multiple Research Approaches and a reviewer for Sage Publishers and for Journal of Mixed Methods Research. She also holds the post of an international correspondent of IASE (Interna- tional Association for Statistical Education). Jaagup Kippar Jaagup Kippar is a programming lecturer and programmer at Tallinn University. He studied in Tallinn Pedagogical University and gradu- ated with a BSc in natural sciences in 2000, and an MSc in didactics of informatics in 2002. He teaches different web technologies and has sig- nificant experience in Java. Mart Laanpere Mart Laanpere is head of the Centre for Educa- tional Technology in the Institute of Infor- matics, Tallinn University. He received his MSc in educational and training systems design from the University of Twente, The Netherlands. He has also studied in the Educational Sciences PhD programme at Tallinn University. He has participated in a number of international R&D projects as a researcher specializing in concep- tual design of virtual learning environments and didactics of e-learning. c� 2008 The Authors. Journal Compilation c� 2008 Blackwell Publishing Ltd Expert Systems, November 2008, Vol. 25, No. 5 483 
work_ocfkwuht7vbinb5hacdo2cw6ky	----	Microsoft Word - PRRES JOURNAL _FORMATTED_.doc Pacific Rim Property Research Journal, Vol 12, No 1 107 EXPERT SYSTEM PORTFOLIOS OF AUSTRALIAN AND UK SECURITISED PROPERTY INVESTMENTS CRAIG ELLIS University of Western Sydney and PATRICK WILSON University of Technology, Sydney ABSTRACT This paper examines whether a rule-based expert system is capable of outperforming the general property market, as well as randomly constructed portfolios from the market. While neural network expert systems have been used in property research, there appears little in the literature on the application of rule-based expert systems. Secondly, the paper considers whether there are benefits to international investment in real estate securities. The perspective of the analysis is that of an Australian investor investing in both the domestic market and the UK property market. Several results ensue including the failure of the rule-based system to significantly outperform the market or the random portfolios on a risk-adjusted returns basis, and the failure of hedging to secure a positive rate of return to the portfolio. Keywords: Expert system, international investment, hedging, portfolios INTRODUCTION One of the roles of the Australian Prudential Regulatory Authority (APRA) is to consider whether Australian financial institutions are obtaining diversification benefits through their international holdings of financial assets including real estate - both direct and indirect. The current paper considers this issue as part of a two-fold purpose as indicated below. This paper is an extension of previous research by the authors into the usefulness of Artificial Intelligence (AI) systems in portfolio construction (Ellis and Wilson, 2005). While that paper analysed the outcomes from neural network based systems in asset selection, this paper seeks to: (i) develop a rule-based expert system that can be used for selecting ‘value’ property stocks from the set of Australian and UK property stocks, and assess the ability of such a portfolio to outperform the general securitised property market in Australia; and (ii) to assess the viability of an Australian property investor receiving positive returns from either hedged or unhedged positions if investing in the Pacific Rim Property Research Journal, Vol 12, No 1 108 UK securitized property market. Why this may be important follows directly from other research that has questioned the viability of benefits arising from international diversification in real estate investment (eg: Ziobrowski and Ziobrowski (1993), Wilson and Zurbruegg (2003, 2004)) and the findings may be of some interest to both APRA as well as to property portfolio managers within Australian financial institutions. Several results follow from the paper’s analysis. While the rule-based expert system ‘value’ portfolios of Australian stocks are shown to outperform a randomly selected portfolio in most cases, they fail to outperform the Australian property market. Interestingly, the expert system ‘value’ portfolios selected from UK property stocks outperform both the Australian property market and the randomly selected portfolio in most instances. A central outcome is that none of these results are statistically significant and attempts to hedge AUD/GBP exchange rate movements result in capital losses to all portfolios, leading us to question the sustainability of long-term hedging in securitised property markets. BRIEF OVERVIEW OF EXPERT SYSTEMS One definition of a rule-based expert system is a computer programme that is capable of using information contained in a knowledge base, along with a set of inference procedures, to solve problems that are difficult enough to require significant human expertise for their solution (Harmon and King, 1985). The set of inference procedures are provided by a human expert in the particular area of interest, while the knowledge base is an accumulation of relevant data, facts, judgments and outcomes. There have been several studies on the use of different types of expert system in business. Wilson (1987) presents a number of applications of expert systems in finance, investment, taxation, accounting and administration, but points to the restrictions on the broader development of expert systems in business posed by the hardware limitations of the time. In their survey articles on the use of expert systems in business from 1977 through 1993, Wong and Monaco (1995a,b) note that expert systems have evolved and have been implemented as practical decision making tools in many businesses, documenting an extensive use of expert systems in various areas of finance such as investment analysis, stock market trading and financial planning. Despite this finding, Ellis Johnson et al (1997) find very little evidence of the application of formal decision support systems, such as expert systems, in the real estate sector. Liao (2005) conducts a review of the use of expert systems across all areas, including finance, over the period 1995 through 2004 and observes that expert systems provide a powerful and flexible means of obtaining solutions to a variety of problems and can be called upon as needed (when a human with expertise in the particular area may not be available). Of the different categories of expert systems as classified by Liao, there has been a broad examination of Neural Network (NN) based expert systems in property Pacific Rim Property Research Journal, Vol 12, No 1 109 research. For instance Borst (1991) examines the usefulness of NNs in predicting the selling price of real estate. Tay and Ho (1991), Do and Grudnitski (1992), Worzala et al (1995) and McGreal et al (1998) all examine the effectiveness of NN systems in property appraisal. Nguyen and Cripps (2001) compare the predictive accuracy of NNs against regression in forecasting housing values, and Wilson et al (2002) use NNs to forecast future trends within the UK housing market. Brooks and Tsolacos (2003) suggest that analysts should exploit the potential of NNs and assess more fully their forecast performance against more traditional models, and more recently, Ellis and Wilson (2005) examine the applicability of a neural network expert system in the construction of portfolios of Australian securitised property. It would appear that there has been little in the way of examination of the usefulness and applicability of rule-based expert systems in property market research. DATA AND SAMPLE The data utilised in this study comprises nominal monthly values for real estate stocks listed on the Australian Stock Exchange (ASX) and the London Stock Exchange (LSE) from January 1996 to February 2004 inclusive, a total of 98 months. The stocks selected are those listed in the Datastream Australian Real Estate Index and the Datastream UK Real Estate Index1. The Datastream Australian Real Estate Index comprises the top 80% of property sector stocks in the Australian market. Stocks included in the Index own property and derive their income from rental returns. As at February 2004, the Datastream Australian Real Estate Index comprised 21 companies and the Datastream UK Real Estate Index, 30 companies. The Datastream Australian Real Estate Index is used as a proxy for the market index, against which the performance of the expert systems ‘value’ portfolios is compared. Market capitalisation (MV), dividend yield (DY), price-to-book-value (PTBV), price-earnings (PE), price-to-cashflow (PC) and total return (TR) data is collected for the two market indices (Australia and the UK). As for the individual stocks, the Index total return represents the cumulative points gain or loss due to changes in the share prices of stocks in the index and normal dividend payments. The total number of company observations over the full sample period is 1766 for Australian real estate stocks, and 2758 for UK real estate stocks. For each company in the sample, the following data has been collected; closing price (P), market capitalisation (MV), dividend yield (DY), price-to-book-value (PTBV), price- earnings (PE) and price-to-cashflow (PC). Closing prices and the dividend yield for each company are used to calculate the total return index for each stock as follows: 1 A Datastream calculated index, the ‘Real Estate’ series is based on the FTSE classification and includes the following sub-sectors: Real Estate Development, Property Agencies, and Real Estate Investment Trusts. Pacific Rim Property Research Journal, Vol 12, No 1 110 1 1 1 1 100 12 t t t t t P DY RI RI P− − ⎛ ⎞ = × × + ×⎜ ⎟ ⎝ ⎠ (1) where Pt and DYt are the price and dividend yield at time t respectively, and RIt is the total return index. This formulation for the total return is identical to that used by Datastream to estimate the Return Index, and adjusts the total return for the monthly frequency used in this study.2 Total return is then calculated as the log difference of the total return index: 1log logt t tR RI RI −= − (2) As not all stocks traded for the full length of the test period (1997 - 2004), this avoids problems associated with survivorship bias that may influence the results (cf. Brown et al, 1992). For the purposes of estimating the total return to an Australian investor from a position in the UK market, monthly closing prices for the Australian Dollar / British Pound (AUD/GBP) exchange rate have also been collected. The total return to an Australian investor from an investment in UK assets includes both a capital gain or loss component, and an exchange rate gain (loss) component. RESEARCH METHODOLOGY The current paper uses as its expert knowledge inference set the outcomes from the research of Haugen (1995), O’Shaughnessy (1998), and Eakins and Stansell (2003) to select a group of fundamental financial ratios that will be used to determine sets (portfolios) of ‘value’ securitised property assets. These fundamental variables form the inputs to a rule-based expert system that has preset constraints to isolate ‘value’ assets. ‘Value’ assets are commonly defined as those whose market value is lower than their intrinsic or liquidating value (O’Shaughnessy, 1998). The attraction of value assets from an investor’s point of view is that the (lower) market value of the asset should rise to meet the (higher) intrinsic value. This being true, portfolios comprised of value assets should outperform portfolios comprised of all assets. Portfolio allocation to value stocks is sometimes referred to as a ‘contrarian investment strategy’. A number of studies indicate that contrarian strategies can outperform a strategy of investing in ‘growth’ 2 The Datastream model for total return is based on a daily frequency and uses 260, rather than 12, in the denominator. Ellis and Wilson (2005) use price and dollar dividend in their construction of the return index, which differs from the return index in this paper. Pacific Rim Property Research Journal, Vol 12, No 1 111 stocks (cf. Fama and French, 1996, 1998; Haugen, 1995; Lakonishok et al, 1994; Levis and Liodakis, 2001). In a comprehensive study of the comparative performance of US equities, O’Shaughnessy (1998) identifies a number of factors as being determinants of value. These include large stocks with low price/earnings ratios; low price/book ratios; low price/cashflow ratios; low price/sales ratios; high dividend yields. ‘Large’ stocks are defined by O’Shaughnessy as those with a higher than average market capitalization. Stocks with ‘low’ (‘high’) ratios are defined as those for which the relevant ratio is lower (higher) than the market average. For instance, a ‘low P/E’ ratio stock is one with a lower than market average P/E ratio, and a ‘high P/E’ ratio stock is one with a higher than market average P/E ratio. Given the lower volatility of large stocks relative to all stocks, value portfolios comprised of large stocks are shown by O’Shaughnessy to typically outperform the market index by a sizeable margin on a risk-adjusted basis. Using neural network techniques to predict value stocks using the ratios identified by O’Shaughnessy, Eakins and Stansell (2003) demonstrate the superior performance of the neural network value portfolio versus the S&P 500 and Dow Jones Industrial Average. Following the work of Eakins and Stansell (2003), Ellis and Wilson (2005) recently investigated the performance of value portfolios comprised of Australian real estate stocks using a similar neural network methodology to identify individual value stocks3. To go some way towards reducing the above mentioned deficiency in the literature on rule-based expert systems in property research, the current paper develops a rule-based expert system to construct portfolios of Australian and UK listed property stocks and compare the performance of these portfolios against the Datastream Australian Real Estate Index. The primary objective of research in this study is to examine the performance of expert system value portfolios comprised of property sector value stocks versus the set of all property sector stocks. The performance of each portfolio is compared first to the Australian market index, and second to mean statistics for 100 randomly diversified portfolios of Australian and UK listed property stocks. For portfolios comprised of UK listed property stocks, consideration is also given to the potential gain or loss from foreign exchange risk borne by Australian investors in the 3 There is an important difference between a rule based expert system, as used here, and a neural network. A rule based expert system operates in a strictly linear fashion following the if….then sequence of rules that have been embedded in the system. This contrasts starkly with that of a neural network in that the neural network is provided with a set of input and output conditions from which it `learns’ behaviour patterns. The input/output relationship is determined by some embedded non-linear function, consequently the `true’ non-linear relationship between the variables is not apparent – owing to the presence of hidden layers. So, the expert system is pre-determined to produce a certain output so long as the set of rules are satisfied. On the other hand, a neural network `learns’ according to the set of input/output conditions that exist in the training phase. The set of if… then expert rules in this paper have been obtained from the literature and are equivalent to the input conditions. Pacific Rim Property Research Journal, Vol 12, No 1 112 UK, and to the cost of hedging such risk and its impact on the profitability of foreign (UK) investments. Rather than the CAPM, which has not been shown to accurately reflect investor sentiment towards risk (Eakins and Stansell, 2003), risk adjusted performance relative to the Australian market index is measured first by the Sharpe ratio (Sharpe, 1966), and second the Sortino ratio (Sortino and Forsey, 1996; Sortino et al, 1997) for adjusting returns on a downside risk basis. The Sharpe ratio is the conventional approach to the measurement of risk in that it includes both upside and downside risk i.e. outperformance of the portfolio is treated exactly the same as underperformance. The Sharpe ratio is given as: r fp RRratioSharpe σ − = (3) where Rf is the risk free rate of return, Rp is the return to the portfolio and σr is the standard deviation of portfolio returns. The Sharpe ratio provides a measure of return per unit of risk with a ratio in excess of 1.0 being considered good. On the other hand the Sortino ratio is calculated as the difference between the portfolio return Rp and the minimum acceptable return MAR, divided by the downside deviation DD of the portfolio return versus the minimum acceptable return DDMAR. Downside deviation is similar to the loss standard deviation with the exception that it (DD) only includes portfolio returns below the MAR, rather than portfolio returns below the mean. The basis of the Sortino ratio is that investors are more concerned with the risk of loss (downside risk), than the risk of gains (upside risk). Standard deviation, as used by the Sharpe Ratio, considers both upside and downside risk. The calculation of the Sortino ratio is given as: p MAR MAR R R Sortino ratio DD − = N L DD N i i MAR ∑ == 1 2)( (4) 0)R(R if L or 0)R(R if RRL MARii MARiMARii >−= <−−= 0 )( Pacific Rim Property Research Journal, Vol 12, No 1 113 For consistency with the calculation of the Sharpe ratio, the Sortino ratio MAR is set equal to the mean monthly risk-free rate of 0.49%, this being the mean of the monthly Australian 10-year Commonwealth Bond rate for the test period.4 Several systems are tested in this study, including both single-factor and multiple-factor systems. Following from Ellis and Wilson (2005), the initial set of input factors employed for both the Australian and UK markets comprises market capitalisation (MV), dividend yield (DY), price-to-book-value (PTBV), price-earnings (PE), and price-to- cashflow (PC). Market average ratios (for comparison) are calculated annually as the average ratio for all stocks in the index each given year. Expert systems testing for UK listed property stocks also includes the effective rate of return (EffR), with the motivation being that Australian investors in the UK will gain from an appreciation of the foreign currency (GBP) net of any capital loss on the asset. To prevent look-ahead bias associated with making investment decisions based on data which is not yet known, portfolios constructed at time t comprise stocks which are identified by the system as being value stocks at time t-1. Despite the fact that the AUD/GBP exchange rate is readily observable on a 24-hour basis, information pertaining to other factors (eg. PTBV, PC) may not be readily known (Eakins and Stansell, 2003). Stock specific information pertaining to the first-order autocorrelation structure and mean excess percentage returns for each factor for the year prior is also included. The additional information is incorporated into the rule-based expert system to determine a priori the raw probability of a value stock at time t-1 being value at time t. While the full sample period comprises company and index data from 1996 - 2004 inclusive, the expert systems under consideration are tested over the period 1997 - 2004. Single-factor systems testing Single-factor expert systems in this study incorporate information pertaining to a single input factor only. Individual systems are tested for each of the five input factors listed above for the Australian market, and six (including information pertaining to monthly changes in the AUD/GBP) for the UK market. The rule-set developed a binary classification system so that the output factor for the single-factor system is defined as either VALUE or NOT according to the value criteria that was previously established for each respective input. Repeating the process for each of the input factors yields a total of five single-factor value portfolios for the Australian market and six single-factor value portfolios for the UK market. 4 Source, OECD Main Economic Indicators. Annual average equals 5.9% Pacific Rim Property Research Journal, Vol 12, No 1 114 Multiple-factor systems testing As opposed to the single-factor systems which classify stocks as being VALUE or NOT on the basis of a single criteria, only the multiple-factor rule-set ranks stocks in terms of the degree of value when measured against all criteria simultaneously. Individual stocks are awarded a cumulative score out of 5 in each period (score out of 6 in the UK) based on how many of the individual value criteria are identified by the expert system as being satisfied. For a stock listed in the Datastream Australian Real Estate Index, a cumulative value score of 5 in any given month would indicate that the stock satisfied all of the relevant value criteria. A value score of 0 alternatively shows that the stock satisfied none of the criteria. A cumulative score of 6 for a stock listed in the Datastream UK Real Estate Index would likewise show that the stock satisfied all of the value criteria in the UK market. Using the multiple-factor value criteria, ‘multi-value’ stocks are then defined as those with a cumulative value score ≥ 4. The testing of multi-value stock portfolios in addition to value stock portfolios (based on a single-value criteria only) is to evaluate the potential value-added by imposing a stricter value criteria on stocks during each period. Given the inclusion of the effective rate of return (EffR) as a value input in the UK market, two multi-value stock portfolios are tested in the UK. The first, Multi 1 comprises information pertaining to market capitalisation (MV), dividend yield (DY), price-to- book-value (PTBV), price-earnings (PE), and price-to-cashflow (PC) only, while the second, Multi 2 comprises these five plus information pertaining to the effective rate of return (EffR). Excluding the effective rate of return from the first UK multi-value portfolio allows the performance of this to be directly compared to the equivalent Australian multi-value portfolio. Its inclusion in the second UK multi-value portfolio allows the contribution of expected exchange rate changes to be measured separately. The binary classification in the multi-value context operates in much the same way as for the single-factor models: stocks scoring a cumulative value score ≥ 4 are classified as VALUE, stocks with a cumulative score < 4, NOT. RESULTS Descriptive statistics Summary statistics pertaining to monthly returns for Datastream Australian Real Estate Index and the Datastream UK Real Estate Index are provided in Table 1. The Datastream Australian Real Estate Price Index started the sample period (January 1996) at 634.77 points and finished on February 2004 at 1109.1. The Datastream UK Real Estate Price Index started at 1780.04 points and finished at 2864.05. The mean return to the Australian market index is approximately 1.02%, and for the UK market index is 0.77%. The AUD/GBP started at 2.0882 AUD per 1 GBP and finished at 2.3952 implying Pacific Rim Property Research Journal, Vol 12, No 1 115 Australian investors with a long position in GBP denominated assets would realize an exchange gain of about 0.3070 AUD per 1 GBP invested. Table 1 : Descriptive Statistics of Australian and UK Real Estate Index Returns, and AUD/GBP Exchange Rate, 1/01/1996 - 1/02/2004 Australia DS Real Estate UK DS Real Estate AUD/GBP Count 98 98 98 Mean 0.0102 0.0077 0.0025 Standard Deviation 0.0355 0.0489 0.0340 Excess Kurtosis 0.3090 0.3644 0.0668 Skewness 0.1190 -0.4406 0.2532 Minimum -0.0765 -0.1195 -0.0804 Maximum 0.1219 0.1340 0.0971 Sum of ln returns 1.0043 0.7557 0.2010 Value: start of period 634.77 1780.04 2.0882 end of period 1109.1 2864.05 2.3952 Gain (loss) 474.33 1084.01 0.3070 Runs test (p-value) 0.2160 0.4740 0.3100 Ljung-Box Q (1) (p-value) 0.0137 0.1351 0.2796 Anderson-Darling (p-value) 0.9490 0.0410 0.5290 Ryan-Joiner (p-value) > 0.1000 0.0840 > 0.1000 To establish the degree of randomness in monthly returns for the Australian and UK indices and the AUD/GBP, a non-parametric runs test is conducted. This test accepts randomness for all three series at the 10% level. Each series is then tested for the presence of first-order autocorrelation, the significance of which is tested using the Ljung-Box Q statistic. The significant test result for the Australian market confirms negative autocorrelation in the index, suggesting a higher than expected probability of consecutive price changes in the opposite direction. The Anderson-Darling test for the normality of returns fails to reject the null hypothesis that Australian market index returns follow a Normal distribution, but rejects the Normal null for the UK market index. The Normal null is accepted for the GBP/AUD at the 10% level. The Ryan-Joiner p-value similarly rejects the Normal null at the 10% level for the UK Market index and accepts the null for both of the Australian market index and the GBP/USD. Pacific Rim Property Research Journal, Vol 12, No 1 116 Expert system value portfolios The performance of each of the expert system value portfolios relative to the Australian market index is shown in a series of tables. Table 2 describes the performance of expert system value portfolios comprised of Australian real estate stocks, and Table 3 the performance of expert system value portfolios comprised of UK real estate stocks. Portfolios in both tables are compared to the Australian market index as proxied by the Datastream Australian Real Estate Index and to mean statistics for 100 randomly diversified portfolios drawn from the set of all Australian and UK listed property stocks. Table 2: Performance of Australian Expert System Value Portfolios: 1997 - 2004 MV DY PC PE PTBV Multi Market Random Mean Monthly Return 0.64% 0.37% 0.22% 0.39% 0.37% 0.36% 0.93% 0.44% Standard Deviation 0.0392 0.0286 0.0247 0.0268 0.0261 0.0285 0.0358 0.0284 Downside Deviation 0.0262 0.0211 0.0191 0.0200 0.0196 0.0204 0.0225 0.0207 Sharpe Ratio 0.0371 -0.0413 -0.1103 -0.0040 -0.0297 0.0542 0.1225 -0.0191 Sortino Ratio 0.0556 -0.0562 -0.1423 -0.0055 -0.0404 0.0813 0.1949 -0.0247 Max Monthly Gain 14.39% 8.91% 6.48% 7.80% 7.53% 10.30% 12.19% 8.23% Max Monthly Loss -9.93% -6.97% -7.40% -7.33% -6.67% -6.61% -7.65% -7.21% Cumulative Mean Return 54.16% 31.72% 18.65% 33.41% 31.71% 30.25% 79.06% 37.19% Compound Return 60.90% 32.57% 17.41% 35.40% 33.34% 30.70% 108.27% 67.59% Pacific Rim Property Research Journal, Vol 12, No 1 117 Nominal monthly mean returns for each of the expert system value portfolios in Table 2 are less than the mean return to the Australian market index (0.930%). The highest mean return is for the market capitalisation portfolio MV (0.640%) and the lowest is for the price-to-cashflow portfolio (0.392%). An analysis of z scores and p-values for the difference between mean returns to the market portfolios and the expert systems value portfolios reveals that the level of underperformance is not significant at the 0.10 level. Only the market value portfolio outperformed the bootstrap mean, though the difference is again insignificant at the 0.10 level. Compared to mean returns for each of the single- factor portfolios, p-values for the difference between the multi-factor portfolio mean return and single-factor mean returns is likewise insignificant.5 Cumulative mean returns for all portfolios over the test period are lower than the cumulative mean market return (79.064%) and the bootstrap mean cumulative return (37.188%) except for the market capitalisation portfolio which returns 54.164% for the 86-months to February 2004. Cumulative mean returns in this study are calculated as the sum of monthly portfolio returns and do not include the effects of monthly compound interest. Compound returns - the percentage return to $1 invested at the beginning of the test period and subsequently reinvested at each period’s monthly rate of return - are also calculated. Consistent with the cumulative returns, none of the expert system value portfolios in Table 2 were able to outperform the market, nor the bootstrap mean on a compound returns basis. Risk-adjusted returns in this study are calculated using the Sharpe ratio and the Sortino ratio. In contrast with findings already discussed, the difference between expert system value portfolio Sharpe ratios and Sortino ratios versus the market portfolio or the bootstrap mean Sharpe and Sortino ratios is statistically significant for all portfolios at the 10% level. Results pertaining to the performance of UK expert system value portfolios are presented in Table 3. Relative to the performance of the Australian market index, results in Table 3 are for the effective rate of return to an Australian investor from a position in UK (GBP denominated) assets. Calculated as /[(1 ) * (1 )] 1AUD GBP AUD GBPR R S= + + ∆ − (5) where RAUD is the AUD denominated total return, RGBP is the GBP denominated capital return, and ∆SAUD/GBP is the change in the spot Australian Dollar/British Pound exchange rate, returns in Table 3 include both a capital gain and exchange gain component. 5 z scores and p-values for the difference between the Multi-Value portfolio mean return and single-factor mean returns, not reported in this study, are available from the authors by request. Pacific Rim Property Research Journal, Vol 12, No 1 118 Table 3: Performance of UK Expert System Value Portfolios: 1997 - 2004 MV DY PC PE PTBV Mean Monthly Return 0.50% 0.52% 0.80% 0.81% 0.44% Standard Deviation 0.0515 0.0442 0.0430 0.0419 0.0434 Downside Deviation 0.0366 0.0319 0.0299 0.0289 0.0328 Sharpe Ratio -0.0579 0.1118 0.1614 0.1309 0.2269 Sortino Ratio -0.0760 0.1658 0.2520 0.2002 0.3874 Max Monthly Gain 12.78% 11.23% 10.01% 10.05% 10.04% Max Monthly Loss -11.46% -13.54% -10.07% -10.75% -11.60% Cumulative Mean Return 42.84% 44.52% 68.27% 68.43% 37.39% Compound Return 37.09% 43.45% 82.54% 83.57% 33.93% Multi1 EffR Multi2 Market Random Mean Monthly Return 0.30% 2.74% 1.43% 0.93% 0.87% Standard Deviation 0.0455 0.0373 0.0424 0.0358 0.0454 Downside Deviation 0.0353 0.0222 0.0265 0.0225 0.0313 Sharpe Ratio -0.0411 0.6030 0.2213 0.1225 0.0844 Sortino Ratio -0.0530 1.0113 0.3539 0.1949 0.1243 Max Monthly Gain 9.63% 12.26% 9.54% 12.19% 11.13% Max Monthly Loss -17.01% -8.71% -11.01% -7.65% -11.96% Cumulative Mean Return 25.89% 248.96% 122.76% 79.06% 74.22% Compound Return 18.38% 772.2% 196.74% 108.27% 105.20% Inclusive of a mean monthly exchange rate gain of approximately 0.186%, mean returns to all single factor portfolios, except the effective rate of return value portfolio EffR, underperform both the Australian market index return and the bootstrap mean return. The multi-factor model Multi 2, which incorporates information pertaining to exchange rate Pacific Rim Property Research Journal, Vol 12, No 1 119 changes, also outperform both the market and the bootstrap mean. The effective rate of return value portfolio earns the highest mean monthly return (2.742%) and the multi- factor portfolio Multi 1, the lowest (0.186%). With the exception of the effective rate of return value portfolio, expert system value portfolio mean monthly returns are not significantly different to either the mean monthly market return or the bootstrap mean return. Cumulative mean and compound returns to the above listed portfolios also exceed the market portfolio and bootstrap mean returns, though the statistical significance of the difference in cumulative mean returns cannot be ascertained. Sharpe and Sortino ratios for all of the portfolios in Table 3, except the price-earnings value portfolio PE, are likewise significantly different to the market portfolio Sharpe and Sortino ratios. Of these, only the effective rate of return value portfolio yields Sharpe and Sortino ratios which are statistically greater than the corresponding market ratios. Hedging foreign exchange exposure For Australian investors with open positions in foreign currency denominated assets, two questions arise; namely: what is the contribution of exchange gains or losses to overall portfolio performance; and whether the underlying exchange rate risk should be actively managed. Appreciation of the foreign currency will increase the domestic currency value of foreign currency denominated assets, and depreciation will decrease the value of foreign currency denominated assets. Results presented in Table 3 comprise both a capital gain (loss) component and an exchange rate gain (loss) component, i.e. the effective rate of return to an Australian investor. To consider the impact of currency variability on portfolio returns in Table 3, Table 4 provides comparative returns for the cases where (i) the AUD/GBP exchange rate is fixed, and (ii) the exchange risk is fully hedged. Pacific Rim Property Research Journal, Vol 12, No 1 120 Table 4: AUD/GBP Exchange Rate Variability and the Performance of UK Expert System Value Portfolio: 1997 - 2004 MV DY PC PE PTBV Multi1 EffR Multi2 Fixed Exchange Rate Mean Monthly Return 0.39% 0.41% 0.69% 0.69% 0.33% 0.18% - - Standard Deviation 0.0540 0.0469 0.0460 0.0449 0.0469 0.0464 - - Downside Deviation 0.0403 0.0370 0.0353 0.0343 0.0384 0.0381 - - Sharpe Ratio -0.0182 -0.0170 0.0436 0.0450 -0.0341 -0.0663 - - Sortino Ratio -0.0244 -0.0215 0.0569 0.0589 -0.0415 -0.0806 - - Max Monthly Gain 15.92% 10.22% 11.32% 10.10% 12.68% 12.39% - - Max Monthly Loss -12.51% -15.67% -11.71% -12.17% -13.77% -19.05% - - Cumulative Mean Return 33.42% 35.04% 58.86% 58.98% 28.22% 15.69% - - Compound Return 23.24% 28.97% 64.10% 65.04% 20.42% 6.42% - - Costless Hedging Mean Monthly Return 2.63%== 1.71% 2.93%== 2.00%= 2.56%== 2.41%== 5.73%== 3.64%== Standard Deviation 0.0546 0.0460 0.0450 0.0437 0.0462 0.0466 0.0404 0.0441 Downside Deviation 0.0287 0.0237 0.0220 0.0210 0.0255 0.0225 0.0093 0.0191 Sharpe Ratio 0.3913 0.4669 0.5418 0.5579 0.4475 0.4124 1.2966 0.7139 Sortino Ratio 0.7456 0.9083 1.1095 1.1612 0.8123 0.7068 5.6326 1.6483 Max Monthly Gain 14.51% 12.31% 13.25% 13.44% 12.73% 11.94% 16.64% 13.51% Max Monthly Loss -11.46% -12.67% -10.07% -10.75% -11.39% -16.17% -8.93% -11.01% Cumulative Mean Return 223.43% 224.52% 248.81% 248.91% 217.61% 205.25% 436.38% 306.77% Compound Return 703.51% 740.83% 969.91% 975.98% 685.120% 593.84% 2017.4% 1176.2% = significantly different to the Datastream Australian Real Estate Index at the 0.10 level = = significantly different to the Datastream Australian Real Estate Index at the 0.05 level Pacific Rim Property Research Journal, Vol 12, No 1 121 While portfolio mean returns net of GBP appreciation may be inferred from Table 3 via subtraction of the mean monthly forex gain from the mean monthly return, Table 4 provides exact mean returns for the UK value portfolios exclusive of exchange rate gains or losses. Assuming that the AUD/GBP exchange rate is at par for the entire test period, the change in the AUD/GBP in Equation (3) is therefore zero. In terms of cumulative returns, appreciation of the GBP over the test period can be seen to add from 9.17% (price-to-book-value) to 10.20% (Multi 1). Compound returns are approximately 11.96% (Multi 1) to 18.53% (price-earnings) higher. Despite the fact that mean returns in Table 3 are not statistically different to those given a fixed exchange rate in Table 4, the difference in compound returns illustrates the value of foreign currency appreciation to Australian investors. Expert systems value portfolios purchased at time t use all available information up to the previous period, t-1. By avoiding look-ahead bias, the investor now bears the risk that changes in the values of expert system input variables (used to designate a stock as being VALUE or NOT) between t and t-1 will result in capital and/or exchange losses. To manage exchange rate risk between time t-I and t, the hedge in Table 4 is constructed along the following lines: at time t-1, the investor purchases an AUD/GBP foreign exchange option with an exercise value equal to the then current spot exchange rate. The option maturity is the next period, time t. If the GBP depreciates between time t-1, and t, the option is exercised and the portfolio effective rate of return is calculated on the change in the AUD/GBP to time t-1. Else if the GBP appreciates between time t-1 and t, then the option expires worthless and the portfolio effective rate of return is calculated on the change in the AUD/GBP to time t. The hedge strategy follows from that employed by Ziobrowski and Ziobrowski (1993) into the benefits to US investors of hedging long- term positions in British Pound and Japanese Yen denominated real estate stocks with currency options. Despite the short-term gains from managing exchange rate risk with currency options, the authors concluded that ‘as a long-term strategy, the currency option offers no real protection against foreign asset devaluation caused by currency devaluation’ (Ziobrowski and Ziobrowski, 1993). Under the initial assumption that the above described hedge is costless, i.e. the option premium is zero, results for the Costless Hedging expert systems value portfolios in Table 4 show a significant degree of outperformance relative to the Datastream Australian Real Estate Index with cumulative returns as high as 336.38%, and compound returns as high as 2017.4% for the effective rate of return portfolio. Furthermore, mean returns given costless hedging in Table 4 are also significantly greater than their respective unhedged values in Table 3 at the 0.10 level, and returns to all except the dividend yield and price-earnings value portfolios are significantly greater at the 0.01 level. Relative to mean monthly returns presented in Table 3, the additional mean monthly portfolio return of approximately 1.19% to 2.99% implies a potentially substantial benefit from hedging foreign exchange risk. Pacific Rim Property Research Journal, Vol 12, No 1 122 Under the assumption of costless hedging with currency options, fully hedged returns in Table 4 have already been shown to be significantly greater than their unhedged values. To determine the impact of the cost of the option premium on portfolio returns, Table 5 provides the return to an initial $1 investment, reinvested each period over the full sample for Australian expert systems portfolios, and for each of the UK expert system value portfolios given the cases where (i) exchange rate risk is unhedged, (ii) exchange risk is costlessly hedged (Hedged Gross), (iii) the real cost of the option premium is deducted from each period’s reinvested value (Hedged Net)6. Real option premiums each period are calculated using the Garman and Kohlhagen (1983) modified Black-Scholes model for valuing foreign currency options: 1 2( ) ( ) fr t rtC Se N d Xe N d− −= − 2 1 1 ln 2f S r r t Xd t σ σ ⎛ ⎞ ⎛ ⎞+ − +⎜ ⎟ ⎜ ⎟ ⎝ ⎠ ⎝ ⎠= (6) 2 1d d tσ= − where C is the call option premium, S = X are the spot AUD/GBP exchange rate at time t-1 and option exercise respectively, r is the mean of the monthly Australian 10-year Commonwealth Bond rate, rf is the mean of the monthly UK 10-year Commonwealth Bond rate, and σ is the annualized standard deviation of AUD/GBP monthly returns from January 1990 to December 1996. The average call option premium over the test period is 0.0419AUD and the total premium paid is 3.6009AUD. 6 The UK - Hedged Net beginning of period value of $0.96 is calculated as the $1 initial investment less the initial option premium. Pacific Rim Property Research Journal, Vol 12, No 1 123 Table 5: Returns from $1Invested in Expert System Value Portfolios: 1997- 2004 MV DY PC PE PTBV Multi 1 EffR Multi2 Australia Beginning-of-period value $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 - - End-of-period value $ 1.61 $ 1.33 $ 1.17 $ 1.35 $ 1.33 $ 1.31 - - UK Unhedged Beginning-of-period value $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 End-of-period value $ 1.37 $ 1.43 $ 1.83 $ 1.84 $ 1.34 $ 1.06 $ 9.43 $ 2.65 UK Hedged Gross Beginning-of-period value $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 End-of-period value $ 8.04 $ 8.41 $ 10.70 $ 10.76 $ 7.58 $ 6.94 $ 21.80 $ 13.46 UK - Hedged Net Beginning-of-period value $ 0.96 $ 0.96 $ 0.96 $ 0.96 $ 0.96 $ 0.96 $ 0.96 $ 0.96 End-of-period value -$ 2.06 -$ 2.59 -$ 2.50 -$ 2.51 -$ 2.50 -$ 3.16 $ 0.87 -$ 1.25 Unhedged Req'd Monthly Return 1 4.42% 4.43% 4.47% 4.47% 4.42% 4.38% 5.06% 4.56% Hedged Gross Req'd Monthly Return 2 4.97% 5.00% 5.13% 5.13% 4.94% 4.90% 5.58% 5.26% Breakeven Req'd Monthly Return 3 4.38% 1 Average monthly required rate of return for which the Hedged Net end-of-period value equals the Unhedged end-of-period value 2 Average monthly required rate of return for which the Hedged Net end-of-period value equals the Hedged Gross end-of-period value 3 Average monthly required rate of return for which the Hedged Net end-of-period value equals $1.00 Pacific Rim Property Research Journal, Vol 12, No 1 124 Consistent with compound returns in Table 2, returns to $1 invested for all of the Australian expert system portfolios are less than would be earned by investment in the general property market. For UK expert system portfolios, end-of-period values for the hedged net portfolios are all lower than their respective unhedged and hedged gross end- of-period values. Except for the effective rate of return portfolio, end-of-period UK hedged net values are negative. As illustrated by Figure 1, this result can be seen to be due to the cumulative impact of the premium on the future value of reinvested returns. Figure 1: Mean of Compound Percentage Returns for UK Expert System Value Portfolios: 1997 – 2004 Despite the fact that UK hedged gross end-of-period values minus the nominal total premium exceeds the UK unhedged end-of-period values for all portfolios and the Australian market index end-of-period value, the subtraction of a premium each period reduces the value reinvested in the next period. This effect is cumulative over time resulting in much lower future values (compound returns) for the portfolios. Based on the average monthly call option premium of 0.0419AUD, estimation of the average monthly required rate of return in Table 5 shows that Australian investors would have to earn C om po un de d R et ur n (% ) Pacific Rim Property Research Journal, Vol 12, No 1 125 approximately 4.375% compounded monthly for the end-of-period UK hedged net portfolio value to equal the initial $1 invested. This result is consistent with the earlier findings of Ziobrowski and Ziobrowski (1993) for US investors and confirms the incapacity of long-term hedged positions to return a yield in excess of the compound value of the option premium. The result does not preclude the viability of a short-term hedge strategy where, for instance a derivatives position is taken only when there is a forecast likelihood of foreign currency depreciation. The determination of a short-term hedge strategy is not attempted in the current study as the analysis of the AUD/GBP (refer Table 1) suggests the exchange rate cannot be accurately forecast. CONCLUSIONS There are a number of conclusions to flow from this study. First, the use of a rule-based expert system is capable of beating the general property market and randomly selected portfolios in select cases, although the outperformance is not statistically significant. This is in contrast to the outcomes from Ellis and Wilson (2005) who use a neural network system to develop portfolios that consistently outperformed the market. A possible explanation for this may be due to the fact that a neural network system is capable of picking up non-linear relationships that are unable to be identified by the rule-based system. The crucial finding in the paper is there appears to be no long term benefit to the Australian investor through hedging exposure to fluctuations in the AUD/GBP exchange rate, which is in broad agreement with the findings of Ziobrowski and Ziobrowski (1993) in relation to hedging US real estate. Further to Ziobrowski and Ziobrowski, we show this result is not due to the total cost of the premium, but rather is due to the continuous impact of the premium on compounded returns. REFERENCES Borst, R.A. 1991. Artificial Neural Networks: The Next Modeling/Calibration Technlogy for the Assessment Community? Property Tax Journal, 10(1), 69-94. Brooks, C. and Tsolacos, S. 2003. International Evidence on the Predictability of Returns to Securitised Real Estate Assets: Econometric Models vs Neural Networks. Journal of Property Research, 20(2), 133-156. Brown, S.J., Goetzmann, W., Ibbotson, R.G. and Ross, S.A. 1992. Survivorship Bias in Performance Studies. Review of Financial Studies, 5(4), 553-580. Do, A.Q. and Grudnitski, G. 1992. A Neural Network Approach to Residential Property Appraisal. The Real Estate Appraiser, December, 38-45. Pacific Rim Property Research Journal, Vol 12, No 1 126 Eakins, S.G. and Stansell, S.R. 2003. Can Value-based Stock Selection Criteria Yield Superior Risk-adjusted Returns: An Application of Neural Networks. International Review of Financial Analysis, 12, 83-97. Ellis, C. and Wilson, P. 2005. Can a Neural Network Property Portfolio Selection Process Outperform the Property Market? Journal of Real Estate Portfolio Management, 11(2),105-121. Ellis Johnson, L., Redman, A.L., and Tanner, J.R. 1997. Utilization and Applications of Business Computing Systems in Corporate Real Estate. Journal of Real Estate Research, 13(2), 211-230. Fama, E.F. and French, K.R. 1996. Size and Book-to-Market Factors in Earnings Returns. Journal of Finance, 51(1), 131-155. Fama, E.F and French, K.R. 1998. Value versus Growth: The International Evidence. Journal of Finance, 53(6), 1975-1999. Garman, M.B. and Kohlhagen, S.W. 1983. Foreign Currency Option Values. Journal of International Money and Finance, 2, 231-237. Harmon, P. King, D. 1985. Artificial Intelligence in Business – Expert Systems. John Wiley and Sons: New York. Haugen, R.A. 1995. The New Finance: The Case Against Efficient Markets. Prentice- Hall: New Jersey. Lakonishok, J., Shleifer, A. and Vishny, R. 1994. Contrarian Investment, Extrapolation and Risk. Journal of Finance, 49(5), 1541-1578. Levis, M. and Liodakis, M. 2001. Contrarian Strategies and Investor Expectations. Financial Analysts Journal, 57(2), 43-56. Liao, S. 2005. Expert System Methodologies and Applications – A Decade Review from 1995 to 2004. Expert Systems with Applications, 28, 93-103. McGreal, S., Adair, A., McBurney, D., and Patterson, D. 1988. Neural Networks: The Prediction of Residential Values. Journal of Property Valuation and Investment, 16(1), 57-67. Nguyen, N. and Cripps, A. 2001. Predicting Housing Value: A Comparison of Multiple Regression Analysis and Artificial Neural Networks. Journal of Real Estate Research, 22(3), 313-336. Pacific Rim Property Research Journal, Vol 12, No 1 127 O’Shaughnessy, J.P. 1998. What Works on Wall Street. Revised Edition. McGraw-Hill: New York. Sharpe, W.F. 1966. Mutual Fund Performance. Journal of Business, January, 119-138. Sortino, F.A. and Forsey, H.J. 1996. On the Use and Misuse of Downside Risk. Journal of Portfolio Management, 22(2), 35-42. Sortino, F.A., Miller, G.A., and Messina, J.M. 1997. Short-Term Risk-Adjusted Performance: A Style-Based Analysis. Journal of Investing, 6(2), 19-28. Tay, D.P.H. and Ho, D.K.K. 1992. Artificial Intelligence and the Mass Appraisal of Residential Apartment. Journal of Property Valuation & Investment, 10, 525-540. Wilson, I.D., Paris, S.D., Ware, J.A., and Jenkins, D.H. 2002. Residential Property Time Series Forecasting with Neural Networks. Journal of Knowledge-Based Systems, 15, 335-341. Wilson, P.J. and Zurbruegg, R. 2004. Contagion or Interdependence? Evidence from Co- movements in Asian Securitised Real Estate Markets During the 1997 Crisis. Journal of Property Investment and Finance, 22(5), 401-413. Wilson, P.J. and Zurbruegg, R. 2003. International Diversification of Real Estate Assets – Is it Worth It? Evidence from the Literature. Journal of Real Estate Literature, 11(3), 259-278. Wilson, P.J. 1987. Expert Systems in Business, Vols. 1 and 2. MTE: Sydney. Wong, B.K. and Monaco, J.A. 1995a. A Bibliography of Expert System Applications in Business (1984-1992), European Journal of Operational Research, 85, 416-432. Wong, B.K. and Monaco, J.A. 1995b. Expert System Applications in Business: A Review and Analysis of the Literature (1977 – 1993). Information Management, 29, 141- 152. Worzala, E., Lenk, M. and Silva, A. 1995. An Exploration of Neural Networks and its Application to Real Estate Valuation. Journal of Real Estate Research, 10(2), 185-201. Ziobrowski, A.J. and Ziobrowski, B.J. 1993. Hedging Foreign Investments in U.S. Real Estate with Currency Options. Journal of Real Estate Research, 8(1), 27-54. 
work_ocjgoxlbkrhmjj7njm3dfudbte	----	Fuzzy linear regression-based detection of earnings management Fuzzy linear regression-based detection of earnings management Henrik Höglund ⇑ Hanken School of Economics, Handelsesplanaden 2, 65101 Vasa, Finland a r t i c l e i n f o Keywords: Earnings management Discretionary accruals Fuzzy linear regression a b s t r a c t A large number of accounting studies have examined the occurrence of earnings management in various contexts. In most of these studies, the earnings management detection model is based on the linear regression model suggested by Jones (1991). A considerable problem with the Jones model is the require- ment of long time series of financial statement data. An alternative to estimating the linear regression model coefficients with ordinary least squares (OLS) is to use fuzzy linear regression (FLR) instead. One of the main advantages with FLR described in the literature is its ability to handle small data sets. The purpose of this study is to compare the performance of the OLS-based Jones model with the perfor- mance of the FLR-based Jones model. The results show that the performance of both types of models decreases when the length of the time series decreases and that there is no significant difference in the estimated discretionary accruals between the models. The results also show that the FLR-based Jones model outperforms the OLS-based Jones model in detecting simulated earnings management when the estimation time series is short. Overall, the results show that the FLR-based Jones model is a feasible alternative to the OLS-based Jones model, especially when the length of the estimation time series is restricted by data availability. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction A large number of accounting studies have examined the occur- rence of earnings management in various contexts. There is, for example, evidence suggesting that firms manage earnings in order to avoid violating debt covenants (DeFond & Jiambalvo, 1994) or to influence initial public offering (IPO) valuations (Teoh, Welch, & Wong, 1998). A major challenge in earnings management studies is to find a measure for how much firms have managed their earn- ings. A common assumption is that earnings are managed through accounting accruals. With this assumption the measure of earnings management is the unexpected part of a firm’s total accruals. Var- ious models have been suggested for dividing the total accruals into non-discretionary (expected) and discretionary (unexpected) accruals. Most of these models are based on the earnings manage- ment detection model suggested by Jones (1991). The Jones model is a linear regression model where the level of total accruals is as- sumed to be explained by property, plant and equipment and the change in sales. In the original Jones model, the regression coeffi- cients are estimated using a firm specific time series comprising data prior to the event year. The expected level of accruals is then calculated using the estimated coefficients with event year data. A considerable problem with the Jones model is the require- ment of long time series of financial statement data. Typically, in studies where the Jones model is used the requirement is set to at least ten years of data prior to the event year (e.g. Dechow, Sloan, & Sweeney, 1995; Thomas & Zhang, 2000). This requirement might lead to several problems, such as survivorship bias and non-station- ary regression coefficients (Peasnell, Pope, & Young, 2000; Young, 1999). An alternative to estimating the firm specific regression coefficients with ordinary least squares (OLS) regression is using fuzzy linear regression (FLR), first introduced by Tanaka, Uejima, and Asai (1982). Contrary to probability theory-based OLS regres- sion, FLR is based on possibility theory and fuzzy set theory. With FLR the objective is to minimize the fuzziness of the model rather than the sum of squared residuals. One of the main advantages with FLR described in the literature is its ability to handle small data sets (Shapiro, 2004). This was corroborated empirically by Kim, Mosko- witz, and Koksalan (1996) who provided evidence that FLR outper- forms statistical linear regression in predictive capability with small data sets. Thus, it is possible that the Jones model regression could be run with shorter time series using FLR. If this is the case, an FLR-based Jones model could reduce several problems originating from the requirement of long time series. The purpose of this study is to compare the performance of the OLS-based Jones model with the performance of the FLR-based Jones model. The focus is on the time series version of the Jones model and the comparison is made using different lengths of the time series. The remainder of this study is organized as follows. The basic operating principle of the linear regression-based accrual models is covered in Section 2. In Section 3 the estimation of FLR 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.05.046 ⇑ Tel.: +358 (0)40 3521768. E-mail address: henrik.hoglund@hanken.fi Expert Systems with Applications 40 (2013) 6166–6172 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.dyndns.org/dialog/?doi=10.1016/j.eswa.2013.05.046&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.05.046 mailto:henrik.hoglund@hanken.fi http://dx.doi.org/10.1016/j.eswa.2013.05.046 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa coefficients and prediction with FLR models are covered. The re- search design is presented in Section 4 and the results from the empirical study are presented in Section 5. Section 6 concludes the study. 2. Time series-based discretionary accrual estimation models During the past 30 years a large number of studies have exam- ined the occurrence of earnings management in various contexts. Typically, the assumption in these studies has been that deviations from a normal or expected level of accruals constitute earnings management. In the early earnings management detection models the expected level of accruals was defined as the average accruals over an estimation period (Healy, 1985) or as the previous year accruals (DeAngelo, 1986). The shortcoming of these models is that they do not consider changing circumstances of the firms. For example, a normal increase in accruals as a result of an increase in sales would turn up as earnings management. Jones (1991) re- laxed the assumptions that accruals remain stationary over time by suggesting a regression approach where the level of total accru- als (TACC) is explained by the change in sales (DREV) and property, plant and equipment (PPE). The change in sales is assumed to ex- plain the current accruals, such as receivables, inventory and pay- ables, whereas property, plant and equipment should mainly explain the level of depreciation. TACCt TAt�1 ¼ a0 1 TAt�1 þ a1 DREV t TAt�1 þ a2 PPEt TAt�1 Typically, the variables in the Jones model are deflated with lagged total assets (TA) in order to reduce heteroscedasticity. The coeffi- cients a0. . .a2 in the Jones model are estimated using a sufficiently long firm specific time series. The coefficients are then used to calcu- late the expected level or the non-discretionary part of total accruals using event period data. The difference between the observed event period total accruals and the calculated non-discretionary accruals is considered as unexpected accruals or discretionary accruals equal- ing earnings management. Jones (1991) used 13 years of data for estimating the regression coefficients whereas subsequent studies have typically settled for a minimum of 10 years (e.g. Dechow et al., 1995; Thomas & Zhang, 2000). A firm specific time series this long might lead to several problems. First, issues with survivorship bias are likely to arise (Bartov, Gul, & Tsui, 2000; Peasnell et al., 2000; Young, 1999). Second, a large number of firms do not have data stretching back ten years, reducing the size of the data set. Third, it is unlikely that the regression model coefficients remain stationary over a long period of time (Peasnell et al., 2000). Finally, the self-reversing nature of accruals can lead to problems with serial correlation of residuals (Peasnell et al., 2000). To remedy some of these problems and shortcomings, DeFond and Jiambalvo (1994) suggested a cross-sectional approach where the regression coefficients are industry and year specific rather than firm specific. This approach does, however, also have its shortcomings. First, the industry membership is usually defined at a 2-digit SIC level. This might result in that some industries do not have a sufficient number of observations to run the regression, reducing the size of the data set. Second, the assumption is that the accrual generating process is similar among the firms in the same industry. There is, however, evidence suggesting that this assump- tion does not necessarily hold (Ecker, Francis, & Olsson, 2011; Höglund, 2013; Kothari, Leone, & Wasley, 2005). For example, Eck- er et al. (2011) showed that the performance of the Jones model improved when running the regression with firms matched on lagged total assets rather than on industry membership. 3. Fuzzy linear regression The concept of fuzzy linear regression (FLR) builds on the fuzzy set theory developed by Zadeh (1965) and it was first introduced by Tanaka, Uejima, and Asai (1982). The idea behind FLR is that the deviations between the observed and estimated values are assumed to originate from imprecise observations or vague relations be- tween the model variables. This is contrary to OLS regression where random errors are assumed to be the reason for differences be- tween observed and estimated values. Thus, the uncertainty in FLR models is fuzziness rather than randomness (Yang & Lee, 2002). 3.1. Estimating fuzzy linear regression coefficients In OLS regression the objective is to minimize the sum of squared residuals. In FLR, on the other hand, the objective is to minimize the fuzziness of the model. The general form of the FLR model is: bY i ¼ ~A0i þ ~A1i x1i þ���þ ~Ajixji where x1. . .xj denotes the independent variables, Ã0 the estimated fuzzy intercept coefficient, Ã1. . .Ãj the estimated fuzzy slope coeffi- cients and Ŷ i the estimated fuzzy output. The fuzzy coefficients are generally represented as symmetric triangular fuzzy numbers where aj equals the center value of Ãj and cj the spread (see Fig. 1). However, other than symmetric triangular membership functions, such as asymmetric triangular and trapezoidal (Ishibuchi & Nii, 2001), are also used. Before estimating the coefficients of the FLR, the fit between the FLR model and the data set has to be defined. This is done by set- ting the value of the h term, also called the target degree of belief (e.g. Chang & Ayyub, 2001), between 0 and 1. Each observed fuzzy ~Y i or crisp Yi output must fall within the estimated fuzzy output Ŷ i at h (see Fig. 2). As the value of h increases, the fuzziness of the FLR model also increases. The h value does, however, not affect the cen- ter value of the fuzzy coefficients (Tanaka & Watada, 1988). The FLR coefficients Ã0. . .Ãj are estimated using linear program- ming with the objective of minimizing the total spread or fuzziness of the FLR model. The following objective function, where S de- notes the total fuzziness of the regression model, m the number of independent variables and n the number of observations, is to be minimized: S ¼ Xm j¼0 Cj Xn j¼0 jxjij The FLR linear programming problem comprises two sets of con- straints. First, the spread cj of the fuzzy coefficients Ãj has to be zero or positive. Second, all observed output variables ~Y i must fall within the estimated fuzzy output variables Ŷ i at level h. Yi is the center αj cj m em b er sh ip 1.0 0.0 Ãj cj Fig. 1. Fuzzy coefficient with center value and spread. H. Höglund / Expert Systems with Applications 40 (2013) 6166–6172 6167 https://isiarticles.com/article/24668 
work_odu3gkspafa3xjn5zpcmla4sqi	----	  Open Peer Review Any reports and responses or comments on the article can be found at the end of the article. SOFTWARE TOOL ARTICLE    ESforRPD2: Expert System for Rice Plant Disease  Diagnosis [version 2; referees: 1 approved] Fahrul Agus ,     Muh. Ihsan , Dyna Marisa Khairina , Krishna Purnawan Candra 2 GIS and Environment Modelling Lab. CSIT, Mulawarman University, Samarinda, East Kalimantan, 75242, Indonesia Department of Agricultural Product Technology, Faculty of Agriculture, Mulawarman University, Samarinda, 75123, Indonesia Abstract One of the factors causing rice production disturbance in Indonesia is that farmers lack knowledge of early symptoms of rice plant diseases. These diseases are increasingly rampant because of the lack of experts. This study aimed to overcome this problem by providing an Expert System that helps farmers to make an early diagnosis of rice plant diseases. Data of rice plant pests and diseases in 2016 were taken from Samarinda, East Kalimantan, Indonesia using an in-depth survey, and rice experts from the Department of Food Crops and Horticulture of East Kalimantan Province were recruited for the project. The Expert System for Rice Plant Disease Diagnosis, ESforRPD2, was developed based on the pest and disease experiences of the rice experts and uses a Waterfall Paradigm and Unified Modeling Language. This Expert System can detect 48 symptoms and 8 types of diseases of rice plants from 16 data tests with a sensitivity of 87.5%. ESforRPD2 is available in Indonesian at http://esforrpd2.blog.unmul.ac.id Keywords Expert System, Rice Plant Disease, Waterfall, Unified Modelling Language   This article is included in the ICTROPS 2018 collection. 1 1 1 2 1 2  Referee Status:   Invited Referees      version 2 published 21 Feb 2019 version 1 published 06 Dec 2018 1 report report , University of Georgia, USAYi Fang1  06 Dec 2018,  :1902 (First published: 7 )https://doi.org/10.12688/f1000research.16657.1  21 Feb 2019,  :1902 (Latest published: 7 )https://doi.org/10.12688/f1000research.16657.2 v2 Page 1 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 https://f1000research.com/articles/7-1902/v2 https://f1000research.com/articles/7-1902/v2 https://orcid.org/0000-0003-2983-137X https://orcid.org/0000-0003-1358-1329 http://esforrpd2.blog.unmul.ac.id https://f1000research.com/collections/ICTROPS2018 https://f1000research.com/collections/ICTROPS2018 https://f1000research.com/articles/7-1902/v2 https://f1000research.com/articles/7-1902/v1 https://orcid.org/0000-0003-2583-328X https://doi.org/10.12688/f1000research.16657.1 https://doi.org/10.12688/f1000research.16657.2 http://crossmark.crossref.org/dialog/?doi=10.12688/f1000research.16657.2&domain=pdf&date_stamp=2019-02-21    Fahrul Agus ( )Corresponding author: fahrulagus@unmul.ac.id   : Conceptualization, Funding Acquisition, Supervision, Writing – Original Draft Preparation, Writing – Review & Editing; Author roles: Agus F : Data Curation, Formal Analysis, Writing – Original Draft Preparation;  : Supervision;  : Writing – Review &Ihsan M Marisa Khairina D Candra KP Editing  No competing interests were disclosed.Competing interests:  The author(s) declared that no grants were involved in supporting this work.Grant information:  © 2019 Agus F  . This is an open access article distributed under the terms of the  , whichCopyright: et al Creative Commons Attribution Licence permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Data associated with the article are available under the terms of the   (CC0 1.0 Public domain dedication).Creative Commons Zero "No rights reserved" data waiver  Agus F, Ihsan M, Marisa Khairina D and Candra KP. How to cite this article: ESforRPD2: Expert System for Rice Plant Disease Diagnosis    2019,  :1902 ( )[version 2; referees: 1 approved] F1000Research 7 https://doi.org/10.12688/f1000research.16657.2  06 Dec 2018,  :1902 ( ) First published: 7 https://doi.org/10.12688/f1000research.16657.1 Page 2 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 http://creativecommons.org/licenses/by/4.0/ http://creativecommons.org/publicdomain/zero/1.0/ https://doi.org/10.12688/f1000research.16657.2 https://doi.org/10.12688/f1000research.16657.1 Amendments from Version 1 • We corrected some typos (Abstract: ‘Modelling’ change to ‘Modeling’) • We have changed the term of ‘accuracy’ to ‘sensitivity’ and added the explanation that the ES is specific to rice plant diseases diagnosis. • We corrected some typos in the abstract and Table 1. • We also agree with the comment from Yi Fang to change some sentences, i.e. o “is the lack of knowledge of farmers on early symptoms…” changed to “is that farmers lack knowledge of early symptoms…” in the Abstract o “accuracy of 87.5%” change “sensitivity of 87.5%” in the Abstract o “….education, and business, including agriculture, problems” changed to “…..education, business, and agriculture problems.” o We added ‘was’ to the sentence “Waterfall paradigm was applied in designing this ES” o We have changed the term of ‘accuracy’ to ‘sensitivity’ in the sentence of “In this case test, the ES gave the accuracy of disease type detection of 91%” changed to “In this case test, the ES gave the sensitivity of disease type detection of 91%” See referee reports REVISED Introduction Correct diagnosis of symptoms in rice plant diseases, caused by bacteria, nematodes, fungi, phythoplasmal and viruses1–4, is very critical in supporting the productivity of rice plants. However, many regions in Indonesia have a huge problem because of a limited number of rice plant pathologists. The large plantation area of rice plants is also a problem due to logistical issues when visiting these sites, leading to difficulty obtaining disease evidence. Along with other rapid technological developments, a technology known as Expert System (ES)5–8 has been developed to solve health9–12, education13, business14, and agriculture15,16, problems. ES is usually designed for a specific condition, i.e. variables of climate in cases of agriculture. This article proposes a new software based on ES for the diagnosis of disease in rice plants in the Samarinda region, Indonesia. Waterfall Paradigm was applied in designing this ES. The prototype, Expert System for Rice Plant Disease Diagnosis (ESforRPD2) is available at: http:// esforrpd2.blog.unmul.ac.id. Methods Data collection and ES development The ES of rice plant disease diagnosis was designed to help farmers and agricultural officials to diagnose rice plant diseases occurring in the Samarinda region, East Kalimantan province, Indonesia. Rice plant experts were recruited from the Seed Tech- nology Development Division at the Department of Food Crops and Horticulture of East Kalimantan Province and from the Department of Agro-eco-technology of Agricultural Faculty of Mulawarman University (one expert from each). The experts were the primary source for information on rice plant symptoms and diseases. The two rice plant experts have experience in diagnosing rice plant disease in the region of East Kalimantan Province for 20 years. Symptoms and diseases specific for rice plant and their relationships (and their ranked impor- tance) were derived from the experts by questionnaire (Supple- mentary File 1). This allows the ES to be specific for rice plant diseases diagnosis. This information was then used to construct the knowledge base for building the ES software. The ES software was developed using the Waterfall paradigm as recommended by Sommerville17 using five stages, i.e. (i) planning and requirement, (ii) analysis and software design, (iii) imple- mentation and unit testing, (iv) integration and (v) system test- ing and operation and maintenance. ES architecture consists of three parts, namely the user interface, the inference engine and the knowledge base as proposed by Lucas and van der Gaag7. The user interface is used as a consulting interface in order to obtain knowledge and advice from the ES, which would be like consulting an expert. In this ES, the inference engine works as a consultation system in processing input data to build a diagnosis based on the knowledge base developed. Implementation The implementation of the ESforRPD2 application is based on Unified Modelling Language (Figure 1) as proposed by Sommerville17, which consists of use case diagrams, activity diagrams, and class diagrams. We constructed two types of “Use case diagram”, namely “Use case for user” consisting of four cases (Article, Consult- ing, Choose Symptoms and Consulting Result); and “Use case for expert” consisting of three cases (Symptoms, Diseases, and Relation). The use case describes the functions of the ES inter- acting with user and expert. The activity diagram illustrates the flow of various activities being designed in the ES, i.e. how the flow starts, the decision that might occur, and the flow end. The activity diagram also describes parallel processes that might occur in some executions. In this ES, we build four data stores (Expert, Symptoms, Relation, and Diagnosis) in the class diagram. The ESforRPD2 application uses four datasets, namely disease- and symptoms-data, knowledge base, and symptoms- disease-weight relationships table (Dataset 2). The construction of decision trees and forward-chaining tracing for diagnosing of rice plant diseases in the ES is shown in Figure 2. ESforRPD2 is the first version of ES (only in Indonesian) to make it user-friendly for Indonesian users. Users use a consul- tation page to choose the symptoms of the rice plant. The ES performs the calculation process to obtain the trust level using the Dempster-Shafer method18. The user page (Figure 3a) is the main web page for users without logging in. In the user page, there is also a home menu that displays articles about ES, rice plant diseases, and the Dempster-Shafer method. The consultation page starts the user consultation about the disease of rice plants (Figure 3b). The ES will provide an output as a display showing the symptoms, diagnosis of disease and the confidence level (Figure 3c). Page 3 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 http://esforrpd2.blog.unmul.ac.id/ http://esforrpd2.blog.unmul.ac.id/ Figure 1. a. Use case diagram of user. b. Use case diagram of expert. Figure 1c. Activity diagram of ESforRPD2 system application. Page 4 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 Figure 1d. Class diagram of ESforRPD2 system application. Operation The ESforRPD2 application is developed using CPU with specifications of Intel Core i3, 4GB RAM, and 300GB HDD. The same specification of CPU is needed to operate this application. Uses case The ESforRPD2 application was tested applying symptom-data inputs by clicking the symptoms selected (Figure 5b). In a single test using the case of four symptom-data inputs selected, namely (i) Spots on leaf midrib, (ii) Little spots are dark brown or slightly purple rounded shape, (iii) Spots on oval-shaped leaves and evenly distributed on the leaf surface, (iv) The size of spots is 2–10 mm long and 1 mm wide, a display of diag- nosis page (Figure 3c) will appear following clicking of the “submit diagnose” button. The diagnosis page shows the confidence level. In this case test, the ES gave the sensitivity of disease type detection of 91%. 16 tests in row were conducted using randomly selected symptoms by user in the ES. The results were approved by the two experts. In total, 14 diagnosis (87.5%) of the 16 results showed by the ES were justified by the two experts (Table 1). Discussion The ESforRPD2 application is showing good reliability. By applying 16 tests, the ESforRPD2 showed a level of performance of 87.5% (Table 1) following justification to two rice plants diseases experts. The performance of the ESforRPD2 during validation was the expected high-performance level of plant diseases diagnosis by the expert system. This performance is much higher than the performance of ES for Chili pepper pest diagnosis invented by Agus et al.16. However other Expert System could show excellent performance of 98.38%19, this evidence advice that the performance of ESforRPD2 could be improved in the next study. Currently, ESforRPD2 has only been tested with data from the Samarinda region. In a future study, we will use data from other regions of East Kalimantan, which have the same climate (tropical rainforest) and soil character as the Samarinda region. In addition, we will test data from other regions in Indonesia, Page 5 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 Figure 2. Decision tree and forward chaining tracing. G1 G2 G3 G4 G5 G6 G7 G8 G9 G10 G11 G12 G13 G14 G15 G16 G17 G18 G19 G20 G21 G22 G23 G24 G25 G26 G27 G28 G29 G30 G31 G32 G33 G34 G35 G36 G37 G38 G39 G40 G41 G42 G43 G44 G45 G46 G47 G48 P1 P2 P3 P4 P5 P6 P7 P8 Page 6 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 Figure 3a. User main page. Figure 3b. Consultation page. Page 7 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 Figure 3c. Diagnosis results page. Table 1. System testing with expert justification. Test No. Experts Justification (English/Indonesian) Results Diagnosis of ESforRPD2 (English/Indonesian) Results 1 Blast/Blas Blast/Blast Suitable 2 Brown Spot (Bercak Coklat) Brown Spot (Bercak Coklat) Suitable 3 Narrow Brown Spot (Bercak Coklat Sempit) Narrow Brown Spot (Bercak Coklat Sempit) Suitable 4 Sheath Bligh (Hawar Pelepah) Sheath Bligh (Hawar Pelepah) Suitable 5 False Smut (Noda Palsu/Gosong Palsu) False Smut (Noda Palsu/Gosong Palsu) Suitable 6 Grassy Stunt (Kerdil Rumput) Grassy Stunt (Kerdil Rumput) Suitable 7 Bacterial leaf blight (BLB-Kresek Hawar Daun) Bacterial leaf blight (BLB-Kresek Hawar Daun) Suitable 8 Tungro (Tungro) Tungro (Tungro) Suitable 9 Blast (Blas) Blast (Blas) Suitable 10 Brown Spot (Bercak Coklat) Brown Spot (Bercak Coklat) Suitable 11 Narrow Brown Spot (Bercak Coklat Sempit) Blast/Blas Unsuitable 12 Sheath Bligh (Hawar Pelepah) Blast/Blas Unsuitable 13 False Smut (Noda Palsu/Gosong Palsu) False Smut (Noda Palsu/Gosong Palsu) Suitable 14 Grassy Stunt (Kerdil Rumput) Grassy Stunt (Kerdil Rumput) Suitable 15 Bacterial leaf blight (BLB-Kresek Hawar Daun) Bacterial leaf blight (BLB-Kresek Hawar Daun) Suitable 16 Tungro (Tungro) Tungro (Tungro) Suitable Page 8 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 which have a different climate. Newbery et al.20 showed that different climate conditions affect symptoms of arable crop disease; therefore, the ESforRPD2 will need continuous evaluation because climate change effects21. Consent Written informed consent was obtained from the two experts for participation in the study. Software availability Software application is available from: http://esforrpd2.blog. unmul.ac.id. Source code: https://github.com/fahrulagus/paper. Archived source code as at time of publication: https://doi. org/10.5281/zenodo.149064122 License: GNU GPL v3.0 Data availability Underlying data Zenodo: Knowledge base for rice plant disease diagnosis, https:// doi.org/10.5281/zenodo.149065823 Extended data Zenodo: Dataset for rice plant diseases expert interview, http:// doi.org/10.5281/zenodo.195338324 Grant information The author(s) declared that no grants were involved in supporting this work. Acknowledgements The authors are grateful to both experts in this research, the Rector of Mulawarman University and Islamic Development Bank Project. References 1. Jagan Mohan K, Balasubramanian M, Palanivel S: Detection and Recognition of Diseases from Paddy Plant Leaf Images. Int J Comput Appl. 2016; 144(12): 34–41. Publisher Full Text 2. Chapuis E, Besnard G, Andrianasetra S, et al.: First report of the root-knot nematode (Meloidogyne graminicola) in Madagascar rice fields. Australas Plant Dis Notes. 2016; 11(1): 32. Publisher Full Text 3. Qudsia H, Akhter M, Riaz A, et al.: Comparative Efficacy of Different Chemical Treatments for Paddy Blast, Brown Leaf Spot and Bacterial Leaf Blight Diseases in Rice (Oryza Sativa L.). Appl Microbiol Open Access. 2017; 3(3). Publisher Full Text 4. Kulmitra AK, Sahu N, Kumar VBS, et al.: In vitro evaluation of bio-agents against Pyricularia oryzae (Cav.) causing rice blast disease. Agric Sci Dig - A Res J. 2017; 37(3): 98–101. Publisher Full Text 5. Todd BS: An Introduction to Expert Systems. Issue 95, Oxford University Computing Laboratory, Programming Research Group; 1992. Reference Source 6. Turban E, Frenzel LE: Expert Systems and Applied Artificial Intelligence. Prentice Hall Professional Technical Reference; 1992. Reference Source 7. Lucas PJF, van der Gaag LC: Principles of Expert Systems. Amsterdam: Addison-Wesley; 1991. Reference Source 8. Gaines BR: Designing Expert Systems for Usability. Knowl Creat Diffus Util. 1986; 1–40. Reference Source 9. Gudu J, Gichoya D, Nyongesa P, et al.: Development of a medical expert system as an expert knowledge sharing tool on diagnosis and treatment of hypertension in pregnancy. Int J Biosci Biochem Bioinforma. 2012; 2(5): 297–300. Publisher Full Text 10. Abu Naser SS, Ola AZA: An expert system for diagnosing eye diseases using clips. Theor Appl Inf Technol. 2008; 923–930. Reference Source 11. Ayangbekun OJ, Jimoh IA: Expert System for Diagnosis Neurodegenerative Diseases. Int J Comput Inf Technol. 2015; 4(4): 694–698. Reference Source 12. Ayangbekun OJ, Bankole FO: An Expert System for Diagnosis of Blood Disorder. Int J Comput Appl. 2014; 100(7): 975–8887. Reference Source 13. Divayana DGH: Utilization of cse-ucla model in evaluating of digital library program based on expert system at universitas teknologi indonesia: A model for evaluating of information technology-based education services. J Theor Appl Inf Technol. 2017; 95(15): 3585–3596. Reference Source 14. Arias-Aranda D, Castro JL, Navarro M, et al.: A fuzzy expert system for business management. Expert Syst Appl. 2010; 37(12): 7570–7580. Publisher Full Text 15. Ihsan M, Agus F, Khairina DM: Penerapan Metode Dempster Shafer Untuk Sistem Deteksi Penyakit Tanaman Padi. Prosiding Seminar Ilmu Komputer dan Teknologi Informasi. 2017; 2(1). Reference Source 16. Agus F, Wulandari HE, Astuti IF: Expert System With Certainty Factor For Early Diagnosis Of Red Chili Peppers Diseases. JAIS. 2017; 2(2): 52–66. Reference Source 17. Sommerville I: Software Engineering. 10th Edition. Pearson; 2016. Reference Source 18. Maseleno A, Mahmud Hasan M: Skin infection detection using Dempster-Shafer theory. In: 2012 International Conference on Informatics, Electronics and Vision, ICIEV 2012. 2012. Publisher Full Text 19. Sabzi S, Abbaspour-Gilandeh Y, García-Mateos G: A fast and accurate expert system for weed identification in potato crops using metaheuristic algorithms. Comput Ind. 2018; 98: 80–89. Publisher Full Text 20. Newbery F, Qi A, Fitt BD: Modelling impacts of climate change on arable crop diseases: progress, challenges and applications. Curr Opin Plant Biol. 2016; 32: 101–109. PubMed Abstract | Publisher Full Text 21. Xu C, Wu W, Ge Q: Impact assessment of climate change on rice yields using the ORYZA model in the Sichuan Basin, China. Int J Clim. 2018; 38(7): 2922–2939. Publisher Full Text 22. fahrulagus, Fajar MI: fahrulagus/paper: ESforRPD2 (Version V.10). Zenodo. 2018. http://www.doi.org/10.5281/zenodo.1490641 23. Agus F, Ihsan M, Khairina DM, et al.: Knowledge base for rice plant disease diagnosis [Data set]. Zenodo. 2018. http://www.doi.org/10.5281/zenodo.1490658 24. Agus F, Ihsan M, Khairina DM, et al.: Dataset for rice plant diseases expert interview [Data set]. Zenodo. F1000Research. 2018. http://www.doi.org/10.5281/zenodo.1953383 Page 9 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 http://esforrpd2.blog.unmul.ac.id/ http://esforrpd2.blog.unmul.ac.id/ https://github.com/fahrulagus/paper https://dx.doi.org/10.5281/zenodo.1490641 https://dx.doi.org/10.5281/zenodo.1490641 https://dx.doi.org/10.5281/zenodo.1490658 https://dx.doi.org/10.5281/zenodo.1490658 http://dx.doi.org/10.5281/zenodo.1953383 http://dx.doi.org/10.5281/zenodo.1953383 http://dx.doi.org/10.5120/ijca2016910505 http://dx.doi.org/10.1007/s13314-016-0222-5 http://dx.doi.org/10.4172/2471-9315.1000138 http://dx.doi.org/10.18805/asd.v37i03.8989 https://www.cs.ox.ac.uk/files/3425/PRG95.pdf https://dl.acm.org/citation.cfm?id=573712&preflayout=flat https://www.cs.ru.nl/P.Lucas/proe.pdf https://pages.cpsc.ucalgary.ca/~gaines/reports/KBS/HFES/HFES.pdf http://dx.doi.org/10.7763/IJBBB.2012.V2.120 http://www.jatit.org/volumes/research-papers/Vol4No10/5Vol4No10.pdf https://www.ijcit.com/archives/volume4/issue4/Paper040413.pdf https://www.academia.edu/27016549/An_Expert_System_for_Diagnosis_of_Blood_Disorder http://www.jatit.org/volumes/Vol95No15/17Vol95No15.pdf http://dx.doi.org/10.1016/j.eswa.2010.04.086 http://e-journals.unmul.ac.id/index.php/SAKTI/article/download/249/pdf https://publikasi.dinus.ac.id/index.php/jais/article/download/1455/1182 https://books.google.co.in/books?id=tW4VngEACAAJ&dq=Software+Engineering,+10th+Edition+PDF&hl=en&sa=X&ved=0ahUKEwij6dX544LfAhVJK48KHUvfD9sQ6AEILjAB http://dx.doi.org/10.1109/ICIEV.2012.6317330 http://dx.doi.org/10.1016/J.COMPIND.2018.03.001 http://www.ncbi.nlm.nih.gov/pubmed/27471781 http://dx.doi.org/10.1016/J.PBI.2016.07.002 http://dx.doi.org/10.1002/joc.5473 http://www.doi.org/10.5281/zenodo.1490641 http://www.doi.org/10.5281/zenodo.1490658 http://www.doi.org/10.5281/zenodo.1953383   Open Peer Review Current Referee Status: Version 2 01 March 2019Referee Report https://doi.org/10.5256/f1000research.19990.r44749    Yi Fang Nano Electrochemistry Laboratory, College of Engineering, University of Georgia, Athens, GA, USA After the language revision and then using the correct terminologies, I agree that the manuscript should be indexed.  No competing interests were disclosed.Competing Interests: Reviewer Expertise: electrochemistry, plant diseases, biosensors, sensor I have read this submission. I believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard. Version 1 24 January 2019Referee Report https://doi.org/10.5256/f1000research.18205.r43488    Yi Fang Nano Electrochemistry Laboratory, College of Engineering, University of Georgia, Athens, GA, USA The author applied Expert System (ES) for rice plant disease and diagnosis; the background information and introduction is sufficient and well organized. The entire manuscript is also presented well. However, the author used the word “accuracy” which is not quite a scientific term. If “accuracy” is defined as sensitivity of the method, how about the specificity of the method? If a method has low specificity, it may not be able to solve the problem from false positive. For other comments, please see below: Page 1: Change “is the lack of knowledge of farmers on early symptoms…” to “is that farmers lack of knowledge of early symptoms…”.   Page 1: “accuracy of 87.5%” accuracy is not a scientific terminology, do you refer to sensitivity or specificity?   Page 3: change “… education, and business, including agriculture, problems” to “…education, Page 10 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 https://doi.org/10.5256/f1000research.19990.r44749 http://orcid.org/0000-0003-2583-328X https://doi.org/10.5256/f1000research.18205.r43488 http://orcid.org/0000-0003-2583-328X   Page 3: change “… education, and business, including agriculture, problems” to “…education, business, and agriculture problems.”   Page 3: please change to “Waterfall Paradigm was applied in designing this ES.”   Page 5: “In this case test, the ES gave the accuracy of disease type detection of 91%”. Do you refer to sensitivity? Please also try to apply this comment to the other “accuracy” you mentioned in the manuscript. Is the rationale for developing the new software tool clearly explained? Partly Is the description of the software tool technically sound? Yes Are sufficient details of the code, methods and analysis (if applicable) provided to allow replication of the software development and its use by others? Yes Is sufficient information provided to allow interpretation of the expected output datasets and any results generated using the tool? Yes Are the conclusions about the tool and its performance adequately supported by the findings presented in the article? Yes  No competing interests were disclosed.Competing Interests: Reviewer Expertise: electrochemistry, plant diseases, biosensors, sensor I have read this submission. I believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above. Author Response 12 Feb 2019 , Mulawarman University, IndonesiaFahrul Agus We agree with your judgment regarding the term of accuracy. We meant the accuracy is the sensitivity, for that reason we change the term accuracy to sensitivity. Regarding the term of specificity, we explain that this system has high specificity for rice plants because all data used in constructing the algorithm were collected specifically for rice plant diseases.   No competing interests were disclosed.Competing Interests: Page 11 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019   The benefits of publishing with F1000Research: Your article is published within days, with no editorial bias You can publish traditional articles, null/negative results, case reports, data notes and more The peer review process is transparent and collaborative Your article is indexed in PubMed after passing peer review Dedicated customer support at every stage For pre-submission enquiries, contact   research@f1000.com Page 12 of 12 F1000Research 2019, 7:1902 Last updated: 01 MAR 2019 
work_oeutam4z2jd45axuyv63jaknh4	----	Human action recognition with sparse classification and multiple-view learning This document is published in: Expert Systems (2014). 31(4), 354-364. DOI: http://dx.doi.org/10.1111/exsy.12040 © 2013 Wiley Publishing Ltd. http://e-archivo.uc3m.es/ http://dx.doi.org/10.1111/exsy.12040 Human action recognition with sparse classification and multiple-view learning Rodrigo Cilla, Miguel A. Patricio, Antonio Berlanga and José M. Molina Applied Artificial Intelligence Group, Universidad Carlos III de Madrid, Madrid, Spain E-mail: rodri.cilla@gmail.com;rcilla@inf.uc3m.es Abstract: Employing multiple camera viewpoints in the recognition of human actions increases performance. This paper presents a feature fusion approach to efficiently combine 2D observations extracted from different camera viewpoints. Multiple-view dimensionality reduction is employed to learn a common parameterization of 2D action descriptors computed for each one of the available viewpoints. Canonical correlation analysis and their variants are employed to obtain such parameterizations. A sparse sequence classifier based on L1 regularization is proposed to avoid the problem of having to choose the proper number of dimensions of the common parameterization. The proposed system is employed in the classification of the Inria Xmas Motion Acquisition Sequences (IXMAS) data set with successful results. Keywords: Human Action Recognition, Multiple View Learning, L1 regularization 1. Introduction The recognition of human actions has received increasing attention by the computer vision community during the last two decades (Lavee et al., 2009; Weinland et al., 2011). The objective is to build computational models to study how humans move and behave. A wide range of applications such as ambient intelligence (Chaaraoui et al., 2012), video surveillance (Foresti et al., 2004) or human–computer interaction (Ren & Xu, 2002) have benefitted from enhancements made in the field. Whereas the first systems built for human action recognition were limited to a single camera (Cedras & Shah, 1995), advances made in visual sensor networks have motivated the inclusion of multiple cameras to enhance recognition robustness (Cilla et al., 2012). This way, wider scenes can be covered, and systems can deal with occlusions caused by walls and furniture that complicate the recognition from a single camera view. Additionally, the perception of the motion from different viewpoints provides supplementary information that facilitates the recognition. A current trend in human action recognition is the efficient combination of observations grabbed from different camera viewpoints. Most of the existing approaches are based on recovering a 3D model of the target of study exploiting multiple-view geometry. Visual hulls (Weinland et al., 2006) or 3D skeletons (Parameswaran & Chellappa, 2006) are examples of these models. However, they have some drawbacks: (1) the need of accurate camera calibration parameters to recover the 3D model and (2) large network bandwidth requirements, as many raw data should be sent to a central node where the reconstruction is performed. All these facts make current 3D reconstruction methods not appropriate for their implementation in visual sensor networks. Experimental evidence has shown that common 2D high- dimensional cues employed in human action recognition, such as silhouettes or optical flow, might be parameterized in low-dimensional manifolds where most of their variance is preserved (Blackburn & Ribeiro, 2007; Wang & Suter, 2008). At the same time, 3D descriptors recovered from multiple camera viewpoints, such as visual hulls or skeletons, are also parameterizable in low-dimensional manifolds (Turaga et al., 2008; Peng et al., 2009b). Both kinds of models contain information about the same real- world phenomena but at different depths, and both share the property of being parameterizable in low-dimensional manifolds preserving most of their variance. A question arising from this fact is if it would be possible to recover a low-dimensional manifold parameterization from multiple 2D descriptors with a similar performance to that recovered from 3D descriptors. A design decision when learning low-dimensional manifold parameterizations is the number of dimensions that the learned low-dimensional space should have. There are different proposals to automatically adjust it. Variance preservance ratios (Jolliffe & MyiLibrary, 2002) or automatic relevance determination priors (Nounou et al., 2002) are some popular choices. They adjust manifold dimensionality to preserve the maximum variance employing a reasonable number of parameters. However, these methods do not ensure a good performance when employing the obtained low-dimensional parameterization in a subsequent task, such as action class prediction. In practice, experimental cross-validation is employed to select the right number of dimensions. It has a high computational cost, as many set- ups have to be evaluated to find the best one. An alternative is to select the right number of dimensions in the subsequent task incorporating feature selection to the model with a sparse prior (Tibshirani, 1996). This paper presents a multi-camera human action recognition system taking into account the considerations presented earlier. A feature fusion algorithm is proposed to learn a common manifold representation of feature 1 descriptors extracted from each camera viewpoint. The manifold is learned with a large number of dimensions. A sparse sequence classifier is proposed to select relevant features from the manifold parameterization. This way, the problem of choosing the right number of dimensions for the manifold vanishes. 1.1. Purpose and contributions The purpose and contributions of this paper are summarized as follows: • The aim of the paper is to provide a method to classify human actions using descriptors computed from different camera views. • A feature fusion approach is proposed to combine action descriptors extracted from each camera view. Canonical correlation analysis (CCA) and kernel CCA (KCCA) are employed to obtain a common manifold param- eterization of the motion descriptors. • To avoid the problem of selecting the manifold dimen- sionality, a sparse hidden conditional random field (SHCRF) is introduced for action sequence classification. Sparsification is achieved by introducing an L1 penalty to the objective function optimized during training. • An efficient online learning method is employed to train SHCRFs, reducing training time. • The proposed system is evaluated in the recognition of IXMAS data set. Experimental evidence shows that the proposed method has a performance similar to state-of- the-art proposals employing 3D models. 1.2. Paper organization This paper is organized as follows: Section 2 reviews relevant work on human action recognition with a special focus on methods employing multiple camera views. Section 3 introduces feature fusion algorithms to recover common manifold parameterizations for the action descriptors extracted from each one of the camera views. Section 4 introduces an SHCRF model employed to predict human actions while selecting relevant features from the manifold parameterization. Experimental results are presented and discussed in Section 5. Finally, Section 6 summarizes the contributions of this work. 2. Related work 2.1. Human action recognition Multiple works have been developed to bridge the semantic gap from pixel intensity values in image sequences to descriptions of human actions performed in them. Each work defines different steps to solve the correspondence, but, in general, the process might be split in two steps: (1) feature extraction and representation and (2) action class prediction. The former step deals with the extraction and efficient encoding of features to describe motions of interest. Multiple features might be extracted for motion modelling. Parametric models have been fitted to the targets, and temporal moments of the recovered parameter values have been employed for recognition (Ribeiro & Santos-Victor, 2005). Recovering model parameters is a difficult task, so these approaches have been deprecated in favour of visual features. Appearance descriptors, such as silhouettes or skeletons, describe how the target looks like (Bobick & Davis, 2001). Local motion descriptors, such as optical flow, describe the apparent motion of the pixels, providing a very strong cue for action recognition (Efros et al., 2003). The main disadvantage of these methods is their lack of robustness towards partial occlusions of the target. Local feature descriptors have been proposed to overcome this limitation, encoding temporal and spatial variations in the neighbourhood of pixels with spatio-temporal saliency properties (Laptev, 2005). The latter step deals with the transformation of features into semantic descriptions. Sequence models capture temporal correlations among feature values and employ them to select the appropriate action label. It is possible to apply exemplar-based models with different distances measures to select action labels (Efros et al., 2003; Blackburn & Ribeiro, 2007; Wang & Suter, 2008), but most of the systems employ probabilistic graphical models. The generative Hidden Markov Model (HMM) is the de facto algorithm for human action recognition (Rabiner, 1989; Piccardi & Perez, 2007). However, discriminative graphical models have shown a better performance in action class prediction. The HCRF (Quattoni et al., 2007) has outperformed HMMs in multiple action recognition tasks. However, there are many open problems related to the usage of HCRF. This work deals with model and feature selection for the HCRF. 2.2. Multi-camera approaches to human action recognition Dasarathy’s (1997) input–output data fusion model provides a categorization framework for data fusion systems according to the level of abstraction where system inputs and outputs are defined (data, features and decisions). Here, it is employed to organize relevant works to human action recognition from multiple camera viewpoints. Data-based levels of the framework are not considered, as the information employed in human action recognition is only defined at the feature and decision levels. Diverse methods have been defined at the feature-in feature-out data fusion level to perform human action recognition from multiple camera viewpoints. It is possible to divide the works at this level into three different categories: (1) projection of 2D features to 3D; (2) feature fusion in a subspace; and (3) selection of the best view. Different 3D representations might be obtained from projecting 2D features to 3D. A popular approach is to project 2D silhouettes to 3D to obtain a visual hull representing 3D appearance (Gkalelis et al., 2009; Peng et al., 2009a; Pehlivan & Duygulu, 2011). Visual hull reconstruction requires accurate silhouette segmentation from the different views. Recent works have proposed to project optical flow to 3D (Holte & Chakraborty, 2012) or to project local interest points (Holte et al., 2011). Other works recover 3D star skeletons from 2D skeleton feature 2 correspondences (Chen et al., 2008). Action sketch correspondence across multiple views has been also proposed (Yan et al., 2008). The main drawback of 3D projection approaches is the need of accurate camera calibration parameters. Other methods compute 2D features for each view and combine them by employing some simple scheme. Averaging of multiple features representing pose, global and local motion has been proposed, improving accuracy when compared with other alternatives (Maatta et al., 2010). A joint bag-of-words histogram might be built with local feature descriptors extracted from each one of the camera views (Wu et al., 2010), but a better performance is reported when other fusion strategies are employed. Projections maximizing cross-covariance have been learned to combine R -transform derivatives extracted from each camera view (Karthikeyan et al., 2011). Two-level linear discriminant analysis has been employed to learn silhouette projections maximizing action class separability (Iosifidis et al., 2012). All these methods provide more flexible solutions for the combination of the features obtained from multiple cameras. However, experimental evidence shows a performance lower than that reported by 3D projection methods. The last class of methods is based on the computation of a quality measure for each camera view, to perform the recognition employing only the data from the best view. An estimation of the orientation of human with respect to the camera (Shen et al., 2007) or the different properties of the silhouette (Määttä & Aghajan, 2010) have been employed to obtain a quality measure. It has been proposed to select the camera with the highest number of detections (Wu et al., 2010) when methods based on interest point location are employed. There are different quality measures for studying saliency, concavity or variations in silhouette stacks (Rudoy & Zelnik-Manor, 2011). The main drawback of these approaches is that they do not exploit comple- mentary and redundant information obtained from multiple camera views. Methods defined at the feature-in decision-out level encode existing correlations between feature descriptors extracted from each camera view and action labels. The concatenation of input features is the most straightforward procedure to perform data fusion in this way (Määttä & Aghajan, 2010; Wu et al., 2010). The fused HMM (Wang et al., 2007) proposes modelling correlations among observations coupling the values of the hidden state chains of parallel HMMs defined for each camera view. Histo- grams of local features are fused, rotating the ordering of the inputs to account for the variations in orientation (Srivastava et al., 2009a). The main drawback of these works is their lack of flexibility, assuming that camera configuration remains unchanged between the train and test steps. A procedure to align camera views when the configuration changes from the train to test steps is defined in Ramagiri et al. (2011), but it requires to know relative camera placement. Decision-in decision-out is the highest level of abstraction. Action prediction is performed for each camera view to later combine the results obtained. Majority voting is the most common method for the fusion of decisions (Määttä & Aghajan, 2010). A weighted voting strategy is proposed in Zhu et al. (2012), correcting each vote according to the value of the observed feature. Errors produced at each camera view might be incorporated into the voting procedure as proposed by Cilla et al. (2012). 3. Feature fusion for human action recognition This section presents feature extraction and feature fusion methods employed in the proposed systems. The 2D human motion descriptor employed is first introduced. Then, CCA, a multiple-view dimensionality reduction method employed to find a common manifold param- eterization of the motion descriptors, is described. This section finishes with a presentation of KCCA, a non-linear extension to CCA. 3.1. Feature extraction Section 2 has shown that it is possible to employ different visual features in the recognition of human actions. This work employs the action descriptor proposed by Tran et al. (2008). It combines motion and appearance infor- mation. It has been selected because it has shown a high experimental performance in predicting the data set that will be employed for system evaluation. It extracted normalizing the bounding box of the observed human to a square box preserving aspect ratio. Shape and optical flow are computed from the normalized box. Vertical and horizontal planes of the optical flow are split and blurred with a median filter. Thus, each box has three channels: silhouette, vertical flow and horizontal flow. The box is divided into four tiles, and a radial 18-bin histogram is computed from each tile and each channel. The obtained histograms are concatenated to obtain a 216-d vector. A Principal Component Analysis (PCA) reduction of the surrounding past, present and future vectors is appended to generate a descriptor of dTRAN = 286 dimensions. Readers are referred to Tran et al. (2008) for more details. 3.2. Canonical correlation analysis The objective of CCA (Hardoon et al., 2004) is to find a pair of linear projections maximizing the correlation in the projected space between a pair of multivariate random variables. Given the zero-mean random variables in the input space x1 and x2 with dimensions d1 and d2, CCA finds a pair of linear transformations w1, w2 such that one component within each set of transformed variables is correlated with a single component in the other set. The correlation between the corresponding components is called canonical correlation, and there can be at most d = min(d1, d2) canonical correlations. The first canonical correlation is defined as r ¼ max w1;w2 wT1 x1�wT2 x2 � �ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi k wT1 x1k2 � � k wT2 x2k2 � �q (1) ¼ max w1;w2 wT1 x1x T 2 � � w2ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi wT1 x1x T 1 � � w1w T 2 x2x T 2 � � w2 q (2) 3 where x1x T 1 � � , x2x T 2 � � and x1x T 2 � � are estimated as eΣ11, eΣ22 andeΣ12 , respectively, that is, the different minors of the empirical covariance matrix eΣ ¼ eΣ11 eΣ12eΣ21 eΣ22 ! of a set of training data x = (x1, x2). The remaining canonical correlation directions are orthogonal to w1 and w2, respectively. They are solutions of the generalized eigenvalue problem: eΣ11 eΣ12eΣ21 eΣ22 ! w1 w2 � � ¼ 1 þ rð Þ eΣ11 0 0 eΣ22 ! w1 w2 � � The standard CCA model is defined for only two random variables x1 and x2. Bach and Jordan (2003) generalize CCA to m random variables. The generalized eigenvalue problem to solve is defined as follows: eΣ11 ⋯ eΣ1m ⋮ ⋮eΣm1 ⋯ eΣmm 0 B@ 1 CA w1⋮ wm 0 B@ 1 CA ¼ l eΣ11 ⋯ 0 ⋮ ⋮ 0 ⋯ eΣmm 0 B@ 1 CA w1⋮ wm 0 B@ 1 CA where eΣ11 ⋯ eΣ1m ⋮ ⋮eΣm1 ⋯ eΣmm 0 B@ 1 CA denotes the empirical covariance matrix of a set of training data x = (x1, . . ., xm) 3.3. Kernel canonical correlation analysis The main limitation of the CCA model is given by the linearity of the projections obtained. Kernel methods (Burges, 1999) provide a procedure to transform linear algorithms based on inner products of the input data to non-linear algorithms, mapping the input data x to a high- dimensional feature space f(x): f : x ¼ x1; . . . ; xnð Þ ! f xð Þ ¼ f1 xð Þ; . . . ; fN xð Þð Þ n < Nð Þ The linear algorithm (CCA in this case) is applied in the transformed feature space. The mapping from the input to feature spaces is not explicitly made. Instead, inner products performed by the linear algorithm in the input space are replaced by inner products in the feature space. Inner products in the feature space are computed by means of kernel functions. A kernel is a function K such that for all x, z 2 X, K x; zð Þ ¼ < f xð Þ�f zð Þ > (3) In CCA, data are introduced into the algorithm through the empirical covariance matrix eΣ . If the input data X = [X1X2] is centred, the minors of eΣ are computed as follows: eΣ11 ¼ XT1 X1 (4) eΣ12 ¼ XT1 X2 (5) Projection directions wx, wy can be replaced as the projection of the data into directions a1 and a2: w1 ¼ XT1 a1 (6) w2 ¼ XT2 a2 (7) Equation (2) can then be rewritten as r ¼ max a1; a2 aT1 X1X T 1 X2X T 2 a2ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi aT1 X1X T 1 X1X Ta1 � aT2 X2XT2 X2XT2 a2 q (8) Let K1 ¼ X1XT1 and K2 ¼ X2XT2 be the Gram matrices computed from the input data. Substituting into equation (8), r ¼ max a1; a2 aT1 K1K2a2ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi aT1 K 2 1a1 � aT2 K22a2 q (9) With this transform, the input data are now introduced into the algorithm through the Gram matrices K1 and K2 instead of the empirical covariance matrix eΣ. If the Gram matrices are obtained using a non-linear kernel function, the CCA algorithm now is non-linear. The generalized eigenproblem to solve in order to obtain the projection directions is 0 K1K2 K2K1 0 � � a1 a2 � � ¼ r K 2 1 0 0 K22 ! a1 a2 � � The eigenproblem in equation (10) has a trivial solution where all the values of r equal to one. That solution is not useful, and it can be avoided with a regularized version of the problem (Bach & Jordan, 2003): 0 K1K2 K2K1 0 � � a1 a2 � � ¼ r K1 þ Nk 2 I � �2 0 0 K2 þ Nk 2 I � �2 0 BBB@ 1 CCCA a1a2 � � where N is the number of samples in the training set, k is a small regularization constant and I is the N � N identity matrix. Finally, the generalization of KCCA to m input variables is given by 4 where Km corresponds to the Gram matrix generated from the samples of the input variable m. The kernel function employed in this paper is the Gaussian radial basis kernel: K x; zð Þ ¼ e� 1s2 x � zk k (10) where s is a parameter controlling the bandwidth of the Gaussian. 4. Sparse sequence classification This section presents the sequence classification algorithm employed to test the performance of the feature fusion method introduced in the previous section. A sparse version of the HCRF model is developed to select relevant features from the result of the feature fusion algorithm. The HCRF model is introduced in the first term, to later present the proposed sparse extension, and the optimization algorithm employed to recover optimal model parameters. 4.1. Hidden conditional random fields The HCRF (Quattoni et al., 2007) extends the CRF (Lafferty et al., 2001) by introducing hidden state variables into the model. An HCRF is an undirected graphical model composed of three different sets of nodes, as shown in Figure 1. The node y represents the class label for an input sequence. X = x1, . . ., xt is the set of nodes corresponding to the temporal observations in the input sequence. H = h1, . . ., ht is the set of hidden variables modelling the relationship between the observations xi and the class label y and the temporal evolution of the sequence. The conditional probability of a sequence label y and a set of hidden part assignments h given a sequence of observations X is defined using the Hammersley–Clifford theorem of Markov random fields: P y; hjx; yð Þ ¼ e Ψ y;h;x;yð ÞX y 0 X h eΨ y;h;x;yð Þ (11) where y is the vector of model parameters. HCRFs belong to the general class of log-linear models. The conditional probability of the class label y given the observation sequence X is obtained by marginalizing over all the possible value assignments to hidden parts h: Figure 1: Graphical model representation of the hidden conditional random field. K1 þ Nk 2 I � �2 K1K2 . . . K1Km K2K1 K2 þ Nk 2 I � �2 . . . K2Km ⋮ ⋮ ⋱ ⋮ KmK1 KmK2 . . . Km þ Nk 2 I � �2 0 BBBBBBBBBBBB@ 1 CCCCCCCCCCCCA a1 a2 ⋮ am 0 BBBBB@ 1 CCCCCA ¼ l K1 þ Nk 2 I � �2 0 . . . 0 0 K2 þ Nk 2 I � �2 . . . 0 ⋮ ⋮ ⋱ ⋮ 0 0 . . . Km þ Nk 2 I � �2 0 BBBBBBBBBBBB@ 1 CCCCCCCCCCCCA a1 a2 ⋮ am 0 BBBBB@ 1 CCCCCA 5 P yjx; yð Þ ¼ X h eΨ y;h;x;yð ÞX y 0 X h eΨ y 0;h;x;yð Þ (12) The potential Ψ(y, h, x; y) is a linear function of the input variables: Ψ y; h; x; yð Þ ¼ X i f xið Þ � y hið Þ þ X i y y; hið Þ þ X j; kð Þ 2 E y y; hj; hk � � (13) The first term, parameterized by y(hi), measures the compatibility of the observation at instant xi with the assignment to the hidden variable hi. The second term measures the compatibility of the hidden part hi with the class label and is parameterized by y(y, hi). Finally, the third term models sequence dynamics, measuring the compatibility of adjacent hidden parts hi and hj with the class y. Values for model parameters y are estimated from training samples {xi, yi} to maximize the L2-regularized conditional likelihood function of the model: L yð Þ ¼ Xn i ¼ 1 logP yijxi; yð Þ þ 1 2s2 jjyjj22 (14) where the parameter s controls the amount of penalization induced by the L2 norm of the model parameters. Different convex optimization techniques have been proposed to find the optimal parameters y* maximizing the conditional likelihood function in equation (14). Among them, Limited Memory - Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) (Liu & Nocedal, 1989) and conjugate gradient (Nocedal & Wright, 1999) are the most popular. Inference of the posterior probability distribution in equation (12) and auxiliary probability distributions needed for the estimation of the conditional likelihood gradient in equation (14) is made, using belief propagation as proposed in Quattoni et al. (2007). 4.2. L1-regularized hidden conditional random fields The L2-regularized training of the HCRF in equation (14) produces solutions y * where all the components have a small value. This L2 regularization approach has some drawbacks from the computational learning perspective: • There is no feature selection during training. Irrelevant features at input sequences are given a non-zero weight caused by the nature of the gradient of the L2 norm, producing model overfitting. • No model selection is performed. Occam’s razor principle of machine learning stands that the best model is the one with a lower complexity best adapting to training data. A consequence of not fulfilling this requirement is overfitting. Similarly to the previous case, parameters y(y, hi) and y(y, hj, hk) corresponding to unnecessary hidden parts obtain a non-zero weight in the HCRF result. In practice, the right number of hidden parts is selected by means of cross-validation, requiring to test multiple configurations to find the best one. A possible way to overcome these limitations is to replace the L2 penalty term in equation (14) by an L1 penalty term. This way, the objective function to minimize for parameter estimation has the form: L yð Þ ¼ Xn i ¼ 1 logP yijxi; yð Þ þ 1 s jjyj 1j (15) The L1 norm causes components in y to have a zero value if they are not needed. If a zero value is given to a parameter y(hi), then the corresponding input feature is not taken into account, producing a feature selection effect. If the zero value is given to a parameter y(y,hi) or y(y,hj,hk), then model selection is performed, reducing the complexity of the model. Although the L1 regularization has been employed to estimate the optimal parameters of log-linear models such as CRFs (Lafferty et al., 2001), this is, to the best of our knowledge, the first time that it has been employed to train HCRFs. Unfortunately, the L1 norm has the property of non- smoothness at 0, and conventional convex optimization techniques are no longer valid to recover optimal models parameters. Different approaches have been proposed for the optimization of L1-regularized log-linear models. A first approach is to reparameterize the model to transform the optimization problem into a smooth one (Vail et al., 2007). Model parameters are split in a pair of vectors y = y+ � y�, such as y+ > 0 and y� > 0. Standard convex optimization techniques are then applied to find the optimal parameters y*. Unfortunately, the convergence speed of the method is very slow requiring many iterations until an optimal solution is found. An alternative to this method is to perform stochastic gradient descent (Tsuruoka et al., 2009). Stochastic gradient descent updates model parameters after presenting each sample. Although the obtained solutions are worse than those obtained using conventional methods, in practice, they are good enough to evaluate the performance of the feature fusion approaches proposed in the previous section. The algorithm performs as follows: y k þ 12 i ¼ @ki þ �k @L j; yk � � @yki (16) yk þ 1i ¼ max 0; y k þ 1 2 i � uk þ qk � 1i � �0@ 1 A if yk þ 12i < 0 min 0; y k þ 1 2 i þ uk � qk � 1i � �0@ 1 A if yk þ 12i > 0 8>>>>>>>< >>>>>>>: (17) where qki is the total L1 penalty that yi has received up to the point: 6 qki ¼ Xk t ¼1 yt þ1i � y t þ 12 i (18) �k controls the learning rate and at each iteration is decreased by an exponential decay: �k ¼ �0a k N (19) where �0 and a are constants and N is the number of training samples. 5. Experiments 5.1. Experimental set-up Algorithms presented in previous sections are going to be evaluated in the prediction of IXMAS data set (Weinland et al., 2006). It contains 11 actions performed by 10 different actors at least three times each. The actions are recorded from five different viewpoints. The actions are (1) check watch; (2) cross arms; (3) scratch head; (4) sit down; (5) get up; (6) turn around; (7) walk; (8) wave; (9) punch; (10) kick; and (11) pick up. A sample frame of the data set observed from the five viewpoints is presented in Figure 2. The protocol to evaluate system performance is the leave- one-actor-out cross-validation. The data set is divided into 10 different sets according to the actor performing each sample. The system is trained using all sets except one used for testing. The procedure is repeated until every actor has been used for testing. The CCA and KCCA are configured to provide 150 projection directions corresponding to the 150 highest canonical correlations. It should be emphasized again that a high value is given to this parameter as the SHCRF will select the relevant projections during training. KCCA is employed with a radial basis kernel of width s = 0.5. The SHCRF has been set up with |H| = 22 hidden parts. A value s = 0.2 is given to the sparsity penalty term. The training algorithm is run for 40 iterations with a batch size of 20 samples. The exponential decay of the learning rate � is configured with an initial value �0 = 0.01 and a decay rate a = 0.95. A Monte Carlo scheme is employed to evaluate the action prediction performance as the objective function employed to train the SHCRF is non-convex, producing different solutions depending on the initial value y0. The accuracy obtained for each test set is averaged over 30 different runs of the SHCRF training algorithm to obtain a real estimate of the algorithm performance. Finally, to measure the improvement produced by feature fusion in the predictive performance, a baseline model is going to be employed. The descriptors computed for cameras 1 and 2 are independently projected using PCA to retain the 150 most significant dimensions. The SHCRF is respectively evaluated using this projected sequences with the same procedure presented earlier. 5.2. Results and discussion 5.2.1. Feature fusion algorithms To visualize the effect of the feature fusion algorithms, the first three most significant features obtained by them and the PCA baseline are shown in Figures 3(a)–3(f). Projections have been obtained with the data of actors 2–10. It can be observed that fused features have a stronger class structure than features obtained by the PCA baselines, although they do not seem to be separated in any case. This is something normal in action recognition domains, as different action sequences usually share common frames, being their temporal evolution the real discriminative factor for action class prediction. Another phenomenon observed in the plots of the linear models (Figures 3(a), 3(b), 3(c) and 3(e)) is that the main cause of class variation is produced by the direction and amount of movement. It can be seen that the variation of the actions sit down, get up and pick up, involving a vertical movement, seems to have a stronger structure than the others. However, in the features from the non-linear models presented in Figures 3(d) and 3(f), that behaviour is not present. Instead, class structure is appreciated for most of the actions. It seems that non-linear algorithms model other hidden factor of variations distinct from the movement direction. 5.2.2. Action classification Table 1 shows the accuracy obtained by the SHCRF predicting class labels for the IXMAS data set. For each evaluation fold, the worst, median and best results have been extracted. Results show that predicting action classes with fused features has an accuracy higher than predicting them with only one camera, although for CCA12 the worst case is worse than the classification with PCA2. (a) Camera 1 (b) Camera 2 (c) Camera 3 (d) Camera 4 (e) Camera 5 Figure 2: The kick action in the IXMAS data set from the five available views. 7 Other interesting fact is that CCA and KCCA have almost the same accuracy when employing the data from the five camera viewpoints, whereas in the two cameras scenario, KCCA performs better than CCA. The feature fusion algorithms employed here extract common information in the data from each camera, so this saturation phenomenon seems reasonable. A non-linear algorithm extracts common information better than the linear algorithm, but as the number of data sources increases, the gap is reduced until they have a similar performance. To better understand the behaviour of the presented algorithms, Figure 4 presents the box plots of accuracies obtained evaluating each actor following a Monte Carlo scheme. It should be noted that the feature fusion accuracy does not increase in all the cases reported. The most striking is when evaluating actor 2 (Figure 4(b)). The disastrous performance of camera 1 makes the fusion perform worse in all the cases than when using only camera 2. The cause for this phenomenon is that the fusion algorithms assume that every data source has a similar quality, not weighting them according to some quality metric. The phenomenon is produced for other actors too (Figures 4(h) and 4(i)), although there is not a big difference in the results using a single camera. That might be produced because the sources have complementary information that are not well represented in the transformed space. (a) Camera 1 PCA (b) Camera 2 PCA (c) CCA Cameras 1-2 (d) KCCA Cameras 1-2 (e) CCA Cameras 1-2-3-4-5 (f ) KCCA Cameras 1-2-3-4-5 Figure 3: Projections of the first three components obtained with the different dimensionality reduction algorithms. CCA, canonical correlation analysis; KCCA, kernel CCA. Table 1: Worst, average and best case accuracy estimations of the different methods obtained using Monte Carlo leave-one-actor-out cross-validation PCA1 PCA2 CCA12 KCCA12 CCA KCCA Minimum 0.634 0.688 0.685 0.709 0.748 0.748 Median 0.754 0.778 0.799 0.814 0.850 0.850 Maximum 0.865 0.868 0.880 0.907 0.934 0.940 8 Finally, Table 2 compares the result of the presented proposal to others. The best case result performs as good as 3D proposals using visual hulls, whereas the average case improves results from decision-in decision-out. 6. Conclusions This work has proposed the usage of multiple-view learning as a feature fusion method for human action recognition from multiple cameras. It has been shown that the usage of the information shared by rich 2D motion descriptors computed from multiple camera viewpoints improves the predictive accuracy. The usage of an L1-regularized sequence classifier has avoided the manual choice of the number of dimensions of the projected space. Testing multiple configurations to find the best dimension has been avoided, saving computational time. The usefulness of the proposed systems has been shown by predicting IXMAS data set with an accuracy similar to that reported by state-of-the-art methods. References BACH, F.R. and M.I. JORDAN (2003) Kernel independent component analysis, Journal of Machine Learning Research, 3, 1–48. BLACKBURN, J. and E. RIBEIRO (2007) Human motion recognition using isomap and dynamic time warping, in Proceedings of the 2nd conference on Human motion: understanding, modeling, capture and animation, Rio de Janeiro, Brazil: Springer-Verlag, 285–298. BOBICK, A.F. and J.W. DAVIS (2001) The recognition of human movement using temporal templates, IEEE Transactions on Pattern Analysis and Machine Intelligence, 23, 257–267. BURGES, C.J.C. (1999) Advances in Kernel Methods: Support Vector Learning, The MIT Press. CEDRAS, C. and M. SHAH (1995) Motion-based recognition: a survey, Image and Vision Computing, 13, 129–155. CHAARAOUI, A.A., P. CLIMENT-PÉREZ and F. FLÓREZ-REVUELTA (2012) A review on vision techniques applied to human behaviour analysis for ambient-assisted living, Expert Systems with Applications, 39(12), 10873–10888. CHEN, D., P.C. CHOU C.B. FOOKES and, S. SRIDHARAN (2008) Multi- view human pose estimation using modified five-point skeleton model. In International Conference on Signal Processing and Communication Systems 2007, 17–19 Dec 2007, Gold Coast, Australia. CILLA, R., M.A. PATRICIO, A. BERLANGA and J.M. MOLINA (2012) A probabilistic, discriminative and distributed system for the recognition of human actions from multiple views, Neurocomputing, 75(1), 78–87. DASARATHY, B.V. (1997) Sensor fusion potential exploitation- innovative architectures and illustrative applications, Proceedings of the IEEE, 85, 24–38. EFROS, A.A., A.C. BERG, G. MORI and J. MALIK (2003) Recognizing action at a distance. IEEE International Conference on Computer Vision, 2, 726–733. FORESTI, G.L., C. MICHELONI and L. SNIDARO (2004) Event classification for automatic visual-based surveillance of parking lots, in Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference on, vol. 3, 314–317, IEEE. GKALELIS, N., H. KIM, A. HILTON, N. NIKOLAIDIS and I. PITAS (November 2009) The i3DPost multi-view and 3D human action/interaction database. 2009 Conference for Visual Media Production, 159–168. HARDOON, D.R., S. SZEDMAK and J. SHAWE-TAYLOR (2004) Canonical correlation analysis: an overview with application to learning methods, Neural Computation, 16, 2639–2664. HOLTE, M.B. B. CHAKRABORTY, J. GONZALEZ and T. B. MOESLUND (2012) A local 3D motion descriptor for multi-view human action recognition from 4D spatio-temporal interest points, Selected Topics in Signal Processing, IEEE Journal of, 6(5), 553–565. HOLTE, M.B., T.B. MOESLUND, N. NIKOLAIDIS and I. PITAS (May 2011) 3D human action recognition for multi-view camera systems. 2011 International Conference on 3D Imaging, Modeling, Processing, Visualization and Transmission, 342–349. IOSIFIDIS, A., A. TEFAS, N. NIKOLAIDIS and I. PITAS (2012) Multi- view human movement recognition based on fuzzy distances 0.4 0.6 0.8 1 0.4 0.6 0.8 1 0.4 0.6 0.8 1 0.4 0.6 0.8 1 0.4 0.6 0.8 1 0.4 0.6 0.8 1 0.4 0.6 0.8 1 0.4 0.6 0.8 1 0.4 0.6 0.8 1 0.4 0.6 0.8 1 (a) Actor 1 (b) Actor 2 (c) Actor 3 (d) Actor 4 (e) Actor 5 (f ) Actor 6 (g) Actor 7 (h) Actor 8 (i) Actor 9 (j) Actor 10 Figure 4: Results for the different evaluation actors. A is the PCA from camera 1. B is the PCA from camera 2. C is the canonical correlation analysis (CCA) fusion of cameras 1 and 2. D is the kernel CCA (KCCA) fusion of cameras 1 and 2. E is the CCA fusion of the five cameras. F is the KCCA fusion of the five cameras. Table 2: Comparison of the accuracy of our method to others Method Accuracy Type Srivastava et al. (2009b) 81.4 Decision-in Decision-out Our Average 85.00 2D feature-in 2D feature-out Weinland et al. (2006) 93.33 2D feature-in 3D feature-out Our Best 94.00 2D feature-in 2D feature-out Peng et al. (2009b) 94.59 2D feature-in 3D feature-out 9 and linear discriminant analysis, Computer Vision and Image Understanding, 116, 347–360. JOLLIFFE, I.T. (2002) Principal Component Analysis. Springer verlag. KARTHIKEYAN, S., U. GAUR, B.S. MANJUNATH and S. GRAFTON (November 2011) Probabilistic subspace-based learning of shape dynamics modes for multi-view action recognition. 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), 1282–1286. LAFFERTY, J., A. MCCALLUM and F. PEREIRA (2001) Conditional random fields: probabilistic models for segmenting and labeling sequence data, in ICML ’01 Proceedings of the Eighteenth International Conference on Machine Learning, 282–289. LAPTEV, I. (2005) On space-time interest points, International Journal of Computer Vision, 64, 107–123. LAVEE, G., E. RIVLIN and M. RUDZSKY (2009) Understanding video events: a survey of methods for automatic interpretation of semantic occurrences in video, IEEE Transactions on Systems, Man and Cybernetics - Part C: Applications and Reviews, 39, 489–504. LIU, D.C. and J. NOCEDAL (1989) On the limited memory BFGS method for large scale optimization, Mathematical Programming, 45, 503–528. MÄÄTTÄ, T. and H. AGHAJAN (2010) On efficient use of multi-view data for activity recognition, 158–165. MÄÄTTÄ, T., A. HÄRMÄ and H. AGHAJAN (2010) On efficient use of multi-view data for activity recognition, in Proceedings of the Fourth ACM/IEEE International Conference on Distributed Smart Cameras, (pp. 158–165), ACM. NOCEDAL, J. and S.J. WRIGHT (1999) Numerical Optimization, Springer Verlag. NOUNOU, M.N., B.R. BAKSHI, P.K. GOEL and X. SHEN (2002) Bayesian principal component analysis. Journal of Chemometrics, 16, 576–595. PARAMESWARAN, V. and R. CHELLAPPA (2006) View invariance for human action recognition, International Journal of Computer Vision, 66, 83–101. PEHLIVAN, S. and P. DUYGULU (2011) A new pose-based representation for recognizing actions from multiple cameras, Computer Vision and Image Understanding, 115, 140–151. PENG, B., G. QIAN and S. RAJKO (August 2009a) View-invariant full-body gesture recognition via multilinear analysis of voxel data, 2009 Third ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC), 1–8. PENG, B., G. QIAN and S. RAJKO (September 2009b) View-invariant full-body gesture recognition via multilinear analysis of voxel data, Third ACM/IEEE Conference on Distributed Smart Cameras. PICCARDI, M. and O. PEREZ (2007) Hidden Markov models with kernel density estimation of emission probabilities and their use in activity recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR’07), 1–8. QUATTONI, A., S. WANG, L.-P. MORENCY, M. COLLINS and T. DARRELL (2007) Hidden conditional random fields, Pattern Analysis and Machine Intelligence, IEEE Transactions on, 29, 1848–1852. RABINER, L.R. (1989) A tutorial on hidden Markov models and selected applications in speech recognition, Proceedings of the IEEE, 77, 257–286. RAMAGIRI, S., R. KAVI and V. KULATHUMANI (August 2011) Real- time multi-view human action recognition using a wireless camera network, 2011 Fifth ACM/IEEE International Conference on Distributed Smart Cameras, 1–6. REN, H. and G. XU (2002) Human action recognition in smart classroom, in Automatic Face and Gesture Recognition, 2002. Proceedings of the Fifth IEEE International Conference on, 417–422, IEEE. RIBEIRO, P.C. and J. SANTOS-VICTOR (2005) Human activity recognition from video: modeling, feature selection and classi- fication architecture, in International Workshop on Human Activity Recognition and Modeling (HAREM). RUDOY, D. and L. ZELNIK-MANOR (2011) Viewpoint selection for human actions, International Journal of Computer Vision, 97, 243–254. SHEN, C., C. ZHANG and S. FELS (2007) A multi-camera surveillance system that estimates quality-of-view measure- ment. 2007 IEEE International Conference on Image Proce- ssing, III–193–III–196. SRIVASTAVA, G., H. IWAKI, J. PARK and A.C. KAK (August 2009a) Distributed and lightweight multi-camera human activity classification, 2009 Third ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC), 1–8. SRIVASTAVA, C., H. IWAKI, J. PARK and A.C. KAK (September 2009b) Distributed and lightweight multi-camera human activity classification, in Third ACM/IEEE Conference on Distributed Smart Cameras, 1–8. TIBSHIRANI, R. (1996) Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B (Methodological), 58(1), 267–288. TRAN, D., A. SOROKIN and D. FORSYTH (2008) Human activity recognition with metric learning, in Proceedings of the 10th European Conference on Computer Vision: Part I, Springer Berlin Heidelberg: Springer-Verlag, 561. TSURUOKA, Y., J. TSUJII and S. ANANIADOU (2009) Stochastic gradient descent training for l1-regularized log-linear models with cumulative penalty, in Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP: Volume 1-Volume 1, Association for Computational Linguistics, 477–485. TURAGA, P., A. VEERARAGHAVAN and R. CHELLAPPA (2008) Statistical analysis on Stiefel and Grassmann manifolds with applications in computer vision, in Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on, 1–8, IEEE. VAIL, D.L., J.D. LAFFERTY and M.M. VELOSO (2007) Feature selection in conditional random fields for activity recognition, in Intelligent Robots and Systems, 2007. IROS 2007. IEEE/RSJ International Conference on, 3379–3384, IEEE. WANG, L. and D. SUTER (2008) Visual learning and recognition of sequential data manifolds with applications to human movement analysis, Computer Vision and Image Understanding, 110, 153–172. WANG, Y., K. HUANG and T. TAN (2007) Multi-view gymnastic activity recognition with fused HMM, In Computer Vision- ACCV 2007 (pp. 667–677). Berlin Heidelberg: Springer. WEINLAND, D., R. RONFARD and E. BOYER (2006) Free viewpoint action recognition using motion history volumes, Computer Vision and Image Understanding, 104, 249–257. WEINLAND, D., R. RONFARD and E. BOYER (2011) A survey of vision-based methods for action representation, segmentation and recognition, Computer Vision and Image Understanding, 115, 224–241. WU, C., A.H. KHALILI and H. AGHAJAN (2010) Multiview activity recognition in smart homes with spatio-temporal features, Proceedings of the Fourth ACM/IEEE International Conference on Distributed Smart Cameras - ICDSC’10, 142. YAN, P., S.M. KHAN and M. SHAH (June 2008) Learning 4D action feature models for arbitrary view action recognition. 2008 IEEE Conference on Computer Vision and Pattern Recognition, 1–7. ZHU, F., L. SHAO and M. LIN (2012). Multi-view action recognition using local similarity random forests and sensor fusion. Pattern Recognition Letters, 34(1), 20–24. The authors Rodrigo Cilla Rodrigo Cilla received his BS, MS and PhD degrees in computer science from the Universidad Carlos III de Madrid in 2007, 2008 and 2012, respectively. His research interest includes, among others, human action recognition, multiple target tracking, high-dimensional data analysis, data fusion and biological image processing. Miguel A. Patricio Miguel A. Patricio received his BS and MS degrees in computer science and his PhD degree in artificial intelligence from the Universidad Politécnica de Madrid in 1991, 1995 and 2002, respectively. He has held an administrative position at the Computer Science Department of Universidad Politécnica de Madrid since 10 1993. He is currently an associate professor at the Escuela Politécnica Superior of the Universidad Carlos III de Madrid and a research fellow of the Applied Artificial Intelligence Group (GIAA). He has carried out a number of research projects and consulting activities in the areas of automatic visual inspection systems, texture recognition, neural networks and industrial applications. Antonio Berlanga Antonio Berlanga received the BA degree in physics from the Universidad Autónoma de Madrid, Spain, and the PhD degree in computer engineering from the Universidad Carlos III de Madrid in 1995 and 2000, respectively. He is currently an associate professor with the Department of Computer Science, Universidad Carlos III de Madrid. His current research interests include evolutionary computation for multi-objective optimization, machine learning and data mining. José M. Molina José M. Molina is a full professor at the Universidad Carlos III de Madrid. He joined the Computer Science Department of the Universidad Carlos III de Madrid in 1993. Currently, he leads the Applied Artificial Intelligence Group (GIAA). His current research focuses on the application of soft computing techniques (neural network, evolutionary computation, fuzzy logic and multi-agent systems) to radar data processing, air traffic management and e-commerce. He is the author of more than 20 journal papers and 80 conference papers. He received his BS and PhD degrees in telecommunications engineering from the Universidad Politécnica de Madrid in 1993 and 1997, respectively. 11 
work_ogq5r3sr6vb2dkhfp4speay7dq	----	A new dynamic modeling framework for credit risk assessment Expert Systems With Applications 45 (2016) 341–351 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa A new dynamic modeling framework for credit risk assessment Maria Rocha Sousa a,∗, João Gama a,b, Elísio Brandão a a School of Economics and Management, University of Porto, Portugal b Laboratory of Artificial Intelligence and Decision Support of the Institute for Systems and Computer Engineering, Technology and Science, Portugal a r t i c l e i n f o Keywords: Credit risk modeling Credit scoring Dynamic modeling Temporal degradation Default concept drift Memory a b s t r a c t We propose a new dynamic modeling framework for credit risk assessment that extends the prevailing credit scoring models built upon historical data static settings. The driving idea mimics the principle of films, by composing the model with a sequence of snapshots, rather than a single photograph. In doing so, the dynamic modeling consists of sequential learning from the new incoming data. A key contribution is provided by the insight that different amounts of memory can be explored concurrently. Memory refers to the amount of historic data being used for estimation. This is important in the credit risk area, which often seems to undergo shocks. During a shock, limited memory is important. Other times, a larger memory has merit. An application to a real-world financial dataset of credit cards from a financial institution in Brazil illustrates our methodology, which is able to consistently outperform the static modeling schema. © 2015 Elsevier Ltd. All rights reserved. 1 m m a p i c h i t s A f a b a i i t j m h c r i a u y s a d u t m i u n m c e u n h h 0 . Introduction In banking, credit risk assessment often relies on credit scoring odels, so called PD models (Probability of Default models).1 These odels output a score that translates the probability of a given entity, private individual or a company, becoming a defaulter in a future eriod. Nowadays, PD models are at the core of the banking business, n credit decision-making, in price settlement, and to determine the ost of capital. Moreover, central banks and international regulation ave dramatically evolved to a setting where the use of these models s favored, to achieve soundness standards for credit risk valuation in he banking system. Since 2004, with the worldwide implementation of regulations is- ued by the Basel Committee on Banking Supervision within Basel II ccord, banks were encouraged to strengthen their internal models rameworks for reaching the A-IRB (Advanced Internal Rating Based) ccreditation (BCBS, 2006; BIS, 2004). To achieve this certification, anks had to demonstrate that they were capable of accurately evalu- ting their risks, complying with Basel II requirements, by using their nternal risk models’ systems, and keep their soundness. Banks own- ng A-IRB accreditation gained an advantage over the others, because hey were allowed to use lower coefficients to weight the exposure of ∗ Corresponding author. Tel.: +351967139811. E-mail addresses: 100427011@fep.up.pt, jsc@inescporto.pt (M.R. Sousa), gama@fep.up.pt (J. Gama), ebrandao@fep.up.pt (E. Brandão). 1 Other names can be used to refer to PD models, namely: credit scoring, credit risk odels, scorecards, credit scorecards, rating systems, or rating models, although some ave different meanings. c e 1 s t r ttp://dx.doi.org/10.1016/j.eswa.2015.09.055 957-4174/© 2015 Elsevier Ltd. All rights reserved. redit at risk, the risk weighted assets, and benefit from lower capital equirements. A lot of improvements have been made in the exist- ng rating frameworks, extending the use of data mining tools and rtificial intelligence. Yet, this may have been bounded by a certain nwillingness to accept less intuitive algorithms or models going be- ond standard solutions being implemented in the banking industry, ettled in-house or delivered through analytics providers. Developing and implementing a credit scoring model can be time nd resource consuming, easily ranging from 9 to 18 months, from ata extraction until deployment. Hence, it is not rare that banks use nchanged credit scoring models for several years. Bearing in mind hat models are built using a sample file frequently comprising 2 or ore years of historical data, in the best case scenario, data used n the models are shifted 3 years away from the point they will be sed. Should conditions remain unchanged, then this would not sig- ificantly affect the accuracy of the models, otherwise, their perfor- ance can greatly deteriorate over time. The recent financial crisis onfirmed that financial environment greatly fluctuates, in an un- xpected manner, posing renewed attention regarding models built pon time-frames that are by far outdated. By 2007–2008, many fi- ancial institutions were using stale credit scoring models built with istorical data of the early-decade. The degradation of stationary redit scoring models is an issue with empirical evidence in the lit- rature (Avery, Calem, & Canner, 2004; Crook, Thomas, & Hamilton, 992; Lucas, 2004; Sousa, Gama, & Gonçalves, 2013b), however re- earch is still lacking more realistic solutions. Dominant approaches rely on static learning models. However, as he economic conditions evolve in the economic cycle, either deterio- ating or improving, also varies the behavior of an individual, and his http://dx.doi.org/10.1016/j.eswa.2015.09.055 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.09.055&domain=pdf mailto:100427011@fep.up.pt mailto:jsc@inescporto.pt mailto:jgama@fep.up.pt mailto:ebrandao@fep.up.pt http://dx.doi.org/10.1016/j.eswa.2015.09.055 342 M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 p h s i i c o c 2 a a i c a u s ( p i s c d f c m t t t l e o t T l t s b A p c a l v t s o e a c c t d w t e c s c i m i m a b ability to repay his debt. Furthermore, the default evolution echoes trends of the business cycle, and related with this, regulatory move- ments, and interest rates fluctuations. In good times, banks and bor- rowers tend to be overoptimistic about the future, whilst in times of recession banks are swamped with defaulted loans, high provisions, and tighten capital buffers turn highly conservative. The former leads to more liberal credit policies and lower credit standards, the later promotes sudden credit-cuts. Hence, default needs to be regarded as time changing. Traditional systems that are one-shot, fixed memory-based, trained from fixed training sets, and static settings are not prepared to process the evolving data. And so, they are not able to continuously maintain an output model consistent with the actual state of environ- ment, or to quickly react to changes (Gama, 2010). These are some of the features of classic approaches that evidence the constraints of the existing credit scoring systems. As the processes underlying credit risk are not strictly stationary, consumers’ behavior and default can change over time in unpredictable ways. A few limitations to the ex- isting approaches, idealized in the classical supervised classification paradigm, can be traced in published literature: • The static models usually fail to adapt when the population changes. Static and predefined sample settings often lead to an incomplete examination of the dynamics influencing the prob- lem (Gama, 2010; Hand, 2006). • Certain assumptions that are implicit to the methods, often fail in real-world environments (Yang, 2007). These assumptions relate to: – Representativeness - the standard credit scoring models rely on supervised classification methods that run on 2-years-old static samples, in order to determine which individuals are likely to default in a future fixed period, 1 year for PD mod- els (Thomas, 2010; Thomas, Edelman, & Crook, 2002). Such samples are supposed to be representative of the potential borrowers consumers of the future, the through-the-door pop- ulation. They should also be sufficiently diverse to reflect dif- ferent types of repayment behavior. However, a wide range of research is conducted in samples that are not representative. – Stability and non-bias - the distribution from which the design points and the new points is the same; classes are perfectly de- fined, and definitions will not change. Not infrequently there are selective biases over time. Simple examples of this occur- rence can be observed when a bank launches a new product or promotes a brand new segment of customers. It can also occur when macroeconomics shifts abruptly from an expansion to a recession phase, or vice versa. – Misclassification costs - these methods assume that the costs of misclassification are accurately known, but in practice they are not. • The methods that are most widely used in the banking industry, logistic regression and discriminant analysis are associated with some instability with high-dimensional data and small sample size. Other limitations regard to intensive variable selection effort and incapability of efficiently handling non-linear features (Yang, 2007). • Static models are usually focused in assessing the specific risk of applicants and obligors. However, a complete picture can only be achieved by looking at the return alongside risk, which requires the use of dynamic rather than static models (Bellotti & Crook, 2013). There is a new emphasis on running predictive models with the ability of sensing themselves and learn adaptively (Gama, 2010). Ad- vances on the concepts for knowledge discovery from data streams suggest alternative perspectives to identify, understand and effi- ciently manage dynamics of behavior in consumer credit in chang- ing ubiquitous environments. In a world where the events are not reordained and little is certain, what we do in the present affects ow events unfold in unexpected ways. So far, no comprehensive et of research to deal with time changing default had much impact nto practice. In credit risk assessment, a great deal of sophistication s needed to introduce economic factors and market conditions into urrent risk-assessment systems (Thomas, 2010). The study presented in this paper is a large extension of a previ- us research that delivered the winning model within the BRICS 2013 ompetition in data mining and finance (Sousa, Gama, Brandão et al., 013a; Sousa et al., 2013b). This competition opened to academics nd practitioners, was focused on the development of a credit risk ssessment model, tilting between the robustness of a static model- ng sample and the performance degradation over time, potentially aused by market gradual changes along few years of business oper- tion. Participants were encouraged to use any modeling technique, nder a temporal degradation or concept drift perspective. In the re- earch attached to the winning model, Sousa, Gama, and Gonçalves 2013b) have proposed a two-stage model for dealing with the tem- oral degradation of credit scoring models, which produced motivat- ng results in a 1-year horizon. The winners first developed a credit coring method using a set of supervised learning methods, and then alibrated the output, based on a projection of the evolution in the efault. This adjustment considered both the evolution of the de- ault and the evolution of macroeconomic factors, echoing potential hanges in the population of the model, in the economy, or in the arket. In so doing, resulting adjusted scores translated a combina- ion of the customers’ specific risk with systemic risk. The winning eam (Sousa, Gama, & Gonçalves) concluded that the performance of he models did not significantly differ among classification models, ike logistic regression (LR), AdaBoost, and Generalized Additive Mod- ls (GAM). However, after training in several windows lengths, they bserved that the model based on the longest window has produced he best performing model over the long-run, among all competitors. his finding allowed to realize that some specifics of the credit portfo- ios and macroeconomic environments may reveal quite stable along ime. For those cases, a model built with a static learning setting may eem appropriate, if tested during stable phases. The question yet to e answered was in which conditions credit risk models degrade? nd when so, if there is any alternative modeling technique to the revailing credit scoring models? The aim of this study is to find a learer understanding on which type of modeling framework allows rapid adaptation to changes, and in which circumstances a static earning setting still delivers well-performing models. With this in iew, we implemented a dynamical modeling framework and two ypes of windows for model training, which enable testing our re- earch questions: (a) In which conditions can a dynamic modeling utperform a static model?; (b) Is the recent information more rel- vant to improve forecasting accuracy?; (c) Does older information lways improve forecasting accuracy? This paper introduces a new dynamic modeling framework for redit risk assessment, imported from the emerging techniques of oncept drift adaptation, in streaming data mining and artificial in- elligence. The proposed model is able to produce more robust pre- ictions in stable conditions, but also in the presence of changes, hile the prevailing methods cannot. This is a promissory tool both o academics and practitioners, because unlike the traditional mod- ls, it has the ability of adjusting the predictions in the presence of hanges, like inversions in the economic cycles, major crisis, or intrin- ic behavioral circumstances (e.g. divorce, unemployment and finan- ial distress). Besides the goal of enhancing the prediction of default n credit, the new modeling framework also enables developing a ore comprehensive understanding of the evolution of the credit rat- ng systems over time and anticipating unexpected events. Further- ore, we study the implications to credit risk assessment of keeping long-term memory, and forgetting older examples, which have not een done so far. M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 343 i s d s f u a fi b c t G m t s f u I a m w c i m p d 2 c fi l m a t e 2 l y t a u u m u i a a k v a a P f v i e 2 t t d c t o t d t L o S t m m i i b p A m c s m A s g w c o j o w o a a r 2 p e d t a c b w T m c o m f i c b f t t Few authors have explicitly tried a dynamic modeling framework n credit risk assessment, or connected concepts. Based on a national ample of a credit reporting agency, Avery et al. (2004) show that tra- itional modeling often fails to consider situational circumstances, uch as local economic conditions and individual trigger events, af- ecting the ability of scoring systems to accurately quantify individ- als’ credit risk. We can trace the few existing contributions in this rena over the most recent years. Sun and Li (2011) formally de- ne financial distress concept drift and build a dynamic modeling ased on instance selection. Saberi et al. (2013) worked on the con- ept of granularity for selecting the optimum size of the testing and raining groups with a sample of credit cards of a bank operating in erman. Pavlidis, Tasoulis, Adams, and Hand (2012) proposed a ethodology for the classification of credit applications with the po- ential of adapting to population drifts. This paper follows in Section 2 with a brief description of the main ettings and concepts of the supervised learning problem and score ormulation. It also presents an overview of the methods typically sed in supervised learning, and specifically in credit score modeling. n Section 3, we introduce the topic of concept drift in credit default nd some adaptation methods that can be promising for dynamic odeling credit risk. In Section 4 we present a case study, where e employ a set of these adaptation methods to a real-world finan- ial dataset. First, we characterize the database and provide some ntuition on the background of the problem. Then, we explain the ethodology of this research. Section 5 provides the fundamental ex- erimental results. Conclusions and future applications of the new ynamic modeling framework are traced in Section 6. . Settings and concepts In this work we import some of the emerging techniques in con- ept drift adaptation into credit risk assessment models. This is a eld of research that has been receiving much attention in machine earning over the last decade, as an answer for suitably shaping the odels and processes to a reality that is ever-changing over contexts nd time. The settings and definitions adopted in this paper replicate he general nomenclature surveyed by Gama, Žliobaitė, Bifet, Pech- nizkiy, and Bouchachia (2014). .1. Supervised learning problem Credit risk assessment can be addressed as a classification prob- em, a subset of supervised learning. The aim is to predict the default ∈ {good, bad}, given a set of input characteristics x. The term at- ribute refers to each of the possible values that a characteristic can ssume; the term bin denotes a set of attributes or an interval of val- es in a continuous characteristics; the term example, or record, is sed to refer to one pair of (x, y). Supervised learning classification ethods try to determine a function that best separates the individ- als in each of the classes, good and bad, in the space of the problem. The model building is carried on a set of training examples - train- ng set – collected from the past history of credit, for which both x nd y are known. The best separation function can be achieved with classification method. These methods include, among others, well- nown classification algorithms such as decision trees (DT), support ector machines (SVM), artificial neural networks (ANN), and Gener- lized Additive Models (GAM). Hands-on software packages are avail- ble to the user for example in R, SAS, Matlab, and Model Builder for redictive Analytics. In credit scoring models, the accuracy of such unctions is typically assessed in separate sets of known examples – alidation or out-of-sample data sets. The idea behind this procedure s to mimic the accuracy of that function in future predictions of new xamples where x is known, but y is not. According to the Bayesian Decision Theory (Duda, Hart, & Stork, 001), a classification can be described by the prior probabilities of he classes p(y) and the class conditional probability density func- ion p(x|y) for the two classes, good (G) and bad (B). The classification ecision is made according to the posterior probabilities of the two lasses, which for class B can be represented as: p(B|x) = p(x|B)p(B)/p(x) (1) where p(x) = p(B)p(x|B) + p(G)p(x|G). Here, it is assumed that he costs for misclassifying a bad customer are the same as for the pposite situation, the equal costs assumption. It is worth recalling hat, in real-world financial environments, the costs of failing the pre- iction in a real bad are by far superior to failing in a real good. In he first case, there is essentially a loss of the exposure at default, oss Given Default (LGD), possibly mitigated with collateral. The sec- nd case affects the business, as it translates into a loss of margin. ousa and da Costa (2008) show several possibilities to overcome his practical issue, by adapting the output of standard classification ethods under the equal costs assumption to imbalanced costs for isclassification, associated with the decision and prediction tasks. It s worth discussing the related issue of class imbalance in credit scor- ng datasets. Quite often, these datasets contain a much smaller num- er of observations in the class of defaulters than in that of the good ayers (Brown & Mues, 2012; Marqués, García, & Sánchez, 2012a). large class imbalance is therefore present which some techniques ay not be able to successfully handle. Baseline methods to handle lass imbalance include oversampling the minority class or under ampling the majority class; Tomek links is an example of the for- er, SMOTE of the latter (Chawla, Bowyer, Hall, & Kegelmeyer, 2002). nother established approach to correct imbalance adopt a cost sen- itive classifier with the misclassification cost of the minority class reater than that of the majority class. Within this approach, it is orth mentioning MetaCost, a general method for making classifiers ost-sensitive (Domingos, 1999). All these methodologies, implicitly r explicitly, optimize the decision process for a specific business ob- ective. In other words, the optimization is made for a specific trade- ff between the error committed in identifying someone as defaulter hen one is in fact a non-defaulter individual and the opposite type f error of diagnosing someone as non-defaulter when one is in fact defaulter. This individualization is unconnected with our study and ny of these methods can be incorporated in the methodology under esearch. .2. Score formulation A credit scoring model is a simplification of the reality. The out- ut is a prediction of a given entity, actual or potential borrower, ntering in default in a given future period. Having decided on the efault concept, conventionally a borrower being in arrears for more han 90 days in the following 12 months, those matching the criteria re considered bad and the others are good. Other approaches may onsider a third status, the indeterminate, between the good and the ad classes, e.g. 15 to 90 days overdue, for which it may be unclear hether the borrower should be assigned to one class or to the other. his status is usually removed from the modeling sample, despite the odel can be used to score them. For simplicity, in this paper we will onsider the problem of two classes, although the proposed method- logy can easily be adapted to the other case. The output is a function of the input characteristics x, which is ost commonly referred as score, s(x). We also consider that this unction has a monotonic decreasing relationship with the probabil- ty of entering in default (i.e. reaching the bad status). A robust score- ard enables an appropriate differentiation between the good and the ad classes. It is achieved by capturing an adequate set of information or predicting the probability of the default concept (i.e. belonging to he bad class), based on previous known default occurrences. The no- ation of such probability, Pr{bad|score based on X}, is: p(B|s(x)) = p(B|s(x), x) = p(B|x), ∀x ∈ X (2) 344 M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 & g 2 c 2 w a a o u v i a l c t t a p S t 2 e m 3 3 p i b a a d e m c s f W e c t m t t u i n i ( t v p u t t o h Since p(G|x) + p(B|x) = 1, it naturally follows the probability of the complementary class: p(G|s(x)) = P(G|x) = 1 − p(B|x), ∀x ∈ X (3) Among researchers and real-world applications, a usual written form of the score is the log odds score: s(x) = ln p(G|x) p(B|x) , and p(G|x) + p(B|x) = 1. (4) In so saying, the score may vary from −∞, when P(G|x) = 0, to +∞, when P(G|x) = 1, i.e. s(x) ∈ R. The probability of the default event can be written in terms of the score: p(B|x) = 1/(1 + es(x)), ∀x ∈ X The most conventional way to produce log odds score is based in the logistic regression. However, other classification algorithms can also be used, adjusting the output to the scale of that function. In so saying, we assume that independently of the method used to de- termine the best separation between the two classes, good and bad, and then the resulting scorecard has the same property of the log odds score. Although a grounded mathematical treatment may be tempting to tackle this problem, it goes beyond the scope of this work. Notwithstanding, we provide some intuitions on the techni- cal material to survey. The basics of credit scoring and the most common approaches to build a scorecard, are further detailed in the operational research literature (Anderson, 2007; Crook, Edelman, & Thomas, 2007; McNab & Wynn, 2000; Thomas, 2009; Thomas et al., 2002). Recent advances in the area also deliver methods to build risk based pricing models (Thomas, 2009) and methodologies towards the optimization of the profitability to the lenders (Einav, Jenkins, & Levin, 2013). 2.3. Supervised classification methods The first approach to differentiate between groups took place in Fisher’s original work in (1936) for general classification prob- lems of varieties of plants. The objective was to find the best sep- aration between two groups, searching for the best combination of variables such that the groups were separated the most in the sub- space. Durand (1941) brought this methodology to finance for distin- guishing between good and bad consumer loans. Discriminant analysis was the first method used to develop credit scoring systems. Altman (1968) introduced it in the prediction of cor- porate bankruptcy. First applications in retail banking were mainly focused on credit granting in two categories of loans: consumer loans, and commercial loans (for an early review and critique on the use of discriminant analysis in credit scoring see Eisenbeis (1978)). The boom of credit cards demanded the automation of the credit decision task and the use of better credit scoring systems, which were doable due to the growth of computing power. The value of credit scoring became noticed and it was recognized as a much better predictor than any other judgmental scheme. Logistic regression (Steenackers & Goovaerts, 1989) and linear programming (see Chen, Zhong, Liao, and Li, 2013 for a review) were introduced in credit scoring, and they turned out to be the most used in financial industry (Anderson, 2007; Crook et al., 2007). The use of artificial intelligence tech- niques imported from statistical learning theory, such as classifi- cation trees (Breiman, Friedman, Olshen, & Stone, 1984; Quinlan, 1986) and neural networks (Desai, Crook, & Overstreet Jr, 1996; Jensen, 1992; Malhotra & Malhotra, 2002; West, 2000) have arisen in credit scoring systems. Support Vector Machine (SVM) is another method based in optimization and statistical learning, that received increased attention over the last decade in research in finance, ei- ther to build credit scoring systems for consumer finance or to predict bankruptcy (Li, Shiue, & Huang, 2006; Min & Lee, 2005; Wang, Wang, & Lai, 2005). Genetic algorithms (Chen & Huang, 2003; Ong, Huang, Tzeng, 2005), colony optimization (Martens et al., 2007), and re- ression and multivariate adaptive regression splines (Lee & Chen, 005) have also been tried. Evolutionary computing (Marqués, Gar- ía, & Sánchez, 2013), including genetic algorithms (Chen & Huang, 003; Ong et al., 2005) and colony optimization (Martens et al., 2007), as also considered for credit scoring. Regression (Lee & Chen, 2005) nd clustering (Wei, Yun-Zhong, & Ming-shu, 2014) techniques have lso been tailored to the problem. The choice of a learning algorithm is a difficult problem and it is ften based on which happen to be available, or best known to the ser (Jain, Duin, & Mao, 2000). The number of learning algorithms is ast. Many frameworks, adaptations to real-life problems, intertwin- ng of base algorithms were, and continue to be, proposed in the liter- ture, ranging from statistical approaches to state-of-the-art machine earning algorithms, from parametric models to non-parametric pro- edures (Abdou & Pointon, 2011; Baesens et al., 2003). As an alterna- ive to using a single method, a trend that is still evolving relates to he use of hybrid systems (Hsieh, 2005; Lee, Chiu, Lu, & Chen, 2002), nd ensemble of classifiers with which the outputs are achieved by a redefined sequence or rule, or a voting scheme (Marqués, García, & ánchez, 2012b; Wang, Hao, Ma, & Jiang, 2011). New concepts for adapting to changes (Adams, Tasoulis, Anagnos- opoulos, & Hand, 2010; Pavlidis et al., 2012; Sousa et al., 2013b; Yang, 007) and modeling the dynamics (Crook & Bellotti, 2010; Saberi t al., 2013) in populations start being exploited in credit risk assess- ent. . Dynamic modeling for credit default .1. Concept drift in credit default Credit default is mostly a consequence of financial distress. A erson, or a company, is in financial distress when is experiencing ndividual financial constraints or is being exposed to external distur- ances. In private individuals, financial constraints may result from brupt or intrinsic circumstances. In the first case, distress is usu- lly an outcome of sorrowful events like unemployment, pay cuts, ivorce, and disease. The second is most commonly related to over- xposure, low assets, erratic behavior, or bad management perfor- ance. In this paper we tackle the phenomenon of concept drift in redit default, which we now briefly explain. In the existing literature, concept drift is generally used to de- cribe changes in the target concept, which are activated by trans- ormations in the hidden context (Schlimmer & Granger Jr, 1986; idmer & Kubat, 1996) in dynamically changing and non-stationary nvironments. As a result of these transformations, the target concept an shift suddenly or just cause a change in the underlying data dis- ribution to the model. This means that with time, optimal features ay drift significantly from their original configuration or simply lose heir ability to explain the target concept. For example, a reduction of he minimum LTV (loan to value), tighten the space of possible val- es, which is noticed with a change in the distribution, and eventually n the credit default concept. When such drifts happen, the robust- ess of the model may significantly decrease, and in some situations t may no longer be acceptable. Some authors distinguish real concept drift from virtual drift Gama et al., 2014; Sun & Li, 2011; Tsymbal, 2004). The former refers o changes in the conditional distribution of the output (i.e., target ariable) given the input features, while the distribution of the in- ut may remain unchanged. The later refers to gradual changes in the nderlying data distribution with new sample data flowing, whereas he target concept does not change (Sun & Li, 2011). Real concept drift refers to changes in p(y|x), and it happens when he target concept of credit default evolves in time. Such changes can ccur either with or without a change in p(x). This type of drift may appen directly as a result of new rules for defining the target classes, M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 345 g d g r p d d o w i i fi s a s M d t p m i ( i i l V a m i t w c t p t 3 l d o c b t w t a d m T t s m t s i c f m g m d r t w d f p d fi i c t s b p s d ( b i i T fi t t b b d p M s 4 c t s t p a 4 c i i o a c i B u B T t 4 t g ood or bad, as those settled by regulators, when new criteria for efault are demanded to the banks. Examples of these include the uidelines for the minimum number of days past due or in the mate- iality threshold for the amount of credit in arrears, issued with the revious Basel II Accord. Another understanding of the real concept rift in credit default is associated with indirect changes in the hid- en context. In this case, credit default changes when evolving from ne stage of delinquency to another. For example, most of the people ith credit until five days past due tend to pay before the following nstallment, as most of them are just delayers. Yet, the part of debtors n arrears that also fail the next installment are most likely to be in nancial distress, possibly as a result of an abrupt or intrinsic circum- tance, and therefore they require more care from the bank. When rrears exceed three installments, the debtor is most certainly with erious financial constraints, and is likely to fail his credit obligations. ore extreme delays commonly translate into hard stages of credit efault, which require intensive tracking labor or legal actions. Virtual drifts happen when there are changes in the distribu- ion of the new sample data flowing without affecting the posterior robability of the target classes, p(y|x). With time, virtual drifts may ove to real concept drifts. Other interpretations can also be found n literature, for describing an incomplete representation of the data Widmer & Kubat, 1993), and changes in the data distribution lead- ng to changes in the decision boundary (Tsymbal, 2004). Accord- ng to some authors, other events can also be seen as virtual drifts, ike sampling shift (Salganicoff, 1997), temporary drifts (Lazarescu, enkatesh, & Bui, 2004), and feature change (Salganicoff, 1997). As n example of virtual drift, we might consider the credit decision- aking along the recent financial crisis. The lenders had to anticipate f a borrower would enter in default in the future (i.e. being bad). Al- hough the macroeconomic factors have worsened, employed people ith lower debt to income remained good for the lenders, and so they ontinued to have access to credit. Although we are mostly interested to track and detect changes in he real target concept, p(y|x), the methodology introduced in this aper attempts to cover both real concept and virtual drifts applied o the default concept drift detection and model rebuilding. .2. Methods for adaptation Traditional methods for building a scorecard consider a static earning setting. In so doing, this task is based in learning in a pre- efined sample of past examples and then used to predict an actual r a potential borrower, in the future. This is an offline learning pro- edure, because the whole training data set must be available when uilding the model. The model can only be used for predicting, af- er the training is completed, and then it is not re-trained alongside ith its utilization. In other words, when the best separation func- ion is achieved for a set of examples of the past, it is not updated for while, possibly for years, independently of the changes in the hid- en context or in the surrounding environment. New perspectives on odel building arise together with the possibility of learning online. he driving idea is to process new incoming data sequentially, so that he model may be continuously updated. One of the most intuitive ideas for handling concept drift by in- tance selection is to keep rebuilding the model from a window that oves over the latest batches and use the learn model for predic- ion on the immediate future. This idea assumes that the latest in- tances are the most relevant for prediction and that they contain the nformation of the current concept (Klinkenberg, 2004). A framework onnected with this idea consists in collecting the new incoming data or sequential batches in predefined time intervals, e.g. year by year, onth by month, or every day. The accumulation of these batches enerates a panel data flow for dynamic modeling. In Finance, it remains unclear whether it is best having a long emory or forgetting old events. If on the one hand, a long memory is esirable because it allows recalling a wide range of different occur- ences, in the other, many of those occurrences may no longer adjust o the present situation. A rapid adaptation to changes is achieved ith a short window, because it reflects the current distribution of efault more accurately. However, for the contrary reason, the per- ormance of models built upon shorter windows worsens in stable eriods. In credit risk assessment modeling, this matter has been in- irectly discussed by practitioners and researchers when trying to gure the pros and cons of using a through-the-cycle (TTC) or point- n-time (PIT) schema to calibrate the output of the scorecards to the urrent phase of the economic cycle. For years, a PIT schema was he only option, because banks did not have sufficient historical data eries. Since the implementation of the Basel II Accord worldwide, anks are required to store the data of default for a minimum 7-years eriod and consider a minimum of 5-years period for calibrating the corecards. An original idea of Widmer and Kubat (1996) uses a sliding win- ow of fixed length with a data processing structure first-in-first-out FIFO). Each window may consist of a single or multiple sequential atches, instead of single instances. At each new time step, the model s updated following two processes. In the first process, the model s rebuilt based on the training data set of the most recent window. hen, a forgetting process discards the data that move out of the xed-length window. Incremental algorithms (Widmer & Kubat, 1996) are a less ex- reme hybrid approach that allows updating the prediction of models o the new contexts. They are able to process examples batch-by- atch, or one-by-one, and update the prediction model after each atch, or after each example. Incremental models may rely on ran- om previous examples or in representative selected sets of exam- les, called incremental algorithms with partial memory (Maloof & ichalski, 2004). The challenge is to select an appropriate window ize. . Case study This research evolves from a one-dimensional analysis, where we ome across the financial outlook underlying the problem, to a mul- idimensional analysis along several points in time. The former, de- cribed in Sections 4.1, 4.2, and 4.3, is tailored to gain intuition on he default predictors and the main factors ruling the context of the roblem. The latter, in Section 4.4, is designed to gradually develop nd test a new dynamic framework to model credit risk. .1. Dataset and validation environment The research summarized here was conducted in a real-life finan- ial dataset, comprising 762,966 records, from a financial institution n Brazil along two years of operation, from 2009 to 2010. Each entity n the modeling dataset is assigned to a delinquency outcome - good r bad. In this problem, a person is assigned to the bad class if she had payment in delay for 60 or more days, along the first year after the redit has been granted. The delinquency rate in the modeling dataset s 27.3%, which is in line with the high default rates in credit cards in razil, one of the countries with the highest default rates in the prod- ct. The full list of variables in the original data set is available in the RICS 2013 official website. It contains 39 variables, categorized in able 1, and one target variable with values 1 identifying a record in he bad class and 0 for the good class. .2. Data analysis and cleansing Some important aspects of the datasets were considered, because hey can influence the performance of the models. These aspects re- ard to: 346 M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 Table 1 Predictive variables summary. Type # Information Numerical 6 Age, monthly income, time at current address, time at current employer, number of dependents, and number of accounts in the bank. Treated as nominal 13 Credit card bills due day, 1st to 4th zip digit codes, home (state, city, and neighborhood), marital status, income proof type, long distance dialing code, occupation code, and type of home. Binary 16 Address type proof, information of the mother and fathers names, input from credit bureau, phone number, bills at the home address, previous credit experience, other credit cards, tax payer and national id, messaging phone number, immediate purchase, overdraft protection agreement, lives and work in the same state, lives and work in the same city, and gender. Date 1 Application date. ID 3 Customer, personal reference, and branch unique identifiers. 2009 2010 0% 20% 40% 60% 80% 100% 72,532 109,200 159,800 225,800 325,700 902,000 C u m u la ti ve f re q u en cy Income (R$) Fig. 1. Cumulative frequency of the monthly income for 2009 and 2010. Table 2 Information values for the tested combi- nations. Combination IV Age × income 0.315 Age × occupation 0.009 Income × marital status 0.208 Income × occupation 0.334 Income × proof of income 0.123 Age × income × occupation 0.007 c t W n c 4 i s t G u g c b r a t w u t 4 t w I d a i 0 0 4 n a w • Significant percent of zero or missing values In exception to the vari- ables ‘lives and work in the same state’ and ‘previous credit expe- rience’, binary flags have 95% to 100% concentrated in one of the values, which turn them practically unworkable. The same occurs for the numerical variables number of dependents and number of accounts in the bank, both with more than 99% zeroes. The re- maining variables were reasonably or completely populated. • Outliers and unreasonable values The variable age presents 0.05% of applications assigned to customers with ages between 100 and 988 years. A small percent of values out of the standard ranges are observable in the variables credit card bills due day, monthly in- come and time at current employer. Unreasonable values are de- tected in the first semester of 2009, suggesting that the data were subjected to corrections from the second semester of 2009 on- wards. • Unreliable and informal information Little reliability on socio- demographic data is amplified by specific conditions in the back- ground of this problem. This type of scorecards is usually based in verbal information that the customer provides, and in most of the cases no certification is made available. In 85% of the applications, no certification for the income was provided, and 75% do not have proof for the address type. Customers have little or no concern to provide accurate information. The financial industry is aware of this kind of limitations. However, in highly competitive environ- ments there is little chance to amend them, while keeping in the business. Hence, other than regulatory imperatives, no player is able to efficiently overcome this kind of data limitations. As cur- rently there are no such imperatives in Brazilian financial market, databases attached to this type of models are likely to keep lack- ing reliability in the near future. • Bias on the distributions of modeling examples The most noticeable bias is in the variable monthly income, where values shift from one year to another, exhibited in Fig. 1. This is most likely related to increases in the minimum wages and inflation. Slight variations are also observable in the geographical variables, which are possibly related with the geographical expansion of the institution. In the remaining characteristics, the correlation between the frequency distributions of 2009 and 2010 range from 99 to 100%, suggesting a very stable pattern during the analyzed period. 4.3. Data transformation and new characteristics 4.3.1. Data cleansing and new characteristics We focused the data treatment on the characteristics that were reasonably or fully populated. Fields state, city, and neighborhood ontain free text, and were subjected to a manual cleansing. At- ributes with 100 or less records were assigned to a new class “Other”. e could observe that there may be neighborhoods with the same ame in different cities; and hence we concatenated these new leansed fields, state and city, into the same characteristic. .3.2. Data transformation Variables were transformed using the weights of evidence (WoE) n the complete modeling dataset, which is a typical measure in credit core modeling (FICO, 2006). W oE = ln g/G b/B , where g and b are respec- ively the number of good and the number of bad in the attribute, and and B are respectively the total number of good and bad in the pop- lation sample. The larger the WoE the higher is the proportion of ood customers in the bin. For the nominal and binary variables we alculated the WoE for each class. Numerical variables were firstly inned using SAS Enterprise Miner, and then manually adjusted to eflect domain knowledge. In so doing we aim to achieve a set of char- cteristics less exposed to overfitting. Cases where the calculation of he WoE rendered impossible - one of the classes without examples – ere given an average value. The same principle was applied to val- es out of the expected ranges (e.g. credit card bills due day higher han 31). .3.3. One-dimensional analysis The strength of each potential characteristic was measured using he information value (IV) in the period, IV = ∑ni=1 (g/G − b/B)W oE, here n is the number of bins in the characteristic. The higher is the V, the higher is the relative importance of the characteristic. In a one- imensional basis, for the entire period, the most important char- cteristics are age, occupation, time at current employer, monthly ncome and marital status, with information values of 0.368, 0.352, .132, 0.117, and 0.116, respectively. Remaining characteristics have .084 or less. .3.4. Interaction terms Using the odds in each attribute of the variables, we calculated ew nonlinear characteristics using interaction terms between vari- bles to model the joint effects. We tested six combinations, for which e present the information value in Table 2. M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 347 4 d o r c e i i d c 4 s m w t u s c b A t t S l b t p e i t o f b p t m a s o s T a w r T i s t f m m c d u a T t m s s i t a t r s 5 t t w t t A s v l f q a y t n w s s 5 m w f o t s t i d i t t s t w .3.5. Time series descriptive analysis Fig. 2a shows the real concept drift along 2009–2010. The highest efault rates are noticed in the first quarter of 2009, and at the end f 2010. Fig. 2b displays the evolution of the business in the same pe- iod. It exhibits two features of the business. First, we can see that the redit cards business follow an annual seasonality, increasing along ach year. Second, the credit cards business is rising over time, which s related with the expansion of the branch network of the financial nstitution. The decrease of default rate during 2009 suggests that the ecision-making process might have been slightly enhanced, when omparing to the beginning of the period. .4. Dynamic modeling framework The dynamic modeling framework presented in this research con- iders that data is processed batch-by-batch. Sequentially, at each onthly window, a new model is learned from a previous selected indow, including the most recent month. To mimic the time evolu- ion, we assumed that the current month gradually shifts from 2009 ntil the third quarter of 2010. Each learning unit for the model building was grounded on a static etting. The training of each unit consists of a supervised classifi- ation procedure, executed in three steps. First, characteristics are inned. Second, the classification model is designed with Generalized dditive Models (GAM) and a 10 fold crossed-validation, upholding he classification algorithm used to develop the winning model in he BRICS 2013 in data mining and finance (BRICS-CCI&CBIC, 2013; ousa et al., 2013b). Concurrently, the best set of characteristics is se- ected until no other characteristic in the training dataset adds contri- ution to the information value (IV) of the model. In this application he threshold was set for a minimum increment of 0.03. Third, the erformance of the model is measured based on the Gini coefficient, quivalent to consider the area under the ROC curve (AUC), which s a typical evaluation criteria among researchers and in the indus- ry (Řezáč & Řezáč, 2011). This coefficient refers to the global quality f the credit scoring model, and ranges between −1 and 1. The per- ect scoring model fully distinguishes the two target classes, good and ad, and has a Gini coefficient equal to 1. A model with a random out- ut has a Gini coefficient equal to zero. If the coefficient is negative, hen the scores have a reverse meaning. The extreme case −1 would ean that all examples of the good class are being predicted as bad, nd vice versa. In this case, the perfect model can be achieved just by witching the prediction. At each month, instances for modeling are selected from all previ- us available batches, according to a selection mechanism. We use in- tance selection methods to test the hypothesis under investigation. wo methods for tackling default concept drift were implemented - full memory time window, and a fixed short memory time window ith a forgetting mechanism. The full memory time window assumes that the learning algo- ithm generates the model based on all previous instances (Fig. 3a). he process is incremental, so every time a new instance arises, it s added to the training set, and a new model is build. This schema hould be appropriate to detect mild concept drifts, but it is unable o rapidly adapt to major changes. Models of this schema should per- orm suitably in stable environments. A shortcoming of this incre- ental schema is that the training dataset quickly expands which ay requires a huge storage capacity, and constrain the use of some lassification algorithms, to be able of processing the expanding ataset. In the fixed short memory time window, the model development ses the most recent window. With this schema, illustrated in Fig. 3b, new model is build in each new batch, by forgetting past examples. he fundamental assumption is that past examples have low correla- ion with the current default concept. Under this setting, the dynamic odeling should quickly adapt to changes. The most extreme case of hort memory time window is when only the current example is con- idered to train the new model, which represents to the online learn- ng without any memory of the past. A deficiency of this method is hat it often lacks of generalization ability in stable conditions that is mplified with extremely short windows. These modeling frameworks enable comparing these configura- ions between themselves, and also compare them with the model eached with a static learning setting. The research questions of this tudy should be answered following the reasoning: • If the full memory time window outperforms the other schema, then more recent data are not fundamental for the prediction; the environment of the decision-making should be in a stable phase. Otherwise, the default concept is drifting, and so the most recent data are more relevant for the prediction. • If a model built with static learning in the first window of the period has the best performance, then older data can improve the prediction. This may happen, for example, when a new credit product is launched, and the credit decision-making criteria are adjusted afterwards. In such case, the oldest data are more rep- resentative, as they can illustrate a more diverse range of risk be- haviors. Otherwise, over the long-run, dynamic modeling should outperform the model learnt with static setting. . Experimental results We assessed the performance of the sequential models built with he dynamic modeling framework introduced in the previous sec- ion, through the period 2009 and 2010. The experimental design as drawn for assessing the performance in the modeling period, in he short-term, and in the farthest-term. In each model rebuilding, he performance in the modeling period was assessed in the test set. dditionally, using two out-of-sample windows, we measured the hort-term performance of the model in the month following the de- elopment, and the farthest-term performance was measured in the ast quarter of 2010. Although we have considered monthly windows or developing the model, for the long-run assessment we chose a uarterly window instead of a single month. In so doing, potential typical properties of the decision-making process at the end of the ear were smoothed. In this section, we provide further evidence on temporal degrada- ion of static credit scoring. Then, we challenge the robustness of the ew concept of dynamic modeling against a static model, developed ith a traditional framework. We finally present and discuss the re- ults for the two sliding-window configurations - full memory and hort memory. .1. Temporal degradation of static credit scoring The temporal degradation of the credit scoring is detected when easuring the performance of each model in the sequence generated ith the dynamic modeling. Fig. 4a and b exhibit the Gini coefficient or each model, measured in the modeling test set and two different ut-of-sample windows, one month after rebuild the model, and in he farthest quarter in the period (2010 Q4). Fig. 4a shows the performance along the entire period with the hort memory configuration. One month after rebuilding the model, he performance curve is always below the performance measured n the modeling period, showing that the performance consistently ecreases one month after rebuilding the model. When evaluat- ng the performance with the full memory configuration, in Fig. 4b, he extent of degradation within a month is not consistent over he period. During the first semester of 2009, performance mea- ured in the month after rebuilding the model is slightly superior to he one measured in the modeling period, and from that point on- ards, it is marginally inferior. This may suggest that the short-term 348 M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 Default Average default 17% 22% 27% 32% 37% 42% 2009 -Jan 2009 -Apr 2009 -Jul 2009 -Oct 2010-Jan 2010-Apr 2010-Jul 2010-Oct (% ) r a te D e fa u lt Application month (a) Default rate. 0 20,000 40,000 60,000 80,000 2009 -Jan 2009 -Apr 2009 -Jul 2009 -Oct 2010-Jan 2010-Apr 2010-Jul 2010-Oct (# ) c a r d s c r e d it U n d e r w r it e d Month (b) Number of new credit cards contracts. Fig. 2. Default rate and new contracts in the period 2009–2010. … Learning 2 Learning 1 Learning 3 Learning n timeMonth 1 Month 2 Month 3 Month n … (a) Full memory time window. … Learning 2 Learning 1 Learning 3 Learning n timeMonth 1 Month 2 Month 3 Month n … (b) Fixed short memory time window. Fig. 3. Configurations for tackling concept drift in credit default. Modeling period 1 month after rebuilding Farthest quarter (2010 Q4) 0.27 0.31 0.35 0.39 0.43 0.47 0.51 2009-Jan 2009-Mar 2009-Jun 2009-Sep 2009-Dec 2010-Mar 2010-Jun 2010 - Sep c o e ff ic ie n t G in i Model rebuilding month (a) Short memory. Modeling period 1 month after rebuilding Farthest quarter (2010 Q4) 0.27 0.31 0.35 0.39 0.43 0.47 0.51 2009-Jan 2009-Mar 2009-Jun 2009-Sep 2009-Dec 2010-Mar 2010-Jun 2010 - Sep c o e ff ic ie n t G in i Model rebuilding month (b) Full memory. Fig. 4. Gini coefficient of the sequence of models produced with the dynamical modeling. e t t c i performance is more similar to the performance in the modeling pe- riod when using the full memory configuration. The extent of degradation is higher, when the performance of the model is measured at the end of the period (2014 Q4). The farthest is the point of the prediction from the point of the development; the highest is the extent of degradation of the performance. These ffects are consistently perceived on the two windowing configura- ions - short memory (Fig. 4a) and full memory (Fig. 4b). Considering the real performance of the models one month after hey were built, the average degradation of the models sequentially onstructed, shown in Table 3, is 0.02 in the short memory and 0.01 n the full memory configuration. In the farthest quarter in the period M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 349 Table 3 Average degradation of the sequence of models produced with the dynamic modeling. Memory type Gini index Degradation Modeling Month after Farthest quarter Month after Farthest quarter period rebuilding (2010 Q4) rebuilding (2010 Q4) Short 0, 40 0, 38 0, 33 −0, 02 −0, 07 Full 0, 39 0, 38 0, 33 −0, 01 −0, 06 ( u q t 5 p s n m s i c q p h o p - c t m m m t h i h t m T o i n u s 5 w o t t H f m t t t c t t d 0 H s b i d 6 a u c r d a b e d d a d S w m t p t p n i t o t l l m d c o ( w c e i i n t e 2010 Q4) the degradation reaches 0.07 in the short memory config- ration and 0.06 in the full memory schema. Although degradation can be observed in all models of the se- uence, updating the model always yields the best discrimination be- ween the target classes - good and bads. .2. Dynamic versus static The proposed dynamic modeling framework enables a major im- rovement of the initial static model, which was trained with the ample in the first month of 2009 (2009 M1). Fig. 5a shows the immediate performance achieved with the dy- amic modeling – full and short term memory – versus the static odel, measured in the month following the development. Fig. 5b hows the performance of the models in each point in time, measured n the farthest quarter of the period. Consistently, for both memory onfigurations and performance criteria, immediate or in the farthest uarter, either the static or the dynamic modeling performances im- rove until the third quarter of 2009, which might reflect the en- ancement of the set of characteristics x that was partially corrected ver that period. In Fig. 5a, we observe a certain overlap between the immediate erformance achieved with the two types of memory configurations short and full. For all the periods, the short-term performance in- reases until the third quarter of 2009, and slightly decreases from hat point onwards. The extent of improvement with the dynamic odeling reaches 0.05 in 2010. Fig. 5b shows that the farthest-term performance of the first odel in the sequence of the dynamic modeling, same as the static odel, is significantly improved with the sequential rebuilding un- il the third quarter of 2009, possibly as a consequence of the en- ancement of the set of characteristics. In this period, performance ncreases from 0.28 to 0.36, meaning that the risk assessment is en- anced with the new dynamic modeling, rather than the static. From hat quarter onwards, the long-run predictions given by the dynamic odeling slightly improve, and always outperform the static frame. his suggests that, the new incoming data allow a better knowledge f the new context. Although we know beforehand that the increase n performance is somewhat a consequence of the training being earer to the out-of-sample validation window, still we can see that sing the newest data improves the initial prediction given by the tatic model (2009 Q1). .3. Memory - keep or lose it The new dynamic modeling framework enables investigating on hether it is preferable keeping a long-term memory or forgetting lder observations, or if they are equivalents in some contexts. From he second semester of 2009 onwards, the best results in the farthest- erm (2014 Q4) are reached with the full memory configuration. owever, we realize that there is a certain overlap between the per- ormances of the sequence of models resulting from the two types of emory configurations, both for the short-term and for the farthest- erm. This suggests that, in the period, the information contained in he older examples remain appropriate for the default target, and that he context is not drifting as a result of particular changes in the set of haracteristics. Hence, drifts in particular characteristics, like income, ranslate into virtual drifts because they did not have an impact in he distribution of target concept, p(y|x). To some extent, the imme- iate performance, exhibited in Fig. 5a, decreases during 2010 from .44 to 0.38%, which could be interpreted as the presence of a drift. owever, as the timeframe is small, it remains uncertain if it is a tran- itory outcome or a persistent drift in the context, potentially caused y changes in features that are not represented in the set of character- stics available for modeling in this application, like macroeconomic ata. . Conclusions This research presents a new modeling framework for credit risk ssessment that extends the prevailing credit scoring models built pon historical data static settings. Our framework mimics the prin- iple of films, by composing the model with a sequence of snapshots, ather than a single picture. Within the new modeling schema, pre- ictions are made upon the sequence of temporal data, and are suit- ble for adapting to the occurrence of real concept drifts, translated y changes in the population, in the economy or in the market. It also nables improving the existing models based on the newest incoming ata. We present an empirical simulation using a real-world financial ataset of 762,966 credit cards, from a financial institution in Brazil long two years of operation. A first conclusion is that monthly up- ates avoid the degradation of the model following the development. econdly, newest data consistently improve the forecasting accuracy, hen compared to the previous models in the sequence of dynamic odeling, both in a short-term as in a full-term memory configura- ion. Particularly, the static model available at the beginning of the eriod is outperformed by every succeeding model, suggesting that he dynamic modeling framework has the ability of improving the rediction by integrating new incoming data. Third, a slight domi- ance is achieved with the full-term memory, suggesting that older nformation remains meaningful for predicting default target within he analyzed period. In banking industry, prevailing credit scoring models are devel- ped from static windows and kept unchanged possibly for years. In his setting, the two basic mechanisms of memory, short-term and ong-term memory are fundamental to learning, but are still over- ooked in current modeling frameworks. As a consequence, these odels are insensitive to changes, like population drifts or financial istress. The usual outcomes are the default rates rising and abrupt redit cuts, as those that were observed in the U.S. in the aftermath f the last Global crisis (as documented by Sousa, Gama, and Brandão 2015)). This problem could be overcome with the proposed frame- ork, since it would allow to gradually relearning along time and hanges. Still, there are some real business problems with rebuilding mod- ls over time. First, lenders have little incentive to enhance the exist- ng rating systems frameworks because there is a recycling idea that t expensive and time-consuming to build new scorecards. Then, they eed to be internally tested and validated, and then regulators need o approve them. Second, regulators still promote models whose co- fficients do not change over time. This is one area where practice is 350 M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 Short memory Full memory Static model (2009 M1) 0.26 0.28 0.30 0.32 0.34 0.36 0.38 0.40 0.42 0.44 2009-Jan 2009-Mar 2009-Jun 2009-Sep 2009-Dec 2010-Mar 2010-Jun 2010 - Sep c o e ff ic ie n t G in i Model rebuilding month (a) Short-term performance. Short memory Full memory Static model (2009 M1) 0.26 0.28 0.30 0.32 0.34 0.36 0.38 0.40 0.42 0.44 2009-Jan 2009-Mar 2009-Jun 2009-Sep 2009-Dec 2010-Mar 2010-Jun 2010 - Sep c o e ff ic ie n t G in i Model rebuilding month (b) Farthest-term performance, 2010 Q4. Fig. 5. Performance with the dynamic modeling – full and short memory – versus the static model (2009 M1). B B B C C C C C C D D D D E E F F G G H H J J K L L far distant from the technical advances, and new thoughts, like sim- plifying current decision layers, need to be encouraged. There are some important topics in default concept drift that we did not consider, which we defer for future research. While this pa- per provides convincing results, some additional simulations using real-world datasets from highly stressed economic environments and longer time frames would be valuable. Second, modeling the delin- quency presents a specificity since a window of time is required in order to measure the outcome, i.e. the true class, before the new model is built. Therefore, for forecasting, it turns out that there will be a time gap of the same length between the values of predictor variables used in the model and the first possible forecast period in the future. Although this is not a problem of the proposed methodol- ogy, future research should bring new insights to overcome this issue, with a view on practicality. Third, some good alternatives to using windows of data blocks are encouraged, which may be based on us- ing ensembles of the models learned in the past, possible combining the two components of memory, short-term and long-term memory, or a forgetting factor method. There is some material on this going back to Adams et al. (2010). Fourth, our empirical study considered a set of fixed predictors. Therefore, future research should consider sets of predictor of variable length. This is important for detecting concept drift, because the set predictors that are being used may be quite limited to exhibit signs of change, even if they are occurring in the environment. Finally, performance is reported in this paper, but the conditions leading to difference in performance are not explored. This is another future research direction. References Abdou, H. A., & Pointon, J. (2011). Credit scoring, statistical techniques and evaluation criteria: A review of the literature. Intelligent Systems in Accounting, Finance and Management, 18(2-3), 59–88. Adams, N. M., Tasoulis, D. K., Anagnostopoulos, C., & Hand, D. J. (2010). Temporally- adaptive linear classification for handling population drift in credit scoring. In Y. Lechevallier, & G. Saporta (Eds.), Proceedings of compstat’2010 (pp. 167–176). Physica-Verlag HD. Altman, E. I. (1968). Financial ratios, discriminant analysis and the prediction of corpo- rate bankruptcy. Journal of Finance, 23(4), 589–609. Anderson, R. (2007). The credit scoring toolkit: Theory and practice for retail credit risk management and decision automation. OUP Oxford. Avery, R. B., Calem, P. S., & Canner, G. B. (2004). Consumer credit scoring: Do situational circumstances matter? Journal of Banking & Finance, 28(4), 835–856. Baesens, B., Van Gestel, T., Viaene, S., Stepanova, M., Suykens, J., & Vanthienen, J. (2003). Benchmarking state-of-the-art classification algorithms for credit scoring. Journal of the Operational Research Society, 54(6), 627–635. BCBS (2006). International convergence of capital measurement and capital standards: A revised framework - comprehensive version. Bank for International Settlements. Bellotti, T., & Crook, J. (2013). Forecasting and stress testing credit card default using dynamic models. International Journal of Forecasting, 29(4), 563–574. BIS (2004). Implementation of Basel II: Practical considerations. Bank for International Settlements. reiman, L., Friedman, J., Olshen, R., & Stone, C. (1984). Classification and regression trees. Belmont, California: Wadsworth International Group. RICS-CCI&CBIC (2013). CI algorithms competition (CIAC): Credit risk assessment sys- tem robustness against degradation and seasonal variation. http://brics- cci.org/ ci- algorithms- competition- ciac/ Accessed 19.06.13. rown, I., & Mues, C. (2012). An experimental comparison of classification algorithms for imbalanced credit scoring data sets. Expert Systems with Applications, 39(3), 3446–3453. http://dx.doi.org/10.1016/j.eswa.2011.09.033. hawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. Journal of artificial intelligence research, 321– 357. hen, D., Zhong, Y., Liao, Y., & Li, L. (2013). Review of multiple criteria and multi- ple constraint-level linear programming. Procedia Computer Science, 17(0), 158– 165. hen, M.-C., & Huang, S.-H. (2003). Credit scoring and rejected instances reassigning through evolutionary computation techniques. Expert Systems with Applications, 24(4), 433–441. rook, J., & Bellotti, T. (2010). Time varying and dynamic models for default risk in consumer loans. Journal of the Royal Statistical Society: Series A (Statistics in Society), 173(2), 283–305. rook, J. N., Edelman, D. B., & Thomas, L. C. (2007). Recent developments in consumer credit risk assessment. European Journal of Operational Research, 183(3), 1447– 1465. rook, J. N., Thomas, L. C., & Hamilton, R. (1992). The degradation of the scorecard over the business cycle. IMA Journal of Management Mathematics, 4(1), 111–123. esai, V. S., Crook, J. N., & Overstreet Jr, G. A. (1996). A comparison of neural networks and linear scoring models in the credit union environment. European Journal of Operational Research, 95(1), 24–37. omingos, P. (1999). Metacost: A general method for making classifiers cost-sensitive. In Proceedings of the fifth ACM SIGKDD international conference on knowledge discov- ery and data mining. In KDD ’99 (pp. 155–164). New York, NY, USA: ACM. uda, R. O., Hart, P. E., & Stork, D. G. (2001). Pattern classification. John Wiley & Sons. urand, D. (1941). Risk elements in consumer installment financing. National Bureau of Economic Research, Inc. inav, L., Jenkins, M., & Levin, J. (2013). The impact of credit scoring on consumer lend- ing. The RAND Journal of Economics, 44(2), 249–274. isenbeis, R. A. (1978). Problems in applying discriminant analysis in credit scoring models. Journal of Banking & Finance, 2(3), 205–219. ICO (2006). Introduction to scorecard for FICO model builder, . isher, R. A. (1936). The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2), 179–188. ama, J. (2010). Knowledge discovery from data streams. London: Chapman & Hall/CRC. ama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., & Bouchachia, A. (2014). A survey on concept drift adaptation. ACM Comput. Surv., 46(4), 44:1–44:37. doi:10.1145/ 2523813. and, D. J. (2006). Classifier technology and the illusion of progress. Statistical Science, 21(1), 30–34. sieh, N.-C. (2005). Hybrid mining approach in the design of credit scoring models. Expert Systems with Applications, 28(4), 655–665. ain, A. K., Duin, R. P. W., & Mao, J. (2000). Statistical pattern recognition: A review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(1), 4–37. ensen, H. L. (1992). Using neural networks for credit scoring. Managerial Finance, 18(6), 15–26. linkenberg, R. (2004). Learning drifting concepts: Example selection vs. example weighting. Intelligent data analysis, 8(3), 281–300. azarescu, M. M., Venkatesh, S., & Bui, H. H. (2004). Using multiple windows to track concept drift. Intelligent data analysis, 8(1), 29–59. ee, T.-S., & Chen, I.-F. (2005). A two-stage hybrid credit scoring model using artificial neural networks and multivariate adaptive regression splines. Expert Systems with Applications, 28(4), 743–752. http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0010 http://brics-cci.org/ci-algorithms-competition-ciac/ http://dx.doi.org/10.1016/j.eswa.2011.09.033 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0023 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0023 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0024 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0024 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0025 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0025 http://dx.doi.org/10.1145/2523813 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0033 M.R. Sousa et al. / Expert Systems With Applications 45 (2016) 341–351 351 L L L M M M M M M M M O P Q R S S S S S S S S S T T T T W W W W W W Y ee, T.-S., Chiu, C.-C., Lu, C.-J., & Chen, I.-F. (2002). Credit scoring using the hybrid neural discriminant technique. Expert Systems with Applications, 23(3), 245–254. i, S.-T., Shiue, W., & Huang, M.-H. (2006). The evaluation of consumer loans using sup- port vector machines. Expert Systems with Applications, 30(4), 772–782. ucas, A. (2004). Updating scorecards: Removing the mystique. In Readings in credit scoring: foundations, developments, and aims (pp. 93–109). New York: Oxford Uni- versity Press. alhotra, R., & Malhotra, D. K. (2002). Differentiating between good credits and bad credits using neuro-fuzzy systems. European Journal of Operational Research, 136(1), 190–211. aloof, M. A., & Michalski, R. S. (2004). Incremental learning with partial instance memory. Artificial intelligence, 154(1), 95–126. arqués, A. I., García, V., & Sánchez, J. S. (2012a). On the suitability of resampling tech- niques for the class imbalance problem in credit scoring. Journal of the Operational Research Society, 64(7), 1060–1070. arqués, A. I., García, V., & Sánchez, J. S. (2012b). Two-level classifier ensembles for credit risk assessment. Expert Systems with Applications, 39(12), 10916–10922. arqués, A. I., García, V., & Sánchez, J. S. (2013). A literature review on the application of evolutionary computing to credit scoring. Journal of the Operational Research So- ciety, 64(9), 1384–1399. artens, D., De Backer, M., Haesen, R., Vanthienen, J., Snoeck, M., & Baesens, B. (2007). Classification with ant colony optimization. IEEE Transactions on Evolutionary Com- putation, 11(5), 651–665. cNab, H., & Wynn, A. (2000). Principles and practice of consumer credit risk manage- ment. CIB Publishing. in, J. H., & Lee, Y.-C. (2005). Bankruptcy prediction using support vector machine with optimal choice of kernel function parameters. Expert Systems with Applica- tions, 28(4), 603–614. ng, C.-S., Huang, J.-J., & Tzeng, G.-H. (2005). Building credit scoring models using ge- netic programming. Expert Systems with Applications, 29(1), 41–47. avlidis, N., Tasoulis, D., Adams, N., & Hand, D. (2012). Adaptive consumer credit classi- fication. Journal of the Operational Research Society, 63(12), 1645–1654. uinlan, J. R. (1986). Induction of decision trees. Machine learning, 1(1), 81–106. ˇ ezáč, M., & Řezáč, F. (2011). How to measure the quality of credit scoring models. Finance a Uver: Czech Journal of Economics & Finance, 61(5), 486–507. aberi, M., Mirtalaie, M. S., Hussain, F. K., Azadeh, A., Hussain, O. K., & Ashjari, B. (2013). A granular computing-based approach to credit scoring modeling. Neurocomput- ing, 122(0), 100–115. alganicoff, M. (1997). Tolerating concept and sampling shift in lazy learning using pre- diction error context switching. Artificial Intelligence Review, 11(1-5), 133–155. chlimmer, J. C., & Granger Jr, R. H. (1986). Incremental learning from noisy data. Ma- chine learning, 1(3), 317–354. ousa, M. R., & da Costa, J. P. (2008). A tripartite scorecard for the pay/no pay decision- making in the retail banking industry. Frontiers in Artificial Intelligence and Applica- tions, 45. ousa, M. R., Gama, J., & Brandão, E. (2015). Links between scores, real default and pric- ing: Evidence from the Freddie Mac’s loan-level dataset. Journal of Economics, Busi- ness and Management, 3(12), 1106–1114. ousa, M. R., Gama, J., Brandão, E., et al. (2013a). Introducing time-changing economics into Credit Scoring. Technical Report. Universidade do Porto, Faculdade de Econo- mia do Porto. ousa, M. R., Gama, J., & Gonçalves, M. J. S. (2013b). A two-stage model for dealing with temporal degradation of credit scoring. In Proceedings of BRICS-CCI & CBIC. teenackers, A., & Goovaerts, M. (1989). A credit scoring model for personal loans. In- surance: Mathematics and Economics, 8(1), 31–34. un, J., & Li, H. (2011). Dynamic financial distress prediction using instance selection for the disposal of concept drift. Expert Systems with Applications, 38(3), 2566–2576. homas, L. C. (2009). Consumer credit models: pricing, profit and portfolios: Pricing, profit and portfolios. Oxford University Press. homas, L. C. (2010). Consumer finance: Challenges for operational research. Journal of the Operational Research Society, 61, 41–52. homas, L. C., Edelman, D. B., & Crook, J. N. (2002). Credit scoring and its applications. Philadelphia: Society for Industrial and Applied Mathematics. symbal, A. (2004). The problem of concept drift: Definitions and related work. Dublin: Computer Science Department, Trinity College. ang, G., Hao, J., Ma, J., & Jiang, H. (2011). A comparative assessment of ensemble learn- ing for credit scoring. Expert systems with applications, 38(1), 223–230. ang, Y., Wang, S., & Lai, K. (2005). A new fuzzy support vector machine to evaluate credit risk. IEEE Transactions on Fuzzy Systems, 13(6), 820–831. ei, G., Yun-Zhong, C., & Ming-shu, C. (2014). A new dynamic credit scoring model based on the objective cluster analysis. In Practical applications of intelligent sys- tems. In Advances in Intelligent Systems and Computing: 279 (pp. 579–589). Springer Berlin Heidelberg. est, D. (2000). Neural network credit scoring models. Computers & Operations Re- search, 27(11), 1131–1152. idmer, G., & Kubat, M. (1993). Effective learning in dynamic environments by explicit context tracking. In Machine learning: Ecml-93 (pp. 227–243). Springer. idmer, G., & Kubat, M. (1996). Learning in the presence of concept drift and hidden contexts. Machine learning, 23(1), 69–101. ang, Y. (2007). Adaptive credit scoring with kernel learning methods. European Journal of Operational Research, 183(3), 1521–1536. http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0036 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0036 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0038 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0038 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0038 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0038 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0039 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0039 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0039 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0039 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0039 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0040 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0040 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0040 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0040 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0040 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0041 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0041 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0041 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0041 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0041 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0043 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0043 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0043 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0043 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0044 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0044 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0044 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0044 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0047 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0047 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0048 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0048 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0048 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0048 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0049 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0049 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0049 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0049 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0049 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0049 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0049 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0049 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0050 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0050 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0051 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0051 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0051 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0051 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0052 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0052 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0052 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0052 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0053 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0053 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0053 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0053 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0053 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0054 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0054 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0054 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0054 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0054 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0055 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0055 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0055 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0055 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0055 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0056 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0056 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0056 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0056 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0057 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0057 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0057 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0057 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0058 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0058 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0059 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0059 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0060 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0060 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0060 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0060 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0060 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0061 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0061 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0062 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0062 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0062 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0062 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0062 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0062 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0063 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0063 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0063 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0063 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0063 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0064 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0064 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0064 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0064 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0064 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0065 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0065 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0066 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0066 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0066 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0066 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0067 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0067 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0067 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0067 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0068 http://refhub.elsevier.com/S0957-4174(15)00684-3/sbref0068 A new dynamic modeling framework for credit risk assessment 1 Introduction 2 Settings and concepts 2.1 Supervised learning problem 2.2 Score formulation 2.3 Supervised classification methods 3 Dynamic modeling for credit default 3.1 Concept drift in credit default 3.2 Methods for adaptation 4 Case study 4.1 Dataset and validation environment 4.2 Data analysis and cleansing 4.3 Data transformation and new characteristics 4.3.1 Data cleansing and new characteristics 4.3.2 Data transformation 4.3.3 One-dimensional analysis 4.3.4 Interaction terms 4.3.5 Time series descriptive analysis 4.4 Dynamic modeling framework 5 Experimental results 5.1 Temporal degradation of static credit scoring 5.2 Dynamic versus static 5.3 Memory - keep or lose it 6 Conclusions References 
work_oi3geyph7bgajkhssj6zkjgjfa	----	Modular design to support green life-cycle engineering Modular design to support green life-cycle engineering Hwai-En Tseng a,*, Chien-Chen Chang b, Jia-Diann Li a a Department of Industrial Engineering and Management, National Chin-Yi Institute of Technology, 35, Lane 215, Section 1, Chung-Shan Road, Taiping City, Taichung County 411, Taiwan, ROC b Department of Industrial Design, Huafan University, No. 1, Huafan Road, Shihtin Hsiang, Taipei Hsien, Taiwan, ROC Abstract The severe competition in the market has driven enterprises to produce a wider variety of products to meet consumers’ needs. However, frequent variation of product specifications causes the assembly and disassembly of components and modules to become more and more complicated. As a result, the issue of product modular design is a problem worthy of concern. In this study, engineering attri- butes were added to the liaison graph model for the evaluation of part connections. The engineering attributes added, including contact type, combination type, tool type, and accessed direction, serve to offer designers criteria for evaluating the component liaison intensity during the design stage. A grouping genetic algorithm (GGA) is then employed for clustering the components and crossover mechanisms are modified according to the need of modular design. Furthermore, a reasonable green modular design evaluation is conducted using the green material cost analysis. According to the results, adjusted design proposals are suggested and materials that cause less pollution are recommended to replace the components with pollution values higher than those in the module. Finally, the authors use Borland C++ 6.0 to evaluate the system and clustering method. To illustrate the methodology proposed in this study, a table lamp is offered as an example. � 2007 Elsevier Ltd. All rights reserved. Keywords: Modular design; Green life-cycle engineering; Grouping genetic algorithm; Liaison intensity; Green analysis 1. Introduction Not until recent years have people realized the impor- tance of environmental protection. People pay more atten- tion to the environment they are living in and the way people deal with the limited resources. Among resource dis- posal methods, recycling and garbage classification are two methods widely applied. These ways, however, are passive methods in the launch, usage and damage of products. Moreover, fierce market competition is shortening the product life cycle and the passive resource recycling approach can no longer cope with the ever-increasing bur- den current products have on the environment. Therefore, it is important to maximize the usage percentage of resources and minimize the damage to the environment in the early product design stage. This kind of more aggres- sive design tendency is referred to as green life-cycle engi- neering design (Otto & Wood, 2001; Tseng & Chen, 2004). The so-called product life cycle refers to the total amount of time from material, manufacturing, assembly, consumer use, and final disposal or recycle of a product, and green life cycle is mainly determined by the last two stages, product use, disposal or recycle. While the use of a product will affect its life span, the disposal and recycle of a product will definitely affect the environment and the resource availability. To prolong the product’s life cycle and to make the most of resources, the end of the product life cycle does not imply the disposal of the components. Instead, we need to solve the problem from the root of the enterprise activities, especially from the R&D of new products (Tseng & Chen, 2004). Many researchers have explored the issue from different points of view, such as design for environment (DFE), design for recy- cling (DFR), and design for disassembly (DFD) (Güngör 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.04.018 * Corresponding author. Tel.: +886 4 23924505x6012; fax: +886 4 23921742. E-mail address: hwai_en@seed.net.tw (H.-E. Tseng). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 34 (2008) 2524–2537 Expert Systems with Applications mailto:hwai_en@seed.net.tw & Gupta, 1999; Lambert, 2003). Due to the fact that well- designed modular structures can improve product life-cycle activities, modularity plays a more important role than the whole product life-cycle approach. For example, not only will common modules increase the chances of efficient reuse and recycling operation, they also feature ease of upgrade and maintenance, ease of product diagnosis, repair, and disposal, and so on. Taking green life cycle into consideration, the authors attempted to apply the green modular concept to product design. Advantages for this study are listed as follows (Gu & Sosale, 1999; Otto & Wood, 2001; Zhang & Ger- shenson, 2003): (1) Reexamination of product functions and specifica- tions ensures that the goal of environmental protec- tion can be achieved. (2) Reduction in product assembly time can enhance the efficiency of production. (3) Products or product components can be recycled, reused and disposed of more easily. (4) The life-cycle cost estimation enables designers to bring product cost into control. There are different perspectives with respect to measur- ing the product modularity. Jose and Tollenaere (2005) had made a comprehensive review regarding the modular design issue. In Section 2, the different viewpoints will be discussed. In the past, most conceptual descriptions have been rendered regarding the green-oriented modular study but a scientific methodology is rarely seen. In our study, a new methodology of green-oriented modular design will be proposed in Section 3. The approach comprises the follow- ing three parts: (1) Clarifying the liaison intensity of components: in addition to clarifying the liaison relationships of com- ponents through visualized diagrams, the liaison intensity of components is decided by their engineer- ing attributes. (2) The clustering algorithm: the goal of clustering is to assign the components whose liaison intensities are stronger in the same module. In this way, the liaison intensities among different modules are relatively weaker, indicating that it is easy to connect the com- ponents if they are assigned to the same module. (3) Green pollution and cost analysis: while changing the design specification, designers need to green pol- lution take and costs into consideration so that they can work out the proper design ideas in accordance with the material property to fulfill the product functions. These three parts will be discussed in Sections 4–6 respectively. In Section 7, the design of a table lamp will be used to illustrate the methodology proposed in this study. Finally, conclusions are made in Section 8. 2. Background of product modularity 2.1. Diverse viewpoints In the descriptive model of product functions, according to Otto and Wood (2001), modules can be defined as the integral physical structures corresponding to specific prod- uct functions. Meanwhile, the product function model is closely related to the customer’s needs. Therefore, a proper modular design is able to reduce the production cost and assemble components effectively into new products to cope with the rapid change of customer’s needs. The approach to meeting the customer’s needs and function specifications is called the function-based modular design. In the past, different methods had been proposed to explore the func- tion-based modular design issue; for example, Huang and Kusiak (1998) and Kreng and Lee (2004). In the assembly-based modular design method, products are generally described by liaison graph proposed by De Fazio and Whitney (1987). Researchers need to deal with modules on the basis of network partition and analysis the subassemblies or modules from the stability viewpoint. Lee (1994) and Tseng, Chang, and Yang (2004) are typical representation for this approach of assembly-based modu- lar design. In general, the manufacturing-based idea does not cover the algorithm for the formation of product modules. Emphasis is often placed upon smooth connections between the design and manufacturing phases. In this domain, He and Kusiak (1996) and Kahoo and Situmdrang (2003) used manufacturing time as the criteria for efficiency evaluation. On the other hand, the traditional low-cost and mass production model has difficulty in meeting the require- ments of the contemporary era. Mass customization aims to provide customer satisfaction with increasing variety and customization without a corresponding increase in cost and lead time. Enterprises have to cope with the frequent variation of product specifications by a stock strategy of bigger numbers of components and modules in the custom- ized environment. Therefore, the exploration of modular design will help with the production and control of mass customization. In terms of mass customization, Mikkola and Gassmann (2003) and Fujita and Yosshida (2004) offered a valuable evaluation method for this issue. According to the green life-cycle-based concept, modu- lar design is focused upon the environmental level. New- comb, Bras, and Rosen (1998) used group techniques to develop modular design; Gu and Sosale (1999) used the simulated annealing algorithms to explore modular design; Qian and Zhang (2003) proposed an environmental analy- sis model for achieving the modular goal. 2.2. Problem formulation As mentioned earlier, the paper attempts to focus on the green life cycle of product design. Three different problems should be taken into consideration in green modular H.-E. Tseng et al. / Expert Systems with Applications 34 (2008) 2524–2537 2525 https://isiarticles.com/article/1508 
work_oi5ldfwr25f2lm4vqqtv6sk5h4	----	1-s2.0-S0957417414002164-main.pdf MIFS-ND: A Mutual Information-based Feature Selection Method N. Hoquea,!, D. K. Bhattacharyyaa,!, J. K. Kalitab,! aDepartment of Computer Science & Engineering, Tezpur University Napaam, Tezpur-784028, Assam, India bDepartment of Computer Science, University of Colorado at Colorado Springs CO 80933-7150, USA Abstract Feature selection is used to choose a subset of relevant features for e!ective classification of data. In high dimensional data classification, the performance of a classifier often depends on the feature subset used for classification. In this paper, we introduce a greedy feature selection method using mutual information. This method combines both feature-feature mutual information and feature- class mutual information to find an optimal subset of features to minimize redundancy and to maximize relevance among features. The e!ectiveness of the selected feature subset is evaluated using multiple classifiers on multiple datasets. The performance of our method both in terms of classification accuracy and execution time performance, has been found significantly high for twelve real-life datasets of varied dimensionality and number of instances when compared with several competing feature selection techniques. Keywords: Features, mutual information, relevance, classification 1. Introduction Feature selection, also known as variable, attribute, or variable subset selection is used in ma- chine learning or statistics for selection of a subset of features to construct models for describing data [1, 2, 3, 4, 5]. Two important aspects of feature selection are: (i) minimum redundancy and (ii) maximum relevance [6]. Besides these, people use feature selection for dimensional- ity reduction and data minimization for learning, improving predictive accuracy, and increasing comprehensibility of models. To satisfy these requirements, two dimensionality reduction ap- proaches are used, i.e., feature extraction and feature selection [7]. A feature selection method selects a subset of relevant features from the original feature set, whereas a feature extraction method creates new features based on combinations or transformations of the original feature set. Feature selection is used to overcome the curse of dimensionality [8] in a pattern recognition !Corresponding authors Email addresses: tonazrul@gmail.com (N. Hoque), dkb@tezu.ernet.in (D. K. Bhattacharyya), jkalita@uccs.edu (J. K. Kalita) 1Nazrul Hoque is a Senior Research Fellow in the Department of Computer Science and Engineering, Tezpur University, Napaam, Tezpur, Assam, India. 2Dhruba Kr. Bhattacharyya is a professor in the Department of Computer Science and Engineering, Tezpur University, India. 3Jugal K. Kalita is a professor in the Department of Computer Science, University of Colorado at Colorado Springs, USA. Preprint submitted to Expert systems with applications (Elsevier) April 21, 2014 system. The main objective of feature selection is to identify m most informative features out of the d original features, where m<d. In the literature, feature selection approaches are widely used prior to or during classification of data in pattern recognition and data mining in fields as diverse as bioinformatics and network security. Feature selection approaches are classified into four categories, such as filter approach, wrapper approach, embedded approach, and hybrid approach. 1. Filter approach [9]: This approach selects a subset of features without using a learning algorithm. It is used in many datasets where the number of features is high. Filter-based feature selection methods are faster than wrapper-based methods. 2. Wrapper approach [10]: This approach uses a learning algorithm to evaluate the accuracy produced by the use of the selected features in classification. Wrapper methods can give high classification accuracy for particular classifiers, but generally they have high computational complexity. 3. Embedded approach [9]: This approach performs feature selection during the process of training and is specific to the applied learning algorithms. 4. Hybrid approach [11]: This approach is a combination of both filter and wrapper-based methods. The filter approach selects a candidate feature set from the original feature set and the candidate feature set is refined by the wrapper approach. It exploits the advantages of these two approaches. Feature selection plays an important role in network anomaly detection. In a network anomaly detection system, anomalies are identified in a network by monitoring the behavior of normal data compared to abnormal ones. The detection system identifies an attack based on behavioral analysis of features of network tra"c data. Network tra"c data objects may contain protocol specific header fields, such as source address, source port, destination address, destination port, protocol type, flags and time to live. Not all this information is equally important to detect an attack. In addition, if the number of features is high, the classifier usually performs poorly. So, the selection of an optimum and relevant set of features is important for the classifier to provide high detection accuracy with low computational cost. 1.1. Applications of Feature Selection Feature selection is an important step in most classification problems to select an optimal subset of features to increase the classification accuracy and improve time needed. It is widely used in many applications of data mining and machine learning, network anomaly detection and natural language processing. 1.1.1. In Data Mining and Machine Learning Machine learning is appropriate when a task is defined by a series of cases or examples rather than in terms of an algorithm or rules. Machine learning is useful in many fields including 2 robotics, pattern recognition and bioinformatics. In learning a classifier, a predictor or a learning algorithm is used to extract information from the behavior of the features of a data object. From the acquired information, the classifier can predict the class label of a new data object. So, a feature selection algorithm is used to find an optimal features set that can be used by the predictor to generate maximal information about the class label of the data object. 1.1.2. In Network Anomaly Detection Anomaly detection in real-time is a challenging problem for network security researchers and practitioners. A network packet contains a large number of features and hence, an anomaly detec- tion system takes a significant amount of time to process features to detect anomaly packets [12]. To overcome the problem we need a feature selection method that can identify the most relevant features from a network packet. The selected features are used by an intrusion detection system to classify network packets either as normal or anomalous. The accuracy, detection time and e!ectiveness of the detection system depend on the input feature set along with other hardware constraints. Therefore, feature selection methods are used to determine a minimal feature set which is optimal and does not contain redundant features. 1.1.3. In Text Categorization In the recent past, content-based document management tasks have become important. Text categorization assigns a boolean value to a document d and the value determines whether the document belongs to a category c or not. In document classification, appearance of any word in a document may be considered a feature. Feature sets used reflect certain properties of the words and their context within the actual texts. A feature set with most relevant features can predict the category of the document quickly. In automatic text categorization, the native feature space contains unique terms (words or phases) of the document which can number in tens, hundreds or thousands. As a result, operation cost for text categorization is not only time consuming but also intractable in many cases [13]. Therefore, it is highly desirable to reduce the feature space for e!ective classification of a document. 1.1.4. In Gene Expression Data Mining Gene analysis is a topic of great interest associated with specific diagnosis in microarray study. Microarrays allow monitoring of gene expression for thousands of genes in parallel and produce enormous valuable data. Due to large dimension and over-fitting problem [14] of gene expression data, it is very di"cult to obtain a satisfactory classification result by machine learning techniques. Also, discriminant analysis is used in gene expression data analysis to find e!ective genes responsible for a particular disease. These facts have given rise to the importance of feature selection techniques in gene expression data mining [15]. Gene expression data is typically high dimensional and is error prone. As a result, feature selection techniques applied in gene expression data analysis especially classification techniques. Feature selection has also been applied as a preprocessing task in clustering gene expression data in search of co-expressed patterns. 3 1.2. Contribution The main contribution of this paper is a mutual information-based feature subset selection method to use for complex time series data classification such as network anomaly detection, pattern classification. The method has been established to perform satisfactorily both in terms of classification accuracy and execution time for a large number of real life benchmark datasets. The e!ectiveness of the method is established in terms of classification accuracy exhibited for several real-life intrusion, text categorization, gene expression and UCI datasets, in association with some well-known classifiers. The rest of the paper is organized as follows. In Section 2, we explain related work in brief. Section 3 formulates the problem. The concept of mutual information and the proposed method are discussed in Section 4. Experimental results are shown in Section 5. Finally, conclusion and future work are discussed in Section 6. 2. Related Work Many feature selection algorithms [16], [17], [18], [19], [20], [21] have been proposed for clas- sification. The common approach for these algorithms is to search for an optimal set of features that provides good classification result. Most feature selection algorithms use statistical measures such as correlation and information gain or a population-based heuristic search approach such as particle swarm optimization, ant colony optimization, simulated annealing and genetic algo- rithms. An unsupervised feature subset selection method using feature similarity was proposed by Mitra et al. [22] to remove redundancy among features. They use a new measure called maximal information compression index to calculate the similarity between two random variables for feature selection. Bhatt et al. [23] use fuzzy rough set theory for feature selection based on natural properties of fuzzy t-norms and t-conorms. A mutual information-based feature selection algorithm called MIFS is introduced by Battiti [24] to select a subset of features. This algorithm considers both feature-feature and feature-class mutual information for feature selection. It uses a greedy technique to select a feature subset that maximizes information about the class label. Kwak and Cho [25] develop an algorithm called MIFS-U to overcome the limitations of MIFS to obtain better mutual information between input features and output classes than MIFS. Peng et al. [26] introduce a mutual information based feature selection method called mRMR (Max- Relevance and Min-Redundancy) that minimizes redundancy among features and maximizes de- pendency between a feature subset and a class label. The method consists of two stages. In the first stage, the method incrementally locates a range of successive feature subsets where a consistent low classification rate is obtained. The subset with the minimum error rate is used as a candidate feature subset in stage 2 to compact further using a forward selection and backward elimination based wrapper method. Estevez et al. [27] propose a mutual information-based fea- ture selection method as a measure of relevance and redundancy among features. Vignolo et al. [28] introduce a novel feature selection method based on multi-objective evolutionary wrappers 4 Table 1: Symbols used in our method Symbols Meaning Symbols Meaning D Dataset d Dimension of dataset F Original feature set F ! Optimal feature set C Class label k Number of features in F ! fi Feature no i fj Feature no j Cd Domination count Fd Dominated count MI Mutual Information FFMI Feature Feature Mutual Information FCMI Feature Class Mutual Information AFFMI Average Feature Feature Mutual Information using genetic algorithm. To support e!ective data classification, several attribute evaluation techniques, such as ReliefF [29], Chi squared [30], Correlation Feature Selection (CFS) [31] and Principal components analysis [32], also have been provided in the Weka software platform [33]. These techniques employ dif- ferent search techniques such as, BestFirst, ExhaustiveSearch, GreedyStepwise, RandomSearch, and Ranker for feature ranking. To describe the algorithms in this paper, we use symbols and notations given in Table 1. 2.1. Motivation Feature selection or attribute selection is an important area of research in knowledge discovery and data mining. Due to rapid increase in the availability of datasets with numerous data types, an e!ective feature selection method is absolutely necessary for classification of data in high dimensional datasets. In many application domains, such as network security or bioinformatics, feature selection plays a significant role in the classification of data objects with high classification accuracy. Although a large number of classification and feature selection techniques have been used in the past, significant reduction of false alarms, especially in the network security domain remains a major and persistent issue. Decrease in the number of irrelevant features leads to reduction in computation time. This has motivated us to design a mutual information based feature selection method to identify an optimal subset of features which gives the best possible classification accuracy. 3. Problem Formulation For a given dataset D of dimension d, with a feature set F = {f1, f2, · · · , fd}, the problem is to select an optimal subset of relevant features F " where (i) F " ! F and (ii) for F ", a classifier gives the best possible classification accuracy. In other words, we aim to identify a subset of features where for any pair (fi, fj) " F ", the feature-feature mutual information is minimum and feature-class mutual information is maximum. 4. Proposed Method The proposed method uses mutual information theory to select a subset of relevant features. The method considers both feature-feature and feature-class mutual information to determine an optimal set of features. 5 4.1. Mutual Information In information theory, mutual information I(X; Y ) is the amount of uncertainty in X due to the knowledge of Y [34, 35]. Mathematically, mutual information is defined as I(X; Y ) = ! x,y p(x, y)log p(x, y) p(x)p(y) (1) where p(x, y) is the joint probability distribution function of X and Y , and p(x) and p(y) are the marginal probability distribution functions for X and Y . We can also say I(X; Y ) = H(X) # H(X|Y ) (2) where, H(X) is the marginal entropy, H(X|Y ) is the conditional entropy, and H(X; Y ) is the joint entropy of X and Y . If H(X) represents the measure of uncertainty about a random variable, then H(X|Y ) measures what Y does not say about X. This is the amount of uncertainty in X after knowing Y and this substantiates the intuitive meaning of mutual information as the amount of information that knowing either variable provides about the other. In our method, a mutual information measure is used to calculate the information gain among features as well as between feature and class attributes. Using a greedy manner, each time we pick a feature from the feature set still left one that provides maximum information about the class attribute with minimum redundancy. 4.2. Our Feature Selection Method Initially, we compute feature-class mutual information and select the feature that has the highest mutual information. The feature is then put into the selected feature subset and removed from the original feature set. Next, for each of the non-selected features, we compute the feature- class mutual information and then calculate the average feature-feature mutual information for each of the selected features. At this point, each non-selected feature contains feature-class mutual information and average feature-feature mutual information. From these calculated values, it selects a feature that has the highest feature-class mutual information, but minimum feature- feature mutual information using an optimization algorithm known as Non-dominated Sorting Genetic Algorithm-II (NSGA-II) [36] with domination count Cd and dominated count Fd for feature-class and feature-feature values, respectively. The domination count of a feature represents the number of features that it dominates for feature-class mutual information and dominated count represents the number of features that it dominates for feature-feature mutual information. We select the feature that has the maximum di!erence of domination count and dominated count. 6 Table 2: Domination count of a feature feature feature Average feature Domination Dominated Cd # Fd class MI -feature MI count Cd count Fd f1 0.54 0.17 2 1 1 f2 0.76 0.09 3 0 3 f4 0.17 0.78 0 3 -3 f5 0.33 0.27 1 2 -1 4.2.1. Benefits of Competitive Ranking The proposed feature selection method selects features based on a ranking procedure using the NSGA-II algorithm. It selects a feature that has the highest feature-class mutual information but minimum feature-feature mutual information, which requires solving an optimization problem. The soundness of the proposed method is derived from the fact that it selects a feature which is either strongly relevant or weakly redundant (explained in Example 1 and Example 2). A strongly relevant feature has high feature-class mutual information but low feature-feature mutual information and a weakly relevant feature has high feature-class mutual information but medium feature-feature mutual information [37]. In addition, the method handles the tie condition (i.e., if two features have the same rank) by selecting the feature that has high feature-class mutual information. 4.2.2. Example 1 Let F = {f1, f2, f3, f4, f5} be a set of five features. First, we compute feature-class mutual information for every feature and select the feature that has the maximum mutual information value. Let us assume feature f3 has the highest mutual information value and hence, f3 is removed from F and is put in the optimal feature set, say F ". Next, for a feature fj " F , we compute feature-feature mutual information with every feature fi " F " and store the average mutual information value as average feature-feature mutual information for fj. This way, we compute feature-feature mutual information for f1, f2, f4 and f5. Again, we compute feature-class mutual information for f1, f2, f4 and f5. Consider the scenario shown in Table 2. Here, feature f2 has the maximum di!erence between Cd and Fd, i.e., 3. Hence feature f2 will be selected. In case of tie for the values of (Cd - Fd), we pick the feature that has maximum feature-class mutual information. This procedure is continued until we get a subset of k features. 4.2.3. Example 2 For the situation shown in Table 3, the method selects the weakly relevant feature f1 instead of feature F5. Though the two features have the same Cd - Fd value, F1 has higher feature-class mutual information but medium feature-feature mutual information. In such a tie situation, the other methods may select either F1 or F5. According to Peng’s [26] and Battiti’s [24] method a tie condition for non-selected features X1, X2 and a selected features set Xk with class label Y is given by the following. I(X1; Y ) # ! "n#1 k=1 I(X1; Xk) = I(X2; Y ) # ! "n#1 k=1 I(X2; Xk) I(X1; Y ) + "n#1 k=1 I(X2; Xk) = I(X2; Y ) + "n#1 k=1 I(X1; Xk) 7 Table 3: Domination count of a feature feature feature Average feature Domination Dominated Cd # Fd class MI -feature MI count Cd count Fd f1 0.96 0.43 3 1 2 f2 0.72 0.45 0 2 -2 f4 0.85 0.78 1 3 -2 f5 0.91 0.10 2 0 2 if I(X1; Y )>I(X2; Y ) $ "n#1 k=1 I(X2; Xk)< "n#1 k=1 I(X1; Xk) In this scenario, the proposed method selects the feature X1, since its feature-class mutual infor- mation is higher than X2. Definition 1. Feature relevance: Let F be a full set of features, fi a feature and Si = F #{fi}. Feature fi is strongly relevant i! I(C|fi, Si) %= I(C|Si) otherwise if I(C|fi, Si) = I(C|Si) and &S"i ' Si such that I(C|fi, Si) %= I(C|Si), then fi is weakly relevant to the class C. Definition 2. Feature-class relevance: It is defined as the degree of feature-class mutual in- formation for a given class Ci of a given dataset D, where data elements described by d features. A feature fi " F ", is an optimal subset of relevant features for Ci, if the relevance of (fi, Ci) is high. Definition 3. Relevance score: The relevance score of a feature fi is the degree of relevance in terms of mutual information between a feature and a class label. Based on this value, a rank can be assigned to each feature fi, (i = 1, 2, 3, · · · , d. For a given feature fi " F ", the relevance score is high. The following properties are trivial based on [26] and [38] Property 1: For a pair of features (fi, fj) " F ", the feature-feature mutual information is low, while feature-class mutual information is high for a given class Ci of a given dataset D. Explanation: For a feature fi to be a member of an optimal subset of relevant features F " for a given class Ci, the feature-class mutual information or the relevance score (Definition 3) must be high. On the contrary, whenever feature-class mutual information is high, feature-feature mutual information has to be relatively low [38]. Property 2: A feature f"i /" F " has low relevance w.r.t. a given class Ci of a given dataset D. Explanation: Let f"i /" F " where F " corresponds to class Ci and let feature-class mutual informa- tion between f"i and Ci be high. If the feature-class mutual information score is high, then as per Definitions 1 and 2, f"i must be a member of set F ", which contradicts. Property 3: If a feature fi has higher mutual information than feature fj with the class label C, then fi will have a smaller probability of misclassification [39]. Explanation: Since fi has higher mutual information score thanfj corresponding to class C, it has more relevance. So, the probability of classification accuracy using fi as one of the feature will be higher than fj. Lemma 1: For a feature fi, if the domination count Cd is larger and the dominated count Fd is 8 smaller than all other features fj, (i %= j), then the feature fi has the highest feature-class mutual information and has more relevant. Proof: For a feature fi, if its domination count Cd is larger and the dominated count Fd is smaller than all other features fj, (i %= j) then NSGA-II method ensures that feature-class mutual infor- mation for Fi is the highest whereas the feature-feature mutual information value is the lowest. Hence, the method selects feature fi as the strongly relevant feature as shown in Example 1. Lemma 2: For any two features, fi and fj, if the di!erence of Cd and Fd for feature fi is the same as the di!erence of Cd and Fd for feature fj, then the feature with the highest feature-class mutual information is relevant. Proof: For any two features fi and fj, if the di!erence of Cd and Fd for feature fi is as same as the di!erence of Cd and Fd for feature fj, then NSGA-II ensures that neither fi nor fj satisfies the Lemma 1 and in this situation either fi or fj has higher feature-class mutual information. Hence, the method selects a feature that has higher feature-class mutual information as shown in Example 2. 4.3. MIFS(MI based Feature Selection) method Input: d, the number of features; dataset D; F = {f1, f2, · · · , fd}, the set of features Output: F ", an optimal subset of features Steps: for i=1 to d, do Compute MI(fi, C) end Select the feature fi with maximum MI(fi, C) F " = F " ) {fi} F = F # {fi} count=1; while count <= k do for each feature fj " F, do FFMI=0; for each feature fi " F ", do FFMI=FFMI+compute FFMI(fi,fj) end AFFMI=Average FFMI for feature fj. FCMI=Compute FCMI(fj, C) end Select the next feature fj that has maximum AFFMI but minimum FCMI F " = F " ) {fi} F = F # {fj} i = j count=count+1; end Return features set F " Algorithm 1: MIFS-ND The proposed feature selection method depends on two major modules, namely Compute FFMI and Compute FCMI. We describe working of each of these modules next. Compute FFMI(fi,fj): For any two features fi, fj " F , this module computes mutual informa- tion between them, i.e.; fi and fj using Equation 1. It computes marginal entropy for variable fi 9 and computes mutual information by subtracting conditional entropy of fi for the given variable fj from marginal entropy. Compute FCMI(fj, C): For a given feature fj " F and a given class label say C, this module is used to find mutual information between fj and C using Shannon’s mutual information formula using equation 1. First, it computes marginal entropy for the variable fj and then subtracts conditional entropy of fj for the given variable C. Using these two modules the proposed method picks up a high ranked feature which is strongly relevant but non-redundant. To select a strongly relevant but non-redundant feature it uses NSGA-II method and computes the domination count (Cd) and dominated count (Fd) for every feature. If a feature has the highest di!erence between Cd and Fd then it selects that feature using Lemma 1 otherwise it uses Lemma 2 to select the relevant feature. 4.4. Complexity Analysis The overall complexity of the proposed algorithm depends on the dimensionality of the input dataset. For any dataset of dimension d, the computational complexity of our algorithm to select a subset of relevance features is O(d2). However, the use of appropriate domain specific heuristics can help to reduce the complexity significantly. 4.5. Comparison with other Relevant Work The proposed feature selection method di!ers from other methods in the following manners. 1. Like Battiti [24], our method considers both feature-feature and feature-class mutual in- formation, but Battiti uses an additional input parameter called ! to regulate the relative importance of mutual information between a feature to be selected and the features that have already been selected. Instead of the input !, our method calculates the domina- tion count and dominated count, which are used in NSGA-II algorithm to select a relevant feature. 2. Like Kwak and Choi [25], our method uses a greedy filter approach to select a subset of features. However, like them our method does not consider joint mutual information among three parameters, such as a feature that is to be selected, a set of features that are already selected and the class output. 3. Unlike Peng et al. [26], who use both filter and wrapper approaches, our method uses only the filter approach to select an optimal feature subset using a ranking statistic, which makes our scheme more computationally cost e!ective. Similar to the Peng’s mRMR method, we use the maximum relevance and minimum redundancy criterion to select a feature from the original feature set. In the subsequent section (Section 5) a detailed comparison of our method with Peng et al.’s method for two high dimensional gene expression datasets and four UCI datasets has been shown. Unlike Brown [40], our method uses Shannon’s mutual information on two variables only but Brown’s method uses multivariate mutual information. Also, our method selects a feature set using a heuristic bottom-up approach 10 and iteratively inserts a relevant but non-redundant feature into the selected feature set. Brown’s method follows a top-down approach and discards features from the original feature set. 4. Unlike Kraskov et al. [35], our method computes Shannon’s entropy whereas they compute entropy using k-nearest neighbor distance. 5. Experimental Result Experiments were carried out on a workstation with 12 GB main memory, 2.26 Intel(R) Xeon processor and 64-bit Windows 7 operating system. We implement our algorithm using MATLAB R2008a software. 5.1. Dataset Description During our experimental analysis, we use several network intrusion, text categorization, a few selected UCI and gene expression datasets. These datasets contain both numerical and categorical values with various dimensionalities and numbers of instances. Descriptions of the datasets are given Table 4. Table 4: Dataset description Dataset Number of instances Number of attributes Intrusion NSL-KDD 99 125973 42 Dataset 10% KDD 99 494021 42 Corrected KDD 311029 42 Wine 178 13 Non-intrusion Monk1 432 6 Dataset Monk2/ Monk3 432 6 IRIS 150 4 Ionosphere 351 34 Sonar 209 61 Text categorization Bloggender-male 3232 101 Bloggender-female 3232 101 Gene Expression Lymphoma 45 4026 Colon Cancer 62 2000 5.2. Results The proposed algorithm first selects a subset of relevant features from each dataset. To eval- uate the performance of our algorithm we use four well known classifiers, namely, Decision Trees, Random Forests, K-Nearest Neighbor (KNN) and Support Vector Machines (SVM). The perfor- mance of our algorithm is compared with five standard feature selection algorithms, namely Chi squared method, Gain Ratio, Information Gain, ReliefF, and Symmetric Uncertainty, which are already available in Weka. We use 10-fold cross validation to evaluate the e!ectiveness of selected features using di!erent classifiers. The comparison of classification accuracy of our algorithm with all the aforesaid feature selection algorithms is shown here for all the mentioned datasets. From the experimental results, we observe that the proposed method gives good classification accuracy on most datasets. Since, the network datasets are very large to be handled in MATLAB, we split the network datasets into smaller partitions which contain class labels from all the classes. 11 Finally, we compute the average classification accuracy of all the partitioned datasets. The proposed feature selection method selects an optimal subset of features for which we achieve the best classification accuracy. However, the cardinality of a feature subset identified by the proposed MIFS-ND method varies for di!erent datasets. This variation in cardinalities of the optimal feature subsets for some of the datasets, namely, Sonar, Ionosphere, Wine and Iris is shown in Figures 3(a), 3(b), 4(a) and 4(b), respectively. Our observation is that (i) except for the Sonar dataset, the proposed MIFS-ND method performs significantly better for other datasets than the competing methods. The performance of our method su!ers in case of Sonar dataset due to lack of su"cient training instances, and (ii) with the increase in size of the original feature sets of di!erent datasets, the cardinality for optimal feature subset identified by MIFS-ND also increases. For example, in case of the UCI datasets, when the size of feature set varies from 4 to 61 (as shown in Table 4) the cardinality of the optimal feature subset varies from 3 to 6, whereas in case of the intrusion and text categorization datasets, the cardinality of the feature subsets varies from 5 to 10 due to the increased size of feature sets. Figure 1 establishes this fact for 12 datasets. Figure 1: Optimal range of the size of feature subsets for di!erent datasets 5.3. Discussion Feature selection is an unavoidable part for any classification algorithm. We use mutual information theory to select a subset of features from an original feature set based on feature- feature and feature-class information values. We evaluate our feature selection method using network security, text categorization, a few selected UCI and gene expression datasets. From experimental analysis, we observe that for most datasets, the classification accuracy is much better for a subset of features compared to when using the full feature set. Due to significant reduction of the number of features, better computational e"ciency is also achieved. As shown in Figure 2, the computational time increases significantly for Random forests but computational 12 time is constant for SVMs. The computational time is relatively increased for Decision Trees and KNN classifiers due to the increase in the number of features. Figure 2: Execution time performance for di!erent classifiers 5.3.1. Performance Analysis on Intrusion Datasets In case of network datasets, we found high classification accuracy for all the classifiers with the 10% KDD dataset as shown in Figures 6(e) to 6(h). We found better classification accuracy using Decision Trees, Random Forests and the KNN classifier than the SVM on the corrected KDD dataset for the proposed algorithm as shown in Figures 6(a) to 6(d). Similarly, as shown in Figures 6(i) to 6(l), our algorithm produces better result on the NSL KDD dataset when k * 6, where k is the number of features in a subset and when k + 5, the classification accuracy is relatively lower. 5.3.2. Performance Analysis on UCI Datasets We perform experiments to analyze the classification accuracy of our proposed MIFS-ND method on UCI datasets. With the Wine dataset, Decision Trees, Random Forests and the KNN classifier give better classification accuracy for the proposed method but the accuracy is a bit lower with the SVM classifier, as plotted in Figures 5(a), 5(b), 5(c) and 5(d), respectively. As shown in Figures 5(e) to 5(t), the used classifiers show high classification accuracy with Monk1, Monk2, Monk3 and Iris datasets. We also compare the performance of our method MIFS-ND with MIFS [24] and MIFS-U [25] for Sonar and Ionosphere datasets as shown in Figure 3(a) and Figure 3(b), respectively. The comparison shows that the proposed algorithm gives better classification accuracy with these two UCI datasets. 5.3.3. Performance Analysis on Text Categorization Datasets The proposed feature selection method is also applied to text datasets to evaluate its per- formance. From experimental results, we observe that the Bloggender female and male datasets 13 show a bit poorer classification accuracy for Decision Trees as shown in Figures 7(a) and 7(d), respectively. But, the method shows almost equal classification accuracy for Random forests and KNN as plotted in Figures 7(b), 7(c), 7(e) and 7(f), respectively. 5.3.4. Performance Analysis on Gene Expression Datasets The proposed MIFS-ND method is also tested on two high-dimensional gene expression datasets, namely, Lymphoma and Colon Cancer. The classification accuracy on the Lymphoma dataset is very good for all the four classifiers as shown in Figures 8(a)-8(d). Compared to mRMR, MIFS-ND gives high classification accuracy for all four classifiers. However, the accuracy is com- promised from feature number 7 for SVM classifier only. Similarly, in Colon Cancer dataset, Random forests give high classification accuracy for MIFS-ND as shown in Figure 8(f) compared to other feature selection methods. But KNN and SVM give high classification accuracy for the mRMR method as plotted in Figure 8(g) and Figure 8(h), respectively. In case of Decision Tree, classification accuracy for mRMR and MIFS-ND is almost similar as shown in Figure 8(e). (a) Ionosphere dataset (b) Sonar dataset Figure 3: Comparison of MIFS-ND with MIFS, MIFS-U and mRMR for Ionosphere and Sonar datasets (a) Wine dataset (b) Iris dataset Figure 4: Comparison of MIFS-ND with Entropy measure, GINI info, ICA-MI, Local ICA-MI and mRMR 5.4. Evaluation of Classification Performance based on t-Test We use the McNemar’s test [41] to evaluate the performance of our method in terms of clas- sification accuracy using four di!erent classifiers. According to McNemar’s test, two algorithms 14 can have four possible outcomes such as TT, TF, FT and FF. Here, T and F stand for true and false respectively. The test computes a value called z-score, defined in Equation 3 that measures how statistically significant the results are. The arrows (,, -) indicate which classifier performs better with a given dataset. z = (|NT F # NF T | # 1). NT F + NF T (3) where, N= number of instances in a dataset. From the results of t-test, it can be observed that for Wine dataset, the Decision Trees performance is better than KNN and SVM but Random Forests perform better than Decision Trees for all the feature selection methods as shown in Table 5. In Monk1 dataset, KNN and Random Forests, both give better performance than Decision Trees, whereas Decision trees performance is better than SVM as shown in Table 6. In Monk2, Monk3 and Ionosphere dataset, Decision Trees classifier show better performance compared to KNN and SVM whereas Random Forests performance is better than Decision Trees as shown in Table 7, 8 and 11. As given in Table 9, Decision Tree performance is better than SVM but KNN and Random Forests show better performance than Decision Trees. In Sonar dataset, the performance of KNN and Random Forests is better than Decision Trees whereas Decision Trees performance is better than that of SVM as shown in Table 10. Table 5: t-test for Wine dataset Dataset Feature selection KNN RF SVM method Wine ChiSquar DT $ 4.3410 1.5539 % $ 8.7818 GainRatio $ 3.1201 2.2159 % $ 8.6927 InfoGain $3.4758 1.3790 % $ 8.8202 ReliefF $ 3.7773 2.4859 % $ 8.6646 SymmetricU $2.9723 1.9829% $ 8.4879 MIFS-ND $2.7097 2.6000 % $ 8.4917 Table 6: t-test for Monk1 dataset Dataset Feature selection KNN RF SVM method Monk1 ChiSquar DT 3.1888 % 4.2504 % $ 12.6320 GainRatio 3.2269 % 4.4338 % $ 12.5119 InfoGain $1.8995 1.8943% $10.7272 ReliefF 5.5811 % 5.5811 % $ 12.2406 SymmetricU 3.7420% 4.3032% $ 12.6151 MIFS-ND 2.7419 % 3.8916 % $ 12.4085 Table 7: t-test for Monk2 dataset Dataset Feature selection KNN RF SVM method Monk2 ChiSquar DT $ 2.3840 $ 1.0071 $ 11.0991 GainRatio $ 2.2699 $ 0.5800 $ 10.9840 InfoGain $ 1.7675 $ 1.2980 $ 10.9856 ReliefF 0.7857 % 0.2196% $ 11.2472 SymmetricU $ 2.8095 $1.4083 $ 11.2910 MIFS-ND 0.6572 % 0.5255 % $ 11.0219 15 Table 8: t-test for Monk3 dataset Dataset Feature selection KNN RF SVM method Monk3 ChiSquar DT $ 2.6415 $ 6.4908 $ 13.8191 GainRatio $ 2.4969 6.4944 % $ 13.9999 InfoGain $ 1.3687 6.0628 % $ 13.7995 ReliefF 0.6791 % 6.5169 % $ 13.7592 SymmetricU $ 2.3657 6.2318 % $ 13.8910 MIFS-ND $1.2637 $ 5.9339 $ 13.7014 Table 9: t-test for Iris dataset Dataset Feature selection KNN RF SVM method Iris ChiSquar DT 0.1215 % 1.5815 % $ 8.2768 GainRatio 0.1250 % 1.7166 % $ 8.2622 InfoGain 0.1768 % 1.6919 % $ 8.2618 ReliefF $ 0.1250 1.2877 % $ 8.3372 SymmetricU 0.6637 % 1.4462 % $ 8.3071 MIFS-ND 0.2754 % 1.6303 % $ 8.2772 Table 10: t-test for Sonar dataset Dataset Feature selection KNN RF SVM method Sonar ChiSquar DT 0.7063 % 1.3416 % $ 7.2530 GainRatio 1.6916 % 1.6897 % $ 7.5730 InfoGain 1.8612 % 1.4681 % $ 6.6444 ReliefF 1.4308 % 1.8329 % $ 7.0064 SymmetricU 1.8296 % 2.5176 % $ 6.9518 MIFS-ND 1.7365 % 1.8857 % $ 7.2041 Table 11: t-test for Ionosphere dataset Dataset Feature selection KNN RF SVM method Ionosphere ChiSquar DT $ 0.8715 1.4005 % $ 9.9160 GainRatio $ 0.8175 1.3189 % $ 9.8643 InfoGain $ 0.4200 1.6768 % $ 9.7896 ReliefF $ 0.7503 1.7440 % $ 9.8320 SymmetricU $ 0.6462 1.9885 % $ 9.7559 MIFS-ND 1.6916 % 1.6897 % $ 7.5730 6. Conclusion and Future Work In this paper, we have described an e!ective feature selection method to select a subset of high ranked features, which are strongly relevant but non-redundant for a wide variety of real-life dataset. To select high ranked features, an optimization criterion used in the NSGA-II algorithm is applied here. The method has been evaluated in terms of classification accuracy as well as execution time performance using several network intrusion datasets, a few UCI datasets, text categorization datasets and two gene expression datasets. We compare the classification accuracy for the selected features using Decision trees, Random Forests, KNN and SVM classifiers. The overall performance of our method has been found excellent in terms of both classification accuracy and execution time performance for all these datasets. Development of an incremental feature selection tool based on MIFS-ND is underway to support wide variety of application domains. Acknowledgment This work is supported by Department of Information Technology (DIT). The authors are thankful to the funding agency. 16 (a) Decision Trees accuracy WINE (b) Random Forests accuracy for WINE (c) KNN accuracy for WINE (d) SVM Accuracy for WINE (e) Decision Trees accuracy MONK1 (f) Random Forests accuracy for MONK1 (g) KNN accuracy for MONK1 (h) SVM Accuracy for MONK1 (i) Decision Trees accuracy MONK2 (j) Random Forests accuracy for MONK2 17 (k) KNN accuracy for MONK2 (l) SVM Accuracy for MONK2 (m) Decision Trees accuracy MONK3 (n) Random Forests accuracy for MONK3 (o) KNN accuracy for MONK3 (p) SVM Accuracy for MONK3 (q) Decision Trees accuracy IRIS (r) Random Forests accuracy for IRIS (s) KNN accuracy for IRIS (t) SVM Accuracy for IRIS 18 (u) Decision Trees accuracy Ionosphere (v) Random Forests accuracy for Iono- sphere (w) KNN accuracy for Ionosphere (x) Decision Trees accuracy Sonar (y) Random Forests accuracy for Sonar (z) KNN accuracy for Sonar Figure 5: Accuracy of di!erent classifiers found in non-intrusion datasets References [1] A. Arauzo-Azofra, J. L. Aznarte, J. M. Beńıtez, Empirical study of feature selection meth- ods based on individual feature evaluation for classification problems, Expert Systems with Applications 38 (7) (2011) 8170–8177. [2] M. Dash, H. Liu, Feature selection for classification, Intelligent data analysis 1 (3) (1997) 131–156. [3] K. Polat, S. Güneş, A new feature selection method on classification of medical datasets: Kernel f-score feature selection, Expert Systems with Applications 36 (7) (2009) 10367– 10373. [4] H. Liu, L. Yu, Toward integrating feature selection algorithms for classification and clustering, Knowledge and Data Engineering, IEEE Transactions on 17 (4) (2005) 491–502. [5] J. M. Cadenas, M. C. Garrido, R. Mart́ıNez, Feature subset selection filter–wrapper based on low quality data, Expert Systems with Applications 40 (16) (2013) 6241–6252. 19 [6] A. Unler, A. Murat, R. B. Chinnam, mr2pso: A maximum relevance minimum redundancy feature selection method based on swarm intelligence for support vector machine classifica- tion, Information Sciences 181 (20) (2011) 4625–4641. [7] D. D. Lewis, Feature selection and feature extraction for text categorization, in: Proceedings of the workshop on Speech and Natural Language, Association for Computational Linguistics, 1992, pp. 212–217. [8] G. Hughes, On the mean accuracy of statistical pattern recognizers, Information Theory, IEEE Transactions on 14 (1) (1968) 55–63. [9] I. Guyon, A. Elissee!, An introduction to variable and feature selection, The Journal of Machine Learning Research 3 (2003) 1157–1182. [10] A. L. Blum, P. Langley, Selection of relevant features and examples in machine learning, Artificial intelligence 97 (1) (1997) 245–271. [11] H.-H. Hsu, C.-W. Hsieh, M.-D. Lu, Hybrid feature selection by combining filters and wrap- pers, Expert Systems with Applications 38 (7) (2011) 8144–8150. [12] K.-C. Khor, C.-Y. Ting, S.-P. Amnuaisuk, A feature selection approach for network intrusion detection, in: Information Management and Engineering, 2009. ICIME’09. International Conference on, IEEE, 2009, pp. 133–137. [13] Y. Yang, J. O. Pedersen, A comparative study on feature selection in text categorization, in: ICML, Vol. 97, 1997, pp. 412–420. [14] J.-T. Horng, L.-C. Wu, B.-J. Liu, J.-L. Kuo, W.-H. Kuo, J.-J. Zhang, An expert system to classify microarray gene expression data using gene selection by decision tree, Expert Systems with Applications 36 (5) (2009) 9072–9081. [15] Y. Saeys, I. Inza, P. Larrañaga, A review of feature selection techniques in bioinformatics, bioinformatics 23 (19) (2007) 2507–2517. [16] R. Caruana, D. Freitag, Greedy attribute selection., in: ICML, Citeseer, 1994, pp. 28–36. [17] H. Frohlich, O. Chapelle, B. Scholkopf, Feature selection for support vector machines by means of genetic algorithm, in: Tools with Artificial Intelligence, 2003. Proceedings. 15th IEEE International Conference on, IEEE, 2003, pp. 142–148. [18] S.-W. Lin, K.-C. Ying, C.-Y. Lee, Z.-J. Lee, An intelligent algorithm with feature selection and decision rules applied to anomaly intrusion detection, Applied Soft Computing 12 (10) (2012) 3285–3290. [19] L. Yu, H. Liu, Redundancy based feature selection for microarray data, in: Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, ACM, 2004, pp. 737–742. 20 [20] D. K. Bhattacharyya, J. K. Kalita, Network Anomaly Detection: A Machine Learning Per- spective, CRC Press, 2013. [21] S. Nemati, M. E. Basiri, N. Ghasem-Aghaee, M. H. Aghdam, A novel aco–ga hybrid algorithm for feature selection in protein function prediction, Expert systems with applications 36 (10) (2009) 12086–12094. [22] P. Mitra, C. Murthy, S. K. Pal, Unsupervised feature selection using feature similarity, IEEE transactions on pattern analysis and machine intelligence 24 (3) (2002) 301–312. [23] R. B. Bhatt, M. Gopal, On fuzzy-rough sets approach to feature selection, Pattern recognition letters 26 (7) (2005) 965–975. [24] R. Battiti, Using mutual information for selecting features in supervised neural net learning, Neural Networks, IEEE Transactions on 5 (4) (1994) 537–550. [25] N. Kwak, C.-H. Choi, Input feature selection for classification problems, Neural Networks, IEEE Transactions on 13 (1) (2002) 143–159. [26] H. Peng, F. Long, C. Ding, Feature selection based on mutual information criteria of max- dependency, max-relevance, and min-redundancy, Pattern Analysis and Machine Intelligence, IEEE Transactions on 27 (8) (2005) 1226–1238. [27] P. A. Estévez, M. Tesmer, C. A. Perez, J. M. Zurada, Normalized mutual information feature selection, Neural Networks, IEEE Transactions on 20 (2) (2009) 189–201. [28] L. D. Vignolo, D. H. Milone, J. Scharcanski, Feature selection for face recognition based on multi-objective evolutionary wrappers, Expert Systems with Applications 40 (13) (2013) 12086–12094. [29] K. Kira, L. A. Rendell, A practical approach to feature selection, in: Proceedings of the ninth international workshop on Machine learning, Morgan Kaufmann Publishers Inc., 1992, pp. 249–256. [30] H. Liu, R. Setiono, Chi2: Feature selection and discretization of numeric attributes, in: Tools with Artificial Intelligence, 1995. Proceedings., Seventh International Conference on, IEEE, 1995, pp. 388–391. [31] M. A. Hall, L. A. Smith, Feature selection for machine learning: Comparing a correlation- based filter approach to the wrapper., in: FLAIRS Conference, 1999, pp. 235–239. [32] Y. Ke, R. Sukthankar, Pca-sift: A more distinctive representation for local image descriptors, in: Computer Vision and Pattern Recognition, 2004. CVPR 2004. Proceedings of the 2004 IEEE Computer Society Conference on, Vol. 2, IEEE, 2004, pp. II–506. [33] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, I. H. Witten, The weka data mining software: an update, ACM SIGKDD explorations newsletter 11 (1) (2009) 10–18. 21 [34] B. Swingle, Rényi entropy, mutual information, and fluctuation properties of fermi liquids, Physical Review B 86 (4) (2012) 045109. [35] A. Kraskov, H. Stögbauer, P. Grassberger, Estimating mutual information, Phys. Rev. E 69. [36] K. Deb, S. Agrawal, A. Pratap, T. Meyarivan, A fast elitist non-dominated sorting genetic algorithm for multi-objective optimization: Nsga-ii, Lecture notes in computer science 1917 (2000) 849–858. [37] P. E. Lutu, A. P. Engelbrecht, A decision rule-based method for feature selection in predictive data mining, Expert Systems with Applications 37 (1) (2010) 602–609. [38] Q. Song, J. Ni, G. Wang, A fast clustering-based feature subset selection algorithm for high dimensional data, IEEE transactions on knowledge and data engineering 25 (1) (2013) 1–14. [39] B. Frénay, G. Doquire, M. Verleysen, Theoretical and empirical study on the potential in- adequacy of mutual information for feature selection in classification, Neurocomputing 112 (2013) 64–78. [40] G. Brown, A new perspective for information theoretic feature selection, in: International Conference on Artificial Intelligence and Statistics, 2009, pp. 49–56. [41] T. G. Dietterich, Approximate statistical tests for comparing supervised classification learn- ing algorithms, Neural computation 10 (7) (1998) 1895–1923. 22 (a) Decision Trees accuracy CKDD (b) Random Forests accuracy for CKDD (c) KNN accuracy for CKDD (d) SVM Accuracy for CKDD (e) Decision Trees accuracy 10%KDD (f) Random Forests accuracy for 10%KDD (g) KNN accuracy for 10%KDD (h) SVM Accuracy for 10%KDD (i) Decision Trees accuracy NSL KDD (j) Random Forests accuracy for NSL KDD 23 (k) KNN accuracy NSL KDD (l) SVM accuracy for NSL KDD Figure 6: Accuracy of di!erent classifiers found in non-intrusion datasets (a) Decision Trees accuracy Bloggender Fe- male (b) Random Forests accuracy for Bloggender Feamle (c) KNN accuracy for Bloggender Female (d) Decision Trees Accuracy for Bloggender Male (e) Random Forests accuracy Bloggender Male (f) KNN accuracy for Bloggender Male Figure 7: Accuracy of di!erent classifiers found in Text Categorization datasets 24 (a) Decision Trees accuracy for Lymphoma (b) Random Forests accuracy for Lym- phoma (c) KNN accuracy for Lymphoma (d) SVM accuracy for Lymphoma (e) Decision Trees Accuracy for Colon Can- cer (f) Random Forests accuracy Colon Cancer (g) KNN accuracy for Colon Cancer (h) SVM accuracy for Colon Cancer Figure 8: Accuracy of di!erent classifiers for Gene Expression datasets 25 
work_oj7jjtk2rnflli5ntvucohg56q	----	١ جهت افزایش کارایی خطی غیرالگوریتم ژنتیککاربرد الگوریتم ژنتیک خطی و ها در بازار سرمایه شرکتبینی بحران مالی پیش 2محمدرضا اولی1زهرا پورزمانی چکیده بنـابراین .کشـد هم از لحاظ اجتماعی و هم از لحاظ اقتصادي کشور به چـالش مـی رویدادي است که ورشکستگی تـوانیم از اگر بتوانیم در مورد امکان وقوع ورشکستگی پیش از رخداد واقعی آن اطالعـاتی بـه دسـت آوریـم، مـی بینـی هدف ایـن تحقیـق بررسـی قـدرت پـیش . پیامدهاي اقتصادي و اجتماعی آن کاسته و یا حتی جلوگیري کنیم لگوریتم ژنتیک غیرخطی جهت باال بردن توان تصمیم هاي الگوریتم ژنتیک خطی و ابحران مالی با استفاده از مدل توجـه بـه نتـایج بدسـت .ها مـی باشـد هاي مالی در پیش بینی بحران مالی شرکتگیري استفاده کنندگان صورت در دسـترس بـر اسـاس اطالعـات و آمارهـاي . آمده، الگوها با یکدیگر مقایسه و بهترین الگو استخراج شـده اسـت مشمول ماده هاي ، از بین شرکت1376-1389دوره طی در در بورس اوراق بهادار تهرانهاي پذیرفته شدهشرکت نمـار بـراي نتـایج آزمـون مـک .شرکت انتخاب شد72ها نیز شرکت و از بین بقیه شرکت72، تجارتقانون141 90(غیرخطی بینی الگوریتم ژنتیکهاي الگوریتم ژنتیک خطی و غیرخطی نشان داد اگرچه که دقت پیشتکنیک . دار نیستاست ولی این تفاوت از لحاظ آماري معنی) درصد80(بیشتر از الگوریتم ژنتیک خطی ) درصد الگوریتم ژنتیک غیر خطیالگوریتم ژنتیک خطی، بحران مالی، متغیرهاي مالی، :هاي کلیديواژه zahra.poorzamani@yahoo.comدانشگاه آزاد اسالمی، واحد تهران مرکزي، تهران،استادیار، , زهرا,ورزمانیپ1- mohammadreza_ola@yahoo.comتهران,واحد علوم و تحقیقات,دانشگاه آزاد اسالمی ,دانشجوي دکتري حسابداري ,اولی,محمدرضا-2 ٢ مقدمه ها دارند زیرا در صورت بحران مالی بنگاهبینیگذاران و اعتباردهندگان تمایل زیادي براي پیشسرمایه هاي در چند دهه اخیر پژوهش).2002، 3ان و دیگرانندا(هاي زیادي به آنها تحمیل می شودورشکستگی هزینه ها در هر چند نخستین تالش. بینی ورشکستگی انجام گرفته استخصوص پیشه اي در زمینه ورشکستگی بگسترده تري و با تحقیق انجام گرفته توسط بیور این مقوله شکل جدي1966گردد اما از سال بر می1930این زمینه به سال بینی بحران مالی یا ورشکستگی یکی از نخستین محققانی است که به مطالعه پیش) 1966(4بیور. به خود گرفت با ) 1968(پس از او، آلتمن . شودپرداخته است و به عنوان منادي تحقیقات آکادمیک در این زمینه محسوب می . هاي آماري پیشرفته توانست به موفقیت چشمگیري دست یابداستفاده از روش دهند ها کارایی خود را از دست میکند و در نتیجه متغیرهاي مورد استفاده در مدلبا گذشت زمان شرایط تغییر می ها اند متفاوت از دیگر بخشریزي شدهپایه آن طرحها بر هاي اقتصادي که این مدلهمچنین سیستم). 2006، 5هابر( بینی بحران مالی یا ورشکستگی هاي سنتی پیشروشاز سوي دیگر ). 2001، 6گریک و دوگان(و یا کشورها است بینی کننده یا داراي برخی مفروضات محدود کننده مانند خطی بودن، نرمال بودن و مستقل بودن متغیرهاي پیش . هاي زیادي هستندهاي سنتی در ارتباط با میزان کارایی و اعتبار، داراي محدودیتروشنابراینب. ها استورودي بینی بحران مالی یا هاي سنتی پیشهاي الگوریتم ژنتیک نسبت به سایر مدلکی از برتريیکهاز انجایی ها یا برابري بودن توزیع نسبتکننده و نرمال هاي آماري محدودعدم وابستگی این الگوریتم بر فرضیه،ورشکستگی بینی بحران مالی با استفاده از ، هدف این تحقیق بررسی قدرت پیشها استواریانس یا کوواریانس ماتریس نسبت جهت باال بردن توان تصمیم و مقایسه آنها بایکدیگر هاي الگوریتم ژنتیک خطی و الگوریتم ژنتیک غیرخطی مدل .می باشدها ي مالی در پیش بینی بحران مالی شرکتهاگیري استفاده کنندگان صورت مبانی نظري و پیشینه پژوهش هاي تکنیک،)کالسیک(هاي آماري بینی ورشکستگی بر اساس ماهیت خود در سه دسته تکنیکهاي پیشتکنیک ها ترین تکنیکترین و رایجهاي آماري از ابتداییتکنیک. اندبندي شدهطبقههاي نظريهوش مصنوعی و مدل ٣ Adnan et al ٤ Beaver ٥ Haber ٦ Grice and Dugan ٣ شوند؛ هاي آماري خود به دو گروه تقسیم میمدل. روندبینی ورشکستگی به شمار میبندي پیشجهت مدل هاي چند متغیره از فراوانی هاي آماري چند متغیره ، که در میان آنها مدلهاي آماري تک متغیره و مدلمدل هاي تعدیل ناقص ، پروبیت، مجموع تجمعی و فرایندتحلیل تشخیصی، احتمال خطی، لوجیت. بیشتري برخور دارند . هاي آماري چند متغیره هستندتشکیل دهنده تکنیک استرشدحالدرو8ايرشتهمیانجدیدحوزهیکAIT(7(هاي هوش مصنوعیکاوي و تکنیکهاي دادهمدل اینکاربردزمینهدرزیاديتحقیقات. )2009، 9تیساي(کنند بینیپیشراوکارهاکسبشکستاستقادرکه و اعتماديتحقیقبهتوانمیمیاناینازکهاستانجام گرفتهو کارهاکسبشکستبینیپیشبرايهاتکنیک و؛ مین)2009(12دیگرانولین؛)2009(11چنو؛ هانگ)2008(10و دیگران؛ هووانگ)2009(دیگران اشاره ) 2010(17و وو) 2008(16لیو؛ سان)2008(15پرامودو؛ راوي)2008(14لیو؛ مین)2009(13جوونگ یاوماشینپشتیبانبردارو،18تصمیمهايدرختعصبی،هايشبکهبهتوانمیهاتکنیکاینجملهاز. کرد ، 20ناهموارمجموعهتئوري،19ژنتیکهايالگوریتممانندمصنوعیهوشهايروشباهاتکنیکاینترکیبی از .کرداشارهغیرهو21فازيمجموعهتئوري شرکت، 500بینی ورشکستگی استفاده کرد نمونه او متشکل از از الگوریتم ژنتیک براي پیش) 1998(22وارتو یک% 93بینینتایج این تحقیق بیانگر دقت پیش. شرکت غیر ورشکسته است264شرکت ورشکسته و 236شامل الگوي مبتنی بر الگوریتم ژنتیک را با ووارت. سه سال قبل از ورشکستگی است% 91.6سال قبل از ورشکستگی و الگوي تحلیل تشخیصی خطی مقایسه نمود ضمن بیان برتري الگوي الگوریتم ژنتیک در پیش بینی یک سال قبل از ع فراتر از الگوي الگوریتم ژنتیک دانست وقوع، الگوي تحلیل تشخیصی خطی را در پیش بینی سه سال قبل از وقو ٧ Artificial Intelligence techniques ٨ Interdisciplinary ٩ Tsai ١٠ Huang, Tsai, Yen, Cheng ١١ Hung, Chen ١٢ Lin and et al. ١٣ Min, Jeong ١٤ Min, Lee ١٥ Ravi, Pramodh ١٦ Sun, Li ١٧ Wu ١٨ Decision trees ١٩ Genetic algorithm ٢٠ Rough set theory ٢١ Fuzzy set theory ٢٢ Varetto ٤ از دیگر مطالعات .استو از طرفی بیان نمود که الگوي تحلیل تشخیصی خطی داراي ثبات و قابلیت تعمیم بیشتري .اشاره کرد) 2002(24و مک کی و لنزبرگ) 2002(23لیتوان به شین وانجام شده در این زمینه می اب، الگوریتم ژنتیک و تحلیل تشخیصی چندگانه را با یکدیگر مقایسه نمود الگوریتم پیش انتخ)2004(کاواکامی که نتایج وي بیانگر برتري الگوریتم پیش انتخاب نسبت به دو الگوي دیگر و از طرفی برتري الگوریتم ژنتیک .نسبت به تحلیل تشخیصی چندگانه می باشد هاي پذیرفته شده در بورس اوراق بهادار را با شرکتبینی ورشکستگی بندي پیشمدل) 1384(زاده دهکردي فرج وي جهت ساخت . ریزي ژنتیک مورد بررسی قرار داداستفاده از دو مدل تحلیل تشخیصی چندگانه و برنامه ها در تهیه کرد و پس از بررسی نسبت) نسبت مالی93(هاي مالی هاي مذکور ابتدا فهرست کاملی از نسبتمدل جهت ساخت مدل استخراج و در نهایت با استفاده از آزمون برابري میانگین دو جامعه، اقدام نسبت مالی42نهایت ریزي ژنتیک و تحلیل تشخیصی چندگانه توانستند در نهایت مدل برنامه. به ساخت دو مدل مورد بحث نمود به صورت % 73و % 90هاي نمونه آزمایشی به ترتیب و شرکت% 77و % 94هاي نمونه آموزشی را به ترتیب شرکت .بندي نمایدصحیح طبقه الگوهاي مبتنی بر (بینی کننده بحران مالیهدف تحقیق خود را ساخت الگوهاي پیش) 1389(پورزمانی و دیگران بینی بحران براي پیش) خطی و شبکه عصبیژنتیک خطی، الگوریتم ژنتیک غیر، الگوریتم MDAروش هاي سنتی الگوهاي مبتنی بر روش هاي (در این تحقیق چهار الگوي پیش بینی بحران مالی. قراردادمالی دو سال قبل از وقوع براي پیش بینی بحران مالی دو ) ژنتیک خطی، الگوریتم ژنتیک غیر خطی و شبکه عصبی، الگوریتم MDAسنتی مقایسه و مشخص سپس با توجه به نتایج بدست آمده، الگوها با یکدیگر. سال قبل از وقوع آن تدوین شده است . گردید الگوي مبتنی بر شبکه عصبی داراي باالترین توان در پیش بینی بحران مالی شرکت ها می باشد هايعرضهقیمت گذاريجهتمناسببینیپیشابزارایجادپژوهشی تحت عنوان) 1388(عرب مازار و قاسمی هايشبکهترکیبدهدمینشاننتایج.انجام دادندژنتیک الگوریتموعصبیهايشبکهوسیلهه اولیه بعمومی میافزایشمحسوسیطوربهرابینیپیشقدرتبهینه،انتخاب متغیرهايمنظوربهژنتیکالگوریتمباعصبی .دهد ورشکستگیپیش بینیافزایش دقتژنتیک درالگوریتمازاستفادهدادندنشان) 1388(نژاد و اسکندريفدایی توان نمیآمارينظرازکهدادنشانذراتتجمعیسازيبهینهژنتیک والگوریتمهايمدلمقایسهاما استموثر بازاراز داده هايکهمدلیدادنشاننتایجهمچنین.داردبرتريدیگريبرهاروشاینازیکیکهنموداثبات ٢٣ Shin and Lee ٢٤ McKee and Lensberg ٥ ورشکستگیدرصد92/6تا تواندمیذرات آموزش ببیندتجمعیسازيبهینهالگوریتمطریقازوکردهاستفاده .نمایدبینیپیشدرستیبهراهاشرکت فرضیه تحقیق :با توجه به اهداف مورد نظر در این تحقیق سه فرضیه تدوین شد داراي توانمندي در پیش بینی بحران مالی استمبتنی بر الگوریتم ژنتیک خطی مدل: فرضیه اول الگوریتم ژنتیک غیر خطی داراي توانمندي در پیش بینی بحران مالی استمبتنی بر مدل: دومفرضیه مبتنی بر الگوریتم مبتنی بر الگوریتم ژنتیک خطی بیشتر از مدلمدلبحران مالی بینی قدرت پیش: سومفرضیه .خطی استغیرژنتیک جامعه و نمونه آماري هاي پذیرفته شده در بورس اوراق بهادار تهران طی دوره جامعه آماري مورد بررسی در این تحقیق، کلیه شرکت -گروه اول شرکته به دو گروه عمده تقسیم می شوند؛ نمونه آماري شرکت هاي مورد مطالع. است1376-1389 -قانون تجارت شده141ی مشمول ماده که براي سه سال متوالمی باشدشرکت72هاي درگیر بحران مالی به تعداد هاي داراي سالمت شرکتبه همان تعداد نیز، هاي درگیر بحران مالی شرکتبراي مطابقت با گروه دوم که . اند گیري تصادفی انتخاب با استفاده از روش نمونهقانون تجارت نبودند141که مشمول ماده مالی یا بدون بحران مالی هاي ورشکسته و غیر هاي پذیرفته شده در بورس اوراق، امکان انطباق شرکتبه دلیل محدودیت تعداد شرکت.شد ها به عنوان یک متغیر بالقوه براي از آنجایی که اندازه شرکت. شدبامیورشکسته از نظر صنعت فعالیت میسر ن . شودمیاساس اندازه شرکت نیز صرف نظر ها بر است، از انطباق شرکتبینی ورشکستگی در نظر گرفته شدهپیش هاي ورشکسته انطباق داده هاي غیر ورشکسته تنها از نظر سال مالی با شرکتگیري، شرکتبنابراین در نمونه . شوندمی به دو گروه آموزشی و ها مجددپس از انتخاب شرکت هاي نمونه به صورت تصادفی هر یک از دو گروه شرکت 51گیرد، شامل مجموعه آموزشی که جهت ساخت و آموزش مدل مورد استفاده قرار می. آزمایشی تفکیک شدند 19شرکت ورشکسته و 21مجموعه آزمایشی که شامل . شرکت غیرورشکسته است53شرکت ورشکسته و . رودشرکت غیرورشکسته است به منظور بررسی میزان قابلیت تعمیم و روایی خارجی مدل حاصله به کار می ٦ ي تحقیقمتغیرها نسبت مالی 23فهرستی متشکل از ،بینی بحران مالیمناسب براي پیشي مالیهادر این تحقیق جهت تعیین نسبت 1شرح جدول شماره ه باندبینی بحران مالی یا ورشکستگی به کار رفتهات قبلی براي پیشقاست که در تحقی .استخراج شد مالی براي ساخت مدل، موجب افزایش پیچیدگی مدل و کاهش کارایی آن نسبت 23از آنجایی که به کارگیري هاي مالی که داراي بیشترین توانایی در براي انتخاب نسبت) SDA(25تحلیل تشخیصی گام به گامشود از روش می . هاي ورشکسته از غیرورشکسته هستند استفاده شدافتراق شرکت به این ترتیب . کنداز نسبت واریانس درون گروهی به واریانس برون گروهی استفاده میSDAبراي انجام این امر هاي ورشکسته و غیر ورشکسته است و همچنین داراي ی که داراي کمترین واریانس در هر یک از گروهنسبت مال هاي اق شرکتبیشترین واریانس بین دو گروه ورشکسته و غیر ورشکسته است، داراي بیشترین قدرت در افتر .هاي غیرورشکسته استورشکسته از شرکت متغیرهاي مستقل استفاده شده در تحقیق. 1جدول نسبت مالیمتغیرنسبت مالیمتغیر Xسرمایه در گردش به حقوق صاحبان سهام١Xهانسبت فروش به کل دارایی١٣ Xنسبت سرمایه در گردش به فروش٢Xهاداراییها به کل نسبت کل بدهی١٤ Xهانسبت سرمایه در گردش به کل بدهی٣Xهانسبت سود و زیان انباشته به کل دارایی١٥ Xهانسبت سرمایه در گردش به کل دارایی٤Xنسبت سود عملیاتی به فروش١٦ X٥ نسبت سود قبل از کسر بهره و مالیات به حقوق صاحبان X١٧سهام نسبت هزینه مالی به سود ناخالص Xنسبت سود قبل از کسر بهره و مالیات به فروش٦Xهانسبت دارایی جاري به کل دارایی١٨ Xهانسبت سود قبل از کسر بهره و مالیات به کل بدهی٧Xنسبت فروش به دارایی جاري١٩ Xهانسبت سود قبل از کسر بهره و مالیات به کل دارایی٨Xنسبت دارایی جاري به بدهی جاري٢٠ Xنسبت حقوق صاحبان سهام به فروش٩Xنسبت سود خالص به فروش٢١ Xهانسبت حقوق صاحبان سهام به کل بدهی١٠Xهانسبت سود خالص به کل دارایی٢٢ Xهانسبت حقوق صاحبان سهام به کل دارایی١١Xهانسبت بدهی جاري به کل دارایی٢٣ Xهانسبت فروش به کل بدهی١٢ ٢٥ Stepwise Discriminant Analysis ٧ SDAدر این مرحله. گردداجرا میجهت انتخاب دومین متغیر مجدداSDAنخستین متغیر، فرایندپس از انتخاب . ها باشدنماید که در کنار نخستین نسبت مالی داراي بیشترین توانایی افتراق گروه شرکترا انتخاب مینسبت مالی این . گیردها مد نظر قرار میروه شرکتدر واقع در این مرحله تاثیر توامان و متقابل دو نسبت مالی در تفکیک گ هاي مالی انتخاب نشده واجد شرایط سطح اهمیت تعیین فرایند هنگامی متوقف خواهد شد که هیچ یک از نسبت . دهدمتغیر را نشان می23بر روي SDAنتایج حاصل از فرایند2جدول .نباشند)0.1(شده SDAماتریس سازه فرایند: 2جدول )a .(هاي مالی انتخاب نشدهنسبت کد میزان اهمیت کد میزان اهمیت کد میزان اهمیت X١٦ .857 X١٤ -.437 X١٩ .243 X١(a) .832 X٨(a) .359 X٢٣(a) .222 X٣(a) .750 X١١(a) .344 X١٥(a) .221 X٢١(a) .745 X٩(a) .333 X١٣(a) -.157 X٢(a) .721 X٢٠(a) .318 X١٧ -.102 X٢٢(a) .636 X١٨ .305 X٤(a) -.044 X٦(a) .550 X١٠(a) .270 X٥(a) .026 X٧(a) .477 X١٢(a) .252 طی نموده است که طی هر یک از 3طبق جدول مرحله را5هاي مالی نهایی جهت انتخاب نسبتSDAفرایند کاهش. گردداضافه میهاي مالی انتخاب شده هاي مالی انتخاب شده و به مجموعه نسبتاین مراحل یکی از نسبت Wilks' Lambdaانتخاب شدهمتغیرپنج.تشخیصی متغیرهاي انتخاب شده استتهر مرحله به معناي افزایش قدر :به ترتیب قدرت تشخیصی عبارتند ازدهدهاي مهم وضعیت مالی شرکت را پوشش مییکی از جنبهکدامهر که X16) سودآوري(فروشسود عملیاتی به نسبت .1 X14) هاتوان پرداخت بدهی(ها ها به کل داراییکل بدهینسبت .2 X18) نقدینگی(ها به کل داراییجاريهاي دارایینسبت .3 X19) کارایی(هاي جاري نسبت فروش به دارایی.4 X17) پوشش بهره(نسبت هزینه مالی به سود ناخالص .5 ٨ هاي انتخاب متغیرهاخالصه گام: 3جدول مراحل کد حد تغییرات Fجهت خروج Wilks' Lambda 1 X١٦ 1.000 71.858 2 X١٦ .973 76.804 .986 X١٤ .973 6.212 .697 3 X١٦ .804 35.601 .787 X١٤ .925 8.740 .681 X١٨ .766 6.281 .671 4 X١٦ .771 26.502 .731 X١٤ .826 12.246 .675 X١٨ .724 8.552 .661 X١٩ .851 4.726 .646 5 X١٦ .738 18.332 .711 X١٤ .814 9.843 .654 X١٨ .714 7.656 .641 X١٩ .801 4.240 .635 X١٧ .708 3.213 .617 بینی بحران مالی مبتنی بر الگوریتم ژنتیک خطی و غیرخطیمدل پیش 4ویرایش GeneXproToolsافزاربینی بحران مالی از نرمبراي اجراي فرایند الگوریتم ژنتیک و ایجاد مدل پیش . اندانتخاب شده0.06و 0.6عملگرهاي تقاطع و جهش به ترتیب در سطح . استفاده شده است باشد ) ايارزش آستانه(0.5هنگامی که نتیجه حاصل از مدل الگوریتم ژنتیک براي یک شرکت، بزرگتر یا مساوي بالعکس، هنگامی که ارزش حاصل از مدل . شودبندي میهاي ورشکسته طبقهاین شرکت در گروه شرکت مقایسه گروه واقعی . شودبندي میاست شرکت در گروه غیرورشکسته طبقه0.5ریزي ژنتیک کمتر از برنامه . کندگیري میبینی شده توسط مدل الگوریتم ژنتیک دقت مدل را اندازهها با گروه پیششرکت این درخت تصمیم معرف یک کروموزم . دهدخطی را نشان میبهترین مدل حاصل از الگوریتم ژنتیک 1تصویر .براي تعیین گروه شرکت، مقایسه شود0.5اي نتیجه این درخت براي یک شرکت بایستی با ارزش آستانه. است ٩ :، مدل حاصل به صورت زیر قابل ارایه است1با توجه تصویر Y=X١٤+X١٨+((X١٩+X١٤)-X١٨)-X١٦+((X١٤-X١٦)-X١٩) این درخت تصمیم معرف یک . دهدبهترین مدل حاصل از الگوریتم ژنتیک غیرخطی را نشان می. 2تصویر براي تعیین گروه شرکت، 0.5اي براي یک شرکت بایستی با ارزش آستانهنتیجه این درخت. کروموزم است . مقایسه شود + - + - -+ - - X١٩ X١٦ X١٤ X١٦ X١٨X١٨X١٤ X١٤X١٩ بینی بحران مالی یا ورشکستگی حاصل از فرایند الگوریتم ژنتیک خطیبهترین مدل پیش. 1تصویر ١٠ :، مدل حاصل به صورت زیر قابل ارایه است. 2با توجه به تصویر Y=(X٢(١٦+(X١٤-X١٦-((X١٦×X١٩)×X١٤×X١٩))+((-X٢(١٨×X١٤×X١٨) آزمون فرضیه هاي تحقیق این دو نمونه به تفکیک تعداد نمونه آموزشی و تعداد نمونه آزمایشی و تعداد خطاها در جهت آزمون فرضیه اول .نشان داده شده است4هاي ورشکسته و غیرورشکسته در مدل الگوریتم ژنتیک خطی در جدول براي شرکت تعداد نمونه و خطاها در مدل الگوریتم ژنتیک خطی. 4جدول نمونه آموزشینمونه آزمایشی تعداد نمونهتعداد خطاهاتعداد نمونهتعداد خطاها غیرورشکسته4191453 ورشکسته421451 جمع84018104 + + -^ * * X١٨ X١٦ ٢ X١٤ X١٤ X١٤ X١٦ X١٩X١٦ ژنتیک غیرخطیبینی بحران مالی یا ورشکستگی حاصل از فرایند الگوریتم بهترین مدل پیش. 2تصویر ^٢ - * * * X١٩ -١ X١٨ * ١١ به صورت صحیح % 83هاي موجود در نمونه آموزشی را با دقت کلی مدل الگوریتم ژنتیک خطی توانست شرکت شرکت موجود در 104به این صورت که از میان . بندي نمایدهاي ورشکسته و غیر ورشکسته طبقهدر گروه دهد که مدل بررسی نتایج این مدل نشان می. اندبندي شدهشرکت به صورت صحیح طبقه86مجموعه آموزشی، درصدي 92هاي ورشکسته در مجموعه آموزشی داراي دقت بندي صحیح شرکتالگوریتم ژنتیک خطی در طبقه ن همچنی). اندبندي شدهشرکت به صورت صحیح طبقه47شرکت ورشکسته در این مجموعه، 51از میان (است 53از میان (است % 74هاي غیرورشکسته در مجموعه آموزشی داراي دقت بندي صحیح شرکتاین مدل در طبقه ).اندبندي شدهشرکت به صورت صحیح طبقه39شرکت غیرورشکسته در این مجموعه، اي مربوط به هبه منظور بررسی توانایی قدرت تعمیم و پایداري مدل الگوریتم ژنتیک خطی، این مدل بر روي داده هاي قرار گرفته در نمونه آزمایشی در فرایند شرکت. شرکت موجود در نمونه آزمایشی آزمون شده است40 مدل .توانند جهت آزمون روایی خارجی مدل به کار رونداند، بنابراین به درستی میساخت مدل دخالت نداشته به صورت صحیح در % 80ایشی را با دقت کلی هاي موجود در نمونه آزمالگوریتم ژنتیک خطی توانست شرکت شرکت موجود در مجموعه 40به این صورت که از میان . بندي نمایدهاي ورشکسته و غیرورشکسته طبقهگروه بندي صحیح مدل الگوریتم ژنتیک خطی در طبقه. اندبندي شدهشرکت به صورت صحیح طبقه32آزمایشی، شرکت ورشکسته در این 21از میان (درصدي است 81ی داراي دقت هاي ورشکسته در مجموعه آزمایششرکت هاي بندي صحیح شرکتهمچنین این مدل در طبقه). اندبندي شدهشرکت به صورت صحیح طبقه17مجموعه، 15شرکت غیرورشکسته در این مجموعه، 19از میان (است % 79غیرورشکسته در مجموعه آزمایشی داراي دقت ).اندبندي شدهیح طبقهشرکت به صورت صح بینی مدل الگوریتم ژنتیک خطی ماتریس اغتشاش نتیجه پیشي اندازه گیري دقت مدل الگوریتم ژنتیک خطی، راب .نشان داده شده است5در جدول مدل الگوریتم ژنتیک خطیبینیپیشنتیجهماتریس اغتشاش. 5جدول واقعی غیرورشکستهورشکسته شدهبینیپیش TP=17FN=4ورشکسته FP=4TN=15غیرورشکسته ١٢ درصد است را به شیوه زیر 80توان دقت مدل الگوریتم ژنتیک خطی را که در نمونه آزمایشی برابر بنابراین می :محاسبه کرد 80%= 15+17 = TP+TN دقت مدل الگوریتم ژنتیک خطی= 4+15+4+17TP+FP+TN+FN تعداد نمونه آموزشی و تعداد نمونه آزمایشی و تعداد خطاها در این دو نمونه به تفکیک براي آزمون فرضیه دوم .نشان داده شده است6هاي ورشکسته و غیرورشکسته در مدل الگوریتم ژنتیک غیرخطی در جدول براي شرکت تعداد نمونه و خطاها در مدل الگوریتم ژنتیک غیرخطی. 6جدول آموزشینمونه نمونه آزمایشی تعداد نمونهتعداد خطاهاتعداد نمونهتعداد خطاها غیرورشکسته019253 ورشکسته421751 جمع4409104 به صورت % 91هاي موجود در نمونه آموزشی را با دقت کلی مدل الگوریتم ژنتیک غیرخطی توانست شرکت شرکت موجود در 104به این صورت که از میان . نمایدبندي هاي ورشکسته و غیرورشکسته طبقهصحیح در گروه دهد که مدل بررسی نتایج این مدل نشان می. اندبندي شدهشرکت به صورت صحیح طبقه85مجموعه آموزشی، 86هاي ورشکسته در مجموعه آموزشی داراي دقت بندي صحیح شرکتالگوریتم ژنتیک غیرخطی در طبقه ). اندبندي شدهشرکت به صورت صحیح طبقه44ورشکسته در این مجموعه، شرکت 51از میان (درصدي است از (است % 96هاي غیرورشکسته در مجموعه آموزشی داراي دقت بندي صحیح شرکتهمچنین این مدل در طبقه ).اندبندي شدهشرکت به صورت صحیح طبقه51شرکت غیرورشکسته در این مجموعه، 53میان هاي مربوط به یی قدرت تعمیم و پایداري مدل الگوریتم ژنتیک خطی، این مدل بر روي دادهبه منظور بررسی توانا هاي قرار گرفته در نمونه آزمایشی در فرایند شرکت. استشرکت موجود در نمونه آزمایشی آزمون شده40 مدل . ه کار روندتوانند جهت آزمون روایی خارجی مدل باند، بنابراین به درستی میساخت مدل دخالت نداشته به صورت صحیح % 90هاي موجود در نمونه آزمایشی را با دقت کلی الگوریتم ژنتیک غیرخطی توانست شرکت شرکت موجود در مجموعه 40به این صورت که از میان . بندي نمایدهاي ورشکسته و غیرورشکسته طبقهدر گروه بندي صحیح ل الگوریتم ژنتیک غیرخطی در طبقهمد. اندبندي شدهشرکت به صورت صحیح طبقه36آزمایشی، شرکت ورشکسته در این 21از میان (درصدي است 81هاي ورشکسته در مجموعه آزمایشی داراي دقت شرکت ١٣ هاي بندي صحیح شرکتهمچنین این مدل در طبقه). اندبندي شدهشرکت به صورت صحیح طبقه17مجموعه، شرکت غیرورشکسته در این مجموعه، 19از میان (است % 100ي دقت غیرورشکسته در مجموعه آزمایشی دارا ). اندبندي شدهبه صورت صحیح طبقههاتمامی شرکت بینی مدل الگوریتم ژنتیک ماتریس اغتشاش نتیجه پیشبراي اندازه گیري دقت مدل الگوریتم ژنتیک غیرخطی، .نشان داده شده است7غیرخطی در جدول مدل الگوریتم ژنتیک غیرخطیبینیپیشنتیجهماتریس اغتشاش7جدول واقعی غیرورشکستهورشکسته شدهبینیپیش ٠=١٧FN=TPورشکسته ١٩=٤TN=FPغیرورشکسته درصد است را به شیوه زیر 90توان دقت مدل الگوریتم ژنتیک غیرخطی را که در نمونه آزمایشی برابر بنابراین می :کردمحاسبه 90%= 17+19 = TP+TN =دقت مدل الگوریتم ژنتیک غیرخطی 0+19+4+17TP+FP+TN+FN هاي الگوریتم ژنتیک خطی و هاي مبتنی بر تکنیکبینی مدلبراي آزمون فرضیه سوم این تحقیق، دقت پیش بینی مدل همانطور که مشاهده شد دقت پیش. گیردمورد مقایسه قرار می26نمارغیرخطی با استفاده از آزمون مک درصد 90درصد و تکنیک الگوریتم ژنتیک غیرخطی برابر 80ایجاد شده با تکنیک الگوریتم ژنتیک خطی برابر هم تفاوت ها با هاي ایجاد شده با این تکنیکبینی مدلحال سوال این است که آیا این تفاوتها بین دقت پیش. است هاي ایجاد شده با همدیگر مقایسه شده تا در مورد وجود نمار ، نتایج مدلدر آزمون مک. دار دارند یا خیرمعنی دار بین نتایج نتایج این آزمون براي آزمون وجود تفاوت معنی. گیري شوددار در این نتایج تصمیمتفاوت معنی .نشان داده شده است8لگوریتم ژنتیک خطی وغیر خطی در جدول هاي اهاي ایجاد شده با استفاده از تکنیکمدل ٢٦ Mc Nemar Test ١٤ هاي الگوریتم ژنتیک خطی و شبکه عصبینمار براي تکنیکنتایج آزمون مک. 8جدول GANL & GAL GANL GAL ٠ ١ ٠ ٤٦ ٢٣ ١ ١٦ ٥٩ Test Statisticsb GANL & GAL N ١٤٤ Chi-Squarea .٨٢١ Asymp. Sig. .٢٣٧ a. Continuity Corrected b. McNemar Test دهد که چون سطح هاي الگوریتم ژنتیک خطی و شبکه عصبی نشان مینمار براي تکنیکنتایج آزمون مک الگوریتم داري بین نتایج الگوریتم ژنتیک خطی و در نتیجه تفاوت معنی) 237/0(درصد است 5داري بیشتر از معنی الگوریتم بیشتر از ) درصد90(خطی غیربینی الگوریتم ژنتیکاگر چه دقت پیش. نداردوجود ژنتیک غیرخطی مبنی بر سومدار نیست و بنابراین فرضیه فرعی است ولی این تفاوت از لحاظ آماري معنی) درصد80(ژنتیک خطی خطی تایید غیرنسبت به الگوریتم ژنتیکالگوریتم ژنتیک خطی بینی مدل مبتنی بر بیشتر بودن قدرت پیش .شودنمی بحث و نتیجه گیري تري نسبت به اطالعات سنتی توان اطالعات دقیقهاي کامپیوتري میامروزه با پیشرفت سریع فناوري و تکنولوژي تري را در خصوص احتمال برگشت هاي مناسبگیريگیرندگان قرار داد، تا بتوانند تصمیمدر اختیار تصمیم .هاي سنگین اتخاذ نمایندبحران مالی قبل از وقوع و تحمل هزینهسرمایه و یا وقوع با بررسی روند مطالعات . بینی بحران مالی یکی از ابزارهاي برآورد وضع آینده شرکت ها می باشدالگوهاي پیش بیشتر هاي آماري در این زمینه کاهش یافته است و مطالعات اخیر شود که امروزه استفاده از مدلمشخص میپیشین شبکه هاي عصبی، درخت تصمیم گیري، استدالل مبتنی بر (هاي مبتنی بر هوش مصنوعی تمایل به استفاده از مدل یکی از مهمترین . دارند) موضوع، الگوریتم ژنتیک، مجموعه هاي سخت، ماشین بردار تکیه گاه و منطق فازي ها اغلب این مدلاین است کهشده است گرانهاي هوش مصنوعی از سوي پژوهشموجب اقبال مدلی که دالیل ناپارامتریک بوده و در بکارگیري آنها نیاز چندانی به فرضیات اولیه و یا اطالعات مربوط به چگونگی توزیع .هاي ورشکسته و غیر ورشکسته نیستهاي شرکتهاي مالی در میان گروهویژگی هاي الگوریتم ژنتیک خطی و مالی با استفاده از مدلبینی بحراندر این راستا هدف این تحقیق بررسی قدرت پیش الگوریتم ژنتیک غیرخطی و مقایسه آنها بایکدیگر جهت باال بردن توان تصمیم گیري استفاده کنندگان صورت دهد که مدل الگوریتم ژنتیک نتایج تحقیق نشان می. ها قرار گرفتهاي مالی در پیش بینی بحران مالی شرکت ١٥ درصد عالوه بر ایجاد نتایج 90بینیدرصد و مدل الگوریتم ژنتیک غیرخطی با دقت پیش80بینی خطی با دقت پیش ها به گروه بندي شرکتها با استفاده از اطالعات مالی آنها، در طبقهبینی وضعیت آتی شرکتمطلوب در پیش . باشنداطمینان میکه قابلکنندهاي غیرورشکسته نتایج متعادلی ایجاد میهاي ورشکسته یا گروه شرکتشرکت . داري بین نتایج الگوریتم ژنتیک خطی و غیرخطی وجود نداردنتایج این تحقیق نشان داد که تفاوت معنیهمچنین بینی الگوریتم ژنتیک غیرخطی از الگوریتم ژنتیک خطی بیشتر است ولی این تفاوت از لحاظ اگر چه دقت پیش . دار نیستآماري معنی فارسیمنابع بررسی توانمندي الگوهاي پیش بینی "، )1389(پورزمانی، زهرا، کی پور، رضا، نورالدین ،مصطفی، .1 الگو هاي مبتنی بر روشهاي سنتی، الگوریتم ژنتیک و شبکه : الگوهاي مورد مطالعه (کننده بحران مالی 28-1اسالمی،صص، دانشگاه آزاد4، فصلنامه مهندسی مالی و مدیریت پورتفوي،شماره ")هاي عصبی هاي بورس نامهمجموعه قوانین و آیین"). 1378(شوراي بورس، -سازمان بورس اوراق بهادار تهران.2 ه، چاپ اولانحالل شرکت76و 75، ماده "اوراق بهادار ترکیب:اولیهعمومیهايعرضهگذاريقیمت"، )1388(یزدي، محمد، قاسمی مهسا مازارعرب.3 هاي حسابداري و حسابرسی، شماره فصلنامه بررسی"ژنتیکالگوریتمومصنوعیعصبیهايشبکه 102-87، صص 58 ، "بینی ورشکستگیکاربرد الگوریتم ژنتیک در الگوبندي پیش"). 1384(زاده دهکردي، حسن فرج.4 .نامه کارشناسی ارشد حسابداري، دانشکده علوم انسانی دانشگاه تربیت مدرسپایان 5. Adnan aziz.M,Humayon A.Da, (2002). “Predicting Corporate Bankruptcy: Where do we Stand”, Department of Economics, Loughborough University, UK. 6. Altman, E.I., (1968). “Financial Ratios, Discriminate Analysis and the Prediction of Corporate Bankruptcy”, Journal of Finance, 23, 589–609. 7. Beaver, W.H., (1966). "Financial Ratios as Predictors of Failure", Journal of Accounting Research 4, Empirical Research in Accounting: Selected Studies, p. 71-111. 8. Etemadi, H., Rostamy, A., and Dehkordi, H. (2009). "A Genetic Programming Model for Bankruptcy Prediction: Empirical Evidence from Iran", Expert Systems with Applications, 36 (2), p. 3199–3207. 9. Grice, John Stephn, Ingram Robert., (2002). “Tests of the Generalizability of Altman’s bankruptcy Prediction Model”, Journal of Busines Research;Vol. 54,Issue 1:53-61 ١٦ 10. Haber, J., (2006), “Theoretical Dvelopment of Bankruptcy Prediction Variables”, the Journal of Theoretical Accounting Research, 2, 82-101 11. Huang, S., Tsai, C.-F., Yen, D., and Cheng, Y. (2008). "A Hybrid Financial Analysis Model for Business Failure Prediction", Expert Systems with Applications, 35(3), p. 1034–1040. 12. Hung, C., and Chen, J., (2009). "A Selective Ensemble Based on Expected Probabilities for Bankruptcy Prediction", Expert Systems with Applications, 36(3), p. 5297–5303. 13. Kawakami ,Becerra ,seada., (2004). ":Ratio Selection for Classification Models",.Data Mining and Knowledge Discovery ,2004: 8,151-170. 14. Lin, R., Wang, Y. and Wu, C., (2009). "Developing a Business Failure Prediction Model via RST, GRA and CBR", Expert Systems with Applications, 36(2), p. 1593–1600. 15. McKee, T.E. and Lensberg, T. (2002). “Genetic Programming and Rough Sets: a Hybrid Approach to Bankruptcy Classification”, European Journal of Operational Research, 138, 436-51. 16. Min, J., and Jeong, C., (2009). "A Binary Classification Method for Bankruptcy Prediction", Expert Systems with Applications, 36(3), p. 5256–5263. 17. Min, J.H., and Lee, Y.C., (2008). "A Practical Approach to Credit Scoring", Expert Systems with Applications, 35(4), p. 1762–1770. 18. Ravi, V., and Pramodh, C., (2008). "Threshold Accepting Trained Principal Component Neural Network and Feature Subset Selection: Application to Bankruptcy Prediction in Banks", Applied Soft Computing, 8(4), p. 1539–1548. 19. Shin, K. and Lee, Y., (2002). “A Genetic Algorithm Application in Bankruptcy Prediction Modeling”, Expert Systems with Applications, 23 (3), 321-8. 20. Sun, J., and Li, H., (2008). "Listed Companies Financial Distress Prediction Based on Weighted Majority Voting Combination of Multiple Classifiers", Expert Systems with Applications, 35(3), p. 818–827. 21. Tsai, C.F., (2009). "Feature Selection in Bankruptcy Prediction", Knowledge-Based Systems, 22, p. 120–127. 22. Varetto F., (2009). " Genetic Algorithm Applications in the Analysis of Insolvency Risk", Journal of Banking and Finance 22 (1998) 1421–1439 23. Wu, W.W., (2010). "Beyond Business Failure Prediction", Expert Systems with Applications, 37(3), p. 2371–2376. Applying the Linear Genetic Algorithm and Non-linear Genetic to Increasing the Efficiency of Predicting Companies Financial Distress in Capital Market Abstract Bankruptcy is event that challenges the social and economic aspects of country. So if we can obtain information about the possibility of bankruptcy before the actual event, economic and social consequences could have reduced or even prevented. ١٧ Purpose of this study is investigating financial crisis prediction strength of linear genetic algorithm and non-linear genetic algorithm to increase the ability of decision making for financial statement users to predicate the financial distress. Then, respecting the obtained results these pattern are compared with each other and the best pattern is extracted. Based on available information and statistics, of all companies listed in Tehran Stock Exchange during the period ١٣٧٦-١٣٨٩, ٧٢ companies have been subject to Article ١٤١ trade law and ٧٢ companies have not been subject to this Article was elected. Results of Mc-Nemar test for linear genetic algorithm and non-linear genetic algorithm showed that although the predictive accuracy of nonlinear genetic algorithm (٩٠%) is more than of the linear genetic algorithms (٨٠%) but this difference is not statistically significant. Keywords: Financial Distress, Financial Variables, Linear Genetic Algorithms, Non- Linear Genetic Algorithm 
work_ojmv2246mrd67dv5xge5au5jxm	----	Cass Knowledge | Cass Business School Skip to main content Study Business services Faculties and research News and events About us Student Hub Alumni Staff Hub Menu Cass Business School Faculties and research Research at Cass Cass Knowledge About Cass Knowledge InBusiness Student Hub Alumni Staff Hub Research at Cass Actuarial Science and Insurance Finance Management Centres Cass Experts PhD students Honorary Visiting Appointments Faculties and research Research at Cass Cass Knowledge Cass Knowledge Free access to our world-renowned research Find Cass Knowledge articles SearchSubmit search Filter Articles Select TopicAll topics Actuarial Science and Insurance Finance Management Select SeriesAll series Asset Management Charity and Non-Governmental Sector Corporate Finance Corporate Governance and CSR Health and Care Human Resource Management Insurance and Pensions Investment and Risk Management Leadership, Entrepreneurship and Innovation Marketing Operations and Supply Chain Management Professional Services Real Estate Shipping and Transport Strategy Select YearAll years 2021 2020 2019 2018 2017 2016 2015 2014 2013 2012 Select MonthAll months January February March April May June July August September October November December Cass Knowledge offers free access to our world-renowned research output to provide an essential resource for business and the professions. Featured Cass Knowledge Browse all articles Ashwin Karthick Natarajan/ Shutterstock.com FinanceInvestment and Risk Management18th February 2021 Bank Business Model Migrations in Europe - Determinants and Effects This paper investigates the determinants of bank business model… ManagementLeadership, Entrepreneurship and Innovation20th January 2021 The Challenge for Business and Management Degree Courses in the 21st Century Assessing the current state of evaluation methods applied in… Prostock-studio/ Shutterstock.com ManagementMarketing22nd December 2020 Examining the role that attention plays in the purchasing behaviour of consumers New research shows that retailers can influence a person's shopping… All Cass Knowledge whiteMocca/ Shutterstock.com Actuarial Science and InsuranceInsurance and Pensions30th March 2021 Prediction of claims in export credit finance - a comparison of four machine learning techniques This study investigates machine learning techniques for claims… VanderWolf Images /Shutterstock.com FinanceCorporate Finance19th March 2021 Does Air Pollution Harm Business Ethics? The harm air pollution causes to one's heath is well documented.… boonchoke / Shutterstock.com ManagementHuman Resource Management28th February 2021 How common are bad bosses? Many employees may be of the view that bad bosses abound. This… Cass Knowledge Registration Subscribe to our mailing list > If you would like to be kept up to date with news from Cass Knowledge please subscribe to our mailing list. Back to top The Business School (formerly Cass) 106 Bunhill Row London EC1Y 8TZ +44 (0)20 7040 8600 Contact us Find us Legal Cookies Accessibility Connect with Cass Business School Twitter Facebook YouTube LinkedIn Instagram Weibo Triple-crown accreditation © 2019 City, University of London 
work_okkzldu7ozgi5bflm7hthf7sgq	----	Discovering Filter Keywords for Company Name Disambiguation in TwitterI Damiano Spina∗, Julio Gonzalo, Enrique Amigó UNED NLP & IR Group Juan del Rosal, 16 28040 Madrid, Spain http: // nlp. uned. es Abstract A major problem in monitoring the online reputation of companies, brands, and other entities is that entity names are often ambiguous (apple may refer to the company, the fruit, the singer, etc.). The problem is particularly hard in microblogging services such as Twitter, where texts are very short and there is little context to disambiguate. In this paper we address the filtering task of determining, out of a set of tweets that contain a company name, which ones do refer to the company. Our approach relies on the identification of filter keywords: those whose presence in a tweet reliably confirm (positive keywords) or discard (negative keywords) that the tweet refers to the company. We describe an algorithm to extract filter keywords that does not use any previously annotated data about the target company. The algorithm allows to classify 58% of the tweets with 75% accuracy; and those can be used to feed a machine learning algorithm to obtain a complete classification of all tweets with an overall accuracy of 73%. In comparison, a 10-fold validation of the same machine learning algorithm provides an accuracy of 85%, i.e., our unsupervised algorithm has a 14% loss with respect to its supervised counterpart. Our study also shows that (i) filter keywords for Twitter does not directly derive from the public information about the company in the Web: a manual selection of keywords from relevant web sources only covers 15% of the tweets with 86% accuracy; (ii) filter keywords can indeed be a productive way of classifying tweets: the five best possible keywords cover, in average, 28% of the tweets for a company in our test collection. Keywords: Twitter, online reputation management, name disambiguation, filtering IThis research was partially supported by the Spanish Ministry of Education (FPU grant nr AP2009- 0507), the Spanish Ministry of Science and Innovation (Holopedia Project, TIN2010-21128-C02), the Regional Government of Madrid and the ESF under MA2VICMR (S2009/TIC-1542) and the European Community’s FP7 Programme under grant agreement nr 288024 (LiMoSINe). ∗Corresponding author Email addresses: damiano@lsi.uned.es (Damiano Spina), julio@lsi.uned.es (Julio Gonzalo), enrique@lsi.uned.es (Enrique Amigó) Preprint submitted to Elsevier March 6, 2013 http://nlp.uned.es 1. Introduction The vast use of social media to share facts and opinions about entities, such as companies, brands and public figures has generated the opportunity - and the necessity – of managing the online reputation of those entities. Online Reputation Management consists of monitoring – and handling – the opinion of Internet users (also referred to as electronic word of mouth, eWOM) on people, companies and products, and is already a fundamental tool in corporate communication [76, 20, 64, 32]. Over the last years, a wide variety of tools have been developed that facilitate the monitoring task of online reputation managers1. The process followed by these tools typically consists of three tasks: 1. Retrieval of potential mentions. The user gives as input a set of keywords (e.g. the company name, products, CEO’s name. . . ), and the service uses these keywords to retrieve documents and content generated by users from different sources: broad- cast news sources, social media, blogs, site reviews, etc. 2. Analysis of results. Retrieved documents are automatically processed in order to get relevant information to the user: sentiment, authority and influence, back- ground topics, etc. 3. Results visualization. Analyzed data is presented to the user in different ways: ranking documents, drawing graphics, generating tag clouds, etc. A major problem concerning the first task (retrieval of potential mentions) is that brand names are often ambiguous. For instance, the query “Ford” retrieves informa- tion about the motor company, but also might retrieve results about Ford Models (the modeling agency), Tom Ford (the film director), etc. One might think that the query is too general, and the user should provide a more specific query, such as “ford motor” or “ford cars”. In fact, some tools explicitly suggest the user to refine possibly ambiguous queries2. This approach has two main disadvantages: (i) users have to make an additional effort when defining unambiguous queries and (ii) query refinement harms recall over the mentions of the brand in the Web, which can be particularly misleading in an online reputation management scenario. Note that filtering out the mentions that do not refer to the monitored entity is also crucial when estimating its visibility. Quantifying the number of mentions on the Web about an entity, and how this number changes over time, is essential to track marketing or Public Relationship campaigns. When the entity name is ambiguous, indicators given by tools such as Google Trends3 or Topsy4 can be misleading. We think that a component capable of filtering out mentions that do not refer to the entity being monitored (specified by the user as a keyword plus a representative URL) 1At the time of writing this paper, some popular reputation management tools are trackur (http: //www.trackur.com/), BrandsEye(http://www.brandseye.com/), Alterian SM2(http://socialmedia. alterian.com/) or SocialMention (http://socialmention.com), among others. 2Alterian SM2 service: http://www.sdlsm2.com/social-listening-for-business/industry/. 3http://www.google.com/trends 4http://www.topsy.com 2 http://www.trackur.com/ http://www.trackur.com/ http://www.brandseye.com/ http://socialmedia.alterian.com/ http://socialmedia.alterian.com/ http://socialmention.com http://www.sdlsm2.com/social-listening-for-business/industry/ http://www.google.com/trends http://www.topsy.com would be a substantial enhancement of current online reputation management tools, and would also facilitate the analysis of the online presence/visibility of a brand. Among the most important information sources monitored by reputation managers are microblogging services [17, 36], As of today, Twitter5 is the most popular microblog- ging service and provides a communication environment that deserves serious attention as a form of eWOM. Users send short posts (tweets) where they can describe things of interest and express attitudes that they are willing to share with others. The work presented in this paper deals with the challenge addressed by the Online Reputation Management Task of WePS-3 [3]. Given a set of Twitter entries containing an (ambiguous) company name, and given the home page of the company, the task is to filter out irrelevant information, providing a binary classification of tweets as related or unrelated to the company. Notice that ambiguity resolution is particularly challenging in Twitter: tweets are minimal (140 characters at most) and little context is available for resolving name ambiguity. The main objective of this paper is to validate an intuitive observation derived by the set-up and the analysis of the results of the WePS-3 Online Reputation Management Task [3]. The observation is that manual annotation can be simplified by picking up spe- cial keywords –henceforth called filter keywords – that reliably signal positive or negative information. For instance, “ipod” is a positive filter keyword for the Apple company, because its presence is a highly reliable indicator that the tweet is about the company. Reversely, “crumble” is a negative filter keyword for Apple, because it correlates with unrelated tweets. The intuition is that automatic detection of such filter keywords can be a valuable signal to design an automatic solution to the problem. Our goal is to provide quantitative evidence supporting (or rejecting) our intuition, and to answer some related questions: • Where should we look for filter keywords in order to find them automatically? • How much recall can we expect from such keywords? • Can we use the notion of filter keywords to build a binary classifier that solves the task? In order to answer these questions we will use the WePS-3 Task 2 test collection6, which is, to our knowledge, the first dataset built explicitly to address this problem. The paper is organized as follows: in Section 2 we start by introducing some back- ground concepts and discussing the State of the Art: we review some related work on Twitter, we summarize previous work on Named Entity Disambiguation and Automatic Keyword Extraction (which is one of the techniques we want to apply to the problem). Then we focus on the goals and results of the WePS-3 Online Reputation Management Task, which is the test collection used in our experiments. In Section 3 we detail and validate our Filter Keywords hypothesis by computing its upper bounds on our test collection. In Section 4, we study how to automatically discover filter keywords. Then, in Section 5 we discuss how to perform the classification task by using automatically extracted filter keywords. Finally, we discuss our results in Section 6 and conclude in Section 7. 5http://twitter.com 6http://nlp.uned.es/weps/weps-3 3 http://twitter.com http://nlp.uned.es/weps/weps-3 2. Related Work Filtering out mentions that do not refer to a given company name can be seen as a named entity disambiguation problem. Named Entity Disambiguation (NED) have been receiving a lot of attention from the Natural Language Processing (NLP) research community in the last decade [30, 11, 19, 13, 39, 6]. Most recent work tackle this problem in contexts where disambiguation is more challenging, as in the case of Twitter: the textual context is reduced (limited to 140 characters) and language is used as chat-speak or SMS style, where the use of the phonetic spelling and acronyms for common phrases is vast [42, 29]. In this section we provide some background about Twitter and its related datasets. Then we introduce some previous work on NED, focusing on scenarios where NED is crucial such as Entity Linking, Document enrichment by Wikipedia links and Web People Search. Related work about automatic keyword extraction – one of the techniques we want to apply to our problem – is also presented, and then we describe the definition and the results of the WePS-3 Online Reputation Management Task – which is the test collection used in our experiments. 2.1. Twitter Twitter7 is a relatively new social networking site [44] which stands out as the proto- typical microblogging service. Its particularity is that posts do not exceed 140 characters and there are no privacy conditions. Therefore, Twitter reflects opinions in real time and is very sensitive to burstiness phenomena [61, 12, 46]. Tweets are particularly challenging for Natural Language Processing tasks, given that (i) tweets are shorts (i.e. 140 characters), and (ii) users post text using a nonstandard language with similarities to SMS style [42, 47, 29]. 2.1.1. Twitter-related Tasks Up to now, most of the work on Twitter has focused on analyzing the microblogging phenomenon [37, 46, 12], modeling the propagation of information in the social network [78, 44, 79] and analyzing the content of tweets [36, 70, 66, 35, 15]. Some previous work address the problem of tweet classification for different purposes such as sentiment analysis [36, 63] and information filtering [70]. Sriram et al. [70] classify tweets in five categories (i.e. news, events , opinions, deals, and private messages) in order to improve information filtering. They compare different tweet representations, including bag-of-words and features like the author name, the presence of mentions of users on the tweet and features regarding the writing style (i.e. opinionated words, currency and percentage signs). Then, they train a Naive Bayes classification model to evaluate the different representations, obtaining high accuracy scores when combining bag-of-words with the other features. Jansen et al. [36] use a multinomial Bayes model to determine the sentiment of a set of tweets. The model is built using a lexicon of around 200,000 unigrams and bigrams of words and phrases that have a probability distribution to determine the sentiment of a topic. The system calculates the probability of each tweet by checking the occurrence of 7http://www.twitter.com 4 http://www.twitter.com polarized words and then picks the class with the greatest probability in a winner-takes- all scenario. In this work we use a similar approach, in the manner that we test, among others, the winner-takes-all strategy in order to classify all the tweets for a company name either as related or unrelated. Zhao et al. [90] propose an automatic keyphrase extraction model to summarize top- ics in Twitter. Firstly, they identify topics using Twitter-LDA [89], which assumes a single topic assignment for each tweet. Secondly, they use a variation of Topical PageR- ank [48] where edges between two words are weighted taking into account co-occurrences in tweets assigned to the same topic. Then, they generate phrase candidates by looking for combinations of the top topic keywords that co-occur as frequent phrases in the text collection. Finally, topical keyphrases are ranked by using a probabilistic model that takes into account (i) the specificity of a phrase given a topic and (ii) the retweet ratio of tweets containing a keyphrase. To the best of our knowledge, the Online Reputation Management Task on WePS-3 held at CLEF 2010 was the first campaign of NLP-centered tasks over Twitter. TREC 2011 and TREC 2012 held a track about Twitter: TREC Microblog Track8. Here, the problem addressed is a realtime search task (Realtime Adhoc Task): given a query at a specific time, systems should return a list both recent and relevant tweets [67]. 2.1.2. Twitter Datasets There are several Twitter datasets that are suitable for a variety of research purposes. A corpus of 900.000 tweets has been provided by the Content Analysis in Web 2.0 Workshop (CAW 2.0)9, to tackle the problem of text normalization on user generated contents. Yang & Leskovec [80] use a dataset of 580 million Tweets to identify tem- poral patterns over the content published on the tweets, and Kwak et al. [46] use a representative sample of 467 million tweets from 20 million users covering a 7 month period from June 1 2009 to December 31 2009 to study the information diffusion and the topological characteristics of Twitter.10. Cha et al. [14] built a dataset comprising 54,981,152 users, connected to each other by 1,963,263,821 social links and including a total of 1,755,925,520 tweets to analyze users’ influence and how to measure it on the Twittersphere. Finally, the Tweets2011 corpus is the dataset that will be used on the TREC Mi- croblog Track. The corpus is a representative sample of the twittersphere, that includes also spam tweets. It consists of approximately 16 million tweets over a period of 2 weeks (24th January 2011 until 8th February, inclusive), which covers both the time period of the Egyptian revolution and the US Superbowl, among others. To the best of our knowledge, WePS-3 Task 2 test collection (described in Section 2.4), is the first dataset specifically built to address the problem of disambiguation of orga- nization names on tweets. Recently, the RepLab evaluation campaign [4] held in CLEF 2012, has addressed the same problem but in a multilingual scenario: the RepLab dataset contains tweets written in English and Spanish. 8https://sites.google.com/site/microblogtrack 9http://caw2.barcelonamedia.org/node/41 10These datasets are no longer available as per request from Twitter 5 https://sites.google.com/site/microblogtrack http://caw2.barcelonamedia.org/node/41 2.2. Named Entity Disambiguation In Natural Language Processing, Named Entity Disambiguation (NED) is a step fur- ther from Named Entity Recognition. The latter consists on identifying and classifying phrases in a text that refer to people, places or organizations, as well as temporal ex- pressions or numeric quantities [30, 62]. The former involves the association of mentions in one or more texts (also called references or surface forms) of an entity with the con- crete object that they are actually referencing [11]. For instance, in the sentence: “The Big Apple’s new Apple retail store was officially opened today.” the two occurrences of the word Apple refer to two different entities. The former refers to New York City (nicknamed as “The Big Apple”11), while the latter refers to the consumer electronics and software company Apple Inc12. Word Sense Disambiguation (WSD) is another NLP task related to NED, which deals with the polysemy of common words, and consists of assigning the appropriate sense to an occurrence of a word in a given context [1]. While in WSD the senses of a word are often looked up in a dictionary, thesaurus or lexical resource such as WordNet [22], we will see that in NED the use of Wikipedia as knowledge source is widely extended. In the following, we review three main problems –closely related with ours– in which entity disambiguation has been applied: Entity Linking, Document Enrichment by Link- ing to Wikipedia Articles and Web People Search. Finally, we conclude this section describing some previous work about named entity disambiguation on Twitter. 2.2.1. Entity Linking Entity linking consists of associating a mention in a text with the corresponding entity in a Knowledge Base. In the Knowledge Base Population scenario (KBP) [53, 39, 21, 38], this is a first step to discover facts about entities and augmenting a knowledge base with these facts and with newly discovered entities. Given an entity name and a background document where the name occurs, systems typically perform three steps to link the entity to the knowledge base: (i) query expansion (enrich the query by mining the Wikipedia structure or resolving co-reference in the background document); (ii) candidate generation (find all possible entries in the knowledge base that the query might link to) and (iii) candidate ranking (rank candidate entities by computing similarity between the represented query and the entities, and fixing a threshold to decide when the entity does not exist in the knowledge base). 2.2.2. Document Enrichment by Linking to Wikipedia Articles There are a variety of systems that automatically annotate a document by linking names appearing in the document with Wikipedia articles [13, 57, 71, 60, 45]. In general, named entity disambiguation in these systems is carried out in three steps: – Mention or surface form representation. In this step, the context of the mention to disambiguate is defined. The most common representations used are vector space model [13, 57, 19], the set of named entities that occur in the text [27], and the resolved Wikipedia links of unambiguous entities next to the mention [60, 45, 24, 54]. 11http://en.wikipedia.org/wiki/Big_Apple 12http://en.wikipedia.org/wiki/Apple_Inc 6 http://en.wikipedia.org/wiki/Big_Apple http://en.wikipedia.org/wiki/Apple_Inc – Candidates entities retrieval and representation. The system retrieves all possible entities that could be referenced by the mention from the knowledge base (i.e. Wikipedia pages) and represents each entity as a bag-of-words from the page [13, 57, 19], extracting features from the page structure (such as the categories which the page belongs to) [13, 19, 27], or syntactic features [57] or also exploiting the hyperlink structure of Wikipedia, retrieving all pages that link to the candidate entity page [60, 45, 24, 54]. – Best candidate selection. In the final step, the best candidate entity is selected by computing a distance or similarity function between the surface form and each of the candidate entities. The most common functions are cosine [13] and other vector similarity functions [57, 19], random walk graph models to compute semantic relatedness [27] and finally relatedness or coherence functions that involve all the entity links made in the text [60, 45, 24]. Unlike document enrichment, our task focuses on a single organization name and does not require linking every mention of entities in the collection, but just deciding whether each mention refers or not to the entity. 2.2.3. Web People Search Web People Search is defined as the task of clustering a set of web pages, which are the result of a Web search for a person name, in as many groups as entities sharing that name [6]. This was the original goal of the Web People Search campaign (WePS) [9], com- plemented in subsequent editions by the tasks of person attribute extraction [10, 7] and company name disambiguation in Twitter [3]. In the web people search scenario, Hierar- chical Agglomerative Clustering seems to be the most competitive (and most frequently used) clustering technique [28, 9, 10, 7]. Documents are typically represented as bag-of- words [11, 40, 9]. Other works use smaller portions of the document, such as sentences where the ambiguous name occurs or pre-defined windows of words [28, 49, 10]. Named entities are also a frequently used feature for people name disambiguation [40, 9, 10], and biographical features (e.g. title, organization, email address, phone number, etc.) are used to a lesser extent [50, 2, 75]. The most common similarity measure in this scenario is cosine [11, 9, 10], while some works also use Kullback-Leibler divergence [28, 52]. 2.2.4. Named Entity Disambiguation in Twitter Most Named Entity Disambiguation techniques have been applied to disambiguate entities on reasonably long texts such as news articles or blog posts. However, little work has been done on NED over microblogging posts [24, 54, 55]. In this scenario, disambiguation is harder due to fact that texts are short (limited by 140 characters) and hence the context of a mention is minimal. Ferragina & Scaiella [24] propose a system capable of annotating short texts (includ- ing tweets) by linking entities to Wikipedia pages. The system relies on the hyperlink structure of Wikipedia, exploiting the links between pages and the anchors texts of the links. When the system receives a text, it detects the anchors on the text and retrieve the possible senses for each anchor. Ambiguous senses are disambiguated by a collective agreement function among all senses associated to the anchors detected on the text [45], and taking advantage of the unambiguous anchors to boost the selection of these senses for the ambiguous anchors [60]. They do not evaluate the accuracy of their disambigua- tion component over tweets. However, they report that almost 95% of the 5,000 analyzed tweets have at least 3 phrases with an entry in Wikipedia (which is not necessarily an 7 entity). This result shows that Wikipedia have a high coverage as a catalog of senses for tweet disambiguation. Different from [24], Meij et al. [54] uses supervised machine learning techniques to refine a list of candidate Wikipedia concepts that are potentially relevant to a given tweet. The candidate ranking list is generated by matching n-grams in the tweet with anchor texts in Wikipedia articles, taking into account the inter-Wikipedia link structure to compute the most probable Wikipedia concept for each n-gram. Michelson & Macskassy [55] focus on discovering the topics of interest for a particular Twitter user. Given the stream of tweets corresponding to the user, they firstly find the Wikipedia page of the entities mentioned on tweets and secondly they build a topic profile from the high-level categories that cover these entities. Entity disambiguation is performed by calculating the overlap between the terms on the tweet and the terms on the page of each candidate entity. In this scenario, the accuracy of the disambiguation process is not critical, since the system takes into account the categories that occur frequently across all the entities found in order to produce a topic profile of a stream. 2.3. Automatic Keyphrase Extraction We include here an overview of Automatic Keyphrase Extraction techniques, as this is a technique that plays a crucial role in our approach to company name disambiguation on Twitter. Automatic Keyphrase Extraction is the task of identifying a set of relevant terms or phrases that summarize and characterize one or more given documents [77]. Most of the literature about automatic keyphrase extraction is focused on (well-written) technical documents, such as scientific and medical articles, since the keywords given by the authors can be used as gold standard [74, 77, 25, 59, 58, 43]. Some authors address automatic keyword extraction as a way of automatically summarizating web sites [86, 87, 88, 85] . In [85] different keyword extraction methods are compared, including tf*idf, supervised methods, and heuristics based on both statistical and linguistic features of candidate terms. While Zhang et al. [85] study the automatic keyword extraction from website descriptions, in our work we explore a semi-supervised keyword extraction approach that extract filter keywords over a stream of tweets. Automatic keyword extraction is typically used to characterize the content of one or more documents, using features intrinsically associated with that documents. In order to detect both positive and negative filter keywords, however, we need to look into external resources in order to discriminate between related and unrelated keywords. Thus, automatic keyword extraction methods are not directly applicable to our filter keyword approach. 2.4. The WePS-3 Online Reputation Management Task Disambiguation of company names in text streams (and in particular in microblog posts) is a necessary step in the monitoring of opinions about a company. However, it is not tackled explicitly in most research on the subject. Rather than this, most previous work assume that query terms are not ambiguous in the retrieval process. The disambiguation task has been explicitly addressed in the WePS-3 evaluation campaign [3]. In this section we summarize the outcome of that campaign, analyzing the test collection and comparing system results. 8 2.4.1. Task Definition and Test Collection The Online Reputation Management task of the WePS-3 evaluation campaign consists of filtering Twitter posts containing a given company name, depending on whether the post is actually related with the company or not. The task is defined to deal with the scenario where an online system accepts any company name as input, and has to learn to disambiguate entries about that company without supervision (i.e. without a set of previously disambiguated tweets). Therefore, the set of organization names in the training corpora is different from the set of companies in the test set. For each organization in the dataset, systems are provided with the company name c (e.g. apple) used as query to retrieve the stream of tweets to annotate Tc, and a representative URL (e.g. http://www.apple.com) that univocally identifies the target company. The input information per tweet consists of a tuple containing: the tweet iden- tifier, the organization name, the query used to retrieve the tweet, the author identifier, the date and the tweet content. Systems must label each tweet as related (i.e. the tweet refers to the company) or unrelated (the tweet does not refer to the given company). The WePS-3 ORM task dataset comprises 52 training and 47 test cases, each of them including a company name, its URL, and an average of 435 tweets manually annotated as related/unrelated to the company. The dataset has been annotated using the Mechan- ical Turk crowdsourcing marketplace13, with the following manual annotation options: related, unrelated and undecidable. Each hit has been redundantly annotated by five Me- chanical Turk workers. A total of 902 annotators have participated in the annotations of 43730 tweets. Finally, an agreement analysis was done in order to decide the final annotation for each tweet. An interesting property of the dataset is that there is a great variability of the degree of ambiguity across the training and test cases. That is, there are companies with low occurrence in tweets (e.g. Delta Holdings, Zoo Entertainment), companies with medium ambiguity (e.g. Luxor Hotel and Casino, Edmunds.com) and companies with high presence in tweets (e.g. Yamaha, Lufthansa). 2.4.2. WePS-3 Results A total of five research groups participated in the campaign. The best two systems were LSIR [81] and ITC-UT [83]. The LSIR system builds a set of profiles for each company, made of keywords extracted from external resources such as WordNet or the company homepage, as well as a set of manually defined keywords for the company and the most frequent unrelated senses for the company name. These profiles are used to extract tweet-specific features that are added to other generic features that give infor- mation about the quality of the profiles to label the tweets as related or unrelated with an SVM classifier. The ITC-UT system is based on a two step classification. Firstly, it predicts the class of each query/company name according to the ratio of related tweets of each company name and secondly applies a different heuristic for each class, basically based on the PoS tagging and the named entity label of the company name. The SINAI system [26] also uses a set of heuristic rules based on the occurrence of named entities both on the tweets and on external resources like Wikipedia, DBPedia 13https://www.mturk.com 9 https://www.mturk.com and the company homepage. The UvA system [72] does not employ any resource related to the company, but uses features that involve the use of the language in the collection of tweets (URLs, hashtags, capital characters, punctuation, etc.). Finally, the KALMAR system [41] builds an initial model based on the terms extracted from the homepage to label a seed of tweets and then uses them in a bootstrapping process, computing the point-wise mutual information between the word and the target’s label. In the WePS exercise, accuracy (ratio of correctly classified tweets) was used to rank systems. The best overall system (LSIR) obtained 0.83, but including manually produced filter keywords. The best automatic system (ITC-UT) reaches an accuracy of 0.75 (being 0.5 the accuracy of a random classification), and includes a query classification step in order to predict the ratio of positive/negative tweets. 2.4.3. Other work using WePS-3 datasets Recently, Yerva et al. [82] have explored the impact of extending the company profiles presented in [81] by using the related ratio and considering new tweets retrieved from the Twitter stream. By estimating the degree of ambiguity per entity from a subset of 50 tweets per entity, they reach 0.73 accuracy14. Using this related ratio and considering co-occurrences with a given company profile, the original profile is extended with new terms extracted from tweets retrieved by querying the company name in Twitter. The expanded profile outperforms the original, achieving 0.81 accuracy. Zhang et al. [84] presents a two stages based disambiguation system. They combine supervised and semi-supervised methods. Firstly, they train a generic classifier using the training set, similarly to [81]. Then, they use this classifier to annotate a seed of tweets in the test set, using the Label Propagation algorithm to annotate the remainder tweets in the test set. Using Näıve Bayes and Label Propagation they achieve a 0.75 accuracy, that matches the performance of the best automatic system in WePS-3. The RepLab evaluation campaign held at CLEF 2012 addressed, as a subtask, the same filtering problem introduced in WePS-3. The main difference was that while WePS- 3 dataset contains only tweets written in English, the RepLab collection [4] contains tweets both in English and Spanish. Finally, the WePS-3 ORM task dataset has been extended with manual annotations for the task of entity profiling in microblog posts, that includes identifying entity aspects and opinion targets [69]. 2.5. Wrap Up There are two main findings on entity name disambiguation that motivate our re- search: 1. Use of filter keywords. Artiles et al. [8] studied the impact of query refinement in the Web People Search clustering task and concluded that “although in most occasions there is an expression that can be used as a near-perfect refinement to retrieve all and only those documents referring to an individual, the nature of these ideal refinements is unpredictable and very unlikely to be hypothesized by the user” [8] 14Note that, unlike the original formulation of the WePS-3 task, this is a supervised system, as it uses part of the test set for training. Hence, their results cannot be directly compared with the results in our work. 10 This is both a positive indication – there are keywords able to isolate relevant information well – and a suggestion that finding optimal keywords automatically might be a challenging task. Another evidence in favor of using filter keywords is that the best results on the WePS-3 ORM Task were achieved by a system that used a set of manually produced –both positive and negative– filter keywords [81]. One of the goals of our work is to analyze query refinement in the scenario of the Company Name Disambiguation on Twitter. In other words, we explore the impact of defining a set of keywords to filter both related and unrelated tweets to a given company. 2. Use of knowledge bases to represent candidate entities. Both entity linking and document enrichment systems use knowledge bases in order to characterize the possible entities that a mention may refer to. Most systems use Wikipedia as knowledge base [57, 13, 19, 27, 60, 45, 24, 39, 21]. In this direction, we believe that looking at Wikipedia pages related to the company to disambiguate could give useful information to characterize positive filter keywords. We also explore the use of other resources such as Open Directory Project (ODP), the company website and the Web in general. Note, however, that not every company is listed in ODP or has an entry in Wikipedia. 11 3. Potential Benefits of Filter Keywords for Name Disambiguation in Twitter A positive/negative filter keyword is an expression that, if present in a tweet, indicates a high probability that the tweet is related/unrelated to the company. In this section, we investigate the upper bound of the filter keyword strategy, in two ways: firstly, we con- sider manual annotations in the WePS-3 collection to derive oracle (optimal) keywords; and secondly, manually extracting keywords from representative Web pages about the company. Finally, we study how to use these filter keywords to solve the WePS-3 ORM task. 3.1. Upper Bound Performance of Filter Keywords The most useful filter keywords are those with a high coverage, i.e., those which appear in as many tweets as possible. As all tweets in the WePS-3 collection are manually annotated as related/unrelated to their respective company name, we can find exactly how many filter keywords there are (by definition, filter keywords are those terms that only appear in either the positive or the negative tweets), and how much recall they provide. Figure 1 shows the coverage of the first n filter keywords (for n = 1 . . . 20) in the test collection. Coverage at step n is the proportion of tweets covered by adding the keyword that filters more tweets among those which were not still covered by the first n−1 keywords. We will hereafter refer to this optimal keyword selection as oracle keywords. Figure 1: Upper bound of the filter keywords strategy. The graph shows that, in average, the best five oracle keywords cover around 30% of the tweets, and the best ten cover around 40% of the tweets. Note that only five discriminative terms directly cover, in average, 130 out of 435 tweets in each stream, and those could in turn be used to build a supervised classifier for the rest of the tweet stream. This indicates that filter keywords are, potentially, a relevant source of information to address the problem. In principle, the natural place to find filter keywords is the Web: the company’s web domain and reference to this domain in Wikipedia, ODP, etc. and the Web at large. Using the company’s URL and web search results for the company name, we performed 12 a manual selection of positive and negative keywords for all the companies in the WePS-3 corpus. Note that the annotator inspected pages in the web search results, but did not have access to the tweets in the corpus. Tables 1 and 2 show some examples for positive and negative keywords, respectively. Note that in the set of Oracle keywords there are expressions that a human would hardly choose to describe a company (at least, without previously analyzing the tweet stream). For instance, the best positive oracle keywords for the Fox Entertainment Group do not include intuitive keywords such as tv or broadcast; instead, they include keywords closer to breaking news (leader, denouncing, etc.). Company name Oracle Positive Keywords Manual Positive Keywords amazon sale, books, deal, deals, gift electronics, apparel, books, computers, buy apple gizmodo, ipod, microsoft, itunes, auction store, ipad, mac, iphone, com- puter fox money, weather, leader, de- nouncing, viewers tv, broadcast, shows, episodes, fringe, bones kiss fair, rock, concert, allesegretti, stanley tour, discography, lyrics, band, rockers, make up design Table 1: Differences between oracle and manual positive keywords for some of the com- pany names on the test collection. Looking at negative keywords (Table 2), we can find occasional oracle keywords that are closely related with the vocabulary used in microblogging services, such as followdaibosyu, nowplaying or wanna, while intuitive manual keywords like wildlife for jaguar are unlikely to occur in the Twitter collection. Company name Oracle Negative Keywords Manual Negative Keywords amazon followdaibosyu, pest, plug, brothers, pirotta river, rainforest, deforestation, bolivian, brazilian apple juice, pie, fruit, tea, fiona fruit, diet, crunch, pie, recipe fox megan, matthew, lazy, valley, michael animal, terrier, hunting, Volk- swagen, racing kiss hug, nowplaying, lips, day, wanna french, Meg Kevin, bang bang, Ryan Kline Table 2: Differences between oracle and manual negative keywords for some of the com- pany names on the test collection. Remarkably, manual keywords extracted from the Web (around 10 per company) only reach 15% coverage of the tweets (compare with 40% coverage using 10 oracle keywords extracted from the tweet stream), with an accuracy of 0.86 (which is lower than expected for manually selected filter keywords). This seems an indication that the vocabulary and topics of microblogging are different from those found in the Web. Our experiments in Section 4 corroborate this finding. 13 3.2. Using Keywords to Solve The Task So far, we have seen that keywords do not cover all tweets in the collection, but a part of them. Then, we need an additional step in order to complete the task: given a seed of tweets, annotate the tweets that remain uncovered by filter keywords on each test case. To this aim, we use a standard bootstrapping method. Tweets are represented as bag-of-words (produced after tokenization, lowercase and stop word removal) and term occurrence is used as weighting function; then we have employed a C4.5 Decision Tree classification model15 [65] –with its default parameters– using the implementation provided by the Rapidminer toolkit [56]. For each stream, we use the tweets retrieved by the keywords as seed (training set) in order to classify automatically the rest of tweets. Table 3 displays results for different amounts of filter keywords: the bootstrapping strategy ranges from 0.81 (with 5 keywords) up to 0.87 with 20 keywords. On the other hand, using the tweets covered by manual keywords as training set, the bootstrapping achieves only a 0.67 accuracy. keyword seed set bootstrapping selection strategy coverage accuracy overall accuracy 5 oracle keywords 28% 1.00 0.81 10 oracle keywords 40% 1.00 0.85 15 oracle keywords 47% 1.00 0.86 20 oracle keywords 53% 1.00 0.87 manual keywords 15% 0.86 0.67 Table 3: Quality of seed sets and the bootstrapping classification strategy when applying oracle/manual filter keywords. 15We also tried with other machine learning methods, such as linear SVM and Naive Bayes, obtaining similar results; therefore we only report results on C4.5 for the sake of clarity. 14 ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● avg: 0.67 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 (a) manual keywords ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● avg: 0.87 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 (b) 20 oracle keywords Figure 2: Fingerprints for the bootstrapping classification strategy when applying manual keywords (left) or 20 oracle keywords (right). In order to better understand the results, Figure 2 shows the fingerprint represen- tation [68] for manual keywords and 20 oracle keywords. This visualization technique consists of displaying the accuracy of the system (vertical axis) for each company/test case (dots) vs. the ratio of related (positive) tweets for the company (horizontal axis). The three basic baselines (all true, all negative and random) are represented as three fixed lines: y = x, y = 1 − x and y = 0.5, respectively. The fingerprint visualization method helps in understanding and comparing systems’ behavior, specially when class skews are variable over different test case. Using 20 oracle keywords (see Figure 2b), the obtained average accuracy is 0.87. The fingerprint shows that the improvement resides in cases with a related ratio around 0.5, i.e. the cases where it is more likely to have enough training samples for both classes. Manual keywords, on the other hand, lead to annotations that tend to stick to the ”all related” or ”all unrelated” baselines, which indicates that they tend to describe only one class, and then the learning process is biased. In summary, our results validate the idea that filter keywords can be a powerful tool in our filtering task, but also suggest that they will not be easy to find: descriptive web sources that can be attributed to the company do not lead to the keywords that are most useful or accurate in the Twitter domain. 15 4. Automatic Discovery of Filter Keywords Our next step is now is to discover automatically the terms which are most strongly associated to the company name (positive filter keywords) and to the alternative meanings of the company name (negative filter keywords), and to discard those which are not discriminative (skip terms). For this, we take a machine learning approach, in which training data corresponds to a set of company names and test data to a different set of companies. Thus, the learning process must be able to generalize across companies. Each term is represented by features that take into account the company’s website, Wikipedia, ODP, the Web at a large and the WePS-3 collection itself. In this section we start discussing the features that we propose to represent terms; then, we perform a statistical analysis of the features, and finally we report the results of experiments with our dataset. 4.1. Term Features Table 4 summarizes the notation that we will use in this section to describe the features. We will only work with terms which are not stop words and appear at least in five different tweets on the set Tc, given a company c. Item Description t,ti term c (ambiguous) name that identifies a company (e.g. jaguar) T set of tweets in the WePS-3 collectiona Tc set of tweets in the collection for a given company name c. dft(Tc) document frequency of term t in the collection Tc. dfweb(q) number of total hits returned by the Yahoo! Search BOSS API (http://developer.yahoo.com/search/boss/) for the query q. M an approximation of the size of the search engine index (30 · 109). domainc domain of the website given as reference for the company c. wikipedia(q) set of Wikipedia pages returned by the MediaWiki API (http://www.mediawiki.org/wiki/API) for the query q. dmoz(q) set of items (composed by an URL, a title, a description and a category) returned by searching q on the Open Directory Project (http://www.dmoz.org/search) a For each company name, only the dataset to which the company belongs is used (either training or test). Table 4: Notation used to describe the features used to represent terms. For each of these terms, we have considered 18 features grouped in three classes: 1. Collection-based features (col * ): Terms that co-occur frequently with the (am- biguous) company name c, or terms written as hashtags should have more proba- bility to be (positive/negative) keywords than others. These features combine in- formation about the occurrence of the term in the collection: document frequency 16 http://developer.yahoo.com/search/boss/ http://www.mediawiki.org/wiki/API http://www.dmoz.org/search in the whole corpus, document frequency in the set of tweets for the company, how many times the term occurs as a twitter hashtag, and the average distance between the term and the company name in the tweets. (a) col c df : Normalized document frequency in the collection of the tweets for the company Tc: dft(Tc) |Tc| (1) (b) col c specificity: Ratio of document frequency in the tweets for the company Tc over the document frequency in the whole corpus T: dft(Tc) dft(T) (2) (c) col hashtag: Number of occurrences of the term as a hashtag (e.g. #jobs, #football) in Tc. (d) col c prox avg, col c prox sd, col c prox median: Mean, standard de- viation and median of the distance (number of terms) between the term and the company name in the tweets. 2. Web-based features (web * ): These features are computed from information about the term in the all the Web (approximated by search counts), the website of the company, Wikipedia and the Open Directory Project (ODP).16 (a) web c assoc: Intuitively, a term which is close to the company name has more chances to be a keyword (either positive or negative) than more generic terms. This feature represents the association, according to the search counts, between the term t and a company name c. dfweb(t OR c)/dfweb(c) dfweb(t)/M (3) (b) web c ngd: The Normalized Google Distance [16] (applied to the Yahoo! search engine), which is a measure of semantic distance between two terms from the search counts. Then, for a term t and a company name c, the Google Normalized Distance is given by (4): max(log(f(c)), log(f(t))) − log(f(t AND c)) M − min(log(f(t)), log(f(c))) (4) where f(x) = dfweb(x) (c) web dom df : Frequent terms in the company website should be meaningful to characterize positive keywords. web dom df is the normalized document frequency of the term in the website of the company. dfweb(t AND site:domainc) dfweb(site:domainc) (5) 16Some companies in the Weps-3 collection have a Wikipedia page as reference page instead of the company website. In these cases, the feature web dom df (that is also the numerator in the feature web dom assoc) is computed as the presence of the term t in the Wikipedia page. Also, the query used to get the values of the features web odp occ and web wiki occ is the title of the Wikipedia page. 17 (d) web dom assoc: degree of association of the term with the website in com- parison with the use of the term in the web. This feature is analogous to web c assoc, using the website domain instead of the company name c. dfweb(t AND site:domainc)/dfweb(site:domainc) dfweb(t)/M (6) (e) web odp occ: Number of occurrences of the term in all the items in dmoz (domainc). Each item is composed by an URL, a title, a description and the ODP category to which it belongs. (f) web wiki occ: Number of occurrences of the term in the first 100 results in wikipedia(domainc). In order to filter pages returned by the API that could be unrelated to the company, only pages that contain the string domainc are considered. 3. Features expanded with co-occurrence: In order to avoid false zeros in web- based features, we expand some of the previous term features with the value ob- tained by the five most co-occurrent terms. Given a feature f, a new feature is computed as the Euclidean norm (7) of the vector with components fti ∗w(t,ti) for the five most co-occurrent terms with t in the set of tweets Tc (8), where fti is the web-based feature value f for the term ti and w(t,ti) is the the grade of co-occurrence of each term (9): cooc agg (t,f) = √ ∑ i∈cooct (f(ti) ∗w(ti))2 (7) cooct = set of the five terms which most co-occur with t (8) w(t,ti) = |co-occurrencesTc (t,ti)| |Tc| (9) f(ti) = value of the feature f for the term ti This formula is applied to web c assoc, web c ngd, web dom df, web dom assoc, web odp occ and web wiki occ, resulting in the features enumerated below: (a) cooc c assoc = cooc agg (t, web c assoc) (b) cooc c ngd = cooc agg (t, web c ngd) (c) cooc dom df = cooc agg (t, web dom df) (d) cooc dom assoc = cooc agg (t, web dom assoc) (e) cooc odp occ = cooc agg (t, web odp occ) (f) cooc wiki occ = cooc agg (t, web wiki occ) 4.2. Feature Analysis The first step for the feature analysis is to develop a gold standard set of positive and negative keywords. In order to get sufficient training data and to deal with possible miss-annotations in the corpus, we set a precision of 0.85 of a term in a related/unrelated set of tweets as a feasible threshold to annotate a term as a keyword. Those terms with precision lower than 0.85 in both classes are labeled as skip terms (10): 18 label (term) =   positive if |related(Tc)| |Tc| > 0.85 negative if |unrelated(Tc)| |Tc| > 0.85 skip otherwise (10) where related(Tc) and unrelated(Tc) are respectively the set of the tweets annotated as related and unrelated in the collection Tc. Labeling all suitable terms of the WePS-3 training dataset we end up with a total of 6410 terms, where 34% were labeled as positive keywords, 44% as negative keywords and the remaining 22% as skip. The test dataset, on the other hand, produces a total of 4653 candidate terms, where 33% were labeled as positive keywords, 40% as negative keywords and 27% as skip. In order to study feature behavior, we calculate the distribution of each feature in the three classes: positive, negative and skip. We rely on box/whisker plots to show these distributions and differences or similarities between classes (see Figure 3). Each box/whisker plot shows the distribution of values of a feature for the three classes. The bottom and top of the box are the 25th and 75th percentile (the Q1 and Q3 quartiles, respectively), and the band near the middle is the 50th percentile (the median, Q2). The whiskers extend to the most extreme data point (1.5 times the length of the box away from the box: −1.5 · IQR and 1.5 · IQR, where IQR = |Q3 −Q1|). 19 keyword col_c_df 0.015 0.020 0.025 0.030 0.035 0.040 0.045 positive negative skip (a) col c df keyword col_c_specificity 10 20 30 40 positive negative skip (b) col c specificity keyword col_hashtag 0 positive negative skip (c) col hashtag 2.5 5.0 7.5 positive negative skip label_0.85 col_c_proximity_median (d) col c prox median 2 4 6 8 positive negative skip label_0.85 col_c_proximity_avg (e) col c prox avg 0 2 4 6 positive negative skip label_0.85 col_c_proximity_sd (f) col c prox sd keyword web_odp_occ 0 positive negative skip (g) web odp occ keyword web_wiki_occ 0 50 100 150 positive negative skip (h) web wiki occ label_0.85 cooc_odp_occ 0 positive negative skip (i) cooc odp occ label_0.85 cooc_wiki_occ 0 100 200 300 400 500 positive negative skip (j) cooc wiki occ keyword web_c_assoc 50 100 150 200 positive negative skip (k) web c assoc keyword web_c_ngd 0.35 0.40 0.45 0.50 0.55 0.60 positive negative skip (l) web c ngd keyword web_dom_df 0.00 0.05 0.10 0.15 0.20 0.25 positive negative skip (m) web dom df keyword web_dom_assoc 0 20 40 60 positive negative skip (n) web dom assoc keyword cooc_c_assoc 100 200 300 400 500 600 positive negative skip (o) cooc c assoc keyword cooc_c_ngd 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 positive negative skip (p) cooc c ngd keyword cooc_dom_df 0.0 0.1 0.2 0.3 0.4 0.5 0.6 positive negative skip (q) cooc dom df keyword cooc_dom_assoc 0 100 200 300 400 positive negative skip (r) cooc dom assoc Figure 3: Box-plots representing the distribution of each of the features in the positive, negative and skip classes. The bottom and top of the box are the Q1 and Q3 quartiles, respectively, and the band near the middle is the Q2 quartile –i.e., the median. The whiskers extend to the most extreme data point (1.5 times the length of the box away from the box: −1.5 · IQR and 1.5 · IQR, where IQR = |Q3 −Q1|).20 These plots help visualizing the range of values for each feature, as well as where most of the values lie, allowing for a qualitative analysis of the features. We can see that features col hashtag (Fig. 3c), web odp occ (Fig. 3g) and cooc odp occ (Fig. 3i) are not informative, because almost all of their values are zero. There are less than 1% of the terms in the test set that occur at least one time as hashtag. Also, less than 1% are terms that appear in descriptions and titles of ODP search results. Features describing term - company distance seem to capture differences between keyword and skip terms: both negative and positive keywords, generally occur closer to the company name than skip terms. While positive and negative keywords share similar median and standard deviation (Figs. 3d and 3f) of proximity to the company name, average distance for positive keywords is slightly smaller than for negative keywords (Fig. 3e). Features col c df, col c specificity, web c assoc, web c ngd and their expanded (by co-ocurrence) versions cooc c assoc and cooc c ngd were defined to discriminate filter keywords from skip terms. The most discriminative feature seems to be col c specificity (Fig. 3b). On the other hand, features web wiki occ, web dom df, web dom assoc, cooc dom df,cooc dom assoc and cooc wiki occ were designed to distinguish between positive and negative filter key- words. At a first glance, positive and negative keywords have different distributions in all the features. Skip terms, on the other hand, tend to have distributions similar to those of positive keywords. The features cooc dom assoc (Fig. 3r) and cooc wiki occ (Fig. 3j) seem to be the best to discriminate positive keywords from negative and skip terms. Remarkably, features expanded by co-occurrence seem to be more informative than the original features, which tend to concentrate on low values (the median is near zero). When expanding the original values by co-occurrence, positive terms receive higher values more consistently. In order to quantitatively evaluate the quality of features, we compute the Mann- Whitney U test [51], which is a non-parametric test used in statistical feature selection when a normal distribution of the features cannot be assumed. The p-value could be used to rank the features, since the smaller the value of the p-value, the more informative the feature is [31]. 21 Filter keywords vs. Skip Terms Positive vs Negative Filter Keywords p-value rank p-value rank col c df 2.11e-19 7 4.55e-31 8 col c specificity 8.18e-50 1 4.40e-22 9 col hashtag 1.49e-02 15 5.40e-04 12 col c prox sd 1.87e-33 2 4.20e-02 15 col c prox avg 2.03e-20 6 4.18e-02 14 col c prox median 5.79e-14 8 1.62e-01 16 web c assoc 4.76e-06 12 8.39e-19 10 web dom df 6.67e-30 4 5.33e-92 6 web dom assoc 7.14e-14 9 7.01e-138 4 web c ngd 5.12e-12 10 3.38e-08 11 web odp occ 1.96e-01 16 3.63e-01 18 web wiki occ 1.68e-20 5 4.86e-115 5 cooc c assoc 2.91e-30 3 2.04e-54 7 cooc dom df 1.53e-05 13 1.42e-189 3 cooc dom assoc 8.35e-01 18 7.27e-233 1 cooc c ngd 3.54e-07 11 2.41e-01 17 cooc odp occ 3.15e-01 17 3.92e-03 13 cooc wiki occ 2.36e-03 14 2.13e-211 2 Table 5: U test p-value and ranking position of the features, comparing filter keywords (both positive and negative) with skip terms and comparing positive with negative filter keywords. Table 5 shows the p-value and the rank of each feature for the U test. The most remarkable aspect of this table is that – in agreement with the boxplot analysis – the col c specificity feature discriminates between filter keywords and skip terms better than other features. In addition, the feature cooc dom assoc, which measures the association between the term and the company website, is the best feature to discriminate between positive and negative filter keywords. These results confirm our assumptions that salient terms in the set of tweets of the company tend to be discriminative and salient terms associated with the company in tweets are also associated with the company website. Although the features analyzed above are signals that help differentiating between positive, negative and skip terms, it seems that the vocabulary that characterizes a company in microblog streams is different from the vocabulary associated to the company in its website, in ODP entries or in Wikipedia. 4.3. Keyword Discovery The features described above have been combined in three different ways to automat- ically discover filter keywords. The first one, (machine learning-all features ), consists of training a positive-negative-skip classifier over the training corpus in WePS-3 by using all the features. We combine two classifiers: positive versus others and negative ver- sus others, using the confidence thresholds learned by the classifiers. Terms which are simultaneously under/over both thresholds are tagged as skip terms. The second approach, (heuristic), is inspired by the analysis of the signal provided by each of the features (in the previous section). It is a simple heuristic which looks 22 at only the best two features according to the Mann-Whitney U test: first, we define a threshold to remove skip terms according to the specificity w.r.t. the collection of the tweets for the company (col c specificity feature). Then we state, for the feature that measures association with the website (cooc dom assoc feature) a lower bound to capture positive filter keywords and an upper bound to capture negative filter keywords. These three thresholds have been manually optimized using the training data set. Finally, we also explored a third option: we apply machine learning using only the best two features instead of the whole feature set. We will refer hereafter to this method as machine learning-2features. We have experimented with several machine learning methods using Rapidminer [56]: Neural Nets, C4.5 and CART Decision Trees, Linea Support Vector MAchines (SVM) and Naive Bayes. All methods have been used with “out-of-the-box” parameters. All the terms labeled over the WePS-3 training dataset were used to train the models. In the same way, terms extracted from the test dataset were used as test set. Table 6 shows the values of the Area Under the ROC Curve (AUC) of each of the binary classifiers evaluated. AUC is an appropriate metric to measure the quality of binary classification models independently of the confidence threshold [23]. We analyzed three different subsets of features to represent the terms: (i) using all but the six features expanded by co-occurrence, (ii) using only the best two features (those used by the heuristic and machine learning-2features classifiers), and (iii) using all the features. machine learning algorithm not expanded by co-occurrence features 2 best features all features pos neg pos neg pos neg Neural Net 0.68 0.67 0.73 0.72 0.75 0.73 CART Dec. Trees 0.58 0.61 0.72 0.71 0.63 0.64 Linear SVM 0.50 0.50 0.73 0.71 0.50 0.50 Näıve Bayes 0.64 0.64 0.71 0.71 0.72 0.72 C4.5 Dec.Trees 0.50 0.61 0.50 0.50 0.59 0.66 Table 6: Area Under the ROC Curve (AUC) values of the five classification models and the three feature sets used to classify positives and negatives keywords. The results obtained are similar for all models, except for C4.5 and SVM that in sev- eral cases do not provide any useful information for classification (AUC = 0.5). Keeping out the “expanded by co-occurrence“ features, the performance is, in general, lower for all the algorithms. This corroborates the results of our previous feature analysis. In the following experiments, we focus on the Neural Net algorithm to train both (positive versus others and negative versus other) classifiers, because it is consistently the best performing algorithm according to the AUC measure. For each of the feature combinations described at the beginning of this section (ma- chine learning-all features, heuristic and machine learning-best 2 features ), below we analyze the obtained results. The methods were trained using terms from the WePS-3 training dataset and evaluated with the WePS-3 test set. Table 7 shows the confusion matrix obtained for the machine learning-all features method. The precision for the positive and negative classes is 62% and 56%, respectively, 23 while recall is 52% and 72%. In order to obtain a few representative keywords, this recall levels are good enough; but the precision may compromise the final accuracy of the filtering process. Actual class positive negative skip Class Precision positive 790 190 304 62% Predicted class negative 483 1334 583 56% skip 242 330 375 40% Class Recall 52% 72% 30% Table 7: Confusion matrix for the machine learning-all features classifier. Table 8 shows the confusion matrix for the heuristic method, that only uses the col c specificity and cooc dom assoc features. This method is more precision-oriented than machine learning - all features : precision values of positive and negative class are higher (68% and 75%), but recall is significantly lower (26% and 19%). Actual class positive negative skip Class Precision positive 391 60 122 68% Predicted class negative 23 352 94 75% skip 1102 1453 1056 29% Class Recall 26% 19% 83.02% Table 8: Confusion matrix for the heuristic classifier. Finally, Table 9 shows the contingency matrix for the machine learning-2features method, that represents terms with the features col c specificity and cooc dom assoc and uses the Neural Net machine learning algorithm to build the model. Actual class positive negative skip Class Precision positive 438 102 139 65% Predicted class negative 85 399 101 68% skip 993 1364 1032 30% Class Recall 29% 21% 81% Table 9: Confusion matrix for the machine learning-2features classifier. As expected, its performance lies between machine learning-all features and heuristic methods, with a precision higher than the former (65% and 68%) and a recall higher than the latter (29% and 21%). These results indicate that automatic detection of keywords is plausible and challeng- ing at the same time. Which of the three approaches is better for our problem depends on their performance on the final task: tweet classification. In the next section we explore how to use these filter keywords to classify tweets. 5. Automatic Tweet Classification After detecting the filter keywords automatically, we directly classify the subset of tweets that contain only negative or only positive keywords. Then, the bootstrapping 24 strategy described in Section 3.2 is used to complete the task. Ii is also interesting to compare the bootstrapping strategy with the näıve winner- takes-all baseline –that directly classifies all the tweets as related or unrelated depending on which is the dominant class in the seed of tweets– and the winner-takes-remainder strategy, which consists of applying the winner-takes-all strategy only to those tweets that were not covered by some of the filter keywords. Figure 4: Fingerprints for each of the keyword selection strategies combined with each of the different tweet classification strategies. Figure 4 shows the fingerprint of each of the combinations tested and Table 10 shows the results. The best automatic method, which combines (machine learning-all features to discover keywords and bootstrapping with the tweets annotated using that keywords) gives an accuracy of 0.73, which is higher than using manual keywords from 25 keyword seed set overall accuracy selection strategy coverage acc. wta wtr bootstrapping 20 oracle keywords 53%N 1.00N 0.80M 0.85N 0.87N manual keywords 15%H 0.86 0.61 0.63 0.67 supervised bootstr. 0.85N m. learning - all feat. 58% 0.75 0.69 0.71 0.73 heuristic 27%H 0.79 0.64 0.65 0.71 m. learning - 2 feat. 39%H 0.78 0.70 0.72 0.72 Table 10: Results for automatic keyword detection strategies (wta=winner-takes-all, wtr=winner-takes-remainder). Statistical significance w.r.t. the ml-all features se- lection strategy was computed using two-tailed Student’s t-test. Significant differences are indicated using N (or H) for α = 0.01 and M (or O) for α = 0.05. the Web (0.67) and is close to the best automatic result reported in the WePS-3 compe- tition (0.75). In addition, the bootstrapping process almost doubles the coverage (from 58% to 100%) with only 2.7% of accuracy loss. In general, the more tweets are covered by filter keywords (seed coverage), the lower is the loss in accuracy: the heuristic keyword selection covers 27% of tweets with .79 accuracy and achieves a .71 accuracy with the bootstrapping process, while machine learning-2 best features covers 39% of the tweets with 0.78 accuracy and finishes with 0.72 accuracy. Remarkably, the bootstrapping process outperforms the winner- takes-all and winner-takes-remainder baselines in all the cases. ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● avg: 0.85 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Figure 5: Fingerprint for the bootstrapping upper bound (10-fold cross-validation). An interesting question is how our approach - which does not use training data from the companies in the test set - compares with a truly supervised counterpart (i.e. one which uses the same machine learning algorithm and the same bag-of-word features, but uses perfect training material – taken from the test set – for each of the companies). To this aim, we carried a 10-fold validation on the test set of the machine learning algorithm. 26 This supervised upper bound achieves 0.85 accuracy, that is only 14% higher than our unsupervised algorithm (0.73). In Figure 5 we can see the fingerprint representation of the supervised upper bound. This fingerprint is similar to the 20 oracle keyword’s fingerprint shown in Section 3.2. Discovery of filter keywords has proved to be challenging using signals from the Web: the accuracy of the resulting seed set ranges between 0.75 and 0.79, with a potentially useful coverage (58%) in the case of the machine learning - all features. Overall, this result reinforces the conclusion that the characterization of companies in Twitter, in terms of vocabulary, is probably different from the characterization that can be derived from the Web. 27 6. Discussion 6.1. How Much of The Problem is Solved? In order to shed light on the trade-off between quality and quantity of filter keywords, here we analyze the relation between accuracy and coverage of the tweets classified by considering different sets of filter keywords. Figure 6 shows the coverage/accuracy curves for oracle, manual and automatic filter keywords. ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0 10 20 30 40 50 60 70 80 90 100 tweet coverage with keywords (%) tw e e t cl a ss ifi ca tio n a cc u ra cy Keyword Classification Strategy ● oracle keywords manual keywords machine learning − all features heuristic machine learning − 2 features Figure 6: Coverage/accuracy curves for oracle, manual and automatic filter keywords. Curves were generated as follows: Oracle keywords. At step n, we consider the nth positive/negative oracle keywords that maximizes accuracy and - in case of ties - coverage of tweets (i.e., in the case of two keywords having same accuracy, the one that covers more tweets is considered first). Manual keywords. At each step n we consider the nth positive/negative manual key- words that maximize coverage of tweets. Machine learning - all features. The set of terms considered in this analysis are those that were classified as positive or negative by using the confidence thresholds learned by the two classifiers in the method “machine learning - all features”. Skip terms are (i) those classified as skip by both binary classifiers; (ii) those classified si- multaneously as positive and negative keywords. Then, we use the maximum of the 28 confidence scores returned by the two classifiers (i.e., max(conf(positive), conf(negative))) to sort the keywords. The keyword with highest confidence score is added at each step. The point in the curve with highest coverage corresponds to the classifier used in the experiments explained in Section 4. Machine learning - 2 features. This curve is generated by using the two classifiers learned in “machine learning - 2 features”, similarly to the curve generated for “machine learning - all features”. Heuristic. Since this classifier consists of manually defining thresholds using the training set, it doesn’t provide any confidence score for the test cases. Hence, in the graphic it is represented as a single point (×). The curve for Oracle keywords provides a statistical upper bound of how many tweets can be directly covered using filter keywords. Considering the best 100 oracle keywords for each test case/company name, it is possible to directly tag 85% of the tweets with 0.99 accuracy. On the other hand, a more realistic upper bound is given by manual keywords. Here, we can observe how the accuracy remains stable around 0.85, while the coverage grows from 10% to 15% approx. In the best possible case, with more keywords the curve would continue as the line y = 0.85. Note that manual keywords have been annotated by inspecting representative Web pages (from Google search results) rather than inspecting tweets. Therefore, an automatic keyword classifier cannot achieve an accuracy above 0.85. Considering this, our automatic approaches establish a strong lower bound of 0.7 accu- racy. In conclusion, it seems that a filter keyword classifier should have reach an accuracy between 0.7 and 0.85 to be competitive. 6.2. Comparing Systems with Different Metrics We have seen that related/unrelated tweets are not balanced in most of the test cases in WePS-3, and the proportion does not follow a normal distribution (extreme values seem to be as plausible as values around the mean). Because of this, accuracy may be not sufficient to understand the quality of systems, and that’s why we have complemented it with the fingerprint representation [68]. In this section, we evaluate (and compare) results with the most popular alternative evaluation metrics found in the literature. Considering the confusion matrix given by each system, where TP=true related tweets, FP=false related tweets, TN=true unrelated tweets, and FN=false unrelated tweets, we compute the following metrics, in addition to accuracy: Normalized Utility. Utility has been used to evaluate document filtering tasks in TREC [34, 33] and is commonly used assigning a relative α weight between true positives and false positives: u(S,T) = α ·TP −FP As in the TREC-8 filtering task [33], here Utility is normalized by means of the following scaling function: u∗s(S,T) = max(u(S,T),U(s)) −U(s) MaxU(T) −U(s) 29 where u(S,T) is the original utility of system output S for topic T , U(s) is the utility of retrieving s non-relevant documents, and MaxU(T) = α · (TP + FN) is the maximum possible utility score for topic T. In this paper, we set α = 2 and U(s) = −25. lam%. lam% (logistic average misclassification percentage) has been used in TREC to evaluate the problem of spam detection [18]. It was defined as the geometric mean of the odds of hm% (ham misclassification percentage) and sm% (spam misclassification percentage). More precisely, lam% is defined as lam% = logit−1 (logit(hm%) + logit(sm%) 2 ) where hm% = FN FN + TP sm% = FP FP + TN logit(x) = log( x 1 −x ) logit−1(x) = ex 1 + ex Note that lam% is an error-based metric –i.e., maximum scores represent minimum quality. One remarkable property of this metric is that, when a system has a non-informative behavior –that is, it classifies the documents randomly, or classifies all documents to a single class– lam% score is around 0.5. Reliability & Sensitivity. Reliability and Sensitivity [5] have been recently proposed as two complementary measures to evaluate document organization tasks involving classification, clustering and ranking. It was the official metric used in RepLab held at CLEF 2012 [4]. When evaluating a binary classification task, Reliability corresponds to the product of the precision of the classes, and Sensitivity to the product of the recall of both classes: R = TP TP + FP · TN TN + FN S = TP TP + FN · TN TN + FP and F1(R,S) = 2 · R ·S R + S As lam%, Reliability & Sensitivity also penalizes systems that do not provide any useful information. Contrary to lam%, that assigns a low but not the minimum score, Reliability & Sensitivity gives zero –the minimum score– to these systems. F1 measure. The most standard combination of Precision and Recall is F1, or balanced F measure. Here we focus on the “related” class, where Precision = TP TP + FP Recall = TP TP + FN 30 and F1 = 2 · Precision · Recall Precision + Recall Table 11 reports the results for the baselines, the WePS-3 systems and our proposed systems for the metrics described above. All metrics were macro-averaged by topics, and undefined scores were considered as zero values. System accuracy utility lam% F1(R, S) F1 Gold standard 1.00N 1.00N 0.00N 1.00N 1.00N supervised bootstr. 0.85N 0.69N 0.28N 0.30N 0.62N WePS-3: LSIR (manual) 0.83N 0.66N 0.28N 0.27 0.62N ml-all feat. + bootstr. 0.73 0.47 0.37 0.27 0.49 ml-2 feat. + wtr 0.72 0.49 0.43 0.16O 0.49 ml-2 feat. + bootstr. 0.72 0.47 0.43 0.17O 0.50 heuristic + bootstr. 0.71 0.45 0.42 0.11H 0.46 ml-all feat. + wtr 0.71 0.44 0.39H 0.21H 0.43H ml-2 feat. + wta 0.70 0.48 0.50H 0.00H 0.39O ml-all feat. + wta 0.69O 0.40O 0.50H 0.00H 0.27H heuristic + wtr 0.65H 0.46 0.42 0.10H 0.46 heuristic + wta 0.64H 0.44 0.50H 0.00H 0.39O WePS-3: ITC-UT 0.75 0.52 0.37 0.20 0.49 WePS-3: SINAI 0.64O 0.38O 0.35 0.12H 0.30H WePS-3: UvA 0.58H 0.22H 0.46H 0.17H 0.36H WePS-3: KALMAR 0.47H 0.35H 0.43 0.16O 0.48 baseline: all unrelated 0.57H 0.20H 0.50H 0.00H 0.00H baseline: random 0.49H 0.20H 0.49H 0.16H 0.37H baseline: all related 0.43H 0.40 0.50H 0.00H 0.52 Table 11: Results for proposed systems, WePS-3 systems and baselines compared with different evaluation metrics. Best automatic runs are in boldface. (ml=machine learning, wta=winner-takes-all, wtr=winner-takes-remainder). Statistical significance w.r.t. the ml-all feat. + bootstrapping run was computed using two-tailed Student’s t-test. Significant differences are indicated using N (or H) for α = 0.01 and M (or O) for α = 0.05. Results show that, according to Reliability & Sensitivity, our best automatic system ml-all features + bootstrapping achieves the same score as the WePS-3 LSIR semi- automatic system (0.27) – which is the best result at the competition and involves manual processing – and outperforms the best automatic system in WePS-3 (ITC-UT=0.20), with a 35% of relative improvement. In terms of lam%, the SINAI system achieves the best automatic score of 0.35, followed by ITC-UT & ml-all features + bootstrapping that reaches 0.37 lam%. Note that lam% and R&S penalize non-informative/baseline- like behaviors. Because of this, the winner-takes-all systems and the “all (un)related” baselines get the worst scores in these metrics. According to utility, ITC-UT is still the best automatic system (0.52). Our best runs are between 0.47 and 0.49, being ml-2 features + bootstrapping the best of them. 31 Finally, F1 rewards systems that tend to return all tweets as related. Indeed, the best score given by an automatic system is achieved by the “all related” baseline, that has perfect recall and enough precision to get the highest score. There are a number of empirical observations that can be made on this comparison of metrics: • In general, high F1(R,S) implies high accuracy, but not vice-versa: F1(R,S) is a stricter metric, at least in this dataset. • Metrics such as lam% and F1(R,S) are suitable to identify baseline-like behaviors, while F1 is not. • ITC-UT and ml-features + bootstrapping perform consistently well across met- rics. • Different metrics illustrate different aspects of the behavior of systems: If we need to penalize non-informative behavior, we should look at results with lam% or F1(R,S). Accuracy and utility directly show misclassification errors, but are sen- sitive to collections where class skews are variable over different test cases, such as our dataset. 6.3. Web vs. Twitter In our experiments, we have found two results indicating that the vocabulary that characterizes a company on Twitter substantially differs from the the vocabulary asso- ciated to the company on the Web: • Low recall of manual keywords. As we saw in Section 3.1, manually selecting around 10 salient terms from Web search results retrieved using the company name and its representative URL only covers 15% of the tweets. • Web-based features are useful but inconclusive. Analyzing the features (see Section 4.2), we found that web-based features that may discriminate positive from negative keywords tend to receive low values. Moreover, the low quality ( ≤ 0.75 AUC) of the automatic classification of filter keywords indicates that building an accurate classifier with features extracted from the Web is challenging to say the least (see Section 4.3). In order to reach a better understanding of the problem, we have explored the asso- ciation between the best 10 oracle keywords for each tweet stream and its occurrences in both the company’s homepage and its Wikipedia article17. The terms from each page have been extracted using the lynx -dump url Linux command. Table 12 shows the average percentage of the best 10 oracle keywords that occur on the company’s homepage, on the Wikipedia page, and both. 17We manually extended the input data of each organization on the WePS-3 dataset with its Wikipedia page (or its homepage in the cases which the Wikipedia page is provided as the representative page) 32 Filter keywords Homepage Wikipedia Both related oracle keywords 36% 68% 33% unrelated oracle keywords 9% 19% 6% Table 12: Percentage of the 10 best oracle keywords extracted from the tweet stream covered by the company’s homepage, its Wikipedia article and both. Overall, the only substantial overlap is for positive keywords in Wikipedia, indicating that representative Web pages are not the ideal place to look for effective filter keywords in Twitter. Note that the overlap of related oracle keywords with the company’s Wikipedia page roughly doubles the overlap with its homepage. The same thing happens with unrelated keywords: almost 20% on Wikipedia and almost 10% on the homepage. The percentage of oracle keywords that occur both in the homepage and in the Wikipedia article is similar to the homepage alone, indicating that Wikipedia basically extends the keywords already present in the homepage. In summary, exploring the nature of filter keywords leads us to the conclusion that the vocabulary characterizing a company in Twitter is substantially different from the vocabulary associated to the company in its homepage, in Wikipedia, and apparently in the Web at large. These findings are in line with the “vocabulary gap“ that has been shown between Twitter and other Web sources such as Wikipedia or news comments [73]. One way of alleviating this problem is using co-occurrence expansion of web-based fea- tures, which allows to better recognize automatically filter keywords. While the com- pany’s Wikipedia article seems to have more coverage of (perfect) filter keywords than the company’s homepage, further investigation is needed on how to automatically infer the company’s Wikipedia page from its homepage URL in order to extract additional keyword features from it. 33 7. Conclusion In this paper we tackled the problem of company name disambiguation in Twitter, defined in WePS-3 as a binary classification problem. We have studied the use of filter keywords: expressions that, if present in a tweet, indicate a high probability that it is related/unrelated to the company. In our experiments, automatically discovered filter keywords are able to classify a subset of 30%-60% tweets with an accuracy range of .75 − .79. We defined features that characterize terms in the Twitter dataset, the company’s website, ODP, Wikipedia and the searchable Web. We found that (i) term specificity in the tweet stream of each company name is a feature that discriminates between filter keywords and skip terms and (ii) the association between the term and the company website is useful to differentiate positive vs. negative filter keywords, specially when it is averaged by considering its most co-occurrent terms. Tweets classified by these filter keywords can be used to feed a supervised machine learning process to obtain a complete classification of all tweets for an overall accuracy of 0.73. In comparison, a 10- fold validation of the same machine learning algorithm provides an accuracy of 0.85, i.e., our unsupervised algorithm has a 14% loss with respect to its supervised counterpart. We also found that, in average, the best five optimal keywords can directly classify around 30% of the tweets. Nevertheless, keywords defined by a human by inspecting web search results relevant to the company name only cover 15% of the tweets and accuracy drops to 0.86. Exploring the nature of filter keywords also led us to the conclusion that that the there is a gap between the vocabulary characterizing a company in Twitter and the vocabulary associated to the company in its homepage, in Wikipedia, and apparently in the Web at large. Note that all our experimentation is based on the WePS-3 dataset, which is an English-only dataset. As immediate future work, we plan to explore cross-lingual strate- gies to deal with multilingual data and so be able to test our approach on the RepLab 2012 dataset, which is the other test collection that fits our problem. 34 References [1] Agirre, E., Edmonds, P., 2006. Word sense disambiguation: Algorithms and applications. Springer. [2] Al-Kamha, R., Embley, D., 2004. Grouping search-engine returned citations for person-name queries. In: Proceedings of the 6th annual ACM international workshop on Web information and data management. pp. 96–103. [3] Amigó, E., Artiles, J., Gonzalo, J., Spina, D., Liu, B., Corujo, A., 2010. WePS-3 Evaluation Cam- paign: Overview of the Online Reputation Management Task. In: CLEF 2010 Labs and Workshops Notebook Papers. [4] Amigó, E., Corujo, A., Gonzalo, J., Meij, E., de Rijke, M., 2012. Overview of RepLab 2012: Evaluating Online Reputation Management Systems. In: CLEF 2012 Labs and Workshop Notebook Papers. [5] Amigó, E., Gonzalo, J., Verdejo, F., 2012. Reliability and Sensitivity: Generic Evaluation Measures for Document Organization Tasks. Tech. rep., UNED. [6] Artiles, J., October 2009. Web people search. Ph.D. thesis, UNED University. [7] Artiles, J., Borthwick, A., Gonzalo, J., Sekine, S., Amigó, E., 2010. Weps-3 evaluation campaign: Overview of the web people search clustering and attribute extraction tasks. [8] Artiles, J., Gonzalo, J., Amigó, E., 2009. The impact of query refinement in the web people search task. In: Proceedings of the ACL-IJCNLP 2009 Conference Short Papers. Association for Compu- tational Linguistics, pp. 361–364. [9] Artiles, J., Gonzalo, J., Sekine, S., 2007. The semeval-2007 weps evaluation: Establishing a bench- mark for the web people search task. Proceedings of Semeval. [10] Artiles, J., Gonzalo, J., Sekine, S., 2009. Weps 2 evaluation campaign: overview of the web people search clustering task. In: 2nd Web People Search Evaluation Workshop (WePS 2009), 18th WWW Conference. [11] Bagga, A., Baldwin, B., 1998. Entity-based cross-document coreferencing using the vector space model. In: Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics-Volume 1. pp. 79–85. [12] Boyd, D., Golder, S., Lotan, G., 2010. Tweet, tweet, retweet: Conversational aspects of retweeting on twitter. In: Proceedings of the 2010 43rd Hawaii International Conference on System Sciences. HICSS ’10. IEEE Computer Society, Washington, DC, USA, pp. 1–10. [13] Bunescu, R., Pasca, M., 2006. Using encyclopedic knowledge for named entity disambiguation. In: Proceedings of EACL. Vol. 6. pp. 9–16. [14] Cha, M., Haddadi, H., Benevenuto, F., Gummadi, K. P., 2010. Measuring User Influence in Twitter: The Million Follower Fallacy. In: In Proceedings of the 4th International AAAI Conference on Weblogs and Social Media (ICWSM). Washington DC, USA. [15] Cheng, Z., Caverlee, J., Lee, K., 2010. You are where you tweet: a content-based approach to geo- locating twitter users. In: Proceedings of the 19th ACM international conference on Information and knowledge management. pp. 759–768. [16] Cilibrasi, R. L., Vitanyi, P. M., 2007. The google similarity distance. IEEE Transactions on Knowl- edge and Data Engineering. [17] Comm, J., Robbins, A., 2009. Twitter power: How to dominate your market one tweet at a time. John Wiley & Sons Inc. [18] Cormack, G., Lynam, T., 2005. Trec 2005 spam track overview. In: The Fourteenth Text REtrieval Conference (TREC 2005) Proceedings. [19] Cucerzan, S., 2007. Large-scale named entity disambiguation based on wikipedia data. In: Proceed- ings of EMNLP-CoNLL. Vol. 2007. pp. 708–716. [20] Dellarocas, C., Awad, N., Zhang, X., 2004. Exploring the value of online reviews to organizations: Implications for revenue forecasting and planning. In: Proceedings of the International Conference on Information Systems. [21] Dredze, M., McNamee, P., Rao, D., Gerber, A., Finin, T., 2010. Entity disambiguation for knowl- edge base population. In: Proceedings of the 23rd International Conference on Computational Linguistics. pp. 277–285. [22] Edmonds, P., Cotton, S., 2001. Senseval-2: Overview. In: Proceedings of The Second International Workshop on Evaluating Word Sense Disambiguation Systems (SENSEVAL-2. pp. 1–6. [23] Fawcett, T., 2006. An introduction to roc analysis. Pattern recognition letters 27 (8), 861–874. [24] Ferragina, P., Scaiella, U., 2010. Tagme: on-the-fly annotation of short text fragments (by wikipedia entities). In: Proceedings of the 19th ACM international conference on Information and knowledge management. pp. 1625–1628. 35 [25] Frank, E., Paynter, G., Witten, I., Gutwin, C., Nevill-Manning, C., 1999. Domain-specific keyphrase extraction. In: International joint conference on artificial intelligence. Vol. 16. pp. 668–673. [26] Garćıa-Cumbreras, M. A., Garćıa-Vega, M., Mart́ınez-Santiago, F., Peréa-Ortega, J. M., 2010. SINAI at WePS-3: Online Reputation Management. In: CLEF 2010 Labs and Workshops Notebook Papers. [27] Gentile, A., Zhang, Z., Xia, L., Iria, J., 2010. Semantic relatedness approach for named entity disambiguation. In: Digital Libraries. pp. 137–148. [28] Gooi, C., Allan, J., 2004. Cross-document coreference on a large scale corpus. In: Proceedings of HLT/NAACL. Vol. 4. [29] Gouws, S., Metzler, D., Cai, C., Hovy, E., Marina del Rey, C., 2011. Contextual bearing on linguistic variation in social media. [30] Grishman, R., Sundheim, B., 1996. Message understanding conference-6: A brief history. In: Pro- ceedings of the 16th conference on Computational linguistics-Volume 1. pp. 466–471. [31] Guyon, I., Gunn, S., Nikravesh, M., Zadeh, L., 2006. Feature extraction: foundations and applica- tions. Springer Verlag, Ch. 2.2, pp. 67–78. [32] Hoffman, T., 2008. Online reputation management is hot?but is it ethical? Computerworld, Febru- ary. [33] Hull, D., Robertson, S., 1999. The trec-8 filtering track final report. In: Proceeding of eighth Text Retrieval Conference (TREC-8). [34] Hull, D., et al., 1998. The trec-7 filtering track: description and analysis. NIST SPECIAL PUBLI- CATION SP, 45–68. [35] Hurlock, J., Wilson, M. L., 2011. Searching twitter: Separating the tweet from the chaff. In: Pro- ceedings of ICWSM 2011, the 5th International AAAI Conference on Weblogs and Social Media. AAAI. [36] Jansen, B., Zhang, M., Sobel, K., Chowdury, A., 2009. Twitter power: Tweets as electronic word of mouth. Journal of the American society for information science and technology 60 (11), 2169–2188. [37] Java, A., Song, X., Finin, T., Tseng, B., 2007. Why we twitter: understanding microblogging usage and communities. In: Proceedings of the 9th WebKDD and 1st SNA-KDD 2007 workshop on Web mining and social network analysis. pp. 56–65. [38] Ji, H., Grishman, R., 2011. Knowledge base population: Successful approaches and challenges. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies. Vol. 1. pp. 1148–1158. [39] Ji, H., Grishman, R., Dang, H., Griffitt, K., Ellis, J., 2010. Overview of the tac 2010 knowledge base population track. In: TAC (Text Analysis Conference) 2010 Workshop. [40] Kalashnikov, D., Mehrotra, S., Chen, Z., Nuray-Turan, R., Ashish, N., 2007. Disambiguation algo- rithm for people search on the web. In: Data Engineering, 2007. ICDE 2007. IEEE 23rd International Conference on. pp. 1258–1260. [41] Kalmar, P., 2010. Bootstrapping Websites for Classification of Organization Names on Twitter. In: CLEF 2010 Labs and Workshops Notebook Papers. [42] Kaufmann, M., Kalita, J., 2010. Syntactic normalization of twitter messages. In: International Conference on Natural Language Processing, Kharagpur, India. [43] Kim, S., Medelyan, O., Kan, M., Baldwin, T., 2010. Semeval-2010 task 5: Automatic keyphrase extraction from scientific articles. In: Proceedings of the 5th International Workshop on Semantic Evaluation. pp. 21–26. [44] Krishnamurthy, B., Gill, P., Arlitt, M., 2008. A few chirps about twitter. In: WOSP ’08: Proceedings of the first workshop on Online social networks. ACM, New York, NY, USA, pp. 19–24. [45] Kulkarni, S., Singh, A., Ramakrishnan, G., Chakrabarti, S., 2009. Collective annotation of wikipedia entities in web text. In: Proceedings of the 15th ACM SIGKDD international conference on Knowl- edge discovery and data mining. pp. 457–466. [46] Kwak, H., Lee, C., Park, H., Moon, S., 2010. What is twitter, a social network or a news media? In: WWW ’10: Proceedings of the 19th international conference on World wide web. ACM, New York, NY, USA, pp. 591–600. [47] Laboreiro, G., Sarmento, L., Teixeira, J., Oliveira, E., 2010. Tokenizing micro-blogging messages using a text classification approach. In: Proceedings of the fourth workshop on Analytics for noisy unstructured text data. pp. 81–88. [48] Liu, Z., Huang, W., Zheng, Y., Sun, M., 2010. Automatic keyphrase extraction via topic decomposi- tion. In: Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing. pp. 366–376. [49] Mann, G., 2006. Multi-document statistical fact extraction and fusion. Ph.D. thesis, THE JOHNS 36 HOPKINS UNIVERSITY. [50] Mann, G., Yarowsky, D., 2003. Unsupervised personal name disambiguation. In: Proceedings of the HLT-NAACL’03. pp. 33–40. [51] Mann, H., Whitney, D., 1947. On a test of whether one of two random variables is stochastically larger than the other. The Annals of Mathematical Statistics 18 (1), 50–60. [52] Martinez-Romo, J., Araujo, L., 2009. Web people search disambiguation using language model techniques. [53] McNamee, P., Dang, H., 2009. Overview of the TAC 2009 knowledge base population track. In: Text Analysis Conference (TAC). [54] Meij, E., Weerkamp, W., de Rijke, M., 2012. Adding semantics to microblog posts. In: Proceedings of the fifth ACM international conference on Web search and data mining. [55] Michelson, M., Macskassy, S., 2010. Discovering users’ topics of interest on twitter: a first look. In: Proceedings of the fourth workshop on Analytics for noisy unstructured text data. pp. 73–80. [56] Mierswa, I., Wurst, M., Klinkenberg, R., Scholz, M., Euler, T., 2006. YALE: Rapid prototyping for complex data mining tasks. In: Proceedings of the 12th ACM SIGKDD. [57] Mihalcea, R., Csomai, A., 2007. Wikify!: linking documents to encyclopedic knowledge. In: CIKM. Vol. 7. pp. 233–242. [58] Mihalcea, R., Tarau, P., 2004. Textrank: Bringing order into texts. In: Proceedings of EMNLP. Vol. 4. pp. 404–411. [59] Milios, E., Zhang, Y., He, B., Dong, L., 2003. Automatic term extraction and document similarity in special text corpora. In: Proceedings of the Sixth Conference of the Pacific Association for Computational Linguistics. pp. 275–284. [60] Milne, D., Witten, I., 2008. Learning to link with wikipedia. In: Proceeding of the 17th ACM conference on Information and knowledge management. pp. 509–518. [61] Milstein, S., Chowdhury, A., Hochmuth, G., Lorica, B., Magoulas, R., 2008. Twitter and the micro-messaging revolution: Communication, connections, and immediacy–140 characters at a time. O’Reilly Media, Inc. [62] Nadeau, D., Sekine, S., 2007. A survey of named entity recognition and classification. Lingvisticae Investigationes 30 (1), 3–26. [63] Pak, A., Paroubek, P., 2010. Twitter as a corpus for sentiment analysis and opinion mining. Pro- ceedings of LREC 2010. [64] Pollach, I., 2006. Electronic word of mouth: A genre analysis of product reviews on consumer opinion web sites. In: Proceedings of the 39th Annual Hawaii International Conference on System Sciences. Vol. 3. [65] Quinlan, J., 1993. C4. 5: programs for machine learning. Morgan Kaufmann. [66] Sakaki, T., Okazaki, M., Matsuo, Y., 2010. Earthquake shakes twitter users: real-time event de- tection by social sensors. In: Proceedings of the 19th international conference on World wide web. WWW ’10. pp. 851–860. [67] Soboroff, I., McCullough, D., Lin, J., Macdonald, C., Ounis, I., McCreadie, R., 2012. Evaluating real-time search over tweets. [68] Spina, D., Amigó, E., Gonzalo, J., 2011. Filter keywords and majority class strategies for com- pany name disambiguation in Twitter. In: CLEF 2011 Conference on Multilingual and Multimodal Information Access Evaluation. Springer Berlin/Heidelberg, pp. 50–61. [69] Spina, D., Meij, E., Oghina, A., Bui, M. T., Breuss, M., de Rijke, M., 2012. A Corpus for Entity Profiling in Microblog Posts. In: LREC Workshop on Language Engineering for Online Reputation Management. [70] Sriram, B., Fuhry, D., Demir, E., Ferhatosmanoglu, H., Demirbas, M., 2010. Short text classification in twitter to improve information filtering. In: Proceeding of the 33rd International ACM SIGIR Conference on Research and Development in Information Retrieval. SIGIR ’10. ACM, New York, NY, USA, pp. 841–842. [71] Strube, M., Ponzetto, S., 2006. Wikirelate! computing semantic relatedness using wikipedia. In: Proceedings of the National Conference on Artificial Intelligence. Vol. 21. p. 1419. [72] Tsagkias, M., Balog, K., 2010. The University of Amsterdam at WePS3. In: CLEF 2010 Labs and Workshops Notebook Papers. [73] Tsagkias, M., de Rijke, M., Weerkamp, W., 2011. Linking online news and social media. In: Pro- ceedings of the fourth ACM International Conference on Web Search and Data Mining. [74] Turney, P., 1999. Learning to extract keyphrases from text, national research council. Institute for Information Technology, Technical Report ERB-1057. [75] Wan, X., Gao, J., Li, M., Ding, B., 2005. Person resolution in person search results: Webhawk. In: 37 Proceedings of the 14th ACM international conference on Information and knowledge management. pp. 163–170. [76] Wilson, R., 2003. Keeping a watch on corporate reputation. Strategic Communications Management 7 (2). [77] Witten, I. H., Paynter, G. W., Frank, E., Gutwin, C., Nevill-Manning, C. G., 1999. Kea: practical automatic keyphrase extraction. In: Proceedings of the fourth ACM conference on Digital libraries. DL ’99. pp. 254–255. [78] Wu, S., Hofman, J. M., Mason, W. A., Watts, D. J., 2011. Who says what to whom on twitter. In: Proceedings of the 20th International Conference on World Wide Web. WWW ’11. ACM, New York, NY, USA, pp. 705–714. [79] Yang, J., Leskovec, J., 2010. Modeling information diffusion in implicit networks. Data Mining, IEEE International Conference on 0, 599–608. [80] Yang, J., Leskovec, J., 2011. Patterns of temporal variation in online media. In: Proceedings of the fourth ACM international conference on Web search and data mining. WSDM ’11. pp. 177–186. [81] Yerva, S. R., Miklós, Z., Aberer, K., 2010. It was easy when apples and blackberries were only fruits. In: CLEF 2010 Labs and Workshops Notebook Papers. [82] Yerva, S. R., Mikls, Z., Aberer, K., 2012. Entity-based classification of twitter messages. IJCSA 9 (1), 88–115. [83] Yoshida, M., Matsushima, S., Ono, S., Sato, I., Nakagawa, H., 2010. ITC-UT: Tweet Categorization by Query Categorization for On-line Reputation Management. In: CLEF 2010 Labs and Workshops Notebook Papers. [84] Zhang, S., Wu, J., Zheng, D., Meng, Y., Xia, Y., Yu, H., 2012. Two stages based organization name disambiguity. Computational Linguistics and Intelligent Text Processing, 249–257. [85] Zhang, Y., Milios, E., Zincir-Heywood, N., 2007. A comparative study on key phrase extraction methods in automatic web site summarization. Journal of Digital Information Management 5 (5), 323. [86] Zhang, Y., Zincir-Heywood, N., Milios, E., 2004. Term-based clustering and summarization of web page collections. Advances in Artificial Intelligence, 60–74. [87] Zhang, Y., Zincir-Heywood, N., Milios, E., 2004. World wide web site summarization. WEB IN- TELLIGENCE AND AGENT SYSTEMS. 2, 39–54. [88] Zhang, Y., Zincir-Heywood, N., Milios, E., 2005. Narrative text classification for automatic key phrase extraction in web document corpora. In: Proceedings of the 7th annual ACM international workshop on Web information and data management. pp. 51–58. [89] Zhao, W. X., Jiang, J., Weng, J., He, J., Lim, E.-P., Yan, H., Li, X., 2011. Comparing twitter and traditional media using topic models. In: Proceedings of the 33rd European conference on Advances in Information Retrieval. [90] Zhao, X., Jiang, J., He, J., Song, Y., Achananuparp, P., LIM, E., Li, X., 2011. Topical keyphrase extraction from twitter. 38 1 Introduction 2 Related Work 2.1 Twitter 2.1.1 Twitter-related Tasks 2.1.2 Twitter Datasets 2.2 Named Entity Disambiguation 2.2.1 Entity Linking 2.2.2 Document Enrichment by Linking to Wikipedia Articles 2.2.3 Web People Search 2.2.4 Named Entity Disambiguation in Twitter 2.3 Automatic Keyphrase Extraction 2.4 The WePS-3 Online Reputation Management Task 2.4.1 Task Definition and Test Collection 2.4.2 WePS-3 Results 2.4.3 Other work using WePS-3 datasets 2.5 Wrap Up 3 Potential Benefits of Filter Keywords for Name Disambiguation in Twitter 3.1 Upper Bound Performance of Filter Keywords 3.2 Using Keywords to Solve The Task 4 Automatic Discovery of Filter Keywords 4.1 Term Features 4.2 Feature Analysis 4.3 Keyword Discovery 5 Automatic Tweet Classification 6 Discussion 6.1 How Much of The Problem is Solved? 6.2 Comparing Systems with Different Metrics 6.3 Web vs. Twitter 7 Conclusion 
work_om3lmqvjone6fcsyruwnlpw3hi	----	A survey on sentiment detection of reviews Expert Systems with Applications 36 (2009) 10760–10773 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A survey on sentiment detection of reviews Huifeng Tang, Songbo Tan *, Xueqi Cheng Information Security Center, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100080, PR China a r t i c l e i n f o Keywords: Sentiment detection Opinion extraction Sentiment classification 0957-4174/$ - see front matter � 2009 Published by doi:10.1016/j.eswa.2009.02.063 * Corresponding author. E-mail addresses: tanghuifeng@software.ict.ac.cn (H e.ict.ac.cn (S. Tan), cxq@ict.ac.cn (X. Cheng). a b s t r a c t The sentiment detection of texts has been witnessed a booming interest in recent years, due to the increased availability of online reviews in digital form and the ensuing need to organize them. Till to now, there are mainly four different problems predominating in this research community, namely, sub- jectivity classification, word sentiment classification, document sentiment classification and opinion extraction. In fact, there are inherent relations between them. Subjectivity classification can prevent the sentiment classifier from considering irrelevant or even potentially misleading text. Document sen- timent classification and opinion extraction have often involved word sentiment classification tech- niques. This survey discusses related issues and main approaches to these problems. � 2009 Published by Elsevier Ltd. 1. Introduction Today, very large amount of reviews are available on the web, as well as the weblogs are fast-growing in blogsphere. Product re- views exist in a variety of forms on the web: sites dedicated to a specific type of product (such as digital camera), sites for newspa- pers and magazines that may feature reviews (like Rolling Stone or Consumer Reports), sites that couple reviews with commerce (like Amazon), and sites that specialize in collecting professional or user reviews in a variety of areas (like Rottentomates.com). Less formal reviews are available on discussion boards and mailing list archives, as well as in Usenet via Google Groups. Users also com- ment on products in their personal web sites and blogs, which are then aggregated by sites such as Blogstreet.com, AllConsum- ing.net, and onfocus.com. The information mentioned above is a rich and useful source for marketing intelligence, social psychologists, and others interested in extracting and mining opinions, views, moods, and attitudes. For example, whether a product review is positive or negative; what are the moods among Bloggers at that time; how the public reflect towards this political affair, etc. To achieve this goal, a core and essential job is to detect subjec- tive information contained in texts, include viewpoint, fancy, atti- tude, sensibility etc. This is so-called sentiment detection. A challenging aspect of this task seems to distinguish it from traditional topic-based detection (classification) is that while top- ics are often identifiable by keywords alone, sentiment can be ex- pressed in a much subtle manner. For example, the sentence ‘‘What a bad picture quality that digital camera has! . . . Oh, this Elsevier Ltd. . Tang), tansongbo@softwar- new type camera has a good picture, long battery life and beautiful appearance!” compares a negative experience of one product with a positive experience of another product. It is difficult to separate out the core assessment that should actually be correlated with the document. Thus, sentiment seems to require more understand- ing than the usual topic-based classification. Sentiment detection dates back to the late 1990s (Argamon, Koppel, & Avneri, 1998; Kessler, Nunberg, & SchÄutze, 1997; Sper- tus, 1997), but only in the early 2000s did it become a major sub- field of the information management discipline (Chaovalit & Zhou, 2005; Dimitrova, Finn, Kushmerick, & Smyth, 2002; Durbin, Neal Richter, & Warner, 2003; Efron, 2004; Gamon, 2004; Glance, Hurst, & Tomokiyo, 2004; Grefenstette, Qu, Shanahan, & Evans, 2004; Hil- lard, Ostendorf, & Shriberg, 2003; Inkpen, Feiguina, & Hirst, 2004; Kobayashi, Inui, & Inui, 2001; Liu, Lieberman, & Selker, 2003; Rau- bern & Muller-Kogler, 2001; Riloff and Wiebe, 2003; Subasic & Huettner, 2001; Tong, 2001; Vegnaduzzo, 2004; Wiebe & Riloff, 2005; Wilson, Wiebe, & Hoffmann, 2005). Until the early 2000s, the two main popular approaches to sentiment detection, espe- cially in the real-world applications, were based on machine learn- ing techniques and based on semantic analysis techniques. After that, the shallow nature language processing techniques were widely used in this area, especially in the document sentiment detection. Current-day sentiment detection is thus a discipline at the crossroads of NLP and IR, and as such it shares a number of characteristics with other tasks such as information extraction and text-mining. Although several international conferences have devoted spe- cial issues to this topic, such as ACL, AAAI, WWW, EMNLP, CIKM etc., there are no systematic treatments of the subject: there are neither textbooks nor journals entirely devoted to sentiment detection yet. mailto:tanghuifeng@software.ict.ac.cn mailto:tansongbo@software.ict.ac.cn mailto:tansongbo@software.ict.ac.cn mailto:cxq@ict.ac.cn http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 10761 This paper first introduces the definitions of several problems that pertain to sentiment detection. Then we present some appli- cations of sentiment detection. Section 4 discusses the subjectivity classification problem. Section 5 introduces semantic orientation method. The sixth section examines the effectiveness of applying machine learning techniques to document sentiment classification. The seventh section discusses opinion extraction problem. The eighth part talks about evaluation of sentiment detection. Last sec- tion concludes with challenges and discussion of future work. 2. Sentiment detection 2.1. Subjectivity classification Subjectivity in natural language refers to aspects of language used to express opinions and evaluations (Wiebe, 1994). Subjectiv- ity classification is stated as follows: Let S = {s1, . . . , sn} be a set of sentences in document D. The problem of subjectivity classification is to distinguish sentences used to present opinions and other forms of subjectivity (subjective sentences set Ss) from sentences used to objectively present factual information (objective sen- tences set So), where Ss [ So = S. This task is especially relevant for news reporting and Internet forums, in which opinions of various agents are expressed. 2.2. Sentiment classification Sentiment classification includes two kinds of classification forms, i.e., binary sentiment classification and multi-class senti- ment classification. Given a document set D = {d1, . . . , dn}, and a pre-defined categories set C = {positive, negative}, binary senti- ment classification is to classify each di in D, with a label expressed in C. If we set C* = {strong positive, positive, neutral, negative, strong negative} and classify each di in D with a label in C*, the problem changes to multi-class sentiment classification. Most prior work on learning to identify sentiment has focused on the binary distinction of positive vs. negative. But it is often helpful to have more information than this binary distinction pro- vides, especially if one is ranking items by recommendation or comparing several reviewers’ opinions. Koppel and Schler (2005a, 2005b) show that it is crucial to use neutral examples in learning polarity for a variety of reasons. Learning from negative and posi- tive examples alone will not permit accurate classification of neu- tral examples. Moreover, the use of neutral training examples in learning facilitates better distinction between positive and nega- tive examples. 3. Applications of sentiment detection In this section, we will expound some rising applications of sen- timent detection. 3.1. Products comparison It is a common practice for online merchants to ask their cus- tomers to review the products that they have purchased. With more and more people using the Web to express opinions, the number of reviews that a product receives grows rapidly. Most of the researches about these reviews were focused on automatically classifying the products into ‘‘recommended” or ‘‘not recom- mended” (Pang, Lee, & Vaithyanathan, 2002; Ranjan Das & Chen, 2001; Terveen, Hill, Amento, McDonald, & Creter, 1997). But every product has several features, in which maybe only part of them people are interested. Moreover, a product has shortcomings in one aspect, probably has merits in another place (Morinaga, Yamanishi, Tateishi, & Fukushima, 2002; Taboada, Gillies, & McFe- tridge, 2006). To analysis the online reviews and bring forward a visual man- ner to compare consumers’ opinions of different products, i.e., merely with a single glance the user can clearly see the advantages and weaknesses of each product in the minds of consumers. For a potential customer, he/she can see a visual side-by-side and fea- ture-by-feature comparison of consumer opinions on these prod- ucts, which helps him/her to decide which product to buy. For a product manufacturer, the comparison enables it to easily gather marketing intelligence and product benchmarking information. Liu, Hu, and Cheng (2005) proposed a novel framework for ana- lyzing and comparing consumer opinions of competing products. A prototype system called Opinion Observer is implemented. To en- able the visualization, two tasks were performed: (1) Identifying product features that customers have expressed their opinions on, based on language pattern mining techniques. Such features form the basis for the comparison. (2) For each feature, identifying whether the opinion from each reviewer is positive or negative, if any. Different users can visualize and compare opinions of different products using a user interface. The user simply chooses the prod- ucts that he/she wishes to compare and the system then retrieves the analyzed results of these products and displays them in the interface. 3.2. Opinion summarization The number of online reviews that a product receives grows rapidly, especially for some popular products. Furthermore, many reviews are long and have only a few sentences containing opin- ions on the product. This makes it hard for a potential customer to read them to make an informed decision on whether to purchase the product. The large number of reviews also makes it hard for product manufacturers to keep track of customer opinions of their products because many merchant sites may sell their products, and the manufacturer may produce many kinds of products. Opinion summarization (Ku, Lee, Wu, & Chen, 2005; Philip et al., 2004) summarizes opinions of articles by telling sentiment polari- ties, degree and the correlated events. With opinion summariza- tion, a customer can easily see how the existing customers feel about a product, and the product manufacturer can get the reason why different stands people like it or what they complain about. Hu and Liu (2004a, 2004b) conduct a work like that: Given a set of customer reviews of a particular product, the task involves three subtasks: (1) identifying features of the product that customers have expressed their opinions on (called product features); (2) for each feature, identifying review sentences that give positive or negative opinions; and (3) producing a summary using the dis- covered information. Ku, Liang, and Chen (2006) investigated both news and web blog articles. In their research, TREC, NTCIR and articles collected from web blogs serve as the information sources for opinion extraction. Documents related to the issue of animal cloning are selected as the experimental materials. Algorithms for opinion extraction at word, sentence and document level are proposed. The issue of relevant sentence selection is discussed, and then top- ical and opinionated information are summarized. Opinion sum- marizations are visualized by representative sentences. Finally, an opinionated curve showing supportive and non-supportive de- gree along the timeline is illustrated by an opinion tracking system. 3.3. Opinion reason mining In opinion analysis area, finding the polarity of opinions or aggregating and quantifying degree assessment of opinions 10762 H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 scattered throughout web pages is not enough. We can do more critical part of in-depth opinion assessment, such as finding rea- sons in opinion-bearing texts. For example, in film reviews, infor- mation such as ‘‘found 200 positive reviews and 150 negative reviews” may not fully satisfy the information needs of different people. More useful information would be ‘‘This film is great for its novel originality” or ‘‘Poor acting, which makes the film awful”. Opinion reason mining tries to identify one of the critical ele- ments of online reviews to answer the question, ‘‘What are the rea- sons that the author of this review likes or dislikes the product?” To answer this question, we should extract not only sentences that contain opinion-bearing expressions, but also sentences with rea- sons why an author of a review writes the review (Cardie, Wiebe, Wilson, & Litman, 2003; Clarke & Terra, 2003; Li & Yamanishi, 2001; Stoyanov, Cardie, Litman, & Wiebe, 2004). Kim and Hovy (2005) proposed a method for detecting opinion- bearing expressions. In their subsequent work (Kim & Hovy, 2006), they collected a large set of hreview text, pros, consi triplets from epinions.com, which explicitly state pros and cons phrases in their respective categories by each review’s author along with the re- view text. Their automatic labeling system first collects phrases in pro and con fields and then searches the main review text in or- der to collect sentences corresponding to those phrases. Then the system annotates this sentence with the appropriate ‘‘pro” or ‘‘con” label. All remaining sentences with neither label are marked as ‘‘neither”. After labeling all the data, they use it to train their pro and con sentence recognition system. 3.4. Other applications Thomas, Pang, and Lee (2006) try to determine from the tran- scripts of US Congressional floor debates whether the speeches rep- resent support of or opposition to proposed legislation. Mullen and Malouf (2006) describe a statistical sentiment analysis method on political discussion group postings to judge whether there is oppos- ing political viewpoint to the original post. Moreover, there are some potential applications of sentiment detection, such as online message sentiment filtering, E-mail sentiment classification, web- blog author’s attitude analysis, sentiment web search engine, etc. 4. Subjectivity classification Subjectivity classification is a task to investigate whether a par- agraph presents the opinion of its author or reports facts. In fact, most of the research showed there was very tight relation between subjectivity classification and document sentiment classification (Pang & Lee, 2004; Wiebe, 2000; Wiebe, Bruce, & O’Hara, 1999; Wiebe, Wilson, Bruce, Bell, & Martin, 2002; Yu & Hatzivassiloglou, 2003). Subjectivity classification can prevent the polarity classifier from considering irrelevant or even potentially misleading text. Pang and Lee (2004) find subjectivity detection can compress re- views into much shorter extracts that still retain polarity informa- tion at a level comparable to that of the full review. Much of the research in automated opinion detection has been performed and proposed for discriminating between subjective and objective text at the document and sentence levels (Bruce & Wiebe, 1999; Finn, Kushmerick, & Smyth, 2002; Hatzivassiloglou & Wiebe, 2000; Wiebe, 2000; Wiebe et al., 1999; Wiebe et al., 2002; Yu & Hatzivassiloglou, 2003). In this section, we will discuss some approaches used to automatically assign one document as objective or subjective. 4.1. Similarity approach Similarity approach to classifying sentences as opinions or facts explores the hypothesis that, within a given topic, opinion sen- tences will be more similar to other opinion sentences than to fac- tual sentences (Yu & Hatzivassiloglou, 2003). Similarity approach measures sentence similarity based on shared words, phrases, and WordNet synsets (Dagan, Shaul, & Markovitch, 1993; Dagan, Pereira, & Lee, 1994; Leacock & Chodorow, 1998; Miller & Charles, 1991; Resnik, 1995; Zhang, Xu, & Callan, 2002). To measure the overall similarity of a sentence to the opinion or fact documents, we need to go through three steps. First, use IR method to acquire the documents that are on the same topic as the sentence in question. Second, calculate its similarity scores with each sentence in those documents and make an average va- lue. Third, assign the sentence to the category (opinion or fact) for which the average value is the highest. Alternatively, for the frequency variant, we can use the similarity scores or count how many of them for each category, and then compare it with a prede- termined threshold. 4.2. Naive Bayes classifier Naive Bayes classifier is a commonly used supervised machine learning algorithm. This approach presupposes all sentences in opinion or factual articles as opinion or fact sentences. Naive Bayes uses the sentences in opinion and fact documents as the examples of the two categories. The features include words, bigrams, and trigrams, as well as the part of speech in each sen- tence. In addition, the presence of semantically oriented (positive and negative) words in a sentence is an indicator that the sentence is subjective. Therefore, it can include the counts of positive and negative words in the sentence, as well as counts of the polarities of sequences of semantically oriented words (e.g., ‘‘++” for two con- secutive positively oriented words). It also include the counts of parts of speech combined with polarity information (e.g., ‘‘JJ+” for positive adjectives), as well as features encoding the polarity (if any) of the head verb, the main subject, and their immediate modifiers. Generally speaking, Naive Bayes assigns a document dj (repre- sented by a vector d�j ) to the class ci that maximizes Pðcijd � j Þ by applying Bayes’ rule as follow, Pðcijd � j Þ¼ PðciÞPðd � j jciÞ Pðd�j Þ ð1Þ where Pðd�j Þ is the probability that a randomly picked document d has vector d�j as its representation, and P(c) is the probability that a randomly picked document belongs to class c. To estimate the term Pðd�j jcÞ, Naive Bayes decomposes it by assuming all the features in d�j (represented by fi,i = 1 to m) are con- ditionally independent, i.e., Pðcijd � j Þ¼ PðciÞ Qm i¼1 PðfijciÞ � � Pðd�j Þ ð2Þ 4.3. Multiple Naive Bayes classifier The hypothesis of all sentences in opinion or factual articles as opinion or fact sentences is an approximation. To address this, multiple Naive Bayes classifier approach applies an algorithm using multiple classifiers, each relying on a different subset of fea- tures. The goal is to reduce the training set to the sentences that are most likely to be correctly labeled, thus boosting classification accuracy. Given separate sets of features F1, F2, . . . , Fm, it train separate Na- ive Bayes classifiers C1, C2, . . . , Cm corresponding to each feature set. Assuming as ground truth the information provided by the docu- ment labels and that all sentences inherit the status of their docu- ment as opinions or facts, it first train C1 on the entire training set, H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 10763 then use the resulting classifier to predict labels for the training set. The sentences that receive a label different from the assumed truth are then removed, and train C2 on the remaining sentences. This process is repeated iteratively until no more sentences can be removed. Yu and Hatzivassiloglou (2003) report results using five feature sets, starting from words alone and adding in bigrams, trigrams, part-of-speech, and polarity. 4.4. Cut-based classifier Cut-based classifier approach put forward a hypothesis that, text spans (items) occurring near each other (within discourse boundaries) may share the same subjectivity status (Pang & Lee, 2004). Based on this hypothesis, Pang supplied his algorithm with pair-wise interaction information, e.g., to specify that two particu- lar sentences should ideally receive the same subjectivity label. This algorithm uses an efficient and intuitive graph-based formula- tion relying on finding minimum cuts. Suppose there are n items x1, x2, . . . , xn to divide into two classes C1 and C2, here access to two types of information: indj(xi): Individual scores. It is the non-negative estimates of each xi’s preference for being in Cj based on just the features of xi alone; assoc(xi, xk): Association scores. It is the non-negative estimates of how important it is that xi and xk be in the same class. Then, this problem changes to calculate the maximization of each item’s score for one class: its individual score for the class it is assigned to, minus its individual score for the other class, then minus associated items into different classes for penalization. Thus, after some algebra, it arrives at the following optimization problem: assign the xi to C1 and C2 so as to minimize the partition cost:X x2C1 ind2ðxÞþ X x2C2 ind1ðxÞþ X xi2C1;xk2C2 assocðxi; xkÞ ð3Þ This situation can be represented in the following manner. Build an undirected graph G with vertices {v1, . . . , vn, s, t}; the last two are, respectively, the source and sink. Add n edges (s, vi), each with weight ind1(xi), and n edges (vi, t), each with weight ind2(xi). Finally, add ðC2nÞ edges (vi, vk), each with weight assoc(xi, xk). A cut (S, T) of G is a partition of its nodes into sets S = {s}US0 and T = {t}UT0, where s R S0, t R T0. Its cost cost(S, T) is the sum of the weights of all edges crossing from S to T. A minimum cut of G is one of minimum cost. Then, finding solution of this problem is changed into looking for a minimum cut of G. 5. Word sentiment classification The task on document sentiment classification has usually in- volved the manual or semi-manual construction of semantic orien- tation word lexicons (Hatzivassiloglou & McKeown, 1997; Hatzivassiloglou & Wiebe, 2000; Lin, 1998; Pereira, Tishby, & Lee, 1993; Riloff, Wiebe, & Wilson, 2003; Turney & Littman, 2002; Wiebe, 2000), which built by word sentiment classification tech- niques. For instance, Das and Chen (2001) used a classifier on investor bulletin boards to see if apparently positive postings were correlated with stock price, in which several scoring methods were employed in conjunction with a manually crafted lexicon. Classify- ing the semantic orientation of individual words or phrases, such as whether it is positive or negative or has different intensities, generally using a pre-selected set of seed words, sometimes using linguistic heuristics (For example, Lin (1998) & Pereira et al. (1993) used linguistic co-locations to group words with similar uses or meanings). Some studies showed that restricting features to those adjec- tives for word sentiment classification would improve perfor- mance (Andreevskaia & Bergler, 2006; Turney & Littman, 2002; Wiebe, 2000). However, more researches showed most of the adjectives and adverb, a small group of nouns and verbs possess semantic orientation (Andreevskaia & Bergler, 2006; Esuli & Sebas- tiani, 2005; Gamon & Aue, 2005; Takamura, Inui, & Okumura, 2005; Turney & Littman, 2003). Automatic methods of sentiment annotation at the word level can be grouped into two major categories: (1) corpus-based ap- proaches and (2) dictionary-based approaches. The first group in- cludes methods that rely on syntactic or co-occurrence patterns of words in large texts to determine their sentiment (e.g., Hatzi- vassiloglou & McKeown, 1997; Turney & Littman, 2002; Yu & Hat- zivassiloglou, 2003 and others). The second group uses WordNet (http://wordnet.princeton.edu/) information, especially, synsets and hierarchies, to acquire sentiment-marked words (Hu & Liu, 2004a; Kim & Hovy, 2004) or to measure the similarity between candidate words and sentiment-bearing words such as good and bad (Kamps, Marx, Mokken, & de Rijke, 2004). 5.1. Analysis by conjunctions between adjectives This method attempts to predict the orientation of subjective adjectives by analyzing pairs of adjectives (conjoined by and, or, but, either-or, or neither-nor) which are extracted from a large unlabelled document set. The underlying intuition is that the act of conjoining adjectives is subject to linguistic constraints on the orientation of the adjectives involved (e.g. and usually conjoins two adjectives of the same-orientation, while but conjoins two adjectives of opposite orientation). This is shown in the following three sentences (where the first two are perceived as correct and the third is perceived as incorrect) taken from Hatzivassiloglou and McKeown (1997): ‘‘The tax proposal was simple and well received by the public”. ‘‘The tax proposal was simplistic but well received by the public”. ‘‘The tax proposal was simplistic and well received by the public”. To infer the orientation of adjectives from analysis of conjunc- tions, a supervised learning algorithm can be performed as follow- ing steps: 1. All conjunctions of adjectives are extracted from a set of documents. 2. Train a log-linear regression classifier and then classify pairs of adjectives either as having the same or as having different ori- entation. The hypothesized same-orientation or different-orien- tation links between all pairs form a graph. 3. A clustering algorithm partitions the graph produced in step 2 into two clusters. By using the intuition that positive adjectives tend to be used more frequently than negative ones, the cluster containing the terms of higher average frequency in the docu- ment set is deemed to contain the positive terms. The log-linear model offers an estimate of how good each pre- diction is, since it produces a value y between 0 and 1, in which 1 corresponds to same-orientation, and one minus the produced value y corresponds to dissimilarity. Same- and different-orienta- tion links between adjectives form a graph. To partition the graph nodes into subsets of the same-orientation, the clustering algo- rithm calculates an objective function U scoring each possible par- tition P of the adjectives into two subgroups C1 and C2 as, UðPÞ¼ X2 i¼1 1 jCij X x;y2Ci;x – y dðx; yÞ ! ð4Þ where jCij is the cardinality of cluster i, and d(x, y) is the dissimilarity between adjectives x and y. http://wordnet.princeton.edu/ 10764 H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 In general, because the model was unsupervised, it required an immense word corpus to function. 5.2. Analysis by lexical relations This method presents a strategy for inferring semantic orienta- tion from semantic association between words and phrases. It fol- lows a hypothesis that two words tend to be the same semantic orientation if they have strong semantic association. Therefore, it focused on the use of lexical relations defined in WordNet to calcu- late the distance between adjectives. Generally speaking, we can defined a graph on the adjectives contained in the intersection between a term set (For example, TL term set (Turney & Littman, 2003)) and WordNet, adding a link between two adjectives whenever WordNet indicates the presence of a synonymy relation between them, and defining a distance measure using elementary notions from graph theory. In more de- tail, this approach can be realized as following steps: 1. Construct relations at the level of words. The simplest approach here is just to collect all words in WordNet, and relate words that can be synonymous (i.e., they occurring in the same synset). 2. Define a distance measure d(t1, t2) between terms t1 and t2 on this graph, which amounts to the length of the shortest path that connects t1 and t2 (with d(t1, t2) = +1 if t1 and t2 are not connected). 3. Calculate the orientation of a term by its relative distance (Kamps et al., 2004) from the two seed terms good and bad, i.e., SOðtÞ¼ dðt; badÞ� dðt; goodÞ dðgood; badÞ ð5Þ 4. Get the result followed by this rules: The adjective t is deemed to belong to positive if SO(t) > 0, and the absolute value of SO(t) determines, as usual, the strength of this orientation (the con- stant denominator d(good, bad) is a normalization factor that constrains all values of SO to belong to the [�1, 1] range). 5.3. Analysis by glosses The characteristic of this method lies in the fact that it exploits the glosses (i.e. textual definitions) that one term has in an online ‘‘glossary”, or dictionary. Its basic assumption is that if a word is semantically oriented in one direction, then the words in its gloss tend to be oriented in the same direction (Esuli & Sebastiani, 2005; Esuli & Sebastiani, 2006a, 2006b). For instance, the glosses of good and excellent will both contain appreciative expressions; while the glosses of bad and awful will both contain derogative expressions. Generally, this method can determine the orientation of a term based on the classification of its glosses. The process is composed of the following steps: 1. A seed set (Sp, Sn), representative of the two categories positive and negative, is provided as input. 2. Search new terms to enrich Sp and Sn. Use lexical relations (e.g. synonymy) with the terms contained in Sp and Sn from a thesau- rus, or online dictionary, to find these new terms, and then append them to Sp or Sn. 3. For each term ti in S 0 p [ S 0 n or in the test set (i.e. the set of terms to be classified), a textual representation of ti is generated by collating all the glosses of ti as found in a machine-readable dic- tionary. Each such representation is converted into a vector by standard text indexing techniques. 4. A binary text classifier is trained on the terms in S0p [ S 0 n and then applied to the terms in the test set. 5.4. Analysis by both lexical relations and glosses This method determines sentiment of words and phrases both relies on lexical relations (synonymy, antonymy and hyponymy) and glosses provided in WordNet. Andreevskaia and Bergler (2006) proposed an algorithm named ‘‘STEP” (Semantic Tag Extraction Program). This algorithm starts with a small set of seed words of known sentiment value (positive or negative) and implements the following steps: 1. Extend the small set of seed words by adding synonyms, ant- onyms and hyponyms of the seed words supplied in WordNet. This step brings on average a 5-fold increase in the size of the original list with the accuracy of the resulting list comparable to manual annotations. 2. Go through all WordNet glosses, identifies the entries that con- tain in their definitions the sentiment-bearing words from the extended seed list, and adds these head words to the corre- sponding category – positive, negative or neutral. 3. Disambiguate the glosses with part-of-speech tagger, and elim- inate errors of some words acquired in step 1 and from the seed list. At this step, it also filters out all those words that have been assigned contradicting. In this algorithm, for each word we need compute a Net Overlap Score by subtracting the total number of runs assigning this word a negative sentiment from the total of the runs that consider it posi- tive. In order to make the Net Overlap Score measure usable in sen- timent tagging of texts and phrases, the absolute values of this score should be normalized and mapped onto a standard [0, 1] interval. STEP accomplishes this normalization by using the value of the Net Overlap Score as a parameter in the standard fuzzy mem- bership S-function (Zadeh, 1987). This function maps the absolute values of the Net Overlap Score onto the interval from 0 to 1, where 0 corresponds to the absence of membership in the category of sentiment (in this case, these will be the neutral words) and 1 re- flects the highest degree of membership in this category. The func- tion can be defined as follows, Sðu; a; b; cÞ¼ 0 if u 6 a 2 u�ac�a � �2 if a 6 u 6 b 1 � 2 u�ac�a � �2 if b 6 u 6 c 1 if u P c 8>>>>>>< >>>>>>: ð6Þ where u is the Net Overlap Score for the word and a, b, c are the three adjustable parameters: a is set to 1, c is set to 15 and b, which represents a crossover point, is defined as b = (a + c)/2 = 8. Defined this way, the S-function assigns highest degree of membership (=1) to words that have the Net Overlap Score u P 15. Net Overlap Score can be used as a measure of the words degree of membership in the fuzzy category of sentiment: the core adjec- tives, which had the highest Net Overlap Score, were identified most accurately both by STEP and by human annotators, while the words on the periphery of the category had the lowest scores and were associated with low rates of inter-annotator agreement. 5.5. Analysis by pointwise mutual information The general strategy of this method is to infer semantic orienta- tion from semantic association. The underlying assumption is that a phrase has a positive semantic orientation when it has good asso- ciations (e.g., ‘‘romantic ambience”) and a negative semantic orien- tation when it has bad associations (e.g., ‘‘horrific events”) (Turney, 2002). H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 10765 The semantic orientation of a given word is calculated from the strength of its association with a set of positive words, minus the strength of its association with a set of negative words. More con- cretely, the strength of the semantic association between words can express by calculating their pointwise mutual information (PMI) value. So, it focuses on inferring the semantic orientation of a word from its statistical association with a set of positive and negative paradigm words. Given a term t, and seed term sets (Sp for positive set and Sn for negative set), the t’s orientation value O(t) (where positive value means positive orientation, and higher absolute value means stronger orientation) is given by: OðtÞ¼ X ti2Sp PMIðt; tiÞ� X ti2Sn PMIðt; tiÞ ð7Þ Pointwise mutual information can be computed based on IR tech- niques. Term frequencies and co-occurrence frequencies are mea- sured by querying a document set by means of a search engine with a ‘‘t” query, a ‘‘ti” query, and a ‘‘t NEAR ti” query, and using the number of matching documents returned by the search engine as estimates of the probabilities needed for the computation of PMI. In the AltaVista search engine (http://www.altavista.com/), the NEAR operator produces a match for a document when its oper- ands appear in the document at a maximum distance of ten terms, in either order. The paradigm words can be selected as following (Turney & Littman, 2003): Sp ¼fgood; nice; excellent; positive; fortunate; correct; superiorg Sn ¼fbad; nasty; poor; negative; unfortunate; wrong; inferiorg In addition, Gamon and Aue (2005) described an extension to the technique for the automatic identification and labeling of sen- timent terms described in Turney and Littman (2003). Besides the basic assumption in Turney and Littman (2003), Turney (2002), Gamon and Aue (2005) adds a second assumption, namely that sentiment terms of opposite orientation tend not to co-occur at the sentence level. This additional assumption allows them to identify sentiment-bearing terms more reliably to some extent. 5.6. Analysis by general inquirer General Inquirer (GI) is a system which lists terms as well as dif- ferent senses for the terms. For each sense it provides a short def- inition as well as other information about the term. This includes tags that label the term as being positive, negative, a negation term, an overstatement, or an understatement. The labels are for each sense of a word. For example, there are two senses of the word fun as seen in Table 1. One sense is a noun or adjective for enjoyment or enjoyable. The second sense is a verb that means to ridicule or tease, to make fun of. The first sense of the word is positive, while the second is negative. The entry also indicates that the first sense is more fre- quent than the second sense (estimated to occur 97% of the times while the second sense occurs only 3% of the times). In addition, the GI dictionary includes negations, intensifiers, and diminishers. Table 2 shows the GI entries of the words not, fantastic and barely which are examples of a negation, an overstatement and an understatement, respectively. The GI dictionary contains 1,915 positive senses and 2,291 neg- ative senses. Kennedy and Inkpen (2006) add more positive and negative senses from Choose the Right Word (Hayakawa, 1994) Table 1 GI entries for the word fun. Fun (sense 1) H4Lvd Positiv Pstv Pleasur Ex Fun (sense 2) H4Lvd Negativ Ngtv Hostile C (hereafter CTRW). CTRW is a dictionary of synonyms, which lists nuances of lexical meaning. After adding them, they obtain 1955 positive senses and 2398 negative senses. There are 696 overstate- ments and 319 understatements in GI. When adding those from CTRW, they obtain 1269 overstatements and 412 understatements. Kennedy and Inkpen (2006) used this approach to classify reviews based on the number of positive and negative terms they contain. They examined the effect of three types of valence shifters: negations, intensifiers, and diminishers. In: GI intensifiers are known as overstate- ments and diminishers are known as understatements. Negations are used to reverse the semantic polarity of a particular term, while intensifiers and diminishers are used to increase and decrease, respec- tively, the degree to which a term is positive or negative. 6. Document sentiment classification based on machine learning methods 6.1. Single domain There are many possible approaches to identifying the actual polarity of a document. Here we talk about using supervised ma- chine learning techniques to identify the likelihood of reviews hav- ing ‘‘positive” or ‘‘negative” polarity with respect to previously hand-classified training data. The key problem of this method in- cludes two aspects: extracting feature and training classifier. 6.1.1. Extracting feature Starting with a raw document (a portion of a web page in test- ing and training, and a complete web page for mining), then strip out HTML tags and separate the document into sentences. These sentences are optionally run through a parser before being split into single-word tokens. A variety of transformations can then be applied to this ordered list of lists. This is called feature extraction. There are several methods for feature extraction: 1. Lexical filtering Lexical filtering can apply to reviews, on the accuracy of statis- tical classifiers trained on such filtered data. There are mainly two different kinds of lexical filters used (Salvetti, Lewis, & Rei- chenbach, 2004): one based on hypernymy as provided by WordNet (Budanitsky & Hirst, 2001; Curran, 2002; Devitt & Vogel, 2004; Edmonds, 1999; Edmonds & Hirst, 2002; Fellbaum, 1998; Grefenstette, 1994; Justeson & Slava, 1993; Kamps & Maarten, 2002; Rapp, 2004; Riloff & Shepherd, 1997; Roark & Charniak, 1998; Taboada, 2006; Thelen & Riloff, 2002; Valitutti, Strapparava, & Stock, 2004), the other based on part-of-speech (POS) tags (Ait-Mokhtar, Chanod, & Roux, 2002; Brill, 1992; Brill, 1994; Brill, 1995; Losee, 2001; Ratnaparkhi, 1996; Schmid, 1994; Wiebe, Wilson, & Bell, 2001; Wiebe, Wilson, & Cardie, 2005; Wilks & Stevenson, 1998). WordNet filter attempted to substitute synonyms in reviews by a set of likely synonyms and hypernymy generalization in WordNet, because it is uncommon to encounter repetitions of identical words in non-technical written text. POS filter considered that individual phrases and words vary in their contribution to opinion polarity. It may even be said that only some part of the meaning of a word contributes to opinion polarity. Any portion that does not contribute to sentiment detection is noise. POS filters were developed to reduce these noises. prsv WlbPsyc WlbTot Noun PFREQ 97% noun–adj: Enjoyment, enjoyable omForm SV RspLoss RspTot SUPV 3% idiom–verb: Make fun (of) – to tease, parody http://www.altavista.com/ Table 2 GI entries for the words not, fantastic and barely. Not H4Lvd negate NotLw LY – adv: expresses negation NotLw LY adv: expresses negation Fantastic H4Lvd Positiv Pstv Virtue Ovrst EVAL PosAff Modif – Virtue Ovrst EVAL PosAff Modif Barely H4Lvd Undrst Quan If LY – Quan If LY 10766 H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 2. Appraising adjective Appraising adjective method (Whitelaw, Argamon, & Garg, 2005; Whitelaw, Garg, & Argamon, 2005) focuses on the extraction and analysis of adjectival appraisal groups headed by an appraising adjective (such as ‘‘beautiful” or ‘‘boring”) and optionally modified by a sequence of modifiers (such as ‘‘very”, ‘‘sort of”, or ‘‘not”). It made a more detailed semantic analysis of attitude expressions, in the form of a well-designed taxonomy of attitude types and other semantic properties. Furthermore, it treats ‘‘atomic units” of such expressions not with individual words, but with appraisal groups: coherent groups of words that express together a particu- lar attitude, such as ‘‘extremely boring”, or ‘‘not really very good”. This method can be described as following steps: 1. Build a lexicon using semi-automatic techniques, gathering and classifying adjectives and modifiers to categories in several taxonomies of appraisal attributes. 2. Extract adjectival appraisal groups from texts and compute their attribute values according to this lexicon. 3. Represent documents as vectors of relative frequency features using these groups. 4. Train a support vector machine algorithm discriminating posi- tively from negatively oriented test documents. Beineke, Hastie, and Vaithyanathan (2004) extend this proce- dure by extract a pair of derived features that are linearly com- bined to predict sentiment. This perspective allows to improving upon previous methods, primarily through two strategies: incorpo- rating additional derived features into the model and, where pos- sible, using labeled data to estimate their relative influence. Matsumoto, Takamura, and Okumura (2005) used text-mining techniques to extract frequent word sub-sequences and depen- dency sub-trees from sentences in a document dataset and used them as features of support vector machines. 6.1.2. Training classifier Several classical text classifiers, such as K-Nearest Neighbor, Winnow, Naı̈ve Bayes, Maximum Entropy and Support Vector Ma- chine (SVM), are used in machine learning method for document sentiment classification. Pang et al. (2002) applied Naı̈ve Bayes, Maximum Entropy and Support Vector Machine classification techniques to the identifica- tion of the polarity of movie reviews. Their best result (82.9% accu- racy) was obtained by using unigrams and Support Vector Machine. In fact, most of the literatures showed that SVM and Naı̈ve Bayes are perfect methods in single domain document sentiment classification (Aue and Gamon, 2005; Beineke et al., 2004; Kennedy and Inkpen, 2006; Lin et al., 2006; Matsumoto et al., 2005; Mullen and Collier, 2004; Pang and Lee, 2004; Pang et al., 2002; Read et al., 2005; Salvetti et al., 2004; Whitelaw, Garg, et al., 2005). 6.2. Multiple domains Document sentiment classification is a very domain-specific problem: classifiers trained in one domain do not perform well in others Charlotta (2004), Nigam, McCallum, and Thrun (2000), Lafferty, McCallum, and Pereira (2001), Avrim and Chawla (2001). The main factor is that the polarity of sentiment words may change with the domain too. For instance, the word ‘‘small” in house review is negative (e.g. ‘‘The bedroom is very small”), while in cell-phone review is positive (e.g. ‘‘The Nokia N3100 is so small as to be put in any pockets”). Tan, Wu, Tang, and Cheng (2007) attempted to tackle domain- transfer problem by combining old-domain labeled examples with new domain unlabeled ones. Their basic idea is to use old-domain- trained classifier (‘‘old classifier” for brevity) to label top n most informative unlabeled examples in new domain and learn a new classifier based on these selected examples (n is a pre-defined number indicating how many examples in new domain shall be picked out as informative ones). Detailed algorithm for proposed scheme is: 1. Train a base classifier using labeled data in old-domain. 2. Label some informative unlabeled ones in new domain. 3. Train a new classifier based on these selected examples. 4. Classify examples in new domain using new classifier. The experimental results demonstrate their proposed scheme can boost the accuracy of the base sentiment classifier on new domain. Aue and Gamon (2005) surveyed four different approaches to customizing a sentiment classification system to a new target do- main with a small amount of labeled data. Read et al. (2005) pro- posed a novel source of training data based on the language used in conjunction with emoticons in Usenet newsgroups. Then they used emoticons-labeled data to train a classifier, which has the po- tential to reduce dependent of domain, topic and time. 7. Opinion extraction Analysis of favorable and unfavorable opinions is a task requir- ing high intelligence and deep understanding of the textual con- text, drawing on common sense and domain knowledge as well as linguistic knowledge. The interpretation of opinions can be debatable even for humans. For example, when we tried to deter- mine if each specific document was on balance favorable or unfa- vorable toward a subject after reading an entire group of such documents, we often found it difficult to reach a consensus, even for very small groups of evaluators. Many researchers think it is too coarse to compute a unit of an opinion for a whole document (Bai, Padman, & Airoldi, 2004; Bet- hard, Yu, Thornton, Hatzivassiloglou, & DanJurafsky, 2004; Bune- scu & Mooney, 2004; Daille, 1996; Didier, 1995; Etzioni et al., 2004; Freitag & McCallum, 2000; Jacquemin & Bourigault, 2001; Luca & Mazzini, 2002; Riloff & Jones, 1999; Wilson, Wiebe, & Hwa, 2004; Zhai & Liu, 2005). Turney (2002) makes a similar point, noting that for reviews, ‘‘the whole is not necessarily the sum of the parts”. Even in a single sentence, a holder might express two different opinions. It seems necessary using more sophisticated techniques to determine the focus of each sentence, so that one can decide whether the author is talking about the topic (Dave, Lawrence, & Pennock, 2003; Fei, Liu, & Wu, 2004; Kim & Hovy, 2004; Ku, Wu, Lee, & Chen, 2005; Mei, Ling, Wondra, Su, & Zhai, 2007). H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 10767 Consequently, opinion extraction plays a very important role in sentiment detection. It not only focuses on extracting opinion information from reviews, but also on extracting relation between opinion and document topic. 7.1. Opinion information extraction The opinion information we mainly discussed in this paper in- cludes opinion-bearing word and opinion holder. An opinion-bear- ing word is a word or a phrase that carries a positive or negative sentiment directly such as ‘‘good”, ‘‘bad”, ‘‘foolish”, ‘‘virtuous”, etc. An opinion holder is an entity (person, organization, country, or special group of people) who expresses explicitly or implicitly the opinion contained in the sentence. For instance: ‘‘According to the review in Internet, the keypad of Ericsson W810i is easy to use”. In above sentence, ‘‘the review in Internet” is an opinion holder; ‘‘easy” is an opinion-bearing word. 7.1.1. Opinion-bearing word extraction Opinions can be recognized from various granularities such as a word, a sentence, a text, or even multiple texts, and each is impor- tant. Here we focus on word level opinion detection, i.e., finding words or phrases that carry a positive or negative sentiment (opin- ion-bearing word) from subjectivity sentences or paragraphs. Actually, opinion-bearing word is the smallest unit of opinion that can thereafter be used as a clue for sentence-level or text-level opinion detection. How to extract opinion-bearing words from document? A straightforward way can be conducted by following steps: 1. Collect sentiment words from sentiment lexicon as opinion- bearing seed words. Opinion-bearing seed words can be col- lected from several sources: General Inquirer (GI), Dictionary of Affect of Language (DAL), and WordNet (Kim & Hovy, 2006; Yi, Nasukawa, Bunescu, & Niblack, 2003). 2. Expand those selected opinion-bearing seed words of each sen- timent class by collecting synonyms from WordNet. However, it cannot simply assume that all the synonyms of positive words are positive since most words could have synonym relation- ships with both positive and negative classes. It should calcu- late the closeness of a given word to each category and determine the most probable class (Kim & Hovy, 2006). 3. Refine some of the sentiment patterns from sentiment lexicon and training datasets (Yi et al., 2003). For GI and DAL, the sen- timent verb extraction is the same as the opinion-bearing seed words extraction. For WordNet, the sentiment verb can be extracted from the emotion cluster. The other sentiment pat- terns can be manually refined from the training datasets. Senti- ment pattern database contains sentiment extraction patterns for sentence predicates. 4. Extract sentences and text fragments from input documents containing subjectivity information (the approach we have dis- cussed in Section 4). Apply sentiment analysis with expanded opinion-bearing seed words and sentiment patterns to those subjectivity sentences and text fragments. At last, the opin- ion-bearing words were extracted. 7.1.2. Opinion holder extraction The goal of opinion holder extraction is to identify direct and indirect sources of opinions, emotions, sentiments, and other pri- vate states that are expressed in text. Identifying opinion sources (or opinion holder) will be especially critical for opinion-oriented question–answering systems (e.g., systems that answer questions of the form ‘‘Who feels about . . .?”) and opinion-oriented summa- rization systems, both of which need to distinguish the opinions of one source from another. There are mainly three research points in opinion holder extrac- tion. The first is: given a sentence, identifying opinion sources in it (Choi, Cardie, Riloff, & Patwardhan, 2005); the second is: given an opinion expression in a sentence, identifying its corresponding opinion sources (Kim & Hovy, 2006); the third is: given an opinion expression and a source, determines whether the source and the opinion expression is correspond (Choi, Breck, & Cardie, 2006). 1. The first research point The first research point learns patterns of opinion sources using a graphical model and extraction pattern learning. It views the opin- ion holder extraction problem as an information extraction task and adopts a hybrid approach that combines CRFs (Conditional Random Fields) and a variation of AutoSlog (a supervised extrac- tion pattern learner that takes a training corpus of texts and their associated answer keys as input) (Riloff, 1996). While CRFs identi- fies source and AutoSlog learns extraction patterns.It starts with a sequence of words (x1, x2, . . . , xn) in a sentence and then: (1) Generate a sequence of labels (y1, y2, . . . , yn) indicating whether the word is a holder or not. (2) Presents a new variation of AutoSlog, AutoSlog-SE, which generates patterns to extract sources. This algorithm is not perfect, however, so the resulting set of patterns needs to be manually reviewed by a person. In order to build a fully automatic system which has no use for manual review, Choi et al. (2005) combined AutoSlog’s heuristics with statistics from the annotated training data to create a fully auto- matic supervised learner. 2. The second research pointThe second research point uses clas- sification and ranking to model the problem with Maximum Entropy (ME) (Kim & Hovy, 2006). Classification allocates each holder candidate to a set of pre-defined classes while ranking selects a single candidate as answer. It chooses the most prob- able candidate via a conditional probability.This method acts as following steps. (1) Generate all possible holder candidates, given a sentence and an opinion expression hEi. After parsing the sentence, it extracts features such as the syntactic path information between each candidate hHi and the expression hEi and a distance between hHi and hEi. (2) Rank holder candidates according to the score obtained by the ME ranking model, and pick the candidate with the highest score. Given a set of holder candidates {h1, h2, . . . , hn} and opin- ion expression e. The conditional probability P(hj{h1, h2, . . . , hn}, e) can be calculated based on K feature functions fk(hj{h1, h2, . . . , hn}, e), as follows, h ¼ argmaxh½Pðhjfh1; h2; . . . ; hNg; eÞ� ¼ argmaxsh XK k¼1 kk fkðhjfh1; h2; . . . ; hNg; eÞ " # ð8Þ where each kk is a model parameter indicating the weight of its feature function. 3. The third research point The third research point (Choi et al., 2006) aimed to identify the opinion holder via extracting relations between opinion expres- sion entities and source entities. That is, given opinion expres- sion Oi and source Sj, it determines whether Sj is the source of opinion expression Oi. The global inference procedure is imple- mented via integer linear programming (ILP) to produce an optimal and coherent extraction of entities and relations. ILP formulation group consists of an objective function and a set of equality and inequality constraints among variables. The fol- lowing formulation is the objective function of binary ILP, 10768 H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 f ¼ X i ðwoi OiÞþ X i ðw�oi O � i Þþ X j ðwsj SjÞþ X j ðw�sj S � j Þ þ X i;j ðwli;j Li;jÞþ X i ðw�li;j L � i;jÞ ð9Þ 8i; Oi þ O�i ¼ 1; 8j; Sj þ S � i ¼ 1; 8i; j; Li;j þ L � i;j ¼ 1 ð10Þ where f is the objective function of ILP. Oi and O � i are two variables for each opinion entity, Oi = 1 means to extract the opinion entity, and O�i ¼ 1 means to discard the opinion entity. woi and w�oi are weights for Oi and O � i , respectively, which are computed based on the labels of the adjacent variables of the CRFs (Choi et al., 2006). Likewise, Sj and S � j are two variables for each source entity to indi- cate whether it had been extracted, their weights wsj and w�sj are computed in the same way as opinion entities; Li,j and L � i;j are two variables for each link relation between Oi and Sj. Li,j = 1 indicate Oi and Sj both be extracted, otherwise, L � i;j ¼ 1. Their weights wlij and w�lij are based on probabilities from the binary link classifier.The following is a set of equality and inequality constraints among variables. 8i; Oi ¼ X j Li;j ð11Þ 8j; Sj þ Aj ¼ X i Li;j ð12Þ 8j; Aj � Sj 6 0 ð13Þ 8i; j; i < j; Xi þ Xj ¼ 1; X 2fS; Og ð14Þ Formula (11) enforces that only one link can emanate from an opin- ion entity. Formulas (12) and (13) together allow a source to link to at most two opinions, where Aj is an auxiliary variable between 0 and 1. Aj can be assigned to 1 only if Sj is already assigned to 1. For- mula (14) is used to restrict all pairs of entities with overlapping spans. 7.2. Opinion-topic relation extraction Opinion-topic relation refers to relationship between opinion expression (opinion-bearing words) and document topic (or fea- ture of topic). A feature of a topic is a term that satisfies one of the following relationships: � A part-of relationship with the given topic. � An attribute-of relationship with the given topic. � An attribute-of relationship with a known feature of the given topic. For instance, ‘‘I found the Sony Ericsson W810i is a good mobile phone. The keypad is easy to use and texting is very simple as the buttons are small but well defined. The screen is bright and clear with good resolution. The software on this phone is excellent. I have had absolutely no problems with it what so ever, it never crashes, freezes or otherwise upsets me. The battery life on this phone is excellent. I have had mine for 18 months, with it rarely being switched off in all this time. In my normal use (only about 30–60 min calls a day) it will last for about 3–4 days before needing a charge. Charging is very quick too taking only around an hour to fully charge from about 10%. All in all an excellent little phone with very few faults. Recommended!” In above text, ‘‘good”, ‘‘easy to use”, ‘‘bright” etc., are opinion expression, ‘‘Sony Ericsson W810i” is topic, and ‘‘keypad”, ‘‘screen”, ‘‘software” etc., are features of topic. 7.2.1. Feature term extraction Yi et al. (2003) put forward the following three candidates as feature term to be extracted: 1. Base Noun Phrases (BNP). BNP restricts the candidate feature terms to one of the following base noun phrase (BNP) patterns: NN, NN NN, JJ NN, NN NN NN, JJ NN NN, JJ JJ NN, where NN and JJ are the part-of-speech (POS) tags for nouns and adjectives, respectively. 2. Definite Base Noun Phrases (dBNP). dBNP further restricts can- didate feature terms to definite base noun phrases, which are noun phrases of the form defined above that are preceded by the definite article ‘‘the”. Given that a document is focused on a certain topic, the definite noun phrases referring to topic fea- tures do not need any additional constructs such as attached prepositional phrases or relative clauses, in order for the reader to establish their referent. Thus, the phrase ‘‘the battery,” instead of ‘‘the battery of the digital camera,” is sufficient to infer its referent. 3. Beginning Definite Base Noun Phrases (bBNP). bBNP refers to dBNP at the beginning of sentences followed by a verb phrase. This heuristic is based on the observation that, when the focus shifts from one feature to another, the new feature is often expressed using a definite noun phrase at the beginning of the next sentence. They developed and tested two feature term selection algo- rithms based on a mixture language model and likelihood ratio, while the Likelihood Test method gets better result. Following is principle of the Likelihood Test method: Let D+ be a collection of documents focused on a topic T, D� those not focused on T, and bnp a candidate feature term extracted from D+. Then, the likeli- hood ratio �2log k is defined as follows, � 2 log k ¼�2 log maxp16p2 Lðp1; p2Þ maxp1;p2 Lðp1; p2Þ p1 ¼ pðd 2 Dþjbnp 2 dÞ p2 ¼ pðd 2 Dþjbnp 2 dÞ ð15Þ where L(p1, p2) is the likelihood of seeing bnp in both D+ and D�. The higher the value of �2log k, the more likely the bnp is relevant to the topic T. For each bnp, compute the likelihood score, �2log k, as de- fined in formula (15). Then, sort bnp in decreasing order by their likelihood score. Feature terms are all bnp’s whose likelihood ratio satisfying a pre-defined confidence level. Alternatively simply only the top N bnp’s can be selected. 7.2.2. Makes (topic j feature term, opinion) association In order to achieve high precision, we need focus on identifying semantic relationships between opinion expressions and topic (or feature of topic) terms, i.e. extracting opinion-bearing words asso- ciated with polarity of positive or negative for specific topic (or fea- ture of topic) from a document (Agrawal & Srikant, 1994; Jon & Tardos, 2002; Liu, Hsu, & Ma, 1998; Rosario & Hearst, 2004). In order to identify opinion expressions and analyze their semantic relationships with the topic (or feature of topic) term, fol- lowing natural language processing techniques play an important role (Nasukawa & Yi, 2003): 1. POS tagging POS tagging can disambiguate some polysemous expressions such as ‘‘like,” which denote sentiment only when used as a verb instead of as an adjective or preposition. 2. Syntactic parsing Syntactic parsing is used to identify relationships between senti- ment expressions and the subject term. Furthermore, in order to H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 10769 maintain robustness for noisy texts from various sources such as the WWW, it need to use a shallow parsing framework that iden- tifies phrase boundaries and their local dependencies in addition to POS tagging, instead of using a full parser that tries to identify the complete dependency structure among all of the terms. Yi et al. (2003) extracts ternary expressions (T-expressions) and binary expressions (B-expressions), in order to make (topic j feature term, opinion) association. There are two types of T-expressions: 1. positive or negative sentiment verbs:h<target, verb, ‘‘”i 2. trans verbs:htarget, verb, sourceiAnd one format of B-expres- sions:hadjective, targeti For each opinion expressions detected, its target and final polar- ity can be determined by sentiment pattern database (Sentiment pattern database contains sentiment extraction patterns for sen- tence predicates.) If no corresponding sentiment pattern is avail- able, the B-expressions can be created for making the sentiment assignment. From a T-expression, sentiment of the verb (for sentiment verbs) or source (for trans verb), and from a B-expression, sentiment of the adjective, is assigned to the target. 8. Evaluation of sentiment detection As for sentiment analysis systems, the evaluation of sentiment classifiers is typically conducted experimentally, rather than ana- lytically. The reason is that, in order to evaluate a system analyti- cally (e.g., proving that the system is correct and complete), we would need a formal specification of the problem that the system is trying to solve (e.g., with respect to what correctness and com- pleteness are defined), and the central notion of sentiment detec- tion is. The experimental evaluation of a classifier usually measures its effectiveness (rather than its efficiency), that is, its ability to take the right classification decisions. 8.1. Sentiment classification effectiveness 8.1.1. Precision and recall Precision is the ratio of correct cases within the system outputs. Recall is the ratio of correct cases that the system assigned com- pared to the base of all cases where a human analyst associated either positive or negative sentiments manually. In other words, precision and recall are calculated with the following formulas: A = number of all cases that the system assigned either a posi- tive or negative sentiment B = number of all cases that the human assigned either a posi- tive or negative sentiment C = number of correct cases in the system output based on the manual judgment Precision ¼ C A Recall ¼ C B ð16Þ The traditional F-measure value can compute as follow formula: F ¼ 2 � Precision � Recall ðPrecision þ RecallÞ ð17Þ This is also known as the F1 measure, because recall and preci- sion are evenly weighted. Leshed and Kaye (2006) use precision and recall to estimate Blogger mood recognition. Ye, Shi, and Li (2006) distinguish the precision and recall ratios of positive and negative reviews. Classification effectiveness is usually measured in terms of the classic IR notions, like correlation coefficient, relative error, precision, recall, accuracy, etc. However, these values did not seem to provide any better differentiation than simple accuracy. In the domain of sentiment classification of reviews it is often acceptable to sacrifice recall for accuracy. This result is particu- larly interesting for applications that rely on web data, because the customer is not always interested in having all the possible reviews, but many times is interested in having just a few posi- tive and a few negative. From this perspective accuracy is more important than recall. 8.1.2. Correlation coefficient and relative error The correlation coefficient indicates how accurate the senti- ment classification is, showing to what degree the fluctuation pat- terns (e.g., detection of peaks and drops) of a sentiment are predicted by the model (Mishne & de Rijke, 2006). A correlation coefficient of 1 means that there is a perfect linear relation be- tween the prediction and the actual values; whereas a correlation coefficient of 0 means that the prediction is completely unrelated to the actual values. The correlation coefficient is a standard mea- sure of the degree to which two variables are linearly related, and is defined as CorrCoefficient ¼ SPA SP � SA ð18Þ where SPA ¼ P iðpi � �pÞ � ðai � �aÞ n � 1 SP ¼ P iðpi � pÞ 2 n � 1 ; SA ¼ P iðai � aÞ 2 n � 1 ð19Þ where pi is the estimated value for instance i, ai is the actual value for instance i, �x is the average of x, and n is the total number of instances. The relative error denotes the mean difference between the ac- tual values and the estimated ones, and is defined as: RelError ¼ P iðjpi � aijÞP iðjai � ajÞ ð20Þ 8.2. Benchmarks for sentiment detection Generally speaking, different experiments used for comparing only if the experiments satisfy following conditions: 1. On exactly the same collection (i.e., same documents and same categories); 2. With the same ‘‘split” between training set and test set; 3. With the same evaluation measure, and the same parameter values (if have). That is to say, a lack of these three conditions may make the experimental results hardly comparable between each other (see Table 3). We cannot cover the experiments performed on one collection, because it is hardly to find enough number of authors used the same collection in same experimental conditions. Tables 4–6 list the results of all experiments known to us performed on four major datasets, which HM and GI datasets are used for word sentiment classification, while IMDb and polarity datasets are used for docu- ment sentiment classification. 8.2.1. Benchmarks for word sentiment classification The HM dataset consists of 1336 adjectives, 657 positive and 679 negative (Hatzivassiloglou & McKeown, 1997). The GI dataset is a list of labeled words extracted from the General Inquirer lexicon (Stone, Dunphy, Smith, & Ogilvie, 1966). Table 4 Comparative results among different literatures obtained on GI dataset (boldface indicates the best performer on the collection). Author & literature Training set Accuracy Turney and Littman (2002) SO-PMI-IR with a two-billion word corpus .894 SO-LSA with a 10-million word corpus .817 Kamps et al. (2004) WordNet .787 Turney and Littman (2003) SO-PMI with three corpus (AV-ENG, AV-CA, TASA) .971 SO-LSA with corpus TASA .820 Esuli and Sebastiani (2005) WordNet .881 Takamura et al. (2005) WordNet .915 Table 5 Comparative results among different literatures obtained on IMDb dataset (boldface indicates the best performer on the collection). Author & literature Classifier used Accuracy Pang et al. (2002) NB .815 ME .810 SVM .829 Salvetti et al. (2004) NB .795 Markov model .805 Mullen and Collier (2004) SVM .860 Beineke et al. (2004) NB .659 Matsumoto et al. (2005) SVM .883 Table 6 Comparative results among different literatures obtained on polarity dataset (bold- face indicates the best performer on the collection). Author & literature Classifier used Accuracy Pang and Lee (2004) SVM .872 Whitelaw, Garg et al., (2005) SVM .902 Matsumoto et al. (2005) SVM .937 Aue and Gamon (2005) SVM .905 Read et al. (2005) NB .789 SVM .815 Kennedy and Inkpen (2006) SVM .862 Table 3 Comparative results among different literatures obtained on HM dataset (boldface indicates the best performer on the collection). Author & literature Training set Accuracy Hatzivassiloglou and McKeown (1997) A corpus of 21 million words .924 Turney and Littman (2003) SO-PMI with three corpus (AV-ENG, AV-CA, TASA) .982 SO-LSA with corpus TASA .889 Esuli and Sebastiani (2005) WordNet .874 10770 H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 It includes 3596 adjectives, adverbs, nouns, and verbs, in which 1614 are positive and 1982 are negative. 8.2.2. Benchmarks for document sentiment classification The Internet Movie Database (IMDb, http://reviews.imdb.com/ Reviews/) dataset consists of 27,000 movie reviews in HTML form, using 35 different rating scales such as A. . .F or 1. . .10 in addition to the common 5 star system. The Polarity dataset contains 1000 positive and 1000 negative reviews all written before 2002, with a cap of 20 reviews per author (312 author total) per category. 9. Conclusion Sentiment detection has a wide variety of applications in infor- mation systems, including classifying reviews, distinguishing syn- onyms and antonyms, extending the capabilities of search engines, summarizing reviews, tracking opinions in online discussions, and analyzing survey responses. There are likely to be many other applications that we have not anticipated. This paper discusses four related problems, i.e., subjectivity classification, word senti- ment classification, document sentiment classification based on machine learning techniques, and opinion extraction problem. Although we were able to obtain fairly good results for the re- view classification task through the choice of appropriate features and metrics, but we identified a number of issues that make this problem difficult. 1. Since sentiments can be expressed with various expressions including indirect expressions that require common sense rea- soning to be recognized as a sentiment, it’s been a challenge to analyze the complex structures of sentences in the input con- text that negates the local sentiment for the whole. 2. Some reviewers use terms that have negative connotations, but then write an equivocating final sentence explaining that over- all they were satisfied. Mixed reviews introduce considerable noise to the problem of scoring words. 3. Although the overall opinion about a topic is useful, it is only a part of the information of interest. Document level sentiment classification fails to detect sentiment about individual aspects of the topic. In reality, for example, though one could be gener- ally happy about his car, he might be dissatisfied by the engine noise. To the manufacturers, these individual weaknesses and strengths are equally important to know, or even more valuable than the overall satisfaction level of customers. 4. The association of the extracted sentiment to a specific topic is difficult. Most statistical opinion extraction algorithms perform poorly in this aspect. They either assume the topic of the docu- ment is known at the fore, or simply associate the opinion to a topic term co-existing in the same context. But in fact, a docu- ment (or even a portion of a document as small as a sentence) may discuss multiple topics and contain sentiment about multi- ple topics. 5. An accurate identification of semantic orientation requires anal- ysis of units larger than individual words; it requires under- standing of the context in which those words appear. To this end, we intend to use Rhetorical Structure Theory to impose on the text a structure that indicates the relationships among its rhetorical units. In future, more work is needed on further improving and refin- ing techniques mentioned in this paper, and to deal with the out- standing problems identified above. The success story of sentiment detection is also going to encourage an extension of its methods and techniques to neighboring fields of application. A variety of steps can be taken to extend this work: http://reviews.imdb.com/Reviews/ http://reviews.imdb.com/Reviews/ H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 10771 1. In addition to using a larger amount of discriminating terms, additional non-content attributes can be used for sentiment detection, e.g., collect more text marked-up with emoticons. Features which seem promising are the use of emoticons – tex- tual representations of facial expressions – as well as sentiment values of the individual words, and other features. 2. Learn ways to separate out review topic and subtopics (or attri- butes) within reviews and treat them differently. 3. Explore the feasibility of integrating a full parser and various discourse processing methods to better sentence structure anal- ysis, thus better relationship analysis, i.e., associating topic with holders and topic with feature terms. 4. Combine manual and automated generation of lexicon tech- niques, build a more general sentiment lexicon, modify and add sentiment terms for further domains with domain information and sentiment intensity (or sentiment strength), which will be used in multi-domain and multi-class sentiment classification. 5. Look into monitoring (or tracing) of opinions, take the time sequence of reviews into account, and draw the outline of sen- timent mutative trend. An opinion tracking system provides not only text-based and graph-based opinion summaries, but also the trend of opinions from many information sources. 6. Sentiment clustering. It is possible to cluster sentiment accord- ing to the correlation between the temporal behavior of certain sentiment and the temporal behavior of opinion-bearing word frequencies in the text (measuring the occurrences of words over time). 7. Opinion question answering. For instance, who is the opinion holder? Why did the holder propose this opinion? How did the holder express this opinion, and what was the final attitude? Acknowledgements This work was mainly supported by special fund of Chinese Academy of Sciences, ‘‘Research on Opinion Mining of Web Text”, under Grant Number 0704021000 and another four Pro- jects, i.e., 2006AA010105, 2007AA01Z416, 2007CB311100 and 2007AA01Z441. References Agrawal, R., & Srikant, R., (1994). Fast algorithm for mining association rules. In VLDB’94. Ait-Mokhtar, S., Chanod, J. P., & Roux, C. (2002). Robustness beyond shallowness: Incremental deep parsing. Natural Language Engineering, 8(2–3), 121–144. Andreevskaia, A., & Bergler, S. (2006). Mining WordNet for a fuzzy sentiment: Sentiment tag extraction from WordNet glosses. In Proceedings EACL-06, Trento, Italy. Argamon, S., Koppel, M., & Avneri, G. (1998). Routing documents according to style. In First international workshop on innovative information systems. Aue, A., & Gamon, M. (2005). Customizing sentiment classifiers to new domains: A case study. In Proceedings of RANLP. Avrim, B., & Chawla, S. (2001). Learning from labeled and unlabeled data using graph mincuts. In Proceedings 18th international conference on machine learning (ICML) (pp. 19–26). Bai, X., Padman, R., & Airoldi, E. (2004). Sentiment extraction from unstructured text using tabu search-enhanced Markov blanket. In Proceedings of the international workshop on mining for and from the semantic web (pp. 24–35). Beineke, P., Hastie, T., Vaithyanathan, & S. (2004). The sentimental factor: Improving review classification via human-provided information. In Proceedings of the 42nd ACL conference. Bethard, S., Yu, H., Thornton, A., Hatzivassiloglou, V., & DanJurafsky (2004). Automatic extraction of opinion propositions and their holders. In AAAI spring symposium on exploring attitude and affect in text. Brill, E. (1992). A simple rule-based part of speech tagger. In Proceedings of ACL conference on applied natural language processing, Trento, Italy. Brill, E. (1994). Some advances in transformation-based parts of speech tagging. In AAAI. Brill, E. (1995). Transformation-based error-driven learning and natural language processing: A case study in part-of-speech tagging. Computational Linguistics, 21(4), 543–565. Bruce, R. F., & Wiebe, J. (1999). Recognizing subjectivity: A case study in manual tagging. Natural Language Engineering, 5(02), 187–205. Budanitsky, A., & Hirst, G. (2001). Semantic distance in WordNet: An experimental, application-oriented evaluation of five measures. In Workshop on WordNet and other lexical resources, second meeting of the NAACL, Pittsburgh. Bunescu, R. C., & Mooney, R. J. (2004). Collective information extraction with relational markov networks. In ACL’2004. Cardie, C., & Wiebe, J., Wilson, T., & Litman, D. (2003). Combining low-level and summary representations of opinions for multi-perspective question answering. In AAAI spring symposium on new directions in question answering (pp. 20–27). Chaovalit, P., & Zhou, L. (2005). Movie review mining: A comparison between supervised and unsupervised classification approaches. In Proceedings of the 38th Hawaii international conference on system sciences (pp.1–9). Big Island, Hawaii: IEEE. Charlotta, E. (2004). Topic dependence in sentiment classification. Master’s thesis, University of Cambridge. Choi, Y., Cardie, C., Riloff, E., & Patwardhan, S. (2005). Identifying sources of opinions with conditional random fields and extraction patterns. In Proceedings of HLT/ EMNLP. Choi, Y., Breck, E., & Cardie, C. (2006). Joint extraction of entities and relations for opinion recognition. In Proceedings of the 2006 conference on empirical methods in natural language processing (EMNLP 2006), Sydney (pp. 431–439). Clarke, C. L. A., & Terra, E. L. (2003). Passage retrieval vs. document retrieval for factoid question answering. In Proceedings of the 26th annual international ACM SIGIR conference on research and development in information retrieval, Toronto, Canada (pp. 427–428). Curran, J. R. (2002). Ensemble methods for automatic thesaurus extraction. In Proceedings of the 2002 conference on empirical methods in natural language processing, Philadelphia, PA, USA (pp. 222–229). Dagan, I., Shaul, M., & Markovitch, S. (1993). Contextual word similarity and estimation from sparse data. In Proceedings of ACL-93, Columbus, Ohio (pp. 164– 171). Dagan, I., Pereira, F., & Lee, L. (1994). Similarity-based estimation of word co- occurrence probabilities. In Proceedings of the 32nd annual meeting of the ACL, Las Cruces, NM (pp. 272–278). Daille, B. (1996). Study and implementation of combined techniques for automatic extraction of terminology. The balancing act: Combining symbolic and statistical approaches to language. Cambridge: MIT Press. Das, S. R., & Chen, M. Y. (2001). Yahoo! for Amazon: Sentiment parsing from small talk on the web. In Proceedings of the 8th Asia Pacific finance association annual conference. Dave, K., Lawrence, S., & Pennock, D. M. (2003). Mining the peanut gallery: Opinion extraction and semantic classification of product reviews. In WWW2003, Budapest, Hungary. Devitt, A., & Vogel, C. (2004). The topology of WordNet: Some metrics. In Proceedings of GWC-04, 2nd global WordNet conference, Brno, CZ (pp. 106–111). Didier, B. (1995). Lexter: A terminology extraction software for knowledge acquisition from texts. In Proceedings 9th Banff knowledge acquisition for knowledge-based systems workshop, Banff (Vol. 5, pp. 1–17). Dimitrova, M., Finn, A., Kushmerick, N., & Smyth, B. (2002). Web genre visualisation. In Conference on human factors in computing systems. Durbin, S. D., Neal Richter, J., & Warner, D. (2003). A system for affective rating of texts. In Proceedings of OTC-03, 3rd workshop on operational text classification, Washington, US. Edmonds, P. (1999). Semantic representations of near-synonyms for automatic lexical choice. Ph.D. thesis, University of Toronto. Edmonds, P., & Hirst, G. (2002). Near-synonymy and lexical choice. Computational Linguistics, 28, 105–144. Efron, M. (2004). Cultural orientation: Classifying subjective documents by cociation analysis. In Proceedings of the AAAI fall symposium on style and meaning in language (pp. 41–48). (Art, music, and design). Esuli, A., & Sebastiani, F. (2005). Determining the semantic orientation of terms through gloss classification. In Proceedings of CIKM-05, the ACM SIGIR conference on information and knowledge management, Bremen, DE. Esuli, A., & Sebastiani, F. (2006). SentiWordNet: A publicly available lexical resource for opinion mining. In Proceedings of LREC 2006, 5th conference on language resources and evaluation, Genova. Esuli, A., & Sebastiani, F. (2006). Determining term subjectivity and term orientation for opinion mining. In Proceedings of EACL-06, 11th conference of the european chapter of the association for computational linguistics, Trento. Etzioni, O., Cafarella, M., Downey, D., Kok, S., Popescu, A.-M., Shaked, T., et al. (2004). Web-scale information extraction in KnowItAll (preliminary results). In WWW’2004. Fei, Z., Liu, J., & Wu, G. (2004). Sentiment classification using phrase patterns. In Proceedings of the fourth international conference on computer and information technology (CIT’04). Fellbaum, C. (Ed.). (1998). WordNet: An electronic lexical database. Cambridge, MA: MIT Press. Finn, A., Kushmerick, N., & Smyth, B. (2002). Genre classification and domain transfer for information filtering. In F. Crestani, M. Girolami, C. van Rijsbergen, & J. Cornelis (Eds.), Proceedings of ECIR-02, 24th European colloquium on information retrieval research, Glasgow, UK. Heidelberg, DE: Springer- Verlag. Freitag, D., & McCallum, A. (2000). Information extraction with HMM structures learned by stochastic optimization. In AAAI’00. 10772 H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 Gamon, M. (2004). Sentiment classification on customer feedback data: Noisy data, large feature vectors, and the role of linguistic analysis. In Proceedings the 20th international conference on computational linguistics (pp. 841–847). Gamon, M., & Aue, A. (2005). Automatic identification of sentiment vocabulary exploiting low association with known sentiment terms. In Proceedings of the ACL workshop on feature engineering for machine learning in NLP, Ann Arbor (pp. 57–64). Glance, N. S., Hurst, M. & Tomokiyo, T. (2004). Blogpulse: Automated trend discovery for weblogs. In WWW 2004 workshop on the weblogging ecosystem: Aggregation, analysis and dynamics. Grefenstette, Gregory (1994). Explorations in automatic thesaurus discovery. Boston, MA: Kluwer Academic Press. Grefenstette, G., Qu, Y., Shanahan, J. G., & Evans, D. A. (2004). Coupling niche browsers and affect analysis for an opinion mining application. In Proceedings of RIAO-04, 7th international conference on recherche d’information assist’ee par ordinateur, Avignon, FR (pp. 186–194). Hatzivassiloglou, V., & McKeown, K. R. (1997). Predicting the semantic orientation of adjectives. In Proceedings of the 35th annual meeting of ACL. Hatzivassiloglou, V., & Wiebe, J. (2000). Effects of adjective orientation and gradability on sentence subjectivity. In Proceedings of the 18th international conference on computational linguistics. Hayakawa, S. I. (1994). Choose the right word (2nd ed.). Harper-Collins Publishers. (revised by Eugene Ehrlich). Hillard, D., Ostendorf, M., & Shriberg, E. (2003). Detection of agreement vs. disagreement in meetings: Training with unlabeled data. In Proceedings of HLTNAACL. Hu, M., & Liu, B. (2004). Mining and summarizing customer reviews. In KDD’04, Seattle, Washington, USA. Hu, M., & Liu, B. (2004). Mining opinion features in customer reviews. In AAAI 2004 (pp. 755–760). Inkpen, D. Z., Feiguina, O., & Hirst, G. (2004). Generating more-positive and more- negative text. In J. G. Shanahan, Y. Qu, & J. Wiebe (Eds.), Computing attitude and affect in text: Theory and applications. The information retrieval series (Vol. 20, pp. 187–196). Dordrecht, The Netherlands: Springer. Jacquemin, C., & Bourigault, D. (2001). Term extraction and automatic indexing. In R. Mitkov (Ed.), Handbook of computational linguistics. Oxford University Press. Jon, K., & Tardos, E. (2002). Approximation algorithms for classification problems with pairwise relationships: Metric labeling and Markov random fields. Journal of the ACM, 49(5), 616–639. Justeson, J., & Slava, K. (1993). Technical terminology: Some linguistic properties and an algorithm for identification in text. Natural Language Engineering, 1(1), 9–27. Kamps, J. & Marx, M. (2002). Words with attitude. In Proceedings of the first international conference on global WordNet, CIIL, Mysore, India (pp. 332–341). Kamps, J., Marx, M., Mokken, R. J., & de Rijke, M. (2004). Using WordNet to measure semantic orientation of adjectives. In Proceedings of LREC-04, 4th international conference on language resources and evaluation, Lisbon, PT (Vol. IV, pp. 1115– 1118). Kennedy, A., & Inkpen, D. (2006). Sentiment classification of movie reviews using contextual valence shifters. Computational Intelligence, 22(2), 110–125. Kessler, B., Nunberg, G., & SchÄutze, H. (1997). Automatic detection of text genre. In Proceedings of the 35th ACL/8th EACL (pp. 32–38). Kim, S.-M., & Hovy, E. (2005). Automatic detection of opinion bearing words and sentences. In Companion volume of the proceedings of IJCNLP-05, Jeju Island, Republic of Korea. Kim, S.-M., Hovy, E. (2006). Identifying and analyzing judgment opinions. In Proceedings of HLT/NAACL-2006, New York City. Kim, S.-M., Hovy, E. (2006). Automatic identification of pro and con reasons in online reviews. In Proceedings of the COLING/ACL 2006 main conference poster sessions, Sydney (pp. 483–490). Kim, S.-M., & Hovy, E. (2004). Determining the sentiment of opinions. Coling, 1367–1373. Kobayashi, N., Inui, T., & Inui, K. (2001). Dictionary-based acquisition of the lexical knowledge for p/n analysis (in Japanese). In Proceedings of Japanese society for artificial intelligence, SLUD-33 (pp. 45–50). Koppel, M., & Schler, J. (2005). The importance of neutral examples for learning sentiment. In Workshop on the analysis of informal and formal information exchange during negotiations (FINEXIN). Koppel, M., Schler, J. (2005). Using neutral examples for learning polarity. In Proceedings of IJCAI (poster). Ku, L.-W., Lee, L.-Y., Wu, T.-H., & Chen, H.-H. (2005). Major topic detection and its application to opinion summarization. In SIGIR 2005 (pp. 627–628). Ku, L.-W., Wu, T.-H., Lee, L.-Y., & Chen, H.-H. (2005). Construction of an evaluation corpus for opinion extraction. In Proceedings of NTCIR-5 workshop meeting, Tokyo, Japan. Ku, L.-W., Liang, Y.-T., & Chen, H.-H. (2006). Opinion extraction, summarization and tracking in news and blog corpora. In AAAI-CAAW’06. Lafferty, J., McCallum, A., & Pereira, F. (2001). Conditional random fields: Probabilistic models for segmenting and labeling or sequence data. In ICML’01. Leacock, C., & Chodorow, M. (1998). Combining local context and WordNet similarity for word sense identification. WordNet: An electronic lexical database (pp. 265–283). MIT Press: Cambridge MA. Leshed, G., & Kaye, J. (2006). Understanding how bloggers feel: Recognizing affect in blog posts. In Extended Abstracts of CHI 2006 (pp. 1019–1024). Li, H., & Yamanishi, K. (2001). Mining from open answers in questionnaire data. In Proceedings of the 7th ACM SIGKDD conference. Lin, D. (1998). Automatic retrieval and clustering of similar words. In Proceedings of COLING-ACL (pp. 768–774). Lin, W.-H., Wilson, T., Wiebe, J., & Hauptmann, A. (2006). Which side are you on? Identifying perspectives at the document and sentence levels. In Proceedings of the 10th conference on computational natural language learning (CoNLL-X), New York City (pp. 109–116). Liu, B., Hsu, W., & Ma, Y. (1998). Integrating classification and association rule mining. In KDD’98. Liu, H., Lieberman, H., & Selker, T. (2003). A model of textual affect sensing using real-world knowledge. In IUI’03: Proceedings of the 8th international conference on intelligent user interfaces (pp. 125–132). New York, NY, USA: ACM Press. Liu, B., Hu, M., & Cheng, J. (2005). Opinion observer: Analyzing and comparing opinions on the web. In Proceedings of the 14th international world wide web conference (WWW-2005) (pp. 10–14). Chiba, Japan: ACM Press. Losee, R. M. (2001). Natural language processing in support of decisionmaking: Phrases and part-of-speech tagging. Information Processing and Management, 37(6), 769–787. Luca, D., & Mazzini, G. (2002). Opinion classification through information extraction. In International conference on data mining methods and databases for engineering, finance and other fields (pp. 299–310). Matsumoto, S., Takamura, H., & Okumura, M. (2005). Sentiment classification using word sub-sequences and dependency sub-trees. In PAKDD 2005 (pp. 301–311). Springer-Verlag: Berlin, Heidelberg. Mei, Q., Ling, X., Wondra, M., Su, H., & Zhai, C. X. (2007). Topic sentiment mixture: Modeling facets and opinions in weblogs. In WWW 2007, Banff, Alberta, Canada. Miller, G. A., & Charles, W. G. (1991). Contextual correlates of semantic similarity. Language and Cognitive Processes, 6, 1–28. Mishne, G., & de Rijke, M. (2006). Capturing global mood levels using blog posts. In AAAI 2006 spring symposium on computational approaches to analysing weblogs (AAAICAAW 2006). Morinaga, S., Yamanishi, K., Tateishi, K., & Fukushima, T. (2002). Mining product reputations on the web. In ACM SIGKDD 2002 (pp. 341–349). Mullen, T., & Collier, N. (2004). Sentiment analysis using support vector machines with diverse information sources. In Proceedings of EMNLP-2004, Barcelona, Spain (pp. 412–418). Mullen, T., & Malouf, R. (2006). A preliminary investigation into sentiment analysis of informal political discourse. In Proceedings of the AAAI symposium on computational approaches to analyzing weblogs (pp. 159–162). Nasukawa, T., & Yi, J. (2003). Sentiment analysis: Capturing favorability using natural language processing. In Second international conference on knowledge capture, Florida, USA. Nigam, K., McCallum, A., & Thrun, S. (2000). Text classification from labeled and unlabeled documents using EM. Machine Learning, 39(2/3), 103–134. Pang, Bo., & Lee, L. (2004). A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings 42nd ACL, Barcelona, Spain (pp. 271–278). Pang, Bo., Lee, & L., Vaithyanathan, S. (2002). Thumbs up? Sentiment classification using machine learning techniques. In Proceedings of the 2002 conference on empirical methods in natural language processing (EMNLP) (pp. 79–86). Pereira, F. C. N., Tishby, N., & Lee, L. (1993). Distributional clustering of English words. In Meeting of the association for computational linguistics (pp. 183– 190). Philip, B., Hastie, T., Christopher M., & Shivakumar V. (2004). Exploring sentiment summarization. In AAAI spring symposium on exploring attitude and affect in text: Theories and applications (AAAI tech report SS-04-07). Rapp, R. (2004). A freely available automatically generated thesaurus of related words. In Proceedings of 4th international conference on language resources and evaluation (LREC 2004), Lisbon, Portugal. Ratnaparkhi, A. (1996). A maximum entropy model for part-of-speech tagging. In Proceedings of the EMNLP conference (pp. 133–142). Rauber, A., & Muller-Kogler, A. (2001). Integrating automatic genre analysis into digital libraries. In First ACM-IEEE joint conference on digital libraries. Read, J. (2005). Using emoticons to reduce dependency in machine learning techniques for sentiment classification. In ACL student research workshop (pp. 43–48). Ann Arbor, MI: Association for Computational Linguistics. Resnik, P. (1995). Using information content to evaluate semantic similarity in a taxonomy. In Proceedings of the 14th international joint conference on artificial intelligence. Morgan Kaufmann. Riloff, E. (1996). An empirical study of automated dictionary construction for information extraction in three domains. In Artificial intelligence (Vol. 85). Riloff, E., & Jones, R. (1999). Learning dictionaries for information extraction by multi-level bootstrapping. In Proceedings of the 16th national conference on artificial intelligence. Riloff, E., & Shepherd, J. (1997). A corpus-based approach for building semantic lexicons. In Proceedings of the second conference on empirical methods in natural language processing (pp. 117–124). Riloff, E., & Wiebe, J. (2003). Learning extraction patterns for subjective expressions. In Proceedings of the 2003 conference on EMNLP (pp. 105–112). Riloff, E., Wiebe, J., & Wilson, T. (2003). Learning subjective nouns using extraction pattern bootstrapping. In Conference on natural language learning (CoNLL), Edmonton (pp. 25–32). Roark, B., & Charniak, E. (1998). Noun-phrase co-occurrence statistics for semi- automatic semantic lexicon construction. In Proceedings of the 36th annual meeting of the association for computational linguistics (pp. 1110–1116). Rosario, B., & Hearst, M. A. (2004). Classifying semantic relations in bioscience text. In ACL’2004. H. Tang et al. / Expert Systems with Applications 36 (2009) 10760–10773 10773 Salvetti, F., Lewis, S., & Reichenbach, C. (2004). Automatic opinion polarity classification of movie reviews. Colorado research in linguistics (Vol. 17, no. 1), Boulder: University of Colorado. Schmid, H. (1994). Probabilistic part-of-speech tagging using decision trees. In International conference on new methods in language processing, Manchester, UK. Spertus, E. (1997). Smokey: Automatic recognition of hostile messages. In Proceedings IAAI. Stone, P. J., Dunphy, D. C., Smith, M. S., & Ogilvie, D. M. (1966). The general inquirer: A computer approach to content analysis. Cambridge, MA: MIT Press. Stoyanov, V., Cardie, C.., Litman, D., & Wiebe, J. (2004). Evaluating an opinion annotation scheme using a new multi-perspective question and answer corpus. In G. James, Y. Q. Shanahan, J. Wiebe (Eds.), Computing attitude and affect in text: Theory and applications. The information retrieval series, (Vol. 20, pp. 77–89). Dordrecht, The Netherlands: Springer. Subasic, P., & Huettner, A. (2001). Affect analysis of text using fuzzy typing, fuzzy systems. IEEE Transactions, 9, 483–496. Taboada, M., Caroline A., & Kimberly V. (2006). Creating semantic orientation dictionaries. In Proceedings of 5th international conference on language resources and evaluation (LREC), Genoa, Italy. Taboada, M., Gillies, M. A., & McFetridge, P. (2006). Sentiment classification techniques for tracking literary reputation. In Proceedings of LREC 2006 workshop ‘‘Towards Computational Models of Literary Analysis” (pp. 36–43). Takamura, H., Inui, T., & Okumura, M. (2005). Extracting semantic orientations of words using spin model. In Proceedings of the 43rd annual meeting of the ACL, Ann Arbor (pp. 133–140). Tan, S., Wu, G., Tang, H., & Cheng, X. (2007). A novel scheme for domain-transfer problem in the context of sentiment analysis. In Proceedings of CIKM’07, Lisboa, Portugal. Terveen, L., Hill, W., Amento, B., McDonald, D., & Creter, J. (1997). PHOAKS: A system for sharing recommendations. Communications of the ACM, 40(3), 59–62. Thelen, M., & Riloff, E. (2002). A bootstrapping method for learning semantic lexicons using extraction pattern contexts. In Proceedings of the 2002 conference on empirical methods in natural language processing. Thomas, M., Pang, Bo., & Lee, L. (2006). Get out the vote: Determining support or opposition from congressional floor-debate transcripts. In Proceedings of the 2006 conference on empirical methods in natural language processing (EMNLP 2006), Sydney (pp. 327–335). Tong, R. M. (2001). An operational system for detecting and tracking opinions in on- line discussion. In SIGIR 2001 workshop on operational text classification. Turney, P. D. (2002). Thumbs up or thumbs down? Semantic orientation applied to unsupervised classification of reviews. In Proceedings of the 40th annual meeting of the association for computational linguistics (ACL), Philadelphia (pp. 417–424). Turney, P. D., & Littman, M. L. (2002). Unsupervised learning of semantic orientation from a hundred-billion-word corpus. Technical Report ERB-1094. National Research Council Canada, Institute for Information Technology. Turney, P. D., & Littman, M. L. (2003). Measuring praise and criticism: Inference of semantic orientation from association. ACM Transactions on Information Systems, 21(4), 315–346. Valitutti, A., Strapparava, C., & Stock, O. (2004). Developing affective lexical resources. Psychology Journal, 2(1), 61–83. Vegnaduzzo, S. (2004). Acquisition of subjective adjectives with limited resources. In Proceedings of the AAAI spring symposium on exploring attitude and affect in text: Theories and applications, Stanford, US. Whitelaw, C., Argamon, S., & Garg, N. (2005). Using appraisal taxonomies for sentiment analysis. In Proceedings of the first computational systemic functional grammar conference, University of Sydney, Sydney, Australia. Whitelaw, C., Garg, N., & Argamon, S. (2005). Using appraisal groups for sentiment analysis. In Proceedings of CIKM-05, 14th ACM international conference on information and knowledge management, Bremen, DE (pp. 625–631). Wiebe, J. (1994). Tracking point of view in narrative. Computational Linguistics, 20(2), 233–287. Wiebe, J. (2000). Learning subjective adjectives from corpora. In AAAI/IAAI (pp. 735– 740). Wiebe, J., & Riloff, E. (2005). Creating subjective and objective sentence classifiers from unannotated texts. In Sixth international conference on intelligent text processing and computational linguistics. Wiebe, J., Bruce, R., & O’Hara, T. (1999). Development and use of a gold standard data set for subjectivity classifications. In Proceedings of the 37th annual meeting of the association for computational linguistics (ACL-99) (pp. 246–253). Wiebe, J., Wilson, T., Bruce, R., Bell, M., & Martin, M. (2002). Learning subjective language. Technical Report TR-02-100. Department of Computer Science, University of Pittsburgh, Pittsburgh, Pennsylvania. Wiebe, J., Wilson, T., & Bell, M. (2001). Identifying collocations for recognizing opinions. In Proceedings of the ACL/EACL workshop on collocation. ACL. Wiebe, J., Wilson, T., & Cardie, C. (2005). Annotating expressions of opinions and emotions in language. Language Resources and Evaluation, 1(2). Wilks, Y., & Stevenson, M. (1998). The grammar of sense: Using part-of-speech tags as a first step in semantic disambiguation. Journal of Natural Language Engineering, 4(2), 135–144. Wilson, T., Wiebe, J., & Hwa, R. (2004). Just how mad are you? Finding strong and weak opinion clauses. In Proceedings of AAAI-04, 21st conference of the american association for artificial intelligence, San Jose, US. Wilson, T., Wiebe, J., & Hoffmann, P. (2005). Recognizing contextual polarity in phrase-level sentiment analysis. In Proceedings of human language technologies conference/conference on empirical methods in natural language processing (HLT/ EMNLP 2005), Vancouver, Canada. Ye, Q., Shi, W., & Li, Y. (2006). Sentiment classification for movie reviews in Chinese by improved semantic oriented approach. In Proceedings of the 39th Hawaii international conference on system sciences. Yi, J., Nasukawa, T., Bunescu, R., & Niblack, W. (2003). Sentiment analyzer: Extracting sentiments about a given topic using natural language processing techniques. In Proceedings of the third IEEE international conference on data mining (p. 427). Yu, H., & Hatzivassiloglou, V. (2003). Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences. In Proceedings of the 2003 conference on empirical methods in natural language processing (pp. 129–136). Zadeh, L. A. (1987). PRUF-a meaning representation language for natural languages. In R. R. Yager, S. Ovchinnikov, R. M. Tong, & H. T. Nguyen (Eds.), Fuzzy sets and applications (pp. 499–568). John Wiley & Sons. Zhai, Y., & Liu B. (2005). Web data extraction based on partial tree alignment. In WWW’05. Zhang, Y., Xu, W., & Callan, J. (2002). Exact maximum likelihood estimation for word mixtures. In ICML workshop on text learning. A survey on sentiment detection of reviews Introduction Sentiment detection Subjectivity classification Sentiment classification Applications of sentiment detection Products comparison Opinion summarization Opinion reason mining Other applications Subjectivity classification Similarity approach Naive Bayes classifier Multiple Naive Bayes classifier Cut-based classifier Word sentiment classification Analysis by conjunctions between adjectives Analysis by lexical relations Analysis by glosses Analysis by both lexical relations and glosses Analysis by pointwise mutual information Analysis by general inquirer Document sentiment classification based on machine learning methods Single domain Extracting feature Training classifier Multiple domains Opinion extraction Opinion information extraction Opinion-bearing word extraction Opinion holder extraction Opinion-topic relation extraction Feature term extraction Makes (topic ? feature term, opinion) associatio Evaluation of sentiment detection Sentiment classification effectiveness Precision and recall Correlation coefficient and relative error Benchmarks for sentiment detection Benchmarks for word sentiment classification Benchmarks for document sentiment classification Conclusion Acknowledgements References 
work_omjjryebmrbd3lmzysae32arzm	----	The head has a smaller bony nucleus and more delayed in appearance than normal but to start with it is normal Raf. J. of Comp. & Math’s. , Vol. 5, No. 2, 2008 155 Design a Fuzzy Expert System for Pediatrics Diseases Diagnosis Nada N. Saleem Nada_N_S@uomosul.edu.iq Adepa Ismaeel Fatin George College of Computer Sciences and Mathematics\ University of Mosul Received on: 25/9/2007 Accepted on: 20/4/2008 ABSTRACT Fuzzy logic is a branch of artificial intelligence techniques, it deals with uncertainty in knowledge that simulates human reasoning in incomplete or fuzzy data. Fuzzy relational inference that was applied in medical diagnosis was used within the medical knowledge base system to deal with diagnostic activity, treatment recommendation and patient's administration. In this research, a medical fuzzy expert system named (PedFES) has been developed for diagnosis and decision making of general pediatrics diseases. The (PedFES) is a rule based fuzzy expert system, the results of laboratory analysis are inserted into the system. This system can define the probable diagnosis depending on these data, and later on it can pick out the most probable one for disease. Keyword: fuzzy logic, fuzzy expert system, Pediatrics diseases. تصميم نظام خبير مضبب لتشخيص امراض االطفال ندى نعمت سليم فاتن جورج اديبة اسماعيل جامعة الموصل/ كلية علوم الحاسوب والرياضيات 20/4/2008: ريخ قبول البحثات 25/9/2007: ريخ استالم البحثات الملخص إن المنطق المضبب هو فرع من تقنيات الذكاء االصطناعي, فهو يستخدم المعلومات المضببة أن العالقات المضببة تستخدم في التشخيصات والمعالجة إذوغير الكاملة في محاكاة أسلوب عمل اإلنسان. شخص المعني.الطبية اعتمادا على المعلومات المستخلصة من الفحوصات الخاصة بال Nada Nimat Saleem & Adepa Ismaeel and Fatin George 156 في هذا البحث تصميم نظام مضبب خبير يدخل في متناول المجال الطبي ويساعد في اتخاذ تم القرار فيما يتعلق بتشخيص أمراض األطفال. يعتمد تشخيص النظام المضبب الخبير لتحديد نوعية المرض المصاب به على نتائج التحاليل وتشخيص األمراض الشائعة الخاصة باألطفال. المختبرية, ويستخدم النظام في اتخاذ القرار الكلمات المفتاحية: المنطق المضبب, النظام الخبير المضبب, امراض االطفال 1. Introduction In recent years, computational intelligence has been used to solve many complex problems by developing intelligent systems, and fuzzy logic has proved to be a powerful tool for decision making systems, such as expert systems and pattern classification systems. Fuzzy set theory has already been used in some medical expert systems [1]. One challenge in medical expert systems is the problem of imprecision and uncertainty in both data and knowledge [2]. In practice, it is important to develop techniques for handling such imprecision and uncertainty to enhance the robustness and performance of medical expert systems. Fuzzy logic and fuzzy set theory [3] provide a good framework for managing uncertainty and imprecision in medicine [4], [5], [6] and have been applied to a number of area [7], [5], [8]. The successful development of a fuzzy model for a particular application domain is a complex multi-step process, in which the designer is faced with a large number of alternative implementation strategies. The principle alternatives are in the selection of: - Inference methodology - Linguistic variables and fuzzy terms - Rule set - Fuzzy operators - Membership functions The effect of a fuzzy system is defined by the set of vectors comprising the fuzzy output variables that are obtained from a set vectors representing input variables. Even the simplest modification to the fuzzy system, may alter the input-output mapping of the fuzzy model. Thus the behavior of a fuzzy system is governed by a combination of all these design choices. In a given application there is the additional process of defuzzification to map the output fuzzy sets to the real world. Design a Fuzzy Expert System for Pediatrics Diseases Diagnosis 157 2. Artificial Intelligence in Medicine Artificial Intelligence (AI) is a study to emulate human intelligence into computer technology. The potential of AI in medicine has been expressed by a number of researchers. [9] summarized the potential of AI techniques in medicine as follows: - Provide a laboratory for the examination, organization, representation and cataloging of medical knowledge. - Produces new tools to support medical decision-making, training and research. - Integrates activities in medical, computer cognitive and other sciences. - Offers a content-rich discipline for future scientific medical specialty. 3. Fuzzy Control in Medicine Fuzzy control techniques have recently been applied in various medical processes. Fuzzy control compared to classical control theory, which is a fuzzy logic approach to control, offers the following advantages [10], [11]: - It can be used in systems, which cannot be easily modeled mathematically. In particular systems with non-linear responses that are difficult to analyze may respond to a fuzzy control approach. - As rule-based approach to control, fuzzy control can be used to efficiently represent an expert's knowledge about a problem. - Continuous variables may be represented by linguistic constructs that are easier to understand, making the controller easier to implement and modify. - Fuzzy controllers may be less susceptible to system noise and parameter changes, in other words, they will be more robust. - Complex process can be controlled by relatively few logical rules permitting an easily comprehensible controller design and faster computation. In other words, fuzzy control can be best applied to production tasks, that heavily rely on human experience and intuition, and which therefore rule out the application conventional control methods. Nada Nimat Saleem & Adepa Ismaeel and Fatin George 158 4. Medical Knowledge as a Fuzzy Relation The relationship between symptoms and diagnosis by the concept of medical knowledge was introduced. "In a given pathology, we denote by S a set of symptoms, D a set of diagnoses and P a set of patients. What we call medical knowledge is a fuzzy relation, generally denoted by R, from S to D expressing of diagnoses". Zadeh's max-min compositional rule is adopted as an inference mechanism. It accepts fuzzy descriptions of the patient's symptoms and infers fuzzy descriptions of the patient's diseases by means of the fuzzy relationships. If a patient's symptom is Si then the patient's state in terms of diagnoses in a fuzzy set Dj with the following membership function [12], [13]:   DdSsdssd RS Ss D ij =  ,,),();(minmax)(  R(s,d) is the membership function of the fuzzy relation "medical knowledge". With P, a set of patients, and a fuzzy relation Q from P to S, and by "max-min" composition" we get the fuzzy relation T=Q×R with the membership function [14].   DdSsPpdsspdp RQ Ss T =  ,,,),();,(minmax),(  5. The Developed Pediatrics Fuzzy Expert System (PedFES) The design structure, application and working principles of a Pediatrics Fuzzy Expert system (PedFES) has been described for diagnosis of general pediatrics diseases. The (PedFES) is a parallel rule-firing system, in which all fireable rules are fired effectively at one time. During the process with the medical expert system (PedFES), laboratory test results are converted into fuzzy compatibility values reaching from zero to unity by consideration of the linguistic medical concepts. These fuzzified data are used to infer diagnosis with knowledge contained in a knowledgebase. Fuzzy relations were calculated for all linguistic medical concepts between test results and diagnosis by using the obtained fuzzy sets with the given set of patient data. Design a Fuzzy Expert System for Pediatrics Diseases Diagnosis 159 5.1 Medical Knowledgebase The examination and laboratory data of the system assigned for the pediatrics patients were collected from Al-Kansaa Teaching Hospital. The data then are inserted into the system (PedFES), for the design process, the Fasting Blood Sugar (FBS), Cholesterol and PH examinations are used for Diabetes Mellitus diagnosis (MELT). The Urine Culture (UC), Serum Creatiniue (S.creatinine) and General Urine Examination (GUE) are used for Urinary Tract Infection disease (UTI) diagnosis. Total Serum Bilirubin (T.S.B_direct), (T.S.B_in_direct) and Hemoglobin (Hb) examination are used for diagnosis of Jaundice disease (JD). The units of the used factors are: FBS (mg/100ml), Cholesterol (mg/100ml), S.creatinine (mg/100ml), T.S.B_direct and T.S.B_in_direct (mg/100ml), Hb (gm/100l). 5.2 Inference Methodology The fuzzy control application structure as shown in figure (1) has nine crisp input variables, these input parameters represent the patients data. The mamdani model of inference was used. All (PedFES) fuzzy rules were of the form "If x1 is A1 and y1 is B1 then z1 is C1", where A1, B1 and C1 are fuzzy sets. The max-min operators were used for implication throughout for implementing a (PedFES). It was necessary to obtain crisp output for the purposes of evaluation of the fuzzy model, center-of-gravity (centroid) defuzzification was used to produce crisp values on an arbitrary scale of the fuzzy output variable Pediatrics Patients Disease (PPD). 5.3 Linguistic Variables and Fuzzy Terms Each of the nine input parameters was assigned a linguistic variable and examination of the data and rules showed that each could naturally be divided into two or three fuzzy terms corresponding to meanings of Low (L), Normal (N) and High (H) for the types of laboratory examinations. One output fuzzy variable was used, from the rules it was determined that the output fuzzy variable PPD had eight linguistic terms in it's term_set, these terms represent the pediatric diseases. Nada Nimat Saleem & Adepa Ismaeel and Fatin George 160 Figure(1): The structure of the pediatrics fuzzy expert system 5.4 Rule Set PedFES must contain a set of rules that can deal with fuzzy tags, fuzzy sets and relations and must provide an appropriate output which corresponds to a particular input. The rules for the fuzzy expert system were obtained by means of knowledge expert-doctor, and had been carefully refined to form a complete and consistent set of classifiers. Part of the developed (PedFES) fuzzy knowledge base rules are shown in the table (1), total of 84 rules are formed as shown in the appendix. Cholesterol FBS PH UC S.creatinine GUE T.S.B-direct T.S.B- indirect Hb PP D Design a Fuzzy Expert System for Pediatrics Diseases Diagnosis 161 Table (1): (PedFES) fuzzy rules Rule No. FBS PH Cholestrol GUE UC S.creatin TSB_ Direct TSB in_direct HB PPD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rule 2 N N N N N N H N L JD Rule 3 N N N H H H N N N UTI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rule 5 N N N H H N N N N UTI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rule 25 H L H N N N H N L MELT & JD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rule 40 H L H N H H N N N UTI& MELT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . For example Rule 25 can be interpreted as follows: Rule 25: if patient’s FBS is High, patient’s PH is Low, patient’s Cholestrol is High and patient’s GUE is Normal, patient’s UC is Normal, patient’s S.crartin is Normal and patient’s TSB_Direct is High, patient’s TSB_indirect is Normal and patient’s HB is Low, then patient has Diabetes Mellitus and Jaundice diseases. 5.5 Fuzzy Operations The probabilistic family of operators was chosen for the reason that all the rules features conjunction of the input linguistic variables use the fuzzy conjunction operators. The min operator is used for conjunction, the overall truth value of such a rule will obviously be determined solely by the reasoning of the clinician in considering all the parameters. After the truth degree of each rule is determined, we calculate the truth degree of all rules by taking max between working rules. 5.6 Membership functions Fuzzy membership functions for the medical factors for the mentioned laboratory parameters are calculated by the following functions as shown in the figure (2). Nada Nimat Saleem & Adepa Ismaeel and Fatin George 162 Figure(2): The membership functions of input-output terms 6. Experimental Results For the output factor PPD, the linguistic expressions represent the patients disease. The truth degree of the rules are determined for each rule by aid of the min and then by taking max between working rules. For example, for the input patient laboratory tests values FBS=80, PH=7.5, Cholestrol=100, Design a Fuzzy Expert System for Pediatrics Diseases Diagnosis 163 100= sampleofnumbertotal diagnosisofnumbercorrecttotal eperformanc T.S.B_indirect=0, T.S.B_Direct=0, Hb=15, GUE=8, UC=200 and S.Crartin=1.2. the rules (3) and (5) will be fired as shown in figure (3). From Mamdani max-min inference we will obtain the membership function of our system, max (rule 3, rule 5)=rule 5, then we calculate the crisp output. The crisp value of the PPD is calculated by the method of center of gravity defuzzifier. As you see from figure (3), the value of PPD=35, this means that the patient has the Urinary Tract Infection (UTI) disease (see figure 2). Another example for implementation of the (PedFES) system for patients disease diagnosis shown at figures (4 and 5) with patients laboratory tests values, fired rules (12) and (13) and values of output term PPD that refer to the diagnosis of the patient complaint. 7. PedFES Performance Measurement The (PedFES) system performance can be evaluated by the equation: The performance of the model is tested for 60 patients, the results show that the correct diagnosis using (PedFES) system is 50 and this achieves performance of (83.3٪) . Nada Nimat Saleem & Adepa Ismaeel and Fatin George 164 . . . . . . . . . . . . Figure(3): Calculation of the value PPD represents the diagnosis of a patient affected by Urinary Tract Infection depending on the following laboratory test values (FBS=80, PH=7.5, Cholestrol=100, T.S.B.in.Direct=0, T.S.B.Direct=0, Hb=15, GUE=8, UC=200, S.Crartin=1.2). Design a Fuzzy Expert System for Pediatrics Diseases Diagnosis 165 . . . . . . . . . . . . Figure(4): Calculation of the value PPD represents the diagnosis of a patient affected by Diabetes Mellitus depending on the following laboratory test values (FBS=50, PH=7.5, Cholestrol=300, T.S.B.in.Direct=0.5, T.S.B.Direct=0, Hb=12, GUE=4, UC=75, S.Crartin=0.3). Nada Nimat Saleem & Adepa Ismaeel and Fatin George 166 . . . . . . . . . . . . Figure(5): Calculation of the value PPD represents the diagnosis of a patient affected by Urinary Tract Infection with Jaundice depending on the following laboratory test values (FBS=100, PH=7.3, Cholestrol=150, T.S.B.in.Direct=3, T.S.B.Direct=0, Hb=8, GUE=8, UC=800, S.Crartin=5). Design a Fuzzy Expert System for Pediatrics Diseases Diagnosis 167 8. Conclusions The developed diagnosis module (PedFES) consists of expert system and fuzzy logic techniques to perform diagnostic tasks. A set of rules will be defined using the patients disease database as well as the expert knowledge on the disease domain (from doctors). The designed expert system uses the rules to diagnose patient's illness base on their laboratory tests. In addition, fuzzy logic is integrated to enhance the reasoning when dealing with fuzzy data. The combination of expert system and fuzzy logic that forms a hybrid (expert-fuzzy) system could increase the system performance and has been implemented successfully. Nada Nimat Saleem & Adepa Ismaeel and Fatin George 168 REFERENCES [1] Nguyen Hoang Phuong and Vladik Kreinovich, 2000, "Fuzzy logic and it's applications in medicine", Institute of Information Technology, National Center for Natural Science and Technology of Vietnam. [2] Hughes, C. "The representation of uncertainty in medical expert systems", Medical Informatics, vol. 14, pp. 269-279, 1989. [3] Zadeh,L.A. "The role of fuzzy logic in the management of uncertainty in expert systems", Fuzzy Set Systems, vol. 11, pp. 199- 227,1983. [4] Adlassnig, K.P. "Fuzzy set theory in medical diagnosis", IEEE Transactions Systems Man Cybernetics, vol. 16, no. 2, pp. 260-265, 1986. [5] Cohen M.E. and D.L.Hudson. "The use of fuzzy variables in medical decision making", in Fuzzy Computing, M.M.Gupta and T.Yamakawa, Eds., pp. 263-271. Elsevier Science, North Holland, 1988. [6] Halim, M. K.M.Ho, and A.Liu, "Fuzzy logic for medical expert systems", Annals Academy Medicine, vol. 19, no. 5, pp. 672-683, 1990. [7] Adlassing, K.P. and G.Kolarz, "CADIAC-2: Computer-assisted medical diagnosis using fuzzy subsets", in Approximate Reasoning in Decision Analysis, MM Gupta and E Sanchez, Eds., pp. 219-247, North-Holland, New York, 1982. [8] Watanabe, H. W.J.Yakowenko, Y.M.Kim, J.Anbe, and T.Tobi, "Application of a fuzzy discrimination analysis for diagnosis of valvular heart disease", IEEE Transactions Fuzzy Systems, vol. 2, no. 4, pp. 267-276, 1994. [9] Hoong N.K., "Medical Information Science-Framework and Potential", International Seminar and Exhibition Computerization for Development- the Research Challenge, Universiti Pertanian Malaysia: Kuala Lumpur, pp. 191-198, 1998. Design a Fuzzy Expert System for Pediatrics Diseases Diagnosis 169 [10] Mamdani, E. H., Qstergaard J.J., Lembessis E., "Use of Fuzzy Logic for Implementing Rule-Based Control of Industrial Processes", In: P.P.Wang, Advances in Fuzzy Sets, Possibility Theory, and Applications, New York: Plenum Press, 1989, pp. 307-323. [11] Klement, E.P., Slany W., "Fuzzy Logic in Artificial Intelligence", Proceedings of the 8th Austrian Artificial Intelligence Conference FLAI '93, Berlin 1995. [12] Sanchez, E., "Medical Diagnosis and Composite Fuzzy Relations", Advances in Fuzzy Set Theory and Applications, Amsterdam: North-Holland 1979, pp. 437-444. [13] Sanchez, E, "Resolution of Composite Fuzzy Relation Equations", Information and Control, 30, 1976, pp. 38-48. [14] William Siler and James J. Buckley, "Fuzzy Expert Systems and Fuzzy Reasoning", Wiley Interscience, 2005. Nada Nimat Saleem & Adepa Ismaeel and Fatin George 170 Appendix: Rule_base Knowledge for (PedFES) System Rule No. FBS PH Cholestrol GUE UC S.creatin TSB_ Direct TSB in_direct HB PPD 1 N N N N N N N N N Normal 2 N N N N N N H N L JD 3 N N N H H H N N N UTI 4 N N N N H H N N N UTI 5 N N N H H N N N N UTI 6 N N N N N N N H N JD 7 H N N N N N N N N MELT 8 H N H N N N N N N MELT 9 H L H N N N N N N MELT 10 H L N N N N N N N MELT 11 L N N N N N N N N MELT 12 L N H N N N N N N MELT 13 N N N H H H H N L UTI& JD 14 N N N N H H H N L UTI& JD 15 N N N H H N H N L UTI& JD 16 N N N N H H N H N UTI& JD 17 N N N H H N N H N UTI& JD 18 N N N H H H N H N UTI& JD 19 H N N N N N H N L MELT& JD 20 H N N N N N N H N MELT& JD 21 H L N N N N H N L MELT& JD 22 H L N N N N N H N MELT& JD 23 H N H N N N H N L MELT& JD 24 H N H N N N N H N MELT& JD 25 H L H N N N H N L MELT& JD 26 H L H N N N N H N MELT& JD 27 L N N N N N H N L MELT& JD 28 L N N N N N N H N MELT& JD 29 L N H N N N N H N MELT& JD 30 L N H N N N H N L MELT& JD 31 H N N N H H N N N UTI& MELT 32 H N N H H N N N N UTI& MELT 33 H N N H H H N N N UTI& MELT 34 H L N N H H N N N UTI& MELT 35 H L N H H N N N N UTI& MELT 36 H N H H H H N N N UTI& MELT 37 H N H N H H N N N UTI& MELT 38 H N H H H N N N N UTI& MELT 39 H L H H H H N N N UTI& MELT 40 H L H N H H N N N UTI& MELT 41 H L H H H N N N N UTI& MELT Design a Fuzzy Expert System for Pediatrics Diseases Diagnosis 171 42 L N N H H H N N N UTI& MELT 43 L N N N H H N N N UTI& MELT 44 L N N H H N N N N UTI& MELT 45 L N H H H H N N N UTI& MELT 46 L N H N H H N N N UTI& MELT 47 L N H H H N N N N UTI& MELT 48 H L N H H H N N N UTI& MELT 49 H N N H H H H N L UTI, JD& MELT 50 H N N N H H H N L UTI, JD& MELT 51 H N N H H N H N L UTI, JD& MELT 52 H N N H H H N H N UTI, JD& MELT 53 H N N N H H N H N UTI, JD& MELT 54 H N N H H N N H N UTI, JD& MELT 55 H N H H H H H N L UTI, JD& MELT 56 H N H N H H H N L UTI, JD& MELT 57 H N H H H N H N L UTI, JD& MELT 58 H N H H H H N H N UTI, JD& MELT 59 H N H N H H N H N UTI, JD& MELT 60 H N H H H N N H N UTI, JD& MELT 61 H L H H H H H N L UTI, JD& MELT 62 H L H N H H H N L UTI, JD& MELT 63 H L H H H N H N L UTI, JD& MELT 64 H L H H H H N H N UTI, JD& MELT 65 H L H N H H N H N UTI, JD& MELT 66 H L H H H N N H N UTI, JD& MELT 67 H L N H H H H N L UTI, JD& MELT 68 H L N N H H H N L UTI, JD& MELT Nada Nimat Saleem & Adepa Ismaeel and Fatin George 172 69 H L N H H N H N L UTI, JD& MELT 70 H L N H H H N H N UTI, JD& MELT 71 H L N N H H N H N UTI, JD& MELT 72 H L N H H N N H N UTI, JD& MELT 73 L N N H H H H N L UTI, JD& MELT 74 L N N N H H H N L UTI, JD& MELT 75 L N N H H N H N L UTI, JD& MELT 76 L N N H H H N H N UTI, JD& MELT 77 L N N N H H N H N UTI, JD& MELT 78 L N N H H N N H N UTI, JD& MELT 79 L N H H H H H N L UTI, JD& MELT 80 L N H N H H H N L UTI, JD& MELT 81 L N H H H N H N L UTI, JD& MELT 82 L N H H H H N H N UTI, JD& MELT 83 L N H N H H N H N UTI, JD& MELT 84 L N H H H N N H N UTI, JD& MELT 
work_omkrjeappzda3e2pamua6hh7pm	----	The development of an intelligent tutoring system on design of liquid retaining structures 1 Expert Systems, Vol. 21, No. 4, 2004, pp. 183-191 AN EXPERT SYSTEM ON DESIGN OF LIQUID RETAINING STRUCTURES WITH BLACKBOARD ARCHITECTURE K.W. Chau Department of Civil & Structural Engineering, Hong Kong Polytechnic University, Hunghom, Kowloon, Hong Kong F. Albermani Department of Civil Engineering, University of Queensland, Australia ABSTRACT: The design of liquid retaining structures involves many decisions to be made by the designer based on rules of thumb, heuristics, judgment, code of practice and previous experience. Structural design problems are often ill structured and there is a need to develop programming environments that can incorporate engineering judgment along with algorithmic tools. Recent developments in artificial intelligence have made it possible to develop an expert system that can provide expert advice to the user in selection of design criteria and design parameters. This paper introduces the development of an expert system in design of liquid retaining structures using blackboard architecture. An expert system shell, Visual Rule Studio, is employed to facilitate the development of this prototype system. It is a coupled system combining symbolic processing with the traditional numerical processing. The expert system developed is based on British Standards Code of Practice BS8007. Explanations are made to assist inexperienced designers or civil engineering students to learn how to design liquid retaining structures effectively and sustainably in their design practices. The use of this expert system in disseminating heuristic knowledge and experience to practitioners and engineering students is discussed. INTRODUCTION Design of liquid retaining structures involves many decisions to be made by the designer based on rules of thumb, heuristics, judgment, code of practice and previous experience. Various design parameters to be chosen include configuration, material, loading, etc. A novice engineer may face many difficulties in the design process. Advances in computer hardware, computer software and engineering methodologies in the recent decades have led to an increased use of computers by engineers. Until recently, the main use of computers by engineers was mainly confined to the number crunching of large volumes of numerical data. In the realm of structural design, this use has been limited almost exclusively to algorithmic solutions (Chau & Lee 1991). Structural design problems are often ill structured. Hence, there is a need to develop programming environments that can incorporate engineering judgment along with algorithmic tools. (Kitzmiller & Kowalik 1987) In the past decade, the potential of artificial intelligence (AI) techniques for providing assistance in the solution of engineering problems has been recognized. Expert systems are considered suitable for solving problems that demand considerable expertise, judgment or rules of thumb. As a result of years of research in AI, expert systems have emerged as a most promising application covering a wide range of applications (Adeli & Hawkins 1991, Chau 1992, 2004a&b, Chau & Anson 2002, Chau & Chen 2001, Chau et al. 2002, Chau & Cheung 2004, Chau & Ng 1996, Chau & Yang 1994, Chau & Zhang 1995, Kumar 1995, Lin & This is the Pre-Published Version. 2 Albermani 1998, Sriram 1987). Expert systems have developed into practical problem solving tools that can reach a level of performance comparable to that of a human expert in some specific problem domains. All these applications can be broadly classified into the following categories: diagnosis; design; data interpretation; planning; and education. Areas of early applications of expert systems technology include medical diagnosis, mineral exploration and chemical spectroscopy. Recent developments in artificial intelligence have made it possible to develop an expert system that can provide expert advice to the user in selection of design criteria and design parameters. (Dym & Levitt 1991) This paper introduces the development of an expert system in design of liquid retaining structures using blackboard architecture with hybrid knowledge representation techniques including production rule system and object-oriented approach. An expert system shell, Visual Rule Studio, is employed to facilitate the development of this prototype system. It is a coupled system in which AI-based symbolic processing is combined with the traditional numerical processing. The expert system developed is based on British Standards Code of Practice BS8007: 1987: Design of concrete structures for retaining aqueous liquids (British Standards Institution 1987). Explanations are made to assist inexperienced designers or civil engineering students to learn how to design liquid retaining structures effectively and sustainably in their design practices. The use of this intelligent tutoring system in disseminating heuristic knowledge and experience to practitioners and engineering students is discussed. KNOWLEDGE ON DESIGN OF LIQUID RETAINING STRUCTURES Liquid retaining structure is a structure which is designed and constructed to retain aqueous liquid. It is subject to lateral water pressure and earth pressure when it is located underground. Most of these structures are constructed by reinforced concrete with design life of 50 years in Hong Kong. Crack widths are to be checked to ensure impermeability of concrete and prevention from corrosion of reinforcement. Normally, two kinds of classification are used regarding liquid retaining structures, i.e. according to the shape or the location. Based on the shape, it is classified as rectangular, circular or polygon. Based on its location, it is classified as underground or above ground. Compared with circular tank structure with the same width, rectangular liquid retaining structure has larger volume. However, because of stress concentration at corners, rectangular structures will be more vulnerable to failure. It also has a weaker deflection control. With a circular tank design, not all the spaces are utilized. Since a circular structure can be constructed monolithically without any construction joints, it has better strength quality. With precise structural analysis, circular structure has a better control in deflection, crack width, bending moment resistance, axial compression resistance, and shear resistance than rectangular structure. Polygon liquid retaining structure is usually used for aesthetic purposes, such as a fountain in a garden and the retaining height is usually not very high. Its major design consideration is on crack width to ensure its impermeability and is seldom used in industrial or domestic fields. Underground liquid retaining structure usually has larger base area which cannot easily be supported by structure above the ground. In some cases, the tank is connected to underground pipe network system, in order to reduce maintenance cost, it will be constructed underground to suit the invert level of the pipe network system. The underground structure is mainly 3 subjected to lateral earth pressure or lateral water pressure which is due to the underground water table. Besides checking structural failure mode, bearing capacity of soil and settlement of structure also need to be checked. If the soil bearing capacity does not satisfy the requirement, pile foundation or raft foundation will be required. Liquid retaining structure above the ground is only subject to liquid pressure due to its own retaining liquid. The structure can either rest on ground concrete slab immediately or rest on supports which can be 3-D steel truss, mass concrete blocks, or concrete beam. There are a great variety of factors affecting the decision in selecting design criteria and design method. These factors are: dimensions, location, ground condition, support condition, groundwater conditions, aesthetic properties, design life, exposure condition, usage, roofing, availability of construction materials. Liquid retaining structures, like any other engineering structures, should not fail to satisfy any of its performance criteria. According to the Code of Practice BS8007, the two main classes of limit state, which should be considered, are ultimate limit state and serviceability limit state. Ultimate limit state is design against structural failure, including bending moment check and shear force check. Serviceability limit state is design against deflection and crack width. Normal crack width control is 0.2mm while, for severe cases, allowable crack width is 0.1mm. For underground liquid retaining structures, serviceability limit state design is used in checking of bearing capacity of soil. Factors of safety are involved in the design in order to increase the reliability that the structures will perform satisfactorily. EXPERT SYSTEM Expert systems are interactive computer programs that incorporate expertise and provide advice on a wide range of tasks. They solve a specific complex problem using reasoning processes that resemble those of human experts, when they solve the same problem. They tend to mimic the decision making and reasoning processes of human experts by providing expert advice, answering questions, and justifying their conclusions. An expert system can contain the purpose of teaching. Only explanations are not sufficient in the system. The problem should be dealt with by communication between the user and the system. In other words, the system has to engage the user in a dialogue actively and systematically. Figure 1 shows the schematic view of an expert system. These systems typically consist of the following three basic components: knowledge base, context and inference mechanism. The heart and core of any expert system is the knowledge base, which is usually a collection of rules, typically in the form of IF..THEN….. The knowledge base is a collection of general facts, rules of thumb and causal models of the behavior of the problem domain. It contains the knowledge specific to the domain of the problem to be solved. Other forms of representations commonly used are logic, frame-based schemes, nets and, more recently, the object-oriented approach. For expert system developers, rule-based system tends to be more easily understood and thus accepted. However, because of its modularity, data abstraction and inheritance characteristics, object-oriented programming will likely subsume other approaches in the very near future. The context is a workspace for the problem constructed by the inference mechanism from the information provided by the user and the knowledge base. It contains facts that reflect the current state of the problem. The organization of the context depends on the nature of the problem domain. The context builds up dynamically as a particular problem is being 4 considered. The context is used by the inference mechanism to guide the decision making process. The inference mechanism monitors the execution of the program by using the knowledge base to modify the context. It manipulates the context using the knowledge base. A number of problem solving strategies exist in current expert systems; such as forward chaining, where the system works from an initial state of known facts to a goal state. In addition to the three main modules described above, the system should also be provided with three other components that are not necessarily part of every expert system but are desirable in a final product: a graceful user interface, an explanation facility and a knowledge acquisition module. The function of the user interface module is to accept a problem description from the user and to interact with the rest of the system in order to analyze the problem or augment the capability of the system. It provides an interface between the user and the expert system, usually as a command language for directing execution. The interface is responsible for translating the input as specified by the user to the form used by the expert system and for handling the interaction between the user and the expert system during the decision making process. The explanation module provides explanations of the inferences used by the expert system. This explanation can be a priori – why a certain fact is requested, or a posteriori – how a conclusion was reached. The knowledge acquisition module serves as an interface between the experts and the expert system. It provides a means for entering domain specific knowledge into the knowledge base and revising this knowledge when necessary. COMPARISON WITH CONVENTIONAL PROGRAMMING Conventional programs consist of a set of statements whose order of execution is predetermined. These programs are very inflexible; updates need considerable effort, because the programmer has to locate the appropriate place to update in the predefined sequence. The programmer must ensure completeness, i.e., that the program performs correctly for all possible combination of conditions, and uniqueness of the solution, i.e., that the output is unique for every possible input. The user perceives the program as a blackbox, where the program outputs results for the input provided; he does not have any idea as to why the program has produced certain results. An expert system eliminates some of the impediments posed by conventional programs by making a clear distinction between the knowledge base and the control strategy. This partitioning allows for incremental addition of knowledge, without manipulating the overall program structure; the programmer need not guarantee completeness. Further, by ranking several alternatives either by an evaluation scheme or by the use of inexact inference methods, several solutions can be provided for a particular set of input conditions, thus relaxing the uniqueness constraint. The user can also question the results produced by the program through the explanation module. It is clear that an expert system offers a powerful programming environment for the development of engineering software, since engineering involves extensive use of heuristics. 5 BLACKBOARD ARCHITECTURE The blackboard architecture is intended to support development of systems in domains characterized by interaction between diverse sources of knowledge and hence provides a framework for integrating knowledge from several sources. The blackboard serves as a global data structure which facilitates this interaction. A common analogy may be made to problem- solving in domains where a number of experts in different areas of specialities co-operate over the solution which any one of them could never achieve alone. In order to facilitate this process, they agree to use a blackboard to post (or write) any partial result they can contribute separately. Each expert takes turns to write on the blackboard and, in case more experts wish to write simultaneously, the conflict is resolved by some pre-defined strategy. The blackboard architecture has been successfully used in solving a wide range of tasks, such as speech recognition, signal processing, and planning. (Engelmore & Morgan 1988) A blackboard system consists of a number of knowledge sources that communicate through a blackboard and are controlled by an inference mechanism. Figure 2 shows the architecture of a blackboard system. The main components of a typical blackboard system are entries, knowledge sources, blackboard, and inference mechanism. Entries are the immediate results produced by the system. In a typical system, each entry has a certainty factor as well as a specification. The knowledge base consists of a number of knowledge sources (KSs), which contain the knowledge. These knowledge sources are independent chunks of knowledge and do not directly communicate with each other. Instead, they participate in the problem solving process by creating entries in a global database – the blackboard. Knowledge modules look at the blackboard to see if suitable data is present to trigger their execution. If they are selected, execution results in new or altered data on the blackboard, which will then trigger other knowledge modules. Solving a problem using a blackboard architecture is based on cooperation of the knowledge modules present. Each knowledge source consists of a condition-action pair. Actions are executed whenever the conditions are satisfied in the blackboard. The blackboard or context consists of the information or entries generated by the knowledge sources during the problem solving process. It is organized into a number of levels each representing different aspects or stages of the solution process. These levels depict an a priori plan for the solution of a problem that can be naturally decomposed into a set of levels. Each level contains objects and attributes that are important to the representation of the problem. Normally, knowledge sources are specific to certain levels in the blackboard, i.e., the activation of a certain knowledge sources depends on the entries generated at certain levels in the blackboard, while the actions of the knowledge source modify entries at some other level. The main units in the blackboard are hypotheses. The hypotheses are either primary guesses about particular aspects of the problem or partial solutions. Hypotheses at various levels are related through structural relationships. The inference mechanism consists of two main components: the agenda or scheduler, and the monitor. The agenda keeps track of all the events in the blackboard and calculates the priority of execution for knowledge sources that were generated as a result of the activation of other knowledge sources. It is a list of knowledge sources or rules to be executed in the next cycle. 6 Based on the success or failure of a particular rule, new rules may get added on to it or some may be deleted from it. The basis of giving priorities to the rules on the agenda may vary from system to system. The monitor takes the element with the highest priority and executes it. Several problem-solving strategies can be implemented using the monitor. Because of its modularity, the blackboard architecture enables easy incremental development of a software system. Developers can integrate different methods of knowledge representation in a single system because of the modularity of knowledge sources. KNOWLEDGE REPRESENTATION Broadly speaking, structural analysis is described as a three-stage process involving: modeling, solution and evaluation. Before knowledge can be represented in structural analysis, the type of knowledge involved must be identified and classified. Static knowledge consists of definitions, axioms and laws. These may be a priori or the result of scientific investigation. Dynamic knowledge refers to heuristics, which is related to the process of search or to knowledge based on experience. Dynamic knowledge is not deducible from any axiom, rather it is generally gained from experience. Dynamic knowledge can also be described as ‘shallow’, meaning that it cannot provide us with explanations of why certain decisions should be made. Static knowledge is ‘deep’ in that it is deduced from fundamentals. There are several approaches for declarative representation of knowledge that are available in the AI literature, the following three, among others, being the major ones: rule-based production system, frames and object-oriented programming. A production system is a collection of rules and is believed to be good at describing heuristic knowledge. A frame system, on the other hand, is suitable for a complex and rich representation of knowledge, such as fundamental principles (categorized as static knowledge). Object-oriented programming concept is used, in which a computer program consists of a number of independent objects that process jobs by exchanging information they need via messages. To apply object-oriented program development, data representations for the model must be specified. There are three steps in the development of an object-oriented program: selection of classes, specification of the classes, implementation of the classes. Here, the word ‘class’ refers to a description of a set of similar objects; one member object in a class is called an ‘instance’. It seems a good idea to utilize both representations to solve structural design problems. USE OF EXPERT SYSTEM SHELL To facilitate the development of expert systems, expert system programming environments or shells have been developed. These system shells contain specific representation methods and inference mechanisms. The knowledge base and explanation facility of the system have been developed using a commercially available expert system shell called Visual Rule Studio which is a hybrid application development tool that integrates object-oriented techniques and expert system technology with traditional, procedural programming (RuleMachine Corporation 1998). Visual Rule Studio installs as an integral part of Microsoft Visual Basic 6.0 and appears within Visual Basic as an ActiveX Designer. As a part of the Visual Basic Integrated Development Environment, using a RuleSet in the application is similar to using a form or other Visual basic Designer. 7 By isolating rules as component objects, separate from objects and application logic, Visual Rule Studio allows developers to leverage the proven productivity of today’s component oriented development tools, such as Visual Basic. With Visual Rule Studio, rule development becomes a natural part of the component architecture development process. The complex and time consuming problems of integrating multiple development tools and managing incompatible object models no longer exist. Visual Rule Studio becomes an integrated part of the Visual Basic development and produces objects that can interact with virtually any modern development product. Visual Rule Studio objects are used to encapsulate knowledge structure, procedures, and values. An object’s structure is defined by its class and attribute declarations within a RuleSet. Object behavior is tightly bound to attributes in the form of facets, methods, rules, and demons. Figure 3 shows the structure of Visual Rule Studio components. Each attribute of a class has a specific attribute type. The Visual Rule Studio attribute types are compound, multicompound, instance reference, numeric, simple, string, interval, and time. Each attribute can have many facets associated with it. Facets provide control over how the inference engines process and use attributes. Each attribute can also have methods associated with it. Methods establish developer-defined procedures associated with each attribute. The set of backward-chaining rules that conclude the same attribute is called the attribute’s rule group. The set of forward-chaining demons that reference the same attribute in their antecedents is called the attribute’s demon group. The Visual Rule Studio inference engines control the strategies that determine how, from where, and in what order a knowledge base draws its conclusions. These inference strategies model the reasoning processes an expert uses when solving a problem. Visual Rule Studio supports these types of inferencing strategies: backward chaining, forward chaining and hybrid chaining. Each of these inferencing strategies acts on specific knowledge base components. Backward chaining inferencing starts with a specific hypothesis, or set of hypotheses, called the agenda. In backward chaining, the inference engine works backward from the agenda, pursuing a hypothesis via its search order strategies, which can be composed of the session context, the knowledge base rules, WHEN NEEDED methods, Query Objects, or the default value. Forward chaining inferencing starts with known data or conditions and determines what can be concluded from that data. Forward chaining uses a knowledge base’s demons and WHEN CHANGED methods. Hybrid inferencing is a mixed strategy that combines backward chaining and forward chaining. EXPERT SYSTEM ON DESIGN OF LIQUID RETAINING STRUCTURES The system being developed combines expert systems technologies, object-oriented programming, relational database models and hypertext/graphics in a windowing environment. It runs under and follows the conventions of Microsoft Windows. In a windowing system, any types of display windows can be represented as objects, each with its own private data or information. By defining various types of windows as different classes, such as checkbox group, hyperregion, promptbox, pushbutton, textbook, etc., they can inherit common characteristics or/and possess their own special properties. The knowledge used has been acquired mostly from written documents such as code of practice, textbooks and design manuals and complemented by experienced engineers involved with the design of liquid retaining structures. The domain knowledge is translated into procedures and methods using object-oriented representation. The system can be 8 compiled and encrypted to create a run-only system. This run-only system can be installed on a microcomputer for office use or on a portable laptop to be used in the field. The user can always overrule any design options and recommendations provided by the system. In other words, it plays the role of a knowledgeable assistant only. The input data provided by the user will be rejected if it is not within the range specified. It can explain its line of reasoning for obtaining an answer. It provides information about any individual member in multi-window graphics text display where graphic images are combined with valuable textual information. This kind of intelligent graphics is extremely valuable to structural designers because it enhances their confidence in the design provided by the expert system. The system offers a state-of-the-art user interface. The use of a mouse or other pointing device makes the data entry a simple task even for novice computer users. As such, users simply point and click their way through the process to appreciate the dynamic behavior of the system. Input data entry is kept at minimum. Input data are provided by the user mostly through selection of appropriate values of parameters from the menus and providing answers to the queries made by the system. An expert system can help the novice engineers or students to learn more knowledge of the topic through using the system. Explanations are made to assist them to learn how to design liquid retaining structures effectively and sustainably in their design practices. Implementation of the system reduces the dependence on experienced designers for routine design works and frees them to do more innovative design works. TYPICAL CONSULTATION SESSION To demonstrate how this expert system assists engineer in the preliminary design of liquid retaining structure, a sample run is demonstrated in this section. A typical example of liquid retaining structure is used to illustrate its application. An rectangular shape liquid retaining structure with a single compartment located underground, having a volume of 100 m3, a depth of 5 m and breadth/width ratio of 1.2, is designed under a severe exposure environment, i.e. the maximum design crack width is 0.2mm. The purpose of the design is to find a feasible optimum structural section in term of minimum unit cost per meter length, based on the standard slab thickness and reinforcement size as well as bar spacing. Figures 4 to 7 show some typical screen displays during the execution of the expert system. Little explanation facility is required since each screen display was designed to be user-friendly to follow. CONCLUSIONS It has been demonstrated that the hybrid knowledge representation approach combining production rule system and object-oriented programming technique to the design of water retaining structures is possible with the implementation of blackboard system architecture. It is appropriate to integrate algorithmic and symbolic programming on structural design into a single computer-aided environment running under a Windows platform. The educational spin-off of expert systems in training novice engineers or in transferring knowledge cannot be overemphasized. REFERENCES Adeli, H. and Hawkins, D.W. (1991) A hierarchical expert system for design of floors in 9 highrise buildings, Computers & Structures, 41(4): 773-778. British Standards Institution (1987) Design of Concrete Structures for Retaining Aqueous Liquids, British Standards Institution, London. Chau, K.W. (1992) An expert system for the design of gravity-type vertical seawalls, Engineering Applications of Artificial Intelligence, 5(4), 363-367. Chau, K.W. (2004a) A prototype knowledge-based system on unsteady open channel flow in water resources management, Water International, 29(1), 54-60. Chau, K.W. (2004b) Knowledge-based system on water-resources management in coastal waters, Journal of the Chartered Institution of Water and Environmental Management, 18(1), 25-28. Chau, K.W. and Anson, M. (2002) A knowledge-based system for construction site level facilities layout, Lecture Notes in Artificial Intelligence, 2358, 393-402. Chau, K.W. and Chen, W. (2001) An example of expert system on numerical modelling system in coastal processes, Advances in Engineering Software, 32(9), 695-703. Chau, K.W., Cheng, Chuntian and Li, C.W. (2002) Knowledge management system on flow and water quality modeling, Expert Systems with Applications, 22(4), 321-330. Chau, K.W. and Cheung, C.S. (2004) Knowledge representation on design of storm drainage system, Lecture Notes in Artificial Intelligence, 3029, 886-894. Chau, K.W. and Lee, S.T. (1991) Computer Aided Design Package `RCTANK' for Analysis and Design of Reinforced Concrete Tanks, Computers and Structures, 41(4): 789-799. Chau, K.W. and Ng, V. (1996) A knowledge-based expert system for design of thrust blocks for water pipelines in Hong Kong, Water Supply Research and Technology - Aqua, 45(2), 96-99. Chau, K.W. and Yang, W.W. (1994) Structuring and evaluation of VP-expert based knowledge bases, Engineering Applications of Artificial Intelligence, 7(4), 447-454. Chau, K.W. and Zhang, X.N. (1995) An expert system for flow routing in a river network, Advances in Engineering Software, 22(3), 139-146. Dym, C.L. and Levitt, R.E. (1991) Knowledge-Based Systems in Engineering, McGraw-Hill, New York. Engelmore, R. and Morgan, T. (1988) Blackboard Systems, Addison_Wesley, Wokingham. Kitzmiller, C.T. and Kowalik, J. S. (1987) Coupling Symbolic and Numeric Computing in Knowledge-Based Systems, AI Magazine, Summer: 5-90. Kumar, B. (1995) Knowledge Processing for Structural Design, Topics in Engineering Vol. 25, Computational Mechanics, Southampton. Lin, S. and Albermani, F., 2001, Lattice-dome design using A knowledge-based system approach, Computer-aided Civil and Infrastructure Engineering, 16(4), 268-286. RuleMachines Corporation (1998) Visual Rule Studio Developer’s Guide, Indialantic. Sriram, D. (1987) Knowledge-Based Approach for Structural Design, Topics in Engineering Vol. 25, Computational Mechanics, Southampton. 10 Figure 1. Schematic View of an expert system User interface Explanation facility Knowledge acquisition facility Context Inference mechanism Knowledge base User Expert 11 Figure 2. The Blackboard Architecture Agenda (Scheduler) Monitor Knowledge base Blackboard Level 1 Level 2 Level n-1 Level n KS1 KS2 KS3 KS4 KS5 KS6 Inference Mechanism 12 Figure 3. Structure of Visual Studio Components Class Name Properties Attributes Name Type Facets Methods Rules Demons Instance Attribute Value 13 Figure 4. Screen showing structural specification in preliminary design 14 Figure 5. Screen displaying imposed load specification 15 Figure 6. Screen showing generation of finite element model 16 Figure 7. Screen prompting a message for iteration on model generation 
work_onji2byofbez7gkwoqzv3macsa	----	PII: 0360-8352(90)90093-2 Computers ind. Engng Vol. 19, Nos I-4, pp. 140-144, 1990 0360-8352/90 $3.00 + 0.00 Printed in Great Britain. All rights reserved Copyright © 1990 Pergamon Press plc INTEGRATED DECISION SUPPORT AND EXPERT SYSTEMS IN A COMPUTER INTEGRATED MANUFACTURING ENVIRONMENT Chla-heo Chang Dept. of Industrial & Systems Eng. University of Michigan-Dearborn Dearborn, Michigan Jimmlng T. LIn Faculty of Adrnlnlstrstion University of Ottawa Ottawa, Ontario ABSTRACT In a computer integrated manufacturing (CIM) environment, well planning, control and operational process require both expert knowledge of the area, and powerful decision support capabilities. This paper discusses the features of decision support systems and expert systems, and their integration to support the major functions from marketing and strategic level considerations to manufacturing operational planning and process. From the hierarchical structure of information flow in a company, this paper attempts to find the best way of combining decision support systems with expert systems in enhancing the planning, control and operational functions in a CIM environment. INTRODUCTION Application of Information Technology has proven significantly improve the productivities in the manufacturing environment. More and more manufacturing functions are computerized, and results are usually encouraging. Among them we can think of CAD (computer-aided design), CAM (computer-aided manufacturing), CAPP (computer-aided process planning), CAT (computer-aided testing), etc.. Then CIM (computer integrated manufacturing), an evolved concept, represents an integration of all the relevant activities. In fact, not only engineering functions are involved, the recent development believes business functions like marketing, sales, distributions, finance and administration are also relevant to the manufacturing production and planning processes. IBM's COPICS (Communication Oriented Production Information and Control System) and MAPIC II (Manufacturing Accounting and Production Information Control System version 2) integrate information across functional areas and manage the resources of manpower, facilities and materials effectively (see [7]). The manufacturing production functions use business data to expedite planning, forecasting, scheduling and control. The trend is to develop a fully integrated system. APPLICATION OF INFORMATION TECHNOLOGY IN ClM Decision Sunoort Svstems Manufacturing planning and control functions were handled manually until the late sixties. Practitioners finally accepted information technology in the manufacturing environment. IBM brought out an integrated concept called COPICS (Communication Oriented Production Information and Control System), which significantly influenced the further development of computer-based production planning and control systems. One group of the early products was referred to as Material Requirements Planning Systems (MRP) which generated material requirements for the manufacturing production lines. Although many successful cases were brought forth, a rather large portion of the attempts to use MRP ended in dissatisfaction and sometimes even failures. After years of experience, the major causes to MRP's failure were gradually found and corrected. First of all, provided every information known, MRP could mechanically calculate material requirements very precisely, but unfortunately most manufacturing environments were constantly undergoing dynamic changes. Secondly, it would not be realistic to develop production and requirement planning without the awareness of support from manufacturing resources. That included manpower, machinery, material and capital requirements. Realizing the drawback, the new version of MRP with more decison support routines was developed under a new name called manufacturing resource planning (MRP II). Compared to MRP, MRP II puts more emphasis on matching resources with demands in order to attain the corporative plans, that which will eventually be converted to production 140 Chang and Lin: Decision Support and Expert Systems in a CIM Environment 141 plans. The process is often a recursive one. Attempts and adjustments are required to ensure the feasibility of the outcomes. Potential capacity problems will be identified. The long term capacity planning pinpoints the discrepancy between the demands and the available resources in the long run, and decides on what kind of adjustment is needed to support a realistic master production scheduling. Getting to the operational level, short term time phased production scheduling is set to observe the predicted loads, job-shop status and bottlenecks. The short term capacity planning will handle the adjustments by temporarily increasing or decreasing its capacities. With those modifications, MRP II becomes more flexible in handling the dynamic environment and the output scheduling is also more realistic (see [5]). MRP II and likewise systems support manufacturing management with necessary information so that the planners can make judgment and decisions. Let's use the long term capacity planning as an example. This function needs input knowledge such as the forecast of demands, the current capacity status, the inventory status and the production status. The forecast of demands and master production schedules generated by the systems will indicate a rough estimated manpower and machine capacity in each level of the assembly process. A long term (rough-cut) time-phased capacity requirement will be generated and compared against the existing capacity status. If a manufacturing firm is found undercapacity, the finding may justify the needs for increasing man and machine power. Alternatives such as hiring new workers, increasing work shifts, opening new production lines, installing new machines, replacing current facilities and equipments with more efficient ones, and even building a new plant would be considered possible solutions. On the other hand, when the firm is found constantly overcapacity, solutions like reducing working shifts, laying off employees, closing clown production lines and plants will be under consideration to correct the problem. Decision support systems will be used to test alternatives in either case. The objective is to develop a realistic and optimum long term capacity plan. The help from the information models often sharpens up such planning and control functions and improves the overall competitive status. Although the information technology is getting more powerful and sophisticated than ever, in many situations human intelligence and expertise still play a determinant role. Often users' limited knowledge in their areas of responsibility would affect their level of performance. In recent years, the development of knowledge based expert systems provides a solution to permit an expert's performance to the ordinary users. Although the commercialization of expert systems began only a few years ago, the progress in developing applications for business and industry has been dramatic. Manufacturing systems, by no exception, also find the invaluable use of expert systems to support their operations and management. For example, it is no news .for manufacturing engineers and operators to face substantial production line shut down caused by the failure of complex equipments, or spend costly process planning for each new manufacturing part, or go through time consuming scheduling at planning and shop floor levels. All these activities need the knowledge and expertise to perform their corresponding functions. The development of expert systems gives a lot of hope in those areas. To construct a knowledge based expert system, human expertise is transferred into means that can be accessed by the system, and forms a knowledge base. Knowledge and expertise often need refinement judged by the testing process, and should constantly be updated. The process will continue to keep the system's performance at the export level. Inference mechanism of the system accesses the knowledge base to generate analysis and judgment. User-friendly interface allows non-technical users to operate the system without extensive training. Some system provides explanations behind the generated output when questioned by the user, or suggests alternatives from which the user can choose. For more practical use, some expert systems are constructed so as to make judgment from incomplete or fuzzy information. Although it is costly and time-consuming to build an expert system, there are a lot of attractive benefits. The duplication cost is very low and easy. The systems are highly mobile, eliminating geographical barriers and are available twenty-four hours a day. Besides, the system can incorporate knowledge from different experts from different time and locations into a common knowledge base. However, expert systems do have limitations. For one, they are usually designed to serve specific problems only. They are not the general problem solver, and it is too costly to be one. Because of that reason, only the specific complex problems are considered the targets of expert systems. Process planning, for example, is a major application of expert systems in the manufacturing environment. The function converts the design information into manufacturing process information in terms of process routes, process equipments and tools, 142 Proceedings of the 12th Annual Conference on Computers & Industrial Engineering ~ j ~ MARKETING ~ ~ i STRATIG|EsCOMPANY &~REQUIREMENTI - ~ 1 & CUSTOMERI~ 1 I ~SPECiFICATION I v I RE.ANOHI I BUSINESS PLANS + DER/AND ~ PLANNING" PROCESS CONTROL ~'~ F T DESIGN PLANNING PLANNING i.o~.~;~ M,,~ro~ MATERIAL BILL OF I I MATERIAL REQUIREMENTS PLANNING I .ORT-TER.I - i . ~ R I PRODUCTION INVENTORY "ANAGE"ENT~--4~M°NGI-I ~ VENDOR I . .=CTION I I SUPPLIEE I l < / W A R E H O U S , N G I IPOST-SALESi - * J , N S P E C T , O . PACKAG,NG • SERVICE • " ) w ~ V l SHIPPING I I FEEDBACK I v F i g u r e 1 I n f o r m a t i o n f l o w in a C I M e n v i r o n m e n t operation sequences, etc., and is considered to be the key element in integrating the computer-aided design (CAD) and the computer-aided manufacturing (CAM). The function requires a considerable amount of expertise, has been the area of expert system application ever since the early eighties. The maintenance of complex equipment is another example. It usually involves fault diagnostic procedures incorporating various maintenance rules and judgment decisions. The personnel expertise and knowledge play an important role in locating the problem and determining the appropriate corrective actions (see [1] and [8]). INTEGRATION OF DECISION SUPPORT AND EXPERT SYSTEMS P_~i_vJmmCEa Decision support system provides the flexibility and adaptability in dealing with ill-defined problems, partial information, conflicting objectives and ad hoc questions. Such a system is found particularly useful to support the solutions of a wide range of semistructured and unstructured problems like strategic and tactic planning. Knowledge-based expert system incorporates knowledge from experts to provide users with expert level considerations. Most expert systems are developed to support specific and narrow application domains. DSS and ES represent two kinds of information technology tools serving for distinct purposes, yet they are not necessarily conflicting to each other. In most situations, they can actually complement each other. A combination of DSS and ES can yield an even better result. User may use DSS regarding information acquisition and information evaluation, and at the same time use ES to get intelligence for a particular domain and a suggestion for the tentative decision. The joint effort can provide users with better information to make final decisions. Information Flow in a ClM Environment To determine which kind of information technology, DSS or ES, should be applied to which function in a CIM environment, or whether the integration of DSS and ES would be appropriate for a particular function, one should first study the information flow in the CIM. Most companies now believe the importance of "the voice of customer", and it is the customer requirements that drive the company's operations. The product life cycle in such companies will start with marketing research, where customer needs and competitors' strategies are studied. Collected information will help to set up long term strategies and develop business plans. That is a typical area where traditional decision support systems demonstrate their abilities using modeling, "what-if" simulation, and quantitative analysis to help user reason and make decisions. The marketing research and business plans determine the kinds of product the company wants to produce. The specific descriptions of the product rely on the design and manufacturing considerations. It used to be an iterative process between the product design division and manufacturing division to ensure that the final design meets customer requirements and will not cause manufacturing Chang and Lin: Decision Support and Expert Systems in a CIM Environment 143 complications. The recent simultaneous engineering concept encourages both design and manufacturing divisions participate in the design process and work together, so the design will be manufacturable and also fully functional in the way customers expect. The final design will be converted to manufacturing process plans by the process planning function. This function requires a lot of knowledge about the manufacturing environment and expertise to develop the route plans. Numerous existing expert systems are able to perform this function. Then the process control plans will be developed to ensure the quality during operation. On the other hand, the forecast of product demands and the information of available capacities and resources are used to generate manufacturing planning schemes. The master production schedules (MPS) lay out a time-phased product demand, and the material requirements planning (MRP) determines the requirements for each level of a product's structure (bill-of-material). At shop floor, the foremen will adjust the production schedule based on the short term available capacities, resources and production orders. Operations at the production lines usually emphasize on efficiency, productivity and high quality. Various expert systems were developed to support the individual operation (see [2] and [3]). Inteoratlon of DSS and ES Following the flow of information in a CIM environment, we may classify the major functions into three levels: long term strategic planning, short term process and operational planning, and production operations. Marketing and customer research, setting company strategies and business plans all belong to the first category. Traditional decision support systems are developed to support such functions. But expert systems can further play the role of consultants by providing expert advice on complex issues that deal with legal problems, tax problems and others. In that case, outputs from expert systems will be directed to the DSS for possible solutions. DSS here is still the main tool for the user, and ES only plays the supporting role (see [12] and [13]). From the development of company strategies and business plans to the production stage, MPS, MRP, capacity planning and shop floor control follow one information flow path that is served by MRP (or MRP II) systems. Product design, process and control planning follow the other information flow path, where numerous expert systems have been developed to support these individual functions. The integration of DSS and ES at this level will keep all the current systems with additional components to enhance the overall result. Along the first information flow path, expert knowledge about the customer behavior and forecasting is required to develop viable master production schedules. To choose a capacity adjustment strategy, an expert's advice will make a significant difference. Most shop floor control requires knowledge of manufacturing and plant operation procedures. Therefore expert systems should be developed to assist DSSs in those aforementioned functions. Along the second information flow path, there are expert systems already developed to support product design function and process planning. The results of quantitative analysis provided by DSSs are often used as input to those expert systems. For example, engineering analysis is required during a design process, and DSSs can be useful in choosing a combination of feeding rate and operation speed, and also other parameters in developing process plans. Under that situation, DSS outputs will be directed to the ES (see [4]). Figure 2. Integration of DSS & ES In an organization At the operational level, many production and operational processes can be benefited by using expert systems achieve expert performance. Most operations at this level require little decision making. For specific operations at a narrow application domain, expert system is the ideal tool. Operations such as welding, material handling, and machining are all becoming the service targets for expert systems (see [11]). Hence it would make a lot of sense to adopt an integrated system if other information systems, when combined with expert systems, would generate more economic and efficient results. Therefore we should use expert systems to perform the operation and use DSS and other information technologies as supplements in the system. r=.O.UgL.U_~Jg~ Decision support systems provide the flexibility and adaptability to support the solution of a wide range of semistructured and unstructured problems like strategic and long 144 Proceedings of the 12th Annual Conference on Computers & Industrial Engineering term planning. Expert systems incorporate [5] knowledge from experts to provide users with expert level considerations in specific and narrow application domains. In a computer integrated manufacturing environment, well planning, control and operation require both expert knowledge of the area, and decision support capabilities. This paper discusses the [6] features of decision support systems and expert systems, and their integration to support the major functions from marketing and strategic level considerations to manufacturing operational planning and process. From the hierarchical structure of information flow in a [7] company, this paper attempts to find the best way of combining decision support systems with expert systems in enhancing the planning, control and operational functions in a CIM environment. [8] [1] [2] [3] [4] BE.EEBEKGE~ Chang, C., "Expert Systems for Manufacturing Operations and Management," Proceedings of 1988 North Amedcan IBSCUG Conference, Oxford, pp. 220-226, 1988. Chang, C., "Quality Function Deployment (QFD) Processes in an Integrated Quality Information System," Int. J. of Computers and Industrial Eng., Vol. 17, Nos. 1-4, pp. 311-316, 1989. Chang, C., "The Structure of Quality Information System in a Computer Integrated Manufacturing Environment," Int. J. of Computers and Industrial Engineering, Vol. 15, Nos. 1.4, pp. 338-343, 1988. Chang, C. and T. Steiner "Computer Aided Manufacturing Planning and Control Expert Model," 1986 lASTED Applied Simulation and Modeling Conference Proceedings, Vancouver, pp. 536-540. 1986. [9] [10] [11] [12] [13] Chang, C. and L. Tsui "Integrated Manufacturing Planning & Control System Using MRP I1," Proceedings of the ISMM Int. Symposium of Computer Applications in Design, Simulation and Analysis, Reno, pp. 302-305, 1989. Chang, C. and K. Yeung "Manufacturing Process Planning Expert Systems," Proceedings of Third Int. Conference of Expert Systems: Theory & Applications, Los Angeles, pp. 126-129, 1988. Groover, M. and E. Zimmers, Jr. CAD/CAM: Computer-aided Design and Manufacturing, Prentice Hall, Englewood Cliffs, 1984. Heragu, S. and A. Kusiak "Analysis of Expert Systems in Manufacturing Design," IEEE Transactions on Systems, Man and Cybernetics, Vol. SMC-17, No. 6, pp. 898-912, 1987. Laudon K. and J. Laudon Management Information Systems, Macmillan Pub., New York, 1988. Long, L, Management Information Systems, Prentice Hall, Englewood Cliffs, 1989. Miska, K. "The New Mavens of Manufacturing," Manufacturing Engineering, Oct. pp. 36-39, 1989 Rouse, W. "Intelligent Decision Support for Advanced Manufacturing Systems," Manufacturing Review, Vol. 1, No.4, pp. 236-243, 1988. Sharit, J., R. Eberts and G. Salvendy "A Proposed Theoretical Framework for Design of Decision Support Systems in Computer-integrated Manufacturing Systems: A Cognitive Engineering Approach," Int. J. of Production Research, Vo. 26, No. 6, pp. 1037-1063, 1988. 
work_oosyya7ztreevjkgbwbnfx5tom	----	Ontology-based sentiment analysis of twitter posts Expert Systems with Applications 40 (2013) 4065–4074 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Ontology-based sentiment analysis of twitter posts Efstratios Kontopoulos a,⇑, Christos Berberidis a,1, Theologos Dergiades a,2, Nick Bassiliades a,b,3 a School of Science and Technology, International Hellenic University, Thessaloniki, Greece b Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece a r t i c l e i n f o Keywords: Micro-blogging Twitter Tweet Sentiment analysis Ontology 0957-4174/$ - see front matter � 2013 Elsevier Ltd. A http://dx.doi.org/10.1016/j.eswa.2013.01.001 ⇑ Corresponding author. Tel.: +30 2310 807538. E-mail addresses: e.kontopoulos@ihu.edu.gr (E. Ko edu.gr (C. Berberidis), t.dergiades@ihu.edu.gr (T. Der (N. Bassiliades). 1 Tel.: +30 2310 807534. 2 Tel.: +30 2310 807501. 3 Tel.: +30 2310 998913. a b s t r a c t The emergence of Web 2.0 has drastically altered the way users perceive the Internet, by improving infor- mation sharing, collaboration and interoperability. Micro-blogging is one of the most popular Web 2.0 applications and related services, like Twitter, have evolved into a practical means for sharing opinions on almost all aspects of everyday life. Consequently, micro-blogging web sites have since become rich data sources for opinion mining and sentiment analysis. Towards this direction, text-based sentiment classifiers often prove inefficient, since tweets typically do not consist of representative and syntactically consistent words, due to the imposed character limit. This paper proposes the deployment of original ontology-based techniques towards a more efficient sentiment analysis of Twitter posts. The novelty of the proposed approach is that posts are not simply characterized by a sentiment score, as is the case with machine learning-based classifiers, but instead receive a sentiment grade for each distinct notion in the post. Overall, our proposed architecture results in a more detailed analysis of post opinions regarding a specific topic. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction The emergence of Web 2.0 has drastically altered the way users perceive the Internet, by improving information sharing, collabora- tion and interoperability. Contrary to the first generation of websites, where users could passively view content, Web 2.0 users are encouraged to participate and collaborate, forming virtual on-line communities. Additional Web 2.0 traits include data openness and metadata, dynamic content, rich user experience and scalability tolerance (Skiba, 2006). Among the most popular Web 2.0 applications, one can come across social-networking sites (e.g. Facebook, MySpace), wikis, blogs, multi-media sharing sites (e.g. YouTube, Flickr), mash-ups and rich web applications. One of these activities is micro-blogging, which initially attracted comparatively less attention, but gradually became a highly popu- lar communication tool for a considerable percentage of users. Micro-blogging is in principle based on blogs (i.e. Web logs), where users can post opinions, experiences and queries on any chosen topic. The main difference between micro- and traditional blogs is the strict constraint in content size (Kaplan & Haenlein, ll rights reserved. ntopoulos), c.berberidis@ihu. giades), nbassili@csd.auth.gr 2011). Currently, the most popular on-line micro-blogging service is Twitter, which enables its users to send and receive text-based posts, known as ‘‘tweets’’, consisting of up to 140 characters. Twit- ter was created in 2006 and currently records over 140 million ac- tive users that generate over 340 million tweets per day. With their rapidly increasing popularity, micro-blogging services, like Twitter, have evolved into a practical means for sharing opinions on almost all aspects of everyday life. The strict character limit of tweets forces users to be concise and eventually more expressive than with social networks and blogs. Thus, micro-blogging posts are im- bued with emotional information and are considered as rich opin- ion mining data sources (Pak & Paroubek, 2010). Additionally, tweets can be processed more effectively than lengthy blog posts and articles. Opinion mining, also known as sentiment analysis, is the process aiming to determine whether the polarity of a textual corpus (doc- ument, sentence, paragraph etc.) tends towards positive, negative or neutral. Sentiment analysis constituted a popular research area even before Twitter and micro-blogging, since it can offer advanta- ges to a variety of domains, from sales predictions (Liu, Huang, An, & Yu, 2007) to politics (Park, Ko, Kim, Liu, & Song, 2011) and investors’ choices (Dergiades, 2012). Furthermore, the automatic detection of sentiment on textual corpora has comprised a research topic for many approaches. Examples include, among others, product and services reviews (Kang, Yoo, & Han, 2012; Prabowo & Thelwall, 2009), articles on the Web (Melville, Gryc, & Lawrence, 2009) and news feeds (Moreo, Romero, Castro, & Zurita, 2012; Wanner, Rohrdantz, Mansmann, Oelke, & Keim, 2009). However, the http://dx.doi.org/10.1016/j.eswa.2013.01.001 mailto:e.kontopoulos@ihu.edu.gr mailto:c.berberidis@ihu.edu.gr mailto:c.berberidis@ihu.edu.gr mailto:t.dergiades@ihu.edu.gr mailto:nbassili@csd.auth.gr http://dx.doi.org/10.1016/j.eswa.2013.01.001 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa asghari Highlight asghari Sticky Note این رو فقط برای این مثال زدم که توی اسلایدای ارائه میتونی بیاری با دسته بندی اینکه هر کدوم روی چی کار کردن 4066 E. Kontopoulos et al. / Expert Systems with Applications 40 (2013) 4065–4074 expressive power and immediacy of Twitter in particular has led researchers to try utilizing it in further approaches related to poli- tics (Tumasjan, Sprenger, Sandner, & Welpe, 2010), tourism (Claster, Cooper, & Sallis, 2010) and many more. One of the most promising applications of sentiment analysis of tweets could be in the field of economic and financial modeling. Econometric specifications that do not incorporate variables, such as the investors’ or consumers’ sentiment, prove to be inefficient. However, the real-time discovery of individuals’ sentiment is a challenging task that often involves analysts digging manually through virtually infinite articles and news feeds. Twitter analysis offers one of the best solutions to-wards the automatic discovery of sentiment. Machine learning-based sentiment classifiers consti- tute a prevailing sentiment analysis tool. However, they can often prove less efficient in the case of tweets (Barbosa & Feng, 2010; He & Alani, 2011), since the latter do not typically consist of repre- sentative and syntactically consistent words, due to the character restriction. An additional limitation is that classifiers usually distin- guish sentiment into classes (positive, negative and neutral), assigning a corresponding score to the post as a whole, regardless the fact that many aspects of the same ‘‘notion’’ may be discussed in a single post. Consider, for example, the sample tweet Tex: ‘‘The screenplay was wonderful, although the acting was rather bad’’. The machine-learning based approaches would return a single quanti- tative (sentiment score) or qualitative (positive, negative or neu- tral) result. In this paper, we propose the deployment of ontology-based techniques towards a more fine-grained sentiment analysis of Twitter posts. According to the proposed approach, tweets are not simply characterized by a sentiment score, but in- stead receive a sentiment grade for each distinct notion in the post. This results overall in a more elaborate analysis of post opinions regarding a specific topic. More specifically, regarding the sample tweet Tex above, our proposed ontology-based approach distin- guishes the features of the domain (in this case, screenplay and act- ing) and assigns respective scores, resulting in a more detailed sentiment analysis of the given statement. The rest of the paper consists of the following: Section 2 reports on the related research efforts focusing on sentiment analysis on mi- cro-blogging data, followed by a section elaborating on the deploy- ment of ontologies in the micro-blogging domain. Section 4 explains the proposed sentiment analysis approach, outlining the distinct phases of the methodology and also includes a baseline sce- nario that better illustrates the whole process. Finally, the paper con- cludes with some final remarks as well as directions for future work. 2. Sentiment analysis in micro-blogging data Researchers in their effort to perform sentiment analysis on mi- cro-blogging posts initially applied mainstream methodologies, used for analyzing the sentiment of ‘‘normal’’ textual corpora, like e.g. product reviews. More specifically, two main approaches were commonly used, in order to conclude, whether a piece of text (sentence, paragraph or document) expresses a positive or negative sentiment: the lexicon-based and the machine learning-based approach. The former approach (Kaji & Kitsuregawa, 2007; Neviarouskaya, Prendinger, & Ishizuka, 2009; Taboada, Brooke, Tofiloski, Voll, & Stede, 2011) is based on opinion words, namely, words that are commonly used in expressing positive or negative sentiment. Opinion words are typically contained in a dictionary called opinion lexicon. However, tweets are not considered ‘‘nor- mal’’ pieces of text, since the 140-character threshold imposes lim- itations in the length of words and phrases. A further peculiarity is the extensive usage of ‘‘every-day’’ (i.e. Jargon) expressions, abbre- viations and emoticons (sequences of symbols representing an emotion). One could suggest adding these words and symbols to the lexicon, but this would nevertheless generate dubious results. These expressions are usually of a dynamic nature, changing con- stantly and being replaced frequently, following each time the pop- ular trends on the Web. An additional disadvantage is the fact that jargon expressions are often domain dependent. These factors lead to low recall, when the lexicon-based method is applied on ‘‘infor- mal’’ corpora of text, like posts from micro-blogs. An alternative approach is the application of machine learning methodologies (Pang & Lee, 2008). According to this approach, a sentiment classifier is trained, in order to distinguish positive, neg- ative and neutral sentiments in textual corpora. Typical features used in training the classifiers are unigrams or bigrams (n-grams of size 1 and 2, respectively (De Kok & Brouwer, 2012). The draw- back of the machine learning-based methods is mainly focused on the manual labeling required over (usually) massive sets of tweets. Additionally, the labeling has to be performed on each distinct domain of interest, in order to achieve satisfactory training levels for the classifier on the given domain (Aue, 2005). As outlined in (Saif, He, & Alani, 2012), in relation to sentiment analysis which is derived from Twitter posts, the following approaches can be distinguished: 1. Working with noisy labels or distant supervision (Read, 2005), by creating, for example, emoticon vocabularies for represent- ing sentiment and for training supervised sentiment classifiers, such as Naïve Bayes (NB), Maximum Entropy (MaxEnt) and Support Vector Machines (SVMs) (Barbosa & Feng, 2010; Pak & Paroubek, 2010). 2. Applying a combination of feature engineering (e.g. using the feature-based model and the tree kernel-based model for senti- ment classification, as well as n-gram and lexicon features) with machine learning methods to improve sentiment classification accuracy on tweets (Agarwal, Xie, Vovsha, Rambow, & Passonneau, 2011; Kouloumpis, Wilson, & Moore, 2011). As described in the next section, there already exist sentiment analysis methods that deploy ontology-based techniques. Never- theless, none of the existing approaches applies ontologies simi- larly to the way we propose in this work that results in specifically evaluating the sentiment for the various distinct no- tions (i.e. properties) at hand. 3. Micro-blogging and ontologies An ontology can be defined as an ‘‘explicit, machine-readable specification of a shared conceptualization’’ (Studer, Benjamins, & Fensel, 1998). Ontologies are used for modeling the terms in a do- main of interest as well as the relations among these terms and are now applied in various fields, like agent and knowledge manage- ment systems and e-commerce platforms (Gómez-Pérez & Corcho, 2002). Other applications include natural language generation, intelligent information integration, semantic-based access to the Internet and extracting information from texts. However, the most important contribution of ontologies is the key role they play in the development of the Semantic Web. The Semantic Web is an extension of the current Web, where information is given a well-defined meaning, encouraging cooper- ation among human users and computers (Berners-Lee, Hendler, & Lassila, 2001). Ontologies serve as the primary means of knowl- edge representation in the Semantic Web. Although various ontol- ogy languages have emerged, the currently dominant standards are RDF/S (Resource Description Framework Schema) and OWL (Web Ontology Language). Additional points of motivation for preferring the use of ontologies in an application include: (a) analyzing domain knowledge and separating the latter from operational asghari Highlight asghari Highlight asghari Sticky Note کلیت داره این رو بیان میکنه که توی روشهای معمولی یادگیری ماشین از یه حدی نمیتونه ریزتر شه ولی اینجا با انتولوژی در واحد "ایده" به نظر رتبه میدیم. asghari Highlight asghari Highlight E. Kontopoulos et al. / Expert Systems with Applications 40 (2013) 4065–4074 4067 knowledge, (b) Enabling the reuse of domain knowledge, (c) Mak- ing domain assumptions explicit, and (d) Sharing a common understanding of the information structure among people and/or software agents. Regarding the deployment of ontologies in the micro-blogging domain, to the best of our knowledge, the most prominent effort belonging to this category is the recent work by Iwanaga, Nguyen, Nakagawa, and Ohsuga (2011). In their approach, the authors present a methodology for populating an existing earthquake evacuation ontology with instances based on tweets. The proposed approach extracts related information like evacuation center names, products offered at the centers and the timestamp of each tweet. Additional information retrieved from the Web is appended, including the evacuation center address (retrieved via Google Maps), the center’s latitude and longitude (via Geocoding) and Japanese-to-English translation of all the above (via Google Translation). This information is obtained in real time, even though it is not included in the initial tweet. Other directions of research include the development of ontol- ogies for representing micro-blog posts and relationships between social-network users as for example FOAF (see Brickley & Miller, 2010), SIOC (see Breslin, Harth, Bojars, & Decker, 2005), OPO (see Stankovic, Passant, & Laublet, 2009), SMOB2 (see Passant, Bojars, Hastrup, & Laublet, 2010), or ontologies for representing levels of emotions (e.g. Baldoni, Baroglio, Patti, & Rena, 2012; Francisco, Germans, & Peinado, 2007). These topics, however, are unrelated to the line of research presented in this paper. Our approach assumes a different direction as far as the usage of ontologies is concerned. While Iwanaga et al. (2011) deploy the ontology as a means for modeling the application domain, we extend the usabil- ity of the ontology in our approach. The domain of reference is semantically divided and sentiment scores are assigned not to whole statements (i.e. tweets), but to the various aspects of the to- pic at hand. For example, according to ‘‘traditional’’ sentiment analysis techniques, the sample tweet Tex presented in the intro- duction would probably acquire a positive score, since ‘‘wonderful’’ is rather stronger than ‘‘bad’’. However, according to our proposed ontology-based approach, the subject (movie) would have two dis- tinct features (screenplay and acting) with respective scores 8 and �3, providing higher granularity in the sentiment analysis of the given statement. 4. Description of the proposed approach As already explained, the basic idea behind the proposed ap- proach is to take advantage of a domain ontology for providing more elaborate sentiment scores regarding the notions contained in a tweet. The aim is to have a system that accepts as input a tweet (or a set of tweets) regarding a specific subject and provides senti- ment scores for every aspect/feature of this subject. The architec- ture of the system we developed is presented in detail in a following section. The proposed methodology is divided in two phases: (a) crea- tion of the domain ontology (Section 4.1), and (b) sentiment anal- ysis on a set of tweets, based on the concepts and properties included in the ontology (Section 4.2). These two phases are fur- ther described next. 4.1. Creating the domain ontology In order to create a domain ontology, one can adopt various methods, like extending existing ontologies or developing the ontology from the ground up. In this work, we examine two alter- native approaches, which are presented subsequently: (a) Formal Concept Analysis, and (b) Ontology Learning. 4.1.1. Formal Concept Analysis Formal Concept Analysis (FCA) is a mathematical data analysis theory, typically used in Knowledge Representation and Informa- tion Management (Ganter & Wille, 1999). Its main characteristic is that is applies a user-driven step-by-step methodology for creat- ing domain models. With the recent emergence of the Semantic Web and the establishing of ontologies as its principal means for knowledge representation (see Section 3), FCA has been accounted as a valuable engineering tool for deriving an ontology from a collection of objects and their properties (Obitko, Snasel, Smid, & Snasel, 2004). Towards this affair, FCA has recently been applied in various occasions (e.g. Fu & Cohn, 2008; Ning, Guanyu, & Li, 2010; Zhang & Xu, 2011) and has been preferred in this work, since it offers the following advantages: 1. Appropriate ontology size: The domain ontology is gradually developed, depending on the given data set (i.e. in this case tweets). Therefore, it does not contain unnecessary concepts and/or properties, which would result in redundancy and potential illegibility on behalf of the user. On the other hand, based on the same principle, the output ontology does not lack essential concepts, either. 2. Better ontology design: Concepts and concept hierarchies are not explicitly defined, but are dynamically designated via the detected properties. This leads to better ontology design and a more distinct classification of concepts. 3. Domain specific ontology: Towards creating a domain ontology, one would reasonably consider utilizing existing ontologies that describe the given domain in more detail. However, the aim is not to exhaustively describe the application domain, but to develop an ontology that corresponds to the given set of tweets, namely, the notions that currently ‘‘matter the most’’ to the users. Otherwise, the result would be an ontology that contains numerous classes and attributes that never appear in the data set and for which no sentiment assessment score can be derived. 4.1.1.1. FCA basic elements. The main building block in FCA is the concept, which is described via two sets: The extension, which is a set of objects and the intension, which is a set of attributes (Ganter & Wille, 1999). Every object that belongs to the concept has all the attributes in the intension and every attribute that belongs to the concept is shared by all objects of the extension. The relationships between the set of objects and the set of attri- butes are represented by a formal context. A formal context K (O;A;I) is a triple where: � O is a set of formal objects, � A is a set of attributes, and, � I is a binary incidence relation between the objects and the attributes; I #O�A, where (o, a) 2I is read as ‘‘object o has attribute a’’. A formal context can be represented as a cross-table, where the rows represent O, the columns represent A and the inci- dence relation I is represented by a series of crosses as shown in Table 1. Concepts can be organized into a concept lattice, which is based on the mathematical notion of lattices (Davey & Priestley, 2002). Concept lattices are visualized via Hasse dia- grams (or line diagrams – see Fig. 5). The nodes in the diagram represent concepts, while attributes and objects are denoted above and below the nodes, respectively. By traversing all the paths leading down from a node, one can retrieve the concept’s extension, while the opposite path leading up from the node re- trieves the intension. Table 1 Smartphone ontology. Model series Properties Android Camera Battery 4G microUSB Processor Windows Display iOS Apple iPhone + + + + + Samsung galaxy + + + + + + + HTC one + + + + + + + Nokia lumia + + + + + + Fig. 1. Ontology creation algorithm via FCA. 4068 E. Kontopoulos et al. / Expert Systems with Applications 40 (2013) 4065–4074 4.1.1.2. Populating the concept cross-table. According to FCA, the aim is to create the concept cross-table that corresponds to the do- main ontology. As mentioned previously, the table corresponds to the incidence relation I #O�A; thus, it is essential to determine sets O and I. These sets can be derived via the algorithm illustrated in Fig. 1. 1. The algorithm accepts as initial input a concept parameter (e.g. ‘‘#smartphone’’), determined manually by the user. The retriev- eTweets method retrieves the first N tweets (default value for N is 100) that belong to the result set corresponding to the concept parameter. Towards this affair, Twitter4J4 was utilized, a Java library that gives access to the Twitter API and assists in integrating the Twitter service into any Java application. The retrieveObject method subsumes that every tweet is inspected for references to objects and, if any such reference is detected, the corresponding attributes are retrieved via method retrieveAttributes. For every attribute asso- ciated to the object, the attribute is appended to the existing set of attributes. The latter two methods are currently performed manually by an ontology engineer. Fig. 2(A) displays two sample tweets used for the creation of the ontology. The detected objects and attributes are represented in boldface fonts. Eventually, the final concept table is populated with the detected objects and attributes. Table 1 displays the resulting cross-table as well as the corresponding concept lattice, after ana- lyzing 100 retrieved tweets for the concept ‘‘smartphone’’. 4.1.2. Ontology learning An alternative to the manual FCA methodology presented above is offered by the various existing semi-automatic ontology learning 4 Twitter4J Java library: http://twitter4j.org/en/index.jsp techniques (e.g. Cimiano & Völker, 2005; Hazman, El-Beltagy, & Rafea, 2009). Ontology learning, also known as ontology extraction, ontology generation or ontology acquisition, refers to the task of automatically creating an ontology, via extracting concepts and relations from a given data set. Nevertheless, the task of creating an ontology in a fully automated manner still remains elusive to a great degree. In this work, we resorted to OntoGen (Fortuna, Grobelnik, & Mladenic, 2007), a semi-automatic, data-driven ontology editor. The software deploys text-mining techniques via an efficient user interface that reduces development time and complexity. Overall, the tool attempts to bridge the gap between complex ontology edi- tors and domain experts, who do not necessarily possess ontology engineering skills. OntoGen interactively offers assistance during the development of domain ontologies, by suggesting concepts and relations and by automatically assigning instances to the concepts. The user can ac- cept or reject the suggestions or perform manual adjustments. Most of the aid provided by the system is based on the provided underlying data. In our case, the role of this data set is played by the initial set of retrieved tweets (see previous subsection), which is fed to OntoGen as a set of named line-documents (i.e. each tweet is stored as a separate text file, with the first word in the line serv- ing as the document title/ID). Fig. 3 illustrates the resulting ontol- ogy visualization via OntoGen, after examining the same set of 100 retrieved tweets for the concept ‘‘smartphone’’. 4.1.3. Augmenting the semantics The ontology created via FCA (Section 4.1.1) or Ontology Learn- ing (Section 4.1.2) is in essence a simple taxonomy of concepts and attributes. In order to augment the underlying semantics, the ontology is enriched with synonyms and hyponyms (subordinate notions) of the detected attributes. For example, in the http://twitter4j.org/en/index.jsp Fig. 2. (A) Sample tweets used for the creation of the ontology, and (B) Sample tweets used for sentiment analysis, accompanied by the corresponding sentiment scores. Fig. 3. Ontology visualization via OntoGen. E. Kontopoulos et al. / Expert Systems with Applications 40 (2013) 4065–4074 4069 ‘‘smartphone’’ universe used throughout this paper, the ‘‘display’’ attribute could also be expressed as ‘‘monitor’’ or ‘‘screen’’, which are synonymous words. For appending the sets of synonyms and hyponyms to the ontology, we used the popular WordNet lexical database (Miller, 1995), which retrieves synsets (groups of synonymous words or collocations) of the synonyms and hyponyms of every given word. Each synonym and hyponym is then added to the ontology and associated with the initial attribute. Syntactically, in the OWL DL representation of the ontology, these associations are expressed via the owl:equivalentProperty and rdfs:subPropertyOf constructs, respectively. 4.2. Sentiment analysis on tweets The previously described process (Section 4.1) results in a for- mulated and populated domain ontology. The second phase of the proposed methodology constitutes the main effort of this work and performs the automatic sentiment analysis on a set of tweets. Fig. 4 displays the overall architecture of the approach we propose in this paper. As can be seen from Fig. 4, the overall process involves retriev- ing a set of tweets that correspond to entities in the ontology and performing sentiment analysis on each of the retrieved tweets. There are three distinct steps in the procedure: (1) querying the ontology for the corresponding attributes of each object, (2) retrieving the relevant tweets, and (3) performing the sentiment analysis. 4.2.1. Step#1: Taking advantage of the ontology In order to take advantage of the domain ontology created dur- ing the previous steps, the retrieved tweets have to contain infor- mation regarding the objects and attributes of reference. This is achieved via JENA (Rajagopal, 2005), a Java API for processing RDF/S and OWL ontologies. Having an ontology-based structured hierarchy of classes and properties, JENA assists in retrieving ob- Fig. 4. Architecture of the proposed approach. Fig. 5. Ontology visualization via a Hasse diagram. 4070 E. Kontopoulos et al. / Expert Systems with Applications 40 (2013) 4065–4074 ject-attribute pairs (oi, aij) – see Fig. 4. More specifically, for every object/class oi, all attributes/properties aij are retrieved via process- ing RDF/S triples of the form: <aij rdfs:domain oi>. 4.2.2. Step #2: Retrieving the relevant tweets For every property aij of an object oi a relevant query is submit- ted to Twitter via the Twitter4J library described previously. The query has the form ‘‘oi aij’’, where different terms are separated by whitespaces, resulting in an intersection query. Alternatively, one could execute a hashtag intersection query, like e.g. ‘‘#oi #aij’’, which would nevertheless drastically reduce the result set, without necessarily increasing the precision. A predefined number of tweets t1i, . . . , t1n is retrieved (default number is 100) that contain the relevant keywords. A secondary phase of preprocessing takes place on the retrieved set of tweets. The preprocessing phase involves removing characters or se- quences of characters that cannot assist during the subsequent sentiment analysis phase, in order to reduce the noise in the data set. More specifically, for each retrieved tweet, the following items of text constitute representative examples to be removed: 1. Replies to other users’ tweets, represented by strings starting with ‘@’. 2. URLs (i.e. strings starting with ‘http://’). 3. Hashtags, which are strings starting with ‘#’ used for categoriz- ing messages are not removed as a whole. Instead, only the ‘#’ character is removed, since the rest of the string often forms a legible word that contributes to better understanding the tweet. The remainder of each tweet is added into a collection of sentences. 4.2.3. Step #3: Sentiment analysis After going through the preprocessing phase during the previ- ous step, the retrieved tweets are submitted to OpenDover for sen- timent analysis. OpenDover5 is a web service that tags the opinions 5 OpenDover sentiment tagging web service: http://opendover.nl/. and sentiments detected in a textual corpus, based on the subject domain, as well as the intensity of the sentiment expression. A sen- timent score s is assigned to each tweet, where s 2 [�10, 10], depending on the appreciation level of the submitted sentence. Fig. 2(B) displays two sample preprocessed tweets retrieved during this phase, as well as their corresponding sentiment score generated from OpenDover. OpenDover was considered an appropriate choice for the pro- posed approach, since it is suitable for extracting sentiment from isolated sentences. An additional advantage is OpenDover’s ontol- ogy-based architecture that offers the capability of detecting each time the domain of reference, adjusting the sentiment scores accordingly. It should be noted, however, that OpenDover could be substituted in the proposed architecture (Fig. 4) by any other equally efficient sentiment analysis tool and approach; the novelty in this work lies in the ontology-based analysis of tweets preceding the sentiment analysis phase. This analysis provides a more fine- grained sentiment evaluation regarding the distinct topics of a spe- cific subject, discussed in each tweet. On the other hand, deploying a third-party sentiment analysis service like OpenDover may be considered as a drawback, since the exact process of extracting the sentiment from a sentence can- not be verified – the source code and methodology behind Open- Dover are not publicly available. Thus, an imminent goal for the future is to integrate a custom sentiment analysis methodology in our approach and compare the resulting sentiment scores. 4.3. Baseline scenario The current subsection describes a baseline scenario that better illustrates the usability of the proposed approach. Suppose that a user wishes to perform a market research regarding smartphones and wants to determine other users’ opinions on Twitter regarding the most popular smartphone models. As the process described in this paper outlines, the first step is the creation of the domain ontology. In the scenario, we are going to adopt the FCA approach for creating the ontology (see Section 4.1.1). Thus, according to the ontology creation algorithm (Fig. 1), a default number of tweets is retrieved, based on the initial concept ‘‘#smartphone’’. After processing the tweet set as the algorithm suggests, the resulting ontology takes the form displayed in Table 1. As mentioned earlier, the ontology can be visualized via a Hasse diagram, illustrated in Fig. 5; the diagram was created with ConExp (Yevtushenko, 2000), a software tool for analyzing formal contexts in FCA, drawing the corresponding concept lattices and exploring dependencies between attributes. Alternatively, one could resort to ontology learning techniques for semi-automatically creating the ontology. Fig. 3 displays the resulting ontology visualization after using the OntoGen software tool. In essence, Table 1 depicts the formal context K (O;A;I) of the scenario, where: � O is the set of retrieved objects, namely, the detected smart- phone model series in the tweets, where O¼fApple iPhone; Samsung Galaxy; HTC One; Nokia Lumiag, http://www.opendover.nl/ Fig. 6. Sentiment values corresponding to the smart phone attributes in the scenario. E. Kontopoulos et al. / Expert Systems with Applications 40 (2013) 4065–4074 4071 � A is the set of properties, where O¼fAndroid; Camera; Battery; 4G; microUSB; Processor; Windows; Display; iOSg, and, � The incidence relation I is represented by a series of crosses as shown in the table, where a ‘+’ in cell (i, j) indicates that object oi has attribute aj. In order for the model series comparison to make sense, we re- move the attributes that are not common for every smartphone. This results in attribute set A0 ¼fajj8oi 2O;ðoi; ajÞ 2 Ig. More spe- cifically, A0 ¼fCamera; Battery; Processor; Displayg and, obvi- ously, A0 �A. However, this modification in the attribute set is taking place only for fairness of comparison among the different smartphone models and does not affect in any sense the method- ology proposed in this work. The ontology is integrated to the proposed system illustrated in Fig. 4. The system automatically collects tweets and submits them to the OpenDover web service, according to the previously de- scribed sequence of phases. The tweets retrieved for the scenario spanned over a 1-week period and are available as a comma- separated file at: http://goo.gl/UQvdx. The resulting sentiment values for each object-attribute pair are stored and the final results are depicted in Fig. 6. More specifically, the graph illustrates the average sentiment scores per attribute and model series. For each score, the total number of retrieved tweets is also displayed and, in parentheses, the corresponding positive- to-total tweet ratio. For reasons of objectiveness, the actual names of the smart phone model series in both tables have been substi- tuted with generic tags. 6 Given a random sample of T observations, the recall ratio for a particular methodology is defined by the ratio of the total number of relevant selected tweets over the sample size. 4.4. Evaluation The purpose of the current subsection is twofold: (a) to esti- mate the recall ratios6 for the two versions of our proposed archi- tecture (see Fig. 4) as well as for the custom-built system (which has been introduced for evaluation purposes only) and (b) to eval- uate whether the observed differences in the way the selections are performed by each method can be characterized as qualitatively analogous or not. The two versions of our proposed architecture are: (a) the full-fledged ontology-based semantically-enabled sys- tem (SEM), and, (b) the same system without the synonym/hyp- onym augmentation (see Section 4.1.3), but still with ontology support (ONT). The custom-built system is stripped of any ontol- ogy-based domain representation and, thus, cannot retrieve tweets referring to specific object-attribute pairs; it is limited to retrieving tweets regarding the superclass of the domain, which is associated to the ‘‘#smartphone’’ tag7 (CUS). The introduction of the CUS method is attributed to the fact that our approach adopts an utterly novel path and therefore it is not feasible to identify other methods that can be used as a comparison base. Additionally, there is no point in evaluating the returned sentiment results, since the Open- Dover sentiment classifier used in this work does not constitute a contribution of ours. The recall ratios for the three examined selection methods are estimated by using 10 randomly taken samples, each comprised of 100 observations. The estimation results, for all taken samples, reveal that SEM achieves steadily higher recall ratios from ONT and both present steadily higher recall ratios from CUS (See Panel A in Table 2). Solely based on the recall ratio as a comparison cri- terion a first round conclusion is that SEM performs better than ONT and both outperform CUS. However, we cannot draw any 7 The domain of reference still remains the world of smartphones, mentioned previously in the baseline scenario. http://www.goo.gl/UQvdx Table 2 Concordance statistics. Sample number Panel A Panel B Recall ratios Concordance statistics SEM CUM ONT SEM–CUS SEM–ONT CUS–ONT Sample1 0.94 0.72 0.26 0.30 0.78⁄⁄ 0.40 Sample2 0.86 0.72 0.40 0.36 0.86⁄⁄⁄ 0.48 Sample3 0.89 0.60 0.26 0.29 0.71⁄⁄ 0.46 Sample4 0.90 0.67 0.27 0.37 0.77⁄⁄⁄ 0.40 Sample5 0.79 0.60 0.32 0.41 0.81⁄⁄⁄ 0.48 Sample6 0.90 0.69 0.29 0.33 0.77⁄⁄⁄ 0.42 Sample7 0.85 0.65 0.26 0.29 0.80⁄⁄⁄ 0.39 Sample8 0.80 0.53 0.26 0.40 0.73⁄⁄⁄ 0.53 Sample9 0.88 0.60 0.36 0.44 0.72⁄⁄⁄ 0.50 Sample10 0.89 0.61 0.28 0.33 0.72⁄⁄⁄ 0.41 Notes: The number of observations for all the examined samples above is equal to 100. ⁄⁄⁄, ⁄⁄ and ⁄ indicate statistical significance at the 0.1 significance level, respectively. 4072 E. Kontopoulos et al. / Expert Systems with Applications 40 (2013) 4065–4074 solid conclusion if we do not first examine the significance of the degree of synchronization between all the investigated methods. Statistically significant synchronization between two methods implies no qualitative difference in the way the selections are performed by the methods under consideration while insignifi- cant synchronization suggests the opposite. For this purpose we employ the non-parametric Concordance Statistic (CS hereafter), as suggested by Harding and Pagan (2002), which evaluates the degree of synchronization among the three alternative selection methods (mj with j = 1, 2, 3). 8 Assuming for every selection method (mj) that the finally pro- duced outcome can be typified by two mutually exclusive states (relevant or irrelevant tweet selection), then the CS signifies (for two compared methods) the fraction of observations that indicate simultaneously the same state. To facilitate the computation of the CS we introduce a binary random variable Sm1;i which receives for observation i the value of one if the m1 method selects a relevant tweet and zero otherwise. For the remaining selection methods (m2 and m3) analogously are introduced the binary random vari- ables Sm2;i ,i and Sm3;i. The CS between m1 and m2 (CSm1;m2 ) mathe- matically is defined as: CSm1;m2 ¼ T �1 XT i¼1 ðSm1;i � Sm2;iÞþ XT i¼1 ð1 � Sm1;iÞ � ð1 � Sm2;iÞ ( ) ð1Þ where, T is the sample size and Sm2;i and Sm3;i are dichotomous ran- dom vectors defined as previously. A noticeable drawback of the above statistic is that it does not allow us to establish whether the identified degree of synchronization is statistically significant. To surmount the said drawback a Generalized Method of Moments (GMM) estimator is conscripted. In particular, the significance level of the CS is identified through the magnitude of the t-Statistic that corresponds to the coefficient a in the following moment condition: E Sm1;i � Sm1;i � � � Sm2;i � Sm2;i � � � a � � ¼ 0 ð2Þ where, Sm1;i and Sm2;i denote the sample means for methods m1 and m2, respectively. Equation (2) is estimated via the GMM technique using the Mar- quardt optimization algorithm, the Bartlett kernel and a fixed band- width equal to five. The estimated CS for the three resulting pairs between the alternative methods, along with their significance le- vel are analytically illustrated in Panel B of Table 2. The results re- veal that all the estimated degrees of synchronization for the first 8 In that way we actually examine whether the observed differences in the estimated recall ratios are statistically significant. (SEM and CUS) and the third pairs (CUS and ONT) are consistently non-significant and lie in a range between 0.29 to 0.44 and 0.39 to 0.53, respectively. On the contrary, the estimated degrees of syn- chronization for the second pair (SEM and ONT) are all statistically significant, mainly at the 0.01 significance level, with values that vary between 0.71 and 0.86. Overall, it can be argued that there is reasonable statistical evidence to support significant synchroniza- tion between the two versions of our architecture, while there is no evidence toward the direction of synchronization between the custom-built system and each one of the two versions of our archi- tecture. Therefore, our proposed architecture, given the higher ob- served recall ratios, appears to perform evidently better than the custom-built system method. 4.5. Difficulties During our work on the given field, a number of difficulties emerged, the most important of which are briefly outlined in this subsection, including suggestions for addressing each issue. A sig- nificant difficulty involves the unpleasantly high ratio of advertis- ing tweets. The latter are not necessarily negative for Twitter users, but unfortunately distort the derived results to a degree. Advertise- ments can either be misunderstood by our system as positive user tweets or are rejected by OpenDover, since the vocabulary they contain conveys neither positive nor negative sentiments. A rec- ommended solution to this problem might involve integrating into the architecture a subjectivity classifier, like the one proposed in (Barbosa & Feng, 2010). Such a classifier can distinguish subjective from objective tweets, isolating the advertisements and offering an additional filter to the set of tweets to be processed. Another critical downside when working with Twitter and mi- cro-blogging services in general involves the extensive use of jar- gon in the published posts. This is an issue that unavoidably occurs partly due to the imposed character limit to posts, but can also be attributed to the every-day communication nature of mi- cro-blogs. The latter are not suitable for posting long pieces of text or fully justified opinions on a matter. A further difficulty was out- lined in a previous section and involves the use of OpenDover as a third-party sentiment analysis service. The fact that the system is not open-source constitutes a drawback, since it is not possible to analyze the software, or even attempt to improve its efficiency. Therefore, one has to ‘‘blindly’’ rely on the derived sentiment scores. Nevertheless, an imminent solution to this issue involves integrating a custom-built sentiment analysis tool to the architec- ture of the proposed system, which constitutes one of our goals for future improvements. 5. Conclusions and future work The paper argued that sentiment analysis constitutes a rapidly evolving research area, especially since the emergence of Web 2.0 and its related technologies (social networks, blogs, wikis etc.) The recent explosion in the usage of micro-blogging services, and particularly Twitter, has shifted attention to sentiment analysis of micro-blogging posts and tweets. There exist various machine learning-based approaches that perform sentiment analysis on tweets, with the drawback that they treat each tweet as one uni- form statement and assigning a sentiment score to the post as a whole. This paper proposes the deployment of ontology-based techniques for determining the subjects discussed in tweets and breaking down each tweet into a set of aspects relevant to the sub- ject. The result is the assignment of a sentiment score to each dis- tinct aspect. A baseline scenario is also presented that deals with the domain of a popular product (smartphones) and results in com- paratively evaluating the distinct features of each model series. E. Kontopoulos et al. / Expert Systems with Applications 40 (2013) 4065–4074 4073 Our goals for future improvements of the proposed approach initially involve the integration of a custom-built sentiment classi- fier that will substitute OpenDover in our architecture. A further aim is to integrate a fully automatic ontology-building functional- ity, potentially through a combination of ontology learning tech- niques. Nevertheless, keeping the manual and semi-automatic ontology creation approaches still remains useful, as they offer a more controlled means for building the domain vocabulary. Exploring various methods for visualizing the resulting sentiment is an additional direction for providing more thorough information to the user. Finally, provided that investors (or consumers) are prone to exogenous sentiment waves, an interesting research direction would be the development of time-series sentiment in- dexes for a range of investment (or consumer) goods. In case these developed indexes contain useful predictive power with respect each time to the future price movements of the investigated good, they may act as valuable tools in forming efficient strategies for all market participants. Acknowledgements The present scientific paper was executed in the context of the project entitled ‘‘International Hellenic University (Operation – Development)’’, which is part of the Operational Programme ‘‘Edu- cation and Lifelong Learning’’ of the Ministry of Education, Lifelong Learning and Religious affairs and is funded by the European Com- mission (European Social Fund – ESF) and from national resources. Additionally, the authors would like to thank Dr. Ioannis Katakis for his valuable comments on the paper. References Agarwal, A., Xie, B., Vovsha, I., Rambow, O., & Passonneau, R. (2011). Sentiment analysis of twitter data. In Proceedings of the ACL 2011 workshop on languages in social (pp. 30–38). Media. Aue, A., & Gamon, M. (2005). Customizing sentiment classifiers to new domains: a case study. In Proceedings of the international conference on recent advances in natural language processing (RANLP-05) (pp. 207–218). Baldoni, M., Baroglio, C., Patti, V., & Rena, P. (2012). From tags to emotions: Ontology-driven sentiment analysis in the social semantic web. Intelligenza Artificiale, 6(1), 41–54. Barbosa, L., & Feng, J. (2010). Robust sentiment detection on twitter from biased and noisy data. In Proceedings of the 23rd international conference on computational linguistics: Posters (COLING ‘10) (pp. 36–44). Stroudsburg, PA, USA: Association for Computational Linguistics. Berners-Lee, T., Hendler, J., & Lassila, O. (2001). The semantic web. Scientific American, 284(5), 34–43. Breslin, J. G., Harth, A., Bojars, U., & Decker, S. (2005). Towards semantically- interlinked online communities. In 2nd European semantic web conference (ESWC 2005) (pp. 500–514). May 29-June 1. Brickley, D., & Miller, L. (2010). FOAF Vocabulary Specification 0.98. Namespace Document 9 August 2010 – Marco Polo Edition, FOAF Project. Available from: <http://xmlns.com/foaf/0.1/>, last access: November 2012. Cimiano, P., & Völker, J. (2005). Text2Onto – a framework for ontology learning and data-driven change discovery. In 10th International conference on applications of natural language to information systems (NLDB) (pp. 227–238). LNCS. Claster, W. B., Cooper, M., & Sallis, P. (2010). Thailand – tourism and conflict: Modeling sentiment from twitter tweets using Naïve Bayes and unsupervised artificial neural nets. In Proceedings of the CIMSIM’10 (pp. 89–94). Washington, DC, USA: IEEE Computer Society. Davey, B. A., & Priestley, H. A. (2002). Introduction to lattices and order. Cambridge University Press. ISBN 978-0-521-78451-1. De Kok, D., & Brouwer, H. (2012). Natural language processing for the working programmer. E-book available under the creative commons attribution 3.0 License (CC-BY). 2011, Available from: <http://www.nlpwp.org/nlpwp.pdf>, last access: November 2012. Dergiades, T. (2012). Do investors’ sentiment dynamics affect stock returns? Evidence from the US economy. Economics Letters, 116(3), 404–407. Fortuna, B., Grobelnik, M., & Mladenic, D. (2007). OntoGen: Semi-automatic ontology editor. In M. J. Smith & G. Salvendy (Eds.), Proceedings of the 2007 conference on human interface (pp. 309–318). Berlin, Heidelberg: Springer- Verlag. Francisco, V., Germans, P., & Peinado, F. (2007). Ontological reasoning to configure emotional voice synthesis. In Proceedings of the web reasoning and rule systems (RR 2007) (pp. 88–102). Springer, LNCS 4524. Fu, G., & Cohn, A. G. (2008). Utility ontology development with formal concept analysis. In C. Eschenbach & M. Grüninger (Eds.), Proceedings of the 2008 conference on formal ontology in information systems (FOIS 2008) (pp. 297–310). Amsterdam, The Netherlands: IOS Press. Ganter, B. B., & Wille, R. (1999). Formal concept analysis, mathematical foundation. Berlin: Springer Verlag. Gómez-Pérez, A., & Corcho, O. (2002). Ontology languages for the semantic web. IEEE Intelligent Systems, 17(1), 54–60. Harding, D., & Pagan, A. (2002). Dissecting the Cycle: A Methodological Investigation. Journal of Monetary Economics, 49(2), 365–381. Hazman, M., El-Beltagy, S. R., & Rafea, A. (2009). Ontology learning from domain specific web documents. International Journal of Metadata, Semantics and Ontologies, 4(1/2), 24–33. He, Y., & Alani, H. (2011). Semantic smoothing for twitter sentiment analysis. In Proceedings of the10th international semantic web conference (ISWC). demo/ poster session. Iwanaga, I., Nguyen, T. M., Kawamura, T., Nakagawa, H., Tahara, Y., & Ohsuga, A. (2011). Building an earthquake evacuation ontology from twitter. In Proceedings of the IEEE international conference on granular computing (GrC) (pp. 306–311). 8–10 November. Kaji, N., & Kitsuregawa, M. (2007). Building lexicon for sentiment analysis from massive collection of HTML documents. In Proceedings of the joint conference on empirical methods in natural language processing and computational natural language learning (EMNLP-CoNLL) (pp. 1075–1083). Kang, H., Yoo, S., & Han, D. (2012). Senti-lexicon and improved Naïve Bayes algorithms for sentiment analysis of restaurant reviews. Expert Systems with Applications, 39(5), 6000–6010. Kaplan, A. M., & Haenlein, M. (2011). The early bird catches the news: Nine things you should know about micro-blogging. Business Horizons, 54(2), 105–113. Kouloumpis, E., Wilson, T., & Moore, J. (2011). Twitter sentiment analysis: The good the bad and the OMG!. In Proceedings of the ICWSM. Liu, Y., Huang, X., An, A., & Yu, X. (2007). ARSA: A sentiment-aware model for predicting sales performance using blogs. In Proceedings of the 30th annual international ACM SIGIR conference on research and development in information retrieval (pp. 607–614). Amsterdam, The Netherlands. July 23–27. Melville, P., Gryc, W., & Lawrence, R. D. (2009). Sentiment analysis of blogs by combining lexical knowledge with text classification. In Proceedings of the 15th ACM SIGKDD international conference on knowledge discovery and data mining (KDD ‘09) (pp. 1275–1284). New York, NY, USA: ACM. Miller, G. A. (1995). WordNet: A lexical database for English. Communications of the ACM, 38(11), 39–41. Moreo, A., Romero, M., Castro, J. L., & Zurita, J. M. (2012). Lexicon-based comments- oriented news sentiment analyzer system. Expert Systems with Applications, 39(10), 9166–9180. Neviarouskaya, A., Prendinger, H., & Ishizuka, M. (2009). SentiFul: Generating a reliable lexicon for sentiment analysis. In Proceedings of the affective computing and intelligent interaction and workshops (ACII 2009), 3rd international conference on affective computing and intelligent interaction and workshops (pp. 10–12). IEEE. September 1–6. Ning, L., Guanyu, L., & Li, S. (2010). Using formal concept analysis for maritime ontology building. In Proceedings of the 2010 international forum on information technology and applications (IFITA ‘10) (pp. 159–162). New York, USA: Springer. Vol. 2. Obitko, M., Snasel, V., Smid, J., & Snasel, V. (2004). Ontology design with formal concept analysis. In V. Snasel & R. Belohlavek (Eds.), Concept lattices and their applications (pp. 111–119). Ostrava: Czech Republic. Pak, A., & Paroubek, P. (2010). Twitter as a corpus for sentiment analysis and opinion mining. In Proceedings of the 7th international conference language resources and evaluation (LREC ‘10). European Language Resources Association ELRA. Pang, B., & Lee, L. (2008). Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2(1–2), 1–135. Park, S., Ko, M., Kim, J., Liu, Y., & Song, J. (2011). The politics of comments: Predicting political orientation of news stories with commenters sentiment patterns. In Proceedings of the ACM 2011 conference on computer supported cooperative work (CSCW ‘11) (pp. 113–122). New York, NY, USA: ACM. Passant, A., Bojars, U., Breslin, J. G., Hastrup, T., Stankovic, M., & Laublet, P. (2010). An overview of SMOB 2: Open, semantic and distributed microblogging. In 4th International conference on weblogs and social media (pp. 303–306). ICWSM. Prabowo, R., & Thelwall, M. (2009). Sentiment analysis: A combined approach. Journal of Informetrics, 3(2), 143–157. Rajagopal, H. (2005). JENA: A Java API for ontology management. Colorado software summit: October 23–28, 2005 � Copyright 2005, IBM Corporation, Available from: <http://goo.gl/QMaQC>, last access: November 2012. Read, J. (2005). Using emoticons to reduce dependency in machine learning techniques for sentiment classification. In Proceedings of the ACL-05, 43nd meeting of the association for computational linguistics. Association for Computational Linguistics. Saif, H., He, Y., & Alani, H. (2012). Alleviating data sparsity for twitter sentiment analysis. In 2nd Workshop on making sense of microposts (#MSM2012): Big things come in small packages at World Wide Web (WWW) 2012 (pp. 2–9). Lyon, France. Skiba, D. J. (2006). Web 2.0: Next great thing or just marketing hype? Nursing Education Perspectives, 27(4), 212–214. Stankovic, M., Passant, A., & Laublet, P. (2009). Status messages for the right audience with an ontology-based approach. In 1st International workshop on collaborative social networks – collaboratesn 2009 – at the 5th international conference on collaborative (pp. 1–7). Computing. http://www.xmlns.com/foaf/0.1/ http://www.nlpwp.org/nlpwp.pdf http://www.goo.gl/QMaQC 4074 E. Kontopoulos et al. / Expert Systems with Applications 40 (2013) 4065–4074 Studer, R., Benjamins, R., & Fensel, D. (1998). Knowledge engineering: Principles and methods. IEEE Trans on Data and Knowledge Eng., 25(1–2), 161–197. Taboada, M., Brooke, J., Tofiloski, M., Voll, K., & Stede, M. (2011). Lexicon-based methods for sentiment analysis. Computational Linguistics, 37(2), 267–307. Tumasjan, A., Sprenger, T. O., Sandner, P. G., & Welpe, I. M. (2010). Predicting elections with twitter: What 140 characters reveal about political sentiment. In Proceedings 4th international AAAI conference on weblogs and social (pp. 178– 185). Media. Wanner, F., Rohrdantz, C., Mansmann, F., Oelke, D., & Keim, D. A. (2009). Visual sentiment analysis of RSS news feeds featuring the US presidential election in 2008. In Workshop on visual interfaces to the social and the semantic web (VISSW) (pp. 1–8). Yevtushenko, S. A. (2000). System of data analysis concept explorer. In Proceedings of the 7th national conference on artificial intelligence KII-2000 (pp. 127–134). Russia. (in Russian). Zhang, R., & Xu, H. (2011). Building the ontology system in semantic web based on formal concept analysis and rough set. Journal of Convergence Information Technology, 6(7), 56–62. Ontology-based sentiment analysis of twitter posts 1 Introduction 2 Sentiment analysis in micro-blogging data 3 Micro-blogging and ontologies 4 Description of the proposed approach 4.1 Creating the domain ontology 4.1.1 Formal Concept Analysis 4.1.1.1 FCA basic elements 4.1.1.2 Populating the concept cross-table 4.1.2 Ontology learning 4.1.3 Augmenting the semantics 4.2 Sentiment analysis on tweets 4.2.1 Step#1: Taking advantage of the ontology 4.2.2 Step #2: Retrieving the relevant tweets 4.2.3 Step #3: Sentiment analysis 4.3 Baseline scenario 4.4 Evaluation 4.5 Difficulties 5 Conclusions and future work Acknowledgements References 
work_ope6navdjndnnebvcd4hv3hj44	----	A comparative study of approaches to forecast the correct trading actions Received: 7 November 2015 Revised: 7 April 2016 Accepted: 8 July 2016 DO I 10.1111/exsy.12169 A R T I C L E A comparative study of approaches to forecast the correct trading actions Luís Baía1,2 | Luís Torgo1,2 1 LIAAD ‐ INESC TEC, Porto, Portugal 2 Departamento de Ciência de Computadores‐ Faculdade de Ciências, Universidade do Porto, Porto, Portugal Correspondence Luís Baía, LIAAD‐INESC TEC, Porto, Portugal. Email: luisbaia_1992@hotmail.com Expert Systems 2017;34:e12169. w https://doi.org/10.1111/exsy.12169 Abstract This paper addresses the problem of decision making in the context of financial markets, more specifically, the problem of forecasting the correct trading action for a certain future horizon. We study and compare two alternative ways of addressing these forecasting tasks: (a) using standard numeric prediction models to forecast the variation on the prices of the target asset and, on a second stage, transform these numeric predictions into a decision according to some predefined decision rules; and (b) use models that directly forecast the right decision thus ignoring the intermediate numeric forecasting task. The objective of our study is to determine if both strategies provide identical results or if there is any particular advantage worth being considered that may distinguish each alternative in the context of financial markets. KEYWORDS classification, extensive experimental comparison, forecast, regression, trading actions 1 | INTRODUCTION Many real‐world applications require decisions to be made based on forecasting some numeric quantity. Sales forecasting may lead to some important decisions concerning the production process. Asset price forecasting may lead investors to buy or sell some financial product. Forecasting the future evolution of some indicator of a patient may lead a medical doctor to prescribe some important treatments. These are just a few examples of concrete applications that fit this general setting: decisions based on numeric forecasts of some variable. Frequently, the decision process is based on a predefined protocol that associates inter- vals of the range of the numeric variable with concrete actions/decisions. This means that once we have a prediction for the numeric variable, we use some deterministic process to reach the action/decision to be made. In spite of the generality of this type of applications, this work is focused on analyzing them in the context of financial markets. In this domain the goal of investors is to make the correct trading decision (Sell, Buy, or Hold) at any given point in time. These decisions are made based on the investor’s expectations on the future evolution of the asset prices. In this work, we approach this decision problem using prediction models. More specifically, we will compare two possible ways of trying to forecast what is the correct trading decision at any point in time. In our target applications, we assume that there are determin- istic decision rules that given the estimated evolution of the prices of the asset will indicate the trading action to be taken. These rules ileyonlinelibrary.com/journal/exsy are typically driven by the investor’s preferences concerning impor- tant aspects like financial risk. For instance, a rule could state that if the forecast of the variation of prices is a 2.5% increase then the correct decision is to buy the asset as this will allow covering trans- action costs and still have some profit. Given the deterministic mapping from forecasted values into decisions, we can define the prediction task in two ways. The first consists on obtaining a numeric prediction model that we can use to obtain predictions of the future variation of the prices, which are then transformed (deterministically) into trading decisions (e.g., Hellstrom (1999); Lu, Lee, and Chiu (2009)). The second alternative consists of directly forecasting the correct trading decisions (e.g., Luo and Chen (2013); Ma, Song, Hung, Su, and Huang (2012); Teixeira and de Oliveira (2010)). Which is the best option in terms of financial results? To the best of our knowledge, no comparative study was carried out to answer this question. This is the goal and the main contribution of this paper: to compare these two approaches to decision making and provide experimental evidence of the advantages and disadvantages of each alternative. 2 | PROBLEM FORMALIZATION The problem of decision making based on forecasts of a numerical (continuous) value can be formalized as follows. We assume there is Copyright © 2016 John Wiley & Sons, Ltd. 1 of 13 https://doi.org/10.1111/exsy.12169 http://wileyonlinelibrary.com/journal/exsy https://doi.org/10.1111/exsy.12169 2 of 13 BAÍA AND TORGO an unknown function that maps the values of p predictor variables into the values of a certain numeric variable Y. Let f be this unknown function that receives as input a vector x with the values of the p predictors and returns the value of the target numeric variableY whose values are supposed to depend on these predictors, f : ℝp→ℝ x↦f xð Þ: We also assume that based on the values of this variable Y some decisions need to be made. Let g be another function that given the values of this target numeric variable transforms them into actions/ decisions, g : ℝ→A ¼ α1; α2; α3…f g Y↦g Yð Þ: where A represents a set of possible actions. In our target applications, functions f and g are very different. Function g is known and deterministic, in the sense that it is part of the domain background knowledge. Function f is unknown and uncertain. The only information we have about function f is a histor- ical record of mappings from x into Y, that is, a data set that can be used to learn an approximation of the function f. Given that the variable Y is numeric, this approximation could be obtained using some existing multiple regression tools. This means that given a data set Dr ¼ xi; Yih ini¼1 � � , we can use some regression tool to obtain a model r ̂ xð Þ that is an approximation of f. From an operational perspective, this would mean that given a test case q for which a decision needs to be made, we would proceed by first using r ̂ to obtain a prediction for Y and then apply g to this predicted value to get the predicted action/decision, that is, q↦ r ̂ qð Þ↦g r ̂ qð Þð Þ. In the con- text of financial markets, the predictors describe the currently observed dynamics of the prices of some financial asset, and the target numeric variable Y represents the future variation of this price. This means that f is the unknown function that maps the currently observed price dynamics into a future evolution of the price. On the other hand, g is a deterministic function (typically based on domain knowledge and risk preferences of traders) that maps the prediction of the future evolution of prices into one of three possible decisions: Sell, Hold, or Buy. Given the deterministic nature of g, we can use an alternative process for obtaining decisions. More specifically, we can build an alternative data set Dc ¼ xi; g Yð Þh ini¼1 � � , where the target variable is the decision associated with each known Y value in the historical record of data. This means that we have a nominal target variable, that is, we are facing a classification task. Once again, we can use some standard classification tool to obtain an approximation bc of the unknown function that maps the predictors into the correct actions/ decisions. Once such model is obtained, we can use it given a query case q to directly estimate the correct decision by applying the learned model to the case, that is, q↦ ĉ qð Þ . This means that given the description of the current dynamics of the price, we will use function ĉ to forecast directly the correct trading action for this context. Independently, of the approach followed, the final goal of the applications we are targeting is always to make correct decisions. This means that whatever process we use to reach a decision, it will be evaluated in terms of the “quality” of the decisions it generates. In this context, it seems that the classification approach, by having as target variable the decisions, would be easier to bias towards optimal actions. However, this approach completely ignores the intermediate numeric variable that is supposed to influence decisions, though one may argue that information on the relationship between Y and the decisions is “encoded” when building the training set Dc by using as target the values of g(Yi). On the other hand, although the regression approach is focused on obtaining accurate predictions of Y, it completely ignores questions like eventual different cost/benefits of the different possible decisions that could be easily encoded into the classification tasks. All these potential trade‐offs motivate the current study. The main goal of this paper is to compare these two approaches in the context of financial markets. 3 | MATERIAL AND METHODS This section describes the main issues involved in the experimental comparison we will carry out with the goal of comparing the two possible approaches described in the previous section. 3.1 | The tasks The problem addressed in this paper is very common in automatic trading systems where decisions are based on the forecasts of some prediction models. The decisions to open or close short/long positions are typically the result of a deterministic mapping from the predicted prices variation. In our experiments, we have used the assets prices of 12 companies. Each data set has a minimum of 7 years of daily data and a maximum of 30 years. In order to simplify the study, we will be work- ing with a 1‐day horizon, that is, take a decision based on the forecasts of the assets variation for 1 day ahead. Moreover, we will be working exclusively with the closing prices of each trading session, that is, we assume trading decisions are to be made after the markets close. The decision function for this application receives as input the forecast of the daily variation of the assets closing prices and returns a trading action. We will be using the following function in our experiments: g : ℝ→A ¼ hold; buy; sellf g Y↦ buy; Y>0:02 sell; Y<−0:02 hold; other cases 8>>< >>: : This means we are assuming that any variation above 2% will be sufficient to cover the transaction costs and still obtain some profit. Concerning the data that will be used as predictors for the forecasting models (either forecasting the prices variation [Y] or directly the trading action [A]), we have used the price variations on recent days as well as some trading indicators, such as the annual volatility, the Welles Wilder style moving average (Wilder, 1978), the stop and reverse point indicator developed by J. Welles Wilder (Wilder, 1978), the usual moving average, and others. The goal of this selection of predictors is to provide the forecasting models with useful information on the recent dynamics of the assets prices. BAÍA AND TORGO 3 of 13 Regarding the performance metrics, we will use to compare each approach; we will use two metrics that capture important properties of the economic results of the trading decisions made by the alterna- tive models. More specifically, we will use the Sharpe Ratio as a mea- sure of the risk (volatility) associated with the decisions and the percentage total return as a measure of the overall financial results of these actions. We will also consider the Macro versions of the Recall and Precision metrics for the buy and sell signals combined. This will allow us to analyze how many trading opportunities are being detected (Recall) and how accurate are the models regarding the prediction of every non‐hold trading action (Precision), respectively. To make our experiments more realistic, we will consider a transaction cost of 2% for each buy or sell decision a model may originate. At this stage, it is important to remark that the prediction tasks we are facing have some characteristics that make them particularly challenging. One of the main hurdles results from the fact that inter- esting events, from a trading perspective, are rare in financial mar- kets. In effect, large movements of prices are not very frequent. This means that the data sets we will provide to the models have clearly imbalanced distributions of the target variables (both the numeric percentage variations and the trading actions). To make this imbalance problem harder, the situations that are more interesting from a trading perspective are rare in the data sets, which creates difficulties to most modeling techniques. In the next section, we will describe some of the measures we have taken to alleviate this prob- lem. We have conducted a thorough analysis to test the impact of these measures on each approach. Evaluating these existing TABLE 1 Regression models used for the experimental comparisons Model Variants SVM cost = {1,5,10}, ε = {0.1,0.05,0.01}, tolerance = {0.001,0.005},kernel = linear SVM cost = {1,10}, ε = {0.1,0.05,0.01}, degree = {2,3,5},kern Random Forest ntree = {500,750,1000,2000,3000},mtry = {4,5,6} Trees (pruned) se = {0,0.5,1,1.5,2},cp = 0, minsplit = 6 KNN k = {1,3,5,7,11,15} NNET size = {2,4,6},decay = {0.05,0.1,0.15} MARS thresh = {0.001,0.0005,0.002}, degree = {1,2,3},minspa AdaBoost dist = {gaussian},n.trees = {10000,20000}, shrinkage = {0.001,0.01},interaction.depth = {1,2)} SVM = support vectorial machines; KNN = K‐nearest neighbours; NNET = neur TABLE 2 Classification models used for the experimental comparisons Model Variants SVM cost = {1,3,7,10},kernel = linear tolerance = {0.001,0.005 SVM cost = {1,10}, ε = {0.1,0.05}, degree = {2,3,4,5},kernel = p Random Forest ntree = {500,750,1000,2000,3000},mtry = {3,4,5} Trees (pruned) se = {0,0.5,1,1.5,2},cp = 0, minsplit = 6 KNN k = {1,3,5,7,11,15} NNET size = {2,4,6},decay = {0.05,0.1,0.15} AdaBoost coeflearn = c(‘Breiman’,‘Freund’,‘Zhu’), mfinal = c(500,10 SVM = support vectorial machines; KNN = K‐nearest neighbours; NNET = neur techniques in the context of financial forecasting problems is the sec- ond main contribution of the paper. 3.2 | The models In this section, we describe all the model variants that will be used in the experimental comparisons. They were selected to ensure that both approaches have the same conditions for a fair comparison and that the results are not biased by some specific characteristics of one particular technique. Several variants for each family of models (SVM [support vectorial machines], Random Forests, etc.) were tested in order to make sure our conclusions were not biased. Tables 1 and 2 show all the model variants used in our experiments (nearly 182 model variants were tested). To facilitate the reproduc- ibility of our results, we have used the free and open source implementations of these techniques available in the R software environment (R Core Team, 2014). The predictive tasks we are facing have two main difficulties: (a) the fact that the distribution of the target variables is highly imbal- anced, with the more relevant values being less frequent; and (b) the fact that there is an implicit ordering among the decisions. The first problem causes most modeling techniques to focus on cases (the most frequent) that are not relevant for the application goals. The second problem is specific to classification tasks as these algorithms do not distinguish among the different types of errors, whilst in our target application, confusing a buy decision with a hold decision is less seri- ous than confusing it with a sell. R package e1071 ‐ Meyer, Dimitriadou, Hornik, Weingessel, and Leisch (2014) el = polynomial e1071 ‐ Meyer et al. (2014) randomForest ‐ Liaw and Wiener (2002) DMwR – Torgo (2010) DMwR ‐ Torgo (2010) Nnet ‐ Venables and Ripley (2002) n = {0,1} Earth – Milborrow (2014) Gbm – Ridgeway (2013) al networks; MARS = multivariate adaptive regression spline. R package ,0.0005,0.002} e1071 ‐ Meyer et al. (2014) olynomial e1071 ‐ Meyer et al. (2014) randomForest ‐ Liaw and Wiener (2002) DMwR – Torgo (2010) DMwR – Torgo (2010) Nnet ‐ Venables and Ripley (2002) 00,2000) Boosting ‐ Ridgeway, Southworth, and RUnit (2013) al networks; MARS = multivariate adaptive regression spline. 4 of 13 BAÍA AND TORGO These two problems led us to consider several alternatives to our base modeling approaches described in Tables 1 and 2. For the first problem of imbalance, we have considered the hypothesis of using resampling to balance the distribution of the target variable before obtaining the models. In order to do that, we have used the SMOTE algorithm (Chawla, Bowyer, Hall, & Kegelmeyer, 2002). This method is well‐known for classification models, consisting basically of oversampling the minority classes and under‐sampling the majority ones. The goal is to modify the data set in order to ensure that each class is similarly represented. Regarding the regression tasks, we have used the work by Torgo, Branco, Ribeiro, and Pfahringer (2015), where a regression version of SMOTE was presented. Essentially, the concept is the same as in classification, using a method to try to balance the continuous distribution of the target variable by oversampling and under‐sampling different ranges of its domain. Regarding the second problem of the order among the classes, we have also considered a frequently used approach to handle this issue. Namely, we have used a cost–benefit matrix that allows to distinguish between the different types of classification errors. Using this matrix, and given a probabilistic classifier, we can predict for each test case the class that maximizes the utility instead of the class that has the highest probability. We have used the following procedure to obtain the cost–benefit matrices for our tasks. Correctly predicted buy/sell signals have a pos- itive benefit estimated as the average return of the buy/sell signals in the training set. On the other hand, in the case of incorrectly predicting a true hold signal as buy (or sell), we assign it minus the average return of the buy (or sell) signals. Basically, the benefit associated to correctly predicting one rare signal is entirely lost when the model suggests an investment when the correct action would be doing nothing. In the extreme case of confusing the buy and sell signals, the penalty will be minus the sum of the average return of each signal. Choosing such a high penalty for these cases will eventually change the model to be less likely to make this type of very dangerous mistakes. Considering the case of incorrectly predicting a true sell (or buy) signal as hold, we also charge for it but in a less severe way. Therefore, the average of the sell (or buy) signal is considered, but divided by two. This division was our way of “teaching” the model that it is preferable to miss an opportunity to earn money rather than making the investor lose money. Finally, correctly predicting a hold signal gives no penalty nor reward, because no money is either won or lost. Table 3 shows an example of such cost–benefit matrix that was obtained with the data from 1981‐01‐ 05 to 2000‐10‐13 of Apple. We have thoroughly tested the hypothesis that using resampling before obtaining the models would boost the performance of the dif- ferent models we have considered for our tasks (both classification TABLE 3 Example cost–benefit matrix for Apple shares Trues S H B S 0.49 −0.49 −0.82 Pred H −0.24 0.00 −0.17 B −0.82 −0.33 0.33 S = share; H = hold; B = buy. and regression) and have also tested the hypothesis that using cost– benefit matrices to implement utility maximization would also improve the performance of the classification models. The results of testing these hypotheses will be presented in Section 4. 3.3 | The experimental methodology In this section, we present the experimental methodology used in our comparative experiments. Because of the temporal nature of the data sets, the usual cross‐validation methodology should not be used to estimate the performance of a certain model. This procedure involves randomly reshuffling the data, which may lead to test cases that are “older” than the training cases, which would lead to unreliable estimates. In this context, we have used a Monte Carlo simulation method consisting of randomly selecting a series of N points in time within the available data set. For each of these random dates, we use a certain consecutive past window as training set for obtaining the alternative models that are then tested/compared in a subsequent and consecutive test window. The Monte Carlo estimates are formed by the average scores obtained on the N repetitions. In our experi- ments, we have used N = 10, 50% of the data as the size of the training window, and 25% of the data as size of the test sets. With respect to testing the statistical significance of the observed differences between the estimated scores, we have used the recommendations of the work by Demšar (2006). More specifically, in situations where we are comparing k alternative models on one specific task, we have used the Wilcoxon signed‐rank test to check the significance of the differences. On the experiments where k models are compared on t tasks, we use the Friedman test followed by a post hoc Nemenyi test to check the significance of the difference between the average ranks of the k models across the t tasks. 4 | EXPERIMENTAL RESULTS This section describes the results of our empirical studies. We have split these results in two main parts. The first has to do with testing the validity of the hypotheses described in Section 3.2 concerning the usage of resampling techniques and also the usage of cost–benefit matrices. The second part is the results of the final comparison between the two modeling approaches to our target decision problems. 4.1 | Addressing the particularities of the prediction tasks As we have mentioned before, the prediction tasks we are facing have some particularities that turn them into particularly challenging problems. We have described two ways of trying to overcome these challenges. In this section, we check the validity of these hypotheses. Specifically, we test the advantages of: (a) using SMOTE on classification and regression models to overcome the problem of imbalanced distributions; and (b) using cost–benefit matrices on classification models to provide information on the different importance of the class values. BAÍA AND TORGO 5 of 13 4.1.1 | Hypothesis 1: Resampling the data sets Our first hypothesis states that resampling the data sets with the goal of balancing the response variable will enhance the performance of both classification and regression models. In the experiments carried out to check the validity of this hypothesis, we have observed that the way the resampling affected each modeling approach was not significantly different. In this context, we will just present the results for the classification models as the conclusions with regression models were similar. We start by comparing the top modeling variant obtained without using resampling against the best one using SMOTE, company by com- pany, applying a Wilcoxon test in each case (for each modeling approach individually). In Figure 1 we analyze recall and precision for the buy and sell signals combined. The results were somewhat expected. Resampling methods balance more the distribution of the target, which means the models will have more examples of the rare cases that are interesting in this application. This has a positive impact on recall because the models end up forecasting more frequently these events because they are not so rare in the re‐sampled training sets. However, due to the inherent difficulty of this financial forecasting tasks, this also involves a higher risk of making wrong predictions and thus the decrease in precision. Because we are in a trading context, we should prefer safer decisions rather than riskier ones. However, only after checking the financial results of these strategies we can confirm this. Nevertheless, we should reinforce that the issue of preferring high precision (less risk) over high recall is something specific to financial trading, and in other application domains, the conclusions on the usefulness of resampling could be different if the preference bias is different. FIGURE 1 Best classification variant without SMOTE against the best classif rics (asterisks denote that the respective variant is significantly better, acco Figure 2 allows us to analyze the financial consequences of the resampling procedure. We can observe that in terms of Total Return and Sharpe Ratio, the models applied to data sets without any resampling achieved significantly better results and, in several cases, by a large margin. This means that pre‐processing the data using resampling is leading to models that make more risky decisions and that these decisions are often wrong, leading to serious financial losses. Considering all the plots at the same time, our results provide evidence that using SMOTE on the training data will make the models predict significantly more trading signals (Buy and Sell). This leads to higher recall but unfortunately also to much lower precision because these signals are frequently wrong, with serious financial consequences. However, we should not forget that the previous comparisons were carried out between the best overall variant of each type (with and without SMOTE). A different question is whether the same conclusions are valid if we consider each modeling technique individually. In this context, we will now check the resampling hypothesis per type of model instead of globally. We will group the variants according to the type of model and analyze the impact of using SMOTE per type of model. The results of this new set of comparisons are shown in Figure 3. The first thing to notice is that the recall was significantly better every single time across all types of models where resampling was applied. On the other hand, the precision was worse almost every time, except with SVMs. So we again observe that the use of the SMOTE algorithm is making the models more capable of detecting the buy and sell (rare) classes at the cost of riskier decisions (except for SMVs). Concerning the financial metrics, we observe a clear ication with SMOTE for the macro versions of Precision and Recall met- rding to a Wilcoxon test with α = 0.05) FIGURE 2 Best classification variant without SMOTE against the best classification with SMOTE for the Total Return and Annualized Sharpe Ratio metrics (asterisks denote that the respective variant is significantly better, according to a Wilcoxon test with α = 0.05) FIGURE 3 Segmented by type of model and by metric, a Wilcoxon test is performed between the best model variant of each modeling tool (Classification without SMOTE versus Classification with SMOTE). Because there are 12 datasets in each segment, then there are 12 results per segment. Each bar shows the results for each segment, where each one of the four colors is associated to a type of win (significant/non‐significant win without SMOTE ‐ strong/light blue, non‐ significant/significant win with SMOTE ‐ light/ strong green). The length of each color describes the number of times that type of win occurred 6 of 13 BAÍA AND TORGO advantage of the standard approach (i.e., no resampling applied), with SVM and AdaBoost being the only algorithms benefiting from the use of resampling on some tasks, with the advantage being more evident for SVMs. In summary, we have collected sufficient evidence to conclude that resampling the data sets for both classification and regression models will have a negative impact on their performance in the context of financial forecasting tasks, namely when considering BAÍA AND TORGO 7 of 13 trading‐related evaluation metrics. Only the SVM, TREE, and AdaBoost algorithms have benefited from the resampling method on a small set of tasks. Overall, our conclusion is that this resampling strategy is not recommended in the context of forecast- ing for financial trading. 4.1.2 | Hypothesis 2: adding cost–benefit matrices The second hypothesis we have put forward was that the use of cost–benefit matrices would boost the performance of the classifi- cation models, because the information on the implicit ordering among the classes would be passed to the models, allowing them to avoid more costly errors (e.g., confusing a buy decision with a sell decision). In order to test this hypothesis, we will follow the same methodol- ogy used for the resampling hypothesis. Firstly, the top modeling variant of each alternative (with and without cost–benefit matrices) will be compared for each data set. The results regarding precision and recall are shown in Figure 4. These results are a bit inconclusive. In terms of recall the use of these matrices led to better results, as we have five significant wins against only one significant loss of the approach using the matrices. On the other hand, in terms of precision, we observed three significant wins of the approach without cost–benefit matrices. Figure 5 shows the results of this same experiment in terms of the financial metrics. The most evident observation is the fact that not a single statistically significant difference was achieved by either alternative. However, both in terms of the Total Return FIGURE 4 Best classification variant without costs against the best classific (asterisks denote that the respective variant is significantly better, accordin and of the Annualized Sharpe Ratio, the approaches using cost– benefit matrices obtained more wins. This is an interesting result, suggesting that the use of these matrices may be beneficial for the classification models in the context of financial trading based on prediction models. Let us now check if these results hold across each different type of learning algorithm. Figure 6 shows the comparison between the top modeling variant of each type of model. Even though there is a slightly higher abundance of green (suggesting that the usage of cost–benefit matrices may be beneficial), these results are quite even. Each type of model is influenced in a different way by the cost–benefit matrices. Although NNET (neural networks), TREE (decision tree models) and SVM are able to take advantage of the information on the matrices, the conclusions for the other models are not so clear. These observations seem to indicate that the poten- tial advantage of the usage of cost–benefit matrices is algorithm‐ dependent, with some techniques being able to capitalize on this extra information while others do not. As future research agenda it, should be interesting to understand if there are some theoretical properties of the algorithms that may be causing this differentiated behavior. A possible explanation has to do with the quality of proba- bility estimates. In effect, the decisions using cost–benefit matrices depend on the estimated probabilities of each class. If these esti- mates are wrong or unreliable, this may lead to unreliable decisions. Still, this explanation needs to be confirmed in practice. In summary, contrary to the first hypothesis involving resampling, the conclusions for the second hypothesis regarding the advantages of ation with costs for the macro versions of Precision and Recall metrics g to a Wilcoxon test with α = 0.05) FIGURE 5 Best classification variant without costs against the best classification with costs for the Total Return and Annualized Sharpe Ratio metrics (asterisks denote that the respective variant is significantly better, according to a Wilcoxon test with α = 0.05) FIGURE 6 Segmented by type of model and by metric, a Wilcoxon test is performed between the best model variant of each modeling tool (Classification without costs versus Classification with costs). Because there are 12 datasets in each segment, then there are 12 results per segment. Each bar shows the results for each segment, where each one of the five colors is associated to a type of win (significant/non‐significant win without costs ‐ strong/light blue, draw ‐ yellow, non‐significant/significant win with costs ‐ light/strong green). The length of each color describes the number of times that type of win occurred 8 of 13 BAÍA AND TORGO cost–benefit matrices are that there is some potential for this alterna- tive way of addressing the classification tasks. Although we have not observed an overwhelming advantage of this usage, we have found that for some modeling algorithms, these matrices provide a clear boost in terms of their results (both in terms of Precision/Recall and financially‐oriented metrics). BAÍA AND TORGO 9 of 13 4.2 | Comparison of classification and regression modeling approaches This section presents the results of the experimental comparisons between the two general approaches to making trading decisions based on forecasting models. In our experiments, we have considered 76 clas- sification models. For each of these models, we have also tried the ver- sion with resampling and the version with cost–benefit matrices, totaling 76 × 3 = 228 different classification variants. In terms of regres- sion, we have a slightly larger set of 97 base models that were then tried with and without resampling, for a total of 97 × 2 = 194 variants. All these variants were compared on the data sets of the 12 companies described in Section 3.1 using the methodology described in Section 3.3. We have divided our experimental analysis in two main parts. In the first one, for each company and for each metric, we have compared the best regression and classification variant using a Wilcoxon singed‐ rank statistical test with a significance level of 0.05 to check if we can reject the null hypothesis that there is no significant difference between the best classification and regression variants. This leads to 12 statistical tests for each metric (one test for each company), where the models compared for each company are not necessarily the same. The motivation of this first part is to compare the best classification variant against the best regression variant for each company and metric. Figure 7 shows the results of this comparison for the Total Return and Sharpe Ratio financial evaluation metrics. The results on these figures are somewhat correlated. In effect, whenever we have found a significant difference in terms of Total Return, the same also FIGURE 7 Best classification variant against the best regression one for respective variant is significantly better, according to a Wilcoxon test with happened in terms of Sharpe Ratio. Regarding the left graph (Total Return), we have one significant win for each approach and 6 against four non‐ignificant wins for classification and regression, respectively. With respect to the right graph (Sharpe Ratio), we can observe a slight advantage of the classification approach, with one more significant win and eight vs one non‐significant wins. Overall, we have observed a very slight advantage of the best classification approach against the best regression variant. From an economical perspective, some results are contradictory. For instance, there is a very high level of Total Return for the Meg company (above 60% return), but the best Sharpe Ratio was very low. This means that the best model for the first metric was taking enormous amounts of risk and that the high level of return achieved was probably due to pure luck. On the other hand, there are some high values for the Total Return accompanied by high levels of Sharpe Ratio, such as for the Exas company. This strongly suggests that the models could actually provide some profit with low risk, thus indicating that the model actually predicted meaningful signals. Given the high variability of the results across companies, taking conclusions solely based on the analysis of the best variant per model and per metric may lead to wrong results. This establishes the motivation for the second part of our experiments. In this second part of our experiments, instead of grouping by metric and company, we will just group by metric and study the average rank of each model across all the companies (top five of each approach are considered). With the use of the Friedman test followed by the post hoc Nemenyi test, we check whether there are statistically the Total Return and Sharpe Ratio metrics (asterisks denote that the α = 0.05) 10 of 13 BAÍA AND TORGO significant differences among these rankings. This way, if a model obtains a very good result for one company but poor for all the others (meaning that it was lucky in that specific company), its average rank- ing will be low allowing the top average rankings to be populated by the true top models that perform well across most companies. Figure 8 summarizes the results in terms of Total Return and Sharpe Ratio. In either case, because we could not reject the Friedman null hypothesis, the post hoc Nemenyi test was not performed. This means that we cannot say with 95% confidence that there is some sig- nificant difference in terms of the average rank of the models for both the Total Return and the Sharpe Ratio between these two modeling approaches. Nevertheless, there are some observations to remark. Regarding Total Return, the model with the best average ranking is a classification model using cost–benefit matrices. All the remaining clas- sification variants are in their original form (without using cost–benefit matrices) and occupying mostly the last positions in terms of average rankings. Moreover, not a single variant obtained with SMOTE appears in this top five for each approach, which means that we confirm that resampling does not seem to pay off for this type of applications due to the economic costs of making more risky decisions. Furthermore, another very interesting remark is that all the top models are using SVMs as the base learning algorithm. Overall, we cannot say that any of the two approaches (forecasting directly the trading actions using classification models or forecasting the price returns using regression) is better than the other regarding the Total Return. The second part of Figure 8 shows the results of the same exper- iment in terms of Sharpe Ratio, that is, the risk exposure of the alterna- tives. The conclusions are quite similar to the Total Return metric. FIGURE 8 The top five average ranking model variants of both modeling a iants, and their average rankings are recalculated. Each model is thus given implies that the Friedmans null hypothesis of all averages being equal was r black line is considered to have their average rankings significantly differen Once again, no significant differences were observed. Still, one should note that the first five places are dominated by the classification approaches. The best variant for the Total Return is also the best variant for the Sharpe Ratio, which makes this variant unarguably the best one of our study when considering the 12 different companies. Hence, ultimately, we can state that the most solid model belongs to the classification approach using an SVM with cost–benefit matrices, because it obtained the highest returns with lowest associated risk. Finally, unlike the results for Total Return, in this case, we observe other learning algorithms appearing in the top five best results. In conclusion, we cannot state that one approach performs definitely better than the other in the context of financial trading decisions. The scientific community typically puts more effort into the regression models, but this study strongly suggests that both have at least the same potential. Actually, the most consistent model we could obtain belongs to the classification approach. Another interest- ing conclusion is that of a considerably large set of different types of models, SVMs achieved better results both when considering classifi- cation or regression tasks. 5 | DISCUSSION In this paper, we have studied the question of classification versus regression on decision making problems based on forecasts of a numeric variable. Our study was focused on financial trading decisions, so the obvious question is as follows: Can these results be generalized to other domains/tasks with similar structure? pproaches (Classification and Regression) are forming a new set of var- an average ranking (x axis). The presence of at least one black line ejected. In that case, every pair of model variants not connected by any t according to a Nemenyi statistical test BAÍA AND TORGO 11 of 13 This question was addressed in Baia (2015). The setup used in that work was rather similar to the one presented here, with two main dif- ferences: (a) the tasks were associated to several distinct contexts (while the current paper focus only on trading tasks); and (b) each task was considered with a different number of possible decisions to be made. While the second point allowed us to evaluate if one modeling approach could gain some specific advantage over the other depending on the number of values of the decision variable, the first point let us check if whatever conclusion we reached regarding trading problems would also apply in a more generic context. Moreover, the tasks considered in Baia (2015) were all non‐temporal, stressing out even more the differences to the trading problem. The main conclusions of the study carried out in Baia (2015) were that, overall, the classification modeling approach can outperform more frequently the approach based on regression tools. Whether the user is willing to make an extensive search for the optimal parameters to model a certain task or not, the results point in the same direction. The methods based on the classification approach tend to be better. The only setup where the regression approach was more competitive was on tasks with a high number of classes/decisions and where the user is not merely interested in the accuracy of the decisions and wants to consider different grades of severity of the decision errors. Contrary to what was observed in Baia (2015), in this paper, we have gathered enough evidence to state that there is no statistically significant difference between both approaches in the context of the trading decision problem. The trading problem has some characteris- tics that make it quite different from the generic tasks studied in Baia (2015), which may explain the different conclusions. Namely, the tasks addressed in the current paper are based on data that is ordered by time (time series data), and the frequency of decisions is rather imbalanced, with the more important decisions being rare. These characteristics raise significant challenges to modeling tools, and even with the use of techniques developed to address these issues, we were not able to obtain conclusive results. In summary, the conclusions of the current paper seem to be specific to the financial trading setting. For more general tasks, existing evidence (Baia, 2015) seems to indicate that there is some advantage on using the classification approach. Future work should try to explain the reason for this different conclusion, namely if it is something specific to the domain or to some of its characteristics (time‐dependent data and imbalanced decisions). 6 | RELATED WORK To the best of our knowledge, this is the first research work to directly compare the regression and classification approaches to financial trad- ing. Existing work in this area essentially uses one of the approaches and compares different variants of it on some concrete financial data sets. Even outside of the financial trading domain, we were not able to find some comparison between these two plausible approaches to decision making based on numeric forecasts. Still, as we have men- tioned in the introduction, we think this is a common and relevant setup for many application domains. Nevertheless, there are some concepts that are strongly related with the problem of making decisions based on numeric forecasts and thus we provide here a short review of some of the main works in these areas. In this paper, the ultimate goal is to predict an action/ decision. However, in many application domains, there are decisions that are more important than others, and there may exist information on some costs and benefits associated with each decision. In this con- text, all area of cost‐sensitive learning (Elkan, 2001) is strongly related with our target applications. Another related problem is that of imbal- anced distributions of the target variable in the context of prediction models. Both these two problems are present in financial trading prob- lems. Branco, Torgo, and Ribeiro (2016) present a survey of existing techniques for handling these situations of imbalanced target variables. Although most of the existing work considers classification tasks (nominal target variables), this work also describes methods designed to handle similar problems within regression tasks (numeric target variables). In terms of the use of regression models for trading systems, neu- ral networks models have shown to be a promising tool for forecasting time series (namely the prices of some assets). However, they typically require a large effort in terms of model tuning and can be computation- ally demanding. Genay (1999) has observed that a simple feed‐forward network fails to statistically outperform the random walk model when the input variables are just the past returns. However, with the simple addition of a moving average, it can statistically outperform this base- line. Ghazali, Hussain, and Liatsis (2011); Shin and Ghosh (1995) and Sermpinis, Dunis, Laws, and Stasinakis (2012) have proposed variants and generalizations of neural networks that present decent improve- ments over the standard versions of these models. Most research seems to indicate that these standard versions of the neural networks models may present poor results, yet some small variations in terms of the structure of the model may greatly improve their performance. Another popular modeling technique in this area is k‐nearest neighbours (KNN). Genay (1999) has observed that the regression KNN model statistically outperforms the random walk model by merely using the past return as predictors, unlike the feed‐forward neural network. Lee, Wei, Cheng, and Yang (2012) have used the nearest‐neighbour‐based approach for churn predictions. Even though this problem is different from financial trading, there are some similar- ities. Detecting an uncommon event the soonest possible can be seen as highly relevant in trading, that is, detecting a buy or sell signal the earliest possible to obtain the maximum possible profit. Support Vector Machines (SVM) and Multivariate Adaptive Regression Splines (MARS) have been the subject of a study by Kao, Chiu, Lu, and Chang (2013). The author also tested these models incorporated with wavelets, where some interesting results were obtained. Furthermore, Lu et al. (2009) has studied the combination of using independent component analysis to reduce the noise and randomness of the data before applying standard SVM models and has observed a slight improvement of the results. In terms of using classification models in the context of financial trading, there are not some many works. Chang, Fan, and Liu (2009) have studied the combination of piecewise linear models with back‐ propagation artificial classification neural networks PLR‐BPN. The experimental results were interesting in terms of the amount of profit 12 of 13 BAÍA AND TORGO obtained. However, Luo and Chen (2013) have seen that PLR‐BPN is outperformed by the combination of PLR with the well‐known Support Vector Machine model. Ma et al. (2012) have used cost matrices with back‐propagation neural networks. In several tasks, an increase of the utility score of a model came at the cost of a decrease in the accuracy level. However, this accuracy decrease may not be relevant for finan- cial trading, where the main goal is to avoid serious errors like forecast- ing a buy signal when we should sell, as this type of errors may have very serious economic impact. Teixeira and de Oliveira (2010) have studied the classification KNN model and also the combination of this model with indicators such as the RSI filter, stop‐gain and stop‐loss criteria. All the tested models outperformed the used benchmark, par- ticularly the models combined with the above indicators that obtained an overall better performance. Atsalakis and Valavanis (2009) contain a very detailed state of art review on stock market forecasting techniques. For each referred work, the authors list the used data sets, the chosen input variables and, most importantly, a summary of all the used modeling techniques as well as which models were compared against each other. Informa- tion regarding the usage of pre‐processing techniques and the training method was also given. In summary, although our review of the related literature has not found any work with objectives similar to the current paper, we were able to observe that the most frequent approach to financial trading is based on regression approaches. Our work has shown that classifica- tion approaches should be given more importance by the research community in this area. 7 | CONCLUSIONS This paper presents a comparative study of two different approaches to financial trading decisions based on forecasting models. The first, and more conventional approach, uses regression tools to forecast the future evolution of prices and then uses some decision rule based on these predictions, to choose the trading action. The second approach tries to directly forecast the “correct” trading decision. Our study is a specific instance of the more general problem of making decisions based on numerical forecasts. In this paper we have focused on financial trading decisions because this is a domain that requires specific trade‐offs in terms of economic results. This means that our conclusions are specific to this area and it remains an open question for future research whether the same conclusions can be drawn in other application domains where the same two approaches to decision making are plausible. Still, our initial experiments (Baia, 2015) with other types of domains provide some evidence for the specificity of financial trading decision making that is addressed in the current paper. Overall, the main conclusion of the study we have described in this paper is that, for this specific application domain, there seems to be no statistically significant difference between these two approaches to decision making. Given the large set of classification and regression models that were considered, as well as the different data sets involved in our study, we claim that this conclusion is supported by significant experimental evidence. Financial forecasting has some particularities that are challenging to most modeling techniques. In our study we have considered the hypothesis of using some existing techniques that were developed to address this type of challenges. Another contribution of this paper involves testing the applicability of these solutions in the context of financial forecasting. Regarding the problem of the imbalance of the distribution of the target variable we have considered the application of resampling strategies both for regression and classification models. This technique has been applied with success to many application domains. Our experiments have shown that although resampling increases the ability of the models to generate trading signals (thus increasing their recall), it also brings a significant amount of financial risk as several of these signals are wrong, frequently leading to cata- strophic financial results. This means that our study clearly indicates that resampling is not recommended in the context of financial fore- casting due to this increase in the risk exposure. The other standard technique we have considered in our study was the use of cost– benefit matrices as a means to make the classification models aware of the different costs of the classification errors. In this case, our study collected sufficient evidence to conclude that this method is promising, although not all modeling techniques were able to capitalize on the extra information provided by these matrices. ACKNOWLEDGEMENTS This work is financed by the European Regional Development Fund (ERDF) through the Operational Programme for Competitiveness and Internationalisation ‐ COMPETE 2020 Programme within project POCI‐01‐0145‐FEDER‐006961 and by the North Portugal Regional Operational Programme (ON.2 ‐ O Novo Norte), under the National Strategic Reference Framework (NSRF), through the European Regional Development Fund (ERDF), and by national funds, through the Portuguese funding agency (FCT) within Project NORTE‐07‐ 0124‐FEDER‐000059. Part of the work of Luís Torgo was supported by a sabbatical scholarship (SFRH/BSAB/113896/2015) from the Portuguese funding agency (FCT). REFERENCES Atsalakis, G. S., & Valavanis, K. P. (2009). Surveying stock market forecast- ing techniques part II: Soft computing methods. Expert Systems with Applications, 36(3, Part 2), 5932–5941. Baia, L. (2015). Actionable forecasting and activity monitoring: Applications to financial trading. Branco, P., Torgo, L., & Ribeiro, R. (2016). A survey of predictive modeling on imbalanced domains. ACM Computing Surveys (to appear). Chang, P.‐C., Fan, C.‐Y., & Liu, C.‐H. (2009). Integrating a piecewise linear representation method and a neural network model for stock trading points prediction. IEEE Transactions on Systems, Man and Cybernetics Part C: Applications and Reviews, 39(1), 80–92 .cited By 39 Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). Smote: Synthetic minority over‐sampling technique. Journal of Artificial Intelligence Research, 16(1), 321–357. Demšar, J. (2006). Statistical comparisons of classifiers over multiple data sets. The Journal of Machine Learning Research, 7, 1–30. Elkan, C. (2001). The foundations of cost‐sensitive learning. In IJCAI’01: Proc. of 17th Int. Joint Conf. of Artificial Intelligence, volume 1, pp. 973–978. Morgan Kaufmann Publishers. BAÍA AND TORGO 13 of 13 Genay, R. (1999). Linear, non‐linear and essential foreign exchange rate prediction with simple technical trading rules. Journal of International Economics, 47(1), 91–107. Ghazali, R., Hussain, A. J., & Liatsis, P. (2011). Dynamic ridge polynomial neural network: Forecasting the univariate non‐stationary and stationary trading signals. Expert Systems with Applications, 38(4), 3765–3776. Hellstrom, T. (1999). Data snooping in the stock market. Theory of Stochastic Processes, (21, 1999b), pp. 33–50. Kao, L.‐J., Chiu, C.‐C., Lu, C.‐J., & Chang, C.‐H. (2013). A hybrid approach by integrating wavelet‐based feature extraction with {MARS} and {SVR} for stock index forecasting. Decision Support Systems, 54(3), 1228–1244. Lee, Y.‐H., Wei, C.‐P., Cheng, T.‐H., & Yang, C.‐T. (2012). Nearest‐neighbor‐ based approach to time‐series classification. Decision Support Systems, 53(1), 207–217. Liaw, A., & Wiener, M. (2002). Classification and regression by random forest. Lu, C.‐J., Lee, T.‐S., & Chiu, C.‐C. (2009). Financial time series forecasting using independent component analysis and support vector regression. Decision Support Systems, 47(2), 115–125 .cited By 112 Luo, L., & Chen, X. (2013). Integrating piecewise linear representation and weighted support vector machine for stock trading signal prediction. Applied Soft Computing, 13(2), 806–816. Ma, G.‐Z., Song, E., Hung, C.‐C., Su, L., & Huang, D.‐S. (2012). Multiple costs based decision making with back‐propagation neural networks. Decision Support Systems, 52(3), 657–663. Meyer, D., Dimitriadou, E., Hornik, K., Weingessel, A., & Leisch, F. (2014). e1071: Misc Functions of the Department of Statistics (e1071), TU Wien. R package version 1.6–4. Milborrow, S. (2014). Earth: Multivariate adaptive regression spline models. R package version 3.2–7. R Core Team (2014). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Ridgeway, G. (2013). GBM: Generalized Boosted Regression Models. R package version 2.1. Ridgeway, G., Southworth, M. H., & RUnit, S. (2013). Package gbm. Sermpinis, G., Dunis, C., Laws, J., & Stasinakis, C. (2012). Forecasting and trading the eur/usd exchange rate with stochastic neural network com- bination and time‐varying leverage. Decision Support Systems, 54(1), 316–329. Shin, Y., & Ghosh, J. (1995). Ridge polynomial networks. IEEE Transactions on Neural Networks, 6(3), 610–622 .cited By 64 Teixeira, L. A., & de Oliveira, A. L. I. (2010). A method for automatic stock trading combining technical analysis and nearest neighbor classification. Expert Systems with Applications, 37(10), 6885–6890. Torgo, L. (2010). Data Mining with R, learning with case studies. London, United Kingdom. Torgo, L., Branco, P., Ribeiro, R. P., & Pfahringer, B. (2015). Resampling strategies for regression. Expert Systems, 32(3), 465–476. Venables, W. N., & Ripley, B. D. (2002). Modern applied statistics with S (4th ed. ISBN 0‐387‐95457‐0). London, United Kingdom. Wilder, J. (1978). New concepts in technical trading systems. Kingston, New York: Trend Research. How to cite this article: BaíaL,TorgoL.Acomparativestudyof approaches to forecast the correct trading actions. Expert Sys- tems. 2017;34: e12169. https://doi.org/10.1111/exsy.12169 AUTHOR BIOGRAPHIES Luis Baia has a master degree in Applied Mathematics and a Bachelor in Pure Mathematics. After some months researching in Machine Learning at LIAAD, he joined a company called “Farfetch” as a Data Scientist. Luis Torgo is an associate professor in the Department of Computer Science at the University of Porto in Portugal. An active researcher in machine learning and data mining for more than 20 years, Dr. Torgo is also a researcher in the Laboratory of Artificial Intelligence and Data Analysis (LIAAD) of INESC Porto LA. https://doi.org/10.1111/exsy.12169 
work_oqlymuuk6ja3lfpzblmme72kyy	----	ProQuest Dissertations Quantization-free parameter space reduction in ellipse detection Kuang Chung Chen A Thesis in The Concordia Institute for Information Systems Engineering Presented in Partial Fulfillment of the Requirements for the Degree of Master of Applied Science (Information Systems Security) at Concordia University Montreal, Quebec, Canada September 2009 © Kuang Chung Chen, 2009 1*1 Library and Archives Canada Published Heritage Branch 395 Wellington Street Ottawa ON K1A 0N4 Canada Bibliotheque et Archives Canada Direction du Patrimoine de I'edition 395, rue Wellington Ottawa ON K1A 0N4 Canada Your file Votre reference ISBN: 978-0-494-63052-5 Our file Notre reference ISBN: 978-0-494-63052-5 NOTICE: AVIS: The author has granted a non- exclusive license allowing Library and Archives Canada to reproduce, publish, archive, preserve, conserve, communicate to the public by telecommunication or on the Internet, loan, distribute and sell theses worldwide, for commercial or non- commercial purposes, in microform, paper, electronic and/or any other formats. L'auteur a accorde une licence non exclusive permettant a la Bibliotheque et Archives Canada de reproduire, publier, archiver, sauvegarder, conserver, transmettre au public par telecommunication ou par I'lnternet, prefer, distribuer et vendre des theses partout dans le monde, a des fins commerciales ou autres, sur support microforme, papier, electronique et/ou autres formats. The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. L'auteur conserve la propriete du droit d'auteur et des droits moraux qui protege cette these. Ni la these ni des extraits substantiels de celle-ci ne doivent etre imprimes ou autrement reproduits sans son autorisation. In compliance with the Canadian Privacy Act some supporting forms may have been removed from this thesis. Conformement a la loi canadienne sur la protection de la vie privee, quelques formulaires secondaires ont ete enleves de cette these. While these forms may be included in the document page count, their removal does not represent any loss of content from the thesis. Bien que ces formulaires aient inclus dans la pagination, il n'y aura aucun contenu manquant. 1 + 1 Canada Abstract Quantization-free parameter space reduction in ellipse detection Kuang Chung Chen Ellipse modeling and detection is an important task in many computer vision and pattern recognition applications. In this thesis, four Hough-based transform algorithms have been carefully selected, studied and analyzed. These techniques include the Standard Hough Transform, Probabilistic Hough Transform, Randomized Hough Transform and Directional Information for Parameter Space Decomposition. The four algorithms are analyzed and compared against each other in this study using synthetic ellipses. Objects such as noise have been introduced to distract ellipse detection in some of the synthetic ellipse images. To complete the analysis, real world images were used to test each algorithm resulting in the proposal of a new algorithm. The proposed algorithm uses the strengths from each of the analyzed algorithms. This new algorithm uses the same approach as the Directional Information for Parameter Space Decomposition to determine the ellipse center. However, in the process of collecting votes for the ellipse center, pairs of unique edge points voted for the center are also kept in an array. A minimum of two pairs of edge points are required to determine the ellipse. This significantly reduces the usual five dimensional array requirement needed in the Standard Hough Transform. We present results of the experiments with synthetic images demonstrating that the proposed method is more effective and robust to noise. Real world applications on complex real world images are also performed successfully in the experiments. in Acknowledgements Most importantly, I wish to thank my advisor Nizar Bouguila for providing me with this great opportunity. I have really learned a lot and I really appreciate that. I would also like to thank Dr. Bouguila for his encouragement, support and guidance when I needed it the most. Additionally, I would like to thank Professor Djemel Ziou from the University of Sherbrooke for rec- ommending using the Hough Transform when we started working on the problem of eye detection and modeling. Without his coaching on the Hough Transform, our research direction would have been signifi- cantly different. Finally, I would like to thank my wife, I-Chun, for always believing in me, and my daughter Ashlin, for opening up a new chapter in my life. IV Table of Contents List of Tables vii List of Figures viii 1 Introduction 1 1.1 Introduction 1 1.2 Algorithm Motivation 3 1.3 Background 3 1.3.1 Standard Hough Transform (SHT) 4 1.3.2 Probabilistic Hough Transform (PHT) 6 1.3.3 Randomized Hough Transform (RHT) 7 1.3.4 Parameter Space Decomposition (PSD) 11 1.4 Contributions 13 2 Experimental results using synthetic and real world images 15 2.1 Description of the experiments 15 2.2 Experiment using single synthetic ellipse 15 2.3 Experiment using single synthetic ellipse with a rectangle 16 2.4 Experiment using synthetic ellipse with a rectangle, triangle, and a letter T 17 2.5 Experiment using synthetic ellipse with a rectangle, triangle, and a letter T, with one percent noise 18 2.6 Summary of the four experiments 19 2.7 Experiment using real world images 20 2.8 Comparison and discussion of the Hough based algorithms 24 3 New proposed algorithm 28 3.0.1 Introduction 28 3.0.2 Recent works 29 3.1 Quantization-free parameter space reduction(QFPSD) 31 3.1.1 Locating the ellipse center using two points 32 v 3.1.2 Equidistant points to the ellipse center 33 3.1.3 Locate the ellipse center with the highest accumulated cell and retrieve edges that have voted for the ellipse center 34 3.1.4 Points projected using pi,P2 and ellipse center 34 3.1.5 Normal distribution 35 3.1.6 Direct Least Square Ellipse fitting 36 3.2 Experiments 37 3.2.1 Experiments with synthetic images 38 3.2.2 Experiments on elliptical shapes in real world environment 42 3.2.3 Real world application: traffic sign detection 45 3.2.4 Real world application: eye detection 51 3.2.5 Analysis of Quantization-free parameter space reduction 54 4 Conclusions 57 List of References 59 v i List of Tables 2.1 Experiments - Single synthetic ellipse detection 16 2.2 Experiment2: Single synthetic ellipse detection using figure 2.2 17 2.3 Experiments - Single synthetic ellipse detection using figure 2.3 18 2.4 Experiment4 - Single synthetic ellipse detection using figure 2.4 with one percent noise. . . 19 2.5 Noise and edge influence in the experiments 20 2.6 Calculation time required in the SHT,PHT, RHT and PSD for eyel-12 24 2.7 Number of major iterations and memory requirements for the Standard Hough Transform, Probabilistic Hough Transform, Randomized Hough Transform and Parameter space de- composition 25 3.1 Theoretical values and QFPSR in figure 3.9d 40 3.2 Theoretical values and QFPSR in figure 3.10 40 3.3 Value comparison of theoretical, SHT, PHT, RHT, PSD and QFPSR 51 3.4 Calculation time required in SHT, PHT, RHT, PSD and QFPSR for eye 1-12 54 vn List of Figures 1.1 Edges missing in ellipse 3 1.2 An eye and its edges 3 1.3 Overlapping ellipses 4 1.4 Ellipse parameters 5 1.5 Ellipse example 6 1.6 Accumulator 7 1.7 PHT and edges 8 1.8 Ellipse example with Probabilistic Hough Transform 8 1.9 Accumulator for Probabilistic Hough Transform 9 1.10 Create tangent lines at point p1 ?p2 and p3 and find their midpoints mi2 and m23 10 1.11 Geometrical relationship between the two randomly selected points and their tangent line . . 11 2.1 Single synthetic ellipse 16 2.2 Experiment2 - Single synthetic ellipse and a single rectangle 16 2.3 Experiment3 - Single synthetic ellipse, rectangle, triangle and a letter T 17 2.4 Experiment4 - Single synthetic ellipse, rectangle, triangle and a letter T with one percent noise 18 2.5 Close edge selected for the random sampling 20 2.6 Ellipse formation using close edge points 21 2.7 Results of each algorithm for eye 1 21 2.8 Results of each algorithm for eye 2 21 2.9 Results of each algorithm for eye 3 22 2.10 Results of each algorithm for eye 4 22 2.11 Results of each algorithm for eye 5 22 2.12 Results of each algorithm for eye 6 22 2.13 Results of each algorithm for eye 7 22 2.14 Results of each algorithm for eye 8 23 2.15 Results of each algorithm for eye 9 23 2.16 Results of each algorithm for eye 10 23 2.17 Results of each algorithm for eye 11 23 vni 2.18 Results of each algorithm for eye 12 23 2.19 A normal eye 25 3.1 Geometrical relationship between the two randomly selected points and their tangent line . . 32 3.2 pi and ^2 equidistant to the ellipse center (xo,yo) 33 3.3 Projected points from pi andp2 35 3.4 Normal distribution of distance between ellipse points and ellipse center 36 3.5 Direct least square fitting for ellipse 37 3.6 Sampling interval for the pairs of points 38 3.7 Accuracy of ellipse using an interval of 50 39 3.8 Accuracy of ellipse detection using one pair of points 39 3.9 Experimental results to demonstrate both filters 40 3.10 Occluded ellipses 40 3.11 Pink bicycle result 41 3.12 Wheels result 42 3.13 Left rear of a red volvo 850 43 3.14 Antique vase 44 3.15 Magnifying glass 44 3.16 Hidden plate 45 3.17 Traffic sign 1 46 3.18 Traffic sign 1 results 46 3.19 Traffic sign 2 46 3.20 Traffic sign 2 results 47 3.21 Traffic sign 3 47 3.22 Traffic sign 3 results 48 3.23 Traffic sign 4 49 3.24 Traffic sign 4 results 50 3.25 Results of each algorithm for eye 1 52 3.26 Results of each algorithm for eye 2 52 3.27 Results of each algorithm for eye 3 52 3.28 Results of each algorithm for eye 4 52 3.29 Results of each algorithm for eye 5 52 3.30 Results of each algorithm for eye 6 53 3.31 Results of each algorithm for eye 7 53 3.32 Results of each algorithm for eye 8 53 3.33 Results of each algorithm for eye 9 53 3.34 Results of each algorithm for eye 10 53 3.35 Results of each algorithm for eye 11 54 3.36 Results of each algorithm for eye 12 54 3.37 Multiple local maxima 56 IX CHAPTER 1 I Introduction 1.1 Introduction Detecting elliptical shapes accurately and immediately has been a major issue in computer vision and pattern recognition. These elliptical shapes are commonly viewed in real world scenes. Some of these elliptical shapes can be found in the human body, such as the head [7], eyes [18] and lips [29]. In the biometric world, the eye is often used as a method of personal identification [10]. In driver vigilance and fatigue monitoring systems, the gaze of an eye can provide us with information about the attention a driver is placing on the road [13] [18] [2] [14]. The appearance of elliptical forms increases on an image through the perspective projection of 3-Dimensional circular or elliptical features. This application includes road sign detection [25] and cancerous cell counting [23]. Some of the methods currently applied to detect ellipses are based on symmetry [28] [17]. However, since ellipses of a given image are rarely in perfect condition, the symmetry is hard to apply. Often, segments of an ellipse are missing and symmetry can't be used. Other methods of ellipse detection include random sampling [20] [22]. These methods have an advantage where memory consumption is kept to the minimum. These methods perform well in a real environment where ellipse edge presence is high [19]. However, in real world images, where the number of ellipse edges is low, sampling the correct ellipse edges becomes more difficult [19]. This difficulty is compounded when random sampling edges must also satisfy a specific edge combination. In other words, either the sampled ellipse edges must not be too close to each other or only a specific combination of ellipse edges can be used to locate the ellipse parameters accurately. These 1 edge point combinations are normally close to the major and minor axis of the ellipse and far apart from each other in order to obtain the best accurate results for ellipse parameters. If they are not, the accuracy of the ellipse detection is compromised. Another issue is that random sampling of edges cannot produce consistent results. This is why a specific number of trials are necessary to produce the final average result, even if it is for the same image. But, if a consistent result is required in a probabilistic model, more trials are needed [16]. To produce a consistent result in terms of ellipse detection (success/failure) and speed, Hough-based transforms can be used. These methods are recognized as a powerful tool in shape detection, or analysis, with good results in a noisy and occluded environment. Although the technique has a straight forward computation, its primary limitations include lengthy computational time and massive storage requirements. These limitations are barely noticeable with today's high power computing, especially for line detection. For a complex shape, such as ellipse, the computation slowness and five dimensional storage requirements are still obvious. Since the introduction of the Hough Transform, many improvements have been made . These improve- ments have targeted performance, software methods, probabilistic models and parallel processing. In the case of performance, the algorithms must perform relatively well in a noisy environment. There are in gen- eral two types of noise: random noise and correlated noise. Random noise occurs randomly in the image. As for correlated noise, it only happens when two features are being placed together to form a third feature. Concerning software methods, algorithms have been rewritten efficiently for specific hardware. One of the most popular methods is the lookup table. Using this method, a number of calculations can be performed in advance and stored in an array index . This method speeds up the calculation time, but uses an enormous amount of memory. Other methods based on Hough transforms vary from edge directional information to parameter space decomposition. Edge directional information can be used to limit the input range of the parameter space. Meanwhile, in the parameter space decomposition, the parameter space can be reduced into separate stages. Each stage passes the result into the next. In the case of probabilistic models, algorithms seek to reduce the number of redundancies using a smaller sample population size. A comprehensive survey of Hough-based transform algorithms can be found in [16]. 2 (a) Image (b) Edges Figure 1.1: Edges missing in ellipse (a) Eye (b) Edges Figure 1.2: An eye and its edges 1.2 Algorithm Motivation Detecting ellipses accurately and quickly is difficult in many computer vision applications. Accurate detec- tion is generally compromised due to the presence of noise, partially hidden ellipse or poor threshold values in edge detection, as illustrated in Figure 1.1. Additionally, overlapped ellipse formation interferes with ellipse accuracy detection. This overlapping adds complexity to the problem. As an example, an image of an eye is represented in Figure 1.2a and its edges are represented by Figure 1.2b. As illustrated in Figures 1.3a and 1.3b, extra edges create additional potential ellipses or features in the image. These ellipses or features can be overlapped one on top of the others. This overlapping is difficult to be spotted even by the naked eyes as illustrated in Figure 1.2b. 1.3 Background An ellipse located at the origin with no axis of rotation can be described as follows: 3 (a) (b) Figure 1.3: Overlapping ellipses [ - ] 2 + [^]2 = l (1) a b where a is semi-major axis and b is semi-minor axis. However, when the axes of ellipse are rotated on an angle <f>, the equations of orientation for the semi- major and semi-minor axes become: x' — x cos 6 + y sin 6 (2) y = —x sin <j> + y cos <p Substituting the previous two equations into equation 1, a new ellipse equation located at the origin with axis orientation can be obtained by: x c o s 0 + y s i n 0 2 -xsin</> + ycoscp 2 _ a b If a translation is performed on the ellipse to (xQ,yo), equation 3 becomes: (x - x0) cos <j) + {y- yo) sin </> 2 - ( x - x0) sin </> + (y - y0) cos </> 2 _ 1 In the following sections, we will present in details the Standard Hough Transform [4], Probabilistic Hough Transform [15], Randomized Hough Transform [23] and finally, Parameter Space Decomposition [ 1 ]. 1.3.1 Standard Hough Transform (SHT) As stated previously, detecting elliptical shape is a difficult problem in computer vision. This difficulty is primarily due to the variables that describe an ellipse. In the case of an ellipse, there are five variable 4 parameters: center (xo,yo), semi-major axis a, semi-minor axis b and orientation of the ellipse angle </> as illustrated in Figure 1.4. y-axas ^ x-axjs Figure 1.4: Ellipse parameters A popular method used to detect ellipses is the Hough Transform. The Standard Hough Transform was first proposed in 1962 [11] and later improved by Duda [4]. The basic concept behind the Standard Hough Transform in ellipse extraction is to define a two dimensional image space mapping in the xy-plane with a five dimensional parameter space mapping in the xoJ/oa^-plane. All possible ellipses in the parameter space can be generated from a given point in the image space. This process is then repeated for every other points in the image space. The parameter set that appears most frequently indicates the possible presence of an ellipse. The counting frequency of a parameter set is constructed using an accumulator cell. This counting technique is often referred to as a voting scheme where each accumulator cell represents a parameter set 5 is normally quantized and is initially set to zero. The collection of all the accumulator cells is called the accumulator. An example of an accumulator, taking into account the ellipse in Figure 1.5 is presented in Figure 1.6. %F axis Cjj=02 • > x-axts Figure 1.5: Ellipse example 1.3.2 Probabilistic Hough Transform (PHT) The Probabilistic Hough Transform was proposed in 1991 by Kiryati [15]. The idea is to reduce the number of points across the image evenly as illustrated in Figures 1.7a and 1.7b. After reducing the number of points in the image, the Standard Hough Transform is used. This reduction of points reduces the computational time. The improvement in the Probabilistic Hough Transform can be explained as follows. Normally, in the Standard Hough Transform, there are two phases. The first phase consists of the generation of all the possible ellipse parameter sets using every point in the image. The second phase consists of searching for the highest accumulator cell. Analyzing these two phases, the first phase clearly takes more time. Using only a small percentage of points to compute the first phase clearly reduces the overall computing time. Figures 1.8 and 1.9 demonstrate an example of the Probabilistic Hough Transform. The comparison between figures 1.6 and 1.9 demonstrate that figure 1.9 is only a scaled down version of Figure 1.6. The overall shape of the histogram is retained. Additionally, according to Kiryati [15], the percentage of points required from the original image falls in the range of 5% to 15%. 6 Representation of an accumulator 1ZU -| c e ll o 0 5 "3 6 .£ > ' * ' . ' " • ' • . : ' . • . ••- a . * b 3 >o I O • •t". . ?*•_•• J . " ' • " ' • • • 0 10 20 10 20 ' PI'"' 0 10 20 20 30 - _ _ _ T • 1 i ^ • " " . • • • ' . ' • : - ' • * * • : • ' - ' ' • * • £ - & : . • • • 0 10 20 30 40 0 10 20 3 * 45 • ' s". ' ' " . • •' ' ' ' i . . . . v ' '• • - . ' - ' J 0 10 2 0 38 50 1 L'. * - . -• ' i i . . . . t | 0 10 20 42 55 £ * v ' I -'S" ; 0 10 20 46 60 • \ J, 0 10 20 56 70 0 10 20 70 too accumulator cell Figure 1.6: Accumulator 1.3.3 Randomized Hough Transform (RHT) The Randomized Hough Transform was originally proposed for curves that can be only expressed by a linear equation. Curves that are nonlinear cannot be applied directly. Reference [31] provides a brief summary of how to apply a Randomized Hough Transform in linear curve while reference [23] details how to apply it to an ellipse. The algorithm described by McLaughlin [23] is as follows. Select any of the three point's pi(x\ ,Vi),P2(x2,y2) and P3(x3,j/3) randomly from the image space and estimate a tangent line at each of the selected points using the neighboring points, as illustrated in Figure 1.10. Take any of the two selected points, in our first case p\ and p2, and locate their midpoint m\2 and the intersection of their tangent lines £12, as described in Figure 7 (a) All edges (b) 30 of edges Figure 1.7: PHT and edges 4f -axs (5038) m 28" A: q*fl's ^ x-axts Figure 1.8: Ellipse example with Probabilistic Hough Transform 1.10. The center (xo,2/o) will lie on the line In that passes through points t\2 and m\2 as described by Yuen et al [32]. Similarly, take points j>2 and ps and locate their midpoint, 77123. The center of the ellipse is situated on the line '23 that passes through £23 and 77123. The approximate center of the ellipse (xo,yo) is situated on the intersection of the lines l\2 and hi as described in Figure 1.10. The ellipse equation used by McLaughlin is: Accumulator for Probabilistic Hough Transform 30 - ? 25- 0 0 e 0 0 | 10- " • $ : • * b a »° I O 0 10 2 0 0 0 ' ; , • - . ' • * • X 0 10 20 0 10 0 10 20 10 20 • . • . : • ' • • s - - > . " in 20 30 0 10 20 30 40 l i it: 0 10 20 34 45 VI i i 1 fc I • 1 . '• t 0 10 20 38 50 | If" • ; . : . !> 0 10 20 42 55 n 0 10 20 46 60 0 10 20 56 70 0 ' 0 20 7 0 100 , i , 0 10 20 80 110 0 10 20 3 0 120 where accum(ilat6i cell Figure 1.9: Accumulator for Probabilistic Hough Transform A(x - x0) 2 + 2B(x - x0){y - y0) + C{y - y0) 2 = 1 AC - B2 > 0 and A, B and C are coefficients. For simplicity, the center of the ellipse is translated to the origin. Thus, Ax2 + 2Bxy + Cy2 = 1 (5) (6) (7) Substituting the coordinates of the points p i , p2 and p 3 into equation 5, a three linear equation system is obtained. 9 Figure 1.10: Create tangent lines at point p\, p2 and p3 and find their midpoints m\2 and 77123. x\ 2xxy\ y\ A 2xi 2/2 y\ x% 2xiy3 y|_ A B C = 1 1 1 (8) Once elliptical (xo,yo,A,B,C) quadratic equations shown in equation 8 are solved, a translation back to the original parameter form is required (x0,yo,a,b,cp) as demonstrated by Inverso [12]. Store the value of the ellipse parameters in the accumulator set, if the accumulator set does not exist previously in the accumulator. If the accumulator set does exist, combine it with the existing accumulator set found in the accumulator. 10 1.3.4 Parameter Space Decomposition (PSD) The idea behind this transform is to use two independent accumulators to gather evidence. The first accu- mulator is used to gather evidence regarding the ellipse center and the second accumulator is used to gather evidence about a, b and <fi. This transform focuses on locating a map using two selected points, p\ and pi, to find the ellipse center (xo,2/o)- Once the ellipse center (xo,2/o) has been located, it can be used as an input to gather additional evidence regarding parameters a, b and (j>. Figure 1.11: Geometrical relationship between the two randomly selected points and their tangent line Take any of the two selected points, in our first case pi and P2, and locate their midpoint m\i and the intersection of their tangent lines £12, as illustrated in Figure 1.11. Let L\ be aline that passes through point P\ and ii2- Let Li be a line that passes through points pi and t\2- Then, equations L\ and L<i are described as the following, where slopei and slopei are the slopes of L\ and Li, respectively. L\ — slope\x + c\ (9) Li = slopeix + ci 11 where c\ and c2 are constants. , Vt-Vi slopei = xt - xi , Vt-V2 siope2 = Xt - X2 Solving c\ and c2 in Equation 9 using point t\2, it can be determined that Vt-Vi c\=Vt xt xt - Xi Vt-Vi C2 = Vt Xt xt - x2 The equations of L\ and Li are as follows: L\ = slope\{x - xt) +yt L2 — slopei(x - xt) + yt By using points p\ and p2 in Equation 12 2/2 = slopei{x2 - xt) + yt j/i = slope2{xi - xt) + yt Solving Equation 13 for xt and yt yields: 2/2 — yi — slope2x2 + slopeixi xt -slope2yi Vt •- slope\ — slope2 slope2y\ + slope\y2 + slope\slope2{x\ — x2) slopei — slope2 Let the midpoint m,\2 of p\ and p2 be defined as (xm, ym), the slope for mu and t\2 slopes as follows: Let xm and ym be defined as slopes xm = ii yt - ym Xt Xm X2 + Xi 2 Vi + 2/1 12 Using xm and ym from Equation 16 and yt and xt from Equation 14 and substituting into Equation 15, it can be found that: AC + 2BD slopes — 2A + BC where slopei = slope2 = A = yi-y2 B = X\ — X2 C — slopei + slope2 D = slopei * slope2 2/i -yt Xi - Xt 2/2 -yt X2 - Xt Therefore, the equation of line that passes through the ellipse center and midpoint m\i can be expressed as: , AC + 2BDf 2/Q = 2 / m + 2A + BC {XO-Xm) (17) 1.4 Contributions The contributions of this thesis to the literature will be many. A thorough analysis of the various Hough Transform based algorithms for ellipse detection will be provided. The focus will primarily be on the Standard Hough Transform, Probabilistic Hough Transform, Randomized Hough Transform and Parameter Space Decomposition. After this thorough analysis, a new proposed method will be provided. This thesis is organized as the follows: Chapter 1 provides essential information about the concepts and definitions that will be referenced throughout the thesis. It also provides information about the different types of Hough Transforms. In chapter 2, synthetic images will be used to validate how quickly and ac- curately ellipses are being detected by each algorithm. Real world images will be introduced afterwards. 13 And a discussion will follow after the experiments. In chapter 3, a new method is proposed by taking into account the strengths and weaknesses of each Hough-based transform. Comparisons between the new pro- posed method and other Hough-based transforms will follow afterwards. Chapter 4 concludes the thesis and provides recommendation for future work. 14 CHAPTER 2 Experimental results using synthetic and real world images 2.1 Description of the experiments The main purpose of this chapter is to analyze each algorithm through experiments. The experiments were designed to bring out the best and the worst of each algorithm under different circumstances. It is impor- tant in our case, to determine how each algorithm deteriorates in terms of detection and calculation. The experiments were first conducted using synthetic ellipses. Additional objects were placed in afterwards to disturb ellipse detection. Finally, noise was scattered and added to the image. In the Probabilistic Hough Transform, only 20% of the edge points were taken. In the Randomized Hough Transform, each trial only had 200 samples. In each sample, three points were given the chance to vote. There were a total of 10 trials. The emphasis was placed mainly on the detection and the calculation time. The experiments were conducted using a DELL laptop D500 centrino 1.3GHz with 1.5Gigabytes of memory running on a Windows XP 32bit and Matlab 2008a system. 2.2 Experiment using single synthetic ellipse The first experiment was carried out using synthetic ellipses. Figure 2.1 illustrates an example. The ellipse center location (xo>yo)> orientation <f>, semi-major axis a and semi-minor axis b vary from picture to picture. There were in total ten images used in the first experiment. On average, there were a total of 77 edge points 15 in the ellipse. In this experiment, all the algorithms performed relatively well for the ten images provided. As expected, the Standard Hough Transform was the slowest and the parameter space decomposition was the quickest. The results of the first experiment are summarized in Table 2.1 Figure 2.1: Single synthetic ellipse Table 2.1: Experiment! - Single synthetic ellipse detection. Number of Trials Number of suc- cessful ellipse detections Percentage of suc- cessful rate Average time in seconds SHT 10 10 100% 500 PHT 10 10 100% 100 RHT 10 10 100% 42 PSD 10 10 100% 2 2.3 Experiment using single synthetic ellipse with a rectangle In the second experiment, an object of rectangular shape was added to the image as illustrated in Figure 2.2. Figure 2.2: Experiment2 - Single synthetic ellipse and a single rectangle Ellipse and rectangle location, shape and orientation were modified in every image. There were a total of ten images used in the second experiment. On average, there were, in total, 168 edge points in each image. 77 belonged to the ellipse and 91 belonged to the rectangle on average. This translates to over 50% of the edges that do not belong to the ellipse. In the second experiment, the Standard Hough Transform and 16 Directional information for parameter space decomposition had the highest accuracy in the ten images pro- vided. However, due to the extra involved edge points in the image, the average calculation has been almost doubled. As expected, the Standard Hough Transform was the slowest and the Directional information for parameter space decomposition was the quickest. It should be noted that the accuracy of the Randomized Hough Transform and Probabilistic Hough Transform have deteriorated slightly. The calculation time for the Randomized Hough Transform has been kept constant, since only a fixed number of samples were used and the number of trials remained the same. The results of the second experiment can be seen in Table 2.2 Table 2.2: Experiment2: Single synthetic ellipse detection using figure 2.2. Number of Trials Number of times of successful el- lipse detection Percentage of suc- cessful rate Average time in sec SHT 10 10 100% 900 PHT 10 7 70% 180 RHT 10 5 50% 42 PSD 10 10 100% 2 2.4 Experiment using synthetic ellipse with a rectangle, triangle, and a letter T In experiment 3, two additional objects were added. These objects consisted of a triangle and a letter T. The purpose of adding these two objects was to distract further each ellipse detection algorithm. The images used were similar to those in Figure 2.3. Figure 2.3: Experiment3 - Single synthetic ellipse, rectangle, triangle and a letter T. The location, shape and orientation of each ellipse, rectangle, triangle and letter T were changed ran- domly in every image. There were a total often images used in the third experiment. On average, there were, 17 in total 227 edge points in each image. 77 belonged to the ellipse and 150 belonged to the rest of the objects. Over | of the edge points did not belong to the ellipse. These extra edge points in the image tripled the orig- inal calculation compared to experiment 1. The Standard Hough Transform and Directional information for parameter space decomposition had the highest accuracy in the ten images provided. As expected, Standard Hough Transform was the slowest and the Directional information for parameter space decomposition was the quickest. It should be noted that the accuracy of Randomized Hough Transform and Probabilistic Hough Transform have further deteriorated. The results of the third experiment are summarized in Table 2.3 Table 2.3: Experiment3 - Single synthetic ellipse detection using figure 2.3. Number of Trials Number of suc- cessful ellipse detection Percentage of suc- cessful rate Average time in seconds SHT 10 10 100% 1220 PHT 10 6 60% 244 RHT 10 2 20% 42 PSD 10 10 100% 2 2.5 Experiment using synthetic ellipse with a rectangle, triangle, and a letter T, with one percent noise In experiment 4, random noise was added to the images to further distract each algorithm from detecting the ellipse as illustrated in Figure 2.4). Figure 2.4: Experiment4 - Single synthetic ellipse, rectangle, triangle and a letter T with one percent noise. 18 Adding the 1% noise has made the ellipse detection more difficult and the computation longer. There were a total of ten images used in the fourth experiment. On average, there were a total of 594 edge points in the image. Only 77 edge points belonged to the ellipse. Over 87% of the edge points did not belong to the ellipse. The Standard Hough Transform and Directional information for parameter space decomposition had the highest accuracy in ellipse detection in the ten images provided. Computational time was longer than ever. Both the Randomized Hough Transform and Probabilistic Hough Transform performed poorly in the fourth experiment. The summary of the fourth experiment is illustrated in Table 2.4. Table 2.4: Experiment4 - Single synthetic ellipse detection using figure 2.4 with one percent noise. Number of Trials Number of suc- cessful ellipse detection Percentage of suc- cessful rate Average time in seconds SHT 10 7 70% 3290 PHT 10 1 50% 630 RHT 10 0 0% 42 PSD 10 10 100% 5 2.6 Summary of the four experiments As percentage of edge points of the ellipse diminished, the probabilistic models were found to perform poorly. This can be explained by the difficulty encountered in the random sampling. Sampling the right ellipse edge points was more difficult than ever, because of the low ellipse edge points presence. This observation can be used to explain the results provided in Table 2.5. To detect an ellipse, edge points belonging to the ellipse need to be used. Random sampling edge points belonging to the ellipse do not guarantee an accurate detection. For instance, this is true given that the edge points belonging to the ellipse were chosen in close proximity to each other as illustrated in Figure 2.5. Given that edge points belonging to the ellipse were chosen in close proximity to each other as seen in Figure 2.5, the partial ellipse detection can be observed in Figure 2.6. This phenomenon can be explained by the lack of information about the ellipse itself such as the ellipse center, semi-major axis, semi-minor 19 Table 2.5: Noise and edge influence in the experiments. Number of Edge points belonging to ellipse Total number of edge points % of edge points belong- ing to ellipse % of successful ellipse detection in SHT % of successful ellipse detection in PHT % of successful ellipse detection in RHT % of successful ellipse detection in PSD Experiment 1 74 74 100% 100 100 100 100 Experiment2 74 168 44% 100 70 50 100 ExperimenG 74 227 32% 100 60 20 100 Experiment4 74 594 12% 70 50 0 100 Figure 2.5: Close edge selected for the random sampling. axis and ellipse orientation. Therefore, ideally, to raise the accuracy of ellipse detection, the points selected should be equidistant apart from each other in the perimeter of the ellipse. This ensures accurate ellipse detection. Another observation made while performing these four experiments in the Standard Hough Transform and Directional information for parameter space decomposition was the computational time. As more and more edge points were found in the image, the computation time increased linearly. When the number of edge points doubled in the image, the computational time also doubled. 2.7 Experiment using real world images The images used in this experiment can be divided into four categories: a normal eye illustrated in Fig- ures 2.7a-2.10a, an eye wearing glasses as illustrated in Figures 2.11a-2.14a and an eye with a different 20 Figure 2.6: Ellipse formation using close edge points (a) Original (b) SHT (c)PHT (d) RHT (e) PSD Figure 2.7: Results of each algorithm for eye 1 orientation as illustrated in Figures 2.15a-2.18a. The results of each algorithm can be seen side-by-side to each other. Figures 2.7b-2.18b show the results for the Standard Hough Transform. Figures 2.7c-2.18c show the results for the Probabilistic Hough Transform. Figures 2.7d-2.18d shoe the results for the Randomized Hough Transform. And finally, Figures 2.7e-2.18e show the results for Parameter Space Decomposition. (a) Original (b) SHT (c) PHT (d) RHT (e) PSD Figure 2.8: Results of each algorithm for eye 2 21 (a) Original (b) SHT (c) PHT (d) RHT Figure 2.9: Results of each algorithm for eye 3 (e) PSD (a) Original (b) SHT (c) PHT (d) RHT Figure 2.10: Results of each algorithm for eye 4 (e) PSD (a) Original (b) SHT (c) PHT (d) RHT Figure 2.11: Results of each algorithm for eye 5 (e) PSD (a) Original (b) SHT (c) PHT (d) RHT Figure 2.12: Results of each algorithm for eye 6 (e) PSD (a) Original (b) SHT (c) PHT (d) RHT Figure 2.13: Results of each algorithm for eye 7 (e) PSD 22 (a) Original (b) SHT (c) PHT (d) RHT Figure 2.14: Results of each algorithm for eye 8 (e) PSD (a) Original 1^ * £€%&, V ;»**Pfi|i|||f|SKPf«5' .-* '-5: ft m •.•? (b) SHT (c) PHT (d)RHT (e) PSD Figure 2.15: Results of each algorithm for eye 9 - : 'y 'feBRi* m • * « ^ : ? o * c » ,J*. pf*»%, (a) Original (b) SHT (c) PHT (d) RHT Figure 2.16: Results of each algorithm for eye 10 (e) PSD (a) Original (b) SHT (c) PHT (d) RHT Figure 2.17: Results of each algorithm for eye 11 (e) PSD (a) Original (b) SHT (c) PHT (d) RHT Figure 2.18: Results of each algorithm for eye 12 (e) PSD 23 In Table 2.6, the calculation time required for each image has been recorded. The average calculation time for each algorithm is also provided at the end. The Standard Hough Transform had the slowest average calculation time. The Parameter Space Decomposition had the quickest. The Randomized Hough Transform was omitted, since, in most pictures, no ellipse was found. Looking through Figures 2.7-2.18, it was found that the Standard Hough Transform and Probabilistic Hough Transform had provided reasonable accuracy. In the Probabilistic Hough Transform, the result was impressive, considering that only a fraction (20%) of the edge points was used. In the Randomized Hough Transform, the result of the ellipse detection has been quite disappointing. This was due to the fact that there were too many edge points not pertaining to the ellipse. In most cases, no ellipse was found. The Random- ized Hough Transform provided the poorest detection/accuracy. The Parameter Space Decomposition was found to have the quickest calculation and best accuracy. Table 2.6: Calculation time required in the SHT, PHT, RHT and PSD for eyel-12 Image Eyel Eye2 Eye3 Eye4 Eye5 Eye6 Eye7 Eye8 Eye9 Eye 10 Eye 11 Eye 12 Average time SHT 614sec 610sec 630sec 615sec 1211sec 834sec 938sec 741 sec 588sec 724sec 627sec 638sec 730sec PHT 123sec 126sec 145sec 130sec 219sec 156sec 178sec 140sec 138sec 173sec 148sec 166sec 154sec RHT 386sec 416sec 357sec 360sec 130sec 99sec 105sec 98sec 92sec lOOsec 95sec 88sec 193sec PSD 4.07sec 4.67sec 4.1sec 4.95sec 7.72sec 5.11sec 5.69sec 4.17sec 4.53sec 5.66sec 4.84sec 4.60sec 5sec 2.8 Comparison and discussion of the Hough based algorithms In Table 2.7, major iterations were calculated in BigOh notation for each algorithm. The major iterations include computing the Hough Transform and locating the highest accumulated value in the accumulator. Memory requirements for each algorithm were also calculated. 24 Table 2.7: Number of major iterations and memory requirements for the Standard Hough Transform, Prob- abilistic Hough Transform, Randomized Hough Transform and Parameter space decomposition Standard Hough Transform Probabilistic Hough Transform Randomized Hough Transform Parameter space de- composition Big Oh notation in terms of major iter- ation loop where M is the total number of edge points m[a]i[b]j[:Eo]fc[0]i, + M i M j M f c M u where m < M and m is the total ran- domly selected point and M is the total number of edge points P pei ' ( M \ (MX centage of points con where P is a ibination P ( f ) [xo]k + MiMMv + N)]fcM« +[a]i [%[</>]« where P is a percentage combination of edges Accumulator storage re- quirement [a] i[%[x0]fc[yo] u[4>]v [a]i[b)j[xo]k[yo]u[(l>]v '(?) [xo}k[yo]u+[a}i[b]j[(t>]v • Figure 2.19: A normal eye In Figure 2.19, there were in total 454 edge points. Only 133 edge points belonged to the eye contour. There were 321 irrelevant points, meaning more than 70% of the points were not needed. These irrelevant edge points contributed to unnecessary calculation in the Standard Hough Transform. Eliminating these irrelevant edge points sped up Standard Hough Transform calculation, at the same time as removing the false ellipse detection. Additionally, these irrelevant edge points also played an important role in the Randomized Hough Trans- form. As the number of irrelevant edge points increased, the precision and the successful ellipse detection rate for the Randomized Hough Transform decreased. This was due to the fact that there were too many irrelevant edge points in the image. Therefore, randomly selecting points belonging to ellipse became more difficult. Hence, the Randomized Hough Transform has more difficulty converging to a possible ellipse 25 parameter set. Using eye contour edge points only improves part of the computational time. The other part of the improvement comes from how each of the accumulated cell was constructed. Here is an example: Let [a]i be all the possible values of the semi-major axis, let [b]j be all the possible values of the semi- minor axis, let [xo]k be all the possible values of the ellipse center in x-component, let [yo]u be all the possible values of the ellipse center in y-component and let [<f>]v be all the possible values of the ellipse orientation. A normal accumulator can be constructed as follows: Accumulator[a]i[b]j[xo}k[4>]v Where [a]i+i - \a]i - 1 [b]j+1 - [% = 1 N]fe+i - txo]fc = i [<f>]v+l - [4>]v = 1 Having the accumulator cell spaced at 1 for every Hough Space parameter xo-yo-a-b-tfi provides the following number of loop operations. M[a]i[6]j[zo]fc[<£k' + Mi[%[zo]fcMi> where M is the number of edge points. The first term M[o]j[fr]j[io]ifeMij is for collecting votes and the second term M[a],[%[a:o]fc[</>]i> is for searching for the highest accumulator cell with the most accumulated votes. If the accumulator cell is spaced at 0.5 for every Hough Space parameter, as illustrated next: 26 H i + i - [a]i = ° - 5 [b]j+1 - [b]j = 0.5 [xo)k+i - [xo)k = 0.5 \4>)v+i - [4>}v = 0.5 Then the number of iteration loops increases by 16. leMlalif&kbolfeM., + 16[a]t-[&fc[a:o]fc \4>]v Hence, the accuracy of each accumulator cell plays an important role in the computational time of any Hough based Transform. In the Probabilistic Hough Transform, every edge point contribution to the final voting accumulator set depends upon whether the edge point is being selected to vote. This is a good way to reduce unwanted edge points since over 70% of the edge points were irrelevant. In our Probabilistic Hough Transform experiment, we randomly selected 10% of the edge points or 45 points from Figure 2.19. Using only 14 edge points in the eye contour, the Probabilistic Hough Transform was able to accurately determine an ellipse. The only drawback with the Probabilistic Hough Transform is that if points were not selected carefully, the ellipses that were not the highest peaks will be chosen. Similarly, if there are too many edge points not belonging to the ellipse in the image, locating the ellipse becomes more difficult. In the Randomized Hough Transform, the primary iteration loop can be found only through the different combination set of the 3-point selection process in pi, p2 and p i . Selecting points belonging to the ellipse / 454 \ has turned out to be difficult using Figure 2.19 since there are or 15493203 combinations in total \ 3 / and only 383306 combinations of 3-points belonging to the ellipse. Only 2.5% of the 3-point combinations belong to the ellipse. 4532969 iterations can be removed if eye contour points are selected, or over than 97.5% of the iteration is unnecessary. 4532969 accumulator cells can be removed with a reduction of memory consumption by 97.5%. Locating the maximum accumulated cell is also quicker, as a result. 27 CHAPTER D New proposed algorithm 3.0.1 Introduction Many real world images contain elliptical shapes. Elliptical shape detection has always been a key problem in computer vision and pattern recognition, especially for real-time applications like human face detection [26], iris detection [30] and driving assistance [9] [6]. Most ellipse detection approaches falls under four categories: Symmetry-based [28] [17], Random sam- pling [3], Genetic algorithm [27] and Hough-based transform [4] [15] [23] [1]. In symmetry-based detection, the ellipse geometry is taken into account. The most common elements used in ellipse geometry are the el- lipse center and axis. Using these elements and edges in the image, the ellipse parameters can be found. In random sampling methods, samples must satisfy a time bound and also a given probability density based on a particular geometry. After sampling, the edge points are used to determine ellipse parameters [3], In Genetic algorithms, the points in the image are divided into smaller subsets. The random samples taken from the subset are used to create a particular ellipse parameter set. A cost function is used to evaluate the presence of an ellipse by counting the number of points in the proximity of the ellipse perimeter. High points presence close to the ellipse perimeter indicates the presence of an ellipse. This cost function acts as an equivalent of an accumulator in the Hough transform. Two subsets (parents) are chosen to evoluate/crossover/mutate based on their cost function value to produce two additional subsets (offsprings). These offsprings are used to replace their parents in the subsets. A solution is found when all subsets converge to a particular ellipse parameter set. Hough-based transform is easy to implement. It uses edges in the image along with different ellipse parameters to create votes for the accumulator. The highest accumulated cell indicates the presence 28 Chapter 3. New proposed algorithm of an ellipse. Hough-based transforms vaguely fall under three different variations. There is the classic standard Hough transform [4]. It is a robust method used to detect ellipses. However, due to the heavy computational time and large amount of storage area requirement, ellipse detection under the standard Hough transform is impractical. Moreover, it may result in innacurate estimates especially in the case of noisy images. The second type is the probabilistic Hough transform [15]. This algorithm seeks to reduce the number of redundant votes in the accumulator. This reduction is accomplished by reducing the edges across the image evenly. The third type is parameter space decomposition [24] [8]. In this method, calculations are split into multiple stages where output in the accumulator in the first stage is used as an input in the next stage. Furthermore, improvement in the original parameter space decomposition is accomplished by the edge directional information level [1]. A comprehensive survey of different Hough-based transforms can be found in [16]. 3.0.2 Recent works Recent developments on the Hough-based transform in [33] have proposed removing spurious points using edge curvature convexity. Edges that do not converge to the surrounding group edges curvature are then removed. This process dramatically reduces the number of edges to be used in any Hough-based transform. Keeping only edges belonging to the curvatures should help locating the ellipse parameters values accurately. In the method proposed in [33], ellipse detection consists then of four stages: removing spurious edge points in the first stage, locating the ellipse center using an accumulator (xo,yo) in the second stage, locating the ratio N of semi-minor axis b over the semi-major axis a and the tangent of ellipse orientation tan(0) using a second accumulator (N,K) in the third stage, where K = t a n ( ^ ) . Rewriting the second accumulator in terms of a, b and </>, the new second accumulator becomes (a/6,tan((/>)). Finally, using a/b and t a n ( ^ ) in the third stage, combined with the ellipse center (xo,j/o)to indirectly locate a, b and 0 through N and K in the fourth stage. This method suffers mostly from input error propagation, where inaccurate results are passed on to detect a, b and <j> due to the error introduced by indirect quantization. The main source of inaccuracy is situated in the second accumulator (N,K) where ellipse parameters a, b and </> are not quantized directly. Ellipse orientation is computed from K. Since K is a quantized value or more precisely 29 Chapter 3. New proposed algorithm an approximation, the orientation <j> will be erroneously calculated. For instance, given the real value of if to be 1.5, the theoretical ellipse orientation value <f> yields 56.31 degrees. If the quantization of 1 is to be used between accumulator cells, the original K will be rounded off to 2. This will yield 63.43 degrees. Narrowing down the quantization cells will only increase the additional memory requirement for the accumulator. If the direct parameter quantization of 1 was applied to the ellipse orientation, the ellipse orientation will only be off by 1. In other words, 56.31 degrees will be rounded off to 56 degrees. By the same explanation, b will be inaccurately calculated given that a is provided and N is retrieved from the highest accumulator cell. Finally, it can be seen in the experimental results in [33] that the values of the semi-major axis and semi-minor axis fall short in terms of accuracy. Other recent developments have focused on improving Randomized Hough transform (RHT). The au- thors in [19] demonstrated that RHT is easily influenced by noise in the image, so an iterated algorithm is proposed by narrowing down a search area in each iteration to remove noise. By focusing on a specific area, the edges belonging to the ellipse have a higher percentage chance of being selected randomly. Higher selection of ellipse edges leads to higher ellipse detection. However, if ellipse edges were not included in the first iteration, ellipse detection becomes difficult. In addition, what happens when noise and other unrelated edges are actually inside the ellipse? Zooming-in to a specific area will not help reduce unrelated edges or noise. Therefore, narrowing down a specific area might not work as originally intended. Furthermore, at each iteration, votes are reset and recollected. This leads to the same combination of edges voting multiple times throughout the iterations. Other proposed algorithms involve extracting line segments at the pixel level [21] and then linking together all the potential line segments to form arc segments. Arc segments of the same ellipse are then grouped together. This method falls short in terms of detecting small ellipses, since the accuracy of multiple arcs cannot be guaranteed. In general, most Hough-based transforms face parameter quantization inaccuracy. This inaccuracy not only affects the ellipse parameter values but also disperses votes into other accumulator cells. Therefore, to improve ellipse accuracy and memory consumption, quantization must be avoided and original edges should be used instead to compute ellipse parameters. However, quantization inaccuracy is not the only issue that faces Hough-based transforms. The computation slowness for Hough-based transforms is caused by the number of different combinations in the parameter space and the number of edges to be used in the image 30 space. This different combinations adds extra memory requirement in the accumulator. In this paper, we propose a method to minimize the number of edges to be used in the image space and at the same time to eliminate the use of parameter space. This improves memory consumption and ellipse detection time. The proposed method uses the approach in [1] to find the ellipse center using different combinations of two points with different values of xo in the image space. These two points were separated by a distance apart as described in [1]. However, instead of searching and solving for the rest of the ellipse parameters, the proposed method stores edges that voted for the ellipse center. Direct least square fitting of ellipse proposed in [5] is used to fit the stored edges. A two-dimensional parameter space will be required to store the ellipse location and one additional three-dimensional array to store edges that have voted for the ellipse center. Using only pairs of edges and the ellipse center, the recreation of the original ellipse can be accomplished and ellipse parameters determined using the Direct Least Square Ellipse fitting. The proposed method requires only a two edge accumulator voting system. The rest of this chapter is organized as follows. In the next section, new proposed algorithm is pre- sented. Section 3 is devoted to the experimental results comparison and analysis. Section 4 contains some concluding remarks. 3.1 Quantization-free parameter space reduction(QFPSD) The algorithm consists of 5 stages: Q Use two different points, pi(xi,yl) and P2{x2,y2) with different values of xo to locate the ellipse center (xo,j/o) a s described by [1]. Store p\ andp2 if they are equidistant to the ellipse center and increase the accumulator for the ellipse center. Otherwise, disregard them. Only a small sample of equidistant pairs is required. Q Locate the highest accumulated ellipse center cell and retrieve the edge points, that voted for the ellipse center. Q Use pi, p2 and the ellipse center to project the new additional points p[ and p'2. 31 • Using the Euclidean distance between ellipse points and the ellipse center, the normal distribution can be used to model the presence probability of ellipse edge points. Q Use filtered scattered edges to fit into Direct Least Square Ellipse fitting. These operations are described in details in each of the following subsection stages. 3.1.1 Locating the ellipse center using two points The procedure used to locate the ellipse center is the same as described in [1] (see figure 3.1). The ellipse center can be found using the following equation: Figure 3.1: Geometrical relationship between the two randomly selected points and their tangent line AC + 1BD. (i) where 32 A = yi-y2 B = X\ — X2 C = slopei + slope2 D = slopeislope2 ?/i -yt Xi 2/2 X2 -xt -yt -xt X2 +X\ 2 2/2 + 2/1 slopei — slope2 = Xm — ym = 3.1.2 Equidistant points to the ellipse center In our algorithm, only the pair of edge points that are equidistant to the ellipse, as illustrated in Figure 3.2, are stored in a three-dimensional array and allowed to vote. All other pairs are disregarded and their votes are also being omitted. Figure 3.2: p\ andp2 equidistant to the ellipse center (XQ, yo)- If the goal is to store as few points as possible, then only two pairs of unique edge points are required for a perfect half ellipse. Throughout our experiments, it has been proven that a single pair was more than enough to detect the ellipse. This is because by having the ellipse center and four edge points, it is possible to solve the ellipse equation. This pair of points should be located with an angle of | from any of the ellipse axis. 33 Normally, in real world images where ellipses are never perfect, a tolerance parameter has to be intro- duced to detect the elliptical form. This tolerance value can be described as the following where d\ and d.2 are distances from ellipse edges to the ellipse center. tolerance = \d\ — cfel The tolerance value will be zero for a perfect ellipse. For a real world elliptical shape, the tolerance value will be higher than zero. This value can be adjusted depending on the image. Furthermore, not all pairs of equidistant points are needed. Only a small portion of the pairs is needed. Pairs of points can be sampled at a specific interval. This is equivalent as to reduce the number of votes to a smaller scale in the Probabilistic Hough transform. 3.1.3 Locate the ellipse center with the highest accumulated cell and retrieve edges that have voted for the ellipse center In this stage, there are two main tasks. The first is to locate the ellipse center in terms of the xy coordinate with the highest accumulated cell by traversing a two-dimensional array: [zoMyoh- Once t n e location of the ellipse center is found, the second task consists of using the location of the ellipse center (xo,2/o) to retrieve the edges that have voted for the center. These edges can be stored in a three-dimensional array, as illustrated by the following: [xo] k b o h [numberOf Edges] 3.1.4 Points projected using p i , P2 and ellipse center Using the original pair of points p\,P2 and the ellipse center (xo,yo). additional points can be generated. The purpose of generating additional points is to save unnecessary storage space and add ellipse curve projection. This technique is illustrated in Figure 3.3. 34 Figure 3.3: Projected points from pi and p2. 3.1.5 Normal distribution Using the Euclidean distance d between the ellipse center and ellipse points, the presence of ellipse points can be modeled using the Normal distribution, as illustrated in Figure 3.4. In this stage, the goal is to remove a small quantity of points that do not belong to the ellipse. Points that are beyond the semi-minor and semi- major axis or points that are too close to the ellipse center are removed. Using a plus or minus one standard deviation, most points beyond the bell curve are rejected. 35 (a) Ellipse Tl* mraial feftihrfai rf Gxbmt bctweei etpsc pmb aid efipsc a pmrifedNtMCtfrM eifttttHtrlxO.yfll ttM-na|Bruits (b) Normal distribution Figure 3.4: Normal distribution of distance between ellipse points and ellipse center 3.1.6 Direct Least Square Ellipse fitting In this stage, the fitting of scattered data into the ellipse is performed using [5]. These scattered data are obtained after ellipse points are modeled using Normal distribution. The main idea is to minimize the sum of the distances of the scattered data from the ellipse curve. This method is especially efficient when data or spurious points to be fitted are in imperfect condition, as illustrated in figure 3.5a. The result of the scattered data fitting can be seen in figure 3.5b. 36 (a) Noisy ellipse (b) Ellipse fitting using Direct least square Figure 3.5: Direct least square fitting for ellipse 3.2 Experiments The experiments were conducted using a Dell D500 Centrino 1.3GHz with 1.5Gigabytes of memory running Matlab 2008a under Windows XP 32bit. The experiments were divided into two primary stages: validating the correctness of the new proposed algorithm using synthetic images and testing the algorithm in real world environment. In the validation process, random noise was added to the images to disturb ellipse detection. Unfortunately, synthetic images can only validate the correctness of the algorithm. In real world images, elliptical shapes are often never in perfect condition. Therefore, the new algorithm must be adaptable to poor elliptical shape. Not having a perfect ellipse is just one of many issues that can arise when detecting elliptical shape. In some cases, elliptical shapes might have other obstructing objects inside itself or partially hidden by the elliptical object itself. To conclude our real world experiments, the new algorithm was tested on traffic sign detection and eye detection. Both of these applications are primarily used in the automotive industry for real-time application. In the traffic sign detection, the detector can warn the driver against danger ahead if the driver has not being diligent on the road. Similarly, in respect of eye detection, the machine can warn the driver if not enough attention has been put on the road, or if fatigue is taking over. The experiments in the real world environment are carried out with comparisons against other Hough- based transforms, such as the Standard Hough transform (SHT), Probabilistic Hough transform (PHT), Randomized Hough transform (RHT) and Parameter Space decomposition (PSD). 37 3.2.1 Experiments with synthetic images Since the proposed algorithm relied on storing edges that voted for the ellipse center, in the next experiment, a demonstration revealing that not all edges are required is given in Figure 3.6. The basic idea is to store edges at a particular interval. In other words, stored edges are never close to each other. An interval of 50 has been used to demonstrate the accuracy of the ellipse detected in Figure 3.7. Figure 3.8 demonstrates that a single pair of edges might be enough to adequately locate the ellipse parameters. Figure 3.9 demonstrates the efficiency of each filter (equidistant and normal distribution) working side-by-side to remove unwanted non- pertinent edges. Overlapping synthetic ellipses were tested in Figure 3.10. Visually fitted ellipse theoretical parameter values and the detected ellipse parameter values using the proposed algorithm for figures 3.9 and 3.10 can be seen in tables 3.1 and 3.2. (a) all votes (b) interval of 10 (c) interval of 20 (d) interval of 30 (e) interval of 40 Figure 3.6: Sampling interval for the pairs of points 38 (a) interval of 50 (b) interval of 50 result Figure 3.7: Accuracy of ellipse using an interval of 50 (a) one pair vote (b) one pair vote with pro- jected points (c) one pair vote with pro- jected points result Figure 3.8: Accuracy of ellipse detection using one pair of points 39 I V-™w". .* I (a) all votes (b) equidistant votes I'M I (c) applying normal distribution (d) direct least square fitting Figure 3.9: Experimental results to demonstrate both filters Table 3.1: Theoretical values and QFPSR in figure 3.9d. Image Figure 3.9d Methods Visually fitted QFPSR Xo 33.8829 35.8783 yo 37.0013 37.0828 a 13.4027 14.5824 6 9.5305 9.9159 * -0.0128 -0.0072 A A (a) Occluded ellipse 1 (b) Occluded ellipse 2 Figure 3.10: Occluded ellipses Table 3.2: Theoretical values and QFPSR in figure 3.10. image Figure 3.10a Figure 3.10b Methods Visually fitted QFPSR Visually fitted QFPSR „ 42.1244 42.9877 46.4839 47.0384 yo 41.9764 41.6621 27.4097 25.5777 a 17.6388 16.4914 16.7158 13.8856 b 9.2941 10.4560 9.7942 10.3807 * -0.0007 -0.1108 1.5663 1.45 Since elliptical objects are found everywhere, elliptical shapes are put to test in Figures 3.11 and 3.12. 40 Clearly, in the results on both cases, our algorithm has performed relatively well with more precision. Edges not pertaining to the ellipse were filtered out using equidistant condition and normal distribution. (a) SHT (b) PHT (c) RHT (d) PSD (e) QFPSR Figure 3.11: Pink bicycle result 41 (a) SHT (b) PHT (c) RHT (d) PSD (e) QFPSR Figure 3.12: Wheels result 3.2.2 Experiments on elliptical shapes in real world environment In this section, our algorithm is tested in real world environment. Side-by-side results are illustrated in Figures 3.13, 3.14, 3.15 and 3.16 42 (a) SHT (b) PHT (c) RHT (d) PSD (e) QFPSR Figure 3.13: Left rear of a red volvo 850 43 (a) SHT (b) PHT (c) RHT (d) PSD (e) QFPSR Figure 3.14: Antique vase (a) SHT (b) PHT (c) RHT (d) PSD (e) QFPSR Figure 3.15: Magnifying glass 44 (a) SHT (b) PHT (c) RHT (d) PSD (e) QFPSR Figure 3.16: Hidden plate 3.2.3 Real world application: traffic sign detection Detecting the appropriate road warning sign can help drivers reduce accidents on the road. Different road signs are classified with different geometry shapes. Special signs that might require strict attention for drivers are "No Entrance", as illustrated in Figure 3.17, "No right turn", as illustrated in Figure 3.19, "No left turn", as illustrated in Figure 3.21 and "No U turn", as illustrated in Figure 3.23. Comparison of the results can be seen in Figures 3.18, 3.20, 3.22 and 3.24. The main purpose of this experiment was to test accuracy. To measure how accurate our algorithm is comparing to other Hough-based transforms, ellipses were visually fitted to create theoretical values. Only the highest accumulator cell was retrieved in this experiment. A summary of the experimental results can be found in Table 3.3. Through experimental results, it can be seen that our proposed algorithm had the best result in terms of accuracy and execution time. 45 (a) Original image (b) Edges (c) Visually fitted Figure 3.17: Traffic sign 1 (a) SHT (b) PHT (c)RHT (d) PSD (e) QFPSR Figure 3.18: Traffic sign 1 results (a) Original image (b) Edges (c) Visually fitted Figure 3.19: Traffic sign 2 46 (a) SHT (b) PHT (c) RHT (d) PSD (e) QFPSR Figure 3.20: Traffic sign 2 results (a) Original image (b) Edges (c) Visually fitted Figure 3.21: Traffic sign 3 47 (a) SHT (b) PHT (c)RHT ——jJIljM (d) PSD (e) QFPSR Figure 3.22: Traffic sign 3 results 48 (a) Original image (b) Edges (c) Visually fitted Figure 3.23: Traffic sign 4 49 (a) SHT (b) PHT '\^< Z&*\ (c) RHT (d) PSD (e) QFPSR Figure 3.24: Traffic sign 4 results 50 Table 3.3: Value comparison of theoretical, SHT, PHT, RHT, PSD and QFPSR Image Traffic Sign 1 Traffic Sign 2 Traffic Sign 3 Traffic Sign 4 Methods Visually titled SHT PHT RHT PSD QFPSR Visually fitted SHT PHT RHT PSD QFPSR Visually filled SHT PHT RHT PSD QFPSR Visually fitted SHT PHT RHT PSD QFPSR *o 81.2850 79 IS 66 82 81.5134 102.7677 92 17 68 102 I01J711 84.9302 30 21 25 85 83.6187 97.9580 52 16 1 94 96.9702 yo 47.5423 41 73 89 47 47.2630 39.0177 46 76 94 38 38.1045 83.1512 68 167 173 114 81.6666 102.8233 74 74 162 76 102.7695 a 15.6248 It 12 14 30 15.8250 17.2039 15 14 6) 24 15.3174 19.8230 11 14 47 30 18.2584 21.553! 15 15 81 30 17.7194 b 16.4715 10 11 4 3 17.1777 16.1118 13 10 9 18 17.5491 19.9808 10 13 33 4 17.7879 20.9893 14 14 45 3 20.2579 C -0.0348 -0.0873 -0.0768 0 -O.OI75 -0.9379 -0.0181 -0.2618 -0.0192 1 -0.0175 0.8238 -0.15S5 -0.3491 0 2 -0.0175 -0.2095 1.5606 -0.3491 -0.1920 1 -0.0175 -0.4772 time in N/A 9271 1854 7835 38 27 N/A 8164 1633 744 23 9 N/A 12470 2495 225 107 82 N/A 36710 7342 205 223 198 In reviewing our experimental results, I found that our algorithm has achieved results very close to the theoretical values. Most ellipse parameters values were found within 1%. This 1% error can be attributed to human error since the theoretical results were created visually. In addition, the algorithm was the fastest to finish up the ellipse detection. 3.2.4 Real world application: eye detection Eye detection presents a good scenario to test the proposed algorithm, since the eye does not have a perfect elliptical shape. Additionally, pupils, irises and sclera generate unwanted disturbance edges. In some of the pictures, eyeglasses frames were also included. Figures 3.25 to 3.36 were used to demonstrate the results of the proposed algorithm against other Hough-based algorithms. These pictures are truly challenging, since the upper left/right curvatures and lower left/right curvatures of the eyes do not necessary reflect to the same ellipse parameters. Locating the best fit ellipse to the eye can be a difficult task. The calculation time for each image is illustrated in Table 3.4. 51 (b)PHT (c)RHT (d)PSD (e)QFPSR Figure 3.25: Results of each algorithm for eye 1 (b)PHT (c)RHT (d)PSD (e)QFPSR Figure 3.26: Results of each algorithm for eye 2 (b)PHT (c)RHT (d)PSD (e) QFPSR Figure 3.27: Results of each algorithm for eye 3 (b)PHT (c)RHT (d)PSD (e) QFPSR Figure 3.29: Results of each algorithm for eye 5 52 (a) SHT (b) PHT (c) RHT (d) PSD Figure 3.30: Results of each algorithm for eye 6 (e) QFPSR (a) SHT (b) PHT (c) RHT (d) PSD Figure 3.31: Results of each algorithm for eye 7 (e) QFPSR (a) SHT (b) PHT (c) RHT (d) PSD Figure 3.32: Results of each algorithm for eye 8 (e) QFPSR -< s : - s « » f v « > WKKM - . m mmua (a) SHT (b) PHT (c) RHT (d) PSD Figure 3.33: Results of each algorithm for eye 9 M M (e) QFPSR iHSiS (b)PHT (c)RHT (d)PSD (c) QFPSR Figure 3.34: Results of each algorithm for eye 10 53 (a)SHT (b)PHT (c) RHT (d) PSD (e)QFPSR Figure 3.36: Results of each algorithm for eye 12 Table 3.4: Calculation time required in SHT, PHT, RHT, PSD and QFPSR for eyel-12 Image E y e l Eye 2 Eye 3 Eye 4 Eye 5 Eye 6 Eye 7 Eye 8 Eye 9 Eye 10 Eye 11 Eye 12 Average time SHT 614 sec 610 sec 630 sec 615 sec 1211 sec 834 sec 938 sec 741 sec 588 sec 724 sec 627 sec 638 sec 730sec PHT 123 sec 126 sec 145 sec 130 sec 219 sec 156 sec 178 sec 140 sec 138 sec 173 sec 148 sec 166 sec 154sec RHT 386 sec 416 sec 357 sec 360 sec 598 sec 456 sec 500 sec 398 sec 400 sec 501 sec 402 sec 412 sec 432 sec PSD 4.07 sec 4.67 sec 4.1 sec 4.95 sec 7.72 sec 5.11 sec 5.69 sec 4.17 sec 4.53 sec 5.66 sec 4.84 sec 4.60 sec 5 sec QFPSR 2.31 sec 2.74 sec 2.27 sec 2.89 sec 4 sec 2.38 sec 2.47 sec 1.98 sec 2.28 sec 3.01 sec 2.60 sec 2.49 sec 2.62 sec 3.2.5 Analysis of Quantization-free parameter space reduction In the new proposed algorithm, there were basically two filters to remove unwanted non-ellipse pertinent edges: one filter focus on the individual edge-pair level and the other focus on the group edge level. In the first filter, individual pairs of edges were targeted and rejected if the equidistant condition was not met. In the second filter, edges that were too far away compared to the average distance to the ellipse center were 54 completely removed in order to improve ellipse accuracy. Additionally, the new proposed algorithm did not add any major loops to the elliptical shape detection. The complexity stayed at cO(n2) where c is a fraction. In terms of memory requirements, only a two dimensional accumulator was needed to store the ellipse center. A three dimensional array was needed to store the edge points from the image. Using the Direct Least Square fitting of the ellipse, the algorithm itself could validate the ellipse center (xo, yo) and eliminate errors in the parameter space quantization. In an elliptical shape where the ellipse is not perfect, the Direct Least Square fitting of the ellipse can be used to best describe the ellipse by minimizing the distance between ellipse noise edges. When comparing the new proposed algorithm with directional information parameter space decompo- sition, the new proposed algorithm had few advantages. In the proposed algorithm, once the ellipse center was located, ellipse parameters were found by fitting into the Direct least square method. Meanwhile for parameter space decomposition, additional combinations of ellipse parameters were required to locate the rest of the parameter values. The number of additional loops was proportional to the number of ellipses found. The new proposed method also eliminates input error propagation in the original Parameter Space De- composition. This input error propagation affects values of semi-major axis a, semi-minor axis b and angle 4> of the ellipse orientation. Additionally, the values of semi-major axis a, semi-minor axis b and angle cf> of the ellipse orientation were quantized. In the current proposed algorithm, no ellipse parameters were quantized. In terms of accuracy, unlike any Hough-based transform, the new proposed algorithm uses original edges. This translates into higher accuracy, as illustrated in the experiments. The accumulator for the ellipse center was only used to locate and collect evidence in order to decide which edges were needed in order to be fed into direct least square method. Therefore, no parameters of the ellipse were actually quantized. Quantized accumulator cells only led to approximated values for the ellipse parameters. Additionally, to detect the ellipse, the range of values of the semi-major axis a, semi-minor axis b and angle (j> of ellipse orientation must be given the opportunity to vote. If, for a particular reason, the range of these values were skipped, accuracy on the ellipse is compromised, or failure in ellipse detection will occur. In any Hough-based transform, the local maxima in the accumulator are often surrounded by other 55 closed values, if they are not equal to the local maxima itself, as illustrated in Figure 3.37. This phenomenon occurs when ellipse edges are noisy. Because Direct Least square methods have been used to locate ellipse parameters, the center local maxima are truly found. Figure 3.37: Multiple local maxima 56 CHAPTER T* I Conclusions The most popular method for ellipse detection in digital imagery is currently the Hough Transform. The weaknesses in Hough-based transforms include inaccuracy due to quantization and memory consumption due to five-dimensional parameter space. As a result, we have proposed a new approach for ellipse detection. Unlike any of the previous Hough-based transforms, where five-dimensional parameters space is required, in the proposed algorithm, the ellipse parameter values were located by reconstructing the original ellipse using a pair of edges and the ellipse center. This method has demonstrated that a five-dimensional parameter space is no longer required by an ellipse. Through rigorous testing using a large number of experimental results involving synthetic images and different real world applications, such as traffic sign detection and eye detection, our proposed algorithm has obtained higher accuracy, lower memory consumption and quicker calculation times. Unlike previous generations of Hough-based transforms, the focus was based on locating quantized parameter values, the new algorithm focused on the edges themselves. Any edges that have the potential of being part of an ellipse were retained for locating the ellipse parameter values. Because focus was placed on the edges themselves, accuracy on the ellipse has greatly improved. Also, since only edges belonging to the ellipse were kept, due to the filters equidistant and normal distribution, the proposed algorithm was really robust to noise. Furthermore, the new algorithm, did not add any additional major iterations to the original parameter space decomposition algorithm. The complexity was kept to cO(n2) where c is a fraction. As for the memory consumption for storing the edges, only a small portion was required. Results have been proven to be quite accurate. 57 Chapter 4. Conclusions In the future works, additional improvement can be done regarding the number of edges to be stored for the potential ellipse based on the ellipse center. Ideally, edges should be stored evenly across the ellipse perimeter. This additional improvement can improve greatly the detected ellipse parameters. Additionally, effective edges storage can improve greatly memory consumption. Furthermore, technically speaking, hav- ing two edges pi and P2, the projected edges p[ and p'2 and the ellipse center, ellipse equation should be solvable without the need of Direct Least Square method. 58 List of References [1] Alberto S. Aguado, Eugenia, and Mark S. Nixon. On using directional information for parameter space decomposition in ellipse detection. Pattern Recognition, 29(3)369-381, March 1996. [2] J. Batista. A drowsiness and point of attention monitoring system for driver vigilance. In Intelligent Transportation Systems Conference, pages 702 - 708, 2007. [3] Y.Y. Tsai C M . Wang, N.C. Hwang and C.H. Chang. Ellipse sampling for monte carlo applications. Electronics Letters, 40(l):21-22, 2004. [4] Richard O. Duda and Peter E. Hart. Use of the hough transformation to detect lines and curves in pictures. Communications of the ACM, 15(1): 11—15, 1972. [5] Andrew Fitzgibbon, Maurizio Pilu, and Robert B. Fisher. Direct least square fitting of ellipses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(5):476-480, 1999. [6] M.A.; Martm-Gorostiza-E. Garcia-Garrido, M.A.; Sotelo. Fast traffic sign detection and recognition under changing lighting conditions. Intelligent Transportation Systems Conference, 17-20:811 - 816, 2006. [7] Nikos Grammalidis and Michael G. Strintzis. Head detection and tracking by 2-d and 3-d ellipsoid fitting. In CGI '00: Proceedings of the International Conference on Computer Graphics, page 221, Washington, DC, USA, 2000. IEEE Computer Society. 59 References [8] N. Guil and E. Zapata. Lower order circle and ellipse hough transform. Pattern Recognition, 30:1729- 1744, 1997. [9] F. Fraunhofer EDMT-Ilmenau Hardzeyeu, V. Klefenz. On using the hough transform for driving as- sistance applications. In 4th International Conference on Intelligent Computer Communication and Processing, pages 91-98, 2008. [10] Ric Heishman, Zoran Duric, and Harry Wechsler. Using eye region biometrics to reveal affective and cognitive states. Computer Vision and Pattern Recognition Workshop, 5:69, 2004. [11] P.V.C. Hough. Method and means for recognizing complex patterns. U.S. Pattent 3069654, 1962. [12] S. A. Inverse Ellipse detection using randomized hough transform. http.V/www.saminverso. com/res/vision/EllipseDetection.pdf, May 2006. [13] Qiang Ji and Xiaojie Yang. Real-time eye, gaze, and face pose tracking for monitoring driver vigilance. Real-Time Imaging, 8(5):357-377, 2002. [14] M. Imran Khan and A. Bin Mansoor. Real time eyes tracking and classification for driver fatigue detection. In ICIAR '08: Proceedings of the 5th international conference on Image Analysis and Recognition, pages 729-738, Berlin, Heidelberg, 2008. Springer-Verlag. [15] N. Kiryati, Y. Eldar, and A. M. Bruckstein. A probabilistic hough transform. Pattern Recognition, 24(4):303-316, 1991. [16] V.F. Leavers. Survey: Which hough transform? CVGIP image understanding, 58(2):250-264, September 1993. [17] Yiwu Lei and Kok Cheong Wong. Ellipse detection based on symmetry. Pattern Recognition Letters, 20(l):41-47, 1999. [18] Xia Liu, Fengliang Xu, and K. Fujimura. Real-time eye detection and tracking for driver observation under various light conditions. Intelligent Vehicle Symposium, 2002. IEEE, 2:344-351 vol.2, 2002. 60 http://http.V/www.saminverso References [19] Wei Lu and Jinglu Tan. Detection of incomplete ellipse in images with strong noise by iterative randomized hough transform (irht). Pattern Recognition, 41(4):1268—1279, 2008. [20] E. Lutton and P. Martinez. A genetic algorithm for the detection of 2d geometric primitives inimages. Proceedings of the 12th IAPR International Conference on Computer Vision and Image Processing, Pattern Recognition., 1:526-528, 1994. [21] F. Mai, Y. S. Hung, H. Zhong, and W. F. Sze. A hierarchical approach for fast and robust ellipse extraction. Pattern Recognition, 41(8):2512-2524, 2008. [22] Tom Mainzer. Genetic algorithm for traffic sign detection. Applied Electronic, 2002. [23] Robert A. Mclaughlin. Randomized hough transform: improved ellipse detection with comparison. Pattern Recognition Letters, 19(3-4):299-305, 1998. [24] Derek Pao, H. F. Li, and R. Jayakumar. A decomposable parameter space for the detection of ellipses. Pattern Recognition Letters, 14(12):951-958, 1993. [25] G. Piccioli, E. De Micheli, P. Parodi, and M. Campani. Robust road sign detection and recognition from image sequences. Intelligent Vehicles '94 Symposium, Proceedings of the, pages 278-283, 1994. [26] A. Pietrowcew. Face detection in colour images using fuzzy hough transform. Opto-Electronics Re- view, ll(3):247-251,2003. [27] G. Roth and M. D. Levine. Geometric primitive extraction using a genetic algorithm. IEEE Trans. Pattern Anal. Mach. Intel!., 16(9):901-905, 1994. [28] Chun ta Ho and Ling-Hwei Chen. A fast ellipse/circle detector using geometric symmetry. Pattern Recognition, 28(1): 117-124, 1995. [29] Yong-Hui Huang Bao-Chang Pan Sheng-Lin Zheng Jianjia Pan Yuanyan Tang. Lip-reading detection and localization based on two stage ellipse fitting. In Wavelet Analysis and Pattern Recognition, 2008. ICWAPR '08. International Conference on, volume 1, pages 168-171, Hong Kong, 2008. 61 References [30] Klaus Toennies, Frank Behrens, and Melanie Aumhammer. Feasibility of hough-transform-based iris localisation for real-time-application. Proceedings on 16th International Conference on Pattern Recognition, 2:1053-1056, 2002. [31] L. Xu, E. Oja, and P. Kultanen. A new curve detection method: randomized hough transform (rht). Pattern Recognition Letters, 11(5):331—338, 1990. [32] H. K. Yuen, J. Illingworth, and J. Kittler. Detecting partially occluded ellipses using the hough trans- form. Image and Vision Computing, 7(1):31—37, 1989. [33] Si-Cheng Zhang and Zhi-Qiang Liu. A robust, real-time ellipse detector. Pattern Recognition, 38(2):273-287, February 2005. 62 
work_ospbi4ykq5hghekbizm3ul2aqi	----	Microsoft Word - revista 1-2015 final.doc Annals of “Dunarea de Jos” University of Galati Fascicle I. Economics and Applied Informatics Years XXI – no1/2015 ISSN-L 1584-0409 ISSN-Online 2344-441X www.eia.feaa.ugal.ro A Formal Definition for Expert Systems used in Real- time Applications Vasile MAZILESCU A R T I C L E I N F O A B S T R A C T Article history: Accepted April 2015 Available online May 2015 JEL Classification C63, M15 Keywords: Intelligent Knowledge Management System, Formal definition of an Expert System The present paper is situated on the grounds of research in the field of symbolic Artificial Intelligence systems, applied to the new Knowledge Management Systems. The basic feature of these systems is represented by the processing of the fuzzy knowledge involved in the synthesis of some decisions. The work reported in this paper serves to promote an Intelligent System that can operate in dynamic and uncertain environments based on a formal definition of an expert system. We can develop and justify thus a series of modelling and design techniques for Intelligent Knowledge Management Systems (IKMS), as well as methods for the analysis of expert systems performance, and, of a fuzzy expert system in particular, between which there are strong similarities. We will also outline a number of differences between conventional problem solving systems and IKMS, the links between expert systems and those of structural and functional planning, the analogy between the model of the problem or business process and the problem domain. © 2015 EAI. All rights reserved. 1. Introduction There are highlighted the ratios between conventional and intelligent management, by the use of planning and multi-agent systems. Knowledge-based systems cu real-time functioning bear features that the majority of classic system do not have: reasoning is evolutionary and non-monotonous due to the dynamic nature of the application, and events can change the status of the expert management system. The management architectures based on symbolic techniques acquire characteristics specific to the domain of the problem and to the type of expert system within the management structure. Planning is a difficult problem and may become much simpler if certain restrictions are applied. Different planning problems may be defined by restricting the type of operators, imposing a series of limitations to the number of preconditions and post conditions. Knowledge-based knowledge management (KBKM) focuses on applications of knowledge-based systems (KBS) tailored to knowledge management (KM) problems. The terms express the use of KBS to enable knowledge management (Bhatt, 2000). KM practitioners and research scientists have been implementing various frameworks to address pragmatic KM problems reusing decades of technology developed for KBS. Therefore, when we talk about knowledge-based knowledge management, we talk about the overlap of knowledge-based systems and knowledge management. It is easy to understand the scope of KM by focusing on a knowledge process that collects, stores and reuses knowledge leveraging it and making organizational the knowledge that once was individual. This knowledge process supports organizational goals by controlling the collection, storage, and use of knowledge (Bhatt, 2001). Accordingly, KM applications can be envisioned along the dimensions of a knowledge process that fundamentally performs knowledge tasks to support, steer and control organizational goals. Expert systems (ES), case-based reasoning (CBR), and ontologies are examples of relevant knowledge-based methodologies that have much to contribute to KM systems because they manipulate knowledge to implement various tasks. Although KM practitioners frequently comment that KM is not a technology problem, it is also the case that most KM solutions include an element of technology. It would seem that using technologies for collecting, storing, and reusing knowledge would be critical to consider in any KM effort. Therefore, it is natural that the understanding of a knowledge process differs in different organizations; e.g., a knowledge process in one organization has a different set of tasks than in another. Sometimes these knowledge processes differ on the surface, though their conceptual model is the same. It is difficult to distinguish a knowledge process from KM and that may be the reason why there are so many definitions on KM as a discipline. On the one hand we represent a knowledge process as a cycle – one in which KM manifests itself but can we also say that each of the tasks that are part of a knowledge cycle represent a kind of KM. We can describe what role they play in the knowledge process and we can also describe how they support the tasks of the knowledge process (another form of KM) (Mazilescu, 2012). Faculty of Economics and Business Administration, “Dunarea de Jos” University of Galati, Romania. E-mail address: vasile.mazilescu@ugal.ro (V. Mazilescu) 28 Expert systems represent one way that expertise can be captured, coded, and reused. They are embedded in IKMS and are based on Artificial Intelligence techniques. Fundamentally, an expert system consists of some representation of expertise, some representation of a problem to be solved, and some mechanism to apply the expertise to a problem. Although expertise may be represented in various forms, one common representation for expertise is in the form of rules. The representation of an expert’s knowledge would be a collection of rules that were derived from the expert. A rule consists of two parts: an antecedent and a consequent. The rule’s antecedent consists of one or more conditions that specify when and where to apply the rule. If the conditions of the rule are met, then the second part of a rule – the consequent – specifies the actions to take when the conditions of the rule are met. The mechanism that chooses rules to see if they can be used in a problem is called an inference engine. The inference engine checks antecedents of rules, and based on their values performs the actions specified in the consequents of the rules. To maintain the state of a problem being solved, the inference engine uses a special structure to store the state of problem solution. The structure is called short-term memory. Rules are scanned from the knowledge base to determine which apply to the current problem state in short-term memory. When the inference engine identifies a rule, the actions of that rule are carried out, which may result in a change to the problem state in short-term memory. The process repeats itself until it solves the problem, or no condition can be fulfilled or the expert system is explicitly stopped. There are in IKMS conception three large groups of problems that should approach in terms of decision-based applications: human-environment interface, qualitative knowledge modeling and time management. Such applications obviously require dated event operations the life-time of which should be managed by the system, which often works asynchronously with the acquisition and control system. A real-time expert system shell must also represent imprecise, time and temporal data, encode temporal knowledge, and manage temporal/fuzzy reasoning for allocating temporally interdependent tasks to homogeneous or heterogeneous cooperative agents in dynamic large-scale networks. Planning systems based on Artificial Intelligence use specific models for the problem field, called problem representations and logic argumentation models. Current expert systems have many characteristics in common with planning systems (representation of knowledge, heuristic inference strategies), conceived specifically for communicating with the exterior, while conventional expert systems are strongly encapsulated. Planning systems execute actions in a dynamic manner to produce modification in the state of the problem field. The planner monitors the problem field to progressively obtain information useful for decision synthesis. An explicit loop is being crossed between the actions done by the planner, the problem field, the measured outputs and the planner that uses the outputs to decide the control actions, with a view of reaching the goal. Within expert systems, a similar loop exists: the knowledge base represents the problem field, while the inference engine represents the planner. We have developed an IKMS as an agent that works like a planner. It is realized as a fuzzy real-time expert system shell to meet the challenges of the dynamic environment, like Virtual Organizations (VOs) or Hierarchical Coalitions (HCs). VOs provide an effective instrument to integrate a company’s operations with those of other enterprises, to work with customers and create a better product or service, to achieve a faster time to market, and to acquire a higher degree of product customization (Burden, 2009; Lin et al., 2007; Bergamaschi et al., 2006). Section 2 is a comparative analysis of planning capabilities for IKMS. At higher levels or at the much more abstract levels, every step of the plan has more effects than a step of the plan at a lower level. Section 3 is dedicated for a presentation of our IKMS based on logical events. Section 4 presents our conclusions and future developments. 2. Basic capabilities for IKMS A Multi-agent planning architecture (MPA) can be an option in the support systems for coalitions because it is a framework for the integration of various technologies into a system able to solve problems which cannot be solved by the independent systems. MPA is widely used in planning the applications, such as coalitions’ operations. MPA agents are organized based on the concept of cells planning. Each cell has a managing member that performs the planning of the cell from a group of available agents and distributes the tasks among selected planning agents. Sharing the information generated during the planning process will be held by a central plan, and each agent of the cell can access it. MPA is important for the integration of the different planning solutions. This integration has limits for the development of support systems coalitions, like any other approach that tries to integrate existing instruments and possibilities in a simple framework. The main reason is collaboration between problem solving on different components and the need to be embedded from the beginning into systems. The problem is how to incorporate collaborative requirements into a distribution of processes planning, so that planning the final connection will not be a sum of independent plans (Barachini, 1990; David et al., 2006; Jonsson et al., 2007). Considering that the most important function of the knowledge-based instruments is to support the human user, we must also understand how they interact in the planning the collaboration process. It should be noted that we do not take into account aspects of the interface which could improve the man-agent interaction. The main idea is that we can join up the groups, associated with the development of hierarchical coalitions as support for the agents, by means of a framework in line with an ontology based on constraints and on the appropriate 29 functions. We can argument a MAP framework brings several advantages such as: a well-known environment to represent and build plans; a transparent manner to integrate collaborative concepts that complement the planning skills; opportunities to develop the human-agent mechanisms; support for customizing the intermediaries. Based on this statement, we can have specific problems (Jonsson and Mattsson, 2006; Kim et al., 2003; Stoop and Wiers, 1996). An expert system is a type of planner, since it emulates the way in which experts make decisions, in the presence of abstract, qualitative and approximate knowledge, in a limited expertise field. Currently, what is necessary is the formulation of a mathematical theory of intelligent planning systems, which operates in a dynamic environment, with real time properties. The state of a planner or an expert system describes the situation that defines the problem solving strategy at a given time. A planning problem for a system may be defined thus: given an initial state, a desired state and an aggregate of possible actions, let it an aggregate of actions (either partially or totally ordered) be determined, which applied to the initial state leads the system to the desired state. Planning is a difficult problem and may become much simpler if certain restrictions are applied. Different planning problems may be defined by restricting the type of operators, imposing a series of limitations to the number of preconditions and post conditions. For these types of restrictions, the planning problem is polynomial or NP-complete. The computational complexity of planning was investigated. Many authors have been analyzed the general problem of deciding the existence of a solution for the planning task in the context of the STRIPS system and it is demonstrated that this general problem is PSPACE-complete (Bylander, 1994). The planner's outputs are inputs for the problem domain, representing control actions (Lunardhi and Passino, 1995). The outputs of the problem field are inputs for the planner. They are measured by a planner and used to determine evolution in the problem solving process. Furthermore, there are non-measurable exogenous inputs for the problem field (disturbances), which represent the uncertainty associated with it. The measured exogenous input of the planner represents the goal. It is a task of the planner which examines the outputs of the problem field, it compares them to the goal function and determines what actions must be undertaken to reach it. Not all planners are completely autonomous. Some have a user interface, through which the goals may be generated, allowing for certain degrees of intervention of the human element in the planning process. A solution for the problem is a sequence of inputs and outputs (possible states), generated with a view to fulfil the purpose. A model for the field of the real problem is thus developed, called the representation of the problem. Planners based on Artificial Intelligence techniques are made from the following important components (McKay and Wiers, 2003): the plan generator, the plan simulator (uses plans in the shape of heuristic decision rules), the execution model for the selected plan, the situation evaluator (optional). The Artificial Intelligence techniques used in the design of planning or expert systems are diverse. They refer especially to representation problems (on must take care in selecting the degree of detail for the mathematical structure used or the allowed modeling power, since too much modeling power may prevent the development of certain components of the planner, the verification and validation of the system), the type of approach (dependant or independent of the field), the type of the planner (hierarchical or not, linear or non-linear, reactive, distributed, encompassing a meta-planner etc.), the type of interactions that may appear in the synthesis of decisions, the forms of searching and re-planning. Due to the strong similarities between the process and the problem field, an analogy may be outlined between the models used for the process and the problem field, on the one side, and between the fundamental systemic concepts on the other. The process is described through stochastic (possibly non-linear) or differential equations, called state and output equations, which describe the dynamics of the process, the structure and its inter-connections (Stadtler and Kilger, 2005). The problem field may be described through symbolic equations. There are still strong mathematical analogies between the two types of systems, studying certain proprieties of special significance in the classic control theory: controllability, observability, stability, system rate. There is a structural analogy between classic control systems and Artificial Intelligence systems (planning, expert system), both in an open loop as well as a closed loop. Generating a plan is the process of synthesis of an aggregate of candidate plans, to fulfil the gi purpose at moment i (the goals can remain fixed or be modified). During the process of plan generation, the system designs (through a simulation based on a model of the problem field) the plausible future plan. The system uses heuristic plans based on decision rules, resource management, success probabilities, in order to choose the plan to execute. The plan execution module, translates the chosen plan into actions over the problem field, using techniques of resource allocation. Situation evaluator uses the inputs, outputs and the representation of the problem in order to determine the state of the field. The estimate state of the field is used for updated the model implicated in the design and generation process of the plan. The term of situation is preferred in this case to designate an abstract, general view of the state. The monitoring of the execution uses the estimated state of the problem field, in order to determine the viable nature of the elaborated plan, in cases of failure launching re-planning actions. In the structure of a planner based on techniques of Artificial Intelligence may also be integrated a world modeler, which does the permanent updating of the representation of the problem field. The evaluation of the situation and the monitoring of the execution in open loop systems cannot be assured, being no possibility of re-planning. In this case, the planner is sensible 30 to the variations which appear in the problem field, open-loop planners not being able to reduce this sensibility. These planners are of interest only for insignificant disturbances, which is not the case with real problems, which need complex and detailed problems. Seeing as the outputs of the system are not detected with inputs, if the problem field is unstable (input-output or internal), then it will never be able to be stabilized in open-loop planning. Open-loop planning absolutely imposes an exact amount of knowledge of the problem field. Since this cannot be obtained, any disturbance may be catastrophic. Open-loop planners present advantages due to their simplicity. If the problem field is stable and the disturbances are insignificant, then these planners may be useful. They also have the advantage of reduced costs, not being necessary measurements of states and outputs related to the problem field. Closed loop planning systems are analogous to closed loop conventional control systems and most of the time, do not use situation evaluation. The monitoring of the execution as well as the re-planning are thus permitted. The planning system examines the difference between the current output situation and the desired goal, with a view to executing certain actions. The error is not so easily obtained, compared to the classics systems of control, seeing as the distance between the fuzzy states is much harder to quantify. Ever more difficult is the evaluation of the similarities between states, based on a given criteria in the case of our fuzzy expert system. These aspects will clearly come into the design of the IKMS. Designing an IKMS as a planning system is a difficult problem, through the variety of the models used (satisfactory from a computational perspective), which must be a reflection of the complexity of the environment in which they operate (Berglund and Karlun, 2007; Chen, 2008). 3. An analysis of our IKMS based on logical events Real time expert systems are important applications both in terms of Artificial Intelligence techniques, and especially in terms of the purpose for which they are designed (Li and Xiao, 2006). They must operate in the presence of real-time restrictions. These systems are often designed and implemented without a formal analysis of how it interfaces with the process and without a careful examination of how the inference engine exploits a domain’s specific knowledge. Regularly, such systems are evaluated by: i) comparison to how experts solve the problems empirically; ii) studying their reliability and degree of user-friendliness of the interface; iii) examining the results of difficult simulations; iv) using software engineering techniques. Our IKMS consists of an expert system (ES) and the process (P), as outlined in figure 2. Efforts are concentrated on the class of expert systems based on imprecise knowledge and on the corresponding inference engine, characterizing the process management strategies. We will reduce the structure of an expert system only to the knowledge base and inference engine, which has the role of decision-making control by using the reference inputs and the outputs of the process . In this way the results of the expert system’s inferential process will be summarized, helping to obtain either the outputs , in the form of hypotheses or conclusions, provided to the user, or the potential commands euk. It is important that the design of the IKMS to allow the synthesis of an expert system based on imprecise knowledge, to perform certain management functions properly. The first step in analyzing the behaviour of a management expert system is to develop a global mathematical model for it, based on partial models, which are used for the representation and qualitative analysis of the expert system and of the process P (problem domain). Must be also analyzed the reasoning process of the rule-based expert system which emphasizes that representation of knowledge involved in the management strategy and the reasoning are more difficult than for controllers, highlighting how to develop a fuzzy expert system and in particular how to choose the conflict resolution strategy. In relation to the integration of the expert system in the structure of management expert system, its results can be used by the human decision-maker in decision-making or can be applied directly in the process. Logical discrete event systems are a class of discrete event systems with discontinuous and nonlinear equations of motion in relation to the occurrence of events. Although we use the expert systems’ terminology established in Artificial Intelligence, there are large differences in the validity of standard models in this field in relation to human cognitive structures specific to management expert systems. Therefore, the objective of the present approach is the analysis and design of a fuzzy expert system embedded in a management structure to support effectively the decision-making process. In this respect, the expert system in this work is able to use the management experience integrated into the process management model, permanently adapting to the current situation through the inference engine. It must be able to plan a finite number of steps to support the appropriate management actions and must be able to reformulate what plan (actions sequence) must be synthesized, by monitoring the current evolution of the process. This type of planning is not the most general possible, but that is not a limitation of the management expert system, since practical considerations indicate that frequently can be planned only a relatively small number of steps (especially in the presence of imprecise knowledge). For a more rigorous planning, expert system state previously defined may be completed the component xbt, representing a vector of state trajectories generated by simulating different plans. In fact, with the planning properties mentioned, the expert system performs at every step a 31 significant amount of reasoning, before synthesizing a decision. These features will be highlighted for the management expert system for the flexible manufacturing system presented in the case study. Figure. 1. Consider a process P (a structure identical to systems interconnected following an imposed scheme) where data and knowledge related to the current state and management strategies are limited because they have a high degree of imprecision. Time is a discrete measure. If the process works as a nondeterministic finite automaton, and expert system will function as a deterministic finite automaton, the closed loop management expert system can be modelled as a nondeterministic automaton, whose output is the output of the process (fig. 1). Let X be the finite set of states of the process and XSE is the finite set of states of the expert system. The model of the process may have, in this case, the following form: P = (X, XSE, δ, X0) where δ : XSE × X → P (X) - {∅} is the transition function of the process, and X0 ⊂ X is the set of initial states of P. Similarly, can be constructed the expert system model in the form SE = (XSE, X, ξ, ), where ξ : X × XSE → XSE and ⊂ XSE is an initial state of the expert system. Due to the non-deterministic character of P, there are generally infinite strings of α formulas composed of states of P, which are generated by the management expert system. These strings are called ES-admissible strings (the reason for attaching ES symbol is that we may consider different structures of expert systems for a given process P in order to achieve a goal). Interpretation of the model. Let N be the set of natural numbers. Activation of the management expert system produces a sequence (infinite) of pairs ( , xk)k∈N where is the expert system’s state, and xk is the state of the process P, at the moment k. Operation of the management expert system is defined as a sequence of pairs ( , xk). For ( ∀ ) k, = ξ (xk, ), and xk+1 = δ (ξ (xk, ), xk). The output of the process is {xk}k∈N. There is little asymmetry in the definition of management expert system’s operation. We may write the transition function of the process so that xk+1 to be the value of δ ( , xk) and not of δ( , xk). If the expert system’s input is xk and the result will be . If is the process’ input, then the management expert system’s output is xk+1. We define, in terms of the selected model, the fulfilment of a temporal formula. We want a set of temporal formulas to describe the functioning of the management expert system. We answer the following questions for a fixed system model: When the expert system works? When we measure the size of the output of the process? There must be a device in the process P to measure the current state xk based on which is calculated . 4. The Model of the Expert System Based on Fuzzy Knowledge A real time expert system must react quickly to a series of events and respond with a certain delay. Expert fuzzy systems involved in decision-making for processes management have embedded in their structure hierarchically organized knowledge that allow associations and reasoning chains based on fuzzy logic inferences. The Expert system has three types of inputs: reference input events ∈ , the outputs of the process ∈ and user inputs IU. Based on these inputs and on its current state, the expert system can synthesize allowed input command events for the process ∈ , emulating in this way the way in which human decision-maker coordinates in loop the use of evidence knowledge from the process, the reference inputs and knowledge from its own knowledge base. Fuzzy expert system models the cognitive process used in decision making, and the interaction between the inference engine and the knowledge base forms the inference cycle. Definition: A fuzzy expert system (SEF) can be modelled as follows: SEF = (XSE, ESE, , δSEF, gSEF, , ) where: XSE = Xb × Xint is the set of states xSE of the expert system, Xb is the set of imprecise states xb from the facts base, and Xint is the set of internal states xint of the inference engine. ESE = ∪ R ∪ is the set of events of the expert system, where takes place the following relationship ⊂ P ( ∪ ∪ IU) - { ∅ }. is the set of all input events of the expert system (reference inputs , the output events of the process that may arise and in relation to which the system must be able to respond in real time, and user inputs). 32 R = {R1, ..., Rn} is the set of fuzzy rules of the expert system. ⊂ P (Eu ∪ IC) - { ∅ } is the set of logical events out of the expert system. They are sets of command events allowed for the process or hypotheses / conclusions (IC) obtained from the inference process. gSE: Xb × Xint → P ( ∪ R) - { ∅ } is the activation function of fuzzy expert system. : Xb × Xint → Xb × Xint for ek ∈ P ( ∪ R) - { ∅ } is the fuzzy states transition function. δSE: Xb × Xint → are the output functions of fuzzy expert system. is the initial state of expert the system, with = ( , ) ∈ ⊂ XSE. ⊂ ESE is the set of allowed trajectories of the expert system in inference loop (physically realizable trajectories due to events). Occurrence of the input event ∈ is always accompanied by the activation of a rule, of several rules or even of more instances of the same rule for imprecise knowledge (Ri ∈ R or iR ⊂ Ri with iR the set of instances of rule Ri). Thus the inference engine will operate by updating the set of states XSE. A rule Ri ∈ R cannot be activated alone, the inferential cycle being launched only if there is a change in the process (which is reflected by the changes in its outputs), at the level of the reference inputs or user inputs. In this case, the expert system’s input events must contain (for the classical case) exactly one rule Ri ∈ R and a single command input event ∈ . Each event contains at most one event ∈ and a reference input event ∈ . These aspects can be modelled using the notion of allowed trajectory of the expert system in inference loop. Operators (xSE) corresponding to events ek ⊂ gSE(xSE) describe the update of the knowledge base and of the internal states of the inference engine when the process’ outputs, the reference, or both change, and certain rules Ri ∈ R are activated. Function δSE(xSE) describes the expert system’s outputs and the possible input commands allowed on the process P. It should be noted that events such as are not allowed as long as the expert system is running (changes its state xSE). This system can control the activation of the input command events of the process but not the set of events edk. Premises of the rules are functions Pi: Xb × → [0,1]; Pi ( , ) ∈ [0,1] indicates the degree of truth of premise Pi at moment k. Functions Pi will be used to measure the degree to which a rule (or an instance of it) may be enabled. Also, a consequent of rules is a function defined on Xb × with values in various sets (Xb, Xb × , Xb × Xint × ). The codomain of these functions is determined for each case against a series of properties of the management expert system: conflict resolution strategy, the type of knowledge. For a rule Ri, the event ek = {Ri, } ⊂ gSE( ) can occur only if the filtering degree of the premise is satisfactory at the moment k, for the state and input event . If event ek ⊂ gSE( ) occurs, then the new state of the expert system = ( ) is obtained by applying the consequent functionto the state ∈ Xb, in order to obtain the state and by updating the inference engine’s state ∈ Xint. Including input events in the rules base, allows the expert system to infer conclusions relative to the outputs and inputs of the process (reference, user commands), perceived directly as a part of the model. This aspect corresponds to the use of variables in OPS5-type conventional expert systems. Inference engine’s modelling is limited to its basic functional stages: restriction, filtering, conflict resolution, execution. For the case of imprecise knowledge, the most important stage in terms of computation is the filtering stage, from which is obtained the set of conflicts: MCk = { iR :{ iR , } ⊂ gSE( ) so that the premise of rule (instance) iR ∈R has a satisfactory filtration degree for at the moment k}. In the selection phase one rule is chosen from the set of conflicts MCk, using different conflict resolution strategies. Below we summarize the basic operations of an expert system for processes management: i) acquisition of events ; ii) Synthesis of the set of conflicts MCk in the filtering stage, using the set of rules R and the events , the facts from the facts base (blackboard) and the updated values of the variables xint; iii) Determining, based on conflict resolution strategies, the rule Ri (its instance) in MCk, in order to be activated; iv) Performing the actions of consequent function of the selected rule (not before achieving unification and the spread of all parameters for the case of fuzzy knowledge). This stage involves updating the knowledge base, the inference engine’s states and generating the inferred conclusions. Events occur in the expert system so that its operation is synchronous with the process (if an event occurs in the process will produce the activation of inference engine) and with reference input events, for which the system can generate events ∈ ⊂ P (Eu ∪ I/C)-{ ∅ }. 33 The IKMS Inference Engine Algorithm Initializations: k = 0, Πk[i] = 0, Nk[i] = 0, Θk[i] = 0, Kk[i] = 0, Λk= 0, Wk = 0, = 0, ak[i, j]. • Step 1. (Identify the current state). Pattern-matching is performed using the compiled fuzzy knowledge structure, specific to IKBSCP system, in relation to the occurrence of a certain type of event at expert system’s input. At the end of this stage we get the set of conflicts MCk. For r = 1 to nr do If (r ∈ MCk) then = 1 {Determine the rules that can be activated} If (there is exactly one rule r′ so that = 1) then execute rule r′ If (there is not a rule r′ so that = 1) then stop {I/C = ∅ or call the control module) • Step 2. (Conflict Resolution). This step uses different strategies but which are not all active for a particular fuzzy knowledge model or for a given situation. • Step 3. (Execution) e′ = {r′, } {we highlight the type of event for the IKMS} ( , ) = fCES( , ) {as a result of the occurrence of event e′, the expert system’s state is updated, i.e. the state component in the facts base and the component , whose update we present below} Λk+1(r′) = 1 {Remove the rules from the set of conflicts, using refraction} For r = 1 to nr do If (r ∈ MCk) then = + 1 {Increment the age of each rule from MCk} For r = 1 to nr do If (A[r, r′] = 1) then Λk +1 (r) = 0; = 0 {Allows the rules affected by the activation of rule r′ to be considered within the set of conflicts and resets the age of these rules to 0} End In the phase of selection based on refraction, if the expert system inference engine, which exploits the management linguistic model developed in accordance with a specific strategy used by certain human decision-maker, cannot respond to the current situation, stop may be replaced by resetting the values of to the previous values, before applying the refraction and continuing the execution, so that the expert system to use the selection based on refraction in conflict resolution only if this strategy effectively reduces the set of conflicts MCk. Note, that ( , ) = fSE( ), as a result of the occurrence of an event of type e′= {r′, }, is the application of the state modifying operator, equivalent to the application of the consequent formula of rule r′, selected for execution in the state . In this way are updated also the components of the internal state vector xint of the inference engine. There was no evidence, within the inference engine algorithm structure, of the mode of action of output function , as it has a character specific to the management model. This will be validated practically for the flexible production system, chosen for a case study to highlight the characteristics of IKMS system. The closed-loop control expert system can be modelled like a nondeterministic state machine, whose outputs are the process outputs. 5. Conclusions The use of temporal aspects refers to the design of those tools to integrate time in control economic applications (e-commerce, Knowledge Management Systems, Virtual Organizations, Multi-agent Systems). These aspects are formally found on the inference engine algorithms, able to make full use of the specific knowledge to the process control. We assume that the process operates like finite nondeterministic state machine, while the expert system will operate like a finite deterministic state machine. The problem domain must be defined as a collection of problems that the expert system desires to solve. In conventional control, the plant is a dynamical system, described with linear or non-linear differential/ difference equations. An Intelligent Knowledge Management System based on knowledge control and planning models consists of the planner or the inference engine, the problem domain, the exogenous inputs, and their interconnections. An expert system can be modelled using predicate or temporal logic or other symbolic techniques such as finite state machine. Although the KMSs are frequently being used to perform complex control functions, most often it is the case that no formal analysis of the dynamics is conducted because mathematical analysis of such systems is often considered to be beyond the scope of conventional control theory. Mathematical Logics have been developed in three great directions: i) The Theory of Models. The model is considered the main concept of first-degree languages semantics. The main objective of the theory of models is to describe the models from certain axiomatic theories, in order to highlight their mathematical structures (algebraic, topological); ii) The Demonstration Theory. The notion of demonstration is important here, and one can study complexity-related problems of various demonstrations on a single theorem, according to different deductive systems or 34 languages (regularity, equivalences, similarities); iii) The Computability Theory. This is the case in which one studies the intrinsic notion of computable function, with its various models: the Turing machine, recursive functions and combinatory logic with a set of properties. Starting from these notions we can study the decidability concept, together with a study of various models used to demonstrate the decidability or non-decidability of mathematical theories. The theory of abstract complexity that generally uses different variants of the Turing machine (as an elementary calculus model to evaluate the necessary calculus resources to solve certain given problems) allows us to establish the complexity classes for decidable problems. We must also highlight the unity of these three great aspects. Calculus and decidability notions are strongly inter-related, and we can thus establish precise correspondences between them. The theory of models can contribute as regards decidability problems and can also provide the semantic justification of the rules of inference and of automated demonstration systems. This unit is an essential characteristic of Logics which represents a fundamental theoretic aspect. From the symbolic Artificial Intelligence perspective, we can find tendencies specific to the three characteristics that were presented above: semantics and the theory of models, demonstration and reasoning, computability and complexity, but often with different methods compared to theories applied only to mathematical logics. As regards the contribution of Logics in terms of Artificial Intelligence systems, we can mention: i) on the theoretical level, logic contributes with a series of creation elements and methods such as the syntax/semantics/decision trilogy, coherence (verifying the coherence of the knowledge-base, maintaining the truth systems), decidability (for example, absence logic is not semi-decidable, contrary to first-degree logic, certain temporal logic are decidable, others are not), complexity of the decision methods. Logic is thus a reference model for the foundations of symbolic Artificial Intelligence systems. ii) The formalisation of different types of reasoning. Logic has a normative role even in the absence of some certain knowledge. Automated demonstration, representing knowledge languages, logical programming languages they all represent means that support and integrate reasoning methods. References 1. Barachini, F. (1990). Match-Time Predictability in Real-Time Production Systems, in Springer-Verlag, Lecture Notes in Artificial Intelligence, pp. 190-203. Retrived May 8, 2012, from http://link.springer.com/chapter/10.1007%2F3-540-53104-1_42. 2. Bergamaschi, S., Gelati, G., Guerra, F., Vincini, M. (2006). An intelligent data integration approach for collaborative project management in virtual enterprises, World Wide Web 9,35–61. doi 10.1007/s11280-005-2322-7. 3. Berglund, M. and Karlun, J. (2007). Human, technological and organizational aspects influencing the production scheduling process, International Journal of Production Economics, Vol. 110, No. 1/2, 160-174. 4. Bhatt, G. (2000). Organizing knowledge in the knowledge development Cycle, Journal of Knowledge Management, Vol. 4 No. 1, 15–26. 5. Bhatt, G. (2001). Knowledge management in organizations: examining the interaction between technologies, techniques, and people, Journal of Knowledge Management. Kempston: Vol. 5 No. 1, 68-75. 6. Burden, D.J. (2009). Deploying embodied AI into virtual worlds, Knowledge-Based Systems 22, pp. 540–544. 7. Bylander, T. (1994). The computational complexity of propositional STRIPS planning, Artificial Intelligence, no 69, 165-204. 8. Chen, T.Y. (2008). Knowledge sharing in virtual enterprises via an ontology-based access control approach, Computers in Industry 59, pp. 502–519. 9. David, F., Pierreval, H., Caux, C. (2006). Advanced planning and scheduling systems in the aluminium conversion industry, International Journal of Computer Integrated Manufacturing, Vol. 19, No. 7, 705-715. 10. Jonsson, P. and Mattsson, S. (2006). A longitudinal study of material planning applications in manufacturing companies, International Journal of Operations and Production Management, Vol. 26, No. 9, pp. 971-995. 11. Jonsson, P., Kjellsdotter, L., Rudberg, M. (2007). Applying advanced planning systems for supply chain planning: three case studies, International Journal of Physical Distribution & Logistics Management, Vol. 37, No. 19, 816-834. 12. Kim, Y.G., Yu, S., Lee, H. (2003). Knowledge strategy planning: methodology and case, Expert Systems with Applications, Volume 24, Issue 3, pp. 295-307. 13. Li, H. and Xiao, R. (2006). A multi-agent virtual enterprise model and its simulation with Swarm, International Journal of Production Research 44, 1719–1737. 14. Lin, C.H., Hwang, S.L., Wang, M.Y. (2007). A reappraisal on advanced planning and scheduling systems, Industrial Management & Data Systems, Vol. 107, No. 8, 1212-1226. 15. Lunardhi, A.D. and Passino, K.M. (1995). Verification of qualitative properties of rule-based expert systems, in Applied Artificial Intelligence, pp. 587-621. 16. Mazilescu, V. (2012). A Knowledge Management System Embedded in the New Semantic Technologies, Chapter 1, in New Research on Knowledge Management Technologies, Intech, Rijeka, pp.1-22. 17. McKay, K.N. and Wiers, V.C. (2003). Planning, scheduling and dispatching tasks un production control, Cognition, Technology and Work, Vol.5, No. 2, 82-93. 18. Stadtler, H. and Kilger, C. (2005). Supply Chain Management and Advanced Planning-Concepts, Models, Software and Case Studies, 3rd ed., Springer, Berlin. 19. Stoop, P.M. and Wiers, C.S. (1996). The complexity of scheduling in practice, International Journal of Operations & Production Management, Vol. 16. No. 10, 37-53. 
work_osxoyxxqwncyjc53pteadeg4ne	----	A Delphi-based approach to developing expert systems with the cooperation of multiple experts A Delphi-based approach to developing expert systems with the cooperation of multiple experts Hui-Chun Chu, Gwo-Jen Hwang * Department of Information and Learning Technology, National University of Tainan, 33, Sec. 2, Shulin Street, Tainan City 70005, Taiwan, ROC Abstract Knowledge acquisition has been a critical bottleneck in building knowledge-based systems. In past decades, several methods and sys- tems have been proposed to cope with this problem. Most of these methods and systems were proposed to deal with the acquisition of domain knowledge from single expert. However, as multiple experts may have different experiences and knowledge on the same appli- cation domain, it is necessary to elicit and integrate knowledge from multiple experts in building an effective expert system. Moreover, the recent literature has depicted that ‘‘time’’ is an important parameter that might significantly affect the accuracy of inference results of an expert system; therefore, while discussing the elicitation of domain expertise from multiple experts, it becomes an challenging and impor- tant issue to take the ‘‘time’’ factor into consideration. To cope with these problems, in this study, we propose a Delphi-based approach to eliciting knowledge from multiple experts. An application on the diagnosis of Severe Acute Respiratory Syndrome has depicted the superiority of the novel approach. � 2007 Elsevier Ltd. All rights reserved. Keywords: Knowledge acquisition; Expert system; Knowledge integration; Medical diagnosis; Delphi method 1. Introduction In the past decades, expert systems have been applied to various applications. Subject domains that are supported by experts systems include bioengineering, defense, educa- tion, engineering, finance, and medical diagnosis. For example, MYCIN project is a well-known medical expert system for diagnosing infectious diseases (Buchanan & Shortliffe, 1985); ISODEPOR was developed to evaluate the muscle strength of Spanish top-competition athletes (Barreiro et al., 1997); FRBS-GP is a fuzzy rule-based sys- tem for diagnosing aphasia’s subtypes and the classification of pap-smear examinations (Jantzen, Axer, & Keyserlingk, 2002). The successful cases of the expert system approach not only demonstrated the benefits of applying expert system approach to coping with medical diagnosis problems, but also depicted the difficulty of applying it. In building an expert system, the critical bottleneck is to obtain the knowledge of the special domain from the domain experts, which is called knowledge acquisition. In past decades, several methods and systems have been proposed to cope with this problem. However, most of these methods and systems were proposed to deal with the acquisition of domain knowledge from single expert. However, as multi- ple experts may have different experiences and knowledge on the same application domain, it is necessary to elicit and integrate knowledge from multiple experts in building an effective expert system. Recent literature also indicated that ‘‘time’’ is an important parameter that might signifi- cantly affect the accuracy of inference results of an expert system; therefore, while discussing the elicitation of domain expertise from multiple experts, it becomes a much more challenging and important issue to take the ‘‘time’’ factor into consideration (Hwang, Chen, Hwang, & Chu, 2006). To cope with these problems, we shall propose a Delphi- based approach to eliciting knowledge from multiple experts. An application of developing a medical expert sys- tem has depicted the superiority of the novel approach. 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.05.034 * Corresponding author. Tel.: +886 915396558; fax: +886 6 2606132. E-mail address: gjhwang@mail.nutn.edu.tw (G.-J. Hwang). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 34 (2008) 2826–2840 Expert Systems with Applications mailto:gjhwang@mail.nutn.edu.tw 2. Relevant researches To cope with the knowledge acquisition problem, many knowledge acquisition tools or methods have been pro- posed to build rapid prototypes and to improve the quality of the elicited knowledge, e.g., ETS (Boose, 1984, 1985), TEIRESIAS (Davis, 1979), MORE (Kahn, Nowlan, & McDermott, 1985), SALT (Marcus, 1987; Marcus, McDer- mott, & Wang, 1985), NeoETS (Boose & Bradshaw, 1986; Kitto & Boose, 1986), KNACK (Klinker, Bentolila, Gen- etet, Grimes, & McDermott, 1987), AQUINAS (Boose & Bradshaw, 1987; Shema & Boose, 1988), KRITON (Diede- rich, Ruhmann, & May, 1987), Student (Gale, 1987), Rule- Cons (O’Bannon, 1987), MOLE (Eshelman, Ehret, McDermott, & Tan, 1987), KITTEN (Shaw & Gaines, 1987), KSSO (Gaines, 1987), ASK (Gruber, 1988), Word- Net (Millar, 1990; Navigli, Velardi, & Gangemi, 2003), KADS (Schreiber, Wielinga, & Breuker, 1993; Wielinga, Schreiber, & Breuker, 1992), MCRDR (Kang, 1996), Med- Frame/CADIAG-IV (Boegl, 1997; Kolousek, 1997; Leitich et al., 2001). Most of these systems were developed based on the repertory grids method originated from Kelly’s Per- sonal construct theory (Kelly, 1955), which assists in iden- tifying different objects in a domain and distinguishing among these objects. A single repertory grid is represented as a matrix whose columns have elements labels and whose rows have con- struct labels. A 5-scale rating mechanism is usually used in filling the grid; i.e., each rating is an integer ranging from 1 to 5, where ’’1’’ represents that the element is very likely to have the trait; ‘‘2’’ represents the element may have the trait; ‘‘3’’ represents ‘‘unknown’’ or ‘‘no relevance’’; ‘‘4’’ represents that the element may have the opposite charac- teristic of the trait; ‘‘5’’ represents that the element is very likely to have the opposite characteristic of the trait. As repertory grid approach has been widely used by researchers, some extensions have been made to enhance its representative ability. For example, Jose, Nicholas, Jennings, Luo, and Shadbolt, 2003 developed a technique using a fuzzy repertory grid for acquiring the finite set of attributes or variables that the expert uses in a classifica- tion problem, characterizing and discriminating a set of elements. In addition, several models have been proposed to generate more meaningful rules from the repertory grid- oriented approaches, such as the EMCUD method, which can generate embedded meanings from repertory grids by defining the impacts of the constructs to each element (Hwang & Tseng, 1990). Recently, Hwang et al. (2006) indicated that, in building medical expert systems, most of the previous knowledge acquisition methods only pay attentions to the relationships between diseases and symp- toms, while, the variant of the symptoms in different time scales of the diseases are not taken into account. Consider the repertory grid given in Table 1 which depicts an exam- ple of eliciting knowledge for diagnosing various kinds of gastrointestinal diseases. Note that the rating of the (Acute bronchitis, Throat pain) entry is 4, which implies highly tendency for Acute bronchitis to have Throat pain. How- ever, in practical situation, Influenza has significant appearance of Throat pain in the early time scale. What has been addressed in the repertory grid is not happened in the last time scale of acute bronchitis. For later time scale, the throat pain symptom will become not so signif- icant. Such variant of disease symptoms with respect to different time scales cannot be precisely presented by those conventional knowledge acquisition approaches. 3. Delphi-based knowledge acquisition approach In developing a knowledge-based system, it is very dif- ficult to elicit and integrate knowledge from multiple experts (Hwang et al., 2006), especially the application domains in which various time scales of elements need to be taken into account. To cope with this problem, a novel approach, Knowledge Acquisition for Multiple Experts with Time scales (KAMET), is proposed in this section, which takes time scales into consideration while eliciting expertise from multiple experts. In addition to time scales, KAMET takes importance degree for each construct to each element in different time scales into account, such that more embedded knowledge can be explicitly presented. Let eti denote tth stage period of element (or disease) ei and cj denote a construct (or symptom), where i = 1 to n, and j = 1 to m. Each KAMET entry is a triplet that con- sists of three values: a rating to indicate the relevance of eti and cj, a certainty degree for giving the rating and an impact factor to represent the importance of cj to e t i , which are represented by the following three functions: (1) Rating (eti , cj): the degree of relevance for element ei in tth time scale to construct cj, ranging from 1 to 5: ‘‘1’’ represents that the element is very likely to have the opposite characteristic of the trait; ‘‘2’’ represents the element may have the opposite characteristic of the trait; ‘‘3’’ represents ‘‘unknown’’ or ‘‘no rele- vance’’; ‘‘4’’ represents that the element may have the trait; ‘‘5’’ represents that the element is very likely to have the trait. (2) Certainty (eti , cj): the degree of certainty for giving Rating (eti , cj), which is either ‘‘S’’ or ‘‘N’’ represent- ing ‘‘sure’’ or ‘‘not sure’’. (3) Impact_factor (eti , cj): the degree of importance for construct cj to element ei in tth time scale. Impact_ factor (eti , cj) can be one of the following values: ‘‘X’’ represents no relationship between the element and the construct; ‘‘D’’ means that the construct dominates the element, i.e., if the value of the con- struct is not matched, it is impossible for the element to be implied; an integer, ranging from 1 to 5, indi- cates that the construct is of some degree of impor- tance to the element, but does not dominate the implication of the element. H.-C. Chu, G.-J. Hwang / Expert Systems with Applications 34 (2008) 2826–2840 2827 https://isiarticles.com/article/978 
work_othruimflnaupb3s56fqcmitji	----	A reduced data set method for support vector regression A reduced data set method for support vector regression Horng-Lin Shieh *, Cheng-Chien Kuo Department of Electrical Engineering, Saint John’s University, Taipei, Taiwan a r t i c l e i n f o Keywords: Support vector regression Outlier Fuzzy clustering Robust fuzzy c-means a b s t r a c t Support vector regression (SVR) has been very successful in pattern recognition, text categorization, and function approximation. The theory of SVR is based on the idea of structural risk minimization. In real application systems, data domain often suffers from noise and outliers. When there is noise and/or out- liers exist in sampling data, the SVR may try to fit those improper data, and obtained systems may have the phenomenon of overfitting. In addition, the memory space for storing the kernel matrix of SVR will be increment with O(N2), where N is the number of training data. Hence, for a large training data set, the kernel matrix cannot be saved in the memory. In this paper, a reduced support vector regression is pro- posed for nonlinear function approximation problems with noise and outliers. The core idea of this approach is to adopt fuzzy clustering and a robust fuzzy c-means (RFCM) algorithm to reduce the com- putational time of SVR and greatly mitigates the influence of data noise and outliers. Crown Copyright � 2010 Published by Elsevier Ltd. All rights reserved. 1. Introduction The theory of support vector machines (SVM) developed by Vapnik (1995) in 1995 is gaining in popularity due to its many attractive features. The SVM is based on the idea of structural risk minimization (SRM) and has been shown to be superior to tradi- tional empirical risk minimization (ERM) principles employed by conventional neural networks (Gunn, 1998). SVM has been suc- cessfully applied to a number of applications, such as classification, time predictions, pattern recognition, and regression (Burges, 1998; Jair, Xiaoou, Wen, & Kang, 2008; Kamruzzaman & Begg, 2006; Kumar, Kulkarni, Jayaraman, & Kulkarni, 2004; Lijuan, 2003; Wong & Hsu, 2006; Zhou, Zhang, & Jiao, 2002). In many intelligent systems, SVM has been shown to provide higher perfor- mance than traditional learning machines, and has thus been adopted as a tool for solving classification issues (Lin & Wang, 2002). Over the past few years, a lot of researchers of neural net- works and machine learning fields are attracted to devoting them- selves to research on SVM (Wang & Xu, 2004). The SVM is systematic and properly motivated by the statistical learning theory (Vapnik, 1998). Training of the SVM involves opti- mization of a convex cost function and globally minimizes to com- plete the learning process (Campbell, 2002). In addition, SVM can handle large input, and can automatically identify a small subset consisting of informative points, namely support vectors (Gustavo et al., 2004). The SVM can also be applied to regression problems by the introduction of an alternative loss function (Gunn, 1998). Such approaches are often called support vector regression (SVR). SVM maps the input data into a high-dimensional feature space, and searches a separate hyperplane that maximizes the margin be- tween two classes. SVM adopts quadratic programming (QP) to maximize the margin as computing tasks become very challenging when the number of data is beyond a few thousand (Hu & Song, 2004). For example, in Fig. 1, there are 500 sampling data gener- ated from a sin wave with Gaussian noise N(0, 0.447). The SVR algo- rithm is adopted to construct this function. The entries of the kernel matrix of SVR are floating-points numbers, and each float- ing-point number requires 4 bytes for storing. Therefore, the total memory required is 500 � 500 � 4 = 1000,000 bytes. The SVR algo- rithm is performed on a Pentium 4, 1.8 GHz with 128 MB of mem- ory running Windows XP. The total execution time of the simulation is 21941 s (above 6 h). This execution time is very long and the memory requirements are very large for real applications of science. Osuna, Freund, and Girosi (1997) proposed a generalized decomposition strategy for the standard SVM, in which the original QP problem is replaced by a series of smaller sub-problems, which are proved able to converge to a global optimum point. However, it is well known that the decomposition process relies heavily on the selection of a good working set of the data, which normally starts with a random subset (Hu & Song, 2004). Lee and Huang (2007) proposed to restrict the number of support vectors by solving re- duced support vector machines (RSVM). The main characteristic of this method is to reduce the matrix from l � l to l � m, where m is the size of a randomly selected subset of training data that are considered as candidates of support vectors. The smaller matrix 0957-4174/$ - see front matter Crown Copyright � 2010 Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.04.062 * Corresponding author. E-mail address: shieh@mail.sju.edu.tw (H.-L. Shieh). Expert Systems with Applications 37 (2010) 7781–7787 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2010.04.062 mailto:shieh@mail.sju.edu.tw http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa can easily be stored in memory, and then optimization algorithms, such as the Newton method, can be applied (Lin & Lin, 2003). How- ever, as shown by Lin and Lin (2003), numerical experiments show that the accuracy of RSVM is usually lower than that of SVM. Because support vectors can state the distributional features of all data, according to the characteristics of SVM, removing trivial data from the whole training set will not greatly affect the out- come, but will effectively increase the training process (Wang & Xu, 2004). A reduced set method based on the measurement of similarities between samples is developed by Wang and Xu (2004). In this paper, the samples similar to some data points will be discarded under a pre-established similarity threshold. In other words, these samples are so similar to the special data point that their influence on the prediction function can be ignored. Accord- ing to this method, a large number of training vectors are dis- carded, and then a faster SVM training can be obtained without compromising the generalization capability of SVM. However, like the K-means clustering algorithm, the disadvantage of this algo- rithm is that the number of clusters must be predetermined, but in some real applications, there is no information to predefine the number of the clusters. In real applications, data is bound to have noise and outliers, and algorithms utilized in engineering and scientific applications must be robust in order to process these data. In system modeling with noise and/or outliers existing in the sampling data, the system models may try to fit those improper data, and the output may have the phenomenon of overfitting (Chung, Su, & Hsiao, 2000; Shieh, Yang, Chang, & Jeng, 2009). SVR has been shown to have excellent performance for both the e-insensitive and Huber’s ro- bust function for matching the correct type of noise in an applica- tion of time series prediction (Mukherjee, Osuna, & Girosi, 1997). However, in this SVR approach, outliers may possibly be taken as support vectors, and such an inclusion of outliers in support vec- tors may lead to serious overfitting phenomena (Chung, 2000). In this paper, in order to overcome the above problems, a robust fuzzy clustering method is proposed to greatly mitigate the influ- ence of noise and outliers in sampling data, and then the SVR method is used to construct the system models. Three experiments are illustrated, and their results have shown the proposed ap- proach has better performance and less execution time than the original SVR method in various kinds of data domains with data noise and outliers. 2. Support vector regression The model of learning from examples can be considered a gen- eral statistic framework of minimizing expected loss using sam- pling data. Suppose there are n random independent identically distributed (i.i.d.) data (x1, y1), (x2, y2), . . ., (xn, yn), where xi 2 Rd, yi 2 R, i = 1, 2, . . ., n drawn according to the uniform probability dis- tribution function P(x, y) = P(x)P(y—x). Given a set of functions f(x, a), a 2 K, where K is a parameter set, from which the goal of the learning process is to choose a function f(x, a0) that can obtain the best relationship between input and output pairs. Consider a measure of the loss L(y, f(x, a0)) between the output y of the sam- pling data to a given input x, and the response f(x, a0), provided by the learning machine. In order to obtain f(x, a0), one has to min- imize the expected risk functional RðaÞ¼ Z Lðy; fðx; a0ÞÞdPðx; yÞ; ð1Þ A common choice for the loss function is L2-norm; i.e., L(e) = e2 = (y � f(x, a0))2. However, because P(x, y) is unknown, R(a) cannot be directly evaluated from Eq. (1). In general, the ex- pected risk function is replaced by the empirical risk functional RempðaÞ¼ 1 n Xn i¼1 Lðy; fðx; a0ÞÞ: ð2Þ There is no probability distribution in Eq. (2). However, in real application systems, data domains often suffer from noise and out- liers. When there is noise and/or outliers exist in sampling data, Eq. (2) may try to fit those improper data and obtained systems may have the phenomenon of overfitting. Let the sampling data be represented as {(xi, yi)jxi 2 Rd, yi 2 {�1, 1}}, i = 1, 2, . . . n. In the SVR method, the regression func- tion is approximated by the following function as: f ¼ Xn i¼1 wiuðxiÞþ b; ð3Þ where fuðxiÞg n i¼1 are the features of inputs, fwig n i¼1 and b are coeffi- cients. The coefficients are estimated by minimizing the regularized risk function (Wang & Xu, 2004) RðCÞ¼ C 1 n Xn i¼1 Lðy; fÞþ 1 2 kwk2; ð4Þ where L(y, f) adopt the e-insensitive loss function, and is defined as follows: Lðy; fÞ¼ jy � f j� e; jy � f j P e; 0; otherwise � ð5Þ and e P 0 is a predefined parameter. In Eq. (4), the second term, 12kwk 2 , is used for the flatness mea- surement of function (3), and C is a regular constant determining the tradeoff between the training error and the model flatness. SVR introduces slack variables n, n* and leads Eq. (4) to the follow- ing constrained function (Wang & Xu, 2004): minimize Rðw; n�Þ¼ C� Xn i¼1 ðni þ n � i Þþ 1 2 kwk2; ð6Þ subject to wuðxiÞþ b � yi 6 e þ n � i ; ð7Þ yi � wuðxiÞ� bi 6 e þ ni; n; n� P 0; where n, n* are slack variables representing upper and lower con- straints on the outputs of the system. Thus, function (3) becomes the explicit form: fðx; ai; a�i Þ¼ Xn i¼1 wiuðxiÞþ b ¼ Xn i¼1 ðai � a�i ÞuðxiÞ T uðxiÞþ b ð8Þ 0 1 2 3 4 5 6 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 X Y Fig. 1. Sin wave with noise N(0, 0.447). 7782 H.-L. Shieh, C.-C. Kuo / Expert Systems with Applications 37 (2010) 7781–7787 https://isiarticles.com/article/25301 
work_oupt5hnjxjdyteqxbkcurzltsm	----	doi:10.1016/j.eswa.2005.09.082 Knowledge acquisition through information granulation for imbalanced data Chao-Ton Su a,*, Long-Sheng Chen b , Yuehwern Yih c a Department of Industrial Engineering and Engineering Management, National Tsing Hua University, 101, Kuang Fu Road, Sec. 2, Hsinchu 300, Taiwan, ROC b Department of Industrial Engineering and Management, National Chiao Tung University, Hsinchu 300, Taiwan, ROC c School of Industrial Engineering, Purdue University, IN 47907-2023, USA Abstract When learning from imbalanced/skewed data, which almost all the instances are labeled as one class while far few instances are labeled as the other class, traditional machine learning algorithms tend to produce high accuracy over the majority class but poor predictive accuracy over the minority class. This paper proposes a novel method called ‘knowledge acquisition via information granulation’ (KAIG) model which not only can remove some unnecessary details and provide a better insight into the essence of data but also effectively solve ‘class imbalance’ problems. In this model, the homogeneity index (H-index) and the undistinguishable ratio (U-ratio) are successfully introduced to determine a suitable level of granularity. We also developed the concept of sub-attributes to describe granules and tackle the overlapping among granules. Seven data sets from UCI data bank, including one imbalanced diagnosis data (pima-Indians-diabetes), are provided to evaluate the effectiveness of KAIG model. By using different performance indexes, overall accuracy, G-mean and Receiver Operation Characteristic (ROC) curve, the experimental results comparing with C4.5 and Support Vector Machine (SVM) demonstrate the superiority of our method. q 2005 Elsevier Ltd. All rights reserved. Keywords: Information granulation; Fuzzy ART; Granular computing; Knowledge acquisition; Imbalanced data 1. Introduction Learning from imbalanced/skewed data is an important topic and rises very often in practice. In such kind of data, one class might be represented by a large number of examples while the other is represented by only a few. Many real world data have these characteristics, such as fraud detection, text classification (Chawla, Bowyer, Hall, & Kegelmeyer, 2002; Chawla, Japkowicz, & Kolcz, 2004) telecommunications management, oil spill detection, risk management, medical diagnosis/monitoring, financial analysis of loan policy or bankruptcy (Batista, Prati, & Monard, 2004; Chawla et al., 2004; Grzymala-Busse, Stefanowski, & Wilk, 2004) and protein data (Provost & Fawcett, 2001). Traditional classifiers seeking an accurate performance over a full range of instances are not suitable to deal with imbalanced learning tasks (Batista et al., 2004; Chawla et al., 2004; Guo & Viktor, 2004) since 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.09.082 * Corresponding author. Tel.: C886 357 429 36; fax: C886 357 222 04. E-mail addresses: ctsu@mx.nthu.edu.tw (C.-T. Su), alexchn.iem91g@ nctu.edu.tw (L.-S. Chen), yih@purdue.edu (Y. Yih). they tend to classify all data into the majority class, which is usually the less important class. Therefore, these traditional algorithms often produce high accuracy over the majority class, but poor predictive accuracy over the minority class. To cope with imbalanced data sets, there are some methods proposed in literatures, such as the methods of sampling (Batista et al., 2004; Chawla et al., 2002; Guo & Viktor, 2004), adjusting the cost-matrices (Cristianini & Shawe-Taylor, 2000), and moving the decision thresholds (Chawla et al., 2002; Huang, Yang, King, & Lyu, 2004; Jo & Japkowicz, 2004). Sampling methods reduce data imbalance—by ‘down- sampling’ (removing) instances from majority class or ‘up- sampling’ (duplicating) the training instances from the minority class or both. The second kind of methods improves the prediction accuracy by adjusting the cost (weight) for each class or changing the strength of rules (Batista et al., 2004). The third school of methods tries to adapt the decision thresholds to impose bias on the minority class. However, these three schools of methods lack a rigorous and systematic treatment on imbalanced data (Huang et al., 2004). For example, down- sampling the data will lose information, while up-sampling will introduce noise. In this study, we introduce the concept of ‘information granulation’ to solve class imbalance problems. There are two Expert Systems with Applications 31 (2006) 531–541 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa Fig. 1. An information-processing pyramid (Bargiela & Pedrycz, 2003). C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541532 reasons why we propose this concept to tackle this issue. The first one is human instinct. As human beings, we have developed a granular view of the world. When describing a problem, we tend to shy away from numbers and use aggregates to ponder the question instead. This is especially true when a problem involves incomplete, uncertain, or vague information. It may be sometimes difficult to differentiate distinct elements, and so one is forced to consider ‘information granules’ (IG) which are collections of entities arranged together due to their similarity, functional adjacency and indistinguishability (Bargiela & Pedrycz, 2003; Castellano & Fanelli, 2001; Yao & Yao, 2002; Zadeh, 1979). A typical example is the theory of rough sets (Walczak & Massart, 1999). The process of constructing IGs is referred to as information granulation. This was first pointed out in the pioneering work of Zadeh (1979) who coined the term ‘information granulation’, and emphasized the fact that a plethora of details does not necessarily amount to knowledge. Granulation serves as an abstraction mechanism for reducing an entire conceptual burden. The essential factor driving the granulation of information is the need to comprehend the problem and have a better insight into its essence, rather than get buried in all the unnecessary details. By changing the size of the IGs, we can hide or reveal more or less details (Bargiela & Pedrycz, 2003). The second reason is about the behavior of data. In many practical datasets, such as medical/diagnosis, inspection, fault monitoring and fraud detecting data, the normal group and abnormal group are considered separate populations. Taguchi & Juoulum (2002) thought every abnormal condition (or a condition outside ‘healthy’ group) is considered unique, since the occurrence of such a condition is different. Tolstoy’s quote in Anna Karenina: ‘All happy families look alike. Every unhappy family is unhappy after its own fashion’ is also noted to illustrate their opinions (Taguchi & Juoulum, 2002). Therefore, we can clearly understand the normal group (i.e. healthy patients, good products) looks alike while the abnormal group (i.e. sick patients, defective products) is unique. If we construct IGs by the similarity of numerical data, the amount of IGs in normal group will be remarkably smaller than the size of normal numerical data. In other words, if we consider IGs instead of numerical data, it might increase the proportion of abnormal data and improve imbalanced/skewed situation of data. In this study, we propose a ‘knowledge acquisition via information granulation’ (KAIG) model which can improve classification performance by controlling the reduction of unnecessary details. In KAIG model, Fuzzy ART (Adaptive resonance theory) neural network is utilized to construct IGs. The two indexes, the homogeneity index (H-index) and the undistinguishable ratio (U-ratio), are developed to determine a suitable level of granularity. The concept of sub-attributes is presented to tackle the overlapping among granules. Six data sets (one for illustrative example) from data bank are employed to illustrate our method and evaluate the effectiveness of our proposed model. Besides, one imbalanced diagnosis dataset, pima-Indians-diabetes, is provided to demonstrate the superiority of our method in solving class imbalance class problem by using the indexes, overall accuracy, G-mean and receiver operation characteristic (ROC) curve. 2. Granular computing Granular computing, which is oriented towards the representation and processing of IGs, is quickly becoming an emerging conceptual and computing paradigm of information processing (Bargiela & Pedrycz, 2003). It is a superset of the theory of fuzzy information granulation, rough set theory and interval computations, and is a subset of granular mathematics. Granular computing as opposed to numeric computing is knowledge-oriented. Numeric computing is data oriented. The main issues (Castellano & Fanelli, 2001) of granular computing are how to construct the IGs, and to describe IGs. One particular question that arises is how to determine the level of granularity. We discuss these issues in the next sections. 2.1. Construction of information granules In the issue of constructing IGs, there are many approaches, such as the Self Organizing Map (SOM) network (Castellano & Fanelli, 2001), Fuzzy C-means (FCM), rough sets, shadowed sets (Bargiela & Pedrycz, 2003) used to do this. Because IGs exist at different levels of granularity, we usually group granules of similar ‘size’ (that is granularity) in a single layer. If more detailed processing is required, smaller IGs are selected. Fig. 1 illustrates this concept of granularity. At the lowest level, we are concerned with numeric processing. This is a domain completely taken over by numeric models, such as differential equations, regression models, neural networks, etc. At the intermediate level, we see larger IGs (viz. those embracing more individual elements). The top level is solely devoted to symbol-based processing, and as such invokes well- known concepts of Petri nets, qualitative simulation, etc. (Bargiela & Pedrycz, 2003). In this study, the Fuzzy ART is utilized to construct IGs. ART is a well established neural network theory developed by Carpenter, Grossberg, and Rosen (1991). The ART network is also a famous method of clustering. Instead of clustering by a given number of clusters, it assigns patterns onto the same cluster by comparing their similarity. The detailed algorithm of Fuzzy ART can be found in (Serrano-Gotarredona, Linares-Barranco, & Andreou, 1998). The major difference between ART and other unsupervised neural networks is the so called vigilance parameter (r) which is viewed as a granularity and can be adjusted by the users to Table 1 The information granule-iris example Condition attributes Decision attribute (classes) A B C D 5.8 2.7 4.1 1 Versicolor 6.2 2.2 4.5 1.5 Versicolor 5.6 2.5 3.9 1.1 Versicolor 5.9 3.2 4.8 1.8 Versicolor 5 3.3 1.4 0.2 Setosa C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541 533 control the degree of similarity of patterns placed on the same cluster. In an ART, the degree of similarity between a new pattern and a stored pattern is defined. This similarity, compared to r, is a measure to ensure whether the new pattern is properly classified or not. The other unsupervised learning neural networks which do not implement vigilance may cause a significantly different input pattern to be forced into an inappropriate cluster. In contrast to some other cluster methods, an ART network will not automatically force all input vectors onto a cluster if they are not sufficiently similar. This is the reason why the ART network is employed in this study to construct the IGs. There are three similar ART architectures, namely ART 1, ART 2, and Fuzzy ART. ART 1 is designed for binary-valued input patterns, and ART 2 is for continuous-valued patterns. Fuzzy ART is the most recent adaptive resonance framework that provides a unified architecture for both binary and continuous valued inputs. There are several factors that motivated us to use Fuzzy ART, and they are as follows (Burke & Kamal, 1995): (1) Unlike ART1, Fuzzy ART does not require a completely binary representation of the parts to be grouped. In addition, Fuzzy ART possesses the same desirable stability properties as ART1 and a simpler architecture than that of ART2. (2) ART2 can experience difficulty in achieving good categorizations if the input patterns are not all normalized to a constant length. However, such normalization can possibly destroy valuable information. Besides, there is a serious dependency of classification results in the case of ART1 on the sequence of input presentation. As a result, the Fuzzy ART network is employed to construct IGs in this study. Table 2 The undistinguishable information granule Condition attributes Decision attribute A B C D 5.4 2.2 3.9 1.2 Versicolor 6.8 3.4 5.6 2.4 Virginica 2.2. Selection of granularity Selecting an appropriate size of IGs is a difficult task. Enough background knowledge is required to determine how similar objects should be gathered together to form one IG. An objective index is needed to select the appropriate similarity of granules. We proposed H-index and U-ratio to solve this problem. The basic assumption of the H-index is that the classes of objects should be equal if their values of attributes are sufficiently similar. This implies that we always make the same decision under a similar condition. Because we form granules by the similarity of objects, the objects in the same granule should have the same class. The H-index is used to measure the consistency of the class of the objects in one IG. The H-index is defined as H � index Z i n (1) where n represents the number of all objects in one granule and i is the amount of objects possessing the majority class. For example, Table 1 shows one IG involving five objects (nZ5). There are four condition attributes (namely A–D) in the iris data. The decision attribute (class) of the first four objects is ‘versicolor’, but the last one has a different decision attribute, ‘setosa’. In this example, ‘versicolor’ is the majority class and iZ4. The H-index of this IG is 4/5. Another index for selecting similarity is the U-ratio. In the preceding example, ‘versicolor’ is the majority of the classes. So it is assigned to be the class of this IG. If there was another granule described as Table 2, and we are unable to distinguish the class of the IG, then we call that granule an ‘undistinguish- able granule.’ The U-ratio is defined as U � ratio Z u m (2) where u represents the number of undistinguishable granules and m represents the quantity of all granules. This index is to calculate the proportion of undistinguish- able granules to all granules. If there are 10 granules and two of them are undistinguishable granules, which means u is equal to 2 and m is equal to 10, then the U-ratio is equal to 0.2. By using these two indexes, we can determine the similarity of the IGs. In the present study, the larger the H-index the better it is, because it means that more objects in one granule possess the same class. There is no need to set up the index to a fixed value. The size of the index depends on the domain knowledge or how large an error you can tolerate. On the other hand, the U-ratio is the opposite. As far as the U-ratio is concerned, the smaller the better. It’s difficult to process an undistinguishable granule, so we need to view them carefully. However, we try to avoid this situation by setting the U-ratio as small as possible. In other words, if we select a specific similarity where the H-index is larger and the U-ratio is smaller, then this similarity is the best solution. Table 3 Two IGs represented by hyperbox form IGs Attributes X1 X2 A ðaK1 ; a C 1 Þ ða K 2 ; a C 2 Þ B ðbK1 ; b C 1 Þ ðb K 2 ; b C 2 Þ Table 4 The IGs with sub-attributes Original attributes X1 X2 Sub-attributes X11 X12 X13 X21 X22 X23 IGs ½aK1 ; b K 1 � ½b K 1 ; a C 1 � ½a C 1 ; b C 1 � ½a K 2 ; b K 2 � ½b K 2 ; a C 2 � ½a C 2 ; b C 2 � A 1 1 0 1 1 0 B 0 1 1 0 1 1 C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541534 2.3. Representing the information granules In this section, we utilize hyperboxes to represent IGs (Pedrycz & Bargiela, 2002). A hyperbox [b] defined in R n is fully described by its lower (b K ) and upper corner (b C ), where b K and b C are vectors in R n . An important and frequently used universal set is the set of all points in the n-dimensional space. This set is denoted as R n . Using b K and b C we can express the hyperbox as [b]Z[bK,bC].Consider two IGs (hyperboxes) AZ [a] and BZ[b] defined in R2. More explicitly, we follow a full notation [a]Z[aK,aC] and [b]Z[bK,bC]. These two granules are described as Table 3. As Fig. 2 shows, there are overlaps between two granules A and B. This makes it difficult to handle by knowledge acquisition tools. This is because most of knowledge acquisition algorithms are not designed to deal with IGs, especially when overlapping occurs between granules. Unfortunately, the overlapping situation always happens in real world. In this study, we introduce the concept of ‘sub- attributes’ to tackle the problem of overlaps between granules. We can explain this idea of ‘sub-attributes’ by using Fig. 2. In axis X1 (attribute 1), the overlapping part of two granules are separated into overlapping part (½bK1 ; a C 1 �) and non- overlapping parts (½aK1 ; b K 1 � and ½a C 1 ; b C 1 �). These sub-intervals, ½aK1 ; b K 1 �, ½b K 1 ; a C 1 � and ½a C 1 ; b C 1 �, are named as X11, X12, X13 which are so called ‘sub-attributes.’ The binary variable which is employed to be the values of sub-attributes represents whether an IG contains these sub-intervals or not. The results of rewriting the IGs by using sub-attributes can be found in Table 4. We divide the original attribute X1 into sub-attributes X11, X12, X13; and attribute X2 into X21, X22, X23. Then, these two granules are rewritten by replacing the original attributes with sub-attributes. By introducing the concept of B A − 1b − 1a + 1a + 1b 1X 2X + 2b + 2a − 2b − 2a Fig. 2. The overlap between IGs. sub-attributes, we can easily extract knowledge from the granules even if the overlapping situation always exists. This method can maintain the complete characteristics of data. The IGs with addition of sub-attributes are suitable for all knowledge acquisition algorithms. It is not required to adjust the computational architecture of these algorithms. However, too many sub-attributes may be generated in the situation of natural overlapping which the values of the condition attributes are continuous and diverse. Therefore, as we often do in data preparation phase of data mining, we suggest descretizing data before implementing KAIG model to control the number of sub-attributes. 3. Proposed methodologies This section describes in detail the procedure of the KAIG model. First, we address how the IGs are formed from numerical data. Secondly, H-index and U-ratio are introduced to determine the level of granularity which can be used to construct IGs in Fuzzy ART. Then, we try to describe IGs by using sub-attributes and extract knowledge from them. The well-known dataset, iris, will serve as an illustrative example. 3.1. The KAIG model Fig. 3 shows the proposed KAIG model. We explain it by the following steps: Knowledge rules Numerical data Select the level of granularity Information granules representation Knowledge acquisition Check granularity by using H-index & U-ratio Not satisfied SatisfiedInformation granulation Fig. 3. knowledge acquisition via information granulation (KAIG) model. 0% 20% 40% 60% 80% 100% 120% 1 0.9 0.8 0.7 0.6 0.5 0.4 Similarity C la ss if i. A cc u Fig. 5. The performance of classification (Iris data). C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541 535 3.1.1. Step 1: Information granulation In step 1, we use Fuzzy ART to construct IGs. But, first thing we need to determine is to select the suitable level of granularity (vigilance). The IGs are formed by the selected granularity. The initial value of granularity is set 1 and then decrease gradually until find one satisfying criteria of H-index and U-ratio. The found suitable granularity is employed to construct IGs. 3.1.2. Step 2: Information granules representation IGs are represented in a suitable form that can be handled by knowledge acquisition tools. As mentioned in Section 2.3, these formed IGs are described in hyperboxes. Then, the sub- attributes are applied in these IGs to solve the problem and finally, we can extract knowledge from these IGs. 3.1.3. Step 3: Knowledge acquisition After describing IGs appropriately and tackling the overlapping situation, we can use knowledge acquisition tools to extract knowledge rules from the granules. In this study, we will compare three famous data mining algorithms, C4.5, Rough sets and neural network (back-propagation), to evaluate their effectiveness in KAIG model. 3.2. llustrative example We apply the KAIG model to the well-known data set, iris. It is comprised of 150 examples. We rearrange it randomly and divide it into two subsets, training set (100 objects) and test set (50 examples). We will illustrate the process of KAIG step by step. 3.2.1. Step 1: Information granulation We input the 100 training examples to the Fuzzy ART to form IGs. We set the parameters of Fuzzy ART aZ0.01 and bZ1. The number of IGs varies with the different level of similarity (vigilance). In this study, similarity value varies gradually from 1 to 0. The similarity 1 represents the numerical data. Next, we need to determine which similarity is suitable by the H-index and the U-ratio. The H-index is ’the larger-the- better’ and the U-ratio is ‘the smaller-the-better’. In Fig. 4, we can find more than one similarity that satisfies this criterion. 0 0.2 0.4 0.6 0.8 1 1.2 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 Similarity H-index U-ratio Fig. 4. The H-index and U-ratio of the iris data. These similarities are 0.95–0.8 and 0.7–0.55, where H-indexZ 1 and U-ratioZ0. Their performances of classification, as described in Fig. 5, are equal to each other. All classification accuracies are equal to 100%. When the performances are equally good, the amount of granules becomes another criterion for selecting the similarity. In this study, we use IGs instead of numerical data to acquire knowledge and make decisions. If the smaller similarity is selected, the lesser the amount of granules will be dealt with. This smaller amount of granules may save some training time during the building of the model. Therefore, we select a similarity of 0.55 and the amount of granules is 3. 3.2.2. Step 2: Representing the IGs We describe these three granules in hyperboxes form and they are shown in Table 5. Li represents the lower bound of attribute values, and Ui represents the upper limit of attribute values in the ith granule. Take granule #1 for example, it contains 33 objects. In condition attribute A, the minimum is 4.4 and the maximum is 5.7. We utilize the low limit and upper limit to describe all examples in the same one granule. Granule 1 possesses the same class, setosa. Granule 2 contains 33 examples which are of the same class, versicolor. Granule 3 is comprised of 34 examples which have the same class, virginica. Next, the original attributes are divided into several sub- attributes. Table 6 shows the IGs and their sub-attributes. The four original condition attributes (A, B, C, D) are divided into 17 sub-attributes (A1,.,D4). These 17 sub-attributes are used Table 5 The IGs with the similarity of 0.55 No. of granules Condition attribute Classes (No. of examples) A B C D Setosa Versicolor Virginica #1 L1 4.4 2.3 1 0.1 33 0 0 U1 5.7 4.2 1.9 0.6 #2 L2 5 2.2 3 1 0 33 0 U2 6.8 3.4 5.1 1.8 #3 L3 5.6 2.2 4.8 1.4 0 0 34 U3 7.9 3.8 6.9 2.5 T a b le 6 T h e IG s w it h su b -a tt ri b u te s O ri g in a l a tt ri b u te s A B C D C la ss e s S u b - a tt ri b u te s A 1 A 2 A 3 A 4 A 5 B 1 B 2 B 3 B 4 C 1 C 2 C 3 C 4 D 1 D 2 D 3 D 4 4 .4 – 5 .0 5 – 5 .6 5 .6 – 5 .7 5 .7 – 6 .8 6 .8 – 7 .9 2 .2 – 2 .3 2 .3 – 3 .4 3 .4 – 3 .8 3 .8 – 4 .2 1 – 1 .9 3 – 4 .8 4 .8 – 5 .1 5 .1 – 6 .9 0 .1 – 0 .6 1 – 1 .4 1 .4 – 1 .8 1 .8 – 2 .5 G ra n u le N o .1 1 1 1 0 0 0 1 1 1 1 0 0 0 1 0 0 0 S e to sa G ra n u le N o .2 0 1 1 1 0 1 1 0 0 0 1 1 0 0 1 1 0 V e rs ic o lo r G ra n u le N o .3 0 0 1 1 1 1 1 1 0 0 0 1 1 0 0 1 1 V ir g in ic a C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541536 as the inputs for the operation of knowledge acquisition algorithms. 3.2.3. Step 3: Knowledge acquisition The rough sets method can be utilized to remove super- fluous sub-attributes and to acquire knowledge. The theory of rough sets emerged as a major mathematical tool for discovering knowledge. A fundamental principle of a rough set-based learning system is to discover redundancies and dependencies between the given features of a problem to be classified (Mitra, Pal, & Mitra, 2002). In the rough set method, a reduct is the minimal subset of attributes that enable the same classification of objects with full attributes. All results of rough sets are operated by Rosseta software. Readers can find additional information on the theory of rough sets in the references (Hu, Cercone, Han, & Ziarko, 2002; Walczak & Massart, 1999). The knowledge rules extracted by rough set method are listed as follows: Rule 1 IF B4Z1 THEN ClassZsetosa; Rule 2 IF D2Z1 THEN ClassZversicolor; Rule 3 IF B4Z0 AND D2Z0 THEN ClassZvirginica; These knowledge rules can be translated as follows: Rule 1 ATTRIBUTE B2(3.8,4.2] THEN ClassZsetosa; Rule 2 ATTRIBUTE D2(1.0,1.4] THEN ClassZversicolor; Rule 3 ATTRIBUTE B2(3.8,4.2] AND ATTRIBUTE D; (1.0,1.4] THEN ClassZvirginica; These knowledge rules are applied to test the remaining 50 examples. Table 7 is the minimal reduct of the testing granules. The sub-attributes of testing granules, B4 and D2, are put into these extracted knowledge rules. The predicted decisions are fully equal to the true ones. Therefore, the classification accuracy is 100%. In this illustrative example, we reduce some unnecessary detailed information by acquiring knowledge from IGs, but the classification accuracy remains high. Also, the knowledge rules for decision-making are fewer than those extracted from numerical data, which may save the response time of a decision. Table 8 shows the comparison of classification performances. 4. Evaluation of KAIG model To evaluate the effectiveness of the KAIG model, five data sets which come from databank of UCI machine learning group Table 7 The minimal reduct of IGs for testing IGs No. B4 D2 Classes 3.8–4.2 1–1.4 Predicted True #1 1 0 Setosa Setosa #2 0 1 Versicolor Versicolor #3 0 0 Virginica Virginica Table 8 The comparison of processing with information granules and numerical data Methods Rough sets KAIG Data type Numerical data (similarityZ1.0) Information granules (similarityZ0.55) Classification accuracy 100% 98% 100% 100% No. of rules 16 3 C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541 537 (http://www.ics.uci.edu/wmlearn/) are considered in this section. Table 9 provides brief explanation about the data background, including data size, number of features, data characteristics (binary/continuous), and defined classes. Before implementing, we divide all data sets into training set and testing set with the proportion of 3:1. With the help of the H-index and the U-ratio shown in Fig. 6, we can find the suitable similarity of these data sets. According to these determined similarities, numerical data is transformed into IGs by Fuzzy ART. Then, three famous knowledge acquisition algorithms, neural network (BP), decision tree (C4.5 algorithm) and the rough set method, are utilized. Professional II PLUS is employed to build neural network in this study. The optimal neural network (BP) parameter settings, structure and learning iterations shown in Table 10 are obtained by trial and error. Decision trees built by C4.5 are models which each node is a test on an individual variable and a path from the root to a leaf is a conjunction of conditions required for a given classification. See5 (C4.5 commercial version) software was utilized to construct a decision tree in this study. In See5, there are two parameters that can be tuned, during the pruning phase: the minimal number of examples represented at any branch of any feature- value test; and the confidence level of pruning. To avoid the occurrence of over-fitting and generating a simple tree, 2 was set as the minimum number of instances at each leaf and the confidence level for pruning was set at 25%. The inputs and outputs of decision tree and the rough set method are condition attributes and defined classes, respectively. The comparisons of implementation results are provided in Table 11. Except WDBC, KAIG model has better classification performances in the other five data sets than those of traditional methods which use numerical data. In average, the Table 9 The background of five data sets Data set Title No. of instances No of attributes WDBC Wisconsin diagnostic breast cancer 683 (699 minus 16 missing data) 9 (remove first attribute—‘ID’ CE Car evaluation database 1728 6 TAE Teaching assistant evaluation 151 5 BUPA BUPA liver disorders 345 6 WINE Wine recognition data 178 12 PIMA Pima Indians diabetes 768 8 classification accuracy increases 2.33% and the number of rules is reduced by 48.67% compared with traditional methods. In KAIG model, we can use different kind of knowledge acquisition tools and the results will be different. The classification accuracy averagely increases 0.86, 2.238, 1.182% by applying Rough sets, C4.5 and BP, respectively. In addition, C4.5 has fewer number of knowledge rules (12 rules in average) than those of Rough sets (84.8 rules in average). Therefore, C4.5 is more suitable to be employed in KAIG model than the other two methods. 5. Implementation in imbalanced data This section will apply KAIG method to overcome the class imbalance problems. C4.5 and SVM are usually utilized as benchmarks or basic learners in related works (Batista et al., 2004; Guo & Viktor, 2004; Huang et al., 2004; Jo & Japkowicz, 2004; Provost & Fawcett, 2001; Radivojac, Chawla, Dunker, & Obradovic, 2004). Therefore, the experimental results of KAIG will be compared with these two methods. A brief introduction about SVM can be found in (Cristianini & Shawe-Taylor, 2000; Wu & Chang, 2005). 5.1. Performance measures The easiest way to evaluate the performance of classifiers is based on the confusion matrix described as Table 12. TP, FP, TN and FN are defined as bellows. TP the number of True Positive examples FP the number of False Positive examples TN the number of True Negative examples FN the number of False Negative examples Traditionally, the performance of a classifier is evaluated by considering the overall accuracy against test cases. However, when learning from imbalanced data sets, the measure is often not sufficient. For example, it is straightforward to create a classifier having an accuracy of 95% in a domain where the majority class proportion corresponds to 95% of the examples, by simply forecasting every new example as belonging to the majority class. Another fact is the metric considers different classification errors to be equally important. But as we know, a highly imbalanced class problem does not have equal error Data characteristics Class distribution ) All discrete Benign (65.5%) malignant (34.5%) All discrete Unacceptable (70.023%) acceptable (22.222%) good (3.993%) very good (3.762%) 1-continuous 4-discrete Low (32.45%) medium (33.11%) high (34.44%) All continuous Class 1 (42.03%) class 2 (57.97%) All continuous Class 1 (33.15%) class 2 (39.89%) class 3 (26.96%) All continuous Healthy (65%) diabetic (35%) http://www.ics.uci.edu/~mlearn/ 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0.95 0.8 0.65 0.5 0.35 0.2 0.95 0.8 0.65 0.5 0.35 0.2 Similarity Similarity H-index U-ratio 0 0.95 0.8 0.65 0.5 0.35 0.2 0.95 0.8 0.65 0.5 0.35 0.2 0.95 0.8 0.65 0.5 0.35 0.2 0 0.2 0.4 0.6 0.8 1 1.2 WDBC CE Similarity Similarity Similarity TAE WINE BUPA H-index U-ratio H-index U-ratio H-index U-ratio H-index U-ratio Fig. 6. The H-indexes and U-ratios of five data sets. C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541538 costs that favor the minority class, which is often the class of primary interest. Therefore, following the available studies (Batista et al., 2004; Estabrooks, Jo, & Japkowicz, 2004; Guo & Viktor, 2004; Provost & Fawcett, 2001; Radivojac, Chawla, Dunker, & Obradovic, 2004), we use overall accuracy (including positive accuracy and negative accuracy), G-mean and receiver operation characteristic (ROC) curve to evaluate our KAIG model. The G-mean is defined as ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Positive Accuracy!Negative Accuracy p (3) Table 10 The setting of parameters in neural network (BP) Data Set Data type Structure WDBC Numerical 9-11-1 Granule(0.85) 90-160-1 CE Numerical 6-9-1 Granule(0.95) 21-35-1 TAE Numerical 5-6-1 Granule(0.85) 69-120-1 BUPA Numerical 6-5-1 Granule(0.85) 26-31-1 WINE Numerical 13-5-1 Granule(0.8) 35-7-1 where positive accuracy and negative accuracy are calculated as TP/(FNCTP) and TN/(TNCFP). This measure is to maximize the accuracy on each of two classes while keeping these accuracies balanced. For instance, a high positive accuracy by a low negative accuracy will result in poor G-mean. Another index is ROC curve, which is a technique for summarizing a classifier’s performance over a range by considering the tradeoffs between TP rate and FP rate. The TP rate and FP rate are calculated as TP/(FNCTP) and FP/(FP CTN). We use the term ROC space to denote the coordinate system used for visualizing classifier’s performance. In ROC space, TP rate is represented on the Y-axis and FP rate is represented on the X-axis. Each classifier is represented by the point in ROC space corresponding to its (FP rate, TP rate) pair. A ROC analysis also allows the performance of multiple classification functions to be visualized and compared simultaneously. The area under ROC curve (AUC) represents the expected performance as a single scalar. The AUC has a known statistical meaning: it equals to the Wilconxon test of ranks, and is equivalent to several other statistical measures for evaluating classification and rank models (Hand, 1997). 5.2. Diagnosis data The imbalance class problems often happen in medical diagnosis data. Therefore, pima-Indians-diabetes whose infor- mation shows in Table 9 is employed to verify effectiveness of our model. Results for this data set, shown in Table 13, were averaged over four-fold cross validation (CV) experiments, which the data set was partitioned into four equal sized sets and each set was then in turn used as the test set. Besides, in order to test the robustness of KAIG model, we reduce the proportion of minority class from 35% to 10% and 5% by removing the number of minor examples randomly. In the experiments of 35%, 10% and 5%, the results indicate that KAIG model has better performance than those of SVM and C4.5 against highly imbalanced data sets, in term of the negative accuracy. In average, KAIG owns 58.08% of negative accuracy far better than 14.55% of SVM and 27.49% of C4.5. It means KAIG has excellent capability of detecting minor examples (diabetic patients). Meanwhile, KAIG does not lose overall accuracy and positive accuracy. They are even better than those of SVM and C4.5 in experiment of 35%. Learning rate Momentum Iterations 0.2 0.9 20,000 0.2 0.9 20,000 0.2 0.9 10,000 0.3 0.9 20,000 0.2 0.9 20,000 0.2 0.9 20,000 0.3 0.8 30,000 0.2 0.9 15,000 0.3 0.7 10,000 0.2 0.8 15,000 Table 11 The comparison of classification performance Methods Classification accuracy Data type Numerical data (similarityZ1.0) Traditional methods Granules (similarityZ0.85) KAIG Train (%) Test (%) No. of rules Train (%) Test (%) No. of rules WDBC Rough sets 100 92.23 212 100 89.47 58 Decision tree (C4.5) 97.5 97.06 10 93.4 94.74 4 Neural network (BP) 96.66 100 – 100 89.64 – CE SimilarityZ1.0 SimilarityZ0.95 Rough sets 100 89.58 385 100 88.96 207 Decision tree (C4.5) 97.4 92.8 75 98.4 95.58 36 Neural network (BP) 91.18 91.09 – 94.04 92.80 – TAE SimilarityZ1.0 SimilarityZ0.90 Rough sets 84.96 84.21 90 95.95 87.37 68 Decision tree (C4.5) 60.2 47.36 13 64.9 48.39 11 Neural Network (BP) 68.23 69.05 – 78.61 84.21 – BUPA SimilarityZ1.0 SimilarityZ0.85 Rough sets 100 63.95 165 100 66 80 Decision tree (C4.5) 76.4 65.1 15 78.2 70 5 Neural Network (BP) 69.35 64.47 – 100 66.15 – WINE SimilarityZ1.0 SimilarityZ0.8 Rough sets 100 93.18 31 100 95.65 11 Decision tree (C4.5) 95.6 90.9 6 96.7 95.7 4 Neural network (BP) 87.63 86.49 – 85.87 84.21 – Table 12 Confusion matrix for binary class problem Predicted positive Predicted negative Actual positive TP (the number of true positive) FN (the number of false negative) Actual negative FP (the number of false positive) TN (the number of true negative) C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541 539 Both G-mean and ROC curves shown in Fig. 7 also demonstrate the superiority of our method. In extreme skewed data (10 and 5%), G-mean is more sensitive than overall accuracy. When negative accuracy decreases dramatically, Table 13 The results in different proportion of minor class examples Methods KAIG SVM Training Test Training Mean (%) SD (%) Mean (%) SD (%) Mean (%) SD (35%) Overall accuracy 91.97 2.5 78.78 2.5 76.82 1.4 Pos. Acc. 93.07 2.3 84.00 4.5 93.07 0.5 Neg. Acc. 85.24 2.1 70.52 8.0 46.52 4.7 G-mean 90.67 2.59 76.46 3.85 65.73 3.1 (10%) Overall accuracy 95.01 1.1 87.05 2.3 89.93 0 Pos. Acc. 99.33 0.8 92.20 1.2 100 0 Neg. Acc. 52.98 11.9 41.08 20.5 0 0 G-mean 72.16 7.91 59.73 17.0 0 0 (5%) Overall accuracy 97.48 0.7 94.89 1.6 94.94 0 Pos. Acc. 98.54 0.9 98.60 1.7 100 0 Neg.Acc. 72.50 13.2 28.57 26.1 0 0 G-mean 84 7.69 44 33.3 0 0 G-mean can indicate these changes but overall accuracy cannot. ROC curves provide visual results which can easily compare these three methods and find KAIG has best performances (AUC) in different experiments. Decision tree (C4.5) Test Training Test (%) Mean (%) SD (%) Mean (%) SD (%) Mean (%) SD (%) 75.52 2.8 81.50 4.28 74.22 3.1 92.60 2.7 87.94 7.50 83.20 2.8 43.66 3.3 71.40 8.73 57.46 8.3 8 63.56 3.29 78.95 3.24 68.99 4.80 89.93 0 91.55 1.7 88.49 1.6 100 0 98.73 1.9 96.80 3.1 0 0 27.38 20.4 14.29 24.0 0 0 44.43 30.6 23.63 32.1 94.70 0 96.52 0.8 93.56 1.0 100 0 99.47 0.7 98.20 0.8 0 0 41.25 14.9 10.72 13.7 0 0 63 12.7 23 26.9 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 ROC Curves of pima-indians-diabetes data(35%) FP rate 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 FP rate 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 FP rate T P r at e 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 T P r at e 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 T P r at e KAIG SVM C4.5 ROC Curves of pima-indians-diabetes data (10%) KAIG SVM C4.5 ROC Curves of pima-indians-diabetes data (5%) KAIG SVM C4.5 Fig. 7. ROC curves of pima-indians-diabetes data. C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541540 6. Conclusions This study introduces the concept of information granula- tion to solve class imbalance problems. A novel method called KAIG model is presented. In this model, we propose two indexes to determine the level of granularity and the ‘sub- attributes’ concept to describe IGs. The experimental results show that the KAIG model can improve classification performance by reducing unnecessary details of information. We also demonstrate that the proposed method has excellent ability of identifying the minority examples in imbalanced learning tasks. In medical diagnosis data, our method can dramatically increase negative accuracy without losing positive accuracy and overall accuracy. ROC curves and G-mean also illustrate the superiority of KAIG model compared with C4.5 and SVM. Construction of IGs is one of many interesting and important issues in granular computing. IGs are aimed at building efficient and user-centered views of the external world and supporting/facilitating our perception of the surrounding physical and virtual world. In our research, we construct IGs by objects’ ‘similarity’, the parameter (vigilance) of Fuzzy ART. It can define the ‘indistinghishable, similar, coherency and alike’ relations of objects. However, other relations whose definitions are not specific/ concrete, such as ‘functional adjacency’, also can employ to construct IGs. But, it is hard to define these ‘not specific’ relations. Therefore, more efforts of studying different relations are necessary in the future researches. Acknowledgements This work was supported in part by National Science Council of Taiwan (Grant No. NSC 94-2213-E007-059). References Bargiela, A., & Pedrycz, W. (2003). Granular computing: An introduction. Boston: Kluwer Academic Publishers. Batista, G., Prati, R. C., & Monard, M. C. (2004). A study of the behavior of several methods for balancing machine learning training data. SIGKDD Explorations, 6(1), 20–29. Burke, L., & Kamal, S. (1995). Neural networks and the part family/machine group formation problem in cellular manufacturing: A framework using fuzzy art. Journal of Manufacturing Systems, 14(3), 148–159. Carpenter, G. A., Grossberg, S., & Rosen, D. B. (1991). Fuzzy art: Fast stable learning and categorization of analog patterns by an adaptive resonance system. Neural Networks, 4, 759–771. Castellano, G., & Fanelli, A. M. (2001). Information granulation via neural network-based learning. IFSA World Congress and 20th NAFIPS international conference, Vol. 5 pp. 3059–3064. Chawla, N. V., Bowyer, K., Hall, L., & Kegelmeyer, W. (2002). SMOTE: Synthetic minority over-sampling technique. Journal of Artificial Intelli- gence Research, 16, 231–357. Chawla, N. V., Japkowicz, N., & Kolcz, A. (2004). Editorial: Special issue on learning from imbalanced data sets. SIGKDD Explorations, 6(1), 1–6. Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector machines and other kernel-based learning methods. Cambridge: Cambridge University Press. Estabrooks, A., Jo, T., & Japkowicz, N. (2004). A multiple resampling methods for learning from imbalanced data sets. Computational Intelligence, 20(1), 18–36. Grzymala-Busse, J. W., Stefanowski, J., & Wilk, S. (2004). A comparison of two approaches to data mining from imbalanced data. Lecture Notes in Computer Science, 3213, 757–763. Guo, H., & Viktor, H. L. (2004). Learning from imbalanced data sets with boosting and data generation: The DataBoost-IM approach. SIGKDD Explorations, 6(1), 30–39. C.-T. Su et al. / Expert Systems with Applications 31 (2006) 531–541 541 Hand, D. J. (1997). Construction and assessment of classification rules. New York: Wiley. Hu, X., Cercone, N., Han, J., & Ziarko, W. (2002). In T. Y. Lin, Y. Y. Yao, & L. A. Zadeh (Eds.), GRS: A generalized rough sets model, data mining, rough sets and granular computing (pp. 447–460). New York: Physica- Verlag. Huang, K., Yang, H., King, I., & Lyu, M. (2004). Learning classifiers from imbalanced data based on biased minimax probability machine. Proceed- ings of the 04’ IEEE Computer Society conference on computer vision and pattern recognition (CVPR’04) pp. 558–563. Jo, T., & Japkowicz, N. (2004). Class imbalances versus small disjuncts. SIGKDD Explorations, 6(1), 40–49. Mitra, S., Pal, S. K., & Mitra, P. (2002). Data mining in soft computing framework: A survey. IEEE Transactions on Neural Networks, 13(1), 3–14. Pedrycz, W., & Bargiela, A. (2002). Granular clustering: A granular signature of data. IEEE Transactions on Systems, Man, and Cybernetics-Part B: Cybernetics, 32(2), 212–224. Provost, F., & Fawcett, T. (2001). Robust classification for imprecise environments. Machine Learning, 42, 203–231. Radivojac, P., Chawla, N. C., Dunker, A. K., & Obradovic, Z. (2004). Classification and knowledge discovery in protein databases. Journal of Biomedical Informatics, 37, 224–239. Serrano-Gotarredona, T., Linares-Barranco, B., & Andreou, A. G. (1998). Adaptive resonance theory microchips-circuit design techniques. Boston, MA: Kluwer Academic Publishers. Taguchi, G., & Juoulum, R. (2002). The Mahalanobis-Taguchi strategy—a pattern technology system. New York: Wiley. Walczak, B., & Massart, D. L. (1999). Tutorial: Rough sets theory. Chemometrics and Intelligent Laboratory Systems, 47, 1–16. Wu, G., & Chang, E. Y. (2005). KBA: Kernel boundary alignment considering imbalanced data distribution. IEEE Transactions on Knowledge and Data Engineering, 17(6), 786–795. Yao, Y. Y., & Yao, J. T. (2002). Granular computing as a basis for consistent classification problems. Proceedings of PAKDD’02 workshop on toward the foundation of data mining pp. 101–106. Zadeh, L. A. (1979). Fuzzy sets and information granularity. In M. M. Gupta, R. K. Ragade, & R. R. Yager (Eds.), Advances in fuzzy set theory and applications (pp. 3–18). Amsterdam: North Holland. Knowledge acquisition through information granulation for imbalanced data Introduction Granular computing Construction of information granules Selection of granularity Representing the information granules Proposed methodologies The KAIG model llustrative example Evaluation of KAIG model Implementation in imbalanced data Performance measures Diagnosis data Conclusions Acknowledgements References 
work_ovc3spc7cvhv7au52qkbn3ozru	----	Публикации Archives - SZMA Главная Профиль компании Новости История Реализованные проекты Сертификаты Публикации Заказчики Вакансии Услуги Проектирование Функциональная безопасность Информационная безопасность АСУ ТП Проведение аудита систем ПАЗ Расчет надежности Шефмонтаж и пусконаладка Ремонт преобразователей частоты Учебный центр Производство Сборочное производство Шкафы управления насосами Системы АСУ ТП Фильтры гармоник и дроссели Испытательный стенд Cертифицированный сервисный центр Toshiba Программные комплексы Программный комплекс АРБИТР Программный комплекс ДЕЛЬТА-СИ Центр продаж Низковольтные преобразователи частоты Toshiba Низковольтные электродвигатели Toshiba Высоковольтные электродвигатели TMEIC Высоковольтные преобразователи частоты TMeic Устройства плавного пуска Toshiba серий TS/TD/TX Вакуумные контакторы/выключатели Toshiba Интегрированные промышленные контроллеры Toshiba Средства измерения Toshiba Контакты EN Публикации 2020 год Можаева И.А., Струков А.В. Тенденция в развитии стандартов МЭК в области управления надежностью технических систем. Сборник трудов двадцать третьей Всероссийской научно-практической конференции «Актуальные проблемы защиты и безопасности». Т. 2: «Средства противодействия терроризму. ФГБУ РАРАН-Москва, НПО СМ  – СПб.  2020 С.253-260. 14.10.2020 Можаева И.А., Нозик А.А., Струков А.В. Методы решения прямой и обратной задачи оценки функциональной безопасности систем ПАЗ //Сборник трудов двадцать третьей Всероссийской научно-практической конференции «Актуальные проблемы защиты и безопасности». Т. 2: «Средства противодействия терроризму». ФГБУ РАРАН-Москва, НПО СМ – СПб. 2020. С. 196-211. 14.10.2020 2019 год Можаева И.А., Нозик А.А., Струков А.В. Типовые примеры расчета функциональной безопасности систем противоаварийной защиты опасных производственных объектов //Сборник трудов двадцать второй Всероссийской научно-практической конференции «Актуальные проблемы защиты и безопасности». Т. 2: «Средства противодействия терроризму». ФГБУ РАРАН-Москва, НПО СМ — СПб.  2019. С. 486-494. 29.10.2019 Можаева И.А., Нозик А.А., Струков А.В. Методология аудита функциональной безопасности эксплуатируемых систем противоаварийной защиты //Сборник трудов двадцать второй Всероссийской научно-практической конференции «Актуальные проблемы защиты и безопасности». Т. 2: «Средства противодействия терроризму». ФГБУ РАРАН-Москва, НПО СМ — СПб.  2019. С. 456-463. 29.10.2019 Рябинин И.А., Струков А.В. Решение одной задачи оценки надежности структурно-сложной системы разными логико-вероятностными методами //Сборник трудов 15-й Международной научной конференции «Моделирование и анализ безопасности и риска в сложных системах» (МАБР-2019). 2019. С. 1-14. 25.06.2019 Можаева И.А., Нозик А.А., Струков А.В. Алгоритмы расчета показателей функциональной безопасности систем противоаварийной защиты //Сборник трудов 15-й Международной научной конференции «Моделирование и анализ безопасности и риска в сложных системах» (МАБР-2019). 2019. С. 1-7. 25.06.2019 2018 год Скворцов М.С. Проектирование систем ПАЗ с учетом анализа опасностей и риска аварий на опасном производственном объекте //Автоматизация в промышленности. 2018. № 3. C. 61-65. 03.04.2018 Ryabinin I.A., Strukov A.V. Quantitative examples of safety assessment using logical-probabilistic methods //International Journal of Risk Assessment and Management (IJRAM). 2018. Vol. 21. № 1/2. P. 4-20. 17.03.2018 Скворцов М.С. Отказоустойчивый протокол реального времени TСnet для промышленных сетей //Автоматизация в промышленности. 2018. № 1. P. 34-38. 03.03.2018 Скворцов М.С. Программный комплекс АРБИТР – инструмент для проектной оценки надежности и функциональной безопасности систем управления //Автоматизация & IT в энергетике. 2018. № 2. С. 2-7. 19.02.2018 Скворцов М.С. Проектная оценка функциональной безопасности систем противоаварийной автоматической защиты //Безопасность труда в промышленности. 2018. № 1. С. 50-57. 19.01.2018 2017 год Струков А.В., Можаева И.А. Метод Монте-Карло в задачах логико-вероятностного моделирования надежности и безопасности структурно-сложных систем //Труды 8-й Всероссийской научно-практической конференции «Имитационное моделирование. Теория и практика» (ИММОД-2017). 2017. С. 299-303. 19.09.2017 Newer posts Главная Новости Профиль компании Реализованные проекты Проектирование Функциональная безопасность Информационная безопасность Аудит систем ПАЗ Сборочное производство Учебный центр ПК АРБИТР ПК ДЕЛЬТА-СИ Преобразователи частоты Toshiba Сервисный центр Toshiba Шкафы управления насосами Испытательный стенд © СПИК СЗМА 1998-2021. Все права защищены. 
work_oycc2mtdyrdkpfquyovfrgyj3m	----	Auditing file system permissions using Association Rule Mining S. Parkinsona,∗, V. Somarakia, R. Warda aDepartment of Informatics, School of Computing and Engineering, University of Huddersfield, Queensgate, Huddersfield, HD1 3DH, UK Abstract Identifying irregular file system permissions in large, multi-user systems is challenging due to the com- plexity of gaining structural understanding from large volumes of permission information. This challenge is exacerbated when file systems permissions are allocated in an ad-hoc manner when new access rights are re- quired, and when access rights become redundant as users change job roles or terminate employment. These factors make it challenging to identify what can be classed as an irregular file system permission, as well as identifying if they are irregular and exposing a vulnerability. The current way of finding such irregularities is by performing an exhaustive audit of the permission distribution; however, this requires expert knowledge and a significant amount of time. In this paper a novel method of modelling file system permissions which can be used by association rule mining techniques to identify irregular permissions is presented. This results in the creation of object-centric model as a by-product. This technique is then implemented and tested on Microsoft’s New Technology File System permissions (NTFS). Empirical observations are derived by making comparisons with expert knowledge to determine the effectiveness of the proposed technique on five diverse real-world directory structures extracted from different organisations. The results demonstrate that the technique is able to correctly identify irregularities with an average accuracy rate of 91%, minimising the reliance on expert knowledge. Experiments are also performed on synthetic directory structures which demonstrate an accuracy rate of 95% when the number of irregular permissions constitutes 1% of the total number. This is a significant contribution as it creates the possibility of identifying vulnerabilities without prior knowledge of how to file systems permissions are implemented within a directory structure. Keywords: Access control, Auditing, Association Rule Mining. 1. Introduction File systems are common amongst the majority of computer operating systems and from a user perspec- tive their primary use is to securely store files. Modern, multi-user computer systems contain high quantities of data that require strong access control mechanisms to restrict data access to intended users. Different operating systems provide different implementations of access control, but common to the most prevalent is that they provide a customisable architecture for access control. This is implemented through the use of both coarse- and fine-grained permissions (De Capitani di Vimercati et al., 2003). Coarse-grained permis- sions are predefined levels (e.g read, write, full control, etc.) and fine-grained permissions are customised permissions created from a set of predefined attributes to represent highly customised access control rules. Using a mixture of both types of permissions provides a flexible architecture which can accommodate a large verity of different access control levels. However, this flexible nature also provides the possibility for complex permission relationships and anomalous permissions which often go undetected. ∗Corresponding author Email addresses: s.parkinson@hud.ac.uk (S. Parkinson), v.somaraki@hud.ac.uk (V. Somaraki), r.ward@hud.ac.uk (R. Ward) Preprint submitted to Expert Systems with Applications February 20, 2016 Administration of file system permissions on large file systems is both a challenging and cumbersome task as it is often difficult to conceptualise the large volumes of information available. Due to the complexities of managing permissions, unforeseen weak and incorrect allocations are often made. These complexities are usually the result of there being a large number of directories to secure, a large number of users that need to be correctly assigned, and a large number of access control rules. There is a wide range of literature discussing these complexities (Cao and Iverson, 2006; Beznosov et al., 2009; De Capitani di Vimercati et al., 2003), and many factual guides have been produced for different operating systems (Thomas, 2010; Solomon, 2005). However, even with such guides, users are left to analyse the large amount of access control information independently. The level of their knowledge regarding access control and their system’s configuration often directly determines the quality of their audit, and thus produces a heavy reliance on expert knowledge. When analysing for vulnerabilities within file system permissions they can be divided into two groups of; (1) known system vulnerabilities, and (2) those relative to the access control structure implemented within a system. A trivial example of a known system fundamental is that users should not have access to an important system directory (e.g C:\windows\system32). These are programmatically easy to find by using a predefined knowledge-base of potential vulnerabilities. Identifying such vulnerabilities is at most O(n×v) where n is the number of access control entries to examine and v is the number of known vulnerabilities. An example of a relative vulnerability is an incorrect assignment of permission with respect to an organisation’s implementation of access control. For example, the anomaly of one user having write privileges on a directory where all other users have read access. Such anomalies are very difficult to identify within access control as there is no quick method of determining potential vulnerabilities. The consequences of both types of vulnerabilities can be severe and can be generated from both user and software actions. For example, if a user has an elevated level of permission on a network directory structure, they could unintentionally modify or remove important data. A more severe consequence would be if the user could access data which is sensitive, as this could result in the organisation breaching the Data Protection Act. It is also possible that software (such as viruses, etc.) executing under the user’s credentials could exploit their file system permissions to perform malicious activity. Trend mining, in the form of Association Rule Mining (ARM) (Ma, 1998), has been extensively used to identifying anomalies and irregularities in many different application areas (Cheng et al., 2015). ARM is a method of automatically identifying interesting relationships amongst variables in large datasets. Interesting relationships are often those that frequently occur; however, in some applications, such as the one presented in this paper, an interesting relationship is one which occurs infrequently. There have been many successful applications of ARM for identifying interesting (both frequent and infrequent) relationships in large datasets. For example, in finance (Yu et al., 2009; Barak and Modarres, 2015), medical data (Somaraki et al., 2011, 2015), and cellular networks (Khatib et al., 2015). The use of ARM is therefore a natural selection for applying to the challenge of identifying irregularities in file system permissions. The work presented in this paper is tailored towards Microsoft’s New Technology File System (NTFS); however, it can be easily adapted to other file systems. The majority of multi-user environments in organ- isations will use Microsoft’s NTFS for providing a distributed mechanism of file storage. There are many complexities associated with administrating and auditing file system permissions on Microsoft’s NTFS which can result in the creation of vulnerabilities. The complexity of identifying these vulnerabilities has resulted in a unhealthy reliance on expert knowledge. The work presented in this paper helps to remove this unhealthy balance on expert knowledge and provide a method of auditing file system permissions for all NTFS users. This paper first provides a review into related work in the area of aiding auditing of file system permis- sions. The next section the provides a description of NTFS access control structure, highlighting auditing complexities. At the end of this section, detail is provides on how file system permissions are modelled in the work presented in this paper, including the use of an algorithm to combine permissions to determine the effective permission. This then leads to a description of association rule mining techniques which are useful for identifying irregularities in file system permissions. In this section the chosen technique is also discussed and justified. Empirical observations are then provided where the developed technique is tested on five diverse real-world file systems. In this empirical observation, the results are compared to those of a domain expert. 2 2. Related work Access control is typically defined as a relational model over the following domains: O the set of objects (i.e users), the set of resources R and the set of permissions P . Access control is a characteristic function on the set A ⊆ S × O × R. A subject s is granted permission r over resource o iff 〈s, o, r〉 ∈ A. Access control models are typically called the access matrix. In many operating systems the access matrix is stored as an access list, which is associated with a resource object and is used to list all subjects and their permissions. In NTFS, access lists are implemented as Discretionary Access Control List (DACL) models, where only the owner of a resource is authorised to change its access permissions. However, in multi-user environments it is possible for any user with sufficient permission to take ownership of a resource or change its permission. Other operating systems also implement access control lists to manage file system permissions. Unix (including Max OSX and Linux) and Portable Operating System Interface (POSIX) compliant systems have a simple system for managing individual file permissions. This is the ability to assign a predetermined set of coarse-grained permissions (often called “traditional Unix permissions”). Most of these systems also support the use of Access Control Lists (ACL), such as POSIX.1e ACLs (coarse-grained) (Nemeth, 2010) or NFSv4 ACLs (fine-grained) (Pawlowski et al., 2000). Interestingly, the implementation between NFSv4 and NTFS ACLs is very similar and also caters for fine-grained permissions. The ability to create fine-grained permissions significantly increases the complexity of the access control implementation. For this reason, and due to strong similarities between NFSv4 and NTFS, the rest of this section will describe in more detail the access control implementation of NTFS. There are many tools available to assist with the examination of NTFS permissions allocation (Microsoft, 2006b,a). However, there is one common weakness in that they all still require expert knowledge to analyse the output of the tools to determine if there are any weaknesses. Previous work in the area of file system permission analysis resulted in the development of a novel tool for permissions administration that allows users to easily view file system permissions for large directories (Parkinson and Crampton, 2013; Parkinson and Hardcastle, 2014). The tool provides features to help the user identify vulnerabilities and view necessary information to make better informed permission allocations. For example, the reduction in reported permis- sions by not displaying inherited permissions, and allowing the user to filter and show effective permission for a specified user, are two prominent features. Identifying anomalies or irregularities in system security is by no means a new topic (Lazarevic et al., 2003; Bhuyan et al., 2014). For decades, researchers have been developing techniques and tools to identify security anomalies. For example, recent works have covered topics from the identification of anomalous user behaviour in social networks (Viswanath et al., 2014), anomalies in network traffic (Catania et al., 2012; Mahoney, 2003), anomaly detection in wireless networks (Islam and Rahman, 2011; Xie et al., 2011), and the anomaly detection in power station security (Ten et al., 2011). All these studies generate good results and demonstrate the potential of using machine learning and statistics to identify anomalies. Recent research in the area of file systems includes the construction of a Bayesian network and neural network from the predetermined knowledge of the manipulation of file system artefacts for file system forensic analysis (Khan, 2012). Research has been carried out into developing tools for identifying vulnerabilities in file system access control based on some predetermined knowledge-base of vulnerability definitions. Nalgurd et. al. (Naldurg et al., 2006) use an inference engine alongside a snapshot of the access control implementation with static knowledge, and in further work they produce a technique where a model checking algorithm (Naldurg and KR, 2011) is used to identify vulnerabilities. Another area that has recently received significant interest in the access control community is that of access control and anomaly detection for cloud computing. For example, recent work by Hu et al. (2013) both motivates the need for identifying anomalies in web access control policies and provides a potential solution through the use of a binary decision diagram for anomaly discovery and resolution. These methods show significant ability at identifying vulnerabilities, but they require statically defining a knowledge-base of vulnerabilities in respect to relative file system permissions which is heavily reliant on expert knowledge. The focus of the work presented in this paper is to provide a mechanism to identify likely vulnerabilities without the reliance on expert knowledge. Literature suggests that there has been no work in the detection of irregularities in file system security without the need to statically define what constitutes an 3 Attribute Number (a) Description 1 Full control 2 Traverse folder \ execute files 3 List folder \ read data 4 Read attributes 5 Read extended attributes 6 Create files \ write data 7 Create folders \ append data 8 Write attributes 9 Write extended attributes 10 Delete subfolders and files 11 Delete 12 Read permissions 13 Change permissions 14 Take ownership Table 1: Individual attributes. irregular permission. This includes the potential application of Association Rule Mining, This is surprising considering the vast user-base and the potential benefits that can be achieved. 3. NTFS Implementation and Modelling In this section, a description of how NTFS permissions are implemented is provided. Alongside this description is a discussion of how its implementation can results in complexities which ultimately can result in vulnerabilities. Following this, a description is provided of how NTFS permissions are translated into a relational object-model. 3.1. Access Mask NTFS implements DACLs by applying a DACL to each object within the file system. Each DACL will contain a Security Identifier (SID) which is a unique key that identifies the owner of the object and the primary associated group. The structure of the DACL is a sequential storage mechanism which contains access control entries (ACEs). An ACE is an element within a DACL which dictates the level of access given to the interacting subject. The ACE contains a SID that identifies the particular subject, an access mask which contains information regarding the level of permissions and the inheritance flags. An ACE within the NTFS is a set of attributes, p, from the predefined set of attributes p ⊆ P , P = {a1, . . . , an}. Table 1 shows the fourteen predefined attributes in the NTFS. The NTFS provides six levels of standard coarse- grained permission that consist of a combination of predefined attributes. The NTFS also allow for the creation of special fine-grained permissions that are constructed from a combination of fourteen permission attributes (Russel et al., 2003). The potential to create a special permission is a useful feature; however, it can introduce complexity as it requires detailed knowledge regarding the authority that each attribute holds (Thomas, 2010). A example of where a special permission could be used is if the user wanted to assign a user or group the standard access level of Modify for all the contents of a shared directory. However, this has adverse consequences as the Modify permissions will allow the user(s) to be able to delete the folder itself. A potential solution here is to assign the user or group the default permissions level of Modify, and then modify the permissions’ attributes so that only subfolders and files can be deleted, and not the folder itself. This modification would result in the creation of a special permission. 4 3.2. Propagation and Inheritance In this section, a discussion is provided on the different mechanisms through which NTFS permissions can propagate through a directory structure. There are two types of ACE entires within the DACL; (1) Explicit and (2) Inherited. Explicit entires are those that are applied directly to a DACL, whereas inherited are propagated from the directory’s parent. The type property of the ACE allows programmatic determination of whether the permission has been explicitly assigned to a directory or if it was inherited from a parent directory. Furthermore, it is also possible to create custom fine-grained inheritance rules. Such special inheritance rules can easily be overlooked during administration and auditing, resulting in theunintended propagation of access. 3.3. Accumulation Permissions accumulation creates the potential for an interacting subject to receive permissions from multiple different policies. Any interacting subject within the NTFS can be assigned to access control groups. This means that they do not have to be explicitly added in the ACE, they could be an associated group member. Such policy combination is controlled by the Local Security Authority Subsystem Service (LSASS). The high-level functionality of this service is to create an ordered union all relevant policies. However, there are a few complexities that can occur in this combination process. These are: (1) explicit permissions take priority over those inherited, (2) explicit deny permissions always take priority over any other assigned permission, and (3) permissions inherited from closer relatives take priority over those further away. It would be logical to assume that deny permissions would always take precedence over apply permissions to ensure that the user operates at the least possible level of access. However, the first point makes it possible for an inherited deny permission to never be reached. This goes against a fundamental aspect of policy combination that a deny permission should never be ignored. If a situation where a user is able to ignore a deny permission was to arise, a system administrator would need to be aware of this so that they could take rectifying action. In addition to explicit permissions taking priority over inherited permissions, inherited permissions closer to the derrived directory will take priority over those more distance. For example, a folder’s inherited permissions will take priority over those inherited from their grandparent. To summarise, the ordered hierarchy for policy combination is: 1. Explicit deny. 2. Explicit allow. 3. Inherited deny. 4. Inherited allow. 3.4. Group Membership A powerful feature of NTFS access control is that of group membership. An interacting subject (group, user, or process) that interacts with the file system can hold membership to another group. This creates the possibility for permissions to be inherited from all of the associated groups. Users will often be grouped together to mange management easier. This separation of duty can often be seen in many organisations. For example, users within a finance department will have different permissions that of management. As Hanner, et. al. (Hanner and Hörmanseder, 1999) identifies, understanding effective file permissions can become increasingly more complex by group association. This is because to evaluate a user’s effective permission would require the knowledge of which groups that they are inheriting from. It should be noted here that it is not a direct mechanism of how NTFS access control is implemented, rather it results from how Microsoft allows for users, groups and processes to be managed and controlled through group memberships. 5 Algorithm 1: Depth-first recursive permission extraction algorithm, returning an ordered list of ef- fective permissions for each object within the directory structure. Input: Initial directory d Output: Set of ordered effective permissions E = e1, e2, . . . , en where en = {d, o, p}. Here d is the directory resource, o is the object, and p is the permission level, p = {a1, a2, . . . , an} 1 Output: Membership relation set, G = {g1, g2, . . . , gn} where gn = {o1, o2, . . . , on} 2 Algorithm algo() 44 P ← proc(d) 66 return 7 1 Procedure proc(directory d) 2 pACL ← d(ACL) 3 foreach subdirectory c of d do 4 cACL ← c(ACL) 5 if cACL ! = pACL then 6 ex = ∅, in = ∅, r = memberships(G) 7 foreach ACE a in cACL do 8 foreach Membership g in r do 9 if isExplicitDeny(a) then 10 E ← (c, getObj(a), getPerm(a)) 11 break 12 else 13 else if isExplicitAllow(a) then 14 E ← (c, getObj(a), getPerm(a)) 15 break 16 else if isInherited(a) then 17 if isInheritedDeny(a) then 18 E ← (c, getObj(a), getPerm(a)) 19 break 20 else if isExplicitAllow(a) then 21 E ← (c, getObj(a), getPerm(a)) 22 break 23 end 24 25 26 end 27 3.5. Object-centric Modelling In order to represent the accumulated permission that each object holds on a file system resource, is it necessary to adopt an object-centred approach to permissions modelling. In this section, the description of NTFS’s access control structure is translated into an object-centred model suitable representation which can be used for association rule mining. Here each interacting object (user, group, etc.), o, holds a combination of permission attributes which are accumulated based on (1) the assigned permissions, (2) propagated and inherited permissions, and (3) group membership. Algorithm 1 has been developed to iterate over all the directories and create a set of effective object permissions for each directory. Each effective permission for an object is the set of permission attributes, p, that o holds on a directory, d. As seen in Algorithm 1, the hierarchy for permission accumulation (Section 3.3) is taken into consideration. The algorithm also ignores permissions which do not 6 Dave Read D : \ has permission on directory Figure 1: Object-based diagram for the permission entry of D : \, Dave, FileReadData differ from those of their parent directory. This helps to remove repetitive information and provides a useful heuristic for increasing performance and reducing the volume of captured data. The data structure returned from executing the algorithm is a list ordered by the directory structure. This list, E = {e1, e2, . . . , en}, contains an ordered list of tuples, en = {o, p, d}, where d is the directory resource object, o is the interacting object, and p is the set of permission level objects, p = {a1, a2, . . . , an}. The returned list (E) contains the object model for each files system entry. For example, consider a directory where a user Dave, o = Dave, has permission of FileReadData, p = FileReadData, on a locally mounted directory, d = D : \. A diagrammatic illustration of this model is provided in Figure 1. This object model is then stored in Comma Separated Value form for future processing. In the provided example, the following data set would be output: Dave, FileReadData, D : \ The example provided is simplistic and only represents one individual file system permission. It is often the case that an interacting object would accumulate access permission from multiple sources (explicit, inherited, etc.), and therefore Algorithm 1 is used to calculate each objects’ effective permission (i.e the permission the user actually receive) before the creation of the object model. The output object model (stored in text form) is then a complete representation of each interacting objects’ effective permission on the specified directory structure. 4. Association Rule Mining In this section, the area of Association Rule Mining (ARM) is discussed, and details of how it is used for detecting irregularities in file system permissions are provided. ARM is a method of discovering associations within data that frequently occur. For example, in the perspective of file system permissions analysis, determining that a user (e.g Dave) frequently has the same level of permissions (e.g Read) on a wide array of directories. In ARM, I is the set of binary items, and D are the data sets. In applying ARM to file system permissions, I is the total object set (users, permissions, directories) and D (previously defined as E) is the total set of effective permissions, D = {e1, e2, en}. Each permission entry en is a subset of I. For example, a potential item set could be I = {Dave, Mike, FileRead, FileWrite, D : \} and the following are example data sets: E1 = {Dave, FileRead, D : \} E2 = {Mike, FileRead, D : \} E3 = {Mike, FileWrite, D : \} Association rule mining is a way to discover interesting relationships in datasets, or in the case of this paper, infrequent relationships. The three following rules are used in association rule mining: Definition 1 (Association rule). X =⇒ Y is an Association Rule (AR), if X and Y are item sets, X is the LHS (Left Handside) or body of the rule, and Y is RHS (Right Handside) or head of rule. A set of 7 data entries, can be informally define an Association Rule (AR) as the rule of X =⇒ Y , where X, Y ⊂ I. In the continuing example, a file system specific association rule would be {FileRead} =⇒ {D : \}. Definition 2 (Support of an AR). The support (supp(X)) of an AR is defined as the percentage of per- mission entries that contain both X and Y . It can also be defined as the probability P ( X ∩ Y ) . In the example, the dataset {FileRead, D : \} has support of 0.6 since is occurs in one two thirds of the permission entries. Definition 3 (Confidence of an AR). The confidence of an AR is defined as the ratio between the num- ber of transactions that contain X∪Y and the number of transactions that contain X. It can also be defined as P ( X ∪ Y |X ) = P ( X ∪ Y ) /P ( X ) . In the example, the confidence (supp(X ⊂ Y )/supp(X)) would equal 0.6/0.6 = 1, meaning that for 100% of the permission entries that have FileRead are related to D : \ ARM procedures contain two stages: (i) the identification of frequent datasets, and (ii) generation of ARs. Piatetsky-Shapiro (1991) defines ARM as a method for the description, analysis and presentation of ARs which are discovered in databases using different measures to determine interesting data. ARM is concerned with the discovery, in tabular databases, of rules that satisfy defined threshold requirements. Of these requirements, the most fundamental is concerned with the support (frequency) of the item sets used to make up the ARs: a rule is applicable only if the relationship occurs sufficiently often in the data. Whether an item set is relevant or not, is determined by its support count confidence value. An item set is deemed to be relevant if its support count is above a user specified support threshold. Similarly, an AR is deemed relevant if the confidence value is above a user specified confidence threshold (Singer and Willett, 2003). Although minimum support and confidence thresholds help remove the generation of uninteresting rules, many of the remaining rules are still not interesting to the users. This work focuses on irregularities which in the context of Association Rule Mining can be assumed to be a non-frequent set of items. Therefore, in this work an interesting rule (I.e an irregular file system permission) is one with both low support and confidence. Frequent data set mining is thoroughly studied by many researchers, but important rare items are often not discovered by these algorithms. In many cases, the contradictions or exceptions also offers useful associations. In the recent past researchers started to focus on the discovery of such kind of associations called rare associations. Rare data sets can be obtained by setting low support; however, this generates huge number of rules. Therefore the key idea is to inverse the support threshold criterion by using a maximum support threshold but also using the confidence as a second criterion to deal with the large amounts of rare relations between items. In the technique presented in this paper, the selection of the minimum support threshold value is based on the frequency of each item in the data set. E.g. how many times an item appears which is turned into a percentage and represents the threshold value. As regards to the confidence threshold, it is selected empirically and usually its value is selected to be equal to the minimum support threshold. SOMA (Somaraki et al., 2010) is a trend-mining framework for knowledge discovery from large databases, the development of a validation framework for trend mining. SOMA was originally developed for the application of trend mining in medical data; however, SOMA itself is not domain dependent and can therefore be used with any type of data in different application areas. The framework exploits three steps: pre-processing, association rule mining, and trend mining; however, given the nature of the data and the aim of the investigation, both preprocessing and trend mining stages are not required. The preprocessing stage is superfluous in this application as the data is complete and discrete and there are no temporal requirements. In SOMA’s original application of medical data, logic was utilised to aid with merging and time-stamping data subsets for analysis. Trend mining is also not required. For this application the most important phase is association rule mining, where rules are extracted by considering a the of variables extracted from the data. In this paper, the Matrix (Yuan and Huang, 2005) algorithm which is part of SOMA framework was used for association rule mining. The algorithm is more efficient than the well-known Apriori (Tsai and Chen, 2004) algorithm as it only requires one pass over the data to generate the matrix. The algorithm is called a matrix algorithm as it creates a binary matrix with entries (0, 1) from passing over the database, resulting in the creation of a set of candidate items from which association rules are produced. In the work 8 Stage 1 (NTFS Processing) Stage 2 (SOMA) Input: NTFS Directory Effective Permissions Output: Infrequent Permissions Figure 2: Schematic illustrating the integration of both the NTFS permissions extraction software and SOMA for association rule mining. presented in this paper, the file system permission data is passed to the matrix algorithm in tabular form for processing and the output is a set of association rules, generated based on the support threshold and confidence value. The output is then processed to identify association rules for infrequent data. Full details of the matrix algorithm can be found in Yuan and Huang (2005). There is an absence of literature detailing the use of ARM to detect anomalies in file systems; however, closely related research on the identification of anomalies and irregular data sets using ARM strongly moti- vate its use (Chandola et al., 2009). For example, Li et al. (2009) present the successful use of ARM to detect anomalies in identifying irregular behaviour in local area network traffic. In addition, intrusion detection is another successful area of research which has seen applications of ARM (Patcha and Park, 2007). 5. Produced Software Software has been developed for the extraction, processing and association rule mining of file system permissions. For the extraction and accumulation of file system permissions (stage one), software was developed in C# using the .NET System.Management namespace. The developed application is a command line application which can be executed on any Windows NT operating system. The application will run under the user’s credentials, and therefore can only analyse directories on which the user has authority to read security permissions. The output from the tool contains the ACL information for each directory in comma value separated form which is subsequently used for association rule mining. In the second stage, the SOMA framework was developed in MATLAB 20014a. In the experiments presented in this paper, both stages were executed on a Intel i7 Processor with a clock speed of 2.2 GHz. Figure 2 provides an overview illustration of the integration of the two stages. This integration of these two stages allows the user to extract NTFS permissions and then identify irregularities through trend mining. 6. Empirical Observations This section contains two types of empirical observations through using the tools developed in Section 5. The first is empirical observations based on synthetically generated file system data, and the second is from analysing five real-world, multi-user directory structures of different configurations. The aim of this experiment is to determine the effectiveness of the proposed technique at identifying irregular and anomalous permissions within the directory structures. In both tests, to asses the performance, the following measures are considered: (1) True Positive Rate (tpr): the fraction of irregular permissions correctly identified as irregular; (2) False Positive Rate (fpr = 1 - tnr): the fraction of regular permissions incorrectly identified as irregular; (2) True Negative Rate (tnr): the fraction of regular permissions correctly identified as regular; (3) False Negative Rate (fnr = 1- tpr): the fraction of irregular permissions incorrectly classified as regular; Finally, the accuracy is reported as the fraction of all samples correctly identified. 9 Algorithm 2: Algorithm for generated synthetic directory structures Input: The maximum number of directories to be created, MaxDir Input: The step size of the increasing number of directories, StepDir Input: The maximum number of anomalies to be created, MaxAnom Input: The step size of increasing the number of anomalies, StepAnom Output: A set of directories, S = {s1, s2, ..., sn}, where sn = {d, p} where d is the directory resource, and p is the permission level 1 Algorithm algo() 2 for n ← 0 to MaxAnom do 3 i = 0 4 while i ≤ MaxDir do 5 S[j] ← createDirectory() 6 i+ = StepSize 7 end 8 j = 0, i = 0 9 while J ≤ MaxAnom do 10 dirNo = genRandomInt(0, i) 11 pLevel = genRandomInt(1, 14) 12 i+ = StepSize 13 S[j] ← pLevel 14 end 15 n+ = StepAnom 16 end 17 return S 6.1. Synthetic Data sets The motivation behind using synthetically generated test data is that although acquiring security in- formation from real-world directory structures is easily achievable, it is difficult to acquire ground-truth knowledge regarding the irregularities within those systems. This is because the identification of irregular permissions is ultimately down to the auditor’s knowledge, and although evaluating the system against such data is useful, it potentially limits the quality of assessment as both correct and incorrect permissions might have been incorrectly identified by the expert. In order to generate synthetic file system data, a technique has been created to automatically generate directory structures, assign permissions, and insert irregular permissions (anomalies). From the created directory structures, it is then possible to use the proposed NTFS processing software to extract access control information, and then use SOMA to identify irregularities. Algorithm 2 describes the process to generate synthetic directory structures. The first aspect of the algorithm is to create a the directory struc- tures by providing a maximum number of directories and a stepsize. For example, providing a maximum directory size (MaxDir) of 50 and a stepsize (StepDir) of 10 would result in the generation of five directory structures (10, 20, 30, 40, and 50). In the creation of directories, the security permission of their parent and system defaults will automatically be inherited. In addition, the same directory structure is created multiple times, from 1 until MaxAnom in the step size of StepAnom. For example, if MaxAnom = 20 and StepAnom = 5, the number of the same directory structure created would be four with 5, 10, 15, and 20 anomalies. Creating the anomalies for each directory structure is performed by identifying a random number between 0 and and the current directory size (i). This random number (dirNo) will be used to assign an anomaly. The permission of the anomaly is determined by generating a random number between 1 and 14 (Table 1), pLevel. The number identified represents the number of permission attribute set. For example, if the generated number is 2, the first two permission attributes are set. In the synthetic experiments presented in this paper, MaxDir = 1, 000, StepDir = 100, MaxAnom = 64, and for each iteration is StepAnom = StepAnom× 2. I.e the anomalies would increase in the sequence 10 User Group Directory Attribute Support Confidence Administrators 100\100 16 Delete 0.05 0.27 Anomalyuser 100\100 16\0 Delete 0.001 0.01 Table 2: Example Association Rules generated when analysing synthetically produced file systems that has 100 directory and 16 anomalies. of 1, 2, 4, 8, 16, 32, 64. Doubling the number of anomalies in each iteration will allow for the comprehensive understanding of the system’s ability to recognise anomalies as their frequency increases. The random approach of assigning permissions, as well as the permission level, helps to simulate a file system that is becoming increasingly complex and helps to replicate the process of ‘ad-hoc’ permissions allocation. This results in the production of 70 different directory structures to test the proposed technique. Figure 6 illustrates the average execution time for analysing each directory size. The difference in execution time for directory structures of the same size but with a different number of irregular permissions is less than 1 second. From the Figure 6 it is noticeable that execution time increases by around 50 seconds until the directory size reaches 800 where the time increases by around 100 seconds for the addition of 100 directories. It is also noticeable that the time required for the associative rule mining stage makes up a larger portion of the duration than the permission extraction stage. The results from all 10 directory structures, and the 7 different permission levels, are demonstrated in three different graphs illustrates the Receiver Operator Characteristics (ROC) space (Hanley and McNeil, 1982). Figure 3 illustrates the tpr and the fpr for the first four directory structures, with a directory size of 100, 200, 300, and 400, respectively. Figure 4 illustrates the results for the directory structures with 500, 600 and 700 directories. Finally Figure 5 illustrates the results for the directory sizes of 800, 900, and 1000. These results are interesting as they demonstrate both an improvement in tpr and a decrease in fpr as the directory size increases. This reason behind this is that the proportion of anomalies in the directory structure containing 100 directories is greater than that of the directory with 200 directories. For example, the directory structure with 100 directories containing 32 anomalies consists of a total of 1008 permissions. The directory structure containing 200 directories and 32 anomalies anomalies consists of 1936 total permissions. This is interesting as it demonstrates that the tools ability improves with an increasing number of directories and a decreasing number of irregular permissions. In the first instance this might appear like a negative characteristic of the system; however, as real-world directory structures will often be well in excess of 200 directories, with a low proportion of anomalies, it can be seen that the tools performance would be suitable for application. It is, however, worth raising the point that the accuracy would reduce in directory structures where permissions have been managed in an ad-hoc manner for a prolonged period of time. However, considering the fact that a large portion of permissions will always be system generated, the technique should still be able to identify anomalies, even if accuracy is reduced. The verbose results from all 70 experiments are omitted from the paper due to space constraints. How- ever, the results and data sets are available from the author upon request. In addition, from the graphs it is not possible to identify the number of irregular permissions each directory structure has; however, from analysing the results it can be stated that the directory structure with fewest regular permissions has the highest tpr an the lowest tnr across all directory sizes. Another observation that is consistent for all directory sizes is that the tpr increases and the tnr decreases incrementally with the increasing number of irregular permissions. Table 2 demonstrates two example Associate Rules for demonstrating both an irregular and regular per- mission extracted from a synthetically generated directory structure with 100 directories and 16 anomalies. The first rule is for a frequently occurring permission, and the second rule is for a irregular permission allocation. The first rule (X =⇒ Y ) demonstrates that the Administrators group implies Delete, X = {Administrators} =⇒ Y = {Delete}. This rule is frequently occurring because the percent- age (support) of all entries that contain both X and Y is at 5 %. This is high considering that there are in excess of five other users and groups within the system, and fourteen different attributes used in total. The confidence value is also quite high (27 %) which indicates that there is a strong relationship between both Administrators and Delete. The second rule demonstrates an irregular permission between 11 Directory Number Directories Permissions levels Permission Entries 1 36 3 248 2 108 3 1142 3 427 2 1708 4 453 4 3184 5 407 4 3719 Table 3: Test NTFS directory structures specifics. The numbers are only considering directories with permissions different from their parent. I.e. not inherited. the Anomalyuser and Delete permission. Here the support is less 0.1 % indicating that a low number of permissions contain both X and Y . In addition, the confidence value is also low at 1 % indicating that a low number of the total Delete permissions are related to Anomalyuser. From this example, it is easy to determine how the second permission has been identified as an anomaly. In the results presented in this section, the average accuracy rate for each dataset is incrementally improving from 0.85 to 0.95 for the directory structure with 100 and 1000 directories, respectively. This further stimulates the assumption that the system performs best when the number of anomalies represents 1% or less than the total number of permissions (100 irregular and 9136 regular). However, even when the number of irregular permissions is around 10%, the accuracy rate is around 85%. This is interesting as it demonstrates the suitability of the technique for identifying file system permission regularity, which increases as the percentage of irregular permissions decreases. This is contradictory to the main aim of this work which is trying to eliminate the requirement for expert knowledge when auditing file system permissions. This is because as the complexity of a file system increases through more ad-hoc permissions, the tool’s ability decreases and the requirement for expert knowledge increases. However, this tool does make progress on reducing the need for expert knowledge by making considerable progress towards removing the requirement for an exhaustive manual audit. 6.2. Real-world Data sets In this section the proposed system is validated against five real-world file systems to identify irregular permissions. As this test is not synthetic, the ground truth is acquired by expert knowledge. The limitation of acquiring expert knowledge alongside access to real-world directory structures limits the number of available directory structures for analysis. Table 3 shows details of the five previously unseen NTFS directory structures that are used to establish the developed techniques’ performance. The diversity of different organisations permissions varies widely; some have good access control which is maintained through adhering to a rigorous structure, whereas some have inconsistent access control resulting from ad-hoc fixes. The number of directories presented in Table 3 is significantly lower than the total number of directories in the file system. This is because the algorithm considers directories which have permission different from their parent. For example, directory number 1 in Table 3 actually contains 2856 directories of which only 55 are are unique. I.e these 55 permissions are inherited to all their sub-directories. For each test case, the ground truth (I.e the correct identification of irregular permissions) is acquired through reports produced by an independent security auditing organisation. Figure 7 illustrates the computation times for each of the directory structures presented in Table 3. From this graph, it is noticeable that the execution time for both stages of the algorithm increase as the size and complexity of the directory structure increase. However, the results demonstrate that the duration of the extraction and processing stage is more sensitive to an increase in both directory size and the number of assigned permissions, whereas stage two (association rule mining) is less affected by an increasing number of allocated permissions. For example, in the results it can be seen that the execution time of both stages increases greater than exponentially until directory structure 4. Interestingly there is a large difference between the number of permission entries for directory structures 3 and 4 (1708 to 3184) which results in a 12 Directory Number Irregular Regular True positive (tpr) False Positive (fpr) True Negative (tnr) False Negative (fnr) accuracy 1 6 248 5 (0.83) 1 (0.17) 203 (0.82) 45 (0.18) 0.82 2 62 1142 50 (0.81) 12 (0.19) 977 (0.86) 165 (0.14) 0.85 3 24 1708 20 (0.83) 4 (0.17) 1538 (0.90) 170 (0.10) 0.90 4 14 3184 13 (0.93) 1 (0.07) 3037 (0.95) 147 (0.04) 0.97 5 144 3719 143 (0.97) 4 (0.03) 3619 (0.97) 100 (0.03) 0.99 Table 4: Accuracy results from empirical analysis. See paragraph 1 in Section 6 for a description of the measures. large increase in computation time for stage one; however, the computation time for stage two only changes a small amount (less than one second). This demonstrates good scalability of SOMA (stage 2) for analysing file system permissions. Following analysis of the performance, it is important to evaluate the ability to correctly identify both irregular and regular permissions. This is performed by comparing the results of processing permissions with SOMA with the results of expert analysis. Figure 8 illustrates the Receiver Operator Characteristics (ROC) space (Hanley and McNeil, 1982) for all five directory structures. The ROC graph demonstrates that although accuracy is lower on the smaller directory structures (directory structure 1 and 2), it is improving as both the directory size and complexity increase. For example, for directory structure 1 has a tpr of 0.83, a fpr of 0.17, and an accuracy of 0.82. Directory structure 5 has a tpr of 0.97, a fpr of 0.03, and an accuracy of 0.99 (full results can be seen in Table 4). Although directory structure 1 demonstrates a reasonable performance where 80 % of classifications are correct, it would still require a reliance on expert interpretation to improve confidence. Directory structures two and three are slightly better where the accuracy rate has increased to 0.85 and 0.90, respectively. It is interesting to discover that directory structures four and five have improved and have accuracy levels of 0.97 and 0.99, respectively. This is a significant finding as it suggests that the accuracy of the proposed techniques are improving as both the directory size and complexity are increasing. This would suggest that the directory structures one, two and three are not big enough to allow the association mining techniques to correctly distinguish between what is normal and irregular. The empirical results presented above demonstrate the potential of the proposed technique to program- matically identify irregular permissions with a minimal reliance on expert knowledge. These results have demonstrated that the proposed technique is sensitive to directory size, and in general performs better on larger directory structures with diverse permissions. 7. Conclusion This paper presents a technique to identify irregular file system permissions using Association Rule Mining. This research focuses on the vulnerabilities of Microsoft’s New Technology File System (NTFS). Motivation for applying the presented technique to the NTFS is justified through a details discussion on the complexities associated with NTFS permissions administration. A detailed discussion is then provided where the structure of NTFS permissions is discussed and modelled in a suitable way to develop algorithms to autonomously identify irregular file system permissions. The developed technique is a two-stage algorithm. The first stage is a permission extraction algorithm for acquiring all relevant information for analysis. Empirical analysis has demonstrated that this stage is sensitive to both the size of the directory structure as well as the complexity of the assigned permissions. However, given that analysis is performed off-line and that permissions are often not changing quickly, the time-scale that is presented (a maximum of just over 2 minutes for the largest directory) is not detrimental. The second stage is to use association rule mining to identify permissions which are irregular and can potentially result in vulnerabilities. Using association rule mining results in the production of a final rule set containing those which have low support and confidence. This stage of the produced system scales well and has an accuracy rate which increases with both directory size and complexity. 13 Empirical observations are then performed which included executing the produced technique against 70 synthetically generated and five real-world, multi-user directory structures. The synthetically generated directory structures created the possibility to rigorously evaluate the system’s performance and understand the ability to identify irregular permissions which have been created in an ‘ad-hoc’ manner. The results from the synthetic directory structures demonstrated that the presented technique has an 90 % average accuracy. The main finding from these experiments was that the ability to correctly identify irregular permissions increases and the incorrect identification of regular permissions decreases as the directory size increases and the number of anomalies stays the same or decreases. It was identified that 95 % accuracy can be achieved when the percentage of irregular permissions is equal to or less than 1 % of the total permissions. The real-world directory structures were different in terms of directory size, the number of assigned permissions, and the number of known irregular permissions. The irregular permissions have been prede- termined through the use of expert knowledge. This expert knowledge was used as ground truth for the empirical analysis. The results are promising with an average accuracy rate of 91 %. It is interesting to discover that the technique has a better accuracy on larger and more complex directory structures. Although an accuracy rate in excess of 90 % has been achieved, it still does not completely remove the requirement for expert knowledge. However, it significantly reduces the reliance on expert knowledge as well as the time consuming process of having to perform an exhaustive manual security audit of a file system. The contribution from this paper to the anomaly detection and file system auditing communities is significant and strongly motivates further research. All the data sets and software presented in this paper are available form the corresponding author upon request. Future research is to further develop the techniques to include different mechanisms of file system access control as well as identifying suitable mechanisms of further improving accuracy. 8. References Barak, S., Modarres, M., 2015. Developing an approach to evaluate stocks by forecasting effective features with data mining methods. Expert Systems with Applications 42 (3), 1325 – 1339. Beznosov, K., Inglesant, P., Lobo, J., Reeder, R., Zurko, M. E., 2009. Usability meets access control: challenges and research opportunities. In: Proceedings of the 14th ACM symposium on Access control models and technologies. SACMAT ’09. ACM, New York, NY, USA, pp. 73–74. URL http://doi.acm.org/10.1145/1542207.1542220 Bhuyan, M., Bhattacharyya, D., Kalita, J., First 2014. Network anomaly detection: Methods, systems and tools. Communica- tions Surveys Tutorials, IEEE 16 (1), 303–336. Cao, X., Iverson, L., 2006. Intentional access management: making access control usable for end-users. In: Proceedings of the second symposium on Usable privacy and security. SOUPS ’06. ACM, New York, NY, USA, pp. 20–31. URL http://doi.acm.org/10.1145/1143120.1143124 Catania, C. A., Bromberg, F., Garino, C. G., 2012. An autonomous labeling approach to support vector machines algorithms for network traffic anomaly detection. Expert Systems with Applications 39 (2), 1822 – 1829. Chandola, V., Banerjee, A., Kumar, V., 2009. Anomaly detection: A survey. ACM computing surveys (CSUR) 41 (3), 15. Cheng, Q., Lu, X., Liu, Z., Huang, J., 2015. Mining research trends with anomaly detection models: the case of social computing research. Scientometrics 103 (2), 453–469. De Capitani di Vimercati, S., Paraboschi, S., Samarati, P., 2003. Access control: principles and solutions. Software: Practice and Experience 33 (5), 397–421. URL http://dx.doi.org/10.1002/spe.513 Hanley, J. A., McNeil, B. J., 1982. The meaning and use of the area under a receiver operating characteristic (roc) curve. Radiology 143 (1), 29–36. Hanner, K., Hörmanseder, R., 1999. Managing windows nt file system permissions— a security tool to master the complexity of microsoft windows nt file system permissions. Journal of Network and Computer Applications 22 (2), 119 – 131. Hu, H., Ahn, G.-J., Kulkarni, K., Nov 2013. Discovery and resolution of anomalies in web access control policies. Dependable and Secure Computing, IEEE Transactions on 10 (6), 341–354. Islam, M. S., Rahman, S. A., 2011. Anomaly intrusion detection system in wireless sensor networks: security threats and existing approaches. International Journal of Advanced Science and Technology 36 (1). Khan, M. N. A., 2012. Performance analysis of bayesian networks and neural networks in classification of file system activities. Computers & Security 31 (4), 391 – 401. Khatib, E. J., Barco, R., Gmez-Andrades, A., Muoz, P., Serrano, I., 2015. Data mining for fuzzy diagnosis systems in LTE networks. Expert Systems with Applications 42 (21), 7549 – 7559. Lazarevic, A., Ertöz, L., Kumar, V., Ozgur, A., Srivastava, J., 2003. A comparative study of anomaly detection schemes in network intrusion detection. In: SDM. SIAM, pp. 25–36. 14 Li, X., Zhang, Y., Li, X., Sept 2009. Local area network anomaly detection using association rules mining. In: Wireless Communications, Networking and Mobile Computing, 2009. WiCom ’09. 5th International Conference on. pp. 1–5. Ma, B. L. W. H. Y., 1998. Integrating classification and association rule mining. In: Proceedings of the fourth international conference on knowledge discovery and data mining. Mahoney, M. V., 2003. Network traffic anomaly detection based on packet bytes. In: Proceedings of the 2003 ACM Symposium on Applied Computing. SAC ’03. ACM, New York, NY, USA, pp. 346–350. Microsoft, 2006a. AccessEnum V1.32. URL http://technet.microsoft.com/en-us/sysinternals/bb897332 Microsoft, 2006b. How to use Xcalcs.vbs to modify NTFS permissions. URL http://support.microsoft.com/kb/825751 Naldurg, P., KR, R., 2011. Seal: a logic programming framework for specifying and verifying access control models. In: Proceedings of the 16th ACM symposium on Access control models and technologies. ACM, pp. 83–92. Naldurg, P., Schwoon, S., Rajamani, S., Lambert, J., 2006. Netra:: seeing through access control. In: Proceedings of the fourth ACM workshop on Formal methods in security. ACM, pp. 55–66. Nemeth, E., 2010. UNIX and Linux system administration handbook. Pearson Education. Parkinson, S., Crampton, A., 2013. A Novel Software Tool for Analysing NT File System Permissions. International Journal of Advanced Computer Science and Applications 4 (6), 266–272. Parkinson, S., Hardcastle, D., 2014. Automated planning for file system interaction. Proceedings of The 32nd Workshop of the UK Planning and Scheduling Special Interest Group (PlanSIG2014). Patcha, A., Park, J.-M., 2007. An overview of anomaly detection techniques: Existing solutions and latest technological trends. Computer Networks 51 (12), 3448 – 3470. URL http://www.sciencedirect.com/science/article/pii/S138912860700062X Pawlowski, B., Shepler, S., Beame, C., Callaghan, B., Eisler, M., Noveck, D., Robinson, D., Thurlow, R., 2000. The nfs version 4 protocol. In: Proceedings of the 2nd International System Administration and Networking Conference (SANE 2000). Vol. 2. p. 50. Piatetsky-Shapiro, G., 1991. Discovery, analysis and presentation of strong rules. Knowledge discovery in databases, 229–238. Russel, C., Crawford, S., Gerend, J., 2003. Microsoft windows server 2003 administrator’s companion. Microsoft Press. Singer, J. D., Willett, J. B., 2003. Applied longitudinal data analysis: Modeling change and event occurrence. Oxford university press. Solomon, D. A., 2005. Microsoft windows internals: Microsoft windows server 2003, windows xp, and windows 2000. Somaraki, V., Broadbent, D., Coenen, F., Harding, S., 2010. Finding temporal patterns in noisy longitudinal data: a study in diabetic retinopathy. Advances in Data Mining. Applications and Theoretical Aspects, 418–431. Somaraki, V., Harding, S., Broadbent, D., Coenen, F., 2011. Soma: A proposed framework for trend mining in large uk diabetic retinopathy temporal databases. Research and Development in Intelligent Systems XXVII, 285–290. Somaraki, V., Vallati, M., McCluskey, L., 2015. Discovering interesting trends in real medical data: A study in diabetic retinopathy. Proceedings of the 17th Portuguese Conference on Artificial Intelligence. EPIA 2015. Ten, C.-W., Hong, J., Liu, C.-C., Dec 2011. Anomaly detection for cybersecurity of the substations. Smart Grid, IEEE Transactions on 2 (4), 865–873. Thomas, O., 2010. Are NTFS and share permissions a bit too complicated. Windows IT Pro. Tsai, P. S., Chen, C.-M., 2004. Mining interesting association rules from customer databases and transaction databases. Information Systems 29 (8), 685–696. Viswanath, B., Bashir, M. A., Crovella, M., Guha, S., Gummadi, K. P., Krishnamurthy, B., Mislove, A., 2014. Towards detecting anomalous user behavior in online social networks. In: Proceedings of the 23rd USENIX Security Symposium (USENIX Security). Xie, M., Han, S., Tian, B., Parvin, S., 2011. Anomaly detection in wireless sensor networks: A survey. Journal of Network and Computer Applications 34 (4), 1302 – 1325, advanced Topics in Cloud Computing. Yu, L., Chen, H., Wang, S., Lai, K. K., 2009. Evolving least squares support vector machines for stock market trend mining. Evolutionary Computation, IEEE Transactions on 13 (1), 87–102. Yuan, Y., Huang, T., 2005. A matrix algorithm for mining association rules. In: Advances in Intelligent Computing. pp. 370–379. 15 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 False Positive Rate T ru e P o si ti v e R a te d = 100 d = 200 d = 300 d = 400 Figure 3: ROC space demonstrating the fpr and tpr of identifying anomalies in the synthetic directory structures where d is the number of directories. 0 0.1 0.2 0.3 0.4 0 0.2 0.4 0.6 0.8 1 False Positive Rate T ru e P o si ti v e R a te d = 500 d = 600 d = 700 Figure 4: ROC space demonstrating the fpr and tpr of identifying anomalies in the synthetic directory structures where d is the number of directories. 0 0.1 0.2 0.3 0.4 0 0.2 0.4 0.6 0.8 1 False Positive Rate T ru e P o si ti v e R a te d = 800 d = 900 d = 1000 Figure 5: ROC space demonstrating the fpr and tpr of identifying anomalies in the synthetic directory structures where d is the number of directories. 0 200 400 600 800 100 200 300 400 500 600 700 800 900 1000 Time in Seconds D a ta se ts NTFS-r SOMA Figure 6: Computation time in seconds for (1) extracting and processing NTFS permissions, followed by (2) the ex- ecution of the trend mining (SOMA) algorithm to identify irregular permissions. 16 0 50 100 150 200 1 2 3 4 5 Time in Seconds D ir ec to ry S tr u ct u re NTFS-r SOMA Figure 7: Computation time in seconds for (1) extracting and processing NTFS permissions, followed by (2) the execution of the trend mining (SOMA) algorithm to identify irregular permissions. 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 123 45 False Positive Rate T ru e P o si ti v e R a te Figure 8: ROC curve for the analysis of the test file systems 17 
work_ozfgaeogofdlroxd65simz73ay	----	Automatic classification of glycaemia measurements to enhance data interpretation in an expert system for gestational diabetes Automatic classification of glycaemia measurements to enhance data interpretation in an expert system for gestational diabetes Estefanía Caballero-Ruiz , Gema García-Sáez , Mercedes Rigla , Maria Villaplana , Belén Pons , M. Elena Hernando A B S T R A C T Expert systems for diabetes care need to automatically evaluate glycaemia measurements in relationship to meals to correctly determine patients' metabolic condition and generate recommendations about ther- apy adjustments. Most glucose meters allow patients to manually label each measurement with a meal tag, but as this utility is not always used, a completion procedure is needed. Classification methods are usually based on predefined mealtimes and present insufficient accuracy that might affect the automatic data analysis. Expert systems in diabetes require a reliable method to manage incomplete glycaemia data so that they can determine if patients' metabolic condition is altered due to a specific meal or due to an extended fasting period. This paper presents the design and application of a classification module to automatically assign the appropriate meal and 'moment of measurement' to incomplete glycaemia data. Different machine learn- ing techniques were studied in order to design the best classification algorithm in terms of accuracy. The selected classifier was implemented with a C4.5 decision tree with 7 input features selected with a wrapper evaluator and the genetic search algorithm, which achieved 95.45% of accuracy with the training set on cross-validation. The classification module was integrated in the Sinedie expert system for gesta- tional diabetes care and was evaluated in a clinical environment for 8 months with 42 patients. A total of 7,113 glycaemia measurements were uploaded by patients into the Sinedie system and were completed by the "classification module". The 98.79% of the measurements were correctly classified, while patients modified the automatic classification of 1.21% of them. Classification results were improved by 21.04% compared to a classification based on predefined mealtimes. The automatic classification of glycaemia measurements minimizes the patient's intervention, allows structuring measurements in relationship to meals and makes automatic data interpretation by expert systems more reliable. 1. Introduction As in other types of diabetes, the prevalence of Gestational diabetes mellitus (GDM) is increasing throughout the world (IDF, 2015). If the new International Association of Diabetes Study Group diagnosis criteria (IADPSG, 2010) -recently proven to be cost effective (Duran, Saenz, & Torrejon, 2014) - are adopted, the prevalence could be doubled. Several adverse outcomes are asso- ciated with hyperglycaemia in pregnancy, as foetal macrosomia, shoulder dystocia or caesarean section (Metzger, Lynn, & Lowe, 2008). Although most cases resolve with delivery, both mother and foetus are at a higher risk of developing type 2 diabetes in the future (Boney, 2005; Franks, Looker, & Kobes, 2006). Maternal glycemic control reduces adverse GDM outcomes (Harding, Dryden, & Guthrie, 2013) so patients are prescribed to self-monitor their blood glucose (BG) levels with a glucose meter around main meals. Although measurements are stored in the glu- cose meter memory file, patients usually note down their results in a paper logbook, structuring measurements in relationship to meals. They indicate the specific meal which the measurement is related to (breakfast, lunch or dinner) and whether it was made before (preprandial) or after the meal (postprandial). Clinicians evaluate the patients' measurements each week or every other week to determine the appropriate treatment, which consists of nutritional prescription, physical activity and, if necessary, insulin therapy. It has been observed that patients commit errors when manually reporting their BG levels, being the mean values signif- icantly higher than the logbooks' ones (Given, 0' Kane, Bunting, & Coates, 2013). Although this could mask a bad glycemic control and make clinicians establish a wrong therapy, they still prefer to examine logbooks instead of meter memory files (Polonsky, Jelsovsky, & Panzera, 2009). The reason might be that logbooks are easier to be reviewed, as they provide structured information that glucose meter memory lacks, like associations of measurements to meals, which are essential to make therapy adjustments. Logbooks also provide additional information such as food intakes, insulin doses or exercise. Telemedicine allows patients to send their BG data to the system to be remotely evaluated, which avoids unnecessary dis- placements (Carral, Ayala, del, & Fernández, 2015; Pérez-Ferré, Galindo, & Fernández, 2010) and improves access to specialized care in rural areas (Mohan & Pradeepa, 2014). Furthermore, by a more exhaustive and frequent evaluation of accurate data, telemedicine is capable of improving glycemic control (Wojcicki, Ladyzynski, & Krzymien, 2001) and reducing GDM adverse out- comes (Dalfra, Nicolucci, & Lapolla, 2009; Ferrara, Hedderson, & Ching, 2012). Monitoring data in telemedicine systems should be presented to clinicians organized as they appear in paper logbooks to facilitate their interpretation. The use of telemedicine could increase clinician's workload as it favors the generation of a greater amount of data to be evaluated by clinicians. Expert systems can solve the potential increment of clinicians' workload (Klonoff & True, 2009) by auto- matically analysing patients' monitoring data according to expert specifications (Hernando, Gómez, Corcoy, & del Pozo, 2000). The automatic analysis of monitoring data could optimize clinician's time by notifying which patients are evolving satisfactorily and which ones need a deeper examination. Expert systems, like clinicians, need to analyze glycaemia data in relation to meals to be able to determine patients' condition and to generate specific recommendations about therapy adjustments. Glycaemia data entry in expert systems can be performed by patients either manually or by uploading the data stored in their glucose meter (El-Gayar, Timsina, Nawar, & Eid, 2013a). The automation of data entry is preferred as it minimizes transcription errors (Given et al., 2013), results in more data captured, sim- plifies the date entry process and increases patients' satisfaction (El-Gayar, Timsina, Nawar, & Eid, 2013b). One of the problems that expert systems in diabetes have to face is the management of incomplete glycaemia measurements. A measurement is consid- ered incomplete if it lacks its association with a meal or with a moment of measurement. Newer glucose meters can include the functionality to allow registering these data manually, but even if they do, it is a time-consuming task and patients sometimes forget to introduce it. Without an accurate method to manage incomplete glycaemia data, expert systems cannot determine if patients' metabolic condition is altered due to a specific meal that should be adjusted or due to an extended fasting period. The majority of studies available in literature about expert sys- tems do not explicit describe the method used to retrieve the as- sociated meal and moment of measurement of glycaemia data or how they manage the lack of this information. Some expert sys- tems allow patients to add a meal tag to glycaemia measurements after uploading data from the glucose meter (Cafazzo, Casselman, & Hamming, 2012; Lim, Rang, & Shin, 2011; Quinn, Clough, & Mi- nor, 2008), but it has been observed that sometimes patients for- get to label some of the measurements (Mackillop et al., 2014) so they cannot be automatically analyzed. To solve this problem, the expert system can preselect a meal tag for each measurement downloaded and allow patients to modify it. Bromuri et al, pre- select glycaemia meal tags based on previous measurements, so, if the last measurement was taken before dinner, an after dinner pe- riod is preselected (Bromuri, Puricel, & Schumann, 2016). However, this method might present problems when dealing with repeated or missing measurements, for example if the patient forgets to measure her glycaemia after dinner, she measures it the following day before breakfast and she, by mistake, accepts the preselected meal tag. Some commercial applications automatically classify the glycaemia measurements downloaded by patients according to patient's predefined mealtimes (Sanofi Diabetes, 2015), but this method might present an elevated rate of errors as we will see in the following sections. We propose an innovative method for gly- caemia meal tag preselection using machine learning techniques. This paper presents the methodology to design an automatic classifier to associate the appropriate meal and moment of mea- surement to each glycaemia data downloaded from a glucose meter, its integration within the Sinedie expert system for GDM and the classification results obtained in a pilot study at Hospital de Sabadell with 47 patients for 8 months. 2. Material and methods This section describes the Sinedie expert system and how the automatic classifier is integrated with the BG levels uploading procedure. We explain the two different classification strategies studied to design the classifier: a simple algorithm based on the patient's mealtime schedules, measurements' time and BG level; and a more complex algorithm based on machine learning tech- niques. Finally, we describe the design of the clinical evaluation experiment. 2.1. Sinedie expert system for GDM Sinedie is a telemedicine platform enhanced by an expert sys- tem to manage the treatment of GDM patients. It aims to improve health care processes by reducing the evaluation time per patient, avoiding unnecessary displacements and improving the access to specialized healthcare. The expert system available in Sinedie computes the patients' metabolic condition and generates advice, to both patients and physicians about treatment changes, including the need to start an insulin therapy. The BG classifier presented in this paper was integrated in the Sinedie system as a classification module, whose functionality is to assign an appropriate mealtime and a moment of measurement to each incomplete measurement uploaded to the system with the glucose meter. The glucose meter memory file provides information about date and time of each measurement, as well as its value (see :ig. 1). Additionally, patients can enter the corresponding moment of measurement (preprandial or postprandial) or the associated mealtime if they have a glucose meter that allows registering such information. The classifier allows structuring the BG levels obtained from the glucose meter file to be visualized in an e-logbook in Sinedie (Fig. 1). This completion procedure is executed in a preprocessing step prior to the automatic data analysis performed by the expert system and it is essential to detect anomalous conditions in patient's health. After each data download, patients verify if the automatic classification is accurate and otherwise correct it. 2.2. Classification problem analysis A preliminary study (study 1) was carried out to determine which would be the optimal classifier to be implemented in Sinedie. As part of this study, we examined measurements' distri- bution along the day according to time and BG level to analyze the GLUCOSE METER MEMORY FILE E - LOGBOOK <IMP0RT> <ACSPIX Type="2106" SN="UI00346822" Ver= <DEVICE Name="Aviva" SN="36815015" Dt=' Tm="10:49" BGUnit="mg/dL"/> <RECENTREC D t = " 2 0 1 4 - l l - 2 4 " Tm="17:07"/> <BG0ATA> <BG Val= <BG Val= <BG Val= <BG Val= <BG Val= <BG Val= <BG Val= <BG Val= <BG Val= <BG Val= <BG Val= <BG Val= '118" D t = " 2 0 1 4 - l l - 2 4 " Tm="17:07' '126" D t = " 2 0 1 4 - l l - 2 4 " T m = " l l : 0 9 ' '99" D t = " 2 0 1 4 - l l - 2 4 " Tm="09:12" '116" D t = " 2 0 1 4 - l l - 2 3 " Tm="23:58' '135" D t = " 2 0 1 4 - l l - 2 3 " Tm="21:25' •132" D t = " 2 0 1 4 - l l - 2 3 " Trrt="17:12' '133" D t = " 2 0 1 4 - l l - 2 3 " T m = " l l : 2 9 ' ' 9 4 " D t = " 2 0 1 4 - l l - 2 3 " Tm="09:47" '116" D t = " 2 0 1 4 - l l - 2 2 " Tm="23:31' • 1 2 1 " D t = " 2 0 1 4 - l l - 2 2 " Tm="17:25' '144" D t = " 2 0 1 4 - l l - 2 2 " T m = " l l : 1 3 ' '144" D t = " 2 0 1 4 - l l - 2 2 " T m = " l l : l l ' <BG V a l = " 8 3 " D t = " 2 0 1 4 - l l - 2 2 " Tm="09:15" " 3 . 0 O . 0 2 " / > 2015-02-06" D = ' , l " / > D = " l " / > D = " l " / > D = " l " / > 0 « " l " / > D = " l " / > D = " l " / > D = " l ' 7 > D = " l " / > D = " l " / > D = " l " / > D = " l " / > D = " l " / > Classification • Date-r 24/11/14 23/11/14 22/11/14 21/11/14 20/11/14 19/11/14 18/11/14 17/11/14 16/11/14 15/11/14 14/11/14 13/11/14 12/11/14 Prep, breakf. 99 94 83 98 91 93 90 91 97 94 105 100 95 Post, breakf. 126 133 114 132 117 117 133 142 133 147 140 Prep, lunch Post, lunch 118 132 121 147 124 154 122 125 129 115 118 117 Prep, dinner 135 Post, dinner 116 116 145 130 168 129 134 113 149 146 132 130 Others 3 5 5 5 4 4 4 4 4 4 4 4 5 Fig. 1. Built patient e-logbook from the glucose meter memory file in the Sinedie system. The Sinedie e-logbook shows the 6 main intervals around main meals, when patients usually perform a BG measurement. Most of the patients also measure their glycaemia in other intervals, or perform several measurements in the same one. The column ("Others") shows the total number of measurements performed by the patient on a specific day, and shows all measurements' details when the cell is clicked. feasibility of using a simple algorithm based on patient's mealtime schedules, measurements' time and BG level. 2.2. J. Glycaemia data description The study 1 dataset (DSG) includes retrospective data from 25 GDM patients from the Hospital de Sabadell in Barcelona, Spain. Fifteen of the patients were treated only with diet and ten of them also with insulin therapy. Patients with insulin therapy adminis- tered either fast insulin associated to one or more meal intakes, and/or slow insulin at night. Patients used the Accu-check Aviva or the Accu-check Nano glucose meter (Roche Diagnostics, 2013a, b) over an average period of 60.9 ±33.5 days, and were requested to perform at least 4 measurements a day: before breakfast (breakfast preprandial) and after each main meal (breakfast postprandial, lunch postprandial and dinner postprandial). Patients also filled in a paper logbook with the four BG measurements requested. The DSG dataset includes a total of 6025 BG measurements from pa- tients' glucose meters (241.0 ± 134.8 per patient) and it is formed by two features obtained directly from the glucose meter memory file: • "time": Time of the day when the measurement was taken in absolute minutes. • "bg": Glycaemia value in mg/dl. BG measurements were labeled with the meal information registered manually in the patients' paper logbooks. The concor- dance detected between meter files and logbooks was 93.50%, considering concordance as the measurements correctly registered in the logbook out of those registered at all (Given et al., 2013). The "underreported" measurements, the ones appearing in the memory file but not in the logbook, were 5.05% of the stored measurements and were labeled by the authors including the following tags: lunch preprandial, dinner preprandial, repeated, morning, afternoon and night. We will refer to these measure- ments in the rest of the document as "additional measurements". BG values differed between the meter file and the logbook in 1.78% of the measurements registered; in that case the meter file value was maintained. We also found that 1.98% of the logbook measurements registered did not appear in the meter file. 0 Jp* ; 'É v" s 1- V Fig. 2. Glycaemia measurements distribution of patient P5 along the day. 2.2.2. Measurements scatter plots Twenty five scatter plots, one for each patient, were created in order to visually analyze the distribution of glycaemia mea- surements along the day. Each plot was generated according to measurement time and BG value. All scatter plots showed that measurements were clearly distributed in three groups corre- sponding to the three main meals: breakfast, lunch and dinner. It was also observed an overlap in breakfast preprandial and postprandial measurements time and, in some cases, BG values as well. An example of these assessments can be observed in Fig. 2, which shows the measurements' distribution of one patient treated only with diet. Fig. 2 shows the time overlap in P5 breakfast measurements (white and black circles) from 9:43 to 11:20. It also presents four additional measurements: one at 10:52 in the morning (+symbol) that falls in the breakfast interval and three in the afternoon (xsymbols), two of them coincide with the postprandial lunch interval and the last one is a preprandial dinner (black triangle symbol). These additional measurements were observed in 18 out of the 25 patients, ranging from 1 to 134 measurements per patient. At the left side of the chart we can see an example of a midnight dinner which should be grouped with the rest of the meals at dinnertime instead of with the breakfast ones. It would be logical to think that the postprandial BG levels would be higher than the postprandial ones, due to carbohydrates intake as we observe in P5 values. But, as shown in Section 2.2.3, we found that this is not always true. 2.2.3. Statistical analysis A further statistical analysis performed with the PSPP tool (PSPP, 2013) confirmed our visual appreciations. Four intervals Table 1 Measurement time variability for each patient in the four mealtime intervals defined, distinguishing between weekdays and weekends (w). Patient PI P2 P3 P4 P5 P6 P7 P8 P9 P10 Pll P12 P13 P14 P15 P16 P17 P18 P19 P20 P21 P22 P23 P24 P25 Mean Bpre IQR 277 39 7 37 60 59 188 29 78 67 86 42 19 33 80 104 85 33 19 43 38 24 60 81 63 66 ± 5 7 IQRw 143 49 144 37 60 57 84 50 67 57 92 58 29 38 80 92 61 81 74 28 75 67 34 73 68 68 ± 2 9 Bpos IQR 94 42 32 33 60 46 135 74 87 65 88 35 19 45 76 96 92 31 27 47 62 38 54 72 60 60 ± 2 8 IQRw 49 34 85 34 62 48 49 49 97 48 87 43 23 83 110 104 82 84 63 77 45 64 47 63 73 64 ± 2 3 Lpos IQR 46 30 13 33 32 49 65 46 43 59 79 59 39 34 49 116 40 11 27 64 47 59 48 47 43 47 ± 2 1 IQRw 45 69 39 10 91 114 73 55 26 106 62 116 47 98 66 49 48 34 13 31 40 66 58 42 63 58 ± 2 9 Dpos IQR 83 33 32 68 27 43 62 48 20 67 47 23 26 35 66 64 30 13 44 57 37 62 70 53 66 47 ±19 IQRw 53 44 41 107 41 96 114 45 43 69 37 68 33 114 46 47 42 62 47 69 60 73 39 49 36 59 ± 2 5 Interquartile range expressed in minutes. were established: preprandial breakfast (Bpre), postprandial break- fast (Bpos), postprandial lunch (Lpos) and postprandial dinner (Dpos) in order to study the variability in measurement time in each one of them. For each interval, we calculated the measure- ment mean time and interquartile range (IQR) for each patient, globally and differentiating between working days and weekends. We used the IQR instead of the standard deviation because it bet- ter shows the variability in the patients' usual measurement times, not considering the extreme values. Midnight measurements, from 0:00 to 3:20, were modified by adding 1440min (24h*60min) to the original value, in order to elude cyclical distances issues. Table 1 shows the results of the statistical study about measure- ment time variability of each patient over the four meal intervals defined. The observed average IQR exceeds 45 min in all the intervals. Focusing our attention on breakfast's intervals, the average IQR in both preprandial and postprandial subintervals is of 66 and 68 min respectively. This elevated IQR, over 60 min, was detected in 13 out of the 25 patients in "Bpos" interval and 12 out of 25 in "Bpre" interval. Furthermore, we detected a delay in the mean measurement time at weekends of 47 min at breakfast, 30 min at lunch and 16 min at dinner. These behavioral changes at weekends could be a criterion to take into account in our classification algorithm. However, in 3 patients no delay was observed. In order to study the glycemic preprandial and postprandial ranges at breakfast time, we calculated the minimum and maxi- mum BG level of the two breakfast intervals establishing for each patient its glycemic range: (Bpregmax - Bpregmin) and (Bposgmax - Bposgmin). We calculated the percentage of preprandial measure- ments whose BG value fell within the postprandial glycemic range and vice versa. The percentage of coincidence found for insulin treated patients (P3, P4, P6, P7, P8, Pll, P16, P17, P18 and P19) was of 47.92% (656 measurements out of 1367) and of 41.71% (599 measurements out of 1436) for non-insulin treated patients, reaching more than 50% in 3 patients using insulin and in 5 non- insulin treated patients. Fig. 3 shows the breakfast preprandial and postprandial measurements' boxplot where preprandial and postprandial BG levels overlap can be appreciated in the patients' breakfast measurements. The variability observed in the statistical analysis indicates that a classification algorithm based only on the time of measurement according to patients' mealtime schedules could present problems when distinguishing between preprandial and postprandial mea- surements in the same mealtime interval, as well as additional measurements. In GDM this is especially problematic at breakfast interval, when it is necessary to discriminate between fasting and postprandial measurements. Neither is enough to add the BG level information to the algo- rithm, since we observed over a 40% of coincidence in glycaemia ranges at breakfast time regardless insulin administration, which makes it really complex to discern whether the measurement is preprandial or postprandial. For these reasons, we consider that a machine learning algorithm is needed to perform a rigorous and more accurate classification of measurements to correctly complete the uploaded measurements with the associated meal and moment of measurement. 2.3. Machine learning for classification We evaluated different machine learning techniques applied to automatic classification. In order to evaluate whether the consideration of more parameters could improve classification accuracy, we obtained a larger set of features (DS18) from the DSG. We tried 6 different feature selection (FS) methods over the DS18 to select the most relevant features and to remove the redundant ones. Three datasets were obtained using FS (DSCFS, DSWB and DSWG) in addition to the initial DSG, the DS18, and a feature subset defined by an expert (DSE). We evaluated two learning algorithms in terms of accuracy, the 1 8 0 - 1 6 0 - 1 4 0 - 1 2 0 - 1 0 0 - class = break-prep class = break-post Fig. 3 . Boxplot of breakfast preprandial and postprandial measurements of all patients. Blood glucose level in mg/dl. C4.5 decision tree (Quinlan & J. Ross, 1993) and the Multilayer perceptron neural network (Haykin, 1994) with the 6 different subsets. We used the Weka data mining tool (Witten, I. H. & Frank, E, 2005) due to its Java library that provides all the algorithms implementation we needed. It was also chosen for its ease of inte- gration in the Sinedie system, as it is also developed in Java. Weka default parameters were maintained in all the algorithms tested and a 10-fold cross-validation was applied for FS and learning algorithms evaluation. 2.3.1. Search of input and output features We obtained a total of 18 input features from the DSG and from the patients' clinical history. In a previous study performed for pa- tients with type 1 diabetes (García-Sáez, Alonso, & Molero, 2009), an expert specified that for measurement's meal assignation, besides time and BG level, it was relevant to know when the pre- vious measurement was taken and whether the patient was under insulin therapy or not. Following these recommendations, we ob- tained the DSE subset with 'time', 'bg' and two additional features: • "insulin": Boolean that indicates whether the patient has in- sulin therapy or not. • "iprev": Time difference in minutes from previous measurement of the same day if any. As the expert considered relevant the insulin treatment and time interval between measurements, we added 6 more features related to those parameters: • "instype": Indicates the type of insulin administered and the moment of administration if any. • "fib": Fast insulin dose at breakfast time. • "fil": Fast insulin dose at lunchtime. • "fid": Fast insulin dose at dinnertime. • "sin": Slow insulin dose at night time. • "ipos": Time difference in minutes until posterior BG measure- ment if any. The consideration of time intervals between measurements led to the idea of evaluating groups of three measurements instead of evaluating isolated ones. So, two more features were calculated: • "bgprev": Previous BG value if any. • "bgpost": Posterior BG value if any. To study if weekends, holidays or even the seasons of the year affect the patterns of the patients' measurements, and to identify which rules could be established related to this, we calculated the following 5 features from the date of measurement: • "work": Boolean value that indicates if it is a working day or not. • "dow": Day of the week. • "doy": Day of the year. • "day": Day of the month. • "month": Month of the year. As we verified in patients' scatter plots, measurements could be grouped in three clusters related to the three main meals, so that a last feature denominated "intake" was obtained. This feature indicates which cluster the measurement belongs to, and can take three different values: Breakfast, lunch or dinner. The feature is calculated clustering each patient's measurements according to its time with the Expectation Maximization algorithm (EM) (Dempster, Laird, & Rubin, 1977), specifying the number of clus- ters. Measurements after midnight were corrected, as explained in Section 2.2.3, to be grouped in the dinner cluster instead of the breakfast one. We also performed the clustering using the K-means (MacQueen, 1967) and X-Means (Pelleg & Moore, 2000) algorithms, but we obtained worse results than with the EM algorithm. Regarding the classifier output features, ten classes were defined: "breakfast-prep", "breakfast-post", "lunch-prep", "lunch- post", "dinner-prep", "dinner-post", corresponding to preprandial and postprandial main meals measurements, and "morning", "af- ternoon", "night" and "repeated" corresponding to the additional measurements that some patients make. 2.3.2. Feature selection In order to discover the optimal feature subset that achieves the best accuracy with the minimum complexity we tested 6 different FS methods by the combination of 2 evaluators and 2 searching algorithms. There are 2 main approaches in FS evaluators, wrappers and filters, so we tested one evaluator of each approach. The wrapper evaluates feature sets by using a learning scheme (Kohavi, 1996) and it is computationally intensive. Among the filter models we chose the Correlation based feature selection (CFS), that evaluates the worth of a subset of attributes by considering the individ- ual predictive ability of each feature along with the degree of redundancy between them (Hall, 1999). As not all searching algorithms perform equally before differ- ent datasets, we tested two algorithms belonging to two of the three main categories available: "Sequential search" and "Random search" (Liu & Yu, 2005). We excluded the "Complete search" category as it might be computationally infeasible when working with a large number of features. From the first category we tested the Best First algorithm (Witten & Frank, 2005), which searches the space of features subsets by greedy hill climbing augmented with a backtracking facility. It is simple and fast. We tried the three searching strategies: forward, backward and bidirectional. The other algorithm tested was the Genetic search (Goldberg, D. E., 1989), which belongs to the second category that can avoid local minimal problems. These 6 FS methods were applied to the DS18 obtaining 6 features subsets. To determine which features should form the resulting subsets from FS, we initially included the features selected the greatest number of times. Then, we followed a sequential forward strategy adding the rest of the features one by one, selecting the most popular each time and stopping when accuracy decreased. 2.3.3. Learning algorithms The algorithms evaluated to build the classifier are the C4.5 decision tree and a neural network with the multilayer perceptron (MLP) architecture. Both algorithms have been widely and suc- cessfully used in diabetes clinical domain with multiple purposes like retinal disease diagnosis (Bourouis, Feham, Hossain, & Zhang, 2014), impaired glucose metabolism or type 2 diabetes mellitus identification (Hische, Luis-Dominguez, & Pfeiffer, 2010; Mohlig, Floter, & Spranger, 2006; Shankaracharya, Mallick, & Shukla, 2012; Upadhyaya, Farahmand, & Baker-Demaray, 2013), or to identify significant factors influencing diabetes control (Huang, McCullagh, Black, & Harper, 2007). However, they have not been used for BG level classification in mealtime intervals. The C4.5 algorithm implementation used is the ¡48 and for the MLP the Multilayer Perceptron, both defined in Weka. 2.4. Clinical evaluation The classifier was integrated in the Sinedie system as part of a classification module, which is also responsible for acquiring the input features required by the classifier. The classifier implemen- tation is based on the results obtained in study 1. We selected the classification algorithm and the set of input features that achieved the highest accuracy in the second part of the study {2.3. Machine learning for classification). The implemented classifier was trained with the same data used in study 1. The clinical evaluation was performed at a Spanish hospital, the "Hospital de Sabadell" for 8 months. During this period, 42 patients with GDM have uploaded a total of 7113 measurements to the Sinedie system, which have been automatically classified by the classification module. In order to evaluate the classification results that would have been obtained using an algorithm based in predefined mealtimes we implemented a second classifier based on usual Spanish mealtimes. At the end of the clinical study, we categorized the measurements that patients had uploaded into the system with Table 2 Features selected by each method in 10-fold cross validation. In bold and underlined the features selected by us to be included in the features subsets. Feature day month doy dow work time iprev ipos bgprev bg bgpos insulin instype flb lil fld sin intake Wrapper Best First J48 1 1 1 2 1 9 10 10 1 9 0 1 1 1 3 1 8 9 MLP 0 1 0 2 0 10 9 10 0 10 10 6 0 4 2 4 0 10 Genetic J48 2 1 0 0 5 10 10 10 4 9 0 4 4 4 4 2 9 10 MLP 0 2 1 0 4 10 2 10 7 10 10 10 1 4 2 8 6 10 CFS 0 0 0 0 0 10 10 10 50 10 10 0 0 0 0 0 10 10 the classifier based on fixed time intervals and compared the re- sults obtained with the labels assigned by the classifier integrated in Sinedie, which were verified by the participants in the clinical study. 3. Results In this section we present the results obtained in the machine learning techniques evaluation, considering FS and learning al- gorithms tests, the classification module structure specification and its performance in the Sinedie expert system where it was integrated. 3.1. Machine learning 3.Í.Í. Feature selection Table 2 shows the number of times (from 0 to 10) that each feature was selected by each FS method during the 10-fold cross validation. The table is divided into three main columns (features name and the two evaluators tested, Wrapper and CFS). The col- umn -corresponding to the Wrapper evaluator- is divided into two sub-columns, one for each searching algorithm used: Best First and Genetic Search. These two columns are subdivided into two sub-columns according to the learning algorithms tested (J48 and MLP). The CFS algorithm, as it evaluates attributes correlation, does not depend on any learning algorithm. CFS obtained the same results with the two searching algorithms so its results are shown in a single column, the last one. In the Best First algorithm, the result obtained with the "Forward" search strategy is shown as it had similar results than the "Bidirectional" strategy and slightly better results than the "Backwards" one in MLP. All FS methods selected 7 features apart from the wrapper- MLP-genetic combination, which selects 9. There are 4 features selected by all the methods in 9 or 10 folds: "time", "ipos", "bg" and "intake". There are two other features that were selected by all the methods except one: the "iprev" feature (not selected by the wrapper-MLP-genetic) and the slow insulin dose "sin" (not selected by the wrapper-MLP-Best First). The MLP combinations selected the boolean feature "insulin" in more folds than the "sin", on the contrary of J48. The posterior BG measurement "bgpos" is selected by MLP in all the folds and CFS but it is never chosen by the J48. The "work" feature seems not to have a high impact as it has been selected only in three methods in half of the 10 folds, but as it is shown in Table 3, it belongs to the combination that achieves the best accuracy. 3.1.2. Learning algorithms performance Table 3 shows the accuracy results obtained by each candidate dataset obtained in the FS process, as well as with the DSG, the DS18 and the DSE with both learning algorithms (J48 and MLP). The results show that the J48 learning algorithm achieves higher accuracy than MLP with all the datasets tested. Accuracy improves when increasing the attributes number from 2 (DSG) to 18 (DS18), and reaches the highest values when removing the redundant ones by feature selection (DSCFS. D S W * ) . We achieved a 4% enhancement with the DSE, adding two features to the initial DSG. When adding the rest of features calculated, a total of 18 forming the DS18, we obtained a slight improvement with the J48 algorithm (1.42%), and of 4.3% in MLP algorithm. This might be explained because the J48 achieved a 3.6% more accuracy than the MLP with the initial DSG, remaining fewer margin for improvement. It is observed that with few features, 2-4, the J48 algorithm performs substantially better than the MLP but when we increase the features number up to 18 the performance is similar, over 94% with just a 0.67% of enhancement over the MLP. The MLP cross-validation execution time shows how the algorithm complexity also increases with the number of features. The J48 achieved the best accuracy result, 95.45% and it is also faster to be trained than the MLP ( I s vs. 2 min comparing the best accuracy results of each algorithm) so we decided to build the classification module based on the J48 learning algorithm discarding the MLP. The selected feature subset was the one with highest accuracy, the Wrapper evaluator and the Genetic search algorithm (DSWG) using the J48 learning algorithm. 3.2. Classification module structure We implemented the automatic BG measurements classifier as a C4.5 decision tree, with the J48 algorithm. The features chosen as inputs were the 7 selected by the FS method formed by the wrapper evaluator and the genetic search algorithm: "time", "ipre", "ipos", "bg", "intake", "work" and "sin". The classification procedure is divided in two stages: 1) ac- quisition of input features and; 2) classification of the glycaemia measurements. In the first stage ("Acquisition of inputs features" ), the set of measurements uploaded by the patient (Mupid), formed by 3 input features (date, time and BG level), is transformed to obtain a measurement set composed by 7 inputs features (M7), which are required by the classifier in the second stage ("Classi- fication" ). The 7 features are obtained from: 1) Mup|d; 2) Mprev: dataset with the glycaemia measurements from the previous download if they exist; and 3) S: dataset with the slow insulin doses related to the measurements uploaded, in case the patient uses insulin treatment. Mprev and Mupid with midnight time rec- tification ("cotime" ) are the inputs to the cluster that obtains the dataset (I) with the "intake" features related to each measurement. Clusters are recalculated in each measurement's upload so the "in- take" feature adapts to possible changes in patients' schedules due to pregnancy progress or to other reasons. When patients upload their glycaemia data for the first time, if the number of measure- ments is lower than three, the "intake" attribute is obtained with a simple logic rule based on fixed mealtime intervals to solve errors detected in the clustering results with a small number of measurements. The "Classification" stage involves the classification of the M7 dataset, obtaining as a result the measurements' relationship to main meals. The classifier output (C) is a set of pairs of values GLUCOSE METER DOWNLOAD PATIENT / n VERIFICATION n / M p r e v ( t i m e ) Mupld( date, tlme.bg) Mclsf (date,time,bg, meal, moment) C L A S S I F I C A T I O N M O D U L E % ACQUISITION OF INPUT FEATURES j FEATURE CALCULATION ( w o r k , i p r e v , i p o s c o t i m e ) Mupld(cotime)' M p r e v ( c o t i m e l 'n+m CLASSIFICATION n : M 7 ( d a t e , w o r k , t i m e , : : Iprev.lpos.bg, M e x t ( d a t e , w o r k , j : intake.sl) t i m e , ¡ p r e , i p o s . b g ) : I CLUSTER (EM) n FEATURES JOIN n : : l(intake) , ' n RELATIONSHIP T O MEAL ASSIGNATION T * n / ( m e a l , moment) CLASSIFIER 0 4 8 ) S ( s i n ) DATA ANALYSIS M r e v ( d a t e , t l m e , b g , meal,moment) Fig. 4. Classincation module structure diagram. The inputs and outputs of submodules are datasets whose features are indicated in brackets. Table 3 Accuracy results of both learning algorithms with the 7 datasets tested with 10 fold cross-validation. Set DSG DSE DS,8 DSCFS DSWB DSWG J48 Accuracy 89.20% 93.44% 94.82% 95.19% 95.20% 95.45% F-Measure 0.876 0.925 0.947 0.951 0.951 0.954 Execution Time Is Is Is Is Is Is MLP Accuracy 85.61% 89.86% 94.15% 94.58% 95.10% 95.00% F-Measure 0.847 0.895 0.938 0.944 0.949 0.947 Execution Time l m l 3 s 1 m23s 8 m 2 2 s l m 5 9 s 2 m 4 s 2 m 2 6 s DSCFS, DSWB y DSWG correspond to features sets obtained by the CFS; and the wrapper evaluator with the Best First (WB) (Forwards) and the Genetic search algorithms (WG) respectively. (meal-moment) related to each measurement uploaded. The four outputs related to the additional measurements "morning", "af- ternoon", "night" and "repeated" have moment "unknown". Only incomplete measurements are automatically classified. Whenever a patient introduces either the meal or the moment in her glucose meter, this data is maintained and the missing value is completed with the automatic classifier output. Classified measurements (Mcisf) are presented to the patient through the Sinedie graphical user interface for verification, offering the possibility of reclassify- ing them (Mrev) before the automatic data analysis is performed by the expert system. The classification module structure is shown in Fig. 4. 3.3. Clinical evaluation The classification module integrated in the Sinedie expert sys- tem achieved a 98.79% of accuracy when being used by patients. Only 1.21% (86) of the measurements was reclassified by patients, out of the 7113 measurements transmitted by them and automati- cally classified by the system. Among the reclassified ones, 53.49% (46) were automatically classified as "repeated", 15.12% (13) as "preprandial-breakfast", 12.79% (11) as "postprandial lunch", 10.47% (9) as "postprandial breakfast", 3.49% (3) as "preprandial dinner", 3.49% (3) as "postprandial dinner" and 1.16%(1) as "preprandial lunch". Table 4 shows the reclassification changes made by the patients. Most classification errors (46 out of 86) are made in the identi- fication of repeated measurements. However, in half of these cases (23 out of 46) patients only reclassify the moment of measurement from "unknown" to "postprandial" or "preprandial" maintaining the meal determined by the classifier ("Repeated"). We consider the automatic classification of these 23 measurements correct, as the classifier does not specify the moment of measurement in repeated measurements. http://tlme.bg http://ipos.bg Table 4 Classifier confusion matrix. Aclsf Bpre Bpos Morn. Lpre Lpos Aftrn Dpre Dpos Night Runk / Prclsf* Bpre 1743 1 0 0 0 0 0 0 0 3 Bpos 5 1724 0 0 2 0 0 0 0 10 Morn. 0 1 0 0 4 0 0 0 0 0 Lpre 0 0 0 1 2 0 0 0 0 0 Lpos 0 2 0 1 1718 0 1 1 0 4 Aftrn 0 0 0 0 0 0 2 0 0 0 Dpre 0 0 0 0 0 0 0 1 0 0 Dpos 6 1 0 0 0 0 0 1706 0 6 Night 0 0 0 0 0 0 0 1 0 0 Runk 0 0 0 0 0 0 0 0 0 107 Rpre 2 0 0 0 0 0 0 0 0 1 Rpos 0 4 0 0 3 0 0 0 0 22 The rows show the tags assigned by the automatic classifier and the columns the tags assigned by the patients. Bpre: Break- fast preprandial; Bpos: Breakfast postprandial; Morn: Morning; Lpre: Lunch preprandial; Lpos: Lunch postprandial; Aftrn: Afternoon; Dpre: Dinner preprandial; Dpos: Dinner postprandial; Runk: Repeated unknown; Rpre: Repeated preprandial; and Rpos: Repeated posprandial. a Aclsf: Automatic classification b Prclsf: Patient reclassification. Other noteworthy errors are the 6 measurements reclassified as dinner postprandial. These measurements belong to the same patient and were taken between 2a.m and 5a.m. The accuracy achieved by the classifier based in Spanish pre- defined mealtimes, was 78.00%, a 21.04% lower than the classifier based in machine learning techniques. 4. Discussion Machine learning techniques have confirmed our initial hypoth- esis about the need of using more variables than the 3 initially proposed (mealtime schedules, BG level and measurement time) to obtain better accuracy results that optimize the automatic classi- fication process. Increasing the number of input features improved classification accuracy in both learning algorithms tested, but it is by the application of feature selection methods how we achieved the highest performance. The common 6 features selected by all the FS methods are: measurement time (time), time interval be- tween measurements (iprev, ipost), BG level, insulin treatment (sin or insulin) and mealtime interval (intake). The selection of time intervals between the measurements confirms the importance of evaluating groups of three measurements instead of isolated ones. The best result in MLP selects the boolean attribute "insulin" instead of the integer "sin" selected in CFS or in C4.5 best results. The slow insulin is administered to lower fasting glycaemia, so its selection could help to differentiate between preprandial and postprandial measurements at breakfast time interval. Apart from the 6 common features mentioned before, while the CFS selects the "bgpos" as relevant, the Wrapper for the C4.5 does not. In- stead, the combination that obtained the best result, wrapper-C4.5 genetic, selects the "work" attribute, which gives information about whether the measurement was taken during the weekend or not. A delay in the average measurement time during weekends was observed in all the patients but three, so this also validates the idea that this information should be considered. The classification module presented in this paper was suc- cessfully integrated in the Sinedie expert system, where it was validated in a clinical environment during 8 months at one hos- pital. In this period, 42 patients sent a total of 7113 measurements which were completed by the automatic classifier developed. Only 1.21% of the measurements were reclassified by patients, so we consider that 98.79% were correctly classified by the module. The classifier presented has a good generalization ability, as the accuracy achieved in the clinical evaluation with new glycaemia data is 3% better than the one achieved in study 1 with the cross-validation of the training dataset. Even if all patients had a glucose meter with the functionality to manually add the meal information to each measurement, the automatic classifier integrated in Sinedie would still be an essential part of the expert system to manage the incomplete measurements that patients may forget to label. In the clinical study, the Sinedie expert system was able to detect all situations that required a therapy adjustment due to altered glycaemia val- ues, and all recommendations generated about diet modifications affected the appropriate meal intake. The published studies on expert systems for diabetes that de- scribe the method used for glycaemia measurements' classification do not report the accuracy achieved with the method used. For that reason, it is not possible to compare our results with the results obtained by other classification methods. Considering our results, the classifier defined in this article presents the strength of obtaining higher accuracy than the algorithms based on patients' mealtimes' schedules or on previous measurements. A greater accuracy in glycaemia classification results in lower patient in- tervention when registering monitoring data into expert systems, which speeds up the process and makes it less tedious. The po- tential limitation of the tested classifier would be determined by the training set used and the generalization ability in the presence of a different population than the one treated at the Hospital de Sabadell. In case of reduced accuracy with different populations the classifier could be trained with a different training set. Possible future research lines include: 1) evaluation of the generalization ability of the classifier; 2) enhancement of the identification of repeated measurements and those belonging to optional moments of measurement (preprandial, morning, after- noon and night); 3) comparison of the results obtained with the classifier proposed and the ones that would be obtained with mealtime schedules specified by patients; 4) extension of the clas- sifier functionality to include other types of diabetes like type 1 and type 2; 5) Integration of the classifier in a mobile application that connects with the Sinedie expert system. To evaluate the generalization ability of the classifier it would be necessary to test it with patients from different regions from Spain or different countries. In case the classifier does not gener- alize correctly, it would require to be trained with a different data set from patients treated at the hospital where the classifier will be used. Training the classifier with a more homogeneous dataset could improve the identification of repeated and optional measurements. In the dataset from study 1, which was used to train the GDM classifier, the measurements labeled as "repeated" only represent the 3.5%, of the whole data set. The training of the classifier can be performed in parallel without requiring a system interruption. Then, it would be necessary to replace the file where the trained classifier is stored with the new retrained one, and reboot the system for the changes to take effect. In any case, to minimize the reclassification of this type of measurements is complicated, as the criteria about which measurement, either the first or the last one, should be considered as "Repeated", is ambiguous. We also observed that, sometimes, when facing several repeated measurements, some patients tend to label the ones with higher BG values as "Repeated". In order to confirm the theory proposed in this article, about the accuracy enhancement that machine learning algorithms present in comparison to the accuracy achieved by algorithms based on patient's mealtime schedules, it would be interesting to compare the performance of both types of algorithms in the Sinedie system. Patients could specify their mealtime's schedules when they are registered in the system, with the possibility of modifying them in the system whenever they want. The evalua- tion could be done comparing the accuracy results obtained with the machine learning classifier (already integrated in Sinedie), and those that would be obtained using the mealtime schedules specified by each patient. A similar methodology as the one presented in this paper could be followed to build a classifier for other types of diabetes. As glycaemia ranges and recommended moments of measurement are different depending on the type of diabetes, it would be necessary to implement a new classifier, in order to classify other types of diabetes measurements, performing a new feature selection with retrospective data of patients with the corresponding type of diabetes and create a new training dataset. Based on the learning curve of the presented GDM classifier, we estimate that the size of the dataset needed to perform a new study for a different type of diabetes should be over 5000 instances. However, from 2500 instances our classifier accuracy is stabilized at around 95% with the training dataset from study 1, achieving the maximum accuracy value with around 5000 instances. The last future research direction identified is the implementa- tion and integration of the classifier presented in the article into a mobile phone application. Mobile phones are smaller than laptops and therefore easier to carry when patients are out of home. Patients would download their monitoring data to their mobile phones and then the application would transmit the classified data to the Sinedie expert system for evaluation. Finally, patients would be able to verify the results of the data classification in the same mobile application. 5. Conclusions Machine learning tools have proved to be effective to design an automatic classifier for glycaemia measurements, required to pro- vide automatic recommendations based on an expert system for gestational diabetes. All methods tested obtained accuracy results above 80%, but the highest performance was achieved with the C4.5 decision tree learning algorithm with 7 inputs selected by a wrapper evaluator and the Genetic search algorithm. The selected classifier not only performed well with the training dataset with cross-validation, achieving a 95.45% of accuracy, but also in the Sinedie expert system, when being used to classify new glycaemia measurements uploaded by patients in clinical practice (98.79%). The scarce methods available in literature for glycaemia classi- fication based on previous measurements or predefined mealtimes present a high rate of errors that prevent expert systems from correctly determining patients' condition or generating recommen- dations about therapy adjustments. If the expert system detects altered glycaemia values but it is not able to determine whether they are caused by the breakfast or the lunch, it will not be able to recommend the reduction of carbohydrates in the appropriate intake or the administration of insulin to metabolize them. Automatic classification is essential in Sinedie expert system to analyze incomplete glycaemia measurements registered, to min- imize patients' intervention, to automatically detect the patients' metabolic condition and to generate recommendations about therapy changes in specific mealtimes. In addition, the information added by the classification module in the Sinedie expert system allows structuring measurements in relationship to meals so that the medical team can visualize and evaluate them as fast as they usually do with paper logbooks. Acknowledgment This work has been funded by the Spanish grant Sinedie (PIIO/01125), co-funded by FEDER. We would like to thank the patients of Hospital de Sabadell for their collaboration and support in this research. References Boney, C. M., Verma, A., Tucker, R., & Vohr, B. R. (2005). Metabolic syndrome in childhood: association with birth weight, maternal obesity, and gestational dia- betes mellitus. Pediatrics, 225(3), e290-e296. doi:10.1542/peds.2004-1808. Bourouis, A., Feham, M., Hossain, M. A., & Zhang, L (2014). An intelligent mobile based decision support system for retinal disease diagnosis. Decision Support Systems, 59, 341-350. doi:10.1016/j.dss.2014.01.005. Bromuri, S., Puricel, S., Schumann, R., Krampf, J., Ruiz, J., & Shumacher, M. (2016). An expert Personal Health System to monitor patients affected by Gestational Diabetes Mellitus: A feasibility study, journal of Ambient Intelligence and Smart Environments, 8(2), 219-237. Cafazzo, J. A., Casselman, M., Hamming, N., Katzman, D. K., & Palmert, M. R. (2012). Design of an mHealth app for the self-management of adolescent type 1 dia- betes: A pilot study, journal of Medical Internet Research, 24(3), e70. doi:10.2196/ jmir.2058. Carral, E, Ayala, M., del, C., Fernández, J. J., González, C., Pinero, A., & Gar- cia, G. (2015). Web-based telemedicine system is useful for monitoring glucose control in pregnant women with diabetes. Diabetes Technology & Therapeutics, 27(5), 349-354. doi:10.1089/dia. 2014.0223. Dalfra, M. G., Nicolucci, A., & Lapolla, A.on behalf of the TISG. (2009). The effect of telemedicine on outcome and quality of life in pregnant women with diabetes. Journal of Telemedicine and Telecare, 25(5), 238-242. doi:10.1258/jtt.2009.081213. Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum likelihood from in- complete data via the EM algorithm, journal of the Royal Statistical Society. Series B (Methodological, 39(1), 1-38. Duran, A., Saenz, S., Torrejon, M. J., Bordiu, E, del Valle, L, & Galindo, M. (2014). Introduction of IADPSG criteria for the screening and diagnosis of gestational diabetes mellitus results in improved pregnancy outcomes at a lower cost in a large cohort of pregnant women: The St. carlos gestational diabetes study. Diabetes Care, 37(9), 2442-2450. doi:10.2337/dcl4-0179. El-Gayar, O., Timsina, P., Nawar, N., & Eid, W. (2013a). A systematic review of IT for diabetes self-management: Are we there yet? International journal of Medical Informatics, 82(8), 637-652. doi:10.1016/j.ijmedinf.2013.05.006. El-Gayar, O., Timsina, P., Nawar, N., & Eid, W. (2013b). Mobile applications for di- abetes self-management status and potential, journal of Diabetes Science and Technology, 7(1), 247-262. Ferrara, A., Hedderson, M. M., Ching, J., Kim, C., Peng, T, & Crites, Y. M. (2012). Re- ferral to telephonic nurse management improves outcomes in women with ges- tational diabetes. American journal of Obstetrics and Gynecology, 206(6), 491.el- 491.e5. doi: 10.1016/j.ajog.2012.04.019. Franks, P. W., Looker, H. C., Kobes, S., Touger, L, Tataranni, A., & Hanson, R. L. (2006). Gestational Glucose Tolerance and Risk of Type 2 Diabetes in Young Pima Indian Offspring. Diabetes, 55, 460-465. García-Sáez, G., Alonso, J.M., & Molero, J., Rigla, M., Martinez-Sarriegui, I., de Leiva A. et al. (2009). Mealtime blood glucose classifier based on fuzzy logic for the diabtel telemedicine system. In Lectures Notes in Artificial Intelli- gence (Vol. 5651, pp. 295-304). Verona, Italy: Springer, http://doi.org/10.1007/ 978-3-642-02976-9_42. Given, J. E, O'Kane, M. J., Bunting, B. P., & Coates, V. E (2013). Comparing patient- generated blood glucose diary records with meter memory in diabetes a sys- tematic review.pdf. Diabetic Medicine, 30, 901-913. doi:10.1111/dme.l2130. Goldberg, D. E. (1989). Genetic algorithms in search, optimization and machine learn- ing. Adison-Wesley. Hall, M.A. (1999). Correlation-based feature selection for machine learning. The Uni- versity of Waikato. Retrieved from https://www.lri.fr/~pierres/donn%E9es/save/ these/articles/lpr-queue/hall99correlationbased.pdf Hartling, L., Dryden, D. M., Guthrie, A., Muise, M., Vandermeer, B., & Dono- van, L. (2013). Benefits and harms of treating gestational diabetes mellitus: A systematic review and meta-analysis for the US preventive services task force and the national institutes of health office of medical applications of research. Annals of Internal Medicine, 259(2), 123-129. http://doi.org/10.1007/ https://www.lri.fr/~pierres/donn%E9es/save/ Haykin, S.S. (1994). Neural networks: A comprehensive foundation. New York: Macmillan College Publishing Company Inc. Hernando, M. E., Gómez, E. J., Corcoy, R., & del Pozo, F. (2000). Evaluation of DI- ABNET, a decision support system for therapy planning in gestational diabetes. Computer Methods and Programs in Biomedicine, 62(3), 235-248. Hische, M., Luis-Dominguez, 0., Pfeiffer, A. F. H., Schwarz, P. E., Selbig, J., & Spranger, J. (2010). Decision trees as a simple-to-use and reliable tool to iden- tify individuals with impaired glucose metabolism or type 2 diabetes mellitus. European Journal of Endocrinology, 163(4), 565-571. doi:10.1530/EJE-10-0649. Huang, Y., McCullagh, P., Black, N., & Harper, R. (2007). Feature selection and classifi- cation model construction on type 2 diabetic patients' data. Artificial Intelligence in Medicine, 42(3), 251-262. doi:10.1016/j.artmed.2007.07.002. International Association of Diabetes and Pregnancy Study Groups Consensus Panel. (2010). International Association of Diabetes and Pregnancy Study Groups Rec- ommendations on the Diagnosis and Classification of Hyperglycemia in Preg- nancy. Diabetes Care, 33(3), 676-682. doi:10.2337/dc09-1848. International Diabetes Federation. (2015). IDF diabetes atlas seventh edition (7th ed). Brussels, Belgium: International Diabetes Federation, Retrieved from www.diabetesatlas.org. Klonoff, D. C, & True, M. W. (2009). The missing element of telemedicine for di- abetes: Decision support software, journal of Diabetes Science and Technology, 3(5), 996-1001. Kohavi, R., & John, G. H. (1996). Wrappers for feature subset selection. Artificial In- telligence, 97, 273-324. Lim, S., Kang, S. M., Shin, H., Lee, H. J., Won Yoon, J., & Yu, S. H. (2011). Improved glycemic control without hypoglycemia in elderly diabetic patients using the ubiquitous healthcare service, a new medical information system. Diabetes Care, 34(2), 308-313. doi:10.2337/dcl0-1447. Liu, H., & Yu, L (2005). Toward integrating feature selection algorithms for clas- sification and clustering. IEEE Transactions on Knowledge and Data Engineering, 17(4), 491-502. Mackillop, L, Loerup, L., Bartlett, K., Farmer, A., Gibson, O. J., Hirst, J. E., et al. (2014). Development of a Real-Time Smartphone Solution for the Management of Women With or at High Risk of Gestational Diabetes, journal of Diabetes Sci- ence and Technology, 8(6), 1105-1114. doi:10.1177/1932296814542271. MacQueen, J. (1967). Some methods for classification and analysis of multivariate observations. Proceedings of the fifth Berkeley symposium on mathematical statis- tics and probability, 1, 281-297. Metzger, B. E., Lynn, M. D., Lowe, P., & Dyer, A. R. Northwestern University Feinberg School of Medicine, Trimble, E. R. The HAPO Study Cooperative Research Group. (2008). Hyperglycemia and Adverse Pregnancy Outcomes. The New England jour- nal of Medicine, 358, 1991-2002. doi:10.1056/NEJMoa0707943. Mohan, V., & Pradeepa, R. (2014). Telemedicine in Diabetes Care: In rural India, a new prevention project seeks to fill in the screening gap. IEEE Pulse, 5(3), 2 2 - 25. doi:10.1109/MPUL2014.2309575. Mohlig, M., Floter, A., Spranger, J., Weickert, M. O., Schill, T, & Schlosser, W (2006). Predicting impaired glucose metabolism in women with polycystic ovary syn- drome by decision tree modelling. Diabetologia, 49(11), 2572-2579. doi:10.1007/ S00125-006-0395-0. Pelleg, D., & Moore, A. (2000). X-means. Extending K-means with efficient estima- tion of the number of clusters. In Proceedings of the Seventeenth International Conference on Machine Learning (pp. 727-734). San Francisco: Morgan Kauf- mann. Retrieved from http://www.cs.cmu.edu/~dpelleg/download/xmeans.pdf. Pérez-Ferré, N., Galindo, M., Fernández, M. D., Velasco, V., de la Curz, M. J., & Martin, P. (2010). A Telemedicine system based on Internet and short mes- sage service as a new approach in the follow-up of patients with gestational diabetes. Diabetes Research and Clinical Practice, 87(2), e l 5 - e l 7 . doi:10.1016/j. diabres.2009.12.002. Polonsky, W H., Jelsovsky, Z., Panzera, S., Parkin, C. G., & Wagner, R. S. (2009). Pri- mary care physicians identify and act upon glycemic abnormalities found in structured, Episodic blood glucose monitoring data from non-insulin-treated type 2 Diabetes. Diabetes Technology & Therapeutics, 11(5), 283-291. (Version 0.7.9). Retrieved from PSPP. https://www.gnu.org/software/pspp/. Quinlan, J. Ross (1993). C4.5: programs for machine learning. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc. Ouinn, C. C, Clough, S. S., Minor, J. M., Lender, D., Okafor, M. C, & Gruber- Baldini, A. (2008). WellDoc mobile diabetes management randomized con- trolled trial: change in clinical and behavioral outcomes and patient and physi- cian satisfaction, iabetes Technology & Therapeutics, 10(3), 160-168. doi:10.1089/ dia. 2008.0283. Retrieved from Diagnostics, Roche. Accu-Check Aviva user's manual https://www. accu-chek.ca/multimedia/documents/aviva2-user-manual.pdf. Retrieved from Diagnostics, Roche. Accu-Check Nano user's manual https://www. accu-chek.com/us/glucose-meters/nano-how-to-use.html. Diabetes, Sanofi (2015). iBGStar(g> Diabetes Manager Application - MyStar (Ver- sion 2.2.1.). Sanofi Diabetes. Retrieved from http://www.mystarsanofi.com/web/ products/softwares/ibgstar-application. Shankaracharya, Odedra, D., Mallick, M., Shukla, P., Samanta, S., & Vid- yarthi, A. S. (2012). Java-based diabetes type 2 prediction tool for better diag- nosis. Diabetes Technology & Therapeutics, 14(3), 251-256. doi:10.1089/dia. 2011. 0202. Upadhyaya, S., Farahmand, K., & Baker-Demaray, T. (2013). Comparison of NN and LR classifiers in the context of screening native American elders with diabetes. Expert Systems with Applications, 40(15), 5830-5838. doi:10.1016/j.eswa.2013.05. 012. Witten, I. H., & Frank, E. (2005). Data Mining, practical machine learning tools and techniques (2nd ed). San Francisco: Morgan Kaufmann Publishers. Wojcicki, J. M., Ladyzynski, P., Krzymien, J., Jozwicka, E., Blacjowic, J., & Janczewska, E. (2001). What we can really expect from telemedicine in inten- sive diabetes treatment: Results from 3-year study on type 1 pregnant diabetic women. Diabetes Technology & Therapeutics, 3(4), 581-589. http://www.diabetesatlas.org http://www.cs.cmu.edu/~dpelleg/download/xmeans.pdf https://www.gnu.org/software/pspp/ https://www https://www http://accu-chek.com/us/glucose-meters/nano-how-to-use.html http://www.mystarsanofi.com/web/ 
work_p2drr5fdwzhgvcdp6idlyl5yo4	----	Digital Library of Expert System Based at Indonesia Technology University (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 4, No.3, 2015 1 | P a g e www.ijarai.thesai.org Digital Library of Expert System Based at Indonesia Technology University Dewa Gede Hendra Divayana 1 Chair of Information Technology Department Indonesia Technology University Bali, Indonesia I Putu Wisna Ariawan 2 Lecture of Mathematics Education Ganesha University of Education Bali, Indonesia I Made Sugiarta 3 Lecture of Mathematics Education Ganesha University of Education Bali, Indonesia I Wayan Artanayasa 4 Chair of Sport & Health Education Department Ganesha University of Education Bali, Indonesia Abstract—Digital library is a very interesting phenomenon in the world of libraries. In this era of globalization, the digital library is needed by students, faculty, and the community in the search for quick reference through internet access, so that students, faculty, and the community does not have to come directly to the library. Accessing collections of digital libraries can also be done anytime and anywhere. Digital Library development also occurred at Indonesia Technology University. That University offers a digital library based of expert system. The concept of digital library is utilizing science expert system in the process of cataloging and searching digital collections. By using this digital library based of expert system, users can search the collection, reading collection, and download the desired collection by online system. The digital library based of expert system at Indonesia Technology University is built using the PHP programming language, MySQL database as a data base management system, and developed the method of forward chaining and backward chaining as inference engine. Keywords—Digital Library; Expert System; Forward Chaining; Backward Chaining I. INTRODUCTION In the current era of globalization, information technology has a very important role in supporting the activities carried out by the community. The development of information technology very rapidly leads us toward the use of documents in digital form. Library as one of the sources of knowledge needs to be organized and presented for the system services can be accessed from anywhere and anytime with the involvement of information technology in the form of an integrated system, so the user does not have to come directly to the library. This phenomenon is supported by the use of the Internet that facilitate processing, dissemination, and accessing information in digital form from anywhere and anytime. The development of increasingly sophisticated technology also affects the formation of a library in a digital form. Digital library is a very interesting phenomenon in the world of libraries. In this era of globalization, the digital library is needed by students, faculty, and the community in the search for quick reference through internet access, so that students, faculty, and the community do not have to come directly to the library. Accessing collections of digital libraries can also be done anytime and anywhere. Digital Library development also occurred at Indonesia Technology University. That University offers a digital library based of expert system. The concept of digital library is utilizing science expert system in the process of cataloging and searching digital collections. By using this digital library based of expert system, users can search the collection, reading collection, and download the desired collection by online system. II. LITERATURE REVIEW A. Digital Library In [1], The library has become digital: processes such as mass digitization, web archiving, and to a smaller extent digital preservation, are no longer isolated but disseminated among relevant production teams within the library. In [2], A digital library (DL) is a library in which collections are stored in digital formats (as opposed to print, microform, or other media) and accessible by computers. In [3], A digital library is a particular kind of information system which consists of a set of components, typically a collection (or collections) of computer system offering diverse services on a technical infrastructure, people, and the environment or usage. In [4], Digital libraries are set of library activities and services which facilitate by electronic means the processing, transmission and display of information. B. Expert System In [5], an expert system is a computer program designed to simulate the problem-solving behaviour of a human who is an expert in a narrow domain or discipline. An expert system is normally composed of a knowledge base (information, heuristics, etc.), inference engine (analyses the knowledge base), and the end user interface (accepting inputs, generating outputs). The concepts for expert system (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 4, No.3, 2015 2 | P a g e www.ijarai.thesai.org development come from the subject domain of artificial intelligence (AI), and require a departure from conventional computing practices and programming techniques. In [6], an Expert system is software that simulates the performance of human experts in a specific field. Today’s expert systems have been used in many areas where require decision making or predicting with expertise. In [7], an expert system is a set of programs that manipulate encoded knowledge to solve problems in a specialized domain that normally requires human expertise. In [8], Expert System is a branch of Artificial Intelligence that makes extensive use of specialized knowledge to solve problems at the human expert level. In [9], the Expert System (ES) is one of the well-known reasoning techniques that is utilized in diagnosis applications domain. In ES, human knowledge about a particular expertise to accomplish a particular task is represented as facts and rules in its knowledge base. In [10], an expert system is the computer system that emulates the behaviour of human experts in a well- specified manner, and narrowly defines the domain of knowledge. It captures the knowledge and heuristics that an expert employs in a specific task. An overview of current technologies applied with an expert system that is developed for Database Management System, Decision Support System, and the other Intelligent Systems such as Neural Networks System, Genetic Algorithm, etc. In [11], expert system is an artificial intelligence system that combines knowledge base with inference engine so that it can adopt the ability of the experts into a computer, so the computer can solve problems such as the often performed by experts. C. Forward Chaining The inference engine contains the methodology used to perform reasoning on the information in the knowledge base and used to formulate conclusions. Inference engine is the part that contains the mechanism and function of thought patterns of reasoning systems that are used by an expert. The mechanism will analyze a specific problem and will seek answers, conclusions or decisions are best. Because the inference engine is the most important part of an expert system, that plays a role in determining the effectiveness and efficiency of the system. There are several ways that can be done in performing inference, including the Forward Chaining. In [12], an inference engine using forward chaining searches the inference rules until it finds one where the IF clause is known to be true. When found it can conclude, or infer, the THEN clause, resulting in the addition of new information to its dataset. In other words, it starts with some facts and applies rules to find all possible conclusions. Therefore, it is also known as Data Driven Approach. In [13], forward chaining is matching facts or statements starting from the left (first IF). D. Backward Chaining Also in [13], backward chaining is matching facts or statements starting from the right (first THEN). In other words, the reasoning starts from the first hypothesis, and to test the truth of this hypothesis to look for the facts that exist in the knowledge base. In [14], an inference engine using backward chaining would search the inference rules until it finds one which has a THEN clause that matches a desired goal. If the IF clause of that inference rule is not known to be true, then it is added to the list of goals (in order for goal to be confirmed it must also provide data that confirms this new rule). In other words, this approach starts with the desired conclusion and works backward to find supporting facts. Therefore, it is also known as Goal-Driven Approach. In [15], backward chaining systems are good for diagnostic and classification tasks, but they are not good for planning, design, process monitoring, and quite a few other tasks. In backward chaining, the search is goal directed, so rules can be applied that are necessary to achieve the goal. E. Digital Library Based of Expert System In [16], Digital Library Based of Expert System is a digital library that implements the basic concept of expert system includes a knowledge base and the inference engine in helping service mechanism. The knowledge base is used for storage and cataloging process of making collections that exist in digital library, while the inference engine is used to search the detail collection that available in digital libraries such as the ability to work like an expert. III. METHODOLOGY A. Object dan Research Site 1) Research Object is Digital Library Based of Expert System 2) Research Site at Indonesia Technology University. B. Data Type In this research, the authors use primary data, secondary data, quantitative data and qualitative data. C. Data Collection Techniques In this research, the authors use data collection techniques such as interviews, observation, and documentation. D. Analysis Techniques Analysis techniques used in this research is descriptive statistical. IV. RESULT AND DISCUSSION A. Result 1) Early Trial At this early trial, the authors conducted a limited scale testing of the digital library based of expert system that have been made previously by involving five staff at Indonesia Technology University to perform white box and black box testing. This test can be done by giving 10 questionnaires (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 4, No.3, 2015 3 | P a g e www.ijarai.thesai.org early trials digital library based of expert system to staff at Indonesia Technology University. Diagram form of answers score percentage given by the respondents in early trial can be described as follows: Fig. 1. Percentage Diagram of Respondents Answer Score In Early Trial Based on the diagram above, it can be seen that the results of early trials of the digital library based of expert system, find a constraint that is the answer to a very bad score by 90% of the questions on the questionnaire 1 st initial trials. This is due to the unavailability of the form for the create of a new username and password for administrator in the future if there is a mutation of the staff who operate the digital library based of expert system. Given these constraints, then the system needs to be revised again. 2) Field Trial At this field trial, the authors tested in a larger scale, involving an expert is understood about the digital library and ten staff at Indonesia Technology University. This test can be done by giving 15 questionnaires field trials for digital library based of expert system to the librarian and staff at Indonesia Technology University. Diagram form of answers score percentage given by the respondents in field trial can be described as follows: Fig. 2. Percentage Diagram of Respondents Answer Score In Field Trial Based on the diagram above, it can be seen that the results of a field trial of the digital library based of expert system, the presence of obstacles that scores are very bad answer the score of 73% to the question 1 st , 3 rd , 5 th , and 15 th , 82% of the questions 4 th , 11 th and 13 th , at 91% of the questions 7 th , 9 th , 10 th , and at 100% of the questions 6 th , 8 th and 14 th on field trial questionnaire. This is due to the unavailability of the collection input or edits form if in the future there is a new journal and article. Of the constraints are found, then the system needs to be revised to obtain collection more interactive and dynamic. 3) Usage Test At this usage test, the authors conducted a trial involving with the use of 50 people (users). The test is performed to test the operation of the overall form available on digital library based of expert system that has undergone revisions to field trials. This test can be done by giving the user satisfaction questionnaire to users who visited Indonesia Technology University. Diagram form of answers score percentage given by the respondents in usage test can be described as follows: (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 4, No.3, 2015 4 | P a g e www.ijarai.thesai.org Fig. 3. Percentage Diagram of Respondents Answer Score In Usage Test Based on the diagram above, it can be seen that the results of testing the use of the digital library based of expert system outline already looks very good and not found again the constraints in terms of technical operation (inputing and editing a new collection) as well as the principle method of expertise (forward chaining and bacward chaining method). This is evidenced by the percentage scoring very good response by 78% of questions 3 rd , and 9 th . Percentage scoring very good response by 82% of questions 7 th . Percentage scoring very good response by 84% of questions 1 st and 5 th . Percentage scoring very good response by 88% of statement 6 th . Percentage scoring very good response by 94% of question 2 nd . Percentage scoring very good response by 96% of questions 10 th . As well as scoring 98% of the questions 4 th and 8 th trial usage. And it would be even better if the digital library based of expert system added help fasility form for written in accordance with the suggestions of the respondents to the improvement of the system, so as to explain the performance of the expert system and the function of the buttons in the design of digital library based of expert system overall with easy to understand and simple language. B. Discussion 1) Knowledge Base Knowledge base is used to build the expert system obtained from multiple sources of knowledge, containing data on digital library based of expert system at Indonesia Technology University. The knowledge base contained in a digital library based of expert system at Indonesia Technology University can be described by the following table. TABLE I. KNOWLEDGE BASE ON DIGITAL LIBRARY BASED OF EXPERT SYSTEM AT INDONESIA TECHNOLOGY UNIVERSITY No Properties/Identity Journal Book Magazine Article 1. Collection Code     2. Collection Name     3. Author     4. Publisher     5. Year of Issue     6. Pages     7. Place of Issue     8. ISSN     2) Shows forward chaining concept in digital library based of expert system Application of forward chaining method in a digital library can be explained by the following chart: For example : search by collection code Collection Code Collection Name Author Year of Issue Place of Issue PagesISSNPublisher Author Year of Issue Place of Issue PagesISSNPublisher Year of Issue Place of Issue PagesISSNPublisher Year of Issue Place of Issue PagesISSN Place of Issue PagesISSN ISSNPage ISSN Block 1 Block 2 Block 3 Block 4 Block 5 Block 6 The model is same with block 1 The model is same with block 2 The model is same with block 3 The model is same with block 4 The model is same with block 5 The model is same with block 6 Explanation : = Search by Fig. 4. Forward Chaining Concept in Digital Library Based of Expert System 3) Shows backward chaining concept in digital library based of expert system Application of backward chaining method in a digital library can be explained by the following chart: For example : search by ISSN (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 4, No.3, 2015 5 | P a g e www.ijarai.thesai.org Collection Code Collection Name Author Year of Issue Place of Issue PagesISSNPublisher Author Year of Issue Place of Issue PagesISSNPublisher Year of Issue Place of Issue PagesISSNPublisher Year of Issue Place of Issue PagesISSN Place of Issue PagesISSN ISSNPage ISSN Block 1 Block 2 Block 3 Block 4 Block 5 Block 6 The model is same with block 1 The model is same with block 2 The model is same with block 3 The model is same with block 4 The model is same with block 5 The model is same with block 6 Explanation : = Search by Fig. 5. Backward Chaining Concept in Digital Library Based of Expert System 4) Trials forward chaining and backward chaining performed by respondents Respondents who did this trial was a expert and 19 staff at Indonesia Technology University conducted the field trials. The trial results are shown in table II. Based on the table results of trials forward chaining and backward chaining performed by respondents mentioned above, it can be analyzed that the forward chaining and backward chaining method has been run in accordance with the rule of expert system inference engine. To view the forward chaining and backward chaining method has been run in accordance with the rules can be seen in the percentage diagram of rules conformance testing. TABLE II. TRIALS FORWARD CHAINING AND BACKWARD CHAINING METHOD Respondent Method Properties/Identity % C C C N A T P L Y I P G P I I S RS.01 FC         100 BC         100 RS.02 FC         100 BC         100 RS.03 FC         100 BC         100 RS.04 FC         100 BC         100 RS.05 FC         100 BC         100 RS.06 FC         100 BC         100 RS.07 FC         100 BC         100 RS.08 FC         100 BC         100 RS.10 FC         100 BC         100 RS.11 FC         100 BC         100 RS.12 FC         100 BC         100 RS.13 FC         100 BC         100 RS.14 FC         100 BC         100 RS.15 FC         100 BC         100 RS.16 FC         100 BC         100 RS.17 FC         100 BC         100 RS.18 FC         100 BC         100 RS.19 FC         100 BC         100 RS.20 FC         100 BC         100 Explanation : FC : Forward Chaining BC : Backward Chaining CC : Collection Code CN : Collection Name AT : Author PL : Publisher YI : Year of Issue PG : Pages PI : Place of Issue IS : ISSN As for the form of percentage diagram of rules conformance testing given by the respondents can be described as follows: (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 4, No.3, 2015 6 | P a g e www.ijarai.thesai.org Fig. 6. Answer Percentage Diagram of Rules Conformance Testing Based on the diagram above, it can be seen that the results of testing the suitability of digital library based of expert system rules is an outline already looks qualify. This is evidenced by the answer percentage of collection code, collection name, author, publisher, year of issue, pages, place of issue and ISSN according to the rules of backward chaining and forward chaining in the field of testing and each get a percentage of 100%. 5) Implementation of Digital Library Based of Expert System a) Main Menu Page Fig. 7. Main Page This main menu page used as link to home menu, about us menu, book case menu, and contact menu. b) Membership Registration Form Fig. 8. Membership Registration Form This form is used by users who will register to become a member, that all fields must be filled in accordance with the user's identity, and then click the register button to register and click cancel button to cancel the registration. c) Member Login Form Fig. 9. Member Login Form This form contains the username and password for the members, so that members can search and download a collection of all the collections that exist in digital library based of expert system at Indonesia Technology University. d) Latest Collection Menu Page After login, the user will go to the latest collection menu, the menu contains the latest collections of Digital Library based of expert system. The following figure is a latest collection menu display. Fig. 10. Latest Collection Menu Page e) Collection List Page Fig. 11. Collection List Page (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 4, No.3, 2015 7 | P a g e www.ijarai.thesai.org This collection list page is a page that is used to display a list of collections that exist in the digital library based of expert system at Indonesia Technology University and displays the complete details of the collection identity. f) Collection Search Page This form contains the search facilities of data on existing collections in digital library based of expert system using a keyword based input the desired category. Fig. 12. Collection Search Page g) Collection Search Result Page Collection Search Result Page is the page serves to display the collection search result page. The following figure is a collection search results page display. Fig. 13. Collection Search Result Page h) Administrator Login Form Fig. 14. Administrator Login Form This form contains the username and password for the Admin, click on login button if you want to login and click cancel button if you want to cancel. i) Administrator Page Administrator page is a page that is used by administrators to system perform processing, such as processing of knowledge base and put the rule into an expert system inference engine. The following figure is a administrator page display. Fig. 15. Administrator Page V. CONCLUSIONS Based on the analysis that has been made and the results of the discussion in the previous section, then some conclusions can be drawn as follows: a) With this digital library based of expert system can help the performance of existing conventional systems towards a computerized system, so as to speed up service. b) With digital library based of expert system makes it easy for users to search the collection (books, journals, magazines, and articles) in digital form through the facilities of the internet without having to come directly to the campus library. c) With the digital library digital library based of expert system, the concept of a real expert system can already be applied and useful in solving the existing problems in human life, especially in the field of library. d) By looking at the optimal use of knowledge base and also the use of forward chaining and backward chaining inference engine implemented as in this digital library based of expert system, it can also be the accuracy result of the system is working at 100%. ACKNOWLEDGMENT The authors express their gratefulness to staff at Indonesia Technology University for inspiring words and allowing them to use the examination data. They generously thank Mr. Dayung, President of Indonesia Technology University, and Mr. Ketut Semadi, Dean of Computer Faculty, Indonesia Technology University. REFERENCES [1] E.Bermès, and L.Fauduet, “The Human Face of Digital Preservation: Organizational and Staff Challenges, and Initiatives at the Bibliothèque (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 4, No.3, 2015 8 | P a g e www.ijarai.thesai.org nationale de France,” in The International Journal of Digital Curation vol.6, 2011,pp.226-237. [2] D.I.Greenstein, Thorin, and S. Elizabeth. The Digital Library: A Biography. Washington: Digital Library Federation, 2002. [3] N. Fuhr, G. Tsakonas, T. Aalberg, M.Agosti, P. Hansen, and S. Kapidakis,“Evaluation of digital libraries,” in International Journal on Digital Libraries vol. 8, 2007, pp.21–38. [4] K. Towolawi, and Oluwakemi, “School Library Media Specialist’s Awareness and Perception of Digital Library Services: A Survey,” in Ozean Journal of Social Sciences vol. 6, 2013, pp.77-89. [5] Y.A. Nada, “Construction of Powerful Online Search Expert System Based on Semantic Web,” in International Journal of Advanced Computer Science and Applications vol.4, 2013, pp.181-187. [6] Y. Qu, F. Tao, and H. Qui, “A Fuzzy Expert System Framework Using Object Oriented Techniques,” in IEEE Pacific-Asia Workshop on Computational Intelligence and Industrial Application, 2008, pp. 474- 477. [7] Y. Erdani, “Developing Recursive Forward Chaining Method in Ternary Grid Expert Systems,” in International Journal of Computer Science and Network Security, vol.11, No.8, 2011, pp.126-130. [8] J.C. Giarratano, and G. Riley, Expert Systems : Principles and Programming 4 th Edition. USA : PWS Publishing Co, 2004. [9] A. A. Hopgood, Intelligent Systems for Engineers and Scientists (2 nd Edition). USA : CRC Press, 2001. [10] E. Turban, and J. E. Aronson, Decision Support Systems and Intelligent System. NJ, USA: Prentice-Hall Inc, 2001. [11] H. Divayana, “Development of Duck Diseases Expert System with Applying Alliance Method at Bali Provincial Livestock Office,” in International Journal of Advanced Computer Science and Applications, Vol. 5, No. 8, 2014,pp.48-54. [12] RC Chakraborty, 2010. Knowledge Representation: AI Course Lecture 15-22.Retrieved December, 2012, from www.myreaders.info/ html/aritificial_intelligence.html [13] S.Kusumadewi, Artificial Intelligence (Technique and Application) 1 st Edition. Yogyakarta : Graha Ilmu, 2003. [14] K. Donna, A Comparison of Forward and Backward Chaining Algorithms For use in a Technical Support Expert System Used for Diagnosing Computer Virus Issues. Chico : Computer Science Department California State University, 2009. [15] T. Sharma, et.al., “Study of Difference Between Forward and Backward Reasoning,” in International Journal of Emerging Technology and Advanced Engineering, vol. 2 No. 10, 2012, pp.271-273. [16] H. Divayana, et.al., Digital Library Based of Expert System at Indonesia Technology University. Bali : Indonesia Technology University, 2015. AUTHORS PROFILE Dewa Gede Hendra Divayana, S.Kom., M.Kom., M.M., Ph.D. was born in Denpasar, Bali, in 1984. He received his Ph.D. in Information Technology from Corllins University, USA. He worked as Lecturer of Expert System, Chair of Information Technology Department, Faculty of Computer, Indonesia Technology University, Bali, Indonesia. Drs. I Putu Wisna Ariawan, M.Si. was born in Ulakan, May 19 th 1968. He worked as Lecturer of Mathematics Education at Ganesha University of Education. And also He worked as Guest Lecturer of Graph Theory at Faculty of Computer, Indonesia Technology University, Bali, Indonesia. Drs. I Made Sugiarta, M.Si. was born in Badung, in 1967. He worked as Lecturer of Mathematics Education at Ganesha University of Education. And also He worked as Guest Lecturer of Dicrete Mathematic at Faculty of Computer, Indonesia Technology University, Bali, Indonesia. I Wayan Artanayasa, S.Pd., M.Pd. was born in Manikliyu, in 1973. He worked as Lecturer of Sport & Health Education at Ganesha University of Education. And also He worked as Guest Lecturer of Human-Computer Interaction at Faculty of Computer, Indonesia Technology University, Bali, Indonesia. http://www.myreaders.info/ 
work_p3wb7fys7ncfvoks3dqhrs7wpq	----	Research Article Fuzzy Logic Expert System for Classifying Solonchaks of Algeria Samir Hadj Miloud ,1,2 Kaddour Djili,1 and Mohamed Benidir3 1Soil Science Department, Higher National Agronomic School (ENSA-ES1603), BP 16200 El Harrach, Algeria 2Department of Agronomics, Faculty of Natural Science and Life, University Saad Dahlab, Soumâa, BP 270 Blida, Algeria 3Unit of Sétif, Institut National de la Recherche Agronomique (INRAA), El Harrach, Algeria Correspondence should be addressed to Samir Hadj Miloud; shadjmiloud@gmail.com Received 29 November 2017; Accepted 13 June 2018; Published 8 July 2018 Academic Editor: Claudio Cocozza Copyright © 2018 Samir Hadj Miloud et al. +is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Under arid and semiarid regions of the North of Africa, the soils considered as Solonchaks contain both calcium carbonate and gypsum. When these elements are presented at high quantities, these Solonchaks are getting close to Calcisol or Gypsisol. +e World Reference Base (WRB) for soil classification does not take into account the soil as a continuum. Instead, this international soil system classification is based on threshold values that define hierarchical diagnostic criteria. Consequently, the distinction between Solonchaks, Calcisol, and Gypsisol is still not clear. To avoid this situation, fuzzy logic based on the Mamdani inference system (MFIS) was used to determine to what extent soil classified as Solonchak in WRB can interfere with Calcisols and Gypsisols. For that purpose, membership values of Solonchaks (Is), Calcisols (Ic), and Gypsisols (Ig) indices were calculated from 194 soil profiles previously classified as Solonchak in WRB. Data analyses revealed that Solonchaks soils were subdivided into Solonchaks (61%), Calcisols (1%), Gypsisols (0.5%), Solonchaks-Calcisols intergrades (29%), Solonchaks-Gypsisols intergrades (5%), and Solonchaks-Calcisols-Gypsisols intergrades (2%). Moreover, Is, Ic, and Ig showed high significant correlations with almost all WRB diagnostic criteria (P<0.05). Under our study, soil classification obtained by employing MFIS was analogous to that provided by WRB; however, MFIS exhibited high precision concerning the membership value between soils and their intergrades. +erefore, the application of MFIS for other soil classifications in the world is possible and could lead to improvement in conventional soil classification. 1. Introduction Soil classification is considered as an important means of communication at both national and international levels [1, 2]. However, few soil classification studies have been published [3]. +e lack of soil classification reduces our knowledge and affects our land use decision. +is difficulty is compounded by the fact that the hierarchical classifications are often built on criteria that vary greatly from one to the other. +e classical reference proposed by Baize and Girard [4] reduces the number of hierarchical levels, mitigates these challenges, and represents a significant improvement. Both conventional classifications (i.e., hierarchical and classical reference) are the most used. However, there are many other national classifications, such as the Australian classification system [5], Canadian [6], and French [7]. Currently, the International Union of Soil Science promotes the development of a universal soil classification system [8] in which all the soils of the world find a place in its hierarchy. Preliminary results suggest that the objective and pedo- metric approaches can support the planned development of a universal soil classification system. As the soil is part of an ecological continuum, the usual classifications are facing major challenges which require a choice, often questionable [9], between the base characters and their importance on the hierarchy of taxonomic units. Conventional classifications define several intergrades be- tween the main units. Hierarchical classification systems are based partly on the judgment of the expert on soil formation. 40 years ago, the first attempts at the numerical soil clas- sification was made [10, 11]. McBratney and Odeh [12] argued that a system of discrete soil classes is not adequate for soil classification and proposed a numerical classification based on fuzzy sets. Numerical approaches can deal with Hindawi Applied and Environmental Soil Science Volume 2018, Article ID 8741567, 11 pages https://doi.org/10.1155/2018/8741567 mailto:shadjmiloud@gmail.com http://orcid.org/0000-0002-7360-4388 https://doi.org/10.1155/2018/8741567 a large number of properties simultaneously. Currently, there are numerical classification systems that attempt to be an objective classification, which are based on the actual dif- ferences between morphological and analytical characteristics of the soils. +eir main goal is to minimize intragroup var- iations and maximize intergroup variations according to objective criteria [13]. WRB [14] has been favored by the International Union of Soil Science and the European Union as a correlation system between the soils [15]. It defines the different groups of soil horizons in terms of references, properties, and diagnostic materials; each criterion is quan- titative and well differentiated. Some studies have shown that WRB compared to Soil Taxonomy [16] is well suited for classifying calcareous soils [17] and gypsum soils [18, 19]. Similarly, at the end of the 16th World Congress of Soil Science, it was recommended to strengthen conventional methods of mapping by other methods, such as fuzzy logic and artificial intelligence to assess the inherent uncertainties in soil mapping [20]. Fuzzy logic has become widespread in recent years in many scientific fields, such as soil science [21–23]. It aims at dealing with the uncertainty due to the imprecision [24]. Fuzzy systems belong to the class of systems based on knowledge or expert systems. +eir main purpose is to implement a human skill, or linguistic rules, by a computer program. Fuzzy logic provides a mathematical formalism with uncertain linguistic concepts. +is mathe- matical method, which is based on set theory, has been introduced by Zadeh [25]. +us, many studies have focused on fuzzy logic to the study of soils and their classifications. It follows that fuzzy logic allows the creation of non- hierarchical continuous classes defined by their centers of gravity [12]. +e notion of intergrade soils has been formally recognized by using the concept of fuzzy sets [12]. +e fuzzy logic-based algorithms can estimate the number of soil intergrades [26]. +e combination of fuzzy clustering to other techniques is currently used in the development of models for the prediction of soil properties [27, 28]. In general, this theory has great potential in soil science [12]. +e use of this system can replace the Boolean variable, which is poorly suitable to the presentation of most natural phe- nomena, by using linguistic concepts that will be transformed into multilanguage values. +ere are several fuzzy inference systems that were used in different applications; the most commonly used is the Mamdani fuzzy inference system (MFIS), which will be used in this research. +e advantages of MFIS are reputed as intuitive; it is the most widespread ac- ceptance and better suited to human cognition [29]. In arid and semiarid regions of the North of Africa, the calcium carbonate accumulations are frequently associated with gypsum and soluble salts. In many cases, soils are at the same time saline, calcareous, and gypsic [30]. Rahmouni [31] studied soil classification in Algeria (semiarid area) using WRB. +is author revealed high similarity between Gypsi- sols and Solonchak and concluded that the soil studied could be considered as an intermediary group between Gypsisols and Solonchaks. To improve WRB soil classification, the fuzzy classification (continuous) is of great importance to group all soil types into continuous classes with membership values [12], to avoid imprecise and ambiguous classification. In this context, the expert system based on MFIS was used to (1) determine the degree of membership between Solon- chaks in northern Algeria and both Calcisols and Gypsisols based on WRB criteria and (2) to see to what extent these Solonchaks may constitute Calcisols or Gypsisols. 2. Materials and Methods 2.1.StudiedSoilCharacteristics. +is research focused on the study of 194 profiles identified by Djili [30] (Figure 1) in the north of Algeria. Using WRB [14] criteria, all these profiles were considered as Solonchaks. +e characteristics of the diagnostic horizons of all profiles are shown in Table 1. N 0° 3° 6° 9° 36° 33° Morocco Tunisia Mediterranean Sea 1/4000 000 O ra n A lg er A nn ab a Profile of Solonchak Administrative border Figure 1: Solonchaks location map. 2 Applied and Environmental Soil Science +e main features of the whole horizons are summarized as follows: (i) Electrical conductivity (ECe) of saturated soil-paste varies between 15 and 96 dS/m with an average of 30 dS/m. +ese Solonchaks are marked by a very high salinity [32] that varies highly (CV � 56%) from a soil to another. (ii) Calcium carbonate equivalent is variable (CV � 56%) and is between 1% and 67% with a mean of 23%. +ese Solonchaks can therefore be very little cal- careous, or, conversely, very heavily filled in calcium carbonate equivalent. Some of these horizons follow the criteria of the calcic horizon. (iii) Gypsum content is highly variable (CV � 115%). +ey range from less than 1% to 73% (by mass) with an average of 9%. +erefore, diagnostic horizons of these soils are extremely gypsic. Some of these horizons follow the criteria of the gypsic horizon. (iv) Soil pH of the saturation extract values is between 7 and 8.9 with an average of 7.83, which indicates an alkaline soil reaction. (v) +e thickness of the diagnostic horizons (E) is highly variable (CV � 55%), and these Solonchaks may have diagnostic horizons moderately thick to very thick. +ese characteristics suggest that these Solonchaks are related to Calcisols than Gypsisols. +e variables or physical values used for the three groups studied soils (Solonchaks, Calcisols, and Gypsisols) are presented in Table 2. All studied Solonchaks have ECe≥15 dS/m, and therefore the pH is no taken into account according to WRB classification. Among the 194 studied Solonchaks profiles, 74% are Calcic Solonchak, 10% are Gypsic Solonchak, 4% are Gypsic Calcic Solonchak, and 12% are Solonchak, which are neither Gypsic nor Calcic. Despite the high content of calcium carbonate equivalent and gypsum, we observed a lack of petrocalcic and petrogypsic qualifiers because of the absence of more diagnostic criteria. 2.2. Decision-Making. +e expert system based on MFIS requires three steps, the fuzzification, inference, and defuzzification, as shown in Figure 2. 2.2.1. Fuzzification. +e fuzzification is a process of con- verting numeric values (or physical parameters of the diagnostic criteria) of each group of soil (Table 2) into fuzzy variables. Compared to other numerical classifications as distance metrics method [33], neither crisp data (nonfuzzy) nor model assumptions are required [34–37], which is considered as one of the major advantages of MFIS. During this step, we firstly defined the membership function of all variables, and then, we proceeded to the passage from the physical quantities to the linguistic vari- ables. +e membership functions describe the membership degree of a fuzzy variable (the ECe in this case) to a fuzzy subset A (little ECe value, medium, or great), and it is noted as μA(x) � [01] six ∈ Aet μA(x) � 0 six ∉ A. Fuzzification of all physical variables has been applied using the Gaussian membership function and the fuzzy set. +e fuzzy variables (input) were divided into three subsets using linguistic variables (little value (L), medium value (M), and great value (G)). On the other hand, the output was also divided into three subsets (little value (L), medium value (M), and great value (G)) (Table 2, Figures 3 and 4). +e advantage of this method, using linguistic variable, is to avoid threshold values that are not adequate for continuum soil. +e boundaries of linguistic variables of calcium car- bonate equivalent, gypsum, electrical conductivity, sec- ondary carbonates (SC), and thickness of diagnostic horizon L (0–30%), M (30–80%), and G (80–100%) are shown in Figure 3. +e electrical conductivity was divided into L (0– 30 dS/m), M (30–80 dS/m), and G (80–100 dS/m) linguistic variables. For the thickness of the diagnostic horizon, three subsets were defined as cited above (L (0–30 cm), M (30– 80 cm), and G (80–100 cm)), and finally, for both thick- ness of the diagnostic horizons × ECe and thickness of the Table 1: Characteristics of diagnostic horizons of studied Solonchaks. Parameters Minimum Maximum Mean SD CV (%) pH 7 8.9 7.83 0.3 4 ECe (dS/m) at 25°C 15 96 30 17 56 Calcium carbonate equivalent (%) 1 67 23 13 56 Gypsum (%) (by mass) 0 73 9 10 115 E (cm) 15 1.5 69 38 55 Note. E: thickness of the diagnostic horizons; ECe: electrical conductivity; SD: standard deviation; CV: coefficient of variation. Table 2: Physical quantities of soil groups recommended by WRB [14]. Soil groups (output variables) Variables used [14] (input variables) Solonchaks (i) ECe (dS/m) (ii) E (cm) (iii) ECe × thickness of the diagnostic horizons (ECe × E) Calcisols (iv) Calcium carbonate equivalent (%) (iv) E (cm) (vi) Secondary carbonate (%) by volume Gypsisols (vii) Gypsum content (%) (by mass) (viii) E (cm) (ix) +ickness of diagnostic horizons × content in gypsum (E × gypsum) Applied and Environmental Soil Science 3 diagnostic horizon × gypsum, the linguistic variables were L (0–2000), M (2000–7000), and G (7000–9000). +ese boundaries of linguistic variables are based on the human knowledge from field experiences. +e same procedure was performed for the output variables which have been translated into Solonchaks indices, Calcisols indices, and Gypsisols indices as shown in Figure 4. 2.2.2. 6e Inference Rules. +e inference rules were de- veloped using the 9 input data (diagnostic criteria or physical variables) previously divided into three subgroups that represent Solonchak, Calcisol, and Gypsisol, re- spectively (Table 2). +e soil was classified Solonchak if all its diagnostic criteria were great (G). +e same was applied for Calcisol and Gypsisol. For example, if our soil presented Input variables Output variables Indices ECe Gypsum Fuzzification Defuzzification Solonchak Gypsisol Calcisol Fuzzy inference system Calcium carbonate equivalent x1 x1 x2 x3 xN y1 y2 y3 yN μ (x1) μ (y1) μ (yN) μ (x2) xj xi μ (xN) Membership function Rule base ... .... .... ..yz Figure 2: Reasoning of the database schema by a fuzzy inference system. 1 0.8 0.6 0.4 0.2 0 0 20 30 40 50 60 70 80 90 100 20 30 40 50 60 70 80 90 100 20 30 40 50 60 70 80 90 100 20 30 40 50 60 70 80 90 100 20 30 40 50 60 70 80 90 100 1000 2000 3000 4000 5000 6000 7000 8000 9000 1000 2000 D eg re e of m em be rs hi p 3000 4000 5000 6000 7000 8000 9000 0 0 0 0 0 0 Little value Medium value Great value Carbonate equivalent (%) Secondary carbonate (%) by volume Gypsum content (%) (by mass) E (cm) ECe (dS/m) (E × gypsum) Ece × E Figure 3: Membership function of input variables. 4 Applied and Environmental Soil Science little calcium carbonate equivalent and (SC) (that char- acterize Calcisol), medium gypsum content (that charac- terize Gypsisol), and great ECe (that characterize Solonchak), the soil will be classified as Solonchak. +ese rules are expressed as single conditions (IF) or combined with other conditions (AND, OR) to achieve a linguistic result. Each rule consists of an antecedent part (condition or input) expressed by IF and a substantial portion (conclusion or output) expressed by THEN. For example, IF ECe is G AND (ECe × E) is G AND E is G (Solonchak criteria), AND calcium carbonate equivalent is L AND SC is L (Calcisol criteria), AND gypsum content is L AND (E × gypsum) is L (Gypsisol criteria) THEN Solonchak is G, Calcisol is L, Gypsisol is L. In this study, the degree of membership between the soils studied was highlighted using 9 physical variables (three for each soil) and 3 linguistic variables (Little, Medium, and Great). (1) In total, 171 inference rules that represent all diagnostic criteria combinations were developed by the following relation: 3 ×∁13 + 3 ×∁ 2 3 +∁ 3 3􏼐 􏼑 × 9􏼐 􏼑 (1) where ∁ is combination. 2.2.3. Defuzzification. Inference methods provide mem- bership function μres(y) for the output variable “y” (Solonchak, Calcisol, and Gypsisol). Defuzzification is the transformation of this fuzzy information into measured information. A centroid (Z) method was employed [38] (Figure 5). +e expression of Z is given by the following equation: Z � 􏽒 D y × μres(y) × dy􏼐 􏼑 􏽒 D μres(y) × dy􏼐 􏼑 ⎛⎝ ⎞⎠. (2) In this study, Z represents the Solonchak indice, Calcisol indice, or Gypsisol indice obtained by MFIS. +e correlations between indices of soil obtained by MFIS and all soil classification criteria considered by WRB (ECe, calcium carbonate equivalent, gypsum content, and thickness of diagnostic horizons) were conducted. From the 171 rules estimated, only 21 inference rules were selected under our conditions. +e selection of these 21 inference rules was based on high significant correlation (P<0.05) between the different Solonchaks (Is), Calcisol (Ic), and Gypsisol (Ig) indices and WRB diagnostic criteria (except for 1 0.8 0.6 0.4 0.2 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 D eg re e of m em be rs hi p 0 0 1 1 1 Little value Medium value Great value Calcisol indices Gypsisol indices Solonchak indices Figure 4: Membership function of the output variables. 1 µres (y) 0 Centroid (z) (Indices of Solonchak) Output (y) (Solonchak) Figure 5: A method of the centroid used for defuzzification. Applied and Environmental Soil Science 5 diagnostic horizon thickness criteria). Finally, the key rules obtained are as follows: Rule 1: IF ECe is G AND (ECe × E) is G AND E is G AND calcium carbonate equivalent is L AND SC is L AND gypsum content is L (E × gypsum) is L THEN Solonchak is G, Calcisol is L, Gypsisol is L. Rule 2: IF ECe is L AND (ECe × E) is L AND E is L AND calcium carbonate equivalent is G AND SC is G AND gypsum content is L (E × gypsum) is L THEN Solonchak is L, Calcisol is G, Gypsisol is L. Rule 3: IF ECe is M AND (ECe × E) is M AND E is M AND calcium carbonate equivalent is G AND SC is G AND gypsum content is L (E × gypsum) is L THEN Solonchak is M, Calcisol is G, Gypsisol is L. Rule 4: IF ECe is L AND (ECe × E) is L AND E is L AND calcium carbonate equivalent is M AND SC is M AND gypsum content is L (E × gypsum) is L THEN Solonchak is L, Calcisol is M, Gypsisol is L. Rule 5: If ECe is L AND (ECe × E) is L AND E is L AND calcium carbonate equivalent is L AND SC is L AND gypsum content is G (E × gypsum)is G THEN Solonchak is L, Calcisol is L, Gypsisol is G. +e same procedure is used, as mentioned above, for the rest of the rules. +e method of min-max [25, 39, 40] was used to cal- culate the fuzzy inference. A weighting coefficient (Wi) is assigned to each inference rule. +is coefficient depends on the structure of the rule, that is to say the combination of OR and AND. AND is used for the min operator and OR for the max operator. +e weighting coefficient is used as a constant clipping of the output membership function. 2.3. Interpretation of Indices. Classification by MFIS is in favor of the higher indices. However, when the indices have the same value, it means that soils have the same degree of membership. +erefore, the soil is considered intergrade. +us, we can interpret the evidence as follows: (i) If Is>Ic and Is>Ig, then the soil is classified Solonchak. (ii) If Ic>Is and Ic>Ig, then the soil is classified Calcisol. (iii) If Ig>Is and Ig>Ic, then the soil is classified Gypsisol. (iv) If Is � Ic � Ig, then the soil is classified intergrade Solonchak-Calcisol-Gypsisol. (v) If Is � Ic and Ig<Ic and Ig<Is, then the soil is classified as an intergrade Solonchak-Calcisol soil. (vi) If Is � Ig and Ic<Gypsisol and Ic<Is, then the soil is classified as an intergrade Solonchak-Gypsisol soil. (vii) If Ic � Ig and Is<Ic and Is<Ig, then the soil is classified as an intergrade Calcisol-Gypsisol soil. 3. Results and Discussion Data analyses showed some differences between the three calculated indices. +ese data varied from 0.15 to 0.53 for Is, from 0.13 to 0.50 for Ic, and from 0.14 to 51 for Ig (Table 3). In general, Solonchaks under our study were more related to Calcisols (0.31 versus 0.25, resp.) than to Gypsisols (0.31 versus 0.19, resp.). +ese results suggest that the degree of membership of Solonchaks with Calcisols was more im- portant compared to the degree of membership of Solon- chaks with Gypsisols. 41% of the indices are assigned to Solonchaks, 33% to Calcisols, and 25% to Gypsisols as shown in Figure 6. +is result revealed that the soils studied are dominated by Solonchaks. +e high similitude between Solonchaks and Calcisols suggests that soils studied could be classified as Calcisols. +e results illustrated in Figures 6 and 7 show the following facts: (i) Solonchaks are the most dominant followed by the Calcisols. (ii) Solonchaks similar to Gypsisols are represented only by soil 89 with an index of 0.5. Soil 89 is qualified in WRB as Gypsic Solonchak. (iii) Only soils 39, 107, and 138 simultaneously exhibit the same degree of similarity with Solonchaks, Calcisols, and Gypsisols because their indices are 0.21, 0.18, and 0.15, respectively. +erefore, soils 39, 107, and 138 are qualified in WRB as Gypsic Calcic Solonchak. Table 3: Statistical parameters of the indices of three soils. Soils Minimum Maximum Mean SD Is 0.15 0.53 0.31 0.12 Ic 0.13 0.50 0.25 0.1 Ig 0.14 0.51 0.19 0.06 33 41 25 0 5 10 15 20 25 30 35 40 45 Soil groups In di ce s ( % ) Calcisol Solonchak Gypsisol Figure 6: Histogram of frequency indices. 6 Applied and Environmental Soil Science (iv) In contrast, soil 78 is classified Gypsisol by the fuzzy classification unlike WRB. We explain this differ- ences that the soil 78 is very rich in gypsum (58% by mass) (Table 4). +is soil is qualified by WRB as Gypsic Solonchak. Soils 15 and 126 have a higher degree of similarity to Calcisols than to Solonchaks (Figure 7). +erefore, these soils were classified by MFIS as Calcisols and not as Sol- onchaks. Moreover, the soils 15 and 126 are very rich in calcium carbonate equivalent (67%), rich in SC, with values ranging from 20% to 25% (by volume) for soils 15 and 126, respectively (Table 4). +ese 2 soils are qualified by WRB as Calcic Solonchak. +e difference observed between the WRB and MFIS was due to fact that the threshold values and priority order of classification were not considered by MFIS. According to the overall trend of the three curves (Figure 7), we concluded that the majority of Gypsisols was affected by values below 0.2. +e indices between 0.2 and 0.4 affect soils that have almost the same dominance between Solonchaks and Calcisols. +ese reveal the presence of a dominant overlap between Solonchaks and Calcisols compared to Solonchaks and Gypsisol. Index values above 0.4 represent essentially Solonchaks. +erefore, we can al- locate indices (I) obtained by MFIS into three groups to determine the frequency of the level of soil membership studied within each group, as shown in the following breakdown: (i) Group 1 (low indices): I<0.2 (ii) Group 2 (average indices): 0.2<I≤0.4 (iii) Group 3 (high indices): I>0.4 3.1. Membership Degree between the Soils Studied. Figure 8 showed that 50% of Gypsisols, 32% of Calcisols, and 19% of Solonchaks in the study area shared the group of low indices (I<0.2). +is result means that in this group, studied Solonchaks have a low degree of membership with Gypsisols and a relatively higher degree of membership with Calcisols. Similarly, some Solonchaks of this group have simulta- neously the same degree of membership with Calcisols and Gypsisols. +e Majority of the soils of group 1 are qualified by WRB as Gypsic Solonchak, and a small proportion of these soils are qualified as Gypsic Calcic Solonchak. In group 2 (0.2<I≤0.4), 43% of the average indices are assigned to Solonchaks, 39% to Calcisols, and 18% to Gypsisols. According to WRB, the qualifier Calcic is predominant for this group comparing the qualifier Gypsic. +is result suggests that Solonchaks, which are also dominant in this group, have a higher degree of membership with Calcisols than with Gypsisols. In group 3 (I>0.4), 67% of the indices are assigned to Solonchaks against 26% and 6% to Calcisols and Gypsisols, respectively. +is result means S15 S126 S39 S78 S89 S107 S138 0.0 0.1 0.2 0.3 0.4 0.5 0.6 1 8 15 22 29 36 43 50 57 64 71 78 85 92 99 10 6 11 3 12 0 12 7 13 4 14 1 14 8 15 5 16 2 16 9 17 6 18 3 19 0 In di ce s Soil Calcisol Solonchak Gypsisol S: soil Figure 7: Classification of soils obtained by the MFIS. Table 4: Characteristics of Solonchaks 15, 126, and 78. Parameters Soil 15 Soil 126 Soil 78 pH 8.6 7.5 7.6 ECe (dS/m) 23 15 15 Calcium carbonate equivalent (%) 67 67 22 Secondary carbonate (%) by volume 20 25 1 Gypsum (%) by mass 5 8 58 E (cm) 70 40 100 Note. E: thickness of the diagnostic horizons; ECe: electrical conductivity. Applied and Environmental Soil Science 7 that Solonchaks clearly dominate this group. It also suggests that, compared to groups 1 and 2, the degrees of mem- bership between soils (Solonchaks with both Calcisols and Gypsisols) in group 3 were low, and there are some Sol- onchaks that have no membership with either Calcisols or with Gypsisols. On the other hand, the proportions of the soil qualifiers in group 3 are distributed in the following way: Calcic (80%), Gypsic (12%), Gypsic Calcic Solonchak (3%), and 5% of Solonchaks, which are neither Calcic nor Gypsic. Overall, these results showed that the studied soils were dominated by Solonchaks; 74% of these Solonchaks are Calcic Solonchak. +e degree of membership of Solonchaks to Calcisols or Gypsisols was different depending on the group considered. +e degree of membership between Solonchaks and Calcisols is stronger than that between Solonchaks and Gypsisols. +us, Solonchaks with a higher degree of membership with Calcisols are qualified by WRB as Calcic Solonchaks. +e data analysis of each group of indices (Figure 8) was computed to determine Solonchaks, Calcisols, and Gypsisols frequencies and all possible intergrade soils. Figure 9 shows that Solonchak frequency increased from 39% in group 1 to 64% in group 2 and to 78% in group 3. Both Calcisols and Gypsisols were only detected in group 3 (3% and less than 2%, resp.). In each group, the soils which are neither Sol- onchaks, nor Gypsisols, nor Calcisols may be considered as intergrade soils with Solonchaks. As conclusion, the MFIS- based algorithms could estimate the number of soil in- tergrades and the degree of membership values (indices) of these soils with more precision [12–25] compared to WRB soil classification. Figure 9 shows three types of intergrade soils, Solonchaks- Calcisols, Solonchaks-Gypsisols, and Solonchaks-Calcisols- Gypsisols. +ese intergrade soils represent 60% of Solon- chaks in group 1 with a clear predominance of Solonchaks- Calcisols (45%) followed by Solonchaks-Gypsisols (11%) and Solonchaks-Calcisols-Gypsisols (3%). +ese intergrade soils represent only 34% of soils in group 2 and 16% in group 3. +e use of MFIS revealed that Solonchaks previously classified by WRB have strong similarities with Calcisols and Gypsisols. According to MFIS, Solonchaks have a higher degree of membership with Calcisols (29%) (Figure 10) than with Gypsisols (5%), and only 1% was Solonchaks-Calcisols- Gypsisols. Consequently, the classification of Solonchaks intergrades will be Solonchaks-Calcisols, Solonchaks-Gypsisols, and Solonchaks-Calcisols-Gypsisols. MFIS also showed that 61% of soils were strictly Solonchaks and only 1% and 0.5% were strictly Calcisols and Gypsisols, respectively. Based on WRB, from the 61% Solonchaks detected by MFIS, only 26% were classified as Solonchaks, 43% as Calcic Solonchaks, 3% as Gypsic Solonchaks, and 1% as Gypsic Calcic Solonchaks. 3.2. Correlation between the Indices of Soil Obtained by MFIS and Diagnostic Criteria of WRB. +e correlation data ana- lyses were conducted to determine the relationship between the indices obtained by MFIS, and its diagnostic criteria are defined by WRB (ECe, calcium carbonate equivalent, SC, gypsum, thickness of horizon (E), (E × ECe), (E × gypsum)). +e statistical parameters of these relationships presented in Table 5 showed that all correlations were positive and significant (0.49<r<0.77; P<0.05) except for (E) (0.01<r<0.06; P>0.05). Solonchaks indices (Is) presented significant and positive correlation with ECe (0.76), and Gypsisols indices (Ig) showed high correlation with gypsum content (0.70), while Calcisols indices (Ic) soil presented significant correlation with calcium carbonate equivalent and SC (0.77 and 0.70, resp.). +e majority of WRB di- agnostic criteria were highly correlated with the indices obtained by MFIS. +ese results suggest that MFIS gives the same soil classification as WBR; however, its application provides more precision concerning the degree of mem- bership values between soils (reference group soil). 32 39 26 19 43 67 50 18 6 0 10 20 30 40 50 60 70 80 ≤0.2 0.2–0.4 0.4–0.6 In di ce s ( % ) Groups of indices Calcisol Solonchak Gypsisol Figure 8: Histogram of frequencies of groups of indices. 3 64 78 32 10 5 0 10 20 30 40 50 60 70 80 90 Group 1 Group 2 Group 3 (% ) Calcisol Solonchak Gypsisol 39 45 11 3 <2 <2 Solonchak-Calcisol Solonchak-Gypsisol Solonchak-Gypsisol-Calcisol Figure 9: Histogram of soil group frequencies and their intergrades. 8 Applied and Environmental Soil Science Data analyses showed that the relationships between the three soil indices were positive and significant (P<0.05) (Table 6). +e relationship Is × Ic (r � 0.7) is stronger than Is × Ig (r � 0.52) and Ic × Ig (r � 0.32). +e high correlation between Is and both Ic and Ig suggests the presence of some intergrade soils under our conditions (Solonchak-Calcisol, Solonchak-Gypsisol, and Solonchak- Calcisol-Gypsisol) (Figure 10). +erefore, Solonchaks studied are actually a mixture of Solonchaks, Calcisols, and Gypsisols. However, the results showed that the indices assigned to Solonchaks were the most dominant (about 41%) followed by those assigned to Calcisols (33%) and Gypsisols (25%). In general, Solonchaks presented the highest indices, followed by Calcisols and Gypsisols. In total, Solonchaks have a higher degree of membership with Calcisols (29% of soil or Solonchak-Calcisol) than with Gypsisols (5%). Data analyses also confirmed these results by the high correlation between Solonchaks and Calcisols (r � 0.7; P<0.05) and between Solonchaks and Gypsisols (r � 0.52; P<0.05). Solonchaks-Calcisols (29%) intergrade soils were the most important in the north of Algeria compared to Solonchaks-Gypsisols (5%). +ese results confirm the conclusion of Hughes et al. [26] who showed that fuzzy logic allows the determination of intergrade groups. Also, Viscarra Rossel et al. [41] use the fuzzy ap- proach to provide information on the group of soil overlaps. MFIS showed that the degree of overlapping membership of these Solonchaks-Gypsisols-Calcisols intergrades was poorly represented (<2%). Similarly, it was found that 1% and 0.5% of 194 Solonchaks classified by WRB are respectively rec- ognized as Calcisols and Gypsisols by MFIS. +is small difference between the two classification systems is due to the fact that the threshold values of diagnostic criteria de- fined by conventional classifications would not be suitable for soil that is considered as a continuum system [9]. Consequently, significant information is lost [42], especially for both taxonomic fragmentation and soil mapping. However, fuzzy classification is continuous and numerical [12] that use the linguistic variables and Gaussian mem- bership functions. +ese results revealed that the two used systems (WRB and MFIS) provide the same classification (predominance of the Solonchak group). Based on fuzzy logic, the soil pre- viously classified by WRB as Calcic Solonchak has high degree of membership to Calcisols, and some of these Calcic Solonchaks are now classified by MFIS as Calcisol. Solon- chaks qualified previously by WRB as Gypsic Solonchak has a high degree of membership to Gypsisol, and some of these soils were classified by MFIS as Gypsisol. +e differences noted between the two soil system classifications are at- tributed to the fact that the WRB depends on the order of priority and weights attributed to diagnostic criteria. MFIS exhibited more precision concerning intergrade soils and degree of membership compared to WRB. +is precision is very useful in soil management practices and land evaluation systems [23–43]. 4. Conclusion +e soil classification by MFIS of 194 profiles previously classified as Solonchaks (Calcic Solonchak, Gypsic Solon- chak, and Gypsic Calcic Solonchak) by WRB revealed 6 different soil groups represented by Solonchaks, Solonchaks- Calcisols intergrades, Solonchaks-Gypsisols intergrades, Solonchaks-Calcisols-Gypsisols intergrades, Calcisols, and Gypsisols. In addition, this study showed that Is, Ic, and Ig were highly correlated with almost diagnostic criteria established by WRB except for horizon thicknesses. More- over, the correlation between Is and Ic (r � 0.7) was more important than Is and Ig (r � 0.52) and Ic and Ig (r � 0.32). Table 5: Correlations between indices of Solonchaks (Is), Calcisols (Ic), and Gypsisols (Ig) and different diagnostic criteria. Relations df r R2 Is, ECe 192 0.76∗ 0.58 Is, E 192 0.01 0.0001 Is, (E × ECe) 192 0.49∗ 0.24 Ig, gypsum 192 0.7∗ 0.49 Ig, E 192 0.06 0.04 Ig, (E × gypsum) 192 0.6∗ 0.36 Ic, E 192 0.04 0.002 Ic, SC 192 0.7∗ 0.5 Ic, calcium carbonate equivalent (CE) 192 0.77∗ 0.6 Note. ∗Significant at probability P<0.05; r: coefficient of correlation; R: coefficient of determination; df: degree of freedom. Table 6: Relations between indices of Solonchaks, Calcisols, and Gypsisols Relations df r R2 Is, Ic 192 0.7∗ 0.46 Is, Ig 192 0.52∗ 0.27 Ig, Ic 192 0.32∗ 0.1 Note. ∗Significant at probability P<0.05. 1 61 0.5 29 5 1 0 10 20 30 40 50 60 70 (% ) Calcisol Solonchak Gypsisol Solonchak-Calcisol Solonchak-Gypsisol Solonchak-Gypsisol-Calcisol Figure 10: Histogram of soil group frequencies and their intergrades. Applied and Environmental Soil Science 9 +ese relationships between indices suggest the presence of intergrade soils. On the other hand, these results confirm that Solonchaks-Calcisols intergrade is more dominant than Solonchaks-Gypsisols intergrade, as previously reported by Halitim [44] and Djili [30]. Our results showed that soil groups determined by WRB are analogous to those de- termined by MFIS. However, the application of MFIS pro- vides us the degree of membership between all these soils and their intergrades and takes into account the continuous complex nature of soil. As general conclusion, MFIS im- proved soil classification by using the degree of membership. +erefore, fuzzy logic could be considered as the basic tool for both classification and soil mapping and an undeniable support in precision agriculture. +e application of the MFIS to other soils of the world is possible because of the flexibility of inference rules. Conflicts of Interest +e authors declare that they have no conflicts of interest. References [1] T. Zádorová and V. Penı́žek, “Problems in correlation of Czech national soil classification and World Reference Base 2006,” Geoderma, vol. 167-168, pp. 54–60, 2011. [2] X. Z. Shi, E. D. Warner, and H. J. Wang, “Cross-reference for relating genetic soil classification of China with WRB at different scales,” Geoderma, vol. 155, no. 3-4, pp. 344–350, 2010. [3] A. E. Hartemink, “+e use of soil classification in journal papers between 1975 and 2014,” Geoderma Regional, vol. 5, pp. 127–139, 2015. [4] D. Baize and M. C. Girard, Référentiel Pédologique, Edition Quae, Versailles, France, 2008. [5] Commonwealth Scientific and Industrial Research Organi- sation, Australian Classification System, August 2015, http:// www.clw.csiro.au/aclep/asc_re_on_line/soilhome.htm. [6] Soil Classification Working Group, 6e Canadian System of Soil Classification, Agriculture and Agri-Food, Ottawa, ON, Canada, 1998. [7] National Institute of Agronomic Research, Référentiel Pédologique, INRA, Paris, France, 1995. [8] FAO, Système Universel de Classification des sols, 2017, http:// www.fao.org/soils-portal/soil-survey/soil-classification/universal- soilclassification/en/. [9] P. Duchaufour, “Réflexions sur les classifications des sols,” Etude et Gestion des Sols, vol. 5, pp. 201–205, 1998. [10] J. H. Rayner, “Classification of soils by numerical methods,” Journal of Soil Science, vol. 17, no. 1, pp. 79–92, 1966. [11] A. W. Moore, J. S. Russell, and W. T. Ward, “Numerical analysis of soils: a comparison of three soil profile models with field classification,” Journal of Soil Science, vol. 23, no. 2, pp. 193–209, 1972. [12] A. McBratney and I. O. A. Odeh, “Application of fuzzy sets in soil science: fuzzy logic, fuzzy measurements and fuzzy de- cisions,” Geoderma, vol. 77, no. 2–4, pp. 85–113, 1997. [13] FAO, Les Systèmes Numériques, 2017, http://www.fao.org/ soils-portal/soil-survey/soil-classification/numerical-systems/ en/. [14] IUSS Working Group WRB, “World reference base for soil resources. International Soil Classification System for Naming Soils and Creating Legends for Soil Maps,” World Soil Re- sources Reports, no. 106, FAO, Rome, Italy, 2014. [15] A. Jones, L. Montanarella, and R. Jones, Soil Atlas of Europe, Joint Research Centre, Ispra, Italy, 2005. [16] Soil Survey Staff, Soil Taxonomy, USDA National Resources Conservation. Services, Washington, DC, USA, 2nd edition, 1999. [17] I. Esfandiarpour, M. H. Salehi, and A. Karimi, “Correlation between Soil Taxonomy and World Reference Base for Soil Resources in classifying calcareous soils: (a case study of arid and semi-arid regions of Iran),” Geoderma, vol. 197-198, pp. 126–136, 2013. [18] N. Toomanian and A. Jalalian Karimian, “Application of the WRB (FAO) and US taxonomy systems to gypsiferous soils in Northwest Isfahan, Iran,” Journal of Agricultural Science and Technology, vol. 5, pp. 51–66, 2003. [19] A. Mojiri, A. Jalalian, and N. Honarjoo, “Comparison between keys to soil taxonomy and WRB to classification of soils in Segzi plain, Iran,” Journal Applied Science, vol. 11, no. 3, pp. 579–583, 2011. [20] M. Jamagne, “Compte rendu succinct des travaux. 16eme Congre Mondial de la Science du Sol, Montpellier, 20 au 26 Août 1998,” Etude et Gestion des Sols, vol. 6, pp. 1–3, 1999. [21] J. V. Vliet, A. Hagen-Zanker, J. Hurkens, and H. V. Delden, “A fuzzy set approach to assess the predictive accuracy of land use simulations,” Ecological Modelling, vol. 261-262, pp. 32–42, 2013. [22] M. Elaalem, “A Comparison of parametric and fuzzy multi- criteria methods for evaluating land suitability for Olive in Jeffara Plain of Libya,” APCBEE Procedia, vol. 5, pp. 405–409, 2003. [23] A. Sharififar, H. Ghorbani, and F. Sarmadian, “Soil suitability evaluation for crop selection using fuzzy sets methodology,” Acta Agriculturae Slovenica, vol. 107, pp. 32–35, 2016. [24] H. J. Zimmermann, Fuzzy Set 6eory and its Applications, Kluwer Academic Publishers, London, UK, 4th edition, 2001. [25] L. A. Zadeh, “Fuzzy sets,” Journal of Information and Control, vol. 8, no. 3, pp. 338–353, 1965. [26] P. Hughes, A. McBratney, and B. Minasny, “End members, end points and extragrades in numerical soil classification,” Geoderma, vol. 226-227, pp. 365–375, 2014. [27] P. Verma, P. Singh, and K. George, “Uncertainty analysis of transport of water and pesticide in an unsaturated layered soil profile using fuzzy set theory,” Applied Mathematical Mod- elling, vol. 33, no. 2, pp. 770–782, 2009. [28] M. Fajardo, A. McBratney, and B. Whelan, “Fuzzy clustering of Vis–NIR spectra for the objective recognition of soil morphological horizons in soil profiles,” Geoderma, vol. 263, pp. 244–253, 2015. [29] E. H. Mamdani, “Application of fuzzy logic to approximate reasoning using linguistic synthesis,” IEEE Transactions on Computers, vol. 26, pp. 1182–1191, 1977. [30] K. Djili, “Contribution à la connaissance des sols du Nord d’Algérie. Création d’une banque de données informatisées et utilisation d’un système d’information géographique pour la spatialisation et la valorisation des données pédologiques,” +èse doctorat, p. 227, ENSA, El-Harrach, Algeria, 2000. [31] A. Rahmouni, “Morphologie et propriétés des gypsisols références du Hodna,” Magister +esis, p. 188, Superior National School of Agronomics, Oued Smar, Algeria, 2010. [32] United States Salinity Laboratory Staf, Diagnosis and Im- provement of Saline and Alkali Soils, Agriculture Handbook, no. 60, Govt. Printing Office, Washington, DC, USA, 1954. 10 Applied and Environmental Soil Science http://www.clw.csiro.au/aclep/asc_re_on_line/soilhome.htm http://www.clw.csiro.au/aclep/asc_re_on_line/soilhome.htm http://www.fao.org/soils-portal/soil-survey/soil-classification/universal-soilclassification/en/ http://www.fao.org/soils-portal/soil-survey/soil-classification/universal-soilclassification/en/ http://www.fao.org/soils-portal/soil-survey/soil-classification/universal-soilclassification/en/ http://www.fao.org/soils-portal/soil-survey/soil-classification/numerical-systems/en/ http://www.fao.org/soils-portal/soil-survey/soil-classification/numerical-systems/en/ http://www.fao.org/soils-portal/soil-survey/soil-classification/numerical-systems/en/ [33] F. Carré and M. Jacobson, “Numerical classification of soil profile data using distance metrics,” Geoderma, vol.148, no. 3- 4, pp. 336–345, 2009. [34] E. H. Mamdani, “Application of fuzzy algorithms for simple dynamic plants,” Proceedings of the IEEE, vol. 121, no. 12, pp. 1585–1588, 1974. [35] A. Uyumaz, A. Altunkaynak, and M. Ozger, “Fuzzy logic model for equilibrium scour downstream of a Dams vertical gate,” Journal of Hydraulic Engineering, ASCE, vol.132, no.10, pp. 1069–1075, 2006. [36] M. Özger, “Comparison of fuzzy inference systems for stream flow prediction,” Hydraulic Science Journal, vol. 54, no. 2, pp. 261–273, 2009. [37] A. Ahumada, A. Altunkaynak, and A. Ashraf, “Fuzzy logic- based attenuation relationships of strong motion earthquake records,” Expert Systeme with Applications, vol. 42, no. 3, pp. 1287–1297, 2015. [38] J. T. Ross, Fuzzy Logic with Engineering Applications, McGraw-Hill, Inc., New York, NY, USA, 1995. [39] C. Negoita, Expert Systems and Fuzzy Systems, Benjamin Cummings, Redwood City, CA, USA, 1985. [40] G. Klir and T. Folger, Fuzzy Sets, Uncertainty and In- formation, Prentice-Hall, Englewood Cliffs, NJ, USA, 1988. [41] R. A. Viscarra Rossel, T. Behrens, and E. Ben-Dor, “A global spectral library to characterize the world’s soil,” Earth-Science Reviews, vol. 155, pp. 198–230, 2016. [42] A. Zhua, E. Lawerence, B. D. Band, D. +omas, and J. Nimlosd, “Automated soil inference under fuzzy logic,” Ecological Modelling, vol. 2, no. 90, pp. 123–145, 1996. [43] E. Van Ranst and H. Tang, “Fuzzy reasoning versus Boolean logic in land suitability assessment,” Malaysian Journal of Soil Science, vol. 3, pp. 39–58, 199. [44] A. Halitim, Sols des régions arides d’Algérie, OPU, Alger, Algeria, 1988. Applied and Environmental Soil Science 11 Hindawi www.hindawi.com Applied & Environmental Soil Science Volume 2018 Hindawi www.hindawi.com Volume 2018 Journal of Chemistry Archaea Hindawi www.hindawi.com Volume 2018 Forestry Research International Journal of Hindawi www.hindawi.com Volume 2018 Environmental and Public Health Journal of Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 Meteorology Advances in Ecology International Journal of Hindawi www.hindawi.com Volume 2018 Marine Biology Journal of Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 Chemistry Advances in Agronomy Hindawi www.hindawi.com Volume 2018 International Journal of Hindawi www.hindawi.com Volume 2018 Advances in Virolog y Hindawi www.hindawi.com Volume 2018 International Journal of Geophysics Hindawi www.hindawi.com Volume 2018 Geological Research Journal of Hindawi www.hindawi.com Volume 2018 Public Health Advances in Biodiversity International Journal of Hindawi www.hindawi.com Volume 2018 Scienti�ca Hindawi www.hindawi.com Volume 2018 Botany Journal of Hindawi www.hindawi.com Volume 2018 Hindawi Publishing Corporation http://www.hindawi.com Volume 2013 Hindawi www.hindawi.com The Scientific World Journal Volume 2018 Agriculture Advances in Hindawi www.hindawi.com Volume 2018 Submit your manuscripts at www.hindawi.com https://www.hindawi.com/journals/aess/ https://www.hindawi.com/journals/jchem/ https://www.hindawi.com/journals/archaea/ https://www.hindawi.com/journals/ijfr/ https://www.hindawi.com/journals/jeph/ https://www.hindawi.com/journals/amete/ https://www.hindawi.com/journals/ijecol/ https://www.hindawi.com/journals/jmb/ https://www.hindawi.com/journals/ac/ https://www.hindawi.com/journals/ija/ https://www.hindawi.com/journals/av/ https://www.hindawi.com/journals/ijge/ https://www.hindawi.com/journals/jgr/ https://www.hindawi.com/journals/aph/ https://www.hindawi.com/journals/ijbd/ https://www.hindawi.com/journals/scientifica/ https://www.hindawi.com/journals/jb/ https://www.hindawi.com/journals/tswj/ https://www.hindawi.com/journals/aag/ https://www.hindawi.com/ https://www.hindawi.com/ 
work_p4nyzm7uz5fwrbumfajdtyk34i	----	MnSymbol7 Kent Academic Repository Full text document (pdf) Copyright & reuse Content in the Kent Academic Repository is made available for research purposes. Unless otherwise stated all content is protected by copyright and in the absence of an open licence (eg Creative Commons), permissions for further reuse of content should be sought from the publisher, author or other copyright holder. Versions of research The version in the Kent Academic Repository may differ from the final published version. Users are advised to check http://kar.kent.ac.uk for the status of the paper. Users should always cite the published version of record. Enquiries For any further enquiries regarding the licence status of this document, please contact: researchsupport@kent.ac.uk If you believe this document infringes copyright then please contact the KAR admin team with the take-down information provided at http://kar.kent.ac.uk/contact.html Citation for published version Kampouridis, Michael and Otero, Fernando E.B. and Kampouridis, Michael (2016) Evolving Trading Strategies Using Directional Changes. Expert Systems with Applications, 73 . pp. 145-160. ISSN 0957-4174. DOI https://doi.org/10.1016/j.eswa.2016.12.032 Link to record in KAR http://kar.kent.ac.uk/59758/ Document Version Author's Accepted Manuscript Evolving Trading Strategies Using Directional Changes Michael Kampouridisa,∗, Fernando Oteroa aSchool of Computing, University of Kent, UK Abstract The majority of forecasting methods use a physical time scale for studying price fluctuations of financial markets, making the flow of physical time discontinuous. Therefore, using a physical time scale may expose companies to risks, due to ignorance of some significant activities. In this paper, an alternative and original approach is explored to capture important activities in the market. The main idea is to use an event-based time scale based on a new way of summarising data, called Directional Changes. Combined with a genetic algorithm, the proposed approach aims to find a trading strategy that maximises profitability in foreign exchange markets. In order to evaluate its efficiency and robustness, we run rigorous experiments on 255 datasets from six different currency pairs, consisting of intra-day data from the foreign exchange spot market. The results from these experiments indicate that our proposed approach is able to generate new and profitable trading strategies, significantly outperforming other traditional types of trading strategies, such as technical analysis and buy and hold. Keywords: directional changes, financial forecasting, algorithmic trading, genetic algorithm 1. Introduction The global financial system, recently rocked by the financial crisis, is open 24 hours a day, 7 days a week and can be defined as a complex network of interacting agents (e.g., corporations, retail traders). With an average daily turnover of 3–4 trillion USD (International Monetary Fund, 2009) and price changes nearly every second, its activity varies at different times of a day and reacts on the announcement of political or economic news. The majority of traditional methods to observe such price fluctuations in financial time series are based on phys- ical time change. For example, what researchers and practitioners tend to do is to use snapshots of the market, taken at fixed intervals; they first decide how often to sample the data, and then they take snapshots at the chosen fre- quency. Therefore, these snapshots create an interval-based summary—e.g. daily closing prices or minute-by-minute summaries. However, important price movements (and thus potential profit) might be lost due to the creation of these artificial price summaries. For example, if we are using daily closing price summaries we would not be able to observe the 6 May 2010 Flash Crash, which was a United States trillion-dollar stock market crash that lasted for approximately 36 minutes.1 Directional Changes (DC) is based on the idea that an event-based system can capture significant points in price movements that the traditional physical time methods cannot. Hence, instead of looking the market from an interval- based perspective, DC record the key events in the market (e.g., changes in the stock price by a pre-specified percent- age) and summarise the data based on these events, moving away from a physical-time view to an event-based-time view. Under this new paradigm, a threshold θ is defined, usually expressed by a percentage of the price. The market is then fragmented and summarised into upward and downward trends. Different thresholds produce different price summaries. Thus, the directional changes paradigm focuses on the size of price change, while time is the varying factor; whereas in the physical-time paradigm, time was fixed (e.g. daily closing prices). This new concept provides ∗Corresponding author. Tel: +44 1634 8888 37 Email addresses: M.Kampouridis@kent.ac.uk (Michael Kampouridis), F.E.B.Otero@kent.ac.uk (Fernando Otero) 1http://blogs.wsj.com/marketbeat/2010/05/11/nasdaq-heres-our-timeline-of-the-flash-crash/ Last access: 12 September 2016. Preprint submitted to Expert Systems with Applications December 23, 2016 traders with new perspectives to price movements, and allows them to focus on those key points than an important event took place, blurring out other price details which could be considered irrelevant, or even noise. The first works to use the concept of directional changes were proposed in Olsen et al. (1997) and Glattfelder et al. (2011). In these works, new empirical scaling laws in foreign exchange data series were discovered. These scaling laws aimed to establish mathematical relationships among price moves, duration and frequency. Then, di- rectional changes and the scaling laws from the above works were used to develop new trading models in Dupuis & Olsen (2012). However, those models were not used for any financial forecasting purposes and were only used to derive statistics from potential trading. Furthermore, Aloud et al. (2012) demonstrated the effectiveness of directional changes in capturing periodic market activities. In addition, Gypteau et al. (2015) presented an approach to forecasting the daily closing price of financial markets by combing directional changes and genetic programming. Lastly, Tsang et al. (2016) introduced new trading indicators for profiling markets under directional changes. As we can observe, the majority of the above works have focused on theoretical aspects of directional changes—e.g. establishing mathemat- ical relationships and developing new indicators. Only Dupuis & Olsen (2012) and Gypteau et al. (2015) attempted to generate trading strategies based on the DC concept. However, Dupuis & Olsen (2012) did not take advantage of the combined knowledge that can exist by using multiple DC thresholds to generate different event-based series; in addition, Gypteau et al. (2015) only tested their approach on 4 datasets on daily closing prices. In this paper our aim is to offer a more complete analysis on the directional changes paradigm from a financial fore- casting perspective. We run extensive tests on intraday data from six currency pairs from the foreign exchange (FX) market: yearly tick data from GBP/JPY, and yearly 10-minute interval data from EUR/GBP, EUR/USD, EUR/JPY, GBP/CHF, and GBP/USD. In total, we run tests on 255 different datasets. In terms of DC, we present two novel types of trading strategies: a single-threshold DC strategy, and a multi-threshold DC strategy. The former uses a sin- gle threshold to generate event-based series. The multi-threshold strategy combines different thresholds, where each threshold generates a different event-based series; then information from each series is aggregated to form a more informed trading strategy. In addition, we use a genetic algorithm (GA), which is a bio-inspired algorithm mimicking an evolutionary process, to optimise the parameters of the multi-threshold strategy. Such evolutionary algorithms have extensively been used in financial forecasting problems and have shown to be extremely effective (Kwon & Moon, 2007; Mani, 1996; Evans et al., 2013; Kampouridis & Otero, 2015). The GA-generated trading strategies are then compared against the single-threshold and the multi-threshold DC strategies. We test the GA-generated trading strategies with other financial benchmarks, such as buy and hold and technical analysis strategies. Overall, our goals in this work can be summarised as follows: (i) demonstrate that the paradigm of DC returns profitable strategies, (ii) provide evidence that the strategies optimised by the GA are more profitable than using standard DC strategies, and (iii) demonstrate that our GA generated strategies outperform classical physical-time based strategies, namely technical analysis and buy and hold. Lastly, it should be acknowledged that directional changes has similarities to the concept of the zigzag indicator (Azzini et al., 2010), which also focuses on an event-based approach, and to the concept of perceptually important point (PIP) identification (Chen & Chen, 2016), which preserves the salient points in a time series to reduce the number of data points. However, as we’ve mentioned above, the recent research on the DC field has led to the development on many new concepts, such as the scaling laws and new trading indicators that only exist under DC price summaries. Thus, in order to take advantage of all these new developments, we are focusing on the DC paradigm. The rest of this paper is organised as follows: Section 2 presents background information in the fields of financial forecasting, directional changes, and genetic algorithms. Then, Section 3 presents the proposed DC-derived trading strategies, and Section 4 discusses how we used the genetic algorithm to optimise the parameters of these strategies. In addition, Section 5 presents the experimental setup, and Section 6 presents and discusses the results. Finally, Section 7 concludes the paper and discusses directions for future work. 2. Background Financial forecasting is a vital area in computational finance (Tsang & Martinez-Jaramillo, 2004). The end goal of financial forecasting is deriving a trading strategy, which makes a recommendation whether to buy, hold or sell. There are numerous works that attempt to return profitable trading strategies—several examples can be found in Chen (2002); Binner et al. (2004); Hu et al. (2015); Jaisinghani (2016). 2 There are several different strategies for the purposes of trading in a financial market. A very common investment strategy, albeit a passive one, is buy and hold, and commonly acts as benchmark for trading algorithms. The principle of this strategy is based on the view that in the long run financial markets give a good rate of return to investors. Thus, in this strategy an investor buys an asset and holds it for a long time, without being concerned about short-term price movements. Then at the end of a given period, s/he sells the stock and potentially makes profit based on the price difference. In contrary to buy and hold, there is also the belief that it is profitable to take advantage of short-term price movements, as long as one can anticipate/predict them. Technical analysis is such a technique, and is discussed in Section 2.1. Then, we present background information on directional changes, a new way of summarising financial data that can lead to new types of trading strategies. This takes place in Section 2.2. Finally, Section 2.3 gives an overview of genetic algorithms, which is the technique used in this paper for optimising the use of directional changes parameters. 2.1. Technical analysis Technical analysis is a methodology for financial forecasting. This method assumes that patterns exist in historical price data and that these patterns will repeat themselves. Consequently, it is worth identifying these patterns, so that we can exploit them in the future and make profit. Several works exist that are using technical analysis—recent examples can be found in Cervelló-Royo et al. (2015); Chourmouziadis & Chatzoglou (2016). As part of technical analysis, several indicators are used. These technical analysis indicators are formulas that measure different aspects of a given financial dataset, such as trend, volatility and momentum. Below in Equations (1)-(6) we present 6 popular indicators that can be found in the literature (Martinez-Jaramillo, 2007; Allen & Kar- jalainen, 1999; Austin et al., 2004). Given a price time series [P(t), t ≥ 0], and a period of length L, these indicators are defined as follows: Moving Average (MA) MA(L,t) = P(t)− 1 L L ∑ i=1 P(t − i) 1 L L ∑ i=1 P(t − i) (1) MA allows traders to observe any changes in the trend of the prices of a stock. Typically, when a short-term MA (e.g., 12 days) goes above a long-term MA (e.g., 60 days), this indicates upward momentum. On the other hand, when a short-term MA goes below a long-term one, this indicates downward momentum. Trade Break Out (TBR) TBR(L,t) = P(t)− max{P(t − 1), . . . ,P(t − L)} max{P(t − 1), . . . ,P(t − L)} (2) In order to understand this indicator better, we first need to explain two terms: support and resistance. Support is the point where the price stops going down any further, whereas resistance is the point where the price does not go up any further. Technical analysis suggests that downward price trends tend to reverse at support points, whereas upward trends tend to reverse at resistance points. However, when these points are breached (breakout), perhaps because of some new information regarding the market, it is likely that the price will continue in the same direction. Hence, traders tend to observe these breakouts and when a stock goes above its point of resistance, they buy; when on the other hand the stock price goes below its point of support, traders sell. Filter (FLR) FLR(L,t) = P(t)− min{P(t − 1), . . . ,P(t − L)} min{P(t − 1), . . . ,P(t − L)} (3) 3 This indicator is used to indicate buy or sell actions, depending on whether the price movement goes in the opposite direction by a predefined percentage. For instance, if the price reverses from a downward trend and rises by a specific percentage from the low price that it was previously, then the trader would perform a ‘buy’ action. Volatility (Vol) Vol(L,t) = σ(P(t), . . . ,P(t − L + 1)) 1 L L ∑ i=1 P(t − i) (4) A period of an increasing volatility could indicate a reversal in the trend or strong downward trends. This would thus give an indication to a trader that s/he should be cautious. On the contrary, when there is a period of decreasing volatility, this indicates upward trends and traders should buy. Momentum (Mom) Mom(L,t) = P(t)− P(t − L) (5) When Mom is positive, this indicates an upward trend. If Mom starts decreasing, this could be an indication that there is going to be a reverse in the previously upwards trend, and hence the traders should be cautious. Finally, when Mom is negative, this indicates a downwards trend. Momentum Moving Average (MomMA) MomMA(L,t) = 1 L L ∑ i=1 Mom(L, t − i) (6) Finally, from Mom we can also calculate its MA, as shown in the above equation, which allows us to obtain summaries of the Momentum movements. While the above indicators can offer valuable information, normally a trader would use many of these indicators together, thus combine their recommendations. It is very common in the literature to use machine learning algorithms to combine technical information indicators (Chiang et al., 2016; Kim & Enke, 2016). In this work, we use EDDIE (Kampouridis & Tsang, 2012, 2010) as a baseline algorithm. EDDIE is a genetic programming (GP) (Koza, 1992) financial forecasting algorithm, which combines the different technical analysis indicators together, in order to form predictions. The advantage of combining technical analysis indicators is that their combined knowledge can lead to better predictions. We should also mention that EDDIE has been used over the years for different types of financial problems, such as stock price movement prediction (Kampouridis & Otero, 2015), arbitrage opportunities detection (Tsang et al., 2005), and market crash detection (Garcia-Almanza et al., 2013). EDDIE has thus extensively and effectively utilised technical analysis for its predictions, and for this reason we have chosen to use it as a benchmark of an algorithm using technical analysis. As EDDIE is a GP algorithm, its trading strategies are represented as trees. A sample tree of EDDIE is presented in Figure 1. As we can see, if the 20 days MA is less than or equal to 6.4, then the user is advised to buy; otherwise, the user is advised to consult another tree, which is located in the third branch (“else-branch”) of the tree. This tree checks if the 50 days MomMA is greater than 5.57; if it is, it advises to not-buy, otherwise to buy. Of course this is simply a sample tree; so additional/different technical analysis indicators could be used in other trees. In summary, what we have presented so far—namely buy and hold, technical analysis and EDDIE—all use in- formation derived from data that is based on physical-time. As we have explained, in this paper we propose using event-based price summaries for creating the trading strategies, based on the concept of directional changes. 4 If <= VarConstructor MA 20 6.64 Buy(1) If > VarConstructor MomMA 50 5.57 Not-Buy(0) Buy(1) Figure 1: Sample tree generated by EDDIE representing a trading strategy using technical indicators. 2.2. Directional changes The directional change (DC) approach is an alternative approach for summarising market price movements. A DC event is identified by a change in the price of a given financial instrument. This change is defined by a threshold value, which was in advance decided by the trader. Such an event can be either an upturn or a downturn event. After the confirmation of a DC event, an overshoot (OS) event follows. This OS event finishes once an opposite DC event takes place. The combination of a downturn event and a downward overshoot event represents a downward trend and, the combination of an upturn event and an upturn overshoot event represents an upturn trend. In other words, a downward trend is a period between a downturn event and the next upturn event and an upturn trend is a period between an upturn event and the next downturn event. Figure 2 presents an example of how a physical-time price curve is transformed to the so-called intrinsic time (Glattfelder et al., 2011) and dissected into DC and OS events. As we can observe, two different thresholds are used, and each threshold generates a different event series. Thus, each threshold produces a unique series of events. The idea behind the different thresholds is that each trader might consider different thresholds (price percentage changes) as significant. A smaller threshold creates a higher number of directional changes, while a higher thresholds produces fewer directional changes. Looking at the events generated by a threshold of θ = 0.01% (events connected via solid and dashed lines), we can observe that any price change less than this threshold is not considered a trend. On the other hand, when the price changes above that threshold, then the market is divided accordingly, to uptrends and downtrends. DC events are in solid lines, and OS events are in dashed lines. So an downturn DC event starts at Point A and lasts until Point B, when the downturn OS events starts. The downturn OS lasts until Point C, when there is a reverse in the trend, and an uptrend starts, which lasts until Point D. From Point D to E we are in an upturn OS event, and so on. As we mentioned, different thresholds generate different event series. Looking at theta = 0.018% (events con- nected via dotted and dot-dashed lines), we can observe that the events generated are different: a downward trend starts from A and lasts until B′, and the downward OS is from Point B′ until C. Then, from Point C until Point E there is an upward DC trend, and from E to E′ there’s an upward OS trend. Algorithm 1 presents the high-level pseudocode for generating directional changes events. It is important to note here that the confirmation of a change of a trend can only be confirmed retrospectively, i.e. only after the price has changed by the pre-specified DC threshold value θ. For example, under θ= 0.01% we can only confirm that we are in a upward trend from Point D onwards. Point D is thus called a confirmation point. Before Point D, the directional change had not been confirmed (i.e. the market price had not changed by the pre-specified threshold value), thus a trader summarising the data by the DC paradigm would continue believing we are in a downward trend, which started from Point A. Similarly, a trader using θ = 0.01% would continue considering being in a upward trend from Point D until the price has reversed by θ = 0.01%, which only takes place at the next confirmation point, i.e., Point F. So what becomes important here is to be able to anticipate the change of the trend as early as possible, i.e. before Points C and E have been reached. In addition, since different thresholds generate different event series, we hypothesise that the combined information from these series would lead to profitable trading strategies. 5 Figure 2: Directional changes for tick data for the GBP/JPY currency pair. The solid and dashed lines denote a set of events defined by a threshold θ = 0.01%, while the dotted and dot-dashed lines refer to events defined by a threshold θ = 0.018%. The solid and the dotted lines indicate the DC events, and the dashed and dot-dashed indicate the OS events. Under θ = 0.01%, the data is summarised as follows: Point A ↦ B (Downward directional change), Point B ↦ C (Downward overshoot event), Point C ↦ D (Upward directional change), Point D ↦ E (Upward overshoot event), Point E ↦ F (Downward directional change). Under θ= 0.018%, the data is summarised as follows: Point A ↦ B’ (Downward directional change), Point B’ ↦ C (Downward overshoot event), Point C ↦ E (Upward directional change), Point E ↦ E’ (Upward overshoot event). The advantage of this new way of summarising data is that it provides traders with new perspectives to price movements, and allows them to focus on those key points that an important event took place, blurring out other price details which could be considered irrelevant or even noise. Furthermore, DC have enabled researchers to discover new regularities in markets, which cannot be captured by the interval-based summaries (Glattfelder et al., 2011). Therefore, these new regularities give rise to new opportunities for traders, and also open a whole new area for research. One of the most interesting regularities that was discovered in Glattfelder et al. (2011) was the observation that a DC of threshold θ is on average followed by an OS event of the same threshold θ. At the same time, it was observed that if on average a DC takes t amount of physical time to complete, the OS event will take an amount of 2t. This observation is summarised in Figure 3, and was only made under DC-based price summaries, and not under phsycical- time summaries. Furthermore, this astonishing observation was made on all of the 13 different currency exchange rates that the authors of Glattfelder et al. (2011) experimented with. This thus leads us to further hypothesise that such statistical properties could lead to profitable strategies, if appropriately exploited, mainly because such properties are not well-known to traders yet. Therefore, the DC area is a rich research area that could potentially lead to significant discoveries. In this work, we will present new trading strategies based on the concept of directional changes. As part of the implementation of this trading strategy we will be using the scaling law presented above; we have also introduced several parameters into the system, which we present in detail in Section 3. Lastly, since a user/trader has to decide on which thresholds to use for generating DC events, a problem that arises is what are appropriate thresholds and how much weight we should give to the information provided by each threshold. We are thus faced with an optimisation problem, where one has to look into the space of different thresholds and focus on the most promising ones. One of the popular optimisation techniques is Genetic Algorithms, discussed next. 2.3. Genetic algorithms Genetic Algorithms (GAs) are bio-inspired algorithms that mimic an evolutionary process to find good solutions to optimisation problems (Godlberg, 1989; Mitchell, 1996; Michalewicz, 2002). GAs have several elements that allow them to perform a robust global search: (a) they work with a population of candidate solutions (individuals), rather than a single candidate solution; (b) individuals of the population are evaluated according to a fitness function, 6 Figure 3: An example of a scaling law presented in Glattfelder et al. (2011), which shows that (1) a DC event (solid line) of threshold θ is followed by an OS event (dotted line) of also threshold θ, and (2) the OS event lasts about the double amount of time that it took for the DC event to take place. Algorithm 1 Pseudocode for generating directional changes events (source: Aloud et al. (2012)). Require: Initialise variables (event is Upturn event, ph = pl = p(t0),∆xdc(Fixed)≥ 0, t dc 0 = t dc 1 = t os 0 = t os 1 = t0) 1: if event is Upturn Event then 2: if p(t)≤ ph ×(1 −∆xdc) then 3: event ← DownturnEvent 4: pl ← p(t) 5: tdc1 ← t // End time for a Downturn Event 6: tos0 ← t + 1 // Start time for a Downward Overshoot Event 7: else 8: if ph < p(t) then 9: ph ← p(t) 10: tdc0 ← t // Start time for Downturn Event 11: tos1 ← t − 1 // End time for an Upward Overshoot Event 12: end if 13: end if 14: else 15: if p(t)≤ pl ×(1 +∆xdc) then 16: event ← U pturnEvent 17: ph ← p(t) 18: tdc1 ← t // End time for a Upturn Event 19: tos0 ← t + 1 // Start time for an Upward Overshoot Event 20: else 21: if pl > p(t) then 22: pl ← p(t) 23: tdc0 ← t // Start time for Upnturn Event 24: tos1 ← t − 1 // End time for an Downward Overshoot Event 25: end if 26: end if 27: end if 7 which measures the quality of the candidate solution represented by an individual—the higher their quality, the more likely that their genetic material will be carried forward to the next population; (c) the solution space is explored using genetic operators, which generate new offspring individuals from individuals of the current population using a stochastic selection procedure based on fitness. Algorithm 2 presents a high-level pseudocode of a GA. The algorithm starts by creating a population of p candidate solutions, where p is referred to as population size. These are evaluated by a fitness function in order to determine their performance in solving the problem at hand. On each iteration (while loop), a new population is then generated by probabilistically selecting the fitter individuals from the current population. Some of the selected individuals undergo crossover and mutation, which introduce modification to explore the search space; other are carried forward without modifications. The new population replaces the old and the new individuals are evaluated. This process is repeated until a maximum number of generations is reached or the (near-)optimal solution is found, acting as a termination condition. Through this evolutionary process, GAs perform a robust global search in the space of candidate solutions, less likely to get trapped into local minima. Representation. Individuals in GAs are usually represented as a string of symbols, either binary or numeric values— the representation is dependant on the problem at hand. Figure 4 shows an illustration of a real-valued representation. Each position in the string is referred to as a gene and it represents a variable to be optimised. At the start of a GA, the population is initialised with random individuals: each gene is initialised with a random value in the domain of the variable. Genetic Operators. Genetic operators manipulate the genetic material of individuals (strings of symbols) to generate new offspring individuals, mimicking a mechanism of inheritance. For example, crossover create two new offspring solutions from two parent individuals by swapping portions of genetic material (or genes) from each parent. To illustrate, consider the uniform crossover operator in Figure 4. This operator combines genes sampled uniformly from two parents. In addition to crossover, GAs usually employ a mutation operator, which produces a new offspring individual from a single parent. In uniform mutation, small changes are introduced to a parent individual by choosing and modifying each gene at random. Both uniform crossover and uniform mutation are controlled by two probabilities rates: the first one is used to decided whether the selected individual will undergo crossover/mutation (pc and pm in Algorithm 2, respectively) or not; the second rate is used to decide if a gene is swapped/mutated or not. Figure 4 shows an illustration of both uniform crossover and uniform mutation operators. Elitism and Selection. Elitism is the process of allowing the best individuals (in terms of fitness) of the current population to be carried over to the next without modification. This guarantees that the solution fitness will not decrease from one generation to the next. The remaining individuals are subject to a probabilistic selection for inclusion in the next generation population. One of the most popular selection strategies is the tournament selection. In tournament selection, k individuals are selected at random, where k denotes the tournament size. The more fit individual of the selected subset is then selected for inclusion in the next population. 2.4. Summary In this section we started by discussing two popular trading techniques, namely buy and hold and technical anal- ysis. Both of these two methods will form our financial benchmarks during our experimental phase. In addition, we presented in detail the concept of directional changes, which our trading strategies are going to be based on. Lastly, we discussed what genetic algorithms are, as they are going to be our optimisation engine. In the next section, we present two new types of trading strategies, which are derived by directional changes. 3. DC-derived trading strategies In this section, we will present how we can use the DC paradigm to derive two different types of trading strategies. The first one is going to be based on a single DC threshold, and is going to be presented in Section 3.1. The second strategy is going to be based on multiple DC thresholds and is going to be presented in Section 3.2. 8 Algorithm 2 High-level pseudocode of a genetic algorithm. GA( p, Fitness, pc, pm) p: population size Fitness: determines the quality of solutions pc: crossover rate pm: mutation rate 1: Initialise population: P ← Generate p individuals (candidate solutions) at random 2: Evaluate: for each i in P, calculate Fitness(i) 3: while termination condition not satisfied do 4: Pg ← Create new population for generation g (a) Elitism: copy the r best individuals from P to Pg (b) Select: probabilistically select (p − r) individuals of P to add to Pg and • Perform crossover between a pair of selected individuals according to pc • Perform mutation on a selected individual according to pm 5: Update: P ← Pg 6: Evaluate: for each i in P, calculate Fitness(i) 7: end while 8: Return the individual with the highest fitness from P 0.7 0.3 0.5 0.4 0.2 0.1 0.9 0.6 0.3 0.1 0.8 0.9 0.4 0.2 0.7 0.6 0.7 0.1 0.8 0.4 0.4 0.1 0.9 0.6 0.5 0.1 0.4 0.2 0.1 0.7 0.2 0.8 0.5 0.1 0.7 0.2 0.6 0.7 0.2 0.1 0.3 0.3 0.5 0.9 0.2 0.2 0.7 0.6 Uniform crossover: Uniform mutation: Parent individuals Offspring Figure 4: Illustration of uniform crossover and uniform mutation operators. Individuals are represented as a string of real values. The dark positions (genes) in parent individuals are the ones that will be swapped/mutated to generate offspring individuals. 9 3.1. Single-threshold DC trading strategy As we have already discussed in Section 2.2, a physical-time price curve can be transformed to intrinsic time, where it’s dissected into DC and OS events. When a DC event is confirmed (either upturn or downturn), the OS period starts. It is worth re-iterating that we can only know the market has changed direction in hindsight; we only detect a DC event only after the DC confirmation point has been observed. After the DC confirmation has taken place, we are during an OS period, which lasts until there is a reverse in the direction, which is again only confirmed once we have reached the next confirmation point. Therefore, if one can act during the OS period and before the reverse of the trend, then this can lead to a profitable trading strategy. The idea is that if we are in an downtrend, we buy at the last point (ideally) of the OS event, which is the lowest observed value. When the trend then reverses and we are in an uptrend, we can then sell at a much higher price and make profit. The same principle applies for uptrend: sell as close as possible to the end of the OS event. To sum up, when there is a downtrend, we buy; when there is an uptrend, we sell. In order to make the above clearer, let us refer back to Figure 2. As we can observe, after the confirmation of the downward trend in Point B, a period of OS starts, which lasts until Point D, which is the next DC confirmation point, confirming that we are now in a upward trend. It is thus important to take a buy action during the OS event and ideally before the reverse of the trend, which as we can see takes place at Point C. The closer to the end of the OS event we can trade (i.e., the closer to Point C), the higher the profit margin a trader can make. Thus, it is crucial to be able to anticipate the reverse in the trend and be able to act before that. In order to tackle this, we use the scaling laws we discussed earlier in Section 2.2. As explained, these laws states that when on average a DC event takes t amount of physical time to complete, the OS event takes an amount of 2t. Because this is an approximation and it can rely on the underlying dataset, we wanted to have our own calculations for the datasets we are dealing with. Thus, what we did was to calculate the average time of each OS event for every period we would use as a training set. We hence created two variables, expressed as the average ratio of the OS event length over the DC event length. These two variables are ru and rd , where ru is the average ratio of the upwards OS event, and rd is the average ratio of the downwards OS event. Our calculations showed that these two variables had indeed values around 2 (ranging between 1.8–2.0), which confirms the scaling law findings. More importantly, this allowed us to have tailored values for ru and rd , for each training set we use. After obtaining these ratios, we were able to anticipate the end of a trend (approximately) and as a result make trading decisions once an OS event had reached the average ratio of ru or rd . Of course, in reality things are not that simple. The ru and rd ratios are just average approximations, so many times the OS event might last longer or shorter than anticipated. In an attempt to address this issue, we have created two user-specified parameters, namely b1 and b2, which define a range of time within the OS period, where trading is allowed. For instance, if a trader expects the OS event to last for 2 hours, then we can define an action range of [b1,b2]= [0.90,1.0], which effectively means we are going to trade at the last 10% of the 2 hours duration, i.e. in the last 12 minutes. By introducing b1 and b2, we are essentially attempting to anticipate the approximation errors that might have been created during the calculation of ru and rd . Equation 7 presents the formulas for these starting and ending for upward and downward OS periods: t U 0 = (t dc 1 − t dc 0 )× ru × b1 t U 1 = (t dc 1 − t dc 0 )× ru × b2 t D 0 = (t dc 1 − t dc 0 )× rd × b1 t D 1 = (t dc 1 − t dc 0 )× rd × b2 , (7) where tU0 , t U 1 are the start and end times for upwards overshoot period, respectively, and t D 0 , t D 1 are the start and end times for downwards overshoot period, respectively. In addition, tdc0 and t dc 1 are the start and the end times of the current DC event, after the confirmation of the event has taken place at time tdc1 . Their difference t dc 1 − t dc 0 returns the length of the current DC event. Also, ru and rd are the average ratios of the upwards and downwards OS period lengths, respectively, over the current DC period. Lastly, b1 and b2 are the two parameters defining the action range within the OS periods, as explained above. Although b1 and b2 define a window for trading, a problem that exists with high-frequency data—particularly tick data—is that there can still be hundreds of points to trade, even if that trading window is very narrow. This could 10 Table 1: DC strategy parameters. Parameter Description ru Average ratio of upwards OS event over the upwards DC event length rd Average ratio of downwards OS event over the downwards DC event length Qtrade Quantity to trade N↓ Number of thresholds recommending to buy N↑ Number of thresholds recommending to sell Nθ Total number of thresholds used in the experiments Q Quantity for trading b1 Start of trading period during an OS event b2 End of trading period during an OS event b3 Range of prices close to the trading price Ptrade that a trade can be perfomed be problematic, because trading at multiple price levels will not return the highest profit. What is more effective is to sell (buy) at a price as expensive (cheap) as possible. To achieve this, we introduced another variable b3, which prevents traders from doing expensive trades. To ensure this, we only allow the system to sell at the most expensive (peak) price Ppeak and buy at the cheapest recorded price (trough) Ptrough, or in prices in close range. This range is determined by the value of b3. Therefore a trader would sell when the price is equal to Ppeak × b3, or buy when the price is equal to Ptrough ×(1 − b3). Essentially, b3 is a value within the range of [0,1] and defines the range of prices close to Ppeak andPtough that the system will perform an action. Furthermore, there is a user-specified parameter Q, which controls the trading quantity. Lastly, it should be mentioned that our system allows short selling. However, in order to avoid excess short selling, which can lead to significant losses, we have introduced a stop loss mechanism that is called short selling allowance. This allowance is a percentage of our budget and allows short selling activities up to this pre-specified percentage. This percentage is decided during parameter tuning. 3.2. Multi-threshold DC trading strategy This strategy builds on the previous one, as it still uses Equation 7 and the b3 variable. But instead of only using a single threshold, it combines information by multiple thresholds. As we discussed in Section 2.2, a DC event is identified by a change in the price by a given threshold value. The use of different DC thresholds provides a different view of the data: smaller thresholds allow the detection of more events and, as a result, actions can be taken promptly; larger thresholds detect fewer events, but provide the opportunity of taking actions when bigger price variations are observed. This proposed trading strategy combines the use of different threshold values in an attempt to take advantage of the different characteristics of smaller and larger thresholds. From the single-threshold strategy we know that under a specific threshold we should buy towards the end of a downtrend and sell towards the end of an uptrend (i.e. towards the end of the respective OS events). Since now we are dealing with multiple thresholds, each threshold summarises the data in a unique way. For example, at one point in time the trading strategy under one threshold could be recommending a buy action, while under a different threshold recommend a sell action. As we have already argued, the advantage of having the multiple thresholds is that we have multiple recommended actions per data point. In order to decide which action to follow, a majority vote takes place. In order to allow for a majority vote, we associate each DC treshold to an equal weight of 1 (vote). Therefore, W1 = W2 = W3 = ... = WNθ = 1, where Nθ is the total number of thresholds used. As a result, at any point in time the trading strategy is able to make a buy/sell/hold recommendation based on the combined recommendations of all thresholds. As we already know, each threshold produces DC events; thus each threshold is able to make this buy/sell/hold recommendation. Since we have Nθ thresholds, this means that at any point in time we receive Nθ 11 recommendations. In order to decide which recommendation to follow, we sum the weights of the thresholds: if the sum of the weights for all thresholds recommending a buy (sell) action is greater than the sum of the weights for all thresholds recommending a sell (buy) action, then the strategy’s action will be to buy (sell). The hold action is a special case of both buy and sell and it happens when we are outside the price range recommended by b3, or when there is not enough quantity to act, see Algorithm 4 lines 8, 11, and 26. In addition, the multi-threshold trading strategy is able to make recommendations on the trading quantity Qtrade. The decision for this quantity is a dynamic decision, taken by the number of DC thresholds that are advising to sell (buy) at a certain point in time: if many thresholds are advising to sell (buy), then the algorithm sells (buys) a higher quantity of the given currency pair. Equations 8a and 8b present the relevant formulas, for buy and sell, respectively: Qtrade =(1 + N↓ Nθ )× Q (8a) Qtrade =(1 + N↑ Nθ )× Q (8b) where Qtrade is the quantity to trade, N↓ and N↑ are the number of thresholds recommending to buy and sell, re- spectively, Nθ is the total number of thresholds used in our experiments, and Q is the user-specified quantity already presented in Section 3.1. As we can see, by taking into account the recommendations given by the DC thresholds, we are giving more or less weight to the Q quantity, resulting to a new quantity Qtrade. This concludes the presentation of the two proposed DC strategies and their respective parameters. For the con- venience of the reader, we have summarised and listed these parameters in Table 1. We have also summarised the trading strategy processes into pseudocodes: Algorithm 3 presents an overview of how the return of a trading strategy is calculated. In addition, Algorithm 4 presents how the buy and sell actions take place. While these algorithms are for the multi-threshold strategy, they can also be applied to the single-threshold strategy, where there is only a single weight (for the single threshold) of W1 = 1 and Qtrade = Q. 4. Optimising multi-threshold strategies via a Genetic Algorithm In the previous section, we presented two novel trading strategies based on the DC paradigm: a single-threshold strategy and a multi-threshold strategy that builds on top of the single-threshold. While the multi-threshold strategy has the advantage of combining recommendations from different thresholds, a problem that exists is that we do not know how much weight we should give to each threshold. Simply assigning an equal weight of 1 to all of the thresholds might be a naive approach. Some thresholds might be more useful than others, hence we should give them more weight. Thus, we use a genetic algorithm (GA) to evolve real values for the weight of each DC threshold. In addition, we also evolve some other DC parameters that are crucial to the success of the trading strategy. All these are discussed next, in Section 4.1, where the GA representation is presented. Then, Section 4.2 presents the GA operators and Section 4.3 presents the fitness function. 4.1. Representation Each chromosome consists of 4 + Nθ genes, where Nθ is the number of different threshold values of the multi- threshold strategy. The number 4 denotes that in addition to the thresholds, there are also 4 additional parameters to be optimised: Q (first gene), b1 (second gene), b2 (third gene), and b3 (fourth gene). Q,b1,b2 and b3 refer to the DC-related parameters presented in Section 3.1. Each remaining gene in the chromosome (positions 5 to [4+Nθ]) represents a weight associated to a given threshold. Thus, after first deciding the DC threshold values (through parameter tuning) and generating the DC events per threshold, each GA gene is assigned the same initial weight. Therefore, W1 = W2 = W3 = ... = WNθ = 1 Nθ . The GA then evolves the weight for each threshold (in addition to the 4 parameter values in positions 1-4). As a result, at any point in time a GA individual is able to make a buy/sell/hold recommendation based on the combined recommendations of all thresholds by using the majority vote mechanism we presented in Section 3.2. An example of an 8-gene GA chromosome is presented in Figure 5. 12 Algorithm 3 Pseudocode for calculating the return of a trading strategy. Require: Initialise variables (cash = budget,Qtrade = 0,current = 0) Require: b1,b2, b3, Q and weight values W1 = W2 = ...= WNθ = 1 for each threshold 1: for i = 0; i < dataset_length; i ++ do 2: Initialise variables weights for buy and sell: WB = WS = 0, number of upturn and downturns: N↑ = N↓ = 0 3: current ← current + 1 4: if PC > Ppeak then //PC is the current price and Ppeak is the highest so-far price. 5: Ppeak ← PC 6: else if PC < Ptrough then 7: Ptrough ← PC 8: end if 9: for j = 0; j < Nθ; j ++ do 10: Calculate tU0 , t U 1 , t D 0 , t D 1 as explained in Equation 7 11: if event is Downturn Event then 12: WB ← WB + W j 13: if current within range of [tD0 , t D 1 ] then 14: N↓ ← N↓+ 1 15: else 16: N↓ ← N↓− 1 17: end if 18: else 19: WS ← WS + W j 20: if current within range of [tU0 , t U 1 ] then 21: N↑ ← N↑+ 1 22: else 23: N↑ ← N↑− 1 24: end if 25: end if 26: end for 27: if WS > WB then 28: Perform the sell action for a given quantity [see Algorithm 4] 29: else if WS < WB then 30: Perform the buy action for a given quantity [see Algorithm 4] 31: end if 32: end for 33: Wealth ← cash + Qtrade × PC 34: Return ← 100 × wealth budget Figure 5: An example of a 8-gene GA chromosome. The first four genes are : Q,b1,b2 and b3, respectively. The remaining four genes are the weights for the DC thresholds: W1,W2,W3, and W4. 13 Algorithm 4 Pseudocode for performing the buy and sell actions. 1: if WS > WB then 2: if N↑ > 0 and PC ≥ b3 × Ppeak then 3: Qtrade ←(1 + N↑ Nθ )× Q 4: if Qtrade > 0 or (Qtrade ≤ 0 and ∣Qtrade∣× PC ≤ shortS ellingAllowance × budget) then 5: Cash ← Cash + Qtrade × PC 6: PF L ← PF L − Qtrade // PFL stands for Portfolio, i.e. the amount/quantity of the currency pair we are currently holding 7: else 8: Hold 9: end if 10: else 11: Hold 12: end if 13: else if WS < WB then 14: if N↓ > 0 and PC ≤ Ptrough +(Ptrough ×(1 − b3)) then 15: Qtrade ←(1 + N↓ Nθ )× Q 16: if cash > Qtrade × PC then 17: Cash ← Cash − Qtrade × PC 18: PF L ← PF L + Qtrade 19: else 20: // Buy only as much as you can afford 21: Q′trade is the new quantity to be traded, up to the amount we can afford 22: Cash ← Cash − Q′trade × PC 23: PF L ← PF L + Q′ 24: end if 25: else 26: Hold 27: end if 28: end if 14 Based on this example, the GA recommends buying/selling a quantity of Q equal to 10, and only acting in the period [0.9,1.0] of the estimated duration of the OS event (i.e., in the last 10% of the length of the OS event). In addition, the fourth gene recommends to only consider prices that are within a 20% range (the value of b3 is 0.8, so 1.0 − 0.8 = 0.20 or 20%) of the highest (lowest) recorded price Ppeak (Ptrough). In addition, to decide the trading action, we would check the recommendation of each individual threshold. For this example, let us assume that the first threshold recommends buy, the second threshold recommends sell, the third threshold recommends buy, and the fourth threshold recommends hold. We would then sum up the weights of the thresholds, according to each action. Therefore, the weight for buying WB is equal to W1 + W3 = 0.2 + 0.2 = 0.4, and the weight for selling WS is equal to W2 = 0.5. 2 Since WS > WB, the GA’s recommendation would be to sell. 4.2. Operators The following three operators are being used during the evolutionary process: elitism, uniform crossover and uniform mutation—as detailed in Section 2.3. In elitism, the best-performing individual (in terms of fitness) is copied to the next generation. In uniform crossover, two parents are selected via a tournament selection. In this type of crossover, the genes between the two parents are swapped with a fixed probability of 0.5. In addition, we ensure that the value of the third gene is always greater than the value of the second gene, i.e. b2 always has to be greater than b1. Lastly, for the uniform mutation operator a single parent is selected, again by tournament selection. With a probability of 0.5, each gene of the chromosome is mutated, and a different value is obtained. It should be clarified here that for the first gene (quantity Q), the mutated value can be any integer up to a pre-specified maximum quantity value; whereas for the remaining genes (i.e., b1,b2,b3 and all weights W), the mutated values are real numbers between 0 and 1, where b2 > b1. 4.3. Fitness function Several different metrics have been used in the literature as fitness function in algorithmic trading problems. Some examples are: wealth, profit, return, Sharpe ratio, information ratio (Brabazon & O’Neill, 2006; Bradley et al., 2009). In this paper, we set our fitness equal to the total return minus the maximum drawdown, presented in Equation 9: f f = Return −α× MDD MDD = Ptrough − Ppeak Ppeak , (9) where Return is the return of the investment, MDD is the maximum drawdown, and α is a tuning parameter. Maximum drawdown is defined as the maximum cumulative loss since commencing trading with the system. It is used to penalise volatile trading strategies in terms of return. Its value is given as the percentage of Ptrough−Ppeak Ppeak , where Ptrough the trough value of the price, and Ppeak is the peak value of the price. Lastly, the tuning parameter α is used to define how much risk-averse the strategy is. The more risk-averse in terms of wishing to avoid a catastrophic loss, the higher the value of α. 5. Experimental Setup This section is divided into three parts: Section 5.1, where we present the data we use for our experiments, Section 5.2, where we present the experimental setup, and lastly, Section 5.3, which presents the experimental parameters. 2As explained in the previous section, the hold action is an exceptional case that is considered as an alternative to buy and sell actions; see Algorithm 4, Lines 8, 11 and 26 for detail. 15 Figure 6: Minimum and maximum daily tick data (transactions) per month for the GBP-JPY currency pair, for the period June 2013 to May 2014. 5.1. Data We use two different types of datasets: (i) tick data and (ii) intra-day data at 10-minute intervals.3 In the first case of tick data, we use a year’s FX spot tick data on a daily basis from the currency pair of GBP/JPY (British Pound and Japanese Yen), for the period June 2013 to May 2014. Thus, we use a daily rolling window, where a single day is used for training the algorithm, and the consecutive day is used for testing the returned model. Exception to this rule was when there is a weekend, which is not taken into account. The number of tick data can vary significantly from day to day, and even more from month to month. Nevertheless, each day has a very high number of observations, giving more than enough training data for the GA to learn and produce a profitable model. As we can observe from Figure 6, where the minimum and maximum number of daily tick data on a given month are presented, a day could have anything between approximately 70,000 transactions (minimum value of April 2014) to above 900,000 transactions (maximum value of June 2013). It should be noted that this high number of data per day should not considered to be a problem for the DC algorithm, i.e. that the algorithm is dealing with too much data to handle; on the contrary, this is one of the strengths of the algorithm, since it will only be focusing on the important events, thus filtering out all ‘noise’ from the data. In addition, we use 10-minute interval high frequency data for the following currency pairs: EUR/GBP (Euro and British Pound), EUR/USD (Euro and US dollar), EUR/JPY (Euro and Japanese Yen), GBP/CHF (British Pound and Swiss Franc), and GBP/USD (British Pound and US dollar). The period is again June 2013 to May 2014. Since the amount of the 10-minute data is significantly less than the tick data (e.g. for the whole of June 2013 for EUR/GBP there’s around 3000 entries for the whole month), we test our algorithms in the following way: every month is split into its own dataset, with the first 70% of the data being the training set, and the remaining 30% being the testing set. 5.2. Algorithmic experimental setup In order to demonstrate the efficiency of our evolutionary event-based DC approach, we will be comparing it with several other benchmarks. This section presents in detail the different algorithms that we use to benchmark our approach. It should be stated that the objective function (i.e., the function that all trading strategies are optimising) for all of these algorithms is Equation 9, which was presented earlier in Section 4.3. Our proposed algorithm is going to be benchmarked against 3 different other types of trading strategies: (i) single- threshold and multi-threshold directional changes, (ii) buy and hold, and (iii) technical analysis. In addition, there 3All data was purchased by OlsenData: http://www.olsendata.com 16 are 4 parameters for all DC configurations, which depending on the experimental setup, we optimise or not. These 4 parameters are: Q, b1, b2, and b3. Please refer back to Section 3.1 for a detailed presentation of these parameters. Therefore, by taking the above parameters into account, we have the following different configurations, which will constitute our different algorithmic experimental setups: Standard directional changes The purpose here is to present the results of the DC paradigm, under a single-threshold and a multi-threshold framework. The values of thresholds were decided during the parameter tuning process. 1. Single-threshold, with no evolution (SDC) This is the trading strategy presented in Section 3.1. The 4 parameters [Q,b1,b2,b3] have not been optimised and have fixed values: Q = 1,b1 = 0,b2 = 1,b3 = 1, which essentially allow trading of a single quantity throughout the length of the OS event for any given price, no matter how expensive or cheap it is. Of course, this is not the optimal setup for these values and this is why in the next setup we evolve these parameters to obtain better values. However, we consider it important to report results from this setup of SDC to demonstrate that it is crucial to optimise the four parameters [Q,b1,b2,b3]. Lastly, in order to decide which (single) threshold to use, we experiment with several different thresholds (one at a time) and report the performance of the best threshold. 2. Single-threshold, with evolution on the 4 parameters (S DCEV O) As above. The difference is that now we use a standard GA to evolve the values of the 4 numeric parame- ters [Q,b1,b2,b3]. The idea behind this setup was to evolve the 4 parameters for a single threshold, so that algorithmic performance is optimised for that specific threshold. 3. Multi-threshold, with no evolution (MDC) This is the trading strategy presented in Section 3.2. The 4 parameters [Q,b1,b2,b3] have not been optimised and have fixed values: Q = 1,b1 = 0,b2 = 1,b3 = 1. 4. Multiple-threshold, with evolution on the 4 parameters (MDCEV O) As above. The difference is that now we use a standard GA to evolve the values of the 4 numeric parameters [Q,b1,b2,b3]. Buy and hold Buy and hold is a common benchmark for trading algorithms. Under this strategy, one would buy at a cer- tain point in time and not act (hold) for a long period. Thus, traders are not concerned with short-term price movements, as they expect that in the long term the value of their portfolio will increase. 5. Buy and Hold (BH) Buy at the beginning of the trading period in August 20134, sell at the end of the period, in May 2014. Technical analysis 6. EDDIE The EDDIE algorithm, which uses technical analysis indicators to evolve decision trees that make suggestions for buying, selling, or holding. Proposed algorithm 7. DC+GA Our proposed algorithm, which evolves both the DC threshold weights [W1,W2, ...,WNθ] and the 4 DC parame- ters [Q,b1,b2,b3]. As we can observe, DC+GA will be benchmarked against 6 different types of traging strategies. Next, we present the parameter tuning process that we undertook. 4The first two months (June and July 2013) were used for parameter tuning, and the remaining ten months were used for our experiments. More details about this in Section 5.3. 17 Table 2: Experimental parameters determined using I/F-Race. Parameter SDC / MDC SDCEV O / MDCEV O / DC+GA EDDIE Population N/A 1000 500 Generations N/A 35 30 Tournament size N/A 7 2 Crossover probability N/A 0.90 0.90 Mutation probability N/A 0.10 0.10 Number of thresholds 5 5 N/A Short selling allowance 0.25 0.25 0.25 MDD weight 0.20 0.20 0.20 5.3. Experimental parameters In order to decide the values for the parameters for the algorithms, we undertook a parameter tuning process by using the I/F-Race package (Lopez-Ibanez et al., 2011). It should be noted that buy and hold is a simple process with no parameters that require tuning. I/F-Race implements the iterated racing procedure, which is an extension of the Iterated F-race process. Its main purpose is to automatically configure optimisation algorithms by finding the most appropriate settings, given a set of instances of an optimisation problem. It builds upon the race package by Birattari et al. (2009). In order to avoid biased results, we used the first two months of our data (June and July 2013) for each currency pair (both tick and 10-minute data) for tuning purposes. Thus, I/F-Race was applied to the data of June and July 2013. The remaining ten months (August 2013–May 2014) were used only with the tuned parameters, after I/F-Race was complete. At the end of the tuning process, we picked the best parameters returned by I/F-Race. These parameters constitute the experimental parameters for our algorithms. These parameters are presented in Table 2. The buy and hold setup did not have any parameters, so it is not present in Table 2. As we can observe, we will be using 5 different thresholds. These thresholds are: 0.01%, 0.013%, 0.015%, 0.018%, and 0.02%. In addition, it should be mentioned that when we use an evolutionary algorithm (S DCEV O, MDCEV O, EDDIE, and DC+GA), the experiments are run 50 times on each dataset and the results presented corre- spond to the average value over the 50 executions; SDC and MDC are run just once per dataset, since they represent deterministic strategies. Similarly, the buy and hold strategy BH is run one time per dataset, as it also represents a deterministic strategy. 6. Results This section presents results for all DC algorithms, and the two physical-time financial benchmarks of BH (buy and hold) and EDDIE (technical analysis), for the currency pair of GBP/JPY under tick data, and for the remaining five currency pairs (EUR/GBP, EUR/JPY, EUR/USD, GBP/CHF, GBP/USD) under the 10-minute interval data. Ex- periments took place for the 10 month period of August 2013–May 2014. As explained in Section 5, we used a daily rolling window for GBP/JPY, where every day was used for training the algorithms, and the following day was used for testing. The above setup resulted in 205 different datasets, i.e. each algorithm was tested at 205 different unseen datasets for the tick data of GBP/JPY. In addition, for each of the remaining five currency pairs we undertook exper- iments for each month during the 10 month period August 2013 - May 2014. Therefore, this returned 50 different datasets for the 10-minute interval currency pairs. Therefore, our experiments were conducted over a total of 255 different datasets. To increase comprehensibility, we divide this section in the following way: Section 6.1 presents results for the tick data dataset (GBP/JPY), Section 6.2 presents results for the 10-minute interval datasets, Section 6.3 presents the computational time results for the algorithms, and Section 6.4 discusses the results. In addition, Sections 6.1 and 6.2 18 Figure 7: Day-to-day average return for the period August 2013 to May 2014 for SDC, SDCEV O, MDC, MDCEV O and DC+GA. Results shown in % values. are futher divided into two subsections: 6.1.1, 6.1.2, and 6.2.1, 6.2.2, respectively. Subsections 6.1.1 and 6.2.1 present a comparison among the DC algorithms only (i.e. SDC, SDCEV O, MDC, MDCEV O, and DC+GA), in order to identify the best DC setup; subsections 6.1.2 and 6.2.2 present results among the best DC algorithm and the two physical-time financial benchmarks (i.e. BH and EDDIE). In addition, we would like to remind the reader that the goal of our experiments is threefold: (i) demonstrate that the paradigm of DC returns profitable strategies, (ii) provide evidence that the DC strategies optimised by the GA are more profitable than using standard DC strategies, and (iii) demonstrate that our GA generated strategies outperform typical physical-time based strategies, namely technical analysis and buy and hold. We demonstrate the fulfilment of (i) and (ii) in Sections 6.1.1 and 6.2.1. We demonstrate the fulfilment of (iii) in Sections 6.1.2 and 6.2.2. Lastly, when an algorithm yields positive return, we will also be commenting on its MDD performance, as an indicator of downside risk.5 In this way, we make a detailed analysis on both performance metrics (return, MDD), which offers a more holistic view on the results of the trading algorithms. 6.1. Tick data results 6.1.1. Comparison among the DC algorithms Since each algorithm was tested on a daily basis (excluding weekends) over the 10-month period, we can calculate the daily return for each algorithm. Figure 7 presents the box and whisker plot for each DC algorithm. As we can observe, SDC, MDC, and MDCEV O and DC+GA show results with very low variance, as all of them are concentrated around the mean that seems to be a value near and above zero. On the other hand, SDCEV O experiences high variance and many extreme values, both positive and negative. These results are summarised in Table 3. What we can observe is that only MDCEV O and DC+GA have positive mean daily return over the 10 month period. We can also observe that their mean returns are relatively close, as MDCEV O’s return is 0.0677% and DC+GA’s is 0.0730%. We should note here that while these return values are relatively low, the reader should keep in mind that trading takes place on a daily basis. Thus, an average daily mean return of the scale of 0.0730% will have a significant cumulative effect in the long run. This is further demonstrated in Section 6.1.2, when we present and discuss with equity curve for DC+GA. 5We are analysing MDD performance only on strategies with positive returns, as a trader would not consider at all a trading algorithm that yields negative returns. 19 Table 3: Mean return results for each DC algorithm. Tick data for GBP/JPY. Results shown in % values. SDC SDCEV O MDC MDCEV O DC+GA Mean -0.0053 -5.6975 -0.0092 0.0677 0.0730 StandDev 0.0536 25.296 0.1069 0.3673 0.3942 Max 0.0963 90.170 0.1820 1.1684 1.2587 Min -0.3873 -136.18 -0.7751 -1.1750 -1.3742 Table 4: Mean return results for EDDIE and DC+GA under GBP/JPY’s tick data. BH’s return (not included in the table, as it does not do daily trading) was -0.1164. EDDIE DC+GA Mean -0.1918 0.0730 StandDev 0.3732 0.3943 Max 0.929 1.26 Min -2.01 -1.37 To investigate whether there is a statistical significance between MDCEV O and DC+GA, we ran the Kolmogorov- Smirnoff non-parametric test, with the null hypothesis being that the data from these two algorithms come from the same continuous distribution. The test showed that indeed they both come from the same distribution with a p-value of 0.9638, thus the difference in the mean values is not statistically sigificant. In addition, we look into the maximum drawdown (MDD) values for these two algorithms, to get insight on the downside risk of the trading strategies generated by each algorithm. The average value for MDCEV O is 0.4156%, whereas DC+GA’s value is slighly higher, at 0.4251%, showing that both algorithms’ strategies have similar downside risk. We further explore the effect of this risk in the next section, when we present the average daily return and its fluctuations. Since DC+GA returned higher mean return, we use this setup for the comparisons with the physical-time financial benchmarks. 6.1.2. Comparison with physical-time financial benchmarks Table 4 presents the mean results for DC+GA and EDDIE. We should also note that BH yielded a return of -0.1164%. As we can observe from the table, EDDIE also has a negative mean daily return of -0.1918%. There- fore, DC+GA was the only algorithm among these three with a positive daily return. In addition, a two-sample Kolmogorov-Smirnov test at 5% significance level returned a p-value of 4.7194e-09, and thus showed that DC+GA significantly outperformed EDDIE. To visualise DC+GA’s results, we present the average (over the 50 runs) daily return for the period August 2013 to May 2014 in Figure 8. As we can observe, the majority of the days experience a positive return. In fact, 58.5% of the tested datasets experienced a positive return (120 out of the total of 205 days). Furthermore, Figure 9 presents the equity curve for DC+GA. Equity curve is a graphical representation of the change in value of a trading account over a time period. An equity curve with a consistently positive slope would generally indicate that the trading strategies of the account are profitable, while a negative slope would indicate that the account is losing money. As we can observe, the given equity starts from an initial budget of £500K and never drops below this threshold. It generally follows a positive slope, with the only exception of around February-March, where there was a decline. Nevertheless, the curve soon returns to its positive slope, demonstrating the long-run effectiveness of the trading strategy. This concludes the results under tick data, which showed that DC+GA was ranked first among the other DC versions, and also outperformed the two physical-time financial benchmarks. Next, we present results under the 20 Figure 8: Day-to-day average return for the DC+GA strategy for the period August 2013 to May 2014 for GBP/JPY’s tick data. Results shown in % values. Figure 9: Equity curve for the DC+GA strategy for the period August 2013 to May 2014 for GBP/JPY’s tick data. 21 10-minute interval data. 6.2. 10-minute interval data results This section presents results for the 10-minute interval data for the five currency pairs: EUR/GBP, EUR/JPY, EUR/USD, GBP/CHF, and GBP/USD. We start again by presenting results for the DC algorithms only, in Section 6.2.1. After identifying the best DC setup, we move on to Section 6.2.2, where we compare this best DC setup with BH and EDDIE. 6.2.1. Comparison among the DC algorithms Table 5 presents the return results for each month, for each DC algorithm, under the 10-minute interval data, for each currency pair. Overall, each DC algorithm appears to be able to show positive returns for each currency pair. SDC and SDCEV O are the two algorithms with the highest frequency of positive returns. Out of the 10 months tested per currency pair, SDC had 7 positive returns for EUR/GBP, 6 positive returns for EUR/JPY, 7 positive returns for EUR/USD, 5 positive returns for GBP/CHF, and 3 positive returns for GBP/USD. Similarly, SDCEV O’s number of positive returns were 6, 5, 6, 6, and 4. Table 6 summarises these results. In terms of currency pairs, it appears that EUR/GBP is the easiest to predict, as all algorithms showed non-negative returns. On the other hand, all other currency pairs had two to three currency pairs with negative returns. Overall, SDC and MDC experience again, as with the tick data, a negative mean return. On the other hand, SDCEV O, MDCEV O and DC+GA experience positive mean returns, with values close to each other (0.01064%, 0.00875%, and 0.01046%). It thus appears that these three algorithms have similar performance. Once again, we would like to note that while these returns appear to be low, their cumulative effect can be much higher when trading over all 50 datasets available for the 10-minute interval data. To further investigate the algorithms’ performance, we applied Friedman’s non-parametric statistical test to com- pare multiple algorithms. We present the results in Table 7. For each algorithm, the table shows the average rank according to the Friedman test (first column), and the adjusted p-value of the statistical test when that algorithm’s average rank is compared to the average rank of the algorithm with the best rank (control algorithm) according to the Hommel post-hoc test (second column) (Demšar, 2006; García & Herrera, 2008). As we can observe from the Friedman test, there is no statistical significance between SDCEV O and DC+GA at the α = 0.05 level. Also, there is no statistical significance between SDCEV O and MDCEV O at the α= 0.05 level, but there is a statistical significance at the α= 0.10 level. Lastly, we compare the MDD results over the three algorithms that yielded positive mean return (i.e., SDCEV O, MDCEV O, and DC+GA). DC+GA and MDCEV O have the lowest mean MDD values, 0.03789% and 0.03308%, re- spectively; on the other hand, SDCEV O’s mean MDD value is higher, at 0.05251%. Overall, all algorithms showed very low MDD values. One interesting observation that can be made is that while SDCEV O returned the highest mean return, as we saw in Table 6, it also returned more volatile trading strategies. This is mainly because of the much higher MDD value for EUR/JPY in Table 8 (0.22863% for SDCEV O, against 0.12386% and 0.14274% for MDCEV O and DC+GA, respectively). Nevertheless, since SDCEV O ranked first in terms of mean return, we are going to be using it with the comparisons with the physical-time financial benchmarks. 6.2.2. Comparison with physical-time financial benchmarks Table 9 presents the mean return for EDDIE and SDCEV O under the 10-minute interval datasets (for completeness, we also present the month-by-month return results in the Appendix in Table A.11.). We should also note that BH’s average return was 0.01274%. As we can observe, EDDIE has again a negative mean return of -0.00873%; it is also worth noting that for all five currency pairs EDDIE’s mean return is negative. On the other hand, SDCEV O has a positive return for three currency pairs: EUR/GBP, EUR/JPY, and EUR/USD. Overall, SDCEV O’s mean return is 0.01064%. A two-sample Kolmogorov-Smirnov test for EDDIE and SDCEV O returned a p-value of 0.0560, showing that there is a statistical significance between these two algorithms at the 10% significance level. However, the fact that EDDIE returned a negative mean return means that it would not be attractive to an investor as a trading algorithm. This leads us to argue that SDCEV O outperforms EDDIE, while it returns a similar average return with BH. 22 Table 5: Monthly return results for each DC algorithm. 10-minute interval data. Results presented per currency pair, in % values. SDC SDCEV O MDC MDCEV O DC+GA EUR/GBP August 0.000004 -0.005459 0.000012 -0.007698 -0.009455 September -0.000014 -0.004532 -0.000032 -0.003943 -0.002992 October 0.000007 0.002058 0.000011 -0.000974 -0.001070 November 0.000002 0.002844 0.000007 0.001022 0.002826 December 0.000001 0.013137 0.000001 0.007134 0.006939 January 0.000005 -0.000073 0.000010 0.004246 0.004940 February 0.000004 0.000962 0.000005 -0.000139 0.000471 March -0.000003 -0.002265 -0.000009 0.000006 -0.001710 April 0.000001 0.000304 0.000002 0.002933 0.003820 May -0.000001 0.000892 -0.000001 0.000797 0.000344 EUR/JPY August -0.000483 0.155999 -0.000490 0.000000 0.000000 September -0.000433 0.050100 -0.001048 0.001103 -0.022498 October 0.000209 0.060285 -0.000711 0.003731 -0.072743 November -0.000822 0.000000 -0.001747 -0.005446 -0.001633 December -0.003502 -0.421926 -0.008150 -0.084816 -0.049468 January 0.000647 0.771926 0.001547 0.584600 0.598602 February 0.000320 -0.006850 0.000485 -0.004669 -0.004390 March 0.000395 -0.118628 0.000842 -0.106838 -0.085178 April 0.000375 0.049121 0.000679 0.015529 0.057428 May 0.000460 0.040874 0.001183 0.062574 0.064694 EUR/USD August -0.000010 0.001232 -0.000022 0.000000 0.000000 September 0.000003 0.001880 0.000004 -0.000063 -0.000062 October 0.000016 -0.007077 0.000030 -0.005307 -0.003792 November -0.000001 -0.000073 -0.000002 -0.001359 -0.001942 December 0.000002 -0.005356 0.000007 -0.007442 -0.008586 January 0.000024 0.014046 0.000052 0.024849 0.018826 February 0.000010 -0.000601 0.000029 -0.000359 -0.004705 March 0.000001 0.003289 0.000003 0.001864 0.004594 April -0.000001 -0.006851 -0.000009 -0.014233 -0.016055 May 0.000003 0.000756 0.000008 0.000176 0.000430 GBP/CHF August 0.000007 -0.004584 0.000009 -0.000334 0.000061 September 0.000000 -0.014025 0.000000 -0.019435 -0.026564 October 0.000002 0.002891 0.000001 0.004546 0.007189 November 0.000004 -0.002861 -0.000008 -0.000798 -0.000490 December -0.000006 0.000477 -0.000010 -0.001521 -0.002420 January -0.000011 0.000968 -0.000035 -0.000030 -0.000033 February 0.000012 0.009052 0.000024 0.015172 0.015056 March -0.000015 -0.027585 -0.000026 -0.032727 -0.034124 April -0.000001 0.000696 -0.000006 -0.000063 -0.000009 May 0.000002 0.005771 0.000004 -0.001288 0.002493 GBP/USD August -0.000004 0.000249 0.000000 -0.000863 -0.000716 September -0.000031 -0.014953 -0.000043 0.005087 0.003959 October 0.000001 0.028377 -0.000002 0.035794 0.052110 November -0.000001 0.000013 -0.000009 0.000000 0.000000 December 0.000001 -0.030266 0.000001 -0.035563 -0.040706 January -0.000009 -0.001725 -0.000024 0.000107 0.001240 February -0.000004 -0.012477 -0.000005 -0.004516 -0.004187 March -0.000015 -0.002918 -0.000026 -0.012024 -0.010595 April 0.000000 -0.005678 0.000001 -0.002531 -0.003123 May 0.000000 0.010379 -0.000001 0.021339 0.028387 23 Table 6: Mean return results for each DC algorithm. 10-minute interval data. Results shown in % values. SDC SDCEV O MDC MDCEV O DC+GA EUR/GBP 0.00000 0.00079 0.00000 0.00034 0.00063 EUR/JPY -0.00028 0.05809 -0.00074 0.04658 0.05387 EUR/USD 0.00000 0.00012 0.00001 -0.00019 -0.00125 GBP/CHF 0.00000 -0.00292 0.00000 -0.00365 -0.00388 GBP/USD -0.00001 -0.00290 -0.00001 0.00068 0.00293 Mean -0.00006 0.01064 -0.00015 0.00875 0.01046 Table 7: Statistical test results according to the non-parametric Friedman test with the Hommel’s post-hoc test. 10-min interval data. Significant differences at the α= 0.1 level are shown in boldface. Algorithm Average Rank Adjusted pHomm SDCEV O (c) 1.86 - DC+GA 1.91 0.80258 MDCEV O 2.23 0.06431 Table 8: Mean MDD results for each DC algorithm. 10-minute interval data. Results shown in % values. SDCevo MDCevo DC+GA EUR/GBP 0.00347 0.00418 0.00520 EUR/JPY 0.22863 0.12386 0.14274 EUR/USD 0.00798 0.01334 0.01402 GBP/CHF 0.01139 0.0114 0.01302 GBP/USD 0.01107 0.0126 0.01449 Mean 0.05251 0.03308 0.03789 Table 9: Mean return results for EDDIE and SDCEV O. 10-minute interval data. BH’s average return (not included in the table) was 0.01274%. Results shown in % values. EDDIE SDCEV O EUR/GBP -0.00141 0.00079 EUR/JPY -0.01644 0.05809 EUR/USD -0.00840 0.00012 GBP/CHF -0.01114 -0.00292 GBP/USD -0.00628 -0.00290 Mean -0.00873 0.01064 24 Table 10: Mean computational times per run for SDCEV O, MDCEV O, EDDIE, and DC+GA. SDC, MDC, and BH are deterministic algorithms and only take 1 second to execute. SDCEV O MDCEV O EDDIE DC+GA Tick 18 secs 20 secs 55 secs 45 secs 10-min 10 secs 12 secs 25 secs 20 secs 6.3. Computational times Table 10 presents the average computational times per run for all algorithms. SDC, MDC, and BH are deterministic algorithms and are thus very fast in executing (around 1 second). All other algorithms have their execution times varying between 10 seconds and 55 seconds. Thus, all algorithms have relatively fast execution times. As we can see, DC+GA ranks third in terms of computational cost, but we believe that this slower execution time is justified by the improvements in the algorithm’s mean return performance. Besides, all these are very minor differences, especially after taking into account that the current forecasting application is an off-line problem. Lastly, evolutionary algorithms can be easily parallelised since each individual (trading strategy) builds and evaluates a candidate solution independently from all other individuals in the population. Therefore, a large speed up could be obtained by running a parallel version of any evolutionary DC version, as it has actually been shown in Brookhouse et al. (2014), where speed ups of up to 21 times were observed. 6.4. Discussion From the above results, we can reach the following conclusions. DC has the potential of returning profitable trading strategies. The single and multi-threshold DC strategies (SDC and MDC) were able to return profitable strategies, as it is evident from the best results of Tables 3 and 6. As we can observe in these tables, all DC algorithms had the maximum return entry as a positive value, which indicates that there was at least one instance per algorithm that had yielded positive return. However, the SDC and MDC paradigm could not consinstently return profitable strategies and thus their mean returns were negative. So it was evident to us that while DC is a promising method, it would benefit from optimising its parameters. Optimising DC parameters and weights increases the mean return. Using a genetic algorithm to optimise the parame- ters SDC and MDC increased the mean return, leading to a positive mean return in 3 out of 4 cases. More specifically, under the tick data (Table 3) moving from MDC to MDCEV O increased the mean daily return from -0.0092% to 0.0677%. In addition, under the 10-minute data (Table 6), moving from SDC to SDCEV O led to an increase of mean return from -0.00006% to 0.01064%. Similarly, moving from MDC to MDCEV O increased the mean return from -0.00015% to 0.00875%. Furthermore, optimising the weights of MDCEV O led to the development of the DC+GA algorithm. DC+GA fur- ther improved the mean return of MDCEV O under the tick data, from 0.0677% to 0.0730%. Under the 10-minute data, DC+GA again improved the mean return from MDCEV O’s 0.00875% to DC+GA’s 0.01046%. However, DC+GA had slightly lower mean return from SDCEV O’s 0.01046%. Nevertheless, statistical tests showed that the difference in DC+GA’s performance and SDCEV O were not statistically significant. The above leads to us conclude that the introduction of both parameter and weight optimisation is beneficial to DC algorithms. The DC paradigm is able to outperform traditional physical-time financial benchmarks. The third and last conclusion we can reach is that the DC paradigm is able to outperform traditional physical-time financial benchmarks, such as buy and hold (BH) and technical analysis (EDDIE). In fact, EDDIE never managed to yield positive returns under the experiments in this work, even though in the past has been very successful in similar financial problems (Kampouridis & Otero, 2015; Kampouridis & Tsang, 2012, 2010). On the other hand, BH yielded a negative return under the tick data and a positive return under the 10-minute data. In addition, DC+GA significantly outperformed EDDIE under 25 the tick data at the 5% level, and SDCEV O significantly outperfomed EDDIE under the 10-minute data at the 10% level. However, SDCEV O is heavily dependent on the single threshold we choose to use, so while sometimes it can perform very well (10-minute data), some other times it can perform extremely poorly (tick data). We made a similar observation for SDCEV O’s mean MDD value under the EUR/JPY (Table 8), where it almost doubled the MDD value to its competitors MDCEV O and DC+GA. For this reason, we believe that DC+GA is a better algorithm, as it is more robust, i.e., it is more consinstent in terms of positive returns and lower MDD. Hence, for completeness we also run a Kolmogorov-Smirnov test between DC+GA and EDDIE, under the 10-minute data. The null hypothesis is that they both come from the same continuous distribution. The p-value of the test is 0.0560, which just misses rejecting the hypothesis at the 5% level, but does reject it at the 10% level. So, DC+GA is able to significantly outperform EDDIE at the 5% level under tick data and at the 10% level under the 10-minute data. In addition, DC+GA returned higher mean return than BH under the tick data, and a similar mean return under the 10-minute data. We can thus conclude that DC+GA is able to perform at least as well as BH and outperforms EDDIE. To summarise, the above three conclusions demonstrate that we have successfully met the the three goals of this paper: (i) the DC paradigm returns profitable strategies, (ii) optimising DC strategies by a GA leads to an increase in profits, and (iii) our proposed algorithm, DC+GA, is able to outperform physical-time financial strategies, such as technical analysis, and buy and hold. 7. Conclusion To conclude, this paper used a new way of summarising high-frequency foreign exchange data, and combined it with a genetic algorithm for optimising its parameters. We used our proposed framework to trade in six different FX markets, and showed that we are able to not only produce average profitable results, but also outperform benchmarks coming from two traditional physical-time approaches (technical analysis, buy and hold). We believe that these results constitute a very promising start, and that further research should take place towards this direction. More specifically, at the moment, we have focused on the core theory of directional changes and we have derived strategies based on that theory. More work could take place in defining new indicators, derived from the concept of directional changes, in a similar manner that technical analysis indicators exist with physical time. In addition, in our current approach we allowed for the generation of multiple thresholds, and then let the GA combine the suggested action of each threshold. A potential improvement to this would be to leave the decision of the generation of thresholds completely to the optimisation algorithm. For example, the Genetic Algorithm could be further extended to not only generate thresholds at the beginning of each evolutionary process, but also mutate them, thus generate new ones, during the process. This is by far a more dynamic technique, which could lead to even better trading results. Lastly, we also plan to test our DC+GA algorithm on more data sets from the FX market and also from other type of markets, for instance the stock market. Appendix A. Monthly return results for EDDIE and SDCEVO References Allen, F., & Karjalainen, R. (1999). Using genetic algorithms to find technical trading rules. Journal of Financial Economics, 51, 245–271. Aloud, M., Tsang, E., Olsen, R., & Dupuis, A. (2012). A directional change events approach for studying financial time series. Economics, 36, (online). Austin, M., Bates, G., Dempster, M., Leemans, V., & Williams, S. (2004). Adaptive systems for foreign exchange trading. Quantitative Finance, 4(4), 37–45. Azzini, A., da Costa Pereira, C., & Tettamanzi, A. G. B. (2010). Modeling turning points in financial markets with soft computing techniques. In A. Brabazon, M. O’Neill, & D. G. Maringer (Eds.), Natural Computing in Computational Finance (pp. 147–167). Berlin, Heidelberg: Springer Berlin Heidelberg. Binner, J., Kendall, G., & Chen, S.-H. (Eds.) (2004). Applications of Artificial Intelligence in Finance and Economics volume 19 of Advances in Econometrics. Elsevier. Birattari, M., Yuan, Z., Balaprakash, P., & Stutzle, T. (2009). F-race and iterated f-race: An overview. Technical Report TR/IRIDIA/2009-018, IRIDIA, Université Libre de Bruxelles, Belgium. Brabazon, A., & O’Neill, M. (2006). Biologically inspired algorithms for financial modelling. Springer. Bradley, R., Brabazon, A., & O’Neill, M. (2009). Dynamic high frequency trading: a neuro-evolutionary approach. In M. Giacobini, & A. Brabazon (Eds.), Applications of Evolutionary Computing: EvoWorkshops 2009, LNCS 5484 (pp. 233–242). Springer. 26 Table A.11: Monthly return results for EDDIE and SDCEV O. 10-minute interval data. Results presented per currency pair. Results shown in % values. EDDIE SDCEV O EUR/GBP August 0.004458 -0.005459 September 0.005456 -0.004532 October -0.013644 0.002058 November 0.006171 0.002844 December -0.011072 0.013137 January -0.010316 -0.000073 February -0.001960 0.000962 March 0.007997 -0.002265 April 0.000799 0.000304 May -0.002026 0.000892 EUR/JPY August -0.513460 0.155999 September -0.170211 0.050100 October -0.088246 0.060285 November 0.101843 0.000000 December 0.613728 -0.421926 January -0.548754 0.771926 February 0.033461 -0.006850 March 0.400234 -0.118628 April 0.048707 0.049121 May -0.041675 0.040874 EUR/USD August -0.026099 0.001232 September -0.004848 0.001880 October -0.053083 -0.007077 November 0.004349 -0.000073 December -0.0004962 -0.005356 January -0.008512 0.014046 February 0.011698 -0.000601 March -0.000885 0.003289 April -0.006766 -0.006851 May 0.000674 0.000756 GBP/CHF August -0.019175 -0.004584 September 0.002983 -0.014025 October -0.036008 0.002891 November -0.030220 -0.002861 December -0.019675 0.000477 January 0.007053 0.000968 February 0.009592 0.009052 March -0.008943 -0.027585 April -0.002353 0.000696 May -0.014646 0.005771 GBP/USD August 0.000832 0.000249 September -0.009259 -0.014953 October 0.002542 0.028377 November 0.001273 0.000013 December -0.020717 -0.030266 January 0.007417 -0.001725 February -0.005826 -0.012477 March -0.029045 -0.002918 April -0.002988 -0.005678 May -0.007080 0.010379 27 Brookhouse, J., Otero, F., & Kampouridis, M. (2014). Working with OpenCL to speed up a genetic programming financial forecasting algorithm: Intial results. In Proceedings of the Companion Publication of the 2014 Annual Conference on Genetic and Evolutionary Computation (GECCO) (pp. 1117–1124). ACM. Cervelló-Royo, R., Guijarro, F., & Michniuk, K. (2015). Stock market trading rule based on pattern recognition and technical analysis: Forecasting the DJIA index with intraday data. Expert Systems with Applications, 4 (14), 5963–5975. Chen, S.-H. (2002). Genetic Algorithms and Genetic Programming in Computational Finance. Springer-Verlag New York, LLC. Chen, T. L., & Chen, F. Y. (2016). An intelligent pattern recognition model for supporting investment decisions in stock market. Information Sciences, 346, 261–274. Chiang, W. C., Enke, D., Wu, T., & Wang, R. (2016). An adaptive stock index trading decision support system. Expert Systems with Applications, 59, 195–207. Chourmouziadis, K., & Chatzoglou, P. D. (2016). An intelligent short term stock trading fuzzy system for assisting investors in portfolio manage- ment. Expert Systems with Applications, 43, 298–311. Demšar, J. (2006). Statistical Comparisons of Classifiers over Multiple Data Sets. Journal of Machine Learning Research, 7, 1–30. Dupuis, A., & Olsen, R. (2012). High frequency finance: using scaling laws to build trading models. In J. James (Ed.), Handbook of exchange rates chapter 20. Wiley, NJ, USA. Evans, C., Pappas, K., & Xhafa, F. (2013). Utilizing artificial neural networks and genetic algorithms to build an algo-trading model for intra-day foreign exchange speculation. Mathematical and Computer Modelling, 58, 1249–1266. García, S., & Herrera, F. (2008). An Extension on “Statistical Comparisons of Classifiers over Multiple Data Sets” for all Pairwise Comparisons. Journal of Machine Learning Research, 9, 2677–2694. Garcia-Almanza, A., Alexandrova-Kabadjova, B., & Martinez-Jaramillo, S. (2013). Bankruptcy prediction for banks: An artificial intelligence approach to improve understandability. In X.-S. Yang (Ed.), Artificial Intelligence, Evolutionary Computing and Metaheuristics (pp. 633–656). Springer Berlin Heidelberg volume 427 of Studies in Computational Intelligence. Glattfelder, J., Dupuis, A., & Olsen, R. (2011). Patterns in high-frequency FX data: Discovery of 12 empirical scaling laws. Quantitative Finance, 11 (4), 599–614. Godlberg, D. (1989). Genetic Algorithms in Search Optimisation and Machine Learning. Addison-Wesley. Gypteau, J., Otero, F. E., & Kampouridis, M. (2015). Generating directional change based trading strategies with genetic programming. In M. A. Mora, & G. Squillero (Eds.), Applications of Evolutionary Computation: 18th European Conference (pp. 267–278). Springer. Hu, Y., Liu, K., Zhang, X., Su, L., Ngai, E., & Liu, M. (2015). Application of evolutionary computation for rule discovery in stock algorithmic trading. Appl. Soft Comput., 36, 534–551. International Monetary Fund (2009). Global financial stability report. https://www.imf.org/external/pubs/ft/gfsr/2009/01/pdf/text.pdf. Jaisinghani, D. (2016). An empirical test of calendar anomalies for the indian securities markets. South Asian Journal of Global Business Research, 5 (1), 53–84. Kampouridis, M., & Otero, F. E. B. (2015). Heuristic procedures for improving the predictability of a genetic programming financial forecasting algorithm. Soft Computing, (pp. 1–16). Kampouridis, M., & Tsang, E. (2010). EDDIE for investment opportunities forecasting: Extending the search space of the GP. In Proceedings of the IEEE World Congress on Computational Intelligence (pp. 2019–2026). Barcelona, Spain. Kampouridis, M., & Tsang, E. (2012). Investment opportunities forecasting: Extending the grammar of a GP-based tool. Int. Journal of Computa- tional Intelligence Systems, 5, 530–541. Kim, Y., & Enke, D. (2016). Developing a rule change trading system for the futures market using rough set analysis. Expert Systems with Applications, 59, 165–173. Koza, J. (1992). Genetic Programming: On the programming of computers by means of natural selection. Cambridge, MA: MIT Press. Kwon, Y.-K., & Moon, B.-R. (2007). A hybrid neurogenetic approach for stock forecasting. IEEE Transactions on Neural Networks, 18, 851–864. Lopez-Ibanez, M., Dubois-Lacoste, J., Stutzle, T., & Birattari, M. (2011). The irace package, iterated race for automatic algorithm configuration. Technical Report TR/IRIDIA/2011-004, IRIDIA, Université Libre de Bruxelles, Belgium. Mani, S. (1996). Financial forecasting using genetic algorithms. Applied Artificial Intelligence, 10, 543–566. Martinez-Jaramillo, S. (2007). Artificial Financial Markets: An agent-based Approach to Reproduce Stylized Facts and to study the Red Queen Effect. Ph.D. thesis CFFEA, University of Essex. Michalewicz, Z. (2002). Genetic algorithms + data structures = evolution programs. Springer. Mitchell, M. (1996). An Introduction to Genetic Algorithms. MIT Press, Cambridge, MA. Olsen, R. B., Muller, U. A., Dacorogna, M. M., Pictet, O. V., Dave, R. R., & Guillaume, D. M. (1997). From the bird’s eye to the microschope: A survey of new stylized facts of the intra-day foreign exchange markets. Finance and Stochastics, 1 (2), 95–129. Tsang, E., Markose, S., & Er, H. (2005). Chance discovery in stock index option and future arbitrage. New Mathematics and Natural Computation, World Scientific, 1(3), 435–447. Tsang, E., & Martinez-Jaramillo, S. (2004). Computational finance. IEEE Computational Intelligence Society Newsletter, (pp. 3–8). Tsang, E., Tao, R., Serguieva, A., & Ma, S. (2016). Profiling high-frequency equity price movements in directional changes. Quantitative Finance, 2016, (online). 28 
work_p5jfo4uolzgablpiooecwk2xu4	----	Software maintenance project delays prediction using Bayesian Networks Software maintenance project delays prediction using Bayesian Networks Ana C.V. de Melo a,*, Adilson J. Sanchez b a University of São Paulo (USP), Department of Computer Science, Rua do Matão, 1010, Cidade Universitária, 05508 090 São Paulo, SP, Brazil b Banco Itaú, Av do Estado, 5533 – Mooca, 03105 000, São Paulo, SP, Brazil Abstract Managing software maintenance is rarely a precise task due to uncertainties concerned with resources and services descriptions. Even when a well-established maintenance process is followed, the risk of delaying tasks remains if the new services are not precisely described or when resources change during process execution. Also, the delay of a task at an early process stage may represent a different delay at the end of the process, depending on complexity or services reliability requirements. This paper presents a knowledge-based representa- tion (Bayesian Networks) for maintenance project delays based on specialists experience and a corresponding tool to help in managing software maintenance projects. � 2006 Elsevier Ltd. All rights reserved. Keywords: Bayesian Belief Networks; Software management; Project delay estimation 1. Introduction The advances in technologies and competitiveness of business products made computational support essential to have products in time and with a desirable quality to the market. Computational development became part of business strategy, instead of an auxiliary development support. One of the reasons that makes computational support strategic to business products rests on the fact that business rules are continuously changing and this can be acquired by computational systems. Changing business rules and the corresponding system implies in adapting existing sys- tems to accomplish new rules, the so called adaptive main- tenance (Pressman, 2004). In this scenario, software maintenance process is essential to business products com- petitiveness to the market. Despite of being widely studied and of interest to mar- ket, software maintenance is still a complex and costly task. Maintenance is pointed out as one of the most expensive tasks in software development and the difficulties inherent to its execution is the main cause of software high costs. Having no maintenance plans or, even when plans are established, having the plan schedule not achieved lead to cost-ineffective maintenance projects (Pressman, 2004). That situation, however, does not come from no effort in trying to amend maintenance problems. Many people have employed a lot of work in developing new software devel- opment and management techniques to make software maintenance cost-effective. Even so, uncertainties in the maintenance process drives to unpredictable situations hard to be recovered during project execution (Ziv, 1997; Ziv, Richardson, & Klosch, 1996). Ordinary uncertainty situations in maintenance processes occur, such as: • Maintenance plans represent only starting and end time of tasks, with no procedures to be applied as tasks are delayed in the middle of the project execution. 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.10.040 * Corresponding author. Tel.: +55 11 30919689; fax: +55 11 30916134. E-mail addresses: acvm@ime.usp.br (A.C.V. de Melo), Adilson.San- chez@itau.com.br (A.J. Sanchez). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 34 (2008) 908–919 Expert Systems with Applications mailto:acvm@ime.usp.br mailto:Adilson.Sanchez@itau.com.br mailto:Adilson.Sanchez@itau.com.br • In general, information related to system restrictions and services are treated in different ways. Users want as many services as possible, even when they are not nec- essary to them. On the other hand, developers have to implement all services without a precise idea of their priorities. • Time to complete tasks in the maintenance schedule is planned accordingly to a set of resources. Such resources may change through project development (the main pro- ject analyst is no longer with the company, for example), causing a delay in certain tasks. Due to these uncertainties and others, project managers have to decide how to reconfigure the project plan as its schedule fails. One of the problems in replanning a project to achieve its schedule and services requirements rests on systems development information. What to do to finish the system in time? How much resource should be added in order to finish a delayed project in time? Can we imple- ment the system in a shorter time without dropping its quality requirements? These and many other issues must be answered when a project is replanned and, in general, managers use their own experience to do that. Such an approach is very human-intensive and based on particular experiences – too subjective. A more objective approach can be supported by a decision system if experience of spe- cialists and historical data of past projects are embedded in such a system. Current project risk management processes have given a restricted focus on management of project uncertainties. The importance of treating uncertainties in project devel- opments has already been recognised in the fields of project management (Chapman & Ward, 2000; Green, 2001; Ward & Chapman, 2003) and software development (Khodaka- rami, Fenton, & Neil, in press; Ziv et al., 1996). In both areas, uncertainties are pointed out as one of the causes of having delayed, unsuccessful and low-quality projects. In all those works, treating uncertainties and their main causes during project execution has been suggested as a way of having successful projects. Several tools and methods have been proposed in order to support expert judgement. To produce such tools, a kind of knowledge about domains must be embedded, and Bayesian Belief Networks (Jensen, 1996; Pearl & Rus- sell, 2003; van der Gaag, 1996) (Bayesian Networks for short) appear as a reasonable representation of uncertain- ties and knowledge. Roughly speaking, Bayesian Networks are cause-effect graphs, based on Bayes infer- ence, that can model uncertainty and have been success- fully deployed in certain domains such as medicine and politics. More recently, Bayesian Networks have also been applied to deal with uncertainties in software develop- ment. Stamelos, Angelis, Dimou, and Sakellaris (2003) show the construction of Bayesian Networks regarding technical and human factors in software development to estimate software productivity, by explicitly modelling uncertainties of factors influencing the productivity. Fan and Yu (2004) present a scheme to incorporate Bayesian Networks in software project risk management to predict potential risks, identify sources of risks and advise dynamic resource allocation. Fenton, Krause, and Neil (2002), Fenton et al. (2004) and Fenton et al. (2007) show how quality metrics incorporated to predictive models (Bayesian Networks) can provide accurate prediction of software defects in a variety of development processes. Bayesian Networks have also been used to estimate costs of object oriented software, as in Bibi and Stamelos (2004). There, a set of Bayesian Networks are defined, concerned with the Rational Unified Process (RUP), to depict software process modelling along with software effort estimation. In Khodakarami et al. (in press), the application of Bayesian Networks enables the incorpora- tion of risk, uncertainty and causality in project planning to identify specific risks and indicate their corresponding resources. Once resources of risk are identified, replanning projects (removing damaging resources) during execution can be performed. Ziv (1997) proposes a Bayesian Net- work to deal with the uncertainties of software testing. In the present work, a technique to support the prediction of software maintenance process delay during project exe- cution is presented in order to help managers in replan- ning projects. To help in managing maintenance project delays, this paper presents a technique to represent specialists knowl- edge about project development regarding delays. The main goal is to be able to calculate the project delay prob- ability during its execution. Managers can use this informa- tion to replan a project in order to finish it in time. The forthcoming sections are as follows: Section 2 summarises Bayesian Belief Networks; Section 3 presents a particular Bayesian Network for the delay propagation problem; Sec- tion 5 gives guidelines of a tool to help managers to deal with maintenance project delays based on the Bayesian Network of the previous section; and the last section is about limitations of the present work and the effective use of this technique in software development. 2. Decision systems and Bayesian Networks Decision systems are concerned with a way of embed- ding knowledge in order to make a mimic of human deci- sions (Jackson, 1990; Lucas & van der Gaag, 1991; Russell & Norvig, 2003). This involves how to represent knowledge and use such an information to make decisions. Those systems, the so called expert systems, are for solving problems of particular domains and the knowledge embed- ded in each domain is acquired from both domain special- ists and systems development information. Expert systems have embedded a knowledge base and techniques to obtain (infer) new knowledge, or take decisions, from the existing base. One of the challenges in experts systems is how to acquire and represent knowledge as accurate as possible. If this is the case, the system can take decisions as close as a domain specialist. A.C.V. de Melo, A.J. Sanchez / Expert Systems with Applications 34 (2008) 908–919 909 https://isiarticles.com/article/28614 
work_pa4cke2pdrde5hq57vbfuw75ui	----	untitled Multi-agent system for knowledge-based event recognition and composition Angel Rivas-Casado,1 Rafael Martinez-Tomás1 and Antonio Fernández-Caballero2 (1) Dpto. Inteligencia Artificial. Escuela Técnica Superior de Ingenierı́a Informática, Universidad Nacional de Educación a Distancia, Juan del Rosal 16, 28040 Madrid, Spain Email: arivas@dia.uned.es (2) Dpto. Sistemas Informáticos, Escuela Técnica Superior de Ingenierı́a Informática, Universidad de Castilla-La Mancha, Albacete, Spain Abstract: This work presents a multi-agent system for knowledge-based high-level event composition, which interprets activities, behaviour and situations semantically in a scenario with multi-sensory monitoring. A perception agent (plurisensory agent and visual agent)-based structure is presented. The agents process the sensor information and identify (agent decision system) significant changes in the monitored signals, which they send as simple events to the composition agent that searches for and identifies pre-defined patterns as higher- level semantic composed events. The structure has a methodology and a set of tools that facilitate its development and application to different fields without having to start from scratch. This creates an environment to develop knowledge-based systems generally for event composition. The application task of our work is surveillance, and event composition=inference examples are shown which characterize an alarming situation in the scene and resolve identification and tracking problems of people in the scenario being monitored. Keywords: knowledge-based systems, high-level vision, image understanding, video surveillance, event-base system, intelligent agent, plurisensorial agent, Arduino 1. Introduction This work presents a multi-agent system for knowledge-based high-level event composition, which interprets activities, behaviour and situa- tions semantically in a scenario with multi- sensory monitoring. The system’s specific aims would be (Bobick, 1997) to monitor the complex scenario, using all the sensors’ capacities (Car- mona, 2009) to interpret the information from these sources as a whole, and (Chleq & Thonnat, 1996) to request further information to confirm or reject the hypotheses raised in the process for diagnosing the situations. The semantic interpretation of scenes must recognize situations, activities and interactions between the different actors. The physical sig- nals captured by the sensors must be linked with the interpretation of their meaning. When a human observer interprets the meaning of a scene, he uses his knowledge of the world, the behaviour of the things that he knows, the laws of physics and the set of intentions governing the agents’ activity. All this additional knowl- edge enables the observer to model the scene and use this model to predict, at least partially, what may happen. In other words, to launch a DOI: 10.1111/j.1468-0394.2010.00578.x Article _____________________________ c� 2011 Blackwell Publishing Ltd Expert Systems 1 mailto:arivas@dia.uned.es mailto:arivas@dia.uned.es mailto:arivas@dia.uned.es hypothesis about the evolution of the events and activities that he is detecting. The aim is to aid the human operator so that he can take objec- tive, consistent real-time decisions about the events detected. Multi-sensory interpretation combines multiple sources of information from different sensors in order to generate a more exact and robust inter- pretation of the environment (Pavón et al., 2007). The interpretation of behaviour and situations is currently based on using several sensors that offer a high number of false alarms. The availability of new types of sensors for monitoring tasks pro- vides new challenges for multi-sensory data fu- sion that they solve or they diminish this problem. It is now possible to construct models of the environment and diagnose situations by analys- ing indefinite sequences of sensory information from different types of sensors. The efficient fusion of multi-sensory information – understood as the fusion of data from several sensors of the same type or, additionally, as the fusion of data from several different types of sensors – is a crucially important task in advanced surveillance. Thus, in recent years there has been a rapid increase in research and development into the combined use of multiple sensors, including vi- sion, audio, heat (thermal), presence (volumetric), vibration sensors, etc. (Zhu & Huang, 2007). A multi-sensory, monitoring system usually re- quires three components of equal importance (Zhu & Huang, 2007): (1) sensors that capture the information for the monitoring system, pro- cess and interpret it; (2) fusion algorithms of data from each of the sensors; (3) architectures for constructing real-time systems. Of all the sensors, the visual ones are ob- viously very important. High-level vision is defined as the interpretation of scenes beyond the mere recognition of objects (Fuentes & Velastin, 2004). Image understanding (image comprehension=interpretation) generally starts by analysing movement and then describes the scene with symbols (Tostsos et al., 1980; Levine et al., 1983). It is widely accepted that it is necessary to inject the expert’s knowledge to complement the low-level information (Neumann & Novak, 1983; Neumann, 1984). In a constructivist model, the event composition is done with spatio-temporal relations, with parallel processes that determine the time inter- vals of other events. Regarding the diagnosis of situations, in Chleq and Thonnat (1996) the hypothesis approach is included, which implies doing parallel explorations for alternative solu- tions. The confirmation of a hypothesis in a description level helps the previous levels. In Rota and Thonnat (2000), the use of declarative models for representation is described. Our work follows this line, constructing declaratory models from the skill of a human expert who knows how to identify critical situations, parti- cularly in surveillance monitoring problems. Surveillance is a perfect application field for both plurisensory monitoring and interpreting video-sequence scenes, where our group has accumulated considerable experience with AVI- SA project works (Folgado et al., 2007; Martı́- nez-Tomás et al., 2008; Carmona, 2009). It is also a multi-disciplinary task affecting an increasing number of scenarios, services and clients. In particular, the use of agents in surveillance systems has some precedents in the bibliography (Abreu et al., 2000; Remagnino et al., 2004; Haesevoets et al., 2007) and the multi-sensory surveillance works by Castanedo et al. (2008). In the cooperative sensor agent (CSA) architecture proposed in Molina et al. (2004) coalitions are formed between agents to perform surveillance tasks. The CSA is broken down into two levels: sensor layer and coalition layer. A coalition is formed when an agent (sensor) needs to cooperate with other agents that have capacities that it does not have. The article is structured as follows: the system structure, its organization into vision agents (VAs) and plurisensory agents (PAs), and the high-level composition agent (CA) are in Section 2. The perception agents (VAs and PAs) identify significant changes in the sensors’ magnitudes and transmit them as simple events to the CA. From these simple events, the CA composes composed events that imply higher-level semantic activities. The system is described and an application exam- ple is given that identifies the abandoning of objects as an alarming situation. Section 3 2 Expert Systems c� 2011 Blackwell Publishing Ltd describes the CA as a knowledge-based system and the structure of its knowledge base. The methodology for creating, updating and modify- ing the base for different application fields is also described. The article finishes by describing the system’s application for monitoring people in different visual fields with different cameras, therefore just visual information. It also describes the problem of access control where plurisensory information is already being used for precise identification and tracking. 2. Intelligent agents of the system The term intelligent agent in artificial intelli- gence refers to any entity that can take decisions from its environment (Russell & Norvig, 2009). As shown in Figure 1, the system schema pro- posed consists of one or several PAs, one or several image interpretation agents (IIAs) and in theory, just one CA. Both the PAs and the IIAs process the sensor information and identify (agent decision system) significant changes in the monitored signals, which the connection interface translates into simple events and sends to the CA via a network connection system. For an efficient system, the agents must be carefully selected and placed by an expert in the best positions of the environ- ment to be monitored. 2.1. Composition agent It is the high-level knowledge-based synthetic software agent that identifies activities or situa- tions (composed events) as a composition from simple events that meet certain spatio-temporal restrictions. We can distinguish between two types of functionalities: the reception of simple events generated by the PA and IIA and the composition from these simple events. Defining the event patterns (composed events) that char- acterize the target activities is an analytical task of knowledge acquisition by a human expert who knows a behaviour pattern’s simple actions perfectly. The actions correspond to significant high-level activities and have associated replies. For this, a methodology was created to develop the knowledge base. The generation of an event from the PA and VA implies the instantiation of the corresponding generic event, where values are given to its attributes. For example, At(h, x, y, t) is instantiated at At((h ‘Peter’)(x 150)(y 120)(t 430)) when Peter is identified as at posi- tion (150, 120) at instant 430. For the composi- tion, the events are included as part of the facts base. The scenario ontology also includes facts describing its inherent characteristics, where the activities occur: doors, windows, passages, no entry areas, adjacent cameras, etc. Figure 2 schematically illustrates an example of an alarming situation. A person leaves an object on the ground and goes away. The first row includes simplified images in the scene instants. The second row contains the simple events that are generated from segmenting, monitoring and identifying each frame of the sequence. The third row shows the pattern for the composition (composition unit) of events occurring at each instant. It is a knowledge unit for a composition: a set of events that must meet specific spatio- temporal relations and the consequent events with higher-level semantics. In the event base there is a previous ‘Walking’ event that com- pletes the pattern. The forth row shows the higher-level event inferenced by this way. ‘Walk- ing’ belongs to a special type of event that can be called ‘event-state’ that occur over a time period t1–t2, unlike ‘instant-event’. At, for example, occurs at a given instant t. Thus, following the example of Figure 2 we can interpret the sequence as follows, step by step: 1. At instant t, human1 is detected on the scene. The identification and monitoringFigure 1: Agent interconnection schema. c� 2011 Blackwell Publishing Ltd Expert Systems 3 agent does not recognize that the human is carrying an object, so the spatio-temporal location of the human is only represented with the event At in that instant t. With this At the event ‘Walking2 is updated to the following instant. 2. At instant tþ1, ‘object1’ is detected near to the position of human1. This situation acti- vates a pre-alarm of possible abandonment of an object with the event ‘Pre-alarm’. Since there are no other humans nearby, it is inferred that ‘human1’ has left the object. An association is created between the object and human, and the pre-alarm is activated. 3. At instant tþ2, ‘human1’ is detected leav- ing the scene. This event and the active pre- alarm identify a situation of abandoning an object. The event ‘Alarm’ goes off. We group composition units into packages that identify a specific situation. In turn, the pack- ages are organized into composition levels, each package is assigned to a composition level. Each composition level sends the composed events that it has generated to their higher composition level. Packages in different composition levels are inter-dependent. Thus, if a package in a specific level is added to the knowledge base, all those packages in the lower composition levels, which are necessary for its functionality, will be added. Our aim is to give the possibility to create libraries of packages rich enough to configure a system easily. Each library of packages has its own corresponding event ontology. 2.2. Plurisensory agent The incorporated systems provide information about the environment where they are located. They can take a high number of readings per second to pinpoint significant alterations. Gen- erally, they perform very simple operations. They can take, analyse and send samples from several sensors. Figure 3 shows a multi-sensory system consisting of a shield ethernet connected to an Arduino plate with an Atmel AVR 8bits Figure 2: Recognition that a person has abandoned an object. Figure 3: Multi-purpose multi-sensory system. 4 Expert Systems c� 2011 Blackwell Publishing Ltd processor. This system can capture, analyse and send information from the sensors once the information has been filtered by the local area network (LAN). At present, as shown in Figure 4, work is being done on integrating into the multi- purpose multi-sensory system different sensors that are particularly relevant for surveillance problems: movement sensors, proximity sen- sors, lighting sensors, temperature sensors, rela- tive humidity sensors, open door and window detector, environmental noise sensors, radio frequency identification (RFID)-based identifi- cation systems. This multi-purpose multi- sensory system provides a PA with several advantages: its low cost and easy reproduction, it provides reliable, rapid information and final- ly, it can be used to monitor private rooms, where for legal reasons, video-surveillance sys- tems cannot be installed. We are currently studying the possibility of incorporating IIAs into the system taking advantage of the infor- mation received from the PA improve the event segmentation and focusing processes (VA), like, for example, detecting scene lighting changes or doors opening in the visual field. 2.3. Vision agent This agent combines a system of perceiving images with vision software that requires high computational performance for near real-time tasks. Its main functions are those of a vision system: segmentation, identification and track- ing of people and=or objects within a monitored visual field (Carmona, 2009). For the identifica- tion process, a block-based system is used (Fol- gado et al., 2007) developed by our group, the same as the level structure to identify events (Martı́nez-Tomás et al., 2008). The connection interface can identify a battery of pre-designed events, which are not specific to the application field, but are especially relevant for identifying (composing) the significant high-level events related to the scene events, as shown in Martinez Tomás and Rivas Casado (2009). Section 5 shows an example. Figure 4: Multi-sensory system schema. c� 2011 Blackwell Publishing Ltd Expert Systems 5 3. Knowledge base structure (KBS) Figure 5 shows the internal structure of the knowledge base. First of all, we differentiate between levels that correspond to abstraction levels, from knowledge to identify more precise activities, more directly associated with mechan- ical actions or movements, to more abstract activities, which define behaviour or situations. The packages are sets of rules (as mechanisms for minimal representation of composition units) that identify specific situations. The con- cept of package is really important in this schema since the system can add or eliminate these from a knowledge base. Thus there is package inter-dependency. If a package is elimi- nated from an upper level, all the lower-level packages providing it with information will be automatically eliminated. Obviously, when a package is eliminated, all its composition units are also eliminated. Each package has a work environment defined by the knowledge level to which it belongs, by the input events and the events that it generates. Each package contains one or more rules. The rules only have access to the events defining the package. Thus each rule is encapsulated within the best environment. Errors and possible crossing of information between different packages are also prevented. The advantage of this structure compared with a traditional KBS is that it is easy to maintain and reconfigure when changes occur in the environment. In each level, the rules are prior- itized for their execution. The most important rules are the ones that process the information from the previous level and the least important are those that generate the information that is transferred to the upper level. This process is performed for each knowledge level as a pipe- line. The composed events that level N generates at instant T are processed in level Nþ1 at instant Tþ1. 3.1. Composition unit In the prototypes that we have developed, the composition of events is represented on rules, but generically we can speak of composition unit. It is regarded as the unit representing the system’s knowledge. It consists of three parts: � Antecedents: It is the event pattern that must be met to be able to check the composition conditions. � Conditions: It is a filter that must be passed to be able to execute the high-level event composition. � Actions: Once all the tests have been passed, the corresponding changes are made in the event base. New events can be created or one or other of the existing events can be eliminated. To create new events, the event data are taken that are the composition unit antecedents. Similarly, only events can be eliminated that belong to the antecedents. This mechanism is able to fuse simple events into more complex events. In the example below, we can see two events, a simple one and a composed one. The aim is to update the composed event information with the simple Figure 5: Breakdown of the knowledge base. Each knowledge level (abstraction level) consists of a number of packages that in turn have rules. Each rule implements, as composition unit, one or several high-level events pattern. 6 Expert Systems c� 2011 Blackwell Publishing Ltd event information: Walkingðh;x1;y1; t1;x2;y2; t2Þ Atðh;x;y; tÞ The event At shows where a human h is at instant t in the position x, y. The event ‘Walk- ing’ represents at what time and position he began to walk to what time and position he stopped walking. Now the example is presented with the instantiated events: In the first event it is observed that the human ‘‘Peter’’ is at position x¼200, y¼50 and t¼500. The conditions would be that the hu- mans were the same person and that instant t2 of the event ‘Walking’ were less than instant t of the event At. With these data, the second event could be updated to Walkingððh PeterÞðx1 100Þðy1 50Þ ðt1 450Þðx2 200Þðy2 50Þðt2 500ÞÞ We now have the event updated to the instant 500, so we can affirm that ‘Peter’ has advanced 100 pixels to the right in 50 instants. After updating all the ‘Peter’ dependent events, the simple event could be erased before executing the following instant. The knowledge base is structured as an ab- straction pyramid (Figure 6). At the bottom is the knowledge level supported by the simple events, which arrive from the PAs and IIAs, and at the top of the pyramid are the upper knowledge levels. To do the event composition in level n (for n>1), a hypothesis is made, which searches the events verifying this hypothesis in level n�1. If the hypothesis’s restrictions are met, the new composed event is generated in level n. Configuration parameters are global variables that can be accessed by all the rules. They store a specific value, usually, numerical. They are used to be able to configure the knowl- edge base in different environments. A video camera’s resolution, for example, the ontology terms like ‘Near’ and ‘Far’ act as configuration parameters. 4. Methodology and development tools One of the AVISADOS project’s aims was to use it in different application fields to facilitate the configuration for new scenarios, always with the structure described in Sections 2 and 3. For this, a number of tools and similar methodology were developed, which means that each new application does not start from scratch. The Figure 6: Inference pyramid. Walkingððh PeterÞðx1 100Þðy1 50Þðt1 450Þðx2 190Þðy2 50Þðt2 499ÞÞ Atððh PeterÞðx 200Þðy 50Þðt 500ÞÞ c� 2011 Blackwell Publishing Ltd Expert Systems 7 result is a system that generates KBSs in general and interpretation KBSs based on the agent structure in particular. 4.1. Methodological and functional schema In a design stage, in Figure 7, the event ontology is configured. This will be used later to create agent interfaces and define the knowledge packages and units. Three stages can be identi- fied in the execution stage. In the first, the agents process the environment information and iden- tify the events corresponding to the actions detected. Then the agent interface sends the set of events to the execution and monitoring sys- tem (EMS) via a LAN. In the second stage, the events received are labelled with the correspond- ing instant in the EMS. Then the execution takes place. The inferred events are sent to the data- base to which the EMS is connected. The last stage is the analysis of the results by the user interface. All the agents are connected to their corresponding interface. This in turn is con- nected to the EMS. In the following sections, we focus in more detail on the system generating the agent interfaces and the knowledge base. 4.2. Knowledge base IDE The flowchart of Figure 8 shows the procedure for constructing a knowledge base and the agent interfaces. The development of a new knowledge base begins by defining the event attributes. Then we implement the simple and composed event ontology that the CAs, PAs and VAs will be able to identify (define events) and that we use to develop the agent profiles (define agent) and to create the knowledge packages (define package). An agent profile contains the characteristic data of this perception agent ‘identifier, position in the scenario, function, description’ and the collection of simple events that it can identify. This informa- tion is necessary to create the connection interface (generate agent interfaces), from the perception agent, which sends the simple events perceived to the CA. The package library is a repository of packages of identifiable situations previously im- plemented. It is not necessary to finish of defining (define package and construct rules) all packages to generate a knowledge base. Through a process of selection, you can configure the knowledge to address different situations (select identify situa- tion in package library and generate knowledge base). So, this process allows for reconfiguring a new system very quickly. Figure 9 shows the user interface of knowl- edge base IDE. The list of simple and compound events developed is shown. The left-side bar shows the information about the project and the selected item. This provides expert assistance in designing the knowledge base. The search for elements developed is one of the utilities included in the tool. On the right of the window we can see Figure 7: System schema. 8 Expert Systems c� 2011 Blackwell Publishing Ltd all elements that form the knowledge base. They are classified into six sections: agents interface, configuration parameters, parameters of the events, events, composition units and packages. Each package has an associated knowledge level and some simple and=or composed events with which to work. The expert, following the activity identification patterns, generates the right composition units so that the package functions correctly. For this the rule manager is used. First of all, the package must be defined for the rule. Each composition unit has an associated package from which it inherits the set of events to which it may refer in the antecedents, conditions and actions. 4.3. Execution system This tool executes the knowledge base created with the knowledge base IDE, stores all the composed events generated in a record and connects all the system’s agents. It consists of: � Network server: It activates the socket so that the agents connect to the system. The connection is bidirectional to be able to request certain information from the agent. � List of connected agents: A list is kept of all the agents that are connected to the system along with their description and related data. � Event synchronization system: It generates time slices to receive and label events. Once the event has been received, it is labelled with the time associated with the time slice. Thus an asyn- chronous system becomes a synchronous one. � Composition motor: Inference motor that evaluates the knowledge unit requirements and adds the inferred ones to the event base. � Statistic and monitoring system: It analyses the number of events that arrive from each Figure 8: Flowchart for developing a knowledge (centre), reconfiguration system and generate knowledge base (right). c� 2011 Blackwell Publishing Ltd Expert Systems 9 agent, the composition motor execution time, the synchronization system buffers and the database connection response times. This information is really useful when cali- brating and configuring all the system’s ex- ecution parameters. � Database connection to store the information: It sends all the events inferred at each instant to the selected database. The information can then be processed later. When these data are analysed, possible errors can be de- bugged from the knowledge base, which is really useful for refining the system. Figure 10 shows the EMS. This window shows the user the state overall system: server status, Figure 9: Knowledge base IDE. Figure 10: Execution and monitoring system. 10 Expert Systems c� 2011 Blackwell Publishing Ltd connection port, number of clients, state of the system timer, state of inference engine, state of the statistics and the state connection with the database. The red cylinder represents the execution time instant and the green cylin- der, average execution time. The bottom bar shows number of the events that are currently on the basis of facts. The left bar buttons are used to change the status of different system functions monitoring and enforcement. The buttons on the top bar show individual from each of the elements system. The upper-left button, displays the system preferences: configuring the connection to the database, maximum number of customers, cycle time im- plementation, connection port, maximum size of the base facts and the selector of the knowl- edge base. 5. Resolving problems in identifying people and tracking Section 2 showed identifying an abandoned object as an application to recognize an alarm- ing situation. In this section, a solution is shown for identifying and tracking people, first with adjacent cameras and then with the support of precise identification from the access control. Both of them are application examples of our event compositions to solving problems for the lower levels events. 5.1. Adjacent camera tracking In the framework of tracking people and objects with artificial vision, the problem is to identify each one in two overlapping or adjacent images. Figure 11 shows a scenario where a person appears in the view angle of two separate cam- eras. Each of the cameras is calibrated according to the global coordinates. The system receives the positioning events of each person visible in each of the camera’s images. If we add the knowledge of the ‘Adjacent(c1, c2)’ scene to this, we can generate a rule that resolves this pro- blem. The example is shown below whereby the problem would be resolved: Atðc1;h345;x1;y1Þ Atðc2;h213;x2;y2Þ Adjacentðc1;c2Þ IFðx2 near x1Þandðy2 near x1Þ Figure 11: Scenario with overlapping cameras. c� 2011 Blackwell Publishing Ltd Expert Systems 11 Then (c1, h345)¼ (c2, h213) and (c2, h213)¼ (c1, h345). Thus it is immediately inferred that the human visible in camera 1 is the same one visible in camera 2. Figure 11 shows the evolution in time of events and by inference rule described above. In Intanto, t h345 shows a human in a position x1, y1 in camera 2. At time t2 is a h345 human in a position x2, y2 in the camera 2 and in turn a human h213 at position x3, y3 in the house 1. In the knowledge of the scene must be the camera 1 and camera 2 are adjacent. With this back- ground runs the rule displayed in row 3, and it appears that the h345 human camera 2 is the same as the h213 human camera 1 by the event ‘Similar’. 5.2. Precise identification from the access control Figure 12 shows an access control RFID identi- fication. Such controls provide, in normal cir- cumstances, a robust system to identify people. One of the problems posed by the artificial vision is the reliable identification of humans within a scene. It is therefore envisaged the merger of the information from the RFID identification system with the identification and tracking information provided by artificial vi- sion. Figure 13 shows an example of agent fusion using the scene knowledge level. It is an RFID card reader access control. A camera with a tracking system monitors people entering. The CA receives the person’s identification informa- tion from the RFID card read by the PA. Moreover, the VA detects another person enter- ing the security area and labels him=her with a self-generated ID. When these two items of information reach the CA, it searches for some correlation between the two. The system changes the self-generated ID for the RFID card ID. Then the person can be tracked via the video-camera inter-connection system. 6. Conclusions This work presents a multi-agent system for composition based on high-level knowledge events, which interprets activities, behaviour and situations semantically in a scenario with multi-sensory monitoring. A perception agent (PA and visual agent)-based structure is pre- sented. The agents process the sensor informa- tion and identify (agent decision system) significant changes in the monitored signals, which they send as simple events to the CA that searches for and identifies pre-defined patterns like higher-level semantic composed events. This Figure 12: Individual access control. Figure 13: Video-surveillance access control. 12 Expert Systems c� 2011 Blackwell Publishing Ltd structure also has a description of the methodol- ogy to develop knowledge-based systems to compose events and a set of tools to facilitate its application. The methodology focuses on dividing knowledge, classifying it into types and encapsulating it into what we call knowl- edge packages, which in turn are organized into abstraction levels. Each knowledge package identifies specific situations or activities from lower-level abstraction events. These ideas were shown using the prototype developed for video surveillance. It consists of two tools: a develop- ment interface for a rule-based knowledge base, and an EMS. The first tool aims to configure the pattern of new alarming situations, either based on standard, usual activities (in the knowledge package repository) or directly from identifying and tracking events in the image sequences. These sequences can be labelled with the tool included or the image analysis module. The EMS offers an execution system to facilitate the tasks to analyse the results. Different applica- tion examples of the event composition from different perception agents have been shown, which identify alarming situations and resolve problems in identifying and tracking people. The system is particularly useful for surveillance but it could also be applied to other fields. Future works will aim to increase the system’s inferential capacities. Obviously, it is necessary to bear in mind and process the sources of indetermination in each abstraction level, and particularly the identification and tracking ones. Acknowledgements The authors are grateful to the CiCYT for financial aid on project TIN-2007-67586-C02-01. References ABREU, B., L. BOTELHO, A. CAVALLARO, D. DOUX- CHAMPS, T. EBRAHIMI, P. FIGUEIREDO, B. MACQ, B. MORY, L. NUNES, J. ORRI, M. TRIGUEIROS and A. VIOLANTE. (2000) A video-based multiagent traf- fic surveillance system, in The IEEE Intelligent Ve- hicles Symposium 45-462. BOBICK, A. (1997) Movement, activity, and action: the role of knowledge in the perception of motion, Royal Society Workshop on Knowledge-based Vision in Man and Machine, London, pp. 1257–1265. CARMONA, E.J. (2009) On the effect of feedback in multilevel representation spaces, Neurocomputing, 72, 916–927. CASTANEDO, F., J. GARCÍA, M.A. PATRICIO and J.M. MOLINA (2008) A multiagent architecture to support active fusion in a visual sensor network, in 2nd ACM= IEE International Conference on Distributed Smart Cameras, Stanford University, CA, USA, pp. 1–8. CHLEQ, N. and M. THONNAT (1996) Realtime image sequence interpretation for video-surveillance appli- cations, IEEE International Conference On Image Processing (ICIP’96), Laussane, pp. 800–804. FOLGADO, E., M. RINCÓN, E.J. CARMONA and M. BACHILLER (2007) A block-based model for moni- toring of human activity, Technical Report AVISA- 12-07. FUENTES, L.M. and S.A. VELASTIN (2004) Vigilancia avanzada: del tracking a la detección de sucesos, IEEE América Latina, 2, 206–211. HAESEVOETS, R., B. VAN EYLEN, D. WEYNS, A. HELLEBOOGH and T. HOLVOET (2007) Context- driven dynamic organizations applied to coordi- nated monitoring of traffic jams, Engineering Environment-Mediated Multiagent Systems, Dres- den, Germany, 1–5 October, 2007, D. Weyns, S. Brueckner andY. Demazeau (eds.), pp. 126–143. LEVINE, M.D., P.B. NOBEL and Y.M. YOUSSEF (1983) A rule-based system for characterizing blood cell motion, in Image Sequence Processing and Dynamic Scene Analysis, T.S. Huang (ed.), Berlin, Germany: Springer-Verlag, 663–709. MARTÍNEZ-TOMÁS, R., M. RINCÓN-ZAMORANO, M. BACHILLER-MAYORAL and J. MIRA-MIRA (2008) On the correspondence between objects and events for the diagnosis of situations in visual surveillance tasks, Pattern Recognition Letters, 29, 1117–1135. MARTINEZ TOMÁS, R. and A. RIVAS CASADO (2009) Knowledge and event-based system for video-sur- veillance tasks, in Proceedings of the 3rd Interna- tional Work-Conference on the Interplay Between Natural and Artificial Computation (IWINAC): Part I: Bioinspired Applications in Artificial and Natural Computation, ISBN:978-3-642-02263-0, J. Mira, J. M. Ferrández, J.-R. A. Sanchez (eds.), Berlin: Springer-Verlag, 386–394. MOLINA, J.M., J. GARCÍA, F.J. JIMÉNEZ and J.R. CASAR (2004) Fuzzy reasoning in a multiagent system of surveillance sensors to manage coopera- tively the sensor-to-task assignment problem, Ap- plied Artificial Intelligence, 18, 673–711. NEUMANN, B. (1984) Natural language description of time-varying scenes. Brericht no. 105, FBI-HH-B- 105=84, Fachberic Informatik, University of Hamburg. c� 2011 Blackwell Publishing Ltd Expert Systems 13 NEUMANN, B. and H. NOVAK (1983) Events models for recognition and natural language description of events in real-world image sequences, Proceedings of the Eighth IJCAI, Karlsruhe, Morgan Kaufmann, San Mateo, California, pp. 724–726. PAVÓN, J., J.J. GÓMEZ-SANZ, A. FERNÁNDEZ-CABAL- LERO and J.J. VALENCIA-JIMÉNEZ (2007) Develop- ment of intelligent multi-sensor surveillance systems with agents, Robotics and Autonomous Systems, 55, 892–903. REMAGNINO, P., A.I. SHIHAB and G.A. JONES (2004) Distributed intelligence for multicamera visual sur- veillance, Pattern Recognition, 37, 675–689. ROTA, N.A. and M. THONNAT (2000) Video sequence interpretation for visual surveillance, in IEEE Inter- national Workshop on Visual Surveillance (VS’00), Dublin, Ireland, pp. 59–68. RUSSELL, S. and P. NORVIG (2009) Artificial Intelli- gence: A Modern Approach, 3rd edn, Pearson: Pre- ntice Hall. TOSTSOS, J.K., J. MYLOPOULOS, H.D. CORVEY and S.W. ZUCKER (1980) A framework for visual motion understanding, IEEE Transactions on Pattern Ana- lysis and Machine Intelligence, 2, 563–573. ZHU, Z. and T.S. HUANG (eds), (2007) Multimodal Surveillance: Sensors, Algorithms and Systems, Artech House Publishers. The authors Angel Rivas-Casado Angel Rivas-Casado received his degree in technical engineer in computer science of systems from the University Polytechnical of Madrid, Spain, in 2009. At present, he attends a master of advanced artificial intelligence in the National University for Distance Education (UNED) in Madrid, Spain. His research interests are in robotics, knowledge engineering, artificial vision, sensors fusion and intelligent agents. Rafael Martinez-Tomás Rafael Martinez-Tomás received his degree in physics from the University of Valencia, Spain, in 1983, and received his PhD from the department of artificial intelligence of the National University for Distance Education (UNED) in Madrid, Spain, in 2000. Since 2001, he is an associate professor with the department of artificial intelligence at the UNED. His research interests are in knowledge engineering, knowl-edge-based systems, spatial– temporal logics, description logics and video- sequence interpretation. Antonio Fernández-Caballero Antonio Fernández-Caballero received his de- gree in computer science from the Technical University of Madrid, Spain, in 1993, and re- ceived his PhD from the department of artificial intelligence of the National University for Dis- tance Education, Spain, in 2001. Since 1995, he is an associate professor with the department of computer science at the University of Castilla- La Mancha, Spain. His research interests are in image processing, computer vision, neural net- works and agent technology. A. Fernández- Caballero is member of the IAPR. 14 Expert Systems c� 2011 Blackwell Publishing Ltd 
work_paoqgodiljbkndasl5e2q3cfny	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy Landmark detection from mobile life log using a modular Bayesian network model Keum-Sung Hwang, Sung-Bae Cho * Department of Computer Science, Yonsei University, 262 Seongsanro, Sudaemoon-ku, Seoul 120-749, South Korea a r t i c l e i n f o Keywords: Modularized probabilistic reasoning Mobile application a b s t r a c t Mobile devices can now handle a great deal of information thanks to the convergence of diverse function- alities. Mobile environments have already shown great potential in terms of providing customized ser- vices to users because they can record meaningful and private information continually for long periods of time. Until now, most of this information has been generally ignored because of the limitations of mobile devices in terms of power, memory capacity and speed. In this paper, we propose a novel method that efficiently infers landmarks for users to overcome these problems. This method uses an effective probabilistic Bayesian network model for analyzing various kinds of log data in mobile environments, which were modularized in this paper to decrease complexity. We also present a cooperative inference method, and the proposed methods were evaluated with mobile log data generated and collected in the real world. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction Mobile environments have very different characteristics from desktop computer environments. First of all, mobile devices can col- lect and manage various kinds of user information, for example, by logging a user’s calls, SMS (short message service), photography, music-playing and GPS (global positioning system) information. Also, mobile devices can be customized to fit any given user’s pref- erences. Furthermore, mobile devices can collect everyday informa- tion effectively. Such features allow for the possibility of diverse and convenient services, and have attracted the attention of researchers and developers. Recent research conducted by Nokia is a good example (Nokia LifeBlog). Especially, the context-aware technique that has recently been widely studied is more applicable to mobile environments, so many intelligent services such as intelligent call- ing services (Schmidt, Takaluoma, & Mntyjrvi, 2000), messaging ser- vices (Lo, Thiemjarus, & Yang, 2003), analysis, collection and management of mobile logs (DeVaul, Sung, Gips, & Pentland, 2003; Gemmell, Williams, Wood, Lueder, & Bell, 2004; Korpipaa, Mantyjarvi, Kela, Keranen, & Malm, 2003; Krause, Smailagic, & Siewiorek, 2006; Raento, Oulasvirta, Petit, & Toivonen, 2005; Siewiorek et al., 2003; Zheng & Ni, 2006) have been actively investigated. However, mobile devices present some limitations. They con- tain relatively insufficient memory capacity, lower CPU power (data-processing speed), smaller screen sizes, awkward input interfaces, and limited battery lives when compared to desktop PCs. In addition, they have to operate in the changeable real world, which means that they require more active and effective adapta- tion functions (Dourish, 2004). In this paper, we propose a novel method of analyzing mobile log data effectively and extracting semantic information and mem- ory landmarks, which can be used as special ways of helping recall specific functions (Horvitz, Dumais, & Koch, 2004). The proposed method adopts a Bayesian probabilistic model to efficiently man- age various uncertainties that can occur when working with mo- bile environments, including real-world irregularities, like varying levels of attention and emotions, inaccuracy of sensors, and uncertain causal factors. The proposed model uses a coopera- tive reasoning method with a modular Bayesian network (BN) in order to work competently in mobile environments. We also dis- cuss how to learn and update the Bayesian inference model by using the proposed BN learning method with training data. The proposed method was applied to several experiments using both synthetic data and real mobile log data collected with a smart- phone for a month in the real world. In Hwang and Cho (2006), we proposed a method for identifying landmarks on mobile de- vices, which was a modularized BN model designed by human manually. There have already been various attempts to analyze log data and to support expanded services by using the probabilistic ap- proach. Li and Ji used a probabilistic model for active affective state detection of user (Li & Ji, 2005). They utilized a dynamic Bayesian network and the utility theory to reason the ‘fatigue,’ ‘nervous,’ and ‘confused’ states. They showed that the probabilistic approach was good at management of uncertain information like affection. Ji et al. used also a dynamic Bayesian network method for real- time monitoring human fatigue (Ji, Lan, & Looney, 2006). They 0957-4174/$ - see front matter � 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2009.03.002 * Corresponding author. Tel.: +82 2 2123 2720; fax: +82 2 365 2579. E-mail addresses: yellowg@cs.yonsei.ac.kr (K.-S. Hwang), sbcho@cs.yonsei.ac.kr (S.-B. Cho). Expert Systems with Applications 36 (2009) 12065–12076 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy concerned and combined many conditions such as light, heat, humidity, time, napping, anxiety, temperature, weather, and work type. Zhang and Ji proposed an active and dynamic information fu- sion method for multi-sensor systems with dynamic Bayesian net- works (Zhang & Ji, 2006). This showed the usefulness of Bayesian approach for information fusion. These works showed the Bayesian probabilistic approach was good tool for handling, reasoning, and combining uncertain information. Krause et al. clustered the sensor and log data collected on mo- bile devices, discovered a context classifier that reflected a given user’s preferences, and estimated the user’s situation in order to provide smart services (Krause et al., 2006). The context classifier was constructed using the BN model, which was based on a general learning method for a small domain of classification subjects. Hor- vitz et al. proposed a method that detected and estimated land- marks by discovering a given human’s cognitive activity model from PC log data based on the Bayesian approach (Horvitz et al., 2004). Their approach showed a good performance for recognizing and learning everyday-PC-life of humans. However, these methods were not suitable for mobile devices that were limited in terms of capacity and power. For larger do- mains, the general BN and BN learning method require highly com- plex computation. This is a crucial problem when it comes to modeling everyday life situations with mobile devices. To over- come these problems, a more appropriate approach was required to reduce the complexity levels. Marengoni, Hanson, Zilberstein, and Riseman (2003) tried to reduce the complexity levels of the BN model by dividing it into several multi-level modules and using procedural reasoning of the connected BNs (just like chain infer- ence). However, this method required procedural and classified properties of the target functions. Tu, Allanach, Singh, Pattipati, and Willett (2006) proposed a hy- brid BN model that allowed hierarchical hybridization of BNs and HMMs. However, it supported only links from lower level HMMs to higher level BNs without consideration of links between BNs of the same level. Hwang, Kim, and Zhang (2006) proposed the hierarchical probabilistic graphical modeling method, which con- structed a hierarchical and distributed BN structure using gener- ated hidden nodes and the links between them. However, this method is improper in mobile and private service environments since it leads to an increase of the number of nodes and does not keep the intuitive causal structure. 2. Landmark inference from mobile log data The overall process of landmark extraction from the mobile log data used in this paper is shown in Fig. 1. Various mobile log data is preprocessed in advance, and then the landmark-reasoning module detects the landmarks. The preprocessing module is oper- ated by the techniques of pattern recognition and simple rule rea- soning. The BN reasoning module performs probabilistic inference. The update module learns the Bayesian network mod- els and adapts to the user and environment by using the accumu- lated data. BNs refer to models that can express large probability distribu- tions with relatively small costs to statistical mechanics. They have the structure of a directed acyclic graph (DAG) that represents the link (arc) relations of the node, and has conditional probability ta- bles (CPTs) that are constrained by the DAG structure (Korb & Nicholson, 2003). Fig. 2 shows an example BN that was designed by human and used for the application of this paper. It shows a DAG structure, node name, state name and inferred probabilities. 2.1. Collection and preprocessing Table 1 shows log information collected on a mobile device and on the internet. The GPS log presents the places that the user vis- ited, and the call and SMS logs show the user’s calling patterns. The MP3 (a music file format) player log offers an idea of the user’s emotions and the photograph log shows when the user wanted to memorize something. Since logs have temporal properties, we considered their time spans, frequencies (per hour, daily, and weekly), and start/end times as well as their impact factors that reflect the density of gi- ven events. For example, the impact factor of time t, IPt can be cal- culated as Eq. (1), where x represents the event and a(x) and b(x) represent the increment/decrement constants and monitoring time-span for event x. We set the value a(x) as 1 for each event x, and the values for b(x) are set manually as follows; b(GPS) = 1 h, b(Call) = 1 h, b(SMS) = 20 min,b(Pic. View) = 5 min, b(Photograph- ing) = 30 min, and b(Mp3 Playing) = 30 min, IPtðxÞ¼ IPt�1ðxÞþ aðxÞ; if event x is occurred IPt�1ðxÞ� aðxÞ; if event x is not occurred in bðxÞ after prior event x 8>< >: ð1Þ The coordinates from the GPS log are used to get place names. In this paper, we divided the domain area into a lattice and then la- beled each region. The user profiles and PIMS (personal informa- tion management system) datasets were used to find the user’s social position (student, worker), gender, the position of their home, and the names and phone numbers of their friends and relatives. Weather User profile GPS log Preprocessing (Statistics, Analysis) P h o n e Call / SMS log PIMS Landmark Decision Photo log Device log Log Context DB W eb Landmark Reasoning Music Playing log Landmark DB Update module BNn BN2… BN1 Modular BNs Fig. 1. The process of the landmark extraction from mobile log data. 12066 K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 Author's personal copy 2.2. Cooperative reasoning of modular Bayesian networks There are two general differences between the proposed BNs and conventional BNs. Firstly, we modularized the Bayesian inference models according to their separated sub-domains (Fig. 3). The re- quired computing power for dealing with BN models is proportional to the number of nodes and arcs. Especially, since the computational complexity of Bayesian inference is approximately proportional to O(cN), where c is the number of states and N is the number of causal (parent) nodes, the modularized BN is more efficient. Secondly, to consider the co-causality of the modularized BN, the proposed method shows 2-pass inference stages (Fig. 3). A vir- tual linking technique is utilized to reflect the co-causal evidence more correctly. The technique is performed to add the virtual nodes and regulate their conditional probability values (CPVs) to apply the probability of the evidence. The process of the virtual linking technique is shown in Figs. 4 and 5. In Fig. 5, the virtual linking technique replaces given probabilistic evidence to the prior probability or conditional probability of virtual node value to keep the connectivity of BNs. For example, when the structure of BN1 is {A ? B ? C}, BN2 is its modularized version {A ? B, B ? C}, and A and C are evidence and result nodes, we can calculate the belief value of node C using the virtual link assumption and chain rule such as Eqs. (2)–(5) Yes 0% No 100% Coffee-shop Yes 0% No 100% Previor:Eating Yes 0% No 100% Previor:Stroll Yes 100% No 0% Lunch-time Yes 0% No 100% Breakfast-time Yes 0% No 100% Dinner-time Yes 90% No 10% Meal-time Yes 2% No 98% Teaing Yes 100% No 0% Restaurant Yes 90% No 10% Eating(Western) Yes 0% No 100% Korean-restaurant Yes 18% No 82% Eating(Eastern) Yes 0% No 100% Fastfood-shop Yes 20% No 80% Eating(Fastfood) Yes 93% No 7% Eating Yes 0% No 100% Ordinary-place Yes 0% No 100% High-class-restaurant Yes 83% No 17% Eating-out whereprevior landmarkswhen/work Fig. 2. The landmark inference BN ‘Activity in restaurant’ designed and used in this paper. This is 1 of 39 BNs designed. The node of gray box is landmark node. This shows the case that the ‘Lunch-time’ and ‘Restaurant’ evidence are set. Table 1 The log information that was collected on a mobile device. Log Information GPS Latitude, longitude, velocity, direction, date, time Call Caller’s phone number, call/receive/absence log, time span, start/end time SMS Sender’s phone number, call/receive/absence log, time span, start/end time Photographing Photo file name, taking time Weather Weather, visibility range (km), cloud degree (%), temperature (�C), discomfort index, effective temperature (�C), rainfall (mm), snowfall (cm), humidity (%), wind direction, wind velocity (m/s), barometer (hPa) MP3 Player Title, time span, start/end time Charging Charging status, time span, start/end time Log context Landmark Place- activity rea- soning BN Emotion/ condi- tion rea- soning BN Circum- stance situation BN event rea- soning BN Virtual evidence 1st Stage Landmark Place- activity rea- soning BN Emotion/ condi- tion rea- soning BN Circum- stance situation BN event rea- soning BN 2nd Stage Log context Fig. 3. The 2-pass inference process for the cooperation of modular Bayesian networks. It uses the result from inference of Bayesian network as the evidences at the 2nd inference stage. The parenthesis indicates the number of BNs contained. There are four kinds of BNs, and totally 39 BNs are used. The BNs are as follows: 19 place-activity BNs = {houses, religion, shopping, photographs, hospitals, nature, meetings, workplaces, sports, movements, food, calls, music, schools, traffic, amusement, busy, watching, resting}, 10 emotion/condition BNs = {joy, hunger, hot temperatures, cold temperatures, nonsense, surprise, tiredness, drunk, anger, worry, gloom, sickness, boredom}, 5 circumstance situation BNs = {space, climate, time, device, group}, 2 event BNs = {anniversaries, other events}. K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 12067 Author's personal copy where the proposed virtual link assumption is P(B) = Bel(B), and Bel(B) means belief value of the node B (Korb & Nicholson, 2003), BN1 ’s BelðCÞ¼ PðA; B; CÞ¼ PðC j BÞPðB j AÞPðAÞ ð2Þ BN2 ’s BelðBÞ¼ PðA; BÞ¼ PðB j AÞPðAÞ ð3Þ BN2 ’s BelðCÞ¼ PðB; CÞ¼ PðC j BÞPðBÞ ð4Þ ¼ PðC j BÞBelðBÞ since PðBÞ ¼ BelðBÞ by virtual link assumption ¼ PðC j BÞPðB j AÞPðAÞ ) BN2’s BelðCÞ¼ BN1 ’s BelðCÞ ð5Þ From the cooperative reasoning process we can get the landmarks and their probability distribution. Fig. 6 shows a pseudo code for the landmark extraction process. 2.3. Landmark decision The final landmark decision is obtained by the belief probabil- ity, thresholds and weights of each landmark node after reasoning the modular BNs. The threshold is used to tune the landmark extraction model. Generally, the training data is not uniform en- ough to obtain an almost-perfect model, so we need to tune the model. Eq. (6) shows the computations to obtain the selected land- mark set LMs, where bi represents the belief probability of node Xi, wi is the weight of node Xi, and LM is the set of landmark nodes. The weights are used to apply the preferences or life-patterns of the given user, so if the landmark is a preferred by an user, the weight has higher value than others and the landmark can be se- lected more easily, LMs ¼fXijbi � wi > thi ^8Xi 2 LM ^ 0 6 wi 6 1g ð6Þ 2.4. Complexity analysis To compare the complexities of the proposed modular BNs and the ordinary BN, we combined the designed BNs into a BN as shown in Fig. 7. A total of 39 BNs were designed with 638 nodes, 623 arcs and 4205 CPVs, as summarized in Table 2. On average, 16.6 nodes and 107.8 CPVs are used at the same time for the BN reasoning pro- cess since they are modularized and the computations are distrib- uted. However, even though the combined BN for them has 462 nodes, which is decreased from 638 after removing the duplicated nodes on the combining process, the number of parents and CPVs are increased. This means that it has much larger complexity. Eq. (7) shows the time complexity of the exact inference of the BN using the Lauritzen Speigelhalter (LS) algorithm (Lauritzen & Spiegelhalter, 1988), which is the most popular exact inference algorithm and a junction tree-based algorithm, where n represents the number of nodes, k represents the maximum number of par- ents for each node, r denotes the number of values for each node, and w represents the maximum clique that each parameter used in the LS algorithm (Namasivayam & Prasanna, 2006), cmpxinf ¼ Oðk 3 nk þ wn2 þðwrw þ rwÞnÞ ð7Þ We have replaced the maximum clique size w with k, since the clique size is proportional to the parents’ size, and the number of values r is about 2 in this paper, so Eq. (7) was reduced as shown by Eq. (8). cmpxinf ffi Oðk 3 nk þ kn2 þ 2kðk þ 1ÞnÞ ð8Þ The ideal modularization technique roughly divides the number of nodes by the number of modules d but it has to compute d times for d modules, so the exact inference complexity of the modular BNs is as shown in Eq. (9) where the new number of nodes is (n/ d). As shown in Table 2, the proposed modularization technique roughly divides the number of nodes n by 27.8 and the number of parents by 1.3, so the exact inference complexity is decreased significantly, cmpx0inf ffi O k3 dk�1 nk þ k d n2 þ 2kðk þ 1Þn ! ð9Þ 3. Learning modular Bayesian networks from data In this section we introduce a method to learn the proposed modular BNs automatically from the training data set. Fig. 4. Virtually linked Bayesian networks. Dotted line denotes some nodes linked by virtual linking technique. Xi'' Virtual Link 2: for root nodes; Replace the prior probability as p Xi Xi' Virtual node added xi=yes Virtual Link 1: for general nodes; Set the evidence xi, and then replace the conditional probability of virtual node as p P (xi''=yes) p x xi=yesx i=no P (xi'=yes|xi) p k' Probabilistic evidence: xi=yes, P (xi=yes)=p p cpt cpt p Fig. 5. The virtual linking method using virtual evidence enabling to use proba- bilistic evidences. Fig. 6. The pseudo codes for landmark extraction processes. L is log data set. 12068 K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 Author's personal copy 3.1. Learning Bayesian network from data The BN model G can be defined as (Bs, H), which means a net- work structure Bs and a probability parameter set H. H = {BU, Bp} is composed of the conditional probability table BU and the prior probability distribution Bp. In general, learning Bayesian networks from data consists of two parts. The first is to learn the structure Bs and the other part is to learn the parameter set. Learning the struc- ture can be solved by searching for the network structure which best matches the given data. The fitness can be measured by some scoring metrics. Two popular metrics are the minimum description length (MDL) score (Lam, 1998) and the Bayesian Dirichlet (BD) score (Heckermann, Geiger, & Chickering, 1995). The BD score for the structure Bs given the training data set D is defined as Eq. (10). To find the best structure with score metric, greedy search algorithms can be used. Fig. 8 shows a general process for the score-metric based structure learning, BDðBs; DÞ¼ PðBs; DÞ ¼ PðBsÞPðDjBsÞ¼ PðBsÞ Z PðDjH; BsÞPðHjBsÞdH ð10Þ Parameter learning can be cast as either estimating the most likely parameter values (maximum likelihood learning) or updating the posterior probability distributions of the parameters given data (Bayesian learning). The parameter h is calculated from the training data set D by using Eq. (11). h� ¼ arg max h PðDjhÞPðhÞ ð11Þ P(h) means prior probability. The general discovery process is shown by Eq. (12) where ZT = {z1, z2, . . ., zT} represents a set of T sta- tus variables, and YT is a set of T observation result variables, PðZT; Y T; hÞ¼ PðY TjZT; BUÞPðZTjBPÞ ð12Þ The conditional probability table BU is calculated from the rela- tional frequency between the observation values and the status variables, meaning the probabilistic histogram of the training data. 3.2. Discovering the modular Bayesian networks The proposed learning process of the modular BNs is shown in Fig. 9. This method includes a parameter setting process (for node Fig. 7. The complex BN structure combined from the 39 designed BNs. This has 588 arcs and 462 nodes. The graph is drawn by GeNIe 2.0 (Graphical BN tool). Table 2 The 39 modular BNs and the monolithic BN of them. MonoBN – the monolithic BN corresponding to the modular BNs, NN – no. of nodes, NNR – no. of root nodes, NNI – no. of intermediate nodes, NNL – no. of leaf nodes, NL – no. of arcs, NPavg – avg. no. of parents, NS – no. of states, NSavg – avg. no. of states, NCPV – no. of CPVs, CPTmax – maximum size of CPT. NN NNR NNI NNL NL NPavg NS NSavg NCPV CPTmax 39 BNs 638 375 135 128 623 0.98 1279 2.00 4205 64 MonoBN 462 235 111 116 588 1.27 927 2.01 4.869 512 Initialize structure Score all possible single changes Any changes better? Return saved structure Perform the best change yes no Fig. 8. A general process for score metric based structure learning. K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 12069 Author's personal copy set X, the topological order set of nodes O, a level set of nodes V and the maximum size of parents p that each node had), a structure learning (for Bs) and a parameter learning (for h). The node set of BN (X) is composed of the union of the collected log context set (L) and the target landmark set (LM). The values of the landmark set are defined by an expert user. The topological order set of nodes (O), the level set of nodes (V), and the maximum size of the parents (p) have been prepared for structure learning (the proposed meth- od; domain-constrained K2 learning). To determine the topological order (O), we ranks each variable based on its total influence on the other variables with mutual information. The domain-constrained K2 learning method is proposed to constrain the parent domain and the hierarchical levels of the nodes. The learning method and the mutual information computation are presented in the next section. The network structures learned are divided into several mod- ules by a modularization process as shown in Fig. 10, where G is the network structure and d denotes the number of modules. Pro- cess 1 defines the nodes of the module-BNs based on the module domain and the log context set. Process 2 defines the arcs of the module BNs based on the network structure (G). Process 3 defines the virtual nodes based on the network structure (G). The virtual node usage used for virtual linking between the BNs is described in Section 2.2. 3.3. Domain-constrained K2 algorithm The domain-constrained K2 method proposed in this paper is based on our prior work (Hwang & Cho, 2004) and aims to learn the structure of the BN hierarchically with a constrained parent do- main of nodes. The algorithm limits the hierarchical level at which a node can be positioned and reduces the search domain for the parent of the nodes. It can control the general direction of causality and decrease the searching complexity. In this paper, we use the algorithm to exclude the arcs between the evidence nodes (but we permits arcs between the landmarks) and maintain the direc- tion of the arcs in order to use the virtual link technique. Table 3 shows the level and domain settings used in this paper. Fig. 11 shows the proposed domain-constrained K2 algorithm, based on the K2 algorithm as proposed by Cooper and Herskovits (1992), which is the most popular BN algorithm and the basis of many advanced learning algorithms. This algorithm adopts a score metric known as the K2 metric, which calculates scores based on the difference between the BN graph and the training data distri- bution. The search process of the K2 algorithm is greedy and heu- ristic, and the K2 metric is calculated using Eq. (13), PðG; DÞ¼ PðGÞ Yn i¼1 Yqi j¼1 ðri � 1Þ! ðNij þ ri � 1Þ! Yri k¼1 Nijk! ð13Þ where ri represents the number of status variables, xi, qi means the number of combinations of variables included in pi, Nijk denotes the count in the dataset D, pi is the jth combination, and xi is the kth Preprocessed log L={l1, l2, …} Target landmark LM={lm1, lm2, …} Defining the topological order O={o1, o2, …} sorted by Mi=Σj Mij Module Domain M1={X2, X5, …} Md={X34, X49, …} Modularization of BN Parameter learning of BN Learned Modular BNs Structure learning of BN G = structure_Learn (X, O, V, p) Defining the node set of BN X={X1, X2, …}=L LM Calculating the link strength between nodes M(Xi, Xj): Mutual Information Defining the level of nodes V={v1, v2, …| vi=1,if Xi L, vi=0, if Xi LM, i=1,…} Fig. 9. The learning procedure of modular Bayesian network. Fig. 10. The modularization process of BN. The right side shows an example, in which the gray nodes denote evidence nodes and the white nodes mean landmark nodes or their virtual nodes. Table 3 The level and domain definition. Level Node domain Parent domain 0 LM = {lm1, lm2, . . .} D0 = U 1 L = {l1, l2, . . .} D1 = L [ LM LM is landmark set, L is log context set. 12070 K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 Author's personal copy value. The Nij value is computed by Eq. (14) and the function g( ) is calculated by Eq. (15), Nij ¼ Xri k¼1 Nijk! ð14Þ gðxi; PaðxiÞÞ¼ Yqi j¼1 ðri � 1Þ! ðNij þ ri � 1Þ! Yri k¼1 Nijk! ð15Þ The K2 algorithm uses a topological order to maintain the graph as the DAG by maintaining that the prior node cannot be the child of the posterior node without any other DAG checking rules. How- ever, we have to optimize the topological order since a different topological order will lead to a different BN structure. In this paper, we compute the influence score of all the nodes by using mutual information (Su & Zhang, 2006), and sort out the topological order with the score. Eqs. (16) and (17) show the influence score and the mutual information calculation, Mi ¼ XMðXi ;XjÞ j ð16Þ MðX; YÞ¼ X x;y Pðx; yÞ log Pðx; yÞ PðxÞPðyÞ ð17Þ 3.4. Threshold and weight The weight value presents the importance of landmarks and can be set by humans in terms of their preferences. In this paper, we set all the weight values as 1.0 to observe the general performance of landmark extraction. The performance of a classifier depends on its threshold since it determines the result status of the node. Especially in BN classifi- ers, it is changeable because the belief probability distribution is based on the number and distribution of training data values, so we have defined the threshold as the value that shows the best hit rate for the training data. Fig. 12 shows an example of the clas- sification results of the node ’moving.’ In the results, the hit rate Fig. 11. The domain-constrained BN structure learning algorithm. Dk denotes the parent domain of kth level, and level(i) means level of xi and MaxParentNum is a limitation of the number of parents. Fig. 12. The classification performance of the node ‘moving’ with threshold transitions (from a modular BN trained with the parameter p = 4). The rest threshold (more than 0.3) is omitted since there is no change. The value 0.1 shows the best performance. Table 4 The category, id and names of nodes. Category IDs Node names Landmark nodes (48) N001, N002, . . ., N048 Moving, using vehicle, school-club activity, eat out, lecture, sleeping, surprising, fret, joy, throb, tired, meet friend, cold, meal (Korean), meet family, busy, hungry, drinking (alcohol), take a walk, troublesome, overflowing joy, bored, ride a vehicle, weight-training, home activity, with leisure, hair-cut, study, yearning, school activity, traffic jam, run, walk, go to school, meal (western), eat (tee), ready to go out, date with my date, sad, test in school, sleepy, meet kin, employment counsel, extracurricular lecture, late for school, concentration, eat (Chinese), supper Log context nodes (110) N049, N050, N051, . . ., N156, N157, N158 SMS sender = unknown, take a picture, listening music, calling with kin or friend, receive a call, listening music for a long time, charging, good scenery place, GPS is operating, average velocity is less then 20 km/h, speed is from 5 km/h to 20 km/ h, amusement grounds, velocity is from 20 km/h to 150 km/h, shopping place, lecture building, velocity is less than 5 km/ h, town, front of school, department store, rest place, bus station, soccer play ground, park, traffic place, a post office, street, public institution, home, 0 o’clock, 1 o’clock, 2 o’clock, . . ., 23 o’clock, night, dawn, morning, AM, PM, evening, daytime, sunshine-time, week, calling, ignore of call, velocity is faster then 150 km, out of town, merrymaking place, subway station, tourist resort, terminal of exp. bus, express bus, student building, place for meal, church, place for special meal, school, coffee shop, out of town, front of YS Univ., main load of YS Univ., front of lecture building, main entrance of school, wedding place, museum, street basket-ball court, coffee shop, Japanese restaurant, playground, library, lecture building, theater, middle school, weather = fine, Thursday, spring, frequent call, many call, electronic and electric products center, commercial center, subway station, Restaurant, many photo, high-school, Dong-dae-moon place, weather = fine, Hospital, dining room, Korean restaurant, house of kin, house of friend, Chinese restaurant The parenthesis denotes the number of nodes. K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 12071 Author's personal copy was the best when the threshold was 0.1, and we adopted the va- lue for the threshold of the nodes. This procedure was conducted automatically during BN training stage. 4. Experiments In this section, we analyze the proposed method with experi- ments with generated and collected data. We evaluate the reason- ing performance of the landmark extraction model designed and learned. Then, we observe the experimental results. Finally, we compare the auto-trained modular BNs and their monolithic versions. 4.1. Experimental data The log data used in this paper include GPS log, call log, SMS log, picture log, music-playing log, device charging log, and weather log obtained from a website. The total number of designed BNs is 39 (19 activity reasoning BNs for 19 kinds of places, 13 emotion/con- dition reasoning BNs for the users, 5 circumstances or situation BNs, and 2 event reasoning BNs). We collected data from three college students (women) with smartphones for a month. These users performed subtasks (such as writing activity diaries, shopping, walking and calling) to make the data more useful. The experimental data was segmented into units of ten min- utes, so there were 2304 entries. Redundant data were excluded resulting in 779 examples remained. We defined 48 landmarks and used 110 life log contexts (as evidence) as shown in Table 4. To learn the modular BNs, we divided the landmarks into four modules based on four categories (Emotion and status, Everyday life, Events, and School life) as shown in Table 5. 4.2. A case study We tested the proposed landmark-reasoning model with the following scenario in order to confirm performance. The left side of Fig. 13 shows the scenario used. The BN set strongly related to Table 5 Module domain definition of landmark nodes. Domain name Number Landmarks Emotion and status 17 Bored, busy, cold, concentration, fret, hungry, joy, overflowing joy, sad, sleepy, surprising, throb, tired, troublesome, with leisure, yearning Everyday life 14 Eat (Chinese), eat (western), eat (Korean), home activity, meet family, moving, ready to go out, ride a vehicle, run, sleeping, supper, using vehicle, walk Event 9 Date with my date, drinking(alcohol), eat (tee), eat out, hair-cut, meet friend, meet kin, take a walk, traffic jam, weight-training School life 9 Employment counsel, extracurricular lecture, go to school, late for school, lecture, school activity, school-club activity, study, test in school Total 48 home park school restaurant downtown coffee shop home 09hr 10hr 11hr 12hr 13hr 14hr 15hr 16hr 17hr 18hr 19hr 20hr 21hr 22hr 23hr 24hr moving moving Activity lecture building student hall lecture building Scenario Today is school day. Morning lecture time. Simple lunch. Walking for a park. Photo with spring flowers. Met friends. Going to the restaurant wanted going a coffee shop. Chatting. Funny & joyful day. Target landmark Going-out- preparation Walking-for Joyful-photo Eating-out Teaing Overflowing- joy Lecture Lecture Eating Sleeping Sleeping (a) (b) Fig. 13. (a) A scenario of an everyday life with mobile device of an undergraduate student for experiments. (b) The observation of the probability values of 11 target landmarks. The denoted time is from 4 o’clock to 27 o’clock (equal to 3 o’clock of the next day). Abbreviations used for landmarks: A – overflowing-joy, B – photo (scenery), C – joyful-photo, D – walking-for, E – tea, F – eating-out, G – eating (western style), H – eating, I – going-out-preparation, J – shower, K – sleeping. 12072 K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 Author's personal copy the scenario is {food, photo, movement, nature, joy, home}. The probabilities were calculated when the related evidence was given. The log context data used as evidence includes 26 artificial log con- texts: going out (in 2 h), lecture-building, daytime, restaurant, pho- tography, eating, entertainment, before breakfast, condition, morning, going-out (before), moving, daylight-hours, sleeping (in 2 h), nature, device-in-use, dinner(before), good-weather, joyful day, home, charging, coffee-shop, ordinary-place, on campus, stu- dent-hall, GPS-running. After generating the log contexts during the day, we tested them. The right side of Fig. 13 shows the inference results. We can see the high probability values of the related landmarks at the corresponding times. For example, there are ‘going-out-prepa- ration’ and ‘shower’ landmarks at 7–9 o’clock, ‘eating’ at 12–13 and 17–19 o’clock, ‘walking’ at 13–14 and 20–21 o’clock, ‘photography at 14–15 o’clock, and ‘eating’ landmarks at 17–19 o’clock. 4.3. Performance evaluation of the designed BNs To evaluate performance of the designed BNs, we operated our landmark extraction model. Table 6 shows the results of landmark extraction using data from eleven days collected from human1. The data of the underlined date was also used in the learning experiment. In the experiment, we set all landmark thresholds as 66% to avoid worthless landmarks, since a human designed the BN models by focusing on true positive relations naturally and without regarding negative relations between the logs and the landmarks. Since the log data were uncertain and the activity diary did not cover all possible life contexts, we evaluated the results (RHIT’) using a wide viewpoint (permitting partial hits), which means that the landmarks could be reasonably detected. We decided the Table 6 The real life log data collected in smartphone. We have excluded the data of unstable GPS log. NCon – the number of input contexts (evidences), NLM – the number of extracted landmark candidates, N0LM – the number of selected landmarks, N 0 HIT – the number of hitting landmarks, N0LM – the number of partial hitting landmarks, NERR – the number of landmarks, RHIT – the rate of NHIT , R 0 HIT – the rate of N 0 HIT , RERR – the error rate. Date NCon NLM N 0 LM NHIT N 0 HIT NERR RHIT R 0 HIT RERR 02-24 116 72 13 3 10 0 23.1 100.0 0.0 02-27 167 49 15 4 11 0 26.7 100.0 0.0 02-28 64 50 8 3 4 1 37.5 87.5 12.5 03-02 202 128 18 8 10 0 44.4 100.0 0.0 03-04 102 53 7 1 5 1 14.3 85.7 14.3 03-06 86 56 12 5 3 4 41.7 66.7 33.3 03-08 114 92 12 3 7 2 25.0 83.3 16.7 03-09 103 45 7 4 2 1 57.1 85.7 14.3 03-15 128 76 13 4 9 0 30.8 100.0 0.0 03-17 46 45 8 3 3 2 37.5 75.0 25.0 03-21 67 40 10 4 4 2 40.0 80.0 20.0 Total 1195 706 123 42 68 13 34.1 89.4 10.6 Table 7 The matching results of extracted landmarks. They are evaluated based on the activity diaries and GPS trajectories of users. Date Hit landmark Partial hit landmark Missing (omission) 02-24 Meal (Korean), meeting, shopping Studying, lecture, lecture time, busy, unpleasant SMS, viewing, eat out, traffic jam, singing, dancing at club – 02-27 Study, joyful photograph, meeting, shopping Lecture time, busy, unpleasant SMS, photograph (food), photograph (product), photograph (scene), viewing, meal (Korean), eat out, singing, dancing at club – 02-28 Busy, joyful call, meeting Lecture time, unpleasant SMS, viewing, studying Shopping 03-02 Lecture time, busy, joyful call, meal (Korean), eat out, meeting, take a walk Unpleasant SMS, photograph (food), photograph (product), photograph (scene), viewing, disappointment, shopping, singing, dancing at club, studying, studying till late – 03-04 Viewing Computer working, washing face, busy, unpleasant SMS, traffic jam Lecture time 03-06 Meeting, unpleasant SMS, lecture time, busy, meal (Korean) Take a walk, viewing, traffic jam Eat out, shopping, singing, dancing at club 03-08 Lecture time, meal (Korean), eat out Unpleasant SMS, busy, traffic jam, meeting, shopping, singing, dancing at club Viewing, take tea 03-09 Study, meeting, lecture time, viewing, meeting Busy, traffic jam Shopping 03-15 Lecture time, mean (Korean), eat out, shopping Cleaning, cooking, dishwashing, busy, viewing, meeting, singing, dancing at club, take a walk – 03-17 Studying, take a walk, lecture time Busy, disappointment, meeting Shopping, viewing 03-21 Study till late, lecture time, meal (Korean), shopping Busy, meeting, singing, dancing at club Viewing, eat out 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 ea t o ut joy mo vin g ha ir- cu t ea tin g( Ko re an ) lec tur e stu dytire d tak e a w alk tes t in sc ho ol me et frie nd us ing ve hic le em plo ym en t c ou ns el sc ho ol ac tiv ity ea t ( we ste rn ) wa lk da te wi th my da te sc ho ol- clu b a cti vit y dr ink ing al co ho l) ru n fre t tra ffic ja m ea t ( tee ) sa d sle ep y me et kin bo re d ye ar nin g ex tra cu rri cu lar le ctu re tro ub les om e thr ob co nc en tra tio n wi th lei su re sle ep ing su rp ris ing co ld me et fam ily we igh t-t ra ini ng ho me ac tiv ity go to sc ho ol re ad y t o g o o ut bu sy ov er flo wi ng jo y ea t ( Ch ine se ) rid e a ve hic le hu ng ry lat e f or sc ho ol su pp er Landmark nodes O pt im iz ed th re sh ol d Yes No Fig. 14. The optimized threshold set for node. The thresholds are sorted by the value. K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 12073 Author's personal copy partial hits by majority of 5 persons. As a result, the rate of the complete hit landmarks (RHIT) was mostly low (34.1% in total) but the rate of the partial hit landmarks (RHIT’) was as high as 89.4%. Especially, when the activities were plenty, the results were better since the contexts were sufficient. Table 7 shows the specific results of landmark reasoning. For example, the user’s real activity on 03-02 was practicing music until late at night, but the land- mark’studying till late’ was detected. We classified landmarks like this as partial hit landmarks. 4.4. Performance evaluation of the trained BNs In this section, we describe the test result of the landmark extraction model using 16 days of mobile life log data obtained Fig. 15. The monolithic BN trained with parameter p = 8. The brighter nodes are landmark nodes and the others are evidence nodes. out of town front of YS Univ. main load of YS U... house of friend home 10 o'clock express bus weddin g place subway station take a picture front of school AM tourist resort school house of kin joy lecture buildin g japane se resta... playgro und V_meet friend V_eat out V_sho ol-club activity lecture buildin g V_lectur e daytime surprisi ng fret17 o'clock throb street 23 o'clock out of town middle school electon ic and electr... tired 15 o'clock cold V_ea ting (Kore an) V_studyV_sch ool activity comme rcial center high-sc hool busy hungry V_drink ing(alc ohol) bored V_rid e a vehic le troubles ome V_weig ht-traini ng V_hom e activity overflo wing joy V_eat (wester n) with leisure yearningsleepy 9 o'clock sad concent ration out of town f ront of Y S Univ . V_mov ing hom e f ront of scho ol tourist resort scho ol house of kin V_joy lecture bui ldin g town japan ese restau rant play gr ound meet f riend mai n entra nce of school eat out 11 o'clock sunshi ne-t ime many photo gra phi ng cof f ee sho p V_studyV_busy rest place dri nking( al coho l) 16 o'clock restau rant take a walk park wei ght-tr ai nin g 13 o'clock hair- cut V_go to scho ol traf f ic jam cof f ee sho p eat (tee) date with my date meet kin GPS is operating out of town f ront of Y S Univ . main load of Y S Univ . house of f riend mov ing traf fic place home 10 o'clock PM express bus wedding place museum subway station using v ehicle tourist resort V_joy lecture building town japanese restaurant V_eat out weather=f i ne night day timesleeping 17 o'clock V_throb V_tired 15 o'clock 12 o'clock eating (Korean) 22 o'clock speed is f aster then 15... terminal of exp. Bus meet family speed is less than 5km/h public institution walk 16 o'clock restaurant V_bored park ride a v ehicle home activity eat (western) 13 o'clock week run V_sleepy 7 o'clock ready to go out Chinese restaurant eat (Chinese) 21 o'clock supper front of YS Univ. mai n load of YS Univ. house of friend PM wedding place AM school V_joy lect ure building V_meet friend V_eat out 11 o'clock sunshine- time shool-clu b acti vity lect ure building weather =f ine many photogra phing lect ure street middle school evening coff ee shop library frequent call weather =f ine st udy V_wal k theater many call school act ivit y go to school test in school calling st udent building place for meal empl oym ent counsel extrac urri cular lect ure V_conc e ntrat ion late for school Emotion & status BN Everyday life BN Event BN School life BN Fig. 16. The BN modules trained using the proposed learning method with parameter p = 8. The dark nodes: virtual nodes, the lighter nodes are landmark nodes, and the others (medium brightness) are evidence nodes. 12074 K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 Author's personal copy by the proposed modular BN learning method. We set the MaxPar- entNum parameters (p) as 4 and 8 in the experiment. The automat- ically-discovered threshold values for the landmarks of the modular BNs are shown in Fig. 14, where the landmarks ‘hungry’, ‘late for school’, and ‘supper’ cannot be detected when the thresh- old is 1. This means the training data do not contain the related evidences. The other landmarks show various threshold values. Figs. 15 and 16 show the monolithic BN and modular BNs learned with (parameter p = 8). The average number of nodes, par- ents, and conditional probability values and the level of complexity are shown in Table 8. The complexities are calculated by Eq. (8) and we can observe the decrement of the complexity of the mod- ular BNs. Table 9 shows the results of the landmark reasoning evaluation. Because the number of training data was small, we used the leave- one-out validation method. We compared the monolithic BN and modular BNs with the parameters p = 4 and p = 8. The computation of the precision rate is (TP/(TP + FP)), and the hit rate is ((TP + TN)/ (TP + TN + FP + FN)). As shown by the results, the performance of the modular BNs is similar to that of the monolithic BN. The measured values are same when the parent size parameter p is 4. This means the proposed method is valuable since it is used to reduce the BN model and in- crease efficiency. The results with p = 4 is better than with p = 8. We think that the bigger p increase complexity by allowing more parents and cause lack of training. How many parameters of BN are not trained from data can be measured by the number of CPVs that is 0.5 since an initial untrained CPV is 0.5. The number of the trained mono BN with p = 4 is 582 and with p = 8 is 4396, so it is reasonable that the BN with p = 4 is better than with p = 8. 4.5. Comparison of the execution speed In order to compare the computation complexities between the ordinary monolithic BN model and the proposed model, we have tested their execution time on mobile device simulation. The sub- jects are designed 39 BNs, their monolithic BN, trained BN, and their monolithic BN. In this experiment, since we used 2-step infer- ence method for modular BN model, the reasoning process is con- ducted twice for each BN. We conducted 10 times of inference for each BN model. The test environment for mobile device is as follows: � Simulation Tool: Microsoft Pocket PC 2003 SDK. � Operating System: Pocket PC 2003. � Memory: 44 MB. Table 10 shows the experimental results. The execution time denotes CPU clock time-span during 10 reasoning processes. The monolithic and combined BN took more time than modular BNs on Pocket PC environments, but some experiments with mono- lithic BN failed to run because of out-of-memory error. This means that the ordinary BN cannot handle the landmark detection task on the constrained environment like pocket PC where the domain is too large while the proposed method can manage the task well. To find the executable size of BN for monolithic BN 2 (for the learned), we tested the BN with progressive deletion of nodes. Fig. 17 shows the experimental result. The learned monolithic BN could operate after deleting the 3rd node. It means that the 3rd node requires much computation complexity. Actually, the 3rd node had 11 parents and has 4096 conditional probability param- eters. It can be understood that some complex nodes cause much computation and make troubles. In the experiments, we can observe that the proposed model re- duced the computation complexity of BN and enabled to handle more complex probabilistic model on the constrained environment even when the original BN model cannot work. 5. Concluding remarks In this paper, we proposed a modular landmark inference mod- el, which was more efficient and suitable to mobile environments. We introduced the modularized BN model for efficient operations in mobile environments, and proposed the 2-step inference meth- od by applying the virtual node concept, and then learned the modular BNs automatically from the given training data. In exper- imental results with artificial and real mobile life log data, the in- tended landmarks were well-extracted and the proposed method was able to reduce the level of complexity. We also discussed how to define the learning parameters and thresholds However, in this paper, we did not sufficiently cover the temporal properties of human landmarks. In the future, we need to continue research using a dynamic BN model that manages temporal features Table 8 The comparison of the number of nodes, parents, and conditional probability values and the complexity of learned BNs. The complexities are calculated based on Eq. (8). BN Nodes # Nodes #avg Parents # Parents #avg Complexity Mono BN 115 115 298 2.59 O (2.6 �106) Modular BNs 160 40 182 1.14 O (0.388 � 106) # – the number of, #avg – the average number of. Table 9 The correctness comparison of extracted landmarks. Same values are underlined. BN P TP TN FP FN Precision Hit rate Mono BN 8 135 35,845 64 1,348 0.678 0.962 Modular BNs 8 133 35,853 56 1,350 0.704 0.962 Mono BN 4 420 35,807 102 1,063 0.805 0.969 Modular BNs 4 420 35,807 102 1,063 0.805 0.969 TP – true positive, TN – true negative, FP – false positive, FN – false negative. Table 10 The comparison of the execution time on Pocket PC emulator. BN Environment Number of BN Execution time (clk) Modular BNs 1 Pocket PC 39 100 combined BN Pocket PC 1 180 Modular BNs 2 Pocket PC 4 50 Monolithic BN Pocket PC 1 N/A – out of memory error Fig. 17. The execution time of the learned monolithic BNs 2 with node deletion on Pocket PC emulation. K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 12075 Author's personal copy well. Also, experiments with sufficient real world data should be conducted for a longer period of time. Acknowledgement This work was supported by MKE, Korea under ITRC IITA-2009- (c1090-0902-0046) and UCN 09C1-T3-11T. References Cooper, G. F., & Herskovits, E. (1992). A Bayesian method for the induction of probabilistic networks from data. Machine Learning, 9, 309–347. DeVaul, R., Sung, M., Gips, J., & Pentland, A. (2003). MIThril 2003: Applications and architecture. In Proceedings of the 7th IEEE international symposium on wearable computers (pp. 4–11). Dourish, P. (2004). What we talk about when we talk about context. Personal and Ubiquitous Computing, 8(1), 19–30. Gemmell, J., Williams, L., Wood, K., Lueder, R., & Bell, G. (2004). Pervasive capture and ensuing issues for a personal lifetime store. In Proceedings of the 1st ACM workshop on continuous archival and retrieval of personal experiences, October (pp. 48–55). GeNIe 2.0. Decision Systems Laboratory, University of Pittsburgh. <http:// genie.sis.pitt.edu>. Heckermann, D., Geiger, D., & Chickering, D. M. (1995). Learning Bayesian networks: The combination of knowledge and statistical data. Machine Learning, 20(3), 197–243. Horvitz, E., Dumais, S., & Koch, P. (2004). Learning predictive models of memory landmarks. In CogSci 2004: 26th annual meeting of the cognitive science society (pp. 1–6). Hwang, K.-S., & Cho, S.-B. (2004). Constrained learning method of Bayesian network structure for efficient context classification. Proceedings of Korea Information Science Society, 31(2), 112–114 (in Korean). Hwang, K.-S., & Cho, S.-B. (2006). Modular Bayesian networks for inferring landmarks on mobile daily life. In The 19th Australian joint conference on artificial intelligence (pp. 929–933). Hwang, K.-B., Kim, B.-H., & Zhang, B.-T. (2006). Learning hierarchical Bayesian networks for large-scale data analysis. International Conference on Neural Information Processing, 1, 670–679. Ji, Q., Lan, P., & Looney, C. (2006). A probabilistic framework for modeling and real- time monitoring human fatigue. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 36(5), 862–875. Korb, K. B., & Nicholson, A. E. (2003). Bayesian artificial intelligence. Chapman & Hall/ CRC. Korpipaa, P., Mantyjarvi, J., Kela, J., Keranen, H., & Malm, E.-J. (2003). Managing context information in mobile devices. IEEE Pervasive Computing, 2(3), 42– 51. Krause, A., Smailagic, A., & Siewiorek, D. P. (2006). Context-aware mobile computing: Learning context-dependent personal preferences from a wearable sensor array. IEEE Transactions on Mobile Computing, 5(2), 113–127. Lam, W. (1998). Bayesian network refinement via machine learning approach. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3), 240–251. Lauritzen, S. L., & Spiegelhalter, D. J. (1988). Local computations with probabilities on graphical structures and their application to expert systems. Journal of the Royal Statistical Society, 157–224. Li, X., & Ji, Q. (2005). Active affective state detection and user assistance with dynamic Bayesian networks. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 35(6), 93–105. Lo, B. P. L., Thiemjarus, S., & Yang, G.-Z. (2003). Adaptive Bayesian networks for video processing. International Conference on Image Processing, 1(1), 889– 892. Marengoni, M., Hanson, A., Zilberstein, S., & Riseman, E. (2003). Decision making and uncertainty management in a 3D reconstruction system. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(7), 852–858. Namasivayam, V. K., & Prasanna, V. K. (2006). Salable parallel implementation of exact inference in Bayesian networks. International Conference on Parallel and Distributed Systems. Nokia LifeBlog. <http://www.nokia.com/lifeblog>. Raento, M., Oulasvirta, A., Petit, R., & Toivonen, H. (2005). ContextPhone: A prototyping platform for context-aware mobile applications. IEEE Pervasive Computing, 4(2), 51–59. Schmidt, A., Takaluoma, A., & Mntyjrvi, J. (2000). Context-aware telephony over WAP. Personal Technologies, 4(4), 225–229. Siewiorek, D. P., Smailagic, A., Furakawa, J., Krause, A., Moraveji, N., Reiger, K., et al. (2003). SenSay: A context-aware mobile phone. In Proceedings of the 7th international symposium of wearable computers, October (pp. 248–249). Su, J., & Zhang, H. (2006). Full Bayesian network classifiers. In Proceedings of the international conference on machine learning (pp. 897–904). Tu, H., Allanach, J., Singh, S., Pattipati, K. R., & Willett, P. (2006). Information integration via hierarchical and hybrid Bayesian networks. IEEE Transactions on Systems, Man, and Cybernetics, Part A: Systems and Humans, 36(1), 19–33. Zhang, Y., & Ji, Q. (2006). Active and dynamic information fusion for multisensor systems with dynamic Bayesian networks. IEEE Transactions on Systems, Man, and Cybernetics, Part B, 36(2), 467–472. Zheng, P., & Ni, L. M. (2006). The rise of the smartphone. IEEE Distributed Systems Online, 7(3). 12076 K.-S. Hwang, S.-B. Cho / Expert Systems with Applications 36 (2009) 12065–12076 
work_pb27dlju2vfnzoojuflkutsnxa	----	doi:10.1016/j.eswa.2005.09.044 Model gene network by semi-fixed Bayesian network Tie-Fei Liu a , Wing-Kin Sung a,*, Ankush Mittal b a Department of Computer Science, National University of Singapore, Singapore b Department of Electronics and Computer Engineering, Indian Institute of Technology, Roorkee, India Abstract Gene networks describe functional pathways in a given cell or tissue, representing processes such as metabolism, gene expression regulation, and protein or RNA transport. Thus, learning gene network is a crucial problem in the post genome era. Most existing works learn gene networks by assuming one gene provokes the expression of another gene directly leading to an over-simplified model. In this paper, we show that the gene regulation is a complex problem with many hidden variables. We propose a semi-fixed model to represent the gene network as a Bayesian network with hidden variables. In addition, an effective algorithm based on semi-fixed structure learning is proposed to learn the model. Experimental results and comparison with the-state-of-the-art learning algorithms on artificial and real-life datasets confirm the effectiveness of our approach. q 2005 Elsevier Ltd. All rights reserved. Keywords: Gene network; Bayesian networks; Hidden variable; Semi-fixed network; Semi-fixed structure EM learning algorithm 1. Introduction The complexity of a living cell is achieved by the concerted activity of many genes and their products. Such activity is often coordinated by the organization of the genome into regulatory modules, or sets of co-regulated genes that share a common function (Segal et al., 2003). The global gene expression pattern is therefore the result of the collective behavior of individual regulatory pathways. The large number of regulation pathways comprise a complex network which is called a gene network (Kolpakov, Ananko, Kolesov, Kolchanov, & Genenet, 1998). Understanding the network as a whole is essential and learning the gene network is an important central theme in post genomic research (Cumiskey, Levine, & Armstrong, 2003; D’haeseleer, Liang, & Somogyi, 2000; Friedman, 2004; Hasty, McMillen, Isaacs, & Collins, 2001). The accurate reconstruction of gene network has many possible benefits: (1) gene network provides information to help the annotation of the genome (Brazhnik, Fuente, & Mendes, 2002); (2) it is an important step to uncover the biochemical network in a cell (Brazhnik et al., 2002); (3) it provides valuable clue and lead to new idea to treat some complex diseases (Friedman, 2004), etc. With the construction 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.09.044 * Corresponding author. Tel.: C65 68746772; fax: C65 67797465. E-mail addresses: liutiefe@comp.nus.edu.sg (T.-F. Liu), ksung@comp. nus.edu.sg (W.-K. Sung), ankumfec@iitr.ernet.in (A. Mittal). of the expert system of gene network, it is possible to predict the regulators of genes, the regulatory relationships among genes and the regulatory pathways, thus learn the regulatory patterns of the organisms. Learning gene network is also a necessary step to build an expert system to predict the outcome of the perturbation of the genome, such as diseases (Brazhnik et al., 2002). Yet, it is impractical to construct a detailed biochemical model of an organism containing hundreds or even tens of thousands of genes by analyzing each gene and determining all the binding and reaction constants one by one manually 1 . Therefore, the methods for automatically reconstructing gene network by computing are needed. Gene network can be reconstructed by analyzing the gene expression data, which is extracted from the mi-croarray (Akutsu, Miyano, & Kuhara, 1999; Chen, He, & Church, 1999; Chen, Filkov, & Skiena, 1999; Cumiskey et al., 2003; D’Haeseleer, Wen, Fuhrman, & Somogyi, 1999; Friedman, Linial, Nachman, & Pe’er, 2000; Imoto, Goto, Miyano, 2002; Imoto et al., 2003; Murphy & Mian, 1999; Someren, Wessels, & Reinders, 2000; Tominaga, Okamoto, Watanabe, & Eguchi, 2001; Wagner, 2002; Watanabe, 1998; Wessels, Someren, & Reinders, 2001; Yaki & Friedman, 2004). Various methods are used for this purpose. The early computational approaches were based on learning the relationships among genes either by studying the mutual information or the correlation among their expression values. The representatives of such approaches are Expert Systems with Applications 30 (2006) 42–49 www.elsevier.com/locate/eswa 1 http://www.cs.unm.edu/wpatrik/networks/networks.html http://www.elsevier.com/locate/eswa http://www.cs.unm.edu/~patrik/networks/networks.html T.-F. Liu et al. / Expert Systems with Applications 30 (2006) 42–49 43 Pair-wise interaction (Arkin, Shen, & Ross, 1997) and clustering (Wahde & Hertz, 2000), which directly find the correlations among genes. After that, Boolean networks (Akutsu et al., 1999; Liang, Fuhrman, & Somogyi, 1998) are used in several works, which assume the gene expression levels are represented by Boolean values and the gene regulatory relationship are represented by a set of Boolean functions. Linear (Someren et al., 2000) and non-linear (Watanabe, 1998) models are followed which assume the regulatory relationships are represented by linear functions and non-linear functions, respectively. In a famous paper by Friedman et al. (2000), an algorithm to learn gene network using Bayesian network is presented: Each gene is a vertex and each regulatory relationship is an edge in the Bayesian network. Bayesian network represents probabilistic multivariate statistical depen- dencies among variables in a compact and easy-to-decipher way (Barrientos & Vargas, 1998; Gustavo, Sucar, & Villavicencio, 1998; Kang & Gola, 1999; Oatley & Ewart, 2003; Sucar & Miriam, 1998; Viaene, Dedene, & Derrig, in press). As it can handle stochastic events, control noise and handle dataset with only a few replicates (Friedman et al., 2000; Imoto et al., 2002), it has been shown to have advantages over other methods in learning gene network (Friedman et al., 2000; Imoto et al., 2002; Murphy & Mian, 1999). Since then, many works based on Bayesian network frame work are proposed and more biological relevant results are obtained. Hartemink, Gifford, Jaakkola, & Young, (2001) extended Friedman’s work by adding the annotations to the edges: ‘C’, ‘K’ or ‘G’ which represent the positive, negative or unknown regulation. Imoto et al., extended Bayesian network by combining non-parametric regression (Imoto et al., 2002) to detect the nonlinear relationship among genes and by making use of relevant biological information (Imoto et al., 2003) to improve the learning performance. Murphy and Mian (Murphy & Mian, 1999) used the dynamic Bayesian network, which is an extension of Bayesian network, to model the time delay in the gene network. However, when referring to gene regulation, most works simply model that a gene is directly regulated by other genes. The regulations of genes in fact occur at various levels including transcription, translation, splicing, posttransla- tional protein degradation, and other processes (Kolpakov et al., 1998). To give a correct description of the gene network, we cannot just consider the gene expression level (that is, the level of mRNA transcription). Based on the biological knowledge, it is known that protein plays a key role in gene network (Wyrick & Young, 2002). In fact, protein–DNA interaction and Protein-Protein interaction are the main activities in the regulatory system (Kolpakov et al., 1998). Therefore, when predicting microarray gene expression data, proteins should be included. Although their expression levels are still difficult to measure in the large scale, it is good if we can treat them as hidden variables. The following are the advantages of modeling proteins as hidden variables: † The model will become more meaningful, more interpretable and closer to real-life system (Barbara, Jameson, & Witting, 1999; Elidan, Lotner, Friedman, & Koller, 2000; Friedman, 1997; Friedman, 1998). † In the model, the proteins are decision-relevant. The network without considering hidden variable may omit some dependencies (Barbara et al., 1999; Elidan et al., 2000). Introducing hidden variables introduces advantages as well as increases the complexity of the network learning. Moreover, microarray gene expression datasets often have missing values. Considering the above challenging issues, Bayesian network forms a natural choice with the advantage of supporting several principled methods for learning the causal relationships with incomplete data, both hidden variables and missing values (Friedman, 1998). Successful application of Bayesian network to learn structure with hidden variables can be found in for several applications (Barbara et al., 1999; Elidan et al., 2000; Friedman, 1997, 1998). The most widely used method for structure learning is EM algorithm (Friedman, 1998). In E step, the algorithm calculates the score of each possible structure using the structure and parameters learnt in M step. The procedure is repeated until the convergence criterion is met. In our problem, such kind of learning is difficult since the algorithm needs to learn the relationship among the hidden variables and the observed variables. In addition it is difficult to predetermine the correct number of hidden variables. To learn the optimum number of hidden variables is computationally complex. Besides, in the presence of hidden variables, the network is no longer decomposable and this makes the learning difficult (Friedman, 1998) (as described in Section 3). We propose a system which models the gene network as a directed graph with hidden variables. In our model, the number of hidden variables is predefined using the biological knowl- edge and in addition, the relationships between hidden variables and observed variables are partially fixed. We propose a modified EM algorithm that takes advantages of the semi-fixed structure to decompose the network, and thus allow us to learn such network efficiently. Also, an approximation method to perform inference on the joint probability of two genes is presented in order to speed up the learning procedure. 2. Model gene network as a semi-fixed network with hidden variables In the gene regulation system, the regulation process can be divided into two main steps. The first step is gene expression, which is represented by gi/ri/pi where gi is a gene, ri and pi are the corresponding mRNA and Protein respectively. gi/ri is the transcription and ri/pi is the translation process. The transcription efficiency that is represented by mRNA level can be measured by microarray. The second step is the gene regulation, which is represented by pi/gj. In this step, the generated protein pi, possibly in collaboration with some other proteins, regulates the target gene gi. These two steps are the most important steps in the gene regulatory system. All other steps such as protein degradation just adjust the expression and the regulation strengths. Let Pa jZ{g1, ., gk} denote a parent T.-F. Liu et al. / Expert Systems with Applications 30 (2006) 42–4944 set of gene gj such that each gi2Pa j regulates gene gj. We note that all proteins {p1,., pk} in Pa j act in a combinative manner to regulate gj. In other words, the proteins may combine together to form a complex then bind to the binding site of the target gene and regulate it, or bind to their own binding sites then collaboratively regulate the target gene, or a mixture of the above two ways. An example network is shown in Fig. 1(a). Considering a single node representing the proteins’ combined influence as the direct regulator to regulate the target gene, the model can be further simplified as shown in Fig. 1(b). Based on the above discussion, for each gene gj, we propose a hidden variable hj, which represents the combi- nation of a set of regulatory proteins expressed from Pa j . Thereupon, Pa j regulate hj, then hj regulates gj. Based on the simplified regulation system, each gene is regulated by at most one hidden variable and the hidden variables are regulated by one or more than one genes. Such model means the interaction exists only between Gene )/ Protein or among proteins, and there is no edges from gene to gene or hidden variable to hidden variable. This model is termed as a semi-fixed network and is formally defined as follows: † A semi-fixed network NZ! V, EO where vertices VZ OgH, OZ{g1, ., gn} is a set of observed variables representing the expression levels of genes, and HZ{hi, ., hn} is a set of hidden variables representing the expression levels of combined proteins) and E is the edge set. † For each ek2E, ekZ(gi, hj) or (hj, gi). Thus, N is a bipartite graph on two partitions O and H. † For each gi2O, there is exactly one incoming edge (hi, gi) while there can be many outgoing edges. † For each hi2H, it has exactly one outgoing edge (hi, gi) while there can be many incoming edges {(gi1, hi), (gi2, hi), ., (gik, hi)} where Pa(hi)ZPajZ{gi1, gi2, ., gik}. Compared to learning general network with several hidden variables, learning the semi-fixed network is easier since the number of hidden variables is fixed and the relationships between hidden variables and observed variables are partially known. Fig. 1. Simplified gene regulation system. (a) Gene expression system can be simplified as the interaction of genes and proteins. (b) The system can be further simplified since the combined proteins are the direct regulator to the target genes. 3. Semi-fixed structure EM learning algorithm When there is no hidden variable in the network and no missing values in the gene expression data, the probability of the variables given a network structure can be expressed as the production of the probabilities of independent sub- networks: PðX1.Xn : No; DÞ Z Y i PðXijPaðXiÞÞ (1) where X1 to Xn are observed variables and No denotes the network without hidden variables. D is a complete dataset. The decomposability property reduces the learning difficulty in the following manner. With score functions such as minimum description length (MDL) and Bayesian scoring metric, learning the structure of a network can be decomposed to learn each independent sub-network separ- ately (Friedman, 1998; Friedman, Nachman, & Peer, 1999). Whereas, in the presence of incomplete data, the decom- posability property is not valid and this makes the learning difficult (Friedman, 1998). However, if all parameters of all hidden variables are assigned, hidden variables are observed and thus in a sense, the incomplete dataset becomes ‘complete’. A useful property of semi-fixed network is that when all parameters of hidden variables have been assigned (as the number of hidden variables and partial relationships of hidden variables and observed variables are known ensuring that all hidden variables can be assigned parameters), the network can be decomposed into independent subnetworks. Each subnet- work comprises of a gene gj, a hidden variable hj and a parent set of hj (denoted as Pa(hj)) (as show in Fig. 1(b)). The probability is decomposed as: Pðg1.gn; h1.hn : No;h; DÞ Z Y i PðhijPaðhiÞÞ Y i PðgijhiÞ (2) where g1, g2, ., gn are observed variables (genes), h1, h2, ., hn are hidden variables and No,h denotes the network with hidden variables. D is an incomplete dataset. Based on different decomposition methods, the decompo- sition of scores is also different. In our work, Bayesian score is used. For the detail of Bayesian score, please refer to (Friedman et al., 2000). When the network No has no hidden variable, the score is decomposed as follows: ScoreðNoÞ Z X i ScoreðgijPaðgiÞÞ (3) For a semi-fixed network No,h, hidden variables are included. We can show that its Score can be decomposed because the hidden variables hi’s are attached to the corresponding genes, as follows: ScoreðNo;hÞ Z X i ScoreðhijPaðhiÞÞ C ScoreðgijhiÞ (4) Proof:. Given all hidden variables, No,h is decomposable. By Formula 4, ScoreðNo;hÞZ P i ScoreðvijPaðviÞÞ where vi is the variables, include both genes and proteins, in No,h: {v1, v2, ., T.-F. Liu et al. / Expert Systems with Applications 30 (2006) 42–49 45 v2n}Z{g1, g2, ., gn, h1, h2, ., hn}. Therefore, it is easy to seeP i ScoreðvijPaðviÞÞZ P i ScoreðhijPaðhiÞÞC P i ScoreðgijPað giÞÞ where P a(gi)Zhi in No,h. Thus, Formula (4) is proved. With the aim of computing a network No,h which maximizes Score(No,h) using the property in Eq. (4), we propose a EM algorithm known as Semi-fixed Structure EM (SSEM) algorithm to learn such network. The principle of the EM algorithm is as follows: In each iteration, given an initial network structure NiK1 and a parameter set qiK1 (which includes both the parameters of observed variables qiK1 and hidden variables qhiK1), the step 1 is to calculate the missing values and hidden variables to build a complete dataset. The second step is to find a better structure Ni and a better parameter set qi based on the new complete dataset. A detailed description of the algorithm is shown at the end of this chapter. For step 1, the missing values and hidden variables can be filled up as follows: Given N and q, for a missing value of a variable gi, supposing Pa(gi) is the parent set of gi in N and DZ{d1, ., dk} is the discrete values of gi, we fill up the missing value of gi by a vector SZ(s1, ., sk) where sjZP(giZdjjPa(gi)). This is illustrated in the following example: Example:. The variable g1 has parents g2 and g3. As shown in Table 1(a), there is a missing value in an instance: {g1, g2, g3}Z{(), 1, 1} where () indicate a missing value. Suppose P(giZ0jg2Z1, g3Z1)Z0.4 and P(giZ1jg2Z1, g3Z1)Z0.6, we fill up the instance by {g1, g2, g3}Z{(0.4, 0.6), 1, 1}. It means, the filled value adds 0.4 count to the case g1Z0 and 0.6 count to the case g1Z1. As shown in Table (b), there are 1.4 cases for which g1Z0 and 2.6 cases for which g1Z1. In other words, P(g1Z0)Z1.4/4Z0.35 and P(g1Z1)Z2.6/4Z 0.65. The values of hidden variables are treated as missing values too and are computed in a similar fashion. Given the filled dataset, NiK1 and qiK1, step 2 can learn Ni and qi using general structure learning method. The general method to learn the structure from a complete dataset is to decompose the network into independent subnetworks. Then, learn the structure of each subnetwork independently. Since we fix Table 1 An example of filling missing values g1 g2 g3 (a) () 1 1 1 0 1 1 1 1 0 1 1 (b) (0.4, 0.6) 1 1 1 0 1 1 1 1 0 1 1 partial structure, we can decompose it as the independent subnetworks, each of which has a target gene, a hidden variable and a parent set of the hidden variable (similar to Fig. 1(b)). The main objective is to find the optimal parents for each hidden variable. Learning parents from a large number of candidates is difficult task. As gene network is a sparse network (Someren et al., 2000), a popular technique is to first measure the dependencies of candidates to target variable and to choose the best k genes as candidate parents. Then, the search for the parents from the candidate parent set can be performed. Friedman et al. (Friedman et al., 1999) proposed to calculate the dependency by KL-divergence (as Eq. (5)): MIðX; YjMÞ Z X X;Y PðX; YÞlogðPðX; YÞ=PMðX; YÞÞ (5) where MI(X, YjM) is the mutual information of variables X and Y with respect to the network M, P(X, Y) is the observed joint probability of X and Y and PM(X, Y) is the estimated joint probability of X and Y given M. Though we can compute PM(gi, hi) given the filled dataset, we cannot compute P(gj, hi) to measure the dependency of gj and hi, as hi is hidden. We propose an alternate way to measure the dependency: in each subnetwork, the target gene is the only descendant of the hidden variable. The probabilities of the hidden variable are passed to the target gene. Therefore, MI(gj, hi jM) can be reflected by MI(gj, gijM). Thus, the MI can be calculated as following: MIðgj; gijMÞ Z X gj;gi Pðgj; giÞlogðPðgj; giÞ=PMðgj; giÞÞ (6) Where P(gj, gi) is the observed probability distribution of genes gj and gi while PM(gj, gi) is the estimated probability distribution of gene gj and gi in network M. Performing inference on joint probabilities in a big dataset is quite time intensive. Thus, we approximate the joint probability of gi and gj as follows: † If gj2Pa(gi), then gj/hi/gi. PM(gj, gi)ZPM(gijgj) PM(gj)zPM(gijhi)PM(hijgj)PM(gj). Note that in this case, PM(gj, gi)sPM(gi, gj). † Otherwise, gi and gj are conditionally independent and PM(gj, gi) can be approximated as PM(gj)PM(gi). By the above approximation, only the joint probabilities of the gene pairs with the parent-child relationships need to be calculated. The detail of the iterative algorithm is as follows: † In iteration i, given NiK1, qi-1 and the original incomplete dataset D 0 . In E step, the missing values and the values of hidden variables of D 0 are filled up based on the inference of NiK1 and qiK1. Then, we obtain a complete dataset DiK1. The structure Ni is learnt based on DiK1 by maximizing Score(Ni: qi-1, Di-1). Because of the decomposability, we learn the parents for each gene gi independently: Fig. 2. Leaning performance comparison in the artificial dataset. (a) Structure of the artificial network, which contains 4 genes and 4 regulatory edges. (c) Structure learned by K2, there is no correct edge. (b) Structure learned by SEM. There is only 1 correct edge. (d) Structure learnt by SSEM. All edges are uncovered. T.-F. Liu et al. / Expert Systems with Applications 30 (2006) 42–4946 (1) Find k genes with best MI score with respect to gi to build the candidate parent set CPS. (2) Find parents PS2CPS which maximum Score(SSi: qiK 1, DiK1), where SSi is a substructure PS/hi/gi. – repeat the procedure until converging. † In M step, qi is learnt based on Ni which maximum Score(Ni: qi, DiK1) † Repeat the procedure until convergence or till the predefined number of iteration is reached. In the above procedure, the network structure and parameters are refined iteratively. To get the initial network structure N0 and parameters q0: † learn the D 0 as a complete dataset. Pa i for each gi and the dependent probabilities are gotten. † for each gene gi, add a hidden variable hi. Set the P a(hi)Z Pa i and P(hijPa(hi))ZP(gijPa i ). † A complete dataset D 0 by filling D 0 based on the network structure and dependent probabilities is obtained. † N0 and q0 can be learned from D 0 . 4. Experimental results and comparison The model and the algorithm are implemented in MATLAB and BNT 2 . Below, we present experimental results on both artificial and real life gene expression datasets. 4.1. Experiment on artificial datasets An artificial dataset is generated by simulating the gene expression with hidden variables. The artificial network contains 5 genes and 4 edges (as shown in Fig. 2(a)). We built a network containing 5 genes, 5 proteins and 2 protein combinations whose topology is similar to Fig. 1(a) and assign parameters randomly. Then we generated the dataset with 150 replicates and delete the data of the proteins and protein combinations. We compare the learning results with traditional learning algorithms K2 (a Bayesian network learning algorithm (Cooper & Herskovits, 1992)) and structural EM (SEM) (an algorithm in learning Bayesian network with hidden variables (Friedman, 1998)). K2 is a learning algorithm to learn the complete dataset. SEM allows hidden variables. However, it does not define either the number of hidden variables or any prior information of the relationships of hidden variables and observed variables. Our algorithm SSEM allows hidden variables. Moreover, the hidden variables are semi-fixed as stated in previous sections. The comparison for artificial dataset is shown in Fig. 2. K2 and SEM recovered 0 and 1 correct edges, respectively, together with 2 false edges for each while SSEM recovered all edges together with 2 false positives. Hence, SSEM results in more correct edges as compared to K2 and SEM. Moreover, 2 Bayes network toolbox, http://www.ai.mit.edu/wmurphyk the two false positive parents of the variable 5 are the ancestors of the variable 5. Each of them is in a pathway with the variable 5. Not like the false edges in Fig. 2(b) and (c), they contain some regulatory information though they are false edges. 4.2. Experiments on real-life data The microarray gene expression data for S. Cerevisiae contains 76 gene expression measurements (Spellman, Sherlock, & Futcher, 1998). To simplify the presentation and implemen- tation, the expression levels were discretized to discrete values 0 and 1: a value is discretized to 0 if it is smaller than 0 and 1 otherwise. The theoretical support behind the binary discretiza- tion of gene expression level is that the genes are most of the time either maximally expressed or virtually not expressed (Louis & Becskei, 2002) ensuring that the binary gene regulatory network is biologically meaningful. The real-life gene network in our work is a Yeast transcriptional cell cycle subnetwork published in (Futcher, 2002), which includes 13 genes and 18 edges. It is a good example to demonstrate the learning power since the subnetwork is the backbone part of Yeast cell cycle gene regulatory network (Simon et al., 2001), which is robustly designed (Li, Long, Lu, Ouyang, & Tang, 2004), and the genes here are the most important transcription factors in Yeast cell cycle (Li et al., 2004; Simon et al., 2001). The directed pairwise regulatory relationships for the subnetwork are verified by YPD (Yeast Proteome Database) 3 (Hodges, McKee, Davis, Payne, & Garrels, 1999) and (Futcher, 2002), as shown in Fig. 3(a), and the cell cycle regulations are shown in Fig. 3(c). Since some time delays exist in the regulation system (Liu, Sung, & Mittal, 2004), a gene in time slice i may regulate another gene in time slice i to iCk (k indicates the maximum time delay). Thus, we learn the edges in a k-window, i.e. the consecutive k time slices (Boyen, Friedman, & Koller, 1999). An approach based on k -DBN (Boyen et al., 1999) is implemented for the comparison purpose. Besides the direct regulatory relationship between genes, we also employed a popular statistic feature, the Markov relation. A 3 http://www.proteome.com/YPDhome.html http://www.ai.mit.edu/~murphyk http://www.proteome.com/YPDhome.html Fig. 3. Leaning performance of SSEM in real-life gene network. (a) The original subnetwork structure, which contain 13 genes and 18 edges. (b) The structure learnt by SSEM. There are sum 29 edges with confidence no smaller than 0.6. Among them, 11 edges are verified as true positives by (a). Considering (c), there are 9 more edges are regarded as true positives. (c) The Yeast cell cycle regulatory network. The genes in a box is a transcription factor or a complex. (d) Markov features with confidence no smaller than 0.6 learned by SSEM. There are 44 Markov features. Among them 38 can be verified by (a) and (c). (e) Learned cell cycle regulatory network which is simplified from (d). T.-F. Liu et al. / Expert Systems with Applications 30 (2006) 42–49 47 Markov relation (Friedman, 2004; Friedman et al., 2000) describes whether a gene Y is in the Markov blanket of another gene X (denoted by Y2MB(X)). Y2MB(X) if and only if there is either an edge between them or both are parents of another variable. It indicates whether two genes are involved in the same biological event and it has been regarded as a criterion by several works to evaluate the performance of gene network reconstruc- tion (Friedman, 2004; Friedman et al., 2000). Bayesian network is a model of dependencies among random variables, rather than causality. A/B and B/A are the alternative ways of describing that A and B are not independent on each other. Markov relation is helpful in discovering the dependencies without caring about the directions of the edges. Moreover, in gene regulatory system, there are a lot of transcription factors and transcription complexes consisting of two or more proteins working collaboratively. Such collaborations cannot be uncovered by directed regulatory edges but can be discovered by the Markov feature. Given the insufficient dataset, Bayesian network may give a set of models, which explain the data equally well (Pe’er, Regev, Elidan, & Friedman, 2001). In order to do further analysis, we used the statistical confidence to measure the likelihood of a learned edge or a statistic feature (Friedman, 2004; Friedman et al., 2000), which can be calculated by a bootstrap approach as below: † For iZ1... m: – Generate a ‘perturbed’ version of the input transformed dataset by re-sampling q instances where q is smaller than the number of instances of the transformed dataset. – Apply the learning algorithm on the perturbed dataset and induce a network Ni. T.-F. Liu et al. / Expert Systems with Applications 30 (2006) 42–4948 † For each Markov relation f, the confidence is calculated as follows: confðf Þ Z 1=m X i f ðNiÞ (7) where f(Ni) is 1 if f is a feature in Ni and 0 otherwise. We initialized q to be 35 and m to be 50 in our experiment. A high confidence value of a feature for two genes indicates that the learning algorithm can consistently recover the relationship between them. The effectiveness of semi-fixed structure EM learning algorithm can be seen as compared to k-DBN algorithm as SSEM takes advantage of its semi-fixed structure and effective learning algorithm. In the final results, we only adopt the edges whose confidence value is not less than 0.6. The comparison is shown in Table 2 and Fig. 3(b) where only the original structure and the structure learnt by SSEM are presented. K2, SEM and K-DBN gave 1, 4 and 5 edges, respectively, while SSEM gave 11 edges based on the knowledge of YPD (Fig. 3(a)). K2, SEM, K-DBN and SSEM gave 11, 12, 22 and 18 false positive edges, respectively. Fig. 3(b) shows the 29 edges learned by SSEM with confidence greater than 0.6. As shown in following discussion, some of the false positive edges of the 18 edges learnt by SSEM are meaningful in terms of the regulatory relationships between transcription factors or complexes instead of between genes. As shown in Fig. 3(c), CLN3–CDC28 kinase activate transcription factors, SBF (comprised by SWI4 and SWI6 and MBF (comprised by SWI6) and MBP 1), to begin the cell cycle. SBF and MBF then activate about 200 genes in late G1 and S phase include NDD1 and CLN2. CLN2 is one of the important factors to activate CLB2–CDC28 kinase. The complex CLB2– CDC28 inactivates SBF and MBP to shut off G1/S events and the cell goes to G2 phase. CLB2–CDC28 continue to activate a transcription factor containing MCM1, FKH1 (or FKH2, in the experiment, FKH2 is not included as it has too many missing values in the dataset.) and NDD1. This factor activates a set of genes include SWI5 and SCE2 which activate SIC1. SIC1 and some other genes inhibit CLB2-CDC28. It is the yeast cell cycle transcriptional regulatory network, which is essential to yeast development and differentiation. For the detail of the genes and the cell cycle regulatory network, please refer to (Simon et al., 2001). Table 2 Comparison of learning performances on real-life dataset Algorithm Yeast gene subnetwork TE TP1 TP2 K2 12 1 3 SEM 16 4 7 k-DBN 27 5 9 SSEM 29 11 20 TE (Total learnt edges) indicates the number of edges the algorithm learned from the dataset, TP1 indicates the number of true positives verified by Fig. 3(a) and TP2 indicates the number of true positives verified by Fig. 3(a) and (c). When a gene (or a complex) activates or inhibits a transcription complex, it is possible that there is no directed edge from the regulator to any gene of the complex and vice versa. For example, for CDC28, it activates both SWI4-SWI6 and SWI6-MBP 1 when it forms complex with CLN3 while it inhibits them when it forms complex with CLB2. Whereas, there is no direct edge from CDC28 to SWI4, SWI6 and MBP1 in YPD. The use of the hidden variables in our model ensures that we can find such regulatory relationships. As shown in Fig. 3(b), with the verification by Fig. 3(c), 10 more edges are proved biological meaningful in Fig. 3(b), such as CDC28/ SWI4, CDC28/SWI6, CDC28/MBP1, SWI4/CLN2, SIC3/CLB2, etc. Using this criterion, K2, SEM, k-DBN and SSEM got 3, 7, 9 and 20 correct edges and 9, 9, 18 and 9 false positive edges, respectively. We listed the Markov features with confidence greater than 0.6, as shown in Fig. 3(d). There are total 44 Markov features, in which 38 can be verified by Fig. 3(a) and Fig. 3(c). It is clear that there are strong inter-actions among the genes comprising a complex and between complexes which are connected in Fig. 3(c) . We then simplify the network by using the transcription factors or complex, which play roles as a unit in the regulatory system, as the node and the Markov relations between them as the undirected edges, as shown in Fig. 3(e). We completely discover the cell cycle regulatory network, together with 3 false edges between CLN3–CDC28 & SIC1 (which is weak since between them there are only one Markov feature between CLN3 and SIC1), CLN2 & SIC1 and SBF /MBF & SWI5-ACE2. 5. Conclusion The semi-fixed hidden variable model introduces hidden variables to model the important components of gene network, i.e. regulatory proteins. The model makes the network decomposable and the parts of the gene network are fixed using biological knowledge. Compared to the current work on gene network, we integrate the biological knowledge to build the semi-fixed hidden variable model, which is effective and reflects the real life system. It increases the learning performance as well as finds several regulatory relationships, which are difficult to be discovered by employing traditional methods. As our model can model protein complex which plays key role in a lot of biological systems, we foresee our model can be applied to other applications in computational biology. In addition to the model, this paper presents an effective learning algorithm to learn our model. The main disadvantage of our model is that it needs more computing resource to model the hidden variables and it requires more iterations to converge. As the model is based on the Bayesian network framework which can be used to build an probabilistic expert system (Castillo, Gutierrez, & Hadi, 1997; EZ, Mira, Iturralde, & Diaval, 1997), it is easy to develop an probabilistic expert system by making use of the learned network structure and parameters to predict the outcome of the perturbation of the organism, such as diseases. This potential application makes the consequent works meaningful. In the future, we would like to further reduce the learning complexity by utilizing more T.-F. Liu et al. / Expert Systems with Applications 30 (2006) 42–49 49 biological facts. Those biological facts include: (1) regulatory proteins are usually much less than the total proteins, and (2) it is common for several genes to share the same parent set. References Akutsu, T., Miyano, S., & Kuhara S. (1999). Identification of genetic networks from a small number of gene expression patterns under the Boolean network model. In PSB (pp. 17–28). Arkin, A., Shen, P., & Ross, J. (1997). A test case of correlation metric construction of a reaction pathway from measurements. Science, 277, 1275–1279. Barbara, G., Jameson, A., & Witting, F. (1999). Learning Bayesian networks with hidden variables for user modeling. In Proceedings of the IJCAI 99 workshop ‘learning about users’ (pp. 29–34). Barrientos, M. A., & Vargas, J. (1998). A framework for the analysis of dynamic processes based on Bayesian networks and case-based reasoning. Expert Systems with Applications, 15, 287–294. Boyen, X., Friedman, N., & Koller, D. (1999). Discovering the hidden structure of complex dynamic systems. In UAI (pp. 91–100). Brazhnik, P., Fuente, A., & Mendes, P. (2002). Gene networks: How to put the function in genomics. Trends in Biotechnology, 1(20), 467–472. Castillo, E., Gutierrez, J., & Hadi, A. (1997). Expert systems and probabilistic network models. New York: Springer. Chen, T., Filkov, V., & Skiena, S. (1999). Identifying gene regulatory networks from exprimental data. In PSB (pp. 94–103). Chen, T., He, H., & Church, G. (1999). Modeling gene expression with differential equations (pp. 29–40). Cooper, G., & Herskovits, E. (1992). A Bayesian method for the induction of probabilistic networks from data. Machine Learning, 9, 309–347. Cumiskey, M., Levine, J., & Armstrong, D. (2003). Gene network reconstruction using a distributed GA with a backprop local search (pp.33–43). D’haeseleer, P., Liang, S., & Somogyi, R. (2000). Genetic network inference: From co-expression clustering to reverse engineering. Bioinformatics, 16(2), 707–726. D’Haeseleer, P., Wen, X., Fuhrman, S., & Somogyi, R. (1999). Linear modeling of mRNA expression levels during CNS development and injury. In PSB (pp. 41–52). Elidan, G., Lotner, N., Friedman, N., & Koller, D. (2000). Discovering hidden variables: A structure-based approach. In NIPS (pp. 479–485). EZ, F., Mira, J., Iturralde, E., & Diaval, Z. (1997). A Bayesian expert system for echocardiography. Artificial Intelligence in Medicine, 59–73. Friedman, N. (1997). Learning belief networks in the presence of missing values and hidden variables. In Proceedings of the 14th international conference on machine learning (pp. 125–133). Friedman, N. (1998). The Bayesian structure EM algorithm. In UAI (pp. 129–138) Friedman, N. (2004). Inferring cellular networks using probabilistic graphical models. Science, 303(6), 799–805. Friedman, N., Linial, M., Nachman, I., & Pe’er, D. (2000). Using Bayesian networks to analyze expression data. In RECOMB (pp. 127–135). Friedman, N., Nachman, I., & Peer, K. (1999). Learning Bayesian network structure from massive datasets: The ‘sparse candidate’ algorithm, In UAI (pp. 206–215). Futcher, B. (2002). Transcriptional regulatory networks and yeast cell cycle. Current Opinion in Cell Biology, 14, 676–683. Gustavo, A., Sucar, L., & Villavicencio, A. (1998). Probabilistic temporal reasoning and its application to fossil power plant operation. Expert Systems with Applications, 15, 317–324. Hartemink, A., Gifford, D., Jaakkola, T., & Young, R. (2001). Using graphical models and genomic expression data to statistically validate models of genetic regulatory networks, In PSB (pp. 437–449). Hasty, J., McMillen, D., Isaacs, F., & Collins, J. (2001). Computational studies of gene regulatory networks: In numero molecular biology. Nature Reviews Genetics, 2(4), 268–279. Hodges, P., McKee, A., Davis, B., Payne, W., & Garrels, J. (1999). The yeast proteome database (YPD): A model for the organization and presentation of genome-wide functional data. Nucleic Acids Research, 27(1), 69–73. Imoto, S., Goto, T., & Miyano, S. (2002). Estimation of genetic networks and functional structures between genes by using Bayesian networks and nonparametric regression. In PSB (pp. 175–186). Imoto, S., Higuchi, T., Goto, T., Tashiro, K., Kuhara, S., & Miyano, S. (2003). Combining microarrays and biological knowledge for estimating gene networks via Bayesian network. In CSB (pp. 104–113). Kang, C. W., & Gola, M. W. (1999). A Bayesian belief network-based advisory system for operational availability focused diagnosis of complex nuclear power systems. Expert Systems with Applications, 17(1), 21–32. Kolpakov, F., Ananko, E., Kolesov, G., & Kolchanov, N. (1998). A gene network database and its automated visualization. Bioinformatics, 14, 529– 537. Li, F., Long, T., Lu, Y., Ouyang, Q., & Tang, C. (2004). The yeast cell-cycle network is robustly designed. Proceedings of the National Academy of Sciences USA, 101(14), 4781–4786. Liang, S., Fuhrman, S., & Somogyi, R. (1998). Reveal, a general reverse engineering algorithm for inference of genetic network architectures. In PSB (pp. 18–20). Liu, T. F., Sung, W. K., & Mittal, A. (2004). Learning multi-time delay gene network using Bayesian network framework. In ICTAI (pp. 640–645). Louis, M., Louis, A., & Becskei (2002). Binary and graded responses in gene networks. Science STKE, 143, pe33. Murphy, K., & Mian, S. (1999). Modelling gene expression data using dynamic Bayesian networks. Technical report, University of California, Berkeley. Oatley, G., & Ewart, B. (2003). Crimes analysis software: ‘Pins in maps’, clustering and bayes net prediction. Expert Systems with Applications, 25, 569–588. Pe’er, D., Regev, A., Elidan, G., & Friedman, N. (2001). Inferring subnetworks from perburbed expression profiles. Bioinformatics, 17(Suppl. 1), S215– S224. Segal, E., Shapira, M., Regev, A., Pe’er, D., Botstein, D., & Koller, D. (2003). Module networks: Identifying regulatory modules and their condition- specific regulators from gene expression data. Nature Genetics, 34(2), 166– 176. Simon, I., Barnett, J., Hannett, N., Harbison, C., Rinaldi, N., Volkert, T., et al. (2001). Serial regulation of transcriptional regulators in the yeast cell cycle. Cell, 106(6), 697–708. Someren, E., Wessels, L., & Reinders, M. (2000). Linear modeling of genetic networks from experimental data. In ISMB (pp. 355–366). Spellman, P., Sherlock, G., & Futcher, B. (1998). Comprehensive identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Molecular Biology of the Cell, 9,(3273), 3297. Sucar, L. E., & Miriam, M. A. (1998). Interactive structural learning of Bayesian networks. Expert Systems with Applications, 15(3–4), 325–332. Viaene, S., Dedene G., & Derrig, R. A. (in press). Auto claim fraud detection using Bayesian learning neural networks. Expert Systems with Applications. Wagner, A. (2002). Estimating coarse gene network structure from large-scale gene perturbation data. Genome Research, 12(2), 309–315. Wahde, M., & Hertz, J. (2000). Course-grained reverse engineering of genetic regulatory networks. Biosystems, 55, 129–136. Watanabe, S. (1998). Algorithms for inference of genetic networks AIGNET. In International workshop on genome informatics (pp. 274–275). Wessels, L., Someren, E. V., & Reinders, M. (2001). A comparison of genetic network model. In PSB, Vol. 6 (pp. 508–519). Wyrick, J., & Young, R. (2002). Deciphering gene expression regulatory network. Current Opinion in Genetics and Development, 12, 130–136. Yaki, M., Tominaga, D., Okamoto, M., Watanabe, S., & Eguchi, Y. (2001). Development of a system for the inference of large scale genetic networks. In PSB (pp. 446–458). Model gene network by semi-fixed Bayesian network Introduction Model gene network as a semi-fixed networkwith hidden variables Semi-fixed structure EM learning algorithm Experimental results and comparison Experiment on artificial datasets Experiments on real-life data Conclusion References 
work_pdphn5l3jbcsznfelp5dlhdaom	----	Elsevier required licence: © <2016>. This manuscript version is made available under the  CC‐BY‐NC‐ND 4.0 license http://creativecommons.org/licenses/by‐nc‐nd/4.0/    A Disaster Management Knowledge Repository (DMKR) and Sharing System Siti Hajar Othman, Ghassan Beydoun Abstract. Disaster Management (DM) is collaborative and complex. Roles involved in DM processes often cut across many organizational boundaries and are dynamic. Knowledge involved is enormous and diverse. It includes information related to varieties of disasters, roles, plans and operations. Many practices may also vary across different regions and authorities. Structuring, storing and reusing DM knowledge is the challenge tackled in this paper. The paper presents a Metamodel-based Disaster Management Knowledge Repository (DMKR) tool, DMKR1.0, to facilitate collaboration and DM knowledge sharing. The knowledge representation that underpins the repository structure is a disaster management modeling language that makes statements about what can be expressed in various DM models. Using the tailored DM language, the tool developed offers a flexible structure to allow the storage and retrieval not only of observed and measured data, but also interpretative and inferred information of the disaster management knowledge. The paper describes the development, functionality and implementation of DMKR. The validation of DMKR illustrates how users can easily instantiate DM models to communicate and generalize their knowledge for the benefit of sharing it within their community. Keywords: Knowledge Sharing; Decision Making; Disaster Management; Metamodels A Disaster Management Knowledge Repository (DMKR) and Sharing System Siti Hajar Othman, Ghassan Beydoun Abstract. Disaster Management (DM) is collaborative and complex. Roles involved in DM processes often cut across many organizational boundaries and are dynamic. Knowledge involved is enormous and diverse. It includes information related to varieties of disasters, roles, plans and operations. Many practices may also vary across different regions and authorities. Structuring, storing and reusing DM knowledge is the challenge tackled in this paper. The paper presents a Metamodel-based Disaster Management Knowledge Repository (DMKR) tool, DMKR1.0, to facilitate collaboration and DM knowledge sharing. The knowledge representation that underpins the repository structure is a disaster management modeling language that makes statements about what can be expressed in various DM models. Using the tailored DM language, the tool developed offers a flexible structure to allow the storage and retrieval not only of observed and measured data, but also interpretative and inferred information of the disaster management knowledge. The paper describes the development, functionality and implementation of DMKR. The validation of DMKR illustrates how users can easily instantiate DM models to communicate and generalize their knowledge for the benefit of sharing it within their community. 1.0 Introduction Disaster Management (DM) processes involve many interacting elements e.g.: people, authority, emergency teams, resources, procedures, uncertain environmental situations and many more. Modeling coordination of DM activities is tremendously hard and complex. DM activities often extend across various government sectors, non-governmental organizations/industry, from international levels down to state or region levels and may also include individual people and various facets of society. It is often unclear what are the exact tasks and responsibilities before, during or after a disaster strikes. The complexity of DM and the failures of DM agencies can be observed in many recent examples, e.g. the management of the Swine-Flu (H1N1) pandemic hitting Australian shores in large numbers through cruise ships in 2009 (Larcombe, Moloney et al. 2009), or the devastating communication failures in the Victorian bushfires in Victoria in 2008 (Cordner, Woodford et al. 2011). Observed failures often surface as the required expertise not being available in a timely manner or a systemic inability to timely recognize and identify the required expertise. Reusing expertise is overlooked as it is often perceived as too specific to kinds of events such as floods, bushfires, tsunamis, pandemics or earthquakes. This research advocates the use of a middle knowledge layer to enable DM practitioners to discern disaster dependent and disaster-independent features in the challenges that they face. For this purpose, the Metamodel-based Disaster Management Knowledge Repository (DMKR) is constructed. The structure of the repository is an essential factor in obtaining good retrieval results for knowledge retrieval system (Shiva and Shala 2007). DMKR structure follows a tailored DM metamodel that recently appeared in (Othman, Beydoun et al. 2014). DMKR facilitates knowledge reuse and timely identification of required DM knowledge sources. It allows a unified access to various DM experiences and offers a common DM description language. The work advocates that combining expertise used in various disasters will enable the best approach in managing a new disaster. It provides the following benefits:  Facilitating communication among different disaster emergency users;  Simplifying the teaching of new DM approaches through a set of semantic rules;  Providing guidelines for creating DM models which can cover specific phases of DM (e.g: Response Phase Earthquake Emergency Response Model, Mitigation Phase Bushfire Risk Reduction Model);  Highlighting scope for improvement in a DM practice through validation against other existing metamodels in DM field. We developed and validated a DM language as a middle layer of knowledge in the form of a disaster-independent metamodel to unify knowledge from different disaster experiences in (Othman, Beydoun et al 2014). This required tuning the metamodelling process to DM to ensure that appropriate knowledge sources (candidate models) are identified and used as input to the process. The resulting language, DMM, was theoretically validated showing how 89 DM models from the literature can be generated from the DMM (Othman, Beydoun et al 2014). In this paper, we deploy the language by demonstrating a viable and a unifying DM knowledge repository. This paper develops an actual knowledge sharing system, the DMKR system which stores DM solution modelling components and guides users on how to reconstruct solutions as new contexts arise. To store DM knowledge as reusable modelling components, DMKR has a metamodel driven interface to guide DM experts to articulate their knowledge appropriately. To illustrate the expressive power of the system, a number of case studies to describe real disaster situations are used in the DMKR implementation. Besides constructing DM solution models based for specific disaster problems presented to the system, DM stakeholders are also provided with the solutions from a range of disaster categories (e.g.: earthquake, landslide, nuclear meltdown and etc.). This enables mixing and matching solutions across different disasters. Typical, users of DMKR can be disaster managers (local, state or federal), monitoring personnel, coordinators, aid agencies or even researchers who may wish to study the domain. The rest of this paper is organized as follows: Section 2 describes the DM model driven background and related work that underpin the conceptual foundation of this paper. Section 3 describes the development and the architecture of a model driven DM system. Section 4 demonstrates the usefulness of the system implementing it and illustrating in real-world disaster problems. Specifically, a bushfire disaster exemplar is used to illustrate the knowledge capture and reuse facilitated by the system. Finally, Section 5 concludes with recommendations for future research. 2.0 Background and Related Work A model is an abstract representation of a real world domain typically used to manage complexity (Levendovszky, Rumpe et al. 2010). It generally consists of a collection of two elements: concepts and relationships. Concepts characterize domain entities and relationships characterize links between them (Trabelsi, Atitallah et al. 2011). A model explicitly expresses the structure, behaviour and other properties in a domain and ideally has a causal connection to the real world (Aßmann, Zschaler et al. 2006). Our approach for DM knowledge sharing can be said to be model-driven. It is inspired by model driven systems development (Colin, Thomas et al. 2003, Stahl, Voelter et al. 2005). Instead of requiring software developers to detail how a system is implemented, in a model-driven approach software models specify what the functionality and the architecture of the system to be used are (Colin, Thomas et al. 2003) via the use of software models developed in a specific language. With syntactically correct models, using only allowed symbols and conforming to rules, the modelling language facilitates sharing of the outcome of the modelling process. Developers are able to abstract and share knowledge using models as starting points. To formally describe the semantics underpining a formal modelling language, a metamodel is required, without which semantics of domain models can be ambiguous (2000). Applying a model driven approach to share DM knowledge has the added benefit of also making knowledge accessible to non-technical specialists and newcomers to DM. Similar to model-driven software development, this also requires a DM modelling language to specify all elements with which any model can be described. Model transformations can then enable precise and timely knowledge sharing communication across the various phases of the process (Trabelsi, Atitallah et al. 2011). Similarly, in DM, these are also required to enable the unified access to various DM knowledge models and to also communicate the knowledge across different disasters and DM activities. The metamodel we use in this paper to specify DM modelling constructs is DMM. This is a DMM specific metamodel which we synthesized and validated created in (Othman, Beydoun et al. 2014). DMM has four sets of concept classes: the Mitigation (shown by Figure 2), Preparedness (Figure 3), Response (Figure 4) and Recovery (Figure 5) class of concepts. Each set represent a corresponding DM phase. This clearly describes the DM domain to its users. The relationship between the metamodel and domain models are described through model transformations (Vytautas Stuikys 2010) which convert one model to another model (OMG 2003). In this work, we deploy the transformations prescribed in the metamodelling framework of MOF, the standard for software metamodelling offered by a OMG (2002). MOF is essentially a meta-metamodel that defines a common way for capturing the diversity of modelling standards and interchange constructs that are used in model driven software engineering. In (Cook 2004), the MOF is defined as a framework designed for defining modelling languages. MOF is based on a hierarchical architecture of four meta-layers as illustrated in Figure 1, MOF is primarily about defining information models for metadata. It uses an object-modelling framework that is essentially a subset of the UML core. The main four modelling concepts in MOF are classes, association, data types and package. The main advantages of OMG standard are wide acceptance, coverage of many domains and cohesion among the standard (M. Picka 2004). MOF has four layers, M0, M1, M2 and M3. It has different views on modelling at different layers of details. It is strictly hierarchical. Concepts at any given layer (below M3) belong to a concept from the layer above. Any concept in any given layer (above M0) can be instantiated at the layer below. M3-Level is reserved for Meta-metamodel element, comprised of the description of the structure and semantics of meta-meta data. M2-Level is for the metamodel layer (instance of meta-metamodel), comprised of the descriptions of the structure and semantics of metadata. M1-Level is designed for the model layer (instance of metamodel), comprised of the metadata that describe data in the information layer. The lowest level, M0, is dedicated for user models (instance of model and also called as information layer). In this paper, M0-Level will cover the data that a disaster model describes at M1. Figure 1 The derivation of various DM solution models in three MOF modelling layers Figure 2 The DMM - Mitigation-phase class of concepts The system introduced in this paper, DMKR 1.0, represents how all the components associated the DMM are operationalised. It is inspired by method engineering, where a metamodel is used to index a software development methodology as required by a software process model (Firesmith 2006). The design of DMKR and the theories underlying it are described in the next section. Figure 3 The DMM - Preparedness-phase class of concepts Figure 4 The DMM - Response-phase class of concepts 3.0 The Disaster Management Knowledge Repository System (DMKR) This section describes the architecture of the knowledge repository, DMKR 1.0. This repository utilises DMM as a representational infrastructure to store and structure complex DM knowledge and processes. DMKR integrates the metamodel with retrieval and modeling support components. It has a model driven interface to ensure that users can easily access and reuse the stored knowledge (see Figure 6). The stored knowledge can be reused to develop varying DM solutions as DM contexts vary. It has a modelling tool to develop variety kinds of new DM models based on past DM problems and knowledge retrieved. New DM models can be derived through operationalisations of vertical model transformations from DMM. The transformations used are: Instantiation (to instantiate a single Concept), and Conformance (to instantiate a Model holistically from a number of concepts). For example, an Emergency Response Model (an M1-Model) is a model conformance from the DMM (Figure 1). At this level, M1-Fragment is used to represent DM models. An M1 solution derived from DMM can in turn derive various models of real-world disasters at the M0-level e.g.: response after a nuclear meltdown, rescue and search after an earthquake, flood rescue model etc. This M0-Level is the DM User Model and represents real-world DM activities sought by DM users. Concepts at this level are represented through M0-Fragment. The combination of all M0-Fragments would form the DM solution model (M0-User Model) from the system. Figure 5 The DMM - Recovery-phase class of concepts 3.1 DMKR Architecture DMKR design has three components (see Figure 7): (a) Module A – DMM and its tools component: This contains four sub-components: Consistency Checker Module, Metamodel Formulation Module, Modelling Constructor Module, DMM Concepts Database. Each of these will be detailed in Section 3.3. (b) Module B – The DM knowledge base (The “DM Knowledge Repository”): The repository schema of this database follows the structure of its metamodel. It stores a collection of emergency organisation and coordination activities executed by different agencies. (c) Module C – The retrieval module (The “DM Retrieval”): This module functions to assist in deriving the best DM model solution according to the disaster in hand. The DMKR is built using a relational database management system (RDBMS). The usage of RDBMSs in metamodelling applications is common, e.g. in geosciences (Viswanathan and Schneider 2010), for geographic data (Shanzhen and Yuntao 2011) and web-based interface components (Panesar-Walawege, Knutsen et al. 2011). A collection of relational database tables is created to manage the variety of modelling tools element in DMKR. These include: Model Fragments (M2-Fragment, M1-Fragment, and M0-Fragment); Model Database (M2-DMM, M1- Model and M0-User Model); Concept and Class elements (Terminology, Notations); DM Knowledge (Data for Model Fragments); User Information (DM Real-user); and Disaster information (Type of disaster). Figure 6 The main page of the DMKR system Figure 7 The DMKR System Architecture 3.2 Knowledge Representation in DMKR In DMKR, modelling structures to support specific DM activities and related past experiences to these activities are available side by side. The experiences underpin the DM knowledge stored as Model Fragments. In DMKR, knowledge is populated in a structured knowledge storage unit known as Model Fragments - the building blocks of reuse. These reusable knowledge units can be mixed and matched as DM context demands. The use of modelling fragments is adapted from method engineering field where a fragment describes a cohesive method component (Firesmith 2006). DMM itself is expressed as a collection of M2-Fragments making statements about what can be expressed in valid DM models. Each concept in DMM has its own M2-Fragment. In this research, a fragment consists of DM operations and modelling structures that are prerequisites for the applicability of the operations. Following the MOF metamodelling framework, a model fragment is available at three levels of abstractions (Table 1): The M2-Metamodel Level, the M1-DM Model, and the M0-DM User Model. The M2-Fragment is used to represent fragment for concept in the M2-Metamodel (DMM), the M1-Fragment is used to represent an instance of concept in the M1-DM Model, and the M0-Fragment is used to represent information of the instance of concept in the M0-DM User Model. When an M1-Model is developed, DM knowledge of every single unit fragments of the model can later be retrieved and reused. The user is guided to user identify relations between unit fragments and DM knowledge. This offers fast and precise knowledge derivation. Object oriented inheritance links these fragments across the three levels of MOF. For DMM, when a model fragment is indexed by the metamodel, the actual operational DM knowledge will need to be generated as M0-Fragments. To enable effective usage of concepts to create combination of concepts at the M1-Level and M0-Level, DMKR users are given explanations about domain concepts in the metamodel level. These explanations provide examples and more details of usage. Table 1 Representation of DMM Concept in MOF modelling paradigm Model Level Level Name Description Elements M2 DMM Metamodel  Defines the language for specifying a DM model  Consists of MetaClasses, MetaAttributes, MetaOperations, MetaRelationships of DMM concepts M2-Fragment (Metamodel fragments) - Contains DMMClass (e.g.: Concept: “Evacuation”, “Aid”, “EmergencyPlan”, “Warning”) M1 DM Model Instance of DMM Metamodel  Models a specific information of DM problems ( e.g.: Emergency Preparedness Model – all the required components for this model are defined in DMM) M1-Fragment (Data Model Fragments) - Contains DMMObject (e.g.: a) setupEvacuationCentre (from “Evacuation”), b) refugeeShelterDonation (from “Aid” concept), c) identifyDamageCost (from “DamageAssessment” concept), d) disseminateAlertWarning (from “Warning” concept) M0 DM User Model Instance of DM Model  Models user’s real world data  DM Model represents a real DM M0-Fragment (User Model Fragments) - Contains DMMInstance of Real World Data (e.g.: Data on Evacuation during Flood disaster in Brisbane, Aid distribution during Japan Tsunami) The object-oriented paradigm is used to design and implement Model Fragment in DMKR. Every DMM concept is represented as Class, DMMClass. DM Concept::= DMMClass (Note: Symbol ::= is an equivalent). Each DMMClass has the following fields: name, ID, terminology, attribute, operations, and relationships. That is: DMMClass::= <DMMClassName, DMMClassID, DMMClassTerminology, lDMMClassAttribute, DMMClassOperation and DMMClassRelationship> where:  DMMClassName represents a DM concept name  DMMClassTerminology represents a DM concept definition  DMMClassAttribute represent a DM requirement  DMMClassOperation represent a DM task  DMMClassRelationship represents the relationship with other concept/s  DMMClassID is a unique identifier for the concept Each DMM concept is represented as a DMMClass which has its own M2-Fragment, a unit fragment. The coupling between the DM operations and the contextual knowledge required for their applicability, in unit fragments (specific for this research, a Model Fragment is used to represent this unit fragments). Each concept will have its own concept’s task, requirement, plan, responsible user etc. These are refined down the MOF hierarchy through the corresponding unit fragments. An example of this is shown in Table 2 giving an example of a DMMClass concept, Evacuation, a Preparedness-phase concept and how it becomes refined from M2 to M0 according to DM context. Table 2 The DMMClass of Evacuation concept Class EVACUATION ClassID BP07 Definition: An organised, phased and supervised withdrawal, dispersal or removal of civilians from dangerous or potentially dangerous areas, and their reception and care in safe areas. Relations PreparednessPlan (BP01) Relation type: Association, Relation name: “Follows” Warning (BP05) Relation type: Association, Relation name: “Activates” A tt ri b u te s evacuationName: The name of the process (e.g.: Queensland Flood Evacuation). evacuationRoad: The network of roads where evacuees will travel to reach a safe evacuation destination. evacuationCentres: The safe destination for people who live in disaster risk areas as a temporary refuge. evacuationTransport: The transportation facilities that have a capacity and need for resources (e.g.: fuel) to travel (e.g.: ambulance, helicopter, fire engine). evacuationAlertProcedure: Communication of public warning. evacuationAuthority: Leaders who have the authority to order evacuations. evacuationExercise: A program to practice the need to evacuate under certain circumstances and play specific actions to be taken when a real disaster happens. evacuationServiceTeam: Availability of emergency services facilities and personnel (e.g.: personnel and equipment for search and rescue, hospital and medical services). hazardZone: The zone areas where the disaster took place which has one or more locations (e.g.: North Queensland during Cyclone Yasi 2010). locationAtRisk: The location where people at risk should be moved (e.g.: Townsville and Innisfail in North Queensland during Cyclone Yasi 2010). peopleAtRisk: People who live in hazard zone areas who must be evacuated. evacuationPlan: The plan of evacuation that needs to be followed by all stakeholders involved (e.g. “Emergency Management Australia (EMA) Flood Evacuation Plan”). O p e ra ti o n s performPublicEvacuationProcedure(): A process to implements the evacuation procedure by all victims and emergency teams. alertWithDisasterWarning(): A situation which requires people to listen to a nominated radio station or watch a nominated TV channel for further advice regarding the evacuation warnings. makeEvacuationDecisionWarning(): A process of the authority making a decision as to whether to evacuate or not, which will be assisted by the availability of timely and relevant information. broadcastPublicDisasterWarning(): A process of disseminating the evacuation information and warning through a variety of media (print, electronic, community announcement, word-of-mouth, etc). evacuatePeopleAtRisk(): A process of evacuating the victims by the response emergency teams with a special consideration of special needs victims. setupAndOrganizeEvacuationCentre(): An establishment and management of the evacuation centres which provides basic human needs including accommodation, food, water, first aid, hygiene facilities, clothing, blankets, information and referral services, among others. returnEvacuee(): A final stage of the evacuation process. It will be necessary to assess the disaster area to determine if return is possible and identify any special conditions which may need to be imposed. 3.3 DMKR Detailed Design We use a database management system to implement the components of DMKR. It has these four components to support modelling, storage and retrieval : a) Consistency Checker Module, b) Metamodel Formulation Module, c) Modelling Constructor Module, d) DMM Concepts Database. Each of these components is detailed in this section. 3.3.1 Consistency Checker Module Model transformations, at M2, M1 and M0 levels, are specified via a set of rules defining the modelling constructs and model composition mechanisms for its DMKR modelling design activities. The Consistency Checker Module is a library of these rules which act as policy guidelines for the correct utilisation of the DMM modelling components. The module ensures compliance of model transformations to these rules. The rules define semantic properties of metamodel transformations as prescribed in Wachsmuth (2007) and Lammel (2001). Thus a consistent transformation from a source model (DMM) to a corresponding target model (either an M1-Model or an M0-User Model) is guaranteed. Figure 8 shows a sample of six rules (for the Rule_ID of R04c, RR04a, R04, RR03a, and RR06). For example Rule RR04a formalises the creation of an “Instance model of DMM”: DMM Rule RR04a Rule Syntax : Ґ(µDMM) := [ ι (µDMM) |= µDMM ] Rule Meaning : The set of all DM model (metamodel instances) conforming to a Disaster Management metamodel, µDMM RR04a rule ensures each newly created DM model (metamodel instances) is compliant to the DMM, (µDMM). This rule will mostly be used in association with other rules according to the type of model being developed. For example, if User A wants to develop a new Preparedness model from the DMM, a rule specific to initialise the element is used: DMM Rule RC01b Rule Meaning : The set of concepts for a Preparedness model Rule Syntax : CM (µDMMp) ::=[PreparednessPlan, PreparednessOrganization, PreparednessTask, SuppliesRegistry, Warning, PreparednessGoal, Evacuation, BeforeDisaster, Event, DecisionMaking, Administration, EmergencyPublicInformation, Pre-Position, DisasterFactor, Exposure, DisasterRisk, Training, PreparednessTeam, Media, MutualAidAgreement, PublicEducation, PublicAwareness, Resource, Monitoring, AidAgency] Figure 8 An extract form the rules repository 3.3.2 Metamodel Formulation Module DMM classes are stored as a Concept_table and DMM corresponding phase classes are stored in Phase_table. Figure 9 shows the relationship between both tables in which: a) shows a table’s relationship, and b) shows data values of each tables. In this figure, the MitigationPlan is a concept in the Mitigation-phase class (Phase ID =1). In Concept_table, the Concept_ID of the concept has the same ID with the M2Fragment_ID, AM01. This shows how one M2-Fragment exists for every single concept in DMM. To formalise the usage of concepts in each M2- Fragment (a DMM class), specific rules in the Consistency Checker Module are used. The following four rules formalise the Instantiation of a concept (Rule R04a and R04b) and a Conformance of model (Rule RR07a and RR07b) for two M2-Fragments (the Mitigation and Preparedness class): DMM Rule R04a Rule Syntax : CM (µDMMm) Rule Meaning : Mitigation Concepts of the Disaster Management Metamodel Rule Construct : MitigationConcept::= (<MitigationDMMClass AND MitigationM2Fragment>) DMM Rule R04b Rule Syntax : CM (µDMMp) Rule Meaning : Preparedness Concepts of the Disaster Management Metamodel Rule Construct : PreparednessConcept::= (<PreparednessDMMClass AND PreparednessM2Fragment>) DMM Rule R07a Rule Syntax : ι(µDMMm) Rule Meaning : Mitigation Model of Disaster Management Rule Construct : MitigationModel::= (<MitigationM1Model> AND <MitigationM1Fragment>) AND (<MitigationM0Model> AND <MitigationM0Fragment>) DMM Rule R07b Rule Syntax : ι(µDMMp) Rule Meaning : Preparedness Model of Disaster Management Rule Construct : MitigationModel::= (<PreparednessM1Model> AND < PreparednessM1Fragment>) AND (<PreparednessM0Model> AND < PreparednessM0Fragment>) Figure 9 Relationship between Concept and Phases; (a) Concept_table and Phase_table, (b) Data values of the Concept_table and Phase_table Rule R04a ensures that every single mitigation-phase concept consists of a DMMClass and a model fragment. Rule R05a ensures that the Mitigation-phase model (DM solution model of DMKR) consists of an M1-Model and an M0-DM User Model derived from the Mitigation-phase class. The different fragments and their associated models are managed through the storing of data for Fragment (M2Fragment_table, M1Fragment_table and M0Fragment_table) and Model (M2Model_table, M1Model_table and M0Model_table). A single DMMClass has only one M2-Fragment and the combination of all M2-Fragments form the M2-Model (Metamodel). As shown in Figure 9 (a), the DMM has many M2-Fragments in the model. As shown in Figure 9 (b), at any one time many M1- models may consist of many M1-Fragments and many M0-User Models may consist of many M0-Fragments. Table 3 shows an example of M2-Fragments with its respective unit fragments that belongs to the MassCasualtyManagement class. These fragments underpin the formulation of M1 and M0-Fragments and models. Table 3 Part of Unit Fragments of the MassCasualtyManagement class Unit ID Unit Name Unit Description CR24a Tactical Procedure The procedure which provides standard guidelines for mass casualty operation in a multi-casualty emergency medical situation. CR24b Organization Structure The organisational structure of multi-responsible teams who are responsible for the overall management of multi-casualty incidents (e.g.: Patient Transportation Group, Medical Supply Group, Medical Communication Group, Victims Treatment Group). CR24c Equipment The equipment needed to perform a mass casualty operation (e.g.: disaster kits containing bar coded wrist bands, envelopes for storing victim's items, sealed bags for storing dead bodies for medical treatment or mortuary). oCR24a Mass Casualty Setup The implementation of mass casualty procedure (e.g.: prepare disaster kits containing: bar coded wrist bands, different sized envelopes for personal effects, a collection of all items taken from the victims, dead or alive, and place them in the bags. Sealed bags are transported with the victims for medical treatment or the mortuary). oCR24b Classify Casualty Classifying the victims of a disaster (e.g.: victims who died in the Medical Post (MP) certified by the MP Doctor. Victims who died in the field need to be taken to the hospital and certified by a qualified doctor). oCR24c Move Patient The movement of patients from the scene of the incident to medical facilities in close coordination within the Treatment Dispatch Manager(s) in the Medical Group and the medical facilities. Figure 10 shows the M2-Fragment of an Evacuation Class with its corresponding Unit Fragment. M1-Models describe a specific DM problem. An M1-Model is formed by many M1-Fragments. M1-Fragments can be instantiated to represent many M0-Fragments (real world data). Examples of suitable candidates to be represented as Figure 10 The M2-Fragment of an Evacuation Class with its respective Unit Fragment Figure 11 The M1-Model of a Disaster Mitigation Model M1-Models are: preparedness operations before floods, the coordination of resources during a flood response, and the reconstruction and rehabilitation following a bushfire. In Section 4.0, real-world DM problems for various bushfire problems are used to implement M1-Models. For example, a Disaster Mitigation Model is the M1-Model conformance from the Mitigation-phase class. This model is created from a combination of a few DMMClass: the StructuralMitigation, the Insurance, the Non-StructuralMitigation, the NeedsPlanning, the RiskAnalysis etc. By using a few unit fragments taken from each class, this model can later be formed. The relationship between the M1- Figure 12 An M0-Fragment of the HomeRadiationMonitoringDetection, is instantiated from an M1-Fragment of the PerformPublicEmergencyPlanModelling Decision associated with Activity Tasks Activity Tasks SetupStructuralMitigationPlan(), a unit fragment of the StructuralMitigation concept Model and M1-Fragment can be described as for each M1-Model; the model has many M1-Fragments. In Figure 11, one unit fragment of a ‘setup structural mitigation plan’ is instantiated from the StructuralMitigation class. For this Instantiation process, a special notation is used to link a relationship between model-metamodel. The stereotype notation of (<<STRUCTURALMITIGATION>>) is used to signify that the SetupStructuralMitigationPlan() is a class operation that comes from the StructuralMitigation class (shown by an arrow). Other instances of unit fragment taken from the DMMClass are: a ‘perform risk analyses’ from the RiskAnalysis, and a ‘perform hazard assessment’ from the HazardAssessment class. Figure 12 shows a sample of one M0-Fragment, Home Radiation Monitoring Detection, in DMKR that represent DM knowledge of a Nuclear Radiation problem. This is instantiated from the M1-Fragment, PerformPublicEmergencyPlan(), which in turn is derived from the M2-Fragment, PreparednessTask. In this M0-Fragment, the real activity of how a home radiation monitoring detection process should be conducted during a nuclear disaster situation is demonstrated. 3.2.3 Modelling Constructor Module The Modelling Constructor Module contains the library of notations/symbols that can be used by system users to develop their semantic DM model in DMKR. In this paper, modelling notations offered by the Unified Modelling Language (UML) are used. A model derived from DMM can be modelled according to UML-based representations: a) ‘Class Diagram-view’ (instantiation of concepts is in form of Class), b) ‘State Transition Diagram-view’ (mapping events – start, stop, condition etc.), c) ‘Use Case Diagram-view’ (model actors and their respective functions), or d) ‘Sequence Diagram-view’ (model activity sequence order). Modelling Constructor Module is the library that provides notations for customisation in DMKR. The representation of a derived M1-Model during a conformance of the model from the DMM is another issue in DMKR. As described by Brottier et al. (2006), a transformation offers an opportunity to improve the readability of the M1-Model rather than use the same notation. Sometimes, improving the readability of the model may involve hiding some relationships in the M1-Model. Therefore, when users create the M1-Model in a class diagram-view, the representation of the model can be seen exactly as its conformant metamodel (because the DMM is developed in this view). However, in other certain cases, users can also present their new M1-Model in a human-friendly representation such as the State Transition Diagram or in Use Case Diagram (UML models). DMKR gives flexibility for a system user to create their model in the modelling environment provided in the system. 3.2.4 DMM Concepts database This database is implemented as the following three smaller databases: 1. The DMMClassTerminology database: This contains DMM concept definitions. 2. The DMMClassRelationship database is the collection of relationships between concepts within and across Mitigation, Preparedness, Response and Recovery-phase class of concepts (M2-Level). These relationships guide users on integration of model components (concepts) as they create a new model for their DM context. For instance, when a user creates a new M1-Model model of the “Earthquake Rescue Model”, one of the important elements that will be derived from the Response-phase class is the Rescue concept. In the M2-Fragment corresponding to this concept, there are seven relationships to other concepts that users need to consider in formulating their M1-Model e.g. “IsPerformedBy” linked to the EmergencyManagementTeam concept or “Follows” linked to the “EmergencyPlan” concept. 3. The third database specifies the attributes and the operations of each class. To illustrate this, Figure 13 shows three attributes for the Evacuation concept: evacuationRoad, evacuationCentres and evacuationAlertProcedure. It also shows three operations: performPublicEvacuationProcedure(), evacuatePeopleAtRisk() and setupAndOrganizeEvacuationCentre(). 4. Demonstrating DMKR in Real-world Disaster Problems: Bushfire Case Study In DMKR, system users are given flexibility to prepare their model. How a model can be formed is varied according to the intention of the user for their new model. This section illustrates how knowledge can be stored and later retrieved and reused using DMKR. An focusing on Bushfire knowledge is shown. In Section 4.1, the storage of Bushfire knowledge is shown. In Section 4.2, retrieving this knowledge according to varying bushfire scenarios is illustrated. 4.1 Storing Bushfire Knowledge in DMKR Bushfire management knowledge is typically available in sources in text form available in reports, books and disaster-related websites. Storing this knowledge begins following the structure of the M2-Fragment, DMMClass). For starters, knowledge for the ResponseTask for a bushfire is stored, based on previous bushfire experiences describing how various DM users operate during the response phase. Their specific actions and roles are stored. Knowledge is collected from the following sources: a website (Johnston 2009); a report (Emergency Management Australia (EMA) 2000); and four books (Paterson 2007), (Troy and Kennedy 2007) (Machin and Hentze 2007), (Stephens and Collins 2007). The stored knowledge combines knowledge of different groups of stakeholders. For example, solution taken from Stephen and Collins (2007) provides knowledge of three different groups: i) media, ii) emergency team, and iii) affected people and community, whereas, a solution from Machin and Hentze (2007) provides knowledge for the emergency team and monitoring user. The knowledge stored in the M0-Fragment of the ResponseTask concept represents tasks of different users. Later other groups of users (e.g.: task of government, task of people at risk or task by media and etc.) will be able to reuse this knowledge. Figure 13 Samples of data values for the Attribute_table and the Operation_table Further bushfire knowledge is sourced from the report of the ‘2009 Bushfire Royal Commission Final Report’ (Parliament of Victoria 2010). This knowledge is presented through Figures 8.1 (a) to (c). The report describes the real experiences of the bushfire tragedy of February 2009 in Victoria, Australia. This was a huge tragedy, claiming 173 lives, destroying at least 1834 homes, rendering 7500 people homeless and releasing millions of tonnes of carbon into the atmosphere. Figure 14 (a) describes knowledge of bushfire factors related to this tragedy. The knowledge is stored as an M0-Fragment of the DisasterFactor concept. In Figure 14(b), the Recommendation 48 from the report describes knowledge for StructuralMitigation. Figure 14(c) contains knowledge for the Legislation concept captured from Recommendation 27. Unlike knowledge stored in the ResponseTask concept which categorises bushfire knowledge according to users: government, people at risk, media. Figure 14 categorises the knowledge according to DMM concepts: DisasterFactor, StructuralMitigation and Legislation. Data presented in Figure 14(a) is a collection of knowledge representing identified factors in the 2009 Victoria Bushfire. This is stored in DMKR as illustrated in Table 4, as DMMClassAttribute of the DisasterFactor concept: the GravityFactor, the TriggeringFactor and the ComplexityFactor. (a) 2009 Victorian Bushfire factors knowledge data, the new case entry for the DisasterFactor concept in Preparedness-phase class (b) Recommendation 48 of ‘2009 Bushfire Royal Commission Final Report’ (Parliament of Victoria 2010), the new case solution for the StructuralMitigation concept (c) Recommendation 27 of ‘2009 Bushfire Royal Commission Final Report’ (Parliament of Victoria 2010), the new case solution for the Legislation concept Figure 14 Bushfire experiences knowledge derived from one knowledge source: 2009 Bushfire Royal Commission Final Report. The experiences knowledge are fragmented into three different DMM concepts: (a) DisasterFactor (b) StructuralMitigation (c) Legislation concept Is stored as an M0-Fragment of the StructuralMitigation  Rapid fire spread followed ignition, which responding crews could not contain.  Fires crowned in forested areas, which made them impossible for ground crews to control.  Powerful convection columns were generated above the fires.  Extensive forward spotting occurred as a result of the fuel type, the weather conditions and the topography.  Late in the day a wind change altered the direction of fire spread and extended the fire front. Is stored as an M0-Fragment DisasterFactor of Preparedness phase Is stored as an M0-Fragment of the Legislation Detailed DisasterFactor knowledge is structured into small unit fragments of M0-Fragments (in the DisasterFactor concept), as follows: The DMMClassAttribute of the GravityFactor: 1) Temperature: 28-30 Jan 2009 - consecutive hot days above 43 C (109 F), with peaking at 45C (113 F) on 30 Jan – 3rd hottest day in the city’s history. Hottest: 46.4 (115F). 2) Hot tropical air to South Eastern Australia is caused by a) Heatwave cause by a slow moving high-pressure system over the Tasman Sea; b) Intense tropical low located off the North West Australian coast; c) Monsoon trough over Northern Australia; and d) the weather conditions and the topography. Late in the day a wind change altered the direction of the fire spread and extended the fire-front. 3) Low humidity. The DMMClassAttribute of the TriggerFactor: 1) Power lines ignition - Rapid fire spread followed ignition, which responding crews could not contain, 2) Lightning, Cigarette butts, 3) Sparks from a power tool - Powerful convection columns were generated above the fires. The DMMClassAttribute of the ComplexityFactor: 1) Major drought more than a decade, 2) 50-year warming trend link to human induced climate change, 3) Extensive pyrocumulus clouds (most intense firestorm ever experienced in Australia history). Table 4 Storing knowledge of DisasterFactor of the 2009 Victorian Bushfire 4.2 Deriving new bushfire solution models from DMKR In this section, the functionality of DMKR as a DM modelling tool is presented. In DMKR based modelling, users are given flexibility to utilise the functionalities facilitated by DMM. In the rest of this section, the modelling capabilities of DMKR are illustrated in two illustrative bushfire scenarios. Scenario 1 presents how knowledge retrieved directly from the specific M0-Fragments can be used as the inputs to create an M1-Model of the Aerial Fire Fighting Strategy model. The models required in this scenario are developed bottom up. In Scenario 2, knowledge of an M1-Model initially created in M1-level is refined into M0-Fragments of two M0-User Models. The models required in this scenario are developed top-down. The 2009 Victorian bushfire factor knowledge is stored to the DisasterFactor concept through its M0-Fragment of the ‘GravityFactor’ Scenario 1: Supporting Aerial Fire Fighting Model process using DMKR (M0) knowledge In this scenario, the reuse of knowledge retrieved from various M0-Fragments to support the user to generate an M1- Model is shown. The user here aims to develop the Aerial Fire Fighting model. The development of this model requires other knowledge that needs to be studied before it can be constructed. For instance, knowledge of many aerial fire suppression factors is crucial for planning an efficient aerial bushfire fire fighting operation as described next. Scenario 1: “The use of aerial fire fighting resources has grown steadily over the past few decades and aircrafts now play an integral role in fire management and suppression in Victoria. In 1997-98 the then Department of Natural Resources and Environment operated the largest fleet of firebombing aircraft ever used in Australia including, for the first time, an Erickson Air crane, a large helicopter that is used for firebombing. An evaluation of the effectiveness of this fleet of aircraft was commissioned by Department of Sustainability and Environment (DSE) and published in 2003. A Bushfire Cooperative Research Centre (CRC) team tries to create a framework model of the effectiveness and efficiency of aerial fire-fighting in Australia. Key findings from this study should include: The effectiveness of aerial fire suppression depends on many factors, including suppressant agent used (retardant, foam or water), aircraft characteristics, drop characteristics, fire intensity, fire size, fuel type, pilot skill, aircraft travel time, distance from fire, ambient conditions, availability of support ground resources, and organisational and infrastructure arrangements.” (2009 Victorian Bushfires Royal Commission 2010) For this kind of problem, a study of different aerial fire suppression factors such as “suppressant agent used, aircraft characteristics, drop characteristics, fire intensity, fire size, fuel type, pilot skill, etc” are required. DMKR can support this by providing DM knowledge of the aerial fire suppression factors from its repository. These factors can be derived through the M0-Fragments of related DMM concepts. For example, through an M2-Fragment of the ResponseTask concept, the DMMClassAttribute of the concept, the ‘ResponseTaskFactor’ stores knowledge of various real-world responses task operation’s factors in the M0-Fragments of the concept. Particularly for the ‘suppressant agent used’ factor, knowledge earlier sourced from Emergency Management Australia (EMA) of can be reused. A sample of the relevant M0-Fragment in DMKR is as follows: ... If Fire-fighting OBJECTIVE = ("For common solid and combustible materials") Then USE = ("WATER") If Fire-fighting OBJECTIVE = ("For extinguishing fires that is flammable and combustible liquids") Then USE = ("FOAM") If Fire-fighting OBJECTIVE = ("Effective as a smothering and chemical chain reaction") Then USE = ("EXTINGUISHING POWDER") If Fire-fighting OBJECTIVE = ("For extinguishing small fires and fires involving electrical equipment") Then USE = ("CARBON DIOXIDE") If Fire-fighting OBJECTIVE = ("For extinguishing chemical chain reaction-inhibiting") Then USE = ("VAPORIZING LIQUID") If Fire-fighting OBJECTIVE = ("To react to form a smothering soap layer on the surface of burning fats") Then USE = ("WET CHEMICAL") If Fire-fighting OBJECTIVE = ("For extinguishing combustible chemical") Then USE = ("SPECIAL AGENTS") … In addition to the above, DMKR contains knowledge of similar processes (choosing a right fire fighting agent) implemented by other countries. This knowledge can support the CRC’s team in their study. In addition, knowledge other factors, for example, the ‘aircraft characteristics’ can also be retrieved for reuse. ‘ResponseTaskEquipment’ is another DMMClassAttribute of the ResponseTask concept. This knowledge is also sourced from practices of other countries and is retrieved straightforwardly as M0-Fragments. These include how: a) China uses “Metropolitan Aerial Fire Fighting System for the EC 225 Helicopter Simplex Model 516”; b) USA through its Los Angeles County Fire Department uses “Sikorsky S-70C Fire Hawk”; c) Italy through Italian Civil Protection Department uses “Bombardier Canadair CL 415 Aircrafts”; d) Ukraine uses the “Antonov 32 Aircraft”; and e) Germany uses the “KFOR helicopters etc (sources from Euro-Atlantic Disaster Response Coordination Centre (EADRCC) 2007)” in their aerial fire fighting operations. Efficiency of an aerial fire-fighting strategy also depends on fire behaviour analysis. Through an M2-Fragment of the Monitoring concept, the DMMClassOperation of the ‘EventAnalysis’ represents knowledge of how to monitor analyses disaster data. Specific for the purpose of the CRC team, the following M0-Fragment for the bushfire fire behaviour analysis is also reusable: --------------------------------------------------------------------------------------------------- ACTIVITY TASKs: EventAnalysis (M0-Fragment of the Monitoring concept: 1) Use current bushfire events data (Humidity Percentage, Air Temperature degree Celsius, Average Wind Speed, Slope Position, Fuels Loads, Fuel Type, Fuel Mixture, Steepness, Vegetation Communities) 2) Based on current events data, perform fire prediction analysis (Fire Prediction, Fire Forecasting, Fire Environmental Model, calculating Fire Spread Speed, estimating probability of Fire Ignition, Fire Evaluation based on weather, calculating Fire Danger Index) and perform Fire Behaviour Prediction Analysis. 3) Send event analysis results to Emergency Operation Centre. FIRE BEHAVIOR PREDICTION ANALYSIS: i) FIRE SPREAD SPEED: [ D = 1.15(5 + 0.5v) ] Description: This subroutine representing fire spread speed to leeward. Where v is a wind velocity meter per second (m/sec) and D is the limit of distance which fire can spread (m). ii) FIRE SPREADING PROBABILITY: [ Fij = a. (Sij . Pij). W Bij. P (tckl)] Description: The fire spreading judgement, that is whether or not cell ij changes from one state to another, is done by using the fire spreading probability of cell ij. The probability is given by the fire spreading judgement index Fij. This is calculated for all cells ij within the neighbourhood of cell kl and defined by this equation. iii) POINT OF FIRE ORIGIN: [ tx = (3 + 3a/8 + 8d/D) / (1+0.1v)] Description: This equation shows the time until an adjacent house to the point of fire origin ignites after catching fire in the origin. Where, a is a length of average of a side in building (m), d is a pitch of building (m) and D is the limit of distance which fire can spread (m). iv) CONTINUATION TIME OF FIRE: [ty = (w/5.5)/(Aw /'H / Af) Description: The time t2 until cell kl burns out after catching fire is defined in this subroutine. Where, w is a fire load (kg/m2), Aw is a opening area (m2), Af is a floor area (m2) and H is a height (m). This equation approximately represents the continuation time of fire as assuming wooden building and constant combustion speed v) PROCESS OF FIRE SPREADING: [ IF Nij(t) = 1 and Fij > ran then nij(t+1) = 2] Description: Fire spreading judgement index Fij of a selected cell ij is calculated in fire spread model. If Fij is satisfied with the following requirements, the set of cell ij becomes nij = 2 (Catching fire). This subroutine calculates nij. Where nij (t) is the state of cell ij at simulation time t and ran is the random number which takes from 0 to 1 vi) FIRE EXTINGUISHED UTILISATION: [ Ce = (Vw/(Cw.tf)). Rw] Description: The number of cells of which the fire can be extinguished with water utilisation, Ce, is given in this equation. This is based on the assumption that one nozzle continues spraying water to a cell of state until extinction of the cell …. Scenario 2: Creating Process Models (M1) and User Models (M0) for a new Bushfire Mitigation Plan process The derivation of an M1-Model using DMKR to setup a bushfire mitigation plan is presented in the following scenario: “The Head of Operation Engineer at a government-owned electric power company (SWEPCO) in Australia has been requested to prepare a bushfire mitigation plan for his company. SWEPCO is a company which has the responsibility for power sector operations such as dispatching electricity to different public local areas, generate electricity, operates power hydro stations and many more electrical operational works. The requirement is to develop a specific bushfire mitigation plan for this company as a result of many recent reported bushfire events in several of the company’s client’s locations. Therefore, a company preparedness strategy is required to overcome any unforeseen bushfire events in the future. As this will be the first time for this SWEPCO head engineer to prepare a plan that he has no prior experience of, this user definitely requires some support for preparing this task. The information of how to start the mitigation plan process, how to set up a strategic planning committee, which rules and legislations are relevant to bushfire problems, or what are the structural or non-mitigation appropriate for the plan are important for this user.” In this scenario, the SWEPCO’s engineer will be directed to the Mitigation-phase class represented in the Metamodel Formulation module. This is because the ‘Setup Mitigation Plan’ is recognised as a part of mitigation activities by the Consistency Checker Module. The engineer is then offered information on all the concepts (DMMClass in Mitigation-phase) from which he is able to derive his own DM model (M1-Model of Setup Mitigation Plan for SWEPCO Company). 21 such concepts exist. Class Diagram-view (Figure 15) is one M1-Model possibility which contains 11 concepts in the Mitigation-phase class selected for his model. This now becomes the M1-Fragments of the SWEPCO Setup Bushfire Mitigation Plan Model. Based on the M1-Model created by this user, the next step is to use DMKR to generate data for his M1-Model’s M1-Fragments. The collection of existing DM knowledge case solutions is stored in a repository that matches the inputs (bushfire knowledge) for the M1-Model entered by this user that will be retrieved. For this example, two such models are available in DMKR based on sources earlier collected from Powercor Australia Ltd (Powercor Australia Ltd 2011) and the United Energy Company (United Energy 2011). These provide bushfire mitigation knowledge specific for an electricity company which match the intention of the M1-Model (Mitigation plan model - Figure 15) as specified by the user. Generally for such users, electricity networks are a serious fire ignition hazard. Prevention is very important due to enormous cost associated with such disasters. For Powercor Company, the bushfire mitigation model includes processes for asset inspection, maintenance, construction, upgrading, replacement, vegetation management, performance monitoring and auditing processes. This strategy plan applies to assets that could cause fire ignition in areas designated as hazardous bushfire risk areas. For the United Energy Company, its bushfire mitigation plan model concentrates more on the policies and procedures used to demonstrate the commitment from all levels of management within the company to the minimisation of bushfire risk due to electricity assets. Figure 15 The Setup Mitigation Plan Model (M1) derived from the Mitigation-phase class The weather conditions for the two M0-User Models retrieved match SWEPCO’s operating conditions. The bushfire conditions in Australia faced by the two companies are equivalent to SWEPCO’s problem situation. Information derived from Powercor and United Energy provide relevant and significant bushfire knowledge for SWEPCO’s plans formulation. For instance, specific to the StructuralMitigation, based on the derived knowledge retrieved from the two sources, different kinds of structural mitigation processes are found to be conducted in their mitigation plans. The structural mitigation plans of Powercor include : a) The management of vegetation around power lines, b) Inspect private overhead electric lines, and c) The implemented new technologies to minimise risk of electricity assets causing fire ignition, whereas, for the United Energy the plan stresses: a) Electric lines clearance, b) Asset management system, and c) Procurement of equipment and services. From the retrieval of this knowledge, it is identified that the procedure used (data in the M0-Fragment) to check the technology and equipments standard in each plan is different. Powercor uses the ‘Technical Standard EA121 - Overhead - Bushfire Mitigation’ and United Energy uses the ‘Equipment Materials Approval Document No.JEN PR 0070 WI 04’. This information gives richness to the bushfire mitigation plans for the SWEPCO M1-Model. Hence, by using the M1-Model created in Figure 15, the real-world bushfire knowledge source of the two companies will be presented as the retrieval result for the SWEPCO’s new M1-Model. The two new M0-User Models are shown in Figures 16 and 17. Figure 16 The retrieval of model solution (M0) for the Powercor Mitigation Plan (M1). Figure 17 The retrieval of model solution (M0) for the United Energy Mitigation Plan (M1). The M0-User Model 1 source from Powercor Australia The M0-User Model 2 source from United Energy Even though sources from Powercor and United Energy are chosen as a solution for the SWEPCO case, the repository of DMKR also has a population of others similar model samples: the Berlin Wildfire Hazard Mitigation Plan from New Hampshire, USA (North Country Council 2007), the Fire Hazards Mitigation Plan from the Navi Mumbai Municipal, India (Navi Mumbai Municipal Corporation 2010) and many other fire mitigation solutions. In this section, the functionalities of the Metamodel-based DMKR system were demonstrated in real world DM scenarios. Specifically, real-world DM bushfire scenarios were visualised. Collecting and storing of bushfire knowledge experience from different bushfire knowledge sources showed the capability of DMKR for DM knowledge storage. The collected knowledge was later reused to solve other incoming bushfire problem scenarios faced by other users. This process demonstrated the potential of this system as a DM modelling tool. Moreover, derived solutions were also reusable and this was also demonstrated. Utilising DMKR, DM users in different bushfire scenarios were presented with new derived DM solutions suitable for their real world problems. 4. Conclusion and Future Works This paper illustrates a new a metamodelling based approach to guide DM practitioners to store and reuse DM based on a Disaster Management Metamodel (DMM). It presents a knowledge sharing system, DMKR 1.0, to uniformally store and structure DM knowledge. This is facilitated by a specific a DM language. DMKR facilitates the derivation of new DM solution models using the DM language, underpinned by the metamodel DMM. DMKR successfully combines the use of DMM for storage, retrieval and modelling to describe modelling capability for any specific disaster. To DM practitioners therefore, the paper offers a coordinated approach and activities to achieve DM goals where DMKR can act as a central tool for creating, organizing and managing DM modelling knowledge. Through fragmenting DM knowledge experience for reuse, DMKR guides DM users to achieve specific DM goal. Many stakeholders are either directly involved (e.g.: disaster manager, emergency coordinator, fire fighters..) or indirectly involved in DM operations (e.g.: local businesses, insurance companies etc..) but they do not have a complete view of how different DM activities can be conducted. DMM through its four sets of classes (response, recovery, mitigation and preparation classes) give an overall picture of how all DM actions are executed. With better awareness and understanding of the whole DM processes, many benefits can be derived because ensuring success in initial DM phase will lead to success in the subsequent phases. The paper also contributes to semi-formal knowledge management practices by developing a rigorous modelling hierarchy managed and enforced by DMKR and underpinned by DMM. As the ability to offer modelling guideline to many domain users, DMM helps various users to make quicker decisions using earlier created semantic models. It facilitates three levels of modeling as in the MOF framework: the M2-level for a Metamodel, the M1-level for a Model, and the M0-level for a User Model. Instance models are derived from DMM based on conformance relationships and DM knowledge is structured in the DMKR system using Model Fragment units. Since there are three MOF modelling levels are used in this research to describe the DM knowledge, three corresponding types of Model Fragment (M2-Fragment, M1-Fragmnet and M0-Fragment) are also used. Further future research will support knowledge evolution in DMKR and add further reasoning capacities to support DM first responders: The paper presents the construction of formal semantics of relating DM models at different metamodelling levels. This can form in the future a fundamental part of an automated decision support in many disaster scenarios. By populating knowledge of multiple disaster cases for the same unit fragment, a reasoning process can be easily applied. This would be of significance to DMM, as the use of the reasoning technique through metamodel for is not intensively being yet investigated. Furthermore, such automated reasoning capabilities can also pave the way for the construction of a real-time DM system using models created from the metamodel for a quick response. Since the DMM describes data that should exist in the Response-phase models, real-time calculation can be made when fresh disaster data that is being “pushed” from the field by first responders. This can for example include locations of responders, information on the status of response activities, data mapping damage, and sensor data. Results from the calculation of the data can be pushed back into the field immediately e.g. responder’s team send field reports, imagery and map updates and requests the current status information on resource deployments. References 2009 Victorian Bushfires Royal Commission (2010). Systemic Issues – Aerial Firefighting Submissions of Counsel Assisting Aßmann, U., S. Zschaler and G. Wagner (2006). Ontologies, Meta-models, and the Model-Driven Paradigm. Ontologies for Software Engineering and Software Technology, Springer Berlin Heidelberg: 249-273. Beydoun, G. and A. Hoffmann (1998). Simultaneous Modelling and Knowledge Acquisition using NRDR. 5th Pacific Rim Conference on Artificial Intelligence (PRICAI98), Singapore, Springer-Verlag. Brottier, E., F. Fleurey, J. Steel, B. Baudry and Y. L. Traon (2006). Metamodel-based Test Generation for Model Transformations: an Algorithm and a Tool. 17th International Symposium on Software Reliability Engineering (ISSRE 2006), Raleigh, North California, USA. Colin, A., K. Thomas and hne (2003). "Model-Driven Development: A Metamodeling Foundation." IEEE Softw. 20(5): 36-41. Cook, S. (2004) "Domain-Specific Modeling and Model Driven Architecture." MDA Journal: A BPT Column. Cordner, S. M., N. Woodford and R. Bassed (2011). "Forensic aspects of the 2009 Victorian Bushfires Disaster." Forensic Science International 205(1-3): 2-7. e-objects Open Source MetaModel. (2012). "MetaModel database compliancy in Open Source MetaModel." Retrieved 20 March 2012. Emergency Management Australia (EMA) (2000). Disaster Preparedness Manual for Commonwealth Agencies. Australian Emergency Manual Series, National Archives of Australia. Euro-Atlantic Disaster Response Coordination Centre (EADRCC) (2007). Forest Fires In Albania, Bulgaria And The Former Yugoslav Republic Of Macedonia. EADRCC Executive Summary. Firesmith, D. (2006). Method Engineering Using OPFRO. 11th Annual European Software Engineering Process Group Conference (EuroSEPG). Johnston, S. (2009). "Independent Community Representatives Advocate." The Australian Retrieved 1-3-2009, from http://www.johnston-independent.com/. Kleijnen, J. P. C. and R. G. Sargent (2000). "A Methodology for Fitting and Validating Metamodels in Simulation." European Journal of Operational Research 120: 14-29. Lammel, R. (2001). Grammar Adaptation. Proceedings of the International Symposium of Formal Methods Europe on Formal Methods for Increasing Software Productivity (FME'01), Springer-Verlag: 550-570. Larcombe, P. J., S. E. Moloney and P. A. Schmidt (2009). Pandemic (H1N1) 2009: a clinical spectrum in the general paediatric population. Archieves of Disease in Childhood. Queensland, Gold Coast Hospital. 96: 96-98. Lauras, M., Truptil S., and Bénaben, F. (2015) Towards a better management of complex emergencies through crisis management meta-modelling, Disasters 29 (4), 687–714 Levendovszky, T., B. Rumpe, B. Schatz and J. Sprinkle (2010). Model Evolution and Management. MBEERTS. Berlin Heidelberg, Springer-Verlag. 6100: 241-270. M. Picka (2004). "Metamodelling and Development of Information System." Agriculture Economics 2(50): 65-70. Machin, B. and M. Hentze (2007). Comments on The Present and Future of Wildland Fire Suppression Decision- Making Processes. Living on the Edge: Economic, Institutional and Management Perspectives on Wildfire Hazard in the Urban Interface, Advances in the Economics of Environmental Resource. Elsevier Ltd. 6: 1-12. Navi Mumbai Municipal Corporation (2010). Fire Hazards Response And Mitigation Plan Mumbai, India. North Country Council (2007). Berlin Wildfire Hazard Mitigation Plan. Bethlehem, New Hampshire, USA. OMG (2002). Meta Object Facility (MOF) Specification, Object Management Group. OMG (2003). MDA Guide Version 1.0.1. http://www.johnston-independent.com/ http://onlinelibrary.wiley.com/doi/10.1111/disa.12122/abstract http://onlinelibrary.wiley.com/doi/10.1111/disa.12122/abstract Othman, S. H., G. Beydoun (2013). Model-driven Disaster Management. Information and Management. 50(5), 218-228. Othman, S. H., G. Beydoun and V. Sugumaran (2014). "Development and validation of a Disaster Management Metamodel (DMM)." Information Processing & Management 50(2): 235-271. Panesar-Walawege, R. K., T. S. Knutsen, M. Sabetzadeh and L. Briand (2011). CRESCO: construction of evidence repositories for managing standards compliance. Proceedings of the 30th international conference on Advances in conceptual modeling: recent developments and new directions. Brussels, Belgium, Springer-Verlag: 338-342. Parliament of Victoria (2010). 2009 Victorian Bushfires Royal Commission. Victoria. I,II,II and IV: 41. Paterson, R. G. (2007). Wildfire Hazard Mitigation As Safe Smart Growth. Living on the Edge: Economic, Institutional and Management Perspectives on Wildfire Hazard in the Urban Interface, Advances in the Economics of Environmental Resource. Elsevier Ltd. 6: 1-12. Powercor Australia Ltd (2011). Bushfire Mitigation Strategy Plan 2011-2012: Electricity Safety (Bushfire Mitigation) Regulations 2003, Document No: 05-M810. F. Audley: 68. Saltor, F., M. García-Solaco and P. Gargallo (1993). Diversity with cooperation in database schemata: Semantic relativism. In Proc. ICIS Conference. Shanzhen, Y. and L. Yuntao (2011). An intelligent information sharing method of heterogeneous geographic data based on unified metamodel and entity matching. Geoinformatics, 2011 19th International Conference on. Shiva, S. G. and L. A. Shala (2007). Software Reuse: Research and Practice. Information Technology, 2007. ITNG '07. Fourth International Conference on. Stahl, T., M. Voelter and K. Czarnecki (2005). Model-Driven Software Engineering, Technology, Engineering, Management, John-Wiley & Sons Ltd. Stephens, S. L. and B. M. Collins (2007). Fire Policy In the Urban Wildland Interface (UWI) In the United States: What Are The Issues and Possible Solutions? Living on the Edge: Economic, Institutional and Management Perspectives on Wildfire Hazard in the Urban Interface, Advances in the Economics of Environmental Resource. Elsevier Ltd. 6: 1-12. Trabelsi, C., R. B. Atitallah, S. Meftali, J.-L. Dekeyser and A. Jemai (2011). "A Model-Driven Approach for Hybrid Power Estimation in Embedded Systems Design." EURASIP Journal on Embedded Systems (Article ID 569031): 15 Tran, N., G. Beydoun and G. Low (2007). Design of a Peer-to-Peer Information Sharing MAS Using MOBMAS (Ontology-Centric Agent Oriented Methodology). In Advances in Information Systems Development, W. Wojtkowski, W.G. Wojtkowski, J. Zupancic, G. Magyar and G. Knapp (eds.). Springer US, 63-76. Troy, A. and R. G. Kennedy (2007). Finding Solutions To The Urban Wildland Fire Problem In a Changing World. Living on the Edge: Economic, Institutional and Management Perspectives on Wildfire Hazard in the Urban Interface, Advances in the Economics of Environmental Resource. Elsevier Ltd. 6: 1-12. United Energy (2011). Bushfire Mitigation Plan 2011-2012, Document No: UE PL 0009. T. Fisher: 130. Vara, J. M., B. Vela, V. A. Bollati and E. Marcos (2009). Supporting Model-Driven Development of Object - Relational Database Schemas: A Case Study. Proceedings of the 2nd International Conference on Theory and Practice of Model Transformations. Zurich, Switzerland, Springer-Verlag: 181-196. Viswanathan, G. and M. Schneider (2010). BigCube: A Metamodel for Managing Multidimensional Data. International Conference on Software Engineering and Data Engineering (SEDE 2010). I. Rahal and R. Zalila- Wenkstern. USA, ISCA: 237-242. Vytautas Stuikys, R. D., Aleksandras Targamadze (2010). "A Model-Driven View To Meta-Program Development Process." Information Technology and Control 39(2). Wachsmuth, G. (2007). Metamodel Adaptation and Model Co-adaptation. Proceedings of 21st European Conference on Object-Oriented Programming (ECOOP), Berlin, Germany, Springer-Verlag. Xu, D., C. Wijesooriya, X.-G. Wang, G. Beydoun (2011), Outbound logistics exception monitoring: A multi- perspective ontologies’ approach with intelligent agents, Expert Systems with Applications, 38 (11), pp. 13604– 13611. 
work_pjmz74mdjbc4thg76nqncha7l4	----	Problems in applying expert system technology to Radiographic Image Interpretation Problems in Applying Expert System Technology to Radiographic Image Interpretation David W. Piraino, Bradford J. Richmond, Masataka Uetani, Thomas Luetkehaus, Daniel Rockey, George Belhobek, Joe Armistead, and Fred Jones A prototype expert system was developed to study the problems applying expert system technology to radiographic image interpretation. The Radiographic Image Interpretation System (RIIS) was developed on a microcomputer using Turbo Prolog, a low cost implementation of the prolog programming lan- guage. The present implementation of RIIS was developed to highlight potential problems in applying expert system technology in the evaluation of radio- graphic images. It was believed that the evaluation of this prototype expert system should include a large number of users unfamiliar with the program's use as this would probably be the case in clinical use of an image interpretation expert system. At present, the expert system deals with a limited domain of focal bony lesions. Twenty cases of pathologically proven bony lesions of varying difficulty were used to evaluate potential problems in the use of this expert system technology. RIIS, with the 20 sample cases, was presented as an exhibit at the 1987 Radiological Society of North America (RSNA) meeting to evalu- ate the potential problems with inexperienced users. These results were compared with those of experi- enced users. When a musculoskeletal radiologist, familiar with the programs use, provided the "proper description," the program averaged the correct diag- nosis in the top five 80% of the time. During the program's use at the RSNA meeting, the program selected the correct diagnosis in the top five 22% of the time. @) 1989 by W.B. Saunders Company KEY WORDS: Diagnosis, computer, expert system, bone tumor, image. EXPERT system technology, an outgrowth ofartificial intelligence research, has been applied in a wide variety of medical situations including radiographic differential diagnosis, hematologic disorders, evaluation of anemia, chemotherapy protocols, and the diagnosis of medical disorders.' Difficulties related to expert system application to radiographic image inter- pretation must address all the standard problems in applying expert system technology in any environment. In addition, radiographic differen- tial diagnosis or image interpretation systems must address the problem of an appropriate method to input image information into the expert system. Evaluation of an expert system is an important Journal of Digita/lmaging, Vol 2. No 1 (February). 1989: pp 21-26 part of its development. Evaluation of expert systems, however, remain difficult and a stan- dard evaluation process has not been developed. In general, a system may be evaluated on its accuracy, method of construction, and user impact. A system's accuracy can be judged in relation- ship to a standard such as pathology. Since an expert system uses information and knowledge used by experts to arrive at a conclusion, its relative accuracy should be judged against and by other experts in its area of expertise.' The techniques used to develop the expert system influence its user interface, methods of deduction, potential conclusions, and explana- tion of its conclusions. If a specific knowledge structure or user interface is chosen that is not appropriate for a expert system's environment, it will not be useful. The goal of expert system technology is to produce a system that can be used by non-experts to help them arrive at conclusions at a level comparable to experts. It is therefore important, when evaluating expert systems, to test the user interface for its ease of use and user acceptance.' It is also necessary to evaluate whether the expert system helps the non-expert to make decisions at an accuracy level comparable with that of an expert. Finally, the non-expert must be confident that the decision he or she reaches with the help of an expert system is appropriate. RIIS was developed on a personal computer to explore problems in implementing a prototype expert system for radiographic image interpreta- tion. The study of the problems in implementing this prototype were directed primarily toward expert system construction, user interface, and From the Cleveland Clinic Foundation, Tripier Army Medical Center, Honolulu and Henry Ford Hospital, Detroit. Address reprint requests to David Piraino. MD, Depart- ment of Radiology. The Cleveland Clinic Foundation, 9500 Euclid Ave, Cleveland, Ohio 44195-5021. © 1989 by W.H. Saunders Company. 0897-1889/89/0201-0004$03.00/0 21 22 accuracy. The investigation evaluated users familiar and unfamiliar with the program. RADIOGRAPHIC IMAGE INTERPRETATION SYSTEM The Radiographic Image Interpretation Sys- tem (RIIS) was developed to produce a differen- tial diagnosis for focal musculoskeletal lesions. Focal bone lesions were chosen because this domain can be defined relatively easily and because several of the authors are musculoskele- tal radiologists. Turbo Prolog, a fourth genera- tion language for microcomputers, was used to construct the program. The program uses rela- tive likelihoods and relative predictive values to produce a list of differential diagnostic possibili- ties. Several psychological models have been devel- oped to explain the process of image interpreta- tion. Differences in psychological models deal primarily with whether there is a preliminary expectation of the observer that affects the pro- cessing of the visual information before the actual inputting of the image information (Table 1).4 The RIIS prototype expert system only deals with the last transformation of the aforemen- tioned model from a conceptional understanding of the image to a decision about the most appro- priate diagnosis related to that image. This deci- sion to start with a language description of the image abnormalities places constraints on the expert system and imposes potential problems concerning image information input. The user of RIIS selects the positive radio- graphic findings from a list of possibilities pre- sented to the user by the program. Typical find- ings included in the initial implementation include bony matrix, chondroid matrix, bony expansion, geographic lesion, permeative pat- tern, and many other findings. The program considers any findings that are not selective as Table 1. Example of an Image Interpretation Psychological Model I. Input: Various light levels projected on retina II. Preprocess: Retina preprocesses information III. Segmentation: Grouping of input information IV. Understanding: Image groups recognized V. Decision: Decision on image diagnosis* *Data adopted from Kundel et al.4 PIRAINO ET AL not being present in the image. An information base relating each diagnostic entity to the radio- graphic findings is contained within the program. All findings associated with any specific diag- nosis such as osteogenic sarcoma are given a relative frequency from I to 5 and a relative predictive value from 1 to 5. A relative frequency of 1 indicates a very unusual finding associated with that disease while a 5 represents a finding that is always associated with that disease. A predictive value of 1 represents a completely nonspecific finding while a value of 5 represents a pathognomonic finding." Relative frequencies and relative predictive values were assigned by the consensus of two musculoskeletal radiologists after reviewing standard radiology textbooks. Each diagnosis in the information base is compared with the findings input by the user. In the first level of evaluation, each diagnosis is compared with the positive image findings. Any positive finding that a specific diagnosis can explain, raises the relative likelihood of that diagnosis. After all diseases have been evaluated using this technique, a second evaluation takes place. In the second evaluation process, the find- ings that are commonly seen in a specific disease entity but were not described as being present on the image, decrease the overall likelihood of that specific diagnosis. Finally, a correction is made for findings described as being present in the image, but do not occur in the specific diagnosis being considered. In this case, since the diagnosis does not explain all the findings on the image, the overall likelihood of that specific diagnosis is decreased. RIIS uses a simple inference engine to accom- plish the above results. A simple consecutive search is performed on the information base comparing the radiographic description of the specific image with the radiographic descriptions of each disease. The relative likelihoods of each diagnosis is then calculated. RIIS arranges the diagnoses in order of most likely first, least likely last, according to their calculated relative likelihoods. The relative like- lihoods are compared and different numeric selection criteria are applied to the relative likeli- hood for appropriate diagnosis selection. Dif- ferent selection criteria include: (1) considering only those whose relative value is >50% of the relative likelihood associated with the most likely PROBLEMS IN APPLYING EXPERT SYSTEMS TO X-RAYS diagnosis, and (2) considering only those diag- noses above any point in which there is a 50% difference in the relative likelihood of any two adjacent diagnoses. PROTOTYPE EVALUATION RIIS was evaluated with experienced and non- experienced users. Twenty cases of focal bony abnormalities in print form were used to evaluate the program. The cases were selected to repre- sent a range of diagnoses and a range of difficul- ties from classic cases to atypical cases. The cases were selected from the Cleveland Clinic teaching file and from clinical practice. The program along with the 20 cases in print form were presented as an exhibit at the 1987 RSNA meeting in Chicago. People viewing the exhibit were instructed how to use the RIIS system in print form and were encouraged to use the 20 sample cases as examples. A monitoring program was produced to retain statistics throughout the meeting to determine how well the users and the program performed for each case. The users entered their experience level, selected what they considered to be the appropri- ate differential diagnosis from a list of diagnostic possibilities, and entered the positive findings they identified on the radiograph. The program then produced its differential diagnosis, with associated relative likelihoods, which the user could compare with their differential diagnosis. Statistics were maintained for the program's use throughout the week as well as for each experi- encelevel. Subsequently, a musculoskeletal radiologist involved in the development of the program and familiar with its syntax and use developed a "proper description" of the abnormalities on each of the 20 cases. While the musculoskeletal radiologist knew the pathologic diagnosis, the proper descriptions were developed indepen- dently of the program. All the descriptions were reviewed before input into the program and any findings that were considered questionable were removed to avoid bias. The proper descriptions were entered into the program and its perfor- mance was then evaluated. The differential diagnosis included only those diagnoses above a cutoff point where the first 50% difference between any two adjacent diag- 23 noses was found in the differential diagnostic list. This same selection was applied both at the 1987 RSNA meeting and when the proper description was provided. RESULTS The results of the program's use at the RSNA meeting, the program's results with a proper description, and the results for RSNA users are presented graphically in Fig 1. At the RSNA meeting, the program was used 268 times by people with varying experience levels. The cor- rect diagnosis was selected in the program's differential diagnostic list in 22% of attempts. The correct diagnosis was selected first in 33 cases (12%); second in 16 cases (6%); third in six cases (2%); fourth in two cases (0.7%); and fifth in two cases (0.7%). The correct diagnosis was not listed in 209 attempts (78%). When the proper description was entered by our musculoskeletal radiologist, the program 90 80 70···· J 60 t 50 .c IV 0\ 4D.s j 30 20···· 10···· o CCMP-CffiR B/J-SPEC FEllOW RES-3f\D RES-1ST CCMP-RSNA GEN RES-41l-1 RES-2ND AVE Experience Leve Fig 1. This figure graphically demonstrates the per- centage by experience level in which the pathologic diag- nosis was selected in the top five diagnoses. RIIS with a proper description (COMP-CORR) selected the pathologic diagnosis in the top five 80% of the time and RIIS at the 1987 RSNA (COMP-RSNA) selected the pathologic diag- nosis 22% of the time. This compares with musculoskeletal specialists (B/J-SPEC) at RSNA who selected the correct diagnosis 71 % of the time. The remaining bars demonstrate the percentage of attempts in which the diagnosis was listed in the top five for the remaining experience levels including general radiologists (GEN), fellows (FELLOW), fourth-year residents (RES-4TH), third-year residents (RES-3RD), second-year residents (RES-2ND), first-year residents (RES-1ST), and the average (AVE) for all experi- ence levels. 24 selected the correct diagnosis in 16 of 20 cases (80%). In 11 of 20 (55%) of these cases, the correct diagnosis was listed as the most likely diagnosis and in five cases (25%), the correct diagnosis was listed as the second most likely. By comparison, musculoskeletal radiologists at the 1987 RSNA meeting included the proper diagnosis in their differential diagnostic list in 34 of 48 attempts (71 %). General radiologists included the correct diagnosis in their differen- tial diagnostic list in 58 of 109 attempts (47%). Statistics for individual cases at RSNA varied widely for the accuracy of both the program and the users. The best performance for the program at RSNA occurred in case no. 3, a chondrosarco- ma, where the program listed the proper diag- nosis as its most likely possibility in over 70% of attempts. This compares with musculoskeletal radiologists who listed chondrosarcoma 100% of the time as the most likely diagnosis but general radiologists listed chondrosarcoma only 30% of the time. The worse performance for the pro- gram at RSNA was case no 6 (Fig 2), a lytic osteogenic sarcoma. The correct diagnosis was never listed in the program's differential diag- nostic list in 15 attempts. In the 15 attempts at RSNA, only one participant listed lytic osteo- genic sarcoma in their differential diagnosis. When the proper descriptions were entered, the program did not include the proper diagnosis in four cases. The first case was again the lytic osteogenic sarcoma. The program did not list a lytic osteogenic sarcoma as a possibility although it included osteogenic sarcoma, chondrosarcoma, metastatic disease, and Ewing's sarcoma in its differential diagnostic list. Another case in which the correct diagnosis was not included by the program when a proper description was input was a case of a parosteal osteogenic sarcoma involving a finger. The program listed a single differential diagnostic choice, a juxtacortical chondroma. DISCUSSION The RIIS program was able to correctly diag- nose the lesions in the 20 sample cases in 22% of attempts with inexperienced users. This accu- racy level is significantly less than that of muscu- loskeletal experts. However, the system's accu- racy improves remarkably with a user familiar with the program and its associated syntax and PIRAINO ET AL Fig 2. Representation of case no 6, a lytic osteogenic sarcoma in the proximal fibula. The correct diagnosis was listed only once out of 16 attempts during the RSNA meeting. RIIS never listed the diagnosis during the RSNA meeting. When the proper description was provided, RIIS's differential diagnosis was osteogenic sarcoma, chondrosar- coma, metastatic disease, and Ewing's sarcoma. However, the correct diagnosis of lytic osteogenic sarcoma that is considered separately by the RIIS system was not included. language. In fact, with an experienced user, RIIS performed at a level comparable with musculo- skeletal radiologists. Also, the results from the RSNA meeting should not be strictly interpreted because (1) several of the users may have input descriptions slightly different from the sample cases to observe how the program performed in these instances and (2) 20 cases are a limited sample for evaluating the performance of an expert system. The RIIS did perform worse on the average than the general radiology user at our 1987 PROBLEMS IN APPLYING EXPERT SYSTEMS TO X-RAYS RSNA meeting and much worse than a muscu- loskeletal specialist. There was marked improve- ment when the proper descriptions were provided by an experienced user. While standard descrip- tions for focal bony lesions were used by the RIIS program, it appears that these descriptions were interpreted differently by different radiologists. For example, in a case of chondrosarcoma with chondroid matrix, several users described the matrix as groundglass or bone. In these instances, this description of the matrix is consid- ered incorrect and this incorrect description would adversely affect the program's diagnostic list. The proper description in such a case should include chondroid matrix. While all incorrect descriptions may not be as obvious as the above example, such descriptions lead the program further away from the correct diagnosis. The possibility of providing an incorrect description to the system raises the question of whether providing a description of x-ray abnor- malities is a legitimate method for inputting abnormal findings on an image or radiograph. Other possibilities for inputting image informa- tion include direct input of the image, using more descriptive terminology to describe the findings, or using graphic drawings of possible findings. Presently, direct input of the radiographic image is a relatively simple technical task with a state of the art video digitizer. However, the process of abstracting anatomic structures and anatomic abnormalities is an extremely difficult task that is only beginning to be investigated.' Therefore, the direct input of radiographic images would not be practical on a microcomputer. Possible interim solutions would include dia- gramatic graphic representations of x-ray abnor- malities and a more descriptive input environ- ment. Simplified graphic representations of radiographic abnormalities might help eliminate potential discrepancies between use of standard radiographic descriptions. The user could simply select the graphic representation of specific ab- normalities that could then be assembled by the system to produce a schematic drawing of the abnormality to provide user feedback. The graphic descriptions could then be assembled by the program for symbolic processing of either matched descriptive information or direct pro- cessing of the schematic diagrams. A second and perhaps simpler solution relates 25 to a more simplified descriptive input environ- ment. In this type of environment, the program itself makes interim conclusions about the pres- ence or absence of a finding from simplified descriptions. For example, instead of inquiring whether a bony matrix is present within a lesion, the system questions the user as to whether the abnormality demonstrates increased, decreased, or mixed density abnormalities. Subsequently, if the user selects increased density, the program questions more specifically about the location of the area of increased density and asks for a description of the characteristics of the increased density. The program then uses the more descrip- tive terminology to arrive at a conclusion about the presence or absence of a bony matrix. Another major problem area involves the inference methodology or program logic. The inference engine in this system is simplistic in nature. The complete sequential match and score technique works relatively well in this small limited domain system as shown by the very good results observed with the proper descriptions. However, in larger more complex domains, this technique becomes computationally intensive and does not lend itself well to explanation of the conclusions made by the system. A second important problem area in the pro- gram's logic is its selection criteria. A cutoff at the first 50% relative likelihood difference may not be appropriate, especially when the relative likelihood of all selected lesions are relatively low. This selection criteria is certainly simplistic in nature and may be inadequate for this type of decision. This problem relates to the expert sys- tem prototype construction and highlights how a specific expert system technique or programming technique can affect the accuracy of perfor- mance of an expert system. An important evaluation criteria of an expert system is to determine that its information or knowledge base is factually and conceptually accurate. This task is quite difficult in the medi- cal domain, especially with relative predictive values and frequency measures. Inaccuracies in determining relative frequencies or relative pre- dictive values would be expected to significantly change the accuracy of the program. It is diffi- cult to determine the relative frequency and predictive value units from standard radio- graphic textbooks. The inability to confirm the 26 accuracy of the knowledge base is a significant drawback to this type of knowledge or informa- tion structure. User acceptibility and user confidence in the RIIS system was not evaluated. User "believabil- ity" in the system is an extremely important evaluation criteria especially if such expert sys- tems are to be used in a clinical setting. It is therefore important to structure further evalua- tions of expert systems to evaluate user accep- tance and to determine if the expert system helps non-experts to perform as experts. Furthermore, it is probably more important to provide the user PIRAINO ET AL with a good explanation as to why certain dis- eases were selected rather than simply presenting the user with a differential diagnostic list. There are several systems at present that provide such an explanation as a critiquing facility in their use.' Finally, in order for expert systems to be used in a clinical setting, they must be widely avail- able, relatively inexpensive, and user friendly. While there are many problems remaining to be solved before expert systems are clinically accepted, there is much potential in specific limited domain areas. REFERENCES 1. Bar A, Feigenbaum EA: The Handbook of Artificial Intelligence (vol 3). Reading, Addison-Wesley, 1982 2. Quaglin S, Stefanelli M, Barosi G, et al: A performance evaluation of the expert system ANEMIA. Comput Biomed Res 21:307-323,1988 3. Hudson DL, Cohen ME: The role of user-interface in a medical expert system, in Ackerman MJ, (ed): Proceedings Symposium on Computer Applications in Medical Care (SCAMC), Washington, DC, Institute of Electrical and Electronics Engineers, (IEEE) Computer Society, 1985, pp 232-236 4. Kundel HL, Nodine CF, Doi K: Human interpretation of displayed images, in Hendee WR, Wells PN (eds): Engi- neering Research in Visual Perception. Chicago, American College of Radiology, 1986 5. Vries JK, Banks G, Mclinden S, et al: Three- dimensional neuro-imaging using octree encoding, in Acker- man MJ (ed): Proceedings SCAMC 1955,697 6. Miller RA, Pople HE, Myer JD: Internist-I, an experi- mental computer-based diagnostic consultant for general internal medicine. N Engl J Med 307:466-476,1982 7. Swett HA, Miller PL: ICON: A computer-based approach to differential diagnosis in radiology. Radiology 163:555-558,1987 
work_pjoeu5ebfzhrlhey3rpdnpyzkq	----	137816 529..550 Expert systems and mass appraisal John Kilpatrick Greenfield Advisors LLC, Seattle, Washington, USA Abstract Purpose – The purpose of this paper is to examine the usefulness of a heuristic expert system, to show its applicability to real-world valuation problems, and to suggest several avenues for statistical testing. Design/methodology/approach – The expert systems follow a traditional sales adjustment grid format, with sufficient data for non-parametric testing. Findings – The paper finds that, while non-parametric statistics provide weaker results than traditional (e.g. hedonic regression) modeling, the technique provides a statistically testable model useful in situations with limited data and/or poorly characterized probability functions. Practical implications – This paper addresses the conundrum faced by real estate valuers on the lack of statistical underpinnings of traditional heuristic models. Originality/value – This is one of the first empirical studies in the valuation literature exploring statistical characterization of heuristic valuation methods. Keywords Expert system, Property values, Sales adjustment grid, Non-parametric statistics, Property, Fair value, Sales management, United States of America Paper type Research paper 1. Introduction The purpose of this paper is to explore what is known about expert systems, a set of methodologies within the category of real estate appraisal sales comparison approaches which combines the heuristic characteristics of the sales adjustment grid with some of the statistical power of regression modeling. Expert systems are useful when the appraiser is confronted with small data sets or the likelihood of non-normality in values, but nonetheless has a sufficiently large array of data to at least take advantage of some non-parametric statistical characterizations. Regression relies on the appraiser’s judgment in the modeling process, but lets the data essentially speak for themselves in the adjustment phase. Expert systems rely on appraiser judgment at both levels; but by applying a larger set of data than can be comfortably managed with a sales grid, this method allows for a heightened degree of accuracy, reliability, and replicability in the process. Expert systems draw from Bayesian estimation, and constitute a maximum likelihood estimator of value, which results in the same coefficients as the least squares estimator derived from a hedonic model, but approaching the problem from a different perspective Real estate occupies a unique place on the asset spectrum. The real estate market is notoriously inefficient, and unlike securities markets, which provide severe penalties for taking advantage of certain kinds of inefficiencies (e.g. insider trading), real estate markets actually foster and encourage participants to in this regard. Transactions require substantial intermediation, high degrees of leverage, and lengthy clearing periods. The assets themselves suffer from locational monopoly and high degrees of The current issue and full text archive of this journal is available at www.emeraldinsight.com/1463-578X.htm Expert systems and mass appraisal 529 Received December 2010 Accepted March 2011 Journal of Property Investment & Finance Vol. 29 No. 4/5, 2011 pp. 529-550 q Emerald Group Publishing Limited 1463-578X DOI 10.1108/14635781111150385 both temporal and spatial autocorrelation. (See, for example, Pace et al., 1998, and Des Rosiers et al., 2000) As such, raw asset prices themselves reveal very little about the true value of real estate; yet an understanding of the actual underlying value is critical for a number of reasons, including business decisions (particularly financing), forced acquisition litigation (either through eminent domain or through trespass, such as encroachment or contamination), and taxation (e.g. – property, estate). This has given rise to a rather stylized appraisal process. In the USA, and in most other countries, appraisal methods have developed heuristically over the years[1]. We can trace appraisal methods and standards in the USA back to the Virginia House of Burgesses in the 1600s, at which time they gave instructions on the assessment of property for tax purposes. Professional appraisal organizations arose in the USA in the 1930s, amid a clamor for better organization of financial markets in general. Various professional appraisal organizations came together in the 1980s to codify the Uniform Standards of Professional Appraisal Practice (USPAP), which were then transferred to the newly-formed Appraisal Foundation, which was empowered via the 1989 Financial Institutions Reform, Recovery, and Enforcement Act to promulgate both appraisal standards and qualifications for state-based licensure. Appraisal methods – as distinct from appraisal standards – continued to be developed heuristically. The “body of knowledge” evolved on basically a two-track system, with academic researchers exploring values via statistically robust methods, such as regression, contingent valuation, or time-series modeling, and practitioners relying more heavily on professional organizations for methodological guidance. Of course, in practice, this sort of bi-modality was not so clearly defined, as many practitioners also held academic appointments, and many academics contributed to practitioner texts and coursework. Nonetheless, for good or bad, the appraisal profession did not follow the same path carved out by accountants, who have a more well-developed integration of academia and accounting practitioners. In recent years, two separate evolutions have given the profession some pause. First, arguably, the three largest uses of appraisals in the USA are for property tax purposes, mortgage financing, and eminent domain “takings”. The first of these is more-or-less governed by supplemental standards and methods promulgated by the International Association of Assessing Officers (IAAO). While the IAAO is nominally a part of the USPAP universe, tax assessors usually adhere to a mass appraisal paradigm (provided for by USPAP Standard 6), which, in its best applications, closely resembles hedonic regression. They generally are required to adhere to a certain degree of statistical rigor, and the IAAO actually promulgates minimum confidence levels, such as maximum acceptable coefficients of dispersion. Second, mortgage finance appraisal, however, has no such published thresholds, and that segment of the profession has been roiled with accusations of inaccuracies, inarguably contributing to the overall problems with residential mortgage finance today. A thorough examination of these problems is beyond the scope of this study, but it is sufficient to say that as of this writing, that segment of the appraisal profession is currently casting about for a better way to do things. To follow-up on Shiller and Weiss (1999), mortgage lending has apparently recently erred on the side of minimizing Type I errors (failure to make a deserved loan) but at the expense of increased Type II errors (making loans that should JPIF 29,4/5 530 not have been made). At either extreme of the bell curve (assuming it IS a bell curve), there is presently a lack of statistical power to assess the likelihood of either type of error. Third, eminent domain is also one of the largest purposes for which appraisal is currently conducted. In most circumstances, eminent domain appraisal follows essentially the same methodological paradigm as financing appraisal. In 1993, the USA Supreme Court issued the now-famous Daubert ruling, in which it set down certain criteria for the admission of expert scientific testimony[2]. It was not actually until 1999 that these criteria were extended to technical experts, including appraisers, via the Kumho Tire ruling[3], but nonetheless the criteria as applied to real estate valuation continues to be referred to as the Daubert test. Many states amended their rules of evidence to conform to Daubert, and as of this writing, most states use Daubert for civil litigation[4]. The emphasis of Daubert is hypothesis testing. In its opinion, the Court stated: Ordinarily, a key question to be answered in determining whether a theory or technique is scientific knowledge that will assist the trier of fact will be whether it can be (and has been) tested. Scientific methodology today is based on generating hypotheses and testing them to see if they can be falsified; indeed, this methodology is what distinguishes science from other fields of human inquiry(emphasis added). Academic researchers will immediately recognize that this places a requirement for statistical characterization of the data, analysis, and results, but the Court went further to establish a four-pronged test to provide guidance for trial courts[5]: (1) Do(es) the method(s) center upon a testable hypothesis? (2) Is there a known or potentially knowable error rate associated with the method(s)? (3) Has the method been subject to peer review? (4) Is the method generally accepted in the relevant scientific community? Clearly, valuation methodology used in the courtroom must carry with it some measure of statistical validity in order to meet these criteria. Thus, all three of the leading uses of appraisals in America now have significant reason to explore statistical characterization. One solution to this problem would be the use of hedonic regression modeling for all real estate valuation. OLS regression has been an important component of the econometric tool box since Gauss and Markov developed their eponymous theorem[6]. The appraisal body of knowledge has discussed the applicability of regression analysis for many years, and Smith (1971) provides a good summary of the early thinking within the profession. While a more widespread use of regression may be appealing, and in fact may be feasible in many areas and for many property types, at the very least it is overkill and at worst the requirements for linear regression make it infeasible in many real estate valuation situations. In the former, hedonic modeling requires a substantial level of data gathering. One might envision a situation such as exists in Germany, where local boards keep track of property values on a mass-basis, but even those situations require an appraiser to handle individual property characteristics. Worse, in a mass appraisal Expert systems and mass appraisal 531 scenario, practical application of ordinary least squares (OLS) regression requires that several assumptions should be satisfied: (1) Data are independently and identically distributed (iid) draws from their joint distribution. (2) Strict exogeneity, that is, the conditional means of the error terms is zero, and the errors are uncorrelated with the regressors: E 1jX � � ¼ 0 ð1Þ and E X 01 � � ¼ 0: ð2Þ (3) No multicollinearity, which requires that all of the regressors are linearly independent, and the matrix Q is positive and semi-definite, with moments up to at least the second: Pr rank Xð Þ ¼ p � � ¼ 1 ð3Þ and Qxx ¼ E X 0X=n � � : ð4Þ (4) Spherical errors, where I is an n £ n identity matrix: Var 1jX � � ¼ s 2In: ð5Þ (5) The most common violations of spherical errors are heteroskedasticity and autocorrelation: E 12i jX � � – s 2 ð6Þ and E 1i1jjX � � – 0; for i – j: ð7Þ (6) Normality: 1j ~XN 0; s 2In � � : ð8Þ JPIF 29,4/5 532 Also, in a practical sense, cross-sectional hedonic modeling requires a much larger data set than is typically available or easily procured. Given the typical practical admonition that appraisers use data which are spatially and temporally as close as possible to the subject (implicitly admitting that assumptions (3) and (4) are violated), then raising an adequate data set is often not practical. While a large enough data set (and, implicitly, a large enough budget) allows a researcher to attack these issues through a variety of well-accepted econometric tools (e.g. weighted least squares, logarithmic transformations, etc.), these tools and techniques are simply beyond the scope of the vast majority of appraisal situations. Further, use of these tools and techniques usually requires a very specialized skill-set in econometrics. Thus, the profession is left with the need for methodologies which have solid statistical characterization but which do not require the strict assumptions of OLS. Prior research has explored neural networks as an alternative, but with a focus on either reviewing “canned” software packages or examining the ways neural networks can improve hedonic pricing models. Worzala et al. (1995), McGreal et al. (1998), Liu et al. (2006), and Peterson and Flanagan (2009) provide a good review of these lines of research. Neither line of research suggests a solution to the adjustment-grid-versus- hedonics conundrum. The exploration of neural networks also introduced the term “fuzzy logic” to the appraisal lexicon, although, in fact, traditional appraisal methodology is simply a heuristic application of fuzzy logic. The term is generally credited to Zadeh (1964)[7], who was extending prior work on fuzzy sets: [. . .] a “fuzzy set” [. . .] extends the concept of membership in a set to situations in which there are many, possibly a continuum, of grades of membership. The concept of fuzzy sets and fuzzy logic is necessary in neural networks, because it is the best way – if not only way – to instruct computers to select not just comparables which are exact matches (an impossibility in real estate) but instead to learn from comparables which are close matches. Experienced real estate appraisers will immediately recognize that this is what they have always heuristically endeavored to accomplish in a sales adjustment grid. Thus, the application of fuzzy logic simply provides a means for computers to mine data sets for the closest comparables. Expert systems, on the other hand, provide the potential to provide the appraiser with statistical properties without the strict assumptions of OLS. Expert systems are adaptive to non-parametric data, and are useful for smaller data sets than would normally be required in hedonic models. The principal shortcoming of expert systems is that they do not come in a “one-size-fits-all” package, and require some modeling skills on the part of the appraiser. The remainder of this paper explores the theoretical basis for a mass appraisal, with an eye to describing the equilibrium condition of residential real estate markets so as to understand the nature of the mass appraisal model. The paper then presents two case studies of expert systems applications to demonstrate the modeling and statistical characterization techniques. For simplicity, this paper focuses on residential valuation. However, we have observed non-residential expert systems in place and used successfully in tax assessment situations. Expert systems and mass appraisal 533 2. The residential real estate market – a brief primer on equilibrium and modeling The nature of the residential real estate equilibrium is one of the least understood theoretical underpinnings of modern appraisal practice. While a thorough exploration deserved a separate stream of research, it is useful to mention the highlights to set the stage for the inputs to an expert systems model. Students of economics begin their studies with a purposefully simplistic example of supply and demand, as shown in Figure 1, to illustrate that an increase in demand (from D1 to D2) results in an increase in the quantity supplied (from Q1 to Q2) along with an increase in price (P1 to P2). Students who pay close attention in class will note that this is a partial equilibrium model, in which the demand, supply, quantity, and price of all complimentary and substitute goods are held constant. Demand and supply curves are assumed to be downward and upward sloping, respectively, universally differentiable (that is, no points of inflection) and concave[8]. Of course, in practice, nothing could be further from the truth. Housing supply is sticky in the short- and intermediate term, substitute goods abound, locational monopolies predominate, and consumer expectations (and resultant demand functions) are frequently incompatible with each other and with existing supply. Thus, at the very least, the actual model in practice more closely resembles Figure 2. In practice, and particularly in the short-run, the residential real estate market is more accurately described with a Nash equilibrium, in which a finite number of “demanders” are competing against each other to optimize their utility in a market of dissimilar but finite supply. Demanders take into account each others’ equilibrium strategy, and no one player can gain an advantage by unilaterally changing his or her strategy. The Nash equilibrium takes into account the notion that different demanders enter the market place with different strategies, but are faced with the same vector of heterogeneous supply. The idea of a mixed-strategy game was not unique to John Nash’s, 1951 paper, and in fact his work built on the earlier work of Von Neumann and Morganstern (1947). However, their earlier work assumed the special case of a zero-sum game. Nash generalized their work to show that a collective optimization could result in a Pareto-optimality that was not zero-sum[9]. Figure 1. Simple model of supply/demand JPIF 29,4/5 534 What are the implications for real estate valuation? Simply put, direct comparison of property transaction prices in order to develop a value must take into account the fact that each transaction price was arrived at via a slightly different “demander” strategy. Heuristic appraisal methods are able to take this into account – see the foregoing discussion of fuzzy logic – albeit in a fashion not well characterized from a statistical perspective. Indeed, explorations into the Daubert-esque characteristics of individual appraisals have revealed significant problems (see Kilpatrick, 2010). This is not to cast dispersions on OLS-type models, which do not incorporate such heuristic accommodation of strategy variation. Indeed, OLS-type models accommodate such idiosyncrasies as long as those idiosyncrasies are reasonably well behaved. Statistical theory tells us that with large sample sizes, the OLS assumptions at least asymptotically are true. Nonetheless, individualized appraisals, such as those conducted for mortgage lending and most eminent domain work, do not have the luxury of large data sets. To substitute for this, appraisers draw on so-called “appraisal judgment” in arriving at a reconciled “opinion of value.” The former of these terms actually draws from a field of statistical inference called Bayesian estimation, and the latter constitutes a maximum likelihood estimator. As will be shown in the next section, an expert system draws from these fields as well, albeit in a statistically characterizable fashion. Hence, a brief exploration of the two topics is in order. The Rev. Thomas Bayes (1702-1761) was a Presbyterian theologian and mathematician, who published on both topics. He is also well known for his defense of Isaac Newton’s development of calculus. In the early part of the eighteenth century, statisticians were fascinated with what were then called inverse probabilities, which we now refer to as conditional probabilities. In essence, the question asked is “What is the probability of something happening if we already know some predecessor information about that event?” The classic example is an urn filled with an equal number of black and white balls (say, ten of each). An introductory statistics student should be able to guess that the probability of drawing either a black or a white ball is 50-50, the same as a coin flip – a very simple binomial distribution. Now, what if five white balls have been drawn out of the urn in a row, but each time they are replaced in the urn and the urn shaken so that subsequent draws are totally random? On each Figure 2. Supply/demand with perfectly inelastic supply Expert systems and mass appraisal 535 subsequent draw, the same binomial distribution continues to apply. Counter-intuitively, the probability of a white versus black ball remains 50-50, despite the seemingly endless but totally random draw of white balls[10]. Conversely, what if the balls are not replaced, so after five draws, the urn is known to only possess five white balls, but ten black ones. What now is the probability of a white draw or a black one? Mathematically, this can be expressed as equation (9): Pr HjE � � ¼ PrðEjHÞ PrðHÞ PrðEÞ ; ð9Þ where: H ¼ The hypothesis being tested. E ¼ Prior knowledge observed by the researcher. Pr(H) ¼ The prior probability of H, before the researcher gained the prior knowledge. Pr(EjH) ¼ The conditional probability of seeing the evidence E if the hypothesis H is actually true. Pr(E) ¼ The marginal probability of observing E with or without H being true (or across all possible outcomes of Hi) – P PrðEjHiÞ PrðHiÞ: Pr(HjE) ¼ The posterior probability of H given the observation of E. Indeed, the student of Bayesian statistics will quickly see that the sales adjustment grid process is simply a very practical manifestation of Rev. Bayes’ theorem. The hypothesis being tested is the value of the property, and the prior knowledge is not only the comparable data, but also the appraiser’s judgment in making adjustments to such data. As a launch-point for the next discussion, Pr(EjH) is would be recognized by a statistician as a maximum likelihood estimator (MLE) and is the mathematical expression for the judgmental process by which the appraiser draws comparables for the sales comparison approach. The MLE asks the simple question, “Given what we know about these data, what probability process best fits”? Ordinary estimation procedures start with some predecessor assumption about the distributional characteristics of the data (i.e. normality). In OLS, the asymptotic properties allow us to make this assumption[11]. MLE, conversely, begins with the data instead, and then looks for a probability distribution that fits. In most cases, the MLE has desirable properties, and these properties fit well with appraisal assumptions: . Consistency. The ML estimator converges asymptotically to the value being estimated. From an appraisal perspective, this means that there is a benefit to experience and professionally-developed judgment. . Asymptotic normality. As sample size increases, the MLE distribution tends toward a normal distribution. To an appraiser, this suggests that the reliability of the value opinion improves with more comparable sales. . Efficiency. There is no asymptotically unbiased estimator that has lower asymptotic mean square error. To an appraiser, this suggests that the MLE minimizes the risk of error. JPIF 29,4/5 536 From a more technical perspective, suppose the appraiser draws a sample of data, x1, x2, x3, . . . xn, where the n observations are all independent and identically distributed. The distribution, f0( *) is unknown, but is believed to be a part of a family of distributions, so that f 0 ¼ fð* jFÞ where F is the true, but unknown value. Thus, it would be desirable to find some estimator, Fx, which would be as close in value to F as possible. The actual likelihood function is as shown in equation (11), and in practice, the log-likelihood (equation (12)) is used for simplicity (and also has certain properties which make it desirable in real estate analyses[12]): L Fjx1; x2:::xn � � ¼ f x1; x2:::xnjF � � ¼ Yn i¼1 f xijF � � ð10Þ ln L Fjx1; x2:::xn � � ¼ f x1; x2:::xnjF � � ¼ Xn i¼1 ln f xijF � � : ð11Þ Note that the iid assumption can be relaxed for the underlying data so long as it is possible to write a joint density function and that the parameter, F, has a finite dimension that does not depend on the sample size. This is a handy simplification in real estate, since the data are most likely not independent, but temporally and spatially correlated. In effect, the ML estimator is also usually the most probable Bayesian estimator. The literature tying MLE to appraisal practice is scant but growing. Assane (2007) uses MLE methods to reconcile spatial and hedonic models, and Ross et al. (forthcoming) extend this work using Monte Carlo simulation to determine the welfare implications one can draw from distance variables used in hedonic regression models. However, both of those papers, and other work in the area, endeavor to make improvements on the hedonic model, and demonstrate the usefulness of more data. The gap in the knowledge base is to apply these statistical tools in a fashion that provides statistical characterization in a limited dataset/non-parametric world. Thus, we are left with three important observations which tie heuristic appraisal methods to underlying economic and econometric theory: (1) The real estate transactional market constitutes a Nash equilibrium, in which all participants take into account the strategies of other participants in the goal of optimizing their utility. Thus, spatial and temporal autocorrelation are part of the process, rather than aberrations to the model. (2) Real estate valuation takes into account what is already known about the market, and indeed temporal and spatial autocorrelation makes it highly likely that the value of the nth property is tied inexorably to the value of the (n þ 1)th and (n 2 1)th properties. Considerable literature, not discussed here, discusses the appraisal smoothing problem (source?). Nonetheless, without the Bayesian priors that give rise to appraisal smoothing, heuristic methods would not be possible. (3) The appraiser, faced with a set of data and a set of prior observations about the underlying market, uses fuzzy logic to formulates a maximum likelihood estimator to determine the true value of the property. S/he is able to do this with a limited data set based on the Bayesian priors already known about the probable behavior of the market. The challenge now is to formulate a way to explain all of this in a statistically valid fashion. The next discusses the theory behind expert systems. Expert systems and mass appraisal 537 3. Expert systems – theoretical and methodological underpinnings Some of what we think of today as “expert systems” grew out of the data mining studies in the 1990s. Increasing amounts of data were being stored in relatively easily accessible databases. Researchers were interested in a variety of data analyses, including classification, discovery of associations, pattern identification, temporal modeling, deviation detection, dependency modeling, clustering, and characteristic rule discovery. Individual data analytical techniques, such as OLS, proved problematic working across disparate data sets. Out of that, fuzzy logic and hybrid systems developed which drew on the various strengths of different techniques. Goonatilake and Khebbal (1995) provide a contemporaneous outline of this emerging data analysis trend. McCluskey and Anand (1999) were among the first to thoroughly outline the application of such expert systems as they apply to mass appraisal. However, as early as 1989, Scott and Gronow outlined models that would simulate appraisal expertise (Scott and Gronow, 1989). Follow-up research in this vein was done by Nawawi and Gronow (1991) and Nawawi et al. (1997). Early research pointed to the problem of transparency in the expert system model. McCluskey and Anand (1999) basically outline two analytical strategies, both of which rely on separating the data (in this case, comparable transactions) into two sets: a comparable set (which will be directly used for valuation) and a separate data set which will be used for the system to “learn”. In their loosely coupled hybrid system (see Figure 3), the learning process is an artificial neural network which estimates attribute values. The weights on various values is computed using an equation they developed. Alternatively, they present a tightly coupled hybrid system (see Figure 4) in which a genetic algorithm starts out with a set of likely solutions. The various solutions are iteratively applied in a survival-of-the-fittest mode. The solution that best fits the data becomes the MLE which is then applied to the comparable data. McCluskey and Anand (1999) also present, and dismiss, what they call the “domain expert” model. In this model, relative weights and factors are determined by an expert, who has prior knowledge of the valuation equation. They state that the principal shortcoming of this model is that it requires an “expert” and thus is not self-learning. Given the foregoing discussions about Bayesian analysis, however, this domain expert, coupled with the two-stage paradigm of their tightly coupled hybrid system (which produces an MLE), provide a powerful underlying basis for a useable expert system. Figure 3. JPIF 29,4/5 538 4. Practical expert system applications – two case studies The challenge at this juncture is to develop expert systems that are actually adaptable in real life for actual appraisal problems. Much like heuristic or OLS models, expert systems will require appraiser input both in the modeling and the data selection. Unlike neural networks, the expert system will not “teach itself” but rather the appraiser will utilize the data, in a fuzzy-logic fashion, to develop adjustment factors. In practice, the expert system closely resembles heuristic sales adjustment grids, but with the added advantage of the ability for statistical characterizations. Case study 1 – Plaquemines Parish, Louisiana Plaquemines Parish has been in the news lately – it is “ground zero” for the April 2010 Deepwater Horizon explosion and subsequent oil spill. Many of the major news reports came out of the towns of Belle Chase and Venice. However, Plaquemines Parish was also “ground zero” or nearly so for Hurricane Katrina in 2005. Kilpatrick and Dermisi (2007) discuss much of the real estate research following that event, and one area of research was the impacts of flooding on property values. Considerable attention was paid to two major proposed class actions versus the Army Corps of Engineers, regarding breeches of the Lake Ponchartrain levee and the Mississippi River Gulf Outlet (MRGO) levee. In the former case, the Federal District Court ruled that the Army enjoyed protection from lawsuit by the Flood Control Act of 1936. In the latter, however, the Court ruled that the Act did not apply, since the MRGO was not dredged for flood control but rather for navigation. In Plaquemines Parish, however, the levees were not built and maintained by the Army, but rather by the Parish itself, which did not enjoy sovereign immunity. Residences were damaged by flooding from allegedly poorly designed, built, and/or maintained levees, and a class action suit was filed (Bermaster v. Plaquemines Parish). The complexity of the case was compounded by four somewhat related factors: (1) Properties in the parish are very thinly traded, and appraisals are “noisy” in statistical terms. (2) Tax assessment data in Louisiana are problematic compared to other jurisdictions, and in the case of Plaquemines Parish, records were destroyed when the Parish Courthouse burned down (ironically, not Katrina related). Figure 4. Expert systems and mass appraisal 539 (3) The case was significantly delayed when a judge died and a new judge had to be appointed. (4) Many properties in the Parish were affected by the flooding and were totally destroyed, and the occupants moved away. Adjudication of a class action required some methodology for establishing base-line values of the homes. Hedonic modeling would require data quality that was simply not available, and individualized appraisals of all of the affected properties would have been prohibitively expensive. For that reason, some sort of expert system was needed. The first step in the expert system was to identify a study area (Figure 5), which encompassed all of the proposed class area. Within this study area, comparables were collected (Figure 6) from transactions over a period of several years prior to the flooding. It was immediately obvious that the comparables were extraordinarily heterogeneous, and data verification was an important component of the study. The next step in the process was to apply appraisal expertise to the data in order to find common themes to the valuation of properties in this market place (see equation (12) below). This is analogous to the model specification process in a hedonic regression. Next, a hold-out sample was selected (Figure 7) which consisted of a cohesive neighborhood of properties which spanned most property types in the region, and for which good baseline values were either known or knowable through good tax assessment records, recent arms-length transactions, recent appraisals, or interviews with local appraisers or brokers: Vi;j ¼ BasePricej þ Qi;j 2 Qbase;j � � bj;1 þ Bri;j 2 Brbase;j � � bj;2 þ Ai;j 2 Abase;j � � bj;3 þ SFi;j 2 SFbase;j � � bj;4 þ Disti;j 2 Distbase;j � � bj;5 þ 1 ð12Þ where: Vi,j ¼ Estimated Value of property i in property type j. BasePricej ¼ Base price of property type j. Qi,j ¼ Quality Level of property i in property type j. Qbase,j ¼ Base Quality level of property type j. bj,1 ¼ Quality Level coefficient for property type j (or price adjustment due to difference between Quality level of property i and base Quality level in property type j). Bri,j ¼ Presence of Brick in property i of property type j. Brbase,j ¼ Base level Presence of Brick in property type j. bj,2 ¼ Presence of Brick coefficient for property type j (or adjustment due to difference between Presence of Brick in property i and base level of Presence of Brick in property type j). Ai,j ¼ Acreage of property i in property type j. Abase,j ¼ Base Acreage of property type j. JPIF 29,4/5 540 bj,3 ¼ Acreage coefficient for property type j (or adjustment due to difference between Acreage of property i and Base Acreage in property type j). SFi,j ¼ Square Footage of property i of property type j. SFbase,j ¼ Base Square Footage of property type j. Figure 5. Plaquemines Parish study area Expert systems and mass appraisal 541 bj,4 ¼ Square Footage coefficient for property type j (or adjustment due to difference between Square Footage of property i and base Square Footage in property type j). Disti,j ¼ Distance from Ferry of property i of property type j. Distbase,j ¼ Base Distance from Ferry of property type j. Figure 6. Plaquemines Parish comparables for expert system JPIF 29,4/5 542 bj,5 ¼ Distance of Ferry coefficient for property type j (or adjustment due to difference between Distance from Ferry of property i and base Distance from Ferry in property type j). 1 ¼ Error term. An example of this methodology is shown in Table I. Accuracy of the model is best described by the coefficient of dispersion (COD), a non-parametric tool recommended Figure 7. Plaquemines Parish sample area for test 1 Expert systems and mass appraisal 543 by the IAAO and widely used by tax assessors as an accuracy benchmark. COD for n properties is computed using equation (13), and for the Plaquemines test sample, the COD is 9.06 percent. This is well within the standards for this property type recommended by the IAAO: COD ¼ P abs Price 2 Valueð Þ=n median value : ð13Þ Table II also shows the results on a property-by-property basis. Case study 2 – Lomax, Illinois Located in western Illinois, just east of the banks of the Mississippi River near the Iowa/Missouri border, this small township had a population of 477 as of the 2000 census. Homes are typically several decades old, and rarely change hands. The few transactions that are recorded are rarely arms-length, since homes pass among family members or friends frequently. People who live there were often born there or married into families there. People leave typically by marrying and moving away. It was recently discovered that an oil pipeline, owned by BP Pipelines North America, has been leaking into the drinking water and soil of the homes in Lomax. The township and many of the property owners filed suit, and one of the analytical challenges was to craft a supportable baseline value for these homes. Again, a hedonic model would have been problematic. Individual appraisals of the homes might have been feasible, but would still have lacked statistical robustness and characterizability. Table III details the 20 homes that were valued, with home types including mobile homes, ranch homes, and typical Midwestern farmhouses. Comparable sales were drawn from the same county (Henderson) as well as adjacent Mercer and Hancock counties. Over 250 comparables were analyzed, broken down into five categories: (1) Double-wide manufactured homes. (2) Single-wide mobile homes, sold with the lot as a unit. (3) Old-style homes category 1. (4) Old-style homes category 2. (5) Ranch homes. Adjustments were made based on factors specific to each group. For example, under the ranch category, adjustments had to be made for lot size, improvements size, full baths (beyond 1), half baths (beyond -0-), garage size as it differed from 575, basement Property type Base value ($) Qual/cond Brick Acres SF Distance Coefficients ($) ! 20,000 5,000 3,000 25 2500 Property type “J” base values 48,000 2 0 0.5 1,200 17 Property type “I” characteristics 2 1 0.3 1,350 21 Notes: Vali ¼ $48,000+((2 2 2)*$20,000)+((1 2 0)*$5,000)+((0.3 2 0.5)*$3,000)+((1,350 2 1,200)*25)+ ((21 2 7)* 2 500) Vali ¼ $48,000+$0+$5,000 2 $600+$3,750 2 $2,000 Vali ¼ $54,150 Table I. Case study no. 1 valuation example JPIF 29,4/5 544 ID S a le d a te S a le p ri ce ($ ) T im e a d ju st ed p ri ce ($ ) Im p r T y p e Q /C B ri ck A cr es S F D is t V a lu e ($ ) E rr o r (% ) 9 3 /4 /2 0 0 0 4 8 ,5 5 1 6 3 ,1 6 4 1 2 0 .5 0 .2 7 1 ,1 5 6 0 .1 3 7 7 ,1 4 7 2 2 2 1 5 5 /1 0 /2 0 0 1 6 0 ,0 0 0 7 6 ,5 2 8 1 2 0 2 .8 5 1 ,5 0 0 6 .4 5 6 7 ,8 4 2 1 1 1 8 .1 1 0 /1 5 /2 0 0 2 2 1 ,0 0 0 2 6 ,0 0 5 1 1 0 3 .8 8 8 3 4 1 8 .1 3 2 8 ,4 1 5 2 9 1 8 .2 8 /1 2 /2 0 0 3 2 1 ,0 0 0 2 5 ,2 4 7 1 1 0 3 .8 8 8 3 4 1 8 .1 3 2 8 ,4 1 5 2 1 3 1 9 8 /2 0 /2 0 0 1 4 0 ,0 0 0 5 1 ,0 1 9 1 2 0 1 .6 2 1 ,0 9 2 1 9 .5 8 4 7 ,3 8 4 7 2 0 8 /3 0 /2 0 0 1 5 0 ,0 0 0 6 3 ,7 7 3 1 2 0 0 .5 8 1 ,2 7 9 2 1 .0 0 4 8 ,2 2 0 2 4 2 3 1 2 /2 0 /2 0 0 2 6 7 ,0 0 0 8 2 ,9 6 7 1 3 0 0 .7 8 1 ,3 8 3 2 2 .4 3 7 0 ,6 8 9 1 5 2 6 5 /1 9 /2 0 0 4 4 5 ,0 0 0 5 1 ,0 3 9 1 2 0 0 .3 4 1 ,4 7 5 2 2 .5 7 5 1 ,6 0 0 2 1 3 2 .1 1 1 /3 /2 0 0 0 3 5 ,0 0 0 4 5 ,5 3 4 1 2 0 0 .2 8 1 ,3 6 8 2 1 .5 1 4 9 ,2 7 0 2 8 2 1 9 /1 1 /2 0 0 2 6 0 ,0 0 0 7 4 ,2 9 9 1 .5 2 0 0 .6 7 2 ,2 9 1 2 1 .8 0 8 0 ,8 3 2 2 9 2 2 1 /1 0 /2 0 0 1 4 0 ,0 0 0 5 1 ,0 1 9 1 .5 2 0 0 .5 0 1 ,0 7 0 2 1 .8 4 4 5 ,9 0 3 1 0 2 4 1 0 /5 /2 0 0 0 5 6 ,6 6 0 7 3 ,7 1 3 1 .5 3 0 0 .2 7 1 ,4 8 5 2 2 .5 5 7 3 ,5 4 1 0 2 5 8 /1 9 /2 0 0 5 1 2 0 ,0 0 0 1 2 8 ,4 0 0 1 .5 3 0 1 0 .5 8 1 ,5 1 2 2 2 .3 7 1 1 5 ,9 5 2 1 0 4 3 1 1 /2 5 /2 0 0 2 1 4 4 ,0 0 0 1 7 8 ,3 1 8 1 .5 4 0 .5 0 .2 6 3 ,1 1 7 3 .9 6 1 8 2 ,1 4 8 2 2 4 4 /1 3 /2 0 0 5 1 3 5 ,0 0 0 1 4 4 ,4 5 0 2 3 1 1 .0 5 1 ,7 0 4 0 .3 9 1 5 3 ,1 6 9 2 6 1 0 6 /1 2 /2 0 0 3 2 0 5 ,0 0 0 2 4 6 ,4 6 2 2 4 1 3 .2 0 1 ,8 4 4 2 .2 8 2 1 2 ,1 2 4 1 4 1 1 .2 4 /5 /2 0 0 4 2 0 0 ,0 0 0 2 2 6 ,8 4 0 2 4 1 8 .8 7 1 ,2 5 4 3 .4 5 2 1 3 ,6 2 3 6 1 2 1 1 /1 /2 0 0 2 1 1 0 ,0 0 0 1 3 6 ,2 1 5 2 3 1 0 .6 8 1 ,9 0 4 3 .3 7 1 5 5 ,3 6 7 2 1 4 2 7 8 /1 4 /2 0 0 0 1 2 6 ,0 0 0 1 6 3 ,9 2 3 2 4 1 0 .5 9 2 ,1 9 9 7 .1 7 2 0 2 ,0 4 3 2 2 3 2 9 1 /3 1 /2 0 0 3 1 3 8 ,0 0 0 1 6 5 ,9 1 1 2 3 1 0 .2 1 2 ,6 3 7 1 .7 5 1 9 2 ,8 8 7 2 1 6 3 0 2 /1 1 /2 0 0 5 9 9 ,5 0 0 1 0 6 ,4 6 5 2 2 1 0 .2 0 1 ,5 8 5 1 .7 8 1 0 5 ,2 0 7 1 3 1 .1 5 /2 9 /2 0 0 2 8 0 ,0 0 0 9 9 ,0 6 6 2 2 1 0 .2 9 1 ,4 8 5 4 .0 1 9 1 ,2 0 5 8 3 9 .1 2 /2 2 /2 0 0 2 1 6 0 ,0 0 0 1 9 8 ,1 3 1 2 4 1 0 .3 4 2 ,1 6 8 3 .7 5 2 1 6 ,1 2 0 2 9 3 9 .2 7 /2 3 /2 0 0 4 2 0 0 ,0 0 0 2 2 6 ,8 4 0 2 4 1 0 .3 4 2 ,1 6 8 3 .7 5 2 1 6 ,1 2 0 5 4 0 4 /3 0 /2 0 0 4 1 8 6 ,0 0 0 2 1 0 ,9 6 1 2 3 1 0 .2 9 2 ,5 6 3 3 .8 8 1 9 0 ,3 3 4 1 0 4 1 1 0 /3 1 /2 0 0 2 2 0 5 ,0 0 0 2 5 3 ,8 5 6 2 4 1 0 .3 7 2 ,1 4 9 3 .8 1 2 1 7 ,6 7 2 1 4 4 2 2 /5 /2 0 0 1 1 5 0 ,0 0 0 1 9 1 ,3 2 0 2 3 1 0 .2 4 1 ,7 5 0 3 .8 5 1 4 9 ,4 9 4 2 2 4 4 1 0 /1 7 /2 0 0 2 1 5 3 ,0 0 0 1 8 9 ,4 6 3 2 3 1 0 .7 0 2 ,3 0 0 3 .9 9 1 7 4 ,0 1 9 8 4 6 3 /3 1 /2 0 0 3 2 3 2 ,0 0 0 2 7 8 ,9 2 2 2 4 1 0 .2 8 2 ,8 0 3 3 .9 2 2 3 7 ,2 3 0 1 5 1 1 1 /7 /2 0 0 3 1 8 2 ,0 0 0 2 1 8 ,8 1 0 3 3 0 .5 0 .3 4 2 ,7 3 5 3 .5 0 2 0 7 ,7 8 3 5 3 .3 1 0 /1 8 /2 0 0 1 3 1 5 ,0 0 0 4 0 1 ,7 7 3 3 5 1 3 .9 7 3 ,4 8 4 0 .4 7 4 0 4 ,4 3 4 2 1 1 3 4 /1 6 /2 0 0 1 3 4 5 ,0 0 0 4 4 0 ,0 3 7 3 4 0 .5 1 9 .9 7 3 ,3 3 6 4 .3 7 4 1 5 ,2 4 2 6 2 8 5 /1 5 /2 0 0 1 1 3 0 ,0 0 0 1 6 5 ,8 1 1 3 3 1 0 .3 4 2 ,8 1 8 7 .1 1 2 0 3 ,0 7 5 2 2 2 3 3 5 /1 /2 0 0 3 2 3 0 ,0 0 0 2 7 6 ,5 1 8 3 4 1 0 .3 4 2 ,7 0 0 3 .8 1 2 6 2 ,6 0 6 5 3 5 8 /3 0 /2 0 0 1 1 9 0 ,0 0 0 2 4 2 ,3 3 9 3 3 0 0 .2 4 3 ,8 2 3 3 .7 9 2 7 4 ,5 2 3 2 1 3 3 6 .1 8 /2 1 /2 0 0 0 2 2 8 ,0 0 0 2 9 6 ,6 2 3 3 3 1 0 .2 4 3 ,6 6 4 3 .7 9 2 6 2 ,4 8 4 1 2 3 6 .2 3 /2 7 /2 0 0 1 2 3 5 ,0 0 0 2 9 9 ,7 3 5 3 3 1 0 .2 4 3 ,6 6 4 3 .7 9 2 6 2 ,4 8 4 1 2 3 6 .3 5 /7 /2 0 0 4 2 4 2 ,5 0 0 2 7 5 ,0 4 4 3 3 1 0 .2 4 3 ,6 6 4 3 .7 9 2 6 2 ,4 8 4 5 3 7 1 0 /2 7 /2 0 0 0 1 2 9 ,0 0 0 1 6 7 ,8 2 6 3 3 1 0 .2 9 2 ,5 0 0 3 .7 5 1 9 0 ,4 5 8 2 1 3 3 8 9 /3 /2 0 0 4 1 9 5 ,0 0 0 2 2 1 ,1 6 9 3 3 0 0 .2 4 2 ,0 7 2 3 .7 8 1 6 9 ,4 6 6 2 3 4 5 1 2 /1 /2 0 0 4 1 7 9 ,0 0 0 2 0 3 ,0 2 2 3 3 1 0 .3 8 2 ,3 4 1 3 .9 9 1 8 3 ,3 6 6 1 0 Table II. Plaquemines test results Expert systems and mass appraisal 545 (full, partial, or finished), fireplace (beyond a base of -0-), deck, porch, enclosed porch, patio, shed, proximity to a railroad, and condition. Note that different adjustments were made for different categories of homes. Table IV shows the CODs for Lomax property types (see also Figure 8). As an example, the median value for ranch homes determined by the comparables was $70,364. The average dispersion was $5,399, resulting in a coefficient of dispersion of 7.67 percent. Across the five categories, the range was 7.67 percent to 35.21 percent. For three of the categories, the properties were within recommendations of IAAO, the fourth category was nearly so, and only one category (mobile homes) was well outside of the range of standards. However, at least these measures of dispersion were determinable, and so in a Daubert setting, the evidence would be acceptable. 5. Summary, conclusions, and recommendations for future research One of the current challenges in real estate valuation is to measure the statistical properties of the appraisal estimates. Colwell et al. (2009) demonstrate that when the data are available, appraisers should always opt for a statistically characterizable GA-ID Lot SF Yr blt Bldg SF Rms BR Ba .5 Ba Gar SF Descr 1 1,936 1820 1,936 7 5 1 0 0 Cape Cd 2 19,602 1998 1,836 7 3 1 0 0 Doubwide 3 43,734 1970 1,064 5 3 1 0 576 Ranch 4 25,570 1973 1,266 5 3 1 0 576 Ranch 5 7,656 1900 929 4 2 1 0 368 1-stry 6 15,312 1920 1,163 6 2 1 0 490 1-stry 7 8,712 1920 1,542 6 2 1 0 637 1.5 stry 8 10,890 1920 1,428 5 3 1 0 576 1-stry 9 55,321 1998 1,232 5 3 2 0 0 m-home 10 7,950 1935 1,186 6 3 1 0 1,080 1-stry 11 8,557 1900 1,554 7 4 1 0 0 1.5 stry 12 33,055 1900 1,114 4 2 1 0 0 1-stry 13 22,704 1920 1,055 5 2 1 0 450 1-stry 14 18,480 1986 1,248 6 3 1 1 576 Ranch 15 16,368 1981 1,068 5 3 1 0 732 Ranch 16 8,712 1900 792 4 1 1 0 720 1-stry 17 12,000 1950 1,184 4 2 1 0 624 1-stry 18 12,000 1954 1,144 5 3 1 1 704 Ranch 19 6,000 1996 1,224 5 3 2 0 0 M-home 20 9,000 1910 742 4 2 1 0 0 1-stry Table III. Homes in Lomax, Illinois COD Property type (%) Double-wide 8.05 Mobile home 35.21 Old style 1 21.81 Old style 2 13.40 Ranch 7.67 Table IV. Lomax coefficients of dispersion JPIF 29,4/5 546 methodology, and in fact, in the courtroom, it is generally required that the appraiser report on the measurable error rate of the estimator. While many advanced methods such as hedonic modeling, meta analysis, or survey research provide for such statistical measurements, these methods require significantly large data sets and computational undertakings. The goal of this study is to present some of the theoretical underpinnings of expert systems, to help define its use, to give some examples of such use, and to briefly discuss ways to statistically characterize the output. The concept of statistical characterization is key – both for acceptability in litigation, but also to aid Figure 8. Lomax, Illinois Expert systems and mass appraisal 547 mortgage-financing appraisers in understanding the error rates of sales adjustment-grid appraisals. Expert systems provide potential for a solution that lies between the purely heuristic sales adjustment grid and the more computationally intensive hedonic model. The benefits of statistical characterization are important, and with risk of pontification, this paper hopes to drive that forward so as to minimize, as Shiller and Weiss put it, the Type II errors in residential mortgage lending. Additional work can quickly be done in two areas: (1) building on Kilpatrick (2010) to find ways to better characterize individual appraisals; and (2) extend this research to add more non-parametric measurement to the expert systems calculations. Notes 1. In some countries, real estate appraisal takes on different characteristics, and is often more highly regulated. For example, Germany has a strictly codified process dependent on local councils which periodically publish land value, to which the depreciated value of the improvements can be added. In most parts of Mexico, a sales comparison is typically not useful and again a variant of a cost approach is preferred. 2. Daubert v. Merrell Dow Pharmaceuticals, 509 USA 579 (1993), now formally captured in Rule 702, Federal Rules of Evidence. 3. Kumho Tire Co. v. Charmichael, 526 USA 137 (1999). 4. Notably, some states which have not adopted Daubert nonetheless have rules of evidence which emulate Daubert. See Kaufman (2006). 5. Of the remaining 20 states, many adhere to what are known as the Frye standards, from Frye v. United States, 293 F. 1013 (DC Cir. 1923). It establishes a bar of general acceptance, but is silent as to the need for a known or knowable error rate. 6. A complete restatement of the theorem is beyond the scope of this paper, and while most econometric students should be familiar with it, a simple summary may be useful: In a linear regression model, if the expected value of the errors is zero, the errors are uncorrelated, and have homoskedastic variances, the best linear unbiased estimator is the ordinary least squares estimator. Note that the errors do not have to be normal or even identically distributed. (As a bit of an aside, Carl Friedrich Gauss (1777-1855) and Andrey Markov (1856-1922) never actually worked together. In the 1800s, the theorem was credited to Gauss alone. Markov’s contributions – as a Russian – were allegedly overlooked in the west, according to a 1934 article in the Journal of the Royal Statistical Society. Most writers attribute the first use of the term “Gauss Markov Theorem” to the book by Scheffe (1959), Analysis of Variance.) 7. Zadeh also published about fuzzy sets in 1965 and 1968 (Zadeh, 1965, 1968). 8. This more-or-less standard graphical representation is generally attributed to Alfred Marshall. Simple presentations often use straight lines to represent demand and supply, but concavity is necessary for there to be a marginal rate of substitution. Fisher suggests that this second-order economic behavior may be biologically imprinted. 9. Note that a Pareto optimality does not imply a singular, unique point of maximum utility, but encompasses the possibility of multiple solutions. In a Nash equilibrium, no one player can change the mix with unilateral decisions, but collectively, multiple solutions could emerge. JPIF 29,4/5 548 10. Casino owners watch with glee as roulette players see a long string of “reds” show up, and thus bet heavily on black, thus further enhancing the profitability to the casino. The opening scene of Tom Stoppard’s classic play, Rosencranz and Gildenstern Are Dead uses this phenomenon to set the stage for the play’s commentary on the fickleness of fate. 11. This is the second time asymptotic normality has been referred to in this paper, and perhaps it useful to define it a bit better. As the number of observations goes to infinity, the distributional characteristics emulate a normal one. This can be mathematically expressed per equation A-1: fðXÞ , N ðas n ) 1Þ: ðA-1Þ 12. Since the log is a monotonic transformation, a log-likelihood solution is equivalent to a likelihood solution. In addition, real estate prices and values are famously nonlinear and have a truncated lower bound at zero. References Assane, D.D. (2007), “Estimating the effect of air quality: spatial versus traditional hedonic price models”, Southern Economic Journal, Vol. 73, April, pp. 1088-111. Colwell, P., Trefzger, J.W. and Heller, J.A. (2009), “Expert testimony: regression analysis and other systematic methodologies”, The Appraisal Journal, Vol. 77, Summer, pp. 253-62. Des Rosiers, F., Thériault, M. and Villeneuve, P. (2000), “Sorting out access and neighbourhood factors in hedonic price modeling”, Journal of Property Valuation & Investment, Vol. 18 No. 3, pp. 291-315. Goonatilake, S. and Khebbal, S. (1995), “Intelligent hybrid systems: issues, classifications, and future directions”, in Goonatilake, S. and Khebbal, S. (Eds), Intelligent Hybrid Systems, John Wiley & Sons, Chichester. Kaufman, M.S. (2006), “The status of Daubert in state courts”, Atlantic Legal Foundation, March 31. Kilpatrick, J.A. (2010), “What is the error rate of an appraisal?”, paper presented at the American Real Estate Society Annual Meeting. Kilpatrick, J.A. and Dermisi, S. (2007), “The aftermath of Katrina: recommendations for real estate research”, Journal of Real Estate Literature, Vol. 15 No. 2, pp. 213-27. Liu, J.-G., Zhang, X.-L. and Wu, W.-P. (2006), “Application of fuzzy neural network for real estate prediction”, Advances in Neural Networks, Springer, Berlin and Heidelberg, pp. 1187-91. McCluskey, W. and Anand, S. (1999), “The application of intelligent hybrid techniques for the mass appraisal of residential properties”, Journal of Property Investment & Finance, Vol. 17 No. 3, pp. 218-38. McGreal, S., Adair, A., McBurney, D. and Patterson, D. (1998), “Neural networks: the prediction of real estate values”, Journal of Property Valuation & Investment, Vol. 16 No. 1, pp. 57-70. Nash, J. (1951), “Non-cooperative games”, The Annals of Mathematics, Vol. 54 No. 2, pp. 286-95. Nawawi, A.H. and Gronow, S. (1991), “Expert systems in rating valuation”, Journal of the Society of Surveying Technicians, Vol. 18 No. 8, pp. 66-72. Nawawi, A.H., Jenkins, D. and Gronow, S. (1997), “Expert system development for the mass appraisal of commercial property”, in McCluskey, W.J. and Adair, A.S. (Eds), Computer Assisted Mass Appraisal: An International Review, Ashgate, London. Pace, R.K., Barry, R., Clapp, J.M. and Rodriguez, M. (1998), “Spatiotemporal autoregressive models of neighborhood effects”, Journal of Real Estate Finance and Economics, Vol. 17 No. 1, pp. 15-33. Expert systems and mass appraisal 549 Peterson, S. and Flanagan, A.B. (2009), “Neural network hedonic pricing models in mass real estate appraisal”, Journal of Real Estate Research, Vol. 31 No. 2, pp. 147-64. Scheffe, H. (1959), Analysis of Variance, John Wiley, New York, NY. Scott, I. and Gronow, S. (1989), “Valuation expertise: its nature and application”, Journal of Valuation, Vol. 8 No. 4, pp. 362-75. Shiller, R.J. and Weiss, A.N. (1999), “Evaluating real estate valuation systems”, Journal of Real Estate Finance and Economics, Vol. 18 No. 2, pp. 147-61. Smith, T.R. (1971), “Multiple regression and the appraisal of single family residential properties”, The Appraisal Journal, Vol. 39 No. 2, April, pp. 277-84. Von Neumann, J. and Morganstern, O. (1947), The Theory of Games and Economic Behavior, Princeton University Press, Princeton, NJ. Worzala, E., Lenk, M. and Silva, A. (1995), “An exploration of neural networks and its application to real estate valuation”, Journal of Real Estate Research, Vol. 10 No. 2, pp. 185-201. Zadeh, L.A. (1964), Memorandum 4307-PR 1, Rand Corporation, Santa Monica, CA. Zadeh, L.A. (1965), “Fuzzy sets”, Information and Control, Vol. 8, pp. 338-53. Zadeh, L.A. (1968), “Probability measures and fuzzy events”, Journal of Mathematical Analysis and Applications, Vol. 23, pp. 421-7. Further reading Barndorff-Nielson, O.E. and Cox, D.R. (1989), Asymptotic Techniques for Use in Statistics, Chapman & Hall, London. McLean, D., Mundy, B. and Kilpatrick, J. (1999), “Summation of evidentiary rules for real estate experts mandated by Daubert v. Merrell Dow Pharmaceuticals, Inc.”, Real Estate Issues, Fall. Mahle, S. (1999), “The impact of Daubert v. Merrell Dow Pharmaceuticals, Inc., on expert testimony: with applications to securities litigation”, The Florida Bar Reporter, April. About the author Dr John Kilpatrick is the CEO of Greenfield Advisors, a 34-year-old firm specializing in complex real estate problem-solving with offices in Seattle and Atlanta. He has a PhD in Real Estate Finance, and prior to joining Greenfield he taught real estate and corporate finance at the University of South Carolina. He was also one of the founders, and the first Administrator, of the South Carolina Supercomputer Network. Dr Kilpatrick is a Visiting Scholar in Real Estate at the Zicklin School of Business, Baruch College, City University of New York, and serves on the Finance Department Advisory Board for Washington State University. He is the author of four books on real estate and contributing author to three others, as well over 100 monographs, scholarly journal articles, and working papers. Dr Kilpatrick is a Fellow of the American Real Estate Society, the Membership Chair of the Real Estate Counseling Group of America, and serves on the Editorial Board of the Journal of Sustainable Real Estate. He is featured in the current editions of Who’s Who in America and Who’s Who in Business and Finance. He has been featured in the Wall Street Journal, the New York Times, the Boston Herald, and numerous other major periodicals. John Kilpatrick can be contacted at: john@greenfieldadvisors.com JPIF 29,4/5 550 To purchase reprints of this article please e-mail: reprints@emeraldinsight.com Or visit our web site for further details: www.emeraldinsight.com/reprints 
work_pjyyzfpgdbaf5mh2ngygy5fbgu	----	A genetic search of patterns of behaviour in OSS communities Expert Systems with Applications 39 (2012) 13182–13192 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A genetic search of patterns of behaviour in OSS communities M.R. Martínez-Torres ⇑ University of Seville, Facultad de Turismo y Finanzas, Avda. San Francisco Javier, s/n 41018 Sevilla, Spain a r t i c l e i n f o a b s t r a c t Keywords: Open source software Virtual communities Social network analysis Genetic algorithm Factor analysis 0957-4174/$ - see front matter � 2012 Elsevier Ltd. A http://dx.doi.org/10.1016/j.eswa.2012.05.083 ⇑ Tel.: +34 954 55 43 10; fax: +34 954 55 69 89. E-mail address: rmtorres@us.es This paper proposes the identification of patterns of behaviour of open source software (OSS) communi- ties using factor analysis and their social network analysis (SNA) features. OSS communities can be modelled as a social network in which nodes represent the community members and arcs represent the social interactions among them, and factor analysis is able to provide the factors that explain the latent patterns of behaviour. Due to the complexity of the problem and the high number of SNA features that can be extracted, this paper proposes a genetic search of an optimum subset of indicators leading to a group of latent patterns of behaviour maximizing the explained data variance and the interpretation of factors. Obtained results illustrate the feasibility of the proposed framework to extract relevant informa- tion from a large set of data. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction The OSS projects differ greatly from commercial software devel- opment models in several aspects. For instance, commercial soft- ware companies prevent access to the source code of their products from outside developers and customers, while OSS allows source code to be freely modified and redistributed under ‘‘open source’’ licenses (Feller & Fitzgerald, 2002). Besides, OSS projects are typically developed in a distributed and decentralized way as a difference to proprietary software, based on closed and formal structures. Precisely, one of the most distinctive characteristics of OSS projects is the fact that they are written, developed, and de- bugged largely by worldwide volunteers, who in most cases are connected and collaborate solely through the Internet. Therefore, the community behind the development of the project plays an essential role for the project to success (Deek & McHugh, 2008). Several authors have described OSS development teams as hav- ing a hierarchical or onion-like structure (Crowston & Howison, 2005), with a central core of highly active individuals, surrounded by other layers of progressively less active individuals. One exam- ple of this is presented in the study by Ye, Nakakoji, et al. (2005) where the central core is composed of the project leaders and core members, with five outer layers containing active developers, peripheral developers, bug reporters, passive users, and stakehold- ers, respectively. It has been demonstrated that much of the OSS development is realized by a small percentage of individuals de- spite the fact that there are tens of thousands of available develop- ers. Such concentration is called ‘‘participation inequality’’ (Kuk, ll rights reserved. 2006), and it can be explained by the different user profiles of open source communities. Participation inequality allows the categori- zation of OSS community members in three groups (Mockus, Fielding, & Herbsleb, 2002; Xu, Gao, Christley, & Madey, 2005). Core members are responsible for guiding and coordinating the development of an OSS project. They are usually involved with the project during a long period of time and have made significant contributions to the development and evolution of the system. Moderators and leaders are included in this group. Active develop- ers are those community members that regularly make contribu- tions to the project. Finally, peripheral developers occasionally contribute with new features to the existing system. This contribu- tion is irregular, and the period of involvement is short and spo- radic. Free riders (people who just are seeking answers without making any contributions) are also included in this group. The social structure of OSS teams directly influences the participation and the decision-making process affecting the overall performance of the project. Therefore, an important research question is extracting the different patterns of behaviour in OSS communities. These patterns leading to successful projects are of great interest both for autonomous and sponsored communities, that share a common aim of retaining and attracting participants to their communities (West & O’mahony, 2008). However, the structure of communities can only be derived from the participa- tion activity of their members. For this purpose, OSS communities have been frequently modelled as a social network, being the nodes of the network the community members while the arcs represent the flow of interactions among users (Toral, Martínez-Torres, & Barrero, 2009a). These networks are then analyzed using Social Network Analysis (SNA) techniques by obtaining a set of SNA fea- tures. For instance, previous studies have considered the size and out-degree of nodes (Valverde, Theraulaz, Gautrais, Fourcassie, & http://dx.doi.org/10.1016/j.eswa.2012.05.083 mailto:rmtorres@us.es http://dx.doi.org/10.1016/j.eswa.2012.05.083 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Sergio Cuadro de texto Sergio Cuadro de texto Sergio Cuadro de texto M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 13183 Sole, 2006), closeness centrality (Panchal, 2009) and betweeness centrality (Hossain, Wu, & Chung, 2006; Toral, Martínez-Torres, Barrero, & Cortés, 2009b), the clustering coefficient (Kwon, Oh, & Jeon, 2007; Singh, 2010), structural holes (Okoli & Oh, 2007) or the brokerage role of nodes (Sowe, Stamelos, & Angelis, 2006; Barcellini, Détienne, & Burkhardt, 2009; Toral, Martínez Torres, & Barrero, 2010), among other SNA features. Moreover, each of these measurements can be computed for the whole network or for several specific sub-networks like the one obtained from active developers or from the core group of the community. Due to the large number of possible SNA indicators that can be obtained, previous studies have been focused on a small number of social net- work features to characterize participation and obtain OSS patterns of behaviour. As a difference, in this study participation is analyzed using the whole set of SNA features that can be obtained from OSS communities. For this purpose, a genetic search of the optimum subset of indicators able of providing the main pattern of behaviour in OSS communities is proposed. Therefore, the main contribution of this paper is the possibility of dealing with all the SNA features of the social networks modelling OSS communities through an evolutionary computation technique like Genetic Algorithms. The remainder of the paper is structured as follows. First, previ- ous studies related to social network structures in OSS communities are reviewed in Section 2. In Section 3 the problem is formulated and the proposed approach described. Section 4 describes the Genetic Algorithm implementation. Obtained results and discus- sion are included in Section 5. Finally, conclusions are detailed in Section 6. 2. Related work Social network theory is based on the idea that social interac- tion patterns reflect the behaviour of individuals (Freeman, 2004). Therefore, the analysis of the social interactions can provide information about how individuals behave or how groups are orga- nized. This is why social network analysis focuses on the relation- ships between people, instead of on characteristics of people. By mapping these relationships, network analysis helps to uncover the emergent and informal communication patterns present in an organization, which in turn can be used to explain several orga- nizational phenomena. A social network can be modelled as a graph with nodes repre- senting people or groups, and links representing relationships or information flows between them. OSS communities constitute clear examples of dynamic social networks, as it is changing over time. Social networks begin when developers join a project, work with others, and form co-working relationships (Xu, Christley, & Madey, 2006). These relationships are important because the sense of belonging to a group is encouraged through social network bonding. The aim of OSS communities is attracting and retaining people, as this will benefit the underlying software. As new mem- bers join the community, different user profiles emerge. Not all the users are interested in participating the same way. Some of them, a small percentage, intensely contribute to the development of the project, while the rest of them make regular, occasional or even no contribution at all. These variety of user profiles lead to a core/periphery structure, where the core group of developers is lo- cated at the center of the community and the rest of them far away from the center depending on their contributions. The core group members are strongly connected to each other while the periphery contains members who are usually weakly connected to each other as well as to the core members (Long & Siau, 2007). Mathematically, a social network can be represented as a graph G = (V, E) where V denotes a finite set of vertices and E denotes a finite set of edges such that E # V � V. Some network analysis methods are easier to understand when graphs are conceptualized as matrices (Nooy, Mrvar, & Batagelj, 2005) as shown in Eq. (1). M ¼ðmi;jÞn�n; where n ¼ jVj; mi;j ¼ 1 if ðv i; v jÞ 2 E 0 otherwise � ð1Þ In case of a valued graph, real valued weight function w(e) is defined on the set of edges, i.e., wðeÞ¼ ExR, and the matrix is then defined as given by Eq. (2). Mi;j ¼ wðeÞ if ðv i; v jÞ 2 E 0 otherwise � ð2Þ In the context of OSS communities, V is given by all the commu- nity members and E is given by the interactions among them. The resulting network is a directed network in the sense that edges are actually arcs from one community member to another one, and the direction of the arcs shows the flow of information between them. It is also a valued network, as it is possible multiple interaction be- tween the same community members. Networks can be partitioned using some discrete characteristics of vertices. For instance, several classes of vertices can be obtained using the function w(e), that is, the strength of arcs. In the case of OSS projects, these kinds of partitions should highlight the core/ periphery (C/P) structure of the community. A C/P structure divides vertices in three distinct subgroups: vertices in the core, densely connected with each other, and vertices on the periphery, which in turn can be divided in active and peripheral vertices depending on their level of interaction. The following list summarizes the characteristics that can be computed for the whole network or each of the mentioned sub- networks as well as the main previous studies focused on them. Size and connectivity: the size of the community is the number of members the community has and the number of arcs represent the connectivity among these community members. Both indica- tors has been frequently used as a measure of the successful devel- opment of a OSS project. Success in a virtual community could be manifested through the level of participation, which can be under- stood as the number of participants (Preece, 2001) or as the level of community activity and quantity of community work output (Hinds & Lee, 2008). Several topological indices based on connec- tivity can be computed, like the Zagreb group index, the Randic connectivity index or the Platt index, all of them defined in terms of connections of nodes (Devillers & Balaban, 1999). Density: it is defined as the number of lines in a simple network, expressed as a proportion of the maximum possible number of lines. The main problem of this definition is that it does not take into account valued lines higher than 1 and it depends on the net- work size. A different measure of density is based on the idea of the degree of a node, which is the number of lines incident with it (Tor- al et al., 2009a). A higher degree of nodes yields a denser network, because nodes entertain more ties. The advantage of average de- gree is that it is a non-size dependent measure of density. As OSS communities are directed networks, several statistical measures of the out-degree distribution can be considered. Finally, density can be measured alternatively using an egocentric point of view; the egocentric density of a node is the density of ties among its neighbours (Nooy et al., 2005). In previous works, density has been used to study the coordination performance of OSS projects. Sev- eral studies conclude there is a negative impact of density over the quality of the project and its coordination (Hossain & Zhu, 2009; Feczak & Hossain, 2011). Components: A strong component is a maximal strongly con- nected subnetwork. A network is said to be strongly connected if each pair of vertices is connected by a path, taking into account the direction of arcs (Nooy et al., 2005). In the context of this study, Sergio Cuadro de texto 13184 M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 components allow the identification of connected substructures in the OSS community. K-cores: a k-core is a sub-network in which each node has k de- gree in that sub-network. The core with the highest degree is the central core of the network, detecting the set of nodes where the network rests on (Toral et al., 2010). Distance: it is defined as the number of steps in the shortest path that connects two vertices. Distance between members in the OSS community can affect how ideas and discussions can spread over the community. Short distances mean information only has to travel a few links to reach anybody in the network (Xu et al., 2006). Closeness centralization: it is an index of centrality based on the concept of distance. The closeness centrality of a node is calculated considering the total distance between one node and all other nodes, where larger distances yield lower closeness centrality scores. The closeness centralization is an index defined for the whole network, and it is calculated as the variation in the closeness centrality of vertices divided by the maximum variation in close- ness centrality scores possible in a network of the same size (Toral et al., 2009b). It has been found that closeness centralization has a positive correlation with coordination mechanism on OSS projects (Pereira & Soares, 2007; Feczak & Hossain, 2011). Betweenness centrality: it is a measure of centrality that rests on the idea that a person is more central if he or she is more important as an intermediary in the communication network (Nooy et al., 2005). The centrality of a node depends on the extent to which this node is needed as a link to facilitate the connection of nodes within the network. If a geodesic is defined as the shortest path between two nodes, the betweenness centrality of a vertex is the proportion of all geodesics between pairs of other vertices that include this vertex, and betweenness centralization of the network is the vari- ation in the betweenness centrality of vertices divided by the max- imum variation in betweenness centrality scores possible in a network of the same size. It has been used to study the hierarchy and centralization in free and open source software team commu- nications (Crowston & Howison, 2006). Brokerage roles: A broker is a middle node in a directed triad (a set of three vertices and the lines among them). Different types of brokerage roles can be distinguished considering mediation be- tween members of the same or different groups. In this context of OSS communities, these groups are given by active and periph- eral members. Therefore, two possibilities of mediation can be con- sidered as shown in Fig. 1. The role of knowledge brokers has been highlighted in mailing lists as community facilitators, helping answer those questions knowledge seekers posted (Sowe et al., 2006). This role has also been assigned to the core team of the OSS project, acting as inter- mediary between expert software developers and peripheral users and helping OSS projects to engage in a discourse and co-learning experience with their user communities (Toral et al., 2010). Clustering coefficient: It measures whether first degree neigh- bours of a particular vertex interact with each other. Basically, clustering coefficient is a measure of local cohesiveness through the neighbour interactions of a vertex (Durugbo, 2012). The clus- tering coefficient of a vertex is defined as the ratio of the number Broker Broker Active member Peripheral member Active member Peripheral member Fig. 1. Brokerage roles. of links to the total possible number of links among its neighbours, and the clustering coefficient of a social network is the average of all the clustering coefficients of the vertices. It has been used to characterize the small-world phenomena in networks (Xu et al., 2006). Besides, highly clustered networks provide better informa- tion propagation (Gao & Madey, 2007). Proximity prestige: It is defined in terms of the output domain of a node, which is the number of nodes for which there is a path to that node. Following Nooy et al. (2005) definition, the proximity prestige of a node is the proportion of all nodes in the network (excluding itself) that are in its output domain divided by the mean distance from all nodes in its output domain. Therefore, a zero va- lue of the proximity prestige means that this node is isolated. 3. Formulation of the problem and proposed framework The problem consists of finding a set of SNA indicators able to explain the different structural patterns of OSS communities. Fac- tor Analysis is a multivariate statistical technique usually em- ployed for the identification of latent dimensions or factors on a dataset. These factors are not directly observable and segment the dataset into relatively homogeneous segments (Rencher, 2002). It is assumed each variable is dependent on a linear combi- nation of the common factors, and the coefficients are known as loadings (Toral & Martínez Torres, 2009c). Mathematically, the fac- tor analysis model expresses each variable as a linear combination of underlying common factors f1, f2 , . . . , fm, with an accompanying error term to account for that part of the variable that is unique (not in common with the other variables). For y1, y2, . . . , yp in any observation vector y, the model is as follows: y1 � l1 ¼ k11f1 þ k12 f2 þ���þ k1mfm þ e1 y2 � l2 ¼ k21f1 þ k22 f2 þ���þ k2mfm þ e2 . . . yp � lp ¼ kp1f1 þ kp2 f2 þ���þ kpmfm þ ep ð3Þ Model (3) can be written in matrix notation as in Eq. (4), where K is the factor loadings matrix. y � l ¼ Kf þ e ð4Þ The coefficients kij are called loadings and serve as weights, showing how each yi individually depends on the underlying fac- tors (Lee and Lee, 2011). With appropriate assumptions, kij indi- cates the importance of the jth factor fj to the ith variable yi and can be used in interpretation of fj. It is expected the loadings will partition the variables into groups corresponding to factors. Ideally, the number of factors m should be substantially smaller than the problem dimension p; otherwise we have not achieved a parsimonious description of the variables as functions of a few underlying factors. In the case of exploratory factor analysis, the lack of theoretical background causes that the number of factors to be extracted is a priori unknown. Therefore, factors must be se- lected attending to the homogeneity of their indicators. The main problem of using factor analysis when considering a large set of indicators is that the final result is conditioned by the number of variables included in the analysis. The described model try to fit the data in set of factors, and the inclusion of non appropriate variables may distort the obtained latent factors. However, it is not easy to decide which variables are or not appro- priate, above all in those situations where the theoretical back- ground cannot guide this process. This paper proposes a computation framework where the selection of indicators is per- formed using Genetic Algorithms (GA). Each element (or chromo- some using typical GA notation) of the population considered by GA represent a subset of all the possible SNA indicators. Fig. 2 shows a brief scheme of the proposed framework. Sergio Cuadro de texto Fig. 2. Proposed framework. Fig. 3. Outline of the genetic algorithm implementation. M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 13185 The developed algorithm consists of two loops. The inner loop consists of the evaluation of the fitness function for each element of the population using factor analysis. The outer loop generates a new generation using genetic operators like reproduction, cross- over and mutation, using a selection based on the individual fitness of each element of the population. The framework stops working when a selected stopping criterion is reached. 4. Genetic algorithm implementation Genetic Algorithms are a family of computational models in- spired by evolution (Holland, 1975; Goldberg, 1989). These algo- rithms encode a potential solution to specific problem on a simple chromosome-like data structure and apply genetic opera- tors to these structures in order to preserve critical information (Martínez-Torres & Toral-Marín, 2010). An initial population Pi composed of Nt chromosomes is considered. Goldberg (1989) stud- ied the optimum number of chromosomes for a population accord- ing to the chromosome’s length. His main conclusion was that the optimum population’s size value gets higher as the chromosome’s length increases. This initial population is generated randomly in order to preserve the diversity in the population and the fitness function is calculated to evaluate the goodness of each chromo- some. The mechanism for generating the subsequent generations is based on the selection scheme from (l + k) evolution strategy (Michalewicz, 1996; Reina, Toral, Johnson, & Barrero, 2012). The l best chromosomes are included directly in the next generation. The crossover and mutation operations are responsible of generat- ing k chromosomes of a new population. The crossover consists of using two members of a population Pj to generate two new mem- bers of the next population Pj+1 by crossing their genetic informa- tion. The new chromosomes contain genetic information from the predecessors. The purpose of mutation is to change the genetic information of a chromosome included in Pj to generate a new chromosome of Pj+1. Fig. 3 illustrates the implemented genetic algorithm. 4.1. Chromosome encoding A chromosome Ci represent the subset of indicators that will be considered to perform factor analysis. The total number of Sergio Cuadro de texto 13186 M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 indicators extracted for the OSS communities according to the de- scribed SNA features is 60 (see appendix). Therefore, each chromo- some is a binary string of length 60 where a value of 1 means that the corresponding variable is part of the considered subset of indicators and a value of 0 means that it is excluded from this list. Notice that the space of possible solutions is formed by 260 = 1.1529e+018 possibilities. That means that we should per- form 260 different factor analyses to completely explore the space of possible solutions. In this kind of problems, GA can perform a guided search of the optimum solution with lower computational cost than exploring one by one all the possibilities. The representa- tion of chromosomes as binary strings has the advantage that this chromosomal encoding is complete and valid. Complete means that the whole space of possible solutions can be represented and valid means that all of them can be computed. The composi- tion of a chromosome is shown in Fig. 4. 4.2. Evaluation function The fitness function quantifies the suitability of each chromo- some as a solution. Genetic operators make selections based on individual fitness. That means that chromosomes with high fitness value have more chance of being selected, passing their genetic material (via reproduction, crossover or mutation) to the next gen- eration. As a result, the fitness function provides the pressure for evolution towards a new generation with chromosomes of higher fitness than the previous ones. In this case, the fitness function should measure how well factor analysis can identify latent factors. However, its capacity to perform such identification depends on the following parameters: � Explained variance: It refers to the percentage of the total sam- ple variance explained by the considered factors. � Correlations between variables: It is the average of the sum of the squared correlation coefficients between indicators. Consid- ered indicators must be correlated as the factor analysis is based on the interrelationships among variables. � Interpretability of factors: A factor is well defined if it is explained by at least three variables (Rencher, 2002). That is to say, at least three different factor loadings should be maxi- mized for each considered factor. As a result, the fitness function requires to be defined as a mul- ti-objective fitness function considering the aforementioned parameters. F ¼ c1 Var þ c2 1 n Xk i¼1 r2i þ c3 Interp ð5Þ Explained variance and correlations among variables exert opposite effects on the evolution of GA. Explained variance makes the GA to evolve towards a minimum number of indicators, as it is easiest to explain the variance of the dataset whenever a lower number of variables are considered. As a difference, correlations among variables makes the GA to search for a higher number of variables, as this parameter is maximized by including as many variables as possible. However, the third parameter, interpretabil- ity of factors, is the most important one, because this parameter guarantees factors are well defined. 1 I1 0 I2 1 I3 0 I4 1 I5 0 I6 . . . 1 I59 0 I60 Fig. 4. Chromosome’s composition. C1, C2, and C3 coefficients in Eq. (5) are used to adjust the rela- tive importance of the three parts of the fitness function. Obvi- ously, the range of them is [0,1], with the restriction of C1 + C2 + C3 = 1. 4.3. Stopping criteria The population’s average fitness function has been chosen as the stop criterion of the genetic algorithm. If Pav,j denotes the pop- ulation’s average fitness function, the stopping criterion can be for- mulated as: Sc ! Pav;jþ1 � Pav;j ð6Þ 4.4. Procedure of transition The procedure used to generate a new population Pj+1 from the previous population Pj is as follows: � The best 20% chromosomes are copied from Pj to Pj+1. This ensures that the best individuals of each population will be included in the next generation. As a consequence, the likeli- hood of using a good chromosome for reproduction operations gets higher. � The 80% of the new chromosomes are generated by using cross- over (75%) and mutation (5%) operations. This aims to favour the diversity of the chromosomes. Pc denotes the probability of a chromosome Ci to take part in a crossover operation and it is calculated as: PCi ¼ ffðCiÞPn i¼1 ffðCiÞ ð7Þ The term ff(Ci) stands for the evaluation of the fitness function for the chromosome Ci. Consequently, the best chromosomes are more likely to be selected. The crossover operation is illustrated in Fig. 5. A one-point crossover operation has been implemented. The point of cross is denoted by pk, where 0 6 k 6 l, and l is the size of the chromosome. The value of k is randomly chosen for each crossover operation. This point divides each chromosome into two parts RGj and LGj. The two new chromosomes are then ob- tained swapping LGj,1 by LGj,2. Similarly, Pm is the probability of a chromosome i to take part in a mutation operation. The purpose of mutation is to make small changes in the chromosomes. These changes consist of modifying one chromosome’s bit. According to De Jong (1975) we have calcu- lated Pm as l �1. The mutation operation is illustrated in Fig. 6. The position of the mutated bit is denoted by pm, where 0 6 m 6 l. The value of pm is randomly chosen for each mutation operation. 5. Results The proposed approach has been applied to twelve virtual communities listed in Table 1. They correspond to Linux Debian ports to different processor architectures. The Debian Project is an association of individuals who have made common cause to create a free operating system called Debian GNU/Linux, or simply Debian for short (Robles, Gonzalez-Barahona, & Michlmayr, 2005; Mateos-Garcia & Steinmueller, 2008). Each community was analyzed from the year in which each community started its activity to 2010. For each year and commu- nity, a social network based on interactions among participants has been extracted. As a result, a total of 134 social networks have been analyzed, extracting the set of data detailed in the appendix. Sergio Cuadro de texto Fig. 5. Crossover operation. Fig. 6. Mutation operation. M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 13187 Once this large dataset has been obtained, GA has been applied to obtain an optimum subset of indicators able to identify OSS commu- nities profiles according to their topological participation structure. A population’s size of 10000 chromosomes will be considered (Alander, 1992). That means that the 2000 best chromosomes are di- rectly included in the next population, 7500 chromosomes are gen- erated by the crossover operation and 500 chromosomes are Table 1 Analyzed Linux Debian communities. URL Debian port to m68k (Debian-68k) http://lists.debian.org/debian-68k/ Debian port to ARM (Debian-ARM) http://lists.debian.org/debian-arm/ Debian port to Intel IA-64 (Debian-ia64) http://lists.debian.org/debian-ia64/ Debian port to Alpha (Debian-alpha) http://lists.debian.org/debian-alpha/ Debian port to AMD64 (Debian-amd) http://lists.debian.org/debian-amd64/ Debian port to BSD (Debian- bsd) http://lists.debian.org/debian-bsd/ Debian port to HPPA (Debian-hppa) http://lists.debian.org/debian-hppa/ Debian port to Hurd (Debian-hurd) http://lists.debian.org/debian-hurd/ Debian port to MIPS (Debian-mips) http://lists.debian.org/debian-mips/ Debian port to PowerPC (Debian-ppc) http://lists.debian.org/debian-powerpc/ Debian port to SPARC (Debian-s390) http://lists.debian.org/debian-s390/ Debian port to SPARC (Debian-sparc) http://lists.debian.org/debian-sparc/ generated by the mutation operation. Factor analysis is used for eval- uating each chromosome. The cost function follows the general structure defined in Eq. (5). The coefficients C1, C2 and C3 represent the relative weigh of each part of the fitness function. GA has been run for different values of the three parameters, obtaining the results shown in Table 2. It can be observed that interpretability only reaches a value of 1 when C3 is overweighed, while the explained variance varies depending on the value of C1. That is why the set of coefficients with the values C1 = 0.1, C2 = 0.1 and C3 = 0.8 has been chosen (This set of value is shown in italics in Table 2). An interpret- ability of factors equal to 1 guarantees that all the latent factors are interpretable. Using this set of values, GA converged after 30 gener- ations, with an explained variance of 70.93%, and 23 indicators grouped in four factors. Time required by genetic algorithm execu- tion was 2873.6 s (47.89 min). This value is much smaller than the alternative option of exploring the whole solution space. According to the chosen encoding, the size of a chromosome is l = 60 bits, so the space of possible solutions is 2l = 260 = 1.1529e+018. The idea of finding the optimum solution exploring all the possibilities is unat- tainable. The time necessary to carry out a single factor analysis is 12.9 ms. That means it would take more than 470 million years to ex- plore the whole space of possible solutions. The genetic algorithm implementation is able to speed up the search of an optimal solution. Description Period Motorola 68k port of Debian GNU/Linux. Debian currently runs on the 68020, 68030, 68040 and 68060 processors 1998–2010 ARM port for Debian GNU/Linux. Debian fully supports a port to little-endian ARM 1999–2010 Discussions on the intel IA64 (aka Itanium, Merced) port of Debian GNU/Linux 2001–2010 The purpose of this project is to assist developers and others interested with the ongoing project to port the Debian distribution of Linux to the Alpha family of processors 1998–2010 Porting Debian to AMD x86-64 architecture 2004–2010 This is a port of the Debian operating system, complete with apt, dpkg, and GNU userland, to the NetBSD kernel 2001–2010 This is a port to Hewlett-Packard’s PA-RISC architecture 2001–2010 The GNU Hurd is a totally new operating system being put together by the GNU group 1999–2010 MIPS port of Debian GNU/Linux, able to run at both endiannesses 1999–2010 PowerPC port of Debian GNU/Linux. The PowerPC architecture allows both 64-bit and 32-bit implementations 1999–2010 Discussions on the IBM S/390 port of Debian GNU/Linux 2001–2010 This port runs on the Sun SPARCstation series of workstations, as well as some of their successors in the sun4 architectures 1998–2010 Sergio Cuadro de texto Table 2 GA results for different values of coefficients c1, c2 and c3. Parameters (c1/c2/c3) Explained variance Indicators Factor number Interpretability 1.00/0.00/0.00 77.73 20 7 0.00 0.80/0.10/0.10 79.80 49 12 0.00 0.60/0.20/0.20 78.03 46 10 0.30 0.40/0.30/0.30 73.87 53 10 0.18 0.20/0.40/0.40 75.70 49 10 0.30 0.00/0.50/0.50 72.40 48 9 0.44 0.00/0.00/1.00 59.65 14 2 1.00 0.10/0.10/0.80 70.93 23 4 1.00 0.20/0.20/0.60 71.60 36 5 0.60 0.30/0.30/0.40 74.94 49 10 0.30 0.40/0.40/0.20 77.28 51 11 0.27 0.50/0.50/0.00 75.16 52 11 0.08 0.00/1.00/0.00 72.53 55 11 0.09 0.10/0.80/0.10 72.52 54 11 0.18 0.20/0.60/0.20 75.37 54 12 0.17 0.30/0.40/0.30 76.79 53 12 0.16 0.40/0.20/0.40 75.31 43 8 0.50 0.50/0.00/0.50 73.24 22 5 0.80 Fig. 7. Fitness distribution over 30 generations of the genetic algorithm. Table 3 Selected set of indicators. Description VAR02 Number of interactions VAR03 Number of repeated interactions VAR04 Density VAR06 The Zagreb group index VAR09 Free riders VAR10 Active members VAR11 Members responsible of more than 50% of contribution VAR17 Normalized size of output-domain (standard deviation) VAR18 Proximity prestige (average value) VAR26 Network betweenness Centralization VAR29 Number of vertices with betweenness centrality >0 VAR30 Egocentric density (average value) VAR31 Egocentric density (standard deviation) VAR33 Number of developed brokerage roles among active me VAR34 Number of vertices developing a brokerage role among VAR35 Number of developed brokerage roles among active me VAR41 Density VAR42 Average degree VAR45 Average out-degree VAR48 Normalized size of output-domain (average value) VAR50 Proximity prestige (average value) VAR53 Closeness centrality (average value) VAR59 Egocentric density (average value) 13188 M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 The evolution of the genetic clustering algorithm is detailed in Fig. 7. The initial population (generation 0) has a low fitness value, which indicates that the individuals of the population are far from the optimum. As the number of generations increase, the fitness of individuals within the population also increases, as the genetic algorithm is biased towards the survival of genetic material con- tained within the individuals with high fitness function values. The optimum subset of indicators provided by GA is listed in Ta- ble 3. In particular, the indicators description and the network over which they have been calculated are detailed. The results from factor analysis using the set of variables se- lected by the genetic algorithm are detailed in Table 4. Usually, a number of factors equal to the number of eigenvalues higher than 1 is selected (Rencher, 2002). Consequently, up to four latent fac- tors can be distinguished as result of factor analysis. The indicators associated to each factor are obtained from the factor loadings using a Varimax rotation. All the indicators associ- ated in this way with the same factor are hypothesized to share a common meaning that the analyst should discover. Table 5 shows which indicators are associated to each factor and their corre- sponding factor loadings. On the other hand, factor scores are used to categorize the ori- ginal sample of OSS communities, which can be approximated to one of the identified latent factors. Consequently, the original sam- ple of OSS communities can be categorized in four groups. An anal- ysis of variance (ANOVA) has been applied to the categorization of Network/subnetwork Complete network Complete network Complete network Complete network Complete network Complete network s Complete network Complete network Complete network Complete network Complete network Complete network Complete network mbers Complete network active members and free riders Complete network mbers and free riders Complete network Active members network Active members network Active members network Active members network Active members network Active members network Active members network Table 4 Explained variance of resulting factor analysis. Factor Eigenvalues Value Variance (%) Cumulative (%) 1 11.881 47.526 47.526 2 4.388 17.553 65.078 3 2.309 9.234 74.313 4 1.455 5.818 80.131 5 0.914 3.658 83.789 6 0.799 3.195 86.984 7 0.652 2.606 89.590 .. . .. . .. . .. . 23 0.000 0.001 100.000 Sergio Cuadro de texto Table 5 Identified factors. Description Loading F1 VAR02 Number of interactions 0.921 VAR03 Number of repeated interactions 0.900 VAR06 The Zagreb group index 0.836 VAR09 Free riders 0.800 VAR10 Active members 0.931 VAR11 Members responsible of more than 50% of contributions 0.703 VAR29 Number of vertices with betweenness centrality >0 0.945 VAR33 Number of developed brokerage roles among active members 0.929 VAR34 Number of vertices developing a brokerage role among active members and free riders 0.925 VAR35 Number of developed brokerage roles among active members and free riders 0.910 F2 VAR17 Normalized size of output-domain 0.854 VAR18 Proximity prestige (average value) 0.847 VAR26 Network Betweenness Centralization 0.764 VAR30 Egocentric density (average value) 0.881 VAR31 Egocentric density (standard deviation) 0.719 F3 VAR48 Normalized size of output-domain (average value) 0.717 VAR50 Proximity prestige (average value) 0.901 VAR53 Closeness centrality (average value) 0.899 VAR59 Egocentric density (average value) 0.740 F4 VAR04 Density (complete network) 0.705 VAR41 Density (active members netw.) 0.897 VAR42 Average degree 0.799 VAR45 Average out-degree 0.703 Table 6 Statistical significance of ANOVA. F Sig F Sig VAR02 41.837 0.000 VAR31 16.644 0.000 VAR03 36.386 0.002 VAR33 41.910 0.000 VAR04 7237 0.003 VAR34 86.377 0.000 VAR06 24.322 0.000 VAR35 46.100 0.000 VAR09 50.931 0.000 VAR41 8018 0.000 VAR10 72.496 0.000 VAR42 10.950 0.000 VAR11 23.912 0.000 VAR45 6489 0.000 VAR17 53.598 0.000 VAR48 26.019 0.000 VAR18 40.664 0.000 VAR50 16.445 0.306 VAR26 31.039 0.000 VAR53 14.434 0.000 VAR29 73.973 0.000 VAR59 7488 0.000 VAR30 37.410 0.000 0.000 Table 7 Mean values of selected indicators. F1 F2 F3 F4 VAR02 11096.3846 2264.8140 1935.4688 1092.2500 VAR03 8102.5385 1553.1860 1464.1250 900.3333 VAR04 0.0201 0.0558 0.0268 0.0663 VAR06 1.9473E7 1.9473E7 1.9473E7 1.9473E7 VAR09 353.6154 353.6154 353.6154 353.6154 VAR10 377.9231 377.9231 377.9231 377.9231 VAR11 8.1538 8.1538 8.1538 8.1538 VAR17 0.2353 0.2353 0.2353 0.2353 VAR18 0.0894 0.0894 0.0894 0.0894 VAR26 0.1048 0.1048 0.1048 0.1048 VAR29 198.3462 198.3462 198.3462 198.3462 VAR30 0.2611 0.2611 0.2611 0.2611 VAR31 0.3163 0.3163 0.3163 0.3163 VAR33 60910.6538 60910.6538 60910.6538 60910.6538 VAR34 55.8846 55.8846 55.8846 55.8846 VAR35 4590.7692 4590.7692 4590.7692 4590.7692 VAR41 0.0742 0.0742 0.0742 0.0742 VAR42 383725.4850 383725.4850 383725.4850 383725.4850 VAR45 152903.9688 152903.9688 152903.9688 152903.9688 VAR48 0.7834 0.7834 0.7834 0.7834 VAR50 0.2969 0.2969 0.2969 0.2969 VAR53 0.2972 0.2972 0.2972 0.2972 VAR59 0.4730 0.4730 0.4730 0.4730 M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 13189 the original sample in the four groups obtained form factor analy- sis. The aim of this analysis consists of checking the null hypothesis of equal population means. Table 6 details the F statistic, the ratio of two different estimators of population variance, which appears together with its corresponding critical level or observed signifi- cance. The results is that the null hypotheses have been rejected in all the cases with a significance value below 0.05. That means the obtained categorization from factor analysis is well defined. Table 7 details the mean value of the considered 23 indicators per each of the distinguished groups. Using this information as well as the factor loadings of Table 5, the following websites struc- ture patterns can be distinguished: Factor 1: This factor considers high sized communities character- ized by a high number of interactions and a clear hierar- chy among its members. The core group is located on top of this hierarchy, and they are responsible of devel- oping a brokerage role among the rest of community members like active members and free riders. Factor 1 includes those communities with the highest number of members belonging to the core group. Fig. 8. Interpretation of identified factors. Sergio Cuadro de texto Size Hierarchy LargeSmall High Low Debian-68k Debian-ARMDebian-ia64 Debian-alpha Debian-amd64 Debian-BSD Debian-hppa Debian-hurd Debian-mips Debian-ppc Debian-s390 Debian-sparc Fig. 9. Patterns of behaviour of Debian Linux ports communities. 13190 M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 Factor 2: Factor 2 refers to those communities with a high local connectivity. Communities following this pattern exhi- bit the highest average value of egocentric density and proximity prestige, which means their nodes are highly connected with their neighbours. Although these com- munities are not so high as those of Factor 1, they also exhibit an important size with a high number of interac- tions. However, the level of hierarchy is considerable lower. The role of the core group is shaded by the high activity of the rest of the community. In fact, F2 commu- nities are the ones with the lowest number of free riders. Factor 3: Factor 3 corresponds to small and centralized networks, where active members exhibit a high involvement and the core group also performs an important brokerage role. This factor includes communities with a strong hierarchy among their members but with a flatter struc- ture compared to F1 communities. Factor 4: Factor 4 considers the smallest communities but with a high density, which means that active members are highly interconnected. Despite of being smaller than F3 communities, they exhibit a higher average degree value. As a difference, the activity of the core group is shad- owed by the active members of the community, so the hierarchy is much weaker than in F3 communities. Obtained factors can be interpreted in terms of size and hierar- chy of communities, as represented in Fig. 8. This figure graphically visualizes the four identified patterns of behaviour of OSS commu- nities. Size represents how many users the OSS project is able to at- tract. Therefore, it can be interpreted as a measurement of how successful the underlying software is. Hierarchy refers to the inter- nal organization of the community. A high hierarchy means that the core group exerts an meaning- ful influence over the rest of the community while a low hierarchy means there is not a dominant group. Fig. 9 details the position of the considered Linux Debian communities in terms of size and hierarchy. The general trend is to be located on the right part of this figure as communities try to attract as many members as pos- sible. As a difference, they follow different strategies regarding their internal hierarchy level. In general, they tend to maintain a certain hierarchy and only one community has a clear low hierar- chy level. One of the main implications of this results is that size and hier- archy are independent dimensions, so it is not incompatible a high size with a strong hierarchy. In fact, the growth in size impels the creation of hierarchical layers. In general, projects gain stability as more or less formal hierarchies emerge. Hierarchy mainly depends on the activity of the core group team. It is their responsibility to canalize the discussion, to provide solutions o alternatives to the posted problems and to facilitate the incorporation of qualified ac- tive members as part of this core group. Hierarchy also causes a rise in the entry barriers to the core group. Outsiders only can ac- quire authority inside the project through the legitimate peripheral participation procedure. Therefore, hierarchy guarantees that the project is under the control of the core group of developers. 6. Conclusions This paper describes a procedure for the extraction of the pat- terns of behaviour of open source communities using a genetic search over a large set of Social Network Analysis indicators. As a difference to previous studies, only focused on small set of SNA fea- tures, the proposed evolutionary computation technique combined with factor analysis allows to consider all the possible metrics able to characterize interactions among community members. Obtained results show four patterns of behaviour that have been extracted as latent factors from the whole data set. These patterns can be in turn classified in terms of size and hierarchy of communities. Analyzed communities can be approached to a point in a bidimen- sional space described by these two dimensions. In general, OSS communities tend to gain as many members as possible, and their internal organization facilitates a certain hierarchy to emerge. Hierarchy can be understood as a stabilizing factor of the commu- nity that assigns the authority to a reduced core group member. Access to the inner circle of the community can only be achieved through a process of participation and interaction with the rest of the community. Sergio Cuadro de texto M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 13191 Appendix A Indicators Description Network VAR01 Number of community members Complete VAR02 Number of interactions Complete VAR03 Number of repeated interactions Complete VAR04 Density Complete VAR05 Average degree Complete VAR06 The Zagreb group index Complete VAR07 The Randic connectivity index Complete VAR08 Index of relinking Complete VAR09 Free riders Complete VAR10 Active members Complete VAR11 Members responsible of more than 50% of contributions Complete VAR12 Average out-degree Complete VAR13 Standard deviation out-degree Complete VAR14 Number of components with more than 1 user Complete VAR15 Per cent of users included in components Complete VAR16 Normalized size of output-domain (average value) Complete VAR17 Normalized size of output-domain (standard deviation) Complete VAR18 Proximity prestige (average value) Complete VAR19 Proximity prestige (standard deviation) Complete VAR20 Maximum k-core (value of k) Complete VAR21 Number of users included in maximum k-core Complete VAR22 Number of k-cores with more than 1 user (k > 0) Complete VAR23 Per cent if users included in k-cores (k > 0) Complete VAR24 Closeness centrality (average value) Complete VAR25 Closeness centrality (standard deviation) Complete VAR26 Network Betweenness Centralization Complete VAR27 Average value of vertices betweeness centrality Complete VAR28 Standard deviation of vertices betweeness centrality Complete VAR29 Number of vertices with betweenness centrality > 0 Complete VAR30 Egocentric density (average value) Complete VAR31 Egocentric density (standard deviation) Complete VAR32 Number of vertices developing a brokerage role among active members Complete VAR33 Number of developed brokerage roles among active members Complete VAR34 Number of vertices developing a brokerage role among active members and free riders Complete VAR35 Number of developed brokerage roles among active members and free riders Complete VAR36 Average brokerage role of core group with active members Complete VAR37 Average brokerage role of core group with free riders Complete VAR38 Average number of core group Complete Appendix A (continued) Indicators Description Network neighbours in the first level VAR39 Average number of core group neighbours in all levels Complete VAR40 Average number of levels in which users are accessed by the core group Complete VAR41 Density Active members VAR42 Average degree Active members VAR43 The Randic connectivity index Active members VAR44 Index of relinking Active members VAR45 Average out-degree Active members VAR46 Standard deviation out-degree Active members VAR47 Per cent of users included in components Active members VAR48 Normalized size of output-domain (average value) Active members VAR49 Normalized size of output-domain (standard deviation) Active members VA050 Proximity prestige (average value) Active members VAR51 Proximity prestige (standard deviation) Active members VAR52 Per cent if users included in k-cores (k > 0) Active members VAR53 Closeness centrality (average value) Active members VAR54 Closeness centrality (standard deviation) Active members VAR55 Number of vertices with closeness centrality > 0 Active members VAR56 Network betweenness centralization Active members VAR57 Average value of vertices betweeness centrality Active members VAR58 Standard deviation of vertices betweeness centrality Active members VAR59 Egocentric density (average value) Active members VAR60 Egocentric density (standard deviation) Active members References Alander, J. T. (1992). On optimal population size of genetic algorithm. In Proceedings CompEuro 1992. Computer systems and software engineering, 6th annual European computer conference (pp 65–70). Barcellini, F., Détienne, F., & Burkhardt, J.-M. (2009). Participation in online interaction spaces: Design-use mediation in an open source software community. International Journal of Industrial Ergonomics, 39(3), 533–540. Crowston, K., & Howison, J. (2005). The social structure of free and open source software development. First Monday, 10(2). Crowston, K., & Howison, J. (2006). Hierarchy and centralization in free and open source software team communications. Knowledge, Technology, & Policy, 18(4), 65–85. De Jong, K. A. (1975). An analysis of behaviour of a class of genetic adaptive systems. Thesis. University of Michigan. Deek, F. P., & McHugh, J. A. M. (2008). Open source: Technology and policy. NY: Cambridge University Press. Sergio Cuadro de texto 13192 M.R. Martínez-Torres / Expert Systems with Applications 39 (2012) 13182–13192 Devillers, J., & Balaban, A. T. (1999). Topological indices and related descriptors in QSAR and QSPR. The Netherlands: Gordon and Breach Science Publishers. Durugbo, C. (2012). Modelling user participation in organisations as networks. Expert Systems with Applications, 39(10), 9230–9245. Feczak, S., & Hossain, L. (2011). Exploring computer supported collaborative coordination through social networks. The Journal of High Technology Management Research, 22(2), 121–140. Feller, J., & Fitzgerald, B. (2002). Understanding open source software development. London, UK: Addison-Wesley. Freeman, L. C. (2004). The development of social network analysis: A study in the sociology of science. Vancouver, Canada: Empirical Press. Gao, Y., & Madey, G. (2007). Network Analysis of the SourceForge.net Community. In J. Feller, B. Fitzgerald, W. Scacchi, A. Sillitti (Eds.), IFIP International Federation for Information Processing, vol. 234, Open Source Development, Adoption and Innovation, Boston, Springer, pp. 187–200. Goldberg, D. E. (1989). Genetic algorithm in search, optimization and machine learning. Reading, MA: Addison-Wesley. Hinds, D. & Lee, R.M., 2008. Social network structure as a critical success condition for virtual communities. In Proceedings of the 41st Hawaii international conference on system sciences (pp. 323–333). Holland, J. (1975). Adaptation in Natural and Artificial Systems. Ann Arbor, MI: University of Michigan Press. Hossain, L., Wu., A., & Chung, K. (2006). Actor centrality correlates to project based coordination. In: P. Hinds, D. Martin (Eds), Proceedings of the CSCW’06 conference, Banff, Canada (pp. 363–372). Hossain, L., & Zhu, D. (2009). Social networks and coordination performance of distributed software development teams. The Journal of High Technology Management Research, 20(1), 52–61. Kuk, G. (2006). Strategic Interaction and Knowledge Sharing in the KDE Developer Mailing List. Management Science, 52(7), 1031–1042. Kwon, D., Oh, W., & Jeon, S. (2007). Broken Ties: The Impact of Organizational Restructuring on the Stability of Information-Processing Networks. Journal of Management Information Systems, 24(1), 201–231. Lee, Y., & Lee, H. (2011). Application of factor analysis for service R&D classification: A case study on the Korean ICT industry. Expert Systems with Applications, 3(3), 2119–2124. Long, Y., & Siau, K. (2007). Social network structures in open source software development teams. Journal of Database Management, 18(2), 25–40. Martínez-Torres, M. R., & Toral-Marín, S. L. (2010). Strategic group identification using evolutionary computation. Expert Systems with Applications, 37(7), 4948–4954. Mateos-Garcia, J., & Steinmueller, W. E. (2008). The institutions of open source software: Examining the Debian community. Information Economics and Policy, 20, 333–344. Michalewicz, Z. (1996). Genetic algorithm + data structures = evolution programs (3rd ed.). Berlin, Germany: Springer-Verlag. Mockus, A., Fielding, T., & Herbsleb, D. (2002). Two case studies of open source software development: Apache and Mozilla. ACM Transactions on Software Engineering and Methodology, 11(3), 309–346. Nooy, W., Mrvar, A., & Batagelj, V. (2005). Exploratory network analysis with Pajek. New York: Cambridge University Press. Okoli, C., & Oh, W. (2007). Investigating recognition-based performance in an open content community: A social capital perspective. Information & Management, 44(3), 240–252. Panchal, J. H. (2009), Co-evolution of products and communities in mass- collaborative product development – a computational exploration. In Proceedings of international conference on engineering design (ICED’09) (p. 147). Pereira, C. S., & Soares, A. L. (2007). Improving the quality of collaboration requirements for information management through social networks analysis. International Journal of Information Management, 27(2), 86–103. Preece, J. (2001). Sociability and usability in online communities: determining and measuring success. Behaviour & Information Technology, 20(5), 347–356. Reina, D. G., Toral, S. L., Johnson, P., & Barrero, F. (2012). An evolutionary computation approach for designing mobile ad hoc networks. Expert Systems with Applications, 39(8), 6838–6845. Robles, G., Gonzalez-Barahona, J. M., & Michlmayr, M. (2005). Evolution of volunteer participation in libre software projects: Evidence from Debian. In Proceedings of the first international conference on open source systems, Genova (pp. 100–107). Rencher, A. C. (2002). Methods of Multivariate Analysis. Wiley Series in Probability and Statistics (2nd ed.). Berlin: Springer. Singh, P. V. (2010). The small-world effect: The influence of macro-level properties of developer collaboration networks on open-source project success. ACM Transactions on Software Engineering and Methodology, 20(2), 1–27. Sowe, S., Stamelos, I., & Angelis, L. (2006). Identifying knowledge brokers that yield software engineering knowledge in OSS projects. Information and Software Technology, 48(11), 1025–1033. Toral, S. L., Martínez-Torres, M. R., & Barrero, F. (2009a). Virtual communities as a resource for the development of OSS projects: the case of Linux ports to embedded processors. Behaviour and Information Technology, 28(5), 405–419. Toral, S. L., Martínez-Torres, M. R., Barrero, F., & Cortés, F. (2009b). An empirical study of the driving forces behind online communities. Internet Research, 19(4), 378–392. Toral, S. L., & Martínez Torres, M. R. (2009). International comparison of R&D investment by European, US and Japanese Companies. International Journal of Technology Management, 49(1/2/3), 107–122. Toral, S. L., Martínez Torres, M. R., & Barrero, F. (2010). Analysis of virtual communities supporting OSS projects using social network analysis. Information and Software Technology, 52(3), 296–303. Valverde, S., Theraulaz, G., Gautrais, J., Fourcassie, V., & Sole, R. V. (2006). Self- organization patterns in wasp and open source communities. IEEE Intelligent Systems, 21(2), 36–40. West, J., & O’mahony, S. (2008). The role of participation architecture in growing sponsored open source communities. Industry & Innovation, 15(2), 145–168. Xu, J., Gao, Y., Christley, S. & Madey, G. (2005). A topological analysis of the open source software development community. In Proceedings of the 38th annual Hawaii international conference on system sciences. HICSS ‘05 (pp. 188–198). Xu, J., Christley, S., & Madey, G. (2006). Application of social network analysis to the study of open source software. In Jürgen Bitzer & Philipp J. H. Schröder (Eds.), The economics of open source software development. Elsevier Press. Ye, Y., Nakakoji, K., et al. (2005). The co-evolution of systems and communities in free and open source software development. Free/open source software development. S. Koch. Hershey, PA, Idea Group Inc. (IGI) (pp. 59–82). Sergio Cuadro de texto A genetic search of patterns of behaviour in OSS communities 1 Introduction 2 Related work 3 Formulation of the problem and proposed framework 4 Genetic algorithm implementation 4.1 Chromosome encoding 4.2 Evaluation function 4.3 Stopping criteria 4.4 Procedure of transition 5 Results 6 Conclusions Appendix A References 
work_pkt6ad356rbahib4cigxolwqrm	----	PII: 0950-7051(93)90039-V Towards competent information acquisition interactions between an expert system and its user E T Keravnou, J Washbrook* and F Dams* Expert systems are consultative, highly interactive systems, and hence the quality of interaction between system and user is important for the acceptability of the system. Acquisition o f data about the problem in hand is a central feature of the interaction between system and user, and it makes a major contribution to the user's perception of the system. It is there- fore crucial that the acquisition of this data (both as individual data items and as a sequence of data inputs) is perceived by the user as competent. This paper identifies central, domain-independent, design goals for the information-acquisition interactions between an expert system and its user, including mixed-initiative interac- tion, flexibility in user input, and system competence in query- ing the user. The paper discusses the realisation of these goals in a diagnostic expert system, Skeletal Dysplasias Diagnos- tician, through an explicit data model which allows for the representation of data with temporal and spatial (locality) aspects and the decide-status function which operates on the data model. Keywords: information-acquisition interactions, intelligent data handling, medical diagnostic expert systems, competent expert systems, skeletal dysplnsins, background knowledge E x p e r t systems, especially those t h a t dispense advice in critical d o m a i n s such as medicine, engineering an d finance, m u s t be perceived by their users as c o m p e t e n t if they are to be accepted. R e c o m m e n d i n g solutions fo r real p r o b l e m s which are subsequently p r o v e n in real life to be c o r r e c t is o f course indisputable evidence o f a Department of Computer Science, University of Cyprus, Kallipoleos 75, PO Box 537, Nieosia, Cyprus *Department of Computer Science, University College London, Gower Street, London WC1E 6BT, UK Paper received 7 January 1992. Revised paper received 7 September 1992. Accepted 8 October 1992 system's competence. H o w e v e r , fo r a system to progress f r o m the research l a b o r a t o r y into a real-life environ- ment, it m u s t n o t o n l y be able to solve test p roblems co rrect l y b u t also to interact intelligently with the user. It is t h r o u g h the dialogue with the system t h a t the user forms i m p o r t a n t first impressions. T h u s the i n t eraction between the system an d its user is o f prime i m p o r t a n c e if the system is to inspire confidence in p o t en t i al users whilst still in the l a b o r a t o r y , a n d to retain their confi- dence once it has progressed to a real-life setting. T h e principal spheres o f interaction are i n f o r m a t i o n acqui- sition a n d explanation, b o t h o f which m u st be carried o u t in a w ay which is acceptable to, an d tailored to the needs of, the p art i cu l ar user ~. Acquisition o f i n f o r m a t i o n a b o u t the p r o b l e m in h a n d m a y be t h r o u g h the system asking questions, o r the user volunteering i n f o r m a t i o n . A t the beginning o f a consultation, initial i n f o r m a t i o n is acquired t h r o u g h a p r e d e t e r m i n e d sequence o f general questions asked b y the system a n d items o f d a t a volun- teered b y the user. T h e initial i n f o r m a t i o n acquisition process is thus mixed-initiative, in t h a t b o t h the system an d the user m a y initiate a p a r t i c u l a r interaction. It is possible f o r an expert system t o require t h a t all p r o b l e m d a t a be entered at the beginning o f the consul- tation, w i t h o u t the necessity (or possibility) o f subse- q u e n t i n f o r m a t i o n acquisition. This, however, is based o n a simplified an d restrictive m o d el o f the co n su ltation process. A m o r e sophisticated m o d el w o u l d allow for system-initiated i n f o r m a t i o n acquisition at subsequent stages in the consultation. H o w e v e r it is arguable (see below) t h a t i n f o r m a t i o n acquisition should c o n t i n u e t h r o u g h o u t the consultation, a n d should be potentially mixed-initiative (i.e. b o t h the user a n d the system should be able to initiate i n f o r m a t i o n input) at all stages. In addition, t o be perceived as c o m p e t e n t , the system should take into a c c o u n t all relevant i n f o r m a t i o n pre- viously acquired b efo re asking questions. This m a y require the m a k i n g o f substantial chains o f inferences. An expert system which accurately models the reason- 0950-7051/931030141-16 © 1993 Butterworth-Heinemann Ltd Knowledge-Based Systems Volume 6 Number 3 September 1993 141 Towards competent information acquisition interactions: E T Keravnou et aL ing and knowledge o f d o m a i n experts will also provi d e a model o f their information-acquisition strategies. This can be used as the basis for system information-acqui- sition interactions. Such interactions are dynamically constructed, being driven by the derivation o f the solu- tionL T h e y can be analysed at two levels. At a high level the sequence is structured in terms o f the hierarchy o f f o r e g r o u n d problem-solving strategies that have been instantiated. At a low level each s u b g r o u p o f interactions c o r r e s p o n d i n g to a terminal problem-solving strategy can be analysed in terms o f b a c k g r o u n d auxiliary strate- gies. T h e f o r e g r o u n d strategies represent d o m a i n - e x p e r t problem-solving processes, whilst the b a c k g r o u n d strate- gies provide indirect b u t indispensable s u p p o r t to these processes by organising the p r o b l e m d a t a and subse- quent system-initiated i n f o r m a t i o n acquisition 3. F o re- g r o u n d problem-solving strategies m a y be domain-speci- fic, while b a c k g r o u n d data-handling strategies, which often represent common-sense knowledge and reason- ing, are usually domain-independent. O t h e r aspects o f an information-acquisition interac- tion model, such as those related to the specifics o f con- ducting a particular task, or technical aspects related to the design o f user interfaces, are outside the scope o f this paper. Although the w o r k presented here is in the context o f a diagnostic expert system, c o m p e t e n c e in the informati o n - acquisition interactions is relevant to every consultative expert system, diagnostic or other. This p a p e r discusses information-acquisition interactions in the context o f the diagnostic system S D D (Skeletal Dysplasias Diagnos- tician) 4. T h e design goals identified constitute essential aspects o f the overall information-acquisition interaction model o f this system s . T h e first section identifies central, domain-indepen- dent, design goals for the information-acquisition inter- actions between an expert system and its user. These design goals have been realised in the medical diagnostic system, Skeletal Dysplasias Diagnostician 4, t h r o u g h b a c k g r o u n d strategies based on an explicit d a t a model and the decide-status function which operates on the d a t a model. The second, third and f o u r t h sections give a detailed description o f the decide-status operation; some o f this material has been a d a p t e d f r o m Reference 6. T h e fifth section discusses h o w the d a t a model an d decide-status function are used to achieve the required design goals for the information-acquisition interactions between an expert system and its user. I N F O R M A T I O N - A C Q U I S I T I O N I N T E R A C T I O N S : D E S I G N G O A L S I m p o r t a n t design goals for the information-acquisition interactions between an expert system and its user are listed below. Achieving these goals is necessary if the system is to be deemed competent. • Mixed-initiative interaction • Flexibility in user-volunteered information: o T h e user must be allowed to use the entire d o m a i n vocabulary. T h e user must be able to revoke information. T h e system must detect a conflict in the user- volunteered i n f o r m a t i o n as soon as it occurs a n d resolve it with the user. • System competence in querying the user." 0 It must be evident to the user t h at the system has t ak en notice o f the i n f o r m a t i o n volunteered by the user. 0 Th e system questions must be focused a n d meth- odical. 0 T h e r e should be n o n - r e d u n d a n c y in system questions. 0 Th e system must detect and eliminate a n y redun- d an cy in the user information. M i x e d - i n i f i a t v e interaction Th e information-acquisition interaction between the system an d the user must be allowed to be mixed-initia- tive where the user wishes it to be so. Th e user must be able to volunteer i n f o r m a t i o n initially a n d be given the o p p o r t u n i t y to volunteer fu rt h er i n f o r m a t i o n at subse- q u en t stages. Similarly, the system m u st be able to request fu rt h er i n f o r m a t i o n by querying the user if the i n f o r m a t i o n available at a n y stage is n o t sufficient for the system to p ro ceed to the next stage, o r to m a k e firm recommendations. W h ere necessary, a p p r o p r i a t e guidance should be given to the user b o t h when volun- teering i n f o r m a t i o n an d responding to a system question 7. T h e system should o f course be able to p e r f o r m a consultation even i f the user does n o t wish to volunteer information. This can be initiated t h r o u g h a predeter- mined sequence o f questions. Th e ex p l an at i o n model o f an expert system is separate f r o m the information-acquisition interaction model s . As m e n t i o n e d above, the sequence o f questions raised by the system is intrinsically related to the system's reasoning 2. Traditionally, the central role o f the ex p l an at i o n model is to reveal this reasoning 9. However, the explanation model has a subsidiary role in relation t o i n fo rm ation- acquisition interactions. This role concerns individual items o f i n f o r m a t i o n rat h er t h an the system reasoning processes. T h e user needs to be able to ask, n o t only why the system is asking a particular question (i.e. h o w does it relate to the reasoning process), b u t also w h at the parti- cular question means. Also, the system needs to be able to recognise when i n f o r m a t i o n input by the user is either unclear o r meaningless in the co n t ex t o f the consultation, and m u st have the ability to q u e r y the user, asking for clarification. F l e x i b i l i t y in user-volunteered information F o r the user to take full ad v an t ag e o f the facility to volunteer i n fo rm at i o n , the system m u st allow the user flexibility in the means by which i n f o r m a t i o n is entered. 142 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou e t al. Flexible vocabulary A menu-driven interface facilitates the entry of user information and ensures that every piece of information is meaningful to the system. However the user may also wish to volunteer information 'freely', without having to go through a menu, especially if the user has domain expertise and does not need guidance in deciding what information to enter and how to word it. This facility is also useful in cases where the domain vocabulary requires a complex network of menus for covering all the possible expressions. The flexibility to enter information directly does not necessarily imply a completely unconstrained natural language interface (which is anyway beyond the capabili- ties of the current technology), but rather the use of the entire domain vocabulary (which will undoubtedly include redundant terminology) albeit within specific syntactical constraints which should be simple and con- venient. The system should be capable of dealing with redun- dant terminology by correlating different expressions which have the same meaning. Internally the system will probably use a non-redundant set of terms, into which it will translate user terms. It is important, though, that when the system interacts with the user it either uses the user's terms, or makes it clear that it has translated the user's terms. Conflict detection and resolution and retraction Another aspect of flexibility in user interaction is for the system to be able to detect and resolve conflicting infor- mation input, and for the user to be able to retract previously volunteered information. Conflicting information input may occur if the user information summarises lower-level data (e.g. sensor readings, images, direct observations etc.). The inherent uncertainty of this lower-level data can lead to misinter- pretations. The user may wish to enter possibly conflicting data, provided that a facility to retract previously volunteered information is available. The system should be able to detect and reveal an inconsistency as soon as it occurs and should attempt to resolve it by consulting the user. This should give the user more confidence in the system; the user who is aware of the inconsistency would be surprised if the system did not detect it, whilst the una- ware user would be favourably impressed. In both cases it would be evident to the user that the system is taking notice and 'making sense' of the information entered. The user should therefore be able to volunteer infor- mation either directly, in the form which he or she is used to in reporting findings, or through menus, depending on inclination and convenience, and should also be given the option to retract information subsequently. System competence in querying user Being unable to use the entire domain vocabulary or revoke information could reduce the usability of the system without necessarily affecting the confidence of the user in the system's judgement. However, that confi- dence will be seriously undermined if the system's ques- tioning is perceived as incompetent or unintelligent. For a system to be deemed as competent in its questioning, the following requirements must be satisfied. Use of previously given information The system must be able to show the user that inform- ation entered has been accepted and processed by the system, and is being used both in the diagnostic process and as a basis for any subsequent information acqui- sition. For this to happen, the generation of (partial) solu- tions (in a diagnostic domain this would be the gene- ration of hypotheses) must be driven by input infor- mation; thus subsequent questioning based on the currently entertained (partial) solutions would relate back to what the system had already been told, making the questioning focused and methodical. (The details of hypothesis generation and exploration refer to the fore- ground problem-solving strategies rather than the back- ground data-handling strategies and hence are outside the scope of this paper.) One significant way of showing the user that the system is utilising input information is for the system to ask questions which naturally follow on from the data volunteered, or to reveal inconsistencies in the user- volunteered information. Furthermore, if some infor- mation volunteered by the user needs clarification or triggers other questions, these clarifying or prompting questions should usually be raised by the system imme- diately after receiving this information; this constitutes intelligent behaviour. Non-redundant questions It is important that questions raised by the system should be non-redundant, in that their answers should not be deducible from information already entered. In response to a redundant question the user is justified in saying 'but I told you so already', or 'I've told you I don't know this', or 'I couldn't possibly know this at the present point in time'. If many redundant questions are asked, the user will get justifiably irritated. To ensure non- redundancy, the system should be able to make intelli- gent inferences on the given information; these should include the correlation of statements with equivalent meanings. Another aspect of intelligent inferencing on input information is the ability to recognise whether a particu- lar item of information would be unknown or not, based on information that the user has already specified as unknown, and, in some domains, based on the temporal context defined by the particular problem. This is par- ticularly important in most diagnostic domains. In the domain of skeletal dysplasias diagnosis, for instance, a feature may be unknown because the relevant radio- graphs are unavailable, or because the patient is too young for certain features to be observable. The system should also show that it understands dependencies between items of input information by removing redundant information in the light of new input. When new information volunteered by the user subsumes previously given information, possibly by being more specific, the system should make it evident to Knowledge-Based Systems Volume 6 Number 3 September 1993 143 Towards competent information acquisition interactions: E T the user that it is aware o f the r e d u n d a n c y and has eliminated it; this should inspire confidence and respect. Since the information volunteered by the user is likely to be c o m m u n i c a t e d back to the user by the system in different contexts (for example in run-time explanations, or in summing up final recommendations), there should be ample scope for the system to d e m o n s t r a t e to the user that it has this capability. Summary The main requirements for c o m p e t e n t information acquisition by an expert system are mixed-initiative interactions, flexibility in the entry o f user-volunteered information, a n d c o m p e t e n c e o n the p a r t o f the system in querying the user. These requirements can be achieved if the system has an explicit and flexible d a t a model which is used b y an intelligent reasoner to d r a w inferences from an actual set o f data. W e have implemented such a d a t a model and a reasoner, the decide-status function. D E C I D E - S T A T U S F U N C T I O N : A N I N T E L L I G E N T D A T A R E A S O N E R The decide-status function is an intelligent reasoner a b o u t problem-specific d a t a o r findings. It provides the platform for achieving the seemingly diverse goals dis- cussed above. Earlier w o r k on an a t e m p o r a l decide-sta- tus function has been reported in Reference 10. Decide-status operation The decide-status operation is first illustrated through an example from the d o m a i n o f skeletal dysplasias. Consider the following p r o b l e m definition. Domain-Model: synonyms((metaphyses flared) (long- bones dumbbell-shaped)) (carpal-centres poor-ossification) =:, (carpal-centres small) (carpal-centres small) =~ (epiphyses small) Problem findings: (long-bones dumbbell-shaped from-birth) (carpal-centres poor-ossification from- birth) (metaphyses irregular from-birth) Query finding." (and (metaphyses flared irregular [ta tb]) (epiphyses small [ta tb])) The d o m a i n m o d e l consists o f t w o s y n o n y m o u s findings and two implications. Each o f the p r o b l e m findings has the qualitative-temporal aspect 'from-birth' which trans- lates to the time interval [0 now]. The query finding is a c o m p o u n d finding consisting o f the t w o simple findings • metaphyses being b o t h flared a n d irregular during the time-interval [ta tb], Keravnou et a l • epiphyses being small during the same time interval. M e t a p h y s e s and epiphyses constitute the subjects o f the respective findings (finding subjects). The decide-status function reasons b a c k w a r d s f r o m the q u e r y finding. A goal tree, having the q u e r y finding at its r o o t is being (implicitly) constructed (see Figure 1). The nodes o f the goal-tree n a m e findings and the branches strategy applications. The leaf n o d e o f a 'suc- cessful' (i.e. n o t pruned) branch names a problem finding. The decide-status o p e r a t i o n is simply described as deciding the truth status o f a finding as true, false, or unknown, given a g r o u p o f findings which collectively hold in some context. A firm answer (true or false) returned b y the decide-status function for a finding quer- ied against the user observations can be directly corro- borated, e.g. through observation. The decide-status function operates on an explicit data model. Formalisation o f data model The d a t a model has a representational aspect, i.e. it provides a formal language for representing findings (data), and an inferential aspect, i.e. it provides a formal set o f relations between findings which form the channels through which intelligent inferences a b o u t an actual set o f findings can be drawn. The representational and infer- ential aspects o f the decide-status d a t a model are as follows. Representational aspect: grammar for expressing findings: (finding):: = (compound-finding)l(simple-finding)1 ( a t o m i c - f i n d i n g ) ( c o m p o u n d - f i n d i n g ) : : = (and (findings))l(or (findings)) ( f i n d i n g s ) : : = < f i n d i n g ) ( f i n d i n g s ) t ( f i n d i n g ) (simple-finding):: = ( ( f i n d i n g - s u b j e c t ) ( a t t r i b u t e - v a l u e s ) ( t i m e - i n t e r v a l ) ) ( a t t r i b u t e - v a l u e s ) : : = n i l t ( a t t r i b u t e - v a l u e ) ( a t t r i b u t e - v a l u e s ) ( f i n d i n g - s u b j e c t ) : : = symbol ( a t t r i b u t e - v a l u e ) : : = (locality-value)l (non-locality- v a l u e ) (locality-value):: = symbol (non-locality-value):: = symbol ( t i m e - i n t e r v a l ) : : = nil l(closed-interval)l(open-interval)l (open-from-left)1 ( o p e n - f r o m - r i g h t ) (closed-interval):: = [ ( b a s e ) (limit)] (open-interval):: = ( ( b a s e ) ( l i m i t ) ) ( o p e n - f r o m - l e f t ) : : = ( ( b a s e ) ( l i m i t ) ) ( o p e n - f r o m - r i g h t ) : : = [ ( b a s e ) (limit)] (ba se):: = integer (limit):: = integer ( a t o m i c - f i n d i n g ) : : = ((finding-subject) (simp-att- v a l s ) ( t i m e - i n t e r v a l ) ) (simp-att-vals):: =nill ( l o c a l i t y - v a l u e ) (non-locality- value )1 ( locality- value )1 ( no n-locality-value ) ( a t e m p o r a l - f i n d i n g ) is a ( f i n d i n g ) w i t h o u t temporal- aspect(s) 144 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou et al. (and (metaphyses flared irregular [ta tb]) (epiphyses small [ta tb]))? (metaphyses flared irregular (epiphyses small [ta tb]) [ta tb]) ~ COMPOSITION (metaphyses flared [ta tb]) MATCHER <SYNONYM> INFERENCER (carpal-centres small [ta tb]) l INFERENCER (metaphyses irregular [ta tb]) I MATCHER <DIRECT*> (metaphyses irregular [0 nowl) (carpal-centres poor-ossification [ta tb]) MATCHER <DIRECT*> ]1 (carpal-centres poor-ossification [0 now]) (long-bones dumbbell-shaped [ta tb]) I MATCHER <DIRECT*> (long-bones dumbbell-shaped [0 nowl) ]1 key: problem-f'mding *matches if [ta tb] is in [0 now] Figure 1 Decide-Status Knowledge-Based Systems Volume 6 Number 3 September 1993 145 Towards competent information acquisition interactions: E T Keravnou e t al. finding subject I ~ f'mdings for subject F y location and any time ) . . . . . . . . . . . . . . . . . . . . . . . s / / I ,'locality t t I t time Figure 2 Space of findings is three-dimensional space of find- ing subject, time and locality Inferential aspect: model entity relationships: is-a((finding-subject ) (finding-subject)) part-of((finding-subject) (finding-subject)) syn-subjects((finding-subject)(finding-subject)); symmetric relation negative-value((finding-subject) (attribute-value)) status-value((finding-subject)(attribute-value)) normal-subject((finding-subject)) abnormai-subjeet((finding-subject )) negative-finding((finding)) positive-finding( (finding) ) most-generai-positive-finding((finding)) status-finding((finding)) prompting-query((atemporal-finding) (atemporal- finding)) elarifying-query((atemporal-finding) (atemporal- finding)) synonyms-value((attribute-value) (attribute-value)); symmetric relation opposites-value((attribute-value)(attribute-values)); symmetric relation synonyms((atemporal-finding) (atemporal-finding)); symmetric relation impHes((atemporal-finding) (atemporal-finding)) alternatively expressed as: (atemporal-finding) => (atemporal-finding) when-to-ask((atemporal-finding) (earliest-time)) primary-trigger((finding) (hypothesis)) secondary-trigger((finding) (context-hypothesis) (alternative-hypothesis)) Decide-status operation The entire space of findings is partitioned along three dimensions, finding subject, time and locality, as shown in Figure 2. The decide-status operation is formally des- cribed as follows. Let Q be a query for some problem, expressed as holds(F)?, where F is a finding. The decide-status func- tion Decide-Status determines whether Q is decidable from the theory P constituting the problem, i.e. whether P ~ Q or P ~ ~ Q If Q is undecidable the reply unknown is returned. The problem theory P = domain-model U problem-findings, where the domain model is the instantiation o f the infer- ential aspect of the data model for the particular domain, and the problem fndings are the instantiations of the representational aspect o f the data model, i.e. assertions about which findings hold for the specific problem. The domain model therefore defines the finding subjects, attribute values and temporal aspects in the domain, as well as the relationships between these entities (e.g. status and negative values for finding subjects, synonymous finding subjects, synonymous findings, opposite attribute values). The meaning o f these relationships (given above) will be clarified when the decide-status operation is detailed. Decide-Status determines whether Q is true, false or unknown from the set of problem findings given the domain model, by attempting to construct a p r o o f for either Q or ~ Q. If this is not possible, Q is unknown. The p r o o f is constructed in a 'backwards-reasoning' or 'goal-driven' fashion where Q forms the goal, yielding a focused search strategy. Q ( ~ Q) follows from a set of problem findings if it is directly subsumed or implied by a subset o f the problem findings. The domain model is used to explicate relationships between the query finding F and the problem findings. Starting with the query Q at the root of the p r o o f tree, the goal-decomposition stra- tegy is applied until the goal is decomposed into simple subgoals, corresponding to simple queries (i.e. F is decomposed to simple findings). The domain model is used to generate subgoals from a given goal when that goal cannot be further decomposed. Subgoals generated from the domain model may be compound (e.g. the antecedents o f implications may be compound findings) which then need to be decomposed. Subgoals are themselves query findings. A tree branch terminates successfully if the query finding constituting its leaf node is directly determined from the problem findings, i.e. it directly matches, or conflicts, with problem findings. On the other hand, a tree branch is pruned if the simple, undetermined query finding corres- ponding to its leaf node cannot be mapped further through the domain model. A simple goal is determined if either itself or at least one o f the subgoals (which the goal can be mapped onto using the domain model) is directly determined from the problem findings, i.e. the subgoal leads to a successful tree branch. For a conjunc- tive compound finding to be determined as true all its component findings must be determined as true. For it to be determined as unknown at least one o f its component findings must be undetermined whilst the remaining component findings are determined as true. Lastly, deter- mining one o f the component findings as false suffices to 146 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou e t a/. determine the c o m p o u n d finding as false. A disjunctive c o m p o u n d finding is determined as true, false or u n k n o w n if at least one o f its c o m p o n e n t findings is determined as true, all o f its c o m p o n e n t findings are false, or at least one o f its c o m p o n e n t findings is u n k n o w n whilst the remaining are false, respectively. Thus if the current subgoal o f some simple goal is undetermined, the d o m a i n model is used to generate an alternative subgoal for t h a t goal. A t the top level there are two generic strategies corresponding to two alterna- tive derivation methods: a matcher a n d an inferencer. The m a t c h e r a n d inferencer strategies are described in detail in the third a n d f o u r t h sections, respectively. M A T C H E R The matcher strategy is invoked for a particular finding F queried relative to a group o f findings G, if F has relevant findings in G or F has relevant s y n o n y m s in the d o m a i n model. finding subject I I t I / , I I I I ) - - t s ~ _ y _ _ - e , , " I X ,locality I ) J S • tl i i I i s S ' s - j J s t t a time / I I I I ,'I / 1 • I • I I I I I t - S " J . - ) J - - J tb R e l e v a n t f i n d i n g s The relevant findings o f a finding F from a group G o f findings is the subset o f G defined as follows: relevant-finding s( F, G) = {P c G I subject(P) = subject(F) and overlapping(time-interval(P), time-interval(F)) a n d overlapping(locality(P), locality(F)} i.e. those elements o f G which have the same subject as F and whose time-interval a n d locality attributes overlap with the corresponding time-interval and locality attri- bute o f F (if the time or locality are n o t given explicitly, the defaults 'now' and 'some-relevant-locality' are assumed, respectively). Finding subjects are discrete, time intervals specify a c o n t i n u o u s period o f time, a n d locality attributes can specify a collection o f discrete, (physically) disjoint localities, or a c o n t i n u o u s space o f adjoining localities. A finding, therefore, defines a subs- pace on the entire space o f findings given in Figure 2. The subspace could be a single point if the finding's time interval refers to a time point a n d its locality to a single discrete locality. The relevant findings o f a finding, from a group G o f findings, are those findings whose subspaces overlap with (e.g. are contained in) the subspace o f the given finding (see Figure 3). L o c a l i t y attribute Consider the following findings f r o m the d o m a i n o f skel- etal dysplasias, expressed in natural language: 'all the vertebrae [of the spine] are irregular'*, 'some o f the verte- brae are fiat, i.e. they have the condition platyspondyly', 'the femoral capital epiphyses are small', and 'the middle o f the face is flat'. In each o f these findings an abnormal- ity is described which occurs in a specific region (local- *A finding will be denoted in quotes, e.g. 'platyspondyly at birth', in natural language or when referring to the user interface, and as a parenthesised expression, e.g. (platyspondyly (atmonths (0 0))) when referring to the internal representation. Key / - - 7 subspace representing finding (S ... [Ix ly] [ta tb]) subspace of a relevant-finding (contained-within or overlapping-with the finding's space) Figure 3 Finding and its relevant findings as subspaces on findings space ity). The locality specifies a generic region, or a compo- nent, or a subtype, o f the subject. The above findings can be expressed in the f o r m a t (subject) (locality) (attri- bute-values) as follows: 'spine all-vertebrae irregular', 'spine some-vertebrae flat', 'epiphyses femoral-capital small', and 'face middle flat'. In addition, generic terms can describe localities rele- vant to m a n y subjects. The above examples use 'all-ver- tebrae' and 'some-vertebrae'. It is also possible to talk a b o u t 'all-long-bones', 'all-epiphyses', 'all-digits' etc., a n d similarly a b o u t 'some-long-bones', 'some-epiphyses' a n d 'some-digits'. All these can be expressed in terms o f the two generic locality descriptions: localised a n d gener- alised. These two localities are opposing (conflicting) localities. Other generic locality descriptions could be 'upper', 'middle' and 'lower', 'left' a n d 'right', 'laterally' a n d 'medially', 'anteriorly' a n d 'posteriorly' etc. These are n o t opposing localities since they specify disjoint regions. Since the d a t a model allows for the explicit represen- tation o f taxonomic relationships between finding sub- jects, the above findings can therefore be rewritten as 'vertebrae irregular generalised', 'vertebrae flat localised' (or 'platyspondyly localised'), 'femoral-capital-epiphyses small' a n d 'face middle flat'. Locality attributes are spe- cial attributes as illustrated by the following example. The finding 'vertebrae flat irregular localised from-birth' can be converted into the two atomic findings 'vertebrae Knowledge-Based Systems Volume 6 Number 3 September 1993 147 T o w a r d s competent information acquisition interactions: E T K e r a v n o u e t al. flat localised f r o m - b i r t h ' and 'vertebrae irregular loca- lised from-birth', which say t h a t some vertebrae are irre- gular f r o m birth and some are flat f r o m birth*. T h e scope o f a locality attribute is a finding subject if the finding subject describes an a b n o r m a l i t y (e.g. 'platyspondyly'); otherwise, in the case o f ' n o r m a l ' (i.e. anatomical c o m p o - nent) subjects like vertebrae, its scope is a (finding-sub- ject attribute-value) pair, where the attribute value is a non-locality value. This is because it is meaningful to say ' p l a t y s p o n d y l y localised' but it is not meaningful to say 'vertebrae localised'. Suppose that the p r o b l e m findings include 'verteb rae irregular localised' and that the query finding is 'verte- brae irregular generalised'. T h e query finding is estab- lished as false because the localities generalised and loca- lised are opposing; in this respect, these two generic localities are treated as any o t h e r attribute value f r o m the perspective o f the matcher. Suppose, further, that the problem findings include 'long-bones short' and that the q u e r y finding is 'ulnae short' (ulnae being two o f the long bones). T h e query finding is established to be true as follows: 'long-bones short' = 'long-bones all-long-bones sh o rt ' 'ulnae short' = 'long-bones ulnae short' {ulna} c all-long-bones I f the query finding were 'long-bones short' and the problem finding were 'ulnae short' the query finding would be u n k n o w n . T h e relevant findings o f a finding could therefore include findings whose subjects are taxonomically related to the subject o f the finding. T a x o n o m i c relationships between finding subjects, however, are better manipu- lated by the inferencer (see the f o u r t h section). T i m e i n t e r v a l s Findings have temporal aspects which default to 'now'. The t e m p o r a l reasoner which supports the decide-status function is discussed in Reference 11. T h e t e m p o r a l aspects o f findings are expressed qualitatively, e.g. 'at- birth', ' f r o m - b i r t h ' , 'under-yrs', ' = + y r s ' , 'infancy', 'mid- childhood' etc. in the S D D system. These qualitative expressions are translated internally into time intervals. Because o f the u n c e r t a i n t y inherent in some o f the quali- tative t e m p o r a l aspects, e.g. ' f r o m a b o u t a certain time' or 'up to a b o u t a certain time', the time intervals can be 'open'. Since an open time interval essentially delineates a smaller closed interval, f o r the purposes o f this p a p e r it is assumed that there is no time uncertainty and hence time intervals are closed intervals. Recently, Console et al? 2 have r e p o r t e d o n a causal t e m p o r a l f r a m e w o r k allowing for 'variable' time intervals. In general m u c h a t t e n t i o n has been given to t e m p o r a l reasoning by the AI c o m m u n i t y in the recent years, b o t h at the theoretical and applicative levels (e.g. References 13-20). R e l e v a n t s y n o n y m s Th e c o n c e p t o f relevant synonyms is illustrated t h r o u g h an example f r o m the d o m a i n o f skeletal dysplasias. T h e S D D d o m a i n m o d el includes the following relations: syn-subjects (vertebrae vertebral-bodies) synonyms ((platyspondyly) (vertebral-bodies flat)) Let the q u ery finding F be (vertebral-bodies flat). T h e relevant s y n o n y m s o f F are (platyspondyly) and (verte- brae fiat). T h e first s y n o n y m is a direct synonym o b tained fro m the synonyms relation in the d o m a i n model. T h e second s y n o n y m is a direct subject synonym o b t a i n e d by directly substituting a s y n o n y m o u s subject fo r the query- finding subject. Again the s y n o n y m o u s subjects are dir- ectly given in the d o m a i n model t h r o u g h the syn-subjects relation. This example illustrates the two direct routes t h r o u g h which relevant sy n o n y m s m a y be derived. In addition there are two indirect routes. I f the subject p l a t y s p o n d y l y h ad an y s y n o n y m o u s subjects, o r if find- ing (vertebrae flat) h ad an y direct synonyms, t h en m o r e relevant sy n o n y m s would have been obtained. T h u s the first indirect ro u t e is to substitute s y n o n y m o u s subjects for the direct s y n o n y m s o f F an d the second indirect ro u t e is to determine direct sy n o n y m s fo r the direct sub- ject s y n o n y m s o f F. T h e relevant s y n o n y m s o f a q u ery finding F are there- fore defined as follows: relevant-synonyms(F) = direct-syn(F) U direct-subj-syn(F) O indirect-subj-syn( F) U indirect-syn( F) direct-syn(F) = {O I synonyms(Oa ~a) an d directly-matches(F, ~ ) an d time-interval(O) = time-interval(F)} where Xa is the a t e m p o r a l finding o f finding X. direct-subj-syn( F) = { O L syn-subjects( subject( O ) subject (F)) an d attribute-values(O) = attribute-values(F) an d time-interval(O) = time-interval(F)} indirect-subj-syn(F) = U direct-syn(03 i where Oi c direct-subj-syn(F) indirect-syn(F) = U direct-subj-syn(Oi) i where Oi e direct-syn(F) U indirect-subj-syn(F) *But not necessarily the same vertebrae which is a limitation of the representation if this is what is required to be expressed. T h e direct sy n o n y m s o f a finding are o b t a i n e d f r o m the synonyms relation in the d o m a i n model. T h e findings in the instances o f this relation are a t e m p o r a l an d hence the 148 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou et al. time interval o f the q u e r y finding is removed. T h e result- ing a t e m p o r a l finding is m a t c h e d against the a r g u m e n t s o f the instances o f the synonyms relation. I f a direct m a t c h occurs with one o f the two a r g u m e n t findings, the o t h e r finding a u g m e n t e d with the relevant time interval is a direct s y n o n y m o f the q u e r y finding. An a t e m p o r a l finding directly matches a n o t h e r a t e m p o r a l finding if they have identical subjects a n d the attribute values o f the f o r m e r are subsumed by the a t t r i b u t e values o f the latter. F o r example (metaphyses flared) directly matches against (metaphyses wide irregular), 'flared' and 'wide' being s y n o n y m o u s values (in the c o n t e x t o f the subject 'metaphyses'). F o r efficiency reasons the synonyms rela- tion is implemented as tuples (equivalence classes) o f s y n o n y m o u s findings, indexed by the relevant subjects. Similarly, the syn-subjects relation is i m p l e m e n t e d as tuples o f s y n o n y m o u s subjects. Completeness of relevant-synonyms definition T h e completeness o f the relevant-synonyms definition is p r o v e n below, starting with the following assumptions: • Canonical subject assumption: One subject f r o m each equivalence class o f s y n o n y m o u s subjects is desig- n a t e d as the canonical subject. • Minimality assumption: T h e equivalence classes o f s y n o n y m o u s findings are expressed in a minimal way, i.e. o n l y the canonical subjects are used. • Completeness assumption: T h e knowledge base is complete with respect to all s y n o n y m relations, i.e. the equivalence classes o f s y n o n y m o u s subjects, the minimal equivalence classes o f s y n o n y m o u s findings, and the equivalence classes o f s y n o n y m o u s attrib u t e values are complete. Suppose t h a t the relevant-synonyms definition is incom- plete, m e a n i n g t h a t one o f the relevant s y n o n y m s o f some finding, which is derivable f r o m the knowledge base, is n o t covered by the definition. T h e a b o v e p r o c e d u r a l * definition o f relevant syno- nyms (which facilitates i m p l e m e n t a t i o n ) is 'equivalent't to the following declarative definition: • P is a relevant synonym o f F if syn-subjects( subject( P), subject(F)) a n d attribute-values(P) = attribute-values(F) a n d time-interval(P) = time-interval(F) • P is a relevant synonym o f F if *Procedural because it imposes a specific sequence in the operations involved. t'Equivalent' because all findings returned by the procedural definition are also returned by the declarative definition and any finding returned by the declarative definition which is not explicitly returned by the procedural definition is directly related to one returned by the proce- dural definition through their attribute values. These findings do not need to be returned explicitly since the matcher can dynamically estab- lish whether such a link exists between two findings. synonyms(X, Y) a n d imatches(F,, X) a n d imatches(P., Y) a n d time-interval(P) = time-interval(F) where X an d Y are a t e m p o r a l findings; F,(P,) is the a t e m p o r a l finding o f F(P), a n d imatches(U., V,) if {subject(U,) = subject(V,) o r syn-subjects(subject( U,), subject(V,))} a n d (attribute-values(V,) => attribute-values(U,)) i.e. imatches(Ua, V,) i f V, => U. (V, subsumes U, where s u b s u m p t i o n includes identity). A relevant s y n o n y m is t h erefo re n o t co v ered i f the syn-subjects relation in the first clause is not satisfied, either the synonyms o r o n e o f the imatches relations in the second clause is n o t satisfied; i f an imatches relation is n o t satisfied this m ean s t h a t either the syn- subjects relation is n o t satisfied o r the => relation is n o t satisfied; the => relation is n o t satisfied i f syno- n y m -v al u e relations are n o t satisfied. In each o f the a b o v e cases there is a violation o f the completeness assu m p t i o n (the s y n o n y m o u s subjects are incomplete, o r the minimal s y n o n y m o u s findings are incomplete o r the s y n o n y m o u s attribute values are incomplete). H e n c e the a b o v e definition o f relevant syno- n y m s is complete. Using non-redundant internal representation T h e ability o f a system to u n d e r s t a n d s y n o n y m o u s expressions ( t h r o u g h s y n o n y m o u s findings, s y n o n y m o u s finding subjects a n d s y n o n y m o u s attribute values) is con- sidered b y the d o m a i n experts o f the S D D system as an i m p o r t a n t req u i rem en t because it gives m u c h flexibility to the users o f the system in entering i n f o r m a t i o n dir- ectly. H o w e v e r the question t h a t justifiably arises is w h et h er this r e q u i r e m e n t should be catered f o r by the user interface o r w h e t h e r the system should s u p p o r t an internal r e d u n d a n t representation, requiring the d y n a m - ic c o r r e l a t i o n o f s y n o n y m o u s expressions. I f the first o p t i o n is t ak en t h en a f r o n t - e n d processor w o u l d trans- late every user finding to its internal canonical (standard) representation; o n o u t p u t , findings w o u l d be translated b ack to the expressions used b y the user. T h e advantages o f this o p t i o n are t h a t it reduces the ru n -t i m e processing and, to a large extent, c o m m u n i c a t e s with users in their o w n terms (with the ad d ed possibility o f expressing their findings in the system's terms). T h e benefit o f the second a p p r o a c h is t h a t the infor- m a t i o n is c o m m u n i c a t e d back to the user as entered w i t h o u t incurring an y overheads, whilst with a s t a n d a r d internal rep resen t at i o n it m a y be difficult t o translate some item o f i n f o r m a t i o n b ack to its original form. T h e canonical f o r m f o r some finding, chosen o n the basis o f u n i f o r m i t y an d o t h e r criteria which aim to eliminate Knowledge-Based Systems Volume 6 Number 3 September 1993 149 Towards competent information acquisition interactions: E T Keravnou e t al. ambiguity, m a y not in fact be the preferred form from the user's point o f view. holds(S). I f 0 [--- Y then return true. I f ® ~ {E and • then return false. Step 3: • is unknown. Synonymous attribute values Attribute values m a y be s y n o n y m o u s in the context o f specific subjects. F o r example (metaphyses flared) m a y be taken as s y n o n y m o u s with (metaphyses wide), and (thorax broad) is s y n o n y m o u s with (thorax wide). How- ever metaphyses are n o t described as broad nor is thorax described as flared. Nonetheless the three terms {broad, wide, flared} are considered s y n o n y m o u s in the SDD system. The tradeoff here is between the simplicity o f this representation, which m a y result in some unnecessary processing, and a more complex representation structure which makes explicit the context under which attribute values are synonymous. Findings which are s y n o n y m o u s through their attri- bute values are dynamically determined by the matcher. Negative, positive and status findings The negative finding for a subject at a given time and locality is an atomic finding whose attribute value is the negative value for the subject (see the decide-status d a t a model above). A negative finding expresses a normal situation and thus excludes the relevant positive findings. Examples o f negative findings are (platyspondyly absent) and (hands normal). A positive finding therefore expresses an abnormality. F o r each subject there is a single negative finding for a given time and locality, a most general positive finding and a number o f more speci- fic positive findings. The most general positive finding for an a b n o r m a l subject simply expresses the presence o f the given subject, and for a normal subject it expresses the fact that the given subject is abnormal, e.g. (platyspon- dyly) and (hands abnormal). Some anatomical-component subjects have a status finding which expresses the absence o f the given finding subject (at a particular time and locality), e.g. (humeri absent) or (skull-vault absent-ossification). The status finding is an atomic finding whose attribute value is the status value for the subject. A status finding is a positive finding since it expresses an abnormality. However it is a special positive finding because it excludes all other find- ings for that subject (for the given time and locality). The matcher operation is as follows Let • be u {holds(E)} i where F i e relevant-findings (Q, G r o u p ) and • be holds- (Q). • Step 1: I f ~ [-- • then return true. I f ~ ~ - - ~ then return false. • Step 2: Let S e relevant-synonyms (Q, Group), ® be I f Step 1 fails, Step 2 is tried, and, if that fails as well, the query finding Q is determined as u n k n o w n by the matcher. In order to decide whether a finding matches, or is refuted, against a group o f findings, the matcher uses the following axioms. • A x i o m 1: negative-finding (N) and positive-fnding (P) and subject ( N) = subject(P) and overlapping(locali- ty(N), locality(P)) and overlapping (time-interval(N), time-interval(P)) and holds(P) ~ ~ holds(N). • A x i o m 2: negative-finding(N) and positive-finding(P) and subject( N) = subject( P) a n d overlapping(locali- ty(N), locality(P)) and overlapping(time-interval(N), time-interval(P)) and holds( N) ~ .,~ holds( P). • A x i o m 3: negative-finding(NO and negative-find- ing(N2) and subject(N=)= subject(N2) and includes(lo- cality(NO, locality(N2)) and includes(time-inter- val(NO, time-interval(N2)) and holds( N O--* holds( N2). • A x i o m 4: positive-finding(P) and most-general-posit- ive-finding(M) and subject(P)= subject(M) a n d inclu- des(locality(P), locality(M)) and includes(time-inter- val(P), time-interval(M)) a n d h o l d s ( P ) ~ h o l d s ( M ) . • A x i o m 5: non-status-finding(N) and status-finding(S) and subject(N) = subject(S) and overlapping(local- ity(N), locality(S)) and overlapping(time-interval(N), time-interval(S)) and holds( N ) ~ ,,~ holds(S). • A x i o m 6." non-status-finding(N) and status-finding(S) and subject(N)=subject(S) and overlapping(local- ity(N), locality(S)) and overlapping(time-interval(N), time-interval(S)) and holds( S ) ~ ~ holds( N). • A x i o m 7: status-finding(SO and status-finding(S2) and subject(SO = subject(S2) and includes(locality(SO, locality(S2)) and includes(time-interval(Si), time- interval( Sz) ) and holds(SO--holds(S2). • A x i o m 8: positive-finding(P) and V V e values (P) q F such that positive-finding(F) and match-value(P, F, lO and holds( F ) ~ holds( P). • A x i o m 9: positive-finding(P) and value (P, V) and positive-finding(F) and refute-value(P, F, V) and holds(F) ~ ~ holds(P). • A x i o m 10: positive-finding(P1) and positive-find- ing(P2) and subject( P O = subject( P2) and includes(lo- cality(P2), locality( P O ) and includes(time-inter- val(P2), time-interval(PO) and value(Pb Vl) and value(P2, V2) and identical-or-synonymous( V1, V2)~match-value(P1, P2, VI). • A x i o m 11: positive-finding(PO and positive-find- ing(P2) and subject(PO=subject(P2) and overlapp- ing(locality(P2), locality(Pl)) and overlapping(time- interval(P2), time-interval(Pl)) a n d value(Pl, Vl) and value(P2, V2) and opposite-values(V1, I12) --*refute- value(Pi, P2, VO. U {holds (F0} i where F i e relevant-findings (S, Group), and E be (includes (X, Y) means that X includes Y. subject, local- ity, time-interval and values are selector functions which apply to findings.) 150 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou et aL INFERENCER The inferencer is invoked for some finding Q queried relative to a group of findings if Q has relevant impli- cants in the domain model. Relevant implicants The relevant implicants of a finding are obtained from the implies relation in the domain model: implies( Fl, F2) i.e. F, =~F2, where F, and F2 are atemporal findings, relevant-implicants( F) = positive-implicants( F) U negative-implicants(F) where positive-implicants(F) = {Pt [ implies(P, C) and matches(Fa, C)} negative-implicants(F) = {Pt I implies(P, C) and m a t c h e s ( ~ F~, C)} Pt is the temporal finding obtained from P and the time interval of F, Fa is the atemporal finding of F and mat- ches(X, Y) is true if the matcher returns 'true' when X is queried against the singleton group of findings consisting of Y. The implies relation collectively covers dependency, definitional, or causality relationships between findings. Generalisations and restrictions The is-a and p a r t - o f domain relations are also used to obtain more relevant implicants according to the con- text-free axioms given below. • A x i o m 1: sholds(P, S) and is-a(s, S ) ~ s h o l d s ( P , s). sholds(P, S) means that proposition (i.e. finding) P holds for subject S, e.g. (long-bones short)~(ulnae short). • A x i o m 2: V s such that is-a(s, S) sholds(P, s ) ~ sholds(P, S), e.g. {(ulnae short) and (radii short)} ~long-bones short). • A x i o m 3: V s such that part-of(s, S) sholds(P, s)--* sholds(P, S), e.g. {(cervical-spine normal) and (thor- aco-lumbar-spine normal)}--* (spine normal). • A x i o m 4: ~ s such that part-otis, S) sholds(abnor- real(s), s) ~ sholds(abnormal(S), S), e.g. (mid-face abnormal)--* (face abnormal). • ? A x i o m 5: sholds(P, S) and part-of(s, S) ~ sholds (P, s). The fourth axiom deals with the most general positive findings. This is because if a part of some subject is abnormal in some specific way, it does not necessarily mean that the whole subject is similarly abnormal, e.g. the trunk and the limbs may be loosely described as parts of one's stature; if the trunk is short it is not necessarily the case that the stature is also short (although the opposite would be unusual). Similarly, Axiom 5 is not really used because it is not truly context-free (except when the property is normality). Axiom 5 may apply under some interpretations only. For example if the spine is abnormal it does not mean that all the vertebrae on each of the regions of the spine (cervical-spine, lum- bar-spine, thoracic-spine) are abnormal. On the other hand, if the spine is irregularly ossified it can be inferred that all the vertebrae are irregularly ossified. In other words if we are dealing with a general abnormality (e.g. 'spine abnormal') it is not possible to infer that all the subcomponents are abnormal. With some specific abnor- malities, however (e.g. 'spine irregularly-ossified'), it can be inferred that the given abnormality applies to all the subcomponents. The scope of this axiom therefore depends on domain specific knowledge. Suppose that the query finding is (ulnae severely- short) and that the given finding is (short-limbed-dwar- fism). The latter finding is synonymous with (long-bones severely-short) which is a generalisation of the query finding. Hence the inferencer, by dynamically generating implications based on taxonomic relations (is-a and part- of), enables 'linking' between the query finding and a given finding which is synonymous with generalisations/ restrictions of the query finding. This example illustrates the co-operation between the matcher and inferencer. Inferencer operation The inferencer operation is simple to express since the inferencer calls Decide-Status recursively to decide the truth status of the antecedents of the relevant implicants of the query finding. I f a positive implicant is entailed to hold then the query finding holds (the inferencer returns true) and if a negative implicant is entailed to hold the query finding does not hold (the inferencer returns false). Otherwise the answer to the query is decided to be unknown by the inferencer. The inferencer axioms are given below. Let F be the assertions about which findings hold for the given problem • A x i o m 1: L e t P c positive-implicants(F). If F ]-- holds(P) then F ~ holds(F). • A x i o m 2: Let N ~ negative-implicants(F). If F [-- holds(N) then F ~-- ,-~ holds(F). If the problem theory is consistent, i.e. both the domain model and problem findings are consistent, then the two axioms are never simultaneously applicable for the same query finding. The chain of implications which is dyna- mically constructed through the recursive calls between Decide-Status and the inferencer is recorded so that a loop can be immediately detected and the relevant infer- ence chain terminated. The matcher and inferencer strategies both compete and co-operate. If the matcher cannot return a firm answer, the inferencer is invoked; however it is the matcher that terminates a successful inference chain generated by the inferencer by determining the relevant terminal antecedent finding as true. Decide-Status is implemented in Franz LISP on a Sun Knowledge-Based Systems Volume 6 Number 3 September 1993 151 Towards competent information acquisition interactions: E T Keravnou e t al. 3/60 running Unix (declarative definitions for the matcher, inferencer and Decide-Status functions are given in Reference 6). The algorithms for the Decide- Status and the inferencer, illustrating the chain recursion between them are as follows. The definition below assumes that the a r g u m e n t Find- ing is a simple finding; the generalisation to include com- p o u n d findings is given in Reference 6. Finding is queried against a group o f findings, G r o u p , which are assumed to hold. Impl-Chain is a chain o f implications, initially empty, which is constructed dynamically t h r o u g h the recursive calls between Decide-Status a n d the inferencer. Decide-Status (Finding, Group, Impl-Chain): let Truth-Value = Matcher(Finding, G r o u p ) if firm(Truth-Value) then return(Truth-Value) else let Rel-Impl = relevant-implicants(Finding) New-Chain = cons(Finding, Impl-Chain) return(Inferencer(Rel-Impl, G r o u p , New-Chain)) A firm truth value means 'true' or 'false', but not ' u n k n o w n ' . The relevant implicants o f Finding are those findings which could be used to infer Finding. Inferencer (Rel-Impl, Group, Impl-Chain): if null(Rel-Impl) return ('unknown') else if member(first(Rel-Impl), Impl-Chain) then return (Inferencer(rest(Rel-Impl), Group, Impl-Chain)) else if Decide-Status(first(Rel-Impl), Group, Impl- Chain) = 'true' then if positive (first (Rel-Impl)) then return ('true') else return ('false') else return(Inferencer(rest(Rel-Impl), Group, Impl-Chain)) The inferencer determines whether any o f the relevant implicants can be determined from the group o f findings. No formal analysis o f the complexity o f the decide-status function has been done. However it is a computationally intensive operation; much o f this processing is attributed to the inferencer operation since the inferencer m a y insti- gate a number o f alternative inference chains which m a y all prove to be unfruitful. The matcher is always tried before the inferencer. As Decide-Status is invoked in m a n y different reasoning contexts in the expert system (see below), its optimisation would be beneficial. Finally Decide-Status is a m o d u l a r piece o f software which faci- litates extensions. New strategies or axioms can be easily added. Such extensions m a y be accompanied by exten- sions to the d a t a model. R E A L I S I N G I N F O R M A T I O N - A C Q U I S I T I O N G O A L S T H R O U G H D E C I D E - S T A T U S Decide-Status operates on an explicit d a t a model. Its operation is conceptually simple ('given a d o m a i n model and a set o f findings which are assumed to hold, does a particular finding follow or not?'), but powerful enough to either directly implement or support the implemen- tation o f the information-acquisition interaction objec- tives discussed in the first section. The d a t a model makes explicit the structure o f the d o m a i n o f findings. Findings are n o t indivisible, atomic entities, i.e. strings o f characters; they have an internal meaningful structure, enabling the decomposition o f individual findings and the correlation between find- ings*. Mixed-initiative interaction The dependencies between findings (synonyms, generali- sations, restrictions, implications) are explicit; Decide- Status, through these dependencies, can correlate expres- sions which mean literally or essentially the same thing, and can detect r e d u n d a n c y and inconsistency in the user- volunteered information. As a result o f this, the user has the option to volunteer information directly rather than through a sequence o f menus. It is still possible that some information volunteered by the user will not be under- stood by the system in which case the system should inform the user. As explained in the first section, the overall sequence o f questions raised by the system can be viewed at two levels o f abstraction. The top level gives the sequence o f strategy applications generated by the workings o f the problem solver and the b o t t o m level gives the question subsequences corresponding to the different strategy applications 2. Thus some strategy application determines that a given set o f information items should be elicited from the user. However it is the d a t a model and the associated relevant reasoning that organises the required information into an intelligible, coherent sequence o f questions. Questions for the same subject are grouped together and preceded by the relevant general questions (most general positive finding and, if applicable, status finding). F o r example if the system wants to ask whether the skull vault is small it will first try to establish whether the skull vault is abnormal and whether it is ossified (i.e. present) in this order. The context for raising specific questions in a meaningful way is established by asking general questions first. I f this were n o t so then an answer could be ambiguous; for example if it is n o t known whether the patient exhibits platyspondyly and the system raises the question 'platyspondyly severe?' a nega- tive answer would either mean that 'there is no platy- spondyly' or that 'there is platyspondyly but it is mild'. In addition to general questions, there m a y be other contextual questions, either associated with the subject or specific findings o f the subject, which need to be raised prior to asking some particular question. At present the d a t a model does n o t include such contextual associa- tions, but their inclusion is straightforward. The ordering o f the groups o f questions for different subjects is n o t necessarily arbitrary. Again contextual associations between different subjects, if any, can indi- *In m a n y s y s t e m s findings h a v e n o s t r u c t u r e a n d internally t h e y c a n be represented b y a n o n y m o u s codes, w h i c h o f c o u r s e speeds u p t h e p r o - cessing; this is o f t e n t h e case w i t h m e n u - d r i v e n interfaces where strings are displayed o n t h e screen b u t u s e r replies a r e t r a n s l a t e d i n t o internal, n o n - d e c o m p o s a b l e codes. 152 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou e t al. cate a p a r t i c u l a r sequence. In the S D D system the spatial p r o x i m i t y o f a n a t o m i c a l c o m p o n e n t s is used to o r d e r questions for adjoining a n a t o m i c a l c o m p o n e n t s consecu- tively, thus saving the user f r o m having to keep swapping between different X - r a y films; hence questions o n the ribs a n d the thoracic spine will be g r o u p e d together. A user reply to a system question is immediately t a k e n into consideration, which is reflected in subsequent questions being screened out. T h e decide-status funct i o n is used p r i o r to asking every question to see w h e t h e r the answer to the question can already be determined. In some expert systems the questions raised by the system are m o r e general t h a n the c u r r e n t l y p u r s u e d goal requires, where the generality is n o t for establishing a meaningful c o n t e x t for raising specific questions b u t fo r acquiring all relevant i n f o r m a t i o n simultaneously (e.g. instead o f asking 'does p a r a m e t e r x o f object y have value z?' it will ask ' w h a t is the value o f p a r a m e t e r x fo r object y?'). This is absolutely necessary if the system does n o t allow the user to v o l u n t e e r i n f o r m a t i o n . In the pro - posed f r a m e w o r k this is n o t so. I m m e d i a t e l y after a questioning r o u n d by the system, the user is given the o p p o r t u n i t y to v o l u n t e e r i n f o r m a t i o n . T h e main objective o f expert systems t e c h n o l o g y is to disseminate (and possibly augment) h u m a n expertise 21. Thus, the m a j o r i t y o f users o f an expert system, a l t h o u g h knowledgeable in the given d o m a i n , will n o t be experts themselves. In this respect, it is possible that i n f o r m a t i o n volunteered by the user will be e r r o n e o u s a n d t h a t the user will n o t u n d e r s t a n d a question raised by the system. Requiring the system to t r a p e r r o n e o u s user input and to give a p p r o p r i a t e guidance to the user when respond i n g to system questions 7,22,23 puts a n o t h e r level o f complex i t y on the system i n f o r m a t i o n - a c q u i s i t i o n initiatives. E r r o n - eous input which conflicts with o t h e r i n f o r m a t i o n can be trapped. Clarifying questions (see below) d o n o t aim to trap e r r o n e o u s i n p u t as such, b u t r a t h e r to t r a p c o m m o n inaccuracies in the interpretations o f actual situations. Serious inaccuracies result in e r r o n e o u s input. In the S D D system an on-line d a t a b a s e o f X - r a y images will be used in the validation o f user input and in provid i n g guidance to the user in answering a system question. System questions m a y be illustrated by displaying a rele- vant image and, similarly, user input m a y be validated b y displaying images which illustrate w h a t the user has said. I f the displayed image does n o t m a t c h the actual case image then it is likely t h a t the user has misinterpreted the particular situation*. Flexibility in user-volunteered information T h e instantiation o f the d a t a m o d e l for a specific d o m a i n aims to c a p t u r e the entire v o c a b u l a r y regarding the expression o f findings, in o r d e r to give the user the required flexibility. In a d d i t i o n the user m a y wish to revoke i n f o r m a t i o n . Solutions to t r u t h m a i n t e n a n c e are n o t cheap since assumptions a n d inference dependencies m u s t be expli- citly represented. On the basis o f w h a t i n f o r m a t i o n is assumed to h o l d at some stage in the consultation, par- tial solutions are g en erat ed a n d pursued. These alterna- tive solutions are evaluated against each other. I f some i n f o r m a t i o n is r e v o k e d t h en its rev o cat i o n will f a v o u r the solutions conflicting with this i n f o r m a t i o n a n d will reduce the promise o f those solutions s u p p o r t e d by the r e v o k e d i n f o r m a t i o n . I f the r e v o k e d i n f o r m a t i o n is criti- cal f o r a p art i cu l ar solution t h en t h a t solution m a y be excluded. Decide-Status does n o t directly enable infor- m a t i o n r e v o c a t i o n b u t it su p p o rt s it to some extent. In a diagnostic c o n t e x t the promises o f the various co m peting hypotheses are dynamically r e c o m p u t e d d u ri n g every diagnostic cycle, always using the i n f o r m a t i o n which is believed to h o l d at t h a t stage (Decide-Status is called to determine w h et h er hypotheses' expectations are satisfied o r refuted a n d w h et h er case findings are a c c o u n t e d by, or are in conflict with, hypotheses). H e n c e revocations will be reflected in the new 'p ro m i se' f o r an active hypothesis. In practice, however, because Decide-Status is a c o m p u - tationally expensive process, some o f its derivations are stored fo r fu t u re use. I f a r e v o c a t i o n results in the refu- t at i o n o f a previous derivation t h en inconsistencies will occur. A n u m b e r o f d o m a i n - i n d e p e n d e n t t ru t h -m ainten- ance f r a m e w o r k s have been p r o p o s e d in the literature 26-28 which present theoretically interesting a p p r o a c h e s b u t which m a y fail in practice because o f the associated c o m p u t a t i o n a l o r o t h e r overheads. In practice, for rea- sons o f viability, a t r u t h - m a i n t e n a n c e f r a m e w o r k will be tailored to the specifics o f a p art i cu l ar p r o b l e m solver, o r m a y even be 'h ard -w i red ' in the workings o f the p r o b l e m solver. In the S D D system the case findings are divided into ' h a r d ' findings a n d 'soft' findings. T h e distinction between h a r d an d soft findings is m a d e in m a n y medical expert systems, often to distinguish between l a b o r a t o r y findings an d co-incidental (circumstantial) findings. S D D ' s m ean i n g an d usage o f these terms m a y be differ- ent f r o m o t h e r systems, a l t h o u g h the c o n c e p t is essen- tially the same: some findings are m o r e i m p o r t a n t t h a n others in a diagnostic contextt. H a r d findings c o r r e s p o n d to diagnostically significant observations o f a b n o r m a - lites. Soft findings describe abnormalities which m a y be a t t r i b u t e d to n a t u r a l causes. T o be acceptable, any hypothesis m u st a c c o u n t fo r a reasonable p r o p o r t i o n o f h a r d findings, while soft findings m a y conceivably be ignored. Obviously, revoking a h a r d finding will affect the solution space b u t the rev o cat i o n o f a soft finding will n o t necessarily affect the solution space. Similarly, some findings are r a t h e r critical fo r certain hypotheses, i.e. the presence o f some finding results in concluding the hypothesis, o r the absence o f some finding results in refuting the hypothesis. Again a n y revocations directly o r indirectly resulting in refuting o r establishing the presence o f such findings m u st be t ak en into consider- at i o n when some i n f o r m a t i o n is revoked. Decide-Status can be used to determine the t r u t h status o f such critical findings, which h av e resulted in considerably enhancing o r reducing the promise o f p o t en t i al alternative solu- tions; the reversal o f potentially irreversible decisions o n the p a r t o f the p r o b l e m solver m a y thus be possible. *The use of images in computer-aided diagnosis in radiology is dis- cussed in References 24 and 25. *Again Decide-Status is used to decide whether a case finding is soft or hard; this reasoning context is outside the scope of the paper. Knowledge-Based Systems Volume 6 Number 3 September 1993 153 Towards competent information acquisition interactions: E T Keravnou e t al. Similarly, conflicts in the user-volunteered infor- mation can be detected by Decide-Status. I f a new obser- vation is determined as false by the previous obser- vations the p r o o f tree constructed by Decide-Status can also be used to explain the conflict (see Reference 10). Decide-Status can be used to determine dependencies between the items o f information volunteered by the user. I f some observation is deducible from other obser- vations then it is redundant. F o r example if the user says that the stature is abnormal and subsequently says that the stature is short, the first observation is eliminated. In a diagnostic context pieces o f evidence should be inde- pendent. Equally, hypotheses' expectations must be non- redundant; again Decide-Status is used to eliminate poss- ible redundancies. In addition the d a t a model (but not Decide-Status per se) is used to temporally screen a hypothesis profile, thus eliminating expectations whose (relative) temporal aspects refer to the future with respect to the patient (see Reference 11). System competence in querying user The system maintains two lists o f findings: Observations and U n k n o w n Findings. The former includes the find- ings which are currently assumed to hold; some o f these findings are directly volunteered by the user whilst others are elicited from the user through questions. Revoking a user observation results in deleting the particular obser- vation from Observations. The U n k n o w n Findings list includes all findings which have been specified by the user, in response to a system question, as unknown. The previously u n k n o w n answer to a question m a y subse- quently become known to the user, who will have the o p p o r t u n i t y to volunteer this information. W h e n new information is volunteered by the user, the U n k n o w n Findings must be revised, since the new information m a y result in determining a firm truth status for a previously u n k n o w n finding. Decide-Status is thus invoked to deter- mine the truth status o f U n k n o w n Findings against Observations. Prior to asking a question, the system first checks whether the truth status o f the query finding can be determined from the Observations list. I f Decide-Status (query-finding, Observations, nil) returns ' u n k n o w n ' , the system checks whether the user has already said that the given item o f information is u n k n o w n , i.e. matcher (query-finding, Unknown-Findings) returns a firm answer (true or false) or if the query finding is more specific t h a n one o f the findings in the U n k n o w n Find- ings list; specificity here is determined using is-a and part- o f relations. F o r example suppose t h a t the user was asked the question 'skull abnormal?' to which the reply ' u n k n o w n ' was given. ( U n k n o w n replies are fairly c o m m o n in the d o m a i n o f SDD since Often the X-ray images for the patient do n o t give a complete skeletal survey and few clinical findings are usually available). I f subsequently the system wants to find whether the skull is small it will n o t ask the question because it knows that the status o f the skull is unknown. Similarly, if the user does not know if the spine is abnormal, the system should n o t ask whether the cervical spine is abnormal. However if the user does not know a b o u t the status o f the cervical spine it is still intelligible on the part o f the system to ask a b o u t the status o f the spine. Since specific questions are always preceded by the relevant general questions (see above), if the U n k n o w n Findings list includes 'platyspondyly severe' then it can be inferred that the presence o f platyspondyly has already been established; otherwise if platyspondyly were absent, 'platyspondyly severe' would be false, not unknown. Finally the d a t a model associates prompting and clari- fying questions, and temporal information (when-to-ask relation) with specific findings. These associations contri- bute much to the naturalness o f the information-acqui- sition interaction. Prompting and clarifying questions are raised in response to information volunteered by the user whilst temporal information is used to stop the system from asking unintelligent questions such as 'is the gait o f y o u r 2 m o n t h old baby waddling?'. When-to-ask relations are used in the context o f temporal screening. The three relations are as follows: • prompting~-query(X, Y,) e.g. prompting-query ((platy- spondyly) (platyspondyly throughout)), • clarifying-query(XaY~) e.g. clarifying-query ((face fiat) (mid-face hypoplastic)), • when-to-ask(W~ T) e.g. when-to-ask ((kyphoscoliosis severe) ( = + yrs 1)). where Xa, Ya and Wa are atemporal findings and T is a relative time point. Suppose that the user volunteers finding F. If mat- cher(X~, [Fa]) returns true, i.e. if Xa holds given Fa, then the temporal version o f the corresponding atemporal prompting query will be generated. Thus if the user volunteers 'platyspondyly' then the question 'platyspon- dyly throughout?' is prompted. I f the user volunteers 'platyspondyly mild' again the same question will be prompted. However if the user volunteers 'platyspondyly t h r o u g h o u t ' no prompting will be given; the prompting query is generated but it is immediately quashed because Decide-Status indicates that it is a r e d u n d a n t question. Clarifying queries are dealt with in the same way. Thus if the user volunteers 'face flat', the clarifying question 'do y o u mean mid-face hypoplastic?' will be raised. A future SDD extension is for the system to infer some prompting and clarifying questions. Regarding when-to-ask associations, suppose that Fa is an atemporal feature o f some hypothesis. I f matcher (F~, [Wa]) returns true, then, if T refers to a future point in time relative to the patient, the particular expectation o f the hypothesis is screened out. Thus if the patient is neonatal then the expectation o f 'severe kyphoscoliosis' for any hypothesis will be screened out. However the more general expectation o f 'kyphoscoliosis' will n o t be screened out, since mild forms o f this abnormality are observable from birth. The d a t a model and the decide-status reasoner can therefore directly support the realisation o f most o f the information-acquisition interactions objectives men- tioned in the first section. 154 Knowledge-Based Systems Volume 6 Number 3 September 1993 Towards competent information acquisition interactions: E T Keravnou e t al. C O N C L U S I O N S T h e n e e d f o r c o m p e t e n t c o n v e r s a t i o n a l s t r u c t u r e s b e t w e e n e x p e r t s y s t e m s a n d their users w a s identified e a r l y in t h e life o f e x p e r t s y s t e m s t e c h n o l o g y ; o v e r t h e y e a r s , as t h e r o l e o f e x p e r t s y s t e m s as intelligent a d v i s e r s h a s a c q u i r e d m o r e p r o m i n e n c e , so h a s t h e n e e d f o r a d e q u a t e i n t e r a c t i o n w i t h users. T h e c o n v e r s a t i o n a l s t r u c t u r e e n c o m p a s s e s t h e e n t i r e i n t e r a c t i o n b e t w e e n t h e s y s t e m a n d t h e user. T h i s i n c l u d e s t h e q u e s t i o n s r a i s e d b y t h e s y s t e m , b o t h i n d i v i d u a l l y as well as a n o r d e r e d s e q u e n c e , the i n s t r u c t i o n s a n d g u i d a n c e given t o t h e u s e r b y t h e s y s t e m f o r p e r f o r m i n g a c t i o n s a n d a n s w e r i n g q u e s t i o n s , t h e s y s t e m ' s e x p l a n a t i o n s r e g a r d i n g its reas- o n i n g a n d / o r s u g g e s t i o n s , a n d t h e k i n d s o f initiatives a n d c h o i c e s w h i c h the u s e r is a l l o w e d t o t a k e , e.g. v o l u n t e e r d a t a , v o l u n t e e r s u g g e s t i o n s , a n d r e t r a c t d a t a o r sugges- tions. T h e issue o f a d e q u a t e e x p l a n a t i o n s h a s a t t r a c t e d m o r e a t t e n t i o n t h a n t h e o t h e r a s p e c t s o f a c o n v e r s a t i o n a l s t r u c t u r e . T h e s e o t h e r aspects, t h e i n f o r m a t i o n - a c q u i - sition ones, a r e e q u a l l y i m p o r t a n t . T h e f o c u s r e g a r d i n g i n f o r m a t i o n - a c q u i s i t i o n i n t e r a c t i o n a s p e c t s is o n intelli- g e n t f r o n t e n d s f o r m a k i n g a piece o f s o f t w a r e o r d a t a - b a s e m o r e u s a b l e r a t h e r t h a n in t h e c o n t e x t o f e x p e r t c o n s u l t a n t s y s t e m s p e r s e 7. H o w e v e r a t t e n t i o n is s h i f t i n g in this d i r e c t i o n 29. T o be able t o c o n d u c t a n intelligent i n f o r m a t i o n - a c q u i s i t i o n i n t e r a c t i o n , b o t h in t h e sense o f a s k i n g intelli- gent, c o h e r e n t q u e s t i o n s , a n d in t h e sense o f u n d e r s t a n d - ing t h e u s e r ' s i n f o r m a t i o n v o l u n t e e r i n g , a n e x p e r t s y s t e m m u s t h a v e c o m m o n sense in a limited w a y specific t o t h e n e e d s o f t h e p a r t i c u l a r a p p l i c a t i o n . T h i s o b j e c t i v e is m u c h easier t o a c h i e v e t h a n t h e c o n s i d e r a b l y m o r e a m b i - t i o u s o b j e c t i v e o f c a p t u r i n g t h e e n t i r e b o d y o f w o r l d k n o w l e d g e , as in t h e C Y C p r o j e c t 3°. T h e a r g u m e n t p r e s e n t e d in this p a p e r is t h a t a n e x p e r t s y s t e m c a n c o m m u n i c a t e m o r e intelligently i f it h a s a d e e p e r u n d e r s t a n d i n g o f t h e d a t a f o r s o m e p r o b l e m case, i.e. i f it is c a p a b l e o f h a n d l i n g these d a t a intelligently; this is t h e limit o f t h e r e q u i r e d c o m m o n sense. T h e p a p e r is n o t a b o u t intelligent d a t a b a s e s o r intelligent f r o n t ends. T h e d a t a m o d e l is n o t a d a t a b a s e s c h e m a , b u t r a t h e r a s c h e m a f o r c a p t u r i n g t h e k n o w l e d g e a b o u t t h e d a t a in t h e d o m a i n . T h e r e l a t i o n s i n c l u d e d in t h e d a t a m o d e l are l a r g e l y c o m m o n - s e n s e r e l a t i o n s s u c h as p a r t - o f , is-a, trig- gers, a n d i m p l i c a t i o n s , as well as p r o m p t s , c l a r i f y i n g q u e s t i o n s a n d w h e n - t o - a s k . T h e a x i o m s u s e d b y t h e infer- e n c e r a n d t h e m a t c h e r a r e c o m m o n - s e n s e a x i o m s , w h i c h in f a c t e n h a n c e s t h e g e n e r a l i t y o f t h e d a t a m o d e l . T h e p a p e r d o e s n o t p r o v i d e a c o m p l e t e g e n e r i c s o l u - t i o n t o t h e p r o b l e m o f intelligent i n f o r m a t i o n - a c q u i s i t i o n i n t e r a c t i o n s f o r e x p e r t s y s t e m s , b u t d e s c r i b e s a k e r n e l w h i c h p o i n t s t o w a r d s a c o m p l e t e s o l u t i o n . T h e s t a r t i n g p o i n t is t o m a k e explicit t h e k n o w l e d g e a b o u t t h e d o m a i n d a t a t h r o u g h a d a t a m o d e l ; t h e d a t a m o d e l is a t t h e k n o w l e d g e m e t a l e v e l , the i n s t a n t i a t i o n o f it is a t t h e k n o w l e d g e o b j e c t level a n d t h e p r o b l e m d a t a a r e a t t h e f a c t u a l level. A C K N O W L E D G E M E N T S W e a r e g r a t e f u l t o t h e L e v e r h u l m e T r u s t f o r their s u p - p o r t o f t h e S D D p r o j e c t , t o t h e referees o f t h e p a p e r f o r their u s e f u l c o m m e n t s o n a n earlier d r a f t o f t h e p a p e r , a n d t o S o p h i a P r e v e z a n o u f o r h e r h e l p in c l a r i f y i n g m a n y p o i n t s . R E F E R E N C E S 1 Gilbert, N 'Explanation and Dialogue' Knowledge Engineering Review Vol 4 No 3 (1989) pp 235-247 2 Keravnou, E T and Johnson, L Competent Expert Systems: A Case Study in Fault Diagnosis Chapman and Hall (1986) 3 Keravnou, E T, Dams, F, Washbrook, J, Dawood, R, Hall, C and Shaw, D 'Background Knowledge in Diagnosis' Artificial Intelligence in Medicine Vol 4 No 4 (1992) pp 1-17 4 Keravnou, E T, Washbrook, J, Dawood, R M, Hall, C M and Shaw, D 'A Model-Based Diagnostic Expert System for Skeletal Dysplasias' Proc. A I M E "89 Springer-Verlag, Germany (1989) pp 47-56 5 Stenton, S P 'Dialogue Management for Co-operative Know- ledge-Based Systems' Knowledge Engineering Review Vol 2 No 2 (1987) pp 99-122 6 Keravnou, E T, Washbrook, J and Dams, F 'Explicit Data- Modelling in Second-Generation Diagnostic Expert Systems' Proc. l l th International Conference on Expert Systems and their Applications- Vol 2 Avignon, France (1991) pp 185-198 7 Kok, A J 'A Review and Synthesis of User Modelling in Intelli- gent Systems' Knowledge Engineering Review Vol 6 No 1 (1991) pp 21-47 8 Keravnou, E T and Washbrook, J 'Deep and Shallow Models in Medical Expert Systems' Artificial Intelligence in Medicine Vol 1 No 1 (1989) pp 1-28 9 Southwick, R W 'Explaining Reasoning: An Overview of Expla- nation in Knowledge-Based Systems' Knowledge Engineering Review Vol 6 No 1 (1991) pp 1-19 10 Keravnou, E T and Johnson, L 'Intelligent Handling of Data by Integration of Commonsense Reasoning' Knowledge-Based Systems Vol 1 No 1 (1987)pp 32-42 11 Keravnou, E T and Washbrook, J 'A Temporal Reasoning Framework used in the Diagnosis of Skeletal Dysplasias' Artifi- ciallntelligence in Medicine Vol 2 No 5 (1990) pp 239-265 12 Console, L, Janin Rivolin, A and Torasso, P 'Fuzzy Temporal Reasoning on Causal Models' Int. J. Intelligent Systems Vol 6 No 2 (1991) 13 Allen, J F 'Towards a General Theory of Action and Time' Artificial Intelligence Vol 23 (1984) pp 123-154 14 Dean, T L and McDermott, D V 'Temporal Data Base Manage- ment' Artificial Intelligence Vol 32 (1987) pp 1-55 15 Kahn, M G 'Model-Based Interpretation of Time-Ordered Medi- cal Data' PhD Dissertation Medical Information Sciences, University of California, USA (1988) 16 Keravnou, E T (Ed.) 'Medical Temporal Reasoning' Artificial Intelligence in Medicine Vol 3 No 6 (1991) (special issue) 17 Kowalski, R and Sergot, M 'A Logic-Based Calculus of Events' New Generation Computing Vol 4 (1983) pp 67-95 18 Long, D 'A Review of Temporal Logics' Knowledge Engineering Review Vol 4 No 2 (1989) pp 141-162 19 Rosser, B, Washbrook, J, Campbell, J, Keravnou, E T and Long, D 'A Framework for Time Dependent Reasoning Systems' Pro- duct Pl11-1 Esprit Project P2409 Equator (1989) 20 Shoham, Y 'Temporal Logics in AI: Semantical and Ontological Considerations' Artificial Intelligence Vol 33 (1987) pp 89-104 21 Slatter, P E 'Cognitive Emulation in Expert System Design' Knowledge Engineering Review Vol 2 No 1 (1987) pp 27-41 22 Wolstenholme, D 'Saying "I don't know" and Conditional Answers' in Moralee, D S (Ed.) Research and Development in Expert Systems I V Cambridge University Press (1987) pp 115- 125 23 Wolstenholme, D 'External Data in Logic-Based Advice Systems' PhD Thesis Dep. Computing, Imperial College London, UK (1990) 24 Mutalik, P G, Fisher, P R, Weltin, G and Swett, H A 'Expert System Advice as a By-product of Image Acquisition and Reporting: Obstacles to Overcome' Proc. C A R '91 Springer- Verlag (1991) pp 315-320 25 Swett, H A 'Computer-Aided Diagnosis in Radiology' Proc CAR" 91 Springer-Verlag (1991) pp 738-743 Knowledge-Based Systems Volume 6 Number 3 September 1993 155 Towards competent information acquisition interactions: E T Keravnou e t al. 26 de Kleer, J ' A n Assumption-Based T M S ' Artificial Intelligence Vol 28 (1986) pp 127-224 27 Doyle, J A ' A Truth-Maintenance System' Artificial Intelligence Vol 12 (1979) pp 231 272 28 Inoue, K 'Pruning Search Trees in Assumption-Based Reason- ing' Proc. 8th International Workshop on Expert Systems and their Applications (1988) pp 133-151 29 Keravnou, E T and Washbrook, J ' W h a t is a Deep Expert System? An Analysis of the Architectural Requirements of Second-Generation Expert Systems' Knowledge Engineering Review Vol 4 No 3 (1989) pp 205 233 30 Lenat, D B, Ramanthan, V G, Pittman, K, Pratt, D and She- pherd, M 'CYC: Towards Programs with Common Sense' C A C M Vol 33 No 8 (1990) pp 30-49 B I B L I O G R A P H Y Neal, I M 'First Generation Expert Systems: A Review of Knowledge Acquisition Methodologies' Knowledge Engineering Review Vol 3 No 2 (1988) pp 105-145 Wilson, M, Duce, D and Simpson, D 'Life Cycles in Software and Knowledge Engineering: A Comparative Review' Knowledge Engineer- ing Review Vol 4 No 3 (1989) pp 189 204 156 Knowledge-Based Systems Volume 6 Number 3 September 1993 
work_pkuiz7nuabfynihbb4g5iq3cfm	----	A Paper for the Journal- Expert Systems with Applications A Methodology to allow Rural Extension Professionals to build Target- specific expert systems for Australian Rural Business Operators Shah J. Miah, Don Kerr, John Gammack Griffith Business School, Department of Management, Griffith University, Logan campus, Queensland, Australia Abstract Expert systems (ES) development technology has been used to build rural business applications in the past but these have usually been developed using traditional expert systems shells. This paper introduces a new architecture for the development of a design environment where the domain experts can build a knowledge base for target-specific ES for rural business operators. The system allows rural business operators to use their own knowledge in building their own, target-specific ES for tailored development to their own specific requirements. At this stage, this reusable design environment caters for the Australian dairy industry but in the long run we claim it will be useful for the other livestock based rural industries such as beef cattle and sheep. This approach of developing target-specific ES contributes new knowledge in that it provides a new way of developing decision support by allowing human domain experts to develop relevant ES for different livestock farming business. An evolutionary prototyping approach was employed for initial development of a proof of concept example and as a method of outlining the solution environment. Multiple qualitative data collection methods were engaged to facilitate knowledge acquisition in the domain of milk protein enhancement for dairy operations. This paper also describes the generic development procedure used in this project. Keywords: target-specific expert systems; design environment; rural application 1. Introduction This article describes development of a new solution architecture in which industry operators can build their own target-specific expert systems (TSES) using their own knowledge. The development methodology ensures end users such as farmers/rural business operator’s involvement in the development process and afterwards encourages the use of contextual knowledge for building their specific expert system in the design environment. We have called this system an end-user enabled design environment (EUEDE). EUEDE represents a specific type of design environment where end users involvement is central. Other reason behind calling the system EUEDE is that our aim is to focus on the end user’s thought processes, relevant technology and their own judgement in order to give them active participation in their own application development. It is hypothesised that this will contribute to increased ownership of the developed application, and help to increase the adoption rate of the developed application. This new methodology takes traditional ES development technology one step further by ensuring active end user involvement in development and by reducing third party involvement. It is also thought that by reducing end-user’s training requirements with respect to analysis and design (Wagner, 2000) and by reducing risk associated (McGill, 2002) with end users development that a better quality product will be produced. McGill (2002) suggested that end user development (in particular spreadsheet applications) does come with some risk and this risk can be in terms of using incorrect design variables and inappropriate design methodology. However, in the proposed design framework used here the end user can exercise their own choices in the context of their own knowledge use and their own operational knowledge with re-use and sharing provisions with respect to building their own TSES. Previous literature in user-developed application (UDA) suggested that, it is appropriate to design a solution environment which could be useful for integrating design elements and processes instead of developing a single solution for the problem in a specific domain (Gammack, 2002). Consequently, many works have been initiated to empower end user solution development. For instance, Clarke (1997) designed a user-enabled framework using a Lotus Note based system which accommodated different inputs from users and knowledge from multiple sources for rapid information processing towards a solution. The type of user-enabled system Clarke (1997) discusses allows for iterations of process rather than enabling changes in the process of solution making in the application. In addition, this Lotus Note based system does not have many user provisions which could allow for the construction of the whole solution application. A few design environments for building decision support tools have been outlined over the past few years. For instance, Pop and Neugru (2003) proposed a development environment for constructing expert system that helped design professionals in the process of improving old implementation using latest technical features; Kim et al. (2000) outlined a programming environment for developing DSS in Prolog, that offered a set of dialogues to build and manage various components of DSS and top-level control components in the Prolog programming environment. In addition, Prebil and Kaiba (2002) described an object-oriented programming environment for originating proprietary expert systems in which designers can take care of input data, technical regulations relevant to customers and other parameters during the design process. However, these types of solution environments are more suitable for technical designers rather than end users who want to develop their own solution through the selection of relevant variables and using their own judgement. Incorporating different AI techniques is widely recognized as a way of empowering users with respect to developing systems in many previous studies of rural application development. Some examples of AI techniques include: ES development using rule based methods by Mahaman et al. (2003) and Plant and Vayssieres (2000); ES development using Knowledge based methods by Cohen and Shoshany (2002) and Girard and Hubert (1999); and ES development using neural network based method by Fu (1998). However, giving end users such empowerment in DSS development is limited. Though, not a novel concept in other domains such as banking applications, an earlier design environment called IDIOMS (Gammack et al. 1992) filled the gap by empowering end users in the development of DSS. The IDIOMS system uses a constraint-based knowledge representation in which decision making rules are derived from databases rather than qualitatively from human experts. The end users could then form a DSS model using the rules derived from the database. In particular, IDIOMS has resolved several limitations associated with traditional ES development. These limitations include a lack of context sensitivity, obligation of intermediary to engineer the system and systems build-in obsolescence. Our work advances the IDIOMS concept by utilizing the method of problem ontology for specifying decision requirements of end users in such a way that domain experts can assist directly. A set of expert rules for decision making is determined from domain experts, which are then used to generate the target-relevant ES according to the end users decision making requirements. Although in our case, the expert rules are often heuristic. This system offered a core technique that gives end users the opportunity for the rules creation or modification in a generic way. Our approach differs from the traditional AI techniques approach as well as IDIOMS approach in that rules are developed from heuristics rather than from databases. Another object of this approach is to discover the feasibility of the proposed design environment in the rural business by focusing on giving as much empowerment as possible to end users in solution building for rural applications. In the main, ES development for rural applications provides decision support solutions in three directions or tracks. One track is the advisory systems which focused on replacing human expertise to enable decision making. In this example we can include the studies of Gillard et al. (1997), Mansingh, et al. (2005), and Mahaman et al. (2003). Another track in ES development is the use of diagnostic support tools. These focuses on symptom/clue based knowledge stored in the system. Some example includes Kramers et al. (1998); Li et al. (2002); Potter et al. (2000); Tomson (2000) and Yialouris & Sideridis (1996). Another track in ES development is planning and management support tools. These focus on management support knowledge. Example of this approach include the studies by Ellison et al. (1998); Lokhorst & Lamaker (1996); Matthews et al. (2006); Mohan & Arumugam (1996) and Nuthall & Hurley (1996). Findings suggested that most ES development uses rules based methodology although they used different AI techniques varying from agent-based to case-based decision support approaches. However, rules based methodology is limited especially when frequent changes are required. We adopted the rule based method using case based reasoning where the decision making rules can be modified, reused and shared in different cases. Girard and Hubert (1999) suggested that strategies that farmers can use to manage their farming operations require advisors to understand and apply improved ways of achieving the required objectives. This system supports this strategy but reduces need for on the spot advice from the advisor. Using this design environment, a domain expert extension professional can provide their expert suggestions with respect to configuration and the setting out of farm-relevant decision making parameters. They can also provide relevant instances, and expert rules of thumb to make the design environment ready/fit within a decision making case of a specific farming operation. Subsequently, the system could be delivered to end-users for further building of their target-relevant ES. In this instance, the system promotes use of domain expertise knowledge to estimate the required/expected level in production using practical ‘rules of thumb’ rather than use of traditional mathematical or simulation based approaches (i.e. study by Hart et al. 1998 and Kerr et al. 1999). The developed target-specific ES contributes to a new way of providing decision support solutions as the human expert provides solutions with different classes of problems. In specific farming operations, the system provides an assessment of the current situation by determining the scope for operational improvement. It also identifies changes in the potential for improvement and gives an estimate of the benefits associated with each improvement as well as providing expert opinions on the current status of the farming situation by exploring possible cost improvements. The proposed approach focuses on specific ES development issues in rural applications. In particular, our solution aims to facilitate a new technology over the conventional ES development approach for rural business. The application addresses such issues as systems rigidity, end users subjectivity in the context of use, obsolescence, intermediary requirements and imbalances in problem solving approach between end users and designers. Our approach also offers solution to reduce the additional rural DSS design issues identified by Cox (1996) such as the need for an analytical phase in decision support tools development, resolution, validation and appropriateness in relation to intended purpose. Moreover, in order to avoid over-engineered solutions, the proposed approach offers customisable but simple and feasible solution guidance in ES development for rural application. It also has a generic capability to fit in other livestock business domains. The rest of the paper is organised as follows. Section 2 presents a general description of the research problem. Section 3 describes research methodology and knowledge acquisition procedures. Section 4 presents the system architecture. Section 5 describes the development procedures for the proposed design environment. Section 6 presents the system operations. Finally, Section 7 contains a brief discussion and summary of the development work. 2. Description of the research problem The Australian Dairy Industry was deregulated in the year 2000 and this has greatly affected dairy operational practices. As a result, dairy business operators have had to explore more effective way of decision making, for example farmers receive incentives from the milk factories for higher quality milk at different seasons of the year and this is a different system to that of the past where the price of product was guaranteed provided minimal standards were met and production was based on a constant year round supply. The post deregulation incentives depend on the content of milk protein and extra payments are made for content above the minimum standard. Before deregulation, the milk factories provided a guaranteed price for quota milk and as a result, dairy farmers had less difficulty in allocating resources for milk production. However, in the recent deregulated environment dairy farmers have had to adopt changed strategic and operational decision making approaches revolving around how to cut the cost of producing milk while increasing production and taking advantage of the incentives offered by factories. As a result, dairy farmers need to think about how to enhance the quality of milk, for example increasing the percentage of milk protein. Expert advice on whole farm management factors such as cow feeding plans, calving plans and cow management can help dairy farmers attain levels of milk protein that can assist in profit maximization. 3. Materials and methods A case study approach was adopted for conducting the research. The object was to acquire an in depth understanding of the decision-making aspects of dairy stakeholders. Multiple qualitative techniques were used for knowledge acquisition and verification, these included techniques such as focus group meetings which were convened over a two to three hours time period. Using heterogeneous approaches, we organised meeting sessions that integrated personnel from different but relevant expertise areas. These areas included dairy practitioners, dairy breed specialist, dairy system specialist and farmers. We found the participants were enthusiastic about the subject matter and were willing to share their views and knowledge in order to develop this tool. A relatively less-structured approach (Morgan, 2002) used for conducting focus group session where the goal was to understand participants thinking and the aim was to explore ideas in new directions rather than getting the answers for specific points. This facilitated interactions between the experts in the meeting and discussions helped us to acquire a clear understanding without any conflicts in expert’s opinions. For knowledge verification, we employed the convergent interviewing technique (Dick, 2002) in which informal interviewing was used to ask experts individually about their specific knowledge. As an initial strategy, we surveyed current literature with the aim of characterising the key issues in dairy operations associated with the problem domain. Studying internal documentations and booklets enable us to find the appropriateness of selected milk protein enhancement techniques for our study. Using this initial knowledge, we convened three focus group meetings on cow nutrition and management each of two to three hours duration. In these sessions we covered the impact of milk protein on feed dry matter intake, metabolisable energy content, and other feed values, farm management including the impacts of selected feed items, feed processes, feed use and limits imposed. Other factors considered included heat stress and breed management; especially the impacts of heat stress and breed types on milk protein. The first focus group session incorporated skilled personnel such as dairy nutritionists, experienced dairy practitioners and dairy farmers. The second focus group session incorporated skills from personnel such as experienced dairy practitioners, dairy physiologists, and dairy farmers. The third focus group session incorporated skilled personnel such as a dairy breed specialist, experienced dairy practitioners and dairy farmers. To verify the extracted knowledge, we used convergent interviewing with domain experts. In this approach, an informal one-to-one meeting was conducted. The elicited expert knowledge allowed us to identify the required decision making parameters, their impacts and relationships with different instances that lead to the development of the general problem ontology. At the same time that the knowledge acquisition process is operating, a prototype development of the solution implementation has begun for end user and expert’s assessment. This process involves two phases. The first being a spreadsheet based development of an expert system for demonstrating the extracted knowledge and concepts to the domain experts. This spreadsheet based ES is used as a vehicle to demonstrate concepts as suggested by Karire (2000). Karire (2000) suggests that the spreadsheet is widely used modelling technique for DSS and ES development. The developed spreadsheet based ES forms the basis of the target-relevant ES which is expected to be produced from the design environment. The other involves the development of a generic template of a design environment from the proof of conceptual knowledge which is demonstrated in the spreadsheet based ES. Development of the design environment is scheduled for four different phases (details in development approach section).By this way, the validated knowledge is used for designing the generic environment as a final deliverable prototype to the dairy industry. As our proposed solution environment incorporates bi-directional activities by the domain experts and end users, we used forward chaining and backward chaining strategies for defining the differentiated tasks in terms of data- driven (forward chaining method) and goal-driven strategies (backward chaining method). For instance, a forward chaining method was used as the domain experts entered the decision making parameters in our development platform. Mansingh et al. (2005) suggested that a forward chaining strategy with data-driven reasoning is simple straightforward when knowledge is required to be modelled. On the other hand, a goal- driven strategy was used as end-users utilised the developed knowledge-base in our development platform. In addition, end users can be expected to use their developed ES for more than one purpose and domain experts may have less control of it, therefore, a backward chaining method is suitable. End user could consider all possible decision making parameters for obtaining expert suggestions in the relevant problems they are facing. For this instance, use of the backward chaining method (goal-driven strategy) for end users is justified in that the system is given a specific solution or expert suggestions after determining the scope that is needed to be improved. 3.1 Knowledge acquisition The knowledge acquisition in this study included a successful knowledge extraction and elicitation from multiple experts. Literature from Stair and Reynolds (2005) suggests that knowledge extraction from multiple experts is problematic when the experts disagreed, however we did not encounter any conflicts between experts during the knowledge acquisition process in this study. Our elicitation process resulted in six main components relating to the problem domain as shown below; 1. Dry Matter levels in feed 2. Feed values 3. Water management 4. Herd management 5. Heat stress management and 6. Breed management. These are consistent with 14 factors (i.e. average body weight, days in milk, major breed, average milk protein percentage, milk per day/cow, ABV1, water availability per day, frequency of feeding, feeding access to pasture, average day temperature, relative humidity, distance between the shade and feeding area, approximate shade length, and sprinklers) in three instances (animal instance, management instance, climate instance) that are dominant for milk protein enhancement in dairy operations (Milk Protein handbook - DPI, 2006). The production rules developed from the expert’s opinions are classified into six main relationships (figures 1 and 2). We identified six main aspects associated with milk protein enhancement each of which has different impacts on potential protein levels in milk. These aspects were considered relevant by all dairy stakeholders. For example, by providing sufficient dry matter in a cow’s feed milk protein was improved by a maximum of 0.3%. Each of these six aspects had a limit to the maximum expected potential milk protein and are used as building-blocks for the problem ontology. 1 ABV-Australian Breed Value is a calculation that estimates a cow’s breeding values or its worth for breeding future dairy cattle. A significant issue experienced during the knowledge acquisition process was caused by communication difficulties with domain experts. Li et al. (2002) also pointed to a similar bottleneck from Liebowitz and Baek (1996), which was mainly cased by the inability of the expert to explain the knowledge and inability of knowledge engineer to obtain expertise. We experienced issues associated with the second cause. Li et al. (2002) suggested that issues using multiple data collection techniques could handle this bottleneck. However, we used three strategies to minimise this issue in our project. These were an earlier survey on current relevant literature, use of an intermedieraty during focus group sessions, and to demonstrate the developed prototype in the focus group sessions for review of the concepts. In the second strategy, we recorded the meeting discussions and used a person who has expertise in the problem domain to assist the knowledge engineer to understand and categorise the recorded knowledge. Though knowledge extraction from multiple experts is problematic, particularly when experts disagree (Stair and Reynolds, 2005), during our knowledge acquisition process, the reasons behind the non-conflict between the experts were. (1) The nature of dairy operational knowledge is justified in the science and this is accepted by all relevant experts and (2) Experts have already realised that the proposed system will have generic capabilities and will be changeable through their own justifications, so the extracted rules can be recreated or modified by any domain expert. 4 System Architecture EUEDE serves two levels of functional processes in that it allows end user DSS development by allowing domain experts to build relevant knowledge bases where the decision making parameters, instances and rules are developed for the design environment. This is an expert’s pre-setting activities which are important for end users who is less experienced in dairy farming overall. It reduces risks associated with selecting wrong parameters and rules from observed relationships. Accordingly, the second process involves development activities that allow end users to build their own decision support tools for evaluating their specific farming situation and obtain expert suggestions for further improvement. In this instance, end users such as farmers have provision to use their own set of decision making parameters to build the target-specific decision support tools. The EUEDE architecture consists of three key modules which ensure the generic integrity of the system. The elements are a knowledge base module, an expert system development module, and a problem ontology module. The following section describes each module in the design environment. The knowledge base module serves as intermediate module between the other two modules. The main role of this module is to assist in knowledge maintenance and the storing of the elicited knowledge. It helps in the creation of new parameters and rules relevant to specific farming conditions for the end user. It also holds the decision making parameters (DMP) table that is calibrated/adjusted by the domain experts. The parameters in this table are relevant to the system with underlying rules that have impact on specific farming aspects. This module displays the DMP table that allows end users to choose their target-relevant DMP’s for building their specific decision support tools. The problem ontology module allows domain experts to create a knowledge base (KB) for the specific decision making context on a farm. Chan (2005) argued that an ontology model can be used as the basis for building knowledge acquisition tools, which in turn gathers domain knowledge for constructing the KB. Functions incorporating four sub-modules are available to help domain experts outline the appropriate problem ontology for the specific domain. These functions are designed to add new parameters or to modify the DMP, defining rules in different relationships, add or modify expert suggestions and test the KB to verify expert outcomes from the system. Knowledge base Module ES Creator Module Problem Ontology Module End-users interact with range of DMP’s D om ain E xperts acquire and create K B DMP table DMP table Add new/modify DMP Define Rules in different relationships Testing KB Add/modify expert suggestions Selecting useful DMP ES building with target-relevant DMP Use Scope Finder Use Potential Estimator & Cost Calculator Obtaining expert treatments on specific farm situations Knowledge Modelling Knowledge reuse and share DMP- Decision Making Parameters KB- Knowledge Base Figure 1: System architecture of EUEDE Finally the ES development module generates the target-relevant ES automatically by corresponding with end users DMP selection. This module involves sequential activities such as DMP selection, ES building with the selected DMP’s, scope finder, potential estimator and cost calculator and obtaining expert suggestion for specific improvements. 5. Development of EUEDE There were three phases involved in the development process. In the first phase of development, a set of decision making parameters and the expert’s ‘rules of thumb’ were estimated from the focus group meeting and verified by DPI’s internal documentation and the Milk Protein handbook – DPI (2006). Subsequently, the project gathered the required expert knowledge to develop a sharing knowledge base for an ES in the domain of milk protein enhancement. The developed ES which was outlined from the extracted parameters and rules has been demonstrated in order to obtain end user verification. In the next phase of development, the parameters used in the ES gave us an opportunity to discover the generic components such as parameter definitions, the generic weight of parameters, the generic relationships/rules and the generic outcomes associated with decision making in the rural business domain. This first ES led to the development of the problem ontology. The developed generic components or building blocks were then transformed into a design environment architecture in which the domain experts could calibrate the system to specific farming conditions. The developed design environment was then verified with the end users. This phase also involved a final demonstration to the industry. The complete development procedure is shown in the following figure (figure 2). Initial documents Evolutionary Prototyping Building of ES: expert solution and different functions and rules Knowledge acquisition from Domain Experts Demonstration on Focus Group Meeting for Discussion Satisfy Finding generic parameters in decision making Verify Develop a Design template of EUEDE Verify Development of generic EUEDE End Users Design Philosophy and standard Feedback on the design Demonstration to relevant/rural experts End users Assessment No Yes Yes No No Yes Experts & end users Assessment Theoretical Assessment Building problem ontology Feedback on the design Specification of generic EUEDE ES development phase Development of generic template phase EUEDE Development phase Generic design Philosophy and standard Theoretical Assessment Figure 2: EUEDE development procedure through rapid development activities Successful application development through evolutionary prototyping has been reported in rural business (Bloom & Chung, 2001; Kerr & Winklhofer, 2005; Quintero et al., 2005). These studies suggest that the key to successful prototyping development is the user’s active participation in the development phases. We utilized this practice of end users participation for the development to ensure design acceptance rather than reporting user’s requirements. However, studies by Bloom & Chung (2001) and Kerr & Winklhofer (2005) confirm that rapid prototyping seems to work well as it supports the key elements of the development processes such as integration, reusability and the finding of generic components which is the core knowledge in our study. In this project, the system requirements and project goals allow us to adopt the approach of evolutionary prototyping. Our development processes also rely on the traditional system development life cycle approach (Alter, 2001) because of the heavy emphasis on iterative process in every phase of development. Figure 3: Expert results from the spreadsheet based ES. The reason behind building the ES initially was to demonstrate the acquired knowledge in a meaningful way so that domain stakeholders can understand how the design environment template will work. In this ES, domain experts acquire the data and create the knowledge base on specific farming condition by adding decision making parameters, rules and expert’s suggestions. Subsequently, the system is ready for end-users who can use those pre-settings to build their specific decision support tools selecting suitable decision making parameters, instances and production rules. Figure 3 shows expert results for farmers while applying preset specific farming conditions in our initial spreadsheet based ES. 5.1 Problem ontology construction We engaged an approach for ontology construction called METHONTOLOGY for ontology development (Fernandez et al. 1997), which advocates the use of a structured informal representation to support the ontology development (Bally et al. 2004). Depending on particular problem task and development point of view, Chen (2005) distinguishes ontology into top level ontologies, domain ontologies, task ontologies and application ontologies. A top level ontology explains general concepts using class, event and process that are independent of the particular problem or domain. This type of ontology consists of common terms which are understood by a large number of users. Domain ontology describes the vocabulary that relates to a generic domain or specific task in the domain. The tasks could be monitoring, diagnosis and/or control. Task ontology is similar to domain ontology. Application ontology explains concepts which are dependent on a particular domain and task that are often a specialization of both the related ontologies. According to the above classification we adopted domain ontology development as our aim was to organise the domain knowledge so that it could be reused and shared. However, development of domain ontology is difficult because it requires detailed knowledge to map the relationships between the knowledge elements. We managed to acquire and configured the knowledge elements to outline a generic way of modelling the domain knowledge. We attempt to reuse this modelling concept for the other domains where similar characteristics of the knowledge will be experienced. Our argument in this paper is to provide this approach for all extensive livestock based rural business. The set for developing domain ontology is associated with the problem ontology development for our solution environment where different independent knowledge elements were functional for configuration. This was used as the basis for the reusable generic EUEDE. The innovative aspect of the software solution is able to fit into other domains. 5.2 Generic feasibility of the system The innovative aspect of the proposed EUEDE solution is that the architecture could be re-useable into other rural business domains such as beef cattle and sheep because, the decision making aspect in the production systems in these types of livestock-based rural business are very similar. For instance, the decision making rules (rules of thumb) associated with water supply in the dairy business domain could be applied to the beef cattle industry by changing the required amount of water for beef cattle. However, in the other domain, our proposed solution architecture consists of four generic levels for domain experts. The first one is associated with the level for adding or removing parameters where the domain expert provides domain knowledge in terms of its building-blocks such as parameters, instances, classes and relationships. The second one is associated with the level for rules production, where the domain expert defines the decision making rules using ratios. The third one is associated with setting potential levels where the domain expert sets up potentials for target outcomes. The fourth one is associated with the ability to add/remove expert suggestions where the domain experts add suggestions for potential improvements. The overall concept is to use this proposed system in such domains where some specific factors are important for potential businesses. The parameters that are obtained from external or internal sources have impacts on the business potential production levels. In most cases in the livestock based business, if we could tailor some factors with different business potential production levels and subsequently adjust the rules between the factors and the potentials using expert suggestions for improvement for each specific potential, then the system architecture could be useful for expert decision making by business operators. This makes the developed knowledge base in the EUEDE system a generic function in which the knowledge components are shareable and changeable. 6. System operation The following figures are screen shots of the EUEDE design environment. For instance figure 4 illustrates the welcome window of the EUEDE where all users are required to agree with the terms and conditions for system use. Figure 5 illustrates the decision making parameters (DMP) table where domain experts update domain- relevant decision making parameters, their instances, identified potential classes and relationships between the parameters and classes. This DMP database is used as storage of knowledge components of domain knowledge entered by experts. This database is also frequently re-used to create production rules by the domain expert. Figure 4: EUEDE version 2 template Figure 6 illustrates the function for creating production rules by the domain experts. Experts select a potential class relevant to a parameter. Subsequently they define the ratio and associated mathematical operators. For example: if an expert creates rules for the required amount of DM for dairy cows, they could select DM level as a potential class from the class drop-down list. Then they can select body weight as a parameter from the parameters drop-down list, after they define a ratio (could be 3.2%) between the class and the parameter. Finally the expert will select mathematical operators for the defined rules. When the expert clicks on the save rules button (figure 6), produced rules will be stored and used as shown below: Required DM = body weight * 3.2 % Figure 5: the DMP table displays the parameters and instances In the developed ES, this required DM level (the above developed rule) will be compared with the current DM level to determine differences. This will indicate that there is either scope or no scope to improve the DM level. Consequently, when the end-user developed ES uses this parameter; it uses expert opinions with respect to DM level improvement. Currently EUEDE is under the process of continuos development. Figure 6: Windows where the production rules are being created in EUEDE 7. Discussion and conclusion This paper described the development of a design environment where dairy professionals assist farmers to build their own TSES. Our proposed design architecture is different from the standard commercial ES shells (for example, EXSYS, ACQUIRE) which often encompass tools for the design, development, and testing of the knowledge base. In most of the cases, the traditional ES shells contain a specified format in which a generic interface engine and user interface is designed to integrate with a domain knowledge base. Knowledge engineers can use these shells to build ESs but they have limited options for the end user to tailor the system to their own requirements. In our case, the diversity of EUEDE not only offers options for the acquisition and structuring of the domain knowledge but also offers a way of tailoring the system to cater for rapidly-changing requirements demanded by end users. The developed ES gives expert support to improve farming operations. One of the objectives of this paper is to document how the domain knowledge is used and modelled for the development of an ES for rural applications. On the other hand, using a generic problem ontology based solutions enables us to modify this model for use in other similar rural business operations, specifically extensive livestock based industries. This approach aims to facilitate a new technology over the conventional DSS development tools used for rural business applications. This new approach also aims to overcome aspects such as systems rigidity, end users subjectivity in the context of use, obsolescence, intermediary requirements and differences in problem solving approaches between end users and designers. Acknowledgements This research was supported by an Australian Research Council linkage grant and the authors wish to acknowledge the financial support from the Australian Research Council. In addition, the authors wish to acknowledge the financial and in kind support provided by the Queensland Department of Primary Industries, the industry partner in this linkage project. References Alter, S. (2001). Which life cycle- work system, information system or software?, Communication of the association for information systems, 7(17), School of business and management, University of san Francisco, USA Bally, J., Boneh, T., Nicholson, A. E., & Korb, K. B. (2004). Developing an ontology for the Meteorological Forecasting Process, Decision Support in an Uncertain and Complex World: The IFIP TC8/WG8.3 International Conference, Retrieved on 20th April, 2007 from URL: http://vishnu.sims.monash.edu.au:16080/dss2004/proceedings/pdf/08_Bally_Boneh_Nicholson_Korb.p df, Bloom, P.C. & Chung, Q.B. (2001). Lessons learned from developing a mission critical expert system with multiple experts through rapid prototyping, Expert systems with Application, 20, 217- 227. Chen, C.W. (2005). An expert decision support system for monitoring and diagnosis of petroleum production and separation processes, Expert Systems with Applications, 29,131-143. Clarke D. (1997). Hurray for real time management! (Response), Retrieved on 20th April, 2007 from URL: http://www.learning-org.com/97.05/0048.html Cohen, Y., & Shoshany, M. (2002). A national knowledge-based crop recognition in Mediterranean environment. International Journal of Applied Earth Observation, 4, 75–87. Dick, B. (2002). Convergent interviewing. Session 8 of Areol - action research and evaluation on line, Retrieved on 20th April, 2007 from URL:http://www.scu.edu.au/schools/gcm/ar/areol/areol- session08.html DPI (2006), Protein plus Check book: A guide to increasing milk protein % and profit, Department of Primary Industries and Fisheries, Queensland Government, Australia Ellison, P., Ash, G., & McDonald C. (1998). An expert system for the management of Botrytis cinerea in Australian Vineyards. I. Development, Agricultural Systems, 56, 185-207. Fernandez, M., Gomez-Perez, A. & Juristo, N. (1997). METHONTOLOGY: From Ontological Art Towards Ontological Engineering, Presented to AAAI97, Workshop on Ontological Engineering, Stanford University, 33-40. Fu, Y., & Shen, R. (2004). GA based CBR approach in Q&A system. Expert Systems with Applications, 26, 167–170. Gammack, J. (2002). Constructive Design environments: Implementing End-users Systems Development, Human Computer Interaction and Management, Idea Group Industries, IRM press, pp 153-173. Gammack, J.G., Fogarty, T.C., Battle, S.A., Ireson, N.S. & Cui, J. (1992). Human Centered Decision Support: The IDIOMS System, Journal of AI & Society, 6, 345 – 366. Gillard, P., Sale, P.W.G., & Tennakoon, S.B.(1997). Building an expert system to advise on the use of reactive phosphate rock on Australian pasture, Australian Journal of Experimental Agriculture, 37,1077- 1084. http://vishnu.sims.monash.edu.au:16080/dss2004/proceedings/pdf/08_Bally_Boneh_Nicholson_Korb.pdf http://vishnu.sims.monash.edu.au:16080/dss2004/proceedings/pdf/08_Bally_Boneh_Nicholson_Korb.pdf http://www.scu.edu.au/schools/gcm/ar/areol/areol-session08.html http://www.scu.edu.au/schools/gcm/ar/areol/areol-session08.html Girard, N. & Hubert, B. (1999). Modelling expert knowledge with knowledge-based systems to design decision aids the example of a knowledge-based model on grazing management, Agricultural Systems, 59, 123-144. Hart, R. P.S., Larcombe, M. T., Sherlock, R. A., & Smith, L. A. (1998). Optimisation techniques for a computer simulation of a pastoral dairy farm, Journal Computing and Electronics in Agriculture, 19:129-153. Kerr, D., and Winklhofer, H. (2005). The effect of rapid rural industry changes on the development of a decision support system for dairy farmers in Australia, Computers and Electronics in Agriculture, 50, 61-69. Kerr, D., Cowan, R.T., & Chaseling, J. (1999). DAIRYPRO- a knowledge based decision support system for strategic planning on sub-tropical dairy farms. I. system description, Agricultural Systems, 59, 245-255. Kim, G., Nute, D., Rauscher, H. M., & Loftis, D. L., (2000). AppBuilder for DSSTools: an application development environment for developing decision support systems in Prolog, Computers and Electronics in Agriculture, 27, 107-125. Kramers, M.A., Conijin C.G.M., & Bastiaansen, C. (1998). EXSYS, an expert system for diagnosing flowerbulb diseases, pests and non parasitic disorders, Agricultural Systems,58,57-85. Kreie J, Cronan T.P., Pendley J. & Renwick J.S. (2000). Applications development by end-users: can quality be improved? Decision Support Systems, 29,143-152. Li, D., Fu, Z & Duan, Y.(2002). Fish-expert: a web based expert system for fish disease diagnosis, Expert systems with Applications, 23,311- 320. Liebowitz, J., & Baek, S. I. (1996). The protocol multimedia expert system, The New Review of Applied Expert Systems,1, 3–17. Lokhorst, C. & Lamaker, E.J.J. (1996). An expert system for monitoring the daily production process in aviary systems for laying hens, Computers and Electronics in Agriculture, 15, 215-231. Mahaman, B.D., Passam, H.C., Sideridis, A.B., & Yialouris, C.P. (2003). DIARES-IPM: a diagnostic advisory rule – based expert system for integrated pest management in Solanaceous crop systems, Agricultural Systems,76, 1119-1135. Mansingh, G., Reichgelt, H., & Bryson, K.M.O. (2005). CPEST: An expert system for the management of pests and diseases in the Jamaican coffee industry, Expert systems with Applications, 29th December (article in press). Matthews, K.B., Buchan, K., Sibbald, A.R., & Craw, S. (2006). Combining deliberative and computer based methods for multi objective land use planning, Agricultural Systems,87, 18-37 McGill, T.J. (2002). User developed applications: can end users assess quality?, Journal of end user computing, 14, 15. Mohan, S., & Arumugam, N. (1996). Expert system applications in irrigation management: an overview, Computer and Electronics in Agriculture, 17,263-280. Morgan, D. L.(2002). Focus Groups as Qualitative research, second edition, Sage Publications Inc, California Nuthall, P.L., & Bishop-Hurley, G.J. (1996). Expert systems for animal feeding management part II: Farmers attitudes, Computers and Electronics in Agriculture,14, 23- 41. Plant, R., E., & Vayssieres, M., P. (2000). Combining expert system and GIS techniligy to implement a state-transition model of oak woodlands, Computer and Electronics in Agriculture, 27, 71-93. Pop, D. & Negru, V. (2003). An Extensible Environment for Expert System Development, Department of Computer Science, University of the West from Timi_oara ,3 V. Pârvan Street, RO-1900 Timioara, Romania, Retrieved on 20th April, 2007 from URL: http://www.info.uvt.ro/~petcu/infosoc3/KES03.pdf Potter, W.D., Deng, X., Li, J., Xu, M., Wei, Y., Lappas, I., Twery, M.J., & Bennett, D.J. (2000). A web- based expert system for gypsy moth risk assessment, Computers and Electronics in Agriculture, 27, 95- 105. Quintero, A., Konare, D., & Pierre, S. (2005). Prototyping an intelligent decision support system for improving urban infrastructure management, European journal of operational research,162, 654-672. Stair, R. & Reynolds, G. (2005). Fundamentals of Information Systems, Third edition, Thomson course technology, United States, 277-279. Tomson, A.J., Allen, E., & Morrison, D. (1998). Forest tree disease diagnosis over the World Wide Web, Computers and Electronics in Agriculture, 21, 19- 31. Wagner (2002) End Users as Expert System Developers?: Journal of End User Computing, 12, (3). Yialouris, C.P., & Sideridis, A.B. (1996). An expert system for tomato diseases, Computers and Electronics in Agriculture, 14, 61-76. http://www.info.uvt.ro/%7Epetcu/infosoc3/KES03.pdf 
work_pqny4mzrtbhwzm3pbmam6e6erq	----	International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 420 Feature Selection for Intrusion Detection System Using Ant Colony Optimization Mehdi Hosseinzadeh Aghdam1, and Peyman Kabiri2 (Corresponding author:Mehdi Hosseinzadeh Aghdam) Department of Computer Engineering and Information Technology, Payame Noor University (PNU)1 P.O. BOX 19395-3697, Tehran, Iran School of Computer Engineering, Iran University of Science and Technology2 Tehran, 16846-13114, Iran (Email: mhaghdam@iust.ac.ir) (Received March 30, 2014; revised and accepted Jan. 16 & Mar. 4, 2015) Abstract Intrusion detection is a major research problem in net- work security. Due to the nonlinear nature of the intru- sion attempts, unpredictable behavior of the network traf- fic and the large number of features in the problem space, intrusion detection systems represent a complicated prob- lem area. Choosing effective and key features for intrusion detection is a very important topic in information secu- rity. The purpose of this study is to identify important features in building an intrusion detection system such that they are computationally efficient and effective. To improve the performance of intrusion detection system, this paper proposes an intrusion detection system that its features are optimally selected using ant colony op- timization. The proposed method is easily implemented and has a low computational complexity due to use of a simplified feature set for the classification. The extensive experimental results on the KDD Cup 99 and NSL-KDD intrusion detection benchmark data sets demonstrate that the proposed method outperforms previous approaches, providing higher accuracy in detecting intrusion attempts and lower false alarm with reduced number of features. Keywords: Ant colony optimization, feature selection, in- trusion detection system 1 Introduction Intrusion Detection Systems (IDSs) have become impor- tant and widely used tools for ensuring network security. In recent years, intrusion detection based on statistical pattern recognition methods has attracted a wide range of interest in response to the growing demand of reliable and intelligent IDSs, which are required to detect sophisti- cated and polymorphous intrusion attacks [7]. In general, IDSs are usually classified into two categories in the liter- ature: signature-based intrusion detection and anomaly- based intrusion detection [33]. Signature-based intrusion detection, misuse detection, approach can reliably iden- tify intrusion attacks in relation to the known signatures of discovered vulnerabilities. Therefore, well-known in- trusions can be detected efficiently with a very low false positive rate. Intrusions are usually polymorphic, and evolve continuously. Signature-based detection will fail easily when facing unknown intrusions and emergent in- tervention of security experts is required to define signa- tures or accurate rules, which limits the application of the signature-based detection approach to build intelli- gent IDSs. Signature-based detection is commonly used in the design of commercial IDSs [38]. Anomaly-based IDS usually deals with statistical anal- ysis and pattern recognition problems. It can detect zero day attacks if the classification model has the generaliza- tion capability to extract intrusion pattern and knowledge during the training period [33]. In the anomaly-based in- trusion detection approach, the system builds models for normal behavior in the network traffic and detects any de- viation from this model as a potential intrusion attempt. This approach commonly suffers from high false positive rate on classifying normal network traffic. Due to the dy- namic nature of the network traffic, discovering bound- aries between normal and abnormal behavior is a major difficulty in this approach [38]. To overcome the anomaly intrusion detection problem, the data mining [14], ma- chine learning [25] and immune system [37] approaches have been proposed in recent years. Considering the available computational power and due to the large amount of audit data with a large number of features that IDS needs to examine, even for a small network, traffic analysis is a difficult task. Audit data captures various features of the connections. For exam- ple, the audit data would show the network service on the destination (http, telnet, etc.), or the number of wrong fragments or length of the connection (second). Selected features in the problem space might be correlated, which International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 421 is difficult for humans to discover. An IDS operating in real-time needs a reduced amount of data for the process- ing. Some of these features are irrelevant and redundant and thus can be eliminated before the processing [7]. Most of these features are not applicable to intrusion detection; even some noise data may badly affect the performance of the process of detecting intrusions. Extra and noisy fea- tures can increase the computation time, and can have a negative impact on the accuracy of the IDS [28]. Hence, we need to select some representative features from the original feature space to reduce the dimensionality of fea- ture space and improve the efficiency and performance of IDS. Feature Selection (FS) is a commonly used step in ma- chine learning, especially when dealing with a high di- mensional space of features. The objective of FS is to simplify a data set by reducing its dimensionality and identifying relevant underlying features without sacrific- ing predictive accuracy. By doing that, it also reduces redundancy in the information provided by the selected features. In real world problems FS is a must due to the abundance of noisy, irrelevant or misleading features. FS is extensive and it spreads throughout many fields, includ- ing text categorization, data mining, pattern recognition, signal processing and intrusion detection [1, 2]. Recently, Ant Colony Optimization (ACO) is gaining popularity as a new approach to the FS [1, 13, 36]. Some literatures provided experimental results that illustrated ACOs superiority over conventional approaches, such as sequential search and the genetic algorithm [13]. Dorigo and colleagues introduced meta-heuristic optimization al- gorithm based on behavior of ants in the early 1990s. ACO is a new solution finding method in artificial intel- ligence called Swarm Intelligence (SI). In SI interaction of cooperative individuals is such that a problem-solving behavior emerges. SI is the property of a system whereby the collective behaviors of unsophisticated individuals in- teracting locally with their environment cause coherent functional global patterns to emerge. Insects such as ants and bees live in colonies. An individual can only do sim- ple behavior on its own, while their colonial cooperative work represents a complex behavior [10]. ACO algorithm is inspired by social behavior of ant colonies. Although they have no sight, ants are capable of finding the shortest route between a food source and their nest by chemical materials called pheromone that they leave when mov- ing [6]. ACO algorithm was originally applied to the travelling salesman problem and later on, it was successfully applied to other optimization problems such as the quadratic as- signment problem [18], routing in telecommunication net- works, graph coloring problems, scheduling, etc. This method is particularly attractive for FS as there seems to be no heuristic that can lead the search to the optimal minimal subset every time [1]. In this paper a novel in- trusion detection system using ACO has been introduced. The classifier performance and the length of selected fea- ture subset are adopted as heuristic information for ACO. Thus, the proposed method needs no prior knowledge of features. Also the classifier performance and the length of selected feature vector are considered for performance evaluation. Finally, the proposed method is applied to KDD Cup 99 and NSL-KDD, the new version of KDD Cup 99, data sets. The rest of this paper is organized as follows. Section 2 outlines the related work about feature selection methods in intrusion detection. The proposed ACO-based method for IDS is described in section 3. Section 4 reports compu- tational experiments. It also includes a brief discussion of the results obtained and finally the conclusion and future works are offered in the last section. 2 Related Work Feature selection is a process that selects a subset of orig- inal features. The significance of FS can be viewed in two facets. The frontier facet is to filter out noise and elimi- nate irrelevant and redundant features. FS is compulsory due to the abundance of noisy, irrelevant or misleading features in a data set. Second, FS can be considered as an optimization problem for an optimal subset of features that better satisfy a desired measure [12]. Quality of the FS optimization can be measured using certain evaluation criteria. Since FS is a NP-hard problem, there is no prac- tical solution to find its optimal feature subset [17]. A typ- ical FS process includes subset generation, subset evalua- tion, termination criteria and result validation. Subset se- lection procedure implements a search method that selects feature subsets for evaluation based on a certain search strategy. It may start with empty subset, full subset, a selected feature subset or some random feature subset. Those methods that start with an initial subset usually select this feature subset using heuristic methods. These search methods include forward selection, backward elimi- nation and forward/backward combination methods. The process of subset selection and evaluation is repeated un- til a given termination condition is satisfied. The selected best feature subset usually needs to be validated using a different test data set [17]. Volume of the network traffic data and the high di- mensionality of the feature space are main causes for pro- hibitively high processing overhead. This is why they are major problems in IDS design. Therefore, FS is an im- portant phase in IDS design. Variation of the normal traffic often hinders anomaly-based IDS ability to accu- rately model normal state of the operation of the network. In any intrusion attempt there are some behavioral pat- terns and interrelations that are unique and recognizable. Since these patterns are hidden within the irrelevant and redundant features it is often difficult to discover them. Eliminating the less relevant features can improve both the speed and the accuracy of the classification [29]. In accordance to the literature, approaches to FS can be divided into filters, wrappers and embedded ap- proaches [12]. The filter model separates FS from learning International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 422 algorithm and selects feature subsets that are independent of any learning algorithm. It relies on various measures of the general characteristics of the training data such as distance, information, dependency, and consistency. In the wrapper approach feature subset is selected using the evaluation function based on the same learning algorithm that will be used later for learning [17]. In this approach the evaluation function finds the optimal feature subset using the subset generation procedure. Finally, compar- ing the result with best subset resulted from the previous iterations subset evaluation procedure only keeps the best subset. Subset evaluation procedure tests the best subset against the termination criteria to determine if the selec- tion process should end. Although, wrappers may gener- ate a better result, their execution cost is high and may encounter problem in dealing with feature spaces with the very high dimensions. This is due to the use of learning algorithms in the evaluation of subsets, some of which can encounter problems while dealing with a high dimen- sional space of features [4, 12]. If the FS and learning algorithm are interleaved then the FS approach is a kind of embedded approach [20]. Feature selection is an important issue in intrusion de- tection. Elimination of the insignificant features may en- hance accuracy of the detection. By concentrating on the most important features execution speed of the process can be increased without significant effect on the accu- racy of detection. Chebrolu et al. employed the effective- ness of IDS in terms of real-time performance and detec- tion accuracy from the feature selection perspective [7]. In the reported work, features were selected using two methods, Bayesian network and classification and regres- sion trees. Four different sets of features were derived and used in their ensemble method for IDS. In their ex- periments, KDD Cup 99 data set is used. A very high detection rate is reported in their experiments. Their ap- proach uses only 5092 records for the training and 6890 records for the test. They have proposed no generic sub- set. Their reported results show that different feature subsets with different length are more effective in detect- ing various types of attacks [7]. Sung and Mukkamala proposed a well-known closed- loop FS method for SVM-based IDS, called SVM-RFE, which recursively eliminated one feature at a time and compared the resulting performance in each SVM test [28]. They also ranked six significant features [29]. They used three methods and compared the performance of these methods in terms of classification accuracy on the test data set. In the reported work, they used support vec- tor decision function ranking, linear genetic programming and multivariate adaptive regression splines [29]. Sung et al. reported a new feature subset that provides accurate results for the detection of different types of attacks [30]. They have used genetic algorithm to maximize inter-class difference and minimize size of the subset. Mukkamala and Sung applied SVM-RFE method to the KDD Cup 99 data and performed the feature ranking for FS [22]. They ranked the features into three categories: important, sec- ondary, and insignificant according to three main perfor- mance criteria: overall accuracy of classification, training time, and testing time. 19 important features were iden- tified and used in the experiments. This heuristic-based technique is time consuming. Additionally, unknown at- tack types are not considered in the reported work since the same data set (kddcup.data-10-percent.gz) was ap- plied to the experiment. Zhang et al. considered the capability of rough set the- ory in coming up with classification rules in detecting the attacks [39]. They showed that rough set classification using GA can produce a high detection rate. Speed of the feature ranking in the reported work is fast. They did not report the selected features used for classification pro- cess. Ohn et al. adopted genetic algorithm to find optimal feature subset for SVM [23]. 31 features were used with radial kernel function in their experiment and a very high detection rate was obtained for the original KDD Cup 99 test data set (corrected.gz). Since their training data set was sampled from the full data set (kddcup.data.gz), the challenge of the problem was reduced. The reason is that the number of attack types in the original training data set (kddcup.data-10-percent.gz) is intentionally em- ployed more than the test data set (corrected.gz). These two data sets can challenge the methods of detecting the unknown type attack. From these reported works, we can conclude that some features are really significant in intrusion detection. Also, it has been proven that there is no single generic classifier that can best classify all the attack types. Instead, in some cases, specific classifier performs better than others. Thus, most of these works lead to an ensemble or fusion of multiple classifier IDS. 3 Proposed ACO-based Method for Intrusion Detection System In the early nineties an algorithm called Ant System (AS) was proposed by Dorigo and colleagues as a novel nature- inspired meta-heuristic approach for the solution of com- binatorial optimization problems. First, algorithm was applied to the traveling salesman problem. Recently, it was extended and/or modified both to improve its per- formance and to apply it to other optimization prob- lems. Improved versions of AS include, among others, Ant Colony System (ACS), MAX-MIN, AS and AS-rank. An ant colony optimization algorithm is essentially a sys- tem based on individuals which simulate the natural be- havior of ants, including mechanisms of cooperation and adaptation. The inspiring source of ACO is the foraging behavior of real ants [9]. The ACO algorithm is based on a computational paradigm inspired by real ant colonies and the way they function. The idea is to use several constructive computational ants. Based on the results of previous experiments stored in the ant dynamic memory structure, each ant is guided to the constructed solution. The paradigm is based on the observation made by ethol- International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 423 ogists about the medium used by ants to communicate information regarding shortest paths to food by means of pheromone trails. While an isolated ant moves practically at random, exploration, an ant encountering a previously laid trail can detect it and decide with high probability to follow it, exploitation, and consequently reinforces the trail with its own pheromone. What emerges is a form of autocatalytic process through which the more the ants follow a trail, the more attractive that trail becomes to be followed. The process is thus characterized by a positive feedback loop, during which the probability of choosing a path increases with the number of ants that previously chose the same path. The mechanism above is the inspi- ration for the algorithms of the ACO family [21]. Before the training phase, a FS phase may also be con- sidered. The FS process identifies which features are more discriminative than the others. This has the benefit of generally improving system performance by eliminating irrelevant and redundant features. In general, FS is not very popular procedure in IDS. However, a few studies use different FS methods for their experiments. This im- plies that FS could improve some certain level of classifi- cation accuracy in IDS. Given a feature set of size N, the FS problem is to find a minimal feature subset of size S (S < N) while retaining the previous accuracy. There- fore, there is no concept of solution path in FS problem. A partial solution, i.e. subset, does not define any or- der among the components, i.e. features of the solution, and the next component to be selected is not necessarily influenced by the last component added to the partial so- lution [5, 15]. FS problem solutions are not necessarily of the same size. The first step in FS using ACO is to address the problem of redefining the way that the ACO representation graph is used. 3.1 Graph Representation The main goal of ACO algorithm is to model a prob- lem as the search for a minimum cost path in a graph. Here nodes can be considered as features, with the edges between them denoting the choice of the next feature. The search for the optimal feature subset is then an ant traversal through the graph where a minimum number of nodes, features, are visited that satisfies the traversal stopping criterion. Nodes are fully connected to allow any feature to be selected next. On the basis of this reformula- tion of the graph representation, the transition rules and pheromone update rules of standard ACO algorithms can be applied. In this case, pheromone and heuristic value are not associated with links. Instead, each feature has its own pheromone value and heuristic value [1]. 3.2 Heuristic Information Generally, the representation of heuristic value is the at- tractiveness of features and the basic ingredient of any ACO algorithm is a constructive heuristic for probabilis- tically constructing solutions. A constructive heuristic assembles solutions as sequences of features from the fi- nite set of features. A subset construction starts with an empty subset. Then, at each construction step the cur- rent subset is extended by adding a feature from the set of features. A suitable heuristic desirability of travers- ing between features could be any subset evaluation. In proposed method classifier performance is mentioned as heuristic information for FS. In other words, the classi- fier accuracy of each feature on training set is considered as heuristic information for each feature. The heuristic information of traversal and node pheromone levels are combined to form the so called probabilistic transition rule, denoting the probability that ant k will include fea- ture i in its subset at time step t: Pki (t) = { [τi(t)] α.[ηi] β∑ u∈Jk[τu(t)] α.[ηu]β ifi ∈ Jk 0 otherwise (1) where Jk is the set of feasible features that ant k can be added to its subset; τi and ηi are respectively the pheromone value and heuristic information associated with feature i. α and β are two parameters that determine the relative weight of the pheromone value and heuristic information. The transition probability used by ACO is a balance between pheromone intensity (i.e. history of previous successful moves), τi, and heuristic information (expressing desirability of the move), ηi. This effectively balances the exploitation-exploration trade-off. The best balance between exploitation and exploration is achieved through proper selection of the parameters α and β. If α=0, no pheromone information is used, i.e. previous search experience is neglected. The search then degrades to a stochastic greedy search. If β=0, the attractiveness (or potential benefit) of moves is neglected. 3.3 Pheromone Update Pheromone updating is an important part for working the ACO algorithm suitably. After all ants have completed their solutions, pheromone evaporation on all nodes is triggered using Equation (2) and then according to Equa- tion (3) all ants deposit a quantity of pheromone, ∆τi(t), on each node that they have used. τi(t) = (1 −ρ)τi(t) (2) τi(t + 1) = τi(t) + ∆τi(t) (3) with ∆τi(t) = m∑ k=1 ∆τki (t) (4) where m is the number of ants at each iteration and ρ ∈ (0, 1) is the pheromone trail decay coefficient. The main role of pheromone evaporation is to avoid stagna- tion, that is, the situation in which all ants constructing the same solution. All ants can update the pheromone ac- cording to Equations (3,4). Where ∆τki (t) is the amount International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 424 of pheromone deposited by ant k on node i at time step t: ∆τki (t) = { ω.γ(Sk(t)) + φ.(n/|Sk(t)|) ifi ∈ Sk(t) 0 otherwise (5) where Sk(t) is the feature subset found by ant k at iter- ation t, and |Sk(t)| is its length. The pheromone is up- dated according to both the measure of the classifier per- formance, γ(Sk(t)), and feature subset length. ω and φ are two parameters that control the relative importance of classifier performance and feature subset length, ω ∈ [0, 1] and φ = 1 − ω. This formula means that the classifier performance and feature subset length have different sig- nificance for FS process. In our experiment we assume that classifier performance is more important than subset length, so they were set as ω = 0.7, φ = 0.3. Algorithm 1 ACO-based Feature Selection Algorithm for IDS 1: Begin 2: Initialize all parameters, i.e. α,β,ρ,m,τ0,φ,ω,T. 3: Let t = 1. 4: for Each node i do 5: τi(t) = τ0. 6: end for 7: Place m ants, k = 1, ,m. // Initialize a population of ants with random positions 8: while t ≤ T do 9: for Each ant k = 1, ...,m do 10: Sk(t) = {} 11: while Ant is able to increase the detection rate do 12: From current node, select next node i using Equation (1). 13: Add node i to subset Sk(t). 14: end while 15: Calculate the subset length |Sk(t)|. 16: Calculate the classifier performance γ(Sk(t)). 17: end for 18: for Each node i do 19: Apply pheromone evaporation using Equa- tion (2). 20: Calculate ∆τi(t) using Equations (4,5). 21: Update pheromone using Equation (3). 22: end for 23: t = t + 1 24: end while 25: Return the subset Sk(t) with highest γ(Sk(t)) as the solution. 26: End 3.4 Solution Construction The overall process of ACO feature selection can be seen in Figure 1. The process starts by generating a number of ants which are then placed randomly on the graph i.e. Figure 1: Overall process of proposed ACO-based IDS each ant begins with one random feature. Alternatively, the number of ants to place on the graph may be set equal to the number of features within the data set; each ant starts path construction at a different feature. From these initial positions, they traverse nodes, features, probabilis- tically until a traversal stopping criterion is satisfied. The resulting subsets are collected and then evaluated. If an optimal subset is found or the algorithm is executed a cer- tain number of times, then the process halts and outputs the best feature subset encountered. If none of these con- ditions occurred, then the pheromone is updated, a new set of ants are created and the process iterates once more. In the proposed approach an IDS consists of several essential parts including feature extraction and FS. Af- ter preprocessing of compressed raw (binary) TCP dump data of 7 weeks of network traffic (DARPA 98), feature extraction is used to transform the input TCP dump data into a feature set, feature vector. FS is applied to the fea- ture set to select more informative features and to reduce the dimensionality of the problem space. ACO algorithm is used to explore the space of all subsets of given fea- ture set. The performance of selected feature subsets is measured by invoking an evaluation function with the cor- responding reduced feature space and measuring the spec- ified classification result. The best feature subset found International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 425 is then output as the recommended set of features to be used in the actual design of the IDS. 4 Experiments and Results The following sections describe the data sets and imple- mentation results. 4.1 Data sets Since 1999, KDD Cup 99 data set from UCI repository is widely used as the benchmark data set for IDS eval- uation [34]. The KDD Cup 99 contained 4,898,431 and 311,029 records in the training set and test set, respec- tively. Each of record contains 41 features and is labeled as either normal or an attack, with exactly one specific attack type. In our experiments, we apply its 10% train- ing data set consisting of 494021 connection records for training. This data set is prepared by Stolfo et al. and is built based on the data captured in DARPA 98 IDS evaluation program [27]. The DARPA 98 intrusion detec- tion evaluation program was prepared and managed by MIT Lincoln Labs. Lincoln Labs set up an environment to acquire nine weeks of raw TCP dump data for a LAN simulating a typical U.S. Air Force LAN. DARPA 98 is about 4 gigabytes of compressed raw (binary) TCP dump data of 7 weeks of network traffic, which can be processed into about 5 million connection records, each with about 100 bytes. The two weeks of test data have around 2 million connection records. All attacks fall in one of the following four categories: • Denial of Service (DoS): Is an attempt to con- sume network resources in such a way that their ser- vices become limited or unavailable for the legitimate users. • User to Root (U2R): Is an attack in which the attacker starts accessing a normal user account on a machine and gains root access to the machine by exploiting vulnerabilities. • Remote to Local (R2L): Occurs when an attacker does not have an account on a remote system, but who has the ability to send packets to a system over a network and exploits vulnerabilities to gain local access as a user of that system. • Probing: An attack in which the attacker scans net- work to collect information about its systems for the apparent purpose of circumventing its security con- trols. KDD Cup 99 features can be classified into three groups: 1) Basic features: this category encapsulates all the attributes that can be extracted from a TCP/IP con- nection. Most of these features leading to an implicit delay in detection. 2) Traffic features: this category includes features that are computed with respect to a window interval. 3) Content features: Most of the DoS and Probing attacks have many intrusion frequent sequential pat- terns, this is due to the fact that these attacks estab- lish many connections to the host(s) in a very short period of time. Unlike these attacks, the R2L and U2R attacks donot have any intrusion frequent se- quential patterns. The R2L and U2R attacks are em- bedded in the payload of the packets, and normally include only a single connection. To identify these kinds of attacks, some relevant features are needed to identify suspicious behavior in the packet payload. These features are called content features [31]. Conducting a thorough analysis of the recent research trend in anomaly detection, one will encounter several machine learning methods reported to have a very high detection rate of 98% while keeping the false alarm rate at 1%. However, when we look at the state of the art IDS solutions and commercial tools, there is few products us- ing anomaly detection, and practitioners still think that it is not a mature technology yet. Tavallaee et al. stud- ied the details of a research in anomaly detection and considered various aspects such as learning and detec- tion approaches, training data sets, testing data sets, and evaluation methods. Their study shows that there are some inherent problems in the KDD Cup 99 data set [31], which, is widely used as one of the few publicly available data sets for network-based anomaly detection systems. The first important deficiency in the KDD Cup 99 data set is the large amount of redundant records. They found approximately 78% and 75% of the records are duplicated in the training and testing sets, respectively. This large number of redundant records in the training set will cause learning algorithms to be biased towards the more fre- quent records, and thus prevent it from learning infre- quent records which are usually more harmful to networks such as U2R attacks. The existence of these repeated records in the test set, on the other hand, will cause the evaluation results to be biased by the methods which have better detection rates on the frequent records [31]. Tavallaee et al. applied 21 learned machines to analyze the difficulty level of the records in KDD Cup 99 data set. They labeled the records of the entire train and test sets, which provide 21 predicted labels for each record. All the 21 methods applied on the data set classified about 98% of the records in the train set and 86% of the records in the test set correctly. Tavallaee et al. reported these statistics on both KDD Cup 99 train and test sets since they have found similar results presented in many papers, random parts of the KDD Cup 99 training set are used as test sets. As a result, these papers obtain about 98% detection rate. Even applying the KDD test set will result in having a minimum detection rate of 86%, which makes the comparison of IDSs quite difficult since they all vary in the range of 86% to 100% [31]. In this paper, both the KDD Cup 99 and more recent and revised version International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 426 of KDD Cup 99 data set, NSL-KDD data set, are used for the experimentation. NSL-KDD data set is publicly available for the researchers and does not suffer from any of the mentioned problems [35]. Additionally, the number of NSL-KDD records in the train and test sets is more reasonable. This advantage makes it a good choice to run the approaches on the complete KDD Cup 99 data set without the need to randomly select a small portion. It should be mentioned that the test set is not from the same probability distribution as the training set, and it includes unknown attack types that do not exist in the training set that makes it more realistic. The data sets contain a total number of 22 training attack types, with an additional 17 types in the test data set [19]. In Table 2, distribution of different attack types in the KDD Cup 99 and NSL-KDD data sets are listed. In the KDD Cup 99 and NSL-KDD data sets there are 41 features (listed in Table 3) suggested for each record. 4.2 Performance Measures The False Positive Rate (FPR) is defined as the number of normal records that are incorrectly detected as intrusions divided by the total number of normal records. The de- tection rate is defined as the number of intrusion records classified by the IDS divided by the total number of intru- sion records present in the test data set. These are good performance measures, since they measure what percent- age of intrusions the system is able to detect and how many incorrect classifications are made in the process. Usually True Positive Rate (TPR) and false positive rate are used for performance measurement. TPR also known as Detection Rate (DR) or sensitivity or Recall and FPR also known as the False Alarm Rate (FAR). They showed in the following equations: Recall = TPR = TP TP + FN (6) Precision = TP TP + FP (7) FDR = FP TN + FP (8) Accuracy = TP + TN TP + TN + FP + FN (9) where TP, TN, FP, FN are the numbers of true positives, true negatives, false positives and false negatives, respec- tively. Another commonly used measure is F-measure that is defined in Equation (10) [26]. F −measure = 2 ×Recall×Precision (Recall + Precision) (10) 4.3 Results A series of experiments was conducted to show the util- ity of proposed method. All experiments are executed on a machine with Intel(R) Core(TM) i7 CPU 3.2 GHz and 4 GB of RAM. The operating system was Windows 7 Professional. All tested models were implemented on MATLAB R2011b. Various values were tested for the pa- rameters of the proposed method. The results show that the highest performance is achieved by setting the param- eters to values as follow: α = 0.4,β = 0.6,ρ = 0.2, the initial pheromone intensity of each feature is equal to 1 (τ0 = 1) the number of ant in every iteration is 50 (m=50) and the maximum number of iterations is 100 (T=100). These values were empirically determined in our prelimi- nary experiments; however, we make no claims that these are optimal values. Parameter optimization is still a topic for future research. Many papers have focused on improving detection rates of the IDSs using efficient classifiers, i.e. this is a quiet difficult approach. This paper puts forward a modified ACO-based FS algorithm aiming at building a classifier to detect intrusion attempts. In the experiments, the whole training set and test set were applied. Each record in the training set or the test set consisted of 41 features. Initially the proposed FS algorithm is used to select im- portant features for each type of the previous discussed attacks. Later on, a classification system was used to clas- sify the attacks were the results are reported. Each fea- ture value is standardized using Equation (11) and then unimportant features are removed, i.e., leaving only the important features, as listed in Table 4. For each type of attacks, using selected features classification of the at- tacked is performed and results are compared with those using all 41 features. Finally, using the results of the com- parisons, their performance in detecting attacks is evalu- ated. StandardScore = X −µ σ (11) where X is a feature value to be standardized, µ is the mean of feature values and σ is the standard deviation of the feature values. Results of classification using the proposed method are reported in Tables 5 and 6. In each class (normal or at- tack), each row shows the performance of the proposed method and baseline approach (using all features) in de- tecting attacks. Experimental results show that FS phase of the process improves the detection rate. In the normal class, studying input features with regard to the output shows that there is no linear relation between the input features and output. Therefore, comparing it versus the baseline approach, implementing the proposed method has significantly improved accuracy of the classification models. The best performance of this system, in terms of its accuracy, is reported to be 98.90%, with only 2.59% false positive rate. Several experiments are performed to compare the two different IDSs. Results show that an IDS combined with the proposed ACO-based algorithm has higher detection rates in detecting attacks than the baseline approach. Figure 2 shows the true positive rate and the false positive rate for the proposed method as we change the number of selected features. The effect of selecting different fea- International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 427 Table 1: KDD Cup 99 and NSL-KDD data sets Data set name # Instances Normal DoS U2R R2L PROBE KDD Cup 99 Training set 494021 97278 391458 54 1124 4107 Testing set 311029 60593 229853 2636 13781 4166 NSL-KDD Training set 25192 13449 9234 11 209 2289 Testing set 22544 9711 7458 533 2421 2421 Table 2: The distribution of attack types in the KDD Cup 99 and NSL-KDD Attack category Attack type KDD Cup 99 NSL-KDD Training set Testing set Training set Testing set kddcup.data10percent corrected.gz KDDTrain+20Percent KDDTest+ Neptune 107201 58001 8282 4657 Smurf 164091 280790 529 665 Pod 264 87 38 41 Teardrop 979 12 188 12 Denial of Service Land 21 9 1 7 (DoS) Back 2203 1098 196 359 Apache2 - 794 - 737 Udpstorm - 2 - 2 Process-table - 759 - 685 Mail-bomb - 5000 - 293 Buffer-overflow 30 22 6 20 Load-module 9 2 1 2 Perl 3 2 0 2 Rootkit 10 13 4 13 User to Root Spy 2 - 1 - (U2R) Xterm - 13 - 13 Ps - 16 - 17 Http-tunnel - 158 - 133 Sql-attack - 2 - 2 Worm - 2 - 2 Snmp-guess - 2406 - 331 Guess-password 53 4367 10 1231 Ftp-write 8 3 1 3 Imap 12 1 5 1 Phf 4 2 2 2 Multihop 7 18 2 18 Remote to Local Warezmaster 20 1602 7 944 (R2L) Warezclient 1020 - 181 - Snmpgetattack - 7741 - 178 Named - 17 - 17 Xlock - 9 - 9 Xsnoop - 4 - 4 Send-mail - 17 - 14 Port-sweep 1040 354 587 157 IP-sweep 1247 306 710 141 Probe Nmap 231 84 301 73 Satan 1589 1633 691 735 Saint - 736 - 319 Mscan - 1053 - 996 International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 428 Table 3: Lists of features in KDD Cup 99 and NSL-KDD data sets No. Feature name Description Type 1 Duration Length of the connection (second) Continuous 2 Protocol-type Type of protocol, e.g. tcp, udp, etc. Discrete 3 Service Network service on the destination, e.g., http, telnet, etc. Discrete 4 Flag Normal or error status of the connection Discrete 5 Src-bytes Number of data bytes from source to destination Continuous 6 Dst-bytes Number of data bytes from destination to source Continuous 7 Land 1 if connection is from/to the same host/port; 0 otherwise Discrete 8 Wrong-fragment Number of wrong fragments Continuous 9 Urgent Number of urgent packets Discrete 10 Hot Number of hot indicators Discrete 11 Num-failed-logins Number of failed login attempts Discrete 12 Logged-in 1 if successfully logged in; 0 otherwise Discrete 13 Num-compromised Number of compromised condition Discrete 14 Root-shell 1 if root shell is obtained; 0 otherwise Discrete 15 Su-attempted 1 if su root command attempted; 0 otherwise Discrete 16 Num-root Number of root accesses Discrete 17 Num-file-creations Number of file creation operations Discrete 18 Num-shells Number of shell prompts Discrete 19 Num-access-files Number of operations on access control files Discrete 20 Num-outbound-cmds Number of outbound commands in an ftp session Discrete 21 Is-host-login 1 if the login belongs to the hot list; 0 otherwise Discrete 22 Is-guest-login 1 if the login is a guest login; 0 otherwise Discrete 23 Count Number of connections to the same host as the current Discrete connection in the past two seconds 24 Srv-count Number of connections to the same service as the current Discrete connection in the past two seconds 25 Serror-rate Percent of connections that have SYN errors Discrete 26 Srv-serror-rate Percent of connections that have SYN errors Discrete 27 Rerror-rate Percent of connections that have REJ errors Discrete 28 Srv-rerror-rate Percent of connections that have REJ errors Discrete 29 Same-srv-rate Percent of connections to the same services Discrete 30 Diff-srv-rate Percent of connections to different services Discrete 31 Srv-diff-host-rate Percent of connections to different hosts Discrete 32 Dst-host-count Count for destination host Discrete 33 Dst-host-srv-count Srv-count for destination host Discrete 34 Dst-host-same-srv-rate Same-srv-rate for destination host Discrete 35 Dst-host-diff-srv-rate Diff-srv-rate for destination host Discrete 36 Dst-host-same-src-port-rate Same-src-port-rate for destination host Discrete 37 Dst-host-srv-diff-host-rate Diff-host-rate for destination host Discrete 38 Dst-host-serror-rate Serror-rate for destination host Discrete 39 Dst-host-srv-serror-rate Srv-serror-rate for destination host Discrete 40 Dst-host-rerror-rate Rerror-rate for destination host Discrete 41 Dst-host-srv-rerror-rate Srv-serror-rate for destination host Discrete International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 429 Table 4: Selected features out of the 41 features Category # Feature Selected features Normal 5 Urgent, Num-failed-logins, Count, Rerror-rate, Dst-host-srv-diff-host-rate (9, 11, 23, 27, 37) DoS 4 Duration, Flag, Root-shell, Dst-host-srv-diff-host-rate (1, 4, 14, 37) U2R 4 Service, Dst-bytes, Count, Serror-rate (3, 6, 23, 25) R2L 3 Count, Srv-count, Diff-srv-rate (23, 24, 30) Probe 8 Protocol-type, Duration, Hot, Logged-in, Num-compromised, Num-access-files, Diff-srv-rate, Dst-host-diff-srv-rate (2, 4, 10, 12, 13, 19, 30,35) Table 5: Proposed IDS performance on the KDD Cup 99 data set Category All features Proposed IDS Normal DoS U2R R2L Probe Normal DoS U2R R2L Probe #Correctly detected 45954 183928 29 537 2800 59024 229347 2465 13667 3110 # Miss detected 14639 45925 2607 13244 1366 1569 506 171 114 1056 Precision 77.74 87.86 46.77 82.23 6.68 69.60 81.66 6.53 13.07 86.41 Recall (TPR) 75.84 80.02 1.10 3.9 67.21 97.41 99.78 93.51 99.17 74.65 F-measure 76.78 83.76 2.15 7.45 12.15 81.19 89.82 12.21 23.10 80.10 tures for the all attack classes in each step of the proposed method is depicted in Figure 2. The maximum difference between true positive rate and false positive rate is in the 5th step. Hence the first five features are selected for modeling the system. A comparison between the test results for the proposed method versus other machine learning methods tested on the KDD Cup 99 test set are presented in Table 7. It can be stated that all the machine learning algorithms tested on this data set offered an acceptable level of de- tection performance for Normal, DoS and Probe attacks but they did not have good performance on the U2R and R2L types. The proposed method shows better TPR for DoS, U2R and R2L attacks and offers an acceptable level of detection rate for Normal and Probe attacks. There are 18729 records of various new attacks in the KDD Cup 99 test set, which have never appeared in the training set. These new attack records make an IDS trained by a training set hard to achieve good performance for test set. Experiments show that the proposed method shows an acceptable detection rate for detecting new attacks. Considering the reported results, ACO is faster in lo- cating the optimal solution. In general, it can find the optimal solution within tens of iterations. If exhaustive search is used to find the optimal feature subset in the KDD Cup 99 data set, there will be tens of billions of candidate subsets, which, makes the search nearly impos- sible. Using ACO, the optimal solution is found after 100th iterations. ACO has powerful exploration ability; it is a gradual searching process that approaches optimal solutions. The execution time for the ACO is mainly af- fected by the dimensionality of the problem (number of the features), and the size of the data. ACO can search in Figure 2: (a) TPR is depicted in each step and (b) FPR is shown in each step International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 430 Table 6: Proposed IDS performance on the NSL-KDD data set Category All features Proposed IDS Normal DoS U2R R2L Probe Normal DoS U2R R2L Probe #Correctly detected 9455 5131 28 454 1434 9483 5613 392 597 1667 # Miss detected 256 2327 505 1967 987 228 1845 141 1824 754 Precision 65.46 87.34 68.29 90.80 85.10 61.31 79.22 8.17 93.14 85.79 Recall (TPR) 97.36 68.80 5.25 18.75 59.23 97.65 75.26 73.55 24.66 68.86 F-measure 78.29 76.97 9.75 31.08 69.85 75.33 77.19 14.71 39.00 76.40 Table 7: Detection rate per record of KDD Cup 99 for the different algorithms performances on the test data set with corrected labels of KDD Cup 99 data set Model Normal(%) DoS(%) U2R(%) R2L(%) Probe(%) Accuracy FPR Proposed method 97.41 99.78 93.51 99.17 74.65 98.9 2.59 PLSSVM [3] 95.69 78.76 30.7 84.85 86.46 Not reported 4.3 Clustering feature [11] 99.3 99.5 19.7 28.8 97.5 95.7 0.7 ESC-IDS [32] 98.2 99.5 14.1 31.5 84.1 95.3 1.9 KDDwinner [24] 99.5 97.1 13.2 8.4 83.3 91.8 0.5 KDD99 runner-up [16] 99.4 97.5 11.8 7.3 84.5 91.5 0.6 the feature space until the optimal solution is found. ACO comprises a very simple concept, and the ideas can be im- plemented in a few lines of computer code. It requires only primitive mathematical operators, and is computationally inexpensive in terms of both memory requirements and speed. Each ant has a separate memory. In ACO, each ant that passes the optimum solutions are tagged so that other ants can use them and knowledge of good solutions is retained by all ants. 5 Conclusions and Future Works This paper addresses the problem of dimensionality re- duction using ACO in intrusion detection problem area. ACO has the ability to converge quickly. It has a strong search capability in the problem space and can efficiently find minimal feature subset. Experimental results demon- strate a competitive performance. More experimentation and further investigation into this technique is required. The pheromone trail decay coefficient and pheromone amount have an important impact on the performance of ACO. Selection of the parameters is proved to be problem- dependent. The deposited pheromone expresses the qual- ity of the corresponding solution. Evaporation becomes more important for more complex problems. If ρ = 0, i.e. no evaporation, the algorithm does not converge. If pheromone evaporates too much (a large ρ is used), the algorithm often converged to sub-optimal solutions. In many practical problems, it is difficult to select the best ρ without trial-and-error. α and β are also key factors in ACO for FS. For large data sets, to speed up the calcu- lation of FS process, a parallel algorithm can be imple- mented. The proposed method uses ACO algorithm and a sim- ple classifier (nearest neighbor classifier) to select impor- tant features and a trained classifier to identify any kind of new attacks. Tests and comparisons are performed on KDD Cup 99 and NSL-KDD data sets, the test sets con- tains 17 kinds of different attacks. The proposed method reduced the number of features by approximately 88% and the detection error reduced by around 24% using KDD Cup 99 test data set. The proposed method will signif- icantly reduce both the memory size and the CPU time required for intrusion detection by reducing number of the features used for the detection. This shows that the proposed method is very reliable for intrusion detection. Results indicate that the proposed ACO-based detection method outperforms other methods since it can provide better and more robust representation of the data. This is due to the fact that it can accurately detect a broader range of attacks using smaller number of features. As for the future work, intention is to apply the pro- posed intrusion detection method using complicated clas- sifiers to improve its performance and to combine the pro- posed method with other population-based algorithms. Analyzing packet payload is recently attracting lots of attention and many researchers report works carried-out in this area. It is notable that feature selection for the payload-based intrusion detection is not mature yet. In- tension will be to extract and selection appropriate fea- tures from the packet payload to improve the detection rate. International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 431 Acknowledgments This study was supported by the Iran University of Sci- ence and Technology. The authors gratefully acknowledge the anonymous reviewers for their valuable comments. References [1] M. H. Aghdam, N. Ghasem-Aghaee, and M. E. Basiri, “Application of ant colony optimization for fea- ture selection in text categorization,” in Proceedings of the IEEE Congress on Evolutionary Computation (CEC’08), pp. 2872–2878, 2008. [2] M. H. Aghdam, J. Tanha, A. R. Naghsh-Nilchi, and M. E. Basiri, “Combination of Ant Colony Optimiza- tion and Bayesian Classification for Feature Selection in a Bioinformatics Dataset,” Journal of Computer Science and System Biology, vol. 2, pp. 186–199, 2009. [3] F. Amiri, M. M. R. Yousefi, C. Lucas, A. Shakery, and N. Yazdani, “Mutual information-based feature selec- tion for intrusion detection systems,” International Journal of Network and Computer Applications, vol. 34, pp. 1184–1199, 2011. [4] J. Bins, Feature Selection from Huge Feature Sets in the Context of Computer Vision, Ph.D. dissertation, Colorado State University, 2000. [5] C. Blum, and M. Dorigo, “The hyper-cube framework for ant colony optimization,” IEEE Transaction on Systems, Man, and Cybernetics, Part B, vol. 34, no. 2, pp. 1161–1172, 2004. [6] E. Bonabeau, M. Dorigo, and G. Theraulaz, Swarm Intelligence: From Natural to Artificial Systems, Ox- ford University Press, New York, 1999. [7] S. Chebrolu, A. Abraham, and J. P. Thomas, “Fea- ture deduction and ensemble design of intrusion de- tection systems,” International Journal of Computers and Security, vol. 24, pp. 295–307, 2005. [8] P. S. Chung, C. W. Liu, and M. S. Hwang, “A Study of Attribute-based Proxy Re-encryption Scheme in Cloud Environments,” International Journal of Net- work Security, vol. 16, no. 1, pp. 1–13, 2014. [9] M. Dorigo, and C. Blum, “Ant colony optimization theory: A survey,” Theoretical Computer Science, pp. 243–278, 2005. [10] A. P. Engelbrecht, Fundamentals of Computational Swarm Intelligence, John Wiley & Sons, London, 2005. [11] S. J. Horng, M. Y. Su, Y. H. Chen, T. W. Kao, R. J. Chen, J. L. Lai, C. D. Perkasa, “A novel intru- sion detection system based on hierarchical clustering and support vector machines,” International Journal of Expert Systems with Applications, vol. 38, no. 1, pp. 306–313, 2011. [12] R. Jensen, Combining Rough and Fuzzy Sets for Fea- ture Selection, Ph.D. dissertation, School of Informa- tion, Edinburgh University, 2005. [13] K. J. Lee, J., Joo, J., Yang, and V. Honavar, “Exper- imental comparison of feature subset selection using GA and ACO algorithm,” in Advanced Data Mining and Applications, LNCS 4093, pp. 465–472, Springer, 2006. [14] W. Lee, S. J. Stolfo, K. W. Mok, “Adaptive intru- sion detection: a data mining approach,” Interna- tional Journal Artificial Inteligence Review, vol. 14, no. 6, pp. 533–567, 2000. [15] G. Leguizamon, and Z. Michalewicz, “A New Version of Ant System for Subset Problems,” In Proceedings of IEEE Congress on Evolutionary Computation, 1999. [16] I Levin, “KDD-99 classifier learning contest LLSofts results overview,” International Journal of SIGKDD Explorations, vol. 1, no. 2, pp. 67–75, 2000. [17] H. Liu, and L. Yu, “Toward integrating feature se- lection algorithms for classification and clustering,” IEEE Transactions on Knowledge and Data Engineer- ing, pp. 491–502, 2005. [18] V. Maniezzo, A. Colorni, “The Ant System Applied to the Quadratic Assignment Problem,” IEEE Trans- action on Knowledge and Data Engineering, vol. 11, no. 5, pp. 769–778, 1999. [19] MIT Lincoln Labs, 1998 DARPA Intru- sion Detection Evaluation, Feb. 2012. (http: //www.ll.mit.edu/mission/communications/ist/ corpora/ideval/index.html) [20] D. Mladeni, “Feature Selection for Dimensionality Reduction,” in Subspace, Latent Structure and Fea- ture Selection, Statistical and Optimization, Perspec- tives Workshop (SLSFS’06), LNCS 3940, pp. 84–102, Springer, 2006. [21] R. Montemanni, L. M. Gambardella, A. E. Rizzoli, and A. V. Donati, “A new algorithm for a Dynamic Vehicle Routing Problem based on Ant Colony Sys- tem,” in Second International Workshop on Freight Transportation and Logistics, pp. 27–30, 2003. [22] S. Mukkamala, and A. H. Sung, “Feature ranking and selection for intrusion detection systems using support vector machines,” in Proceedings of the Inter- national Conference on Information and Knowledge Engineering, pp. 503–509, 2002. [23] S. Y. Ohn, N. H. Nguyen, D. S. Kim, and J. S. Park, “Determining Optimal Decision Model for Sup- port Vector Machine by Genetic Algorithm,” in Com- putational and Information Science, LNCS 3314, pp. 895–902, Springer, 2005. [24] B. Pfahringer, “Winning the KDD 99 classifica- tion cup: bagged boosting,” International Journal of SIGKDD Explorations, vol. 1, no. 2, pp. 65–66, 2000. [25] M. Ramadas, S. Ostermann, and B. Tjaden, “De- tecting anomalous network traffic with self-organizing maps,” in Proceedings of Sixth International Sympo- sium on Recent Advances in Intrusion Detection, pp. 36–54, 2003. [26] C. J. Rijsbergen, Information Retrieval (2nd ed.), London, UK: Butterworth, 1979. [27] S. J. Stolfo, W. Fan, W. Lee, A. Prodromidis, P. K. Chan, “Cost-based modeling for fraud and intru- sion detection: Results from the jam project,” in In International Journal of Network Security, Vol.18, No.3, PP.420-432, May 2016 432 Proceedings DARPA Information Survivability Con- ference and Exposition (DISCEX’00), pp. 130–144, 2000. [28] A. H. Sung, and S. Mukkamala, “Identifying impor- tant features for intrusion detection using support vec- tor machines and neural network,” in Proceedings of International Symposium on Applications and the In- ternet, pp. 209–216, 2003. [29] A. H. Sung, and S. Mukkamala, “The Feature Se- lection and Intrusion Detection Problems,” in Pro- ceedings of Advances in Computer Science - ASIAN: Higher-Level Decision Making. 9th Asian Computing Science Conference, LNCS 3321, pp. 468–482, 2004. [30] W. S. Sung, H. L. Chi, “Using Attack-Specific Fea- ture Subsets for Network Intrusion Detection,” in Ad- vances in Artificial Intelligence, LNCS 4304, pp. 305– 311, Springer, 2006. [31] M. Tavallaee, E. Bagheri, W. Lu, and A. A. Ghor- bani, “A Detailed Analysis of the KDD CUP 99 Data set,” in Proceeding of the IEEE Symposium on Com- putational Intelligence in Security and Defense Appli- cations, pp. 57–62, 2007. [32] A. N. Toosi, and M. Kahani ”A new approach to in- trusion detection based on an evolutionary soft com- puting model using neuro-fuzzy classifiers,” Interna- tional Journal of Computer Communications, vol. 30, pp. 2201–2212, 2007. [33] C. Tsang, S. Kwong, and H. Wang, “Genetic-fuzzy rule mining approach and evaluation of feature selec- tion techniques for anomaly intrusion detection,” In- ternational Journal of Pattern Recognition, vol. 40, pp. 2373–2391, 2007. [34] UCI, KDD Cup 1999 Data, The UCI KDD Archive Information and Computer Science University of Cali- fornia, Irvine, Oct. 2014. (http://kdd.ics.uci.edu/ databases/kddcup99/kddcup99.html) [35] UNB, NSL-KDD Data Set for Network-based Intru- sion Detection Systems, Mar. 2014. (http://nsl.cs. unb.ca/NSL-KDD/) [36] S. M. Vieira, J. M. C. Sousa, and T. A. Runkler, “Fuzzy classification in ant feature selection,” in IEEE International Conference on Fuzzy Systems, pp.1763– 1769, 2008. [37] P.D. Williams, K. P. Anchor, J. L. Bebo, G. H. Gun- sch, and G. D. Lamont, “CDIS: towards a computer immune system for detecting network Intrusions,” in Proceedings of Fourth International Symposium on Recent Advances in Intrusion Detection, pp. 117–133, 2001. [38] S. X. Wu, and W. Banzhaf, “The use of computa- tional intelligence in intrusion detection systems: A review,” International Journal of Applied Soft Com- puting, vol. 10, pp. 1–35, 2010. [39] L.H. Zhang, G. H. Zhang, L. Yu, J. Zhang, and Y. C. Bai, “Intrusion Detection Using Rough Set Classifi- cation,” International Journal of Zheijiang University Science, vol. 5, no. 9, pp. 1076–1086, 2004. Mehdi Hosseinzadeh Aghdam is a Ph.D. candidate in the computer engineering department at the Iran University of Science and Technology, he is also a lecturer at Payame Noor University (PNU). He graduated with a master’s in computer engineering Artificial Intelligence from University of Isfahan (UI) in 2008. At UI, he worked on swarm intelligence-based method for feature selection. His main research interests are: Data Mining (Feature Selection, Recommendation Systems), Computational Intelligence, Pattern Recognition and Social Network Analysis. Peyman Kabiri is assistant professor of computer engineering at the Iran University of Science and Tech- nology (IUST). He received a B.Eng. degree in computer hardware engineering from IUST in 1992 and the M.Sc. and Ph.D. degrees in computer science at the Nottingham Trent University, UK, in 1996 and 2000 respectively. He is the director of the Intelligent Automation Laboratory at the Department of Computer Engineering-IUST. Dr. Kabiri has authored a number of publications including journal articles, book chapters, and conference papers. He is editor of a book in intrusion detection work area. 
work_ps7ltaslkvb7hj7qgt6uhrcbni	----	PII: 0957-4174(95)00013-Y Pergamon Expert Systems With Applications, Vol. 9. No. 3. pp. 407-421, 1995 Copyright 0 1995 Elsevier Science Ltd Printed in the USA. All tights reserved 0957-4 I74/95 $9.50 + .Oa 0957-4174(95)00013-5 Building a Fuzzy Expert System for Electric Load Forecasting Using a Hybrid Neural Network I? K. DASH AND A. C. LIEW National University of Singapore, 10, Kent Ridge Crescent, Singapore S. RAHMAN Department of Electrical Engineering, Virginia Polytechnic Institute and State University, Blacksburg, VA G. RAMAKRISHNA Regional Engineering College, Rourkela-769 008, India Abstract-This paper presents the development of a hybrid neural network to model a fuzzy expert system for time series forecasting of electricc load. The hybrid neural network is trained to develop fuzzy logic rules andjind optimal inputloutput membership values of load and weather parameters. A hybrid learning algorithm consisting of unsupervised and supervised learning phases is used for training the fuzzifred neural network. In the supervised learning phase, both back-propagation and linear Kalman jilter algorithms are used for the adjustment of weights and membership functions. Extensive tests have been performed on a 2-year utility data for the generation of peak and average loadprojiles in 24 h. 48 h. and 168 h ahead time frame during summer and winter seasons. From the simulation results, it is observed that the fuzzy expert system using the Kalman $lter-based algorithm gives faster convergence and more accurate prediction of a load time series. 1. INTRODUCTION IT IS WELL KNOWN that fuzzy logic provides an inference strategy that enables approximate human reasoning capabilities to knowledge-based systems. Also, it pro- vides a mathematical morphology to emulate certain perceptual and linguistic attributes associated with human cognition. Although fuzzy theory provides an inference mechanism under cognitive uncertainty, com- putational neural networks offer exciting advantages such as learning, adaptation, fault tolerance, parallelism, and generalization. The computational neural networks, comprising processing elements called neurons, are capable of coping with computational complexity, non- linearity, and uncertainty. To enable a system to deal with cognitive uncertain- ties in a manner more like humans, one may incorporate Requests for reprints should be sent to P K. Dash, Centre for Intelligent Systems, Electrical Engineering Department of Regional Engineering College, Rourkela 769008, India. the concepts of fuzzy logic into neural network. Although fuzzy logic is a natural mechanism for modelling cognitive uncertainty, it may involve an increase in the amount of computation required. This can be readily offset by using fuzzy neural network approaches having the potential for parallel computa- tions with high flexibility. The application of an artificial neural network (ANN) and fuzzy logic-based decision support system to time- series forecasting has gained much attention recently. ANN-based load forecasts give large errors when the weather profile changes very fast. Also extremely slow training or even training failure occurs in many cases due to difficulties in selecting proper structures of the neural network paradigm being used, and due to the errors in associated parameters such as learning rates, activation functions, etc., which are fundamental to any back- propagation neural network. On the other hand, the development of a fuzzy decision system (fuzzy expert system) for load forecasting requires detail analysis of data, and the fuzzy rule base has to be developed 407 408 P. K. Dash et al. heuristically for each season. The rules fixed in this way may not always yield the best forecast. Thus, a hybrid neural network model is used that combines the idea of fuzzy logic-based decision system and neural network structure and learning abilities into an integrated frame- work. Such a hybrid model provides human understandable meaning to the normal feedforward neural network in which the internal units are always opaque to the users. This structure also avoids the rule matching time of the inference engine in the traditional fuzzy logic system and results in enhanced learning speed and prediction accuracy. The present work is aimed at achieving the said objective of a robust load forecast with much improved accuracy using a fuzzy expert system modelled by the hybrid ANN architecture. The two types of fuzzy expert system models, abbreviated FES, and FES, are based on the argument that any fuzzy expert system employing one block of rules may be approximated by a neural network (feedforward, multilayered). The input vector to FES, and FES, consists of differences in weather parameters between the reference and the forecasted day or hour. The output of the FES, and FES, gives the load correction that, when added to the past load, gives the forecasted load. The learning algorithm for FES, and FES, combines unsupervised and supervised learning procedures to build the rule nodes and train the membership functions. The supervised learning proce- dure for FES, uses a gradient back-propagation algorithm for finding the optimum weights and member- ship functions. Although for FES,, the supervised learning procedure uses a linear Kalman filter-based learning algorithm suggested by Scaler0 and Tepede- lenlioglu (1992), which is similar to the least square adaptive learning techniques. The least square adaptive filtering techniques are known to have rapid convergence properties over the back-propagation algorithm. A few examples of peak load forecasting and daily average load forecasting of a typical utility with 24 h and 168 h lead time are shown in this paper. The approach presented in this paper is also applicable to load forecasts with longer lead times. 2. FUZZY EXPERT SYSTEM 2.1. Fuzzy Statements This section portrays inexact information presented by fuzzy statements, and explains both fuzzy conditional statements and inference mechanisms. Typically, an engineering model requires the use of exact or mathe- matical statements. These statements correspond to precise information such as x=3.0,2<8<0, ory=3t+24. The value of x = 3.0 has a grade membership comparable to 100% (= l), for all other values (2.8,2.9, 3.1,3.2), the grades of membership in the solution is zero. In the case of real world values, however, this grade of membership is not true due to the imprecision of tools, the influence of the observer, etc. Additionally, human reasoning is often imprecise, that is, inexact statements such as, “the value of x is not big” or “the temperature is about 4”.” Therefore, a theory to correctly express the grade of membership is desirable. The Fuzzy Set Theory allows manipulation of both exact and inexact (fuzzy) statements. This is very important for load forecasting because there are so many fuzzy factors that are difficult to characterize by a number. An instance of this could be weather conditions such as temperature, humidity, cloud cover, etc. For example, “the temperature this morning will be about ll’.” Temperature is an important factor in load forecasting, but is not easy to characterise by an exact numerical quantity. In addition, the linguistic hedges (such as small, medium, large) can be modified by other linguistic hedges (such as not, very, very very) (Zadeh, 1965) For example, “value of x is not big,” where big is a linguistic hedge and not a modifier. 2.2. Fuzzy Rule Base A fuzzy expert system consists of a collection of fuzzy IF-THEN rules. For example, the rule base R is written as R = [R’, R=, . . ., R”] (1) with R’:IF (x, is F{ and . . and xp is FL) THEN (y, is G: and . . . y4 is Gb) (2) where x = (x, , . . ., x,JT and y= (yl, . . ., y,)r are the input and output vectors to the fuzzy system, respectively. Fi and GI are labels of fuzzy sets Vi and V,, respectively, and l<p~n;l~q~m,andl=1,2 ,..., M. 2.3. Fuzzy Inference Engine A fuzzy inference engine uses the rules in the fuzzy rule base to determine a mapping from fuzzy sets in U to fuzzy sets in V based on fuzzy logic operations. In the fuzzy inference engine, the IF part of Rj defines a Cartesian product of F{, . . ., Fi and the Rj itself is viewed as a fuzzy implication F: x . . . x Fi*G:. If A is an arbitrary fuzzy set in U, then each R’ of eqn. (2) determines a fuzzy set AR’ in V based on the sup-star composition WW(YA.RO = supr; t JCLA(X . Y,) F,* . . . *F;*G; = supx E &,4(~*rLL,;(x,)* . . . *~Fpp)*PG~(Y,>l (3) where y, E V,. Hybrid Neural Nehvork for Fuzzy Expert System 409 The final fuzzy set AoR, (R,= [R,!, . . . , RF]) in V, determined by the fuzzy inference engine is obtained by combining eqn. (2) for 1= 1,2, . . ., M using the triangu- lar co-norm ~~dt,(Y,) = PA-R,! (Y,) + . . . + PA++ (YJ>. (4) 2.3.1. Fuzzijication and Defuzzification. In many cases it is convenient to express the membership function of a fuzzy subset in terms of a standard nonlinear function. The following Gaussian membership function suggested by Lin and Lee (1991) is used for the fuzzification of input and output linguistic parameters of the fuzzy expert system: PA,(x) = exp -Iv i I (5) where a and h are the center and width of the Gaussian membership function, respectively. The defuzzifier is needed to give a crisp output for any practical application like forecasting. The output of the fuzzy inference engine is a fuzzy set AQR in V, therefore, the defuzzifier maps AoR into a crisp point y E V. The most commonly used centroid defuzzifier is defined as (6) 3. FUZZY EXPERT SYSTEM MODEL USING NEURAL NETWORK APPROACH An alternative to the neural network-based load forecast is the expert system approach. A fuzzy expert system approach for load forecast consists of rules similar to the rules R’, . . ., RM given in Section 2.2. One of the difficulties with the fuzzy expert system is the rule matching and composition time, apart from the time- consuming process of adapting the rules. The neural network eliminates the rule matching process and stores the knowledge in the link weights. The decision signals can be pumped out immediately after the input data are fed in. Figure 3 shows the proposed hybrid neural network to model the fuzzy expert system using the ANN architecture. The fuzzy expert system clusters the differential temperatures and humidities of the ith and (i + n)th days into fuzzy term sets. Here n is the lead time for the forecast (i.e., n = 24 for 24 h ahead forecast, n = 48 for 48 h ahead forecast, and so on). The output of the expert system is the final load correction (ELc). Hence, the forecasted load on (i + n)th day (PAi + n)) is given by: P,(i + n) = P(i) +6,,(i). (7) The fuzzy expert system modelled as a hybrid neural network has a total of five layers. Nodes at layer 1 are the input linguistic nodes. Layer 5 is the output layer and consists of two nodes [one is for the actual load correction (&) and the other is the desired load correction (e&l. Nodes at layer 2 and 4 are term nodes that act as membership functions to represent the term sets of the respective linguistic variable. Each node at layer 3 represents the preconditions of the rule nodes, and layer 4 links define the consequence of the rules. The functions of each layer is described as follows: a> b) cl Layer 1: The nodes in this layer just transmit the input feature x,, i = 1, 2 to the next layer. Layer 2: Each input feature x,, i = 1, 2 is expressed in terms of membership values &a,,, l~,~), where i corresponds to the input feature and j corresponds to the number of term sets for the linguistic variable x,. The membership function &, uses the Gaussian membership function (5). Layer 3: The links in this layer are used to perform precondition matching of fuzzy logic rules. Hence, the rule nodes perform the product operation (or AND operation). pRuRp = n pJ I, (8) where R,= 1, 2,. . ., n. R, corresponds to the rule node and n is the maximum number of rule nodes. However, if the fuzzy AND operation is used d) Layer 4: The nodes in this layer have two operations (i.e., forward and backward transmission). In forward transmission mode, the nodes perform the fuzzy OR operation to integrate the fired rules, which have the same consequence: p4=C14; (10) ,=I where p corresponds to the links terminating at the node. In the backward transmission mode, the links function exactly the same as the layer 2 nodes. e) Layer 5: There are two nodes in this layer (i.e., for obtaining the actual and desired output load correc- tion, respectively). The desired output load correction (e,J is fed into the hybrid neural network during learning whereas the actual load correction (a,,) is obtained by using the centroid defuzzification method (6). 410 4. HYBRID LEARNING ALGORITHM FOR THE FUZZY EXPERT SYSTEM (FES) The hybrid learning scheme consists of two phases. In phase I, the initial membership functions of the input and output linguistic variables are fixed and an initial form of the network is constructed. Then, during the learning process, some nodes and links of this initial network are deleted or combined to form the final structure of the network. In phase II learning, the input and output membership functions are optimally adjusted to obtain the desired outputs. Phase I represents the unsupervised learning phase. Phase II represents the supervised learning phase. 4.1. Phase-I: Unsupervised Learning Phase Given the training input data x,(r), i = 1,2, the desired output load correction (e,(r)) and the fuzzy partitions I&,1. We want to locate the membership functions (i.e., aij and b,) and find the fuzzy logic rules. The Kohonen’s feature maps algorithm (Lin & Lee, 1991) is used to find the values for ai, and b,: ai, clo~est(~ + l) = ui, doscAr) + tit) [ (‘)I xi(r)-ui, closest u,(r + 1) = uij(t) for uii f a,, C,Oscst (11) (12) (13) where rl(f) is the monotonically decreasing learning rate and t= IT( (i.e., the number of term sets for the linguistic variable xi). This adaptive formulation runs independently for each input linguistic variable xi: The width, b, is determined heuristically at this stage (Lin & Lee, 1991) as follows: b,j = Ia, jwui, closest1 r (14) where r is an overlap parameter. After the parameters of the membership functions have been found, the weights in layer 4 are obtained by using the competitive learning algorithm (Rumelhart & Zipses, 1985) as follows: w, = Ll;<u; - wi,, (1% where LJ; serves as the win-loss index of the rule node at layer 3 and Ll;’ serves as the win-loss index of the jth term node at layer 4, respectively. After the competitive learning through the whole training data set, the link weights at layer 4 represent the P. K. Dash et al. strength of the existence of the corresponding rule consequence. If a link weight between rule node and the term node of the output linguistic node is very small, then all the corresponding links are deleted, meaning that this rule node has little or no relation to the output. After the consequences of rule nodes are determined, the rule combination is performed to reduce the number of rules in the following manner. The criteria for the choice of rule nodes are: 1. 2. 3. they have the same consequences, some preconditions are common to all the rule nodes in this set, the union of other preconditions of these rule nodes composes the whole term set of some input linguistic variables. The rule nodes that satisfy these criteria are replaced by a new rule node with common preconditions. 4.2. Phase II: Supervised Learning Phase Once the fuzzy logic rules have been found, the phase II learning is used to find the optimum weights and the input and output membership functions by using the gradient descent backpropagation algorithm (Scaler0 & Tepedelenlioglu, 19921. a) The weight update equation,for layer 4 is: Wij(f) = W;,(t)+ 7) g . [ I ‘I The error function E is given by Because c (@i&d c,, = c bijd using centroid defuzzification method (6) P /+p+L; w, i=l where W,= 1 fort= 1. Now, dE aE ae^,* -=-- a w, a;,, a CL,’ a w;,’ (16) (17) (18) (19) (20) Hybrid Neural Network for Fuzzy Expert System 411 Therefore, $ = [eLcW - 4&)1 V Training the output membership functions at layer 5: -dE a,,0 + 1) = U;,(f) + 771 [ 1 __ a 4, (22) where (a,,, b,,) correspond to the output term set. The error signal at layer 3 is found by performing the summation over the consequences of a rule node (i.e., layer 4). Therefore. @=Xcsp. (2% The adaptive rule for a;, (layer 2) is derived as: -l!iE u&t + 1) = ujjt) + r), [ 1 __ a u,, where qj is the learning rate. Now, where n1 is the learning rate. Now, aE aE ae^, G= a;,, au,,’ Therefore, using eqns. (17) and (18) we get (23) Also, S2= CS3, (30) (31) (32) a,,(t + 1) = a,(r) + ~l(eK(~)-4_c(t)) . (24) in a similar way as eqn. (29). Therefore, Similarly, (33) b,,(t + 1) = h,(t) = 712 (25) Similarly, the adaptive rule for b,, (layer 2) is derived as where v2 is the learning rate. Now, aE aE ac, a==ae^,. (26) Therefore, using eqns. (17) and (18) we get h,,(t + 1) = h,(t) + ~2h_C(t) -%&)) c) Tuning the input membership functions at layer 2: The error signal # at layer 4 is given by b,,(t+ l)=bi,(~)--~~#e( -(~~~‘)‘)r~$~}. (34) The convergence speed of the phase 11 supervisory learning scheme for fuzzy expert systems is found to be superior to the supervisory learning scheme for ordinary fuzzy neural network approach (Dash et al., 1993) because the phase I unsupervised learning process had done much of the learning in advance. The convergence speed of the supervised learning process can be improved further by solving the weight update equations at layer 3 and the input and output membership function at layer 1 and layer 2 by linear Kalman filter equations (Singhal & Wu, 1989). 5. FUZZY EXPERT SYSTEM USING LINEAR KALMAN FILTER (FE&) Referring to Figure 1, the tuning of Gaussian member- ship function at layers 2 and 4 (a,,, b,) is similar to the weight update equations at layer 3. In (Scaler0 & Tepedelenlioglu, 1992) it has been shown that the 412 P. K. Dash et al. III-I- Layer- 1 Layer-2 Layer-3 Layer-4 Layer-5 (input linguistic (input (rule nodes) (output term (output linguistic nodes) terms nodes) nodes) nodes) AtJ - Differential temperature t - Iteration No. AH- Differential humidity Wij - weight OLC - Actual load correction eLC - Desired load correction FIGURE 1. Hybrid neural network modelled as a fuzzy expert system. presence of nonlinear function (in this study the Gaussian membership function) between the weights and the error makes the back-propagation algorithm different from standard least square adaptive filtering techniques, which are known to have rapid convergence (Haykin, 1986). The FE& model uses the linear Kalman filter equations for updating the weights and the membership function. Unlike the back-propagation technique, this algorithm assumes that the estimated weight matrix is nonstationary and hence will allow the tracking of a time varying data like that of load forecasting. This algorithm defines locally at each node a gradient based on present and past data, and updates the weights of each node using the linear Kalman filter equations to bring this gradient identically to zero whenever an update is made. Performing the update and defining the gradient in this manner ensures that maximum use is made of available information. The gradient for the linear combiner in each is defined as G=R W-C. (35) The weight vector that makes G = R W-C zero is the solution to the normal equations. Here R is the auto correlation matrix for each layer and is calculated as NP R = c fNPmnpxnp XI;, flp=l and C is the cross-correlation matrix and is given by NP C = c f”‘-” dnp XI;, flp=l where NP denotes the total number of patterns. dnp and xnp are the summation output and the output of the nonlinearity (Gaussian membership function), respec- tively, for the layer 2 and layer 5 nodes, respectively. As the layer 4 nodes contain no nonlinearity term, dnp = xnp. The weights given by eqn. (35) are updated by using a Kalman filter at each layer with a variable forgetting factor (J as shown in eqns. (36) and (37). A variable forgetting factor is used to take care of the new estimates giving less weightage to old estimates. The Kalman gain vector K, (t) at each layer j is given by K,(t) = RF'(f)xi(r) +2(r) R,-‘(o-w I I (38) where x,(t) corresponds to the previous layer, ct, is the learning rate, and t denotes the iteration number. The error, E, at each iteration is obtained from eqn. (17). The first term in eqn. (38) expresses the Kalman gain, K,, in Hybrid Neural Network for Fuzzy Expert System 413 TABLE 1 The Learned Fuzzy Logic Rules for 24 Hour Ahead Peak Load Forecasting Using FE!& in Winter _--_ Term Sets Preconditions Consequence Rule A &,,(i, i+ 1) AH,,,(i, i+ 1) &(I) -- 0 0 3 7 1 0 4 7 2 1 0 8 3 1 1 7 4 1 2 7 5 1 3 6 6 1 4 6 7 2 0 8 8 2 1 7 9 2 2 7 10 2 3 6 11 2 4 7 12 3 1 2 13 3 2 4 14 3 3 6 15 4 2 5 16 4 3 6 17 4 4 1 18 5 0 3 19 5 1 2 20 5 2 1 21 5 3 1 22 5 4 0 23 6 0 1 24 6 1 1 25 6 2 0 terms of the inverse autocorrelation matrix R, and the input to the previous layer x,. The second term is used to track the error along a gradient descent surface to attain the final convergence (similar to tuning of membership functions). This helps in a faster rate of convergence. During training the network is randomly initialized. The original training pairs become obsolete as the adjacent layers adapt their weights according to their perceived observations. Hence, the R,(t) matrix for a particular layer, containing the sum total of all prior observations, becomes outdated. Because the Kalman filter minimizes the prediction effort of all prior observa- tions, parameter estimates become biased towards initial conditions and the solution to this problem is the modification of the covariance matrix and the forgetting factor in the following manner: f,(r + 1) =fi- Y, w> R,-‘U+ 1) = [R,-‘W-K,(O$(r) R,- ‘(W/J; where (39) (40) and ‘y, is a constant that is inversely proportional to the prediction error, e(r), at that layer. As a result,& remains close to 1 when R, is already large and the weight estimates are sensitive to parameter variations. A lower bound fO and f, is used to prevent the forgetting factor from becoming excessively small, resulting in large estimate fluctuations in spite of small prediction errors. The strategy for employing this forgetting factor is to initially allow it to assume small values during learning such that the initial conditions are quickly forgotten. However, as the average prediction error decreases, it is important to cause 6 to remain near one to avoid the winding up of the weight estimates. This can be simply implemented by setting y, smaller as the mean squared error decreases. The unsupervised learning phase of FES, is same as for FES,. The supervised learning phase of FES, is modified using the linear Kalman filter equations instead of the back-propagation method used for FES,. There- fore, the update equations are modified as follows: a) The weight update equation for layer 4 is: P(f) = - r(O 1 +x;(t)R,-‘(t)x,(t) (41) (42) 414 P. K. Dash et al. where 8 El 8 W, is given by eqn. (21) and K,(t) is given by eqn. (38). b) The update equations for a,, and b,, at layer 5 are: b;,(t + 1) = b,,(t) + rl,K,(O i 1 2 V (4.4) where a El a d,, is given by eqn. (26) and K,(t) is given by eqn. (38). c) The update equations for a,, and b,j at layer 2 are: a,,@ + 1) = a,,(t) + rll K,(t) -aE [ 1 a (43) u&t + 1) = a,(t) + T3K,(t) [ 1 c (45) V where d El i_l a,_ is given by eqn. (23) and K,(t) is given by eqn. (38). Maximum temperature difference (after supervised learning) Maximum temperature difference (after unsupervised learning) 1.0 2 0.8 E : 0.6 2 c 2 0.4 5 E 0.2 0 -12 -8 -4 0 4 8 12 -12 -8 -4 0 4 8 12 (W 09 Maximum humidity difference (after unsupervised learning) Maximum humidity difference (after supervised learning) 0 -55 -15 25 (%I (%) Peak load difference Peak load difference (after unsupervised learning) (after supervised learning) 1 .o 1 4 0.8 E M .> 0.6 c :: 2 E 0.4 E 0.2 0 1 .o 4 0.8 E M 2 0.6 @ 2 E 0.4 E 0.2 0 -300 -180 -60 60 180 300 -300 -180 -60 60 180 300 (MW (MW FIGURE 2. Learned membership functions for 24 h ahead peak load forecasting, in January (winter) using FES,. Hybrid Neural Network for Fuzy Expert System 415 25 r n - ANN ----- FES, . . . . . . . . . . FESZ ‘!, r I i 1 i L .."\ '; I, :\ .'. . . I 400 800 1200 1600 2000 2400 Number of iterations FIGURE 3. Comparison of the mean absolute percentage errors versus the iteration number for 24 h ahead forecast in January (winter). where, SE/ 8 a,] is given by eqn. (31) and K,(t) is given by eqn. (38). Similarly, b&r + 1) = b,,(t)- ~4K,W~e (46) where df is given by eqn (32) and K,(t) is given by eqn. (38). The convergence properties of the supervisory leam- ing scheme for FES, are found to be superior to FES, using the Kalman filter equation in the weight and membership update equations, as shown in Sections 6.3 and 6.4. Also, the accuracy of the forecast increases for a similar reason. R,-‘(t+ 1) =$(t)R,-‘(t) +$(t)x,(t) and the forgetting factorA is modified as f;(t+l)=l-yt?(t+l) (46) (47) TABLE 2 Peak Load Forecasting in June (Summer) Using 24 Hour Ahead Forecast Peak Peak Peak Actual Forecast Forecast Forecast Load NJ) ,K) (FEW &SE,, (FEW Day (MW (MN (MN WW &SE,, 1 2 : 5 ; 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 1848 1850 1852 1687 1573 1752 1679 1674 1712 1778 1614 1520 1730 1663 1722 1710 1806 1730 1622 1752 1658 1672 1701 1736 1588 1482 1720 1699 1684 1693 30 MAPE Iterations required 1872.8 -1.34 1883.7 -1.82 1827.8 1.31 1627.6 3.52 1499.6 4.67 1752.4 -0.02 1661.4 1.05 1726.6 -3.14 1750.3 -2.24 1858.3 -4.52 1660.4 -2.87 1456.7 4.16 1705.0 1.45 1709.5 -2.80 1755.8 -1.96 1745.5 -2.08 1838.3 -1.79 1788.2 -3.37 1584.9 2.29 1775.7 -1.35 1648.6 0.57 1684.1 -0.72 1643.7 3.37 1669.5 3.83 1556.1 2.01 1413.4 4.63 1672.0 2.79 1750.4 -3.02 1666.6 1.03 1757.5 -3.81 2.45 970 1823.3 1844.7 1833.6 1669.7 1539.4 1747.6 1671.1 1652.6 1686.0 1752.7 1620.6 1474.5 1747.4 1674.5 1755.7 1681.4 1805.1 1757.7 1649.1 1772.9 1671.4 1685.3 1718.8 1702.9 1621.0 1463.9 1695.7 1708.2 1705.2 1676.9 1.34 0.29 0.99 1.03 2.14 0.25 0.47 1.28 1.52 1.42 -0.41 3.00 -1.01 -0.69 -1.96 1.67 0.05 -1.60 -1.67 -1.20 -0.81 -0.80 -1.05 1.91 -2.08 1.22 1.41 -0.53 -1.26 0.95 1.20 520 1825.7 1831.8 1868.0 1688.9 1544.1 1749.1 1668.5 1698.5 1715.3 1761.8 1602.7 1559.6 1719.5 1657.8 1741.7 1693.5 1803.3 1718.3 1644.7 1738.9 1662.6 1680.3 1727.7 1714.4 1619.2 1462.9 1705.4 1665.9 1667.5 1721.2 1.21 0.99 -0.86 -0.11 1.84 0.17 0.62 -1.46 -0.19 0.91 0.70 -2.60 0.61 0.31 -1.14 0.97 0.15 0.68 -1.40 0.75 -0.28 -0.50 -1.57 1.24 -1.96 1.29 0.85 1.95 0.98 -1.66 1.00 440 416 P. K. Dash et al. TABLE 3 Peak Load Forecasting in January (Winter) Using 48 Hour Ahead Forecast Peak Peak Peak Actual Forecast Forecast Forecast Load U'JN) Day (M'W (MW 1 2 3 4 5 6 7 8 9 IO 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 MAPE Iterations required 2690 2553.1 2628 2554.4 2703 2657.0 2592 2631.9 2530 2583.9 2574 2590.2 2389 2383.7 2513 2482.6 2500 2372.0 2450 2366.7 2551 2504.3 2763 2889.0 2603 2563.4 2914 2781.7 2761 2865.1 2514 2425.0 2543 2647.8 2435 2500.0 2496 2468.8 2551 2484.2 2813 2882.8 2537 2427.4 2381 2282.2 2459 2378.3 2505 2487.0 2429 2500.2 2438 2320.7 2748 2771.4 2388 2503.3 2175 2132.4 2539 2466.1 5.09 2663.6 0.98 2.80 2577.5 1.92 1.70 2672.7 1.12 -1.54 2560.6 1.21 -2.13 2509.5 0.81 -0.63 2591.2 -0.67 0.22 2400.0 -0.46 1.21 2493.4 0.78 5.12 2595.8 -3.83 3.40 2490.9 -1.67 1.83 2569.9 -0.74 -4.56 2849.2 -3.12 1.52 2623.3 -0.78 4.54 2875.8 1.31 -3.77 2780.9 -0.72 3.54 2567.3 -2.12 -4.12 2511.5 1.24 -2.67 2486.6 -2.12 1.09 2487.0 0.36 2.62 2505.6 1.78 -2.48 2773.3 1.41 4.32 2591.3 -2.14 4.15 2323.1 2.43 3.28 2432.7 1.07 0.72 2521.3 -0.65 -2.93 2451.3 -0.92 4.81 2379.0 2.42 -0.85 2711.7 1.32 -4.83 2347.2 1.71 1.96 2208.7 -1.55 2.87 2520.0 0.75 2.82 1.42 1560 880 2673.3 2663.2 2728.7 2561.4 2544.9 2555.5 2367.0 2497.7 2568.0 2478.9 2557.4 2811.6 2620.2 2895.9 2783.4 2552.0 2516.3 2414.8 2488.3 2522.9 2794.2 2574.8 2333.1 2433.2 2517.3 2450.6 2403.1 2702.9 2350.5 2142.2 2521.0 0.62 -1.34 -0.95 1.18 -0.59 0.72 0.92 0.61 -2.72 -1.18 -0.25 -1.76 -0.66 0.62 -0.81 -1.51 1.05 0.83 0.31 1.10 0.67 -1.49 2.01 1.05 -0.49 -0.89 1.43 1.64 1.57 1.51 0.71 1.07 700 t?(t+ l)= kw12 1 +x’(QR,(t- 1)x(t) (48) where y is a constant inversely proportional to the prediction error e(t). 6. IMPLEMENTATION RESULTS In order to evaluate the performance of the fuzzy expert system, load forecasting is performed on a typical utility data. The models ANN, FES,, and FES, are tested on a 2-year utility data for generating peak and average load profiles and some of the results are given in the subsequent subsections. In (Bunn and Farmer (1989) and Brace (199 1)) it has been shown that ANN gives the best prediction and accuracy compared to conventional approaches. So in this case the results of FES, and FES, are compared to that of the ANN approach. The training sets are formed separately for each of the seven day types (i.e., Tuesdays through Thursdays, Mondays, Fridays, Saturdays, Sundays, Holidays). The selection of training patterns and the selection of variable ranges are given in Sections 6.1 and 6.2. 6.1. Optimum Selection of Training Patterns The utility data studied here are susceptible to large and sudden changes in weather and load, so selection of appropriate training cases plays a vital role in training the network. Several techniques for the selection of training patterns have been suggested in Peng, Hubele, and Karady (1992). The following load model is used for peak load forecasting: P,,(i) =f(P,,,(i--n), P,,,(i-n-l), . . ., P,,(i-n-n,), O,,,(i--n), O,,,(i), O,,,(i- l), . . ., OmaxG-dr H,,,(i-n), H,,,(i), H&i-l), Hybrid Neural Network for Fuzzy Expert System TABLE 4 Peak Load Forecasting in January (Winter) Using 166 Hour Ahead Forecast 417 Peak Peak Peak Actual Forecast Forecast Forecast Load (NV Day (MW (MY 1 2690 2534.5 2 2628 2556.0 3 2703 2610.8 4 2592 2651.4 5 2530 2616.5 6 2574 2503.5 7 2389 2433.2 8 2513 2433.1 9 2500 2353.8 IO 2450 2324.3 11 2551 2620.1 12 2763 2953.9 13 2603 2531.9 14 2914 2761.6 15 2761 2894.6 16 2514 2389.6 17 2543 2697.4 18 2435 2494.7 19 2496 2463.8 20 2551 2456.6 21 2813 2886.4 22 2537 2382.0 23 2381 2279.8 24 2459 2540.6 25 2505 2575.4 26 2429 2500.7 27 2438 2281.2 28 2748 2839.0 29 2388 2513.8 30 2175 2125.4 31 2539 2456.2 MAPE Iterations required 5.78 2.74 3.41 -2.29 -3.42 2.74 -1.85 3.18 5.85 5.13 -2.71 -6.91 2.73 5.23 -4.84 4.95 -6.07 -2.45 1.29 3.70 -2.61 6.11 4.25 -3.32 -2.81 -2.95 6.43 -3.31 -5.27 2.28 3.26 3.87 2050 2630.0 2564.7 2649.5 2636.6 2580.1 2519.9 2351.5 2444.9 2370.8 2340.5 2510.7 2886.0 2555.9 2853.1 2658.3 2408.4 2653.9 2454.7 2509.2 2493.9 2864.5 2441.9 2285.0 2518.5 2538.8 2480.5 2351.0 2808.2 2481.8 2198.9 2476.3 2.23 2.41 1.98 -1.72 -1.98 2.10 1.57 2.71 5.17 4.47 1.58 -4.45 1.81 2.09 3.72 4.20 -4.36 -0.81 -0.53 2.24 -1.83 3.75 4.03 -2.42 -1.35 -2.12 3.57 -2.19 -3.93 -1.10 2.47 2.61 1170 2742.2 2571.0 2666.5 2551.0 2480.7 2628.8 2354.8 2449.7 2378.5 2360.1 2519.6 2874.1 2649.6 2972.3 2660.2 2415.2 2651.8 2424.8 2481.5 2514.5 2762.9 2462.2 2285.0 2507.4 2537.8 2462.3 2370.2 2784.3 2474.2 2159.3 2489.2 -1.94 2.17 1.35 1.58 1.95 -2.13 1.43 2.52 4.86 3.67 1.23 -4.02 -1.79 -2.00 3.65 3.93 -4.28 0.83 0.58 1.43 1.78 2.95 4.03 -1.97 -1.31 -1.37 2.78 -1.32 -3.61 0.72 1.96 2.29 890 . ., H&i-4)). (49) For average load forecast a similar model as eqn. (49) is chosen, given by: P,,(i) =KP,,(i-n), P,,(i- l), . . ., PJi-n,), %di-n), O,,,W~ O,,,G- I), . . ., Omax(i-Ut OmJi-nh O,i,(& O,,,(i- 11, . . ., O,,,U-n,L %,,(i-n), H,,,(i), ff,,A-I), . . ., ~,,,(i-~J). H,,,(i-fi), HmilI(i)7 H,i,(i- I), . . ., H,,,,(i- 0. (50) Here, n indicates the lead time of the forecast (n * n,, n2, n?). P, 0, and H stand for load, temperature, and humidity, respectively. Also different values of n,, n,, and n3 were tested to find the effect of the past data on the load forecast. With n,, n,, n,>O there was no marked improvement in the results. Therefore, we chose, n1 = n2 = n3 = 0 in this study. However, the choice of n,, n2, n3 depends entirely on the utility data concerned. 6.2. Scaling of the Variable Range Because the input variables and the estimated ones from the hybrid neural network have wide variations in magnitudes, they will cause convergence problem and the system will behave completely erratic. To circumvent this problem, the variables are scaled between 0.1 and 0.9 (Rahman, Drezga, & Rajagopalan, 1993). This is performed so as to maximise accuracy and minimise training time. The following scheme is adopted to scale the variables between 0.1 and 0.9. Let X,,,,, and X, m,n denote the upper and lower bounds of the observed range of feature X, in all the patterns in the historical data base considering numerical values only. Similarly, O,,,,, and 0 ,, m,n denote the upper and lower bounds of the observed range of outputs. Then X, is normalised as X: = 0.1 + WW-X. ,,,MX.,,,-X,. n,,n II (51) where (i, j) corresponds to the jth pattern of the ith training set. 418 P. K. Dash et al. TABLE 5 Average Load Forecasting in January (Winter) Using 24 Hour Ahead Forecast Peak Peak Peak Actual Forecast Forecast Forecast Load WJ) Day WV (M'W (%) 'FG (FpE& 'Kz &I,, 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 23: 31 MAPE Iterations required 2166 2223.4 2153 2209.8 2233 2218.1 2093 2092.1 2050 2041.0 2109 2109.3 1937 1933.4 1983 1974.9 2045 1991.6 2003 2036.6 1982 1954.9 2117 2116.2 2064 2035.6 2237 2262.3 2198 2239.1 2068 2065.3 2021 2057.0 1976 1953.7 2019 2019.9 2073 2090.0 2195 2224.6 2015 1982.8 1917 1866.8 1977 1943.5 2021 2017.5 1977 1961.6 1964 2010.8 2086 2124.2 1876 1889.6 1772 1751.3 1960 200.1 -2.65 -2.64 0.67 0.04 0.44 -0.01 0.18 0.41 2.61 -1.68 1.37 0.04 1.37 -1.13 -1.87 0.13 -1.78 1.13 -0.04 -0.82 -1.35 1.60 2.62 1.70 0.17 0.78 -2.38 -1.83 -0.72 1.17 -2.05 1.21 1700 2161.2 2192.4 2242.3 2092.6 2050.1 2107.9 1934.2 1969.4 2008.7 2007.1 1980.2 2116.8 2042.8 2241.5 2171.0 2065.2 2047.9 1970.8 2030.1 2053.3 2175.8 2041.4 1935.1 1961.3 2017.1 1965.6 1921.8 2078.7 1889.4 1757.2 1928.1 0.22 -1.83 -0.42 0.02 -0.00 0.05 0.14 0.68 1.78 -0.21 0.09 0.01 1.03 -0.20 1.23 0.14 -1.33 0.26 -0.55 0.95 0.87 -1.31 -0.94 0.80 0.20 0.58 2.15 0.35 -0.72 0.83 1.63 0.69 1020 2163.5 2120.8 2216.6 2089.0 2050.0 2108.2 1936.9 1976.2 2080.1 1998.2 1980.5 2116.6 2049.9 2240.6 2221.3 2067.1 1994.9 1982.1 2007.3 2087.2 2203.5 2003.3 1936.8 1961.0 2024.1 1970.9 1939.0 2077.9 1883.0 1765.5 1924.6 0.12 1.49 0.73 0.19 -0.00 0.04 0.00 0.34 -1.72 0.24 0.07 0.02 0.70 -0.17 -1.06 0.04 1.29 -0.31 0.58 -0.68 -0.39 0.58 -1.03 0.81 -0.15 0.31 1.27 0.39 -0.37 0.37 1.81 0.56 680 Similarly, the output 0, can be expressed 0~=0~1+[0~8(0~~0~,~~)/(0~,~~~~0~,~~~)] (52) where X7 and 07 denote the normalised input and output vectors, respectively. The normalised predicted values can be converted back to the actual values using the above expressions. The ideas discussed in sections 6.1. and 6.2 are used for obtaining the peak and average load forecasts. 6.3. Peak Load Forecasting For n h ahead peak load forecasting, the following training data are used for the ANN: Input pattern: P,,,(i), O,,(i), H,,,(i), o’,,,(i + n), H/,,( i + n) Output pattern: P ,,,(i + n) for the ANN. Superscript f denotes the forecasted values for @ and H. The forecasted values for (i + n)th day are used to get a more realistic forecast. For FES, and FES, the training patterns used are: Input pattern: A@,,,(i, i + n) and AH_(i, i + n) Output pattern: erc(i), the desired load correction. The P,,(i + n) for FES, and FES, is obtained using eqn. (7). Table 1 gives the learned membership functions for 24 h ahed peak load forecasting in winter using FES,, for example, rule 0 is interpreted as: RO: IF A@,,, is term 0 and AH_ is term 3 THEN e^, is term 7. Figure 2 gives the learned membership functions after the first phase (unsupervised learning phase) and the Hybrid Neural Network for Fuzzy Expert System TABLE 6 Average Load Forecast in June (Summer) Using 46 Hour Ahead Forecast -- Peak Peak Peak Actual Forecast Forecast Forecast Load NV Day (MW) WW 1 1521 2 1531 3 1473 4 1336 5 1285 6 1433 7 1406 a 1403 9 1428 10 1431 11 1301 12 1235 13 1418 14 1408 15 1421 16 1447 17 1440 la 1352 19 1305 20 1430 21 i 383 22 1414 23 1423 24 1406 25 1267 26 1221 27 1397 28 1401 29 1403 30 1411 MAPE Iterations required 1541.1 -1.32 1562.5 -2.06 1455.2 1.21 i 380.9 -3.36 1254.0 2.41 1446.2 -0.92 1375.5 2.17 1442.3 -2.80 1461.0 -2.31 1467.3 -2.54 1272.8 2.17 1206.7 2.29 1394.7 1.64 1395.5 0.89 1451.3 -2.13 1480.1 -2.29 1420.7 1.34 1313.9 2.82 1279.8 1.93 1407.5 1.57 i 387.8 -0.35 1434.1 -1.42 1386.9 2.54 1361.6 3.16 1230.6 2.87 1254.3 -2.73 1371.7 i .ai 1374.1 1.92 1415.2 -0.87 1376.1 2.47 2.01 1630 1509.7 1512.3 1453.1 1370.3 1260.8 1443.5 1386.6 1437.0 1450.7 i 459.8 1286.0 1217.8 1419.3 1401.9 1443.0 1468.4 1425.3 1324.8 1282.6 1416.6 1395.3 1425.0 1404.1 1370.4 1250.3 1207.9 1401.7 I 384.3 1416.5 1393.6 0.74 1.22 1.35 -2.57 1.88 -0.73 1.38 -2.42 -1.59 -2.01 1.15 1.39 -0.09 0.43 -1.55 -1.48 1.02 2.01 1.72 0.94 -0.89 -0.78 1.33 2.53 1.32 1.07 -0.34 1.19 -0.96 1.23 1.31 1520 1516.7 1549.5 1456.5 1305.9 1272.9 1423.5 1418.5 1418.0 i 448.8 1411.3 1322.7 1219.9 1419.8 I 399.8 1439.6 1429.9 i 428.3 1330.5 1290.3 1440.3 i 389.2 1401.1 1410.9 1373.1 1248.6 1234.4 1395.9 I 384.9 I 389.3 1390.1 0.28 -1.21 1.12 2.25 0.94 0.66 -0.89 -1.07 -1.46 i .3a -1.67 1.22 -0.13 0.58 -1.31 1.18 0.81 1.59 1.13 -0.72 -0.45 0.91 0.85 2.34 1.45 -1.10 0.08 1.15 0.98 1.48 i .oa 1070 419 second phase (supervised learning phase). Figure 3 gives the mean absolute percentage errors (MAPE’s) versus the number of iterations. The results in Figures 2 and 3 are obtained for 24 h ahead load forecasting in January (winter). From Figure 3 we find that the FES, gives fastest convergence followed by FFS, and ANN. The convergence speed of the FES, was found to be superior because of the linear Kalman filter equations used for weight update and the error-dependent forgetting factor was responsible for driving the MAPE low during the first few hundred iterations until the bias introduced by the initial conditions was eliminated. Table 2 gives the peak load forecasting results in terms of mean absolute percentage errors (MAPEs) for ANN, FES, and FES, in the month of June (winter) using 24 h ahead forecast. Tables 3 and 4 give the results for the month of January (winter) using 48 h and 168 h ahead forecast, respectively. From Tables 2,3 and 4 we see that the FES, and FES, give better prediction accuracy compared to ANN. Also we find that the results for 48 h and 168 h predictions are comparable with that of the 24 h ahead predictions. This is because the load forecasting was performed as a one- step process (i.e. looking 24 h ahead, 48 h ahead, and so on). However, as the lead time increased to 168 h the PEs were found to be greater than 4% even with FES,. As the primary aim of this paper was to make a comparative assessment between ANN, FES, and FES,, no attempt was made to improve the forecast errors further. 6.4. Daily Average Load Forecast For n h ahead average load forecasting, the following training data are used for ANN. Input pattern P,,(i), O,,(i), O,,,(i), H,,,(i), &&). @L(i + n), o’,,,(i + n). Hf,,,(i + n), Hti,(i + n) Output pattern: P,,(i + n) for ANN. For FES, and FES,, the training patterns used are: 420 R K. Dash et al. TABLE 7 Average Load Forecasting in June (Summer) Using 168 Hour Ahead Forecast Peak Peak Actual Forecast Forecast Load VW Day (MW) (MW (K, 'Kz cFPE:,, 1 z 4 5 ; : 10 11 12 13 14 15 16 17 18 19 20 21 22 23 z; 26 27 28 z: MAPE Iterations required 1521 1558.6 1531 1558.9 1473 1436.0 1336 1409.3 1285 1244.1 1433 1409.4 1406 1377.5 1403 1441.2 1428 1484.1 1431 1469.6 1301 1358.0 1235 1202.5 1418 1390.3 1408 1394.9 1421 1484.4 1447 1482.3 1440 1402.7 1352 1308.5 1305 1269.2 1430 1404.0 1383 1350.4 1414 1463.1 1423 1371.5 1406 1359.5 1267 1219.4 1221 1256.7 1397 1372.7 1401 1373.1 1403 1414.4 1411 1359.6 -2.47 -1.82 2.51 -5.49 3.18 1.65 2.03 -2.72 -3.93 -2.70 -4.38 2.63 1.95 0.93 -4.46 -2.44 2.59 3.22 2.74 1.82 2.36 -3.47 3.62 3.31 3.76 -2.92 1.74 1.99 -0.81 3.64 2.78 2650 1542.9 1551.8 1452.7 1388.2 1250.8 1428.4 1385.3 1435.3 1471.8 1447.0 1345.8 1212.5 1403.0 1399.7 1453.7 1473.2 1412.1 1357.7 1290.0 1406.3 1363.2 1448.6 1393.0 1372.1 1237.6 1241.6 1384.6 1379.7 1405.5 1374.9 -1.44 -1.36 1.38 -3.91 2.66 0.32 1.47 -2.30 -3.07 -1.12 -3.44 1.82 1.06 0.59 -2.30 -1.81 1.94 -0.42 1.15 1.66 1.43 -2.45 2.11 2.41 2.32 -1.69 0.89 1.52 -0.18 2.56 1.76 2480 Peak Forecast 1538.2 1550.6 1456.5 1364.9 1261.6 1434.6 1385.9 1431.1 1448.7 1416.5 1330.0 1221.7 1432.0 1411.0 1451.8 1463.8 1415.5 1358.1 1295.0 1441.9 1399.7 1396.6 1404.8 1389.7 1254.2 1236.1 1387.9 1380.5 1399.5 1378.1 -1.13 -1.28 1.12 -2.16 1.82 -0.11 1.43 -2.00 -1.45 1.01 -2.23 1.08 -0.99 -0.21 -2.17 -1.16 1.70 -0.45 0.77 -0.83 -1.21 1.23 1.28 1.16 1.01 -1.24 0.65 1.46 0.25 2.33 1.23 1310 Input pattern: A@,,,,(& i + n), AQmin(ir i + n), AH&i, i + n), AHmi,(i, i + n) Output pattern: e,,(i), the desired load correction. Again forecasted temperature and humidity values are used for the day of the forecast. The P,,(i+n) for FES, and FES, is obtained using eqn. (7). Table 5 gives the average load forecasting results, number of iterations for convergence, PEs and the MAPEs for the ANN, FES,, and FES, models in the month of January (winter) using 24 h ahead forecast. Tables 6 and 7 give the results for 48 h and 168 h ahead predictions in the month of June (summer). Again, from Tables 5, 6, and 7 we find the improved performance of FES, in terms of faster convergence and improved overall accuracy over the ANN and FES,. 7. DISCUSSION The fuzzy expert system presented in this paper is constructed from training examples by machine learning techniques, and the neural network model is trained to develop fuzzy logic rules and find input/output member- ship functions. By combining both unsupervised and supervised learning schemes, the learning speed for time series forecasting problems converges much faster than original back-propagation learning algorithm. Further, by using a linear Kalman filter for the supervised learning phase, the learning time is shortened with rapid con- vergence due to the adaptive nature of the learning algorithm. The fuzzy expert system using a linear Kalman filter produces a more accurate load forecast for lead times varying from 24 h to 168 h in comparison to the one using gradient descent back-propagation algo- rithm. The forecasting results presented in Section 6 are based upon only true temperature information because all the data for this utility are historical. However, in reality the temperature information will be the forecasted value and this will add to the forecast error. The developed expert systems are used to forecast the load during weekends but no attempt is made to forecast for days with unusual events. Such events will require Hybrid Neural Network for Fuzzy Expert System 421 extensive data analysis to track the events in that day and to select training cases accordingly from the previous days with similar events. Such activities are continuing, and will be reported in a future paper. As a final note, a shorter period is used for training as load patterns change very fast. But training and predic- tion can be done over longer periods with less chaotic data having strong correlation to any particular weather and environmental variable. This flexibility permits the adaptation of the proposed fuzzy expert systems for producing accurate load forecasts for different geo- graphic areas. 8. CONCLUSIONS This paper presents the development of a fuzzy expert system using a hybrid neural network approach to predict peak and daily average load profiles in an energy management system. The fuzzy expert system is mod- elled as a hybrid neural network and uses fuzzy membership values of load and weather parameters. A hybrid learning scheme consisting of both self-organized and supervised learning phases is used for training the hybrid neural network. Further, the paper presents simulation results using both back-propagation and Kalman filter-based algorithms, the latter producing faster convergence during training and more accurate predictions. Acknowledgement-The authors gratefully acknowledge the funds from the National Science Foundation (NSF Grant No. INT-9209103 and INT-9 117624). USA for undertaking this research. REFERENCES Box, G. E. P., & Jenkins, G. M. (1976). Time series onolysisforecosring and control, Oakland, CA: Holden-Day. Brace, M. C. (1991). A comparison of the forecasting accuracy of neural networks with other established techniques, Firsr Inrer- national Forum on Applications on Neural Networks to Power Systems, Seattle, WA, July, 1991, pp. 23-26. Buckley, J. J., Hayashi, Y., & Czogala, E. (1993) On the equivalence of neural nets and fuzzy expert systems, Fuzzy Sets Systems, 53 129-134. Bunn. D. W.. & Farmer, E. D. (1989). Comparative models for electric loadforecasting, New York: John Wiley and Sons. Dash, P. K., Dash, S., Rama Krishna, G., & Rahman, S. (1993). Forecasting of a load time series using a fuzzy expert system and fuzzy neural networks, Engineering International Systems, l(2), 103-117. Dash, P. K., Satpathy, J. K., Rama Krishna, G., & Rahman. S. (1993). An improved kalman filter based neural network approach for short-term load forecasting, Proceedings of the Third International Symp. on Electricity Distribution and Energy Management, Vol. 1, pp. 3-8. El-Sharkawi, M. A., Oh, S., Marks, R. J., Dambourg. M. J., & Brace, C. M. (1991). Short term electric load forecasting using an adaptively trained layered perceptron, First International Forum on Applications of Neural Networks to Power Systems, Seattle WA., July 23-26, 199 1, pp. 3-6. Hayashi, Y., Buckley, J. J., & Czogala, E. (1992). Fuzzy expert systems versus neural networks, Proceedings of International Joint Con- ference on Neural Networks, Baltimore, Vol. II, June 7-12 pp. 720-726. Haykin, S. S. (1986). Adaptive filter theory (pp. 3 12-3 14). Englewood Cliffs, NJ: Prentice-Hall. Ho, K. L., Hsu, Y. Y., &Yang, C. C. (1992). Short term load forecasting using a multilayer neural network with an adaptive learning algorithm, IEEE Transactions on Power Systems, 7( 1). 141-I 50. Lee, C. C. (1990). Fuzzy logic in control systems: Fuzzy logic controller, part 2. IEEE Transactions Systems, Man & Cybernetics, 20(3), 266-275. Lee, K. Y., Cha, Y. T., & Park, J. H. (1992). Short term load forecasting using an artificial neural network, IEEE Transactions on Power Systems, 7(I), 124-133. Lin, C. T., & Lee, G. C. S. (1991). Neural network-based fuzzy logic control and decision system, IEEE Transactions on Computers, 40(12), 1320-1336. Moghram, I., & Rahman. S, (1989). Analysis and evaluation of five short-term load forecasting techniques, lEEE Transactions on Power Systems, 4(4), 1484-1490. Pal, S. K., & Mitra, S. ( 1992). Multilayer perceptrons, fuzzy sets and classifications, IEEE Transactions on Neural Networks, 3(5), 683-698. Park, D. C., El-Sharkawi, M. A., & Marks, R. J., (1991). Adaptively trained neural networks, IEEE Transactions on Neural Networks, 224-245. Park, J. H., Park, Y. M.. & Lee, K. Y. (1991). Composite models for adaptive short term load forecasting, lEEE Transactions on Power Systems, l(2), 450-157. Peng, T. M., Hubele, N. F., & Karady, G. G. (1992). Advancement in the application of neural networks for short term load forecasting, IEEE Transactions on Power Systems, 7(l), 250-258. Rahman, S., Drezga, I., & Rajagopalan, J. (1993). Knowledge enhanced connectionist models for short-term electric load fore- casting, International Conference on ANN applications to Ponaer Systems, Yokohama, Japan, April. Rumelhart, D. E., & Zipses, D. (1985). Feature discovery by competitive learning. Cognitive Science, 9, 75-l 12. Scalero. R. S., & Tepedelenlioglu, N. (1992). A fast algorithm for training feedforward neural networks, IEEE Transactions on Signal Processing, 40( 1). 202-2 10. Singhal, S., & Wu, L. (1989). Training feed-forward network with the extended Kalman algorithm, Proceedings of ICASSP, pp. 1187- 1190. Singhal, S., & Wu, L. (1989). Training feed-forward network with the extended Kalman algorithm. Proceedings of ICASSP, pp. 1187-1190. Wang, L.-X. & Mendel, J. M. (1992). Generating fuzzy rules by learning from examples, IEEE Transacrions on Systems Man & Cybernetics, 22(6), 1414-1427. Zadeh. L. A. (1965). Fuzzy sets, Information and Control. 8, 338-358. 
work_psml6m62qjcxvnyhwfreu372ma	----	Design of an optical content-addressable parallel processor for expert systems Design of an optical content-addressable parallel processor for expert systems Ahmed Louri and Jongwhoa Na The slow execution speed of current rule-based systems 1RBS’s2 has restricted their application areas. To improve the speed of RBS’s, researchers have proposed various electronic multiprocessor systems as well as optical systems. However, the electronic systems still suffer in performance from the large amount of required time-consuming pattern-matching and comparison operations at the core of RBS’s. And optical systems do not fully exploit the available parallelism in RBS’s. We propose an optical content- addressable parallel processor for expert systems. The processor executes the three basic RBS operations, match, select, and act, in a highly parallel fashion. Additionally, it extracts and exploits all possible parallelism in a RBS. Distinctive features of the proposed system include the following: 112 two-dimensional representation of data 1knowledge2 and control information to exploit the parallelism of optics in the three RBS units; 122 capability of processing general-domain knowledge expressed in terms of variables, numbers, symbols, and comparison operators such as greater than and less than; 132 the parallel optical match unit, which performs the two-dimensional optical pattern matching and compari- son operations; 142 a novel conflict-resolution algorithm to resolve conflicts in a single step within the optical select unit. The three units and the general-knowledge representation scheme are designed to make the optical content-addressable parallel processor for expert systems suitable for any high-speed general-purpose RBS. Key words: Optical content-addressable parallel processor, rule-based system, inference engine, optical parallel conflict resolution. 1. Introduction Rule-based systems 1RBS’s2 are one of the problem- solving methodologies that have been developed by artificial-intelligence researchers.1 RBS’s have a vast potential in several application areas because of the modularity, maintainability, and expressibility of the knowledge base as well as the simplicity of control. The RBS types include classification, selection, diag- nosis, design, planning, and interpretation systems for such problems as computer-aided design, medi- cine, configuration tasks, manufacturing, and oil ex- ploration.2–4 In spite of these large potential applica- tion areas, RBS’s have not been used as widely as conventional problem-solving methods that use a programming language such as C. One major rea- son is the slow execution speed of the underlying architectures implementing the RBS. The problem The authors are with the Department of Electrical and Computer Engineering, University ofArizona, Tucson,Arizona 85721. Received 21 March 1994; revised manuscript received 12 Septem- ber 1994. 0003-6935@95@235053-11$06.00@0. r 1995 Optical Society ofAmerica. stems from the many pattern-matching and compari- son operations necessary to solve a given query in a RBS. To overcome this fundamental speed problem, re- searchers have developed new algorithms and new architectures tailored for RBS implementations. The new algorithms include the TREAT optimizing compiler for sequential RBS’s,5 the PSM-E parallel compiler for RBS’s,6 and the SWARM parallel- programming environment.7 The new architectures include shared memory and message-passing multi- processor systems such as the DADO multiprocessor,8 the NON-VON multiprocessor,9 and data-flow comput- ers such as the data-driven parallel-production sys- tem.10 Although these proposed systems incorporate new algorithms and new hardware designed to im- prove performance over that of sequential RBS’s, they do not show any significant performance increase at present owing to the communication overhead and synchronization problems.11,12 Optics has been introduced as an alternative to improve the speed of RBS’s because of the ability to represent knowledge in two-dimensional 12-D2 space and because of natural implementation of the parallel pattern-matching and comparison operations.12–15 10 August 1995 @ Vol. 34, No. 23 @ APPLIED OPTICS 5053 Previously proposed optical expert systems include a matched-filter inference engine using a classical VanderLugt matched filter14 and an optical expert system based on an optical vector–matrix multiplier.15 However, these systems do not fully utilize all the available parallelism in a RBS, particularly rule-level parallelism in which more than one rule is fired at a time. Recently the authors introduced a new optical system called the electro-optical rule-based system 1EORBS2 for the parallel implementation of RBS’s.12 Although the EORBS fully exploits the available parallelism in RBS’s, the system still lacks generality. In this paper we extend the concept of the optical content-addressable parallel processor16 1OCAPP2 to a novel architecture designed specifically for parallel EORB’s, known as the optical content-addressable parallel processor for expert systems 1OCAPP-ES2. The OCAPP-ES executes the three basic RBS opera- tions 1match, select, and act2 in a highly parallel fashion. Additionally, it extracts and exploits all possible parallelism in a RBS. Distinctive features of the proposed system include the following: 112 2-D representation of data 1knowledge2 and control infor- mation; 122 capability of processing general-domain knowledge expressed in terms of variables, numbers, symbols, and comparison operators such as greater than and less than; 132 the parallel optical match unit, which performs the 2-D optical pattern matching and comparison operations; 142 a novel conflict-resolution algorithm to resolve conflicts in a single step within the optical select unit. The three units and the general-knowledge representation scheme are de- signed to make OCAPP-ES suitable for any high- speed general-purpose RBS. 2. Background An expert system can be defined as an intelligent system that can mimic some part of human intelli- gence. As shown in Fig. 1, an expert system is composed of 112 a RBS and knowledge-acquisition facility, 122 an explanation facility, and 132 a user interface.1 The RBS performs inferencing using the knowledge base and the inference engine. The Fig. 1. Logical block diagram of an expert system. 5054 APPLIED OPTICS @ Vol. 34, No. 23 @ 10 August 1995 knowledge-acquisition facility and the user interface act as an interface unit between the RBS and the user. The explanation facility explains to the user the way results have been obtained. The knowledge base is composed of rules representing universal knowledge and facts representing current knowledge. A rule conjoins condition elements Ci and action elements Ai with the following format: condition elements action elements6 6 if 1C1 ` C2 ` C3 . . . ` Cn2 ⇒ then 1A1, A2, . . . , Am2, where ` is a logical AND operator. The condition elements and the action elements are represented by facts, which are the basic units of knowledge in a RBS. The inference engine uses the modus ponens theo- rem as an underlying principle. The modus ponens theorem states that, if there is an axiom of the form E1 ⇒ E2 and there is another axiom of the form E1, then E2 logically follows.17 The inference engine applies modus ponens as follows: current fact rule inferred fact6 6 6 51A is true2 ` 1if A then B26 ⇒ 1B is true2. Thus with known facts and rules the inference engine can produce new facts in either a forward- chaining system or a backward-chaining system or both.17 In the forward-chaining system the system tries to find final 1goal2 states by comparing the known facts, which describe the initial states, with the condition-specifying if parts of the rules. In the backward-chaining system the system tries to find initial hypotheses by comparing the known facts, which now describe the final 1goal2 states, with the action-specifying then parts of the rules. The basic operations of a RBS are as follows: Match. For each rule, determine whether the con- dition part of the rule matches the current facts. If a rule satisfies the condition part, the rule is added to the conflict set. Otherwise, the rule is discarded. A conflict set is a set of triggered rules that have satisfied condition parts. Select. If the conflict set is empty, the inference engine stops inferencing and reports the failure to the user. If the conflict set is not empty and contains more than one rule, the inference engine selects one rule from the conflict set by applying a conflict- resolution strategy such as rule ordering.18 Act. The inference engine fires the selected rule by executing its action. Because the current facts are changed by the fired rule, it is important to check whether the changes agree with predefined goals. If the goals are satisfied, the inference engine stops inferencing and reports results to the user. Other- wise, the inference engine must continue until the goal is satisfied or there are no more matching rules. In the following we discuss knowledge representation and operation of the OCAPP-ES. 3. Overview of the Optical Content-Addressable Parallel Processor for Expert Systems The main objective of the OCAPP-ES is to exploit the maximum possible parallelism in a RBS. Because optical devices are two dimensional in nature, 2-D data can be processed simultaneously with optics. To exploit the parallelism available in optics fully, we represent knowledge-base and control information in a 2-D optical plane. In what follows we describe the OCAPP-ES and a knowledge-representation scheme for OCAPP-ES. A. Description of the Optical Content-Addressable Parallel Processor for Expert Systems Figure 2 shows a block diagram of the OCAPP-ES. The OCAPP-ES consists of an optical subsystem and an electronic subsystem. The optical subsystem is intended to implement the most time-consuming op- erations of RBS, namely, pattern-matching and com- parison operations, and the electronic subsystem implements the flexible control of RBS’s. The two subsystems complement each other. The optical sub- system consists of an optical match unit and an optical select unit. The optical match unit, utilizing an OCAPP, compares in a single step all the input facts to the input conditions of the rules and outputs a list of selected rules. The optical select unit then resolves conflicts among the selected rules and sends a list of conflict-free triggered rules back to the electronic subsystem. The electronic subsystem con- sists of a detector controller, an electronic act unit, and a spatial-light-modulator 1SLM2 controller. The detector controller converts the optical signal from the optical select unit into an electronic format. The act unit then executes the action elements of the triggered rules and sends the execution results to both the front-end computer and the SLM controller. The SLM controller is used to convert the electronic information into an optical signal. The front-end computer determines whether the desired solution has been obtained by checking the newly inferred facts from the OCAPP-ES. If the desired state is Fig. 2. Block diagram of the OCAPP-ES. reached, the front-end computer stops further opera- tion of the OCAPP-ES and returns control to the user. B. Optical Knowledge Representation For the optical subsystem of the OCAPP-ES, facts are represented in a one-dimensional 11-D2 fact vector 1FV2 and the condition parts of rules are represented in a 2-D plane called the condition template 1CT2, as shown in Fig. 3. The CT and the FV are used as inputs to the optical match unit. As an illustration of knowl- edge representation and the operation of the OCAPP-ES throughout this paper, consider the follow- ing going-to-the-theater example knowledge base: @* ‘Going to the theater’ input rulebase *@ Rule 1: If distance . 5 miles Then means 5 drive Rule 2: If distance . 1 mile and time , 15min Then means 5 drive Rule 3: If distance . 1 mile and time . 15min Then means 5 walk Rule 4: If distance . 3 miles and time . 30min Then means 5 walk Rule 5: If means 5 drive and location 5 down- town Then action 5 take_a_cab Rule 6: If means 5 drive and location 5 suburb Then action 5 drive_your_car Rule 7: If means 5 walk and weather 5 bad Then action 5 take_a_coat_and_walk Rule 8: If means 5 walk and weather 5 good Then action 5 walk This knowledge base decides how one can go to the theater under various circumstances. In this knowl- edge base the rule condition parts require three different logical operations: equality, greater than, and less than comparisons. It has been shown that the equality comparison can be easily done with optical devices performing an XOR operation.19 However, optical implementation of the magnitude comparison operations such as greater than and less than turns out to be a difficult task. Although these comparison operations have been shown to be feasible for optical implementation, they require a fair num- ber of optical logic devices,16,20 which can dominate cost and propagation delay. Fig. 3. Templates for optical knowledge representation: Each cell of the fact vector 1FV2 represents a fact variable, and the contents of the cell represent the value of the variable. In the condition template 1CT2 each row represents a rule and each column represents a condition element. Each fact variable, FVj, is used as a condition-element variable CT1i, j 2, where i represents the rule number and j represents the FV index. 10 August 1995 @ Vol. 34, No. 23 @ APPLIED OPTICS 5055 To reduce the number of required optical logic devices and the design complexity while enhancing system feasibility, we introduce a solution that can evaluate the greater-than and less-than operations with only the equality-checking optical hardware. In the proposed method, instead of performing a greater-than or less-than comparison with a specific number at the left-hand side of each condition ele- ment, we can perform a logically equivalent operation. As an illustration, consider the condition element 1distance . 1 mile2 of rule R1 of the going-to-the- theater example rule base. This condition element becomes true when the variable distance receives any number greater than 1. In other words the condition element compares whether the input value of the variable distance is in the range from 1 mile to ` miles. Thus by assigning a unique symbolic interval constant distance_set_A for the interval from 1 mile to ` miles and by performing the equality comparison with the symbolic interval constant distance_set_A, we can obtain the greater-than comparison result. The above translation method can be generalized as follows: First, we assign an appropriate symbolic constant for both the condition element related to the greater-than or less-than comparison and the value of the fact element used as the condition element. Next, we perform an equality comparison instead of a greater-than or less-than comparison between the input value of the variable 1left-hand side of a condi- tion element2 and the prescribed symbolic constant 1right-hand side of a condition element2. With the translation method the greater-than and less-than checking condition elements of the going-to-the- theater example rule base are translated into the condition elements with the symbolic interval con- stants and equality comparators, as shown in Table 1. This method replaces greater-than and less-than comparison operations using exact values for a vari- able with equality comparison with a symbolic inter- val constant representing the corresponding interval. However, owing to the nature of expert systems 1i.e., symbolic computation in the decision-making environ- ment rather than numerical computation2, the trans- lation method will affect neither the overall perfor- mance of the system nor the generality of the proposed system. The number of symbolic interval constants created for a variable will be small considering that the major application domain of expert systems is in Table 1. Translation of Condition Elements with Greater-Than or Less-Than Comparison Operators into Condition Elements with Corresponding Symbolic Constants Input Condition Elements with Magnitude Comparison Operators 1., ,2 Translated Condition Elements with Symbolic Interval Constant distance . 1 mile distance 5 dist_set_A distance . 3 mile distance 5 dist_set_B distance . 5 mile distance 5 dist_set_C time , 15 min time 5 time_set_A time . 15 min time 5 time_set_B time . 30 min time 5 time_set_C 5056 APPLIED OPTICS @ Vol. 34, No. 23 @ 10 August 1995 business and management3 rather than numerical computation. Even if the number of created sym- bolic interval constants exceeds the capacity that one variable can represent, the front-end computer can easily create another variable to share the remaining symbolic interval constants. The translation method requires preprocessing in the host computer during the rule compile time. Hence this will not affect the execution time of the OCAPP-ES. On the other hand, the additionally created variable will take up one additional variable slot in the optical match@select unit. However, owing to the small number of origi- nal symbolic interval constants and the nonnumerical nature of expert systems, the additionally occupied variable slot should not cause any problem. This translation process can be regarded as a special case of fuzzy logic.21,22 From Table 1, the original going-to-the-theater example rule base is modified into the translated rule base with equality-checking elements only as follows: @* Translated rulebase for the ‘Going to the the- ater’ example *@ Rule 1: If distance 5 distance_set_C Then means 5 drive Rule 2: If distance 5 distance_set_A and time 5 time_set_A Then means 5 drive Rule 3: If distance 5 distance_set_A and time 5 time_set_B Then means 5 walk Rule 4: If distance 5 distance_set_B and time 5 time_set_C Then means 5 walk Rule 5: If means 5 drive and location 5 down- town Then action 5 take_a_cab Rule 6: If means 5 drive and location 5 suburb Then action 5 drive_your_car Rule 7: If means 5 walk and weather 5 bad Then action 5 take_a_coat_and_walk Rule 8: If means 5 walk and weather 5 good Then action 5 walk Once the translated rule base is ready, fact vari- ables related to condition elements with a greater- than or less-than operator can be modified. For these variables the values of facts are translated into the corresponding symbolic constants instead of using numbers given by the user. For example, when the fact variable distance receives the number 2 as input, the translated value of the fact becomes dist_set_A. When the knowledge base is ready, the electronic host computer modifies the translated knowledge base into the execution format for the optical infer- ence engine by using templates, shown in Fig. 3. In Fig. 4 each cell of the CT and the FV represents the name of a variable. The cell contains the value of the variable. As an illustration, assume that we have a fact 1location 5 downtown2 and that it is assigned to the ith cell of the FV, FVi. Then FVi represents the variable name location, and the content of FVi be- comes downtown. Next, assume that the condition element of rule R1 1distance 5 distance_set_C 2 is to be assigned to the jth column of the ith row of the CT. Then, cell CT1i, j 2 will represent the variable distance, and the contents of CT1i, j 2 will have the number distance_set_C. Each cell of Fig. 4 is further divided into n pixels, creating an n-bit word-representation system. Fig- ure 5 shows the 8-bit binary assignment scheme for numbers and symbols 1including symbolic interval constants, symbolic variables, and values2 used in the going-to-the-theater example. We create the num- bers and symbols by distinguishing the most-signifi- cant bit, b7. If b7 is 1102, the word represents a symbol 1number2. Thus a number in the range from 00000000122 to 01111111122 can be represented. We further divide the symbols into the symbolic variable, the symbolic value, and the symbolic interval con- stant by distinguishing the next-most-significant bit, b6. If b6 is 1102, the word represents either a sym- bolic variable or a symbolic value 1a symbolic interval constant2. Then, we represent each unique symbol by assigning a unique binary number to the symbol, using the remaining bits. Once the symbol and number representations are set, the host computer builds the FV and the CT, as shown in Fig. 6. Each fact is assigned to a unique cell of the FV and a unique column of the CT. Then each cell of the FV is filled with the current value of the facts known to the user. In this example, the Fig. 4. Optical knowledge-representation example using a fact 1location 5 downtown2 for the FV and a condition element 1distance 5 dist_set_C 2 of the rule R1 for the CT. Fig. 5. Example of 8-bit binary value assignment for the vari- ables, symbolic values, symbolic interval constants, and numbers for the going-to-the-theater example. Numbers are identified by the 0 in the most-significant bit 1bit b72. Symbolic interval constants for intervals are identified by the 11 in bits b7 and b6. b5 and b4 are used to identify the name of symbolic interval constants for intervals, and the remaining bits are used to specify various intervals in a set. Symbolic variables and values are identified by the 10 in bits b7 and b6. The rest of the bits are used to identify various symbolic variables and values in this case. currently known value of fact distance is 10, the fact time is 20, the fact location is downtown, and the rest are unknown. Because the variables time and dis- tance belong to the interval evaluating variables, the host computer must assign symbolic constants for the values of these variables. For example, the value 20 of the variable time can be translated into time_set_B (11100010) because time_set_B becomes true for input value of time .15. On the other hand, the input number 10 of the variable distance will be translated into distance_set_D (11010ddd), where d represents don’t care. The three d bits represent the symbolic interval constant, distance_set_D, as a union of the three sets, distance_set_A (11010001), distance_set_B (11010010), and distance_set_C (11010100). The rea- son why we use the don’t care bit for a union of multiple symbolic interval constants is discussed later in this paper. The variable location keeps its original value, downtown. In the CT the values described in the condition parts of the translated rule base are written to cells of the CT. For example, in the case of the condi- tion element 1distance 5 distance_set_C 2 of rule R1, dist_set_C is assigned to the cell corresponding to the variable distance of the first row of the CT. For a variable that is not used as a condition element of a rule, d is assigned to the corresponding cell. Next, we discuss the OCAPP-ES units. Fig. 6. FV and CT with data for the going-to-the-theater example. It is assumed that the input value of the fact distance from the user is 10, time is 20, location is downtown, and the rest are unknown. The host computer translates the interval evaluating variables into corresponding symbolic interval constants. For example, the input number 10 of the fact distance is translated into a symbolic interval constant, distance_set_D (11010ddd), d:don’t care. Distance_set_D represents the union of the three symbolic interval constants distance_set_A (11010001), distance_set_B (11010010), and distance_set_C (11010100), as shown in Table 1. Likewise, the input number 20 of the fact time is translated into a symbolic interval constant, time_set_B (11100010). If a fact does not belong to the interval evaluating variables, the binary number for the symbol is written in the FV with the assignment of Fig. 4. We construct the CT by writing the value specified at the right-hand side of each condition element of the translated rule base at the cell specified by the variable of the left-hand side of each condition element. The d bits in the CT represent the variables that are not used in the rule. 10 August 1995 @ Vol. 34, No. 23 @ APPLIED OPTICS 5057 C. Description of the Optical Match Unit The optical match unit of the OCAPP-ES is designed around the original OCAPP design.16 The OCAPP is an optical parallel data-base@knowledge-base proces- sor based on an optical content-addressable memory. The system implements symbolic computing tasks such as searching, sorting, and information retrieval in a highly parallel fashion. The OCAPP is com- posed of a selection unit, a match@compare unit, a response unit, an output unit, and a control unit. The match@compare unit of an OCAPP is used as the optical match unit of the OCAPP-ES. A detailed explanation and implementation of each OCAPP unit and the algorithms implemented on the OCAPP are presented in Refs. 16 and 23. The optical match unit compares an input FV of given facts with a CT of the condition parts of given rules to generate a selected-rule list. As shown in Fig. 7, the comparison is performed by a vector– matrix multiplier performing a 2-D XOR logic function followed by a masking operation. Figure 8 explains how the selected-rule vector 1SRV2 is obtained for a given FV and CT. First, each bit of FVi is vertically expanded to cover a column of the CT. The expanded 2-D FV is then XOR’ed with the CT to produce a 2-D intermediate XOR result plane. On the 2-D intermediate XOR result plane, matches between the FV and the CT form dark output pixels, whereas unmatched pixels create a bright output. This intermediate XOR result plane is then imaged onto the mask. The latter blocks the contributions from the unused condition elements of a rule as well as the bits specified as d bits in the mask. Recall from the example that the d bits in the distance_set_D (11010ddd) are used to represent a union of the three symbolic interval constants distance_set_A(11010001), distance_set_B (11010010), and distance_set_C (11010100). Now, by blocking of the three d bits, the output d bits will become dark 1matched2. Thus if the distance_set_D is compared with any of the three symbolic interval constants, the three d bits will become dark 1matched2 for any input. The XOR result after the mask is then logically OR’ed to produce a selected-rule vector 1SRV2. In the 2-D XOR result image, if there is at least one unmatched condition element in a rule 1represented by a row2, there should be some light in the row of that rule. On the other hand, if the condition elements consist Fig. 7. Logical block diagram of the OCAPP-ES match unit: The FV is expanded vertically and XOR’ed with the CT. The 2-D XOR result is then masked by the DM to filter out unnecessary data. The XOR result after the DM is then OR’ed rowwise to produce a selected-rule column vector. 5058 APPLIED OPTICS @ Vol. 34, No. 23 @ 10 August 1995 of only either matched or don’t care pixels, then the absence of light indicates the row is matched. Therefore in the SRV, bright pixels indicate un- matched rules and dark pixels represent matched rules. The match operation is performed in parallel with an execution time, independent of the number of matches to be performed. D. Description of the Optical Select Unit Once the SRV is available, the optical select unit resolves conflicts among the selected rules and pro- duces a triggered-rule vector 1TRV2, which represents all the rules that can be fired in parallel. In the SRV and the TRV each pixel represents a rule. Whereas the SRV represents candidate rules that can be fired, the TRV represents a list of rules to be fired. Instead of traditional conflict-resolution strategies that en- able only one rule among the selected rules to be fired at a time, the optical select unit of the OCAPP-ES utilizes a new parallel conflict-resolution scheme to maximize performance by firing as many rules as possible. The new parallel conflict-resolution scheme is based on a dependency analysis among the rules so that rules can be fired simultaneously without any undesir- able side effects.24 For optical implementation of this parallel conflict-resolution scheme, an algorithm creates a conflict-resolution control matrix 1CRCM2 for any given rule base. The CRCM performs parallel rule selection by taking a 1-D SRV as an input, controlling specific SRV rules according to the depen- dencies, and generating a 1-D output TRV. The test necessary for detecting dependency between rules is exhaustive, as each rule of the rule base must be tested against all the remaining rules. However, because these analyses can be done at compile time for a given rule base, the dependency testing should not incur any run-time overhead. For a detailed explanation of the proposed new algorithm and for an example of constructing a CRCM, please refer to Ref. 30. Figure 9 shows the CRCM and an execution ex- ample of the optical select unit for the eight-rule going-to-the-theater example knowledge base. The figure illustrates how the 1 3 8 TRV 11 0 0 0 0 0 0 02 is generated with the 8 3 1 11 0 1 0 0 0 0 02T SRV from the optical match unit and the 8 3 8 CRCM. First, the input selected-rule column vector is ex- panded horizontally so that each column of the CRCM can receive the SRV. Each CRCM pixel is configured such that the pixel will perform an XNOR 1equivalence2 operation between incoming input data and the control data specified in the CRCM pixel. If a control datum is set to don’t care, the pixel should produce a 1 regardless of input data. Then the intermediate image, as shown in Fig. 9, is AND’ed columnwise to produce a row vector representing the desired TRV, which is 11 0 0 0 0 0 0 02 in this example. Again, the conflict resolution 1rule selection2 is per- Fig. 8. Match operation result for the going-to-the-theater example. The input FV is expanded to cover the CT. The XOR logic function is then performed with the expanded FV and CT to produce a 2-D XOR result image. The XOR result is imaged onto the dynamic mask 1DM2, which blocks the d bits, b2 2 b0, and the unused cells 1condition elements2 in the rule. Finally, the masked XOR result image is OR’ed rowwise to produce a selected-rule vector 1SRV2. formed in parallel and is independent of the number of rules to be fired. E. Description of the Act Unit When the triggered-rule vector 1TRV2 becomes avail- able, the front-end computer converts the optical TRV into an electronic form by means of the detector array. Then the electronic act unit executes the action part of the triggered rules and updates the changes caused by the rule-firing operation for the front-end computer as well as for the optical match unit. The act unit is implemented with electronics because of the following reasons: whereas the match and select operation requires computationally inten- sive operations 1e.g., comparing each rule of the rule base to all other rules2, the act operation requires simple assertion operations for only the triggered rules of the TRV. Also, the electronic act unit will ease the implementation of extra services such as explanations about the decisions being made. In addition, an optical act unit will add more hardware complexity while providing only simple operations that can be easily performed with electronics. In our example we have only one triggered rule, R1, which asserts the value of the action variable means to drive. Next, we discuss the scalability problem in which the given knowledge-base 1rules and facts2 size ex- ceeds the capacity of a given optical implementation. We solve this scalability problem in the RBS by taking advantage of the modular characteristic of the RBS.18 This characteristic stems from the fact that the structure of human knowledge can be modeled in a modular and hierarchical structure. Thus if the size of the given knowledge base is greater than the size of the available hardware, the problem should be parti- tioned into a set of subproblems whose size fits the available hardware. 10 August 1995 @ Vol. 34, No. 23 @ APPLIED OPTICS 5059 Fig. 9. 8 3 8 CRCM array for the eight-rule going-to-the-theater example. 4. Implementation of the Optical Content-Addressable Parallel Processor for Expert Systems The OCAPP-ES consists of an optical match unit, an optical select unit, an electronic act unit, and 2-D optical sources and detectors for input–output inter- facing with the front-end computer. A 1-D optical fact vector and a 2-D optical rule plane are created either by use of a 2-D laser diode array or modulation of a single laser with a 2-D spatial light modulator 1SLM2. The SLM can be a ferroelectric liquid-crystal SLM,25 an electrically addressed microchannel SLM,26 or a silicon lead-lanthanum-zirconate-titanate SLM.27 Because an OCAPP-ES must also return results to the electronic host, an optical detector array is needed to convert optical signals into electronic ones. An OCAPP-ES also needs a 2-D optical logic gate array to perform the XOR logic function. The XOR logic func- tions are performed by a pair of liquid-crystal SLM’s and by control of the polarization. Optical implemen- tation of the match and select units constituting the OCAPP-ES and electrical implementation of the act unit are discussed in the following. A. Implementation of the Optical Match Unit The optical match unit consists of three SLM’s 1SLM1, SLM2, and SLM32, three cylindrical lenses 1CL1, CL2, and CL32, and two polarizers 1P1 and P22, as shown in Fig. 10. The three SLM’s are pixellated, electrically addressed, ferroelectric liquid-crystal Fig. 10. Implementation of the optical match unit: The image of SLM1 is vertically expanded and XOR’ed with the data of SLM2 by means of cylindrical lenses CL1 and CL2. The XOR result image is then filtered with polarizer P1. The DM disables the transmis- sion of unnecessary pixels. The intermediate image after the DM is then focused into a vertical line 1SRV2. Each pixel of the SRV represents a logical OR of all the pixels in a row of the image after the DM. 5060 APPLIED OPTICS @ Vol. 34, No. 23 @ 10 August 1995 SLM’s configured to rotate the incident light by 90° in bit positions containing a logical value of 1 and 0° for those containing a logical value of 0. SLM1, SLM2, and P1 of Fig. 10 are configured as a vector–matrix multiplier to perform 2-D XOR logic functions, as shown in Table 2. A uniform beam of vertically polarized light illuminates SLM1. The logical val- ues of SLM1 are encoded such that vertically polar- ized light emanating from the device represents a 0 and horizontally polarized light represents a 1. The combinations of possible polarization rotations experi- enced by a beam passing through SLM1 and SLM2 are functionally equivalent to a Boolean XOR operation. Because CL1 and CL2 image a bit position of SLM1 onto a bit slice 1column2 of SLM2, the resulting 2-D data plane after SLM2 expresses the result of a bitwise matrix of XOR gates. Polarizer P1 then dark- ens any horizontally polarized light so that 1’s are represented by the absence of light and the 0’s continue to be represented by vertically polarized light. Table 2. Truth Table for the Optical XOR Logic Function Using SLM1 and SLM2 and Polarizer P1of Fig. 10a SLM1 Input SLM2 Input XOR Result After P1: Pass 1F2 0 1F2 0 1No rotation2 0 1F2 Bright 1F2: unmatched 0 1F2 1 1Rotate 90°2 1 1&2 Dark: matched 1 1&2 0 1No rotation2 1 1&2 Dark: matched 1 1&2 1 1Rotate 90°2 0 1F2 Bright 1F2: unmatched aA logical value of 0 is assigned to a vertically polarized beam 1F2, and a logical value of 1 is assigned to a horizontally polarized beam 1&2. SLM’s in the system are assumed to be electrically address- able liquid-crystal SLM’s. In the SLM’s the pixels with a logical value of 1 rotate the incoming light polarization by 90°, and the pixels with a logical value of 0 pass light without rotation. First, SLM1 is illuminated with a vertically polarized beam so that the pixels with a logical value of 0 have vertical polarization, and the pixels with a logical value of 1 have horizontal polarization. SLM2 then performs the XOR logic function by rotating the pixels with a logical value of 1 102 by 90° 10°2. Then, the pixels with horizontally polarized light 1logical 1: matched2 of the XOR result are blocked by use of polarizer P1, which passes vertically polarized light 1logical 0: unmatched2 only. The vector–matrix multiplier is then followed by SLM3, which behaves as a dynamic mask 1DM2, shown in Fig. 10. The purpose of the DM is to block the transmission of the don’t care pixels while passing the other pixels, as shown in Table 3. After the DM, the matched pixels and the don’t care pixels are dark, and the unmatched pixels will remain bright. CL3 focuses the image after the DM into a vertical line, which represents the SRV. The SRV is routed to the optical select unit. B. Implementation of the Optical Select Unit Figure 11 shows the implementation of the optical select unit. The optical select unit consists of three SLM’s 1SLM1, SLM2, and SLM32, three cylindrical lenses 1CL1, CL2, and CL32, two beam splitters 1BS1 and BS22, and two polarizers 1P1 and P22. In the optical select unit the three SLM’s of Fig. 11 are configured to rotate the polarization of the incident light by 90° in bit positions containing a logical value of 0 and 0° for those containing a logical value of 1. SLM1 1an optically addressable SLM2 and BS1 are used to convert the intensity-encoded SRVi from the optical match unit into the polarization-encoded SRVp, as shown in Table 4. The vertically polarized colli- mated beam going into BS1 is sent into the reflective side of SLM1 to modulate the incoming SRVi, which simultaneously writes its value to the photoconduc- tive side of SLM1. If the input pixel of SRVi is bright 1logical 02, SLM1 will rotate the polarization of the vertically polarized light incident upon it at the photoreflective side by 90° to produce a horizontal polarization. On the other hand, if the input is dark 1logical 12, the output pixel will have a vertical polar- ization. Thus SRVp is encoded such that a logical value of 0 1an unselected rule2 is represented by horizontal polarization, whereas a logical value of 1 1a selected rule2 is represented by vertical polarization. Next, SLM2 and SLM3 of Fig. 11 work together to realize the CRCM. The logical values of SLM2 are encoded such that vertically polarized light emanat- ing from the device represents a 1 and horizontally polarized light represents a 0, as shown in Table 5. Table 3. Truth Table for the DM and P2 of Fig. 10a After P1: Pass 1F2 DM Input After P2: Pass 1F2 F 0 1No rotation2 Bright 1F2 F 1 1Rotate 90°2 Dark dark 0 1No rotation2 Dark dark 1 1Rotate 90°2 Dark aThe DM is assumed to be an electrically addressable liquid- crystal spatial light modulator 1SLM2, and F represents a verti- cally polarized beam. In the table a logical value of 1 is assigned to the don’t care pixels of the DM, and a logical value of 0 is assigned to the other pixels. In the DM the don’t care pixels rotate the incoming light polarization by 90°, and the other pixels remain unchanged. With vertically polarized input light a vertical polar- izer, P2, placed behind the DM selectively blocks the don’t care pixels. After this unit, the matched pixels and the don’t care pixels will be dark and the unmatched pixels will be bright. The combinations of possible polarization rotations experienced by a beam passing through SLM2 are functionally equivalent to a Boolean XNOR operation. Because CL1 and CL2 image a bit position of SRVp onto a bit slice 1column2 of SLM2, the resulting 2-D data plane after SLM2 expresses the result of a bitwise matrix of XNOR gates. Polarizer P1 then darkens any horizontally polarized light so that, after P1, 0’s are represented by the absence of light and the 1’s continue to be represented by vertically polarized light. SLM3 provides light for the don’t care pixels of the CRCM. With a vertically polarized collimated input beam, SLM3 controls the polarization angle of each pixel as follows: If the pixel corresponds to the don’t care pixel, it will have 90° polarization rotation to have a horizontal polarization. On the other hand, if the pixel does not correspond to the don’t care pixel, it will preserve the vertical polarization. Then, by use of P2, which passes only horizontally polarized light, only the don’t care pixels will become bright. Finally, the two intermediate data planes from SLM2 and SLM3 are merged by the use of BS2 and focused into a horizontal line to form a TRV. C. Implementation of the Act Unit The act unit is implemented with software because of the reasons given in Subsection 3.E. One implemen- tation method for the act unit is keeping a simple table that records the rule number, the name of the variable to be triggered, and its value. To perform Fig. 11. Implementation of the optical select unit: The polariza- tion-encoded selected-rule vector 1SRV2 is expanded and XOR’ed with the CRCM in SLM1. Then the two data planes from SLM2 and SLM3 are spatially added by the use of beam splitter BS2. The added image is then collimated with cylindrical lens CL3 to form a TRV. Table 4. Conversion of the Intensity-Encoded SRVi to Polarization-Encoded SRVp at SLM1 of Fig. 11a Input SRVi 1before SLM12 SRVp 1after SLM12 Logical 1 No light F Logical 0 Light & aAssuming a vertically polarized collimated beam is incident upon the photoreflective side of SLM1, the reflected light will be controlled by the intensity of the input pixel at the photoconductive side 1SRVi 2 of SLM1. If the input pixel of SRVi is bright 1logical 02, the SLM1 will rotate the polarization by 90° to produce a horizontal polarization. If the input is dark 1logical 12, the pixel polarization at the photoreflective side of SLM1 remains vertical. 10 August 1995 @ Vol. 34, No. 23 @ APPLIED OPTICS 5061 the act operation, the detector unit of the electrical subunit of the OCAPP-ES converts the optical TRV into an electrical TRV, which is a set of the triggered- rule numbers. Then executing the action parts of the rules will be a matter of accessing the table with the rule numbers as indices and of updating the values of the corresponding variables. 5. Theoretical Estimation of Execution Time and Number of Processed Rules per Second In this section the execution time and the maximum number of rules that can be processed per second in the OCAPP-ES are estimated. In what follows we use the following terms and assumptions for estimat- ing the execution speed: c The dimension of available 2-D SLM’s is n 3 n. c The response time of the SLM’s and optical logic gate arrays is Tr. c The setup time for a SLM of size n 3 n is Ts2. c The setup time for a SLM of size n is Ts1. c The detector response time is Td. c The propagation delay of optical passive devices such as lenses and mirrors is negligible compared with that of the SLM’s. A. Execution Time In general the execution time of a system can be regarded as the sum of the system initialization time Ti and the system processing time Tp. Ti includes the SLM setup times in the optical match and select units. Although these two units have six SLM’s in total, these SLM’s can be accessed simultaneously: hence Ti becomes Ts2. In an OCAPP-ES, Tp can be further divided into the optical system processing time Top, the conversion time from optical signal to electrical signal Toe, the act operation processing time Tap, and the conversion time from electrical signal to optical signal Teo. Top is equal to 5Tr because there are five SLM’s in the major optical signal path in the OCAPP-ES. Toe is the detector readout time Td; Teo is the n 3 n SLM modulation time Ts2. Thus Tp becomes 5Tr 1 Td 1 Ts2 1 Tap. Therefore the overall execution time of the OCAPP-ES, Te, becomes Te 5 Ti 1 Tp 5 Ts2 1 x15Tr 1 Td 1 Ts2 1 Tap2, 112 where x represents the number of iterations neces- sary to solve a given query. Table 5. Summary of XNOR Logic Function at SLM2 of the Optical Select Unit of Fig. 11a SLM1 Output SLM2 Setup XNOR Result After P1: Pass F 0 1&2 0 1Rotate 90°2 1 1F2 Bright: F 0 1&2 1 1No rotation2 0 1&2 Dark 1 1F2 0 1Rotate 90°2 0 1&2 Dark 1 1F2 1 1No rotation2 1 1F2 Bright: F aThe logic performed is similar to that of Table 2 except for the assignment of polarization for the logical values and for the polarization rotation angle control at the second SLM. 5062 APPLIED OPTICS @ Vol. 34, No. 23 @ 10 August 1995 B. Number of Processed Rules per Second Next, the maximum number of rules that can be processed per second is estimated. For the n 3 n SLM the number of rules that can be represented in the OCAPP-ES becomes n. Because the processing time of the OCAPP-ES, Tp, is equal to 5Tr 1 Td 1 Ts2 1 Tap, the maximum number of rules R that can be processed per second is R 5 n 5Tr 1 Td 1 Ts2 1 Tap . 122 As an example, assuming that the technology per- mits, n 5 256, Tr 5 1026 s, and Ts2 5 Tap 5 Td 5 1023; then an OCAPP-ES can execute roughly 8.52 3 104 rules@s. As a comparison, the NON-VON is a multi- processor system that is composed of 32 powerful large processing elements and 16,000 small process- ing elements and is estimated to process 850 rules@s.9 Another multiprocessor called RUBIC 1rule-based in- ference computer2 is estimated to process 4 3 103 rules@s.28 With the processing capability of 8.52 3 104 rules@s, the OCAPP-ES is expected to achieve a system throughput an order of magnitude better than any electronic RBS can achieve. 6. Conclusion Although many optical systems have been proposed to improve the performance of electronic rule-based expert systems, these systems still suffer from their limited exploitation of the available parallelism in RBS’s or from being too application specific. To overcome these limitations and to take advantage of optics parallelism, we have proposed an optical con- tent-addressable parallel processor for expert sys- tems 1OCAPP-ES2 in this paper. Distinctive features of the OCAPP-ES include the following: 112 2-D representation of data 1knowledge2 and control infor- mation; 122 capability of processing general-domain knowledge expressed in terms of variables, numbers, symbols, and comparison operators such as greater than and less than; 132 the parallel optical match unit 1designed around the match@compare unit of the OCAPP2, which performs the 2-D optical pattern matching and comparison operations; 142 a novel con- flict-resolution control-matrix 1CRCM2 algorithm to resolve conflicts in a single step within the optical select unit. We have detailed the implementation of various units of the system. We have also shown that the estimated maximum number of rules that can be processed per second is 8.52 3 104. It is expected that further developments in optical devices technology 1logic, memory, smart SLM’s2 will further increase the performance of the OCAPP-ES. This research was supported by National Science Foundation grant MIP-9113688. References 1. P. Harmon and D. King, Expert System: Artificial Intelligence in Business 1Wiley, New York, 19852, Sec. 1, pp. 13–76. 2. H. Schorr and A. Rappaport, Innovative Applications of Artifi- cial Intelligence 1AIAA Press, Menlo Park, Calif., 19892, p. 137. 3. K. Pedersen, Expert Systems Programming 1Addison-Wesley, Reading, Mass., 19892, Chap. 14, pp. 124–139. 4. R. I. Levine, AI and Expert Systems: a Comprehensive Guide, C Language 1McGraw-Hill, New York, 19902, Chap. 9. 5. D. P. Miranker, ‘‘TREAT: a new and efficient algorithm for AI production systems,’’ Ph.D. dissertation 1Columbia University, New York, N.Y., 19872. 6. A. Gupta, C. Forge, D. Kalp, A. Newell, and M. S. Tambe, ‘‘Parallel OPS5 on Encore Multimax,’’ in Proceedings of the 1988 International Conference on Parallel Processing, F. A. Briggs, ed. 1Pennsylvania State Univ. Press, University Park, Pa., 19882, 271–280. 7. R. Gamble, ‘‘Transforming rule-based programs: from the sequential to the parallel,’’ in International Conference on Industrial and Engineering Applications of Artificial Intelli- gence and Expert Systems 1Association of Computing Machin- ery, Charleston, S.C., 19902, pp. 854–863. 8. S. J. Stolfo and D. P. Miranker, ‘‘DADO: a parallel processor for expert systems,’’ in Proceedings of the International Confer- ence on Parallel Processing, R. M. Keller, ed. 1Institute of Electrical and Electronics Engineers, New York, 19842, pp. 74–82. 9. B. W. Wah and G. J. Li, ‘‘Design issues of multiprocessors for artificial intelligence,’’ in Parallel Processing for Supercomput- ers, K. Hwang and D. Degroot, eds. 1McGraw-Hill, New York, 19892, Chap. 4, pp. 107–159. 10. J. Gaudiot and A. Sohn, ‘‘Data-driven parallel production systems,’’ IEEE Trans. Software Eng. 16, 281–293 119902. 11. S. Kuo and D. Moldovan, ‘‘The state of the art in parallel production systems,’’ J. Parallel Distrib. Comput. 15, 1–26 119922. 12. A. Louri and J. Na, ‘‘Parallel electro-optical rule-based system for fast execution of expert systems,’’Appl. Opt. 32, 1863–1785 119932. 13. Y. Li and G. Eichmann, ‘‘Conditional symbolic modified signed- digit arithmetic using optical content-addressable memory logic elements,’’Appl. Opt. 26, 2328–2333 119872. 14. A. B. VanderLugt, ‘‘Signal detection by complex spatial filter- ing,’’ IEEE Trans. Inf. Theory IT-10, 139–145 119642. 15. J. Y. Jau, F. Kiamilev, Y. Fainman, and S. H. Lee, ‘‘Optical expert system based on matrix-algebra formulation,’’ Appl. Opt. 27, 5170–5175 119882. 16. A. Louri, ‘‘Optical content-addressable parallel processor: architecture, algorithms, and design concepts,’’ Appl. Opt. 31, 3241–3258 119922. 17. P. H. Winston, Artificial Intelligence 1Addison-Wesley, Read- ing, Mass., 19842, Chap. 6, pp. 159–204. 18. L. Brownston, R. Farrell, E. Kant, and N. Martin, Program- ming Expert Systems in OPS5 1Addison-Wesley, Reading, Mass., 19852, Chap. 1, pp. 4–18. 19. A. D. McAulay, Optical Computer Architectures 1Wiley, New York, 19912, Chap. 8, pp. 193–222. 20. A. Louri and J. Hatch, ‘‘Optical implementation of a single- iteration thresholding algorithm with application to parallel data-base@knowledge-base processing,’’ Opt. Lett. 18, 992–995 119932. 21. M. Togai and H. Watanabe, ‘‘An inference engine for real time approximate reasoning: toward an expert on a chip,’’ IEEE Expert 1132, 55–62 119862. 22. C. C. Lee, ‘‘Fuzzy logic in control systems: fuzzy logic control- ler. Part 1,’’ IEEE Trans. Syst. Man Cybern. 20, 404–415 119902. 23. A. Louri and J. Hatch, ‘‘An optical associative parallel proces- sor for high-speed database processing: theoretical concepts and experimental results,’’ IEEE Computer 271112, 65–72 119942. 24. S. Kuo and D. Moldovan, ‘‘Implementation of multiple rule firing production systems on hypercube,’’ J. Parallel Distrib. Comput. 13, 383–394 119912. 25. J. Liu, K. M. Johnson, and M. G. Robinson, ‘‘Room-tempera- ture 10-MHz electro-optic modulation in ferroelectric liquid crystals,’’Appl. Phys. Lett. 62, 934–936 119932. 26. A. D. McAulay, X. Xu, and J. Wang, ‘‘Optical implementation of a multi-layer neural network by SLR and MSLM,’’ in IEEE International Conference on Systems Engineering, 1Institute of Electrical and Electronics Engineers, New York, 19912, pp. 197–200. 27. A. Ersen, S. Dasgupta, T. H. Lin, S. Esner, and S. H. Lee, ‘‘Dual-beam recrystallization of silicon on PLZT,’’ in Optical Computing, Vol. 9 of 1989 OSATechnical Digest Series 1Optical Society of America, Washington, D.C., 19892, pp. 54–57. 28. D. Moldovan, ‘‘RUBIC: a multiprocessor for rule-based sys- tems,’’ IEEE Trans. Syst. Man Cybern. 19, 699–706 119892. 29. J. A. Na, ‘‘Design and simulation of digital optical computing systems for artificial intelligence,’’ Ph.D. dissertation 1Univer- sity ofArizona, Tucson,Ariz., 19942. 10 August 1995 @ Vol. 34, No. 23 @ APPLIED OPTICS 5063 
work_puqhglqbwrhgvnt2rdo43d4zpm	----	Clustering-based location in wireless networks Expert Systems with Applications 37 (2010) 6165–6175 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Clustering-based location in wireless networks Luis Mengual, Oscar Marbán *, Santiago Eibe Facultad de Informática, Universidad Politécnica de Madrid (UPM), Spain a r t i c l e i n f o Keywords: Indoor location Wireless lan 802.11 Signal strength Fingerprint techniques Clustering 0957-4174/$ - see front matter � 2010 Elsevier Ltd. A doi:10.1016/j.eswa.2010.02.111 * Corresponding author. Address: Campus de Monte de Informática, Universidad Politécnica de Madrid Monte, Madrid, Spain. E-mail addresses: lmengual@fi.upm.es (L. Meng Marbán), seibe@fi.upm.es (S. Eibe). a b s t r a c t In this paper, we propose a three-phase methodology (measurement, calibration and estimation) for locating mobile stations (MS) in an indoor environment using wireless technology. Our solution is a fin- gerprint-based positioning system that overcomes the problem of the relative effect of doors and walls on signal strength and is independent of network device manufacturers. In the measurement phase, our sys- tem collects received signal strength indicator (RSSI) measurements from multiple access points. In the calibration phase, our system utilizes these measurements in a normalization process to create a radio map, a database of RSS patterns. Unlike traditional radio map-based methods, our methodology normal- izes RSS measurements collected at different locations (on a floor) and uses artificial neural network models (ANNs) to group them into clusters. In the third phase, we use data mining techniques (cluster- ing) to optimize location results. Experimental results demonstrate the accuracy of the proposed method. From these results it is clear that the system is highly likely to be able to locate a MS in a room or nearby room. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction The use of wireless technology has expanded over recent years. Universities, private companies, as well as stations and airports all provide wireless communication facilities. This type of networks can be implemented using what are known as access point devices that provide total connectivity over a surface and connection to Internet. Several such devices are sometimes necessary to cover an office, a university or an airport. One of the services to be implemented on these networks is user location on the premises. Through location and the visualiza- tion of a plan of the coverage area, users can navigate the premises as they would do using GPS navigation in an outdoor environment. Additionally, a network manager would be able to pinpoint users in such environments and interact with them through their posi- tioning. Indoor positioning technology can also complement mo- bile–telephone-based GPS user positioning to determine the users exact location indoors. In another context, location tracking is an essential feature for enterprises building business-critical wireless networks. If infor- mation technology (IT) staff can identify and track the location of wireless clients and highly mobile assets, they can improve the accuracy of WLAN planning and deployment, optimize ongoing ll rights reserved. gancedo s/n. DLSIIS, Facultad (UPM), 28660 Boadilla del ual), omarban@fi.upm.es (O. network performance, enhance wireless security, and improve the usefulness and value of important business applications. Loca- tion tracking provides enhanced visibility and control of air space, helping IT staff to deploy wireless networks that are as easy to manage and as effective to deploy as traditional wired networks. To address lack-of-visibility problems, organizations need a cost-effective, easy-to-deploy solution for tracking and managing thousands of Wi-Fi devices and tags across a variety of business environments. In heterogeneous environments, e.g. inside a building, though, the received power is a very complex function of distance, wall geometry, building infrastructures and obstacles. Even if you have a detailed model of the building, it takes a lengthy simulation to solve the direct problem of deriving signal strength given the loca- tion. This is what has motivated us to consider flexible models based on functions networks (neural networks) to implement a system to locate MSs. The received signal strength (RSS), a measure of the power re- ceived by the client from an access point (AP), is the key parameter for establishing the position of a MS in an indoor environment. Tra- ditionally, WLANs have used different measurement techniques to derive the position of MSs (Kaemarungsi, 2005; Raniwala & Chiueh, 2002). There are three major categories: closest access point-based techniques, location pattern or fingerprint and distance or angle measurement (see Fig. 1). The closest AP method finds devices within the total coverage area of a single AP. It is the simplest but least accurate way to lo- cate a device or user. With the closest AP method, the location tracking system identifies only devices within the total coverage http://dx.doi.org/10.1016/j.eswa.2010.02.111 mailto:lmengual@fi.upm.es mailto:omarban@fi.upm.es mailto:seibe@fi.upm.es http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. WLAN positioning techniques. 6166 L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 area of a single AP. This area can be quite large and include multi- ple rooms. This system considers an AP to which a terminal con- nects as the user location. The main distance measurement-based approach is known as TDOA (Time Difference of Arrival). This technique relies on the timing precision between the signal transmitter and receiver. It uses the propagation delay to calculate the distance between transmitter and receiver (Golden & Bateman, 2007). Therefore, a precise synchronization is very important in such systems. By combining at least three distances from three reference positions, triangulation can be used to estimate the MS’s location. Such techniques will require a high accuracy clock in the communica- tion system. Angle of arrival (AOA) is a technique based on angle measure- ment. This methodology locates the MS by determining the signal angle of incidence. The location can be estimated using simple geo- metric relationships, like triangulation (Sanchez, Afonso, Macias, & Suarez, 2006). In indoor environments, however, the distance between trans- mitter and receiver is usually shorter than the time resolution that the system can measure. Additionally, the MS is surrounded by scattered objects, which results in multiple angles of signal recep- tion. Therefore, the AOA and TDOA approaches are difficult to implement in indoor environments. Fingerprinting techniques generally only require measurement of received signal strength or other non-geometric features at several locations to form a database of location fingerprints (Kaemarungsi, 2006; Widyawan, Klepal, & Pesch, 2007). To estimate the mobile location, the system needs to first measure the received signal strength at particular locations and then search for the pattern or fingerprints with the closest match in the database. Generally, the deployment of fingerprinting-based positioning systems can be divided into two phases. First, the location finger- prints are collected in the off-line or calibration phase by perform- ing a site survey of the received signal strength (RSS) from multiple APs. Enough RSS measurements are taken to set up a database or a table of RSS patterns for predetermined points of the coverage area. The database of RSS patterns with their respective locations is called a radio map. Second, a MS will report a sample measured vector of RSSs from different APs to a central server (or a group of APs will collect the RSS measurements from a MS and send it to the server) in the on- line or estimation phase. The server uses a positioning algorithm to estimate the location of the MS based on the radio map and reports the estimate back to the MS (or the application requesting the po- sition information). At this point we have several possibilities. The most common algorithm used to estimate the location computes the Euclidean distance between the sample measured RSS vector and each finger- print in the database. For indoor positioning systems, other ad- vanced algorithms and techniques, such as neural networks (Battiti, Le Nhat, & Villani, 2002; Debono & Buhagiar, 2004; Yiming, Biaz, Pandey, & Agrawal, 2006; Youssef & Agrawala, 2005), probabilistic methods (Chai & Yang, 2007; Kontkanen et al., 2004; Pan, Kwok, Yang, & Chen, 2006; Robinson & Psaromiligkos, 2005; Seigo & Kawaguchi, 2005) or fuzzy logic (Astrain, Villadan- gos, Garitagoitia, González-Mendivil, & Cholvim, 2006), have been introduced to determine the relationship between RSS samples and the location fingerprint in the radio map. Other approaches have been presented. For example, some authors have proposed simulating the calibration phase with a ray-tracing model (Nezafat, Kaveh, Tsuji, & Fukagawa, 2004; Nuño & Páez-Borrallo, 2006; Zhong, Bin-Hong, Hao-Xing, Hsing- Yi, & Sarkar, 2001). However, this model requires a highly detailed model of the building. It could be tricky, if not unfeasible, to get such a detailed model. One of the main drawbacks of both of these models is that signal strength prediction is dependent not just on building layout, but also on the position of many other hard-to- model components, such as furniture, equipment and human beings. So far, fingerprinting techniques with a real calibration phase have attracted more attention because they are a simple and the most effective solution for indoor environments. Of the solutions that use RSSI to locate a mobile user, some scan the radioelectric spectrum of every square metre of the coverage area. Additionally, they only use one card brand. It is unfeasible to position another mobile user with a different Wi-Fi receiver using this solution. The solution proposed here is a fingerprint-based positioning system that gets significantly better results than other similar sys- tems because it normalizes measurements and optimizes clusters formed according to the physical features of the ground plan. In this way, the proposed solution is valid for any Wi-Fi receiver, as we account for the relative effect of walls and obstacles as a rela- tive power increase or decrease. Additionally, our technique does not require an exhaustive scan of the coverage area. Actually we present the results of sampling just one point per location. Our system can adapt to the geometry of each scenario and operate with any Wi-Fi device (irrespective of the manufacturer). Finally, we present the results of implementing our system in practice and the evaluation tests run in a School of Computing L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 6167 building (Department of Computer Languages and Systems and Software Engineering). The experimental results confirm the effec- tiveness of our methodology. 2. Location estimation system issues Unfortunately, commercial 802.11 hardware does not provide positioning functionality. Even so, positioning can be implemented using RSS measurements provided by an 802.11 wireless card. RSS is a measure of the power received by the client from an AP and provides information as to the clients location. In heterogeneous environments, e.g. inside a building, though, the received power is a very complex function of distance, geome- try, and materials. This is further complicated by the fact that the signal propagation is influenced by environmental factors like, for example, the number of people in the working area, the posi- tion of walls and other building infrastructures and the materials they are made of, as well as the multi-path effect (Ali & Nobles, 2007). Another problem is the RSSI gathered from wireless stations using their device driver. RSSI is an optional parameter that has a value of 0 through to RSSI_Max. This parameter is a measure of the physical sublayer of the energy observed at the antenna used to receive the current frame. RSSI shall be measured from the beginning of the start frame delimiter (SFD) to the end of the frame, i.e. header error check (HEC). RSSI is intended to be used in a relative manner. Absolute RSSI reading accuracy is not specified. There is nothing in the 802.11 standard that stipulates a rela- tionship between the RSSI value and any particular energy level as could be measured in mW or dBm. Individual vendors have cho- sen to provide their own levels of accuracy, granularity, and actual power range (measured as mW or dBm) and their RSSI value range (from 0 to RSSI_Max). Some vendors do not provide an RSSI value and instead convert directly from dBm to percent. In this study we used our own RSS collecting software to test different wireless cards. Each card collected 300 samples over a period of 5 minutes (1 sample/s) on the 3rd Floor of Building 4 at the School of Computing (Universidad Politcnica Madrid). Fig. 2 shows the measurable ranges for the different wireless cards in a room. We found that there was a difference of 20 dB between dif- ferent cards (3Com/Ovislink). There is even a card (USROBOTICS) that implements RSSI as a percentage and not in dBm units. From these results we conclude that the mapping between the actual RF energy and the RSSI range in a wireless device varies Fig. 2. Measurable RSS ra from one vendor to another. Since the range and the measurement of RSS depend on the WLAN card, most fingerprint-based position- ing systems have used the same wireless card vendor to collect the location fingerprints and determine the location. The results of our experimental tests match the study run in Kaemarungsi (2006), showing the spread of values for five card manufacturers. In Kaemarungsi (2006), Kaemarungsi even goes as far as to recommend the use of the same card manufacturer for radio map construction-based positioning systems. However, our methodology is capable of working with different wireless cards because we normalize all RSS measurements in the calibra- tion phase. 3. Location estimation methodology The methodology that we propose can locate any mobile user in any indoor environment equipped with several symmetrically dis- tributed APs. In the following we describe the proposed system. 3.1. Outline of the system Our system is a three-phase fingerprint-based positioning solu- tion, as shown in Fig. 3. 1. In phase 1, RSS measurements are collected from multiple APs positioned in different rooms on one floor. These will be the location fingerprints used to create the RSS database, called the radio map, of that floor. 2. This information is processed in a calibration phase. The cali- bration phase is divided into several stages. First, we normalize the collected measurements to identify the relative effect of walls and obstacles. Second, we use neural networks and the computed normalized values to group the measurements in clusters. Finally, we use the physical topology to optimize the clusters formed. 3. In phase 3, called the estimation phase, our system can locate any MS from its current RSS measurements in one or two rooms on the floor. This is possible because of the cluster optimization logic and the applied normalization methodology. In our system we chose to use Kohonen networks as a clustering tool. Koho- nen networks represent a type of self-organizing map (SOM), which is actually a special class of neural networks. SOMs are based on competitive learning, where the output nodes com- pete with each other to be the winning node (or neuron), the only node to be activated by a particular input observation. This m nge of WLAN cards. RadioMap Builder Wireless DeviceDriver Normalization MeasuresClustering RadioMap OptimizedClusters Estimator Optimization ClusteringLogic EstimatedLocation Measurement Phase Estimation PhaseCalibrationPhase Normalization Measures Signal Strength Measurements Physical Distribution Plant RadioMap Builder RadioMap Builder Wireless DeviceDriver Normalization Measures Normalization MeasuresClusteringClustering RadioMap Estimator Optimization ClusteringLogic Optimization ClusteringLogic EstimatedLocation Measurement Phase Estimation PhaseCalibrationPhase Normalization Measures Normalization Measures Signal Strength Measurements Physical Distribution Plant Physical Distribution Plant Fig. 3. Location estimation methodology structure. 6168 L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 clustering technique has the advantage of converting a complex high-dimensional input signal into a two-dimensional discrete map. In the next section we detail the location estimation system. 3.2. Location estimation algorithm Our location estimation system consists of the following phases (see Fig. 3). 3.2.1. Measurement phase As already mentioned, the RSSI measurements database is built in this phase. To do this, we systematically measure the signal val- ues relative to the available APs. The calculations, which we will describe later, call for at least three APs arranged to form a triangle. If there are more than three APs, they should be positioned sym- metrically to optimize coverage. As many measurements will be taken as necessary to assure that the results are significant. At the end of the process, a vector S is output for each measure- ment. This vector represents the signal power values received from k surrounding APs (in this case k ¼ 4): S ¼ðs1; s2; . . . ; skÞ ð1Þ In this paper we will demonstrate that, according to our meth- odology, the make of the card used to build the radio map is irrel- evant. Generally, as explained in Section 2, the values output would be different. However, later data processing makes them va- lid for estimating the position of any card. 3.2.2. Calibration phase This phase can be divided into the following stages illustrated in Fig. 4. Normalization. The normalization of measurements is a key as- pect in our methodology. Thanks to normalization, we can take into account the relative effect of walls and obstacles on the signal loss from the surrounding APs. Additionally, the normalization of measurements allows us to use any wireless device to estimate the position of a mobile terminal. To understand this key point, lets look at the measurements in Fig. 2. The table data are evidence of the spread of the received RSSI values. This rules out the use of a clustering technique. However, you can solve this problem by referencing all the power values to fixed coordinate sources on the map. The normalization of measurements involves first setting the normalization reference points. We will set both the reference points and APs existing in the region under analysis will be set. Actually, the coordinates of the measurement point nearest to each AP will be taken as the normalization reference. Consequently, we considered four reference point coordinates x1; x4; x6 and x7 , respectively, in this study. They correspond to the symmetric posi- tions in which the APs were located. Second, we calculate the mean value of the measures taken using the different cards in each of the rooms with the set reference points. This way, we get the following data. Let k be number of APs = number of reference points, and lets take n measurements as follows: mi ¼ðS1; S2; . . . ; SnÞ where i ¼ 1; . . . ; n ð2Þ measured at reference point: mRj ¼ðS R 1; S R 2; . . . ; S R kÞ ð3Þ e.g. at R ¼ P1: mP1j ¼ðS P1 1 ; S P1 2 ; . . . ; S P1 k Þ ð4Þ Mean of the measurements at the reference points: MR ¼ �mR ¼ Pn j¼1 S R 1j n ; Pn j¼1 S R 2j n ; . . . ; Pn j¼1S R kj n ! ð5Þ e.g. at R ¼ P1: MP1 ¼ �mP1 ¼ Pn j¼1S P1 1j n ; Pn j¼1 S P1 2j n ; . . . ; Pn j¼1 S P1 kj n ! ð6Þ Fig. 4. Detailed calibration phase. L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 6169 Then we proceed to expand the measures by making: 8mi ¼ðS1; S2; . . . ; SkÞ; 8R ¼ 1; . . . ; k eRi ¼ �m R � miÞ � � ¼ MR � mi � � ð7Þ This would give, for one reference point, the vector eRi of size 1 � k and, for k reference points, we would have the vector Ei of size 1 �ðk � kÞ Ei ¼ðe R1 i ; e R2 i ; . . . ; e Rk i Þ ð8Þ Example. If R ¼ P1; P4; P6; P7: eP1i ¼ðM P1 � miÞ; eP4i ¼ðM P4 � miÞ eP6i ¼ðM P6 � miÞ; eP7i ¼ðM P7 � miÞ ð9Þ the expansion vector of the measurements would be Ei ¼ðeP1i ; e P4 i ; e P6 i ; e P7 i Þ ð10Þ This expansion process is motivated by the analysis of the experimental results, as we will see in the evaluation section. Clustering and optimized clustering. Having built the radio map of normalized measures our system applies artificial intelligence techniques to estimate the location. We will actually use clustering techniques to form different SOMs from the values of the normal- ized measurements with respect to each reference point. This way, Fig. 5. Cluster vi we will get as many SOMs as reference points were set in the system. A classification method validation process can be included in SOM construction, although, as we will see in the examples, this is not always necessary. The process involves dividing the input data into two sets: a training set to form the SOMs and a test data set to test the SOMs and check they are capable of locating a device whose measurements have not been used to form the SOM. With this test set, we can evaluate the performance of the SOMs in terms of percentage of correct locations. The original data from the experiments run are partitioned at random as follows: 90% for the training set and 10% for the test set. The next step is to optimize the above clusters. To get a preli- minary evaluation of the cluster obtained from the positions with respect to their arrangement in the physical space, the cluster is plotted on a chart. Fig. 5 shows this chart. As the chart shows, all the clusters contain measures taken at different positions. As our ultimate aim is to define regions in which to locate measurements, we decompose the above chart into planes. Each plane contains the measurements of just one position, as shown in Fig. 6. The regions containing points from each posi- tion are easier to identify in Fig. 6. The result of the classification will now be associated with a region instead of a position, as had been the case up to this point. Obviously, the drawback of this solution is that it outputs an area rather than pinpointing an exact location. Therefore, the location estimation problem has need of further processing. sualization. Fig. 6. Decomposition by cluster positions. 6170 L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 3.2.3. Estimation step The MS location estimation phase gathers information about the power actually received from the nearby APs. With this infor- mation and the information about the optimized regions, we posi- tion the MS in one of these regions. There are three possible ways of doing this: 1. Exact location estimation: if the n SOMs output one position, this is the MS location. If the chart plots a point located within a region, we do not know the exact location of the MS. 2. Region estimation: the n SOMs output a region, and this region is possibly different for each SOM. This conflict is settled by a voting algorithm, where the position/region that is the most common across all the SOMs is the final result. 3. Optimized region estimation according to the physical distribu- tion of the space: as above, but only the regions that have points that are close together in the physical space are taken into account before the result is shown. For this purpose, rules are built into the system to reduce the number of possible regions for each result.The experimental results (see Section 4) show that the likelihood of locating a station is high. Therefore, it is possible to build optimized regions with a single wireless card brand and use this infrastructure to locate any other brand of wireless device. 4. Results of evaluation experiments We applied our location estimation methodology shown in Sec- tion 4 to implement an estimation location system and conducted evaluation experiments on the 3rd floor of building 4 at the School of Computing. Fig. 7 gives an overview of the experimental envi- ronment. The figure shows the floor of part of our university build- ing. The total surface area ranges from 300 to 500 m2. Faculty members work in separate rooms. The walls are very thin (plaster). This is an obstacle especially for locating MS in adjoining offices because of the reception of similar powers from the APs. Red circles show the location of each AP that we used in our tests. Green squares show the locations (12) at which we took measurements of the four AP signal strengths. The normalization points are P1; P4; P6 and P7. In the survey step, we observed signal strength for 4 min and took 240 measurements at each location. The hardware that we used in our experiments included seven different IEEE 802.11 wireless adapters; three wireless PC cards (3Com, Enterasys and US Robotics); three wireless USB 2.0 adapt- ers (SMC, Zoom and Ovislink) and one integrated internal Intel Wi-Fi. APs are manufactured by US Robotics and 3Com. We have divided the results into two parts for presentation, depending on the measurements used to create the clusters. The first part presents the results of using the measurements collected from all the wireless cards. The second part uses measurements from just one SMC wireless card. In this case, the data do not have to be divided into a training and a test set, as the SMC card data form the training set and the data of the other cards (not used to define the SOMs) form the test set. In both cases, we have created four different SOMs using the normalized measurements with re- spect to the normalization points P1; P4; P6 and P7 in Fig. 7. As the first step in the explanation of our method, we are going to discuss an example using the experimental values (absolute or pre-normalized values) obtained by different cards in the same place (any place) on the floor in question: mi;3Com ¼ð�92;�74;�67;�87Þ mi;Enterasys ¼ð�86;�70;�52;�83Þ mi;Ovislink ¼ð�75;�54;�49;�64Þ mi;SMC ¼ð�82;�71;�53;�79Þ mi;USRobotics ¼ðþ56;þ70;þ86 þ 63Þ ð11Þ Note that, as mentioned in Section 2, there is a wide spread of values. This makes it impossible to directly apply the clustering technique. However, we can transform these absolute values into relative values as explained in Section 3.2.2. This way, by normalizing the Fig. 7. Floor plan. L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 6171 values with respect to the reference value P1 in each case, we get the expansion vectors: eP1i;3Com ¼ð�0:75;�0:02;�6:21;�2:69Þ eP1i;Enterasys ¼ð�2:19;�0:94;�2:40;þ1:83Þ eP1i;Ov islink ¼ðþ1:83;�0:11;�7:13;þ4:72Þ eP1i;SMC ¼ð�0:29;þ0:67;þ2:27;þ1:99Þ eP1i;USRobotics ¼ðþ0:63;�1:48;þ0:94;�0:78Þ ð12Þ By normalizing the values with respect to the reference point P4, we get: eP4i;3com ¼ð�5:52;�21:87;þ19:63;þ0:12Þ eP4i;Enterasys ¼ð�6:61;�39:14;þ21:03;þ1:18Þ eP4i;Ov islink ¼ð�6:46;�24:61;þ19:49;þ3:04Þ eP4i;SMC ¼ð�7:07;�31:10;þ22:51;�0:01Þ eP4i;USRobotics ¼ð�8:18;�34:09;þ22:76;þ1:58Þ ð13Þ By normalizing the values with respect to the reference point P6, we get: eP6i;3com ¼ð�7:34;þ0:83;þ22:18;�28; 67Þ eP6i;Enterasys ¼ð�6:9;�7; 59;þ25; 34;�22:46Þ eP6i;Ov islink ¼ð�13:61;þ1:34;þ23:74;�19:89Þ eP6i;SMC ¼ð�12:67;�5:43;þ26:58;�25:12Þ eP6i;USRobotics ¼ð�11:83;�7:51;þ29:24;þ22:24Þ ð14Þ and, finally, by normalizing the values with respect to the reference value P7, we get: eP7i;3com ¼ð�23:92;þ6:12;þ35:59;þ2:01Þ eP7i;Enterasys ¼ð�30:28;þ3; 21;þ40:38;þ1:02Þ eP7i;Ov islink ¼ð�35:82;þ9:74;þ34:58;þ3:03Þ eP7i;SMC ¼ð�31:08;þ4:74;þ37:37;þ1:78Þ eP7i;USRobotics ¼ð�33:73;þ2:09;þ34:87;þ2:24Þ ð15Þ After analysing the experimental results, we find that we have managed to get very similar relative, card-independent values using this method. This is because we measure the relative differ- ence between two points across all the cards. This way, we can en- ter these new relative data into our clustering system. The deviations in the measurements, as shown in Fig. 8, are insignifi- cant compared with the absolute values of the analysed cards. Additionally, we could undertake the clustering process from k perspectives (see Section 3.2.2). This will make the mobile user location more precise. 4.1. Experiment 1: clustering data for all wireless cards This is the first experiment we did by calibrating seven different wireless devices and collecting received power data from the spec- ified APs at the locations of interest. After building the four SOMs, we named the clearly separate re- gions in each one. For naming purposes, the SOM is plotted on a chart. The measurement position indicated in the data is super- posed, and, as you can see, the positions are clearly grouped (see Fig. 9). Having named the regions, each SOM will classify the measure- ments in a particular location region or area where the wireless de- vice could be located. Therefore, as we have created four SOMs, we will have four regions or locations for the device depending on the AP on which the measurement is normalized. This way, we have located the wireless device. If the result of the four SOMs is the same position, this is the device’s location. Otherwise we have to analyse the different SOM responses to be able to locate the device. Fig. 10 shows the confusion matrix for the SOMs using the test set data. The table rows list the positions predicted by the system, whereas the columns list the real positions. Clearly, the system correctly predicts position P1, for example, in 98.15% of the cases, and these points are incorrectly classified in 1.85% of the cases. These results are quite acceptable for some positions, like P1; P12 or P4, for example, but not for others, like P3 or P7, where most of the test cases are classified incorrectly. Many of these incorrect classifications correspond to the region defined as multi- ple. This means that the four SOMs indicate that the wireless device is equally likely to be located in several positions. Fig. 10. Confusion matrix for SOM predictions. Fig. 8. Comparison of absolute/normalized measurement STD. Fig. 9. Region naming. 6172 L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 A voting algorithm has been used to find out the position or positions (also termed regions) of the wireless device that the SOMs together locate in multiple positions. In this algorithm, each SOM indicates the region that it assigns to the device. The final re- gion is the most voted by all the SOMs. The percentage of correctly assigned regions is calculated using the following formula: L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 6173 XReg3Pos Red %CorrectedReg ð16Þ Fig. 11 shows the result of using the location regions instead of the exact measurement position. Comparing the total results (cor- rected line) in Figs. 10 and 11, we find that the predictions have improved in all cases as a result of using the regions. Taking posi- tion P3, for example, we find that the system gives a correct re- sponse in 1.85% of the cases using positions, whereas the success rate is 90.74% if we use regions. Fig. 11. Regions con Fig. 12. Position vs. region $Region Count Position $Region P1 P12 P3 P4 P1 98,15% P4,P5 98,15% P7,P9,P12 5,56% P3,P6,P12 94,44% 90,74% P6,P9,P12 83,33% P8,P9 Corrected 98,15% 94,44% 90,74% 98,15% Fig. 13. Optimized Fig. 12 is a graph comparing the success rate for positions and regions. It shows that the system’s success rate increases signifi- cantly using regions. The system’s success rate increases using the regions, but the approach has the drawback of there being very extensive regions covering a large part of the physical space in which the wireless de- vice can be located, e.g. ½P4; P5; P8; P9� or ½P5; P9; P12�. To reduce the positioning area, at this point we enter information about the layout of the physical space of the premises where the system is going to be used, e.g. information related to the proximity of the fusion matrix. prediction comparison. P5 P6 P7 P8 P9 61,11% 97,78% 90,74% 92,59% 16,67% 40,74% 75,56% 61,11% 92,59% 97,78% 75,56% 90,74% region results. Fig. 14. Comparison between estimated position, region and optimized region. 6174 L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 regions. The results with the optimized regions are shown in Fig. 13. Comparing the results with the non-optimized regions (see Fig. 14), we find that the success rates drops in some cases and increases in others. The advantage is, though, that the regions are quite a lot smaller, and the positioning is much more accurate in most cases. Fig. 14 is a comparison of the three wireless device location esti- mation methods. In view of these results, we conclude that the best and most accurate way to do things is to use regions optimized according to the physical arrangement of the premises. 4.2. Experiment 2: clustering data for one wireless card This experiment addresses a real operating environment where there is at first just one type of wireless device on which to take the $Region Count Position $Region P1 P12 P3 P1 97,78% P7,P8,P9 81,11% P3,P6 94,89 P4,P5 P7,P9 64,00% P8,P9 60,00% Corrected 97,78% 81,11% 94,89 Fig. 15. Optimized Fig. 16. Comparison between estimated p measurements. As wireless devices of other brands are acquired, they are located without modifying the SOMs or the regions de- fined for the first device. Therefore, unlike the last experiment, where all the measure- ments on the available wireless devices were used, we have not di- vided the data into a training set and a test set. In this case, the SMC card data form the training set and the data on the other cards (not used to define the SOMs) make up the test set. Fig. 15 shows the re- sults of locating different devices from the data used to create the SOMs. As you can see, success rate is over 90% in most cases. Fig. 16 shows the comparison between using the position, the region and the optimized region if a single card is used to create the SOMs. Finally, we can compare the results of using data from a single wireless device and using the data from all the devices to create P4 P5 P6 P7 P8 P9 % 95,33% 96,89% 74,00% 88,06% 91,11% 91,39% % 96,89% 74,00% 95,33% 88,06% 91,39% 91,11% region results. osition, region and optimized region. Fig. 17. One wireless device vs. all wireless devices used to build SOMs. L. Mengual et al. / Expert Systems with Applications 37 (2010) 6165–6175 6175 the SOMs. Fig. 17 shows that the results are quite similar. The out- come is sometimes better using all the cards, and the use of just one card gets better results in other cases. From these results we can conclude that the system can be used in any environment with a wireless device and at least three APs. 5. Conclusions We conceived and developed a system that is highly likely to correctly estimate the location of a mobile terminal indoors in a room or adjacent rooms. Unlike many other implemented systems, the resulting system is independent of the card brand and hardware technology (USB, PCMCIA) manufacturer. Our system evaluates the position by rela- tive differences in the measurements taken at places close to the APs. The system is also adaptable to any environment. All it takes is system training with any card at locations on the floor, the nor- malization of the measurements and the creation of clusters. As of this point, the system is ready to locate any user with any wireless card. Role of the funding source This work was conducted as part of the CYCIT-funded Project No. TIN2008-05924. References Ali, S., & Nobles, P. (2007). A novel indoor location sensing mechanism for IEEE 802.11 b/g wireless lan. In 4th Workshop on positioning, navigation and communication (WPNC ’07) (pp. 9–15). Astrain, J. J., Villadangos, J., Garitagoitia, J. R., González-Mendivil, J. R., & Cholvim, V. (2006). Fuzzy location and tracking on wireless networks. In International workshop on mobility management and wireless access (pp. 84–91). Battiti, R., Le Nhat, T., & Villani, A. (2002). Location-aware computing: A neural network model for determining location in wireless lans. Technical Report DIT- 5. Universita di Trento, Dipartimento di Informatica e Telecomunicazioni. Chai, X., & Yang, Q. (2007). Reducing the calibration effort for probabilistic indoor location estimation. IEEE Transactions on Mobile Computing, 6(6), 649–662. June. Debono, C. J., & Buhagiar, J. K. (2004). Neural location detection in wireless network. In 7th European conference on wireless technology. Golden, S. A., & Bateman, S. S. (2007). Sensor measurements for Wi-Fi location with emphasis on time-of-arrival ranging. IEEE Transactions on Mobile Computing, 6(10), 1185–1198. Kaemarungsi, K. (2005). Design of indoor positioning systems based on location fingerprinting technique. PhD thesis. University Pittsburg. Kaemarungsi, K. (2006). Distribution of WLAN received signal strength indication for indoor location determination. In 1st International symposium on wireless pervasive computing. Kontkanen, P., Myllymäki, P., Roos, T., Tirri, H., Valtonen, K., & Wettig, H. (2004). Topics in probabilistic location estimation in wireless networks. In 15th IEEE international symposium on personal, indoor and mobile radio communications, Barcelona, Spain. IEEE Press. Nezafat, M., Kaveh, M., Tsuji, H., & Fukagawa, T. (2004). Localization of wireless terminals using subspace matching with ray-tracing-based simulations. In Sensor array and multichannel signal processing workshop proceedings (pp. 623– 627). Nuño, G., & Páez-Borrallo, J. M. (2006). New location estimation system for wireless networks based on linear discriminant functions and hidden Markov models. EURASIP Journal on Applied Signal Processing. Pan, J. J., Kwok, J. T., Yang, Q., & Chen, Y. (2006). Multidimensional vector regression for accurate and low-cost location estimation in pervasive computing. IEEE Transactions on Knowledge and Data Engineering, 18(9), 1181–1193. Raniwala, Ashish, & Chiueh, Tzi-Cker (2002). Deployment issues in enterprise wireless lans. Wireless Communications. Robinson, M., & Psaromiligkos, I. (2005). Received signal strength based location estimation of a wireless lan client. IEEE Wireless Communications and Networking Conference, 4, 2350–2354. Sanchez, D., Afonso, S., Macias, E. M., & Suarez, A. (2006). Devices location in 802.11 infrastructure networks using triangulation. In The 2006 IAENG international workshop on wireless networks, Hong Kong. Seigo, I., & Kawaguchi, N. (2005). Bayesian based location estimation system using wireless lan. In Proceedings of the third IEEE international conference on pervasive computing and communications workshops (pp. 273–278). Widyawan, K., Klepal, M., & Pesch, D. (2007). Influence of predicted and measured fingerprint on the accuracy of RSSI-based indoor location systems. In 4th Workshop on Positioning, Navigation and Communication (WPNC ’07) (pp. 145– 151). Yiming, J., Biaz, S., Pandey, S., & Agrawal, P. (2006). Ariadne: A dynamic indoor signal map construction and localization system. In International conference on mobile systems, applications and services (pp. 151–164). Youssef, M., & Agrawala, A. K. (2005). The Horus WLAN location determination system. In Third international conference on mobile systems, applications, and services (MobiSys 2005), Seattle, WA, USA. Zhong, J., Bin-Hong, L., Hao-Xing, W., Hsing-Yi, C., & Sarkar, T. K. (2001). Efficient ray-tracing methods for propagation prediction for indoor wireless communications. IEEE Antennas and Propagation Magazine, 43(2). Clustering-based location in wireless networks Introduction Location estimation system issues Location estimation methodology Outline of the system Location estimation algorithm Measurement phase Calibration phase Estimation step Results of evaluation experiments Experiment 1: clustering data for all wireless cards Experiment 2: clustering data for one wireless card Conclusions Role of the funding source References 
work_pvhpgmyphjaytf5rsqpo4744nu	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy Bootstrap testing fuzzy hypotheses and observations on fuzzy statistic Mohammad Ghasem Akbari a,*, Abdolhamid Rezaei b a Department of Statistics, Faculty of Sciences, University of Birjand, South Khorasan, Iran b Department of Statistics, School of Mathematical Sciences, Ferdowsi University of Mashhad, Mashhad, Iran a r t i c l e i n f o Keywords: Fuzzy canonical number Yao–Wu signed distance Fuzzy data Fuzzy hypotheses Fuzzy confidence Testing hypotheses Bootstrap method a b s t r a c t The bootstrap is a simple and straightforward method for calculating approximated biases, standard deviations, confidence intervals, testing statistical hypotheses, and so forth, in almost any nonparametric estimation problem. A new approach for bootstrap testing fuzzy hypotheses based on fuzzy test statistic is introduced. In this paper we describe bootstrap method that is designed directly for hypothesis testing for fuzzy data based on Yao–Wu signed distance. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction Statistical analysis, in traditional form, is based on crispness of data, random variable, point estimation, hypotheses, parameter and so on. As there are many different situations in which the above mentioned concepts are imprecise. The point estimation ap- proaches frequently are used in statistical inference. On the other hand, the theory of fuzzy sets is a well known tool for formulation and analysis of imprecise and subjective concepts. Therefore the testing hypotheses with fuzzy data can be important. The problem of testing statistical hypotheses when the hypotheses are crisp or fuzzy have been studied by a few authors. Arnold (1996, 1998) presented an approach to test fuzzily for- mulated hypotheses, in which he considered fuzzy constraints on the type I and II errors. Taheri and Behboodian (1991) state and prove a Neyman–Pearson Lemma for testing fuzzy hypotheses. Their approach has been extended by Torabi, Behboodian, and Taheri (2006) for the cases when the data are fuzzy, too. For some other recent works in testing hypothesis using fuzzy approaches, see Buckley (2005, 2006), Desimpelaere and Marchant (2007), Fil- zmoser and Viertl (2004), Hyniewicz (2006), Thompson and Geyer (2007), Viertl (2006). The bootstrap using fuzzy data, is developed in different approaches. Montenegro, Colubi, Casals, and Gil (2004) have presented asymptotic one-sample procedure. Korner’s asymptotic develop- ment (2000) concerns general fuzzy random variables (taking on way-either finite or infinite-number of values in the space of com- pact convex fuzzy sets of a finite-dimensional Euclidean space). Gonzalez-Rodriguez, Montenegro, Colubi, and Gil (2006) have shown that the one-sample method of testing the mean of a fuzzy random variable can be extended to general ones (more precisely, to those whose range is not necessarily finite and whose values are fuzzy subsets of finite-dimensional Euclidean space). In this paper we construct a new method for bootstrap testing hypotheses in fuzzy environment which is completely different from those mentioned above. For this purpose we organize the matter in the following way: in Section 2 we describe some basic concepts of canonical fuzzy numbers, Yao and Wu (2000) signed distance and fuzzy hypothe- ses. In Section 3 we come up bootstrap testing fuzzy hypotheses based on fuzzy test statistic for mean based on Yao–Wu signed dis- tance. Section 4 provide a Bootstrap testing fuzzy hypotheses based on fuzzy test statistic for variance based on Yao-Wu signed distance. A brief conclusion is provided in Section 5. 2. Preliminaries In this section we study canonical fuzzy numbers, Yao–Wu signed distance and fuzzy hypotheses. 2.1. Canonical fuzzy numbers Let X be the universal space, then a fuzzy subset ~x of X is defined by its membership function l~x : X !½0; 1�. We denote by ~xa ¼ fx : l~xðxÞ P ag the a-cut set of ~x and ~x0 is the closure of the set fx : l~xðxÞ > 0g, and (1) ~x is called normal fuzzy set if there exist x 2 X such that l~xðxÞ¼ 1; (2) ~x is called convex fuzzy set if l~xðkx þð1 � kÞyÞ P min ðl~xðxÞ; l~xðyÞÞ for all k 2 ½0; 1�; 0957-4174/$ - see front matter � 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.02.030 * Corresponding author. Tel./fax: +98 511 8828605. E-mail address: g_z_akbari@yahoo.com (M.G. Akbari). Expert Systems with Applications 37 (2010) 5782–5787 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy (3) the fuzzy set ~x is called a fuzzy number if ~x is normal convex fuzzy set and its a-cut sets are bounded 8a – 0; (4) ~x is called a closed fuzzy number if ~x is fuzzy number and its membership function l~x is upper semicontinuous; (5) ~x is called a bounded fuzzy number if ~x is a fuzzy number and its membership function l~x has compact support. If ~x is a closed and bounded fuzzy number with xLa ¼ inffx : x 2 ~xag and xUa ¼ supfx : x 2 ~xag and its membership function is strictly increasing on the interval xLa; x L 1 � � and strictly decreasing on the interval xU1 ; x U a � � for any a 2 ½0; 1�, then ~x is called canonical fuzzy number. Let ‘‘�” be a binary operation � or � between two canonical fuzzy numbers ~a and ~b. The membership function of ~a � ~b is de- fined by l~a�~bðzÞ¼ sup x�y¼z minfl~aðxÞ; l~bðyÞg 8z 2R for �¼� or � and �¼þ or -. In the following, let �int be a binary operation �int or �int between two closed intervals ~aa ¼ aLa; aUa � � and ~ba ¼ b L a; b U a h i . Then ~aa�int ~ba is defined by ~aa�int ~ba ¼fz 2R : z ¼ x � y; x 2 ~aa; y 2 ~bag: If ~a and ~b be two closed fuzzy numbers. Then ~a � ~b and ~a � ~b are also closed fuzzy numbers. Furthermore, we have ð~a � ~bÞa ¼ ~aa�int ~ba ¼ a L a þ b L a; a U a þ b U a h i ; ð~a � ~bÞa ¼ ~aa�int ~ba ¼ a L a � b U a ; a U a � b L a h i : 2.2. Yao–Wu signed distance Now we define a signed distance between fuzzy numbers which is used later. Several ranking methods have been proposed so far, by Cheng (1998), Modarres and Sadi-Nezhad (2001) and Nojavan and Gha- zanfari (2006). In this paper we use another ranking system for canonical fuzzy numbers which is very realistic and is defined by Yao and Wu as the following: Definition 2.1. For each a; b 2R, define the signed distance d� of a and b by d�ða; bÞ¼ a � b. Thus, we have the following way to define the rank of any two numbers on R. For each a; b 2R, d�ða; bÞ > 0 () d�ða; 0Þ > d�ðb; 0Þ() a > b; d�ða; bÞ < 0 () d�ða; 0Þ < d�ðb; 0Þ() a < b; d�ða; bÞ¼ 0 () d�ða; 0Þ¼ d�ðb; 0Þ() a ¼ b: Definition 2.2. For each ~a; ~b 2FðRÞ, define the signed distance of ~a and ~b as follows: dð~a; ~bÞ¼ Z 1 0 Mað~aÞ� Mað~bÞ � � da ¼ Z 1 0 d� Mað~aÞ; Mað~bÞ � � da; where Mað~aÞ and Mað~bÞ are equal to aLaþa U a 2 and bLaþb U a 2 , respectively, furthermore dð~a; ~bÞ means the distance of ~a to ~b. Definition 2.3 (Yao and Wu, 2000). For each ~a; ~b 2FðRÞ, define the ranking ; and � of ~a and ~b by dð~a; ~bÞ > 0 () dð~a; 0Þ > dð~b; 0Þ() ~a ~b; dð~a; ~bÞ < 0 () dð~a; 0Þ < dð~b; 0Þ() ~a ~b; dð~a; ~bÞ¼ 0 () dð~a; 0Þ¼ dð~b; 0Þ() ~a � ~b: 2.3. Fuzzy hypotheses We define some models, as fuzzy sets of real numbers, for modeling the extended versions of the simple, the one- sided, and the two-sided ordinary (crisp) hypotheses to the fuzzy ones. Testing statistical hypothesis is a main branch of statistical inference. Typically, a statistical hypothesis is an assertion about the probability distribution of one or more random variable(s). Traditionally, all statisticians assume the hypothesis for which we wish to provide a test are well-defined. This limitation, some- times, forces the statistician to make decision procedure in an unrealistic manner. This is because in realistic problems, we may come across non-precise (fuzzy) hypothesis. For example, suppose that h is the proportion of a population which has a disease. We take a random sample of elements and study the sample for having some idea about h. In crisp hypothesis testing, one uses the hypotheses of the form: H0 : h ¼ 0:2 versus H1 : h – 0:2 or H0 : h 6 0:2 versus H0 : h > 0:2, and so on. However, we would sometimes like to test more realistic hypotheses. In this example, more realistic expressions about h would be considered as: small, very small, large, approximately 0.2, and so on. Therefore, more realistic formulation of the hypotheses might be H0 : h is small, versus H1 : h is not small. We call such expressions as fuzzy hypotheses. We define some models, as fuzzy sets of real numbers, for mod- eling the extended versions of the simple, the one-sided, and the two-sided crisp hypotheses to the fuzzy ones. Definition 2.4. Let h0 be a real number and known. (i) Any hypothesis of the form (H : h is approximately h0) is called to be a fuzzy simple hypothesis. (ii) Any hypothesis of the form (H : h is not approximately h0 ) is called to be a fuzzy two-sided hypothesis. (iii) Any hypothesis of the form (H : h is essentially smaller than h0 ) is called to be a fuzzy left one-sided hypothesis. (iv) Any hypothesis of the form (H1 : h is essentially larger than h0 ) is called to be a fuzzy right one-sided hypothesis. We denote the above definitions by (w) H0 : h is approximately h0 H1 : h is not approximately h0 � or H0 : h is eH 0 H1 : h is eH 1; � (ww) H0 : h is approximately h0 H1 : h is certainly larger than h0 � or H0 : h is eH 0L H1 : h is eH 1; � (www) H0 : h is approximately h0 H1 : h is certainly smaller than h0 � or H0 : h is eH 0R H1 : h is eH 1: � The above areas are shown in Figs. 1–3. 3. Bootstrap testing fuzzy hypotheses based on fuzzy test statistic for mean Suppose that we have canonical fuzzy random sample ~x ¼ð~x1; ~x2; . . . ; ~xnÞ. 3.1. Testing fuzzy simple hypothesis fuzzy against the fuzzy two-sided hypothesis We want to test the fuzzy null hypotheses eH 0 : the mean of observations ðhÞ is approximately h0 versuseH 1 : the mean of observations ðhÞ is not approximately h0 . M.G. Akbari, A. Rezaei / Expert Systems with Applications 37 (2010) 5782–5787 5783 Author's personal copy We generate B bootstrap fuzzy random sample ~x� 1 ; ~x� 2 ; . . . ; ~x� B (i.e., each ~x� b is a fuzzy sample of size n drawn randomly and replacement from ~x). We need a distribution that estimates the population of treat- ment times under H0 . Note first that the empirical distribution (i.e., putting probability 1n on each member of ~x) is not an approx- imate estimate for distribution because it does not obey H0. Some- how we need to obtain an estimate of the population that has mean ~h0. A simple way is to translate the empirical distribution so that it has the desired mean. In other words, we use as our esti- mated null distribution the empirical distribution on the values ~xci ¼ ~xi � ex �fh0; i ¼ 1; 2; . . . ; n; because 1n� n i¼1~xci ¼ fh0 . We compute t ~x� b c � � ¼ d ex�bc ;eh0� �cSe�bðexcÞ b ¼ 1; 2; . . . ; B; where (1) d is Yao–Wu signed distance. (2) ex�bc ¼ 1n�ni¼1~x�bci b ¼ 1; 2; . . . ; B. (3) cSe�bðexcÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1nðn�1ÞPni¼1 d2 ~x�bci ; ex�bc� � r b ¼ 1; 2; . . . ; B. The cth percentile of tð~x�bc Þ is estimated by the value t̂c such that # t ~x� b c � � 6 t̂c n o B ¼ c: Finally, the fuzzy bootstrap confidence interval at the level 1 � 2c using fuzzy data is ePa ¼ 1n X n i¼1 xi � t̂c sð~xÞffiffiffi n p ; 1 n Xn i¼1 xi � t̂1�c sð~xÞffiffiffi n p " # ; ( xi 2 ~xia; i ¼ 1; 2; . . . ; n ) ; where s2ð~xÞ ¼ 1 n�1 Pn i¼1 d 2ð~xi; ~�xÞ. If B � c is not an integer, the following procedure can be used. Assuming c 6 0:5, let k ¼ ½ðB þ 1Þc�, the largest integer 6 ðB þ 1Þc. Then we define the empirical c and 1 � c quantizes by the kth larg- est and ðB þ 1 � kÞth largest values of tð~x�bc Þ, respectively. Example 3.1. Suppose that we have taken a fuzzy random sample of size n ¼ 9 from a population and we observed the following triangular fuzzy data: and using of the ability of package ‘‘it MINITAB 13”, we show the percentiles and histogram of t ~x� b c � � in the following using 10,000 bootstrap samples. If B ¼ 10; 000, the estimate of the 5% point is the 500th largest value of the t ~x� b c � � s and the estimate of the 95% point is the 9500th largest value of the t ~x� b c � � s. The a-cuts of fuzzy bootstrap confidence interval (c ¼ 0:05 or 90%) using fuzzy data is ePa ¼ 19 X 9 i¼1 xi �1:781 20:36ffiffiffi 9 p ; 1 9 X9 i¼1 xi þ2:088 20:36ffiffiffi 9 p " # ; ( xi 2 ~xia; i ¼ 1;2; . . . ;9 ) ¼½46:153þ2:22a;76:167�1:556a�: We show the following the distribution of tð~x�bc Þ computed using 10000 bootstrap samples (Fig. 4). We obtain the a- cuts of the so-called fuzzy test statistics -10 0 10 0 100 200 300 400 500 600 700 800 900 1000 Fr eq ue nc y Fig. 4. bootstrap distribution of t ~x� b c � � . Fig. 2. The fuzzy hypotheses eH 0L versus eH 1 . Fig. 3. The fuzzy hypotheses eH 0R versus eH 1 . Fig. 1. The fuzzy hypotheses eH 0 versus eH 1 . 5784 M.G. Akbari, A. Rezaei / Expert Systems with Applications 37 (2010) 5782–5787 Author's personal copy eZ a ¼ ePa � ~h0asð~xÞffiffi n p ; and use the fuzzy test statistics to provide an approach for testing above fuzzy hypotheses, based on the following assumptions (see Fig. 5). ASSUMPTIONS 1. CT be the total area under eZ . 2. C1 and C2 be the areas according to Fig. 5. 3. CR ¼ C1 þ C2 DECISION RULE If CRCT 6 2c, then we accept H0 . If CRCT P 2c, then we reject H0 . Example 3.2. Consider Table 1. Now suppose that we want to test the following fuzzy hypotheses eH 0 : h isð57; 60; 62ÞeH 1 : h is notð57; 60; 62Þ: ( Here, eH 0 suggests that h is approximation 60, and eH 1 suggests that h is away from 60. We have eZ a ¼ ½�2:339 þ 0:622a; 2:899 � 0:671a�; CR ¼ C1þ C2 ¼ 0:061 þ 0:18 ¼ 0:241; CT ¼ 3:39. since CRCT ¼ 0:071 6 0:1, thus accept H0 (Table 2). 3.2. Testing fuzzy simple hypothesis fuzzy against the fuzzy left one- sided hypothesis We want to test the fuzzy null hypotheses H0 : h is approximately h0; H1 : h is certainly smaller than h0: � We obtain the a-cuts of the so-called fuzzy test statistics eZ a ¼ ePa �eh0as ðexÞffiffi n p ; and use the fuzzy test statistics to provide an approach for testing above fuzzy hypotheses, based on the following assumptions (see Fig. 6) ASSUMPTIONS 1. CT be the total area under eZ . 2. C1 be the area according to Fig. 6. 3. CR ¼ C1 DECISION RULE If CRCT 6 c, then we accept H0. If CRCT P c, then we reject H0. Example 3.3. Consider Table 1. Now suppose that we want to test the following fuzzy hypotheses eH 0 : h is approximatelyð:; 60; 62Þ;eH 1 : h is certainly smaller thanð:; 60; 62Þ: ( We have eZ a ¼ ½�2:339 þ 0:622a; 2:382 � 0:23a�; CR ¼ C1 ¼ 0:061; CT ¼ 2:79. since CRCT ¼ 0:022 6 0:05, thus accept H0 . 3.3. Testing fuzzy simple hypothesis fuzzy against the fuzzy right one- sided hypothesis We want to test the fuzzy null hypotheses H0 : h is approximately h0; H1 : h is certainly larger than h0: � We obtain the a-cuts of the so-called fuzzy test statistics eZ a ¼ ePa �eh0asð~xÞffiffi n p ; and use the fuzzy test statistics to provide an approach for testing above fuzzy hypotheses, based on the following assumptions (see Fig. 7) Fig. 5. eZ in testing fuzzy simple hypothesis versus fuzzy two-sided hypothesis. Table 1 Fuzzy random sample of size n ¼ 9 from a population. N Observation N Observation N Observation 1 (32, 35, 40) 4 (60, 63, 63) 7 (70, 73, 75) 2 (80, 82, 82) 5 (41, 45, 47) 8 (54, 56, 59) 3 (60, 60, 60) 6 (93, 95, 96) 9 (34, 35, 36) Table 2 Percentiles of the t distribution with 8 degree of freedom, the Nð0; 1Þ distribution and the bootstrap distribution of tð~x�bc Þ. Percentile 5% 10% 90% 95% t8 �1.86 �1.40 1.40 1.86 Normal �1.65 �1.28 1.28 1.65 Bootstrap �2.088 �1.537 1.319 1.781 Fig. 6. eZ in testing fuzzy simple hypothesis versus fuzzy left one-sided hypothesis. M.G. Akbari, A. Rezaei / Expert Systems with Applications 37 (2010) 5782–5787 5785 Author's personal copy ASSUMPTIONS 1. CT be the total area under eZ . 2. C2 be the area according to Fig. 7. 3. CR ¼ C2 DECISION RULE If CRCT 6 c, then we accept H0. If CRCT P c, then we reject H0. Example 3.4. Consider Table 1. Now suppose that we want to test the following fuzzy hypotheses eH 0 : h is approximatelyð28; 30; :Þ;eH 1 : h is certainly larger thanð28; 30; :Þ: ( We have eZ a ¼ ½2:38 þ 0:327a; 7:097 � 0:524a�; CR ¼ C2 ¼ 2:772; CT ¼ 16:9. since CRCT ¼ 0:164 P 0:05, thus reject H0 . 4. Bootstrap testing fuzzy hypotheses based on fuzzy test statistic for variance Suppose that we have canonical fuzzy random samples ~x ¼ð~x1; ~x2; . . . ; ~xnÞ. We generate B bootstrap fuzzy random sample ~x� 1 ; ~x� 2 ; . . . ; ~x� B and for each we compute (Fig. 8) v2� b ¼ ðn � 1Þs2�bð~xÞ s2ð~xÞ b ¼ 1; 2; . . . ; B; where (1) s2� b ð~xÞ ¼ 1 n�1 Pn i¼1d 2ð~x�bi ; ex�Þ. (2) d is Yao–Wu signed distance. (3) ex� ¼ 1n�ni¼1~x�bi . (4) s2 ðexÞ ¼ 1n�1 Pni¼1 d2ð~xi; exÞ. The cth percentile of v2�b is estimated by the value t̂c such that #fv2�b 6 t̂cg B ¼ c: Finally, the a-cuts of bootstrap confidence interval using fuzzy data is eP�a ¼ Pn i¼1ðxi � �xÞ 2 t̂1�c ; Pn i¼1ðxi � �xÞ 2 t̂c " # : xi 2 ~xia; i ¼ 1; 2; . . . ; n ( ) ; whenever its membership function is given by leP�ðyÞ¼ sup06a61 aIeP�aðyÞ: Example 4.1. Suppose that we have taken a fuzzy random sample of size n ¼ 12 from a population and we observed the following triangular fuzzy data: (Table 3). The last line of Table 4 shows the percentiles of v2� b for variance computed using 10000 bootstrap samples. The fuzzy bootstrap confidence interval (c ¼ 0:05 or 90%) using fuzzy data is eP�a ¼ P12 i¼1ðxi � �xÞ 2 15:27 ; P12 i¼1ðxi � �xÞ 2 4:523 " # : xi 2 ~xia i ¼ 1; 2; . . . ; n ( ) ; and we have for some a0’s a 0 0.1 0.2 Confidence interval [317.68, 1115.83] [318.1, 1112.66] [318.39, 1109.59] a 0.3 0.4 0.5 Confidence interval [318.8, 1106.63] [319.24, 1103.24] [319.71, 1101.02] a 0.6 0.7 0.8 Confidence interval [320.21, 1098.38] [320.74, 1095.84] [321.3, 1093.41] a 0.9 1 Confidence interval [321.9, 1091.08] [322.52, 1088.86] We obtain the a-cuts of the so-called fuzzy test statistics eZ�a ¼ðn � 1ÞeP�aeh0a ; and now we can similarly apply the previous section for analyzing of this section. 0 10 20 0 100 200 300 400 500 Fr eq ue nc y Fig. 8. Bootstrap distribution of v2�b . Table 3 Fuzzy random sample of size n ¼ 12 from a population. N Observation N Observation N Observation 1 (33, 35, 36) 5 (60, 63, 66) 9 (100, 103, 105) 2 (80, 82, 84) 6 (70, 70, 72) 10 (54, 56, 58) 3 (85, 87, 87) 7 (70, 73, 76) 11 (40, 40, 42) 4 (90, 90, 90) 8 (65, 70, 73) 12 (94, 96, 99) Fig. 7. eZ in testing fuzzy simple hypothesis versus fuzzy left one-sided hypothesis. 5786 M.G. Akbari, A. Rezaei / Expert Systems with Applications 37 (2010) 5782–5787 Author's personal copy 5. Conclusion The proposed procedure is based on a significance level. Exten- sion of the proposed method to test the parameters of linear mod- els, such as regression models, design of experiment is a potential area for the future work. We can use the Yao–Wu signed distance for estimation of coefficients of linear models and apply the pro- posed method in this paper. Acknowledgements Authors are grateful to the referees of the journal for their sug- gestions and would like to thank the Co-Editor-in-Chief Doctor Vijayakumar for the helpful comments and useful helps. References Arnold, B. F. (1996). An approach to fuzzy hypothesis testing. Metrika, 44, 119–126. Arnold, B. F. (1998). Testing fuzzy hypothesis with crisp data. Fuzzy Sets and Systems, 94, 323–333. Buckley, J. J. (2005). Fuzzy probabilities: New approach and applications. Berlin, Heidelberg: Springer-Verlag. Buckley, J. J. (2006). Fuzzy probability and statistics. Berlin, Heidelberg: Springer- Verlag. Cheng, C. (1998). A new approach for ranking fuzzy numbers by distance method. Fuzzy Sets and Systems, 95, 307–317. Desimpelaere, C., & Marchant, T. (2007). An empirical test of some measurement- theoretic axioms for fuzzy sets. Fuzzy Sets and Systems, 158, 1348–1359. Filzmoser, P., & Viertl, R. (2004). Testing hypotheses with fuzzy data: The fuzzy p- value. Metrika, 59, 21–29. Gonzalez-Rodriguez, G., Montenegro, M., Colubi, A., & Gil, M. A. (2006). Bootstrap techniques and fuzzy random variables: Synergy in hypothesis testing with fuzzy data. Fuzzy Sets and Systems, 157, 2608–2613. Hyniewicz, O. (2006). Possibilitic decisions and fuzzy statistical tests. Fuzzy Sets and Systems, 157, 2665–2673. Korner, R. (2000). An asymptotic a-cut for the expectation of random fuzzy variables. Journal of Statistical Planning and Inferences, 83, 331–346. Modarres, M., & Sadi-Nezhad, S. (2001). Ranking fuzzy numbers by preference ratio. Fuzzy Sets and Systems, 118, 429–436. Montenegro, M., Colubi, A., Casals, M. R., & Gil, M. A. (2004). Asymptotic and bootstrap techniques for testing the expected value of a fuzzy random variable. Metrika, 59, 31–49. Nojavan, M., & Ghazanfari, M. (2006). A fuzzy ranking method by desirability index. Journal of Intelligent and Fuzzy Systems, 17, 27–34. Taheri, S. M., & Behboodian, J. (1991). Neyman–Pearson Lemma for fuzzy hypotheses testing. Metrika, 49, 3–17. Thompson, E. A., & Geyer, C. J. (2007). Fuzzy p-values in latent variable problems. Biometrika, 94, 49–60. Torabi, H., Behboodian, J., & Taheri, S. M. (2006). Neyman–Pearson Lemma for fuzzy hypotheses testing withy vague data. Metrika, 64, 289–304. Viertl, R. (2006). Univariate statistical analysis with fuzzy data. Computational Statistics and Data Analysis, 51, 133–147. Yao, J. S., & Wu, K. (2000). Ranking fuzzy numbers based on decomposition principle and signed distance. Fuzzy Sets and Systems, 11, 275–288. Table 4 Percentiles of the v2 distribution with 7 and 11 degrees of freedom and the bootstrap distribution of v2�b . Percentile 0.005 0.01 0.025 0.05 0.95 0.975 0.99 0.995 v2;7 0.989 1.239 1.69 2.167 14.067 16.013 18.475 20.278 v2;11 2.603 3.053 3.816 4.575 19.675 21.92 24.725 26.757 Bootstrap 2.699 3.077 3.85 4.523 15.27 16.889 18.29 21.349 M.G. Akbari, A. Rezaei / Expert Systems with Applications 37 (2010) 5782–5787 5787 
work_pvp64m4phrctlisjvtkchdhwju	----	paper4.dvi Towards an expert system for  enantioseparations: induction of rules  using machine learning Bryant, CH, Adam, AE, Taylor, DR and Rowe, RC http://dx.doi.org/10.1016/0169­7439(96)00016­0 Title Towards an expert system for enantioseparations: induction of rules using  machine learning Authors Bryant, CH, Adam, AE, Taylor, DR and Rowe, RC Type Article URL This version is available at: http://usir.salford.ac.uk/1772/ Published Date 1996 USIR is a digital collection of the research output of the University of Salford. Where copyright  permits, full text material held in the repository is made freely available online and can be read,  downloaded and copied for non­commercial private study or research purposes. Please check the  manuscript for any further copyright restrictions. For more information, including our policy and submission procedure, please contact the Repository Team at: usir@salford.ac.uk. mailto:usir@salford.ac.uk Towards an Expert System for Enantioseparations� Induction of Rules Using Machine Learning C�H�Bryant�� A�E�Adam �Computation Department� D�R�Taylor �Chemistry Department� University of Manchester Institute of Science and Technology� PO Box ��� Manchester� M�� QD� United Kingdom� R�C�Rowe Zeneca Pharmaceuticals� Alderley Park� Maccles eld� Cheshire� SK � �NA� United Kingdom� Abstract A commercially available machine induction tool was used in an attempt to automate the acquisition of the knowledge needed for an expert system for enan� tioseparations by High Performance Liquid Chromatography using Pirkle�type chi� ral stationary phases �CSPs�� Various rule�sets were induced that recommended particular CSP chiral selectors based on the structural features of an enantiomer pair� The results suggest that the accuracy of the optimal rule�set is ��� � � �� which is more than ten times greater than the accuracy that would have resulted from a random choice� �Correspondence to� C�H�Bryant� School of Computing and Mathematics� The University of Hudder� s�eld� HD� �DH� United Kingdom� � � Introduction This paper presents the rst results of a project concerned with the development of an expert system for enantioseparations that is the separation of enantiomers� It describes an attempt to automate the rst step in the process of developing such as system using a technique of arti cial intelligence known as machine induction� Although machine induction has been applied to analytical chemistry before �see Section �� the authors believe that this is the rst published work to describe a validated application of machine induction to enantioseparations� The separation of enantiomers by High Performance Liquid Chromatography �HPLC� using chiral stationary phases �CSPs� is based on the formation of tran� sient diastereomeric complexes between the enantiomers of the solute and a chiral selector that is an integral part of the stationary phase� The di�erence in stability between these complexes leads to a di�erence in retention time the enantiomer that forms the less stable complex will be eluted rst� If the di�erence in stabil� ity is too small no separation is observed� Such enantioseparations are important in many scienti c disciplines including stereoselective synthesis mechanistic and catalytic studies agrochemistry medicine and pharmacology� �See ��� for a review of enantioseparations�� Since enantioseparations are performed in many disciplines and since there is a choice of over �� commercially available CSPs guidelines are needed on the choice of materials for enantioseparations by HPLC� A computer system which could guide analysts in the choice of materials for enantioseparations by HPLC would be bene� cial because there are currently few guidelines on how to choose the materials and they are di�cult to access the papers describing them are spread across a wider range of scienti c journals than analysts can be reasonably expected to survey� CHIRBASE ��� ��� ��� is a conventional database which makes data on enantiosep� arations accessible but it is expensive� Furthermore it does not tell an analyst how to use such data that is guide an analyst in the selection of materials for a particular enantioseparation� CHIRULE is a computer system that was designed to provide such guidance� CHIRULE was developed by Stau�er and is described in PhD thesis ���� It uses similarity searching on molecular properties to retrieve a list of enantiomer pairs that are chemically similar to a given enantiomer pair together with columns that have been reported in the literature to have successfully separated them� However in his thesis Stau�er does not report testing CHIRULE � to see which CSPs it would recommend when it was given enantiomer pairs which have been reported in the literature as having been separated on Pirkle�type CSPs� Pirkle�type CSPs are so named because their invention is credited to W�H�Pirkle�s group at the University of Illinois� They are also referred to as the �brush� or �multiple interaction� type� They are chiral selectors of moderate molecular weight covalently bonded to silica� As far as the authors of this work are aware this is the rst published work to have taken a validated rst step towards a computer system that gives guidance on the selection of materials for enantioseparations on Pirkle�type CSPs� The remainder of this paper describes the rst results of a project concerned with the development of an expert system for enantioseparations by HPLC� An expert system is a computer program that represents and reasons with knowledge of some specialist subject with a view to solving problems or giving advice���� The characteristics of expert systems are described in ��� together with previous expert systems for chromatography� � Machine Induction This section introduces a technique of arti cial intelligence called machine induc� tion a branch of machine learning and explains why it has been used as a rst step towards developing an expert system for enantioseparations� The original na� ture of the work described in this paper is illustrated by brie�y reviewing previous applications of machine induction to analytical chemistry� The process of acquiring the knowledge needed for an expert system is called knowledge acquisition� The knowledge acquisition process is usually divided into three stages deciding what knowledge is needed variously referred to as the def� inition stage or initial analysis� getting knowledge predominantly from human experts and interpreting it usually called elicitation� and �writing� the knowledge in the internal language of the system encoding it usually called representation� Knowledge acquisition as described above is a notoriously slow process and has become known as the �bottle�neck� in the process of developing expert systems� ��� The knowledge acquisition problem for this project initially appeared particularly severe because no human experts in the selection of materials for enantioseparations were available to work on the project� This paper describes an attempt to over� � come this problem by automating the knowledge acquisition process using machine induction� The motivation for using machine learning was the expectation that a machine learning technique might enable a computer to learn how to recommend one or more suitable CSP chiral selectors for a given enantiomer pair� The subject matter of machine learning is the study and computer modelling of learning processes� There are two fundamental reasons for studying learning to understand the pro� cess itself and to provide computers with the ability to learn� ��� One of the results of research aimed at providing computers with the ability to learn has been a num� ber of widely known machine induction algorithms such as ID� ���� ����� Some of these algorithms have been incorporated into commercially available tools such as Ex�Tran �st�Class and the one used in this project DataMariner �see Sec� tion ��� Machine induction algorithms such as that used by DataMariner take as input a set of examples known as the training set and produce as output a set of classi cation rules� These rules are of the form IF description THEN class These rules can then be used to predict the class of previously unseen examples� Each example in the training set represents an example from the domain as a set of attribute values� The same attributes must be used for all the examples� One attribute is the classi er and its values are the classes to which particular examples belong� The other attributes are known as the predicting attributes� The description in the rule antecedent usually comprises conditions on the predicting attributes� In this work the classes were CSP chiral selectors and the predicting attributes were chemical structural features� The aim was to develop a set of classi cation rules that would recommend one or more CSP chiral selectors given particular details of structural features of a given enantiomer pair�� The original nature of this work is illustrated in the remainder of this section by brie�y reviewing previous applications of machine induction to analytical chemistry� �The authors realise that an expert system for enantioseparations by HPLC would need to provide the Users of such a system with more information than just which CSP chiral selector to use� However in this work� the �rst step in the development of such a system� the recommendations were limited to CSP chiral selectors so that the experiments with machine induction would remain tractable� � Only a few references to the application of domain independent machine in� duction algorithms to induce rules for analytical chemistry domains were found in the literature� Two papers describe systems for classifying organic pollutants given their GC�MS data� Both describe the use of commercially available tools that incorporate induction algorithms based on ID� ���� ����� Derde et al� ���� used Ex�Tran to induce classi cation rules� Scott ���� successfully used �st�Class to induce classi cation and identi cation decision trees� Recently Mulholland et al� ���� used C��� an extension of ID� to induce a de� cision tree for chosing a detector when performing ion interaction chromatography� The decision tree was validated in two ways� Firstly a similar tree was generated using only ��� of the data for training and this tree was tested using the other ��� of the data� Secondly by using another test�set which was provided by a domain� expert and comprised �� pertinent examples of the ideal choice of detector as selected by that expert� The validation showed that ��� of the recommendations made by the decision tree were an exact match with the published methods and a further ��� were acceptable to the domain expert in that s�he thought that they would perform well for the given separation� The data used by Mulholland et al� originated from a database of published methods for ion chromatography� The database contained information on almost ���� applications including most of the chromatographic conditions employed� Part of this data was input to the C��� algorithm after being preprocessed� Mul� holland et al� reported that this preprocessing was the most time consuming part of the work� It is widely known within the eld of machine induction that prepro� cessing of data is often necessary� Later sections of this paper describe how the data used in this work was preprocessed� The most famous example of a machine induction system in analytical chemistry is Meta�Dendral� The work on Meta�Dendral was di�erent to the other work in analytical chemistry in that it did not utilise any domain independent induction algorithms� a machine induction system was developed as part of the project� The role of Meta�Dendral was to help a chemist determine the relationship between molecular fragmentations and the structural features of the compounds� Meta� Dendral produced rules which could be used by Dendral an expert system which uses a set of rules to reason about the domain of mass�spectrometry� The quality of the rules generated by Meta�Dendral were assessed by testing them on structures � not in the training set by consulting mass spectroscopists and by comparing them with published rules� The program succeeded in rediscovering known rules of mass� spectrometry that had already been published as well as discovering new rules� Its ability to predict spectra for compounds outside the original sets of instances was impressive� ��� � Experimental This section describes the tool used for the experiments the data input to the tool and the experiments themselves� The tool that was used in this project is called DataMariner �Release ������ ���� ����� It incorporates a rule induction algorithm which can be used to generate rules for membership of classes� The classes must be disjunctive that is membership of classes is mutually exclusive and non�hierarchical� DataMariner induces rules with the following syntax� classname rule no IF clause � clause � � � � THEN conclusion � �probability �� conclusion � �probability �� � � � The rule consequent is an implicit disjunction of clauses where each clause is a conclusion about class membership and has a probability associated with it� The rule antecedent is an implicit conjunction of clauses that is a set of clauses that are implicitly logically ANDed together� Each one of these clauses can only involve one attribute� Thus rules in which there is a disjunction involving two or more attributes are not allowed� A clause of the rule antecedent can specify the value�s� of a discrete� attribute as one of the following � discrete value �eg� detector � uv� disjunction of discrete values �eg� detector � uv OR fluorescence� negation of a discrete value �eg� detector �� uv� DataMariner comprises a number of tools� A description of some of these is given below� �Numeric attributes are allowed but they are outside the scope of this paper� � Merge This can be used to merge values of attributes� Divide Divide can be used to split the data into several training and test les so that a K�fold cross�validation can be performed� Induce This produces a set of rules describing each class in turn where the classes are sorted by the number of examples belonging to each class in descending order� The induction process continues for each class until all the examples that belong to that class are covered by the induced rules� The order of the induced rules describing each class is important� Once the rst rule has been induced for a class then all the examples which are covered by that rule are ignored when inducing the next rule� Thus an example obeys a second induced rule only if it does not obey the rst rule and does obey the second rule� Induce uses an algorithm� developed from the PRISM algorithm ����� The PRISM algorithm is described below� For each class in turn �� For each attribute�value pair calculate the probability that an example which has that value for that attribute belongs to the class� �� Select the attribute�value pair which has the largest probability and cre� ate a subset of the training set comprising all the examples which contain this attribute�value pair� �� Repeat steps � and � for this subset until it contains only examples of the class� The induced rule is a conjunction of all the attribute�value pairs used in creating the homogeneous subset� �� Remove all the examples covered by this rule from the training set� �� Repeat steps � to � until all the examples of the class have been removed� The PRISM algorithm is based on the ID� algorithm but instead of producing a decision tree it produces production rules directly� The major di�erence between ID� and PRISM is that ID� is concerned with nding the attribute which is most relevant whilst PRISM is concerned with nding the attribute� value pair which is most relevant� The problem with nding the attribute �Details of the speci�c algorithm used by Induce are not given because they could not be released by Logica� � which is most relevant is that this attribute may have some values which are irrelevant� Thus PRISM avoids a drawback of ID�� Prune This can be used to prune rules� It examines each clause in each rule starting with the last clause in a rule to test whether a clause signi cantly improves the proportion of examples correctly allocated to the class� If a clause fails this test then it is removed and the preceding clause is tested� If it does not fail then the preceding rule is tested� If all the clauses of a rule are found to make an insigni cant contribution then the whole rule is removed� Prune uses the Fisher one�tailed statistic to decide whether a clause signif� icantly improves the proportion of examples correctly allocated to the class� no domain knowledge is used to support its actions� The level of pruning can be controlled using a parameter known as the prune� level� The level can be regarded as a lter where a high gure implies that more should be retained� Pruning with the prune�level set to �� would remove all of the rules� Pruning with the prune�level set to ���� would not remove any clauses or rules although this would remove redundant conditions� Evaluate The rules induced by DataMariner can be tested using Evaluate� Evaluate uses the induced rule�set to classify some examples and compares the results with the actual classi cations that is those classi cations which are known before the rules are induced� Evaluate generates a variety of other information that guides the data analyst in identifying any problems or omissions in the rules� This information may include for example suggestions on how the values of attributes could be merged� The way in which DataMariner interprets the data given to it can be con� trolled in a number of ways� Some examples of these are described below� Data� Mariner can be instructed to � � ignore one or more attributes and their values� � treat one or more discrete attributes as ordinal types and prevent the gen� eration of disjunctive clauses containing non�contiguous values of these at� tributes� DataMariner treats a discrete variable as nominal unless it is given this instruction� � use a speci ed attribute as the classi er� � � only generate rules for a number of speci ed classes� The data that were input to DataMariner were limited to a subdomain of enantioseparations as follows� � Only analytical separations not preparative ones were considered� � Only enantioseparations by HPLC were considered� � Only the use of CSPs was considered as opposed to the addition of a chiral additive to the mobile phase� � Only successful separations� on commercially�available Pirkle�type CSPs were considered� The data were extracted from chemistry journals and literature obtained from suppliers of CSPs� The data were stored as the values of attributes� One of the attributes was es�name which represented the name of a chiral selector of a CSP� All of the remaining attributes represented instances of chemical features of an enantiomer pair� The chemical features selected and the names that were used for them are shown in Figure �� There are some features which distinguish between substructures where one or more aromatic groups are attached to a functional group and substructures where none are attached to the same type of functional group� The former are referred to as aromatic and the rst letter of the corresponding attribute name is B� The latter are referred to as aliphatic and the rst letter of the corresponding attribute name is R� There were three attributes for each chemical feature�� Each attribute contained a single character which was a digit representing the distance of an occurrence �A separation was judged to be a success if one of the following mutually exclusive conditions were true� The percentage of the separations represented by the data input to DataMariner that satis�ed each of these conditions is shown in parenthesis after each one� �� The separation factor� �� had been recorded and was greater than or equal to ��� � �� � �� The separation factor had not been recorded but resolution� Rs� had and was greater than or equal to ���� � � �� Neither the separation factor or resolution had been recorded but the literature either stated that a separation was a success or illustrated this using a chromatogram� �� � �except the number of chiral centres from the nearest chiral centre in terms of the number of connecting bonds� The three attributes for each feature were numbered � � and � to indicate that they represented the rst second and third closest occurrences respectively� This did not allow for molecules where a feature occurred more than three times a compromise had to be drawn between having a practical number of attributes and allowing for a larger number of instances� Rules were devised to ensure that structural features were represented uni� formly� These rules which are described below were obeyed for all the data that were input to DataMariner� The distance from the chiral centre was the number of connecting bonds between the nearest chiral centre and the atom of the structural feature which was closest to that chiral centre� If there were two or more chiral centres equidistant then one was arbitrarily chosen as the choice was of no consequence� For structural features which were functional groups it was the atom of the functional group itself and not an atom in a connected ring or chain which was closest� For structural features which were a double bond between carbon atoms in an alkyl chain it was whichever one of the two atoms connected by the bond was closest� With the exception of alkyl chains if a structural feature occurred at the chiral centre the distance was considered to be zero� An alkyl chain which started with a carbon atom at the chiral centre was repre� sented as that chain of carbon atoms less the one at the chiral centre the distance distance from the chiral centre being entered as one� Alkyl chains which passed through the chiral centre were conceptually split at the centre and represented as two alkyl chains each one being treated as though it had started there� The alkyl chain attributes represented all alkyl chains regardless of the degree of saturation they did not represent this� Branched chains were conceptually split into the longest straight chain and the side chains originating from it� If any of the side chains were branched then they too were split in the same manner� Thus branched side chains were split recursively until there were none remaining� Each conceptually�formed chain was represented separately� Thus branched chains were represented as a series of substituent straight chains� The way in which these substituent chains were inter�connected was not represented� The following rules were devised for functional groups� If an occurrence of a � functional group was part of a ring as distinct from attached to a ring then it was not represented as a functional group in the database� If an occurrence of a functional group was part of an occurrence of a larger functional group then the occurrence of the smaller group was not represented in the database� If two occur� rences of the same functional group or two occurrences of two di�erent functional groups shared some but not all of the same atoms then both occurrences were represented� Only amides which were derivatives of carboxylic acids in which the OH por� tion of the COOH group had been replaced by NH� �as such or substituted� were represented as amides� Thus amides could take the following forms � RCONH� primary RCONHR� primary RCONR�R�� primary RCONHCOR� secondary RCON�COR��COR�� tertiary An amide was considered to be aromatic if R R� or R�� was an aromatic group� Whenever a NH� �as such or substituted� occurred which was not part of an amide as de ned above it was represented as an amine� An amine was considered to be aromatic if one or more aromatic groups were attached to the nitrogen� Otherwise an amine was considered aliphatic� Once the data had been stored in accordance with these rules experiments were performed� DataMariner was instructed to use the attribute es�name as the classi er for all the experiments that were performed using Induce and Merge� Table � summarises the experiments performed using the tools Induce and Merge� The experiments are identi ed by numbers which correspond to the chronological order in which the experiments were performed� The rst experi� ment that was performed is referred to as test � the second as test � and so on� Tables � and � list the experiments� in such a way that similar ones are grouped �When the experiments were designed the fact that the attributes representing the alkyl chains would never have a value of � was overlooked� Consequently values such as � or at the centre or � that appear in some of the clauses generated by DataMariner that involve the alkyl chain attributes are misleading� However this oversight is of no consequence with respect to the validations performed since both the data used to test and train will not have a value of � for any of the alkyl chain attributes� �� together rather than in chronological order� The di�erence between the orders re�ects the exploratory manner in which DataMariner was used� The purpose of tests � � � and �� was to investigate the e�ect of increasing the number of classes for which DataMariner was instructed to induce rules� Tests � and �� investigated how the induced rules would di�er if DataMariner was instructed to ignore the attributes for the second and third occurrences of chemical features� Tests � �� �� and �� investigated the e�ects of merging the values of the chemical feature attributes� Tests � �� and �� explored whether the values of these attributes should be ordered� Tests �� and �� investigated what the e�ect would be of ordering the values created by merging the original values of the chemical feature attributes� Prune was used on some of the rule�sets induced during the experiments de� scribed above� Prune was used in two ways � �� To remove redundant conditions from rule�sets� This was done by setting the prune�level to ����� Table � indicates for which rule�sets Prune was used in this way by adding the extension �p��� to the name of the experiments concerned� �� To investigate the e�ects of pruning the rule�sets� Most of both the pruned and unpruned rule�sets were tested using Evaluate� All the examples from the example� le had to be used for training to ensure that the accuracy of the induced rules would be acceptable there were ��� examples belonging to �� classes giving an example to class ratio of just �� �� Since none of the examples could be used exclusively for testing Evaluate could only be used to calculate the classi cation success�rates of the rule�sets on their training sets and to cross�validate the rule�sets� The type of cross�validation performed was a K�fold cross�validation where K was equal to ten� Table � shows some of the statistics that were calculated when the le used for testing was identical to that which had been used for training and Table � shows the the statistics that were estimated using cross�validation� In addition to being cross�validated the rule�set induced during test �� was manually validated� That is a paper exercise was used rather than Evaluate� This exercise will be referred to as the external validation because the rule�set was tested on �� enantioseparations that were not stored in the example��le used �� by DataMariner� These enantioseparations were reported in sources similar to those from which the data in the example� le originated� The choice of enantiomer pairs was restricted to those which had been separated on one of the CSP chiral selectors for which rules had been generated by DataMariner� The external validation compared for some enantioseparations not stored in the example��le the CSP chiral selectors recommended by the rule�set� induced during test �� with the choice of selector reported in the literature� The aim of the external validation was to prove that the cross�validation correctly simulated the e�ects of testing with unseen data� � Results and Discussion In tests ��� DataMariner induced rules whose clauses speci ed not only whether a particular occurrence of a chemical feature was present and if so how far it was from the chiral centre but also whether the occurrence was the rst second or third closest occurrence of that chemical feature� This author believes that in some cases it may not matter whether a chemical feature is the closest second closest or third closest occurrence of that feature as long as the feature is present at a particular distance or within a range of distance values� However DataMariner could not have induced rules that represented this because it could not induce rules in which there was a disjunction of attributes� For example DataMariner could not have induced a clause such as cooh� OR cooh� OR cooh� � � In tests ���� DataMariner was instructed to ignore all the attributes that represented the second and third occurrences so that rules would be induced that reasoned about the presence of the nearest occurrences only� The e�ects of ignoring the second and third occurrences can be analysed by comparing tests � and � as these were identical in every other respect� When the second and third occurrences were ignored the number of rules increased very slightly whilst the classi cation success�rate on the training set remained at ����� This suggested that providing DataMariner with data on the second and third occurrences did not result in �Only the recommendations of the �rst of the rules in the rule�set that could �re were considered� �� better rules� Consequently DataMariner was instructed to ignore the second and third occurrences in all the remaining experiments� The e�ects of using di�erent ordinal types are shown by tests � � and ��� These tests were identical except for the data types used for the chemical feature attributes� The attributes had discrete values in all three tests but in tests � and �� the values were ordered� The use of the ordinal types reduced the number of rules from �� in test � to � in both tests � and ��� The average number of clauses per rule rose from three in test � to �� in tests � and ��� In test � the order was not present � � � � � � �� The order in test �� was the same except that not present came after �� This makes more chemical sense because not present can be considered to be the case where a chemical feature is an in nite number of bonds away� Tests � and �� suggested that some of the values of the chemical feature attributes should be merged� Consider the rst rule induced during test � which is shown in Figure � � All the clauses that are disjunctions include the values � and �� This was re�ected across the rest of the rule�set� �� out of the �� disjunctions in the rule�set included these values which suggests that they should be merged� Test �� suggested that the values � and � should be merged and that the values � � � and � should be merged� the rst rule from test �� is shown in Figure � and illustrates this� There were �� disjunctions in the rule�set for test��� Eight of these suggested that � and � should be merged and �� that � � � and � should be merged� The e�ects of merging the values of the chemical feature attributes can be analysed by comparing tests �� �� �� and ��� These tests were identical except for the way in which values were merged� The rules that were induced in test �� in which no values were merged are very speci c� They include many precise statements about the distance of chemical features from the chiral centre� Consider the rst rule that was induced which is shown in Figure �� The clauses involving bx� state that bx� should not equal � � or �� The clause involving boh� states that boh� should equal not present or �� The rules generated during test �� seem chemically implausible because they are very precise about the distances� �This rule� and the one shown in Figure �� has a redundant clause� cen � � serves no purpose because it appears after another clause cen � � OR �� Such redundancy could have been removed see Section �� but this was would have been irrelevant to the purpose of the experiment� �� In test �� the values � � � � � � � of all the chemical feature attributes were merged to the value present� In test �� the values � and � were merged to � or � bonds away and � � � and � were merged to more than ve bonds away� In test �� the merges performed in test �� were repeated and in addition the values � and � were merged to at the centre or �� The accuracies calculated by cross�validation for all four tests were indistinguishable but the number of rules for the classes did vary� Merging all the values to the value present increased the number of rules from �� to �� that is by ���� Merging some of the values led to a slight increase �� in test �� and �� in test ��� The e�ects of merging the values can also be seen by comparing tests �� and ��� Test �� was similar to test ��� identical merges were performed in both but in test �� the values were ordered after they were merged� The e�ects caused by the merge used in test �� can be considered in isolation by comparing the results of tests �� and �� given the merges performed the ordinal types used in these tests are e�ectively the same� Figures � and � show two comparable rules from tests �� and �� respectively� These gures show that merging values results in more general rules� Consider the respective clauses for the attribute rconh�� In test �� the clause is as follows� rconh� � � OR OR not�present In test �� this is generalised to the following� rconh� � more�than�five�bonds�away OR not�present Table � lists the results of the cross�validation� It shows that the accuracies were all more than ten times greater than the accuracy that would result from choosing one of the selectors at random� The tests that were cross�validated di�ered only in the merges that were per� formed and the ordinal types that were used� Table � shows that for any two of the tests that were cross�validated �pA � �pA �� �pB � �pB where A is the test with largest �p value and B is the other test� Hence the estimates of accuracy for these tests are indistinguishable the values for �p are too This is explained later in this section� �� close given the values of �p� This suggests that using merges or ordinal values did not a�ect the accuracy of the resulting rules� Table � shows some of the results of the external validation performed on the rule�set induced during test ��� It indicates the extent of the agreement on the choice of CSP chiral selector between the literature and the rule�set induced during test ��� Tables � to �� list the names and structures of the enantiomer pairs used in the external validation and show the diverse range of structures used� Only for two of the �� enantiomer pairs ���� did the rule�set fail to recommend the choice of CSP chiral selector reported in the literature� The two enantiomer pairs concerned are Labetolol and N���FMOC� ��benzoylglycine �N�phenylamide�� In both these cases the rule�set failed to recommend any CSP chiral selector� The choice of CSP chiral selector reported in the literature was either the rst or second choice recommendation of the rule�set for �� of the �� enantiomer pairs ������ The choice of selector reported in the literature was the rst choice of the rule�set for �� of the �� enantiomer pairs ������ The accuracy calculated using just the rst choice of the rule�set is most comparable to the cross�validation result for test �� since Evaluate calculates accuracy by assigning each example to the class with the highest probability associated with it amongst all the rules that can re� The cross�validation result for test �� was ��� � � �� and the accuracy calculated during the external validation using just the rst choice was ���� Hence the cross�validation and external validation are mutually corroborative the di�erence between the upper limit of the cross�validation result and the external validation result is only ��� The analysis of the experiments with Prune was di�cult� The developers of DataMariner acknowledge that a possible consequence of pruning is that excep� tion relationships that are correct but rare can be eliminated� They recommend that pruned and unpruned rules should always be checked to con rm that no valu� able information has been lost ����� It is not easy to provide a chemical justi cation for the rules that were induced as part of this work by looking at the rules them� selves� Consequently it is impossible to check that Prune did not result in the loss of valuable information� Prune can be used to remove clauses or rules that are induced as a result of noise in an example� le ����� However the rule�sets could not have been improved signi cantly by Prune the example� le was carefully and meticulously prepared� �� Prune has as great a potential to have an adverse e�ect as it does to have a bene cial one because it relies solely upon a statistical test to support its actions� it can not distinguish between a clause whose presence is due to noise and one whose presence is due to an exceptional relationship which is correct but rare� Recall that this paper is concerned with the knowledge acquisition phase of de� veloping an expert system for enantioseparations rather than the implementation of such a system� Therefore a detailed discussion of the phases that must follow the knowledge acquisition phase is consigned to further work� the remainder of this section brie�y indicates how the optimal rule�set induced during test �� could be used� The authors believe that the con�ict resolution strategy ��� that follows should be adopted given the induction algorithm used by Induce �see Section ��� �� Try to re each rule in turn until a rule res� �� Let the rst choice recommendation of the rule�set be the CSP chiral selector in the consequent of the rule that red which has the highest probability associated with it� �� If the consequent of the rule that red is a disjunction of CSP chiral selectors then let the second choice recommendation be the selector in the consequent that has the second highest probability associated with it� Let the third choice be the one with the third highest probability and so on� Such a strategy could be used to generate an ordered list of recommended CSP chiral selectors whenever the consequent of the rule that res is a disjunction� This would suit analysts as they would then be free to either try each selector in the list in turn starting with the rst choice of the rule�set or to choose selectors from the list using other criteria such as cost or availability in their laboratory� � Conclusions The optimal rule�set must � � have rules for membership of all the classes that is CSP chiral selectors� � be induced using an ordinal type which re�ects the inherent order in the distance values and allows not present to be considered as the case where a chemical feature is an in nite number of bonds away� �� Rule�sets induced when such an ordinal type is used are smaller and have rules where the average number of clauses is much larger than the corre� sponding rule�sets which are induced when ordinal types are not used but the experimental conditions are otherwise identical� � be induced using the following merges that were suggested by Evaluate� � and � merged to at the centre or � � and � merged to � or � � � � and � merged to more than ve� Unless merges are performed the induced rules include clauses that are too precise about the distances� The merges suggested by Evaluate make chem� ical sense and result in more general and plausible rules� The rule�set induced during test �� ful lls these requirements and so is the optimal rule�set� DataMariner was successfully used to induce and validate rules that rec� ommended particular CSP chiral selectors based on the structural features of an enantiomer pair� Although it is not easy to provide a chemical justi cation for the rules by looking at them the results suggest that they have a high degree of accu� racy� The cross�validation performed on the optimal rule�set induced suggests that this rule�set would recommend as its rst choice a correct CSP chiral selector for ��� � � �� of enantiomer�pairs that can be separated on Pirkle�type CSPs� The ex� ternal validation which used test data that had not been input to DataMariner supported the results of the cross�validation� The accuracy of the optimal rule�set is more than ten times greater than the accuracy that would result from choosing one of the selectors at random� The external validation suggests that either the rst or second choice recommendation of the optimal rule�set would be correct for ��� of enantiomer pairs that can be separated on Pirkle�type CSPs� � Acknowledgements The funding was provided by EPSRC under the remit of the Total Technology programme and by Zeneca Pharmaceuticals� R� Dallaway and I�T�Nabney of Logica Cambridge Ltd� provided helpful advice on the use of DataMariner� C� Bryant is grateful for this and the hospitality shown to him by all at Logica Cambridge Ltd� �� G�V� Conroy from the Computation Department at UMIST made some valu� able comments on the machine induction aspects of this work� D�J� Williams from the Chemistry Department at UMIST prepared the dia� grams of the chemical structures� � References ��� D�R� Taylor and K� Maher Chiral Separations by High�Performance Liquid Chromatography� Journal of Chromatographic Science �� ������ ������ ��� C� Roussel and P� Piras CHIRBASE A Molecular Database for Storage and Retrieval of Chromatographic Chiral Separations� Pure � Applied Chem� istry �� ������ �������� ��� B� Koppenhoefer A� Nothdurft J� Pierrot�Sanders P� Piras C� Popescu C� Roussel M� Stiebler and U� Trettin CHIRBASE a Graphical Molecular Database on the Separation of Enantiomers by Liquid� Supercritical Fluid� and Gas Chromatography� Chirality � ������ �������� ��� B� Koppenhoefer R� Graf H� Holzschuh A� Nothdurft U� Trettin P� Piras and C� Roussel CHIRBASE a Molecular Database for the Separation of Enantiomers by Chromatography� Journal of Chromatography ��� ������ �������� ��� S�T� Stau�er� Expert System Shells in Chemistry� CHIRULE� a Chiral Chromatographic Column Selection System using Similarity Searching and Personal Construct Theory� PhD Thesis� Virginia Polytech Ins� State Univ� USA� ������ ��� P� Jackson Introduction to Expert Systems� �nd Ed� Addison�Wesley ����� ��� C�H� Bryant A�E� Adam D�R� Taylor and R�C� Rowe A Review of Expert Systems for Chromatography� Analytica Chimica Acta ��� ������ �������� ��� D� Diaper Knowledge Elicitation� Principles� Techniques and Applications� Ellis Horwood ����� ��� S� Kocabas A Review of Learning� Knowledge Engineering Review � ������ �������� ���� J�R� Quinlan Discovering Rules from Large Collections of Examples a Case Study� in D� Michie� �Ed� Expert Systems in the Micro Electronic Age� Edinburgh University Press� Edinburgh ����� ���� J�R� Quinlan Learning E�cient Classi cation Procedures and their Appli� cation to Chess End Games� in R�S� Michalski� J�G� Carbonell� and T�M� Mitchell� �Ed�s Machine Learning An Arti cial Intelligence Approach� Palo Alto� Tioga ����� � ���� M� Derde L� Buydens C� Guns and D�L� Massart Comparison of Rule� Building Expert Systems with Pattern Recognition for the Classi cation of Analytical Data� Analytical Chemistry �� ������ ���������� ���� D�R� Scott Classi cation and Identi cation of Mass Spectra of Toxic Com� pounds with an Inductive Rule�Building Expert System and Information Theory� Analytica Chimica Acta ��� ������ �������� ���� M� Mulholland D�B� Hibbert P�R� Haddad C� Sammut Application of the C��� Classi er to Building an Expert System for Ion Chromatography� Chemometrics and Intelligent Laboratory Systems �� ������ ������� ���� Logica UK Limited� DataMariner� User Manual� Version B ����� ���� I�T� Nabney and O� Grasl Rule Induction for Data Exploration� in Pro� ceedings of Avignon �� Expert systems and their applications � ������ �������� ���� J� Cendrowska PRISM An Algorithm for Inducing Modular Rules� Inter� national Journal of Man�Machine Studies �� ������ �������� ���� W�H� Pirkle T�C� Pochapsky G�S� Mahler and R�E� Field Chromato� graphic Separation of the Enantiomers of ��Carboalkoxyindolines and N� Aryl���amino Esters on Chiral Stationary Phases Derived from N��� �� Dinitrobenzoyl����amino Acids� Journal of Chromatography ��� ������ ��� ��� ���� I�W� Wainer and M�C� Alembik Steric and Electronic E�ects in the Res� olution of Enantiomeric Amides on a Commercially Available Pirkle�Type High�Performance Liquid Chromatographic Chiral Stationary Phase� Jour� nal of Chromatography ��� ������ ������ ���� L�E� Weaner and D�C� Hoerr Separation of Fatty Acid Ester and Amide Enantiomers by High�Performance Liquid Chromatography on Chiral Sta� tionary Phases� Journal of Chromatography ��� ������ �������� ���� R� Dernoncour and R� Azerad High Performance Liquid Chromatographic Separation of the Enantiomers of Substituted ��Aryloxypropionic Acid Methyl Esters� Journal of Chromatography ��� ������ �������� ���� A� Berthod H�L� Jin A�M� Stalcup and D�W� Armstrong Interactions of Chiral Molecules With an �R��N��� ��Dinitrobenzoyl� Phenylglycine HPLC Stationary Phase� Chirality � ������ ������ �� ���� W�H� Pirkle and J�E� McCune Separation of the Enantiomers of N� Protected ��amino Acids as Anilide and � ��dimethylanilide Derivatives� Journal of Chromatography ��� ������ �������� ���� Phenomenex Ltd� U�K� The Arsenal Heapy Street Maccles eld Cheshire SK�� �JB� Table � Summary of experiments that used the Induce and Merge tools of DataMariner� Table � Results of experiments that used the Induce Merge and Prune Tools of DataMariner� Statistics calculated by Evaluate for each rule�set as a whole� �The le used for testing was identical to that which had been used for training�� Table � Results of cross�validation� Statistics on the accuracy with which the rule�sets induced from the training les classify examples from the test les� Table � Results of external validation� Number of occurrences of di�erent rankings� Table � Some of the data that were used in the external validation� Some of the enantiomer pairs for which �R��N��� ��dinitrobenzoyl�phenylglycine was both the �rst choice recommendation of the optimal rule�set and the chiral selector used in the separations reported in the literature� Table � Some of the data that were used in the external validation� Some of the enantiomer pairs for which �R��N��� ��dinitrobenzoyl�phenylglycine was both the �rst choice recommendation of the optimal rule�set and the chiral selector used in the separations reported in the literature� Table � Some of the data that were used in the external validation� Some of the enantiomer pairs for which �R��N��� ��dinitrobenzoyl�phenylglycine was both the �rst choice recommendation of the optimal rule�set and the chiral selector used in the separations reported in the literature� Table � Some of the data that were used in the external validation� The enantiomer pairs for which �S��N��� ��dinitrobenzoyl�leucine was both the second choice rec� ommendation of the optimal rule�set and the chiral selector used in the separations reported in the literature� Table � Some of the data that were used in the external validation� The enan� tiomer pairs for which the chiral selector used in the separations reported in the literature was neither the rst or second choice recommendation of the optimal rule�set� Table �� Some of the data that were used in the external validation� The enan� tiomer pairs for which the optimal rule�set did not make any recommendations� Figure � The chemical features of enantiomer pairs that were input to Data� Mariner and the names that were used for them� Figure � One of the rules induced during test �� Figure � One of the rules induced during test ��� Figure � One of the rules induced during test ��� Figure � One of the rules induced during test ��� Figure � One of the rules induced during test ��� � Test No� of �nd � �rd Values Merged Ordinal chiral occurrences values selectors of chemical All a Some b Some c None for which features of rules were enantiomer induced pairs ignored � � no no no no yes none � � no no no no yes none � � no no no no yes none � � yes no no no yes none �� �� yes no no no yes none � � yes no no no yes yesd �� � yes no no no yes yese �� �� yes no no no yes yese � � yes yes no no no none �� �� yes yes no no no none �� � yes yes no no no none �� �� yes no yes no no none �� �� yes no yes no no yesf �� �� yes no no yes no none �� �� yes no no yes no yesg aThe values �� �� � � � � � of the chemical feature attributes were merged to the value present� bTwo merges were performed on the chemical feature attributes� and � were merged to or � bonds away �� �� �� and � were merged to more than �ve bonds away� cThree merges were performed on the chemical feature attributes� � and � were merged to at the centre or � and � were merged to or � �� �� �� and � were merged to more than �ve� dThe following order was speci�ed for the values of each of the chemical feature attributes� not present� �� �� �� � � � �� eThe following order was speci�ed for the values of each of the chemical feature attributes� �� �� �� � � � �� not present� fThe following order was speci�ed for the values of each of the chemical feature attributes� �� �� �� �� or � bonds away� more than �ve bonds away� not present� gThe following order was speci�ed for the values of each of the chemical feature attributes� at the centre or �� �� �� or � bonds away� more than �ve bonds away� not present� Table �� Summary of experiments that used the Induce and Merge tools of Data� Mariner Test No� of Average Overall No� of No� of Rules no� of accuracy miscla� uncla� clauses � ssi ed ssi ed per rule rings eg�s � �� � ��� � ��� � �� � ��� � ��� ���p��� �� � �� � �� � � �� �� �� �� �� � �� �� �� �� ���p��� �� � �� �� �� ��p��� �� � ��� � ��� ���p��� �� � �� � �� ���p��� �� � �� � �� ���p��� �� � �� �� �� ���p��� �� � �� � �� ���p��� �� � �� �� �� Table �� Results of experiments that used the Induce� Merge and Prune Tools of DataMariner Statistics calculated by Evaluate for each rule�set as a whole �The �le used for testing was identical to that which had been used for training � Test Accuracy � Mean Standard Error �� �� � �� �� � �� �� � �� �� � �� �� � �� �� � �� �� � Table �� Results of cross�validation Statistics on the accuracy with which the rule�sets induced from the training �les classify examples from the test �les Ranka Enantiomer Pairs with this Ranking Number �Number x ���� � �� � �� �� � � �� � � � � � � � � � � � � � � � � No rules red � � aRank that was assigned by the rule�set induced during test �� for the choice of CSP chiral selector reported in the literature� Table �� Results of external validation Number of occurrences of di�erent rankings Enantiomer Pair Refa Name Structure ethyl N�phenyl phenylglycine ���� N O O H ���ethoxycarbonyl�indoline ���� N H H O O N�phenyl���methylheptanamide ���� N O H N����phenylethyl����naphthylamide ���� O N H N����naphthoyl���aminohexan���ol ���� O N HO CH3 ( CH 2 ) 3 H N����naphthoyl����methylbut���ylamine ���� N H O aLiterature reference for the separation that was reported in the literature� Table �� Some of the data that were used in the external validation Some of the enan� tiomer pairs for which �R��N������dinitrobenzoyl�phenylglycine was both the �rst choice recommendation of the optimal rule�set and the chiral selector used in the separations reported in the literature Enantiomer Pair Refa Name Structure benzoylmethyl ��tetradecylglycidate ���� TDGA ester derivative O O O O CH 3 ( CH 2 ) 1 2 CH 2 methyl ��hexylglycidate ���� TDGA analogue O O O n - C 6 H 1 3 methyl ���� ��dichlorophenoxy�propanoate ���� OC l C l O O methyl �����methylphenoxy�propanoate ���� O O O methyl �����naphthoxy�propanoate ���� O O O aLiterature reference for the separation that was reported in the literature� Table �� Some of the data that were used in the external validation Some of the enan� tiomer pairs for which �R��N������dinitrobenzoyl�phenylglycine was both the �rst choice recommendation of the optimal rule�set and the chiral selector used in the separations reported in the literature Enantiomer Pair Refa Name Structure N�������naphthyl��ethyl� � ��dichloroacetamide ���� N O C l C l H N�acetyl �����naphthyl�ethylamine ���� N O H N�chloroacetyl���aminoindane ���� H N H O C l N��acetylmethionine �N�����naphthyl�amide� ���� N O HN H O S N�������naphthyl�ethyl� acetamide ���� N O H aLiterature reference for the separation that was reported in the literature� Table �� Some of the data that were used in the external validation Some of the enan� tiomer pairs for which �R��N������dinitrobenzoyl�phenylglycine was both the �rst choice recommendation of the optimal rule�set and the chiral selector used in the separations reported in the literature Enantiomer Pair Refa Name Structure N���CBZ�alanine �N�phenylamide� ���� O O N H H O N N���BOC�alanine �N� � ��dimethylphenylamide� ���� N ON O O H H N���FMOC�alanine �N� � ��dimethylphenylamide� ���� O O N O N H H aLiterature reference for the separation that was reported in the literature� Table �� Some of the data that were used in the external validation The enantiomer pairs for which �S��N������dinitrobenzoyl�leucine was both the second choice recommendation of the optimal rule�set and the chiral selector used in the separations reported in the literature Enantiomer Pair Refa Name Structure Metoprolol ���� O O OH N H Bepridil ���� N O N Timolol ���� O N S NN O HO N H aLiterature reference for the separation that was reported in the literature� Table � Some of the data that were used in the external validation The enantiomer pairs for which the chiral selector used in the separations reported in the literature was neither the �rst or second choice recommendation of the optimal rule�set Enantiomer Pair Refa Name Structure N���FMOC� ��benzoylglycine �N�phenylamide� ���� O O N H O O N H Labetolol ���� O H 2 N HO OH N H aLiterature reference for the separation that was reported in the literature� Table � � Some of the data that were used in the external validation The enantiomer pairs for which the optimal rule�set did not make any recommendations Chemical Feature Name Chemical Feature Name number of chiral centres Cen alkyl chain of length � C aliphatic OH Roh alkyl chain of length � Cc aromatic OH Boh alkyl chain of length � Ccc COOH Cooh alkyl chain of length � Cccc ester Ester alkyl chain of length � � C���c aldehyde Ald alicyclic � membered ring Rg� ketone Ket alicyclic membered ring Rg aliphatic amide Rconh alicyclic � membered ring Rg� aromatic amide Bconh other alicyclic ring Rg aliphatic amine Rnh aromatic membered ring Bg aromatic amine Bnh aromatic � membered ring Bg� nitro No� other aromatic ring Bg cyanide�nitrile Cn bicyclic ring Bic thio Rsr tricyclic ring Tri sulphinyl Rsor polycyclic ring Ply sulphonyl Rso�r hetero N Nhe aliphatic X Rx hetero O Ohe aromatic X Bx hetero S She ether Ror other hetero atom �he carbon carbon double bond Cdbc Figure �� The chemical features of enantiomer pairs that were input to DataMariner and the names that were used for them �R� N ��� dinitrobenzoyl�phenylglycine rule � IF rnh� � not�present OR � OR � no��� � not�present cdbc�� � not�present OR � OR � est� � not�present OR � OR � ket� � not�present OR � OR � OR � ald� � not�present nhe� � not�present OR � OR � OR � OR � rso�r�� � not�present she� � not�present cooh� � not�present OR � OR � rx� � not�present OR � OR � OR � cc�� � not�present OR � OR � OR � OR � bx� � not�present OR � OR � OR � OR � OR � OR OR � cen � � OR � c�� � not�present OR � OR � OR � OR � OR � OR cen � � rg��� � not�present OR � OR � cn� � not�present THEN es�name � �R� N ��� dinitrobenzoyl�phenylglycine ��� �� es�name � �S� N ��� dinitrobenzoyl�leucine ������ es�name � �R� N � �alpha naphthyl�ethylaminocarbonyl �S�indoline � carboxylic�acid ������ Figure �� One of the rules induced during test �R� N ��� dinitrobenzoyl�phenylglycine rule � IF no��� � not�present boh� � � OR � OR � OR OR not�present ket� � � OR � OR � OR OR � OR � OR � OR OR not�present bx� � � OR OR � OR � OR � OR OR not�present rso�r�� � not�present she� � not�present cen � � OR � bg �� � � OR � OR � OR OR � OR � OR � OR OR not�present cen � � cdbc�� � not�present cn� � not�present THEN es�name � �R� N ��� dinitrobenzoyl�phenylglycine ������ es�name � �S� N ��� dinitrobenzoyl�leucine ������ es�name � �R� N � �alpha naphthyl�ethylaminocarbonyl �S�indoline � carboxylic�acid ������ Figure �� One of the rules induced during test � �R� N ��� dinitrobenzoyl�phenylglycine rule � IF no��� � not�present ket� � not�present OR � cooh� � not�present cdbc�� � not�present OR � boh� � not�present OR � rnh� � not�present ohe� � not�present OR � bic� �� � she� � not�present bx� �� bx� �� � cc�� �� � nhe� �� � c�� �� � cn� � not�present bx� �� rconh� � not�present c���c�� �� � bic� �� � cen � � bg��� �� not�present ror� �� � cc�� �� � THEN es�name � �R� N ��� dinitrobenzoyl�phenylglycine ������ Figure �� One of the rules induced during test �� �R� N � �alpha naphthyl�ethylaminocarbonyl �S�indoline � carboxylic�acid rule � IF bic� � � OR � OR OR � OR � OR � OR OR not�present est� � not�present nhe� � � OR � OR � OR OR not�present cooh� � not�present bconh� � not�present rconh� � � OR OR not�present cdbc�� � not�present tri� � not�present ply� � not�present ror� � � OR � OR � OR OR � OR � OR � OR OR not�present cen � � THEN es�name � �R� N � �alpha naphthyl�ethylaminocarbonyl �S�indoline � carboxylic�acid ������ Figure �� One of the rules induced during test �� �R� N � �alpha naphthyl�ethylaminocarbonyl �S�indoline � carboxylic�acid rule � IF bic� � � OR ��or� �bonds�away OR more�than� �bonds�away OR not�present est� � not�present nhe� � more�than� �bonds�away OR not�present cooh� � not�present bconh� � not�present rconh� � more�than� �bonds�away OR not�present cdbc�� � not�present tri� � not�present ply� � not�present ror� � � OR � OR ��or� �bonds�away OR more�than� �bonds�away OR not�present cen � � THEN es�name � �R� N � �alpha naphthyl�ethylaminocarbonyl �S�indoline � carboxylic�acid ������ Figure �� One of the rules induced during test �� 
work_pvqnj77vmvhlxio2xevwsc3you	----	doi:10.1016/j.compchemeng.2008.10.006 L J a b a A R R 1 A A K H C O P 1 i m a p i n b T i r t w h o m o V e p t t i 0 d Computers and Chemical Engineering 33 (2009) 371–378 Contents lists available at ScienceDirect Computers and Chemical Engineering j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / c o m p c h e m e n g earning HAZOP expert system by case-based reasoning and ontology insong Zhao a,∗, Lin Cui b, Lihua Zhao b, Tong Qiu a, Bingzhen Chen a Department of Chemical Engineering, Tsinghua University, Beijing, 100084, China College of Information Science and Technology, Beijing University of Chemical Technology, Beijing, 100029, China r t i c l e i n f o rticle history: eceived 1 November 2007 eceived in revised form 3 September 2008 a b s t r a c t To improve the learning capability of HAZOP expert systems, a new learning HAZOP expert system called PetroHAZOP has been developed based on the integration of case-based reasoning (CBR) and ontology that can help automate “non-routine” HAZOP analysis. PetroHAZOP consists of four modules including case base module, CBR engine module, knowledge maintenance module and user graphical interface ccepted 4 October 2008 vailable online 22 October 2008 eywords: AZOP ase-based reasoning module. Within the case base, HAZOP analysis knowledge is represented as cases which are organized with a hierarchical structure. Similarity-based case retrieval algorithm is also depicted to find the closest- matching cases. In order to enhance the case retrieval, a new set of ontologies for CBR-based HAZOP analysis is created by integration of existing ontologies reported in literature. Finally the application of PetroHAZOP is demonstrated by two case studies of industrial processes. t “ a w l t a H m e u s p H t t i s r ntology rocess safety . Introduction Safety is an important issue in process design and operation n the chemical process industry (CPI). It is even more critical for odern chemical manufacturing processes which are often oper- ted under extreme conditions to achieve maximum economic rofit, or have to undergo changes of customer demands. The mportance of safety analysis in process operation is well recog- ized after occurrence of several tragic accidents that could have een avoided if adequate process safety analysis had been done. o ensure safe operation, process hazard analysis (PHA) is very mportant to proactively identify the potential safety problems and ecommend feasible mitigation actions. Among the available PHA echniques, hazard and operability (HAZOP) analysis is the most idely used one in the CPI. HAZOP analysis done by human teams, owever, has the following shortcomings: time consuming, labori- us, expensive and inconsistent. To solve these problems, various odel and/or rule-based HAZOP expert systems have been devel- ped during the last decade, which was respectively reviewed by enkatasubramanian, Zhao, and Viswanathan (2000) and McCoy t al. (1999). These systems, however, can only address “routine” or rocess-generic HAZOP analysis. “Routine” HAZOP analysis means hat its reasoning logic can be applied to different processes while he “non-routine” HAZOP analysis means that its reasoning logic s process specific or plant specific. Generally analysis of devia- ∗ Corresponding author. Tel.: +86 10 62783109. E-mail address: jinsongzhao@tsinghua.edu.cn (J. Zhao). i w c v a p a f 098-1354/$ – see front matter © 2008 Elsevier Ltd. All rights reserved. oi:10.1016/j.compchemeng.2008.10.006 © 2008 Elsevier Ltd. All rights reserved. ions generated by using guidewords “other than”, “as well as” and part of” are “non-routine”. As a result, these kinds of deviations re hardly addressed in literature about HAZOP expert systems. In the CPI, “routine” HAZOP analysis roughly occupies 60–80% hile “non-routine” HAZOP analysis occupies 20–40%. Due to the ack of self-learning capability in existing HAZOP experts systems, he knowledge of “non-routine analysis” can be hardly formulized nd reused for similar chemical processes, and the “non-routine” AZOP analysis still needs to be addressed by human experts. To evaluate the output quality of the signed directed graph (SDG) odel based HAZOP expert system HAZID developed by McCoy t al. (1999), five industrial plant systems which had not been sed during the model development stage were selected as a test et (McCoy, Wakeman, Larkin, Chung, & Rushton, 2000). The out- ut of HAZID was compared against the results of conventional AZOP study which was done by human teams. Table 1 shows he test results which are quite interesting. According to Table 1, he percentage of scenarios identified by conventional HAZOP also dentified by HAZID ranged from 33% to 60%, and the percentage of cenarios identified by HAZID which were judged to be corrected anged from 33% to 83%. Moreover, the percentage of scenarios dentified by HAZID which were judged to be correct and of interest as much lower, ranging from 9.5% to 29%. In other words, severe ompleteness and correctness issues existed in HAZID. The unfa- orable performance of HAZID was attributed to a few factors such s the quality of the unit model, lack of fluid property data, the lant complexity and human judgment variance. In fact, there is nother more important factor that was unrecognized and that actor is knowledge representation limitation. The “non-routine” http://www.sciencedirect.com/science/journal/00981354 http://www.elsevier.com/locate/compchemeng mailto:jinsongzhao@tsinghua.edu.cn dx.doi.org/10.1016/j.compchemeng.2008.10.006 bahamut 下划线 bahamut 下划线 372 J. Zhao et al. / Computers and Chemical Engineering 33 (2009) 371–378 Table 1 Summary analysis of trial results of HAZID (McCoy et al., 2000). Plant systems Absorber system �-trichloroethane system Propane rectification system Benzene storage Separation system Scenarios identified by conventional HAZOP also identified by HAZID (%) 36 33 60 50 53 Scenarios identified by HAZID which were judged to be correct (%) 49 33 69 83 53 Scenarios identified by HAZID which were judged to be correct and 9.5 29 24 27 N/A P H T k c e d 2 a 1 t s & i m t o e b m a I e t w c 5 2 t H t n t “ a H P H s r t f b a p F t w i a P r c k p t time, effort and money involved in performing HAZOP, there exists considerable incentive to develop intelligent systems for automat- ing the process hazards analysis of chemical process plants. An intelligent system can reduce the time, effort and expense involved of interest (%) rotections identified by HAZID which were judged to be correct (%) 9.5 29 AZOP analysis could not be represented by the existing models. herefore, it is clear that there is much room for improvement in nowledge representation for HAZOP expert systems. It is worth noting that consistence and completeness are criti- al in HAZOP analysis because neglect of any potential hazard may ven result in disasters. Investigation results of past industrial acci- ents, e.g. the tragic BP Texas city plant accident occurred in March 005, have proved that poor quality of PHA is a major root cause of ccidents occurred in the CPI. Recently case-based reasoning (CBR) technology (Kolodner, 993; Aha, 1998) has been integrated into HAZOP automation echnology by researchers at Purdue University to enhance the elf-learning capability of HAZOP expert systems (Zhao, Bhushan, Venkatasubramanian, 2005). However, the case-based reason- ng they proposed aimed to facilitate modification of the existing odels and creation of new models based on the knowledge in he existing models. The “non-routine” HAZOP analysis still relied n the human team. To solve this problem, a new learning HAZOP xpert system called PetroHAZOP has been developed in this paper ased on the integration of CBR and ontology that can help auto- ate both “routine” and “non-routine” HAZOP analysis. This article is organized as follows. In Section 2, HAZOP analysis nd related work on the automation of HAZOP are briefly described. n Section 3, the integrated methodology for HAZOP analysis is xplained. Section 4 contains two industrial application examples hat illustrate how the proposed HAZOP expert system can help ith improvement of “routine” and “non-routine” analysis. Finally, ontributions of this work are summarized and discussed in Section . . HAZOP HAZOP was firstly introduced by ICI (Imperial Chemical Indus- ries, UK) for identifying hazards in chemical plants in 1960s. AZOP study is accomplished by a HAZOP team through a collec- ive brainstorming effort that stimulates creativity and brings about ew ideas of the potential hazards including their cause–effect rela- ionships. Generally the chemical process is divided into sessions called analysis nodes” before study. Then meaningful deviations in every nalysis nodes are generated by combining process parameters and AZOP guidewords including MORE OF, LESS OF, NONE, REVERSE, ART OF, AS WELL AS and OTHER THAN. For each deviation, the AZOP team has to identify all of its credible causes and all of pos- ible adverse consequences. Once the causes and consequences are ecorded, the team has to list the existing safeguards for the iden- ified hazards and give necessary recommendations accordingly or hazard mitigation if the required risk level cannot be achieved N/A 77 N/A y the safeguards. The process is repeated deviation by deviation nd node by node until the analysis of the whole process is com- leted. The conventional HAZOP study procedure is presented in ig. 1. To complete the HAZOP analysis of a typical chemical process, it akes about 1–8 weeks for a HAZOP team with 4–8 members. It is idely accepted that HAZOP analysis is an extremely time consum- ng and effort consuming process. An estimated including direct nd indirect costs $5 billion is spent annually by the CPI to perform HAs and related activities. The estimated cost of process hazards eviews is about 1% of sales or about 10% of profits for a big chemical ompany. Moreover, the quality of HAZOP analysis depends on the nowledge and experience of the HAZOP team. Therefore, incom- leteness and inconsistence usually are the drawbacks with regards o HAZOP done by human teams. Given the enormous amounts of Fig. 1. HAZOP study procedure. bahamut 下划线 mical i s o a H “ l a m a 3 o 3 t s “ a r t a s I r p R r a t b s m u I v b r F a f M b v 3 c d d d a r t a f i s f a g J. Zhao et al. / Computers and Che n a HAZOP, make the analysis more thorough, detailed, and con- istent, minimize human errors, and free the team to concentrate n the more complex aspects of the analysis which are unique nd difficult to automate (Venkatasubramanian et al., 2000). The AZOP analysis that is difficult to automate generally refers to non-routine” analysis discussed in the above section. In what fol- ows, case-based reasoning and ontology are integrated for the utomation of HAZOP analysis that was considered difficult to auto- ate before by using the traditional model-based or rule-based pproaches. . HAZOP expert system by the integration of CBR and ntology .1. Case-based reasoning (CBR) Experts often find it easier to relate stories about past cases han to formulate rules. Similarly it is true in the HAZOP analy- is domain that rules or models are hard to construct to automate non-routine” analysis. To overcome this problem, an important rtificial intelligence technique – CBR is adopted to augment the easoning machines embedded in the existing HAZOP expert sys- ems. CBR is both a pattern for computer-aided problem solvers and model of human cognition. The central idea is that the problem olver reuses the solution from past cases to solve a new problem. n this way, valuable experiences that are difficult to formulate into ules or models could be utilized for solving new problems. In a CBR system, the problem solving process includes four hases (Aamodt and Plaza, 1994), namely 4R’s: Retrieve, Reuse, evise and Retain as shown in Fig. 2. Basically, knowledge and expe- ience are stored in the form of cases in a case base. The content of case is made up of three parts: the problem/situation description, he solution, and the outcome. The outcome is not needed but could e added to suggest solutions that work and use cases with failed olutions to warn of potential failures. When a new problem is sub- itted, CBR system retrieves the similar cases from the case base i a e o Fig. 2. Problem solving Engineering 33 (2009) 371–378 373 sing certain similarity algorithm based on the predefined indexes. ndexes should be abstract enough to retrieve a relevant case in a ariety of future situations, and also should be concrete enough to e easily recognizable in future situations. Then the solutions of etrieved cases are adapted to solve the new problem if necessarily. inally, the new problem description and its solutions are retained s a new case in the case base for future use. Most of the CBR applications do not go through all the above our phases (López-Arévalo, Bańares-Alcántara, Aldea, Rodríguez- artínez, & Jiménez, 2007). In this paper, the adaptation is done y the users because it is highly domain dependent and requires erification of the solution performance. .2. Ontology Human experts are indispensable in HAZOP analysis of any hemical processes even though various expert systems can be esigned to facilitate the process. Different experts especially from ifferent organizations have different jargons with regards to the escriptions of the analysis objects and results including causes nd consequences of hazards. That is to say, there is no standard to epresent the HAZOP analysis domain information. This increases he difficulty of CBR for different users. To settle the terminological nd conceptual incompatibility problem, a new set of ontologies or CBR-based HAZOP analysis (CHA) is created in this paper by ntegration of existing ontologies reported in literatures. “Ontology” is a term of philosophy originally, which refers to the ubject of existence. Artificial intelligence (AI) borrows the old term rom philosophy and gives it new wonderful meanings. In AI, there re a number of definitions of ontology. However, the definition iven by Gruber is accepted by majority of researchers: an ontology s an explicit specification of a conceptualization (Gruber, 1993). HAZOP analysis of chemical process needs knowledge from reas such as chemistry, chemical process engineering, safety ngineering, electrical engineering and so on. A large-scale ontol- gy OntoCape was constructed by the research group of Professor phases with CBR. bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 374 J. Zhao et al. / Computers and Chemical Engineering 33 (2009) 371–378 ow of M M t r ( s D o 2 1 ( t H u r c c s o p o f d s r d i c c d P t 3 a t a t s o l o a u e 3 i s o g Fig. 3. Workfl arquardt for chemical process engineering (Morbach, Yang, & arquardt, 2007). To facilitate the information sharing among he HAZOP analysis expert systems developed by the labo- atory of Professor Venkatasubramanian at Purdue University Venkatasubramanian et al., 2000), process simulation packages uch as Aspentech’s BatchPlus and documentation tool such as yadym’s PHAPro, operational related ontologies and safety related ntologies were created (Zhao, Bhushan, & Venkatasubramanian, 003). Batres et al. created an upper level ontology based on ISO 5926 that had already been used for knowledge queries in HAZOP Batres et al., 2007). Based on the major concepts and ideas from he above ontologies, six ontologies are created for case-based AZOP analysis. They respectively are process ontology, process nit ontology, unit operation ontology, equipment ontology, mate- ial ontology and HAZOP ontology. Within each of the ontologies, oncepts are organized in a hierarchy where concept nodes are onnected by is-a links. Synonyms are given for concepts whose ynonyms are available in the CPI. Process ontology is built based n the classification of chemical plant types such as hydrocracking lant, FCCU plant, ethylene plant, ammonia synthesis plant and so n. Process unit ontology and unit operation ontology are basically rom OntoCape. In equipment ontology, equipment is specified by esign properties such as design temperature, pressure and some tructural specifications. Since MSDS information of each mate- ial is indispensable in HAZOP analysis, material ontology not only epicts the chemical species information, but also contains MSDS nformation for each material. In HAZOP ontology, HAZOP related onceptions such as nodes, parameters, guidewords, deviations, auses, consequences, safeguards, recommendations and risk are escribed. After the ontologies were created, the ontology editor d g s d t PetroHAZOP. rotégé (Stanford Medical Informatics, 2006) was used for verifica- ion. .3. Integrated reasoning framework for HAZOP analysis As stated above, the existing HAZOP expert systems could only ddress “routine” and generic process analysis. Due to the lack of he ability of machine learning, they could not “remember” the nalysis, especially “non-routine” analysis that has been done so hat the HAZOP team has to do the analysis once again even if imilar HAZOP analysis scenarios have been discussed before. To vercome this problem, a HAZOP expert system PetroHAZOP with earning capability is built by integrating CBR and the above six ntologies. The HAZOP analysis process of the case-based expert system pproach is illustrated in Fig. 3. PetroHAZOP consists of four mod- les, as shown in Fig. 4. In what follows, the four modules will be xplained. .3.1. Case base module Construction of the case base to a large extent determines the ntelligence level of a CBR system. Each case instance generally con- ists of two parts: the problem and the solution. Inside the case base f PetroHAZOP, the problem part contains the HAZOP analysis back- round information of a particular deviation while the solution part escribes its abnormal causes, adverse consequences, risk, safe- uards, recommendations and some other auxiliary information uch as the name of HAZOP team members and HAZOP analysis ate. Stored in a relational database, each case holds a unique iden- ification number. It is not uncommon that there are hundreds or J. Zhao et al. / Computers and Chemical e i c s s a m s o t c fi s t t H f c t c s d a o p i m a c s i t a 3 p p T t t a t T a a s a d d c ( ( Fig. 4. Configuration of PetroHAZOP. ven thousands of deviations that need to be combed through dur- ng a HAZOP analysis of a typical chemical process. Therefore, the ase base will grow to a very large size with time. To facilitate the imilarity-based case retrieval that is described in the following ection, a hierarchical case structure is introduced in this paper s an indexing method to partition a huge number of cases into ultiple hierarchical subordinate case bases (SCB). HAZOP analy- is cases can be categorized into the first level SCBs by the types f the chemical processes specified in the process ontology while hose cases within a same type of chemical process can be further lassified into the second level SCBs by the equipment types speci- ed in the equipment ontology. Cases inside a second level SCB can till be divided into the third level SCBs according to the deviation ypes specified in the HAZOP ontology. This hierarchical case struc- ure can be regarded as a knowledge-based indexing method where AZOP domain-specific knowledge is applied, important features or quick and accurate retrieval of past cases (Barletta, 1991). Each case in the case base is defined by indexes of four major ategories: equipment with its design parameters, materials con- ained in the equipment, operating conditions, and stream context onditions. The equipment design parameters such as design pres- ure and design temperature describe the equipment where the eviation being analyzed occurs. The equipment type must be vailable in the equipment ontology. Each case contains a list f materials present in the equipment. Hazardousness related hysical–chemical characters of materials such as flash point, boil- ( Fig. 5. Similarity algorithms for diff Engineering 33 (2009) 371–378 375 ng point and toxicity are distilled from MSDS to represent the aterial characters. Operating conditions include parameters such s operating temperature, pressure and level. The stream context onditions reflect the equipment types of both up stream and down tream of the equipment. According to the standard IEC61882, the case solution which s the HAZOP analysis results should include information such as he deviation’s root causes, adverse consequences, the safeguards vailable in the P&IDs, the recommendations for hazard mitigation. .3.2. CBR engine module The CBR engine module (CEM) is the core of the indispensable hase of CBR systems, i.e. retrieval. When a new deviation analysis roblem is presented to the system, the CBR engine is activated. he engine starts from selecting the corresponding SCB that fits he problem through the hierarchical indexing mechanism. Within he chosen SCB, all past cases are compared with the new problem, nd scored based on the similarity-based case retrieval algorithm hat is described in the following to find the closest-matching cases. o define the similarity between the past case and new problem, measure is needed first to assess the closeness between the ttributes belonging to them. Basically there are five types of attributes for each case: object uch as equipment, string such as the material name, numeric such s operating temperature of equipment, interval-numeric such as esign parameters and set object such as materials. Fig. 5 shows ifferent similarity algorithms that are employed in this paper to alculate the similarities of different types of attributes. 1) The similarity of string attributes is simply calculated by string matching algorithm. If string attributes are same, then their similarity is 1, otherwise 0. 2) In a HAZOP case, all numeric attributes are transformed to non- negative values. For example, the temperature unit is Kelvin temperature while the absolute pressure is used to represent pressure attributes. The mathematics similarity formula for non-negative numeric attributes xi and yi is: sim(xi, yi) = 1 − ∣∣xi − yi ∣∣ max(xi, yi) (1) 3) Similarity between sets Usually, there is more than one material involved in a piece of process equipment. Comparison of HAZOP cases usually requires comparison of the material sets present in the equip- ments where the cases originate. We proposed a new approach erent types of case attributes. bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 3 mical ( ( m o r l P s n T t m s w B s i e O Q e i e t 3 i k q t t H c c d b p “ a e j t f 3 t • • • • • • • • • h I e s g u S t fi done, the caption of button is changed to “Show similar cases”. PetroHAZOP also has a user menu for project management, public 76 J. Zhao et al. / Computers and Che (Set-Similarity method) to compute the material set similarity between different cases. The approach is expressed as follow. Assume there are two material sets A and B, which belong to two different cases, respectively. Set A contains m materials: MA1, MA2, . . ., MAm, and set B contains n materials: MB1, MB2, . . ., MBn. Then the similarity of A and B could be computed by Eq. (2). sim(A, B) = N∑ i=1 Si max(m, n) (2) where N = Min(m, n) Si is the maximum similarity between the ith material in one set and each material in the other material set, 1 ≤ i ≤ N. if N = m then, Si = n Max j=1 {sim(MAi, MBj )} (3) if N = n then Si = m Max j=1 {sim(MAj , MBi)} (4) In Eqs. (3) and (4), sim(MAi,MBj) represents the similarity between material MAi and material MBj, which can be computed by Eq. (5) sim(MAi, MBj ) = K∑ k=1 (Wksim(attAik, attBjk)) (5) where K represents the number of index attributes of a material, Wk represents the weight of the kth index attribute, 1 ≤ k ≤ K, attAik, attBik are respectively the kth numeric index attributes of materials MAi and MBj, and sim(attAik, attBik) can be calculated by Eq. (1). 4) The similarity algorithm of interval-numeric feature is extended-Euclidian algorithm. Suppose there are two interval- numeric attributes A = [a1,a2], B = [b1,b2], then their similarity can be calculated as follows: sim(A, B) = (a1b1 + a2b2) max((a1)2 + (a2)2, (b1)2 + (b2)2) (6) 5) Object similarity HAZOP cases have object attributes such as equipment and aterials. The object similarity calculation takes advantage of the ntologies described in Section 3.2. Ontology based similarity algo- ithms have been reported in literature. In this paper, the path ength measure is used to calculate the object similarity (Pedersen, akhomov, Patwardhan, & Chute, 2007). It essentially computes the imilarity between two object nodes by counting the numbers of odes on the shortest path between them in the ontology hierarchy. he shortest path includes both the object nodes. Mathematically, he similarity of two object nodes A and B using the path-length easure (path) is defined as: im(A, B) = 1 (7) p here p is the number of nodes on the shortest path between A and within an ontology hierarchy. For example, in the equipment ontology, if equipment A is a ubclass of equipment P (subclass is equivalent to a is-a relationship d v m C n Engineering 33 (2009) 371–378 n ontology), and equipment B is a subclass of equipment Q while quipment P and equipment Q are two subclasses of equipment . The shortest path from equipment A to equipment B is A-P-O- -B. There are five nodes on the path. Therefore, the similarity of quipment A and equipment B is 1/5 (see Fig. 5). Finally the case similarity is the sum of each case attribute sim- larity multiplied by its weight which is determined by domain xperts. The weights are adjustable through the knowledge main- enance module which is described below. .3.3. Knowledge maintenance module The effectiveness of a CBR system depends largely on the qual- tative and quantitative richness of its stored cases which are the nowledge repository of past experiences. That is to say, the more uality cases stored in the case base, the more effectively the sys- em reasons. Initially, there are few cases in the case base. Through he knowledge maintenance module (KMM), cases from the past AZOP analysis records can be manually input to the subordinate ase bases to facilitate CBR. Case retaining is another feature of KMM. Revised cases can be onverted to new cases and retained in the corresponding subor- inate case bases. Before a new case is stored, term translation ased on ontology is to be done if necessary. For example, “tem- erature too high” and “high temperature” will be translated into more temperature” in the case base. Since weight factors used in the similarity based case retrieval re predefined according to the knowledge of authors and a few xpert consultants, it will not be surprised that there is a need to ustify them according to the feedbacks of users in industrial prac- ices. A user interface is designed in KMM to modify the weight actors through a certain level of authorization. .3.4. Graphical user interface (GUI) module The GUI module contains graphical user interfaces to perform he following functions: Creating HAZOP analysis project Specifying equipment Specifying materials Selecting parameters and HAZOP guidewords Editing HAZOP analysis results Retrieving and reusing similar cases Retaining cases Reporting HAZOP analysis results Administrating users Fig. 6 is a snapshot of a main GUI of PetroHAZOP. On its left- and side is a treeview like the Microsoft’s Windows Explorer tree. t displays a hierarchical collection of labeled items such as nodes, quipments, process variables and deviations. On its right-hand ide is the area where causes, consequences, risk, safeguards, sug- estions and comments for a certain deviation can be edited by sers or automatically loaded from selected similar cases. Case earching button on the top-right corner of the snapshot implies hat the system is searching similar cases based on the speci- cations of the deviation being analyzed. Once the searching is ata entry and user administration. On the bottom of the tree- iew, there are two more nodes, i.e. materials and report. In the aterials node, the process materials can be specified by users. ustomizable HAZOP reports can be generated by using the Report ode. J. Zhao et al. / Computers and Chemical Engineering 33 (2009) 371–378 377 hical 4 t u t i w p p “ 4 i t u a a p p r r h r s e f o f m c o o a r t c t 4 Fig. 6. Snapshot of a main grap . Application examples PetroHAZOP is programmed with Java allowing concurrent mul- iple users to manipulate the system through intranet. This multiple ser mode greatly improves the work efficiency of the HAZOP eam. Recently, PetroHAZOP has been successfully installed and mplemented at one of the largest oil companies in China. There ere more than 900 cases in the case base at the time when this aper was written. The following examples demonstrate how the roposed HAZOP expert system PetroHAZOP can help with both routine” and “non-routine” analysis. .1. Example 1: “non-routine” analysis Acrylonitrile production process is a highly hazardous process n the petrochemical industry. Most of the 16 materials involved in he process including raw materials, intermediate products, prod- cts and by-products are flammable, toxic or/and volatile. There re about 14 major equipments such as reactor, chilling tower, bsorption column, recycle column, distillation columns. The whole rocess was divided into six nodes. Totally 87 deviations of 59 key arameters had been analyzed by a HAZOP team. We transfer the esults through the knowledge maintenance module into cases, esulting in 87 cases in the case base. H t s Fig. 7. Snapshot of case ‘Ignition failure user interface of PetroHAZOP. The vinyl chloride production process (VCPP) is another highly azardous chemical process which consists of three sections: chlo- ine/hydrogen processing section (CHPS), hydrochloride synthesis ection (HSS) and vinyl chloride synthesis section (VCSS). One xample of the ‘non-routine’ analysis of this process was ‘ignition ailure of the HCl synthesis furnace’ since it is hard to be modeled r generalized with rules. With PetroHAZOP, a similar case ‘ignition ailure of acrylonitrile startup furnace’ from the case base was auto- atically retrieved with the similarity degree of 0.602. The found ase is analogous in the equipment type in the equipment ontol- gy and the deviation type in the HAZOP ontology. If the user clicks n the button “Show similar cases”, he/she can find out the causes nd consequences of similar case (Fig. 7). This similar case can be eused if the button “Reuse” on Fig. 7 is pressed. Here reuse means hat the causes and consequences are automatically loaded to the auses and consequences of the new case. The user then can edit he causes and consequences if necessary. .2. Example 2: “routine” analysis Another example is the completeness checking even for routine AZOP analysis by means of CBR. A HAZOP team was assigned o HAZOP a polymer production process. Node 1 containing a tirred tank was analyzed first and it took the team about 2 days to of acrylonitrile startup furnace’. 3 mical c o w t r t s c T l S c s 5 p T a r t T t l c C c e E o i o t l a t t p a t A T a R A A B B C G K L M M M P S V Zhao, C., Bhushan, M., & Venkatasubramanian, V. (2003). Roles of ontology in 78 J. Zhao et al. / Computers and Che omplete its analysis. In the third day, the team started the analysis f Node 2 which also contained a stirred tank. When the team as about to close the analysis of the ‘Low Pressure’ deviation in he stirred tank of Node 2, the leader asked the recorder who was esponsible for manipulating PetroHAZOP to retrieve similar cases o check if anything was missed. A similar case ‘Low pressure in the tirred tank of Node 1’ was discovered in the case base and one of its auses, ‘axial sealing leak’, was not considered for the current case. he consequence of the cause was “oxygen entering the stirred tank eading to contamination and deactivity of the catalyst in the tank”. ince the cause and consequence could also happen in the current ase, they decided to reuse the found case, and minor modification uch as change of the catalyst name was made for the new case. . Conclusions HAZOP analysis requires high accuracy, consistency and com- leteness because any ignorance would lead to catastrophic losses. herefore, the HAZOP team must ensure that it would not lose ny resources that are available to help them meet the above equirements. As a solution, this article offers an integrated solu- ion for the complex problems in the path of automating HAZOP. he proposed HAZOP expert system PetroHAZOP not only facili- ate “routine” analysis but also “non-routine” analysis due to its earning capability by which the HAZOP analysis quality can be ontinuously improved during practice. Due to the adoption of BR mechanism, PetroHAZOP can map past experiences to the new ases. Therefore, it is more adaptive in HAZOP analysis and greatly ases the knowledge management and knowledge dissemination. ven though HAZOP has been widely practiced in the CPI of devel- ped countries, it just starts to get recognized by the practitioners n China. Lack of HAZOP experts hinders the wide implementation f HAZOP in the chemical plants. Hopefully, PetroHAZOP can facili- ate the industrial exercises of HAZOP in China and contribute to the oss prevention in the largest developing country where chemical ccidents are posing a serious threat to its fast development. Automatic adaptation for HAZOP analysis is an outstanding task hat could be addressed by introducing other artificial intelligence echnologies to CBR. Layered digraph model (LDG) has been pro- osed by authors to perform both “non-routine” and routine HAZOP nalysis (Cui, Zhao, Qiu, & Chen, 2008). Future work will be oriented o integrate CRB with LDG based reasoning. Z Engineering 33 (2009) 371–378 cknowledgements This work is supported by Program for New Century Excellent alents in University. Anonymous reviewers of this paper are highly ppreciated for their helpful comments and suggestions. eferences amodt, A., & Plaza, E. (1994). Case-based reasoning: Foundational issues. Methodological variations, and system approaches. Artificial Intelligence Com- munications, 7(1), 39–59. ha, D. W. (1998). The omnipresence of case-based reasoning in science and appli- cation. Knowledge-Based Systems, 11(5–6), 261–273. arletta, B. (1991). An introduction to case-based reasoning. AI Expert, 6(8), 42–49. atres, R., West, M., Leal, D., Price, D., Masaki, K., Shimada, Y., et al. (2007). An upper ontology based on ISO 15926. Computers & Chemical Engineering, 31(5–6), 519–534. ui, L., Zhao, J., Qiu, T., & Chen, B. (2008). Layered digraph model for HAZOP analysis of chemical processes. Process Safety Progress, 27(4), 293–305. ruber, T. R. (1993). A translation approach to portable ontology specifications. Knowledge Acquisition, 5(2), 199–221. olodner, J. (1993). Case-based reasoning [M]. San mateo, CA: Morgan Kaufmann Publishers, Inc. ópez-Arévalo, I., Bańares-Alcántara, R., Aldea, A., Rodríguez-Martínez, A., & Jiménez, L. (2007). Generation of process alternatives using abstract models and case- based reasoning. Computers & Chemical Engineering, 31(8), 902–918. cCoy, S. A., Wakeman, S. J., Larkin, F. D., Jefferson, M. L., Chung, P. W. H., Rushton, A. G., et al. (1999). HAZID, a computer aid for hazard identification 1. The STOPHAZ package and the HAZID code: An overview, the issues and the structure. Process Safety and Environmental Protection, 77(6), 317–327. cCoy, S. A., Wakeman, S. J., Larkin, M. L., Chung, P. W. H., & Rushton, A. G. (2000). HAZID, a computer aid for hazard identification: 4. Learning set, main study system, output quality and validation trials. Process Safety and Environmental Protection, 78(2), 91–119. orbach, J., Yang, A., & Marquardt, W. (2007). OntoCape—A large-scale ontology for chemical process engineering. Engineering Applications of Artificial Intelligence, 20(2), 147–161. edersen, T., Pakhomov, S. V. S., Patwardhan, S., & Chute, C. (2007). Measure of semantic similarity and relatedness in biomedical domain. Journal of Biomedical Informatics, 40(3), 288–299. tanford Medical Informatics. (2006). The protégé ontology editor and knowledge acquisition system. Available at http://protege.stanford.edu. enkatasubramanian, V., Zhao, J., & Viswanathan, S. (2000). Intelligent systems for HAZOP analysis of complex process plants. Computers & Chemical Engineering, 24(9–10), 2291–2302. automated process safety analysis. Computer Aided Chemical Engineering, 14, 341–346. hao, C., Bhushan, M., & Venkatasubramanian, V. (2005). PHASUITE: An automated HAZOP analysis tool for chemical processes Part I: Knowledge Engineering Framework. Process Safety and Environmental Protection, 83(B6), 509–532. http://protege.stanford.edu/ Learning HAZOP expert system by case-based reasoning and ontology Introduction HAZOP HAZOP expert system by the integration of CBR and ontology Case-based reasoning (CBR) Ontology Integrated reasoning framework for HAZOP analysis Case base module CBR engine module Knowledge maintenance module Graphical user interface (GUI) module Application examples Example 1: "non-routine" analysis Example 2: "routine" analysis Conclusions Acknowledgements References 
work_pxrj4u4bkbdaxngyw7p7vedjne	----	Fuzzy-Expert System for Voltage Stability Monitoring and Control BY O Bodapatti Nageswararao, B .E. A thesis submitted to the School of Graduate Studies in partial fulfillment of the requirements for the degree of Master of Engineering Faculty of Engineering and Applied Science Memonal University of Newfoundland Febniary, 1998. St. John's Canada National Library Bibliothèque nationale du Canada Acquisitions and Acquisitions et Bibliographie Services services bibliographiques 395 Wellington Street 395. nie Wellington OtGawaON K 1 A W OttawaON KtAON4 Canada Canada The author has granted a non- exclusive licence allowing the National Library of Canada to reproduce, loan, distribute or sel copies of this thesis in microform, paper or electronic formats. The author retains ownership of the copyright in this thesis. Neither the thesis nor substantial extracts fkom it may be printed or othexwise reproduced without the author's permission. L'auteur a accordé une licence non exclusive permettant à la Bibliothèque nationale du Canada de reproduire, prêter, distribuer ou vendre des copies de cette thèse sous la fome de microfiche/nlm, de reproduction sur papier ou sur format électronique. L'auteur conserve la propriété du droit d'auteur qui protège cette thèse. Ni la thèse ni des extraits substantiels de celle-ci ne doivent être imprimés ou autrement reproduits sans son autorisation. Abstract In recent years, electric power utilities are forced to transmit maximum possible power through existing networks due to environmental, economic and regulatory changes. Due to these constraints, voltage instability has emerged as one of the most important areas of concem to modem power utilities. Voltage instability has been responsibie for several system collapses in North America, Europe and Asia This thesis presents fùndamental concepts of voltage stabiliîy. it describes three traditionai voltage stability indices namely singular value decomposition, L index and QV curves. A simple five bus system is used to highlight the limitations of these traditional methods. A more widely accepted technique like modal analysis along with continuation power flow is studied and simulations are carried out on the IEEE 30 bus system and the New Engfand 39 bus system. The test results clearly indicate areas prone to voltage instability and also identify groups of buses and cntical bus that participate in the instability and thereby eliminate the problems associated with traditional methods. Hence, modal d y s i s technique is not only used as a benchmark tool for the developrnent of the proposed f h q - expert system, but also as an important tool for validating its accuracy. To understand this new approach, fundamental concepts of funy logic based on the theory of approximate reasoning is dealt in detail. To get fûrther insight into this altemate approach, a simple method using fùzq sets for the voltage-reactive power control to improve the system voltage level is presented. A modified IEEE 30 bus system is used as an example to illustrate this method. Simulation resuits of this simple problem is encouraging and has been a useful stamng point for the proposed fùzzy-expert system for voltage çtability evaluatioa The proposed fuay-expert system consists of two main wmponents. nie knowledge-base and the inference engine. Here, the key system variables Like load bus voltage, generator MVAR reserve and generator terminal voltage which are used to monitor the voltage stability are stored in the database. Changes in the system operating conditions are reflected in the database. The above key variables are fuzzified using the theory of uncertamty. The debase comprises a set of production d e s which fom the basis for logical reasoning conducted by the inference engine. The production d e s are expressed in the form of F-THEN type, that relates key system variables to stability. The New England 39 bus system is taken as a case study to illustrate the proposed procedure. The expert system output is compared with the simulation results of a commercially available software ( VSTAB 4.1 ) output through modal analysis. The proposed system is fast and more efficient than conventional voltage stability methods. Acknowledgments 1 would like to express my sincere gratitude and appreciation for the invaluable help and guidance given to me by Dr. Benjamin Jeyasurya during al1 stages of this work. My thanks are also to Dr. Jeyasurya, Facuity of Engineering and Applied Science and the School of Graduate Studies for the fuiancial support provided to me during my M-Eng. program. 1 acknowledge the assistance received fiom faculty members, fellow graduate midents and other staff of the Faculty of Engineering and Applied Science. Finally, I express my sincere gratitude to my family for their encouragement, support and understanding. Contents A c h o w ledgments Contents List of Figures List of Tables iv v vui IC 1 INTRODUCTION 1 2 VOLTAGE S T A B I ' ANALYSIS IN P O W R 4 SYSTEMS 2.1 Introduction 4 2.2 Voltage Stability Phenomenon 5 2.3 Voltage Collapse incidents 10 2.4 Factors Contributing to Voltage uistability/Collapse 1 1 2.5 Voltage Stabilify Analysis 12 2.5.1 S ingular value decomposition 13 2.5.2 L index 15 2.5.3 Q V c w e s 17 2.6 Simulation Resuits and Discussions 19 2.6.1 Singular value decomposition 20 2.6.2 L index 22 2.6.3 QV curves 24 2.7 Limitation of Traditional Methods in Voltage Stability 28 Anal ysis 2.8 Need for Expert Systems in Electric Power System 29 2.9 Slrmmary 3 CONTINUATION POWER FLOW AND MODAL ANALYSPS 3.1 Introduction 3.2 Continuation Power Flow Technique 3.2.1 Basic prùlciple 3.2.2 Mathematical formulation 3.3 Modal Analysis 3.4 Simuiation Results and Discussions 3 -4.1 lEEE 30 bus system 3.4.2 New England 39 bus system 3.5 Summary 4 FUZZY-EXPERT SYSTEMS 4.1 Introduction 4.2 Expert Systern Structure 4.3 Theory of Approximate Reasoning 4.3.1 Fuay set theory 4.4 Application of Funy-Set Theory to Power Systems 4.5 sulfl~zlary 5 FUZZY CONTROL APPROACH TO VOLTAGE PROFILE ENEANCEMENT FOR POWER SYSTEMS 5.1 Introduction 5.2 Problem Statement 5.3 Fuzzy Modeling 5.3.1 Bus voltage violation level 5.3 -2 Controlling ability of controllhg devices 5.3.3 Control strategy 5.4 Methodology 5 -5 Simulation Results and Discussions 5.6 Summary 6 FUZZY-EXPERT SYSTEM FOR VOLTAGE STABItITY MONITORING AND CONTROL Introduction Expert S ystem and Design 6.2.1 The Database 6.2.2 TheRulebase Merence Engine Simulation Results and Discussions Inteption of Fuzzy-Expert Systenl into an Energy Management System s-ary 7 CONCLUSIONS 7.1 Contributions of the Research 7.2 Recommendations for Future Work REFERENCES APPENDICES A Line and Bus data for the 5 bus systern B Rules relating key variables to stability masure of the New England 39 bus system C Data used for the debase formation for the monitoring stage of the New England 39 bus system D Complete list of expert system output for the monitoring stage of the New England 39 bus system E Participation factors for the cntical case ( at the voltage stability limit ) o f the New England 39 bus system List of Figures Sample two bus power systern Vottage-Power characteristics for difEerent V1 and power factors Radial system with some of the elements that play key role in the voltage stability Line mode1 for two bus system Sample QV curve Single line diagram of the 5 bus system Singular value decomposition for 5 bus system L index of individual load buses for 5 bus system System L hdex for 5 bus system Family of QV curves for the base case of the 5 bus systern Family of QV curves near the stability Iimit of the 5 bus system Family of QV curves at the collapse point of the 5 bus system An illustration of the continuation power flow technique Single line diagram of the IEEE 30 bus system PV c u v e for bus 30 of the IEEE 30 bus system Single line diagram of New England 39 bus system PV curve for bus 12 of the New England 39 bus system Expert system structure Mernbenhip function for the fuPy set TALL The S-function The II-fûnction F u n y mode1 for bus voltage violation level Fuzzy mode1 of controlling ability of controlling device 6.1 Membership h c t i o n for the worst load bus voltage 76 6.2 Membenhip fiinction for the worst MVAR reserve 77 6.3 Membership fiindon of the generator terminal voltage 78 corresponding to the worst MVAR reserve 6.4 Voltage stability margui ( VSM ) for pre and p s t control cases 92 6.5 Funy-expert system as a part of new EMS 98 List of Tables Five smallest eigenvalues for the three loading condition of the EEE 30 bus system Participation factors for base case and critical case corresponding to the Ieast stable mode of the IEEE 30 bus system Voltage stability margin for the three loading condition of the IEEE 30 bus system Five smallest eigenvalues for different loading conditions of the New England 39 bus system Participation factors for criticai case corresponding to the least stable mode of the New England 39 bus system Voltage stability margin for the three loading condition of the New England 39 bus system Fuzzy logic operators Base case voltage profiles of the modified IEEE 30 bus system Lower and upper limits of the controllers Load bus voltage profiles of the modified IEEE 30 bus system before and after control actions Optimal fuzzy controi solution Panuneters "a" and "A" for the membership fûnction of the key variables Operating conditions for the 32 cases listed in Appendk D Expert system output - Monitoring Expert system output - control stage ( case A ) Expert system output - control stage ( case B ) Expert system output - control stage ( case C ) Voltage stability marpin ( VSM ) for pre and p s t control cases L h e &ta for the 5 bus system Bus data for the 5 bus system Rdes relating worst load bus voltage to d i l i t y measure under group 1 Rules relating key generator MVAR reserve to stability measure under group 2 Rdes relating key generator terminal voltage to stability measure under group 3 Rules relating combineci key variables to stability measure under group 4 Data used for the nilebase formation Expert systern output for various neighborhood points - Monitoring stage Bus and Generator participation factors for critical case Chapter 1 INTRODUCTION The phenornenon of voltage instability in electric power systems is characterized by a progressive decline of voltage, which can occur because of the inability of the power system to meet increasing demand for reactive power. The process of voltage instability is generally triggered by some fom of disturbance or change in operating conditions that create increased demand for reactive power in excess of w b t the system is capable of supplyuig. The dynarnic characteristics of loads, location of reactive compensatiori devices and other control actions such as those provided by load tap changing transformers, automatic voltage regulating equiprnent, speed goveming mechanism on generators are important factors which affect voltage stability. in recent years, electric power utilities are forced to transmit maximum possible power through existing networks due to environmental, economical and regulatory changes. As a result of load growth without a corresponding increase in either the generation or transmission capacity, many power systems operate close to their voltage stability boundaries. Many utilities around the world have experienced major black-outs caused by voltage instability. When a power system is operating close to its limits, it is essential for the operators to have a clear knowledge of its operating state. A number of speciai algorithms and methods have been proposed in the literature for the analysis of voltage instability. But these traditional methods require significantiy large wmputations and are not efficient enougb for reai-time use in energy management system. Hence, there is a need for an alternative approach, which cm quickiy detect potentiaily dangerous situation and alleviates the power system nom possible collapse or blackout. To meet the above challenge, this thesis proposes a new and cost-effective solution based on fby-expert system. The aim of the thesis is as follows: O To investigate three traditional methoâs of voltage stability indices with the help of a simple system. The three methods are singular value decornposition, L index and QV curves. The analysis of these methods will bnng out the bsic concepts involved in voltage stability along with their limitations. O To apply modal anaiysis along with continuation power flow for voltage stability analysis of a power system. These methods aim to overcome the limitations of the traditional methods. To review expert systems and fuzn/ logic concepts. This will lay a foundation for the understanding of the proposed fûzzy-expert system. To get M e r insight into this alternative approach, a simple method using f i n y sets for voltage-reactive power control to improve the system voltage level is investigated To design a fùzzy-expert system for voltage stability monitoring and control. The designed funy-expert system is investigated and compared with modal anaiysis output Extensive simulations are carried out to validate the usefulness of this new approach. The thesis is organued as follows. Chapter two reviews the concepts of voltage stability followed by the description of the three most widely & techniques for the analysis of voltage stability in power systems. A simple 5 bus system is used to highiight the limitations of these traditional methods. Chapter three gives a detailed description of modal analysis for voltage stability evaiuation followed by the simulation results of the IEEE 30 bus system and the New England 39 bus system. This chapter forms the basis for the development of hizzy-expert system. Chapter four presents the theory behind funy-expert systems. Ln chapter five, a simple application of ~~I.ZZY logic to power system is s h o w to emphasize the theory developed in chapter four. A moâified IEEE 30 bus system is used as an example to illustrate this m e t h d Chapter six gives a detailed description of the fùzzy-expert system for voltage stability monitoring and control. The New England 39 bus system is taken as case study to illustrate the proposed procedure. It highlights the database and nilebase design. In the database design, the key system variables that are used to monitor the voltage stability are hansformed into fuay domain to obtain their appropriate membership hctions. The rulebase comprises a set of production rules that f o m the basis for logical reasoning conducted by the inference engine. These production d e s relate key system variables to stability. To validate the proposed system, simulation results are compared with modal analysis output Finally, chapter seven concludes the thesis with sorne recommendation for friture work. Chapter 2 VOLTAGE STABILITY ANALYSIS IN POWER SYSTEMS Introduction Voltage control and stability problems are not new to the electric utility industry but are now receiving special attention. Maintaining an adequate voltage level is a major concem because many utilities are loading their bulk transmission networks to their maximum possible capacity to avoid the capital cost of building new lines and generation facilities. Load growth without a correspondhg increase in transmission capacity has brought many power systems close to their voltage stability boundaries. In this conte* the terms 'koltage stability", "voltage collapse" occur fiequently in the literature. EEE cornmittee report [l] defines the following terrninology related to voltage stability : "Voltage Stability" is the ability of a system to maintain voltage so that when load admttance is increased, load power will increase and that both power and voltage are controllable. "Voltage Collapse" is the process by which voltage instability leads to very low voltage profile in a significant part of the system. "Voltage instability" is the absence of voltage stability and results in progressive voltage decrease ( or increase ). Voltage instability is a dynamic process. A power system is a dynamic system. in contrast to rotor angle stability, the dynamics mainly involve the loads and the means for voltage control. A system enters a state of voltage instability when a disturbance, increase in load, or system change causes voltage to drop quickly or drift downward and operator, automatic system controis fail to halt the decay. The voltage decay may take just a few seconds or 10 to 20 minutes. If the decay continues unabated, steady-state angular instability or voltage collapse will occur. 2.2 Voltage Sta bility Phenornenon [2] Voltage collapse is in general caused by either load variations or contingencies. The following illustration considers voltage collapse due to load variations. The basic configuration used to explain voltage collapse is shown in Fig. 2.1. V2: receiving end load voltage X: reactance of the line V 1 : sending end voltage 6 : load angle Fig.2.1 Sample two bus power system In this circuit, a synchronous genenitor is connecteci to a load through a lossless transmission line. The load is described by its red and reactive powers P, Q and the load voltage V2 [3]. The goveming algebraic relations are Under steady-state conditions, equations (2.1) and (2.2) represent the voltage/power relation at the load end of the circuit- V I = l .O PF=0.95 laa \ // 0.2 0.4 0 - 6 0.8 1 1 . 2 1.4 Load Real Power (p.u) Fig.2.2 Voltage-Power characteristics for diEerent V1 and power factors [3] Fig. 2.2 shows the plot of load voltage versus red power for several power factors and different sending-end voltages. The graph of these equations is a parabola. In the region corresponding to the top half of the curves, the load voltage decreases as the receiving-end power increases. The nose points of these curves represents the maximum power that can theoretically be delivered to the load. If the load dernand were to increase beyond the maximum tramfer level, the amount of actual load which can be supplied as well as the receiving-end voltage will decrease. These curves Uidicate that there are two possible values of voltage for each loading. The system cannot be operated in the lower portion of the curve even though a mathematical solution exists. Consider an operating point in the lower portion of the c w e . If an additional quantity of load is added under this condition, this added Ioad wil draw additional current fkom the system. The resulting &op in voltage in this operating state would more than offset the increase in current so that the net effect is a &op in delivered pwer. If the load attempts to restore the demanded power by sorne means, such as by increasing the current, the voltage will decrease even M e r and faster. The process will eventually lead to voltage collapse or avalanche, possibly leading to l o s of synchronism of generating units and a major blackout. in a larger system, apart from load dynamics, other dynamics such as generator excitation control, on load tap changer ( OLTCs ), static var compenwitor ( SVC ) controls, thermostat controlled loads, etc, play an important role in the voltage stability of the system. The radial system shown below presents a clear picture of the voltage stability problem and its associated dynamics [l]. Genemtor to trip Residential Load 6 ~ i n e to trip Industrial not on LTC -4 Primary Capacitors LTC 7 Industrial Loat Fig.2.3 Radial system with some of the elements that play key role in the voltage stability [ 1 1 In the above system, two types of loads are considered: residential and industrial. Residential load has a relatively hi& power factor and tends to drop with voltage. On the other han4 industrial load has low power factor and does not Vary much with voltage. if this system is heavily loaded and operating near its voltage stability l e t , a mal1 increase in load ( active or reactive ), a loss of generation or shunt compensation, a drop in sending end voltage cm bring about voltage instability. Assuming that one of the above mentioned changes happen and receiving end voltage falls, several rnechanisms corne into play. As residential Ioads are voltage dependent, the active and reactive load drops with drop in voltage, while industrial active and readve loads which are dominated by induction motoa change linle. nius, the overall effect may be the stabilization of voltage at a value slightly less than the ratai value. The next action is the operation of distribution transformer load tap changers ( LTCs ) to restore distribution voltages. The residential active ioad will increase while the industrial reactive load will decrease. The increasing residential load will initially outweigh the decrease in reactive load causing the prirnary voltage to fdl further. in this scenario, the on- load tap changers ( OLTCs ) may be close to their limitç, primary voltage at around 90% and distribution voltage below normai. Industrial loads served fiom the prUnary system without LTCs will be exposed to the reduced voltage levels. This greatiy increases the stalling of induction moton. When a motor stalls, it will draw increasing reactive cunent, bringing down the voltage on the bus. This results in a cascade stalling of other induction motors resulting in a localized voltage collapse. Since, most large induction motors are controlled by magnetically held contactors, the voltage collapse would cause most motos to drop off from the system. This l o s of load will cause the voltage to recover. However, the recovered voltage will again result in the contactor closing, motor stalling and another collapse. Thus, this loss and rwvery of the load can cause aitemate collapse and recovery of voltage. From the above discussion, it is clear that voltage stability is essentially a slow dynamics and is affected by the nature and type of load and other control actions. Power systems have become more complex and are k i n g operated closer to their capability limits due to economic and environmental reasous. While these trends have contributed to angle instability, it is clear fiom a midy of recent incidents of system failures [1,4] that, it is voltage instability that is the major factor in these failures. 2.3 Voltage Collapse Incidents [4] Throughout the world, there have been disturbances involving voltage collapse over the last twenty years with the majority of these occurrences since 1982. Two typical examples of voltage collapse occurrences are the 1987 French and Tokyo power system failures. in France on January 12, 1987 at 10:30 am, one hour before the incident occurred, the voltage level was normal despite very low temperature outside. For various reasons, three themal units in one generating station failed in succession between 1055 and 1 1 :4 1 am. Thirteen seconds later, a fourth unit tripped as the result of operation of the maximum field cument protection circuit. This sudden loss in generation led to a sharp voltage &op. This &op in voltage increased thirty seconds later and spread to adjacent areas, resdting in the trippkg of other generating units on the system within a span of few minutes. As a result, 9000 MW were Iost on the French system between 11:45 and 1150 am. Normal voltage levels were restored rapidly after some load shedding took place on the system. In the same year on July 23, Tokyo's power system also experienced the voltage collapse phenornena The temperature rose to39"C and as a result, there was a sharp increase in the demand due to the extensive use of airconditionen. The 500 KV voltages began to suik ( to 460 KV ) and eventually the over current protection on the 500 KV Iines operated As a result, seven 500 KV nibstations were without supply. This resulted in the loss of 8000 MW of load for about three houn. It is interesting to note that during ail this tirne, there was no indication or abnomal operation which would have alerted the operaton to the impendlng disaster. The only indication that sornething out of the ordinary happening was that the rate of rise of the load was 400 MWhinute, which was twice as much as ever recorded before. 2.4 Factors Contributing to Voltage Instability/Colfapse [SI Based on the voltage collapse incidents described in references 1 and 4, the voltage collapse can be characterized as follows: The initiaikg event may be due to a variety of causes: smali gradua1 changes such as naturai increase in system load, or large sudden disturbances such as loss of a generating unit or a heavily loaded line. Sometimes, a seemingly uneventful initial disturbance may lead to successive events that eventually cause the system to collapse. The heart of the problem is the inability of the system to meet its reactive demands. Usually, but not dways, voltage collapse involves system conditions with heavily loaded lines. When the transport of reactive power fiom neighboring areas is difficult, any change that calls for additionai reactive power support may lead to voltage collapse. The voltage collapse generally manifests itself as a slow decay of voltage. It is the result of an accumulative process involving the actions and interactions of many devices, controls and protectïve systems. The t h e £iame of collapse in such cases could be of the order of several minutes. Thus, some of the major factors contributing to voltage instability c m be summarized as follows: sudden increase in load. O rapid on-load transformer tap changing. O level of series and shunt compensation. reactive power capability of generaton. response of various control systems. 2.5 Voltage Sta bility Analysis As incidents of voltage instability become more common and systems continue to be loaded closer to their stability limits, it becomes imperative that system operators be provided with tools that can identiS potentially dangerous situation leading to voltage collapse. A nurnber of special algorithm have been proposai in the literature [6] for the analysis of voltage instability. Few of hem are: (1) Singuiar Value Decomposition. (2) 'L' Index. (3) PV and QV c w e s . (4) Eigenvahe Decomposition. (5) V-Q Sensitivity. (6) Energy Based Measure. In ths chapter, the first three methods are discussed in detait because they are commonly used in the elecûic power industry. Whle, singular value decornposition and 'L' Ludex provide the necessary andytical tools in identifjmg voltage collapse phenornena, PV and QV curves are the more traditional methods used as voltage collapse proximity indicaton ( VCPI ) in indu- today. PV curves have already been discussed with respect to the simple two bus power systern in section 2.2. 2.5.1 Singular value decomposition In 1988, Tiranuchit and Thomas [ A proposed a global voltage stability index based on the minimum singular value of the Jacobian. They showed that a mesure of the neamess of a matnx A to singularity is its minimum singular value. An important aspect to be considered when deriving corrective control measure is the question " how close is the Jacobian to being singular? ". To examine the above question, consider the following basic problem: given a matrk A, determine conditions on perturbation matrix A such that A + AA is singular. Note that if A is non-singular, one may write A + A A = ( 1 + A - ' A A ) A ( 2.3 ) The terni ( I + A-' A A ) can be shown to have an inverse if To get a better understanding of the use of this voltage stability index, consider a set of non-linear algebraic equations in maîrix format given by Given y and Rx), we want to solve for x This can be done by Newton- Rhapson iterative method, where old values of x are used to generate new values of x i.e. where, J is an invertible rnatrix, called the Jacobian matrix. The Newton-Raphson method c m be applied to solve the load flow problem. The power flow equations in polar coordinates are as follows: P, , Q , : scheduled r d and reactive power supplieci to bus K V, , V, : bus voltages at bus K and n 6 ,, 6 , : phase angles of the bus voltages at bus K and n 8,, :admittance angle Y,, : elements of the bus admittance matrix For the power fiow probiem The Jacobian matrix for the power flow is o f the fonn: The inverse of the matrix J is given by Lf I J I = O, then the Jacobian is non-invertible or singular. The singular value decomposition of the Jacobian matrix gives a measure of the system closeness to voltage collapse. As the power system moves towards voltage collapse, the minimum singular value of the Jacobian matrix approaches zero. 2.5.2 L index In 1986, Kessel and Glavitsch [8] proposed a fast voltage stability indicator. This method provides a means to assess voltage stability without actually computing the operating point. Consider a two bus system as s h o w in Fig.2.4. Y : Series adminance Y : Shunt admittance V , . V, - Nodd v o t t q p Fig.2.4 Line mode1 for two bus system The properties of node 1 can be described in tems of the admittance matru< of the system. s ; Y,, v1 + Y,, v* = 1, = - v; Y,, . Y,, , Y,, . Y, form the admittance matrix [Y] and S, is the complex power. s, = v1 1; Equation 2.13 can be written as s : v,' + vo v; = - v where, It is shown in reference (81 that the solution of equation 2.15 indicates the stability limit of the power system. At t h i s point This relation can be used to define an indicator Lj at each bus for the assessrnent of the voltage stability. It's range is O 5 Lj < 1. The global indicator describing the stability of the complete system is the maximum of al1 Lj values at each and every bus. The indicator L, is a quantitative measure for the estimation of the distance of the actual state of the system to the stability tirnit. The local indicaton Lj can be used to determine the buses fiom which voltage collapse rnay originate. 2.5.3 QV Curves [9] QV curves are presently the workhorse method of voltage stability analysis at many utilities. The QV curves show the sensitivity and variation of bus voltages with respect to reactive power injections. They are used for the assessrnent of the voltage stability of the system. They show the mega volt ampere reactive ( MVAR ) and voltage rnargins to instability and provide information on the effectiveness of reactive power sources in controlling the voltage in different parts of the system. Vo-Vc : Voltage stability mvgin Qc : MVAR stabili~y margin Voltage Instability point Fig.2.5 Sample QV curve For each QV curve, a reactive power source is placed at the selected bus ( QV bus ) to move its voltage in a given range, Vmin to V- by a given step size V*. At each voltage step, the power flow is solved to compute the required MVAR injection Qi, at the QV bus for holding its voltage at Vi. The points of QV curves are computed by starting from the existing voltage, VO and zero MVAR injection and increasing the voltage until V- is reached or the puwer flow fails to solve. Then the system is reset to the initial condition at V. and the QV computation proceeds in the opposite direction by decreasing the voltage until Vmin is reached or the power Bow becornes unsolvable. Fig.2.5 shows a typical QV curve. The voltage difference V A L and the value of Qc provides the voltage and MVAR stability rnargins at the bus and the dope of the cuve provides the sensitivity information. 2.6 Simulation Results and Discussions In this section, a simple 5 bus system is taken as a case study to ver@ the algorithms discussed in the previous sections. The single line diagram is shown in Fig.2.6 [IO]. The line and bus data for the 5 bus system is s h o w in Appendix A. In this system, bus 1 is the reference ( slack ) bus while buses 2 J,4 and 5 are considered load ( PQ ) buses. The test system is studied without considering any limit on the generators. Fig.2.6 Single line diagram of the 5 bus system The system is dnven from an initial operating point ( base case ) up to the collapse ( bifurcation ) point by changmg the loading factor. Loading factor is a scalar parameter used to simuiate the system load changes that drives the systern from base case to the bifurcation point The tem ""bifurcation" [ i l ] in power systems can be explained as follows: For a load condition, in addition to normai load-flow solution, which is typically the actual operating point or stable equilibrium point ( s.e.p ), several solutions may be found for the load-flow equations. The 'closest" one to the s.e.p. is the unstable equilibrium point ( u.e.p ) of interest for voltage collapse studies [12]. These equilibrium points approach each other as the system is loaded, up to the point when only one soiution exists. If the system is loaded m e r , al1 system equilibria disappear. The "lasf7 equilibrium has been identified as the steady-state voltage collapse point This point is known as saddle-node bifurcation point. At this bifurcation point, the real eigenvalue of the load-fiow Jacobian becomes zero, that is., the Jacobian becomes singular. 2.6.1 Singular value decomposition Simulation results of the singular value decornposition method for the 5 bus system is s h o w in Fig.2.7. In this method, the system could not be loaded beyond a loading factor of 3.3 because of the divergence of the load flow solution. Loading factor of 3.3 corresponds «, a total load of 544.5 MW. .................................................................... ( C . . . . . . . . . . . . . . --.' Fig.2.7 Singular value decomposition for 5 bus system From Fig.2.7, it cm be said that as the loading factor is increased From the base case, the minimum singuiar value decreases and finally approaches zero at the collapse point For example, the minimum singular index is 0.8986 at a loading factor of 3.3 ( near the stability limit ), vev close to the collapse point. Hence, the minimum singular value gives a measure of the n e m e s s to instability or in other words 'distance to collapse'. The system is said to have collapsed when the minimum singular value of the Iacobian is zero. It can be seen nom the above figure that at or near the collapse point, the singdar index is very sensitive to load changes and hence there is a sharp drop in its value. Thus, this index fails when the system is operating close to its limits. Another major disadvantage of this method, is the large computation time required to calculate the minimum singular value for larger systems. 2.6.2 L Index Simulation results of the "Lw index method for the 5 bus system is shown in Fig.2.8 and Fig.2.9. Fig.2.8 L index of individual load buses for 5 bus system Fig.2.9 System L index for 5 bus system From Fig.2.8, as the loading factor is increased, the L index of bus 5 approaches more rapidly towards mity than other buses. When the L index is unity, the system collapses. For instance, near the collapse point ( loading factor of 3.3 ), the L index of bus 5 is 0.7165, while for buses 2.3 and 4 the L indices are O.1927,O. 1 90 1,O. 1926 respectively. The maximum of these is 0.7165. This represents the system L index, which gives a quantitative rneasure for the estimation of the distance of the actuai state of the system to the stability b i t Fig.2.9 shows the system L index and the corresponding bus voltage for the 5 bus system The local L index permits the determination of those buses from which collapse May originate. In the above simulation, voltage collapse originates from bus 5. This is considered as the critical bus for this system. Thus, the stability indicator L is able to characterize the load tlow solution and the potential of the system to become unstable. This is bound to the load fiow mode1 and the assumption of a PQ node. The mode1 does not reflect any dynamic behavior. It is based on single node calculations and is therefore easy to calculate. However, it may require substantial engineering judgment in a large system. 2.6.3 QV curves The farnily of QV curves for the three operating cases: basecase ( Load = 165 MW ), near the stability limit ( Load = 565 MW ), at the point of instability ( Load = 584.53 MW ) are s h o w in Fig.2.10, Fig.2.11, Fig.2.12 respectively. Fig.2.10 Farnily of QV curves for the base case of the 5 bus system Fig.2.11 Family of QV curves near the stability limit of the 5 bus system Fig.2.12 Family of QV curves at the collapse point of the 5 bus system The above QV curves are obtained from a commercially available software VSTAB 4.1 [9]. In this program, lists of buses for which QV curves are to be generated and the s p of each Q V curve must be provided. The span of QV curve is specified by providing the maximum and minimum voltage level at a bus. The program is initiated by assigning the parameters QVCRVS = 1 and SELQVC = FALSE in the parameter file of the VSTAB program. For example, consider the base case operating condition of Fig.2.10. The lists of load buses for which QV curves are to be generated are 2 , 3 , 4 and 5 . The span of each QV curve is specified by providing the maximum voltage level V- = 1.1 p . u and minimum voltage level V,, = 0.2 p.& The step voltage is O. 1 p.u. For the bus 2, the base case voltage V, = 1.0475 p.u. For this bus, QV curves are generated by solving the series of power flow equations 2.6 and 2.7 d l V,, is reached Then the system is reset to the initial condition at V, and the QV computation is carried out in the opposite direction by decreasing the voltage. Here, eventhough V,, is assigued at 0.2 p-u, the p w e r fl ow becomes unsolvable at 0.4 p . u itself Hence, compuîations are stoppeci at this value. Thus, the voltage stability margin V, - V,= 1.0475 - 0.6 = 0.4475 p. u. and MVAR stabiIity margin is -497.06 MVAR. From the above results, the following interpretations can be made. At the base case operating point, al1 buses have significant reactive margin, while at the point of instability, the buses have vimially no reactive margin. The reactive margin is the MVAR distance from the operating point tu the bottom of the curve. There is a significant voltage stability margin for the base case operating point compared to the margin at or near the stability limit. For example, corresponding to the most cntical bus i.e. bus 5, the voltage stability margin for the base case and near the stability lirnit is 0.5 18 p.u. and 0.1205 p.u respectively. QV curves cm thus be used in the assessrnent of voltage stability of the system. Another advantage of this rnethod is the characteristics of test bus shunt reactive compensation can be plotted directly on the QV curve. The operating point is the intersection of the QV characteristic and the reactive compensation charactenstic. This is useful since reactive compensation is ofien a solution to voltage stability problem. On the other han& since the method artificially stresses a single bus, conclusions should be confirmed by more reaiistic methods. Also, the c w e s are obtained by a senes of power flow simulations which makes it more time consuming. 2.7 Limitation of Traditionai Methods in Voltage Stabüity Analysis The steady-state techniques described in the previous section have not found widespread practical applications due to the following shortcomings: they do not provide practical rnargins to measure the stability of the system. they provide little or no information regarding the mechanisrn of instability . they focus on the strength of individual buses and not fkom the system wide perspective. they rnake significant modeling assumptions such as constant generator voltages or constant impedance loads and such assumptions may eliminate factors which ifluence system stability. From the analysis of the traditional rnethods, it cm be concluded that the most important aspect of a practical voltage stability index is the computation effort and hence not usefid for rd-time use in energy management system ( EMS ). Moreover, these rnethods depend largely on conventional power fiows to determine the voltage collapse levels of various points in a network. Furthemore, this approach is laborious, t h e consuming and requires analysis of massive amounts of data as the size of the power system increases. When the power system is operating close to its limits, it is essential for the operator to have a clear knowledge of the operating state of the power system. Thus, for on-line applications, there is a need for tools which can quickly detect the potentially dangerous situations of voltage uistability and provide guidance to steer the system away from voltage collapse. Recently, efforts to improve the speed and ability to h a d e stressed power system have led to the development of intelligent systems. N d for Expert Systems in Eleetric Power System In recent years, research in the field of artificiai intelligence ( AI ) has achieved sigmficant success. Arnong the most signifiant of these is the development of "Expert" or "Knowledge-Based" systems and "Artificial Neural Networks ( ANNs )Y Expert systerns and ANNs are subsets of AI. Both these areas have attracted a widespread interest in the field of power system applications. Reference [13] presents a global view of knowledge engineering applications and tools available in the field of power systerns. Expert systems is a branch of AI that makes extensive use of specialized knowledge to solve problerns at the level of a human expert. That is, an expert system is a cornputer system which emuiates the decision-rnaking ability of a human expert. It consists of two main components: knowledge-base and the inference engine. The expert systems will be treated in detail in chapter 4. Expert systems offer a number of advantages [14]: a As& Human Experts An expert system cm reduce tedious and redundant manual tasks and thereby enhance productivity leading to efficient operation. Flexibility Each production d e represeats a piece of knowledge relevant to the task. Hence it is very convenient to add, remove and modiQ a rule in the knowledge base as experience is gained a Rapidity The expert system can provide more rapid reaction to emergency events than human operators. This is very usefbl in power system operation. in the power system area, expert system applications has been m d y devoted to power systern operations [14]. This is because the operation of power systems has become very complex and operaton are unable to deal with the large amount of data associated with the modem energy management systern. Also, power system operation experience is lost when experienced operators retire or change jobs. It is important to preserve this valuable experience as it is not contained in any textbook or manual. Thus, expert systems can assist in decision-making and minimize erron by human operators. 2.9 Summary In this chapter, the basic concepts, d y t i c a l tools and industry experience related to voltage stability analysis of power systems are reviewed. Advantages and disadvantages of the traditional methods are highlighted by simuiating a simple 5 bus system. From the simulation results, it cm be concluded that the existing methods require significantly large cornputations and are not usefid for real-time use in energy management system. Hence, an altemate approach, such as the applications of artificial intelligent techniques to power systems in general is discussed. To enhance the capabilities of the EMS, an integrated approach of knowledge-based systems ( amficial neural networks and expert systems ) and conventional power system solution methodologies has to be adopted. This approach has the potential to achieve secure and economic operation of the p w e r system. Chapter 3 COlYTINUATION POWER FLOW GND MODAL ANALYSE 3.1 Introduction With the growing concem for voltage instability in the power system industry, much attention has been given to investigating this phenornenon. As a result, a number of techniques have been developed to study the problem. However, well accepted cntena and study procedures regarding system voltage stability do not yet exist Many utilities continue to use pst-contingency voltage decline as an indicator of voltage stability, while some others use performance criteria based on either PV or QV curves. Thus, there is a need to develop a practical procedure which utilities can apply as part of their routine studies. Also, results of voltage stability anaiysis using power flow based static techniques need to be validated using detai led time-domain simuiation tools, such as the Transient- Midtem Stability Programs which are similar to EPR17s ETMSP [15]. Static analysis is based on the solution of conventional or modifiai power flow equations. A popular tool for static based voltage stability analysis is EPRI's voltage stability ( VSTAB ) program [9]. The VSTAB, developed by Ontario Hydro, is a voltage stability assessrnent package for large complex power systems. It provides idormation regarding both the proximity to and mechanisms of voltage instability. The main features of the VSTAB package are: Automatic modification of the system state which includes increasing the load in a predefined manner, creating new dispatches, performing sets of contingencies. Comprehensive voltage stability anaiysis like PV curves, QV curves and modal analysis. Determination of Megawatt ( MW ) or Megavar ( MVAR ) distance to voltage instability, which can indicate modes of instabi lity characterized by groups of buses, branches and generators which participate in the imtability. Capability to obtain the nose of the PV c w e s using the continuation power flow. Steady state approximations to tirne fiarnes associated with ULTC action, governor response, Automatic Generator Control ( AGC ) action and economic dispatc b Voltage stability analysis technique based on the eigen-analysis of the reduced steady state Jacobian rnatrix is referred to as modal analysis. It provides masures of proximity to instability and clearly indicates areas ( modes ) which have potential stability problems. The modal analysis dong with continuation power 80w technique [16,17] have been applied to voltage stability analysis of practicd systems. In the following sections, these two techniques are discussed in detail, supplemented by the simulation resuits of typical test systems using VSTAB and EPRi's Interactive Power Flow ( IPFLOW ) [18] tools. 3.2 Continuation Power Flow Technique A parhcular difficulty encountered in indices mentioned in section 2 . 5 , is that the Jacobian of a Newton-Raphson power 80w becomes singuiar at the voltage stability limit. Consequently, conventional power-flow algorithms are prone to convergence problems at operating conditions near the stability Iirnit The continuation power-flow anaiysis overcomes this problem by reformulating the power-flow equations so that they remain well-conditioned at al1 possible loading conditions. 3.2. l Basic principle This technique was proposed by Ajjarapu and Christy in 1992 [19]. The continuation power-fiow analysis uses an iterative process involving a predictor-corrector scherne as shown in Fig.3.1. Fig.3.1 An illustration of the continuation power flow technique From a known initial solution (A), a tangent predictor is used to estimate the solution (B) for a specified pattern of load increase. The corrector step then detemines the exact solution (C) using a conventional power-flow anaiysis with the system load assumed to be fixed. 3.2.2 Mathematical formulation When the objective is to obtain the maximum loadability point of a system, the problem can be fomulated as ~(6, v ) = AK ( 3 . 1 where, S is the vector of bus voltage angles v is the vector of bus voltage magnitude K is the vector representing percent load change at each bus A is the load parameter Equation 3.1 can be re-written as ~ ( 6 , v , R ) = O In the predictor step, a linear approximation is used to predict the next solution for a change in one of the state variables ( i.e. 6, v, A ). Taking the derivatives of both sides of the equation and evaluating the derivatives at the initial solution, will result in the following set of linear equations in To account for one more equation due to the introduction o f an additional variable in the power flow equations ( namely A. ), one of the components of the tangent vector is set to +1 or - 1. This component is referred to as the continuation parameter. Equation 3.2 now becomes L -1 where, e , is a vector where al1 the entries are zero, except for the variable which is chosen as the continuation parameter, where the entry is ' 1 ' . initially, the continuation parameter is chosen as A. ( load parameter ) and the sign of its slope is positive. However, once the tangent vector is found, the continuation parameter is selected as the bus voltage with the iargest entry in the tangent vector and its sign is used during the next predictor step. Once the tangent vector is found, the next solution can be approxùnated as: The step size a should be chosen so that a solution would exist for the specified continuation parameter. I f for a gïven step size, a solution cannot be found in the corrector step, the step size should be reduced d l a successfiil solution is obtained. in the corrector step, the original set of equations F (6, v, A ) = O is augmenteci by one equation that specifies the value of the state variable selected as continuation parameter. The new set of equations can be written as where x, is the state variable that has been selected as the continuation panimeter and 7 is equal to x" . This set of equations is solved using Newton-Raphson power flow method. The introduction of the extra equation rnakes the Jacobian to be non-singular at the maximum Ioadability point and hence a solution can be obtained at that point. 3.3 Modal Analysis [SI The Modal analysis for voltage stability is briefly discussed below. Consider the linearized power flow equation expressed as wtiere, AP is the incrernental change in bus real power. AQ is the incremenuil change in bus reactive power injection. Aû is the inmemental change in bus voltage angle. AV is the incremental change in bus voltage magnitude. J e , , J , ., J Q e , J are the power flow Jacobian elements. System voltage stability is affected by both real and reactive power. However, at each operating point the real power is kept constant and voltage stability is evaluated by considering the incremental relationship between reactive power and voltage. This is analogous to the Q-V cuve approach. Although incrementai changes in real power are negiected in the formulation, the effects of changes in system load or power transfer levels are taken into account by studying the incremental relationship between reactive power and voltage at different operating conditions. To reduce ( 3.6 ), let AP = O , then J , is called the reduced Jacobian mat* of the system. J , is the matrix which direaly relates the bus voltage magnitude and bus reactive power injection. Eliminating the real power and angle part fiom the system steady state equations allows one to focus on the hidy of the reactive demand and supply problem. Voltage stability characteristics of the system can be identified by computing the eigenvalues and eigenvectors of the reduced Jacobian rnatrix J,. Let, J , = t A v where, 5 = right eigenvector matrix of J, . q = left eigenvector matrix of J , . A = diagonal eigenvalue matrix of J , . and J," = 5 A-' 7 From ( 3 . 7 )and ( 3.9 ), we have, A V = 6 A-' q A Q Each eigenvalue A , and the correspondhg right and left eigenvectors and r ) define the i mode of the Q-V response. Since 5 -' = q , equation 3.10 can be written as v A V = A-' q A Q or v = ~ ' q (3.12 ) where, v = 7 A V is the vector of modal voltage variaticm. q = 9 A Q is the vzctor of modal reactive power variation. Thus for the i" mode, i f A , > 0, the i' modal voltage and the i' modal reactive power variation are along the same direction, indicating that the systern is voltage stable. if A , < 0, the i' modal voltage and the i' modal reactive power variation are along opposite direction, indicating that the systern is unstable. In this sense, the magnitude of R detemiines the degree of stability of the i h modai voltage. The smaller the magnitude of positive A , , the closer the i" modal voltage is to k i n g unstable. WhenA , = O, the i' modal voltage collapses because any change in that modal reactive power causes infinite change in the modal voltage. The magnitude of the eigenvalues can provide a relative measure of the proximity to instability. Eigenvalues do not, however, provide an absolute measure because of the non-linearity of the problem. The application of modal analysis helps in identifying the voltage stability critical areas and elements which participate in each mode. The bus participation factors detennine the contribution of A. , to t h e V-Q sensitivity at bus k They cm be expressed in terms of the left and right eigenvectors of J , as P k i = Çti Bü ( 3 . 1 4 ) Bus participation facton detennine the areas associated with each mode. The size of bus participation in a given mode indicates the effêctiveness of remedial actions applied at that bus in stabilizing the mode. The branch participation fkctors Pji which give the relative participation of b m c h j in mode i are given by - - A Q loss for b m c h j 'J ' max. A Q loss for al1 branches The A Q loss can be caiculated by calcuiating the A V and A 0 change at both ends of the branch. Branch participation factors indicate for each mode, which branches consume the most reactive power in respome to an incremental change in reactive load Branches with high participations are either weak links or are heavily loaded. Branch participations are usefd for identi f j m g remedial mesures to alleviate voltage stability problems and for contingency selection. The generator participation facton Pm which give the relative participation of machine m in mode i are given by - A Q for machine m P m i - ( 3. 16 ) max. A Q for al1 machines The generator participation factors can be used to determine the generators that supply the most reactive power on demand. Generator participation provide important information regarding proper distribution of reactive reserves among al1 the machines in order to maintain an adequate voltage stability margin. 3.4 Simulation ResuIts and Discussions In this section, two typical test systems: the New England 39 bus system [9] and the IEEE 30 bus system [20], are simulated to illustrate the modal analysis along with continuation power flow technique used for the voltage stability evaluation. 3.4.1 IEEE 30 bus system The single line diagram of the IEEE 30 bus system is shown in Fig.3.2. In this system, the load is increased at buses 3,12,21,26 and 30, while generation at buses 2,5,8,11 and 13 are scaled accordingiy to meet the increased demand. In this system, the base case system load corresponding to the selected buses is 45.2 MW. Fig.3.2 Single line d i a m of the IEEE 30 bus system I f the load in this system i s u n i f o d y increased ( constant power factor ), and the power scaled up accordingiy, the PV curve show in Fig.3.3 is obtained fiom which it is seen that the voltage stability limit is 179.38 MW. r i . .......... > V.V O 20 40 60 80 100 120 140 180 180 200 Total Laad et Sclected 8uws (MW) Fig.3.3 PV curve for bus 30 of the IEEE 30 bus system Here, the PV curve is obtained up to the nose point using the continuation power flow technique. On this PV curve, three operating points are identified: f oint A: base case condition LOAD = 45.2 M W . Point B: near the voltage stability limit LOAD = 170.2 MW. Point C: at the voltage stability limit LOAD = 179.3 MW. Now modal analysis is performed at the three operating points to find five least stable modes for each case. Tabie 3.1 shows the results of the analysis for the three operating cases. Table 3.1 Five smallest eigenvaiues for the three operating cases Eigenvaiues 1 From Table 3.1 ., it is seen that the minimum eigenvdue for operating points A, B, C are 0.5145, 0.1538 and 0.0107 respectively. Table 3.2 shows bus, branch and generator participation factors for the base case ( point A ) and at the voltage stability limit ( point C ) corresponding to the least stable mode ( mode # 1 ). Table 3.2 Participation factors ( P.F ) for base case and critical case Operating Point A B C corresponding to the least stable mode. mode # 1 0.5145 0.1538 0.0107 Total Load ( M W ) 45.2 170.2 179.3 mode # 2 1.0644 0.7068 0.5697 Operating Point Point A Point C mode # 3 1.7985 0.9707 0.8492 Bus Participation Bus No P.F Branch Participation Branch No P.F 30 29 26 30 29 4 - 12 9 - I l 28 - 27 27 - 30 28 - 27 Generator Participation Generator No P.F mode # 4 3.7182 0.2 145 0.1940 0.1740 0.4789 0.2463 1 .O000 0.6622 0.569 1 1 .O000 0.9834 8 13 8 mode # 5 4.2255 L 1 .O000 0.5733 1.0000 2.9505 2.3959 3.2571 2.8459 The above analysis shows that as the system load increases, the magnitude of the eigenvalue decreases and at the point of instability, it becomes essentially zero. Table 3.2 indicates that, at the point of instability, buses 30,29 and branches 27-30,28927 are prone to voltage instability. This cm be justified fkom the fact that, for the least stable mode, buses 30 and 29 participate significantly while the rest of the buses do not contribute to voltage instability. This is evident fiom their participation factors. Table 3.3 gives the voltage stability margin for the three operating cases. Table 3.3 Voltage stability margin for different operating conditions 1 Operathg Point ( Total Load ( MW ) Voltage Stability Margin ( MW ) l Voltage stability rnargin ( VSM ) is a measure of how close the system is to voltage instability. It is defined as the difference between the values of a Key System Parameter ( KSP ) at the current operating condition and the voltage stability critical point [lq. Here, KSP is defined as the total load increase in the selected buses. For example, the voltage stability margin for the base case operating condition is 134.1 MW and near the stability limit ( operating point B ) is 9.1 MW. 3.4.2 New England 39 bus system In this section, modal analysis for the New Englcad 39 bus system [9] is perfomed. This systern is widely w d for voltage stability evaluation. The single iine dia- of the New England 39 bus system is shown in Fig.3.4. Here, the load is increased at buses 3.4, 12, 15 and 2 1. Generation at 3 0 , 3 2 , 3 5 and 37 are scaled accordingly to meet the increased demand. Fig.3.4 Single line diagram of New England 39 bus system The PV curve for bus !2 of the New England 39 bus system is shown in Fig.3.5. Table 3.4 shows five least stable modes for the three operating cases. Operating point A corresponds to 1424.5 MW, operating point B corresponds to 4949.5 MW and operating point C corresponds to 4976.8 M W * 0.4 1 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 Total Load at çelected Buses (MW) Fig.3.5 PV curve for bus 12 of the New England 39 bus system Table 3.4 Five smallest eigenvaiues for different loading conditions 1 Point 1 ( M W ) 1 mode 1 mode 2 mode 3 mode 4 mode 5 1 Operating Total Load Point A Eigenvalues Point B 1424.5 10.1145 1 22.5085 1 39.8117 1 45.4208 1 61.3524 4949.5 0.9424 10.5001 16.3547 28.2210 35.0747 Table 3.5 gives bus, branch and generator participation factors for critical case ( at the point of instability ) corresponding to the least stable mode. Table 3.5 Participation factors ( P.F ) for critical case ( point C ) corresponding to the least stable mode ( mode 1 ) From the a b v e results, it c m be concluded that the weakest buses and branch associated with this system are 12, 4, 14 and 10-32 respectively. Table 3.6 shows the voltage stability margin for the three operating cases considered. Table 3.6 Voltage stability margin for the three operating conditions Bus Participation Bus No P.F 1 o p e r a ~ g Point 1 Total Load ( MW ) 1 Voltage Stability Mar@ ( MW ) 1 12 4 14 1 1 t Point A 1 1424.5 I 3552.3 I O. 17 13 O. 1042 0.0949 Branch Participation Branch No P.F 1 1 1 Point B 1 4949.5 1 27.3 1 10 - 32 Generator Participation Generaîor N o P.F I i Point C 4976.8 O 1 1 .O000 32 3 1 3.5 Summary This c hapter has presented a voltage stability assessrnent technique for large power systems using modal analysis in conjunction with continuation power flow technique. The method cornputes a specified number of the smallest eigenvalues of a reduced Jacobian matrix and the associated bus, branch and generator participations. Two typical test systems namely IEEE 30 bus system and New England 39 bus system were used for the purpose of analysis. 1 .O000 0.73 17 The above two examples clearly indicate how the modes represent areas prone to voltage instability. Based on the simulation resuits, the following conclusions can be reached: Each eigenvalue corresponds to a mode of voltage/reactive power variation. - The modes correspondhg to s m a l l eigenvalue represent the modes most proue to loss of stability. - The magnitude of each srnall eigenvalue provides a relative measure of proximity to loss of voltage stability for that mode. Bus, branch and generator participations provide useful information regarding the mechanism of loss of stability. - Bus participations indicate which buses are associated with each mode. - Branch participations show which branches are important in the stability of a given mode. - Generator participations indicate which machines must retaui reactive reserves to ewure the stability of a given mode. The modal analysis along with continuation power flow can be used to determine the voltage stability margin for the base case and a large number of operating cases. Participation factors for the critical case conesponding to the least stable mode is most useful for any remedial actions. It clearly identifies groups of buses and critical bus that participate in the iastability and thereby elirninate the problems associated with traditional methods. Identifjmg the cntical bus will help in taking appropnate control actions. Hence, in the subsequent chapters, this method is used as basis for developing an expert system for voltage stability evaluation. Chapter 4 FUZZY-EXPERT SYSTEMS 4.1 Introduction As rnentioned in chapter two, expert systems are built based on the knowledge acquired nom domain experts. The knowledge of an expert system may be represented in a number of ways. One common method of representing knowledge is in the fonn of IF-THEN type d e s , such as IF the light is red THEN stop If a fact exists that the light is red, this matches the pattern "the light is red". The nile is satisfied and perfonns its action of "stop". Expert systems can be considered declarative programming because the programmer does not specifjr how a program i s to achieve its goal at the level of an algorithm. Expert systems are generally designed very differently fiom conventionai programs because the problems usuaily have no algorithme solution and rely on inferences to achieve a reasonable solution. The strength of an expert system can be exploited fully when it is used in conjuncàon with a database. Changes in the system operating conditions are reflected in the database. The expert system that draws its data fiom the database automatically tracks the system operating conditions. The purpose of this chapter is to introàuce the basic theory of expert systems and funy logic. The concept of membership h c t i o n plays an important role in the design of the proposed fiiW-expert system for voltage stability monitoring and control. Hence, it is relevant to explore these concepts in detail to understand its applicability to power system problems. One of the objective of this thesis is to arrive at a solution that is fasf reliable and usefùi for the power system operators. This chapter will lay a foundation in understanding the proposed fbzzy-expert system in order to achieve the desired objective. in this chapter, the theory of expert system is described in section 4.2. The fundamental concepts of f u n y set theory is explained in section 4.3. Section 4.4 highlights potential applications of fuzzy-set-based approaches and their relevance to power system problems. 4.2 Expert System Structure The structure of an expert system in a general block diagram i s shown in Fig.4.1 [2 11. Expert System Fig.4.1 Expert system structure As shown in Fig-Cl., the expert system consists of two cornponents. The knowledge-base and the inference engine. The kuowledge-base comprises knowledge that is specific to the dornain of application, incluâing simple facts about the domain, d e s that describe relations or phenornena in the domain and hewistics. The inference engine actively uses the knowledge in the kaowIedge-base and draws conclusions. The user interface provides smooth communication between the user and the system. It also provides the user with an insight into the problem-solving process executed by the inference engine. It is convenient to view the inference engine and the interface as one module, usually called an expert system shell. Expert systems are often designeci to deal with uncertainty because reasoning is one of the best tools that have been discovered for dealing with uncertainty. The uncertainty rnay arise in the input data to the expert system and even the knowledge-base itself. At first this rnay seem surprising to people used to conventionai programming. However, much of human knowledge is heuristic, which means that it rnay oniy work correctly pari of the time. In addition, the input data rnay be incorrect, incomplete, inconsistent and have other errors. Algorithmic solutions are not capable of dealing with situations like ths. Depending on the input &ta and the knowledge-base, an expert system rnay corne up with the correct answer, a good answer, a bad answer or no answer at all. While this rnay seem shocking at first, the alternative is no answer al1 the tirne. 43.1 Fnayset theory The theory of uncertainty is based on funy logic. The traditional way of representing which objects are membea of a set is in terms of a characteristic function, sometimes called a discrimination fiuiction If an object is an element of a set, then its characteristic function is 1. If an object is not an element of a set, then its characteristic function is O. This definition is summarized by the following characteristic fûnction: pA(x) = 1 if x is an element of set A O if x is not an element of set A where the objects x are elements of some universe X. In funy sets, an object may belong partially to a set The degree of membenhip in a fwzy set is measured by a generalization of the characteristic fiinciion called the membership function or compatibility function defined as The characteristic function rnaps al1 elements of X into one of exactly two elements: O or 1 . In contrast, the membership function maps X into the codomain of red numbers defined in the interval fiom O to 1 inclusive. That is, the membership fûnction is a reai number 0 1 p , S l where O means no membership and 1 means full rnembership in the set. A particuiar value of the membership hction, such as 0.5, is called a grade of membership. For example, if a person is an adult, then myone about 7 feet and tailer is considered to have a rnembership fhction of 1.0. Anyone Iess than 5 feet is not considered to be in the fuzzy set TALL and so the membenhip function is O. Between 5 and 7 feet, the membenhip h c t i o n is monotonically increasing with height ï h i s paticular rnembership h c t i o n is ody one of many possible functions. The membership h c t i o n for the Fuzzy set TAU is shown in Fig.4.2 Fig.4.2 Membership function for the fuay set TALL The S-function and ri function are two important mathematical hnctions that are often used in fuzLy sets as membership bctions. They are defhed as follows: for x < a 2(%) for a s x l / 3 for x~ y Fig.4.3 The S-f'iinction Fig.4.4 The ïI - h c t i o n The II-bction is shown in Figure 4.4. Notice that the ,8 parameter is now the bandwidth or total width at the crossover points. The Il -function goes to zero at the points x = y t f l while the crossover points are at Funy logic forms the basis of funy-expert systems. In fupy-expert system, the knowledge is captured in natural language which have arnbiguous meanings, such as tall, hot, danprous and so on, which is concerned with the theory of uncertainty based on fuPy logic. The theory is primarily concemed with quantifying the linguistic variable into possible fuzy subsets. A linguistic variable is assigned values which are expressions such as words, phrases or sentences in a naniral or artificial language. For exarnple, the linguistic variable "height" has typical values like "dwarf", "short", "average", "tall", tbgiant". These values are referred to as funy subsets. Every element in these fuPy subsets has its own degree of membership. Besides dealing with uncertainty, furzy-expert systems are also capable of modeling cornmonsense reasoning, which is very dificult for conventional systems. Fuzzy logic is the logic of approximate reasoning. Essentially, approximate or fuzsr reasoning is the inference of a possibly imprecise conclusion fiom a set of possibly imprecise premises. One of the important type of fuzzy logic is based on Zadeh's theory of approximate reasoning 1221, which uses a fuzzy logic whose base is Lukasiewicz L , logic. In this fuzzy logic. truth values are linguistic variables that are ultimately represented by fuey sets. Fuzzy logic operators based on the Lukasiewicz operators are defmed in the Table 4.1. x(A) is a numeric tmth value in the range [ 0,l ] representing the tnith of the proposition " x is A", which c m be interpreted as the membership grade p, (x) . Table 4.1 Fuzzy logic openitors 4.4 Application of Fuzzy-Set Theory to Power Systems One of the fiindamental objectives in the management of power systems is to provide safe and reliable electric power at the lowest possible cost. To achieve this objective, rapid advances in the contml and management technology of power systems has been made. For example, ABB's Energy Management Systems ( EMSYS ) [23] is an innovative, cornputer-based information and control system designed to provide full range of utility solutions, fiom basic SCADA to advanced transmission system network security and distribution automation applications. During this process, power systems have become even more complex in structure and statu. This growing complexity is causing problems to power system operaton. Some of the more evident problems are: rapid increase in the number of real-tirne messages has made operator response more difficult. current numerical processing software m o t meet the operational requirements of power systems in some situations. Typical example is processing difficulty during emergency conditions. moa design, planning and control problems encountered are complex and time consuming because of multiple objective functions, multiple constraints, and complex system interactions. Analytical solution methods exist for many power systems operation, planning and control problems. However, the mathematical formulations of real-world problems are derived under certain restrictive assumptions and even with these assumptions, the solutions of large scale power systems problems is not trivial. On the other han& there are many uncertauities in various power systems problems because power systems are large, complex and infiuenced by unexpected events. These fxts make it difficult to effectively deal with many power systems problems through strict mathematical formulations alone. Therefore, expert systems apjxoaches have emerged in recent yean as a complement to mathematical approaches and have proved to be effective when properly coupled together. Reference [24] gives a bibliographical survey of the research, development and applications of expert systems in electric power systems. As already mentioned, expert systems are built based on the knowledge acquired fiom domain experts. The expert's empirical knowledge is generally expressed by language containing ambiguous or fuzy descriptions. As a resdt, a number of researchers have applied fuay logic concept to power sy stem applications. Some of the applications include fuzzy approach to power system security [25], dynamic generation rescheduling [26], reactive powerhroltage control [27l, short-term load forecasting [28], prioritiüng emergency control [29], contingency consaained optimal power flow [30]. A more comprehensive list of applications of fuPy set theory to power systems is aven in [3 11. When fuzzy set theory is used to solve red problems, the following steps are generally followed [3 1 1: step 1 Description of the ori@ problem. The problem to be solved should first be stated mathernatically/linguistically. step 2 Definition of thresholds for variables. step 3 Furry quantization. Based on the threshold values from step 2, proper foms of membership fwictions are coflstfllcted Many forms of membership fiuictions are available, such as linear, piece-wise linear, trapezoidal and parabolic. The membenhip hctions should refiect the change in degree of satisfaction with the change in variables evaluated by experts. step 4 Selection of the proper fuzzy operations, so that the results obtained is like those obtained by experts. 4.5 Summary This chapter has reviewed some of the hdamental concepts of expert systems and fuPy logic. It introduced the generic expert system çtnicture and the theory of approximate reasoning. The S and n functions were highlighted Funy sets has the potentiai to play an important role in power system operations and control. Some of the drawbacks of conventional expert systems [32] are They process sequential procedures by matching a set of d e s to execute an operation for a given system condition. For firing rules, a complete matching of predetemined conditions for a given input is needed, resulting in limited effective operation when applied to a practical situation. a Lack of practical knowledge and suitable means to represent heurîstics. Funy-expert systems based on approximate reasoning overcomes the above mentioned problems. The following are the advantages of funy-expert systems over conventional expert systems. a It offers flexible processing of lmowledge expressed by d e s . In approximate reasoning, the attributes included in the rules are given by linguistic variables. A linguistic variable measures the proximity of a given value to a fuay set by the grade of membership to the set. Such grading of attributes are known as uncertainty factors. Approximate reasoning pennits mdti-attnhte evaiuation of an input because every condition included in a d e has a numerical value, rather than tnie or false state as in conventional expert system. In the overall context of the research, this chapter introduces the theory behind the proposed fuay-expert system for voltage stability monitoring and control. Chapter 5 FUZZY CONTlROL APPROACH TO VOLTAGE PROFILE ENHANCEMENT FOR PO'WER SYSTEMS 5.1 Introduction The application of funy set theory to power systems is a relatively new area of researck Chapter 4 has highlighted the basic pnnciples involved in this a m However, before attempting to develop a fuay concept for voltage stability evaluation, the f ù z q set theory is first applied to voltage- reactive power control for power systems. As the voltage profile of the electric power system could be constantly affecteci, either by the variations of load or by the changes of network configuration, a real-time control is required to alleviate the problems caused by the perturbations. The problem is how to accurately compose a voltage control stmtegy during emergency conditions when complete system information is not available. To overcome this problern, adjustments to the control devices are needed af€er the perturbations to alleviate voltage limit violations. This can be achieved by determining the sensitivity coefficients of the control devices. Several papers in the literature explore voltage/reactive power control by means of fûzq sets [33], heuristic and algorithmic [34], rule-based systems [35] etc. in this chapter, a voltage-reactive power control model using fuPy sets is described which aims at the enhancement of voltage security. In this model, two linguistic variables are applied to measure the proximity of a given quantity to a certain condition to be satisfied. The two linguistic variables are the bus voltage violation level and the controlling ability of control devices. These are tnuislated to fuay domain to formulate the relation between them. A feasible solution set is first attained using min- operation of fiiw sets, then the optimal solution is determined employing the max- operation. The method was proposed by Ching-Tzong Su and Chien-Tung Lin [36]. 5.2 Problem Statement In power system operation, any changes to system topology and power demand can cause a voltage violation. When a voltage emergency has been identifie4 control actions mut be initiated, either automatically or manually, to alleviate the abnomai condition. The reaction time is critical. For example, when a destructive storm passes through an area, communication systems can be disrupted, reducing the information received at the system control center. If bus voltages are beyond desirable limits during such an emergency, the voltage control problem cannot be solved by conventional methods i.e., using load flow solution techniques. This is due to incomplete information needed to construct the system network modeI. Optimal control of voltage and reactive power is a significant technique for improvements in voltage profiles of power system. The objective is to improve the system voltage profile, such that it will lead as closely as possible to the desired system condition. The network constraints include the upper and lower bounds of bus voltage magnitudes as well as the adjustable minimum and maximum quantities of controlling devices. When a load bus voltage violates the operational limits, control actions m u t be taken to alleviate the abnormal condition. Consider an N-bus system, with buses 1 to L as load buses, buses L+ 1 to N-1 as generator buses. By adjusting the controlling device on bus j, the voltage improvement of bus i is given by AVi = S i j A U j ( 5 - 1 ) 1 2 . L ; j = 1 2 ,......, N-1 where, AV, isthevoltagechangeofbus i. S i is the sensitivity coefficient of bus j on bus i. A U is the adj ustment of controlling device at bus j . Sensitivity coefficient in general gives an indication of the change in one system quantity ( eg: bus voltage, MW flow etc. ) as another quantity is varied ( VAR injection, generator voltage magnitude. transfomer tap position etc. ). Here, the assumption is that as the adjustable variable is changecl, the power system reacts so as to keep al1 of the power flow equations solved As such, sensitivity coefficients are linear and are expressed as partial derivatives. For exarnple, expression 5.2 represents the sensitivity of voltage at bus i for reactive power injected at bus j. Adjustment of the controlling device is constrained as m i n < AU, - A U j I A UJ'" ( 5.3 ) where, A Ur " and A Ur ' " represent the adjustable minimum and maximum reactive power. The Ioad bus voltage deviation should be conûolled within I 5% of the nominal voltage Vn O . which can be expressed as yin 5 vi vpiax where, ymi" ( = 0.95 V n o m )and Tax ( = 1.05 VnO" ) are the lower and upper voltage limits of bus i respectively and Vi is the input of the system. 5.3 Fnzzy Modeling Two linguistic variables namely, the bus voltage violation level and the controlling ability of controlling devices are modeled in the fuzzy domain as folIows 53.1 Bus voltage violation level The membership function of the bus voltage violations is shown in F i g 5 1 . The maximum deviation of the bus voltage is A V," ' " = Vy ' " - VO " " , the minimum deviation of bus voltage is A v,"' " = V," ' " - VoO " . V i m a x , V , m i n and v n o m take the values 1.05, 0.95 and 1.0 p.u. Here, it is desirable to control the bus voltage deviation within f 5% of the nominal voltage. 0 i 408 4 0 6 O M 4 0 2 0 0 0 2 0 0 . 0.06 0 0 6 0 1 Daviatioo of Bua Voltage AV @.a) F i g 5 1 Fuzzy model for bus voltage violation level 5.3.2 Controüisg ability of controllhg devices The funy representation of the controlling ability of the controlling device is show in Fig.5.2. The controlling ability is Cij = Si, M, ( 5.5 where, M , is the controlling margin of the controlling device ai bus j . Si is the sensitivity coefficient of bus j on bus i. Ci is the controlling ability for controlling device of bus j on bus i. Fig-5.2 F u ~ y mode1 of controlling ability of controlling device 5.33 Control Strategy The violation of bus voltages and the controlling ability of the controlling devices are first fupified with the fuPy models defined in Fig.5.1 and Fig.5.2. To alleviate the bus voltage violatioc a suitable control strategy is adopted The strategy coosists of selecting an optimal controlling device and the adjustment of that device. The two most effective watrol devices are Static Var Compensators ( SVC ) and generator controls. Since, voltage violation is a local phenornena, SVC's are given a greater priority compared to generator controls [37l. ïhe control strategy coosists of three steps as descnbed below Let p and p be the membership value of voltage violation and controlling ability of the controlling device respectively for a pam'cular controlled bus i. Select a controlling device j with conesponding membership value p of controlling ability, take the min- operation R i j = m h ( ~ J * , / f c ) ( 5.6 Repeat the above min- operation for ail controlling devices. Take the max- operation to the j terms of R,, obtained above. K,, = m a x ( R i i , R ,,...., RÏ) ( 5-7 ) where, j represents the total number of controlling devices. O For each of the L controlled buses, repeat the min-max operations L terms of K i , are obtained Take the max- operation totheLtermsof K i , . Pi, = rnax(K,,, K,, . . . ., K, ) ( 5 - 8 where, Pi, represents the membenhip value of controlling ability for controlling device at controlling bus j on controlled bus i. Note that the above three operations are to be done independently for SVC and generator controllers. 5.4 Methodology The following are the computational steps involved in the film control approach for the voltage profile enhancement. For the given system and loading conditions, perfom the base case load flow using a comrnercially available software package, for example, the Interactive Power Flow ( I P E O W ) version 4.1. IdentiQ those affected load bus voltages that violate either the upper or lower bound voltage limits. For the available controllers, find their sensitivity coefficients. Calculate the available control margin which is the difference between the present setting and the maximum possible setang of a parhcuiar controller. Find the mernbership value of bus voltage violation level and coatrolling ability. Evaluate the optimal control solution as described under control strategy in section 5.3. ModiQ the value of the control variables. Perform the load flow study and output the resdts. 5.5 Simulation Results and Discussions To verie the above method, a modified IEEE 30 bus test system is taken as a numerical example. Modifications are made to the original IEEE 30 bus system at bus 13 and bus 1 1. Their pre-modified initial terminal voltages are 1.071 and 1.082 p.u respectively. Base case voltage profiles of the modified IEEE 30 bus system are shown in Table 5.1. In this system, the reactive power sources are assumed to be at buses 10, 19, 24, 29 and generator voltage regulators at buses 2, 5, 8. The following three cases are investigated case 1: Load is increased at bus 26, causing bus 26 to violate the voltage constraint, but the violation is not serious. case 2: Load is increased at buses 7 and 15 causing voltage violation at buses 15, 18 and 23. case 3 : For load increase at bus 15 and outage of line 12- 15, buses 15 and 23 violate the voltage constraint. Table 5.1 Base case voltage profiles of the modified IEEE 30 bus systern Table 5.2 shows the lower and upper limits of the available controllers. control action is initiated are 0,0,0,0, 1 - 0 6 , 1 .O 1, 1 .O 1 respectively. Table 5.2 Lower and upper limits of the controllers Controller Type Q 10 Q 19 Q 24 Q 29 V2 V5 V8 Lower Limit ( p.u ) -1.0 - 1 .O -1.0 -1.0 0.95 0.95 0.95 Upper Limit ( p.u ) 1 .O 1 .O 1 .O 1 -0 1 .O5 1 .O5 1 -05 Table 5.3 compares the load bus voltage profiles before and after the control actions for the three cases considered and Table 5.4 shows the optimal fuPy control solution. Table 5.3 Load bus voltage profiles before and after control actions Load Bus No Case 3 VoItages ( p.u ) initial F i 1 Case 1 Voltages ( p.u ) initial Final Case 2 Voltages ( p.u ) Initial Final Table 5.4 Optimal fuay control solution Finai Value of the The results for case I in Table 5.3 show that the load increase at bus 26 causing bus 26 to violate the voltage limits. The lower and upper voltage limits are 0.95 and 1.05 p.u respectively. The voltage violation at bus 26 is 0.007 1. From Fig. 5.1 ., the membership value corresponding to the voltage violation, which in this case is 0.142, is obtained. Then, for al1 the available controllen, from the sensitivity anal ysis and fiom controlling margin of the controlling devices, the controlling ability of the controlling devices ( Ci ) is detertnined. From Fig 5.2, the membership values for these controlling abilities are obtained. Findly, by applying min-max operations, controllers 429 and V8 are obtained as indicated in Table 5.4. To obtain the final value of the controller 429, sensitivity analysis is perfomed. The sensitivity coefficient of the controller 429 with respect to bus 26 is 0.0024. Hence, the final value of Q29 is 0.0295 p-u. If there is a voltage violation at more than one bus as in case 2, then the above procedure has to be repeated for each violated bus independently. This is referred to as layer-1 operation. Then fiom the selected controllen of layer-1, max operation is perfomed to anive at the proper solution This is known as layer-2 operation. 5.6 Summary In this chapter, a simple method using fuPy sets for the voltage-reactive power control to improve the system voltage level is presented. The control strategy is obtained by employing max-min operatiom. A modified EEE 30 bus system is used as an example to illustrate this m e t h d From the simulation results, it can be iderred that fuPy models are indeed effective in power system control applications. One of the desirable feature of this fuzzy mode1 is that if the operator is not satisfied with the grading of the fuzzy model, the operator can adjust the parameters associated in the definition of the membership function to suite the needs of the desired system operation. Chapter 6 F'UZZY-EXPERT SYSTEM FOR VOLTAGE STABlLITY MONITORING AND CONTROL 6.1. Introduction In the preceding chapters, considerable attention bas been aven to the concepts of voltage stability, modal analysis technique, fuzzy-expert systems and their application to typical test systems. Simulation results of these test systems are encouraging and have ken a motivation to investigate this concept for the voltage stability problem. In this chapter, a new fuPy-expert system is proposed for voltage stability monitoring and control. The phenornena o f voltage sîability can be attributed to slow variation in system load or large disturbances such as loss of generators, transmission lines or transformers. The impact of these changes leads to progressive voltage degradation in a significant part of the power system causing instability . In the context of real-time operation, voltage stability analysis should be perfomed on-line. A nurnber of special algorithms and rnethods have been discussed in chapter 2. However, these methods require significant computation time. When the power system is operating close to its limits, it is essential for the operator to have a clear knowledge of the operating state of the power system. For on-line applications, there is a need for tools that can quickly detect the potentially dangerous situations of voltage instability and provide guidance to stem the system away from voltage coilapse. In an effort to improve the speed and ability to handle stressed power systems, a funy-expert system approach is proposed A number of researchers have made an attempt in this direction. An expert system prototype was developed to correct for voltage steady state stability [38]. In this approach, an expert system arrives at a fast solution by considering an overall system threshold indicator to decide on the degree of VAR shortage and the vulnerability of the system to voltage instability. Yuan- Yih Hsu and ChungChing Su [39] proposed a de-based expert system for srnall-signal stability analysis. They developed an efficient on-line operational aid, wherein the expert system perfoms deductive reasoning to arrive at the degree of stability without the need to calculate the eigenvalues. In this chapter, a fiiny-expert system approach to steady- state voltage stability monitoring and control is proposed. The results of chapter five have been a usefbi starhg point for this new approach- The proposed expert system evaluates system state through deductive reasoning by operating on a set of fÙzzy rulebase. The funy d e b a s e is system dependent, but once fomed, it cm handle any operating condition. An integrated approach of expert systems and conventional power system solution rnethodologies has the potential to provide real-time monitoring and control. The chapter is organized as follows. Section 6.2 highlights the concepts of fuzy-expert system, design of database and debase specifically for the New Englaad 39 bus system. Sections 6.3 and 6.4 present the inference engine and simulation results of the New England 39 bus system respectively. Section 6.5 provides a summary of this new approach. 6.2 Expert System and Design Expert system consists of two main components. The knowledge-base and the inference engine. A knowledge-base is a collection of facts and d e s describing how the facts are linked Based on the facts, the expert system draws conclusions by the inference engine. rii fiizzy-expert system, the knowledge is captured in nahnal language which have arnbiguous meaoings, such as tall, hot and dangerous, and is concerned with the theory of uncertainty based on fwry logic. The theory is primarily concemed with quantimg the linguistic variable say % i l " into possible fuzzy subsets like "low", "high", "medium". Every element in these fûzzy subsets has its own degree of rnembership. The main reasons for the use of funy logic to voltage stability m o n i t o ~ g and control are The approach is simple, straightfoward and fast, where it only needs key system variables to arrive at the solution state. Due to the veiy nature of the problem, the imprecision of the linguistic variables can easily be transferred into funy domain, which would otherwise be difficult to manage. Fuzzy logic can hancile non-linearity of power system problem and does not require complex cornputations as in traditional methods. Thus, it is more efficient than conventionai methods for voltage stability anaiysis. In the proposed fkq-expert system, the key system variables like load bus voltage, generator MVAR reserve and generator terminal voltage which are used to monitor the voltage stability are stored in the database. Changes in the system operating conditions are reflected in the database. The above key variables are funified using the theory of uncertainty. The debase comprises a set of production rules which form the basis for logicai reasoning conducted by the inference engine. The production d e s are expressed in the form of IF-THEN type, that relates key system variables to stability. The strength of the above fùzzy-expert system can be exploited Mly when the rulebase is used in conjunction with the database. The modal analysis technique dong with continuation power flow discussed in chapter two, play an important d e in the development of the proposed fuzzy-expert system. It is not a part of the fuzzy-expert system. It is used as an aid in the formulation of database and as a benchmark tool for validating the accuracy of the proposed approach. 6.2.1 The Database The key variables identified are load bus voltage, generator MVAR reserve and generator temiinal voltage. The three key variables are selected based on the solution obtained by repeated load flow and modal analysis performed for various operating conditions. Based on the simulations carried out for different loading factors, it is found that load bus voltage and reactive generation reserve are significantly affected as the system approaches collapse point for a specific loading pattern. Aiso, the terminal voltage of the generaton has an effect on voltage stability margin. Hence, the above three variables are used to monitor the voltage stability of the system. Each of these variables is M e r divided into four linguistic variables and then transfonned into fuzzy domain: Low, Tolerable, Moderate and Safe. The membership f'unctions for these linguistic variables are of the general form given by where, K may be any one of the key variables. A and a are constants. The membenhip fiuictions of the key variables are s h o w in Fig.6.1, Fig.6.2 and Fig.6.3. Fig.6.l Membenhip fiuiction for the worst load bus voltage. Lem 0.8 ......--...... Fig.6.2 Membenhip h c t i o n for the worst MVAR reserve. Ganeretor Terminal Voltage (p-u) Fig.6.3 Membership function of the generator terminal voltage corresponding to the worst MVAR reserve. For example, in Fig.6.1 corresponding to the load bus voltage of 0.6 p.u, the fuay sets Low, Tolerable, Moderate and Safe have membership grades 1.0, 0.2, 0.1 and 0.02 respectively. The union of the above hiay sets represents the total uncertainty in the stability of the system. The membeahip functiom of the key variables can take the form monotonie, bell-shaped, trapezoidal or even tnangular. The selection of the type, depends on the application and how ciosely it c m descnbe the system behavior. In the following studies that involve non-linear equations, bell-shaped coupled with S-shape is the closest that describes the behavior of power system in general. The constants 'a" and "A" are selected such that the membership fiinction covers the entire range of the key variables. For example in Fig.6. l., the rnembeahip h c t i o n of the four linguistic variables should cover lower limit ( 0.6 p.u ) and upper limit ( 1 .O p. u ). Panuneten "a" and "A" are selected such that the desired stability condition is çatisfied for the test cases. In this way, proper membership grades are obtained. In an utility environment, the peak values of these linguistic variables of the key variable is obtained from the operator's experience [39]. In the studies supported here, peak values have been found by trial and error. Extensive simdations are performed to identiQ the range of key variables. The constants "ay7 and "A" in equation 6.1 are selected appropnately to give proper membership grades. For instance, the parameters "a" and "A" for the rnembenhip h c t i o n of the worst load bus voltage, worst MVAR reserve, generator terminal voltage corresponding to the worst MVAR reserve are shown in Table 6.1. Table 6.1 Parameters a and A for the mernbenhip function of the key variables. Linguistic Variables 1 1 For membership hction of the worst load bus voltage 1 L a ( MVAR) [ 8000 1 8500 1 9250 1 9500 Low L 1 1 For rnembership bction of the generator terminal voltage a ( P-u 1 A ( Phu Having represented the inputs to the expert system as linguistic variables, the output of the expert system is the degree of voltage stability represented by four linguistic variables: Very stable ( VS ), stable ( S ), critically stable ( CS ) and unstable ( US ). The four linguistic variables correspond to the following situations. VS: Â. > 10.0 S : 10.0 2 A > 6.5 CS: 6.5 >, A. > 2.0 us: A 5 2.0 where,A is the minimum of the absolute real part of the eigenvalues of the reduced Jacobian matrix J , of equation 3.7. Toierable For membership fiuiction of the w o m MVAR reserve 0.7 0.06 Moderate Safe 0.8 O. i 0.9 0.1 0.95 O. 06 6.2.2 The Rulebase Two separate sets of rulebase are formed, one for monitoring and other for control stage. In t h s stage, the rule base is divided into four groups. group 1:ruies that relate worst load bus voltage to stability measure. group 2:rules that relate key generator MVAR reserve to stability measure. group 3:niles that relate key generator terminal voltage to stability measure. group 4:rules that combine stability masures of the above three parameters. Typicai d e for groups 1 J and 3 is of the fom: IF K is X1 THEN S is X2, with rnembership value p (XI, X2)., where K is any one of the key variables. X1 is any one of the four possible fuey subset ( low, tolerable, moderate, safe ) characte~ng the key variables. S is membership grade. X2 is any one of the four possible fuay subset (VS, S, CS, US) characterizing stability measure. There are 48 d e s relating key variables to stability measure, 16 d e s for each variable. Among the 16 rules, there are 4 rules which will yield g ( VS ). The resdtant p ( VS ) is obtained by max-min compositional rule of inference as follows The membership values for p ( S ), p ( CS ) and p ( US ) c m be derived similarly. For group 4, the stability measures derived for wont load bus voltage ( psb (VS),psb (S),jsb (CS) and psb (US) ), worst generator MVAR reserve ( psm (VS), psm (S), psm (CS) and pm (US) ), corresponding generator terminal voltage of wont MVAR reserve ( psg (VS), psg ( S ) , psg (CS) and psg (US) ), are combined together using max-min composition to yeild the overall stability of the power system. Thus, in this group, there are 256 d e s . The stability rneasure with the greatest membership value gives the state of the systern. The procedure is exactly the same as mentioned in the monitoring stage, but the only difference is that the number of inputs to the expert system is two instead of three. Here, the generator terminal voltage is not taken as one of the key input variable because of its value k i n g fïxed throughout the control phase. This stage has 96 rules. The reasoning process of the inference engine that involves both monitoring and control is implemented in MATLAB [40]. Appendix B shows the rules relating key variables to stability rneasure. Appendix C shows the data used for the formulation of the nilebase with reference to the monitoring stage. 6.3 Inference Engine The inference engine takes the key variables such as load bus voltage, reactive generation reserve and terminal voltage of the generator as its inputs and uses the production d e s to perfom deductive reasoning. nie steps involved in the reasoning process are s h o w below Monitoring stage: (1) Read in load bus voltage, reactive generation reserve and terminal voltage of the generator. (2) F d f y each of these key elements into four linguistic variables using equation 6.1 to obtain their membership grades. (3) input the membership grades into the fuzzy debase. (4) From the niles formed under group 1, obtain the stability measures corresponding to the worst load bus voltage. ( 5 ) From the d e s formed under group 2, obtain the stability measures corresponding to generator MVAR reserve. ( 6 ) From the rules formed under group 3, obtain the stability measures corresponding to key generator terminal voltage. (7) From the niles formed under group 4, evaluate the global stability membership grade of the system. lf the system state is either critically stable or unstable, then perform the control action. Control stage: (8) To improve the voltage profile of the worst load bus, VAR compensation andor raise in generator terminai voltage is performed at the exisring wntrollers. ( 9 ) Fuzzie these new improved values and combine with stability d e s to obtain the new global stability state. 6.4 Simulation Reaolts and Discussions To show the effectiveness of the proposed fuay-expert syçtern, the New England 39 bus system is taken as an example. The single line diagram of the system is shown in Fig.3.4. Here, the load is increased at buses 3, 4, 12, 15, 21 and generation at buses 30, 32, 35 and 37 are scaled accordingly to meet the increased demand. The simulation is camed out under six different operating conditions. condition 1 : No contingency and generator 32 terminal voltage is maintaineci at 0.983 1 p.u. condition 2 : No contingency and generator 32 terminal voltage is maintained at 0.95 p.u. condition 3 : No contingency and generator 32 terminal voltage is maintained at 1-05 p.u. condition 4 : line 6-1 1 outage. condition 5 : generator 36 outage. condition 6 : line 6-1 1 and generator 36 outage. To simplify the analysis, only six conditions are considered. In reality, for the proposed fuay-expert systern to be valid, al1 possible conditions have to be taken into accowzt. Note tiiat some of the worst cases have been considered to validate the correctness of the proposed system. Extensive number of operating conditions are tested to verifL the correctness of the expert system output during the monitoring stage. A cornplete list of expert system output for al1 cases is show in Appendk D. Table 6.2 below shows the comesponding operating condition for the 32 cases listed in Appendix D. Each case is for a specific Ioading condition. Table 6.2 Operating conditions for the 32 cases listed in Appendix D Table 6.3 Expert system output - Monitoring Index No 1-7 Operathg Condition 1 I CONTINGENCY Index No NO CONTINGENCY Gen.32 Volt (p.u) Bus 12 Load Volt b u ) Gen32 M'VAR Global state VSTAB output Eigen System vaiue state For the purpose of analysis, consider the eleven cases s h o w in Table 6.3. The expert system output for these eleven operating cases correspond to the monitoring stage. Here, the fuzzy-expert system output ( Global state ) is compared with the simulation results given by VSTAB 4.1 output through modal anaiysis. In Table 6.3, index Nos.1 and 2 correspond to operating condition 1, index No.3 refers to operating condition 2, index Nos.4 and 5 correspond to operating condition 3, index Nos.6 and 7 correspond to operating condition 4, index Nos.8 and 9 correspond to operating condition 5 and finally index Nos.10 and 11 correspond to operating condition 6. As shown in Table 6.3, the input variables to the expert system are bus 12 load voltage, generator 32 MVAR reserve and generator bus 32 terminal voltage. The basis for their selection is from îheir m c i p a t i o n factors described in chapter 3. Appendix E shows the participation factors for the critical case ( at the voltage stability limit ) for each of those six conditions. As seen in Table 6.3, there are some operating cases that are either unstable or cntically stable. These are the cases that need control. To show the difference between conventional methods and the proposed approach, consider the index No.7 of Table 6.3. The input variables to the fupy-expert system are bus 12 load voltage = 0.6059 peu, generator 32 terminal voltage = 0.983 1 p.% generator 32 MVAR reserve = 8 102.4 MVAR* From Fig.6. l., the membership function of the linguistic variables - low, tolerable, moderate, safe for the worst load bus voltage are mu-Io = 1.0 mub_tol= 0.2098 mub-mod = 0.1036 mub - saf= 0.0295 From Fig.6.2., the membenhip fiuiction of the linguistic variables for the worst MVAR reserve are muv 10 = 0.8657 - muv-tol = 0.2995 muv-mod = 0.0295 muv-saf = 0.020 1 From Fig.6.3., the membership function of the linguistic variables for the generator terminai voltage are m e m g l o = 0.0836 merngtoI= 0.5497 ( 6.6 m e m g m o d = 0.8936 merngsaf = 0.0684. Once the membership fùnction of the linguistic variables are determined, the next step is to relate these variables to the stability measure. From the d e s formed under group 1, the stability measure for the very stable (VS) case is obtained as follows vsb 1 = min ( mub-10, murb (4,l ) ) vsb2 = min ( mub-tol, murb (3,l ) ) vsb3 = min ( mub-mod, murb (2,1 ) ) vsM= min ( m u b - d , murb (1,l ) ) where, murb matrix gives the required d e s relating the load bus voltage to the stability measure of group 1 as shown in Appendix B. Thus, the overall stability measure for 'tery stable" is given by mb - vs = max ( vsbl, vsb2, vsb3, vsb4 ) = 0.1036 ( 6-8 SimiIarly, the stability measures for stable, critically stable and unstable cases are obtained as mb-s = 0.2098 mb-CS = 0.5 ( 6-9 mb-us = 1 .O Following the same procedure as mentioned above, the d e s for group 2 can be derived. The çtability measures corresponding to key generator MVAR are rnr-vs = 0.0295 mr-s = 0.2995 mecs = 0.5 - u s = 0.8657 From d e s forrned under group 3, the stability measures corresponding to generator terminai voltage are m g v s = 0.6 mg-s = 0.8936 mg-CS = 0.8936 mg-us = 0.4 Finally, fiom the niles formed under group 4, the stability measures o f the system are s- = 0.2098 ss-s = 0.4 ss-CS = 0.8 ss-us = 0.8657 The overall system stability is given by gl-sta = max ( ss vs, ss-s, ss-CS, ss-us ) = 0.8657 - Hence, the system staôility is unstable. From the above analysis, it is seen that the solution is obtained by simple max-min d e s of fuey logic and thereby avoiding detailed computatiom of conventionaf methods. For the purpose of control, consider the following three cases. Case A: In this case, it is assumed thai Stafic Var Compensators (SVC) are available at buses 6,8,13 with maximum capacity of 100 MVAR each, while the voltage of generaton at buses 32 and 37 c m be adjusted up to a maximum of 1.05 p.& Case B: Here, assume improved lirnits on controllers in the same buses as in case A. SVC controllers at buses 6, 8, 13 operate with a maximum capacity of 400 MVAR each, while the upper limit of generator voltage at buses 32 and 37 is 1.06 p.u. Case C: In h s case, SVC controllers are available at buses 6, 8, 13, 21 with a maximum capacity of 400 MVAR each, the upper limit of generator voltage at buses 30,32,35,37 is 1 .O6 p u and the upper limit of the taps of tap changing transformer in branches 12- 1 1 and 19-20 is 1.07 p.u. Thus, ten controllers are considered for this case. Table 6.4, Table 6.5 and Table 6.6 show the expert system output after the control action for those cases needing control. Table 6.4 Expert system output - control stage ( case A ) Table 6.5 Expert system output - control stage ( case B ) index No 2 3 ' 5 7 8 9 *11 * misclassification * misclassification Bus 12 Load Volt (P.@ O. 7944 0.8192 0.8673 0.8229 0.9349 0.8307 0.8362 Index No '2 3 5 7 8 9 11 Gen 32 W A R 8490.7 86 1 O. 5 8843.5 8656.3 9176.7 8682.6 8720.4 Bus 12 Load Volt (P-u) 0.9 179 0.9336 0.967 1 0.9325 0.9349 O. 9475 0.9443 Global state CS CS - CS S CS - Gen. 32 M V A R 8988.1 9067.5 9234.5 9087.8 91 76.7 9147.7 9145.5 Gfobal state - VS/S S S S S S VSTAB output Eigen System value state VSTAB output Eigen System d u e state 4-12 4.76 6.03 4.63 7.1 1 4.44 4.48 6.74 7.2 8.19 7.2 7.1 1 6.79 6.83 CS CS CS CS S CS CS S S S S S S S A Table 6.6 Expert system output - control stage ( case C ) It can be seen from Tables 6.4 to 6.6 that the global state of the system has Index No 2 3 5 7 * 8 9 11 changed after the control action. For example, in case C, the operating condition corresponding to index number 9 is unstable (US) before the * misclassification Bus 12 Load Volt (PJ) 0.9620 0.9768 1 .O092 0.9744 1.0656 0.992 1 0.9879 control action and stable after the control action. Note that the output of the expert system is either very stable or stable for the operating condition G e n 32 MVAR 9060.8 9133.7 9292.0 9148.4 9566.3 92 17.9 921 1.7 corresponding to the index 2. This is a typical case where, the expert system is able to corne out with an approximate solution only, considenng Global state VS/S S S S VS S S the fact that no solution may be possible by other conventional methods Note: The tenn "misclassification" refers to the error in the chsification state by the fitay-expert system output. The four classification states are Very Stable, Stable, Critically Stable and Unstable. If there is a discrepancy between the fûzzy-expert system output ( GIobai state ) and the system state of the VSTAB outpu& then a "misclassification" is said to OCCUT. VSTAB output Eigen System value state 7.44 7.86 8.78 7.77 9.47 7.52 7.5 S S S S S S S From the above Tables, it is seen that there is a considerable improvement in the voltage stability rnargïn of the system, limited oniy by the number of controllen available and their operational lirnits. Voltage stabiliîy m a r e ( VSM ) is a measure of how close the system is to voltage instability. It is defineci as the difference between the values of a Key System Parameter ( KSP ) at the current operating condition and the voltage stability critical point [IV. Here, KSP is defined as the total load increase in the selected buses. Fig. 6.4 shows the voltage stability margin improvement afler the control action for the cases A, B and C. TabIe 6.7 shows the voltage stability margin corresponding to Fig. 6.4. Fig. 6.4 Voltage stability margin (VSM) for pre and p s t control cases Table 6.7 Voltage Stability Margin ( VSM ) for pre and post control cases 1 No 1 Sdected Buses 1 VSM 1 I Total Laad at From the above simulation results, it is seen that given the key variables ( load bus voltage, generator MVAR reserve and generator terminal voltage ), the expert system arrives at the global state without the need for complex computations. With reference to the selection of the number of input variables on the size and complexity of fupy-expert system, consider the monitoring stage of the proposed expert system. Here, three input variables are selected. The total nurnber of d e s under the four groups is 304. if only two input variables are selected, the number of d e s is reduced to 96. Thus, the size of the d e s will have an effect on the computational complexity of the fÙzzy-expert system. Pn-Coatrol The designed expert system is tested for a total of 68 cases covering a wide range of operating conditions. 32 cases are listed under Appendix D, 21 cases listed in the control stages - Case A, Case B and Case C, 15 cases Post-Contml VSM ( MW ) i listed under Appendur C, totalling 68 cases. Out of these 68 cases, 4 cases are misclassifieci. The source of error for the misclassified cases is probabiy due to the inadequate tky-expert system framework. Within this framework, two key factors are the membership function of the key variables and inappropnate mernbership value assigned to the linguistic variables descniing the imprecision of the d e . Since the membership function of the key variables described by the equation 6.1 is fonned based on the extensive simulations, the only other parameter responsible for the misclassification is the latter. For example, consider the operating case conesponding to the index No. 5 of Table 6.4. The input variables to the furzy-expert system are bus 12 load voltage = 0.8673 p u and generator 32 MVAR reserve = 8843.5 MVAR. From Fig.6.l. and applying ma-min compositional nile of inference. the membership fiinction of the stability measures VS, S, CS and US for the load bus voltage are ml-vs = 0.6 ml-s = 0.9034 (6.14 ) ml-CS = 0.6882 ml-us = 0.4 Similady, fiom Fig.6.2 and applying max-min compositional mie of inference, the membership function of the stability measures VS, S, CS and US for the genemtor MVAR reserve are mg-vs = 0.1948 mg-s = 0.3642 mg-cs = 0.3642 mg-us = 0.3642 The rules nlating load bus voltage. generator MVAR reserve and stability measure is given by the "mugi" ma& as s h o w below From the above matrix, the stability mesures for the entire system are SS-vs = 0.3642 SS-s = 0.3642 ( 6.17) ss-CS = 0.3642 ss-us = 0.3642 From the above solution, one cannot infer whether the system is very stable, stable, critically stable or unstable. For the above operating condition, the fimy-expert system did not arrive at the proper solution. The fuPy-expert system output is significantly kdïuenced by the "mugl" matrix. The membership values in the "mugl" matrix describes the imprecision of the d e s and is obtained by trial and error method. In achlal field situation, the "mugl" matrix is fonned tlirough operator experience. Hence, the projected 10% classification error may not be taken as a sole criteria for the proposed fuay-expert system to be acceptable in a field situation. The limitations of the proposed funy-expert system in accurately classifjmg the voltage stability condition are The sire of the debase is large due to the selection of thtee input variables for the monitoring stage and two input variables for the control stage. The total number of mies for the above two stages is 400. The viability of the proposed fupy-expert system in an actual field situation is questionable due to the projected 10% classification error and non-availability of irnprovement in the computational speed compared to the conventional voltage stability methods. The proposed funy-expert system assumes a constant power load mode1 and test results limited to sixty four operating conditions. To make a proper voltage stability assessrnent for a practical power system, suitable load models have to be incorporated. Initially, the knowledge-base of fuzzy-expert system starts with 15 operating cases as indicated in Appendix C. Once the data used for the formation of the rulebase is established, the 32 cases listed in Appendix D are tested. The füzzy-expert system updates its knowledge-base to 47 cases and so on. Each tirne a new operating case is tested and verified, it is stored in the knowledge-base and thereby its performance c m be improved. In a utility operation, the tnie potential of the proposed fuzzy- expert system with its vast knowledge-base can be realized Mly only after a considerable period of time and experience. It is possible to investigate the optimum selection of the number of key variables for evaluating the voltage stability of any complex power system under "any operating condition", if sufficient resources and system &ta are available. in order to establish firm conclusions based on the concepts developed in the preceding chapten, extensive simulations on various power utility systems need to be perfonned, and factors that affect the performance of funy-expert system identified In the studies reported in this thesis, only limited number of cases were tested on the sample New England 39 bus system because of non-availability of resources and real utility system data. 6.5 Integration of f u z y s r p e r t system into an Energy Management S y stem The main challenge for the implementation of an on-line voitage stability evaluation in Energy Management System ( EMS ) is the cornputational burden and the ability to arrive at the operating state of the system. Also, it is essentiai for the operator to bave a clear Imowledge of the o p e r a ~ g state when the power system is operating close to its limits. This is where the fuzzy-expert system plays an important role. Since, the proposeci fuay- expert system uses input pammeten already monitored by the EMS, additional data acquisition equi pment and other associated communication systems become unnecessary. Hence, speed c m be improved w osiderab1 y. Man Machine Intcrfücc To design an efficient fuzzy-expert system, key variables that affect the system voltage stability have to be identified first either by off-Iine simulations or through operator's experience. in order to establish database that reflects al1 possible operating conditions of the system, numerous simulations are to be performed and verified by a standard bench mark tool. A set of decision rules relating key system variables to stability are formed. This is a continuous process wherein the funy-expert system updates its knowledge-base and thereby improve its performance. Fig 6.5 shows a block diagram of the proposed scheme integrating fuzy- expert system for voltage stability evaluation as part of an Energy Management System. Power System A r Fig 6.5 Fuzzy-expert system as a part of new EMS R T U R m RW A v e SCADA r Fu ny - Expert S ystcm The Remote Terminal Units (RTUs) collect data fiom various locations in the power system and reiay them to the S u p e ~ s o r y Control and Data Acquisition (SCADA). The SCADA is comected to the Man Machine Interface (MMI), which allows the operator to interact with the EMS. The funy-expert system gets its inputs Erom the SCADA. The main fùnction of the SCADA is to perform various control actions like switching on and off of circuit breaker, transformer taps, capacitor banks etc. Based on the production d e s developed which form the basis for logical reasoning conducted by the inference engine, the expert system arrives at the system state and alerts the system operator to any potentidly dangerous situations. Before taking the control action, the operator performs the load flow solution by incorpo rating appropriate VAR compensation or other available control devices in the secondary analysis to obtain improved load bus voltage and reactive generation reserve. The secondary andysis contains application fùnctions like contingency analysis, load flow, short- circuit analysis, stability analysis and optimal power flow. When the operator is satisfied with the seconàary analysis output, the improved load bus voltage and reactive generation reserve serve as input to the SCADA after verifjmg the system state fiom the fuay-expert system output. The SCADA takes the necessary control action to alleviate the voltage stability problem. The proposeci fiizzy-expert system does not replace any of the well developed algonthmic solutions of the secondary aoalysis. However, it offen a powemil and effective tool for the use of these programs. 6.6 Summary This chapter addresses the issues conceming the selection of the number of input variables and its impact on the size and cornplexity of the furzy-expert system. It also includes factors to be taken into account for proper selection of parameters "a" and "A7' of the generk equation describing the membership b c t i o n s of the key variables. Issues like accuracy of the method and erroa are also addresseci It also substantiates the claims for the use of fuzzy logic approach by showing a reliable assessrnent of stability for the monitoring stage without perforrning the detailed voltage stability calculations for a given operating point. In the proposed funy-expert system, the key variables like load bus voltage, generator MVAR reserve and generator terminal voltage w hich are used to rnonitor the voltage stability are stored in the database. Changes in the system operating conditions are reflected in the database. The rulebase comprises a set of production rules which form the basis for logical reasoning conducted by the inference engine. The reasoning process of the inference engine is implemented in MA=. Given the key variables, the expert system arrives at the global state without the need for complex computations. The New England 39 bus system is taken as a case study to illustrate the proposed procedure. Extensive operating conditions have been tested to validate the proposed system. The results that are k i n g compared are the fuzzy-expert system output ( global state ) and the system state of the VSTAB output for the 68 operating cases. The misclassification is f d to be less than 10%. Hence, it has the potential to be integrated for on-line implementation in Energy Management System. In this new system. the rnembership fiinctions of the key variables and the debase may be defined based on system requirements and operator's experience. Thus, it offers flexibility and satisfactory results in a very efficient way. Chapter 7 CONCLUSIONS 7.1 Contribution of the Research The concept of voltage stability phenornena in power systems has been thoroughly reviewed As power systems continue to be loaded closer to their stability limit, there is a need for suitable voltage stability indices. Three simple stability indices were investigated with the help of a sample 5 bus power system. They are singular value decomposition, "L" Index and QV curves. The simulation resdts showed that singular value decomposition and "L" Index indicate proximity or neamess to the voltage collapse point. But the cooslraint in these indices is that the load flow solution does not converge at the bifurcation point. Regarding QV cuves, the rnethod artificially stresses a single bus, hence conclusions should be confirmed by more realistic methods. Also, the curves are obtained by a series of power 80w simulations that make it more time consuming. Modal analysis technique in conjunction with continuation power flow, which is a better performance criteria for the assessrnent of voltage stability was investigated. Simulations were carried out for the IEEE 30 bus system and the New England 39 bus system. The above two examples indicate how modes represent areas prone to voltage instability. Thus, modal analysis clearly identifies groups of buses and critical bus that participate in the instability and thereby eliminate the problems associated with traditional rnethods. Though time consuming, this method is used as a benchmark for developing an expert system for voltage stability evaluation. Expert sysiems, a subset of artificiai intelligent system, have attracted wide spread interest to power system applications. This is due to their ability to bandle stressed power system and improve speed. in this regard, the concept of fuzzy-expert systems has k e n described in detail. A modified IEEE 30 bus system applied to voltage control was simulated to emphasize the concepts developed in the fuzzy-expert systems. To M e r explore its suitability to voltage stability monitoring and control, the New England 39 bus system was taken as case study. Extensive operating conditions were tested to validate the proposed scheme. The percentage error was found to be less than 10%. Based on the simulation results, this new approach was found to be simple and straight fonivard, where it ody needs key system variables to mive at the solution state. In general, the method is able to handle non-linearity of power systern problem and does not require complex computations as in traditional methods. Thus, it is more efficient than conventional methods for voltage stability analysis. The proposed hiny-expert system was tested for 68 different operating conditions. Four cases were misclassified totally and three cases partially misclassified. The source of error for these misclassified cases was probably due to the inappropriate values assigned to the linguistic variables of the key variables in the nilebase of group 4. To give a guarantee or bound on the error with this method, d l possible operating cases are to be tested and verified with standard bench mark tool. With the three input variables, the total number of d e s for the monitoring stage was determined to be 304. The size of the rulebase can be reduced to 96 if two input variables are selected judicially. Thus, the sue of the rulebase will have an effect on the computational compiexity of the futzy-expert system. There is considerable interest among utilities in developing on-line voltage stability twls that will enable the power systerns to be operated at higher loads without risking voltage collapse. The fbq-expert system proposed in this thesis has the potential to be integrated for on-line implementation in Energy Management System to achieve the goals of secure power system operation. This will allow power system operators to continuously monitor the system state and thereby obviate any impending dangers of voltage collapse. However, to realize this challenge, an efficient database that reflects al1 possible system operating conditions should be formed. 7.2 Recommendations for Future Work The voltage stability analysis considered in this thesis assume a constant power load model, which is not the case with a practical power system. Suitable load models can be incorporateci in the stability assessment. The present study is lirnited to around si* four operating conditions. To assess the viability of fuzzy-expert systems to a larger realistic power systems, al1 possible operating conditions should be considered. In order to enhance the efficiency of the present fiipv-expert system, better funy models and rulebase that descnbe both monitoring and control under one urnbrella is recommended. REFERENCES IEEE Speciai Publication 90TH0358-2-PWR, Voltage Stability of Power Systeis: Concepts, Analfical Tools and Industry Experience, 1990. B. Jeysurya, Artificial neural networks for power system steady-state voltage instability evaluation, Electric Power Systems Research., 29 ( 1 994) 85-90. A.R. Bergen, Power System Analyss, Prentice-Hall, Englewood Cli ffs, NJ, 1986. C. W. Taylor, Power System Voltage Stabil-y, McGraw-Hill, 1 994. P. Kundur, P ower System Strrbility and Conîrof, McGraw-Hill, 1994. IEEE Special Publication 93THa620-5-PWFt, Suggested Techniques for Voltage Stability Anaiysis, 1993. A. Tiranuchit and R.J. Thomas, "A Poshinng Stnitegy Against Voltage hstabilities in Electric Power Systems", IEEE Trans. Power Systems, Vol. 3, No. 1, Febnmy 1988, pp. 87-93. P. Kessel and H. Glavitsch, "Estimating the Voltage Stability of a Power System", IEEE Trans. Power Delivery, Vol. PWRD-1, No. 3, July 1986, pp. 346-354. Voltage Stability Ana[ysis Program - VSTAB Version 4.1, Powertech Labs Inc., B.C, Canada, November 1994. G. W. Stagg and A.H. El-Abiad, C o m p t e r Methodî in Power Systems Anolysis, McGraw-Hill International, 1 968. C.A. Canizares, A.C.Z. de S o m and V.H. Quintana, "Cornparison o f Performance Indices for Detection of Proximity to Voltage Collapse", IEEE Trans. Power Systems, Vol. 1 1, No. 3, August 1996, pp. 144 1- 1450. [12] T.J. ûverbye and C.L. DeMarco, "Voltage Security Enhancement using Energy Based Sensitivities", EEE T m . Power Systems, Vol. 6, No. 3, August 199 1, pp. 1 196- 1202. [13] M. Huaeault, C. Rosu, R. Manoliu and F.D. Galiana, "A Study of Knowledge Engineering Tools in Power Engineering Application", IEEE Trans. Power Systems, Vol. 9, No. 4, November 1994, pp. 1825-1 832. [14] 2.2. Zhang, G.S. Hope and O.P. Malik, "Expert Systems in Elecaic Power Systems- A Bibliographicd Survey", EEE Tram. Power Systems, Vol. 4, No. 4, October 1989, pp. 1355-136 1. [15] Extenried Tronrienfi'Midtenn Srabifity P rogram - ETMSP Version 3.1, Ontario Hydro, Ontario, Canada, May 1994. [16] B. Gao, G.K Morison and P. Kundur, "Voltage Stability Evaluation using Modal Analysis", IEEE Trans. Power Systems, Vol. 7, No. 4, November 1992, pp. 1529- 1542. [17] B. Gao, G.K. Monson and P. Kundur, "Towards the Development of a Systematic Approach for Voltage Stability Assessrnent of Large-ScaIe Power Systems", IEEE T m . Power Systems, Vol. 11, No. 3, August 1996, pp. 13141324. [18] Interactive Power Flow- lPFLOW Version 4.1, Powertech Labs Inc., B.C, Canada, May 1994. [19] V. Ajjarapu and C. Christy, 'The Continuation Power Flow: A Tool for Steady-State Voltage Stability Analysis", IEEE Trans. Power Systems, Vol. 7, No. 1, February 1992, pp. 4 16-423. [20] H.B. Wm, Y.H. Song and A.T. Johns, "Effective Control for Power System Voltage Stability Enhancement", 28th North Amencan Power Symposium, M.LT, November 1996, pp. 257-263. [2 11 J. Gianatano and G. Riley, Expert System: Principles and Programming, PWS Publishing Company, Boston, 1994. [22] L.A. Zadeh, "A Theory of Approximate Reasoning", Machine Intelligence, 9, e d S.E. Hayes, D. Michie and L.I. Mikulich, 1979, pp. 149-194. [23] ABB leads the way with solutions for total systems integration, ABB Systems Control, Marketing Department, Santa Clam* CA 9505 1. [24] Proceedings of Third Symposium on Expert Systems Applications to Power Systems, Japan, A M 1 199 1. [25] K. Tomsovic and J.M. Ling, " A hoposed Fuzzy uifomation Approach to Power System Security", Proceedings of Third Symposium on Expert Systems Applications to Power Systems, Japan, Apil 199 1, pp. 427-432. [26] T. Maekawa, K. Yasuda, R. Yokoyama, H. Ohtsuki and Y. Minikami, "Fuzzy Coordination of Multi-Objectives Reflecting Operator's Intention for Dynamic Generation Rescheduling", Proceedings of Third Symposium on Expert Systems Applications to Power Systems, Japan, April 199 1, pp. 470-477. [27l K. Tomsovic, "A Furry Linear Programrning Approach to the Reactive PowerNoltage Control Problem", IEEE Trans. Power Systems, Vol. 7, No. 1, February 1992, pp. 287-293. [28] H. Mori and H. Kobayashi, 'Optimal F u n y ùiference for Short-terni Load Forecasting", IEEE Trans. Power Systems, Vol. 1 1, No. 1, Febniary 1996, pp. 390-396. [29] Rahul mare and R.D. Christie, "Prioritizing Emergency Control Roblems with Fuzzy Set Theory", IEEEPES Summer Meeting, Denver, August 1996. [30] V.C. Ramesh and X u n Li, "A F u a y Multiobjective Approach to Contingency Constrained OPF, (EEE/PES Sumrner Meeting, Denver, August 1996. J.A. Momoh, X W . Ma and K. Tomsovic, " O v e ~ e w and Literature S w e y of Fuzzy Set Theory in Power Systems*, IEEE Trans. Power Systems, Vol. 10, No. 3, August 1995, pp. 16764690. T. Niimura, Y. Nakanishi, K Yasuda and R Yokoyarna, "Multi-Attribute Voltage-Reactive Power Control Based on Approximate Reasoning", Proceedings of Third Symposium on Expert Systems Applications to Power Systems, Japn, April 199 1, pp. 462469. KK Abdul-Rahman and S.M. Shahidehpour, ''A Fuzzy-Based Optimal Reactive Power Contro 1", IEEE Trans. Power Systems, Vol. 8, No. 2, May 1993, pp. 662-670. A.G. Exposito, J.L.M. Ramos, J.L.R. Macias and Y.C. Salinas, "Sensitivity - Based Reactive Power Contro 1 for Voltage Profile hprovement", IEEE Trans. Power Systems, Vol. 8, No. 3, August 1993, pp. 937-945. W.R Wagner, A. Keyhani, S. Hao and T.C. Wong, "A Rule Based Approach to Decentralized Voltage Control", IEEE Trans. Power Systems, Vol. 5, No. 2, May 1990, pp. 643-651. Ching-Tzong Su and Chien-Tung Lin, "A New Fuzzy Control Approach to Voltage Profile Enhancement for Power Systems", EEE Trans. Power Systems, Vol. 1 1, No. 3, August 1996, pp. 1654-1 659. Chen-Ching Liu and Kevin Tomsovic, "An Expert System Assisting Decision-Making of Reactive PowerNoltage Control", IEEE Trans. Power Systerns, Vol. 1, No. 3, August 1986, pp. 195-20 1. R. Mohamedi, C. Belhadji, S. Lefebvre and Xuan-Dai Do, "An Expert System for S teady-State Voltage Stability ", Canadian Conference for Electrical and Cornputer Engineering, Hafi fax, September 1994, pp. 696- [39] Yuan-Yih Hsu and ChungChing Su, "A RulaBaseci Expert System for Steady-State Stability Analysisw, IEEE T m . Power Systems, Vol. 6 , No. 2, May 199 1, pp. 77 1-777. 1401 M T . Version 4.1, The MathWorks Inc., Natick, Massachusetts 01760, August 1992. APPENDIX - A Line and Bus data for the S bus system Table. A l Line data for the 5 bus system Table. A2 Bus &ta for the 5 bus systern Bus I Code 1 2 Assumed Bus V o h p ( p.u ) 1 .O6 1 .O Load MW MVAR O 1 O 20 1 1 0 Gentration MW MVAR O 40 O 30 APPENDIX - B Rules relating key variables to stabüity meesure of the New England 39 bus system Table. B 1 Rdes relacing worst load bus voltage to stability measure under group 1 0 Moderate 1 0.6 1 1.0 1 0.4 1 0.0 Table. B2 Rdes relating key generator MVAR reserve to stability Tolerable rneasure under group 2 Safc 1 .O O. 8 0.3 0.0 0.0 1 1 1 1 Moderate 1 0.6 1 1.0 1 0.4 1 0.0 Tolerable 0.0 O. 7 1 .O 0.4 0.7 I Low 0.0 0.2 0.5 1 .O Table. B3 Rules relating key generator terminal voltage to stability 1 .O measure under group 3 0.4 1 Safe 1 .O 0.8 0.3 0.0 Tolerable I l l Low 0.0 0.0 0.7 1 1.0 1 0.4 0.2 o. 5 1 .O Table. 84 Rules relating combined key variables to stability measure under group 4 Data used for the rulebase formation for the monitoring stage of the New Eogland 39 bus system Table. C 1 Data used for the miebase formation Case No 1 Totai Load at Selected Buses ( MW ) Gen 32 Volt ( p.u ) Bus 12 Load Voit ( P - U Gen 32 MVAR Complete list of expert system output for the monitoring stage of the New Engiand 39 bus system Table. Dl Expert system output for various neighborhood points - Monitoring stage h d a No Gen 32 Volt (PJ) Total Load at Selected B u s ~ s ( M W ) NO CONTINGENCY Bus 12 Load Voit (PW 1 2 3 4 5 6 7 8 9 1 O 11 12 13 14 15 16 17 Gen 32 W A R 1424.5 1824.5 2624. 5 4624. 5 4958.8 4968.2 497 1.4 3824.5 4774.5 48 12.0 1424.5 1824.5 4424.5 4824.5 5 124.5 5174.5 5183.8 0.983 1 0.983 1 0.983 1 0.9831 0.983 1 0.9831 0.9831 0.95 0.95 0.95 1.05 1 .O5 1 .O5 1 .O5 1-05 1 .O5 1 .O5 18 1424.5 S GIobal State 9823.0 1.0072 0.9879 0.9428 0.7399 0.6250 0.6141 0.6093 0.8276 0.6580 0.6399 1.0374 1.0190 0.8263 0.7644 0.6801 0.6482 0.6378 CONTINGENCY VSTAB output Eigen System vaiue State A 9.89 VS S 10.1 1 9746.5 9556.6 8628.3 81 19.3 8073.1 8052.4 9199.7 8439.0 8361.4 9630.3 9554.9 8663.0 8367.8 7976.0 7832.3 7786.4 VS 9.63 8.50 3.40 0.76 0.52 0.39 5.77 1.71 1.30 10.55 10.09 5.10 3.55 1.53 0.79 0.56 S S CS US U S US CS US US VS VS CS CS US US US S S CS US US US CS US US VS VS CS CS US US US Participation factors for the criticai case 1 Case 1 ( at the voltage stability Iimit ) of the New England 39 bus system Table. E 1 Bus and Generator participation factors for criticai case Bus Participation Factors Bus N o Participation Factors Genefator Participation Factors Gen No Participation Factors 12 4 14 Case 5 Case 2 O. 1713 O. 1042 0.0949 3 2 3 1 12 4 14 Case 6 1.0000 0.73 17 Case 3 0.1788 0.1015 0.094 1 32 31 12 4 14 13 1 .O000 0.7390 Case 4 O. 1562 O. 1093 0.0965 0.0795 32 3 1 1 -0000 0.7150 l MAGE NALUATION TEST TARGET (QA-3) APPLIED - lN1AGE. lnc = 1653 East Main Street - -. , , Rochester, NY t 4604 USA ,=-* Phone: i l 6 / ~ 8 ~ ~ -- -- - - Fax: 71 a12û8-5989 
work_pz4j3x2gwfcodfqaa2m6wga6h4	----	Auto claim fraud detection using Bayesian learning neural networks Auto claim fraud detection using Bayesian learning neural networks S. Viaene a,b,*, G. Dedene b,c , R.A. Derrig d a Applied Economic Sciences, K. V. Leuvei, Naamsestraat 69, B-3000 Leuven, Belgium b Vlerick Leuven Gent Management School, Reep1, B-9000 Gent, Belgium c Economics and Econometrics, University of Amsterdam, Roetersstract 11, 1018 WB Amsterdam, The Netherlands d Automobile Insurers Bureau of Massachusetts & Insurance Fraud Bureau of Massachusetts, 101 Arch Street, Boston MA 02110, USA Abstract This article explores the explicative capabilities of neural network classifiers with automatic relevance determination weight regularization, and reports the findings from applying these networks for personal injury protection automobile insurance claim fraud detection. The automatic relevance determination objective function scheme provides us with a way to determine which inputs are most informative to the trained neural network model. An implementation of MacKay’s, (1992a,b) evidence framework approach to Bayesian learning is proposed as a practical way of training such networks. The empirical evaluation is based on a data set of closed claims from accidents that occurred in Massachusetts, USA during 1993. q 2005 Elsevier Ltd. All rights reserved. JEL classification: C45 Keywords: Automobile insurance; Claim fraud; Neural network; Bayesian learning; Evidence framework SIBC: IB40 1. Introduction In recent years, the detection of fraudulent claims has blossomed into a high-priority and technology-laden problem for insurers (Viaene, 2002). Several sources speak of the increasing prevalence of insurance fraud and the sizeable proportions it has taken on (see, for example, Canadian Coalition Against Insurance Fraud, 2002; Coalition Against Insurance Fraud, 2002; Comité Européen des Assurances, 1996; 1997). September 2002, a special issue of the Journal of Risk and Insurance (Derrig, 2002) was devoted to insurance fraud topics. It scopes a significant part of previous and current technical research directions regarding insurance (claim) fraud prevention, detection and diagnosis. More systematic electronic collection and organization of and company-wide access to coherent insurance data have stimulated data-driven initiatives aimed at analyzing and modeling the formal relations between fraud indicator combinations and claim suspiciousness to upgrade fraud detection with (semi-)automatic, intelligible, accountable tools. Machine learning and artificial intelligence solutions are increasingly explored for the purpose of fraud prediction and diagnosis in the insurance domain. Still, all in all, little work has been published on the latter. Most of the state-of- the-art practice and methodology on fraud detection remains well-protected behind the thick walls of insurance compa- nies. The reasons are legion. Viaene, et al. (2002) reported on the results of a predictive performance benchmarking study. The study involved the task of learning to predict expert suspicion of personal injury protection (PIP) (no-fault) automobile insurance claim fraud. The data that was used consisted of closed real-life PIP claims from accidents that occurred in Massachusetts, USA during 1993, and that were previously investigated for suspicion of fraud by domain experts. The study contrasted several instantiations of a spectrum of state-of-the-art supervised classification techniques, that is, techniques aimed at algorithmically learning to allocate data objects, that is, input or feature vectors, to a priori defined Expert Systems with Applications 29 (2005) 653–666 www.elsevier.com/locate/eswa 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.04.030 * Corresponding author. Tel.: C32 16 32 68 91; fax: C32 16 32 67 32. E-mail address: stijn.viaene@econ.kuleuven.ac.be (S. Viaene). http://www.elsevier.com/locate/eswa object classes, based on a training set of data objects with known class or target labels. Among the considered techniques were neural network classifiers trained according to MacKay’s (1992a) evidence framework approach to Bayesian learning. These neural networks were shown to consistently score among the best for all evaluated scenarios. Statistical modeling techniques such as logistic regression, linear and quadratic discriminant analysis are widely used for modeling and prediction purposes. How- ever, their predetermined functional form and restrictive (often unfounded) model assumptions limit their usefulness. The role of neural networks is to provide general and efficiently scalable parameterized nonlinear mappings between a set of input variables and a set of output variables (Bishop, 1995). Neural networks have shown to be very promising alternatives for modeling complex nonlinear relationships (see, for example, Desai et al. 1996; Lacher et al. 1995; Lee et al. 1996; Mobley et al. 2000; Piramuthu, 1999; Salchenberger et al. 1997; Sharda & Wilson, 1996). This is especially true in situations where one is confronted with a lack of domain knowledge which prevents any valid argumentation to be made concerning an appropriate model selection bias on the basis of prior domain knowledge. Even though the modeling flexibility of neural networks makes them a very attractive and interesting alternative for pattern learning purposes, unfortunately, many practical problems still remain when implementing neural networks, such as What is the impact of the initial weight choice? How to set the weight decay parameter? How to avoid the neural network from fitting the noise in the training data? These and other issues are often dealt with in ad hoc ways. Nevertheless, they are crucial to the success of any neural network implementation. Another major objection to the use of neural networks for practical purposes remains their widely proclaimed lack of explanatory power. Neural networks are black boxes, it says. In this article Bayesian learning (Bishop, 1995; Neal, 1996) is suggested as a way to deal with these issues during neural network training in a principled, rather than an ad hoc fashion. We set out to explore and demonstrate the explicative capabilities of neural network classifiers trained using an implementation of MacKay’s (1992a) evidence framework approach to Bayesian learning for optimizing an automatic relevance determination (ARD) regularized objective func- tion (MacKay, 1992; 1994; Neal, 1998). The ARD objective function scheme allows us to determine the relative importance of inputs to the trained model. The empirical evaluation in this article is based on the modeling work performed in the context of the baseline benchmarking study of Viaene et al. (2002). The importance of input relevance assessment needs no underlining. It is not uncommon for domain experts to ask which inputs are relatively more important. Specifically, Which inputs contribute most to the detection of insurance claim fraud? This is a very reasonable question. As such, methods for input selection are not only capable of improving the human understanding of the problem domain, in casu the diagnosis of insurance claim fraud, but also allow for more efficient and lower-cost solutions. In addition, penalization or elimination of (partially) redundant or irrelevant inputs may also effectively counter the curse of dimensionality (Bellman, 1961). In practice, adding inputs (even relevant ones) beyond a certain point can actually lead to a reduction in the performance of a predictive model. This is because, faced with limited data availability, as we are in practice, increasing the dimensionality of the input space will eventually lead to a situation where this space is so sparsely populated that it very poorly represents the true model in the data. This phenomenon has been termed the curse of dimensionality. The ultimate objective of input selection is, therefore, to select a minimum number of inputs required to capture the structure in the data. This article is organized as follows. Section 2 revisits some basic theory on multilayer neural networks for classification. Section 3 elaborates on input relevance determination. The evidence framework approach to Bayesian learning for neural network classifiers is discussed in Section 4. The theoretical exposition in the first three sections is followed by an empirical evaluation. Section 5 describes the characteristics of the 1993 Massachusetts, USA PIP closed claims data that were used. Section 6 describes the setup of the empirical evaluation and reports its results. Section 7 concludes this article. 2. Neural networks for classification Fig. 1 shows a simple three-layer neural network. It is made up of an input layer, a hidden layer and an output layer, each consisting of a number of processing units. The layers are interconnected by modifiable weights, represented by the links between the layers. A bias unit is connected to each unit other than the input units. The function of a processing unit is to accept signals along its incoming connections and (nonlinearly) transform a weighted sum of these signals, termed its activation, into a single output signal. In analogy with neurobiology, the units are sometimes called neurons. The discussion will be restricted to the use of neural networks for binary classification, where the input units represent individual components of an input vector, and a single output unit is responsible for emitting the values of the discriminant function used for classification. One then commonly opts for a multilayer neural network with one hidden layer. In principle, such a three-layer neural network can implement any continuous function from input to output, given a sufficient number of hidden units, proper nonlinea- rities and weights (Bishop, 1995). We start with a description of the feedforward operation of such a neural S. Viaene et al. / Expert Systems with Applications 29 (2005) 653–666654 https://isiarticles.com/article/17681 
work_q6hwg6uubffdvieyd3xnivrg3e	----	Modeling of an Inverted Pendulum based on Fuzzy Clustering Techniques International Journal of Computer Applications (0975 – 8887) Volume 9– No.4, November 2010 23 Modeling of an Inverted Pendulum based on Fuzzy Clustering Techniques E.Sivaraman Assistant Professor, Annamalai University Annamalainagar, Tamilnadu India - 608002 S.Arulselvi Associate Professor, Annamalai University Annamalainagar, Tamilnadu India - 608002 S.P.Natarajan Professor, Annamalai University Annamalainagar, Tamilnadu India - 608002 ABSTRACT The inverted pendulum is a highly nonlinear and open loop unstable system. To develop an accurate model of the inverted pendulum, different linear and nonlinear methods of identification will be used. However one of the problems encountered during modeling is the collection of experimental data from the inverted pendulum system. Since the output data from the unstable system does not show enough information or dynamics of the system. This can be overcome by designing a feedback controller , which stabilize the system before identification can takes place. Recently Takagi-Sugeno (T-S) fuzzy modeling based on clustering techniques have shown great progress in identification of nonlinear systems. Hence in this paper, Takagi-Sugeno (T-S) model is proposed for an inverted pendulum based on fuzzy c- means , Gustafson-Kessel (G-K) and Gath-Geva (G-G) clustering techniques. Simulation results show that Gustafson-Kessel (G-K) clustering technique produces satisfactory performance. General Terms Fuzzy logic, Modeling. Keywords Nonlinear, Clustering, Fuzzy, Inverted Pendulum, Takagi-Sugeno. 1. INTRODUCTION Fuzzy control system, identification and design have been a subject of intense research in recent years. The heart of a fuzzy control system consists of a set of fuzzy “if – then” rules, which models the process state-control action relationship in the control domain. The design of appropriate fuzzy rules and membership is a critical bottleneck in developing powerful fuzzy control systems. Membership functions and fuzzy rules can be defined by the model developer (expert), using prior knowledge, or by experimentation, which is a typical approach in knowledge based fuzzy control [1]. Knowledge acquisition however is, a cumbersome task, and for (partially) unknown systems, human experts are not available. Therefore, data-driven construction of fuzzy membership function and rules from measured input - output data has received a lot of attention. Such modeling approaches typically seek to optimize some numerical objective function, while less attention is paid to the complexity of the resulting model in terms of the number of membership functions and rules, Babuska [2]. Takagi- Sugeno [3] proposed a heuristic method for designing Takagi - Sugeno type of fuzzy rules from input-output data, using some performance index. Subsequently, Sugeno and Tanaka reported [4] a method for on-line identification of a fuzzy system. Recently, several proposals are made to use fuzzy clustering for fuzzy system identification. An effective approach is to partition the available data into subsets and approximate each subset by a simple piecewise linear model. Fuzzy clustering can be used as tool to partition the data where transitions between the subsets are gradual [5]. Sugeno and Yasukawa [6] uses the fuzzy c-means algorithm to cluster the output data points to determine the fuzzy rules, and use some heuristics to select the most relevant input variables in fuzzy system design. Bezdek [7] combines the ideas in [6] and [8] to develop a neural network fuzzy control system which also utilizes clustering. Typically, the specification of a real-world control problem is given by a finite set of input-output data points zk = [ xk, yk], and the design task is to extract optimum membership functions and good set of fuzzy rules that best represent the input-output relationship in the data. This paper address the design issue by demonstrating good performance of the fuzzy clustering-based methods, which identifies clusters from the input-output data and constructs membership function thereby generating a fuzzy rule for each cluster. In this paper, the inverted pendulum problem has been considered as a typical representative of inherently unstable nonlinear control systems, thus it is an ideal choice for testing the modeling capability of the fuzzy c-means, Gustafson-Kessel (G-K) and Gath–Geva(G-G) clustering algorithms. Simulation results are presented using fuzzy modeling and identification toolbox [2]. The performance of the models obtained by fuzzy c- means ,G-K and G-G methods are compared in terms of RMSE. 2. TAKAGI-SUGENO FUZZY MODEL Consider the identification of an unknown nonlinear system )( xfy = (1) based on some available input-output data T k,nk,1 k ]x,.....x[x = and ky . The index k=1,…,N denotes the individual data samples. While it may be difficult to find a model to describe the unknown system globally, it is often possible to construct local linear models around selected operating points. The modeling framework that is based on combining local models valid in predefined operating region is called operating regime-based modeling. In this frame work, the model is generally given by ∑ += = c 1i i T ii )bxa)(x(y φ ) (2) International Journal of Computer Applications (0975 – 8887) Volume 9– No.4, November 2010 24 where )(i xφ is the validity function for the i th operating regime and T i T ii ]ba[=θ is the parameter vector of the corresponding local linear model. The operating regimes can also be represented by fuzzy sets, in which case the TS fuzzy model is obtained [6] as ]w[,bxay)x(AxR ii T iii += ) thenisIf: i=1,….c. (3) here, the number of rules is denoted by c, )x(Ai is a multivariable membership function, ai and bi are parameters of the local linear model as shown in Fig.1, and ]1,0[∈iw is the weight of the rule. Fig. 1. Local linear model and the membership function diagram. The value of iw is usually chosen by the designer of the fuzzy system to represent the belief in the accuracy of the i th rule. When such knowledge is not available iw =1, iA is used. The degree of fulfillment of the rule can be calculated as the product of the individual membership degrees and the rule’s weight ∏== = n 1j jj,iiiii )x(Aw)x(Aw)x(β (4) The rules are aggregated by using the fuzzy-mean formula ∑ ∑ + = = = c 1i i c 1i i T ii )( )ba)(( y x xx β β ) (5) From (2) and (5), one can see that the fuzzy model is equivalent to the operating regime-based model when the validity function is chosen to be the normalized rule degree of fulfillment as ∑ = = c 1i i i i )( )( )( x x x β β φ (6) Prior to clustering, the regression structure of the model is selected, in order to properly represent the system dynamics. Problems where little prior knowledge is available are usually represented in an input-output form. After the structure is determined, clustering in the product space of the regressors and of the regressand can be applied to partition the data. Since each cluster serves as a local linear model of the system, the fuzzy c- means , Gustafson-Kessel (G-K) and Gath-Geva clustering algorithms are capable of detecting clusters which lie in linear subspaces. 3. FUZZY MODEL IDENTIFICATION BASED ON FUZZY C-MEANS, G-K AND G-G CLUSTERING ALGORITHMS Forward and inverse modeling techniques helps to design model based control techniques like direct inverse, Internal Model Control and Model Predictive Control for nonlinear processes. These control techniques give satisfactory response when conventional control techniques fail. Hence, this necessitates the development of different modeling techniques based on fuzzy clustering. 3.1 Fuzzy Partition From the available input–output data pairs, the regression matrix X and the output vector y are constructed as given by ].N1 T N1 T y,...y[y]x,...x[X == and (7) where N>>n is the number of samples used for identification. The antecedent fuzzy sets Ai in (3) are determined by means of fuzzy clustering in the product space of the systems input and outputs. Hence, the data set N)1n( RZ ×+ = to be clustered is represented as a (n+1) × N data matrix composed from X and y ]y,X[Z T = (8) where each column , k z N,...,2,1k = of Z contains an input- output data pair as given by T k T kk ]y,X[Z = Given Z and an estimated number of clusters ‘c’, fuzzy clustering partitions Z into ‘c’ fuzzy clusters. A fuzzy partition can be represented as an ‘N × c’ matrix U, whose elements )x(iφ ∈ [0, 1] represents the membership degree of k Z in clusters ‘i’. Hence, the i th column of U contains the values of the membership function in the fuzzy partition, which is taken to be a point wise representation of the antecedent fuzzy set Ai of the i th rule as given in (3). The sum of each column of U is constrained to one, but the distribution of membership among the ‘c’ fuzzy subsets is not considered. Also, there can be no empty clusters and no cluster International Journal of Computer Applications (0975 – 8887) Volume 9– No.4, November 2010 25 may contain all the objects. This means that the membership degrees in the partition matrix U are normalized, and for the given identification data, the membership values )x(iφ correspond to the normalized degree of fulfillment of the rule antecedents as given in (6). In this paper, membership functions extracted using fuzzy c-means , G-K clustering and Gath - Geva algorithms are discussed in detail. Each cluster forms one fuzzy rule. 3.2 Cluster Validation Cluster validity refers to the problem whether a given fuzzy partition fits to the data all. The clustering algorithm always tries to find the best fit for a fixed number of clusters and the parameterized cluster shapes. However this does not mean that even the best fit is meaningful at all. Either the number of clusters might be wrong or the cluster shapes might not correspond to the groups in the data, if the data can be grouped in a meaningful way at all. Clustering data for different values of c, and using validity measures to assess the goodness of the obtained partitions. Different types of validity measure to find optimum number of clusters are given below 3.2.1 Partition Coefficient (PC) It measures the amount of “overlapping” between cluster. It is defined by Bezdek as follows ∑ ∑= = = c 1 N 1j 2 ij i )( N 1 )c(PC µ where µij is the membership of data point j in cluster i. The disadvantage of PC is lack of direct connection to some property of the data themselves. The optimal number of cluster is at the maximum value. 3.2.2 Classification Entropy (C) It measures the fuzziness of the cluster partition only, which is similar to the partition coefficient. )log( N 1 )c(CE ij c 1i N 1j ij µµ∑ ∑−= = = 3.2.3 Partition Index(SC) It is the ratio of the sum of compactness and separation of the clusters. It is a sum of individual cluster validity measures normalized through division by the cardinality of each cluster ∑ − −∑ ∑= = = = c 1k 2 i 2m ij N 1jc 1i ivk vN ivjx)( )c(SC µ SC is useful when comparing different partitions having equal number of clusters. A lower value of SC indicates a better partition. 3.2.4 Separation Index(S) On the contrary of partition index (SC), the separation index uses a minimum-distance separation for partition validity 2 ik k,i min 2 ij 2 ij N 1j c 1i vvN vx)( )c(S − −∑∑ = == µ 3.2.5 Xie andBeni’s Index (XB) It aims to quantify the ratio of the total variation within clusters and the separation of clusters 2 ijj,imin 2 ij m ij N 1j c 1i vxN vx)( )c(XB − −∑∑ = == µ The optimal cluster should mimimize the value of the index. 3.2.6 Dunn’s Index (DI) This index is originally proposed to use at the identification of “compact and well separated clusters”. So the result of the clustering has to be recalculated as it is a hard partition algorithm. { { } }         ∈∈ ∈∈ ≠∈∈ = )y,x(dmaxmax )y,x(dmin min)c(DI cy,xck jcy,icx ji,cj minci The main drawback of Dunn’s index is computational since calculating becomes computationally very expensive as c and N increase. 3.2.7 Alternative Dunn’s Index (DI) The aim of modifying the original Dunn’s index was that the calculation becomes more simple, when the dissimilarity function between two clusters ))y,x(dcy,cx(min ji ∈∈ is rated in value from beneath by the triangle-non equality: )v,x(d)v,y(d)y,x(d jj −≥ Where j v is the cluster centre of the j-th cluster. { } }           ∈∈ ∈∈ ∈ − ≠∈ = )y,x(dmaxmax )v,x(d)v,y(dmin ji,cj minmin)c(ADI cy,xck jijjcjx,icix ci International Journal of Computer Applications (0975 – 8887) Volume 9– No.4, November 2010 26 Note that the only difference of SC, S and XB is the approach of the separation of clusters. In the case of overlapped clusters the values of DI and ADI are not really reliable because of re- partitioning the results with the hard partition method. The above performance indexes are used to get optimum number of clusters for the following algorithms. 3.3 Fuzzy C-means Clustering Algorithm Most analytical fuzzy clustering algorithms are based on optimization of the basic c-means objective function, or some modification of it. The advantage of fuzzy c-means algorithm is simplicity and easy implementation. Given the data set Z, choose the number of clusters by cluster validation as 1<c<N, the weighting exponent m > 1. The termination tolerance 0>ε and the norm-inducing matrix A. Initialize the partition matrix randomly such that fc M∈ (0) U . The following steps are repeated using MATLAB software for i =1,2,…l, until the termination tolerance is achieved, then the values of membership functions [9] are obtained as given in (11). Step (1): Compute the cluster prototypes (means) ci1, )( Z)( V N 1k m)1l( ik N 1k k m)1l( ik )l( i ≤≤ ∑ ∑ = = − = − µ µ (9) Step (2): Compute the Euclidian distances ,)VZ(A)VZ(D )l( ik T)l( ik 2 ikA −−= (10) Nk1,ci1 ≤≤≤≤ Step (3): Update the partition matrix If 0DikA > for ,Nk1,ci1 ≤≤≤≤ ∑ = = −c 1i )lm/(2 jkAikA )l( ik )D/D( 1 µ otherwise ]1,0[,0Dif0 )l( ikikA )l( ik ∈>= µµ and (11) 1 c 1i )l( ik =∑ = µwith until 1)(l(l) UU −− < ε 3.4 G-K Fuzzy Clustering Algorithm Gustafson and Kessel extended the standard fuzzy c-means algorithm by employing an adaptive distance norm, in order to detect clusters of different geometrical shapes in one data set [7]. Each cluster has its own norm-inducing matrix Ai. The matrices Ai are used as optimization variables in the c-means functional, thus allowing each cluster to adapt the distance norm to the local topological structure of the data. Given the data set Z, choose the number of clusters by cluster validation as 1<c<N, the weighting exponent m > 1 and the termination tolerance 0>ε . Initialize the partition matrix randomly such that fc M∈ (0) U . The following steps are repeated using MATLAB software for l=1,2,…. ,.The G-K algorithm is same as that of fuzzy c-means, only the computation of distance measure is calculated using co- variance matrices as follows: Co-variance matrix , N 1k )( N 1k )VZ)(VZ()( F m)1l( ik T)l( ik )l( ik m)1l( ik i ∑ = ∑ = −− = − − µ µ (12) where ci1 ≤≤ Euclidian distance [ ]*1in/1iiT)l(ikiikA2 F)F((*)VZ(D −−= detρ ,)VZ( )l( ik − Nk1,ci1 ≤≤≤≤ (13) 3.5 G-G Fuzzy Clustering Algorithm The fuzzy maximum likelihood estimates (FMLE) clustering algorithm employs a distance norm based on the fuzzy maximum likelihood estimates proposed by Bezdek and Dunn [9]. The Euclidian distance functions is chosen as                  −−∑− ∑ =∑ )l( i v k z 1 i T )l( i v k z 2 1 exp* i P ) i det( i ikD , Nk1,ci1 ≤≤≤≤ (14) The priori probability of selecting cluster i is given by .ci1, N 1 P N 1k )1l( ik i ∑ ≤≤= = −µ (15) Given the data set Z, choose the number of clusters by cluster validation as 1<c<N and the termination tolerance 0>ε . The following steps are repeated using MATLAB software for l=1,2,…. ,.The Gath - Geva algorithm is same as that of G-K, only the computation of distance norm involves an exponential term and thus decreases faster than the inner-product norm. The fuzzy covariance matrix of the i th cluster is given by , N 1k w)1l( ik N 1k T)l( ik )l( ik w)1l( ik i )( )VZ)(VZ()( ∑ ∑ −− =∑ = − = − µ µ .ci1 ≤≤ (16) International Journal of Computer Applications (0975 – 8887) Volume 9– No.4, November 2010 27 The difference between the matrix Fi in GK algorithm and the i ∑ define above is that the latter does not involve the weighting exponent m, instead of this it consists of w = 1. 4. THE INVERTED PENDULUM PROBLEM Very often the quality of a control algorithm is tested and demonstrated on the inverted pendulum problem due to its inherent instability and dynamic characteristics. Figure 2 shows the free-bodied diagram of the inverted pendulum system. It is the problem of learning, how to balance an upright pole. Its solution consists of finding the horizontal force to be applied to the cart in order to balance the pole. The cart is moving on the track with no friction. Also, the pole is tied up to the cart by a frictionless hinge. Both the cart and the pole have only one degree of freedom, i.e. each of them can move in vertical plane only. Lagrange equations can be used to derive dynamical system equations for a complicated mechanical system such as the inverted pendulum. The Lagrange equations use the kinetic and potential energy in the system to determine the dynamical equations of the inverted pendulum system. Fig. 2 Inverted pendulum where, θ - Angle of the pole with respect to the vertical axis θ& - Angular velocity of the pole with respect to the vertical axis F - Force applied to the cart M - Mass of the cart m - Mass of the pole x - Position of the cart x& - Velocity of the cart The kinetic energy of the system is the sum of the kinetic energies of each mass. The kinetic energy Ec of the cart is 2 c xM 2 1 E &= (17) The pole can move in both the horizontal and vertical directions. So the pole kinetic energy is ) 2 2 2 2p zx(m 2 1 E && += (18) From the free bodied diagram x2 and z2 are equal to ) θ sinL(xx 2 += (19) ) θ cos(Lz 2 = (20) θ )θsin(Lz 2 && −= (21) θ)θcos(Lxx 2 &&& += (22) The total kinetic energy, E of the system is equal to pc EEE += (23) Substitute Eq. (17),(18), (21)and (22) in (23) θ )cosθ ( xmLθmL 2 1 xm 2 1 xM 2 1 E 2222 &&&&& +++= (24) The potential energy, V of the system is stored in the pendulum so 2 mgzV = (25) Substitute equation (20) in (25) )cosθ(mgLV = (26) The Lagrangian function is V-EP = )θ cos(mgL θmL 2 1 θx)θ (cosmLxm)(M 2 1 P 222 − +++= &&&& (27) The state-space variables of the system are x and θ , so the lagrange equations are F x P x P dt d = ∂ ∂ − ∂ ∂       & (28) 0 θ P θ P dt d = ∂ ∂ − ∂ ∂       & (29) Solving equation (28) and (29) the dynamic equations of the inverted pendulum is obtained given below 21 xx =& (30) ( ) 1 2 1 2 211 1 2 xmcosm)(ML )Fcosx (-)x)(cosx (sinx mL )(sinx m)g(M x −+ − + =         & (31) 43 xx =& (32) International Journal of Computer Applications (0975 – 8887) Volume 9– No.4, November 2010 28 ( ) ( )12 1 2 11 2 21 4 xmcosm)(M F xmcosm)(M ) (cosx )(sinx mg-)x (sinx mL x −+ + −+ =& (33) Let the state variables θx1 = ; Angle of the pole with respect to the vertical axis θx 2 &= ; Angular velocity of the pole with respect to the vertical axis xx3 = ; Position of the cart xx4 &= ; Velocity of the cart 5. SIMULATION RESULTS To simulate the above equations, the mass of the cart, M is set to 1.2 kg, mass of the pendulum is set to 0.1 kg, length of the pendulum is 0.4 meters, gravitational force, g is set to 9.81 m/s. The above state equations are simulated using SIMULINK software. It is observed that the angle of the inverted pendulum shown in Fig. 3 does not give us enough information on the inverted pendulum system. The pendulum falls over quickly and 0 500 1000 1500 2000 -150 -100 -50 0 Time (samples) A n g le ( ra d ) Fig. 3. Open loop response of the inverted pendulum it found to be unstable. One of the requirements in system identification is the collection of ‘information rich’ input-output data. In order to adequately model the inverted pendulum it is necessary to stabilize it using a nonlinear feedback controller. 5.1 Design of Nonlinear Feedback Controller Using a nonlinear feedback controller, the output data will contain more information for describing the process. The following equations are the control law developed for the inverted pendulum. θ gsin 4l 3 h1 = (34) cosθ 4l 3 h2 = (35)       −= gsin2θ 8 3 θlsinθmf 2 1 & (36)       −+= θcos 4 3 1mMf 2 2 (37) 1 2 d12d11 2 2 f xc )x(xcθk)θ(θkh h f u −      + −++−+ = & & (38) Equation (38) calculates the required force u, to keep the pendulum stable. For simulation, k1=25, k2=10, c1=1 and c2=2.6 [10]. Also xd = 0 meters and d θ = 0 rad, which are the desired position of the cart and angle of the pendulum, respectively. The inputs to this controller are the four output states of the non-liner pendulum model. The correct magnitude and force to keep the pendulum stable is calculated by the control law. The control law and closed loop model of the inverted pendulum developed by simulink. The closed loop response of the inverted pendulum with controller and force applied to the cart are shown in Fig.4 and Fig.5 respectively. 0 500 1000 1500 2000 -0.015 -0.01 -0.005 0 0.005 0.01 Time (Samples) A n g le ( ra d ) Fig. 4. Closed loop response of the inverted pendulum with Controller. 0 500 1000 1500 2000 -1.5 -1 -0.5 0 0.5 1 Time (samples) F o rc e ( N e w to n s ) Fig.5 Force applied to the cart International Journal of Computer Applications (0975 – 8887) Volume 9– No.4, November 2010 29 5.2 Generation of Input-Output Data with Nonlinear Controller Fig.6 shows that the control law is working properly. The data matrix Z was constructed from the identification data set as: T 200021 200021 ,...., u,....u,u Z       = θθθ 5.3 Simulation results for Cluster Validation Validity measures given in section 3.2 were used with Gustafson- Kessel algorithm to validate the partitioning of the Inverted pendulum data with the current number of clusters. During the optimization parameters were fixed to the following values: m=2, 1,001.0 ==∈ ρ for each cluster, ]142[c∈ . The values of the validity measures depending from the number of cluster are plotted and embraced in table.1. 2 4 6 8 10 12 14 0 2 4 x 10 -4 Number of clusters D I 2 4 6 8 10 12 14 0 5 x 10 -4 Number of clusters A D I Fig.6 Values of Dunn’s Index and Alternative Dunn Index. 2 4 6 8 10 12 14 0.4 0.6 0.8 Number of clusters P C 2 4 6 8 10 12 14 0 1 2 Number of clusters C E Fig. 7 Values of Partition coefficient and classification entropy. 2 4 6 8 10 12 14 0 0.5 1 Number of clusters S C 2 4 6 8 10 12 14 0 2 4 x 10 -4 Number of clusters S 2 4 6 8 10 12 14 0 50 100 Number of clusters X B Fig. 8 Values of Partition index, Separation Index and Xie and Beni’s Index. Fig.9. Results showing number of clusters for Gustafson-Kessel algorithm applied to Inverted Pendulum. From Fig.8 SC and S hardly decreases at c=4 point. The xb index reaches this local minimum at c=4. considering that SC and S are more useful, when comparing different clustering methods with the same c, we choose the optimal cluster to 4, which is confirmed by Dunn’s index and Alternative Dunn index in Fig.6. 5.4 Simulation results implementing Fuzzy C- means algorithm Initialization parameters of fuzzy c-means algorithm are given as: number of clusters c = 4, fuzziness of clustering m = 2, termination tolerance ε = 0.001. These parameters are fixed by trial and error. Using these initialization parameters, the fuzzy c- means algorithm given in Section 3.3 is simulated. The modeling of inverted pendulum using fuzzy c-means algorithm is shown in Fig. 10. It does not exactly matching with actual process output. International Journal of Computer Applications (0975 – 8887) Volume 9– No.4, November 2010 30 0 500 1000 1500 2000 -0.015 -0.01 -0.005 0 0.005 0.01 Time (Samples) A n g le ( ra d ) Fuzzy c-means Actual output Fig.10 Simulated angle of fuzzy C-means and actual process output. 5.5 Simulation results implementing G-K algorithm Initialization parameters of G-K algorithm are given as: number of clusters c=4, fuzziness of clustering m=2, termination tolerance ε = 0.001, expected cluster volumes ρ =1. These parameters are fixed by trial and error. Using these initialization parameters, the G-K algorithm given in Section 3.4 is simulated. The model of inverted pendulum using G-K algorithm is shown in Fig.11. It is evident from the response the model output is exactly matching with the process output. 5.6 Simulation results implementing G-G algorithm Initialization parameters of Gath - Geva algorithm are given as: number of clusters c=4, fuzziness of clustering m=2, termination tolerance ε =0.001. Using these initialization parameters, the Gath-Geva algorithm given in Section 3.5 is simulated. The modeling of inverted pendulum using Gath - Geva algorithm is shown in Fig. 12. It does not exactly matching with actual process output. 0 500 1000 1500 2000 -0.015 -0.01 -0.005 0 0.005 0.01 Time (Samples) A n g le r a d ) G-K algorithm Actual output Fig. 11. Simulated angle of G-K algorithm and actual process output. 0 500 1000 1500 2000 -0.015 -0.01 -0.005 0 0.005 0.01 Time (Samples) A n g le ( ra d ) G-G algorithm Actual output Fig.12 Simulated angle of Gath – Geva algorithm and actual process output. Table 1. The numerical values of validity measures. No. of Cluster 2 3 4 5 6 7 8 9 10 11 12 13 14 SC 0.802 0.279 0.150 0.118 0.1204 0.125 0.101 0.103 0.103 0.106 0.113 0.113 0.094 PC 0.726 0.609 0.562 0.534 0.5115 0.486 0.472 0.462 0.469 0.453 0.442 0.429 0.440 CE 0.425 0.681 0.824 0.922 1.0156 1.109 1.176 1.227 1.244 1.307 1.357 1.415 1.390 XB 50.71 19.00 7.229 7.843 7.201 10.201 10.181 9.565 5.512 6.790 5.824 4.509 6.905 International Journal of Computer Applications (0975 – 8887) Volume 9– No.4, November 2010 31 5.7 Comparison of performance of fuzzy c- means, G-K and G-G Algorithm The model outputs obtained using c-means, G-K and Gath – Geva algorithms are compared with actual output is shown in Fig.10, Fig 11 and Fig.12. From the result and performance table 2, it is observed that G-K algorithm gives better result by minimizing the Root Mean Square Error (RMSE). The performance index measured on the free-run test experiment is given by ∑ = −= N 1k 2 ))k(y)k(y( N 1 ) RMSE where, N represents the number of input-output data used for validating the model, y(k) is the actual output and )k(y ) is the model output. Table 2. Performance measure (RMSE) for a inverted pendulum. 6. CONCLUSION The design and implementation of fuzzy model based on fuzzy c- means, g-k algorithm and Gath – Geva algorithms for the inverted pendulum are discussed in this paper. The effectiveness of G-K algorithm is verified through simulation studies and the results obtained are compared with that of actual process output. By means of product-space fuzzy c-means, G-K and Gath – Geva clustering, a data set generated by a inverted pendulum system can be partitioned into fuzzy subsets of data that are locally described by linear sub models are demonstrated in this paper. Also different cluster validation methods are discussed and presented to choose optimum number of cluster. From the simulation results it is observed that the G-K clustering gives better performance for the inverted pendulum. Hence, it can be concluded that the proposed G-K algorithm can be applied to decompose any nonlinear system into local linear models and can be used to design model based control techniques such as Internal Model Control. 7. REFERENCES [1] Abonyi, R. Babuska and F. Szeifert, 2001, Fuzzy modeling with multivariate membership functions: gray-box identification and control design, IEEE Trans. On Systems, vol. 31, pp. 755-767. [2] R. Babuska, 1998, Fuzzy Modeling for Control, Boston, MA: Kluwer. [3] T. Takagi and M.Sugeno, 1985, Fuzzy identification of systems and its application to modeling and control, IEEE Trans. Syst., Man, Cybern., vol SMC-15, pp. 116-132. [4] M. Sugeno, K. Tanaka, 1991, Successive identification of a fuzzy model and its application to prediction of a complex system, Fuzzy sets and system, Vol. 42, pp. 335-344. [5] S. Arulselvi, G.Uma and M. Chidambaram, 2004, Design of Takagi-Sugeno fuzzy model based on Gustafson Kessel clustering technique for a dc-dc converter, 4 th International conference on TIMA, CSIR complex, Chennai, pp. 221-227. [6] M. Sugeno, T. Yasukawa, 1993, A fuzzy logic based approach to qualitative modeling, IEEE Trans.on fuzzy systems, Vol. 1, No. 1, pp. 7-31. [7] J. Bezdek, hybrid models for fuzzy control, 1994, Proc. Of the first Int. Symp. On Integrating Knowledge and Neural Heuristics, pp. 3-14. [8] H. Berenji, Y.Y. Chen, C.C. Lee, J.S. Jang, S. Murugesan,1990,A hierarchical approach to designing approximate-reasoning based controllers for dynamical physical systems, Proc. Of 6 th Conf. on Uncertainityin AI, pp. 362. [9] J.C. Bezdek and J.C. Dunn, 1975,Optimal fuzzy partitions:A heuristic for estimating the parameters in a mixture of normal distributions, IEEE Transactions on Computers, pp. 835- 838. [10] A.Guez, J. Selinsky, 1988, A trainable neuromorphic controller, Journal of robotic systems, Vol.5, No.4, pp 363- 388. Modeling Fuzzy c-means Gath- Geva G – K algorithm RMSE 0.0017 0.0011 0.00049598 
work_q7ajlfuhcfcmffnxawe6llzbxe	----	doi:10.1016/j.eswa.2004.12.007 An expert system for manufacturing systems machine selection Hédi Chtourou a,b,*, Wassim Masmoudi a , Aref Maalej a a Laboratoire des Systèmes Electro-Mécaniques, Ecole Nationale d’Ingénieurs de Sfax, BP W 3038, Sfax, Tunisia b Département de technologie, Institut Préparatoire aux études d’ingénieurs de Sfax, Rte Menzel Chaker Km 0.5, Sfax 3000, Tunisia Abstract This work presents the development of a prototype expert system (ES) for the machine selection of manufacturing systems. This tool, called ESMRS (Expert System for Manufacturing Resource Selection) is used in a simulation based approach in order to structure the solution search mechanism and to overcome the try and error aspect. In fact, in such an approach a number of ‘simulation—ES optimization’ cycles are run until obtaining non-improvable performance measures. The ES main role is to suggest resource modifications based on due date related performance measures obtained through simulation. So, this paper introduces the ‘ES—simulation approach’ that constitutes the utilization scope of ESMRS and then describes the ES static and dynamic knowledge representation before presenting the basic ES features as well as its development using a commercial ES shell. Finally a simple case study illustrates the validity of the approach and its potential applicability for real cases. q 2005 Elsevier Ltd. All rights reserved. Keywords: Machine selection; Expert system; Manufacturing system sizing; Simulation; Performance measure 1. Introduction The resource selection problem is defined as the specification of the number of each type of resources to use in a manufacturing system (MS) for a given planning horizon. Mainly, the approaches that dealt with this problem are either analytical or simulation-based methods. The latter are iterative approaches strongly associated with ‘try and error’ (Bullinger & Sauer, 1987; Peng, Skinner, & Mason, 2001), whereas the former are based on mathematical models connecting parameters like production needs and resource capacities to the required resource quantities (De Matta, Hsu, & Lowe, 1999; Lin & Yang, 1996; Miller & Davis, 1977). Unlike simulation based approaches, the analytical ones are limited to small size problems due to the difficulty of handling the mathematical formulations. In addition they are deterministic and tackle the resource selection problem independently from other problems such as scheduling and layout. Moreover, they do not consider 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2004.12.007 * Corresponding author. Tel.: C216 98 667 530; fax: C216 74 246 347. E-mail addresses: hedi.chtourou@ipeis.rnu.tn (H. Chtourou), wassim. masmoudi@webmails.com (W. Masmoudi), aref.maalej@enis.rnu.tn (A. Maalej). the dynamic and the stochastic aspects inherent to certain manufacturing factors like demand. Hence the results obtained by these methods lack robustness. Besides, artificial intelligence techniques such as expert systems (ES) were used in several MS design fields such as layout and capacity planning (Jayaraman & Srivastava, 1996) but very rarely in the resource selection context. In fact, an ES was used by Kusiak (1987, 1990) to decide, according to the available data and to the problem size, which mathematical model and which resolution algorithm to use. This work presents the development of an ES for MSs machine selection, called ESMRS (Expert System for Manufacturing Resource Selection). This tool is used in a simulation based approach in order to structure the solution search mechanism and to overcome the try and error aspect. Its main role is to suggest resource modifications based on performance measures obtained through simulation. So, the remaining of this paper is organized as follows. First, Section 2 introduces the Expert System Simulation Approach (ESSA) that constitutes the utilization scope of ESMRS. Then, Section 3 describes the ES static and dynamic knowledge representation, whereas Section 4 presents the basic ES features as well as its development Expert Systems with Applications 28 (2005) 461–467 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa H. Chtourou et al. / Expert Systems with Applications 28 (2005) 461–467462 using a commercial ES shell. Afterwards, an application intended to validate the developed ES is discussed in Section 5. Finally, Section 6 is dedicated to conclusion and future work perspectives. 2. Expert system simulation approach 2.1. Concept In this approach, the simulation tool utilizes the data of an arbitrary feasible initial MS as well as the demand pattern to simulate the realization of a typical manufacturing order (MO) set over a period of time corresponding to a given planning horizon. Steady state simulation results are then considered as performance measures of the system. These results, in addition to user prescribed performance limits as well as the MS data and the demand pattern constitute the ES required inputs. This tool is in charge of analyzing the MS situation. If the simulated system performance is found to be improvable, the ES recommends a modification to its resources in order to overcome the problem considered to be the most responsible for the low performance. Conse- quently, a new cycle is run until the ES is no more able to suggest any modifications (see Fig. 1). Despite the fact that the proposed approach is not restricted neither to one type of resources nor to a particular layout, this paper is focused on the specific machine selection problem of a job shop type MS with no labor constraints. Moreover, all products are manufactured by batches of a constant size following the launching of MOs. Finally, the material handling system is bi-directional and covers the totality of the departments. Fig. 1. Overview of the ESA. 2.2. Performance measures In a competitive ‘make to order’ manufacturing context, the main priority is to minimize mean batch tardiness (MT) in order to avoid associated penalties. On the other hand, it is also important to minimize mean batch earliness (ME) and its related extra storage costs. It is worth mentioning here that each MO due date is obtained by multiplying its total work content (TWK) by a user defined factor K. TWK is the sum of all processing and transportation times required to complete the MO in an ideal situation, where neither waiting nor setup are required, whilst K expresses the DD strictness as required by the user. Moreover, another performance measure is crucial for the determination of the department representing a potential bottleneck. It is the average number of batches waiting in machine queue (nw). Finally, each department utilization rate (ur) should be calculated in order to make sure it is bounded by a higher limit expressing the average machine availability and a lower limit depending on its acquisition cost. 3. Knowledge representation Two main kinds of knowledge constitutes the core of any ES, also called knowledge based system: the static and the dynamic knowledge. The former is the group of concepts describing the expertise domain, whereas the latter is the reasoning mechanism. Both are exploited by the ES inference engine either to inductively answer a question or to deductively generate new facts. The next sub sections describe the static knowledge representation through a set of objects and the dynamic knowledge representation using sets of production rules. 3.1. Static knowledge In the job shop machine selection problem using the ESSA, the main concepts are: optimization objective, performance measures, machine departments, performance limits and optimization history. Theses concepts are modeled using an object based formalism in which each object is defined by a set of private data called attributes and a set of intrinsic functions called methods. Objects are also categorized into classes and sub-classes and both attributes and methods are inheritable. The main objects that served to model the problem concepts are depicted in Table 1 that also presents their main attributes. For the sake of brevity, this table presents only the objects and attributes necessary to understand the expertise domain as well as the reasoning mechanism presented in the following section. For the same reason, the objects methods are not discussed here. Besides, it is worth noting here that the number of ‘Machine department i’ objects corresponds to the effective number of machine departments of the job shop. Moreover, Table 1 Main objects description Object Attribute Name Description Machine department i nbr Number of machines nw Mean number of batches waiting wt Mean waiting time of a batch ur Mean machine utilization rate tumax Maximum allowable utilization rate tumin Minimum allowable utilization rate us Utilization state (ok, over utilized, under utilized) pb Diagnosed problem (lack or surplus of machines) Department a Res List of all machine departments (descendents) ResLackO List of machine departments with a machine lack problem (sorted by decreasing nw and wt as tiebreaker) ResSurplusO List of machine departments with a machine surplus problem (sorted by increasing UR) Global performance MT Mean tardiness (actual cycle value) MA Mean advance (actual cycle value) MSS Manufacturing system state according to the no performance deterioration constraint (ok, NO) History MT Mean tardiness (last cycle value) MA Mean advance (last cycle value) Solj Solution of approach cycle number j a Plus all the ‘Department i’ attributes, defined for inheritance purpose only. H. Chtourou et al. / Expert Systems with Applications 28 (2005) 461–467 463 all these objects are descendents of the ‘Department’ object and hence, all their attributes and methods are inherited from this object in which they are defined but not used. Furthermore, the ‘History’ object should have as many Solj as previous ESSA cycles. These attributes are intended to verify that the proposed solution do not correspond to any previously obtained one in order to avoid infinite looping. 3.2. Dynamic knowledge The dynamic knowledge, also called ‘know-how’ can be schematized by a general resolution procedure in which each function or step is realized by a set of production rules. The resolution procedure for the job shop machine selection problem through the ESSA is shown in Fig. 2. In fact, the first ES task is to make sure that the last cycle recommendation did not significantly deteriorate the MS performance. If this test is conclusive, a new cycle starts by updating the best solution if a significant improvement is observed. If not, a new iteration of the previous cycle will take place by trying to recuperate non- explored recommendations from the previous cycle. Subsequently, the ES first tries to diagnose all tardiness problems before trying to make the MS compliant with utilization rate constraints. The earliness minimization objective comes then in the third importance rank. This ranking also governs the first test. Therefore for example, if a significant improvement of the utilization rates implies a significant increasing of tardiness, the corresponding recommendation is simply cancelled. In such a case, the ES recuperates and then proposes the next element from the last cycle recommendation list. Once all problems are diagnosed, the ES tries to establish the list of corresponding feasible recommendations. Such recommendations should ensure the feasibility of all MOs without generating a previously obtained solution. Then, the ES ranks the feasible recommendations according to the severity of the related problems. Thus, a ‘lack of resource’ problem is more critical than a surplus situation. In addition, ‘lack of resource’ problem solutions are ranked by decreasing nw order of their corresponding departments, whereas the more severe ‘surplus of resource’ problem corresponds to the department with the lowest utilization rate. It is worth mentioning here that the absence of feasible recommendation means the end of the procedure. Besides and as previously mentioned, every element of this general resolution procedure is realized by a set of production rules. These are written by means of the previously declared ‘object: attribute’ pairs. Table 2 gives four examples of generic rules constituting the core of the problem diagnosis task. These rules apply for every object D belonging to the class ‘Machine department i’ and use the writing convention 4. Expert system development An ES shell was used to develop a prototype of the hybrid Expert System for Manufacturing Resource Selec- tion (ESMRS). This development tool uses an object oriented knowledge representation formalism as well as production rules (Intellicorp, 1991). It also features an inference engine of the first order capable of forward as well as backward chaining. The first mode was used in the present application in order to solve the machine selection Fig. 2. ES general problem solving procedure. H. Chtourou et al. / Expert Systems with Applications 28 (2005) 461–467464 problem in a deductive way. Furthermore, the ‘breadth first’ option was preferred over the ‘depth first’ in order to diagnose a maximum number of problems before the chaining stops. The shell also permitted the development of a user friendly interface allowing the introduction of MS data and performance measures as well as the visualization of recommendations together with appropriate explanation. A context dependent help was also developed for the application. Fig. 3 depicts the main interface screen of the prototype ES realized for a maximum of five machine departments. 5. Application A simple case application is here presented for a preliminary validation of developed ES. The MS to be sized is a job shop composed of five machine departments (M1–M5). The typical demand pattern consists of single product MOs launched at given frequencies (Poisson law) and characterized by a certain batch size (BS). In addition each type of product (P1–P3) is characterized by a specific routing resulting in a specific total work content (TWK) obtained by summing up the processing times (PTs) on all workstations and the needed transport times. The setup Table 2 Generic diagnosis production rules Function (context) Rule Lack diagnosis (Tardiness) If [GlobalPerformance:MSS Z ZOK] Then AND [GlobalPerformance: MTOGlobalPerformance: Thsld] D:pbZlack AND [D:nwO0] Lack diagnosis (urOurmax) If [GlobalPerformance:MSSZ ZOK] Then AND [D:urOZD:urmax] D:pbZlack Surplus diagnosis (Earliness) If [GlobalPerformance:MSSZ ZOK] Then AND [GlobalPerformance: MEOGlobalPerformance: Thsld] D:pbZSurplus AND [D:nwZ Z0] AND [D:ur!ZD:urmax] AND [D:nbrO1] Surplus diagnosis (Earliness) If [GlobalPerformance:MSSZ ZOK] Then AND [GlobalPerformance:ME OGlobalPerformance: Thsld] D:pbZSurplus AND [GlobalPerformance: MT!GlobalPerformance: Thsld] AND [D:ur!ZD:urmin] AND [D:nbrO1] H. Chtourou et al. / Expert Systems with Applications 28 (2005) 461–467 465 times (STs), governed by a triangular law, are not included in the TWK since this parameter is calculated for the most favorable situation, where the machines are already set for the same product. For each MO, the DD is set to be the product of factor expressing its tightness by its corresponding TWK. The main characteristics of the MS and the demand pattern are recapitulated in Table 3. Fig. 3. ESMRS us An arbitrary initial solution, consisting of assigning n machines to machine department number n, was considered. And the first simulation run showed that most likely, department M1 is the most critical bottleneck (see first line of Table 4). Consequently, the first of the ES issued recommendation list was ‘addition of a machine in depart- ment 1’. This recommendation was taken into account and a second simulation was run. Since no deterioration was noticed er interface. Table 3 Application characteristics Product type Manufacturing data Demand data Routing (dept) PT (min) ST (min) TWK (min) BS Inter-arrival law (min) P1 M3 20 Triangular (115,120,125) 790 10 Poisson (120) M1 25 Triangular (20,25,30) M2 30 Triangular (25,30,35) P2 M2 5 Triangular (25,30,35) 840 20 Poisson (120) M3 15 Triangular (95,100,105) M4 20 Triangular (100,105,110) P3 M1 10 Triangular (20,25,30) 1240 30 Poisson (120) M4 10 Triangular (145,150,155) M5 20 Triangular (70,75,80) H. Chtourou et al. / Expert Systems with Applications 28 (2005) 461–467466 the addition of a machine in department M1 is then accepted and the ES could start finding other problems. Furthermore, since a significant improvement has been noticed the best solution was also updated. Hence 13 cycles were needed in order to obtain a non-improvable solution consisting of (5; 4; 5; 7; 5) machines in departments M1–M5, respectively. It is worth noting here that the last cycle did not generate any conclusive recommendation. Besides and for the sake of clarity, all unfeasible recommendations were displayed in a crossed format. For instance, removal of a machine from Table 4 Results of the various cycles of optimization Cycle Machine number nw ur M1 M2 M3 M4 M5 M1 M2 M3 M4 M5 M1 1 1 2 3 4 5 566 0 311 0 0 100 2 2 2 3 4 5 299 0 282 0 0 100 3 3 2 3 4 5 104 77 200 44 0 100 4 3 2 4 4 5 188 143 0 68 0 100 5 4 2 4 4 5 0 264 0 71 0 95 6 4 3 4 4 5 0 0 48 213 0 93 7 4 3 4 5 5 0 0 48 62 0 93 8 4 3 4 6 5 0 0 45 0 0 94 9 4 3 5 6 5 0 0 0 0 0 94 10 4 4 5 6 5 0 0 0 0 0 93 11 4 4 5 7 5 0 0 0 0 0 94 12 5 4 5 7 5 0 0 0 0 0 75 13 5 3 5 7 5 0 0 0 0 0 77 13 4 4 5 7 5 0 0 0 0 0 93 13 5 4 5 6 5 0 0 0 0 0 75 13 5 4 5 7 4 0 0 0 0 0 74 13 5 4 4 7 5 0 0 0 0 0 75 C, machine addition; K, machine removal. department M4 in cycle 8 was judged unfeasible since it would bring the MS to its previous state and hence will cause infinite looping. Finally, in order to illustrate the robustness of the approach regarding the initial solution, another optimization study is started using the same data but with a different initial MS. In fact, departments M1–M5 started with (7; 6; 5; 4; 3) machines, respectively. As it could be seen from Table 5, the same final solution was reached using 11 cycles. MT MA ES recommen- dations Solution update M2 M3 M4 M5 52 100 61 24 27,298 0 CM1CM3K M5KM2KM4 Yes 78 100 88 50 18,660 115 CM1CM3K M5KM2KM4 Yes 100 100 100 64 12,118 691 CM3CM1C M2CM4KM5 Yes 100 91 100 55 11,464 378 CM1CM2C M4CM3KM5 Yes 100 81 100 70 8700 2072 CM2CM4C M1KM5KM3 Yes 89 100 100 59 5431 1695 CM4CM3C M1CM2KM5 Yes 86 100 100 73 674 3813 CM4CM3C M1KM5KM2 Yes 90 100 88 77 0 6298 CM3CM1C M2KM5KM4 Yes 95 82 91 77 0 8447 CM2CM1C M4KM5KM3 Yes 69 83 93 77 0 8451 CM4CM1K M2KM5KM3 Yes 69 83 77 77 0 8498 CM1KM2K M5KM4KM3 Yes 69 83 76 77 0 8510 KM2KM1K M4KM5KM3 Yes 95 81 76 76 0 8479 KM1KM4K M5KM3 No 69 83 77 77 0 8493 KM4KM5K M3 No 68 83 91 77 0 8498 KM5KM3 No 68 83 77 100 0 8462 KM3 No 63 100 76 77 0 5613 END No Table 5 Results of the various cycles of optimization with different initial system Cycle Machine number nw ur MT MA ES recommen- dations Solution update M1 M2 M3 M4 M5 M1 M2 M3 M4 M5 M1 M2 M3 M4 M5 1 7 6 5 4 3 0 0 0 250 0 53 45 83 100 95 5819 3157 CM4CM5K M2KM1KM3 Yes 2 7 6 5 5 3 0 0 0 87 67 53 45 84 100 100 2109 3510 CM4CM5K M2KM1KM3 Yes 3 7 6 5 6 3 0 0 0 0 112 52 46 84 92 100 1799 5496 CM5CM4K M2KM1KM3 Yes 4 7 6 5 6 4 0 0 0 0 2 53 45 83 91 100 0 8413 CM5CM4K M2KM1KM3 Yes 5 7 6 5 6 5 0 0 0 0 0 53 46 84 91 77 0 8486 CM4KM2K M1KM5KM3 Yes 6 7 6 5 7 5 0 0 0 0 0 53 45 82 77 76 0 8513 KM2KM1K M5KM4KM3 Yes 7 7 5 5 7 5 0 0 0 0 0 53 54 82 77 76 0 8513 KM1KM2K M5KM4KM3 Yes 8 6 5 5 7 5 0 0 0 0 0 62 54 82 77 76 0 8513 KM2KM1K M5KM4KM3 Yes 9 6 4 5 7 5 0 0 0 0 0 63 69 82 76 77 0 8507 KM1KM2K M4KM5KM3 Yes 10 5 4 5 7 5 0 0 0 0 0 75 69 83 76 77 0 8510 KM2KM1K M4KM5KM3 Yes 11 5 3 5 7 5 0 0 0 0 0 77 95 81 76 76 0 8479 KM1KM4K M5KM3 No 11 4 4 5 7 5 0 0 0 0 0 93 69 83 77 77 0 8493 KM4KM5K M3 No 11 5 4 5 6 5 0 0 0 0 0 75 68 83 91 77 0 8498 KM5KM3 No 11 5 4 5 7 4 0 0 0 0 0 74 68 83 77 100 0 8462 KM3 No 11 5 4 4 7 5 0 0 0 0 0 75 63 100 76 77 0 5613 END No C, machine addition; K, machine removal. H. Chtourou et al. / Expert Systems with Applications 28 (2005) 461–467 467 6. Conclusion This work presented the development of an ES used in a simulation based approach in order to structure the solution search mechanism. A prototype of this ES has been developed and tested. This rule based and object oriented tool permitted to obtain satisfactory preliminary results in the sizing of a simple MS belonging to predefined domain of application. However, many aspects of the approach, are currently being developed. In fact, the main issue is the enlargement of the domain of application by incorporating other types of resources and by considering resource reliability and routing flexibility. This should allow applying and validat- ing the approach on real cases. Consequently, the ES reasoning mechanism should be enriched by incorporating new knowledge acquired from sets of planned simulations. Finally, a thorough investigation of the approach robustness should be conducted in order to assess its applicability in various scenarios. References Bullinger, H. J., & Sauer, H. (1987). Planning and implementing a flexible assembly system supported by simulation. International Journal of Production Research, 25(11), 1625–1634. De Matta, R., Hsu, V. N., & Lowe, T. J. (1999). Capacitated selection problem. Naval Research Logistics, 46(1), 19–37. Intellicorp, Kappa users guide, version 1.2, 1991. Jayaraman, V., & Srivastava, R. (1996). Expert systems, in production and operations management. International Journal of Operations and Production Management, 16(12). Kusiak, A. (1987). Artificial intelligence and operation research in flexible manufacturing systems. Information processing and operation research, 25(1), 2–12. Kusiak, A. (1990). Intelligent manufacturing systems. Englewood Cliffs, NJ: Prentice-Hall. Lin, Z. C., & Yang, C. B. (1996). Evaluation of machine selection by the AHP method. Journal of Materials Processing Technology, 57(3–4), 253–258. Miller, D. M., & Davis, R. P. (1977). The machine requirements problem. International Journal of Operation Research, 15, 219. Peng, Q., Skinner, G.E., & Mason, S.J. (2001). Sizing a pilot production line using simulation. Proceedings of the 2001 winter simulation conference. An expert system for manufacturing systems machine selection Introduction Expert system simulation approach Concept Performance measures Knowledge representation Static knowledge Dynamic knowledge Expert system development Application Conclusion References 
work_q7rkao5a2fgllf2ctwv7imjw4a	----	INMAMUSYS: Intelligent Multiagent Music System Miguel Delgado, Waldo Fajardo, Miguel Molina-Solana∗ Department of Computer Science and Artificial Intelligence, Universidad de Granada, Daniel Saucedo Aranda s/n, 18071 Granada, Spain Abstract Music generation is a complex task even for human beings. This paper describes a two level competitive/collaborative multiagent approach for autonomous, non-deterministic, computer music composition. Our aim is to build a high modular system that composes music on its own by using Experts Systems technology and rule-based systems princi- ples. To do that, rules issued from musical knowledge are used and emotional inputs from the users are introduced. In fact, users are not allowed to directly control the composition process. Two main goals are sought after: investigating relationships between computers and emotions and how the latter can be represented into the former, and developing a framework for music composition that can be useful for future experiments. The system has been successfully tested by asking several people to match compositions with suggested emotions. Key words: Computer music, Music composition, Multiagent systems, Rule-based system 1. Introduction Surely music is one of the most difficult human dis- ciplines. It requires creativity, specific knowledge and eventually, some manual abilities. It is commonly said that music is something humans do, but something we do not understand. There are several mechanisms in- volved in music and unfortunately, many of them are still unknown. Others are so complex that we cannot manage them with current computer tools. On the other hand, Computer Science has suffered a huge evolution in just a small period of time. However, despite all the great applications and problems solved during the last decades, there are many problems that computers are unable to deal with nowadays. Among others, we can point fuzzy representation of abstract concepts and, in general, dealing with feelings and emotions. In fact, cognitive processes involving reasoning, knowledge and experience are hardly repre- sented in a computer and are current hotspots for Artifi- cial Intelligence. Music is a good example of applied AI related with those topics. It has been demonstrated that music com- position is a hard task even for humans, so using ma- chines appears as a very interesting and fascinating area ∗Corresponding author Email addresses: mdelgado@ugr.es (Miguel Delgado), aragorn@ugr.es (Waldo Fajardo), miguelmolina@ugr.es (Miguel Molina-Solana) of research. AI techniques are going to be applied to the musical domain with the aim of understanding human musical abilities. Because there are so many challenges to deal with, a bottom-up approximation is required for solving all of them in a modular way. This paper is organized as follows. In Section 2, we examine some previous works done in the field. Sec- tion 3 deals with the architecture of the proposed sys- tem. In Section 4, we discuss design and implementa- tion of system components. Section 5 presents an evalu- ation method and summarizes some users’ impressions about the output; and in the last section, future work to be done is presented and discussed. 2. Background Since first computers were developed many people have tried to apply them to musical tasks. There are two main classes in which computer mu- sic projects could be classified: analysis and composi- tion. The first one consists on extracting information from the music itself (or the associated data) in order to learn some rules, or go to a model that describes the concrete examples. Because this is not the main field of this paper, so we are not going to go further. However, [1], and [2] could be reviewed for more information. Composition is about generating new music from the rules. In fact, is doing the process in the other way: Preprint submitted to Expert Systems with Applications August 30, 2016 ’from rules to music’ instead of ’from music to rules’. According to [3], the final objective of most of the com- positional prototypes is to demonstrate that standard musical techniques could be handled by computer pro- gramming, and also to validate generative music theo- ries. Initial approaches in algorithmic composition con- sisted on randomly selecting notes (mainly pitch and rhythm) with some constraints in order to generate com- positions. This vision produced limited results but was a great point of departure for later works. Experts systems has been widely used to compose music. Rules concerning pitch, duration and volume have tried to apprise the knowledge involved in human composition. [4] introduced an early example of rule system. Many works on computer music generation are based on the approach that the composition rules are specified by the composer. So that, an expert is needed in order to give all the knowledge. The problem here is that Music Theory is not as formal as it should be to be easily represented in a computer. However, [5] proposed an interesting model to generate music using grammars. Another approach to compose music with computers is to use genetic algorithms. Genetics have been suc- cessfully applied to several problems with difficulty in defining the solution process, or where searching a huge solution space is needed. Composition falls into this class of tasks, and some works in this direction can be found in [6] and [7]. Although many advances have been done during last decades in Computer Music in general and in algorith- mic composition in particular, it is true that the greatest moment was when [8] presented EMI Project. Many prototypes, throughout the years, have demon- strated that computers algorithms cannot be compared with human minds. Machines just produce very simple compositions in a quite mechanical manner. Computer programs that perform music with proper stylistic considerations (like a human expert would) are very scarce and only work for a few examples and in very specific domains may be found. The main problem is that algorithms rarely deal with feelings. And music without emotions is unworthy and not natural. [9] expressed his conviction that the unique way for creating a machine whose creations transmit something to listeners, should start by simulating emotions in com- puters. [10] addressed this topic (expression in comput- ers) from a musicological point of view, while [11] pro- posed a cognitive model for understanding melodies. This topic is a current hotspot in computer music, and many groups are working on it right now. [12], [13], Figure 1: First level architecture [14], [7], [15] and [16] can be mentioned as some rele- vant examples. For further information about compos- ing music with computers, [17] and [18] works should be reviewed. In recent years, agent paradigm has become quite common. Almost everybody is using it for everything. The reason is that agents are a very powerful way of implementing distributed AI. Indeed, they can be com- bined with others tools, such as rule-based systems, case-based systems or searching algorithms. Actually, Agent Theory just defines interactions between agents, not how they are internally built. Because of that, we can have a multiagent system where agents can be implemented with just an algorithm, using CBR tech- niques, rules, genetics... Interesting work related with multiagent systems can be found in [19] and [20]. 3. Architecture The architecture we propose in this paper is a two- layer multiagent system. The current application of multiagent systems in real-time environments is an area of increasing interest. In general, multiagent systems are an appropriate approach for solving inherently dis- tributed problems, whereby clearly different and inde- pendent processes can be distinguished. The first level is the competitive one, where agents (called composers) compete among themselves to be the one chosen for composing. This layer allows us to make an initial separation between composition styles. We think it does not make any sense to have a one-for-all agent, so a collection of simple agents specialized in some task is proposed. Each composer announces its abilities, and the system, according with user inputs, se- lects the composer that better fits. See Fig. 1. 2 Figure 2: Second level architecture This agent chooses, between other parameters, rhythm, number of voices, and instruments to be used in the compositions. However, the selected composer only acts as a director, being useless on its own. Going further with the modularization principle, the composer agent finally asks some others agents from the second level for their collaboration in order to get a solution 1. This second level contains auxiliary agents that col- laborate between them, so this layer should be under- stood as a collaborative level. We call these agents voice generators, because mainly that is their job. See Fig. 2. In general, there are just three voices in normal com- positions: melody, accompaniment and harmony. How- ever, the system presents no limitations in relationships between agents. A composer agent can employ as many voices as desired, and in the other hand, several com- posers could use the same voice generator. When some intelligence is needed, agents are to be designed as intelligent rule-based systems. Otherwise, they just implement a simple algorithm. At the end, we have a big set of simple agents that together, manage to find a whole solution for the problem. 4. Design INMAMUSYS current prototype has several com- posers that generate music in different ways: from just a random composer to more elaborate ones where aes- thetic principles are mainly searched. 1Solution in this context means just a composition that the system offers as output, without evaluating its quality An important design principle that is going to guide our system is creating a tool that can be used by every- one. Most prototypes present a very complex interface that only experts can understand, an even for them, it is very annoying to fill a huge amount of data in order to get a composition. In fact, when people speak about music, they do not usually use technical terms such as granularity, rhythm or tonality; they use emotions, feel- ings and abstract concepts. Actually, they are not speak- ing in low level, but in a high level, where technical pa- rameters do not exist. Keeping that in mind, our aim is the users ought to be faced with a friendly interface with only a few questions; and not really difficult ones, just questions about the wished music style. In other words, a high level music interface is to be introduced. In fact, we are integrating into the system itself the knowledge about how to use tonality, rhythm and in- struments in order to get, for instance, a sad music. Until now, this task has been left in the hands of hu- mans. We propose to go further providing the machine with the suitable resources it should use to compose in a certain way. The human does not have to worry about dozens of parameters and how they should be combined to get a certain result. Java has been chosen as programming language due to its object oriented style and ease of use. jMusic li- brary is used for music representation and processing. This library is described by [21]. 4.1. Knowledge representation We have made use of several ways to represent in- formation. Firstly, because most agents are designed as rule-based systems, we need to represent some informa- tion through IF-THEN rules. They are implemented in XML in order to achieve an universal way of represen- tation. Fig. 3 shows the XML schema for rules in the DTD language. Basically, a knowledge base is a set of several rules. A rule has two parts: IF and THEN. IF part consists of some attribute-value pairs; and THEN has a list of consequents, that can be viewed as actions to be done. We have decided to employ several rules subsets in order to respect the modularity, and to make the inference engine’s job easier. Higher performance is also reached with this approach. To write code, we use the representation that jMu- sic API offers us. This is an object oriented one, very close to normal Western classical music representation. There are manuals and several examples in jMusic web- site (http://jmusic.ci.qut.edu.au). As agents are devel- oped for a concrete musical style or concrete task, there is obviously a lot of heuristic information about how to compose coded in each agent. Agents are implemented 3 Figure 3: XML schema in DTD language for representing rules in the system with certain parameters (that only an expert knows) in order to accomplish its goals. This kind of knowl- edge should be minimized and moved to a unique and standard repository, in order to obtain an independent knowledge database. However, many times it is not easy to code procedural information in a data format, so we should be aware that much knowledge about the topic is into the composition process itself. System output is a sound file in MIDI format. This representation is preferred rather than others audio for- mats such as wav or mp3, because it is easier for sym- bolic manipulation. Even more, MIDI could be easily converted into the others two (if needed), but the op- posite is not true. Eventually, a music sheet could be generated if needed. 4.2. Composition Agents Composer agents, as we have previously said, are lo- cated in the first layer of the architecture. They compete with others in order to be elected for composing. Basi- cally, all composers are in charge of defining the num- ber of voices the composition will have, instruments to be used, the measure and tonality. They also indicate which second level agents should be used for generating voices. In other words, composers act as directors. In current system there are four implemented composers: Muzak, Dark, Scales and Random. Muzak Composer generates music in a Muzak2 way. This kind of music, also called ambient or elevator mu- sic, could be described as a soft and quiet one. Brian Eno wrote ”Ambient Music must be able to accommo- date many levels of listening attention without enforcing one in particular; it must be as ignorable as it is interest- ing”. From this point of view, ambient music does not need to be very elaborate or complex. Its only requisite is to be pleasant and agreeable to the human ear. Dark composer aim is to compose music that pro- vokes fear in listeners. To do that a set of suitable re- 2MUZAK is the name of a company specialized in this kind of music. People often refer to ambient music in this way. sources is used. Dissonances and diminished fifth in- terval (known as tritone) are heavily employed by this agent. This interval is called diabolus in musica (the Devil in music) and has been historically avoided. Great distances between voices and deep bass chords are also useful resources that are employed. Scales Composer is quite simple. It just produces some ascendant and descendent scales in random tonal- ities and modes. The mixing of several voices doing the same in different moments, velocities and tones pro- duces interesting effects. Finally, we introduce the Random Composer with its two variants: the differential, and the independent. In the first one, intervals are randomly generated, so a note pitch depends on previous note. The other option is to randomly generate pitches, so a note is independent of the rest. Durations are also generated with a random generator. As the reader can imagine, a composer act- ing this way produces chaotic compositions without any internal coherence. So that, chaos is the tag that better defines the music this composer generates. 4.3. Voice generators Voice Generators are in charge of producing sounds with different volumes, durations and intonations for ev- ery melodic line. Many kinds of voice generators could be imple- mented, but the main ones fall into any of the following classes: melody, harmony, accompaniment and drums. However, the system architecture is flexible enough to allow any other desired voices. In fact, the architecture does not care about the semantic meaning of a certain melodic line. 4.3.1. Harmony generator Harmony generator makes use of a set of rules that indicates which tonal movements are allowed. In this way, chord progressions can be generated. With this, we have the skeleton of compositions. We can under- stand this as a Markov chain model. In Fig. 4 we can see the set of rules in the current implementation, that correspond with Fux’s counterpoint rules ([22]). 4.3.2. Melody generator Melody generators produce the melody of composi- tions. Current design uses a big set of motives which are one measure long. A motive is randomly selected in each measure and put into the melody voice. However, preliminary tests reveal that acting this way, composi- tions will not be coherent, in the sense that they will not seem as a unique entity, but rather, like little pieces stuck together. 4 Figure 4: Rules used by harmony generator to generate chord progres- sions The trick here consists in to randomly select (from the whole set) a subset of motives to be used in each execu- tion. To have a small number of motives for each exe- cution makes the composition appear as a whole, giving the impression of melodic phrases. In fact, there is no structure at all to be followed, but the repetition over and over again of similar motives induces the listener to think so. Employing this simple trick the output seems to be organized in phrases (but as seen there is no more than random selection) and the impression is quite real- istic. Moreover, due to the fact that the subset is randomly selected, it is different for each execution. This is an important point because it produces a great variety be- tween two compositions, even thought when they are generated with same inputs. 4.3.3. Drums generator Drums generators are in charge of generating a con- tinuous rhythm. There are several composers that can be classified within this class. They implement different rhythms: some offers more density, others are lighter; ones use various sounds (drums, snare, cymbal, hat...), others just use a simple drum. At the end, it is the first level agent (the composer) the one in charge of selecting Figure 5: INMAMUSYS’s user interface the appropriate Drum Generator (or even none of them) for each occasion. 4.4. User interface Our main aim was to develop an easy and simple in- terface that anyone could use without problems. Before, we talked about the purpose of implementing an easy-to-use system. The objetive here is presenting to the users an interface that anyone can use, even if they are not experts. To do that, we need to include within the system all the information related to technical parameters. This is a great deal because the matching between emotions and music parameters is not direct, and to define parameters themselves is definitively not a trivial task. The interface (see Fig. 5) is very dynamic and not fixed because its contents depend on the composers im- plemented in the system and the capabilities they an- nounced. The form only contains terms that agents have declared in some rule. With this, we avoid unexpected inputs and assure the system will always be able to give us a valid output. Also, it lets us add, replace, or mod- ify agents when any problems appear, making changes easier. In addition, we have added the possibility of selecting which instruments should be used in each voice. How- ever, if we do not specify an instrument, the system selects one from those it thinks is a better fit. Not all instrument combinations sound good, and even if they do, it is possible that they are not compatible with the style of the melody. These options exist in order to demonstrate the complexity of the knowledge involved. Choosing rhythm, instruments, tonality or granularity are not trivial tasks, because they should be consid- ered all together, and related with some other concepts. By allowing selection of instruments, users can test if they are able to find combinations that really make great compositions, beating system choices. 4.5. The composition process In this subsection, the composition process of Muzak Composer is described to illustrate how a composer runs 5 because the process is quite similar for the rest of com- posers. When Muzak composer is selected by the Manager Agent, it firstly decides the tempo of the composition attending to user inputs. The mapping between linguis- tics tags and exact bpm (beats per minute) is done by us- ing a normal random number generator with given mean and variance. The means values for each tag have been statistically obtained from both a set of classical sheet music and experts’ feedback. As said before, composer agents are also in charge of selecting the number of voices to be used, and the second level agents are responsible for their generation. To begin with, the harmony generator is executed, producing a chord progression which is the skeleton of the composition. After that, it is the melody generator’s turn. For this agent, and in the current implementation, a mo- tive database is decided to be used (as commented be- fore), as well as employing one of its elements in each measure. Even thought there is no kind of structure of phrases, or a grammar to be followed, the global com- position seems to have an internal structure. This effect is suggested by frequent motive repetitions. Next step is to execute the accompaniment generator. Goal here is to complete the harmonization of the com- position. Accompaniment generator introduces some new notes and completes chords. It could just be a note every measure, an arpeggio or perhaps something more elaborate. It is important to indicate that melody and accompa- niment are generated always in C major. This is not a handicap and greatly facilitates the task. However, it is necessary to transpose them to the correct position in each measure, according to the tonality and the current grade. Also, the note in the first time of each measure must be accentuated. It is up to voice generators to as- sure that. 5. Experiments and evaluation The evaluation of any musical work is a complex task and often comes down to individual subjective opinion. It depends not only on formal aspects but also on some stylistics ones. Because of that, it is hard to empirically evaluate music compositions, and therefore it is difficult to evaluate the effectiveness of a computer music com- position system. Many metrics could be developed, but all of them will fail as at the end, music (understanding the term as much more than just a chain of sounds) cannot be reduced to a number. Due to this difficulty, many authors will typically con- clude their papers with a vague comment such as “com- positions generated by the system are quite impressive and very promising” or “sometimes, melody seems to be a bit simple and unelaborated; but many, results are very human like”. However, this is unsatisfactory for two reasons: first, evaluating the music produced by the system reveals little about its utility as a compositional tool; and second, qualitative and subjective evaluation by the designers of the system reveals little about the value of the tool to other composers ([3]). This author affirms that there is an historical malaise in adopting suitable evaluation procedures for judging the degree to which the aims have been satisfied. As we agree with this assertion, we have also proposed an evaluation method to assess compositions in the same way that they are usually appraised: through audience reactions and critical reviews. So that, a short test has been design to carry out that task. 5.1. Experiments To begin with, we asked some people for listening and evaluating some examples generated by INMA- MUSYS. Most of them could not believe that a machine was the real author. For us, this is an important point be- cause capturing the essence of a human’s composition method is a great deal. This fact makes us also think that our system could have passed, in some sense, the Turing test. The second part of the proposed evaluation method was more formal and consists in testing what emotions were induced on listeners when listening to some com- positions. The objective was to probe whether composi- tions generated by the system provokes in listeners emo- tions such as those that guided the composition process and were given as inputs to the system. This test will give us a measure of how the system is able to com- pose music that successfully matches user emotional re- quests. Four compositions were generated with different in- put values. Songs A, B, C and D are respectively gen- erated by using the inputs worry, happiness, chaos and worry again (this input is used twice). System mapped these emotions into the use of Dark, Muzak, Random and Dark composers. The four examples were presented to the listeners without giving them information about the origin of the songs, and the listeners were asked to select emotions that better fit what they listen. Users could tick as much tags as they want from the list: sad- ness, happiness, fear, worry, chaos and indifference. Twenty people, involving a huge range in musical ex- 6 Figure 6: Results from the test: Feelings that the four tested songs provoked in listeners pertise, participated in our experiment by answering the questions. Results can be found in Fig. 6. We were really sur- prised by them because we did not expect such a clear confirmation, even more when the emotion-composer matching mechanism is quite primitive. In all the four cases, the most repeated options is the right one (the one that was given as input to the system). It is true that the list of potential options is small, (we would like to make it bigger the next time), but it is also clear that differences between some tags are very small, so that it would have been easy to get some wrong answer. It is also noticeable, that many people (12 out of 20) point out that the first song and the last one were very similar. They are definitively different but these answers make us think that Dark composer outputs do not cover the whole area at the solution space that we supposed, but just a little zone. We think this should be corrected in the future, because is important to accomplish the ob- jective that all the compositions from the same inputs fall in the same area, but it is also crucial to maximize this area, not just focusing in a point. To conclude, we would like to comment that an equi- librated system has been developed: on the one hand, there is a great variety between different executions; on the other, we have managed that each composition sounds as a whole not like several pieces stuck together. As said, the whole solution space for a tag (e.g. fear) is not completely covered, but at least we can say that almost all the system outputs are classified (by humans) under the right tag. 6. Conclusions and future work Composing music is a very complex process that in- volves many disciplines and tasks. In this paper we have presented a new approach for composing music using computers. Because there are so many challenges to deal with, a bottom-up approximation is required for solving all of them in a modular way. We are inter- ested in building a framework for successfully compos- ing music that provokes some feelings in listeners. In order to achieve this goal, we have proposed here a two-level architecture that successfully deals with the complex problem of music composition, so that, a huge problem can be broken into smaller tasks. This ap- proach makes use of several agents and rule-based sys- tems. It also permits that users were provided by an easy-to-use interface that hides all the complexity of music composition. Even more, inputs for this inter- face are emotional inputs from the users, so that we are able to address the problem of music expressiveness. Several design decisions that affect the system were taken and we have explained them. We have focused in knowledge representation because it is a main topic. In that section we also describe different agents, as well as the role they play in the whole system. At the mo- ment, just four kinds of composers have been devel- oped: Dark, Muzak, Scales and Random. Finally, results obtained after evaluating the current system are showed. This evaluation has been carried out with an experiment and in a formal way. Even though the system is in a quite early stage of development, re- sults are promising enough to encourage us to continue working with this framework. 6.1. Future work The current prototype is quite rigid in the way it deals with user inputs. By now, we are just using classic rule matching for selecting agents, and that is a poor ap- proach for dealing with fuzzy concepts, such as emo- tions and musical terms. So that, including fuzzy logic in the inference system is probably the first modification to be done. This change would requiere major modifica- tions in the inference engine, as well as the definition of fuzzy domanisn and liguistic tags. However, we think it would be worth the effort. Secondly, we would like to develop more composers in order to get a bigger collection of these kinds of agents. Our aim is to obtain as much diversity as possi- ble not only to compare differents algorithms and com- positional mechanisms, but for implementing and test- ing new ideas. These ideas could be related with new theories of musical styles or cognitive processes. Even more, as the set of composers growths, we will be able to reproduce a wider range of human emotions. Another interesting project would be to develop a module to automatically obtain composition rules. In 7 the current prototype, this knowledge is given by hu- mans, and coded into the agents. It would be very pow- erful if the system could analyze a music sheet, then to extract some rules, and thus to compose according to them. The system will win a great flexibility with this ability. Not in vain, this one is the key idea in Cope’s system, and an interesting hotspot in current data min- ing. Doing this, the system would be quite complete, in the sense that it would include both composition and analysis tasks. As seen, there is plenty of room for researching in the computer music area in general and in composition in particular. At our department, we feel computer music is a great opportunity for modern Artificial Intelligence. We have developed this project with the aim of it being used as a framework for future experiments and works. We also believe that Expert Systems and Agent The- ory have many things to say in this field. Composing music is a huge problem that should be divided into several tasks, and it definitively needs intelligence to be done. Acknowledgements This research has been partially supported by the Spanish Ministry of Education and Science under the project TIN2006-15041-C04-01. References [1] C. Anagnostopoulou, G. Westermann, Classification in music: A computational model for paradigmatic analysis., in: Interna- tional Computer Music Conference, 1997, pp. 125–128. [2] M. Balaban, The music structures approach in knowledge repre- sentation for music processing, Computer Music Journal 20(2) (1996) 96–111. [3] M. T. Pearce, D. Meredith, G. A. Wiggins, Motivations and methodologies for automation of the compositions process, Mu- sicae Scientiae 6 (2) (2002) 119–147. [4] A. Friberg, Generative rules for music performance: A formal description of a rule system, Computer Music Journal 15 (2) (1991) 56–71. [5] F. Lerdahl, R. Jackendoff, A generative theory of tonal music, Cambridge, Mass: MIT Press, 1983. [6] M. Marques, V. Oliveira, A. Rosa, Music composition using ge- netic evolutionary, in: 2000 Congress on Evolutionary Compu- tation, 2000, pp. 714–719. [7] E. R. Miranda, At the crossroads of evolutionary computation and music: Self-programming synthesizers, swarm orchestras and the origins of melody, Evolutionary Computation 12 (2) (2004) 137–158. [8] D. Cope, Experiments in music intelligence, Madison, WI: A-R Editions, 1996. [9] M. Minsky, A conversation with Marvin Minsky. Understand- ing musical activities: Readings in AI and music, Menlo Park: AAAI Press, 1991. [10] L. Meyer, Emotion and meaning in music, Chicago, IL: Univer- sity of Chicago Press, 1956. [11] E. Narmour, The analysis and cognition of basic melodic struc- tures: The Implicacion-Realization model, Chicago, IL: Univer- sity of Chicago Press, 1990. [12] H. Kiendl, T. Kiseliova, Y. Rambinintsoa, Use of fuzzy strate- gies for interpretation of music, Journal of Multiple-Valued Logic and Soft Computing 12 (1) (2006) 149–169. [13] R. L. de Mantaras, J. L. Arcos, Ai and music. from composition to expressive performance., AI Magazine Fall (2002) 43–58. [14] Q. Zhang, E. R. Miranda, Towards and evolution model of ex- pressive music performance., in: Sixth International Conference on Intelligent Systems Design and Applications, ISDA’06, 2006, pp. 1189–1194. [15] G. Widmer, The musica expression project: A challenge for machine learning and knowledge discovery, in: 12th European Conference on Machine Learning, EMCL 2001, 2001, pp. 603– 614. [16] A. Wieczorkowska, Towards extracting emotions from music, in: Intelligent media technology for communicative intelli- gence, IMTCI, 2004, pp. 228–238. [17] D. Cope, Computers models of musical creativity, The MIT Press, 2005. [18] E. R. Miranda, Composing music with computers, Oxford, UK: Focal Press, 2001. [19] P. M. Todd, G. M. Werner, Frankensteinian methods for evo- lutionary music composition. Musical networks: Parallel dis- tributed perception and performance., Cambridge, MA: MIT Press/Bradford Books, 1999. [20] R. D. Wulfhorst, L. Nakayama, R. M. Vicari, A multiagent ap- proach for musical interactive systems, in: Second International Joint Conference on Autonomous Agents and Multiagents Sys- tems, 2003, pp. 584–591. [21] A. R. Brown, A. C. Sorensen, Introducing jmusic, in: Aus- tralasian computer music conference, 2000, pp. 68–76. [22] W. Piston, Counterpoint, W.W. Norton, 1947. 8 
work_qakoronh2rgybldob74jhpfznu	----	which bring software engineers (broadly t e r m e d to include expert s y s t e m developers) in c o n t a c t w i t h students of those other disciplines. While the c o m p u t e r science s t u - dents refine their k n o w l e d g e about particular issues in c o m p u t e r science, they also reap the benefits of exposure to other fields such as, psyschology, sociology, or any of the humanitits. They will extend their rational capabilities and will have increased s y m p a t h y for o t h e r fields of study. Likewise, the students of the d o m a i n field w i l l i m p r o v e their c o m m a n d of that study, and at the same t i m e they gain appreciation of c o m p u t e r science. Learning i n t e n - sifies. In Socrates, even the expert claimed to have benefited through the conflicts that she found in DSM-III. CITATIONS Bobrow, Daniel G., Sanjay Mittal, and Mark J. Stefik. Expert Systems: Perils and Promise," in C o m m u n i c a t i o n s of th_e ACM. V o l u m e 29,9 (September, 1986.) Buchanan, Bruce G., David Barstow, Robert Bechtal, et al "constructing an Expert System," in Building. E_~P_~ S_ystems. Hayes-Roth, Waterman, Lenat eds. Reading, Mass: Wesley Publishing Company, Inc. (1983) 127-67. Diag.nostic and Statistical Manual of Mental Disorders, Third Edition. Washington, DC: A m e r i c a n Psychiatric As- sociation, (1980). Fairley, Richard. S o f t w a r e Enci~._eerin_9 Concepts. New York: M c G r a w - H i l l Book Company, (1985). Fellers, Jack W. "Key I-actors in Knowledge Acquisition," in C o m p u t e r Personnel. A quarterly p u b l i c a - tion of the Special Interest Group on C o m p u t e r Personnel Research: ACM Press. Volume II, 1 (May, 1987). Frank, Edwina D., Phd.D., Professor of Psychology, Loyola University, New Orleans, LA: i n f o r m a l interviews and p r o j e c t reviews. (January, 1987 - - August, 1987) Harmon, Paul, and David King. Expert Systems: A r t i f i - cial Intelligence in Business. New York: John Wiley & Sons. (1985). Hayes-Roth, Frederick, Donald A. Waterman, and Douglas B. Lenat. "An O v e r v i e w of Expert Systems," in Building Expe~ Systems. Hayes-Roth, Waterman, Lenat eds. Reading, Mass: Wesley Publishing Company, Inc. (1983) 3-29. Page-Jones, Meilir. The Practical Guide to Structured Systems Desi.gLn ~. E n g l e w o o d Cliffs, N J: Y o u r d o n Press. (1980). Sagalowicz, Daniel. Quoted by Edith Meyers in "Expert Systems: Not for Everyone," in Datamation. (May 15, 1987). Welburn, T y l e r Advanced Structured Cobol: Batch, O n - l i n e , and Data-Base Concepts. Palo Alto, CA: Mayfield Publishing Company, 1983. Zuckerman, Edward, Ph.D., Professor of Psychology, Wellness Resource Center, University of Pittsburg. Pit- tsburg, PA: c o r r e s p o n d e n c e and t e l e p h o n e interviews. (January 1987). AN EXPERT SYSTEM FOR AN I D I O S Y N C R A T I C D O M A I N : LOVE, I N T I M A C Y , A N D FRIENDSHIP J o s e p h S. Fulda Hofstra University (C) 1987 Joseph S. Fulda Expert systems conceived and d e v e l o p e d thus far have dealt with p r o b l e m s , such as medical diagnosis, c i r - cuit design, and mineral exploration, which have o b j e c t i v e solutions. Humans, h o w e v e r , are also experts in reasoning about domains where the key concepts are not o b j e c t i v e l y fixed. We do f o r m subjective, but sound, j u d g m e n t s about i d i o s y n c r a t i c - - i . e . , p e r s o n - s p e c i f i c - - domains. One such d o m a i n is that of love, intimacy, and friendship. Despite the fact that our n o t i o n s of love, intimacy, and friendship vary greatly and that these relationships have been the subject of poets, novelists, biographers, p s y c h o l o g i s t s , and o t h e r students of human nature for centuries, there is enough of a c o m m o n consensus to make c o m m u n i c a t i o n and, very frequently, c o m m o n j u d g m e n t s a b o u t this d o m a i n possible. We present a r u l e - b a s e d expert system, c u r r e n t l y b e - ing i m p l e m e n t e d in PROLOG, which provides a general m e c h a n i s m for handling idiosyncratic domains. Our s y s - t e m also provides a t h o r o u g h example of h o w the a t t a i n - m e n t of c o g n i t i v e - a f f e c t i v e states and resulting r e l a t i o n - ships between people can be m o d e l l e d using a r u l e - b a s e d system. Thus this w o r k should interest c o g n i t i v e s c i e n - tists as well as c o m p u t e r scientists seeking t o e n d o w c o m p u t e r s with truly h u m a n - l i k e thought. The key to our design is, as m i g h t be expected, t o a b s t r a c t out w h a t e v e r is c o m m o n to m o s t people's u n d e r - standing of the nature of a particular class of r e l a t i o n - ships. Nevertheless, there will be a substantial failure rate if idiosyncratic criteria f o r the a t t a i n m e n t or b l o c k i n g of relationships are s i m p l y ignored. Therefore, w e a d o p t the f o l l o w i n g f o r m a l i s m : Iff X (a) Y (b) Then Z (c). This b i d i r e c t i o n a l rule is t o be u n d e r s t o o d as f o l l o w s : If X & Y then Z with p r o b a b i l i t y c; If Z then X w i t h p r o b a b i l i t y a; If Z then Y w i t h p r o b a b i l i t y b. (Of course, " I f Z then X" is identical to "Only if X then Z.") Thus a, b, and c are w h a t have been called c e r t a i n t y factors or, m o r e gracefully, a t - t e n u a t i o n factors (1). (Note that we are nod. associating certainty factors with the a n t e c e d e n t c o n d i t i o n s of the rules in either direction, t h o u g h it is trivial to add these.) The c o n c e p t of a t t e n u a t i o n factors was d e v e l o p e d f o r the MYCIN system, a r u l e - b a s e d expert system f o r m e d i c a l diagnosis (2). This f o r m a l i s m using b i d i r e c t i o n a l rules w i t h attenuation factors o e r a t i v e in both d i r e c t i o n s (cf. 3, 4) neatly handles the p r o b l e m s inherent in reasoning about an idiosyncratic domain. These are t w o : s o m e s o n e may have idiosyncratic r e q u i r e m e n t s for the a t t a i n m e n t or SIGART N e w s l e t t e r , July 1988, N u m b e r 105 Page 30 b l o c k i n g of r e l a t i o n s h i p Z or he m a y i d i o s y n c r a t i c a l l y i g - n o r e s o m e of t h e e l e m e n t s that m o s t do use in d e c i d i n g w h e t h e r to e n t e r r e l a t i o n s h i p Z. Thus, in the e x a m p l e above, if c is .7 this m e a n s that 30% of t h e i n s t a n c e s w h e r e X and ¥, n e v e r t h e l e s s n o t Z b e c a u s e of i d i o s y n c r a t i c r e q u i r e m e n t s f o r t h e a t t a i n m e n t or b l o c k i n g or t h e r e l a t i o n s h i p . Likewise, if a is .8, then in 20% of t h e i n s t a n c e s w h e r e the p a r t i e s have r e l a t i o n s h i p Z, t h e y do so d e s p i t e t h e p r e v a l e n t v i e w t h a t X o u g h t to h o l d f o r Z t o o b t a i n but d o e s no__t. We deal here, i n t e n t i o n a l l y , w i t h a small n u m b e r of rules and r e l a t i o n s h i p s ; nuances of t h e s e are n o t t r e a t e d , since t h e y are far t o o i d i o s y n c r a t i c f o r g e n e r a l a g r e e m e n t , e v e n b r o a d l y c o n s i d e r e d . The r e l a t i o n s h i p s c o n s i d e r e d are, in o r d e r of i n c r e a s i n g c o m m i t m e n t and initial intensity, c o m p a n i o n s h i p , f r i e n d s h i p , confidant(e)s, i n t i m a c y , love, m a r r i a g e , and f a m i l y . Before p r e s e n t i n g t h e rules t h e m s e l v e s , w h i c h are the u l t i m a t e t e s t of t h e m e r i t of the idea, s o m e p e r f a t o r y r e m a r k s on t h e s y s t e m as a w h o l e are in order. First, we have d e s i g n e d the s y s t e m with t w o p u r p o s e s in mind: ease of i m p l e m e n t a t i o n and t e s t i n g and n a t u r a l - ness of r e p r e s e n t a t i o n , w i t h the latter easily t h e m o r e i m - p o r t a n t d e s i d e r a t u m . Thus, e x c e p t f o r c o m p a n i o n s h i p , each level of i n t e r p e r s o n a l c l o s e n e s s i n c l u d e s and s u b - s u m e s t h e p r e v i o u s (and t h e r e f o r e all prior) level(s). Thus if X and Y have a f a m i l y t o g e t h e r , t h e y are m a r r i e d , t h e y l o v e each other, t h e y are intimates, t h e y are c o n f i d a n t s , and t h e y are f r i e n d s ( s u b j e c t t o t h e usual i d i o s y n c r a t i c e x c e p t i o n s ) . This s e e m s natural e n o u g h f o r t h e a t t a i n m e n t of r e l a t i o n s h i p s . H o w e v e r , r e l a t i o n s h i p s are d y n a m i c and w h e n t h e y d e g r a d e , as t h e y so often do, t h e y do n o t do so n i c e l y and m o n o t o n i c a l l y . So w h i l e it m a y be t r u e t h a t a p a r t f r o m t h e i d i o s y n c r a t i c e x c e p t i o n f r i e n d s h i p a l w a y s p r e c e d e s m a r r i a g e , after m a n y years it is c e r t a i n l y p o s s i b l e f o r X and Y to l o v e each o t h e r and miss each o t h e r w h e n apart; b u t n o t t o like each o t h e r or t a k e any p l e a s u r e in each other's c o m p a n y : An'~odd, but c o m m o n , s t a t e of a f - fairs. So t h e s y s t e m p r e s e n t e d here s h o u l d be see e i t h e r as a r e l a t i o n s h i p - a t t a i n m e n t model or as an ideal. Second, t h e s p e c i f i c a t t e n u a t i o n f a c t o r s s e l e c t e d w e r e n o t o b t a i n e d by e m p i r i c a l research. T h e y are s i m p l y l e f t - brain e s t i m a t e s of f r e q u e n c y or r i g h t - b r a i n e s t i m a t e s of p r o b a b i l i t y . T h e y d o seem to c o n f o r m to our ( l i m i t e d ) e x - p e r i e n c e and o u r i n t u i t i o n , b u t a m e t h o d o l o g i c a l c o g n i t i v e s c i e n t i s t m i g h t wish to use o t h e r n u m b e r s , m o r e r i g o r o u s l y o b t a i n e d . T a n t mieux. Since w e are, a f t e r all, d e a l i n g w i t h and p r e s e n t i n g a m e t h o d o l o g y f o r i d i o s y n c r a t i c d o m a i n s , t h e author's o w n p e r s o n - s p e c i f i c b i a s e s or i n s i g h t s m a y o c c a s i o n a l l y be visible. COMPANIONSHIP. Iff X and Y: like e a c h o t h e r 7 (.7) X and Y: w i l l i n g to share t i m e w i t h each o t h e r (1) X and Y: l o n e l y (.8) Then X and Y: c o m p a n i o n s (.5) In t h e f o r w a r d d i r e c t i o n , t h e r e are t h r e e a n t e c e d e n t c o n d i t i o n s w h i c h are r e p r e s e n t a t i v e of t h e t h r e e c l a s s e s of c o n d i t i o n s on w h i c h r e l a t i o n s h i p s m a y be c o n t i n g e n t . First, t h e r e is t h e " e n g i n e , " t h e e m o t i o n a l i m p e t u s f o r t h e r e l a t i o n s h i p , here loneliness. S e c o n d , t h e r e is t h e d e f i n i n g c o n d i t i o n , a c o n d i t i o n w h i c h h o l d s u n i v e r s a l l v w h e n t h e r e l a t i o n s h i p exists, here " w i l l i n g t o share free t i m e w i t h each o t h e r . " Third, t h e r e is a p r e c o n d i t i o n , here s i m p l e a f - f e c t i o n , w i t h o u t w h i c h t h e r e l a t i o n s h i p is b l o c k e d (5), To t h e c o m p u t e r s c i e n t i s t i n t e r e s t e d in t h e f o r m a l i s m i n t r o d u c e d and t h e f r a m e w o r k p r o v i d e d for i d i o s y n c r a t i c d o m a i n s , t h e r e is no d i s t i n c t i o n of i m p o r t a n c e b e t w e e n t h e s e t h r e e classes of c o n d i t i o n s . To the c o g n i t i v e s c i e n - tist, t h e s p e c i a l i s t i m p a r t i n g his k n o w l e d g e to t h e e × p e r t s y s t e m and the k n o w l e d g e e n g i n e e r a s s i s t i n g him, t h e user a p p l y i n g t h e e x p e r t s y s t e m , and t h e g e n e r a l reader, on t h e o t h e r hand, t h e s e d i s t i n c t i o n s are p r e g n a n t w=th m e a n i n g . In p r e s e n t i n g t h e rules, t h e r e f o r e , we shall c o n - t i n u e t o d r a w them. A rule m u s t have o n e or m o r e engines, f o r t h e r e is no r e l a t i o n s h i p w h i c h b e g i n s w i t h no e m o t i o n a l i m p e t u s of any sort; usually, but as we shall see n o t a l w a y s , t h e r e is j u s t one. When t h e a t t e n u a t i o n f a c t o r s of the e n g i n e c o n - d i t i o n s sum t o less t h a n one, as is t h e case here, w i t h c o m p a n i o n s h i p , it is an i n d i c a t i o n t h a t t h e e m o t i o n a l i m - p e t u s f o r t h e r e l a t i o n s h i p class is s o m e t i m e s i d i o s y n c r a t i c . A rule m a y or m a y n o t have a d e f i n i n g c o n d i t i o n , b u t if it has m o r e than one, it is an i n d i c a t i o n of s l o p p y d e s i g n - - in particular, f a i l u r e to g e n e r a l i z e . N o t e t h a t a d e f i n i n g c o n d i t i o n d o e s n o t g u a r a n t e e a r e l a t i o n s h i p (as here), a l t h o u g h the r e l a t i o n s h i p g u a r a n t e e s it. As w e shall see, h o w e v e r , m o s t a n t e c e d e n t c o n d i t i o n s are p r e c o n - ditions. The l o w a t t e n u a t i o n f a c t o r f o r c o m p a n i o n s h i p r e f l e c t s t h e f a c t that, unlike m o r e c o m m i t t e d r e l a t i o n s h i p s , it a r i s e s f r o m a g e n e r a l i z e d l o n e l i n e s s on t h e part of, t o use t h e s y s t e m ' s names, X and Y, and m a n y o t h e r r a n d o m or i d i o s y n c r a t i c f a c t o r s h e l p d e c i d e w h e t h e r t h e t w o w i l l b e - c o m e c o m p a n i o n s . As w e a s c e n d the h i e r a r c h y of r e l a t i o n s h i p s , r a n d o m i n f l u e n c e s w i l l b e c o m e i n c r e a s i n g l y n e g l i g i b l e and i d i o s y n c r a t i c f a c t o r s i n c r e a s i n g l y v i s i b l e . FRIENDSHIP. Iff X and Y: like each o t h e r 2 (.8) X and Y: r e s p e c t each other. (.7) X and Y: w a n t to share free t i m e w i t h each o t h e r (1) T h e n X and Y: f r i e n d s ( 9 ) For f r i e n d s h i p , w a n t i n g t o share free t i m e w i t h each o t h e r is b o t h t h e e n g i n e and t h e d e f i n i n g c o n d i t i o n . Note t h a t the p r e c o n d i t i o n s are still n e c e s s a r y t o e n s u r e that t h e m o t i v a t i o n f o r w a n t i n g each o t h e r ' s c o m p a n y is f r i e n d s h i p and n o t (say) business. T h o s e w h o a t t a i n the c o g n i t i v e - a f f e c t i v e state c o m m o n t o friends w i t h o u t liking or r e s p e c t i n g each o t h e r are, of course, a c c o u n t e d f o r by t h e a t t e n u a t i o n factors. For s o m e people, all i m p o r t a n t b u s i n e s s a s s o c i a t e s m a y be friends. M a n y w i l l find t h e p r e c o n d i t i o n s g i v e n here i n a d e - q u a t e ; t h e i r n o t i o n of f r i e n d s h i p is p r o b a b l y akin t o o u r n o t i o n of c o n f i d a n t s , w h i c h will be p r e s e n t e d n e x t and a m o u n t s t o classical, t r u e f r i e n d s h i p ( " f r i e n d s h i p w i t h a c a p i t a l 'F'"). As t h e w o r d is n o r m a l l y used t o d a y , h o w e v e r , " f r i e n d " m e a n s (at most) w h a t is g i v e n above. N o t i c e t h e d i f f e r e n c e b e t w e e n the d e f i n i n g c o n d i t i o n s f o r c o m p a n i o n s h i p and f r i e n d s h i p : " w i l l i n g " vs. " w a n t " to 1A rule elaborating this will be presented later. 2A rule elaboratin 9 this will be presented later. SIGART N e w s l e t t e r , J u l y 1988, N u m b e r 105 Page 31 s h a r e f r e e t i m e w i t h e a c h o t h e r . A p e r s o n - s p e c i f i c n e e d has r e p l a c e d a g e n e r a l i z e d l o n e l i n e s s . CONFIDANTS. Iff X and Y: f r i e n d s (.85) X and Y: r e s p e c t e a c h o t h e r h i g h l y (.7) X and Y: w i l l i n g / w a n t t o s h a r e t h e i r t h o u g h t s and f e e l i n g s w i t h e a c h o t h e r (1) X and Y: m i s s e a c h o t h e r w h e n a p a r t (.5) X and Y: t r u s t e a c h o t h e r 3 (.8) T h e n X and Y: c o n f i d a n t ( e ) s (.7) O n c e a g a i n , t h e e n g i n e and t h e d e f i n i n g c o n d i t i o n are t h e s a m e : s h a r i n g t h o u g h t s and f e e l i n g s . This is f r i e n d s h i p o n a m o r e i n t i m a t e l e v e l (it s u b s u m e s f r i e n d s h i p w i t h i d i o s y n c r a t i c e x c e p t i o n s ) t h a n s h a r i n g t i m e a l o n e can e v e n reach. It is a l s o t h e l o w e s t l e v e l o f t h e h i e r a r c h y of t h e r e l a t i o n s h i p c l a s s e s w e c o n s i d e r at w h i c h m i s s i n g e a c h o t h e r , a c o g n i t i v e - a f f e c t i v e s t a t e c e n t r a l t o s t r o n g r e l a t i o n s h i p s , c o m e s i n t o play. For s o m e , it is a p r e c o n - d i t i o n t o b e i n g a c o n f i d a n t (or a v e r y q u i c k r e s u l t ) ; f o r s o m e , it m a y a l s o be an e n g i n e c o n d i t i o n . O n e m i g h t s u p p o s e t h a t " t r u s t i n g e a c h o t h e r " and b e i n g " w i l l i n g t o s h a r e t h e i r t h o u g h t s and f e e l i n g s w i t h e a c h o t h e r " w o u l d be c o i n c i d e n t . E x p e r i e n c e t e a c h e s t h a t t h i s is n o t a l w a y s so, as t h e n e e d t o c o n f i d e , in m a n y so o f t e n o v e r p o w e r s d o u b t s a b o u t t h e o t h e r ' s t r u s t w o r t h i n e s s . It is a l s o a r e f l e c t i o n o f t h e m e c h a n i s m o f t r u s t b u i l d u p , w h i c h is n o t u n l i k e t h e b u i l d u p o f b a n d c r e d i t . O n e t r u s t s a l i t t l e ; if it w o r k s out, o n e t r u s t s a l i t t l e m o r e . Of c o u r s e , h o w e v e r n a t u r a l t h i s s y s t e m m a y be and h o w e v e r p r e v a l e n t , it l e a v e s t h e c o n f i d a n t ( o r bank) o p e n t o d i s a s t e r in u n f o r - t u n a t e i n s t a n c e s . P e r h a p s t h a t ' s w h y a n o t h e r m e t h o d - - f o r m a l a n d i n f o r m a l r e f e r e n c e s - - h a s e v o l v e d f o r t h e r e s o l u t i o n o f t h e t r u s t i s s u e (by p o t e n t i a l c r e d i t o r s and p o t e n t i a l f r i e n d s alike). INTIMATES. Iff X and Y: c o n f i d a n t s ( e ) s (.95) X a n d Y: k n o w e a c h o t h e r f o r s o m e t i m e (.5) X a n d Y: e m o t i o n a l l y c o m p a t i b l e e v e n in c l o s e q u a r t e r s (.5) X a n d Y: find e a c h o t h e r a t t r a c t i v e (.8) T h e n X and Y: i n t i m a t e s (.5) I n t i m a c y , as w e are c o n s i d e r i n g it, is a s e r i o u s r e l a t i o n s h i p o n e s t e p shy o f r o m a n t i c l o v e , n o t m e r e p h y s i c a l i n t i m a c y w i t h o u t a v e r y p o w e r f u l e m o t i o n a l c o m - p o n e n t . O b v i o u s l y , m e r e p h y s i c a l i n t i m a c y is p o s s i b l e w h e n X and Y are s t r a n g e r s , h a r d l y i n t i m a t e s as w e are u s i n g t h e t e r m . T h e i n t i m a c y t o w h i c h w e r e f e r is a s t a t e o f h e a r t and m i n d , a c o g n i t i v e - a f f e c t i v e s t a t e , j o i n e d t o a s t a t e o f a f f a i r s (in t h e p h i l o s o p h e r ' s s e n s e o f t h e p h r a s e ) w h i c h can be r e a c h e d w i t h o u t a c o n f i d a n t - t y p e r e l a t i o n - ship i d i o s y n c r a t i c a l l y o n l y o n e t i m e in t w e n t y . For the f o r w a r d - c h a i n i n g a n t e c e d e n t "X and Y: p h y s i c a l l y i n t i m a t e " t h e c o n s e q u e n t s w o u l d h a v e v e r y d i f f e r e n t a t t e n u a t i o n f a c t o r s , say, .05, .25, .02, .6, r e s p e c t i v e l y . A n o t h e r i s s u e t h a t m u s t e v e n t u a l l y be a d d r e s s e d h e r e is t h a t o f s a m e - s e x r e l a t i o n s h i p s . For h e t e r o s e x u a l s , t h e f o u r t h a n t e c e d e n t c o n d i t i o n r u l e s o u t s a m e - s e x r e l a t i o n - ships. For o t h e r s , it d o e s not, n o r d o e s a n y t h i n g e l s e in t h e rule. S t u d i e s s h o w t h a t o n e - o n e r e l a t i o n s h i p s are m u c h m o r e c o m m o n a m o n g h e t e r o s e x u a l s , b u t it is n o t t o be s u p p o s e d t h a t t h i s c r e a t e s a s e r i o u s o v e r s t a t e m e n t o f t h e c h a n c e s o f an i n t i m a t e r e l a t i o n s h i p d e v e l o p i n g (50%) g i v e n all t h e s t a t e d a n t e c e d e n t c o n d i t i o n s . H a v i n g d e a l t w i t h t h e s e t w o s i d e i s s u e s , w e r e t u r n t o an e x p o s i t i o n o f t h e rule. The e n g i n e h e r e is m u t u a l a t - t r a c t i o n and as a n y o n e w i l l a d m i t u n d e r p e r s i s t e n t q u e s - t i o n i n g it is n o t a s u r e t h i n g ; t e c h n i c a l l y , t h a t is, it d o e s n o t g i v e "X and Y: i n t i m a t e s " an a t t e n u a t i o n f a c t o r o f 1. W h a t w i l l m a k e t w o p e o p l e d e c i d e t o u p g r a d e a c o n f i d a n t r e l a t i o n s h i p t o an i n t i m a t e r e l a t i o n s h i p is h i g h l y i d i o s y n c r a t i c , so m u c h so t h a t f u l l y h a l f t h e t i m e t h a t t h e a n t e c e d e n t s are t r u e , i n t i m a c y w i l l n e v e r t h e l e s s n o t r e s u l t . If it s e e m s t o t h e r e a d e r t h a t t h i s is s u r e l y a g r o s s e x - a g g e r a t i o n , it is l i k e l y t h a t he is u n w i t t i n g l y u s i n g i n t i m a c y in a w e a k e r s e n s e b y n o t b e i n g s e n s i b l e o f t h e f u l l r a n g e o f a n t e c e d e n t c o n d i t i o n s f o r a c o n f i d a n t - t y p e r e l a t i o n s h i p , w h i c h r e l a t i o n s h i p is, in all b u t a h a n d f u l o f i d i o s y n c r a t i c cases, p r e r e q u i s i t e t o i n t i m a c y as h e r e d e f i n e d . (And t h e r e a r e of c o u r s e t h e r e q u i r e m e n t s o f t h e e n g i n e c o n d i t i o n and t h e t w o p r e c o n d i t i o n s . ) R O M A N T I C LOVE, Iff X and Y: i n t i m a t e s (.95) X and Y: w i l l i n g / w a n t m u t u a l e m o t i o n a l c o m m i t m e n t ( . 7 5 ) X and Y: w i l l i n g / w a n t a b s o r p t i o n in e a c h o t h e r ' s life(.7) X and Y: m i s s e a c h o t h e r q u i t e i n t e n s e l y w h e n a p a r t (1) T h e n X and Y: l o v e e a c h o t h e r ( r o m a n t i c a l l y ) (.9) The e n g i n e s h e r e are t h e m u t u a l d e s i r e f o r e m o t i o n a l c o m m i t m e n t , o n e - o n e l o y a l t y , a n d t h e m u t u a l d e s i r e f o r t o t a l i n v o l v e m e n t in e a c h o t h e r ' s life ( i n v o l v e m e n t d o e s n o t m e a n d o m i n a t i o n , u n w a n t e d i n t e r f e r e n c e , etc.) and t h e d e f i n i n g c o n d i t i o n f o r t h i s c o g n i t i v e - a f f e c t i v e s t a t e is g i v e n b y t h e f i n a l a n t e c e d e n t c o n d i t i o n . It w i l l be i s t r u c - t i r e t o c o m p a r e l o v i n g and l i k i n g ( t h e l a t t e r is r e q u i r e d f o r c o m p a i o n s h i p and f r i e n d s h i p (and, as e x p l a i n e d e a r l y in t h i s paper, b y i n c l u s i o n in all m o r e i n t e n s e r e l a t i o n s h i p s ) ) : Iff X and Y: e n j o y e a c h o t h e r ' s c o m p a n y (.95) X and Y: e m o t i o n a l l y c o m p a t i b l e (.8) T h e n X and Y: like e a c h o t h e r (1) T h u s t h e e m p h a s i s in l o v i n g is on a l o n g i n g w h e n a p a r t , w h i l e t h e e m p h a s i s in l i k i n g is o n an a g r e e a b l e n e s s w h e n t o g e t h e r . Of c o u r s e , e x c e p t f o r i d i o s y n c r a t i c c a s e s , l o v e i n c l u d e s i n t i m a c y w h i c h i n c l u d e s . . . . f r i e n d s h i p a n d , t h e r e f o r e , l i k i n g . N e v e r t h e l e s s , t h i s n e g a t i v e e m p h a s i s h e l p s e x p l a i n w h y l o v e w i t h o u t a f f e c t i o n as m a y o c c u r w h e n a r e l a t i o n s h i p d e g r a d e s can b e a c l e a r l y d e s t r u c t i v e f o r c e and w h y l e c t u r e s e n t i t l e d " H o w t o Fall o u t o f L o v e " d r a w l a r g e a u d i e n c e s . MARIAGE. Iff X and Y: l o v e e a c h o t h e r (.65) X and Y: c u l t u r a l l y c o m p a t i b l e (.7) X and Y: k n o w e a c h o t h e r f o r a c o n s i d e r a b l e t i m e (.75) X and Y: w i l l i n g / w a n t m u t u a l i n s t r u m e n t a l c o m m i t m e n t X and Y: f i n d e a c h o t h e r ' s i d i o s y n c r a t i c b e h a v i o r s t o l e r a b l e t o p l e a s i n g (.5) X and Y: l o n e l y ( . 7 ) / X a n d Y: w a n t kids (.9) T h e n X and Y: m a r r i e d ( . 7 ) / X and Y: m a r r i e d (.75) 3A rule elaboratm 9 this will be presented later. SIGART N e w s l e t t e r , July 1988, N u m b e r 105 Page 32 M a r r i a g e c o n d i t i o n s are f a i r l y i d i o s y n c r a t i c ; h e n c e o n l y 7 0 - 7 5 % of t h e t i m e t h a t t h e a n t e c e d e n t c o n d i t i o n s are s a t i s f i e d is t h e c o n s e q u e n t r e a l i z e d . S i n c e n o c o n d i t i o n as t o g e n d e r is i m p o s e d , if a s a m e - s e x r e l a t i o n s h i p m e e t s all t h e a n t e c e d e n t c o n d i t i o n s , it is a m o n g t h e 3 0 % ( f o r g e t a b o u t w a n t i n g kids as an a n t e c e d e n t ! ) o f s u c h c a s e s w h e r e t h e r u l e is i d i o s y n c r a t i c a l l y n o t t r i g g e r e d . A l s o l o w a r e t h e a t t e n u a t i o n f a c t o r s in t h e o t h e r d i r e c t i o n . This m a y i n d i c a t e w h y so m a n y m a r r i a g e s s o u r . In o t h e r w o r d s , if t h e c o n s e q u e n t w e r e "X and Y: h a p p i l y m a r r i e d " t h e a t t e n u a t i o n f a c t o r s w o u l d be f a r h i g h e r , p e r h a p s .98, .95, .8, .7, .9, .7/.9 r e s p e c t i v e l y . T h e r e are t w o o r p o s s i b l y t h r e e e n g i n e s h e r e w h i c h o v e r l a p ; w a n t i n g i n s t r u m e n t a l c o m m i t m e n t , w a n t i n g kids, a n d b e i n g l o n e l y . As t h e a t t e n u a t i o n f a c t o r s f o r m a r r i a g e s h o w , w a n t i n g kids is a s t r o n g e r s p u r t o m a r r i a g e t h a n l o n e l i n e s s , w h i c h can m o r e e a s i l y be s a t i s f i e d b y m e a n s s h o r t of m a r r i a g e . As i m p l e m e n t e d , t h e r u l e s f o r m a r r i a g e in t h e s y s t e m n u m b e r t h r e e . One w i t h "X and Y: w a n t kids (.9)" l e a d i n g t o "X and Y: " m a r r i e d (.75)"; o n e w i t h "X and Y: l o n e l y (.7)" l e a d i n g t o "X and Y: m a r r i e d (.7)"; a n d o n e w i t h b o t h "X and Y: l o n e l y (.7)" and "X a n d Y: w a n t kids (.9)" l e a d i n g t o "X and Y: m a r r i e d (.75)." T h e r e is no r e d u n d a n c y in t h e p r e s e n c e o f t h e t h i r d r u l e , s i n c e all r u l e s are b i d i r e c t i o n a l . F u r t h e r m o r e , t h a t t w o r u l e s say t h a t 75% o f c o u p l e s w i l l e n d up m a r r i e d d o e s n o t g u a r a n t e e t h a t t h e y w i l l be t h e s a m e 75%, so f r o m an i n - t e l l e c t u a l p o i n t of v i e w , it is w e l l t o i n c l u d e it, e v e n if it m a k e s no d i f f e r e n t p r o b a b i l i s t i c p r e d i c t i o n . T h e r e are t h r e e d e g r e e s o f e m o t i o n a l c o m p a t i b i l i t y r e q u i r e d by t h e s y s t e m : f i n d i n g e a c h o t h e r ' s i d i o s y n c r a t i c b e h a v i o r s t o l e r a b l e t o p l e a s i n g - - f o r m a r r i a g e , e m o t i o n a l c o m p a t i b i l i t y e v e n in c l o s e q u a r t e r s - - f o r i n t i m a c y , and (basic) e m o t i o n a l c o m p a t i b i l i t y - - f o r c o m p a n i o n s h i p (via liking). The l e v e l s o f c o m p a t i b i l i t y r e q u i r e d m a y v a r y on a p e r s o n - s p e c i f i c basis and that, o f c o u r s e , is d e a l t w i t h b y t h e a t t e n u a t i o n f a c t o r s . A w o r d a b o u t l o n e l i n e s s , a r e q u i r e m e n t f o r c o m - p a n i o n s h i p and m a r r i a g e - - b u t n o t h i n g in b e t w e e n , is in o r d e r . The s y s t e m r e f l e c t s t h e i d e a t h a t e i t h e r c o m - p a n i o n s h i p w i l l s a t i s f y l o n e l i n e s s o r if it d o e s n o t , n e i t h e r w i l l f r i e n d s h i p , a c o n f i d a n t , o r e v e n i n t i m a c y and l o v e . T h e s e i n - b e t w e e n r e l a t i o n s h i p s s a t i s f y p e r s o n - s p e c i f i c and t y p e - s p e c i f i c (to c o n f i d e , t o be i n t i m a t e , etc.) n e e d s , r a t h e r t h a n l o n e l i n e s s in g e n e r a l . M a r r i a g e , o n t h e o t h e r h a n d , a d d s t o l o v e ( w i t h its s a t i s f a c t i o n o f a w i d e v a r i e t y o f p e r s o n - s p e c i f i c a n d t y p e - s p e c i f i c n e e d s ) c o n s t a n t c o m - p a n i o n s h i p . FAMILY Iff X and Y: m a r r i e d (1) X and Y: w a n t kids (.9) X a n d Y: s u f f i c i e n t m e a n s and s p a c e (.5) T h e n × and Y: h a v e a f a m i l y t o g e t h e r (.95) T h e e n g i n e h e r e , o b v i o u s l y , is w a n t i n g kids. V i r t u a l l y all m a r r i e d c o u p l e s w h o w a n t kids d o h a v e o r a d o p t t h e m . The o n l y s o m e w h a t g e n e r a l c o n s t r a i n t s e e m s t o be s u f - f i c i e n c y o f m e a n s a n d s p a c e and, as t h e a t t e n u a t i o n f a c t o r s h o w s (50% o f f a m i l i e s d o n ' t h a v e s u f f i c i e n t m e a n s and space), it is n o t t e r r i b l y c o n s t r a i n i n g . It is i n s t r u c t i v e t o c o m p a r e t h e f i n a l t h r e e r u l e s o f t h e h i e r a r c h y : f a m i l y g i v e n m a r r i a g e , m a r r i a g e g i v e n l o v e , l o v e g i v e n i n t i m a c y . T h e c o m p a r i s o n s u g g e s t s t h a t t h e c o n - d i t i o n s f o r m a r r i a g e are f a r m o r e i d i o s y n c r a t i c t h a n a r e t h o s e f o r a f a m i l y g i v e n m a r r i a g e a n d l o v e g i v e n i n t i m a c y . P e o p l e w h o a r e m a r r i e d d o a l m o s t a l w a y s h a v e f a m i l i e s ; p e o p l e w h o a r e i n t i m a t e and m e e t t h e r e m a i n i n g c o n - d i t i o n s d o fall in l o v e . But e v e n if l o v e is t r i e d a n d t r u e , m a r r i a g e is o f t e n n o t t h e result. (By o f t e n , w e m e a n , o f c o u r s e , 2 5 - 3 0 % o f t h e t i m e . ) The n o n - i d i o s y n c r a t i c c o n - d i t i o n s f o r m a r r i a g e are a l s o a g o o d d e a l m o r e d e m a n d i n g t h a n t h o s e f o r l o v e g i v e n i n t i m a c y a n d f a m i l y g i v e n m a r - riage. Thus w i t h s u c h a m o d e l o f r e l a t i o n s h i p a t t a i n m e n t , it is n o s u r p r i s e t h a t m a r r i a g e s are h a r d t o m a i n t a i n . T h a t c o n c l u s i o n f o l l o w s f r o m t h e s i n g l e p r e m i s e t h a t t h e h a r d e r a r e l a t i o n s h i p is t o f o r m , t h e e a s i e r it is t o d a m a g e . W h i l e w e c o u l d see an i n t e l l e c t u a l c a s e e i t h e r w a y o n t h e p r e m i s e , e m p i r i c a l e v i d e n c e s u p p o r t s b o t h t h e p r e m i s e and t h e c o n c l u s i o n . T h i s is o n l y o n e o f m a n y d o m a i n - s p e c i f i c c o m m e n t s m a d e in t h i s w o r k . It is n a t u r a l w h e n d e s i g n i n g a s y s t e m t o be an e x p e r t and d i s c u s s i n g o n e ' s i d e a s w i t h c o l l e a g u e s and a s s o c i a t e s t o l e a r n a c o n s i d e r a b l e a m o u n t a b o u t t h e d o m a i n as w e l l as a b o u t e x p e r t s y s t e m d e s i g n f o r i d i o s y n c r a t i c d o m a i n s in g e n e r a l . R e t u r n i n g t o t h e l a t t e r , an a t t e m p t w a s m a d e t h r o u g h o u t t h e s y s t e m t o r e s t r i c t c o n s i d e r a t i o n t o c o n - d i t i o n s w i t h a t t e n u a t i o n f a c t o r s t h a t w o u l d m a k e at l e a s t .5. Thus, in t h e c a s e of t h e f a m i l i a l r e l a t i o n s h i p , t h e 5% o f c o u p l e s w h o d o n ' t h a v e o r a d o p t c h i l d r e n d e s p i t e t h a t desire, t h e i r b e i n g m a r r i e d t o e a c h o t h e r , and t h e i r h a v i n g s u f f i c i e n t m e a n s a n d s p a c e h a v e " r e a s o n s o f t h e i r o w n " - - i d i o s y n c r a t i c o r p e r s o n - s p e c i f i c f a c t o r s w h i c h i n f l u e n c e t h e r e l a t i o n s h i p s t h e y f o r m . L i k e w i s e , t h e 10% o f c o u p l e s w h o h a v e a f a m i l y b u t d o n o t w a n t c h i l d r e n m u s t a l s o h a v e " r e a s o n s o f t h e i r o w n " - - p e r h a p s r e l i g i o u s b e l i e f s o r c o n c e r n w i t h w h a t t h e n e i g h b o r s t h i n k , etc. It is i m p o r t a n t in an e x p e r t s y s t e m d e s i g n e d f o r an i d i o s y n c r a t i c d o m a i n n o t t o t r y t o d i s c o v e r e v e r y i d i o s y n c r a s y and a s s i g n it an a p p r o p r i a t e a t t e n u a t i o n f a c t o r (less t h a n .5), o r t h e s y s t e m w i l l r a p i d l y be t r a n s f o r m e d f r o m o n e w h i c h h a n d l e s i d i o s y n c r a s i e s in its d o m a i n t o o n e w h i c h m e r e l y r e f l e c t s t h e i d i o s y n c r a s i e s o f its a u t h o r . The l a t t e r is t o s o m e e x - t e n t i m p o s s i b l e t o a v o i d in a n y d o m a i n w h e r e t h e k e y c o n c e p t s are n o t o b j e c t i v e l y f i x e d ; w e h o p e w e h a v e d o n e a c r e d i b l e job. REFERENCES 1. W i n s t o n , P.H., A r t i f i c i a l I n t e l l i g e n c e , 1984, R e a d i n g , MA; A d d i s o n W e s l e y . 2. S h o r t l i f f e , E. H. and B u c h a n a n , B.G., "A M o d e l o f I n - e x a c t R e a s o n i n g in M e d i c i n e , " M a t h e m a t i c a l Bio__ s c i e n c e s 23(1975): 3 5 1 - 3 7 9 . 3. D u d a , R.O., et al., " D e v e l o p m e n t o f t h e P r o s p e c t o r C o n s u l t a t i o n S y s t e m f o r M i n e r a l E x p l o r a t i o n , " S e p - t e m b e r 1978, M e n l o Park, CA: SRI. 4. Duda, R.O, e t al., "A C o m p u t e r - B a s e d C o n s u l t a n t f o r M i n e r a l E x p l o r a t i o n , " S e p t e m b e r 1979, M e n l o Park, CA: SRI. 5. V e r e , S.A., " R e l a t i o n a l P r o d u c t i o n S y s t e m s , " A r t i f i c i a l I n t e l l i g e n c e 8(1977): 4 7 - 6 8 . APPENDIX A. DEFAULT RULE. Iff t r u e (1) T h e n X a n d Y: n o t h i n g (.8) T h e s y s t e m ' s u n i v e r s e o f d i s c o u r s e is all c o u p l e s , X and Y, w h o k n o w e a c h o t h e r , h o w e v e r s l i g h t l y . T h i s r u l e s t a t e s t h a t 8 0 % o f s u c h c o u p l e s a r e n e i t h e r c o m p a n i o n s n o r f r i e n d s ( n o r a n y t h i n g c l o s e r ) ; it s e r v e s as a d e f a u l t . B. ADOPTION RULES (UNDIRECTIONAL). 1. If X and Y: m a r r i e d Y a n d Z: f r i e n d s ( o r m o r e ) T h e n X and Z: c o m p a n i o n s (.6) SIGART Newsletter, J u l y 1988, N u m b e r 105 Page 33 2. If X and Y: married Y and Z: c o n f i d a n t s ( m o r e w o u l d p r o b a b l y lead t o d i v o r c e ) Then X and Z: friends (.4) 3. If X and Y: married Y and Z: c o n f i d a n t s Then X and Z: c o m p a n i o n s (.7) There is little p o i n t in w r i t i n g rules f o r × and Y: less than married, since the a t t e n u a t i o n f a c t o r s w o u l d be v e r y l o w ( v e r y f e w such "adoptions"). it is p o s s i b l e t h a t for X and Y: have a f a m i l y t o g e t h e r , the a t t e n u a t i o n factors w o u l d be lower, since p e o p l e with f a m i l i e s have less time. Finally, s u b j e c t to i d i o s y n c r a t i c e x c e p t i o n s , a d o p t i o n s never result in m o r e than a friend or, if t h e y do, t h e y cease b e - ing an a d o p t i o n and m e e t the standard a n t e c e d e n t c o n - ditions. N o t e t h a t the a n t e c e d e n t s of (2) and (3) o v e r l a p as do t h o s e of (1) and (2)/(3) t o s o m e e x t e n t (because, w i t h i d i o s y n c r a t i c e x c e p t i o n s , c o n f i d a n t s are friends). One m i g h t t r y t o r i g o r o u s l y use a v a i l a b l e m e t h o d s (see refs. 2,3,4) to o b t a i n b e t t e r n u m b e r s . But t h e a b o v e w i l l do. C. TRUST RULE. IFF X and Y: believe t h e y can repose t h e i r t h o u g h t s and f e e l i n g s in each o t h e r w i t h o u t facing ridicule, r e j e c t i o n , etc. (.85) X and Y: believe ~hey can repose t h e i r t h o u g h t s and f e e l i n g s in each o t h e r w i t h o u t facing e x p o s u r e l e a d i n g to m t e r p e r s o n a l a n d / o r i n s t r u m e n t a l harm (.85) X and Y: believe t h e y can repose t h e i r t h o u g h t s and f e e l i n g s in each o t h e r w i t h o u t fearing t h a t t h e i r c o n f i d e n c e s will e v e n t u a l l y be used a g a i n s t t h e m e m o t i o n a l l y i.e., t h r o w n up at t h e m , etc. (.7) X and Y: trust each o t h e r (1'.) The difference in a t t e n u a t i o n f a c t o r s r e f l e c t s t h e fact t h a t n e a r l y e v e r y o n e w o u l d s o o n e r risk l o n g - t e r m e m o - t i o n a l hurt ( c o n d i t i o n 3) t h a n s h o r t - t e r m e m o t i o n a l hurt ( c o n d i t i o n 1) or e x p o s u r e ( c o n d i t i o n 2). SELECTED AI-RELATED DISSERTATIONS Assembled by: Susanne M. Humphrey National Library of Medicine Bethesda, M D 2 0 8 9 4 and Bob Krovetz University of Massachusetts Amherst, MA 01002 The f o l l o w i n g are c i t a t i o n s s e l e c t e d by t i t l e and a b s t r a c t as being related t o AI, resulting f r o m a c o m p u t e r search, using the BRS I n f o r m a t i o n T e c h n o l o g i e s r e t r i e v a l service, of t h e Dissertation Abstracts I n t e r n a t i o n a l (DAI) d a t a b a s e produced by U n i v e r s i t y M i c r o f i l m s I n t e r n a t i o n a l . The online file includes abstracts, w h i c h are n o t published in this listing, but t h e c i t a t i o n s b e l o w do include the DAI reference f o r f i n d i n g t h e a b s t r a c t in t h e published DAI. Other e l e m e n t s of the c i t a t i o n are a u t h o r ; u n i v e r s i t y , degree, and, if available, n u m b e r of pages; title; UM o r d e r n u m b e r and y e a r - m o n t h of DAI; and DAI s u b j e c t c a t e g o r y chosen by the a u t h o r of t h e dissertation. References are s o r t e d first by the initial DAI subject c a t e g o r v and second by the author. In a d d i t i o n t h e d a t a b a s e i n c l u d e s m a s t e r s abstracts, d e n o t e d by MAI (Masters A b s t r a c t s I n t e r n a t i o n a l ) instead of DAI as t h e r e f e r e n c e to the a b s t r a c t in t h e published MAI. Unless o t h e r w i s e specified, paper or m i c r o f o r m copies of d i s s e r t a t i o n s m a y be o r d e r e d from U n i v e r s i t y M i c r o f i l m s International, Dissertation Copies, Post Office Box 1764, Ann Arbor, MI 48106; t e l e p h o n e for U.S. ( e x c e p t Michigan, Hawaii, Alaska): 1 - 8 0 0 - 5 2 1 - 3 0 4 2 , f o r Canada: 1 - 8 0 0 - 2 6 8 - 6 0 9 0 . Price lists and o t h e r o r d e r i n g and s h i p - ping i n f o r m a t i o n are in the i n t r o d u c t i o n t o t h e p u b l i s h e d DAI. DIAZ, JULIAN, I I I . Georgia State U n i v e r s i t y - C o l l e g e of Business A d m i n i s t r a t i o n Ph.D 1987, 310 pages. PROCESS TRACING INVESTIGATION INTO PROBLEM- SOLVING WITHIN RESIDENTIAL REAL ESTATE A P - PRAISAL. DAI V48(09), SecA, pp2378. U n i v e r s i t y M i c r o f i l m s Order N u m b e r ADG87-26477. Business A d - m i n i s t r a t i o n , General. MATHIESON, KIERAN DONALDI. Indiana U n i v e r s i t y Ph.D 1987, 159 pages. A MACHINE LEARNING APPROACH TO FLEXIBLE MANUFACTURING SYSTEM SCHEDULING. DAI V48(09), SecA, pp2378. U n i v e r s i t y M i c r o f i l m s Order N u m b e r ADG87-27515. Business A d m i n i s t r a t i o n , General. BOCK, DOUGLAS BRIAN. Indiana U n i v e r s i t y Ph.D 1987, 201 pages. SETTING DUE DATES FOR COMPUTER BASED SYSTEM DEVELOPMENT PROJECTS. DAI V48(09), SecA, pp2384. U n i v e r s i t y M i c r o f i l m s O r d e r N u m b e r ADG87-27484. Business A d m i n i s t r a t i o n , M a n a g e m e n t ABRAMS, DAVID HOWARD. Georgia I n s t i t u t e of T e c h n o l o g y D.B. 1987, 220 pages. A REFERENCE MODEL FOR HETEROGENEOUS DATA BASE MANIPULATION AND AN EXPERT SYSTEM PROTOTYPE FOR HOMOGENEOUS DATA BASE MANIPULATION. DAI V48(09), SecB. U n i v e r s i t y M i c r o f i l m s Order N u m b e r A D G 8 7 - 2 7 1 8 4 . C o m p u t e r Science. S I G A R T N e w s l e t t e r , J u l y 1988, N u m b e r 105 P a g e 34 
work_qazchnt2ebbbzg5332xif7e2t4	----	Inducing safer oblique trees without costs Vadera, S http://dx.doi.org/10.1111/j.1468­0394.2005.00311.x Title Inducing safer oblique trees without costs Authors Vadera, S Type Article URL This version is available at: http://usir.salford.ac.uk/944/ Published Date 2005 USIR is a digital collection of the research output of the University of Salford. Where copyright  permits, full text material held in the repository is made freely available online and can be read,  downloaded and copied for non­commercial private study or research purposes. Please check the  manuscript for any further copyright restrictions. For more information, including our policy and submission procedure, please contact the Repository Team at: usir@salford.ac.uk. mailto:usir@salford.ac.uk Inducing Safer Oblique Trees without Costs Sunil Vadera S.Vadera@salford.ac.uk School of Computing, Science and Engineering, University of Salford, Salford M5 4WT, UK Abstract Decision tree induction has been has been widely studied and applied. In safety appli- cations, such as determining whether a chemical process is safe, or whether a person has a medical condition, the cost of misclassification in one of the classes is significantly higher than the other class. Several authors have tackled this problem by developing cost-sensitive decision tree learning algorithms or suggested ways of changing the distribution of training examples to bias the decision tree learning process so as to take account of costs. A pre-requisite for applying such algorithms is the availability of costs of misclassifica- tion. Although this may be possible for some applications, in the area of safety, obtaining reasonable estimates of costs of misclassification is not easy. This paper presents a new algorithm for applications where the cost of misclassifications can not be quantified, though the cost of misclassification in one class is known to be significantly higher than another class. The algorithm utilises linear discriminant analysis to identify oblique relationships between continuous attributes and then carries out an appropriate modification to ensure that the resulting tree errs on the side of safety. The algorithm is evaluated with respect to ICET, one of the best known cost-sensitive algorithms, OC1 a well known oblique decision tree algorithm and an algorithm that utilises robust linear programming. 1 Introduction Research on decision tree learning is one of the major successes of AI, with many commercial products, such as case based reasoning systems (Kolodner, 1993) and data mining tools utilising decision tree learning algorithms (Berry & Linoff, 2004). Early research in this area developed algorithms that aimed to maximize overall accuracy. More recent research is being influenced by recognising that human decision making is not focused solely on accuracy, but also takes account of the potential implications of a decision. For example, a chemical engineer considers the risks of explosion when considering the safety of a process plant, a bank manager carefully considers the implications of a customer defaulting on a loan and a medical consultant does not ignore the potential consequences of misdiagnosing a patient. This motivation has led to research in the area of decision tree learning progressing from considering accuracy alone, to the development of algorithms that minimize costs of misclassification (e.g., Breiman, Friedman, Olsen, & Stone, 1984; Turney, 1995). Research on this problem can be categorised into two broad approaches: algorithms that aim to reduce a cost-sensitive problem to an equivalent accuracy maximization problem, and those that extend accuracy based algorithms to handle costs of misclassification. 1 1There are, of course some approaches that don’t use decision trees at all, but these are outside the scope of this paper. 1 The first category of systems includes the proposal by Breiman et al. (1984), who sug- gested changing the distribution of the training examples to be equivalent to the ratio of costs of misclassification. More recently, Zadrozny, Langford, and Abe (2003) formalize this idea, called stratification, and show how the distribution of examples can be changed to implicitly reflect the costs and thereby enable the application of accuracy based classi- fiers. However, Elkan (2001) shows that an impurity measure, with a similar basis to those adopted in many of the decision tree learning algorithms (such as CART (Breiman et al., 1984) and C4.5 (Quinlan, 1993)) is insensitive to stratification, therefore casting doubt on the effectiveness of using such systems after changing the distribution of examples. The second category of systems is based on the observation that decision tree learning algorithms that aim to maximize accuracy construct trees in a greedy, top-down manner and utilize an information-theoretic measure to select decision nodes. Hence, a natural approach to cost-sensitive decision tree construction is to alter the information-theoretic selection measure so that it includes costs. A number of systems take this approach, including the EG2 (Núñez, 1991), CS-ID3 (Tan, 1993) and ICET (Turney, 1995). The studies by Pazzani, Merz, Murphy, Ali, Hurne, and Brunk (1994) and Turney (1995) include empirical evaluations of these types of algorithms. Pazzani et al. (1994) conclude that these algorithms are no more effective than algorithms that ignore costs. Turney (1995) develops ICET, an algorithm based on using Genetic Algorithms (GAs) to evolve cost-sensitive trees that aim to take account of both costs of misclassification as well as costs of tests. Turney evaluates ICET with respect to several algorithms and shows that ICET performs better than the other algorithms, though he also notes the difficulty ICET has for domains where costs of misclassification are not symmetric: “it is easier to avoid false positive diagnosis (a patient is diagnosed as being sick, but is actually healthy) than it is to avoid false negative diagnosis (a patient is diagnosed as healthy, but is actually sick). This is unfortunate, since false negative diagnosis usually carry a heavier penalty, in real life.” 2 Thus, the problem of developing an effective cost-sensitive decision tree algorithm when given costs of misclassification remains a difficult and open problem. Even if there is an effective algorithm for minimising costs of misclassifications and costs are available, there is a further potential problem which we now illustrate. Consider the application of a de- cision tree induction algorithm on the Diabetes data set, which has been widely used for benchmarking. The data set has 768 examples of which 65% are in the safe class (i.e., no diabetes) and 35% are in the unsafe class (i.e., has diabetes). Suppose we use 50% of this data for training and 50% for testing, then decision tree algorithms would typically produce the results given in table 1.3 If, say, the cost of misclassifying unsafe cases was four times that of misclassifying safe cases, then the expected cost and accuracy are: Expected Cost = 50 ∗ 1 + 59 ∗ 4 384 = 0.75 2Turney refers to a simple cost matrix as one where the costs of misclassifications are equal, and a complex cost matrix as one where the costs are different. 3These particular results were obtained with 10 random trials using an axis-parallel algorithm similar to C4.5 and the literature contains similar results. 2 Class Algorithm Algorithm returns Safe returns Unsafe Actually Safe 200 50 Actually Unsafe 59 75 Table 1: Typical results produced for the diabetes data set Accuracy = 275 384 = 71.6% However, a trivial algorithm that always returns the unsafe class would result in: Expected Cost = 250 ∗ 1 + 0 ∗ 4 384 = 0.65 Accuracy = 125 384 = 32.6% Thus, as this example illustrates, if a learning algorithm aims to minimise the cost, there is a danger that it will achieve this by sacrificing overall accuracy significantly. Of course, if there was an algorithm that produced high accuracies for both classes, then there is no need to sacrifice accuracy in one of the classes and indeed, no need for cost-sensitive algorithms, since accuracy based algorithms would suffice. Given this motivation, this paper presents an alternative algorithm that aims to improve the accuracy of the unsafe class whilst aiming to retain overall accuracy and which does not require the costs of misclassifications. The paper is organised in two main parts: section 2 develops the algorithm and section 3 presents an evaluation of the algorithm with respect to three other decision tree algorithms on six data sets. 2 Development of a Safer Induction Algorithm The central idea of the algorithm is to err on the side of caution and aim to avoid concluding that a process or situation is safe without supporting examples. This aim is not easy to implement with current algorithms as we now illustrate. When faced with continuous attributes, most decision tree learning algorithms discretise the attributes resulting in what are known as axis-parallel splits (such as in C4.5 (Quinlan, 1993) and Remind (Althoff, Auriol, Barletta, & Manago, 1995)). 4 Figure 1 illustrates what happens when axis-parallel splits are used in an example where there are two classes denoted by ‘x’ and ‘o’. Line 1 represents the first division, line 2 the second division, and so on. Consequently, the decision tree for this example would have 5 non-leaf nodes (one for each dividing line) and 6 leaf nodes for the conclusions (one for each region). Thus, to achieve our aim of not drawing conclusions about a region being safe, the only freedom we have is to move the lines along each of the axes. 4The exceptions, CART and OC1 are considered later. 3 x x x x x x x x x x x x x x x x x x x x o o o o o o o o o o o o o o o o o o o o o o o o o o o ox x X2 X1 1 2 3 o 4 5 Figure 1: An ID3 division x x x x x x x x x x x x x x x x x x x x o o o o o o o o o o o o o o o o o o o o o o o o o o o ox x X2 X1 1 o Figure 2: A linear or oblique division However, a greater degree of freedom would be possible if we could use oblique divi- sions, which are linear combinations of the attributes (i.e., hyperplanes), to identify a more appropriate division of the space. For figure 1, we could adopt the oblique division shown in figure 2. Such a division would result in a decision tree that has only one non-leaf node and the conclusions would be based on more examples. Of course, this example is for illustrative purposes and is simple enough to allow a single linear relationship to divide the classes. In general, the problem will be much more complex: a clean separation may not be possible, more attributes are usually involved and there may be some discrete, as well as continuous attributes present. The following describes how the greater degree of freedom available with linear relationships is used to obtain safer divisions and incorporated as part of a top-down tree induction algorithm. 4 Following the intuition suggested by the above example, the aim is to identify a linear combination of the form: m∑ j=1 aj xj ≤ c (1) where m is the number of continuous attributes, aj are the coefficients, xj are the values of the continuous attributes and c is the threshold. This linear combination should be such that it “best” separates the two classes. By “best” we mean the division that minimizes the probability of incorrectly classifying an example. With this interpretation of “best”, we can use a well known statistical technique, called the linear discriminant method, for identifying the initial form of such relationships. As described below, the identified relationship is then modified so that the resulting decision tree errs on the safe side. The linear discriminant method enables us to find relationships like (1) by using the following equation (Morrison, 1976): xtΣ−1(µ1 − µ2) − 1 2 (µ1 + µ2) ≤ ln ( P (C2) P (C1) ) (2) where x is a vector representing the new example to be classified, µ1, µ2 are the mean vectors for the two classes, Σ is the pooled co-variance matrix, and P (Ci) is the probability of an example being in class Ci. Theoretically, it can be shown that this equation minimizes the misclassification rate when x has a multivariate normal distribution and when the co-variance matrices for each of the two groups are equal (this is sometimes known as the homogeneity condition). Although in practice, this equation is sometimes used without testing these optimality conditions, there remains a danger that it could give poor results if the data departs far from these conditions. Hence later, when we incorporate such linear discriminants into a decision tree learning algorithm, we will guard against this possibility. In the context of safety, we need to ensure that the linear discriminant obtained by the above equation is on the safe side. In particular, there may be situations where the classes overlap and it is not possible to separate them using a hyperplane. Since the linear discrimination method minimizes the probability of an error, it may misclassify some of the examples of both classes. Figure 3 illustrates this problem for a two-attribute and two-class problem. If we adopt such a linear split, each of the regions would be further sub-divided by the ID3 based algorithms and therefore result in a smaller version of the kind of problem we illustrated in figure 1. To reduce this, we can adjust the linear discriminant so that only one of the regions requires further sub-divisions. In making such an adjustment, we have a choice: 1. Adjust the discriminant so that one of the regions has just unsafe examples. 2. Adjust the discriminant so that one of the regions has just safe examples. If we choose the second alternative, there may be a few safe examples in a region that has numerous unsafe examples. Hence, at some point, the learning algorithm will further 5 x x x x x x x x x x x x x x x x x x x o o o o o o o o o o o o o o x x X2 x X1 A ox o Unsafe x Safe o x xx o o x x Figure 3: An illustration of overlapping classes sub-divide the unsafe region and isolate part of that region as safe. That is, a region would be inferred as safe based on just a small number of examples which is contrary to our aims. As an illustration of this effect, figure 4(a) shows the regions identified when we apply this approach on figure 3. Notice that the region on the right of line A and below line B is considered as safe based on just two examples. In contrast, if we select the first o X2 x xx o o X1 A B x x Safe Region UnsafeSafe Region Region (a) An unsafe linear adjustment. o X2 x xx o o X1 B A x x C Safe Unsafe UnsafeSafe Region Region Region Region (b) A safer linear adjustment. Figure 4: An illustration of adjustments alternative, some regions may be inferred as unsafe based on a small number of examples. As an illustration, figure 4(b) shows the effect of applying this approach on figure 3. Hence, since we wish to err on the side of safety, we prefer to adjust the discriminant towards the first alternative. To achieve this, the linear discriminant is adjusted by changing a coefficient (i.e., one of ai, or c in equation (1)) in a way that captures the furthest misclassified safe example and results in the best separation (as measured by the information gain). That is, equation (1) is used to compute a new value for each coefficient so that 6 it excludes the furthest misclassified safe example and by keeping the other coefficients constant. The new coefficient that results in the best division is then adopted. Once we have this method for identifying suitable linear divisions and adapting them to err on the safe side, we need to incorporate it within the usual inductive learning algorithm used by the ID3 family. This is done in a manner that remains faithful to the principle be- hind ID3 based algorithms. That is, from the available divisions: axis-parallel splits, linear division and the other attributes, we select the one that maximizes the information gained. Adopting this view gives added protection against situations where the data does not satisfy the theoretical conditions that guarantee a minimum probability of misclassification (i.e., the homogeneity condition described above), since the algorithm can default to the usual ID3 process. Figure 5 summarizes the modification. IF all examples belong to the same class THEN stop ELSE BEGIN Find a linear discriminant on the safe side. Find an axis-parallel split. Evaluate all discrete attributes. Select the test with the best measure from the above three. Partition the examples with respect to the selected test. Repeat the process for each partition. END Figure 5: Top-level of LDT The process of constructing axis-parallel splits is similar to that found in algorithms such as C4.5 except that the threshold is chosen to be on the safe side. More precisely, instead of selecting the mid-point between two consecutive examples of an interval, selecting the value of an example that lies on the safe boundary reduces the risk of misclassifying an unsafe example as safe. Decision tree induction algorithms of the kind proposed above are known to result in trees that overfit the data and are unnecessarily complex. One way of reducing this problem, which is widely used, is to adopt post-pruning methods (Quinlan, 1987; Mingers, 1989). The central operation in most post-pruning methods is to replace a subtree by a leaf node which takes the value of the majority class of the examples in the subtree. Since in practice, the unsafe class usually has fewer cases than the safe class, post-pruning methods that do not take account of safety may result in more accurate trees but without improving their safety. Hence, we now consider how to adapt an existing post-pruning method so that it takes account of safety and can be coupled with the tree induction method proposed in this paper. The literature contains numerous methods for post-pruning including: minimum cost- complexity pruning (Breiman et al., 1984), critical value pruning (Mingers, 1989), error- based pruning (Quinlan, 1993) and cost sensitive pruning (Knoll, Nakhaeizadeh, & Tausend, 1994). The reader is referred to the survey paper by Breslow and Aha (1997), a very 7 good comparison by Frank and Witten (1998) and an analysis of the reduced-error pruning method by Elomaa and Kaariaine (2001) for further information on pruning methods. Knoll et al.’s work is particularly relevant since it suggests modifying the accuracy based replacement criteria by one that takes account of the cost of misclassification. Thus, we have a ready made method that can take account of safety: simply allocate a large cost of misclassification to the unsafe class relative to the safe class. Unfortunately, it is not that easy, since assigning a relatively large cost to one class simply has the effect of pruning trees so that the accuracy in the other class is sacrificed to an unacceptable level. This behaviour is confirmed in Knoll et al.’s empirical evaluation, where in a credit allocation problem they adopt a misclassification cost of 1 for the “no risk” class and a misclassification cost of 13 for the “risk” class. In analyzing the results obtained for this problem, they conclude (Knoll et al., 1994, p. 385): “Concerning the credit data, the cost matrix is strongly asymetric. As a con- sequence, the pruning methods tried to classify all the test examples as “risk”, i.e. granting no credit at all. Obviously such a classifier is useless in practice.” In a related study, Bradford, Kunz, Kohavi, Brunk, and Brodley (1998) carry out an empirical evaluation of the effect of cost-based versions of pruning algorithms, similar to those proposed in (Knoll et al., 1994), on trees produces by systems such as C4.5. Surpris- ingly, they conclude that such pruning methods appear not to be effective in reducing costs, even relative to results from unpruned trees. Thus, instead of using Knoll’s approach, one of the other techniques, known as the minimum cost-complexity pruning method (Breiman et al., 1984), is adapted for our purposes. This technique works in two stages. First, it generates a series of pruned trees. This is done by considering the effect of replacing each subtree by a leaf and working out the reduction in error per leaf (α): α = R(t) − R(Tt) Nt − 1 (3) where R(t) is the expected error rate if the subtree is pruned, R(Tt) is the expected error rate if the subtree is not pruned, and Nt is the number of leaf nodes in the subtree. The subtree with the least value per leaf (smallest α) is pruned by replacing the subtree by its majority class, and the process repeated recursively to generate a series of increasingly pruned trees. Second, once a series of increasingly pruned trees have been generated, it uses an independent testing set to select a tree that is the most accurate. 5 In its pure form, the minimum cost-complexity pruning (MCCP) method does not take account of costs of misclassification or safety. To take account of safety, the MCCP method is adapted as follows. Given that the primary goal is to minimize the number of misclassified unsafe examples, the weakest subtrees are those that have least effect on the misclassification of unsafe examples irrespective of the number of leaves in the subtree. This leads to an adaptation of equation (3) so that R(t) is the expected error rate on the unsafe class if the subtree is pruned, R(Tt) is the expected error rate on the unsafe class if the subtree is not pruned, and the divisor Nt − 1 is omitted. In addition, the second part of MCCP is modified to select the safest pruned tree instead of the most accurate. Section 3 will include 5More precisely, it selects the smallest tree within one standard error of the most accurate tree. 8 the results of coupling this simple adaptation of the MCCP method with the developed algorithm. This concludes the development of the algorithm and an associated pruning method. The next section presents an evaluation with respect to several algorithms. 3 Empirical Evaluation with Respect to Existing Systems As suggested by the introduction, decision tree learning is a substantial research area, with many algorithms. The choice of which algorithms we should compare with is guided by the following potential counter claims: • Existing cost-sensitive algorithms can handle this problem by simply allocating a larger nominal cost to the more expensive class. • Any potential improvement may be due to the use of oblique divisions only and algorithms that utilise such splits may be just as effective. Thus, the empirical evaluation needs to be with respect to a cost-sensitive algorithm and an algorithm that induces oblique decision trees. But which ones and why? There are several systems that aim to take account of the costs (e.g., those proposed by (Pazzani et al., 1994; Draper, Brodley, & Utgoff, 1994; Provost & Buchanan, 1995; Turney, 1995)). 6 Despite some differences in their approaches, most of these systems extend the attribute selection measure to include cost and aim to optimize a function that measures the cost of a decision tree. Amongst these systems, the ICET system is possibly the best known system and one which Turney has demonstrated to be superior to several systems (Turney, 1995, p384). In another empirical study, ICET performs better than LMDT (Vadera & Ventura, 2001). Algorithms that utilise oblique splits include CART (Breiman et al., 1984), OC1 (Murthy, Kasif, & Salzberg, 1994) and RLP (Bennett, 1999). Murthy developed OC1 as an improve- ment of CART and shows OC1 to be empirically superior. Hence, given these considerations, we include an evaluation with respect to ICET, OC1, RLP and an axis parallel (AP) algorithm. The following subsection summarises the main characteristics of the selected algorithms and subsection 3.2 presents the empirical evalua- tion. 3.1 Summary of ICET, OC1 and RLP Summary of ICET The ICET system takes an evolutionary approach to inducing cost effective decision trees (Tur- ney, 1995). It utilizes a genetic algorithm, GENESIS (Grefenstette, 1986), and an extension of C4.5 in which a cost function is used instead of the information gain measure. The cost function used is borrowed from the EG2 system (Núñez, 1991) and takes the following form 6The reader can browse the following for a more complete bibliography: http:// mem- bers.rogers.com/peter.turney/ml cost.html 9 for the ith attribute: ICFi = 2∆i − 1 (Ci + 1)ω Where ∆i is the information gain, Ci the cost of carrying out the ith test and ω is a bias parameter used to control the amount of weight that should be given to the costs. The central idea of ICET is summarized in figure 6. The genetic algorithm GENESIS GENESIS C4.5 Biases Ci, ω, CF Trees Fitness Figure 6: The ICET system begins with a population of 50 individuals where each individual consists of the parameters Ci, ω, CF whose values are randomly selected initially. The parameters Ci, ω are utilized in the above equation and the parameter CF is used by C4.5 to decide the amount of pruning. Notice that in ICET, the parameters Ci are biases and not true costs as defined in EG2. Given the individuals, C4.5 is run on each one of them to generate a corresponding decision tree. Each decision tree is obtained by applying C4.5 on a randomly selected sub-training set, which is about half the available training data. The fitness of each tree is obtained by using the remaining half of the training data as a sub-testing set. The fitness function used by ICET is the cost of the tests required plus the cost of any misclassifications averaged over the number of cases in the sub-testing set. The individuals in the current generation are used to generate a new population using the mutation and cross over operators. This cycle is repeated 20 times and the fittest tree is returned as the output. Summary of OC1 The OC1 algorithm is motivated by the process adopted by CART. The following first summarises how CART finds the linear relationships and then how OC1 improves upon CART. The method used by CART works in two stages. First it uses a heuristic to find a linear combination of the continuous attributes that best divides the space. Then it eliminates those continuous attributes that have little influence (a process called backward elimination in CART). This latter process of eliminating variables is carried out by simply comparing the benefit of retaining a variable over deleting it and is aimed at reducing the complexity of the equation rather than improving the accuracy of a decision tree. The first step, that of finding the linear combination, is central to CART and is now described in more detail. CART aims to find a hyperplane by searching for the coefficients and a threshold (for equation 1) that divides the space of examples so that its evaluation function (GINI 10 index in CART) is minimized. Initially, the linear form takes an arbitrary set of coefficients and an arbitrary threshold. It then cycles through the continuous attributes one by one, updating the coefficients and the threshold to find an improved linear form with respect to the evaluation function and the training examples. This is repeated until the change in measures between the combination obtained in the previous cycle and the new division is less than a pre-specified amount. An individual coefficient, ak is updated by viewing equation 1 as a function of ak so that an example is above the current hyperplane if (for positive xk): ak > c − ∑j 6=k aj xj xk (4) A value for ak can then be found by evaluating the right hand side of this equation for each training example, sorting them and selecting the mid-point between two successive values that results in the best division. Murthy et al. (1994) argue that CART’s approach to identifying a hyperplane is focused on finding hyperplanes that are locally optimal and can therefore result in the identification of “poor” hyperplanes. Hence, although OC1’s approach to finding a linear combination is similar to CART ’s cycling process, it provides the following two key refinements that enable it to escape local minima: • When a locally optimal linear form is obtained, it is perturbed in a random direction. The amount by which it is perturbed in the random direction is then optimized with respect to the evaluation function. If this new hyperplane is better than the previous one, then the cycling process is repeated with this new hyperplane. The number of such random jumps it performs is a user specified parameter. • The linear combination to which the iterative process will converge to, depends mainly on the initial starting hyperplane. Hence, OC1 also takes different random starting hyperplanes to see if they converge to better hyperplanes. The number of restarts it makes is also a user specified parameter. Summary of RLP A number of authors have proposed the use of linear programming to obtain linear discrim- inants (see (Bennett, 1999) for a good overview). The central idea is to formulate a linear programming problem where the constraints are based on discriminating the classes and an objective function that minimises the distances between the misclassified points and the discriminant. Given matrices A and B that consist of m and k examples from their respective classes, and a vector of ones e, Bennett and Mangasarian (1992) develop the following linear pro- gramming formulation, called Robust Linear Programming (RLP) to identify a discriminant wx = γ. 11 Data Set No of examples %Safe %Unsafe No of attributes Pima Indians Diabetes 768 65 35 9 Intensive Therapy 400 77 23 16 Breast Cancer 683 65 35 11 Housing 506 49 51 14 Cleveland Heart 297 54 46 14 Hepatitis 155 79 21 20 Table 2: Characteristics of the data sets Objective: min w,γ,y,z 1 m ey + 1 k ez Subject to: Aw − eγ + y ≥ e −Bw + eγ + z ≥ e y, z ≥ 0 This minimisation problem can be solved using standard linear programming methods and incorporated as part of the top down procedure outlined in figure 5 resulting the RLP algorithm. 3.2 Empirical Evaluation This section presents the results of an empirical evaluation of the tree induction algorithm developed in this paper relative to the algorithms outlined above. The section also includes the results obtained when axis parallel splits are utilised, since such split are used as a comparative benchmark by a number of authors. The OC1 system is available and RLP was relatively easy to implement with the aid of the Linear programming package, loqo (Vanderbe, 2002). Implementation of ICET is more involved and some of the experiments in Turney(1995) were repeated to ensure that the implementation was faithful (see appendix B). Since post-pruning potentially adds its own effects on the results, experiments were carried out both before and after post-pruning was employed. The experiments utilized six data sets whose basic characteristics are summarized in table 2. The diabetes, breast cancer, housing, heart and hepatitis data sets are from the UCI repository of machine learning databases where they are fully described (Blake & Merz, 1998). For the housing data set, we followed Murthy et al’s (1994) discretisation of the class so that values below $21000 are of class 1 and otherwise of class 2. Arbitrarily, values below $21000 were considered safe. The intensive therapy data set was obtained from Manchester Royal Infirmary in the UK.7 It consists of a database where each record describes a patient’s 7This data set is available from the author. 12 state by measures such as age, temperature, mean arterial pressure, heart rate, respiratory rate, and has a class variable with values “critical” or “not critical”. Bearing in mind our earlier comments about the optimality conditions for linear dis- criminant analysis (p. 5), it is worth noting that none of the selected data sets satisfy the homogeneity condition for the covariance matrices. This was tested by Box’s M test using SPSS (Norusis, 1994, p73). The diabetes data, which was the closest to satisfying the optimality requirements, resulted in Box’s M =229, F = 6.2 with (36,1049414) DF, and zero significance. That is, these data sets are not, somehow biased towards the use of linear discriminant analysis, and therefore the algorithm developed in this paper. Results before pruning For each data set, we randomly partitioned the data into a training and a testing set where the training set was used to learn the decision tree and the testing set was used to evaluate its accuracy. To ensure robustness of the results, the experiments involved three different proportions of training/testing sets: 30/70, 50/50, and 70/30. For each of these training/testing splits, 50 experiments with randomly selected examples were carried out. Figure 7 shows the results obtained for the average accuracy on the unsafe class using OC1, axis-parallel (AP), Robust Linear Programming (RLP) and the algorithm presented in this paper (LDT). It presents the results for each of the data sets and across each of the training/testing splits. 8 Table 3 presents the average results for the 50/50 training/testing experiments (the complete results are given in appendix A, and the conclusions drawn apply to all the results). The table gives the averages for the overall accuracy, the accuracy of the safe class, the accuracy of the unsafe class, the depth of trees, and the time taken on a Sun SPARCStation2. The OC1 results were obtained by using the system supplied by Murthy et al. (1994) with the default parameters: number of restarts set to 20, and the random directional jumps set to 5. The results for the axis-parallel approach were obtained by using the axis-parallel option of OC1.9 In all cases, the Information Gain measure, which is probably the most widely used, was employed. These results show that, in general, when pruning is omitted, the developed algorithm produces safer trees than both OC1, AP and RLP without a significant loss in the overall accuracy. The only exception to this is on the 70/30 Hepatitis trial where the AP algorithm does produce better results than all the other algorithms. The degree of improvement in safety varies from a significant improvement for the intensive therapy data set to only a very small improvement in the breast cancer data set. The fact that there is only a small improvement in the breast cancer data set is probably due to the “ceiling effect”. That is, since the accuracy of the algorithms is already very high and close to the upper bound, the level of improvement obtained by any algorithm will be limited (Cohen, 1995). The behaviour of ICET is also of interest, since although Turney (1995) notes the relative difficulty of the problem of inducing safer trees, ICET is shown to perform better 8The use of ROC curves was also considered but is not appropriate given the lack of variation in the results and the more direct presentation is preferred. Readers interested in ROC can refer to (Provost, Fawcett, & Kohavi, 1998; Martin, Doddington, Kamm, Ordowski, & Przybocki, 1997) for some of the non-trivial issues with their use for such problems. 9This produces similar results to C4.5 in the empirical trials presented in (Murthy et al., 1994). 13 Diabete s 50.0 52.0 54.0 56.0 58.0 60.0 62.0 30/70 50/50 70/30 Training/Tes ting se t % A c c u ra c y Intensive Therapy 0.0 10.0 20.0 30.0 40.0 50.0 30/70 50/50 70/30 Training/Tes ting se t % A c c u ra c y Breast Cance r 87 88 89 90 91 92 93 30/70 50/50 70/30 Training/Tes ting se t % A c c u ra c y Hous ing 78 79 80 81 82 83 84 85 86 87 30/70 50/50 70/30 Training/Tes ting se t % A c c u ra c y Hepatitis 44.0 46.0 48.0 50.0 52.0 54.0 56.0 58.0 60.0 62.0 30/70 50/50 70/30 Training/Tes ting se t % A c c u ra c y LDT OC1 AP RLP Heart 66.0 68.0 70.0 72.0 74.0 76.0 30/70 50/50 70/30 Training/Testing set % A c c u ra c y Figure 7: Summary of accuracy for the unsafe class without pruning 14 Data set System Accuracy Tree Time Overall Safe Unsafe Depth (secs) Diabetes LDT 69.4±1.5 73.8±0.9 61.3±1.4 17.1±0.7 3.3±0.09 OC1 68.5±1.2 75.5±1.0 55.4±1.6 11.2±0.3 192.3±5.10 AP 69.7±1.1 76.4±0.9 57.2±1.4 14.2±0.4 1.8±0.03 RLP 67.4±0.7 72.5±1.1 57.9±1.5 46.0±3.8 27.3±0.31 Breast Cancer LDT 95.1±1.0 96.4±0.3 92.6±1.1 7.1±0.4 1.1±0.03 OC1 94.3±0.5 96.4±0.3 90.4±1.2 6.3±0.3 91.5±7.27 AP 94.5±0.8 96.4±0.4 90.8±1.0 7.7±0.4 0.7±0.02 RLP 94.8±0.4 96.4±0.3 91.9±1.2 16.5±5.1 13.8± 0.52 Heart LDT 74.1 ± 1.1 73.9 ± 1.9 74.7 ± 1.6 9.8± 0.5 1.3± 0.05 OC1 74.0 ± 1.0 74.4 ± 1.7 73.9 ± 1.7 6.8± 0.3 44.8± 1.26 AP 72.9 ± 0.9 74.4 ± 1.7 71.3 ± 1.6 7.8± 0.3 0.7± 0.02 RLP 76.1 ± 1.1 77.8 ± 2.2 74.6 ± 1.7 26.6± 6.1 7.4± 0.25 Hepatitis LDT 80.5 ± 1.1 87.0 ± 1.5 56.0 ± 3.6 5.1 ± 0.4 0.7 ± 0.1 OC1 80.7 ± 1.5 87.4 ± 1.6 54.6 ± 4.4 4.7 ± 0.4 22.1 ± 1.7 AP 80.0 ± 1.5 87.1 ± 1.7 52.7 ± 3.9 5.6 ± 0.4 0.4 ± 0.0 RLP 79.4 ± 1.3 86.1 ± 1.4 53.3 ± 4.2 4.8 ± 3.3 3.4 ± 0.1 Housing LDT 83.3±1.2 81.3±1.3 85.2±1.1 10.2±0.5 2.3±0.07 OC1 82.4±0.8 82.3±1.2 82.6±1.2 7.7±0.3 104.6±4.41 AP 81.7±0.8 81.3±1.2 82.1±1.1 9.2±0.4 1.4±0.03 RLP 83.1± 0.5 81.2 ± 1.2 85.0 ±1.1 25.3 ±5.9 18.3 ±0.40 Intensive Therapy LDT 70.4±1.0 78.2±1.3 44.2±2.4 11.2±0.5 3.0±0.09 OC1 72.5±0.6 83.0±1.0 37.8±2.8 8.2±0.4 99.1±2.87 AP 72.2±0.7 81.9±1.0 39.7±2.3 9.7±0.4 1.4±0.04 RLP 70.5± 1.0 80.6 ±1.4 37.2± 2.4 25.0±5.9 17.3±0.46 Table 3: Results for 50/50 experiments before pruning together with 95% confidence interval 15 than other cost-sensitive algorithms in such situations (Turney, 1995, figure 5, p389). Thus it is worth investigating ICET’s behaviour as the cost of misclassification in the unsafe class is increased relative to the safe class. Figure 8 shows the results obtained by ICET prior to pruning when the 50/50 experiments are carried out and as the ratio of misclassification costs is increased. As the results show, there is little variation and the trees produced are not as safe as those produced by LDT. 10 It would be reasonable to suggest that, given the evolutionary nature of ICET, removing the pruning component restricts the search space unfairly and is bound to result in the lack of variation displayed in this figure. Hence, discussion of this behaviour is postponed until the results after pruning are presented. Diabetes 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Breast Cancer 88% 90% 92% 94% 96% 98% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Hepatitis 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Intensive Therapy 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Housing 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Heart 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Overall Safe Unsafe Figure 8: ICET behaviour before pruning, as the misclassification ratio unsafe/ safe is increased 10Note that the large ratio of costs is chosen to give it every chance to show variation. 16 Bre as t Cance r 86 88 90 92 94 96 98 30/ 70 50/ 50 70/ 30 Training/Te sting s e t % A cc u ra cy Diabe te s 0 10 20 30 40 50 60 70 80 30/ 70 50/ 50 70/ 30 Training/Tes ting s et % A cc u ra cy Hous ing 70 75 80 85 90 95 30/ 70 50/ 50 70/ 30 Training/Tes ting s e t % A cc u ra cy Inte ns ive The rapy 0 10 20 30 40 50 60 30/ 70 50/ 50 70/ 30 Training/Tes ting s et % A cc u ra cy He art 60 65 70 75 80 85 30/ 70 50/ 50 70/ 30 Training/Te sting s e t % A cc u ra cy He patitis 0 10 20 30 40 50 60 70 30/ 70 50/ 50 70/ 30 Training/Tes ting s e t % A cc u ra cy LDT OC1 AP RLP Figure 9: Summary of accuracy for the unsafe class after pruning Results after pruning As mentioned earlier, post-pruning methods can result in improved accuracy, so the above experiments were repeated with pruning. The pruning methods employed were the default methods associated with the systems: OC1 and AP utilize the MCCP method, and LDT is coupled with the safer variation of MCCP described in section 2. Figure 9 presents the results for the unsafe class, and table 4 gives the results for the 50/50 experiments (the complete results are given in appendix A). The results obtained show that LDT produces significantly safer trees than the other methods on almost all the datasets. The only dataset where it does not improve the safety of the trees relative to the other methods is the breast cancer data, where both LDT and OC1 obtain an accuracy of 94.3% for the unsafe class. When comparing overall accuracy, the only significant difference occurs on the intensive therapy dataset, where the performance of both OC1 and AP is better than LDT. But since the accuracy on the unsafe class is 17 Data set System Accuracy Tree Time Overall Safe Unsafe Depth (secs) Diabetes LDT 69.9±0.8 67.6±1.7 74.6±2.1 15.3±0.8 3.1±0.08 OC1 71.4±0.8 80.3±1.8 55.0±3.6 5.1±1.0 167.8±4.25 AP 73.8±0.6 84.2±1.5 54.3±2.4 5.0±1.1 1.7±0.03 RLP 75.1 ± 0.7 81.4 ± 1.3 63.3 ± 2.1 3.2 ± 0.6 23.0 ± 0.30 Breast Cancer LDT 95.7±0.3 96.5±0.3 94.3±1.0 5.5±0.6 1.0±0.03 OC1 94.6±0.5 94.7±0.7 94.3±1.2 1.9±0.3 82.2±5.04 AP 94.1±0.5 94.9±0.7 92.6±1.2 3.1±0.5 0.7± 0.02 RLP 96.5 ± 0.2 96.9 ± 0.3 95.6 ± 0.6 1.2± 0.2 12.0 ± 0.50 Heart LDT 75.0 ± 1.1 72.8 ± 2.1 78.0 ± 1.5 9.8 ± 0.6 1.2 ± 0.06 OC1 75.8 ± 1.2 77.9 ± 2.5 73.6 ± 2.5 2.4 ± 0.5 38.0 ± 1.04 AP 74.4 ± 1.1 77.8 ± 2.1 70.7 ± 2.5 3.6 ± 0.7 0.6 ± 0.01 RLP 79.3 ± 0.9 82.1 ± 1.5 76.2 ± 1.9 1.8 ± 0.4 6.4 ± 0.18 Hepatitis LDT 79.4 ± 1.4 85.6 ± 1.8 56.2 ± 4.2 5.0 ± 0.4 0.6 ± 0.04 OC1 79.3 ± 1.5 85.4 ± 2.1 56.2 ± 4.5 1.6 ± 0.3 18.5 ± 1.48 AP 80.2 ± 1.3 89.4 ± 2.1 45.6 ± 6.6 1.8 ± 0.3 0.4 ± 0.02 RLP 78.5 ± 1.4 84.9 ± 1.4 54.7 ± 4.4 1.0 ± 0.0 2.9 ± 0.10 Housing LDT 82.7±0.6 76.2±1.8 88.8±1.2 8.5±0.5 2.1±0.07 OC1 81.0±1.0 82.0±2.9 80.3±2.0 3.3±0.6 84.7±2.24 AP 81.5±0.6 84.1±1.9 79.3±2.0 3.7±0.8 1.3±0.03 RLP 85.2 ± 0.6 86.0 ± 1.4 84.5 ± 1.1 2.4 ± 0.6 15.8 ± 0.42 Intensive Therapy LDT 70.0±1.0 76.8±1.3 47.8±2. 9 9.9±0.6 2.6±0.08 OC1 74.8±0.9 89.8±2.4 25.1±5.3 2.5±0.6 82.0±3.51 AP 75.9±0.9 92.8±1.8 19.2±4.7 3.8±0.9 1.3±0.04 RLP 73.8 ± 1 88.5 ± 2.7 25.7 ± 5.9 2.1 ± 0.4 14.9 ± 0.4 Table 4: Results for 50/50 experiments together with 95% confidence interval after pruning 18 significantly poorer than LDT’s performance, such trees are useless in practice. Indeed, the overall accuracy of 74.8% obtained by OC1 masks the information that in more than a quarter of the trials, OC1 produces simple, one node trees, labelled with the safe class, that are 100% accurate on the safe class but 0% accurate on the unsafe class. A comparison of these results with those obtained without pruning (given in table 3) reveals the effects of pruning. As expected, pruning does improve accuracy in most cases. More interestingly, the use of pruning with OC1 and AP results in improving the accuracy of the unsafe class for just the breast cancer dataset. For the other 3 data sets, the use of pruning with OC1 and AP actually results in a decrease in the accuracy of the unsafe class. The adoption of the safer MCCP method does improve the accuracy of the unsafe class, but, this is at the expense of a reduction in the accuracy of the safe class. Although this is reasonable in the context of safety, these results show that pruning methods do not necessarily result in balanced improvements to the accuracy of all the classes. Earlier, discussion of ICET’s results prior to pruning (figure 8) was postponed in case the lack of variation in accuracy was due to the omission of pruning. As the results in figure 10 show, even after pruning is employed, the accuracy does not vary much on either class as the ratio of misclassification costs is increased. This behaviour is not too surprising when we consider how ICET works. At the bottom level (figure 6), it uses C4.5 to generate trees based on the post-pruning parameter CF, and a selection function ICF that has biases Ci and ω. The pruning algorithm adopted is the error based pruning method that does not directly take account of costs but may be influenced indirectly by the parameter CF. Thus the extent to which the trees generated by C4.5 are influenced by costs depends on how well the genetic algorithm learns biases that relate to the costs. The results show that in its current form, the genetic algorithm is unable to learn the biases in a way that is sufficient for generating trees that are biased towards a particular class. Prior to these experiments, this author (as well as Turney (1995), p 390) thought that ICET’s inability to improve the accuracy of the unsafe class was due to the fewer number of examples available relative to the safe class. So it is worth noting that, as the results of the housing data (where each class has approximately 50% of the data) show, ICET’s lack of sensitivity to unbalanced misclassification costs is independent of the proportion of examples available for training from each class. 4 Conclusion and Future Work This paper has presented a decision tree induction algorithm aimed at domains where safety is paramount, costs of misclassification are not easy to estimate accurately and where it is known that the cost of misclassification for one of the classes is significantly higher than another. The algorithm errs on the side of caution by aiming to avoid concluding that a process is safe without supporting examples. To improve the chances of achieving this goal the algorithm utilises oblique divisions in preference to the less flexible axis-parallel splits. The algorithm, called LDT, utilises the linear discriminant method to identify the initial division which is then adapted to favour the development of a tree that errs on the safe side. An alternative way of tackling the problem could be to utilise an existing cost-sensitive algorithm and use a high cost for the more costly class. Given the use of oblique splits, 19 Diabetes 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Intensive Therapy 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Breast Cancer 88% 90% 92% 94% 96% 98% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Housing 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Hepatitis 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Heart 0% 20% 40% 60% 80% 100% 0 .1 1 1 0 1 0 0 1 0 0 0 1 0 0 0 0 Ratio A c c u ra c y Overall Safe Unsafe Figure 10: ICET’s behaviour after pruning, as the misclassification ratio unsafe/safe is increased it could also be argued that existing oblique decision tree algorithms may be adequate. Hence, the algorithm is evaluated relative to OC1, the axis parallel approach (AP), Robust Linear Programming (RLP) and an implementation of ICET. Since pruning can have its own effects on accuracy, experiments were carried out both before and after pruning. The default pruning methods were adopted for OC1, AP, and ICET, while a simple variation of the MCCP method that took account of safety was used with LDT. Both before and after pruning, LDT produces trees that are significantly safer than OC1, AP, RLP and ICET. The only exception occurs on the breast cancer data set after pruning, where OC1 performs as well as LDT, and where, with accuracies of about 95%, there is not much scope for improvement. Since ICET is an approach that takes account of costs of misclassification, a reasonable proposal for inducing safer trees is to increase the cost of misclassification in the unsafe class relative to the safe class. Hence experiments, in which the ratio of costs of misclassifications was varied were carried out on an implementation of ICET. The results revealed that ICET is not particularly sensitive to imbalanced costs of misclassification and 20 that the trees do not become any safer as the ratio is increased. Further, the experiments show that this lack of sensitivity is not simply due to an imbalance in the proportion of examples available in the classes as previously thought. In general, the improvements in safety obtained by LDT are without sacrificing the overall accuracy. For the diabetes, breast cancer, and housing datasets, the overall accuracy obtained by LDT is similar to that obtained by OC1, both before and after pruning. For the intensive therapy data set, the overall accuracy obtained by OC1 and AP is significantly higher than that obtained by LDT. However, this is achieved by taking advantage of the low ratio of unsafe cases in the data set to the extent that over a quarter of the trees degenerate into always returning the safe class. Such degenerate trees are, of course, not much use in any application. Comparing the results obtained before and after pruning reveals that post-pruning meth- ods can have an unbalanced effect on the accuracy of the classes. The use of the default pruning strategies with OC1, AP, and ICET resulted in no significant improvement to the safety of the trees, while the use of the adapted pruning method improved the safety of the trees, but with a reduction in the accuracy of the safe class. There are several areas for further work. In a related study, non-linear divisions have been used to develop a cost-sensitive algorithm (Vadera, 2005) but which relies on the availability of costs. Adapting non-linear divisions for safety applications is likely to be more difficult but represents a natural extension of this work. Several authors have also experimented with the use of boosting for cost-sensitive problems (Domingos, 1999; Fan, Stolfo, Zhang, & Chan, 1999; Ting, 2000; Merler, Furlanello, Larcher, & Sboner, 2003). Attempts to use boosting coupled with the algorithm presented in this paper also represents a further potential enhancement. To conclude, this paper has developed a tree induction algorithm that results in safer trees whilst maintaining overall accuracy and which is more appropriate for safety applica- tions. 5 Acknowledgements The work presented in this paper is a significant development of work that was initiated with the help of a Research Assistant, Said Nechab in 1994-95. In particular, the motivation and examples in section 2 appear in a previous joint paper written and presented by the author at Knowldege Based Systems Conference in 1996. The author is grateful to David Ventura for implementing ICET as part of his MSc project, and for the results included in appendix B. The author is also grateful to Sreerama Murthy for making his OC1 system available and the anonymous referees for their comments. References Althoff, K., Auriol, E., Barletta, R., & Manago, M. (1995). A review of industrial case-based reasoning tools. Oxford: AI Intelligence. 21 Bennett, K. P. (1999). On mathematical programming methods and support vector ma- chines. In Schoelkopf, A., Burges, C., & Smola, A. (Eds.), Advances in Kernel Methods – Support Vector Machines, chap. 19, pp. 307–326. MIT Press, Cambridge, MA. Berry, M., & Linoff, G. (2004). Data Mining Techniques (2nd edition). John Wiley. Blake, C., & Merz, C. (1998). UCI Repository of Machine Learning Databases. [http: //www.ics.uci.edu/ mlearn/ MLRepository.html], Irvine, CA: University of California, Department of Information and Computer Science, USA. Bradford, J., Kunz, C., Kohavi, R., Brunk, C., & Brodley, C. (1998). Pruning decision trees with misclassification costs. In Proceedings of the Tenth European Conference on Machine Learning, pp. 131–136. Breiman, L., Friedman, J. H., Olsen, R. A., & Stone, C. J. (1984). Classification and Regression Trees. Belmont: Wadsworth. Breslow, L., & Aha, D. (1997). Simplifying decision trees: A survey. Knowledge Engineering Review, 12, 1–40. Cohen, R. (1995). Empirical Methods for Artificial Intelligence. Massachusetts : MIT Press. Domingos, P. (1999). MetaCost: A general method for making classifiers cost-sensitive. In Proceedings of the Fifth International Conference on Knowledge Discovery and Data Mining, pp. 155–164. Draper, B., Brodley, C. E., & Utgoff, P. E. (1994). Goal-directed classification using lin- ear machine decision trees. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16 (9), 888–893. Elkan, C. (2001). The foundations of cost-sensitive learning. In Proceedings of the 17th International Joint Conference on Artificial Intelligence, pp. 973–978. Elomaa, T., & Kaariaine, M. (2001). An analysis of reduced error pruning. Journal of Artificial Intelligence Research, 15, 163–187. Fan, W., Stolfo, S., Zhang, J., & Chan, P. (1999). AdaCost: Misclassification cost-sensitive boosting. In Proceedings of the 16th International Conference on Machine Learning, pp. 97–105. Morgan Kaufmann, San Francisco, CA. Frank, E., & Witten, I. (1998). Reduced-error pruning with significance tests, [http://citeseer.ist.psu.edu/frank98reducederror.html].. Grefenstette, J. (1986). Optimization of control parameters for genetic algorithms. IEEE Transactions on Systems, Man, and Cybernetics, 16, 122–128. Knoll, U., Nakhaeizadeh, G., & Tausend, B. (1994). Cost-sensitive pruning of decision trees. In Proceedings of the Eight European Conference on Machine Learning, Vol. 2, pp. 383–386, Berlin, Germany. Springer-Verlag. Kolodner, J. (1993). Case-based Reasoning. Palo Also, CA:Morgan Kaufman. Martin, A., Doddington, G., Kamm, T., Ordowski, M., & Przybocki, M. (1997). The DET curve in assessment of detection task performance. In Proceedings of Eurospeech ’97, pp. 1895–1898, Rhodes, Greece. 22 Merler, S., Furlanello, C., Larcher, B., & Sboner, A. (2003). Automatic model selection in cost-sensitive boosting. Information Fusion, 4, 3–10. Mingers, J. (1989). An empirical comparison of pruning methods for decision tree induction. Machine Learning, 4, 227–243. Morrison, D. (1976). Multivariate Statistical Methods (Second edition). New York : McGraw-Hill. Murthy, S., Kasif, S., & Salzberg, S. (1994). A system for induction of oblique decision trees. Journal of Artificial Intelligence Research, 2, 1–32. Norusis, M. (1994). SPSS Advanced Statistics 6.1. SPSS Inc., Chicago, Illinois 60611, USA. Núñez, M. (1991). The use of background knowledge in decision tree induction. Machine Learning, 6, 231–250. Pazzani, M., Merz, C., Murphy, P., Ali, K., Hurne, T., & Brunk, C. (1994). Reducing misclassification costs: Knowledge-intensive approaches to learning from noisy data. In Proceedings of the Eleventh International Conference on Machine Learning, pp. 217–225. Provost, F. J., & Buchanan, B. G. (1995). Inductive policy: The pragmatics of bias selection. Machine Learning, 20, 35–61. Provost, F. J., Fawcett, T., & Kohavi, R. (1998). The case against accuracy estimation for comparing induction algorithms. In Proceedings of the Fifteenth International Conference on Machine Learning, pp. 445–553. Quinlan, J. R. (1987). Simplifying decision trees. International Journal of Man-Macine Studies, 27, 221–234. Quinlan, J. R. (1993). C4.5 : Programs for Machine Learning. California:Morgan Kauffman. Tan, M. (1993). Cost-sensitive learning of classification knowledge and its applications in robotics. Machine Learning, 13, 7–33. Ting, K. (2000). A comparative study of cost-sensitive boosting algorithms. In Proceedings of the 17th International Conference on Machine Learning, pp. 983–990. Turney, P. (1995). Cost sensitive classification: Empirical evaluation of a hybrid genetic decision tree induction algorithm. Journal of Artificial Intelligence Research, 2, 369– 409. Vadera, S. (2005). Inducing Cost-Sensitive Nonlinear Decision Trees, Technical Report. School of Computing, Science and Engineering, University of Salford, Salford, M5 4WT, UK, [http://citeseer.ist.psu.edu]. Vadera, S., & Ventura, D. (2001). A comparison of cost-sensitive decision tree learning algo- rithms. In Proceedings of the Second European Conference on Intelligent Management Systems in Operations, pp. 79–86. Vanderbe, R. (2002). Loqo: Optimisation and Applications Web Site. Princeton University, NJ 08544, [http://orfe.princeton.edu/ loqo/]. Zadrozny, B., Langford, J., & Abe, N. (2003). A simple method for cost-sensitive learning. IBM Technical Report RC22666, [http://citeseer.ist.psu.edu/zadrozny03simple.html]. 23 A Results of Empirical Evaluations of LDT, AP, RLP and OC1 This appendix presents the complete set of results for LDT, AP,RLP and OC1 both before and after pruning. The experimental method used is described in the body of the paper. Diabetes (unpruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 68.1±1.6 73.1±1.2 58.9±2.0 13.3±0.5 1.7±0.05 OC1 67.7±0.8 75.2±1.2 54.0±1.6 9.4±0.4 92.0±2.46 AP 68.5±1.0 76.0±1.1 54.8±1.9 12.2±0.5 1.0±0.02 RLP 66.9 ± 0.7 71.8 ± 1.1 58.0 ± 1.7 38.4 ± 5.4 13.60 ± 0.20 50/50 LDT 69.4±1.5 73.8±0.9 61.3±1.4 17.1±0.7 3.3±0.09 OC1 68.5±1.2 75.5±1.0 55.4±1.6 11.2±0.3 192.3±5.10 AP 69.7±1.1 76.4±0.9 57.2±1.4 14.2±0.4 1.8±0.03 RLP 67.4 ± 0.7 72.5 ± 1.1 57.9 ± 1.5 46.0 ± 3.8 27.3 ± 0.31 70/30 LDT 68.7±2.2 73.8±1.0 59.5±1.8 19.1±0.6 5.0±0.08 OC1 68.4±1.3 75.8±1.1 55.1±1.7 12.7±0.4 313.0±9.60 AP 69.6±1.4 77.5±1.0 55.3±1.6 15.4±0.5 2.7±0.05 RLP 68.0 ± 0.8 74.0 ± 1.1 57.1 ± 1.5 46.9 ± 3.3 42.8 ± 0.30 Breast Cancer (unpruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 94.9±0.8 96.4±0.3 92.0±0.9 5.5±0.3 0.6±0.03 OC1 93.6±0.5 95.8±0.5 89.5±1.4 4.4±0.3 43.0±2.09 AP 93.9±0.7 95.6±0.4 90.8±0.8 6.2±0.5 0.3±0.01 RLP 93.9 ± 0.4 96.2±0.37 89.64 ± 1.41 4.6 ± 2.7 6.3 ± 0.40 50/50 LDT 95.1±1.0 96.4±0.3 92.6±1.1 7.1±0.4 1.1±0.03 OC1 94.3±0.5 96.4±0.3 90.4±1.2 6.3±0.3 91.5±7.27 AP 94.5±0.8 96.4±0.4 90.8±1.0 7.7±0.4 0.7±0.02 RLP 94.8± 0.4 96.4 ± 0.3 91.9± 1.2 16.5±5.1 13.8±0.52 70/30 LDT 95.1±1.0 96.7±0.4 92.2±1.2 8.3±0.4 1.8±0.05 OC1 94.5±0.7 96.3±0.4 91.0±1.0 6.8±0.3 145.2±6.77 AP 94.6±0.8 96.2±0.5 91.7±1.0 8.7±0.4 1.2±0.04 RLP 93.5± 0.5 95.8± 0.4 89.2 ±1.5 14.6 ± 4.5 23.4 ± 0.60 24 Heart(unpruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 73.8 ± 1.0 74.0 ± 1.8 73.9 ± 2.3 7.5± 0.5 0.6± 0.03 OC1 73.1 ± 0.9 73.8 ± 2.0 72.5 ± 1.9 5.6± 0.4 19.9± 0.76 AP 72.9 ± 1.0 75.9 ± 1.6 69.6 ± 2.5 6.7± 0.4 0.4± 0.01 RLP 72.8 ± 1.0 74.9 ± 1.6 70.7 ± 2.2 9.0± 4.0 3.8± 0.15 50/50 LDT 74.1 ± 1.1 73.9 ± 1.9 74.7 ± 1.6 9.8± 0.5 1.3± 0.05 OC1 74.0 ± 1.0 74.4 ± 1.7 73.9 ± 1.7 6.8± 0.3 44.8± 1.26 AP 72.9 ± 0.9 74.4 ± 1.7 71.3 ± 1.6 7.8± 0.3 0.7± 0.02 RLP 76.1 ± 1.1 77.8 ± 2.2 74.6 ± 1.7 26.6± 6.1 7.4± 0.25 70/30 LDT 73.6 ± 1.2 72.0 ± 1.9 75.2 ± 2.0 11.2± 0.5 2.1± 0.07 OC1 74.8 ± 1.3 75.6 ± 2.1 74.2 ± 1.9 7.6± 0.3 75.3± 2.18 AP 73.4 ± 1.3 75.1 ± 2.0 71.6 ± 1.7 9.2± 0.5 1.1 ±0.02 RLP 74.1 ± 1.2 74.4 ± 2.3 74.1 ± 2.3 33.3± 5.8 12.2± 0.31 Hepatitis(unpruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 79.6 ± 1.3 86.0 ± 1.8 55.9 ± 3.7 3.0 ± 0.4 0.3 ± 0.0 OC1 80.0 ± 1.2 87.3 ± 1.8 53.3 ± 4.4 3.2 ± 0.3 8.6 ± 0.6 AP 79.7 ± 1.1 87.4 ± 1.5 50.5 ± 3.9 4.1 ± 0.3 0.2 ± 0.0 RLP 77.1 ± 1.6 83.4 ± 1.8 54.3 ± 4.2 1.1 ± 0.1 1.6 ± 0.0 50/50 LDT 80.5 ± 1.1 87.0 ± 1.5 56.0 ± 3.6 5.1 ± 0.4 0.7 ± 0.1 OC1 80.7 ± 1.5 87.4 ± 1.6 54.6 ± 4.4 4.7 ± 0.4 22.1 ± 1.7 AP 80.0 ± 1.5 87.1 ± 1.7 52.7 ± 3.9 5.6 ± 0.4 0.4 ± 0.0 RLP 79.4 ± 1.3 86.1 ± 1.4 53.3 ± 4.2 4.8 ± 3.3 3.4 ± 0.1 70/30 LDT 80.9 ± 1.7 86.9 ± 1.9 58.4 ± 4.5 7.1 ± 0.5 1.4 ± 0.1 OC1 82.9 ± 1.5 89.5 ± 1.7 57.4 ± 4.9 5.5 ± 0.3 38.3 ± 2.0 AP 81.9 ± 1.4 88.0 ± 1.8 59.6 ± 4.2 6.4 ± 0.4 0.6 ± 0.0 RLP 80.1 ± 1.4 87.4 ± 1.6 52.5 ± 4.5 13.2 ± 5.0 5.9 ± 0.4 25 Housing (unpruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 82.8±1.0 81.3±1.5 84.3±1.0 7.9±0.5 1.1±0.04 OC1 80.9±0.5 80.8±1.4 81.1±1.1 6.6±0.3 45.5±1.31 AP 80.9±0.7 80.2±1.3 81.6±1.1 8.1±0.4 0.7±0.02 RLP 82.3 ± 0.6 82.3 ± 1.3 82.4 ± 1.1 18.1± 5.8 9.2 ± 0.18 50/50 LDT 83.3±1.2 81.3±1.3 85.2±1.1 10.2±0.5 2.3±0.07 OC1 82.4±0.8 82.3±1.2 82.6±1.2 7.7±0.3 104.6±4.41 AP 81.7±0.8 81.3±1.2 82.1±1.1 9.2±0.4 1.4±0.03 RLP 83.1 ± 0.5 81.2 ± 1.2 85.0 ± 1.1 25.3 ± 5.9 18.3 ± 0.40 70/30 LDT 83.4±1.3 81.1±1.4 85.6±1.0 11.4±0.5 3.6±0.09 OC1 82.3±0.8 82.4±1.3 82.3±1.1 9.0±0.3 164.5±4.10 AP 82.6±0.8 82.8±1.5 82.6±1.2 11.0±0.4 2.3±0.04 RLP 83.2 ± 0.7 80.5 ± 1.6 85.7 ± 1.1 34.2 ± 5.8 30.1 ± 0.56 Intensive Therapy (unpruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 70.0±0.9 77.7±1.3 44.3±2.4 8.9±0.5 1.5±0.07 OC1 72.1±0.5 82.1±1.3 38.7±2.2 6.6±0.3 42.2±2.50 AP 71.5±0.7 81.5±1.4 38.1±2.0 7.9±0.4 0.7±0.03 RLP 69.5 ± 0.8 78.6 ± 1.4 39.1 ± 2.6 20.2± 5.9 8.3 ± 0.27 50/50 LDT 70.4±1.0 78.2±1.3 44.2±2.4 11.2±0.5 3.0±0.09 OC1 72.5±0.6 83.0±1.0 37.8±2.8 8.2±0.4 99.1±2.87 AP 72.2±0.7 81.9±1.0 39.7±2.3 9.7±0.4 1.4±0.04 RLP 70.5 ± 1.0 80.6 ± 1.4 37.2 ± 2.4 25.0 ± 5.9 17.3 ± 0.46 70/30 LDT 71.4±0.9 79.1±1.2 45.3±3.2 12.4±0.5 4.4±0.11 OC1 71.3±0.8 81.0±1.1 38.5±3.1 9.3±0.4 169.5±4.35 AP 72.0±0.8 81.5±1.2 39.6±3.3 11.0±0.5 2.2±0.05 RLP 70.1 ± 1.2 79.0 ± 1.5 39.7 ± 3.4 40.2 ± 5.1 29.4± 0.49 26 Diabetes (pruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 68.8±0.8 67.9±2.1 70.8±2.8 11.8±0.8 1.6±0.04 OC1 70.6±0.7 79.5±2.7 54.2±5.1 3.4±0.7 79.0±2.16 AP 71.7±0.7 82.5±2.4 52.0±3.5 4.9±1.1 0.9±0.02 RLP 74±0.54 80.82± 1.76 61.52 ± 3.12 4.78 ± 1.73 10.2 ± 0.47 50/50 LDT 69.9±0.8 67.6±1.7 74.6±2.1 15.3±0.8 3.1±0.08 OC1 71.4±0.8 80.3±1.8 55.0±3.6 5.1±1.0 167.8±4.25 AP 73.8±0.6 84.2±1.5 54.3±2.4 5.0±1.1 1.7±0.03 RLP 75.1 ± 0.7 81.4 ± 1.3 63.3 ± 2.1 3.2 ± 0.6 23.0 ± 0.30 70/30 LDT 69.9±0.8 67.6±1.7 74.6±2.1 15.3±0.8 3.1±0.08 OC1 71.7±0.9 79.9±2.1 56.6±3.3 5.2±0.9 276.5±6.98 AP 73.6±0.8 85.1±1.6 52.8±2.5 5.9±1.3 2.6±0.05 RLP 74.5 ± 0.9 81.9 ± 1.5 60.9 ± 2.8 3.1 ± 0.6 37.5 ± 0.37 Breast Cancer (pruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 95.8±0.5 96.0±0.5 95.3±1.2 5.5±0.9 1.6±0.05 OC1 94.0±0.4 93.8±0.8 94.3±0.8 1.3±0.2 36.1±1.95 AP 92.7±0.5 93.7±0.9 90.8±1.8 2.2±0.4 0.3±0.02 RLP 95.5 ± 0.4 96.3 ± 0.4 93.8 ± 1.0 5.1 ± 0.3 5.1 ± 0.30 50/50 LDT 95.7±0.3 96.5±0.3 94.3±1.0 5.5±0.6 1.0±0.03 OC1 94.6±0.5 94.7±0.7 94.3±1.2 1.9±0.3 82.2±5.04 AP 94.1±0.5 94.9±0.7 92.6±1.2 3.1±0.5 0.7± 0.02 RLP 96.5 ± 0.2 96.9± 0.3 95.6± 0.6 1.2± 0.2 12.0± 0.50 70/30 LDT 95.8±0.5 96.0±0.5 95.3±1.2 5.5±0.9 1.6± 0.05 OC1 95.4±0.4 95.6±0.6 95.2±1.0 2.2±0.5 123.4±5.88 AP 94.4±0.5 95.4±0.7 92.8±1.2 3.9±0.6 1.1±0.04 RLP 96.4 ± 0.4 96.6 ± 0.5 96.6 ± 0.5 2.5 ± 2.0 9.8±0.60 27 Heart(pruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 74.0 ± 1.0 72.7 ± 2.5 75.8 ± 2.1 7.2 ± 0.7 0.6 ± 0.05 OC1 74.3 ± 1.3 75.9 ± 3.2 72.6 ± 3.0 1.4 ± 0.2 18.0 ± 0.66 AP 73.9 ± 0.9 78.1 ± 2.0 69.1 ± 2.6 2.3 ± 0.5 0.3 ± 0.01 RLP 78.3 ± 1.0 80.1 ± 1.6 76.1 ± 1.9 1.1 ± 0.1 3.4 ± 0.13 50/50 LDT 75.0 ± 1.1 72.8 ± 2.1 78.0 ± 1.5 9.8 ± 0.6 1.2 ± 0.06 OC1 75.8 ± 1.2 77.9 ± 2.5 73.6 ± 2.5 2.4 ± 0.5 38.0 ± 1.04 AP 74.4 ± 1.1 77.8 ± 2.1 70.7 ± 2.5 3.6 ± 0.7 0.6 ± 0.01 RLP 79.3 ± 0.9 82.1 ± 1.5 76.2 ± 1.9 1.8 ± 0.4 6.4 ± 0.18 70/30 LDT 75.4 ± 1.4 70.1 ± 2.2 81.7 ± 1.8 11.0 ± 0.6 1.9 ± 0.07 OC1 75.4 ± 1.3 78.0 ± 2.6 72.6 ± 2.8 2.8 ± 0.5 62.7 ± 1.59 AP 75.7 ± 1.4 79.1 ± 2.2 71.8 ± 2.4 3.7 ± 0.7 1.0 ± 0.02 RLP 79.9 ± 1.2 82.0 ± 1.9 77.7 ± 1.8 2.2 ± 0.6 10.4 ± 0.29 Hepatitis(pruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 78.8 ± 1.4 85.4 ± 2.0 54.5 ± 4.0 2.3 ± 0.4 0.2 ± 0.01 OC1 78.5 ± 1.3 85.9 ± 2.2 50.8 ± 5.4 1.2 ± 0.2 7.2 ± 0.53 AP 78.2 ± 1.3 87.5 ± 2.5 44.3 ± 6.5 1.3 ± 0.2 0.2 ± 0.01 RLP 76.6 ± 1.4 83.3 ± 1.7 52.4 ± 4.3 1.0 ± 0.0 1.7 ± 0.03 50/50 LDT 79.4 ± 1.4 85.6 ± 1.8 56.2 ± 4.2 5.0 ± 0.4 0.6 ± 0.04 OC1 79.3 ± 1.5 85.4 ± 2.1 56.2 ± 4.5 1.6 ± 0.3 18.5 ± 1.48 AP 80.2 ± 1.3 89.4 ± 2.1 45.6 ± 6.6 1.8 ± 0.3 0.4 ± 0.02 RLP 78.5 ± 1.4 84.9 ± 1.4 54.7 ± 4.4 1.0 ± 0.0 2.9 ± 0.10 70/30 LDT 81.7 ± 1.3 87.1 ± 1.8 61.9 ± 4.2 6.5 ± 0.5 1.2 ± 0.08 OC1 82.4 ± 1.5 88.8 ± 1.9 58.2 ± 5.1 1.6 ± 0.3 31.4 ± 2.13 AP 82.3 ± 1.7 91.9 ± 1.7 47.1 ± 6.7 2.7 ± 0.5 0.6 ± 0.03 RLP 80.3 ± 1.5 86.0 ± 1.8 59.7 ± 4.6 2.3 ± 2.0 5.3 ± 0.30 28 Housing (pruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 81.9±0.6 76.9±1.5 86.6±1.1 7.2±0.7 1.1±0.04 OC1 80.9±0.6 82.6±1.8 79.2±1.9 2.8±0.6 38.2±1.55 AP 81.0±0.6 83.4±2.5 78.8±2.3 2.9±0.7 0.6±0.02 RLP 84.1± 0.7 84.4 ± 5.8 83.8 ± 4.5 1.5 ± 1.0 8.3 ± 0.14 50/50 LDT 82.7±0.6 76.2±1.8 88.8±1.2 8.5±0.5 2.1±0.07 OC1 81.0±1.0 82.0±2.9 80.3±2.0 3.3±0.6 84.7±2.24 AP 81.5±0.6 84.1±1.9 79.3±2.0 3.7±0.8 1.3±0.03 RLP 85.2 ± 0.6 86.0±1.4 84.5 ± 1.1 2.4±0.6 15.8 ± 0.42 70/30 LDT 82.1±1.9 71.9±4.5 91.5±1.4 8.9±0.7 3.2±0.11 OC1 82.3±1.1 82.4±2.3 82.4±1.6 4.3±0.7 146.6±5.64 AP 81.6±0.8 84.4±1.7 79.3±1.7 5.4±1.1 2.1±0.05 RLP 85.1 ± 0.8 85.3 ± 1.5 85.0 ± 1.3 3.2 ± 0.7 26.2 ± 0.60 Intensive Therapy (pruned) Train/Test System Accuracy Safe Unsafe Depth Time (s) 30/70 LDT 70.2±0.9 77.7±1.6 45.0±2.9 8.6±0.5 1.4±0.08 OC1 74.7±0.8 88.8±2.4 27.6±5.5 2.2±0.4 36.5±2.08 AP 75.3±0.7 93.9±2.0 13.3±4.2 1.7±0.4 0.7±0.03 RLP 72.4 ± 1.1 84.7 ± 3.2 31.9 ± 6.4 1.7 ± 0.3 7.1 ± 0.26 50/50 LDT 70.0±1.0 76.8±1.3 47.8±2.9 9.9±0.6 2.6±0.08 OC1 74.8±0.9 89.8±2.4 25.1±5.3 2.5±0.6 82.0±3.51 AP 75.9±0.9 92.8±1.8 19.2±4.7 3.8±0.9 1.3±0.04 RLP 73.8 ± 1 88.5 ± 2.7 25.7 ± 5.9 2.1 ± 0.4 14.9 ± 0.4 70/30 LDT 69.4±1.2 75.3±1.7 49.6±3.2 11.3±0.5 3.8±0.09 OC1 75.1±1.1 91.0±2.3 21.4±5.2 3.4±0.8 142.5±4.30 AP 76.4±1.0 93.4±1.9 18.7±5.2 3.7±0.9 2.0±0.05 RLP 74.1 ± 1.4 91.3 ± 2.7 15.8 ± 5.0 2.2 ± 0.4 25.2 ± 0.53 29 B Repeated Experiments using Re-implementation of ICET Figure 11 shows the results obtained by both the implementation of ICET utilized in this paper and the published results in (Turney, 1995, figure 3). The results give confidence that the implementation is a fair and accurate representation of ICET. The measure ’Average % Standard Cost’ labelling the Y axis is a normalized average cost of classification. This measure and details of the experimental method are as presented in Turney (1995). ICET results from Turney (1995). Results from implementation used in the paper. Diabetes 0 20 40 60 80 100 10 100 1000 10000 Misclassification Costs A v. % S ta n d ar d C o st Bupa Liver Disease 0 20 40 60 80 100 10 100 1000 10000 Misclassification Costs A v. % S ta n d ar d C o st Thyroid Disease 0 20 40 60 80 100 10 100 1000 10000 Misclassification Costs A v. % S ta n d ar d C o st Figure 11: Published results and those obtained by the implementation of ICET used in this paper 30 
work_qb53iyxcmvdwfeqedfthnek7xi	----	F2_BobilloEtAl DeLorean: A Reasoner for Fuzzy OWL 1.1 Fernando Bobillo, Miguel Delgado, and Juan Gómez-Romero Department of Computer Science and Artificial Intelligence, University of Granada C. Periodista Daniel Saucedo Aranda, 18071 Granada, Spain Phone: +34 958243194; Fax: +34 958243317 Email: fbobillo@decsai.ugr.es, mdelgado@ugr.es, jgomez@decsai.ugr.es Abstract. Classical ontologies are not suitable to represent imprecise or vague pieces of information, which has led to fuzzy extensions of De- scription Logics. In order to support an early acceptance of the OWL 1.1 ontology language, we present DeLorean, the first reasoner that supports a fuzzy extension of the Description Logic SROIQ, closely equivalent to it. It implements some interesting optimization techniques, whose usefulness is shown in a preliminary empirical evaluation. 1 Introduction Ontologies are a core element in the layered architecture of the Semantic Web. The current standard language for ontology representation is the Web Ontology Language (OWL). However, since its first development, several limitations on the expressiveness of OWL have been identified, and consequently several extensions to the language have been proposed. Among them, the most significant is OWL 1.1 [1] which is its most likely immediate successor. Description Logics (DLs for short) [2] are a family of logics for representing structured knowledge. They have proved to be very useful as ontology languages, and the DL SROIQ(D) is actually closely equivalent to OWL 1.1. It has been widely pointed out that classical ontologies are not appropriate to deal with imprecise and vague knowledge, which is inherent to several real- world domains. Since fuzzy logic is a suitable formalism to handle these types of knowledge, several fuzzy extensions of DLs have been proposed [3]. The broad acceptance of the forthcoming OWL 1.1 ontology language will largely depend on the availability of editors, reasoners, and other numerous tools that support the use of OWL 1.1 from a high-level/application perspective [4]. With this idea in mind, this work reports the implementation of DeLorean, the first reasoner that supports the fuzzy DL SROIQ. We also present a new optimization (handling superfluous elements before applying crisp reasoning) and a preliminary evaluation of the optimizations in the reduction. This paper is organized as follows. Section 2 describes the fuzzy DL SROIQ, which is equivalent to the fuzzy language supported by DeLorean. Then, Section 3 describes the reasoning algorithm, based on a reduction into crisp SROIQ. Section 4 presents our implementation, some of the implemented opti- mizations (including handling of superfluous elements), and a preliminary eval- uation. Finally, Section 6 sets out some conclusions and ideas for future work. 2 A Quick View to Fuzzy SROIQ In this section we recall the definition of fuzzy SROIQ [6], which extends SROIQ to the fuzzy case by letting concepts denote fuzzy sets of individuals and roles denote fuzzy binary relations. Axioms are also extended to the fuzzy case and some of them hold to a degree. We will assume a set of degrees of truth which are rational numbers of the form ! ! (0, 1], " ! [0, 1) and # ! [0, 1]. Moreover, we will assume a set of inequalities $% ! {", <, #, >}, ! ! {", <}, " ! {#, >}. For every operator $%, we define: (i) its symmetric operator $%!, defined as "!=#, >!=<, #!=", <!=>; (ii) its negation operator ¬ $%, defined as ¬ "=<, ¬ >=#, ¬ #=>, ¬ <=". Syntax. In fuzzy SROIQ we have three alphabets of symbols, for concepts (C), roles (R), and individuals (I). The set of roles is defined by RA $ U $ {R!|R ! RA}, where RA ! R, U is the universal role and R! is the inverse of R. Assume A ! C, R, S ! R (where S is a simple role [7]), oi ! I for 1 # i # m, m " 1, n " 0. Fuzzy concepts are defined inductively as follows: C, D % & | ' | A | C (D | C )D | ¬C | *R.C | +R.C | {!1/o1, . . . , !m/om} | (" m S.C) | (# n S.C) | +S.Self . Let a, b ! I. The axioms in a fuzzy Knowledge Base K are grouped in a fuzzy ABox A, a fuzzy TBox T , and a fuzzy RBox R1 as follows: ABox Concept assertion !a : C " !#, !a : C > "#, !a : C $ "#, !a : C < !# Role assertion !(a, b) : R " !#, !(a, b) : R > "#, !(a, b) : R $ "#, !(a, b) : R < !# Inequality assertion !a %= b# Equality assertion !a = b# TBox Fuzzy GCI !C & D " !#, !C & D > "# Concept equivalence C ' D, equivalent to {!C & D " 1#, !D & C " 1#} RBox Fuzzy RIA !R1R2 . . . Rn & R " !#, !R1R2 . . . Rn & R > "# Transitive role axiom trans(R) Disjoint role axiom dis(S1, S2) Reflexive role axiom ref (R) Irreflexive role axiom irr(S) Symmetric role axiom sym(R) Asymmetric role axiom asy(S) Semantics. A fuzzy interpretation I is a pair (&I, ·I), where &I is a non empty set (the interpretation domain) and ·I a fuzzy interpretation function mapping: (i) every individual a onto an element aI of &I; (ii) every concept C onto a function CI : &I % [0, 1]; (iii) every role R onto a function RI : &I , &I % [0, 1]. CI (resp. RI) denotes the membership function of the fuzzy concept C 1 The syntax of role axioms is restricted to guarantee the decidability of the logic [6]. (resp. fuzzy role R) w.r.t. I. CI(x) (resp. RI(x, y)) gives us the degree of being the individual x an element of the fuzzy concept C (resp. the degree of being (x, y) an element of the fuzzy role R) under the fuzzy interpretation I. Given a t-norm -, a t-conorm ., a negation function / and an implication function 0 [8], the interpretation is extended to complex concepts and roles as: &I(x) = 1 'I(x) = 0 (C ( D)I(x) = CI(x) - DI(x) (C ) D)I(x) = CI(x) . DI(x) (¬C)I(x) = /CI(x) (*R.C)I(x) = infy"!I{RI(x, y) 0 CI(y)} (+R.C)I(x) = supy"!I{RI(x, y) - CI(y)} {!1/o1, . . . , !m/om}I(x) = supi | x=oIi !i (" m S.C)I(x) = supy1,...,ym"!I [(- n i=1{S I(x, yi) - CI(yi)}) ! (-j<k{yj 1= yk})] (# n S.C)I(x) = infy1,...,yn+1"!I [(- n+1 i=1 {S I(x, yi) - CI(yi)}) 0 (.j<k{yj = yk})] (+S.Self )I(x) = SI(x, x) (R!)I(x, y) = RI(y, x) UI(x, y) = 1 For example, Z SROIQ uses Zadeh logic: minimum t-norm (! - " = min{!, "}), maximum t-conorm (! . " = max{!, "}), !Lukasiewicz negation (/! = 1 2 !), and Kleene-Dienes implication (! 0 " = max{1 2 !, "}). G SROIQ uses Gödel logic: minimum t-norm (! - " = min{!, "}), maxi- mum t-conorm (! . " = max{!, "}), Gödel negation (/! = 1 if ! = 0, or 0 otherwise) and Gödel implication (! 0 " = 1 if ! # ", or " otherwise) [8]. A fuzzy interpretation I satisfies (is a model of): – 3a : C $% #4 i" CI(aI) $% #, – 3(a, b) : R $% #4 i" RI(aI, bI) $% #, – 3a 1= b4 i" aI 1= bI, – 3a = b4 i" aI = bI, – 3C 5 D ! #4 i" infx"!I{CI(x) 0 DI(x)} ! #, – 3R1 . . . Rn 5 R ! #4 i" supx1...xn+1"!I ! [RI1 (x1, x2), . . . , R I n(xn, xn+1)] 0 RI(x1, xn+1) ! #, – trans(R) i" *x, y ! &I, RI(x, y) " supz"!I RI(x, z) - RI(z, y), – dis(S1, S2) i" *x, y ! &I, SI1 (x, y) = 0 or SI2 (x, y) = 0, – ref (R) i" *x ! &I, RI(x, x) = 1, – irr(S) i" *x ! &I, SI(x, x) = 0, – sym(R) i" *x, y ! &I, RI(x, y) = RI(y, x), – asy(S) i" *x, y ! &I, if SI(x, y) > 0 then SI(y, x) = 0, – a fuzzy KB K = 3A, T , R4 i" it satisfies each element in A, T and R. Irreflexive, transitive and symmetric role axioms are syntactic sugar for every R-implication (and consequently it can be assumed that they do not appear in fuzzy KBs) due to the following equivalences: irr(S) 6 3& 5 ¬+S.Self " 14, trans(R) 6 3RR 5 R " 14 and sym(R) 6 3R 5 R! " 14. In the rest of the paper we will only consider fuzzy KB satisfiability, since (as in the crisp case) most inference problems can be reduced to it [9]. 3 A Crisp Representation for Fuzzy SROIQ In this section we show how to reduce a fuzzy Z SROIQ fuzzy KB into a crisp KB (see [5, 6] for details). The procedure preserves reasoning, in such a way that existing SROIQ reasoners could be applied to the resulting KB. The basic idea is to create some new crisp concepts and roles, representing the !-cuts of the fuzzy concepts and roles, and to rely on them. Next, some new axioms are added to preserve their semantics, and finally every axiom in the ABox, the TBox and the RBox is represented using these new crisp elements. Adding new elements. Let AK and RK be the set of atomic concepts and roles occurring in a fuzzy KB K = 3A, T , R4. The set of the degrees which must be considered for any reasoning task is defined as NK = #, 1 2 # | 3' $% #4 ! K. Now, for each !, " ! NK with ! ! (0, 1] and " ! [0, 1), for each A ! AK and for each RA ! RK, two new atomic concepts A#", A># and two new atomic roles R#", R># are introduced. A#" represents the crisp set of individuals which are instance of A with degree higher or equal than ! i.e the !-cut of A. The semantics of these newly introduced atomic concepts and roles is pre- served by some terminological and role axioms. For each 1 # i # |NK| 2 1, 2 # j # |NK|21 and for each A ! AK, T (NK) is the smallest terminology containing these two axioms: A#$i+1 5 A>$i , A>$j 5 A#$j . Similarly, for each RA ! RK, R(NK) contains these axioms: R#$i+1 5 R>$i , R>$i 5 R#$i . Example 1. Consider the fuzzy KB K = {'}, where ' = 3StGenevieveTexasWhite : WhiteWine " 0.754}. We have that NK = {0, 0.25, 0.5, 0.75, 1} and T (NK) = {WhiteWine#0.25 5 WhiteWine>0, WhiteWine>0.25 5 WhiteWine#0.25, WhiteWi- ne#0.5 5 WhiteWine>0.25, WhiteWine>0.5 5 WhiteWine#0.5, WhiteWine#0.5 5 WhiteWine>0.25, WhiteWine>0.5 5 WhiteWine#0.5, WhiteWine#0.75 5 WhiteWi- ne>0.5, WhiteWine>0.75 5 WhiteWine#0.75, WhiteWine#1 5 WhiteWine>0.75}. () Mapping fuzzy concepts, roles and axioms. Concept and role expressions are reduced using mapping (, as shown in the first part of Table 1. Given a fuzzy concept C, ((C, " !) is a crisp set containing all the elements which belong to C with a degree greater or equal than ! (the other cases are similar). For instance, the 1-cut of the fuzzy concept *madeFromFruit.(NonSweetFruit ) SweetFruit) is ((*madeFromFruit.(NonSweetFruit ) SweetFruit), " 1) = *madeFromFruit>0.Non- SweetFruit#1 ) SweetFruit#1. Finally, we map the axioms in the ABox, TBox and RBox. Axioms are re- duced as shown in the second part of Table 1, where ) maps fuzzy axioms into crisp assertions, and * maps fuzzy TBox (resp. RBox) axioms into crisp TBox (resp. RBox) axioms. Recall that we are assuming that irreflexive, transitive and symmetric role axioms do not appear in the RBox. For example, assuming NK = {0, 0.25, 0.5, 0.75, 1}, the reduction of the fuzzy GCI 3Port 5 RedWine " 14 is *(3Port 5 RedWine " 14) = {Port>0 5 RedWine>0, Port#0.25 5 RedWine#0.25, Port>0.25 5 RedWine>0.25, Port#0.5 5 RedWine#0.5, Port>0.5 5 RedWine>0.5, Port#0.75 5 RedWine#0.75, Port>0.75 5 RedWine>0.75, Port#1 5 RedWine#1}. Table 1. Mapping of concept and role expressions, and reduction of the axioms. The semantics of &G uses Gödel implication, that of &KD uses Kleene-Dienes implication. Fuzzy concepts #((, !$) ( #((, "$) ) #(), !$) ) #(), "$) ( #(A, !$) A!! #(A, "$) ¬A¬"! #(¬C, %& $) #(C, %&! 1 * $) #(C + D, !$) #(C, !$) + #(D, !$) #(C + D, "$) #(C, "$) , #(D, "$) #(C , D, !$) #(C, !$) , #(D, !$) #(C , D, "$) #(C, "$) + #(D, "$) #(-R.C, !$) -#(R, !$).#(C, !$) #(-R.C, "$) .#(R, ¬" $).#(C, "$) #(.R.C, {", >}$) .#(R, {>, "}1 * $).#(C, {", >}$) #(.R.C, "$) -#(R, "!1 * $).#(C, "$) #({!1/o1, . . . , !m/om}, %& $) {oi | !i %& $, 1 $ i $ m} #(" m S.C, !$) " m #(S, !$).#(C, !$) #(" m S.C, "$) $ m*1 #(S, ¬ " $).#(C, ¬ " $) #($ n S.C, {", >} $) $ n #(S, {>, "} 1 * $).#(C, {>, "} 1 * $) #($ n S.C, "$) " n+1 #(S, "! 1 * $).#(C, "! 1 * $) #(-S.Self, !$) -#(S, !$).Self #(-S.Self, "$) ¬-#(S, ¬ " $).Self Fuzzy roles #(RA, !$) RA!! #(RA, "$) ¬RA¬"! #(R!, %& $) #(R, %& $)! #(U, !$) U #(U, "$) ¬U Axioms '(!a : C %& $#) {a : #(C, %& $)} '(!(a, b) : R %& $#) {(a, b) : #(R, %& $)} '(!a %= b#) {a %= b} '(!a = b#) {a = b} ((C &G D " !) ! !"N f K \{0} | !#"{#(C, " $) & #(D, " $)}! !"N f K | !<"{#(C, > $) & #(D, > $)} ((C &G D > ") ((C & D " ") / {#(C, > ") & #(D, > ")} ((C &KD D " !) {#(C, > 1 * !) & #(D, " !) } ((C &KD D > ") {#(C, " 1 * ") & #(D, > ") } ((!R1 . . . Rn &G R " !#) ! !"N f K \{0} | !#"{#(R1, " $) . . . #(Rn, " $) & #(R, " $)}! !"N f K | !<"{#(R1, > $) . . . #(Rn, > $) & #(R, > $)} ((!R1 . . . Rn &G R > "#) ((!R1 . . . Rn & R " "#) / {#(R1, > ") . . . #(Rn, > ") & #(R, > ")} ((!R1 . . . Rn &KD R " !#) {#(R1, > 1 * !) . . . #(Rn, > 1 * !) & #(R, " !)} ((!R1 . . . Rn &KD R > "#) {#(R1, " 1 * ") . . . #(Rn, " 1 * ") & #(R, > ")} ((dis(S1, S2)) {dis(#(S1, > 0), #(S2, > 0))} ((ref (R)) {ref (#(R, " 1))} ((asy(S)) {asy(#(S, > 0)} Properties. Summing up, a fuzzy KB K = 3A, T , R4 is reduced into a KB crisp(K) = 3)(A), T (NK) $ *(K, T ), R(NK) $ *(K, R)4. The following theorem shows that the reduction preserves reasoning: Theorem 1. A Z SROIQ fuzzy KB K is satisfiable i! crisp(K)) is [6]. The resulting KB is quadratic because it depends on the number of relevant degrees |NK|, or linear if we assume a fixed set. An interesting property is that the reduction of an ontology can be reused when adding a new axiom. If the new axioms does not introduce new atomic concepts, atomic roles nor a new degree of truth, we just need to add the reduction of the axiom. 4 DeLorean Reasoner This section describes the prototype implementation of our reasoner, which is called DeLorean (DEscription LOgic REasoner with vAgueNess). Initially, we developed a first version based on Jena API2 [6]. This version was developed in Java, using the parser generator JavaCC3, and DIG 1.1 in- terface [10] to communicate with crisp DL reasoners. An interesting property is the possibility of using any crisp reasoner thanks to the DIG interface. However, DIG interface does not yet support full SROIQ, so the logic supported by De- Lorean was restricted to Z SHOIN (OWL DL). From a historical point of view, this version was the first reasoner that supported a fuzzy extension of the OWL DL language. It implemented the reduction described in [11], and applied the optimization in the number of new elements and axioms described below. With the aim of augmenting the expressivity of the logic, in the current version we have changed the subjacent API to OWL API for OWL 24 [4]. Now, DeLorean supports both Z SROIQ(D) and G SROIQ(D), which correspond to fuzzy versions of OWL 1.1 under Zadeh and Gödel semantics, respectively. Since DIG interface does not currently allow the full expressivity of OWL 1.1, our solution was to integrate directly DeLorean with a concrete crisp ontology reasoner: Pellet [12], which can be directly used from the current version of the OWL API. This way, the user is free to choose to use either a generic crisp reasoner (restricting the expressivity to SHOIQ) or Pellet with no expressivity limitations. DeLorean is the first reasoner that supports a fuzzy extension of OWL 1.1. Figure 1 illustrates the architecture of the system: – The Parser reads an input file with a fuzzy ontology and translates it into an internal representation. The point here is that we can use any language to encode the fuzzy ontology, as long as the Parser can understand the representation and the reduction is properly implemented. Consequently we will not get into details of our particular choice. 2 http://jena.sourceforge.net/ 3 https://javacc.dev.java.net 4 http://owlapi.sourceforge.net Fig. 1. Architecture of DeLorean reasoner. In order to make the representation of fuzzy KBs easier, DeLorean also al- lows the possibility of importing OWL 1.1 ontologies. These (crisp) ontologies are saved into a text file which the user can edit and extend, for example adding membership degrees to the fuzzy axioms or specifying a particular fuzzy operator (Zadeh or Gödel family) for some complex concept. – The Reduction module implements the reduction procedures described in the previous section, building an OWL API model with an equivalent crisp on- tology which can be exported to an OWL file. The implementation also takes into account all the optimizations already discussed along this document. – The Inference module tests this ontology for consistency, using either Pel- let or any crisp reasoner through the DIG interface. Crisp reasoning does not take into account superfluous elements as we explain below. – Inputs (the path of the fuzzy ontology) and outputs (the result of the rea- soning and the elapsed time) are managed by an User interface. The reasoner implements the following optimizations: Optimizing the number of new elements and axioms. Previous works use two more atomic concepts A$# , A<" and some additional axioms A<$k 5 A$$k , A$$i 5 A<$i+1 , A#$k(A<$k 5 ', A>$i(A$$i 5 ', & 5 A#$k)A<$k , & 5 A>$i ) A$$i , 2 # k # |NK|. In [6] it is shown that they are unnecessary. Optimizing GCI reductions. In some particular cases, the reduction of fuzzy GCIs can be optimized [6]. For example, in range role axioms of the form 3& 5 *R.C " 14, domain role axioms of the form 3& 5 *R!.C " 14 and functional role axioms of the form 3& 5# 1 R.& " 14 we can use that *(3& 5 D $% #4) = & 5 ((D, $% #). Also, in disjoint concept axioms of the form 3C ( D 5 ' " 14, we can use that *(C 5 ' $% #) = ((C, > 0) 5 '. Furthermore, if the resulting TBox contains A 5 B, A 5 C and B 5 C, then A 5 C is unnecessary since it can be deduced from the other two axioms. Allowing crisp concepts and roles. Suppose that A is a fuzzy concept. Then, we need NK 2 1 concepts of the form A#" and another NK 2 1 concepts of the form A># to represent it, as well as 2 · (|NK| 2 1) 2 1 axioms to preserve their semantics. Fortunately, in real applications not all concepts and roles will be fuzzy. If A is declared to be crisp, we just need one concept to represent it and no new axioms. The case for fuzzy roles is exactly the same. Of course, this optimization requires some manual intervention. Reasoning ignoring superfluous elements. Our reduction is designed to promote reusing. For instance, consider the fuzzy KB K in Example 1. The reduc- tion of K contains )(' ) = StGenevieveTexasWhite : WhiteWine#0.75, but also the axioms in T (NK). It can be seen that the concepts WhiteWine>0, WhiteWine#0.25, WhiteWine>0.25, WhiteWine#0.5, WhiteWine>0.5, WhiteWine>0.75, WhiteWine#1 are superfluous in the sense that cannot cause a contradiction. Hence, for a satisfia- bility test of crisp(K), we can avoid the axioms in T (NK) where they appear. But please note that if additional axioms are added to K, crisp(K) will be di"erent and previous superfluous concept and roles may not be superfluous any more. For example, if we want to check if K $ 3StGenevieveTexasWhite : WhiteWine " 0.54 is satisfiable, then the concept WhiteWine#0.5 is no longer superfluous. In this case, it is enough to consider T %(NK) = {WhiteWine#0.75 5 WhiteWine#0.5}. The case of atomic roles is similar to that of atomic concepts. 5 Use Case: A Fuzzy Wine Ontology This section considers a concrete use case, a fuzzy extension of the well-known Wine ontology5, a highly expressive ontology (in SHOIN (D)). Some metrics of the ontology are shown in the first column of Table 2. In an empirical evaluation of the reductions of fuzzy DLs to crisp DLs, P. Cimiano et al. wrote that “the Wine ontology showed to be completely intractable both with the optimized and unoptimized reduction even using only 3 degrees” [13]. They only considered there what we have called here “optimization of the number of new elements and axioms”. We will show that the rest of the optimizations, specially the (natural) assumption that there are some crisp elements, reduce significantly the number of axioms, even if tractability of the reasoning is to be verified. A fuzzy extension of the ontology. We have defined a fuzzy version of the Wine ontology by adding a degree to the axioms. Given a variable set of degrees NK, the degrees of the truth for fuzzy assertions is randomly chosen in NK. In the case of fuzzy GCIs and RIAs, the degree is always 1 in special GCIs (namely concept equivalences and disjointness, domain, range and functional role axioms) or if there is a crisp element in the left side; otherwise, the degree is 0.5. In most of the times fuzzy assertions are of the form 3' ! "4 with " 1= 1. Clearly, this favors the use of elements of the forms C!# and R!# , reducing the number of superfluous concepts. Once again, we are in the worst case from the 5 http://www.w3.org/TR/2003/CR-owl-guide-20030818/wine.rdf point of view of the size of the resulting crisp ontology. Nonetheless, in practice we will be often able to say that an individual fully belongs to a fuzzy concept, or that two individuals are fully related by means of a fuzzy role. Crisp concepts and roles. A careful analysis of the fuzzy KB brings about that most of the concepts and the roles should indeed be interpreted as crisp. For example, most of the subclasses of the class Wine refer to a well-defined geographical origin of the wines. For instance, Alsatian wine is a wine which has been produced in the French region of Alsace: AlsatianWine 6 Wine ( +locatedAt.{alsaceRegion}. In other applications there could exist examples of fuzzy regions, but this is not our case. Another important number of sub- classes of Wine refer to the type of grape used, which is also a crisp concept. For instance, Riesling is a wine which has been produced from Riesling grapes: Riesling 6 Wine ( +madeFromGrape.{RieslingGrape}( " 1 madeFromGrape.&. Clearly, there are other concepts with no sharp boundaries (for instance, those derived from the vague terms “dry”, “sweet”, “white” or“heavy”). The re- sult of our study has identified 50 fuzzy concepts in the Wine ontology, namely: WineColor, RedWine, RoseWine, WhiteWine, RedBordeaux, RedBurgundy, RedTa- bleWine, WhiteBordeaux, WhiteBurgundy, WhiteLoire, WhiteTableWine, WineSugar, SweetWine, SweetRiesling, WhiteNonSweetWine, DryWine, DryRedWine, DryRies- ling, DryWhiteWine, WineBody, FullBodiedWine, WineFlavor, WineTaste, LateHar- vest, EarlyHarvest, NonSpicyRedMeat, NonSpicyRedMeatCourse, SpicyRedMeat, PastaWithSpicyRedSauce, PastaWithSpicyRedSauceCourse, PastaWithNonSpicyRed- Sauce, PastaWithNonSpicyRedSauceCourse, SpicyRedMeatCourse, SweetFruit, Sweet- FruitCourse, SweetDessert, SweetDessertCourse, NonSweetFruit, NonSweetFruit- Course, RedMeat, NonRedMeat, RedMeatCourse, NonRedMeatCourse, PastaWith- HeavyCreamSauce, PastaWithLightCreamSauce, Dessert, CheeseNutsDessert, De- ssertCourse, CheeseNutsDessertCourse, DessertWine. Furthermore, we identified 5 fuzzy roles: hasColor, hasSugar, hasBody, hasFla- vor, and hasWineDescriptor (which is a super-role of the other four). Measuring the importance of the optimizations. The reduction under Gödel semantics is still to be published [14], so we focus our experimentation in Z SROIQ (omitting the concrete role yearValue), but allowing the use of both Kleene-Dienes and Gödel implications in the semantics of fuzzy GCIs and RIAs. Table 2 shows some metrics of the crisp ontologies obtained in the reduction of the fuzzy ontology after applying di"erent optimizations. 1. Column “Original” shows some metrics of the original ontology. 2. “None” considers the reduction obtained after applying no optimizations. 3. “(NEW)” considers the reduction obtained after optimizing the number of new elements and axioms. 4. “(GCI)” considers the reduction obtained after optimizing GCI reductions. 5. “(C/S)” considers the reduction obtained after allowing crisp concepts and roles and ignoring superfluous elements. 6. Finally, “All” applies all the previous optimizations. Table 2. Metrics of the Wine ontology and its fuzzy versions using 5 degrees Original None (NEW) (GCI) (C/S) All Individuals 206 206 206 206 206 206 Named concepts 136 2176 486 2176 800 191 Abstract roles 16 128 128 128 51 20 Concept assertions 194 194 194 194 194 194 Role assertions 246 246 246 246 246 246 Inequality assertions 3 3 3 3 3 3 Equality assertions 0 0 0 0 0 0 New GCIs 0 4352 952 4352 1686 324 Subclass axioms 275 1288 1288 931 390 390 Concept equivalences 87 696 696 696 318 318 Disjoint concepts 19 152 152 19 152 19 Domain role axioms 13 104 104 97 104 97 Range role axioms 10 80 80 10 80 10 Functional role axioms 6 48 48 6 48 6 New RIAs 0 136 119 136 34 34 Sub-role axioms 5 40 40 40 33 33 Role equivalences 0 0 0 0 0 0 Inverse role axioms 2 16 16 16 2 2 Transitive role axioms 1 8 8 8 1 1 We have put together the optimizations of crisp and superfluous elements because in this ontology handling superfluous concepts is not always useful, due to the existence of a lot of concept definitions, as we will see in the next example. Example 2. Consider the fuzzy concept NonRedMeat. Firstly, this concept ap- pears as part of a fuzzy assertion stating that pork is a non read meat: )(3Pork : NonRedMeat ! !14) = Pork : NonRedMeat!"1 . Secondly, non read meat is de- clared to be disjoint from read meat: *(3RedMeat ( NonRedMeat 5 ' " 14) = RedMeat>0 ( NonRedMeat>0 5 '. Thirdly, non read meat is a kind of meat: *(3NonRedMeat 5 Meat " !24) = NonRedMeat>0 5 Meat. If these were the only occurrences of NonRedMeat, then the reduction would create only two non- superfluous crisp concepts, namely NonRedMeat>0 and NonRedMeat!"1 , and in order to preserve the semantics of them we would need to add just one axiom during the reduction: NonRedMeat!"1 5 NonRedMeat>0. However, this is not true because NonRedMeat appears in the definition of the fuzzy concept NonRedMeatCourse. In fact, *(NonRedMeatCourse 6 MealCourse ( *hasFood.NonRedMeat) introduces non-superfluous crisp concepts for the rest of the degrees in NK. Consequently, for each 1 # i # |NK| 2 1, 2 # j # |NK|21, the reduction adds to T (NK) the following axioms: NonRedMeat#$i+1 5 NonRedMeat>$i ; NonRedMeat>$j 5 NonRedMeat#$j . () Note that the size of the ABox is always the same, because every axiom in the fuzzy ABox generates exactly one axiom in the reduced ontology. The number of new GCIs and RIAs added to preserve the semantics of the new elements is much smaller in the optimized versions. In particular, we reduce from 4352 to 324 GCIs (7.44%) and from 136 to 34 RIAs (25%). This shows the importance of reducing the number of new crisp elements and their corresponding axioms, as well as of defining crisp concepts and roles and (to a lesser extent) handling superfluous concepts. Optimizing GCI reductions turns out to be very useful in reducing the num- ber of disjoint concepts, domain, range and functional role axioms: 152 to 19 (12.5 %), 104 to 97 (93.27 %), 80 to 10 (12.5 %), and 48 to 6 (12.5 %), respec- tively. In the case of domain role axioms the reduction is not very high because we need an inverse role to be defined in order to apply the reduction, and this happens only in one of the axioms. Every fuzzy GCI or RIA generates several axioms in the reduced ontology. Combining the optimization of GCI reductions with the definition of crisp con- cepts and roles reduces the number of new axioms, from 1288 to 390 subclass axioms (30.28 %), from 696 to 318 concept equivalences (45.69 %) and from 40 to 33 sub-role axioms (82.5 %). Finally, the number of inverse and transitive role axioms is reduced in the optimized version because fuzzy roles interpreted as crisp introduce one inverse or transitive axiom instead of several ones. This allows a reduction from 16 to 2 axioms, and from 8 to 1, respectively, which corresponds to the 12.5 %. Table 3 shows the influence of the number of degrees on the size of the resulting crisp ontology, as well as on the reduction time (which is shown in seconds), when all the described optimizations are used. The reduction time is small enough to allow to recompute the reduction of an ontology when necessary, thus allowing superfluous concepts and roles in the reduction to be avoided. Table 3. Influence of the number of degrees in the reduction. Crisp 3 5 7 9 11 21 Number of axioms 811 1166 1674 2182 2690 3198 5738 Reduction time - 0.343 0.453 0.64 0.782 0.859 1.75 6 Conclusions and Future Work This paper has presented DeLorean, the more expressive fuzzy DL reasoner that we are aware of (it supports fuzzy OWL 1.1), and the optimizations that it implements. Among them, the current version enables the definition of crisp concepts and roles, as well as handling superfluous concepts and roles before applying crisp reasoning. A preliminary evaluation shows that these optimiza- tions help to reduce significantly the size of the resulting ontology. In future work we plan to develop a more detailed benchmark by relying on the hyper- tableau reasoner HermiT, which seems to outperform other DL reasoners [15], and, eventually, to compare it against other fuzzy DL reasoners. Acknowledgements This research has been partially supported by the project TIN2006-15041-C04-01 (Mi- nisterio de Educación y Ciencia). Fernando Bobillo holds a FPU scholarship from Minis- terio de Educación y Ciencia. Juan Gómez-Romero holds a scholarship from Consejeŕıa de Innovación, Ciencia y Empresa (Junta de Andalućıa). References 1. Patel-Schneider, P.F., Horrocks, I.: OWL 1.1 Web Ontology Language overview (2006) [Online] Available: http://www.w3.org/Submission/owl11-overview/. 2. Baader, F., Calvanese, D., McGuinness, D., Nardi, D., Patel-Schneider, P.F.: The Description Logic Handbook: Theory, Implementation, and Applications. Cam- bridge University Press (2003) 3. Lukasiewicz, T., Straccia, U.: Managing uncertainty and vagueness in Description Logics for the Semantic Web. Journal of Web Semantics (To appear) 4. Horridge, M., Bechhofer, S., Noppens, O.: Igniting the OWL 1.1 touch paper: The OWL API. In: Proc. of the 3rd International Workshop on OWL: Experiences and Directions (OWLED 2007). Volume 258., CEUR Workshop Proceedings (2007) 5. Straccia, U.: Transforming fuzzy Description Logics into classical Description Log- ics. In: Proceedings of the 9th European Conference on Logics in Artificial Intelli- gence (JELIA-04). Volume 3229 of Lecture Notes in Computer Science., Springer- Verlag (2004) 385–399 6. Bobillo, F., Delgado, M., Gómez-Romero, J.: Optimizing the crisp representation of the fuzzy Description Logic SROIQ. In: Proceedings of the 3rd ISWC Work- shop on Uncertainty Reasoning for the Semantic Web (URSW 2007). Volume 327., CEUR Workshop Proceedings (2007) 7. Horrocks, I., Kutz, O., Sattler, U.: The even more irresistible SROIQ. In: Pro- ceedings of the 10th International Conference of Knowledge Representation and Reasoning (KR 2006). (2006) 452–457 8. Hájek, P.: Metamathematics of Fuzzy Logic. Kluwer (1998) 9. Straccia, U.: Reasoning within fuzzy Description Logics. Journal of Artificial Intelligence Research 14 (2001) 137–166 10. Bechhofer, S., Möller, R., Crowther, P.: The DIG Description Logic interface: DIG / 1.1. In: Proc. of the 16th Int. Workshop on Description Logics (DL 2003). (2003) 11. Bobillo, F., Delgado, M., Gómez-Romero, J.: A crisp representation for fuzzy SHOIN with fuzzy nominals and General Concept Inclusions. In: Proceedings of the 2nd ISWC Workshop on Uncertainty Reasoning for the Semantic Web (URSW 2006). Volume 218., CEUR Workshop Proceedings (2006) 12. Sirin, E., Parsia, B., Cuenca-Grau, B., Kalyanpur, A., Katz, Y.: Pellet: A practical OWL-DL reasoner. Journal of Web Semantics 5(2) (2007) 51–53 13. Cimiano, P., Haase, P., Ji, Q., Mailis, T., Stamou, G.B., Stoilos, G., Tran, T., Tzouvaras, V.: Reasoning with large A-Boxes in fuzzy Description Logics using DL reasoners: An experimental valuation. In: Proceedings of ARea2008. Volume 350., CEUR Workshop Proceedings (2008) 14. Bobillo, F., Delgado, M., Gómez-Romero, J., Straccia, U.: Fuzzy Description Logics under Gödel semantics. (Submitted) 15. Motik, B., Shearer, R., Horrocks, I.: Optimized reasoning in Description Logics using hypertableaux. In: Proc. of the 21st International Conference on Automated Deduction (CADE-21). Lecture Notes in Artificial Intelligence 4603 (2007) 67–83 
work_qbzexnx22zfibibhrvs5slvbk4	----	JASIS 9/23/92 - reformat An Expert System for Automatic Query Reformulation Susan Gauch* and John B. Smith+ *To whom correspondence should be sent. Biological Knowledge Laboratory College of Computer Science Northeastern University Boston, MA 02115 sgauch@flora.ccs.northeastern.edu (617) 437-3850 +Department of Computer Science University of North Carolina Chapel Hill, NC 27599-3175 This work was supported in part by ONR contract N00014-86-K-0680. ABSTRACT Unfamiliarity with search tactics creates difficulties for many users of online retrieval systems. User observations indicate that even experienced searchers use vocabulary incorrectly and rarely reformulate their queries. To address these problems, an expert system for online search assistance was developed. This prototype automatically reformulates queries to improve the search results, and ranks the retrieved passages to speed the identification of relevant information. Users' search performance using the expert system was compared with their search performance on their own, and their search performance using an online thesaurus. The following conclusions were reached: 1) The expert system significantly reduced the number of queries necessary to find relevant passages compared with the user searching alone or with the thesaurus. 2) The expert system produced marginally significant improvements in precision compared with the user searching on their own. There was no significant difference in the recall achieved by the three system configurations. 3) Overall, the expert system ranked relevant passages above irrelevant passages. 1 INTRODUCTION 1.1 Driving Problem Technology to produce, store, and distribute massive quantities of electronic information has matured. Textbases, online full-text databases, are being created in Gauch -2- many fields. Personal workstations have become common. To make use of the information accessible from their desks, technology must be developed which allows end-users to search effectively. One user study found that whereas system mechanics are rarely a problem for any but very inexperienced and infrequent users, even experienced searchers have significant problems with search strategy and output performance (Borgman, 1986a). Another found experienced searchers lost sight of the search logic, missed obvious synonyms, and searched too simply (Fenichel, 1981). In spite of low recall, half of the searchers never modified the original query in an attempt to improve their results. Studies of inexperienced searchers find even more problems with search strategy. In one study, a quarter of the subjects were unable to pass a benchmark test of minimum searching skill (Borgman, 1986b). In an experiment contrasting the searching of novices versus experienced searchers, the novices found some relevant documents easily, but they failed to achieve high recall and were unable to reformulate queries well (Oldroyd, 1984). The experienced searchers in this study were more persistent and willing to experiment than the novices. Blair and Maron (1985) paint an even bleaker picture for searching full-text databases. Legal assistants searching a legal database achieved only 20% recall, although they were attempting to do a high recall search. The factors, as identified by the authors, leading to this poor performance were poor searching technique (failure to use stemming and synonyms), stopping the query iteration too soon, and the inability to search on inter-document relationships. The authors argued that vocabulary problems make high recall impossible on full-text databases. 1.2 Related Work In our system, queries are reformulated by a knowledge-based online search assistant acting as the front-end to existing retrieval systems. Research in this area is summarized in Section 1.2.1. For a more detailed discussion, see (Gauch, 1992). The knowledge base for our system is be built on existing searching practice. Current knowledge on good search technique is presented in Section 1.2.2. 1.2.1 Expert Systems CANSEARCH (Pollitt, 1984; Pollitt, 1987) is one of the earliest expert systems for bibliographic retrieval. It is designed to enable doctors to search the MEDLINE medical database for cancer literature. The expert system contains knowledge of a single domain, cancer, rather than search strategies in general. During the query reformulation process, the expert system guides the searcher through a hierarchy of menus. IR-NLI II (Brajnik, Guida, & Tasso, 1988) incorporates user modeling into a domain- independent bibliographic retrieval expert system. Domain knowledge is supplied separately by an online thesaurus. A user model is built based on the user's amount of domain knowledge and search experience. This model is used to tailor the Gauch -3- dialogue between the system and the user. Initially, the user lists some terms which describe his interests. The expert system, through a lengthy dialogue, clarifies its model of the query, proposes terms to expand the query, and comments on the user's search strategy. No automatic query reformulation is done. Shoval (1985) developed an expert system to assist users in selecting the right vocabulary terms for a database search. The knowledge base of words, concepts, and phrases and their semantic relationships is stored in a semantic network. Decision rules, based on common search practice, are used to locate candidate vocabulary terms in the semantic network and suggest them to the user for possible query expansion. The user then decides whether or not the candidate terms are relevant and should be used to replace the terms in the nodes which point to it. PLEXUS (Vickery & Brooks, 1987) is an expert system to help novice users find information about gardening. The initial query formation consists of a dialogue with the user. Natural language queries are accepted, and information is extracted to fill in frames. The system has a knowledge base of search strategies and term classifications similar to a thesaurus. Most of the domain knowledge is in the classification, but some appears in the rule base, itself. If queries are too broad (defined as more than 10 references) , no narrowing is attempted. The references are displayed 5 at a time to the user. If the query is too narrow (defined as nothing retrieved at all), three strategies are attempted: 1) if two or more terms appear in the same subcategory, OR them together rather than AND; 2) drop one of the terms; 3) replace a term by its parent. IOTA (Chiaramella & Defude, 1987) is an expert system which incorporates a natural language interface. Passages are retrieved from an online book based on keywords which index each passage. Much of the research effort has gone into processing the user's queries, but some simple query reformulation is also done. Specifically, queries are broadened by replacing a term by its parent from an online thesaurus and narrowed by removing O R terms. Their results, based on a small-scale experiment, indicate an increase in precision and recall using the expert system. I3R (Croft & Thompson, 1987) incorporates user modeling and relevance feedback. The query formation process is a dialogue between the user and the system, during which the user supplies a short natural language query or an initial relevant document. The domain knowledge expert infers related concepts from to the query and presents them to the user for confirmation. If the thesaurus-like knowledge base does not contain related information and the initial query contained too many high-frequency terms the user may be asked to provide additional keywords. A ranked list of documents is presented to the user. The user then indicates which terms in each document are interesting. These new terms may be used to modify the query. EP-X (Krawczak et al, 1987; Smith et al, 1989) is a prototype knowledge-based system that assists users in conducting bibliographic searches of the environmental Gauch -4- pollution literature. This system makes extensive use of domain knowledge, represented as hierarchically-defined semantic primitives and frames. The user enters a query as a list of keywords and the system interacts with him to suggest possible broadening or narrowing operations. In spite of the rich domain knowledge, the final search results were mediocre. In particular, users did not take advantage of the available query refinement strategies when they should have. Fewer projects have attempted to provide intelligent assistance for full-text searching. One such system is RUBRIC (McCune et al, 1985; Tong et al, 1987), which has the user describe his query in terms of rules. These rules describe the domain knowledge for the system as a hierarchy of topics and subtopics. Rules may have weights representing the certainty and/or importance of the defined relationships. The lowest level subtopics define patterns in the text which indicate the presence of that subtopic. Although the query process is very powerful, it places a heavy burden on the user. Ongoing projects are developing models on which future expert systems will be based (Belkin & Marchetti, 1990; Chen, 1990). Recent projects focus on incorporating other aspects of artificial intelligence, particularly natural language processing (Jacobs & Rau, 1990) and probabilistic inference over networks of documents (Croft & Turtle, 1992). 1.2.2 Search Strategies The automatic query reformulation incorporated in the systems described in the previous section are, in general, very primitive. However, search strategies employed by both novice and experienced searchers have been widely studied. These studies formed the basis of our expert system's searching knowledge base, which is described in detail in Section 3. Searching Studies Bates (1979) has compiled a thorough catalogue of search tactics. She outlines 29 search tactics in four areas: monitoring, file structure, search formulation, and term manipulation. The tactics for search formulation and term manipulation describe the available techniques to broaden and narrow queries. The search formulation tactics include the selection of appropriate initial search terms and the manipulation of query structure; the term manipulation tactics describe the use of context, thesaural terms, and stemming to modify queries. The tactics she lists provide the basic operations for our expert system; however, she includes no guideline as to when each tactic is appropriate. By analyzing discourses between an expert intermediary and 17 real information seekers, Smith et al (1989) identify a set of search tactics. They noted when each of these tactics was applied, and whether the intermediary used the tactic spontaneously or in response to some cue in the retrieved document. They concluded that the intermediary makes extensive use of domain knowledge to Gauch -5- suggest topic refinesments, generates most knowledge-based suggestions spontaneously, and rarely changes logical opeartors. The results of this study are being used as the basis for EP-X, an online search intermediary (see section 1.2.1). Based on observation of 47 professional online searchers, Fidel has developed a formal decision tree that represents the intuitive rules searchers use when they select search terms. The options available to the searchers are enumerated, as are the conditions under which each option is selected. The author defines rules for deciding when to use textwords or descriptors or both to search indexed databases, and guidelines for including thesaural relationships. In contrast to the tactics proposed by Smith et al (1989), the emphasis is on identifying techniques which are domain independent. Effects of Query Expansion A study of the effects of query expansion on retrieval performance found that automatically adding terms based on their statistical relationships to the user's search terms degrades retrieval performance (Smeaton & van Rijsbergen, 1983). Clearly, a better criteria for selecting terms to add is needed. Another study (Harman, 1988) also showed performance degradation when adding terms from a statistically constructed thesaurus. However, when only those thesaural terms which occur in relevant documents are added, retrieval performance improves over that achieved by the original query. However, the best performance is achieved when user filtering of three types of candidate terms (thesaural, term variants, and statistically selected from relevant documents) is simulated. 2 SYSTEM ARCHITECTURE The prototype system consists of five major components (see Figure 1): 1) MICROARRAS (Smith, Weiss, & Ferguson, 1987), which serves as the full- text search and retrieval engine 2) a full-text database of over 188,000 words 3) a hierarchical thesaurus of approximately 7,424 words specific to the textbase's domain 4) an expert system of 85 OPS83 rules and over 5,000 lines of C code, which interprets the user's queries, controls the search process, analyzes the retrieved text, and ranks the search results 5) a user interface, which accepts the user's queries, presents requests for information from the expert system, and displays the search results. Gauch -6- User Interface Expert System MICRO- ARRAS TextbaseThesaurus User Figure 1. System Architecture The system is implemented on a Sun 3 workstation. MICROARRAS and the thesaurus construction and access routines are written in the C language. The expert system consists of a knowledge base of production rules, written in OPS83, and a set of C language functions to carry out the actions prescribed by the rule-base. The textual database for the current demonstration project consists of an unpublished manuscript on computer architecture written by Gerrit A. Blaauw and Frederick P. Brooks, Jr. (1986). The search process consists of a dialogue between the user and the expert system. The user enters the initial Boolean query and the number of passages (i.e. paragraphs) he would like to retrieve. The expert system parses the query and translates it into a request for information from MICROARRAS. MICROARRAS retrieves text passages from the full-text database and informs the expert system of the number of passages that satisfy the request. The expert system compares the number retrieved with the target number to decide whether or not to reformulate the query, and, if so, how. Once the target number has been reached, or the expert system has run out of reformulations to try, the retrieved passages are presented to the user in rank-order. A major advantage of this architecture is the separation of strategic knowledge, contained in the knowledge base for the expert system, from domain knowledge, contained in the thesaurus. Now that the search strategy rules have been developed and tested with the existing textbase, the expert system can be tested with other content domains by simply providing a suitable thesaurus for the new textbase. 2.1 MICROARRAS MICROARRAS is a full-text retrieval and analysis system. The system provides immediate access to any passage in the textbase, regardless of the length of that document. Contexts for searches can be indicated in terms of words, sentences, paragraphs, etc., for the entire search expression or for different parts of it. To be inserted into MICROARRAS' textbase, documents must first be inverted (i.e. a dictionary is created with an entry for each word in the text. Each entry contains the word and the numerical position in the text of each occurrence of that word). However, they require no semantic preprocessing. Once stored in the textbase, they can be examined individually or in groups. They can also be moved from one Gauch -7- textbase to another. Thus, documents can be processed on a workstation or microcomputer, uploaded into a textbase on a mainframe or textbase server, searched and analyzed there, or downloaded for local use once again. 2.2 Textbase The textbase contains the Fall, 1986 draft of Computer Architecture, Volume 1 - Design Decisions by Blaauw and Brooks. The manuscript consists of 188,278 words comprising 8 chapters, titled: "Introduction", "Machine Language", "Addresses", "Data", "Operations", "Instruction Sequence", "Supervision", and "Input/Output". TeX format marks were already present and were used to display of the retrieved text (line, italics, label), as well as provide structural information (chapter, section, subsection, subsubsection, paragraph, sentence, item). 2.3 Thesaurus All domain-specific knowledge is contained in a hierarchical thesaurus. The expert system uses this information to reformulate queries. The thesaurus was built by the author from the Brooks and Blaauw text, and it strongly reflects the word usage of that textbase. In general, it should not be necessary to provide a unique thesaurus for each textbase. An existing thesaurus for the domain could be used, as long as there is a good match between thesaurus classes and textbase word usage. Word types which share a common stem are grouped into s t e m g r o u p s . The members of a given stemgroup are called stemwords. Each word type in the Blaauw and Brooks text appears in exactly one stemgroup. Thesaurus classes contain stemgroups which are synonyms for each other. Stemgroups may appear in zero, one, or more than one thesaurus class. Because the thesaurus classes are linked together with parent-child links, they are also referred to as n o d e s . The arrangement of the words into stemgroups, stemgroups to thesaurus classes, and the classes into a hierarchy is discussed in (Gauch, 1991). To capture the relationships between thesaurus classes for use by the expert system and the user, the high- frequency terms are included in the thesaurus. 2.4 Query Language A Boolean query language was chosen because it is the most common type available on existing systems. The operators provided, in decreasing order of operator precedence, are: ANDNOT, AND, and OR. A default context of one sentence is used for the A N D and A N D N O T operators. When a query is parsed, the expert system interprets each search term to represent a unique concept. The concepts, and the operators, are flagged as positive or negative based on whether they are specifying information the user does, or does not, wish to receive. For example, the query 'i/o A N D N O T (device O R interrupt)' contains three concepts: i/o, device, and interrupt. I/O is a concept on which the user wishes information, so it is considered a positive concept. Device and interrupt indicate concepts on which the user does not wish information, so they are considered negative concepts. The ANDNOT and O R Gauch -8- operators are followed by negative concepts, so they too are flagged as negative. When the user is searching with the expert system, the expert system controls the context. Initially, the default of one sentence is used, but the expert system may adjust the context during query reformulation. However, when the user is searching without the expert system, the A N D and A N D N O T operators may be augmented with a user-specified context. The user may define the search context for AND or ANDNOT in terms of words, sentences, and/or paragraphs. 2.5 Knowledge Base The expert system performs three main functions: 1) it controls the operation of the system as a whole; 2) it reformulates the Boolean query based on previous search results; 3) it ranks the retrieved passages in decreasing order of estimated relevance for presentation to the user. To perform these functions the expert system contains a knowledge base of the search process, search strategies, and passage ranking procedures. Domain knowledge is contained in the hierarchically structured thesaurus. The system has no knowledge of the user's true information needs, other than the target number they specify to indicate how many passages they wish to retrieve. 2.5.1 Query Reformulation Queries are reformulated based on the target number, the number of passages retrieved, and the history of broadening and narrowing techniques already applied. The expert system has a collection of reformulation tactics at its disposal. Bates (1979) and others have identified successful search tactics and Fidel (1991) discusses when to use free-text terms versus descriptors. However, no one has outlined an overall query reformulation strategy for free-text searching. The guiding principles for the expert system's query reformulation knowledge base were: 1) each search term in the initial query represents one concept on which the user does, or explicitly does not, want information; 2) the user's initial search terms are the best indication of the user's areas of interest; 3) some terms from the thesaurus may be helpful, but others will not; 4) the expert system should never discard concepts in which the user has indicated an interest. Query Reformulation Techniques The expert system reformulates queries using three different techniques: 1) expanding concepts; 2) adjusting context; and 3) changing the query structure. Gauch -9- Expanding Concepts To broaden a query, search terms are added to the positive concepts, whereas narrowing a query adds search terms to negative concepts. Concepts may be expanded by stemming, adding synonyms, and adding related search terms from the thesaurus. The order in which the terms are added from the thesaurus is: parents, then siblings, then children. Replacing a term with its parent to broaden a query is a common practice, both by searchers (Bates, 1979; Salton, 1986), and in systems which automatically reformulate queries (Chiaramella & Defude, 1987; Vickery & Brooks, 1987). The rationale, based on experiences with keyworded bibliographic systems, is that since parent terms represent broader concepts, adding the parent term should broaden the scope of the query. In full-text databases, we believe that the reverse order may make more sense. Broadening a concept containing virtual memory with children terms, yielding 'virtual_memory OR paging OR segmentation', seems more likely to retrieve relevant passages than broadening with the parent terms, yielding 'virtual_memory OR memory'. Crouch (1988) found that augmenting a query with thesaurus terms, rather than replacing the original search terms, lead to improved results. With this in mind, concepts are expanded by adding thesaural terms (O R ing them with the terms already in the concept) rather than by replacing the terms already present. The belief that some stemgroups from the thesaurus will be useful, while other will not, is the basis for providing user filtering of the candidate thesaurus terms. One expert system uses domain-dependent search strategies to choose the appropriate terms from a thesaurus (Smith et al, 1989). In addition, Harman (1988) showed that search results improved when thesaural terms were filtered by the user. Based on these two studies, we decided to allow the users to select which stemgroups to add from a set of thesaural candidates. Finally, candidate search terms selected from the thesaurus are filtered to remove those which already occur in the query and extremely high frequency terms. The remaining terms are added one at a time, in reverse order of frequency, and the new number of retrieved passages is compared to the target number. Adjusting Context The expert system manipulates four different contexts; it adjusts the distance between words in positive and negative multiword phrases as well as the distance between positive and negative search concepts. The expert system broadens queries by increasing the positive contexts and decreasing the negative ones. Narrowing is done by decreasing the positive contexts and increasing the negative ones. Changing Query Structure The final variable the expert system can manipulate to reformulate the query structure. The query can be broadened in two different ways: first, the positive AND Gauch -10- -ve synonyms decrease context -ve parents -ve siblings -ve children decrease context +ve synonyms increase context +ve parents +ve siblings +ve children loosen operators Initial Query remove -ve -ve stemwords SuccessSuccess +ve stemwords increase context B N B B N B N N S S S S B B N N S S B B B N N N B B B S S S N N N S S S B B S N N N B B N N S S B B S N Legend B: broaden N: narrow S: success +ve: positive concepts -ve: negative concepts S tighten operators Failureincreasecontext decrease context S N B N SB Figure 2. Query Reformulation Techniques Gauch -11- operators can be switched to OR operators (and the negative OR operators switched to ANDs); second, the negative parts of the query can be dropped altogether. All of the AND operators are replaced at the same time. A better strategy would be to replace them one at a time, in inverse order of the frequency of occurrence of the concepts. Similarly, the query can be narrowed by replacing OR operators with A N Ds. The expert system does not have enough information about the user's information needs to decide which positive parts of the query to drop, so this technique is not employed to narrow queries. Because manipulating query structure causes major changes to the user's original query, these techniques are only tried as a last resort. It is not likely that the new query will find passages that the user will find highly relevant, but the goal is to find somewhat relevant passages that users can read in order to reformulate their own queries and try again. Flow of Control Figure 2 diagrams the flow of control among the reformulation techniques. The left side of the Figure 2 diagrams the broadening techniques, the right side the narrowing techniques. This figure is somewhat simplified since it does not show the use of context to converge to the target number once queries have been found which bracket the target number from above and below. The expert system records the type of initial query reformulation as the global objective. If the reformulations in the original direction overshoot the target number without achieving success, reformulations in the opposite, or local, direction are tried, beginning at the top node on that side of the diagram. Reformulation never continues in the local direction farther than it reached in the global direction. At this point, queries have already been formed which bracket the target number from below and above, otherwise the system would not have tried both narrowing and broadening techniques. Rather than using techniques which are considered less likely to produce good results, the expert system adjusts the context. Stopping Knowing when to stop a search is a difficult problem. We partially side-step this problem by having the user explicitly state the number of passages he wishes to retrieve. Since the target number he supplies is likely to be a rough guess, a range of 20% is considered successful. A larger range may be desirable, but since the user is able to stop the reformulation process himself, the size of the range is not important. Left on its own, the expert system stops the reformulation process when it achieves success, or it has run out of techniques to try. Sample Scenario The following sample scenario illustrates how the reformulation rules are applied. Since our current textbase concerns the domain of computer architecture, the example describes the interactions of the system with a user searching for Gauch -12- information on the alignment of word boundaries in memory. The user might enter a query 'boundary AND word ANDNOT page', which indicates that he wishes to retrieve passages containing information on word boundaries but not page boundaries. Assume a target number of 15. Applied to this textbase, the original query would retrieve only one passage, so the expert system would attempt to broaden the query. The first step would be to replace the word types boundary and word with their stemgroups. The resulting query would be 'Boundary A N D Word A N D N O T page', where the capitalized search terms indicate the whole stemgroup is included. Notice that page has not been expanded to its stemgroup, as it is a negative, or excluded, concept. Four passages would now be retrieved. The next step would be to broaden the query by including synonym stemgroups for each of the positive search terms, in turn. From the thesaurus it is found that Boundary has one synonym, Limit, however there is no synonym for Word. The query now becomes '(Boundary O R Limit) A N D Word A N D N O T page', which retrieves seven passages. Relaxing the context around the AND operator to adjacent sentences while decreasing the context around the A N D N O T operator to within 5 words increases the number of passages retrieved to nine. To further broaden the query, the parent stemgroups for the positive concepts are added. Block and Segment are added to the concept Boundary. The Word concept remains unchanged, since Word has no parent in the thesaurus. The query becomes '(Boundary O R Limit O R Block O R Segment) A N D Word A N D N O T page', which retrieves twelve passages. Twelve is within 20% of the fifteen passages requested, so the reformulation stops. If the user requests to see the retrieved passages, the expert system would rank the retrieved passages and present them to the user in decreasing rank-order. 2.5.2 Passage Ranking Rules The dialogue between the expert system and MICROARRAS normally produces a set of passages to be displayed to the user. The last task performed by the expert system is to rank order those passages in terms of their probable interest to the user. To do this, it performs an elementary content analysis on each passage and computes a weight representing probable interest. Ranking algorithms for document retrieval systems have been extensively studied (Harman, 1986). There has been less work done on ranking for passage retrieval systems. The FAIR system (Chang & Chow, 1988) performs a simple ranking based on the distance between word pairs, the number of search terms represented and the number of occurrences of the terms. Ro (1988) did an extensive comparison of different ranking algorithms for full-text, but failed to demonstrate significant differences in performance. Al-Hawamdeh (1991) ranks full-text paragraphs based on their similarity to a query expressed as an unstructured list of keywords, finding that the nearest-neighbor based searching is as effective as Boolean searching. Gauch -13- The ranking algorithm used by our expert system considers the following factors: the number of different concepts represented in the passage; the number of different word types for each concept; the relationship of the concept's word types to the user's original search terms; the number of occurrences for each word type from the search expression appearing in the passage; and the contextual distance between search terms. The passages are then ranked according to their respective index values and presented to the user in order of decreasing rank. Calculating Passage Weights The weight Wpq of passage p for query q, 0 <= Wpq <= 1, is a function of the weight Cip of each query concept i in p, the relationship between the concepts (determined by the parse tree), and the contextual closeness between the concepts. The concept weights are combined by applying the rules for fuzzy logic (Zadeh, 1965) to the Boolean structure of the query. Additionally, a closeness factor is associated with each of the AND and ANDNOT operators. The closeness factor for the AND operator is set to one of three values (1.0 for same sentence, 0.9 for adjacent sentences, 0.8 for same paragraph). The closer two positive concepts appear in the passage, the higher weight that passage receives. Complementary closeness values are used for the A N D N O T operator (0.8 for same sentence, 0.9 for adjacent sentences, 1.0 for same paragraph). Wp(Ci AND Cj) = min(Cip, Cjp) * PositiveCloseness (1) Wp(Ci OR Cj) = max(Cip, Cjp) (2) Wp(NOT Cj) = (1 - Cjp) (3) From (1) and (3) Wp(Ci ANDNOT Cj) = min(Cip, 1 - Cjp) * NegativeCloseness (4) The concept weights and closeness factors fall in the range [0,1], therefore the passage weights also fall in the range [0,1]. Calculating Concept Weights The weight of concept i in passage p, Cip, is a function of the weight of each concept term T in query q, denoted Tjq for search term j, and the weight of each concept term in the passage, denoted Tjp for search term j, and the number of search terms for the concept. The weight of a search term in the passage is multiplied by the weight of that search term in the query. Thus, the highest weight search terms are those which are important in the query as well as the passage. The weights for all the concept's search terms are summed together and normalized by the number of search terms for the concept, N. Gauch -14- N Cip = 1/N S Tjq * Tjp where term j is in concept i (5) j =1 The term weights fall in the range [0,1], therefore, the concept weights also fall in the range [0,1]. Calculating Term Weights Two different term weights, T, are calculated: the weight of the search term i in query q, Tiq, and the weight of the search term i in passage p, Tip. Query Term Weights The weight of the search term i in query q, Tiq, reflects the relationship of the search term to the user's original term. The relationships, from closest to most remote, are: same word, stemgroup, synonym, parent, sibling, child. These distances reflect the order in which search terms are added to the concepts, which in turn reflects confidence in the closeness of the relation of the search term to the original term. Tiq = 1.0 (word), 0.9 (stemgroup), 0.8 (synonym), 0.6 (parent), 0.5 (sibling), 0.4 (child) (6) The query term weights fall in the range [0,1] as required, with the original word receiving a weight of 1.0. Terms added by the expert system receive weights which decrease by 0.1 for every step away from the original term, except for the step from synonym to parent terms. This step decreases the term weight by 0.2, reflecting the large decrease in confidence which occurs when terms are added from outside the thesaurus class. Passage Term Weights The weight of the search term i in passage p, Tip, reflects the frequency of the search term in the passage, fip, and the frequency of the search term in the textbase, fit. An evaluation several full-text ranking algorithms and concluded that those based on relative document frequency provided the acceptable performance (Ro, 1988). Since this is also simple to calculate, we chose relative frequency for the term passage weights. Ti p = fip / fit (7) The term passage weights fall in the range [0,1], as required. Gauch -15- 3 EVALUATION Our primary goal is to demonstrate that using an expert system to reformulate queries can improve search performance for novice searchers. To evaluate the system, users queried the textbase using three interfaces with different capabilities: an interface whose only function was to accept contextual Boolean queries and display search results; a similar interface which also allowed the user to explore the online thesaurus; and a third which incorporated the searching expert system. Each subject's search performance with the three interfaces was monitored and compared. 3.1 Hypotheses Hypothesis 1: The expert system improves the search effectiveness for a novice searcher. Hypothesis 2: The expert system improves the search efficiency for a novice searcher. Hypothesis 3: The expert system can rank the passages retrieved by the search in decreasing order of relevance. The effectiveness of the retrieval output is evaluated by looking at recall (the number of relevant items found / the total number of relevant items in the database) and precision (the number of relevant items retrieved / the number of items retrieved). Two estimates of the number of relevant items retrieved are examined: the number of passages the users mark as relevant and the number of passages retrieved from the set of passages deemed relevant by the author. The efficiency of the systems is measured by the number of Boolean queries the subjects entered for each of several high-level questions, and by the amount of time they spent searching for relevant passages for each question. The ranking algorithm was evaluated by comparing the order of appearance of relevant passages after they have been ranked with a random order of appearance. 3.2 Method 3.2.1 Subjects Twelve computer science graduate students participated as subjects in the study. All subjects were knowledgeable in the use of computers, but unfamiliar with online searching. Thus, they were representative of the anticipated users of future information retrieval systems. Gauch -16- 3.2.2 Apparatus Information Retrieval Systems The user-alone configuration consisted of a Sun 3 running MICROARRAS and a rudimentary expert system. This expert system performed only the system control function, and did no query reformulation or ranking of retrieved passages. The user was prompted for a contextual Boolean query, this query was sent to MICROARRAS, and the number of passages retrieved was reported back to the user. The user could display the passages retrieved, if there were fewer than 25, or try another query. Typing was minimized by using the Sun's windowing package to cut and paste the previous query, edit, and re-run it. The u s e r - t h e s a u r u s version consisted of a Sun 3 with one window running MICROARRAS, as in the user-alone system, and a second window running a thesaurus access function. In the thesaurus window the user had access to all the thesaurus information available to the expert system. He could find out the stemname for a specific word's stemgroup. For any stemname, he could ask for the stemnames of the corresponding synonym, parent, sibling, or child stemgroups. These stemnames could be used in the user's query to MICROARRAS. Typing was minimized by using the Sun's windowing package to cut the stemgroup from the thesaurus window and paste it into the appropriate concept of the query. In the user-expert system version the user did not have access to the online thesaurus. Context and the addition of stemgroups were controlled by the expert system. Thus, the user entered a Boolean query and a target number of passages and the expert system reformulated the user's query to attempt to get close to the target number. The user was prompted to filter search terms found in the thesaurus, and to continue or abandon the current reformulation. To keep the response time approximately the same as for the other two configurations it was necessary to run MICROARRAS remotely on the Sun 4 file server containing the textbase. The user worked with one window on a Sun 3 which ran the full version of the query reformulation expert system. The expert system communicated with MICROARRAS over the network. This setup was approximately twice as fast as when MICROARRAS was run on the user's Sun 3. This speed up was necessary, not because the expert system code itself was slow, but rather because the expert system tended to form very long queries involving many MICROARRAS categories, and MICROARRAS slows down linearly with the number of search terms in a query. Questions Three sets of five questions were devised. Each set contained one training question and four questions on which the subjects were monitored. The questions covered material ranging over the whole textbase. For the fifteen questions, the number of relevant passages in the textbase ranged from a low of 7 to a high of 23 with a mean of 15.4. Gauch -17- 3.2.3 Procedure Subjects were asked to try to find on the order of ten relevant passages from the textbase in response to the questions they would be given. They were informed that they might not always be able to find that many, and they were allowed to stop working on a query whenever they were satisfied that they had found as much as they could. The target number of ten was chosen because it was large enough to require a high recall search, yet small enough that the users would not become tired reading passages. Each subject worked with each of the three systems, in turn. This was done to compensate for the large individual differences found in searching ability (Borgman, 1987). To compensate for learning during the experiment, the order of presentation of the three systems was counterbalanced among subjects. The order of presentation of the question sets was the same for all subjects (Set A first, then B, then C). Thus, each question set was searched on each system four times. The subjects received a training session with each system before they began their monitored searches. When they had completed all three sessions, they were asked to fill out the questionnaire stating their preferences and opinions. 3.2.4 Data Collection Raw Data Data was collected in a trace file while the subjects worked with the system. Each communication from the subject to the retrieval system, and vice versa, was stored with a time stamp. Thus, timing information was collected along with the history of queries entered by the subject and the search results. When the subject chose to display the retrieved passages, those passages and the subject's relevance judgment of them were also stored. Definitions A unique query was any error-free query entered by a subject. If a subject entered a query which contained a typographic or logical error, and he indicated that he noticed the error by aborting the search and re-entering a corrected version, then the erroneous query was not considered a unique query. However, if the subject gave no indication that he was aware of the error, but instead moved on to a different query altogether, then the erroneous query was considered unique. The relevance weight of a passage is the relevance number assigned to the passage by the subject. A very relevant (user) passage is one assigned a relevance weight of two. A somewhat relevant (user) passage has a relevance weight of one. A relevant passage (user) is one that is either very relevant or somewhat relevant, as judged by the user. An irrelevant passage (user) is a passage given a relevance Gauch -18- number of zero. It is necessary to have an estimate of the total number of relevant passages available for each question, in order to calculate recall. This estimate was calculated by forming the union, for each question, of the set of passages judged very relevant by any subject. Passages in this set judged irrelevant by the author were removed. The remaining passages form the absolute retrieval set and are called the r e l e v a n t passages. It was necessary to remove some passages marked very relevant by a subject because, perhaps due to a misinterpretation of the question or a misunderstanding of the passage, some subjects gave a relevance weight of two to irrelevant or marginally relevant passages. This tendency to overestimate the relevance of passages may also be because, in some cases, subjects were unable to find the truly relevant passages, and thought that they had retrieved the best passages available when in fact they had not. A successful retrieval set is a retrieval set containing at least five relevant passages. Since the subjects were attempting to find ten relevant passages, a successful retrieval set contains at least half the number for which they were looking. The textbase contained approximately the same number of relevant passages for each question, allowing the target number and size of the successful retrieval set to be held constant. The final retrieval set was chosen as the last successful retrieval set. If a subject never retrieved a successful retrieval set for a given question, the retrieval set with the highest number of relevant passages, as judged by the subject, was chosen. The final query is the query input by the user which resulted in the final retrieval set. Variables Total time per question is calculated from the entry of the subject's first query for the question until after the display, or decision not to display, of the final set of retrieved passages. Number of queries per question is determined by counting the number of unique queries the subject entered for a given question. Number of relevant passages (user) found per question is determined by counting the number of user indicated relevant passages in the final retrieval set for the question. User precision is calculated for the final retrieval set using the standard formula of: number of relevant passages (user) retrieved / number of passages retrieved. Number of relevant passages found per question is determined by counting the number of passages in the final retrieval set for the question that are members of the absolute retrieval set. Gauch -19- Precision is calculated for the final retrieval set using the standard formula of: number of relevant passages retrieved (absolute) / number of passages retrieved. Recall is calculated for the final retrieval set using the standard formula of: number of relevant passages retrieved (absolute) / total number of relevant passages available. The ranking balance point (R) for each retrieval set (not just the final one) is calculated by n S i * relevancei i=1 -------------------- n where n = number of passages in the retrieval set S relevancei i = position of the passage in the retrieval set i=1 relevancei = relevance weight of passage i This calculates where the midpoint of the relevant passages lies, accounting for the relevance weight. The earlier in the retrieval set the relevant passages occur, the smaller their midpoint. For example, consider a retrieval set of five passages of which one is very relevant (weight = 2), two are somewhat relevant (weight = 1) and two are irrelevant. An example ranking might present the very relevant passage first, then a somewhat relevant passage, an irrelevant one, the other somewhatr relevant passage and then the final irrelevant passage, represented as (2, 1, 0, 1, 0). Using this formula, the example balance point would be (1*2) + (2*1) + (3*0) + (4*1) + (5*0) ---------------------------------------------- = 2 2 + 1 + 1 + 0 + 0 Similarly, the worst-case balance point would be 4.25, and the example balance point 2. The best case balance point (BC) for each retrieval set is calculated by applying the ranking balance point formula to the case where all very relevant passages preceded all somewhat relevant passages which in turn preceded all non-relevant passages in the set. In this case, the best case ranking would be (2, 1, 1, 0, 0), yielding a balance point of 1.75. For comparison, the worst case ranking, (0, 0, 1, 1, 2), would yield a Gauch -20- balance point of 4.25. The midpoint (M) for each retrieval set is calculated by (n+1)/2 where n is the number of passages in the retrieval set. The midpoint is used for comparison with the ranking balance point. A ranking algorithm which distributed the relevant passages evenly throughout the set, in our example (1, 0, 2, 0, 1), would yield a ranking balance point of 3. This is also the midpoint of the set. A good ranking algorithm would produce distributions with balance points less than the midpoint (i.e. relevant passages presented earlier). The normalized ranking balance points were calculated from the ranking balance points by moving the midpoint to 0 and adjusting the range so that the best case balance point fell on 1, and the worst case balance point at -1. The normalization performed was: Normalized ranking balance point (NR) = ( M - R ) / ( M - BC ). For the example retrieval set, the normalized ranking balance point would be: (3 - 2) / (3 - 1.75) = 0.8. Note that this normalization is not possible when BC equals M, since a division by zero will be attempted. This can arise when all the retrieved passages receive the identical relevance weight (e.g. (2, 2, 2, 2, 2)). Since any order of presentation is as good as any other when the passages are all of equal relevance, it would be safe to ignore these cases. However, this situation never arose in the experiment. Summaries Calculated for Each System For each system the means calculated were: • number of queries per question • time per question (seconds) • number of relevant passages (user) per question • user precision • number of relevant passages (from absolute retrieval set) • precision • recall For each ranking algorithm (the expert system's, and randomness) the normalized balance points were calculated. Gauch -21- 3.3 Results The means were compared to determine if their differences were statistically significant. Pairwise two-tailed t-tests were performed. A difference was considered significant if its probability of occurring due to chance was less than 5% at the 95% confidence level (a 10% chance at the 95% confidence level was considered marginally significant). Pairs of means with statistically significant differences are flagged with asterisks. 3.3.1 Search Effectiveness All three systems retrieved comparable numbers of relevant passages. Whereas there seemed to be higher recall with the thesaurus, shown by a mean of 7.688 compared to a mean of 7.292 with the expert system, this difference was not significant (p = 0.5333). • number of relevant passages (user) per question • user alone 7.375 • user and thesaurus 7.688 • user and expert system 7.292 All three systems produced comparable precision, based on the subject's relevance judgments. • user precision • user alone 0.763 • user and thesaurus 0.786 • user and expert system 0.761 All three systems retrieved approximately the same number of passages from the absolute retrieval set. • number of passages from absolute retrieval set • user alone 5.521 • user and thesaurus 5.708 • user and expert system 5.729 Gauch -22- Recall was comparable across all three systems. There was a slight improvement in recall for the user and expert system configuration, but the advantage over the user- alone configuration was not significant (p < 0.6988). • recall • user alone 0.364 • user and thesaurus 0.368 • user and expert system 0.379 The user and expert system configuration produced marginally significant improvements in precision when compared with the user-alone configuration. • precision • user alone 0.530 * (p < 0.0817) • user and thesaurus 0.576 • user and expert system 0.604 * 3.3.2 Search Efficiency The user was marginally significantly slower when using a thesaurus. The expert system was not significantly slower than the other two systems. However, MICROARRAS was being executed by a Sun 4 with the user-expert system configuration resulting in approximately a doubling of its speed. • mean time per question (seconds) • user alone 474.5 * (p < 0.101) • user and thesaurus 571.5 * • user and expert system 539.8 The expert system improved search efficiency, as measured by number of user queries over both the user alone and user plus thesaurus. • number of queries per question • user alone 4.833 * (p < 0.0001) • user and thesaurus 5.458 * * (p < 0.0001) • user and expert system 2.354 * , * * Gauch -23- 3.3.3 Ranking The expert system ranked relevant documents more highly than would be predicted by randomness. The expert system's ranking was compared to a random distribution for 74 sets of retrieved passages. • balance points • random 5.00 * (p < 0.0165) • expert system 4.53 * • normalized balance points ( on range of -1 to +1 ) • random 0.000 * (p < 0.0025) • expert system 0.195 * 3.4 Analysis Effectiveness The first hypothesis, that the expert system can improve the search effectiveness for a novice user was not supported by this study. However, the expert system produced marginally significant improvements in precision, and seemed to indicate improvements in recall. Providing the online thesaurus produced no improvement in search effectiveness. The suggested improvements in precision may result from the expert system applying better broadening techniques. The subjects, when searching unassisted, would often stop with a very broad query and examine a large set of retrieved passages (over fifteen) looking for relevant information. This type of strategy results in the lower precision observed when the subjects search on their own. The subjects' browsing strategy may account for their ability to produce recall comparable to the expert system when there were a large number of relevant passages in the textbase. For example, in two questions with large absolute retrieval sets the subjects were able to retrieve, on average, 10 and 10.25 relevant passages on their own compared with the expert system's retrieval of 8 and 7.75 passages respectively. By using a target number of 10 for these broader questions, the expert system was operating at a disadvantage. More relevant information was easily found, judging by the high recall of the subjects, but the expert system did not even attempt to further broaden the query. Clearly, 10 is not the ideal target number for all queries. Efficiency The second hypothesis, that the expert system can improve the search efficiency of novice searchers, was supported. Using the expert system significantly reduced the Gauch -24- number of queries subjects needed to answer a given question. Subjects required fewer than half as many queries per question on average versus systems in which the user queried without it, a substantial improvement. The expert system reduced the amount of user effort required by decreasing the number of queries a user needs to design to express their information needs. If efficiency is measured in terms of total user time the expert system fares less well. The expert system was not significantly slower than either of the other two systems but it was necessary to run MICROARRAS on a faster machine to achieve this. However, this version of the expert system was designed with correctness rather than efficiency in mind, and there are several ways that it could be sped up. In particular, when a stemgroup is added to a concept, the entire query is re-evaluated against the textbase. A large speed improvement could be gained by unioning the passages retrieved by the new stemgroup with those retrieved by the rest of the concept (which has already been calculated). Allowing the subjects to access the online thesaurus actually decreased the subjects' efficiency. They took significantly more time than when they searched on their own and generated as many queries. This seems to indicate that the improvement in efficiency seen above was due to the expert system's searching knowledge base, not just access to an online thesaurus. Ranking The third hypothesis that the expert system could rank passages in decreasing order of relevance was supported. Although the expert system did present relevant passages significantly earlier than would be predicted by randomness, the improvement was not large enough to be considered truly successful. The current algorithm needs to be evaluated with different weights or a somewhat different algorithm needs to be tried in order to further improve the ranking function. Decreasing the query term weights more quickly as the query terms move farther from the original may improve the ranking by placing more emphasis on the user's original search terms. Using a more sophisticated closeness factor, one that took into account to how many words apart the search terms were in the passage, as well as sentence and paragraph measures considered in this version, could also lead to improved ranking. Reformulations Finally, some discussion of the number of reformulations performed by the expert system seems appropriate. The number of reformulations performed for a given user on a given question varies since some unsuccessful starting queries were reformulated before (and sometimes after) a successful starting query is found. The following statistics are given for the final query. The average number of reformulations performed on the starting query for the twelve questions, in order, were: 4.25, 5.75, 4.25, 2.75, 6, 5.75, 3.25, 3.25, 4.25, 2.75, 4, 1. This gives an average of 3.65 reformulations over all final retrieval set queries. It is interesting to note that the highest average number of reformulations is six. If the expert system is Gauch -25- continually broadening the query (which is the most common case), this means that even on the question requiring the most reformulations it stops, on average, just after adding child stemgroups. In fact, examining the 48 final retrieval set queries reveals that only in 3 cases did the expert system go past this point. Twice it went one more step and adjusted the context, and once it performed all ten reformulations on the broadening side before the user was satisfied with the number of passages retrieved. 3.5 Questionnaire The twelve subjects were asked which features of the expert system they liked best. The automatic addition of terms from the thesaurus was the most frequently mentioned (8 subjects), whereas the automatic context adjustment was the second most popular feature (3 subjects). Many subjects (8) mentioned the decreased amount of work needed to perform a search, with three of them specifically mentioning that they did not have to think as much. Other features mentioned which decreased the user effort were the simplified syntax, decreased typing, and the fewer queries to remember. System slowness was the feature most disliked (6 subjects). Although the amount of time necessary to answer a question was no greater with the expert system (see Section 3.3.2),there was less work for the user so time seemed longer. The other main complaints concerned the user interface. The subjects were fairly evenly split between wanting the system to proceed more automatically, with less prompting from them (4 subjects), whereas others wanted the system to explain what it was doing and/or allow the user to direct it (5 subjects). These comments lead to the conclusion that if a usable system is to be built based on the success of this research prototype, the execution of the system must be sped up and more work on interface design is needed. Almost all the subjects (10) found the user-expert system version the easiest to use, with the remaining two subjects split between the other two versions. Not surprisingly, given the comparable effectiveness of the three systems, the subjects were split on which system they felt gave the best results. Three voted for the user- alone version, two for the user-thesaurus, and three for the expert system. Three said it was a tie between the user-thesaurus and the expert system, and one abstained. 4 FUTURE WORK Running the experiment suggested several possible refinements to the system. The experimental subjects had many useful comments, the bulk of which dealt with the desire for a more sophisticated user interface. Desirable changes include: provision of a non-Boolean query language; allowing users to adjust the amount of system interaction; having the user specify the type of search desired, rather than having him give a specific target number; and increasing the speed of the system by improving the way the expert system uses MICROARRAS. Gauch -26- Observing the expert system reformulate real queries gave invaluable insight into which types of queries it handled well, and which it did not. A common type of query that required broadening was one that contained the intersection of three or more concepts. In this case, broadening context and adding search terms to each concept fails to address the fundamental problem of intersecting too many concepts. The next step of replacing the ANDs with ORs is too drastic a change. It invariably leads to too broad a query. Instead, the expert system should take the original query and drop each of the concepts in turn. The most common type of query requiring narrowing consisted of a single, high- frequency concept. None of the current reformulation techniques were of any use in this case. There were no operators to change, no context to adjust, and adding search terms just makes the query broader. This type of query should be treated as a special case. The concept's child concepts from the thesaurus should be presented as alternative, more specific, queries. The user could also be encouraged to AND this concept with another. The treatment of multiword phrases entered by the user which do not appear in the thesaurus should be changed. Currently, the only expansions done are expanding each word to its stemgroup and loosening the context allowed between the words of the phrase. It would be preferable to treat the words of the phrase as separate concepts which are ANDed together with adjacent context. Each phrase word could then be expanded using the full range of thesaural relationships, as is the case with regular search terms. Finally, more work is needed to improve the ranking of the retrieved passages. The current ranking algorithm should be tried with different weights for the query search terms and the closeness factor. It may be necessary to try entirely different algorithms, possibly incorporating syntactic or semantic information, to achieve high quality ranking. In conclusion, we have demonstrated that an expert system can provide online search assistance to improve the efficiency of novice searchers. Whereas more research is necessary to develop a better search assistant, I have been able prove that a useful search assistant can be developed which separates the search strategies from the domain knowledge, and that implementation of such a system is feasible now. REFERENCES Al-Hawamdeh, S., de Vere, R., Smith, G., and Willett, P. (1991). Using nearest- neighbour searching techniques to access full-text documents. Online Review, 15 (3/4), 173-191. Bates, M.J. (1979). Information search tactics. Journal of the ASIS, 30(4), 205-214. Belkin, N.J. & Marchetti P.G. (1990). Determining the functionality and features of an intelligent interface to an information retrieval system. Proceedings of the Gauch -27- Thirteenth Annual International ACMSIGIR Conference on Research & Development in Information Retrieval, Brussels, ACM Press, 151-174. Blair, D.C., and Maron, M.E. (1985). An evaluation of retrieval effectiveness for a full-text document-retrieval system. Communications of the ACM, 28(3), 289- 299. Blaauw G.A. and Brooks, F.P., Jr. (Fall, 1986). Computer Architecture, Volume 1- Design Decisions, Draft. Borgman, C.L. (1986a). Why are online catalogs hard to use? Journal of the ASIS, 37(6), 387-400. Borgman, C.L. (1986b). The user's mental model of an information retrieval system: an experiment on a prototype online catalog. International Journal of Man- Machine Studies, 24(1), 47-64. Borgman, C.L. (1987). Individual differences in the use of information retrieval systems: Some issues and some data. Proceedings of the Tenth Annual International ACMSIGIR Conference on Research & Development in Information Retrieval, New Orleans, ACM Press, 61-69. Brajnik, G., Guida, G., and Tasso, C. (1988). IR-NLI II: Applying man-machine interaction and artificial intelligence concepts to information retrieval. Proceedings of the Eleventh Annual International ACMSIGIR Conference on Research & Development in Information Retrieval; Grenoble, ACM Press, 387-399. Chang, S. and Chow, A. (1988). Towards a friendly adaptable information retrieval system. Proceedings of RIAO 88, M.I.T., 172-182. Chen, H. & Dhar V. (1990). Online query refinement on information retrieval systems: A process model of searcher/system interactions. Proceedings of the Thirteenth Annual International ACMSIGIR Conference on Research & Development in Information Retrieval, Brussels, ACM Press, 115-133. Chiaramella, Y. and Defude, B. (1987). A prototype of an intelligent system for information retrieval: IOTA. Information Processing & Management, 2 3 ( 4 ) , 285-303. Croft, W. B. and Thompson, R.H. (1987). I3R: A new approach to the design of document retrieval systems. Journal of the ASIS, 38(6), 389-404. Croft, W. B. and Turtle, H.R. (1992). Text retrieval and inference. In P.S. Jacobs (Ed.), Text-Based Intelligent Systems: Current Rsearch and Practice in Information Extraction and Retrieval. Hillsdale, N.J.: Lawrence Erlbaum Associates. Gauch -28- Crouch, C.J. (1988). A cluster-based approach to thesaurus construction. Proceedings of the Eleventh Annual International ACMSIGIR Conference on Research & Development in Information Retrieval, Grenoble, ACM Press, 309-320. Fenichel, C.H. (1981). Online searching: measures that discriminate among users with different types of experience. Journal of the ASIS, 32(1), 23-32. Fidel, R. (1991). Searcher's selection of search keys. Journal of the ASIS, 43(7), 490- 500. Gauch, S. (1991). Search improvement via automatic query reformulation. A C M Transactions on Information Systems, 9(3), 249-280. Gauch, S. (1992). Intelligent information retrieval: An introduction. Journal of the ASIS, 43(2), 175-182. Harman, D. (1986). Individual differences in the use of information retrieval systems: Some issues and some data. Proceedings of the 1986 ACM Conference on Research & Development in Information Retrieval, Pisa, ACM Press, 61-69. Harman, D. (1988). Towards interactive query expansion. Proceedings of the Eleventh Annual International ACMSIGIR Conference on Research & Development in Information Retrieval, Grenoble, ACM Press, 321-331. Jacobs, P.S. & Rau, L.F. (1990). SCISOR: Extracting information from on-line news. Communications of the ACM, 33, 88-97. Krawcsak, D., Smith, P.J., & Shute, S.J. (1987). EP-X: A demonstration of semantically-based search of bibliographic databases. In C.T. Yu and C.J. van Rijsbergen (Ed.), Proceedings of the Tenth Annual International ACM SIGIR Conference on Research & Development in Information Retrieval (pp. 263- 271). New Orleans, LA. McCune, B.P., Tong, R.M., Dean, J.S., Shapiro, D.G. (1985). RUBRIC: A system for rule-based information retrieval. IEEE Transactions on Software Engineering, SE-11(9), 939-945. Oldroyd, B.K. (1984). Study of strategies used in online searching 5: differences between the experienced and the inexperienced searcher. Online Review, 8(3), 233-244. Pollitt, A.S. (1984). A 'front-end' system: An expert system as an online search intermediary. ASLIB Proceedings, 36(5), 229-234. Pollitt, A.S. (1987). CANSEARCH: An expert systems approach to document retrieval. Information Processing & Management, 23(2), 119-136. Ro, J.S. (1988). An evaluation of the applicability of ranking algorithms to improve Gauch -29- the effectiveness of full-text retrieval. II. On the effectiveness of ranking algorithms on full-text retrieval. Journal of the ASIS, 39 (3), 147-160. Salton, G. (1986). Another look at automatic text-retrieval systems. Communications of the ACM, 29 (7), 648-656. Gauch -30- Shoval, P. (1985). Principles, procedures and rules in an expert system for information retrieval. Information Processing & Management, 21 (6), 475-487. Smeaton, A.F., and van Rijsbergen, C.J. (1983). The retrieval effects of query expansion on a feedback document retrieval system. The Computer Journal, 26 (3), 239-246. Smith, J.B., Weiss, S.F., and Ferguson, G.J. (1987). MICROARRAS: An advanced full-text retrieval and analysis system. Proceedings of the Tenth Annual International ACMSIGIR Conference on Research & Development in Information Retrieval, New Orleans, ACM Press, 187-195. Smith, P.J., Shute, S.J., Galdes, D., and Chignell, M.H. (1989). Knowledge-based search tactics for an intelligent intermediary system. ACM Transactions on Information Systems, 7(3), 246-270. Tong, R.M., Applebaum, L.A., Askmann, V.N., and Cunningham, J.F. (1987). Conceptual information retrieval using RUBRIC. Proceedings of the Tenth Annual International ACMSIGIR Conference on Research & Development in Information Retrieval; New Orleans, ACM Press, 247-253. Vickery, A., and Brooks, H.M. (1987). PLEXUS-The expert system for referral. Information Processing & Management 23 (2), 99-117. Zadeh, L.A. (1965). Fuzzy sets. Information and Control, 8, 338-353. 
work_qc3hpktherfxbjlacpurupzti4	----	Random Prism: a noise-tolerant alternative to Random Forests Article Accepted Version Stahl, F. and Bramer, M. (2014) Random Prism: a noise- tolerant alternative to Random Forests. Expert Systems, 31 (5). pp. 411-420. ISSN 1468-0394 doi: https://doi.org/10.1111/exsy.12032 (special issue on innovative techniques and applications of artificial intelligence) Available at http://centaur.reading.ac.uk/32914/ It is advisable to refer to the publisher’s version if you intend to cite from the work. See Guidance on citing . To link to this article DOI: http://dx.doi.org/10.1111/exsy.12032 Publisher: Wiley-Blackwell All outputs in CentAUR are protected by Intellectual Property Rights law, including copyright law. Copyright and IPR is retained by the creators or other copyright holders. Terms and conditions for use of this material are defined in the End User Agreement . http://centaur.reading.ac.uk/71187/10/CentAUR%20citing%20guide.pdf http://centaur.reading.ac.uk/licence www.reading.ac.uk/centaur CentAUR Central Archive at the University of Reading Reading’s research outputs online http://www.reading.ac.uk/centaur Random Prism: A Noise Tolerant Alternative to Random Forests Frederic Stahl and Max Bramer Abstract Ensemble learning can be used to increase the overall classification accuracy of a classifier by generating multiple base classifiers and combining their classification results. A frequently used family of base classifiers for ensemble learning are decision trees. However, alternative approaches can potentially be used, such as the Prism family of algorithms which also induces classification rules. Compared with decision trees Prism algorithms generate modular classification rules that cannot necessarily be represented in the form of a decision tree. Prism algorithms produce a similar classification accuracy compared with decision trees. However, in some cases, for example if there is noise in the training and test data, Prism algorithms can outperform decision trees by achieving a higher classification accuracy. However, Prism still tends to overfit on noisy data; hence ensemble learners have been adopted in this work to reduce the overfitting. This article describes the development of an ensemble learner using a member of the Prism family as the base classifier in order to reduce the overfitting of Prism algorithms on noisy datasets. The developed ensemble classifier is compared with a stand-alone Prism classifier in terms of classification accuracy and resistance to noise. 1 Introduction Ho’s Random Decision Forests (RDF) [15] and Breiman’s Random Forests (RF) [7], which was inspired by RDF, are two of the best known ensemble Frederic Stahl University of Reading, School of Systems Engineering, PO Box 225 Whiteknights, Reading, RG6 6AY e-mail: F.T.Stahl@reading.ac.uk Max Bramer University of Portsmouth, School of Computing, Buckingham Building, Lion Terrace, Portsmouth, PO1 3HE e-mail: Max.Bramer@port.ac.uk Frederic Stahl and Max Bramer learning methods. Ho claims that a decision tree cannot grow beyond a certain level without risking a loss of generalisation due to overfitting and proposes inducing multiple trees on randomly selected features. According to his the- ory, a combined classification will produce a better classifier as the individual trees generalise better on their subset of the feature space. Ho evaluates his theory empirically. Breiman’s RF is based on Ho’s RDF approach but also incorporates bagging (Bootstrap aggregating) [6], which is used to improve the classifier in terms of its stability and classification accuracy. An unstable classifier shows considerable variations to small changes in the training set. However, alternative partitioning strategies to bagging have been developed, such as [33, 31]. Like most rule based classifiers and ensemble classifiers, RDF and RF are based on the ‘divide and conquer’ rule induction approach. ‘Divide and con- quer’ induces classification rules in the intermediate form of a decision tree [21, 22]. A more general approach to inducing classification rules is the ‘sep- arate and conquer’ [32] approach, which induces a set of IF..THEN rules that do not necessarily fit into a decision tree representation. This alterna- tive approach to decision trees can be traced back to the AQ system in the 1960s [17]. However, modern representatives exist such as the Prism family of algorithms [25]. This paper presents and evaluates an ensemble classifier, based on the ‘separate and conquer’ approach, called Random Prism [24]. Random Prism is inspired by RF, in order to improve Prism’s classification accuracy further. The prototype is evaluated empirically regarding its classification accuracy in general and in noisy domains. This paper is structured as follows: Section 2 highlights related work in the context Random Prism. Section 3 introduces the Prism family of algorithms in comparison with decision tree classifiers and describes the newly developed Random Prism ensemble learner; Section 4 evaluates Random Prism on sev- eral datasets and compares it with a standalone Prism classifier in terms of classification accuracy and tolerance to noise. Section 5 highlights some ongo- ing and future work, notably some variations of the Random Prism approach and a parallel version of the Random Prism ensemble classifier. Section 6 closes the paper with a brief summary and some concluding remarks. 2 Related Work Besides the RF and RDF approaches mentioned in Section 1, ongoing re- search on ensemble learners, undertaken with the aim of overcoming over- fitting, comprises Dzeroski’s work on generating ensembles of heterogeneous classifiers using stacking [13]. The authors of [11] propose a parallel approach to generating hundreds of thousands of classifiers on small samples of the dataset in a distributed computing environment. Chan and Stolfo propose Random Prism: A Noise Tolerant Alternative to Random Forests the Meta-Learning framework [9, 10] which combines classifiers using various algorithms in order to create a final classifier, which could easily be paral- lelised using the multi-sample mining approach [20]. A recent development in ensemble learning is ‘Pocket Data Mining’ for mining data streams using heterogeneous base classifiers executed on different devices in an ad hoc net- work of smart phones [28]. Pocket Data Mining’s combining method is based on weighted majority voting. Also regarding data stream mining, the authors of [16] and [18] have developed ensemble techniques aiming to detect concept drift. Ensemble learning methods are not necessarily classifiers. However, in the literature the terms ensemble learner and ensemble classifier are often used interchangeably referring to an ensemble classifier, and this paper is no exception. As mentioned in Section 1 most ensemble learning methods are based on ‘divide and conquer’ algorithms for generating the base classifiers. There are heterogeneous ensemble approaches such as Meta-Learning [9, 10] that use different base classifiers. However, in practice these heterogeneous base classifiers are different versions of decision trees and hence use the ‘divide and conquer’ approach. A notable family of algorithms that follows the alternative ‘separate and conquer’ approach is the Prism family of algorithms [8, 3, 4]. A recent development of the Prism family is the parallelisation of Prism including a parallelised pre-pruning facility [27], an improved pre-pruning method, Jmax-pruning [26] and an adaptive version of Prism for data stream classification called ‘eRules’ [30]. Generally members of the Prism family produce a comparable classification accuracy to decision trees. However, in some cases they outperform decision trees, especially when the training and test datasets are noisy or there are missing values [3]. The fact that Prism classifiers tend to feature less overfitting in comparison with decision trees is the motivation for developing the Random Prism ensemble classifier aiming to improve Prism’s predictive accuracy further. 3 Random Prism This section describes our Random Prism ensemble learner. It first introduces the Prism Family of algorithms in Section 3.1 and compares them with the ‘divide and conquer’ approach used by RF. Section 3.2 then highlights the Random Prism approach based on the RF ensemble learner. 3.1 The Prism Family of Algorithms As mentioned in Section 1, the representation of classification rules differs between the ‘divide and conquer’ and ‘separate and conquer’ approaches. Rule Frederic Stahl and Max Bramer sets generated by the ‘divide and conquer’ approach are in the form of decision trees whereas rules generated by the ‘separate and conquer’ approach are modular. Modular rules do not necessarily fit into a decision tree and normally do not. The rule representation of decision trees is the main drawback of the ‘divide and conquer’ approach, for example rules such as: IF A = 1 AND B = 1 THEN class = x IF C = 1 AND D = 1 THEN class = x cannot be represented in a tree structure as they have no attributes in com- mon. Forcing these rules into a tree will require the introduction of additional rule terms that are logically redundant, and thus result in unnecessarily large and confusing trees [8]. This is also known as the replicated subtree prob- lem [32]. Cendrowska illustrates the replicated subtree using the two example rules above in [8]. Cendrowska assumes that the attributes in the two rules above comprise three possible values and both rules predict class x, all re- maining classes being labelled y. The simplest tree that can express the two rules is shown in Figure 1. The total set of rules that predict class x encoded in the tree is: IF A = 1 AND B = 1 THEN Class = x IF A = 1 AND B = 2 AND C = 1 AND D = 1 THEN Class = x IF A = 1 AND B = 3 AND C = 1 AND D = 1 THEN Class = x IF A = 2 AND C = 1 AND D = 1 THEN Class = x IF A = 3 AND C = 1 AND D = 1 THEN Class = x Fig. 1 Cendrowska’s replicated subtree example. Cendrowska argues that situations such as this can cause trees to be need- lessly complex make the tree representation unsuitable for expert systems and may require unnecessary expensive tests by the user [8]. ‘Separate and conquer’ algorithms avoid the replicated subtree problem by directly inducing sets of ’modular’ rules, avoiding unnecessarily redundant rule terms that are induced just for the representation in a tree structure. Random Prism: A Noise Tolerant Alternative to Random Forests The basic ‘separate and conquer’ approach can be described as follows, where the statement Rule_set rules = new Rule_set(); creates a new rule set: Rule_Set rules = new Rule_set(); While Stopping Criterion not satisfied{ Rule = Learn_Rule; Remove all data instances covered from Rule; rules.add(rule); } The Learn Rule procedure generates the best rule for the current subset of the training data, where best is defined by a particular heuristic that may vary from algorithm to algorithm. The stopping criterion is also dependent on the algorithm used. After inducing a rule, the rule is added to the rule set, all instances that are covered by the rule are deleted and a new rule is induced on the remaining training instances. In the Prism approach each rule is generated for a particular Target Class (TC). The heuristic Prism uses in order to specialise a rule is the probability with which the rule covers the TC in the current subset of the training data. The stopping criterion is fulfilled as soon as there are no training instances left that are associated with the TC. Cendrowska’s original Prism algorithm selects one class as the TC at the beginning and induces all rules for that class. It then selects the next class as TC and resets the whole training data to its original size and induces all rules for the next TC. This is repeated until all classes have been selected as TC. Variations exist such as PrismTC [5] and PrismTCS (Target Class Smallest first) [4]. Both select the TC anew after each rule induced. PrismTC always uses the majority class and PrismTCS uses the minority class. Both variations introduce an order in which the rules are induced, whereas there is none in the basic Prism approach. However unpublished experiments by the current authors show that the predictive accuracy of PrismTC cannot compete with that of the basic Prism and PrismTCS. PrismTCS does not need to reset the dataset to its original size and thus is faster than Prism. It produces a high classification accuracy and also sets an order in which the rules should be applied to the test set. The basic PrismTCS algorithm is outlined below where Ax is a possible attribute value pair and D is the training dataset. The statement Rule_set rules = new Rule_set(); creates a new rule set which is a list of rules and the line Rule rule = new Rule(i); creates a new rule with class i as classification. The statement rule.addTerm(Ax); Frederic Stahl and Max Bramer will append a rule term to the rule and rules.add(rule); adds the finished rule to the rule set. D’ = D; Rule_set rules = new Rule_set(); Step 1: Find class i that has the fewest instances in the training set; Rule rule = new Rule(i); Step 2: Calculate for each Ax p(class = i| Ax); Step 3: Select the Ax with the maximum p(class = i| Ax); rule.addTerm(Ax); Delete all instances in D’ that do not cover rule; Step 4: Repeat 2 to 3 for D’ until D’ only contains instances of classification i. Step 5: rules.add(rule); Create a new D’ that comprises all instances of D except those that are covered by all rules induced so far; Step 6: IF (D’ is not empty){ repeat steps 1 to 6; } We will concentrate here on the more popular PrismTCS approach but all techniques and methods outlined here can be applied to any member of the Prism family. 3.2 Random Prism Classifier The basic Random Prism architecture is illustrated in Figure 2. Random Prism uses the R-PrismTCS base classifier, which is outlined below in this Section. The prefix R denotes the random component of the base classifier which is Ho’s [15] random feature subset selection. Random Prism uses a further random component: Breiman’s bagging [6]. Both random components have been chosen in order to make Random Prism more robust towards noise in the training and test data. A further facility that has been implemented in the R-PrismTCS base classifiers, in order to improve their tolerance to noise, is J-pruning [4], a rule pre-pruning facility. Random Prism induces multiple R-PrismTCS classifiers and all of their rule sets are aggregated and used to classify new data instances employing a weighted majority voting. As the Random Prism classifier is inspired by the principles of the RF ensemble learner, this section first introduces the Random Forests classifier briefly and then discusses the new Random Prism ensemble classifier. As mentioned in Section 1, RF are inspired by the RDF approach from Ho [15]. RDF induces multiple trees in randomly selected subsets of the feature space in order to make the trees generalise better. RF uses RDF’s approach plus bagging [6] in order to improve the classifiers’ accuracy and stability. Bagging means that a sample with replacement is taken for the induction of each tree. Random Prism: A Noise Tolerant Alternative to Random Forests Fig. 2 The Random Prism architecture comprising Bagging, R-PrismTCS base clas- sifiers and weighted majority voting. The basic principle of RF is that it grows a large number of decision trees (a forest) on samples produced by bagging, using a random subset of the feature space for the evaluation of splits at each node in each tree. If there is a new data instance to be classified, then each tree is used to produce a prediction for the new data instance. RF then labels the new data instance with the class that achieved the majority of the ‘votes’. The Random Prism ensemble learner’s ingredients are RDF’s random fea- ture subset selection, RF’s bagging and PrismTCS as base classifier. Using Prism as a base classifier is motivated by the fact that Prism is less vulnerable to clashes, missing values and noise in the dataset and in general tends to produce less overfitting compared with decision trees [3], which are used in RF and RDF. In particular PrismTCS is used, as PrismTCS is also computationally more efficient than the original Prism while in some cases producing a better accuracy [27]. A good computational efficiency is needed, because ensemble classifiers induce multiple classifiers and thus place a high computational demand on CPU time. In the context of Random Prism, the terms PrismTCS and Prism may be used interchangeably in this paper, both referring to PrismTCS unless stated otherwise. Given a training dataset D, a sample Di is created, where i is the ith clas- sifier. For the creation of Di bagging and random sampling with replacement [6] are used. This means that the data samples may overlap, as in Di a data instance may occur more than once or may not be included at all. From each Di, a classifier Ci is induced. In order to classify a new data instance, each Ci predicts the class, and the bagged classifier counts the votes and labels the data instance with the class that achieved the majority of the votes. An ensemble classifier created using bagging often achieves a higher accuracy compared with a single classifier induced on the whole training dataset D and Frederic Stahl and Max Bramer if it achieves a worse accuracy it is often still close to the single classifier’s accuracy [14]. The reason for the increased accuracy of bagged classifiers lies in the fact that the composite classifier model reduces the variance of the individual classifiers [14]. The most commonly used bootstrap model for bag- ging is to take a sample of size n if n is the number of instances. This will result in samples that contain on average 63.2% of the original data instances. The fact that bagged classifiers can achieve a higher accuracy than a single classifier induced on the whole dataset D, as mentioned above, motivates the use of bagging in the Random Prism ensemble classifier proposed here. The RDF approach by Ho [15] induces multiple trees on randomly selected subsets of the feature space. Again a composite model is generated and it has been shown in [15] that they generalise better than a single tree induced on the complete feature space, as they are less prone to overfitting. Inspired from RDF, Breiman’s RF randomly selects a subset of the feature space for each node of each tree. Feature subset selection similar to the one used in RF is incorporated in Random Prism as well, inspired from the fact that random feature subset selection generalises the ensemble classifier better and thus makes it likely to produce a higher classification accuracy. The pseudo code below describes the adapted version of PrismTCS for use in Random Prism based on PrismTCS’s pseudo code in Section 3.1. M is the number of features in D: D’ = random sample with replacement of size n from D; Rule_set rules = new Rule_set(); Step 1: Find class i that has the fewest instances in the training set; Rule rule = new Rule(i); Step 2: generate a subset F of the feature space of size m where (M>=m>0); Step 3: Calculate for each Ax in F p(class = i| Ax); Step 4: Select the Ax with the maximum p(class = i| Ax); rule.addTerm(Ax); Delete all instances in D’ that do not cover rule; Step 5: Repeat 2 to 4 for D’ until D’ only contains instances of classification i. Step 6: rules.add(rule); Create a new D’ that comprises all instances of D except those that are covered by all rules induced so far; Step 7: IF (D’ is not empty){ repeat steps 1 to 7; } The pseudo code above is essentially PrismTCS incorporating RDF’s and RF’s random feature subset selection. For the induction of each rule term for each rule, a fresh random subset of the feature space is drawn. Also the number of features considered for each rule term is a random number between 1 and M. The PrismTCS version above is called R-PrismTCS, with the initial R denoting Random sample and feature selection. The basic Random Prism approach is outlined in the pseudo code below, where k is the number of R-PrismTCS classifiers to be induced and i is the ith classifier: double weights[] = new double[k]; Random Prism: A Noise Tolerant Alternative to Random Forests Classifiers classifiers = new Classifier[k]; for(int i = 0; i < k; i++){ Build R-RrismTCS classifier r; TestData T = instances of D that have not been used to induce r; Apply r to T; int correct = number of by r correctly classified instances in T; weights[i] = correct/(number of instances in T); } Please note that the Random Prism pseudo code above, creates not only a set of classifiers but also a set of weights. Random Prism does not employ a simple voting system like RF or RDF, but a weighted majority voting system as in the Pocket Data Mining System [29], where each vote is weighted accord- ing to the corresponding classifier’s accuracy on the test data. As mentioned earlier in this section, the sampling method used for each classifier selects approximately 63.2% percent of the total number of data instances, which leaves approximately 36.8% of the total number of data instances that are used to calculate the individual R-PrismTCS classifier’s accuracy and thus weight. Also the user of the Random Prism classifier can define a threshold N, which is the percentage of classifiers to be used for prediction. Random Prism will always select those classifiers with the highest weights. For example consider the classifiers and weights listed in Table 1. Table 1 Example data for weighted majority voting Classifier Weight A 0.55 B 0.65 C 0.55 D 0.95 E 0.85 Assume that the classifiers in Table 1 are the best classifiers selected ac- cording to the user’s defined threshold. Further assume that for a new unseen data instance classifiers A, B and C predict class Y and classifiers D and E predict class X. Random Prism’s weighted majority vote for class Y is 1.75 (i.e. 0.55 + 0.65 + 0.55) and for class X is 1.80 (i.e. 0.95 + 0.85). Thus Random Prism will label the data instance with class X. The R-PrismTCS pseudo code above does not take pruning into considera- tion, however a pre-pruning method J-pruning presented in [4] is implemented in R-PrismTCS in order to further generalise the base classifiers. J-pruning is based on the J-measure. According to Smyth and Goodman [23], the average information content of a rule of the form IF Y = y THEN X = x can be quantified by the following equation: Frederic Stahl and Max Bramer J(X; Y = y) = p(y) · j(X; Y = y) (1) The J-measure is a product of two terms. The first term p(y) is the proba- bility that the antecedent of the rule will occur. It is a measure of hypothesis simplicity. The second term j(X;Y=y) is the j-measure or cross entropy. It is a measure of the goodness-of-fit of a rule and is defined by: j(X; Y = y) = p(x | y) · log2( p(x|y) p(x) ) + (1 −p(x | y)) · log2( (1−p(x|y)) (1−p(x)) ) (2) If a rule has a high J-value, then it tends to have a high predictive accuracy as well. The J-value is used to identify when a further specialisation of the rule is likely to result in a lower predictive accuracy due to overfitting. The basic idea is to induce a rule term and if the rule term would increase the J-value of the current rule then the rule term is appended. If not then the rule term is discarded and the rule is finished. 4 Evaluation of Random Prism Classification Random Prism’s performance has been evaluated in terms of classification accuracy on several datasets highlighted in Section 4.1 and in terms of its tolerance to noise in Section 4.2. 4.1 Random Prism’s Classification Accuracy Random Prism’s classification accuracy has been evaluated on 15 different datasets retrieved from the UCI data repository [2] [24]. For each dataset, a test and a training set has been created using random sampling without replacement. The training set comprises 70% of the total data instances. Please note that the training set is sampled again by each R-PrismTCS base classifier, in order to incorporate bagging. Hence, as stated in Section 3.2 approximately 63.2% of the training data is used for the actual training and 36.8% is used to calculate the individual classifiers’ weights. The percentage of the best classifiers to be used was 10% and the total number of R-PrismTCS classifiers induced was 100 for each dataset. Table 2 shows the accuracy achieved using the Random Prism classifier and the accuracy achieved using a single PrismTCS classifier. As can be seen in Table 2, Random Prism outperforms PrismTCS in 9 out of 15 cases; in two cases Random Prism achieved the same accuracy as PrismTCS; and in only 4 cases Random Prism’s accuracy was below that of PrismTCS. However, looking into these four cases with a lower accuracy, i.e. for datasets ‘car evaluation’, ‘segmentation’, ‘soybean’ and ‘contact lenses’, it can be seen Random Prism: A Noise Tolerant Alternative to Random Forests Table 2 Accuracy of Random Prism compared with PrismTCS [24]. Dataset Accuracy PrismTCS Accuracy Random Prism monk1 0.79 1.0 monk3 0.98 0.99 vote 0.94 0.95 genetics 0.70 0.88 contact lenses 0.95 0.91 breast cancer 0.95 0.95 soybean 0.88 0.65 australian credit 0.89 0.92 diabetes 0.75 0.89 crx 0.83 0.86 segmentation 0.79 0.71 ecoli 0.78 0.78 balance scale 0.72 0.86 car evaluation 0.76 0.71 contraceptive method choice 0.44 0.54 that the accuracies for ‘car evaluation’ and ‘contact lenses’ are still very close the PrismTCS’s accuracy. In general, Random Prism outperforms its single classifier version PrismTCS in most cases and in the remaining cases, its accuracy is often very close to PrismTCS’s accuracy. 4.2 Evaluation of Random Prism’s Noise Tolerance Noise in an unavoidable problem in many domains of data mining, hence it is important that data mining and thus classification rule induction algorithms exhibit tolerance to noisy datasets. The evaluation of Random Prism’s noise tolerance has been conducted on 3 datasets from the UCI repository, the vote, contact lenses and breast cancer datasets. Several versions of the training and test datasets have been generated introducing different levels of noise in both the test and the training set. For example if the noise level was 10% then for each instance in the test or training set every attribute, including the classification, is replaced by a noise value with a probability of 10%. A noise value is any valid value for this attribute including its original value. For each level of noise in the training set (0%, 10%, 20%, 30%, 40% and 50%) a Random Prism classifier and a PrismTCS classifier was induced using the same parameter setting used for the experiments described in Section 4.1. This classifier was then used to classify the test sets with different noise levels (0%, 10%, 20%, 30%, 40%, 50%, 60%, 70% and 80%). The number of correct classifications for both Random Prism and PrismTCS has been counted and is compared for all three datasets in Figures 3, 4 and 5. Frederic Stahl and Max Bramer Fig. 3 Random Prism’s behaviour when introducing noise into the ‘vote’ training and test sets. Figure 3 compares the number of correct classifications of Random Prism and PrismTCS for increasing levels of noise in the training and test datasets of the ‘vote’ dataset. It can be observed that Random Prism outperforms PrismTCS in almost all cases. Random Prism: A Noise Tolerant Alternative to Random Forests Fig. 4 Random Prism’s behaviour when introducing noise into the ‘contact lenses’ training and test sets. Figure 4 compares the number of correct classifications of Random Prism and PrismTCS for increasing levels of noise in the training and test datasets of the ‘contact lenses’ dataset. Again, it can be observed that Random Prism outperforms PrismTCS in almost all cases. Frederic Stahl and Max Bramer Fig. 5 Random Prism’s behaviour when introducing noise into the ‘breast cancer’ training and test sets. Figure 5 compares the number of correct classifications of Random Prism and PrismTCS for increasing levels of noise in the training and test datasets of the ‘breast cancer’ dataset. Once again, it can be observed that Random Prism outperforms PrismTCS in almost all cases. In general it can be observed that for both Random Prism and PrismTCS the number of correct classifications decreases with an increasing noise level in the training and test set. However, as expected, in most cases Random Prism showed a higher tolerance to noise compared with PrismTCS. Random Prism: A Noise Tolerant Alternative to Random Forests 5 Ongoing and Future Work Ongoing and future work comprises a distributed / parallel version of Random Prism and several variations of the Random Prism approach itself. 5.1 Parallel Random Prism Classifier Random Prism, like any other ensemble learner, has a higher demand on CPU time compared with its single classifier version. Table 3 lists the runtimes of PrismTCS and Random Prism for the evaluation experiments outlined in Section 4.1. As ensemble learners are designed to reduce overfitting, they should be able to be executed on larger datasets as well, as the likelihood that noisy data instances are present is higher the larger the training data is. Table 3 Runtime of Random Prism on 100 base classifiers compared with a single PrismTCS classifier in milliseconds. Dataset Runtime PrismTCS Runtime Random Prism monk1 16 703 monk3 15 640 vote 16 672 genetics 219 26563 contact lenses 16 235 breast cancer 32 1531 soybean 78 5078 australian credit 31 1515 diabetes 16 1953 crx 31 2734 segmentation 234 15735 ecoli 16 734 balance scale 15 1109 car evaluation 16 3750 contraceptive method choice 32 3563 It can be seen, that as expected, the runtimes are much larger for Random Prism than for PrismTCS. This is because Random Prism induces 100 base classifiers whereas PrismTCS is only a single classifier. One would expect the runtimes of Random Prism to be 100 times greater than for PrismTCS as Random Prism induces 100 base classifiers, however the runtimes are much shorter than expected. The reason for this is that the base classifiers use a subset of the feature space and thus have fewer features to scan for the induction of each rule term. Future work will address the problem of scalability of the Random Prism classifier. Google’s Parallel Learner for Assembling Numerous Ensemble Trees (PLANET) system [19] addresses this problem in the context of decision tree Frederic Stahl and Max Bramer based ensemble classifiers using the MapReduce [12] model of distributed computation. MapReduce builds a cluster of computers for a two-phase dis- tributed computation on large volumes of data. First, in the map-phase, the dataset is split into disjoint subsets, which are assigned together with a user specified mapping function to workers (mappers) in the MapReduce cluster. Each mapper then applies the mapping function on its data. The output of the mapping function (a key-value pair) is then grouped and combined by a second kind of worker, the reducer, using a user defined reducer function. Fig. 6 Random Prism parllelised using the MapReduce framework. Figure 6 highlights the parallelisation of Random Prism using MapRe- duce. For Random Prism, the MapReduce model will be used to distribute the induction of the R-PrismTCS base classifiers using map functions. The individual R-PrismTCS classifiers are then combined using a reducer function to create a final Random Prism classifier. Executing each mapping function on a different computer in the MapReduce cluster results in the computa- tional expensive induction of R-PrismTCS based classifiers being distributed over a whole cluster of workstations. An open source implementation of the MapReduce model called Hadoop is available [1]. 5.2 Variations of the Random Prism Ensemble Classifier There are many possible variations of the Random Prism approach that may achieve better classification accuracy, for example different versions of Prism could be used as base classifiers. Also it would be possible to use a diverse mix of all existing Prism classifiers, such as Prism, PrismTC or PrismTCS. Some Prism classifiers may perform well on certain samples, some may perform worse, thus a larger variety of Prism classifiers per sample may well increase Random Prism’s classification accuracy. Also, it is possible to use several Prism and decision tree base classifiers for each sample. Random Prism: A Noise Tolerant Alternative to Random Forests 5.3 Intelligent Voting System Random Prism’s classification accuracy may be further improved by employ- ing a more intelligent voting system. For example a classifier may have in general a moderate predictive accuracy. However, concerning its predictions for class A, the classifier may have a very high predictive accuracy. Such cases could be addressed by calculating individual weights for each class for this particular classifier. Implementing more than one weight for a classifier must also be addressed in the selection of the best classifiers according to a user defined percentage. A similar approach called ‘Combining’ has been used by the Meta-Learning system [9, 10]. 6 Conclusions This work presents the Random Prism ensemble classifier based on the Prism family of algorithms as the base classifier. Most ensemble learners are based on decision trees as base classifiers and aim to reduce the overfitting of the model in order to achieve a higher classification accuracy. However, alterna- tive base classifiers exist, such as the Prism family of algorithms. It has been discussed that Prism algorithms already perform better on noisy datasets compared with decision trees, as they tend to overfit less. The motivation be- hind Random Prism is that an ensemble classifier based on the Prism family of algorithms may further reduce the overfitting and thus achieve a higher classification accuracy compared with single Prism classifiers. First, the Prism family of algorithms has been introduced and compared with decision trees and next the well known Random Forests approach has been reviewed. Random Prism is inspired by the Prism family of algorithms, the Random Decision Forests and the Random Forests approaches. Random Prism uses the PrismTCS classifier as base classifier with some modifications called R-PrismTCS. The modifications were made in order to use the Random Decision Forests’ feature subset selection approach. Random Prism also incor- porates J-pruning for R-PrismTCS and Random Forests’ bagging approach. Contrary to Random Forests and Random Decision Forests, Random Prism uses a weighted majority voting system instead of a plain majority voting system, in order to take the individual classifier’s classification accuracy into account. Also, Random Prism does not take all classifiers into account, and the user can define the percentage of classifiers to be used for classification. Random Prism will select only the classifiers with the highest classification accuracy for the classification task. Random Prism has been evaluated on 15 datasets from the UCI repository and has been shown to produce a better classification accuracy on 9 cases compared with PrismTCS. In two cases, the classification accuracy was the same as for PrismTCS. In two further cases the classification accuracy was Frederic Stahl and Max Bramer slightly below PrismTCS’s accuracy and only in two cases was it much worse than PrismTCS’s accuracy. Random Prism has also been evaluated in terms of its tolerance to noisy training and test datasets in three cases from the UCI repository. In most cases observed, Random Prism achieved a higher classification accuracy compared with PrismTCS. Ongoing work on Random Prism comprises the development of a parallel version in order to make Random Prism scale better on large datasets. For this the MapReduce framework for utilising a computer cluster is considered. Furthermore a more intelligent voting system taking different accuracy levels of the classifier for different classifications into account is planned. Addition- ally, a variety of Random Prism versions are planned, comprising different Prism classifiers as base classifiers. Author Biographies Dr Frederic Stahl Frederic Stahl joined the University of Reading , UK in December 2012 where he is a lecturer. Prior to joining Reading he worked as a lecturer at Bournemouth University and as Senior Research Associate at the University of Portsmouth. He obtained his PhD from the University of Portsmouth in 2006 with the thesis title ”Parallel Rule Induction”. He remained an active researcher in the field of ”Big Data” analytics, in particular in parallel data mining approaches. Besides parallel and distributed data mining, Dr Stahl’s research interests are also in data stream mining, data mining in resource constraint environments, machine learning and artificial intelligence. Since 2011 Dr Stahl is a technical program committee member of the annual SGAI ”International Conference on Artificial Intelligence”. He is also author and co-author of over 24 technical peer reviewed publications. Prof. Max Bramer Professor Max Bramer is Emeritus Professor of Information Technology at the University of Portsmouth, UK. He has been an active researcher into different areas of Artificial Intelligence since the 1970s (PhD awarded 1977). In recent years he has focused particularly on Knowledge Discovery and Data Mining and is the author of a popular textbook on ’Principles of Data Mining’. Pro- fessor Bramer is the Chair of the Specialist Group on Artificial Intelligence of the British Computer Society and Vice-Chair of the Technical Committee on Artificial Intelligence of the International Federation for Information Pro- cessing. He was a member of the founding steering committee for the IEEE Random Prism: A Noise Tolerant Alternative to Random Forests ICDM series of data mining conferences. He is the author, co-author or editor of over 200 technical publications. References 1. Hadoop, http://hadoop.apache.org/mapreduce/ 2012. 2. C L Blake and C J Merz. UCI repository of machine learning databases. Technical report, University of California, Irvine, Department of Information and Computer Sciences, 1998. 3. M A Bramer. Automatic induction of classification rules from examples using N- Prism. In Research and Development in Intelligent Systems XVI, pages 99–121, Cambridge, 2000. Springer-Verlag. 4. M A Bramer. An information-theoretic approach to the pre-pruning of classifica- tion rules. In B Neumann M Musen and R Studer, editors, Intelligent Information Processing, pages 201–212. Kluwer, 2002. 5. M A Bramer. Inducer: a public domain workbench for data mining. International Journal of Systems Science, 36(14):909–919, 2005. 6. Leo Breiman. Bagging predictors. Machine Learning, 24(2):123–140, 1996. 7. Leo Breiman. Random forests. Machine Learning, 45(1):5–32, 2001. 8. J. Cendrowska. PRISM: an algorithm for inducing modular rules. International Journal of Man-Machine Studies, 27(4):349–370, 1987. 9. Philip Chan and Salvatore J Stolfo. Experiments on multistrategy learning by meta learning. In Proc. Second Intl. Conference on Information and Knowledge Management, pages 314–323, 1993. 10. Philip Chan and Salvatore J Stolfo. Meta-Learning for multi strategy and par- allel learning. In Proceedings. Second International Workshop on Multistrategy Learning, pages 150–165, 1993. 11. Nitesh V. Chawla, Lawrence O. Hall, Kevin W. Bowyer, and W. Philip Kegelmeyer. Learning ensembles from bites: A scalable and accurate approach. J. Mach. Learn. Res., 5:421–451, December 2004. 12. Jeffrey Dean and Sanjay Ghemawat. Mapreduce: simplified data processing on large clusters. Commun. ACM, 51:107–113, January 2008. 13. Saso Dzeroski and Bernard Zenko. Is combining classifiers with stacking better than selecting the best one? Machine Learning, 54:255–273, 2004. 14. Jiawei Han and Micheline Kamber. Data Mining: Concepts and Techniques. Morgan Kaufmann, 2001. 15. Tin Kam Ho. Random decision forests. Document Analysis and Recognition, International Conference on, 1:278, 1995. 16. J. Zico Kolter and Marcus A. Maloof. Dynamic weighted majority: An ensemble method for drifting concepts. Journal of Machine Learning Research, 8:2755– 2790, December 2007. 17. R S Michalski. On the Quasi-Minimal solution of the general covering problem. In Proceedings of the Fifth International Symposium on Information Processing, pages 125–128, Bled, Yugoslavia, 1969. 18. Leandro L. Minku and Xin Yao. Ddd: A new ensemble approach for dealing with concept drift. IEEE Trans. on Knowl. and Data Eng., 24(4):619–633, April 2012. 19. Biswanath Panda, Joshua S. Herbach, Sugato Basu, and Roberto J. Bayardo. Planet: massively parallel learning of tree ensembles with mapreduce. Proc. VLDB Endow., 2:1426–1437, August 2009. 20. Foster Provost. Distributed data mining: Scaling up and beyond. In Advances in Distributed and Parallel Knowledge Discovery, pages 3–27. MIT Press, 2000. Frederic Stahl and Max Bramer 21. R J Quinlan. C4.5: programs for machine learning. Morgan Kaufmann, 1993. 22. Ross J Quinlan. Induction of decision trees. Machine Learning, 1(1):81–106, 1986. 23. P. Smyth and R M Goodman. An information theoretic approach to rule in- duction from databases. Transactions on Knowledge and Data Engineering, 4(4):301–316, 1992. 24. Frederic Stahl and Max Bramer. Random prism: An alternative to random forests. In Max Bramer, Miltos Petridis, and Lars Nolle, editors, Research and Development in Intelligent Systems XXVIII, pages 5–18. Springer London, 2011. 25. Frederic Stahl and Max Bramer. Computationally efficient induction of classifi- cation rules with the pmcri and j-pmcri frameworks. Knowledge-Based Systems, 35(0):49 – 63, 2012. 26. Frederic Stahl and Max Bramer. Jmax-pruning: A facility for the information the- oretic pruning of modular classification rules. Knowledge-Based Systems, 29(0):12 – 19, 2012. 27. Frederic Stahl, Max Bramer, and Mo Adda. Parallel rule induction with informa- tion theoretic pre-pruning. In Max Bramer, Richard Ellis, and Miltos Petridis, editors, Research and Development in Intelligent Systems XXVI, pages 151–164. Springer London, 2010. 28. Frederic Stahl, Mohamed Gaber, P. Aldridge, D. May, H. Liu, Max ramer, and P. Yu. Homogeneous and heterogeneous distributed classification for pocket data mining. In A. Hameurlain, J. Kung, and R. Wagner, editors, Transactions on large-scale data and knowledge-centered systems V, volume 5 of Lecture notes in computer sciences. Springer, Berlin, 2012. 29. Frederic Stahl, Mohamed Medhat Gaber, Han Liu, Max Bramer, and Phillip S. Yu. Distributed classification for pocket data mining. In 19th International Symposium on Methodologies for Intelligent Systems, Warsaw, Poland, 2011. Springer. 30. Frederic Stahl, MohamedMedhat Gaber, and ManuelMartin Salvador. erules: A modular adaptive classification rule learning algorithm for data streams. In Max Bramer and Miltos Petridis, editors, Research and Development in Intelligent Systems XXIX, pages 65–78. Springer London, 2012. 31. Haixun Wang, Wei Fan, Philip S. Yu, and Jiawei Han. Mining concept-drifting data streams using ensemble classifiers. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, KDD ’03, pages 226–235, New York, NY, USA, 2003. ACM. 32. I H Witten and F Eibe. Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations. Morgan Kaufmann, 1999. 33. Indre Zliobaite. Three data partitioning strategies for building local classifiers. In G. Valentini, M. Re, and O. Okun, editors, Ensembles in Machine Learning Applications. Springer, Heidelberg, 2011. 
work_qcxfniobyne2bjkncjf3tadu7m	----	A Collaborative Fuzzy Expert System for the Web Tod A. Sedbrook U n i v e r s i t y of N o r t h e r n C o l o r a d o Abstract A convergence of Internet and fuzzy logic technolo- gies provides an opportunity for experts and end users to collaborate in developing, refining, and test- ing knowledge-based systems. Internet technology removes geographical and time-based restraints, and fuzzy rule bases are easier to understand and maintain. This paper describes an architecture and a prototype for developing, defivering, and maintain- ing expert systems on the World Wide Web. The system's collaboration components allowed experts to monitor user consultations remotely, view ~,~ summaries of responses, and trace-rule inference chains. Experts and users participated in real-time chat sessions or posted questions on extended dis- i!! cussion lists. The system allowed experts and us- ers to experiment with real-time enhancements of ~ knowledge bases. Fuzzy rules resulted in semanti- cally richer knowledge bases that flexibly handled ~ complex and uncertain knowledge. A fuzzy infer- ~,~. ence engine supported hedges and partial match- ing to assist users in applying knowledge and ex- ploring Web-based data. Keywords: expert system, fuzzy logic, Internet, collaboration, design ACM Categories: H.4.2, 1.2.1,1.2.5 Introduction: Challenging Expert System Assumptions Expert systems attempt to clone human expertise to avoid geographical and time-based limitations of consulting with human experts. Turban and Aronson :i;i (1998, p. 17) present the main idea behind expert systems: "Expertise is transferred from the expert to the computer...users call on the computer for spe- iii cific advice as needed." The expert's role is to as- sist knowledge engineers in developing, refining, and testing the knowledge base. Once the knowledge base is delivered, the expert is not involved in as- sisting specific users, and users are not involved in ..~i maintaining the knowledge base. ':~' This paper explores a convergence of Internet and fuzzy logic technologies that are challenging these I~ assumptions. Internet technology removes geo- ~ graphical and time-based constraints to improve users' access to human expertise. With Web-based ;; expert systems, experts and users can more easily i >, collaborate in problem solving and knowledge base The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) 19 development. Fuzzy logic technologies reduce cog- nitive dissonance by coding knowledge in English- like expressions which provide an opportunity to better involve users in maintaining knowledge bases. Compared to standard-rule representations, the numbers and complexity of rules required in fuzzy models are much less than with traditional knowl- edge bases (Cox, 1995). The following scenario demonstrates the potential for improved communication and ongoing mainte- nance of globally distributed knowledge bases. As the user, you are responsible for troubleshooting printer problems on your network. You tie into a help- desk site with your Web browser that feeds you a Java fuzzy expert system. You select a knowledge base for your printer problems and start the consul- tation. During the consultation the system requires that you evaluate a document's print quality. You specify that the print quality is between poor and average. Unfortunately, the consultation's recom- mendations prove inadequate to solve your prob- lem and you request an on-line chat with a human expert. You are in luck because the system is un- dergoing a maintenance review, and the experts are currently available to monitor consultations. You take your place in a queue, and an audible alarm alerts you that an expert is available. The expert remotely observes as you repeat your previous consultation. During the consultation you and the expert review the fuzzy rules concerning print quality. A rule's fuzzy premise states - - if print quality is s i g n i f i c a n t l y poor. You have an on-line chat with the expert describing your print quality. The expert agrees to change tem- porarily the premise of the rule to - - if print quality is s o m e w h a t poor. Another consultation session with the revised rule then resolves the printer prob- lem. The expert's change, and the consultation re- sults are logged for the knowledge base's next main- tenance review. This paper explores architecture for supplementing knowledge-based expert systems with Internet tech- nology and fuzzy models. The following sections review research issues, describe a system archi- tecture for development and delivery, discuss our experiences with prototype consultations delivered over theWeb, and conclude and raise future research possibilities. Research Issues Over 12,500 expert systems are deployed in manu- facturing, medicine, and business; and the number of experts system development tools has been grow- ing at about 16% per year (Durkin, 1996). Neverthe- less, expert system developers continue to struggle with design issues such as knowledge acquisition, testing, and maintenance. The Internet and local intranets offer new ways to deliver knowledge-based advice. Traditional shells, however, do not support openness and interoperability required for deploy- ing expert systems over wide-area networks. In ad- dition, globally accessible knowledge bases are dif- ficult to maintain and update. Researchers are ex- ploring new Web-based applications of expert sys- tems, and the following highlights their current ef- forts. One technique for deploying experts systems on the Internet relies on Common Gateway Interfaces (CGI) to coordinate client/server interactions (MultiLogic Inc., 1998; Inference Corp., 1996; Bello and Ribeiro, 1995). The CGI server is responsible for controlling logical inferences, maintaining the system's knowl- edge, and dynamically constructing and distributing HTML forms to conduct consultations. Java tech- nology offers a new way to deliver expert systems where the user's browser serves as an interface for the consultation applet (Ernest Friedman-Hill, 1996). Java-based shells deliver knowledge and provide au- tomatic access to Internet resources but otherwise structure consultations in the same manner as tra- ditional expert systems. Autonomous software agents provide another form for Internet delivery of knowledge bases. Software agents cooperate to exchange knowledge to solve user problems. Researchers are focusing on agent communication structures and interoperability to improve cooperation and support networks of dis- t r i b u t e d k n o w l e d g e b a s e s ( G e n e s e r e t h and Ketchpel, 1994). In contrast to improving communications among software agents, others are focusing on applying the Internet to improve interpersonal communications for knowledge acquisition. Shaw and Gaines (1997) have developed REPGRID - - an interactive Web- based system to elicit personal constructs for col- laborative learning. The system supports knowledge acquisition through interactive repertory grid analy- sis, where knowledge engineers and experts col- laborate from different geographical locations. Oth- ers have proposed that knowledge-based systems on the Internet can support collaboration efforts by assisting in constructing user interfaces and sup- porting s y s t e m d e s i g n s (Benford et al. 1993; Nakakoji, and Fischer, 1995; Far and Koono, 1996). 20 The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) The Internet offers an opportunity to combine the synergies of expert systems and groupware tech- nology. Expert system tools improve groupware by assisting in information retrieval, simplifying system operations, and structuring interactions (Aiken et al. 1991 ). Groupware tools support expert systems by promoting collaboration and maintaining distributed knowledge bases. Maintaining and enhancing Web-based expert sys- tems require a method to help users and designers manage uncertainty caused by lack of available data, ambiguous classification categories, and semantic misunderstandings (Cox, 1995; Walz, Elam and Curtis, 1993). End users, designers, and experts must make judgments, assess the appropriateness for applying rules, and share responsibly for the safety and legal implications that result from misdi- agnosis or misclassifications (Mira, Yanez, and Barreiro, 1991). Fuzzy models respond to these concerns by providing a means to represent and adapt to inherent vagueness and ambiguities that can result when applying a general rule to a specific situation (Klir and Folger, 1988). Fuzzy expert systems reason through fuzzy logic membership functions. Membership refers to the degree to which a particular attribute's value belongs to a set. For example, someone with an age at- tribute equal to 21 years would have high member- ship in the set 'young' and a low membership in the set "old." Membership functions allow degrees of membership to be related to linguistic terms (Zadeh, 1965). tial growth, users are challenged to locate data re- sources and create queries that are aligned to sup- port specific decisions. Internet retrieval tools such as search engines, query-by-form, and intelligent "wizards" provide assistance in the mechanics of information retrieval but offer limited support for help- ing users comprehend a database's context and contents (George, Buckles, Petry, & Radhakrishnan, 1996). A knowledge base combined with a fuzzy query system helps users obtain knowledge-based assistance to form queries in commonly understood terms. Internet-aware knowledge bases improve commu- nication and productivity for a variety of information systems. Figure 1 shows the diversity of IS environ- ments requiring support where knowledge bases can assist in database connectivity and offer mobile agents - - "traveling e x p e r t s " - - to assist in access- ing enterprise and middleware resources across wide-area networks. This review suggests that combining fuzzy expert systems and groupware tools improve knowledge development, delivery, and maintenance. Fuzzy techniques result in knowledge bases that are flex- ible and semantically rich. Internet-based groupware allows experts and end users to collaborate in man- aging and accessing knowledge and data resources. The following describes a prototype development and delivery environment to investigate collabora- tive fuzzy expert systems delivered on the Web. Fuzzy logic in expert systems allows fuzzy proposi- tions of the form: if size is more or less small then investment is rather large, where small and large are linguistic variables denoting fuzzy memberships and more or less and rather are hedges that modify memberships. Fuzzy expert systems also apply fuzzy numbers representing degrees of membership over intervals. Table I defines fuzzy variables, hedge and fuzzy numbers. Fuzzy expert systems allow partial matching of a rule's antecedents to provide a systematic way of managing imprecision and uncertainty. Compared to traditional expert systems, fuzzy expert systems take less time to develop, reduce maintenance cost, and improve user understanding (Cox, 1995; Schneider et al. 1996). As the amounts and diversity of information in data- bases and on the Internet continue their exponen- Instant Fuzzy Traveling Expert Advice ITEA (Instant Traveling Expert Advice) is both an integrated development environment (IDE) applica- tion and Java delivery applet for producing, deliver- ing, and collaborating to maintain fuzzy knowledge bases on the Web. The entire system is implemented in Java to promote portability and ease of network- ing. Figure 2 shows the system's components. The delivery applet conducts user consultations by applying knowledge bases created by the IDE. Con- sultations take place within a Web browser, where users respond to system generated questions. Based on the user's responses and fuzzy logical in- ferences, the system retrieves data, derives prob- lem solutions, and presents recommendations. The following describes system components and pre- sents examples of fuzzy rules. The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) 21 Fuzzy Variable Hedges Fuzzy Numbers A fuzzy variable is an attribute described by a set of linguistic values. Sets of values are related by a membership function to express possibilities of set membership for particular instances. For example, terms representing increasing membership such as narrow, medium, and wide may describe a shoe's broadness attribute. An instance of a shoe may be described as medium broadness, indicating that shoe's membership in "broadness" is intermediate between narrow and wide. A hedge is a linguistic term to qualify fuzzy variables by increasing, reducing or restricting membership levels. For example, hedges such as "very," "strongly," and "really" denote increased membership. Hedges such as "more or less," "slightly," and "few" denote decreased membership. Hedges such as "relatively," "technically," and "strictly" denote restrictions on membership levels. A shoe with slight medium broadness indicates reduced membership compared to a shoe with medium broadness (Bouchon-Meunier,1992). A fuzzy number relates numeric intervals to degrees of possibility associated with a proposition. For example, a fuzzy number may relate a temperature level to the proposition that the temperature is warm. A temperature of 32 ° F. or (0 ° C.) would have low membership in warm. While 60 ° F. would have a higher membership. Two common forms of fuzzy numbers are triangular and trapezoidal. Table 1. Definitions of fuzzy terms. I n t r a n e t ~ ' ~ ~ ~ ~ ' ~ P o ~ z e r d ~ . . ~ ~ Groupware Server I ~ ~ ~ d . p _ ~ , , I hA.~p..q I / Web Server / Laptop c~ompuM;; Quadra Figure 1. Knowledge bases can support database retrieval, and decision making across Intranets and the Internet. 22 The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) Expe~ < ~ User Collaboration Tools Database Retrieval Tools Developn Tools Fuzzy Delivery Reasoning Environment Java Internet Technology Figure 2. System components for ITEA. The Development Environment The Java IDE features provide an object-oriented editing environment for creating classes, subclasses, attributes, and values. Developers create IF-THEN rules that support logical operations ("AND," "OR," "NOT" "ELSE"), fuzzy numbers, fuzzy variables, and fuzzy reasoning. The system provides a fuzzy mem- bership function generator that supports hedges for fuzzy linguistic variables and allows developer to specify triangular or trapezoidal memberships func- tions for fuzzy numbers. Developers set the system's inference strategy to pursue single or multiple val- ues for goals and specify confidence thresholds for firing rules. The Delivery Environment Figure 3 displays the applet interface for the deliv- ery environment. The delivery environment consists of l i b r a r i e s of Java c l a s s e s t h a t can be deliveredacross any platform. The expert system is a Java applet that is executed within a user's browser. Users may select from a list of knowledge bases residing at a Web site. The system conducts the consul-tation and may retrieve additional Web-based resour-ces to present questions, provide explana- tions, explore databases, and retrieve solutions. At the conclusion of a consultation, the user may re- view the session's rule inference chains, repeat the consultation, or select another knowledge base. Collaboration and Explanation Environment The delivery of expert system consultations on the Web provides an opportunity to improve interaction between experts and users. Figure 4 displays the system's interface for remote collaboration. The collaboration components allow experts to monitor users consultations remotely, view summaries of us- ers responses, and trace inference chains. Further, experts and users can participate in either real-time chat sessions or post questions on extended dis- cussion lists. During real-time chat, experts make on-the-fly changes to rule bases to allow users to remotely experiment with knowledge base changes. Knowledge base changes and enhancements are maintained at a central site and automatically dis- tributed user sites. Reasoning with Fuzzy Variables The following demonstrates how ITEA manages in- ferences with fuzzy variables. For instance, con- sider the following rules: The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) 23 T h i s s y s t e ~ wiZZ ~eooznmend a computer s y s t e m t h a t b e s t f i t s y o u z n e e d s . Adjust the sllder b a r to best answe~ t h e q u e s t i o n s . Ok, f i r s t q u e s t i o n : Wheze w i l l y o u u s e t h e co~q~ute~? Figure 3. User interface for the delivery environment. R u l e I n f e r e n c e s ] d e l i v e r y i n v e n t o r y == slowly falling CF = 5 5 --- c o n c l I <RULE> fall i n v e n t o r y I <IF> c u s t o m e r d e m a n d p r e d i c t i o n == rising A N D d e l i v e r y c o n f i d e n c e == u n s u r e < T H E N > l d e l i v e r y i n v e n t o r y == falling ELSE d e l i v e r y i n v e n t o r y == slowly rising I < R E F E R E N C E > User R e s p o n s e s . . . . c u s t o m e r sales == ? a c c e l e r a t i n g CF = 3 0 .... c u s t o m e r sales == slightly d e c l i n i n g CF = 7 0 . . . . c u s t o m e r d e m a n d p r e d i c t i o n = = s t e a d i l y rising CF .... c u s t o m e r d e m a n d p r e d i c t i o n == ? falling CF = 3 .... orders this w e e k == ? a c c e p t a b l e CF = 0 . . . . orders this w e e k == v e r y much limited CF = 1 0 0 . . . . orders this w e e k == ? l a c k i n g CF = 0 Comments ~<Rolie> d o e s c u s t o m e r d e m a n d refer to l o n g term or short term d e m a n d |<Chris> if refers to t h e n e x t f e w m o n t h s ~,~,: J<Rolie> t h e shipment w a s short this w e e k C o n s u l t a t i o n D i s c o s , i o n , Ic.,t0me, saies t h a t are o n l y s l i g h t l y d e c l i n i n g h a v e o n l y a 3 0 Z ~ i I change of a eler,ting Figure 4. Collaboration tools for remote session monitoring and chat. 24 The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) Rule 1: If environment humidity == more or less low AND environment wind == breezy AND environment temperature == mostly warm THEN weather condition == more or less nice Rule 2: IF weather condition == slightly nice AND snow condition == mostly good OR snow condition == very ok THEN recommend activity == ski Assume the following fuzzy attributes have been defined for application classes - e n v i r o n m e n t , w e a t h e r and snow. The attributes for e n v i r o n m e n t are: humidity with fuzzy values - - (high, medium, low) and hedges (very, more of less, slightly) wind with fuzzy values - - (breezy, calm) and hedges (strongly, slightly) temperature with fuzzy values - - (hot, warm, cold) and hedges (mostly, not very) The attributes for w e a t h e r are: condition with fuzzy values - - (the best, great, nice, ok, terrible, the worst) and hedges (very, more or less, slightly) The attributes for s n o w are: condition with fuzzy values - - (excellent, good, ok, bad, awful) and hedges (very, mostly, slightly) The goal of the system is to provide a recommended activity. Consider the fuzzy inferences for a consul- tation where the user provides the following fuzzy facts: humidity is very low wind slightly breezy temperature is mostly warm snow condition is very good The system reasons by backward chaining to as- sign fuzzy memberships that depend on the user's responses and the possibility distributions defined by a rule's hedges and fuzzy attributes. For the re- sponses above, the following can be concluded: The fuzzy certainty level was determined by the fol- lowing procedure: Given the user response that humidity is very low, a membership of 100% is assigned to the Rule 1 first proposition requiring humidity to be more or less low. The user response that the wind is slightly breezy allows the system to assert that wind has an 85% membership in breezy and a 15% membership in calm. The rule's statement contains no hedges so that membership assignment is not changed. The user response that the temperature is mostly warm warrants that the third statement is assigned a 100% membership in mostly warm. Since the Rule 1 premise's fuzzy certainty of 85% exceeds the system's threshold level of 50%, Rule 1 fires to conclude the weather condition is nice. The rule fires by adding the fact that the weather is nice to the system's working knowledge and adjusting the confidence according to the conclusion's hedge. Nice at a membership level of 85% (assigned by the premise) warrants a 100% membership for the hedged fuzzy set more or less nice (see Figure 5). So, the system asserts that weather is more or less nice adjusted to a membership of 100%. Rule 2 assigns the attribute - - recommended activ- ity with the values - - ski with a confidence level of 100% for the following reasons: Rule 2 premises IF weather condition == slightly nice AND snow condition == mostly good OR snow condition == very ok More or less nice weather is known by the system to be true at a level of 100% from Rule 1. Since the first statement of the Rule 2 premise requires only a weather condition of slightly nice, this statement re- sults in an assignment of 100% membership for slightly nice.That is, a weather condition that is more or less nice is also slightly nice. The first premise statement of Rule 1 is true to a fuzzy certainty level of 85%, the minimum member- ship attained among the first rule's three premise statements: The user's assessment that the snow condition is very good results in assigning a 100% fuzzy cer- tainty for the premise's second statement, where the statement requires a slightly good snow condition. IF environment humidity == more or less low AND environment wind == breezy AND environment temperature == mostly warm The rule membership assignment depends on both the user responses and the construction of a rule's statements. The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) 25 ~the /,'-I~terrible , ~ o k /~nice ~ g r e a t fl~the I I I 100%- -- - - membership in more or less nice F i g u r e 5 . Membership functions for fuzzy values of attribute weather. A fuzzy membership of 85% for the fuzzy value nice translates in to 100% membership for more or less nice. If the user had answered that the snow condition was mostly ok, the rule would fail since Rule 2 re- quires the snow to be at least very ok. If the first statement in the Rule 2 premise had re- quired the weather condition to be very nice, the rule would fail since the Rule 1 conclusion warrants only that the weather condition is more or less nice. Database Retrieval Module ITEA supports exploration of Internet databases by providing assistance in defining information needs, determining relevant information sources, and for- mulating fuzzy queries. The knowledge base con- trols database retrieval by dynamically constructing initial queries, retrieving data, calculating fuzzy mem- bership, and presenting query results. The system's knowledge base applies an expert's domain knowledge to define a retrieval vocabulary that reflects qualitative distinctions within a domain. For example, in a medical domain, physicians dis- tinguish the onset of a disease by fuzzy terms such as acute, rapid, or normal. Users are better pre- pared for future collaboration as they explore and learn about complex relations by applying a domain's terminology (Larsen and Yager, 1997; Xu and Ichikawa, 1992). The knowledge base pre-classifies users and rec- ommends query profiles that are likely to satisfy a specific category of users. For example, users may be classified according to level of experience, where beginners receive detailed guidance in formulating queries. Experienced users may dispense with con- sultation and instead directly explore databases. The knowledge base communicates with a Fuzzy Database Retrieval Module to recommend data sources, vocabulary, and attribute sets for construct- ing queries, and explanations. The Retrieval Mod- ule is then responsible for connecting with Internet databases and retrieving information. The user may then refine the query by selecting different combi- nations of attributes, applying different query opera- tors, or raising and lowering acceptable thresholds. After database exploration, the user returns to the consultation session where the knowledge base's explanation model summarizes items retrieved and may provide suggestions for future explorations. Discussion This section describes our experiences and dis- cusses the application of ITEA for delivery and main- tenance of Web knowledge bases.The following de- scribes a prototype application that helps users to rank sets of banks according to financial ratios. Financial analysts value banking stocks through common financial ratios, such as return on invest- ment, price-earnings, price-book, and other ratios unique to the banking industry such as net interest 26 The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) margin and efficiency ratio. Net interest margin measures the spread between interest rates for loans and deposits, divided by total earning assets (loans plus securities). The efficiency ratio measures the proportion of operating expense to revenue. The knowledge base helps users form queries that express financial ratios through fuzzy terms instead of quantitative limits. For example, financial ana- lysts consider a bank well managed if its efficiency ratio is low, and they suggest that 60% is a good target. The analysts also know that a price-earn- ings (P/E) ratio of 21 is good and that 24 is very good (Schlegel, 1997). The analyst may then per- form queries and implicitly rank stocks based on their expertise in construing financial ratios. ample, during query formulation the following advice is offered: "A price-to-earnings ratio is the share price divided by earnings per share for the company's most re- cent four quarters. A very high P/E ratio may indi- cate an overpriced stock. But remember there may be a good reason for the premium in stock price. You probably should select a lower ratio for conser- vative stocks." Users may decide to follow the knowledge base's recommendation or formulate customized queries. The system applies the user's responses to define a query and locate Internet databases and commu- nicate results. A fuzzy query, in contrast, requires only that novice users specify banks with "low" efficiency ratios and "very good" P/E ratios. The system is responsible for assigning fuzzy memberships to banks through fuzzy comparisons within a group of bank stocks. The knowledge base assists users to formulate que- ries initially by explaining financial ratios, suggest- ing preformulated queries that best match a user's investment style, and guiding query construction and automating retrieval. The knowledge base classifies users according to risk and investment goals. Conservative investors are willing to accept lower returns in exchange for safer investments. Moderate investors seek mod- eration in returns and risks, while aggressive inves- tors are willing to risk losses in exchange for the possibility of larger returns. The knowledge base contains rule-based advice for selecting banks based on a user's profile. The following rule defines a pre- formulated fuzzy query for aggressive investors: <RULE> aggressive investor <IF> investor desire == match a query to my style AND investor preferences == aggressive <THEN> query! price == none AND query! price-book == high ratio AND query! price-earnings == high ratio AND query! return on equity == very high ROE AND query! net interest margin == very wide AND query! efficiency ratio == very low The knowledge base also offers guidance to help users interactively construct fuzzy queries. For ex- The knowledge base dynamically constructs an ini- tial query, connects with database resources, offers query refinement, and presents results. Figure 6 presents the interface where a user reviews query results and formulates fuzzy queries to further ex- plore bank stocks. At the completion of user explorations, query results are returned to the knowledge base for further analy- sis. The knowledge base provides additional infor- mation concerning individual banks and suggests other Internet resources to further evaluate compa- nies. The initial evaluations of the prototype focused on user interaction and understanding. An exploratory evaluation suggests that users' final investment de- cisions and explanations were consistent with ex- pert recommendations for selecting bank stocks. Results indicate that users valued support for con- structing queries and easily explored databases with fuzzy queries. Specifically, the tools capture and apply expertise to assist novices in aligning queries with user preferences, promoting understanding of the information contained in a domain, and allowing users to discover data patterns interactively and bet- ter interpret results. Collaboration Strategy The collaboration subsystem was designed to im- prove user interactions and assist in knowledge-base development and maintenance. The computer se- lection knowledge base was placed on a Web server January 1997, and since then there have been thou- sands of consultations sessions from users repre- senting numerous countries. The following summa- The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) 27 Figure 6, Fuzzy query results. rizes early experiences with the collaboration sub- system. Users found it easy to provide feedback, and their suggestions included adding additional rules for Macintosh and other computer brands, adding price considerations, reworking question syntax, provid- ing new Internet resources links, and modifying hedges in rules (e.g., changing moderately unim- portant to slightly important). Other users alerted designers to interface and display problems associ- ated with running the system on their platform and suggested new features such as improved graphic handling and enhanced reasoning facilities. We have recently added on-line session monitoring and real-time chat to increase levels of collabora- tion between users and experts. Initial experiences suggest that users are eager to share real-world ex- periences to assist knowledge designers and ex- perts. Collaborations allowed experts to experiment by modifying and adding rules to respond to user concerns immediately. Experimental versions of the knowledge bases were then maintained at a central site along with the established version. Other user sites could then access the experimental versions and provide comments. The combination of expert systems delivery and groupware addressed real-time operational and maintenance concerns. Experts provided answers to user questions to assist users during consulta- tions. Users provided designers with insight into user problem-solving approaches. The systems collabo- ration features offered an opportunity to increase user participation, enhance the knowledge base, and centralize distributions of updates. Currently there are several ITEA knowledge bases deployed at several locations on the Web including (see http://www.instanttea.com for demonstration systems): Movie a d v i s o r - - recommends a movie and provides access to resources Sports advisor - - an intelligent index to sports re- lated internet sites Fish disease diagnosis - - a diagnosis system with links to remedies Mutual fund advisor - - a classification system for mutual funds Snake identification - - identifies regional snakes Conference planning - - offers intelligent indexes to conference information Car diagnosis - - a troubleshooting system Golf club selection - - recommends clubs for vari- 28 The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) ous course situations Computer selection - - recommends computer brands with links to suppliers These systems are available on the Web 24 hours a day and take advantage of distributed Internet re- sources to present explanations and conclusions. ITEA's delivery and development environment is freely available to researchers and educators by contacting the author. Conclusion The convergence of technologies including Java, the Internet, and fuzzy logic provides a new solution for delivery and maintenance of expert systems. Internet technology allows knowledge to be deliv- ered anywhere in the world. Fuzzy logic supports construction of rule bases that are easier to under- stand and maintain. This paper presented ITEA as a Web-based fuzzy expert system to promote collaboration and improve expert system delivery and maintenance. Web- based expert systems allow users easy access to Internet resources and provide designers a way to maintain and distribute knowledge from a central location. Fuzzy rules result in semantically richer knowledge bases that flexibly handle complex and uncertain knowledge. Integrated collaboration fea- tures that support real-time chat and rule modifica- tions show promise for improving knowledge base usability and rule maintenance. The first prototype of ITEA was developed in the fall of 1996. The collaboration components were added in early 1997. We are currently prototyping three- tier modules, object-oriented databases, and new protocols to improve delivery and collaboration. Our plans include testing usability among experts and users to evaluate how ITEA meets knowledge engineering goals, refining an expert system devel- opment methodology that provides increased levels of ongoing collaboration between end users and experts, and integrating and investigating Web knowl- edge bases and databases. The integration of the Web resources with the field of fuzzy expert systems offers new ways of sharing and distributing knowledge in an organization. Us- ers and experts can jointly access corporate re- sources to explore fuzzy models of complex situa- tions collaboratively. This also offers an opportunity for simulating and refining the modeling of business processes (Yu et al, 1996; Fishwick, 1991). A col- laborative fuzzy expert system for the Web supports a new design for distributed decision making. References Aiken, M.W., Sheng, O, and Vogel, D. (1991 ). "Inte- grating Expert Systems with Group Support Sys- tems" ACM Transactions on Information Systems, Vol. 9, No. 1, January, pp. 75-95. Benford, Steve, Bullock, Adrian, and Harvey, P., and Grace, P. (1993). "A System to Support the Devel- opment and Use of Global Computer-Supported Cooperative Work (CSCW) Applications" Internet Research, Vol. 3, No. 1, Spring, p. 25. Belo, O., and Ribeiro, A. (1995). "A Web-Based Framework for Distributed Expert Systems," Confer#ncia Nacional WWW (CNW3'95), Universidade do Minho, Braga, Portugal, Julho. Bouchon-Meunier, B. (1992). "Fuzzy Logic and Knowledge Representation Using Linguistic Modi- fiers." Fuzzy Logic for the Management of Uncer- tainty," (Zadeh and Kacprzyk, eds.), NewYork, NY: Wiley. Cox, E. (1995). Fuzzy Logic for Business and Indus- try. Charles River Media. Rockland, MA. Durkin, J. (1996). "Expert Systems - A View of the Field;' IEEE Expert~Intelligent Systems and Their Applications, Vol. 11, No. 2, April, pp. 56-63. Far, B.H., and Koono, Z. (1996). "Ex-W-Pert System: A Web-Based Distributed Expert System for Groupware Design," World Congress on Expert Systems '96, Seoul, Korea, February, pp. 545- 552. Fishwick, P. (1991). "Fuzzy Simulation: Specifying and Identifying Qualitative Models," Int. J. General Systems, Vol. 19, pp. 295-316. Friedman-Hill, Ernest and Sandia National Labora- tories (1996). Jess Java Expert System Shell. URL:http://herzberg.ca.sandia.gov/jess. MultiLogic Inc. (1998). URL: http://www.multilogic.com. Genesereth, M., and Ketchpel, S. (1994). "Software Agents," Communications of the ACM, Vol. 37, No. 7, July, pp. 72-76. Inference Corporation's Case-Base Reasoning Search Engine on the Web (1996). URL: http:// m5.inference.com/wcp/. Klir, G.J., and Folger, T.A. (1988). Fuzzy Sets, Un- certainty, and Information, Englewood Cliffs, N J: Prentice-Hall. Larsen, H.L., and Yager, R. (1997). "Query Fuzzification for Internet Information Retrieval." Fuzzy Information Engineering (eds., Dubois, The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) 29 Prade, & Yager). New York: John Wiley & Sons, Inc., pp. 253-264. Mira, J., Yanez, A., and Barreiro, A. (1991). "Fuzzy Specification in Data Capture and Knowledge Representation For an Expert System in Oncol- ogy" Fuzzy Sets and Systems. Vol. 44, No. 3, De- cember, pp. 431-448. Nakakoji, K., and Fischer, G. (1995). "Intertwining Knowledge Delivery and Elicitation: a Process Model for Human-computer Collaboration in De- sign" Knowledge-Based Systems., Vol. 8, Nos. 2- 3, April, pp. 94-106. Shaw, M., and Gaines, B.R. (1997). "Comparing Con- structions Through the Web" Knowledge Science Institute, University of Calgary, Calgary, Alberta, Canada T2N 1N4h, ftp://ksi.cpsc.ucalgary.ca/ar- ticles/CSCL95WG/. Schneider, M., Kandel, A., Langholz, G., and Chew, G. (1996). Fuzzy Expert System Tools, NewYork, NY: Wiley. Turban, E., and Aronson, J. (1998). Decision Sup- port Systems and Intelligent Systems, Englewood Cliffs, N J: Prentice Hall. Walz D.B., Elam, J., and Curtis, B. (1993). "Inside a Software Design Team: Knowledge Acquisition, Sharing and Integration" Communication of the ACM. Vol. 36, No. 10, October, pp. 63-76. Xu, Wu, and Ichikawa, Tadao. (1992)."KDA: A Knowl- edge-based Database Assistant with a Query Guiding Facility" IEEE Transactions on Knowledge and Data Engineering, Vol 4, No. 5, pp. 443-453. Yu, E., Mylopolulos, J., and Lesperance, Y. (1996). "AI Models for Business Process Reengineering" IEEE Expert~Intelligent Systems and their Appli- cations. Vol. 11, No. 4, August, pp. 16-23. Zadeh, L.A. (1965). "Fuzzy Sets" Information and Control, Vol. 8, pp. 338-353. About the Author Tod A. Sedbrook is an associate professor at the College of Business Administration, University of Northern Colorado, Greeley, Colorado. He received his Ph.D. in management information systems from the University of Colorado. His research interests include Internet delivery of instructional resources and methodologies of object-oriented analysis and design. He has published in the areas of knowledge- based systems that apply fuzzy logic, genetic algo- rithms, neural networks and statistical techniques. He is the editor of the International Business Schools Computing Quarterly and President of Instant Trav- eling Expert Advice, Inc. E-mail: tasedbr@ unco.edu 30 The DATA BASE for Advances in Information Systems - Summer 1998 (Vol. 29, No. 3) 
work_qfm3rrdj3vhvtgib5fh3argega	----	doi:10.1016/j.eswa.2004.12.012 BioMen: an information system to herbarium M. Delgado, W. Fajardo, E. Gibaja, R. Pérez-Pérez* Department of Computer Science and Artificial Intelligence, Universidad de Granada, E.T.S.I. Informática. C\ Periodista Daniel Saucedo Aranda, 18071 Granada, Spain Abstract This paper presents the development of BioMen (Biological Management Executed over Network), an Internet-managed system. By using service ontologies, the user is able to perform services remotely from a web browser. The services are managed by means of a Multi-Agent System, i.e. an Input/Output system, which interacts with the web server. In addition, artificial intelligence techniques have been incorporated so that the necessary information may be obtained for the study of biodiversity. We have built a tool which will be of particular use to botanists and which can by accessed from anywhere in the world thanks to Internet technology. In this paper, we shall present the problems we encountered when building this tool and how we managed to overcome them. q 2004 Published by Elsevier Ltd. Keywords: Information system; Botanist; Biodiversity; Multi-agent system; Semantic web; Service ontology 1. Introduction A herbarium is defined as a place where collections of dried, classified plants are stored before being used as material for the study of botany. The specimens contained in herbariums are and always have been the essential base for performing systematic, floral and biogeographical studies; in addition, as a collection of perfectly identified and ordered dried plants these represent a permanent record of biodiversity. It is currently calculated that more than 2.5 billion specimens are to be found in natural history museum collections and herbariums throughout the world (Duck- worth, Genoways, & Rose, 1993). Biological diversity research and study requires satisfactory access to this biological information. As this complex information is currently distributed among herbariums all over the world, this makes it practically inaccessible (Berendsohn et al., 1999). After a time when the declining interest in global questions entailed a certain neglect of floral studies, 0957-4174/$ - see front matter q 2004 Published by Elsevier Ltd. doi:10.1016/j.eswa.2004.12.012 * Corresponding author. Tel.: C34 95 824 9562; fax: C34 95 824 6251. E-mail addresses: mdelgado@ugr.es (M. Delgado), aragorn@ugr.es (W. Fajardo), gibaja@decsai.ugr.es (E. Gibaja), ramon@decsai.ugr.es (R. Pérez-Pérez). and consequently of the work of herbariums, the current preoccupation for the deterioration of the environment has resulted in new interests being defined which has again brought to the forefront the concern for the particular composition of our surroundings. The study and conserva- tion of plant biodiversity is undoubtedly one of the most important challenges facing botanists, naturalists, environ- mentalists, etc. in the next millennium. Let us take the example of GBIF (Global Biodiversity Information Facility) (GBIF, 2004), an international project which aims to make all important data about biodiversity freely and universally available on Internet. The approach of GBIF shall contribute to economic growth, ecological sustainability, social effects and scientific research by increasing the usefulness, availability and completeness of the main scientific source of available biodiversity information on Internet. Most of the information necessary for GBIF comes from information stored in the herbariums all over the world. Therefore, by its own nature, the herbarium once again becomes an essential piece for the development of these objectives and those in charge of it are responsible for providing the response called for by the research community. Consequently, one of the prime current needs is to acquire updated, relevant, scientifically contrasted and easily accessible information as the basis for conservation, Expert Systems with Applications 28 (2005) 507–518 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518508 the handling and the sustainable use of biodiversity. However, the complexity and variability of studies carried out in this field has forced these institutions to adopt new techniques and protocols which are capable of responding to the ever growing demands (Berendsohn, 2001). One first approach at tackling this problem of manage- ment was developed by Pando (1991) with the Herbar program. Due to the fact that the management problem is one shared by all herbariums, individual software packages have been developed for each institution in order to solve their problems, and there is no single, duly developed, standard solution. Generally speaking, the solutions advocated by each herbarium are partial solutions for their problems and have been developed by the herbar- ium’s own staff with little experience of system development. Below, we shall present some of the most important software packages and their features: 1. HERBAR (Real Jardı́n Botánico de Madrid, 1991): free software, entirely developed in Microsoft Access. This reflects some of the special characteristics common to herbarium management such as for example, the entering of information, genera, countries, provinces, information filtering. It creates labels. 2. Virtual Herbarium Express (the New York Botanical Garden, 1996): free software which enables data input by means of forms created in Microsoft Access XP. This software allows labels and certain reports to be created. 3. BRAHUS (University of Oxford, 2002): free software developed in Visual FoxPro using DBASE as the databases. This is one of the most complete software packages and allows a better control of information, printing of labels, user policy, etc. The following characteristics are common to all of the above software packages: 1. Free software. 2. Use of not particularly powerful database managers (Access, Dbase). 3. Data entered using templates. 4. Information filtering. 5. Label creation. 6. Decentralized use of software. After analyzing a herbarium’s needs, it can be seen that the systems developed so far have not been able to incorporate a large number of requirements. For this reason, BioMen was developed, and by taking advantage of modern communication technologies, the information is available online. The center’s information can therefore be consulted without having to be requested from the center itself, in line with the philosophy of GBIF (2004) whereby all infor- mation should be public knowledge. This system adminis- ters all herbarium tasks. In particular, this has been implemented in the prototype stage in the herbarium at the University of Granada. This center was chosen for its complexity and completeness. It meets all the characteristics to be modeled and presents all the problems to be resolved of an extremely important center historically among herbariums. The University of Granada’s herbarium currently belongs to the executive of the Iberian and Macaronesian Herbarium Association. After this brief introduction, we shall go on to explain the work carried out. In Section 2, we shall present an analysis of the problem. In Section 3, we shall show the characteristics of the developed system. In Section 4, we shall show how the information about the treatment of biodiversity has been obtained. Finally, Section 5 ends with a series of conclusions about the system. 2. System analysis A herbarium, as we have already mentioned, is a place where collections of dried, classified plants are stored, so that these can later be used as material for botanical study. From this definition, we can highlight the concepts of storage and study. The stored material is studied in order to obtain information which will be used for the conservation, handling and sustainable use of biodiversity (Lane et al., 2000). The need therefore arises for the information to be available in a suitable, standardized form so that it may be studied by researchers. These specimens are constantly consulted both by researchers from the same center and from other centers, so that relevant studies can be made, i.e. hot-biodiversity, specific richness, etc. The specimens stored in the center come from the collection campaigns carried out by the center and from exchanges with other centers. Once the specimen has been collected, it undergoes a process of preparation in folds and of identification. The identification process reflects the classification of the specimen and how it will be entered into the system databases. The specimens are lent to other researchers so that these may carry out their research work. These loans must be administered by the center’s administrative staff. It is therefore necessary to be able to control the loans made and their period of validity. Lending material entails: 1. The possibility of making a discovery in the samples which the center makes available to researchers. 2. Time being wasted for the center staff in charge of administering and controlling these loans and returns. Once the material has been studied, the researchers issue rectifications, if necessary, about its identification. Because of the complexity of nature, there are samples for which there are discrepancies in the identification or simply due to changes in the way the specimen has been identified. Consequently, the rectifications issued must be entered by M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518 509 the center’s operators, indicating that the specimen has been updated. Due to the value of the information in these centers, the problem arises of restricted access for users to certain types of information. As a result, different levels of access are created and each user can be allocated a level thereby enabling or refusing access to certain types of information. Having arrived at this point, we can see a series of characteristics which the system must offer in order for the objectives marked by a herbarium to be fulfilled from the administrative point of view: 1. Information administration (introduction of folds, revi- sions, genera, etc.) Fig. 1. Client/server architecture of the system. 2. Information consultation (folds, etc.) 3. Multimedia administration 4. User administration 5. Report creation 6. Issuing of labels 7. Loan control Another of the most important points for a herbarium, besides administration, is to be able to satisfy the demands of biodiversity studies. These studies use databases, and this can at times pose quite a complex task because of the way the data is displayed. We should therefore point out that the system must provide a series of services so that studies may be obtained about: 1. Specific richness (this is the number of species in a certain region or location) 2. Taxonomic complexity (complexity when it comes to identifying a specimen) 3. Study of the alpha/beta/gamma diversity (diversity within the habitats: alpha diversity; between the habitats: beta diversity; and for all the habitats being studied: gamma diversity) (Rosenzweig, 1995). 4. Orientation in the collection campaigns. This would entail a complex treatment of the data since this information is not directly accessible and would require a data acquisition and processing process in order to reach the desired objectives (Research Directions in Biodiversity and Ecosystem Informatics, 2001). 3. System design Among all the possibilities which currently exist to tackle the problem, the convenience of information systems was thought of because of the intrinsic nature of the problem. According to Lucas (1987), we can define information system as a set of organized procedures, which when performed, provide information for decision- making and/or control of the organization. The general theory of systems on which the information system analysis and design is based, indicates that it is necessary to consider the system to comprise smaller subsystems. The connection of the smaller systems with the larger systems forms a hierarchy which is characteristic of the theory of systems. It also shows us that we must have an overall view of the system, knowing that all the system components are interrelated and interdependent, with this being one of the most important tasks. We can therefore say that BioMen is an information system with a client–server architecture (Fig. 1) designed for herbarium management. Researchers and those inter- ested in this subject matter can gain online access to a virtual center which models the real behavior of the units which comprise the research center, and they are able to obtain all the information offered totally dynamically. The virtual users request (remote and/or hybrid) services which will enable them to perform all the intended operations within the system. BioMen offers a series of services which enable the users to have: – All the centralized information – Security protocols – Greater computational power The services offered might be: 1. Remote services: remote execution of processes and return of the results to the user. 2. Hybrid services: interaction between local and remote processes, e.g. integration of barcode readers. The majority of the services are remote, although there are some multimedia services which will need hybrid services (remote image processing and inclusion). As BioMen needs a representation of the domain knowledge, our system uses a service ontology described by means of the DAMLCOIL terminological system, and the services are described using OWL-S (OWL-S is an OWL-based web service ontology). This enables us to Fig. 2. Hierarchy of services. M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518510 organize the services on a graph and to provide a description of the services including the characteristics of each service (Fig. 2). The ontology is used by the system to enable the user to select the desired service, provide the necessary parameters, and the system is therefore in charge of executing the selected service. Due to the characteristics of the system, implemen- tation has been carried out using agent technology (Fig. 3). In the last 25 years, we have seen the appearance of several paradigms to design software systems such as procedural programming, structured programming, object orientation and component-based software. Agents (Weiss, 1999; Wooldridge, 2002) are now being championed as the next generation paradigm to design and build complex and distributed software systems. An agent-based architecture provides additional robustness, scalability, flexibility, and is particularly appropriate for problems with a dynamic, uncertain, and distributed nature. In particular, they seem to be the ideal computational model for developing software for Internet, and open networked system with no single controlling organization (Jennings, 2000). Lastly, Fig. 3. Way system acts given user interaction. architectures allow the incremental development of modular systems not only because of the modular nature of the agents, but also because of the possibility to incorporate legacy code by wrapping it within an agent interface. In a multi-agent system (MAS), agents interact with one another to achieve their individual objectives by exchanging information, cooperating to achieve common objectives, or negotiating to resolve conflicts. Alternative flexible patterns of interaction have been used such as the Contract Net Protocol (Reid & Smith, 1980), where a task is advertised by a coordinating agent and is assigned to the agent that makes the best bid. However, details of all possible interactions between agents cannot be foreseen a priori and consequently: 1. Agents need to be able to make decisions about their interactions at run-time, and 2. The organizational relationships between agents need to be represented explicitly (e.g. peer member in a team, manager, coordinator) by means of constructs such as roles, norms, and social laws. An agent is anything which can be observed sensing its environment using sensors and acting on this environment by means of effectors/actuators. The programming language which has been used is Java, enabling us to include a greater number of mobile devices and operating systems. The way the user interacts with the user is even simpler. The user interacts with the server using the HTTP protocol, performing the operations desired by means of a totally pleasing interface and without needing to have any additional tool installed. Once the web server has gathered the user’s request, it interacts with the multi-agent system in order to carry out the service requested by the user (Fig. 4) and returns the results of the service. The multi-agent system is made up of the following agents: – User – Request manager or coordinator – Service execution agents: i. Remote Fig. 4. Way of acting in a remote operation. M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518 511 ii. Hybrid The agents which form the multi-agent system use a blackboard architecture (Hayes-Roth, 1985; Kowalski and Kim, 1991; Nii, 1986a,b) for communication. The black- board is implemented by a series of tables. The agents use the blackboard to exchange the necessary information. Having looked at the operational logic of the system, we shall describe the characteristics of the system and how the requirements of a system for a herbarium have been solved. Given the security restrictions called for by the complex- ity and the intrinsic value of the information managed by the center, it is necessary to create at least three different levels of access, which will match the following user types: Fig. 5. Consultation without identification. – Internet user. Anyone wishing to consult the center’s data. The information available for this group of users is limited and controlled by the center’s staff. This is due to the quantity of useful information which can be misused. For example, knowing the location of protected species by conducting environmental impact studies without the respective authorization. We can therefore speak of a data user who does not receive information processed by the system. – Researcher. System user who can access more infor- mation than the Internet user and who in turn cannot perform any of the characteristic operations of the center management, such as the modification of information, insertion of loans in the system, etc. This user has access to the knowledge obtained by the system such as the representation of a specimen in the center, hot-biodiver- sity, etc. obtained automatically by the system, by means of the application of artificial intelligence techniques. – Center staff. Within this role we can in turn distinguish two different users according to their functions within the system: i. Operators. All the data contained in the system must be entered by someone, this would be the case of the operators. These will be responsible for inserting new folds into the system, new revisions, managing the images associated to the folds so that they can be consulted online, issuing labels for the preparation of folds, etc. ii. Managers. User in charge of supervising and mana- ging the center. This user will be authorized to access all the parts of the system, manage the accesses of the different users, obtain useful information for the center, such as for example, help in the planning of the collection campaigns by knowing which are the least represented specimens in the center. Fig. 6. Identified access in the system. 3.1. From the client’s point of view As we have already mentioned, the client communicates with the server using the HTTP protocol in order to execute the desired (remote and/or hybrid) services. In addition, the HTTP protocol will show the desired information and enter the parameters. There are two ways to access the system: – Without identification (see Fig. 5). The system is accessed as an Internet user and therefore access to the information is severely limited. – With identification (see Fig. 6). The system controls the different authenticated user types using login/password and allows more or less sophisticated services to be carried out according to the level of security allocated by the managers. The identified users access the system from the authorization window (see Fig. 6), beginning a new session. Having been identified in the system, menu systems are created which show the services allowed according to the level of security (see Fig. 7). Through the menu systems and the I/O interfaces, the system will receive requests and will provide the user with the requested information. By means of this access, whether identified or not, the client connects to the virtual center and the number of operations which can be performed will depend on the authorizations granted by the center managers. In this virtually centralized way, we allow the system user to overcome the problems of distance and time-wasting: 1. Loss of time. If we need any of the material contained in the center we will have two options: a. Ask the center managers for a loan. These loans will only be granted to institutions or researchers with Fig. 7. Different options according to the user. M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518512 a certain degree of confidence in the treatment of specimens. b. Visit the center so as to personally consult the desired specimens. As a result, the second of the centraliza- tion problems arises-that of distance. 2. Distance problems. The user/researcher will not always be in the same physical place as the center where the information needed is. Another of the problems which exists in both options is that only those materials which are physically available and which have not already been lent can be borrowed or accessed. In order to solve these problems, the need arises to centralize all the center’s information so that any user/ researcher can access the information (with more or fewer restrictions) simply, quickly, and comfortably. Conse- quently, a client/server has been used where all the information is virtually centralized, so that any user can consult the information needed without having to waste time by visiting the center or waiting for a loan. 3.2. From the server’s point of view From the server, pages are dynamically generated for each of the users, enabling all of the services required of the center to be performed. By maintaining a client/server structure, we provide solutions to the location problems which have previously been mentioned. Therefore, the server will act as a virtual center enabling as many services as those allowed to each user by the center managers. All these services are carried out and are managed by the multi-agent system totally transparently to the user so that a dynamic system is obtained with excellent features from the client’s and the server’s point of view. Any system of these characteristics must present a series of security features which guarantee the durability of the system. Merely thinking of the operator hours needed to enter the 130,000 records in the database from the large quantity of intrinsic knowledge to these records, which in their day have been provided by taxonomical experts who have identified and checked the specimens. For this reason, and as a safety measure, both to guarantee access to the information and to guarantee the safety of this in view of severe, enormous problems, the system is replicated identically on two servers which work with a policy of a daily update. This policy has been chosen rather than a real-time update as in this way we can take action if errors have been entered by the users. Errors are not therefore propagated in real time, and so there is a better treatment of the information. Let us look at the description of each of the services managed by the server. These will be accessible according to the user’s profile. From these services, we shall jointly deal with all the information system. 3.2.1. Access without authentication In this case, it will only be possible to access the system as a data user. This user will only be able to obtain from the system a set of lists obtained from the database. From the access without identification, any Internet user can consult the system and obtain a set of data entirely online. Not all the specimen information is shown as there is information about the specimens classified in the center which due to misuse of the information, and given its quality, could endanger the conservation of the species referenced. 3.2.2. Authenticated access In order to enable more important services to be performed within the system, an identified access has been provided. From the server, the security certificate is issued which the user must accept before continuing with the access. Once this has been accepted, the information is exchanged between the client and the server using a 128-bit SSL secure protocol. In this way, we avoid the possibility of data being filtered by malicious people. Once the user has been identified in the system, the system returns the series of services permitted. M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518 513 Below we shall describe some of the more important tasks which can be carried out within the virtual center. The presentation of the different services is organized from the introduction of the data in the system until the information processed by the information system is obtained. The following modules are described below: 1. Introduction of folds 2. Management of revisions 3. Label creation. 4. Management of images 5. Advanced consultation and/or consultation of folds 6. Management of loans 7. Administration 8. Treatment for the biodiversity Introduction of folds The herbarium works with plants and consequently it is necessary for these to be treated and conserved. The container with one or more specimens is called a fold. The fold univocally identifies the sample in the center and can contain one or more specimens. The information about the fold (number of specimens, taxonomic names, collectors, exicata, location, UTM, height, etc.) is entered into the system by means of the corresponding service. Once the user has selected the desired service, the set of parameters to be provided is returned. When these parameters are entered and returned to the system, the agent in charge performs the remote service, returning the result of the operation to the user. Management of revisions Because of the interest required by other researchers for samples from other centers, folds are lent to other centers or researchers. The taxonomic name of a specimen is not easy to identify due to the complexity of nature. For this reason, there may be discrepancies between different researchers about the taxonomic denomination given to the specimen. Each of the researchers’ proposals, in time, is recorded in the system as the current taxonomic denomination (the last in the instant t). This information will be very important Fig. 8. Sample o when it comes to studying the taxonomic changes in order to be able to make the biodiversity studies. Creation of labels Once the fold has been entered into the system, the center staff can obtain a label which is placed on the fold, so that part of the information and the fold identifier can be seen. For the identification of the fold, an alphanumeric chain has been used. Each fold is assigned a barcode (Code 39) so that the subsequent treatment of the fold in the system will be a much simpler process (see Fig. 8). These labels will mostly be used for loan management. By adding the barcodes, certain hybrid services (barcode reader and server) can be carried out. Multimedia management As we have already mentioned, one of the features of the service which distinguishes it from others is the incorpor- ation of a multimedia service which enables the user to obtain much more detailed information. There is a multi- media element associated to each fold, for example, fold images, video of the habitat, etc. which any authorized user can consult. This service would require the execution of local services (reproducer, image viewer, etc.). From the point of view of information management and incorporation, users can pre-process the information which they wish to incorporate for a given fold. For example, scanning the center’s folds and associating an identifier with the scanned image, recording a video about the collection of the specimen in the field, etc. This is one of the solutions offered by the system to improve the work of the center’s staff. This multimedia service is available to all system users. In this way, the multimedia consultation of a specimen is made possible without the need for a corresponding loan request. Therefore, 1. f a The researcher can consult the specimen’s multimedia information the moment the center’s staff carry out the operation in the system. 2. There is a reduction in the number of loans which the center must make to the researchers. label. M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518514 3. In turn, there is a better conservation of the center’s material. Advanced consultation and/or consultation of folds If we make a consultation using the fold’s identifier, we will only obtain information about that fold, but what happens with the information contained in these? From the researcher’s point of view it is much more necessary to be able to make a consultation using the information contained within a fold than for the existence of a fold. For example, existing specimens for a UTM and/or above a certain height, etc. This is, as we have already mentioned, the main distinction between a library and a herbarium, in that the information which is useful to the researcher is the information which there is within the fold and not the fold in itself. It is therefore as if we were asking about the information contained in each book in a library. From this service, any type of information existing in the system can be searched for, and the result can be obtained both in HTML and PDF so that it can be easily exported. The online access to the information when consultations are made bestows the power that the herbarium staff and researchers need. From the advanced consultation, the center’s staff can begin to make a loan to the center requesting it. If the center needs to lend all the folds containing the specimen Pinaceae Pinus baciano, in a normal, non-computerized process, the staff would need to go to the storeroom, look through the folds one by one in order to select those requested by the researcher. If the system is used, the fold identifiers containing this specimen can be obtained and in turn, the loan service of the system can be activated merely by entering the loan recipient’s data and recording this in the system. Loan management As we have already mentioned, the center’s material is vitally important for everyone, center and researchers alike, Fig. 9. User adm and it is therefore essential to control the loans made so that these are returned within the stipulated period. Therefore, the loan service in turn carries out the following services: 1. inis Loan registration. The loan recipient and all the materials lent are recorded in the system. This new loan has an identifier within the system so as to be able to control the loan in consecutive processes. 2. List of loans. List of loaned material for a loan identifier. 3. Loan return. This enables the complete or partial return of the loan to be entered into the system. From this section, we control which of the loaned folds have not been returned. 4. Loan control. Control of the validity period of the loan. In this way, we will be able to demand that the loan be returned. It is possible that this operation is performed automatically by the system once the return period has ended and if there is no record of the return. The system sends an e-mail to the loan recipient informing them of the need to process the return of the loan. Administration Another just as important service is that of system administration. Given the architecture of the system and the way in which it is accessed, it is clearly necessary to have an exhaustive control of the operations allowed to each user within the system. The system contains a list of operations and users. Using the checkbox, the system’s service administrator can control the services which these can make in the system (Fig. 9). From this service, we can also request services to import data from other databases, user registrations/cancellations, etc. Treatment for biodiversity By means of a series of remote services, the user can request information about: tration. M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518 515 1. Taxonomic complexity (Magurran, Moreno) (Halffter et al.) 2. Specific richness (Magurran, Moreno). 3. Orientation in collection campaigns 4. Study of the alpha/beta/gamma diversity These remote services show the user the desired information, using the information provided by other agents who are constantly processing the data contained in the databases. We shall now see how this information can be obtained and we solve the existing problems. 4. Biodiversity As we have mentioned already, there is a large amount of interesting information in the center’s databases. This information enables important improvements to be made in the quality of botanists’ work. Nevertheless, the information is not usually directly accessible since it needs to be processed from the database. As a first approach to the solution of this problem, we can recover and process information in order to obtain new knowledge and determine: 1. Taxonomic complexity (Magurran, Moreno) (Halffter et al.) 2. Specific richness (Magurran, Moreno). 3. Orientation in collection campaigns 4. Study of the alpha/beta/gamma diversity The main disadvantages of obtaining the taxonomic complexity are: 1. Existence of a large volume of data 2. Redundancies present in the information 3. The existence of synonyms in the database Because of these problems, it is not possible to perform the taxonomic complexity studies directly. In order to look at this problem in more detail, we shall consider the following example. Below we shall show the identification of a small sample of the specimens contained in the database. The specimen’s name, in this case, comprises the family, genus and species: – Cruciferae Alyssum spinosum – Cruciferae Hormathophylla spinosa – Cruciferae Ptilotrichum spinosum If we want to know the number of different specimens, when the count is made in the database, we would obtain three specimens. However, according to Flora Ibérica (1996), the three names refer to the same specimen (Cruficerae Hormathophylla spinosa). In addition, the order by which the name (identification) has evolved (Alyssum/Ptilotrichum/Hormathophylla) is established. For this reason, as we mentioned before, there are synonyms in the database. This makes it impossible for us to obtain the information necessary for biodiversity studies (different number of specimens in one area, e.g. for specific richness studies). Below, and in view of the importance which the problem of synonymy has within the research center, we shall attempt to resolve the problem. In order to do so, there are two possible courses of action: 1. By creating a synonym database. This alternative accelerates the processing work. However, it offers a series of drawbacks: a. The size of the synonym table is very large, since there is a great variety of species. b. The table would have to be compiled by an expert. The expert would have to carry out a repetitive and tedious task. 2. By studying the evolutions. The name we give to the change in the denomination of a specimen is evolution. We shall explore this in greater depth later. This task can be carried out without an expert having to intervene and enables us to obtain the sequence of the change in the identification. Another piece of extremely interesting information relates to orientation in the collection campaigns. This provides the center with advantages both in terms of finances and documentation. The idea is to provide information about the types of specimens needed for the center to be complete and well represented. For example, if the number of specimens in the center is low, it might be that: 1. There really are few specimens. 2. The specimen has been lent to other research centers. It is therefore necessary to inform the center of the specimens which need to be collected so that the center is complete and well-represented. The information might be: 1. Specimens which are not particularly represented and/or below minimum levels. 2. The best path to follow in order to collect the specimens. In order to obtain this knowledge, three intelligent agents have been used (according to Wiener’s definition of intelli- gence) which will act in turn within the multi-agent system described above. These agents would constantly be observing the media (databases) (Konolige, 1982), acquiring and proces- sing the information in order to achieve the necessary information. The agents deposit the information in the system, using the blackboard, so that the users who so desire can access it by means of the previously described corresponding services. The first of the agents, called the revision agent, is responsible for studying all the revisions for a specimen. This result is taken advantage of by the agent called M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518516 the specific richness agent. This agent obtains the set of synonyms contained in the database. This information is necessary in order to count the different specimens which there are in the database. In turn, the information obtained by the two previous agents is used by the agent called the collection campaign orientation agent. This agent issues a report of those little represented specimens in the center. 4.1. Description of the agents The main innovation of the proposed solution is the treatment of the synonyms existing in the database. This is an important problem and one covered by an international project—Species 2000. In order to solve this problem locally, we shall study the concept of synonymy and how this can be solved. The method we propose does not need the expert to intervene directly as it is based on the study of the history of the specimens contained in the database. In order to identify the synonyms, we use the concept of evolution. We shall define evolution as the change in the identification of a specimen. This evolution means that specimens are stored in the database with different names (i.e. synonyms). This means that the possibility of counting how many different species there are in the database is a complicated task without any real values. In order to obtain the correct number of different specimens, we use the following agents. 4.1.1. Revision agent This agent is responsible for studying the number of revisions which each specimen has. In this way, the material most studied by botanists is obtained. Once we have the number of revisions for each specimen, we can conduct the relevant evolution studies. The agent places this information on the blackboard. The AgenteRevisiones table is used to enter the results. 4.1.2. Taxonomic complexity and specific richness agent Information about taxonomic complexity The solution to the synonymy problem was found by studying the evolutions which a certain specimen under- goes. Due to the formal complexity of the evolution and synonymy problems, a botany-based complexity which is therefore outside the field of this work, we shall present the concept using an example: The different blocks marked as A, B, C, D and E are the different identifications which a specimen undergoes. The arrows indicate the evolution to another identification. Let us examine the example more closely. Specimen 1 was identified as A but later, after being researched by other researchers, it has been identified as B, C and D. These identifications are treated as Evolutions and therefore, the identifications A, B and C are synonyms of D. These values are entered in the EvolutionAlert table. This table is managed by the agent. The following information is entered: Specimen 1: A/D, evolution 0. Specimen 1: B/D, evolution 0. Specimen 1: C/D, evolution 0. The agent enters the following fields: 1. Taxon. This indicates the number of the specimen. In this case 1. 2. Antecedent. Previous names of the specimen. For example: A, B or C. 3. Consequent. This indicates the name of the evolution. In this case D. 4. Evolution. This field can take two values: a. When evolution is 0. This indicates that the specimen will not be identified in another way. b. When evolution is 1. This indicates that if the specimen is studied, it may change its identification to that indicated by the Consequent field. This enables us to inform the botanist of the specimens to be revised so that the center’s material is completely up- to-date. This information is based on the revisions which have already been made to other specimens. Below we shall place this information on the blackboard, in particular in the taxones table, so that it may be consulted by other agents. In this way, we shall enter the following taxon in this table: Specimen 1: D. Specimen 2 is a synonym since there is another specimen which has changed from state C to state D. Therefore, we would enter in the AlertasEvolución table: Specimen 2: C/D, evolution 1. In this case, the evolution value of 1 indicates that if specimen 2 were studied, it would certainly be identified as D. Therefore, we enter it as a possible evolution. The following specimen, specimen 3, has some revisions identified as A and B. As in the AlertasEvolución table, we have stored the fact that state B can become state D, due to the sequence of evolutions which specimen 1 has. We therefore add the following tuple to the alertasEvolución table: Specimen 3: B/D, evolution 1. M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518 517 We should remember that in the taxones table there is a single tuple, indicating that of the specimens studied we only have one different taxon. In the following specimen, number 4, we can see that as state D is revised to state E, state D becomes a synonym of E, and it is therefore necessary to revise the data stored in the AlertasEvolucion table and in turn, to update the identified taxa. As a result, the tables would remain as follows: AlertaEvolución table taxones table Specimen 1: A/E, evolution 0 Specimen 4: E Specimen 1: B/E, evolution 0 Specimen 1: C/E, evolution 0 Specimen 1: D/E, evolution 1 Specimen 2: C/E, evolution 1 Specimen 3: B/E, evolution 1 Specimen 4: D/E, evolution 0 Having taken a general look at the example, we shall use the data which we have shown previously in order to see how the agents would act: Fold GDAC2745: Revision 1: Cruciferae Ptilotrichum spinosum (L.) Boiss. Revision 2: Cruciferae Hormathophylla spinosa (L.) Küpfer Fold GDA28909: Revision 1: Cruciferae Alyssum spinosum L. Revision 2: Cruciferae Alyssum spinosum L. Revision 3: Cruciferae Ptilotrichum spinosum Boiss. When the multi-agent system acts, we would obtain the following final situation. AlertaEvolución table: – Fold GDA28909: Cruciferae Alyssum spinosum/Cru- ciferae Hormathophylla spinosa, evolution 0. – Fold GDA28909: Cruciferae Ptilotrichum spinosum/ Cruciferae Hormathophylla spinosa, evolution 1. – Fold GDAC2745: Cruficerae Ptilotrichum spinosum/ Cruciferae Hormathophylla spinosa, evolution 0. Taxa table: – Fold GDAC2745: Cruciferae Hormathophylla spinosa As we can see, we obtain the desired result without needing to produce any table which contains these synonyms and which would involve a great deal of work for the specialist. So far, all the solutions provided for the synonymy problem have involved the construction of the synonym table Species2000. We therefore believe that we have provided an easy and innovative way for the researcher to obtain very important information which does not entail any expense for the center using it. This information determines the taxonomic complexity and the richness of species for any area and consequently enables biodiversity studies to be made. Information about specific richness Once the synonym problem has been solved, we can study the specific richness of the areas covered by the center. In this way, botanists have a large amount of useful information for their hot-biodiversity studies. In turn, the information which the specific richness agent provides can be crossed with the previously mentioned project macro, Species2000. This project attempts to solve the problem of synonymy as a whole. If we cross the information developed by the agent with the information provided by Species2000, the category which a research center has can be established, indicating the number of series of species which this center covers. Suitably planned centers can therefore be combined and in accordance with the needs of the project to be embraced and covering the desired species. 4.1.3. Collection campaign orientation agent If we combine the information obtained by the existing agents, we can obtain yet more advantages. We obtain the necessary specimens to be collected so that we may have a complete center. The systems issues reports filtering the synonymy problems. In order to achieve a complete center, both on a geographical level and in terms of plant groups, we cross the information obtained by the agents with a Geographical Information System. This fact offers a number of advantages which will enable the center to obtain better results from the information provided by the multi-agent system. M. Delgado et al. / Expert Systems with Applications 28 (2005) 507–518518 When the crossed information is obtained with a geographical information system, we obtain different ways of solving the problem of having a complete center: 1. Obtaining information from the areas which have not been visited for collection, thereby enabling each geographical point to be reflected in our center. 2. To obtain the necessary routes and itineraries in order to collect specimens which are little represented in the center. 5. Conclusions In this paper, we have described our experiences of constructing BioMen, an information system executed on Internet and developed for herbariums. The constructed system incorporates all the center’s needs, uses a multi- agent system which makes the system much more dynamic and easy to maintain. In addition, this is done entirely independently of the user who does not need to know how an ontology operates or how the agents must communicate with one another. BioMen is a totally operational system which uses the newest technologies: – Access to the system by means of a web browser. – Javae Servlet technology (Sun Microsystems Corpor- ation, 2001; Hall, 2001) – Apache Server (The Apache Software Foundation, 2001a) – Apache Jakarta Proyect Tomcat (The Apache Software Foundation, 2001b) – Java Agents (Borland Corporation, 2001; Eckel, 2000). – JDBC for communication with the databases. – Any JDBC-compatible database manager (MySql, Oracle). – Semantic web and ontologies. DAMLCOIL and OWL-S The majority of software tools are free and this makes the system much more attractive and enables it to be more standardized. In this document, we propose a solution to the problem of studying hot-biodiversity. By means of artificial intelligence techniques, such as agents, complex and extremely inter- esting tasks have been performed such as the study of hot- biodiversity, cost minimization and orientation in the collection campaigns. We are currently working on transferring the concept of collection from the virtual center where it is not necessary for there to be a physical place to store the specimens, thereby creating a totally virtual center. In order to do so, the latest technologies are being studied: PDA’s, GPRS, GPS, multimedia transmission using mobile devices, etc. References Borland Corporation. (2001). JBuilder Enterprise 6.0.438.0 Documen- tation. Available at: http://www.borland.com, last visited December 26th 2003. Castroviejo, et al. (1996). Flora Ibérica. Real Jardı́n Botánico, CSIC, Vol. II. Cruciferae-Monotropaceae pp. 193–195. Eckel, B. (2000). Thinking in Java2nd ed. Prentice-Hall, Englewood Cliffs, NJ.. Available at: http://planetpdf.com/. Halffter, Moreno y Pineda, Manual para evaluación de la biodiversidad en Reservas de la Biosfera, MT and SEA. Hall, M. (2001). Core Servlets and JavaServlet pages. Englewood Cliffs, NJ: Sun Microsystems Press/Prentice Hall. Hayes-Roth (1985). A blackboard arquitecture from control. Artificial Intelligence, 26(3), 251–321. Henry, C. & Lucas, J. R. (1987). Sistemas de Información, Análisis, Diseño y Puesta a punto. Paraninfo. Jennings, N. R. (2000). On agent-based software engineering. Artificial Intelligence, 177, 277–296. Konolige. (1982). A first-order formalization of knowledge and action for a multi-agent planning system. Kowalski, K. (1991). A metalogic programming approach to multi-agent belief. In V. de Lifschitz (Ed.), Artificial intelligence and mathematical theory of computation: Papers in Honor of John McCarthy (pp. 231– 246). Boston, MA: Academic Press, 231–246. Lane, M. A., Edwards, J. L., & Nielsen, E. (2000). Biodiversity informatics: the challenge of rapid development, large databases, and complex data (keynote). In VLDB 2000, Proceeding of the 26th international conference on very large data bases. Magurran. Diversidad. Ecologı́a y su Medición. Moreno. Métodos para medir la biodiversidad, MT & SEA. Nii (1986a). Blackboard systems: The blackboard model of problem solving and the evolution of blackboard architectures. Nii (1986b). Blackboard systems (part two): Blackboard application systems, blackboard systems from a knowledge engineering perspective. Pando, F. (1991). El Herbario de Criptógamas del Real Jardı́n Botánico y sus bases de datos. IX Simposio Nacional de Botánica Criptogámica, Salamanca. Reid, G., & Smith, T. (1980). The contract net protocol: high load communication and control in a distributed problem solver. IEEE Transactions on Computers, 29(12), 1104–1113. Research Directions in Biodiversity and Ecosystem Informatics. (2001). Report of an NSF, USGS, NASA workshop on biodiversity and ecosystem informatics. Rosenzweig, M. L. (1995). Species diversity in space and time. Cambridge: Cambridge University Press. Species2000, Available at: www.species2000.org, last visited March 26th 2004. Sun Microsystems Corporation (2001). Servlet Specification v 2.3. Available at: http://www.sun.com/servlet/, last visited December 26th 2003. The Apache Software Foundation. (2001a). Apache HTTP Server Version 1.3.29 Documentation. Available at: http://www.apache.org, last visited December 26th 2003. The Apache Software Foundation. (2001b). Apache Jakarta Project Version 4.0 Documentation. Available at: http://jakarta.apache.org, last visited December 26th 2003. Weiss, G. (1999). MultiAgent systems. A modern approach to distributed artificial intelligence. Cambridge, MA: MIT Press. Wooldridge, M. (2002). An introduction to multiagent systems. New York: Wiley. http://www.borland.com http://jakarta.apache.org http://jakarta.apache.org http://jakarta.apache.org http://jakarta.apache.org http://jakarta.apache.org BioMen: an information system to herbarium Introduction System analysis System design From the clients point of view From the servers point of view Biodiversity Description of the agents Conclusions References 
work_qgmytu4atfhsvf6zpxcsivfqni	----	Implicit Untangling: A Robust Solution for Modeling Layered Clothing HAL Id: hal-02129156 https://hal.archives-ouvertes.fr/hal-02129156 Submitted on 16 May 2019 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Implicit Untangling: A Robust Solution for Modeling Layered Clothing Thomas Buffet, Damien Rohmer, Loic Barthe, Laurence Boissieux, Marie-Paule Cani To cite this version: Thomas Buffet, Damien Rohmer, Loic Barthe, Laurence Boissieux, Marie-Paule Cani. Implicit Un- tangling: A Robust Solution for Modeling Layered Clothing. ACM Transactions on Graphics, Asso- ciation for Computing Machinery, 2019, Proc. ACM SIGGRAPH, 38 (4), pp.Article No. 120:1-12. �10.1145/3306346.3323010�. �hal-02129156� https://hal.archives-ouvertes.fr/hal-02129156 https://hal.archives-ouvertes.fr Implicit Untangling: A Robust Solution for Modeling Layered Clothing THOMAS BUFFET, Inria Rhônes-Alpes, France and LIX, Ecole Polytechnique, CNRS, IP Paris, France DAMIEN ROHMER, LIX, Ecole Polytechnique, CNRS, IP Paris, France LOÏC BARTHE, University of Toulouse - IRIT - CNRS, France LAURENCE BOISSIEUX, Inria Rhônes-Alpes, France MARIE-PAULE CANI, LIX, Ecole Polytechnique, CNRS, IP Paris, France Fig. 1. Our method is able, from a set of garments possibly exhibiting deep interpenetration, to compute an untangled state, ie. a guaranteed intersection-free configuration, even while considering the thicknesses of cloth layers. Animations can then be launched. We propose a robust method for untangling an arbitrary number of cloth layers, possibly exhibiting deep interpenetrations, to a collision-free state, ready for animation. Our method relies on an intermediate, implicit rep- resentation to solve the problem: the user selects a few garments stored in a library together with their implicit approximations, and places them over a mannequin while specifying the desired order between layers. The intersecting implicit surfaces are then combined using a new family of N-ary composition operators, specially designed for untangling layers. Garment meshes are finally projected to the deformed implicit surfaces in linear time, while best preserving triangles and avoiding loss of details. Each of the untangling operators computes the target surface for a given garment in a single step, while accounting for the order between cloth layers and their individual thicknesses. As a group, they guarantee an intersection- free output configuration. Moreover, a weight can be associated with each layer to tune their relative influence during untangling, such as leather being less deformed than cloth. Results for each layer then reflect the combined effect of the other layers, enabling us to output a plausible configuration in Authors’ addresses: Thomas Buffet, Inria Rhônes-Alpes , Montbonnot-Saint-Martin, France, 38330, 655 Avenue de l’Europe , LIX, Ecole Polytechnique, CNRS, IP Paris, France, 91120, Palaiseau, 1 rue Honoré d’Estienne d’Orves Alan Turing building; Damien Rohmer, LIX, Ecole Polytechnique, CNRS, IP Paris, France, 91120, Palaiseau, 1 rue Honoré d’Estienne d’Orves Alan Turing building; Loïc Barthe, University of Toulouse - IRIT - CNRS, France, Toulouse; Laurence Boissieux, Inria Rhônes-Alpes , Montbonnot- Saint-Martin, France, 38330, 655 Avenue de l’Europe; Marie-Paule Cani, LIX, Ecole Polytechnique, CNRS, IP Paris, France, 91120, Palaiseau, 1 rue Honoré d’Estienne d’Orves Alan Turing building. © 2019 Copyright held by the owner/author(s). Publication rights licensed to ACM. This is the author’s version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in ACM Transactions on Graphics, https://doi.org/10.1145/3306346.3323010. contact regions. As our results show, our method can be used to generate plausible, new static shapes of garments when underwear has been added, as well as collision-free configurations enabling a user to safely launch animations of arbitrarily complex layered clothing. CCS Concepts: • Computing methodologies → Collision Detection; Volumetric Models. Additional Key Words and Phrases: Implicit surfaces, Cloth untangling, Collision processing, Cloth animation ACM Reference Format: Thomas Buffet, Damien Rohmer, Loïc Barthe, Laurence Boissieux, and Marie- Paule Cani. 2019. Implicit Untangling: A Robust Solution for Modeling Layered Clothing . ACM Trans. Graph. 38, 4, Article 120 (July 2019), 12 pages. https://doi.org/10.1145/3306346.3323010 1 INTRODUCTION The efficient animation of dressed virtual characters is an essential step in 3D production pipelines. Tremendous efforts have been dedicated to achieve high quality dynamics for individual pieces of clothing, based on physically-based simulation models and efficient collision detection and response. Extending these animations to multiple layers of garments is however difficult, starting with the challenging problem of providing an initial collision-free state— which is mandatory for the animated cloth surfaces to behave as expected. In standard production, 3D garments are modeled on top of generic mannequin and poses. Reusing garments for slightly different mannequin, or launching the simulation in different poses, requires manual editing to set them in the expected collision free ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. https://doi.org/10.1145/3306346.3323010 https://doi.org/10.1145/3306346.3323010 120:2 • Buffet, Rohmer, Barthe, Boissieux, and Cani state, and avoid a so-called tangling effect. This becomes very tedious when several layers of garments with extended contact surfaces are involved, since any deformation of a layer to avoid a collision could cause an intersection with another layer. For this reason, many clothing simulations, from video games to film production, still make use of a single garment layer. This results in both inaccurate rest states—since the deformation due to the underlying cloth layers are not accounted for, and wrong dynamics since the friction between neighboring cloth layers is neglected. In this paper, we propose a robust solution to automate the te- dious process of untangling intersecting garment layers. It can be used to quickly model plausible static shapes of garments given the underwear, and to safely initialize layered clothing animations (see Figure 1). Our input is an arbitrary number of garment models. We pre- compute their implicit approximation as the intersection between an iso-surface in a scalar field and an influence zone in a co-variant field. When placed around a character, these garments typically exhibit various levels of interpenetration. We introduce an automatic method to compute a collision-free state, where contact is precisely modeled between initially colliding layers while taking into account the desired theoretical thickness of material (used to set up minimal distances between meshes). This is done using a set of closed-form N-ary composition operators, which deform the input fields while accounting for the user-specified order and thicknesses of the layers. In addition, the user can set weights in order to play with the relative stiffness of layers during deformation. After scalar field composition, garment meshes are projected in linear time with respect to their number of vertices, while tangentially relaxed to better preserve triangle shapes. This yields new shapes for top layers that take all the layers underneath into account, and provides interpenetration-free initial states for all garments, enabling a user to launch a simulation. In summary, our contributions include: • A new method for approximating a garment using a pair of scalar fields, given that garments are open surfaces that only span a part of the associated closed iso-surface. • A closed-form solution for the N-ary operators enabling us to output untangled surfaces from a number of possibly in- tersecting layers. • An adaptation of the projection/relaxation method proposed by Vaillant et al. [2014] for accurately projecting garment meshes within the scalar fields, yielding penetration-free con- figurations. 2 RELATED WORK Our method is dedicated to the modeling of layered clothing, where inter-penetrations should be changed into contact regions with epsilon distance between meshes. We therefore first review contact modeling methods in cloth animation. Since our solution relies on implicit surfaces in scalar fields, we then review the implicit methods that have been used for modeling contact between non- nested volumetric bodies. 2.1 Contact modeling in cloth animation While lots of methods are dedicated to efficient collision detection between rigid bodies [Fares and Hamam 2005] and deformable mod- els [Teschner et al. 2004] , we focus here on approaches dedicated to cloth animation, and enabling not only the detection of collisions but also their resolution. A first category of methods are dedicated to handling collision between a cloth surface and a volumetric body model [Guan et al. 2012], [Chen et al. 2013] ,[Sun et al. 2016] and [Wong et al. 2018]. They use some projection mechanism to locally deform the cloth and move it outside of the volume, therefore robustly handling cloth- to-body collision. However, they cannot directly extend to collisions between multiple cloth layers. A second category of methods, first introduced by Bridson et al. [2002], and then further extended [Harmon et al. 2008; Selle et al. 2009], prevent collisions based on the history of deformations: penetrations detected at a given frame are corrected by looking back- ward in time to a past collision-free state. In particular, Harmon et al. [2008] achieve asynchronous correction leading to extremely ro- bust and stable collision handling, although computationally costly. History-based correction have also been combined with penalty forces to increase stability [Tang et al. 2012], ray-tracing for accu- rate detection of intersection [Lehericey et al. 2015], and fast GPU computation [Tang et al. 2016, 2018]. Müller et al. [2015] propose an efficient unified approach dynamically handling collision detection and contact modeling by combining tetrahedrization of thin air layers between cloth surfaces and position based dynamics. Purely geometric approaches were also used to correct for existing colli- sions [Ye et al. 2017]. Most of these approaches are able to handle multiple layers of cloth and avoid self-intersection, however, they require an initial collision-free state as input. Therefore, they are not applicable to our problem. The last category of approaches are dedicated untangling meth- ods, able to correct existing penetrations without any knowledge of past states. In this case, a heuristic must be chosen in order to correct the ambiguous surface geometry. Previous methods have studied collisions between two cloth layers using repulsive force-based correction [Baraff et al. 2003], as well as geometric displacement in order to minimize the length of the collision contour [Volino and Magnenat-Thalmann 2006; Ye et al. 2017, 2015; Ye and Zhao 2012]. While repulsive force-based approaches can be expressed on arbitrary numbers of layers, they do not provide any guarantee of efficient convergence toward a collision-free state. Geometrical approaches provide this guarantee, but were restricted so far to collisions between two cloth layers only. A naive solution for gen- eralizing to the multi-layers case consists in iteratively applying the untangling process to consecutive pairs of cloth layers, from the deepest layer in contact with the character’s body to the most external one. In contrast with our N-ary approach, this strategy cannot, however, take the influence of upper layers into account when deforming a given layer. 2.2 Implicit surfaces Implicit surfaces are well known for their ability to ease collision detection and model contact between non-nested volumetric bodies: ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. Implicit Untangling: A Robust Solution for Modeling Layered Clothing • 120:3 Fig. 2. Processing pipeline: Implicit reconstruction of all garments are computed as a preprocess (left). At run-time, the user selects a number of garments, sets their order, thicknesses and weights (center). The associated fields are composed using our new untangling operators, leading to a variety of possible results after mesh projection (right). For instance, the T-shirt (low weight during deformation) can either be worn below or above the thicker leather jacket. A first remark is that collisions can be detected in linear time: Let f1 and f2 be two scalar fields andS1,S2 their associated 0-isosurfaces, the inside of the volume being defined by Vi = {p ∈ R3/fi(p) < 0}. Then for any point p on S1, evaluating the sign of f2(p) provides a simple collision test with V2, while the value of |f2(p)| gives some information about the penetration depth. Another benefit of implicit surfaces is their ability to be composed into new surfaces [Bloomenthal 1997], using a variety of combina- tion operators between fields. In addition to the well known blending operators, composition has been used to accurately model contact surfaces between colliding volumetric bodies. Cani [1993] linearly combines the input field functions in the region of interpenetra- tion. Vaillant et al. [2014] introduce more complex binary operators for modelling contact between colliding skin parts of an animated character. Building on advanced implicit composition using field gradients [Gourmel et al. 2013], Angles et al. [2017] show that a variety of operators can be created on the fly from sketch-based user input, enabling, among other applications, the modeling of a variety of contact situations between volumes. Most of these works focus on binary operators numerically precomputed in a 2D or 3D discrete grid, which cannot be easily extended to n-ary composition. Moreover, none of them handles nested implicit volumes. In addition to trivial operators such as maximum, minimum [Sabin 1968] and sum of field functions [Blinn 1982], n-ary composition operators include set-theoretic operations [Pasko et al. 1995], super- elliptic blends [Ricci 1973], extended convolution operators for topol- ogy control [Zanni et al. 2015] as well as various n-ary blends with range control [Barthe et al. 1998; de Groot et al. 2009; Hsu 2018; Hsu and Lee 2003]. The definition of n-ary operators can be tedious and their theory is not as deeply studied as the one of binary composi- tions. To our best knowledge, none are tackling our core challenge, namely contact modeling between formerly intersecting, nested implicit surfaces. 3 OVERVIEW We aim at untangling an arbitrary number of intersecting garments. The key insight is to take benefits from the ability of volumetric representations to model contact through the composition of scalar fields. To achieve this, garments need to be converted to some im- plicit representation, while contact modeling by iso-surface compo- sition needs to be extended to the untangling of an arbitrary number of nested surfaces. 3.1 Notations and input In all this work, the implicit surfaces Si we consider are closed sur- faces defined as the 0-iso-surface of a scalar field fi . By convention, we define the interior of the volume within Si by {p ∈ R3 | fi(p) < 0}. This is the convention used by HRBFs (Hermite Radial Basis Functions) [Wendland 2005] on which we will build on. They enable a user to reconstruct closed implicit surfaces from a set of sample points and the associated normal directions. We consider a set of N predefined garment models given by their input meshes, which are already wrapped over an input character body. The user specifies the desired order between these garments, where layer 1 is the closest to the body and layer N is worn above the other layers. Virtual thickness values ti are specified for each layer, enabling the tuning of the minimal distance between the centered meshes representing each garment, while weights wi > 0, acting as a stiff- ness parameter, enable the tuning of their relative influence when deformed during contact. Small values of wi mimic very flexible material that tends to match the geometry of other surfaces in con- tact regions, while large values mimic stiffer layers that tend to keep their shape. Note that the body is handled as a fully rigid layer wi =+∞. The deformation we want to output for each garment is to account for the combined effect of all the other layers (above and under) both in terms of thicknesses and weights. 3.2 Processing pipeline and challenges Our processing pipeline, depicted in fig. 2, includes a prepossessing step—namely implicit approximation of all garments in a library, which can be done independently from a specific set-up. At run time, the user selects layers, sets their parameters, and untangles them. More precisely, for each mesh vertex, gradient descent and tangential relaxation are interleaved to deform each garment to- wards a target implicit surface, computed by applying an untangling operator to the input fields. Let us detail the main problems to be solved. ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. 120:4 • Buffet, Rohmer, Barthe, Boissieux, and Cani Implicit representation of garments: While implicit surface recon- struction from sample points is a solved problem for closed surfaces such as the character’s body, extending it to open surfaces with boundaries such as garments is a challenge. In particular, the new representation needs to generalize the notion of inside and outside to these open surfaces, to be able to detect the interpenetration regions between layers, in which corrections need to be applied through the field composition operator. Therefore, our first contribution, presented in Section 4, is a gen- eral solution for approximating open surfaces with boundaries us- ing a double implicit representation: we combine the field fi which embeds the mesh among iso-surfaces with a co-variant field hi , computed from the open borders of the mesh. This enables us to define the region where a garment layer could cause intersections, accounting for the fact that it does not cover the full 0-iso-surface of fi . Untangling operators for nested implicit surfaces: Once the user sets a new configuration with a character and some ordered lay- ers of garments from the library, implicit composition is used to untangle the associated fields, generating contact surfaces through deformation in the regions where garments interpenetrate (see fig. 2- middle). Here, we are not looking for a single composition operator but for N operators Oi , parameterized by the layers weights wi , and able to generate N new fields f̂i whose respective 0-iso-surfaces coincide in the previously intersecting regions while remaining intersection-free elsewhere. We introduce a closed-form solution for these operators, enabling us to apply them whatever the number of colliding layers and their relative rigidities. Section 5 first explains how these operators can be defined in the general case of N layers, before detailing it for the specific case of two and three layers. Using the deformed fields for updating the garments: In our case, the untangling composition is not to be applied to some abstract, nested implicit surface, but to meshes representing garments. Simi- larly to Vailllant et al [2014], output meshes are computed by dis- placing vertices of the original input meshes along the gradient of their respective field ∇ f̂i until reaching the zero value of f̂i . Several changes however need to be made: first, the meshes should not lie on the same 0-iso-surface in contact regions, but at some offset distance from each other, to avoid visual artifacts and allow subsequent physically-based animation. Secondly, some tangential relaxation is required to ensure that the local mesh deformation remains minimal while well approximating the 0-iso-surface of f̂i . Indeed, over-elongated triangles could cause geometric intersection even while implicit layers do not intersect, and could lead to unstable subsequent simulations. Section 6 explains our practical solution for applying the deforma- tion to meshes, which involves an extended composition operator accounting for the required thicknesses ti of each layer to guaran- tee nested target surfaces, and a new relaxation method that we interleave with gradient descent to reduce tangential distortions. 4 IMPLICIT APPROXIMATION FOR GARMENTS While we use standard HRBFs to approximate the character’s body, we need to extend implicit reconstruction to the case of open sur- faces with boundaries for handling garments. The main challenge is to characterize regions where the garment might interpenetrate with others during field composition, used to test for interpenetra- tion: indeed, using the full 0-iso-surface of fi to approximate a cloth layer would not work, since extra collisions could then be wrongly detected, leading to unwanted deformations of the other layers, such as in the region below the jacket and t-shirt in Fig. 3. The key idea to achieve this is to use a second implicit volume, defined by a co-variant field hi . Fig. 3. From left to right: a jacket and a T-shirt, with the associated field fi and co-variant field hi . The iso-values of each field are depicted on a 2D plane set to intersect the model, where positive isovalues are in blue, negative ones in red, and the 0-iso-value in green. Note that the negative part of the covariant field (in red) enables us to characterize the part of space where a garment might interact with other layers. This enables to discard, for instance, the rounded part at the bottom of fi ’s 0-iso-surface for the T-shirt. More precisely, the input mesh is first sampled using Poisson dart-throwing. Samples are used to compute a HRBF approxima- tion fi ([Iske 2002] , [Macedo et al. 2011]), such that for each pair point/normal (ps,ns) in the sample, fi(ps) = 0 and ∇fi(ps) = ns . The implicit surface defined by {p ∈ R3 | fi(p) = 0} is then close to the initial mesh anywhere near one of its vertices. As in for- mer works combining meshes and implicit modeling [Vaillant et al. 2013], each mesh vertex keeps track of its exact iso-value in the reconstructed field fi , so that no detail is lost (eg. thicker parts such as seam-lines or pockets would be adequately reconstructed after deformation). Note that in practice, we pre-store the values of fi(p) and ∇fi(p) in a grid. Evaluating fi and its gradient at any spatial position is then efficiently approximated using a tri-linear interpolation. Due to the continuity of the HRBF model, the 0-iso-surface of fi is a closed surface. In order to be able to discard parts where interactions with the other garments should not be considered, we also compute a co-variant field hi , aimed at capturing the region "inside borders". While the use of two scalar fields was already introduced in order to reconstruct open surfaces from incomplete point-set acquisition using evolving level-sets [Solem and Heyden 2004], we propose a new efficient formulation for the field hi , based on HRBF, specifically adapted to take advantage of meshes with boundaries . More precisely, the HRBF is computed from the open- boundary samples (ps′,ns′), where the gradient ∇hi(ps′) = ns′ is ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. Implicit Untangling: A Robust Solution for Modeling Layered Clothing • 120:5 set to the unit vector both orthogonal to the open-boundary curve and in the tangent plane with respect to the surface (see Fig.3). The region where hi(p) > 0 is called the zero-influence region of the layer, meaning that it should not be considered for processing interactions with this layer. Conversely, the influence region of a layer is defined by hi(p) < 0. In consequence, the inside of the garment, which should be used to test for collision with outer cloth layers is defined as {p ∈ R3|fi(p) < 0 & hi(p) < 0} while the outside of the garment, where collisions with inner cloth layers are detected, is given by {p ∈ R3|fi(p) > 0 & hi(p) < 0}. 5 UNTANGLING NESTED IMPLICIT SURFACES In this section, we describe our implicit untangling operator in the general case of N implicit surfaces defined as the 0-iso-surfaces of fields (fi)i ∈[1,N ]. Without loss of generality, we consider in the following that i refers to their nesting order. The application to garments defined by both a field and a co-variant field is presented in Section 6. The objective is to define a set of closed-form composition operators enabling the deformation of each of the surfaces so that contact is modeled in regions where they previously interpenetrated. Untangling is performed by replacing each field fi by f̂i such that ∀p ∈ R3, f̂i(p) = Oi(f1(p), .., fN (p)), where Oi is a composition op- erator i.e. a function from RN to R, and where each of the resulting isosurfaces {p ∈ R3 | f̂i(p) = 0} does not intersect any of the other ones. To simplify notation, we consider the fi , ie. functions applied to a position in space and returning fi(p) ∈ R, as a set of indepen- dent real variables. This enables us to define the N-dimensional vector f = (f1, . . . , fN ) ∈ RN , lying in the so-called fields-space, and to write the operator as Oi(f ). Inspired from Angles et al. [2017], the operators are built in two steps: First, we build a desired zero-set Zi in fields-space, i.e. in RN in our case. Secondly, we build Oi from Zi such that Oi(f ) is the signed Euclidean distance between f and Zi , thus ensuring that Zi is the zero-set of Oi . Finally, applying this operator to a position p in 3D space, consists in evaluating Oi(f (p)), where f is now considered as a function of the input position p. In the following, we present the construction of the zero sets Zi , first by defining their general closed-form expression in the N-dimensional case (sec. 5.1), and then by explaining their intuitive construction in the special cases N = 2 (sec. 5.2) and N = 3 (sec. 5.3). The case N = 4 is detailed in the supplementary material associated with this paper. 5.1 General formulation in the N-dimensional case Let us consider a given layer i. We define its zero-set Zi as the union of d = i(N − i +1) sub-spaces (Hb,ci )b ∈[0,i−1],c ∈[0,N −i]. ∀i ∈ [1,N] , Zi = ⋃ b,c H b,c i . (1) Each subspace Hb,ci is itself defined by a hyperplane with normal vector nb,ci ∈ R N , and N − 1 additional inequalities expressed as linear combinations of the (fi), formally described using a matrix M b,c i . ∀i ∈ [1,N], ∀b ∈ [0,i − 1], ∀c ∈ [0,N − i], H b,c i = { f ∈ RN �� nb,ci · f = 0 and Mb,ci f > 0} . (2) As further detailed in the next sections, Hb,ci corresponds to a subset of RN that expresses the interactions between the layer i and all other layers within the interval [i − b,i +c] \ i. Intuitively, the equality constraint (related to the normal n) shifts the layers to be coincident rather than interpenetrating, and the inequality constraints (related to the matrix M) select whether this shift should be applied based on what interpenetrations are present. The vector nb,ci is defined as ∀j ∈ [1,N], nb,ci [j] = { wj if j ∈ [i − b,i +c] 0 otherwise, (3) where the wj are the weights associated with each layer, and acting as stiffness parameters during the untangling process. Note that nb,ci can be multiplied by a scalar value without changing the zero-set. The matrix Mb,ci of size (N − 1) × N defining the inequality constraints on Hb,ci can be expressed using four squared triangular blocks, respectively, A of size (i − b − 1)2, B of size b2, C of size c2, D of size (N −i −c)2. These matrices have the following expression, when layer i is in collision with its b inner, and c above layers. M b,c i = ©­­­« A 0 0 0 0 0 B 0 0 0 0 0 0 C 0 0 0 0 0 D ª®®®¬ , with (4) A = ©­« w1 w2 . . . wi−b−1 0 w2 . . . wi−b−1 . . . ... 0 . . . 0 wi−b−1 ª®¬ B = ©­« −wi−b 0 . . . 0 ... . . . −wi−b . . . −wi−2 0 −wi−b . . . −wi−2 −wi−1 ª®¬ C = ©­« wi+1 wi+2 . . . wi+c 0 wi+2 . . . wi+c . . . ... 0 . . . 0 wi+c ª®¬ D = ©­« −wi+c+1 0 . . . 0 ... . . . −wi+c+1 . . . −wN −1 0 −wi+c+1 . . . −wN −1 −wN ª®¬ Matrices B and C represent the conditions for layer i and its direct neighbors to be in collision, while matrices A and D represent the absence of collision between layer i and the other layers. Note that each line of these matrices is defined up to a positive scaling factor, since Mb,ci is used to express the N-1 conditions M b,c i f > 0. 5.2 Case of two layers (N = 2) To give some intuition, let us explain in more details the construction of the zero-sets in the simple case of two layers N = 2, and show the correspondences with the variables previously defined. Let f1 and f2 be the two fields representing the two implicit surfaces to be nested, where f1 corresponds to the inner one and f2 to the outer one. Let us detail the creation of the zero-set Z1, which drives the creation of the operator O1 (see Fig. 4 for a visual illustration in field-space). Let us consider a point p ∈ R3 on the inner implicit surface such that f1(p) = 0. We can then consider two cases: ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. 120:6 • Buffet, Rohmer, Barthe, Boissieux, and Cani (1) f2(p) < 0: p is inside the outer layer. This is a legal position in field-space, ie. an expected configuration compatible with the absence of collision. In such case, the iso-surface from layer 1 defined by f1 = 0 should remain unchanged around p. (2) f2(p) > 0: p is outside the outer layer which is an illegal position, meaning that there is some intersection between the two layers. Intuitively, in this case we would want the inner iso-surface to be pushed inward by the outer one to solve such collision. We choose to correct the iso-surface by making it lie on w1f1 +w2f2 = 0. Indeed, this expression can be written as f1 = − w2 w1 f2 < 0: following the conventions expressed in Section 3, the resulting surface is in fact placed inward with respect to f1 = 0. As depicted in Fig.5, modifying the relative weights (w1,w2) moves the corrected iso-surface continuously from f1 = 0 to f2 = 0. The zero-set Z1 can therefore be expressed as the following subset of R2: Z1 = { f = (f1, f2) ∈ R 2 | (f2 < 0 & f1 = 0) || (f2 > 0 & w1f1 +w2f2 = 0) } . (5) This relation can be rewritten under the general form described previously as:  Z1 = H 0,0 1 ⋃ H 0,1 1 H 0,0 1 = {f ∈ R 2 | n0,01 · f = 0 and M 0,0 1 f > 0} H 0,1 1 = {f ∈ R 2 | n0,11 · f = 0 and M 0,1 1 f > 0} , (6) with n0,01 = (1,0),n 0,1 1 = (w1,w2), and M 0,0 1 = (0,−1),M 0,1 1 = (0,1). As explained previously, vector n and the lines of matrix M can be multiplied by positive scalars, and can equivalently be defined to match with the general notation of the equations (3) and (4), as n 0,0 1 = (w1,0),n 0,1 1 = (w1,w2), and M 0,0 1 = (0,−w2),M 0,1 1 = (0,w2), respectively. In these cases, H0,01 and H 0,1 1 represent half-lines in the 2D fields- space. M0,01 = (0,D), with D = (−w2) (no collision with the single outer layer), while the other matrices A, B, C are empty. Similarly, M 0,1 1 = (0,C) with C = (w2) (expressing collision with the single outer layer). Z2 is defined in a symmetric way. For a point p lying on the outer implicit surface, for which f2(p) = 0, two similar cases can arise: (1) f1(p) > 0, denoting a legal position for which the second implicit surface should remain unchanged; (2) f1(p) < 0, denoting an illegal position for which we set the new iso-surface to lie on w1f1 + w2f2 = 0, which can be rewritten as f2 = − w1 w2 f1 > 0. The corrected iso-surface then lies outward the input one, which models the action of the inner implicit surface pushing the outer one away. This leads to the following definition for the zero-set: Z2 = H 0,0 2 ⋃ H 1,0 2 H 0,0 2 = {f ∈ R 2 | n0,02 · f = 0 and M 0,0 2 f > 0} H 1,0 2 = {f ∈ R 2 | n1,02 · f = 0 and M 1,0 2 f > 0} , (7) with n0,02 = (0,1), n 1,0 2 = (w1,w2), M 0,0 2 = (1,0) and M 1,0 2 = (−1,0). Using the convention from Section 5.1, this can be re-expressed as n 0,0 2 = (0,w2), n 1,0 2 = (w1,w2), M 0,0 2 = (w1,0) and M 1,0 2 = (−w1,0). Fig. 4. N=2 with fixed w1 and w2: Left and middle: Visualisation of the subparts forming Zi . Right: Visualisation of Zi and some iso-lines of Oi , computed as the distance to Zi . Notice how H 0,1 1 and H 1,0 2 are similar, leading eventually to the contact between the two corrected iso-surface. As we can check, both operators push points formerly on one of the implicit surfaces, but in the intersection region, to the same iso-surface w1f1 +w2f2 = 0, which untangles the two volumes and creates a contact surface instead. Note that the weightsw1 andw2 do not need to sum to one: only their relative value is important, since the contact surface they define can be re-written as f1 + w1 w2 f2 = 0. Playing with their relative value enables us to tune the position of the contact surface anywhere between f1 = 0 (layer 1 not changed, ie. behaving rigidly during untangling) and f2 = 0 (layer 2 not changed, ie. behaving rigidly, as illustrated in Figure 5). Fig. 5. Case of two layers. From left to right: two intersecting implicit surfaces in cross section, where the red one should be kept at the right; view of Zi and iso-lines of the operators Oi in field-space ; untangled results. Modifying the orientation of the half-lines by tuning the ratio between w1 and w2 brings the contact surfaces after untangling from the former position of the blue layer to the one of the red layer (right, top to bottom) 5.3 Case of three layers (N = 3) The case of three layers is slightly more complex. Indeed, a collision between layer 1 and layer 3 should not be handled directly without ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. Implicit Untangling: A Robust Solution for Modeling Layered Clothing • 120:7 taking into account the influence of layer 2 in-between. To solve such inter-dependence, we consider first the collision of layer 2 and 3 independently from layer 1, and set, in case of collision, an intermediate corrected iso-surface defined as w2f2+w3f3 = 0. Then the collision is handled between the intermediate iso-surface and layer 1. To define Z1, we are left with three cases when classifying the possible interpenetration states at a point p on the first layer such that f1(p) = 0: (1) No intersection occurs if f2(p) < 0 (layer 2 is above, see Fig. 6.a ), and w2f2(p)+w3f3(p) < 0 (layer 3, after intermedi- ate correction, is above, see Fig. 6.b ). In this case, we keep layer 1 non-deformed by maintaining the zero-set on f1 = 0. (2) Only layer 2 is intersecting layer 1 if f2(p) > 0 and f3(p) < 0 (Fig. 6.c). In this case, similarly to N = 2, we chose to model coherently the corrected zero-set using w1f1 +w2f2 = 0. (3) Layers 2 and 3 are both interacting with layer 1 when f3(p) > 0 (p is outside layer 3) and w2f2(p) + w3f3(p) > 0. In this case, even if layer 2 is not intersecting layer 1 in its initial configuration, it is pushed by layer 3, and the resulting in- termediate iso-surface ends up in collision with layer 1. (see Fig. 6 (e)). In this case, we set the corrected 0-set of layer 1 to w1f1 +w2f2 +w3f3 = 0 Fig. 6. Partial view of three implicit layers in cross section. The inner side is at the right of f1, and the outer side is on the left of f3. The dashed line represents, in case where it is needed, the intermediate surface profile between layer 2 and 3 represented by w2f2 + w3f3 = 0. (a) and (b): No interaction with layer 1; (c): layer 1 intersects only layer 2; (d) and (e): All three layers are interacting (note: the corrected state between 1 and the other layers is not represented). These three cases turn into the 3 sub-spaces forming Z1: Z1 = H 0,0 1 ⋃ H 0,1 1 ⋃ H 0,2 1 H 0,0 1 = {f ∈ R 3 | n0,01 · f = 0 and M 0,0 1 f > 0} H 0,1 1 = {f ∈ R 3 | n0,11 · f = 0 and M 0,1 1 f > 0} H 0,2 1 = {f ∈ R 3 | n0,21 · f = 0 and M 0,2 1 f > 0} , (8) with n0,01 = (1,0,0), n 0,1 1 = (w1,w2,0), n 0,2 1 = (w1,w2,w3), and M 0,0 1 = ( 0 −1 0 0 −w2 −w3 ) , M 0,1 1 = ( 0 1 0 0 0 −1 ) , M 0,2 1 = ( 0 w2 w3 0 0 1 ) . Using the convention from Section 5.1 where 1 is replaced by the corresponding weight, this can be rewritten as: n 0,0 1 = (w1,0,0), n 0,1 1 = (w1,w2,0), n 0,2 1 = (w1,w2,w3), and M 0,0 1 = ( 0 −w2 0 0 −w2 −w3 ) , M 0,1 1 = ( 0 w2 0 0 0 −w3 ) , M 0,2 1 = ( 0 w2 w3 0 0 w3 ) In this case, we have M0,01 = ( 0 D ) with D = ( −w2 0 −w2 −w3 ) , M 0,1 1 = ( 0 C 0 0 0 D ) with C = (w2) and D = (−w3), and M 0,2 1 =( 0 C ) with C = ( w2 w3 0 w3 ) . While the constraints associated to layer 3 can be derived symmet- rically to the ones we just described for layer 1, the set of constraints to set up for the second layer are slightly different. Indeed, layer 1 and 3 can be directly compared to layer 2 without requiring the consideration of any intermediate corrected surface. We therefore identify four possibles cases: (1) No collision: f1 > 0 and f3 < 0, leading to the isosurface f2 = 0. (2) Collision with layer 1 only: f1 < 0 and f3 < 0 leading to the isosurface w1f1 +w2f2 = 0. (3) Collision with layer 3 only: f1 > 0 and f3 > 0 leading to the isosurface w2f2 +w3f3 = 0. (4) Collision with layer 2 and 3: f1 < 0 and f3 > 0 leading to the isosurface w1f1 +w2f2 +w3f3 = 0. Using the previous formulations, vector nb,ci and matrices M b,c i are defined as follows: n 0,0 2 = (0,w2,0), n 1,0 2 = (w1,w2,0), n 0,1 2 = (0,w2,w3), n 1,1 2 = (w1,w2,w3), M 0,0 2 = ( w1 0 0 0 0 −w3 ) , M 1,0 2 = ( −w1 0 0 0 0 −w3 ) , M 0,1 2 = ( w1 0 0 0 0 w3 ) , M 1,1 2 = ( −w1 0 0 0 0 w3 ) . and  Z2 = H 0,0 2 ⋃ H 1,0 2 ⋃ H 1,0 2 ⋃ H 1,1 2 H 0,0 2 = {f ∈ R 3 | n0,02 · f = 0 and M 0,0 2 f > 0} H 1,0 2 = {f ∈ R 3 | n1,02 · f = 0 and M 1,0 2 f > 0} H 0,1 2 = {f ∈ R 3 | n0,12 · f = 0 and M 0,1 2 f > 0} H 1,1 2 = {f ∈ R 3 | n1,12 · f = 0 and M 1,1 2 f > 0}. (9) Again, note that the equalities n0,11 = n 1,0 2 and n 0,2 1 = n 1,1 2 allow to model the same contact iso-surface between, respectively, layer 1 interacting with layer 2, and all layers 1, 2, 3 interacting. As in the case of two layers, the relative values given to the weights can be used to tune the amount of deformation applied to each layer in interpenetration situations. For instance, in the deep intersection case when all three surfaces interact and end up on w1f1 +w2f2 + w3f3 = 0, this contact surface can be pushed towards f1 = 0 (w1 >> w2 &w3), f2 = 0 (w2 >> w1 &w3), or f3 = 0 (w3 >> w1 &w2). The case N > 3 can be further developed by considering more intermediate layers for higher values of N , to reach the formulation given in Section 5.1. To provide a better insight on the construc- tion of the matrices in the general case, N = 4 is detailed in the supplementary material. ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. 120:8 • Buffet, Rohmer, Barthe, Boissieux, and Cani 5.4 General evaluation of the operator The operator Oi can be evaluated at any position f ∈ RN as the signed Euclidean distance between f and Zi . To evaluate this dis- tance, we compute, as an intermediate step, the closest point f b0,c0i such that (b0,c0) = minb,c ||f − f b,c i ||, where f b,ci is the closest point to f on (H b,c i ). The algorithm to compute efficiently f b,ci for any (b,c) is pro- vided as supplementary material. It consists of computing an ap- propriate orthogonal basis defined for each Hb,ci from the values of nb,ci and all the constraint vectors given by the rows of M b,c i . (b0,c0) is eventually computed in O(N 3) operations. The complexity is related to the use of a Gram-Schmidt orthogonalization iterated over every Hb,ci . The final signed distance is then computed as: Oi(f ) = sign(n b0,c0 i · f ) ||f − f b0,c0 i || . (10) Once f b0,c0i is known, we can note that computing the gradient of the operator ∇Oi = ( ∂Oi/∂fj ) j ∈[1,N ] (which will be used in Section 6.3) is straightforward. Indeed, Oi is the Euclidean distance function, therefore its gradient at position f is given by: ∇Oi(f ) = sign(n b0,c0 i · f ) (f − f b0,c0 i )/∥f − f b0,c0 i ∥. (11) 6 APPLICATION TO UNTANGLING GARMENT MESHES In this Section, we present the extension of field composition to open surfaces using the co-variant fields introduced in Section 4, the modification of the operators to take cloth thickness into ac- count, and the computation of the resulting deformation of garment meshes. 6.1 Using co-variant fields to detect active layers In practice, garments on top of a mannequin are only locally nested. For instance a jacket and trousers may only be overlapping around the hips of the character. As a result, detected inter-penetrations between the associated implicit surfaces should be discarded, except within the overlapping region. More precisely, we define the active layers at a point p in space as the set of garment layers j whose influence region includes p, ie. such that hj(p) < 0 (see Section 4). This enables us to compute the corrected 0-iso-surface of a given layer i — giving the deformation to be applied to a mesh point p of layer i, while only considering the other locally-active layers j. In field space F , canceling-out the effect of a layer j consists in not taking its corresponding field fj into account within the compu- tation of the operator, while keeping the nesting order unchanged. Thus, the operator must be applied at each point p on the subset of fields (fj)j ∈J(p), with J = {j ∈ [1,N] | hj(p) < 0}. The corrected field for garment i at p is thus given by: f̂i(p) = Oi ( (fj(p))j ∈J(p) ) . (12) Note that the closed-form solution for our operators enable us to seamlessly switch to this lower-dimensional field space. In the re- mainder of this section, for sake of simplicity, we re-number the M active layers at p within [1...M]. 6.2 Taking cloth thickness into account Applying the operators on values f = (f1, . . . , fM) leads to 0-iso- surfaces that are in exact contact when inter-penetrations are cor- rected. While this property was desirable to design our untangling operator, we actually aim at modeling cloth layers that may have a non-negligible physical thickness, and that will anyway need to be located on different surfaces for launching an animation. In the fol- lowing, we include the required void space between garment layers aimed at avoiding ill-conditioned simulation within the notion of "thickness" of a cloth layer. A naive approach to handle thickness would be to leave some geometric gaps between garment meshes and their target implicit surfaces, during the process of meshes projection to the associated implicit surfaces. Unfortunately, this could result in new collisions with neighboring layers due to this extra thickness. In contrast, our approach relies on directly integrating thickness values (ti)i ∈[1,N ] within the expression of the operators, allowing to seamlessly and robustly handle collisions between thick layers. Let us consider that the cloth surface, and thus the iso-surfaces defined by Oi(f ) = 0, is centered within the associated, thick cloth layer. As a result, layer i should remain at a minimal distance of ti/2+ti+1/2 from layer i+1, at a minimal distance ti/2+ti+1+ti+2/2 from layer i +2, etc. We model this effect by applying offsets in field space which con- vert into adequate displacements of the iso-surfaces in the 3D space. More precisely, applying an offset to a field fj leads to a geometric displacement of the layer j along the gradient of the field ∇fj . In our case, fj is computed as a HRBF having, by construction, a unit gradient norm on the sampled surface points, and thus, at first ap- proximation, ≃ 1 in the neighborhood of the 0-isosurface. Therefore applying a small offset δ in field space leads to displacement of the layer j along its normal by a length ≃ δ. Thanks to this property, we take into account the offset on the M active layers with respect to the current layer i by applying a change of variables before applying the operator. More precisely, we replace Oi(f ) in Equation (12) by Oi( f̃ ), with f̃ = ( f̃1, . . . , f̃M) defined as: f̃j =  fj i f i = j fj + ti 2 + tj 2 + ∑j−1 k=i+1 tk i f j > i fj − ti 2 − tj 2 − ∑j+1 k=i−1 tk i f j < i. (13) 6.3 Projecting mesh vertices to their iso-surface The modified operators we just presented are the ones used for deforming the meshes towards their corrected, untangled configu- ration. This is done by interleaving two relaxation processes at each vertex of the mesh: gradient relaxation consists in moving points along the gradient ∇ f̂i , toward their original iso-value, while tan- gential relaxation tends to make them slide along the iso-surfaces of f̂i so that the distortion of mesh triangles is minimized. ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. Implicit Untangling: A Robust Solution for Modeling Layered Clothing • 120:9 Fig. 7. Various untangled clothing on top of the same mannequin body (the initial colliding configuration is shown on the left of each example). Note that the resulting silhouette is strongly influenced by the worn layers. Right: application of our method on a different pose. Gradient relaxation is computed using Newton iterations using the gradient value given by ∇ f̂i = ∑ j ∂Oi ∂fj ∇fj , (14) where the ∇fj are obtained using trilinear interpolation of field values pre-stored in a grid and ∂Oi ∂fj is computed from the closest point f b0,c0i as mentioned in Section 5. In practice, we use a step length = 0.3. Tangential relaxation is inspired from As Rigid As Possible de- formations (ARAP) [Sorkine and Alexa 2007]. The key adaptation to our case is to compute per-edge rotations and length changes: the rotation computed for each edge is the minimal rotation around its center that makes it tangential to the field, and the additional length change is a symmetric displacement of the vertices enabling the edge to restore its original length. These are used as target displacements for the two vertices defining the edge. At each tan- gential relaxation step, each mesh vertex applies an average of the displacements assigned the the adjacent edges. Interleaving this second relaxation process with the more stan- dard gradient descent enables-us to avoid over-elongated or inverted triangles, ensuring that the corrected iso-surface will be well ap- proximated in deformed regions. 7 RESULTS AND DISCUSSION 7.1 Implementation All times measured in this paper were taken on a standard laptop computer with an Intel quad Core i7 CPU, clocked at 3.1GHz with 32GB of RAM. Our software uses up to 1Gb of RAM memory at run- time for the presented examples (including the storage of all fields and their gradients stored in uniform grid of size 256×256×64) As our approach treat each vertex independently during mesh deformation described in Section 6, we use OpenMP to trivially parallelize our code. The rendered images were computed off-line using Blender and 3DSMax renderer for the animations. A real-time screen-capture of our software is shown in the accompanying video. 7.2 Qualitative results Several results of our method are depicted in Fig. 1 and Fig. 7, show- ing it can be used to model a variety of layered clothing while ensuring collision-free states. We note that although aimed at pro- viding a collision-free configuration for the garments, our method is also able to generate a quite plausible initial configuration, en- abling the user to test the look of the virtual character even before launching animations. Fig. 8-left illustrates a change of layer order between a rigid jacket and a flexible t-shirt. Note how the strong rigidity of the jacket influences the visible silhouette of the t-shirt when the latter is above. Another example of exchange of layer order is shown in Fig. 1-middle between a t-shirt and a trouser. Fig. 8-right shows the action of the mannequin body which is modeled as a layer of infinite rigidity and has visible action around the hips. Fig. 8. Influence of the interior layers on the visible silhouette. Left: Exchange between a rigid leather jacket and a flexible t-shirt. Right: Result with and without the mannequin body. We tested our approach on the extreme case of a character wear- ing 9 layers (including the body) and show the result in Fig. 9. This validates the robustness of the method in high dimensional fields space, with strongly tangled initial configuration, and deep inter-penetration. Vertical and top-to-bottom cuts are provided to illustrate the well behaved collision-free geometry of all internal layers, in comparison to the initial state. As explained previously, user defined weights wi can model the relative influence between layers when collision is corrected. Chang- ing these weights allows us to tune the relative amount of deforma- tion of the layers from fully rigid to fully flexible. Fig. 10 shows a ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. 120:10 • Buffet, Rohmer, Barthe, Boissieux, and Cani horizontal cut through the dresses of the model shown in Fig. 1-left when weights are modified. Finally, as shown in Fig. 1-right, our untangled model can be di- rectly plugged in as the initial condition of common cloth simulators to be animated without requiring any manual modification on the surface geometry. Note that the animated version of this model is provided in the accompanying video. Fig. 9. Untangling an extreme configuration made of nine initially colliding layers. Cuts along layers are shown on both the initial input (left) and on the untangled result (right). A horizontal cut is shown across the dress in bottom, while the left-most and right-most cuts are performed on a 45◦ corner and zoomed-in to check that the resulting surfaces are fully exempt of collisions. Fig. 10. Horizontal cut through the red dress layers shown in Fig. 1-left. Top- left: initial colliding configuration. Top-right: untangled configuration when all layers have the same weights. Bottom: layers weights are set, respectively from left to right, to (w1, w2, w3) = (2.5, 1, 1), (w1, w2, w3) = (1, 2.5, 1), (w1, w2, w3) = (1, 1, 2.5). Note how the most-rigid layer mostly-keeps its original shape and deforms the surrounding ones. 7.3 Quantitative results The overall number of operations to untangle a model is O(nKN 3), where N is the number of cloth layers, n is the number of vertices, and K is the number of steps required in the iterative deformation. Indeed, the cubic complexity is brought by the field evaluation, while this evaluation has to be performed for every vertex until converging toward the 0-iso-surface. For all our examples we have K ≤ 15, while N is at most 9. As a result for a constant number of layers and iterations, our untangling algorithm is linear with respect to the number of vertices. Time variation with respect to N in shown in Table 1. Table 2 shows our computation time for different cases and validates the roughly linear dependency of the computation time with respect to n. We can also note that, in practice, timings strongly depend on the number of vertices in collisions that need to be corrected. Therefore deep penetration of multiple layers are more computationally costly that correcting slight superficial ones. Table 3 provides a measure of the error in the user-defined layer thickness due to the approximation detailed in Section 6.2. Lastly, the precomputation of the fields for each garment takes less than 4 seconds for all our examples. Table 1. Runtime in seconds with respect to the number of layers and of vertices, for two examples of increasing complexity. The 3 left (resp. right) columns correspond to the example depicted Fig. 1-middle (resp. Fig. 9) on which we added layers one by one. On the right, notice how deep penetrations involving the 3 first layers cause the runtime to drop for the entire example, while on the left the penetrations are superficial and are resolved in a few iterations. N #Vertices n Time(s) N #Vertices n Time(s) 3 6460 0.12 3 7751 1.04 4 9658 0.20 4 9200 1.29 5 15543 0.36 5 13673 2.35 6 19672 0.87 6 16871 3.45 7 18740 1.45 7 23385 4.54 8 22453 6.80 Table 2. Each pair of lines correspond to the same example, on which we subdivided the meshes two times. Line one and two : example Fig. 1-middle, with one less layer. Line three and four : example Fig. 1-right. Line 5 and 6 : example Fig. 8-second picture. We can note that our method exhibits a linear complexity with respect to the number of vertices. N #Vertices1 #Vertices2 #Vertices3 3 5062 20062 79816 Time 0.22 s 0.76 s 2.71 s 4 7502 29635 117785 Time 0.28 s 1.06 s 3.56 s 5 10694 42315 168275 Time 0.57 s 2.1 s 7.15 s 7.4 Limitations Although our method tends to generate plausible configurations in most cases — thanks to the proper handling of relative weights, and thicknesses — it stops considering the influence of a layer as soon as a point is out of the associated influence region. While this is not a problem with respect of getting correct collision-free configurations, ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. Implicit Untangling: A Robust Solution for Modeling Layered Clothing • 120:11 Table 3. Measures of error between user defined thickness value and com- puted one. Note that the error increases with the thickness value and local curvature of the surface. Mesh δtarдet δobtained Relative error Jacket 0.20 0.1983 0.85% PullV 0.09 0.0901 0.11% TShirt 0.04 0.0406 1.5% Petticoat 0.04 0.0421 5.3% this results in an overly distorted shape for the underlying layers, such as we can see in Fig. 11 for the parts of the skirts immediately below the jacket. If our method is an initialization before animation this is not a problem, but if our results are to be directly used as an illustration, some local relaxation is required, as shown in Figure 11 (right). Fig. 11. Left : limitations. The salient distortion that occurs near the border of the influence zone of a top layer (left) can be attenuated using relaxation (center), if our results are to be directly used as plausible shapes. Right : Even when launched in collision-free states, standard cloth simulators often fail to generate collision-free motion for layered garments. In future works, our approach could also be used to correct such collision happening in a dynamic context. To achieve a better level of plausibility, our method would also need to consider the fact that garments deform isometrically with the associated 2D pattern: even is the cloth is slightly extensible, they tend to fold rather than compress. Consequently, folds should be added to the new shapes of inner layers, when they get compressed by a stiffer layer on top. This could be done by using inspiration from Rohmer et al. [2010], which makes use of an implicit model for folds. The latter could be integrated as an additional displacement within our operator. In addition, our method suffers from a few failure cases. Indeed, it only applies when a well-defined nesting between different cloth lay- ers can be defined. This is not the case for a single, self-intersecting garment (see Fig. 12-left). In such case, the self-intersecting cloth surface cannot be reconstructed as some zero-sets of implicit fields to be untangled, which prevent the use of our method. Enabling im- plicit untangling to be applied on local surface patches computed on the fly would be a nice extension of our method, since it could allow us to handle the challenging case of self-collision states [Ainsley et al. 2012]. Fig. 12. Failure cases: on the left, a shawl folded back onto itself and a large skirt in deep intersection with itself cause the implicit reconstruction to fail: no clean 0-isosurface can be defined, which makes our approach inadequate for processing such self-collision case. On the right, the co-variant field hi computed for a shorty does not capture the influence zone of the shorty between the legs, leading the shorty to be ignored as an outer layer when the vertices of the dress are processed in this zone. Another possible failure case is illustrated in Fig. 12-right, where a shorty is worn on top of a long dress. In this case, the nesting order is not captured globally by the field hi , which leads to a final result where collision still occurs. Extending the definition of hi , possibly taking into account user indications, could also be handled as a future work. Lastly, being able to apply our untangling method at each step of an animation would be highly desirable, since cloth simulation en- gines may not be able to maintain collision-free states during highly dynamic motion, even if we provide one to start with. While apply- ing our current method is feasible, it would require reconstructing the field and co-variant field of each garment at each frame, leading to a few tenths of seconds of computational time. Taking temporal coherence into account to allow us a more efficient reconstruction over time would be an interesting direction for future work. 8 CONCLUSION We described an implicit solution for untangling layered garments. Our method allows to robustly convert penetration states into collision-free contact regions, while taking into account both a thickness parameter for cloth layers and their relative rigidity. Key to our approach are the use of a pair of HRBFs, which are closed implicit surfaces, to approximate open garment surfaces, the in- troduction of closed-form N-ary untangling operators for layered implicit surfaces, and a new method for accurately projecting cloth meshes within the associated implicit field. The configuration we generate for the dressed character can serve as valid initial state for launching simulations as layers are guar- antee to not intersect, and are even separated by a user defined distance. Moreover, since relative cloth rigidity can be considered, this configuration remains close to the rest state in contact regions. This enables our method to be used for quickly generating plau- sible static shapes of garments that take any number of layers of underwear into account. ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. 120:12 • Buffet, Rohmer, Barthe, Boissieux, and Cani Our work opens up several new directions of research. First, the combination of the computed untangling deformations with the automatic generation of implicit folds could lead to more realis- tic shapes which would be useful to start simulation in a close to equilibrium state. Secondly, finding an efficient way to use implicit untangling at each animation step would enable robust multi-layer clothing animations, even for highly dynamic motions. ACKNOWLEDGMENTS This work was partially funded by the FOLD-Dyn project (ANR- 16-CE33-0015). We would also like to thank the reviewers for their valuable advices and help to improve the paper clarity. REFERENCES Samantha Ainsley, Etienne Vouga, Eitan Grinspu, and Rasmus Tamstorf. 2012. Specu- lative Parallel Asynchronous Contact Mechanics. ACM Trans. Graph., Proc. ACM SIGGRAPH Asia 31, 6 (2012). Baptiste Angles, Marco Tarini, Brian Wyvill, Loïc Barthe, and Andrea Tagliasacchi. 2017. Sketch-based Implicit Blending. ACM Trans. Graph. (2017). David Baraff, Andrew Witkin, and Michael Kass. 2003. Untangling Cloth. ACM Trans. Graph., Proc. ACM SIGGRAPH 22, 3 (2003), 862–870. L. Barthe, V. Gaildrat, and R. Caubet. 1998. Combining implicit surfaces with soft blending in a CSG tree. In Proc. of CSG Conference Series. 17–31. James F. Blinn. 1982. A Generalization of Algebraic Surface Drawing. ACM Trans. Graph. 1, 3 (1982), 235–256. Jules Bloomenthal (Ed.). 1997. Introduction to Implicit Surfaces. Morgan Kaufmann. Robert Bridson, Ronald Fedkiw, and John Anderson. 2002. Robust Treatment of Colli- sions, Contact and Friction for Cloth Animation. ACM Trans. Graph., Proc. ACM SIGGRAPH 21, 3 (2002), 594–603. Marie-Paule Cani. 1993. An implicit formulation for precise contact modeling between flexible solids. In ACM SIGGRAPH. 313–320. Zhili Chen, Renguo Feng, and Huamin Wang. 2013. Modeling Friction and Air Effects Between Cloth and Deformable Bodies. ACM Trans. Graph. 32, 4 (2013). Erwin de Groot, Brian Wyvill, and Huub van de Wetering. 2009. Locally restricted blending of Blobtrees. Computers & Graphics 33, 6 (2009), 690–697. Charbel Fares and Ar Hamam. 2005. Collision detection for rigid bodies: A state of the art review. In GraphiCon. Olivier Gourmel, Loic Barthe, Marie-Paule Cani, Brian Wyvill, Adrien Bernhardt, Math- ias Paulin, and Herbert Grasberger. 2013. A Gradient-based Implicit Blend. ACM Trans. Graph. 32, 2 (2013). Peng Guan, Loretta Reiss, David A. Hirshberg, Alexander Weiss, and Michael J. Black. 2012. DRAPE: DRessing Any PErson. ACM Trans. Graph., Proc. ACM SIGGRAPH 31, 4 (2012). David Harmon, Etienne Vouga, Rasmus Tamstorf, and Eitan Grinspun. 2008. Robust Treatment of Simultaneous Collisions. ACM Trans. Graph., Proc. ACM SIGGRAPH 27, 3, Article 23 (2008). P.-C. Hsu. 2018. K-ary Implicit Blends with Increasing or Decreasing Blend Ranges for Level Blend Surfaces. Journal of Advances in Information Technology (2018). P. C. Hsu and C. Lee. 2003. Field Functions for Blending Range Controls on Soft Objects. Proc. of Eurographics, Computer Graphics Forum 22, 3 (2003), 233–242. Armin Iske. 2002. Scattered Data Modelling Using Radial Basis Functions. Tutorials on Multiresolution in Geometric Modelling. Mathematics and Visualization (2002). François Lehericey, Valérie Gouranton, and Bruno Arnaldi. 2015. GPU Ray-traced Collision Detection for Cloth Simulation. In ACM Symp. on VRST. 47–50. Ives Macedo, Joao Paulo Gois, and Luiz Velho. 2011. Hermite Radial Basis Functions Implicits. Comput. Graph. Forum 30 (2011), 27–42. Matthias Müller, Nuttapong Chentanez, Tae-Yong Kim, and Miles Macklin. 2015. Air Meshes for Robust Collision Handling. ACM Trans. Graph. 34, 4 (2015). A. Pasko, V. Adzhiev, A. Sourin, and V. Savchenko. 1995. Function representation in geometric modeling: concepts, implementation and applications. The Visual Computer 11, 8 (1995), 429–446. A. Ricci. 1973. Constructive Geometry for Computer Graphics. Computer journal 16, 2 (1973). Damien Rohmer, Tiberiu Popa, Marie-Paule Cani, Stefanie Hahmann, and Sheffer Alla. 2010. Animation Wrinkling: Augmenting Coarse Cloth Simulations with Realistic- Looking Wrinkles. ACM Transactions on Graphics, Proc. SIGGRAPH Asia 29, 5 (2010), 157. M-A Sabin. 1968. The Use of Potential Surfaces for Numerical Geometry. In Tech. Report VTO/MS/153, British Aerospace Corp., Weybridge, U.K. Andrew Selle, Jonathan Su, Geoffrey Irving, and Ronald Fedkiw. 2009. Robust High- Resolution Cloth Using Parallelism, History-Based Collisions, and Accurate Friction. IEEE TVCG 15, 2 (2009), 12. J. E. Solem and A. Heyden. 2004. Reconstructing open surfaces from unorganized data points. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., Vol. 2. II–II. Olga Sorkine and Marc Alexa. 2007. As-rigid-as-possible Surface Modeling. In Proc. Symposium on Geometry Processing. Liming Sun, Timo R. Nyberg, Gang Xiong, and Juntao Ye. 2016. Minimum Displace- ments for Cloth-obstacle Penetration Resolving. In Proceedings of the 37th Annual Conference of the European Association for Computer Graphics: Short Papers (EG ’16). 53–56. Min Tang, Dinesh Manocha, Miguel A. Otaduy, and Ruofeng Tong. 2012. Continuous Penalty Forces. ACM Trans. Graph. 31, 4 (2012). Min Tang, Huamin Wang, Le Tang, Ruofeng Tong, and Dinesh Manocha. 2016. CAMA: Contact-Aware Matrix Assembly with Unified Collision Handling for GPU-based Cloth Simulation. Computer Graphics Forum, Proc. Eurographics 35, 2 (2016). Min Tang, Tongtong Wang, Zhongyuan Liu, Ruofeng Tong, and Dinesh Manocha. 2018. I-Cloth: Incremental Collision Handling for GPU-Based Interactive Cloth Simulation. ACM Trans. Graph. Proc. ACM SIGGRAPH Asia 37, 6 (2018). M. Teschner, S. Kimmerle, Gabriel Zachmann, B. Heidelberger, Laks Raghupathi, A. Fuhrmann, Marie-Paule Cani, François Faure, N. Magnenat-Thalmann, and W. Strasser. 2004. Collision Detection for Deformable Objects. In Eurographics 2004, State-of-the-Art Report. Eurographics Association, 119–135. Rodolphe Vaillant, Loïc Barthe, Gaël Guennebaud, Marie-Paule Cani, Damien Rohmer, Brian Wyvill, Olivier Gourmel, and Mathias Paulin. 2013. Implicit Skinning: Real- time Skin Deformation with Contact Modeling. ACM Trans. Graph. 32, 4 (2013). Rodolphe Vaillant, Gäel Guennebaud, Loïc Barthe, Brian Wyvill, and Marie-Paule Cani. 2014. Robust Iso-surface Tracking for Interactive Character Skinning. ACM Trans. Graph. 33, 6 (2014). Pascal Volino and Nadia Magnenat-Thalmann. 2006. Resolving Surface Collisions Through Intersection Contour Minimization. In ACM Trans. Graph., Proc. ACM SIGGRAPH. Holger Wendland. 2005. Scattered Data Approximation, Cambridge University Press. Audrey Wong, David Eberle, and Theodore Kim. 2018. Clean Cloth Inputs: Removing Character Self-intersections with Volume Simulation. In ACM SIGGRAPH Talks. Juntao Ye, Guanghui Ma, Liguo Jiang, Lan Chen, Jituo Li, Gang Xiong, Xiaopeng Zhang, and Min Tang. 2017. A Unified Cloth Untangling Framework Through Discrete Collision Detection. Computer Graphics Forum 36, 7 (2017). J. Ye, T. R. Nyberg, and G. Xiong. 2015. Fast Discrete Intersection Detection for Cloth Penetration Resolution. In IEEE Int. Conf. on Multimedia Big Data. 352–357. Juntao Ye and Jing Zhao. 2012. The Intersection Contour Minimization Method for Untangling Oriented Deformable Surfaces. In Symposium on Computer Animation. C. Zanni, M. Gleicher, and M.-P. Cani. 2015. N-ary Implicit Blends with Topology Control. Comput. Graph. 46 (2015). ACM Trans. Graph., Vol. 38, No. 4, Article 120. Publication date: July 2019. Abstract 1 Introduction 2 Related work 2.1 Contact modeling in cloth animation 2.2 Implicit surfaces 3 Overview 3.1 Notations and input 3.2 Processing pipeline and challenges 4 Implicit approximation for garments 5 Untangling nested implicit surfaces 5.1 General formulation in the N-dimensional case 5.2 Case of two layers (N=2) 5.3 Case of three layers (N=3) 5.4 General evaluation of the operator 6 Application to untangling garment meshes 6.1 Using co-variant fields to detect active layers 6.2 Taking cloth thickness into account 6.3 Projecting mesh vertices to their iso-surface 7 Results and discussion 7.1 Implementation 7.2 Qualitative results 7.3 Quantitative results 7.4 Limitations 8 Conclusion Acknowledgments References 
work_qjfjookvc5chvio762y3j6qd4q	----	An Expert System for Modelling Wave - Height Time - Series An Expert System for Modelling Wave - Height Time - Series Rodolfo Piscopia Freelance Rome, Italy Abstract— This paper describes an expert system designed for the analysis of an incomplete, non-stationary and non- Gaussian, long-term, time series of wave significant heights by means of specific linear parametric model. Using this system makes it possible to complete missing-value gaps, forecast wave- height short-term evolution or simulate arbitrarily long sequences of wave data preserving the key statistical properties of the observed series, including autocorrelation, persistence over threshold, non-Gaussian distribution and seasonality. The implemented improvements bear on specific key tasks of ARMA setup procedure, i.e. preliminary analysis, parameter estimation and optimal model-configuration identification. Specifically, a Seasonal Trend decomposition based on Loess robust method is applied to compute more stable and detailed seasonal trend, allowing assuming more confidently its deterministic nature. Moreover, aiming at accurately estimating the model parameters, a proficient method is taken in, which is based on the robust Whittle’s approximation of the maximum log-likelihood function as well as on the direct-search, non- linear, multi-parameter, constrained, optimization technique called complex modified. Finally, an automatic expert system is developed, able to identify, almost correctly, ARMA orders by selecting the model with the smallest residuals variance and parameter numbers. Confident applicability of the suggested procedure is tested by means of both Monte Carlo simulations and comparisons of generated series with observed one, this latter measured offshore Alghero – Italy. Analysis of results clearly indicate that the accuracy in identifying the correct ARMA model is improved; furthermore, it is shown that the simulated time series exhibit all the primary statistical properties of the observed data, including winter and summer seasonal patterns as well as sea states sequencing, persistence and severeness. Keywords — Wave climate; ARMA model; Wave forecast; Storm duration; Sea state persistence; Sea severeness I. INTRODUCTION For marine human activities and engineering applications, the understanding of sea-state sequences is important as well as the knowledge of extreme wave parameters, e.g. to evaluate a maritime traffic line efficiency, to guess a port/terminal operativeness or to assess risks of engineering processes. Actually, marine intervention and installation works involve long-lasting and complex operations. In these cases, the analysis of effects related to meteorological changes during specific operations is utmost relevant to disclose any possible critical situations and their related costs-growth. To these aims, linear models can be very helpful, being able to provide large database of information statistically equivalent to the observed one. Additionally, recorded time series are usually incomplete due to several reasons, e.g. to instrument failures, accidental data loss or spikes rejection. Considering that the data incompleteness can seriously bias statistical inferences, makes obvious the relevance of a procedure able to recover missing values by ensuring same statistical sample properties. Autoregressive, moving-average models (ARMA) are a specific class of the linear parametric family that, in few words, replicate time processes by combining some their outcomes with a white noise. In ocean engineering applications [1] and [2] have used ARMA to simulate individual waves in short-term elevation record, supposed to be stationary in time. Reference [3] have used ARMA to model the non-stationary, long-term, time- series of significant wave-height, whereas [4] proposed a new methodology for the analysis, missing-value completion and simulation of an incomplete, non-stationary, time-series of wave data. Further researches were pointed at verifying data transferability between two wave-measuring stations [5]. Generally, two main problems have to be solved in order to apply ARMA models to long-term series of wave parameters. One is the presence in the series of missing-value gaps, which can sometimes be relatively long, and the other is the series non-stationarity and non-normality. Accordingly, gap filling as well as data transformation procedures are required. Furthermore, the common and challenging task of the model identification, i.e. selecting the most suitable ARMA order, has to be tackled. Here, the work of [4] is extended by improving the following tasks: the seasonal component assessment, the model parameter estimation and the choice of the optimal ARMA configuration. Namely, a technique called STL robust (Seasonal Trend decomposition based on Loess) able to compute more accurate seasonal components is adopted [6]. Furthermore, the more robust Whittle’s approximation of the maximum log-likelihood function is used to estimate ARMA coefficients, the set of which is found out by a proficient, nonlinear, constrained, multi-parameter, optimization technique. Finally, an expert system has been developed allowing automating the struggle step of model identification that, for mixed ARMA process, is quite tricky and someway affected by subjective interpretation. Different Monte Carlo simulations have been carried out with a double purpose: the validation of the parameter estimation procedure and the verification of the automatic expert system proficiency. The obtained results have been gratifying, making possible to confidently say that the proposed enhancements are very efficient. International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 688 In the following, each step of the adopted ARMA modelling procedure is accurately described and the results of Monte Carlo simulations are illustrated. Afterwards, the application to real wave data is fully described and the comparison between two different techniques to normalize- denormalize the series is plainly outlined. Finally, the conclusions are drawn out. II. LINEAR PARAMETRIC MODELS Starting from the Box and Jenkins definition [7], the family of linear models has been developed with the conception of several subtypes that roughly follow a common setup procedure. Here only the ARMA model is considered. Regarding a second-order stationary and Gaussian time series zt, the autoregressive and moving average parts of an ARMA(p,q) model define zt respectively as the combination of p previous terms of the series plus the combination of q+1 terms of a white noise (i.e. a stationary random process with zero mean and variance equal to 2r ). Introducing the back- shift operator defined as ntt n zzBB : , an ARMA(p, q) can be written as     tt aBzB   , being pp BBB   ....1)( 1 and q q BBB   ....1)( 1 . The standard ARMA setup procedure can be resumed as follows [7]: preliminary analysis, model identification, parameter estimation, model verification and optimal model- configuration selection. In what follows each task is accurately delineated. III. PRELIMINARY ANALYSIS With reference to the stationarity of significant wave- height series, it is typically not satisfied and, according to the common knowledge of the environmental process, a seasonal component is expected to exist. According with [8], a non- stationary time series zt can be decomposed as ttttt Xzz   ~ , where tz ~ , t , t are the deterministic functions, respectively, of long-term trend, seasonal mean and standard deviation. In what follows, the long-term trend is not considered. The seasonal mean and variance can be defined as follows by introducing the Buys-Ballot double index, i.e. by re- indexing the time series zt as [9]:      Y j j t MzNtz 1 1, 1,          Y j jz K 1 1   M1          Y j jz K 1 22 1 1    M1   where Y and M are integers respectively equal to the series length in year and the annual number of observations, N=YM is the series numerosity,  is the index within the annual cycle, K is the number of existing values per each observation index  (if no missing values affects the wave series then clearly K=Y). The deseasonalized series is computed according to the following expression:      jj zy with Yj 1 and M1   This approach, however, produces seasonal components possibly affected by large sample variability, especially when the time series length is not enough extended or when many missing-value gaps exist. This large variability contrasts with the assumed deterministic nature of the seasonal component and cannot therefore be accepted by both physical and stochastic points of view. Following [6], here is preferred a more robust method, derived from the STL one (Seasonal-Trend decomposition based on Loess). This technique, used for both the mean and variance seasonal components, is split into five tasks (here, only the evaluation of the mean component is illustrated as the variance computation is straightforwardly derivable). 1. Identify the seasonal mean series * by (2) and reduce it to zero average. Compute the new time series *  jj zX . 2. Define the scaling factor   cXu jj  , where  is the median value of jX and c is a constant (equal to 6 and 36 respectively for the mean and variance seasonal components). 3. Estimate the weighting factor j , for each jX , according to:            1 0 1 -1 2 2 j jj j uif uifu      4. Evaluate the weighted seasonal mean component, for each index  in the annual cycle, as:     Y j j Y j jjXm 11 *     5. Smooth *m by means of the interpolator called Loess. Specifically, considering a time window centered at , with amplitude W, *m is smoothed according to:       2 2 **ˆ W Wi iimm     with 2 3 2 1)(                   W i i     For completing the missing-value gaps, the following procedure has been adopted. When the gap length is very small (dealing with one or two observations), the missing values are interpolated from neighbors. Otherwise, the smoothed mean seasonal component (7) is transformed in a Fourier series. The missing values are therefore replaced by [3]:      sincos ~ baz j    being  the average value of the smoothed mean seasonal component, =2/M, a and b the Fourier coefficients given by:      M M a 1 cos 2         M M b 1 sin 2      International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 689 With reference to normality of the significant wave-height series, it is verified by performing a t-student test; if the tested hypothesis is rejected, a series transformation is applied. Two different approaches were implemented and compared: the first involves the classical Box-Cox transformation [10]; the second entails the Probability Level-Equivalence Transformation (PLET) used by [11]. Using the Box-Cox formula, the time series is transformed as:          0 log 0 1 )(ˆ BCt BCBCt BCt whenz whenz z BC       Differently, the PLET method is based on the percentile equivalence among the standardized normal distribution () and the wave-height best-fitting distribution (Pz). The standardized normal series is therefore obtained by:    tzt zPz 1ˆ    The inverse transformation used in the simulation task is:    tzt zPz ˆ1     IV. MODEL PARAMETERS ESTIMATION This task is here completed in two steps: the preliminary estimation and the accuracy refining. The method of moments [7] is adopted for the former, whereas the maximum log- likelihood method along with the Whittle’s function approximation is implemented for the latter. Actually, for a Gaussian stationary process, the approximated expression of log-likelihood function is [12]:     zAz N dSzL TW )( 1 ,log 2 1 )|( ~ ψψψ          being =(1,…,p,1,…,q, r) the parameters vector to estimate, () the inverse of the process covariance-matrix and S(, ) the parametric power spectrum. The latter is the Fourier transform of the autocorrelation function and can be seen as expressing the energy level of each periodicity composing the time series. Its expression, computed by parametric method, can be written as [13]:                      p j jj q i ii r jj ii S 1 22 1 22 2 2sin2cos1 2sin2cos1 2;    ψ   Equation (13) can be rewritten in terms of the series periodogram (P), i.e. the series power-spectrum computed by Fourier method, as follows [14]:                           d S P dSzLW , ,log 2 1 )|( ~ ψ ψψ   The maximization of (15) is here carried out by a direct- search, non-linear, multi-parameter, constrained, optimization technique called complex modified [15] [16], which has been proved to perform very efficiently [17]. Aiming to enlighten the proficiency of the maximum log- likelihood method, two set of tests were carried out. In the first one, three different spectra were firstly defined by assigning p, q, r, j and  in (14) and then randomly perturbed by adding a white noise drawn out from U[-0.05r; 0.05r]. The resulting frequency distributions were assumed as periodograms to be fitted by the complex modified method with (15) as target function. Fig. 1 shows the results along with those achieved using the spectral least-square target- function, given by [4]:                     d P PS ; 2 1 2 ψ   The fig. 1 and the herein reported table clearly state that, even if both (15) and (16) fulfil the optimization by producing a nearly perfect fitting of ARMA spectra, (15) is more efficient to accurately estimate each parameter value, reducing the maximum relative approximation-error form 200% to 2% (for AR(4) – AR(10)) or from 20% to 3% (ARMA(10,10)). In the second test case, a Monte Carlo simulation was carried out by modelling 500 synthetic series, generated from an ARMA(1,1) with r = 1.0 and both (,  ranging from 0.0 to 1.0, step 0.2. For each generated series, (15) was used to estimate the ARMA parameter values. The difference between the assigned values and the finally estimated one have been represented, in fig. 2, as relative errors in a box-plot form. The relative errors obtained by using the widespread method of moments [18] along with those achieved by minimizing (16) are also reported. The illustrated results, revealing a great error variance reduction (at least halved) as well as an unbiased zero averages, confidently confirm the proficient improvement achieved by the herein implemented method. V. VERIFICATION OF ESTIMATED MODELS AND SELECTION OF THE OPTIMAL ONE To verify ARMA stationarity and invertibility, respectively, all roots of following (17) should lie externally to the unit circle (here IMSL® ZPORC routine is used):  01 1   p i i i x  01 1   q i i i x   If one of the tested hypothesis is rejected the model is not considered further, otherwise a Portmanteau test is completed. Namely, if an ARMA is stationary and invertible as well as properly identified with accurately estimated parameters, the model residuals, given by                q j jj p i ii azzr 11     have nearly null random values. International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 690 Model = AR(4) Model = AR(10) Model = ARMA(10.10) fixed MLM LSM fixed MLM LSM fixed MLM LSM r = 1.00 r = 1.01 r = 2.98 r = 1.00 r = 0.98 r = 3.04 r = 1.00 r = 1.01 r = 1.03 j j j j j j j i j i j i 0.94 0.95 2.81 -0.82 -0.78 -2.38 -0.41 0.87 -0.42 0.87 -0.47 0.84 -0.62 -0.63 -1.87 0.12 0.12 0.41 -0.15 -0.17 -0.14 -0.18 -0.13 -0.14 0.03 0.02 0.10 0.16 0.21 0.71 0.80 -0.38 0.84 -0.37 0.86 -0.33 -0.20 -0.20 -0.59 0.62 0.56 1.54 0.31 0.03 0.36 0.07 0.38 0.09 - - - 0.18 0.20 0.83 0.80 -0.21 0.84 -0.19 0.89 -0.19 - - - 0.02 0.01 0.28 0.92 0.58 0.92 0.56 1.13 0.68 - - - 0.75 0.72 2.06 -0.67 0.38 -0.68 0.39 -0.62 0.46 - - - 0.99 1.00 3.03 0.72 0.09 0.71 0.07 0.84 0.16 - - - 0.45 0.42 1.19 0.81 0.18 0.86 0.20 0.92 0.22 - - - 0.93 0.89 2.72 -0.41 -0.81 -0.44 -0.83 -0.45 -0.79 0.2 0.3 0.3 0.4 0.4 0.5 0.5 0.6 0.6 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 Randomly perturbed theoretical spectrum Adapted spectrum Theoretical spectrum 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.0 0.1 0.2 0.3 0.4 0.5 Randomly perturbed theoretical spectrum Adapted spectrum Theoretical spectrum 0.05 0.10 0.15 0.20 0.25 0.30 0 0.1 0.2 0.3 0.4 0.5 0.8 1.3 1.8 2.3 2.8 3.3 0 0.1 0.2 0.3 0.4 0.5 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.0 0.1 0.2 0.3 0.4 0.5 0.05 0.10 0.15 0.20 0.25 0.30 0 0.1 0.2 0.3 0.4 0.5 0.8 1.3 1.8 2.3 2.8 3.3 0 0.1 0.2 0.3 0.4 0.5 LSM target function (16) Fig. 1. Comparison between synthetics periodograms and spectral distributions obtained by the complex modified optimization technique. International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 691 Fixed parameter values a b c d e f g h i j k l n o p q r s t u p 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 q 1 1 1 1 0 1 1 1 0 1 1 1 0 1 1 1 0 1 1 1 1 # # # # 0.2 0.2 0.2 0.2 0.2 0.4 0.4 0.4 0.6 0.6 0.6 0.6 0.8 .8 .8 .8 1 0.2 0.4 0.6 0.8 # 0.4 0.6 0.8 # 0.2 0.6 0.8 # 0.2 0.4 0.8 # 0.2 0.4 0.6 3.0 2.5 2.0 1.5 1.0 0.5 0.0 -0.5 -1.0 -1.5 -2.0 -2.5 -3.0 M.L.M. (Whittle) Least square method Method of moments 3.0 2.5 2.0 1.5 1.0 0.5 0.0 -0.5 -1.0 -1.5 -2.0 -2.5 -3.0 a b c d e f g h i j k l n o p q r s t u a b c d e f g h i j k l n o p q r s t u a b c d e f g h i j k l n o p q r s t u 3.0 2.5 2.0 1.5 1.0 0.5 0.0 -0.5 -1.0 -1.5 -2.0 -2.5 -3.0 m odel m odel ˆ  i m odel m odel ˆ  i m odel m odel ˆ r rir   Estimates of AR coefficients Estimates of MA coefficients Estimates of residual variance Fig. 2. Results of the Monte Carlo simulation for the ARMA(1,1) parameter estimation obtained by means of three different techniques: the methods of moments (right panels), the spectral Least Square Method (central panels) and the Whittle’s approximation of the Maximum log-Likelihood Function (left panels). International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 692 To verify this hypothesis, the Ljung-Box Portmanteau-test is used [19]. A weighted sum (Q) of residual autocorrelation coefficients is computed according to the following expression:       s k k kN NNQ 1 2 2    where N is the residuals number, k is the time lag, k is the autocorrelation coefficient (computed by IMSL® ACF routine) and s is the maximum lag (here s=75 is adopted). Q is then compared to the quantile of a 2 distribution, with s degrees of freedom, at the level of probability P. If Q is greater than 2 (s), the test is rejected and at least one of the examined autocorrelation coefficients is statistically different from zero, to the fixed significance level P. If the tested ARMA is stationary, invertible, with random residuals, it will be considered for the final task of optimal model selection, i.e. the choice of model order (p, q). Aiming at automating this specific task, many methods based on some patterns of different ACF functions have been proposed [20]. Here a different approach is used. Starting from both the definition and meaning of a linear model, the more efficient configuration can be defined as the one that outlines the process correlation structure by using the lowest parameter number and, at the same time, produces random residuals with the lowest variance. Considering that the latter generally decreases as the former increases, makes it necessary to choose the “optimal configuration” on the basis of statistical indices. In the present work, the following is taken into consideration [21]:                        qp σσN qp+ N-p-q σN N-p-qBIC= rrr 222 ˆ ln ˆ ln   where 2ˆ r σ is the variance of series-residuals. The combination of p and q that minimize (20) is regarded as the ARMA “optimal configuration”. To show the consistency of the proposed procedure, two series of tests were carried out. The first one was performed analyzing the identified ranking of ten different AR, MA and mixed ARMA models of arbitrary orders. The obtained results are reported in Table I and support, although not on a statistical basis, the developed expert-system robustness. Namely, the true model order is ranked for seven times in the first two positions and it is always high-ranked. Furthermore, only one attempt gave the sum p+q of the fitted model underestimated by more than one order (settled ARMA(2,3) – selected MA(3)). The second Monte Carlo simulation was carried out with the goal of comparing the proposed expert-system proficiency with that of the three best-performing ACF pattern-selection methods; namely, the Corner, the EACF and the SCAN methods. The efficiency of these latter methods was found out from [20]. The ARMA(2,1)      tt aBzBB 5.015.018.01  with 0.12 r , was used for simulating 1000 series of 1000 terms, which were successively analyzed. TABLE I. RESULTS OF THE “OPTIMAL MODEL IDENTIFICATION” TEST FOR TEN DIFFERENT AR, MA AND ARMA MODELS OF ARBITRARY ORDER. FIXED Identified Fixed ranking p q p q BIC 1 1 1 1 1st 1 2 1 2 1st 1 3 1 3 1st 2 1 1 1 3rd 2 3 0 3 4th 3 1 3 0 2nd 3 2 3 2 1st 5 7 4 8 2nd 6 4 6 5 6th 20 2 18 20 2nd The occurrence of the identified combinations (p, q) obtained by the different methods are summarized in Table II. All the methods fairly spread out the identified configurations but, in this specific case, any of them significantly underestimate the model total order (p+q). The expert system performs better than the corner method. Namely, the former achieves nearly equal results in selecting the true model configuration (scoring just a 1% less than the latter), but it shows greater sensitivity in both recognizing the minimum model orders and identifying the correct influence of the AR and MA model parts. Actually, when one of the identified model order is equal or greater than the fixed one, the expert system 35% of times overestimates the other one whereas the corner method underestimates it 49% of times. Moreover, the expert system 21% of times bias the autoregressive character of the process with the MA one whereas the corner method makes the same error for the 41% of times. The ESACF and SCAN methods have instead a worse hit percentage for the correct model identification and have the same biasing character of the Corner one. On these bases, it could be stated that the implemented expert system works nicely well, slightly better than the best performing pattern selection method here considered. Moreover, it has to be highlighted that the expert system automatically provides in output a list of ranked models, opening to chances of trying different configurations having similar statistical index values. TABLE II. NUMBER OF IDENTIFIED COMBINATIONS (p, q) FOR THE 1000 SERIES GENERATED FROM AN ARMA (2,1). Expert system Corner method ESACF method SCAN method (1,0) - - - - - - - - (0,1) - - - - - - - - (2,0) 3 0% - - - - 88 9% (0,2) - - - - - - - - (3,0) 59 6% 86 9% - - 309 31% (0,3) - - - - - - - - (1,1) - - - - - - - - (2,1) 472 47% 483 48% 339 34% 206 21% (3,1) 80 8% 7 1% 30 3% 6 1% (1,2) 12 1% 155 15% 110 11% 195 20% (2,2) 108 11% 7 1% 92 9% 10 1% (3,2) 22 2% 2 0% 59 6% 4 0% (1,3) 45 4% 253 25% 303 30% 179 18% (2,3) 156 16% 7 1% 66 7% 2 0% (3,3) 42 4% - - - - - - International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 693 TABLE III. RATES OF UNDER-SPECIFICATION OF THE TOTAL ARMA ORDER.  0 3 5 7 9  Expert system 0% 32% 40% 59% 62% 100% Corner method 0% 0% 2% 47% 49% 100% ESACF method 0% 0% 0% 7% 3% 87% SCAN method 9% 9% 9% 46% 49% 100% Finally, the influence of outlier occurrences on the methods performance has been analyzed. Accordingly, a new Monte Carlo simulation similar to the previous one was carried out; the same number of series, with equal numerosity, were generated and some spiked values were introduced at fixed time lags (i1=0.25N, i2=0.5N, i3=0.75N). The Monte Carlo simulation was replicated five times, varying the outlier magnitude  2r . The obtained results were analyzed to compute the total order underestimation percentage. The results reported in Table III show that the expert system is the more affected one, as could be expected, inasmuch as it is based on indices being functions of the series and residual variances. VI. SERIES SIMULATION An infinite moving average representation [22] is here adopted with 100 terms. The implemented procedure has been compared against the widely used Splus software and results have shown the statistical equivalence of generated series. With reference to the series inverse transformation, the following is noteworthy; when the hydrologist preferred Box- Cox transformation is considered, depending on BC, the generated series can contain data with no physical correspondence. Actually, if BC=0, an exponential inverse transformation is required, making the simulated peak values significantly increase and so possibly involving maximum wave-height of 50m, value that actually have never been observed in the Central Mediterranean Sea [24] or by any ocean buoy all over the world [23]. Moreover, if BC  0, the transformed series could have negative values, again with no physical sense. To overcome in some extends these problems, a method to remove negative value is applied retaining the original occurrence of calms (generally defined as sea states having Hmo0.20m); namely, a constant quantity is added to the generated series so that the original number of calms Nc is equal to the number of series elements xt0.2. Afterwards, the loess-smoothing procedure is used to trim off the peak values. These shortcomings are eliminated by using PLET. VII. APPLICATION TO OBSERVED WAVE DATA The analysed time series of significant wave-height was recorded by the RON directional wave-buoy located one mile offshore the Alghero coast – Sardinia (Italy), at a depth of about 100m. The analysed wave record was observed from 1 July 1989 to 31 December 2000, with a time interval of three hours, resulting in a series numerosity of 33615 observations. For a deeper description of both the Italian Data Buoy Network (RON), managed by ISPRA – Oceanographical Service, and the measured data see respectively [25] and [26]. The missing values are 1224, equal to 3.6% of the data. The frequency distribution of the gap-length showed that almost all gaps cover less than one day and that many of them (equal to the 75% of gaps) can be simply recovered by neighbouring interpolation. The remaining 80% of missing data have been recovered by (8) with  -0.169, a = -0.190, b = -0.043 The time series shows no significant daily non-stationarity but there is a clear seasonal component (fig. 3) having periodicity of about three months (2200 hours), which were identified and removed according to the herein illustrated procedure. Fig. 4 shows that the seasonal component is markedly affecting the mean value of the observed time series, whereas its standard deviation is nearly invariant within the averaged year. It is noteworthy that the application of the STL robust method give seasonal components much more stable than those computed by the classical averaging method as well as more detailed than those estimated by the Fourier representation (fig. 4). Lag k [hh] k [-] Fig. 3. Autocorrelation function of the observed time series at Alghero. -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0 3 0 6 0 9 0 1 2 0 1 5 0 1 8 0 2 1 0 2 4 0 2 7 0 3 0 0 3 3 0 3 6 0 -3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 mean component - LOSS smoothed mean component 1 year FFT mean component Std component - LOSS smoothed Std component 1 year FFT Std component -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0 3 0 6 0 9 0 1 2 0 1 5 0 1 8 0 2 1 0 2 4 0 2 7 0 3 0 0 3 3 0 3 6 0 -3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 mean component - LOSS smoothed mean component 1 year FFT mean component Std component - LOSS smoothed Std component 1 year FFT Std component -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0 3 0 6 0 9 0 1 2 0 1 5 0 1 8 0 2 1 0 2 4 0 2 7 0 3 0 0 3 3 0 3 6 0 -3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 mean component - LOSS smoothed mean component 1 year FFT mean component Std component - LOSS smoothed Std component 1 year FFT Std component -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0 3 0 6 0 9 0 1 2 0 1 5 0 1 8 0 2 1 0 2 4 0 2 7 0 3 0 0 3 3 0 3 6 0 -3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 mean component - LOSS smoothed mean component 1 year FFT mean component Std component - LOSS smoothed Std component 1 year FFT Std component  [m] t [dd]  [m] Fig. 4. Seasonal components (mean and standard deviation) of the observed time series at Alghero estimated by the classical hydrological method, by low order Fourier transform method and by the STL robust method (here, the window amplitude used by the loess smoothing is equal to 240 steps, which is nearly equal to one month). Time scale ranges from July to June. International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 694 Lag k [hh] k [-] Fig. 5. Autocorrelation function of the detrended, deseasonalized series at Alghero. F) [-] ) [-] Fig. 6. Cumulate frequency distributions of the observed time series as well as those (overlapped) of the standardized series by both Box-Cox and Probability Level-Equivalence transformations, as a function of the standard variable (ζ). Fig. 5 shows the autocorrelation of the detrended time series giving clear evidence of the performance of the applied method. Fig. 5 also shows a small residual fluctuation of the autocorrelation with a periodicity of about half a week. No physical causality could be disclosed in this cycling; accordingly, it has been considered tolerable and no more effort to remove it from the detrended series has been done. The time series turned to be non-Gaussian as well, and both (10) and (11) have been applied to recover the time series normality; the obtained results are shown in fig. 6. Both transformations have performed accurately and efficiently by turning the non-Gaussian time series into a Gaussian distributed one; nonetheless, the inverse Box-Cox transformation has showed to be failing by recovering all the storm parametric characteristics when the reverse task of generation is considered. Namely, several attempts were carried out by varying the Box-Cox exponent value BC in the range between 0 and 1, with a step of 0.01, as well as the width of the loess-smoothing window (see fig. 7 where some of the achieved results are reported). Unfortunately, none of the tested combinations gave fully satisfying results by obtaining a simultaneous reasonable agreement between non- exceedance cumulate frequency distributions of both storm duration and its peak-value wave-height. Actually, fixing BC = 0.0 gives a good agreement between the duration cumulate frequencies but produces too many sea state with unreal giant waves. Conversely, setting BC = 0.5 gives a fairly good agreement between the storm peak-value wave-height distributions but produces overestimated durations, with sea storminess greatly increased; actually, sea states with Hmo over the 5m threshold persist 40% more than the observed one. Hmo [m] F(Hmo) [-]  [hh] F() [-] Fig. 7. Comparison between the non-exceedance cumulate frequency distributions of the storm duration (up) and its peak-value wave-height (down), computed from the observed and simulated time series using the Box-Cox transformation, with different λBC and different width of loess window. With the aim to improve the generation task, the probability equivalence transformation was adopted. Accordingly, a probability law should be chosen to properly model the significant wave-height distribution. To this aim, several different probability laws were considered in literature, mainly focused on the distribution upper-tails (see, among others [18], [27], [28]). Thought the efforts made, the achieved results do not give any clear evidence of a true distribution [29]. Generally, it is considered that GEV type III describes better the upper tail, at the cost of larger deviations for small Hmo values, while the log-normal distribution fits better the distribution mode. Thus, GEV seems more appropriate for extreme-value analysis (see [26]), while the lognormal distribution seems more suitable for moderate- value analysis (e.g. fatigue-life analysis, estimation of the wave-energy resource, operativeness analysis, etc.). Here, several probability functions were tested and fitted to the observed data by using the complex modified method to minimize the overall least square error. The goodness of fit was verified by using the Kolmogorov-Smirnov test. The distribution functions having higher K-S confidence level were: GEV type III, three parameters lognormal, four parameters power-lognormal, Beta, gamma, 2 and f (see [30] for distribution details). Each of the above functions was then used in the series simulation, but only the GEV type III and the Beta distributions gave fully satisfying results. Namely, the lognormal and power-lognormal distributions involve exponential transformation and therefore present drawbacks International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 695 similar to the Box-Cox transformation with too many sea state with unreal giant waves. Hmo [m] F(Hmo) [-]  [hh] F() [-] Fig. 8. Comparison between the non-exceedance cumulate frequency distributions of storm duration (up) and its peak-value wave-height (down), computed from the observed and simulated time series using PLET with Beta and GEV III wave-height theoretical probability distributions. Lag k [dd] k [-] k [dd/hh] k [-] Fig. 9. Comparison between the non-exceedance cumulate frequency distributions of storm duration (up) and its peak-value wave-height (down), computed from the observed and simulated time series using PLET with Beta and GEV III wave-height theoretical probability distributions. Conversely, the gamma, 2 and f distributions gave slightly biased distribution lower tails, giving rise to storm persistence too long. Accordingly, only the results obtained implementing the GEV and Beta distributions are reported in fig. 8; the attained improvements are evident inasmuch as generated series comply both frequency distributions of storms duration and wave-height peak-value quite well. Finally, the Beta distribution was chosen for the generation task at Alghero owing to its slightly better performance in replicating the duration of the storm with extreme peak-value wave-height (upper tail of the frequency distribution). The parameters of the best fitting Beta distribution resulted A=0.0, B=96.6, k1=1.275, k2 =100.0. The optimal configuration of the linear parametric model at Alghero resulted in a second order Auto Regressive one, with parameters equal to 1 =1.017, 2 =-0.068 and residual variance 2r =0.087. Hmo [m] F(Hmo) [-] F(Hmo) [-] Fig. 10. Comparison between the non-exceedance cumulate frequency distributions of Hmo observed and simulated by the AR(2) model. Hmo [m] f(Hmo )[-] Hmo [m] f(Hmo )[-] Fig. 11. Comparison between the non-exceedance cumulate frequency distributions of storm duration (up) and its peak-value wave-height (down), computed from the observed and simulated time series using PLET with Beta and GEV III wave-height theoretical probability distributions. Hmo [m]  [hh] Hmo [m] Fig. 12. Comparison between the non-exceedance cumulate frequency distributions of storm duration (up) and its peak-value wave-height (down), computed from the observed and simulated time series using PLET with Beta and GEV III wave-height theoretical probability distributions. Fig. 9 shows the autocorrelation functions of the generated and observed series, revealing a nice matching for lags less or equal to a week as well as a nearly perfect seasonal trend. Figs. 10 and 11 respectively show the not-exceedance and occurrence frequencies of the significant wave-height. The achieved results are quite gratifying, in that distributions are in very nice agreement starting from values greater than 0.5m, value generally assumed as the lower threshold below which data are discarded both in extreme and climatic analysis. In addition, the over-threshold persistence curves are pretty overlapped (fig. 12); it is relevant to stress that the persistence curves almost exactly overlap for Hmo > 3m, given that many International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 696 engineering activities are limited or broken down by the occurrence of such sea states. TABLE IV. STATISTICAL SUMMARY OF BOTH THE SERIES OBSERVED AT ALGHERO AND THE AR(2) SIMULATED ONE. M in 1 st Q u . M e a n M e a n lo w e r c o n f. M e a n u p p e r c o n f. S td E r r M e a n M e d ia n 3 r d Q u . M a x V a r ia n c e S td D e v . S k e w n e ss K u r to sis Observed 0.06 0.40 1.23 1.22 1.24 0.01 0.90 1.60 9.10 1.21 1.10 1.81 4.09 Simulated 0.00 0.47 1.25 1.23 1.26 0.01 0.92 1.65 13.16 1.28 1.13 2.06 6.43 f(Hp) [-] f() [-] Hp [m]  [gg] Fig. 13. Comparison between the occurrence frequencies of storm-peak significant wave-height (Hp - on the left) and storm duration ( - on the right) for the observed time series and the simulated one. All these fine agreements reflect themselves into the descriptive statistics of the observed and simulated series (see Table IV). The only dissonant note is the mismatch between observed and simulated data kurtosis in Table IV, which points out a greater peakedness of observed frequency mode (fig. 11). Taking all these features in mind makes possible to say that the correlation structure of the observed data is very well reproduced into the synthetic time series. By considering the derived dataset of the storm features, characterized by the peak-value of the significant wave-height and by its duration over the threshold of 1m, the agreement is slightly less gratifying but still suitable. Namely, fig. 13 shows the occurrence frequencies of the storm duration and peak-value. The achieved results indicate that the simulated mild storms, characterized by a peak-value of the significant wave-height in the range of 1÷3m, are more frequent than the observed ones. On the contrary, the simulated violent storms, characterized by a peak-value of the significant wave-height in the range of 3÷5m, are fairly less frequent than the observed one. For the extreme storms, characterized by a peak-value of the significant wave-height greater than 5m, both simulated and observed ones exhibit nearly the same frequencies. Furthermore, the simulated storms show a little bit shorter duration. Actually, results indicate that the simulated short storms, characterized by a duration less or equal to 3 days, are more frequent than the observed ones. On the contrary, the simulated persistent storms, characterized by a duration in the range of 3÷10gg, are fairly less frequent than the observed one. Finally, the very long storms, characterized by a duration greater than 10gg, show nearly the same frequencies for both the simulated and observed series. 1 .1 2 5 2 .7 5 4 .3 7 5 6 7 .6 2 5 9 .2 5 1 0 .8 7 5 0.25 3.75 7.25 10.75 14.25 17.75 21.25 24.75 Simulated 1 .1 2 5 2 .7 5 4 .3 7 5 6 7 .6 2 5 9 .2 5 1 0 .8 7 5 0.25 3.75 7.25 10.75 14.25 17.75 21.25 24.75 Observed f(Hp,) [-] Hp [m]  [gg] f(Hp,) [-] Hp [m]  [gg] Fig. 14. Comparison between bivariate occurrence frequencies of storm- peak significant wave-height (Hp) and storm duration () for the observed series (on the right) and the simulated one (on the left). Hmo [m]  [hh] Hmo [m]  [hh] Fig. 15. Evolutions of real and simulated sample storms for severe (left) and mild (right) sea-state conditions. These trends reflect themselves into the bivariate density of occurrence frequency for the simulated series (fig. 14), which appears more peaked near the origin. Fig. 14 states also that the simulated rare storm events (small occurrence frequency) last shorter than the observed one, independently from the peak wave-height. The described deviations can be explained considering the evolution of sample storms for both the observed and generates series, in mild and severe conditions, reported in fig. 15; as shown, the storm decreasing-tails decay slightly faster for the generated events than for the observed ones. Moreover, the decay stops when generated wave-heights became very small while the observed ones show a more persistent behaviour; accordingly, a possible successive event reaches faster the sea-storm censoring-threshold of 1m. These elements concur to give a slightly greater duration of the observed storms. VIII. CONCLUSION This paper describes an improved methodology for the analysis, missing-value completion and simulation of an incomplete, non-stationary and non-Gaussian time series of wave significant height. The method analyses a finite-length time series to identify an ARMA model, which can be used to recover missing values, forecast short-term wave evolution or to generate arbitrarily long sequences of wave data, preserving the primary statistical properties of original dataset, including persistence over threshold, autocorrelation, non-Gaussian distribution and seasonality. Three main improvements to the general ARMA fitting procedure are introduced: a robust estimation of the seasonal components, and accurate method to compute model parameters along with an automatic expert system for the model optimal-configuration selection. These implementations International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 697 are effectively verified to be fine improvements. Namely, STL method computes very regular as well as detailed seasonal components; in addition, the Whittle’s approximation coupled with the complex-modified optimization-procedure give parameter-estimates with lower variance and unbiased mean when different series drawn out from the same process are analysed. Finally, if the observed series is unaffected by significant outliers, the expert system is able to automatically and properly identify the true model with rates very close to 50% and, generally, slightly overestimating the model total order and correctly identifying the right prevalence of the AR and MA model parts. The proficiency of the herein proposed methodology is demonstrated in this paper through comparisons of simulated series with data observed by the directional buoy of RON, located offshore Alghero coast – Sardinia, Italy. The achieved results point out that statistical properties of the observed and simulated time series are almost nearly equivalent; this nice agreement embraces winter and summer seasonal patterns, sea state sequencing, over-threshold persistence, occurrence and cumulate frequency distribution of significant wave-height as well as both the cumulate frequency distribution of the storm duration and its wave-height peak-value. Accordingly, the described ARMA-modelling procedure is an efficient tool in representing the wave-height climate. Its straightforward application is accordingly associated to the comprehension of sea state conditions, which is of central importance for many offshore and nearshore activities. Actually, estimates of risk for critical scenarios are often defined as some over-threshold responses of complex and interlaced systems; Monte Carlo can therefore be the only way to derive the probabilities of interest. Accordingly, even if there is a huge amount of data collected on ocean waves, which is jet geographically sparse and time limited, ARMA models can be very helpful, being able to provide large database of observed statistically-equivalent information. Moreover, taking into consideration the modern engineering-area of wave-energy conversion, a fresh and promising application of linear model is delineated. Actually, according with [31], a real-time control of converters is required to approach the optimal efficiency of wave-energy extraction; to this aim the knowledge of future incident wave elevation is mandatory. Treating wave surface fluctuations as a time series and applying an ARMA model, makes possible to predict incoming wave elevation only from its past history. Results achieved on real data from Galway Bay and Pico Island showed the proficiency of a linear model to render a very accurate prediction of the incoming swell waves for a lag up to two wave periods. The herein presented methodology can be promptly adapted to wave elevation time series excluding the seasonal components estimation. REFERENCES [1] P.TD. Spanos, “ARMA algorithms for ocean modelling”. Trans. ASME, J. Energy Res. Tech., vol. 105, pp. 300-309, 1983. [2] R.J. Sobey, “Correlation between individual waves in a real sea state”, Coastal Eng., vol. 27, pp. 223-242, 1996. [3] C. Guedes Soares, A.M. Ferreira, C. Cunha, “Linear models of the time series of significant wave height in the Portuguese coast”, Coastal Eng., vol. 29, pp. 149-167, 1996. [4] C.N. Stefanakos, G.A. Athanassoulis, “A unified methodology for the analysis, completion and simulation of nonstationary time series with missing values, with application to wave data”, App. Ocean Res., vol. 23, pp. 207-220, 2001. [5] P.C. Ho, J.Z. Yim, “A study of the data transferability between two wave-measuring stations”, Coastal Eng., vol. 52, pp. 313- 329, 2005. [6] S. Grimaldi, “Linear parametric models applied to daily hydrological series”. Journal Of Hydrologic Engineering, vol. 9, pp. 383-391, 2004. [7] G.E.P. Box, G.M. Jenkins, “Time Series Analysis: Forecasting and Control”, Holden-Day, San Francisco, 1976. [8] J.D. Salas, “Analysis and Modelling of Hydrologic Time Series”, in Handbook of hydrology, Maidment Editor, Eds. McGraw-Hill, 1993. [9] G.A. Athanassoulis, C.N. Stefanakos, “A nonstationary stochastic model for long-term time series of significant wave height”, J. Geoph. Res., vol. 100, pp. 16149-16162, 1995. [10] G.E.P. Box, D.R. Cox, “An analysis of transformation”, J. Roy. Stat. Soc., Ser B., vol. 26, pp. 211-252, 1964. [11] C. Cunha, C. Guedes Soares, “On the choice of data transformation for modelling time series of significant wave height”, Ocean Eng., vol. 26, pp. 489-506, 1999. [12] J. Beran, “Statistics for Long-Memory Processes”, Chapman and Hall, New York, 1994. [13] S.M. Kay, S.L. Marple, “Spectral analysis – a modern perspective”, Proc. of the IEEE, vol. 69, n° 11, pp. 1380-1419, 1981. [14] M.S. Taqqu, V. Teverovsky, “Robustness of Whittle-type estimators for time series with long-range dependence”, Stochastic Models, vol. 13, n° 4, pp. 723-757, 1997. [15] M.J. Box, “A new method of constrained optimization and a comparison with other methods”. Computer Journal, vol. 8 (1), pp. 42-52, 1965. [16] P.E. Gill, W. Murray, “Numerical Methods for Constrained Optimization”. Academic Press. 1975. [17] R. Piscopia, “On the optimal fitting of a ten-parameter model to observed wave spectra”. ISOPE 13th conf., vol. III, pp. 234- 240, 2003. [18] J.A. Ferreira, C. Guedes Soares, “Modelling the long-term distribution of significant wave height with the beta and gamma models”, Ocean Eng., vol. 26, nº 8, pp. 713-725, 1999. [19] G.M. Ljung, G.E.P. Box, “On a measure of lack of fit in time series models”, Biometrika, vol. 65, pp. 297-303, 1978. [20] W.S. Chan, “A comparison of some pattern identification methods for order determination of mixed ARMA models”, Statistics & Probability Letters, vol. 42, pp. 69-79, 1999. [21] P.J. Brockwell, R.A. Davis, “Introduction to Time Series and Forecasting”. Springer Verlag, New York, 1996. [22] T.W. Anderson, “The Statistical Analysis of Time Series”. John Wiley & Sons, New York, 1971. [23] R. Piscopia, R. Inghilesi, S. Corsini, L. Franco, 2004. “Italian Wave Atlas”, ed. Univ. Roma III, US Lib. Congr. G1989.21.C7, pp. 134, 2004. [24] P. Petrova, Z. Cherneva, C. Guedes Soares, “Distribution of crest heights in sea states with abnormal waves”. Appl. Ocean Res., vol. 28, pp. 235-245, 2006. [25] M. Bencivenga, G. Nardone, F. Ruggiero, D. Calore, “The Italian data buoy network (RON)”, Proc. Advances in Fluid Mechanics IX, ed. WIT, pp. 321-332, 2012. [26] R. Piscopia, R. Inghilesi, A. Panizzo, S. Corsini, L. Franco, “Analysis of 12 - year wave measurements by the Italian wave network”, Proc. 28th ICCE (Cardiff, UK), pp. 121-133, 2002. [27] M. Isaacson, N.G. Mackenzie, “Long term distribution of ocean waves: a review”, J. Waterways, Port, Coastal and Ocean Eng., vol. 10, pp. 93-109, 1981. International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 698 [28] Y. Goda, K. Konube, “Distribution function fitting for storm wave data”, Proc. 22nd ICCE Conf., vol. 1, pp. 1-14, 1990. [29] C. Guedes Soares, M. Scotto, “Modelling uncertainty in long- term predictions of significant wave height”, Ocean Eng., vol. 28, pp. 329-342, 2001. [30] M. Evans, N. Hastings, B. Peacock, “Statistical Distributions”, 3rd ed., John Wiley and Sons, 2000. [31] F. Fusco, J.V. Ringwood, “Short-term wave forecasting for real-time control of wave energy converters”, Sustainable Energy, IEEE Transactions, vol. 1, n° 2, pp. 99 – 106, 2010. International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-0181 www.ijert.orgIJERTV4IS090656 (This work is licensed under a Creative Commons Attribution 4.0 International License.) Vol. 4 Issue 09, September-2015 699 
work_qkvyiqezdvbz3b32xeckjoauo4	----	International Journal of Rotating Machinery, 9(2): 63–79, 2003 Copyright c© 2003 Taylor & Francis 1023-621X/03 $12.00 + .00 DOI: 10.1080/10236210390147399 Expert System Source Identification of Excessive Vibration R. Gordon Kirk and Zenglin Guo Virginia Polytechnic Institute and State University, Blacksburg, Virginia, USA The importance of vibration data in determining the con- dition of rotating machinery is well established in both the aircraft and the heavy equipment industries. Installation of noncontact displacement probes for shaft motion and either velocity or acceleration sensors for bearing cap or foun- dation motion is standard practice for equipment manu- facturers. Automation of the diagnostic evaluation of cer- tain simple faults can be easily implemented. The advances in computer languages in recent years have made it diffi- cult to keep pace with the graphical capabilities available to the programmers. One major concern is the content of the knowledge base and the method of modifying the knowl- edge base or the procedure of evaluating the confidence of a given identified possible cause of a problem. This article ad- dresses the application of one such expert system to a recent vibration problem on a 7-megawatt steam turbine–driven generator located at the power plant of Virginia Polytechnic Institute. Keywords diagnostics, expert, stability, vibration Vibration data is important in determining the condition of rotating machinery. Automation of the diagnostic evaluation of certain simple faults can be easily implemented. The power and speed of modern computer chips and data acquisition systems have opened a new range of capabilities not possible even 5 years ago. The mass-produced items are the most likely candidates for small expert systems. In the automotive industry, driver- viewable screens are common, and more technical diagnostics of a power plant’s performance are available to trained mechan- ics with access to the proper equipment for connecting to the Received 20 May 2001; accepted 18 October 2001. Address correspondence to R. Gordon Kirk, Department of Me- chanical Engineering, Randalph 0238, Virginia Polytechnic Institute, Blacksburg, VA 24061, USA. E-mail: gokirk@vt.edu onboard data-acquisition system. While it is possible for each user installation to develop a unique expert system, that is highly inefficient and, in general, undesirable. The advances in computer languages in recent years have made it difficult to keep pace with the graphical capabilities available to programmers. The end-users expect the system to have the latest look and feel. Many older technicians and experts may see little need for more advanced capabilities that go be- yond the raw vibration signatures, but younger technicians and engineers without the same vast personal experience may not be familiar with the faults of the machinery and may require a more automated system to solve problems. Over the past 15 years, there has been an explosive growth in the use of expert systems in engineering applications. More than 150 commercial expert systems are currently on the mar- ket for engineering-specific tasks. The only areas for which a greater number of expert systems have been developed are the business, medical, and manufacturing fields (Durkin 1996). Many of these systems are meant for use in power plants and for other heavy industrial applications. Most of them (partic- ularly those designed for nuclear power plants) have been de- signed to aid plant personnel in sorting through and interpreting the assorted alarms and warnings the plant equipment gener- ates (Corsberg 1987; Vale and Mows 1993). Few, however, have been designed to be low-cost systems for smaller power plants, capable of reading data from standard plant sensors and creating its own set of alarms and bringing them to the attention of the operator. Major efforts were made at the Virginia Polytechnic Institute (Virginia Tech) Rotor Dynamics Laboratory to develop an inex- pensive system to monitor, evaluate, and diagnose rotating ma- chinery in industrial applications such as power plants through the use of a personal digital computer. That research began with the development of an off-line rule-based expert system (Kirk et al. 1989a, 1989b; Typrin et al. 1992). The extension of this off-line system into an on-line expert system capable of monitor- ing multiple machines with trending and diagnostic capabilities was the basis of additional work at Virginia Tech (Hoglund 1989; Kirk et al. 1991; Pawtowski 1996; Typrin 1992). That research 63 64 R. G. KIRK AND Z. GUO was similar to the VARMINT system that was developed at Design Maintenance Systems in North Vancouver, BC (Liddle et al. 1993). A common feature of many expert systems is the use of a hard-wired knowledge base that was probably developed on the basis of experience and is capable of detecting known problems having known systems. Others systems have the ability to learn from the system under evaluation and add to the existing knowl- edge base in some fashion. It is now common for automobiles to detect low water, low oil, open or not fully closed doors, and improper lighting elements. The level of sophistication contin- ues to increase in many areas of mass-produced items with cost structures able to support the computer logic essential for de- tecting and displaying the results of the system analysis. Similar dedicated systems will soon be developed for all critical-path rotating machinery. EXAMPLES OF FAULTS OF INDUSTRIAL MACHINERY It is essential to have knowledge of many known machin- ery faults to be able to diagnose a new problem in a different, but perhaps similar, piece of rotating machinery. In addition to 15 years in academics and in consulting for machinery analy- sis and problem solving, the author’s (RGK) own personal in- dustrial experience consists of 3 years in aircraft engine devel- opment work and another 10 years in working with industrial turbomachinery, including centrifugal compressors, steam tur- bines, axial compressors, screw compressors, power turbines, hot gas expanders, and centrifugal pumps. A large portion of this experience involved problems with high-pressure centrifu- gal compressors because of the attention given these machines as a result of the lost value of product in the event of an unscheduled shutdown. Gas pipelines must stay in service all winter in most regions of the United States and United Kingdom, for example. High penalties are paid if the proper volume of gas is not deliv- ered. Many problems with compressors are related to bearings and seals. Others are caused by balance shift due to impeller fits, improper installation of couplings, misalignment, poor founda- tions, entrapped liquids, incipient surge, rotating stall, pressure pulsation reflection in pipe sections, light rubs, seal rubs, poor lubrication to bearings, and so forth. Each problem must be described to distinguish the vibration signatures and phase re- lationships or other measurable parameters from other possible problems. Figure 1 shows a classic case of oil seal reexcitation of a machine at its first critical speed. This condition is also known as shaft whip. The instability frequency locks onto the shaft’s first critical speed, and the level of subsynchronous vibration climbs quickly in many cases, causing the machine to trip on high vibration. In this case, the oil seals are essentially locked and are acting as 360-degree bearings to drive the instability. The remedy is to modify the seals and soft-start the compressor so as to center the oil seal rings (Kirk, 1986). The resulting stable machine frequency spectrum is shown in Figure 2. FIGURE 1 Oil seal–excited shaft whip instability in a centrifugal compressor. A similar spectrum can result from the driving mechanism of a gas labyrinth seal or from aerodynamic excitation caused by pressure distortion in a high-pressure compressor. If the oil seals are not locked but are freely floating, a different spectrum of low-ratio subsynchronous vibration that tracks run- ning speed results. A classic example is shown in Figure 3. In this case (Kirk and Simpson 1985), the problem was a warped seat- ing endplate for the oil seal ring–lapped radial sealing surface. This caused the seal-lapped surface to be destroyed in a short time. The frequency spectrum, with the seal seating corrected, is given in Figure 4. At high speed and full pressure, with live gas, the same com- pressor as that referred to in Figures 3 and 4 shows the shaft whip instability, as shown in Figure 5. In this case, the excita- tion mechanism is not the oil seals, but the gas labyrinth seal of the balance drum. The fix was to provide a purge flow at the balance drum’s entry so as to block and prevent the forward- swirling flow from last-stage impeller leakage to the balance drum. FIGURE 2 Elimination of oil seal–excited shaft whip instability in a centrifugal compressor. EXPERT SYSTEMS 65 FIGURE 3 Oil seal ring instability driving compressor vibration at low order of rotor speed after 5000 rpm. Other examples of machinery faults in actual shop or field operations of full-scale machinery have been published (Kirk 2000). The problem with fault identification based on limited data is illustrated by Figures 1 through 5. Similar spectra have been observed in numerous other pieces of machinery but result FIGURE 4 Spectrum after seating of oil seal ring and elimination of instability. FIGURE 5 (a) Shaft whip instability caused by balance drum labyrinth. (b) Reduced shaft whip instability with balance drum labyrinth radial shunt line. from different causes. The expert must be capable of deciding which mechanism is causingthe instability. Many problems are specific to a particular class of machinery or to machinery with certain components. Only if full details are known can an accu- rate prediction of cause be made with confidence. 66 R. G. KIRK AND Z. GUO TABLE 1 Possible Problems in a Centrifugal Compressor Seal oil excitation Insert bearing oil whirl Heavy rub Thrust-bearing damage Seal oil whirl Loose rotor component Seal whirl Rotor critical Shaft whip Coupling critical Light rub Overhang critical Loose bearing shell Electrical excitation Unbalance Ground loop in instrumentation Electrical or mechanical Torsional resonance glitch Misalignment Process-related problem Bowed rotor Gear problem Axial rotor rub Labyrinth seal excitation EXAMPLE OF EXPERT SYSTEM INFORMATION For a problem to be diagnosed, it is essential to know the signature, or characteristics, of the problem. It is also desirable that the characteristics be unique to the given problem. This re- quires that a human expert know a vast amount of information or that such information exist in a retrievable data structure in a computer system. A major hindrance to generation of such a knowledge base is the possible perception by the human expert working in industry that putting all of his knowledge into a com- puter database would reduce his own usefulness to his company. While this may be true in some cases, the majority of problems requiring a human expert have not been solved previously and, hence, are not in any computer knowledge base. Table 1 lists possible problems in a compressor. Tables 2 through 4 provide lists of conditions that can be used to determine whether one of these particular rules is a likely to be the cause of a current problem. Lists of problems and conditions in a steam turbine are shown in Tables 5 and 6. The corrective action for misalignment is given in Table 7. Table 8 illustrates the desirable and necessary option for a knowledge base editor. A knowledge base must have an editor to update and add new information to allow the latest known vi- bration characteristics to be available in the expert system. The neural network of this system is the human expert. TABLE 2 Conditions for Unbalance Rule: Unbalance The machine does not have subsynchronous whirl. It has strong one-per-revolution vibration. It is actually running. The shaft vibration is dominant, that is, stronger than that of the casing or bearing cap. The amplitude increases uniformly with speed. The amplitude is not influenced by a change in operating conditions. TABLE 3 Conditions for Shaft Whip Rule: Shaft whip The machine operates above first critical. It has a flexible shaft. It has tight clearance bearings. It has an operating speed greater than 2.0 × Ncr. It has nonsynchronous whirl. TABLE 4 Conditions for Labyrinth Seal Excitation Rule: Labyrinth seal excitation The machine operates above the first critical frequency. There is no shunt line on the balance piston. The vibration is dominant subsynchronous at the first critical frequency. The compressor has a high-pressure drop across the balance piston. The machine has tilting-pad bearings. The unit has no problems at reduced loading. An increase in was the source of the vibration increase TABLE 5 Possible Problems in a Steam Turbine Rotor unbalance Misalignment Loss of material Slipped coupling Rub (on-line/at speed) Instability Seal rub (on rollup or rolldown) Steam pressure loading Bad (fluid-film) bearing Cracked rotor Soft foot condition Operating near a critical Rotor bow (on-line bow or rub) Overhang critical Resonant pedestal Poor (lost) insulation on inlet Coupling unbalance Electrical glitch or scratch Bent rotor/excessive runout under probe TABLE 6 Conditions for Misalignment in a Steam Turbine Rule: Misalignment The machine is a turbine. The 1× vibration changes with loading or temperature. The vibration characteristics are generally repeatable. The bearing temperatures are in the normal range for this machine. A precision level check across adjacent bearings shows a large change with load. The pedestal axial vibration is greater than 40% of the radial vibration. The machine has occasionally had what appeared to be stability problems. EXPERT SYSTEMS 67 TABLE 7 Corrective Action for Misalignment in a Steam Turbine Rule: Misalignment Limit unit loading to where vibration is acceptable (for the short term). Measure hot-to-cold alignment change and then realign unit. EVALUATION LOGIC Theexistence of a knowledge base is necessary but not suffi- cient for ranking the possible causes of a particular problem. It is essential to rank each possible cause; the resulting confidence factor for each rule is then used to rank the most likely causes. The work at the Virginia Tech Rotor Lab adopted a three-phase logic in which a condition is true, false, or unknown. The equa- tion used in a prolog expert program is given by Hoglund (1989). Each rule requires an original confidence factor, which is part of the knowledge base, and each evaluation of a problem generates a modified confidence factor for each rule. The results of apply- ing such logic during a typical session are shown in Table 9. It is necessary to have an original confidence factor for each rule. The result of the evaluation of all information then produces a modified original confidence factor (MOCF). For automated expert systems, the response must be deter- mined to be true, false, or unknown by further computer logic. The Virginia Tech on-line system used automated standard eval- uation to determine whether vibration levels were low, nor- mal, or excessive. Other features are documented by Liddle and colleagues (1993). AUTOMATION OF DIAGNOSTICS The on-line automation of an expert system similar to that just described would seem to be a simple task. First, some ques- tions must be answered by the control room operator of a given machine. Then the data are collected from a connected data TABLE 8 Key Options Required for a Knowledge Base Rule Menu F1 Create a new database F2 Views all rules in the database F3 Add a new rule to the database F4 Choose an existing rule to edit F5 Choose a database to edit F6 Exit the Editor program F7 Display remaining RAM memory What part of this rule would you like to edit? F1 Conditions F2 Suggestions for corrective procedures F3 Related standards F4 References F6 Exit to previous menu (Rule Menu) TABLE 9 Modified Original Confidence Factors Problem MOCF I have concluded that your machine’s difficulty may be due to: Seal whirl 0.90 Insert-bearing oil whirl 0.90 Seal-oil whirl 0.89 Seal-oil excitation 0.87 Axial rotor rub 0.85 Shaft whip 0.81 acquisition system and the answers obtained. For a specific ma- chine, this could be simple. If the on-line expert system is to operate for all machine systems, it is necessary to configure the data acquisition in a simple manner, so a well-written program is required. In 1989, specific instruments had to be purchase for each function needed, for example, a Bently Nevada Corp Digital Vector Filter (BNC DVF) to pull data concerning the amplitude and phase of shaft probes, and the spectrum required a specific box to collect and obtain the frequency content. By 1994, all these instruments were virtual monitors but were still programmed. Today, it is possible to use a laptop to collect data via available hardware and MATLAB to compute the spectrum, response, and phase. Desirable Features The desirable features of an on-line expert diagnostic system are as follows: The ability to run in off-line mode if no machine data is available The ability to update the rule condition list The ability to add new rules The ability to recommend solutions The ability to refer to publications concerning the problem The ability to refer to standards relating to the problem The ability to compute spectrum content and response and phase by using standard calculations The ability to allow initial setup to be retained in the control room of the computer system The ability to hold files of machine specifications The ability of the operating system and graphics to be compatible with state-of-the-art languages EXAMPLE OF APPLICATION: OFF-LINE EXPERT SYSTEM A recent and, in fact, the first application of the off-line ex- pert from Virginia Tech to a real problem with an unknown cause will be discussed briefly to further illustrate the use and applica- tion of such an expert system. The power plant at Virginia Tech has a 7-megawatt steam-turbine–driven generator that operates to provide power for peak and normal loading from the student dorms and, in addition, supplies power to a small part of the town of Blacksburg. The 15-year history of the turbogenerator 68 R. G. KIRK AND Z. GUO includes problems with misalignment resulting in bearing insta- bility due to the unloading of the close coupled inboard bear- ings. The original journal bearings were converted to a pressure dam design to help reduce this instability problem. The system had been running with acceptable vibration for 6 years. In the summer of 2000, the turbine was upgraded to produce a more dependable and efficient driver. The new turbine rotor plus new bearings and seals for both the turbine and generator were being installed in early autumn, and start-up was needed prior to cold weather. When the rotor was to be installed, it was discovered that an original seal fix had not been documented, so the rotor could not be installed. The OEM had modified seal parts on site in 2 days, and thus this potentially major problem with the $1.2 million re-rate was averted. The rotor was installed, aligned, and slow-rolled. The start-up was then apparently of no concern and no in- strumentation or help was requested of the Virginia Tech Rotor Lab staff. Previously, BNC DAI 108 data collection had been used when the bearing instability problems were causing near trips some 6 years earlier. In addition, the on-line expert sys- tem was installed and developed at Virginia Tech just after that, but since the problems had been solved, the final system was never upgraded and put into service; the data collection boards are still installed behind the control panel and the lines still run to the instrumentation area where the computer monitoring was to be conducted. The turbine was down each summer, which made verification of the on-line system more difficult. The final on-line system was never installed—completion was delayed by numerous personnel problems and funding problems—and the final computer system was never installed at the power plant. On the Monday following the first attempted start of the new turbine on the previous weekend, a phone call was received from the power plant’s instrumentation engineer. They had not been able to get the unit to function because the vibration was excessive—above the warning level and near trip. The initial suspicion was high subsynchronous vibration due to misalign- ment, suspicion was or maybe a rub in the turbine where the retrofit seals had been installed. The first visit to the power plant, less than 25-min walk from the Mechanical Engineering Department, revealed that the vibration, in the turbine was actually very low. It was in the generator that the vibration was high, especially in the outboard probes. The generator bearing caps and turbine foundation had excessive vibration that was evident even without taking read- ings but that was confirmed by hand-held velocity probe read- ings taken using the small BNC TK-83 hand-held field-balance monitor. The concrete block foundation under the generator had in excess of 1 mil of vibration at running speed. The generator bearing caps were covered at that time, so initial confirmation of the bearing caps levels was not made. Concerns shifted to potential causes of the high vibration. Problems with generator windings or other damage from the re- rate work were suspected; misalignment and coupling problems were not high probabilities because the turbine was running at TABLE 10 Results Based on the Rockland Knowledge Base Problem MOCF I have concluded that your machine’s difficulty may be due to: Eccentric motor mass (no electrical problem) 0.88 Unbalance (centerhung machine) 0.83 Misalignment (parallel) 0.70 such a low vibration. A call to the engineering department of the generator’s manufacturer was planned by the power plant manager, if it could be arranged. The use of the off-line expert system was tried first, using a general knowledge base for machinery that had been developed in the late 1980s based on the Rockland diagnostic table. The summary of possible causes of this problem is shown in Table 10. The results were directly related to how the questions from the expert system were answered. Observation of this information helps to clarify why a specific cause has been selected. The utility of the ability to include the influences of elements not known allows a greater number of possible problems to be shown by the system. Table 11 displays the query and the user response for this specific session. The results are likely to be of no surprise and TABLE 11 Responses of the Rockland Knowledge Base to Results Shown in Table 10 The conditions used to prove eccentric motor mass (no electrical problem) are: Y The machine is a motor unit. Y The frequency is 1×, i.e., running speed. Y It also has vibration at 1× and 2× line frequency. Y The radial vibration is dominant. D The vibration is steady. Y The vibration envelope is narrowband. The conditions used to prove unbalance (centerhung machine) are: Y This machine is centerhung. Y The frequency is 1×, i.e., running speed. Y The radial vibration is dominant. D The vibration is steady. D The inboard and outboard vibrations are in phase. Y The vibration envelops is narrowband. The conditions used to prove misalignment (parallel) are: D It has both 1× and 2× vibration. D The vibration is steady. Y The radial vibration is dominant. D The inboard and outboard vibrations are 180 degrees out of phase. Y The vibration envelope is narrowband. Y, yes; N, no; D, don’t know. EXPERT SYSTEMS 69 TABLE 12 Results of the Motor Knowledge Base Problem MOCF I have concluded that your machine difficulty may be due to: Seal rub 0.87 Weak/light foundation 0.76 Bad (fluid film) bearing 0.71 Loose bearing/pedestal 0.68 Loose part 0.64 Misaligned journal bearing 0.63 Resonant pedestal/system 0.62 Loss/gain of material 0.61 Rotor unbalance 0.61 are, in fact, similar to what any experienced machinery engineer would place high on the list of possible causes. Another knowledge base designed specifically for motors had been developed and was tried next. The results of this knowledge base’s evaluation are shown in Table 12. Here, more possible causes of the excessive vibration were offered, so the knowledge base was of greater help in resolving the problem. The motor-specific knowledge base gave one answer not ex- pected by the authors. This could indicate a lack of experience in motor or generator vibration problems, and it precisely demon- strates the importance of having an expert system in place to double-check and, more important, to enhance personal experi- ence and knowledge. The search also revealed concern for the foundation, the bear- ings, lost parts, and general rotor unbalance. These results were generated prior to the second visit to the power plant (Figs. 6 and 7). Another consultant had been called in and had taken data using his data acquisition system. Those results pointed to possible alignment problems and there was also an indication of impacting or contact in a bearing. Upon comparison of the results, a recommendation of inspection for a seal rub on both ends of the generator was agreed upon and FIGURE 6 The power plant at Virginia Polytechnic Institute. FIGURE 7 A Worthington steam turbine. accepted by the plant manager. A seal rub was predicted to have the highest probability of being the cause of the problem. The plant manager also insisted on a check of the alignment during the teardown and seal-check work. The response listing for problem, seal rub, are presented in Table 13. The resolution of a problem is a very complicated process as illustrated by the response list for this problem. At this point, the machine had not been opened and no hard-copy data were available from a start-up or shut-down. The question on the table was Is this a seal rub? The next day, it was confirmed that a 4- to 6-mil interference had been detected during the seal check in the outboard generator TABLE 13 Responses of the Motor Knowledge Base to Results Shown in Table 12 The conditions used to prove Seal rub are: Y The 1× vibration is nearly constant with loading and/or temperature. Y There is more 2× vibration in the motor than usual. Y The 1× vibration on the casing is higher than usual. D The phase angle either wanders or is cyclical at constant loading. Y The vibration characteristics are generally repeatable. D Recent balancing attempts have not been as predicted. Y The vibration on the motor has changed recently. Y The motor vibration is higher than the driven-equipment vibration. Y An oil sample shows that the oil is relatively clean. Y The bearing temperatures are in the normal range. Y There is no unusual noise coming from the motor. N There is an unusual noise in the bearings. Y Dial indicator check shows that the shaft is true. D The vibration phase angle changes opposite to rotation. Y Seal work has recently been completed. D The shaft relative vibration waveform appears flattened or nonuniform. Y, yes; N, no; D, don’t know. 70 R. G. KIRK AND Z. GUO FIGURE 8 Seal rub area at the generator’s outboard bearing, showing seal rub marks in Shaft. FIGURE 9 Scraping in the seal clearance. FIGURE 10 Steam end, showing probe location. FIGURE 11 Replacing the coupling cover after the fix. FIGURE 12 DAI-108, FFT, DVF-2, orbit O-scopes, switchbox, ADRE computer, and printer. FIGURE 13 Vibration meters after the seal fix. EXPERT SYSTEMS 71 FIGURE 14 Shaft probes on the turbine’s outboard. FIGURE 15 Inboard probe locations. FIGURE 16 Location of outboard generator. FIGURE 17 Gear coupling hub and sleeve. FIGURE 18 Final data-acquisition system for orbits. FIGURE 19 Steam whirl peak hold and Instantaneous spectrum. FIGURE 20 OB generator inspection port. FIGURE 21 Warning sign from generator shown in Figure 20. 72 R. G. KIRK AND Z. GUO FIGURE 22 Turbine inboard vertical, 11-09-2000. bearing. The outer seal was scraped to secure the proper clear- ance (Figs. 8 and 9). The inner air seal was not removed nor was the clearance increased. The alignment had been checked and was as desired for the hot machine; hence, no change was made to the existing align- ment. The machine was readied for start-up (Figs. 10 and 11). Before start-up, a data collection cart was moved from the rotor lab of Virginia Tech using DAI-108, and cables were run for FIGURE 23 Generator inboard vertical, 11-09-2000. all eight bearings from the BNC control panel meters, as shown in Figure 12. The set-up was just finished and the first check of the data acquisition system was the start-up. Unfortunately, a high level of 60-cycle ground loop was in the instrumentation, so the at-speed readings were originally in error as they were recorded by the DAI-108. The meters were correctly grounded and isolated. The start-up was recorded using the DAI-108 for documentation of the system’s vibration. The meters were all EXPERT SYSTEMS 73 FIGURE 24 Generator outboard vertical, 11-09-2000. below the warning levels at that time (Fig. 13), but the high- est was the generator’s inboard probes, so the mechanics were wishing the inboard bearing seals had also been scraped. The levels were close to the warning setting, but the actual vibration was lower because the meters overall were taking into account the scratches on the shaft. FIGURE 25 Turbine outboard bearing, horizontal polar plot. The probe locations are shown in Figures 14, 15, and 16. The outboard generator probes were under the enclosure on that end and the shaft was not easy to see, nor was it easy to take stick readings. The gear-coupling hub and flange that were changed are shown in Figure 17. Continuous running at full power was monitored using instrumentation as shown in Figure 18; it is 74 R. G. KIRK AND Z. GUO FIGURE 26 Turbine inboard bearing, horizontal polar plot. FIGURE 27 Generator inboard bearing, horizontal polar plot. EXPERT SYSTEMS 75 FIGURE 28 Generator outboard bearing, horizontal polar plot. FIGURE 29 Turbine trend plot, outboard horizontal. 76 R. G. KIRK AND Z. GUO FIGURE 30 Generator trend plot, outboard horizontal. FIGURE 31 Generator spectrum, outboard vertical. EXPERT SYSTEMS 77 FIGURE 32 Generator orbit plots, outboard 1×, compensated. FIGURE 33 Turbine orbit plots, inboard and outboard, uncompensated. 78 R. G. KIRK AND Z. GUO FIGURE 34 Generator’s orbit plots, inboard and outboard, uncompensated. interesting to note that the vibration levels of the turbine in- creased slightly for 1× readings. Of interest was the outboard turbine bearing, which was showing an unsteady subsynchronous whirl that was dependent on steam temperature. The colder the steam, the larger would be the whirl component of the vibration. The levels of peak hold over a 24-h period were consistent from day to day at a maximum of near 0.5 mil, but when monitoring during a day, it was usually found to be lower, at about 0.2 mil. This limit-cycle steam whirl is not unexpected (see Figs. 19–21). The bearings in the machine are of pressure dam design and were installed to control bearing- or steam-induced instability. A small portion of the reduced start-up data is shown in Figures 22 through 34. The continuous monitoring data shown in Figures 29 and 30 indicate that a load or steam source was causing a step increase–decrease and then a decrease–increase in overall vibration at the generator’s outboard location. Closer evaluation would have been required to determine the exact cause. The answer is not important, unless the step changes were going to near trip condition, so it was left as an exercise for future attention. CONCLUSIONS This application of an expert system illustrates both the po- tential power and the lingering problem associated with the use of expert systems for general machinery diagnostics. It can be EXPERT SYSTEMS 79 stated that the first real application was a success in that one of the knowledge-base files did indeed identify the source of the problem. Other problems were also identified, and they may continue to be a source of vibration. However, it is also true that the system identified potential problems that were proven not to exist. Obviously, the system would be a failure if only incorrect solutions were identified. This success of this first real application of the Virginia Tech’s off-line expert system renews excitement about the potential power of such systems. It also points out the need for additional and continual update of the information in the current knowledge base. The attempts to start a databank in the early 1980s were total failures due to lack of willingness or failure to grant permis- sion for release of data by large companies. A request for data and machine-specific information resulted in only two replies and they both contained incomplete information. More recently, discussions were in progress to establish an Internet site to col- lect information and solve problems. The projected cost was substantial, so the progress was zero. Others have built expert system engines with no knowledge base. Little further progress has been reported. The simple rules and conditions need to be improved to better determine specific problems and then build more complex rules and machine-specific rules. The task seems simple. REFERENCES Corsberg, D. 1987. Alarm filtering: practical control room upgrade using expert system concepts. Intech 34:39–43. Durkin, J. 1996. Expert Systems: A view of the field. IEEE Expert Intelligent Systems and their Applications 11:56–63. Hoglund, J. R. 1989. An expert system for off-line analysis of rotat- ing equipment. (MS thesis, Virginia Polytechnic Institute and State University) Blacksburg, VA: Virginia Polytechnic Institute and State University. Kirk, R. G., and Simpson, M. 1985. Full-Load testing of an 18,000 H.P. gas turbine driven centrifugal compressor for offshore platform service. Symposium on Stability, NASA CP-2409. Kirk, R. G. 1986, November. Oil seal dynamics: considerations for analysis of centrifugal compressors, Proceedings of the 15th Turbomachinery Symposium. College Station, TX: Texas A&M University. Kirk, R. G. 2002, August. Condition monitoring and diagnosis of rotat- ing machin. International Symposium on Machine Condition Mon- itoring and Diagnosis, Paper Presented at Annual Meeting of the Japan Society of Mechanical Engineers. Meijo University, Nagoya, Japan. Kirk, R. G., Hoglund, J., and Keesee, J. 1989a. Application of arti- ficial intelligence for rapid evaluation of turbomachinery dynamic response and stability. 149–154. Proceedings of the JSME Inter- national Symposium on Advanced Computers for Dynamics and Design. Tsuchiura, Japan: Kirk, R. G., Hoglund, J., and Mondy, R. E. 1989c, September. Devel- opment of a PC-based off-line expert system for evaluation of turbo- machinery response, 439–444. Proceedings of the 1st International Machinery Monitoring and Diagnostic Conference. Las Vegas, NV: Union College. Kirk, R. G., Pawtowski, E. C., and Typrin, M. 1991, April. Intelligent filter for a PC-based expert system, 109–116. Proceedings of 45th Meeting of the Mechanical Failures Prevention Group. Annapolis, MD: . Liddle, ., Lan, ., Reilly, ., and Steven, . 1993. Diagnosing vibra- tion problems with an expert system. Mechanical Engineering —: 54–55. Pawtowski, E. C. 1996. Development of the Second-Generation IMTS (Intelligent Monitoring and Trending System) and WOT (Wizard of Tech) Expert System for Rotating Machinery. (MS thesis, Virginia Polytechnic Institute and State University). Typrin, M., Pawtowski, E. C., and Kirk, R. G. 1992, April. A PC-based expert system for rotating machinery,–. International Conference on Rotordynamics. Venice, Italy : . Typrin, M. 1992, February. IMTS (Intelligent Monitoring and Trend- ing System): A PC-based Monitoring System for Rotating Ma- chinery. (MS thesis, Virginia Polytechnic Institute and State University). Vale, Z. A., and Mows, A. M. 1993, August. An expert system with temporal reasoning for alarm processing in power system control centers, 1307–1313. IEEE Transactions on Power Systems 8:3. International Journal of Aerospace Engineering Hindawi Publishing Corporation http://www.hindawi.com Volume 2010 Robotics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Active and Passive Electronic Components Control Science and Engineering Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Rotating Machinery Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Journal of Engineering Volume 2014 Submit your manuscripts at http://www.hindawi.com VLSI Design Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Shock and Vibration Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Civil Engineering Advances in Acoustics and Vibration Advances in Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Electrical and Computer Engineering Journal of Advances in OptoElectronics Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Sensors Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Modelling & Simulation in Engineering Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Chemical Engineering International Journal of Antennas and Propagation International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Navigation and Observation International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Distributed Sensor Networks International Journal of 
work_qlkly4g5n5gvhgy4z3lc5sgq7u	----	Identifying web sessions with simulated annealing Expert Systems with Applications 41 (2014) 1593–1600 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Identifying web sessions with simulated annealing 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.08.056 ⇑ Corresponding author. Tel.: +56 2 7180900. E-mail addresses: tomas.arcec@usach.cl (T. Arce), proman@dii.uchile.cl (P.E. Román), jvelasqu@dii.uchile.cl (J. Velásquez), victor.parada@usach.cl (V. Parada). 1 Tel.: +56 2 7180900. Tomás Arce a,1, Pablo E. Román b, Juan Velásquez c, Víctor Parada d,⇑ a Departamento de Ingeniería Informática, Universidad de Santiago de Chile, Av. Ecuador 3659, Estación Central, Santiago, Chile b Center of Mathematical Modelling (CMM) UMI CNRS 2807, Universidad de Chile, Av. Blanco Encalada 2120, Piso 7, Santiago, Chile c Departamento de Ingeniería Industrial, Universidad de Chile, República 701, Santiago, Chile d Departamento de Ingeniería Informática, Universidad de Santiago de Chile, Av. Ecuador 3659, Estación Central, Santiago, Chile a r t i c l e i n f o Keywords: Web usage mining Web session Simulated annealing a b s t r a c t Delivery of efficient service through a web site makes it compulsory in the redesigning stage to take into account the behavior of the users, which can be studied by means of a web log file that partially records information about user visits. The reconstruction of all of the sequences of pages that are visited by users who browse a web site is known as the web sessionization problem, and it has been formulated by means of an integer programming model; however, because a web log can accumulate a large amount of infor- mation, it is necessary to reconstruct the sessions over a period of weeks or months, thus the solution to this problem requires a long computational processing time. This paper presents a heuristic approach based on simulated annealing for the sessionization problem. Using this approach, it has been possible to reduce the processing time up to 166 times compared to the time that is required for the integer pro- gramming model. Furthermore, the metaheuristic solution finds new optimum values, which achieve increases on the order of 17% in the best cases. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction The Internet has become a flourishing source of data for differ- ent research fields related to social networks, such as sociology (Lin, Jheng, & Yu, 2012; Nohuddin et al., 2012), marketing (Fong, Zhou, Hui, Tang, & Hong, 2012; Wang, Ting, & Wu, 2013) and com- puter science (Devi, Devi, Rani, & Rao, 2012; Yin & Guo, 2013). Also, the massive usage of the Internet drives many lucrative businesses such as e-commerce. Thus, it is increasingly necessary for web sites to be designed with a structure that makes it easy for the user to obtain the service (Wang & Ren, 2009). To that end, it is necessary to study the behavior of the users which is partially kept in a pri- vacy-compliant data file saved on the web server known as a web log. Legislation in several countries forbids the storage of personal information in order to safeguard personal privacy (Mayer & Mitchell, 2012). In order to follow national laws, internet compa- nies must rely on the anonymity of web logs for extracting infor- mation about their user preferences and browsing behavior (Velásquez, 2013; Velásquez & Palade, 2008). The generated browsing sequences represent an input source for discovering the behavior patterns of users who visit a web site (Cooley, Mobasher, & Srivastava, 1999; Kosala & Blockeel, 2000; Tao, Hong, Lin, & Chiu, 2009). Every time a user requests a web page or some resource con- tained on it such as images, videos, and sounds, a new record is created in the web log. The generated information allows the detection of the most visited pages, the common page access sequences, the users’ preferred contents and even the generation of user profiles (Choi & Lee, 2009; Nasraoui, Soliman, Saka, Badia, & Germain, 2008). Such detection is called web usage mining oriented toward extracting knowledge from web logs (Román, L’Huillier, & Velásquez, 2010), which requires the extraction of individual sequences of a user’s web interaction while maintaining anonymity (Mayer & Mitchell, 2012). The following data are typi- cally recorded in the file: the IP address from which the inquiry is made, the time and date of the inquiry, the requested resource, a code indicating the result of the operation, the number of bytes transferred in the request, a string that contains the address of the site from which the present request was originated and the browser and operating system that was used. The generated browsing sequences represent an input source for discovering the behavior patterns of users who visit a web site (Cooley et al., 1999; Kosala & Blockeel, 2000; Tao et al., 2009). With expert assistance the information stored in the web log can be used to support decision-making that can help in restruc- turing a web site to improve access to the preferred contents; the information can also be used to detect market segments based on the buying behavior of the users and to implement resource suggestion systems for the users. The sequence of individual inter- actions is called a session and the reconstruction of all of the page http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2013.08.056&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.08.056 mailto:tomas.arcec@usach.cl mailto:proman@dii.uchile.cl mailto:jvelasqu@dii.uchile.cl mailto:victor.parada@usach.cl http://dx.doi.org/10.1016/j.eswa.2013.08.056 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 1594 T. Arce et al. / Expert Systems with Applications 41 (2014) 1593–1600 sequences visited by the users during their browsing through a web site is known as the web sessionization problem (WSP) (Berendt, Mobasher, Spiliopoulou, & Wiltshire, 2001; Chitraa & Selvdoos, 2010; Huynh & Miller, 2009; Poženel, Mahnic, & Kukar, 2010; Velásquez & Palade, 2008). A correct reconstruction of the sessions must take into account that the validity of the generated patterns depends largely on the credibility of the sessions obtained (Berendt et al., 2001). In spite of that, the reconstruction of sessions is not a trivial process because of factors that hinder reconstruc- tion, such as proxy servers or the use of cache memory in the client’s web browser. The techniques that solve the WSP from a web log must deal mainly with the nonexistence of a clear identification of the users. Matching the IP address as a single criterion is not sufficient, because when multiple users have access to a site through a proxy server, they are registered in the web log with the proxy server’s address. To partially remove the identification error, the tech- niques use the IP address and the web browser that was used by the web site visitor as identification criteria (Pirolli, Pitkow, & Rao, 1996). Other techniques directly track user operations by using cookies with client-side scripts obtaining accurate sessions, but they are not recommended since they violate user privacy laws (Mayer & Mitchell, 2012). Therefore, the web usage mining on web logs is a safe way to analyze user preferences, as long as we take into account that in general a high quality sessionization is not possible by means of machine learning techniques that are known to be subject to data error (Román et al., 2010). Current improvement of the web usage mining processing relies on accurately solving the WSP, for which several heuristics have been proposed. The most widely used technique to tackle the WSP is the time heuristic, which considers that the sessions have a time limit (Catledge & Pitkow, 1995; Cooley, Mobasher, & Srivastava, 1997; Huynh & Miller, 2009). This technique groups the records according to the IP address and the web browser that was employed by the user; it arranges the records of each resultant group in a temporal order and then obtains the sessions, ensuring that the first and last records of a given session do not exceed a time limit. The major drawback of this algorithm consists of the null consideration of the hyperlink topology of the web site. A revised version of this heuristic considers the site’s link structure together with a temporal criterion (Pirolli et al., 1996), in addition to the criteria that the consecutive records of the same session must refer to pages between which there is a link. Nowadays, with modern web browser navigation it becomes hard to track the infor- mation due to the multiple ways that pages are cached, loaded, and navigated. Multitab navigation, back button browsing and history jumps are commonly not reflected on web logs (Román et al., 2010), so hyperlink topology restriction is relaxed from being a strong restriction on session properties. A recent approach addresses the WSP as an optimization prob- lem by defining an integer programming model (Román, Dell, & Velásquez, 2010). All of the possible reconstructions of sessions from a given web log constitute the feasible solution space of the problem. The choice of a specific reconstruction implies a search for the reconstruction that has a maximum value in terms of a spe- cific objective function. Web hyperlink topology and time sequenc- ing are incorporated as restrictions in the model. Relaxed browsing behavior could even be easily included (Dell, Román, & Velásquez, 2009). However, because a web log in a single day can accumulate an enormous number of records and the supporting decision mak- ing requires the reconstruction of sessions over a period of weeks or months, the number of variables of the integer programming models is huge for the currently available capacity for solving inte- ger programming problems. Solving even small instances of the problem requires several hours of computing time. The perfor- mance is even worse when considering more common browsing behaviors such as parallel tabs and back buttons. The ideal situa- tion would consider the existence of an algorithm that can process the information in real time by requiring a short computing time. A novel algorithm for solving the WSP was presented in Bayir, Toroloslu, Demirbas, and Cosar (2012). The algorithm is based on graph modeling of the sessions that are constructed considering maximal path length, hyperlink topology and back button browsing. They found improved accuracy in recovering sessions relative to previous models. However, their major drawback comes from the theoretical justification of the model affecting its reproducibility. Instead of generating the optimum solution by means of an algorithm that solves the integer programming problem, we obtain a good quality solution using a small amount of computing time, by means of Simulated Annealing (SA) which is a metaheuristic that emulates the physicochemical process that takes place in the cooling of pure substances, systematically generating a new solution from the current solution and allowing, at some instant, the choice of a poor solution, with a certain probability that de- creases with time (Kirkpatrick, Gelatt, & Vecchi, 1983; Talbi, 2009). This method has been shown to be very efficient for solving problems belonging to the NP-Hard class, such as the design of electronic circuits, the reconstruction of images, the generation of roads, set partition problems and planning problems (Suman & Kumar, 2005). This paper presents an approach that is based on simulated annealing for the WSP for the purpose of reconstructing each session of the users that visit a web site. The second section presents the WSP representation under an SA, identifying the elements needed for the experimental phase. The third section presents the main results, and in the last section, the main conclusions are given. 2. Methods and materials 2.1. The web sessionization problem The model founded on integer programming (Dell, Roman, & Velasquez, 2008) is based on ensuring that a session is constituted by a set of records that share the same IP address and the same web browser. For consecutive records within a session, the con- straints ensure the maintenance of the link structure of the site and that the session does not go beyond a certain time limit. Thus, if a record is found directly after another record within the same session, then the following are true: j Both share the same IP address and web browser. j There is a link between the referenced page in both records. j The time difference between the recorded application for both records does not exceed a certain mtp value or time window. Let j r and r’ be records of a web log; j obe the order of a record in a given session, o = 1, 2, . . ., O, where is the maximum size that a session can have; j s be the identifier of a session; j Co be the value of the coefficient of the objective function when there is a record assigned to position o; j xros be the binary decision variable, which has a value of 1 when record r is assigned to position o in a given session s and a value of 0 in any other case; j Bpage(r) be the set of records that can be directly before record r in the same session, according to the criteria of hav- ing the same IP address, corresponding to the same web browser, having a link between consecutive records and maintaining the time window between consecutive records; SA algorithm Input: Min f(x) s.t. x ∈ Ω; Output: x*; Generate an initial solution x0 ∈ Ω; x* = x0; Define T > 0; t = 0; Repeat T. Arce et al. / Expert Systems with Applications 41 (2014) 1593–1600 1595 j first be the set of records that are always found in the same position in a session. It is possible to calculate this set by first getting bpager for all of the web log records. Therefore, a record r 2 first if bpage(r) = £. The integer programming model is the following: ISP : Maximize z ¼ X ros Co xros ð1aÞ s.t. X os xros ¼ 1; 8r ð1bÞ X r xros � 1; 8o; s ð1cÞ xr;oþ1;s � X r02bpageðrÞ xr0os; 8r; o; s ð1dÞ xros ¼ 0; 1 8r; o; s ð1eÞ xros ¼ 0 8r;2 first; o > 1; s ð1fÞ The first constraint establishes that a record can only be in a single session and in a given order within it. The second constraint estab- lishes that there can only be a unique record in a given session and a position within it. The third constraint establishes the correct arrangement of the records of the same session, complying with the time, site structure and the same IP address and web browser combination conditions. The value of Co is used to give a bonus to sessions of a given size. For example, by setting a value of Co = 1 when o = 2 and Co = 0 " o – 2, we are favoring a large number of size 2 sessions in the final reconstruction. The value of the objective function for a sessionization is calculated taking into account the length of each of the generated sessions (l) and then adding the va- lue of Co that corresponds to each position of the session. For exam- ple, for a single session of length l = 5, the value of the objective function is the following: X Co ¼ C1 þ C2 þ C3 þ C4 þ C5 ð2Þ Several measures can be used to reproduce the sessions. Functions C1o through C 4 o are used in this paper and are presented in Table 1. All of the functions are monotonically increasing in o with the pur- pose of maximizing small-size sessions or else maximizing large- size sessions. 2.2. Quality of the reconstruction The quality of the reconstruction is measured with respect to the power law for the size of the sessions on a web site (Huberman, Pirolli, Pitkow, & Lukose, 1998; Oliveira et al., 2006). This law states that most visits to a web site are concentrated on a small number of pages and the rest of the pages receive a smaller number of vis- its. Thus, the quality of the reconstruction can be verified by revis- ing the coefficient of correlation r2 and the standard error S, obtained from a linear regression performed on the logarithm of Table 1 Different coefficients Co for the objective function. Co C1o = log(o) C2o = 1.5log(o) + (o � 3) 2/12o C3o = o C4o = o 2 the size of the session and the number of sessions generated by the reconstruction. 2.3. Simulated annealing (SA) The SA algorithm is a metaheuristic optimization procedure that is inspired by the cooling process of molten metals, and it has been used to find good quality solutions to various combinato- rial optimization problems (Kirkpatrick et al., 1983; Talbi, 2009; Černý, 1985). In the metallurgical process, the initial state is the molten metal, and the temperature is decreased in a controlled manner, allowing the formation of intermediate equilibrium states until a solid crystalline structure is achieved. During the process, an internal form of the metal’s energy is decreased. Following this analogy, to solve an optimization problem, the solution space is visited, varying a parameter T that corresponds to the temperature so that, at a high value of T, the probability of accepting solutions worse than the current solution is high. As T decreases, the algo- rithm is gradually converted into a local search algorithm. Regula- tion of the acceptance probability takes place by means of a Boltzmann distribution. The adequate representation of an optimi- zation problem that is solved by SA requires defining a rule to gen- erate neighboring solutions of the current solution, the function that must be optimized and an initial solution. The proper algo- rithm parameters must also be defined: the initial value of T, a way of decreasing it gradually and the number of iterations that are performed over the internal cycle N(t). The procedure for solv- ing a maximization problem is presented in the pseudocode de- picted in Fig. 1. 2.4. Representation of a solution with SA A solution of a WSP is represented using a matrix X(mxn), where m is equal to the maximum size of a session and n is equal to the number of records contained in the web log partition. Then, each element xij of the matrix corresponds to the identification number of the record that is in position i of session j. Thus, for a given solu- tion, the column index is the identifier of the reconstructed session. 2.5. Generation of an initial solution The initial solution generated uses a variation of the Links Heu- ristic (Pirolli et al., 1996), ensuring that all of the consecutive re- cords within a session do not exceed a maximum time window. For two records, r1 and r2, to be located consecutively in the same session, the following must be satisfied: Repeat Generate solution xj neighboring xi; δ = ƒ(xj) - ƒ(xi); If δ > 0 then xi = xj; Otherwise If random (0, 1) < exp (-δ / T) then xi = xj; If xi < x* then x* = x0; Until N(t) iterations are completed t = t +1; T = T(t); Until the stop criterion is reached. Fig. 1. SA for a maximization problem. Function GeneratesInitialSolution () MTP : Maximum time difference between consecutive records. SMAX : Maximum length of a session. newSession=1; session=1; For each record r in the web log If (newSession = 1)then Create a new session with record r in the first position. Or else If (The size of the session is Not equal to SMAX) AND (The last record of the session, r - 1, has the same address as record r) AND (There is a link between r - 1 and r) AND (r.timeStamp – (r – 1).timeStamp ≤ MTP) then Add r to the current session. Or else newSession =1; r --; End If End If End For End Function Fig. 2. Algorithm for generating the initial solution. 1596 T. Arce et al. / Expert Systems with Applications 41 (2014) 1593–1600 j Both records r1 and r2 have the same IP address. j Record r1 has a link with r2. j The records have visiting times within a predefined time window mtp. j A predefined maximum session time is not exceeded. Fig. 2 presents the algorithm for generating the initial solution. In the algorithm, the records are assigned to a session following the viability conditions that are described above. 2.6. Generation of the neighboring solution The generation of a neighboring solution consists of three steps: choosing a session randomly, deleting the chosen session and reas- signing each of the records of the deleted session. In the last step, there are two possibilities; the record is either assigned to other existing sessions that comply with the maximum number of re- cords per session, the maximum time difference and the existence of a link between consecutive records, or it is assigned to a new session where the record occupies the first position. A record can be relocated as follows: j At the beginning of the session. The chosen record re must have a link to the first record r1, and the time difference between re and r1 must not exceed the maximum time win- dow mtp. j In the middle of the session. If ri and ri+1 are two consecutive records within the session, there must be a link that goes from ri to re and another one that goes from re to ri+1 for the chosen record re to be located between ri and ri+1; the time difference between ri, re and between re, ri+1 must not be greater than the maximum time window mtp. j At the end of the session. The existence of a link between the chosen record re and the last record rf must be verified, and the time difference between rf and re must not exceed the time window mtp. 2.7. Evaluation function The evaluation function considered is the proper objective func- tion (1a) of the integer programming problem, keeping in mind the four possibilities for Co that are presented in Table 1. 2.8. Selection of SA parameters With the aim of searching for the adequate adaptation of SA for the WSP, we set the parameters according to the following criteria for each case: j T0. A scheme has been adopted with a variable T0 that allows changing the execution time of the algorithm depending on the effort needed for the search, measured according to the size of the instance to be solved. In this way, T0 is chosen based on the maximum approximate difference in the objective function (Aarts & Lenstra, 1997), as follows: DFmax ¼ fmax � fmin ð3Þ In the problem’s context, the greatest decrease of Co occurs when the largest session is fragmented into smaller unit sessions. The dif- ference gives Eq. (4) as a result: DFMax ¼ XL o¼1 Co � LC1 ð4Þ where L is equal to the largest session of the initial solution generated. Considering the natural logarithm of the acceptance probability p, ln(p) = �DFmax/T0, with p = 0.99, we get T0 such that when a solu- tion gets worse by a magnitude of DFmax, it will be accepted with a probability of 0.99. Then, T0 is calculated from T0 = �DFmax/0.01. j Cooling function. A geometric temperature drop is used, given by Ti = �a Ti � 1, with a 2 (0, 1). j The number of iterations of the internal cycle. The number of iterations is dynamic as the search continues. If a neigh- boring solution that improves the current solution is gener- ated, the cooling function is applied to decrease T (Cardoso, Salcedo, & Azevedo, 1994; Ravi & Shukla, 2009). To avoid stagnation of the SA at a single T level, the maximum num- ber of iterations equal to the number of possible neighbor- ing solutions of a given solution is considered. An approximation of this number is given by the number of sessions at each level of T; when a neighboring solution is generated, the number of possible movements is equivalent to the number of sessions that can be chosen for their records to be reassigned. j T _f A fixed value is used that is found during the calibration process. 2.9. Algorithm for WSP The internal cycle operates with a constant value of T visits; each time, a neighbor solution of the current solution is generated by means of a reassignment of a record from one session to another. It accepts this solution automatically when a better sessionization is generated and accepts it with a specific probabil- ity when the sessionization is worse than the current one. 2.10. Equipment used The implementation of SA was written in the C language and was compiled with GCC version 4.3.2. The computer used for both the implementation and the execution of the experiment had an Intel Core 2 Duo 1.8-Ghz processor, with 1 GB of RAM, an 80-GB hard disc, and the Linux Operating System, kernel version 2.6.26, Debian version 5.0 distribution. T. Arce et al. / Expert Systems with Applications 41 (2014) 1593–1600 1597 2.11. Test instances To verify the operation of the implementation of the proposed solution, a web log generated by a web site that considered a storage period of one month was used. The site has 172 pages and 1228 links between them. The web log was previously preprocessed to delete the requirements of multimedia objects, access to web mail and error records. After the preprocessing, the total number of records is 102,303. Of the 16,985 IP addresses con- tained in the file, 16,785 have less than 50 records that represent 98% of the total. In turn, 14,265 IP addresses visit 3 or fewer differ- ent pages, which represent 84% of the total IP addresses. Reconstructing sessions using the mathematical model and starting from a web log leads to an oversized requirement of the hardware resources. However, because a session is composed of records that share the same IP address, it is possible to partition the web log into smaller chunks. In this way, the possible solutions space is reduced, and the problem’s natural restrictions are applied to divide it into multiple smaller size problems. By applying the algorithm to each chunk and then integrating the partial results, the resource requirement is reduced considerably. Not all of the partitions have the same interest; there are sets of records that are more interesting for the sessionization, given the possibility that the recorded requests correspond to multiple user accesses that are masked under a single address (which corresponds to the proxy server or the firewall that is being used). To identify these sets, two measures of difficulty are used that are established in relation to a given IP address, the diversity of the record and the number of records. The record diversity of a given IP address is measured in terms of the entropy of the IP address (Dell et al., 2008), defined as the following: X i2PðIPÞ ðfi=nRÞlogbðnR=fiÞ ð5Þ where fi corresponds to the number of times in which page i was accessed by a given IP address, nR is the number of records contain- ing that IP address, b is the number of different pages that have been visited from a given IP address and P(IP) is the number of pages visited from the IP address. The entropy takes values between 0 and 1. When the entropy is close to 0, most of the records for a given address correspond to accesses to a single page; when it is close to 1, all of the pages accessed by the address are visited with the same frequency. A large diversity for an IP address is a good indicator of the dif- ficulty of the records associated with that address, but it is not a sufficient condition, because there can be a large diversity with very few records. That is why, together with entropy, a criterion of a minimum number of records per IP address is used; an IP address with high entropy and a large number of records ensures that the records associated with that address are referred to differ- ent pages and that they are numerous, capturing what happens in a web log when multiple users gain access to the site by means of proxy servers. To construct the test instances, the records with IP addresses that have entropy greater than or equal to 0.5 and a number of records in the web log greater than or equal to 50 are chosen. For the web log used, the number of records chosen under this criterion was 17,709. 2.12. Metrics used To evaluate and analyze the results obtained, the following metrics are used: j Percent difference of the objective function, Gap: Gap ¼ jfðxÞ� f �ðxÞj=f �ðxÞ ð6Þ where f⁄(x) is the reference value for the comparison. j Power Law fit. The quality of the reconstruction is verified by means of the coefficients r2 and S, obtained from a linear regression over the logarithm of the size of a session and the number of sessions. The values 0 6 r2 6 1 and S allow the evaluation of the degree of fit of the number of sessions ver- sus the size of the sessions with the Power Law. When r2 takes values close to 1, the model fits perfectly with the Power Law. On the other hand, S measures the standard error of the model in terms of the difference between the real value and the estimated value. The closer it is to 0, the better the model’s fit is with real data (Montgomery & Runger, 2006). 2.13. Execution of the experiment To reconstruct the sessions with SA, we used the values of a = 0.99 and Tf = 0.08 for the parameters. For each set of instances, an execution of the algorithm considers 10 attempts for each way of assigning coefficient Co of the objective function. An attempt must be understood as the execution of the algorithm over all of the chunks of the web log. The assignment functions of Co presented in Table 1 are consid- ered for the experiment. Subsequently there are 40 attempts for the set of ISP instances and 40 attempts for the set of instances of the metaheuristic, with a total of 80 attempts. Counting the application of the solution to each chunk, a total of 21,200 execu- tions of the SA are performed. 3. Results The execution of SA obtains the following for each assignment function of Co and each set of test instances: j Average attempt: The reconstruction closest to the average value of the objective function calculated from the 10 attempts executed. j Best attempt: A reconstruction based on the best attempt of the 10 executed, in terms of the value of the objective function. j Best reconstruction: A reconstruction that is obtained by choosing, for each chunk, the best sessionization in terms of the value of the objective function. The solutions of each of the best chunks are joined to form the best possible solu- tion from the attempts that are executed up to some point. This reconstruction is obtained from the processing of the results, after the execution of the implementation. Furthermore, SA provides information on the value of the objec- tive function that obtained the solution, the computer time and the quality of the reconstruction of the sessions in terms of the Power Law in relation to the size of the sessions. 3.1. General results The general results that are obtained during the execution of the algorithm are presented in Table 2. The smallest average num- ber of sessions is obtained with C1o . The largest session average is obtained with C4o . The function C 1 o yields the shortest processing times, with an average of 2.15 min per attempt and a total test time for the execution of the 10 attempts of 21.52 min. Function C4o takes an average of 3.86 min per attempt, with a total test time of 38.58 min. Table 3 presents the best attempts for each of the objective functions. The lowest number of sessions was 11,312, obtained Table 2 Results of SA for WSP. Co Maximum # of sessions Minimum # of sessions Average # of sessions Standard deviation zmax zmin zpromedio Average computer time [min] Total test time [min] Co1 11321 11312 11317.7 2.83 6481.3 6460.4 6469.9 2.15 21.52 Co2 11334 11321 11326.3 4.42 13990.0 13918.5 13957.2 2.41 24.14 Co3 11349 11320 11332.5 7.76 32561.0 32391.0 32489.8 2.47 24.74 Co4 11352 11335 11340.5 4.81 134926.0 131012.0 132746.8 3.86 38.58 Table 3 Results of the best attempts for each of the objective functions. Co Id attempt z r 2 S No. of sessions Computer time [min] C1o 10 6481.3 0.943 0.577 11312 2.15 C2o 1 13990.0 0.931 0.666 11327 2.41 C3o 7 32561.0 0.919 0.707 11326 2.48 C4o 6 134926.0 0.941 0.581 11338 3.86 Table 4 Evolution of the value of the objective function for different executions of SA. Co1 Co2 Co3 Co4 Attempt z Dz Gap [%] z Dz Gap [%] z Dz Gap [%] z Dz Gap [%] 1 6460.4 – – 13990.0 – – 32541.0 – – 134012.0 – – 2 6490.7 30.29 0.467 14032.4 42.44 0.302 32688.0 147.0 0.450 136611.0 2599.0 1.90 3 6505.0 14.23 0.219 14041.4 9.03 0.064 32744.0 56.0 0.171 137883.0 1272.0 0.93 4 6511.2 6.25 0.096 14052.4 10.93 0.078 32770.0 26.0 0.079 138246.0 363.0 0.26 5 6514.3 3.12 0.048 14063.7 11.33 0.081 32794.0 24.0 0.073 138811.0 565.0 0.41 6 6519.6 5.30 0.081 14068.6 4.92 0.035 32839.0 45.0 0.137 139885.0 1074.0 0.77 7 6521.4 1.77 0.027 14072.1 3.50 0.025 32847.0 8.0 0.024 140056.0 171.0 0.12 8 6524.0 2.57 0.039 14080.2 8.02 0.057 32893.0 46.0 0.140 140056.0 0.0 0.00 9 6525.1 1.10 0.017 14080.2 0.00 0.000 32898.0 5.0 0.015 140126.0 70.0 0.05 10 6528.0 2.97 0.046 14080.3 0.18 0.001 32898.0 0.0 0.000 140126.0 0.0 0.00 Table 5 Results obtained with SA and CPLEX for WSP. Co z (ISP) z (SA) av. attempt z (SA) best value Gap [%] C1o 5953.3 6470.4 6528.0 9.66 C2o 13229.4 13956.5 14080.3 6.43 C3o 30783.0 32482.0 32898.0 6.87 C4o 119450.0 132485.0 140126.0 17.31 1598 T. Arce et al. / Expert Systems with Applications 41 (2014) 1593–1600 with C1o , and the largest number of sessions, 11,338, was obtained with C4o . The shortest and longest times occur with C 1 o and C 4 o , with 2.15 and 3.86 min, respectively. The best fit with the Power Law occurs with C1o . 3.2. Evolution of the best reconstruction Because the solution is based on a random search method, it is possible that a new execution over a specific chunk will find a bet- ter solution than one found in a previous execution. Executing SA several times over the same web log, it is possible to choose the best solution for each of the chunks. Because the sessionization of each chunk is independent of the others, the best reconstruction can be generated based on a certain number of attempts. This way of operating allows refining the quality of a sessionization by means of successive executions. Table 4 shows the value of the objective function for the best possible reconstruction up to a given attempt. The value of z is presented for every function C1o through C4o , as is the difference in z that is produced by the new execution of SA with respect to the previous execution. It can be seen that the values of the objective function tend to converge at a given point, causing the largest increase in the first four attempts, with more than 70%. Specifically, the increases remain until attempt number 10, except for C3o and C 4 o , which con- verge after the eighth attempt. The best refining is produced with C4o , which shows an increase of 4.56%. If a unique reconstruction is sought that is stable in terms of the value of the objective function, it is seen that the road to follow is to execute SA more than once with C4o . Alternatively, if efficiency is sought while sacrificing the stability of the sessionization obtained, it is recommended to execute SA only once. 3.3. Comparison with the exact method The integer programming model was solved with CPLEX («IBM – IBM – Mathematical Programming», 2011) version 10.1.0, together with GAMS («GAMS», s.f.), a software system used for defining optimization models. The experiment was executed on a computer with 1.6 Ghz and 2 GB RAM. CPLEX was executed with a time limit of 300 s for each chunk. When this limit was reached, the software returned the best solution found up to that moment for this chunk. Table 5 shows the value of the objective function for each of the assignment functions. The results show that SA improves the opti- mum found by CPLEX for all of the assignment functions, with an increase with respect to the optimum value that goes from 9.66% to 17.31%, respectively. Table 6 shows the processing time of CPLEX, tISP, and the aver- age execution time for each SA attempt, tSA. SA has a significantly Table 6 Computer time for ISP and SA. Co tISP [min] tSA [min] Speedup C1o 359.1 2.15 166.78 C2o 372.5 2.40 154.90 C3o 379.9 2.47 153.51 C4o 96.3 3.86 24.96 Table 7 Quality of the reconstruction of sessions based on Power Law fit. Method r2 S Time heuristic 0.915 0.663 ISP-CPLEX model (C2o ) 0.978 0.383 SA (Average attempt, C4o ) 0.943 0.587 SA (Best attempt, C1o ) 0.943 0.577 SA (Best reconstruction, C4o ) 0.943 0.573 T. Arce et al. / Expert Systems with Applications 41 (2014) 1593–1600 1599 shorter execution time than the integer programming model in all of the assignment functions of Co. The speedup reaches 166.78 for C2o , which is the maximum value of this indicator. The smallest speedup is obtained with C4o , with a 24.96-fold reduction in pro- cessing time. To study the quality of the reconstruction based on the fit with the Power Law, we obtain the values of r2 and S for the time heuris- tic. In this case, the records are arranged temporarily in increasing order for each group. Each record is chosen and added to a session in such a way that the consecutive records of a given session do not have a time difference greater than 300 s. Table 7 shows the best- fitting values obtained by each of the methods studied. Based on the above values, the best fit is obtained with the ISP model solved with CPLEX using C2o . Both of the reconstructions that are obtained with ISP and those obtained by SA exceed the time heuristic in terms of that metric, which presents the worst values in both the standard error and in r2. 4. Discussion and conclusions The implementation of Simulated Annealing to solve the WSP improves the value of the objective function for each of the assign- ment functions, with a gap that varies between 9.66% and 17.31%. The speed for obtaining the results is one of the main strengths of SA over the exact method. In the best case, the reduction is of the order of 166 times the execution time of the exact method. Executing SA several times on the same set of instances allows new solutions to be obtained that are better than those that can be obtained with a single execution. The study of the magnitude of the increments in relation to the number of executions of the solu- tion yielded an interesting convergence of the solutions to a given value of the objective function. That allows the achievement of a potentially stable and unique reconstruction depending on the number of executions, compared with only a single execution of the solution. Clearly, this implies an increase in processing time that must be evaluated according to the accuracy required for the reconstruction of the sessions. From the standpoint of the sessionization problem, the fit of the reconstructions obtained by SA in terms of the Power Law for the size of the sessions is better using CPLEX but comes at a high com- putational cost. Acknowledgements This research was partially supported by the Millennium Institute Complex Engineering Systems ICM: P-05-004-F, FBO16. References Aarts, E., & Lenstra, J. K. (Eds.). (1997). Local search in combinatorial optimization. John Wiley & Sons, Inc. Bayir, M. A., Toroloslu, I. H., Demirbas, M., & Cosar, A. (2012). Discovering better navigation sequences for the session construction problem. Data & Knowledge Engineering, 73, 58–72. Berendt, B., Mobasher, B., Spiliopoulou, M., & Wiltshire, J. (2001). Measuring the accuracy of sessionizers for web usage analysis. In Proceedings of the workshop on web mining at the first SIAM international conference on data mining, 7–14, Chicago, IL. Cardoso, M., Salcedo, R., & Azevedo, S. (1994). Nonequilibrium simulated annealing: A faster approach to combinatorial minimization. Industrial Engineering Chemical Research, 33, 1908–1918. Catledge, L. D., & Pitkow, J. E. (1995). Characterizing browsing strategies in the world-wide web. Computer Networks and ISDN Systems, 27(6), 1065–1073. Černý, V. (1985). Thermodynamical approach to the traveling salesman problem: An efficient simulation algorithm. Journal of Optimization Theory and Applications, 45(1), 41–51. Chitraa, V., & Selvdoos, A. (2010). A survey on preprocessing methods for web usage data. IJCSIS International Journal of Computer Science and Information Security, 7(3), 78–83. Choi, J., & Lee, G. (2009). New techniques for data preprocessing based on usage logs for efficient web user profiling at client side. In Proceedings of the 2009 IEEE/WIC/ ACM international joint conference on web intelligence and intelligent agent technology (Vol. 3, pp. 54–57). Washington, D.C. Cooley, R., Mobasher, B., & Srivastava, J. (1997). Web mining: information and pattern discovery on the world wide web. In Proceedings of the 9th international conference on tools with artificial intelligence (pp. 558–567). Newport Beach, California. Cooley, R., Mobasher, B., & Srivastava, J. (1999). Data preparation for mining world wide web browsing patterns. Knowledge and Information Systems, 1(1), 5–32. Dell, R., Roman, P., & Velasquez, J. (2008). Web user session reconstruction using integer programming. In IEEE/WIC/ACM international conference on web intelligence and intelligent agent technology (Vol. 1, pp. 385–388). Sydney. doi:http://dx.doi.org/10.1109/WIIAT.2008.181. Dell, R., Román, P., & Velásquez, J. (2009). Web User Session Reconstruction with back button browsing. In Knowledge-based and intelligent information and engineering systems, lecture notes in computer science (Vol. 5711, pp. 326–332). doi:http://dx.doi.org/10.1007/978-3-642-04595-0_40. Devi, B. N., Devi, Y. R., Rani, B. P., & Rao, R. R. (2012). Design and implementation of web usage mining intelligent system in the field of e-commerce. Procedia Engineering, 30, 20–27. Fong, A. C. M., Zhou, B., Hui, S., Tang, T., & Hong, G. (2012). Generation of personalized ontology based on consumer emotion and behavior analysis. IEEE Transactions on Affective Computing, 3(2), 152–164. GAMS. GAMS Home Page. Retrieved from http://www.gams.com/. Huberman Pirolli Pitkow & Lukose (1998). Strong regularities in world wide web surfing. Science, 280(5360), 95–97. Huynh, T., & Miller, J. (2009). Empirical observations on the session timeout threshold. Information Processing & Management, 45(5), 513–528. IBM – IBM – Mathematical programming: linear programming, mixed-integer programming and quadratic programming – IBM ILOG CPLEX optimizer – software. Retrieved from: http://www-01.ibm.com/software/integration/ optimization/cplex-optimizer/. Kirkpatrick, S., Gelatt, D., & Vecchi, M. P. (1983). Optimization by simulated annealing. Science, 220(4598), 671–680. Kosala, R., & Blockeel, H. (2000). Web mining research: A survey. SIGKDD Explorations, 2, 1–15. Lin, S., Jheng, Y., & Yu, C. (2012). Combining ranking concept and social network analysis to detect collusive groups in online auctions. Expert Systems with Applications, 39(10), 9079–9086. Mayer, J. R. & Mitchell, J. C. (2012). Third-party web tracking: Policy and technology, security and privacy. In IEEE symposium on security and privacy (pp. 413–427). San Francisco. Montgomery, D. C., & Runger, G. C. (2006). Applied statistics and probability for engineers (4th ed.). Wiley. Nasraoui, O., Soliman, M., Saka, E., Badia, A., & Germain, R. (2008). A web usage mining framework for mining evolving user profiles in dynamic web sites. IEEE Transactions on Knowledge and Data Engineering, 20(2), 202–215. Nohuddin, P. N. E., Coenen, F., Christley, R., Setzkorn, C., Patel, Y., & Williams, S. (2012). Finding ‘‘interesting’’ trends in social networks using frequent pattern mining and self organizing maps. Knowledge-Based Systems, 29, 104–113. Oliveira, J. G., Dezso, Z., Goh, K.-I., Kondor, I., Barabasi, A.-L., & Vazquez (2006). Modeling bursts and heavy tails in human dynamics. Physical Review E – Statistical, Nonlinear and Soft Matter Physics, 73(3 Pt 2), 036127. Pirolli, P., Pitkow, J., & Rao, R. (1996). Silk from a sow’s ear: Extracting usable structures from the web. In Proceedings of the SIGCHI conference on human factors in computing systems: common ground (pp. 118–125). Vancouver. Poženel, M., Mahnic, V., & Kukar, M. (2010). Separation of interleaved web sessions with heuristic search. In Proceedings of the ICDM ’10, IEEE international conference on data mining (pp. 411–420). Sidney. Ravi, S., & Shukla, S. (Eds.). (2009). Fundamental problems in computing: An analysis of several heuristics for the traveling salesman problem. Dordrecht, Netherlands: Springer Netherlands. http://refhub.elsevier.com/S0957-4174(13)00677-5/h0005 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0005 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0010 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0010 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0010 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0015 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0015 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0015 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0020 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0020 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0025 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0025 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0025 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0030 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0030 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0030 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0035 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0035 http://dx.doi.org/10.1109/WIIAT.2008.181 http://dx.doi.org/10.1007/978-3-642-04595-0_40 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0040 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0040 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0040 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0045 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0045 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0045 http://www.gams.com/ http://refhub.elsevier.com/S0957-4174(13)00677-5/h0050 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0050 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0055 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0055 http://www-01.ibm.com/software/integration/optimization/cplex-optimizer/ http://www-01.ibm.com/software/integration/optimization/cplex-optimizer/ http://refhub.elsevier.com/S0957-4174(13)00677-5/h0060 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0060 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0065 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0065 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0070 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0070 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0070 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0145 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0145 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0075 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0075 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0075 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0080 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0085 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0085 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0085 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0090 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0090 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0090 1600 T. Arce et al. / Expert Systems with Applications 41 (2014) 1593–1600 Román, P. E., Dell, R. F., & Velásquez, J. D. (2010). Advanced technique in web data pre-processing and cleaning. In J. D. Velásquez & L. Jain (Eds.). Advances in techniques in web intelligence-1 (Vol. 311, pp. 19–48). Heidelberg: Springer Verlag. Román, P. E., L’Huillier, G., & Velásquez, J. D. (2010). Web usage mining. In J. D. Velásquez & L. Jain (Eds.). Advances in techniques in web intelligence-1 (Vol. 311, pp. 143–165). Heidelberg: Springer Verlag. Suman, B., & Kumar, P. (2005). A survey of simulated annealing as a tool for single and multiobjective optimization. Journal of the Operational Research Society, 57(10), 1143–1160. Talbi, E. G. (2009). Metaheuristics: From design to implementation. John Wiley and Sons. Tao, Y. H., Hong, T. P., Lin, W. Y., & Chiu, W. Y. (2009). A practical extension of web usage mining with intentional browsing data toward usage. Expert Systems with Applications, 36(2), 3937–3945. Velásquez, J. D. (2013). Web mining and privacy concerns: some important legal issues to be consider before applying any data and information extraction technique in web-based environments. Expert Systems with Applications, 40(13), 5228–5239. Velásquez, J. D., & Palade, V. (2008). Adaptive web sites: A knowledge extraction from web data approach. Amsterdam: IOS Press. Wang, T., & Ren, Y. (2009). Research on personalized recommendation based on web usage mining using collaborative filtering technique. WSEAS Transactions on Information Science and Applications, 6(1), 62–72. Wang, K., Ting, I., & Wu, H. (2013). Discovering interest groups for marketing in virtual communities: An integrated approach. Journal of Business Research, 66(9), 1360–1366. Yin, P., & Guo, Y. (2013). Optimization of multi-criteria website structure based on enhanced tabu search and web usage mining. Applied Mathematics and Computation, 219(24), 11082–11095. http://refhub.elsevier.com/S0957-4174(13)00677-5/h0100 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0100 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0100 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0100 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0095 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0095 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0095 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0105 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0105 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0105 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0110 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0110 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0115 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0115 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0115 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0120 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0120 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0120 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0120 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0125 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0125 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0140 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0140 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0140 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0135 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0135 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0135 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0130 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0130 http://refhub.elsevier.com/S0957-4174(13)00677-5/h0130 Identifying web sessions with simulated annealing 1 Introduction 2 Methods and materials 2.1 The web sessionization problem 2.2 Quality of the reconstruction 2.3 Simulated annealing (SA) 2.4 Representation of a solution with SA 2.5 Generation of an initial solution 2.6 Generation of the neighboring solution 2.7 Evaluation function 2.8 Selection of SA parameters 2.9 Algorithm for WSP 2.10 Equipment used 2.11 Test instances 2.12 Metrics used 2.13 Execution of the experiment 3 Results 3.1 General results 3.2 Evolution of the best reconstruction 3.3 Comparison with the exact method 4 Discussion and conclusions Acknowledgements References 
work_qplkyxj7angddifjawucgv7umu	----	Development of knowledge Base Expert System for Natural treatment of Diabetes disease (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3, No. 3, 2012 44 | P a g e www.ijacsa.thesai.org Development of knowledge Base Expert System for Natural treatment of Diabetes disease Sanjeev Kumar Jha University Department of Mathematics, Babasaheb Bhimrao Ambedkar Bihar University, Muzaffarpur - 842001 Bihar (India) D.K.Singh University Department of Mathematics, Babasaheb Bhimrao Ambedkar Bihar University, Muzaffarpur - 842001 Bihar (India) Abstract—The development of expert system for treatment of Diabetes disease by using natural methods is new information technology derived from Artificial Intelligent research using ESTA (Expert System Text Animation) System. The proposed expert system contains knowledge about various methods of natural treatment methods (Massage, Herbal/Proper Nutrition, Acupuncture, Gems) for Diabetes diseases of Human Beings. The system is developed in the ESTA (Expert System shell for Text Animation) which is Visual Prolog 7.3 Application. The knowledge for the said system will be acquired from domain experts, texts and other related sources. Keywords- Expert System; ESTA; Natural treatment; Diabetes. I. INTRODUCTION This article presents the conceptual framework of natural treatment methods available for diabetes. The main goal of this research is to integrate all the natural treatment information of diabetes in one place. Expert System named as Sanjeevani is developed using ESTA (Expert System Shell for Text Animation) as knowledge based system to describe the various Natural therapy methods for treatment of Diabetes disease and various other diseases. The main purpose of the present study is in the design and development of an expert system which provides the information of different types of natural treatment (Massage, Acupuncture, Herbal/Proper Nutrition and gems) of Diabetes. The system background starts with the collection of information of different methods of treatment available for Diabetes diseases. The acquired knowledge is represented to develop expert System. II. DEVELOPMENT OF EXPERT SYSTEM We are in the process of development of Sanjeevani Natural therapy expert system that is developed in ESTA Application. It is designed for assisting in treatment of the Diabetes disease and various other diseases which can be cure naturally with different methods (Massage, Herbal/Proper Nutrition, Acupuncture, Gems) and provide treatment solutions. The natural therapy consists of a variety of natural body therapies, soul therapies and energy therapies for healing, that we found helpful for peoples health and wellness and which don't cost the earth The natural treatment is the process of healing/curing diseases through natural, drugless and the most harmless process. Human beings are an intrinsic part of the nature. Therefore, nothing except nature can facilitate a complete cure for human machinery disorders. These are the various methods of Natural therapy:--  Massage  Acupuncture  Herbal/Proper Nutrition  Gems Figure 1: Representation of Natural Treatment of Diabetes Diseases III. DESIGN OF EXPERT SYSTEM The Expert System can be created by using ESTA by building the knowledge base. Expert System -- ESTA + Knowledge Base The Quality of Expert System is depends on its knowledge base. The process of developing knowledge base is:- a) Identifying the input of Problem b) Gaining Knowledge c) Representation of Knowledge A. Identifying the Input of Problem For developing the expert system first we have to identify the problem and its behaviors. The Input for our system is regarding identifying the different types of natural treatments (Massage, Acupuncture, Diabetes Diseases Treatment Methods 1) Natural Care(Herbal / Proper Nutrition) 2) Acupuncture 3) Homeopathic 4) Massage 5) Gems Treatment Solutions Advice (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3, No. 3, 2012 45 | P a g e www.ijacsa.thesai.org Herbal/Proper Nutrition and gems) available for Diabetes Disease. B. Gaining Knowledge Gaining Knowledge is very important in developing the expert system. The acquired gained knowledge is analyzed and processed to give the best solution of the problem. The knowledge of different types of Natural treatments (Massage, Acupuncture, Herbal/Proper Nutrition and gems) of Diabetes has been gained by reading books, browsing Internet and also got information from consultation of Physician in their respective areas. C. Representation of Knowledge Representation of knowledge is the last phase of the development of knowledge base system. There are various approaches for representation of Knowledge into knowledge base. Each Knowledge base contains rules for a specific domain. Thus, for a natural treatment of diabetes expert system the knowledge base will contain rules relating certain natural treatment methods of diabetes such as Massage, Acupuncture, Herbal/Proper Nutrition and gems. ESTA has all facilities to write the rules that will make up a knowledge base. Further, ESTA has an inference engine which can use the rules in the knowledge base to determine which advice is to be given to the expert system user or to initiate other actions Representation in ESTA is the rule based in logical paradigm of simple if-then rules in backward or forward chaining. We have chosen here the backward chaining for knowledge representation with simple if-do pair in place of if- then rules. Here we have considered two major knowledge representations namely Sections and Parameters. The top level of representation of knowledge in ESTA is section. It contains the logical rules that direct the expert system how to solve problem, actions to perform such as giving advice, going to other sections, calling to routines etc. The first section in ESTA is always named as start section. The advice is given when condition(s) in the section is (are) fulfilled. Parameters are used as variable and it determines the flow of control among the sections in the Knowledge Base. A parameter can be one of the four types: Boolean or logical, Text, Number and Category parameters. We have developed the various Parameters and Sections for developing this expert system. IV. REPRESENTATION OF SANJEEVANI EXPERT SYSTEM The Knowledge representation in ESTA is based on the items: a) Section b) Parameters c) Title A. Representation of Parameters Used in Developing Expert System In FIGURE 1: Here we have defined the disease parameter which is of type category describing the various types of disease. Figure 1. Screen Shots of Parameter Disease In FIGURE 2: Here we have defined the diabetesop parameter which is of type category describing the various types of natural treatment available for diabetes disease. Figure 2. Screen Prints of Parameter diabetesop B. Representation of Sections used in developing expert system In FIGURE 3: Here we have a main section start which is developed to transferring controls in accordance with the user’s response about disease Figure 3. Screen Print of Start Section In FIGURE 4: Here causeofdiabetes Section describes the diabetes disease and its symptoms (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3, No. 3, 2012 46 | P a g e www.ijacsa.thesai.org Figure 4. Screen Print of Causeofdiabetes Section In FIGURE 5: Here in the diabetesoption section describes the various natural treatment options available for diabetes disease. Figure 5. Screen Print of Section of diabetesoption In FIGURE 6: Here in the treatdiabetesnatural section describes the Natural Care (Herbal / Proper Nutrition) treatment solution of diabetes disease Figure 6. Screen Print of Section treatdiabetesnatural 1.1. Representation of Title In FIGURE 7: It describes the Title of Sanjeevani expert system. Figure 7. Screen Print of Title of expert System V. CONSULTATION OF EXPERT SYSTEM In FIGURE 8: It describes the beginning of Consultation of Sanjeevani Expert System. It will ask the users to select the disease (Diabetes) for which they want different type of natural treatment solution. Figure 8. Screen Print of ESTA Consult of Diabetes In FIGURE 9: It describes the diabetes diseases and its symptoms Figure 9. Screen Print of Diabetes description and Symptoms In FIGURE 10: It describes the different types of Natural treatment methods available for Diabetes disease. It will ask users to select one of the Natural treatment methods for getting details treatment advice. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3, No. 3, 2012 47 | P a g e www.ijacsa.thesai.org Figure 10. Screen Print of selection of Natural Treatment Method In FIGURE 11: It describes the Natural Care (Herbal / Proper Nutrition) treatment solution of diabetes disease Figure 11. Screen Print of Natural Treatment Solution VI. CONCLUSIONS As the Sanjeevani project shows, applied work in developing of expert system describing the various Natural therapy methods for treatment of Diabetes disease and various other diseases. The field of medical artificial intelligence is particularly appealing to the physicians and computer scientists working in the area. The long-term challenges are well recognized-such as the need for mechanisms that will assure completeness and shared knowledge bases for different natural methods treatment of different diseases. This Expert System development is one of the steps to get the integrated knowledge base system which will help in getting information of various natural treatment methods available for disease to general users (Patient and Physician) REFERENCES [1] Rajkishore Prasad, Kumar Rajeev Ranjan, and A.K. Sinha, “AMRAPALIKA: An expert system for the diagnosis of pests, diseases, disorders in Indian mango,” Knowl.-Based Syst. 19(1): 9-21 (2006). [2] Francisco Elazegui, and Zahirul Islam, “Diagnosis of common diseases of rice,” © 2003, International Rice Research Institute. [3] S.S.Patil, B.V.Dhandra, U.B.Angadi, A.G.Shankar, and Neena Joshi, “Web based Expert System for Diagnosis of Micro Nutrients Deficiencies in Crops”, Proceedings of the World Congress on Engineering and Computer Science 2009 Vol I WCECS 2009, October 20-22, 2009, San Francisco, USA. [4] Turban, E., Expert Systems and Applied Artificial Intelligence. New York: Macmillan Publishing Company, 1992 [5] Prolog Development Center, Expert System Shell for Text Animation (ESTA), version 4.5, A/S. H. J. Volst Vej 5A, DK-2605 Broendby Denmark, copy right©1992-1998 Book [6] Expert System: Principles and Programming by Joseph C.Giarratano [7] PROLOG –Programming for Artificial Intelligence by Ivan Bratko Software [8] Visual Prolog 7.3 ESTA Application AUTHOR’S PROFILE Sanjeev Kumar Jha has received M.C.A from Madras University in 2002; Presently He is pursuing PhD from Ambedkar Bihar University, Muzaffarpur. He has a Software Industry experience of more than 7 years. His areas of interest include Artificial Intelligence, Expert System and Software Testing. 
work_qpwknkiy6fe53n2xd5klp2ww54	----	Expert System for Determining the Level of Stress before Pediatric Dental Treatment Procedia Technology 12 ( 2014 ) 548 – 557 2212-0173 © 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY-NC-ND license. Selection and peer-review under responsibility of the Petru Maior University of Tirgu Mures. doi: 10.1016/j.protcy.2013.12.528 The 7th International Conference Interdisciplinarity in Engineering (INTER-ENG 2013) Expert system for determining the level of stress before pediatric dental treatment Sorana-Maria Bucura,*, Manuela Chibeleana, Adrian Gligorb, Mariana Pacurara a University of Medicine and Pharmacy Tg-Mures, 38 Gh. Marinescu Street, Tg-Mures 540139, Romania b“Petru Maior” University, 1 Nicolae Iorga street, 540088 Tirgu Mures, Romania Abstract The paper is focused on finding a solution based on an expert system designed to ease the dentist’s treatment decision by evaluating the patient’s anxiety and stress status. To identify the requirements and functionalities of this system there was performed an extensive research based on psychological questionnaires and the use of a powerful device made to determine body energy and stress variables. The experimental results were summarized and conclusions related to the proposed solution were discussed. Then we presented the software specially designed to establish the opportunity of clinical intervention considering the patient’s stress and anxiety. © 2013 The Authors. Published by Elsevier B.V. Selection and peer-review under responsibility of Department of Electrical and Computer Engineering, Faculty of Engineering, “Petru Maior” University of Tîrgu Mureș. Keywords: expert system; medical decision support system for dentists; anxiety; dental treatment. 1. Introduction Before initiating any dental treatment it is relevant to evaluate the clinical cooperation of the patients [1]. Getting a successful result in pediatric dentistry needs a good communication with the young patient. This is not possible when fear intervenes in the confrontation with the doctor [2]. * Corresponding author. Tel.: +0-040-753-756-551 E-mail address: bucur_sorana@yahoo.com Available online at www.sciencedirect.com © 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY-NC-ND license. Selection and peer-review under responsibility of the Petru Maior University of Tirgu Mures. ScienceDirect http://creativecommons.org/licenses/by-nc-nd/3.0/ http://creativecommons.org/licenses/by-nc-nd/3.0/ 549 Sorana-Maria Bucur et al. / Procedia Technology 12 ( 2014 ) 548 – 557 Less compliance could be easily associated to the person’s fear. Any dentist could expect a certain anxiety or fear from the young patient but these become a problem when they are exaggerated [1, 2]. Many patients faced with unpleasant or painful past experiences consequently manifest an unjustified fear when confronted with the doctor, the dental equipment, different smells of the dental office or sounds produced by dental drills. Exaggerated fear can lead to poor collaboration with the dentist, a difficult treatment, bad results and undesirable events [3, 4]. The goal of the study was to determine the exact anxiety level in patients using a computer program specially designed for this purpose. This software establishes the opportunity of medical intervention considering the patient’s anxiety and related symptoms. 2. Theoretical background Anxiety includes state anxiety which is defined as “a transitory emotional condition that varies in intensity and fluctuates over time” whereas” trait anxiety is a personality trait that remains relatively stable“[5]. It refers to the tendency to be anxious without external stress. First, we determined the patient’s anxiety level by using a psychological questionnaire, Spielberger’s State-Trait Anxiety Inventory - STAI - which has been approved and validated in many studies [6]. We have adapted this questionnaire in order to better match our study. The Romanian translation of the questionnaire was used. STAI is composed of two scales, A-state and A-trait, measuring distinct anxiety concepts: state anxiety (how someone is feeling in a certain condition) and trait anxiety (common fear of someone). ” Both scales consist of twenty items for which a person rates anxiety on a scale from one (almost never) to four (very much so)” [5]. It has been shown that A-state scores increase after various kinds of stress and decrease after relaxation training [7]. Therefore, this scale can be used to measure changes in the intensity of the anxious condition occurring in certain situations. Generally, those who score high in A-trait will show increases in A-state more frequently than individuals who score low in A-trait because they tend to react to a large number of cases, judging them dangerous or threatening [8]. State anxiety is a transitory emotional state or condition of the body characterized by subjective feelings, consciously perceived tension and fear and increased activity of ANS-Autonomic Nervous System [9]. This nervous system controls vital functions such as heart activity, blood pressure, digestion, salivation, perspiration and exchanges between body and environment [10]. The sympathetic nervous system plays a role in stress situations with all aspects of increased secretion of adrenaline: peripheral vasodilatation, acceleration of heart rate, rise in blood pressure, sweating, hyper salivation; skeletal muscles are well supplied with blood at the expense of internal organs, the pupil of the eye is increased, the whole body being in a state of alarm. The second step of the study was to determine the stress status of the patient with a powerful device used in traditional Chinese medicine and homeopathy [11] approved in E.U. Over the last decades scientists have shown that electromagnetic fields govern biological systems. The device measures the energy level of the whole body shown by Energy parameter which is the average of energy values determined in 24 energy points located on the skin surface of hands and feet, corresponding to various organs and systems. It also evaluates the stress status of the autonomic nervous system associated with symptoms and gives us indications of potential risks for the body health when these values are increased [12]. 3. Research methodology In order to investigate the correspondence between the values given by anxiety questionnaires and those shown by device’s parameters we formulated the following hypotheses: Research hypotheses: We assume that changes in the normal values of trait anxiety can lower the body energy shown by Energy parameter of the device and lead to diseases; 550 Sorana-Maria Bucur et al. / Procedia Technology 12 ( 2014 ) 548 – 557 We suppose that high values of state anxiety are related with high values of the ANS parameter which is shown in a strong relation with stress symptoms by the device we used. 3.1. Research variables In the present research we established an independent variable and dependent variables measured with digital and nominal scales, thus: Age of the patients was the independent variable Dependent variables were: State anxiety expressed by A-state score Trait anxiety expressed by A-trait score Energy parameter which shows the energetic level of the body; we called it E for ease of expression. ANS parameter which reveals the stress at a certain moment; we called it ANS for ease of expression because it’s related to the autonomic nervous system activity. 3.2. Methods To undergo this study, we used the following methods: 1. Psychometric methods STAI questionnaire where we adapted A-scale to our research by adding the formulation “because of the state of my teeth” to the items 2, 3, 6, 7, 9, 10, 12, 14, 16-20. It is allowed even indicated by the authors to adapt their A- state scale to a specific situation for more accurate results. energetic measurements using the device described above 2. Statistical methods descriptive statistical indices correlations The interpretation of the results was performed by the SPSS statistical program. 3.3. Sample of subjects. The research was based on a group of subjects consisting of 30 patients, girls and boys aged between 11 and 18, randomly selected. Toolbox used: clinical observation psychological questionnaire measuring device for highlighting energy level and stress status 3.4. Working procedure The experimental study began with a clinical examination of each patient. Following discussion and selection, the subjects were classified according to their age and sex. 3.5. Presentation and data analysis Fig.1 shows the patients’ distribution according to sex criterion. 551 Sorana-Maria Bucur et al. / Procedia Technology 12 ( 2014 ) 548 – 557 Fig. 1. Patient’s distribution according to sex criterion chart So the randomly chosen sample contains mostly girls with a difference of 46.67% between girls and boys. Fig.2 shows the distribution of patients according to age. 16,66% 23,33% 23,33% 36,66% 0,00 5,00 10,00 15,00 20,00 25,00 30,00 35,00 40,00 11-12 13-14 15-16 17-18 age % Fig. 2. Distribution of patients according to age From the chart above we can see that the predominant adolescent age group (17-18 year old) represents 36.66% of the subjects investigated. The patients from our sample filled in the Spielberger’s State -Trait Anxiety Inventory STAI with those two scales described above - A-trait and A-state - and the results are presented by the next charts: Fig. 3. Distribution of patients according to the values obtained at A-trait questionnaire From the chart above we see that patients who scored values between 35 and 45 represent 69.99% of the research sample. The patients’ distribution according to the values obtained at A-state questionnaire is presented in Fig.4. Girls 73.33% Boys 26.66% 6.66% 36.66% 33.30% 10% 13.33% Score 30-35 Score 35-40 Score 40-45 Score 45-50 Score 50-75 552 Sorana-Maria Bucur et al. / Procedia Technology 12 ( 2014 ) 548 – 557 6.66% 36.66% 33.30% 10% 13.33% Score 30-35 Score 35-40 Score 40-45 Score 45-50 Score 50-75 Fig. 4. Distribution of patients according to the values obtained at A-state questionnaire Fig.5 shows the distribution of patients according to the energetic level of the body, “E”. Fig. 5. Distribution of patients according to the energetic level of the body, “E” From the above chart we can see that the majority of patients, 59.98% which represents more than a half of the research group, were located between 25-55 normal values of this parameter. The ANS normal values are from 0 to 2. More than 2 means we are dealing with a stress status. The patients’ distribution according to the ANS parameter’s values is presented in Fig.6. 23.33% 26.66% 33.33% 3.33% 3.33% value 1-2 value 2-3 value 3-4 value 4-5 value 6-7 Fig. 6. Distribution of patients according to the ANS parameter 6.66% 26.66% 13.33% 19.99% 20% 6.66% 6.66% value 10-20 value 20-30 value 30-40 value 40-50 value 50-60 value 60-70 553 Sorana-Maria Bucur et al. / Procedia Technology 12 ( 2014 ) 548 – 557 4. Results. Main findings 4.1. Results related to Hypothesis 1 We assume that changes in the normal values of trait anxiety can lower the body energy shown by Energy parameter (E) of the device and lead to diseases. In order to verify this hypothesis we used correlations and t-test from statistical program SPSS. We correlated Variable 2-A trait with Variable 3-E and we saw that their means are relatively close which leads to the conclusion that there is a strong relation between the trait anxiety and the energetic level of the body. Table 1: Means Cases Percent Excluded VAR00002 VAR00001 VAR00003 VAR00001 Included Percent Percent N N 30 49.2% 31 50.8% 100.0% 30 49.2% 31 50.8% 100.0% Legend: Var0001=age, Var0002=-A-trait, Var0003= E Table 2: Relations between Variable 2 and Variable 3 Total Mean 42.7333 41.4667 N 30 30 Std. Deviation 6`.4108 16.4981 Correlations Table 3: Correlations between Variable 2=A-trait and Variable 3=A-state VAR00002 VAR00003 VAR00002 Pearson Correlation 1.000 .342 Sig. (2-tailed) . .064 N 30 30 VAR00003 Pearson Correlation .342 1.000 Sig. (2-tailed) .064 . N 30 30 Legend : Pearson=correlation, Var 0002=Variable A-trait, Var0003=Variable E One-Sample Statistics Table 4: t-test N Mean Std. Deviation Std. Error Mean VAR00002 30 42.7333 6.4108 1.1705 VAR00003 30 41.4667 16.4981 3.0121 Legend : Variable 1=age, Variable 2=A-trait, Variable3=E, N=subjects t= -0.342 statistically significant at a significance of 0.05, which shows that the difference is significant and hypothesis 1 is valid. So, we found out by statistical methods that hypothesis 1 is valid. 4.2. Results related to Hypothesis 2 We suppose that high values of state anxiety are related with high values of the ANS parameter which is shown by the device used in a strong relation with stress symptoms. 554 Sorana-Maria Bucur et al. / Procedia Technology 12 ( 2014 ) 548 – 557 Case processing summary: Table 5: Means Cases Included Excluded Total N Percent N Percent N Percent VAR00001 VAR00002 30 100.0% 0 .0% 30 100.0% We correlated Variable 2-A trait with Variable 3-E and we saw that their means are relatively close which leads to the conclusion that there is a strong relation between the trait anxiety and the energetic level of the body. Table 6: Report VAR00001 VAR00002 Mean N Std. Deviation Total 48.2000 30 8.2813 From the table above we see a standard deviation of 8.2813 between the two variables A-state and ANS their mean being 48.200 One-Sample Statistics Table 7: t-test N Mean Std. Deviation Std. Error Mean VAR00001 30 48.2000 8.2813 1.5119 VAR00002 30 2.7333 1.0694 .1952 We see that mean between the two variables is 48.200 bigger than 45.4667 and standard deviation is 1.3167, which means that there is a connection between the stress status shown by ANS parameter of the device and anxiety as a state evidenced by A-state questionnaire. Table 8: One-Sample Test Test Value = 0 t df Sig. (2-tailed) Mean Difference 95% Confidence Interval of the Difference Lower Upper VAR00001 31.879 29 .000 48.2000 45.1077 51.2923 VAR00002 14.000 29 .000 2.7333 2.3340 3.1326 Table 9: Correlations VAR00001 VAR00002 VAR00001 Pearson Correlation 1.000 .667 Sig. (2-tailed) . .000 N 30 30 VAR00002 Pearson Correlation .667 1.000 Sig. (2-tailed) .000 . N 30 30 ** Correlation is significant at the 0.01 level (2-tailed). Legend : Variable 0001=A-state, Var0002=ANS 555 Sorana-Maria Bucur et al. / Procedia Technology 12 ( 2014 ) 548 – 557 t = -0.667 statistically significant at a significance less than 0.01 which shows that the difference is significant and hypothesis1 is valid. So, we found out by statistical methods that hypothesis 2 is valid. 5. System implementation and testing We choose to implement a web based expert system prototype. The architecture of the application is shown in Fig. 7. Fig. 7. Application architecture. We determined strong correlations between A-trait and E, A-state and ANS; those allowed the design of an expert system developed to ease treatment decision. Following we present the knowledge base of the expert system built to assist the dentist to take the most adequate therapeutic decision [13, 14, 15]. The rule base of the developed expert system is built on the following [16]: If E is between 25-55 and ANS less than 2, we start the treatment because stress status of the patient is normal. If E is less than 25 and ANS less than 2, we start the treatment because state anxiety has normal values but we suspect chronic diseases due to the lack of the body energy; so we advice the patient to see a doctor. If E is more than 55 and ANS less than 2, we start our treatment very carefully because any time the patient may become stressed. If E is between 25-55 and ANS more than 2, we postpone our treatment and apply psychological relaxation techniques. If E is less than 25 and ANS more than 2, we postpone our treatment and apply psychological relaxation techniques. We advice the patient to investigate his/her health because of lack of the body energy. If E is more than 55 and ANS more than 2, the treatment is compromised because of the exaggerated stress status of the patient. The system was tested extensively, using historical medical records. The small number of rules and their semantic consistency generates fast and reliable solutions, with real support for the decision-making process. For example: Patient A.I., female, 13 years, E=35, ANS=2.74, we postponed our treatment and applied psychological relaxation techniques. Then we could successfully complete treatment. Patient S.V., male, 14 years, E=22, ANS=1.71, we applied the treatment without any problems but in anamnesis the child reported diabetes. Application Server Inference module Storage Web-based client Security and access module Domain & Knowledge Database 556 Sorana-Maria Bucur et al. / Procedia Technology 12 ( 2014 ) 548 – 557 Patient B.I., female, 17 years, E=70, ANS=3.56, we could not even start our treatment because of the patient’s high level of stress. Patient T.R., male, 15 years, E=67, ANS=1.87, we finished our treatment with some difficulties. Patient A.S., female, 17 years, E=45, ANS=1.62, we completed our treatment without any problems. Patient T.R., male, 16 years, E=21, ANS=3.11, we postponed our treatment and applied psychological relaxation techniques. We advised the patient to investigate his health because of lack of the body energy. 6. Conclusions The present work was carried out a research on anxiety which led to the development of a medical decision support application. We found out that changes in the normal values of trait anxiety can lower the body energy shown by Energy parameter of the device and lead to diseases. We found that high values of state anxiety are related with high values of the ANS parameter which is shown in a strong relation with stress symptoms by the device used. So, we can correlate the two device’s parameters values with those given by the most common and used anxiety questionnaire, practically we validated the device’s measurements. Then became possible to develop a software application designed to ease doctor’s work by helping him with the treatment decision. This expert system greatly eases clinical work by helping the dentist to take the best medical decision at the beginning of the young patient’s treatment. Acknowledgement This study was entirely supported by the Private Medical Office belonging to Dr. Bucur Sorana-Maria and approved by Decision 95 of the Ethics Committee of the University of Medicine and Pharmacy Tg-Mures. We respected the confidentiality of personal data of the patients. The parents of the minor patients were informed about the study and agreed with their children participation by writing an informed consent. They agreed with the use of the data obtained in scientific purpose and for publication. We are grateful to the office staff for their help in conducting the study. The paper was partly supported by the Sectorial Operational Programme Human Resources Development (SOP HRD), financed from the European Social Fund and by the Romanian Government under the contract number POSDRU 80641. References [1] Vassend O, Willumsen T, Hoffart A. Effects of Dental Fear Treatment on General Distress: The Role of Personality Variables and Treatment Method. Behav Modif 2000, 24: 580-599. [2] Townend E, Dimigen G, Fung D. A clinical study of child dental anxiety. Behav Res Ther 2000, 38: 31-46. [3] Kent G, Warren P. A Study of Factors Associated with Changes in Dental Anxiety. J of Dental Research, November 1, 1985, 64: 1316-1318. [4] Arnrup K, Broberg AG, Berggren U, Bodin L. Temperamental reactivity and negative emotionality in uncooperative children referred to specialized paediatric dentistry compared to children in ordinary dental care. Int J Paediatric Dentistry 2007, 17: 419-429. [5] Spielberger CD, Gorsuch RL, Lushene R. Manual for the State-Trait Anxiety Inventory. Consulting Psychologists Press, Palo Alto, California, 1983. [6] Okun A, Stein RE, Bauman LJ, Silver EJ. Content validity of the State-Trait Anxiety Inventory from the perspective of DSM-IV. Psychol Rep 2006, 79: 1060-1069. [7] Eppley KR, Abrams AI, Shear J. Differential effects of relaxation techniques on trait-anxiety: A meta-analysis. J of Clinical Psychology, 1989, 45: 957-974 . [8] Grossman P, Niemann L, Schmidt S, Walach H. Mindfulness – based stress reduction and stress benefits: A meta-analysis. J of Psychosomatic Research 2004, 57: 35-43. [9] West GA, Reid KH, Bastawi AE. Clinical Science: Autonomic Responses to Dental Procedures in Pedodontic Patients During a Standard Restoration Session. J of Dental Research 1985 62: 728-732. [10] Watson D, Pennebaker, JW.Health complaints, stress and distress: Exploring the central role of negative affectivity. Psychol Review, 1989, 96(2): 234-255. 557 Sorana-Maria Bucur et al. / Procedia Technology 12 ( 2014 ) 548 – 557 [11] Sherman KI, Hogeboom CI, Cherkin DC. How traditional Chinese medicine acupuncturists would diagnose and treat chronic low back pain: results of a survey of licensed acupuncturists in Washington State. Complementary Therapies in Medicine, 2001, 9: 146-153. [12] Boss A, Hoogstraten J, Prahl-Andersen B. On the use of personality characteristics in predicting compliance in orthodontic practice. Am J Orthod Dentofacial Orthop 2003, 123: 569-572. [13] Russello G, Chaudron M R, van Steen M, Bokharouss I – An experimental evaluation of self-managing availability in shared data spaces. Science of Computer Programming 2007, 64: 246-262. [14] Krikhaar R, Crukovic I – Software Configuration Management. Science of Computer Programming 2007, 65: 215-221. [15] Colman A, Han J - Using role-based coordination to achieve software adaptability. Science of Computer Programming 2007, 64: 223-245. [16] American Academy of Pediatric Dentistry - Guidelines for behavior management. Pediatr Dent, Supplement issue 2002-2003, 24: 68- 74.14Krikhaar R, Crukovic I – Software Configuration Management. Science of Computer Programming 2007, 65: 215-221. 
work_qqbfeohubbcqpahq4btcqwzt6e	----	PII: 0898-1221(90)90117-3 Computers Math. Applic Vol. 20. No 9.10. pp. 12~140. 1990 0097-4943+90 $3.00+0.00 Printed m Great Britain. All rights rese~'ed Copyright '._C, 1990 Pergamon Press plc D I A G N O S I N G J A U N D I C E EXPERT SYSTEM L. U. YAL~?INALP a n d L. STERLING Department of Computer Engineering and Science, Case Western Reserxe University. Cleveland, OH 44106, U.S.A. Abslract--DIJEST (Diagnosing Jaundice Expert SysTem) is a medical expert system which produces a differential diagnosis of a patient presenting with jaundice. DIJEST is written in Prolog. and illustrates the use of the language for clearly expressing knowledge. Specifically, the expert system contains explicit declarative knowledge of anatomy and physiology which is used by clinicians when diagnosing obstructive jaundice. The inference engine matches patient records against expected manifestations of symptoms in diseases. Novel m DIJEST is the uncertainty reasonmg scheme, using contribution and absence factors, which places equal importance to symptoms present, absent and unknown in the patient's medical record. Domain specific reasoning and domain specific knowledge are clearly separated from general inference capabilities and knowledge representation schemes. DIJEST has performed well in preliminary tests, being particularl2, impressive for patients with multiple diseases. I. I N T R O D U C T I O N Research in medical expert systems, a m a j o r a p p l i c a t i o n a r e a o f AI, has led to the d e v e l o p m e n t of, a n d e x p e r i m e n t a t i o n with, new schemes for representing knowledge. N o universal tool o r technique has emerged. Each research g r o u p has its o w n style affected by the p r o b l e m d o m a i n a n d the medical expertise being used. T h i s p a p e r describes a new medical expert system, D I J E S T ( D i a g n o s i n g J a u n d i c e Expert SysTem), which is c o n c e r n e d with the differential diagnosis o f patients with o b s t r u c t i v e jaundice. D I J E S T evolved with its m a j o r objective to explore present k n o w l e d g e r e p r e s e n t a t i o n techniques a n d to i n t r o d u c e a declarative style for modelling clinical p r o b l e m solving. Subsequently a n o t h e r issue b e c a m e critical, n a m e l y the modelling o f u n c e r t a i n t y r e a s o n i n g d u r i n g the m a n y stages o f c o n s u l t a t i o n a n d diagnosis o f a disease. D I J E S T has yet a n o t h e r f r a m e based scheme for k n o w l e d g e r e p r e s e n t a t i o n a n d r e a s o n i n g with uncertainty. It is d e v e l o p e d using Prolog. We are able to c o m b i n e different knowledge represen- tation techniques in a single f r a m e w o r k due to the flexibility o f P r o l o g in the design o f different d a t a structures for the system. Specifically features o f f r a m e - b a s e d a n d rule-based r e p r e s e n t a t i o n s were integrated with a new calculus for u n c e r t a i n t y reasoning. G e n e r a l medical k n o w l e d g e a b o u t the d o m a i n was easily represented in P r o l o g declaratively. It also enabled a clear r e p r e s e n t a t i o n o f inference as a specialized interpreter handling the d a t a structures. O u r scheme for u n c e r t a i n t y reasoning is novel due to a s t r o n g d e p e n d e n c e on the i n t e r p r e t a t i o n o f present a n d a b s e n t d a t a , i n f o r m a t i o n that is k n o w n to exist, known not to exist a n d i n f o r m a t i o n that is unknown at the time o f c o n s u l t a t i o n with the p r o g r a m . This was a c o n s t r a i n t imposed by o u r medical experts. It allows c o n t e x t - d e p e n d e n t e v a l u a t i o n o f the patient d a t a . T h e scheme uses c o n t r i b u t i o n a n d a b s e n c e factors which are a t t a c h e d to p a r t i c u l a r m a n i f e s t a t i o n s o f a disease. These f a c t o r s c o n s t i t u t e a numerical r e p r e s e n t a t i o n which c o m p l e m e n t s the qualitative descriptions in D I J E S T . These qualitative descriptions m i r r o r the medical e x p e r t s ' definition o f the characteristics o f m a n i f e s t a t i o n s . O u r work has been influenced by the f a m o u s medical expert systems, M Y C I N , Internist a n d PIP. T h e r e p r e s e n t a t i o n o f k n o w l e d g e using f r a m e s is similar to P I P ' s [I]. T h e c o n c e p t o f c o n t r i b u t i o n a n d absence f a c t o r s evolved f r o m investigation o f the confidence f a c t o r s o f M Y C I N [2] a n d the e v o k i n g strengths a n d frequencies in Internist [3]. In addition, representing c o m m o n - s e n s e k n o w l e d g e in D I J E S T is affected by the r e p r e s e n t a t i o n o f properties in Internist [4] a n d the use o f logical decision criteria in P I P [5]. T h e basic system was designed to be e x p a n d e d a n d e n h a n c e d to i n c o r p o r a t e the stages o f clinical r e a s o n i n g d u r i n g the c o u r s e o f a p a t i e n t ' s t r e a t m e n t . D I J E S T has been tested o n s a m p l e p a t i e n t cases t a k e n f r o m medical t e x t b o o k s a n d patient records. It p e r f o r m s at an a c c e p t a b l e level 125 126 L.O. YAL~INALP and L. STERLING according to o u r experts. An interesting feature is its handling o f multiple diseases contributing to the jaundice. T h e paper is organized as follows. After a brief overview o f D I J E S T s scope, we present D I J E S T s architecture and the multi-layered knowledge representation in the system. Th e next section describes the uncertainty reasoning mechanism which underlies the modelling o f diagnosis, followed by o u r conclusions. We emphasize in this a c c o u n t how P ro l o g can be used to develop an expert system. 2. S C O P E O F D I J E S T 2. I. The problem o f obstructit.'e jaundice Jaundice is the yellow pigmentation o f the skin o r sceleras by bilirubin. This in turn is a result o f elevated levels o f bilirubin in the blood stream [6]. T h e r e are several reasons for this elevation. Most o f the bilirubin is derived from the catabolism o f hemoglobin present in the red blood cells. T h e bilirubin is t r a n s f o r m e d into bile and the liver plays a central role in this metabolism o f the bile pigments. T h e d e r a n g e m e n t s o f this metabolism cause several diseases which have jaundice as a c o m m o n symptom. T h e elevation o f the bilirubin might be related to pathogenetic mechanisms o r disease processes. We are c o n c e r n e d a b o u t a subset o f these diseases which cause obstructieejaundice. This is jaundice due to the mechanical o b s t r u c t i o n o f the biliary radicles o r functional factors that cause impaired hepatic excretion o f bilirubin into bile. Figure I is a simplified diagram o f the organs that are related to the flow o f bile to the intestine after its excretion from the liver. T h e enlargement o f any organ near the bile ducts can block the flow o f bile, thereby causing obstructive jaundice. T h e principal examples are inflammation o f the gallbladder, liver or pancreas, a t u m o r or a cystlike mass in the head o f the pancreas. Obstructive jaundice can also be caused by gallstones leaving the gallbladder, lodging in the bile ducts and blocking the flow. Diagnosing the most c o m m o n causes o f obstructive j au n d i ce as mentioned ab o v e is the p r i m a r y focus o f D I J E S T . Specifically, D I J E S T considers viral hepatitis, alcoholic hepatitis, cirrhosis, cholecystitis, choledocholithiasis, pancreatitis, pancreatic can cer and pancreatic pseudo- cyst. Hepatitis is the inflammation o f the liver. We are co n cern ed with two types o f hepatitis, one is caused by excessive c o n s u m p t i o n o f alcohol and the o t h e r by virus. Cirrhosis is the ch ro n i c irreversible injury o f the liver. Cholecystitis is the inflammation o f the gallbladder where choledocholithiasis refers to the obstruction o f the bile duct by gallstone(s). Pancreatitis is inflammation o f the pancreas. Pancreatic cancer refers to a cancerous growth, while pancreatic pseudo cyst refers to the cystlike masses at the head o f the pancreas. It is critical to differentiate the mechanism that causes the o b st ru ct i o n and the site o f the o b s t r u c t i o n in clinical practice. T h e r e f o r e , D I J E S T is designed to p r o d u c e possible diagnosis by , ~. ~ I n ~ r o h e p O t l C ncrreos Cyst,C ouct Deoclenurn----..___ " ~ , ~ l lO of voter ~UCtS Fig. I. The anatomy of organs participating in the bile flow. DIJEST 127 rPat*ent profile l ~--4 (.,sto~y, I I cLinicaL e~am, I L t a d t e s t s ) .~ I Disease desCriptors I General medicaL knowledge MATCHER Candidate d;seases I EvaLuation of pat=ent I_ (dynamic patient data) J- .__L . . . . . -=i t _f -L SCREENING LikeLihood estimates for candidate diseases " . . . . . . . ] t i ; I I L . . . . . . . . . . . . . . . l r . . . . . . . . . . . . - : t t [ . . . . . U ~ r . . . . . . . interaction Fig. 2. The architecture of DIJEST. indicating the likelihood of each of these diseases and the differentiating factors that leads to the diagnosis. 2.2. Architecture o f D I J E S T DIJESTs system structure as initially planned is shown in Fig. 2. The system is constructed around its most important component, the specialized interpreter that we call the MATCHER. In this section we describe the function of the system components. The boxes above the MATCHER indicate the knowledge used by DIJEST. Diseases are represented by disease descriptors. The candidate diseases are the list of diseases that are to be considered for differential diagnosis. The patient profile consists of all the knowledge related to a particular patient. The MATCHER analyzes the patient producing an evaluation of the patient and likelihood estimates for candidate diseases. The MATCHER evaluates the current mixture of known, uncertain, partially satisfied and unknown findings of a patient with respect to the candidate diseases. The evaluation of each patient includes any contradictory evidence and suggestions for additional tests that should be performed. The results will provide feedback for the next stage of diagnosis, More details of the MATCHER are given in Section 4. A full evaluation of the output of the MATCHER was intended to be considered by the screening process. Currently the patient profile is examined for the manifestations expected by the disease descriptors. The screening process would also evaluate significant patient data that is not explained by the differential diagnosis. 3. K N O W L E D G E R E P R E S E N T A T I O N IN DIJEST The knowledge base of DIJEST consists of medical knowledge about jaundice and information about the patient. Knowledge of jaundice is divided into descriptions of the diseases which cause 128 L. 0. YAL(~INALP and L. STERLING disease_history(noL_appl, disease_desc(choledocholithiasis, history(noL.appl, ~previous illness not_appl, ~pre~ious tests not.appl, %e~osed_to not_appl, ~family_background [ %symptoms (pain ,[site(abdomen,S), severity(none_to_severe,I), continuity(intermittent,S), duration(short,3), coupled_by(nausea, 1 I, coupled_by(vomiting, 1 I, threshold(131]. contribution_absence_factors(0.9,-O. 111, normalization J'actor(O.gl], not.appl, ~obser~ations not_appl, J~drng.use not_appl ~surgery 1, clinical( [(jaundice,[pace(fast,2), pace(medium_slow, 1.31, threshold(1.31], contribution_absence_factors(0.6,-2.011, (gallbladder,present,contribution_absence.factors(O.9,-2.011 , (tender_abdomen,[site(upper.quadrant,S I, condition(attacks,S), threshold(lO)], contribution_absence_factors(0.7,-2.0)), normalization.Saetor(2.21] 1, labtests( [(obstructive.tests,disease_related,contribution_absence_factors(O.7,-2.011, (gallbladder, gallstones(presentl,contribution_absence_factors(0.8,0.211, (common.bile_duct, obstruction(presentl,contribution_absence_factors(O.9,-2.0/), (common.bile_duct, dilatation(abnormal),contribution_absence_factors(O.CJ,-2.0/), normalization.factor(2.7)/] 1. Fig. 3. Disease descriptor of choledocholithiasts. jaundice, and general medical knowledge. This section describes the significant points in o u r representation. 3. I. Disease descriptors The diseases that have j a u n d i c e as a c o m m o n s y m p t o m are represented individually in D I J E S T by disease descriptors or DDs. A D D describes the p r o t o t y p i c a l characteristics o f a patient who has the disease and it is represented by a framelike structure in Prolog. Figure 3 shows the disease descriptor for choledocholithiasis. Each disease descriptor is a q u a d r u p l e indexed by the disease name. T h e expected characteristics related to the history, clinical examination and the laboratory tests are the c o m p o n e n t s o f this structure. We will refer to those three c o m p o n e n t s as contexts. Each co n t ex t consists o f a n u m b e r o f slots that show the logical subdivision within that context. F o r example, the history context has eight slots showing previous diseases, previous tests, en v i ro n m en t al and clinical factors which suggest the disease if the patient has been exposed to them, facts that are related to family DIJ EST 129 b a c k g r o u n d , expected previous symptoms, expected physical observations, usage o f particular drugs and previous a b d o m i n a l surgery, respectively. T h e co n t ex t s for b o t h clinical ex am i n at i o n and l a b o r a t o r y tests have only one slot. Slots that are not applicable fo r a specific disease are shown by no_appl as illustrated by Fig. 3. Each slot consists o f specific characteristics related to the slot. T h e y are called elements. T h e y are indicated by slanted uppercase letters in Fig. 3. An element is either a single characteristic o r a disjunction o f characteristics. A single characteristic is called a key tuple and is indexed by a key. A key is the smallest c o m p o n e n t o f this layered structure and it can be either a direct key or an extractable key. A key tuple consists o f key attributes, and a pair c o n t a i n i n g a contribution factor and an absence factor. The key attributes o f a key tuple show the characteristics o f a key which are related to the disease. T h e y are defined with respect to the key type. T h e c o n t r i b u t i o n and absence factors are related to the uncertainty reasoning handled by the M A T C H E R and will be dis- cussed in the next section. T h e y are represented as contribution~absence_factors(CF, AF) for clarity in Fig. 3. T h e type o f a key determines how the diagnosis is handled by D I J E S T . T h e different key types are known by the M A T C H E R and handled differently. T h e y are described below. !. Direct keys. A simple concept or a finding is represented by a direct key. T h e key attri- butes o f a direct key are given by a list o f qualitatit,e characteristics defining the key. F o r each attribute, a n u m b e r showing the i m p o r t a n c e o f this qualitative description with respect to the key is given as an integer between I and 5. It is called the significance ~,alue o f the attribute. T h e qualitative descriptions along with significance values describe the co n cep t or finding fully. F o r example, pain is a direct key, but its severity, d u r a t i o n or location differs from one disease to a n o t h e r as well as their relative significance for defining the pain. In Fig. 3, the respective attribute values o f pain for choledocholithiasis are shown. F o r example, a patient is expected to have pain with intermittent continuity. Since intermittent pain is an i m p o r t a n t indicator for choledocholithiasis, it is given a significance value o f 5. Every direct key is defined with a threshold that is used by the M A T C H E R for uncertainty reasoning. If. Extractable keys. An extractable key represents a medical c o n c e p t that can not be described as a simple finding. T h e concept needs to be extracted fro m the patient data, In o r d e r to simplify their representation, they are shown as a single key in the disease descriptors. T h e y are divided into four categories to simplify the diagnosis: (i) Some o f the keys represent anatomical o r physiological states o r concepts. T h e patient is expected to be in or have such states if he is likely to have the disease. F o r example, gallbladder is a key in the l a b o r a t o r y tests context, and the presence o f gallstones is its defining at t ri b u t e as shown in Fig. 3. Th e value o f this a t t r i b u t e is used as an aid for determining the site, o r the exact co n d i t i o n for such a key. Since there m a y be m a n y tests that would determine whether the patient has the specified state, this representation allows the M A T C H E R to determine the necessary diagnostic tests that would indicate the co n d i t i o n given in this key. (ii) T h e names o f blood tests are used as keys in D I J E S T to show the tests required in the diagnosis, for example bilirubin, amylase and sgot. Th e analysis o f the results o f a particular b l o o d test is disease dependent. Th e same blood test may indicate different likelihoods o f the presence o f a disease for different diseases. Possibili O, distribution cun,es are used to represent those ranges o f results o f b l o o d tests. T h e y are separate from the disease descriptors and were provided by o u r medical experts. H o w they are used with respect to the disease descriptors will be covered in the next section. (iii) Some keys refer to a collection o f simple o r complex findings. T h e individual findings in the collection might not be very significant on their own o r affect the 130 L. 0. YAL~INALP and L. STERLING (iv) diagnostic process. However, their c o m b i n a t i o n constitutes a medical co n cep t and should be considered as a composite finding. We call such keys compound keys. T h e collection o f individual findings are represented in separate tables for each c o m p o u n d key. C o m p o u n d keys are represented as a c o m b i n a t i o n o f direct or extractable keys, but the c o n t r i b u t i o n and absence factors are defined for the c o m p o s i t e meaning. Prodrome is an example o f a c o m p o u n d key which is used for diagnosing hepatitis. Figure 3 does not contain an example o f a c o m p o u n d key. Some keys represent rules that D I J E S T has to activate in o rd er to check the presence o f a disease. T h e y are used by the M A T C H E R to evaluate the patient d a t a that are related to different contexts or to c o m p a r e the tests results. F o r example, inflammation and obstructive tests are two rule names. T h e latter is shown in Fig. 3. T h e key attributes for this kind o f key are not used since the c o n c e p t to which they refer is e m b e d d e d in the rules. 3.2. Patient data T h e i n f o r m a t i o n a b o u t a patient is given as input to D I J E S T . All the i n f o r m a t i o n related to a patient is indexed by a unique patient number. It is given in four different frames, analogous to the contexts in the disease descriptors. We refer to the i n f o r m a t i o n a b o u t the patient as the patient profile. I. Patient ID record. Consists o f identification information. 2. Medicalhistorv. This frame is similar to the history co n t ex t o f a DD. An example frame for a patient is shown in Fig. 4. It consists o f six different slots c o r r e s p o n d i n g to the first six slots in a D D . D ru g and surgery i n f o r m a t i o n are represented as separate slots in the knowledge base if the patient has relevant data. medical.history(lO001, % pre~ious diseases [(jaundice,[occur rence(negative)]), (alcoholism,[occurrence(negative)I)], previous_tests [(wbc,[date(2.2,1987),site(blood),result(10200)])], %exposed to [(hepatotoxins,[exp_to(negative)]), (jaundiced_people,[exp_to(negative)])], ~famzly background [Cjau ndice, [occurrence(negative)])], %symptoms [(pain ,[date(15,1,1987), site(abdomen), severity(severe), continuity(intermittent), duration(6, days), coupled_by(nausea), coupled_by(vomiting)I). (.a-sea.[date(iS,1,198Z)]), (vomiting,[date(15,1,1987),cause(la rge_clinner)]), (intolerance_fatty_foods,[reaction(negative)])], %observations [(skin,[color(yellow)I), (u rine,[color(d ark)I), (stool ,[color(llght.brown)])])). Fig. 4. Medical history for patient IO00l. DIJ EST 131 . . Only simple findings with their related attribute values are represented in this frame. The attributes have both quantitatit,e and qualitatit,e descriptions. For example, the attribute duration for the key pain shows how long the patient has been in pain, and might have the value "6 days". Clinical examination. This frame consists o f two slots: the knowledge related to the actual physical examination o f the patient and the results o f the cardiovascular tests routinely taken. As in the history frame o f the patient, the simple findings are represented as binary key tuples. Tests. Each test frame for a patient is indexed by the test name and the patient ID number. It consists o f information about the date o f the test and its result with respect to the site where it is taken. For tests such as ultrasound, the results are given as a collection o f findings related to a site, its state and its condition, because those tests are used to determine the condition o f different parts o f the body. For blood or urine tests, their specific site is indicated along with a single result. Examples for patient 10001 are shown below. test(10001 ,date (2,3,1987),soot, [ (blood,80) ] ). test (10001 ,date (2,3,1987),alk_phosp, [ ( b l o o d , 1 2 0 ) ] ). test (10001 ,date(2,3,1 987),ultrasound, [ (gallbladder,edema,present), (common_bile_duct,dilations,s) (gallbladder,gallstones, present), (pancreas,swelling,normal), (pancreas_head,dilatation,normal), (pancreas,state,normal) ] ). 3.3. General medical knowledge Medical knowledge in D I J E S T is represented independently from any particular patient. Examples o f such knowledge are the general characteristics o f jaundice, what the available tests measure along with their possible sites, the restricted a n a t o m y o f the human body that concerns the d o m a i n diseases, the d o m a i n specific qualitative representation o f quantitative terms and the possibility distribution curves o f blood test results. This knowledge is used by the M A T C H E R when creating the differential diagnosis. F o r example, ultrasound is used to determine the presence o f gallstones in the gallbladder or the size o f the bile ducts, or sgot is a blood test and its primary function is to detect liver injury. This information is represented declaratively in DIJEST and illustrated below with a few examples. lab_test(ultrasound, [ (gallbladder,gallstones, [ present,a bsent] ), (gallbladder,edema, [present,absent] ), (pa ncreas_head,dilation, [normal,s,inc] ), (extra hepatic_d ucts,dilatation, [ normal,s, inc] ), (intrahepatic_d ucts.dilatation, [ normal,s, inc] ), (liver, hepatic_texture, [ homogen,not homogen ] ), (pancreas,swelling, [head,diffuse, normal] ), (pancreas,state, [atrophic,ind urated,cyst, normal] )] ). lab_test (u ribirilogen, [ (u rine,excretion_bile, [present.absent.decreased,increased] )] ). 4. P A T I E N T E V A L U A T I O N IN D I J E S T 4. !. A Prolog-based MA T C H E R As its name suggests, M A T C H E R compares a patient profile with the disease descriptors present in the system. It is a special interpreter written in Prolog which compares the frame structures, takes into account present, absent and u n k n o w n factors and establishes likelihood scores for the presence o f a disease. The findings o f a disease, namely its DD, is matched against a patient's profile in the three different contexts o f history, clinical exam and laboratory test data. A likelihood score is 132 L. 0. YAL~INALP and L. STERLING calculated for each context, and the overall likelihood score for the disease is c o m p u t e d as the average o f the scores for the three contexts. diagnose( Patient, History,Clinical,Tests,disease( Disease, DiseaseProb) ) ,-- disease_desc(Disease, DH,DC, DT), eval_history(Disease, Patient, History, ClinicaI,Tests, D H,H ist Prob), eval_clinical (Disease, Patient.H istory,ClinicaI,Tests, DC,ClinicalProb). eva I_tests ( D isease, Patient, H istory. C li n icaI,Tests, DT,Tests Prob), combine_prob(Hist Prob,ClinicaIProb,TestsProb, DiseaseProb). combine_prob(H P.CP,TP, FinalProb) ,- FinalProb is ( H P + C P + T P ) / 3 . 0 . L o o k i n g at the sample disease descriptor in Fig. 3, and a patient's medical history record fro m Fig. 4, it should be clear that the matching is not a direct unification o f expected values o f attributes for keys for a slot in a particular context. Nevertheless, the interpreter uses unification to determine key types to handle different type o f keys. Details o f the M A T C H E R will be covered in the follo~.ing sections. The findings o f a patient are evaluated with respect to a list o f d o m a i n diseases, called the candidate disease list. T h e c a n d i d a t e disease list in the c u r r e n t version o f D I J E S T is all the known diseases that are present in the knowledge base. Heuristic rules could be added as a front-end to generate a shorter list. F o r example, some sets o f s y m p t o m s suggest very strongly viral hepatitis and nothing else. At the m o m e n t , all the c a n d i d a t e diseases are processed in a straightforward m a n n e r and for each D D on the c a n d i d a t e disease list, a likelihood score is calculated which represents the possibility that a patient has the disease. 4.2. Calculation oJ" likelihood scores A cot!fidence measure (CM) is calculated separately for each slot in a context. T h e likelihood score for the context is a weighted sum o f the CM for all the slots in the context. T h e weighting is affected by the n u m b e r o f relet,ant slots in a context. T h e relevance o f a slot is disease-dependent. F o r example, the family b a c k g r o u n d o f the patient is not relevant for choledocholithiasis as shown in Fig. 3 and it is indicated by a not_appl value o f the slot. T h e CM for a slot is calculated from the C M s o f all the elements in the slot. T h e confidence measure for an individual slot element represents how much the patient profile satisfies the requirements o f that element o f the disease descriptor. T h e calculation o f individual CMs is tied to o u r use o f c o n t r i b u t i o n and absence factors to be described below. Recall that an element is either a single key tuple or a disjunction o f them. T h e CM calculation o f a key tuple is determined by the key type, and the requirements satisfied by the patient profile which is related to the c o n t r i b u t i o n and absence factors. Th e CM o f a disjunction o f key tuples is the largest CM o f one o f the disjuncts. If the M A T C H E R can find the manifestations defined for a key tuple that are expected to be present in a patient with a particular disease, then this key tuple is t, alidated. If only some o f the findings are existent, then this key is partially ~,alidated. When the patient profile is known not to have those findings or there is evidence against the presence o f the findings, then the key is im'alidated. The M A T C H E R considers the keys to be unknown if it can not find the related attributes from the patient profile in the case o f a direct key, or extract it in the case o f extractable keys. F u r t h e r details a b o u t the validation process are presented after the discussion o f the use o f c o n t r i b u t i o n and absence factors. 4.3. The role o f contribution and absence factors C o n t r i b u t i o n and absence factors are the essence o f the mechanism for reasoning under uncertainty in D I J E S T . A contribution factor (CF ) and an absence factor (AF) are defined for each key in exert, key tuple o f a slot in the DD. T h e C F determines the degree o f i m p o r t a n c e o f the presence o f the specific concept represented by the key n am e to the slot in which it occurs. It indicates the expectation that a patient has the specific disease when the i n f o r m a t i o n in his/her profile validates the requirements o f this key. T h e c o n t r i b u t i o n fact o r is defined as a real n u m b e r between 0 and I, inclusive. F o r example, the c o n t r i b u t i o n factor o f the direct key pain is 0.9 for DIJEST 133 choledocholithiasis as shown in Fig. 3. It shows that the presence o f pain as defined by its respective values is very i m p o r t a n t for choledocholithiasis. T h e A F determines the i m p o r t a n c e o f the absence o f the co n cep t in the patient profile. It effectively measures the likelihood o f a patient to have o r not to have a disease given the absence o f the key. It is represented on a scale o f ( - . ~ , I). T h e wide scale o f absence factors is used to influence the i m p o r t a n c e o f a specific key to the entire slot within which it is defined. F o r example, the absence factor o f pain is - 0 . 1 as shown in Fig. 3. C F values fo r a key are always greater than the A F values. T h e analysis to determine whether the patient has the disease depends purely on CFs and AFs. O u r scheme is similar to the scoring mechanism in P IP where the scores are given in the frames [I]. The CFs and AFs are actually the quan t i t at i v e representation o f the qualitative terms. such as "usually p r e s e n t " , " c o n f i r m i n g " , "'critical", "'more likely", "'less likely" and "'contra- dicting", that were used by o u r medical experts. Th e terms have been distributed on two different scales by using CFs and AFs. Each application has a base value, BV, which partitions the c o n t r i b u t i o n factors into two sets, those a b o v e the BV and those below it. The BV is used as a point o f reference for the distribution o f c o n t r i b u t i o n and absence factors o f the keys. F o r D I J E S T , a BV o f 0.5 was used. T w o principles underly o u r choice o f values for c o n t r i b u t i o n and absence factors from their respective scales for specific keys: • C F / > BV indicates that the key is i m p o r t a n t to establish that the patient has the disease under consideration. • A F < 0 indicates that the absence o f the key is i m p o r t a n t to c o n t r a d i c t that the patient has the disease under consideration. Confidence measure values are classified into four categories based on these two principles: I. C F > B V , AF~>O. These keys are confirming. A confirming key in the patient profile contributes significantly to the likelihood score. Its validation will lead to a high score. H o w e v e r even if the key is not validated, the disease m ay still figure p r o m i n e n t l y in the final differential diagnosis. 2. CF>~ BV, A F ( O . These keys are critical. Critical keys have the most impact on determining the likelihood score. T h e validation o f a critical key contributes to a high score. T h e invalidation o f a critical key co n t ri b u t es negatively to the score by using the AF. If a critical key is un k n o w n , a neutral position is taken. 3. CF ( B V . A F < O. These keys are contradicting. T h e validation o f a co n t rad i ct i n g key does not strongly confirm the existence o f the disease. T h e invalidation o f a c o n t r a d i c t i n g key can lead to a very low likelihood score. 4. C F < BV, A F i> O. These keys are minor. M i n o r keys are used for fine tuning the differential diagnosis and will play a greater role in the future screening process. This classification scheme a p p r o x i m a t e l y c o r r e s p o n d s to the following use o f eroking strength and frequency values in Internist's scoring mechanism. • Critical keys: eroking strength, 4 frequency 4. • C o n t r a d i c t i n g keys: et,oking strength, l frequency 4. * Confirming keys: et,oking strength, 4 frequency 2. • M i n o r keys: et'oking strength, 2 frequency I. Each element in a slot list is evaluated accord i n g to the a b o v e classification. T h e M A T C H E R determines how well the patient profile fits the structure that is determined for this element. Using the state o f the patient profile with respect to the attributes o f each element in this slot list and using the C F and A F factors, the m a t c h e r determines the CM o f this element. Repeating this iterative process, all CM values o f the elements in a slot list are accumulated and normalized by the unique normalization_factor for the slot. T h e overall sum o f the slots determines the score o f a particular context and then the likelihood score o f the disease. T h e matching process for the slot values is illustrated below. In the code, Context refers to the current name o f the context, Hypothesis refers to the n am e o f the disease currently investigated 134 L. 0. YAL~INALP and L. STERLING and PatientSlot has all the values that are currently k n o w n for the Patient for a particular slot, such as symptoms. T h e first clause illustrates that the slots which are not applicable are n o t skipped over, with the assumption that they are completely satisfied for probability calculations. satisfy_slots(Context, Hypothesis, Patient, PatientSlot, noLappl,1.0). satisfy_slots(Context, Hypothesis,Patient, PatientSlot, Slot,SlotProb) ,- Slot' = = not_appl, satisfy_slot (Context, Hypothesis, Patient, PatientSlot, Slot,Slot Prob,O). satisfy_slot (Context,Hypothesis, Patient, PatientSlot,[normalization_factor(N F)], SlotProb,AccProb) *- SlotProb is AccProb/NF. % normalize for a slot satisfy_slot (Context, Hypothesis, Patient, PatientSlot, [ Keyl Key List], Slot Prob,Acc Prob) ,-- Key = = normalization_factor(NF), satisfy_key(Context, Hypothesis, Patient, PatientSlot, Key,CM), accumulate (Acc Prob,C M,AccProbNext), satisfy_slot (Context,Hypothesis, Patient,PatientSlot, KeyList, SlotProb,AccProbNext). satisfv__key determines whether a key is a single key o r a disjunction o f keys. T h e code for processing single keys is given below. Th e find predicate extracts the values for a particular key from the patient profile. h a n d l e _ s i n g l e _ k e y ( C o n t e x t , H y p o t h e s i s , Patient, P a t i e n t S l o t , K e y , KeyValues,C F,AF, CM) ,--- i s_cl i r e c L k e y ( C o n text, Key, KeyVal u es), find (Key, PatientVals, P a t i e n t S I o t ) , d i r e c t _ k e y ( C o n t e x t , H y p o t hesis, Patient, PatientVals, Key, KeyValues, C F,AF, C M). hand•e-sing•e-key( C•ntext'Hyp•thesis•Patient•PatientS••t•Key'Key•a•ues'•F•AF'CM) ,-- n o t is_direct_key(Context, Key, KeyValues), extract_from (Context, Hypothesis, Patient. PatientSlot, Key, KeyValues, CF,AF,CM). hand•e-sing•e-key( C•ntext•Hyp•thesis•Patient•Patients••t•Key•Keyva•ues•CF•AF•CM) .-- base_value(BV), not_known (Context, Hypothesis, Patient, Key, KeyValues, BV, C F,AF, CM) direct_key( Context, Hypothesis, Patient, PatientVals, Key, KeyValues, CF,AF,CM). check_whether_absent(PatientVals), absent_key(Hypothesis,Patient,PatientVals, Key,KeyValues, CF,AF,CM), direcLkey(Context, Hypothesis, Patient, PatientVals, Key, KeyValues, CF,AF,CM) .-- check_whether_present(PatientVals), match_compare (Context, Hypothesis, Patient, PatientVals, Key, KeyValues, CF,AF,CM). 4.4 Calculating confidence measures for keys This subsection describes how the individual C M s are calculated for individual keytuples. Both direct keys and extractable keys are treated in detail. O u r description here is qualitative in nature. T h e exact formulae used can be found in Ref. [7]. The confidence measure o f a direct key is calculated t h r o u g h an extended c o m p a r i s o n o f the values in the key attribute list o f the D D with the patient values as shown below. T h e first stage is to calculate the patient sum. that is a score indicating how well the patient values match the attribute values. Patient sums are only calculated for keys which actually a p p e a r in the patient profile. match_compare (Context, Hypothesis, Patient, PatientVals, Key, KeyVals, C F,AF, C M ) ,-- compute_patient_sum (Context, Hypothesis, Patient, PatientVals,CF,AF, Key, KeyVals,O,PatientSum,Contradiction Flag), member (threshold (Threshold),KeyVals), find_normalization (KeyVals,Norm Factor), compute_key_prob(Contradiction Flag, PatientSum,Norm Factor,Threshold,CF,AF,CM). The M A T C H E R calculates patient sums as follows. First the terms used in the patient profile, which may be a mixture o f qualitative and quant i t at i v e terms such as 6 days, are converted to the DIJEST 135 d o m a i n d e p e n d e n t qualitative terms which are used in the D D s, fo r example short o r medium. T h e terms are then c o m p a r e d with the actual terms in the D D and exact matches an d c o n t r a d i c t i o n s are noted. T h e terms which exactly match are s u m m e d using weights which are given in the D D with respect to each attribute. This is illustrated with the co d e presented below. compute_patient-sum (Context, Hypothesis, Patient, PatientVals, CF,AF, Key, [threshold (T) ] ,TotalSu m,TotalSum,no). compute_patient_sum (Context, Hypothesis, Patient, PatientVals, CF,AF, Key, [ ElementIKeyVals] ,InterSum,NextSum,Contradiction Flag) ,-- Element'.,. = = threshold(T), match ( PrevContrad iction Flag, Context, Hypothesis, Patient, Key, PatientVals, Element, InterSum,TotalSum), check_contradiction (PrevContradiction Flag,Context, Contradiction Flag,Hypothesis, C F,AF, Key,TotalSum,NextSum). match (no,Context, Hypothesis, Patient, Key, PatientVals, Element, I nterSu m,AccSum) 4-- match-single_val(Hypothesis, Element, PatientVals,AttrContr), AccSum is InterSum + AttrContr. match (yes, Context, Hypothesis, Patient, Key, PatientVals, Element, lnterSum,AccSum) , - is_a_contradiction (Hypothesis, Patient, Key, PatientVals), record_contradiction (Context, Hypothesis, Patient, Key). match_single_val ( H ypothesis,site (Site, Contr),AIIValues, Contr) ,- member(site(Patsite),AIIValues), appropriate_.site(Hypothesis, Patsite). match_single_val (AnyConcept, Parameter,AIIValues, Contr) , - %generalized matching Parameter =.. [Name, ParVaI,Contrl, FindVal = .. [Name,SomeVal], member(FindVaI,AIIValues), match_from_tables(AnyConcept, Name, ParVaI,SomeVal). % Sample facts appropriate_site(choledocholithiasis, righLupper_quadrant). appropriate_site (choledocholithiasis,epigestrium). match_from_tables(_,duration, DAYS,short) *- number(DAYS), DAYS > =1, DAYS < 11. match_from_tables(_,duration, DAYS,moderate) , - number(DAYS), DAYS > 10, DAYS < 36. match_from_tables(_,duration, DAYS,long) ,- number(DAYS), DAYS > 35. Every direct key has a threshold, which is the m i n i m u m value o f the patient sum considered to adequately match the key. T h e second stage o f the M A T C H E R is to c o m p a r e the patient sum with the threshold set for this key. On the basis o f this c o m p a r i s o n , the M A T C H E R concludes whether the patient profile satisfies the a t t r i b u t e values completely, partially, o r c o n t r a d i c t s them, and calculates the C M accordingly. If the patient sum exceeds the threshold value, then we say that the direct key has been t'alidated. T h e CM value is this case is the C F value. F o r example, the at t ri b u t e values o f the patient in Fig. 4 indicates a sum o f 13 points. This is equal to the threshold value for this key, t h erefo re the direct key pain is validated for this patient. T h e C M is then set to 0.9. I f the patient sum is less than the threshold, and no c o n t r a d i c t i o n has been noted, the key has been partially z,alidated. T h e confidence measure is a normalized fraction o f the C F value. This is handled by compute-key_prob. M o r e details are in Ref. [7]. If a c o n t r a d i c t i o n has been noted, the value o f the CM differs depending w h et h er the absence factor o f the key is positive or negative. If the A F is negative, it is returned as the CM. Otherwise the CM is the negative o f the C F value. This is handled by check_contradiction. 136 L. I~. YAL(~INALP and L. STERLING We describe each o f the four categories o f extractable keys in turn, where the key appears in the patient profile: (i) Special-purpose knowledge is used to handle the an at o m i cal o r physiological states that are indexed as a key, such as c o m m o n bile duct o b s t r u c t i o n as in Fig. 3 o r swelling o f the pancreas. Some sample facts are given below. extract_from(Context, Hypothesis, Patient, PatientSlot, Key,KeyValuas, CF,AF,CM) ,- anatomy(Key), anatomy_test (Context, Hypothesis, Patient, PatientSIot, Key, KeyValues,CF,AF,CM). anatomy (Key) ,- organ (Key). anatomy(Key) ,-- system (Key, SystemComponents). anatomy (Key) ,- system (Sys, SystemComponents), part_of (Key,SystemComponents). system (intra hepatic_ducts, [left_intra hepatic_duct,rig ht_intra hepatic_d uct] ). system (extra hepatic_ducts, [common_bile_duct,cystic_duct,pancreatic_duct] ). anatomy_test (Context, Hypothesis, Patient, PatientSIot,Organ,present,CF,AF,CF) ,- organ(Organ), surgery(Patient,SurgeryList), not taken (SurgeryList,Organ). anatomy_test (Context, Hypothesis, Patient, PatientSIot,TestContext, (Specification,FacttoDetermine),C F,AF,CM) ,- test_illustrates (TestContext, Specification, ListofTests), prioritize(ListofTests,FinalTests), patient_satisfies (Hypothesis, Patient, PatientSIot,TestContext, Specification.FacttoDetermine, FinalTests, CF,AF,CM). F o r each state, the set o f relevant tests is determined along with their o rd er o f preference. T h e representation o f anatomical knowledge in D I J E S T has been designed to allow the M A T C H E R to find the necessary tests that would indicate the presence o f the specified state. F o r example, the M A T C H E R finds that ultrasound and C T tests are indicative for understanding the co n d i t i o n o f the c o m m o n bile duct when checking choledocholithiasis [7]. After the necessar,v tests are found, it is determ i n ed whether the patient has taken the test. If he has not, the C M for this key is calculated using the C F and A F values, and varies depending in which o f the four categories the C F and A F values lie. I f the patient has taken the test, d o m a i n specific knowledge is used to determine whether the patient's test results satisfy the specified state. If so, the CM is set to the CF. Otherwise, the C M is eq u at ed to A F because a conflict exists between the expected condition o f the patient and the patient profile. T h e r e is no possibility to partially validate these keys. F o r example, the results o f the u l t raso u n d for the patient in Fig. 4 are c o m p a r e d with the expected o u t c o m e s for the key c o m m o n bile duct. F o r the test results [7], it is found that the c o m m o n bile duct o f the patient is very dilated. T h e r e f o r e , CM is eq u at ed to 0.9. Th e ultrasound also shows there are gallstones in the gallbladder. CM for this key is set to 0.8. Planning optimal order o f tests, prioriti-e, is a complicated issue, and could be the d o m a i n o f a n o t h e r expert system that would p e r f o r m in parallel to D I J E S T . Currently, the tests are checked in sequential order. Studies in decision analysis for developing clinical strategies similar to the one for the diagnosis o f extrahepatic obstructive jau n d i ce can be useful for the d ev el o p m en t o f this module. Especially, the sensitivity, specificity, complications and the cost o f the individual tests have been investigated to devise different adaptive strategies for tests taking, represented as decision trees in Ref. [8]. We have used a~ailability as o u r criteria for ordering. (ii) Keys referring to blood tests, such as amylase and bilirubin, are evaluated using possibility distribution curves which are graphs provided to us by o u r experts. First the M A T C H E R checks whether this is a key that requires curve fitting analysis by seeing whether a patient has taken the particular test. If not, the calculation o f the C M is carried out by considering the four classes o f C F and A F values as for the anatomical states. If the patient has the test, the patient value is checked by a disease specific possibility distribution curve, where each curve estimates the DIJEST 137 likelihood that a patient with the particular test value has the disease being considered. The resulting possibility value is used along with the CF and AF to determine the CM of this key. For example, if the patient's test result shows a particular positive possibility, this value is used to normalize the CF specified for this key. Normalization is needed since the importance of this test result is specified with the CF, and how well the patient's result fits the expected value for the disease is determined by the curve. If the patient's test result contradicts the presence of the disease, invalidating the key, then the full AF value is used as the CM value. Curve fitting is actually not very suitable with Prolog if speed and accuracy is required. It should be implemented as an external procedure. extract_from (Context, Hypothesis, Patient, PatientSlot, Key, KeyVals, CF,AF, CM) ,-- possibility_curve( Key), curve_fitting (Hypothesis, Patient, PatientSIot, Key, KeyVals,C F,AF, CM). possibility_curve(Key) ,-- blood_test(Key). curve_fitting(Hypothesis, Patient, PatientSIot, Key, KeyVals, eF,AF, e M ) ,-- blood_test_analysis( Hypothesis, Patient, PatientSIot, Key, CF,AF, CM ). blood_test_analysis(Hypothesis, Patient,PatientSIot, BloodTest, CF,AF, CM) ,-- (get_patient_val(BloodTest,serum,PatientSIot, Result); get_patient_val(BloodTest, blood,PatientSIot, Result)), blood_test (BloodTest, Hypothesis, Result, Prob), calculate_CM (Prob. Hypothesis, Patient, BloodTest, eF,AF, e M ) . (iii) Recall that compound keys refer to a collection of findings, for example prodrome. Their analysis requires the MATCHER to consider each finding in the collection similar to the consideration of each attribute of a direct key. Each finding for compound keys, though, has to be analyzed separately similar to an element of a slot. The collected result of all the findings determines the overall CM for this key. extract_from (Context, Hypothesis, Patient, PatientSIot, Key, KeyValues, C F,A F,C M ) ,- c o n c e p L t a b l e ( C o n t e x t , Key, Keyeoncepts), satisfy_concept (Context, Hypothesis, Patient, PatientSIot, Keyeoncepts, Prob), concept_prob(Prob,CF,AF, CM). The sum of all the confidence measures of the findings that are related to this key is denoted CMs. CMs is tested with respect to an interval [0,Threshold) where the value of the threshold for compound keys is application-dependent. If CMs lies within this interval, the presence of a finding can be neither validated nor invalidated, and is considered to be unknown. If the value is to the left of this region, the finding is invalidated and the overall CM is set to the AF. Otherwise, it is considered to be fully validated and the overall CM is set to the CF. This is handled by concept _prob. (iv) The rule names that are used within key tuples are evaluated by activating each rule, for example for li~'er tests and obstructit,e tests. These rules, which represent for example a group of tests, need to be evaluated considering domain specific dependencies of the tests. Each rule is interpreted separately and the CM calculation varies for each. Default behavior if the patient has not taken the test is similar to the default behavior for anatomical states and blood tests. For example, the rule obstructit'e tests in Fig. 3 is activated for the patient in Fig. 4. The values of the tests of this patient is found to be sufficient for this rule. Therefore, the CM for this key is set to the CF value, which is 0.7. extract_from (Context, Hypothesis, Patient, PatientSIot, Key, KeyValues, C F,AF, C M ) ,-- call_proc ( [ Key, Context, Hypothesis, Patient, PatientSIot, KeyValues, C F,A F,C M ] ). / * C A L L A N Y PROCEDURE PASSED AS P A R A M E T E R ' / call_proc([ProcNamelList]) , - Proc = .. [ProcNamelList],Proc. The MATCHER has a default behavior for evaluating keys which are not covered by the above discussion, for example a direct key in the DD which does not appear in the patient profile, or a compound key for which no information is known. The CMs of these keys are determined with 138 L. CI. YAL(;INALP and L. STERLING respect to the f o u r categories o f C F a n d A F values. T h e crucial categories o f critical keys a n d c o n t r a d i c t i n g keys are chosen so as not to c o n t r i b u t e to the overall sum. T h e confidence m e a s u r e C M is calculated as follows: n o L k n o w n (Context, Hypothesis, Patient, gey, KeyVals, BV,CF,AF,CM) ,-- % minor keys CF < BV A F > = 0 , CM is(CF + AF)/2. not_known (Context, Hypothesis, Patient, Key, KeyVals, BV, C F,A F,0) ,- % critical keys CF > = BV, A F < 0 , record_question ( [ Context, H ypothesis, Patient, Key, KeyVals] ). not_known (Context, Hypothesis, Patient, Key, KeyVals, BV,CF,AF,0) ,- % contradicting keys A F < O , CF < BV record_possible_contra (Context, Hypothesis, Patient, Key). not_known( Context, Hypothesis,Patient, Key, KeyVals, BV, CF,AF,AF) ,- % Confirming keys A F > = 0 C F > = BV, record_u nknown ( Context, Hypothesis, Patient, Key). F o r e x a m p l e , the exact location o f the o b s t r u c t i o n c a n not be d e t e r m i n e d by u l t r a s o u n d for the patient in Fig. 4 [7]. This key is a critical key. T h e r e f o r e , the C M value is set to 0 by the default values as described a b o v e . 4.5. Orerall likelihood score T h e o~erall likelihood score o f a slot Ls~o,, as m e n t i o n e d earlier, is the sum o f the C M for each key a n d n o r m a l i z e d by the specific n o r m a l i z a t i o n f a c t o r o f the slot. T h e n o r m a l i z a t i o n factor, N F , is defined as follows ~ h e r e n is the n u m b e r o f elements in a slot. We a s s u m e that not all absence f a c t o r s are zero. " ,~AF,, A F >/0, N F = E f = , -I.CF,, A F < 0 . T h e weighted sum, WS,, can be defined as the best case where all the elements o f the slot i is validated. T h u s , WS, = E~'= t C F , . T h e r e f o r e , N F , ~< WS,. With this relation, the n o r m a l i z a t i o n helps to increase the c o n t r i b u t i o n o f slot i to the likelihood o f the overall context. T h e score might be g r e a t e r t h a n I with d a t a that c o n f i r m s all the expected values o f a slot in a disease descriptor. T h e overall likelihood o f a c o n t e x t is thus defined as N \ T Ls o~, L c = J=l N V where N V equals the n u m b e r o f valid slots in a context. Let us illustrate this calculat,on b~ using the e x a m p l e disease in Fig. 3 a n d the patient in Fig. 4. I f the l a b _ t e s t s slot is considered, it is seen that the n o r m a l i z a t i o n _ f a c t o r . 2.7 is calculated as described above. T h e calculation o f C M s for each o f the keys in this slot is illustrated in Section 4.3. Respective b , the~ are 0.7, 0.8, 0 a n d 0.9. T h e sum o f these C M s is 2.4. Using these values, Lsk,,, is set to 0.89. Since there is only one slot in this context, L~b_, .... is equal to 0.89. 4.6. The patient anaO,sis When the M A T C H E R calculates the likelihood scores o f a disease, special i n f o r m a t i o n related to the patient with respect to each disease is recorded a l o n g with the likelihood scores, T h i s D I J E S T 139 information is used to produce an evaluation report a b o u t the status o f a patient. It consists o f the list o f findings which are expected but not present in the patient data, which are contradictory to the evaluated disease, and the important concepts which have not been validated during the analysis o f the M A T C H E R . The findings o f the evaluation are divided into four categories, questions, contradictions, possible contradictions and unknowns. To record this information, again the four categories o f C F and A F values are used. The code in the previous section is suitably adapted. The evaluation report can be used to guide the subsequent stages o f clinical diagnosis in the screening process shown in Fig. 2. For example, the missing necessary tests to check a specific condition that have not been performed are suggested by questions for a disease. Contradictions are the set o f facts in the patient profile that contradict the existence o f the disease. Possible contradictions are the unknown classes o f information which might be critical. They can contradict the disease if their definite absence is proven. U n k n o w n s is the category o f data that can be used for confirmation but are u n k n o w n at the time o f evaluation, 5. P E R F O R M A N C E O F D I J E S T The development time for D I J E S T was a b o u t nine m o n t h s including our learning a b o u t aspects o f jaundice, the diseases and the related a n a t o m y and the physiology. The knowledge represen- tation scheme and the uncertainty reasoning mechanism reflect our perception o f medical concepts and clinical reasoning provided by our experts. D I J E S T has been tested with cases taken from medical text books and real patient records. For example, Table I shows a differential diagnosis produced by DIJEST for a patient with choledocholithiasis. The medical history o f this patient is shown by Fig. 4. During testing, the evaluation o f all the d o m a i n diseases were included. In clinical use, a threshold m a y be used to inhibit unlikely diseases. The analysis shows that choledocolithiasis is given the highest likelihood score by DIJEST, even though it does not get the highest score in each context. The score for acute cholecystitis shows the way large absence factors can prevent a disease from being considered seriously as explaining the jaundice. The scores from the contexts o f clinical examination and lab tests strongly suggest that cholecystitis could explain the jaundice, more so than choledocolithiasis, but the patient's history strongly contradicts the disease. The evaluation report o f this patient points out for example the lack o f information about critical findings o f hepatitis, such as the presence o f a prodrome, or the exposure to the use o f needles in the past. The evaluation report is not shown here. Later on in the course o f the disease, the same patient contracted pancreatitis, directly caused by the choledocholithiasis. We added new test results to the patient profile and re-ran DIJEST. The result o f the second differential diagnosis is given in Table 2. The only changed scores are o f those diseases related to the pancreas. Note especially that the likelihood score o f pancreatitis has significantly increased. Table 2 demonstrates the ability o f DIJEST to cope with multiple diseases. Knowledge is still necessary, for example, to realize that hepatitis and choledocholithiasis do not in general co-exist, whereas choledocholithiasis m a y cause pancreatitis. Such reasoning, which would form part o f the screening process, allows us to place more significance on the score for pancreatitis than for hepatitis even though it is actually marginally lower. Table l Table 2 Likelihood Scores for Patient I0001 Disease History Clinical choledocholithissis i .00 0.84 viral hepatitis -0.05 0.99 hepatitis -0.75 0.99 acute cholecystitis -1.50 1.17 pancreatitis 0.28 0.20 pancr, pseudo cyst 0.16 0.00 cirrhosis 0.89 0.90 paJncreatic cancer -0.50 0.17 Tests Total Score 0.89 0.91 0.81 0.58 1.00 0.41 1.08 0.25 0.21 0.23 0.13 0.10 - 1 . 5 0 0.03 - 0 . 8 7 - 0 . 4 0 II II Likelihood Scores for Patient 10001 Disease History Clinical Tests Total Score choledocholithiasis 1.00 0.84 0.89 ().91 viral hepat, itis - 0 . 0 5 0.99 0.81 0.58 pancreatitis 0.28 0.20 1.06 0.51 hepatitis -0.75 0.99 1100 0.41 acute cholecystitis - 1 . 5 0 1.17 1.08 0.25 cirrhosis 0.69 0.90 - 1 . 5 0 0.03 pancr, pseudo cyst 0.16 0.00 - 0 . 3 0 - 0 . 0 5 pancreatic cancer - 0 . 5 0 0.17 - 1 . 2 3 - 0 . 5 2 140 L. I~. YAL~INALP and L. STERLING D I J E S T was i m p l e m e n t e d by using P r o l o g c o n s t r u c t s which are s t a n d a r d in a l m o s t all Prologs. It currently runs under Sicstus a n d Q u i n t u s Prologs. In terms o f speed, p r o d u c i n g a table such as a b o v e a n d the e v a l u a t i o n r e p o r t takes only a few seconds on the average. 6. C O N C L U S I O N S T h e features o f D I J E S T in its c u r r e n t state can be s u m m a r i z e d as follows. Medical k n o w l e d g e is represented declaratively. T h e d o m a i n specific knowledge a n d d o m a i n specific r e a s o n i n g is clearly distinguished f r o m d o m a i n i n d e p e n d e n t k n o w l e d g e by the M A T C H E R by using different types o f keys. T h e c o m p l e x medical k n o w l e d g e related to the diseases, the characteristics o f different testing p r o c e d u r e s a n d the basic a n a t o m i c a l a n d physiological structure o f the b o d y are all represented i n d e p e n d e n t l y o f p a t i e n t i n f o r m a t i o n a n d illustrate characteristics o f jaundice. Using P r o l o g e n a b l e d us to reach o u r objective, to have this s e p a r a t i o n and write a specialized interpreter very easily. T h e interpreter has also been generalized to handle d o m a i n s o t h e r t h a n D I J E S T by c u s t o m i z i n g the general m a t c h i n g capabilities o f the interpreter. R e p r e s e n t i n g the likelihood estimates by using two s e p a r a t e factors, c o n t r i b u t i o n and absence factors, can distinguish between valid, invalid, u n k n o w n and a b s e n t data. D I J E S T presents very realistic likelihood estimates o f the presence o f the c a n d i d a t e diseases by e v a l u a t i n g the patient profiles, which m a y be incomplete. O f special i m p o r t a n c e is the calculation o f likelihood scores o f the individual c o n t e x t s a n d their effect on the final diagnosis. D I J E S T also e m p h a s i s e s significant f a c t o r s in the e v a l u a t i o n o f each disease. C o n t r a d i c t o r y findings a n d i m p o r t a n t d a t a which m a y be required for further e v a l u a t i o n o f the patient are noted. D I J E S T is very p r o m i s i n g in the early detection o f co-existing diseases in a patient a n d provides g o o d likelihood estimates in the cases with multiple diseases. T h e m o s t difficult task in D I J E S T is to o b t a i n the c o n t r i b u t i o n a n d absence factors for different keys. Especially, representing the experts" qualitative view o f the subject by using those factors needs successive e x p e r i m e n t s a n d a d j u s t m e n t . A w e a k p o i n t o f D I J E S T is its neglect o f unexplained factors that are c o n t a i n e d in the patient profile. T h e presence o f a screening process for presenting the results o f M A T C H E R in a user-oriented m a n n e r a n d for r e m o v i n g r e d u n d a n t i n f o r m a t i o n would e n h a n c e the p e r f o r m a n c e o f D I J E S T . T h e consistency checking is also only partially complete. At this stage, however, D I J E S T is e n c o u r a g i n g in its expressive p o w e r for medical knowledge a n d by p r o v i d i n g useful likelihood estimates to indicate the presence o f d o m a i n diseases. It has p o t e n t i a l for detecting the co-existence o f multiple diseases. It is unique in b o t h its knowledge r e p r e s e n t a t i o n scheme a n d r e a s o n i n g with uncertainty. .4cknowledgements--We would like to thank our medical expert, Professor David Ransohoff, for providing the medical knoaledge embodied m DIJEST. We are grateful for his ~aluable time in attending the knowledge engineering sessions and ackno~ledge his influence in shaping DIJEST. Drs Arnold Shmerling and Lawrence Widman also supplied valuable medical insights. Dr Len SamueLs commented on an earher draft and pro,,ided the correspondence between contribution and absence factors of DIJEST and the scoring mechamsm of Internist. We also thank Yuval Lirox for inviting us to submit this paper to the specml isssue of Computers & Mathematics with Apphcations. R E F E R E N C E S S. Pauker, A. Gorry. J. Kassier and W. Schwartz, To,~ards the simulation of clinical cognition: taking a present illness by computer. 4m. d. Med. 60, 981-996 (1976). 2. B G. Buchanan and E. Shortliffe. Rule Based Expert Systems The 31 YCIN Ex'perlments o f the Stanford Programming Prolect. Addison-Wesley, Reading. Mass. (1984). 3. R. Miller, H. Pople and J D. Meyers, Internist I, An experimental computer based diagnostic consultant for general mternal medicine. New Engl. J, Med. 307, 468-476 (1982). 4. F. E. Mesarie, R. Miller and J. D. Meyers, Intermst-I properties: representing common sense and good medical practice in a computerized medical knowledge base. Computers Biomed. Res. 18, 458~,79 (1985). 5. P. Szolovitz and S. Pauker, Categorical and probabilistic reasoning in medical diagnosis. Art!L Intell. II, 115-144 (1978), 6. Harrtson's Principles o f Internal Medicine, I Ith edn. McGraw-Hill, Net York (1987). 7. L I~I. Yalqmalp. Uncertainty reasomng in a medical expert system: DIJEST M.S. Thesis, Department of Computer Engineering and Science, Case V~estern Reserve University, Cle~,eland, Ohio, (1987). 8. J. Richter, M. Silverstein and R. Shapiro. Suspected obstructive jaundice, A decision analysis of diagnostic strategies. Ann. Internal Med. 99, 46-51 (1983). 
work_qqpy7vvyarctvk32abryvwciwe	----	Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 Exemplar Driven Development of Software Product Lines Ruben Heradio ∗1, David Fernandez-Amoros†1, Luis de la Torre‡1, and Ismael Abad§1 1ETS de Ingenieria Informatica,Universidad Nacional de Educacion a Distancia, Madrid, Spain Abstract The benefits of following a product line approach to develop similar software systems are well docu- mented. Nevertheless, some case studies have revealed significant barriers to adopt such approach. In order to minimize the paradigm shift between conventional software engineering and software product line engineering, this paper presents a new development process where the products of a domain are made by analogy to an existing product. Furthermore, this paper discusses the capabilities and limita- tions of different techniques to implement the analogy relation and proposes a new language to overcome such limitations. 1 Introduction Software Product Line (SPL) engineering has become an important and widely used approach for the effi- cient development of whole portfolios of software products [1, 2]. The fundamental idea of the approach is to undertake the development of a set of products as a single, coherent development activity. Although the benefits of using a SPL in matter of quality, productivity and time-to-market are well documented [3, 4, 5], some case studies have revealed significant barriers to adopt the SPL approach. For example, in its successful Diesel Engine SPL, Cummins stopped all product deployments for six months [6]. As Krueger argues [7], many organizations can not afford to slow or stop production for six months, even if the poten- tial return of investment is huge. In order to minimize the paradigm shift between conventional software engineering and SPL engineering, Krueger identifies three prominent adoption models: 1. The proactive approach, also named big bang approach [8], is like the waterfall approach to conven- tional software: all product variations on the foreseeable horizon up front are analyzed, designed, and implemented. 2. The reactive approach is like the spiral or extreme programming approach to conventional software: one or several product variations on each development spiral are analyzed, designed and imple- mented. 3. The extractive approach reuses one or more existing software products for the product line initial baseline. ∗rheradio@issi.uned.es †david@lsi.uned.es ‡ldelatorre@bec.uned.es §iabad@issi.uned.es 1 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 The reactive approach is appropriated when it is difficult to predict the requirements for product variations or if organizations must maintain aggressive production schedules with few additional resources during the transition to a product line approach. The extractive approach is very effective for an organization that wants to quickly transition from conventional to SPL engineering. In this paper, we propose the Exemplar Driven Development (EDD) process to develop SPLs, which adopts reactive and extractive approaches. EDD takes advantage of the similarities among domain prod- ucts to make them by analogy. The starting point of EDD is any domain product built using conventional software engineering. Implicitly, this exemplar implements the intersection of all the domain product re- quirements. Next, the exemplar is flexibilized to satisfy the domain variable requirements that are out of the intersection. That is, an analogy relation is defined in a formal way to derive products automatically from the exemplar. The result of the exemplar flexibilization is a Domain Specific Language (DSL) compiler [9, 10], which is used during application engineering to get the products automatically. EDD is extractive because reuses existing exemplars as product line initial baseline and is reactive because proposes do- main engineering [11] as an incremental activity where flexibilization layers, which implement variable SPL requirements, are added to the exemplar in successive development cycles. Furthermore, this paper shows how to implement the exemplar flexibilization. Code Generation is an increasing popular technique for implementing SPLs that produces code from abstract specifications written in DSLs [12, 13, 14]. The next paradox usually comes up when a DSL compiler is developed: a DSL is a specialized, problem-oriented language. From the point of view of the DSL user, it is interesting that DSL is as abstract as possible (supporting the domain terminology and removing the low-level implementation details). On the other hand, from the point of view of the compiler developer, the DSL abstraction makes harder to build the compiler. That is, the further DSL specifications are from the final code, the more difficult is to transform them into final code. We propose to solve such paradox by taking advantage of a common property to all DSL compilers: the similarities among the final products (i.e., domain product commonalities are the main reason to develop the products jointly as a family, instead of one by one). Instead of synthesizing the final code from scratch or transforming a distant input specification, we suggest to obtain the final products adapting a previously developed domain product to satisfy the input DSL specifications: the exemplar. Figure 1 illustrates this approach, where the generator of a DSL compiler is another compiler which is used to adapt an exemplar according to the DSL source specifications. The figure also represents a possible decomposition of this subcompiler into subgenerators responsible of different sorts of variability. Template languages, such as XPand of openArchitectureWare1[15] or XVCL [16], implicitly use this ap- proach, since a text template can be viewed as a piece of an exemplar with slots. The exemplar code that is common to all the domain products is maintained in the template, whereas the variable code is replaced by slots, that are filled with metacode which specifies how code must change. Unfortunately, code and metacode are strongly coupled in templates. Indeed, as argued in [17], some domain variability should be implemented as crosscutting concerns. When a template engine does not support Aspect Ori- ented Programming (AOP) [18, 19, 20], templates may suffer metacode tangling (multiple variable concerns implemented simultaneously in a template) or metacode scattering (a variable concern implemented in multiple templates). To overcome the templates coupling problem, the metacode should be kept out of the exemplar code. In this case, the exemplar might be processed at: • lexical level, using regular expressions [21]. Nevertheless, though regular expressions can manage text in an agile way, they have serious limitations because they are internally implemented as state machines without memory and cannot manage nested or balanced constructs [22]. • syntactical level, using a metaparser such as ANTLR [23] or a transformation language such as Strat- 1see “openArchitectureWare User Guide Version 4.3.1”, available at http://www.openarchitectureware.org 2 http://www.openarchitectureware.org Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 Figure 1: DSL compiler based on the transformation of a domain exemplar ego [24] or Tom [25]. However, in most cases the simplicity of the exemplar changes does not justify to waste time neither defining the exemplar language grammar nor working with Abstract Syntax Trees (ASTs) [26, 27]. We propose an intermediate solution, the Exemplar Flexibilization Language (EFL) [28], that provides new operators to overcome the regular expressions limitations. EFL also supports the integration with parsers to manage marginal complex exemplar modifications. Besides, EFL supports the implementation of crosscutting generators, that manage variability scattered over the exemplar, and the decomposition and combination of generators. The remainder of the paper is structured as follows. Section 2 summarizes EDD. Section 3 describes EFL. Section 4 exemplifies the flexibilization capabilities and limitations of different techniques that are currently used to generalize code, and how EFL overcomes such limitations. Finally, section 5 sums up the conclusions of our work. 2 Exemplar Driven Development The decision of building a family of systems by using a SPL approach is usually taken when repetitive work is detected in a domain or when business opportunities are identified in the extension of a successful product. Therefore, when the SPL development starts there is often available an exemplar of the domain. Figure 2 depicts EDD, which tries to maximize the reuse of this exemplar applying intensively the idea of analogy in all the domain engineering activities. 1. Domain analysis. EDD domain analysis is based on the exemplar analysis. Since mandatory features, common to all domain products, are implemented by the exemplar, domain analysis is focused on 3 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 identifying the variable features, looking for the differences between the exemplar requirements and the requirements of the remaining products. 2. Domain design. EDD derives SPL architecture from the exemplar architecture. Domain design spec- ifies what adaptations of the exemplar design are necessary to transform it into any other product design. 3. Domain implementation. The exemplar is flexibilized to provide its automatic adaptation to satisfy input DSL specifications. The implementation techniques to flexibilize the exemplar should support the next desirable capabilities: (a) Non–invasiveness. Figure 3 shows how EDD integrates with the Boehm’s spiral lifecycle model [29]. In the first development cycle, an exemplar is built using conventional software engineering. In the next successive cycles, flexibilization modules are added to the exemplar in order to introduce the domain variability. When flexibilization modules are not invasive to the exemplar, there is no manual modification of it, facilitating the SPL evolution. (b) Crosscutting flexibilizations. As argued in [17], some domain variability should be implemented as crosscutting concerns. In the EDD context, the case where one flexibilization module adapts many modules of the exemplar implementation should be supported. (c) Applicable to any kind of software artifact. It should be supported the flexibilization of the exemplar documentation, test cases... (d) Run–time efficient management of the variability. For example, providing the parametrization of the inter-product variability before the products runtime. In order to implement the exemplar flexibilization, different techniques commonly used to generalize code may be used. Such techniques can be classified as internal and external. 1. With internal techniques the flexibilization is implemented using the mechanisms available in the language where the exemplar is written (e.g., using inheritance, genericity, aspects...). 2. With externals techniques the flexibilization is implemented using a different language or tool. In section 4, we will show how to flexibilize an exemplar using some internal and external techniques typically used to generalize code, identifying their limitations. To overcome such limitations we propose EFL, which is described in the following section. 3 Exemplar Flexibilization Language A technique for quickly developing a DSL interpreter is embedding it into a dynamic general purpose language [9]. This way, all the host language capabilities are implicitly available from the DSL. However, the pay-off is that the DSL concrete syntax has to fit in the host language concrete syntax. EFL is currently implemented applying this technique: it is a library of the Ruby object oriented language. EFL is freely available at http://rubyforge.org/projects/efl. As we will see, thanks to Ruby’s extensibility, the EFL concrete syntax is reasonably usable. 4 http://rubyforge.org/projects/efl Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 Figure 2: Overview of the EDD process Figure 3: EDD integration with the Boehm’s spiral lifecycle model 5 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 3.1 Defining Generators Figure 4 shows a simplified EFL metamodel. EFL supports the writing of generators that transform input exemplar files into output final product files according to input DSL specifications. EFL generators are written as Ruby classes that extend from the Generator class. This way, generators can be easily reused by mean of the Ruby composition and inheritance capabilities. Alternatively, the following syntactic sugar to write generators as objects of the Generator class is also available: my_generator = generator { << generator definition >> } A generator definition is composed of substitutions, productions and generations: 1. A substitution describes the interchange of an exemplar code pattern, expressed with a regular ex- pression, to new code. Due to EFL is embedded in Ruby, regular expressions are written in the Ruby notation (delimited with the / symbol). For instance, my_regexp = /code/. Crosscutting generators often apply the same substitutions over different exemplar files. To avoid the repetitive writing of substitutions and support their reuse, substitutions are independent from the exemplar files and the final product files. The main Generator methods to define substitutions are: • sub(reg_exp, text, name = nil) Optionally, substitutions, productions and generations can be named using the name string. • gsub(reg_exp, text, name = nil) A local substitution (sub) expresses the interchange of the first occurrence of the reg_exp regular expression to the text string. A global substitution (gsub) expresses the interchange of all the reg_exp occurrences. Additionally, the Generator class provides the next methods: • del and gdel to delete code from the exemplar. • before and gbefore to insert code before the reg_exp occurrences. • after and gafter to insert code after the reg_exp occurrences. Besides, the EFL substitution capabilities can be easily extended adding the correspondending meth- ods to the Generator class. 2. A production describes the application of a substitution list to an exemplar file to produce a final product file. Generator provides the next method to define productions: • prod(input_file, output_file, sub_list = nil, name = nil) The order of the substitutions in sub_list is irrelevant. If the sub_list is not specified, it will contain implicitly all the substitutions defined before the current production. EFL supports the detection of undesirable overlaps among the code patterns of the sub_list substi- tutions. 3. A generation executes a list of productions. Generator provides the next method for generations: • gen(prod_list = nil) The order of the productions in prod_list is irrelevant. If prod_list is not specified, it will contain implicitly all the productions defined before the current genera- tion. EFL supports the detection of undesirable collisions among the productions of a generation. 6 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 Figure 4: Simplified EFL metamodel 3.2 Combining Generators For writing complex exemplar transformations, EFL provides the next binary operators to combine two generators g1 and g2: 1. Sequence. Executes g1 first and g2 later: g1.gen g2.gen 2. Add. Returns a new generator which substitutions and productions are the union of the substitutions and productions of g1 and g2: (g1 + g2).gen 3. Superposition. Updates the substitutions and productions of g1 with the substitutions and produc- tions of g2. Those with the same name are overwritten and the remaining ones are added: (g1 << g2).gen Moreover, these operators can be combined among them. For example, you can write: ((g1 << g2) + g3 + g4).gen 3.3 EFL Capabilities to Overcome the Regular Expressions Limitations 3.3.1 The Zoom Operator There are two main types of regular expressions engines: the Deterministic Finite Automaton (DFA) and the Nondeterministic Finite Automaton (NFA). Being irrelevant for DFA engines how the regular expressions are written, the behaviour of NFA engines, however, depends on the representations of the regular expres- sions (e.g., a NFA engine follows different ways to match the equivalent regular expressions regexp1 = /to(ni(ght|te)|knight)/ and regexp2 = /tonite|toknight|tonight/ against the “tonight" string). According to Friedl [30], most of the programming languages implement NFA engines because give more 7 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 control to the programmer, since the representation of a regular expression sets the way the NFA engine backtracks during the matching resolution. Besides, NFA engines provide interesting features, such as capturing parentheses and the associated backreferences ($1, $2...), and lazy quantifiers. Writing a complex and time-efficient regular expression for a NFA engine may be quite hard. To simplify this work, EFL provides the zoom operator (>) that supports the step-by-step writing of regular expressions. Thanks to this operator, regular expressions can be chained to progressively specify a text pattern; i. e., the expression: regexp1 > regexp2 > regexp3 > ... > regexpN matches the regexp2 against the text matched by the regexp1, the regexp3 against the text matched by the regexp2, etcetera. 3.3.2 Anti-patterns Sometimes, it is useful to express a pattern in negative terms: instead of specifying the features we are interested in, describing characteristics to exclude some matching candidates. To support the writing of such anti-patterns, many regular expression engines provide the next constructs: • The negated character class [^...], which matches any character that is not listed into the character class. • The negatives look-ahead (?!...) and look-behind (?<!...) These look-around constructs do not “consume" any text, but they look forward or backward to “see" if their subexpressions cannot be matched. For example, the evaluation of the /Ruben (?!Heradio)/ regular expression against the “Ruben Garcia" string only matches the text “Ruben" (i.e. the negative look-ahead queries if anything different of “Heradio" follows “Ruben", but does not consume “Garcia"). Unfortunately, Ruby does not support the negative look-behind construct. EFL provides two new constructs for writing anti-patterns: 1. Complement (o) is an unary-operator that inverts the matching of a regular expression. That is, o(regexp1) > regexp2 matches the regexp2 out of the text matched by the regexp1. 2. Minus (-) is a binary-operator that excludes candidates for matching. That is, regexp1 - regexp2 captures the text that is matched by the regexp1 but not matched by the regexp2. Sometimes, it is quite hard “to find the precise regular expression", general enough to match all the text of interest and particular enough to ignore the rest. Using the minus operator, this problem can be solved in several steps: first, a more general regular expression is written without worrying about catching some undesirable text and, then, the matching is progressively adjusted by subtracting one or more particular regular expressions. Figure 5 illustrates several examples of the zoom, complement and minus operators. The top row shows several regular expressions built combining these operators and the bottom row highlights the result of matching the regular expressions against a given text. 3.3.3 Managing Nested Constructs As it was mentioned in the introduction, regular expressions cannot actually manage nested or balanced constructs because they are internally implemented as state machines without memory. For example, a regular expression for matching any number of balanced parentheses cannot be written, because when the 8 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 Figure 5: Examples of the zoom, complement and minus operators state machine finds the first close-parenthesis, is not able to “remember" how many open-parentheses has processed before. However, it is possible to write a regular expression for matching until a fixed number of balanced parenthesis. For example, the next three regular expressions match until one, two and three balanced parentheses respectively: 1. /[(]([^()])*[)]/ 2. /[(]([^()] | [(]([^()])*[)] )*[)]/ 3. /[(]([^()] | [(]([^()] | [(]([^()])*[)])*[)])*[)]/ Writing a /[(] ...[)]/ regular expression for each particular case is quite hard and repetitive. For- tunately, this work can be automatized using the Ruby meta-programming capabilities. For example, the nested_parentheses function in Figure 6 receives a levels number of balanced parentheses and gener- ates the corresponding regular expression. For example, a regular expression for ten balanced parentheses would be obtained with nested_parentheses(10). Internally, this method makes a string that contains the Ruby code for the corresponding regular expression and, then, calls the eval method for asking to the Ruby interpreter to evaluate the string (i.e., we are writing Ruby code that: (i) writes more Ruby code and (ii) executes the new code). 1 def nested_parentheses(levels) 2 eval(’@level0 = "[(]([^()’ + ’])*[)]"’) 3 (1..(levels-2)).reject {|i| 4 eval("@level#{i}" + ’= "[(]([^()’ + 5 ’] | #{@level’ + "#{i-1}" + ’} )*[)]"’) 6 } 7 if levels > 1 then 8 eval("@level#{levels-1}" + 9 ’= /[(]([^()’ +’] | #{@level’ + 10 "#{levels-2}" + ’} )*[)]/mx’) 11 eval "return @level#{levels-1}" 12 else 13 eval "return /#{@level0}/mx" 14 end 15 end Figure 6: Generating regular expressions to match nested parentheses 9 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 3.3.4 EFL Integration with Parsers and Text Template Engines Sometimes, regular expressions are not the best way to write certain exemplar changes. You may want to work at syntactical level (i.e., against a AST) or to use a text template. Since EFL is embedded in Ruby, EFL generators can integrate parsers (Racc2 and Rockit3) are two cur- rently available metaparsers for Ruby) and text templates (ERB4 is a valuable text template engine for Ruby). Figure 7 shows how to do this inside substitutions, maintaining the EFL support for detecting undesirable overlaps among the substitutions of a production and the possible collisions among the productions of a generation. The first substitution parameter is a very general regular expression that sets the exemplar scope for the parser or the template. The second parameter calls the parser or the template engine for processing the scoped text and producing the new code. Note that the scope is captured with parentheses and then is passed to the parser or the template through the associate backreference ($1). Figure 7: Example of how to integrate a parser or a text template with EFL 4 Example: a SPL for List Containers This section solves a simplified version of the “list container” example proposed in [31] using different flexibilization techniques. The aim of the example is to develop a portfolio of list containers written in 2http://rubyforge.org/projects/racc/ 3http://sourceforge.net/projects/rockit/ 4http://raa.ruby-lang.org/project/erb/ 10 http://rubyforge.org/projects/racc/ http://sourceforge.net/projects/rockit/ http://raa.ruby-lang.org/project/erb/ Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 C++. The PL scope is modeled in Figure 8 by a Feature Diagram (FD), which is a widely used notation to depict the commonalities and variabilities of the products supported by a PL [32, 33]. A FD is represented as a hierarchically arranged set of features with different relations among those features. Figure 8 includes three kinds of relations: mandatory (pointed by simple edges ending with a filled circle; e.g., all Lists in the portfolio have a Ownership feature), alternative (pointed by edges connected by an arc; e.g., External ref- erence, Owned reference and Copy are the mutually exclusive values for Ownership) and optional (pointed by simple edges ending with an empty circle; e.g., a List may have the Tracing feature). In this particular example, the meanings of the features are: 1. ElementType specifies the type of the elements stored in the list. 2. Ownership specifies how a list stores its elements: (a) External reference: the list keeps references to the original elements and is not responsible for element deallocation. (b) Owned reference: the list keeps references and is responsible for element deallocation. (c) Copy: the list keeps copies of the original elements and is responsible for their allocation and deallocation. 3. LengthCounter specifies if there is available a counter of type LengthType to know the length of the list. 4. Tracing indicates if a list traces its operation by logging function calls to the console. Figure 8: Feature diagram that models a SPL of list containers Lets suppose that the exemplar of Figure 9, which implements in C++ the shadowed features in Figure 8, is available at the beginning of the SPL development. According to the EDD approach, an analogy relation will be defined to automatically derive the remaining products from the exemplar. 11 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 1 class List { 2 private: 3 MyClass* head_; 4 List* tail_; 5 int length_; 6 public: 7 List(MyClass&h, List *t=0): 8 head_(0), tail_(t), length_(computedLength()) 9 { setHead(h); } 10 ~List() 11 { delete head_; } 12 void setHead(MyClass& h) 13 { 14 cout << "setHead(" << h <<")" << endl; 15 head_ = new MyClass(h); 16 } 17 MyClass& head() 18 { 19 cout << "head()" << endl; 20 return *head_; 21 } 22 void setTail(List *t) 23 { 24 cout << "setTail(t)" << endl; 25 tail_ = t; 26 length_ = computedLength(); 27 } 28 List *tail() const 29 { 30 cout << "setTail(t)" << endl; 31 return tail_; 32 } 33 const int& length() const 34 { return length_; } 35 private: 36 int computedLength() const 37 { return tail()?tail()->length()+1:1; } 38 }; Figure 9: An existing exemplar: a list of elements of type MyClass that implements features Copy Owner- ship, int LengthCounter and Tracing 12 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 4.1 Flexibilization with internal techniques Internal techniques provide an easy way to express the exemplar adaptations (i.e., they avoid the manipula- tion of intermediate representations of the exemplar such as abstract syntax trees) and take full advantage of the built-in facilities of the exemplar implementation language (for example, the host language type system can help us to detect flexibilization errors). However, this approach has several disadvantages: • The flexibilization of any kind of software artifact should be supported. Unfortunately, some artifacts are written in languages with reduced capability to manage the variability (for example, think about the flexibilization of HTML documentation using HTML). • Most of the variability mechanisms available in the current programming languages are able to per- form only a certain kind of flexibilizations. On the other hand, one flexibilization can be made using different mechanisms. As a consequence, many flexibilizations get complicated because require the combined use of several mechanisms or choosing among alternative mechanisms (the difficulties to choose the best variability mechanism for a given problem are illustrated in the chapter 6 of [34], where Coplien tries to systematize such election). C++ provides two ways of parametrizing types: genericity and inheritance. Nevertheless, both mecha- nisms are invasive (i.e., the lines 3, 5, 7, 12, 15, 17, 33 and 36 in the Figure 9 have to be changed). Whereas the inheritance flexibilization uses abstract classes and late binding, genericity manages the inter-product variability at compile-time. Figure 10 shows a new version of the exemplar, where feature ElementType is implemented using genericity (i.e., C++ templates). Sections 4.1.1 and 4.1.2 discuss how to implement the exemplar flexibilization for the features Owner- ship, LengthCounter and Tracing by using inheritance and aspects respectively. 4.1.1 Inheritance The flexibilizations related to Ownership, LengthCounter and Tracing can be implemented in classes that inherit from the exemplar. These classes will non-invasively adapt the exemplar adding and overwriting methods and attributes. However, the flexibilization requires the following adaptations, not supported by the inheritance mechanism: 1. Removing attributes, expressions, sentences and methods from the base class. For example, to gen- erate a list container without length counter the next elements should be deleted in Figure 9: at- tribute length_ in line 5, expression length_(computedLength()) in line 8, sentence length_ = computedLength() in line 26 and, methods length and computedLength in lines 33–37. 2. Changing the private attributes and methods on the base class to protected, to make them accessible from the derived classes. 3. Modifying the base class destructor. The C++ compiler ensures that all destructors are always called (see chapter 14 in [35]). Base class destructor ~List should be modified to prevent the element deallocation in list containers with external reference. Therefore, inheritance has a limited flexibilization power that involves changing the exemplar by hand (i.e., it is invasive). Figure 10 shows the new exemplar resulting from such manipulation. Figure 11 shows a flexibilization of the new exemplar based on multiple inheritance, where classes ExtRefList, OwnRefList and CopyList implement feature Ownership; class LengthCounterList imple- ments feature LengthCounter; and class TracingList class implements feature Tracing. Figure 11 also exemplifies how to get a list container MyList with Copy ownership, LengthCounter and Tracing. 13 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 1 template <class ElementType> 2 class List { 3 protected: 4 ElementType* head_; 5 List<ElementType>* tail_; 6 public: 7 List(ElementType&h, List<ElementType> *t=0) 8 { 9 setHead(h); 10 setTail(t); 11 } 12 virtual void setHead(ElementType& h) = 0 {}; 13 virtual ElementType& head() 14 { 15 return *head_; 16 } 17 virtual void setTail(List<ElementType> *t) 18 { 19 tail_ = t; 20 } 21 List<ElementType> *tail() const 22 { return tail_; } 23 }; Figure 10: Exemplar manipulated to make possible the flexibilizations based on inheritance and AOP (using AspectC++) Figure 11: Exemplar flexibilization based on multiple inheritance 14 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 However, this solution has a drawback: multiple inheritance introduces ambiguity that the compiler can not solve. For example, when method setTail of the class MyList is called, the compiler is unable to decide between the execution of the method setTail of LengthCounterList or the execution of setTail of TracingList. This ambiguity can be solved using single inheritance instead of multiple inheritance. But, as Figure 12 shows, such solution introduces excessive redundancy (i.e., a combinatory explosion of classes). Finally, Figure 13 shows how the multiple inheritance and the single inheritance redundancy can be avoided applying parametrized inheritance (i.e., using genericity to parametrize the base classes of LengthCounterList and TracingList). Figure 12: Exemplar flexibilization based on single inheritance 15 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 Figure 13: Exemplar flexibilization based on parametrized inheritance 16 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 4.1.2 Aspect Oriented Programming The flexibilizations related to Ownership, LengthCounter and Tracing can be implemented as aspects that crosscut the exemplar. However, the available Aspect Oriented Programming (AOP) extensions to pro- gramming languages often have limitations that hinder totally non-invasive flexibilizations. For example, the only structural changes that AspectC++ supports are: (1) the addition of attributes and methods to a class, and (2) the change of the base class of a given class5. So, an ApectC++ flexibilization would imply to manipulate the exemplar to transform it into Figure 10. Besides, AspecC++ weaving is restricted to non- templated code6. Therefore, the implementation of features ElementType and LengthType using genericity should be substituted by an inefficient implementation based on inheritance. 4.2 Flexibilization with external techniques 4.2.1 Text templates A template can be viewed as a piece of an exemplar with slots. The exemplar code that is common to all the domain products is maintained in the template, whereas the variable code is replaced by slots, that are filled with metacode which specifies how code must change. Figure 14 shows a piece of the exemplar flexibilization applying the ERB Ruby library for text templates, where metacode (i.e., the Ruby code that implements the variable domain features) is embraced within symbols <% and %>. The main drawback of this template–based solution is that code and metacode are strongly coupled. Indeed, if the template engine does not support AOP, templates may suffer metacode tangling (multiple variable concerns implemented simultaneously in a template) or metacode scattering (a variable concern implemented in multiple templates). For example, the flexibilization showed in Figure 11 suffers metacode tangling because metacode in lines 1, 3, 4, 9, 10 and 11 manages the ElementType variability, whereas the metacode in lines 5, 6, 7, 13 and 15 manages the LengthCounter variability. 1 class <%=@list_specificacion[’Element Type’]%>List { 2 private: 3 <%=@list_specificacion[’Element Type’]%>* head_; 4 <%=@list_specificacion[’Element Type’]%>List* tail_; 5 <% if @list_specificacion[’Length Counter Type’]%> 6 <%=@list_specificacion[’Length Counter Type’]%> length_; 7 <% end %> 8 public: 9 <%=@list_specificacion[’Element Type’]%>List 10 (<%=@list_specificacion[’Element Type’]%>&h, 11 <%=@list_specificacion[’Element Type’]%>List *t=0): 12 head_(0), tail_(t) 13 <% if @list_specificacion[’Length Counter Type’]%> 14 , length_(computedLength()) 15 <% end %> 16 { setHead(h); } 17 ... Figure 14: Exemplar flexibilization using text templates 4.2.2 EFL Figure 15 shows an exemplar flexibilization using the available EFL implementation in Ruby, where the Ele- mentType, Ownership, LengthCounter and Tracing are managed by the generators ElementType, Ownership, 5see page 28 of “AspectC++ Language Reference. Version 1.6”, available at http://www.aspectc.org/fileadmin/documentation/ac-language 6see page 19 of “AspectC++ Compiler Manual. Version 1.1”, available at http://www.aspectc.org/fileadmin/documentation/ac-compilerma 17 http://www.aspectc.org/fileadmin/documentation/ac-languageref.pdf http://www.aspectc.org/fileadmin/documentation/ac-compilerman.pdf Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 LengthCounter and Tracing. Such flexibilization has good modularity, is concise, non-invasive and manages the inter-product variability before runtime. Generators in Figure 15 include substitutions (e.g., line 3 defines the substitution of all the occurrences of the MyClass code pattern for the value of the element_type string) and productions (e.g., line 18 defines a production that applies the substitutions defined in lines 11, 12 and 14 to the exemplar file to produce the out file). 1 class ElementType < Generator 2 def initialize(exemplar, out, element_type) 3 gsub(/MyClass/, element_type) 4 prod(exemplar, out) 5 end 6 end 7 class Ownership < Generator 8 def initialize(exemplar, out, ownership_type) 9 case ownership_type 10 when ’External reference’ 11 sub /delete head_;/, ’’ 12 sub /new MyClass\(h\)/ , ’&h’ 13 when ’Owned reference’ 14 sub /new MyClass\(h\)/ , ’&h’ 15 when ’Copy’ 16 # no change 17 end 18 prod(exemplar, out) 19 end 20 end 21 class LengthCounter < Generator 22 def initialize(exemplar, out, length_counter_type) 23 if length_counter_type 24 gsub(/int/, length_counter_type) 25 else 26 gsub(/^.*length.*$/i, ’’) 27 gsub(/\)\:/, ’\0head_(0), tail_(t)’) 28 end 29 prod(exemplar, out) 30 end 31 end 32 class Tracing < Generator 33 def initialize(exemplar, out, tracing) 34 gsub(/cout.+$/, ’’) if !tracing 35 prod(exemplar, out) 36 end 37 end Figure 15: Exemplar flexibilization using EFL Finally, Figure 16 depicts how the generators are combined to adapt the exemplar cooperatively. Be- cause there are overlaps between the substitutions of generators ElementType and Ownership, they are sequentially combined. On the other hand, generators Ownership, LengthCounter and Tracing are com- bined with the add operator. 5 Conclusions We have introduced the EDD process to develop SPLs, which minimizes the product line adoption barrier by means of a reactive and extractive approach. The EDD starting point is any domain product built using conventional software engineering. EDD pursues the reuse of this exemplar applying intensively the idea of analogy to all the domain engineering activities. We have described how to implement analogy relations to automatically derive all the domain products from existing exemplars by using techniques widespread applied to generalize code. The limitations of 18 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 Figure 16: Combination of the EFL generators such techniques have been exposed and the new language EFL has been proposed to overcome those limitations. At the moment, EDD and EFL have been successfully used to develop: (i) a Data Acquisition SPL for the Astrophysics Institute of the Canary Islands [36] and (ii) a generative model that produces, from abstract specifications, change notifications written in PL/SQL for Oracle databases [37]. References [1] K. Pohl, G. Bockle, F. Linden, Software Product Line Engineering: Foundations, Principles and Tech- niques, Springer, 2005. [2] X. Peng, S.-W. Lee, W.-Y. Zhao, Feature-oriented nonfunctional requirement analysis for software prod- uct line, Journal of Computer Science and Technology 24 (2) (2009) 319–338. [3] C. Dinnus, K. Pohl, Software Product Line Engineering, Springer Berlin Heidelberg, 2005, Ch. Experi- ences with Software Product Line Engineering, pp. 413–434. [4] J. Liu, S. Basu, R. R. Lutz, Compositional model checking of software product lines using variation point obligations, Automated Software Engineering 18 (2011) 39–76. [5] R. Heradio, D. Fernandez-Amoros, L. Torre-Cubillo, A. P. Garcia-Plaza, Improving the accuracy of coplimo to estimate the payoff of a software product line, Expert Systems with Applications 39 (9) (2012) 7919–7928. [6] P. Clements, L. Northrop, Software Product Lines: Practices and Patterns, Addison-Wesley, 2001. [7] C. Krueger, Eliminating the adoption barrier, IEEE Software 19 (4) (2002) 29–31. [8] K. Schmid, M. Verlage, The economic impact of product line adoption and evolution, IEEE Software 19 (2002) 50–57. [9] M. Fowler, Domain–Specific Languages, Addison–Wesley, 2010. 19 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 [10] M. E. Edge, P. R. F. Sampaio, The design of ffml: A rule-based policy modelling language for proactive fraud management in financial data streams, Expert Systems with Applications 39 (11) (2012) 9966– 9985. [11] I. Reinhartz-Berger, Towards automatization of domain modeling, Data and Knowledge Engineering 69 (2010) 491–515. [12] J. Herrington, Code Generation in Action, Manning Publications, 2003. [13] C. Pohl, A. Rummler, V. Gasiunas, N. Loughran, H. Arboleda, F. de Alexandria Fernandes, J. Noye, A. Nunez, R. Passama, M. S. Jean-Claude Royer and, Survey of existing implementation techniques with respect to their support for the requirements identified in m3.2, Tech. rep., AMPLE, Aspect ŰOriented, Model-Driven, Product Line Engineering, Specific Targeted Research Project: IST- 33710 (2007). [14] J. Lee, J. Park, G. Yoo, E. Lee, Goal-based automated code generation in self-adaptive system, Journal of Computer Science and Technology 25 (6) (2010) 1118–1129. [15] J. Gallardo, A. I. Molina, C. Bravo, M. A. Redondo, C. A. Collazos, An ontological conceptualization approach for awareness in domain-independent collaborative modeling systems: Application to a model-driven development method, Expert Systems with Applications 38 (2) (2011) 1099–1118. [16] H. Zhang, S. Jarzabek, A hybrid approach to feature-oriented programming in xvcl, in: 14th Interna- tional Software Product Line Conference, Jeju Island, South Korea, 2010, pp. 440–445. [17] M. Voelter, I. Groher, Product line implementation using aspect-oriented and model-driven software development, in: 11th International Software Product Line Conference, Washington, DC, USA, 2007, pp. 233–242. [18] T. Elrad, M. Aksit, G. Kiczales, K. Lieberherr, H. Ossher, Discussing aspects of aop, Communications of the ACM 44 (2001) 33–38. [19] S. A. Vidal, C. A. Marcos, Building an expert system to assist system refactorization, Expert Systems with Applications 39 (3) (2012) 3810–3816. [20] F. Berzal, F. J. Cortijo, A. Jimenez, Tminer aspects: Crosscutting concerns in the tminer component- based data mining framework, Expert Systems with Applications 37 (9) (2010) 6675 – 6681. [21] D. F. Barrero, M. D. R-Moreno, D. Camacho, Adapting searchy to extract data using evolved wrappers, Expert Systems with Applications 39 (3) (2012) 3061–3070. [22] A. V. Aho, M. S. Lam, R. Sethi, J. D. Ullman, Compilers: Principles, Techniques, and Tools, Prentice Hall, 2006. [23] T. Parr, The Definitive Antlr Reference: Building Domain-Specific Languages, Pragmatic Bookshelf, 2007. [24] M. Bravenboer, K. T. Kalleberg, R. Vermaas, E. Visser, Stratego/xt 0.17. a language and toolset for program transformation, Science of Computer Programming 72 (1-2) (2008) 52–70. [25] E. Balland, P. Brauner, R. Kopetz, P.-E. Moreau, A. Reilles, Tom: Piggybacking rewriting on java, in: 18th International Conference on Term rewriting and applications, Paris, France, 2007, pp. 36–47. [26] C. Catal, U. Sevim, B. Diri, Practical development of an eclipse-based software fault prediction tool using naive bayes algorithm, Expert Systems with Applications 38 (3) (2011) 2347 – 2353. 20 Draft version of the paper published in Expert Systems with Applications, doi:10.1016/j.eswa.2012.05.004 [27] V. F. Lopez, R. Aguilar, L. Alonso, M. N. Moreno, Data mining for grammatical inference with bioinfor- matics criteria, Expert Systems with Applications 39 (3) (2012) 2330–2334. [28] R. Heradio, J. A. Cerrada, J. C. Lopez, J. R. Coz, Code generation with the exemplar flexibilization language, Electronic Notes in Theoretical Computer Science 238 (2) (2009) 25–34. [29] B. Boehm, A spiral model of software development and enhancement, ACM SIGSOFT Software Engi- neering Notes 11 (1986) 14–24. [30] J. E. Friedl, Mastering Regular Expressions, O’Reilly Media, 2006. [31] K. Czarnecki, U. Eisenecker, Generative Programming: Methods Tools and Applications, Addison- Wesley, 2000. [32] K. Kang, S. Cohen, J. Hess, W. Novak, S. Peterson, Feature-oriented domain analysis (foda) feasibility study, Tech. rep., CMU/SEI-90-TR-21, Software Engineering Institute (1990). [33] J. Guo, Y. Wang, P. Trinidad, D. Benavides, Consistency maintenance for evolving feature models, Expert Systems with Applications 39 (5) (2012) 4987–4998. [34] J. O. Coplien, Multi-Paradigm Design for C++, Addison-Wesley, 1998. [35] B. Eckel, Thinking in C++: Introduction to Standard C++, Volume One, Prentice Hall. [36] J. C. Lopez-Ruiz, R. Heradio, J. Cerrada, J. Coz, P. L. Ramos, A first-generation software product line for data acquisition systems in astronomy, in: Advanced Software and Control for Astronomy. Marseille, France, 2008. [37] J. Coz, R. Heradio, J. Cerrada, J. Lopez-Ruiz, A generative approach to improve the abstraction level to build applications based on the notification of changes in databases, in: International Conference on Enterprise Information Systems. Barcelona, Spain, 2008. 21 Introduction Exemplar Driven Development Exemplar Flexibilization Language Defining Generators Combining Generators EFL Capabilities to Overcome the Regular Expressions Limitations The Zoom Operator Anti-patterns Managing Nested Constructs EFL Integration with Parsers and Text Template Engines Example: a SPL for List Containers Flexibilization with internal techniques Inheritance Aspect Oriented Programming Flexibilization with external techniques Text templates EFL Conclusions 
work_qrl5mxz37fertdmqjhcuavftvy	----	Brief survey of crowdsourcing for data mining Expert Systems with Applications 41 (2014) 7987–7994 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Review Brief survey of crowdsourcing for data mining http://dx.doi.org/10.1016/j.eswa.2014.06.044 0957-4174/� 2014 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. E-mail address: wangzh@hit.edu.cn (W. Hongzhi). Guo Xintong a, Wang Hongzhi a,⇑, Yangqiu Song b, Gao Hong a a School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150006, China b Department of Computer Science and Technology, University of Illinois at Urbana-Champaign, IL, USA a r t i c l e i n f o Article history: Available online 11 July 2014 Keywords: Data mining Crowdsourcing Quality control Survey a b s t r a c t Crowdsourcing allows large-scale and flexible invocation of human input for data gathering and analysis, which introduces a new paradigm of data mining process. Traditional data mining methods often require the experts in analytic domains to annotate the data. However, it is expensive and usually takes a long time. Crowdsourcing enables the use of heterogeneous background knowledge from volunteers and distributes the annotation process to small portions of efforts from different contributions. This paper reviews the state-of-the-arts on the crowdsourcing for data mining in recent years. We first review the challenges and opportunities of data mining tasks using crowdsourcing, and summarize the framework of them. Then we highlight several exemplar works in each component of the framework, including question designing, data mining and quality control. Finally, we conclude the limitation of crowdsourcing for data mining and suggest related areas for future research. � 2014 Elsevier Ltd. All rights reserved. 1. Introduction classification problem, labeled data is used for training the classi- People from different fields analyze a variety of datasets to understand human behaviors, find new trends in society, and possibly formulate adequate policies in response. Typically, we address the problem of finding interesting and unknown patterns via data mining methodology. Data mining enables people to extract information from a data set and convert it into a comprehensible structure for further use. Typical data mining techniques, however, are not suitable for current applications. First, when mining the datasets, we must have access to all relevant information. In fact, it is impossible to obtain all these transactions, which mainly because of the proper- ties of the human memory. People’s memories are prone to remember summaries, rather than exact details (Boim et al., 2012). Consider the following case. A social scientist wants to ana- lyze life habits of people. The database includes leisure activities (watching TV, jogging, reading, etc.) correlated with time of the day, weather and so on. But it is unrealistic for people to recall an exhaustive list of all cases they did. People can make assump- tions in order to compensate the loss of information by crowd- sourcing the mining task. Second, some mining algorithms are time-consuming, especially used for large datasets, which also leads to much more extra cost. Finally, raw data mining technologies are lack of related information. Algorithm has to be taught the knowledge before mining. For example, for the fier to have the ability of classifying new coming test data. How- ever, acquiring the labeled data is time consuming and costly. In the circumstances, we can solve this problem by crowdsourc- ing. As crowdsourcing is based on the people who have the incen- tives to work on small tasks, the mining tasks can benefit from the aggregation of labeling work which is time-controllable, flexible, easy to implement due to the current crowdsourcing platform. Crowdsourcing is an emerging and powerful information pro- curement paradigm that has appeared under many names, includ- ing social computing, collective intelligence and human computation (Quinn & Bederson, 2011). Requesters decompose the whole task into several small tasks and push them to the crowd, and workers accomplish questions for intrinsic or extrinsic reasons (Von Ahn & Dabbish, 2008). Although people may not remember all of transactions precisely, many current studies prove that simple summaries can still achieve a positive result, and even more complicated questions (Boim et al., 2012). Crowdsourcing has played important roles in data mining. In some kind of scenarios, it can help people resolve the problems in a more efficient way and give them deeply understanding to apply crowdsourcing. Here we give some situations for the applica- tions of crowdsourcing techniques in various real-world data mining tasks. Crisis Map: Crisis map is one of the most representative applica- tions of crowdsourcing. It is a platform, designed to do information collection, analysis of mass data and display in a straightforward way in real time during a crisis. It has become a powerful mecha- nism for a large number of people to contribute about crisis events. http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.06.044&domain=pdf http://dx.doi.org/10.1016/j.eswa.2014.06.044 mailto:wangzh@hit.edu.cn http://dx.doi.org/10.1016/j.eswa.2014.06.044 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 7988 G. Xintong et al. / Expert Systems with Applications 41 (2014) 7987–7994 People not only provide useful information about the crisis situa- tion, but also cluster materials into meaningful categories. Then people with no field-specific skills filter out the irrelevant parts, do analysis and assemble reports. What is more, these crisis maps can visualize a large amount of data and give the rescue teams bet- ter insights of the relief situation (Goolsby, 2010). People gener- ated numerous messages and photos after the devastating earthquake in Haiti happened on 2010, through social media net- working (Gao, Barbier, & Goolsby, 2011). The use of crisis map in disaster accelerates the application development to leverage the value of crowdsourcing. Homeland Security: Crowdsourcing can also benefit homeland security. We can use crowds to contribute to delivering quality information and identifying the suspects. The Boston Marathon bombing, happened on April 15, 2013, caused many injuries and deaths. On the next day, an appeal went out to the public, urging the citizens to submit all photos and videos that they might have of the Boston Marathon environment (Spenser, 2013). A number of sites (Reddit & 4chan) were set up to aggregate the photos and videos. And then the crowd helped to identify the suspects in the flooded materials collected from the first appeal. The public responded quickly and provided valuable intelligence both times (Markowsky, 2013). Facebook: Facebook is another popular website that can be used for crowdsourcing. Compared to twitter, Facebook has more infor- mation sources, including blog and picture. So Facebook can fulfill some sophisticated tasks, such as character analysis, financial anal- ysis (Libert & Spector, 2007), activity planning (Brabham, Sanchez, & Bartholomew, 2009), and product repository generation (Budde & Michahelles, 2010). Facebook builds up various applications for individuals to design their own crowdsourcing tasks. Lots of crowdsourcing platforms have sprung up during these years, such as CloudCrowd (used to write and edit the project), Crowdflower and so on. The Amazon Mechanical Turk is one of the most famous and largest in scale. The Amazon Mechanical Turk (MTurk) allows individuals or business corporations (known as requester) post various tasks, such as image clustering, document labeling, creative designing and so on. The workers (known as Tur- ker) choose HITs (Human Intelligence Tasks) to accomplish for monetary incentive. The requesters connect web applications and MTurk through open interfaces (APIs), which benefit customizing task design and analysis. Managing and analysis data on the crowdsourcing platform have recently become a wide-spread phenomenon, leading to explosion of research activity in recent years (Amsterdamer, Grossman, Milo, & Senellart, 2013a, 2013b). What we cannot ignore are the challenges arising from the real-world applications. The challenges include, but not limit to adaptive question deliver system, recommendation framework for requester to design task, as well as specific mining algorithm. We will discuss them in detail in Section 7. Apparently, a better understanding in crowdsourcing for data mining can help us tap into this powerful new resource in a more efficient way. That is why our survey is important. The contribu- tions of this paper are summarized as follows. 1. We review the state-of-the-art work on the crowdsourcing for data mining in recent years. As we know, this is the first survey about crowdsourcing techniques for data mining. 2. We point it out the difference between raw data mining algo- rithms and those based on crowdsourcing. There are more fac- tors to be considered when data mining task accomplished by crowdsourcing. 3. According to existing work, we summarize a general framework of crowdsourcing for data mining, which includes question design, mining process and quality control. 4. We review the highlight works in each component of the framework. 5. We give some generic tips about task design and discuss the quality control method selection strategies for data mining tasks. It is quite instructive and meaningful for requester to follow. 6. We investigate challenges and opportunities of data mining tasks in crowdsourcing. We also conclude the limitation of crowdsourcing for data mining and suggest related areas for future research. Crowdsourcing is a powerful tool for government to collect and analyze data. It provides new opportunities for data mining which has widely applications in expert systems. The methods proposed in this article are not only fit for data mining, but also good refer- ences for other related work, such as information retrieval, machine learning and crisis management. Such fields also have close relationship to expert systems. The rest of this paper is organized as follows. Section 1 provides a general framework for crowd mining. Section 2 tells how to design a task for data mining. Section 3 proposes various data mining tasks that can be performed by means of crowdsourcing. Section 4 pre- sents the study on quality control for data mining, which is closely related to the result of data mining. Section 5 introduces situations that crowdsourcing method is not suitable for tackling tasks. Sec- tion 6 describes what we can do in the future. At the end of the arti- cle, we give a discussion and conclusion of our work. 2. Framework Traditional data mining methodologies and technologies are sometimes time-consuming, inflexible, expensive to implement, and poor scalable. Crowdsourcing can be applied to manage data and extract interesting patterns from the data sets more efficiently and intelligently by comparison. From existing work of crowd- sourcing techniques (Boim et al., 2012; Amsterdamer et al., 2013a, 2013b; Barbier, Zafarani, Gao, Fung, & Liu, 2012; Karger, Oh & Shah, 2011; Weaver, Boyle & Besaleva, 2012), we conclude that using crowdsourcing for data mining can be performed by following a three-step procedure: question design, mining and quality control. Question Design: Well-designed tasks can obtain high-quality answers. Questions should be designed based on the purpose of the data mining task. We address the problem of effective crowd- sourcing, namely gathering data from the crowd in a way that is economical in time and expense. Mining: The mining phase absolutely takes the center stage in the whole process. Data mining tasks can be divided into the multi- ple kinds: classification, clustering, semi-supervised learning, and association rules mining. Classification has been widely used in many fields, such as face recognition, disaster rescue. Some research takes advantages of crowdsourcing to identify association rules between relating signs and symptoms to diseases (Wright, Chen, & Maloney, 2010). Crowdsourcing appears to have several important merits compared with other automatic knowledge- based approaches. Quality Control: Due to the nature of crowdsourcing task, a qual- ity control step is necessary for the result after the mining step. Malicious workers, who are only attempting to maximize their income or lack of necessary training, are detrimental to the mining result (Venetis & Garcia-Molina, 2012). Quality control step uses vote system, redundant workers, worker’s reputation and other methods to pick out the irresponsible workers. In the following sections, we will discuss these three steps, respectively. G. Xintong et al. / Expert Systems with Applications 41 (2014) 7987–7994 7989 3. Question design In a crowdsourcing system, a requester has to decide how to split the whole task into several small tasks, each of which could be distributed as a unit. For some kind of data mining tasks, such as clustering, each worker has only a partial view of the data. How to decompose the whole task into small pieces is a primary factor to be considered. According to the data mining task, we should ask the correct questions to achieve good performance, and design different ques- tions to different workers to make further progress. In this process, we address the problem of gathering data from the crowd with fewer questions and more information. Here Alonso and Lease (2011) give some generic tips for task design. 1. Experiments should be self-contained. 2. Instructions for task must be short and simple, brief and con- cise. Too complex task or ambiguous instructions for workers to understand will generate poor quality answers. 3. Be very clear with the relevance task. If a similar task has been published, there is no need to replicate it. 4. Engage with the worker. Cancel useless stuff. We can run a test with a very small data set and gather feedback from workers to enhance the task design. 5. Always ask for feedback (open-ended question) in an input box. Iterate and modify accordingly. 6. UI design. Highlight important concepts and pay attention to the layouts. Eickhoff and de Vries (2011) investigate the commonly cheating methods of malicious crowdsourcing workers. They establish a ser- ies of experimental schemes, including task-dependent valuation, interface-dependent evaluation and audience-dependent evalua- tion. They draw a conclusion that bad workers are rarely appeared in novel tasks that contain innovation and information extraction, and the number of poor-quality workers will be considerably reduced if we apply filtering process. Besides, the properties of the question are closely related to the workers’ answers. Well designed task is conductive to high quality mining result. Defensive task design is a tool for quality control, such as employing spammers to judge the result and referring to worker’s reputation. We can add qualifying questions which can block the unqualified workers, gold standard questions for which malicious answers can be filtered out in the process of crowdsourc- ing, and checking completion time. The easiest implementation method to distinguish correct answer is to hire redundant workers to finish the same HIT, and then aggregate them by applying a majority rule (Kazai, Kamps, Koolen, & Milic-Frayling, 2011). 4. Mining data from crowdsourcing Various types of data mining tasks can be accomplished by means of crowdsourcing, e.g. classification, clustering, semi- supervised learning, and association rule mining. Traditional algo- rithms have difficulties in tackling these problems, for the lack of knowledge. In these situations, the powerful crowds can perform more accurately, flexibly and efficiently than the existing auto- matic algorithms. We will discuss how crowdsourcing can be used to solve these problems and provide some application scenario. 4.1. Classification Crowdsourcing can be utilized to solve classification problem. Crowdsourcing method has much more advantages over the general data mining technology. A typical classification problem is to distinguish male and female from a social network users. In the original data mining perspective, a classifier is built to extract features from the given datasets in the first step. In the second step, the classifier is used for classification. Compared to this, if we dis- tribute this task to the crowd, everyone can classify the objects immediately. Sometimes we can say that the wisdom of the crowd allows for more accuracy than any other classification algorithm. Documents categorization can be accomplished by users on the website as well. This approach has been applied on many domains successfully, such as Digg and Yahoo! Directory. Digg is an aggre- gator to customize the user’s news front page. Digg also allows users to tag to the submitted links, which widen the scope to include more relevant articles the users may be interested in. We find that accuracy increases as the more and more users participate in. Another widely known example is CAPTCHA. ‘‘CAPTCHA (Von Ahn, Maurer, McMillen, Abraham, & Blum, 2008) is a challenge- response test used on the World Wide Web to determine whether a user is a human or a computer’’. Technologies at present cannot rec- ognize distorted text as fast as humans can. Human enter the char- acters to digitize handwriting text (Von Ahn, 2009). Chilton, Little, Edge, Weld, and Landay (2013)) present an algo- rithm, named CASCADE, to create an overall consistent taxonomy by distributing HITs to many individuals, each of whom has only a partial view of the data. CASCADE hires many unskilled labors to produce taxonomies. The quality of classification is approximate to that of human experts, while the cost of CASCADE is very cheap. Furthermore, Bragg and Weld (2013) present DELUGE, an improved workflow on the basis of CASCADE. DELUGE produces taxonomies with equivalent quality in spite of reducing the work- force. The categorization step, which is the most consuming, is optimized by decision theory. The experiment result demonstrates that less than 10% of the workers are required by the original approach. As we mentioned before, crowdsourcing helps a lot in disaster rescue. Zhai, Kijewski-Correa, Hachen, and Madey (2012) establish an online framework that generates human computation resources to tackle an image labeling task, classifying post-disaster photos according to damage extent. In real life, such type of information is needed to manage risks in disaster-prone areas, both in pre- disaster risk reductions and post-disaster damage assessments. Other related works that considering the budget allocation con- tributed to the classification work as well. Tran-Thanh, Venanzi, Rogers, and Jennings (2013) raise the issue of how to allocate the budget for redundant workers when dealing with classification tasks, where the key challenge is to find a proper balance between the total cost and quality. They propose CrowdBudget, a bud- get allocation algorithm, aiming to minimize estimation error with the limited fund. 4.2. Clustering Clustering is more complicated than classification problem. One of the aspects, there are lots of ways to define the similarity between items. Different measure of similarity may lead to differ- ent result. Similarly, we can perform cluster task on the crowd- sourcing platform. Many recent social networking sites give humans permission to create categories. In the case of Twitter, users assign tags to their tweets in order to follow up the trending topics. This facilitates quick retrieval when searching for tweets and again, and forms a discussion groups about the hot issue automatically (Barbier et al., 2012). What is more, a large set of Tweets relevant to a par- ticular cluster can be an excellent source for professionals to analyze. 7990 G. Xintong et al. / Expert Systems with Applications 41 (2014) 7987–7994 Chen, Wang, and Tan (2012) create a friendly environment, using crowd to visualize web images into clusters. The method has two stages. The first stage separates an image set into multiple clusters and the second stage purifies each generated cluster inde- pendently. During the whole stage, computers select informative images and the crowds help to label the images to improve the quality. The experimental results demonstrate the combinations of computers and a large number of human workers benefit high-quality visual clusters. Here we require addressing some challenges in crowdsourced clustering. (1) Each worker has only a fraction of the data, so we need additional algorithm to merge the results. (2) Different work- ers may have different clustering standard, leading to produce dif- ferent numbers of categories. (3) The underlying category structure may be hierarchical. According to the intractable problems men- tioned above, Gomes, Welinder, Krause, and Perona (2011) propose a model, based on Variational Bayes method, of how crowdsourc- ing can be applied to clustering. First, divide the dataset into over- lap subsets. And then workers propose partial clustering. Finally, use Bayes model to aggregate the partial clusters into one cluster. 4.3. Semi-supervised learning In semi-supervised learning, we use labeled data to acquire nec- essary knowledge, and then label the unlabeled data. This proce- dure significantly increases the learning accuracy. Similar to previous tasks, semi-supervised learning can also be performed with crowdsourcing. Tang and Lease (2011) aim to achieve more accuracy when inferring consensus labels, with correspondingly less labeled train- ing data for estimating worker accuracy by a Naive Bayes approach. We can apply this method in the situation when we have large amount of unlabeled items and a very small set of expert- labeled items. As human have better learning ability than the algorithm, requesters can provide essential knowledge of how the given task can be performed correctly. This is a well-practiced labeling tech- nique for sophisticated labeling tasks in the data mining field (Sorokin & Forsyth, 2008). 4.4. Sampling Sampling is in correlation with the selection of a subset with sufficient information so that people can easily verify hypotheses devised from the sample information in the whole datasets. And it is one of the complex tasks in data mining. A challenging problem to be solved here is how the requester should select the appropriate distribution so that benefit of the information gathered from the sample is maximized. Crowds have been proved to be trustworthy data samplers, and they keep working on enhancing the precision of the results in maximally informative samples (Von Ahn, 2009). 4.5. Association rule mining Data mining techniques have been developed for discovering and identifying underlying association rules among data items. The typical application is shopping baskets analysis. That is, a mar- ket analyst can explore relation about which items are purchased together by analyzing purchasing records. However, when refer- ring to human behavior, it is impossible to get access to all the transactions. This is because, typically, the everyday actions of people are not recorded in detail, except in their own memories, which are limited in terms of exact recollection. Indeed, social studies show that instead of full details, people often tend to recall sufficient information in the form of summaries when asked the appropriate questions. Amsterdamer et al. (2013a, 2013b) lay the foundations of crowd mining for the first time. They define the basic concepts of mining the association rules. Then, they present an integrated system con- sisted of general-purpose components, incorporating interactive selection of questions to ask, effective mining component, error estimation and so on. Another article from Amsterdamer et al. (2013a, 2013b) present a demo named CrowdMiner. The essence of CrowdMiner is an algorithm enabling the mining of appealing data patterns from the crowd. It allows flexible choice about appropriate questions to ask the crowd as well, with the purpose of gathering more information with fewer questions. 4.6. Validation Similarly, we can perform task to human to validate the correct- ness of mining algorithm and predict the mining result of the auto- mated method on large dataset (Barbier et al., 2012). Agarwal, Liu, Tang, and Yu (2008) want to identify influential bloggers at a blog site. As we know, there is no training and testing data for them to evaluate the efficiency of the proposed model. They use crowdsourced result generated on Digg as a reasonable reference to compare with their automatic techniques. The crowd- sourcing results validate their hypothesis. 5. Quality control Crowdsourcing is a powerful platform for many mining task. Workers may misunderstand the tasks, make mistakes, or deliber- ately cheat the system, which can cause errors or bad results. Bad answers consume a lot of time and money to filter them out. The reason is twofold. On one hand, many workers fulfilled tasks to kill time or gain sense of achievement in the beginning, with the payment being only a minor attraction. Nowadays, the overwhelming majority of workers are attracted by the financial reward (Eickhoff & de Vries, 2011). Payment is the easiest method to motivate people. However, monetary incentives can effectively increase participa- tion, but cannot improve quality. As a consequence, a high number of malicious users arise. They try to finish HITs as quickly as possi- ble in order to maximize their profit. This leads to a mass of generic answers. One the other hand, the worker may lack expertise or skills to handle some kind of complex job (Liu et al., 2012). Incorrect answers may be provided because of this. To tackle this problem, we can provide some basic knowledge to workers before the work, or require some qualifications to prove themselves qualified to fin- ish the task. In a nutshell, the quality of crowdsourced data has great influ- ence on the mining result, so researchers have to pay careful atten- tion to it. To improve the trustworthiness of mining result, various techniques could be employed. They are discussed as follows, respectively. 5.1. Vote Voting system hires additional spammers to judge the crowd- sourcing outcome, and follows majority rule, which is simple to implement in real world application. Voting system is quite successful and easy-implement in deter- mining the credibility of messages on the web. In social media sites, people use thumb up or thump down to express their atti- tudes for or against. For instance, on YouTube, users can provide feedback (positive or negative) for user comments. The website G. Xintong et al. / Expert Systems with Applications 41 (2014) 7987–7994 7991 will hide the comments with too many negative feedbacks auto- matically (Barbier et al., 2012). Another example is eBay, the buyer vote to seller to give other buyers a reference to the product. And the seller vote according to the buyer’s trustworthiness. Although voting approach has its advantages, we have to admit this approach does have drawbacks. Minority voters have less access to express their views and so researchers would be less likely to benefit from these special ideas. 5.2. Redundant work Apparently, redundant work means the requester hire redun- dant workers to finish the same crowdsourcing task. Actually, redundancy work is widely used to identify the correct answers by the requesters. However, what we should pay attention to is that redundancy is not a panacea. Large-scale redundancy is expensive, and redundant workers sometimes may not lead to good result. Therefore, we can apply redundant workers to controversial item to save money, not to all the items. 5.3. Worker’s reputation Worker’s reputation is composed primarily of the worker’s accuracy on previously submitted HITs. Reputation is a practical judgment on workers’ trustworthiness, which urges people to complete work with high quality continuously. It has become a popular method for evaluation of the quality of workers not only on crowdsourcing platforms, but also on a lot of online forums. In Amazon Mechanical Turk, requesters can require the level of worker’s HITs Approval Rate and some other qualifications, such as language skill. When cheating is detected, the reputation reduces and the system forbids low reputation workers, e.g. users who fail two tasks may be put into blacklist (Heimerl, Gawalt, Chen, Parikh, & Hartmann, 2012). Allahbakhsh et al. (2012) propose a reputation management framework, which adequately takes into account the values of the tasks completed, the trustworthiness of the assessors, the results of the tasks and the time of evaluation in order to achieve more credible quality metrics for workers and assessors. Ipeirotis, Provost, & Wang (2010) propose an algorithm that improve the existing advanced techniques of the labeling process in crowdsourcing platform and can be applied when the workers should answer a multiple choice question to complete a task. The algorithm enables the separation of intrinsic error rate from the bias worker. Finally, the algorithm produces a scalar score to mea- sure the intrinsic quality of each worker. Worker’s reputation has potential problems. First, the major determinant factor of human’s reputation is the acceptance rate of HITs. The requesters always accept all answers and do not dis- pose the noisy data right now. Afterwards, requesters do not give feedback to the workers, respectively. The malicious users take advantages of the loophole to receive the increase in reputation and start to complete next tasks. Second, reputation system could not avoid cheating. Ipeirotis (2010) create a scam, which is designed to accelerate the reputation improvement, named rank boosting on his weblog. With this strategy the worker creates a requester account, distributes a number of simple HITs and imme- diately completes them with his worker account. The worker almost spends no money in boosting his rank. 5.4. Gold standard Gold standard is the benchmark that is the best available in par- ticular situation. It does not have to be necessarily the best answer for the condition in giving terms. In crowdsourcing domain, we may pay expert to use small group of data to set gold standard, and then we can compare the gold standard data with the HIT‘s work to judge the reliability and filter out the poor-quality workers. Bernstein, Teevan, Dumais, Liebling, and Horvitz (2012) build a system on the idea of gold standard questions that the requester has labeled as the ground truth. When encountering the ground truth questions, the worker’s answer must include at least one choice from an inclusion list and none from an exclusion list. Le, Edmonds, Hester, and Biewald (2010) insert gold standard data into questions and robustly rejected bad answers to ensure quality when workers made mistakes in those gold standard data. Bernstein, Brandt, Miller, and Karger (2011) extend gold standard questions with novel technique. Also, some trap questions can be mixed with real questions and the system can easily notify the bad answers. Moreover, Callison-Burch and Dredze (2010) suggest that the tasks should be well-priced and make clear enough for workers to follow. They also propose several approaches to restrain cheating, such like using images of sentences instead of text in order to prohibit copying and pasting in translation tasks. For some situations, like kinds of creative jobs, gold standard seems difficult to set. 5.5. Other solutions for quality control There are some other solutions for controlling the quality of mining result, which may be more complicated than the methods above, or just the mixture of them. Liu et al. (2012) design and implement a Crowdsourcing Data Analytics System, CDAS, which is a framework allowing task design and deployment in various crowdsourcing scenarios. Allowing for the human workers’ historical performances, the estimation com- ponent calculates the accuracy of each generated result. The core part of CDAS commands overall arrangements to process and mon- itor the human tasks to satisfy user required accuracy. Liu, Luo, and Li (2013) propose a fancy model, called robust per- sonal classifier (RPC), to improve robustness in crowdsourcing pro- cess. The model can create an expertise fraction for each worker automatically, which reflects the intrinsic quality of each worker. The final component of RPC model raises weights for good workers and reduces weights for malicious workers or spammers, which is more proper than same weights for all workers. Some models and patterns can avoid cheating workers, elimi- nate irrelevant answers and improve quality. For example, Chilana, Ko, and Wobbrock (2012) identify malicious workers by adding ground truth problem and calculating the differences of global labor rates through an economical model. Chen, Wu, Chang, and Lei (2009) utilize a probabilistic choice model to remove controversial inputs by checking individual consistency and overall consistency of workers. Bernstein et al. (2010) create Find-Fix-Verify crowd programming pattern to separate tasks into three phases to improve the quality. First, the system recruits one set of workers to find underlying areas for improvement. Then it collects a set of possible improvements, and finally filters out incorrect candidates. As currently employed methods have failed in filtering fraud, Almendra & Schwabe (2009) apply crowdsourcing to improve pre- cision and recall of fraud detection techniques for online trading sites. The experiment showed that workers could distinguish fraudsters from honest sellers precisely and rapidly, according to the personal profiles. Hirth, Hoßfeld, & Tran-Gia (2010) propose two methods to detect cheating workers based on crowdsourcing: a majority deci- sion (MD) approach and a control group (CG) approach to cross 7992 G. Xintong et al. / Expert Systems with Applications 41 (2014) 7987–7994 check the main task. MD is used to eliminate incorrect results. They hire redundant workers to finish the same tasks and compare the results. The majority of the results are considered to be correct. For CG, a single worker works on a main task and a control group consisting of several other workers re-checks the result. There is an assumption that the re-check task has different costs with the main task. Usually the main task is considered to be expensive, while the re-check task is cheap. If the majority of the control group reaches an agreement, the task is supposed to be correctly done. 5.6. Discussion about applicable occasion In the previous parts, we summarize some tools to control the mining quality. Here we will discuss which quality control mecha- nism is suitable for which type of mining mechanism. For classification and clustering tasks, voting and redundant workers are both suitable. And they are easy to implement. If the task needs more field-specific knowledge, we can apply gold stan- dard questions in the process of the crowdsourcing. The gold stan- dard answers are coming from experts in that domain. For semi-supervised learning, the workers that we need must be able to induct characters from labeled data, which means they have strong ability to learn fast. Hence worker’s reputation is a proper reference. The reputation system can reflect the history where this person has done some relevant works before. For mining the association rule, it is usually about mining the life habits, which are difficult to judge whether they are true or just fake. Voting seems to be an effective way. In this situation, using redundant workers is not a good choice because people differ from behaviors. The previous experience in working with data mining task on crowdsourcing platform highlights the importance of quality con- trol. Quality control phase can present available information better and serve as an important tool allowing companies and organiza- tions to get what they want most. 6. When not to use crowdsourcing in data mining The performance of crowdsourcing for data mining is praised in a number of situations, including social hotspot tracking, disaster relief, and homeland security. However, we have to admit that crowdsourcing is not a panacea for solving tasks in all data mining situations. It may lead to poor efficiency and low-quality result if we always adopt the idea of crowdsourcing without thinking over the feasibility. We should consider whether crowdsourcing could achieve better results than traditional algorithms for the current situation before deciding to use crowdsourcing. Barbier et al. (2012) summarized several scenarios when we cannot use crowdsourcing to solve data mining task. They are listed as follows: 1. The necessity of specific background knowledge. If the back- ground knowledge is too specific and not known by most of people, then using crowdsourcing to do data mining task may not be the proper approach. 2. Improper definition of the problem. Before applying crowd- sourcing, we must clarify a single target. If the problem has multiple targets, it’s better to separate them into a series of sub-problems. 3. The need of long-term dedication. The most prominent example is software development. Requester could not guarantee that the workers do not leave from the beginning to the end. Crowdsourcing is a highly mobile form which is not suitable for long-term dedication. 7. Future work There are still many challenges remaining to be solved. 1. Adaptive question delivering system A real time system that wisely delivers tasks to workers is of great importance. A wise system chooses appropriate next question to ask according to the previous answer that worker provides. It maximums the information gathered from crowd. Not all HITs are equal, and we want to gain more information from the special and ‘‘professional’’ people. Thus different HITs should be allocated dif- ferent questions according to the level of their knowledge. Adaptive question delivering system targets the problem of reducing uncer- tainty of the data and saving total cost on both time and money. 2. Recommendation framework We may design several recommendation frameworks to improve availability of the system. When designing a task, organi- zations should consider lots of factors, which include, but not limit to, precision, privacy, budget, and priority. Sometimes, a trade-off between these factors is the top concerned problem for the whole mining process. The recommendation frameworks guide request- ers to design crowdsourcing task based on their research target and the expected cost. 3. Specific mining algorithms We need more efforts on specific data mining algorithm design for crowdsourcing. Existing applications and services concentrate on building platforms for crowdsouricng. Algorithms used on crowdsourcing should not be the simple transplant of the ones in being. As far as we know, few efforts have been taken on specific data management and algorithm design for the new data mining process. We should consider the time delay, balance between pre- cision and cost, task design and many other attributes when deli- ver data mining task on crowdsourcing platform. For example, when posting an image clustering task on the MTurk, it is very common to ask questions in several iterations and wait for feed- back from workers. The feedback time is influenced by the task dif- ficulty and complexity. Obviously, more questions will lead to better result. But we intend to ask questions as few as possible to save money. Actually, crowdsourcing itself, as a booming social network application, creates important and precious knowledge during the interactions. And we can use the knowledge to design algorithms that are specialized for the crowdsourcing. 4. New quality guaranteed schemes The data derived from crowdsourcing process is often noisy and incomplete. Quality must be controlled before working, during working and after working. It is convinced that strict filtering based on task design is a potential method, because the experiment shows the type of the work has great impact on the quality. Creative works will attract less malicious workers. Constant optimization of the worker’s reputation system, instead of a simple prior acceptance rate, is another line of thought to protect the quality of crowdsourc- ing. What is more, a better understanding of human’s behavior will be beneficial to improve reliability of the system. 5. Scalability of data mining tasks Few works has been done to consider the scalability of crowdsourcing for large scale data mining. Actually, the power of G. Xintong et al. / Expert Systems with Applications 41 (2014) 7987–7994 7993 crowdsourcing is fully embodied in managing large datasets. To efficiently harnessing the floods of information will become a great challenge as the scale increases, especially when the information needs to be gathered in some logic sequences. 8. Conclusion This paper reviews recent work on crowdsourcing-based data mining techniques. Crowdsourcing can do data mining and extract addition information from the datasets more efficiently and intel- ligently than traditional methods. It has to deal with lots of chal- lenges like the low quality of answers from the crowds to apply crowdsourcing to data mining. In this paper, we point out these challenges and introduce the general procedures of an integrated data mining task in crowdsourcing. The task is often partitioned into three phases: question designing, data mining and quality controlling. We take a deep overview of work in each phase and conclude their contributions. Besides those discussed in Section 7, there are some promising directions for future research, such as designing crowdsourcing platform especially for governments or companies, algorithms in various steps to process large scale mining task, the background system for requesters to perform detailed analysis. Another future research task is to apply crowdsourcing to discovery knowledge for real expert systems with specific applications. Acknowledgments This paper was partially supported by NGFR-China 973 Grant 2012CB316200, NSFC-China Grant 60933001, 61003046, 61111130189 and NGFR-China 863 Grant 2012AA011004. References Agarwal, N., Liu, H., Tang, L., & Yu, P. S. (2008). Identifying the influential bloggers in a community. In Proceedings of the 2008 international conference on web search and data mining (pp. 207–218). ACM. Allahbakhsh, M., Ignjatovic, A., Benatallah, B., Beheshti, S. M. R., Bertino, E., & Foo, N. (2012). Reputation management in crowdsourcing systems. In 2012 8th international conference on collaborative computing: networking, applications and worksharing (CollaborateCom) (pp. 664–671). IEEE. Almendra, V., & Schwabe, D. (2009). Fraud detection by human agents: A pilot study. In E-commerce and web technologies (pp. 300–311). Berlin Heidelberg: Springer. Alonso, O., & Lease, M. (2011). Crowdsourcing for information retrieval: Principles, methods, and applications. In Proceedings of the 34th international ACM SIGIR conference on research and development in information retrieval (pp. 1299–1300). ACM. Amsterdamer, Y., Grossman, Y., Milo, T., et al. (2013b). CrowdMiner: Mining association rules from the crowd. Proceedings of the VLDB Endowment, 6(12). Amsterdamer, Y., Grossman, Y., Milo, T., & Senellart, P. (2013a). Crowd mining. In Proceedings of the 2013 international conference on management of data (pp. 241–252). ACM. Barbier, G., Zafarani, R., Gao, H., Fung, G., & Liu, H. (2012). Maximizing benefits from crowdsourced data. Computational and Mathematical Organization Theory, 18(3), 257–279. Bernstein, M. S., Brandt, J., Miller, R. C., & Karger, D. R. (2011). Crowds in two seconds: Enabling realtime crowd-powered interfaces. In Proceedings of the 24th annual ACM symposium on user interface software and technology (pp. 33–42). ACM. Bernstein, M. S., Little, G., Miller, R. C., Hartmann, B., Ackerman, M. S., Karger, D. R., et al. (2010). Soylent: A word processor with a crowd inside. In Proceedings of the 23nd annual ACM symposium on user interface software and technology (pp. 313–322). ACM. Bernstein, M. S., Teevan, J., Dumais, S., Liebling, D., & Horvitz, E. (2012). Direct answers for search queries in the long tail. In Proceedings of the SIGCHI conference on human factors in computing systems (pp. 237–246). ACM. Boim, R., Greenshpan, O., Milo, T., Novgorodov, S., Polyzotis, N., & Tan, W. C. (2012). Asking the right questions in crowd data sourcing. In 2012 IEEE 28th international conference on data engineering (ICDE) (pp. 1261–1264). IEEE. Brabham, D., Sanchez, T., & Bartholomew, K. (2009). Crowdsourcing public participation in transit planning: preliminary results from the next stop design case. Transportation Research Board. Bragg, J., & Weld, D. S. (2013). Crowdsourcing multi-label classification for taxonomy creation. In First AAAI conference on human computation and crowdsourcing. Budde, A., & Michahelles, F. (2010). Towards an open product repository using playful crowdsourcing. In G. I. Jahrestagung (Ed.), (Vol. 1, pp. 600–605). Callison-Burch, C., & Dredze, M. (2010). Creating speech and language data with Amazon’s mechanical Turk. In Proceedings of the NAACL HLT 2010 workshop on creating speech and language data with Amazon’s mechanical Turk (pp. 1–12). Association for Computational Linguistics. Chen, Q., Wang, G., & Tan, C. L. (2012). Web image organization and object discovery by actively creating visual clusters through crowdsourcing. 2012 IEEE 24th international conference on tools with artificial intelligence (ICTAI) (Vol. 1, pp. 419–427). IEEE. Chen, K. T., Wu, C. C., Chang, Y. C., & Lei, C. L. (2009). A crowdsourceable QoE evaluation framework for multimedia content. In Proceedings of the 17th ACM international conference on multimedia (pp. 491–500). ACM. Chilana, P. K., Ko, A. J., & Wobbrock, J. O. (2012). LemonAid: Selection-based crowdsourced contextual help for web applications. In Proceedings of the 2012 ACM annual conference on human factors in computing systems (pp. 1549–1558). ACM. Chilton, L. B., Little, G., Edge, D., Weld, D. S., & Landay, J. A. (2013). Cascade: Crowdsourcing taxonomy creation. In Proceedings of the 2013 ACM annual conference on human factors in computing systems (pp. 1999–2008). ACM. Eickhoff, C., & de Vries, A. (2011). How crowdsourceable is your task. In Proceedings of the workshop on crowdsourcing for search and data mining (CSDM) at the fourth ACM international conference on web search and data mining (WSDM) (pp. 11–14). Gao, H., Barbier, G., & Goolsby, R. (2011). Harnessing the crowdsourcing power of social media for disaster relief. IEEE Intelligent Systems, 26(3), 10–14. Gomes, R. G., Welinder, P., Krause, A., & Perona, P. (2011). Crowdclustering. In Advances in neural information processing systems (pp. 558–566). Goolsby, R. (2010). Social media as crisis platform: The future of community maps/crisis maps. ACM Transactions on Intelligent Systems and Technology (TIST), 1(1), 7. Heimerl, K., Gawalt, B., Chen, K., Parikh, T., & Hartmann, B. (2012). CommunitySourcing: Engaging local crowds to perform expert work via physical kiosks. In Proceedings of the 2012 ACM annual conference on human factors in computing systems (pp. 1539–1548). ACM. Hirth, M., Hoßfeld, T., & Tran-Gia, P. (2010). Cheat-detection mechanisms for crowdsourcing. University of Würzburg, Tech. Rep, 4. Ipeirotis, P. (2010). Be a top mechanical Turk worker: You need $5 and 5 minutes. <http://behind-the-enemy-lines.blogspot.com/2010/10/be-top-mechanical-turk- worker-you-need.html>. Ipeirotis, P. G., Provost, F., & Wang, J. (2010). Quality management on Amazon mechanical Turk. In Proceedings of the ACM SIGKDD workshop on human computation (pp. 64–67). ACM. Karger, D. R., Oh, S., & Shah, D. (2011). Iterative learning for reliable crowdsourcing systems. In NIPS (pp. 1953–1961). Kazai, G., Kamps, J., Koolen, M., & Milic-Frayling, N. (2011). Crowdsourcing for book search evaluation: Impact of hit design on comparative system ranking. In Proceedings of the 34th international ACM SIGIR conference on research and development in information retrieval (pp. 205–214). ACM. Le, J., Edmonds, A., Hester, V., & Biewald, L. (2010). Ensuring quality in crowdsourced search relevance evaluation: The effects of training question distribution. In SIGIR 2010 workshop on crowdsourcing for search evaluation (pp. 21–26). Libert, B., & Spector, J. (2007). We are smarter than me: How to unleash the power of crowds in your business. Wharton School Publishing. Liu, X., Lu, M., Ooi, B. C., Shen, Y., Wu, S., & Zhang, M. (2012). Cdas: A crowdsourcing data analytics system. Proceedings of the VLDB Endowment, 5(10), 1040–1051. Liu, Z., Luo, L., & Li, W. J. (2013). Robust crowdsourced learning. In 2013 IEEE international conference on big data (pp. 338–343). IEEE. Markowsky, G. (2013). Crowdsourcing, big data and homeland security. In 2013 IEEE international conference on technologies for homeland security (HST) (pp. 772–778). IEEE. Quinn, A. J., & Bederson, B. B. (2011). Human computation: A survey and taxonomy of a growing field. In Proceedings of the SIGCHI conference on human factors in computing systems (pp. 1403–1412). ACM. Reddit website dedicated to solving the Boston Marathon Bombing, no longer a public site (0000). <http://www.reddit.com/r/findbostonbombers>. Sorokin, A., & Forsyth, D. (2008). Utility data annotation with Amazon mechanical Turk. Urbana, 51(61), 820. Spenser Ackerman. (2013). Data for the Boston Marathon Investigation Will Be Crowdsourced, Wired, April 16. <http://www.wired.com/dangerroom/2013/04/ boston-crowdsourced>. Tang, W., & Lease, M. (2011). Semi-supervised consensus labeling for crowdsourcing. In SIGIR 2011 workshop on crowdsourcing for information retrieval (CIR). Tran-Thanh, L., Venanzi, M., Rogers, A., & Jennings, N. R. (2013). Efficient budget allocation with accuracy guarantees for crowdsourcing classification tasks. In Proceedings of the 2013 international conference on autonomous agents and multi-agent systems (pp. 901–908). International Foundation for Autonomous Agents and Multiagent Systems. Venetis, P., & Garcia-Molina, H. (2012). Quality control for comparison microtasks. In Proceedings of the first international workshop on crowdsourcing and data mining (pp. 15–21). ACM. http://refhub.elsevier.com/S0957-4174(14)00398-4/h0005 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0005 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0005 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0010 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0010 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0010 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0010 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0015 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0015 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0015 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0020 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0020 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0020 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0020 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0025 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0025 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0030 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0030 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0030 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0035 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0035 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0035 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0040 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0040 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0040 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0040 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0045 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0045 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0045 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0045 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0050 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0050 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0050 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0055 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0055 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0055 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0075 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0075 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0075 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0075 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0080 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0080 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0080 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0080 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0085 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0085 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0085 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0090 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0090 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0090 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0090 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0095 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0095 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0095 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0105 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0105 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0115 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0115 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0115 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0130 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0130 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0130 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0130 http://behind-the-enemy-lines.blogspot.com/2010/10/be-top-mechanical-turk-worker-you-need.html http://behind-the-enemy-lines.blogspot.com/2010/10/be-top-mechanical-turk-worker-you-need.html http://refhub.elsevier.com/S0957-4174(14)00398-4/h0155 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0155 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0155 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0165 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0165 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0165 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0165 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0175 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0175 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0180 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0180 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0185 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0185 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0190 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0190 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0190 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0195 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0195 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0195 http://www.reddit.com/r/findbostonbombers http://refhub.elsevier.com/S0957-4174(14)00398-4/h0205 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0205 http://www.wired.com/dangerroom/2013/04/boston-crowdsourced http://www.wired.com/dangerroom/2013/04/boston-crowdsourced http://refhub.elsevier.com/S0957-4174(14)00398-4/h0215 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0215 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0215 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0215 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0215 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0220 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0220 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0220 7994 G. Xintong et al. / Expert Systems with Applications 41 (2014) 7987–7994 Von Ahn, L. (2009). Human computation. In 46th ACM/IEEE design automation conference, DAC’09 (pp. 418–419). IEEE. Von Ahn, L., & Dabbish, L. (2008). Designing games with a purpose. Communications of the ACM, 51(8), 58–67. Von Ahn, L., Maurer, B., McMillen, C., Abraham, D., & Blum, M. (2008). Recaptcha: Human-based character recognition via web security measures. Science, 321(5895), 1465–1468. Weaver, A. C., Boyle, J. P., & Besaleva, L. I. (2012). Applications and trust issues when crowdsourcing a crisis. In 2012 21st international conference on computer communications and networks (ICCCN) (pp. 1–5). IEEE. Wright, A., Chen, E. S., & Maloney, F. L. (2010). An automated technique for identifying associations between medications, laboratory results and problems. Journal of Biomedical Informatics, 43(6), 891–901. Zhai, Z., Kijewski-Correa, T., Hachen, D., & Madey, G. (2012). Haiti earthquake photo tagging: Lessons on crowdsourcing in-depth image classifications. In 2012 seventh international conference on digital information management (ICDIM) (pp. 357–364). IEEE. 4chan website dedicated to solving the Boston Marathon Bombing, http://imgur.com/a/sUrnA. http://refhub.elsevier.com/S0957-4174(14)00398-4/h0225 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0225 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0230 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0230 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0235 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0235 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0235 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0240 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0240 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0240 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0245 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0245 http://refhub.elsevier.com/S0957-4174(14)00398-4/h0245 http://imgur.com/a/sUrnA Brief survey of crowdsourcing for data mining 1 Introduction 2 Framework 3 Question design 4 Mining data from crowdsourcing 4.1 Classification 4.2 Clustering 4.3 Semi-supervised learning 4.4 Sampling 4.5 Association rule mining 4.6 Validation 5 Quality control 5.1 Vote 5.2 Redundant work 5.3 Worker’s reputation 5.4 Gold standard 5.5 Other solutions for quality control 5.6 Discussion about applicable occasion 6 When not to use crowdsourcing in data mining 7 Future work 8 Conclusion Acknowledgments References 
work_qrnasltopndvzmirwb3wqlfrda	----	exsy648_LR Article DOI: 10.1111/j.1468-0394.2012.00648.x Business goals, user needs, and requirements: A problem frame-based view Luigi Lavazza Università degli Studi dell’Insubria, Dipartimento di Scienze Teoriche e Applicate, Varese, Italy Email: luigi.Lavazza@uninsubria.it Abstract: It is well known that the analysis of requirements involves several stakeholders and perspectives. Very often several points of view at different abstraction levels have to be taken into account: all these features make requirements analysis a complex task. Such intrinsic complexity makes it difficult to understand several of the basic concepts that underlie requirements engineering. Actually, there is some confusion – especially in industry – about what really a user requirement is, what are the differences between user requirements and user needs, and what are their relationships with business processes. The paper aims at clarifying the aforementioned issues, by providing a systematic and clear method for establishing requirements hierarchies. The problem of describing requirements hierarchies is tackled using the problem frames concepts and notation. A case study is used throughout the paper to illustrate the proposed approach. The description of requirements at different levels of abstractions and requirements hierarchies are illustrated. The resulting models are coherent with the reference model for requirements specifications and the problem frames. An analysis process that is aware of the differences between user needs and requirements is also provided, to illustrate the process of refining high-level goals into requirements that can be satisfied by a hardware/software machine. The proposed method appears promising to model, study, and evaluate the relationships between business processes and the strategies for achieving business goals based on the usage of information technology. Keywords: user needs, user requirements, problem frames, requirements analysis, knowledge elicitation 1. Introduction In the analysis and documentation of user requirements several perspectives are often involved. The multiplicity of scopes and points of view is reflected in the terminology: it is quite usual to hear about ‘business goals’, ‘user needs’, ‘user requirements’, etc. Actually, there is some confusion – espe- cially in industry – about what a user requirement really is, what the differences (if any) are between a user requirement and a user need, and what is their relationship with business processes. Moreover, how effectively and correctly to model all these aspects is not always clear. The result is that user requirements tend to encompass multiple different aspects of the systems to be developed, and to be represented in very different manners, often mixing requirements with other issues. The confusion of goals, needs, and requirements is problematic because the same goal can give place to different – often incompatible – sets of needs, and the same need can originate different requirements. Therefore, in proceeding from goals to requirements, evaluations have to be performed and choices made. Moreover, the ‘distance’ that separates goals from requirements is often great, thus ‘jumping’ directly from goals to requirements is generally infeasible in practice: intermediate refinement steps are required. Examples of these issues are given in Section 3. In this paper, the differences between user needs and user requirements are highlighted; the relation of user needs to business processes is also described; finally, the usage of prob- lem frames to model the mentioned issues is illustrated by means of a case study. More specifically, the paper aims on the one hand at clar- ifying the nature of requirements that are conceived and de- scribed at different levels (from the highest business-related goals to software requirements specifications), on the other hand at providing a description of requirements that is ho- mogeneous through the various levels. This description takes the form of a hierarchy of requirements and has a set of de- sirable characteristics: 1. It is coherent with previous work on goal-oriented re- quirements engineering (RE). 2. It is based on the reference model for requirements and specifications (Gunter et al., 2000). 3. It uses problem frames, thus enabling the usage of the concepts and methods described in Jackson (2001). 4. The definition of a homogeneous hierarchy of require- ments makes clear that the nature of requirements defi- nition is basically the same at different levels. 5. It connects high-level business-oriented analysis with Problem-Oriented Software Engineering (POSE) (Hall et al., 2008; Hall & Rapanotti, 2011), which typically focuses on software requirements. A process is defined to provide guidelines to analysts in applying the proposed concepts, and to make explicit that the process involves (a) refinements and (b) feasibility and cost evaluations, since problems often have several possible solutions, of which some could be very hard or expensive to achieve. The paper is organized as follows: the problem of require- ments modelling is introduced in Section 2. The involved C© 2012 Wiley Publishing Ltd Expert Systems, July 2013, Vol. 30, No. 3 215 Figure 1: The problem and solution spaces. issues are presented in Section 3 with the help of a case study. The distinction between needs and requirements is further discussed in Section 4, and an analysis process aware of this distinction is sketched in Section 5. Section 6 reports related work, while Section 7 draws some conclusions and provides an outline of future work. Throughout this paper, the concept and notation of prob- lem frames (Jackson, 2001) are used. The reader is expected to know them. 2. Requirements modelling Requirements modelling is of fundamental importance in the software development process. Among others, Jackson et al., have contributed a problem-oriented view of RE (Gunter et al., 2000; Jackson, 2001; Hall et al., 2008; Hall & Rapan- otti, 2011). The material reported in this section is mainly based on such work. 2.1. The nature of requirements In order to develop a software solution it is necessary to study and understand the problem. In other words, we need to understand where the problem is and where the solution is. Figure 1 schematically represents the world (i.e. the envi- ronment) where the problem exists, and the machine (i.e. the hardware/software solution we have to build to solve the problem). The relationships between the machine and the world are highlighted. In fact, it is fundamental that the ma- chine gets the information it needs about the world, and that is provides outputs (either information or control) to the surrounding environment. In order to define precisely the meaning of requirements modelling it is fundamental to notice that there is an Envi- ronment in which the problems exist, and where the machine (i.e. the hardware/software solution to be developed) will op- erate. The problem domain is the portion of the Environment that is visible by the machine. Of course, the Environment interacts with the machine, since a machine that does not get any input and does not provide any output would be hardly useful. The relationship between the Environment and the machine (which is seen as a black box in this phase) is de- scribed in terms of ‘phenomena’ that are shared between the Problem Domain and the Machine. These concepts are properly defined in the ‘Reference Model for Requirements and Specifications’ (Gunter et al., 2000). The model is based on five artefacts, some of which pertain mostly to the system, while others pertain mostly to the environment (Figure 2 (Gunter et al., 2000)). These artefacts are: Figure 2: The elements of the reference model for requirements and specifications. 1. Domain knowledge that describes environment facts; this artefact is denoted as W (for the World, which in- cludes the environment). 2. Requirements that indicate the desired situation that must be achieved in the environment as an effect of the introduction of the system in the environment; this arte- fact is denoted as R. 3. Specifications of the system, which provide enough in- formation for a programmer to build the system: specifi- cations do not need to communicate to the developer un- necessary information about the environment; this arte- fact is denoted as S. 4. A program that implements the specification using the programming platform; this artefact is denoted as P. 5. A programming platform that provides the basis for pro- gramming the system; this artefact is denoted as M (for Machine, which is the programmable platform). Of course, we should guarantee that the requirements are satisfied (i.e. we build the right system) and the program cor- rectly implements the specification (i.e. we build the system right). Concerning P and M, it is useful to note that in case of embedded software the distinction between P and M on one side and W and R on the other side is quite clear, as the interaction between the environment and the system occurs via sensors and actuators; thus, people belong to the envi- ronment. On the contrary, when traditional ‘business’ appli- cations are concerned, the ‘system’ usually includes people who use the software applications and execute ‘manual’ pro- cedures. In these cases, the ‘machine’ is actually composed by both a hardware/software platform and people, and the ‘pro- grams’ are partly software and partly rules and procedures that people have to follow. In 2005, Hall and Rapanotti ex- tend the reference model to explicitly represent the ‘social component’, that is the people, who can be ‘instructed’ – as machines are programmed – for the purpose of satisfying the requirements. The result is that a third ellipse is added to the model in Figure 2. Figure 2 (Gunter et al., 2000) shows that W and R belong to the environment, while M and P belong to the system. The specifications S, instead, characterize the intersection of the environment and the system, that is the elements that are shared by the environment and the system. In the reference model, designations identify classes of phenomena (states, events, individuals, etc.) in the system and the environment and assign names to them. Some phenomena (denoted as e in Figure 3, Gunter et al., 2000) belong to the environment, 216 Expert Systems, July 2013, Vol. 30, No. 3 C© 2012 Wiley Publishing Ltd Figure 3: Environment’s and system’s phenomena: control and visibility. which controls them. Phenomena denoted as s belong to the system. Some of the e phenomena are visible to the system, while others are not: the former are denoted as ev, while the latter are denoted as eh (e = eh ∪ ev). The s phenomena are similarly classified as sv and sh. Phenomena eh, ev, and sv are visible to the environment and used in W and R. Phenomena sh, sv, and ev are visible to the system and used in P and M. Only ev and sv are visible to both the environment and the system: specifications S are defined only in terms of ev and sv. How can we define and describe a problem in order to enable the development of a software-based solution (or just to evaluate if a software-based solution is at all possible)? We have to keep in mind that developers have to be given all the relevant information concerning the desired behaviour of the system: this information constitutes the requirements of the software to be developed. Therefore, requirements are the criteria, conditions, or constraints that a software artefact has to satisfy in order to be acceptable as a solution of the given problem. This conception of requirements implies that: 1. Requirements are usually expressed with reference to the environment phenomena only. Users know well their environment, while they often do not have a clear idea of what computers and software can provide. Therefore they tend to describe the situation they want to achieve, independently of the fact that the result is obtained via a computer-based solution or not. 2. The effects of the system on the environment depend not only on sv, but also on the reactions that these phenom- ena generate. It is thus necessary that the rules that de- termine the behaviour of the environment are described. These descriptions belong typically to W and are ex- pressed in terms of eh. 3. The goal of the whole software development effort is the construction of a machine having specifications S such that its interaction with the environment causes the requirements to be satisfied. The latter issue is typically summarized in the formula: W, S � R Figure 4: A problem diagram. which states that the satisfaction of the requirements follows from the combination of the world and machine’s behaviour. The task of software analysts is to define S such that, given W and R, W, S � R holds. The task of designers, programmers, testers, etc., is to build a program P such that, given M and S, M, P � S holds. If both the analysts and the implementers accomplish such tasks, then both W, S � R and M, P � S hold, thus we get that also W, M, P � R holds, that is the goal of the development is achieved. Problems are generally described by means of problem di- agrams (Jackson, 2001), like the one in Figure 4. Problem diagrams show the domains involved in the problem (i.e. the domains belonging to the environment), the machine that is expected to solve the problem, possible designed do- mains (i.e. domains whose structure and content can be – to some extent – designed, as opposed to environment do- mains, that are given and cannot be modified). Domains are shown as boxes in the diagrams. Interfaces between domains are represented as lines connecting the domain boxes. In- terfaces indicate that some ‘phenomena’ (e.g. values, states, signals, etc.) are shared between the connected domains. Sets of shared phenomena are labelled with letters, whose meaning is explained separately: for instance, in Figure 10 c: SP!{store_operation} indicates that the domain whose ini- tials are SP (storage processor) controls a phenomenon ‘store operation’ that is shared with domain Repository. Problem diagrams also show the requirements in terms of properties that are not given, but have to be achieved by means of a proper machine. Requirements are given in terms of rela- tionships between problem domains. 2.2. The role of analysts As mentioned above, the user tends to talk about what the problem is in the real world, not about the role of the machine in solving it. In fact, the user knows the problem well at a quite high, business-related level. This level of understanding of the business problem provides a starting point for business analysis, which leads to a better understanding of the user’s problem and to a more complete and detailed description of the problem, especially concerning the problem environment. Usually the user does not have a clear idea of the role of the machine in solving his/her problem. In fact, the user often C© 2012 Wiley Publishing Ltd Expert Systems, July 2013, Vol. 30, No. 3 217 has not the slightest clue of how to solve the problem, or, even worse, has completely wrong ideas of how to solve the problem, because he/she does not know what is easy or hard to achieve by means of a computer system. In general, even business analysis does not yield a final indication about the role of software in providing a solution to the given problem. This is due to the fact that business problems allow for several possible solutions, characterized by different costs and effectiveness. Therefore, requirements have to be proactively defined on the basis of the description – provided by business analysis – of the problem and the environment where the machine will have to operate. RE is thus the creative activity that leads to defining the effects that the hardware/software machine will have on the problem domain so that the problem is solved. To this end, it is necessary to remember that it is the behaviour of the system encompassing the domain and the machine that has to solve (or at least to reduce) the problems that induced the user to ask for a computer-based solution. Accordingly, the analyst has a double role in describing requirements: 1. He/she must understand the needs of the user and the nature and behaviour of the environment. 2. He/she must (pro)actively cooperate with the user in defining requirements that are effective (i.e. the machine actually contributes to solve the problem), technically feasible, and reasonable with respect to economic con- straints. The final product of the analyst is the requirement specifi- cations. The goals of the requirement specifications are: 1. to describe what are the responsibilities of the system in satisfying the user’s needs. 2. To define what are the responsibilities of the environment in supporting and using the hardware/software system. For instance, the environment has to provide input data and to understand the machine outputs. Therefore, the analyst cannot be a passive recorder of the user’s wishes. Instead, he/she has to be active, make propos- als, and explore alternatives. In fact, as described in Section 3, the responsibilities of the system and the interaction between the machine and environment are not always determined uni- vocally by the user needs. These observations partly contradict what was written in Section 2.1, since R are not just given, but are defined by the user and the analyst together. However, this conclusion leads to a new question, since the user surely has some needs or goals that are not negotiable (e.g. because they are funda- mental for achieving a competitive advantage), as opposed to the concrete formulation of requirements, which is actu- ally established with the analyst. In fact, we said that users have ‘computer-independent’ needs that are typically related to their business and that cannot (or at least should not) be changed by the introduction of computer-based systems. Therefore, we would like to be able to answer questions like the following: 1. What are the relationships between non-negotiable user needs and goals on one side and requirements on the other? Figure 5: Problem diagram representing requirements at the business goal level. 2. Is the knowledge about the world that appears in W, S � R the same that leads to the formulation of non- negotiable user needs and goals? 3. How should we deal with the part of the world whose behaviour is not fixed once and for all, but can be to some extent adapted to the user needs? 3. Case study In the rest of the paper, the proposed concepts are illustrated by means of an example, described in this section. Let us start with the high-level needs of the ‘user’. In our case study, a company needs that its agencies act in a coordi- nated way, in order to maximize efficiency, avoid work dupli- cations, avoid inconsistencies, etc. The case study reported here is a simplification of a real-world problem, where a lo- gistic company had to keep track of container movements managed by local agencies in order to optimize freight plan- ning. We can classify these needs as part of the ‘business goals’ (BG) of the company (throughout the paper the term ‘busi- ness goal’ is used as a shorthand for ‘business-originated user goals’). Figure 5 represents the business goal by means of a prob- lem diagram. Since the goal of the company is clearly in- dependent of the existence of a computer-based system, a machineless problem diagram illustrates the goal. The ‘re- quirement’ Coherent activity is actually the BG statement that agencies have to act in a coordinated way, that is a and b must be coherent: for instance, the action performed at agency 1 – denoted by phenomenon a – must be coherent with the actions – denoted by phenomenon b – performed at agency 2. Of course, the domain must be capable of the expected coherent behaviour among all others possible be- haviours; the solution is thus to make sure that the expected behaviour actually occurs. Phenomena a and b denote the in- formation needed to specify the coherency of activities. What does actually mean that a and b must be coherent depends on the detailed specification of the properties of a and b, and on the definition of ‘behaviour coherency’: these details are not relevant for our purposes. Now we have to understand how the business goal can be actually achieved. To this end, we rewrite the diagram intro- ducing a machine (there must be a machine, otherwise the goals could only be achieved by a ‘spontaneous’ behaviour of the environment, but if such spontaneous behaviour existed, no needs and goals would arise . . . ). 218 Expert Systems, July 2013, Vol. 30, No. 3 C© 2012 Wiley Publishing Ltd The “machine”The “machine” Figure 6: Problem diagram representing business goal-level requirements. Figure 6 describes the same problem as Figure 5, except that we have introduced the part of the system that is re- sponsible for the achievement of the business goals. Since at the moment it is not yet known whether the solution will be based on a computer, the solution is represented in a cloud, which possibly contains some sort of machine (no reference to cloud computing here!). In fact, the stakeholder(s) – typ- ically belonging to top management – who expressed the business goal may have no clue about this part of the system. In order to make clear that the part of the system in the cloud is actually part of the solution, we can write the usual formula: WBP, NB � BG (1) where BG is the business goal, WBP describes the business process environment, rules, etc., and NB specifies the be- haviour of the part of the system (possibly including a ma- chine) that will cause the BG to be satisfied. Formula (1) above states that the business goal BG (agencies have to act in a coordinated way) is satisfied, in the environment de- scribed by WBP, by some ‘machine’ described by NB. Note that at this level, the ‘machine’ is not necessarily a hardware machine equipped with suitable software. Rather, it can be a sort of mechanism – typically involving people – which ‘en- acts’ the business processes. In our case study, WBP specifies that the agencies store a set of business-relevant information, upon which they take decisions and operate. Moreover, WBP guarantees that agencies act in a coordinated way if they operate according to the same information concerning the business. It is easy to conclude that the business need (NB) is that information concerning the business has to be shared. In other words: if we perform some actions in the business environment that guarantees that the business information is shared, the coordination goal is achieved. Of course, this brings to a new problem: now we have to decide how the in- formation sharing need can be actually achieved. To this end, we redraw the diagram of Figure 6 as reported in Figure 7. The diagram in Figure 7 still features the cloud, since the nature of the solution is still not known. The requirements NB state that globally interesting information available at agency 1 (a) is available also at agency 2 (b), and vice versa. For the sake of simplicity, in the rest of the paper we always consider agency 1 as the originator of the relevant information and agency 2 as the place where the relevant information must be reproduced. Of course, dealing with the case when agency 2 originates the relevant information, or when multiple agen- cies are involved does not pose conceptual difficulties. The problem described in Figure 7 can be formalized as follows (note that NB, which was the ‘solution’ in (1) is the goal in (2)): W ′, R � NB W ′ = W ∪ WBP (2) where R describes the role of the solution, while W′ is an extension of WBP, since it is likely that in dealing with this new problem we need additional information concerning the environment than was provided for the original problem (1). Figure 7: Problem diagram representing business needs. C© 2012 Wiley Publishing Ltd Expert Systems, July 2013, Vol. 30, No. 3 219 Figure 8: Problem diagram representing business needs according to (2). It is clear that in order to have the globally interesting information available at all agencies, it is necessary to com- municate the information. Thus, W specifies that the agencies are connected via a communication infrastructure. Accord- ingly, a possible specification of R states that the information generated at agency 1 is communicated – via the communi- cation infrastructure – to agency 2, and that the received information is stored locally. The fact that W′ includes a communication infrastructure can be taken into account as in Figure 8. In order to tackle the new problem described in Figure 8 (W ′, R � NB) we need to define W′ and R. To this end, we need to say what is in the cloud, that is what communica- tion infrastructure is available. Of course, several possibilities exist: 1. The agencies are already connected, maybe using rela- tively old technology (e.g. plain email) and we are con- strained to achieve the goal using the available connec- tion. In this case, the communication infrastructure is part of the given environment. 2. The agencies are not connected, or the management de- cided that a new connection can be built anyway. In this case, the infrastructure is part of the solution. At this stage of the analysis process, the analyst should: 1. Evaluate the environment, in order to highlight its rele- vant characteristics. In our case, the communication in- frastructure is the relevant characteristic to be identified and described. 2. Sketch possible solutions, some of which will be based on the environment as is, while others will involve chang- ing the environment, possibly introducing new elements. For instance, if the environment does not contain com- munication infrastructures, the analyst can think of in- troducing one (maybe sketching different new scenarios, each characterized by a different communication infras- tructure). 3. Evaluate which ones of the identified solutions (and re- lated environment) are viable. In our case study we assume that the analyst has performed the tasks mentioned above, and has found that the agencies are connected via plain email. This situation is described in Figure 9 (from now on, the domains are no longer labelled as R, W, N, etc., to avoid overloading the pictures). We also suppose that the analysis of the possible solu- tions has led to the decision that the business goal should be achieved using the available connections. In this case, the communication infrastructure is part of the given environ- ment. Having chosen the email-based infrastructure, W′ specifies the usual functions of the email system (server and clients), while R states that: 1. Any interesting information stored at an agency is sent to the other agencies (W′ assures that the communication is carried out properly). 2. Any received information is stored in the repository. Figure 9: Problem diagram representing business needs and details of the environment. 220 Expert Systems, July 2013, Vol. 30, No. 3 C© 2012 Wiley Publishing Ltd Figure 10: Problem diagram representing the work done at agency 1. These specifications of W′ and R assure that the business need NB involving information sharing is satisfied. Therefore, the problem is now how to satisfy R. As is often the case, there are several ways for satisfying R. In the next sections three possible ways of satisfying R are illustrated. Before proceeding to see how R can be satisfied, it is useful to note that the nature of the problem and the availability of the communication infrastructure suggest that the problem is split into two sub-problems: 1. The sub-problem at agency 1 requires that the locally generated (and locally stored) information is sent to agency 2. 2. The sub-problem at agency 2 requires that the informa- tion received from agency 1 is stored locally. In principle, both sub-problems can be solved in different ways. However, for the sake of simplicity, here we explore only the alternatives for agency 2. 3.1. Email-based case: sub-problem 1 (information generation) The main work done at agency 1 (as far as we are concerned) consists of generating and storing information. This activity is described by the problem diagram in Figure 10. The data storage requirements simply state that the stored information (d) is coherent with the data provided by the operator (b). Figure 12: Problem diagram representing the problem at agency 2. The information stored locally has to be transmitted to agency 2. This requirement is modelled in the problem dia- gram reported in Figure 11. The data transmission requirements state that whenever a new piece of information (c) appears in the repository, an email message (a) containing the same information is sent to agency 2. As already mentioned, there are several machines (the communication processor in Figure 11) that can satisfy these requirements. For instance: 1. The operator sends manually an email whenever he/she inputs new data in the repository; 2. the data input processor, upon receiving data from the operator, both stores it in the repository and sends an email to agency 2 containing the same information; 3. a specific program periodically checks the repository, detects new information and emails such information to agency 2. We do not explore in detail the solutions for agency 1. In- stead, we concentrate on the problem at agency 2, described in Figure 12. This problem diagram finally introduces the machine (the update processor) that is responsible for satis- fying the data transmission requirements. The satisfaction of these requirements, together with the satisfaction of the re- quirements for the problem at agency 1, guarantee that needs NB in equation (2) are satisfied. Figure 11: Problem diagram representing the communication requirements for agency 1. C© 2012 Wiley Publishing Ltd Expert Systems, July 2013, Vol. 30, No. 3 221 Here we have to take into account that the environment is not perfect. The description of the environment (W) states that: 1. Email messages (b) contain the communicated informa- tion (a). 2. However, it is possible that – because of some error – the message (b) does not contain comprehensible informa- tion. The requirements data update requirements (R) state that: 1. The communicated information (a) is stored in the repos- itory; that is d = a. 2. Exception: if the content of the received email mes- sage (EC!{received_email_message}) is unclear, a clar- ification request (DIO!{request_message}) is sent to the originating agency. The different possible ways of specifying the update pro- cessor machine are described in Sections 3.2–3.4. In fact, for the problem in Figure 12 we have to assure that W, S � R is true. There are several possible machines whose specifica- tions make W, S � R true. In the following sections we shall see three machine specifications that cause R to be satisfied. 3.2. Email-based case: scenario 1 In this scenario, the update processor is made of the same data input processor available at agency 1 (i.e. the machine that is used to store useful information in the local reposi- tory) and a data input operator. The latter actually behaves as the ‘program’ of the update processor, that is he/she de- termines the behaviour of the ‘machine’ that guarantees the satisfaction of the data input requirements. In practice, this means that no specific software has to be written; instead, a manual procedure has to be established. R: the requirements data input requirements state that: 1. The communicated information (a) is stored in the repos- itory, i.e. f = a. 2. Exception: if the content of the received email mes- sage (EC!{received_email_message}) is unclear, a clar- ification request (DIO!{request_message}) is sent to the originating agency. W: the environment description states that: 1. Email message (EC!{received_email_message}) contains the communicated information (a). However, it is possi- ble that – because of some error – the message does not contain comprehensible information. 2. The data input processor and Repository guarantee that the stored information (f) is coherent with the input in- formation (d). S: specifications 1. Email messages (EC!{received_email_message}) are read by the operator, who feeds manually the informa- tion (c) to the system via suitable forms (d). 2. Exception: if the content of the received email mes- sage is unclear, the operator sends a clarification request (DIO!{request_message}) to the originating agency. It is easy to see that a ‘machine’ implementing specifica- tions S satisfies requirements R. The data input operator is essential: he/she is part of the ‘machine’. This solution is clearly suitable if the information to be shared is generated rather infrequently. In this case, the work of the operator avoids the need of building an ad-hoc computer-based system. However, if the frequency of incoming email messages be- comes sufficiently high, the operator could have to perform a large amount of work, thus making this solution exces- sively expensive. Moreover, being the solution completely demanded to the operator, it is error prone. These consider- ations suggest that other solutions are explored. 3.3. Email-based case: scenario 2 In this scenario we consider the automation of the work that in scenario 1 was done manually by the operator. Accord- ingly, in Figure 14, the machine (still named data input pro- cessor) is connected directly to the email server, from which it retrieves the incoming email messages. In case the content of the message is not clear, the data input processor issues a warning on the console of the data input operator, who can either provide the correct interpretation of the received data to the processor, or send a clarification request to agency 1. The specification S of the update processor is as follows (note that the ‘machine’ involves also a manual procedure, as in scenario 1, although in this case it is just for solving unclear messages): 1. Email messages (b2) are read by the data input processor, which automatically stores the communicated informa- tion (a) to the repository. 2. Diagnostics (d) are issued when the informative content of email messages is unclear. 3. The operator reads diagnostics (d) and, if able to inter- pret the message, provides to the data input processor the information (d) to be stored, otherwise it sends an email message (b1) requiring a correct input. The machine specified in this scenario is clearly more suit- able for dealing with big volumes of email messages. How- ever, the solution is effective if only a few messages cannot be understood correctly by the machine; otherwise, most of the work would still have to be done manually. 3.4. Email-based case: scenario 3 In this scenario we consider the full automation of the pro- cedure; that is also the handling of unclear messages is per- formed automatically. Accordingly, in Figure 15 the machine update processor is connected directly to the email server, from which it retrieves the incoming email messages. In case the content of the mes- sage is not clear, the update processor sends a clarification request to agency 1 via email. The specification S of the update processor is as follows (note that the ‘machine’ no longer involves a manual proce- dure): 222 Expert Systems, July 2013, Vol. 30, No. 3 C© 2012 Wiley Publishing Ltd Figure 13: Problem diagram for agency 2: scenario 1. Figure 14: Problem diagram for agency 2: scenario 2. Figure 15: Problem diagram for agency 2: scenario 3. 1. Email messages (EC!{received_email_message}) are read by the processor, which automatically stores the communicated information (a) into the repository. 2. If unable to interpret the message the processor sends an email message (UP!{request_message}) requiring a correct input. The machine specified in this scenario is clearly suitable for dealing with big volumes of email messages, both correct and incorrect. 4. The relationship between needs and requirements As already mentioned, according to the reference model (Gunter et al., 2000) the relationship that involves the ma- chine, the environment, and the requirements is: W, S � R where W indicates the problem world properties; R indicates the requirements (what the customer needs from the system, described in terms of its effect on the environment); S are the specifications that provide enough information for a pro- grammer to build a machine that satisfies the requirements. This description allows for several machine specifications that satisfy the requirements. With this description, it is not immediate to distinguish the user needs, which lay completely in the environment, and the requirements, which involve, to some extent, the responsibility of the machine (in fact, the very satisfaction of requirements depends upon the machine). 4.1. Needs versus requirements In order to make the difference between needs and require- ments explicit, we suggest the following model: W, R � N W, S � R C© 2012 Wiley Publishing Ltd Expert Systems, July 2013, Vol. 30, No. 3 223 where N indicates the user needs. This new description stresses that the user requirements concerning a (usually) computer-based solution are functional to the satisfaction of the user needs. Highlighting the user needs is also useful to stress that in general there are several formulations of the requirements that satisfy the user needs, in the same way as there are several different machines specifications that sat- isfy the requirements. In fact, separating the notions of user needs and requirements makes it clear that writing require- ments is a design activity, since one has to choose how to define proper requirements that satisfy the user needs. Now, we can reason about the origins of user needs. It is easy to observe that user needs are generally conceived in some sort of business process; more specifically, the user needs are functional to achieving some goal in the context of a business process. For instance, in our case study the re- quirement of having the email messages stored in the email repository was determined by the need of sharing informa- tion, which in its turn was determined by the goal of letting the agencies act in a coordinated way. In conclusion, we may say that W, N � BG where W is the knowledge of the business process and BG represents some business goal. It is interesting to note that also the relationship between business goals and needs is one-to-many; in fact, the goal of making agencies act in a coordinated manner could be achieved in different ways: for instance, a coordination board could collect action proposals from agencies and control that no conflicts exist. Similarly, information sharing could be translated into dif- ferent requirements, according to the available communica- tion infrastructure. In practice, we are reproducing the hierarchical nature of requirements: the business process has some goals, which de- termine some needs, which can be achieved if some require- ments are satisfied. On their turn, requirements are satisfied if a suitable interaction between a machine and the envi- ronment is achieved. This hierarchy corresponds to different abstraction levels in the knowledge of the environment: 1. At the topmost level, the formulation of the goal depends only on the knowledge of the business process, and needs are functional to the business. 2. At the intermediate level, the mechanisms supporting the process are known: an email system is used; therefore, we need to save emails in repositories at the agencies. 3. At the lowest level, we finally take into consideration the local organization of the process (i.e. the fragment of the environment that is involved) and the responsibilities of the machine. Therefore, knowledge of the availability and cost of the personnel, as well as the cost of the machine development is required. Quite interestingly, while the knowledge of the world that is needed increases in extent and detail while we proceed from goals to requirements, it is also true that at the lower levels several details of the business knowledge can be safely ig- nored: for instance, when dealing with email message storing we could ignore that this is due for work coordination. The hierarchical organization described above could be beneficial to properly define monitoring and measurement activities. For instance, at the topmost level one is interested in knowing how long it takes to a piece of information to reach an agency; at the intermediate level one is interested in knowing the speed and throughput of the email system; at the lowest level several different measures can be defined depending on the chosen requirements: for example if the sit- uation described in Section 3.3 (scenario 2) holds, interesting measures could be the percentage of messages considered in- correct by the machine and the percentage corrected locally by the operator. 4.2. Requirements as transformations of needs and goals The proposed approach can be regarded as an application of software problem transformations (Hall et al., 2008; Hall & Rapanotti, 2011). In these transformations, problem P2 is related to problem P1 in that, in it, there is a detailed descrip- tion of the solution for P1. This relationship is also justified by an argument of the formal correctness of the solution, so that we can say that a solution of P2 is actually a solution of P1. A software problem transformation transforms a sin- gle problem – the conclusion – to a set of problems – the premises – and records an argument – the adequacy argu- ment step – that justifies the relationship of the premises to the conclusion. So, for instance, we could write equation (3) to formalize that problem W ′, R � NB (the information sharing problem) is solved if problems Wa1, Sa1 � Ra1 (information communi- cation at agency 1) and Wa2, Sa2 � Ra2 (information updat- ing at agency 2) are solved. The adequacy argument J (not detailed here) exploits the properties of the environment, namely of the agencies’ environments and the email connec- tion that guarantees that messages sent from agency 1 are received at agency 2. Wa1, Sa1 � Ra1Wa2, Sa2 � Ra2 W ′, R � NB � J (W ′ ∪ Wa1 ∪ Wa2 ) � (3) Another example of the use of transformations in our approach is given by the application of domain interpre- tation, ‘the problem transformation schemata by which a non-solution domain description is interpreted, that is re- expressed in some way to make it more suitable for prob- lem solving’ (Hall et al., 2008; Hall & Rapanotti, 2011): in our case study, the introduction of the description of the email system connecting agencies into the business do- main description can be seen as a domain interpretation transformation. Another construct proposed in Hall et al. (2008) and Hall and Rapanotti (2011) that can be effectively used in our approach is context expansion (the combina- tion of a number of problem domains in a problem’s context). In general, it seems possible to say that the construction of a hierarchy of requirements can be built and manipulated according to the principles of Problem Oriented Software Engineering illustrated in Hall and Rapanotti (2011) and Hall et al. (2008). 224 Expert Systems, July 2013, Vol. 30, No. 3 C© 2012 Wiley Publishing Ltd 4.3. Generalization of the approach Although in the case study we had requirements on three lev- els (the business goals, the user needs, and the requirements) we can easily generalize the proposed approach to deal with requirements hierarchies encompassing any number of lev- els. In such case, the description is given by a set of formulae like the following: W0, R1 � R0 . . . Wi, Ri+1 � Ri . . . Wn, S � Rn (4) In practice, we start with a high-level requirement R0 in an environment described by W0 and – through a sequence of refinement steps (or transformations, if you like better the terminology of POSE (Hall et al., 2008; Hall & Rapanotti, 2011) – we reach the usual formulation of a problem (Wn, S � Rn) as in the reference model (Gunter et al., 2000). The number of refinement steps needed to reach an imple- mentable machine specification depends on several factors, including how abstract was the specification of R0 and how complex is the environment. A constraint that has to be satisfied for the validity of (4) is that Wi+1 must be compatible with Wi , that is all the statements that were true in Wi are still true in Wi+1. However, it is possible that Wi+1 provides a more extended and/or detailed description of the environment than Wi, thus there can be statements that are true in Wi+1 and were simply not present in Wi. Note that the descriptions in (4) are expected to be con- sistent with the properties expressed in the reference model (Gunter et al., 2000). So, for instance, relative consistency should be assured, that is ‘any choice of values for the envi- ronment Wi variables visible to the system should be consis- tent with Ri+1 if it is consistent with assumptions about the environment’ (Gunter et al., 2000). The requirements hierarchy described by (4) can also be useful to classify requirements analysis activities. For in- stance, system-level analysis and component-level analysis are easily distinguished as it is possible to note that Wi con- tains a description of the system with little or no details about the system’s components, while Wi+1 contains more details about the components. As an example, in the description of an intensive care unit (such as the one widely discussed in Jackson (2001) at a given level the specification of the problem might not mention sensors at all, while the refined description could introduce them. The relationship between Wi and Wi+1 can also deal with the unspecified parts that may appear in problem descrip- tions (such parts are represented as grey clouds in Figures 6–8). In general, the lack of details is justified by the fact that the description is at a high level in the hierarchy. Sometimes, however, it may be that we have some degrees of freedom in shaping (and then describing) the environment: hence we can introduce in Wi+1 elements not present in Wi. Consider again the case of the intensive care unit: if the original de- scription of the environment does not specify the presence of any alarm device, we can introduce a device of our choice in the next level description. These considerations provide an answer to the last question reported in chapter 2: when it is clear that a description of the environment does not sup- port the requirements (e.g. it is not feasible that the patients themselves introduce their data into the intensive care unit system) then you can explore the possibility of extending or modifying the environment, for example by introducing sensors that connect patients with the machine. As far as terminology is concerned, the hierarchy of refine- ments described by (4) suggests some possible definitions: 1. Business goals are the highest Ri: for instance ∀ i, 0 ≤ i ≤ k → Ri is a BG. However, the hierarchy could start with a R0 that is not a BG, but some lower level require- ment; therefore, we need more stringent conditions for defining a BG: a possibility is to say that Ri is a BG if the corresponding environment description Wi does not include a computer machine as a necessary element. 2. Similarly, ∀ i, k < i ≤ m → Ri is a need. Needs are characterized by the fact that they do not include any reference to the final goal that led to their definition. In other words, given a need, it is usually not evident why that need arose. The description of the environment can involve a computer system, but the latter is usually described in an abstract way, that leaves several degrees of freedom to the analyst for outlining a solution. 3. Finally, ∀ i, m < i ≤ n → Ri is a requirement. This is the case most often described in literature: requirements are about the responsibility of the machine, whose detailed description is explicitly given (of course, only in terms of shared phenomena, the definition of the machine inter- nals being left to designers). According to this terminology the higher steps dealing with BG can be classified as ‘business analysis’, while lower steps are ‘requirements analysis’. Finally, it is necessary to stress that (4) actually describes a hierarchy, rather than a sequence. In fact, at each step there are generally multiple definitions of Ri+1 that satisfy Ri. 4.4. Validation and limitations This paper is meant to clarify the nature of concepts (such as user requirement, user needs, and business goals) and activities (business analysis, requirements analysis) of the ‘higher phases’ of the software development process. An- other objective of the paper is to put the concepts and ac- tivities mentioned above in the context of well-established approaches (the reference model for requirements and spec- ifications (Gunter et al., 2000), problem frames (Jackson, 2001), POSE (Hall et al., 2008; Hall & Rapanotti, 2011)). These clarifications can be used in several ways. At the two extremes of the range of options are the following pos- sibilities: 1. Analysts can use the suggested approach implicitly, as a discipline to keep order in their ideas while they proceed through the different phases of the analysis. 2. Analysts shape the analysis activities according to a pro- cess that explicitly takes into account the proposed con- cepts. So, for instance, artefacts classified as business goals, user needs and requirements are produced, and refinement/transformation steps are also planned and performed. C© 2012 Wiley Publishing Ltd Expert Systems, July 2013, Vol. 30, No. 3 225 In both cases, Section 5 provides guidelines about the or- ganization of the analysis process. Of the two possible ways of using the proposed approach, only the former was actually used by the author. Up to now, the argumentation of the validity of the proposed approach rests on the fact that it is coherent with the techniques in- volved (the reference model for requirements and specifica- tions (Gunter et al., 2000), problem frames (Jackson, 2001), POSE (Hall et al., 2008; Hall & Rapanotti, 2011)), and the previous research on goal-based RE. Concerning the possible limitations of the approach, it is expected that it is applicable as long as the underlying tech- niques – mainly problem frames – are applicable. A possible hindrance to the full-fledged application could be that the need to produce several classes of documents (business goals, user needs, and requirements) may be excessively expensive for small developments. On the contrary, in bigger develop- ment efforts, the possibility of exploring and assessing the cost of alternatives (see Section 5.3) could be a decisive ad- vantage. 5. An analysis process driven by user needs The observations reported above can also affect the analysis process: being able to tell the difference between user needs and user requirements can help structuring the development process more conveniently and effectively. In fact, when the user needs are known, we face two possibilities: 1. Devising a machine and a domain behaviour such that needs are satisfied, or 2. devising user requirements that satisfy the needs, and then specifying a machine that satisfies the requirements (and, hence, the needs). Option (1) corresponds to defining the machine specifi- cations S and W such that W, S � N. Although possible, this operation may be difficult, especially if the needs are expressed at a high abstraction level. In fact, in such cases the ‘distance’ between needs and machine is so large that the figuring out the features of the machine that satisfy the requirements involves building a quite complex argument. Moreover, the degrees of freedom in choosing the machine features are many, and call for criteria to make choices that are often not at the ‘machine level’. Option (2) is expected to be simpler, because it allows splitting the analysis phase into two steps: 1. Defining R such that R, W � N. 2. Defining the machine specification so that W, S � R. In this way, one does not have to ‘jump’ from needs di- rectly to the machine specifications: the distance between needs and machine is covered in two shorter steps: from needs to requirements and from requirements to machine specifications. However, it must be noted that during the first step one should take into account the cost of implementing the requirements, that is one should estimate the effectiveness and cost of the machine(s) that can satisfy the requirements. We can thus conclude that the analysis process needs to be organized into several activities. In a first step the various possible R that – together with W – satisfy N should be ex- plored; in this phase we consider the possibility of modifying the environment as needed to define more easily achievable requirements. The possibility of modifying the environment is justified by the fact that the higher is the level of needs, the more suitable are hybrid solutions, composed of both hard- ware/software machines and people. Of course, the costs involved in modifying the environment (e.g. for hiring or training people) have to be taken into account when evalu- ating the most convenient definition of R. The second step consists in choosing one of the require- ments definitions produced in step 1 and defining the specifi- cations of a machine that satisfies such requirements. Also in this case the cost of implementing the machine and possibly modifying the environment to correctly operate the machine should be evaluated, in order to assess the viability of the solution. If the costs are too high, it may be the case of going back to step 1 and look for further alternative requirements definitions. 5.1. A process model The considerations reported in Sections 3 and 4 may provide something of interest to analysts. However, it may not be obvious how effectively to use in practice the concepts em- bedded in the proposed hierarchical requirements model. In order to enable analysts to take advantage of such model, we propose a specific problem analysis process. The main merit of the proposed process model is that it highlights that anal- ysis has to proceed through a sequence of refinements, and that feasibility and cost evaluation activities are necessary, since some solutions could be hardly feasible or excessively expensive. Figure 16 shows a UML activity diagram representing the process of dealing with a business-level problem. The first ac- tivity (Define problem) involves describing the environment and the business goal, thus resulting in providing the defini- tions of W and BG. The rest of the process is devoted – as expected – to defining a ‘machine’ N such that W, N � BG. This is clearly the most difficult part of the process, which requires both creativity and a systematic and rigorous ap- proach. Activity Specify N involves defining a specification N that not only lets achieving the business goal BG, but al- lows for feasible and reasonably expensive solutions. Activity Specify N can fail altogether, thus leading to a failure of the whole analysis. If a specification is produced, the following activity (Feasibility and cost evaluation) involves estimating the cost of the specification N. Three types of results are envisaged: 1. The specification is feasible and the estimated cost for implementing it is acceptable; 2. the specification is unfeasible or too expensive; 3. the specification is too complex to get reliable evaluations about its feasibility or the implementation cost. In the first case, we can either consider the analysis com- plete, or we can explore different definitions of N, searching for cheaper specifications. In the second case, we go back to activity Specify N, in order to look for different – more feasible – definitions of N. In the last case, we need to tackle the complexity of the specifications by refining the business goal. This is done by 226 Expert Systems, July 2013, Vol. 30, No. 3 C© 2012 Wiley Publishing Ltd Figure 16: The process of defining how to achieve a business goal. means of the activity Refine business problem, which – if successful – produces a specification R that is feasible and reasonably expensive. Such R makes R, W � N true, so that also W, N � BG becomes true. Therefore, the cost of achiev- ing BG is the cost of implementing R. If the activity Refine business problem fails, either we are able to devise an alterna- tive definition of N, or we must give up, and declare BG not achievable, at least at reasonable costs. By applying the process in Figure 16 to our case study, we would get as a problem definition the specification described in Figure 6. Activity Specify N would bring to the definition of NB as described in Figure 7 (i.e. NB specifies the need for information sharing as a solution of the coordination problem). The cost evaluation of NB would reveal the need for refining the problem; in fact – as discussed in Section 3 – information sharing can be achieved in different ways and at different costs, depending on the available communication infrastructure and how it is used. Activity Specify N includes the construction of a ‘correct- ness argument’. This means that the resulting definition of N is not just stated, but it is shown (convincingly, if not formally) to be true (Jackson, 2001). Activity Refine business problem – whose goal is to de- fine R such that R, W � N and R is feasible and reasonably expensive – is described in Figure 17. It is easy to see that the activity is organized very like the process of dealing with business-level problems. A noticeable difference is that before proceeding to devise any definition of R, the given descrip- tion of the environment (W) is properly extended, in order to include details that are not relevant at the business level, but become essential when dealing with the processes that support the business. It could also be possible to consider the extensions of the environment description as strictly con- nected with the specification of R, thus activity Refine prob- lem definition could be made part of the loop that involves defining alternative specifications of R and evaluating them. In our case study, Refine problem definition would bring to the extended representation of the environment as in Fig- ure 9. Finally, Specify R produces the specifications of R for the two sub-problems (information creation and trans- mission at agency 1 and information receipt and storage at agency 2). More precisely, the first execution of Specify R would produce the specifications reported in Figures 11 and 13. If this set of requirements is considered feasible and rea- sonably expensive the process proceeds to the identification of a machine that satisfies the requirements. Otherwise, a new iteration of Specify R is performed, which produces the spec- ifications reported in Figure 14. This new set of requirements is evaluated with respect to feasibility and cost, and so on, until either a satisfactory set of requirements is found, or the process fails, having explored all the possible requirements that lead to satisfying NB. Of course, we may recognize the need for refining the spec- ifications of R (as is often the case, as discussed in Section 3) in order to test with several specification S of the actual machine. In such cases, activity Refine needs is executed. Its definition (not reported here) could be very similar to that of activity Refine business problem. As a final note, even though three levels of requirements were often mentioned (i.e. goals, needs, and requirements), C© 2012 Wiley Publishing Ltd Expert Systems, July 2013, Vol. 30, No. 3 227 Figure 17: The process of refining a business problem. the requirements hierarchy can be further expanded towards higher and lower levels, for example considering that several machines can implement the same specification, etc. 5.2. Dealing with failures The process described in the section above acknowledges the possibility that refinements fail. When a failure occurs, it is often possible to ‘backtrack’ and look for an alternative refinement. However, when no refinement appears possible you have to modify the problem, for example by negotiating the needs to be satisfied. In this section, the possibilities of failures and how to deal with them are discussed. The most critical case is probably when different business goals are somehow contradictory. In the formula W, N � BG, we have a set of new goals BG and an environment W that already satisfies the ‘older’ business goals. In fact, W is generally suited to support the current business: typically, W describes an organization (and its environment) as it is at the moment, and BG represents the improvements in the business the organization wants to achieve. Now, it is possible that BG contain some goals that may lead to contradictory requirements: in this case by considering the business goals together, either we find a set of requirements that satisfies all the goals, or we come to the conclusion that some goal has to be changed or dropped. It is also possible that some goal in BG is in contrast with W, that is it cannot be achieved as long as W is maintained as it is. This is the case when a new goal is in contrast with some decision that has been previously taken to satisfy older goals. In this case we are confronted with the choice between a conservative behaviour – that is keep W and drop part of BG – or a more innovative behaviour: critically analyse W to understand if the older goals are still valid and if they can be achieved differently; in such case drop the features or W that contrast with the achievement of BG, and rebuild it so that it can support both the new and the older goals. Schematically, the procedure to deal with this issue is as follows: The initial situation is Wol d , Nol d � BGol d . The new problem is Wol d , Nol d , NtoBeDefined � BGnew. If NtoBeDefined cannot be defined: If BGol d is no longer important, then restate the problem as Wol d , NtoBeDefined � BGnew Else, consider restating the problem as Wol d , NtoBeDefined � BGnew, BGol d (this involves dropping the implementation of Nol d , which may be quite expensive). 5.3. Estimating the cost of implementing specifications In the discussions reported above it has been stressed that being able to evaluate specifications is very important. The cost of achieving a business goal is made of two components: the cost of building (or buying) a machine that implements the specifications, and the cost of running it. In order to 228 Expert Systems, July 2013, Vol. 30, No. 3 C© 2012 Wiley Publishing Ltd Figure 18: Activity diagram specifying the behaviour of the data input operator in scenario 1. evaluate such costs it is necessary to remember that the ma- chine is usually made of both a computer-based application and people who execute procedures that are part of the spec- ifications (as in scenario 1 in Section 3.2, where an operator reads emails and fills database forms with the received infor- mation). Therefore, in order to fully evaluate the cost of a specifi- cation, it is necessary to consider and estimate the following factors: 1. The cost of implementing the hardware/software ma- chine; 2. The cost of setting up a human-based “machine” (i.e. defining “manual” procedures, hiring people that are able to carry out such procedures, training people, etc.); 3. The cost of execution of the software applications; 4. The cost of running the manual procedures, which in- cludes the cost of the involved people and of the devices and resources that people have to use in order to carry out the specified procedures. Costs (1) and (2) can be estimated by means of widely available techniques. Concerning cost (1) we should be able to apply the techniques that have been proposed for the cost estimation of software development. These techniques usu- ally require that the size of the software to be developed is measured (or estimated). In particular, modern techniques and tools (like for instance COCOMO II (Boehm et al., 2009) and SEER-SEM (Galorath & Evans, 2006)) accept that the size is given in function points (FP) (Albrecht, 1979) or – sometimes – in COSMIC FP (CFP) (COSMIC, 2007). Both these functional size measures are applied to user require- ments and define the size measure as a function of a set of ‘base functional components’: it is thus necessary to identify base functional components by examining the requirements specifications. Such issue has been addressed in Lavazza and del Bianco (2008) and del Bianco and Lavazza (2009), where mappings between problem frame modelling elements (like shared phenomena, for instance) and both FP and CFP are defined: the methods proposed in Lavazza and del Bianco (2008) and del Bianco and Lavazza (2009) allow for measur- ing the functional size of machine specifications, and then to use consolidated techniques and tools to get estimates of the cost of building the specified machines. As far as costs (4) are concerned, it is possible – as in the case of costs (1) – to exploit the information contained in problem frames. As an example, let us consider the specifica- tion of the data input operator, which is an essential part of the update processor in scenario 1. These specifications can be represented by means of a UML activity diagram, as in Figure 18. By examining the diagram, it is easy to express the cost of executing the specifications as follows: CDI O = K NRM (TRem + PumTScr + (1 − Pum )TF i f ) (5) Equation (5) expresses CDI O the operating cost of the data input operator with respect to: NRM: the number of received email messages; TRem: the average time taken to read and interpret a message; TScr: the average time taken to write and send a clarification request message; TFif: the average time taken to fill the input form with the information from the email message (the time taken to store the data is assumed to be practically nil); Pum: the probability that a received email message cannot be reliably interpreted; K: a constant that transforms the time employed into a cost. 6. Related work The RE community has produced a large amount of work on the role of user goals in RE. C© 2012 Wiley Publishing Ltd Expert Systems, July 2013, Vol. 30, No. 3 229 Ross and Schoman stated that requirements must specify ‘why a system is needed, based on current or foreseen condi- tions, which may be internal operations or an external mar- ket’, ‘what system features will serve and satisfy this context’, and ‘how the system is to be constructed’ (Ross & Schoman, 1977). Goals were identified as the originators of require- ments (Lee, 1991). Conversely, requirements emerge because of some underlying goal (Ross, Schoman, 1977; Dardenne et al., 1991; Sommerville & Sawyer, 1997). Mylopulos suggested the transition from more traditional forms of analysis to goal-oriented analysis, which involves the description and evaluation of alternatives (Mylopoulos et al., 1999). He also pointed out that goals and the way they are tackled are related to ‘the organisztional objectives behind a software development project’ (which correspond quite closely to the ‘business goals’ mentioned above). Goal refinement leads from high-level goals to low-level technical requirements, according to business-related con- siderations (Yu, 1993). Goal refinement involves choosing among different alter- natives; actually it is fundamental for detecting the existence of alternatives (Lamsweerde, 2000). The identification of requirements from goals has been ex- tensively studied (Dardenne et al., 1991; Rubin & Goldberg, 1992; Anton & Potts, 1998; Dubois et al., 1998; Kaindl, 2000; Lamsweerde, 2000). A fundamental step in goal-based RE was the definition of goals as properties to be achieved, described via optative statements as opposed to indicative ones (Jackson, 1995; Zave & Jackson, 1997). A quite comprehensive review of the research carried out in goal-oriented RE can be found in Lamsweerde (2001). Several authors have published requirements acquisition strategies that start from an analysis of goals (e.g. Dar- denne et al., 1993; Yu & Mylopoulos, 1994; Moffett & McDermid, 1995). In 1995, Potts proposes a schema sim- ilar to story schemata to help analysts develop a limited set of representative scenarios. In 1996, Darimont and van Lamsweerde presented an approach to goal refinement that provides formal constructive support while hiding the under- lying mathematics. Kujala et al. propose an approach to representing user needs and translating them into user requirements in indus- trial product development cases (Kujala et al., 2001). They used users’ task sequence diagrams and use cases to model the user requirements. Although Kujala et al., do not provide a precise set of definitions or a methodology, they highlight the importance of documenting requirements in a simple and representative way. None of the papers mentioned above used problem frames for the representation of requirements. Cañete-Valdeón et al. (2008) propose a characterization of the concept of ‘value’ of use cases based on the catalogue of frames for real software problems proposed by Jackson (2001). For each frame they discuss which results from the system could be regarded as valuable for the user. The proposed approach is possibly helpful in integrating the usage of problem frames with use case based RE, but does not help in the identification and management of user needs, or even higher level goals. Some work aimed at transforming requirements into spec- ifications in the context of problem frames appears in the literature (Rapanotti et al., 2006; Seater & Jackson, 2006; Li, 2007; Seater et al., 2007). With respect to the work reported here, all the mentioned papers focus on the lower section of the hierarchy described by (4); that is they tend to disregard the connection of requirements with higher level business goals. On the contrary, they focus on the process of deriving specifications from requirements. The idea of progressing from the problem to the solution via transformations is at the base of POSE: see, for instance, Hall et al. (2008) and Hall and Rapanotti (2011). Here a sim- ilar approach is suggested, but mainly focused on the very high-level descriptions of problems, when the users’ goals and needs are described entirely in the problem world, and the machine specification can be seen as the target of the pro- cess. In other words, while POSE aims – using very similar principles – at finding solutions to given problems, our ap- proach is more concerned on defining problems themselves, given the high-level user goals. In 2009, Hall and Rapanotti propose assurance-driven de- sign in the context of POSE. They also propose a process model that addresses the issues of assurance and trade-offs. Such process stresses the importance of problem and solu- tion validations. An important aspect of development that is made explicit in Hall and Rapanotti (2009) is that of back- tracking: design steps that come to a dead end because they cannot be validated cause to backtrack to a point where a dif- ferent approach can be taken. This also applies in the process proposed here: for instance, in Figure 17 when a refinement fails we go back to specifying requirements in a different manner (but always in a way that satisfies W ′, R � N). In Section 1, it was stated that it would be useful to clarify what are the differences between a user requirement and a user need, and what are the relations with business processes. The growingly strong interconnection between IT and busi- ness processes has made clear the need to link higher level (i.e. business related) and lower level (i.e. at the software develop- ment level) goals. This need appears especially often when it comes to evaluating IT-supported business strategies (Becker & Bostelman, 1999; Buglione & Abran, 2000; Bianchi, 2001; Card, 2003). The relation of higher with lower level goals has been recently described – in connection with the problem of evaluating the effectiveness of business strategies and IT so- lutions – by Basili et al. (2010). In IT solutions are viewed as (part of) the strategies that in a given context and un- der given assumptions lead to achieving given business goals (Basili et al., 2010). However – also because they focus on the evaluation problem – Basili et al. do not propose a clear and systematic way of modelling the requirements hierarchy. The proposal contained herein of systematic refinement of busi- ness goals in a given environment into needs to be achieved could help clarify the concepts of ‘strategy’ and ‘context’ proposed in Basili et al. (2010). 7. Conclusion This paper proposes the definitions of a few concepts that are needed to properly identify and describe user needs and requirements. User needs lie entirely in the user’s problem domain space. No machine responsibility needs to be specified. This appears quite evident when a problem diagram is used to model the 230 Expert Systems, July 2013, Vol. 30, No. 3 C© 2012 Wiley Publishing Ltd user needs: in fact, the part of the diagram that includes the machine can be left unspecified, as in Figures 5–7. There are several requirements definitions that can satisfy the users’ needs. The analysis process has to take this issue into account: RE is a design activity, since it must lead to choosing one of the several possible requirements definitions that satisfy the user needs, and one of the several possible ma- chine specifications that satisfy the requirements. Moreover, problem domains are deeply involved in the analysis: it is often possible, as in our case study, that part of the solution depends on the behaviour of the domains (often humans) that interact with the machine. User needs are themselves determined by the goals of a business process. This observation suggests that there are not just two levels (the needs and requirements levels); rather, we may have a hierarchy of requirements, where each level of the hierarchy corresponds to an abstraction level in the descrip- tion of the problems and solutions. Dealing with this hier- archy calls for an articulated analysis process, in which the problem frames and related methodology play an important role. In Section 5, a process for dealing with requirements hierarchies has been outlined, using UML activity diagrams. Since it has been shown that the methodology of problem frames can be successfully used in combination with UML (Lavazza & Bianco, 2004, 2006; Bianco & Lavazza, 2008), it will be useful to define UML-based representations of the needs and requirements, and a UML-compliant analysis pro- cess. Future work includes the definition of a methodology for building UML-based descriptions of goals, needs, and re- quirements and using problem frames methodology to carry out in the most seamless and integrated way refinements, correctness argument constructions, cost analysis, and any other useful activity. As briefly discussed at the end of Section 4, the transi- tion from goals to needs and from needs to requirements can be considered similar to the transformations at the base of POSE (Hall et al., 2008; Hall & Rapanotti, 2011): formal- izing the proposed concepts in the context of POSE is thus another goal of future activities. Actually, the proposed hi- erarchy of requirements can be seen as the base for defining methodological guidelines for using POSE transformations in dealing with high-level user goals and needs. Finally, we can note that even though in industry there is still some confusion about the nature or requirements and their relationships with business goals, the situation is im- proving: for instance, some companies are starting to follow consistent guidelines, such as the Business Analysis Body of Knowledge (International Institute of Business Analysis), a collection of guidelines that reflects current generally ac- cepted practices of business analysis. Acknowledgements The research presented in this paper has been partially funded by the project ‘Metodi e tecniche per l’analisi, l’implementazione e la valutazione di sistemi software’ funded by Università degli Studi dell’Insubria. References ALBRECHT, A. (1979) Measuring application development productiv- ity. IBM Application Development Symposium, IBM Press, 83–92. ANTON, A.I. and C. POTTS (1998) The use of goals to surface re- quirements for evolving systems, 20th International Conference on Software Engineering, 19–25 April 1998, Kyoto. BASILI, V.R., M. LINDVALL, M. REGARDIE, C. SEAMAN, J. HEIDRICH, J. MUNCH, D. ROMBACH and A. TRENDOWICZ (2010) Linking soft- ware development and business strategy through measurement, IEEE Computer, 43, 57–65. BECKER, S.A. and M.L. BOSTELMAN (1999) Aligning strategic and project measurement systems, IEEE Software, 16, 46– 51. BIANCHI, A.J. (2001) Management indicators model to evaluate per- formance of IT organisations, International Conference on Man- agement of Engineering and Technology, Vol. 2, IEEE Press, 217– 229. BOEHM, B.W., C. ABTS, A.W. BROWN, S. CHULANI, B.K. CLARK, E. HOROWITZ, R. MADACHY, D.J. REIFER and B. STEECE (2009) Soft- ware Cost Estimation with COCOMO II. Upper Saddle River, NJ, USA: Prentice Hall Press. BUGLIONE, L. and A. ABRAN (2000) Balanced scorecards and GQM: what are the differences? 3rd European Software Measurement Con- ference, Federation of European Software Measurement Associa- tion, 18–20 October 2000, Madrid, Spain. CAÑETE-VALDEÓN, J.M., F. ENRÍQUEZ, J. ORTEGA and E. VELÁQUEZ (2008) Clarifying the semantics of value in use cases through Jack- son’s Problem Frames, Information Processing Letters, 107, 221– 229. CARD, D. (2003) Integrating practical software measurement and the balanced scorecard, 27th Annual International Computer Software and Applications Conference, IEEE CS Press. COSMIC – Common Software Measurement International Con- sortium (2007) The COSMIC Functional Size Measurement Method – version 3.0 Measurement Manual (The COSMIC Implementation Guide for ISO/IEC 19761: 2003). Available at http://www.cosmicon.com/dl_manager4.asp?id=73 (accessed August 22, 2012). DARDENNE, A., A. VAN LAMSWEERDE and S. FICKAS (1993) Goal- oriented requirements acquisition, Science of Computer Program- ming, 20, 3–50. DARDENNE, A., S. FICKAS and A. VAN LAMSWEERDE (1991) Goal- directed concept acquisition in requirements elicitation, 6th Inter- national Workshop on Software Specification and Design, Como, 14–21. DARIMONT, R. and A. VAN LAMSWEERDE (1996) Formal refinement patterns for goal-driven requirements elaboration, 4th ACM Sym- posium on the Foundations of Software Engineering, October 1996, San Francisco, 179–190. DEL BIANCO, V. and L. LAVAZZA (2008) Enhancing problem frames with scenarios and histories in UML-based software development, Expert Systems – The Journal of Knowledge Engineering, Vol. 25, no. 1, Blackwell Publishing. DEL BIANCO, V. and L. LAVAZZA (2009) Applying the COSMIC func- tional size measurement to problem frames, 14th IEEE Interna- tional Conference on Engineering of Complex Computer Systems – ICECCS ’09, 2–4 June 2009, Potsdam (Germany): IEEE Com- puter Society. DUBOIS, E., E. YU and M. PETIT (1998) From early to late formal re- quirements: a process-control case study, 9th International Work- shop on Software Specification and Design, 16–18 April 1998, Mie, Japan: IEEE CS Press, 34–42. GALORATH, D.D. and M.W. EVANS (2006) Software Sizing, Estimation, and Risk Management, Auerbach Publications. GUNTER, C., E. GUNTER, M. JACKSON and P. ZAVE (2000) A reference model for requirements and specifications. IEEE Software, 3, 37– 43. HALL, J.G. and L. RAPANOTTI (2005) Problem frames for socio- technical systems, In Requirements Engineering for Socio-Technical Systems, A. Silva and J.L. Mate (eds), chapter 19, Idea Publishing Group, 318–339. HALL, J.G. and L. RAPANOTTI (2009) Assurance-driven design in prob- lem oriented engineering, International Journal on Advances in Sys- tems and Measurements, 2, 119–130. HALL, J.G. and L. RAPANOTTI (2011) Software engineering as the design theoretic transformation of software problems. Innovations in Systems and Software Engineering, London: Springer. C© 2012 Wiley Publishing Ltd Expert Systems, July 2013, Vol. 30, No. 3 231 HALL, J.G., L. RAPANOTTI and M. JACKSON (2008) Problem Oriented Software Engineering: solving the package router control problem, IEEE Transactions on Software Engineering, 34, 226–241. INTERNATIONAL INSTITUTE OF BUSINESS ANALYSIS, The Guide to the Business Analysis Book of Knowledge, version 2. http://www. thei- iba.org. JACKSON, M. (1995) Software Requirements & Specifications – A Lex- icon of Practice, Principles and Prejudices. ACM Press, Addison- Wesley. JACKSON, M. (2001) Problem Frames – Analysing and Structuring Soft- ware Development Problems, Addison-Wesley, ACM Press. KAINDL, H. (2000) A design process based on a model combining scenarios with goals and functions, IEEE Transactions on Systems, Man and Cybernetic, 30, 537–551. KUJALA, S., M. KAUPPINEN and S. REKOLA (2001) Bridging the gap between user needs and user requirements, in Advances in Human- Computer Interaction I, N. Avouris and N. Fakotakis (eds), Rio Patras, Greece: Typorama publications, 45–50. LAVAZZA, L. (2010) User needs vs. user requirements: a problem frame-based view, International Workshop on Advances and Appli- cations of Problem Orientation (IWAAPO’10), 8 May 2010, Cape Town, South Africa. LAVAZZA, L. and V. DEL BIANCO (2004) A UML-based approach for representing problem frames, 1st International Workshop on Advances and Applications of Problem Frames (IWAAPF), 24 May 2004, Edinburgh: IEE. LAVAZZA, L. and V. DEL BIANCO (2006) Combining problem frames and UML in the description of software requirements, Fundamen- tal Approaches to Software Engineering (FASE06), 25 March–2 April 2006, Vienna: Springer. LAVAZZA, L., V. DEL BIANCO (2008) Functional size measurement based on problem frames: a case study, 3rd International Workshop on Advances and Applications of Problem Frames (IWAAPF), 10 May 2008, Leipzig: ACM press. LEE, J. (1991) Extending the Potts and Bruns Model for recording de- sign rationale, 13th International Conference on Software Engineer- ing, 13–17 May 1991, Austin, Texas, USA: IEEE-ACM, 114–125. LI, Z. (2007) Progressing problems from requirements to specifica- tions in problem frames, PhD Thesis, Department of Computing, The Open University, Walton Hall, Milton Keynes, UK. MOFFETT, J.D. and J.A. MCDERMID (1995) Goals as a focus for con- flict management in requirements engineering. RE ’95: Second In- ternational Symposium on Requirement Engineering, 28–30 March 1995, York, UK: IEEE Computer Society. MYLOPOULOS, J., L. CHUNG and E. YU (1999) From object-oriented to goal-oriented requirements analysis, Communications of the ACM, 42, 31–37. POTTS, C. (1995) Using schematic scenarios to understand user needs, 1st Conference on Designing Interactive Systems: Processes, Prac- tices, Methods, & Techniques, 23–25 August 1995, Ann Arbor, Michigan: ACM. RAPANOTTI, L., J.G. HALL and Z. LI (2006) Deriving specifications from requirements through problem reduction, IEE Proceedings: Software, 153, 183–198. ROSS, D.T. and K.E. SCHOMAN (1977) Structured analysis for require- ments definition, IEEE Transactions on Software Engineering, 3, 6–15. RUBIN, K.S. and A. GOLDBERG (1992) Object behavior analysis, Com- munications of the ACM, 35, 48–62. SEATER, R. and D. JACKSON (2006) Problem frame transformations: deriving specifications from requirements, 2nd International Work- shop on Advances and Applications of Problem Frames, 23 May 2006, Shanghai, China. SEATER, R., D. JACKSON and R. GHEYI (2007) Requirement progres- sion in problem frames: deriving specifications from requirements, Requirements Engineering Journal (REJ’07), 12, 77–102. SOMMERVILLE, I. and P. SAWYER (1997) Requirements Engineering: A Good Practice Guide. Chichester: Wiley. VAN LAMSWEERDE, A. (2000) Requirements engineering in the year 00: a research perspective, keynote, 22nd International Conference on Software Engineering, 4–11 June 2000, Limerick, Ireland: ACM Press, 5–19. VAN LAMSWEERDE, A. (2001) Goal-oriented requirements engineer- ing: a guided tour, 5th IEEE International Symposium on Require- ments Engineering, 27–31 August 2001, Toronto, Canada: IEEE Computer Society, 249–263. YU, E.S.K. (1993) Modelling organisations for information systems requirements engineering, 1st International Symposium on Require- ments Engineering, January 1993, San Diego, CA, USA: IEEE, 34–41. YU, E.S.K. and J. MYLOPOULOS (1994) From E-R to ‘A-R’: modelling strategic actor relationships for business process reengineering, 13th International Conference on the Entity-Relationship Approach, Manchester, UK. ZAVE, P. and M. JACKSON (1997) Four dark corners of require- ments engineering, ACM Transactions on Software Engineering and Methodology, 6, 1–30. The author Luigi Lavazza Luigi Lavazza is associate professor at the University of In- subria at Varese. He also cooperates with CEFRIEL (ICT Center of Excellence For Research, Innovation, Education & Industrial Labs partnership). His research interests in- clude: software process modelling, measurement, and im- provement; the measurement of software products, especially concerning quality and the functional size; model-based de- velopment, especially concerning real-time and embedded software; requirements engineering and software develop- ment environments and tools. He participated in several research projects co-funded by the European Union or by the Italian government, often guiding the research unit of CEFRIEL or of the researchers of the University of Insub- ria at Varese. He is co-author of over 100 scientific articles, published in international journals, or in the proceedings of international conferences or in books. He serves on the program committees of a several international conferences. Luigi Lavazza is an IARIA fellow and member of the IEEE computer society. 232 Expert Systems, July 2013, Vol. 30, No. 3 C© 2012 Wiley Publishing Ltd 
work_qskmcgextfc4bf72c3pcmxbz64	----	untitled Multi-agent system for knowledge-based event recognition and composition Angel Rivas-Casado,1 Rafael Martinez-Tomás1 and Antonio Fernández-Caballero2 (1) Dpto. Inteligencia Artificial. Escuela Técnica Superior de Ingenierı́a Informática, Universidad Nacional de Educación a Distancia, Juan del Rosal 16, 28040 Madrid, Spain Email: arivas@dia.uned.es (2) Dpto. Sistemas Informáticos, Escuela Técnica Superior de Ingenierı́a Informática, Universidad de Castilla-La Mancha, Albacete, Spain Abstract: This work presents a multi-agent system for knowledge-based high-level event composition, which interprets activities, behaviour and situations semantically in a scenario with multi-sensory monitoring. A perception agent (plurisensory agent and visual agent)-based structure is presented. The agents process the sensor information and identify (agent decision system) significant changes in the monitored signals, which they send as simple events to the composition agent that searches for and identifies pre-defined patterns as higher- level semantic composed events. The structure has a methodology and a set of tools that facilitate its development and application to different fields without having to start from scratch. This creates an environment to develop knowledge-based systems generally for event composition. The application task of our work is surveillance, and event composition=inference examples are shown which characterize an alarming situation in the scene and resolve identification and tracking problems of people in the scenario being monitored. Keywords: knowledge-based systems, high-level vision, image understanding, video surveillance, event-base system, intelligent agent, plurisensorial agent, Arduino 1. Introduction This work presents a multi-agent system for knowledge-based high-level event composition, which interprets activities, behaviour and situa- tions semantically in a scenario with multi- sensory monitoring. The system’s specific aims would be (Bobick, 1997) to monitor the complex scenario, using all the sensors’ capacities (Car- mona, 2009) to interpret the information from these sources as a whole, and (Chleq & Thonnat, 1996) to request further information to confirm or reject the hypotheses raised in the process for diagnosing the situations. The semantic interpretation of scenes must recognize situations, activities and interactions between the different actors. The physical sig- nals captured by the sensors must be linked with the interpretation of their meaning. When a human observer interprets the meaning of a scene, he uses his knowledge of the world, the behaviour of the things that he knows, the laws of physics and the set of intentions governing the agents’ activity. All this additional knowl- edge enables the observer to model the scene and use this model to predict, at least partially, what may happen. In other words, to launch a DOI: 10.1111/j.1468-0394.2010.00578.x Article _____________________________ c� 2011 Blackwell Publishing Ltd Expert Systems 1 mailto:arivas@dia.uned.es mailto:arivas@dia.uned.es mailto:arivas@dia.uned.es hypothesis about the evolution of the events and activities that he is detecting. The aim is to aid the human operator so that he can take objec- tive, consistent real-time decisions about the events detected. Multi-sensory interpretation combines multiple sources of information from different sensors in order to generate a more exact and robust inter- pretation of the environment (Pavón et al., 2007). The interpretation of behaviour and situations is currently based on using several sensors that offer a high number of false alarms. The availability of new types of sensors for monitoring tasks pro- vides new challenges for multi-sensory data fu- sion that they solve or they diminish this problem. It is now possible to construct models of the environment and diagnose situations by analys- ing indefinite sequences of sensory information from different types of sensors. The efficient fusion of multi-sensory information – understood as the fusion of data from several sensors of the same type or, additionally, as the fusion of data from several different types of sensors – is a crucially important task in advanced surveillance. Thus, in recent years there has been a rapid increase in research and development into the combined use of multiple sensors, including vi- sion, audio, heat (thermal), presence (volumetric), vibration sensors, etc. (Zhu & Huang, 2007). A multi-sensory, monitoring system usually re- quires three components of equal importance (Zhu & Huang, 2007): (1) sensors that capture the information for the monitoring system, pro- cess and interpret it; (2) fusion algorithms of data from each of the sensors; (3) architectures for constructing real-time systems. Of all the sensors, the visual ones are ob- viously very important. High-level vision is defined as the interpretation of scenes beyond the mere recognition of objects (Fuentes & Velastin, 2004). Image understanding (image comprehension=interpretation) generally starts by analysing movement and then describes the scene with symbols (Tostsos et al., 1980; Levine et al., 1983). It is widely accepted that it is necessary to inject the expert’s knowledge to complement the low-level information (Neumann & Novak, 1983; Neumann, 1984). In a constructivist model, the event composition is done with spatio-temporal relations, with parallel processes that determine the time inter- vals of other events. Regarding the diagnosis of situations, in Chleq and Thonnat (1996) the hypothesis approach is included, which implies doing parallel explorations for alternative solu- tions. The confirmation of a hypothesis in a description level helps the previous levels. In Rota and Thonnat (2000), the use of declarative models for representation is described. Our work follows this line, constructing declaratory models from the skill of a human expert who knows how to identify critical situations, parti- cularly in surveillance monitoring problems. Surveillance is a perfect application field for both plurisensory monitoring and interpreting video-sequence scenes, where our group has accumulated considerable experience with AVI- SA project works (Folgado et al., 2007; Martı́- nez-Tomás et al., 2008; Carmona, 2009). It is also a multi-disciplinary task affecting an increasing number of scenarios, services and clients. In particular, the use of agents in surveillance systems has some precedents in the bibliography (Abreu et al., 2000; Remagnino et al., 2004; Haesevoets et al., 2007) and the multi-sensory surveillance works by Castanedo et al. (2008). In the cooperative sensor agent (CSA) architecture proposed in Molina et al. (2004) coalitions are formed between agents to perform surveillance tasks. The CSA is broken down into two levels: sensor layer and coalition layer. A coalition is formed when an agent (sensor) needs to cooperate with other agents that have capacities that it does not have. The article is structured as follows: the system structure, its organization into vision agents (VAs) and plurisensory agents (PAs), and the high-level composition agent (CA) are in Section 2. The perception agents (VAs and PAs) identify significant changes in the sensors’ magnitudes and transmit them as simple events to the CA. From these simple events, the CA composes composed events that imply higher-level semantic activities. The system is described and an application exam- ple is given that identifies the abandoning of objects as an alarming situation. Section 3 2 Expert Systems c� 2011 Blackwell Publishing Ltd describes the CA as a knowledge-based system and the structure of its knowledge base. The methodology for creating, updating and modify- ing the base for different application fields is also described. The article finishes by describing the system’s application for monitoring people in different visual fields with different cameras, therefore just visual information. It also describes the problem of access control where plurisensory information is already being used for precise identification and tracking. 2. Intelligent agents of the system The term intelligent agent in artificial intelli- gence refers to any entity that can take decisions from its environment (Russell & Norvig, 2009). As shown in Figure 1, the system schema pro- posed consists of one or several PAs, one or several image interpretation agents (IIAs) and in theory, just one CA. Both the PAs and the IIAs process the sensor information and identify (agent decision system) significant changes in the monitored signals, which the connection interface translates into simple events and sends to the CA via a network connection system. For an efficient system, the agents must be carefully selected and placed by an expert in the best positions of the environ- ment to be monitored. 2.1. Composition agent It is the high-level knowledge-based synthetic software agent that identifies activities or situa- tions (composed events) as a composition from simple events that meet certain spatio-temporal restrictions. We can distinguish between two types of functionalities: the reception of simple events generated by the PA and IIA and the composition from these simple events. Defining the event patterns (composed events) that char- acterize the target activities is an analytical task of knowledge acquisition by a human expert who knows a behaviour pattern’s simple actions perfectly. The actions correspond to significant high-level activities and have associated replies. For this, a methodology was created to develop the knowledge base. The generation of an event from the PA and VA implies the instantiation of the corresponding generic event, where values are given to its attributes. For example, At(h, x, y, t) is instantiated at At((h ‘Peter’)(x 150)(y 120)(t 430)) when Peter is identified as at posi- tion (150, 120) at instant 430. For the composi- tion, the events are included as part of the facts base. The scenario ontology also includes facts describing its inherent characteristics, where the activities occur: doors, windows, passages, no entry areas, adjacent cameras, etc. Figure 2 schematically illustrates an example of an alarming situation. A person leaves an object on the ground and goes away. The first row includes simplified images in the scene instants. The second row contains the simple events that are generated from segmenting, monitoring and identifying each frame of the sequence. The third row shows the pattern for the composition (composition unit) of events occurring at each instant. It is a knowledge unit for a composition: a set of events that must meet specific spatio- temporal relations and the consequent events with higher-level semantics. In the event base there is a previous ‘Walking’ event that com- pletes the pattern. The forth row shows the higher-level event inferenced by this way. ‘Walk- ing’ belongs to a special type of event that can be called ‘event-state’ that occur over a time period t1–t2, unlike ‘instant-event’. At, for example, occurs at a given instant t. Thus, following the example of Figure 2 we can interpret the sequence as follows, step by step: 1. At instant t, human1 is detected on the scene. The identification and monitoringFigure 1: Agent interconnection schema. c� 2011 Blackwell Publishing Ltd Expert Systems 3 agent does not recognize that the human is carrying an object, so the spatio-temporal location of the human is only represented with the event At in that instant t. With this At the event ‘Walking2 is updated to the following instant. 2. At instant tþ1, ‘object1’ is detected near to the position of human1. This situation acti- vates a pre-alarm of possible abandonment of an object with the event ‘Pre-alarm’. Since there are no other humans nearby, it is inferred that ‘human1’ has left the object. An association is created between the object and human, and the pre-alarm is activated. 3. At instant tþ2, ‘human1’ is detected leav- ing the scene. This event and the active pre- alarm identify a situation of abandoning an object. The event ‘Alarm’ goes off. We group composition units into packages that identify a specific situation. In turn, the pack- ages are organized into composition levels, each package is assigned to a composition level. Each composition level sends the composed events that it has generated to their higher composition level. Packages in different composition levels are inter-dependent. Thus, if a package in a specific level is added to the knowledge base, all those packages in the lower composition levels, which are necessary for its functionality, will be added. Our aim is to give the possibility to create libraries of packages rich enough to configure a system easily. Each library of packages has its own corresponding event ontology. 2.2. Plurisensory agent The incorporated systems provide information about the environment where they are located. They can take a high number of readings per second to pinpoint significant alterations. Gen- erally, they perform very simple operations. They can take, analyse and send samples from several sensors. Figure 3 shows a multi-sensory system consisting of a shield ethernet connected to an Arduino plate with an Atmel AVR 8bits Figure 2: Recognition that a person has abandoned an object. Figure 3: Multi-purpose multi-sensory system. 4 Expert Systems c� 2011 Blackwell Publishing Ltd processor. This system can capture, analyse and send information from the sensors once the information has been filtered by the local area network (LAN). At present, as shown in Figure 4, work is being done on integrating into the multi- purpose multi-sensory system different sensors that are particularly relevant for surveillance problems: movement sensors, proximity sen- sors, lighting sensors, temperature sensors, rela- tive humidity sensors, open door and window detector, environmental noise sensors, radio frequency identification (RFID)-based identifi- cation systems. This multi-purpose multi- sensory system provides a PA with several advantages: its low cost and easy reproduction, it provides reliable, rapid information and final- ly, it can be used to monitor private rooms, where for legal reasons, video-surveillance sys- tems cannot be installed. We are currently studying the possibility of incorporating IIAs into the system taking advantage of the infor- mation received from the PA improve the event segmentation and focusing processes (VA), like, for example, detecting scene lighting changes or doors opening in the visual field. 2.3. Vision agent This agent combines a system of perceiving images with vision software that requires high computational performance for near real-time tasks. Its main functions are those of a vision system: segmentation, identification and track- ing of people and=or objects within a monitored visual field (Carmona, 2009). For the identifica- tion process, a block-based system is used (Fol- gado et al., 2007) developed by our group, the same as the level structure to identify events (Martı́nez-Tomás et al., 2008). The connection interface can identify a battery of pre-designed events, which are not specific to the application field, but are especially relevant for identifying (composing) the significant high-level events related to the scene events, as shown in Martinez Tomás and Rivas Casado (2009). Section 5 shows an example. Figure 4: Multi-sensory system schema. c� 2011 Blackwell Publishing Ltd Expert Systems 5 3. Knowledge base structure (KBS) Figure 5 shows the internal structure of the knowledge base. First of all, we differentiate between levels that correspond to abstraction levels, from knowledge to identify more precise activities, more directly associated with mechan- ical actions or movements, to more abstract activities, which define behaviour or situations. The packages are sets of rules (as mechanisms for minimal representation of composition units) that identify specific situations. The con- cept of package is really important in this schema since the system can add or eliminate these from a knowledge base. Thus there is package inter-dependency. If a package is elimi- nated from an upper level, all the lower-level packages providing it with information will be automatically eliminated. Obviously, when a package is eliminated, all its composition units are also eliminated. Each package has a work environment defined by the knowledge level to which it belongs, by the input events and the events that it generates. Each package contains one or more rules. The rules only have access to the events defining the package. Thus each rule is encapsulated within the best environment. Errors and possible crossing of information between different packages are also prevented. The advantage of this structure compared with a traditional KBS is that it is easy to maintain and reconfigure when changes occur in the environment. In each level, the rules are prior- itized for their execution. The most important rules are the ones that process the information from the previous level and the least important are those that generate the information that is transferred to the upper level. This process is performed for each knowledge level as a pipe- line. The composed events that level N generates at instant T are processed in level Nþ1 at instant Tþ1. 3.1. Composition unit In the prototypes that we have developed, the composition of events is represented on rules, but generically we can speak of composition unit. It is regarded as the unit representing the system’s knowledge. It consists of three parts: � Antecedents: It is the event pattern that must be met to be able to check the composition conditions. � Conditions: It is a filter that must be passed to be able to execute the high-level event composition. � Actions: Once all the tests have been passed, the corresponding changes are made in the event base. New events can be created or one or other of the existing events can be eliminated. To create new events, the event data are taken that are the composition unit antecedents. Similarly, only events can be eliminated that belong to the antecedents. This mechanism is able to fuse simple events into more complex events. In the example below, we can see two events, a simple one and a composed one. The aim is to update the composed event information with the simple Figure 5: Breakdown of the knowledge base. Each knowledge level (abstraction level) consists of a number of packages that in turn have rules. Each rule implements, as composition unit, one or several high-level events pattern. 6 Expert Systems c� 2011 Blackwell Publishing Ltd event information: Walkingðh;x1;y1; t1;x2;y2; t2Þ Atðh;x;y; tÞ The event At shows where a human h is at instant t in the position x, y. The event ‘Walk- ing’ represents at what time and position he began to walk to what time and position he stopped walking. Now the example is presented with the instantiated events: In the first event it is observed that the human ‘‘Peter’’ is at position x¼200, y¼50 and t¼500. The conditions would be that the hu- mans were the same person and that instant t2 of the event ‘Walking’ were less than instant t of the event At. With these data, the second event could be updated to Walkingððh PeterÞðx1 100Þðy1 50Þ ðt1 450Þðx2 200Þðy2 50Þðt2 500ÞÞ We now have the event updated to the instant 500, so we can affirm that ‘Peter’ has advanced 100 pixels to the right in 50 instants. After updating all the ‘Peter’ dependent events, the simple event could be erased before executing the following instant. The knowledge base is structured as an ab- straction pyramid (Figure 6). At the bottom is the knowledge level supported by the simple events, which arrive from the PAs and IIAs, and at the top of the pyramid are the upper knowledge levels. To do the event composition in level n (for n>1), a hypothesis is made, which searches the events verifying this hypothesis in level n�1. If the hypothesis’s restrictions are met, the new composed event is generated in level n. Configuration parameters are global variables that can be accessed by all the rules. They store a specific value, usually, numerical. They are used to be able to configure the knowl- edge base in different environments. A video camera’s resolution, for example, the ontology terms like ‘Near’ and ‘Far’ act as configuration parameters. 4. Methodology and development tools One of the AVISADOS project’s aims was to use it in different application fields to facilitate the configuration for new scenarios, always with the structure described in Sections 2 and 3. For this, a number of tools and similar methodology were developed, which means that each new application does not start from scratch. The Figure 6: Inference pyramid. Walkingððh PeterÞðx1 100Þðy1 50Þðt1 450Þðx2 190Þðy2 50Þðt2 499ÞÞ Atððh PeterÞðx 200Þðy 50Þðt 500ÞÞ c� 2011 Blackwell Publishing Ltd Expert Systems 7 result is a system that generates KBSs in general and interpretation KBSs based on the agent structure in particular. 4.1. Methodological and functional schema In a design stage, in Figure 7, the event ontology is configured. This will be used later to create agent interfaces and define the knowledge packages and units. Three stages can be identi- fied in the execution stage. In the first, the agents process the environment information and iden- tify the events corresponding to the actions detected. Then the agent interface sends the set of events to the execution and monitoring sys- tem (EMS) via a LAN. In the second stage, the events received are labelled with the correspond- ing instant in the EMS. Then the execution takes place. The inferred events are sent to the data- base to which the EMS is connected. The last stage is the analysis of the results by the user interface. All the agents are connected to their corresponding interface. This in turn is con- nected to the EMS. In the following sections, we focus in more detail on the system generating the agent interfaces and the knowledge base. 4.2. Knowledge base IDE The flowchart of Figure 8 shows the procedure for constructing a knowledge base and the agent interfaces. The development of a new knowledge base begins by defining the event attributes. Then we implement the simple and composed event ontology that the CAs, PAs and VAs will be able to identify (define events) and that we use to develop the agent profiles (define agent) and to create the knowledge packages (define package). An agent profile contains the characteristic data of this perception agent ‘identifier, position in the scenario, function, description’ and the collection of simple events that it can identify. This informa- tion is necessary to create the connection interface (generate agent interfaces), from the perception agent, which sends the simple events perceived to the CA. The package library is a repository of packages of identifiable situations previously im- plemented. It is not necessary to finish of defining (define package and construct rules) all packages to generate a knowledge base. Through a process of selection, you can configure the knowledge to address different situations (select identify situa- tion in package library and generate knowledge base). So, this process allows for reconfiguring a new system very quickly. Figure 9 shows the user interface of knowl- edge base IDE. The list of simple and compound events developed is shown. The left-side bar shows the information about the project and the selected item. This provides expert assistance in designing the knowledge base. The search for elements developed is one of the utilities included in the tool. On the right of the window we can see Figure 7: System schema. 8 Expert Systems c� 2011 Blackwell Publishing Ltd all elements that form the knowledge base. They are classified into six sections: agents interface, configuration parameters, parameters of the events, events, composition units and packages. Each package has an associated knowledge level and some simple and=or composed events with which to work. The expert, following the activity identification patterns, generates the right composition units so that the package functions correctly. For this the rule manager is used. First of all, the package must be defined for the rule. Each composition unit has an associated package from which it inherits the set of events to which it may refer in the antecedents, conditions and actions. 4.3. Execution system This tool executes the knowledge base created with the knowledge base IDE, stores all the composed events generated in a record and connects all the system’s agents. It consists of: � Network server: It activates the socket so that the agents connect to the system. The connection is bidirectional to be able to request certain information from the agent. � List of connected agents: A list is kept of all the agents that are connected to the system along with their description and related data. � Event synchronization system: It generates time slices to receive and label events. Once the event has been received, it is labelled with the time associated with the time slice. Thus an asyn- chronous system becomes a synchronous one. � Composition motor: Inference motor that evaluates the knowledge unit requirements and adds the inferred ones to the event base. � Statistic and monitoring system: It analyses the number of events that arrive from each Figure 8: Flowchart for developing a knowledge (centre), reconfiguration system and generate knowledge base (right). c� 2011 Blackwell Publishing Ltd Expert Systems 9 agent, the composition motor execution time, the synchronization system buffers and the database connection response times. This information is really useful when cali- brating and configuring all the system’s ex- ecution parameters. � Database connection to store the information: It sends all the events inferred at each instant to the selected database. The information can then be processed later. When these data are analysed, possible errors can be de- bugged from the knowledge base, which is really useful for refining the system. Figure 10 shows the EMS. This window shows the user the state overall system: server status, Figure 9: Knowledge base IDE. Figure 10: Execution and monitoring system. 10 Expert Systems c� 2011 Blackwell Publishing Ltd connection port, number of clients, state of the system timer, state of inference engine, state of the statistics and the state connection with the database. The red cylinder represents the execution time instant and the green cylin- der, average execution time. The bottom bar shows number of the events that are currently on the basis of facts. The left bar buttons are used to change the status of different system functions monitoring and enforcement. The buttons on the top bar show individual from each of the elements system. The upper-left button, displays the system preferences: configuring the connection to the database, maximum number of customers, cycle time im- plementation, connection port, maximum size of the base facts and the selector of the knowl- edge base. 5. Resolving problems in identifying people and tracking Section 2 showed identifying an abandoned object as an application to recognize an alarm- ing situation. In this section, a solution is shown for identifying and tracking people, first with adjacent cameras and then with the support of precise identification from the access control. Both of them are application examples of our event compositions to solving problems for the lower levels events. 5.1. Adjacent camera tracking In the framework of tracking people and objects with artificial vision, the problem is to identify each one in two overlapping or adjacent images. Figure 11 shows a scenario where a person appears in the view angle of two separate cam- eras. Each of the cameras is calibrated according to the global coordinates. The system receives the positioning events of each person visible in each of the camera’s images. If we add the knowledge of the ‘Adjacent(c1, c2)’ scene to this, we can generate a rule that resolves this pro- blem. The example is shown below whereby the problem would be resolved: Atðc1;h345;x1;y1Þ Atðc2;h213;x2;y2Þ Adjacentðc1;c2Þ IFðx2 near x1Þandðy2 near x1Þ Figure 11: Scenario with overlapping cameras. c� 2011 Blackwell Publishing Ltd Expert Systems 11 Then (c1, h345)¼ (c2, h213) and (c2, h213)¼ (c1, h345). Thus it is immediately inferred that the human visible in camera 1 is the same one visible in camera 2. Figure 11 shows the evolution in time of events and by inference rule described above. In Intanto, t h345 shows a human in a position x1, y1 in camera 2. At time t2 is a h345 human in a position x2, y2 in the camera 2 and in turn a human h213 at position x3, y3 in the house 1. In the knowledge of the scene must be the camera 1 and camera 2 are adjacent. With this back- ground runs the rule displayed in row 3, and it appears that the h345 human camera 2 is the same as the h213 human camera 1 by the event ‘Similar’. 5.2. Precise identification from the access control Figure 12 shows an access control RFID identi- fication. Such controls provide, in normal cir- cumstances, a robust system to identify people. One of the problems posed by the artificial vision is the reliable identification of humans within a scene. It is therefore envisaged the merger of the information from the RFID identification system with the identification and tracking information provided by artificial vi- sion. Figure 13 shows an example of agent fusion using the scene knowledge level. It is an RFID card reader access control. A camera with a tracking system monitors people entering. The CA receives the person’s identification informa- tion from the RFID card read by the PA. Moreover, the VA detects another person enter- ing the security area and labels him=her with a self-generated ID. When these two items of information reach the CA, it searches for some correlation between the two. The system changes the self-generated ID for the RFID card ID. Then the person can be tracked via the video-camera inter-connection system. 6. Conclusions This work presents a multi-agent system for composition based on high-level knowledge events, which interprets activities, behaviour and situations semantically in a scenario with multi-sensory monitoring. A perception agent (PA and visual agent)-based structure is pre- sented. The agents process the sensor informa- tion and identify (agent decision system) significant changes in the monitored signals, which they send as simple events to the CA that searches for and identifies pre-defined patterns like higher-level semantic composed events. This Figure 12: Individual access control. Figure 13: Video-surveillance access control. 12 Expert Systems c� 2011 Blackwell Publishing Ltd structure also has a description of the methodol- ogy to develop knowledge-based systems to compose events and a set of tools to facilitate its application. The methodology focuses on dividing knowledge, classifying it into types and encapsulating it into what we call knowl- edge packages, which in turn are organized into abstraction levels. Each knowledge package identifies specific situations or activities from lower-level abstraction events. These ideas were shown using the prototype developed for video surveillance. It consists of two tools: a develop- ment interface for a rule-based knowledge base, and an EMS. The first tool aims to configure the pattern of new alarming situations, either based on standard, usual activities (in the knowledge package repository) or directly from identifying and tracking events in the image sequences. These sequences can be labelled with the tool included or the image analysis module. The EMS offers an execution system to facilitate the tasks to analyse the results. Different applica- tion examples of the event composition from different perception agents have been shown, which identify alarming situations and resolve problems in identifying and tracking people. The system is particularly useful for surveillance but it could also be applied to other fields. Future works will aim to increase the system’s inferential capacities. Obviously, it is necessary to bear in mind and process the sources of indetermination in each abstraction level, and particularly the identification and tracking ones. Acknowledgements The authors are grateful to the CiCYT for financial aid on project TIN-2007-67586-C02-01. References ABREU, B., L. BOTELHO, A. CAVALLARO, D. DOUX- CHAMPS, T. EBRAHIMI, P. FIGUEIREDO, B. MACQ, B. MORY, L. NUNES, J. ORRI, M. TRIGUEIROS and A. VIOLANTE. (2000) A video-based multiagent traf- fic surveillance system, in The IEEE Intelligent Ve- hicles Symposium 45-462. BOBICK, A. (1997) Movement, activity, and action: the role of knowledge in the perception of motion, Royal Society Workshop on Knowledge-based Vision in Man and Machine, London, pp. 1257–1265. CARMONA, E.J. (2009) On the effect of feedback in multilevel representation spaces, Neurocomputing, 72, 916–927. CASTANEDO, F., J. GARCÍA, M.A. PATRICIO and J.M. MOLINA (2008) A multiagent architecture to support active fusion in a visual sensor network, in 2nd ACM= IEE International Conference on Distributed Smart Cameras, Stanford University, CA, USA, pp. 1–8. CHLEQ, N. and M. THONNAT (1996) Realtime image sequence interpretation for video-surveillance appli- cations, IEEE International Conference On Image Processing (ICIP’96), Laussane, pp. 800–804. FOLGADO, E., M. RINCÓN, E.J. CARMONA and M. BACHILLER (2007) A block-based model for moni- toring of human activity, Technical Report AVISA- 12-07. FUENTES, L.M. and S.A. VELASTIN (2004) Vigilancia avanzada: del tracking a la detección de sucesos, IEEE América Latina, 2, 206–211. HAESEVOETS, R., B. VAN EYLEN, D. WEYNS, A. HELLEBOOGH and T. HOLVOET (2007) Context- driven dynamic organizations applied to coordi- nated monitoring of traffic jams, Engineering Environment-Mediated Multiagent Systems, Dres- den, Germany, 1–5 October, 2007, D. Weyns, S. Brueckner andY. Demazeau (eds.), pp. 126–143. LEVINE, M.D., P.B. NOBEL and Y.M. YOUSSEF (1983) A rule-based system for characterizing blood cell motion, in Image Sequence Processing and Dynamic Scene Analysis, T.S. Huang (ed.), Berlin, Germany: Springer-Verlag, 663–709. MARTÍNEZ-TOMÁS, R., M. RINCÓN-ZAMORANO, M. BACHILLER-MAYORAL and J. MIRA-MIRA (2008) On the correspondence between objects and events for the diagnosis of situations in visual surveillance tasks, Pattern Recognition Letters, 29, 1117–1135. MARTINEZ TOMÁS, R. and A. RIVAS CASADO (2009) Knowledge and event-based system for video-sur- veillance tasks, in Proceedings of the 3rd Interna- tional Work-Conference on the Interplay Between Natural and Artificial Computation (IWINAC): Part I: Bioinspired Applications in Artificial and Natural Computation, ISBN:978-3-642-02263-0, J. Mira, J. M. Ferrández, J.-R. A. Sanchez (eds.), Berlin: Springer-Verlag, 386–394. MOLINA, J.M., J. GARCÍA, F.J. JIMÉNEZ and J.R. CASAR (2004) Fuzzy reasoning in a multiagent system of surveillance sensors to manage coopera- tively the sensor-to-task assignment problem, Ap- plied Artificial Intelligence, 18, 673–711. NEUMANN, B. (1984) Natural language description of time-varying scenes. Brericht no. 105, FBI-HH-B- 105=84, Fachberic Informatik, University of Hamburg. c� 2011 Blackwell Publishing Ltd Expert Systems 13 NEUMANN, B. and H. NOVAK (1983) Events models for recognition and natural language description of events in real-world image sequences, Proceedings of the Eighth IJCAI, Karlsruhe, Morgan Kaufmann, San Mateo, California, pp. 724–726. PAVÓN, J., J.J. GÓMEZ-SANZ, A. FERNÁNDEZ-CABAL- LERO and J.J. VALENCIA-JIMÉNEZ (2007) Develop- ment of intelligent multi-sensor surveillance systems with agents, Robotics and Autonomous Systems, 55, 892–903. REMAGNINO, P., A.I. SHIHAB and G.A. JONES (2004) Distributed intelligence for multicamera visual sur- veillance, Pattern Recognition, 37, 675–689. ROTA, N.A. and M. THONNAT (2000) Video sequence interpretation for visual surveillance, in IEEE Inter- national Workshop on Visual Surveillance (VS’00), Dublin, Ireland, pp. 59–68. RUSSELL, S. and P. NORVIG (2009) Artificial Intelli- gence: A Modern Approach, 3rd edn, Pearson: Pre- ntice Hall. TOSTSOS, J.K., J. MYLOPOULOS, H.D. CORVEY and S.W. ZUCKER (1980) A framework for visual motion understanding, IEEE Transactions on Pattern Ana- lysis and Machine Intelligence, 2, 563–573. ZHU, Z. and T.S. HUANG (eds), (2007) Multimodal Surveillance: Sensors, Algorithms and Systems, Artech House Publishers. The authors Angel Rivas-Casado Angel Rivas-Casado received his degree in technical engineer in computer science of systems from the University Polytechnical of Madrid, Spain, in 2009. At present, he attends a master of advanced artificial intelligence in the National University for Distance Education (UNED) in Madrid, Spain. His research interests are in robotics, knowledge engineering, artificial vision, sensors fusion and intelligent agents. Rafael Martinez-Tomás Rafael Martinez-Tomás received his degree in physics from the University of Valencia, Spain, in 1983, and received his PhD from the department of artificial intelligence of the National University for Distance Education (UNED) in Madrid, Spain, in 2000. Since 2001, he is an associate professor with the department of artificial intelligence at the UNED. His research interests are in knowledge engineering, knowl-edge-based systems, spatial– temporal logics, description logics and video- sequence interpretation. Antonio Fernández-Caballero Antonio Fernández-Caballero received his de- gree in computer science from the Technical University of Madrid, Spain, in 1993, and re- ceived his PhD from the department of artificial intelligence of the National University for Dis- tance Education, Spain, in 2001. Since 1995, he is an associate professor with the department of computer science at the University of Castilla- La Mancha, Spain. His research interests are in image processing, computer vision, neural net- works and agent technology. A. Fernández- Caballero is member of the IAPR. 14 Expert Systems c� 2011 Blackwell Publishing Ltd 
work_qv6jntxxevamlahgaiuxsxufbu	----	PII: 0898-1221(90)90116-2 Computers Math .4pplit Vol. 20. No 9 10, pp. I I 1-123, 1990 009%4943 90 $3.00 + 0.00 Printed in Great Britain. All rights reserved Copyright ,£' 1990 Pergamon Press pie I N T E L L I G E N T I N T E R A C T I O N I N D I A G N O S T I C E X P E R T S Y S T E M S Y. LIROV and S. RAVIKUMAR AT&T Bell Laboratories, Holmdel. NJ 07733, U.S.A. Abstract--Advisory systems help to improve quality m manufacturing. Such systems, how, ever, both human and computerized, are less than perfect and frequently not welcome, Sharp separation between working and learning modes is the mare reason for the apparent hostility of advisory systems. Intelligent interaction deploys computerized advisory capabilities by merging working and learning modes. We have developed a knowledge-based interactive graphic interface to a circuit pack diagnostic expert system. The graphic interface integrates both the domain knowledge (i.e. circuit pack) and the troubleshooting knowledge (i.e. diagnostic trees). Our interface dynamically changes the amount of detail presented to the user as well as the input choices that the user is allowed to make. These changes are made usmg knowledge-based models of the user and of the circuit pack troubleshooting domain. The resulting system, McR, instead of guiding the user by querying for input, monitors users actions, analyzes them and offers help when needed. McR is able both to advise "how-to-do-it" by reifying shallow knowledge from the deep knowledge, and to explain intelligently "'how-does-it-work'" by abstracting deep knowledge from the shallow knowledge, McR is used in conjunction with the STAREX expert sy.stem which is installed at A T & T factory. I. I N T R O D U C T I O N In this paper we discuss the development o f advisory interfaces for electronic circuit pack troubleshooting. Advisory interface is a term used to describe collectively any training and reference material, on-line help and guidance, and any other user support means. A diagnostic expert system is an example o f a computer-based advisor. Computerized advisory systems are hard to build and difficult to use. Since a consultation is a break in work continuity, new users try to skip around in a training sequence or dismiss training altogether [I]. Additionally, as Ref. [2] noted: "'Studying advisory problems and developing advisory solutions for the leading edge o f h u m a n - c o m p u t e r interaction is hampered by the fact that the leading edge must have already been codified and deployed before advisory problems can even exist." A n o t h e r basic difficulty in developing advisory interfaces is the lack o f understanding o f coaching and interacting. For example, people often seek advice by making claims about the possible answer to their (unannounced) query. Does it mean that in such cases a direct question is less efficient? A n o t h e r important issue is the frequency o f an advice, or how often should the advice be offered? A good parent does not teach the children at every occasion when something is not up to the standards o f the parent. How detailed should be the advice? When is it more important to provide the declarative advice (i.e. how-does-it-work) and w h e n - - t h e procedural advice (i.e. how-to-do-it)? In addition, it is i m p o r t a n t to note that the advisor is constantly under suspicion o f giving the wrong advice. The advisor loses credibility after displaying even the smallest deficiencies. W h a t are the techniques to restore credibility? And what are the answers to all these questions, when the advisory interface is a computer program. I.I. Research and Development Topic In this subsection we refine the above outlined research topic to a task: we specify what exactly we want a c o m p u t e r to do. Having developed an expert system for a m a n u f a c t u r i n g facility (STAREX, see Lirov [3]), we experiment with a possible solution to the most important question: how to deploy a less than perfect computerized advisor? Basically we are trying to improve the classical expert systems which patronize the user by issuing queries, explaining only when asked, and giving the result only at the end o f the session. Such systems treat the user as an object for providing the details which are necessary to complete the reasoning done by the machine. These systems also maintain a sharp separation between the learning and working modes. III 112 Y. LIROV and S. RAVIKUMAR I. I.I. The task Carroll and A a r o n s o n [2] address the same question by simulating an intelligent advisory system for an interactive software design package (the "'Wizard o f Oz" technique). We build on their work in that we consider the same issue, but we restrict ourselves to a relatively narrow application domain. By considering a narrow application domain we expect to be able to go a step further from simulating an interface to actually writing its code and observing its behavior. We build an add-on software module, called McR, which merges working and learning modes. I. 1.2. The domai n We chose our application d o m a i n to be the electronic circuit pack diagnostics for two reasons: first, diagnostics is the widest expert systems application domain; and second, diagnostics o f electronic circuit packs is the better understood diagnostic problem because o f the availability o f deep knowledge [3,4]. We build on S T A R E X expert system in that we regard circuit pack diagnostic models to be available. Since S T A R E X is a deployed expert system, we expect our actual intelligent advisory system to be deployable at the factory floor level. We note here that to ease on our p r o g r a m m i n g efforts, ~.e have used a simplified S T A R E X version, which does not include truth maintenance. This issue is postponed for future research. On the other hand we reuse the softw.are modules performing the optimization o f the diagnostic sequences [5]. For the sake of better paper readability, we digress briefly to explain the basic concepts in electronic circuit pack diagnostics. For an overview o f the topic the reader is referred to Ref. [4]. I. 1.2. I. Diagnosis. Diagnosis o f an electronic circuit usually refers to the process of determining the fault~r component(s) that cause an undesired behavior (output) of the given circuit for some (correctly) given input. Diagnosis can be regarded as a problem o f economic optimization: there is a value associated with every c o m p o n e n t in the pack, as well as a fiscal value associated with the process o f assembly and soldering o f the pack. Thus, when a pack is declared faulty, it is desirable to replace only the faulty components. Moreover, the process o f identification o f the faulty components (the diagnostics) consists o f a series o f tests each o f which has an associated cost, expressed in such parameters as test setup time, c o m p o n e n t destruction, etc. The successful troubleshooter must be able to select an appropriate strateg), conduct measurements and replace components. Morris and Rouse [6] report that humans are not good in judging failure rates, h u m a n performance degrades as systems become larger and more complex and, or in the lace o f time constraints, presentation o f theory o f operation does not improve performance and proceduralization improves performance. Diagnosis is a difficult problem for the beginners. An unskilled troubleshooter has difficulties in making an indictment, doing it correctly and doing it sufficiently fast. As a result, a beginner causes buildup in work in process, unnecessary test-repair cycles, unresolved manufacturing problems (e.g. a fault) robot or a bad batch o f components), and even sometimes overlooked design problems. Therefore, the construction o f a u t o m a t e d means for troubleshooting guidance is justified both as an intellectual and as an economic challenge. However, the diagnostics problem li.e. the construction o f a c o m p u t e r algorithm which is capable to perform diagnostics) is NP-hard [7, 8]. I. 1.2.2. Troubleshooting strategies. Two basic approaches and their combinations are currently utilized in industry for conventional diagnostic software development: the fault dictionary approach and the guided probe approach. The fault dictionary approach requires simulation o f the fault) behavior of the system and subsequent storage o f the simulation results (as well as the fault assumptions~ in the fault dictionary. A " m i s b e h a v e d " o u t p u t o f the pack under test can therefore be looked up in the fault dictionary. The fault dictionaries are not being used x~idely since the) are usually incomplete and ambiguous. The guided probe approach requires only simulation o f the correctly functioning circuit pack. The basic tool in utilizing this approach is the blame-shifting mechanism applied after each measurement. The upstream (from the test point view) components in the circuit are blamed only when the measurements do not agree with the expected results. The efficient sequencing of measurements becomes the main problem when implementing the guided probe approach. The troubleshooter must employ some kind o f strategy in searching for the source o f the diffi- cult). Ref. [9] noted that poor troubleshooters mader fewer tests before accepting hypothesis as Intelligent interaction in diagnostic expert systems 113 correct, they had more incorrect hypothesis, and they pursued incorrect hypothesis longer than did the better troubleshooters. Glaser and Phillips [10] associated more than 20% o f strategic shortcomings (e.g. insufficient testing) with faulty inferences (e.g. misinterpretation o f a test). Additionally, poor troubleshooters tend to have incomplete lists o f hypothesis and to be frequently overconfident about their completeness [I I]. I. 1.3. The ciew McR behaves like a diary for the user, where the user enters the measurements and their results. The system collects this data and tries to reconstruct the reasoning of the user. If McR discovers a significant reasoning fault on the user's behalf, then it offers guidance. If the fault is a technical fault (e.g. using wrong measurement device), then the advice is a low-level, "how-to-do-it" advice, Otherwise, it is a "how-does-it-work" advice. Such advice can be about the test strategy (e.g. a diagnostic tree), or about the circuit pack (e.g. signal path). This combination of two approaches (the procedural and the declarative system explanations) is most likely to be the most effective means of troubleshooting advice [6], since the system merges the Socratic and the "'learning by doing" methods of instruction. I. 1.4. A d d i t i o n a l benefits An important expert systems characteristic is its flexibility to acquire knowledge. Expert systems usually mimic the behavior of experts which combine knowledge about several domains. Experienced circuit pack troubleshooters, for example, have some understanding about the circuit pack, know how to use measurement equipment (e.g. oscilloscope), have knowledge about the manufacturing process and its weak spots (e.g. "that robot loses its calibration frequently and thus inserts wrong components", or "'this transistor is often bad"), and know about generally good troubleshooting strategies (e.g. "divide and conquer"). Shallow knowledge bases contain rules which represent the combined expert knowledge. It is difficult to maintain such knowledge bases as it is difficult to comprehend all the interdependencies between the rules. Deep knowledge bases, on the other hand, promote easier maintenance by segregating different kinds of knowledge in separate knowledge bases. The price for this convenience is the need for an integrated inference engine which may take a form of a meta-interpreter [3, 4]. Construction of user interfaces for such knowledge bases is complicated by the necessity to maintain all the knowledge bases simultaneously. Our system is able to acquire different kinds o f knowledge (e.g. electronic signal path or troubleshooting tree) and immediately show the implications in the compiled Ishallow) form. We demonstrate this flexibility using an integrated graphics interface. From the methodological point o f view, this experiment helps to understand better the taxonomy of problems related to artificially intelligent diagnostics. In particular, we show that the contem- porary subdivision of the issue to seven subproblems [4] is rather superficial: the user interface problem is not a stand-alone problem, to be solved separately. Its solution includes solving all the typical problems for the knowledge-based applications, e.g. choosing knowledge representation, acquiring knowledge, etc. I . Z The Approach 1.2. I. User m o d e l Roughly, we develop intelligence o f the advisory system by creating a user model. The model can be used as a reference with which the actual user performance can be compared. The task of building intelligent advisory system through a user model seems to be tractable for three reasons: first, we can inventory the errors o f the analysts; second, we expect to develop shallow knowledge disassembly mechanisms to map from observations to user errors; and finally, we have developed in Ref. [3] a multi-source deep knowledge integration mechanism (reification), which we use to check the success o f the disassembly o f knowledge. Systematic shallow knowledge disassembly (extracting deep knowledge) is a new interesting problem having much in common with machine learning. 114 Y. LlROV and S. RAVIKUMAR 1.2.2. Reification and abstraction C o d i n g an intelligent advisory system is a large project requiring solutions to a variety o f problems: designing knowledge representation, c o n t r o l strategy, building knowledge base, integrat- ing multiple sources o f deep knowledge a b o u t electronic circuitry, developing user model, acquiring knowledge and displaying advice in a variety o f ways. O u r previous rep o rt [3] addressed the first four problems, while in this paper we deal with the last three. We show that the algorithms for knowledge reification and abstraction are the co rn erst o n es in the intelligent advisory systems. This view holds promise for further d e v e l o p m e n t because it provides better understanding o f b o t h reification and its i n v e r s e - - a b s t r a c t i o n . F u r t h e r m o r e , we introduce a new use for a b s t r a c t i o n - - w e view it as a tool to test user b e h a v i o r patterns. 1.2.3. Performance criteria The p e r f o r m a n c e o f the p r o p o s e d m e t h o d can be evaluated by testing it for correct and timely user e r r o r classification. While correct classification depends on the user model and diagnostic algorithm, the timeliness o f classification depends on how m u ch o f the relevant information has already been preprocessed. Being able to preprocess most o f the i n fo rm at i o n b efo reh an d holds promise to be able to deliver timely and correct advice. 1.2.4. Programming techniques T h e p r o g r a m m i n g a p p r o a c h is similar to that o f knowledge-based p r o g r a m m i n g since we use knowledge to supplement the observed user actions to generate o u r interpretations. T h e difference, o f course, is that the p r o d u c t o f the system is not a p r o g r a m code, but an advice. We use object-oriented logic p r o g r a m m i n g to develop graphic interface, and m e t a p r o g r a m m i n g - - f o r reification and abstraction. 1.2.5. The scope O u r m e t h o d might scale up to bigger systems, if deep and differentiable shallow kinds o f knowledge can be integrated, and a complete and finite inventory o f h u m a n errors are available. O f course, certain types o f intelligent help will always be missing because o f the inherent brittleness o f such systems. T h e system is honest a b o u t its limitations when it knows them: the system will tell that something is wrong but will not extend a t e m p o r a r y solution. We ad v o cat e such an a p p r o a c h to all advisory systems (machine and h u m a n ) in o rd er to maintain their credibility. O u r next question is: H o w well is the m e t h o d u n d e r s t o o d ? We confess that we d o not understand the m e t h o d very well, T h e reason is that the m e t h o d depends on having complete knowledge a b o u t the system under diagnosis and a b o u t the h u m a n errors. We are unable to p ro v e that we have completed the acquisition o f either kind o f knowledge, let alone the mapping between the h u m an e r r o r inventory and electronic circuitry deep knowledge. Thus, paradoxically, the only way to complete knowledge acquisition is to write d o w n a p r o g r a m with incomplete knowledge to deploy it, to p r o v o k e the difficulties, and to learn experimentally a b o u t the limitations o f the method. As a result, we expect not only to expand the scope o f the knowledge base but also to refine the knowledge a b o u t h u m a n analyst errors. 1.3. Design for Experimentation Once ~.e decided to use p r o g r a m m i n g as the main exploration tool, we must verify that o u r p r o g r a m is going to s u p p o r t e x p e r i m e n t a t i o n in addition to having the required p e r f o r m a n c e level. T o ensure "'experimentability", we write the code in Prolog, in o rd er to have all the inference mechanisms readily available for alteration. We will be able to observe the p r o g r a m ' s b eh av i o r both externally via graphics display and textual messages, and i n t e r n a l l y - - v i a Prolog execution trace. We d e m o n s t r a t e its correct knowledge t h r o u g h a set o f test cases c o r r e s p o n d i n g to the user faults. T h e p r o g r a m generally supports both " h o w - t o - d o - i t " and "'how-does-it-work" kinds o f advice, it is not tuned to one particular kind o f help. P e r f o r m a n c e o f the p r o g r a m has two aspects: its functional and its time characteristics. Functionally, we expect to observe increased n u m b e r o f rules and improved t r o u b l e s h o o t i n g procedures. O u r understanding a b o u t interaction, together with a short review o f some prevoius relevant work, is laid out in Section 2. An interface, implementing Intelligent interaction in diagnostic expert systems 115 o u r ideas o f interaction, is developed by efficiently combining knowledge and metaknowledge in Section 3. 2. I N T E R A C T I V E A P P R O A C H A c o m p u t e r program is considered interactive when it requires some user participation in order to conclude its processing. Diagnostic expert systems are obviously interactive systems both at the knowledge acquisition and at the diagnosis phases. The degree o f interactivity, the a m o u n t o f freedom that the user is allowed varies depending on the skill o f the user. If the user is a programmer, then an editor is sufficient. Otherwise, more sophisticated tools are required. Some guidelines for constructing h u m a n - c o m p u t e r interfaces have been published by D O D [12-15], but they do not provide the necessary information for determining the effectiveness o f specific interfaces and their sophistication. As a rule o f thumb, the program sophistication level is directly proportional to the user's familiarity with the c o m p u t e r programming techniques. S O P H I E (SOPHisticated Instructional Environment) is probabl~ the best example o f an intelligent interactive computer-based instructional system used for teaching electronics trouble- shooting [16]. The system can be used in two modes: a team troubleshooting game and an interaction with an expert. In the game mode the players o f one team inflict faults into a simulated electronics system and troubleshoot the simulated faults o f the other team. S O P H I E is an example o f an experimentally implemented learning-b)'-~h~ing environment with a t a x o n o m y o f user errors and an advice feedback mechanism to the user. Unfortunately, the system lacks graphic interface, does not handle significantly complex electronics equipment, and does not evaluate troubleshooting strategies. 2. I. Graphic Interaction Interactive graphics systems require both the inputs and outputs be specified graphically. Graphics editing o f the knowledge base has been proposed recently [17-19] as interaction means with the knowledge base. An expertise transfer system (ETS) [20] has been proposed to generate the rules from the user-supplied elements and their attributes and corresponding ratings. The system identifies conflicts and ambiguities in the rules (but not omissions), and asks the expert for modifications. Since ETS does not use a model o f the unit under test, numerous similar rules may need to be entered (e.g. pertaining to the same faults on different channels) and at the same time ETS may still not notice some o f the missing rules. Sand K A S T (Sandia Knowledge Acquisition System, Hill et al. [19]) uses a directed acyclic graph structure as a formalism for knowledge acquisition. In this graph, a node represents a state o f the troubleshooting session with an associated set o f possible faults. An arc represents a test to be applied in the context o f the source node. The graph structure provides the means for viewing the knowledge acquisition and truobleshooting processes as well as for graphic editing of the knowledge base. S a n d K A S T , although alleviates most o f the problems that arise with knowledge base maintenance, runs only in the K E E environment on a Symbolics computer and relies heavily on KEEs graphics utilities and object orientation. I M P U L S E [18] permits the user to interact with the knowledge base via various multiple windows and graphic displays o f the knowledge. Editing, however, is allowed only at the textual level and not graphically. The user o f sophisticated graphics interfaces in computer aided design, simulation, or instruction systems has been also suggested in S T E A M E R [21], Omega [22], Pecan [23], G a r d e n [24], Wlisp [25], PV [26], PAW [27], Thing Lab [28] and G U I D O N - W A T C H [17]. All o f the above systems emphasize and make use o f the abilities o f the brain to detect spatial patterns and reason upon them [29]. 2.2. Flexible htteraction Few o f the above mentioned systems, however, call attention to the questions o f evaluating advisory strategies and adapting them, and all leave those questions open. Consequently, all o f the above systems lack the flexibility o f adjusting the user interface to the sophistication level o f the user. As a result, the systems are either too complex at the initial stages or too detailed at 116 Y. L m o v and S. RAVlKUMAR the subsequent stages. In any case the users becomes a n n o y e d with the system and frequently stop using it. T h e concept o f an adaptive interface [30-32] is an extension o f an interactive user interface idea. An interface system becomes an adaptive system in two ways: active and passive. T h e passive way allows users to modify the interface, so it is tailored to meet the specific users needs. A l t h o u g h resulting in a m o r e suitable interface, the burd en o f its ad ap t i n g is left to the user. An active interface modifies itself. Architecture o f an active interface [33-35] addresses explicitly the issues o f dialogue [36, 37] and user modeling [38, 39]. M c R is an active user interface. C r o c k f o r d [40] identified the user involvement as the most i m p o r t a n t principle in the design o f interactive programs. He characterized user involvement as having " m o r e to d o with taking part than in making decision". T h e user choices must affect the presentation. T h e user is viewed as a part o f the program. Following this a p p r o a c h , o u r system maintains a model o f the user. According to the user-model, the sytem defines the kind o f d at a and rules to o p erat e on. Such an a p p r o a c h allows the user to enter incomplete specifications o f the problem an d let the system make knowledge-based interpretations o f user intentions. Thus, instead o f a one-directional interface at a single level o f complexity, the system interface is flexible. Consider, for example, the arcade games. Dep en d i n g on the n u m b e r o f accumulated points, the speed and the complexity o f the game increases. But, c o n t r a r y to arcade games, where the options given to the user to input into the system remain fixed, we require the user options to be d y n am i c and fit the user needs. T h e user model is const an t l y reevaluated by challenging the user at every step. Such a d y n a m i c a p p r o a c h not only keeps the p r o g r a m to be u p d at ed a b o u t the level o f the user's proficiency, but it also allows some d o u b t in the user's mind a b o u t the o u t c o m e o f the interaction. T h e r e f o r e , the user remains interested in the interaction. T h e final diagnosis then is the "'happ~ e n d " which reinforces the user's interest in using the p ro g ram . Recently Fischer et al. [41] reported a b o u t the need for co m b i n i n g the advice-giving strategies. The). distinguish between actice strategies, where the system provides advice by interrupting the dialogue, and passive strategies, where the user must explicitly ask for advice. Passive strategies usually employ a Socratic style o f interaction where the system poses questions and the user is expected to provide answers. Such systems often patronize the users, treating them as the i n f o r m a t i o n providing tools, needed only to conclude the reasoning by the co m p u t er. Active strategies often e m p l o y learning-by-doing environments, where user's actions are c o m p a r e d with the ideal actions and feedback is provided to i m p ro v e the user's responses towards expert p r o t o t y p e . M c R combines the strategies o f interaction. 2.3. Adaptive bzterface Architecture Knowledge-based adaptive interfaces include four kinds o f knowledge [36, 42,43]: (a) user model: (b) interaction and dialogue management; (c) knowledge o f the task and (d) system characteristics. In the next section we describe user model and interaction management. 2.3. I. L'ser model Kass and Finin [44] claim that individualized user models are essential for g o o d explanations ~ h e n the users differ in their knowledge o f the d o m ai n . Loosely speaking, they perceive the user model as a knowledge source c o n t a i n i n g explicit assumptions on all aspects o f the user that m ay be relevant for the interaction o f the system with the user. A n y interactive system has a user model. However. most systems maintain an implict user model because o f the assumptions a b o u t the user made during system design. Representing user model explicitly allows to maintain it dynamically during the p r o g r a m execution, and thus to achieve the desired level o f system interface flexibility tcf. C o o m b s and Alty [45], who stored e r r o r patterns as a n o rm at i v e user models for users o f Prolog). The simplest technique for building user models involves classifying users as novices and updating their status as they d e m o n s t r a t e imp ro v em en t s [46]. A m o re discriminating technique, allowing more efficient teaching, involves comparing the user's p e r f o r m a n c e with that o f the expert. It is assumed that the user knows a b o u t the underlying concept, if its derivative is used correctly [47, 48]. On the o t h e r hand, by classifying the errors that are m ad e by the user, the underlying deficiencies may be uncovered. T h e most sophisticated technique is the stereotype user modeling I n t e l l i g e n t i n t e r a c t i o n in d i a g n o s t i c e x p e r t s y s t e m s I 17 Table I. Typical a n a l ) s t errors and corresponding user stereotypes Errors Interpretauon Unnccessar) measurement Wrong measurement Wrong conclusion Earl)' conclusion Late concluston No conclusion Lack of technological suggestions Lack of plan Misunderstanding of schematic. lack of sktll to use eqmpment Misunderstanding of schematic Laziness Lack of self-confidence Lack of persistence "l'~ory tower" complex, lack of understanding of manufacturing process which involves describing the user by a set o f characteristics [49]. Examples o f such systems include a bibliographical system G R U N D Y [39], and a real estate r e c o m m e n d a t i o n system [50]. M c R uses a t r o u b l e s h o o t e r ' s model which is a c o m b i n a t i o n o f stereotyping and user e r r o r classification techniques. T h e user model is useful during the t r o u b l e s h o o t i n g session to select the way in which a t r o u b l e s h o o t i n g advice is constructed and displayed, Once such an e r r o r - s t e r e o t y p e table (Table I ) is constructed, the t r o u b l e s h o o t e r s can be ranked, depending on a linear c o m b i n a t i o n o f their scores in the table. When a novice analyst is interacting with the p r o g r a m , the entire is displayed, including the c o m p o n e n t location, description o f the test e q u i p m e n t used, m easu rem en t p r o c e d u r e and the meaning o f the readouts. As the user becomes m o r e proficient in the use o f the system, such a detailed display becomes annoying. A part o f the advice is now sufficient. At the next level o f proficiency, it is enough to display just the location o f the measured c o m p o n e n t instead o f the complete advice. This i n fo rm at i o n now can be supplemented by the display o f the relevant part o f the fault tree. Finally, for the expert, it is en o u g h to display just the list o f suspected c o m p o n e n t s and a scaled dov~.n fault tree. We notice that, as the user sophistication increases, we may condense more and m o re i n f o r m a t i o n at the increasingly abstract levels. We note here that in o r d e r to diagnose correctly the user's errors, it is not sufficient to observe the last user's action. T h e trace o f the entire process that the user traversed in arriving to the current situation is needed in o r d e r to make a correct conclusion a b o u t user's errors [51-53]. F o r example, if a user, t r o u b l e s h o o t i n g a signal path, consisting o f the c o m p o n e n t s C1, C2, C3, C4, C5 in that order, and having observed good input to CI and bad o u t p u t from C4, measures the o u t p u t from C5, then the user most likely m i s u n d e r s t o o d the signal path. On the o t h e r hand, if the user measures the input to C2 then it is most likely that she does not have a good t r o u b l e s h o o t i n g plan. 2.4. Logic Programming Implementation Before continuing with the discussion a b o u t user modeling and interaction, we digress briefl2, to highlight several points on system implementation. O u r graphic interactive interface has been constructed by efficiently co m b i n i n g knowledge and metaknowledge. Aiello et al. [54] describe three a p p r o a c h e s o f embedding m et ak n o w l ed g e in a system. The most primitive a p p r o a c h is to "'hardwire'" metaknowledge by simply writing it as pieces o f code in the system. Such an a p p r o a c h results in extending the implementation o f the system with the procedures that actually instantiate the variables o f metaknowledge. An example o f such an a p p r o a c h is the early rule-based expert systems where each separate case has to be covered by a specific rule instance. A second a p p r o a c h to c o m b i n e knowledge and m et ak n o w l ed g e is the metalanguage a p p r o a c h as in ML [55, 56]. However, using this a p p r o a c h prevents the higher levels o f metareasoning. A third a p p r o a c h allows the user to access both the object and the metalevel simultaneously, as in F O L [57]. This a p p r o a c h requires that both the language and metalanguage have the same form o f expression. A n o t h e r requirement in the third a p p r o a c h is that both levels have access to the inference engine, allowing for proving both theo rem s and metatheorems. T h e user is referred to Ref. [54] for further discussion a b o u t the ways o f a m a l g a m a t i n g knowledge and metaknowledge. 118 Y, LIROV and S. RAVIKUMAR It is suffice to note here that M c R implements the third approach using the m e t a p r o g r a m m i n g techniques o f Prolog. Accordingly, we may state that a subcircuit X is fault}' as follows: indict(X, Z):- suspect(X), generate_suspect (X,Suspects), generate_ideas(Suspects,Ideas), try_ideas( Ideas, Suspects,Z). T h e predicate suspect/lis true if its argument X has g o o d input and bad o u t p u t (this fact may be acquired from the user). T h e predicates g e n e r a t e _ s u s p e c t s / 2 and g e n e r a t e _ i d e a s / 2 are the knowledge base access predicates, where the first is used to subdivide the circuit X into a set o f faulty subcircuits, and the second to derive the set o f indictment m e t h o d s associated with the subcircuits. Here the object-let, el fact a b o u t the c o m p o n e n t Z being the reason for circuit X to fail is derived provided that the metalerel condition ofprol'ability holds between the set o f relevant facts and the goal try-ideas (Ideas,Suspects,Z): try_ideas([IdealRest],Suspects,BadGuy):- do_or (Idea,Suspects, BadG uy), try_ideas(Rest,BadG uy). do_or (A,[XlY],Z):- T = .. [X,A,Z],T, do_or (A,Y,Z). T h e meta-predicate t r y _ i d e a s / 3 creates and executes ( d o _ o r / 3 ) the goals required to indict the suspected subcircuits, as long as there are ways to indict (ideas) and as long as all the subcircuits are not indicted. N o t e that a circuit pack can be faulty for different reasons and hence we not not declare X indicted by the following object-lel,el sentence: indict(X,Z):- s u s p e c t ( X ) , i d e a ( Z ) . T h e reader is referred to Ref. [3] for further i n f o r m a t i o n a b o u t the code o f the circuit pack diagnostic expert s~stem. O u r interface consists o f two basic modules: the test tree m a n a g e r and the user model manager. Test tree m a n a g e r consists o f the modules to reify the diagnostic tree and to handle the knowledge bases o f advice and o f signal path and pins in the circuit pack. It also manages the d at ab ase o f relations which define the mapping between the nodes o f the test tree and the graphical objects representing the top view o f the c o m p o n e n t s [58]. Th e user model m a n a g e r maintains a c o u n t e r representing the level o f the user's proficiency and the conditions for achieving the next level o f proficiency. When the value o f the c o u n t e r exceeds a preset n u m b e r o f allowable errors, the system issues the c o r r e s p o n d i n g advice. 2.5..4daptire Troubleshooting hlterface Any user modeling system must provide the following i m p o r t a n t functions: user model representation scheme, interface between the user model and the rest o f the expert system, and user model acquisition. O u r c o m p u t a t i o n a l paradigm for implementing intelligent user interfaces is a loop consisting o f the following four activities: m o n i t o r , abstract, co m p are, and react. Th erefo re, M c R consists o f a m o n i t o r to acquire knowledge a b o u t the user, a fault identifier to m ap from user actions to user descriptions, and a proficiency table to m ap from user model to advice level. The m o n i t o r simply presents the user with the top view o f the circuit pack and accepts from the user the i n f o r m a t i o n a b o u t the t r o u b l e s h o o t i n g session. This i n f o r m a t i o n contains the location o f the measurements, and their results. T h e user fault identifier matches the a b o v e i n f o r m a t i o n to one o f the diagnostic trees developed at the knowledge acquisition stage. When a significant discrepanc}' is identified, the user is offered I n t e l l i g e n t i n t e r a c t i o n i n d i a g n o s t i c e x p e r t s y s t e m s 1 1 9 I u s e r actions S~gnal p a t h m a n i p u l a t i o n t r e e D m g n o s l i c tree decompo'>lllOrl Ll~er Inlerface I I r o u b l e s h o o t i n g ad'. ice IA+er ] M o d e l +ngnal p a t h F i g . I . A d a p t i v e t r o u b l e s h o o t i n g k n o w l e d g e b a s e s c h e - m a t i c . l i T , . . . . . I L ~ d L o ¢ l t i o n : I c 9 - p 8 | [ ] V a l u l : r o o d mi~t! | [ ] I ~ v l c e : D V N I F ++ : + + I -':" ."" '~?. :" ' ':: i':~.': N~.-'~ ~ .:.~....... ,,,,..:.. ~_" I1.....:.:...K+~ • :: :: : ~..~::~:"::. ':...'.'..:..:.':....:'" :'..': ":" .'.:...." ..',.:.:..:~::.:::..: ".'.'.... ~ ,:.:x."::.':':.q~:::::.~,'~.~ , . . . . . . . . . . ~ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ~ . , , . - , , . , . . ; . ; . . . , . . . . . . . . . . . • ~ . . ~ : . . : : . ~ . ~,.. • . A . . : . c . , : : ; . ~ ' q < , . : . . . . : . . : . ' . . . . : : ' : . : . ' : . . : : . . . • . . . . . ~ ~:..~.~: : ; : > . ' : . . . : " . . . : . ~ . . : . c ~ : : . : . . . . . : . ~ : . h ::.. +-~ ~ . . ~ .... :...: ............ : . . . . . . . . .........~...,..,.,:+~, ~ , :;.. :.:.....: . . . . :...:.~,:, . ~ : ~ . :: • ~ . ~ . , , . , . . . • , . . ; . . . . . k . ~ . & : . ~ ; : : ; : ~ . . . ~ : ~ f...:....:.:..:...~,;~.~..-..:~, ~ v . 0 ; . : : : : , : . " - ' - ; . - ~ , ~ . ~ . . . . . ~ : . : ' . : ? : ' i : ~ " '. ,'...~"'x~ ..~.~:.:.+~::...:~:,+: .:.:.:...: .+..:. ................ k%.:.'..'. : : . . ~ . . ~ . . ~ . . . . . . . . ," ~ : ~ . ~ " - " ' ~ ' . ~ -:.:~ ........... .:'::: .. :.:~.~::,~:i:i.: + i:i:i: .~.iil i:..,,: :ii: ::: .:i:i.'?:'i:..'.::::i:.:..+ +:'.::~:~'~:'~+ii:.'.'ii:~:gi:~:~:~::~:k.~'+~ ~ ~ : ~ :~.+.:.+... ~ , : ~ : • ~. ~z.~.:.'.: .~. +:.: :.." ::. :..::: .;:: +:.-" . ~:. ~ ~-" :++:~:'~: ~: :t:-: :.~ ~:~.~ + + ' . . ~ . ; , ' ~ + ,'~.~',~ • ~ . : . . " .. ~ . . . . " : . ' - - - " . : : . . : : . . : . . . . . / v :.'.::....:...'.: " z : . × . ; . . . ' ~ . ~ . . v x . . : v ' . : . . . : . : . . , . : . : . . . . . ~ . . . . ~ : • . . . . ~ . . : . ~ . . ' : . , . ~ + • ~:'.'~:.'::, :':: ~"-~::~!i..~::~::. • ::::: :~'~.:: ~.~".¢.". :.~t. ~, ~',,.'~: :~:~: :.-~ .~. ~ e . ~ . ~ . ~ i . ~ . • :.:~.:. , . . : ~ , ' : : c ' ~ . . , ~ " : ' ~ ' : : : ' : ' :: ' : . :.'..'..'.':.~ x,";::. ':'.'::':: : . : : ~ ;;.>.':;" : : . ' , :" x,..'.:.. ' • :::~ .:.....,~'~.~ ..... i?,,x--~',~ . . . . .... +. • ~ +.,..:.:.:~:,.~.~.i~+ , ::.. ~,~,.,~"~:.'Y~+:..~.:,,~ ,-,,x~ F i g . 2. A d v i c e a t e l e m e n t a r y l e v e l . help and the proficiency c o u n t e r is decremented. T a b l e I in Section 3.3 lists possible user faults. We differentiate between three basic kinds o f t r o u b l e s h o o t i n g errors: misunderstanding o f schematic, misunderstanding o f t r o u b l e s h o o t i n g strategy and misinterpretation o f the measure- ment. T h e proficiency level is matched also for the a p p r o p r i a t e level o f abstraction which should be used when presenting i n f o r m a t i o n to the user. Low proficiency level c o r r e s p o n d s to high level o f detail and vice versa. The schematic and m e a s u r e m e n t related erro rs can be ident- ified by c o m p a r i n g the true signal path with the signal path abstracted from user measurement sequence observations. The abstractions rules are presented in Section 3.6. Th e t ro u b l esh o o t i n g strategy related errors can be identified by c o m p a r i n g the m easu rem en t sequence with the ideal t r o u b l e s h o o t i n g tree. Using a rule-based knowledge representation, we m a y subdivide o u r system into a hierarchy o f rules (Fig. 1). At the first level are the rules which allow to d e c o m p o s e the knowledge base into a sequence o f c o m p o n e n t s c o r r e s p o n d i n g to the path o f signal flow in the unit under test. At the second level are the rules which use the user supplied measurements to manipulate the sequence o f c o m p o n e n t s obtained at the first level. Every m an i p u l at i o n triggers a rule in the third set o f rules which describe the h u m a n t r o u b l e s h o o t e r s behavi o r patterns. Every rule in the third set m ay issue a t r o u b l e s h o o t i n g advice. 2.6. Diagnostic Tree Decomposition Rules T h e diagnostic tree d e c o m p o s i t i o n rules are used to generate an abstract signal path which will serve as a model for the actual signal path. This model serves the p u rp o se o f m o n i t o r i n g and analyzing in a convenient way the sequence o f actions o f the h u m a n t ro u b l esh o o t er. Rule 2.6. I If c u r r e n t diagnostic rule has subrules then invoke Rule 2.6. I for each o f t h e subrules and collect the current rule ld in the abstract signal path and invoke Rules 2.6.2 and 2.6.3. Rule 2.6.2 Rule 2. 6.3 If current diagnostic rule is a replace rule then tag the c o r r e s p o n d i n g entry in the signal path as a replace entry. If current diagnostic rule is a test rule then tag the c o r r e s p o n d i n g entry in the signal path as a test entry. 120 Y. Lmov and S. RAVmUrdAR Note. T h e resulting signal path m a y contain n o n - u n i q u e elements which are ignored by Rules 2.6,2 and 2.6.3. 2. 7. Abstract Signal Path Manipulation Rules T h e abstract signal path m a n i p u l a t i o n rules maintain the cu rren t status o f the model derived by the previous set o f rules. T h e model is updated either by deleting its parts which b eco m e obsolete due to the results o f test observations, or by replacing it by a new model due to a charge in the t r o u b l e s h o o t i n g strategy. Rule 2. Z I If the o u t c o m e o f the test is g o o d then delete from the c u r r e n t abstract signal p at h all upstream entries. Rule 2. Z2 If the o u t c o m e o f the test is bad then delete from the current abstract signal p at h all d o w n s t r e a m entries. Rule 2. 7.3 If c u r r e n t test point is the r o o t o f the diagnostic tree then select new diagnostic tree 2.8. User Modeling Rules T h e last set o f rules actually selects the a p p r o p r i a t e t r o u b l e s h o o t i n g advice. These rules are used to decide whether to dispense a high level advice due to a m i n o r t r o u b l e s h o o t i n g strategy mistake, or to advice at a more detailed level. Such an advice is given because o f a significant lack o f circuit pack u n d e r s t a n d i n g on the part o f the h u m a n t r o u b l e s h o o t e r discovered by one o f the rules. Rule 2.8. I If current test point does not belong to abstract signal path then display the signal path. Rule 2.8.2 If c u r r e n t test point differs from the ro o t o f the cu rren t diagnostic tree then display the diagnostic tree. Rule 2.8.3 If activated Rules 2.8. I or 2.8.2 m o r e than three times then disperse a c o n v e n t i o n a l t r o u b l e s h o o t i n g advice. 2.9. Intelligent Interaction Example T h e following p a r a g r a p h s illustrate the different kinds o f advice the system can issue depending on the user. The most detailed advice which would be given to a notice user, is shown in Fig. 2. In this m o d e the system advices the user on what-to-do by telling her the location to be tested, the value to be expected, the measuring device to use and the p r o c e d u r e to be followed for making the measurement. This advice would be given to a user who does not have a t ro u b l esh o o t i n g strategy and whose proficiency c o u n t e r value is low. T h e next advanced level o f advice, which would be given to a t r o u b l e s h o o t e r who has a basic understanding o f the test set, various measuring devices, and the circuit pack in this mode, is shown in Fig. 3. N o w the system displays the entire diagnostic tree which was constructed based on good troubleshooting strategies (e.g. "divide and c o n q u e r " ) for this particular pack. Th e system identifies the user as one who d o e s n ' t have a g o o d t r o u b l e s h o o t i n g plan and who might need a longer time to diagnose a fault. T h e tree displays only the location to be tested. Error m test stratecJy-consult, d,ognostlc tree Intelligent interaction in diagnostic expert systems 121 Fig. 3. A more advanced advice. Finally, the most advanced level of advice is shown in Fig. 4. Such an advice would be given to an experienced troubleshooter. In this mode the system gives an advice on how-does-it-work, by displaying the signal path. The user is responsible for the troubleshooting strategy and also for completing any necessary details. 3. D I S C U S S I O N Since consultation interrupts working, advisory systems, both human and computerized, are frequently not welcome. Intelligent interaction deploys computerized advisory capabilities by merging working and learning modes. In this paper we explore the user modeling technique to implement intelligent interaction. If explanation--communicating knowledge to the user by the program--is viewed as a process of human knowledge acquisition, then a user model must be maintained by the program as a presumed human knowledge representation scheme. Carroll and Aaronson [2] showed how it is possible both to frustrate and to help people by providing "intelligent" help. We are dealing with the question of how to deploy a less than perfect advisory capability without having a sound theoretical ground. We believe that a flexible interaction environment based on the human model, maintained by the computer program can alleviate some of the user frustrations. The flexibility of the interaction is achieved by allowing to query the advisory knowledge base at different levels of detail, depending on the level of user sophistication. To obtain realistic experience and insight we chose to work in the area of engineering, mental, and economic significance: circuit-pack diagnostics. In this paper we describe the technical aspects of what we perceive to be an intelligent user interface to an advisory system of sufficiently economic impact. McR is a knowledge-based interactive graphic interface to a circuit pack diagnostic expert system. Scllernot,c error CORSULt $1qrlOL potl't bipolar bipolar blpoLar = ~ clemu~ bipolar J demux bipoLar Fig. 4. The most advanced and the least detailed advice. 122 Y. Lmov and S. RAVIKUMAR McR integrates three kinds of knowledge: domain knowledge (i.e. circuit pack), troubleshooting knowledge (i.e. diagnostic trees) and user knowledge (i.e. stereotype table), McR is an active interface which dynamically changes the amount of detail presented to the user as well as the input choices that the user is allowed to make. These changes are made using a knowledge-based model of the user and of the circuit pack troubleshooting domain. McR combines the strategies of interaction: it is able to advise both on "'how-to-do-it" and on "how-does-it-work". While "'how-to-do-it" advice is concerned with the measurement procedures [e.g. instrument (scope, etc.) and location (ic, pin)], the "'how-does-it-work" deals with the deeper knowledge representing the flow of the signal and the sequence of measurements. We have demonstrated that the analysis of user actions requires disassembly of observations along the different kinds of knowledge. Thus we develop special knowledge abstraction algorithms along with reification algorithms. The resulting system, instead of guiding the user by querying for input, monitors users actions, and offers help when needed. McR is used in conjunction with the STAREX expert system which is currently installed at an AT&T factory. Methodologically. we have shown a deep relationship between the problems of user interface construction, the problem of knowledge acquisition, and the problem of efficient diagnostic reasoning. All of these problems involve optimal selection of the sequences of measurements, and in fact, all of our implementations use the same software module for this purpose [5]. We have also proposed the basic intelligent interaction paradigm to be a loop, consisting of monitor, abstract, compare, and react activities (the resulting system, McR, demonstrates its software implementability and its name is a mnemonic to the paradigm). Additionally, we have demon- strated that metaprogramming techniques have great potential in implementing intelligent inter- action systems. McR, in particular, has been developed entirely in Prolog--a logic programming language with an especially convenient environment to develop metainterpreters. An important area for future research is that of improving a communication backchannel to allow the user to respond to the help in a less constrained way. Another possible way to improve our system is to develop an automated user activities monitoring system. We also foresee future intelligent advisors being able to perform "'what if" analysis of user actions, A way to do such a task is to substitute the "'correct" parts of the model with the assumed user actions and results, reify the knowledge base and compare the resulting tree with the ideal one. Additionally, we plan to incorporate the truth maintenance mechanisms in intelligent user interfaces. And finally, we foresee using intelligent user interfaces for knowledge acquisition. .4cknowledgements--We gratefully acknowledge the support of Walt Lawrence who made man~, invaluable comments and suggesttons: On-Ching Yue for guidance and insight, and Leon Levy for patient reading and comments. R E F E R E N C E S I. J. Carroll and J. McKendree, Interface design issues for advice-giving expert systems. Commun..4CM 30(1h 14-31 (1987). 2. J. Carroll and A. Aaronson, Learning by doing with simulated intelligent help. Commun..4CM, 31(9), 1064-1079, (1988) 3. Y. L~rov. S-fAREX--s~muhaneous test and replace c~rcmt pack troubleshooting expert system prototypmg and implementation. Engng .4pplic Art(£ Intell. 2, 3-18 (1989). 4. Y. LIrox. Circuit pack diagnostic expert systems--a sur,,'ey. Computers Math. Apphc. 18(4), 381-398 (1989), 5. Y. Lirov and O. Yue, Circmt pack troubleshooting via semantic control I: goal selection. Proc. IEEE Wkshp .4rtificial Intelligence lbr Industrial Application, Hitachi, Japan, pp. II8-122 (1988). 6. N. Morris and W. Rouse, Revie~ and evaluation of empirical research in troubleshooting. Human Factors 27(5), 503-530 11985) 7. L. H.~afil and R. L. Rl~,est, Constructing optimal binary decision trees is NP-complete. In~. Process Lett. 51, 15-17 (1976) 8. O. H. Ibarra and S. Sabra. Pol.~nomially complete fault detection problems. IEEE Trans. Comput. C-24(3), 242-250 (1976). 9. J. L. Saupe, Troubleshooting electronic equipment: an empirical approach to the ~dentificauon of certain requirements of a maintenance occupation. Ph.D. Dissertauon, University of Illinois (19541. 10. R. Glaser and J. Philhps, An analysis of proficiency for guided missde personnel: Ill, patterns of troubleshooting behaxior. Technical Bulletin, 55-16. American Institute for Research, Washington D.C., Aug. (1954). I I. C. Getty, s, C. Manning, T. Mehle and S. Fisher, Hypothes~s generation: a final report of three years of research. Techmtal Report 15-10-80. Decision Process Laboratory, Universit3 of Oklahoma, Norman, Okla, Oct. (1980). 12. Human engineering criteria for mihtary systems, equipment, and facihties. Report MIL-STD 1472C, Department of Defense. Washmgton, D.C. (itS81). Intelligent interaction in diagnostic expert systems 123 13. W. Buxton, M. R. Lamb, D. Sherman and K. C. Smith, Towards a comprehensive user interface management system. Comput. Graph. 17(3), 35-42 (1983). [4. S. L. Smith and Y. N. Mosier, Design guidelines for the user-system interface software; The Mitre Corporation Technical Report ESD-IR-48-190, Bedford. Mass. (1984), 15. A. F. Norcio and J. Stanley, Adaptive h u m a n - c o m p u t e r interfaces. Technical Report # NRL-9148, Naval Research Laboratory, Washington, D.C., Sept. (1988). 16. J. Brown, R. Burton and J. deKleer, Pedagogical, natural language, and knowledge engineering techniques in SOPHIE I, I! and Ill. In Intelligent Tutoring Systems (Eds D. Sleeman and J. Brown), pp. 227-282. Academic Press, New York (1982). 17. M. H. Richer and W. J. Clancey, G U I D O N - W A T C H : a graphic interface for viewing a knowledge based system. Technical Report, STAN-CS-85-1068, Stanford University, Calif. Aug. (1985). 18. E. Schoen and R. G. Smith, IMPULSE: a display oriented editor for STROBE. Proc..4AA1"86, pp. 356-358 (1986). 19. F. N. Hill, J. D. Ward and A. L. Yates, SandKAST: An automated knowledge acquisition system. SAND-87-0364C, Sandia, Dec. (1987). 20. J. H. Boose, Personal construct theo D and the transfer of human expertise. Proc. A A A I ' 8 4 (1984). 21. A. Stevens, B. Roberts and L. Stead, The use of a sophisticated graphics interface in computer-assisted instruction. IEEE Comput. Graph. Applic. Mar./Apr., 25-30 (1983). 22. M. L. Powell and M, A. Linton, Visual abstractions in an interactive programming environment. Proc. A C M SIGPLAN, Sigplan Notices, Vol. 18, pp. 14-21, Jun. (1983). 23. S. P. Reiss, Graphical program development with PECAN program development systems. Proc. A C M SIGSOFT/ S I G P L A N Software Engineering Syrup. Practical Software Development Environmems. Apr. (1984), also printed as Sigplan Not. 19(5), 30-41 (1984). 24. S. P. Ross, A conceptual programming environment. Proc. 9th Int. Conf. Software Engineering, Monterey. Calif. Mar,Apr. (1987). 25. C Rathke, Human-computer communication meets software engineering. Proc. 9th Int. Con(. Software Engineering, Monterey, Calif. Mar Apr. (1987). 26. G. P. Brown, R. T. Carling, C. F. Herot, D. A. Kramlich and P. Souza, Program visualization: graphical support for software development. Computer Aug., 27-34 (1985). 27. B. Melamed and R. J T. Morns. Visual simulation: the performance analysis workstauon. Computer. Aug. 87-94 (1985J. 28. A. Borning, The programming language aspects of ThingLab, a constraint-oriented simulation laboratory. A C M Trans. Progrng Lang. Syst. 3(4}, 353-387 11981). 29. T. Dudley, Graphics in software design. Computers. Graph. WId Feb. 35-42 (1986). 30. S. Greenberg and L. H. Witten, Adaptive personalized interfaces--a question of viability, Behavior Inf. Tech. 4( I ). 31-45 (1985). 31. E. A. Edmonds. Adaptive man-computer interfaces. In Computing Skills and the L'ser Inte(lace (Eds M. J. Coombs and Y. L. Ahy), pp. 389-426. Academic Press. London (1981). 32. M. V. Mason and R. C. Thomas. Experimental adaptive interface. Inf. Technol. Res. Des. Applic. 3(3), 162-167 (1984). 33. W. Sherman, SAUCI: self-adaptive user-computer interfaces. Ph. D. Dissertation, Umversity of Pittsburgh, Pittsburgh, Pa 119861. 34. R. Reichman-Adar, Extended person-machine interface. ,4rttf. In(ell. 22, 157-218 (1984). 35. W. B. Rouse, Human-computer interaction in the control of dynamic systems. Computing Sun'. 13, 71-100 (1981). 36. E. Risland. Ingredients of intelhgent user interfaces Int. J. Man-Mach. Stud. 21, 377-388 [1984). 37. W. Wahlster and A. Kobsa, Dialogue-based user models. Proc. IEEE 74(7), 948-960 (1986). 38. H. Mozelco, A human;computer interface to accommodate user learning stages. Communs .4CM 25(2), 100-104 (I 982). 39. E. Rich, User modeling ,,ia sterot~pes Cogn. Sei. 3, 329-354 (1979). 40. D. Crockford, Stand by for fun. In Interactit'e Multimedia (Eds S. A m b r o n and K. Hooper). Microsoft Press (1988). 41. G. Fischer, A. Lemke and T. Schwab, Kno~,ledge-based help systems. Proc. CH185 Human Factors tn Computing Systems. San Francisco, Calif.. 14-17 Apr. pp. 161-167. ACM, New York (1985). 42. I. Monarch and J. Carbonell, Coal SORT: A knowledge-based interface. IEEE Expert 2(I), 39-53 (1987). 43. W. B. Croft, The role of context and adaptation in user interfaces. Int. J. Man-Maeh. Stud. 21, 283-292 (1984). .,t4. R. Kass and T. Finin, The need for user models in generating expert system explanations. Int. J. Expert. Syst. (m press). 45. M. J Coombs and J. L. Alty, Expert systems: an alternate paradigm. Int. J. Man-Maeh. Stud. 20, 21-43 (1984). 46. D. A. Norman, Design rules based upon analyses of human error. Commun. ACM 26, 254-258 {1983}. 47. M. Matz. Tov, ards a process model for high school algebra errors. In Intelhgent Tutoring Systems (Eds. D. Sleeman and J. S. Bunon). Academic Press, New York (1982}. 48. R. Burton and J. S. Brown, A tutoring and student modehng paradigm for gaming environments. Proe. A C M S I G C S E , SIGCUE Joint S.vmp. (1976). 49. T. Finn and D. Darger, G U M S : A general user modeling system. University of Pennsylvania School of Engineering and Apphed Science, Technical Report MS-CIS-86-35, May (1986). 50. K. Morik and C. Pollinger, The real estate agent--modeling users by uncertain reasoning..41 Mag., pp. 44-52 (1985). 51. J. J. Allen and C. R. Perrault. Analyzing intentions in utterances. Art!£ Intell. 15, 143-178 (19801. 52. R Burton and J. S. Brown. An investigation of computer coaching for informal learning activities. In Intelligent Tutoring Systems (Eds. D. Sleeman and J. S. Brown}, pp. 137-155. Academic Press, New York ~1982). 53. M. R. Genesereth, The role of plans in intelligent teaching systems, in Intelligent Tutoring Systems (Eds D. Sleeman and J. S. Brown), pp. 137-[55. Academic Press, N e t ' York (1982). 54. L. Aid(o, C. Cecchi and D. Sartini, Representation and Use of Metaknotledge. Proc. IEEE 74( 10L 1304-1321 (1986). 55. M. Gordon, R. Milner and C. Wadsworth, Edinburg LCF: a mechanized logic for computation. Lecture ,Votes Computer Stience 78. Springer, N e t ' York (1979). 56. A. Wikstrom, Functional Programming t.'sing Standard ,l,lL. Prentice-Hall, Englewood Chffs, N.J. (19871. 57. R. Weyhranch, Prolegomena to a theory of mechamzed formal reasoning. AI J. 13, 133-170 (1980). 58. Y. L~rox, Computer aided software engineering of expert s.~stems. E,xpert Systems with Applications. Pergamon Press, Oxford [in press). 
work_qviavjkg3vh33n2acrcmsuxuve	----	A fuzzy group multi-criteria enterprise architecture framework selection model Expert Systems with Applications 39 (2012) 1165–1173 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A fuzzy group multi-criteria enterprise architecture framework selection model Faramak Zandi a, Madjid Tavana b,⇑ a Information Technology Management and Industrial Engineering, Faculty of Technology and Engineering, Alzahra University, Vanak, Tehran 19938-91176, Iran b Management Information Systems, Lindback Distinguished Chair of Information Systems, La Salle University, Philadelphia, PA 19141, USA a r t i c l e i n f o Keywords: Multi-criteria decision making Enterprise architecture Fuzzy logic Risk failure mode and effects analysis 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.07.120 ⇑ Corresponding author. Tel.: +1 (215) 951 1129; fa E-mail addresses: zandi@iust.ac.ir (F. Zandi), tavan URL: http://lasalle.edu/~tavana (M. Tavana). a b s t r a c t A large body of intuitive and analytical models has evolved over the last several decades to assist deci- sion makers (DMs) in enterprise architecture (EA) framework selection. While these models have made great strides in EA framework evaluation, the intuitive models lack a structured framework and the analytical models do not capture intuitive preferences of multiple DMs. Furthermore, crisp data are fundamentally indispensable in traditional EA framework selection methods. However, the data in real-world problems are often imprecise or ambiguous. The prior research in EA framework selection does not embrace both qualitative and quantitative criteria exhibiting imprecise and ambiguous value judgments. In this paper, we propose a novel fuzzy group multi-criteria model for EA framework eval- uation and selection. The contribution of the proposed model is fourfold: (1) it takes into consideration the qualitative and quantitative criteria and their respective value judgments; (2) it considers verbal expressions and linguistic variables for qualitative judgments which lead to ambiguity in the decision process; (3) it handles imprecise or vague judgments; and (4) it uses a meaningful and robust multi- criteria model to aggregate both qualitative and quantitative data. We use a real-world case study to demonstrate the applicability of the proposed framework and exhibit the efficacy of the procedures and algorithms. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction The increasing global competition and the rapid advances in information systems have led organizations to search for more effective and efficient ways to manage their business. The enter- prise architecture (EA) frameworks can ensure interoperability of information systems and improve the effectiveness and effi- ciency of business organizations. However, the increasing turbulence in the business environment and the larger number of alternatives with conflicting criteria has made the selection of EA frameworks a difficult and complex task. The EA frame- work selection problems are multi-criteria problems that em- brace both qualitative and quantitative criteria. Nevertheless, the traditional selection methods overemphasize quantitative and economic analysis and often neglect to consider qualitative and non-economic data in the formal selection process. When facing such multi-criteria problems, the literature and research show that the following difficulties may be encountered in the decision process: ll rights reserved. x: +1 (267) 295 2854. a@lasalle.edu (M. Tavana). 1. Decision makers (DMs) often use verbal expressions and linguistic variables for qualitative judgments which lead to ambi- guity in the decision process (Poyhonen, Hamalainen, & Salo, 1997). 2. DMs often provide imprecise or vague information due to lack of expertise or unavailability of data (Kim & Ahn, 1999). 3. Meaningful and robust aggregation of qualitative and quantita- tive data causes difficulties in the decision process (Valls & Torra, 2000). 4. A decision process is not complete without fully taking into consideration all the criteria and their respective value judg- ments (Belton & Stewart, 2002; Yang & Xu, 2002). Before an organization takes up a particular EA framework, there is need to consider and evaluate the possible alternative frameworks, and then select an appropriate one through a collab- orative effort involving all key stakeholders (Gammelgêard, Simonsson, & Lindstrèom, 2007). Over the past several years, the selection of EA frameworks has garnered considerable attention from both practitioners and academics. Fayad and Hamu (2000) have presented a detailed list of guidelines and attributes for selec- tion of EA frameworks. Tang, Han, and Chen (2004) have studied and compared the EA frameworks analytically and provided a model of understanding through analyzing the goals, inputs and outcomes of six architecture frameworks. Nonetheless, these stud- ies did not consider any uncertainties such as uncertainty in the http://dx.doi.org/10.1016/j.eswa.2011.07.120 mailto:zandi@iust.ac.ir mailto:tavana@lasalle.edu http://lasalle.edu/~tavana http://dx.doi.org/10.1016/j.eswa.2011.07.120 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa eI The fuzzy weighted collective impact matrix with respect to the impact of m risk criteriaeL The fuzzy weighted collective likelihood matrix with respect to the likelihood of m risk criteriaeD The fuzzy weighted collective detection matrix with respect to the detection of m risk criteriaeI K The individual fuzzy impact matrix with respect to the impact of m risk criteria evaluated by the EA framework selection team member (EAFT)keLK The individual fuzzy likelihood matrix with respect to the likelihood of mrisk criteria evaluated by the EA framework selection team member (EAFT)keDK The individual fuzzy detection matrix with respect to the detection of m risk criteria evaluated by the EA framework selection team member (EAFT)k w(vp)K The voting power of the EA framework selection team member (EAFT)k for scoring (K = 1, 2, . . . , l) wj The importance weight of the jth criterion ~eijðIÞ The fuzzy weighted collective impact value of the EA framework with respect to the risk criterion j evaluated by the EA framework selection team member (EAFT)k ~eijðLÞ The fuzzy weighted collective likelihood value of the EA framework with respect to the risk criterion j evaluated by the EA framework selection team member (EAFT)k ~eijðDÞ The fuzzy weighted collective detection value of the EA framework with respect to the risk criterion j evaluated by the EA framework selection team member (EAFT)k ~ekijðIÞ The individual fuzzy impact value of the EA framework with respect to the risk criterion j evaluated by the EA framework selection team member (EAFT)k ~ekijðLÞ The individual fuzzy likelihood value of the EA framework with respect to the risk criterion j evaluated by the EA framework selection team member (EAFT)k ~ekijðDÞ The individual fuzzy detection value of the EA framework with respect to the risk criterion j evaluated by the EA framework selection team member (EAFT)k cj The jth criterion Ai The ith EA framework m The number of risk criteria l The number of EA framework selection team members n The number of EA frameworks ReP N The fuzzy risk priority numbers (RPN) matrix R The ordinal rank matrix R½Eðr~pnijÞ� The ordinal rank of Eðr~pnijÞ Eðr~pnijÞ The possibilistic mean value of r~pnij r~pnij The fuzzy risk priority number (RPN) for the ith EA framework with respect to the risk criterion j V The vector of the EA frameworks risks vi The risk of the jth EA framework 1166 F. Zandi, M. Tavana / Expert Systems with Applications 39 (2012) 1165–1173 subjective judgments or uncertainties due to lack of data and incomplete information. Multi-criteria decision making methods have also been suc- cessfully applied in various decision making situations that have many similar features with EA framework selection problems. The multi-criteria methods used to evaluate EA frameworks based on quantitative measurements include Technique for Or- der Preference by Similarity to Ideal Solution (TOPSIS) (Hwang & Yoon, 1981; Kahraman, Ates�, Çevik, Gülbay, & Erdoğan, 2007; Shih, 2008); Simple Additive Weighting (SAW) (Chou, Chang, & Shen, 2008; Zavadskas, Turskis, Dejus, & Viteikiene, 2007); Linear Programming Techniques for Multidimensional Analysis of Preference (LINMAP) (Xia, Li, Zhou, & Wang, 2006); COmplex PRoportional Assessment (COPRAS) (Kaklauskas et al., 2006; Zavadskas, Turskis, Tamošaitienė, & Marina, 2008). The multi-criteria methods used to evaluate EA frameworks based on qualitative measurements not converted to quantitative vari- ables include methods of verbal decision-making analysis (Berke- ley, Humphreys, Larichev, & Moshkovich, 1991; Andre’eva, Larichev, Flanders, & Brown, 1995; Larichev, 1992; Larichev, Brown, & Andre’eva, 1995; Larichev & Moshkovich, 1997; Flanders, Brown, Andre’eva, & Larichev, 1998). Comparative pref- erence methods based on pairwise comparison of alternatives include ELECTRE (Costa, Almeida, & Miranda, 2003; Roy, 1996) and PROMETHEE I and II (Brans, Vincke, & Mareschal, 1986; Diakoulaki & Koumoutsos, 1991; Wang & Yang, 2007). The multi-criteria methods used to evaluate EA frameworks based on qualitative measurements converted to quantitative variables include two widely known groups of methods, i.e. Analytic Hierarchy Process (AHP) (Ghodsypour & O’brien, 1998; Saaty, 1994) and fuzzy set theory methods (Roztocki & Weistroffer, 2006; Zimmermann, 2000). Traditional assessment techniques overemphasize quantitative and economic analysis and often neglect to consider qualitative and non-economic factors in the formal selection process. Fur- thermore, crisp data are fundamentally indispensable in tradi- tional EA framework selection methods. However, the data in real-world problems are often imprecise or ambiguous. The prior research in EA framework selection does not embrace both qual- itative and quantitative criteria exhibiting imprecise and ambig- uous value judgments. In this paper, we propose a novel fuzzy group multi-criteria model for EA framework evaluation and selection that takes into consideration (1) the qualitative and quantitative criteria and their respective value judgments; (2) the verbal expressions and linguistic variables for qualitative judgments which lead to ambiguity in the decision process; and (3) imprecise or vague judgments. The five EA frameworks considered in this study include: Federal Enterprise Architecture Framework (FEAF), Zachman Framework for Enterprise Architec- ture (ZACHMAN), The Open Group Architecture Framework (TOGAF), Treasury Enterprise Architecture Framework (TEAF), and Department of Defense Architecture Framework (DoDAF). See Urbaczewski and Mrdalj (2006) for an excellent review of the five EA frameworks considered in our study. The next section presents the mathematical notations and definitions used in our model. In Section 3, we illustrate the details of the proposed model followed by a real-life case study to demonstrate the applicability of the proposed model and exhibit the efficacy of the procedures and algorithms. In Sec- tion 5, we conclude with our conclusions and future research directions. 2. The mathematical notations and definitions Let us introduce the following mathematical notations and definitions: 3. The proposed model The model depicted in Fig. 1 is proposed to assess the alterna- tive EA frameworks and consists of several steps modularized into five phases: Fig. 1. The proposed framework. F. Zandi, M. Tavana / Expert Systems with Applications 39 (2012) 1165–1173 1167 In summary, we. � have group members estimate the impact, probability of occur- rence and probability of detection of certain risks involved in the selection of EA frameworks, � aggregate across group members by forming a weighted average, � calculate expected impact by multiplying impact and probabilities, � use an additive weighting scheme to aggregate across impact categories, � defuzzify results by forming expected values, � transform evaluations into Borda scores, and � select the alternative with the lowest risk score. Phase 1: Establishing the EA framework selection team. In the first phase, we establish an EA framework selection team. Let us assume that l EA framework selection team members are chosen to rank the EA frameworks: EAFT ¼ ½ðEAFTÞ1;ðEAFTÞ2; . . . ;ðEAFTÞk; . . . ;ðEAFTÞl� Phase 2: Identifying the EA frameworks. In this phase, the selec- tion team identifies a set of EA frameworks. Let us further assume that the selection team has identified n EA frameworks: A ¼ ½A1; A2; . . . ; An� Phase 3: Identifying the qualitative and quantitative risk criteria. In this step, the EA framework selection team identifies a set of qual- itative and quantitative risk criteria associated with the EA frame- works. Let c1,c2, . . . , cm and w1,w2, . . . , wm be the risk criteria and their importance weights, respectively. ð3Þ Phase 4: Prioritizing the EA framework risks. In this phase, the EA framework selection team prioritizes the EA frameworks according to the following five steps: Step 4.1: Constructing the individual fuzzy EA framework risk matrices. In this step, the fuzzy individual EA framework risks are assessed based on the following three attributes associated with a risk event: the impact, likelihood and detection of the qualitative and quantitative risks identified in Phase 3. The goal is to identify, quantify and remove or reduce risks in an EA framework. 4.1.1. Calculating the impact value of the EA frameworks. A project risk is defined as ‘‘an uncertain event or condition that, if it occurs, has a positive or negative effect on a project’s objectives’’ (PMI, 2009). The EA framework risk is represented by the impact attri- bute in our model. The EA framework selection team uses a fuzzy set [1–10] to assign an impact number (I) to each EA framework. The impact value guidelines in Table 1 suggested by Carbone and Tippett (2004) are used to assign the impact values. These fuzzy numbers are used to prioritize the EA frameworks with respect to the impact of the risk criteria. The individual fuzzy impact matrix with respect to the impact of m risk criteria evaluated by the EA framework selection team member (EAFT)k will be as follows: ð1Þ Let ~ekijðIÞbe the following trapezoidal fuzzy number: ~ekijðIÞ¼ e k ijðIÞ � �o ; ekijðIÞ � �a ; ekijðIÞ � �b ; ekijðIÞ � �c� � ð2Þ Consequently, substituting Eq. (2) into matrix (1), it can be rewrit- ten as: 4.1.2. Calculating the likelihood values of the EA frameworks. Next, we consider the risk occurrence frequencies in our model by assigning a likelihood ranking (L) to each EA framework Table 2 The likelihood value guidelines. Very unlikely Probably will not occur Equal chance of occurring or not Will probably occur Very likely to occur Fuzzy number 1 or 2 Fuzzy number 3 or 4 Fuzzy number 5 or 6 Fuzzy number 7 or 8 Fuzzy number 9 or 10 Table 1 The impact value guidelines. Very low Low Average High Very high Fuzzy number 1 or 2 Fuzzy number 3 or 4 Fuzzy number 5 or 6 Fuzzy number 7 or 8 Fuzzy number 9 or 10 1168 F. Zandi, M. Tavana / Expert Systems with Applications 39 (2012) 1165–1173 using fuzzy numbers 1–10. The likelihood value guidelines in Table 2 suggested by Carbone and Tippett (2004) are used to assign the impact values. These fuzzy numbers are used to prioritize the EA frameworks with respect to the impact of the risk criteria. The individual fuzzy likelihood matrix with respect to the likelihood of m risk criteria evaluated by the EA framework selection team member (EAFT)k will be as follows: ð4Þ Let ~ekijðLÞ be the following trapezoidal fuzzy number: ~ekijðLÞ¼ e k ijðLÞ � �o ; ekijðLÞ � �a ; ekijðLÞ � �b ; ekijðLÞ � �c� � ð5Þ Consequently, substituting Eq. (5) into matrix (4), it can be rewrit- ten as: ð6Þ 4.1.3. Calculating the detection values of the EA frameworks. Detection values represent the organization’s ability to detect the risk event with enough time to plan for a contingency and act upon the risk (Carbone & Tippett, 2004). The detection value is a mea- sure of being able to predict the specific risk event in the future. The goal is to detect the risk as early as possible. Those risks with high detection values may need one additional control mechanism for early warning. Each EA framework is assigned a detection value (D), again using fuzzy numbers 1–10. This ranks the ability of planned tests in each EA framework to detect risk events in time. The detection value guidelines in Table 3 suggested by Carbone and Tippett (2004) are used to assign the detection values. Certainly, the detection assignment is subjective, but no more so than the assignment of the likelihood and impact values for the common risk matrix method. The individual fuzzy detection matrix with respect to the detection of m risk criteria evaluated by the EA framework selection team member (EAFT)k will be as follows: ð7Þ Let ~ekijðDÞ be the following trapezoidal fuzzy number: ~ekijðDÞ¼ e k ijðDÞ � �o ; ekijðDÞ � �a ; ekijðDÞ � �b ; ekijðDÞ � �c� � ð8Þ Consequently, substituting Eq. (8) into matrix (7), it can be rewrit- ten as: ð9Þ Table 3 The detection value guidelines. Highly effective and it is almost certain that the risk will be detected with adequate time Moderately high effectiveness Medium effectiveness Unproven or unreliable; or effectiveness of detection method is unknown to detect in time There is no detection method available or known that will provide an alert with enough time to plan for a contingency Fuzzy number 1 or 2 Fuzzy number 3 or 4 Fuzzy number 5 or 6 Fuzzy number 7 or 8 Fuzzy number 9 or 10 F. Zandi, M. Tavana / Expert Systems with Applications 39 (2012) 1165–1173 1169 Step 4.2: Constructing the fuzzy weighted collective EA framework risk matrices. In this step, the fuzzy weighted collective EA frame- work risks are assessed based on the likelihood, impact and detec- tion of the identified risks as follows: 4.2.1. Calculating the impact values of the EA frameworks. The fuz- zy weighted collective impact matrix with respect to the impact of m risk criteria will be as follows: ð10Þ or: ð11Þ ð16Þ where: ~eijðIÞ¼ Pl k¼1ðwðvpÞkÞ ~e k ijðIÞ h i Pl k¼1 w vpð Þk ð12Þ 4.2.2. Calculating the likelihood values of the EA frameworks. The fuzzy weighted collective likelihood matrix with respect to the likelihood of m risk criteria will be as follows: or: ð14Þ where: ~eijðLÞ¼ Pl k¼1ðwðvpÞkÞ ~e k ijðLÞ h i Pl k¼1wðvpÞk ð15Þ 4.2.3. Calculating the detection values of the EA frameworks. The fuzzy weighted collective detection matrix with respect to the detection of m risk criteria will be as follows: or: ð17Þ ð13Þ 1170 F. Zandi, M. Tavana / Expert Systems with Applications 39 (2012) 1165–1173 where: ~eijðDÞ¼ Pl k¼1ðwðvpÞkÞ ~e k ijðDÞ h i Pl k¼1 wðvpÞk ð18Þ Step 4.3: Constructing the fuzzy RPN matrix. Next, we calculate the fuzzy RPN by multiplying the three fuzzy impact, likelihood and detection values. This has to be done for each EA framework with respect to the risk criteria. The EA frameworks that have the highest RPN should be given the lowest selection priority and those with the lowest RPN should be given the highest selection priority. Therefore, the fuzzy RPN matrix will be as follows: ð19Þ or: ð20Þ where: r~pnij ¼ ~eijðIÞ:~eijðLÞ:~eijðDÞ ð21Þ Step 4.4: Constructing the ordinal rank matrix. Next, we rank the n EA frameworks based on Borda’s score. That is, for each strategic criterion of preference ordering of the EA frameworks, scores of n � 1,n � 2, . . . , 1, 0 are assigned to the first, second . . . and the last ranked framework: ð22Þ Since r~pnij is a trapezoidal fuzzy number, r~pnij ¼ððrpnijÞ o ; ðrpnijÞ a ;ðrpnijÞ b ;ðrpnijÞ cÞ, its expected value can be derived as follows: Eðr~pnijÞ¼ ðrpnijÞ o þðrpnijÞ a 2 þ ðrpnijÞ c �ðrpnijÞ b 6 ð23Þ Step 4.5: Calculating the EA frameworks risks. The vector of the EA frameworks risks will be calculated as follows: V ¼ ½v 1 v 2 � � � v n � T ð24Þ where: ~v i ¼ Pm j¼1ðwjÞR½Eðr~pnijÞ�Pn i¼1 Pm j¼1wj R½Eðr~pnijÞ� ð25Þ Phase 5: Selecting the optimal EA framework. In this phase, the optimal EA frameworks ranking is determined based on the risks values obtained in phase (4). For this purpose, these values are con- sidered as the coefficients of the objective functions in the follow- ing proposed mathematical model with a series of applicable constraints such as critical impact and RPN values. Min Z2 ¼ v 1:x1 þ v 2:x2 þ�� �þ v n:xn ðModel PÞ Subject to : f1ðx1; x2; . . . ; xnÞ6 0 f2ðx1; x2; . . . ; xnÞ6 0 .. . frðx1; x2; . . . ; xnÞ6 0 x1 þ x2 þ�� �þ xn ¼ 1 xi ¼ 0; 1ði ¼ 1; 2; . . . ; nÞ where: fi(x1,x2, . . . , xn) is a given function of the n EA frameworks. In the next section, we present a real-life case study to demonstrate the applicability of the proposed model and exhibit the efficacy of the procedures and algorithms. 4. The case study The Institute for Energy and Hydro Technology (IEHT) is the largest energy institute in Iran. IEHT is located on a 65 acre campus and has a 50,000 person training capacity per month. The institute employs over 80 full-time and 100 part-time faculties. We used the group EA approach proposed in this study to select an EA for IEHT. A committee of nine DMs from marketing, finance, and information technology was formed to participate to evaluate the following five frameworks: ZACHMAN, TEAF, TOGAF, DoDAF and FEAF. Phase 1: In this phase, we worked with the management team and established an EA framework selection team which included: (ITPT)1: The information technology manager (ITPT)2: The research and development manager (ITPT)3: The capital budgeting manager (ITPT)4: The quality assurance manager Phase 2: In this phase, the EA framework selection team agreed to consider the following five EA frameworks suggested by Urb- aczewski and Mrdalj (2006): A1: ZACHMAN A2: TEAF A3: TOGAF A4: DoDAF A5: FEAF Phase 3: In this phase, the selection team decided to consider the strategic criteria presented in Table 4 for the evaluation and selection of the EA frameworks. Phase 4: In this phase, we used Eqs. (1)–(18) and calculated the impact, likelihood and detection values of the EA frameworks pre- sented in Tables 5–7: In step 4.3, we used Eqs. (19)–(21) and calculated the possibilis- tic mean values of the fuzzy risk priority numbers (RPN) provided in Table 8. In step 4.4, we used Eqs. (22) and (23) and constructed the ordinal rank matrix given in Table 9: F. Zandi, M. Tavana / Expert Systems with Applications 39 (2012) 1165–1173 1171 In step 4.5, we used Eqs. (24) and (25) and calculated the vector of the EA framework risks given in Table 10. These risks values were determined according to the following importance weight vector: ðw1; w2; . . . ; w8Þ¼ ð0:15; 0:5; 0:5; 0:1; 0:13; 0:1; 0:17; 0:25Þ Phase 5: In this phase, the selection team identified the follow- ing constraints for each EA framework with respect to the risk im- pacts. All constraints were set less than or equal to the critical impact value (8). Next, the optimal EA framework was identified using the following mathematical model: Max:Z ¼ 0:29xFEAF þ 0:21xZACHMAN þ 0:16xTOGAF þ 0:06xTEAF þ 0:28xDODAF ðModel PÞ Subject to : Eð~eZACHMAN1ðIÞÞ:xZACHMAN 6 8; Eð~eZACHMAN2ðIÞÞ:xZACHMAN 6 8; Eð~eZACHMAN3ðIÞÞ:xZACHMAN 6 8; Eð~eZACHMAN4ðIÞÞ:xZACHMAN 6 8; Eð~eZACHMAN5ðIÞÞ:xZACHMAN 6 8; Eð~eZACHMAN6ðIÞÞ:xZACHMAN 6 8; Eð~eZACHMAN7ðIÞÞ:xZACHMAN 6 8; Eð~eZACHMAN8ðIÞÞ:xZACHMAN 6 8; Eð~eTEAF1ðIÞÞ:xTEAF 6 8; Eð~eTEAF2ðIÞÞ:xTEAF 6 8; Eð~eTEAF3ðIÞÞ:xTEAF 6 8; Eð~eTEAF4ðIÞÞ:xTEAF 6 8; Eð~eTEAF5ðIÞÞ:xTEAF 6 8; Eð~eTEAF6ðIÞÞ:xTEAF 6 8; Eð~eTEAF7ðIÞÞ:xTEAF 6 8; Eð~eTEAF8ðIÞÞ:xTEAF 6 8; Eð~eTOGAF1ðIÞÞ:xTOGAF 6 8; Eð~eTOGAF2ðIÞÞ:xTOGAF 6 8; Eð~eTOGAF3ðIÞÞ:xTOGAF 6 8; Eð~eTOGAF4ðIÞÞ:xTOGAF 6 8; Eð~eTOGAF5ðIÞÞ:xTOGAF 6 8; Eð~eTOGAF6ðIÞÞ:xTOGAF 6 8; Eð~eTOGAF7ðIÞÞ:xTOGAF 6 8; Eð~eTOGAF8ðIÞÞ:xTOGAF 6 8; Eð~eDODAF1ðIÞÞ:xDODAF 6 8; Eð~eDODAF2ðIÞÞ:xDODAF 6 8; Eð~eDODAF3ðIÞÞ:xDODAF 6 8; Eð~eDODAF4ðIÞÞ:xDODAF 6 8; Eð~eDODAF5ðIÞÞ:xDODAF 6 8; Eð~eDODAF6ðIÞÞ:xDODAF 6 8; Eð~eDODAF7ðIÞÞ:xDODAF 6 8; Eð~eDODAF8ðIÞÞ:xDODAF 6 8; Eð~eFEAF1ðIÞÞ:xFEAF 6 8; Eð~eFEAF2ðIÞÞ:xFEAF 6 8; Eð~eFEAF3ðIÞÞ:xFEAF 6 8; Eð~eFEAF4ðIÞÞ:xFEAF 6 8; Eð~eFEAF5ðIÞÞ:xFEAF 6 8; Eð~eFEAF6ðIÞÞ:xFEAF 6 8; Eð~eFEAF7ðIÞÞ:xFEAF 6 8; Eð~eFEAF8ðIÞÞ:xFEAF 6 8; xFEAF þ xZACHMAN þ xTOGAF þ xTEAF þ xDODAF ¼ 1 xFEAF; xZACHMAN; xTOGAF; xTEAF; xDODAF ¼ 0; 1 The results from model (P) identified TEAF framework as the opti- mal EA framework. We communicated our findings to the IEHT management who proceeded with the implementation of our recommendation. Table 4 Strategic risk criteria associated with the enterprise architecture frameworks. Risk criteria Description Organizational risk The stability of the management; organizational support for User risk The lack of user involvement during the EA framework deve Requirement risk Frequently changing requirements; incorrect, unclear, inade Structural risk The strategic orientation of the application; quite number o Team risk Insufficient knowledge or inadequate experience among tea Complexity risk Whether new the enterprise architecture is used; the compl of links to existing systems is required Competition risk Strong competitor reactions that may prevent the enterprise Market risk The acceptance of customers, vendors and business partners the application become obsolete due to the introduction of Table 5 The impact values of the EA frameworks. Criteria FEAF ZACHMAN Organizational risk (7.6, 8.4, 0.1, 0.1) (8.6, 9.4, 0.1, 0.1) User risk (9.6, 10.4, 0.1, 0.1) (8.6, 9.4, 0.1, 0.1) Requirement risk (8.6, 9.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) Structural risk (8.6, 9.4, 0.1, 0.1) (8.6, 9.4, 0.1, 0.1) Team risk (6.6, 7.4, 0.1, 0.1) (6.6, 7.4, 0.1, 0.1) Complexity risk (8.6, 9.4, 0.1, 0.1) (6.6, 7.4, 0.1, 0.1) Competition risk (4.6, 5.4, 0.1, 0.1) (5.6, 6.4, 0.1, 0.1) Market risk (3.6, 4.4, 0.1, 0.1) (4.6, 5.4, 0.1, 0.1) 5. Conclusions and future research directions In this paper, we proposed a novel fuzzy group multi-criteria model for EA framework evaluation and selection. The contribution of the proposed model is fourfold: (1) it takes into consideration the qualitative and quantitative criteria and their respective value judgments; (2) it considers verbal expressions and linguistic vari- ables for qualitative judgments which lead to ambiguity in the decision making process; (3) it handles imprecise or vague judg- ments; and (4) it uses a meaningful and robust multi-criteria mod- el to aggregate both qualitative and quantitative data. We applied the model to select the best EA framework in a com- plex system with multiple and competing criteria and values. Organizations often fail in practice to follow a systematic and well-structured decision-making process for assessing potential EA frameworks. We have shown that our model considers the mul- ti-dimensional nature of such problems and generates vital development of the Enterprise Architecture lopment; unfavorable attitudes of users towards a new enterprise architecture quate, or ambiguous requirements f department are to be involved; the business process needs to be changed a lot m members exity of the processes being automated; whether a large number from obtaining the expected outcome of the application; unanticipated change in the industry or market; new enterprise architecture TOGAF TEAF DoDAF (9.6, 10.4, 0.1, 0.1) (6.6, 7.4, 0.1, 0.1) (9.6, 10.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (8.6, 9.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (6.6, 7.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (8.6, 9.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (7.6, 8.4, 0.1, 0.1) (6.6, 7.4, 0.1, 0.1) (6.6, 7.4, 0.1, 0.1) (4.6, 5.4, 0.1, 0.1) (6.6, 7.4, 0.1, 0.1) (2.6, 3.4, 0.1, 0.1) (2.6, 3.4, 0.1, 0.1) (4.6, 5.4, 0.1, 0.1) Table 8 The possibilistic mean values of the fuzzy risk priority numbers (RPN). Criteria FEAF ZACHMAN TOGAF TEAF DoDAF Organizational risk 144 90 150 42 280 User risk 450 324 320 224 144 Requirement risk 108 80 32 40 126 Structural risk 486 576 320 224 567 Team Risk 196 245 320 192 320 Complexity risk 189 98 144 80 140 Competition risk 270 240 280 200 378 Market risk 192 175 105 105 315 Table 9 The ordinal rank matrix. Criteria FEAF ZACHMAN TOGAF TEAF DoDAF Organizational risk 2 1 3 0 4 User risk 4 3 2 1 0 Requirement risk 3 2 0 1 4 Structural risk 2 4 1 0 3 Team risk 1 2 3.5 0 3.5 Complexity risk 4 1 3 0 2 Competition risk 2 1 3 0 4 Market risk 3 2 0.5 0.5 4 Table 10 The vector of the EA framework risks. The risks vector FEAF ZACHMAN TOGAF TEAF DoDAF V 0.29 0.21 0.16 0.06 0.28 Table 6 The likelihood values of the EA frameworks. Criteria FEAF ZACHMAN TOGAF TEAF DoDAF Organizational risk (5.75, 6.25, 0.25, 0.25) (4.75, 5.25, 0.25, 0.25) (4.75, 5.25, 0.25, 0.25) (2.75, 3.25, 0.25, 0.25) (6.75, 7.25, 0.25, 0.25) User risk (8.75, 9.25, 0.25, 0.25) (8.75, 9.25, 0.25, 0.25) (7.75, 8.25, 0.25, 0.25) (6.75, 7.25, 0.25, 0.25) (7.75, 8.25, 0.25, 0.25) Requirement risk (5.75, 6.25, 0.25, 0.25) (4.75, 5.25, 0.25, 0.25) (3.75, 4.25, 0.25, 0.25) (4.75, 5.25, 0.25, 0.25) (5.75, 6.25, 0.25, 0.25) Structural risk (8.75, 9.25, 0.25, 0.25) (7.75, 8.25, 0.25, 0.25) (7.75, 8.25, 0.25, 0.25) (6.75, 7.25, 0.25, 0.25) (8.75, 9.25, 0.25, 0.25) Team risk (3.75, 4.25, 0.25, 0.25) (4.75, 5.25, 0.25, 0.25) (4.75, 5.25, 0.25, 0.25) (3.75, 4.25, 0.25, 0.25) (4.75, 5.25, 0.25, 0.25) Complexity risk (2.75, 3.25, 0.25, 0.25) (1.75, 2.25, 0.25, 0.25) (2.75, 3.25, 0.25, 0.25) (1.75, 2.25, 0.25, 0.25) (3.75, 4.25, 0.25, 0.25) Competition risk (8.75, 9.25, 0.25, 0.25) (7.75, 8.25, 0.25, 0.25) (7.75, 8.25, 0.25, 0.25) (7.75, 8.25, 0.25, 0.25) (8.75, 9.25, 0.25, 0.25) Market risk (7.75, 8.25, 0.25, 0.25) (6.75, 7.25, 0.25, 0.25) (6.75, 7.25, 0.25, 0.25) (6.75, 7.25, 0.25, 0.25) (8.75, 9.25, 0.25, 0.25) Table 7 The detection values of the EA frameworks. Criteria FEAF ZACHMAN TOGAF TEAF DoDAF Organizational risk (2.9, 3.1, 0.1, 0.1) (1.9, 2.1, 0.1, 0.1) (1.9, 2.1, 0.1, 0.1) (1.9, 2.1, 0.1, 0.1) (3.9, 4.1, 0.1, 0.1) User risk (4.9, 5.1, 0.1, 0.1) (3.9, 4.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) (3.9, 4.1, 0.1, 0.1) (1.9, 2.1, 0.1, 0.1) Requirement risk (1.9, 2.1, 0.1, 0.1) (1.9, 2.1, 0.1, 0.1) (0.9, 1.1, 0.1, 0.1) (0.9, 1.1, 0.1, 0.1) (2.9, 3.1, 0.1, 0.1) Structural risk (5.9, 6.1, 0.1, 0.1) (7.9, 8.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) (3.9, 4.1, 0.1, 0.1) (6.9, 7.1, 0.1, 0.1) Team risk (6.9, 7.1, 0.1, 0.1) (6.9, 7.1, 0.1, 0.1) (7.9, 8.1, 0.1, 0.1) (5.9, 6.1, 0.1, 0.1) (7.9, 8.1, 0.1, 0.1) Complexity risk (6.9, 7.1, 0.1, 0.1) (6.9, 7.1, 0.1, 0.1) (5.9, 6.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) Competition risk (5.9, 6.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) (5.9, 6.1, 0.1, 0.1) Market risk (5.9, 6.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) (4.9, 5.1, 0.1, 0.1) (6.9, 7.1, 0.1, 0.1) 1172 F. Zandi, M. Tavana / Expert Systems with Applications 39 (2012) 1165–1173 information for selecting the most appropriate framework. While previous studies have valued multi-criteria frameworks, they have failed to consider both subjective and objective judgments in a sys- tematic and consistent model. The proposed framework supports both qualitative and quantitative decision criteria, as well as differ- ent types of risk for both quantitative and qualitative criteria. In addition, the DMs may be able to provide only imprecise or va- gue information because of time constraints or lack of data. Further- more, the DM may feel more comfortable evaluating qualitative criteria by using linguistic variables resulting in two potential prob- lems: (1) how to reconcile quantitative and qualitative criteria and (2) how to deal with imprecise and vague information rationally and consistently. We showed that the proposed framework is able to address these problems and can assist the DMs reach a robust decision. Finally, the case study showed that a combined analysis can generate valuable insight that can help the DMs to select the most suitable framework from a range of competing alternatives. Our model is intended to assist the DMs in the EA framework selection process. In fact, human judgment is the core input in this process. Our approach helps the DMs to think systematically about complex multi-criteria problems and improves the quality of the decisions. We decompose the EA framework selection process into manageable steps and integrate the results to arrive at a solution consistent with managerial goals and objectives. This decomposi- tion encourages the DMs to carefully consider the elements of uncertainty. The proposed structured framework does not imply a deterministic approach in EA framework selection problems. While our approach enables the DMs to assimilate the information and organize their beliefs in a formal systematic approach, it should be used in conjunction with management experience. Man- agerial judgment is an integral component of EA framework selec- tion decisions; therefore, the effectiveness of the model relies heavily on the DM’s cognitive capabilities. There are a variety of extensions to this research. First, EA is a new discipline and it will not mature without substantial new re- search (Langenberg & Wegmann, 2004). We hope the framework proposed in this study will stimulate new research in the fields of EA and multi-criteria decision making. Second, by identifying the unique features of the proposed framework, researchers can further study the applicability of the proposed method to other multi-criteria problems embracing both qualitative and quantita- tive criteria with imprecise or ambiguous value judgments. It is our hope that the discussions, issues and ideas set forth in this paper will motivate further enhancement to this framework. References Andre’eva, Y., Larichev, O., Flanders, N., & Brown, R. (1995). Complexity and uncertainty in Arctic resource decision: The example of the Yamal pipeline. Polar Geography and Geology, 19, 22–35. Berkeley, D., Humphreys, P., Larichev, O., & Moshkovich, H. (1991). Aiding strategic decision making: Derivation and development of ASTRIDA. In Y. Vecsenyi & H. Sol (Eds.), Environment for supporting decision processes. Amsterdam: North- Holland. F. Zandi, M. Tavana / Expert Systems with Applications 39 (2012) 1165–1173 1173 Belton, V., & Stewart, T. J. (2002). Multiple criteria decision analysis: An integrated approach. Norwell, MA: Kluwer. Brans, J. P., Vincke, Ph., & Mareschal, B. (1986). How to select and how to rank projects: The PROMETHEE method. European Journal of Operational Research, 24, 228–238. Carbone, T. A., & Tippett, D. D. (2004). Project risk management using the project risk FMEA. Engineering Management Journal, 16(4), 28–35. Chou, S., Chang, Y., & Shen, C. (2008). A fuzzy simple additive weighting system under group decision-making for facility location selection with objective/ subjective attributes. European Journal of Operational Research, 189(1), 132–145. Costa, A. P. C. S., Almeida, A. T., & Miranda, C. M. G. (2003). Multicriteria support to sort information systems portfolio. Journal of the Academy of Business and Economics, 2(1), 237–247. Diakoulaki, D., & Koumoutsos, N. (1991). Cardinal ranking of alternatives actions: Extension of PROMETHEE method. European Journal of Operational Research, 53, 337–347. Fayad, M., & Hamu, D. (2000). Enterprise frameworks: Guidelines for selection. ACM Computing Survey, 32(1), 1–23. Gammelgêard, M., Simonsson, M., & Lindstrèom, E. (2007). An IT management assessment framework: Evaluating enterprise architecture scenarios. Information Systems and E-Business Management, 5(4), 415–435. Ghodsypour, S. H., & O’brien, C. (1998). A decision support system for supplier selection using an integrated analytical hierarchy process and linear programming. International Journal of Production Economics, 56, 199–212. Flanders, N. E., Brown, R. V., Andre’eva, Y., & Larichev, O. (1998). Justifying public decisions in Arctic oil and gas development: American and Russian approaches. Arctic, 51(3), 262–279. Hwang, C. L., & Yoon, K. (1981). Multiple attribute decision making: Methods and applications. Berlin, Heidelberg: Springer. Kahraman, C., Ates�, N. Y., Çevik, S., Gülbay, M., & Erdoğan, S. A. (2007). Hierarchical fuzzy TOPSIS model for selection among logistics information technologies. Journal of Enterprise Information Management, 20(2), 143–168. Kaklauskas, A., Zavadskas, E. K., Raslanas, S., Ginevicius, R., Komka, A., & Malinauskas, P. (2006). Selection of low-e windows in retrofit of public buildings by applying multiple criteria method COPRAS: A Lithuanian case. Energy and Buildings, 38(5), 454–462. Kim, S. H., & Ahn, B. S. (1999). Interactive group decision making procedure under incomplete information. European Journal of Operational Research, 116(3), 498–507. Langenberg, K., Wegmann, A. (2004). Enterprise architecture: What aspects is current research targeting. Technical Reports in Computer and Communication Sciences, no. 2004-77, École Polytechnique Fédérale de Lausanne, Faculté I& C, School of Computer and Communication Sciences, Lausanne, Switzerland. Larichev, O. (1992). Cognitive validity in design of decision-aiding techniques. Journal of Multi-Criteria Decision Analysis, 1(3), 127–138. Larichev, O., Brown, R., & Andre’eva, E. (1995). Categorical decision analysis for environmental management: A Siberian gas distributing case. In J.-P. Caverni, M. Bar-Hillel, F. H. Barron, & H. Jungermann (Eds.), Contribution to decision making (pp. 55–286). Amsterdam: North-Holland. Larichev, O., & Moshkovich, H. (1997). Verbal decision analysis for unstructured problems. Boston: Kluwer Academic Publishers. Poyhonen, M. A., Hamalainen, R. P., & Salo, A. A. (1997). An experiment on the numerical modelling of verbal ratio statements. Journal of Multi-Criteria Decision Analysis, 6(1), 1–10. Project Management Institute. (2009). A Guide to the Project Management Body of Knowledge (PMBOK� Guide) – 3rd Ed. Project Management Institute. Roztocki, N., & Weistroffer, H. R. (2006). Evaluating information technology investments: A fuzzy activity-based costing approach. Journal of Information Science and Technology, 2(4), 30–43. Roy, B. (1996). Multicriteria methodology for decision aiding. Dordrecht: Kluwer Academic. Saaty, T. L. (1994). How to make a decision: The analytic hierarchy process. Interfaces, 24(6), 19–43. Shih, H. S. (2008). Incremental analysis for MCDM with an application to group TOPSIS. European Journal of Operational Research, 186(2), 720–734. Tang, A., Han, J., Chen, P. (2004). A comparative analysis of architecture frameworks, Technical Report, Swinburne University of Technology. Urbaczewski, L., & Mrdalj, S. (2006). A Comparison of Enterprise Architecture Frameworks. Issues in Information Systems, 7(2), 18–23. Valls, A., & Torra, V. (2000). Using classification as an aggregation tool in MCDM. Fuzzy Sets and Systems, 15(1), 159–168. Wang, J. J., & Yang, D. L. (2007). Using a hybrid multi-criteria decision aid method for information systems outsourcing. Computers and Operations Research, 34(12), 3691–3700. Xia, H. C., Li, D. F., Zhou, J. Y., & Wang, J. M. (2006). Fuzzy LINMAP method for multiattribute decision making under fuzzy environments. Journal of Computer and System Sciences, 72(4), 741–759. Yang, J. B., & Xu, D. L. (2002). On the evidential reasoning algorithm for multiattribute decision analysis under uncertainty. IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, 32(3), 289–304. Zavadskas, E. K., Turskis, Z., Dejus, T., & Viteikiene, M. (2007). Sensitivity analysis of a simple additive weight method. International Journal of Management and Decision Making, 8(5-6), 555–574. Zavadskas, E. K., Turskis, Z., Tamošaitienė, J., & Marina, V. (2008). Multicriteria selection of project managers by applying grey criteria. Technological and Economic Development of Economy, 14(4), 462–477. Zimmermann, H. J. (2000). An application-oriented view of modelling uncertainty. European Journal of Operational Research, 122(2), 190–198. A fuzzy group multi-criteria enterprise architecture framework selection model 1 Introduction 2 The mathematical notations and definitions 3 The proposed model 4 The case study 5 Conclusions and future research directions References 
work_qvv26xmtsbdulp6dyp5qod2yne	----	A new approach for rule extraction of expert system based on SVM A new approach for rule extraction of expert system based on SVM Ai Li, Guo Chen ⇑ College of Civil Aviation, Nanjing University of Aeronautics & Astronautics, Nanjing 210016, PR China a r t i c l e i n f o Article history: Received 18 October 2012 Received in revised form 9 August 2013 Accepted 19 August 2013 Available online 14 September 2013 Keywords: Support Vector Clustering Support Vector Machine Rule extraction Knowledge acquisition Expert system Genetic Algorithm Feature selection a b s t r a c t Based on the SVM’s excellent generalization performance, a new approach is proposed to extract knowledge rules from Support Vector Clustering (SVC). In this method, the first step is to choose the features of the sample data by using Genetic Algorithm for improving the comprehensibility of the knowledge rules. Then the SVC algorithm is adopted to obtain the Clustering Distribution Matrix of the sample data whose features have been chosen. Finally, hyper-rectangle rules are constructed using the Clustering Distribution Matrix. To make the rules more concise, and easier to explain, hyper-rectangle rules are simplified further by using rules combinations, dimension reduction and interval extension. In addi- tion, the SMOTE (Synthetic Minority Over-sampling Technique) algorithm is adopted to resample fault samples in order to solve the serious imbalance problem of samples. The UCI datasets are used to validate the new method proposed in this paper, the results com- pared with other rules extraction methods show that the new approach is more effective. The new method is used to extract knowledge rules for aero-engine oil monitoring expert system, and the results show that the new method can effectively extract knowledge rules for expert system, and break through the bottleneck in expert system knowledge dynamic acquisition. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction At present, knowledge acquisition through data mining [1,2] occurs mainly through machine learning or statistics. Correlation analyses [3], artificial neural networks [4], rough sets [5], and decision trees [6] are extensively em- ployed for data mining. If data mining is applied to an ex- pert system and if the knowledge rules are extracted automatically from real data, then the intelligence level and knowledge acquisition ability of the expert system will be greatly improved. In recent years, the Support Vector Machine (SVM) [7] has become an emerging classification technology in data mining. The SVM can approximate any continuous bounded nonlinear function because of the perfect general- ization theory and strong nonlinear mapping ability. The SVM has several advantages over the neural network, such as better generalization ability, no local minimum prob- lem, the ability to automatically construct the learning ma- chine, no dimension curse, and the ability to deal with small samples. These advantages have caused data mining technology based on SVM to receive the attention of researchers worldwide. Furthermore, a number of promis- ing SVM rule extraction algorithms published to date [8– 14] are not only simple but also broadly applicable. Nunez et al. [9] introduced a rule extraction approach based on the SVM, in which K-means clustering is used to obtain clustering centers, which are then combined with support vectors (SVs) to define ellipsoid rules. Finally, the ‘‘if-then’’ rules can be obtained when the ellipsoid rules are mapped to the input space. However, the generated ellipsoid rules seriously overlap. In addition, the solution quality of K- means strongly depends on the initial values for the cen- ters, and it is difficult to control the quantity and quality 0263-2241/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.measurement.2013.08.028 ⇑ Corresponding author. Tel./fax: +86 025 84891850. E-mail address: cgzyx@263.net (G. Chen). Measurement 47 (2014) 715–723 Contents lists available at ScienceDirect Measurement j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / m e a s u r e m e n t http://crossmark.crossref.org/dialog/?doi=10.1016/j.measurement.2013.08.028&domain=pdf http://dx.doi.org/10.1016/j.measurement.2013.08.028 mailto:cgzyx@263.net http://dx.doi.org/10.1016/j.measurement.2013.08.028 http://www.sciencedirect.com/science/journal/02632241 http://www.elsevier.com/locate/measurement of the obtained rules. In a similar study, Zhang et al. [10] introduced the hyper-rectangle rule extraction (HRE) algo- rithm to extract rules from the trained SVM. The authors used the Support Vector Clustering (SVC) algorithm to find prototype vectors for each class, and then used those vec- tors with the SVs to generate hyper-rectangles. A nested generalized exemplar algorithm is utilized to first con- struct small hyper-rectangles around the prototypes, which are then grown incrementally until the stopping cri- teria based on a user-defined minimum confidence thresh- old (MCT) or minimum support threshold (MST) are met. If-then rules are then generated by projecting these hy- per-rectangles onto coordinate axes. The published results for this method show that the rules provide good accuracy. However, all the features are present as antecedents of these rules. This limits their explanation capability, since no indication is given about the most important features for the classification. Based on the aforementioned limitations, here, a new method is proposed to extract knowledge rules from SVC. The first step in this method is to choose the features of the sample dataset using a Genetic Algorithm (GA) for improving the comprehensibility of the knowledge rules. The next step is to map the chosen features of the training samples into a high-dimensional feature space to get opti- mal separating hyper-planes and SVs. Finally, the hyper- rectangles are constructed using the Clustering Distribu- tion Matrix of the data obtained by the SVC, and the if-then rules are generated by projecting these hyper-rectangles onto coordinate axes. In order to make the rules more con- cise and easier to explain, hyper-rectangle rules are further simplified using a combination of rules, dimension reduc- tion, and interval extension. In addition, the SMOTE (Syn- thetic Minority Over-sampling Technique) algorithm is adopted to resample fault samples in order to solve the serious imbalance problem of samples. Experimental re- sults show that it is easy to control the number and the support degree of the generated rules; feature selection and simplification of rules can greatly improve their expla- nation capability. Spectral oil diagnosis expert system is the advanced stage of aero-engine wear fault diagnosis. At present, some oil monitoring expert systems have been developed, such as, the advanced rapid analysis system PFALink developed by the United States Mobil oil company, lubricating oil analysis expert system Lube Analyst and Atlas developed by the United States and Canada. But these software only provides a framework and management system, and the users need to develop its core knowledge base and provide the monitored wear element threshold value. In the intel- ligent diagnosis expert system, these problems, such as weak knowledge acquisition, hard knowledge updating and poor knowledge adaptability, still did not get effective to be overcome. The expert system knowledge acquisition is basically by means of the mechanical learning methods based on the experiences. The knowledge is hard to update and the rules exist serious problems such as inconsistent, redundancy, and combination explosion. Therefore, in this paper, the new method is applied to the knowledge acqui- sition of aero-engine spectral oil diagnosis expert system. Experimental results to real dataset show the effectiveness and the correctness of the new method. 2. Knowledge rules extracting method based on GA_SVC The rule extraction process includes data preprocessing, SVC, hyper-rectangle rule extraction and rule simplifica- tion. The entire rule extraction procedure is shown in Fig. 1. 2.1. Data preprocessing 2.1.1. Dalancing to unbalance data In data mining experiments, the datasets are usually as- sumed to balance distribution, which is the number of var- ious types of samples is almost the same, while it is almost non-existent in the real. In many real datasets, the number of class with different label is unequal. These datasets are called unbalanced datasets. Usually, the minority class samples will be taken out as noise so that no rules about the minority class can be extracted. Therefore, in order to extract rules of various types of samples completely, and improve the recognition rate of the rules, the first step is to preprocess the unbalance data into balance data before rules extraction. In this paper, we resample fault samples by using Syn- thetic Minority Over-sampling Technique (SMOTE) which is the typical sampling algorithm. SMOTE [15] algorithm is an over-sampling method put forward by Chawla. In or- der to make the dataset be equilibrium, the main concept of the method is to use k neighbor method and linear inter- polation method to insert new samples according to cer- tain rules between the two closer samples of minority class. In Fig. 2, a two dimensional example {X = (x1, x2)} is enlarged by using SMOTE over-sampling method. It can be seen from Fig. 2 that the new re-sampling samples focus Fig. 1. Rules extraction procedure. 716 A. Li, G. Chen / Measurement 47 (2014) 715–723 https://isiarticles.com/article/52631 
work_qy6bxkk4tje53le55f5bggkglm	----	An expert system manipulating knowledge to help human learners into virtual environment Expert Systems with Applications 38 (2011) 8446–8457 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a An expert system manipulating knowledge to help human learners into virtual environment Cédric Buche ⇑, Ronan Querrec 1 Université Européenne de Bretagne, Ecole Nationale d’Ingénieurs de Brest, Laboratoire Informatique des Systèmes Complexes, Technopôle Brest-Iroise, 29238 Brest Cedex 3, France a r t i c l e i n f o Keywords: Knowledge representation Multi-agent system Intelligent tutoring system Classifiers system Virtual learning environment 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.01.040 ⇑ Corresponding author. Tel.: +33 298058966; fax: E-mail addresses: buche@enib.fr (C. Buche), querre 1 Tel.: +33 298058966; fax: +33 298058979. 2 PEGASE: PEdagocial Generic and Adaptable SystEm. a b s t r a c t This research is situated within the context of the creation of human learning environments using virtual reality. We propose the integration of a generic and adaptable intelligent tutoring system (PEGASE). The aim is to instruct the learner, and to assist the instructor. The multi-agent system emits a set of knowl- edge (actions carried out by the learner, knowledge of the field, etc.) used by an artificial intelligence to make pedagogical decisions. Our study focuses on the representation of knowledge about the environ- ment, and on the adaptable pedagogical agent providing instructive assistance. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Many fields of learning, such as driving, or the professional training undergone by firemen, requires the learners to experience the setting in which they will work. These learners must therefore acquire not only knowledge, but real hands-on skills. Virtual envi- ronments (VE) immerse the learners in situations such as these. For example, Fig. 1 represents a road safety application (ARÉVIROAD) (Herviou & Maisel, 2006), a SEVESO plant application (Edward, Lourdeaux, Lenne, Barthes, & Burkhardt, 2008) and GASPAR for logistics on aircraft carriers (Marion, Septseault, Boudinot, & Querrec, 2007). The work is designed for teaching decision-making in VE. Tutor- ing system to instruct the learner and to assist the instructor al- ready exists (Lourdeaux, Burkhardt, Bernard, & Fuchs, 2002; Rickel & Johnson, 1999), but are dedicated to a specific VE. In this paper, we propose an independent VE tutoring system, called PE- GASE,2 in the domain of procedural and collaborative work. 2. Context: acquisition of skills using virtual environments Traditionally, most training programmes aim to teach knowl- edge. In this context, we propose the use of Intelligent Tutoring Systems (ITS) in which this knowledge is used in conjunction with the training setting. In this case, knowledge can be manipulated (Romero & Ventura, 2007), to automatically question the learner, for example. To be competent does not only mean acquiring knowledge, but also how to use that knowledge. In order to facilitate ll rights reserved. +33 298058979. c@enib.fr (R. Querrec). the acquisition of knowledge, we must provide the learner with the right setting. In order to do so, we suggest using interactive sys- tems by which the learners can be immersed in VEs in which they can make trial attempts, take initiatives, make mistakes, and try again in similar situation (which may not be possible in reality). The simulation therefore provides an environment common to the learner, the instructor, and to the skill to be acquired. It medi- ates the learning relationship (learner-skill) as well as the instruc- tive relationship (instructor-learner). Thus, computer-generated simulations, combined with an ITS, create the opportunity to improve learners’ skills by associating knowledge with the oppor- tunity to put their skills into practice. ITS have already been used without being associated to virtual reality. As Wenger (1987) has shown, they usually conform to one of four models. The first, known as the domain model, contains a representation of the knowledge linked to the skill to be acquired. ITS also use a learner model which defines his/her per- sonal characteristics and ascertains the condition of the knowledge at a given moment. Using the domain and learner models, the ITS is able to evaluate the knowledge acquired by the learner by compar- ing his/her activity with information about the field (Keles, Ocak, Keles, & Gülcü, 2009). However, the main objective of the ITS is to the provide appropriate assistance to the learner (Wang, 1997) or the instructor, depending on the setting (following the activities or offering assistance). In this context, the pedagogical model can be used to make choices with regards to the training objective, with the aim of facilitating learning. Finally, an interface model is used to exchange information between the system and the user. For now, this model has not been reified in existing VEs designed for learning. Within the context of our VE, we consider an ITS as a system which is part of the human Virtual Learning Environment (VLE). We propose to evaluate the extent to which ITS are integrated http://dx.doi.org/10.1016/j.eswa.2011.01.040 mailto:buche@enib.fr mailto:querrec@enib.fr http://dx.doi.org/10.1016/j.eswa.2011.01.040 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. From left to right: screenshots from the ARÉVIROAD, VIRTHUALIS and GASPAR applications. C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 8447 within existing VLE. We have grouped VLE according to three categories: 1. VLE as classical simulators This first category groups together those applications which include none of the four models, such as the application designed to assist both in the maintenance and control of mobile cranes (Levesque, 2003). In this kind of VE, the system provides no explanations about the task to be performed, which would require a domain model. The environment is therefore unable to adapt to the learner, as this would require a learner model. Finally, the teaching method is the instructor’s responsi- bility. The system is not able to make decisions regarding instructive interventions, however, a system like this one can help learners to improve or modify pre-existing skills. 2. VLE with domain and/or learner models This second category of VE is made up of applications which include a domain model and/or a learner model (Hubal, 2008). The most well-known example of this type of VE is STEVE, a virtual character who assists in both teaching and learning procedural tasks (Rickel & Johnson, 1999). Using the domain model, STEVE can demonstrate and explain the procedure and above all, verify the learner’s actions. However, STEVE intervenes on demand. He is incapable of knowing when, how and why to intervene, which would require a pedagogical model. In a sys- tem such as this it is possible to acquire skills, but the participa- tion of the instructor is still required for all pedagogical interventions. 3. VLE with domain, learner, and pedagogical models This final category groups together the VEs presenting not only domain and learner models, but also a pedagogical model (Amokrane, Lourdeaux, & Burkhardt, 2008). Let us examine the example of the educational agent HAL, from the FIACRE sys- tem (Lourdeaux et al., 2002). The application is designed to instruct individuals in learning to drive TGV trains (Train Grande Vitesse), using virtual reality (intervention on railways). As well as having all of STEVE’s abilities, HAL assists the instruc- tors in structuring pedagogical discourse. In concrete terms, each anticipated behavior corresponds to a different instructive assistance (additional information, explanation of an object, etc.). The instructor must therefore list the possible errors for each piece of knowledge to be acquired. Furthermore, for each of these errors, the instructor must specify the way in which these pedagogical strategies should be conducted through instructive assistance, and indeed he must do so for each exercise. The main advantage of this kind of VLE lies in the assistance to the instructor in terms of the educational relation- ship linked to the learner, and in the didactic relationship linked to the skill to be learnt. However, the instructor must specify all of the knowledge to be acquired for each exercise. Thus, most VLE only include the representation of knowledge about one specific domain. Systems proposing a diagnostic compo- nent only provide a mechanism for instructive assistance. HAL seems to us to be the most successful of these systems. However, the instructor must still make a list of the possible errors and spec- ify the pedagogical strategies for each exercise. Furthermore, the impact of the instructive assistance on the learner is not taken into consideration. In concrete terms, any proposed assistance which does not help the learner to make progress will be reproposed each time that specific situation occurs. In order to resolve these shortcomings, we propose the integra- tion of an intelligent tutoring system within a VE. This system must propose a flexible pedagogical model, i.e. a model in which instruc- tive concepts can be easily added, modified or deleted. Further- more, a model such as this must be generic, in so much as the pedagogical model must be exploitable independently of the task to be performed. Finally, the knowledge of the pedagogical model, along with its past experience, could be used to automatically sug- gest the appropriate interventions by taking into account both the learner and the context of the simulation: the system therefore be- comes adaptive. Our model is called PEGASE (PEdagogical Generic and Adaptive SystEm). In the next section, we will describe the global architecture of PEGASE. We will then go onto present our domain model (see Section 4). This will then lead to our description of our pedagogical model (see Section 5). We shall then go onto discuss the advantages of our proposed models (see Section 6). It must be noted that the proposal described here is applicable within the context of the learning of procedural and collaborative tasks and cannot be used in general learning situations. 3. Proposing an intelligent tutoring system: PEGASE Our proposal consists of reifying the four classical ITS models (domain, learner, pedagogical, interface), within a VE. We believe that errors can provide crucial information and thus decided to introduce a model called ‘‘error model’’. It is through the use of this new model that we will be able to generalize (where HAL could not do so). Furthermore, we have also added an ‘‘instructor model’’, in which the instructor specifies the knowledge about the exercise to be performed. He defines the guidelines which describe the procedure(s) to be carried out, and the role(s) played by the learner (and consequently those which must also be acti- vated automatically). These models must provide solutions to the shortcomings of the existing systems as detailed above and must therefore display two important characteristics: genericity and adaptability. We there- fore suggest that, from a VE, it is possible to transplant a generic and adaptive ITS by reifying the 6 ITS models. In order for each model to be able to share its information and conduct its analyzes autonomously (independently of both the sit- uation and of other models), an autonomous entity (known as an agent) is associated to each model. The agents interact by exchang- ing messages containing data (see Fig. 2). This data can be ex- tracted from the situation or inferred from the agent’s internal reasoning from its knowledge (the model to which it is linked). Fig. 2. The instructive process of our five-stage system. Table 1 Layers of modeling (Mi): the positioning of Veha within the Mof guidelines, in parallel to Uml. M4 Mof (Uml limitation) M3 Meta–model Uml Veha metamodel M2 Uml user model VE1 model . . . M1 User object VE1a VE1b . . . . . . 8448 C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 Step 1. Observation: Using the interface model, the system analyzes the learner’s activity. The elements that are important for learning are sup- plied to the learner model. This information concerns the lear- ner’s actions, those elements which the learner can observe, and the learner’s movements. Step 2. Detecting and identifying an error: The system analyzes the learner’s actions (learner model) and compares them to the actions to be performed (domain model). This confrontation is used in order to detect errors. If an error has been detected, an error identification mechanism is set up (using the error model). Step 3. Proposing instructive assistance: Using the learner model (characteristics, activities, errors, etc.), and the domain model (knowledge of the organisational struc- tures), a mechanism stimulating instructive reasoning recom- mends the instructive assistance for the given situation. It must be noted that this step is not optional; it intervenes even if no error is detected. Step 4. Choosing instructive assistance: The instructor can choose one specific instructive assistance amongst those proposed. Step 5. Representing instructive assistance: The instructive assistance selected is presented in the VE. To use the information from the VE, we must inform the envi- ronment in order to obtain controllable knowledge: we thus achieve an informed VE (see Section 4). We must therefore reify the environment. This knowledge is complemented by additional information contained within the 6 ITS models. This data makes up a knowledge base for the pedagogical model which we call the pedagocical situation. This knowledge fuels the ITS’s motor for making instructive decisions (see Section 5). An example of the way in which the rules governing this motor are specified is pre- sented in Section 5.3. 4. Domain model For reification, we define the VEHA (VE for Human Activity) metamodel. The aim of Veha is to provide a metamodel to describe the VE, not in terms of geometric space, but by providing the semantics required for the artificial agents (ITS, autonomous char- acters) or humans (learners or instructors) to be able to construct for themselves a representation of the environment and to act to- gether to reach their goals. The Veha metamodel is based on Uml and enables the construction of domain patterns from VEs and from the corresponding concrete VEs (see Table 1). However, the metamodel does not allow us to define the specific concepts of vir- tual reality. In Veha, we propose to extend Uml in order to repre- sent these concepts. The aim of this extension is to reify the realistic VE in which humans carry out activities. However, it is not adapted to all VE (metaphoric, phenomenological, etc.). 4.1. The Veha metamodel The ITS needs to know which objects make up the VE, how to access it, its properties, its behavior and how to interact with it. Three kinds of knowledge can be expressed using Veha: C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 8449 1. Domain concepts: This entails the semantic description of the concepts relating to the field of activity concerned. This repre- sents some of the knowledge that the learner must acquire (Section 4.1.1). 2. The possibility of structuring and interacting with the environment: These concepts resemble those suggested in smart objects (Kallmann & Thalmann, 1998) which reify those properties required for interactions. The means available to the learner or to the ITS must be specified in order to modify the environ- ment (Section 4.1.2). 3. Entities’ behavior: Within the framework of a VLE the environ- ment’s reactions to the learner’s actions must be simulated. Entities’ behavior also represents one of the elements of the knowledge to be transmitted and must be enforceable (Section 4.1.3). In the following part of this section, we shall explain how Veha can be used to express these three kinds of knowledge. 4.1.1. Domain concepts Knowledge of the domain is expressed both at the model (con- cept) level, and at the level of the occurrences of these concepts (tangible objects populating the environment). In Veha as in Uml, this knowledge is represented by classes (Class) and instances (InstanceSpecification). In Veha, the notion of class is used to define a type of object Fig. 8 from domain-specific ontology. The aim is to be able to apply semantics to each of the domain concepts, whether or not they are tangibly represented in the VE (concepts vs concrete objects). All classes stem from the Element class. This class enables the identi- fication of each of the elements of a domain model from its name and the addition of a textual comment. This can be useful when providing the user with explanations regarding the significance of an object. The structural properties (Property) and behavioral features (BehavioralFeature) of the classes are assigned to the Classi- fier 3 via the Feature class. The Property class represents the structural component of the Classifier (as much the attributes as the relationships with other domain concepts). As in Uml, the Operation class is the only tangible sub-class of Behavioral- Feature. It is used to express the effect that an object or a user can have on another object. It does this by defining the object’s ac- tual behavior rather than the method used to achieve that behavior. The way in which the behaviors associated with Operation are modeled is described using behavioral models (see Section 4.1.3). After the notion of Class, Instance, synonymous of object, is Veha’s second key concept. The InstanceSpecification, Slot and AssociationInstance classes represent the instantiation of de Class, Property and Association respectively. The term InstanceSpecification indicates that here, we represent an M1 level entity Table 1, independently of the circumstances in which it is implemented. The set of knowledge about the environment as specified in Veha can be accessed by the ITS and by the users (learners or instructors). The ITS can, for example, suggest to the learner a list of operations to be performed on one specific kind of object. Likewise, the instructor can modify the environment during the simulation by changing the attribute values of a tangible object. 4.1.2. The possibility of structuring and interacting with the environment Within the context of VEs, most of the tangible objects within those environments are represented geometrically, and are situ- 3 UML metamodel class which generalizes the concept of class. ated within the environment. The learner must be able to observe, recognize and manipulate these objects. The ITS also needs to be able to manipulate them within the context of the instructive assistance that it will implement (transparency, refocusing from the learner’s point of view, etc.). Knowledge about the geometry of these objects must also be specified so that the ITS will be able to recontextualize its suggestions within the VE. These objects are entities and all have the properties of the instances veha::Class as well as geometric and topological properties Fig. 7. Each entity is located at a global reference point. The Shape class is used to assign an instance of EntityClass to a graphical representation in the VE. It is possible to assign many forms to one class of entity. The ITS can use this knowledge to highlight an ob- ject or, on the contrary, to hide it. The TopologicalProperty class supports the notion of location (position and orientation) and is used to describe the topological properties of the elements within the VE Fig. 7. It is possible to assign informed points to an entity (Point) which can be used to create an interaction. This information is used by the ITS to turn the learner’s attention to a specific object, for example. Any entity within the VE is an instance of EntityClass: the Entity class, which derives from Kernel::InstanceSpecifi- cation. The values of an entity’s properties are defined by its slots. These therefore depend on the semantic, morphological, geometric and topological properties of the objects within the VE (supplied by InstanceSpecification). 4.1.3. Entities’ behaviors When the learner carries out an action in the environment, that environment must react in a realistic way for the learner to be able to understand the consequences of his actions. The learner there- fore constructs a representation of the entities’ behavior. For the ITS to be able to regulate this representation, the knowledge of entities’ behaviors must also be specified, as for the two previous kinds of knowledge, and it must also could be executed. The role of the Behavior package is to model the possible behaviors of the entities within the VE; the objective being for the model to be interpreted in real-time by a behavioral controller, and to be introspected online. As for the structural aspects, intro- spection relies both on the behavioral model (M2) and on its ‘‘instantiation’’, that is to say, the way it is carried out (M1). The two classes which support these notions are Behavior and BehaviorExecution (Fig. 6). The Veha entities have reactive behaviors which are triggered by events that can be caused either by the learner or by another of the VE’s entities. Traditionally, behaviors are assigned pre-conditions and post- conditions concerning the entities and the environment. Behav- ioral modeling relies on state machines and the Uml activity model. Finally, it can also be based on functions written in programming language that can be consulted online (OpaqueBehavior). The first two methods are introspectable; the ITS can therefore describe or check the way the behavior is carried out. The tutor can thus analyze, explain or check the context in which an entity’s behavior is carried out by the learner. Better still, if a particular behavior has been specifically described (state ma- chine or activity) it can also explain the way it will be carried out. 4.2. Example of an environment in Veha The Veha metamodel can automatically interpret a model de- scribed in Uml. Fig. 5 shows the class diagram for an example of a VE in Veha. This example comes from an application created in Veha, but which has been greatly simplified for demonstration purposes. The application (GASPAR, Marion et al., 2007) is made up of around fifty classes and more than one thousand entities. This model shows the classes Deflector and CatapultControlPod Fig. 4. Visualisation of the instances of CatapultControlPod and Deflector. 8450 C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 (left window). The catapult control pod shields the operators work- ing on the catapult deck of an aircraft carrier. A pod can open (raise above the deck) or close (drop back down into the deck). The do- main model specifies all of the pod’s properties (height, speed, etc.). The reactive behavior of a pod is specified by a state machine (top right-hand window). This state machine is sensitive to the sig- nals Open and Close. Therefore, when the pod is Closed, if it re- ceives the signal to Open, it changes to the Open state and performs the operation: Open (). Within the context of this appli- cation, this operation is detailed by an OpaqueBehavior, a C++ code which carries out the visual displacement of the pod depend- ing on the speed attribute, and updates the height attribute. In much the same way, deflectors also react sensitively. Due to the additional needs of this demonstration, we added a testing operation (Test). This operation takes its settings from a catapult control pod. The behavior of this operation is specified by an activ- ity diagram (bottom right-hand window). Therefore, when a Test operation is evoked in an instance of the Deflector class, the operation sends the signal Open to the predefined pod. This model is defined using Objecteering modeling software. It is then exported in an XMI file. The first proposal is to add an inter- preter to the Veha metamodel within the AReVi virtual reality plat- form. The interpreter reads the XMI file and, for each class of Uml metamodel, creates an instance of the corresponding class in the Veha metamodel. Thus, for each domain class defined in the XMI file, the interpreter creates a new instance of the class class from the Veha metamodel. In the context of our example, an instance of the Class class is created for the Deflector class, and another created for the CatapultControlPod class. The interpreter en- ables the reification of the domain model and provides an set of methods facilitating the introspection of this model. It is therefore possible to ask the interpreter for the set of a class’s properties, the signal which enables the passing from one state to another, and the operation which will then be conducted, all independently of any tangible object. The VE is populated with entities; instances of the Entity class of the Veha metamodel. From a technical standpoint, these Fig. 3. Example of a procedure wri instances are defined in an XML file. Using Uml, class instances can also be described and exported in the XMI file. However, no Uml modeler can make it simple to attribute a geometry and a po- sition to these instances. The geometric design of the VE is, in gen- eral, the result of specialist modelers such as 3DS MAX or Blender. We therefore suggest using an export plugin for 3DS Max which would generate the instance file read by the interpreter. Fig. 4 show the visual result of the file defining the model (XMI) and the instance file (XML) in an application implemented using ARéVi. The interpreter also provides the methods for interrogating and manipulating the entities. It is therefore possible to ask an entity for its property values, to carry out an operation, or to send it a sig- nal in order to change its state. 4.3. Procedure and collaboration Here we shall examine the acquisition of skills. The domain model not only contains knowledge about the environment in use, but also knowledge about the task which must be performed within that environment by the learners. Within this context, activities are defined by the procedures describing the Actions to be performed by a number of entities, each with specifically de- fined roles. The procedures are therefore defined by UML activity diagrams. This kind of diagram uses the traditional possibilities for organizing its Actions (parallelism, sequence, junction, tten using an activity diagram. C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 8451 condition, etc.) As we are dealing with representing human activ- ity, we consider that the sequence of activities takes place in an asynchronous manner. The organization roles are represented by activity corridors. The name of the corridor defines its role and its type, as well as the type of agent that is authorized to take this role. As in Uml 2.1, there are many different types of activity. This could be the execution of an agent’s operations, a basic virtual action (playing an animation, reaching a given position, etc.) or sending a signal to a specific re- source. The resources are played by the environment’s entities and represented by objects in Uml. The conditions are expressed in Ocl and stem from the roles and resources participating in the proce- dure. Fig. 3 illustrates the example of a procedure expressed using an activity diagram. This procedure solicits the intervention of three roles (such as CabinOperator) which must be played by characters of a pre-defined type (PEH for example). The characters which play these roles are those which are effectively instantiated in the environment. This procedure aims to make the airplane which is to be catapulted advance towards a given point by manip- ulating the deflector (a protective plate). The example of the proce- dure in Fig. 3 illustrates the complementary nature of the state machines used to define the reactive behavior of the objects in the environment and the activity diagrams defining a procedure. A procedure’s action can be represented by sending an event to a given object to be manipulated, and the conditions of moving onto the following action can depend on the current state of the object. We implemented agents’ behaviors using knowledge about the procedures to select their actions. The learner plays one or more roles in the context of these procedures. The ITS also draws on this knowledge in order to choose which assistance to suggest. As for the environment, there are two levels of modeling available to the agents (including the ITS) and the users (instructors): the orga- nizational structure and the organisational instances. The intelli- gent tutor is therefore able to recognize the sequence of actions independently of all organization. It can also follow the precise progress of the procedure being carried out in the team in which the learner plays one or more roles. It is therefore able to detect the learner’s errors regarding the order of the actions to be com- pleted and the respect for the conditions defined in the procedure (Trinh, Buche, Querrec, & Tisseau, 2009). 5. Pedagogical model Knowledge about the environment (the entities and about the task to be performed) are represented with the Veha model. Our ITS can thus manipulate them in order to construct its own knowl- edge, as shown in Section 5.1, and can simulate pedagogical rea- soning (Section 5.2). Finally, a tangible implementation of the ITS is proposed in Section 5.3 (specification of the rules of simulated pedagogical reasoning). 5.1. Pedagogical situation It must here be emphasized that our work takes place within the context of in situ learning. Within this theoretical framework, the contextual elements are paramount in the ITS’s decision- making (Pomerol & Brézillon, 2001; Turner, 1993). In our case, we refer to context as the pedagogical situation which serves as a basis for decision-making. The aim is to define a context such as this from a ‘‘generic’’ standpoint, which would enable us to alter information without having to take into account the specific task being carried out. In order to do this, we must separate knowledge on the task to be performed (see Section 5.1.1) from knowledge about the learner (see Section 5.1.2). 5.1.1. Information concerning the task to be performed We positioned our work within the context of training for pro- cedural work. The aim of the ITS is to assist the learner in his/her progression through the procedure. First of all, we can consider the procedure as a sequence of ac- tions defined by an expert. The elements to be considered are therefore subject to sequencing which cannot be questioned, and sometimes cannot be explained. Secondly, we think that memori- zation of the sequence of actions could be facilitated through understanding. In this context Richard (1990) suggests adding the notion of sub-objectives to the procedure. To meet this aim, i.e. the completion of the procedure, an set of causally linked sub-objectives must be conducted. The procedure must therefore be studied taking into the account the distance to the procedure’s goal from a causal, rather than a chronological standpoint. The above analysis thus highlights two ways of dealing with procedural learning: the study of the domain sequencing links which are strongly linked to the roles in the procedure, and the study of causal links between sub-objectives: 1. Sequencing links Sequencing links conduct the relationships between the actions using the strict description of the procedure. They are the direct consequence of the sequencing of actions as defined by the expert. We are interested in the information linked to the actions closest to the action requested by the learner. More precisely: � the last correct action completed before that which the lear- ner has just solicited; � the action which has just been solicited by the learner; � the correct actions to be carried out, taking into account the role(s) to be played (which are potentially different to the solicited action); � the correct actions to be performed, when considering that all roles are played by the learner; � those actions following all the correct actions. We chose the actions closest to that which is solicited in order to try to reduce the ‘‘distance’’ between the goal (the end of the procedure) and the learner’s location in the procedure. Techni- cally, this is done by carrying out plan recognition based on the Veha activity diagram shown in Section 4.3. The pedagogical situation thus retains the knowledge linked to the actions that are chronologically close to that which is requested. 2. Causal links between sub-objectives The procedure can be considered like a graph representing the sequence of causal sub-objectives. We therefore contemplate all of the actions linked to that which the learner is performing. In concrete terms, this means the actions requiring the effect of the correct desired action (usage conditions, state of a resource, etc.). A distinction must be made between these links, which correspond to individual logic, and sequencing, whose links cor- respond to the organization of a collective procedure. Techni- cally, we are dealing with the links between post-conditions and pre-conditions as mentioned in Section 4.1.3. It must be stressed that our objective here is to extract knowl- edge relating to the work to be carried out in order to assist peda- gogical decision-making. Within this context, we shall consider the knowledge detailed in Table 2. All the actions which have been identified up to this point (sequential and causal links) make up the pedagogical situation. More specifically, we are interested in the information that relates to the selected actions. At this point, we must specify the knowledge relating to the concept of action. From this perspective, the ‘‘action context’’ is made up of knowl- edge that is directly linked to the Action (description, resources, etc.), knowledge relating to the Operation, which is the target Table 2 The pedagogical situation: knowledge about the task to be performed. Knowledge Nature Description r Context of the previous action Sequential The last correct action to have been performed. This action serves as a point of reference from which one can position oneself in the procedure s Context of the requested action Sequential The requested action. This action could be correct or incorrect. The action has not necessarily been performed, in accordance with the pedagogical model t Context of the correct action(s) without considering their roles Sequential In considering the last correct action, we can determine the actions to be performed within the context of the current procedure u Context of the correct action(s) Sequential A sub-group of the previous item which does not take the roles played by the learner into account v Context of the following action(s) Sequential For each correct action, we determine the actions which follow it according to the current procedure w Context of related action(s) Causal In considering the actions to be performed following the last correct action, we retrieve the ‘‘causal’’ links between the actions independently of the procedure. We therefore obtain the actions which are related 8452 C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 of the Action, as well as knowledge relating to the agent that has requested the action, as that agent is the central character. We therefore use action contexts in order to represent the knowledge associated with particular actions (a sub-group of the environment made up of the entities and agents considered relevant in the con- text of the action). It is the responsibility of the pedagogical agent to construct this set of knowledge. The pedagogical agent retrieves or constructs the knowledge required about the task to be performed when it re- ceives a message from the interface agent detailing an action which has been requested. This choice is be debatable and indeed another possible solution is to update the knowledge when an error occurs. We chose to reconstruct the knowledge of the actions in order to retain the option to intervene, even if the learner’s actions are correct. This means that we can provide pedagogical assistance in order to reassure the learner about the decisions that they make, or inversely to imply doubt if it looks like they are about to make a mistake (example: confirm false rules which contradict the choices the learner has made). 5.1.2. Information concerning the learner The information about the learner comes from a number of sources, but are all collected by the learner model. This information relates both to static data (such as age) and dynamic data (such as elements of memory at a given time). It must be noted that the learner’s errors are recorded and are analyzed. Our error model is based on the Cognitive Reliability and Error Analysis Method (CREAM). This approach proposed a classification scheme which makes a distinction between observa- tions of errors (phenotypes) and its causes (genotypes). The causal links between phenotype-genotype are represented using a num- ber of consequent-antecedent links. Finally, the scheme could be associated with a method of retrospective analysis (the search for causes). The most probable cause-effect links is found using Demp- ster-Shafer’s theory presented in El-Kechaï and Després (2006). In much the same way, the contexts relating to the actions are also recorded. This information allows us to see whether or not learner has already used a particular resource, for example. In concrete terms, we have just defined the input information and the pertinent elements from which pedagogical decisions can be made. 5.2. The pedagogical agent The pedagogical situation (Section 5.1) gives us the option of triggering pedagogical assistance relating to the elements detailed within it. It thus provides the possible outcomes of the pedagogical decision-making process. We shall now go onto define a model for simulating the behavioral decision-making of the pedagogical agent providing instructive assistance, that is to say, a model link- ing knowledge and the proposed assistance. It must be noted that we are working within the context of learning procedural and col- laborative tasks. We must therefore consider: � The atypical nature of the knowledge involved (knowledge stemming from basic pedagogical methods to virtual reality). � Adaptability (the agent’s reasoning processes must auto-adapt in order to take past experience into account). � This reasoning must be specified prior to the event (initial spec- ifications can therefore be made by an instructor). The criteria which arise from these considerations are as fol- lows: expressiveness, hierarchy, modularity, reactivity and adapt- ability. After having examined the existing families of behavioral architecture (connectionist, automata-based, rule-based), we opted for the rule-based families (Goran & Vladan, 2003) which best respond to the criteria outlined above. More precisely, we chose classifier systems (Sigaud & Wilson, 2007). This is a reactive and adaptive form of architecture, based on conditional rules. We propose the use of a model based on a hierarchical classifier system. This system organizes knowledge while taking into ac- count the abstraction of the data involved. It structures knowledge according to three levels, from rules based on abstract knowledge of educational methods (the pedagogical approach), to the rules based on concrete knowledge of virtual reality (pedagogical tech- niques), via an intermediary level (pedagogical attributes). Each level of abstraction contains sets which group together a number of rules. One set represents a way of dealing with a partic- ular approach, attitude or pedagogical technique. The rules are conditioned by the elements of the pedagogical situation, and favor the sets from the lower level. The system therefore uses a diffusion mechanism in all three levels which considers the rules matching the pedagogical situation. This gives rise to a list which then arranges the different suggestions for pedagogical assistance. Fig. 9 illustrates the structure and the dynamics of the pedagog- ical model controlling the pedagogical agent’s behavior. The infor- mation taken into account in the conditional part of the rules are retrieved by our ITS (pedagogical situation). These ‘‘inputs’’ are available at the three levels of data abstraction (approach, attitudes and pedagogical techniques). The rules whose conditional ele- ments are satisfied in terms of input favor some of the sets of ped- agogical rules from the lower level. The upper level (techniques), directly favors those pedagogical suggestions which can be applied within the environment. These suggestions are made to the instructor who chooses that which they consider to be the most relevant. Simulating pedagogical reasoning has two advantages: 1. As instructors are not always teachers, they too are being given pedagogical assistance. Fig. 5. Class diagram for a deflector and a catapult control pod. Fig. 6. Class diagram from the Veha metamodel: Behavior::Common package, the BehaviorExecution class. C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 8453 2. The instructors are not simulation software experts, so the ped- agogical agent will offer assistance to the learner who will have the opportunity to make the most of the VE. 5.3. Specifications of the pedagogical model In order to implement the pedagogical model, the teacher must specify: 1. The sets of rules for the three levels of abstraction. 2. The pedagogical rules for each of the sets of rules. Here, we will discuss information from the literature which can be used when specifying the pedagogical model. 5.3.1. Specifying the sets of pedagogical rules We worked from the studies by Lourdeaux et al. (2002) in order to define the sets of pedagogical rules. We obtained the following Tables 3–5 corresponding to the three levels; approaches, attitudes and techniques, respectively. This information represents the opportunity to specify sets of rules at each of the three levels (see Fig. 10). 5.3.2. Specifying the pedagogical rules Once the sets of pedagogical rules are defined, the teacher must specify the associated rules. A rule is represented by a sequence of characters. The effect and condition parts are based on the elements of the pedagogical situation. In the following example, we position ourselves at the Pedagog- ical Methods abstraction level, with an set of rules called Active. A first rule for this set is fulfilled is the learner is a novice (Learner.Level==novice), if they have performed an organiza- tion error (Learner.Error.type==procedural) and if the action performed is different from the correct action (!Task.Requeste- Fig. 7. Class diagram from the Veha metamodel: EntityClass. Fig. 8. Class diagram from the Veha metamodel: Features of a Classifier. 8454 C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 dAction in Task.CorrectActions). In this case, the rule favors the Explain set from the following level. if (Learner.Level == novice && Learner.Error.type== procedural && ! Task.RequestedAction in Task. CorrectActions) then (Explain) 5.3.3. Use A specific pedagogical model was created from the structure de- scribed above, and from the papers by Lourdeaux et al. (2002). These sets are described in Fig. 10, with each set containing an average of five rules. This pedagogical model was applied to two distinct VEs designed for learning collaborative procedures. No modification of the pedagogical model (sets and rules) is required for either of these applications, which are very different, but which are both based on the same kind of learning. However, we believe that these changes would only need to be made at the intermediate level. For other types of learning (for example for a scientific prac- tical), these rules would probably need to be changed. 5.4. Artificial learning Thanks to artificial learning, the weight of the rules can be re- fined for adapting to the instructor’s preferences. The learning algorithm is inspired by the Bucket Brigade (Holland, 1986; Wilson, 1987). This system distributes remunera- tions to the rules which enabled them to be obtained. It is adapted to classifier systems (Sigaud & Wilson, 2007) with a list of rules which, when followed one after the other, lead to an action. In our case, this sequence of events corresponds to the passing from one level to another. Remuneration is reflected by the instructor’s choice: the pedagogical technique which they choose defines the rules in the third level which will be compensated. By back-chain- ing, the rules in levels one and two are also compensated. The weights of the rules which match the pedagogical situation, but which participate in activating a technique other than that chosen by the instructor will decrease. The algorithm shares out the remu- neration, including a tax which means that the rules which rarely match are not put at a disadvantage, and that the strong rules are penalized in order to retain the adaptive nature of the system. Therefore, as the exercises progress, the pedagogical agent must make suggestions which correspond more and more closely to the instructor’s decisions. The pedagogical agent could therefore tem- porarily take over and directly apply the assistance that it has cho- sen itself, should there be more than one learner at a time. 6. Discussion Before concluding, we would like to discuss the benefits of our proposal. The study put forward in this article begins by a examin- ing previous studies in this field and analyzing the uses of pre-existing ITS within VLE. We then went onto show that the HAL system is the most successful, and underlined the elements which could be improved. Indeed, in this system, the pedagogical model depends partly on the exercise, and the errors and pedagogical strategies must be defined. Furthermore, the instructor can only choose between two pedagogical methods (active or explanatory). We believe that it is possible to resolve the pedagogical model’s problems of genericity and modularity. Without re-examining every element of our work, we can show how our proposal could solve some of the difficulties of existing models. The knowledge used for pedagogical reasoning does not Fig. 9. Complete representation of the pedagogical model. Table 3 Examples of set definitions for the ‘‘pedagogical approach’’ level of abstraction based on (Lourdeaux et al., 2002). Pedagogical approach Description Active/constructivist An active approach is learner-centered, considering him/her the main actor in the learning process. This approach suggests techniques through which they can produce, create and search. The knowledge required can be found in the environment Expositive/affirmative This is the most traditional approach which uses the display technique. It is based on a content-transfer approach. Knowledge is external Interrogative This approach makes recommendations to the learner, guiding him/her towards the desired outcome. Learners may have the impression that they have discovered something new, but it is the instructor who will have guided the thought process. Knowledge is internal Table 4 Examples of set definitions for the ‘‘pedagogical attitudes’’ level of abstraction based on (Lourdeaux et al., 2002). Pedagogical attitudes Description Perform Perform the task in the place of the learner. This strategy can be used by the instructor to show the learner the correct technique or move Disruption Some instructors tease and disrupt the learners by giving them incorrect information or potentially incorrect solutions in order to test the learners’ conviction of their ability to reason independently Suggest Showing where the learners can find theoretical information or where to find information within the environment. These attitudes allow the instructor to show the learners that they can find the required information independently and therefore deal with the situation in a calm manner Independent learning This attitude encourages the instructor to remain in the background as an observer rather than to intervene Explain The explanations and information are also designed, quite simply, to explain the functioning of certain devices, rules of analysis, safety rules, etc. Encourage Encouraging the learners when they perform a task correctly C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 8455 Table 5 Examples of set definitions for the ‘‘pedagogical techniques’’ level of abstraction based on (Lourdeaux et al., 2002). Pedagogical techniques Description Improvement Addition of visual and audio symbols or animated films Deterioration Unrealistic images (erased points of reference, feed-back, proprioceptifs dégradés, altered colors, blurred background/surround, reduction of objects, iconization, etc.) Upscaling Exageration of reality (representing objects on a larger scale, surreal, brighter or shinier objects, etc.) Simplification Simplification of the virtual scene (a crowd can be represented by people with simplified movements, simplified objects, simplified kinetic systems, wireframe images, etc.), schematic representations of certain devices Restriction Limitation of certain movements or actions (limiting the area within which the learner can move around, etc.) Animation Animated sequence (automatic positioning, keys which turn automatically once in place, etc.) Perspective Altering the learner’s normal viewpoint (view from behind, above, etc.) Modification Changes in appearance and texture (colors, flickering objects, etc.) Modeling The representation of abstract concepts, of physical phenomena invisible to the naked eye, types of errors, etc. Visualisation Hidden mechanisms (the inside of a motor, gears, etc.) Fig. 10. Specifying the three levels of the pedagogical decision-making model. 8456 C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 depend on the specifics of the task to be performed. Therefore ped- agogical rules do not and indeed do not need to consider specific information, (‘‘if the learner can see aeroplane 2 then. . .’’), but will rather use general knowledge independently of the exercise (‘‘if the resources of the correct actions are visible, then. . .’’). In much the same way, although the pedagogical assistance proposes tangi- ble solutions to the instructor (’’make the fireman flicker’’), generic knowledge is also manipulated independently of the exercise (’’make the characters involved in the following actions flicker’’). Thus, the genericity of our proposal is one of its strongest charac- teristics, as illustrated by the inclusion of our ITS at the heart of numerous applications: the learning of collaborative procedures on aircraft carriers (GASPAR) (Marion et al., 2007) and for firemen intervening in Seveso areas (SÉCURÉVI) (Querrec, Buche, Maffre, & Chevaillier, 2004). In addition, the pedagogical model of our ITS has strong modularity, as it offers the option of adding, deleting or modifying each of its components that participates in pedagog- ical decision making (rules or sets of rules). Moreover, the artificial learning mechanism adapts the proposed pedagogical assistance to the learner-instructor pair. Therefore, our proposition offers solu- tions to the problems raised in our introduction. Finally, it must be emphasized that PEGASE is directly based on the learner-instruc- tor relationship. However, we must not forget that there will undoubtedly be limitations linked to the use of our ITS in the context of non-proce- dural learning. To be able to deal with this kind of training, we C. Buche, R. Querrec / Expert Systems with Applications 38 (2011) 8446–8457 8457 would have to rethink the elements which are so strongly linked to the notion of procedure, i.e. knowledge of the pedagogical situation. References Amokrane, K., Lourdeaux, D., & Burkhardt, J. (2008). Hera: Learner tracking in a virtual environment. IJVR: International Journal of Virtual Reality, 7(3), 23–30. Edward, L., Lourdeaux, D., Lenne, D., Barthes, J., & Burkhardt, J. (2008). Modelling autonomous virtual agent behaviours in a virtual environment for risk. IJVR: International Journal of Virtual Reality, 7(3), 13–22. El-Kechaï, N., & Després, C. (2006). A plan recognition process, based on a task model, for detecting learner’s erroneous actions. In Intelligent tutoring systems (pp. 329–338). Goran, S., & Vladan, D. (2003). Building an intelligent system using modern internet technologies. Expert Systems with Application, 25(1), 211–230. Herviou, D., & Maisel, E. (2006). ARéViRoad: A virtual reality tool for traffic simulation. In Proceedings of urban transport (pp. 297–306). Holland, J. H. (1986). Escaping brittleness: The possibilities of general-purpose learning algorithms applied to parallel rule-based systems. In R. S. Michalski, J. G. Carbonell, & T. M. Mitchell (Eds.). Machine learning: An artificial intelligence approach (Vol. 2, pp. 593–623). Los Altos, CA: Kaufmann. Hubal, R. (2008). Embodied tutors for interaction skills simulation training. IJVR: International Journal of Virtual Reality, 7(1), 1–8. Kallmann, M., & Thalmann, D. (1998). Modeling objects for interaction tasks. In Proceedings of computer animation and simulation’98 (pp. 73–86). Keles, A., Ocak, R., Keles, A., & Gülcü, A. (2009). ZOSMAT: Web-based intelligent tutoring system for teaching-learning process. Expert Systems with Application, 36(2), 1229–1239. Levesque, P. (2003). Creation and use of 3d as-built models at EDF. In FIG working week 2003. Lourdeaux, D., Burkhardt, J., Bernard, F., & Fuchs, P. (2002). Relevance of an intelligent agent for virtual reality training. International Journal of Continuous Engineering and Life-long Learning, 12(1/2/3/4), 131–143. Marion, N., Septseault, C., Boudinot, A., & Querrec, R. (2007). Gaspar: Aviation management on an aircraft carrier using virtual reality. In Cyberworlds 2007 proceedings (pp. 15–22). Pomerol, J., & Brézillon, P. (2001). About some relationships between knowledge and context. In V. Akman, P. Bouquet, R. Thomason, & R. A. Young (Eds.), Modeling and using context: Third international and interdisciplinary conference, context 2001 (pp. 461–464). Berlin: Springer-Verlag. Querrec, R., Buche, C., Maffre, E., & Chevaillier, P. (2004). Multiagents systems for virtual environment for training. Application to fire-fighting. International Journal of Computers and Applications (IJCA), 1(1), 25–34 [Special issue ‘‘Advanced technology for learning’’]. Richard, J. F. (1990). Les activités mentales: Comprendre, Raisonner, Trouver des solutions. Paris: Armand Colin. ISBN: 2-200-2187-0. Rickel, J., & Johnson, W. L. (1999). Animated agents for procedural training in virtual reality: Perception, cognition, and motor control. Applied Artificial Intelligence, 13. Romero, C., & Ventura, S. (2007). Educational data mining: A survey from 1995 to 2005. Expert Systems with Application, 33(1), 135–146. Sigaud, O., & Wilson, S. W. (2007). Learning classifier systems: A survey. Journal of Soft Computing, 11(11), 1065–1078. Trinh, T., Buche, C., Querrec, R., & Tisseau, J. (2009). Modeling of errors realized by a human learner in virtual environment for training. International Journal of Computers, Communications and Control, IV(1), 73–81. Turner, R. M. (1993). Context-sensitive reasoning for autonomous agents and cooperative distributed problem solving. In Proceedings of the IJCAI workshop on using knowledge in its context (pp. 141–151). Wang, H. (1997). LearnOOP: An active agent-based educational system. Expert Systems with Application, 12(2), 153–162. Wenger, E. (1987). Artificial intelligence and tutoring systems. Los Altos, California: Morgan Kaufmann. Wilson, S. W. (1987). Hierarchical credit allocation in a classifier system. In Tenth international joint conferences on artificial intelligence (IJCAI’87), Milan, Italy (pp. 217–220). An expert system manipulating knowledge to help human learners into virtual environment Introduction Context: acquisition of skills using virtual environments Proposing an intelligent tutoring system: Pegase Domain model The Veha metamodel Domain concepts The possibility of structuring and interacting with the environment Entities’ behaviors Example of an environment in Veha Procedure and collaboration Pedagogical model Pedagogical situation Information concerning the task to be performed Information concerning the learner The pedagogical agent Specifications of the pedagogical model Specifying the sets of pedagogical rules Specifying the pedagogical rules Use Artificial learning Discussion References 
work_qzvhmckl7bdx3na2hecgn5nkuy	----	404 File not Found - ARRO - Anglia Ruskin Research Online Home About Accessibility Browse Browse by Year Browse by Faculty Browse by Author Advanced Search 404 File not Found Could not find the file: The specified file could not be found on this server. If you reached this page by following a link within the repository, please contact the ARRO - Anglia Ruskin Research Online administration. Otherwise, please check that you have typed the URL in correctly, or contact the person or site that supplied you with this URL. ARRO - Anglia Ruskin Research Online is powered by EPrints 3 which is developed by the School of Electronics and Computer Science at the University of Southampton. More information and software credits. 
work_r3cw4oxnnzfsrdzsslofkvck7e	----	79 Distributed Expert Systems for Planners MANFRED KOCHEN CHARLES BARR University of Michigan Expert systems can increase the effectiveness of planning. Planning information can be exchanged through intelligent computer/communication aids. More effective knowledge- based management of planning projects is possible through the use of such aids in distributed networks. We discuss the potential and limitations of such networks. When the Aswan Dam was planned, there was recorded as well as tacit knowledge about its unwanted side effects, such as the spread of schistosomiasis and diminished soil fertility. The fact that Egyptian’ experts now regard its benefits to greatly outweigh the costs and disutilities does not excuse lack of better planning. The knowledge may not have been retrieved and brought to bear at the right time. If it was, it may not have played an appropriate role in the planning process. We have yet to learn how to fully utilize relevant knowledge in planning. We need first to understand what knowledge it takes to plan various courses of actions that pursue a given policy. We need to determine whether that needed knowledge exists. If we believe it to exist, we need to learn how to retrieve it, possibly transform it so we can use it, screen and evaluate it, and bring it to bear at the right time. We must learn how to inject it into the planning process as high-technology, knowledge-based expert systems become available to planners. In the next section, we ask if the effective utilization of knowledge in planning is too low and, if so, why. What would its more effective use mean? Our main goal in the third section is to explore the potential of distributed intelligence in networks that link machines, people, and problem-solving actions. We ask what kind of intelligent knowledge- Knowledge: Creatim Diffusion, Utilization, Vol. 8 No. I, September @ 1986 Sage Publications, Inc. 80 networking system could be provided to increase the effectiveness of planning. Multiactorplanning demands a blend of consensus formation and negotiation, for which the complexity of organization of informa- tion will be much greater than for a plan executed by a single responsible &dquo;planner.&dquo; In the latter case, a logical specification of steps and contingencies will suffice for the goals selected by the planner. For large systems of actors and consequences, goals and means will be far more changeable. Communication will be needed if more effective planning is to be done. In the final section we discuss some of the problems involved in bringing such networks about, and conclude with recommendations for realistic action. Effectiveness of Knowledge- Use in Planning To plan is to specify spatial and temporal correspondences of actions believed to lead to a desiredfuture state. Many stakeholders, profes- sionals, politicians, and others are involved in planning any complex activity. A major source of complexity in planning, which creates barriers to effective knowledge utilization, is that no two planning participants interpret the key concepts in the same way, Desired future states are popularly called ends, and actions intended to attain them are called means. Feldt (1986) has proposed classifying planning situations according to the extent to which (1) the ends am clearly known, and (2) the means are clearly known. Dichotomizing these two degrees of clarity, he considers four types of planning situations for which four types of &dquo;theory&dquo; (actually belief systems or &dquo;isms’~ are appropriate: rationalism, when both means and ends are clearly known; incrementalism, when neither means nor ends are clearly known; utopianism, when ends are clearly known but means are not; and methodism, when means are clearly known but ends are not. &dquo;Clarity of knowledge&dquo; is, however, subject to a great deal of variability in interpretation. The concepts of structured, semistructured, and unstructured have been used by Simon. It is probably safe to assume,, that most writers would agree that the precise specification of relatively few variables to characterize a state or an action corresponds to clear knowledge, whereas vaguely felt but unarticulated states and actions correspond to unclear knowledge. In between lie descriptions involving many interrelated, poorly specified variables close to the unclear pole 81 and variables specified probabilistically or by fuzzy theoretic grades of membership closer to the clear pole. We cannot expect the same degree of effectiveness in knowledge use for different planning situations. Situations in which rationalism is appropriate call for a high degree of knowledge utilization. But incrementalism, whether it employs reflection-in-action (Schon, 1984) or a &dquo;learning&dquo; planning style (Friedman and Abonyi, 1976; Michael, 1973), also requires knowledge. The former uses deep knowledge recorded in the literature and is in the form of general principles and long chains of reasoning; the latter relies on experience and relatively broad and shallow knowledge in the form of heuristics for how to do something in a specialized domain. Idealized states and actions in the rationalistic approach correspond to those studied by a physicist when, for example, he or she seeks to accelerate a particle to a high-energy state by applying a sequence of small forces to it, as it accelerates, much as a parent adds small pushes to a child on a swing. The set of all possible states, as well as of all possible actions that can be taken in each and the resulting change of state, is what is to be clearly known and specified as a first step. The value of each state or of regions in state-space to the planners, or at least some preference ordering over the states, must be clearly specified and known as a second step. There must also be clear knowledge of the present state from which planning is tor proceed. The same holds for the desired future state. Lest it be thought that incrementalism fits the above model with the desired future state reachable from the starting state in one -incremental step, that is not the intent of writers about this approach. Neither the present state nor any future state is clearly specified; it is felt to be preferable to the starting state after it is encountered, but heuristics sometimes guide the planner about which action to choose. Thus, there are experts in &dquo;muddling through,&dquo; and it is their expertise that constitutes knowledge. Given the limitations to how much planning effectiveness can be expected with this approach, this kind of knowledge has probably been used effectively. Effective planning requires more interpretive aids to guide the use of various kinds of knowledge for these various planning environments. It requires sophisticated aids for discriminating among types of knowl- edge. Knowledge is needed by those playing roles as communicators, interpreters, and decision makers. Value-laden choices are evident at all levels of planning: For example, at the level of small projects, controversy arises in neighborhoods and cities over the siting of halfway 82 houses for ex-convicts or mental patients. But the information needed for planning large complex projects, and evaluating their effects in social, economic, and environmental systems, is greater than for small ones. Large projects are a challenge to information management as well as knowledge interpretation. Knowledge-transfer innovation is needed because of the diversity and complexity of the institutions participating in such projects. Low effectiveness of policymaking and planning is also partly a function of a project or planning timeframe being greater than the &dquo;visible horizon&dquo; or &dquo;planning foresight.&dquo; Oversimplification of com- plex issues may also contribute to low effectiveness, but knowledge management may counter the negative aspects of such reductionism. Uncertainties and changes in the &dquo;out years&dquo; undermine the long-term effectiveness of decisions made now. Second, parties to planning have unequal time horizons, unequal needs to economize, and unequal pressures to represent constituencies rigidly. Third, any decision, plan, or policy addresses an artificially constructed problem. This may be viewed as a closed system. It may also be seen as a system with oversimplified cause and effect and feedback loops to the rest of the environment. A solution to these problems is designing learning systems that detect and correct error and thus support adaptive planning (Michael, 1973). This avoids a combinatorial explosion of forecasting all possible world states. Still, investments of real, limited resources must be made in the present. Feedback should be available within a planner’s timeframe to help inform choices of resource investments. Measures of effectiveness chosen to meet these general needs must come to grips with the complexity of decision making in different styles of planning. Increasing Effectiveness Through Use of ‘ Distributed Expert Systems Our proposal for an interactive distributed expert-system network aims at raising the effectiveness of planning by extending mind power, bridging conceptual gaps, exchanging insights among planning stake- holders, integrating diverse knowledge, and facilitating satisfactory consensus formation and negotiation. Conveying research as well as routine knowledge through planning information systems would be one 83 activity such networks could support. The kind of system we propose eonsists of (1) people among whom consensus is required for a policy and a plan to work, and who are to be supported by the remaining five components; (2) procedures for their use of the subsystems; (3) models for their use, particularly inference engines (deductive and inductive); (4) data bases, knowledge bases, and bibliographic data bases combined into expert systems connected to the inference engines; (5) software, such as on-line advisory expert systems used as consultants, automated tutorial assistants, and utility programs that make the system so easy to use that it does not divert novice or casual users from their main task; and (6) a distributed computer communication system with storeand-forward facilities, conferencing, and other capabilities needed to support 1-5. Figure 1 sketches the overall structure of the system. Only a fraction of all the actors involved have sufficient power and influence to shape the policies and plans at the highest level. This group also motivates others to participate, and it sanctions the debate. Identifying these people is a critical success factor in planning. Any one of them can veto or sabotage a proposed policy or plan, and the agreement of each one is essential to success. Each one of them is therefore assumed to be a participant in the system, with ready access to it. Each owns as many as three terminals or personal computers usable as terminals: one in the office, if appropriate; a portable one for use in the field, car, farm, or site; one at home, where appropriate. Suppose that a motion for a particular proposal has been put before this group by one of its members, for resolution within a few weeks: for example, to assess the impacts of a plan and analyze its uncertainties, risks, and potential environmental and economic effects. The choice of agenda and time limit is flexible. If adding another month is felt to change the outcome, it may be done. If they use their terminals, they will be notified of this action item; but even if they don’t sign on or turn on their terminal, it will ring and blink at them until they do. (The MacIntosh, for example, has a continually active clock.) The proposal presents as objective and well-researched a case, taking into account the benefits and costs to all concerned, as the proposer can make without disguising his or her self-interest, if any. At first, everyone is asked to detect and reduce possible errors of commission and omission or of 84 04 I s BOS? :2 0 GIfila 0 <0 z 14 d F -~ 0 !2 m 0 40£ a) !A 85 stating the wrong problem or representing it improperly. They should check if all the key decision makers are participating. A major objective at this stage is to look at a set of possible solutions in the formulation of the problem that is qualitatively and quantitatively appropriate-not too many, not too few, not too rich and sophisticated, not too poor. This requires consensus building by synthesis of diverse worldviews or representations of the situation and effecting some shifts of represen- tation and some compromises. An educational subsystem embodying expertise in &dquo;getting to yes&dquo; (Fisher and Ury, 1981) about represen- tations could help do this. Games involving role playing and serving to improve communication-somewhat like corporate group therapy- have proved to be valuable tools in effecting such changes (Duke, 1984). Of course, generating a custom-made game in a few weeks is unlikely unless new computer aids to game construction are developed. The system is to help the participants detect differences in their representations, to diagnose from their expressions how they differ or concur intheir understanding of the issues. Each participant may have a personalized expert system to better help him or her understand the issue, but it is connected to a systemwide expert system in somewhat the way the various views of a data base relate to the data base adminis- trator’s view (i.e., to the conceptual schema). The overall system is to suggest ways of reducing differences. A simpler function is to check factual and logical inputs. Participants may challenge an input and request the checker to act. But the checker can be turned on at all times. Both deductive and inductive (statistical) inferences are checked, with results played back to all participants and used to update the record. The checking program is a consensually approved module, based on its performance on test cases. The next phase of system use, guided by a human discussion leader or, optionally, by turning on an automatic agenda maker/ presider, is consensus building of properties and constraints a solution to the problems at issue should satisfy. This requires shifts and compromises among values and interests after differences and commonalities have been made explicit. This is to be aided by an educational expert (sub) system that embodies &dquo;getting to yes&dquo; about values. This is similar in structure to the system for facilitating consensus building about representations. This is far less knowledge and logic-intensive and much more politically, psychologically, and socially sensitive, but at least one experiment has demonstrated the promise of computer conferencing for shifts in value (Kochen and Crickman, 1978). 86 Constraints are constantly added and modified during both policy- making and planning. Deadlines and strategic timing-and elastic constraints about these-are important factors. So are flexible budgets and resource constraints, including capacities for computation and communication. Cultural, psychological, social, and political con- straints are at least as important. All these must be stored in a continually growing and reorganized knowledge base, interfaced by a system that brings these constraints to bear at the right time. Recent publications explore the idea of the need for better systematiza- tion of information about project environmental effects. Dissemination of expertise in problem solving and planning processes is perceived as an effectiveness-increasing management tool. The National Academy of Science recommends that large-scale, knowledge-growth experiments be integrated opportunistically into environmental impact analysis efforts (National Research Council, 1986). The next phase in system use, probably overlapping with unternn- nated earlier phases, is the search for and use of knowledge that is necessary and sufficient to find a solution within the set of possible solutions and satisfying the agreed-upon constraints. We need to consider seven possibilities. * First, the needed knowledge may be readily available, but the participants cannot integrate as well the noncognitive factors that play a tots. It may seem to diminish what a participant thinks be or she knows. &dquo; In the second case, useful knowledge exists, but not one thinks of itt existence or of using it at the time it would help. An automatic reminder aid could help in this regard (Kochen and Floyd, 1983). . In the third case, useful knowledge exists, is thought of, but cannot be retrieved in time or in a form to be used. An improved information document retrieval system can help here. w In the fourth case, potentially useful knowledge has been retrieved. But because it may have been generated for a different purpose and require transformation, a system to assist in such a transformation may be needed. It may be automatically called on or, if special difficulties in transfor- mation arise, a subsystem offers advice to this effect. . Fifth, potentially useful knowledge may exist, but no one can specify what it is or recognize it. . Sixth, the needed knowledge does not yet exist. That, presumably, is believed to be the case for the water project for which a study of alternatives has been launched. * Seventh, we cannot know all we need to know to formulate perfectly effective policies and plans. That might hold for most cases. System use at 87 any time is characterized at best by a probability distribution over these seven possibilities and perhaps quite often by confusion. The planning abilities and actions supported by the network system described in this section are likely to increase in effectiveness for the following reasons. First, the system components provide for much better acquisition and interpretation of knowledge, which are needed for meeting the different kinds of planning challenges described earlier. Many kinds of planning have an arational character. Second, we pro- pose to decentralize and ease access to the technology for all planning participants, thus seeking to assure that the knowledge and knowledge- processing resources of planning are not in scarce supply for crucial planning participants. Third, the system design is intended to carry out a search for knowledge, of the type and at the times needed, more effec- tively than humans acting alone. We need now to address the feasibility of such a system. Feasibility Issues Assessing feasibility of the system requires appreciating the difficulty of problems yet to be solved and understanding some basic issues. The distinction between data, information, knowledge, understanding, and wisdom is relevant (Kochen, 1975). This article has so far blurred it by calling all these information. Observing or using an instrument generates data. If the data can be interpreted in an uncertainty reducing context, in which the scales of the instrument are explicitly presented with the data, the result is informa- tion. Of particular relevance to planning are three kinds of uncertainty: factual (technical or social); value-laden, based on opinions and reflecting priorities; and relating to actions by other affected organi- zations in the same area. Much as the units of knowledge in a knowledge system are clusters of information items imbued with meaning, the units of understanding are clusters of meta-information items that refer to the overall organization of a knowledge system. Van Lohuizen has introduced &dquo;insight&dquo; along with understanding. We take this to be the process by which an anomaly (gap, contradiction, dissonance, an emerging new pattern) in the fabric of knowledge is detected and stimulates the generation of a question. 88 Judging involves the use of values to assign priorities to units of understanding and knowledge. Deciding is essentially selecting or recognizing the instant at which to commit to a choice among viable alternatives, each of which the decider is prepared to accept. Policymaking and planning both involve problem-solving and decision-making processes. We have analyzed problem solving into three aspects: (1) specifying a set of possible solutions, alternative objectives for policymaking, and alternative subgoals for planning; (2) specifying a set of properties or constraints a solution should satisfy. These are constraints on, or desiderata for, planned paths in the case of policymaking and optional paths between subgoals for planning; others should also be induced to do 1 and 2. (3) Specifying the knowledge necessary and sufficient to select and justify objectives and constraints for policymaking and to select a path and persuade those who are to ensure it is followed in planning. We analyze decision making, after Simon, into three aspects. (1) Intelligence (scanning the environment and problem finding): This means identifying and diagnosing the key issues for policymakers and defining and analyzing the state-space for planners. (2) Design (invent, discover, and test solutions): This involves creating a strategy for policymakers and sketching the overall strategy and tactics to find a path in state-space (the universe of possible &dquo;states’* of the system being planned for) by planners. (3) Choice (selecting and implementing a solution): This means prescribing a recommended approach by policy- makers and working out a path in detail by planners. We could ask, in each case, to what extent the proposed system is likely to (1) increase effectiveness and (2) incur expected costs. The costs include the effects of overload and diversion that are opportunity costs, the direct costs of having built, maintained, and operated the system, and the expected risks arising from failures to build the system on schedule, cost overruns, and technical shortfalls. An important step in the design of the decision support system could be an evaluation framework measuring these and related factors. Because of the exploratory nature of this proposed system, many problems have not been clearly identified. Some major problems need at least to be analyzed in a preliminary way. One such problem is the extent to which computer conferencing (Turoff and Hiltz, 1977) enables a group to engange in a &dquo;multilogue&dquo; (a term used first by Kochen, and independently by Duke). Complex issues characterized by numerous nonquantitative random variables, interrelated probabilistically in 89 irrational, unprecedented situations with high stakes (possibly with irreversible consequences) characterize such multilogues. Conferencing can probably be done, at a cost. Kochen (1981) has proposed testing new tools for inquiring conferencing communities but notes that participa- tion parties possessing more or less issue-specific information, differing substantially in values, were probably infeasible for such an experiments This proposal is more ambitious because it attempts to break down those barriers to linking the parties to policymaking and planning. A research agenda to help identify and test mechanisms for achieving this includes applying incentives for successful networking. The agenda would also include the design and operation of a planning-support expert system to monitor and control the potential for information overload, information distortion, or misinterpretation. Of greatest interest, perhaps, is the potential for such networks to facilitate user learning and synergistic integration of problem-solving abilities among the network participants. The explicit role of user learning in planning processes will have to be better clarified, because there are time costs (at least) incurred by learning parties. Such costs will also be distributed unequally among the participants. Ways to measure these processes against some control standard are desirable. What are the grounds for believing that intelligence and expertise made available through a distributed computer network of the kind suggested here will improve the utilization of knowledge by planners and thus the effectiveness of planning? First, consider rationalistic planning styles in which these are appropriate: when both ends and means are fairly clearly known. To begin with, knowledge support can help planners to transform the planning situations they face into those with even more clearly known means and ends. Thereafter, the planner is provided with responses from a knowledge base and an inference engine to his or her needs for knowledge to use in specifying temporal and spatial correspondences of actions. Second, consider incremental planning styles in situations in which both ends and means are wholly unclear. Here, no system is likely to help the planner select variables, properties that future states should have, and so on. Instead, he or she is supplied by expert systems with advice about which incremental steps to take. One of the most significant barriers to success in the use of distributed expert systems to support planning requiring consensus among diverse stakeholders is that of incentives. It is necessary to motivate key persons to participate in the first place and to maintain communication at times 90 of flagging entbusiasm and competing concerns. Attainment and maintenance of revenues, credibility, power, and influences, supporting votes are motives of various stakeholders. Well-substantiated promises to help participants achieve these goals could be adequate incentives. Another problem concerns control over the proliferation of ideas. A large number of ideas of mediocre or bad currency may tend to drive out or drown the few ideas in good currency. Research on computer-based aids to filter out inputs of marginal value in computer conferencing or electronic mail could contribute to alleviating this problem. Peer review and other quality filters to exert some birth control on the generation of ideas may help. Although these and similar problems confront users of such a system, on balance it is feasible, at least as a learning experiment in increasing effectiveness. Conclusion Our proposal to develop a planning-aid expert system has shed light on the utilization of knowledge by suggesting a role for expert systems networks. The advantages of such an approach are that it helps to bring to bear methods and concepts from information science, which is a source of critical appraisals of information system technology and potential. Problem solving in the public sector, as within business enterprises, must generate plans; competitive pressures to do so effectively are ubiquitous. The extent to which information systems and artificial intelligence will improve planning effectiveness is a topic ripe for investigation. We make the following recommendation for realistic action: ex- ploiting an already planned network project to realize the potential of the system we describe. European universities had an offer from IBM to set up a multipurpose network, the European Academic Research Network (EARN), connecting mainframe computers in more than 60 academic and research centers, which would also connect to BITNET in North America (Dickson, 1984). European governments trying to decide between public-sector and private-sector information network developmental pathways have delayed launching the proposed network. The value-added character of the European Academic Research Network is instructive as a prototype for widely distributed planning 91 networks. We recommend investigating its use as a link between scientific research and decision making in planning. It would be best to restrict its use in such an investigation to a limited domain, probably as an experiment in the kind of mobilization of ecological knowledge called for by the recent NAS (1986) report on environmental manage- ment. An EARN-based distributed environmental management plan- ning expert network, capitalized with knowledge-based information technologies, will, we believe, provide European environmental planners with added leverage. Knowledge enters planning in subtle ways. Knowledge-based systems are therefore likely to be invisible and transparent: We will use tools to aid in the planning process-for example, an environmental scanner- without awareness that computer/communication-based expert sys- tems are hidden in the tool. The use of these systems must become so unobtrusive and intrinsic to tool use as not to divert from or interfere with performance of the task at hand. Increases in planning effectiveness are the goals and benefits of a knowledge-based distributed planning expert system. The goals and benefits are within reach of our design abilities now. References BENVENISTE, G. (1977) The Politics of Expertise San Francisco: Boyd & Fraser. CAPLAN, N. (1984) "Research on knowledge utilization lessons and observations." Report, CRUSK. Ann Arbor. Univ of Michigan Press DICKSON, D (1984) "Europeans look IBM gift horse m the mouth." Science 226(5 Octo- ber) 4670. DRAKE W , R MILLER, and D. SCHON (1983) "The study of community-level nutrition interventions. an argument for reflection-in-action " Human Systems Management 4, 2 DUKE, R (1984) Personal Communication FELDT, A (forthcoming) "Planning theories," in J. Snyder and A Catanese (eds) Introduction to Urban Planning FISHER, R and W URY (1981) Getting to Yes Negotiating Agreement Without Giving In Boston: Houghton Mifflin. FRIEDMANN, J and G ABONYI (1976) "Social learning a model for policy research " Environment and Planning GORDON, M., R BELEW, M KOCHEN and R. LINDSAY (1986) "Information construction through computer conferencing " Presented at National Onlme Meeting HART, H (1957) The Dark Missouri Madison Univ of Wisconsin Press 92 JACOBS, H. (1984) "The knowledge household." Presented at the Third International Conference, Seattle. KOCHEN, M. (1954) "An information-theoretical model of organizations." Trans. Inst. Radio Eng. 4: 67-75. &mdash;&mdash;&mdash;[ed.] (1975) Information for Action: From Knowledge to Wisdom. New York: Academic Press. &mdash;&mdash;&mdash;(1980) "Opportunity costs in computer conferencing during and for economic development." Proceedings of Workshop on the Economics of Communications. Honolulu: East-West Economic Institute. &mdash;&mdash;&mdash;(1981) "Inquiring communities on new tools for science-based policy-making." Working paper. Berlin: Wissenschaftszentrum. &mdash;&mdash;&mdash;(1982) "Barriers to planning in business." Human Systems Management 3, 4: 301-306. &mdash;&mdash;&mdash;(1984) "Information about how to retrieve information: its relation to the information retrieved." Presented at second symposium on Empirical Foundations of Information and Software Sciences, Atlanta. &mdash;&mdash;&mdash;and C. BARR (1986) "Planning theory: a cognitive approach," in Interdisciplinary Planning. New Brunswick, NJ: Center for Urban Policy Research, Rutgers State University. KOCHEN, M. and R. CRICKMAN (1978) "Citizen participation through computer conferencing." Technological Forecast. and Social Change, 1-32. KOCHEN, M. and B. FLOYD (1983) "Degree of constraint in use of an on-line time management system." Proceedings of National Online Meeting. Medford, NJ: Learned Information. KOCHEN, M., R. RAJAN, and Y. YUAN (1984) "Distribution of information processing for fiscal management in developmental planning." Proceedings National Online Meeting. Medford, NJ: Learned Information. MICHAEL, D. (1973) On Learning to Plan-and Planning to Learn: The Social Psychology of Changing Toward Future-Responsive Societal Learning. San Fran- cisco : Jossey-Bass. National Research Council (Application of Ecological Theory to Environmental Prob- lems Committee) (1986) Ecological Knowledge and Environmental Problem-Solving. Washington, DC: National Academy Press. National Research Council (Geophysics Study Committee) (1982) Scientific Basis of Water Resource Management. Washington, DC: National Academy Press. POLSBY, N. (1984) Political Innovation in America: The Politics of Policy Initiation. New Haven, CT: Yale Univ. Press. RIDGEWAY, M. (1955) The Missouri Basin’s Pick-Sloan Plan. Urbana: Univ. of Illinois Press. SCHON, D. (1984) The Reflective Practitioner. New York: Basic Books. TUROFF, M. and S. R. HILTZ (1977) "The process of communications and the problem of goals." Presented at 21st annual meeting, Society for General Systems Research, Denver. VAN LOHUIZEN, W. (1984) "The knowledge household and policy making." Working Paper. The Netherlands: Eindhoven University. 93 MANFRED KOCHENwas educated at MIT(1950)tmd at CohJmbia University in,New York (Ph. D., appliedmathe11fl1tics,19S5). Since 1972 he has been Professor of Information Science, with a joint appointmem as Adjunct Professor of Computertmd Information Systems at the Gra~t~Schooioj‘Br~i~,r A~ninist~ tion, University af Michigan. His major research interests are in problems of representation of knowledge. He is the author or editor of six boola. including Information for Action; he has over 150 scientific papers in technicaljoU17lllh: or books. In 1974 he was awarded the Award of Merit by the American Society for Information Science. CHARLES BARR (B.A., history, Yale College) is a Research Associate with the Information Science Program at the University of Michigan. His interests are in planning and policy analysis in the field of natural resources. He has worked for environmental organizations snd- in 1fUI1’Ulgement science research, with an emphasis on information systems development. 
work_r3n6zd6qknebjm5jfpcdhkc6t4	----	DISEASE MARKERS, VOL. 11,283-284 (1993) ANNOUNCEMENT INTERNATIONAL CONFERENCE ON NEURAL NETWORKS AND EXPERT SYSTEMS IN MEDICINE AND HEALTHCARE 24-26 August, 1994 Plymouth, England Organised by Plymouth Postgraduate Medical Schools and Centre for Intelligent Systems, University of Plymouth Honorary Conference Chairman: Professor Bruce McA. Sayers WHO and Imperial College, London In co-operation with The Institute of Physical Sciences in Medicine The Royal Society of Medicine Conference Co-Chairman: Dr. Emmanuel Ifeachor Centre for Intelligent Systems, University of Plymouth In the last few years there has been an explosion of interest in a number of intelligent systems techniques, notably in Neural Networks and Expert Systems. Medicine and healthcare have emerged as fertile areas for the application of these techniques because they require expertise (which is not always available), often involve non-trivial pattern recognition tasks, and because of difficulties with conventional methods. The aims of the conference are to bring together research engineers, medical informaticians, scientists, and medical experts, to discuss actual and potential applica- tions of neural networks and expert systems in medicine and healthcare, and to identify their benefits, limitations, current and future roles. The organisers invite papers from prospective authors in academia, industry, govern- ment and private organisations world wide in all areas of medicine and healthcare including, but not limited to biosignal processing, medical image processing, cardiology, obstetrics and gynaecology, paediatrics, neurology, radiology, occupational medicine, drugs and administration of medicine, patient management, therapy, prevention of diseases, care and rehabilitation, public health, and data security. Papers should come under one of the following general headings: • Neural Networks & Expert Systems Techniques • Neural Networks & Expert Systems Applications • Integration of Neural Networks & Expert Systems into Clinical Environment • Software/Hardware Tools & Environments © 1993 ASFRA B.Y. 284 INTERNATIONAL CONFERENCE ON NEURAL NETWORKS AND EXPERT SYSTEMS Practical papers are strongly encouraged, especially those that demonstrate signifi- cant potential improvements in medicine or healthcare, or cover topics such as the use of consensus expertise; limitations of neural networks or expert systems, safety and legal issues; hybrid neural networks-expert systems, integration with other techniques, e.g. genetic algorithms and multimedia; decision support tools; uncertainty handling with fuzzy logic; clinical verification, validation and assessment, integration of neural networks and expert systems into clinical information systems and environments, user interface, intelligent databases; computer aided tutoring; novel software tools and techniques. TUTORIAL COURSE on Expert Systems and Neural Networks 23 August 1994 Plymouth, England The tutorial will provide a practical introduction to the principles and techniques of neural networks and expert systems. It will be presented by respected experts in the fields, and should be of particular interest to clinicians, academics and industrialists who are considering the application of intelligent systems techniques to their problems or who want a good appreciation of the principles involved. Further information available from: Mrs. Jackie Masters Plymouth Postgraduate Medical School Derriford Hospital Plymouth, United Kingdom PL6 8DH Tel/Fax +44-752-792711 Submit your manuscripts at http://www.hindawi.com Stem Cells International Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 MEDIATORS INFLAMMATION of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Behavioural Neurology Endocrinology International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Disease Markers Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 BioMed Research International Oncology Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Oxidative Medicine and Cellular Longevity Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 PPAR Research The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Immunology Research Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Journal of Obesity Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Computational and Mathematical Methods in Medicine Ophthalmology Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Diabetes Research Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Research and Treatment AIDS Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Gastroenterology Research and Practice Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Parkinson’s Disease Evidence-Based Complementary and Alternative Medicine Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com 
work_r53uf4oktvbcpeybrhp4zbvgfq	----	Microsoft Word - IJBIS060402 DARAMOLA.doc 444 Int. J. Business Information Systems, Vol. 6, No. 4, 2010 Copyright © 2010 Inderscience Enterprises Ltd. A fuzzy expert system (FES) tool for online personnel recruitments J.O. Daramola*, O.O. Oladipupo and A.G. Musa Department of Computer and Information Sciences, College of Science and Technology, Covenant University, PMB 1023, Ota, Ogun State, Nigeria Fax: +234-1-7936529 E-mail: dwande@gmail.com E-mail: frajooje@yahoo.com E-mail: bola_musa@yahoo.com *Corresponding author Abstract: The advent of the internet has facilitated greater access to the myriad of job opportunities available globally. Currently there exist many job application submission portals that are being used for online job recruitment purposes. However, the task of many of these job submission portals is limited to matching the professional and academic qualifications of applicants with the requirements of employers and several organisations and does not involve the ranking of applicants’ credentials according to their relative suitability for the jobs applied for. In this paper, we describe the implementation of fuzzy expert system (FES) tool for selection of qualified job applicants with the aim of minimising the rigour and subjectivity associated with the candidate selection process. A performance evaluation of the FES tool that was conducted confirmed the viability of a FES-based approach in handling the fuzziness that is associated with the problem of personnel recruitment. Keywords: online job recruitment; fuzzy logic; fuzzy expert systems; FES; personnel selection; performance evaluation; job application portal; human resource management; HRM; organisation. Reference to this paper should be made as follows: Daramola, J.O., Oladipupo, O.O. and Musa, A.G. (2010) ‘A fuzzy expert system (FES) tool for online personnel recruitments’, Int. J. Business Information Systems, Vol. 6, No. 4, pp.444–462. Biographical notes: J.O. Daramola holds a PhD in Computer Science. Currently he is a member of faculty of the Department of Computer and Information Sciences, Covenant University, Nigeria. He also holds a Temporary Research Fellowship of the Centre for Mobile e-Services for Development, University of Zululand, South Africa. His research interest includes software architecture, software tools, expert systems, and knowledge-based intelligent systems. He has published many research articles locally and internationally. O.O. Oladipupo is currently a PhD research candidate in Computer Science of the Department of Computer and Information Sciences, Covenant University, Nigeria. She is also a member of faculty of the same department. Her research interest includes fuzzy logic, fuzzy expert systems and data mining. A fuzzy expert system (FES) tool for online personnel recruitments 445 A.G. Musa is currently a PhD research candidate in Computer Science of the Department of Computer and Information Sciences, Covenant University, Nigeria. He is also a member of faculty of the same department. His research interest includes artificial intelligence, grid computing and intelligent systems. 1 Introduction One of the opportunities provided by the internet is the opening up of the global job space through the increasing prominence of online job advertisement portals. This has greatly enhanced access of prospective job seekers to opportunities beyond national or continental boundaries. Also, it has provided greater opportunities for employers to recruit the most ideal candidates from a larger pool of applicants. This mode of online recruitment is becoming more popular not just because of its obvious advantages, in terms of cost reduction and increased visibility, but also because of the need to source for the most qualified candidates irrespective of sentiments. The eligibility criteria in most standard organisations are that candidates must 1 have a referee that introduces them 2 satisfy the minimum job requirements 3 possess other related qualities that can give them some kind of advantage (Werner, 2000). Today, there exist many job application submission portals such as job.com, alljobsearch.com, job-hunt.org, carebuilder.com and many more, where job seekers can submit their resume. Many of these portals have the capability to match the professional and academic qualifications of job applicants with the requirements of existing job vacancies in the available job databases. These job databases are created by aggregating the several job vacancies harvested from the submissions of many corporate organisations that have affiliations with the job submission portal for the purpose of recruitment. However, the functionality of many of the existing online job submission portals does not include the ranking of applicants’ credentials in order to determine their relative suitability for the jobs applied for (viz. they are ‘dumb’ job portals). This implies that these ‘dumb’ job portals lack the capability to alleviate the rigour and inherent subjectivity that is usually associated with the process of personnel selection. Personnel selection approaches so far reported in literature can be broadly classified into two categories: the analytic hierarchy process (AHP) approaches and the artificial intelligence approaches. The AHP makes use of a multi-hierarchical structure that is based on the theory of constructing hierarchies, setting priorities, and reasonable consistency (Saaty, 1995). The AHP is a powerful and flexible decision-making process to assist people set priorities and make the best decision when both qualitative and quantitative aspects of a decision need to be considered. The work of Liu and Shih (2005), and Scholl et al. (2005) are among the recent examples of AHP approaches. On the other hand, artificial intelligence approaches embrace the use of techniques such as: expert systems (Mehrabad and Brojeny, 2007); fuzzy expert system (FES) (Golec and Kahya, 2007; Canós and Liern, 2008; Canós and Liern, 2004); data mining (Chien and Chen, 2008; Wei-Shen and Chung-Chian Hsu, 2006); rough set (Chien and Chen, 2007); artificial neural networks (Drigas et al., 2004; Huang et al., 2006) for personnel selection. 446 J.O. Daramola et al. However, none of the efforts earlier reported in literature has been applied specifically to solve the problem of ‘dumb’ job application portals, such that they become enabled for relevance in the actual candidate selection process. In this work, a FES approach that can enable existing job application portals with personnel selection capability in a way that minimises subjectivity is presented. The issue of subjectivity can have a negative effect on the quality of a selection process if not properly controlled. This is particularly true in instances when the personal sentiments of the decision-maker come into play. Some of the real life instances that can ordinarily warrant subjective judgments include: 1 instances when the core competency needed by an organisation is not directly available in the database and there is the need to choose subjectively from candidates in closely related discipline 2 instances when two or more candidates have the same or almost similar qualifications 3 instances when the assessment parameters being used are fuzzy, inexact and qualitative (e.g. excellent, very good, good, fair, bad). Fuzzy reasoning is a model of human intelligence that can be introduced into computer systems using the mathematical concept of fuzzy logic. It empowers a system to be able to handle instances of approximate reasoning and fuzziness just like humans will do. Therefore, as a solution approach to the problem of fuzziness in the candidate selection process, we present in this paper, the design and implementation of FES that is capable of initiating intelligent decision making in the ranking and evaluation of the credentials of job candidates. To do this, the FES uses a set of objective parameters relative to specific job requirements to evaluate the relative suitability of the job candidates for the specific job vacancy concerned in way that minimises subjectivity. The rest of this paper is structured as follows. In Section 2, is a report of some related work, while an overview of relevant fuzzy logic concepts is given in Section 3. Section 4 presents an illustration of the personnel selection problem, while in Section 5, we give a detailed description of the design and implementation methodology of the FES, and the procedure adopted for its evaluation. The paper is concluded in Section 6 with a brief note. 2 Related work Recruitment is an important issue of human resource management (HRM). The main purpose of recruitment is to engage the best available talents from within or outside the organisation for available job positions. A number of AI-based approaches in the area of personnel selection have been reported in literature which is our bias in this paper. Wei-Shen and Chung-Chian (2006), proposed a personnel selection tool based on fuzzy data mining method to assist business managers to find eligible talent more efficiently. In the work, a fuzzy data mining method was applied to discover those predictors or attributes which are valid to predict the organisational behaviour of applicants in the future. This was done in contrast to the traditional personnel selection methods such as personality, interview, assessment centre, and biodata that were until then usually engaged in businesses. The fuzzy data mining method was applied to obtain valid fuzzy A fuzzy expert system (FES) tool for online personnel recruitments 447 association rules from existing transaction database using the Apriori algorithm. An experiment was implemented in a real case to validate the method proposed. The important contribution of this work is that it goes beyond personnel selection to predicting the future organisational behaviour of employees. Chien and Chen (2008) posited that human capital is one of the core competences for high-tech companies to maintain their competitive advantages in the knowledge economy, and that personnel recruitment and selection directly affect the quality of employees. In their research, decision tree analysis was employed to discover latent knowledge and extract the rules to assist in personnel selection decisions in a high-tech company. Furthermore, using the information gathered, domain experts from the company can also generate recruiting and HRM strategies. Specifically, a framework for human resource data mining was constructed to explore the relationships between personnel profiles and work behaviours. The Chi-squared automatic interaction detection (CHAID) algorithm was used as the data mining tool to explore the latent relationships among the input employee profiles and target variables of work behaviours such as job performance, retention, and turnover reasons. Their approach is one of the rare applications of data mining approaches for personnel selection so far reported in literature. Also, the work of Ranjan et al. (2008) outlines the role of data mining in HRM systems (HRMS) and shows how data mining can be used to discover new knowledge and extracts useful patterns from large data set for HR decision making. The work clearly demonstrates the ability of data mining to improve the quality of the decision-making process in HRMS and gives propositions regarding whether data-mining capabilities should lead to increased performance to sustain competitive advantage. There exist other instances in literature where fuzzy logic and fuzzy expect systems have been employed for personnel selection. Petrovic-Lazarevic (2001) presents a two-level employee personnel selection fuzzy model for short listing and hiring decision in order to minimise subjective judgment associated with the process of filling vacant job positions. The model used consists of an AHP of three levels. The first level involves the preliminary selection of candidates and the identification of the main decision factors which are represented as fuzzy linguistic variables (LVs) to revise multi-objective models of decision-making. The values of the variables are calculated as expected values of the fuzzy variables. The second level entails the process of selecting a final candidate for a job position. Here, the personal expectations of each employee are used as the basis to evaluate their fitness for the job, while the third level assesses the utility of engaging the appropriate employee. The work attempted to minimise subjectivity in the process of distinguishing between an appropriate employee and an inappropriate employee for a job vacancy using multi-objective models of decision making. Golec and Kahya (2007) opined that the employee evaluation and selection system is an important problem that can significantly affect the future competitiveness and the performance of an organisation. In their work, a comprehensive hierarchical structure for selecting and evaluating a right employee for a job was presented. The structure used allow to systematically build the goals of employee selection process, identify the suitable factor and measure indicators, and set up a consistent evaluation standard for facilitating a decision process. A competency-based fuzzy model was used to match an employee with a specific job requirement. An example was used to demonstrate the feasibility of the approach canvassed. As a contribution, the approach signifies a capable way to accommodate the imprecision, qualitative factors inherent in attempting to validate employee selection based on competency at the strategic level in an organisation. 448 J.O. Daramola et al. In Canós and Liern (2004), some fuzzy logic models for decision-making in HRM issues like personnel selection and optimal staff design were implemented. The flexibility of the models built allows for the ranking of the applicants for a job by using distance measures (Hamming distance) or similarity degrees with an ‘ideal candidate’. In cases, where an external valuation is available, the ordered weighted averaging (OWA) aggregation operators was used to assign different weights to selection criteria in order to imitate human expert valuation. In instances of merger or acquisition when the staff design is more complicated two fuzzy linear programming methods were proposed to make viable mathematical model for decision making in such instances. The fuzzy models were implemented in MS EXCEL to demonstrate the simplicity of their implementation. Also, Canós and Liern (2006) reported the development of a flexible decision support system (DSS) to help managers in their decision-making functions. The DSS simulates experts’ evaluations using OWA aggregation operators, which assign different weights to different selection criteria. An aggregation model based on efficiency analysis was used to rank the candidates into an order. This is premised on their belief that the main goal of managers is to obtain a ranking of a set of candidates who have been evaluated according to different competences. Therefore, the development of efficient and flexible information aggregation methods is a main issue in information access methods. The main point of difference of our work compared to these earlier approaches is that our effort is primarily addressed to solve the problem of ‘dumb’ job application portals, such that they become enabled for relevance in the actual candidate selection process. Hence this work offers as its contribution a case study of the application of FES approach in furthering the capabilities of existing online job application portals. 3 Overview fuzzy logic Fuzzy logic is a mathematical technique for dealing with imprecise data (attribute) and problems that have many solutions or values rather than one. The concept was conceived by Zadeh in 1965 (Zadeh, 1965, 1975). It represents knowledge based on the degree of membership. It is a theory of fuzzy sets, the set that calibrates vagueness and provides an approach to approximate reasoning in which the rules of inference are approximate rather than exact. Fuzzy set theory provides a systematic calculus to deal with information linguistically and perform numerical computation by using linguistic labels stipulated by membership function (MF). A fuzzy set A in a universe of discourse X is characterised by MF which associates with each element x in X (x ∈ X) a real number in the interval [0, 1], which indicates the extent to which x belongs to the fuzzy set X (Xu et al., 2005). A fuzzy number is a fuzzy subset in the universe of discourse X that is both convex and normal. Fuzzy set is different from classical set in the sense that while a classical set considered its elements based of Boolean logic; that is an element is either a member or not (1 or 0), fuzzy set considers all its elements based on fuzzy logic (viz. it considers every element based on their degree of membership in X) (Schneider et al., 1996). The definitions of some core fuzzy logic concepts that are relevant to the context of this paper are given as follows: Definition 1: A LV is a variable whose values are expressed in linguistic terms. In other words variables whose values are not numbers but words in a natural or artificial A fuzzy expert system (FES) tool for online personnel recruitments 449 language [3]. For example ‘class of degree’ is a LV whose values are 1st class, 2–1 (second class upper division), 2–2 (second class lower division), 3rd class and pass. Definition 2: The MF μA (x) can be defined as the membership of the elements x of the base set X in the fuzzy set A. The grade of membership μA (xo) of a MF μA (x) describes to which grade, the special element x = xo belong to the fuzzy set A. The MF maps the elements of the universe on to numerical values in the interval [0,1]. A fuzzy set A is a set of ordered pairs: [ ] [ ](( , ( )) : 0,1 , , (x) 0,1 .A AA x x X x Xμ μ= → ∈ ∈ ( ) ( ) ( ) 1 if is totally in 0 if is not in 0 1 if is partly in . µA x x A µA x x A µA x x A = = < < Definition 3: Assume A is a fuzzy subset of X, the support of a fuzzy set A is the set of all points x in X such that μA(x) > 0: ( ) ( ){ 0 and }Support A x A x x Xμ= > ∈ Definition 4: The core of a fuzzy set A is the set of all point x in X such that μA(x) = 1: ( ) ( ){ =1 and }core A x A x x Xμ= ∈ 3.1 MF formulation and parameterisation A fuzzy set is completely characterised by its MF, and as such the classes of parameterised functions commonly used to define MFs of one dimension are: triangular MF, trapezoidal MF, Gaussian MF, generalised bell MF (Padhy, 2005) a Triangular MFs: This is specified by three parameters (a, b, c) where (a < b < c). It can be represented in either of the following: ( ; , , ) max min ' , 0 x a c x triangle x a b c b a c b ⎛ − − ⎞⎛ ⎞ = ⎜ ⎟⎜ ⎟− −⎝ ⎠⎝ ⎠ (i) 0, , ( ; , , ) , 0, x a x a a x b b atriangle x a b c c x b x c c b c x ≤⎧ −⎪ ≤ ≤⎪ −= ⎨ − ⎪ ≤ ≤ −⎪ ≤⎩ (ii) b Trapezoidal MFs: This is specified by four parameters {a, b, c, d} where (a < b < = c < d). It can be represented in either of the following: ( ; , , , max min ,1, , 0 x a d x trapezoid x a b c d b a d c ⎛ − − ⎞⎛ ⎞ = ⎜ ⎟⎜ ⎟− −⎝ ⎠⎝ ⎠ (i) 450 J.O. Daramola et al. 0, , ( ; , , , ) 1, , 0, x a x a a x b b a trapezoid x a b c d b x c d x c x d d c d x ≤⎧ −⎪ ≤ ≤⎪ −⎪ = ≤ ≤⎨ −⎪ ≤ ≤⎪ − ⎪ ≤⎩ (ii) c Gaussian MF: A Gaussian MF is specified by two parameters {c, σ}; c represents the MF’s centre and σ determines the MF’s width. 2 2( ) / ( ; , ) exp 2 x c gaussian x c σ σ ⎛ ⎞− − = ⎜ ⎟⎜ ⎟ ⎝ ⎠ d Generalised bell MFs: A generalised bell MF or simply called bell MF is specified by three parameters {a, b, c}: 2 1 1 ( ; , , ) bx c a bell x a b c − + = where parameters a, and c represents the width and the MF’s centre respectively while b represents the slope of the crossover point, and it is usually positive (if b is negative, the shape of this MF becomes an upside-down bell). 3.2 Basic operations in fuzzy logic The basic operations that could be performed on fuzzy sets include: 3.2.1 Compliment This shows to what extent an element does not belong to the set. For example, if we have the set of tall men, its complement is the set of NOT tall men. The membership representation is shown in Figure 1. Figure 1 Compliment of fuzzy set Membership grade 1.0 0.5 0 5 10 15 20 Tall Not tall Height (ft) A fuzzy expert system (FES) tool for online personnel recruitments 451 If A is the fuzzy set, in a universe X. Its complement A can be found as follows: ( ) ( )1 for all x XA x A xμ μ= − ∈ 3.2.2 Intersection This shows the extent to which an element is in both sets (see Figure 2). For example, the intersection of the set of tall men and the set of fat men is the area where these sets overlap. A fuzzy intersection is the lower membership in both sets of each element. The fuzzy intersection of two fuzzy sets A and B on universe of discourse X: ( ) ( ) ( ) ( ) ( )min[ , ] , for all A B x A x B x A x B x x Xμ μ μ μ μ∩ = = ∩ ∈ Figure 2 Intersection of fuzzy sets 3.2.3 Union Figure 3 Union of fuzzy sets 1.0 0.5 Membership grade 5 10 15 200 Tall Fat Tall ∩ Fat Membership grade 1.0 0.5 0 5 10 15 20 Tall Fat Height (ft) Tall ∪ Fat 452 J.O. Daramola et al. This shows how much of the element is in either set. In fuzzy sets, the union is the reverse of the intersection. That is, the union is the largest membership value of the element in either set (see Figure 3). For example, the union of tall men and fat men contains all men who are tall OR fat. The fuzzy operation for forming the union of two fuzzy sets A and B on universe X can be given as: ( ) ( ) ( ) ( ) ( )max[ , ] , for all A B x A x B x A x B x x Xμ μ μ μ μ∪ = = ∪ ∈ 3.3 Fuzzy expert system A FES is an expert system that employs the concepts of fuzzy logic instead of the Boolean logic to reason about data as its inference mechanism (Aly and Vrana, 2006). It consists of fuzzification, inference, knowledge base, and defuzzification subsystems, and uses a collection of fuzzy MFs and fuzzy rules. It has the capability to solve decision making problems for which no exact algorithm exists by relying on human-like models of approximate reasoning that are expressed in form of fuzzy IF-THEN rules. FESs are well suited to problems that exhibit uncertainty resulting from inexactness, vagueness or subjectivity. The structure of a FES is illustrated in Figure 4. Figure 4 Standard FES architecture Source: Kosko (1992) 3.3.1 Fuzzy inference system This is a popular computing framework based on the concept of fuzzy set theory, fuzzy IF-THEN rules and fuzzy reasoning. The fuzzy inference engine is responsible for the evaluation of fuzzy rules to produce an output for each rule (Wang and Mendel, 1995; Padhy, 2005). The basic structure of fuzzy inference system consists of three conceptual components: a rule base, which contains a selection of fuzzy rules. A database (or dictionary), which defines the MFs used in the fuzzy rules; and a reasoning mechanism, which perform the inference procedure upon the rules, and the given facts, to derive a reasonable output or conclusion. A fuzzy inference system can take either fuzzy inputs or crisp inputs but its outputs are almost always fuzzy sets. Sometimes it is necessary to have a crisp output, especially in a situation where a fuzzy inference system is used as a controller. In such cases a method of defuzzification is needed to extract a crisp value that best represents a fuzzy A fuzzy expert system (FES) tool for online personnel recruitments 453 set. The three most popular fuzzy inference systems are the: Mamdani fuzzy model, Tagaki-Sugeno fuzzy model, and Tsukamoto model. The FES tool for personnel recruitment that is implemented in this work is based on the Mamdani fuzzy model. 3.3.2 Mamdani fuzzy inference The Mamdani-style fuzzy inference process is performed in four steps: 1 fuzzification: this involves definition of fuzzy sets, and determination of the degree of membership of crisp inputs in appropriate fuzzy sets 2 inference: this involves evaluation of fuzzy rules to produce an output for each rule 3 composition: this entails aggregation or combination of the outputs of all rules 4 defuzzification: this involves computation of crisp output. 4 An illustration of the recruitment problem In order to demonstrate the relevance and the applicability of the FES for personnel recruitment, an illustration of the recruitment problem scenario is presented as follows: Suppose an organisation wishes to secure a candidate for a particular job position and desires to analyse and evaluate the resume of various candidates submitted online in respect of the specific job advertised (note that since the focus is online submission, there are some factors that may not be considered since personal interaction with each candidate might not be possible, however, considerably crucial information can be collected through the resume which can be a basis for evaluation). Given that the required qualifications and other desirable factors are clearly stated, or stated imprecisely such as: Qualification, class of degree, years of experience (YOE), certifications, age, course of study (COS) etc. We assume that the post of a system engineer is considered to be vacant and needed to be filled with the requirements stated as follows in Table 1. Table 1 Requirements for system engineer S/no Requirements parameters Fuzzy values 1 Qualification MSc or equivalent 2 Class of degree Second class lower (2–2) (minimum) 3 YOE Not less than five years and not more than ten years 4 Certification Professional certification (offers advantage) 5 Age Within the age range of 25–30 6 COS Computer engineering or related discipline If the permissible range of input values for candidate per requirement is given as follows: qualification 1 BSc*/HND*/0.8 2 BSc’/HND’ + PGD*/0.4 3 MSc*/1.0 4 MSc’ + PGD*/0.6 454 J.O. Daramola et al. 5 PhD*/0.4 where * connotes that degree is in computer engineering and ’ connotes that degree is not in computer engineering but in a related discipline class of degree 2.4 < x < = 5.0 (for a maximum of 5.0 grade point average) YOE 4 < x <= 10 certification yes/no (1 or 0) age 25 < = x < = 30 COS 1 computer engineering/1 2 electrical/electronics/1 3 computer science + PGD (comp eng)/0.8 4 physics (electronics)/0.8 5 computer science/0.6 6 any other engineering courses + PGD (comp eng)/0.4. 4.1 Solution approach In solving this example problem, the different types of qualifications and COS, have been assigned fuzzy values in the interval [0, 1] based on their relative closeness to the specific job requirement. In this specific case, computer engineering and electrical engineering have the value 1.0 because they perfectly match the COS requirement for the ‘ideal’ candidate for the system engineer position, hence the values are used as direct fuzzy number inputs. Also, other real valued LVs that are part of the requirement parameters are assigned appropriate membership values by fuzzification once the fuzzy set for that parameter has been defined. As an example, the case of two candidates that have applied for the post of system engineer with differing credentials is shown in Table 2. Table 2 Credentials of candidates S/no. Qualification Class of degree/ cumm. grade point YOE Professional certification Age COS 1 BSc 2–2 (3.7) 5 Yes 26 Computer engineering 2 HND + PGD 2–2 (2.95) 6 Yes 29 Computer science The fuzzy set representation for the 6 basic requirements is given as: qualification (BSc*/HND*)/0.8 + (BSc’/HND’, PGD*)/0.4 + (MSc*)/1 + (MSc’ + PGD*)/0.6 + (PhD*)/0.4 class of degree 2.49/0 + y = –0.99 + x/2.51 + 5.00/1 (using the linear MF) YOE 4/0 + y = - 0.67 +x/6 + 10/1 A fuzzy expert system (FES) tool for online personnel recruitments 455 certification yes/1 + no/0 age 25/0 + 26/0.33 +27/0.67 +28/1 + 29/0.5 + 30/0 (using triangular MF). Then, the membership values for the requirement parameters of the first and second candidates are given as: C1 1 / 0.80 2 / 0.48 3 / 0.16 4 / 1.00 5 / 0.33 6 / 1.0= + + + + + C2 1 / 0.40 2 / 0.18 3 / 0.30 4 / 1.00 5 / 0.50 6 / 0.60= + + + + + The fuzzy hamming distance measure for the two candidates under consideration using all six parameters is shown in Table 3. Table 3 Fuzzy distance measure for candidate ranking ( ) ( )O i R iµ x µ x− Parameters C1 C2 Qualification 0.2 0.6 Class of degree 0.52 0.82 YOE 0.84 0.7 Certification 0.0 0.00 Age 0.67 0.5 COS 0.0 0.4 ( ) ( ) 1 ( , ) n O i R i i O R µ x µ xδ = = −∑ 2.23 3.0 Based on the result obtained, it can be concluded that the first candidate (C1) is closer to the requirement for the job of system engineer than second candidate (C2). Hence C1 is ranked higher than C2 in the order of job consideration. The solution approach demonstrated in this example is the one emulated by the FES tool for job recruitment that is presented in this paper. 5 Description of architecture of FES tool The architecture of the FES tool (see Figure 5) consists of following: 1 System input: This consist of the candidates’ profile database, and the job vacancy database. The candidates’ profile database contains information that pertains to individual candidates as obtained from the candidate profile form filled by each candidate online. It provides the pool of information on job-seeking candidates that are subsequently evaluated for job placements. The candidates’ profile database grows continually as more candidates subscribe to the FES. The job vacancy database is the other input into the FES which contains the specific employment requirements which a prospective job seeker must fulfil in order to qualify for consideration for the available jobs. 2 The fuzzy inference system: This is the fuzzy logic engine of the FES which consists of: 456 J.O. Daramola et al. • A fuzzifier component: This handles the conversion of input values obtained from candidates’ profile into fuzzy values within requirements parameter fuzzy set. • A fuzzy rule base: This contains the set of fuzzy IF-THEN rules that embodies the knowledge used by the fuzzy inference system. It represents the knowledge base of the expert system. Each of the rules has an antecedent (IF) and a conclusion (THEN) part that prescribes what should be done when certain conditions are true. The specific criteria for selection for each specific job position are represented as IF-THEN rules which determine the eligibility of candidates. • A fuzzy inference engine: This implements the fuzzy reasoning of the FES by combining the fuzzified inputs with the fuzzy rules using the Mamdani-style fuzzy inference process. 3 Candidate ranking components: This computationally evaluates the closeness of each candidate’s fuzzified credentials to the ideal requirements for a specific job using the fuzzy hamming distance function. This fuzzy distance metric is used to rank candidates’ profile in the order of their eligibility for the job. The fuzzy hamming distance is given as (Padhy, 2005): ( ) ( ) 1 ( , ) n O i R i i O R µ x µ xδ = = −∑ where ( ) 1O iµ x = if x is totally in O ( )0 1R iµ x< < if x is partly in R for each parameter included in a specific job requirement. 4 Decision and result interface: This is a logic component that ensures that only candidates that score above the pre-determined minimum cut-off for a particular job vacancy are recommended. It also has a GUI interface that displays the final result. Figure 5 Conceptual view of FES for candidate selection Job vacancy database Fuzzifier Fuzzy rule base Candidate ranking component Candidates’ profile database Fuzzy inference system Job seeker qualification Specific job requirements Inference engine Decision and result interface A fuzzy expert system (FES) tool for online personnel recruitments 457 Figure 6 FES interface showing fuzzified values from candidate credentials (see online version for colours) Figure 7 FES interface showing data captured from candidates (see online version for colours) 5.1 Implementation details The implementation of the FES tool was based on Java servlet technology, running on Sun Application Web Server 9.0 using the NetBeans 5.5.1 Java application development environment. The web interface for application submission was implemented using macro media flash and dream weaver web design tools, and Java 458 J.O. Daramola et al. server pages (JSP). The candidate profile database and job vacancy database were implemented as data tables with MySQL using the JDBC connector as the middleware for client-database connectivity. The MATLAB Builder JA 2.02 (www.mathworks.com/products/javabuilder) which allows functional modules that are developed using MATLAB to be automatically packaged as classes was used to build and wrap the fuzzy logic components of the FES (fuzzifier, fuzzy inference engine, fuzzy rule base) in Java codes so that they could be referenced as a standard Java classes. Figures 6 and 7 are snapshots from our implementation. 5.2 Performance evaluation of FES tool Evaluation is the process of determining the appropriateness of a specific system relative to its functional requirements and objectives. Validation of an expert system is conducted by determining whether the system’s outcome is consistent with the conclusions of the human experts. Validation focuses on evaluating the outcomes rather than the process by which the outcomes are determined. In our specific case, the output of the FES tool was validated by using a direct method following the approach of Salim et al. (2002). The direct method of expert system validation is a quantitative assessment that requires the human expert to obtain a copy of the software to be evaluated and to perform a simple benchmark problem on the software. Based on past experience, the human expert answers a set of 14 questions about the software. The questions are quantitative and are based on a 0 (very false) to 5 (very true) numerical scales. After some arithmetic manipulation of these numbers, a single numerical factor results called the ‘satisfaction level’ that ranges from 0 (least satisfied user) to 5 (most satisfied user) is used to rank the expert system software in terms of its likelihood to satisfy a prospective end user. 5.2.1 Description of the evaluation experiment In conducting a validation test on the FES with the primary aim of determining the users’ satisfaction level, a small experiment to test the system’s conclusions against those of human experts was conducted using the direct method. Ten human resource experts from Akintola Williams group (a leading human-resource management firm in Nigeria) were asked to participate in the survey. Each received an identical set of questionnaire alongside a copy of FES software. We installed the MATLAB component runtime (MCR) library on each of the target machine to ensure that the embedded MATLAB files in the FES function normally. The MCR consist of the full set of shared libraries required for executing MATLAB based components and provides complete support for all features of the MATLAB language and its toolboxes. The human experts were asked to give a rating of 0 to 5 of their assessment of the performance of the FES tool covering seven aspects and a total of fourteen questions. Table 4 shows a sample of a filled FES questionnaire test instrument and a typical computation of satisfaction score for one evaluator. A fuzzy expert system (FES) tool for online personnel recruitments 459 Table 4 A typical FES tool evaluation report Question Assessment value (0–5) Weight Value X weight Correctness of answer 1 Is there enough information to evaluate the FES tool? 4 (2) 8 2 Does the FES tool reach the same conclusion similar to that of a human expert? 5 (2) 10 3 Does the FES tool provide the right answer for the right reasons? 5 (2) 10 Accuracy of answer 4 Is the FES tool accurate in its answer(s)? 5 (2) 10 5 Is the answer complete? Does the user need to do additional work to get a usable result? 4 (2) 8 Correctness of reasoning technique 6 Does the answer change if new but irrelevant data is entered into the FES tool? 5 (1) 5 7 Does the system require a lot of irrelevant question to reach its answer? 0 (1) 0 Sensitivity 8 Does the answer change if irrelevant changes are made to the system rules? 5 (1) 5 Reliability 9 Does the FES tool crash or hang up in its host computer? 2 (1) 2 10 Does the system give warnings for instances of incomplete data or rules? 5 (1) 5 Confidence 11 Is one comfortable using the FES tool? 4 (1) 4 12 Does the conclusion of the FES tool give adequate satisfaction? 4 (2) 8 Limitations 13 Can limitations of the FES tool be detected at this point in time? 4 (1) 4 14 Can the FES tool learn from increased data or experience? 2 (1) 2 20 81 ( ) ( )Result weight value weightx=∑ ∑ 4.05 460 J.O. Daramola et al. The weighting factors used for the computation of the satisfaction score in the experiment was based on the mutually agreed values that were established by the human expert evaluators that participated in the experiment. At the end of our experiment, the mean satisfaction level as computed from the performance evaluation of the ten human expert evaluators was 3.9 out of the possible maximum score of 5.0, which is indicative a good performance (see Table 5). Table 5 Result of evaluation experiment Evaluator Computed satisfaction level 1 4.05 2 4.16 3 4.21 4 3.52 5 3.84 6 4.22 7 4.21 8 3.89 9 4.2 10 3.57 Mean satisfaction level 3.987 6 Conclusions In this paper, in contrast to many previous approaches, an attempt has been made to enable online job application portals with intelligent capability in order to make them more relevant in the actual process of candidate selection. Specifically, a FES tool was implemented to address the subjectivity and fuzziness associated with the manual process of human resource recruitment. The feasibility of a FES approach was demonstrated using a specific case study. An evaluation of the FES tool that was undertaken yielded satisfactory result to confirm the viability of a FES approach to enhancing the existing capabilities of many online recruitment systems. In future work, the possibility of engaging hybrid intelligent models for candidate selection process, and future personnel behaviour prediction would be explored. We will be looking to leverage the combination of fuzzy logic, rough set and some instance-base learning approach such as artificial neural networks and case-based reasoning. Acknowledgements The authors of this paper wish to deeply appreciate the thoroughness and objectivity of IJBIS referees that were used to review this work. We sincerely acknowledge that their informed comments during the peer-review process helped a great deal to improve the quality of this work, we urge them to keep up the good work. A fuzzy expert system (FES) tool for online personnel recruitments 461 References Aly, S. and Vrana, I. (2006) ‘Toward efficient modeling of fuzzy expert systems: a survey’, Agriculture Economics, Vol. 52, No. 10, pp.456–460. Canós, L. and Liern, V. (2004) ‘Some fuzzy models for human resource management’, Int. J. Technology, Policy and Management, Vol. 4, No. 4, pp.291–308. Canós, L. and Liern, V. (2008) ‘Soft computing-based aggregation methods for human resource management’, European Journal of Operation Research, Vol. 189, No.3, pp.669–689. Chien, C.F. and Chen, L.F. (2007) ‘Using rough set theory to recruit and retain high-potential talents for semiconductor manufacturing’, IEEE Transactions on Semiconductor Manufacturing, Vol. 20, No. 4, pp.528–541. Chien, CF. and Chen, LF. (2008) ‘Data mining to improve personnel selection and enhance human capital: a case study in high-technology industry’, Expert Systems and Applications, Vol. 34, No. 1, pp.380–290. Drigas, A., Kouremenos, S., Vrettaros, S. and Kouremenos, J.D. (2004) ‘An expert system for job matching of the unemployed’, Expert Systems with Applications, Vol. 26, No.1, pp.217–224. Golec, A. and Kahya, E. (2007) ‘A fuzzy model for competency-based employee evaluation and selection’, Computer and Industrial Engineering, Vol. 52, No.1, pp.143–161. Huang, M.J., Tsou, Y.L. and Lee, S.C. (2006) ‘Integrating fuzzy data mining and fuzzy artificial neural networks for discovering implicit knowledge’, Knowledge-Based Systems, Vol. 19, No. 6, pp.396–403. Kosko, B. (1992) Neural Networks and Fuzzy Systems. A Dynamical Systems Approach to Machine Intelligence, 1st ed., Prentice-Hall, Englewood Cliffs, NJ. Liu, D.R. and Shih, Y.Y. (2005) ‘Integrating AHP and data mining for product recommendation based on customer lifetime value’, Information and Management, Vol. 42, No. 3, pp.387–400. MATLAB Builder JA (for Java Language) Available at www.mathworks.com/products/javabuilder. Mehrabad, M.S. and Brojeny, M.F. (2007) ‘The development of an expert system for effective selection and appointment of the jobs applicants in human resource management’, Computers & Industrial Engineering, Vol. 53, No. 2, pp.306–312. Padhy, N.P. (2005) Artificial Intelligence and Intelligent Systems, pp.1–632, Oxford University Press, New Delhi. Petrovic-Lazarevic, S. (2001) ‘Personnel selection fuzzy model’, International Transactions in Operational Research, Vol. 8, No. 1, pp.89–105. Ranjan, J., Goyal, D.P. and Ahson, S.I. (2008) ‘Data mining techniques for better decisions in human resource management systems’, International Journal of Business Information Systems, Vol. 3, No. 5, pp.464–481. Saaty, T.L. (1995) Decision Making for Leaders: The Analytic Hierarchy Process for Decision in a Complex World, RWA Publications, Pittsburgh, PA. Salim, M.D., Villavicencio, A. and Timmerman, M.A. (2002) ‘A method for evaluating expert system shells for classroom instruction’, Journal of Industrial Technology, Vol. 19, No. 1, pp.1–11. Schneider, M., Lagholz, G., Kandel, A. and Chew, G. (1996) Fuzzy Expert System Tools, 3rd ed., John Wiley & Sons, New York, NY. Scholl, A., Manthey, L., Helm, R., and Steiner, M. (2005) ‘Solving multi-attribute design problems with analytic hierarchy process and conjoint analysis: an empirical comparison’, European Journal of Operational Research, Vol. 164, No 3, pp.760–777. Wang, L.X. and Mendel, J.M. (1992) ‘Generating fuzzy rules by learning from examples’, IEEE Transactions on Systems, Man, and Cybernetics, Vol. 22, No. 6, pp.1414–1427. 462 J.O. Daramola et al. Wei-Shen, T. and Chung-Chian, H. (2006) ‘A realistic personnel selection tool based on fuzzy data mining’, Proceedings of Joint Conference on Information Sciences, available at www.atlantis- ress.com/php/download_paper?id=46. Werner, J.M. (2000) ‘Implications of OCB and contextual performance for human resource management’, Human Resource Management Review, Vol. 10, No. 1, pp.3–24. Xu, Z., Gao, K. and Khoshgoftaar, T.M. (2005) ‘Application of fuzzy expert system in test case selection for system regression test’, IEEE International Conference on Information Reuse and Integration, IRI– 2005, Nos. 15–17, pp.120–125. Zadeh, L.A. (1965) ‘Fuzzy sets’, Information and Control, Vol. 8, No. 3, pp.338–353. Zadeh, L.A. (1975) ‘The concept of a linguistic variable and its application to approximate reasoning’, Information Science, Vol. 8, No. 3, pp.199–249. 
work_r6qxnxypgjeohca7vm2tslrnou	----	Fiabilitate si Durabilitate - Fiability & Durability No 1/ 2018 Editura “Academica Brâncuşi” , Târgu Jiu, ISSN 1844 – 640X 246 USING EXPERT SYSTEMS TO DESIGN MECHANISMS Ph. D. Student Anca-Olimpia MOSNEAGA (TALAAT-HAMID), Ph. D. Iosif TEMPEA, ―Sfantul Sava‖ National College, 23 General H. M. Berthelot St., Bucharest, Romania University POLITEHNICA Bucharest, 313 Spl. Independentei, Bucharest, Romania e-mail: iosiftempea@yahoo.com, ancamtal@yahoo.com Abstract: Expert systems, as a component of Artificial Intelligence. are used on an increasingly wider scale to solve certain problems in various fields. In this paper, we are describing the structure of an expert system used to design mechanisms. Keywords: mechanism, database, expert system, knowledge acquisition, inference mechanism. 1. Introduction Expert systems are a distinct category of informatics systems incorporating artificial intelligence that can simulate the reasoning of human experts using computer programmes and databases [1]. Just like any other programme, expert systems have an algorithmic structure, but they are open to improvement by developing a knowledge database and a set of inference rules [3]. Since the system is capable of acquiring new knowledge by including the results of the inferences made in its own knowledge database, the solutions provided by the system for the same problem, may vary, due to the new knowledge acquired. The experts system‘s capacity to respond to a requirement within a certain time interval that has to be as short as possible, depends on the size of its knowledge database and the way this database has been structure. But, the larger the knowledge database, the more accurate the results obtained. Also, the correctness of the inference mechanism, the accuracy with which the knowledge is combined in order to issue a solution, shall determine the extent to which the result of an inference can help solve the initial problem. Thus, there must be a balance between the size of the knowledge database + the knowledge acquired and the speed in finding a solution to the problem. 2. Expert system’s structure In order for an expert system to function, one must go through the following stages:  Knowledge acquisition  Knowledge representation  Knowledge processing  Displaying the results According to the stages listed above, an expert system shall comprise the following modules: [2]  Knowledge acquisition module  Module database  Facts database  Inference mechanism mailto:iosiftempea@yahoo.com mailto:ancamtal@yahoo.com Fiabilitate si Durabilitate - Fiability & Durability No 1/ 2018 Editura “Academica Brâncuşi” , Târgu Jiu, ISSN 1844 – 640X 247  Result explaining module  User interface 2.1. Knowledge acquisition module The core element of an expert system is the body of all specialised knowledge accumulated in the knowledge database. In order to acquire knowledge, one must set up some teams made of knowledge engineers and human experts in the field of mechanisms design. The knowledge engineers shall design the knowledge database, they shall build the knowledge acquisition module and they shall update the knowledge database. The human experts shall provide factual knowledge based on specialised books as well as heuristic knowledge based on the good practice principles learned during their working experience in the field. Fig. 1 Elements of a knowledge acquisition module 2.2. Knowledge database and facts database The body of all knowledge used to determine any solution represents the knowledge database. Following the acquisition of such knowledge, the knowledge engineer reflects this knowledge by designing the knowledge database as a set of data structures. In case of an expert system using inference mechanisms based on production rules, the knowledge database will be made of rules and facts database. In the rules database, one will store the rules that will be usable to solve the initial problem. [5] The facts database will comprise the problem‘s initial data, as well as intermediary data obtained while applying the production rules, via the inference mechanism. One fact will be described using three characteristics:  object - a mechanism component found in a certain state that can be deduced.  state – the state in which the object is found or the object‘s relationship with other components of the mechanism.  quantification – a value indicating the size of the object‘s state. The intermediary data represents the rules‘ snapshots. The use of these snapshots depends on the rules application priority and the conflict resolution system. Thus, the intermediary data is arranged in the facts database according to its priority. Fiabilitate si Durabilitate - Fiability & Durability No 1/ 2018 Editura “Academica Brâncuşi” , Târgu Jiu, ISSN 1844 – 640X 248 2.3. Inference mechanism The inference engine is a programme applying the knowledge stored in the knowledge database on the facts stored in the facts database, in order to generate new facts or a solution to the problem, by induction or, in order to confirm or infirm a hypothesis by deduction. . Fig. 2 Block sketch of an inference mechanism 2.4. Result explanation method The stages the inference mechanism goes through, the rules applied during each of these stages, shall be displayed in front of the user via the result explanation module. This way, any eventual errors within the knowledge database shall be discovered, as well as any eventual incorrect or wrongly applied production rules. 2.5. User Interface The user interface is used to insert the input data, to display the results and the explanations related to the way the expert system has reached its current solution. It is important for the interface to be as friendly as possible. Fiabilitate si Durabilitate - Fiability & Durability No 1/ 2018 Editura “Academica Brâncuşi” , Târgu Jiu, ISSN 1844 – 640X 249 3. Knowledge representation The knowledge that must be represented in the expert system‘s knowledge database consists of the data and information received by the knowledge engineer from the human expert, via the specialised literature [2]. The knowledge representation methods are specific to the chosen programming language and the inference mechanism. The procedural methods used to represent the knowledge are based on algorithms and they may be: production rules, decisional trees and neuronal networks. Out of the declarative methods, the most widespread is the frame method [6]. 3.1. Production rules The knowledge representation method based on production rules is similar to human reasoning, relying on the cause-effect principle. The expert system‘s knowledge base shall be built on the factual and heuristic knowledge held by the human expert and it shall be represented as a set of type [4] rules: IF condition THEN result The result may be:  a set of parameters representing a new fact set to be added to the facts database [4]  the display of a message requiring the introduction of new parameters  adding a new rule in the rules stack/queue . The production rules must simulate human reasoning, carrying out the same inferences based on facts. A deductive expert systems, will start from the initial data (facts), then the rules in the knowledge database will applied one by one in order to determine a result. An inductive expert system will start from the result (purpose) and it will look for rules that can be applied in order to demonstrate the proposed result. 3.2. Decisional trees If the knowledge is represented using decisional trees, one shall build the space for the states that the mechanism can be found in and, in order to determine the solution to the problem, one shall use graphs and graph analysis strategies. The graph nodes will be the mechanisms‘ states while the graph curves will be the mechanism‘s transit from one state to another, following an action represented by logic operators. In this case, one will also be able to calculate the minimum cost solution, by attaching an application cost to each operator. The search strategies will be in depth or in wide searches, depending on the problem requirements. 3.3. Frames This knowledge representation method is based on sketches comprising declarative and procedural knowledge. a frame is a structured object comprising information on the object, rules that can be applicable to the object and the way such rules are applied. Frames are an object oriented formalism [6] the connection between various objects being made by applying a method (rule). In expert systems based on frames, the facts are correlated, synthesised band structured in distinct, ranked entities, which leads to simplified algorithms, due to inheritance and to an easy way to add new knowledge. [7] Fiabilitate si Durabilitate - Fiability & Durability No 1/ 2018 Editura “Academica Brâncuşi” , Târgu Jiu, ISSN 1844 – 640X 250 4. Conclusions An expert system is based on the factual and heuristic knowledge held by certain human experts and it is made of a computer programme that uses its knowledge database in order to simulate human reasoning. The efficiency invested in designing the knowledge database and the informatics algorithm created, as well as the way knowledge is represented, shall determine the size and the time needed by an expert system to solve a problem. The development of integrated systems such as CAD/CAE/CAM, as well as the ever increasing volume of data processed, led to the creation of intelligent design systems based on computational algorithms and geometric modelling, as well as the algorithms used in the knowledge representation methods.[7] References 1. Zaharie, Dorin, “Sisteme expert”, (Expert System) DUAL TECH Publishing House, Bucharest, 2005, page 9-15 2. Poole David, Mackworth Alan, Goebel Randz, “Computational Intelligence a logical approach”, Oxford University Press, Oxford, New York, 1998, pag. 9-15, 200-220 3. Carstoiu, D., I., “Sisteme expert” (Expert System) , ALL Publishing House, Bucharest, 1994 page 16-26 4. Gavrilas, Mihai, “Note de curs”, (Course Notes) Universitatea Tehnica ―Gh.Asachi‖, Iasi, http://iota.ee.tuiasi.ro/~mgavril/Simpe/L2.htm 5. Zaharie Dorin, Nastase Pavel, “ Sisteme expert de gestiune”, (Management Expert Systems), Romcart Publishing House, Bucharest, 1993, page 60-87 6. ***, “Note de curs, Inteligenta artificiala si retele neuronale” Universitatea Politehnica Timisoara (Course Notes, Artificial Intelligence and Neuronal Networks) http://www.mpt.upt.ro/doc/curs/gp/Sisteme_inteligente_in_electrotehnica/Inteligenta_artificia la_si_Retele_neuronale_cap1.pdf 7. Butila, Eugen, Valentin, “Rezumat teza de doctorat” (PhD Thesis Summary) Brasov, 2009, pag.6-9 http://www.rrv.ro/rrv/rezumate_teze/2009.09.15_rez_butila_eugen_valentin.pdf http://iota.ee.tuiasi.ro/~mgavril/Simpe/L2.htm 
work_rbflc7gdy5hy5jjvjpxbihcnxm	----	© 1985 Thomas Blair DeGreve A WORKPLACE DESIGN EXPERT SYSTEM by THOMAS BLAIR DeGREVE, B.S. in I.E. A THESIS IN INDUSTRIAL ENGINEERING Submitted to the Graduate Faculty of Texas Tech University in Partial Fulfillment of the Requirements for the Degree of MASTER OF SCIENCE IN INDUSTRIAL ENGINEERING Approved C h a i r m a n of tzhe C o m m i t t e e /^ ^̂ Ŵ̂ LJJAJM-^UV^ Inn • h i OV''^^ Deaiy of the Graduate School December, 1985 1 • / f ^ ^ TABLE OF CONTENTS LIST OF TABLES iv LIST OF FIGURES v I. INTRODUCTION 1 Scope 1 Review of Previous Research 4 Expert Systems 4 M^orkplace Design 10 II. SYSTEM CONCEPTUALIZATION 15 System Components 15 The Context Tree 19 System Prototype 23 III. KNOWLEDGE ENGINEERING 24 Knowledge Acquisition 24 Knowledge Representation 26 IV. DEVELOPMENT OF THE SYSTEM 40 The Development Engine 40 Description of the Data 43 Construction of the Knowledge Base . . . . 46 The Inference Engine 47 V. CONCLUSION 51 LIST OF REFERENCES 54 1 1 APPENDIX A. DEFINITION OF KNOWLEDGE BASE PROPERTIES . . . . 56 B. LISTING OF THE KNOWLEDGE BASE 59 C. SAMPLE CONSULTATIONS 95 D. DEFINITION OF ANTHROPOMETRIC MEASUREMENTS . . . 108 1 1 1 LIST OF TABLES Table 1 : An Overview of the Key Events in the History of Artificial Intelligence Table 2: Ten Generic Categories of Expert Systems . . . 11 Table 3: Rules for Determining the Workstation Type . . 28 Table 4: Corrected Anthropometric Measurements . . . . 45 IV LIST OF FIGURES Figure 1 Figure 2 Figure 3 The Context Tree The Architecture of an Expert System . . . 20 41 Major Categories of Search Strategies used by Inference Engines 48 Figure 4: Sketch of Anthropometric Measurements . . . . 109 CHAPTER I INTRODUCTION Scope The field of ergonomics is rapidly becoming a key area of interest to employers who are concerned with providing a comfortable, safe, and pleasant working area for their employees as well as for themselves. The interest in apply- ing ergonomic principles to industrial workplaces is most likely a result of recent correlations made between the design of a workplace and employee health or productivity (Tichauer 1975, Konz 1979, Grandjean 1982). However most companies do not possess the expertise necessary for apply- ing ergonomic principles to the design of workplaces. Even with the abundance of reference material available, most facility engineers find it difficult to develop a proficien- cy in defining the work classification, analyzing the task requirements, quantifying the work force population, and finally coming up with an ergonomically sound workplace design. While theoretical research needs to continue, a practical method of putting the present knowledge to use is also important. The intention of this research is to develop a computer program called an expert system which will provide practical proficiency or expertise for the workplace designer. Expert system technology is an application of a branch 1 of artificial intelligence research concerned with develop- ing programs that use symbolic knowledge to simulate the behavior of human experts. Professor Edward Feigenbaum (1977) of Stanford University, one of the leading research- ers in expert systems, has defined an expert system as: ... an intelligent computer program that uses knowledge and inference procedures to solve prob- lems that are difficult enough to require significant human expertise for their solution. Knowledge necessary to perform at such a level, plus the inference procedures used, can be thought of as a model of the expertise of the best practi- tioners of the field. The knowledge of an expert system consists of facts and heuristics. The "facts" constitute a body of information that is widely shared, public- ly available, and generally agreed upon by experts in a field. The "heuristics" are mostly private, little discussed rules of good judgment (rules of plausible reasoning, rules of good guessing) that characterize expert-level decision making in the field. The performance level of an expert system is primarily a function of the size and the quali- ty of a knowledge base it possesses. Feigenbaum calls those who build knowledge-based expert systems "knowledge engineers" and refers to their technology as "knowledge engineering." Expert systems are knowledge- intensive computer programs with the following dimensions: expertise, symbol manipulation, uncertainty, complexity, reasoning, and explanation. They use rules of thumb, or heuristics, to focus on .the key aspects of a particular problem, manipulate symbolic descriptions of the problem, and apply reasoning to the knowledge they have been given concerning the problem, to reach a conclusion. They often consider a number of competing hypotheses simultaneously, and frequently make tentative recommendations or assign weights to alternatives. Current expert systems are confined to well- circumscribed tasks. They are not able to reason over a broad field of expertise. They cannot reason from axioms or general theories. They do not learn and, thus they are limited to using the specific facts and heuristics that they were "taught" by a human expert. They lack common sense, they cannot reason by analogy, and their performance deter- iorates rapidly when problems extend beyond the narrow task they were designed to perform. On the other hand, knowledge systems do not display biased judgments, nor do they jump to conclusions. They always attend to details, and they always systematically consider all of the possible alternatives (Harmon and King 1985). It is reasonable, then, to consider developing an expert system to assist facility engineers in the design of industrial workplaces. Workplace design is a problem domain relatively narrow enough, yet unstructured and uncertain enough to merit development of a dedicated expert system. Workplace design literature offers many guidelines and prin- ciples to the potential designer but there are no clear-cut procedures designed for direct application in an industrial setting, A workplace design expert system has the potential to utilize guidelines from the literature, borrow experience from a human expert, and apply this knowledge in a practical sense. The progress of this project was aided by reference material (Hertzberg 1972, Ayoub 1973, Konz 1979, Grandjean 1982) and the advice of Professor M. M. Ayoub, an experi- enced workplace design expert. Review of Previous Research In order to proceed with the development of a workplace design expert system, an extensive review of previous research in the areas of both expert systems and workplace design is appropriate. The following sections will summar- ize previous research in these two areas, respectively. Expert Systems As World War II ended, separate groups of British and American scientists were working to develop what is now called a computer. Each group wanted to create an electro- nic machine that could be guided by a stored program of directions and made to carry out complex numerical computa- tions. The principal British scientist, Alan Turing, argued that such a general-purpose machine would have many differ- ent uses. Turing argued that the fundamental instructions given to such a machine ought to be based on logical opera- tors, such as "and," "or," and "not." These general operators could then be used to assemble the more special- ized numerical operators needed for arithmetic calculations. Moreover, programs based on logical operators would be capable of manipulating any type of symbolic material that one might want to work with, including statements in ordi- nary language. The more practical American scientists, knowing the machine was going to be expensive to build and assuming they would not build very many, decided against using logical operators. They were confident they were building a machine that would do only arithmetic calculations; therefore, they chose to use numerical operators, such as "+", "-", and ">". This decision, which the British subsequently followed as well, resulted in large computers that are essentially very fast calculating machines. Until very recently, the decision to use numerical operators seemed like a reasonable one to most people involved with computers. However, a small group of computer scientists continued to explore the ability of computers to manipulate non-numerical symbols. At the same time, psycho- logists concerned with human problem solving sought to develop computer programs that would simulate human behav- ior. Over the years, these people have formed the inter- disciplinary subfield of computer science called artificial intelligence (AI). AI researchers are concerned with developing computer systems that produce results that would normally be associated with human intelligence (Harmon and King 1985). The applied side of artificial intelligence research includes the areas of natural language processing, robotics, and expert systems. The latter has enjoyed much success in recent years and initiated a rush to find practical applica- tions for this new technology. Table 1 presents an overview of the key events in the history of artificial intelligence with particular attention focused on expert systems research, The earliest acknowledged expert system, DENDRAL, is a chemistry expert system designed to examine a spectroscopic analysis of an unknown molecule and predict the molecular structures that could account for that particular analysis. In 1964, Joshua Lederberg, a Nobel prize-winning chemist, developed the DENDRAL algorithm. In 1965, with the aid of Edward Feigenbaum and Bruce Buchanan, the DENDRAL expert system was programmed directly in the LISP computer language (Lindsay, Buchanan, Feigenbaum and Lederberg 1980). Probably the most well-known expert system, MYCIN, is a computer program designed to provide attending physicians with advice comparable to that which they would otherwise get from a consulting physician specializing in bacteremia and meningitis infections. MYCIN was the first large expert system to perform at the level of a human expert and to pro- vide users with an explanation of its reasoning. MYCIN was developed at Stanford University in the mid-1970's (Buchanan Table 1. An Overview of the Key Event in the History of Artificial Intelligence (from Harmon and King 1985) P E R I O D Pre-World War II Pre-AI The formative years, 1955-1960 Initiation of AI The years of development, 1961-1970 Search for general problem solvers The years of specialization, 1971-1980 Discovery of knowledge based systems Rush to applications 1981- International competition and commercialization KEV E U E N T S Formal. Logic Cognitive Psychology Comjp.uters Developed N. Wiener, "Cybernetics" A.M. .Turing, ""Computing And Intelligence^ Information processing Lansuase I Dartmouth Seminar on AI,l95o General Problem Solver (GPS) LISP Heuristics Robotics Chess programs Dendral (Stanford) MYCIN (Stanford) Hearsay II (Carnegie-Mellon) MACSYMA (MIT) EMYCIN (Stanford) PROLOG PROSPECTOR (SRI) Japan's Fifth-Generation Project E.Fergenbaum, "The Fifth Generation" INTELLECT (AIC) 8 and Shortliffe 1984). A line of expert system development beginning with SAINT (Slagle 1961) culminated with MACSYMA. The design for MACSYMA was originally laid out in 1968 by Carol Engleman, William Martin, and Joel Moses at MIT and has been under continual development since 1969. MACSYMA performs differ- ential and integral calculus symbolically and excels at simplifying symbolic expressions. It incorporates hundreds of rules, each of which expresses one way to transform an expression into an equivalent. The solution to any problem requires finding a chain of rules that transforms the origi- nal expression into one that is suitably simplified (Martin and Fateman 1971). HEARSAY was developed to demonstrate the possibility of a speech-understanding system. Development began in the late 1960's at Carnegie-Mellon University. By the time the HEARSAY II project was finished in 1975, a system had been developed that could deal with a limited amount of spoken grammar and a vocabulary of about 1,000 words. HEARSAY II demonstrated the clear superiority of symbolic, heuristic methods over statistical methods in dealing with problems involving meaning (Erman et al. 1980), PROSPECTOR was developed in the late 1970's at Stanford Research Institute, International by a team" that included Peter Hart, Richard Duda, R. Reboh, K. Konolige, P. Barrett, and M. Einandi. The development of PROSPECTOR was funded by the U.S. Geological Survey and by the National Science Foun- dation. PROSPECTOR was designed to provide consultation to geologists in the early stages of investigating a site for ore-grade deposits. The program informs users of possible data interpretations and identifies additional geological observations that would be valuable to reach a more definite conclusion (Duda and Reboh 1984). PUFF was built in 1979 using the EMYCIN knowledge sys- tem building tool at Stanford University. PUFF was designed to interpret measurements from respiratory tests admini- stered to patients in a pulmonary function laboratory. PUFF interprets a set of test results and produces a written statement that includes a set of interpretations and a diag- nosis for the patient (Aikens, 1984). Recent systems built since 1980 are mostly commercial systems and include the following: XCON (RI) and XSEL— knowledge systems that help configure computer systems for clients, GENESIS--a package of systems that help molecular geneticists plan DNA experiments, DELTA/CATS--an expert sys- tem that aids locomotive maintenance personnel, and DRILLING ADVISOR--an expert system that helps oil rig supervisors to solve problems. There are many tools available commercially to help build expert systems. Most tools provide an inferencing mechanism along with the means for the user to enter his knowledge base. Some tools provide help screens and debugg- 10 ing utilities. A list of some of the current expert system development tools include: EXPERT, KES, 0PS5, S.l, TIMM, ART, KEE, LOOPS, ES/P ADVISOR, Expert Ease, INSIGHT, M.l, and Personal Consultant, Most expert systems fall into ten generic categories of knowledge engineering applications. These ten categories, presented in Table 2, are based on the types of problems that different expert systems address. The subject system of this thesis falls into the Design-type category. v̂ Workplace Design Interest in industrial workplace design can be traced back to the 1880's when Frederick Taylor and Frank and Lillian Gilbreth began their work with time and motion studies. The primary objective of their work involved mea- suring and improving worker productivity. Consequently, they recognized the importance of workplace design, but they did not recognize its full importance. Das and Grady (1983) recognize the following five objectives in designing indus- trial workplaces: /I. Measure and improve worker productivity; 2. Enhance worker satisfaction and job attitudes; 3. Reduce operator fatigue; 4. Improve working environments; 5. Minimize worker safety hazards.\ 11 TABLE 2: Ten Generic Categories of Expert Systems (from Hayes-Roth et al. 1983) CRTEGORV INTERPRETATION PREDICTION DIAGNOSES DESIGN PLANNING MONITORING DEBUGGING REPAIR INSTRUCTION CONTROL PROBLEMS THEV RDDRESS Inferring situation descriptions from sensor data Inferring likely consequences of given situations Inferring system malfunctions from observables Configuring objects under constraints Designing actions Comparing observations to plan vulnerabilities Prescribing remedies for malfunctions Executing a plan to adminiter a prescribed remedy Diagnose, debug, and repair student behavior Interpret, pred., repair, and monit. syst. behavior 12 Maynard (1934) proposed two general concepts of indus- trial workplace design. The first concept was to reduce all motions used in the performance of the task to the lowest possible class. He defined five general classes ranging from the highest: finger, wrist, forearm, upper arm, and body motion to the lowest class which involved only finger motion. Maynard's second concept was to define the normal and maximum work areas in the horizontal and vertical planes. The normal working area in the horizontal plane was determined by arcs drawn with a sweeping motion of the arms. Only the forearms were extended, and the upper arms hung at the sides of the body. The maximum horizontal work area was determined in a similar fashion with the arms fully extend- ed. The normal work area in the vertical plane included the area defined by arcs in the sagittal plane drawn by the forearms hinged at the elbows. Similarly, the maximum ver- tical work area was determined by drawing arcs with the arms fully extended and hinged at the shoulders. The workplace layouts provided by Maynard were dimensionless and, there- fore, of little use to a designer until Barnes (1940) applied dimensions to Maynard's layouts. Woodson and Conover (1964) recommended that anthropo- m etric measurements of the largest worker should be used to determine workplace clearances, while those of the smallest worker should be used to determine limits of reach. They also cautioned the designer to make allowances for clothing, 13 which add to the clearance requirements and cause restric- tion of movement. Konz (1967) conducted an experiment to investigate the effects of work surface heights on performance. Konz con- cluded that the best working height for a standing operator is about 2.5 cm below the elbow. Ayoub (1973) recommended work surface heights for the standing and seated operator as a function of the work classification. He proposed three classifications of standing work and four classifications of seated work. Grandjean (1982) has made similar work surface height recommendations. V In order to design a successful industrial workplace, Ayoub (1973) encouraged designers to follow ten ergonomic guidelines: 1. Reduce the static component of work. 2. Do not overload the muscular system. 3. Strive to achieve the best mechanical advantage. 4. Eliminate extreme positions of the joints. 5. Reduce unusual and stressful postures. 6. Maintain a good seating arrangement. 7. Permit change of posture on the job. 8. Accommodate large operators and give them enough space. 9. Train the operator to use the facilities. 10. Match operator capacities with the job demands. In addition, Ayoub et al. (1982) examined the situa- tions in which the operator should sit at the workplace, 14 stand at the workplace, or alternately sit and stand at the workplace. A. Seating is recommended at the workplace when: (1) All items needed in the short-term task cycle can be easily supplied and handled within the seated workspace. (2) The items being handled do not require the hands to work at an average level of more than 15 cm above the work surface. (3) No large forces are required, such as handling weight greater than 4.5 kg. (4) Fine assembly or writing is done for a majori- ty of the shift. B. Standing is recommended at the workplace when: (1) The workplace does not have knee clearance for a seated operator. (2) Objects weighting more than 4.5 kg are handled. (3) High, low, or extended reaches, such as those in front of the body, are required frequently. (4) Operations are physically separated and require frequent movement over a large area. (5) Downward forces must be exerted, as in packing and wrapping operations. C. Sit/Stand workplaces are considered in these instances: (1) Repetitive operations are done with frequent reaches more than 41 cm forward and/or more than 15 cm above the work surface. (2) Multiple tasks are performed, some best done sitting and others best done standing. CHAPTER II SYSTEM CONCEPTUALIZATION In order to conceptualize a complex, knowledge-based workplace design system one must establish many design goals and ideals. In the case of the subject workplace design system, goals and ideas were chosen to outline the system's potential functionality, usability, and accuracy. Develop- ing the ideal workplace design system would involve hardware, expertise, time, and resources beyond the scope of a master's thesis project; however, the aim of this project is to conceptualize an ideal system and outline the proce- dures required to develop the system. After performing the essential research and ground work required to begin development of the system, a subset of the conceptualized system was selected for prototyping. The working prototype can be tested and can serve as a basis for further research and development toward the completion of the ideal system. System Components The conceptualized workplace design system referred to above consists of three main components. These components are independent in a sense that they are not all necessary to build a workplace design system; yet the integration of all three components will provide the optimum environment for development of the ideal system. The three system com- 15 16 ponents are listed below. 1. Intelligent Software 2. Symbolic Processing Hardware 3. Interactive Graphic Interface The intelligent software component of the system ensures that, during the design phase, the user adheres to accepted ergonomic concepts and principles. The software's functionality and interactiveness are such that even a novice unfamiliar with workplace design criteria can use the system and achieve optimum results. The inherent intelli- gence of the software aids the user in solving problems requiring human judgment, experience, reasoning, and/or expertise. The intelligence of the software lies in the set of knowledge-based rules contained in the system. These rules allow the system to reason and to "think" like an expert in the field of workplace design. The system can interact with the user, make conclusions during the consultation session, and respond by providing an ergonomically sound workplace design. In addition, the system can justify the design for the user by recalling the internal knowledge that the con- clusions were based on. The software acts as an aid to the workplace design engineer. By modeling the engineer's particular workplace characteristics and associated constraints, the software can 17 provide expert advice early in the design phase. The advice can range from suggestions to increase mechanical advantage to an actual layout drawing of the proposed workplace facil- ity. Use of the system will ensure that ergonomic problems do not arise during the workplace implementation phase. This intelligent consulting service can not only save design time, but it can save money by eliminating design errors. The second component of the ideal workplace design sys- tem is symbolic processing hardware. This type of computer hardware enables optimum, efficient functioning of intelli- gent software coded in symbolic languages such as LISP or PROLOG. Conventional computers excel at problems that can be expressed in numerical terms and which lend themselves to repetitive, algorithmic solutions. However, traditional computing has not been effective in dealing with unstruc- tured problems, interpreting information, using rules of thumb gained by experience, or dealing with uncertain or incomplete information. Symbolic processing is a technique that has been developed to address these problems. It refers to the utilization by computers of information and knowledge represented by symbols, analogous to the way humans reason with knowledge they possess. The special features that set symbolic processing com- puters apart from conventional computers typically include a dedicated processor designed to optimize execution of sym- bolic code, large virtual address space, large amounts of 18 physical memory, a high-resolution graphics display, and a high-performance mass storage device. Symbolic* processing hardware might be customized for the ideal workplace design system by reducing some of the knowledge sets to microchips, which might be inserted in the system's hardware. This would increase the speed and efficiency of the system. At present, no computer system exists that could be easily customized as a dedicated workplace design system. However, there are now computer systems available that util- ize symbolic processing technology. Some of these systems include the Texas Instruments Explorer, the Xerox 1100, the Symbolics 3600, the LMI Lambda, and the Apollo system. An interactive graphic interface is the third component of the ideal workplace design system. Conceptually, this device enables a graphic interactivity between the system and the user. The primary graphic output of the system is a dimensioned drawing of the workplace layout. A side view, a top view, and a three-dimensional representation showing how the operator fits into the workplace facility is included. After the system has provided the initial design, the user can customize or modify the design, based on any physical constraints he may be confronting. If any significant changes are made to the design, the system responds by high- lighting the problem areas such as reduced thigh clearance or reduced knee clearance. The severity of the problem is noted graphically and any consequential reduction in the 19 accommodated work force population will be indicated. The Context Tree The conceptualized ideal system can be pictured as a tree of interlocking knowledge sets or contexts. Texas Instruments (1985) defines a context as a structural unit used to separate the information in a knowledge base accord- ing to main concepts or parts. If a knowledge base is thought of as a file cabinet full of information pertaining to one subject area or domain, the contexts are the drawers that are organized according to major divisions of informa- tion. Contexts have parameters associated with them that store pieces of information used in the knowledge base. A special set of parameters, called goals, is attached to each context. After the values of all the goal parameters have been found, the context is exited. The context tree for the conceptualized workplace design system is presented in Figure 1. In a context tree a parent context is the context one level higher than the child context. The parent context of the subcontext entitled Physiological Design is the subcontext entitled Biomechanical Design. The root context. Dimensional Design, has no parent. The context called Other Ergonomic Consider- ations has no child context but has the root context as its parent. A context's ancesto-rs include all the contexts that are in the direct line to the root context and including the 20 D I M E N S i O N R L D E S I G N B I O M E C H R N I C R L D E S I G N OTHER E R G O N O M I C C O N S I D E R R T I O N S P H V S I O L O G I C R L D E S I G N FIGURE 1 . The C o n t e x t T r e e 21 root context. A child context automatically inherits all the parameters of its ancestor contexts. When a context inherits the parameters of another context, it means that it can share the use of those parameters with the other con- text . The root context. Dimensional Design, has many goal parameters associated with it. These include the type of workstation, the work surface height range, the chair height range, the footrest height range, the normal and maximum vertical reach, and the normal and maximum horizontal reach. In order to ascertain these values the system uses informa- tion such as the defined work force population, anthropometric measurements from that population, the class- ification of the work to be performed at the workplace, the physical constraints of the workplace, and the mathematical model developed by the author. The Biomechanical Design context more closely examines the dynamic components of the work. Factors such as the number of objects to be handled, object size, object weight, frequency of lift or movement, vertical or horizontal dis- tance of movement, and distance between the object and the body are analyzed in this context. Goals for this context include the acceptable production rate at the workplace, the efficient arrangement of objects within the workplace, and any restrictions on the work force pertaining to operator fitness, muscular strength, or age. (An expert system that 22 analyses manual lifting tasks has been developed by Karwowski et al. [1985].) The purpose of this context is to provide optimum mechanical advantage for the operator and to ensure that he is in no immediate danger of over-exertion, injury, or disability. The Physiological Design context examines the energy requirements of the task to be performed at the workplace. Based upon a description of the motion components of the task and information gained from ancestor contexts, this context calculates the net energy requirements of the task in kilocalories per minute. Several models (Garg et al. 1978, Asfour 1980) have been developed that enable this type of analysis. An evaluation of worker capacity is also included in this context. The user can rely on the internal physiological capacity data base (Astrand 1977) or he can enter data collected from his specific work force popula- tion. The goal parameters for this context include the expected endurance time at this task, the required resting time, and the recommended number of rest periods for an eight hour shift. The final context. Other Ergonomic Considerations, is designed to provide informational guidelines that should be followed during the workplace design and implementation phases. This context provides guidelines concerning design of the interface between the operator and the workplace. Advice about the machine controls, displays, warning indica- 23 tors, and other ergonomic and safety factors are presented. In addition, this context is knowledgable in national and international safety standards regarding the handling of toxic, hazardous, or radioactive materials; heat stress; cold stress; noise; lighting; and vibration. System Prototype After conceptualization of the ideal workplace design system, a subset of the system was selected for prototyping. Neither the equipment nor the expertise to customize symbol- ic processing hardware or to design an interactive graphic interface device was immediately available. Therefore, it was decided that development of the system's intelligent software would be pursued. This pursuit was aided by the availability of a Texas Instruments Portable Professional Computer and the Personal Consultant Expert System Develop- ment Tool. The system's root context. Dimensional Design (as shown in Figure 1), was the subset selected for proto- typing. Subsequent chapters of the thesis describe the data that was used, outline the knowledge engineering activities that were performed, and detail the development of the pro- totype knowledge base and the use of its inference engine. CHAPTER III KNOWLEDGE ENGINEERING Knowledge engineering is perhaps the most important phase of developing an expert system. Knowledge engineers acquire knowledge from a human expert and other sources and then embed that knowledge in an expert system. The process of knowledge engineering involves knowledge acquisition and knowledge representation. Knowledge is acquired through an extensive interview with a human expert to determine the thought processes he has when he solves a problem in the defined domain. Knowledge is represented by modeling those thought processes in the form of a knowledge base which in turn can be used by an inference engine to artificially reach the same solutions. This chapter describes the know- ledge engineering procedures that were used to develop a workplace design expert system. Knowledge Acquisition The primary source of knowledge was Professor Ayoub, an experienced workplace design expert. In order to acquire his knowledge, it was necessary to discuss his methods of approaching an industrial workplace design project. Inter- view questions were designed to establish the initial data that is required to begin a design project and what the out- put of such a project should consist of. In addition, example design projects were presented to Professor Ayoub so 24 25 that he could apply his expertise to these problems and reach satisfactory solutions. His knowledge and thought processes were examined in a step by step fashion as he made intermediate conclusions in solving the example projects. It is most helpful to have the human expert state his knowledge and thought processes in the form of IF/THEN statements. Consequently, the expert's knowledge can be acquired in a systematic, logical manner. If the expert's knowledge is acquired in this manner, it is relatively easy to represent his knowledge in a dynamic knowledge base that can be utilized by the inference engine. It was established that anthropometric measurements from the population to be accommodated at the workplace are the most important initial data required to begin a design. In addition, the classification of the work to be performed at the workplace is important, whether it is precision work, light work, heavy work, or VDT/keyboard operation. In order to establish what type of workstation to design (sit, stand, or sit/stand combination) it is necessary to find out if the operator is required to use foot controls, if he is required to cover a large work area, and to specify his reach requirements in terms of distance and frequency. Finally, it is important to decide what range of operators will be accommodated at the workplace. It is decided that a realis- tic goal of a workplace design expert system is to design for a range that would accommodate the U. S. industrial pop- 26 ulation from the 5th to the 95th percentiles, unless the design is restricted by physical constraints or specific anthropometric measurements entered by the user. Knowledge Representation Once the workplace design knowledge had been acquired it was necessary to represent the knowledge in a usable form. Thought processes were modeled by logical rules called modus ponens rules. The application of these rules is such that when A is known to be true and if a rule states, "If A, then B," it is valid to conclude that B is true. Stated differently, when we discover that the premis- es of a rule are true, we are entitled to believe the conclusions. Modus ponens is the basic inferencing strategy used by most expert systems. In addition, most expert systems apply these rules to knowledge represented as pbject-attribute-value (0-A-V) triplets. In the case of the workplace design expert sys- tem, the object is the root context, Dimensional Design. Attributes are equivalent to parameters. An example of a parameter might be Chair-Height. Values are the text or numerical denotations assigned to parameters. For example, the parameter, Chair-Height, might have the value 49.5 centimeters. 0-A-V triplets are used to store informa- tion in a knowledge base, while rules are used to infer conclusions from the knowledge base. 27 In order to represent the knowledge required to make a decision on what type of workstation to design, the set of rules presented in Table 3 were developed. The outcome of a rule depends on the value of the work classification para- meter, the foot controls parameter, and the large work area parameter. In general, the aim of this set of rules is to conclude the workstation type as a sit/stand combination unless restricted by one or more of the premise parameters. The knowledge required to determine the physical dimen- sions of the workplace is represented by another set of rules. In order to discuss these rules, the following para- meters must be defined. Goal Parameters: WSU - the upper work surface height; WSL - the lower work surface height; CHU - the upper chair height; CHL - the lower chair height; FRU - the upper footrest height; FRL - the lower footrest height; NVR - the normal vertical functional reach; MVR - the maximum acceptable vertical reach; NHR - the normal horizontal functional reach; MHR - the maximum acceptable horizontal reach; Other Parameters: SH0ULDER5 - the smallest standing shoulder height; 28 TABLE 3. Rules for Determining the Workstation Type IF: Work-Class Foot-Controls Precision yes Precision Precision Precision P Light-Work 1 Light-Work Light-Work Light-Work Heavy Work Heavy Work Heavy Work Heavy Work VDT/Keyboard VDT/Keyboard VDT/Keyboard VDT/Keyboard yes no no yes yes no no yes yes no no yes yes no no Lar^e-Area yes no no yes yes no no yes yes no no yes yes no no yes THEN: Station-Tvoe Sit/Stand Sit Sit/Stand Sit/Stand Sit/Stand Sit Sit/Stand Sit/Stand Sit/Stand Sit/Stand Stand Stand Sit/Stand Sit Sit Sit/Stand / / / / ^ / / / / / -^ 29 SH0ULDER95 - the largest standing shoulder height; / SITTING-SH0ULDER5 - the smallest dimension from buttocks to shoulder; B0DY-DEPTH95 - the largest body depth dimension; THIGH-WIDTH5 - the smallest thigh clearance dimension; THIGH-WIDTH95 - the largest thigh clearance dimension; F0REARM5 - the smallest forearm length; /̂ ARM5 - the smallest arm length; ^ ELB0W5 - the smallest standing elbow height; J ELB0W95 - the largest standing elbow height; SITTING-ELB0W5 - the smallest dimension from buttocks to elbow; SITTING-ELB0W95 - the largest dimension from buttocks to elbow; P0PLITEAL5 - the smallest popliteal height; P0PLITEAL95 - the largest popliteal height; CI - correction factor applies to a Sit work surface; C2 - correction factor applies to Stand or Sit/Stand '^ work surfaces. The aim of the following set of rules is to accommodate the entire workforce population range at the workplace. It must be kept in mind that an operator with the smallest measurement for one parameter does not necessarily have the smallest measurement for the other parameters. For example, an operator might have the smallest popliteal height but the largest body-depth dimension. Therefore, these rules were developed with this knowledge in mind. The first rule applies to all workplaces. 30 If: STATION - TYPE = SIT OR STAND OR SIT/STAND Then: NHR = F0REARM5 - 1/2 (B0DY-DEPTH95) MHR = ARM5 - 1/2 (B0DY-DEPTH95) The next group of rules apply to fully adjustable work- places in which the work surface height, chair height, and footrest height are all adjustable. If: STATION-TYPE = SIT Then: WSU = P0PLITEAL95 + SITTING-ELB0W95 + CI WSL = P0PLITEAL5 + SITTING-ELB0W5 + CI CHU = P0PLITEAL95 CHL = P0PLITEAL5 FRU = CHU - P0PLITEAL5 FRL = 0 NVR = CHL + SITTING-ELB0W5 + F0REARM5 MVR = CHL + SITTING-SH0ULDER5 + ARM5 I If: STATION-TYPE = STAND Then: WSU = ELB0W95 + C2 WSL = ELB0W5 + C2 CHU, CHL, FRU, FRL = NONE NVR = ELB0W5 + F0REARM5 MVR = SH0ULDER5 + ARM5 If: STATION-TYPE = SIT/STAND Then: WSU = ELB0W95 + C2 WSL = ELB0W5 + C2 CHU = ELB0W95 - SITTING-ELB0W5 31 CHL = ELB0W5 - SITTING-ELB0W95 FRU = CHU - P0PLITEAL5 FRL = 0 NVR = ELB0W5 + F0REARM5 MVR = SH0ULDER5 + ARM5 The next group of rules apply to workplaces in which a workplace that includes a chair with no height adjustment feature. In order to accommodate the entire range of opera- tors, the work surface height and the footrest height must be adjustable. Since the chair height is stationary, the CHU parameter will hold the value for the fixed chair height and the CHL parameter will be neglected. If: STATION-TYPE = SIT Then: CHU = P0PLITEAL95 CHL = NONE WSU = CHU + SITTING-ELB0W95 + CI WSL = CHU + SITTING-ELB0W5 + CI FRU = CHU - P0PLITEAL5 FRL = 0 NVR = CHU + SITTING-ELB0W5 + F0REARM5 MVR = CHU + SITTING-SH0ULDER5 + ARM5 If: STATION-TYPE = STAND Then: CHU, CHL = NONE WSU = ELB0W95 + C2 32 WSL = ELB0W5 + C2 FRU, FRL = NONE NVR = ELB0W5 + F0REARM5 MVR = SHOULDERS + ARM5 If: STATION-TYPE = SIT/STAND Then: CHU = ELB0W95 - SITTING-ELBOWS CHL = NONE WSU = CHU + SITTING-ELB0W95 + C2 WSL = CHU + SITTING-ELB0W5 + C2 FRU = CHU - P0PLITEAL5 FRL = 0 NVR = ELB0W5 + FOREARMS MVR = SHOULDERS + ARMS The next group of rules apply to workplaces in which the work surface height is fixed. These rules enable the design of a workplace that includes a work surface or table with no height adjustment feature. In order to accommodate the entire range of operators, the chair height and the footrest height must be adjustable. Since the work surface height is stationary, the WSU parameter will hold the value for the fixed work surface height, and the WSL parameter will be neglected. For the cases in which the workstation type is stand or sit/stand, the footrest dimensions refer to an adjustable platform on which the operator can adjust him- self to the fixed work surface. Determining workplace 33 dimensions for this group of rules is more difficult because the relative position of the elbow, buttocks, or foot must be solved for in relation to the fixed work surface. If: STATION-TYPE = SIT Then: WSU = P0PLITEAL9S + SITTING-ELB0W95 + CI WSL = NONE Equating at the elbow, CHU + SITTING-ELBOWS = WSU - CI and by substitution, CHU = P0PLITEAL9S + SITTING-ELB0W9S - SITTING-ELBOWS CHL = P0PLITEAL9S FRU = CHU - POPLITEALS FRL = 0 Equating at the elbow, NVR - FOREARMS = WSU - CI and by substitution, NVR = P0PLITEAL95 + SITTING-ELB0W95 + FOREARMS Equating at the buttocks, MVR - SITTING-SHOULDERS - ARMS = CHL and by substitution, MVR = P0PLITEAL9S + SITTING-SHOULDERS + ARMS If: STATION-TYPE = STAND Then: WSU = ELB0W9S + C2 WSL = NONE 34 CHU, CHL = NONE Equating at the foot, FRU = WSU - C2 - ELBOWS and by substitution, FRU = ELB0W9S - ELBOWS FRL = 0 Equating at the elbow, NVR - FOREARMS = WSU - C2 and by substitution, NVR = ELB0W95 + FOREARMS Equating at the foot, MVR - SHOULDERS - ARMS = WSU - C2 - ELB0W9S and by substitution, MVR = SHOULDERS + ARMS STATION-TYPE = SIT/STAND Then: WSU = ELB0W9S + 0 2 " ^ WSL = NONE Equating at the elbow, CHU + SITTING-ELBOWS = WSU - C2 and by substitution, .CHU = ELB0W9S - SITTING-ELBOWS Similarly, CHL = ELB0W95 - SITTING-ELB0W9S FRU = CHU - POPLITEALS FRL = 0 35 Equating at the elbow, NVR - FOREARMS = WSU - C2 and by substitution, NVR = ELB0W9S + FOREARMS Equating at the buttocks, MVR - SITTING-SHOULDERS - ARMS = CHL and by substitution, MVR = ELB0W9S - SITTING-ELB0W95 + SITTING-SHOULDERS + ARMS The next group of rules apply to workplaces in which all dimensions are fixed. These rules enable the design of a workplace that includes a work surface, chair, and foot- rest with no height adjustment features. Obviously, this type of design cannot accommodate the entire range of opera- tors. Specifically, a percentage of the workforce population will be excluded from the workplace due to thigh clearance restrictions. In addition, closer examination of the previous rules indicate that possible restrictions could result depending on the thickness of the work surface A set of rules that calculates the severity of the thigh clearance restriction will be presented after the following group of rules. Since the height dimensions of the follow- ing three rules are stationary, the WSU, CHU, and FRU parameters will hold the values for those fixed dimensions, and the WSL, CHL, and FRL parameters will be neglected. 36 If: STATION-TYPE = SIT Then: WSL, CHL, FRL = NONE WSU = P0PLITEAL95 + SITTING-ELB0W9S + CI CHU = WSU - CI - 1/2 (SITTING-ELBOWS + SITTING-ELB0W9S) FRU = CHU - 1/2 (POPLITEALS + P0PLITEAL95) NVR = CHU + SITTING-ELBOWS + FOREARMS MVR = CHU + SITTING-SHOULDERS + ARMS If: STATION-TYPE = STAND Then: WSL, CHL, FRL = NONE WSU = ELB0W9S + C2 CHU = NONE FRU = WSU - C2 - 1/2 (ELBOWS + ELB0W9S) NVR = FRU + ELBOWS + FOREARMS MVR = FRU + SHOULDERS + ARMS If: STATION-TYPE = SIT/STAND Then: WSL, CHL, FRL = NONE WSU = ELB0W95 + C2 CHU = WSU - C2 - 1/2 (SITTING-ELBOWS + SITTING-ELB0W9S) FRU = CHU - 1/2 (P0PLITEAL5 + P0PLITEAL95) NVR = CHU + SITTING-ELBOWS + FOREARMS MVR = CHU + SITTING-SHOULDERS + ARMS The correction factors, CI and C2, are used to modify the work surface height based on the classification of the 37 work to be performed at the workplace. For example, preci- sion work requires a work surface above the elbows and closer to the eyes, while heavy work requires a work surface below the elbows to allow exertion of greater forces. The rules that determine the CI and C2 values are listed below. If: WORK-CLASS = PRECISION-WORK Then: CI = +12 cm. C2 = +8 cm. If: WORK-CLASS = LIGHT-WORK Then: CI = +5 cm, C2 = -5 cm. If: WORK-CLASS = HEAVY-WORK Then: CI = 0 cm. C2 = -13 cm, If: WORK-CLASS = VDT/KEYBOARD-OPERATION Then: CI = 0 cm. C2 = -5 cm. As discussed previously, the completely stationary design and the other adjustable designs present possible thigh clearance restrictions. The next set of rules deter- m ines what portion of the work force population will be excluded from the workplace due to a thigh clearance problem by calculating the approximate reduction percentage. In order to make this estimation, it is assumed that the 38 distribution of thigh clearance measurements over the popu- lation range is normally distributed. Before these rules can be discussed, the following parameters must be defined, TABLE-THICKNESS - vertical thickness of the work surface; THIGH-TEST - determines the existence of a thigh clearance problem; THIGH-CUT - severity of the restriction in centimeters; THIGH-POPS - the smallest thigh width/popliteal combination; THIGH-P0P9S - the largest thigh width/popliteal combination; MU-HAT - the estimated distribution mean; SIGMA-HAT - the estimated distribution standard deviation; THIGH-CRITICAL - the calculated distribution Z value. If: DESIGN-TYPE = STATIONARY, and THIGH-WIDTH9S > (WSU - TABLE-THICKNESS) - CHU Then: THIGH-TEST = YES If: DESIGN-TYPE = FIXED-WORK-SURFACE, and STATION-TYPE = SIT/STAND, and THIGH-WIDTH9S > (WSU - TABLE-THICKNESS) - CHU Then: THIGH-TEST = YES If: THIGH-TEST = YES Then: THIGH-CUT = THIGH-WIDTH9S - (WSU - TABLE-THICKNESS - CHU) THIGH-POPS = THIGH-WIDTHS + POPLITEALS THIGH-P0P95 = THIGH-WIDTH95 + P0PLITEAL9S MU-HAT = 1/2 (THIGH-POPS + THIGH-P0P9S) 39 SIGMA-HAT = (THIGH-P0P95 - MU-HAT) / 1.645 THIGH-CRITICAL = THIGH-P0P9S - THIGH-CUT If: (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT <= 1.645 (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT > 1.28 Then: POPULATION-REDUCTION = < 5% If: (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT <= 1,28 (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT > 1.035 Then: POPULATION-REDUCTION = 5-10% If: (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT <= 1.035 (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT > .0675 Then: POPULATION-REDUCTION = 10-20% If: (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT <= 0.675 (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT > 0.385 Then: POPULATION-REDUCTION = 20-30% If: (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT <= 0.385 (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT > 0.125 Then: POPULATION-REDUCTION = 30-40% If: (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT <= 0.125 (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT > 0 Then: POPULATION-REDUCTION = 40-50% If: (THIGH-CRITICAL - MU-HAT) / SIGMA-HAT <= 0 Then: POPULATION-REDUCTION = >50% CHAPTER IV DEVELOPMENT OF THE SYSTEM Once the knowledge for the Dimensional Design context was acquired and represented in rule form, it was possible to begin development of the system prototype. A personal computer-based expert system building tool was used to assist in the development of the system. The tool chosen was the Personal Consultant Expert System Development Tool. Personal Consultant is implemented in the IQLISP computer language. In other words, the inference engine and the knowledge engineer/user interfaces are coded in the underly- ing IQLISP language. The tool does not offer a compiler to make completed programs faster to operate. However, the existing configuration is quite adequate for the development of an expert system. Personal Consultant is composed of a development engine and an inference engine. The development engine enables the knowledge engineer to develop and main- tain the knowledge base. The inference engine enables the knowledge engineer to test the knowledge base, and it enables the user to execute a consultation. The architec- ture of a completed expert system is presented in Figure 2. The Development Engine The development engine is an interactive, window- oriented interface to the knowledge engineer that helps him to build his knowledge base into an expert system. In 40 41 K N O U I L E D G E BRSE RULES FRCTS I N F E R E N C E E N G I N E I N F E R E N C E CONTROL — z — D E U E L O P M E N T E N G I N E L D O R K I N G M E M O R V 1 1 1 1 1 1 1 1 1 1 1 1 EHPERT OR K N O U I L E D G E E N G I N E E R USER FIGURE 2 . T h e A r c h i t e c t u r e of a n E x p e r t S y s t e m ( f r o m Harmon a n d K i n g 1 9 8 5 ) 42 addition, the development engine's interactive structure editor allows on-line modification of an existing knowledge base, A knowledge base is entered by responding to prompts. The first prompt asks for the name of the knowledge base heading or domain. Then values for root context properties are prompted for, including a list of initial data and goal parameters. Once these parameter names have been entered, the development engine asks for information about the para- meters. Next, the context tree must be described by entering the descendants of the root context. At this point, rules can be inserted into the knowledge base. Rules can be entered directly in LISP or in an Abbreviated Rule Language (ARL). ARL is a BASIC-like rule specification lan- guage that simplifies the rule specification procedure for those unfamiliar with the LISP computer language. During the definition of the rules, if a parameter name is entered that has not been specified, the development engine prompts for information about the parameter. In this manner con- text, parameter, and rule properties are specified by the knowledge engineer until the knowledge base is completed. Definitions of knowledge base properties are explained in Appendix A. The main activity menu of the development engine pre- sents a list of 13 options designed to help modify or debug a knowledge base. The first option. Go, begins a consulta- tion with the specified knowledge base, and the second 43 option, Quit, ends a Personal Consultant session. The third option, Lisp, allows the knowledge engineer to activate the IQLISP interpreter if he wants to write special operations to augment the knowledge base. The Parameters, Rules, Con- texts, and Variables options allow properties associated with these entities to be entered or modified. The Func- tions option enables the knowledge engineer to define special functions written in LISP that are not immediately available in the development engine. The List option pro- vides a complete listing of the knowledge base. The Save option writes the knowledge base to a computer file. The Trace option is a debugging utility that produces a listing of the rules applied and the parameter values used in a given cojisultation. The final two options. Record and Play- back, are used to record a set of prompts with their respective responses from a given consultation and then use those responses in an automated session to detect changes in the logical flow of the knowledge base. Description of the Data A conventional computer program that accesses a data base usually stores the data in an external data file or in internal arrays. However, data used by an expert system can be inserted directly into the knowledge base with the aid of the development engine. Numerical data can be included as parameter values in rules that are needed to access the 44 data. The data base used by the workplace design expert sys- tem is a collection of anthropometric measurements taken from the U. S. military population (Hertzberg 1972). The data were derived on the basis of measurements from nude subjects. Therefore, the data were duly adjusted for cloth- ing, shoe, and other allowances which were added or subtracted to the various measurements. For the standing shoulder height, 2.5 cm were added for men's and women's shoes and 1.4 cm were added for clothing. The sitting shoulder height required an additional 0,6 cm for clothing under the buttocks. 1.0 cm was added to the body depth measurement and 2.0 cm were added to the thigh width measurement for clothing allowances. For the forearm length, 7.6 cm were subtracted for thumb and forefinger manipulation. For the arm length, 5.1 cm were subtracted to account for measurement from the back of the shoulder and 7,6 cm were subtracted for thumb and forefinger manipula- tion, 2,5 cm were added for men's and women's shoes to the standing elbow height. No allowances were given for the sitting elbow height and 2.5 cm were added to the popliteal height for men's and women's shoes (Hertzberg 1972). The corrected anthropometric data base is presented in Table 4. The workplace design expert system user can select the population that he wishes to accommodate at his workplace. The valid populations are male, female, and mixed. If the 45 TABLE 4. Corrected Anthropometric Measurements (from Hertzberg 1972, Das and Grady 1983) PARAMETER Shoulder 5 Shoulder 95 S i t t i n g - S h o u l d e r 5 B o d y - D e p t h T h i g h - l U i d t h Thigh-LDidth F o r e a r m 5 Rrm 5 Elboui 5 Elbouj 95 S i 11 i n g - E1 b 0 uj S i 11 i n g - E1 b 0 uj P o p l i t e a l 5 P o p l i t e a l 95 95 5 95 5 95 Male 1 3 8 . 0 1 5 6 . 8 5 4 . 7 3 4 . 0 1 4 . 2 1 8.5 3 7 . 1 6 8 . 3 1 0 5 . 6 1 2 0 . 4 1 8.8 2 7 . 4 4 2 . 4 4 8 . 7 VALUE (cm) Female 1 2 7 . 9 1 4 6 . 6 5 2 . 2 2 7 . 6 1 2 . 4 1 6.5 3 2 . 5 6 0 . 2 9 9 . 0 1 1 1 . 2 1 8.8 2 7 . 4 3 7 . 3 4 3 . 6 46 male population is chosen, 95th percentile male anthropo- metric measurements are used to determine clearance dimensions while Sth percentile male measurements are used to determine reach dimensions. The method is similar for the female population, /if a mixed population is chosen, 9Sth percentile male measurements are used for clearance determinations, and Sth percentile female measurements are used to calculate the reach requirements. Construction of the Knowledge Base The workplace design expert system knowledge base was constructed using the development engine as described at the beginning of this chapter. The knowledge base heading pro- perty was specified to give the knowledge base a name. Henceforth, the knowledge base will be referred to by its name, "The Ergonomist." All of the parameters used by "The Ergonomist" were defined in terms of their associated properties. See Appen- dix A for an explanation of knowledge base properties. A total of 41 parameters were defined for use by "The Ergono- mist," All parameters belong to the parameter group called Dimensional-Design-Farms. The workplace design knowledge represented in rule form and presented in Chapter III was entered directly into the knowledge base with the aid of the development engine. Since the knowledge was already organized as IF/THEN state- 47 ments, the rules were entered with very little modification A total of 57 rules were embedded in "The Ergonomist" and were grouped as follows: 6 Anthro-Rules, 16 Station-Type- Rules, 16 Dimensioning-Rules, 12 Thigh-Clearance-Rules, and 7 Data-Checking-Rules. The Inference Engine "The Ergonomist" knowledge base, when used in conjunc- tion with the inference engine, comprises the functioning workplace design expert system prototype. The inference engine controls the expert system's reasoning process as it uses the facts and rules stored in the knowledge base, as well as the information it acquired from the user. The inference engine performs two major tasks (Harmon and King 1985). First, it examines existing facts and rules, and adds new facts when possible. Second, it decides the order in which inferences are made. In doing so, the inference engine conducts the consultation with the user. To reach a conclusion about the goal parameters, the inference engine employs an inferencing strategy. Figure 3 presents the inferencing strategies that are utilized by various inference engines. Backward and forward chaining strategies can be utilized, as well as depth-first and breadth-first search strategies. Backward chaining infer- ence engines, or goal-directed systems, work backward through the knowledge base in an effort to choose an answer 48 p 2 & « (0 OQ j 2 i n 00 0) I—I W 3 O O C CO - H d) ^ • H 0 0 T 3 (U C 4-> CO CO Ji C 4-1 O CO e u x : CO a K u CO E cu o C O S-i M - l M - l > ^ O CO CO (1) CU C • H - H V̂ 0 0 O C O O W CU 4-) (U a c ( U M <U CO o CO C CO w o I—I 49 For example, if a goal parameter is D, and there is a rule that states "if C, then D" the inference engine will attempt to ascertain the value for C. The inference engine may dis- cover another rule that states "if A and B, then C." Then the values for A and B must be found. This is also an example of a depth-first search strategy. A breadth-first search sweeps across all of the goal parameters before digg- ing for greater detail. In cases in which the number of possible outcomes is large, a forward chaining strategy is often used. In a forward chaining system, or data-driven system, premises of rules are examined to see whether or not they are true, given the information available at the time. If so, then the conclusions are added to the list of facts known to be true and with the new information, the system examines the rules again. The inference engine supplied with the Personal Consultant employs a depth-first, backward chaining strategy. The inference engine controls the flow of a consultation with "The Ergonomist." Inference control is actually spread subtly throughout "The Ergonomist." since the order of clauses within a rule determines which clause is examined in what order. For example, there is a rule in the knowledge base that states, "if WORK-CLASS = PRECISION and FOOT-CONTROLS = YES and LARGE-AREA = NO, then STATION- TYPE = SIT." As a result, the inference engine tries to find answers for those three premise clauses in the order in so which they appear. Therefore, a degree of inference control is contained in the rule structure. During a consultation, the user has the ability to query the system, A help key is provided that can help clarify the meaning of a question presented to the user. If the user does not understand the meaning of a question, he can use the help key for further clarification. The resulting help message is the value of a knowledge base parameter property called reprompt. The why key is used to clarify the logic being used by the expert system. For example, if the user is presented with a question, he has the ability to ask the system why the answer to that question is needed. The system responds by listing the rule being examined for which the answer is required in order to apply the rule. Finally, the how key enables the user to view the logic that was used to determine the value of parameters already found by the expert system. Use of this key lists a chain of rules that determined the value for a selected parameter. Sample consultations with "The Ergonomist" are presented in Appendix C. The sample consultations simulated actual workplace design projects that could be encountered in an industrial setting and solved by "The Ergonomist" expert system. CHAPTER V CONCLUSION The development of a workplace design expert system that acts as an intelligent aid to workplace designers has been described. The expert system will potentially benefit facility engineers, human factors engineers, safety engin- eers, and industrial engineers. In addition, the system will be useful as an educational tool for students in those areas of specialization. The expert system will be most beneficial early in the design phase of a new workplace. Potential ergonomic prob- lems with the design can be avoided at great savings in time and money. The system ensures that the workplace is designed to fit the operator; it does not force the operator to fit the workplace. Use of the expert system to assist in the design of workplaces should reduce incidents of over- exertion and injury at the workplace. This result will in turn reduce compensation expenditures and avoid man hour losses . In addition to designing new workplaces, the expert system has the potential to evaluate existing workplaces. After the user has responded to questions asked about the existing workplace, the system will provide a description of what the workplace characteristics should be. Comparisons can be made to the existing workplace and necessary modifi- cations can be identified. The system can be used 51 52 iteratively to evaluate all existing workplaces in an indus- trial plant. The effort that went into the completion of this thesis project serves as a foundation for further research and development, "The Ergonomist" is a prototype of the ideal workplace design system conceptualized in Chapter II. As better expert system building tools and symbolic processing hardware become available, the development of a workstation dedicated to the design of industrial workplaces will become a reality. The expert system development tool used to create "The Ergonomist" has many useful features; however, there are areas that need improvement. The tool does not permit the use of complex rules premises. For example, a rule premise such as "if A and (B or C or D)" is not valid. Premises must contain all "ands" or all "ors." In addition, it would be desirable to be able to use other logical operators such as "not," "nand," and "nor" which are not recognized by the tool. The tool does not provide a means for using a graphic user interface. It would be highly desirable to provide a graphical output of the finished workplace design. In addi- tion, the meaning of certain questions asked by "The Ergonomist" could be clarified with graphics. For example, if the user is asked to enter the popliteal height measure- ment, a graphic screen could indicate how this measurement should be taken on the image of an anatomical manikin. S3 The Ergonomist" expert system has yet to undergo extensive testing in industry. Only the direct application of this system over an extended period of time will prove whether the system is beneficial in avoiding ergonomic prob- lems and reducing injury at the workplace. LIST OF REFERENCES Aikens, J. S,, "Prototypes and production rules: A knowledge representation for computer consultations." Unpublished Ph, D. dissertation, Stanford University, 1980, Asfour, S, S., "Energy cost prediction models for manual lifting and lowering tasks." Unpublished Ph, D. dissertation, Texas Tech University, 1980, Astrand, P. 0., Textbook of Work Physiology. McGraw-Hill, 1977. Ayoub, M. M., J, Fernandez, J, Smith, "Design of workplace." Unpublished report, Texas Tech University, 1982. Ayoub, M. M., "Workplace design and posture." Human Factors, 15, No. 3, 265-268, 1973. Barnes, R. M., Motion and Time Study, 2nd Edition, Wiley, 1940. Buchanan, B. G., E. Shortliffe, Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, Addison-Wesley, 1984. Das, B., R. Grady, "Industrial workplace layout and engineering anthropology." In T. Kvalseth (ed.) Ergonomics of Workstation Design, Butterworth, 1983. Duda, R. 0., R. Reboh, "AI and decision making: The PROSPECTOR experience." In W. Reitman (ed.) Artificial Intelligence Applications for Business, Ablex, 1984. Erman, L. D., F. Hayes-Roth, V. Lesser, D. Reddy, "The HEARSAY II speech-understanding system: Integrating knowledge to resolve uncertainty." Computing Surveys, 12 No. 2, 213-253, 1980. Feigenbaum, E. A., "The art of artificial intelligence: Themes and case studies of knowledge engineering." IJCAI 5, 1014-1029, 1977. Garg, A., D. Chaffin, G. Herrin, "Prediction of metabolic rates for manual materials handling jobs." Am. Ind. Hyg. Assoc. J., 13, 661-674, 1978. Grandjean, E., Fitting the Task to the Man: An Ergonomic Approach, Taylor and Francis, 1982. 54 55 Harmon, P., King, D., Expert Systems: Artificial Intelligence in Business. Wiley, 1985. Hayes-Roth, F., D. Waterman, D. Lenat (eds.). Building Expert Systems, Addison-Wesley, 1983. Hertzberg, H. T. E., "Engineering anthropology." In Van Cott H., R, Kincade (eds,) Human Engineering Guide to Equipment Design. McGraw-Hill, 1972. Karwowski, W., N. Mulholland, T. Ward, V. Jagannathan, R. Kirchner, "LIFTAN: A knowledge-based expert system for analysis of manual lifting tasks." Unpublished report. University of Louisville, 1985. ,X6nz, S., "Design of work stations." J. Industrial Eng. 18, No. 7, 413-423, 1967. Konz, S., Work Design. Grid, 1979. Lindsay, R. K., B. Buchanan, E. Feigenbaum, J. Lederberg, Chemistry: The DENDRAL Project. McGraw-Hill, 1980. Martin, W. A., R. Fateman, "The MACSYMA system." Proceedings of the 2nd Symposium on Symbolic and Algebraic Manipulation. 59-75, 1971. Maynard, H. B., "Workplace layouts that save time, effort and money." Iron Age 134, No. 23, 28-30, 1934. Smith, J. L., J. Ramsey, "Designing physically demanding tasks to minimize levels of worker stress." Industrial Engineering, May, 1982. Slagle, J. R., "A heuristic program that solves symbolic integration problems in freshman calculus." Symbolic Automatic Integrator (SAINT), Unpublished Ph. D. dissertation, MIT, 1961. Texas Instruments, Inc., "Personal Consultant: Expert system development tools." User's Guide, T. I., 1985. Tichauer, E. R., Occupational Biomechanics: The Anatomical Basis of Workplace Design, New York University Medical Center, No. 51, 1975. Utfoodson, W. E., D. W. Conover, Human Engineering Guide for Equipment Designers. 2nd Edition, University of California Press, 1964. APPENDIX A DEFINITION OF KNOWLEDGE BASE PROPERTIES This appendix contains explanations of the concepts and terminology used by the Personal Consultant in describing properties of a knowledge base. This information is provided so that structure and context of a knowledge base can be better understood. Refer to Appendix B, Listing of the Knowledge Base, for a detailed description of "The Ergonomist" knowledge base properties. A knowledge base consists of three main components: contexts, parameters, and rules. The properties pertaining to these components are explained below as defined by Texas Instruments (1985). Contexts * PARMGROUP - The name of the parameter group associated with a context. The parameter group contains the names of all parameters in the group. There is only one parameter group associated with a context. * RULETYPES - The names of all rule groups associated with a context. Like the parameter group, the rule group contains the names of all rules in the group. There can be several rule groups associated with a context. * GOALS - A list of the parameters in a context's parameter group whose values, when determined, describe the solution to the context problem. * DISPLAYRESULTS - A value that specifies if the value of the parameter(s) listed in the GOALS property is to be displayed when the context is exited. * INITIALDATA - A list of the parameters in the context's parameter group whose values are always prompted for when the context is instantiated during the 56 57 consultation. * PROMPTEVER, PROMPTIST, PR0MPT2ND, TRANS - Text that is used to describe the type of information needed in the consultation. These properties are used by the infer- ence engine to introduce or announce the instance of a context or describe the purpose of the context. * SYN, UNIQUE - Values that are used to translate the con- text instance name into a meaningful form for the client. PRINTID - Property whose value is used to name a context instance. * ASSOCWITH, OFFSRING - Values used to store information about the context tree. Parameters * PROMPT - This property is usually the question that asks the client for the value of the parameter. If this property contains the question, the question is displayed to the client. * REPROMPT - This property is an explanation or help message for the prompt that is displayed to the client during a consultation. * TRANS - This property is an English phrase that describes the purpose or use of the parameter. This phrase is used to generate questions, explain the reasoning or inferencing of the inference engine, or describe the results of a consultation. Correct form in the TRANS property helps make a knowledge base easier to understand and to correct if problems exist. * ASKFIRST - This property tells the inference engine to ask the client for the value before trying to infer it using rules. * EXPECT - This property describes the types of input values expected from the client. This property can be any of the following: - A list of all appropriate values - the response can be any of the values in the list. - NIL - Expect a yes or no response from the client 58 - ANY - Any type of response is legal. - NUMB - The input value must be a number. - POSNUMB - The input value must be a number greater than or equal to zero. - FIXP - The input value must be an integer, LEGALVALS - This property describes all possible values of a parameter. When a parameter does not have a LEGALVALS property, the inference engine assumes that the EXPECT property specifies all the legal values. This property is required for multivalued parameters. This property can be a list of all legal values, or one of these special values: - ANY - Any value is a legal parameter value. - TEXT - The parameter's value can be any text phrase. Parameters with this LEGALVALS property value can be used to store text which can explain the conclusion or recommendations of a consultation. * MULTIVALUED - This property is used to define the type of a parameter. If this property does not exist, the parameter is single-valued. * USED-BY, UPDATED-BY, CONTAINED-IN, ANTECEDENT-IN, UPDATED-IN - These system properties are used to identify rules that use or modify the parameter's values. These properties are automatically maintained by the development engine. These properties are useful during the debugging process for the knowledge base. Rules * PREMISE - This property defines all the conditions to be met before the rule action will be taken. * ACTION - This property defines the actions taken if the premise is true. ^ SUBJECT - This property is the rule group that the rule belongs to. This determines the context(s) in which the rule must be tried. •5̂ ANTECEDENT - This property is optional and indicates that the rule is a forward-chaining rule. APPENDIX B LISTING OF THE KNOWLEDGE BASE Rule Group ANTHRO-RULES RULEOOl [ANTHRO-RULES] If the workf Then 1) 2) 3) ^ ) 5) 6) 7) 8) 9) 10) 11) 12) it i Stan it i Stan it i Stan It 1 Stan it i popl it i popl It 1 sitt It 1 sitt it i sitt it bod it per it per orce s de ding s de ding s de ding s de ding s de itea s de itea s de ing s de ing s de ing is d y de is d cent is d cent populat finite ( elbow finite elbow finite shoulde finite ( shoulde finite ( 1 height finite ( 1 height finite ( elbow he finite ( elbow he finite ( shoulder efinite pth is 3 efinite lie fore efinite ile arm ion IS 100%) eight 100%) eight 100%) r heig 100%) r heig 100%) is 48 100%) is 42 100%) ight i 100%) ight i 100%) heigh (100%) 4.0, a (100%) arm is (100%) is 68. MALE, that t is 120 that t is 105 that t ht is that t ht is that t . 7 , an that t . 4 , an that t s 27.4 that t s 18.8 that t t is 5 that nd that 37.1, that 3. he 9 .4, he 5 .6, he 9 156. he 5 138. he 9 d he 5 d he 9 , an he 5 , an he 5 4.7, the 5 th percentile and th percentile and 5 th percentile 8 , and th percentile 0, and 5 th percentile th percentile 5 th percentile d th percentile d th percentile and 95 th percentile the length of the Sth and the length of the Sth RULE002 [ANTHRO-RULES] If the workforce popula Then 1) it is definite standing elbow 2) it is definite standing elbow 3) it is definite standing should 4) it is definite standing should 5) it is definite popliteal heigh 6) it is definite popliteal heigh tion is FEMALE, (100%) that the 95 th percentile height is 111.2, and (100%) that the Sth percentile height is 99.0, and (100%) that the 95 th percentile er height is 146.6, and (100%) that the Sth percentile er height is 127.9, and (100%) that the 95 th percentile t is 43.6, and (100%) that the Sth percentile t is 37.3, and 59 60 7) it is definite (100%) that the 95 th percentile sitting elbow height is 21.k, and 8) it is definite (100%) that the Sth percentile sitting elbow height is 18.8, and 9) it is definite (100%) that the Sth percentile sitting shoulder height is 52.2, and 10) it is definite (100%) that the 95 th percentile body depth is 27.6, and 11) it is definite (100%) that the length of the Sth percentile forearm is 32.5, and 12) it is definite (100%) that the length of the Sth percentile arm is 60.2, ' RULE003 [ANTHRO-RULES] If the workf Then 1) it i Stan it i Stan I it i Stan lit i /̂ v-" V 4 5 6 7 8 9 10) 12) Stan it i popl it i popl it i sitt It 1 sitt it i sitt it bod it per it per orce s de ding s de ding s de ding s de ding s de itea s de itea s de ing s de ing s de ing is d y de is d cent is d cent pop f ini elb f ini elb f ini sho f ini sho f ini 1 he f ini 1 he f ini elbo f ini elbo f ini shou ef in pth ef in ile ef in ile ulati te (1 ow te ow te he (1 he (1 ulder te (1 ulder te (1 ight te (1 ight te (1 w hei te (1 w hei te (1 Ider ite ( is 34 ite ( f orea ite ( arm i on 1 00%) ight 00%) ight 00%) hei 00%) hei 00%) is 4 00%) is 3 00%) ght 00%) ght 00%) heig 100% .0, 100% rm i 100% s 60 s MI tha is tha is tha ght tha ght tha 8.7, tha 7.3, tha is 2 tha is 1 tha ht i ) th and ) th s 32 ) th .2. XED, t the 9 120.4, t the 5 99.0, a t the 9 is 156. t the 5 is 127. t the 9 and t the and t the 9 4 , an the 5 8 , an the 5 52.2, the 7. t 8. t s at 5 th percentile and th percentile nd 5 th percentile 8 , and th percentile 9, and 5 th percentile Sth percentile 5 th percentile d th percentile d th percentile and 95 th percentile at the length of the Sth .5, and at the length of the Sth RULE043 [ANTHRO-RULES] If the workforce population is MALE, Then 1) it is definite (100%) that the Sth percentile thigh width is 14.2, and 2) it is definite (100%) that the 95 th percentile thigh width is 18.5. 61 RULE044 [ANTHRO-RULES] If the workforce population is FEMALE, Then 1) it is definite (100%) that the Sth percentile thigh width is 12.4, and 2) it is definite (100%) that the 95 th percentile thigh width is 16.S. RULE04S [ANTHRO-RULES] If the workforce population is MIXED, Then 1) it is definite (100%) that the Sth percentile thigh width is 12,4, and 2) it is definite (100%) that the 95 th percentile thigh width is 18,5, Rule Group DATA-CHECKING-RULES RULE049 [DATA-CHECKING-RULES] If 1 2 the Sth percentile standing shoulder height is less than 138, or the 95 th percentile standing shoulder height is greater than 156,8, or 3) the Sth percentile sitting shoulder height is less than 54 , 7 , or 4) the 95 th percentile body depth is greater than 34, or 5) the Sth percentile thigh width is less than 14.2, or 6) the 95 th percentile thigh width is greater than 18,5, or 7) the length of the Sth percentile forearm is less than 37.1, or 8) the length of the Sth percentile arm is less than 68.3, or 9) the Sth percentile standing elbow height is less than 105.6, or 10) the 95 th percentile standing elbow height is greater than 120.4, or 11) the Sth percentile sitting elbow height is less than 18.8, or 12) the 95 th percentile sitting elbow height is greater than 27.4, or 13) the Sth percentile popliteal height is less than 42.4, or 14) the 95 th percentile popliteal height is greater than 48.7, Then it is definite (100%) that there is a probable error in the male anthropometric data. 62 RULEOSO [DATA-CHECKING-RULES] If 1 2 3 4 5 6 7 8 9 the Sth percentile standing shoulder height is less than 127.9, or the 95 th percentile standing shoulder height is greater than 146.6, or the Sth percentile sitting shoulder height is less than 52.2, or the 95 th percentile body depth is greater than 27.6, or the Sth percentile thigh width is less than 12.4, or the 95 th percentile thigh width is greater than 16.5, or the length of the Sth percentile forearm is less than 32.5, or the length of the Sth percentile arm is less than 60.2, or the Sth percentile standing elbow height is less than 99, or 10) the 95 th percentile standing elbow height is greater than 111.2, or 11) the Sth percentile sitting elbow height is less than 18.8, or 12) the 95 th percentile sitting elbow height is greater than 27.4, or 13) the Sth percentile popliteal height is less than 37.3, or 14) the 95 th percentile popliteal height is greater than 43.6, Then it is definite (100%) that there is a probable error in the female anthropometric data. RULEOSl [DATA-CHECKING-RULES] If 1 2 3 4 5 6 7 8 9 the Sth percentile standing shoulder height is less than 127.9, or the 95 th percentile standing shoulder height is greater than 156.8, or the Sth percentile sitting shoulder height is less than 52.2, or the 95 th percentile body depth is greater than 34, or the Sth percentile thigh width is less than 12.4, or the 95 th percentile thigh width is greater than 18.5, or the length of the Sth percentile forearm is less than 32.5, or the length of the Sth percentile arm is less than 60.2, or the Sth percentile standing elbow height is less than 99, or 63 10) the 95 th percentile standing elbow height is greater than 120.4, or 11) the Sth percentile sitting elbow height is less than 18.8, or 12) the 95 th percentile sitting elbow height is greater than 27.4, or 13) the Sth percentile popliteal height is less than 37.3, or 14) the 95 th percentile popliteal height is greater than 48.7, Then it is definite (100%) that there is a probable error in the mixed anthropometric data. RULE0S2 [DATA-CHECKING-RULES] If 1) the workforce population is SPECIFIC, and 2) the sex of your specific population is MALE, and 3) there is a probable error in the male anthropometric data, Then it is definite (100%) that DATA CHECKING is The following conclusions were made by THE ERGONOMIST concerning the current workplace design. However, a probable error was detected in the anthropometric data that you entered. Please re-check your data and run another consultation.. RULE053 [DATA-CHECKING-RULES] If 1) the workforce population is SPECIFIC, and 2) the sex of your specific population is FEMALE, and 3) there is a probable error in the female anthropometric data. Then it is definite (100%) that DATA CHECKING is The following conclusions were made by THE ERGONOMIST concerning the current workplace design. However, a probable error was detected in the anthropometric data that you entered. Please re-check your data and run another consultation.. RULE0S4 [DATA-CHECKING-RULES] If 1) the workforce population is SPECIFIC, and 2) the sex of your specific population is MIXED, and 3) there is a probable error in the mixed anthropometric data , Then it is definite (100%) that DATA CHECKING is The following conclusions were made by THE ERGONOMIST concerning the current workplace design. However, a probable error was detected in the anthropometric data 64 that you entered. Please re-check your data and run another consultation.. RULEOSS [DATA-CHECKING-RULES] If 1) the workforce population is MALE, or 2) the workforce population is FEMALE, or 3) the workforce population is MIXED, or 4) there is not a probable error in the male anthropometric data, or 5) there is not a probable error in the female anthropometric data, or 6) there is not a probable error in the mixed anthropometric data. Then it is definite (100%) that DATA CHECKING is The following conclusions were made by THE ERGONOMIST concerning the current worplace design,. Rule Group DIMENSIONING-RULES RULE020 [DIMENSIONING-RULES] If the classification of the task is PRECISION-WORK, Then 1) it is definite (100%) that the work classification correction factor-1 is 12, and 2) it is definite (100%) that the work classification correction factor-2 is 8. RULE021 [DIMENSIONING-RULES] If the classification of the task is LIGHT-WORK, Then 1) it is definite (100%) that the work classification correction factor-1 is 5, and 2)\it is definite (100%) that the work classification correction factor-2 is -5. RULE022 [DIMENSIONING-RULES] If the classification of the task is HEAVY-WORK, Then 1) it is definite (100%) that the work classification correction factor-1 is 0, and 2) it is definite (100%) that the work classification correction factor-2 is -13. 65 RULE023 [DIMENSIONING-RULES] If the classification of task is VDT/KEYBOARD-OPERATION, Then 1) it is definite (100%) that the work classification correction factor-1 is 0, and 2) it is definite (100%) that the work classification correction factor-2 is -5. RULE024 [DIMENSIONING-RULES] If 1) the workstation type is SIT, and 2) the design type is FULLY-ADJUSTABLE, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ [ the 95 th percentile popliteal height plus the 95 th percentile sitting elbow height ] plus the work classification correction factor-1 ], and 2) it is definite (100%) that work surface lower ht. in centimeters is [ [ the Sth percentile popliteal height plus the Sth percentile sitting elbow height ] plus the work classification correction factor-1 ], and 3) it is definite (100%) that chair upper height in centimeters is [ the 95 th percentile popliteal height minus 0 ], and 4) it is definite (100%) that chair lower height in centimeters is [ the Sth percentile popliteal height minus 0 ], and 5) it is definite (100%) that footrest upper height in centimeters is [ the 95 th percentile popliteal height minus the Sth percentile popliteal height ] , and 6) it is definite (100%) that footrest lower height in centimeters is 0, and 7) it is definite (100%) that normal vertical reach in centimeters is [ [ the Sth percentile popliteal height plus the 5th percentile sitting elbow height ] plus the length of the Sth percentile forearm ], and 8) it is definite (100%) that maximum vertical reach in centimeters is [ [ the Sth percentile popliteal height plus the Sth percentile sitting shoulder height ] plus the length of the Sth percentile arm ] , and 9) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth percentile forearm minus [ .5 times the 95 th percentile body depth ] ], and 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus [ .5 times the 95 th 66 percentile body depth ] ]. RULE025 [DIMENSIONING-RULES] If 1) the workstation type is STAND, and 2) the design type is FULLY-ADJUSTABLE, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ the 95 th percentile standing elbow height plus the work classification correction factor-2 ] , and 2) it is definite (100%) that work surface lower ht. in centimeters is [ the Sth percentile standing elbow height plus the work classification correction factor-2 ], and 3) it is definite (100%) that chair upper height in centimeters is NONE, and 4) it is definite (100%) that chair lower height in centimeters is NONE, and 5) it is definite (100%) that footrest upper height in centimeters is NONE, and 6) it is definite (100%) that footrest lower height in ^-^ centimeters is NONE, and 7) it is definite (100%) that normal vertical reach in centimeters is [ the Sth percentile standing elbow height plus the length of the Sth'^percentile forearm ], and 8) it is definite (100%) that maximum vertical reach -̂ in centimeters is [ the Sth percentile standing shoulder height plus the length of the Sth percentile arm ], and 9) it is definite (100%) that normal horizontal reach . -yln centimeters is [ the length of the Sth percentile forearm minus [ .5 times the 95 th percentile body depth ] ], and ^ 10) it is definite (100%) that maximum horizontal ^ ̂ )reach in centimeters is [ the length of the Sth percentile arm minus [ .5 times the 95 th percentile body depth ] ] . RULE026 [DIMENSIONING-RULES] If 1) the workstation type is SIT/STAND, and 2) the design type is FULLY-ADJUSTABLE, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ the 95 th percentile standing elbow height plus the work classification correction factor-2 ] , and 2) it is definite (100%) that work surface lower ht. in centimeters is [ the Sth percentile standing elbow height plus the work classification 67 correction factor-2 ], and 3) it is definite (100%) that chair upper height in centimeters is [ the 95 th percentile standing elbow height minus the Sth percentile sitting elbow height J, and 4) it is definite (100%) that chair lower height in centimeters is [ the Sth percentile standing elbow height minus the 95 th percentile sitting elbow height ] , and 5) it is definite (100%) that footrest upper height in centimeters is [ [ the 95 th percentile standing elbow height minus the Sth percentile sitting elbow height ] minus the Sth percentile popliteal height ] , and 6) it is definite (100%) that footrest lower height in centimeters is 0, and 7) it is definite (100%) that normal vertical reach in centimeters is [ the Sth percentile standing elbow height plus the length of the Sth percentile forearm ], and 8) it is definite (100%) that maximum vertical reach in centimeters is [ the Sth percentile standing shoulder height plus the length of the Sth percentile arm ] , and 9) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth percentile forearm minus [ .5 times the 95 th percentile body depth ] ], and 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus the 95 th percentile body depth ] . RULE027 [DIMENSIONING-RULES] If 1) the workstation type is SIT, and 2) the design type is FIXED-WORK-SURFACE, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ [ the 95 th percentile popliteal height plus the 95 th percentile sitting elbow height ] plus the work classification correction factor-1 ] , and 2) it is definite (100%) that work surface lower ht. in centimeters is NONE, and 3) it is definite (100%) that chair upper height in centimeters is [ [ the 95 th percentile popliteal height plus the 95 th percentile sitting elbow height ] minus the Sth percentile sitting elbow height ], and 4) it is definite (100%) that chair lower height in centimetjers is [ the 95 th percentile popliteal 68 height minus 0 ], and 5) it is definite (100%) that footrest upper height in centimeters is [ [ [ the 95 th percentile popliteal height plus the 95 th percentile sitting elbow height ] minus the Sth percentile sitting elbow height ] minus the Sth percentile popliteal height ] , and 6) it is definite (100%) that footrest lower height in centimeters is 0, and 7) it is definite (100%) that normal vertical reach in centimeters is [ [ the 95 th percentile popliteal height plus the 95 th percentile sitting elbow height ] plus the length of the Sth percentile forearm ], and 8) it is definite (100%) that maximum vertical reach in centimeters is [ [ the 95 th percentile popliteal height plus the Sth percentile sitting shoulder height ] plus the length of the Sth percentile arm ], and 9) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth percentile forearm minus [ .5 times the 95 th percentile body depth ] ] , and 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus [ .5 times the 95 th percentile body depth ] ]. RULE028 [DIMENSIONING-RULES] If 1) the workstation type is STAND, and 2) the design type is FIXED-WORK-SURFACE, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ the 95 th percentile standing elbow height plus the work classification correction factor-2 ], and 2) it is definite (100%) that work surface lower ht. in centimeters is NONE, and 3) it is definite (100%) that chair upper height in centimeters is NONE, and 4) it is definite (100%) that chair lower height in centimeters is NONE, and 5) it is definite (100%) that footrest upper height in centimeters is [ the 95 th percentile standing elbow height minus the Sth percentile standing elbow height ] , and it is definite (100%) that footrest lower height in centimeters is 0, and it is definite (100%) that normal vertical reach in centimeters is [ the 95 th percentile standing elbow height plus the length of the Sth percentile 69 forearm ], and 8)iit is definite (100%) that maximum vertical reach in centimeters is [ the Sth percentile standing shoulder height plus the length of the Sth ^̂v̂̂js percentile arm ] , and 9) it is definite (100%) that normal horizontal reach \,\in centimeters is [ the length of the Sth 'percentile forearm minus [ .5 times the 95 th ^percentile body depth ] ], and 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth ^ percentile arm minus [ .5 times the 95 th percentile body depth ] ] . RULE029 [DIMENSIONING-RULES] If 1) the workstation type is SIT/STAND, and 2) the design type is FIXED-WORK-SURFACE, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ the 95 th percentile standing elbow height plus the work classification correction factor-2 J, and 2) it is definite (100%) that work surface lower ht. in centimeters is NONE, and 3) it is definite (100%) that chair upper height in centimeters is [ the 95 th percentile standing elbow height minus the Sth percentile sitting elbow height ], and 4) it is definite (100%) that chair lower height in centimeters is [ the 95 th percentile standing elbow height minus the 95 th percentile sitting elbow height ], and 5) it is definite (100%) that footrest upper height in centimeters is [ [ the 95 th percentile standing elbow height minus the Sth percentile sitting elbow height ] minus the Sth percentile popliteal height ], and 6) it is definite (100%) that footrest lower height in centimeters is 0, and 7) it is definite (100%) that normal vertical reach in centimeters is [ the 95 th percentile standing elbow height plus the length of the Sth percentile forearm ], and 8) it is definite (100%) that maximum vertical reach in centimeters is [ [ [ the 95 th percentile standing elbow height minus the 95 th percentile sitting elbow height ] plus the Sth percentile sitting shoulder height ] plus the length of the Sth percentile arm ], and 9) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth 70 percentile forearm minus [ .5 times the 95 th percentile body depth ] ], and 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus [ .5 times the 95 th percentile body depth ] ]. RULE030 [DIMENSIONING-RULES] If 1) the workstation type is SIT, and 2) the design type is FIXED-CHAIR, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ [ the 95 th percentile popliteal height plus the 95 th percentile sitting elbow height ] plus the work classification correction factor-1 ], and 2) it is definite (100%) that work surface lower ht. in centimeters is [ [ the 95 th percentile popliteal height plus the Sth percentile sitting elbow height ] plus the work classification correction factor-1 ], and 3) it is definite (100%) that chair upper height in centimeters is [ the 95 th percentile popliteal height minus 0 ], and 4) it is definite (100%) that chair lower height in centimeters is NONE, and 5) it is definite (100%) that footrest upper height in centimeters is [ the 95 th percentile popliteal height minus the Sth percentile popliteal height ] , and 6) it is definite (100%) that normal vertical reach in centimeters is [ [ the 95 th percentile popliteal height plus the Sth percentile sitting elbow height ] plus the length of the Sth percentile forearm ], and 7)- it is definite (100%) that maximum vertical reach in centimeters is [ [ the 95 th percentile popliteal height plus the Sth percentile sitting shoulder height ] plus the length of the Sth percentile arm ], and 8) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth percentile forearm minus [ .5 times the 95 th percentile body depth ] ], and 9) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus [ .5 times the 95 th percentile body depth ] ] . 71 RULE031 [DIMENSIONING-RULES] If 1) the workstation type is STAND, and 2) the design type is FIXED-CHAIR, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ the 95 th percentile standing elbow height plus the work classification correction factor-2 ], and 2) it is definite (100%) that work surface lower ht. in centimeters is [ the Sth percentile standing elbow height plus the work classification correction factor-2 ], and 3) it is definite (100%) that chair upper height in centimeters is NONE, and 4) it is definite (100%) that chair lower height in centimeters is NONE, and 5) it is definite (100%) that footrest upper height in centimeters is NONE, and 6) it is definite (100%) that footrest lower height in centimeters is NONE, and 7) it is definite (100%) that normal vertical reach in centimeters is [ the Sth percentile standing elbow height plus the length of the Sth percentile forearm ], and 8) it is definite (100%) that maximum vertical reach in centimeters is [ the Sth percentile standing shoulder height plus the length of the Sth percentile arm ], and 9) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth percentile forearm minus [ .5 times the 95 th percentile body depth ] ], and 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus [ .5 times the 95 th percentile body depth ] ]. RULE032 [DIMENSIONING-RULES] If 1) the workstation type is SIT/STAND, and 2) the design type is FIXED-CHAIR, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ [ [ the 95 th percentile standing elbow height minus the Sth percentile sitting elbow height ] plus the 95 th percentile sitting elbow height ] plus the work classification correction factor-2 ], and 2) it is definite (100%) that work surface lower ht. in centimeters is [ the 95 th percentile standing elbow height plus the work classification correction factor-2 ], and 72 3) it is definite (100%) that chair upper height in centimeters is [ the 95 th percentile standing elbow height minus the Sth percentile sitting elbow height ], and 4) it is definite (100%) that chair lower height in centimeters is NONE, and 5) it is definite (100%) that footrest upper height in centimeters is [ [ the 95 th percentile standing elbow height minus the Sth percentile sitting elbow height ] minus the Sth percentile popliteal height ], and 6) it is definite (100%) that footrest lower height in centimeters is 0, and 7) it is definite (100%) that normal vertical reach in centimeters is [ the Sth percentile standing elbow height plus the length of the Sth percentile forearm ], and 8) it is definite (100%) that maximum vertical reach in centimeters is [ the Sth percentile standing shoulder height plus the length of the Sth percentile arm ], and 9) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth percentile forearm minus [ .5 times the 95 th percentile body depth ] ], and 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus [ .5 times the 95 th percentile body depth ] ]. RULE046 [DIMENSIONING-RULES] If 1) the workstation type is SIT, and 2) the design type is STATIONARY, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ [ the 95 th percentile popliteal height plus the 95 th percentile sitting elbow height ] plus the work classification correction factor-1 ], and 2) it is definite (100%) that work surface lower ht. in centimeters is NONE, and 3) it is definite (100%) that chair upper height in centimeters is [ [ the 95 th percentile popliteal height plus the 95 th percentile sitting elbow height ] minus [ [ the Sth percentile sitting elbow height plus the 95 th percentile sitting elbow height ] divided by 2 ] ], and 4) it is definite (100%) that chair lower height in centimeters is NONE, and 5) it is definite (100%) that footrest upper height in centimeters is [ chair upper height in centimeters 73 minus [ [ the Sth percentile popliteal height plus the 95 th percentile popliteal height ] divided by 2 ] ], and 6) it is definite (100%) that footrest lower height in centimeters is NONE, and 7) it is definite (100%) that normal vertical reach in centimeters is [ [ chair upper height in centimeters plus the Sth percentile sitting elbow height ] plus the length of the Sth percentile forearm ], and 8) it is definite (100%) that maximum vertical reach in centimeters is [ [ chair upper height in centimeters plus the Sth percentile sitting shoulder height ] plus the length of the Sth percentile arm ], and 9) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth percentile forearm minus [ the 95 th percentile body depth divided by 2 ] ], and 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus [ the 95 th percentile body depth divided by 2 ] ]. RULE047 [DIMENSIONING-RULES] If 1) the workstation type is STAND, and 2) the design type is STATIONARY, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ the 95 th percentile standing elbow height plus the work classification correction factor-2 ], and 2) it is definite (100%) that work surface lower ht. in centimeters is NONE, and 3) it is definite (100%) that chair upper height in centimeters is NONE, and 4) it is definite (100%) that chair lower height in centimeters is NONE, and 5) it is definite (100%) that footrest upper height in centimeters is [ the 95 th percentile standing elbow height minus [ [ the Sth percentile standing elbow height plus the 95 th percentile standing elbow height ] divided by 2 ] ] , and 6) it is definite (100%) that footrest lower height in centimeters is NONE, and 7) it is definite (100%) that normal vertical reach in centimeters is [ [ footrest upper height in centimeters plus the Sth percentile standing elbow height ] plus the length of the Sth percentile forearm ], and 8) it is definite (100%) that maximum vertical reach 74 in centimeters is [ [ footrest upper height in centimeters plus the Sth percentile standing shoulder height ] plus the length of the Sth percentile arm ], and 9) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth percentile forearm minus [ the 95 th percentile body depth divided by 2 ] ], and 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus [ the 95 th percentile body depth divided by 2 ] ] . RULE048 [DIMENSIONING-RULES] If 1) the workstation type is SIT/STAND, and 2) the design type is STATIONARY, Then 1) it is definite (100%) that work surface upper ht. in centimeters is [ the 95 th percentile standing elbow height plus the work classification correction factor-2 ], and 2) it is definite (100%) that work surface lower ht. in centimeters is NONE, and 3) it is definite (100%) that chair upper height in centimeters is [ the 95 th percentile standing elbow height minus [ [ the Sth percentile sitting elbow height plus the 95 th percentile sitting elbow height ] divided by 2 ] ], and 4) it is definite (100%) that chair lower height in centimeters is NONE, and 5) it is definite (100%) that footrest upper height in centimeters is [ chair upper height in centimeters minus [ [ the Sth percentile popliteal height plus the 95 th percentile popliteal height ] divided by 2 ] ] , and 6) it is definite (100%) that footrest lower height in centimeters is NONE, and 7) it is definite (100%) that normal vertical reach in centimeters is [ [ chair upper height in centimeters plus the Sth percentile sitting elbow height ] plus the length of the Sth percentile forearm ], and 8) it is definite (100%) that maximum vertical reach in centimeters is [ [ chair upper height in centimeters plus the Sth percentile sitting shoulder height ] plus the length of the Sth percentile arm ] , and 9) it is definite (100%) that normal horizontal reach in centimeters is [ the length of the Sth percentile forearm minus [ the 95 th percentile body depth divided by 2 ] ], and 75 10) it is definite (100%) that maximum horizontal reach in centimeters is [ the length of the Sth percentile arm minus [ the 95 th percentile body depth divided by 2 ] ]. Rule Group STATION-TYPE-RULES RULE004 [STATION-TYPE-RULES] If 1) the classification of the task is PRECISION-WORK, and 2) the requirement to use foot controls, and 3) the requirement to cover a large work area. Then it is definite (100%) that the workstation type is SIT/STAND. RULEOOS [STATION-TYPE-RULES] If 1) the classification of the task is PRECISION-WORK, and 2) the requirement to use foot controls, and 3) the requirement to cover a large work area is not true , Then it is definite (100%) that the workstation type is SIT. RULE006 [STATION-TYPE-RULES] If 1) the classification of the task is PRECISION-WORK, and 2) the requirement to use foot controls is not true, and 3) the requirement to cover a large work area is not true , Then it is definite (100%) that the workstation type is SIT/STAND. RULE007 [STATION-TYPE-RULES] If 1) the classification of the task is PRECISION-WORK, and 2) the requirement to use foot controls is not true, and 3) the requirement to cover a large work area. Then it is definite (100%) that the workstation type is SIT/STAND. RULE008 [STATION-TYPE-RULES] If 1) the classification of the task is LIGHT-WORK, and 2) the requirement to use foot controls, and 3) the requirement to cover a large work area. 76 Then it is definite (100%) that the workstation type is SIT/STAND. RULE009 [STATION-TYPE-RULES] If 1) the classification of the task is LIGHT-WORK, and 2) the requirement to use foot controls, and 3) the requirement to cover a large work area is not true , Then it is definite (100%) that the workstation type is SIT. RULEOIO [STATION-TYPE-RULES] If 1) the classification of the task is LIGHT-WORK, and 2) the requirement to use foot controls is not true, 3) the requirement to cover a large work area is not true, Then it is definite (100%) that the workstation type is SIT/STAND. RULEOll [STATION-TYPE-RULES] If 1) the classification of the task is HEAVY-WORK, and 2) the requirement to use foot controls, and 3) the requirement to cover a large work area. Then it is definite (100%) that the workstation type is SIT/STAND. RULE013 [STATION-TYPE-RULES] If 1) the classification of the task is HEAVY-WORK, and 2) the requirement to use foot controls, and 3) the requirement to cover a large work area is not true , Then it is definite (100%) that the workstation type is SIT/STAND. and If 1) the classification of the task is LIGHT-WORK, and 2) the requirement to use foot controls is not true, and 3) the requirement to cover a large work area. Then it is definite (100%) that the workstation type is SIT/STAND. RULE012 [STATION-TYPE-RULES] 77 RULE014 [STATION-TYPE-RULES] If 1) the classification of the task is HEAVY-WORK, and 2) the requirement to use foot controls is not true, and 3) the requirement to cover a large work area is not true , Then it is definite (100%) that the workstation type is STAND. RULEOIS [STATION-TYPE-RULES] If 1) the classification of the task is HEAVY-WORK, and 2) the requirement to use foot controls is not true, and 3) the requirement to cover a large work area. Then it is definite (100%) that the workstation type is STAND. RULE016 [STATION-TYPE-RULES] If 1) the classification of the task is VDT/KEYBOARD-OPERATION, and 2) the requirement to use foot controls, and 3) the requirement to cover a large work area. Then it is definite (100%) that the workstation type is SIT/STAND. RULE017 [STATION-TYPE-RULES] If 1) the classification of the task is VDT/KEYBOARD-OPERATION, and 2) the requirement to use foot controls, and 3) the requirement to cover a large work area is not true, Then it is definite (100%) that the workstation type is SIT. RULE018 [STATION-TYPE-RULES] If 1) the classification of the task is VDT/KEYBOARD-OPERATION, and 2) the requirement to use foot controls is not true, and 3) the requirement to cover a large work area is not true , Then it is definite (100%) that the workstation type is SIT. 78 RULE019 [STATION-TYPE-RULES] If 1) the classification of the task is VDT/KEYBOARD-OPERATION, and 2) the requirement to use foot controls is not true, and 3) the requirement to cover a large work area. Then it is definite (100%) that the workstation type is SIT/STAND. Rule Group THIGH-CLEARANCE-RULES RULE033 [THIGH-CLEARANCE-RULES] If there is not a thigh clearance problem, Then it is definite (100%) that the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is NO-REDUCTION. RULE034 [THIGH-CLEARANCE-RULES] If there is a thigh clearance problem. Then 1) it is definite (100%) that the thigh clearance difference is [ the 95 th percentile thigh width minus [ [ work surface upper ht. in centimeters minus the table thickness ] minus chair upper height in centimeters ] ], and 2) it is definite (100%) that the Sth percentile thigh-width / popliteal combination is [ the Sth percentile thigh width plus the Sth percentile popliteal height ], and 3) it is definite (100%) that the 95 th percentile thigh width / popliteal combination is [ the 95 th percentile thigh width plus the 95 th percentile popliteal height J, and 4) it is definite (100%) that the estimated thigh width mean is [ [ the Sth percentile thigh-width / popliteal combination plus the 95 th percentile thigh width / popliteal combination ] divided by ] , and 5) it is definite (100%) that the estimated thigh width standard deviation is [ [ the 95 th percentile thigh width / popliteal combination minus the estimated thigh width mean ] divided by 1.645 ], and 6) it is definite (100%) that the critical thigh-clearance height is [ the 95 th percentile thigh width / popliteal combination minus the thigh clearance difference ]. 79 RULE03S [THIGH-CLEARANCE-RULES] If 1) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is less than or equal to 1.645, and 2) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is greater than 1.28, Then it is definite (100%) that the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is <S%. RULE036 [THIGH-CLEARANCE-RULES] If 1) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is less than or equal to 1.28, and 2) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is greater than 1.035, Then it is definite (100%) that the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is 5-10%. RULE037 [THIGH-CLEARANCE-RULES] If 1) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is less than or equal to 1.035, and 2) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is greater than .675, Then it is definite (100%) that the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is ilO- 20%1I. RULE038 [THIGH-CLEARANCE-RULES] If 1) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is less than or equal to .675, and 80 2) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is greater than .385, Then it is definite (100%) that the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is i20- 30%^. RULE039 [THIGH-CLEARANCE-RULES] If 1) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is less than or equal to ,385, and 2) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is greater than .125, Then it is definite (100%) that the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is i30- 40%1I. RULE040 [THIGH-CLEARANCE-RULES] If 1) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is less than or equal to .125, and 2) [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is greater than 0, Then it is definite (100%) that the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is i40- 50%1I. RULE041 [THIGH-CLEARANCE-RULES] If [ [ the critical thigh-clearance height minus the estimated thigh width mean ] divided by the estimated thigh width standard deviation ] is less than or equal to 0, Then it is definite (100%) that the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is >50%. 81 RULE042 [THIGH-CLEARANCE-RULES] If 1) the design type is FULLY-ADJUSTABLE, or 2) the design type is FIXED-CHAIR, or 3) the workstation type is SIT, or 4) the workstation type is STAND, Then it is definite (100%) that the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is NO-REDUCTION. RULE0S6 [THIGH-CLEARANCE-RULES] If 1) the design type is STATIONARY, and 2) the 95 th percentile thigh width is greater than [ [ work surface upper ht. in centimeters minus the table thickness ] minus chair upper height in centimeters ], Then it is definite (100%) that there is a thigh clearance problem. RULE057 [THIGH-CLEARANCE-RULES] If 1) the design type is FIXED-WORK-SURFACE, and 2) the workstation type is SIT/STAND, and 3) the 95 th percentile thigh width is greater than [ [ work surface upper ht. in centimeters minus the table thickness ] minus chair upper height in centimeters ], Then it is definite (100%) that there is a thigh clearance problem. Parameter Group CONTEXTTYPES DIMENSIONAL-DESIGN [CONTEXTTYPES] PROMPTEVER: (THE ERGONOMIST is a knowledge-based workplace design program developed to provide sound ergonomic advice to the industrial workstation designer. :line :line :line The current objective is to: :line (1) Decide what type of workstation to design and :line (2) Provide critical dimensions required to construct the workstation and :line (3) Determine if there is a reduction in the accommodated population due to physical constraints.) PRINTID: WORKPLACE- 82 PARMGROUP: DIMENSIONAL-DESIGN-PARMS RULETYPES: (DATA-CHECKING-RULES THIGH-CLEARANCE-RULES DIMENSIONING-RULES STATION-TYPE-RULES ANTHRO-RULES) INITIALDATA: (WORK-CLASS POPULATION) GOALS: (HEADER STATION-TYPE WSU WSL CHU CHL FRU FRL NVR MVR NHR MHR POPULATION-REDUCTION) DISPLAYRESULTS: T UNIQUE: T Parameter Group DIMENSIONAL-DESIGN-PARMS ARMS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the length of the Sth percentile arm) PROMPT: (For your specific population, enter the SMALLEST Arm Length in centimeters.) REPROMPT: (The Arm Length - is the horizontal distance from the posterior surface of the shoulder to the tip of the extended middle finger.) EXPECT: POSNUMB CONTAINED-IN: (RULE024 RULE025 RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) B0DY-DEPTH9S [DIMENSIONAL-DESIGN-PARMS] TRANS: (the 95 th percentile body depth) PROMPT: (For your specific population, enter the LARGEST Body Depth Measurement in centimeters.) REPROMPT: (The Body Depth Measurement - is the maximum horizontal distance between the vertical planes passing through the most anterior and posterior on the trunk.) EXPECT: POSNUMB CONTAINED-IN: (RULE024 RULE025 RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) CI [DIMENSIONAL-DESIGN-PARMS] TRANS: (the work classification correction factor-1) EXPECT: NUMB CONTAINED-IN: (RULE024 RULE027 RULE030 RULE046) UPDATED-BY: (RULE020 RULE021 RULE022 RULE023) 83 C2 [DIMENSIONAL-DESIGN-PARMS] TRANS: (the work classification correction factor-2) EXPECT: NUMB CONTAINED-IN: (RULE02S RULE026 RULE028 RULE029 RULE031 RULE032 RULE047 RULE048) UPDATED-BY: (RULE020 RULE021 RULE022 RULE023) CHL [DIMENSIONAL-DESIGN-PARMS] TRANS: (chair lower height in centimeters) EXPECT: POSNUMB DICTIONARY: INTERNAL CONTAINED-IN: NIL UPDATED-BY: (RULE024 RULE02S RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) CHU [DIMENSIONAL-DESIGN-PARMS] TRANS: (chair upper height in centimeters) EXPECT: POSNUMB DICTIONARY: INTERNAL CONTAINED-IN: (RULE034) USED-BY: (RULE056 RULE057) UPDATED-BY: (RULE024 RULE02S RULE026 RULE027 RULE029 RULE030 RULE031 RULE032 RULE047 RULE048) RULE028 RULE046 DESIGN-TYPE [DIMENSIONAL-DESIGN-PARMS] TRANS: PROMPT ASKFIRST EXPECT: USED-BY: (the design type) (What type of des :line :line FULL worksurface, cha SURFACE - with a :line FIXED CHAI and footrest :li adjustable featu : T (FULLY-ADJUSTABLE STATIONARY) (RULE024 RULE025 RULE030 RULE031 RULE048 RULE042 ign do you wish to construct? Y ADJUSTABLE - with adjustable ir, and footrest :line FIXED WORK djustable chair and footrest R - with adjustable worksurface ne STATIONARY - with no res ) FIXED-WORK-SURFACE FIXED-CHAIR RULE026 RULE027 RULE032 RULE046 RULE0S7) RULE028 RULE047 RULE029 RULE056 84 ELBOWS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the Sth percentile standing elbow height) PROMPT: (For your specific population, enter the SMALLEST Standing Elbow Height in centimeters.) REPROMPT: (The Standing Elbow Height - is the vertical distance from the floor to the depression at the elbow between the bones of the upper arm and forearm,) EXPECT: POSNUMB CONTAINED-IN: (RULE025 RULE026 RULE028 RULE031 RULE032 RULE047) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) ELB0W9S [DIMENSIONAL-DESIGN-PARMS] TRANS: (the 95 th percentile standing elbow height) PROMPT: (For your specific population, enter the LARGEST Standing Elbow Height in centimeters,) REPROMPT: (The Standing Elbow Height - is the vertical distance from the floor to the depression at the elbow between the bones of the upper arm and forearm.) EXPECT: POSNUMB CONTAINED-IN: (RULE025 RULE026 RULE028 RULE029 RULE031 RULE032 RULE047 RULE048) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) FEMALE-DATA-ERROR [DIMENSIONAL-DESIGN-PARMS] TRANS: (there is a probable error in the female anthropometric data) USED-BY: (RULE053 RULEOSS) UPDATED-BY: (RULEOSO) FOOT-CONTROLS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the requirement to use foot controls) PROMPT: (For this task, is the operator required to use any foot controls?) REPROMPT: (Is it necessary for the operator to operate footpedals, foot switches, or otherwise use his feet at the workplace?) ASKFIRST: T USED-BY: (RULE004 RULEOOS RULE006 RULE007 RULE008 RULE009 RULEOIO RULEOll RULE012 RULE013 RULE014 RULEOIS RULE016 RULE017 RULE018 RULE019) 85 FOREARMS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the length of the Sth percentile forearm) PROMPT: (For your specific population, enter the SMALLEST Forearm Length in centimeters.) REPROMPT: (The Forearm Length - is the horizontal distance from the tip of the elbow to the tip of the longest finger.) EXPECT: POSNUMB CONTAINED-IN: (RULE024 RULE02S RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) FRL [DIMENSIONAL-DESIGN-PARMS] TRANS: (footrest lower height in centimeters) EXPECT: POSNUMB DICTIONARY: INTERNAL UPDATED-BY: (RULE024 RULE02S RULE026 RULE027 RULE028 RULE029 RULE031 RULE032 RULE046 RULE047 RULE048) FRU [DIMENSIONAL-DESIGN-PARMS] TRANS: (footrest upper height in centimeters) EXPECT: POSNUMB DICTIONARY: INTERNAL UPDATED-BY: (RULE024 RULE02S RULE026 RULE029 RULE030 RULE031 RULE047 RULE048) RULE027 RULE032 RULE028 RULE046 HEADER [DIMENSIONAL-DESIGN-PARMS] TRANS: (*) EXPECT: ANY UPDATED-BY: (RULE0S2 RULE053 RULE054 RULEOSS) LARGE-AREA [DIMENSIONAL-DESIGN-PARMS] TRANS: (the requirement to cover a large work area) PROMPT: (Does the short-cycle performance of this task require the operator to cover a large work area?) REPROMPT: (Is the operator required to get up and move about a large work area frequently to stock parts, check equipment, etc. as part of his duties at the workplace?) 86 ASKFIRST-: T USED-BY: (RULE004 RULEOOS RULE006 RULE007 RULE008 RULE009 RULEOIO RULEOll RULE012 RULE013 RULE014 RULEOIS RULE016 RULE017 RULE018 RULE019) MALE-DATA-ERROR [DIMENSIONAL-DESIGN-PARMS] TRANS: (there is a probable error in the male anthropometric data) USED-BY: (RULE0S2 RULEOSS) UPDATED-BY: (RULE049) MHR [DIMENSIONAL-DESIGN-PARMS] TRANS: (maximum horizontal reach in centimeters) EXPECT: POSNUMB DICTIONARY: INTERNAL UPDATED-BY: (RULE024 RULE025 RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) MIXED-DATA-ERROR [DIMENSIONAL-DESIGN-PARMS] TRANS: (there is a probable error in the mixed anthropometric data) USED-BY: (RULE054 RULEOSS) UPDATED-BY: (RULEOSl) MU-HAT [DIMENSIONAL-DESIGN-PARMS] TRANS: (the estimated thigh width mean) EXPECT: POSNUMB USED-BY: (RULE035 RULE036 RULE037 RULE038 RULE039 RULE040 RULE041) UPDATED-BY: (RULE034) MVR [DIMENSIONAL-DESIGN-PARMS] TRANS: (maximum vertical reach in centimeters) EXPECT: POSNUMB DICTIONARY: INTERNAL UPDATED-BY: (RULE024 RULE025 RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) 87 NHR [DIMENSIONAL-DESIGN-PARMS] TRANS: (normal horizontal reach in centimeters) EXPECT: POSNUMB DICTIONARY: INTERNAL UPDATED-BY: (RULE024 RULE025 RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) NVR [DIMENSIONAL-DESIGN-PARMS] TRANS: (normal vertical reach in centimeters) EXPECT: POSNUMB DICTIONARY: INTERNAL UPDATED-BY: (RULE024 RULE02S RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) POPLITEALS [DIMENSIONAL-DESIGN-PARMS] TRANS: PROMPT: REPROMPT (the Sth percentile popliteal height) (For your specific population, enter the SMALLEST Sitting Popliteal Height in centimeters.) : (The Sitting Popliteal Height - is the vertical distance from the floor to the underside of the thigh immediately behind the knee.) POSNUMB EXPECT: CONTAINED-IN: (RULE024 RULE026 RULE027 RULE029 RULE030 RULE032 RULE034 RULE046 RULE048) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) P0PLITEAL95 [DIMENSIONAL-DESIGN-PARMS] TRANS: (the 95 th percentile popliteal height) PROMPT: (For your specific population, enter the LARGEST Sitting Popliteal Height in centimeters.) REPROMPT: (The Sitting Popliteal Height - is the vertical distance from the floor to the underside of the thigh immediately behind the knee.) EXPECT: POSNUMB CONTAINED-IN: (RULE024 RULE027 RULE030 RULE034 RULE046 RULE048) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) 88 POPSPEC [DIMENSIONAL-DESIGN-PARMS] TRANS: (the sex of your specific population) PROMPT: (In regard to the anthropometric measurements that you entered, what is the sex of your specific work force population?) ASKFIRST: T EXPECT: (MALE FEMALE MIXED) USED-BY: (RULE0S2 RULE0S3 RULE0S4) POPULATION [DIMENSIONAL-DESIGN-PARMS] TRANS: (the workforce population) PROMPT: (Describe the population of people who will use this workplace as: :line GENERAL MALE POPULATION - Sth to 95 th male percentiles :line GENERAL FEMALE POPULATION - Sth to 95 th female percentiles :line GENERAL MIXED POPULATION - Sth percentile female to 95 th percentile male :line YOUR SPECIFIC POPULATION - anthropometric measurements entered by you.) ASKFIRST: T EXPECT: (MALE FEMALE MIXED SPECIFIC) USED-BY: (RULEOOl RULE002 RULE003 RULE043 RULE044 RULE04S RULE0S2 RULE053 RULE054 RULEOSS) POPULATION-REDUCTION [DIMENSIONAL-DESIGN-PARMS] TRANS: (the approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem) EXPECT: (NO-REDUCTION <S% 5-10% ilO-20%^ i20-30%1[ i30-40%«i[ i40-50%11 >50%) DICTIONARY: INTERNAL UPDATED-BY: (RULE035 RULE036 RULE037 RULE038 RULE039 RULE040 RULE041 RULE042 RULE033) SHOULDERS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the Sth percentile standing shoulder height) PROMPT: (For your specific population, enter the SMALLEST Standing Shoulder Height in centimeters.) REPROMPT: (The Standing Shoulder Height - is the vertical distance from the floor to the upper-most point on the lateral edge of the shoulder with the operator standing erect.) EXPECT: POSNUMB CONTAINED-IN: (RULE025 RULE026 RULE028 RULE031 RULE032 RULE047) 89 USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) SH0ULDER9S [DIMENSIONAL-DESIGN-PARMS] TRANS: (the 95 th percentile standing shoulder height) PROMPT: (For your specific population, enter the LARGEST Standing Shoulder Height in centimeters.) REPROMPT: (The Standing Shoulder Height - is the vertical distance from the floor to the upper-most point on the lateral edge of the shoulder with the operator standing erect.) EXPECT: POSNUMB USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) SIGMA-HAT [DIMENSIONAL-DESIGN-PARMS] TRANS: (the estimated thigh width standard deviation) EXPECT: POSNUMB USED-BY: (RULE03S RULE036 RULE037 RULE038 RULE039 RULE040 RULE041) UPDATED-BY: (RULE034) SITTING-ELBOWS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the Sth percentile sitting elbow height) PROMPT: (For your specific population, enter the SMALLEST Sitting Elbow Height in centimeters.) REPROMPT: (The Sitting Elbow Height - is the vertical distance from the sitting surface to the bottom of the elbow.) EXPECT: POSNUMB CONTAINED-IN: (RULE024 RULE026 RULE027 RULE029 RULE030 RULE032 RULE046 RULE048) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) SITTING-ELB0W95 [DIMENSIONAL-DESIGN-PARMS] TRANS: (the 95 th percentile sitting elbow height) PROMPT: (For your specific population, enter the LARGEST Sitting Elbow Height in centimeters.) REPROMPT: (The Sitting Elbow Height - is the vertical distance from the sitting surface to the bottom of the elbow.) EXPECT: POSNUMB CONTAINED-IN: (RULE024 RULE026 RULE027 RULE029 RULE030 90 RULE032 RULE046 RULE048) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) SITTING-SHOULDERS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the Sth percentile sitting shoulder height) PROMPT: (For your specific population, enter the SMALLEST Sitting Shoulder Height in centimeters.) REPROMPT: (The Sitting Shoulder Height - is the vertical distance from the sitting surface to the upper-most point on the lateral edge of the shoulder with the operator sitting erect.) EXPECT: POSNUMB CONTAINED-IN: (RULE024 RULE027 RULE029 RULE030 RULE046 RULE048) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULEOOl RULE002 RULE003) STATION-TYPE [DIMENSIONAL-DESIGN-PARMS] TRANS: (the workstation type) EXPECT: (SIT STAND SIT/STAND) DICTIONARY: INTERNAL USED-BY: (RULE024 RULE025 RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048 RULE042 RULE0S7) UPDATED-BY: (RULE004 RULEOOS RULE006 RULE007 RULEOOS RULE009 RULEOIO RULEOll RULE012 RULE013 RULE014 RULEOIS RULE016 RULE017 RULEOIS RULE019) TABLE-THICKNESS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the table thickness) PROMPT: (Enter the work surface table thickness in centimeters.) REPROMPT: (The thickness of the work surface table is the vertical depth dimension of the table that is to be used as the work surface.) ASKFIRST: T EXPECT: POSNUMB CONTAINED-IN: (RULE034) USED-BY: (RULE0S6 RULE057) THIGH-CRITICAL [DIMENSIONAL-DESIGN-PARMS] TRANS: (the critical thigh-clearance height) 91 EXPECT: POSNUMB USED-BY: (RULE035 RULE036 RULE037 RULE038 RULE039 RULE040 RULE041) UPDATED-BY: (RULE034) THIGH-CUT [DIMENSIONAL-DESIGN-PARMS] TRANS: (the thigh clearance difference) EXPECT: POSNUMB UPDATED-BY: (RULE034) THIGH-POPS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the Sth percentile thigh-width / popliteal combination) EXPECT: POSNUMB UPDATED-BY: (RULE034) THIGH-P0P9S [DIMENSIONAL-DESIGN-PARMS] TRANS: (the 95 th percentile thigh width / popliteal combination) EXPECT: POSNUMB UPDATED-BY: (RULE034) THIGH-TEST [DIMENSIONAL-DESIGN-PARMS] TRANS: (there is a thigh clearance problem) USED-BY: (RULE034 RULE033) UPDATED-BY: (RULE0S6 RULE0S7) THIGH-WIDTHS [DIMENSIONAL-DESIGN-PARMS] TRANS: (the Sth percentile thigh width) PROMPT: (For your specific population, enter the SMALLEST thigh width measurement in centimeters.) REPROMPT: (The Thigh Width - is the vertical distance from the sitting surface to the top of the thigh at its intersection with the abdomen.) EXPECT: POSNUMB CONTAINED-IN: (RULE034) USED-BY: (RULE049 RULEOSO RULEOSl) UPDATED-BY: (RULE043 RULE044 RULE045) 92 THIGH-WIDTH95 [DIMENSIONAL-DESIGN-PARMS] TRANS: (the 95 th percentile thigh width) PROMPT: (For your specific population, enter the LARGEST thigh width measurement in centimeters.) REPROMPT: (The Thigh Width - is the vertical distance from the sitting surface to the top of the thigh at its intersection with the abdomen.) EXPECT: POSNUMB CONTAINED-IN: (RULE034) USED-BY: (RULE049 RULEOSO RULEOSl RULE0S6 RULE057) UPDATED-BY: (RULE043 RULE044 RULE04S) WORK-CLASS [DIMENSIONAL-DESIGN-PARMS] TRANS: PROMPT: REPROMPT ASKFIRST EXPECT: USED-BY: (the classification of the task) (How would you classify the tas performed at this workplace?) : (Example tasks categorized by Classification: :line :line Inspection, Fine Assembly, S :line LIGHT WORK - Manual As of Machine, Objects < 4.5 kg - Packing, Wrapping, Objects VDT/KEYBOARD OPERATION - Dat Secretarial, Programming, et : T (PRECISION-WORK LIGHT-WORK HEAV VDT/KEYBOARD-OPERATION) (RULE004 RULEOOS RULE006 RULEO RULEOIO RULEOll RULE012 RULEO RULE016 RULE017 RULEOIS RULEO RULE022 RULE023) k that is to be Work PRECISION WORK - oldering, etc. sembly, Load/Unload . :line HEAVY WORK > 4 . 5 kg. :line a Processing, c.) Y-WORK 07 RULEOOS 13 RULE014 19 RULE020 RULE009 RULEOIS RULE021 WSL [DIMENSIONAL-DESIGN-PARMS] TRANS: (work surface lower ht. in centimeters) EXPECT: POSNUMB DICTIONARY: INTERNAL CONTAINED-IN: NIL USED-BY: NIL UPDATED-BY: (RULE024 RULE02S RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) WSU [DIMENSIONAL-DESIGN-PARMS] TRANS: (work surface upper ht. in centimeters) EXPECT: POSNUMB 93 DICTIONARY: INTERNAL CONTAINED-IN: (RULE034) USED-BY: (RULE056 RULE0S7) UPDATED-BY: (RULE024 RULE025 RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) Domain Variables $$TITLE [DOMAIN.VARIABLES] ^^^^^- (" . ,, THE ERGONOMIST" :line •li^^ Workplace Design Expert System" :line :line :line :line :line " copyright Thomas B. DeGreve 1985") System parameters DOMAIN [SYSVARS] Value: "THE ERGONOMIST" TREEROOT [SYSVARS] Value: DIMENSIONAL-DESIGN System parameters ANTHRO-RULES [RULEGROUPS] SVAL: (ANTHROPOMETRIC DATA) CONTEXT: (DIMENSIONAL-DESIGN) Value: (RULEOOl RULE002 RULE003 RULE043 RULE044 RULE045) DATA-CHECKING-RULES [RULEGROUPS] CONTEXT: (DIMENSIONAL-DESIGN) SVAL: (DATA CHECKING) Value: (RULE049 RULEOSO RULEOSl RULE052 RULE053 RULE054 RULEOSS) DIMENSIONING-RULES [RULESGROUPS] CONTEXT: (DIMENSIONAL-DESIGN) SVAL: (WORKPLACE DIMENSIONS) Value: (RULE020 RULE021 RULE022 RULE023 RULE024 RULE025 94 RULE026 RULE027 RULE028 RULE029 RULE030 RULE031 RULE032 RULE046 RULE047 RULE048) STATION-TYPE-RULES [RULEGROUPS] CONTEXT: (DIMENSIONAL-DESIGN) SVAL: (WORKSTATION TYPE) Value: (RULE004 RULEOOS RULE006 RULE007 RULEOOS RULE009 RULEOIO RULEOll RULE012 RULE013 RULE014 RULEOIS RULE016 RULE017 RULEOIS RULE019) THIGH-CLEARANCE-RULES [RULEGROUPS] CONTEXT: (DIMENSIONAL-DESIGN) SVAL: (THIGH CLEARANCE) Value: (RULE033 RULE034 RULE03S RULE036 RULE037 RULE038 RULE039 RULE040 RULE041 RULE042 RULE0S6 RULE057) System parameters CONTEXTTYPES [PARMGROUPS] Value: (DIMENSIONAL-DESIGN) DIMENSIONAL-DESIGN-PARMS [PARMGROUPS] Value: (WORK-CLASS POPULATION STATION-TYPE WSL CHL CHU FRL FRU NVR MVR NHR MHR WSU ARMS FOREARMS B0DY-DEPTH9S SITTING-SHOULDERS SITTING-ELB0W95 POPLITEALS P0PLITEAL9S SHOULDERS SH0ULDER9S ELBOWS ELB0W9S SITTING-ELBOWS C2 CI DESIGN-TYPE FOOT-CONTROLS LARGE-AREA TABLE-THICKNESS THIGH-WIDTH9S POPULATION-REDUCTION THIGH-CUT MU-HAT SIGMA-HAT THIGH-WIDTHS THIGH-P0P95 THIGH-CRITICAL THIGH-POPS MALE-DATA-ERROR FEMALE-DATA-ERROR MIXED-DATA-ERROR POPSPEC HEADER THIGH-TEST) APPENDIX C SAMPLE CONSULTATIONS • Knowleage base .*: iHt tRGuNunrsT ^ CURRENT OBJECTIVE: THE ERGONOMIST is a knowledge-based workplace design program developed to provide sound erqonomic advice to the' industrial workstation designer. The current objective is to: (1) Decide what type of workstation to design and (2) Provide critical dimensions required to construct the workstation and (3) Determine it* there is a reduction in the accommodated population due to physical constraints. ... end — press RETURN Up Down CF Unknwn Done — Why How Help Undo Stoo , Fl F2 F3 F4 F5 F6 F7 F3 F9 FIO ESC , How would you classify the task that is to be performed at this workplace? Select one of the following: (Press T 3 ' for selection help.) > PRECISION-WORK LIGHT-WORK HEAVY-WORK VDT/KEYBOARD-OPERATION Up Down CF Unknwn Done — Why How Help Undo Stop Fl F2 F3 F4 F5 Fb Ff F8 F9 FIO ESC 95 Describe the population of people who will use this workplace as: GENERAL HALE POPULATION - Sth to 95 th male percentiles GENERAL FEMALE POPULATION - Sth to 95 th female percentiles GENERAL MIXED POPULATION - Sth percentile female to 95 th percentile male YOUR SPECIFIC POPULATION - anthropometric measurements entered by you. Select one of the following: (Press T 3 ' for selection help.) ^ MALE >FEMALE MIXED SPECIFIC Uc I p uown LP Unknwn Done — Why How Help Undo Stop . Fl F2 F3 F4 F5 F6 FT F8 F9 FIO ESC knowledge For this task, is the operator required to use any foot controls? Select one of the following: (Press ^F3' for selection help.) >YES NO Uc Stot p Down CF Unknwn Done — Why How . Help Undo stop . Fl F2 F3 F4 F5 F6 FT F8 F9 FIO ESC Does the short-cycle performance of this task require the operator to cover a large work area? Select one of the following: (Press T S " for selection help.) YES > N 0 UG Down F2 CF Unknwn F3 F4 Done F5 — F6 fF How F8 Help Undo F9 FIO Stop ESC ^^^A^^^^^^^MW^/v^M^^A^^MVM^MM^^MM^n^^MWinA'/jv^jvwd 96 97 What type of design do you wish to construct? FULLY ADJUSTABLE - with adjustable worksurface, chair, and footrest FlXg WpRK^SURFACE - with adjustable chair and footrest PIJED (>WIR - with adjustable worksurface and footrest STATIONARY - with no adjustable features Select one of the following: (Press T 3 ' for selection help.) FULLY-ADJUSTABLE ^ FIXED-WORK-SURFACE > FIXED-CHAIR STATIONARY Up Down CF Unknwn Done — Why How Help Undo Stop , Fl F2 F3 F4 F5 F6 F T F8 F9 FIO ESC MMFLACE-1 is as follows: The following conclusions were made by THE ERi^NOHIST concerning the current workplace design. The workstation type is as follows: SIT Work surface upper ht. in centimeters is as follows: 83.0 Work surface lower ht. in centimeters is as follows: 74.4 Chair upper height in centimeters is as follows: 43.6 Chair lower height in centimeters is as follows: NOI€ Footrest upper height in centimeters is as follows: 6.3 I was unable to make any conclusions regarding footrest lower height in centimeters. Normal vertical reach in centimeters is as follows: 94.9 Maximum vertical reach in centimeters is as follows: 156.0 Normal horizontal reach in centimeters is as follows: 18.7 Maximum horizontal reach in centimeters is as follows: 46.4 The approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is as follows: NO-REDUCTION CURRENT OBJECTIVE: TIC ERGONOMIST is a knowledge-based workplace design program developed to provide sound ergonomic advice to the industrial workstation designer. The current objective is to: 1 Decide what type of workstation to design and (2) Provide critical dimensions required to construct the workstation and (3) Determine if there is a reduction in the accommodated population due to physical constraints. ... end — press RETURN 98 Up Down CF Unknwn Done — Why How Help Undo Stop , Fl F2 F3 F4 F5 F6 FT F8 F9 FIO ESC How would you classify^ the task that is to be performed at this workplace? Select one of the following: (Press ''F3' for selection help.) PRECISIOfHJORK LIGHT-WORK >HEAVY-WORK VDT/KEYB0AR1>-0PERATI0N Up Down CF Unknwn Done — . W h y How , Fl F2 F3 F4 F5 F6 FT F8 ., . ., Help Undo F9 FIO r.v//.\wi Describe the population of people who will use this workplace as: __ GENERAL MALE POPULATION - Sth to 95 th male percentiles GENERAL FEMALE POPULATION - Sth to 95 th female percentiles GENERAL MIXED POPULATION - Sth percentile female to 95 th percentile male YOUR SPECIFIC POPULATION - anthropometric measurements entered by you. Select one of the following: (Press T 3 ' for selection help.) >MALE FEMALE MIXED SPECIFIC Up Down CF Unknwn Done — wny Fl F2 F3 F4 F5 F6 FT jFS F9 ' FIO ESC Why How ^too Help Undo ^ n r^ rj r«t rg FQ r/ jro F9 FIO coî ^ 99 For this task, is the operator required to use any foot controls? Select one of the following: (Press ^F3' for selection help.) ^ YES > N 0 Up Down CF Unknwn Done — Why How Help Undo Stop Fl F2 F3 F4 P5 F6 FT F8 F9 FIO ESC • Know ledge Base :: THt tRGuNUnrbT ^ Does the short-cycle performance of this task require the operator to cover a large work area? Select^one of the following: (Press ^F3' for selection help.) >YES * NO Up Down CF Unknwn Done — Why How Help Undo Stop , Fl F2 F3 F4 F5 F6 FT F8 F9 FIO ESC What type of design do you wish to construct? FULLY ADJUSTABLE - with adjustable worksurface, chair, and footrest . , . . FIXED WORK SURFACE - with adjustable chair and footrest FIXED CHAIR - with adjustable worksurface and footrest STATIONARY - with no adjustable features , , .̂ ^ , ^ Select one of the following: (Press T 3 ' for selection help-.) FULLY-ADJUSTABLE FIXED-WORK-SURFACE FIXED-CHAIR >STATIONARY Up Down CF Unknwn Done — Why How Help Undo Sto , Fl F2 F3 F4 F5 F6 F7' F8 F9 FIO £SL 100 *"«SM5a55'ifi5?«'fln WAW/WWJWA'A'JWf Enter the work surface table thickness in centimeters. Enter a positive number. 10 Down CF Unknwn Up uown u- unKnwn uone — wny Fl F2 F3 F4 F5 F6 F T Done Wh\ How Help Undo F9 FIO Sto M W//^JVWJV/J WORKPLACE-1 is as follows: The following conclusions were made by Tl£ ERGONOMIST concerning the current workplace design. The workstation type is as follows: STAND Work surface upper ht. in centimeters is as follows:. 107.4 Work surface lower ht. in centimeters is as follows: NONE Chair upper height in centimeters is as follows: NONE Chair lower height in centimeters is as follows: NOI€ Footrest upper height in centimeters is as follows: 7.4 Footrest lower height in centimeters is as follows: NOfE Normal vertical reach in centimeters is as follows: 150.1 Maximum vertical reach in centimeters is as follows: 213.7 Normal horizontal reach in centimeters is as follows: 20.1 Maximum horizontal reach in centimeters is as follows: 51.3 The approximate reduction to the population that is to be accommodated at this workplace, due to a thigh clearance problem is as follows: NO-REDUCTION y iCnowieoge Base '•'• i r t twOuNUrfriT ^ J CURRENT OBJECTIVE: J THE ERGONOMIST is a knowledqe-based workplace design proaram , developed to provide sound ergonomic advice to the' ^ industrial workstation designer. y J The current objective is to: . (1) Decide what type of workstation to design and . (2) Provide critical dimensions required to construct the r workstation and . (3) Determine if there is a reduction in the accommodated ^ population due to physical constraints. press RETURN h h h « U f Down F2 CF Unknwn F3 F4 Done F5 F6 Whv How F8 Help Undo F9 FIO Stoo ESC .•WAV^MV.VAVA"A"/WA'A"/^.VAV/AV.V/^/MV/AV.V.»AV.V.V.\VAV.*JVAV.V.VJ ^.VV'jTANVV/AW/AV.-M^AVAViffVi'JAVAVAS'AV^^^^ r Ĵiowiedqe Base :: THt tKbiiNunrsi ? J How would you classify the task that is to be performed at • this workplace? • S e l e c t one of the followinq: (Press 'FS' for selection help.) •• PRECISION-WORK LIGHT-WORK HEAVY-WORK VDT/KE'fBOARD-OPERATION Stop ESC J Up Down CF Unknwn Done — Why How >He!0 Undo J Fl F2 F3 F4 FS F6 F7^ F3 F9 -10 •V/^WAV//^M^WAV/^^*VAVAVA»AV^/A»^AVAVAWiVAVA»AVA»AV.VA«AVAVA'JIl >AVAVAViV;'A»ANVAWA"iCi%Y,!A%%V^^AlAVAVA»A'.VAVA%V."^.V/AVA%VAVA»^ 7 Knowledge case :: THE tRGUNijnibi J Example tasks categorized by Work Classification: ?• PRECISION WORK - Inspection, Fine Assembly,'Soldering, jtc. r LIGHT WORK - Manual Assembly, Load/Unload of Hachinel • Objects C 4.5 kq. * HEHVY WORK - Packing. Wrapping. Objects > ̂ .5 kg. J VDT/KEYBOARD OPERATION - Data'Processing, Secretarial, r Programming, etc. h k h h k h k k k k k t elect one of the following: (Press T3'' for selection help.) PRECISION-WORK LIGHT-WORK ^ HEAVY-WORK > VDT/KE/B0ARDH3PERATI0N Up Fl Down C-7 CF Unknwn F3 F4 L'one F5 F6 Whv F7' How F8 He ID 'Jndo F9 FIO zzo:; .•AV/AWiTAW.VAV.VAWAV.VAV.V.'AVJ'̂ .VAV̂ AV.' I •J 101 J'A'AVAV.yA-ANV.V.yCAVA&W.ViW/AV.VAW^^^^ ,. N.r.cijji3gge case : : IHE tNijuNtjMiST 7 Describe the population of people who will use this aorKpiace as: GENERAL ?1ALE ?i3PULATI0N - 5th to 95 th male percentiles IJENERAL FEMALE POPULATION - Sth to 95 th female percentile^ GENERAL MIXED POPULATION - Sth percentile female to 95 th percentile male ' YOUR SPECIFIC POPULATION - anthropometric measurements entered by you. Select one of the following: (Press 'FS' for selection help.) MALE ^FEMALE > MIXED 102 SPECIFIC Done F5 — Why How F6 F7^ F8 Help Undo F9 FIO StOD ESC Up Down CF Unknwn Fl F2 r3 F4 VAVAVA%VA"AV/AVJ'J'^AVA'/A'.VAVAr/J»d'/^/^A'AV^AVAVA\VAWJV/AV,V.V^^^ • m For this task, is the operator required to use any foot controls? Select one of the following: (Press ''FS' for selection help.) YES > N 0 Done F5 — - F6 Why How F8 Help undo F9 FIO Stoo ESC Up Down CF Unknwn Fl F2 F3 F4 AVAVA»J"^iPJ'//^.»A'A»tV/A'AVAV/AViV^AVAVAW/AV.'yVAV^/.^A»//A».V^/i^AVA*J Does the short-cycle performance of this task require the operator to cover a large work area? S-̂ iect one of the following: (Press "FS' for selection help.) > N 0 JC'ijjn ijnKnwn -4 Done ^5 — Why How F6 n' Fo AVA'AVAVAV.V nelo 'jrico i t CO •i 103 Ĵ VA%%VytVA'AVAVAWA»<A%V»ViVW»VAVA%«̂ ^̂ ^̂ ^ • Knowledgebase :: l^tRedfltjKiSi * What type of design do you wish to construct? PJLLY AD.JUSTABLE - with adjustable worksurface, chair, and footrest FIXED WORK SURFACE - with adjustable chair and footrest FIXED CHAIR - with adjustable worksurface and footrest STATIONARY - with no adjustable features Select^one of the following: (Press "FS" for selection help.) > FULLY-ADJIJSTABLE FIXED-WORK-SURFACE FIXED-CHAIR STATIONARY r^ Down F2 CF Unknwn F3 F4 Done F5 F6 Why F7 How F8 Help Undo F9 FIO Stop ESC .W.W/AV/A%ViVJ'/^/AV^^AWJ'A»/A"JVJ»ir/A"AW/^AV^/J"A%VAVAVAVAV.VAV.'J WORKPLACE-1 is as follows: The following conclusions were made by THE ERGONOMIST concerning the currenfworkplace design. The workstation type is as follows: SIT Work surface upper ht. in centimeters is as follows: 76.1 Work surface lower ht. in centimeters is as follows: 56.1 Chair upper height in centimeters is as follows: 48.7 Chair lower height in centimeters is as follows: 37.3 Footrest upper height in centimeters is as follows: 11.4 Footrest lower height in centimeters is as follows: 0 Normal vertical reach in centimeters is as follows: 88.6 Maximum vertical reach in centimeters is as follows: 149.7 Normal horizontal reach in centimeters is as follows: 15.5 Maximum horizontal reach in centimeters is u follows: 43.2 The acco pr e approximate reduction to the population that is to be commodated at this workplace, due to a thiqh clearance obi em is as follows: NO-REDUCTION ^ ..nojjiijqe case i : 'rE Eh;buNOni';i . CURRENT DSJEC'IVE: 'rtE ERGONOMIST is a knowledge-based 'oiorkplace design program developed to provide sound ergonomic idvici to '"''»' industrial workstation designer. The current objective is to: •U S®*^ide what.type of workstation to design and ui Provide cntitai dimensions required to'construct the •^lorkstation and 13) Determine if there is a reduction in the accommodated population due to physical constraints. ... end — press RETIJRN UP Down CF Unknwn Fl F2 F3 F4 Done FS — Why How F6 F7' He ID Undo ^m AVt"A'.VAV.VbV/A"AV^/^/AVA%V.VA«A'A%VA%VA%VAVAVASWAVAVAV.V.VAVJ J'.'AVVAVNy.VAV.V.WiVA'iAiV.W.W^^ • •̂ •nowiedge Base :: iHE arnwIoHrSi ^ How would you classify the task that is to be perforTied at this workplace? Select one of the followinq: (Press 'FS'' for seiecTion help.) ^ PRECISION-WORK > LIGHT-WORK HEAVY-WORK VDT/KE'/BOARD-OFERATION Up Down CF Unknwn Done Fl F2 F3 F4 F5 — Whv How F6 F7^ F3 ^iiD Undo • - , ) • • • " ":-r I. /A"AVi'AV/AVAV«»AVJ'J'AV///AVAViVJ»AVAS%VAVAVAVA'A"AVAVAV.V.'.VAVJ ^VAVA^VA•A^^VAVA'A•^iVA>^y^VAVAVAVAVAVAV.^^^VAVA^^v.^^^'^^^^^^^ K.nvJjjiecce case '•' THE tKbijNtjniji Describe the population of pec-p:̂ -jho :ni:! use this wor-itD'ace as: '3ENEF:AL flALE POPULATION - Sth t : 95 th inai^ p-i-Cirtiies GE'MERAL FEMALE 'POPULATION - Sth t:i 'ij th femaie osf^centi.es GENERAL HIXED POPULATION - oth oercantiie female to ^5 t" percentile male YOUR SPECIFIC POPULATION - anthropometric measurements entered py vou. Select one of the fclloniina: iPf̂ ess "'Fl' for selection h e b . • <̂ALE > (IIXED cpcr TPTP UD Do'a/ri CF unKnwn Done - ~ ^hy Fl *2 Fl r^ F5 Fk p " ' i i - 11' I 104 1 I I f t 1 1 1 t i t I I f • i i i t t i i i i i i a i i i i J i i i i i i i i i i i i i i i t i t i i i t t i i i . i i i i i i i i t t t i i i i i i i a l |AVAN%SVA%%yAVA%Vyfy,!,yy/,W/AVAVAVAV.VAV,V,%VAV.V.SV.V.VAV.V/AV^ • ^r!Owledge base :: \Kt iKi-UNUMi;;:i ^ J For t n i s t i S K , is trie cp<iraror r e q u i r e d to use any foot 7 -lontrois*' "Select one of the followinq: (Press 'F3' for seiectior'. r,ai:'.f I NO' fa k k k k k k k k k k k k Y Up Down CF Unknwn Done — > Why How J Fl F2 F3 F4 FS F6 F7" F3 .•A'.VAVA"AVASSVAVAVAV.VASVA%VAVAViV.VAVAViVAV.VAV.VAV,VtV,VAVJ >tVAViV/iyi^/J^AVAVA'A'.Vi%V«iVjV«VAVAVAVAVAV.VAVAV.VAV.V.'^ J KT.cwie'̂ geBase : : I h ^ crtiJj^jdnriT z. J For this task, is the operator required to use anv foot r controls? f Why this question is needed: 105 F? FiO •'129 ^/Jhether the requirement to *. * the w o r k s t a t i o n type use f o o t controls i s needed to determine ; RULEOOS J If 1) the classification of the task is LIGHT-WORK, and * 2) the requirement to use foot controls, and * 2) the requirement to cover a larqe aorK area. * Then it is definite (1007.) that the workstation tvpe is J SIT/STAND. I ... end - press RETURN k k k »• lip " F l Down CF Unknwn F2 F4 Done - - - Why How Heip J^J fO F7 l-V \vAVA'iPAVAV.VA'^^^^AVAVAVAV.VAVAV.VAyAVAViVAV.y,y.y.V.V.yAVA%^^^ >AVAVAVyA'C^AV.V.-AW4i,>ytW^ViW'*"W^^^ Knowî qqe aase : : irE tnijijTTiJMibi ^ ** For this task, is the operator required to use my foot ^ "on^Tois? ^ Sei^'-t one of the followinq: (Press "FS-' tor selection nel?. J >N0 k k k k k k k k k k k k k ^ l!p D''iyri i"F 'Til-nujn Done — '-'Jhv roiij fc pi ' r - r-.-, F4 Done — ''Jhv F5 F6 F7' .'r j . : - : u . . . • • • • • • • . . . • • . . • • • • • • • • i i i i i i i i i i i * i i ' * * * * ' * * * * * * * ' * ' * ' * i * * i i i * * * < i * * i i i i i i 1 g V i i i V i V i i 1 1 1 1 i / t ^ i y A A i n I 1 1 1 I • i i i t i i i i i n i i t i i i i i i i i i t i i i i t i t t i i i i i i i i . i j i twd ^AVAVAVpVi*AVAVA'A'A'AiyAViV»NViVAy.VAVAV.VAV.VAy.'.yi'.V.V.V.'.VAy.J • r.riowieqge sase '•'• :rtE-.-.biJwjniST ^ ^ Does the short-cycle jerforn^ance of this task requiri: t'l^ . operator to cover a large 'jjork area? . S e l e c t one of tr.e foliowinq: (Press "F3' for S'jiection n e l p . ' J >M0 k k k k k k k k k k k k k J Up Down CF Unknwn Done — Why How Help Undi J Fl F2 F3 -4 FS F6 F7' ^3 F9 FIO ASV.»AViy.V//A%SSViV.VAVAV^AVAVAV.V«V/AVAVAVAVAVAViVAVAVAyAVi J'AVAVAV/Asiv.VAVA*AyAyjCf!«iViyAv,v.yiViy,y.v.v.vAv.y,VAV.yAv.«.yAv| J knowledge ease :: i H r ^ r a j ? u : ? i ^ ^ What type of design do you wish t: constr'jct'' J RJLL/ AD'JUSTABLE - with adjustable work:sur-ace, chai:-. an: J FIXED WiDRK SURFACE - t'ith adjustable chair and footres- r FIXED CHAIR - with adjustable !i)0'"-ksur̂ "ace anc fvitrest STHTIO! 106 :. f,-, •-, BV - k k k k k k k k k k k k *> e i liiith no adjustabie_feat:r^s '-^iiart one <^^ th(» t̂'?:'••'wing: -Fr-ess ''"2' " ' PJLLY-ArJUSTABLE ' >'IXcD-wORK-SURFACE -iXtD-'-nAiH STATICNARY ; l i • li r *• •« A f: Down CF Unkri'jjn F4 r2 F3 Done — Wiy FS F6 10 w r5 -:U AVAy.v.vAVAy.yAyAv^Av.vAV,v/.y.v.vAVisy.yAyAv.vAv,y.vAy.y,sv.v,vj >,vAVAVAy.'iVAy.yi'A'.y,>wyiyj:»yAv,y.Nv.sv.vA^^^^^ Entir the '-'or-'k surface table tî 'iikr-i! Enter a oositive '•lumber. - i r . f ' T. u ••• • j .* " k ii r. ir. Mown CF u n n i j n Do^* — '''•'V HO Ĵ; -ii-' • ^ 2 =•: P i ~z -z - '• - 3 F'' ,.̂ .,,,.,.,.,̂ v,.,̂ v,v,v,̂ y.̂ v.y.̂ ^ ŷ.y.̂ v̂.̂ y.y.sy.̂ v̂.̂ ^^^^^^^^^^^^^^ •̂̂ 107 WORKFLACE-i is as fvl'ows: The ^'oi'owing I'.nclusions aiere '^a;* by '¥£ ERGONOMIST concerning the currint'Lorkolace Jesign. The workstation type is as ^o'lowss SI'/STAND Work surface upper ht. in lentiiTiet^rs is as f o l l o ' i s : 115.4 Work surface 'ower !-it. in centirteters is as follows: NONE f^ C'-'air uooe''- height in centimeters is as *v;lows: iOl.c Chair lo'der height in centinieters is as fo::0'js: '3.0 Foot'est ijppe-- nei'int in 'isntiTieters is as ^OMCWSJ :"*.? "ootr-ist ioujer heioht in ^lentimeters is as fol^oiis? 0 M.ir'i»ai i - « - i r a ! icai reacr- in 'lentiiTiete's i; 1 5 -0l]0ii!S5 \52.-' I^aximum vertical reacri in centimeters is as foiiowsJ 205.4 Normal horizontal reach in centimeters is as fjllows: 15.5 ^'aximum horizontal reach in centimeters is as follows: *:.'. "hi approximate reduction to the population that is to :-i accommodated at this workglace, due to a thigh clearance orobiem is as ̂ 'oTows: y'jOl APPENDIX D DEFINITION OF ANTHROPOMETRIC MEASUREMENTS (A) Shoulder Height - the vertical distance from the floor to the upper most point on the lateral edge of the shoulder with the operator standing erect. (B) Sitting Shoulder Height - the vertical distance from the sitting surface to the upper most point on the lateral edge of the shoulder with the operator sitting erect. (C) Body Depth - the maximum horizontal distance between the vertical planes passing through the most anterior and posterior points on the trunk. (D) Thigh Clearance - the vertical distance from the sitting surface to the top of the thigh at its intersection with the abdomen. (E) Forearm Length - the horizontal distance from the tip of the elbow to the tip of the longest finger. (F) Arm Reach - the horizontal distance from the posterior surface of the shoulder to the tip of the extended middle finger. (G) Elbow Height - the vertical distance from the floor to the depression at the elbow between the bones of the upper arm and forearm. (H) Sitting Elbow Height - the vertical distance from the sitting surface to the bottom of the elbow. (I) Popliteal Height - the vertical distance from the floor to the underside of the thigh immediately behind the knee. Refer to Figure 4 for a sketch of these measurements. 108 109 FIGURE 4. Sketch of Anthropometric Measurements 
work_rcxb7tpvibdetk4dirtapg2vjy	----	doi:10.1016/j.pss.2007.09.006 ARTICLE IN PRESS 0032-0633/$ - se doi:10.1016/j.ps � Correspond E-mail addr Please cite thi j.pss.2007.09.0 Planetary and Space Science ] (]]]]) ]]]–]]] www.elsevier.com/locate/pss The search for life beyond Earth through fuzzy expert systems R. Furfaro a,� , J.M. Dohm b,c , W. Fink d , J. Kargel b , D. Schulze-Makuch e , A.G. Fairén f , A. Palmero-Rodriguez h , V.R. Baker b,c , P.T. Ferré b , T.M. Hare g , M.A. Tarbell d , H. Miyamoto i , G. Komatsu j a Aerospace and Mechanical Engineering Department, University of Arizona, Tucson, AZ 85721, USA b Department of Hydrology and Water Resources, University of Arizona, Tucson, AZ 85721, USA c Lunar and Planetary Laboratory, University of Arizona, Tucson, AZ 85721, USA d California Institute of Technology, Visual and Autonomous Exploration Systems Research Laboratory, Division of Physics, Mathematics and Astronomy, Pasadena, CA 91125, USA e Department of Geology, Washington State University, Pullman, WA, USA f NASA Ames Research Center, Space Science and Astrobiology Division, Moffet Field, CA 94035, USA g United States Geologic Survey, Flagstaff, AZ, USA h Department of Earth and Planetary Science, University of Tokyo, Tokyo 113-0033, Japan i Department of Geosystem Engineering, University of Tokyo, Tokyo, Japan j International Research School of Planetary Sciences, Università d’Annunzio, Pescara, Italy Received 19 October 2006; received in revised form 2 September 2007; accepted 14 September 2007 Abstract Autonomy will play a key role in future science-driven, tier-scalable robotic planetary reconnaissance to extremely challenging (by existing means), locales on Mars and elsewhere that have the potential to yield significant geological and possibly exobiologic information. The full-scale and optimal deployment of the agents employed by tier-scalable architectures requires the design, implementation, and integration of an intelligent reconnaissance system. Such a system should be designed to enable fully automated and comprehensive characterization of an operational area, as well as to integrate existing information with acquired, ‘‘in transit’’ spatial and temporal sensor data, to identify and home in on prime candidate locales. These may include locales with the greatest potential of containing life. Founded on the premise that water and energy are key to life, we have designed a fuzzy system that can (1) acquire the appropriate past/present water/energy indicators while the tier-scalable mission architecture is deployed (first layer), and (2) evaluate habitability through a specialized fuzzy knowledge-base of the water and energy information (second layer) acquired in (1). The system has been tested through hypothetical deployments at two hypothesized regions on Mars. The fuzzy-based expert’s simulation results corroborate the same conclusions provided by the human expert, and thus highlight the system’s potential capability to effectively and autonomously reason as an interdisciplinary scientist in the quest for life. While the approach is demonstrated for Mars, the methodology is general enough to be extended to other planetary bodies. It can be readily modified and updated as our interdisciplinary understanding of planetary environments improves. We believe this work represents a foundational step toward implementing higher-level intelligence in robotic, tier-scalable planetary reconnaissance within and beyond the solar system. r 2007 Elsevier Ltd. All rights reserved. Keywords: Fuzzy logic; Tier-scalable mission autonomy; Astrobiology; Search for life; Robotic planetary exploration; Autonomous planetary reconnaissance e front matter r 2007 Elsevier Ltd. All rights reserved. s.2007.09.006 ing author. Tel.: +1 520 3127440; fax: +1 520 6218191. ess: robertof@email.arizona.edu (R. Furfaro). s article as: Furfaro, R., et al., The search for life beyond Ear 06 1. Introduction Even though orbiting spacecraft and ground-based landers and rovers have successfully collected significant data through instrument suites, working hypotheses are yet th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ www.elsevier.com/locate/pss dx.doi.org/10.1016/j.pss.2007.09.006 mailto:robertof@email.arizona.edu dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]2 to be confirmed, such as in the case of Mars, whether sites of suspected hydrothermal activity are indeed hydrothermal environments, or whether prime candidate sites of potential life-containing habitability (Dohm et al., 2004) actually contain life. In answering the question of how best to explore planetary environments such as the ancient mountain ranges of Mars that contain complex structure and magnetic signatures (Acuña et al., 1999; Dohm et al., 2001a, b, 2002; Connerney et al., 2005), the expansive martian canyon system of Valles Marineris (Scott and Tanaka, 1986), the putative ocean beneath the ice of Jupiter’s moon, Europa (Carr et al., 1998), a scientific mission concept for remote planetary surface and subsurface reconnaissance has been recently devised (Fink et al., 2005a–c, 2006a, b, 2007) to replace the engineering and safety constrained mission designs of the past. This novel mission concept is a paradigm shift from traditional missions that perform either local ground-level reconnaissance via immobile landers or rovers of limited mobility, or global mapping through orbiters, to what is termed tier-scalable reconnaissance in scientific missions. The paradigm integrates multi-tier (orbit–atmosphere–surface/subsurface) and multi-agent (orbiter–blimps–rovers–immobile sensors) hierarchical mission architectures to enable unconstrained, science- driven and intelligent planetary exploration (Fig. 1, Fink et al., 2005a). The full-scale and optimal deployment of agents as part of a tier-scalable mission architecture requires both the integration of design and architecture and the implementa- tion of an intelligent reconnaissance system. Such a system should (1) include software packages that enable fully automated and comprehensive identification, characteriza- tion, and quantification of feature information within an operational region (e.g., with the Automated Geologic Field Analyzer (AGFA) by Fink et al., 2005d; now Automated Global Feature Analyzer, see Fink et al., 2008) with subsequent target prioritization and selection for close-up reexamination (e.g., Fink, 2006c); and (2) integrate existing information with such AGFA-acquired, ‘‘in transit’’ spatial and temporal sensor data to automatically perform smart planetary reconnaissance and identify and home in on prime candidate life-containing targets on planetary bodies. Our goal is to define the framework for the design of an expert system (intelligent reconnaissance system) to be integrated into the tier-scalable reconnaissance mission architecture and more generally to provide a basis for implementing intelligent algorithms capable of autono- mously performing reasoning over data collected during planetary exploration. Here, we propose a fuzzy logic- based expert system capable of simulating the combined approach of a geologist, biologist, and chemist to identify, in one instantiation, prime candidate locales of life- containing potential through tier-scalable reconnaissance, which includes comparative analysis of the spatial and temporal spaceborne-, airborne-, ground-, and subsurface- Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 based observations, i.e., synthesis of stratigraphic, paleotectonic, topographic, geomorphologic, hydrologic, geophysical, geochemical, spectral, and elemental informa- tion. As explained in the subsequent sections, the system is designed to acquire a sequence of life-habitat indicators whose values are inferred from the sensor data streaming through the system. The indicators, which are specific to the planet under observation, are the input to the system and are manipulated by the fuzzy rules, which form the core of the knowledge-base, in order to infer new facts. The number and type of indicators vary with the planet being explored. For example, indicators suitable for identifying life-potential locales on Mars may be different than those required by the ice-covered Europa where conditions for life may be drastically dissimilar. The fuzzy logic knowledge-base we are creating for tier- scalable autonomous reconnaissance is based on the geologic mapping-based reconnaissance and related synthesis of information of Mars, but can be modified/ revised and updated for other planetary bodies. The intelligent system presented here may be considered a critical part of the ‘‘brain’’ of the tier-scalable reconnais- sance mission architecture, though there may be other possible designs and implementations schemes. Impor- tantly, we intend to show the potential of the methodology, leaving an open forum for the community to discuss about what are the appropriate life indicators for different planetary scenarios. Here, Mars is used as case study to show how the fuzzy logic framework can be employed to design autonomous systems capable of assessing planetary habitability. Starting from the premise that water and energy are key ingredients to life, we devised a two-layer fuzzy system equipped with appropriate knowledge (i.e. rules) as derived from field expertise. The two layers are inherently interconnected: the first layer is comprised of four independent fuzzy systems, each devised to acquire remote and in-situ information and evaluate the potential for past/present water and energy. The second layer is comprised of rules that arrange water and energy (both past and present) in order of importance. For example, if the first layer infers that potentials for past/present water and energy are all high, then the potential for the site to harbor life will be very high. Life- related rules at lower level of importance have been included to provide the system with a comprehensive understanding of Mars habitability for effective and intelligent prediction. The paper is organized as follows: Section 2 discusses the general problem of designing expert systems for planetary exploration. Section 3 makes a case for using fuzzy logic and provides a review of the subject. In Section 4, the rationale for designing the fuzzy expert is outlined and its implementation is shown in Section 5. Simulations and system testing for two hypothesized martian scenarios are illustrated in Section 6. Conclusions and future efforts are outlined in Section 7. th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Fig. 1. Schematic illustration of the ‘‘tier-scalable’’ mission concept (Fink et al., 2005a). The paradigm integrates multi-tier and multi-agent hierarchical mission architectures to enable unconstrained, science-driven planetary exploration. The depicted three-tier architecture is comprised of: (1) a ground tier composed of multiple miniaturized and expendable agents with complementary sensor suites that can transmit the sensor information back to a main data compilation hub for comparative analysis; (2) an airborne tier comprised of blimps and/or balloons to control and command the surface/subsurface reconnaissance agents; (3) a spaceborne tier comprised of multiple orbiters which command the airborne tier, receive the data transmitted from the airborne- and ground-based agents, and interact with Earth-based systems (Fink et al., 2005a–c, 2006a, b, 2007). This novel paradigm introduces new implications for space exploration, including increased mission safety, mission reliability, and science return, which in turn enable new scenarios for the exploration of various planetary bodies. R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 3 2. Expert system design for planetary reconnaissance: problems and solutions The design of an expert system to assess the potential for habitability (PH) through a tier-scalable reconnaissance approach is a knowledge engineering problem: given the domain-containing input data, find a suitable solution Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 among all possible candidates occurring in the solution space. In our case, ‘‘input data’’ consist of information collected via multi-scale and multi-sensor deployment coupled with existing information, while ‘‘solution’’ is the answer to the following question: ‘‘What is the potential for a particular locale under investigation to harbor life?’’. th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]4 If the answer to this question must be given in an automatic fashion, specific field-based knowledge must be implemented on a computer. In defining an expert-system approach to the problem, knowledge is one of the major aspects of the procedure. Knowledge is ‘‘condensed information’’ and for our purposes consists of a set of rules or methods from which it is possible to perform plausible reasoning to obtain new facts, and to formulate new hypotheses while testing existing hypotheses. Gener- ally, the knowledge domain consists of a set of definite rules that can be coded into computer language. One other important aspect that must be considered is the inference mechanism. Inference can be defined as the process of matching data and knowledge to determine a solution to the problem (i.e., infer new facts and/or map around a problem, ultimately addressing the problem). Knowledge is usually acquired and integrated via the inference mechanism. To implement reasoning over knowl- edge, a theoretical framework is required (i.e., the need to define the appropriate logic). The selected logic will help determine the chain of matching results created to provide a suitable solution to the problem. Fig. 2 shows a flow diagram that provides the general architecture of an expert system suitable for the tier- scalable reconnaissance mission concept. The various blocks in the diagram represent the essential ingredients. The knowledge-base contains information acquired via interaction with a field expert required to solve the Knowledge Acquisition Inference Engine Explanation Database Knowledge Base Space Control Expert System Architecture Planetary Geologist (Expert) Fig. 2. General architecture of an expert system p Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 problem. The database contains information about existing and in-transit acquired facts (i.e., information coming from the sensor after processing coupled with existing informa- tion). The inference engine is the place where the inference mechanism for reasoning over knowledge and data is implemented. The results of the inference are provided to the spaceborne command and control module, which selects the course of action based on the inferred information. The latter is usually implemented in the spaceborne tier and distributed through commands sent to the tiers beneath according to a hierarchical scheme. The explanation module is important because it provides explanation about how and why new information has been generated. This module can be interfaced with Earth-based centers where the explanation provided by the reconnais- sance system is monitored. The Earth-based center can override the spaceborne control at ‘‘any’’ time (aside from the communication time lag) if there is an error in the inference. The expert system paradigm defined above is general and requires appropriate design, which includes definition of knowledge, data inputs, and inference mechanisms. In the following, we will focus on the two major parts of the system: knowledge-base and inference. It should be pointed out that different choices of those two ingredients may yield different solutions for the intelligent reconnaissance system. In this regard, we believe that an expert system based on fuzzy logic may be a suitable candidate for borne System roposed for tier-scalable mission deployment. th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 5 an autonomously operating tier-scalable reconnaissance mission. 3. Fuzzy logic paradigm for autonomous fuzzy expert system design The proposed system bases its foundation on the concept of fuzzy sets and fuzzy logic as an inference mechanism. In the following, an overview of fuzzy logic concepts, required for an effective system design, is provided for a basic understanding of the subject. 3.1. Why fuzzy logic for planetary exploration? Planetary exploration is about understanding planetary bodies via remote and in-situ sensing. New discoveries are made using newly acquired data to infer new and unknown facts, as well as to test working hypotheses and to formulate new hypotheses. Formulated working hypoth- eses become closer to reality when diverse layers of information corroborate the formulated hypothesis (con- ciliation of diverse information), upping the potential of discovery, especially through interdisciplinary investiga- tion. The tier-scalable reconnaissance missions strategy provides the means to acquire multi-scale and multi-sensor information (e.g., stored in a multi-layer-information system (MLIS), introduced by Fink et al., 2005a–c) that must be appropriately synthesized to evaluate certainty about possible new facts and findings. Nevertheless, general statements such as ‘‘the observed site contains life’’ cannot be stated as true or false, as classical and propositional logics require, until certainty is guaranteed. Unfortunately, global scale measurements might be incon- clusive and require the concentration of efforts in a defined area where newly acquired information coupled with existing knowledge gives us clues (but not necessarily certainty) about, for example, presence of water and/or presence of life. Moreover, a statement such as ‘‘elevated hydrogen content is indication of an elevated presence of water’’, although rather general, is a linguistic expression of the common reasoning performed by planetary geolo- gists in analyzing data. The knowledge system of planetary scientists, expressed in such statements, can be incorpo- rated in the knowledge-base of an expert system for automatic reasoning. The fuzzy logic semantic provides an ideal framework to deal with independent layers of information of varying degrees of confidence such us elevated hydrogen content, low sulfate level, and medium number of sapping channels. The absolute values of the input data can be transformed into fuzzy values and incorporated in rules that deal with concepts/working hypotheses of varying degrees of con- fidence (depending on the layers of information that are collectively consistent with the hypothesis), typical for planetary exploration. Thus, fuzzy logic provides a power- ful framework that can be exploited to design an expert Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 system to be embedded in tier-scalable reconnaissance mission architectures for assessment of the PH. 3.1.1. Deploying fuzzy experts on planetary bodies: Titan, Enceladus and Mars One of the questions that naturally arise while consider- ing autonomous systems for habitability assessment is how the effectiveness of the system varies as function of the observed planetary body. Indeed, the usefulness of a fuzzy expert system for autonomous interpretation of global and local habitability is influenced by two major and inter- connected factors: (1) the ability of the system to acquire and store data and (2) the distance of the observed planet from Earth. Generally speaking, the larger the distance between the deployed observing platform(s) and Earth, the higher is the level of needed autonomy. Consider Titan and Enceladus exploration scenarios. The current Cassini mission has been successful in collecting and sending back to Earth a great deal of information unveiling novel features of these two saturnian satellites. For example, Cassini radar observations and the Huygens descent/ landing probe unveiled apparent methane lakes, riverbeds, coastlines (Elachi et al., 2006; Perron et al., 2006) and dunes (Lorenz et al., 2006) on Titan. Enceladus geological activity is under scrutiny as well: closer observations of the saturnian satellite as imaged by Cassini, revealed an interesting world where tectonic activity and cryo-volcan- ism dominated the satellite geologic history and where geyser-like eruptions continued in real-time during Cassi- ni’s exploration (Porco et al., 2006; Kargel, 1995, 2006). Cassini is an orbiting platform equipped with instruments for remote data collection and most of its operations are constrained by instrument resolution and atmospheric effects (Tomasko et al., 2005). Science data are stored in the solid-state recorder (SSR) which has the ability to retain 2 Gigabits of data coming from the 12 on-board instruments. Clearly, data storage becomes a factor since if the collected bits exceed the storage capability, information must be erased. Currently, since the spacecraft is located between 8.2 and 10.2 astronomical units (AU) from Earth, it takes about 68–84 min to communicate with Earth. Real- time communication is impossible and as a matter of fact, it takes 3 h for the system (including human beings) to react to any on-board problem. Limited on-board software can diagnose occurring problems, but if they appear too severe, the spacecraft is put into a safe mode and essentially shut down until human’s intervention. Similar approach to fault tolerance caused the Galileo spacecraft at Jupiter to lose nearly half of its data collected at the most interesting close-approach phases. The absence of software capable of autonomous decision-making and control jeopardizes the ability of the system to perform optimal reconnaissance. Far more problematic than engineering monitoring and fault detection/protection (where the system is nominally known and understood) is the task of autonomous science which consists in autonomously and intelligently discover- ing key natural phenomena in environments that are th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]6 neither well known nor understood. These phenomena may be either transient or may be observable for only brief periods. The dynamic nature of some phenomena or the brevity of the observing windows, coupled with concerns for the engineering safety of the spacecraft are two key aspects limiting the ability of current systems to respond to serendipitous discoveries and to modify their observing strategies. More fundamental are the problems of (1) autonomous recognition of what may constitute an important discovery, (2) autonomous prioritization of the discovery relative to other observing and engineering maintenance tasks, (3) the autonomous development of an observing strategy if the phenomenon should be deemed to be important enough to formulate and insert new observations. Current systems are entirely unable to make such decisions. The consequence is that many profoundly important phenomena probably go undiscovered or under- observed. This is a severe handicap when dealing with phenomena that may be difficult to detect due to a transient nature, due to low frequency or low concentra- tion relative to detection thresholds, or due to observability only from special vantages. When the phenomenology is not yet known, and thus is not predictable by traditional means of spacecraft observation, the ability to capture a discovery observation is especially challenging; indeed, some of the things we most search for in the Solar System may be essentially impossible to find with current mission and observing approaches. Is life or fossil detection on distant worlds among these phenomena? We suggest that it may be; thus astrobiology is an important and exception- ally challenging application of the fuzzy logic-based approach to planetary exploration. In the case of Titan, while we believe that a tier-scalable deployment on the saturnian satellite is the most effective way to explore it (Fink et al., 2006b, 2007), any system equipped with a fuzzy-based expert system for habitability assessment can effectively overcome the limitation imposed by Earth communication delays and data storage. We envision a scenario where a multi-tier system (e.g., orbiter(s) plus blimp(s) plus ground agent(s)) is deployed on Titan. The platforms are equipped with various sensors for in-transit and continuous data acquisition. Despite the fact that the overall amount of data collected during the course of a tier-scalable mission could be extremely large, a fuzzy-based expert system for life assessment and inter- pretation can effectively retain only data that yields significant findings; thus, the integrated system can concentrate its actions on locales worth of exploration using the platforms and sensors best suited to make the observations. Such expanded platform/sensor systems may work in the traditional human-controlled mode, but they also may work in a semi-autonomous mode. For example, on Titan, the system might select potentially interesting information- rich sites concentrated around zones of mobile condensed hydrocarbons (e.g., lakes), send signals back to Earth and initiate a human–machine interaction where the humans Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 analyze both system results and explanations (see Fig. 2). Subsequently, humans could command further inves- tigation or alternatively disregard the selected areas. At Enceladus, the system might be designed to look specifically for transient geyser eruptions and geother- mal emissions and focus observations on the phenom- ena deemed most interesting or most suitable for safe observations. In this paper, we focus on designing a fuzzy system for habitability assessment taking Mars as a case study. Since Mars is an extremely rich and complex environment, our work may serve (a) to illustrate the ideas behind developing such fuzzy expert and (b) to provide a platform for system design and implementation that can be readily tested and verified for consistency on hypothesized and (in the future) real Mars scenarios. Mars is relatively close to Earth and one could argue about the need for such system. Indeed, one could envision two alternative scenarios where (1) one deploys landers and/or rovers in a carefully selected locale (e.g., forth coming Phoenix mission 2008) or (2) one deploys a large number of inexpensive agents looking for life everywhere. Case 1 suffers the classical limitations associated with the relative mobility of the deployed agents and past and present missions (e.g., Pathfinder, Mars Exploration Rovers Spirit and Opportunity) have been unsuccessful in looking for life. For case 2, either if agents are deployed following a tier-scalable scheme or if multiple agents are indiscriminately sent onto the Martian surface, data management becomes a critical factor. As the number of agents increases, the amount of data increases exponen- tially. Fuzzy experts may be extremely functional in understanding and interpreting the collected data and again, if the semi-automatic mode with human in the loop is preferred, Earth-based fuzzy systems could help humans to perform synthesis of information quickly and effectively. The Mars reconnaissance orbiter (MRO) has been recently inserted into a martian orbit and its on-board instruments are sending an unprecedented amount of data back to Earth. For example, the high resolution imaging science experiment (HiRISE) camera system is expected to send 12 Terabytes over the course of two years (McEwen et al., 2007). Careful human analysis of large datasets is generally problematic and expert systems capable of understanding Martian geology and geochemistry as well as assess locales’ habitability may operate from Earth in a quasi real-time mode and be established as an aiding tool to help coordinating the sequence of observations. Therefore, while we believe that the proposed methodology is most effective in tier-scalable architectures, it may be considered as foundational in developing an Earth-based semi- autonomous system for decision making (with humans in the loop). Such a system could help humans coordinating the observing platforms already deployed around/on Mars. 3.1.2. Fuzzy logic versus alternative AI schemes Any artificial intelligence (AI) scheme must be able to deal with the major issues arising in knowledge th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 7 engineering. For example, representation, inference, learn- ing, generalization, explanation and adaptation capabilities must be carefully analyzed and considered when choosing a suitable scheme for habitability assessment. Symbolic AI, fuzzy logic-based and neural networks schemes provide different solutions to the life-searching problem. Neural networks are connectionist systems where knowledge is distributed among various nodes. Thus, neural networks use unstructured knowledge (i.e. they learn by examples, by doing or by analogy), and they are capable of good generalization and adapta- tion. Conversely, symbolic AI and fuzzy systems provide an effective mean to represent structured knowledge. Moreover, inference is exact in symbolic AI while it is approximate in fuzzy-based and neural network systems. Because of their nature, symbolic AI systems do not deal very well with missing corrupted and inexact data. When dealing with planetary exploration one of the major problems to be addressed is how to represent uncertain knowledge and uncertain data. Besides fuzzy methods that are inherently capable of dealing with uncertainties, probabilistic methods could be also em- ployed. By relying on the axioms of probability as a mathematical framework, coherent knowledge-bases can be built using (a) data collected in the course of past missions and (b) statistical methods to determine the appropriate conditional probabilities (objective probabil- ity). Moreover, conditional probabilities that define beliefs unsupported by data (subjective probability) can be implemented as well. In short, probabilistic methods are based on estimating the posterior probability for a conclusion (defined by a rule) to be accepted as correct. In practice, probabilistic methods cannot deal with ambiguous and contradictory scenarios. Indeed Bayes’ theorem fails whenever multiple rules reach different conclusions if a condition is true. Conversely, fuzzy systems can deal with contradictory and ambiguous rules by naturally providing a trade-off during the inference process (rules are fired at the same time). Such scenarios are expected to be found when assessing habitability in complex environment. At this stage of development, we believe that fuzzy-based systems are an ideal solution to the problem of assessing planetary habitability. In such systems, structured knowl- edge is directly implemented by intuitive, easy-to-devise, fuzzy rules that can facilitate the interaction between planetary scientists and computing, autonomous machines. The basis of fuzzy logic is the basis for human commu- nication and therefore in unknown and uncertain planetary environments, fuzzy systems will be able to provide good and solid explanation for any autonomous decisions. Planetary experts (e.g., geologists, astrobiologists, cosmo- chemists) will be required to provide their knowledge and to contribute both in the design and testing phase while working on a common ground with computer and AI engineers. Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 3.2. Fuzzy logic basics Fuzzy logic was introduced by Lotfi Zadeh (1965, 1975) as multi-valued logic capable of dealing with intermediate values between traditional, absolute evaluations such as ‘‘true or false’’, ‘‘hot or cold’’, ‘‘yes or no’’, and so on. Fuzzy logic-based systems were born as an alternative to the traditional notion of set membership and the Greek classical logic. One of the famous propositions of the Aristotelian logic is the ‘‘law of the excluded middle’’ according to which any proposition must be either true or false. The semantic meaning of a proposition is usually determined via a truth table. Fuzzy logic allows the possibility of degree of membership to a set, opening the door to a new way of defining knowledge using statements that can be true to a certain degree (the ‘‘excluded middle’’ coming to the forefront). Fuzzy logic is an excellent tool for implementing common sense knowledge on a digital computer. It is based on simple mathematical concepts and is supported by a well-defined logical framework. Moreover, fuzzy logic is tolerant of noisy, imprecise, and faulty data. Fuzzy logic expert systems deal with un- certainty by reasoning over the general understanding of the problem rather than on exactness of the process (e.g., the natural world vs. the confines of a lab-based controlled experiment). Next, we define the basic ingredients and concepts that are required to design a fuzzy expert system for assessing life habitability. 3.3. Fuzzy sets and membership function The basic ingredient of fuzzy logic design is the simple notion of a fuzzy set (Zadeh, 1965; Dubois and Prade, 1980; Kasabov, 1996). In conventional classical and/or propositional logic, given a set U (called universe of discourse) and a subset A (included in U), any element x in U can either belong or not belong to A. Formally, this can be stated defining a function, called membership function, that takes any element x in U and assigns either 1 (the element belongs to A) or 0 (the element does not belong to A). In this sense, the set is distinctly constrained since it has well-defined boundaries (‘‘crisp’’ in Fuzzy Logic terms). A fuzzy set is a set without crisp boundaries, i.e., an element x can belong to the set with a certain degree of membership. The fuzziness of the set is translated into a new definition of the membership function, which can now take any value between 0 and 1. If the value of the membership function over x is 0 or 1, then the element belongs or does not belong to the set. For any intermediate value, say 0.7, we state that the element x belongs to A with a degree of membership of 0.7. It is clear that fuzzy sets are extensions of ordinary (crisp) sets. Operations between fuzzy sets can be extended in similar fashion. Generally, it is desirable to intersect, unify, and negate fuzzy sets. These operations are allowed by using the logical operators AND, OR, and NOT (see Zadeh, 1988; Kaufmann and Gupta, 1985; Ross, 2004, for th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]8 a detailed explanation of these operators and how they are applied). 3.4. Fuzzy logic paradigm: fuzzy rules and inference mechanism The fuzzy logic paradigm is based on the appropriate definition of fuzzy propositions, fuzzy rules, and an inference mechanism, which are built upon the concept of fuzzy sets. Fuzzy propositions are linguistic statements linked by fuzzy connectivities. In general, the statements are not true or false but their truth-value is defined by the membership value associated with the respective fuzzy variable represented in the proposition. The connectivities are the classical logical operators such as AND, OR, and NOT, but they are applied according to the way operations between fuzzy sets are defined (see Zadeh, 1988, 1989). If complex propositions must be evaluated, this can be done by the computation of a new membership function constructed using the operational rules between fuzzy sets. Fuzzy rules are defined by linguistic statements expres- sing the expert field knowledge required to infer new facts. The rules express conditional statements, which form the basis of the fuzzy logic knowledge system. A large part of the presented effort is devoted to defining the appropriate rules that condense the methodology employed in ap- proaching the problem of identifying suitable prime candidate sites of high potential for life in planetary environments. Thus, the ‘‘field expertise’’ and ‘‘knowledge’’ of a field geologist will be translated into appropriate rules for a direct implementation of an intelligent fuzzy-based expert system. Two major classes of rules are commonly used by fuzzy systems: Sugeno-type and Mamdani-type. The Mamdani- type rules (Mamdani and Assilian, 1975; Mamdani, 1977) have the following structure: IF ða is AÞ THEN ðb is BÞ. Here, we distinguish between two fuzzy propositions: antecedent or premise (a is A) and consequent or conclusion (b is B). The truth-value of these propositions is defined via their membership functions. A and B are two linguistic values defined by their associated membership function in the universes X (for variable a) and Y (for variable b). For example, in the context of Mars explora- tion, a could be the number of water-carved valleys present in a region, b is the potential for life habitability, and A and B could be representative of linguistic statements such as ‘‘high’’, ‘‘low’’, etc. The Mamdani-type rules are therefore IF-THEN rules. If many inputs are present, many propositions can be formed and linked using logical operators such as AND and OR. For example, the rules may have more than one antecedent: IF ða1 is A1Þ AND ða2 is A2Þ . . .AND ðan is AnÞ THEN ðb is BÞ. Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 In general, many rules are required to implement the desired knowledge. In fuzzy logic, the inference mechanism determines the sequence used in firing the rules to obtain the desired solution. In typical settings, the rules are fired at the same time or cyclically. In certain Mamdani-type rules, a confidence factor (CF) can be introduced to deal with uncertainty. In essence, the factor is used to weight the importance of a rule relative to the others. We will see how the confidence factor associated to one rule or to a group of rules will play a key role in planetary settings. In our design, we will focus exclusively on Mamdani-type rules (for more information on Surgeno-type rules see, e.g., Sugeno, 1974, 1985). The Inference Mechanism is a process of matching the domain with the solution space. In some sense, the process can be seen as given some data X and a set of rules, to infer, through a chain of matching, a new value Y0. This is a key process since new facts are inferred using some plausible reasoning over knowledge and data. In a broader sense, the inference mechanism is defined by specifying an implica- tion operator, a composition rule, and else-links between rules. Examples of else-links are the AND- and OR-link. In the AND-link, two propositions are linked together by the AND operator, which consists of determining a new membership function through the MAX operator (i.e., the new membership function is obtained by comparing the two original functions point-by-point and taking the maximum value). In the OR-link, two propositions are linked via the OR operator, which consists of determining the new membership function via the MIN operator (i.e., the new membership function is obtained by comparing the two original functions point-by-point and taking the minimum value). The implication operator provides a way to perform inference using the rule. The antecedent is linked using AND/OR operators and the result is transformed by the implication operator, which shapes the fuzzy set of the consequent. In practice, the Mamdani inference system (Mamdani and Assilian, 1975) is generally employed, which includes the MIN function, (i.e., the fuzzy set is chopped off to a degree specified by the antecedent) and the PROD (scaling) function (i.e., the fuzzy set is squashed to a degree specified by the antecedent). (For more information see text books such as Ross, 2004.) Importantly, if many rules are considered, the fuzzy inference mechanism combines the results of all rules to provide the output. 4. Knowledge-base for the fuzzy-expert system: mimicking the scientific/operational approach of a geologist, biologist, and chemist This section provides the basis for building the knowl- edge-base required to construct a fuzzy-based expert system capable of autonomously determining the potential exhibited by the planetary locale under observation to harbor life. The knowledge-base construction is not only an attempt to simulate the approach of a geologist, but also th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 9 an attempt to merge other disciplines with geology such as biology and chemistry, thereby establishing connections between presence of life and/or biosignatures and a specific geological environment. As explained in Section 3.1.1, we focus on Mars, which affords an opportunity to establish the basic methodology that can be eventually extended to other planetary scenarios, and we discuss the rationale that underlies the proposed two-layer fuzzy system. 4.1. Rationale for building the knowledge-base to assess potential for life habitability In a recent paper (Dohm et al., 2004), a novel approach was proposed for selecting prime candidate sites for future Mars exploration. Indeed, published geological maps portray Mars as an episodically active, dynamic planet with magmatic, tectonic, water/ice, and wind activity in its recorded geologic history. It may still be internally active, which includes local elevated heat flow and potential magmatic-driven activity (Dohm et al., 2001b). In the martian context, a prime candidate site can be defined as a locale that has the greatest potential to yield significant geologic, paoleohydrogeologic, paleoclimatic, and exobio- logic information (Dohm et al., 2004). More importantly, in answering the fundamental question of life on Mars, the site should possess the greatest potential for extant and/or fossilized life. In the proposed approach, which has been built upon past efforts (e.g., Greeley and Thomas, 1994), the optimal candidate site is established through a synthesis of all published information, including stratigraphic, topographic, geomorphic, paleotectonic, paloehydrologic, geophysical, thermal, spectral, and elemental data. Based on this approach, Dohm et al. (2004) concluded that the martian Northwestern Slope Valleys (NSVs) region has elevated probability to yield significant geologic, climatic, and possible exobiologic information. The elevated con- fidence came from compiling layers of information, which jointly and consistently pointed in the direction of the region being a prime candidate for future science-driven exploration. The geologic approach of synthesizing layers of in- formation from various perspectives, including orbital, airborne, ground, and subsurface, can in principle be used to determine the potential that a specific locale on Mars comprises extant and/or fossilized life. The appropriate synthesis is usually performed by the field expert (i.e., the planetary geologist/biologist/chemist) who collects avail- able information and, using existing field- and remote- based knowledge and expertise, determines and explains if a certain region has the potential for past and/or present life. To the extent possible, the terrestrial geologic approach has been adopted and then adapted to meet the needs and special challenges inherent in producing planetary geologic maps that rely on spacecraft data, lacking ground truth. In the specific case, the approach used to determine the life potential for a certain locale via synthesis of layers of information, requires the field expert Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 to perform comparative analysis among various elements in the field, employing some level of intuition/judgment to establish when certain data collectively indicate that a site yields the highest probability for the presence of life. If the amount and diversity of information is great, this is indeed difficult and time consuming for any geologist/biologist/ chemist (e.g., while in transit, maintaining location, synthesizing spatial and temporal information at various scales, and testing and formulating hypotheses, etc.). Moreover, in the context of the tier-scalable reconnaissance mission concept, it is desirable to have an autonomous system that mimics the geologist approach, while readily coupling information gathered from various disciplines such as biology or chemistry. Such a system would provide independent inference over a set of basic rules to establish the level of confidence that a certain region is harboring life. Relying on this foundational approach, we aim at defining a set of independent rules that (a) mimic the layers-of-information-synthesis approach as defined above, and (b) can be directly implemented in the fuzzy expert system paradigm. The key element of the knowledge-base approach is to establish how the search for sites with elevated potential for life can be conducted on the basis of layers of synthesis and comparative analysis of diverse information. One of the premises is that water is the key to life, and as such, the system that performs tier-scalable reconnaissance must identify, characterize, and quantify past and present water. Water is considered a first-order indicator of life- containing potential (including surface and subsurface water) in the form of solid (ice/rock with interstitial ice), liquid (including brines), and water vapor (e.g., moisture to the surface such as in the case of fog embankment). In the case of the evolution of the Tharsis magmatic complex (Dohm et al., 2006a, in press; Komatsu et al., 2004), geologic information points to an episodically active complex, which includes magmatic-driven floods (including hydrothermal activity), ponding forming water bodies ranging from lakes to oceans, and transient hydrological cycles with possible extensive periods of inactivity (Baker et al., 1991; Fairén et al., 2003). During the pulses of catastrophic activity, does life originate in the subsurface and potentially thrive at the surface, as hypothesized (e.g., Schulze-Makuch et al., 2005b; Fairén et al., 2005)? In the Atacama desert, Chile, there are investigations underway that are indicating promising results that life may have sprung during magmatically and hydrothermally active periods. The history of water on Mars is recorded not only in geomorphological features such as valley networks, but also in sedimentary facies, and as such, a variety of sedimentary rocks are candidates for exobiological inves- tigation (Komatsu and Ori, 2000; Ori et al., 2000). Furthermore, groundwater is a potential habitat for life in the past (Mahaney et al., 2004) and present. Based on published geologic, paleohydrologic, and geomorphic information, Mars has been active and water enriched until geologically recent times (Late Amazonian th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]10 epoch; e.g., Scott and Tanaka, 1986; Scott et al., 1995; Fairén et al., 2008, currently under review). This was recently confirmed through gamma ray and neutron spectrometers (Boynton et al., 2002; Feldman et al., 2002, etc.) to the extent that it could be currently enriched in water in places. In addition, Mars Global Surveyor-, Mars Express-, and Mars Exploration Rovers-based informa- tion, which includes the identification of rock materials such as hematite, sulfates, and phyllosilicates, points to histories involving water. Moreover, Mars demonstrates internal heat release, both past and possibly present (e.g., see Dohm et al., 2004, 2006b). Another first-order indicator for life-containing potential is energy. Voids in the martian subsurface (e.g., Rodriguez et al., 2005, 2006), which are expressed at the surface by a diverse suite of features such as collapsed and broken terrain, referred to as chaotic terrain, may provide suitable habitats for life, but without energy into the system, life may exist in dormant form; once energy is injected into the environment in the form of heat and/or nutrients, then life may spring (e.g., Boston et al., 1992; Boston, 2004; Schulze-Makuch et al., 2005a, b; Fairén et al., 2005). Anything that indicates present energy is extremely important (e.g., elevated heat flow related to the internal heat release of a planet). There are multiple lines of evidence that collectively point to an active Mars (Anderson, 2001; Mitchell and Wilson, 2003; Márquez et al., 2004), and thus there may be heat release to subsurface and surface environments (including associated aquifers). The thermal emission imaging system (THEMIS), an instrument aboard the Mars Odyssey spacecraft, was intended to identify thermal anomalies on Mars related to internal heat release, but has not identified such. One possible explanation is the low resolution of the instru- ment, or that the heat releases are low grade, and thus exceedingly difficult to identify by remote sensing (Dohm et al., 2006b). The tier-scalable reconnaissance mission paradigm originated by Fink et al. (2005a–c, 2006a, b, 2007) would afford an optimal opportunity to test whether Mars is still active through higher resolution spectroscopy (both multispectral and nuclear) acquired via airborne and field-based reconnaissance (with the advantage of mini- mizing noise from electromagnetic radiation and atmo- spheric conditions). Other means of identifying energy, which could be part of a tier-scalable architecture, include heat sensing, seismic, radar sounding, etc. Our primary objective is to autonomously identify, characterize, and quantify environments that indicate both water and energy (past and present). For example, by identifying a region of heat release with associated near- surface water, the life potential would be at an elevated level. If the presence of water and energy is long-lived, however, then the life potential is even further increased. These considerations are useful in assessing the utility/ importance of the indicators (ranking). We now distinguish between past long-lived water/energy and present water/ energy. The simultaneous presence of these four conditions Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 has the highest life potential. For example, if volatiles release from fractures and geologic contacts, a ponding of brines in a topographic low, elevated heat flow, and/or near-surface aquifers are observed, then the probability for life is elevated. If such indications of potential life are coupled with, for example, evidence of past water and energy such as vent structures, lava flows, joints, faults, and rift systems, valleys (including networks), etc., then the life potential is further increased. Importantly, water alone, though significant, may not be enough. For example, the observation that valley forms such as sapping channels sourcing from fractures, as noted for Mars (e.g., Mouginis-Mark, 1985, 1990), is an indication of past water-related activity, but not necessarily life. The basic indicators for water-related activity (tran- sient and long-lived) are typically geomorphologic, strati- graphic, and topographic. Other indications of past water- related activity include drainage basins, valleys that debouch into basins, stratigraphic sequences (water and rock materials that are deposited into basins), and valley forms (e.g., valley networks, sapping channels). Specifi- cally, if a locale is to be identified as the highest life potential, it must record long-lived, water- and energy- related activity (e.g., ancient-, middle-, and recent-planet magmatic and tectonic activity). Thus, present water and energy coupled with indications of past water and energy has the highest life potential. The terrestrial- and planetary-based analysis of the relationship between water, energy, and life presented above, allows us to categorize the areas of interest according to past/present water/energy, leading to indica- tors of life potential. Detecting past AND present water plus past AND present energy is the first order of importance, whereas the second order of importance are areas exhibiting present water AND present energy; a third order of importance are areas with EITHER past and present water OR past and present energy (water and energy separately are at the same level), and a forth order of importance is past water AND past energy. To implement the above outlined approach in a fuzzy- based expert system, a two-step process is required to connect life to water and energy indicators. First, the expert system acquires water and energy indicators collected via streaming of sensorial data and assesses the past/present water and energy potential of the observed locale. Subsequently, the system reasons over past/present water/energy to extract life-containing potential. Thus, a two-layer fuzzy system is required. The first layer contains knowledge-bases made of rules linking the appropriate water/energy indicators to past/present water and energy potential. For example, if the system detects channels and streamlined bed forms, this is an indication of past aqueous conditions. In addition, if spectral data report high chlorine, sulfates, and evaporites, everything collectively points to past aqueous conditions. The same reasoning can be applied to the indicators of present water and past and present energy. th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 11 The second layer contains a knowledge-base comprised of rules linking past/present water/energy to life-containing potential. The second layer of the expert system reasons over the output of the first layer using rules combining order of importance of past/present water and energy to effectively infer PH of the observed locale. For example, if potential for past/present water and energy is high, the potential for life-containing will be high to the highest degree. Although the presented analysis outlines the essential approach that considers many lines of evidence and demonstrates the path that must be followed to construct the appropriate knowledge-bases, some extensive work must be done to translate the expert knowledge in appropriate rules that can be implemented on a digital computer (IF-THEN Mamdani-type rules, see Section 3.4). Similar to human experts, fuzzy systems are capable of learning and self-adapting, or being manipulated by humans based on experience. Depending on system performance on Mars, or performance in test simulations on Earth, it might be decided to change some rules or to introduce additional expert systems (e.g., warm vs. cold environmental conditions on Mars; e.g., see Kargel, 2004). The following sections illustrate the details of the process Multi-Tier Data ingestion Pre-processing & Categorization Present water indicators Past water indicators Present energy indicators Past energy indicators 1st Fu Fig. 3. Schematic of the two-layer fuzzy logic-based expert system for the tie reconnaissance system can be integrated into the tier-scalable architecture. The distributed over two layers and operate according to the scheme outlined in further deployment to the locale (e.g., drill rig that is optimally placed to sam Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 that creates effective knowledge-bases and implements the rules required by the two-layer fuzzy expert system. 5. Building a fuzzy logic expert system for life habitability assessment Here, we describe in detail the basic structure of the proposed two-layer, fuzzy-logic-based expert system. The life indicators and the proposed set of rules should be foundational and open to revision as our understanding of life on other planetary bodies develops. As such, the approach is generic enough to accommodate any possible modification while still employing the same framework. Fig. 3 shows a schematic of the proposed multi-layer, fuzzy-logic-based expert system. The system is defined by two layers: according to the deployed multi-tier architec- ture, data collected by multiple sensors at multiple scales from several vantage points observing an operational area of interest are pre-processed by AGFA-like software (Automated Geologic Field Analyzer by Fink et al., 2005d; now Automated Global Feature Analyzer, see Fink et al., 2007, to be submitted) to extract feature information in the form of numerical values for the respective indicators of interest. The indicators are arranged according to Rules & Inference Engine Rules & Inference engine Rules & Inference engine Rules & Inference engine Rules & Inference engine zzy-Layer 2nd Fuzzy-Layer Potential For Habitability r-scalable mission architecture. The illustration shows how an intelligent ‘‘brain’’ of the system consists of a set of five interconnected fuzzy systems the text. Depending on the life-assessment result, the system might order ple identified near-surface groundwater in a locale of elevated heat flow). th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Table 2 Fuzzy rules for fuzzy system #1 (first layer) Present water fuzzy system: rules Indicator and confidence factor Rules SLW, CF ¼ 1 IF SLW is H THEN PrW is H IF SLW is M THEN PrW is M SSLW, CF ¼ 1 IF SSLW is H THEN PrW is H IF SSLW is M THEN PrW is M SB, CF ¼ 1 IF SB is H THEN PrW is H IF SB is M THEN PrW is M SSB, CF ¼ 1 IF SSB is H THEN PrW is H IF SSB is M THEN PrW is M Low-rule, CF ¼ 1 IF SLW is L AND SSLW is L AND SB is R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]12 specific categories of scientific interest (e.g., elemental, spectral, geomorphologic, paleohydrologic, topographic, stratigraphic, seismic, paleotectonic, atmospheric, spectral, elemental, etc.) and fed to a set of four independent fuzzy systems that are designed to generate past/present water/ energy information. The first layer processes the indicators according to defined fuzzy rules and generates the past/ present water and energy potential. The latter is forwarded to the second fuzzy-layer to provide, after appropriate fuzzy inference over a pre-defined set of fuzzy rules, the PH exhibited by the locale under investigation. Next, the rules for both layers are defined. 5.1. First fuzzy-layer: past/present water/energy assessment for life-containing potential The first layer is defined by four independent fuzzy systems. Each fuzzy system has an embedded knowledge- base composed of a sequence of IF-THEN, Mamdani-type rules. Using the appropriate rules, each fuzzy system independently assesses the present/past water and energy potential. 5.1.1. Fuzzy system #1: present water assessment The first fuzzy system is designed to acquire indicators of present water at a locale under investigation as signs for potential life. In essence, the first fuzzy system evaluates the presence of water, yielding a parameter (PrW) in the range 0–100, the latter being the maximum possible level. Table 1 shows a set of (intuitive) indicators of the presence of water, and thus relevant to the potential for life. We consider water in all forms (solid, liquid, and vapor/ moisture), occurring in the subsurface, on the surface, and in the atmosphere. The mixture of salt and water is also considered (i.e., brines). The confidence factor (varying between 0 and 1) is descriptive of the importance of the Table 1 Present water indicators Present water Category Indicators Confidence factor Elemental/spectral Surface liquid water (SLW) 1 Elemental/spectral Surface solid water (SSW) 0.9 Elemental/spectral Surface brine (SB) 1 Elemental/spectral Subsurface brine (SSB) 1 Elemental/spectral Subsurface liquid water (SSLW) 1 Elemental/spectral Subsurface solid water (SSSW) 0.9 Atmospheric Atmospheric moisture (AM) 0.9 Atmospheric/spectral Soil moisture (SM) 0.9 Atmospheric/ geomorphology Embankments moisture (EM) 0.9 The parameters, which are selected to characterize the presence of water on Mars, are the input to the fuzzy system #1 (first layer) to determine the ‘‘present water’’ potential (PrW). Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 indicators relative to each other. Moisture embankment has been considered separately, and when combined with other indicators, is critical to identifying potential life- containing locales. The associated rules are shown in Table 2. The set of IF- THEN, Mamdani-type rules are defined and organized to show the impact of the present water potential on harboring life. In the semantic description of the rules, ‘‘H’’, ‘‘M’’, ‘‘L’’ stand for high, medium, and low, respectively. The rules have been organized in two sets where the indicators have the same confidence factor. The confidence factor weights the importance of the rules associated with the corresponding indicators. Particular care has been taken in defining the lower limit. For example, the locale under observation might show high liquid water content but low solid and atmospheric water (ice and moisture, respectively). In this situation, the locale still exhibits an elevated presence of water. To adapt the rule to this circumstance, the AND connector is utilized for low content, i.e., ‘‘present water’’ is low only if all indicators are low (see Table 2). LyAND SSB is L THEN PrW is L SSW, CF ¼ 0.9 IF SLW is H THEN PrW is H IF SLW is M THEN PrW is M SSSW, CF ¼ 0.9 IF SSW is H THEN PrW is H IF SSW is M THEN PrW is M AM, CF ¼ 0.9 IF SLW is H THEN PrW is H IF SLW is M THEN PrW is M SM, CF ¼ 0.9 IF SLW is H THEN PrW is H IF SLW is M THEN PrW is M EM, CF ¼ 0.9 IF SLW is H THEN PrW is H IF SLW is M THEN PrW is M Low-rule, CF ¼ 0.9 IF SSW is L AND SSSW is L AND AM is LyAND SM is L AND EM is L THEN PrW is L The rules define the knowledge base required for automatic inference of the present water potential (PrW) Mamdani. The rules structure is fairly intuitive. For example, ‘‘IF SLW is H THEN PrW is H (CF ¼ 1)’’ should be read as ‘‘If the surface liquid water is high then the present water potential exhibited by the observed area is high (with the highest confidence)’’. Each confidence factor group (CF ¼ 1, 0.9) has an associated Low-rule. The latter has been constructed linking water indicators with same confidence factor via the AND connector. The Low-rule reflects the circumstance that potential for water containing is low only if a reconnaissance system observes a concurrent low value for any of the considered present water indicators. th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Table 4 Fuzzy rules for fuzzy system #2 (first layer) Past water fuzzy system: rules Indicator and confidence factor Rules CHL, CF ¼ 1 IF CHL is H THEN PsW is H IF CHL is M THEN PsW is M SU, CF ¼ 1 IF SU is H THEN PsW is H IF SU is M THEN PsW is M HE, CF ¼ 1 IF HE is H THEN PsWI is H IF HE is M THEN PsWIis M PT, CF ¼ 1 IF PT is H THEN PsW is H IF PT is M THEN PsW is M Low-rule, CF ¼ 1 IF CHL is L AND SU is L AND HE is LyAND PT is L THEN PsW is L VN, CF ¼ 0.9 IF VN is H THEN PsW is H IF VN is M THEN PsW is M SC, CF ¼ 0.9 IF SC is H THEN PsWIis H IF SC is M THEN PsW is M OC, CF ¼ 0.9 IF OC is H THEN PsW is H IF OC is M THEN PsW is M Low-rule, CF ¼ 0.9 IF VN is L AND SC is L AND OC is LyTHEN PsW is L AP, CF ¼ 0.8 IF AP is H THEN PsW is H IF AP is M THEN PsW is M PP, CF ¼ 0.8 IF PP is H THEN PsW is H IF PP is M THEN PsW is M AF, CF ¼ 0.8 IF AF is H THEN PsW is H R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 13 5.1.2. Fuzzy system #2: past water assessment The second fuzzy system accepts past water indicators as inputs and yields a parameter that assesses the potential for presence of water in the past (PsW). Table 3 shows the major parameters that can be used as indicators of past water-related activity and that are relevant to the possible presence of life at the locale. The indicators are mostly categorized as elemental, spectral, geomorphologic, and topographic. As in the previous case, the indicators are associated with confidence factors that weight the relative impact of the indicators on the potential for past water. The rules, shown in Table 4, are constructed following patterns similar to the previous case, but some variations, due to the nature of the categories, are considered. For example, in defining the rules, the elemental and spectral indicators for past water, generally indicating the chemistry of the observed locale, are not mixed with geomorphologic, topographic, and stratigraphic indicators, which are physical markers of the geologic and paleohydrologic histories of the region. The indicators are organized such that each group possesses the same confidence factor. Each group is evaluated with rules that are based on the level of indicator value (e.g., the higher the indicator value the higher the potential of past water at the locale). Again, ‘‘past water’’ is assumed low if all indicators exhibit low values at the same time, i.e. we used the AND logical connector to implement this condition (see Table 4). Table 3 Past water indicators Past water Category Indicators Confidence factor Elemental/spectral Chlorine (CHL) 1 Elemental/spectral Sulfates (SU) 1 Elemental/spectral Hematite (HE) 1 Elemental/spectral Potassium/thorium (PT) 1 Geomorphology Valley networks (VN) 0.9 Geomorphology Sapping channels (SC) 0.9 Geomorphology Anastomosing patterns (AP) 0.8 Geomorphology Outflow channels (OC) 0.9 Geomorphology Poligonal patterns (PP) 0.8 Geomorphology Alluvial fans (AF) 0.8 Geomorphology Stream-lined bedforms (SLB) 0.8 Geomorphology Scarps/terraces (ST) 0.6 Geomorphology Parallel/concentric ridges (PCR) 0.6 Geomorphology Summit pits (SP) 0.6 Geomorphology Pit crater chains (PCC) 0.5 Topographic Basins (B) 0.6 Topographic Topographic highs (TH) 0.8 Stratigraphy Layered stratigraphy (LS) 0.6 The parameters, which are selected to characterize the possible presence of past water on Mars, are the input to the fuzzy system #2 (first layer) to determine the ‘‘past water’’ potential (PsW). The CF values have been established using our long-live experience in studying elemental, geomor- phologic, topographic structures of Mars. The indicators list is not intended to be comprehensive but reports the parameters our group believes have the major impact on past water potential. IF AF is M THEN PsW is M SLB, CF ¼ 0.8 IF SLB is H THEN PsW is H IF SLB is M THEN PsW is M TH, CF ¼ 0.8 IF TH is H THEN PsW is H IF TH is M THEN PsW is M Low-rule, CF ¼ 0.8 IF AP is L AND PP is L AND AF is LyAND SLB is L AND TH is L THEN PsW is L ST, CF ¼ 0.6 IF ST is H THEN PsW is H IF ST is M THEN PsW is M PCR, CF ¼ 0.6 IF PCR is H THEN PsW is H IF PCR is M THEN PsW is M SP, CF ¼ 0.6 IF SP is H THEN PsW is H IF SP is M THEN PsW is M B, CF ¼ 0.6 IF B is H THEN PsW is H IF B is M THEN PsW is M LS, CF ¼ 0.6 IF LS is H THEN PsW is H IF LS is M THEN PsW is M Low-rule, CF ¼ 0.6 IF ST is L AND PRC is L AND SP is LyAND B is L AND LS is L THEN PsW is L PCC, CF ¼ 0.5 IF PCC is H THEN PsW is H IF PCC is M THEN PsW is M IF PCC is L THEN PsW is L The rules define the knowledge base required for automatic inference of the past water potential (PsW). The structure of and the rationale for building the past-water knowledge base is conceptual similar to the other first-layer knowledge base. Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 5.1.3. Fuzzy system #3: present energy assessment The third fuzzy system is designed to evaluate the potential for present energy (PrE) via thermal, spectral, atmospheric, and seismic indicators that have a high impact on habitability. The indicators are shown in Table 5. From a life-searching point of view, the fuzzy th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Table 6 Fuzzy rules for fuzzy system #3 (first layer) Present energy fuzzy system: rules Indicator and confidence factor Rules GHA, CF ¼ 1 IF GHA is H THEN PrE is H IF GHA is M THEN PrE is M AHA, CF ¼ 1 IF AHA is H THEN PrE is H IF AHA is M THEN PrE is M VL, CF ¼ 1 IF VL is H THEN PrE is H IF VL is M THEN PrE is M RN, CF ¼ 1 IF RN is H THEN PrE is H IF RN is M THEN PrE is M Low-rule, CF ¼ 1 IF GHA is L AND AHA is L AND VL is LyAND RN is L THEN PrE is L TI, CF ¼ 0.8 IF TI is H THEN PrE is H IF TI is M THEN PrE is M AL, CF ¼ 0.8 IF AL is H THEN PrE is H IF AL is M THEN PrE is M Low-rule, CF ¼ 0.8 IF TI is L AND AL is L THEN PrE is L The rules define the knowledge base required for automatic inference of the present energy potential (PrE). R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]14 system must indicate the level of current energy that is fed into the system and that may support life. The tier-scalable exploration architecture should look for locales where thermal anomalies are present due to the internal heat release of the planet, both from manifesting in the ground/ subsurface and in the atmosphere. The ground heat anomaly (GHA) and atmospheric heat anomaly (AHA) are defined as the difference between the mean ground and atmospheric temperature of the locale and possible localized thermal spikes. The spectral thermal profile must be processed to filter the thermal signature coming from other possible energy sources, i.e., electromagnetic and radioactive. Thermal inertia and albedo are important parameters indicating the ability of the locale to retain energy. Volatile release might be associated with internal heat release, while seismic events may also record internal activity (e.g., magmatic-driven tectonic activity). The rules are shown in Table 6 and follow the same pattern shown in the previous cases, i.e., the higher the indicator value, the higher the potential for energy to be fed into the locale. The AND logical connector is used for the rules defining the condition that when every indicator is low, the energy present in the system is low. Table 7 Past energy indicators Past energy Category Indicators Confidence factor Stratigraphic Number of rock types (RT) 0.7 Stratigraphic Rock Sorting Index (RSI) 0.7 Paleotectonic Number of Faults (NF) 1 Paleotectonic Number of Joints (NJ) 0.7 Paleotectonic Number of intersection between faults (NIF) 1 Paleotectonic/ geomorphology Number of channels connected to faults (NCF) 1 Paleotectonic/ geomorphology Basins (B) 0.7 5.1.4. Fuzzy system #4: past energy assessment The fourth fuzzy system is designed to indicate the potential for energy flowing in the past. The indicators are shown in Table 7, and placed into the ‘‘past energy’’ category based on stratigraphic, paleotectonic, and geo- morphologic information. Past energy might have been fed into the region through endogenic-driven injection of magma, Marsquakes, and/or water flow (including hydro- thermal activity). The corresponding rules are shown in Table 8. The rules follow the same pattern defined for the previous fuzzy systems based on indicators value (e.g., the higher the indicator value, the greater the potential that energy was flowing into the locale in the past). The Table 5 Present energy indicators Present energy Category Indicators Confidence factor Thermal Ground Heat Anomaly (temperature spike) GHA 1 Thermal/ atmospheric Atmospheric heat anomaly (temperature spike) AHA 1 Thermal Thermal inertia (TI) 0.8 Thermal/spectral Albedo (AL) 0.8 Elemental/ Atmospheric Volatiles (VL) 1 Seismic Richter number (RN) 1 Energy related indicators are categorized as thermal, spectral elemental and seismic. The selected CF values are not absolute and in the future they can be modified to reflect new acquired knowledge and understanding of present energy mechanisms on Mars. Paleotectonic/ geomorphology Number of different lava flow in contact (NLC) 1 Paleotectonic/ geomorphology Lava/basaltic material (LBM) 1 Such indicators are categorized as stratigraphic, paleotectonic and geomorphic. Each CF value has been established using our team’s expertise in understanding the impact of the selected indicators on the past energy potential. Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 indicators are accompanied by confidence factors that weight their relative importance. The AND connector is used in the ‘‘low’’ rule, consistently following the scheme outlined for the other systems. 5.2. Second fuzzy layer: water and energy assessment for life-containing potential The second fuzzy layer is comprised of one fuzzy system. It accepts four inputs describing, respectively, the potential th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 15 for past/present water and energy (i.e., the outputs of the above-discussed fuzzy systems forming the first layer), and outputs the PH. Its knowledge-base has been constructed following the rationale outlined in Section 4.1 where water Table 8 Fuzzy rules for fuzzy system #4 (first layer) Past energy fuzzy system: rules Indicator and confidence factor Rules NF, CF ¼ 1 IF NF is H THEN PsE is H IF NF is M THEN PsE is M NIF, CF ¼ 1 IF NIF is H THEN PsE is H IF NIF is M THEN PsE is M NCF, CF ¼ 1 IF NCF is H THEN PsE is H IF NCF is M THEN PsE is M NLC CF ¼ 1 IF NLC is H THEN PsE is H IF NLC is M THEN PsE is M LBM, CF ¼ 1 IF LBM is H THEN PsE is H IF LBM is M THEN PsE is M Low-rule, CF ¼ 1 IF NF is L AND NIF is L AND NCF is LyAND NLC is L AND LBM is L THEN PsE is L RT, CF ¼ 0.7 IF RT is H THEN PsE is H IF RT is M THEN PsE is M RSI, CF ¼ 0.7 IF RSI is H THEN PsE is H IF RSI is M THEN PsE is M NJ, CF ¼ 0.7 IF NJ is H THEN PsE is H IF NJ is M THEN PsE is M B, CF ¼ 0.7 IF B is H THEN PsE is H IF B is M THEN PsE is M Low-rule, CF ¼ 0.7 IF RT is L AND RSI is L AND NJ is LyAND B is L THEN PsE is L The rules define the knowledge base required for automatic inference of the past energy potential (PsE). Table 9 Second-layer fuzzy knowledge base Second-layer fuzzy knowledge-base Order Ruler no. Confidence factor Rule s First 1 CF ¼ 1 IF PrW First 2 CF ¼ 1 IF PrW Second 3 CF ¼ 0.8 IF PrW Second 4 CF ¼ 0.8 IF PrW Second 5 CF ¼ 0.8 IF PrW Third 6 CF ¼ 0.6 IF PrW Third 7 CF ¼ 0.6 IF PrE Third 8 CF ¼ 0.6 IF PrW Third 9 CF ¼ 0.6 IF PsE Fourth 10 CF ¼ 0.4 IF PrW Fourth 11 CF ¼ 0.4 IF PrW Fifth 12 CF ¼ 0.2 IF PsW Fifth 13 CF ¼ 0.2 IF PsW The table shows the structure of the fuzzy rules selected to form the backbone groups each with decreasing CF factor defining the relative importance of the r CF value (first order the highest with CF ¼ 1, 5th order the lowest with CF ¼ 0 one, ‘‘IF PrW is H AND PsW is H AND PrE is H AND PsE is H THEN PH present energy and past energy is shown to be high (as evaluated by the first fu observation is very high with the highest level of confidence. The order of impor with high past water and present water content are more important than area Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 and energy contributions to possible extinct and extant life have been highlighted. Table 9 shows the full set of rules. The rules have been organized by order of importance, which is generally defined by a confidence factor that weights the importance of the rule relative to the others. The confidence factor is selected to have one out of five possible values (e.g., first order: CF ¼ 1, second order: CF ¼ 0.8, etc., see Table 9). There are 13 possible rules which form the backbone of the second fuzzy-layer. Rules number 1 and 2 are of first order of importance. They state that if all the values for past/present water/energy are high (low), the PH is very high (very low). To second order, if the locale under observation exhibits high (low) value of only present water and energy at the same time, the PH is high (low), but the rule is weighted with a lower confidence factor. Decreasing the order of importance further, there are situations where given high (low) potential of past water, or alternatively high (low) potential of past energy, the PH is still high, but the rule is weighted with CF ¼ 0.2, which indicates that the importance of the rule, in our framework, is the lowest possible. The medium condition for the parameters is expressed in rule number 5, where all water and energy indicators are medium. In this case, the PH is medium. The rule is assumed to be of second-order importance. Importantly, the constructed second-layer knowledge base is directly connected to the experience acquired by our team regarding understanding the connection between life and water/energy indicators. The rules might be modified for various planetary scenarios where the expertise acquired by planetary geologists, astrobiologists, and chemists may require a different set of rules that work better for the particular planetary body under investiga- emantic is H AND PsW is H AND PrE is H AND PsE is H THEN PH is VH is L AND PsW is L AND PrE is L AND PsE is L THEN PH is VL is H AND PrE is H THEN PH is H is L AND PrE is L THEN PH is L is M AND PsW is M AND PrE is M AND PsE is M THEN PH is M is H AND PsW is H THEN PH is H is H AND PsE is H THEN PH is H is L AND PsW is L THEN PH is L is L AND PsL is L THEN PH is L is H OR PrE is H THEN PH is H is L OR PrE is L THEN PH is L is H OR PsE is H THEN PH is H is L OR PsE is L THEN PH is L of the second-layer fuzzy system Mamdani. The rules are organized in five ule. Each group is coupled with an order of importance depending on the .2). The rules can be read in a very intuitive way. For example, rule number is VH (CF ¼ 1)’’, states that if the content of present water, past water, zzy-layer) then the potential for habitability exhibited by the locale under tance of the rules can be understood by observing that for example regions s with only high water or energy content. th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Table 10 Life indicators for two hypothesized martian regions Life indicators for region #1 and #2: first layer input Life indicator Category Region #1 Region #2 Hydrogen content Present water 6% (weight) 2% (weight) Sapping channels Past water 3 (number) 1 Valley networks Past water 10 (number) 1 Sulfates content Past water 19% (weight) 10% (weight) Heat release Present energy 250 (differential) 100 (differential) Richter number Present energy 6 (Richter scale) 0.5 (Richter scale) Local faults Past energy 7 (number) 2 (number) Local volcanoes Past energy 8 (number) 2 (number) Local lava flows Past energy 6 (number) 1 (number) It is assumed that an autonomous tier-scalable system performs reconnaissance over two distinct regions and collects data to derive nine (9) life indicators. The two areas are markedly different with the first region being the place with the richest elemental, geomorphologic, paleotectonic, and thermal features. R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]16 tion. Similarly, the life indicators may have to change in number and variety, and may have to be custom-tailored to provide the most appropriate information on habitability potential for the locale under investigation. The fuzzy logic framework is general enough to incorporate new expertise into the system without changing its underlying structure. 6. Testing the system on planetary scenarios: two examples In this Section, we construct and simulate a fuzzy logic expert system based on a simplified version of the above- mentioned knowledge-bases (1) to analyze the behavior of the expert system on reconnaissance-based, hypothesized planetary scenarios, and (2) to demonstrate the potential of the overall approach. The projected scenario is the following. It is assumed that a tier-scalable reconnaissance mission architecture has been deployed on Mars (Fig. 1): the architecture includes potentially three tiers, i.e., spaceborne, airborne and ground-based, all equipped with multiple agents/sensors to collect sensor data at multiple scales, resolutions, and vantage points. The goal of the mission is to look for signs of life over a large region of Mars. For this simulation, we assume that only nine indicators are available to assess the life potential exhibited by the hypothesized locale under investigation. More specifically, we assume that the gamma ray spectrometer (GRS)-based hydrogen abundance in- dicates weight percentage of water (Boynton et al., 2002, 2004) in this region. The GRS experience has already shown how gamma ray and neutron spectrometers can be used to map the hydrogen content of an area. Although the knowledge-bases constructed in the previous sections contain a variety of indicators for present water assessment (see Table 1), we assume that the system is able to assess only the relative abundance of hydrogen, which has been interpreted to be related to water in various forms (ice, brine, liquid, moisture, and/or hydrated minerals). More- over, it is assumed that the system is able to assess abundance of sapping channels, valley networks, and sulfates (in weight percentages) as indicators for past water; heat energy release (thermal spikes/anomalies) and Richter number (i.e. it quantifies the amount of seismic energy released during an earthquake) as indicators for present energy, and number of local faults, volcanoes, and lava flows as indicators for past energy. The system performs reconnaissance during hypothetical deployments to two martian regions by collecting data en- route (orbiting, hovering, roving, and homing in on) while determining the absolute value of the nine indicators of past/present water and energy (the regions are inferred, though they are not considered unrealistic based on existing geologic, paleohydrologic, geomorphic, geophysi- cal, spectral, and elemental information—e.g., the NSVs regions is hypothesized to approximate region 1, though we do not know whether there is elevated heat flow and groundwater lurking near the surface). Table 10 shows the values of the indicators across the observed regions. Data Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 collected during the hypothetical martian deployments portray a highly interesting region 1: the elemental information is fairly high, with medium-high water content and very high sulfate content. The low number of valley networks is compensated by a high number of sapping channels. Moreover, the area exhibits internal heat release (detected thermal anomalies not attributed to atmospheric and/or electromagnetic radiation) and a high number of faults, lava flows, and volcanoes. During the time of observation, it is assumed that intense seismic activity was recorded. Region 2 is less interesting from a life-potential point of view based on elemental, stratigraphic, paleotec- tonic, and geomorphologic indicator information: the hypothetical region exhibits low hydrogen abundance interpreted to be low weight percentage of water (based on Boynton et al., 2002, 2004), medium sulfate abundance, low number of faults, minor seismic activity, and medium heat release from the ground. For each of the considered hypothetical regions, the indicators are categorized and fed into the first fuzzy layer to assess the past/present water/ energy potential. 6.1. First layer fuzzy system simulation The four fuzzy systems forming the first fuzzy layer have been designed and implemented using the MATLAB fuzzy logic toolbox, which allowed expeditious modeling/asses- sing of the fuzzy-design parameters. Fig. 4 shows the conceptual scheme of one of the four first-layer fuzzy systems (i.e., the system capable of assessing past water from the value of the corresponding indicators). Here, the ‘‘past water’’ system is used as an example to show how the parameters have been selected and how the system operates. Initially, the indicators have a crisp value, and they must be fuzzified by a defined membership function. Fig. 5 shows the selected membership functions for the ‘‘past water’’ case. The membership functions, which map th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Sapping-Channels Valley-Networks Sulfates PastWater (mamdani) Past-Water Fig. 4. Conceptual scheme for fuzzy system #2 on the first layer. The system is designed to ingest past water indicators and output the past water potential. In the current simulation, sapping channels, valley networks and sulfates content have been identified to be past water indicators. R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 17 the crisp values of the ‘‘past water’’ indicators into truth values, ranging between 0 and 1, are fundamental design parameters of the system. The membership functions quantify our concept of ‘‘HIGH’’, ‘‘MEDIUM’’, and ‘‘LOW’’. For example, a locale containing sulfates in the 0–10% range is characterized as ‘‘LOW’’, 10–16% as ‘‘MEDIUM’’, and values higher than 16% as ‘‘HIGH’’, quantified using a generalized bell curve. The same technique was applied to sapping channels and valley networks. Note that the latter parameter is mapped by the number of networks in terms of order of magnitude (0 ¼ no network, 1 ¼ 10 networks, 2 ¼ 100 networks, 3 ¼ 1000 networks, etc.), again emphasizing that such parameters are open to revision and flexible enough to accommodate changes (one planetary surface may merit a different parameter than another planetary surface). Fig. 5 also shows the membership functions quantifying HIGH, MEDIUM, and LOW for the output for the ‘‘past water’’ fuzzy system. The remaining three fuzzy systems, consti- tuting the backbone of the first layer, have similar membership functions defined for both inputs and outputs, which are appropriately scaled to define HIGH, MED- IUM, and LOW for the corresponding range of the input parameters. After fuzzification, the inputs are processed by the IF- THEN rules associated with the corresponding fuzzy system. The knowledge-base, established in Sections 5.1.1–5.1.4, has been reduced to accommodate the nine available indicators. Table 11 summarizes the rules for the ‘‘past water’’ fuzzy system included in the first layer. For any of the four first-layer fuzzy systems, the rules are fired concurrently to map out and characterize the potential for past/present water and energy. After fuzzification, if the antecedent of any rule is composed of several parts connected by logical operators (e.g., AND, OR), the fuzzy logical operator is applied to combine the antecedent and provide the support for the rule. Conventional MIN and MAX constructs (see Section 3.4) have been used for AND/OR operators, respectively. Next, the implication method is applied to every rule to appropriately shape the output functions. The MIN implication method was used Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 whenever the support of the rule is employed to cut the output membership function (see Section 3.4). All rules are finally aggregated via summation of all possible reshaped output functions, and the system is defuzzified using the centroid method (i.e., the fuzzy output is defuzzified by computing the centroid of the area under the output membership function). Figs. 6 and 7 show the overall implication process that yielded the fuzzy-aggregated output and the defuzzified output (crisp number) for the ‘‘past water’’ system in regions 1 and 2, respectively. Interestingly, the past-water characterization, autonomously evaluated by the fuzzy expert system, is different for both regions, since region 1 exhibits high past-water content (80.1/100) whereas region 2 shows a low past-water content (34.8/100). The outputs of the four fuzzy systems (first fuzzy-layer) are reported in Table 12. 6.2. Second layer fuzzy system simulation The fuzzy system’s first layer was designed to provide an assessment of the potential of energy and water both currently present and once existing within the regions. While this information might be used independently by the users via a direct interface with the system, a second layer has been implemented to perform autonomous assessment of the expressed PH in the regions under investigation. Using the knowledge-base established in Section 5.2, and employing similar methods and software outlined in Section 6.1, the second layer fuzzy system was designed and connected to the first layer for rapid streaming of information/assessment. Fig. 8 shows the second layer schematic. Each of the four inputs is fuzzified using a set of membership functions quantifying the relation between input crisp values and corresponding degree of truth for linguistic terms such as HIGH, MEDIUM and LOW. Fig. 9 shows the employed input membership functions. Fig. 10 also shows the membership functions that have been used to quantify the output. VERY HIGH and VERY LOW statements were also included in the characterization process. The th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Fig. 5. Membership functions for fuzzy system #2 (Past water, first-layer). The four panels show the selected membership functions for three inputs (sapping channels, valley networks and sulfates) and one output (past water potential). For each of the input-output parameters, the ‘‘Low’’, ‘‘Medium’’ and ‘‘High’’ functions have been established using a generalized bell curve. While the shapes are defined by selecting the function-type, additional considerations and team’s experience in working with the input indicators were required to select a suitable parametric form. Table 11 Fuzzy rules summary for the fuzzy system #2 (past water, first layer) employed in the simulation Sapping channel Valley networks Sulfates Past water Connector CF Rule no. H H None 1 1 M M None 1 2 H H None 0.9 3 M M None 0.9 4 H H None 0.9 5 M M None 0.9 6 L L L L AND 1 7 The extended past-water fuzzy knowledge base (Table 4) has been modified and adapted to operate with the three available inputs (i.e. sapping channels, valley networks and sulfates). Seven (7) rules are required to operate the system. The rules are presented in a convenient and synthetic form. For example the first row show two H (under sulfates and past water), no connector, and confidence factor equal to 1. The rule should be read as follow: ‘‘IF Sulfates content is High THEN Past Water Potential is High (with the highest confidence factor of 1)’’. Because of the reduced number of input parameters we decided to implement one cumulative Low-rule (rule #7) which states that if all the past water indicators are low then past water potential is low. R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]18 generalized bell curve was employed for any of the possible functions. After fuzzification, the system follows a proce- dure analogous to the one described for the ‘‘past water’’ case. Table 13 reports a symbolic version of the rules employed by the system already defined in Table 9. The rules are implemented in terms of IF-THEN statements Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 and inference is performed using the MIN implication method (see Section 3.4). Defuzzification is accomplished using the centroid method (see Section 6.1). The simulation results are reported in Table 14. Figs. 10 and 11 show how the system operates to assess the potential for life habitability. Fuzzification, rule evaluation, implication, th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Sapping-Channels=3 Valley-Networks = 1 Sulfates = 18 Past-Water = 80.3 1 2 3 4 5 6 7 0 10 0 20 0 100 0 4 Fig. 6. Fuzzy rules interpretation process for past-water fuzzy system (first-layer) as applied to region #1. The figure illustrates how the overall rule interpretation/implication operates. Each row represents one rule, according to the scheme presented in Table 11. The values for the input parameters are reported at the top of each column. First, the ‘‘crisp’’ input values are processed by the rules using the appropriate membership function (i.e. every fuzzy rule is resolved in the antecedent where the input values are fuzzified into a number between 0 and 1, see yellow regions in the panels belonging to the first three columns). The latter value is generally called ‘‘degree of support for the rule’’. Subsequently, fuzzy operators (AND/OR) are applied if the antecedent is composed of multiple parts (see rule #7). In the next step, implication is applied by taking the degree of support for the rule and using it to shape the past water membership function (blue regions under the past water columns). Finally, the shaped membership functions for the output are aggregated (see bottom-right blue panel), the area computed and its centroid defined to be the output crisp value (defuzzyfication). Past water potential for region #1 is evaluated to be 80.1/100. R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 19 and aggregation are reported to show the step-by-step procedure as explained above. 6.3. Discussion of the simulation results The simulations were aimed at designing two intercon- nected fuzzy layers to mimic the operations of the proposed fuzzy logic-based expert system for tier-scalable reconnais- sance. The overall system was tested on hypothetical scenarios based on the assumption that real-time streaming data were converted into the absolute value of a limited number of indicators available to the system for proper habitability assessment. It is clear from the reported data (i.e., the values of the indicators for the observed regions) that region 1 is a potential candidate locale for further exploration/deployment. The system confirms the human prediction. In fact, the system exhibits high levels of geomorphologic, thermal, elemental, spectral, and paleo- tectonic information that collectively points toward a high life potential. Based on the fuzzy assessment, the intelligent system will transmit the results to the central navigation and control system of the tier-scalable reconnaissance mission architecture, which will define a new plan of action and send the appropriate commands for further deploy- Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 ment. For example, based on the fuzzy expert system results, further ground deployment of more agents may be ordered to collect higher resolution data that in turn may yield additional proof for life existence. In the region 2 scenario, the system assesses a low PH, and therefore, the fuzzy expert will inform the central navigation and control system not to consider the area as a candidate for further investigation. In fact, the life indicators of region 2 do not collectively describe a locale of elevated life potential. Water and energy indicators are generally low and the fuzzy logic expert system corroborates the same conclu- sions as provided by the human expert. Importantly, the design and simulation of the two-layer fuzzy expert system have been accomplished through proof-of-concept. Design parameters have been chosen to show the basic ideas underlying the approach required to successfully design and implement an intelligent system suitable for the integration into a tier-scalable reconnais- sance mission architecture. For example, particular mem- bership functions, implication methods, and defuzzifications were selected to complete the design, while realizing that other alternatives are available (e.g., Gaus- sian membership functions, singleton method for defuzzi- fication, etc., see also Ross, 2004). These parameters can be th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Sappings-Channels = 1 Valley-Networks = 1 Sulfates = 10 Past-Water = 34.8 1 2 3 4 5 6 7 0 10 0 4 0 20 0 100 Fig. 7. Fuzzy rules interpretation process for past-water fuzzy system (first-layer) as applied to region #2. The overall rule interpretation and implication methodology is applied to the second region of interest to determine its past water potential. The fuzzy system ingests the three reported input values and output 34.8/100 as assessment of the past water potential. Table 12 Fuzzy first-layer output First-layer fuzzy systems: outputs First-layer fuzzy systems Output region #1 Output region #2 Present water potential (PrW) 57.7/100 15.5/100 Past water potential (PsW) 80.3/100 34.8/100 Present energy potential (PrE) 77.9/100 51.1/100 Past energy potential (PsE) 86.4/100 40.9/100 The observed life indicators are categorized as past/present water and energy indicators and fed to the four independent fuzzy systems for past/ present water and energy potential assessment. The table reports the results for the two hypothesized regions. R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]20 generally changed to tune the system and improve performance. Indeed, a full-scale fuzzy expert system design (including all ‘‘known’’ (to date) life-containing indicators) requires trade studies on the parameters and a multiple-iteration approach to converge to the optimal system design. Nevertheless, the performed design/simula- tion lays the foundation for and outlines the basic structure of an autonomous, intelligent expert system that may be integrated into tier-scalable reconnaissance mission archi- tectures. 7. Conclusions, implications, and future efforts If there is any chance of finding life beyond Earth, as well as testing overarching theories concerning the evolu- Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 tion of planetary bodies (e.g., GEOMARS proposed by Baker et al., 2007, Superplume book), a paradigm shift in planetary exploration is required. The novel tier- scalable reconnaissance mission paradigm of Fink et al. (2005a–c, 2006a, b, 2007) will fundamentally change the way current missions are conducted by allowing less constrained, science-driven planetary exploration via deployment of a multi-tier system of sensor platforms (yielding spaceborne, airborne, surface, and subsurface perspectives), approximating the approach of a geologist/biologist/chemist. ‘‘Autonomy’’ is a key factor in achieving the desired success in future planetary reconnaissance (i.e., detecting, characterizing, and homing in on features of special interest, including transient events). Here, a fuzzy expert system is proposed and described to autonomously identify locales with the greatest PH. This system may serve as the part of a multi-tier architecture that elaborates newly acquired information, compiles it with existing information, and performs comparative analysis of the compiled spatial and temporal information while in-transit. The proposed system has been conceived specifically with the goal of mimicking the approach of a planetary geologist, while coupling the expertise of a terrestrial biologist or chemist. The system assesses and evaluates the life-containing potential of a locale through utilization of all possible clues coming from collecting multiple layers of information, which may include elemental, spectral, atmospheric, geophysical, hydrologic, geomorphologic, th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Low 1 0.8 0.6 0.4 0.2D e g re e o f m e m b e rs h ip D e g re e o f m e m b e rs h ip 0 1 0.8 0.6 0.4 0.2 0 0 10 20 30 40 Medium High Low Medium High Very-HighVery-Low 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Present Water Potential for Habitability Fig. 9. Membership functions (MFs) employed for the second-layer fuzzy system. The top panel shows the MFs used for one of the input parameters (PrW, range between 0 and 100). Low, medium and high have been implemented using the generalized bell curve. The other input potentials have been associated with membership functions with identical structure. The bottom panel illustrates the MFs associated with the output parameter (PH). In addition to the standard ‘‘High’’, ‘‘Medium’’ and ‘‘Low’’, two extra MFs have been considered to introduce the ‘‘Very High’’ and ‘‘Very Low’’ statements. Present-Water Present-Energy Past-Water Past--Energy Water-Energy04 (mamdani) Potential-for-Habitabilty Fig. 8. Conceptual scheme for the second-layer fuzzy system. The system accepts four inputs (present water, past water, present energy and past energy potential) and output the potential for habitability (PH). R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 21 stratigraphic, topographic, and paleotectonic information. The compiled information must be interpreted to char- acterize life potential. Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 In order to effectively design an intelligent system that can autonomously execute the desired task to look for life, the knowledge-base that condenses the multidisciplinary th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Present-Water = 57.7 Past-Water =80.3 Present-Energy=77.9 Past-Energy=68.4 Potential-for- Habitability=81.2 1 2 3 4 5 6 7 8 9 10 11 12 13 0 100 0 100 0 100 0 100 0 100 Fig. 10. Fuzzy rules interpretation process for the second-layer fuzzy system as applied to region #1. The figure illustrates how the overall rules interpretation/implication operates. The fuzzy knowledge-base is comprised of 13 rules, which are operated concurrently. The fuzzy rule interpretation and implication method has been already explained (see Fig. 6). The second-layer system proceeds in similar fashion using the same mechanism to evaluate PLH. Table 13 Fuzzy rules summary for the second-layer fuzzy system employed in the simulation PrW PsW PrE PsE PH Connect CF Rule no. H H H H VH AND 1 1 L L L L VL AND 1 2 H H H AND 0.8 3 L L L AND 0.8 4 M M M M M AND 0.8 5 H H H AND 0.6 6 H H H AND 0.6 7 L L L AND 0.6 8 L L L AND 0.6 9 H H H OR 0.4 10 L L OR 0.4 11 H H H OR 0.2 12 L L L OR 0.2 13 The extended version the fuzzy knowledge base is reported in Table 9. R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]22 expertise must be defined and constructed. The fuzzy logic framework provides a way to translate expertise (including both practical and theoretical knowledge) into simple rules that can be understood by a digital computer. The main goal of this paper was to lay the foundation for a fuzzy knowledge-base that can be constructed and implemented in a digital form to autonomously evaluate the PH of the Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 locale under investigation through tier-scalable reconnais- sance. Hypothetical deployments to two different regions of Mars were examples for demonstrating proof-of-concept and for showing how fuzzy-based rules can be implemented to characterize life-containing potential. For hypothesized martian scenarios the system autonomously reached the same conclusions as a field expert, showing self-consis- tency. It is important to stress that the presented approach is an essential part of the basic foundation currently built by our team of what we think can be the most effective way to perform planetary exploration in the future. Fuzzy logic, applied to tier-scalable reconnaissance mission architec- tures, can effectively provide the intelligence that is required for autonomous less constrained and science- driven planetary reconnaissance. The fuzzy logic expert system is shown to be flexible enough to accommodate any finite number of indicators and rules. This foundational work on autonomy for tier-scalable reconnaissance mis- sions could initiate an open forum (likened to a field camp for geologists) for the entire planetary science community as well as for any other scientific community (e.g., Math, Engineering, Biology, Chemistry, Physics, Geology, Hy- drology) interested in space exploration, to discuss the role of the indicators, the number of indicators, the most appropriate rules that should be used to search for life, or any other objective on planetary bodies. As shown in the th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS Table 14 Second-layer fuzzy system simulation results Fuzzy expert system potential for life habitability assessment Input region #1 PH region #1 Input region #2 PH region #2 Present water 57.7/100 81.2/100 15.5/100 24.6/100 Past water 80.3/100 34.8/100 Present energy 77.9/100 51.1/100 Past energy 86.4/100 40.9/100 The table shows the input values presented to the second-layer fuzzy system for region #1 and region #2 and the output of the system for both scenarios. The table shows that region #1 is the area with the highest potential for habitability and therefore a region worth of further deployment/investigation. Present-Water = 15.5 Present-Energy = 51.1Past-Water = 34.8 Past-Energy = 40.9 Potential-for- Habitabilty=24.6 1 2 3 4 5 6 7 8 9 10 11 12 0 100 0 100 0 100 0 100 0 100 Fig. 11. Fuzzy rules interpretation process for the second-layer fuzzy system as applied to region #2. The figure illustrates how the overall rules interpretation/implication operates (see Fig. 6 for details about the rule interpretation and implication method). R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 23 above sections, the system is modular, i.e., is constructed using a sequence of interconnected fuzzy systems and therefore has the flexibility to add and delete elements depending on the goal to be achieved. Future challenges will therefore include the extension of the concept to multiple real-world planetary scenarios to confirm and test working hypotheses, to review and evaluate the rules and indicators to tailor-fit to various planetary bodies, as well as to implement the designed software in tier-scalable Earth-based deployed hardware to test the effectiveness of the system in selected environments on Earth. The latter will be a critical step to promote the transition of tier- scalable reconnaissance architectures from a concept to an operational system for autonomous planetary reconnais- sance. Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 References Acuña, M.H., Connerney, J.E.P., Ness, N.F., Lin, R.P., Mitchell, D., Carlson, C.W., Mcfadden, J., Anderson, K.A., Reme, H., Mazelle, C., Vignes, D., Wasilewski, P., Cloutier, P., 1999. Global distribution of crustal magnetization discovered by the Mars Global Surveyor MAG/ER experiment. Science 284, 790–793. Anderson, R.C., Dohm, J.M., Golombek, M.P., Haldemann, A., Franklin, B.J., Tanaka, K.L., Lias, J., Peer, B., 2001. Significant centers of tectonic activity through time for the western hemisphere of Mars. J. Geophys. Res. 106, 20,563–20,585. Baker, V.R., Strom, R.G., Gulick, V.C., Kargel, J.S., Komatsu, G., Kale, V.S., 1991. Ancient oceans, ice sheets and the hydrological cycle on Mars. Nature 352, 589–594. Baker, V.R., Maruyama, S., Dohm, J.M., 2007. Superplume and the geological evolution of early Mars. In: Yuen, D.A., Maruyama, S., th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]]24 Karato, S., Windley, B.F. (Eds.), Superplume: Beyond Plate Tectonics. Springer, Berlin. Boston, P.J., 2004. Extraterrestrial caves. In: Gunn, J. (Ed.), Encyclopedia of Cave and Karst Science. Fitzroy-Dearborn Publishers, Ltd., London, pp. 355–358. Boston, P.J., Ivanov, M.V., McKay, C.P., 1992. On the possibility of chemosynthetic ecosystems in subsurface habitats on Mars. Icarus 95, 300–308. Boynton, W.V., Feldman, W.C., Squyres, S.W., Prettyman, T.H., Brückner, J., Evans, L.G., Reedy, R.C., Starr, R., Arnold, J.R., Drake, D.M., Englert, P.A.J., Metzger, A.E., Mitrofanov, I., Trombka, J.I., d’Uston, C., Wänke, H., Gasnault, O., Hamara, D.K., Janes, D.M., Marcialis, R.L., Maurice, S., Mikheeva, I., Taylor, G.J., Tokar, R., Shinohara, C., 2002. Distribution of hydrogen in the near surface of Mars: evidence for subsurface ice deposits. Science 297, 81–85. Boynton, W.V., Feldman, W.C., Mitrofanov, I.G., Evans, L.G., Reedy, R.C., Squyres, S.W., Starr, R., Trombka, J.I., d’Uston, C., Arnold, J.R., Englert, P.A.J., Metzger, A.E., Wänke, H., Brückner, J., Drake, D.M., Shinohara, C., Fellows, C., Hamara, D.K., Harshman, K., Kerry, K., Turner, C., Ward, M., Barthe, H., Fuller, K.R., Storms, S.A., Thornton, G.W., Longmire, J.L., Litvak, M.L., Ton’ Chev, A.K., 2004. The Mars odyssey gamma-ray spectrometer instrument suite. Space Sci. Rev. 110, 37–83. Carr, M.H., et al., 1998. Evidence for a subsurface ocean on Europa. Nature 391, 363–365. Connerney, J.E.P., Acuña, M.H., Ness, N.F., Kletetschka, G., Mitchell, D.L., Lin, R.P., Reme, H., 2005. Tectonic implications of Mars crustal magnetism. Science 102, 14970–14975. Dohm, J.M., Tanaka, K.L., Hare, T.M., 2001a. Geologic map of the Thaumasia region of Mars. USGS Misc. Inv. Ser. Map I-2650, scale 1:5,000,000. Dohm, J.M., Ferris, J.C., Baker, V.R., Anderson, R.C., Hare, T.M., Strom, R.G., Barlow, N.G., Tanaka, K.L., Klemaszewski, J.E., Scott, D.H., 2001b. Ancient drainage basin of the Tharsis region, Mars: potential source for outflow channel systems and putative oceans or paleolakes. J. Geophys. Res. 106, 32,943–32,958. Dohm, J.M., Maruyama, S., Baker, V.R., Anderson, R.C., Ferris, J.C., Hare, T.M., 2002. Plate tectonism on early Mars: diverse geological and geophysical evidence. Lunar Planet. Sci. Conf. XXXIII #1639 (abstract) [CD-ROM]. Dohm, J.M., Ferris, J.C., Barlow, N.G., Baker, V.R., Mahaney, W.C., Anderson, R.C., Hare, T.M., 2004. The Northwestern Slope Valleys (NSVs) region, Mars: A prime candidate site for the future exploration of Mars. Planet. Space Sci. 52, 189–198. Dohm, J.M., Maruyama, S., Baker, V.R., Anderson, R.C., 2006a. Traits and evolution of the Tharsis Superplume, Mars. In: Yuen, D.A., Maruyama, S., Karato, S., Windley, B.F. (Eds.), Superplume: Beyond Plate Tectonics. Springer, Berlin. Dohm, J.M., Anderson, R.C., Baker, V.R., Barlow, N.G., Miyamoto, H., Davies, A.G., Taylor, G.J., Boynton W.V., Keller, J., Kerry, K., Janes, D., Fairén, A.G., Schulze-Makuch, D., Glamoclija, M., Marinangeli, L., Ori, G.G., Strom, R.G., Williams, P., Ferris, J.C., Rodrı́guez, J.A.P., de Pablo Hdez, M.A., Karunatillake, S., 2006b. Tharsis/Elysium corridor: a marker for an internally active Mars? [Abstract 1131] In: 37th Lunar Planetary. Science Conference, [CD- ROM]. Lunar and Planetary Institute, Houston. Dubois, D., Prade, H., 1980. Fuzzy Sets and Systems: Theory and Applications. Academic Press, New York. Elachi, C., Wall, S., Janssen, M., Stofan, E., Lopes, R., Kirk, R., Lorenz, R., Lunine, J., Paganelli, F., Soderblom, L., Wood, C., Wye, L., Zebker, H., Anderson, Y., Ostro, S., Allison, M., Boehmer, P., Callahan, P., Encrenaz, P., Flamini, E., Francescetti, G.Y., Gim, Y., Hamilton, G., Hensley, S., Johnson, W., Kelleher, K., Muhleman, D., Picardi, G., Posa, F., Roth, L., Seu, R., Shaffer, S., Stiles, B., Vetrella, S., West, R., 2006. Titan Radar Mapper observations from Cassini’s T3 flyby. Nature 441, 709–713. Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 Fairén, A.G., Dohm, J.M., Baker, V.R., de Pablo, M.A., Ruiz, J., Ferris, J., Anderson, R.M., 2003. Episodic flood inundations of the northern plains of Mars. Icarus 165, 53–67. Fairén, A.G., Dohm, J.M., Uceda, E.R., Rodrı́guez, A., Baker, V.R., Fernández-Remolar, D., Schulze-Makuch, D., Amils, R., 2005. Prime candidate sites for astrobiological exploration through the hydro- geological history of Mars. Planet. Space Sci. 53, 1355–1375. Fairén, A. G., Schulze-Makuch, D., Davila, A., Rodrı́guez, A. P., Fink, W., Furfaro, R., Uceda, E. R., Amils, R., McKay, C. P., 2008. Evidence for Amazonian acidic liquid water on Mars from the Mars Exploration Rovers mission. Planetary and Space Science, submitted. Feldman, W.C., Boynton, W.V., Tokar, R.L., Prettyman, T.H., Gasnault, O., Squyres, S.W., Elphic, R.C., Lawrence, D.J., Lawson, S.L., Maurice, S., McKinney, G.W., Moore, K.R., Reedy, R.C., 2002. Global distribution of neutrons from Mars: results from Mars Odyssey. Science 297, 75–78. Fink, W., 2006. Generic Prioritization Framework for Target Selection and Instrument Usage for Reconnaissance Mission Autonomy. In: Proceedings of IEEE World Congress on Computational Intelligence (WCCI) 2006, Vancouver, Canada, pp. 11116–11119. Fink, W., Dohm, J.M., Tarbell, M.A., Hare, T.M., Baker, V.R., 2005a. Next-generation robotic planetary reconnaissance missions: a para- digm shift. Planet. Space Sci. 53, 1419–1426. Fink, W., Dohm, J.M., Tarbell, M.A., Hare, T.M., Baker, V.R., 2005b. Next-generation robotic planetary surface/subsurface reconnaissance missions: a paradigm shift [abstract 1977]. In: 36th Lunar and Planetary Science Conference Abstracts [CD-ROM], Lunar and Planetary Institute, Houston. Fink, W., Dohm, J.M., Tarbell, M.A., Hare, T.M., Baker, V.R., 2005c. Next-generation robotic planetary reconnaissance missions: a para- digm shift. In Abstracts of the 15th Annual V.M. Goldschmidt Conference, Moscow, Idaho. Geochimica et Cosmochimica Acta, Vol. 69, No. 10S, p. A533. Fink, W., Datta, A., Baker, V.R., 2005d. AGFA: Automated geologic field analyzer. In: Abstracts of the 15th Annual V.M. Goldschmidt Conference, Moscow, Idaho. Geochimica et Cosmochimica Acta, Vol. 69, No. 10S, p. A535. Fink, W., Dohm, J.M., Tarbell, M.A., Hare, T.M., Baker, V.R., Schulze- Makuch, D., Furfaro, R., Fairén, A.G., Ferré, T.P.A., Miyamoto, H., Komatsu, G., Mahaney, W.C., 2006a. Multi-tier multi-agent Auton- omous robotic planetary surface/subsurface reconnaissance for life [abstract 1433]. In: 37th Lunar and Planetary Science Conference Abstracts [CD-ROM], Lunar and Planetary Institute, Houston. Fink, W., Dohm, J.M., Tarbell, M.A., Hare, T.M., Baker, V.R., Schulze- Makuch, D., Furfaro, R., Fairén, A.G., Ferré, T.P.A., Miyamoto, H., Komatsu, G., Mahaney, W.C., 2006b. Autonomous tier-scalable reconnaissance missions for remote planetary exploration. In: Pro- ceedings of the Forth International Planetary Probe Workshop 2006, Pasadena. Fink, W., Dohm, J.M., Tarbell, M.A., Hare, T.M., Baker, V.R., Schulze- Makuch, D., Furfaro, R., Fairén, A.G., Ferré, T.P.A., Miyamoto, H., Komatsu, G., Mahaney, W.C., 2007. Tier-scalable reconnaissance missions for the autonomous exploration of planetary bodies. IEEE Aerospace Conference Proceedings, Big Sky, Montana. Fink, W., et al., 2008. Automated global feature analyzer (AGFA)—a driver for tier-scalable reconnaissance. IEEE Aerospace Conference Proceedings, Big Sky, Montana. Greeley, R., Thomas, P.E. (Eds.), 1994. Mars landing site catalog. NASA Ref. Pub. 1238, 392pp. Kargel, J.S., 1995. Cryovolcanism on the icy satellites. Earth Moon Planet 67, 101–113, 10.1007/BF00613296. Kargel, J.S., 2004. Mars: A Warmer Wetter Planet. Praxis-Springer (Publ.), 557pp. Kargel, J.S., 2006. Enceladus: cosmic gymnast, volatile miniworld. Science 311, 1389–1391, 10.1126/science.1124495. Kasabov, N., 1996. Foundations of Neural Networks, Fuzzy Systems, and Knowledge Engineering (Computational Intelligence). The MIT Press, Boston MA. th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 ARTICLE IN PRESS R. Furfaro et al. / Planetary and Space Science ] (]]]]) ]]]–]]] 25 Kaufmann, A., Gupta, M.M., 1985. Introduction to Fuzzy Arithmetic. V.N. Reinhold. Komatsu, G., Ori, G.G., 2000. Exobiological implications of potential sedimentary deposits on Mars. Planet. Space Sci. 48/11, 1043–1052. Komatsu, G., Dohm, J.M., Hare, T.M., 2004. Hydrologeologic processes of large-scale tectonomagmatic complexes in Mongolia—southern Siberia and on Mars. Geology 32, 325–328. Lorenz, R.D., Wall, S., Radebaugh, J., Boubin, G., Reffet, E., Janssen, M., Stofan, E., Lopes, R., Kirk, R., Elachi, C., Lunine, J., Paganelli, F., Soderblom, L., Wood, C., Wye, L., Zebker, H., Anderson, Y., Ostro, S., Allison, M., Boehmer, R., Callahan, P., Encrenaz, P., Ori, G.G., Franceschetti, G., Gim, Y., Hamilton, G., Hensley, S., Johnson, W., Kelleher, K., Muhleman, D., Picardi, G., Posa, F., Roth, L., Seu, R., Shaffer, S., Stiles, B., Vetrella, S., Flamini, E., West, R., 2006. The sand seas of Titan: Cassini RADAR observations of longitudinal dunes. Science 312, 724–727. Mahaney, W.C., Milner, M.W., Netoff, D.I., Malloch, D., Dohm, J.M., Baker, V.R., Miyamoto, H., Hare, T.M., Komatsu, G., 2004. Ancient wet aeolian environments on Earth: clues to presence of fossil/live microorganisms on Mars. Icarus 171, 39–53. Mamdani, E.H., 1977. Applications of fuzzy logic to approximate reasoning using linguistic synthesis. IEEE Trans. Comput. 26 (12), 1182–1191. Mamdani, E.H., Assilian, S., 1975. An experiment in linguistic synthesis with a fuzzy logic controller. Int. J. Man-Machine Stud. 7 (1), 1–13. Márquez, A., Fernández, C., Anguita, F., Farelo, A., Anguita, J., de la Casa, M.-A., 2004. New evidence for a volcanically, tectonically, and climatically active Mars. Icarus 172, 573–581. McEwen, A.S., Eliason, E.M., Bergstrom, J.W., Bridges, N.T., Hansen, C.J., Delamere, W.A., Grant, J.A., Gulick, V.C., Herkenhoff, K.E., Keszthelyi, L., Kirk, R.L., Mellon, M.T., Squyres, S.W., Thomas, N.T., Weitz, C.M., 2007. Mars Reconnaissance orbiter’s high resolution imaging science experiment (HiRISE). J. Geophys. Res. 112, E05S02. Mitchell, K.L., Wilson, L., 2003. Mars: a geologically active planet. Astron. Geophys. 44 (4), 4.16–4.20. Mouginis-Mark, P.J., 1985. Volcano/ground ice interactions in Elysium Planitia, Mars. Icarus 64, 265–284. Mouginis-Mark, P.J., 1990. Recent water release in the Tharsis region of Mars. Icarus 84, 362–373. Ori, G.G., Marinangeli, L., Komatsu, G., 2000. Martian paleolacustrine environments and their geological constrains on drilling operations for exobiological research. Planet. Space Sci. 48 (11), 1027–1034. Perron, J.T., Lamb, M.P., Koven, C.D., Fung, I.Y., Yager, E., Ádámkovics, M., 2006. Valley formation and methane precipitation rates on Titan. J. Geophys. Res. 111, E11001. Porco, C.C., Helfenstein, P., Thomas, P.C., Ingersoll, A.P., Wisdom, J., West, R., Neukum, G., Denk, T., Wagner, R., Roatsch, T., Kieffer, S., Please cite this article as: Furfaro, R., et al., The search for life beyond Ear j.pss.2007.09.006 Turtle, E., McEwen, A., Johnson, T.V., Rathbun, J., Veverka, J., Wilson, D., Perry, J., Spitale, J., Brahic, A., Burns, J.A., DelGenio, A.D., Dones, L., Murray, C.D., Squyes, S., 2006. Cassini observes the active south pole of Enceladus. Science 311, 1393–1401. Rodriguez, J.A.P., Sasaki, S., Dohm, J.M., Tanaka, K.L., Strom, B., Kargel, J., Kuzmin, R., Miyamoto, Spray, J.G., Fairen, A.G., Komatsu, G., Kurita, K., Baker, V., 2005. Control of impact crater fracture systems on subsurface hydrology, ground subsidence and collapse, Mars. J. Geophys. Res. 110, E06003. Rodriguez, J.A.P., Kargel J., Crown, D., Bleamaster III, L.F., Dohm, J.M., Miyamoto, H., Baker, V., Sasaki, S., Komatsu, G., 2006. Headward growth of chasmata by volatile outbursts, collapse, and drainage, Mars. Geophysical Research Letters, 33, L18203, doi:10.1029/2006GL026275. Ross, T.J., 2004. Fuzzy Logic with Engineering Applications, second ed. JohnWiley, Hoboken, NJ. Schulze-Makuch, D., Dohm, J.M., Fairén, A.G., Baker, V.R., Fink, W., Strom, R.G., 2005a. Venus, Mars, and the ices on Mercury and the Moon: astrobiological implications and proposed mission designs. Astrobiology 5, 778–795. Schulze-Makuch, D., Irwin, L.N., Lipps, J.H., LeMone, D., Dohm, J.M., Fairén, A.G., 2005b. Scenarios for the evolution of life on Mars. J. Geophys. Res.—Planets 110, E12S23 (special edition on Early Mars). Scott, D.H., Tanaka, K.L., 1986. Geologic map of the western equatorial region of Mars. USGS Misc. Inv. Ser. Map I-1802-A (1:15,000,000). Scott, D.H., Dohm, J.M., Rice Jr., J.W., 1995. Map of Mars showing channels and possible paleolake basins. US Geol. Survey Map I-2461. Sugeno, M., 1974. Theory of Fuzzy integrals and its Applications, Ph.D. Thesis, Tokyo Institute of Technology, Tokyo. Sugeno, M., 1985. Industrial Applications of Fuzzy Control. Elsevier Science Pub. Co. Tomasko, M., Archinal, B., Becker, T., Bézard, B., Bushroe, M., Combes, M., Cook, D., Coustenis, A., de Bergh, C., Dafoe, L., Doose, L., Douté, S., Eibl, A., Engel, S., Gliem, F., Grieger, B., Holso, K., Howington-Kraus, E., Karkoschka, E., Keller, H., Kirk, R., Kramm, R., Küppers, M., Lanagan, P., Lellouch, E., Lemmon, M., Lunine, J., McFarlane, E., Moores, J., Prout, M., Rizk, B., Rosiek, M., Rueffer, P., Schröder, S., Schmitt, B., See, C., Smith, P., Soderblom, L., Thomas, N., West, R., 2005. Rain, winds, and haze during the Huygens probe descent to Titan’s surface. Nature 438, 765–778. Zadeh, L.A., 1965. Fuzzy sets. Inform. Control 8, 338–353. Zadeh, L.A., 1975. The concept of a linguistic variable and its application to approximate reasoning, Parts 1, 2, and 3. Inform. Sci. 8,199–249; 8,301–357; 9,43–80. Zadeh, L.A., 1988. Fuzzy logic. Computer 1 (4), 83–93. Zadeh, L.A., 1989. Knowledge representation in fuzzy logic. IEEE Trans. Knowledge Data Eng. 1, 89–100. th through fuzzy expert systems. Planet. Space Sci. (2007), doi:10.1016/ dx.doi.org/10.1029/2006GL026275 dx.doi.org/10.1016/j.pss.2007.09.006 dx.doi.org/10.1016/j.pss.2007.09.006 The search for life beyond Earth through fuzzy expert systems Introduction Expert system design for planetary reconnaissance: problems and solutions Fuzzy logic paradigm for autonomous fuzzy expert system design Why fuzzy logic for planetary exploration? Deploying fuzzy experts on planetary bodies: Titan, Enceladus and Mars Fuzzy logic versus alternative AI schemes Fuzzy logic basics Fuzzy sets and membership function Fuzzy logic paradigm: fuzzy rules and inference mechanism Knowledge-base for the fuzzy-expert system: mimicking the scientific/operational approach of a geologist, biologist, and chemist Rationale for building the knowledge-base to assess potential for life habitability Building a fuzzy logic expert system for life habitability assessment First fuzzy-layer: past/present water/energy assessment for life-containing potential Fuzzy system #1: present water assessment Fuzzy system #2: past water assessment Fuzzy system #3: present energy assessment Fuzzy system #4: past energy assessment Second fuzzy layer: water and energy assessment for life-containing potential Testing the system on planetary scenarios: two examples First layer fuzzy system simulation Second layer fuzzy system simulation Discussion of the simulation results Conclusions, implications, and future efforts References 
work_reqecactfbfmtncjxukrd3yts4	----	Fuzzy Logic Driven Expert System for the Assessment of Software Projects Risk (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 2, 2019 153 | P a g e www.ijacsa.thesai.org Fuzzy Logic Driven Expert System for the Assessment of Software Projects Risk Mohammad Ahmad Ibraigheeth 1 , Syed Abdullah Fadzli 2 Faculty of Informatics and Computing, Universiti Sultan ZainalAbidin, 21300 Kuala Terengganu, Malaysia Abstract—This paper presents an expert risk evaluation system developed and based on up-to-date empirical study that uses a real data from huge number of software projects to identify the most factors that affect the project success. Software project can be affected by a range of risk factors through all phases of the development process. Therefore, it has become necessary to consider risk concerns while developing the software project. Risk assessment and management play a significant role in avoiding failure of the software project, and can help in mitigating the effect of the undesirable events that could affect the project outcomes. In this paper, the researchers have developed a novel expert fuzzy-logic tool that can be used by project decision makers to evaluate the expected risks .The developed tool helps in estimating the risk probability based on the software project’s critical success factors. A user-friendly interface is created to enable the project managers to perform general risk evaluation during any stage of the software development process. The proposed tool can be helpful in achieving effective risk control, and therefore improving the overall project outcomes. Keywords—Risk assessment; critical success factors; fuzzy expert systems; fuzzy rule-base; risk probability I. INTRODUCTION Risk is a probable event that might lead to undesirable impact on software project outcomes. Software risk is an unexpected problem occurs during software operations that might cause software failure [1]. Project risk assessment and management can help in mitigating the effect of the undesirable events. Identification of probable risk factors is one of the major issues in software project management. Today, the software systems are widely used by people to control and manage their daily routines, due to this fact; it has been a must to consider risk concerns when developing any software project. Developing tools to assess and manage software risks have become increasingly important for measuring the health of the software project during all phases of the software development process. All organizations should focus on managing risks related to their software projects. When risk factors are reported, risk mitigation strategies should be developed in order to avoid potential project failure. Although considering software risk concerns has become critical, there is a limited number of developed tools that can be used by the project decision makers in evaluating and mitigating the probable risks. This paper aims to develop a new expert fuzzy tool that can help project managers to evaluate the expected project risk. This tool evaluates the project risk probability based on ten critical success factors. Using fuzzy set theory is advantageous for recording linguistic variables that are usually used by project managers to describe parameters in the project development environment. A fuzzy based user-friendly tool to evaluate “risk probability” of the software project is developed to support general software project risk assessment through any phase of the software development process. The percentages of presence of ten success factors identified in CHAOS report are used as input to the model. A linguistic variable used for each input, and two membership functions are defined: NO and YES. Fuzzification process then is used to map the crisp values specified by the model users to the fuzzy space Mamdani interference system with rules base includes 1024 if-then rules used to evaluate the project risk as a fuzzy number. Finally, Defuzzification module converts this number into crisp value that represents risk probability of the software project. The developed model can be used as a tool to guide the software project decision makers in making critical decisions in early stages throughout the software development process, and in identifying alternative strategies to avoid the software probable risks. This research presents two contributions: First, it develops an expert risk evaluation system based on up-to-date survey conducted by Standish organization that uses a real data from 50,000 projects to identify the most factors that affect the project success. Second, it provides general and easy-to-use tool with user-friendly interface that enables project managers to assess the project risk during any phase of software development process. The rest of this paper is organized as follow: Section 2 describes software project success factors. Section 3 reviews the related literature. Section 4 explains the proposed model. Section 5 describes the risk evaluation tool design. Section 6 provides experimental work and analyses the behavior of the system. Section 7 concludes the research, describes its limitations, and suggests future work. II. SOFTWARE PROJECT SUCCESS FACTORS Many software project success and failure factors have been described in the literature [2-5]. In this paper, we investigate the effect of project success factors identified in CHAOS report. The report identifies ten software project success factors ranked according to their influence on the project success as shown in Table 1 [6]. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 2, 2019 154 | P a g e www.ijacsa.thesai.org TABLE I. SOFTWARE PROJECT SUCCESS FACTORS Factors of Success Impact Executive Sponsorship 15% Emotional Maturity 15% User Involvement 15% Optimization 15% Skilled Resources 10% Standard Architecture 8% Agile Process 7% Modest Execution 6% Project Management Expertise 5% Clear Business Objectives 4% The CHAOS success factors presented in Table 1 can be defined as a following [6]:  Executive sponsorship: when the executives provide a suitable financial and emotional supports, they will increase the opportunity to implement a successful software project.  Emotional maturity: this relates to project environment and how the project team work together. Having the skills to manage relationships, self-managed and socially aware, can help in producing more successful projects.  User involvement: when users are not involved, the project will perform poorly. User participation in project decision making, and through requirements understanding phase has a major positive effect on project success.  Optimization: optimization of some project aspects can maximize the project efficiency. This includes optimization the scope based on the project sponsorship capabilities, and identifying the optimal team size.  Skilled resources: the project success is made up by staff who have the necessary skills to understand and perform the project requirements.  Standard architecture management environment (SAME): SAME is defined by the Standish Group as a collection of consistent behaviors including the integration of services, practices, and products in software development process.  Agile process: it describes a set of values including adaptive planning, flexible response to change, early delivery, and continuous improvements. These principles support producing successful projects.  Modest execution: it takes place when the process has few and simple moving parts, and when the tools used in project development process have few features used sparingly.  Project management expertise: is the use of knowledge, skills, procedures and techniques in the project development activities to achieve the desired project goals, and meet the organization requirements. III. RELATED WORK Numerous techniques have been used to address and manage the software risks. A software risk management framework is proposed by Boehm [7]. He defined list of top software risks depending on his experience. There were some limitations in his study. No theoretical foundations were presented in his work. Also, as he identified the risks in 1991, these risks have become inadequate as the software development environment has increasingly become more complex and diverse. Another survey was conducted by Barki et al. [8]. A list of 23 software risks is identified and classified into five sets. The complexity of assessment scale that was used for each risk posed a limitation. Schmidt et al. [9] also conducted a survey by integration of many experts opinion to identify 53 software risks. These risks were grouped into 14 sets. As the experts were from different countries, the study declared that the list could be affected and have become inapplicable. Wallace et al. [10] defined 27 software risks and classified them into 6 dimensions (i.e., user, requirements, complexity, planning, staff, and development environment) by performing cluster analysis to develop model that measure the software project risk.. Performing cluster analysis is helpful in finding variable similarities to perform accurate prediction. Artificial intelligent approaches also used widely to counter and manage the software risks. A regression analysis method is used in research proposed by Jiang and Klein [11] to define the most risk factors that affect the process of project development. The impact of applying a certain management activities on the software project outcomes is considered [12]. A genetic algorithm combined with decision trees is an approach for risk prediction by using certain software metrics developed by Xu z et al. [13]. A fuzzy logic is used in developing system to evaluate the software risks through earlier phase of software development cycle [14]. Yavari et al. [15] proposed a method based on Wallace’s [10] work to assess software risk using fuzzy logic. Neural networks are used to identify software projects with high risk [16]. Hu Y et al. [17] proposed a framework for risk analysis based on risk causality using Bayesian networks. Each of these techniques has its own advantages. For example, regression analysis is suitable for risk prediction as it can find the relationships between variables. Applying decision trees is fast and simple while neural network is suitable when the relationships between the system variables are non-linear. Applying Bayesian network with considering causality dependencies can perform better prediction. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 2, 2019 155 | P a g e www.ijacsa.thesai.org The main advantage of our approach is developing a novel tool for software risk assessment based on critical success factors. The primary objective of our work is to perform general risk evaluation that can be done through any stage of software development life cycle (SDLC). The proposed tool can be helpful in achieving effective risk control, and therefore improving the overall project outcomes. IV. PROPOSED SOFTWARE RISK ASSESSMENT MODEL In this paper, ten success factors that are identified in CHAOS report [6] are used (refer to Table 1). Fig. 1 shows our model. The final output of this model is the software risk probability due to the mentioned ten factors. Fuzzy Logic toolbox in MATLAB is used to implement Mamdani inference system. The following steps (shown in Fig. 2) explain how the model works: Step 1: Fuzzification In this step, crisp values (within the range of 0 to 100) for the ten input variables are measured. A scale mapping then performed for these inputs to obtain their membership values within the range of 0 to 1.Two trapezoidal membership functions (similar to Fig. 3). We might interpret NO as: input percentage of presence below 50%, and YES as: input percentage of presence higher than 50%. Fig. 1. Risk Evaluation Model. Fig. 2. Risk Evaluation Steps. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 2, 2019 156 | P a g e www.ijacsa.thesai.org Fig. 3. Trapezoidal Membership Functions (Trapmf). Step 2: Rules Evaluation The rule base includes 1024 IF-THEN rules. The following are samples of the created rules:  Rule 1:If (Executive_Sponsorship is YES) and (Emotional_Maturity is YES) and (User_Involvement is YES) and (Optimization is YES) and (Skilled_Resources is YES) and (Standard_Architecture is NO) and (Agile_Process is YES) and (Modest_Execution is YES) and (Project_Management_Expertise is YES) and (Clear_Business_Objectives is YES) then (RiskProbability is NONRISKY)  Rule 10: If (Executive_Sponsorship is YES) and (Emotional_Maturity is YES) and (Us)  er_Involvement is YES) and (Optimization is YES) and (Skilled_Resources is YES) and (Standard_Architecture is NO) and (Agile_Process is NO) and (Modest_Execution is YES) and (Project_Management_Expertise is YES) and (Clear_Business_Objectives is NO) then (RiskProbability is NONRISKY)  Rule 127: If (Executive_Sponsorship is YES) and (Emotional_Maturity is YES) and (User_Involvement is YES) and (Optimization is NO) and (Skilled_Resources is NO) and (Standard_Architecture is YES) and (Agile_Process is NO) Optimization is YES) and (Skilled_Resources is YES) and (Standard_Architecture is NO) and (Agile_Process is NO) and (Modest_Execution is NO) and (Project_Management_Expertise is YES) and (Clear_Business_Objectives is YES) then (RiskProbability is RISKY)  Rule 397:If (Executive_Sponsorship is YES) and (Emotional_Maturity is NO) and (User_Involvement is NO) and (Optimization is YES) and (Skilled_Resources is YES) and (Standard_Architecture is NO) and (Agile_Process is NO) and (Modest_Execution is NO) and (Project_Management_Expertise is YES) and (Clear_Business_Objectives is YES) then (RiskProbability is RISKY)  Rule 1024:If (Executive_Sponsorship is NO) and (Emotional_Maturity is NO) and (User_Involvement is NO) and (Optimization is NO) and (Skilled_Resources is NO) and (Standard_Architecture is YES) and (Agile_Process is NO) and (Modest_Execution is NO) and (Project_Management_Expertise is NO) and (Clear_Business_Objectives is NO) then (RiskProbability is RISKY) The fuzzified inputs that are obtained in step 1 are applied to the antecedent parts of rules in the rule base. As the fuzzy rule has multiple antecedents, we apply AND operator, with product (prod) method to produce single value that represents the evaluation of each rule antecedent parts .A fuzzy implication operator (minimum method) then is applied to clip the membership values of the rule consequent parts based on membership values of antecedents. The model output is categorized in two linguistic variables that are Risky and Non- Risky. Also, two linguistic variables are used for each input, namely: NO and YES. Step 3: Outputs aggregation In this step, the previously truncated membership functions of rule consequents are combined to obtain single fuzzy set. Step 4: Defuzzification Defuzzification is used to calculate the output as numerical value. Centroid method is applied to obtain the value that represents the software project risk probability. Fig. 4 shows the model’s fuzzy inference system (FIS) represented by using MATLAB FIS editor. It includes ten input variables, and one output named RiskProbability. Fig. 4. FIS Input and Output Variables. V. TOOL DESIGN The graphical user interface (GUI) shown in Fig. 5 is developed to enable software project decision makers to easily access our risk assessment tool. The user first specifies percentages of presence of the ten items (inputs) in his project, and then he presses “Estimate Risk Score” button to evaluate the project risk probability. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 2, 2019 157 | P a g e www.ijacsa.thesai.org Fig. 5. Software Risk Assessment. It is strongly recommended that the tool to be used earlier or during any phase of the project development process to determine the current state of the software project and to identify the possible improvements to mitigate risk and avoid the project failure. The goal of this tool is assigning one of the following labels to the project under evaluation:  Low risk (LOW): If the risk probability is less than 40%, this indicates that the project is healthy and expected to be successfully completed. However, there are no fully guaranteed successful projects, therefore project managers should be aware of individual items with low scores, and these items should be tracked and controlled during all stages of the project development process.  Medium risk (MED): If the risk probability in range of 40 to 60%, the project should be identified as a medium risk, and efforts must be made to avoid occurrence of the undesirable events. Improvement should be applied to those individual items with low scores that can mitigate the overall project probability of risk.  High Risk (HIGH): If the risk probability is greater than 60%, this indicates that the project has run into serious risks that can cause failure if process improvement methods are not applied. The stakeholder should be reported that there is imminent danger of project failure. All project phases have to be kept under monitoring, and a quality reports should be regularly carried out. If the risk is still high after applying mitigation methods, it could be better to decide not to proceeding with this project implementation. VI. EXPERIMENTAL WORK AND DISCUSSION To analyze the behavior and sensitivity of our risk assessment tool, we assumed that it is applied on eight virtual projects. Descriptions of these projects and results of their assessments by the risk tool are presented in Table 2. TABLE II. TOOL ASSESSMENT RESULTS FOR EIGHT SOFTWARE PROJECTS Success Factor Project ID Project A Project B Project C Project D Project E Project F Project G Project H Executive Sponsorship 80% 47% 88% 55% 40% 80% 80% 40% Emotional Maturity 75% 38% 90% 50% 35% 90% 40% 40% User Involvement 75% 60% 85% 53% 30% 80% 33% 30% Optimization 70% 50% 86% 40% 40% 77% 30% 30% Skilled Resources 85% 70% 70% 70% 40% 60% 71% 60% Standard Architecture 80% 66% 70% 55% 48% 63% 45% 63% Agile Process 60% 55% 50% 50% 44% 44% 40% 33% Modest Execution 85% 70% 50% 80% 70% 50% 66% 70% Project Management Expertise 87% 55% 60% 72% 70% 60% 46% 60% Clear Business Objectives 80% 50% 60% 75% 72% 40% 51% 61% Risk Probability 20.63% 48.62 20.05 50 60.92 29.3 56.8 62.44 Risk Classification LOW MEDIUM LOW MEDIUM HIGH LOW MEDIUM HIGH (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 2, 2019 158 | P a g e www.ijacsa.thesai.org Fig. 6. Effect of “Executive Sponsorship” on the Project Risk Probability. In all projects, we observed that some factors have higher impacts on the project risk than others. For example in project E, even it has some factors with high score (i.e. factors with percentage of presence higher than 60%), namely, Modest Execution Project Management Expertise, and Optimization. Similarly, the risk probability for project F is “LOW” even it has six factors with low scores (i.e. factors with percentage of presence lower than 60%). This is due to the high scores of the first four factors that have higher effect on the project success. Also, a sensitivity analyses can be performed for the individual factors. Fig. 6 presents sensitivity analysis for “Executive Sponsorship” factor to show its effect on the total project risk probability. The percentage of presence of the considered factor is changed within the range of 0 to 100% while all other input parameters are kept fixed. VII. CONCLUSION In this paper, a fuzzy based user-friendly tool to assess “risk probability” for the software projects is presented. This tool is developed based on software project success factors identified by Standish organization. These factors correspond to real data collected through survey involved about 50,000 projects. The developed tool supports a general assessment of project risk at any phase of development process. The percentages of presence of ten success factors are used as input to the system that produces a numerical value which presents the total project risk probability. The result can be used to assign one of three labels namely: low, medium, or high risk to the software project. The developed tool can be used to guide the decision makers in making critical decisions early to avoid undesired events that might cause project failure. The system behavior and sensitivity are analyzed using eight virtual projects and the impacts of various factors are observed. The presented tool has two limitations. First, the proposed approach did not consider correlations between factors. For example, the percentage of presence of “User Involvement” factor may be correlated with the percentage of presence of “Clear Business Objectives” factor. Second, we have not applied the tool to actual software projects. Even though our model is implemented based on actual empirical data to be a supportive tool that can be used for observing the current state of the project “during development process” (i.e. this tool is not designed to be applied on already released projects), it might be useful to involve software companies to verify the results of the proposed tool. Future research work will investigate how the system prediction accuracy can be improved by using learning algorithms based on historical data from previous projects. REFERENCES [1] Xu Z, Khoshgoftaar TM, Allen EB, Application of fuzzy expert systems in assessing operational risk of software, Info.Soft. tech., 45 (7) (2003) 373-388. [2] Ewusi-Mensah, K. (2003). Software Development Failures: Anatomy of Abandoned Projects. Cambridge: MIT Press. [3] GAO Report (14-705T). (2014). Preliminary Results of Undercover Testing of Enrollment Controls for Health Care Coverage and Consumer Subsidies Provided Under the Act. [4] The Standish Group. (2013). The Chaos Manifesto.The Standish Group. [5] Ibraigheeth, M., & Fadzli, S. A. (2019). Core Factors for Software Projects Success. JOIV: International Journal on Informatics Visualization, 3(1). [6] Hastie S, Wojewoda S. , Standish group 2015 chaos report-q&a with Jennifer Lynch, (2016). [7] Boehm BW, Software risk management: principles and practices, IEEE software 8(1)(1991), 32-41. [8] Barki H, Rivard S, Talbot J.,Toward an assessment of software development risk, J. manag. Info. sys, 10( 2) (1993) 03-25. [9] Schmidt R, Lyytinen K, Keil M, Cule P, Identifying software project risks: An international Delphi study”. J. manag. Info. Sys,17(4) (2001)5- 36. [10] Wallace L, Keil M, Rai A, How software project risk affects project performance: An investigation of the dimensions of risk and an exploratory model, Deci. sc., 35(2)(2004) 289-321. [11] Jiang JJ, Klein G. ,Risks to different aspects of system success, Info. &Manag., 36(5) (1999) 263-272. [12] García MN, Román IR, Peñalvo FJ, Bonilla MT, An association rule mining method for estimating the impact of project management policies on software quality, development time and effort. Exp. Sys. App., 34(1) (2008)522-529 [13] Xu Z, Yang B, Guo P, Software risk prediction based on the hybrid algorithm of genetic algorithm and decision tree.in Int. Conf. on Intelligent Computing, Berlin, 2007( Springer, Berlin, Heidelberg)pp. 266-274. [14] Xu Z, Khoshgoftaar TM, Allen EB, Application of fuzzy expert systems in assessing operational risk of software, Info.Soft.Tech., 45(7) (2003)373-88. [15] Yavari A, Golbaghi M, Momeni H, DAssessment of Effective Risk in Software Projects based on Wallace's Classification Using Fuzzy Logic, Int. J. Info. Eng. Electronic Bus., 5( 4) (2013) 58 [16] Neumann DE. , An enhanced neural network technique for software risk analysis, IEEE Trans. on Software Eng.,28 (9) (2002) 904-912. [17] Hu Y, Zhang X, Ngai EW, Cai R, Liu M. ,Software project risk analysis using Bayesian networks with causality constraints, Deci.Sup. Sys., 56 (2013) 439-49. 
work_rewmle3fpbbtxekaoibpbng3fe	----	Course Generation Based on HTN Planning Carsten Ullrich German Research Center for Artificial Intelligence, DFKI GmbH Stuhlsatzenhausweg 3, D-66123 Saarbrücken, Germany cullrich@dfki.de Abstract A course generator automatically assembles learning objects retrieved from one or several repositories to a greater unit. Different frame- works for course generation exists, but only a few allow the representation of sophisticated peda- gogical knowledge. In this paper, we describe course generation based on pedagogical tasks and methods, formalized in a hierarchical task network (HTN) planner. We describe the mod- erate constructivist learning strategy we imple- mented in this framework, problems that arouse from using HTN planning and their solutions. A first evaluation supports the practical value of our approach. 1 Introduction Course generation automatically assembles learning ob- jects retrieved from one or several repositories to a greater unit, a course. The learning objects are not assembled ran- domly, but support the user to reach a learning goal. Such goals require a sophisticated representation and us- age of pedagogical knowledge. Although course genera- tion has been a topic of research since long, only a few frameworks exist that offer a representation of pedagogical knowledge powerful enough to encode a variety of generic pedagogical approaches. In this paper, we will describe our approach to course generation, which is based on hierarchi- cal task network (HTN) planning. We will start by review- ing related work (Section 2). Then, Section 3 provides a short introduction to HTN planning. The main part of the paper describes and exemplifies the course generation as developed for LEACTIVEMATH. 2 Related Work Early approaches on course generation date back to the eighties [Peachy and McCalla, 1986; Murray, 1989]. With the Generic Tutoring Environment (GTE), Van Mar- cke [van Marcke, 1992; van Marcke, 1998] introduced the separation between instructional tasks (representations of pedagogical activities) and instructional methods (repre- sentations of different ways of achieving the activities). Most of this instructional knowledge is independent of the learning domain. Vassileva [Vassileva, 1995] build on this approach and added several layers of rules that handled dif- ferent aspects of course generation (e.g., selection of the main strategy or of the appropriate media types, etc.). Subsequent course generation frameworks (e.g., [Specht and Oppermann, 1998; Steinacker et al., 1999]) did not reach such a high level of representing pedagogical knowl- edge (or at least never provided sufficiently detailed de- scriptions). Today’s course generation focuses less on pedagogical knowledge, but more on Semantic Web and metadata, e.g., on using ontologies of the subject domain to automati- cally calculate the best path through the learning mate- rial [Karampiperis and Sampson, 2004], or on calculating a learner specific ranking of and trail through learning ob- jects retrieved according to his query [Keenoy et al., 2004]. These approaches use rather simplified pedagogical knowl- edge, e.g., to select those learning objects with the lowest typical learning time. However, to generate a course which is adapted to the individual learner’s goals and needs and which is based on state of the art pedagogical strategies re- quires more elaborate expertise. In a former project, we used an expert system (JESS, [Friedman-Hill, 1997]) to represent elaborated course gen- eration knowledge as rules [Ullrich, 2003]. However, this approach had severe scaling problems: in case more than 10 users used the system simultaneously, the latency time increased over an acceptable amount. While technical op- timizations might have alleviated this problem, using rules made it difficult to use the hierarchical structure inherent in pedagogical knowledge. Therefore, we decided to investi- gate a new approach based on HTN planning. 3 Hierarchical Task Network Planning In HTN planning a planning problem is represented by sets of tasks (task networks); methods decompose non- primitive tasks into sub-tasks until a level of primitive tasks is reached, which can be solved by operators [Erol et al., 1994]. Because HTN incorporates heuristic knowl- edge in the form of the decomposition methods, it is a very efficient planning technique. It also offers a rela- tively straight-forward way for representing human expert knowledge. The current version of our course generator uses the HTN planner JShop2 [Ilghami and Nau, 2003], a domain-independent planner that compiles domain de- scriptions into domain-specific planners. Figure 1 contains an example of a method (for the time being, we will focus on its HTN related aspects, the un- derlying pedagogical strategy will be explained in Sec- tion 4.1). The method is applicable in case an open task ex- ists that matches with teachConcept ?c, with ?c be- ing a variable (indicated by the prefix ?). Then, because the method has no preconditions (the empty parentheses in line two), the task will be decomposed in seven subtasks. The subtasks annotated with the prefix ! are primitive tasks, i.e., tasks that will not be further decomposed. (:method (teachConcept ?c) () ((!startSection NewBook ?c) (introduce ?c) (develop_concept ?c) (practice ?c) (connect ?c) (reflect ?c) (!endSection) ) Figure 1: An HTN method for teaching a concept 3.1 Challenges of HTN Planning HTN planning faces several difficulties if employed in a Web-based context. However, today most learning en- vironments are Web-based. As an example, the compo- nents of the learning environment ACTIVEMATH [Melis et al., 2001], e.g., the learner model, the learning object databases, and learning supporting tools such as a concept map [Melis et al., 2005], are distributed over the Web and available as Web services. Such a setting poses the follow- ing challenges: Distributed Content In a distributed environment, a course generator should be able to integrate content from several databases. This involves the difficulty that tradi- tional AI planning requires evaluating a method’s precondi- tions against the planner’s world state. However, this would require mirroring all the content stored in the repositories (or its metadata) in the world state. This is simply infeasi- ble. Dynamic User Information Similar problems arise with respect to user information. Similar to most systems AC- TIVEMATH uses information about the learner stored in a learner model. Part of this information, e.g., the knowl- edge state, is frequently updated. Thus, the naive solution of adding the learner information in the world state of the planner is unpractical, especially as the learner model can cover a vast range of content and it is unknown exactly which of the information will be relevant. Learning Services A vast range of services that support the learning process in various ways have been developed, for instance tools that stimulate meta cognition [White and Shimoda, 1999] or perform assessment [Conejo et al., 2004]. A course should integrate these services, not ar- bitrarily but in a pedagogically sensible way: during the learning process, at specific times the usage of a tool will be more beneficial than at some other time. For instance, reflecting over the learned content may be most effective at the end of a lesson. Additionally, access to these services may vary, depending on the configuration of the learning environment and their general availability. Therefore, the planning needs to take into account dynamic information about these services. In the following, we will describe how we represent ped- agogical knowledge in an HTN planning framework and how we overcome the above limitations. 4 Representing Course Generation Knowledge in an HTN planner 4.1 Pedagogical Background In LEACTIVEMATH, the University of Augsburg and the DFKI are jointly developing moderate constructivist ped- agogical strategies that rely on the “Programme for Inter- national Student Assessment” (PISA) framework, [OECD, 1999]. Instead of exclusively assessing curriculum based knowledge, PISA focuses on mathematical literacy and competencies. Mathematical competencies are general mathematical skills such as problem solving, the use of mathematical language and mathematical modeling. In LEACTIVEMATH, authors annotate exercises and exam- ples with the competencies they train (in addition to a sub- set of standard LOM metadata [IEEE Learning Technology Standards Committee, 2002]). The pedagogical strategies take these competencies into account: first, they ensure that every course contains a diversity of competencies. Sec- ondly, they primarily select exercises of competency levels just outside the learner’s current mastery state. We identified six different pedagogical scenarios, each corresponding to specific needs and high-level learning goals of the user: LearnNew, Rehearse, Overview, train- Competency, Workbook, and ExamSimulation. For in- stance, the scenario LearnNew provides the student with an in-depth course that teaches him all necessary material to fully understand the target concepts right from the start. For each scenario, we formalized the pedagogical knowl- edge required to generate courses supporting the learner to achieve the respective learning goals. 4.2 Course Generation by HTN Planning Course generation knowledge lends itself to being repre- sented hierarchically. Additionally, as van Marcke [van Marcke, 1992] showed, by representing the pedagogical objectives as tasks and the ways of achieving the objects as methods, sophisticated pedagogical strategies can be en- coded. In HTN planning, tasks are a basic principle, too. Therefore, in a first step, we will bring together the notion of pedagogical and HTN task. A pedagogical task corresponds to the goals a learner wants to achieve. It combines the two dimensions domain (content) and educational goal. Formally, a task t is de- fined as a tuple t = (l, c), where l is a pedagogical objec- tive and c a unique identifier of a learning object (content goal). While c specifies the concept the course will primar- ily target, the pedagogical objective determines the kinds of learning objects selected for c. In JShop, a task is defined as a predicate symbol fol- lowed by several arguments. Hence, a pedagogical task can be mapped to a HTN task by mapping the pedagogical ob- jective onto the predicate symbol, and the content goal onto the arguments. Because of this straightforward mapping, in the following we will no longer distinguish between peda- gogical and HTN tasks. A top-level task that serves as a starting point of course generation is teachConcept id. The goal of this task is to assemble a structured sequence of learning objects that help the learner to understand the content goal c. Us- ing different collections of tasks and methods (i.e., tutorial strategies), this task can be planed differently. Hence, the task-based approach can serve to represent a variety of ped- agogical strategies. Figure 1 illustrates a method developed in LEACTIVE- MATH based on pedagogical principles. It describes how the task of teaching a concept is decomposed in the sce- nario LearnNew. This decomposition follows a course of action in a classroom which implements constructivist ele- ments and distinguishes between different stages that occur while learning a new concept. Similar solutions were de- veloped for other high-level learning goals. !startSection and !endSection are primitive tasks, which serve to provide additional structure. They open and close sections within a course. The introduction (introduce ?c) makes the aims of the learning process explicit. It provides an overview to the learner and makes the learning process more transparent. The stage of devel- oping a concept serves to provide the essential information about the concept to the learner, and will be explained in more detail in the next paragraph. The fourth subtask of teaching a concept trains the application of the concept. The final stages, connect and reflect, serve metacognitive purposes. They support the learner to put the concept in its context with respect to the other learning material, and help him to get an overview of his knowledge state. As a detailed example, consider the two methods for de- veloping a concept shown in Figure 2. The upper method uses information about the user that is retrieved from a learner model and is applicable if the learner exhibits a high math anxiety. In that case, first the concept is inserted in the course via the primitive task !insertElement ?c. Then, the insertion of an explanation and an example illus- trating the application of the mathematical concept aim at alleviating the anxiety. The bottom method is applied in case the first method does not match, and just inserts the concept itself. (:method (develop_concept ?c) ((learnerProperty anxiety ?an) (>= ?an 3)) ((!insertElement ?c) (explain ?c) (examplify ?c))) (:method (develop_concept ?c) () ((!insertElement ?c))) Figure 2: Two methods for developing a concept 4.3 Critical and Optional Tasks Course generation has to distinguish between critical and optional tasks. Critical tasks represent necessary elements of the pedagogical strategy that have to be fulfilled for the course generation not to fail. For instance, in a problem- based approach it is mandatory to present a learning ob- ject is of the type realWorldProblem, hence the cor- responding task insertProblem! is marked as criti- cal (indicated by the suffix !). In contrast, optional tasks should be achieved if possible, but failing to do so will still result in a course. As an example, the subtask explain of the upper method in Figure 2 is optional, e.g., if there are no learning object that can serve to explain concept c the concept will still be considered as developed. 4.4 Ideal and Fallback Methods In practice, it is impossible to encode course generation knowledge that covers every potential combination of in- formation about the learner. Oversimplified, if a learner (:method (selectAppropriateExercise ?c) ;; ideal method: if motivation is ;; high, then select exercise with ;; higher competency level. (;; preconditions: (learnerProperty field ?field) (learnerProperty educationalLevel ?el) (learnerProperty motivation ?m) (>= ?m 3) (learnerProperty competencyLevel ?c ?cl) (try-all-assign ?exercise (call GetElements ((class exercise) (relation for ?c) (property learningcontext ?el) (property competencylevel (+ 1 ?cl)) (property field ?field))))) (;; sub-task: (insertElement ?exercise))) (:method (selectAppropriateExercise ?c) ;; alternative method: ;; select exercise from any field ( (learnerProperty allowedEducLevel ?el) (learnerProperty competencyLevel ?c ?cl) (try-all-assign ?exercise (call GetElements ((class exercise) (relation for ?c) (property learningcontext ?el) (property competencylevel ?cl) )))) ((insertElement ?exercise))) Figure 3: Selecting an exercise model represents n different kinds of information with m possible values each, then the number of possible combina- tions is nm. Writing rules or methods for each case would keep a lot of teachers busy for a long time. Therefore, in LEACTIVEMATH, methods encode best possible and alternative ways of selecting learning objects. An ideal method typically takes a larger number of user properties into account and poses a larger number of con- straints on the learning object. As an example, consider the methods in Figure 3. They encode the pedagogical knowl- edge about exercise selection, and are called as a sub-task of practice. The upper method checks whether the user is highly motivated. Then, it searches for an exercise that would be slightly too difficult under normal circumstances (competency level plus 1). Additionally, the learning con- text of exercise should correspond to the educational level of the learner, and, similar, its field should correspond to the field of the learner. In case no ideal method is applicable, fallback methods come into play. They encode the least constraining con- ditions on the learning objects possible and serve to insert elements in case no ideal method can be fulfilled. The bot- tom method of Figure 3 matches if any exercise for concept ?c exists that is of the correct competency level and has a learning context equal or lower as the educational level of the learner. 4.5 The Output of Course Generation The result of the planning is a sequence of learning objects called course structure. Similar to the organization element of an IMS Content Package [IMS Global Learning Consortium, 2003], it consists of nested sections with the leaves being pointers to learning objects. Because tasks represent a vast range of pedagogical goals, the size of the generated courses ranges from a sin- gle element to a complete curriculum. For instance, while the task teachConcept results in sequences of several learning objects, other methods may select only a single element. Frequently occurring examples are the methods for exercise selection shown in Figure 3 and for example selection, shown in Figure 4. The example selection is additionally interesting because of its differences to exercise selection. Examples serve to increase motivation by presenting the application of a con- cept. Therefore, the primary requirement is that the exem- plary application is taken from the field of interest of the learner, even in case its assigned competence level is higher than the current competence level of the learner. (:method (selectAppropriateExample ?c) ;; ideal method: if motivation is ;; low, select example with ;; matching field. ( (learnerProperty field ?field) (learnerProperty educationalLevel ?el) (learnerProperty motivation ?m) (call < ?m 2) (learnerProperty competencyLevel ?c ?cl) (equivalent ?cl ?ex_cl) (try-all-assign ?example (call GetElements ((class example) (relation for ?c) (property learningcontext ?el) (property competencylevel ?ex_cl) (property field ?field) )))) ((insertElement ?example))) (:method (selectAppropriateExample ?c) ;; fallback method: select example ;; from any competency level. ( (learnerProperty allowedEducLevel ?el) (learnerProperty field ?field) (try-all-assign ?example (call GetElements ((class example) (relation for ?c) (property learningcontext ?el) (property field ?field) )))) ((insertElement ?example))) Figure 4: Selecting an example 4.6 Solutions to the Challenges of HTN Planning In Section 3.1 we described the problem that traditional AI planning requires evaluating a method’s preconditions against the planner’s world state, which is practically im- possible in case the learning objects are stored in one or several databases. Therefore, we extended the HTN plan- ner so that queries about learning objects in a method’s pre- conditions result in a call to a mediator. Mediators are well known services used in information integration (for a recent overview on this topic, see [Doan et al., 2004]). They act as links between the knowledge processing and knowledge storing components, and offer a uniform query interface to a multitude of autonomous data sources. The LEACTIVEMATH mediator passes incoming queries to the available databases and combines the results. JShop2’s call command allows integrating exter- nal function calls during planning. The function call GetElements conditions (e.g., in Figure 4) queries the mediator to search the repositories for elements that fulfill the given conditions. try-all-assign ?p ?identifiers generates all bindings of ?p to the el- ements of the returned identifier list. Thereby subsequently all possible values can be tried when backtracking. Similarly, preconditions involving information about the learner result in queries to the learner model. These queries are encapsulated in a planning axiom as shown in Figure 5. Axioms are used to infer preconditions not directly asserted in the world state. In the figure, the result of the query is bound to the variable value using the additional ax- iom same. The axiom same returns true in case both provided parameters are instantiated and equal, and false otherwise. If one parameter is uninstantiated, the axiom assigns the value of the instantiated parameter to the unin- stantiated parameter. (:- (learnerProperty ?property ?value) (same ?value (call queryLM ?property))) Figure 5: Querying the learner model The integration of learning services happens by service- adding methods. These methods encode the pedagogical knowledge at what time in the course the learner should use a tool and insert calls to the tool at the appropriate place in the course structure. Later, when the learner navigates through the course, an invitation to use the service is pre- sented to her. Checking for the availability of a service is done by introducing an external function, too. Thereby a method’s preconditions can test for the availability of a ser- vice. In case the service is available, its call is inserted in the course. Otherwise, an alternative method is applied. This way, the pedagogical knowledge remains reusable, re- gardless of the actual configuration of the learning environ- ment. (:method (assess) ;; if Siette is available, ;; then use it for assessment ((call ServiceAvailable Siette)) ((!insertServiceCall Siette))) (:method (assess) ;; fallback: insert some exercises () ((insertText assessment) (insertExercises 5))) Figure 6: Integration of an assessment service Figure 6 illustrates the integration of the assessment tool Siette [Conejo et al., 2004]. The upper, ideal method checks whether Siette is available and if so inserts a call to the service in the course structure. In the fallback case, the assessment will be simulated manually. A short text is added in the course that describes that the following ex- ercises will assess his knowledge, and five exercises are inserted. 5 First Results and Conclusions The results reported in this section focus on technical as- pects. Evaluations regarding learning gains are underway and will be reported at a later time. We performed a first comparison between the old, expert system based (EXSCG) and the new course generator (HT- NCG). The tests and all necessary components (ACTIVE- MATH server and a database) ran on a 2.8GH Intel Pentium 4 CPU with 2GB RAM. Both course generators repeatedly generated an average length course about the mathematical concept derivation. EXSCG showed a slightly slower performance for sin- gle user course generation (average time of 23 and 20 sec- onds). However, the EXSCG course generation did not ter- minate in case of the simultaneous generation of more than 20 users. HTNCG slowed down, but still terminated (aver- age of 2 minutes). We suspect the differences to be caused by the on- demand retrieval of the metadata. In EXSCG, the meta- data of the complete content was loaded in the fact-based of the expert system and then the rules were evaluated. In HTNCG, only content relevant for the current plan was re- trieved if necessary, i.e., when queried. It is interesting that the new approach is faster even though an additional indirect step via the mediator was in- serted. In the future, we plan to extend the mediator with caching possibilities. First of all, it can optimize queries and secondly avoid queries over the Web. An additional evaluation compared an XML-RPC based connection and a direct connection between the mediator and one of our database (based on Lucene [Foundation, 2003]). Because of the overhead of processing XML-RPC, we expected the direct connection to outperform XML- RPC. Surprisingly, the direct connection proved not to be much faster. A possible cause may be the limited number of queries (about 270). We plan to further investigate on this topic. However, a significant speed-up was achieved by using a cached version of Lucene database. There, in case the very same query is posed more than once, the result elements were not retrieved from index but directly from memory. Course generation for a single user decreased to the average time of three seconds, which is about factor 7 faster than compared to the non-cached version. The speedup holds in case of 30 user planning, which took about 20 seconds. This speedup is even more surprising in regard of the fact that due the early stage of implementation the cache was local to a single user, which means instead of one “global” cache, the system used up to 30 “local” caches. We expect the speedup to be caused mainly be the large amount of similar queries, which occur even in a single course genera- tion. For instance, if several examples for the same concept are inserted in a course, a large number of the same queries are processed. In this paper, we described how HTN planning can be used for representing and executing course generation knowledge that is based on tasks and methods. Besides the described advantages, we plan to use tasks to describe the service offered by a course generator Web service. Addi- tionally, they may be used for communication between user and system, for instance by providing the user the possibil- ity of executing tasks on demand with respect to a given course and thereby intelligently extending the course. In- vestigating these possibilities will be our next work. Acknowledgments The work described in this publication is implemented in the learning environment ACTIVEMATH 1. ACTIVEMATH serves as the technical basis used in the LEACTIVEMATH project 2, where a consortium of eight European partners is developing an innovative third generation eLearning sys- tem for high school, college, and university level class- rooms. The system adapts to the learner and learning con- text and comprises personalization, tutorial dialogues, open student modeling and interactivity that is tool-supported for active and exploratory learning. The LeActiveMath project is funded under the 6th Framework Programm of the European Community - (Con- tract Nr IST-2003-507826). The author is solely respon- sible for its content, it does not represent the opinion of the European Community and the Community is not re- sponsible for any use that might be made of data appearing therein. References [Conejo et al., 2004] R. Conejo, E. Guzman, E. Millan, M. Trella, J. L. Perez de-la Cruz, and A. Rios. Siette: A web-based tool for adaptive teaching. International Journal of Artificial Intelligence in Education (IJAIED 2004), 14:29–61, 2004. [Doan et al., 2004] AnHai Doan, Natalya F. Noy, and Alon Y. Halevy. Introduction to the special issue on se- mantic integration. SIGMOD Rec., 33(4):11–13, 2004. [Erol et al., 1994] K. Erol, J. Hendler, and D. S. Nau. HTN planning: Complexity and expressivity. In Proceedings of the Twelfth National Conference on Artificial Intelli- gence (AAAI-94), volume 2, pages 1123–1128, Seattle, Washington, USA, 1994. AAAI Press/MIT Press. [Foundation, 2003] The Apache Software Foundation. Jakarta lucene, July 2003. [Friedman-Hill, 1997] E. Friedman-Hill. Jess, the java ex- pert system shell. Technical Report SAND98-8206, Sandia National Laboratories, 1997. [IEEE Learning Technology Standards Committee, 2002] IEEE Learning Technology Standards Committee. 1484.12.1-2002 IEEE standard for Learning Object Metadata, 2002. [Ilghami and Nau, 2003] Okhtay Ilghami and Dana S. Nau. A general approach to synthesize problem-specific planners. Technical Report CS-TR-4597, Department of Computer Science, University of Maryland, October 2003. [IMS Global Learning Consortium, 2003] IMS Global Learning Consortium. IMS content packaging informa- tion model, June 2003. [Karampiperis and Sampson, 2004] P. Karampiperis and D. Sampson. Adaptive instructional planning using on- tologies. In Proc. of the 4th IEEE International Confer- ence on Advanced Learning Technologies, ICALT 2004, pages 126–130, 2004. [Keenoy et al., 2004] K. Keenoy, A. Poulovassilis, V. Christophides, P. Rigaux, G. Papamarkos, A. Magka- naraki, M. Stratakis, N. Spyratos, and P. Wood. Personalisation services for self e-learning networks. In 1http://www.activemath.org 2http://www.leactivemath.org http://www.activemath.org http://www.leactivemath.org N. Koch, P. Fraternali, and M. Wirsing, editors, Proc. of 4th International Conference on Web Engineering, volume 3140 of Lecture Notes in Computer Science, pages 215–219. Springer, 2004. [Melis et al., 2001] E. Melis, E. Andrès, J. Büdenbender, A. Frischauf, G. Goguadze, P. Libbrecht, M. Pollet, and C. Ullrich. ACTIVEMATH: A generic and adaptive web-based learning environment. International Journal of Artificial Intelligence in Education, 12(4):385–407, 2001. [Melis et al., 2005] E. Melis, P. Kärger, and M. Homik. In- teractive concept mapping in ACTIVEMATH. In Delfi 2005, 2005. [Murray, 1989] W. R. Murray. Control for intelligent tutor- ing systems: A blackboard-based dynamic instructional planner. In D. Bierman, J. Breuker, and J. Sandberg, editors, Proc. 4th International Conference of AI and Education, pages 150–168, Amsterdam, Springfield VA, Tokyo, 1989. IOS. [OECD, 1999] OECD, editor. Organisation for Economic Co-operation and Development: Measuring Student Knowledge and Skills – A New Framework for Assess- ment. 1999. [Peachy and McCalla, 1986] D. R. Peachy and G. I. Mc- Calla. Using planning techniques in intelligent tutoring systems. International Journal of Man-Machine Stud- ies, 24:77–98, 1986. [Specht and Oppermann, 1998] M. Specht and R. Opper- mann. ACE - adaptive courseware environment. The New Review of Hypermedia and Multimedia, 4:141– 162, 1998. [Steinacker et al., 1999] A. Steinacker, C. Seeberg, K. Re- ichenberger, S. Fischer, and R. Steinmetz. Dynamically generated tables of contents as guided tours in adaptive hypermedia systems. In Proceedings of the EdMedia & EdTelecom, pages 167–175, June 1999. [Ullrich, 2003] C. Ullrich. Pedagogical rules in ACTIVE- MATH and their pedagogical foundations. Seki Report SR-03-03, Universität des Saarlandes, FB Informatik, 2003. [van Marcke, 1992] K. van Marcke. Instructional exper- tise. In C. Frasson, G. Gauthier, and G.I. McCalla, editors, Proceedings of the Second International Con- ference on Intelligent Tutoring Systems, number 608 in Lecture Notes in Computer Science, pages 234–243. Springer, 1992. [van Marcke, 1998] K. van Marcke. GTE: An epistemo- logical approach to instructional modelling. Instruc- tional Science, 26:147–191, 1998. [Vassileva, 1995] J. Vassileva. Dynamic courseware gen- eration: at the cross point of CAL, ITS and authoring. In Proceedings of ICCE’95, Singapore, pages 290–297, December 1995. [White and Shimoda, 1999] B.Y. White and T.A. Shi- moda. Enabling students to construct theories of collab- orative inquiry and reflective learning: Computer sup- port for metacognitive development. International Jour- nal of Artificial Intelligence in Education, 10:151–182, 1999. Introduction Related Work Hierarchical Task Network Planning Challenges of HTN Planning Representing Course Generation Knowledge in an HTN planner Pedagogical Background Course Generation by HTN Planning Critical and Optional Tasks Ideal and Fallback Methods The Output of Course Generation Solutions to the Challenges of HTN Planning First Results and Conclusions 
work_rfdvns3gd5djdmkbwoodu767ku	----	&&& t McMASTER UNIVERSITY 1280 Main Street West Hamilton, Ontario, Canada LBS 4M4 Telephone (416) 525-9140 k& iffi FACULTY OF BUSINESS RESEARCH AND WORKING PAPER SERIES "� ------��==��-��-- '"-.,,�. ""� ... l:.:;"'*\;;,.-_ &W* W 0 - WW ;; M&WWW& M ·�1 \ EXPIM: A Knowledge-Based Expert System for Prcduction I Inventory Mo:lelling by Mahmut Parlar Faculty of Business MctA.aster University Hamilton, Ontario, L8S 4M4 · Canada Working Paper #283 August 1987 I" EXPIM: A Knowledge-B ased Expert System for Production / Inventory Mode l l ing M ahmut Parlar Faculty of Bu s ines s McMaster Univer s ity Hami lton, Ontar io, L8S 4M4 Canad a August 1 987 ( EXPERT SYSTEMS ; MODELLING; PRODUCTION/INVENTORY) I . Abstract This paper discu s s e s the development of a knowledge based expert system (EXPIM ) which can identify and recommend up to 30 p roduction - inventory model s . The user (m�nager) i s asked a s equence o f questions regarding the problem situation. The system uses backward-chaining to identify what i s the correct model t o u s e given the situation des cribed b y the user. The paper a l s o discusses the general methodology for building expert systems in m an agement science for th� purpose of identifying mode l s in other areas , such·as queueing theory and location theory. 1 1. Introduction Management s cientists and industrial engineers have, in the p ast forty years, produced hundreds of production-inventory models ranging from the simple extensions of clas sical EOQ ( Hadley and Whitin, 196 3 ) to c omplex multi-level control systems (S�hwarz, 1981). Two decades ago, Eilon and Lampkin ( 1968) summarized more than 500 p apers in inventory control developed and/or published between 1953 and 1965 . Thi s proliferation o f production-inventory (PI) models have prompted many researcher s to publish surveys of "current" status of thes e models in p articular and inventory theory in general at a rate of approximately two per per dec ade (Veinott 1966; Iglehart 1967; Aggarwal 1974; Nahmias 1978; Wagner 1980; Silver 1981. ) With every new review p aper, the readers were informed ab out recent developments in inventory modelling and p o s sible extensions of the current models along with a "wi sh list" of future research activities . Clearly, all these review articles written by the experts in their field h ave helped the research c ommunity get an excellent overview of inventory m odelling and theory, although they prob ably have not helped the actual users of inventory model s very much In recent years, s ome researchers have started inve stigating the p o s s ibility of clas sifying the inventory systems. In p articular, Hollier and Vrat ( 1978) have used a structure similar to that proposed by Kendall ( 1952) f o r queueing models. · They have attempted to des cribe any given inventory system using four aspects , i.e. stru cture , environmental p arameters, operating policies and c o sts . For example an inventory system with Erlang demand , exponential replenishment time, general shelf-life distribution and complete b acklogging was symbolized as E1/M/G/�. Given the 2 ,. apparent duality between queueing and inventory processes ( Prabhu 1 965) this classification seemed to be a natural one t o adopt for inventory systems . But, it appears that this method has not caught on and no other published papers �eem to refer to it. Recently Menipaz ( 1 982) has prop osed another class ification of inventory models based on a systematic use of twelve basic assumptions. W ith all the assumptions holding , one obtains the classical EOQ , but as the assumptions are relaxed a better approximation to the problem is obtained. In ·� maj or research effort started in 1976 and supported by the Hungarian Academy of Sciences ( Barancsi et al. 1 980; Chikan et al. 1 982 ; and Barancs i et al . 1983) another a ttempt is made to classify inventory models. These researchers' obj ect ive was to make the many of the currently ava ilable inventory models more useful for the practi t ioners. By compar ing a c ode system with 4 5 have classified codes each having between two to ten p ossible values they all these models with the purpose of providing a manager with a m o del that would be most suitable for his problem. G iven the fact that maj ority of inventory ( and management science/operations research) models are never used in implemantat ion of the ir results , the correct choice of these models becomes very important . A manager may have several inventory models available to him, but if he is not sure which is the r ight one for the given situat i on obviously he will not be able to solve his pr oblem with the wrong model. The use of classical EOQ in many inventory situati ons although it is the wrong model is an often repeated example. The importance of selecting the r ight ( inventory) model and an anecdote related to that issue is given in Ravindran et al. ( 1 987 , pp . 3 6 9 -370) . 3 It i s a n expert then he clear that when a manager is faced with a PI problem and he has available for choo s ing the r ight model ( and perhaps solving it) can be conf ident that the efforts put into analy s i s will not be wasted. But , as in m any d i s c ipline , experts in inventory modelling are also s c arce and expens ive . In recent years , the emergence of the field of expert systems hope for (ES) as a branch of artifi c i al intell igence ( AI) has provided some increas ing the ava ilab ility of expert knowledge , albeit in a computerized �orm . The premise of thi s p aper is the following. Although the attempts to create systems for class ify ing inventory models as d i s cussed in the p aragraphs ab ove m ay have provided useful information for dec i s ion makers who want to use them , these systems do not have expertise. They all seem to l ack an ab il ity to provide explanations of the ir l ine of reasoning , i . e. how d id they decide to recommend the use of , say , the newsboy model . Except possibly for the Hungari an system, they lack flexib ility , i . e . integration of new knowledge incrementally into its existing store of knowledge. We attempt to go one step farther , and build a knowledge b ased expert computer system which can accurately choo se a PI model after consult ing the user ( m anager) by asking h im a m inimum number of quest ions pert inent to the problem s ituat ion. Our system, named EXPIM (for EXpert Product ion/Inventory Modeller) has the ab ility to explain WHY it is asking a particular question and also HOW it reached a conclus ion and recommended a particular model . By incorporating uncertainties of both the human expert in h i s reason ing, and the manager in his responses , EXPIM can also recommend models with reduced confidence factors and can warn the manager of this fact . It also has the ab ility to automat i c ally run a program once a model choice is made and solve the problem numeri cally. It permits sens itivity analy s i s by letting the 4 i' manager change one or more of his responses to system's queries and come with (pos sibly) different model recoiilrnendations . So far, EXPIM can identify approximately 30 models ranging from the simplest EOQ to more complex coordinated replenishment models. EXPIM runs on the widely available IBM-PC microcomputer (or compatibles) with 512K memory. An expert system , s uch as EXPIM , is a computer program whose behaviour duplicates , in s ome sense, the abilities of a human expert in his area of expertise. Expert systems have been succes sfully developed in field3 such a s medical diagnosis (Shortliffe 1976) , mineral exploration (Duda et al. 1979) and chemical data interpretation (Lindsay et al. 1980) . Excellent reviews of expert systems and their principles have been provided elsewhere (Buchanan and Duda 1983; Stefik et al. 1982; Hayes-Roth et al. 1983; Waterman 1986) and the reader is referred to these sources for detailed description s. In recent years, we have also witnes sed an interest in the MS/OR community towards AI/ES and this is exemplified in the large number of papers presented on these topics in TIMS/ORSA Joint National Meeting s. A few papers have also been published reviewing AI/ES concepts for the management science community and providing pos sible research topic s for interfacing expert systems and management s cience (Kastner and Hong 1984; A s sad and Golden 1986). Other applications of ES in traditional MS/OR are reported in the areas of design optimization (Azarm and Pecht 1985), univer sity admi s sion decisions (vanBreda 1986), simulation modelling (Khoshnevi s and Chan 1986 ; Douk.idis and Paul 1985) and operations analysts (Biswas et al. 1987) . In A s sad and Golden (1986 ) the authors present some possible directions for using ES in management s cience. These includ� i) choice of models which _\ l 5 is what EXPIM does and ii) model generation where systems queries prompt the user to add c onstraints and formulate the objective function in an optimization p roblem. Binbasioglu and Jarke ( 1986) provide a discussion of such a model generation in linear programming using an AI language called Pro log. In an important arti c le , Hahn ( 1985 ) discusses the interface between statistical methods and expert systems. As the statistical computer programs are becoming increasingly available to people who are not statistical experts, he proposes that expert statistical systems be constructed which would embed expert guidance in statistical programs. Some programs that run on IBM-PCs which �o just that are already available for expert forecasting (Wisard 1986). To summarize, we see that problems which could previously be solved by human experts only, are now being attacked by using the expertise embedded in expert computer systems. Provided that these·systems are constructed properly and are thoroughly tested and validated, they will be useful in many areas of man�gement science as the above examples illustrated. As the development of these systems frequently require the use of a specialized AI language or an expert system shell, in the next section we will describe the shell we used in our resear ch. In Section 3 , the methodology we used in the development of EXPIM will be discussed , where we will try to explain the steps that should be followed in developing systems such as EXPIM . Since there are many other areas in management s cience (location theory, queueing theo ry) which may benefit from expert system technology , our discussion may be useful for the developers of such systems. 6 S ection 4 contains a discussion o f the experience gained whi l e structuring and bu ilding this expert system. Last Se ction provides a summary , suggestions for refinements of the mod e l and some possibl e avenues for further research in this new area. 2. The Expert System Shell Expert systems c an b e deve loped in almost any computer language , including FORTRA.i.� (Weiss and Kulikowski 1984) and any of the popular AI l anguages LISP and Prolog . When a computer language is used the developer (known in the AI c ircles as the "knowledge engineer " ) has more control over the flow o f log i c , search mechani sm etc. , but the development time may be very long. On the other hand , using a r e adily avai lable expert system shell r�duc e s the deve lopment time , but the knowledge eng ineer l im its hims e l f to work within the previously determ ined confines of the shell someone e l s e has written. Sinc e our ob j e c t iv e in th i s res e arch was to produc e an expert system which should be easily availab l e , we chose Texas Instruments Personal Consultant E asy as our deve lopment env ironment which runs on IBM-PC. After exam ining a few other shells we d e c ided to use PC Easy b e c ause of its IBM compat i b i lity, fac i lity of deve lopment , sensitiv ity analys i s feature , explanat ion facility in Engl ish language and in corporation of c erta inty factors . PC Easy is a production rule based she l l where the knowledge base cons ists o f IF- THEN type ru les which are chained to each other. S inc e the s e production systems were first proposed by Post (1943) as a general computational mechanism they have been applied to wide variety of problems 7 ( D avis and King 197 7) including the expert systems. A rule in PC E asy specifies a deduction that can be made in a p arti cu l ar situation, and the premise ( I F ) and action (THEN) parts can be composed of compound statements connected with and and or. For example; RULEOl l I f 1) number o f products consi�ered is multiple, and 2) supplier is the same for a l l the items purchased, and 3) mode o f transportation is same Then It is definite that c oordination of replenishments is possible. is a rule which tests whether c o o rdination is possible in an inventory situation. PG E asy uses a combination of backward-chaining and forward-chaining contro l strategies. The former, also known as the g o al-directed control ( B arr and Feigenbaum 1981) reduces the number o f questions the user h as to answer and eliminates quickly many of the irrelevant choices. Forward chaining , or d at a-driven c ontrol can be used to trigger acti ons based on spe�ial conditions. strategies. ) (Appendix discusses an example of these two control When the system reaches a conc lusion regarding the. mode l to choose , the user can ask HOW this conclusion was reached. The explanation facility provides the response by referring to the rule it has used to reach the conc lusion, and prints the English trans l ation o f that rule. Sensitivity analysis is also easily done using the REVIEW feature , where the user is given the option to change one or more o f his responses 8 which h e provided during the consultation . This m ay be an important factor in problems where conditions may change from time to time. PC E asy supports certainty factors ( C F ) which are numerical values indic ating a m e asure o f confidence in the value of a parameter. These factors let a knowledge base deal with the r e ality that facts and opinions are sometimes known with l ess than abso lute c ertainty . Certainty factors c an be incorporated into PC Easy during deve lopment and during consultation when th-e user m ay feel a doubt in responding to a prompt. With this bri e f introdu ction to PC Easy ' s features , we now describe in some detail the m ethodology used in develop ing EXPIM . 1, nevelopm�nt of EXPIM An expert system such as EXPIM consists of three e l em ents: the knowledge base, the inference engine and the user interface (Feigenbaum and McCorduck 1983). The knowledge base contains the knowledge extracted from the expert .by the "knowledge engineer " who is usually a computer s cientist trained in artificial intelligenc e or someone who is we l l informed about the expert systems techno logy and the particular tool he is using. Inference engine part o f the system decides which portions of the knowledge base apply to a given problem and what additional data must be supp lied by the user . The inference engine also provid es explanations of the system's behaviour, such as r esponding to the WHY and HOW questions asked by the user. Final ly , the user interface consists of tools that c an translate the questions asked or conclusions r e ached into plain English. Currently, there are several methods available for constructing expert systems, e. g. rule-based dedu ction, statistical patterns c l assifi c ation , 9 10 etc. (Assad and Golden 1986). As discussed in Reggia and Ahuj a ( 1986), rule-based underlying type o f knowl edge , M any books deductive expert systems should be employed when i ) the knowledge is already organized as (I F-THEN ) rules and ii ) the classification i.e . inventory on inventory is predominantly categoric al. The domain of modelling does s atisfy these two r e quirements: managem ent ( Silver and Peterson 1985) give long lists o f assumptions leading to the p articular model und er discussion. For example , the assumptions le ading to the basic EOQ model are listed as i ) the d em and rate is constant and deterministic, ii) the unit v ariabl e cost does not dep end on repl enishment quantity, etc . which can e asily be put in an IF­ THEN rul e . Also, the cl assi fication of the inventory models is c ategorical in the s ense that every inventory model c an be placed in a particular model group a fter the user responding to a series of questions with two or at most thr e e alternative answers. For these re asons, we decided to use a rule- based system and adopted PC Easy as our expert system shell The rules in PC E asy are constructed using parameters which are structures that identify or contain a piece of information that the system uses to arrive at a conclusion . For exampl e, the parameter QUANTITY- DISCOUNT contains information about the applic ability of quantity discounts . W e have identified twenty three parameters with which all the rules in the knowl edge base are constructed . These parameters , in their abbreviated form are : COORDINATION, DEMAND-LEVEL, DEMAND-PROCES S, DI SCOUN T - TYPE, EOQ­ MODEL , FIXED -COS T , GENERAL - TYPE , INFLATION , LEAD-TIME , L I FE, MODEL , MON I TORING , NUMBER-OR-PRODUCT S , QUANTITY-DISCOUNT, REFERENCE - LI S T-ARTICLES , REFERENCE - LI S T-BOOKS, REPLEN I SHMENT , SHORTAGES, SPACE, SUBSTITUTE , SUPPLIERS , TRANSPORT and UNIT-VARIABLE-COST. For example , the parameter DEMAND-LEVEL has the following properties as used by PC Easy. DEMAND-LEVEL (PARAMETER) TRANSLATION : the level of the future demand PROMPT: What is the level of the future forecasted demand? HELP: :tab 2 " I f the forecasted future level o f demand is b asic ally constant (uniform over time) , choose CONSTA..�T-OVER-TIME. " : line 1 : tab 2 "If the forecast indicates lumpy demand where demand varies from time to time, choose VARIABLE - OVER-TIME. " : line 1 : tab 2 " Refer to (SIL-PET] , pp . 237-239, Section 6.6.4 . for an expl anation of a numer i c al technique which can distinguish between these two cases. " : line 2 COMMENTS : ASKED TYPE : SINGLEVALUED EXPECT: CONSTANT-OVER-TIME VARIABLE-OVER-TIME USED - BY: RULE 0 03 When PG E asy translates rules into Engl ish it uses the TRANSLATION property wh i ch construct ion. I f o f DEMAND-LEVEL, the system developer supp l ies dur ing expert system at a part i cular stage, the system needs to know the type it uses the PROMPT "What is the level of the future forec asted demand?", and prompts the user tc choose between CONSTANT-OVER­ TIME and VARIABLE-OVER-TIME. If the user is not exactly sure what the question is asking, he may obtain further HELP by pressing a function key and get an expl anat ion as displayed in Figure 1 . 11 12 Insert Figure 1 here Sometimes , the values of some parameters are determined internally by the s y s tem . This may be neces sary when the user may not be able to answer questions very easily pertaining to the se parameters. For example , following rule determines in an indirect manner , whether the general model type is EOQ: RULE004 (INFERRING-RULES ) If 1) the forecast for the future demand i s DETERMINISTIC , and 2) the level of the future demand i s CONSTANT-OVER - TIME , Then it is def inite (100%) that general model type i s EOQ . IF' : DEMAND-PROCESS THEN: GENERAL-TYPE DETERMINISTIC AND DEMAND-LEVEL EOQ DESCRIPTION: General type ECQ. CONSTANT-OVER-T!ME This is achieved in an indirect way , by a s k ing the user two eas ier to answer questions regard ing the demand for the product under consideration. Three of the parameters in the above l i st, i.e. MODEL , REFERENCE-LIST­ BOOKS, REFERENCE-LIST-ARTICLES are the goals in the knowledge bas e which the PC Easy tries to prove. Us ing the backward-chaining inference mechanism des cribed in the Appendix , system tries to prove the above three goal s , one by one. When the goal MODEL is proven , the other two are al so automat i cally proven and the system reaches a conclus ion , and l is t s a l l the books and articles used in the knowledge base for the user's information . In its present vers ion , EXPIM has 68 rules with 12 rule groups constructed for ease of development. These rule groups are : i) EOQ Related Groups 1 . EOQ and Extensions 2 . "Not Clas sical EOQ" 3. EOQ Po s s ib i lities ii) 4. Determini s t ic Non-EOQ i i i) Newsboy Model Related Groups iv) 5. Newsboy 6. Newsboy Pos sibilities Stocha stic " Non-Newsboy" 7. Cont inuous Review Groups 8. Cont inuous Review Po s s ib i l ities 9. Per iodic Review 10. Periodic Review Pos s ib i l ities v) Auxil iary Rule Groups 11. Inferring 12. Conclus ions The first rule group has 14 rules each of which g iving rise to an EOQ type model ranging from the c la s s ical EOQ to production lot s ize with lost sales. As an example the fol lowing Rule #5, if sat i s f ied , recommends the cla s s ical EOQ model with a mes s age to the user warning him to check h i s a s s umption before using th i s model. The rule also provides a pub l i shed 1 3 14 reference for th i s model along with the name of the bas ic l anguage computer program ( EOQ.BAS ) the user can run after the consultation is over. ( The program s are prov ided with the knowledge base ) . RULE005 [EOQ - AND-EXTENSIONS-RULES] I f 1 ) general mode l type is EOQ , and 2 ) lead-t ime i s ZERO , and 3) number of products cons idered i s ONE , and 4 ) inventory shortage i s NOT-PERMISSIBLE , and 5 ) replenishment rate o f the order <or product ion quant ity> i s ALL-AT-ONCE, and 6) amount paid to the suppl ier for each unit purchased is INDEPENDENT-OF-ORDER-QUA..�TITY and 7) 1 ) future inflat ion expectations i s LOw , or 2) future inflation expectations is STABLE, and 8 ) item's l i fetime is SUFFICIENTLY-LONG, and 9) storage space availabi l ity is SUFFICIENT, and 10) coordination of replenishments is NOT-FEASIBLE , Then it i s definite ( 100% ) that the model you should use is Class ical EOQ. Th i s is the oldest inventory model available in inventory literature. Because o f its s imp l icity in computations, it has frequently been misused. It i s a good idea to re-check your assumptions be fore adopting this mode l. REFERENCE : [SIL-PET] , pp. 174-180 . EOQ . BAS Sect ion s 5.1.5.3 . COMPUTER PROG&:-'\M: IF GENERAL-TYPE EOQ AND LEAD-TIME = ZERO AND NUMBER-OF-PRODUCTS = ONE AI'ID SHORTAGES = NOT-PERMISSIBLE AND REPLENISHMENT = ALL-AT­ ONCE AND UNIT-VARIABLE-COST = INDEPENDENT-OF-ORDER-QUANTITY AND INFLATION LOW OR INFLATION = STABLE AND LIFE = SUFFICIENTLY-LONG AND SPACE SUFFICIENT AND COORDINATION = NOT-FEASIBLE THEN: MODEL TEXTVAL :LINE 2 :TAB 2 "C las s ical EOQ . " :LINE 2 :TAB 2 "This is the o ldest inventory model available in inventory l iterature. Because of its s imp l icity in computations, it has frequent ly been m i sused . It is a good idea to care fully I recons ider your as sumptions be fore adopting th i s model . " :LINE 2 :TAB 2 "REFERENCE : [SIL-PET] , pp. 174-180 . Sections 5. 1-5.3 . " :LINE 2 : TAB 2 "COMPUTER PROGRAM : EOQ . BAS" : LINE 2 DESCRIPTION : · C l as sical EOQ [ 100] It is important to note that, along with the EOQ type general model requirement nine other requirements are present in order for the model to recommend the C l as s ical EOQ mode l . These are (LEAD-TIME = Zero) , NUMBER-OF- PRODUCTS ( =One), SHORTAGES (=Not Perm i s s ible), REPLENISHMENT , (= A l l at Once), UNIT-VARIABLE-COST (= Independent of Order Quantity), INFLATION (= Low or Stable) , LIFE (=Sufficiently long), SPACE (= Suffic ient) and COORDINATION (= Not Fe a s ib le) . If any one of these is not s at i s f ied , but the others are , then the system recommends a mode l wh ich is the extens ion of the c lassical EOQ . When the general type is EOQ, and if any two of the above nine parameters are not s at i s fied in Rule #5 above, then we can no longer have a model which qua l i fies as an EOQ extens ion. The second rule group "Not Class ical EOQ" contains a total of e ight new rules e ach reach ing the conclus ion that po s s ible choice for an EOQ model is not c las s ical EOQ. Although it was pos s ible to make up a s ingle rule w ith 3 6 condit ions 15 in its I F part , we preferred to use eight different rules for facility o f d ebugging. Following Rule #44 is an example to the rules in this group . RULE004 [NOT- CLA S S I CAL-EOQ-RULES ] I f 1 ) general model type i s EOQ , and 2) 1) 1) l ead-time is not ZERO , and 2). number o f products considered is MULTIPLE , or 2) 1) lead - time is not ZERO , and 2) 1) inventory shortage is PERMISSIBLE-BACKLOGGING , or 2) inventory shortage is PERMISSIBLE-LOST-SALES , or 3) 1) lead-time is not ZERO , and 16 2) replenish..�ent rate of the order <or production quantity> is GRADUALLY , or 4) 1) lead-time is not ZERO , and 2) amount paid to the supplier for each unit purchased is DEPENDENT-ON-ORDER - QUANTITY , or 5 ) 1) 2) 6) 1) lead-time is not ZERO , and future inflation expectations is HIGH , or lead-time is not ZERO , and 2) item's lifetime is LIMITED, or 7) 1) lead time is not ZERO , and 2) storage space availability is LIMITED , or 8) 1) l ead-time is not ZERO , and 2) coordination o f replenishments is POSSIBLE , Then it is d e finite ( 100%) that possible choice for an EOQ model is Not Classical EOQ . If: GENERAL - TYPE EOQ AND (LEAD - TIME != ZERO AND NUMBER - O F - PRODUCTS MULTIPLE) OR (LEAD - TIME != ZERO AND (SHORTAGES PERMI S SIBLE - BACKLOGGING OR SHORTAGES = PERMI S SIBLE-LOST - SALES)) OR (LEAD-TIME != ZERO AND REPLENISHMENT GRADUALLY) OR (LEAD-TIME != ZERO AND UNIT- VARIABLE - CO S T = DEPENDENT - ON - ORDER-QUANTITY) OR (LEAD-TIME != ZERO AND INFLATION = HIGH) OR (LEAD - TIME != ZERO AND LIFE = LIMITED) OR (LEAD- TIME != ZERO AND S PACE LIMITED) OR (LEAD-TIME != ZERO AND COORDINATION = POS SIBLE) THEN: EOQ-MODEL· = "Not Classical EOQ" Here, the symbol != means " is not". The third rule group has a collection of rules which would be act ivated when the general type i s EOQ and the pos s ible choice for an EOQ model is not the classical EOQ. When these conditions are satisfied , and if any one of .- the above nine parameters has a value wh ich would make the model approximate an EOQ extension , then these rule s would recommend that model , but with a s ubstantially reduced confidence factor . An example is Rule # 5 6 below , wh ich recommends us ing EOQ with quant ity d iscounts (conf idence factor 30) , but cautions the user to be careful. RULE056 [EOQ - PO S SIBILITIES-RULES] If 1) general model type is EOQ, and 2) possible cho ice for an EOQ model is Not Class ical EOQ, and 3) amount paid to the supplier for each unit purchased i s DEPENDENT - ON-ORDER - QUANTITY, Then there is weakly suggestive evidence (30%) that the model you should use is EOQ with quantity d i scounts . Caution must be exerc ised in u s ing this model as not all the assumpt ions lead ing to the model are satisf ied . Plea se 17 check your input values using the REVIEW command. REFERENCE pp. 62-66. COMPUTER PROGRAM EOQDISC . BA S . [HAD-WHI] , IF : GENERAL - TYPE VARIABLE-COS T EOQ AND EOQ-MODEL "Not Cla s s ical EOQ" AND UNIT- DEPENDENT-ON-ORDER-QUANTITY 18 THEN : MODEL TEXTVAL : LINE 2 : TAB 2 "EOQ with quantity discounts . " : LINE 1 : TAB 3 " Caution must be exercised in using this model a s not all the a s sumptions leading to the model are satis fied . Plea se check your input values using the REVIEW command . " : LINE 1 : TAB 3 " REFERENCE :[HAD-WHI] , pp. 62-6611 : LINE 1 : TAB 3 "COMPUTER PROGRAM : EOQDISC.BA S " C F 30 DES CRIPTION : EOQ with quantity discounts [ Unit-variable - cost 30 The fourth rule group contains rules which are related to determin istic non-EOQ type model s , such model (Wagner and Whitin recommending stochastic as the time vary ing, periodic review lot-size 1958) . Rule groups 5 to 10 contain rules models only. All of these rules start by determining the general model type and then ask other questions trying to reach the goal MODEL and make a recommendation. It should perhaps be clear by this time that our expert system clas sifies all of the inventory models into four mutually exclus ive (and collectively exhau stive) groups . This is achieved by identi fying the value of the parameter GENERAL-TYPE as one o f 1. EOQ , 2 . Deterministic Non-EOQ , 3. Newsboy , and 4 . S tochastic Non-Newsboy . At the outset , to assign a value to GENERAL - TYPE , the user is asked a question on DEMAND-PROCES S and DEMAND-LEVEL or LIFE depending on the answer given to the first question. Answering these two prompts , imme�iately places the system in one of the four groups , thereby effectively eliminating all the other irrelevant rules pertaining to the three groups. For example, i f GENERAL - TYPE is found to be NEWSBOY , the rules in EOQ and Extensions group, etc. are not considered at all since they would not be relevant to a stochastic demand model. This process is described in Figure 2. Insert Figure 2 here Now , the systems would attempt to prove whether the classical Newsboy or any one of its extensions is the applicable model , using e.g . rules such as Rule #68. RULE068 [NEWSBOY-RULES) If 1) general model type is NEWSBOY , and 2) number of products considered is MULTIPLE , and 3) substitute products exist , and 4 ) fixed order <or set-up> cost i s ZERO , 19 Then it is definite (100%/ that the model you should use is Classical Newsboy with Substitutable Products REFERENCES : [ PAR-GOY] , Opsearch, Vol . 21 , pp . 1-15, (1984 ) for the two item case with exact solution: [ PAR] , 20 Opsearch, Vol . 23, pp . 250-257, (1986) for the multi-item case with an heuristic . COMPUTER PROGRAM : NEWSBSUB . BAS IF: GENERAL-TYPE = NEWSBOY AND NUMBER-OF-PRODUCTS = MULTIPLE AND SUBSTITUTE AND FIXED-COST = ZERO - THEN : MODEL TEXTVAL :LINE 2 :TAB 2 "Classical Newsboy with Substitutable Products" :LINE 1 :TAB 3 " REFERENCES : [ PAR-GOY], Opsearch, Vol. 21 , pp . 1-15 , ( 1984 ) for the two item case with exact solution: [PAR] , Opsearch , Vol . 23, pp . 2 5 0-257, (1986 ) for the multi - item case with an heuristic" :LINE 2 :TAB 2 " COMPUTER PROGRAM : NEWSBSUB.BAS" DES CRIPTION : Classical Newsboy with Substitutable Products [ 100 ] Although it seems obvious, it is very important to emphasize that a structure such as above would cut down the search space of the problem and reduce the time required to f ind the correct model. For any expert system , the first few questions the user answers , should eliminate many irrelevant outcomes and make the search for the goal more efficient . Clearly , there are some cases where the user's responses are so unusual that , a relevant model does not exist for the given situation. An example consultation for such a case is provided in Figure 4 , where the system concludes that no (known) models are available. A result such as this may be useful to researchers working in inventory management by providing them with possibly new research topics not analyzed be fore . ---------------- ---- - Insert F i gure 3 Here Finally , the following f igure describes the results of a part i cular consultation session where the system recommends the Per iod i c lot size model. The reference [SIL - PET] is (S ilver and Peterson 1985). In the second part , the user can REVIEW his previous cho ices and re-run the system . Insert F igure 4 Here As we ment ioned be fore, the s�stem can identify appr ox imately 30 product ion- inventory models found in the current l iterature. Most of the models in every maj or production- inventory textb o ok , including S i lver and Peterson (1985), Hax and Candea (1984), Love (1979) , J o hnson and Montgomery ( 1974) and Naddor (1966) are included . We should note that the current system can be easily extended and ref ined as new models bec ome ava ilable. This is done by add ing new rules to the knowledge base and by poss ibly changing some o f the older rules wh ich interact with the new ones . It is obvious that as long as new inventory models are published and become ava ilable, this expert System w ill continue growing and will be r i cher in content and expert ise. 21 2 2 4 . Les s ons Learned As the use of expert systems in management sc ience applicat ions i s relat ively new , it would be worthwhile to summarize our exper iences gained dur ing the development of EXPIM. It is hoped that other researchers who are working on s imilar proj ects , e . g . queueing expert systems would f ind these hints u s eful. We recommend that the developer u s e an expert system shell instead of an AI language such as LIS P or Prolog. The selection of the shell should be based on its capability to do backward chaining with a forward cha ining option . qu ickly Us ing backward chaining , many irrelevant cho ices can be eliminated in the early stages. It is advisable to use a shell which can interface with out s i de programs and files . For example , many expert systems may have to do heavy mathemat ical computat ions dur ing or at the conclus ion of a consultation (e . g . running a program to solve the EOQ model in EXPIM). S imilarly , they may need to read data from outs ide files or from databases instead of ask ing the user to supply the se data value s . A s management sc ience application would invar iably involve some computat ions , these features become important. Many shells have the capab ility to answer WHY and HOW quest ions , and do sensit ivity analy ses. It is cruc ial that the developer choose one with these capab ilities. As for the development of the knowledge base , the developer should group the models into mutually exclus ive and collectively exhau s t ive groups so that any available model can be placed in exactly one of the groups . For example , in the case of queueing expert sy stem , one may init ially create two groups , i) s ingle server models and i i) multi-server models. In each group a rule should be written for the b e st known model e . g. classical EOQ in deterministic EOQ and M/M/1 in s ingle s erver queues. Then , extensions of this fundamental model should be c onsidered by writing slightly different variations o f the rule for the fundamental model . Finally, new rules should be created which would cho o s e the s e known models with lower confidence factors because of s om e o f the unsatisfied a s sumptions , e.g . Rule #56 discu s s ed in S ection 3. It is als o important t o have s ome idea who the u s er s might b e befcre the development starts. It would be us eful to have rul es which c o uld reach conclu sions by asking the user easy to understand questions , e . g. Rule #4 discu s s ed in S ection 3. When particular article) model in the model . the expert system reaches a conclusion by recommending a model, it should al so give the us er the reference (book or where the model can be found. This way , the u s er can analyze the more detail and perhaps s e e a few numerical example s r el evant to 5. Summary and Conclusions In this paper d evelopment of an we have di s cus s ed in detail , the conc eptualization and expert system (EXPIM) which can identify up to 30 production and inventory models. In the abs ence of a hu.�an expert , the manager who is interested in using thes e models can a c c e s s the expert system and a ft e r r e s ponding to a few questions , he can get a recommendation from the system . EXPIM is written in Texas Instruments' Pers onal Consultant Easy expert system shell and runs on IBM-PC micro computers with at least 512K memory . 23 24 As mentioned in Section 4, there is a potential for creating other expert systems in management science for classifying different collection of models. For example, queueing theory, with its plethora of models is a prime �andidate for this research. Location theory is another possibility. We hope that our points summarized in section 4 will aid other researchers who want to enter this new and interesting research field which combines management science and artificial intelligence. We note that to make these expert systems more "intelligent", they should have the ability to reach some intermediate conclusions on their own. For example, in EX.PIM, instead of asking the user about the DEMAND-LEVEL , the system should ideally be able to determine it on its own. This would require tha system to access external data files, do some computations as discussed in Silver and Peterson (1985, p.238) to obtain a measure of the variability of demand pattern and then determine the value of the parameter DEMAND-LEVEL. Current version of EXPIM does not have this capacibility, but it would be possible to include it in a later version. In the early days of expert system development, artificial intelligence researchers used to recommend that a knowledge engineer should not be his own expert (Nii 1983). This is still true to a certain extent but these days the availability of easy to use shells have made it easier for the domain experts to become proficient in the use of these tools and have assumed the role of a knowledge engineer. Perhaps the convenience and low price of microcomputers which can run AI software have played an important role in this development. To conclude, it is our belief that expert systems will be playing an important role in management science especially in classification and choice of models as we discussed in this paper and also in automatic model building (Binbasioglu and Jarke 1986). ACKNOWLEDGEMENT Financial support from the Natural Sciences and Engineering Research Council of Canada under Grant No. A5872 is gratefully acknowledged. 25 List of Figures Page Figure 1. The HELP Facility 12 Figure 2 . Initial Classification of Models 19 Figure 3. The case of No Conclusions 21 Figure 4 . The " Conclusions" and " Review" Screens 21 Backward Chaining Append ix Backward and Forward Cha ining in PC Easy (From PC Easy Reference Guide , pp. 3-19/ 3 - 30) Al ·Backward cha ining is the primary means by which PC Easy arrives at a solution to a problem- - the knowledge base goal . The driving force behind a PC Easy consultation is the attempt to set a value for each of the one or more parameters l isted in the knowledge base GOALS property. PC Easy begins the attempt to set the value of a goal parameter by looking for a rule whose THEN sta tement ass igns a value to the parameter. When it finds a rule , it tests it , or determ ines whether the condit ions expressed in the rule's IF statement are true. To determine whether the condit ions expressed in the rule's IF statement are true , PC Easy may need to f ind the value of one or more parameters included in the IF statement. Lhis search may lead to another rule that sets the value of one or more of these parameters. If the conditions in the IF statement are true , the rule passes. If the cond itions in the IF statement are not true , the rule fails. If a rule passes , PC E asy f ires the rule--car r ies out the action specified in its THEN statement. PC Easy fires a rule only once dur ing a consultation . A2 Example In this example , PC Easy find s the value of the goal parameter GIFT by backward chaining. GIFT-TYPE and RECEIVER are parameters in the knowledge base. 1. The consultation begins. 2 . P C Easy find a rule that a s signs a value to GIFT: IF statement: GIFT-TYPE = PERSONAL THEN statement: GIFT = FLOWERS To determine whether to apply the action of this rule and set the value of GIFT TO FLOWERS, PC Easy must search for the value of GIFT-TYPE. 3. PC Easy finds a rule that assigns a value to GIFT-TYPE: IF statement : RECEIVER = SPOUSE THEN statement : GIFT-TYPE = PERSONAL To test this rule, PC Easy must find the value of RECEIVER. 4. Because no rule assigns a value to RECEIVER and RECEIVER has a PROMPT property , PC Easy prompts the client: Is the receiver o f the gift your spouse? A3 5 . I f the cl ient answers yes, P C Easy.a s s igns RECEIVER the value S POUSE. Because the condition stated in the IF statement of the rule i s true, PC Easy carries out the action of the THEN statement and a s s igns GIFT-TYPE the value PERSONAL. 6 . Because GIFT-TYPE has the value PERSONAL , the condition stated in the I F statement of the original rule i s true , and P C Easy determines that the value of GIFT i s FLOWER S. Forward Cha ining In forward cha ining , PC Easy tries an antecedent rule when it has a s s igned a value to one of the parameters in the antecedent rule's IF s tatement. In EXPIM an antecedent rule is used to invoke a rule whose THEN s tatement does not conclude values for any parameter. For example , when the model type i s determ ined to be STOCHASTIC-NON-NEWSBOY and if the user selects NON-PERMIS SIBLE for inventory shortage paramete r , then he gets a me s sage indicat ing the unava ilability of any models for the current s ituation. The follow ing antecedent Rule #74 des cr ibes an example to forward chaining in P C Easy . RULE074 [ PERIODIC-REVIEW-RULE S/antecedent] If 1) general model type is STOCHASTIC-NON - NEWSBOY , and 2) inventory shortage is NON-PERMI S SIBLE , A4 Then it is definite (100%) that the model you should use is There is no available model . Please note that you- chose PERMI S S IBLE when the computer asked you about the SHORTAGES . P r io r to that you had spec i f ied that DEMAND was S TOCHASTIC , and LIFETIME was SUFFICIENTLY-LONG. Under these conditions it is impossible to require that - inventory should always be posit ive. Please REVIEW your responses and change them accord ingly. IF: GENERAL-TYPE = STOCHASTIC-NON-NEWSBOY AND SHORTAGES = NOT .. P ERMI S SIBLE THEN: MODEL = TF.XTVAL :LINE 1 : TAJ3 3 " There is no ava ilable model. Please note that you chose PERMI S SIBLE when the computer asked you about the SHORTAGES . Prior to that you had specified that DEMAND was STOCHASTIC, and LIFETIME was SUFFICIENTLY�LONG. Under these conditions it is impossible �o require that inventory should always be posit ive. responses and change them accord ingly". Please REVIEW your DESCRIPTION: Warning the user if posit ive inventory is requ ired in stochastic case ANTECEDENT: YES ® Bl Bibliography AGGARWAL, S.C., "A Review of Current Inventory Theory and Its Applications", Int. J. Prod. Res., 12, #4, (1974), 443-482. ASSAD. A.A. and B.L. GOLDEN, "Expert Systems, Microcomputers and Operations Research", Comput. & Oper. Res., 13, #2/3, (1986), 301-321. AZARM, S. and M. PECHT, "An Approach for Building a Rule-Based System for Design Optimization", Working Paper No. TR-85-51, Dept. of Mechanical Engineering, The U. of Maryland, 1985. BARANCSI, E., G. BANKI, R. BORLOI, A. CHIKAN, P. KELLE, T. KULCSAR and GY. MESZENA, "Analysis of a System of Inventory Models", in Proc. First Int. Symp. on Inventories, Budapest, Hungary, pp. 297-308, (1980). BARANCSI, E., , , , , , , "A Report of Research on Inventory Models, 11 En�sts�Prod.�on., 7, (1983), 127-136. BARR, A. and E.A. PEIGENBAUM, The Handbook of Artificial Intelligence, Vol. 1, William Kaufmann, Los Altos, Calif., 1981. BINBASIOGLU, M., and M. JARKE, "Domain Specific DSS Tools for Knowledge Based Model Building", Dec. Sup. Syst., 2, (1986), .213-223. BISWAS, G. , R. ABRAMCZYK and M. OLIFF, "OASES: An Expert System for Operations Analysis The System for Cause Analysis", IEEE Trans. on Sys. Man and Cvb., SMC-17, #2, (1987), 1 3 3-145. BUCHANAN, B.G. and R.0. DUDA, "Principles of Rule-Based Expert Systems", in M. YOVITS, (Ed.), Advan�es in Computers, Vol. 22, (1983), 16 3-216. CHIKAN, A., P. KELLE and GY. MESZENA, "A Comparison of Various Classification Systems of Inventory Models", in Proc. Second Int. Symp. on Inv., Budapest, Hungary, 661-667, 1982. DAVIS, R. and J. KING, "An Overview of Production Systems", in E. ELCOCK and D. MICHIE (Eds.), Machine Intelligence, Ellis Horwood, Chichester, England, 300-332, 1977. DOUKIDIS, G.I. Simulation 325. DUDA, R. 0. , J. consultant Systems in 167' (1979). and R.J. PAUL, "Research Into Expert Systems to Aid Model Formulation", J. Opl. Res. Soc., 36, #4, (1985), 319- GASCHNIG and P.E. HART, "Model Design in the PROSPECTOR system for Mineral Exploration", in D. MICHIE (Ed.), Expert the Microelectronic Age, Edinburgh Univ., Edinburgh, 153- EILON, S. and W. LAMPKIN, Inventory Control Abstracts. (1953-1965), Oliver and Boyd, Edinburgh, 1968. B 2 FEIGENBAUM , E.A. and P . MCCORDUCK , The Fifth Generat ion , Add i s on - Wes ley , Read ing , Mas s. 1 9 8 3 . HADLEY , G. and T.M. WHITIN , Ana ly s i s o f Inventory Systems , Prent ice - Ha l l , Englewo od C l iffs , N.J . 1963 . HAHN , G . J. , System s : " Mo re Intel l igent Stat i st ical Software and Stat i s t ical Expert Future D i rections" , Amer . Stat i s t i c ian , 3 9 , # l , ( 1 9 85) , 1 - 8. HAX , A . G. and D . CANDE;A , Product ion and Inventory Management , Prentice-Hal l , Englewood C l iffs , · NJ , 1 9 84 . HAYES - ROTH , F . , D . A. WATERMAN and D.B . LENAT , ( Eds.) , Build ing Expert Systems , Addi s on-We s ley , Reading , Mas s. , 1 9 8 3. HOLL IE R , R . H. and P . VRAT , "A Propo s a l for C la s s ification of Inventory System s " , Omega , 6 , #3 , (1978) , 277-279. IGLEHAR T , D . L . , "Recent Results in Invent o ry Theory" , J . o f Ind. Eng. , 1 8 , #l , ( 1967)' 4 8 - 51. JOHNSON , L . A. and D.C. MONTGOMERY , Operations Research in Product i on Pl anning , Schedul ing and Invent ory Contro l , J ohn W i ley , New York , 1974. KASTNE R , J.K. and S . J. HONG , "A Review of Expert Systems " , Euro. Jnl. of Onnl? Res. , 1 8 , #3 ( 1 9 84) , 2 85 - 2 9 2. KENDALL , D.G. " Some P r ob lems in the Theory of Queue s " , J . R . S t at . S o c . , B- 1 3 , #2 (1951) , 151-1 85. KHOSHNEVIS , B . and A. CHEN , " An Expert S imulat i o n Mode l Builder " , Wo rking Paper No. 8 4-04 , Dept of ISE , U . of S outhern Ca l ifornia , 1 9 86. LINDSAY , R . , B.G . BUC!lANAN , E.A. FEIGENBAUM and J . LEDERBURG , D ENDRAL , MCGRAW - HILL , New Y o rk , 1 9 80 . .. LOVE , S . F. , Inventory C ontr o l , McGraw-H i l l , New Y o rk , 1 979 . MENIPAZ , E. , "A Taxonomy and Shibb o leth of Inventory M o de l s " , in Proc. Second Int . Symp. on Invent o r ies , Budape s t , Hungary , 745-75 2 , 1 9 8 2. NADDO R , E. , Inventory Systems , J ohn W i ley , New Y o rk , 1966. NAHMIAS , S . , (Eds.) , 1 97 8 . " Inventory The Encycl. Models " , in J. BELZE R , A.G. HOLZMAN and A. KENT of Comp . Sci. and Techn . , V o l. 9 , Marcel Dekker , N I I , H . P. , "Heur i s t ic s f o r Knowledge Eng ineer ing" , reported in ( Feigenbaum and McCorduck , p . 8 3 , 19 8 3) . POST , E . , " Formal Reducti ons of the General Combinational Prob lem " , Am . Jnl . of Math. , 65 , (1943) , 1 97 - 268. PRABHU , N.V. , Queues and Invento ries , J ohn W i ley , New Y o rk , 1 9 65. B 3 RAVINDRAN , A. , D . T . PHILLIPS and J.J. SOLBERG , Operations Research ; Principles and Pract i ce , 2nd Ed., John Wi ley , New York , 1987.- R EGGIA , J . A. and S . B. AHUJA , " Select ing an Approach to Knowledge Proces s ing " , Comp. and Oper . Res. , 1 3 , #2/3, ( 1986 ) , 329 - 3 3 1. SCHWARZ., L . B. ( Ed . ) , Mult i - level Production/Inventory Control Systems : Theory and Practice , North - Ho l land , Amsterdam , 1981 . SHORTLIFFE , E . H. , Computer Based Medical Consultat ions : Myc in , Elsevier , New York , 1976. SILVER , E . A. and R. PETERSON , Dec is ion Systems for Inventory Management and Production Planning, 2nd Ed., John Wiley, New York , 1985. SILVER , E . A. , " Operations Research in Inventory Management : A Review and Critique " , Oper. Res. , 29, #4 , ( 1981) , 628 - 645 . STEFIK , M . , J. AIKINS, R . BALZER , J . BENOIT , L. BIRNBAUM , F . HAYES-ROTH and E. SACERDOTI, " The Organizat ions of Expert: Sys tems , A Tut o r ia l " , Art i f i c ial Intell i gence, 18 , ( 1982) , 1 35-173. Vl> .. N BRED."., M. F. , "Expert Systems in Univers ity Admi s s i on De c i s i ons " , Presented at the TIM S/ORSA Meet ing, L o s Angeles, (Ap r il 1986). VEINOTT, A . F . , "The Status of Mathemat i cal Inventory Theory " , Hgt. Sc i . , 1 2 , # 1 1 , ( 1966) , 745 - 777. WAGNER, H. and T.M. WHITIN , " Dynami c Ver s i on M odel" , ..,M._,g�t_,_.--=-S-"'c=i. , 5 , #l , ( 1958) , 8 9 - 96. of the Ec onomic Lot S i ze WAGNER , H . , "Research Portfol i o for Inventory Management and Product i on Planning Systems" , · Oper. Res ; , 28 , ( 1980) , 445-475. WATERMAN , D . A., A Guide to Expert Systems , Add i s on Wesley , Read ing , M a s s . 1986. WEISS , S . M . and C . A. KULIKOWSKI, A Practi cal Guide to Des igning Expert Sy stems , R owman and Allanheld , Totowa , N . J . 1984 . WISARD So ftware Co . , Wisard Forecaster for IBM-PC , Green Bay , WI , 1986. EXP I M Expert P r o duct i on - I nv e nt o ry Mode l l er r--�---�������--�����-�����--�-��-����-���-� ) What is the l ev e l o f the f uture f o r e c a s t e d demand? \ l CONSTAN T -OVER - T I ME l H e lp : ���������������������-. I VARIABLE-OVER - T IME J I I f the f o r e c a s t e d f ut u r e l ev e l of demand i s b a s i c a l l y c o n s t ant ( un i f o rm over t im e ) , c ho o s e CONSTANT - OVER - T I ME . I f the f or e c a s t i nd i c at e s lumpy demand where demand var i e s f r om t im e to t im e , cho o s e VAR I ABLE - OVER- T IME : R e f e r t o [ S I L -PET ] , p p . 2 3 7 - 2 3 9 , S e c t i on 6 . 6 . 4 . f o r an exp l anat i o n o f a numer i c a l t e c hn i qu e whi ch can d i st in gu i sh b etween t h e s e two c a s e s . ** End - RETURN /ENTER to cont inue 1 U s e the a r r ow k e y s or f i r s t l et t e r o f i t em t o p o s i t i on the c ur s o r . I _ ___ :_: Pr e s s RETURN/ENTER t o _c_o_n_t_i_n_u __ e_. ____________________ _ , F :i;g.u;r e .J. I _} DEMAND - PROCE S S : D e t e rm i n i s t i c � DEMAND- LEVEL C o n s t ant V a r i ab l e L I FE I 'f EOQ D e t e rmini s t i c n o n -EOQ Figure 2 S t o cha s t i c I L im i t e d + N e w s b o y L o n g + Sto cha s t i c n o n - N e w s b o y ( " C o n s u l t at i on r e c o rd f o r : EXP I M : Expert Product i on - I nventory Mode l l er " " th e f o re c a s t f or the f u t u r e demand . . S TOCHAS T I C " " i t em ' s l i f et ime SUFF I C I E N T LY - L ONG " " pe r i o di c mon i t o r i n g of the invent o ry . . . YE S " " l e a d - t i m e STOCHAS T I C " " f ix e d o rd e r < o r s e t - u p > c o st . . POS I T I VE " " i nventory s h o r t a g e PERM I S S I BLE - LOS T - SALES " " nuI:lb e r o f p r od u c t s c o n s i de r e d MULT I PLE " " s u b s t i tu t e p r o d u c t s exi st YES " ) CONCLU S I ON S : I was un ab l e to make any c o nc l u s i on s r e g a r d i n g the m o d e l y o u s h o u l d u s e L i s t o f b o o k s u s e d i n thi s know l e dg e b a s e i s a s f o l l ows : N o t ava i l ab l e s i n c e t h e r e are no known ( o r e a s i l y imp l ementab l e ) mo d e l s f o r s o lv i ng the inventory p r ob l em d e s c r i b e d by the u s e r in thi s c o n s u l t at i on s e s s i on . P o t e n t i a l r e s e a rch t o p i c s may b e o b t a i n e d f rom t h i s c o n s u l t a t i on ' s r e s u l t s b y r e f e r r i n g t o t h e c o m b i nat i o n o f i nput s p r o v i d e d b y t h e u s e r . P l e a s e u s e the REVIEW f ea t u r e to r e c a l l your input s , i . e . the paramet e r s you s upp l i e d d u r i n g t h e c o ns u l t at i on L i s t of art i c l e s u s ed in thi s kno w l e d g e b a s e i s as f o l l ows : Agai n , n o t ava i l ab l e f o r the r e a s o n s g i v e n above 'F:i:gu:re . .3 EXP IM Expert P roduct i on - I nvent o ry M od e l l e r r C on c l us i o n s :�����������������������������������-. The mo d e l you s h o u l d u s e i s . a s f o l l ow s : P e r i o d i c r e v i ew l ot - s i z e I Thi s m o d e l c o u l d be s o lved u s i n g the Wagn e r - Wh i t i n exact a l go r i thm or b y ! u s i n g S i l v e r - M e a l heur i s t i c o r by u s i n g a f ew o t h e r heur i s t i c s reported i n the l i t e r ature . l P.EFERENCE : [ S I L -PET ) , pp . 2 2 7 - 24 3 , S e c t i o n s 6 . 5 - 6 . 7 L i s t o f b o oks r e f e r e n c e d i n thi s knowledge . b a s e i s a s f o l l ow s : BOOKS : [ BAN - FAB J : J . Banks and W . J . Fab rycky , ' Procurement and I nvent o ry S y s t em s ! An a l y s i s ' , Prent i c e - H a l l , 1 9 8 7 . [ BE R J : D . P . B e rt s eka s , ' Dynam i c P r ogramming and S t ocha s t i c C o n t r o l ' , A c a d e m i c P r e s s , 1 9 7 6 . [ E L S - BOU ) : E . A . E l s ay e d and T . O . Boucher , ' An a l y s i s and C o n t ro l o f \ Produc t i o n S y s t e m s ' , Prent i c e - H a l l , 1 9 8 5 . I [ HAD-WH I J : G . H a d l ey and T . M . Whi t i n , ' An a l y s i s of I nvent o ry S y s tems ' , ! Pr e nt i c e - Ha l l , 1 9 6 3 . I ** M o r e - RETURN/ENTER to c o n t i n u e EXP I M : Expert Product i on - I nv e nt o ry M o de l l e r -Conc l u s i ons : \ Management and Production P l ann i n g ' , 2�d E d i t i on , John W i l ey , 1 9 8 5 . I ( L i .-Rev i e w : - J C o 1 9 S u 1 - Ye s ,_,,____,,����-.,,..��::----::--�����...,....����-=====�:-:==-=::-.::-� CEJ · l the f o re c a s t f o r the future demand DETERM I N I S T I C I � the l eve l of the f uture demand VAR I ABLE - OVE . . . � • p e r i o d i c mon i t o r i n g of the inven t o ry . . . YES r l e ad - t ime ZERO r • numb e r -of p r o d u c t s c on s i d e r e d ONE � • inventory shortage NOT - PERM I S S IBLE � • r e p l e n i shment rate o f the o r d e r < o r A L L - AT -ONCE wi r · supp l i er ' s quan t i t y d i s c ount . . NOT -AVAI LABLE 2 5 i I- • future i n f l at i on e xp e c t a t i o n s . : : LOW 1 . U s e arrow keys or f i r s t l et t e r of i t em t o p o s i t i o n c u r s o r . I 2 . S e l e c t a l l app l i c ab l e r e s p o ns e s . 1 l s h ...__3_. �A_f_t_e_r ...,- m_a_k_i_· n�g�s_e_l_e_c_t_i_o_n_s...,-, �p-r_e_s_s�RE�T-U_Ri_N_ !E�N-1_T_E_R�t-o�c-o_n_ t�i -n_u_e_.�...,-����-' Figure 4 Faculty of Busine s s McMaster University WORKING PAPERS - RECENT RELEASES 257 . Joseph B . Ro s e , " S tatutory Expedited Grievance Arbitration : The Case of Ontario" , July, 1986. 258 . 259 . 260 . 261. 262. 263 . 264 . 265 . 266. 267 . 268 . 269. 270. 271. John W . Medco f , " The Effects o f Event Valence and Degree of Attribution Upon Action Recommendations " , September 1986. Paul D . Dowling and Robert F. Love , " One - D imensional Floor Layout Solutions Us ing S quared Distances" , September , 1986. Harish C . Jain , "Affirmative Action/Employment E quity in Canada : Findings of a S urvey" , September , 1986 . Min Bas adur , "Catalyzing Interfunctional Efforts to Find and Creat ively Solve Important Bu s ines s Problems", Sep tember, 1986. Roy J . A dams and Isik Zeyt inoglu, " Labour-Management D i spute Resolution in Cana dian Heavy Industry : The H ilton Works Case " , September, 1986. ' R ick D. Hackett, " New Directions in The Study of Employee Absenteeism : A Research Example", October 1986. Ali-Reza Montazemi, "An Analy s i s o f Information Technology A s s e s sment and Adoption in Small Bus iness Environment s", October , 1986. I sik U. Zeytinoglu, " Part - t ime Worker s : Unionization and Collect ive Bargaining in Canada", November , 1986 . Norman P. Archer and M . W . Luke Chan, " The Architecture of Chines e - English Microcomputer Systems " , December , 1986. Norman P . Archer, " Structuring Problems for Analy s i s with Electronic Spreadsheets " , January, 1987 . Norman P . Archer , "A Life Cycle Approach to the Implementation of Integrated Information Technology " , January, 1987. Joseph B . Rose, "An Overview of the Canadian and Australian Industrial Relations Systems " , February, 1987. Joseph B . Ros e , "A North American Perspective on Multi - Employer Cohesion in Australian Construction", February, 1987 . Christopher K. Bart , "Budgeting Gamesmanship : S urvival Tactics in a .Hostile Environment", February, 198 7 . 2 7 2 . 273 . 274 . 275 . 276. 277 . 278. 279. 280. 281. 282. - 2 - Thomas E . Muller , "Model ing the Touri st ' s Urban Experience : A New Approach to Irtternat ional Touri sm Product Development Based On Urban Quality of Life , February , 1987. A. William R i chard s on , "Cost-Vo lume - Profit Analy s i s and the Value of Informat ion : An Evaluation for the Normal and Lognormal D i stribut ions" , 1'.'ebruary , 19.87 . Roy J . Adams , "Industrial - Relations Systems : An Internat ional Comp ar i s on" , March , 1987 . John W. Medco f , "An Integration of S ome Attribution Theor ies", March , 1987. Y . L . Chan and A . R. Montazemi, "An Experimental Study of the Informat ion Cho i ces of Security Analysts" , March , 1987. I s ik Zeyt inoglu , "Part-t ime Workers and Collect ive Bargaining in Internal Labour Markets" , .May , 1987. Abraham Mehrez , "A Mult i-Obj e c t ive Approach to a M ixed Integer Non-Line ar Cover ing Prob lem", July , 1 987. Abraham Mehrez , Yufei Yuan , Arniram Gafni , "S table S o l ut i ons Vs . Multipl i c ative Util ity S olut ions For the A s s ignment Prob lem" , July , 1987. H ar i s h C. J a i n , R i ck D . Hackett , "Empl oyment Equ ity Programs In Canada : Publ i c Po l i cy And A Survey", July , 1987. Norm P. Arche r , "MBA Information Systems Curricu lum Needs : A Bus ine s s S urvey", August , 1987 . Chr i s K. Bart , "New Venture Units : Organiz ing for New Products" , September, 1987 . DSB283 
work_rh2ldlogbze5dldx4ce5fsu4ee	----	Experts’ adjustment to model-based forecasts: 1 Combining SKU-level sales forecasts from models and experts We study the performance of SKU-level sales forecasts which linearly combine statistical model forecasts and expert forecasts. Using a large and unique database containing model forecasts for monthly sales of various pharmaceutical products and forecasts given by about fifty experts, we document that a linear combination of those forecasts usually is most accurate. Correlating the weights of the expert forecasts in these linear combinations with the experts’ experience and behaviour shows that more experience and modest deviation from model forecasts gives most weight of the expert forecast. When the rate of bracketing increases, we notice a convergence to equal weights. We show that these results are robust across twelve different forecast horizons. Key words: Human judgement and decision making; Forecasting May 26 2009 2 1. Introduction There is abundant literature on the relative performance of model forecasts, expert forecasts and their combination, see Lawrence et al. (2006) and Fildes et al. (2009), and the earlier work of Blattberg and Hoch (1990). The most common findings are that expert forecasts can improve on model forecasts and that a linear combination of the model forecast with an expert forecast is often even better. The literature so far mainly considers a few single-product, single-horizon and single-expert cases. In our present paper we aim to extend the currently available literature by considering various products in various product categories, twelve different forecast horizons and about fifty experts. A main additional feature of our analysis is that we know a few characteristics of these experts and we also observe their behaviour. This allows us to correlate the optimal balance between the model and the experts with their characteristics and their behaviour, which in turn gives guidelines from a managerial perspective, and this is new to the literature. In this paper we empirically analyze a unique and very large database with model forecasts, expert forecasts and realizations concerning monthly SKU-level sales of a range of pharmaceutical products for a large Netherlands-based firm. At the headquarters office, the model forecasts are automatically created by a statistical package, where the program each month allows for a re-specification of the model and it also re-estimates all parameters each time. The experts, located in local offices in thirty-seven countries, receive these forecasts and, after that, create their forecasts using their own expertise. We will see that expert forecasts often differ from the model forecasts, which is perhaps not unexpected given the fact that the automatic program includes as input only lagged monthly sales values, and that this fact is known to the experts, see Goodwin (2000, 2002). The question the firm faces is whether the model forecasts and the expert forecasts can be improved by taking a linear combination of the two. A related question is whether this linear combination should follow an unconditional 50%-50% rule, or whether the weights shall depend on the characteristics of the experts. The literature on combining forecasts in for example Clemen (1989) and Timmermann (2006) suggests that linear combinations of forecasts may improve on each of its contributors. So the first question we consider in this paper is whether there are optimal weights for each of the experts. And, if so, is that robust across forecast horizons and does it differ across experts? 3 The second question that we try to answer is whether these optimal weights can be explained by characteristics of the experts. This question is very relevant from a managerial perspective as it facilitates training of experts and also their selection prior to their appointment. Blattberg and Hoch (1990) claim that a 50%-50% rule would be best but this claim corresponds with unconditional weights as it is not correlated with experts’ characteristics. Lamont (2002) demonstrated that age (experience) has a positive effect on the quality of an expert, but also that this effect is parabolic. There are also studies like Barber and Odean (2001) and Beyer and Bowden (1997) which find gender differences in (over-) confidence levels, so perhaps there are also such differences across the relative weights of the experts in the combined forecasts. Finally, the degree of bracketing shall be important for the quality of the combined forecast. Larrick and Soll (2006, p. 112) state that when the rate of bracketing increases, the power of averaging forecasts does too. Their findings were based on experiments, and in the present study we shall seek empirical evidence for this statement based on factual data. The outline of our paper is as follows. In Section 2 we outline the main features of our unique database. Section 3 deals with the methodology and gives the details of our empirical findings. Section 4 concludes with various implications for managers who need to evaluate the qualities of the experts. . 2. Data Our data concern a firm that creates model forecasts and which has almost fifty experts allocated in thirty-seven countries 1 are allowed to report their own forecasts additional to the model forecasts they receive from the headquarter’s office. Average characteristics of these experts are available. The question the firm has is whether specific combinations of these two sets of forecasts are better in terms of point-forecast accuracy, and whether such combinations can be associated with observable characteristics and recent behaviour of these experts. To start, we have data on MFi,j,t+h|t, denoting a model forecast created at time t for horizon t+h for sales in country i for product j. The forecasts concern monthly SKU-level sales of pharmaceutical products, and the sample covers October 2004 to and including October 2006. The countries range from the US, UK, Korea, Austria, Thailand, to Malaysia 1 In some countries there are two experts, and then we average their characteristics in the computations to come. 4 and Mexico. In our analysis below we label the countries as i = 1, 2, .., I = 37 for confidentiality reasons, so we allocate one (average) expert to each of the countries. The index j runs from 1, 2, to Ji, which means that for each country we have a different set of products that belongs to the responsibility of the local expert. The products are associated with seven different product categories. The smallest number of Ji is 10, the largest is 85. Finally, we have forecasts for horizons 1, 2, 3, to 10, 11 and 12 (a year ahead). The headquarters’ office uses an automated statistical package to create these forecasts, where the input contains only lagged sales data. This is known to the experts. The model selection process is rerun each month and also parameter estimation is redone each month, and also this is known to the experts. . Additional to the evidently enormous amount of model forecasts we have access to expert forecasts, to be denoted as EFi,j,t+h|t. The experts make these forecasts upon receipt of the model forecasts, and they are aware of the fact that the automated program only includes lagged sales. To the manager, it is unknown who of the experts takes the model forecast as input for their own forecasts. Typically, as we will see below, the expert forecasts differ from the model forecasts, which is not unexpected given that experts may see various reasons to adjust pure own-history-based projections, see Goodwin (2000, 2002). Unfortunately we have no information on what exactly drives the expert to deviate from model forecasts and also not on which factors they look at when creating their final expert forecasts. We therefore simply take these expert forecasts as a second set of forecasts additional to the model forecasts. Finally, we have actual SKU-level sales data denoted by Si,j,t+h corresponding with the two sets of forecasts. In this paper we aim to draw generalizing conclusions on combining model-based forecasts with expert forecasts. We will analyze the data in the dimensions i and h, and thus to aggregate across the products for each expert. Unreported experimentation with the data at a more detailed level indicated that differences across the products are not relevant, and so we can safely aggregate along that dimension. Insert Table 1 about here To get a first impression of the type of data that we have, we report on some basic statistics in Table 1. For the twelve forecast horizons, we compute the fractions where the model forecasts and expert forecasts exceed or do not exceed each other, and also where they differ from the actual sales data. The first two columns of Table 1 correspond with what is 5 called bracketing, meaning that the expert forecasts and model forecasts are on distinct sides of the realizations. The last two columns of Table 1 concern the cases where the expert deviates from the model forecast into the wrong direction, that is, further away from the actual realization. If we consider the first few forecast horizons, we observe that the fraction that experts deviate from the model in the wrong direction is highest (close to 0.350 by summing the last two columns), where it most often happens that forecasts of experts are too high. Looking at the first two columns we see that bracketing occurs only in around 0.270 of the cases. When we compare the various horizons, we see a slight increase in the fraction of bracketing cases, and a decrease in the fraction of wrong direction cases. The middle two columns seem to be rather constant around 0.330 across the forecast horizons. Insert Table 2 about here That experts have a tendency to create forecasts that exceed model forecasts is made more explicit in the second column of Table 2. This over-optimism bias has also been documented in Fildes et al. (2009). This tendency clearly decreases with increasing forecast horizon. At the same time, and as expected, the model forecasts are around 50% above and 50% below the realizations. This of course corresponds with the mean-reversion tendency of regression-type models with symmetric error assumptions, as they are implemented in the automatic program used by the headquarters’ office. In short, model forecasts are unbiased by their very nature. Insert Table 3 about here Finally, in Table 3 we give the available information we have on each of the thirty- seven experts. As said, in some countries there are two experts, and then the data are averaged across these two experts. For each expert we know the (average) age, gender and the (average) number of years that experts occupy their (forecasting) position within the firm. We see that the average age is close to 40, that there are about as many men as women and that the experts are in office for an average of 9 years. 6 3. Methodology and results To address the managerial questions of the firm, which are typical questions any firm would have to manage a range of experts, we aim to compute the optimal value of the weights in a combined forecast. This combined forecast for each expert i given a horizon h is given by (1) thtjiithtjii EFaMFa |,,|,, )1(   where we compute the value of ai across all products within an expert-horizon combination. To achieve this aim, we compute the root mean squared prediction error (RMSPE) as (2)     iJ j htjithtjiithtjii i SEFaMFa J 1 2 ,,|,,|,, ]))1([ 1 for ai = 0.00, 0.05 (with steps of size 0.05), .. 1.00. This gives 21 RMSPE values for each horizon, and we choose the value of ai with the smallest value of RMSPE. Of course, when ai = 0.00, the RMSPE for the pure expert forecast is lowest across all 21 cases and when ai = 1, the RMSPE for the pure model forecast is lowest. Insert Table 4 about here The results of this exercise appear in Table 4. For example, for expert 5 it holds that for the forecast horizon of 5 months, the optimal value of a5 is 0.75, meaning that her contribution to the combined forecasts is 0.25. Details of the computations can be obtained from the authors, but for now it suffices to say that almost all sequences of 21 RMSPE values show a (slight) parabolic curve, meaning that an optimum most often is reached within the range of considered ai values. A closer look at the optimal weights in Table 4 shows that in only 4.73% of the 444 (37 times 12) cases the value of ai equals 0.00 (expert forecasts only), and that in only 5.86% of the cases it equals 1.00 (model forecasts only). This strongly confirms the common finding that combined forecasts are more accurate than their individual components. Here, in 89.41% of the cases model forecasts combined with expert forecasts yield improvement. 7 When we compute the average of the optimal weights, we get values around 0.50 (see the last row of Table 4), with a slight tendency to increase with increasing forecast horizon. This suggests that the relative weight of the model increases with the horizon. Hence, the unconditional weights 50%-50%, as suggested by Blattberg and Hoch (1990), seem to be a good choice indeed, at least, unconditionally. Optimal weights and experts A further impression from the numbers in Table 4 is that there is substantial variation across the experts, and hence it seems worthwhile to examine whether the optimal weights can be explained by experts’ characteristics and their behaviour. The conditional model that we use for this purpose is (3) ii iii iiiii deviationno directionwrongfemaleproductsofnumber ageagepositionpositionaoptimal       _ ___ _ 8 765 2 43 2 210 where we estimate the parameters using OLS. As the sample size is only 37, we rely on a 10% significance level. The optimal ai value per expert follows from Table 4. The variables wrong_direction and no_deviation are fractions of the total amount of expert forecasts which are on the wrong side of the model forecast relative to the realization and which do not differ from the model forecasts, respectively. In the first round we estimate all parameters, and then subsequently delete the least significant ones, until we have at least 10% significant parameters. Before we estimate the parameters, we formulate some prior thoughts on the possible relevance and sign of the parameters in (3). We measure the experience of an expert by the number of years he or she is in that particular position and by his or her age. The results in Lamont (2002) suggest that the effect of experience is positive for the quality of the expert, which here means that the parameters β1 and β3 would have a negative sign (giving smaller values of the optimal ai and hence more weight to the expert. Lamont (2002) also documents that much younger or much older experts perform not as good as medium-aged experts, and hence we expect that β2 and β4 are positive. Additionally, we include the variable that counts the number of products an expert has to deal with as a measure of experience. We expect β5 to 8 be negative too. The studies in Barber and Odean (2001) and Beyer and Bowden (1997) suggest that female experts have a lower tendency to be overconfident, and hence might deviate less from the model, and this may also give the model more weight, so β6 is expected to be positive. On the other hand, female experts may quote forecasts that differ less from the model forecasts, and this in turn may lead to more weight of the model, and then β6 would be negative. Concerning the actual behaviour of experts, we include two variables in (3). Deviating from the model in the wrong way would of course lead to less weight of the expert in the combined forecast, so we expect β7 to be positive. Also, more cases with no deviation would make the model more relevant, and so we expect β8 to be positive. As far as forecast horizons are concerned we have no particular prior hypotheses, except perhaps that, based on Table 2, a smaller deviation of the expert forecast from the model forecast might be beneficial to the weight of the expert forecast. This would mean that the parameter β8 becomes more relevant for further away horizons. Insert Table 5 about here A first immediate conclusion that can be drawn from Table 5, where we only report on the 10% significant parameters, is that the variables position, position 2 , female and number of products do not matter at all. Further, the results in Table 5 indicate that the horizon matters to fit the conditional mean of the optimal value of a. For the short term horizons like 1 to 5 experience matters while for further-away horizons the degree of no deviation matters. Also, for some horizons, the degree of wrong signed deviations influences the optimal value of a. We obtain the expected sign for experience, for experts’ forecasts on the wrong side of the model forecasts and for the degree of no deviation. Forecasts of older experts have more weight in the optimal combined forecast, and, as the squared variable is significant too, too young or too old gives less weight. This estimated quadratic effect clearly supports the findings in Lamont (2002). The age which gives the minimum values of optimal ai, and hence gives most weight to the expert, is around 40 years. Interestingly, experience does not seem to matter much for further-away horizons. For horizons 6, 7, 8 and 10, we find that the optimal weight cannot be predicted by the explanatory variables used in this paper and hence the best predictor is the unconditional mean. Looking at the standard errors for the intercept parameter, we see that 0.50 is within the 90% confidence bounds, which supports the claim in Blattberg and Hoch (1990). 9 Finally, for further-away horizons we see that the degree of no deviation becomes relatively more important when gaining weight for the expert in the combined forecast. This suggests that the degree in which the expert forecast does not differ from the model forecast is indeed relevant. Given that we find a useful prediction model for the optimal weight for various horizons, we conclude that the unconditional 50%-50% rule can sometimes better be replaced by a conditional rule. This insight adds to the insights given in Blattberg and Hoch (1990), and is new to the literature. Bracketing Theory predicts that bracketing, that is, model forecasts and expert forecasts each are on one side of the realization, makes combining forecasts a fruitful exercise. More so, as is argued in Larrick and Soll (2006), if the rate of bracketing increases, the power of simply taking averages increases. This argument rests on the assumption that the location on both sides of the realization of both the model forecast and the expert forecast obeys a uniform distribution. So, if MF is on one side of S and EF is on the other side, and the location of MF and EF is uniformly distributed, meaning that it does not occur that say MF is always closer to S that EF is, then on average MF and EF are equally close to S, and in that case the 50-50 rule should be optimal. To examine this conjecture for actual data and not for experiments as in the study of Larrick and Soll (2006), we run the following simple regression, that is, (4) iii bracketingaoptimal   10 2 )5.0_( where the explanatory variable is the fraction of forecasts that bracket the realization. We argue that when the conjecture is valid, that then β1 in (4) is significant and negative. The relevant estimation results are displayed in Table 6. Insert Table 6 about here Similar to the results in Table 5, we observe that for intermediate horizons the distribution of the optimal value of ai is hard to predict (see the low fit values for the horizons 10 6 to 10 in the last column of Table 6). On the other hand, for horizons 1 to 5 and for 11 and 12, there is strong evidence that the difference between the optimal ai and 0.5 gets smaller for higher rates of bracketing. So, bracketing makes simple averaging more powerful. 4. Discussion Our paper analyzed a very large and unique database with model forecasts and expert forecasts to see if combining these forecasts would be beneficial. Blattberg and Hoch (1990) predicted that unconditional weights of 50%-50% would be best. One of the novelties of our study is that we examined if these weights could be predicted by experts’ characteristics and actual behaviour or performance, that is, whether there are perhaps conditional weights. Main findings We first documented convincingly that the model forecasts are unbiased, at least on average. This is very important as if that would not be the case, all subsequent analysis should have been modified. Evidence indicated that model forecasts are indeed unbiased, and also that experts have a tendency to deliver forecasts that exceed model forecasts in particular for nearby forecast horizons, see also Fildes et al. (2009). Additionally, we documented that the fraction of bracketing is substantially smaller than the fraction of expert forecasts being on the wrong side of the model and of reality. In fact, bracketing occurs in only around one-fourth of all cases. When we computed the optimal weights of combining model forecasts and expert forecasts, we found that the unconditional weights (across horizons and on average) are indeed close to a 50%-50%, but that there also is a strong variation across the experts. In fact, we showed that combined forecasts improve on the component forecasts in about 90% of the (large amount of) cases. Next, the optimal weights were shown to depend on experience (age), the degree of wrong-signed expert forecasts, and the degree of no deviation from the model forecast in the hypothesized way for various forecast horizons. Hence the unconditional 50%- 50% rule can be improved by including experts’ characteristics and actual behaviour. Finally, we found that more bracketing leads to more indication that the 50%-50% rule is optimal. 11 Managerial implications Our findings have various managerial implications. The first is that it is almost always best to combine model forecasts and expert forecasts. When the manager has no information on what the expert does or who he or she is, then the unconditional weights 50%-50% seem to be the best choice. However, when experts more often take extreme positions (on the wrong side of the model forecast) or deviate too much or have less experience the weights of the model forecast in the combination should be higher. On the other hand, when experts’ forecasts and model forecasts would bracket the realizations, then the 50%-50% becomes more useful again. When training new experts it is important to inform them that bracketing makes their contribution more relevant and that taking extreme positions (that is, the wrong side of the model forecast) does not. What also would help is to demonstrate to these experts that the model forecasts are in general unbiased, and hence that persistently quoting above or below a model forecast simply cannot be appropriate. When hiring new experts it makes sense to ask for their past credentials in terms of their forecasts relative to the model forecasts. Note that simply choosing for the expert or for the model because the associated RMSPE is smaller than that of the other is not the best strategy, as we have seen that combined forecasts are almost always better. So, their degree of bracketing matters, and as we saw, their experience does too. Literature suggests that too novice or too established experts have a tendency to take more extreme positions, and our findings suggest that this makes their relative contribution in a combined forecast smaller, at least for nearby horizons. 12 Table 1: The differences between the model forecasts, the expert forecasts and the corresponding realization, measured in fractions across all cases Ji and I Horizon Deviation in Bracketing wrong direction ------------------------------ ------------------------------ MF<S<EF MF>S>EF MF<EF<S MF>EF>S EF<MF<S EF>MF>S 1 0.162 0.100 0.168 0.166 0.133 0.229 2 0.164 0.102 0.162 0.166 0.132 0.232 3 0.164 0.108 0.160 0.168 0.130 0.226 4 0.165 0.108 0.164 0.166 0.133 0.221 5 0.162 0.108 0.159 0.167 0.139 0.220 6 0.160 0.109 0.163 0.171 0.133 0.216 7 0.163 0.115 0.157 0.173 0.136 0.208 8 0.165 0.116 0.152 0.171 0.141 0.205 9 0.160 0.118 0.151 0.174 0.140 0.207 10 0.160 0.122 0.152 0.173 0.140 0.203 11 0.160 0.127 0.153 0.166 0.146 0.196 12 0.154 0.129 0.160 0.173 0.144 0.187 13 Table 2: Expert forecasts relative to model forecasts and model forecasts relative to realizations, measured in fractions across all cases Ji and I Horizon EF > MF MF > S 1 0.559 0.495 2 0.557 0.500 3 0.550 0.502 4 0.549 0.495 5 0.541 0.495 6 0.539 0.497 7 0.527 0.496 8 0.521 0.492 9 0.517 0.499 10 0.516 0.498 11 0.509 0.488 12 0.502 0.489 14 Table 3: Characteristics of the experts (mean values if there are two experts) (and sample average) Expert Age (years) Gender (1 = female) Position (years) 1 30 0 10 2 37.5 0 15 3 55 1 15 4 40 0 5 5 35 1 5 6 30 0 5 7 47.5 0 17.5 8 45 0 10 9 60 0 20 10 30 1 10 11 20 1 1 12 45 1 5 13 35 1 5 14 40 1 15 15 45 1 5 16 50 0 15 17 40 1 3 18 40 1 1 19 30 0 5 20 45 0.5 10 21 35 1 5 22 25 1 2 23 50 0 20 24 32.5 0 7.5 25 45 0 10 26 35 0 10 27 45 0 7 28 35 0 3 29 35 1 5 30 30 1 3 31 55 0 10 32 35 0 5 33 60 1 15 34 30 0.5 4 35 55 1 20 36 30 1 5 37 30 0 2 Mean 40.07 0.47 8.61 15 Table 4: Optimal weights in the combined forecast, that is, thtjiithtjii EFaMFa |,,|,, )1(   , when aggregated across products Expert Horizon 1 2 3 4 5 6 7 8 9 10 11 12 1 0.00 0.00 0.00 0.00 0.00 0.25 0.60 0.65 0.75 0.65 0.85 0.85 2 0.20 0.20 0.35 0.25 0.30 0.30 0.15 0.45 0.50 0.50 0.55 0.50 3 0.65 0.65 0.70 0.55 0.45 0.40 0.35 0.35 0.30 0.30 0.25 0.25 4 0.25 0.50 0.40 0.20 0.20 0.20 0.70 0.65 0.35 0.45 0.50 0.35 5 0.20 0.20 0.30 0.50 0.75 0.70 0.65 0.70 0.80 0.80 0.95 0.90 6 0.65 0.35 0.30 0.30 0.10 0.20 0.30 0.15 0.25 0.15 0.05 0.00 7 0.00 0.05 0.25 0.05 0.00 0.05 0.25 0.30 0.35 0.25 0.50 0.45 8 0.55 0.60 0.50 0.50 0.45 0.45 0.45 0.45 0.45 0.45 0.45 0.45 9 0.65 0.95 0.95 0.90 0.90 0.85 0.55 0.55 0.50 0.50 0.50 0.35 10 0.55 0.40 0.45 0.40 0.40 0.45 0.45 0.50 0.50 0.50 0.50 0.55 11 0.90 0.90 0.90 0.90 0.90 0.90 0.95 0.95 0.90 0.95 1.00 1.00 12 0.75 0.65 0.85 0.90 0.85 0.65 0.75 0.75 0.70 0.65 0.55 0.40 13 0.10 0.15 0.15 0.25 0.15 0.20 0.40 0.35 0.40 0.35 0.05 0.05 14 0.65 0.65 0.55 0.45 0.40 0.40 0.35 0.35 0.30 0.35 0.40 0.35 15 0.45 0.45 0.45 0.45 0.55 0.55 0.45 0.40 0.65 0.75 0.70 0.70 16 0.35 0.35 0.30 0.30 0.30 0.30 0.35 0.35 0.40 0.45 0.50 0.45 17 0.45 0.50 0.45 0.40 0.45 0.40 0.35 0.35 0.30 0.35 0.35 0.20 18 0.25 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.20 0.15 0.20 0.65 19 0.85 1.00 0.60 0.60 0.45 0.05 0.00 0.30 0.15 0.25 0.40 0.75 20 0.65 0.55 0.55 0.45 0.50 0.60 0.65 0.60 0.65 0.70 0.75 0.55 21 0.80 0.85 0.80 0.95 0.95 0.90 0.95 1.00 0.95 1.00 0.80 0.95 22 0.50 0.40 0.35 0.20 0.90 0.10 0.15 0.10 0.20 0.10 0.10 0.10 23 0.75 0.75 0.75 0.75 0.75 0.75 0.75 0.70 0.70 0.70 0.70 0.70 24 0.70 0.65 0.70 0.60 0.55 0.60 0.60 0.60 0.70 0.75 0.70 0.55 25 0.55 0.70 0.70 0.60 0.65 0.75 0.75 0.85 0.80 0.80 0.80 0.85 26 0.25 0.50 0.45 0.35 0.25 0.25 0.40 0.45 0.10 0.10 0.05 0.05 27 1.00 1.00 1.00 1.00 1.00 0.80 0.65 0.65 0.45 0.60 0.80 0.75 28 0.45 0.35 0.50 0.45 0.55 0.55 0.60 0.65 0.50 0.50 0.50 0.40 29 0.50 0.40 0.15 0.20 0.45 0.55 0.50 0.70 0.55 0.65 0.85 0.90 30 0.50 0.65 0.75 0.70 0.70 0.65 0.70 0.65 0.60 0.60 0.50 0.55 31 0.40 0.05 0.00 0.00 0.00 0.35 0.50 0.55 0.50 0.10 0.00 0.00 32 0.50 0.50 0.50 0.45 0.50 0.50 0.50 0.55 0.55 0.50 0.50 0.45 33 0.55 0.60 0.65 0.70 0.70 0.75 0.75 0.75 0.75 0.75 0.80 0.70 34 1.00 0.90 0.95 1.00 1.00 0.95 0.60 0.55 0.90 1.00 1.00 1.00 35 0.30 0.40 0.65 0.90 0.00 0.15 0.30 0.65 0.55 1.00 0.25 0.25 36 0.20 0.30 0.50 0.40 0.50 0.60 0.40 0.65 0.60 0.80 0.50 0.80 37 0.95 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.90 0.90 Mean 0.51 0.52 0.52 0.50 0.50 0.49 0.51 0.55 0.54 0.55 0.53 0.53 16 Table 5: Estimation results for model (3), where insignificant parameters are deleted sequentially (using a 10% significance level). Parameters have been estimated by OLS, with a White-type correction for potential heteroskedasticity Horizon Variable (parameter and standard error in parentheses) Fit Intercept Age Age^2 Wrong No Direction Deviation 1 0.580 -0.044 0.0005 2.248 0.246 (0.574) (0.021) (0.0002) (0.846) 2 0.727 -0.043 0.0005 1.724 0.158 (0.671) (0.026) (0.0003) (0.984) 3 1.617 -0.057 0.0007 0.079 (0.579) (0.028) (0.0003) 4 1.668 -0.062 0.0008 0.082 (0.654) (0.031) (0.0004) 5 1.115 -0.058 0.0007 1.609 0.199 (0.729) (0.030) (0.0004) (0.736) 6 0.489 0.000 (0.046) 7 0.508 0.000 (0.040) 8 0.547 0.000 (0.037) 9 0.508 0.547 0.063 (0.040) (0.297) 10 0.553 0.000 (0.045) 11 0.502 0.614 0.062 (0.051) (0.204) 12 -0.102 1.690 1.398 0.174 (0.304) (0.840) (0.369) =================================================================== 17 Table 6: Estimation results for model (4). Parameters have been estimated by OLS, with a White-type correction for potential heteroskedasticity (standard errors are in parentheses). The sample size is 36. The data on expert 11 are not included as they amount to an outlier. Horizon Intercept Fraction of bracketing Fit 1 0.191 (0.053) -0.486 (0.178) 0.141 2 0.229 (0.055) -0.575 (0.195) 0.164 3 0.192 (0.063) -0.442 (0.212) 0.104 4 0.216 (0.061) -0.489 (0.210) 0.115 5 0.246 (0.057) -0.573 (0.198) 0.149 6 0.105 (0.031) -0.127 (0.093) 0.014 7 0.035 (0.032) 0.065 (0.100) 0.004 8 0.030 (0.037) 0.069 (0.126) 0.005 9 0.055 (0.031) -0.015 (0.103) 0.003 10 0.110 (0.046) -0.138 (0.150) 0.017 11 0.163 (0.046) -0.317 (0.151) 0.095 12 0.174 (0.044) -0.334 (0.132) 0.107 18 References Barber, B. and T. Odean (2001), Boys will be boys: Gender, overconfidence, and common stock investment, Quarterly Journal of Economics, 116, 261-292. Beyer, S. and E. Bowden (1997), Gender differences in self-perceptions: Convergent evidence from three measures of accuracy and bias, Personality and Social Psychology Bulletin, 23, 157-172. Blattberg, Robert C. and Stephen J. Hoch (1990), Database models and managerial intuition: 50% model + 50% manager, Management Science, 36, 887-899. Clemen, Robert T. (1989), Combining forecasts: A review and annotated bibliography (with discussion), International Journal of Forecasting 5, 559-583. Fildes, R., P. Goodwin, M. Lawrence and K. Nikopoulos (2009), Effective forecasting and judgmental adjustments: an empirical evaluation and strategies for improvement in supply- chain planning, International Journal of Forecasting, 25, 3-23. Goodwin, P. (2000), Improving the voluntary integration of statistical forecasts and judgement, International Journal of Forecasting, 16, 85-99. Goodwin, P. (2002), Integrating management judgement with statistical methods to improve short-term forecasts, Omega, 30, 127-135. Lamont, O.A. (2002), Macroeconomic forecasts and microeconomic forecasters, Journal of Economic Behavior & Organization, 48, 265-280. Larrick, Richard P. and Jack B. Soll (2006), Intuitions about combining opinions: Mis- appreciation of the averaging principle, Management Science, 52, 111-127. Lawrence, M., P. Goodwin, M. O’Connor and D. Onkal (2006), Judgemental forecasting: A review of progress over the last 25 years, International Journal of Forecasting, 22, 493-518. 19 Timmermann, Allan (2006), Forecast combinations, Chapter 4 in Graham Elliott, Clive W.J. Granger and Allan Timmermann (eds.), Handbook of Economic Forecasting Volume I, Amsterdam: Elsevier, pp 135-196. 
work_rh3fbbbqgrfnbiovibtnrvibyy	----	NASA Technical Reports Server (NTRS) 
work_rhgi3yz4jfaq7ev6lbeirq2c7u	----	Microsoft Word - 08001 Report production coordinator: Suan Choi, suan@comp.hkbu.edu.hk, +852 3411 7079 Report Web Site: http://www.comp.hkbu.edu.hk/tech-report Report Cover Designer: So Kin Ming 計算機科學系 Department of Computer Science Self-Organized Combinatorial Optimization Jiming Liu, Yu-Wang Chen, Yong-Zai Lu, Gen-Ke Yang COMP-08-001 Release Date: 8 May 2008 Department of Computer Science, Hong Kong Baptist University Abstract In this paper, we present a self-organized computing approach to solving hard combinatorial optimization problems, e.g., the traveling salesman problem (TSP). First of all, we provide an analytical characterization of such an approach, by means of formulating combinatorial optimization problems into autonomous multi-entity systems and thereafter examining the microscopic characteristics of optimal solutions with respect to discrete state variables and local fitness functions. Next, we analyze the complexity of searching in the solution space based on the representation of fitness network and the observation of phase transition. In the second part of the paper, following the analytical characterization, we describe a decentralized, self-organized algorithm for solving combinatorial optimization problems. The validation results obtained by testing on a set of benchmark TSP instances have demonstrated the effectiveness and efficiency of the proposed algorithm. The link established between the microscopic characterization of hard computational systems and the design of self-organized computing methods provides a new way of studying and tackling hard combinatorial optimization problems. Self-Organized Combinatorial Optimization Jiming Liu, Yu-Wang Chen, Yong-Zai Lu, Gen-Ke Yang Department of Computer Science, Hong Kong Baptist University, Kowloon Tong, Hong Kong School of Electronic, Information and Electrical Engineering, Shanghai Jiaotong University, Shanghai 200240, China {jiming,ywchen}@comp.hkbu.edu.hk Abstract In this paper, we present a self-organized computing approach to solving hard combinatorial optimization problems, e.g., the traveling salesman problem (TSP). First of all, we provide an analytical characterization of such an approach, by means of formulating combinatorial optimization problems into autonomous multi-entity systems and thereafter examining the microscopic characteristics of optimal solutions with respect to discrete state variables and local fitness functions. Next, we analyze the complexity of searching in the solution space based on the representation of fitness network and the observation of phase transition. In the second part of the paper, following the analytical characterization, we describe a decentralized, self-organized algorithm for solving combinatorial optimization problems. The validation results obtained by testing on a set of benchmark TSP instances have demonstrated the effectiveness and efficiency of the proposed algorithm. The link established between the microscopic characterization of hard computational systems and the design of self-organized computing methods provides a new way of studying and tackling hard combinatorial optimization problems. Keywords: Combinatorial optimization; Traveling salesman problem; NP-complete; Microscopic characteristics; Fitness network; Phase transition; Self-organized computing 1. Introduction Combinatorial optimization is pervasive in most fields of science and engineering. Its aim is to optimize an objective function on a finite set of feasible solutions. For example, the traveling salesman problem (TSP) (Gutin and Punnen 2002), one of the classical combinatorial optimization problems, can be described as a problem of search for the shortest tour among a set of cities. Generally speaking, most of combinatorial optimization problems in practice are deemed as computationally intractable, and have been proven to belong to the class of NP-complete problems (Garey and Jonhson 1979), where NP stands for “Non-deterministic Polynomial time”. For NP-complete problems, although the optimality of a possible solution can be verified in polynomial time, the computational time for finding the optimal solution grows exponentially with the dimension of input variables in the worst case. Furthermore, if we can find a polynomial time algorithm to solve one NP-complete problem, then all the other NP-complete problems will become solvable in polynomial time (Papadimitriou, 1994). However, in modern computer science, it has been commonly conjectured that there are no such polynomial time algorithms for any NP-complete problems (Cormen et al. 2001). Alternatively, a variety of nature-inspired optimization techniques have been developed for finding near-optimal solutions of NP-complete problems within a reasonable computational time. Examples of such algorithms include: simulated annealing (Kirkpatrick, Gelatt Jr., and Vecchi 1983), genetic algorithm (Forrest 1993), ant colony optimization (Bonabeau, Dorigo, and Theraulaz 2000), particle swarm optimization (Kennedy and Eberhart 1995), among others. It is interesting to note that most of the existing optimization methods employ a centralized control model, and rely on a global objective function for evaluating intermediate and final solutions. In dealing with hard combinatorial optimization problems, however, it would be difficult and time-consuming to collect the global information due to the interaction in, and the dimensionality of, computation. In order to overcome the limitation of existing centralized optimization, several decentralized, self-organized computing methods have been studied in recent years with the help of complex systems and complexity science. Boettcher and Percus (2000) presented a stochastic search method called extremal optimization (EO). The method is motivated by the Bak-Sneppen evolution model (Bak and Sneppen 1993; Sneppen 1995), in which the least adapted species are repeatedly mutated following some local rules. To further improve the adaptability and performance of the EO algorithm, a variation of EO called τ-EO (Boettcher and Percus 2000) was subsequently presented by introducing a tunable parameter τ. So far, EO and its variants have successfully addressed several physical systems with two-degree-of-freedom entities, i.e., the combinatorial optimization problems with binary state variables, such as graph bi-partitioning problems (Boettcher, 2000), Ising spin glasses (Middleton, 2004), community detection in complex networks (Duch and Arenas 2005), etc. Recently, Liu et al. (2005; 2006) presented a general autonomy-oriented computing (AOC) framework for the optimization of self-organized distributed autonomous agents. AOC is also a bottom-up computing paradigm for solving hard computational problems and for characterizing complex systems behavior. Computational approaches based on AOC systems have been applied to distributed constraint satisfaction problem (Liu, Han, and Tang 2002), image feature extraction (Liu, Tang, and Cao 1997), network community mining problem (Yang and Liu 2007), etc. In addition, Han (2003) introduced the concept of local fitness function for evaluating the state of autonomous agents in computational systems. In order to design a reasonable self-organized computing method, there are generally several important issues that should be considered: 1. How can we construct a mapping from combinatorial optimization problems to multi-entity complex systems? 2. What types of local fitness function should be used to guide the computational systems to evolve towards the global optimum? 3. What are the microscopic characteristics of optimal solutions? 4. How can we characterize the computational complexity of such a method in searching a solution space? A number of literatures have been focused on one or some of the above questions over past few years (Mézard, Parisi, and Zecchina 2002; Han and Cai 2003; Goles et al. 2004; Achlioptas, Naor, and Peres 2005). In this paper, we attempt to answer each of those questions, and furthermore, provide a solid theoretical foundation for the self-organized computing method under study. First of all, we will model combinatorial optimization problems into multi-entity systems, in which a large number of self-organizing, interacting agents are involved. We then statistically examine the microscopic characteristics of optimal solutions with respect to the notions of discrete state variable and local fitness function. Moreover, we will analyze the complexity of search in a solution space based on the representation of fitness network and the observation of phase transition. Finally, based on the analysis, we will describe a self-organized computing algorithm for solving hard combinatorial optimization problems. The rest of the paper is organized as follows: In Section 2, we describe the mapping between combinatorial optimization problems and multi-entity systems, and then provide the analytical results of microscopic distribution on the optimal TSP solutions. In order to systematically analyze the solution space, we present the concept of fitness network, and further discuss the search complexity from the characteristics of phase transition. In Section 3, we present a self-organized combinatorial optimization method, as well as the validation results on a set of benchmark TSP instances. In Section 4, we conclude this paper. 2. Analytic characterization of combinatorial optimization problems In this section, we present an analytic characterization of combinatorial optimization from the point of view of a self-organizing system. The characterization will provide us with insights into the design of a decentralized, self-organized computing approach to tackling combinatorial optimization problems. In particular, it points out ways of improving the efficiency of solution searching based on such an approach. 2.1. Modeling combinatorial optimization problems into multi-entity systems Combinatorial optimization is concerned with finding the optimal combination of a set of discrete constrained variables (Papadimitriou and Steiglitz 1998). As summarized by Blum and Roli (2003), a combinatorial optimization problem, P = (S, F), can be defined in terms of: a set of discrete variables, X = {x1, …, xn}, and variable domains, D1, …, Dn; multiple constraints among variables; an objective function F to be optimized, where F: D1×···×Dn → R. Correspondingly, the solution space can be represented as: S = {s = (x1, …, xn) | xi ∈ Di, s satisfies all the constraints}, in which each element s is a candidate solution. Given a combinatorial optimization problem, the aim is to find the optimal solution, s*∈S, such that the global objective, F(s*) ≤ F(s) (∀s ∈ S). Since combinatorial solutions often depend on a non-trivial combination of multiple elements with specific states, a combinatorial optimization problem can be straightforwardly translated into a multi-entity computational system, if we formulate it from the viewpoint of a complex system. In a complex computational system, each entity can be viewed as an autonomous agent, and the agent can collect local (limited) information from the environment, and acts with other interacting agents for achieve the system objective (Liu, Jin, and Tsui 2005). Specifically, an autonomous computational agent, a, can be characterized based on the following properties (Liu, Han, and Tang 2002): a finite number of possible states and some behavioral rules; local information about the computational system and its conditions; certain local objectives. Thus, a multi-entity computational system can be further modeled based on: a group of agents, A = {a1, …, an}; interactions between agents; a set of reactive rules, governing the interactions among agents. Drawing on the above intuitive idea of mapping between the representation of multi-entity computational systems and the definition of combinatorial optimization, various methodologies for complex systems modeling and analysis can be explored and applied to handle combinatorial computing problems. Generally speaking, the modeling and analysis techniques for complex systems can be classified into top-down and bottom-up approaches (Liu, Jin, and Tsui 2005). Top-down approaches follow the idea of centralized control, and mainly study the macroscopic system-level characteristics of a complex system. On the contrary, bottom-up approaches, which were developed more recently, employ the concept of decentralized control, and apply the entity-oriented models to characterize the microscopic emergent properties. In order to address the complexity of practical computational systems, we will, in the paper, focus on the study of a novel decentralized, self-organized computing method, as opposed to earlier work on centralized combinatorial optimization. 2.2. Local fitness function In a multi-entity computational system, the process of self-organization in individual entities is crucial for searching the global optimal solution of the whole system. At any moment, each entity must assess its local conditions, and thus decide its primitive behavior. In doing so, a local fitness function (here, the term “fitness function” is synonymous with objective function) is essential for evaluating the specific states of local entities. For the sake of explanation and illustration, in what follows we will take the TSP as a walk-through example of combinatorial optimization problems. The TSP has attracted much interest in the computing communities in the past decades due to the fact that it provides insights into a wide range of theoretical questions in discrete mathematics, theoretical computer science, computational biology, among others, and offers models and solutions to many real-world practical problems, ranging from scheduling, transportation, to genome mapping. The TSP is usually stated as an optimization problem of finding the shortest closed tour that visits each city only once. In other words, given a set of n cities and the distance measure, dij, for all pairs i and j, the optimization goal is to find a permutation π of these n cities that minimizes the following global fitness function: ∑ − = + += 1 1 )1(),()1(),()( n i nii ddsF ππππ (1) For a TSP solution s in the solution space S, it can be exclusively represented by the discrete states of all cities. In this paper, if city i, which can be represented as entity xi, is connected to its kth (1 ≤ k ≤ n–1) nearest neighbor, the discrete state variable of city i is defined as: 1,,2,1,)( −== nikxs i L (2) In an ideal computational system, all cities will be connected to their first nearest neighbors, i.e., s(xi) = 1 for all i (1 ≤ i ≤ n). However, such ideal microstates are often frustrated by the interacting effects between cities, i.e., competitions on the selection of forward-directed edges. As a result, a city may not always be connected to its first nearest neighbor, and possibly dwell on a specific energy state except the ground state, i.e., s(xi) > 1. Correspondingly, a local fitness function can be formulated to evaluate the microscopic dynamical characteristics of all cities. Now let us assume di to be the length of the forward-directed edge starting from city i in a feasible solution s, the local fitness of city i can be defined as: niddxf ijijiis ,,2,1,min)( L=−= ≠ (3) Correspondingly, the global fitness function for any possible solution s can be represented as: ∑∑ == ≠ += n i is n i ijij xfdsF 11 )(min)( (4) Obviously, the first term in Eq. (4) is a constant, which can serve as a lower bound for a given TSP instance. The second term is the sum of the local fitness for all entities. That is to say, the global fitness of a combinatorial solution can be represented as a function of distributed local fitness, and the optimal solution is equivalent to the microstate with the minimum sum of local fitness for all entities. Intuitively, we can optimize the computational system through updating the states of those entities with worse local fitness. In a complex computational system, the relationship between discrete states and local fitness is usually nonlinear, and the value of local fitness is not necessarily equal even if two entities have the same state value. Since local fitness is not as explicit as global fitness, as discussed in our earlier work (Chen and Lu 2007), the following questions will inevitably arise: 1. Can the optimization guided by local fitness achieve the global optimum of computational systems? 2. What will be the relationship between local fitness and global fitness? For answering the above questions, in the following we introduce the consistency and equivalence conditions between global fitness and local fitness. Definition 1. Suppose that solution s' is constructed from s (∀s ∈ S) by updating the state of entity i using a neighbor rule, local fitness is consistent with global fitness, iff it is true that )]}()([sgn{)}()'(sgn{ ),',( ' xfxfsFsF sissXx s −=− ∑ ∈ (5) where sgn{⋅} is the symbolic function, and X(s,s',i) denotes the set of entities whose states are changed as a result of the interacting effects of updating the state of entity i. If the neighbor rule is a reversible move class, and then X(s,s',i) = X(s',s,i), the solution space can be represented as an undirected graph, otherwise a directed graph. Definition 2. Local fitness is equivalent to global fitness iff (i) local fitness is consistent with global fitness and, (ii) ∃α ∈ R+, β∈ R such that it is true that for all s ∈ S, βα += ∑ ∈Xx s xfsF )()( (6) According to Eq. (5), it is evident that if local fitness is consistent with global fitness, improving one or some entities’ local fitness will also optimize the global fitness of the whole computational system, and the process of self-organization will be effective in solving hard computational problems. Equivalence is a special case of consistency. These definitions make it easier for us to understand what types of local fitness function should be defined in designing self-organized computing methods. 2.3. Microscopic analysis of optimal solutions In a computational system, searching the optimal microstate will no doubt become simpler if we have a definite object in view. Therefore, in this section we will focus on the microscopic analysis of optimal solutions based on the preceding definitions of discrete state variable and local fitness function. As we know, the optimal solutions for a large number of benchmark TSP problems have been given in Reinelt’s TSPLIB (Reinelt 1991; TSPLIB). Therefore, for the sake of illustration and testing, we use the TSP instance, called kroA100 (100 cities) as an example. In order to find a specific TSP tour, we sequentially calculate the discrete state value and local fitness of each city. As a result, the microscopic characteristics of the optimal tour and a random tour can be found and have been comparatively shown in Fig. 1. 0 10 20 30 40 50 60 70 80 90 100 0 50 100 n S ta te Random Optimal 0 10 20 30 40 50 60 70 80 90 100 0 1000 2000 3000 4000 n Lo ca l f itn es s Random Optimal Fig. 1. Microscopic characteristics of the optimal tour and a random tour (kroA100). As can be noted in Fig. 1, in the optimal solution the values of the state variables for all cities as well as their local fitness are far lower than the upper bound of the variables, e.g., s(xi) << n–1 for all i (1 ≤ i ≤ n). This implies that those cities with high local fitness in an initial solution should update their states, until the computational system self-organizes itself into a well-organized microscopic state, i.e., the optimal solution. In order to demonstrate the universality of this microscopic distribution, here we further characterize the statistical properties of the optimal TSP solutions by the kth nearest neighbor distributions (NNDs) (Chen and Zhang 2006). Without loss of generality, in this paper the kth-NND for any possible TSP tour s is defined as follows: 1,,2,1, )( )( −== nk n kr kp L (7) where r(k) is the total number of the kth nearest neighbors for all forward-directed edges in a feasible tour, i.e., r(k) = |X(k)| if X(k) = {xi | s(xi) = k}. Obviously, p(k) ∈ [0, 1], and its sum will be equal to one. Fig. 2 shows the kth-NNDs p(k) of optimal tours on a set of 10 benchmark Euclidean TSP instances, with sizes ranging from 51 to 2392 nodes. 0 2 4 6 8 10 12 14 16 18 20 0 0.1 0.2 0.3 0.4 0.5 k p( k) eil51 pr76 kroA100 ch130 ch150 tsp225 a280 pcb442 pr1002 pr2392 Fig. 2. The kth nearest neighbor distributions p(k) of optimal tours from eil51 to pr2392. As shown in Fig. 2, each kth-NND p(k) is approximately an exponential decay function of neighboring rank k. Numerous experiments have indicated that the optimal solutions of almost all benchmark TSP instances in TSPLIB conform to this qualitative characterization of microscopic exponential distribution. It is worth pointing out that the above analytical results of microscopic distributions are very useful both to mathematicians and to computer scientists. From the viewpoint of mathematics, the maximum neighboring rank Kmax = max{s(xi)} for all i (1 ≤ i ≤ n) in the optimal solution is very useful to reduce the dimension of the state space. Taking pr2392 as an example, if we can approximately estimate it maximum neighboring rank Kmax = 20, the solution space represented by discrete state variables can be significantly reduced from (n–1)n to 20n, and thus a low-dimensional state space can be obtained before being fed into mathematical tools. From the viewpoint of computer science, these analysis results are also of great value on the design of an effective self-organized computing method. 2.4. Neighborhood and fitness network After unveiling the microscopic characteristics of optimal solutions, we can now further analyze the solution space of combinatorial optimization problems, and try to find a reasonable searching dynamics to reach the optimal solution. In a fundamental sense, the effectiveness of combinatorial optimization methods has a close relationship with the underlying neighborhood definition. The neighborhood mapping, also called move class, is usually used to construct better solutions from the current solution. For any possible solution s ∈ S, if N(s) ⊆ S is defined as a neighborhood function, all those feasible solutions in N(s) are called neighbors of the solution s. With respect to a specific neighborhood N(s), we call s' a strict locally-optimal solution if the objective function F(s') ≤ F(s), ∀ s ∈ N(s'). Obviously, the globally-optimal solution s*, F(s*) ≤ F(s), ∀ s ∈ S, belongs to the set of locally-optimal solutions. If we can find out all the locally-optimal solutions, the global optimum will immediately obtained. In order to analyze the computational complexity of search in a solution space, the concept of fitness landscape has been extensively applied to combinatorial optimization problems (Hordijk 1997; Altenberg 1997; Merz and Freisleben 2002; Reidys and Stadler 2002). Generally speaking, if the dimension of search space is δ, i.e., the cardinality of neighborhood |N(s)| = δ, by adding an extra dimension that represents the fitness of each solution, the dimension of fitness landscape will be equal to δ+1. Consequently, a hyper-dimensional landscape with peaks and valleys can be used to characterize the search space of optimization problems. Its ruggedness directly reflects the search complexity from an initial solution to the global optimum. If a fitness landscape is highly rugged, it will be difficult to search it for a better solution due to the lower correlation among neighboring solutions (Hordijk 1997; Merz 2004). To study the properties of fitness landscapes, Kauffman (1993) developed a problem-independent NK-model with a tunable parameter. In NK-model, the symbol N represents the number of entities of a computational system, e.g., cities in a TSP instance or genes in a genotype, and K reflects the degree of interactions among entities. More specifically, the local fitness contribution of each entity i is decided by the binary states of itself and K other interacting entities {i1,…, iK} ⊂ {1,…, i–1, i+1,…, N}. Thus, the global fitness of a solution s = x1, …, xN can be mathematically represented as follows: ∑ = = N i iiii K xxxf N sF 1 ),...,;( 1 )( 1 (8) The local fitness fi of entity i depends on its value xi and the values of K other entities Kii xx ,..., 1 . In simulation, the local fitness function fi: {0, 1} K+1 → IR assigns a random number from a uniform distribution between 0 and 1 to each of its inputs (Merz 2004). With this NK-model, the ruggedness of fitness landscapes can be tuned from smooth to rugged by increasing the value of K from 0 to N–1. When K is small, the global fitness difference between neighboring solutions will be relatively small, and heuristics can easily find better solutions using correlated gradient information. When K is large, a large number of entities between neighboring solutions will possess different states that will greatly influence the global fitness. When K = N–1, the landscape will be completely random, and the local fitness of all entities will be changed, even if we update only one entity’s state. In this extreme case, no algorithm is substantially more efficient than exhaustive search. However, in practice, almost all computational systems are probably not as complex as the worst case condition of K = N–1 in the NK-model, e.g., in the TSP the maximum neighboring rank Kmax << N. Therefore, it leaves us sufficient space to develop reasonable and efficient optimization methods, and hence to tackle those theoretically intractable computational problems. Although the notion of fitness landscape is useful to characterize the complexity of search space, it is impossible to visualize a solution space when the dimension of neighborhood is higher than 2. Thus, in this paper, we utilize the concept of fitness network to represent the structure of solution spaces. For a given combinatorial optimization problem, the fitness network can be defined by a triple (S, F, N): a set of nodes (solutions) s on solution space S; a fitness function F: s → R; a neighborhood function N(s), which decides the structure of fitness network. Thus, the fitness network can be interpreted as graph G = (V, E) with vertex set V = S and edge set E = {(s, s') ∈ S × S | s' ∈ N(s)}. If the move class is a reversible, i.e., if s' ∈ N(s) then also s ∈ N(s'), the fitness network is an undirected network; otherwise it is a directed network. On this complex network of search space, the out-degree of each nodes s is the cardinality of its neighborhood |N(s)|. Let us take the illustrative fitness network of Fig. 3 as an example. The hypothetical fitness network corresponds to a solution space with (0, 1) binary sequences of length 4. In this fitness network, we have the Hamming distances d(s, s') = 1 for all neighboring solutions s and s', the height in the vertical direction reflects the fitness value, i.e., the height of a node represents the fitness of the solution associated with it. Thus, the objective of optimization is to find the globally optimal solution ‘0011’ in the bottom of the fitness network. Fig. 3. A hypothetical fitness network on the Hamming graph of binary sequences of length 4. Based on the representation of fitness network, the optimization dynamics can be described in terms of a searching trajectory navigating through the network in order to find the lowest-fitness node. Therefore, a neighborhood-based algorithm for combinatorial optimization can be essentially characterized by the neighborhood structure N(s) and its searching dynamics in the fitness network. In order to discuss whether or not an algorithmic solution can obtain the optimal solution from an arbitrary initial solution, let us first consider some properties of the fitness network. Definition 3. In a fitness network, solution s' is reachable from s, if there exists at least a searching trajectory, s = s0, …, sk, sk＋1, …, sm = s', satisfying that sk＋1 ∈ N(sk), k = 0, …, m–1. If a node is reachable from another node, then there exists a set of consecutive nodes along a sequence of edges between them. Also, if node s2 is reachable from nodes s1, and node s3 is reachable from node s2, then it follows that node s3 is reachable from node s1. The definition of reachability is applicable to both directed and undirected fitness networks. Definition 4. A fitness network is called ergodic, if all other solutions are reachable from an arbitrary solution s (∀s ∈ S). Here, the term “ergodic”, as adopted from physics, refers to all microstates that are accessible. Theoretically speaking, a probabilistic searching method should guarantee that the optimal solution is reachable from its initial solution s. Since we usually construct initial solutions by non-deterministic methods, the ergodicity of a fitness network is a necessary condition for finding the optimal solution. In a fitness network as discussed above, local optima (such as solution ‘0101’ in Fig. 3) are undoubtedly barriers on the way to the optimal solution. Suppose that solution s1 and s2 are two local optima, and St denotes the solution set in a reachable trajectory from s1 to s2, the fitness barrier separating s1 and s2, as discussed by Reidys and Stadler (2002), can be defined as follows: } to from ies trajectorreachable :|]|)(min{max[],[ 2121 ssSSssFssF tt b ∈= (9) A critical point s ∈ S satisfying the mini-max condition (9) is called as a saddle point of the fitness network (e.g., solution ‘0001’ in Fig. 3). And then, the fitness barrier enclosing a local minimum s1 can be evaluated by the height of the lowest saddle point between s1 and a more favorable local minimum s2 (Kern 1993; Reidys and Stadler 2002). Mathematically, it can be represented as follows: )}()(:|)(],[min{)( 1221211 sFsFssFssFsFb <−= (10) This depth indirectly reflects the difficulty of escaping from the basins of attraction of the current local optimum s1. As a rule of thumb, the quality of local optimal solutions will be improved, if we increase the size of neighborhood. It can also be reflected by Eqs. (9) and (10) that an increased set of reachable trajectories may also decrease the depth of a local optimum. However, a larger neighborhood does not necessarily produce a more effective optimization method due to computational complexity, unless one can search the larger neighborhood in a more efficient manner. 2.5. Computational complexity and “phase transition” As mentioned before, NP-complete combinatorial optimization problems theoretically require exponential time to find a globally optimal solution. However, it is possible for us to solve a typical case in a more efficient manner. In order to characterize the typical-case computational complexity, the concept of phase transition in statistical physics has been successfully applied to the analysis of combinatorial optimization problems (Cheesman, Kanefsky, and Taylor 1991; Monasson 1999; Achlioptas, Naor, and Peres 2005). Phase transition refers to a drastic change in some macroscopic properties, when a control parameter crosses a critical value, e.g., water changing from ice (solid phase), to water (liquid phase), and to steam (gas phase) abruptly, as the temperature increases. Analogous to physical systems, it has been shown that many combinatorial decision problems also possess the phenomena of phase transition. Typical examples include: graph coloring problem (Cheesman, Kanefsky, and Taylor 1991), K-satisfiability problem (Martin, Monasson, and Zecchina 2001), traveling salesman problem (Gent and Walsh 1996; Zhang 2004), among others. As a manifestation of phase transition theory in combinatorial optimization, here we discuss how the fundamentals of phase transition can be used to statistically analyze the solution space of combinatorial optimization problems. As an illustration, we take a random two-dimensional Euclidean TSP instance for example, with placing n = 10 cities on a square of area A, and enumerate the length L of all (n–1)!/2 possible tours. For each solution, we calculate its dimensionless control parameter AnL ⋅/ (Gent and Walsh 1996). Fig. 4 shows the phase transition on the probability distribution of all dimensionless control parameters. In the figure, the cumulative distribution function directly reflects the probability that the solution space takes on solutions less than or equal to a specific control parameter. 0 0.5 1 1.5 2 2.5 3 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Control parameter P ro ba bi lit y Cumulative distribution function optimal Fig. 4. Phase transition on the cumulative distribution of dimensionless control parameters. Observing the phenomenon of phase transition as shown in Fig. 4, we can approximately divide the solution space into two regions: (i) one region is under-constrained among entities, in which the density of solutions is high, thus making us relatively easy to find a solution; (ii) another region is over-constrained, in which the probability of the existence of a solution is towards zero, and it is very difficult to obtain a near-optimal solution, especially the optimal solution, which is located on the boundary of phase transition. Since the cumulative distribution is analogous to Gaussian distribution, it means that most of candidate solutions locate in the middle-level of the fitness network. Based on the above theoretical modeling, simulation, and analysis, we can note that the computational complexity of combinatorial optimization problems always occurs on the boundary of phase transition, and therefore a majority of solution samplings in a searching trajectory should focus on the bottom of the fitness network. A reasonable and effective optimization process for addressing hard combinatorial optimization problems should consist of two types of steps: one for finding a new-optimal solution as efficiently as possible; and another for repeatedly escaping from one local optimum to another with enough fluctuated explorations. 3. Self-organized computing Following the line of analysis and discussions in the preceding section, in this section we present a self-organized computing method for solving combinatorial optimization problems. This optimization algorithm is inspired by the characteristics of collective emergent behavior in a dynamical system through the self-organization of its interacting entities. 3.1. Self-organized optimization algorithm In recent years, self-organization has been recognized as one of the most important research themes in studying the evolving process of complex dynamical systems. In self-organizing systems, global behaviors seem to emerge spontaneously from local interactions among entities. A simple self-organization model was proposed by Bak and Sneppen (1993) for studying biological evolution. In the Bak-Sneppen model, the evolution of a multi-entity system is driven by its extremal dynamics. At each iterative step, the entity with the worst fitness is selected to update its state, and simultaneously, the states of their neighbors will also be changed due to the triggered local interactions. After a number of update steps, the local fitness of almost all entities will transcend a threshold value, and this system will reach a highly correlated critical state, in which a little change of one entity may result in chain reactions and re-organize major parts of the complex system (Bak 1996). As we have analyzed in Sections 2.2 and 2.3, the combinatorial optimization solution can also be optimized by improving extremal local variables. Thus, the self-organized process driven by extremal dynamics can be used as an effective optimization dynamics. In the case of the TSP, the self-organized optimization algorithm under study can be given as follows: Self-organized Optimization Algorithm Step-1. Initialization Randomly construct an initial TSP tour s, which can be naturally represented as a permutation of n cities (path representation). Let sbest represent the best solution found so far, and set sbest ← s. Step-2. For the “current” solution s: (i) Calculate the local fitness fi according to Eq. (3) for each city i (1 ≤ i ≤ n); (ii) Determine the order of all cities {π(1), …, π(k), …, π(n)}, in which fπ(1) ≥ … ≥ fπ(k) … ≥ fπ(n), i.e., the rank k is going from k = 1 for the city with the maximum local fitness to k = n for the city with the minimum local fitness; (iii) Select a city π(k) with the kth “high” local fitness by a scale invariance probability distribution Pk ∝ k-α (1 ≤ k ≤ n), where α is an adjustable parameter; (iv) Update the state of the selected city π(k) by the 2-opt move class (Gutin and Punnen 2002), and then a neighborhood N(s) of the current solution s is constructed; (v) Calculate the global fitness F(s) of all solutions in the neighborhood N(s) and sequence all solutions by the descending order of global fitness. If the best solution s' in the neighborhood is better than the current solution, i.e., F(s')−F(s) < 0, accept s ← s', and set sbest ← s'; Otherwise, selecting a neighbor s' having the lth “low” global fitness by another power-law distribution Pl ∝ l-β (1 ≤ l ≤ |N(s)|) to replace the current solution, where β is another power parameter, and |N(s)| is the cardinality of the neighbor N(s). Step-3. Repeat at Step-2 until the termination criterion is satisfied. Step-4. Return sbest and F(sbest). Generally speaking, the power parameter α is used to control the self-organized optimization process. For α = 0, the updated entities are randomly selected, and for α → ∞, only the extremal entity with the worst local fitness is forced to change at each step. Numerous experiments have shown that a value of α ≈ 1+1/ln(n) seems to provide the best results on self-organized systems (Boettcher and Percus 2003; Chen, Lu, and Chen 2007). Another parameter β is used to adjust the hill-climbing ability on the basin of local optima. It is worth noting that calculating the global fitness of a solution at Step-2 (v) can be done more efficiently by catching and computing those entities being updated rather than computing all of them from scratch. Usually, a predefined iterative number or CPU time can be selected as a termination criterion. By applying the above proposed self-organized optimization algorithm to the TSP instance of kroA100, where the parameters α = 1.25, β = 2.75, the typical searching dynamics of self-organized optimization algorithm has been plotted in Fig.5. 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 2 2.5 3 3.5 4 4.5 5 x 10 4 Number of updates G lo ba l f itn es s 0 200 400 600 800 1000 5 10 15 x 10 4 2000 4000 6000 8000 10000 2.1 2.2 2.3 2.4 2.5 x 10 4 SOA Optimal Fig. 5. The searching dynamics produced by the proposed self-organized optimization algorithm. It can be seen from Fig. 5 that a near-optimal solution can be obtained sufficiently fast (as shown in the inset of Fig. 5) by the initial greedy self-organizing process, and then the enough fluctuated search near the global optimum can help to escape from the basin of the current local optimum and explore new local optima in the fitness network. The extent of fluctuations, which reflects the ability of hill-climbing near the bottom of fitness network, can be adjusted by the parameter β. As we increase the value of β, the fluctuations become gradually reduced. While β → ∞, there are no fluctuations, and the algorithm will directly converge to a local optimum. The experiments also show that the proposed algorithm provides superior performance when β ≈ 2.75±0.25 for the example problems under simulation. According to the analytical results in Section 2, the self-organized optimization dynamics combining the greedy self-organizing process and the fluctuated explorations is an effective search process in complex fitness networks. Furthermore, in order to validate the self-organized behavior of the proposed algorithm, we have randomly generated a large-scale Euclidean TSP instance, with n = 2048, and then statistically analyzed the optimized results of our optimization algorithm. The kth-NND of an optimized result is presented in Fig.6. 2 4 6 8 10 12 14 16 18 20 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 k p( k) p(k) vs. k Exponential fitted curve Fig. 6. Optimization result from the self-organized optimization algorithm. As shown in Fig. 6, the kth-NND for the optimized result obtained by the self-organized optimization algorithm is a nearly perfect exponential distribution. In the exponentially fitted function p(k) = ae–bk, the coefficients a = 0.6791±0.0142, b = 0.5173±0.0115, with 99% confidence bounds. 3.2. Comparison with related methods In past few decades, different nature-inspired or stochastic methods have been proposed for tackling combinatorial optimization problems. In what follows we revisit three classical stochastic search algorithms for purposes of comparison and analysis. – Simulated annealing Simulated annealing (SA) is a stochastic search technique that simulates the thermo-dynamical process of annealing in metallurgy. At each step, the SA algorithm decides whether to replace the current solution according to the Boltzmann probability exp(-∆E/T), where ∆E is the fitness difference between the current solution and a neighboring solution. When the control temperature T is large, a system configuration with worsening energy may be accepted with high probability. As the parameter is sufficiently close to zero by an exponential cooling schedule, the SA algorithm will no longer accept the uphill moves, and then an approximate solution is assumed to be obtained. Since numerous microstates are sampled far from the optimal solution, where the probability of solutions is relatively high, and a feasible solution is easy to be obtained, as discussed in Section 2.5, the SA optimization dynamics is probably not as efficient as the presently studied self-organized optimization dynamics. – Genetic algorithm Genetic algorithm (GA) is a population-based searching algorithm based on the evolutionary theory of natural selection. Generally speaking, the algorithm starts from a population of candidate solutions, and each individual is evaluated in terms of the global fitness. At each generation, a set of individuals are probabilistically selected from the current population according to their global fitness, and updated through crossover and mutation operators for creating a new population of candidate solutions. As compared with our self-organized optimization algorithm, the GA optimization dynamics seems to be a purposeless generation-test process since it mimics evolution at the global system-level solution, and does not go deep into the microscopic entities of computational system. – Extremal optimization Although extremal optimization and our proposed algorithm are both inspired by the self-organized process of the Bak-Sneppen evolution model, two significant differences should be pointed out: (i) The definition of local fitness differs. For example, in a TSP tour, the local fitness of city i is defined as fi = 3/(p+q), if it is connected to its p-th and q-th nearest neighbors, respectively (Boettcher and Percus 2000). Obviously, it is not consistent with the global fitness, as we have discussed, and the global optimum is hardly possible to be obtained. (ii) The method for selecting a new solution s' to replace the current solution s is different. It is more like a physical process rather than optimization. 3.3. Experimental validation In order to further validate the effectiveness and efficiency of the proposed self-organized optimization algorithm, we have selected a set of 10 benchmark Euclidean TSP instances, with sizes ranging from 51 to 2392 nodes, as the test benchmarks. They all are available from Reinelt's TSPLIB (Reinelt 1991). In our computational tests, SA, τ-EO, and the proposed self-organized algorithm (SOA) use the same 2-opt move class at each iterative step. SA usually requires to tune several parameters case by case for obtaining a proper decreasing temperature schedule. The parameters of τ-EO and SOA are approximately set as: τ = α ≈ 1+1/ln(n). GA employs the superior edge recombination crossover operator and displacement mutation operator as concluded in Refs. (Potvin 1996; Larraňaga et al. 1999). All these algorithms are encoded in C++ and implemented on Pentium IV (2.4 GHz). Table 1 shows the comparative computational results. The mean value F and standard deviation σ of error(%)=100×(F(s)–F(s*))/F(s*), where F(s) denotes the solution found by a specific algorithm in each of 10 trials, are used to measure the performance and effectiveness of optimization algorithms. The computation time (tCPU) is used to evaluate the efficiency of optimization algorithms. Table 1 Comparison of the computational results from SA, GA, τ-EO and SOA. SA GA τ-EO SOA Problem F(s*) F σ tCPU F σ tCPU F σ tCPU F σ tCPU eil51 426 0.00 0.00 2 0.00 0.00 25 0.00 0.00 2 0.00 0.00 2 pr76 108159 0.21 0.24 3 0.00 0.00 30 0.00 0.00 3 0.00 0.00 3 kroA100 21282 0.69 0.45 5 0.05 0.07 65 0.12 0.09 5 0.02 0.04 5 ch130 6110 0.74 0.54 10 0.29 0.25 102 0.31 0.34 10 0.13 0.15 10 ch150 6528 0.92 0.61 12 0.38 0.40 149 0.51 0.42 12 0.22 0.20 12 tsp225 3916 1.42 0.77 21 0.66 0.35 212 0.65 0.49 21 0.45 0.34 21 a280 2579 1.45 0.71 45 0.64 0.51 362 0.53 0.46 45 0.32 0.31 45 pcb442 50778 2.07 1.05 125 1.53 0.62 1011 1.52 0.68 125 1.15 0.53 125 pr1002 259045 3.16 1.24 300 1.89 0.80 2957 1.95 0.76 300 1.77 0.69 300 pr2392 378032 4.45 2.23 900 3.83 1.51 5326 3.55 1.54 900 3.21 1.25 900 It can be seen from Table 1 that the proposed self-organized optimization method provides superior optimization performances in both effectiveness and efficiency, and outperforms the state-of-the-art simulated annealing, genetic algorithm, and extremal optimization for all 10 test instances. 4. Concluding remarks The main contributions of this work are mainly twofold: First, we have developed an analytic characterization for combinatorial optimization problems from the self-organizing system’s point of view. Based on the definitions of discrete state variable and local fitness, in this paper we have discussed the empirical observation of the microscopic distribution with respect to an optimal solution. We have also introduced a notion of fitness network in order to characterize the structure of the solution space, and have statistically analyzed the searching complexity of hard optimization problems from the characteristics of phase transition. Second, we have provided as well as demonstrated the performances of a self-organized optimization method for solving hard computational systems. For different optimization problems, only the definitions of a consist local fitness function and the neighboring rules are required. Based on our analytic and algorithmic discussions, it is noted that the presented work offers new insights into, as well as may inspire, further studies on the systematic and microscopic analysis of NP-complete problems. The work paves the way for utilizing self-organization in solving combinatorial optimization problems. Our future work will involve in-depth studies on the optimization dynamics of computational systems by means of introducing the fundamentals of self-organizing systems (Liu and Tsui 2006). Acknowledgements The work reported in this paper has been supported in part by a HKBU Faculty Research Grant (FRG) (FRG/06-07/II-66) and in part by a Hong Kong Research Grant Council (RGC) Central Allocation Group Research Grant (HKBU 1/05C). References Achlioptas, D., A. Naor, and Y. Peres. 2005. Rigorous Location of Phase Transitions in Hard Optimization Problems. Nature, 435: 759–764. Altenberg, L. 1997. NK Fitness Landscapes. In Bäck, T., D.B. Fogel, and Z. Michalewicz, (Eds.) Handbook of Evolutionary Computation, Oxford University Press, New York. Bak, P. 1996. How Nature Works: The Science of Self-Organized Criticality. Copernicus Press, New York. Bak, P., and K. Sneppen. 1993. Punctuated Equilibrium and Criticality in a Simple Model of Evolution. Physical Review Letters, 71(24): 4083–4086. Blum, C., and A. Roli. 2003. Metaheuristics in Combinatorial Optimization: Overview and Conceptual Comparison. ACM Computing Surveys, 35(3): 268–308. Boettcher, S. 2000. Extremal Optimization: Heuristics via Co-Evolutionary Avalanches. Computing in Science & Engineering, 2(6): 75–82. Boettcher, S., and A.G. Percus. 2003. Optimization with Extremal Dynamics. Complexity, 8: 57–62. Boettcher, S., and A.G. Percus. 2000. Nature's Way of Optimizing. Artificial Intelligence, 119: 275–286. Bonabeau, E., M. Dorigo, and G. Theraulaz. 2000. Inspiration for Optimization from Social Insect Behaviour. Nature, 406: 39–42. Cheesman, P., B. Kanefsky, and W.M. Taylor. 1991. Where the Really Hard Problems Are. Proc. IJCAI'91, pp. 163–169. Chen, Y.W., and Y.Z. Lu. 2007. Gene Optimization: Computational Intelligence from the Natures and Micro-mechanisms of Hard Computational Systems. LSMS 2007, Lecture Notes in Computer Science, 4688: 182–190. Chen, Y.W., Y.Z. Lu, and P. Chen. 2007. Optimization with Extremal Dynamics for the Traveling Salesman Problem. Physica A, 385: 115–123. Chen, Y.W., and P. Zhang. 2006. Optimized Annealing of Traveling Salesman Problem from the nth-Nearest-Neighbor Distribution. Physica A, 371: 627–632. Cormen, T.H., C.E. Leiserson, R.L. Rivest, and C. Stein. 2001. Introduction to Algorithms. MIT Press. Duch, J., and A. Arenas. 2005. Community Detection in Complex Networks using Extremal Optimization. Physical Review E, 72: 27104. Forrest, S. 1993. Genetic Algorithms: Principles of Natural Selection Applied to Computation. Science, 261(5123): 872–878. Garey, M.R. and D.S. Jonhson. 1979. Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman, New York. Gent, I.P., and T. Walsh. 1996. The TSP Phase Transition. Artificial Intelligence, 88: 349–358. Goles, E., M. Latapy, C. Magnien, M. Morvan, H.D. Phan. 2004. Sandpile Models and Lattices: A Comprehensive Survey. Theoretical Computer Science, 322: 383–407. Gutin, G., and A.P. Punnen. 2002. The Traveling Salesman Problem and Its Variations. Kluwer Academic Publishers. Han, J. 2003. Local Evaluation Functions and Global Evaluation Functions for Computational Evolution. DOI: SFI-WP 03-09-048. Han, J., and Q.S. Cai. 2003. Emergence from Local Evaluation Function. Journal of Systems Science and Complexity, 16(3): 372–390. Hordijk, W. 1997. A Measure of Landscapes. Evolutionary Computation, 4(4): 335–360. Kauffman, S.A. 1993. The Origins of Order: Self-Organization and Selection in Evolution. Oxford University Press. Kennedy, J., and R. Eberhart. 1995. Particle Swarm Optimization. IEEE International Conference on Neural Networks, 4: 1942–1948. Kern, W. 1993. On the Depth of Combinatorial Optimization Problems, Discrete Applied Mathematics, 43(2): 115–129. Kirkpatrick, S., C.D. Gelatt Jr., and M.P. Vecchi. 1983. Optimization by Simulated Annealing. Science, 220(4598): 671–680. Larraňaga, P., C.M.H. Kuijpers, R.H. Murga, I. Inza, and S. Dizdarevic. 1999. Genetic Algorithms for the Travelling Salesman Problem: A Review of Representations and Operators. Artificial Intelligence Review, 13: 129–170. Liu, J., J. Han, and Y.Y. Tang. 2002. Multi-Agent Oriented Constraint Satisfaction. Artificial Intelligence, 136: 101–144. Liu, J., X. Jin, and K.C. Tsui. 2005. Autonomy Oriented Computing: From Problem Solving to Complex Systems Modeling. Kluwer Academic Publishers. Liu, J., Y.Y. Tang, and Y. C. Cao. 1997. An Evolutionary Autonomous Agents Approach to Image Feature Extraction. IEEE Trans. on Evolutionary Computation, 1(2): 141–158. Liu, J., and K.C. Tsui. 2006. Toward Nature-Inspired Computing. Communications of the ACM, 49(10): 59 – 64. Martin, O.C., R. Monasson, and R. Zecchina. 2001. Statistical Mechanics Methods and Phase Transitions in Optimization Problems. Theoretical Computer Science, 265: 3–67. Merz, P. 2004. Advanced Fitness Landscape Analysis and the Performance of Memetic Algorithms. Evolutionary Computation, 12(3): 303–325. Merz, P., and B. Freisleben. 2000. Fitness Landscape Analysis and Memetic Algorithms for the Quadratic Assignment Problem. IEEE Trans. on Evolutionary Computation, 4(4): 337–352. Mézard, M., G. Parisi, and R. Zecchina. 2002. Analytic and Algorithmic Solution of Random Satisfiability Problems. Science, 297: 812–815. Middleton, A.A. 2004. Improved Extremal Optimization for the Ising Spin Glass. Physical Review E, 69: 055701(R). Monasson, R., R. Zecchina, S. Kirkpatrick, B. Selman, and L. Troyanskyk. 1999. Determining Computational Complexity from Characteristic ‘Phase Transitions’. Nature, 400: 133–137. Papadimitriou, C.H. 1994. Computational Complexity. 1st edition, Addison Wesley. Papadimitriou, C.H., and K. Steiglitz. 1998. Combinatorial Optimization: Algorithms and Complexity. Courier Dover Publications. Potvin, J.Y. 1996. Genetic Algorithms for the Traveling Salesman Problem, Annals of Operations Research, 63: 339–370. Reidys, C.M., and P.F. Stadler. 2002. Combinatorial Landscapes. SIAM Review, 44(1): 3–54. Reinelt, G. 1991. TSPLIB-a Traveling Salesman Problem Library. ORSA Journal on Computing 3(4): 376–384. Sneppen, K. 1993. Extremal Dynamics and Punctuated Co-Evolution. Physica A, 221: 168–179. TSPLIB, http://www.iwr.uni-heidelberg.de/groups/comopt/software/TSPLIB95/. Weinberger, E.D. (1996). NP Completeness of Kauffman's N-k Model, A Tuneable Rugged Fitness Landscape. Technical Report 96-02-003, Santa Fe Institute. Yang, B., and J. Liu. 2007. An Autonomy Oriented Computing (AOC) Approach to Distributed Network Community Mining. Proc. 1st International Conference on Self-Adaptive and Self-Organizing Systems, pp. 151–160. Zhang, W. 2004. Phase Transitions and Backbones of the Asymmetric Traveling Salesman Problem. Journal of Artificial Intelligence Research, 21: 471–497. 
work_rif2akzab5gxffmzav5mcf5uva	----	A hybrid computational approach for seismic energy demand prediction This is a repository copy of A hybrid computational approach for seismic energy demand prediction. White Rose Research Online URL for this paper: http://eprints.whiterose.ac.uk/156298/ Version: Accepted Version Article: Gharehbaghi, S, Gandomi, AH, Achakpour, S et al. (1 more author) (2018) A hybrid computational approach for seismic energy demand prediction. Expert Systems with Applications, 110. C. pp. 335-351. ISSN 0957-4174 https://doi.org/10.1016/j.eswa.2018.06.009 © 2018, Elsevier. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http://creativecommons.org/licenses/by-nc-nd/4.0/. eprints@whiterose.ac.uk https://eprints.whiterose.ac.uk/ Reuse This article is distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs (CC BY-NC-ND) licence. This licence only allows you to download this work and share it with others as long as you credit the authors, but you can’t change the article in any way or use it commercially. More information and the full terms of the licence here: https://creativecommons.org/licenses/ Takedown If you consider content in White Rose Research Online to be in breach of UK law, please notify us by emailing eprints@whiterose.ac.uk including the URL of the record and the reason for the withdrawal request. mailto:eprints@whiterose.ac.uk https://eprints.whiterose.ac.uk/ A HYBRID COMPUTATIONAL APPROACH FOR 1 SEISMIC ENERGY DEMAND PREDICTION 2 3 S. Gharehbaghia, A.H. Gandomib,*, and S. Achakpourc, M.N. Omidvard 4 5 aDepartment of Civil Engineering, Behbahan Khatam Alanbia University of Technology, Behbahan, Iran 6 bSchool of Business, Stevens Institute of Technology, Hoboken, NJ 07030, USA 7 cDepartment of Civil Engineering, Iran University of Science and Technology, Narmak, Tehran, Iran 8 dCenter of Excellence for Research in Computational Intelligence and Applications, School of Computer 9 Science, University of Birmingham, Birmingham B15 2TT, U.K. 10 11 Abstract. In this paper, a hybrid genetic programming (GP) with multiple genes is implemented for 12 developing prediction models of spectral energy demands. A multi-objective strategy is used for 13 maximizing the accuracy and minimizing the complexity of the models. Both structural properties and 14 earthquake characteristics are considered in prediction models of four demand parameters. Here, the 15 earthquake records are classified based on soil type assuming that different soil classes have linear 16 relationships in terms of GP genes. Therefore, linear regression analysis is used to connect genes for 17 different soil types, which results in a total of sixteen prediction models. The accuracy and 18 effectiveness of these models were assessed using different performance metrics and their performance 19 was compared with several other models. The results indicate that not only the proposed models are 20 simple, but also they outperform other spectral energy demand models proposed in the literature. 21 Keywords: Evolutionary computation; Genetic programming; Regression Analysis; Input energy; 22 Hysteretic energy; Seismic energy spectra. 23 24 *Corresponding author. Email:a.h.gandomi@stevens.edu Tel.: +1 (201) 216-5029; Fax: 201-216-5385 URL: http://gandomi.beacon-center.org/ https://mail.google.com/mail/u/0/h/mur9z0mq4ulr/?&cs=wh&v=b&to=a.h.gandomi@stevens.edu http://gandomi.beacon-center.org/ Table of Contents 25 26 1. Introduction 27 2. Seismic Energy Concept and Formulation 28 3. Genetic Programming 29 3.1. Multi-Gene Symbolic Regression 30 3.2. Multi-Objective Genetic Programming (MOGP) 31 3.3. Accelerating GP Process 32 4. Predicting Seismic Energy Demand Spectra Using MOGP 33 4.1. Preparing Exact Data 34 4.1.1. Inelastic SDOF Systems 35 4.1.2. Earthquake Ground Motions 36 4.2. Model Development Using MOGP 37 5. Results and Discussion 38 5.1. Mathematical Model for EDP1 39 5.2. Mathematical Model for EDP2 40 5.3. Mathematical Model for EDP3 41 5.4. Mathematical Model for EDP4 42 5.5. Accuracy of the Models 43 5.6. Comparative Study 44 5.6.1. Assumptions 45 5.6.2. Comparison 46 6. Summary and Conclusion 47 48 1. Introduction 49 The approaches currently used for the seismic analysis and design of structures need to be improved 50 through considering appropriate engineering demand parameters that would represent the 51 characteristics of a structure and the design earthquake. In current approaches, either the structural 52 members are designed based on satisfying the balance between the force demand and the corresponding 53 strength supply while providing an adequate level of ductility (based on e.g. ASCE/SEI7 2010), or 54 based on the concept whether they are force- or deformation-controlled (based on e.g. FEMA-356 55 2000). Such approaches disregard the frequency content and duration of earthquake ground motion, as 56 well as the velocity response and hysteretic behavior (Gupta 1990). The importance of considering 57 these factors lies in evidences suggesting that, for instance, dissipated hysteretic energy due to repeated 58 inelastic excursions could result in a certain amount of seismic damage (Fajfar and Vidic 1994). In fact, 59 in addition to the force and deformation, the energy demand is of great importance in capturing the 60 mentioned seismic factors as the inelastic behavior is expected to occur due to the design and maximum 61 earthquakes. Housner (1956) was first to introduce these factors through defining the energy concept. 62 This concept requires that the energy dissipation capacitybe less than the input energy demand. 63 Both structural properties and earthquake characteristics affect the seismic energy demand. 64 Determining the spectral values of energy demand is beneficial due to its connection tothe amount of 65 structural damage (Fajfar and Vidic 1994, Gharehbaghi 2018). The design codes have not explicitly 66 implemented the energy demand parameter in predicting seismic demands yet. Moreover, the priority 67 of the energy-based design approach compared with the conventional strength-based design approach 68 needs further studies. Although, previous studies (e.g. Housner 1956, Fajfar and Vidic 1994, Kalkan 69 and Kunnath 2007, Akiyama 1985, Bertero and Uang 1988, Decanini and Mollaioli 2001, Manfredi 70 2001) have stipulated that the seismic energy parameters are of great importance in seismic design of 71 structures. It was shown that the hystertic energy demand is directly connectedto the structural damage 72 (Fajfar and Vidic 1994, Kalkan and Kunnath 2007, Akiyama 1985, Bertero and Uang 1988, Decanini 73 and Mollaioli 2001, Manfredi 2001, Benavent-Climent et al. 2010, Gharehbaghi 2018). After more than 74 two decades that the Housner’s proposal was almost neglected, it was received considerable attention 75 among researchers (Akiyama 1985, Kuwamura and Galambos 1989), and became the key issue of a 76 conference held in Bled city of Slovenia (Fajfar and Krawinkler 1992). It was recognized that input 77 energy and hysteretic energy are the indicators of ground motion and have a correlation with the 78 structural damage, and the quantity related to cumulative damage is the hysteretic energy (Fajfar and 79 Vidic 1994, Bertero and Uang 1988, Decanini and Mollaioli 2001). Most recently, Deniz et al. (2017) 80 also found that the most appropriate and reliable intensity measure for the seismic fragility analysis of 81 buildings is the seismic energy demand. 82 Estimation of input and hysteretic energy demands using mathematical models could be considered 83 as one of the important steps aligned with the extension of the energy-based seismic analysis and design. 84 As previously mentioned, both structural and earthquake characteristics need to be accounted for the 85 issue. Some earthquake characteristics such as soil type, earthquake magnitude, peak ground 86 acceleration (PGA), peak ground velocity (PGV), cumulative energy index, fault type, distance from 87 the hypocenter, were used by researchers in determining the energy spectra (e.g. Zahra and Hall 1984, 88 Uang and Bertero 1988, Fajfar et al. 1989, Uang and Bertero 1990, Sucuoğlu and Nurtuğ 1995, 89 Khashaee 2004). In addition to the earthquake characteristics, ductility ratio, damping ratio, and 90 hysteretic behavior model (e.g. elastic-perfectly plastic, bilinear, pinching, Takeda, and Clough models) 91 were the influential structural properties involved in the estimation of seismic energy demand spectra 92 (e.g. Sucuoğlu and Nurtuğ 1995, Decanini and Mollaioli 2001, Benavent-Climent et al. 2011).Several 93 works have been carried out on the estimation of the seismic energy demand parameters. Housner 94 (1956) presented a model to determine input energy based on the spectral velocity of SDOF system. 95 Kuwamura and Galambos (1989) presented energy demand spectra considering the soil type and 96 dominant period of the earthquake. Chou and Uang (2000) estimated absorbed energy for an inelastic 97 system by using an attenuation relation. They used nonlinear regression analysis considering both 98 structural and earthquake variables. Manfredi (2001) proposed simple/efficient mathematical models 99 to estimate input and hysteretic energy spectra. A dimensionless seismic index that is a function of 100 PGA, PGV and cumulative energy was proposed to estimate the seismic energy spectra. Although the 101 estimation models were simple and effective, the effect of soil behavior was not considered, and the 102 number of earthquake ground motions was rather limited. Decanini and Mollaioli (2001) proposed the 103 formulation of elastic seismic energy spectra. They also presented a comprehensive study to propose 104 the design inelastic energy spectra by introducing the response modification factor for the input energy. 105 Several structural variables (e.g. ductility ratio and hysteretic behavior) and earthquake characteristics 106 such as soil type, source-to-site distance, and earthquake magnitude were considered in the proposed 107 spectra. Arroyo and Ordaz (2007) estimated the hysteretic energy demand spectra from elastic response 108 parameters in accordance with the earthquake events recorded in Mexico City. Their mathematical 109 models were a function of pseudo-acceleration, velocity and displacement spectra. Elastic design input 110 energy spectra based on Iranian earthquakes were also presented by Ghodrati Amiri et al. (2008). 111 Recently, Dindar et al. (2015) proposed two regression-based simple mathematical models to estimate 112 the input and hysteretic energy spectra. A database of earthquake ground motion records composed of 113 near- and far-fault ones, PGA, soil types, earthquake magnitude, ductility ratio, and hysteretic behavior 114 model was included in the proposed models. Using the regression analysis, Quinde et al. (2016) also 115 proposed mathematical models to estimate the seismic energy spectra of inelastic systems located on 116 the soft soil for Mexico City. They captured the effect of ductility ratio of inelastic systems and 117 dominant period of the probable earthquakes on the presented models. More recently, Zhai et al. (2016) 118 proposed an expression to account for the effect of after-shock on the input energy spectra using an 119 equivalent velocity. Alıcı and Sucuoğlu (2016) carried out a regression analysis to estimate inelastic 120 input energy spectrum. The prediction equations for the input energy spectra were expressed in terms 121 of an equivalent velocity. Some crucial earthquake characteristics including soil type, epicentral 122 distance, moment magnitude, and the fault type were considered in the proposed models. All the 123 previously mentioned works use conventional regression methods to estimate their energy parameters 124 of interest. 125 Based on the capability of soft computing approaches and their recent advances, it is worthwhile to 126 use such efficient approaches for seismic demand prediction. Computational complexity of the 127 conventional methods and their limitations has made soft computing techniques, such as evolutionary 128 algorithms, artificial neural networks, support vector machines, and fuzzy logic, popular for solving 129 complex engineering problems. A common application of these tools is in predictive analysis for 130 modeling the nonlinear dependency of the input parameters to the output value(s) where the 131 conventional approaches (e.g. regression analysis) fail or perform poorly (Khan et al. 2003, Gandomi 132 and Roke 2015). Despite the success of artificial neural networks (ANNs) in prediction, they are 133 inappropriate to develop practical intelligible equations. In addition to ANNs, support vector machines 134 (SVMs) are another primary class of soft computing methods used to discover patterns and approximate 135 relationships when large quantities of data is available. Although both ANNs and SVMs have received 136 significant attention (e.g. Salajegheh and Heidari 2005, Gholizadeh and Salajegheh 2009, Papadopoulos 137 et al. 2012, Gharehbaghi and Khatibinia 2015, Khatibinia et al. 2015, Yazdani et al. 2016), they require 138 a pre-defined and initial structure for the equation and network architecture to be determined by the 139 user. Genetic programming (GP), a learning algorithm originated from genetic algorithms, is another 140 well-known and successful technique for developing nonlinear mathematical models for the complex 141 problems. GP and its variants have been effectively used for solving various problems in civil 142 engineering (e.g. Kayadelenet al. 2009, Alavi et al. 2011, Mirzahosseini et al. 2011, Gandomi et al. 143 2012, Vardhan et al. 2016; Lim et al. 2016). Several variants of GP have been proposed in the literature, 144 such as gene expression programming (Ferreira 2006) and multi-stage genetic programming (Gandomi 145 and Alavi 2011).One of the robust variants of GP is multi-gene genetic programming (MGGP) that 146 adds the capability of conventional regression to the standard GP ability in parameter estimation. The 147 effectiveness of MGGP has been proved in the works reported by Gandomi et al. (e.g. Gandomi and 148 Alavi 2012a,b, Gandomi et al. 2013, Babanajad et al. 2013, Gandomi et al. 2016). 149 Structural and earthquake engineering has benefited from the soft computing techniques in different 150 applications. For instance, ANNs and SVMs have been widely used for risk assessment, seismic 151 response prediction, control and health monitoring (Tsompanakis and Topping 2011). In this paper, 152 MOGP is used for predicting the seismic energy demand spectra considering both typical structural and 153 earthquake characteristics. For this purpose, eighteen set of the single-degree-of-freedom (SDOF) 154 systems with the structural properties of different hardening ratios of bilinear hysteretic behavior model, 155 damping ratios, and ductility ratios are used to determine the energy demand spectra mentioned. Also, 156 four different sets of earthquake ground motion records based on their soil types (soft, firm, stiff and 157 rock) with the source-to-site distances of more than 17.5 km and the magnitudes of greater than 5.5 158 were used. It was assumed that the different soil classes have linear relationships in terms of GP genes 159 which help to find one equation with different coefficients for different soil types. The records were 160 scaled to two PGA levels 0.5g and 1.0g. Finally, four mathematical models corresponding to the four 161 engineering demand parameters (EDPs) of spectral input and hysteretic energy, spectral hysteretic to 162 input energy ratio, and spectral energy modification factor, are proposed using MOGP. Then, the 163 effectiveness of the models is revealed using the performance metrics compared with those of available 164 in the literature. 165 In this study, section 2 describes the seismic energy concept and its formulation. Also this section 166 introduces the seismic energy based EDPs which can be useful in seismic design of inelastic structures. 167 Section 3 express a hybrid computational approach based on genetic programming used as a predictive 168 tool herein. Section 4 describes a framework for prediction of the EDPs. A set of mathematical models 169 are proposed and their accuracy are examined using some performance metrics in section 5. Finally, 170 the developed models are discussed and compared with some other methods proposed in the literature. 171 172 2. Seismic Energy Concept and Formulation 173 Housner (1956) first proposed the idea of the energy-based seismic design approach. When ground 174 motion transmits energy into a structure, some of the energy is dissipated through the damping and 175 inelastic behavior. The remained energy of the structure is stored in the form of kinetic energy and 176 elastic strain energy. Housner stipulated that the energy supply should be more than the energy demand 177 during an earthquake in the form of this principle that energy supply < energy demand for controlling 178 and avoiding the structural collapse (Housner 1956). 179 When a structure is subjected to earthquake excitation, its governing equation for the dynamic 180 behavior of an inelastic SDOF system could be written as (Chopra 2012): 181 ( ) ( ) ( ( ), ( )) ( )s gmu t cu t f u t u t mu t+ + = − (1) where m, c and fs represent the mass, damping coefficient and lateral resisting force of SDOF system, 182 respectively; ü(t), ( )u t and u(t) are acceleration, velocity and displacement relative to the ground with t 183 representing time in SDOF system, respectively; and üg(t) is the earthquake ground acceleration. 184 Governing equation of energy equilibrium is obtained by the integration of Eq. (1) with respect to u 185 (Uang and Bertero 1990, Chopra 2012): 186 ( ) ( ) ( ( ), ( )) ( )s gmu t du cu t du f u t u t du mu t du+ + = −    (2) In fact, Eq. (2) expresses the energy balance for a structural system during an earthquake while Eq. 187 (1) explains the force balance. Substituting displacement unit (du) by velocity term and integrating it 188 over the time of the earthquake ground motion t, Eq. (2) is expressed as: 189 ( ) ( ) ( ) ( ) ( ( ), ( )) ( ) ( ) ( )s gmu t u t dt cu t u t dt f u t u t u t dt mu t u t dt+ + = −    , (3) where t is the time of interest across the earthquake ground motion. Eq. (3) can also be written in a 190 general form as follows: 191 ( ) ( ) ( ) ( )K D A IE t E t E t E t+ + = , (4) where EI(t), EK(t) and ED(t) are the input energy demand, kinetic energy, and damping energy, 192 respectively; EA(t) is encompassed the recoverable elastic strain energy ES(t) and the irrecoverable 193 plastic hysteretic energy EH(t). The amount of EH(t) is equal to zero in the elastic systems and is 194 appeared in the inelastic systems. Therefore, Eq. (4) can be expressed as: 195 ( ) ( ) ( ) ( ) ( )K D S H IE t E t E t E t E t+ + + = , (5) where EK(t) and ES(t) are cumulative during the earthquake ground motion and are vanished at the end 196 of motion of the inelastic systems. In effect, these two terms are very small in comparison with ED(t) 197 and EH(t). Since the most portion of energy demand is dissipated through the damping energy and 198 hysteretic energy, the Eq. (5) can be approximately written as (Uang and Bertero 1990): 199 ( ) ( ) ( )D H IE t E t E t+  . (6) According to Eq. (6), the main portion of input energy demand is converted to the damping energy 200 and hysteretic energy. Besides, if the structural system remains in the elastic range during an 201 earthquake, EH(t) is trivial, and the energy-based analysis is not useful for seismic design (Uang and 202 Bertero 1990).As mentioned before, for design basis earthquakes, it is expected that a structure will 203 experience inelastic cyclic deformations resulting in hysteretic energy dissipation. As previously 204 mentioned, the hysteretic energy dissipation (EH) is directly attributed to the structural damage where 205 the EH /EI ratio has been introduced as a good indicator of expected damage (Fajfar and Vidic 1994, 206 Sucuoğlu and Nurtuğ 1995, Decanini and Mollaioli 2001, Manfredi 2001). For a given ductility ratio 207 (), EH /EI ratio is defined as follow: 208 H I E HI E    = , (7) where the EI and EH are the input and hysteretic energy corresponding to the ductility ratio of . 209 Another parameter that is of crucial importance for earthquake resistant design based on the energy 210 concept could be the response modification factor of the input energy that can be expressed as follow: 211 I I I E RE E   = . (8) The literature suggests that to have a practical energy based seismic design, the computation of the 212 input and hysteretic energy (EI and EH), hysteretic energy to input energy ratio (HI), response 213 modification factor of input energy (REI), and energy dissipation capacity is useful. Since the design 214 method requires inelastic dynamic analyses resulting in the expensive computational efforts, the use of 215 soft computing techniques is of great importance in predicting the practical mathematical models 216 (formulations). Except for the energy dissipation capacity that needs comprehensive experimental and 217 theoretical studies, the spectral values of the mentioned energy-based EDPs were predicted using 218 MOGP. The next section describes the MOGP. 219 3. Genetic Programming 220 There are two groups of models which can be used for modeling the complex nonlinear engineering 221 systems: phenomenological and behavioral (Gandomi et al. 2016). Phenomenological models need a 222 predefined structure obtained from the physical laws requiring a previous understanding about the 223 system. Concerning the complex systems, sometimes it is hard to find such models. Unlike the 224 phenomenological models, behavioral models can be simply generated by finding a reasonable 225 approximate relation between input variables and the output value(s) for a collection of data 226 (experimental or theoretical) irrespective of their governing physical principles. Although one of the 227 main advantages of behavioral modeling techniques is their independence on prior knowledge about 228 the governing physical relationships of input and outputs (Walter and Pronzato 1997, Metenidis 2004), 229 most of these models need the user to pre-assign a formulation pattern requiring optimization of its 230 unknown coefficients. Concerning the complex engineering systems, the use of conventional 231 techniques such as regression analysis cannot be guaranteed to find an accurate and reliable behavioral 232 model (Gandomi et al. 2016). It has been well recognized that most of the structural earthquake 233 engineering problems such as determining earthquake response of inelastic structures could be 234 considered as such complex problems. 235 Genetic programming (GP) (Koza 1990) is a novel behavioral modeling methodology with 236 completely new characteristics. GP is an extension of the genetic algorithm capable of functionalizing 237 data using tree structures. In fact, unlike classic regression models and ANNs, GP is capable of 238 generating a prediction equation irrespective of a predefined structure. The successful application of 239 GP and its variants have been reported in solving the various real-world problems (e.g. Gandomi and 240 Alavi 2012a,b, Gandomi et al. 2013, Babanajad et al. 2013, Gandomi et al. 2016). One of the robust 241 variants of GP is MGGP that adds the capability of conventional regression to the standard GP ability 242 in parameter estimation, which has been proposed recently (Searson et al. 2007).The initial MGGP 243 studies show that it can outperform other GP variants (Gandomi and Alavi 2012a,b). MGGP is 244 described in the next subsection. 245 3.1. Multi-Gene Symbolic Regression 246 GP can generally be defined as a supervised machine learning technique that searches a program space 247 instead of a parameter space. One of the useful variants of GP is Multi-gene genetic programming 248 (MGGP) (Searson et al. 2007, Gandomi and Alavi2012a). MGGP is used to design mathematical model 249 predictions which are inherently multi-gene, i.e., those models consisting of linear combinations of low 250 order nonlinear transformations of the input variables. Unlike conventional GP that is based on an 251 evaluation of single tree expression, MGGP uses a single GP particle swarm model selection program 252 constructed from a number of genes where each gene has a tree expression (Searson et al. 2010). In 253 effect, the model development procedure is decomposed by MGGP consisted of some relatively simple, 254 fixed-depth sub-models. 255 To develop a population of trees, GP typically uses symbolic regression. Compared with the 256 conventional GP, a weighted linear combination of outputs from a number of GP trees is considered as 257 a symbolic model in which each of these trees could be considered as a “gene” (Searson et al. 2010). 258 The number of genes and tree depth of any gene can be specified by the user which their maximum 259 values depend on the complexity of the models developed by MGGP. The evolved models are linear 260 combinations of low-order nonlinear transformations of the predictor variables (Searson et al. 2010). 261 The ordinary least squares method is used to estimate the linear coefficients for each of the evolved 262 genes of an individual. A more detailed explanation of MGGP is available in Refs. (Searson et al. 2007, 263 Searson et al. 2010). 264 A fixed linear illustration associated with binary encoding of all parameters is used in GA, which 265 results in a string of numbers as output. In GP however, the optimization problems are solved 266 irrespective of a pre-defined solution structure. Depending on the problem domain, the first generation 267 consisting of a population of possible solutions is randomly created by GP, and variation operators 268 generate new candidate solutions then. During the evolutionary process, crossover selects a node from 269 the parental individuals and exchanges the subtrees under the selected nodes randomly and creates a 270 new individual. Mutation truncates and replaces one node of a tree with another randomly generated 271 node from the same set, and then creates a new individual from an existing tree in the population. The 272 individuals with higher fitness values have higher survival likelihood in the successive generation. To 273 find the best fitting solution, the individual solution of a population is updated after executing a number 274 of runs in each generation, and the parental individuals are selected from the population based on a 275 good fitness function. To develop a population of genes, symbolic regression method can be 276 implemented using standard GP in which a symbolic mathematical expression is directly encoded by 277 each of the genes. Based on Figure 1 showing a typical multi-gene model, known as multi-gene 278 symbolic regression, three input variables (x1, x2, and x3) are used to predict the response. As shown in 279 the figure, although the nonlinear terms such as “sin” and “log” are used, the overall model is a weighted 280 linear combination of each gene utilizing the coefficients go, g1, and g2. Mathematically, the general 281 formulation of the multi-gene symbolic regression model can be written as (Gandomi et al. 2016): 282 1 ˆ( , , ) ( ( )) n o i i i y g g G = = + x g x  (9) where go is a bias term; gi is the gene weight, and Gi(ș,x) is the outputs vector from the ith gene 283 encompassing a multi-gene individual; ș is the vector of the unknown parameters for each gene; and n 284 is the number of genes. It should be noted that the algorithmic structure of MGGP, and standard GP is 285 the same, except for crossover and mutation of multi-gene individuals. MGGP does not necessitate any 286 simplifying assumption in the model development process, and it is more accurate and efficient than 287 the standard GP for modeling complex nonlinear problems (Gandomi and Alavi 2012a,b). To construct 288 an initial population in MGGP, random individuals are created by using different nonlinear functions, 289 input variables, and a range of random constants and each individual includes 1 and Gmax number of 290 genes. The algorithm attempts to maximize diversity by ensuring that no individuals contain duplicate 291 genes. The genes are randomly selected and the least squares normal equation is used to estimate the 292 vector of unknown coefficients g as follows (Searson et al. 2007, Searson et al. 2010, Hii et al. 2011, 293 Searson 2014): 294 ( ) 1T T−=g G G G y (10) where G = [1 G1…Gn] is the gene response matrix. Since the columns of matrix G can be collinear, the 295 Moore–Penrose pseudo-inverse (GTG)# can be computed by means of the singular value decomposition 296 instead of the standard matrix inverse (GTG)-1.Genes can be employed or eliminated using a tree 297 crossover operator (high-level crossover) during the MGGP run which is in addition to the standard GP 298 sub-tree crossover (a low-level crossover). The low-level crossover chooses a gene randomly from each 299 parent individual. Next, the standard sub-tree crossover is employed and the generated trees replace the 300 parent trees in the otherwise unaltered individual in the next generation. The high-level crossover allows 301 the exchange of one or more genes with another selected individual subject to the Gma x constraint. The 302 maximum number of genes of an individual is limited to Gmax.If any individual contains more genes, 303 the additional genes are randomly selected and deleted (Searson et al. 2007, Searson et al. 2010, Hii et 304 al. 2011, Searson 2014). 305 3.2. Multi-Objective Genetic Programming (MOGP) 306 Generally, both tree-based GP and MGGP deal with a single objective optimization problem for each 307 individual considering the defined fitness function in which, for symbolic regression, the goodness-of-308 fit to the training data is considered as the only objective to be maximized. Although MGGP yields 309 more compacted models compared with standard GP (Searson 2014), ineffective genes may be acquired 310 by multi-gene models and the single objective optimization problem results in the evolution of overly 311 complex, impractical and non-robust models (Gandomi et al. 2016). The simplest solution to eliminate 312 such shortcoming can be provided by limiting G in a model to Gmax which is a hard-to-determine unique 313 value for any given problem (Searson 2014). One good solution is the use of multi-objective concepts 314 into symbolic regression which is commonly referred to as multi-objective genetic programming 315 (MOGP). Using this methodology, both the goodness-of-fit and the complexity of the developed models 316 can be optimized simultaneously by searching the so-called Pareto front (non-dominated solutions) set. 317 Herein, the GPTIPS 2 toolbox (Searson 2014) associated with the related subroutines coded in 318 MATLAB (2013) is used to solve the MOGP by using a non-dominated sorting technique (Deb et al. 319 2002). To sort the non-dominant solutions by their complexity and precision, the non-dominated sorting 320 method is applied at the end of each generation of the MGGP algorithm. At first, the individuals are 321 classified from both the new and old population based on their position on the Pareto front. Pareto front 322 of each level encompasses a set of Pareto optimal solutions which other solutions do not dominate. In 323 addition, the solutions that include Pareto front of each level are not dominated by any other solution, 324 apart from those of in its previous Pareto front level. After that, a “crowding factor” (i.e., the average 325 distance of a solution from the nearest solutions (either side) on the same Pareto front) is computed for 326 separately all individual to increase population diversity, giving lower priority to the solutions that are 327 crowded together during the ranking process. Finally, the position of solutions is used to rank them 328 (those on level 1 are ranked above those on level 2, and so on),and a crowding factor of each solution 329 is used to rank those within the same level. The top 50% of the population is remained to participate in 330 the next generation, whilst the rest are eliminated (Searson 2014). 331 3.3. Accelerating GP Process 332 Typically, the data sets used in engineering studies are complex and do not include a very large number 333 of records particularly for experimental studies (Gani et al. 2016). While the successful application of 334 GP in modeling engineering systems has been reported in the literature (e.g., (Sajjadi et al. 2016)), it 335 can be difficult to model the systems with big data using GP. The evolutionary approaches are often 336 slower than statistical data mining. Since GP is usually used to find the structure of solution(s), it is one 337 of the slowest evolutionary algorithms. In addition, the extra process of non-dominated sorting of a 338 multi-objective GP magnifies the problem. To improve this weakness, in this paper, two strategies are 339 used in the prediction process: 340 • 60% of data (2160 samples) were randomly selected and used for training process, and the rest 341 (1440 samples) used as training set for each run; 342 • The final Pareto front was determined from merging the Pareto fronts for all runs. 343 In general, there are two classes of machine learning algorithms including trajectory based 344 algorithms and population-based algorithms. ANNs and regression analysis are two well-known 345 examples of trajectory algorithms. In contrast, GP is one of the mostly used population-based 346 algorithms which it deals with a set of the solution in each generation. This feature makes it flexible to 347 adopt with parallel processing, therefore, the paralleled computations can be used to accelerate MOGP 348 procedure using a distributed computing machine in order to deal with the Big Data issue in GP. 349 Although only twelve cores were used to evolve and evaluate new models herein, the number of cores 350 can be increased up to the population size using this framework. The schematic of parallel processing 351 in the GP process is shown in Figure 2. 352 4. Predicting Seismic Energy Demand Spectra Using MOGP 353 4.1. Preparing Exact Data 354 Since the inelastic responses of an SDOF system highly depend on both structural and earthquake 355 ground motion variables, the most influential ones are contributed in predicting spectral seismic energy 356 demand. The input variables are described in the next subsections in detail. 357 4.1.1. Inelastic SDOF Systems 358 The structural variables used in this study are the hardening ratio of bilinear hysteretic behavior model 359 (), damping ratio (), and the displacement ductility ratio () which are the structural properties used 360 to determine the energy demand spectra mentioned. The values assumed for  were 0.0, corresponding 361 to the elastic-perfectly plastic model, and 0.1 indicating bilinear model. Three values of 0.05, 0.10 and 362 0.15 were also used for  of the inelastic SDOF systems. In addition, three common values of 2, 4 and 363 6 were taken into account for . The periods range studied for the prediction of the energy spectra was 364 between 0.01 to 5.0 second for every 0.05 second. These considered variables (, , and ) resulted in 365 18 inelastic SDOF systems used for the prediction. 366 4.1.2. Earthquake Ground Motions 367 Three factors of the site class, source-to-site distance, and PGA are the three variables considered for 368 the earthquake ground motion records used. Based on the shear wave velocities corresponding to the 369 30 m in depth (Vs,30) of more than 750, 360 to 750, 180 to 360 and less than 180 m/s, four soil types of 370 S1, S2, S3, and S4 were assumed for the records used. Site types of S1, S2, S3, and S4 are respectively 371 representing the soft, firm, stiff and rock soil types. To consider the source-to-site distance, the records 372 having the Joyne-Boor distance (RJB) in the range of more than 17.5km and less than 150km were used 373 (known as far-fault records). In addition, two values of 0.5g and 1.0g were used for the PGA of the 374 records. All records are non-pulse-like and have the magnitude (M) of greater than 5.5. All the records 375 were downloaded from NGA-West-II project of PEER ground motion database (2017). The diversity 376 of M, RJB and Vs,30, and the number of records are shown in Figure 3. As shown in this figure, the records 377 are selected in a way that they have a large variety of the properties mentioned. The individual pseudo-378 spectral acceleration of each record of each soil type and their mean spectra are also shown in Figure 379 4. 380 Four main seismic energy-based EDPs were chosen to be predicted: (i) EDP1: spectral EI/m; (ii) 381 EDP2: spectral EH/m; (iii) EDP3: spectral HI; and (iv) EDP4: spectral REI For this purpose, based 382 on the values assumed for the structural variables, the 18 SDOF systems were modeled and subjected 383 to the mentioned earthquake records. A large number of inelastic time history dynamic analyses (more 384 than 1 million) were carried out, and the exact EDPs were determined. The entire process was simulated 385 in MATLAB platform (2013). 386 For each EDP, individual spectral energy responses of the SDOF systems under each set of 387 earthquake records were obtained. Then, based on the normal distribution, the mean plus one standard 388 deviation (mean+) for each EDP under each set of the earthquake records were determined to be 389 predicted. 390 4.2. Model Development Using MOGP 391 To develop powerful models, the suitable parameters ought to be utilized as a part of the MOGP 392 predictive algorithm. To obtain the optimum MOGP models, basic arithmetic operators (+, -, ×, /) and 393 mathematical functions (e.g. tanh, ln) were used. The models are formed by randomly combining the 394 components from the functional set and the terminal set. The number of programs (solutions) in the 395 population is determined by the population size, and the number of levels that the calculation would 396 apply before the run ends are resolved as per the number of generations.The nature of thedata set, 397 problem complexity, and the number of variables are the three determining factors for the population 398 size and the number of generations. Note that, the upper bounds of an individual (Gmax) and the 399 maximum tree depth (Dmax) need to be defined in order to restrict the complexity. To conduct a trade-400 off between the running time and the complexity of the evolved solutions, optimal values of 3 and 5 401 were respectively assumed for Gmax and Dmax. The parameter settings used for the MOGP 402 implementation listed in Table 1 are based on the previously suggested values in the literature (Searson 403 et al. 2007, Searson et al. 2010, Hii et al. 2011, Searson 2014) and employing a case-dependent trial-404 and-error process. 405 In order to generate new genes for individuals as well as to decrease the overall number of genes for 406 one model and increase the total number of genes for the other, a “rate-based high-level crossover” 407 through the use of a crossover rate parameter (CR) is employed herein. A uniform random number 408 between 0 and 1 with a default value of 0.5 is generated separately for each gene in the parents. If r 409 isless than CR, the corresponding gene is moved to the other individual. In case the exchanging process 410 results in offspring that contain more genes than the Gmax, the gene is randomly eliminated such that the 411 constraint is no longer violated (Searson 2014). Two data sets are needed for the analyses. Therefore, 412 data are randomly divided into two subsets for training and validation. The training dataset is used for 413 learning and the validation set for determining the quality of the evolved programs on unseen data. 414 Several combinations of training and validation sets were considered to determine a consistent data 415 division. To evaluate the evolved expressions and finding the best-encoded one, the minimum of the 416 root mean square error (RMSE) is used as the fitness function. RMSE can be expressed as follows: 417 2 1 1 ( ) n i ii RMSE e p n = = − , (11) where pi and ei are the predicted, and exact output values for the ith output, respectively; and n is the 418 number of samples. Two objectives of maximizing the correlation coefficient (R) and minimizing the 419 model complexity are used in MOGP approach in order to select the best final model. 420 5. Results and Discussion 421 Using MOGP, all EDPs (EDP1, EDP2, EDP3, and EDP4) were predicted, and their optimal 422 mathematical models (formulations) were determined. Four cases (S1 to S4) based on different soil 423 types of the earthquake records were considered in the prediction. Although it is quite possible that the 424 obtained model formulations be different for different soil types, it is more practical to develop a unique 425 mathematical model for an EDP with different coefficients. Therefore, in this paper, the complete 426 database (which includes four soil types of S1 to S4) is employed to develop a unique prediction model 427 for each EDP. The final mathematical model is selected based on a compromise between the prediction 428 accuracy (as measured by the correlation coefficient R) and the model complexity (as measured by the 429 number of input variables). After that, the complete database is divided into four groups based on the 430 four soil types. Using each group of data, the predicted coefficients of the final model (gi) are re-431 evaluated by conducting the regression analysis to reflect the influence of the soil type. Finally, four 432 mathematical models including structural variables (and PGA of earthquake records) with four different 433 groups of coefficients corresponding to the four cases mentioned above (S1 to S4) obtained are 434 presented herein. The contribution of each input variable in the mathematical prediction models and the 435 Pareto front obtained by using a nondominated sorting method at the end of an MOGP run are presented. 436 The results of all EDP models developed by MOGP for EDP1 to EDP4 have been shown in Figures 437 5(a)-(d). The Pareto front sets are shown in green circles and the rest of the models are shown in solid 438 blue circles. As mentioned earlier, the Pareto front set is obtained by using a non-dominated sorting of 439 populations at the end of all MOGP runs. This process simultaneously optimizes the accuracy and the 440 complexity of all developed models. The final model in each Pareto front set is selected and highlighted 441 in a red circle. 442 In order to benchmark the MOGP models, they were compared with gene expression programming 443 (GEP) models as a well-known and widely used GP algorithm. The GEP model also uses multiple gene 444 structure, which makes it similar to the MGGP in this respect. GEP requires 10 times more generations 445 to converge. It is because the MOGP algorithm converges quickly since it uses regression analysis 446 beside the evolutionary process. GEP’s parameter setting is similar to that of MOGP (shown in Table 447 1). The final results of GEP are presented in Figure 5. The results show that none of the models found 448 by GEP are among the Pareto front sets for any of the EDP1 to EDP4 problems. 449 The contribution of each input variable in the mathematical prediction models can be investigated 450 through their frequencies where a frequency value of 1.0 for a variable indicates that it has the maximum 451 contribution within the best-generated models (Gandomi et al. 2010). It was assumed that the models 452 with R2> 0.8 are the best-generated models. The frequency histograms of the input variables for all 453 predicted EDPs are shown in Fig. 6. As shown, for the selected database of EDP1, PGA and 454  respectively show the most and the least statistically significant contributions in the best-generated 455 MOGP models. For the collected database of EDP2, although all the input variables almost have 456 significant contributions in the best generated MOGP models, T and  are the most and the least 457 influential variables on the EDP2 prediction model. According to the literature (e.g. Kuwamura and 458 Galambos 1989, Manfredi 2001, Dindar et al. 2015), there is no report about the role of PGA on EDP3 459 https://en.wikipedia.org/wiki/Gene_expression_programming (hysteretic to input energy ratio). Moreover, based on the physics of the problem, when the damping 460 ratio is increased, the portion of hysteretic energy dissipation of the imparted seismic input energy is 461 decreased. These issues have been confirmed by the frequencies of  and PGA for EDP3 where they 462 have the largest and smallest statistically significant contributions, respectively. Moreover, as can be 463 seen in the figure, for the collected database relevant to EDP4, T and PGA have the largest and smallest 464 statistically significant contributions in the best-generated MOGP models, respectively. 465 5.1. Mathematical Model for EDP1 466 The mathematical model obtained for EDP1 is expressed as follows: 467 ( )0 1 1 2 2 I V E a a X a X S m  = + + , (12-a) 1 .tanh(tanh( ))X PGA T T  = , (12-b) 1.5 2 ( ) e T PGA X     − = , (12-c) where a0, is the bias term, a1 and a2 are the gene weight for the EDP1 prediction model. These 468 coefficients are listed in Table 2. SV is the spectral velocity of the elastic SDOF system. 469 5.2. Mathematical model for EDP2 470 The mathematical model derived for EDP2 is expressed as follows: 471 ( )0 1 1 2 2 H D E b b Y b Y S m  = + + , (13-a) 1 Y PGA= , (13-b) 2 2 ( 2 )( ) T Y e PGA   −= − + − , (13-c) where b0, is the bias term, b1 and b2 are the gene weight for EDP2. These coefficients are listed in Table 472 3. SD is the spectral displacement of the elastic SDOF system. 473 5.3. Mathematical model for EDP3 474 The mathematical model derived for EDP3 is expressed as follows: 475 0 1 1 2 2 H I E HI c c Z c Z E    = = + + , (14-a) 2 0.151 1 TZ e e  − − − −= , (14-b) 2 log ( 7.75) tanh( ) T Z T    = +    , (14-c) where c0 is the bias term, c1 and c2 are the gene weight for EDP3. These coefficients are listed in Table 476 4. 477 5.4. Mathematical model for EDP4 478 The mathematical model obtained for EDP4 is expressed as follows: 479 0 1 1 2 2 I I I E RE d d U d U E   = = + + , (15-a) 1 ( 0.422)( 1.4074) tanh( ) T U e T T   −= + + + − − , (15-b) 2 2 1 ln( )U T T   = + + , (15-c) where d0 is the bias term, d1 and d2 are the gene weight for EDP4. These coefficients are listed in Table 480 5. 481 It should be noted that EDP2 can also be determined by the following relationship: 482 I H I E E H m   = (16) In fact, there are two ways to compute EDP2: one is by using Eq. (13) directly (called as EDP2-1), and 483 another is by using Eq. (16) (called EDP2-2). 484 5.5. Accuracy of the Models 485 It should be noted that all the models are only valid in the range of actual database (discussed in section 486 4.1). In order to investigate the effectiveness and accuracy of the EDP models in the range of our 487 database, the common performance metrics, including the mean absolute percentage error (MAPE), the 488 relative root mean square error (RRMSE), the linear correlation coefficient (R), the performance index 489 (PI), coefficient of determination (r2), coefficient of efficiency (E), and the index of agreement (d) 490 corresponding to the predicted formulation of each EDP are obtained. MAPE, RRMSE, R, PI, r2, E, and 491 d are expressed as follows: 492 1 1 n i i i i e p MAPE n e= − =  , (16) 2 1 ( )1 n i ii e p RRMSE e n = − =  , (17) 1 2 2 1 1 ( )( ) ( ) ( ) n i i i ii n n i i i ii i e e p p R e e p p = = = − − = − −    , (18) 1 RRMSE PI R = + , (19) 2 2 1 2 1 ( ) 1 ( ) n i ii n ii e p r p = = − = −   , (20) 2 1 2 1 ( ) 1 ( ) n i ii n i ii e p E e e = = − = − −   , (21) 2 1 2 1 ( ) 1 ( ) n i ii n i i i ii e p d e e p e = = − = − − + −   , (22) where ie and ip are the average values of the exact and predicted outputs, respectively; and n, ei, and 493 pi were defined before. Lower MAPE and RRMSE, higher R (or R2), r2, E and d indicate the accuracy 494 and effectiveness of the prediction model used. Based on Eq. (19), higher R values and lower RRMSE 495 values result in lower PI and, subsequently, indicate a more precise model. It should be noted that PI 496 varies from 0 to ∞ and its values close to 0 indicate the model fits very well to the exact (actual) values. 497 It is worth noting that two sources of complexity affect the accuracy of the models. First, the structures 498 used are inelastic which leads to high nonlinearity, and second, the earthquake ground motion records 499 have some influential characteristics, such as frequency content, which make a structure to experience 500 different cyclic excursions associated with complex behavior. These problems are accentuated when 501 PGA is increased. 502 The abovementioned performance metrics of the predicted EDP1, EDP2 (including EDP2-1 and 503 EDP2-2), EDP3 and EDP4 models using MOGP are listed in Table 6. As can be seen in this table, 504 EDP1 and EDP2-2 have MAPE values respectively less than almost 16% and 17%, and it is less than 505 33.1% for EDP2-1 all of which are in an acceptable/reasonable range of the performance metric for 506 inelastic and complex systems. The MAPE values are very low for both EDP3 and EDP4. Lower 507 RRMSE values (near-zero usually less than 50%) also confirm the accuracy of the mathematical models. 508 According to Table 6, except for EDP2-1 with soil type of S4 which has an approximate RRMSE of 509 54%, EDP1, EDP2-1 and EDP2-2 respectively have RRMSE values less than 36.3%, 44%, and 42.1%, 510 and the remaining EDPs have RRMSE values less than 6.4%, indicating the accuracy of the predicted 511 models. The higher accuracy of predicted EDP3 and EDP4 is evident. The R2, E, d and r2 values near 512 1.0 (e.g. more than 0.8) indicate a good correlation, efficiency and agreement of the predicted values 513 with the exact ones. Based on Table 6, the R2, E, d, and r2 values are more than 0.8 for EDP1, EDP2-514 2, EDP3, and EDP4. The minimum values of R2, E, d, and r2 of EDP2-1are almost equal to 0.68, 0.55, 515 0.83 and 0.65 which belong to the soil type S4 whilst for other soil types almost all of them are more 516 than 0.8. Regarding the PI, that is a combination of R and RRMSE, the values less than 0.3 and 0.2 517 indicate a good and an excellent prediction, respectively. The PI values are less than 0.2 for EDP1, less 518 than 0.3 for EDP2-1, and less than 0.22 for EDP2-2 which indicate a good prediction capability of their 519 corresponding proposed MOGP models. This index is less than 0.04 for both EDP3 and EDP4, 520 confirming the excellent prediction capability of the proposed MOGP models. 521 To make a more informative and general evaluation of the proposed MOGP models, the average 522 values of the performance metrics of all soil types, for each of EDPs, are presented in Table 6. The 523 average results of EDP2 demonstrate that EDP2-2 has a better performance than EDP2-1. In fact, 524 MAPE, RRMSE, and PI of EDP2-2 model have about 39%, 34.5%, and 37% smaller values as well as 525 R2, E, d and r2 have almost 17%, 21%, 5.6%, and 13% larger values as compared to those of EDP2-1. 526 Finally, considering the stochastic nature of earthquake engineering problems and the nonlinear 527 relations governing the inelastic SDOFs behavior, the results indicate the good capability of MOGP 528 prediction models for EDP1, EDP2-1 and EDP2-2, and excellent capability of MOGP prediction 529 models for EDP3 and EDP4. 530 5.6. Comparative study 531 5.6.1. Assumptions 532 In the previous section, it was shown that all the mathematical models precisely predict each EDP of 533 interest. In this section, the accuracy of the proposed mathematical models of all EDPs (EDP1 to EDP4) 534 is compared with those of some available models from the relevant literature. The works presented by 535 Housner (1956), Kuwamura and Galambos (1989), Fajfar and Vidic (1994), Manfredi (2000), Bakhshi 536 and Tavallali (2006), and Dindar et al. (2015) are selected to make the comparison. All the works 537 selected herein have dealt with the inelastic SDOF systems having elastoplastic behavior model ( = 0) 538 and damping ratio () of 0.05. The PGAs of 0.5 and 1.0g and displacement ductility ratios () of 2, 4 539 and 6, are considered for the comparison. The well-known performance metrics including MAPE, 540 RRMSE, R2, PI, E, d, and r2 are used for comparison. To make an informative comparison, all the 541 performance metrics are listed in separated Tables for EDP1, EDP2, (including EDP2-1 and EDP2-2), 542 EDP3 and EDP4 (Tables 7-10 respectively). 543 In order to show the varying trend of the predicted models using MOGP and the best model from 544 literature compared with the exact results, a graphical comparison is also made. As mentioned in the 545 Introduction section and as can be concluded in the above comparison, Manfredi (2000) model is one 546 of the best models presented in the literature. Therefore, this is the only model selected to make a more 547 clear comparison for EDP1, EDP2 (including EDP2-1 and EDP2-2) and EDP3. Moreover, Bakhshi and 548 Tavallali (2006) model is also used for making a comparison for EDP4. In addition, the inelastic SDOF 549 systems having  = 0,  = 6 and  = 0.05 are used and are subjected to earthquake records corresponding 550 to all soil types S1 to S4 with PGAs of 0.5 and 1.0g. The exact and predicted energy-based spectrum 551 using MOGP and another selected model from the literature are shown in Figures 7-10 respectively for 552 EDP1 to EDP4. 553 5.6.2. Comparison 554 The mathematical model of Eq. (12), proposed for prediction of EDP1 using MOGP, is compared with 555 the seismic input energy equation presented by Housner (1956), Kuwamura and Galambos (1989), 556 Manfredi (2000), and Dindar et al. (2015). As shown in Table 7, Eq. (12) and Kuwamura and Galambos 557 (1989) models almost have similar performance for the soil type of S1 which are better than the other 558 models. Although Eq. (12) works very well for the soil type of S2, performance metric values of 559 Manfredi (2000) model, excepting MAPE, show that it works better than Eq. (12) and the other models. 560 Concerning the soil types of S3 and S4, Eq. (12) has an outperformance compared with the other models 561 in total. Based on the average values of performance metrics listed in Table 7, Eq. (12) has lower MAPE, 562 RRMSE and PI, higher R2 and E compared with the other models (and slightly lower d and r2 compared 563 with Manfredi (2000) model), indicating that the predicted model using MOGP totally outperforms the 564 other models available in the literature. The varying trend of the Eq. (12) and Manfredi (2000) model 565 compared with the exact results is also shown in Figure 7. As shown, the varying trend of Eq. (12) is 566 almost conform to the exact results trend excepting a difference for soil type of S1 at medium-periods 567 (see Fig. 7(a)) which is not considerable. This is also true for Manfredi (2000) model except for the 568 long-periods of soil types of S2 and S4 shown in Fig. 7(d). 569 The mathematical models of Eq. (13) and Eq. (16), proposed for prediction of EDP2 using MOGP, 570 are compared with the presented models by Kuwamura and Galambos (1989), Manfredi (2000), and 571 Dindar et al. (2015). The results of the comparison are listed in Table 8. As shown in this table, MOGP-572 2 model (Eq. (16)) significantly outperforms MOGP-1 model (Eq. (13)). Eq. (16) is resulted to lower 573 MAPE, RRMSE and PI, and higher R2, E, d and r2 compared with those of for other models with the 574 exception of RRMSE, PI, R2, E and r2 of soil type of S2 and r2 of soil type of S3 for Manfredi (2000) 575 model and soil type of S1 for Kuwamura and Galambos (1989) which are slightly better than those of 576 for Eq. (16). To make an overall comparison, the average values of the performance metrics are listed 577 in Table 8. These results confirm the superiority of the EDP2 prediction model using MOGP-2 (Eq. 578 (16)) as compared to those from the literature. The varying trend of the Eq. (13), Eq. (16) and Manfredi 579 (2000) models compared with the exact results is also shown in Figure 8. As shown, all models 580 relatively have a consistent varying trend to the exact results except for Eq. (13) at medium- and long-581 periods and for Manfredi model at long-periods of soil types of S1 and S4. Moreover, as shown in Figs. 582 7 and 8, the effect of PGA on both EDP1 and EDP2 is obvious as it is confirmed by their frequencies 583 in MOGP models depicted in Fig. 6(a) and Fig.6 (b). 584 A similar comparison was also carried out for EDP3 prediction model of Eq. (14) using MOGP. 585 The predicting equations presented in the works by Kuwamura and Galambos (1989), Fajfar and Vidic 586 (1994), Manfredi (2000), and Dindar et al. (2015) were selected for comparison. The results of this 587 comparison are shown in Table 9. According to the table, although some of the mentioned works show 588 relatively appropriate results, the performance metrics of the prediction model based on Eq. (14) show 589 lower MAPE, RRMSE and PI, and higher R2, E, d and r2 as compared to the other presented models, 590 except for R2 for soil type S1 of the model used in Manfredi (2000). This indicates the novel capability 591 of MOGP for accurate prediction of the EDP3 as compared with those presented in the literature. To 592 make a general comparison on the performance of the MOGP prediction model of EDP3, the average 593 values of the performance metrics are listed in Table 9. The results show that Eq. (14) highly 594 outperforms the other presented models in the literature. Figure 9 also shows the varying trend of Eq. 595 (14) and Manfredi (2000) model as compared to exact values. As depicted in the figure, the varying 596 trend of Eq. (14) is in good agreement with exact values except for soil type S1 which has inconsiderable 597 differences. Despite the exact values and Eq. (14), the Manfredi (2000) model has a constant trend 598 which has significant differences in the majority of periods. It should be noted that, as depicted in Fig. 599 9, EDP3 is not affected by PGA as evidenced by its very low frequency in MOGP prediction model for 600 EDP3 shown in Fig. 6(c). 601 EDP4 prediction model of Eq. (15) using MOGP was also compared with the Bakhshi and Tavallali 602 (2006) model. The comparative results are shown in Table 10. According to the table, for all soil types, 603 lower MAPE, RRMSE and PI, and higher R2, E, d and r2 is obtained for Eq. (15) with respect to the 604 Bakhshi and Tavallali (2006) model. As shown in the table, average performance metrics are computed 605 for making a general comparison, demonstrating the superiority of Eq. (15) using MOGP in order to 606 predict EDP4 rather than the Bakhshi and Tavallali (2006) model. Figure 10 shows the varying trend 607 of Eq. (15) and Bakhshi and Tavallali (2006) model compared with exact values. The shown graphs 608 confirm the considerable differences between the Bahshi and Tavalali (2006) model and the exact 609 values. In contrast, the MOGP model of Eq. (15) has the closest fit to the exact results. In addition, as 610 shown in this figure, EDP4 is not influenced by PGA as evidenced by its very low frequency in MOGP 611 prediction model for EDP4 shown in Fig. 6(d). 612 It should be noted that most models in the literature are developed for a system with a limited 613 number of SDOF and a low number of earthquake ground motion records. However, the proposed 614 MOGP-based models can deal with SDOF systems with a wide range of structural features subjected 615 to moderate-to-severe earthquake ground motions.. 616 6. Summary and Conclusion 617 Formulation of the seismic energy demand of inelastic SDOF systems is one of the main steps of 618 extending the energy-based seismic analysis and design approach. A comprehensive study was carried 619 out to propose accurate and simple mathematical models for predicting seismic energy demand spectra. 620 Multi-objective genetic programming (MOGP) was employed to formulate some main energy-based 621 EDPs, i.e., spectral input and hysteretic energy, spectral hysteretic to input energy ratio, and spectral 622 energy modification factor. Maximizing the accuracy and minimizing the complexity of the predictive 623 models were considered as two objectives of the multi-objective optimization procedure. Both 624 structural and earthquake characteristics were included in the proposed mathematical models. 625 Regarding each EDP, one equation with different coefficients was proposed for various soil types 626 assuming that different soil types have linear relationships. The frequency of each input variable in the 627 best-generated models was also presented to measure the importance of the variable. 628 Finally, the capability of the proposed models was examined using several common performance 629 metrics. The results indicate the accuracy and effectiveness of the proposed mathematical models in 630 predicting the seismic energy demand spectra compared with some of the models presented in the 631 literature. 632 633 References 634 Akiyama, H., 1985. Earthquake-resistant limit-state design for buildings, The University of Tokyo Press, Tokyo, 635 Japan. 636 Alavi, A.H., Aminian, P., Gandomi, A.H., and Esmaeili, M.A., 2011. Genetic-based modeling of uplift capacity 637 of suction caissons, Expert Systems with Applications 10, 12608–12618. 638 Alıcı, F.S., and Sucuoğlu, H., 2016. Prediction of input energy spectrum: attenuation models and velocity 639 spectrum scaling, Earthquake Engineering and Structural Dynamics 45, 2137–2161. 640 Amiri, G.G., Darzi, G.A., and Amiri J.V., 2008. Design elastic input energy spectra based on Iranian earthquakes, 641 Canadian Journal of Civil Engineering 35, 635–646. 642 ANSI/AISC, 2010. Specification for Structural Steel Buildings, American Institute of Steel Construction, 643 Chicago, IL. 644 Arroyo, D., and Ordaz, M., 2007. On the estimation of hysteretic energy demands for SDOF systems, Earthquake 645 Engineering and Structural Dynamics 36, 2365–82. 646 Babanajad, S.K., and Gandomi, A.H., Mohammadzadeh, D., Alavi, A.H., 2013. Numerical modeling of concrete 647 strength under multiaxial confinement pressures using linear genetic programming, Automation in Construction 648 36, 136–144. 649 Benavent-Climent, A., and Lopez-Almansa, F., and Bravo-Gonzalez, D.A., 2010. Design energy input spectra for 650 moderate-to-high seismicity regions based on Colombian earthquakes, Soil Dynamics and Earthquake 651 Engineering 30, 1129–1148. 652 Bertero, V.V., and Uang, C.M., 1988. Implications of recorded earthquake ground motions on seismic design of 653 building structures, Research Report, UCB/EERC-88/13, University of California at Berkeley, Los Angeles, CA. 654 Cabalar, A. F., and Cevik, A., 2011. Triaxial behavior of sand–mica mixtures using genetic programming. Expert 655 Systems with Applications 38, 10358–10367. 656 Chopra, A.K., 2012. Dynamics of structures: theory and applications to earthquake engineering, Fourth Edition, 657 Prentice Hall, CA. 658 Chou, C.C., and Uang, C.M., 2000 Establishing absorbed energy spectra – an attenuation approach,Earthquake 659 Engineering and Structural Dynamics 29, 1441–1455. 660 Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T., 2002. A fast and elitist multi objective genetic algorithm: 661 NSGA-II, IEEE Transactions on Evolutionary Computation6, 182–197. 662 Decanini, L.D., and Mollaioli, F., 2001. An energy-based methodology for the assessment of seismic demand, 663 Soil Dynamic and Earthquake Engineering 21, 113–137. 664 Deniz, D., Song, J., and Hajjar, J.F., 2017. Energy-based seismic collapse criterion for ductile planar structural 665 frames, Engineering Structures 141, 1–13. 666 Dindar, A.A., Yalçın, C., Yüksel, E., Özkaynak, H., and Büyüköztürke, O., 2015. Development of earthquake 667 energy demand spectra, Earthquake Spectra 31, 1667–1689. 668 Fajfar, P., and Vidic, T., 1994. Consistent inelastic design spectra: hysteretic and input energy, Earthquake 669 Engineering and Structural Dynamics 23, 523–537. 670 Fajfar, P., Vidic, T., and Fischinger, M., 1989. Seismic demand in medium- and long-period structures, 671 Earthquake Engineering and Structural Dynamics 18, 1133–1144. 672 Fajfar. P., and Krawinkler. H., 1992. Nonlinear seismic analysis and design of reinforced concrete buildings, 673 Elsevier Applied Science, New York. 674 Federal Emergency Manage Agency (FEMA), 2000. Prestandard and commentary for the seismic rehabilitation 675 of buildings. FEMA-356, prepared by the American Society of Civil Engineers, Washington, D.C. 676 Ferreira, C., 2006. Gene expression programming: mathematical modeling by an artificial intelligence (Vol. 21). 677 Springer. 678 Gandomi, A. H., and Alavi, A. H., 2011. Multi-stage genetic programming: a new strategy to nonlinear system 679 modeling. Information Sciences 181, 5227-5239. 680 Gandomi, A. H., Babanajad, S. K., Alavi, A. H., and Farnam, Y. (2012). Novel approach to strength modeling of 681 concrete under triaxial compression. Journal of Materials in Civil Engineering 24, 1132–1143. 682 Gandomi, A.H., Alavi, A.H., Mirzahosseini, M.R., and Nejad, F.M., 2010. Nonlinear genetic-based models for 683 prediction of flow number of asphalt mixtures, Journal of Materials in Civil Engineering 23, DOI: 684 10.1061/(ASCE)MT.1943-5533.0000154. 685 Gandomi, A.H., and Alavi, A.H., 2012a. A new multi-gene genetic programming approach to nonlinear system 686 modeling. Part I: materials and structural engineering problems, Neural Computing and Applications 21, 171–687 187. 688 Gandomi, A.H., and Alavi, A.H., 2012b. A new multi-gene genetic programming approach to nonlinear system 689 modeling. Part II: geotechnical and earthquake engineering problems, Neural Computing and Applications 21, 690 189–201. 691 Gandomi, A.H., and Roke, D.A., 2015. Assessment of artificial neural network and genetic programming as 692 predictive tools, Advances in Engineering Software 88, 63–72. 693 http://ascelibrary.org/journal/jmcee7 https://doi.org/10.1061/(ASCE)MT.1943-5533.0000154 Gandomi, A.H., Roke, D.A., and Sett, K., 2013. Genetic programming for moment capacity modeling of 694 ferrocement members, Engineering Structures 57, 169–176. 695 Gandomi, A.H., Sajedi, S., Kiani, B., and Huang, Q., 2016. Genetic programming for experimental big data 696 mining: A case study on concrete creep formulation, Automation in Construction 70, 89–97. 697 Gani, A., Siddiqa, A., Shamshirband, S.,and Hanum, F., 2016. A survey on indexing techniques for big data: 698 taxonomy and performance evaluation, Knowledge and Information Systems 46, 241–284. 699 Gharehbaghi, S., 2018. Damage controlled optimum seismic design of reinforced concrete framed structures, 700 Structural Engineering and Mechanics 65, 53–68. 701 Gharehbaghi, S., and Khatibinia, M., 2015. Optimal seismic design of reinforced concrete structures subjected to 702 time–history earthquake loads using an intelligent hybrid algorithm, Earthquake Engineering and Engineering 703 Vibration 14, 97–109. 704 Gholizadeh, S., and Salajegheh, E., 2009. Optimal design of structures for time history loading by swarm 705 intelligence and an advanced metamodel, Computer Methods in Applied Mechanics and Engineering 198, 2936–706 49. 707 Gupta, A.K., 1990. Response spectrum method in seismic analysis and design of structures. CRC Press, Boca 708 Raton, FL. 709 Hii, C., Searson, D.P., and Willis, M., 2011. Evolving toxicity models using multigene symbolic regression and 710 multiple objectives, International Journal of Machine Learning and Computing 1, 30–35. 711 Housner, G.W., 1956. Limit design of structures to resist earthquakes, in Proceedings, of the First World 712 Conference on Earthquake Engineering, 5, 1–13. 713 Kalkan, E., and Kunnath, S.K., 2007. Effective cyclic energy as a measure of seismic demand effective cyclic 714 energy as a measure of seismic demand, Journal of Earthquake Engineering 11,725–751. 715 Kayadelen, C., Günaydın, O., Fener, M., Demir, A., &Özvan, A. (2009). Modeling of the angle of shearing 716 resistance of soils using soft computing systems. Expert Systems with Applications 36, 11814-11826. 717 Khan, S.A., Shahani, D.T., and Agarwala, A.K., 2003. Sensor calibration and compensation using artificial neural 718 network, ISA Trans 42, 337–352. 719 Khashaee, P., 2004. Energy–based seismic design and damage assessment for structures, Ph.D. Dissertation, 720 Department of Civil Engineering, Southern Methodist University, USA. 721 Khatibinia, M., Gharehbaghi, S., and Moustafa, A., 2015. Seismic reliability-based design optimization of 722 reinforced concrete structures including soil-structure interaction effects, Earthquake Engineering- From 723 Engineering Seismology to Optimal Seismic Design of Engineering Structures, Chapter 11, Moustafa A (ed.), 724 InTech, 267–304. 725 Koza, J. R., 1990. Genetic programming: A paradigm for genetically breeding populations of computer programs 726 to solve problems (Vol. 34). Stanford, CA: Stanford University, Department of Computer Science. 727 Kuwamura, H., and Galambos, T.V., 1989. Earthquake load for structural reliability, Journal of Structural 728 Engineering 115, 1446-1462. 729 Manfredi, G., 2001. Evaluation of Seismic Energy Demand, Earthquake Engineering and Structural Dynamics 730 30, 485–499. 731 MATLAB, 2010. The language of technical computing, Math Works Inc. 732 Metenidis, M.F., Witczak, M., and Korbicz, J., 2004. A novel genetic programming approach to nonlinear system 733 modelling: application to the DAMADICS benchmark problem, Engineering Applications of Artificial 734 Intelligence17, 363–370. 735 Mirzahosseini, M.R., Aghaeifar, A., Alavi, A.H., Gandomi, A.H., and Seyednour, R., 2011. Permanent 736 deformation analysis of asphalt mixtures using soft computing techniques, Expert Systems with Applications 38, 737 6081–6100. 738 Papadopoulos, V., and Giovanis, D.G., Lagaros, N.D., Papadrakakis, M., 2012. Accelerated subset simulation 739 with neural networks for reliability analysis, Computer Methods in Applied Mechanics and Engineering 223–224, 740 70–80. 741 PEER Strong Motion Database. 2017; http://ngawest2.berkeley.edu/. 742 Quinde, P., and Reinoso, E., Terán-Gilmore, A., 2016. Inelastic seismic energy spectra for soft soils: Application 743 to Mexico City, Soil Dynamics and Earthquake Engineering 89,198–207. 744 Sajjadi, S., Shamshirband, S., Alizamir, M., Yee, L., Mansor, Z., Manaf, A.A. et al. 2016, Extreme learning 745 machine for prediction of heat load in district heating systems, Energy and Buildings 122, 222–227. 746 Salajegheh, E., and Heidari, A., 2005. Optimum design of structures against earthquake by wavelet neural network 747 and filter banks, Earthquake Engineering and Structural Dynamic 34, 67挑82. 748 Searson, D., Willis, M., and Montague, G., 2007. Co-evolution of non-linear PLS model components, Journal of 749 Chemometrics 21, 592–603. 750 Searson, D.P., 2014, GPTIPS 2: an open-source software platform for symbolic data mining, (arXiv preprint 751 arXiv:14124690). 752 https://www.infona.pl/resource/bwmeta1.element.elsevier-14dffa12-d307-3f3e-a661-a4637cd14bf6/tab/jContent https://www.infona.pl/resource/bwmeta1.element.elsevier-14dffa12-d307-3f3e-a661-a4637cd14bf6/tab/jContent http://www.sciencedirect.com/science/article/pii/S0045782512000552 http://www.sciencedirect.com/science/article/pii/S0045782512000552 http://ngawest2.berkeley.edu/ Searson, D.P., Leahy, D.E., and Willis, M.J., 2010. GPTIPS: an open source genetic programming toolbox for 753 multigene symbolic regression, in Proceedings of the International Multiconference of Engineers and Computer 754 Scientists: Citeseer , 77–80. 755 Sucuoğlu, H., and Nurtuğ, A., 1995. Earthquake ground motion characteristics and seismic energy dissipation, 756 Earthquake Engineering and Structural Dynamics 24, 1195–1213. 757 Tsompanakis, Y., and Topping, B.H.V., 2011. Soft computing methods for civil and structural engineering, Saxe-758 Coburg Publications, Stirlingshire, UK. 759 Uang, C.M., and Bertero, V.V., 1988. Implications of recorded earthquake ground motions on seismic design of 760 building structures, Earthquake Engineering Research Center , Report No. UCB/EERC-88/13, University of 761 California, Berkeley. 762 Uang, C.M., and Bertero, V.V., 1990. Evaluation of seismic energy in structures, Earthquake Engineering and 763 Structural Dynamics 19, 77–90. 764 Vardhan, H., Garg, A., Li, J., and Garg, A., 2016. Measurement of stress dependent permeability of unsaturated 765 clay, Measurement 91, 371–376. 766 Walter, E., and Pronzato, L., 1997. Identification of parametric models: from experimental data , 767 Communications and Control Engineering, Springer, 413 pp. 768 Yazdani, H., Khatibinia, M., Gharehbaghi, S., and Hatami, K., 2016. Probabilistic optimum seismic design of 769 reinforced concrete structures considering soil-structure interaction effects, ASCE-ASME Journal of Risk and 770 Uncertainty in Engineering Systems, Part A: Civil Engineering 3, G4016004–1–12; DOI: 771 10.1061/AJRUA6.0000880. 772 Zahra, T.F., and Hall W.J., 1984. Earthquake energy absorption in SDOF structures, Journal of Structural 773 Engineering 110, 1757–1772. 774 Zhai, C., Ji, D., Wen, W., Lei, W., Xie, L., and Gong, M., 2016. The inelastic input energy spectra for main shock–775 aftershock sequences Earthquake Spectra 32, 2149–2166. 776 A HYBRID COMPUTATIONAL APPROACH FOR SEISMIC ENERGY DEMAND PREDICTION S. Gharehbaghia, A.H. Gandomib, , and S. Achakpourc, M.N. Omidvard 1. Introduction 2. Seismic Energy Concept and Formulation 3. Genetic Programming 3.1. Multi-Gene Symbolic Regression 4. Predicting Seismic Energy Demand Spectra Using MOGP 5. Results and Discussion 
work_rjehwmoalvgknnzgl5b2jfpziq	----	Kernel methods for short-term portfolio management Kernel methods for short-term portfolio management Huseyin Ince a , Theodore B. Trafalis b,* a School of Business Administration Studies, Gebze Institute of Technology, Çayirova Fab., Yolu, No: 101 P.K:141 41400, Gebze/Kocaeli, Turkey b School of Industrial Engineering, The University of Oklahoma, 202 West Boyd, Ste 124, Norman, OK 73019, USA Abstract Portfolio optimization problem has been studied extensively. In this paper, we look at this problem from a different perspective. Several researchers argue that the USA equity market is efficient. Some of the studies show that the stock market is not efficient around the earning season. Based on these findings, we formulate the problem as a classification problem by using state of the art machine learning techniques such as minimax probability machine (MPM) and support vector machines (SVM). The MPM method finds a bound on the misclassification probabilities. On the other hand, SVM finds a hyperplane that maximizes the distance between two classes. Both methods prove similar results for short-term portfolio management. q 2005 Elsevier Ltd. All rights reserved. Keywords: Support vector machines; Minimax probability machine; Kernel methods; Portfolio management; Earning announcements 1. Introduction The objective of this study is to develop a model for stock selection by using state of the art machine learning techniques, such as minimax probability machine (MPM) and support vector machines (SVMs). Stock selections decisions are derived from earning announcements and volatility of stocks around the earning time. The majority of the studies on stock markets and earning announcements focus on the information content and timeliness- measured by changes in the character- istics of stock return distribution (e.g. mean, variance, and serial correlation) and trading volume—when corporate earn- ings are announced to the market (Eilifsen, Knivsflae, & Sættem, 2001). Numerous studies, including Ball & Brown (1968), Chari, Jagannathan, and Ofer (1988), Easton & Zmijewski (1989), Gennotte & Truemann (1996), and Kross & Schroeder (1984), find that stock prices respond positively to announcements of increase in earnings and negatively to announcements of decrease in earnings for the USA firms. Beaver (1968) argued that earnings announcements possess information content if stock price volatility and/or trading volume increase around the time of the announcement (see also Atiase, 1985; Bamber, 1987; Bamber, Barron, & Stober, 1997; Barron, 1995; Holthausen & Verrecchia, 1990). According to this view, stock prices reflect the market’s aggregated (or average) interpretation of the information, while trading volume measures investor activity, and thus reflects differential beliefs among investors. Three previous studies have examined short-term price movements around earnings announcements—Chari et al. (1988), covering the 1976–1984 period, Ball & Kothari (1991), covering the years 1980–1988, and Trueman et al. (2003) covering January 1998 to August 2000. There are two primary differences between the findings of the first two studies and the findings of Trueman et al. (2003). First, while the stocks in the first two studies’ samples do increase in price prior to earnings announcements, there is no consistent price movement (either positive or negative) afterwards. Second, the magnitude of the pre-announcement returns they found is very small in comparison to what the third study reports (less than one tenth in size). Furthermore, those returns are significant only for the day before and the day of the earnings announcement. Quarterly earnings announcement is probably one of the most highly anticipated events and receives significant media and investor attention. Research consistently shows that the market assimilates more and more of the information in earnings as the announcement date approaches (see Kothari, 2001; for a survey of this research starting with Ball & Brown, 1968). Hence, the extent to which an earnings announcement will provide ‘useful’ information to market participants should be a function not only of the nature of information released but Expert Systems with Applications 30 (2006) 535–542 www.elsevier.com/locate/eswa 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.10.008 * Corresponding author. Tel.: C1 405 325 3721; fax: C1 405 325 7555. E-mail addresses: h.ince@gyte.edu.tr (H. Ince), ttrafalis@ou.edu (T.B. Trafalis). http://www.elsevier.com/locate/eswa also when it is released. Firms tend to release good news earnings reports earlier than those containing bad news (Chambers & Penman, 1984; Kross & Schroeder, 1984). If this is the case, then the longer a firm goes without reporting its earnings, the less positive (or, alternatively, more negative) will investors expect the ultimate earnings news to be. This implies that a firm should experience negative stock returns from the end of its fiscal quarter up until the time of its earnings announcement. When the earnings are finally released by the manager, the stock price should rise. Based on these studies, we propose a classification model with two classes; ‘buy a certain stock whose earning/eps is higher than some threshold’ and ‘do not buy a certain stock if its earning is less than some certain threshold’. The threshold value is subjective and depends on the investor. As far as we know, there have been no studies that propose a classification model by using the earning announcements and anomalies. Data mining techniques such as SVMs, MPM and multilayer perceptron (MLP) would be good candidates addressing this problem instead of using classical statistical techniques. Data mining is the process of discovering and analyzing hidden patterns in data sets. It has wide application areas from engineering to commerce. Most widely used data mining techniques in finance are artificial neural networks (ANNs), SVMs and their variations. Bankruptcy prediction, portfolio management (choosing individual stocks for portfolio), option pricing, and forecasting indices or individual stock prices are typical application areas in finance (Chen, Leung, & Daouk, 2003; Dourra & Siy, 2002; Galindo, 1998; Hutchinson, Lo, & Poggio, 1994; Trafalis & Ince, 2000). The classification problem is to learn a discriminating hyperplane or hypersurface, which separates two classes from examples by minimizing the generalization error (Vapnik, 1995). MPM attempts to control misclassification probabilities for two-class classification problems. Specifically it minimizes the worst case (maximum) probability of misclassification of future data points under all possible choices of class densities with a given mean and covariance matrix, which is positive definite. MPM minimizes directly an upper bound on the generalization errors (Lanckriet, El Ghaoui, Bhattacharyya, & Jordan, 2002). In contrast to MPM, SVMs achieve low generalization error by maximizing an associate quantity termed margin that describes the distance between two classes. In this study, we will provide a framework for short-term portfolio selection problem by using MPM and SVMs. This can be done by analyzing the behavior of the stocks around their earning announcement. We believe that some variables such as volume, difference between expected eps and actual eps, and expectation have big effect on return. Several studies have shown that the stock market is inefficient during the earning announcement in terms of individual stocks. The paper is organized as follows: in Sections 2 and 3 we present the basics of the MPM and SVM techniques, respectively; in Section 4, a short-term portfolio classification model is explained; computational results are given in Section 5; and finally, Section 6 concludes the paper. 2. Minimax probability machine The MPM attempts to control misclassification probabilities for two-class (binary) classification problems. It minimizes the worst case (maximum) probability of misclassification of future data points under all possible choices of class densities with a given mean and covariance matrix (Lanckriet et al., 2002). MPM uses the following theorem, by Popescu & Bertsimas (2001): sup ywð�y; P yÞ Prfy2Sg Z 1 1 Cd2 ; and; d 2 Z inf y2S ðyK�yÞT XK1 y ðyK�yÞ; (1) where y is a random vector, S is a given convex set, and supremum is taken over all distributions for y having mean �y and covariance matrix Sy, assumed to be positive definite. Let x and y denote random vectors in a binary classification problem with means and covariance matrices given by ð�x; P xÞ andð�y; P yÞ. Note that x and y also denote the two classes. MPM seeks a hyperplane H(w,b)Z{zjwTzZb} where w 2 R nnf0g and b 2R, which separates the two classes of points with maximal probability with respect to all distributions having these means and covariance matrices. It minimizes the generalization error by finding the hyperplane for which the misclassification probabilities Pr(w T x%b) and Pr(wTy%b) are minimized. The optimization problem becomes max a;ws0;b a (2) s.t. inf xwð�x; P x Þ Prðw T xR bÞR a; inf ywð�y; P y Þ Pr w T y% b � � R a The quantity (1Ka) can be interpreted as an upper bound on the generalization error. Furthermore, this quantity is the worst-case (maximum) misclassification probability. Infimum in both constraints are computed via a theorem stated in Popescu & Bertsimas (2001) given in (1). Problem (2) can be restated as max k;ws0 k (3) s.t. w T ð�xK�yÞR k ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi wT X x w r C ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi wT X y w s0@ 1 A The problem (3) can be simplified as (Lanckriet et al., 2002): min w k X 1=2 x wk2 C k X 1=2 y wk2 (4) s.t. w T ð�xK�yÞ Z 1 H. Ince, T.B. Trafalis / Expert Systems with Applications 30 (2006) 535–542536 https://isiarticles.com/article/21688 
work_rjygsln4abck3e5ise2c6pfo3u	----	PII: 0888-613X(93)90013-4 Consistency Checking for Fuzzy Expert Systems K. S. Leung and Y. T. So T h e C h i n e s e University o f H o n g K o n g ABSTRACT Inconsistency frequen@ exists in a rule-based expert system. Detecting the existence o f inconsistency in a fuzzy rule-based environment is difficult and may be different from that o f traditional rule-based systems. A n affinity measure, which is based on the similarity measure, is introduced to determine the likeness o f two fuzzy terms. By using the affinity measure, the techniques f o r consistency checking in a non-fuzzy environment can be easily applied to a fuzzy environment. A consistency checker (CCFE) is implemented to detect possible inconsistency in a mixed fuzzy and non-fuzzy environment. K E Y W O R D S : E x p e r t s y s t e m , f u z z y theory, i n c o n s i s t e n c y , possibility distribution I. INTRODUCTION In the development of a rule-based expert system, the knowledge engineering process is an iterative one. The knowledge base needs to be refined many times. When new rules are added or the existing rules are altered, inconsistency in the knowledge base frequently occurs. Although H a y e s - R o t h [1] pointed out that a missing key feature of rule-based systems is "a suitable verification methodology or a technique for testing the consistency of completeness of a rule set," many software tools such as TERIRESIAS [2], ONCOCIN's rule-checker [3], INSPEC- T O R [4], and ESC [5] have been developed to identify inconsistencies in non-fuzzy knowledge bases. The consistency checking tools help both the domain experts and knowledge engineers to build the expert systems more easily, accurately, and quickly. However, checking inconsistency in a fuzzy rule-based environment [6-8] is not as simple as that in a non-fuzzy environment and may also be different from that of traditional rule-based systems. Let "height" be a Address correspondence to K. S. Leung, Department of Computer Science, The Chinese University of Hong Kong, Shatin, N. T., Hong Kong. Received April 30, 1991; accepted March 12, 1993. International Journal of Approximate Reasoning 1993; 9:263-282 © 1993 S. K. Leung 263 264 K.S. Leung and Y. T. So fuzzy object having three fuzzy terms (tall, medium, short) as its possible values. I f the user enters the following two rules: R u l e 1" I F height is tall T H E N C R u l e 2: I F height is short T H E N C D o e s r e d u n d a n c y occur? Consider a n o t h e r case: R u l e 3: I F A T H E N height is more-or-less tall R u l e 4: I F A T H E N height is n o t very tall D o e s conflict occur? Thus, it would be b e t t e r if we could apply a quantitative m e a s u r e on the two fuzzy terms (e.g., tall, not short) in o r d e r to d e t e r m i n e to what extent these two fuzzy terms m e a n the same thing. A n "affinity" m e a s u r e based on the similarity m e a s u r e [6] is c r e a t e d to d e t e r m i n e the degree o f matching between two fuzzy expressions. Using the affinity measure, the techniques for consistency check in a non-fuzzy e n v i r o n m e n t can be transported to a fuzzy environment. T h e following section summarizes different kinds of inconsistencies that exist in a rule-based environment. Section 3 introduces the similarity and affinity measures. Section 4 describes a consistency checker, C C F E , imple- m e n t e d for fuzzy rule-based expert systems. Its results in detecting incon- sistency i n a fuzzy e n v i r o n m e n t are illustrated by an example and discussed in the fifth section. 2. I N C O N S I S T E N C Y I N R U L E - B A S E D S Y S T E M S T h e r e are several kinds o f inconsistencies in a non-fuzzy rule-based e n v i r o n m e n t [9-12]. T h e s u m m a r y is given as follows: L e t us consider the rules: R u l e 1: I F A 1 T H E N B1 (CF1) R u l e 2: I F A 2 T H E N B 2 ( C F 2 ) where A 1 and A 2 m a y be any combination of propositions, B1 and B 2 are single propositions, and CF1 and CF2 are the certainty factors, ranging f r o m 0 to 1, of rule 1 and rule 2 respectively. 2.1. R e d u n d a n t R u l e s • Two rules succeed in the same situation and have the same results. i.e., i, A 1 = A 2 & B1 = B 2 & sign(CF1) = sign(CF2) or ii, A 1 = A 2 & B1 ~ B2 & sign(CF1) ~ sign(CF2) • T h e two rules m a y cause the same i n f o r m a t i o n to be c o u n t e d twice, leading to e r r o n e o u s increases in the certainty factor o f the conclusion. Consistency Check for Fuzzy Expert Systems 265 • C o n d i t i o n ii above m a y n o t be r e d u n d a n t if the expert wants to confirm o n e value (e.g., B 1 with C F 1 > 0) and at the same time disconfirm a n o t h e r value (e.g., B 2 with C F 2 < 0) of the conclusion. 2.2. Conflicting Rules (Contradiction) • Two rules succeed in the same situation b u t with conflicting results. i.e. i, A 1 = A 2 & B 1 = B 2 & s i g n ( C F 1 ) --/: s i g n ( C F 2 ) or ii, A 1 = A 2 & B 1 --/: B 2 & s i g n ( C F 1 ) = s i g n ( C F 2 ) • It is a c o m m o n occurrence in rule sets. However, it m a y cause no problems (inconsistency) because the expert m a y want to conclude different values with different certainty factors. • F o r example, the following are the rules extracted from a medical expert system E S R O M [8]: (Rule m l 7 a I F (((mdcs is anycd) O R (mdcs is gmd)) A N D (gestation > = 30) A N D (gestation < = 34)) T H E N m a n a g e m e n t is delivery Certainty is 0.7 (Rule m18 I F (((mdcs is anycd) O R (mdcs is gmd)) A N D (gestation > = 30) A N D (gestation < = 34) T H E N m a n a g e m e n t is observation Certainty is 0.3 M a n a g e m e n t is a single-valued object with expected values (delivery, observation). T h e two rules listed above have the same a n t e c e d e n t part but with conflicting conclusions. However, they are set on pur- pose by the expert and by no m e a n s a mistake. 2.3. Subsumed Rules • Two rules have the same result, b u t one contains additional restric- tions on the situation in which it will succeed. W h e n e v e r the m o r e restrictive rule succeeds, the less restrictive rule also succeeds, resulting in redundancy. i.e., ( A 1 c A 2 or A 2 c A 1 ) & B 1 = B 2 or ( A 1 c A 2 or A 2 c A 1 ) & B 1 4~ B 2 • e.g., in E S R O M (Rule i o l I F (diagnosis is u n r u p t ) T H E N cx is uninf) Certainty is 0.8 & s i g n ( C F 1 ) = s i g n ( C F 2 ) & s i g n ( C F 2 ) ~ s i g n ( C F 1 ) (Rule iil0 I F ((diagnosis is u n r u p t ) A N D (ctg is reactive)) T H E N cx is uninf) Certainty is 0.95 W h e n e v e r rule iil0 is triggered, rule i o l will also be triggered. Thus, there is redundancy. 266 K.S. Leung and Y. T. So • H o w e v e r , the k n o w l e d g e e n g i n e e r m a y wan t t o write these kinds o f rules so t h a t t h e m o r e restrictive rules will ad d weight to t h e conclu- sions. Thus, t h e experts should b e w a r n e d an d r e q u e s t e d t o clarify t h e i r meanings. In the above example, if the e x p e r t ' s i n t e n t i o n is to give m o r e weight to "cx is u n i n f " w h e n "ctg is r e a c t i v e " is also true, t h e n t h e c e r t a i n t y f a c t o r o f rule i i l 0 should b e c h a n g e d t o 0.75. If b o t h "diagnosis is u n r u p t " and "ctg is r e a c t i v e " are true, t h e cert ai n t y o f conclusion "cx is u n i n f " b e c o m e s 0.95 a f t e r b o t h rules are fired an d the e v i d e n c e c o m b i n a t i o n calculation is app l i ed [6]. 2.4. Sub-contradiction Rules It is a special case o f c o n t r a d i c t i o n . T w o rules have d i f f e r e n t results, b u t o n e contains additional restrictions o n the situation in which it will succeed. W h e n e v e r the m o r e restrictive rule succeeds, t h e less restrictive rule also succeeds: i.e., ( A 1 c A 2 o r A 2 c A 1 ) & B 1 4 : B 2 & s i g n ( C F 1 ) = s i g n ( C F 2 ) o r ( A 1 c A 2 o r A 2 c A 1 ) & B 1 = B 2 & s i g n ( C F 2 ) 4: s i g n ( C F 1 ) 2.5. Unnecessary IF Conditions • T w o rules have the same conclusion, an I F co n d i t i o n in o n e rule is in conflict with an I F c o n d i t i o n in the o t h e r rule, an d all o t h e r I F conditions in t h e two rules are equivalent. e.g., ( A 1 = p / x q ) , ( A 2 = p A -1 q), B 1 = B 2 & C F 1 = C F 2 • T h e e x a m p l e described above actually indicates t h a t only o n e rule is necessary. T h e s e c o n d I F conditions (q a n d -1 q ) o f b o t h rules are unnecessary. A l t h o u g h it will n o t cause any e r r o r in consultation, it would be b e t t e r to i n t e g r a t e the two rules into one: I F p T H E N B I ( C F 1 ) • e.g., ( A 1 = p A q ) , ( A 2 = ~ q ) , B 1 = B 2 & CF1 = C F 2 T h e s e c o n d I F c o n d i t i o n in the first rule is unnecessary, a n d t h e two rules could b e c o m b i n e d into: I F p V -1 q T H E N B I ( C F 1 ) • I f C F 1 4: C F 2 , n o u n n e c e s s a r y I F c o n d i t i o n occurs b e c a u s e t h e ex p ert m a y want to c o n c l u d e a value with differen t cert ai n t y factors in d i f f e r e n t situations, e.g., in E S R O M ( R u l e m20 I F (((wcc > 10) A N D (wcc < 15)) O R ((crp > 20) A N D (crp < 40) a n d (gestation > = 32))) T H E N m a n a g e m e n t is delivery) C e r t a i n t y is 0.8 ( R u l e m21 I F (((wcc > 10) A N D (wcc < 15)) O R ((crp > 2 0 ) A N D (crp < 40) an d (gestation < 32))) T H E N m a n a g e m e n t is delivery) C e r t a i n t y is 0.5 Consistency Check for Fuzzy Expert Systems 267 T h e c o n d i t i o n (gestation > = 32) in rule m20 is conflicting with t h e c o n d i t i o n (gestation < 32) in rule m21. H o w e v e r , t h e e x p e r t wants to assign d i f f e r e n t c e r t a i n t y factors to d i f f e r e n t situations. 2.6. C i r c u l a r R u l e s A set o f rules is circular if the chaining o f these rules in t h e set fo rm s a cycle. i.e., A1 = B2 & A 2 = B2 & sign(CF1) -- sign(CF2) 2.7. S e l f - r e f e r r i n g R u l e s T h e c o n d i t i o n and conclusion clauses o f a rule r e f e r to t h e sam e subject. e.g., I F A = 0 T H E N A = 1 (for A is n o t a multi-valued object) 2.8. I n c o n s i s t e n t I F - C l a u s e T h e clauses in the c o n d i t i o n are c o n t r a d i c t o r y to each o t h e r . e.g., I F A = 1 and A = 2 T H E N B (for A is n o t a multi-valued object) In 2.7 and 2.8, if A is a multi-valued object, it c a n n o t b e c o n c l u d e d that inconsistency occurs b e c a u s e a multi-valued object can have m o r e t h a n o n e value at anytime T h e above discussion only m e n t i o n s superficial inconsistency b e t w e e n two rules. H o w e v e r , inconsistency m a y arise a f t e r a s e q u e n c e o f infer- ring steps, e.g., R u l e l : I F A T H E N B (CF1 > 0 ) R u l e 2: I F B T H E N C (CF2 > O) R u l e 3: I F C T H E N D (CF3 > O) R u l e 4: I F A T H E N D (CF4 > O) R e d u n d a n c y occurs b e t w e e n rule set ( 1 - 3 ) an d rule 4. M o r e o v e r , the d e t e c t i o n o f circular-rule chains is a f f e c t e d by t h e threshold. T h e c e r t a i n t y factors m a y cause a circular chain o f rules to b e " b r o k e n " if the c e r t a i n t y f a c t o r o f conclusion falls below t h e t h r e s h o l d (e.g., 0.2). e.g., R u l e 1: R u l e 2: R u l e 3: R u l e 4: I F A T H E N B (CF = 0.4) I F B T H E N C (CF = 0.7) I F C T H E N D (CF = 0.7) I F D T H E N A (CF = 0.8) As (0.4)(0.7)(0.7) = 0.19 < 0.2, the circular-rule chain is b r o k e n . 268 K.S. Leung and Y. T. So 3. C O N S I S T E N C Y C H E C K S I N A F U Z Z Y E N V I R O N M E N T First, let us consider the possibility, necessity, and similarity measures in possibility theory, and t h e n the affinity m e a s u r e of two fuzzy terms will be introduced. L e t the fuzzy terms be r e p r e s e n t e d by a list o f 11 real numbers that are grades o f m e m b e r s h i p o f the points on an imaginary psychological c o n t i n u u m with an interval scale. The fuzzy terms " t a l l , " " m e d i u m , " and " s h o r t " are r e p r e s e n t e d as follows: tall: 0 0 0 0 0 0 .0280 .1176 .3277 .6561 1 medium: 0 0 0 .0588 .328 1 .328 .0588 0 0 0 short: 1 .6561 .3277 .1176 .0280 0 0 0 0 0 0 H e d g e s like " v e r y " or " m o r e - o r - l e s s " may be applied to the above fuzzy terms to f o r m fuzzy expressions such as " v e r y tall" and "more-or-less short." A square and a square root functions are the fuzzy concepts handling functions for the hedges " v e r y " and " m o r e - o r - l e s s " respectively [13]. The modified representations are shown: very tall: 0 0 0 0 0 0 .0008 .0138 .1074 .4305 1 more-or-less short: 1 .8100 .5725 .3429 .1673 0 0 0 0 0 0 T h e m e m b e r s h i p values described above can be viewed as a possibility distribution [14]. But this distribution has only an indicative or subjective m e a n i n g and m a y n o t be i n t e r p r e t e d directly. Figures 1-3 depict the curves o f the above fuzzy t e r m s / e x p r e s s i o n s . Note that the curves are normalized and convex. 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 v e r y 0 . 1 0 I I I I ; , , , , Figure 1. Possibility distribution for the concepts "tall" and '"very tall." Consistency Check for Fuzzy Expert Systems 269 1 0.9 ~ x 08 m o r e o r l e s s 0.7 0.6 0.5 0.4 0.3 0.2 0 ~ I I I Figure 2. Possibility distribution for the concepts '"short" and "more-or-less" short. 3.1 Possibility and Necessity T h e f o r m u l a e o f t h e possibility ( P ) and necessity ( N ) m e a s u r e s b e t w e e n a fuzzy d a t u m and a fuzzy p a t t e r n are given as follows [15]: P ( F [ F ' ) = max(min(~F(W), I.LF,(W))) N ( F I F ' ) = 1 - e ( ~ F I E ' ) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 m e d i u m , i Figure 3. Possibility distribution for the concept "medium." 270 K.S. Leung and Y. T. So w h e r e P(FIF') is t h e possibility o f the fuzzy d a t u m F ' given the fuzzy p a t t e r n F , and N(F[F') is the necessity o f the fuzzy d a t u m F ' given t h e fuzzy p a t t e r n F . /~(w) is the m e m b e r s h i p function o f w in the universe o f discourse, an d ~ F is the c o m p l e m e n t o f F . T h e possibility m e a s u r e s the d e g r e e o f overl ap p i n g b e t w e e n t h e d a t u m and the p a t t e r n ; o r in o t h e r words, it m e a s u r e s t h e n o n - e m p t i n e s s o f F A F ' . F o r example: P(tall~very tall) = 1 P ( s h o r t l m o r e - o r - l e s s short) = 1 P ( t a l l l m e d i u m ) = 0.028 P ( s h o r t l t a l l ) = 0 Possibility is an optimistic m e a s u r e : it only m e a n s s o m e t h i n g is possi- ble b u t d o e s n o t g u a r a n t e e it will h a p p e n . O n t h e o t h e r hand, necessity is a drastic m e a s u r e t h a t is equal to the impossibility b e t w e e n two o p p o s i t e events. F o r example: N(tall~very tall) = 0.6561 N ( s h o r t l m o r e - o r - l e s s short) = 0.4275 N ( t a l l l m e d i u m ) = 0 N(shortltall) = 0 U n d e r n o r m a l circumstances, the necessity has t h e following p r o p e r t i e s : I F N(FIF') > 0.5 T H E N F ' has a m o r e c o n c e n t r a t e d o r n a r r o w e r distribution t h a n F. e.g., tall and very tall. I F N(FIF') < 0.5 T H E N F ' has a m o r e dilated o r b r o a d e r distribution t h a n F . e.g., short and m o r e - o r - l e s s short, tall a n d m e d i u m . I F N(FIF') = 0.5 T H E N F has the same distribution as F. Consistency Check for Fuzzy Expert Systems 271 Also f r o m the definition o f b o t h possibility a n d necessary, it can be shown that: i, P ( A I B ) = P ( B I A ) .... (1) ii, P(AIB u C ) - - max(P(AIB), P ( A I C ) ) .... ( 2 ) iii, N(AIB U C ) -- m i n ( N ( A I B ) , N(AIC)) .... (3) iv, N ( A I A ) > = 0.5 .... (4) 3.2. Similarity T h e similarity M is calculated by the following algorithm [6]: I F N ( F I F ' ) > 0.5 T H E N M = P ( F [ F ' ) E L S E M = ( N ( F I F ' ) + 0.5)*P(FIF') where * d e n o t e s multiplication. T h a t m e a n s w h e n necessity is greater t h a n or equal to 0.5, the similar- ity b e t w e e n p a t t e r n and d a t u m is s a t u r a t e d and is forced to equal the possibility. I f the similarity is n o t s a t u r a t e d (necessity is less t h a n 0.5), the similarity should d e p e n d o n b o t h possibility and necessity. T h e similarity m e a s u r e is proven to be a useful tool to d e t e r m i n e the similarity b e t w e e n the fuzzy d a t a in a fact base a n d the fuzzy patterns in the premise part of a rule [8]. However, t h e r e is o n e shortcoming in similarity; it is n o t commutative, i.e, M ( p l q ) ~ M(qlp). F o r example, P ( n o t mediumltall) = 1 and P ( m e d i u m l t a l l ) = 0.0588 N ( n o t mediumltall) = 1 - 0.0588 = 0.9412 T h e r e f o r e , M ( n o t m e d i u m ltall) = 1. But P(talllnot m e d i u m ) = 1 and P ( n o t talllnot m e d i u m ) = 1 =* N(talllnot m e d i u m ) = 1 - 1 = 0 T h e r e f o r e , M(talllnot m e d i u m ) = 0.5. H e n c e , the o r d e r o f rule checking will influence the value o f similarity. Owing to this limitation, the affinity ( A ) m e a s u r e is introduced. 3.3. A f f i n i t y T h e Affinity of two simple propositions p and q is defined as follows: A ( p , q ) = M ( p A q[p v q) where Izp ̂ q(W) = m i n ( t z p ( w ) , tZq(W)), tzp v q(W) = max(tZp(W), tZq(W)) a n d /z(w) is the m e m b e r s h i p function o f w in the universe o f discourse. 272 K.S. Leung and Y. T. So A(p, q) measures the similarity o f p v q given p A q. T h a t is, to what extent does the whole part o f b o t h p a n d q ( p v q) m a t c h the shared part o f t h e m ( p A q). T h e following are some properties o f the affinity measures: i, I f p and q are identical, A(p, q) will be equal to 1 (because p V q = p A q ) . ii, A(p, q) is commutative, i.e., A(p, q) = A(q, p). B e c a u s e p A q = q A p a n d p V q = q V p . iii, A(p, q) = min(M(plq), M(qlp)) .... (5) P r o o f o f a s s e r t i o n (iii): Consider the possibility first P ( p A q ] p V q) =max(min(IXp^q(W),lXpvq(W))) = max(I.£pAq(W)) = max(min( i.tp(w), tXq(W))) = P ( p l q ) = P ( q l p ) and Now consider Similarly, N ( p A qlp V q) = m i n ( N ( p A qlp), N ( p A q l q ) ) N ( p A qlp) = l - - P ( ~ ( p A q ) I p ) = 1 - P ( p [ ~ ( p A q ) ) = 1 - P ( p l ~ p V ~ q ) = 1 - m a x ( P ( p [ ~ p), P(p[ ~ q)) = 1 - m a x ( P ( ~ pip), P( ~ q l p ) ) = min(N(p[p), N(q[p)) N ( p A qlq) = min(N(qlq), N ( p l q ) ) [by eqn (1)] .... (6) [by eqn (3)] [by eqn (1)] [by eqn (2)] [by eqn (1)] Consistency Check for Fuzzy Expert Systems 273 So t h a t N ( p /x qlp v q) = m i n ( N ( p l p ) , N(p]q), N(q]p), N(qlq)) .... (7) By e q n (4), it can be c o n c l u d e d that N ( p l p ) an d N(qlq) are b o t h g r e a t e r t h a n o r e q u a l to 0.5. Now, we c o n s i d e r all f o u r cases: case 1: If b o t h N(p[q) and N(qlp) > = 0.5, t h e n M(plq) = P(plq) = P(q[p) = M(qlp) = min(M(plq), M(qlp)) [by e q n (6)] and N ( p A qlP V q > = 0.5 [by e q n (7)] T h u s Z ( p , q ) = M ( p A qlp V q) = P ( p A qlp V q) = e ( p l q ) = P(qlp) = min(M(plq), M(qlp)) [by e q n (6)] Thus, the assertion ( e q n (5)) holds. case 2: If N(plq) < 0.5 and N(qlp) > 0.5, t h e n N ( p A qlp V q) = N(plq) < 0.5 and A ( p , q) = (N(p[q) + 0.5)*P(p[q) = M(plq) ( m i n i m u m o f M(plq) and M(q[p)) case 3: If N(qlp) < 0.5 and N(p[q) > 0.5, t h e n N ( p A qlP v q) = N(qlp) < 0.5 and A ( p , q) = (N(qlp) + 0.5)*P(q[p) = M(q]p) ( m i n i m u m o f M(pJq) and M(q[p)) case 4: If b o t h N(plq) and N(qlp) < 0.5, t h e n N ( p A qlP V q) = min(N(p[q), N(qlp)) So A ( p , q) = M ( p A qlp v q) = ( N ( p m qlp v q) + 0 . 5 ) * e ( p m qlp v q) = (min(N(plq), N(qlp)) + 0.5)*P(plq) = min(M(plq), M(qlp)) T h e r e f o r e , A(p, q) = min(M(plq), M(q]p)) holds fo r all cases. [by eqns (6) & (7)] 274 K.S. Leung and Y. T. So I m p l e m e n t i n g A ( p , q ) using o u r definition (ie., A ( p , q ) = M ( p A ql p v q)) will result in two comparisons, two additions, and one multiplica- tion less t h a n that o f using A ( p , q) = min(M(plq), M(qlp)) each time. H e n c e the f o r m e r is actually used for i m p l e m e n t a t i o n in o u r system. A m o n g the above properties, the c o m m u t a t i v e o n e is the most impor- tant because it implies t h a t the o r d e r o f rule-checking has no influence on the value of the affinity. In o r d e r to investigate the suitability o f affinity in consistency checks, we c o m p u t e all the results o f the affinity for different values o f p and q on a fuzzy object such as " h e i g h t . " H e d g e s like " v e r y " and " m o r e - o r - l e s s " and logical n e g a t i o n " n o t " are a d d e d to the possible fuzzy terms o f the object. So the values of p and q could be [not] [very Imore-or-less] {tall Imedium Ishort} where the choice inside the [ ] is optional. (Although the expressions like "more-or-less m e d i u m " or " v e r y m e d i u m " are n o t very meaningful, we include t h e m to complete the investigation.) T h e results o f the affinity o f any two fuzzy expressions p and q can be divided into four groups. (1) A ( p , q ) = 1 e.g., A(tall, tall) = 1 A(very m e d i u m , m e d i u m ) = 1 (2) A ( p , q) > 0.5 e.g., A(tall, very tall) = 0.9305 A(more-or-less short, short) = 0.9275 A(more-or-less m e d i u m , very m e d i u m ) = 0.9273 (3) A ( p , q ) = 0.5, P ( p /x qlp v q) = 1 a n d N ( p A q[p v q) = 0 e.g., A(tall, n o t m e d i u m ) = 0.5 A ( n o t short, m e d i u m ) = 0.5 A ( n o t very tall, more-or-less short) = 0.5 (4) A ( p , q ) < 0.5, P ( p A qlp V q) < 1 and N ( p A qlp V q) = 0 e.g., A(tall, n o t tall) = 0.1720 A(more-or-less tall, m e d i u m ) = 0.0837 A(short, tall) = 0 The affinities o f groups (1) and (2) are both greater t h a n 0.5. This m e a n s that their degrees of matching are greater t h a n their differences. There- fore, we conclude that the two fuzzy concepts being c o m p a r e d are the same or similar (to some extent). The affÉnities in group (4) are less t h a n 0.5, we say that the two terms are different. The affinities in group (3) equal to 0.5 which m e a n s that the t r u t h and falsity o f the likeness between two fuzzy terms c a n n o t be indicated by the affinity measure. As a conclu- Consistency Check for Fuzzy Expert Systems 275 sion, if the affinity between two fuzzy terms is greater than 0.5, they can be treated as equal. For case (3), we could swing to either side depending on the implementation. The definition of affinity could be extended to detect the likeness of the antecedent part of two rules. Let P1 and P2 be the antecedent parts of two rules that contain multiple propositions connected by A N D operators, i.e., P1 - - P l l A P12 A P13 A "'" A Plk P2 = P21 A P22 A P23 A .-- A P2k Then A ( P 1 , P2) = f i A ( P l k , P2k) k = l If Plk and P2k are non-fuzzy and Plk = P2k, then A ( P l k , p 2 k ) = 1; otherwise, A ( P l k , P2k) = 0. The affinity of an antecedence is equal to the product of the affinity of its propositions. 4. A C O N S I S T E N C Y C H E C K E R IN F U Z Z Y E N V I R O N M E N T Using affinity, we could detect inconsistency in a fuzzy environment. The detection methods can be the same as those used in a nonfuzzy environ- ment except that we use affinity to measure the degree of matching of two fuzzy expressions. i.e., (P1 = P2) and (P1 4: P2) could be replaced by A ( P I I P 2 ) >_>_ 0.5 and A ( P I I P 2 ) < 0.5 in a fuzzy environment respectively. A consistency checker in a fuzzy environment (CCFE) has been imple- mented to check the rule bases built from a fuzzy expert system shell Z-III that allows any mix of fuzzy and non-fuzzy terms in the rules [8]. C C F E is implemented using Turbo C 2.0 in a microcomputer environment. Affinity measure is employed to determine the degree of matching of two fuzzy propositions. At the first stage of the development of CCFE, we set the following assumptions: (a) Only one proposition is allowed in the consequence part of each rule. (b) Disconfirmation of a consequent proposition is represented by a negative certainty factor to a rule. i.e., For the rule: IF height is tall T H E N weight is not light certainty is 0.8 276 K.S. Leung and Y. T. So (c) (d) (i) can be represented as follows: IF height is tall T H E N weight is light certainty is - 0.8 Single-valued objects are assumed. CCFE only searches for the following most common inconsis- tencies: Redundancy Let: rule 5: IF P1 T H E N C1 (CF1) rule 6: IF P2 T H E N C2 (CF2) where P1 and P2 may be any combination of propositions, and CF stands for the certainty factor of a rule. Redundancy may occur when: A ( P I l P 2 ) > 0.5 & A ( C l J C 2 ) > 0.5 & sign(CF1) = sign(CF2) or A ( P I I P 2 ) > 0.5 & A ( C I l C 2 ) < 0.5 & sign(CF1) 4~ sign(CF2) ( i i ) Contradiction Contradiction may occur when: A ( P I l P 2 ) > 0.5 & A ( C I l C 2 ) > 0.5 & sign(CF1) v~ sign(CF2) or A ( P I l P 2 ) > 0.5 & A ( C I l C 2 ) < 0.5 & sign(CF1) = sign(CF2) ( i i i ) Subsumption A special case of redundancy. The details have been described in section 2. (iv) Sub-contradiction A special case of contradiction. The details have been described in section 2. A decision table is used for verification of the rule base. A decision table facilitates the testing of a set of rules for inconsistent conditions. Each column of a decision table represents the conditions of a rule. Building a decision table from the rules is relatively simple [5]. CCFE creates a decision table by parsing the rules stored in the rule base, entering each corresponding value in the appropriate row and column. The table is implemented using arrays of pointers. Thus, the storage for the table is dynamically allocated. Besides, a dynamically allocated one-dimensional array is created to store the certainty factor of each rule. Consistency Check for Fuzzy Expert Systems T a b l e 1. A Sample R u l e Set 277 (rule rl (if (height is tall then weight is heavy) certainty is 0.7 (rule r3 (if (eating is much) and (activity is little) then weight is moreorless heavy) certainty is 0.9 (rule r2 (if (height is tall) and (eating is very much) then weight is very heavy) certainty is 0.65 T h e algorithm for creating the decision table and the array storing the certainty factors is as follows: R e a d all the rules f r o m a file into the rule base. C o u n t the total n u m b e r of objects and rules in the rule base. Allocate spaces for the table a n d the certainty factors. Initialise all the cells in the table to U N O C C U P I E D . F o r all the rules in the rule base Parse the rule a n d e n t e r the values to the corresponding row and column. Set t h a t cell to O C C U P I E D . F o r instance, f r o m the rules in Table 1, C C F E will create the corre- sponding decision table a n d the array storing certainty factors as shown in Figure 4. height eating activity weight r l r2 r3 tall tall very much much little heavy very heavy moreorless heavy Certainty factor 07 I 065 I Figure 4. The sample decision table. 0.9 278 K.S. Leung and Y. T. So A f t e r constructing the decision table, all the rules are checked for inconsistencies. Affinity m e a s u r e is employed to d e t e r m i n e the likeliness b e t w e e n two expressions, either fuzzy or non-fuzzy. T h e following is the algorithm for consistency checking: F o r i = 1 to no. o f rules - 1 F o r j = i + 1 to no. o f rules case compare(antecedence(i), antecedence(j)) E Q U A L : if ( s a m e ( c o n s e q u e n c e ( i ) , consequence(j)) t h e n if (sign(CF(i)) = sign(CF(j))) t h e n report r e d u n d a n c y else report contradiction else if (sign(CF(i)) = sign(CF(j))) t h e n report contradiction else report r e d u n d a n c y endif continue SUBSET: if ( s a m e ( c o n s e q u e n c e ( i ) , consequence(j)) t h e n else if (sign(CF(i)) = sign(CF(j))) t h e n report subsumption else report sub-contradiction if (sign(CF(i)) = sign(CF(j))) t h e n report sub-contradiction else report subsumption e n d i f continue D I F F E R E N T : continue endcase e n d f o r e n d f o r where compare is a function using affinity m e a s u r e to d e t e r m i n e w h e t h e r two antecedences are the SAME, D I F F E R E N T , or one o f the ante- cedence is a S U B S E T o f the other. Consistency Check for Fuzzy Expert Systems 279 same is a Boolean function using affinity measure to determine whether two consequences are the same. sign returns the sign of the certainty factor. 5. RESULTS AND DISCUSSIONS CCFE is used to verify the consistency of fuzzy rule bases. The following rule base with fuzzy and non-fuzzy objects is verified by CCFE and the results are presented. (The example is an inconsistent testing case, so some rules may not make too much sense.) Let height is a fuzzy object with possible values (tall, medium, short). weight is a fuzzy object with possible values (heavy, fit, light). eating is a fuzzy object with possible values (much, balanced, little). sleeping is a non-fuzzy single-valued object with possible values (enough, fair, bad). sport is a non-fuzzy single-valued object with possible values (big-2, football, basketball). Rule Base: (rule r l If (height is tall) then weight is heavy) Certainty is 0.7 (rule r2 if ((height is tall) and (eating is much)) then weight is very heavy) Certainty is 0.75 (rule r3 if (height is tall) then weight is balanced) Certainty is - 0 . 5 (rule r4 If ((eating is much) and (sport is big-2)) then weight is very heavy) Certainty is 0.6 280 K.S. Leung and Y. T. So (rule r5 if (eating is very much) then weight is light) Certainty is 0.5 (rule r6 if (eating is very much) and (sport is big-2) and (sleeping is not bad) then weight is very heavy) Certainty is - 0 . 7 (rule r7 if (eating is much) and (sport is big-2) and sleeping is enough) then weight is heavy) Certainty is 0.6 (rule r8 If (sleeping is not enough) then weight is light) Certainty is 0.5 (rule r9 If (sleeping is bad) then weight is light) Certainty is 0.4 Diagnosis: Warning 1: Subsumption between rules r l and r2 Warning 2: Redundancy between rules r l and r3 Warning 3: Subsumption between rules r2 and r3 Warning 4: Sub-contradiction between rules r2 and r5 Warning 5: Sub-contradiction between rules r4 and r5 Warning 6: Sub-contradiction between rules r4 and r6 Warning 7: Subsumption between rules r4 and r7 Warning 8: Subsumption between rules r5 and r6 Warning 9: Sub-contradiction between rules r5 and r7 Warning 10: Contradiction between rules r6 and r7 Warning 11: Redundancy between rules r8 and r9 Total, 11 warnings. Consistency Check for Fuzzy Expert Systems 281 F r o m the results, we can see that affinity is useful to d e t e r m i n e the degree of matching o f two fuzzy expressions for consistency checking. T h e techniques used in a non-fuzzy e n v i r o n m e n t can be easily trans- p o r t e d to a fuzzy environment. However, in an e n v i r o n m e n t full of mixtures o f approximate and exact reasonings, there are no absolute definitions of consistency as discussed in sections 2 and 3. A fuzzy consistency checker such as the one i m p l e m e n t e d can only serve as a provider for warnings of possible inconsistency to aid b e t t e r knowledge engineering. T h e levels o f warnings can be adjusted by varying p a r a m e - ters such as the threshold affinity in the definition and levels o f tracking. 6. C O N C L U S I O N Using affinity, it has b e e n d e m o n s t r a t e d that we can detect possible inconsistency in a fuzzy environment. T h e detection techniques are very similar to those used in a non-fuzzy e n v i r o n m e n t except that we use. affinity to m e a s u r e the likeness o f two fuzzy expressions, i.e., for two non-fuzzy expressions A1 and A2, A1 = A 2 and (A1 v~ A2) could be replaced by A ( A 1 , A2) > 0.5 and A ( A 1 , A2) < 0.5 respectively for fuzzy expressions A1 and A 2 in a fuzzy environment. The commutative property of affinity makes the checking o r d e r of new and old rules totally indepen- dent. The affinity m e a s u r e has been proven to be a good tool to m e a s u r e the likeness o f two fuzzy terms. The consistency checker C C F E has been successfully i m p l e m e n t e d for aiding knowledge engineering in a heteroge- neous fuzzy e n v i r o n m e n t with mixed fuzzy and non-fuzzy terms. References 1. Hayes-Roth, F., Rule-based systems, Communication of ACM, 28(9), 921-932, 1985. 2. Shortliffe, E. H., Computer-Based Medical Consultations: MYCIN, Elsevier, New York, 1976. 3. Suwa, M., Scott, A. C., and Shortliffe, E. H., An approach to verifying completeness and consistency in a rule-based expert system, AI Magazine, 3(3), 16-21, 1982. 4. KES General Description Manual, Software Architecture and Engineering, Inc., Arlington, V.A., 33, 1983. 5. Cragun, B. J., and Steudel, H. J., A decision-table-based processor for checking completeness and consistency in rule-based expert systems, Int. J. Man-Machine Studies, 26, 633-648, 1987. 282 K.S. Leung and Y. T. So 6. Leung, K. S., and Lam, W., Fuzzy concepts in expert systems, IEEE Computer, 2l(9), 43-56, Sept. 1988. 7. Leung, K. S., and Lam, W., A fuzzy expert system shell using both exact and inexact reasoning, J. Automated Reasoning, 5(2), 207-233, 1989. 8. Leung, K. S., So, Y. T., Leung, A., and Wong, W. S. F., Z-Ill: A PC-based fuzzy expert system shell, Proc. Int. Computer Symp., 657-662, 1990. 9. Beauvieux, A., and Dague, P., Interactive checking of knowledge base consis- tency, Int. Computer Sci. Conf. 567-574, 1988. 10. Buchannan, B. G., and Shortliffe, E. H., Knowledge Engineering, in Rule-Based Expert Systems, Addisley-Wesley Publishing Company, 1984. 11. Nguyen, T. A., Perkins, W. A., Laffey, T. J., and Pecora, D., Knowledge base verification, A I Magazine, 69-75, summer 1987. 12. Tsang, W. W., Wan, Y. S., Lim, E. L., and Hioe, W., A space searching method for checking the consistency and completeness of a rulebase, Int. Computer Sci. Conf. 575-579, 1988. 13. Zadeh, L. A., A fuzzy-set-theoretic interpretation of linguistic hedges, J. Cybernetics, 2(3), 4-34, 1972. 14. Zadeh, L. A. P R U F - - A meaning representation language for natural lan- guages, Int. J. Man-Machine Studies, 10, 395-460, 1978. 15. Cayrol, M., Farreny, H., and Prade, H., Fuzzy Pattern Matching, Kybernetes, 11, 103-116, 1982. 
work_rmcq2fj2gjgjpaemfohxb4igwe	----	ATISA: Adaptive Threshold-based Instance Selection Algorithm Expert Systems with Applications 40 (2013) 6894–6900 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a ATISA: Adaptive Threshold-based Instance Selection Algorithm 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.06.053 ⇑ Corresponding author. Address: Universidade Federal de Pernambuco (UFPE), Centro de Informática (CIn), Av. Jornalista Anibal Fernandes, s/n. Cidade Univer- sitária 50740-560, Recife, PE, Brazil. Tel.: +55 81 2126 8430x4346; fax: +55 81 2126 8438. E-mail addresses: gdcc@cin.ufpe.br (G.D.C. Cavalcanti), tir@cin.ufpe.br (T.I. Ren), clp@cin.ufpe.br (C.L. Pereira). URL: http://www.cin.ufpe.br/~viisar (G.D.C. Cavalcanti). George D.C. Cavalcanti a,⇑, Tsang Ing Ren a,b, Cesar Lima Pereira a a Centro de Informática, Universidade Federal de Pernambuco, Brazil b iMinds – Vision Lab, Department of Physics, University of Antwerp, Universiteitsplein 1, B-2610 Wilrijk, Belgium a r t i c l e i n f o Keywords: Instance selection Instance-based learning algorithms a b s t r a c t Instance reduction techniques can improve generalization, reduce storage requirements and execution time of instance-based learning algorithms. This paper presents an instance reduction algorithm called Adaptive Threshold-based Instance Selection Algorithm (ATISA). ATISA aims to preserve important instances based on a selection criterion that uses the distance of each instance to its nearest enemy as a threshold. This threshold defines the coverage area of each instance that is given by a hyper-sphere cen- tered at it. The experimental results show the effectiveness, in terms of accuracy, reduction rate, and computational time, of the ATISA algorithm when compared with state-of-the-art reduction algorithms. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction In supervised learning, a training set of instances is given as in- put to a machine learning algorithm. Each instance is formed by a descriptor vector and a label that represent its properties and class respectively. After the learning process, unknown instances are presented to the algorithm that must infer their classes through generalizations (Mitchell, 1990). Instance-based learning algorithms use the whole set of in- stances from the training set to construct inference structures. The Nearest Neighbor Rule (Cover & Hart, 1967) is a well-known example of this kind of algorithm. It requires a large memory space since the whole training instance set has to be stored. Each query pattern is compared with every pattern in the training set. An un- known instance is assigned to the class of the most similar instance in the training set. Instead of using only the nearest instance, this rule can be generalized (in the sense of proximity) by using the classes of the k most similar neighbors; this algorithm is called k-Nearest Neighbors (kNN). There are other algorithms that use this neighborhood concept, for example: the Center-based Nearest Neighbor classifier (Gao & Wang, 2006). In spite of the efficiency of instance-based learning algorithms, they suffer from problems such as: large storage requirements, low prediction performance and sensitivity to noisy and redundant data (David, Aha, & Albert, 1991). Instance reduction techniques are used as a pre-processing step on the data set. One of its objectives is to overcome the limitations faced by instance-based learning algorithms. This is obtained by reducing the size of the training set by removing ‘‘irrelevant’’ in- stances. Reduction techniques can be classified in three types: con- densation, edition and hybrid (Garcia, Derrac, Cano, & Herrera, 2012; García, Marqués, & Sánchez, 2012; Nanni & Lumini, 2011). Edition methods aim to remove noisy instances and Condensation methods search for a consistent subset of the training set. A consis- tent subset is formed by eliminating instances in the training set that do not affect the classification accuracy of the whole training set. Hybrid methods compute a subset of the training set combin- ing the characteristics of Edition and Condensation methods; so, noisy and superfluous instances are eliminated. The Condensed Nearest Neighbor Rule (CNN) (Hart, 1968) is one of the first method that attempts to reduce the number of in- stances in a training set. It removes any well classified instance using the nearest neighbor rule. The CNN algorithm maintains bor- der instances and eliminates redundant ones. Another example of a reduction technique is the Edited Nearest Neighbor Rule (ENN) (Wilson, 1972) that keeps central instances by smoothing the deci- sion boundaries. CNN is a condensation algorithm while ENN is a noise filter. There are many algorithms that can not be classified neither as condensation nor noise filter; they are hybrid selection approaches. An example is the Iterative Case Filtering algorithm (ICF) (Brighton & Mellish, 1999) that iteratively removes instances according to a filtering rule. Wilson and Martinez (2000) proposed a series of algorithms called Decremental Reduction Optimization Procedure. Marchiori (2008) proposed a network-based represen- tation of a training set, called Hit Miss Network (HMN) that pro- vides a compact description of the nearest neighbor relation over pairs of instances from each pair of classes. http://crossmark.dyndns.org/dialog/?doi=10.1016/j.eswa.2013.06.053&domain=pdf http://dx.doi.org/10.1016/j.eswa.2013.06.053 mailto:gdcc@cin.ufpe.br mailto:tir@cin.ufpe.br mailto:clp@cin.ufpe.br http://www.cin.ufpe.br/~viisar http://dx.doi.org/10.1016/j.eswa.2013.06.053 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa G.D.C. Cavalcanti et al. / Expert Systems with Applications 40 (2013) 6894–6900 6895 This paper introduces an instance reduction algorithm called Adaptive Threshold-based Instance Selection Algorithm (ATISA, for short). It aims to find the best set of instances in order to reduce the size of the stored set of instances while maintaining or even improving the generalization accuracy. ATISA does not eliminate an instance based on its spacial location - central or border points. It selects important points based on a threshold that defines a cov- erage area per instance. This work is organized as follows. The Adaptive Threshold- based Instance Selection Algorithm or ATISA is described in Sec- tion 2. The experimental results are presented in Section 3. Finally, Section 4 shows the conclusions. 2. Adaptive Threshold-based Instance Selection Algorithm (ATISA) The Adaptive Threshold-based Instance Selection Algorithm (ATISA) is an incremental instance selection technique. Based on a threshold, the algorithm decides whether or not instances are in- cluded in the final subset. ATISA has three different approaches: ATISA1, ATISA2 and ATISA3. These three versions are described in the next subsections. Algorithm 1 : Adaptive Threshold-based Instance Selection Algorithm 1 (ATISA1) Require: T {The training set} 1: Tf = T 2: for xi 2 T do 3: if class(xi) – classmax(NNk(xi,T)) then 4: Tf = Tfn{xi} 5: end if 6: end for 7: for xi 2 Tf do 8: h(xi) = d(xi,NE(xi,Tf)) 9: end for 10: S = ; 11: for xi 2 Tf do 12: N = NNk(xi,S) 13: if class(xi) – classmax(N) or d0(xi,N) > h0(N) then 14: S = S [ {xi} 15: end if 16: end for 17: return S {The subset of the selected instances from the training set} Fig. 1. Definition of the threshold h for the instance x. NE (x) represents the nearest enemy of the instance x. 2.1. Adaptive Threshold-based Instance Selection Algorithm 1 (ATISA1) In general terms, the algorithm applies a noise filter to the set and afterwards it calculates one threshold for each instance. Each remaining instance is classified based on its nearest neighbor. In the case of an erroneous classification, the instance is inserted into the final set. Otherwise, it is added only if the distance to its near- est neighbor is greater than the threshold of that neighbor. Algo- rithm 1 shows the pseudocode of ATISA1. NNk(a,A) represents the set of the k nearest neighbors of the in- stance a in the training set A, while class(a) represents the class of the instance a. classmax(B) receives the set of instances B and re- turns its predominant class. NE(a,A) represents the nearest enemy of a in the set A. That nearest enemy is the nearest neighbor of a that has a different class. The distance between instances a and b is given by d(a,b). The threshold of each instance is represented by the function h(a), where a corresponds to an instance. To use this algorithm with k > 1, the distance and threshold functions have the following variation: d0(a,A) and h0(A). The former returns the average distance between instance a and all instances in the set A, and the latter returns the average threshold calculated over the set of instances A. The selection solution of ATISA1 has three steps: pre-processing (filter), thresholds assignment and reduced set generation. The pre-processing is a filtering process that aims to remove noisy data from the training set (lines 1 to 6 in Algorithm 1). The Edited Nearest Neighbor Rule (ENN) (Wilson, 1972) is the algorithm used for this purpose. ENN removes any instance that is not cor- rectly classified by its nearest neighbor. In other words, given an instance xi and its nearest neighbor NNk(xi), k = 1, this instance is removed if its class class(xi) is different from the class of its nearest enemy class(NNk(xi)). The algorithm ENN enlarges the spaces be- tween classes, smoothing the decision boundaries. Threshold assignment: The threshold of an instance xi is de- fined by the distance to its nearest enemy NE(xi) (lines 7 to 9 in Algorithm 1). Fig. 1 illustrates the threshold h of the instance x. One threshold is assigned for each instance, it defines an area that is covered by the instance. All the instances laying inside this delimited area have the same class of the instance x. The reduced set generation (lines 10 to 16 in Algorithm 1) works as follows; after pre-processing, all remaining instances are randomly selected. One selected instance xi is inserted in the set S if it is correctly classified by its nearest neighbors in S. In a particular case, a misclassified instance can be inserted in S. How- ever, this only occurs if its distance to the nearest neighbor is greater than the threshold of this neighbor (for k > 1, average dis- tances and thresholds are used, regardless their classes). Each threshold can be interpreted as a radius that forms a hy- per-sphere centered at the instance position. If one instance is cor- rectly classified, even standing outside this hyper-sphere, it must be added to the final set, otherwise, it can be misclassified in the upcoming steps of the process. Fig. 2 shows an example of the ATISA1 third step: the procedure to obtain the reduced set. In Fig. 2(a), random indexes were assigned to each instance aiming to create an order. In Fig. 2(b), the first two instances were inserted, because they were not correctly classified and, at that point of the algorithm, the set S is empty. After, the instance with index 3 is not inserted in S. This happens because it is correctly classified, besides, it is inside the circle created by the threshold of its nearest neighbor (Fig. 2(c)). In this toy sample, the threshold of instance 1 is the distance between instances of indexes 1 and 9. In the next verification, instance 5 is checked. Although correctly classified, it is out of Fig. 2. An example of how ATISA1 obtain the reduced set. 6896 G.D.C. Cavalcanti et al. / Expert Systems with Applications 40 (2013) 6894–6900 G.D.C. Cavalcanti et al. / Expert Systems with Applications 40 (2013) 6894–6900 6897 the area covered by its nearest neighbor (Fig. 2(d)); so, it is added to the set S. At the end, the final set S is shown in Fig. 2(e). Algorithm 2: Adaptive Threshold-based Instance Selection Algorithm 2 (ATISA2) Require: T {The training set} 1: Tf = T 2: for xi 2 T do 3: if class(xi) – classmax(NNk(xi,T)) then 4: Tf = Tfn{xi} 5: end if 6: end for 7: for xi 2 Tf do 8: h(xi) = d(xi,NE(xi,Tf)) 9: end for 10: Tf = sort(Tf) 11: S = ; 12: for xi 2 Tf do 13: N = NNk(xi,S) 14: if class(xi) – classmax(N) or d0(xi,N) > h0(N) then 15: S = S [ {xi} 16: end if 17: end for 18: return S {The subset of the selected instances from the training set} 2.2. Adaptive Threshold-based Instance Selection Algorithm 2 (ATISA2) In the second version of the ATISA algorithm, a step before the process of creating the final reduced set is added. Instead of randomly selecting the instances, as done by ATISA1, an ordering procedure is required by ATISA2. So, the distance between each in- stance and its nearest enemy (threshold value) is calculated. The first selected instance is the one having the largest distance; the second largest distance corresponds to the second selected in- stance; and so on. Initially, instances with the largest thresholds are inserted in S. This means that these instances cover a representative area of the feature space. Based on that, it is probable that the last inserted instances (the ones with small thresholds) will be placed inside their neighbors coverage areas. So, they will not be inserted in S. In comparison with ATISA1, this procedure reduces the size of the final instance set. Fig. 3. (a) Show the instance labeled; (b) The reduced Algorithm 2 shows the pseudocode of ATISA2. The difference between ATISA1 and ATISA2 is in line 10. There is a sort function in this line and it is responsible for ordering the instances from the largest to the smallest threshold value. Fig. 3 presents the behavior of ATISA2 for the same toy problem shown in Section 2.1. In Fig. 3(a), the instances are labeled follow- ing the ordering obtained by ATISA2. The first visited instances are the ones more distant from their nearest enemy. In this example, the first two inserted instances are the ones with labels 1 and 2. These instances can be viewed in Fig. 3(b), each with its respective coverage area. They are the instances with greater threshold in the instance set, and, in this problem, all other instances are placed in- side the coverage area of these two instances. As they are correctly classified by instances 1 and 2, they are not included in S. There is only one exception: the instance with index 12 that has instance 1 as its nearest neighbor. Instance 12 is added to S because the dis- tance between this instance and instance 1 is greater than the threshold of the instance 1. The final set of instances generated by ATISA2 is presented in Fig. 3(b). Algorithm 3: Adaptive Threshold-based Instance Selection Algorithm 3 (ATISA3) Require: T {The training set} 1: Tf = T 2: for xi 2 T do 3: if class(xi) – classmax(NNk(xi,T)) then 4: Tf = Tfn{xi} 5: end if 6: end for 7: for xi 2 Tf do 8: h(xi) = d(xi,NE(xi,Tf)) 9: end for 10: Tf = sort(Tf) 11: S = one_by_class(Tf) 12: update_thresholds(S) 13: for xi 2 Tf do 14: N = NNk(xi,S) 15: if class(xi) – classmax(N) or d0(xi,N) > h0(N) then 16: S = S [ {xi} 17: update_thresholds(S) 18: end if 19: end for 20: return S {The subset of the selected instances from the training set} instance set obtained after the ATISA2 procedure. Fig. 4. An example of the ATISA3 procedure to updated the threshold dynamically. Table 1 Accuracy rates of experiments on the datasets using the 3-Nearest Neighbor classifier. R represents the reduction percentage. Databases 3-NN ATISA1 R ATISA2 R ATISA3 R annealing 92.48 92.61 82.01 91.85 87.18 88.47 91.20 australian 85.94 85.80 79.82 85.36 85.27 85.80 91.01 breast cancer 95.42 95.42 90.53 94.71 94.91 93.71 96.92 crx 85.51 86.09 77.91 82.90 83.20 82.32 89.97 hepatitis 83.87 82.58 81.21 80.00 85.59 81.94 90.46 image segmentation 96.23 94.20 86.25 93.07 88.82 91.52 91.90 ionosphere 85.19 86.61 79.61 84.62 82.46 81.48 94.49 iris 94.67 98.00 81.56 94.00 84.00 92.00 88.74 liver 65.22 62.32 65.67 60.58 73.94 62.03 83.09 pima diabetes 74.35 72.53 79.25 72.01 84.46 71.48 90.39 promoters 84.91 83.02 61.64 83.02 72.96 72.64 78.09 sonar 83.65 79.81 56.46 79.81 65.81 73.08 76.66 soybean 89.58 81.76 66.77 78.18 77.24 79.15 77.20 tic tac toe 88.83 89.14 78.84 91.44 83.59 89.87 83.30 voting 92.64 93.79 87.28 93.79 92.03 91.95 95.76 wine 96.63 94.94 77.59 96.07 78.40 90.45 87.58 85.09 82.49 82.99 87.92 84.99 83.79 84.06 90.18 Wilcoxon � n/a n/a + � + � 6898 G.D.C. Cavalcanti et al. / Expert Systems with Applications 40 (2013) 6894–6900 2.3. Adaptive Threshold-based Instance Selection Algorithm 3 (ATISA3) Average 87.19 86.16 77.03 Median 87.39 86.35 79.43 ATISA3 is an evolution of ATISA2 that calculates the threshold dynamically. In other words, when the set S is modified, a proce- dure is started to update all the instances thresholds. ATISA3 reaches even higher reduction rates than ATISA2. Initially, thresh- olds are more likely to have higher values. This happens due to the ordering inherited from ATISA2 combined with the fact that thresholds are dynamically calculated. This dynamic threshold up- date increases the chance of an instance to be within the coverage area of another instance. The pseudocode of ATISA3 is presented in Algorithm 3. Its main difference in comparison to ATISA2 is in lines 11 to 19. In line 11, the function S = one_by_class(A) receives a set of instance A as input and returns a set having one instance of each class. Each class is represented by the instance with the largest threshold. After, the thresholds of the instances in S are updated and this is performed by the function update_thresholds(S) – line 12. ATISA3 starts with one instance per class, as shown in Fig. 4(a). In this case, the instances labeled 1 and 2 have the same threshold value (Fig. 4(b)) that corresponds to the distance between them. This toy problem has only two different classes. It is possible to observe that all other instances are correctly classified and they are inside the boundaries of their respective neighbor threshold. So, the reduced set constructed by ATISA3 has only two instances: 1 and 2. 3. Experimental results Experiments were carried out using 16 databases from the UCI Machine Learning Repository (Frank & Asuncion, 2010): annealing, australian, breast cancer, crx, hepatitis, image segmentation, iono- sphere, iris, liver, pima diabetes, promoters, sonar, soybean, tic tac toe, voting and wine. All numerical values were normalized to a number between 0 and 1. These databases have different number of classes, attributes (categorical or numerical) and instances; also, some of them have instances with missing values. ATISA was compared with the following techniques: Decremen- tal Reduction Optimization Procedure 3 (DROP3) (Wilson & Martinez, 2000), Hit Miss Network for Editing Iterated (HMN-EI) (Marchiori, 2008) and Iteractive Case Filtering (ICF) (Brighton & Mellish, 1999). Garcia et al. (2012) showed that DROP3 and ICF are between the methods most used for comparison purposes and that HMN-EI has found very competitive accuracy rates. It is also important to Table 2 Results of experiments on the data sets. R is the percentage of training points removed. The symbols +, � and � show the number of times ATISA1 average accuracy (storage reduction) is significantly better (+), significantly worse (�) or similar (�) than the other algorithm, according to a paired t-test at 0.05 significance level. In the ‘‘Wilcoxon’’ row, + indicates ATISA1 was significantly better than the other algorithm according to a Wilcoxon test for paired samples, � indicates it was significantly worse and � indicates no significant difference. Databases DROP3 R HMN-EI R ICF R ATISA1 R annealing 92.73� 88.36� 89.47+ 38.64+ 86.34+ 84.18� 92.61 82.01 australian 84.49� 85.04� 83.33+ 48.41+ 84.64� 82.00� 85.80 79.82 breast cancer 94.56� 92.99� 96.14� 52.03+ 94.42� 90.00� 95.42 90.53 crx 85.07� 86.33� 84.49� 50.68+ 82.90+ 82.69� 86.09 77.91 hepatitis 83.23� 89.18� 84.52� 54.48+ 83.87� 87.24� 82.58 81.21 image segmentation 94.94� 90.35� 92.03+ 37.88+ 90.65+ 79.86+ 94.20 86.25 ionosphere 85.76� 95.28� 86.61� 59.54+ 72.36+ 94.90� 86.61 79.61 iris 94.00+ 87.70� 96.00� 44.44+ 96.00� 71.19+ 98.00 81.56 liver 63.77� 74.56� 57.68� 57.01+ 56.52� 77.55� 62.32 65.67 pima diabetes 72.27� 83.44� 74.22� 50.30+ 70.44� 85.63� 72.53 79.25 promoters 84.91� 71.58� 84.91� 59.85� 74.53� 64.98� 83.02 61.64 sonar 77.88� 70.67� 75.48� 57.96� 73.56+ 67.89� 79.81 56.46 soybean 82.41� 65.58+ 80.78� 37.79+ 79.48� 55.77+ 81.76 66.77 tic tac toe 83.61+ 83.59� 81.00+ 82.73� 86.53+ 31.53+ 89.14 78.84 voting 92.87� 94.35� 90.11+ 55.40+ 90.80+ 87.77� 93.79 87.28 wine 96.07� 84.83� 96.63� 66.72+ 94.38� 87.64� 94.94 77.59 Average 85.54 83.99 84.59 53.37 82.34 76.93 86.16 77.03 Median 84.99 85.68 84.71 53.25 84.25 82.34 86.35 79.43 Wilcoxon � � + + + � n/a n/a 82 83 84 85 86 87 88 Accuracy (%) 50 60 70 80 90 100 R e d u ct io n ( % ) ICF ATISA3 HMN-EI ATISA2 DROP3 ATISA1 Fig. 5. Accuracy x Reduction chart. G.D.C. Cavalcanti et al. / Expert Systems with Applications 40 (2013) 6894–6900 6899 emphasize that there are in the literature previous methods that construct a rank in order to find the best instances, for example: the InstanceRank method (Vallejo, Troyano, & Ortega, 2010). How- ever, this method did not perform as well as the DROP3, ICF and HMN-EI methods. So, it was not used for comparison purposes. The Heterogeneous Euclidean Overlap Metric (HEOM) (Wilson & Martinez, 1997) was used in every required situation, i.e., in the presence of hybrid attributes (numerical and categorical). HEOM uses the Euclidean distance in numeric attributes and the Ham- ming distance in categorical ones. Every distance between attri- butes is calculated and the value returned by HEOM corresponds to the square root of the sum of these distances, as it is described in Eq. (1). HEOMðx; yÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXm a¼1 daðxa; yaÞ s ð1Þ where x and y are instances, m is the number of attributes and xa and ya represent the attributes of the instances x and y respectively. daðx; yÞ¼ ðx � yÞ2; numerical 0; categorical ðx ¼ yÞ Max; categorical ðx – yÞ or missing 8>< >: ð2Þ When the attribute values are numeric, the squared difference is calculated as defined in Eq. (2). If they are categorical and have the same value, the attribute distance is minimal (zero). But, if they are categorical and do not have the same value, the attribute dis- tance is maximal (Max = 1, because the whole dataset was normal- ized to a number between 0 and 1). All experiments reported use the k-Nearest Neighbor classifier (k = 3) and the 10-fold cross-validation method. After preliminary experiments, the number of nearest neighbor of the function NNk(�) was also set to 3. Tables 1 and 2 show the experimental results. For 6900 G.D.C. Cavalcanti et al. / Expert Systems with Applications 40 (2013) 6894–6900 each technique, classification accuracy and reduction rate are presented (R denotes reduction). In order to assess whether differences in accuracy and storage reduction are significant, a non-parametric Wilcoxon signed ranks test for zero median at a 0.05 significance level was used. The symbols +, � and �, in the row labeled ‘‘Wilcoxon’’, indicate that ATISA1 is significantly bet- ter, worse or equivalent compared with the other algorithms, respectively. Table 1 presents a comparison between the three versions of ATISA. In this table, the 3-NN classifier is used as a reference result. It shows that ATISA3 obtains better reductions rates than the other ones. Observing the Wilcoxon test, ATISA1 is significantly better than ATISA2 and ATISA3 and similar to 3-NN in terms of accuracy rate. On the other hand, ATISA1 is significantly worse than ATISA2 and ATISA3, in terms of reduction. Table 2 presents a comparison between ATISA1 (the best of the three proposed methods) and three reduction techniques: DROP3, HMN-EI and ICF. ATISA1 is significantly better than ICF and HMN- EI in classification performance. Observing the reduction, ATISA1 is better the HMN-EI and similar to ICF. In terms of accuracy rates, ATISA1 and DROP3 show similar results. Fig. 5 shows a two dimensional plot: accuracy versus reduction. Each point in this figure represents the average value showed in Tables 1 and 2. The top right corner is the best possible solution. ATISA1 is ahead in terms of accuracy and ATISA3 reaches the best reduction. The average running time in seconds shows that the ATISA algo- rithms are faster than the other ones. The reduction process of the DROP3 algorithm took, in average, 4.29 s, while ATISA1 was almost three times faster, 1.56 s. ICF and HMN-EI took 3.16 and 2.44 respectively. The values showed above represent the average time of the reduction process per algorithm using all the datasets. The algorithms were implemented in C++ using the GCC 4.1.2 compiler. All the experiments were executed in a Linux (Kernel 2.6.18) con- trolled environment using an Intel Xeon 3 GHz with 4 GB of RAM. 4. Conclusions This paper proposed three algorithms to perform instance selec- tion: Adaptive Threshold-based Instance Selection Algorithm 1,2,3 (ATISA, for short). ATISA calculates a hypersphere centered at each instance and this hypersphere defines the coverage area of the in- stance. The radius of the hypersphere is calculated based on the distance between the instance and its nearest enemy. Instances that belong to other instances area of coverage are removed. The proposed algorithms maintain important points to define each class, it does not matter if these points belong to the border or are inner points of the class. The three versions of the algorithm were evaluated using UCI databases. In general terms, ATISA1 obtains the best classification performance, ATISA3 obtains the best rates in the reduction task and ATISA2 is a balanced choice between accuracy and reduction rates. ATISA algorithms were compared with the state-of-the-art instance selection methods. This empirical analysis shows the effectiveness of the proposed techniques in terms of accuracy and reduction rates. Moreover, when the computational cost was evaluated, ATISA was the faster algorithm when compared with DROP3, ICF and HMN-EI. Acknowledgment This research was partially supported by CNPq, Capes and Facepe. References Brighton, H., & Mellish, C. (1999). On the consistency of information filters for lazy learning algorithms. In European conference on principles of data mining and knowledge discovery (pp. 283–288). Cover, T. M., & Hart, P. E. (1967). Nearest neighbor pattern classification. IEEE Transactions on Information Theory, 13, 21–27. David, W., Aha, D. K., & Albert, M. K. (1991). Instance-based learning algorithms. Machine Learning, 6, 37–66. Frank, A., & Asuncion, A. (2010). UCI machine learning repository. Gao, Q.-B., & Wang, Z.-Z. (2006). Center-based nearest neighbor classifier. Pattern Recognition, 40, 346–349. Garcia, S., Derrac, J., Cano, J., & Herrera, F. (2012). Prototype selection for nearest neighbor classification: Taxonomy and empirical study. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34, 417–435. García, V., Marqués, A. I., & Sánchez, J. S. (2012). On the use of data filtering techniques for credit risk prediction with instance-based models. Expert Systems with Applications, 39, 13267–13276. Hart, P. E. (1968). The condensed nearest neighbor rule. IEEE Transactions on Information Theory, 14, 515–516. Marchiori, E. (2008). Hit miss networks with applications to instance selection. Journal of Machine Learning Research, 9, 997–1017. Mitchell, T. M. (1990). The need for biases in learning generalizations. In J. Shavlik & T. Dietterich (Eds.), Readings in machine learning (pp. 184–191). Morgan Kaufman. Nanni, L., & Lumini, A. (2011). Prototype reduction techniques: A comparison among different approaches. Expert Systems with Applications, 38, 11820–11828. Vallejo, C. G., Troyano, J. A., & Ortega, F. J. (2010). Instancerank: Bringing order to datasets. Pattern Recognition Letters, 31, 133–142. Wilson, D. L. (1972). Asymptotic properties of nearest neighbor rules using edited data. IEEE Transactions on Systems, Man and Cybernetics, 2, 408–421. Wilson, D. R., & Martinez, T. R. (1997). Improved heterogeneous distance functions. Journal of Artificial Intelligence Research, 6, 1–34. Wilson, D. R., & Martinez, T. R. (2000). Reduction techniques for instance-based learning algorithms. Machine Learning, 38, 257–286. http://refhub.elsevier.com/S0957-4174(13)00448-X/h0005 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0005 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0010 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0010 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0015 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0015 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0020 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0020 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0020 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0025 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0025 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0025 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0030 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0030 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0035 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0035 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0040 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0040 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0040 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0045 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0045 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0050 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0050 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0055 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0055 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0060 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0060 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0065 http://refhub.elsevier.com/S0957-4174(13)00448-X/h0065 ATISA: Adaptive Threshold-based Instance Selection Algorithm 1 Introduction 2 Adaptive Threshold-based Instance Selection Algorithm (ATISA) 2.1 Adaptive Threshold-based Instance Selection Algorithm 1 (ATISA1) 2.2 Adaptive Threshold-based Instance Selection Algorithm 2 (ATISA2) 2.3 Adaptive Threshold-based Instance Selection Algorithm 3 (ATISA3) 3 Experimental results 4 Conclusions Acknowledgment References 
work_rmoyqarpofdehavtg6affi7un4	----	Article Distributed conflict resolution among cooperating expert systems Faruk Polat Shashi Shekhar University of Minnesota, Computer Science Department, Minneapolis, MN55455, USA H. Altay Guvenir Bilkent University, Department of Computer Engineering and Information Science, Bilkent, 06533 Ankara, Turkey Abstract: Cooperating experts approach attempts to integrate and coordinate the activities of multiple specialised problem solvers that come together to solve complex tasks such as design, medical diagnosis, business management and so on. Due to the different goals, knowledge and viewpoints of agents, conflicts may arise at any phase of the problem-solving process. Managing diverse expertise requires well-organised models of conflict resolution. In this paper, a model for cooperating experts is described which openly supports multi-agent conflict detection and resolution. The model is based on the idea that each agent has its own conflict knowledge which is separated from its domain level knowledge, and each agent has its own conflict resolution knowledge which is not accessible and known by others. Furthermore, there are no globally known conflict resolution strategies. Each agent involved in a conflict chooses a resolution scheme according to its self interest. The model is described by using an example in the domain of office design and it is compared with other systems. 1. Introduction Distributed Artificial Intelligence (DAI) is a subfield of AI which attempts to integrate existing problem-solving methods used in classical AI in order to develop systems that benefit from multiple agents' points of view. A solution developed by multiple agents incorporates aspects of each agent's problem- solving capabilities and perspectives rather than just the view of an individual agent's analysis of the problem (Cammarata et al. 1983; Chaib-Draa et al. 1992; Durfee et al. 1989; Gasser 1991; Malone & Crowston 1991; Polat & Guvenir 1991a; Smith & Davis 1988). One of the application domains of DAI is cooperating expert systems. The cooperating expert system approach is concerned with solving complex tasks which re- quire diverse expertise to generate comprehensive solutions. Applications of cooperating expert systems can be seen in human problem-solving tasks such as design, medical diag- nosis, research, business management and human relations. Several systems reflecting the cooperating expert systems ap- proach such as Hearsay-II (Erman et al. 1980), Contract Net (Smith & Davis 1988), Distributed Vehicle Monitoring Test- bed (Lesser & Corkill 1988), MDX (Chandrasekeran 1982, Gomez & Chandrasekeran 1984) and Coop (Shekhar & Ramamoorthy 1992) are described in the literature. Manag- ing diverse expertise is difficult because one has to take into account the problems which will arise in working out solu- tions in the face of conflicting goals, constraints, viewpoints and knowledge of heterogeneous experts. In this paper, we describe a mod.el in which a set of knowl- edge-based agents cooperates to solve design problems. The model is based on the resolution of conflicting solutions gen- erated by experts having different goals, priorities and evalu- ation criteria. Existing approaches to conflict management (Adler et al. 1989; Klein & Lu 1989; Lander & Lesser 1990; Werkman et al. 1990) rely on coordinated resolution strate- gies which require resolution of a conflict based on a globally agreed strategy. In existing systems, conflict resolution knowledge is either maintained centrally or replicated by all agents. In any case, one of the disputants is given the power to take control of the conflict and use a resolution scheme known to everybody. In the approach described in this paper, how- ever, agents are free to choose the most appropriate action, given their understanding of the global and local situations and their own capabilities. They maintain their own set of conflict resolution knowledge which is not globally known. Using their own conflict knowledge, the participants may reach an agreement on a revised solution. There are several reasons why there is a need for resolving conflicts by means of agents' private conflict resolution knowledge instead of global knowledge about conflict resolu- tion. First of all, this is much more similar to the resolution of conflicts that occur among human beings in solving complex Expert Systems, November 1993, Vol. 10, No 4 _________________________ 227 problem tasks in domains like design, diagnosis and business management. When a conflict is detected, it is not resolved by a central authority with global conflict resolution knowledge; rather, specialists involved in the conflict negotiate a revised solution that will be acceptable to all of them, using their own conflict resolution knowledge and perspectives. Second, forming global conflict resolution knowledge requires merg- ing the conflict resolution knowledge of each agent obtained through knowledge acquisition in a consistent manner. This makes the maintenance of the global knowledge difficult be- cause, when a new agent is added to or removed from the system, or the conflict resolution knowledge of an agent is revised, the global conflict resolution knowledge must be re- built accordingly. The last reason is related to the advantages of having distributed knowledge at all agents for reliability and fault-tolerance, instead of maintaining this knowledge in a centrally organised manner. In Section 2, an overview of conflict management in coop- erating expert systems is presented and the existing ap- proaches are summarised. Section 3 describes a new model for cooperating experts and explains how the problem solving proceeds within this model. Section 4 describes how conflict resolution takes place in the new model. Section 5 includes a design example to illustrate how problem solving proceeds in this approach. Finally, Section 6 summarises the new model, emphasising its characteristics. 2. Conflict detection and resolution among cooperating experts In the cooperating experts approach, several specialised agents combine to solve a common problem. During any phase of the problem-solving process, conflicts might appear as a result of incorrect and incomplete local knowledge, dif- ferent goals, priorities and solution evaluation criteria. When there are several conflicting proposed solutions for a (sub)problem, the agents involved in a conflict must either agree to choose one proposal, cooperatively revise one, or search for a new solution that will be acceptable to everyone. A common practice in building knowledge-based systems is to avoid potential conflicting situations by analysis and checking the consistency of the knowledge base at develop- ment time (Gingberg 1988; Nguyen et al. 1987; Polat & Gu- venir 1991 b; Rousset 1988; Vignollet & Ayel 1990). This ap- proach, although effective, is very costly as the amount and diversity of knowledge increases. Resolving all conflicts, no matter how unlikely, at development time can be prohibi- tively time-consuming. Moreover, dividing the domain knowledge into smaller internally consistent collections is difficult. The problems encountered when resolving conflicts in de- velopment time can be avoided by allowing conflicts to occur and be resolved at run-time. In other words, participating agents are allowed to generate conflicting solutions to the subproblems at run-time. In case of a conflict, a set of strate- gies could be used to resolve the conflict. Some examples of strategies include backtracking, compromise negotiation (a solution is iteratively revised by sliding a value or set of val- ues along some dimension until a mutually acceptable middle point is found), integrative negotiation (identify the most im- portant goals of each agent and find a solution which fulfils all of them), constraint relaxation, case-based and utility rea- soning methods, etc. (Adler et al. 1989; Klein & Lu 1989; Lander & Lesser 1990; Polat & Guvenir 1992; Sycara 1989; Werkman et al. 1990). Work in this class comes closest to providing conflict resolution expertise with first class status. Next, we will summarise studies which emphasise the use of conflict resolution within a cooperating expert systems para- digm. The first is conducted by Klein & Lu ( 1989), who pro- pose a model for cooperative design that emphasises the par- allel interaction of design agents. This work addresses the problem of how conflicts among different experts can be re- solved. In their model, there are design experts and a particu- lar conflict resolution expert. Given a design problem, design experts solve the subproblems relevant to their expertise. When a conflict is detected, the conflict resolution expert takes control and tries to resolve it. This expert maintains the global conflict management knowledge which contains con- flict classes and corresponding resolution strategies. The second work is introduced by Lander & Lesser (1990), who propose a Cooperating Expert Framework (CEF) to sup- port cooperative problem-solving among sets ofknowledge- based systems. The participating agents solve subproblems relevant to their specific expertise and integrate their efforts using conflict resolution strategies that are appropriate to the problem-solving context. All of the agents have a global knowledge of conflict resolution strategies. When a conflict is detected, agents involved in the conflict propose their alterna- tive resolution strategies. Eventually they agree on a resolu- tion scheme. Later, the conflict is resolved by one of the cho- sen agents based on that scheme. Werkman et al. (1990) developed a system called Design Fabricator Interpreter (DFI) which is a framework for dis- tributed cooperative problem-solving among construction agents. The DFI system reflects the distributed nature of the construction industry by providing a multi-agent architecture that models design, fabrication and erection processes. Con- flicting recommendations issued by design agents are re- solved by a third-party arbitrator agent, which makes sugges- tions based on the globally known conflict resolution knowl- edge. It operates in both passive and active mode. In passive mode, the arbitrator monitors the agent proposal process and intercedes when a problem is evident; in active mode, it medi- ates during the agent's proposal process when called upon by the agents. Adler et al. (1989) discuss methods of conflict resolution in the domain of telephone network traffic control. A homogene- ous group of agents has geographically divided responsibili- ties with no overlap. The basic problem that the agents are to solve is excessive demand for the resources in some parts of 228 ------------------------ Expert Systems, November 1993, Vol. 10, No 4 the network. Two negotiation protocols are described: con- flict-driven plan merging, a bottom-up approach to resolving a conflict that has already occurred, and shared plan develop- ment, a top-down approach to avoiding conflicts as plans are developed and refined. Their research addresses how con- flicts on the usage of resources could be resolved. 3. A distributed conflict resolution-based model for cooperating experts The cooperating experts environment is organised as a com- munity of cooperating problem-solving agents, where each agent is represented as a fully functional and autonomous knowledge-based system. The model is designed for solving problems in the domain of design. This model is based on the idea that each design agent has its own conflict resolution expertise separate from its domain-level design expertise, and that in the context of particular conflicts this expertise can be instantiated into specific advice for resolving these conflicts. The model allows a new problem-solver to be added or an existing one to be removed without requiring any modifica- tion to the rest of the system. The model can therefore be considered to achieve Open Systems Semantics (Open Sys- tems deal with large quantities of diverse information and exploit massive parallelism) (Hewitt 1986, 1991) in the sense that it not only allows scaleability (ability to increase scale of commitments) but also robustness (ability to keep commit- ments in face of conflicts) - two primary indicators of Open Systems Semantics. 3. 1. Architecture of the model The cooperative design environment (Figure 1) is composed of a set of design agents, which are fully functional knowl- edge-based systems, and a shared blackboard. The agents communicate by posting assertions in a shared language. This requires translation capabilities to be included within the agents. The shared blackboard is a public repository available to all agents; this gives one the ability to store 'global' infor- mation, although the information can only be used locally by the agents. Alternatively, it would be possible to convey infor- mation directly through point-to-point communication chan- nels or reserved-spot communication (Winston 1984). The shared blackboard is partitioned into four chunks, allowing fast access, delete and update operations of units. They are called problem, solution, proposal and conflict areas. The problem area of the shared blackboard contains the in- itial problem definition and overall requirements that must be taken into account by the design agents. A problem instance is a tuple of the form P = <0, G, C, l>, where O denotes the problem originator (an agent that defines the problem to be solved); G is the set of goals that must be satisfied for a design to be accepted; C is the set of constraints that design agents should not violate (some of the constraints may be violated through negotiation with the problem originator); and / de- SHARED BLACKBOARD l • • ~ ' ' Agent 1 Agent 2 Agent n Figure 1: The architecture of the problem-solving environment. notes the initial problem information (such as the layout of a room if the problem is to design an office). The solution area of the shared blackboard includes the evolving design tem- plate to which non-conflicting design commitments produced by agents are added. The proposal area includes partial and incomplete solutions at several layers of abstraction issued by design agents. De- sign agents assert their solutions as proposals into this area. A proposal instance is a tuple of the form Q = <0, A, R, Cc>, where O denotes the owner of the proposal; A is the set of proposed actions to update the current design template; R con- sists of the reasons justifying each of the actions in A (R could be empty because an agent may not provide reasons for its proposal due to its inaccurate or incomplete knowledge); and Cr is the confidence factor that indicates the confidence of the owner in generating such a proposal. This information would be useful for other agents if the owner utilised inaccurate or incomplete knowledge in producing its solution. Conflict area is the place where agents put their critiques related to a new design commitment. A portion of this area provides a communication medium with agents that are in- volved in a conflict situation. This area holds evaluation re- sults and conflict resolution recommendations issued by de- sign agents. An evaluation result instance is a tuple of the form ER= <0, Q, Ac, Ra, Re>, where O denotes the owner of the evaluation result tuple; Q is the identity of the proposal evalu- ated; Ac is the set of actions criticised (Ac i;;;; A in Q); Ra is the set of ratings ( evaluation results) for each action in Ac; and Re is the overall result of the proposal Q which is either 'conflict- ing proposal' or 'nonconflicting proposal'. A conflict resolu- tion instance is a tuple of the form CR= <0, Q,Ar,Rr>, where 0 is the owner of the conflict resolution tuple; Q is the identity of the related proposal; Ar is the set of refinements for those conflicting actions of Q; and Rr includes the reasons for the proposed refinements. A description of the internal structure of an agent in the model is given in Figure 2. An agent supports a knowledge base, a database and a general controller. The knowledge base Expert Systems, November 1993, Vol. 10, No 4 ________________________ 229 includes domain and control knowledge, just like in a classi- cal knowledge-based system. It also contains conflict resolu- tion knowledge to be used in cooperatively managing con- flicts with other agents. This knowledge is not known globally and varies with respect to the agent's beliefs and under- standing of the environment. The database includes facts, goals and constraints specific to the domain of that agent. Agents also maintain two types of history information related to the solution generation phase and conflict resolution phase. This not only makes backtracking possible but also allows case-based information to be used later for solving similar problems encountered. The general controller includes proce- dures for generating and evaluating design commitments, managing conflicts and translating messages into the com- mon language. KNOWLEDGE BASE DATABASE GENERAL CONTROUER Control Knowtedge Goal1 Proposal GaneraUon Constraints Proposal Evaluatlon Domain Knowledge Facts Conrllct RasoluUon Handler Conflict RasoluUon Knowledge Histories Translallon Procedures Figure 2: Internal structure of a design agent in the model. The agents in the model are actually heterogeneous agents in the sense that they might use different knowledge repre- sentation techniques and inference mechanisms. Each agent is assumed to generate proposals (solutions for subproblems) according to its knowledge. They cooperate to achieve the common goal of solving the global problem. The model can be augmented to support special agents like database systems. For the time being, we assume that each agent is a knowledge- based system which offers to solve subproblems (produce proposals) and cooperates with others in resolving conflicts through negotiation. Local knowledge is represented in what- ever language desired and cannot be accessed by any agent except its owner. Several knowledge representation tech- niques have been developed for the domain of design (Akman et al. 1990; Gero 1988; Goel & Pirolli 1989; Smithers & Troxell 1990; Sriram & Adey 1986; Tokoro & Ishkawa 1984). If the internal language is not the same as the shared language, translation procedures are incorporated within agents. In case of a conflict, agents might make some or all of their goals, constraints and even knowledge available to others. 3. 2. Problem-solving phases The problem-solving in the model is initiated by one of the agents asserting a problem definition P = <0, G, C, I> into the problem area of the shared blackboard. All interested agents are instantiated after examining the problem definition and they start producing design proposals related to their ex- pertise, knowledge and viewpoints. When a design agent gen- erates a design proposal Q = <0, A, R, Ct>, it is put into the proposal area. The agent producing this proposal also in- cludes explanation information that indicates which of the agent's goals and constraints caused such a proposal. This explanation allows other agents to understand why such a proposal has been asserted. After the generation of a proposal, all of the agents are sig- nalled. An agent does not interrupt its proposal generation process if it is already working on another proposal, but it immediately awakens another process - the evaluation process - that will run in parallel with the proposal genera- tion process. The evaluation process first informs the owner of the proposal whether it is going to criticise the proposal. If not, it will go to sleep and wait for another proposal to be asserted. If the agent is interested in the proposal, it evaluates and posts the result ER = <0, Q, Ac, Ra, Re> on the conflict area of the blackboard. The owner of the proposal also evalu- ates its proposal, usually as part of the solution generation process. It is necessary for the owner to indicate its confidence because it might use incomplete or inaccurate knowledge in producing its solution. The evaluation process results in a rat- ing to be produced which shows the 'quality' of a solution with respect to the goal criteria used to judge. The agents use their internal evaluation criteria and therefore may not share a common rating scale for their findings. After all the interested parties finish evaluating the newly asserted proposal, those agents which identify the proposal under consideration as conflicting with their beliefs come to- gether to resolve the conflict (those interested agents that put ER's in which Re part is 'conflicting proposal'). Agents in our model do not have a global knowledge of conflict resolution strategies, though in existing systems agents are assumed to be knowledgeable about the global conflict resolution exper- tise. In our model, each agent has its own conflict resolution knowledge that allows it to participate in the process of con- flict resolution. The result of conflict resolution is either revi- sion or abandonment of the proposed solution. When none of the interested agents detect any conflict re- lated to a proposal, the partial design template residing in the solution area is updated by using the design contribution ex- isting in the proposal. The process continues until the design template meets requirements specified by the agent that put the initial problem definition in the problem area of the shared blackboard. The design process may also be terminated, al- though the agent that put the problem definition is not satis- fied. This may happen in cases where none of the agents can generate a non-conflicting design proposal any more. 4. Conflict resolution in the model When an agent detects a conflict, it participates in the resolu- tion process based on its own conflict resolution knowledge. Each agent may utilise different conflict resolution strategies. For example, suppose that we are given the problem of de- 230 ------------------------ Expert Systems, November 1993, Vol. 10, No 4 signing an office. The functionality agent suggests that the PC desk be put close to the window so that a PC user could have a took outside when bored. On the other hand, the computer specialist, detecting a conflict, argues that sunshine could damage the PC. The computer specialist uses a conflict reso- lution strategy which says 'put electrical devices far away from windows'. The functionality agent, however, uses a do- main-independent resolution scheme 'try other subgoal alter- natives'. Eventually two experts revise the proposal such that the PC desk is put into a place in the office which is not ex- posed to sunshine, by using different resolution schemes. In deciding which strategy to apply, an agent uses information gathered up to the time the conflict has occurred as well as its conflict knowledge. This information includes: • Explanation embodied within the proposal (this will al- low other interested agents to understand the intent of such a proposal). • Critiques made by the interested parties to the proposal (after examining outcomes of evaluation procedures of other agents, an agent chooses an appropriate resolution strategy taking into account different viewpoints). • The relevance of the agent to a particular problem being solved (if an agent is more knowledgeable and capable compared to others, it should participate in resolution of a conflict according to its relevance). • Flexibility of agents involved in conflicts (this is impor- tant for an agent to decide how to behave in a compro- mise type of conflict resolution). • Behaviour and actions of other agents in resolving the conflict (by examining this information, an agent might decide to alter,the conflict resolution strategy it has been using). • Conflict resolution history information (if a similar con- flict situation was encountered beforehand, an agent could utilise history information in resolving the con- flict. This allows case-based reasoning in conflict reso- lution). • Number of agents involved in the conflict (depending on the domain, if the number of agents involved in a conflict situation exceeds a certain amount then some of the agents, thinking that they coµld not be effective for resolving the conflict compared to others, may continue to generate alternative solutions rather than participate in conflict resolution). • Available problem-solving resources. When a proposal is issued, each interested agent evaluates it to detect whether or not any conflicting recommendations ex- ist in the proposal. For understanding the kind of conflict that has occurred, each agent examines its conflict knowledge to see which of its conflict situations match (not necessarily a perfect match) the current conflict (if existing). Upon decid- ing on the conflict situation, an agent uses its conflict resolu- tion knowledge to overcome the conflict from its perspective. Conflict resolution knowledge is composed of general strate- gies (domain-independent) and specific strategies (domain- dependent ). Domain dependent strategies are gathered during the knowledge acquisition phase. Agents prefer to use the most specific strategies first for resolving a conflict. General strategies are resorted to last since they are computationally expensive and may lead to poor solutions. When an agent detects a conflict and chooses a strategy for resolving the conflict, it does not mean that the agent may not alter the resolution strategy it has chosen. That is, upon ob- serving the actions of other agents during conflict resolution phase, it may improve its understanding of the overall prob- lem and the particular conflict encountered. This allows agents to alter strategies if they think that they will benefit from doing so. When an agent proposes a revised solution based on its resolution scheme, it also explains why the new solution is a good candidate. This enables other agents in- volved in conflict to choose the most appropriate action on behalf of the resolution process. 5. A cooperating experts' problem: office design The following example is taken from the domain of office design to exemplify the problem solving process of the coop- erating experts that is used in our implementation. The reason for choosing this example is that it is in a concrete rather than an abstract domain and it can be understood easily because of its suitability for simple two-dimensional graphical repre- sentation. Here, we present a simplified layout problem for an office design and describe design agents and their interac- tions. A well-designed office encompasses different areas of expertise concerning aesthetics, functionality, energy effi- ciency, etc. In this example, we have incorporated four agents in the problem-solving framework. They are • the client agent; • the functionality agent; • the electricity agent; • the cost agent. The client agent is the one that puts forth the problem defini- tion specifying general constraints and the global design goal to be satisfied. This agent may be the one to use the office being designed or be the department chairman who is having the office designed for a prospective faculty member. The functionality agent uses specific heuristic search techniques in the area of space planning. The electricity agent is con- cerned with all the electrical and electronic devices and wir- ing including computers, telephones, facsimile systems, etc. The cost agent is required to control the overall cost of the design and avoid wasteful use of resources. When a proposal is generated, each interested agent evaluates it to detect a pos- sible conflict from its own perspective. A conflict is detected when an agent finds a conflict situation (upon examining its knowledge base) that matches the proposal under considera- tion. In this domain, some possible conflict situations are lo- cation of an object, dimensions of an object, quality of an Expert Systems, November 1993, Vol. 10, No 4 ________________________ 231 object, cost of an object, usage of an object, existence of an object and so on. The design process is initiated by the client agent who puts the following problem definition into the problem area of the shared blackboard: Problem Definition= < Client-Agent, { goal: (design office) subgoals: ((minimize amount-of-walking) (customize components-to-the- size-of-office) (maximize efficiency) (must-have PC) ... ) ) ) , { constraints: ( (to-be-used-by faculty-member) (number-of-occupants 1) (cost-of-design< $6000) ... ) ), { layout: shape: rectangular dimensions: ((length 10) (width 8) (height 2.5)) coordinates: (upper left (x O y O)) window: type: ((frame wood) (glass glass!)) dimensions: ((height 1)) coordinates: ((x0y3) (x0y5)) door: type: ((made-up-of wood)) dimensions: ((height 2)) coordinate: ( (x 8 y 7) (x 8 y 8)) electrical-plug: coordinates: ((x 6 y 0)) phone-plug: coordinates: ((x 6.5 y 0)) ) > Figure 3a shows the global layout of an office. In this exam- ple, we ignore the third (z) dimension; instead the height at- tribute of objects is used when necessary. Also, we are not concerned with the precise locations of objects. After examin- ing the problem definition, all of the interested parties start producing design commitments. First, the functionality agent, according to its expertise and understanding of the problem, asserts the following proposal into the proposal area of the shared blackboard which updates the template as shown in Figure 3b. Proposal-0 =<Functionality-Agent, { (put : object des kl : type desk : location (2 2)), (put :object chairl :type chair :location (3 0.5)), (put :object pcdeskl :type pcdesk :location (2.5 6)), (put :object chair2 { (utilize sunshine) (have better-view) nil > :type chair :location (3 5)) ), The functionality agent has decided to put a desk and a PC desk along with two chairs nearer to the window so that the occupant can not only have a good view but also utilise the (0,0) plug - 6m 4m -X 3m ;:: 6m 0 2m "O C: 'i 3m .... 0 0 10m 2~ 'O Ir y (3,0.5) Q chair! (2,2)EJ deskl (3,5) Qchair2 (2.5,6) lpcdeskl I (b) (a) Figure 3: (a) Global layout of the office; (b) Layout of the office after Proposal-0. sunshine. In generating this proposal the agent used the fol- lowing piece of its domain knowledge: if number of occupants is 1 and the occupant of the room is academic person then activate SINGLE-ACADEMIC-KB SINGLE-ACADEMIC-KB (only utilized piece of knowledge) find a place to put a desk (desk!) if the occupant is to use computer then include a computer desk (pcdeskl) and find a place to put the computer desk (pcdeskl) if a desk (desk!, pcdeskl) is to be put then put it near to the window reasons: utilize sunshine have a better view This proposal triggers the evaluation procedures within other interested agents. The client agent detects a conflict af- ter evaluating the proposal. With this configuration, the client 232 ------------------------- Expert Systems, November 1993, Vol. 10, No 4 agent notices that occupants must walk too much because they might need to use the PC (which will be put on the PC desk) a lot. The functionality and client agents combine to resolve the conflict encountered. The client agent uses a spe- cific resolution scheme which states 'keep frequently used objects close to each other', and the functionality agent uses a general conflict resolution strategy which is 'try other loca- tion alternatives'. The client agent has two alternatives to re- solve the conflict from its perspective; it may put the PC desk either to the left or to the right of the other desk. Taking into account the explanation within the functionality agent's pro- posal, the client agent proposes to put the PC desk to the left of the other desk so that the PC desk will be close to the window and hence the occupant can utilise sunshine and have a better view. The revised and agreed solution is shown in Figure 4a. Below is the description of how conflict is detected and resolved by both agents: Conflict Detection Tuple= < Client-Agent, Proposal-a, { (all-actions) } , { (increased amount-of-walking) }, "Conflicting-Proposal"> The Client Agent: Conflict Situation: Location of Objects Resolution : If the objects involved in conflict are to be used very often then try to keep them close to each other The Functionality Agent: Conflict Situation: Location of Objects Resolution : try other configuration alternatives The agreed resolution actions: (move :object deskl :position (3. 5 2)) (move :object chairl :position (4 0. 5)) (move :object pcdeskl :position (0. 8 2)) (move :object chair2 :position (1 0. 5)) Then the cost agent realises that there is no need to have two chairs close to each other. The cost, client and functionality agents decide to remove one of the chairs as shown in Figure 4b. In the resolution phase, agents agree on rotating the PC desk such that a chair could be used for both desks: Conflict Detection Tuple= <Cost-Agent, Proposal-0', { (existence-of :object chair2) }, { }, "Conflicting-Proposal"> The Cost Agent: Conflict Situation: Cost of Objects Resolution : remove them The Client Agent: Conflict Situation: Existence of Objects Resolution : If existence of some objects causes some problems then remove them (1,0.5) (4,0.5) <:)chair2 (::)chairl (0.8,2) (3.5,2) 1pcdeskl 1 B (a) (0.5, l) (4 1) [I] . ochairl "' (3.5,2.5) B (b) Figure 4: ( a) Layout of the office after resolving the conflict in Proposal-0. ( b) Layout of the office after resolving the conflict in Proposal-0'. The Functionality Agent: Conflict Situations Location of Objects - Existence of Objects try other configuration alternatives Resolution upon removing unwanted objects The agreed resolution actions: (remove :object chair2) (move :object deskl :location (3.5 2.5)) (move :object chairl :location (4 1)) (rotate :object pcdeskl -90) (move :object pcdeskl :location (0.5 1)) Later, the electricity agent decides to put the PC on the PC desk and use an extension cord to connect the PC to the plug, and puts a proposal related to these modifications into the proposal area (Figure 5a): Proposal-1 =<Electricity-Agent, { (put :object pcl :type ibmpc :location (0.7 1.5) ), (put :object cordl :type cord :path ( (0.8 1.5) (0.8 0.5) (6 0.5) (6 0.5)) }, { }, nil > Expert Systems, November 1993, Vol. 10, No 4 _________________________ 233 The knowledge used to generate Proposal-1: if there is a suitable place (pcdeskl) to put an electrical device (pell on top of then put the electrical device (pell on top of it (pcdeskl) if an electrical device (pell is not connected to the plug and the electrical device's (pcl's) own cable is short then use an extra and long enough cord (cordl) to connect the electrical device (pell to the plug The cost agent detects another conflict related to the elec- tricity agent's proposal, which is the cost of using an exten- sion cord. The electricity, cost and functionality agents are involved in the resolution of the conflict. The cost agent de- cides to remove the extension cord and is very inflexible in its decision. The functionality agent does not have a specific resolution scheme in its mind, but uses a general one and (6,0) (0.8,0.5' : r--J--------------------·c6.o.5) l!lpcl O chm,! pcdeskl B (a) (6,0) (5.5,0.5) ·----~ (3,1) (5,1) fril (6,0.5) (2,3) Q_,, ~pc! B pcdeskl (b) Figure 5: (a) Layout of the office after Proposal-I. ( b) Layout of the office after resolving the conflict in Proposal-I. decides to put PC desk to the right of the other desk. Although the functionality agent knows that it is better to keep the PC desk close to the window for reasons of utilising sunshine and having a better view, it does not insist on this preference for resolving the conflict. The electricity agent has a domain spe- cific resolution strategy which states 'keep all of the electrical devices close to the plugs'. Below is how this particular con- flict is detected and resolved: Conflict Detection Tuple= < Cost-Agent, Proposal-1, I (cost-of :object cordl) ), I ), "Conflicting-Proposal"> The Cost Agent: Conflict Situation Cost of an Object Resolution - remove the object The Electricity Agent: (preferred resolution alternative) - if some other object of the same type but cheaper exists then try to remove the expensive one - if there exists some other unnecessary or less important object in the template then try to remove that object Conflict Situations: Location of Objects and Existence of Objects Resolution - try other configuration alternatives upon removing unwanted objects - if an electrical device (pell is to be put then put the device (pell close to the plug (preferred resolution alternative) The Functionality Agent: Conflict Situations: Location of Objects and Existence of Objects Resolution try other configuration alternatives The agreed resolution actions: (remove :object cordl) (move (move (move (move :object deskl :object chairl :object pcdeskl :object pcl :location (2 3)) :location (3 1)) :location (5 1)) :location (5.5 1.5)) (plug-in :object cord of pcl :to plug) By explaining their resolution steps the three agents come to an agreement to put the PC desk close to the plug, as shown in Figure 5b. Later, the electricity agent connects the PC's cable to the plug without using an extension cord. The design pro- 234 ---~--------------------- Expert Systems, November 1993, Vol. 10, No 4 ceeds in this manner until reaching the requirements specified by the client agent. In this example, we only gave a segment of the problem- solving process, emphasising the resolution of conflicts with- out considering precise locations of objects. 6. Conclusions The cooperating experts' approach has an important role in the field of Distributed Artificial Intelligence because many of the problems that are being encountered in real life require the application of complex and diverse expertise. One of the important problems faced in a cooperating community of ex- perts is how to detect and resolve conflicts occurring at any phase of problem solving. Existing approaches to conflict resolution rely on coordinated conflict resolution strategies (Adler et al. 1989; Klein & Lu 1989; Lander & Lesser 1990; Werkman et al. 1990). In these approaches, each agent is as- sumed to have a global knowledge of conflict resolution in- formation. When conflicts happen, agents agree on a conflict resolution scheme and one agent resolves the conflict using a globally agreed resolution strategy. In this paper, we introduce a new cooperating experts envi- ronment for solving problems that openly support multi-agent conflict detection and resolution. In this environment, each agent is free to choose the most appropriate action, given its understanding of the global and local situation and its own capabilities. Each agent has its own conflict resolution knowl- edge which is not accessible and known by others. Further- more, there are no globally known conflict resolution strate- gies. Each agent involved in a conflict chooses a resolution scheme according to its self-interest. Agents might use differ- ent strategies of their own and might still agree on a solution. The model achieves flexibility in its problem-solving, which is the most compelling argument for building modular multi- agent systems. Anew agent can be added or an existing one can be removed without any modification of the rest of the system. This characteristic of the model satisfies the requirements of Open Systems Semantics. However, in the existing approaches, addition or removal of an agent requires that the global conflict resolution knowledge be reformed accordingly. This approach is very similar to conflict resolution in human problem-solving. Existing approaches are too restrictive and applicable only to the problems where experts must agree on a known strategy for resolving conflicts. The new approach also allows agents to alter strategies in the resolution phase if they think it is wise to do so. The model also requires agents to explain why a particular action is made in order for other agents to reason about it. This means that agents may explain the reasoning behind their proposed actions in putting their proposals and in detecting and resolving conflicts. Currently, we are implementing the model on intercon- nected SUN-4 workstations. All the problem-solvers - agents - are modelled as processes running on different workstations that communicate over Ethernet. In order to ii- lustrate the problem-solving phases in the model, we have chosen to solve design problems. We will test the conflict resolution-based model on various other examples in the do- main of design and medical diagnosis. References ADLER, M.R. et al. (1989) Conflict resolution strategies for non-hi- erarchical distributed agents, Distributed Artificial Intelligence, 2, M.N. Huhns, pp. 139-161. AKMAN, V., P.J.W. TEN HAGEN and T. TOMIYAMA (1990) A funda- mental and theoretical framework for an intelligent CAD system, Computer-Aided Design, 22(6), 352-367. CAMMARATA, S., D. McARTHUR and R. STEEB (1983) Strategies for cooperation in distributed problem solving, in Proc. Int. Joint Conf Artificial Intelligence, Karlsruhe, Germany, August. pp. 767-770. CHAIB-DRAA, B. et al. (1992) Trends in distributed artificial intelli- gence, Artificial Intelligence Review, No.6, pp. 35--66. CHANDRASEKERAN, B. ( 1982) Decomposition of domain knowledge into knowledge source: The MDX approach, Fourth National Conf of Canadian Society for Computational Studies of Intelli- gence, Canada. DURFEE, H.D., V.R. LEsSER and D.D. CORKILL (1989) Trends in cooperative distributed problem solving, IEEE Transactions on Knowledge and Data Engineering, 1(1), March, 63-83. ERMAN, L.D., F. HAYES-ROTH, V.R. LESSER and D.R. REDDY (1980) The Hearsay-II speech understanding systems: integrating knowledge to resolve uncertainty, Computing Surveys, 12, June, 213-253. GASSER, L., (1991) Social conceptions of knowledge and action: DAI foundations and open systems semantics, Artificial Intelli- gence, 41, 107-138 GERO, J.S. (Ed.) (1988) Artificial Intelligence in Engineering De- sign, Elsevier, UK. GINGBERG A. (1988) Knowledge base reduction: a new approach to checking knowledge bases for inconsistency and redundancy, AAA/. GoEL, V. and P. PIROLLI ( 1989) Design within information-processing theory: the design problem space, Al Magazine, Spring, 19-36. GoMEZ, F. and B. CHANDRASEKARAN (1984) Knowledge organiza- tion and distribution for medical diagnosis, Technical Report, 84-FG-FGBC, Department of Computer and Information Sci- ence, The Ohio State University. HEwm, C. ( 1986) Offices are open systems, ACM Transactions on Office Information Systems, 4(3), 271-287. HEwm, C. ( 1991) Open information systems semantics for distrib- uted artificial intelligence,Artificial Intelligence, 47, 79-106. KLEIN, M. and S.C-Y. Lu (1989) Conflict resolution in cooperative design, Artificial Intelligence in Engineering, 4( 4 ), 168-180. LANDER, S. and V.R. LESSER ( 1990) Conflict resolution strategies for cooperating expert agents, International Conference on Cooper- ating Knowledge-Based Systems, Keele University, October. LESSER, V.R. and D.D. CORKILL, (1988) The distributed vehicle monitoring testbed: a tool for investigating distributed problem solving networks, in R.S. Engelmore and A. Morgan (Eds.), Blackboard Systems, Addison-Wesley, pp. 353-386. MALONE, T.W. and K. CROWSTON (1991) Towards an interdiscipli- nary theory of coordination, Technical Report, CCS-TR-120, Center for Coordination Science, MIT, Cambridge, MA. NGUYEN, T.A. et al. (1987) Verifying consistency of production sys- tems, in Proceedings of the Third Conference on Artificial Intel- ligence Applications, pp. 4-8. POLAT, F. and H.A. GUVENIR (1991a) Coordination issues in distrib- uted problem solving, in Proceedings of the Sixth International Expert Systems, November 1993, Vol. 10, No 4 ________________________ 235 Symposium on Computer and Information Sciences, Elsevier, Vol. 1, Antalya, pp. 585-594. POLAT, F. and H.A. GUVENIR (1991b) A unification based approach for knowledge base verification, Expert Systems, 8(4), 251-259. POLAT, F. and H.A. GUVENIR (1992) A conflict resolution based co- operative distributed problem solving model, in Proceedings of AAAI-92 Workshop on Cooperation among Heterogeneous In- telligent Agents, San Jose, CA, July. RoussET, M.C. (1988) On the consistency of knowledge bases, in Proceedings of European Conference on Artificial Intelligence. SHEKHAR, S. and C.V. RAMAMOORTHY (1992) Coop: a self-assess- ment based approach to cooperating expert systems, Interna- tional Journal on Artificial Intelligence Tools, 1(2), 175-204. SMITH, R.G. and R. DAVIS (1988) Frameworks for cooperation in distributed problem solving, in A.H. Bond and L. Gasser (Eds.), Readings in Distributed Artificial Intelligence, Morgan Kauf- mann, pp. 61-70. The authors SMITHERS, T. and W. TROXELL ( 1990) Design is intelligent behav · but what's the formalism? AI EDAM, 4(2), 89-98. iour SRIRAM, D. and R. ADEY (1986) Applications of Artificial Intel[". gence in Engineering Pr~bl~ms? Vols._ 1-2,. Springer-Verlag. 1 SYCARA, K.P. (1989) Negotiation m design, m Proceedings of th MIT-JSME Workshop on Cooperative Product Developmen7 MIT, Cambridge, MA. ' TOKORO, M. and Y. ISHKAWA (1984) An object-oriented approach to knowledge systems, in Proc. of the Int. Conj on Fifth Generation Computer Systems, pp. 623--031. VIGNOLLET'. L. and M. A YEL (1990) SYCOJECT: A tool for building automatically sets of test for knowledge bases, Applications of Artificial Intelligence, No. 8, pp. 192-201. WERKMAN K. et al. (1990) Design and fabrication problem solving through cooperative agents, NSF-ERC-ATLSS Technical Re- port 90--05, Lehigh University, Bethlehem, PA. WINSTON, P.H. ( 1984) Artificial Intelligence, Addison-Wesley. Faruk Polat Faruk Polat is a PhD candidate in Computer Engineering and Information Science at Bilkent University in Ankara, Turkey. He has been working on his dissertation as a visit- ing NATO science scholar at the University of Minnesota, Minneapolis, for the 1992-93 academic year. His research interests include knowledge-based systems, knowledge base verification and distributed artificial intelligence. He received his BS in Computer Engineering from the Middle East Technical University, Ankara, in 1987 and his MS degree in Computer Engineering ·and Information Science from Bilkent University in 1989. H. Altay Guvenir H. Altay Guvenir is an Assistant Professor of Computer Engineering and Information Science at Bilkent University in Ankara, Turkey. His research interests include expert systems, machine learning and computational linguistics. He received his MS in Electrical Engineering from Istanbul Technical University in 1981 and his PhD in Computer Science from Case Western Reserve University in 1987. He is a member of AAAI, ACM, ACM SIGART and the Inter- national Association of Knowledge Engineers. Shashi Shekhar Shashi Shekhar received the B.Tech degree in Computer Science from the Indian lnstituJe of Technology, Kanpur, India in 1985, and the MS degree in Business Administra- tion and the PhD degree in Computer Science from the University of California, Berkeley in 1989. He is currently an Assistant Professor in the Department of Computer Sci- ence at the University of Minnesota, Minneapolis. His re- search interests include databases, artificial intelligence and software engineering with emphasis on important applica- tions such as manufacturing. Dr Shekhar is a member of the IEEE Computer Society, ACM and AAAI. 236 ------------------------- Expert Systems, November 1993, Vol. 10, No 4 MX-M316NV_20190222_101402_Page_1_2R MX-M316NV_20190222_101402_Page_2_1L MX-M316NV_20190222_101402_Page_2_2R MX-M316NV_20190222_101402_Page_3_1L MX-M316NV_20190222_101402_Page_3_2R MX-M316NV_20190222_101402_Page_4_1L MX-M316NV_20190222_101402_Page_4_2R MX-M316NV_20190222_101402_Page_5_1L MX-M316NV_20190222_101402_Page_5_2R MX-M316NV_20190222_101402_Page_6_1L 
work_rmyv4p2swra7jhyawjyxsgolki	----	Internet of Things based Expert System for Smart Agriculture (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 Internet of Things based Expert System for Smart Agriculture Raheela Shahzadi Department of Computer Science COMSATS Institute of Information Technology Sahiwal, Pakistan Javed Ferzund Department of Computer Science COMSATS Institute of Information Technology Sahiwal, Pakistan Muhammad Tausif Department of Computer Science COMSATS Institute of information Technology Vehari, Pakistan Muhammad Asif Suryani Department of Computer Science COMSATS Institute of Information Technology Sahiwal, Pakistan Abstract—Agriculture sector is evolving with the advent of the information and communication technology. Efforts are being made to enhance the productivity and reduce losses by using the state of the art technology and equipment. As most of the farmers are unaware of the technology and latest practices, many expert systems have been developed in the world to facilitate the farmers. However, these expert systems rely on the stored knowledge base. We propose an expert system based on the Internet of Things (IoT) that will use the input data collected in real time. It will help to take proactive and preventive actions to minimize the losses due to diseases and insects/pests. Keywords—Internet of Things; Smart Agriculture; Cotton; Plant Diseases; Wireless Sensor Network I. INTRODUCTION Internet of Things (IoT) is a broad term that describes the interconnection of different daily life objects through the internet. In the concept of IoT every object is connected with each other through a unique identifier so that it can transfer data over the network without a human to the human interaction [1, 2]. IoT has referred as a network of everyday objects having ubiquitous computing. The ubiquity of the objects has increased by integrating every object with embedded system for interaction [21]. It connects human and devices through a highly distributed network. IoT is basically the world wide interconnection of devices. The aim of IoT is to connect every person and every object through the internet. In IoT ,every object is assigned a unique identifier, so that every object is accessible through the internet [22][23]. Every object in the IoT has the following three capabilities: awareness, representation, and interaction. Awareness is the ability of the smart objects to understand and sense other objects. Representation is the ability of the objects to present, according to the programming concept. Interaction is the ability to communicate with each other The IoT is evolving, growing and becoming popular day by day; in the today’s world, around 5 billion objects have connected through the internet. In 2020, it has estimated that near about 50 billion objects will be connected to the internet [24]. IoT is providing tremendous opportunities for novel applications, which is now widely used in many aspects of life such as intelligent home monitoring system, products supply chain management, precision agriculture and much more. Every object in IoT is addressable, recognizable, readable and locatable through the internet by using RFID (Radio Frequency Identification), Wireless Sensor Network (WSN) or other means. The concept of IoT is using many in different domains such as; precision agriculture [1, 2], products supply chain management [3], Smart Grid [4] , environmental monitoring [5], cloud computing [6] and many more. IoT is gaining much importance these days as every object in the network will become a computer [7]. The idea of IoT has become successful due to the invention of recent technologies like sensors, RFID and WSN. Pakistan is an agricultural country. Although the industry and services sector has transformed Pakistan into a diversified country, still a major part of GDP is contributed by the agriculture sector. The foreign exchange of Pakistan has depended on agricultural products. More than half of the population of Pakistan lives in rural areas and major source of earnings of this population has based on agriculture. Most of the industry in Pakistan is also dependent on agriculture like textile industry, sugar industry, flour industry, juice industry, furniture industry, dairy industry, etc. [2]. Farmers in Pakistan are not aware of the technology and lack agricultural knowledge. They rely on traditional methods and practices. However, the agriculture field in the advanced world has evolved a lot due to the advances in technology and equipment. Pakistan faces huge losses in agriculture due to the following factors. • Delayed sowing and poor seed quality. • Environmental hazards. • Insect and disease attacks. • Unplanned irrigation and water losses. • Untimely harvesting. • Misuse of fertilizer and insecticides. 341 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 • Lack of machinery and equipment. • Mishandling of ripened crops. Cash crops have a major share in the economy of Pakistan and cotton crop is very important among them. It is also called as ‘white gold’. Heavy losses occur every year due to poor farming practices, attack of pests at different stages and attack of diseases. According to a survey [8], 38 percent loss of cotton crop occurred due to the attack of insects in 2013. According to another survey, due to viral attack ,15 percent loss of cotton crop occurs every year. Cotton crop is affected by some of bacterial and fungal diseases as well as pests and insects. Temperature and humidity requirements for the cotton crop are different from other crops before and after sowing. The timing of spray of insecticides, pesticides, and application of fertilizer also affects the crop growth. So, the continuous monitoring is required to keep the crop healthy. Most of the farmers in Pakistan are illiterate, and they are unaware of the latest research in the field of agriculture. The farmers normally take guidance from agriculture experts and other experienced farmers. However, the experts are not always available every time and everywhere [9]. So, expert systems have been developed for different crops, fruits and vegetables in the world. Basically an expert system (ES) is a computer program which solves the problems as human being solves the problem. It is a tool which generates output using its knowledge base, so it replicates the behavior of the human. The ES can pinpoint the problems as well as figure out the solutions. It combines the same domain knowledge of different experts. In the ES accumulation of knowledge from different sources is a very important factor. The ES can provide output whenever it has given input. It means it should be easily available to the farmers. In this paper, we present an Expert System based on the concept of Internet of Things. Sensors will collect data and automatically send it to the ES. The ES will process the information and send the results or decisions to the farmer’s mobile phone. In this way, crops can continuously monitor ,and timely decisions can be taken. It will help to minimize the losses due to sudden disease and pest attacks through timely proactive and preventive actions. The proposed ES is initially developed for the Cotton crop and evaluated by the farmer community. The proposed system can be used for. • Efficient Crop Management. Irrigation Control. • Environment Warnings and Guidance. • Optimal usage of fertilizers, insecticides and pesticides The rest of the paper has organized as follow , in section 2, we discuss the related work. In section 3 we describe proposed ES based on IoT, Section 4 we describe implementation and evaluation of proposed ES based on IoT, In section 5 we describe the results , In section 6, we elaborate the validity of proposed solution and , In section 7, we conclude the whole work and describe future work. II. RELATED WORK Fan TongKe [10] proposed the smart agriculture based on cloud computing. The author presented the architecture for the smart agriculture based upon the concept of the IoT and cloud computing. Agriculture information cloud was combined with Internet of Things to achieve the dynamic distribution of resources and balance of the load. Ji-Chun Zhao et al. [7] studied the applications of IoT in agriculture. The authors proposed a monitoring system based on internet and wireless sensor networks. An information management system was designed to provide the data for research in agriculture. The authors developed software for monitoring of the fields like data acquisition about the fields, data processing models, and system configuration module. The developed application provides accurate control for the monitoring of the green house. Agrawal and Lal Das [2] discussed the possible future applications and challenges faced by the IoT technology. They presented some key challenges in IoT applications such as: standards, privacy, security, authentication and identification, trust and ownership, integration, coordination, and regulation. They stated that the use of RFID (Radio Frequency Identification), Wireless Sensor Network (WSN) and mobile communication technologies would reduce the gap between theoretical and practical implementations of IoT applications. Chen and Jin [11] proposed the ‘Digital Agriculture’ based upon IoT. The working of the digital agriculture is divided into two steps: in the first phase, the information about the temperature, the wind, the soil contents, etc. is collected by different sensors. In the second phase, ZigBee transfers information. The agricultural products has labeled with EPC code. The EPC code reader reads the code of the products. Li Li et al. [4] discussed the application of smart and Wi-Fi based Wireless Sensor Network in IoT. The authors discussed the applications of IoT-based upon Wi-Fi, WSN and smart grid. Smart grid provides the intelligent data collection application, improving reliability of data collection and providing accurate information. IoT provides the intelligent environment monitoring application; water data and air data collected through sensors and sent to server for further processing. They proposed the concept of the precision agriculture. The authors stated that new WSN technology is better as compared to ZigBee. Hussain et al. [12] proposed the application of Internet of Things (IoT) technology in animal stock chain management. By using the RFID technology, anyone can be tracked or monitored. They discussed some operational principle of IoT. RFID technology is used for the unique identification of objects; each object in the RFID is labeled with EPC code. The authors proposed the use of this technology for maintaining all records for livestock management. Kosmatos et al. [13] proposed the architecture based on the RFID and smart objects. RFID objects will perform the primitive functionality in the proposed architecture while the smart objects will perform the complex functionality. The architecture was proposed based on the integration of RFID and smart objects. RFID tags have widely used for the identification of objects. So the RFID is used in the proposed architecture for tracking of the objects. The authors used 342 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 service oriented architecture and semantic model-driven approach in the proposed architecture. Carvin et al. [14] proposed the ubiquitous cognitive management system based on IoT and ubiquitous computing. They presented the problems as well solutions of problems. The basic idea was to use ambient intelligence provided by smart objects to serve the human, improving communication by context information. Zhou and Zhou [15] proposed a management model based on IoT for visualization and traceability of agricultural products. The aim was to ensure food safety and promote sustainable development of modern agriculture. The authors used the products logistic information along with Internet of Things for effective products supply chain management. Prasad et al. [16] proposed an expert system for the diagnosis of pests, diseases, and disorders in mango. The system had developed in ESTA (Expert System Shell for Text Animation). In the proposed system, the first step is the knowledge acquisition; the second step is the diagnosis of disease based on the input. They briefly described the type of mango diseases and recommendations for the disease control on the basis by visual symptoms. Sarma et al. [17] proposed an expert system for diagnosis of disease in rice plant in India. The purpose of the expert system is to assist farmers in solving the problems. The first step is the development of the knowledge base in the form of condition rules. The proposed system is easy to use and will be useful to those people who are unable to get the assistance of some agriculture expert. Kaliuday et al. [18] proposed rule based expert system for the prevention of pest diseases in rice and wheat crops. They built an expert system (ES) called AgPest for the diagnosis of pest disease in wheat and rice. They developed AgPest in CLIPS. It consists of the IF then else rule for finding the disease. They formulated the rules about the diseases of wheat and rice from different online sources. Negied [19] proposed the expert system for the protection of the wheat yield in Egypt. The proposed system is developed using following steps; the first step is the problem identification of the domain, the second step is information . They developed the system in MATLAB. The proposed system is helpful for improving the yield of wheat crop and for providing assistance to the farmers in the remote areas. Kaur et al. [20] proposed the expert system for the detection and diagnosis of the leaf diseases in cereals. It is quite difficult for the farmers to identify the leaf diseases without the assistance of the experts. For the identification of the diseases, they proposed image comparison techniques in JAVA. They used techniques like affine transformation and edge detection for this purpose. It is web based expert system so that it can be accessible from any web-enabled system A. Expert System in Agriculture Expert system (ES) is the branch of artificial intelligence that deals with the development of computer programs which can solve the problems as the human beings solve the problems [25]. The application of expert system in agriculture is increasing widely since many years. A number of expert systems have been developed in the field of agriculture such as • AMRAPALIKA is the ES for the diagnosis of disease, pest and disorders in Indian mango [16] • An expert system for diagnosis of disease in Indian rice plant[4]. • CITEX: An expert system for citrus crop management[26]. • CUPTEX: An expert system proposed for the management of cucumber[27]. • An expert system for the olive crop diseases and weed identification in Spain[9]. • LIMEX: An expert system for the management of lime crop[28]. • CALEX: An expert for the diagnosis of peach and nectarine disorders[29]. • CITPATH: An expert system for the diagnosis of fungal disease in citrus fruit[30]. The first step in the development of any ES is the problem identification. For example, if we are developing the ES for cotton crop then symptoms are identified. The description of the diseases can be textual or in the form of images. After that rules are formulated based on the concept of if then else structure. Most of the farmers in the remote areas are illiterate not having proper knowledge of dealing with diseases. Some diseases of the crops are difficult to distinguish because two or more diseases have the same symptoms. So, it creates problems for the farmers. This problem can overcome with the help of ES by combining the knowledge of different experts in one application. Most of the researchers are trying to develop the ES for fulfilling the needs of the farmers. If farmers. The farmer gets the assistance in time, the productivity rate of the crops will increase III. IOT BASED EXPERT SYSTEM For overcoming the problems of agriculture, we develop an initial framework based on IoT. The proposed solution consists of three main components the first component is the deployment of sensors in the field; we deploy soil sensors, humidity sensors, and temperature and leaf wetness sensors in the fields. Sensors collect the data and send it to sever, on the serve side we deploy the expert system, which processes the data and send the recommendations to the farmers about crops. A. Deployment of Sensors The sensors have deployed in the fields for the collection of data about the environment, humidity, soil moisture and leaf wetness. For the collection of data waspmote agriculture sensor board is used because it is specially designed for handling agriculture activities. The sensor board consists of AT mega 1281 microprocessor and 2GB micro SD –card. Every sensor board consists of four different types of sensors, the soil sensor, humidity sensor, temperature sensor and leaf wetness sensors. 343 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 We use three soil sensors, temperature sensors, humidity sensors and leaf wetness receptively for a precise and accurate measure of soil contents, environmental temperature, the humidity level in the environment and leaf wetness at the same time. The communication module XBee-802.15.4 is present in wasp-mote agriculture board. It can communicate with microcontroller at the rate of 38400 bps. The range of transmission is near about 500 meters. The gateway is the bridge between sensor nodes and server. It can communicate wirelessly with the sensor and through USB port with a computer. We conducted this experiment under controlled environmental conditions. The overall architecture of sensors communication has described in Fig 1 Fig. 1. Communication of Sensors Data in IoT Based ES B. IoT- Based Expert System for Cotton Crop IoT-based Expert System is different from traditional Expert Systems regarding inputs. It uses real-time input data gathered with the help of sensors. The sensor nodes send data to the gateway after the defined interval of time. The server receives data through the USB port. For storing and copying the cool term software, is used. The expert system deploys in the server process the data and sends the recommendation to the farmer cell phone. For solving this problem, the expert system based on the concept of IoT has proposed in Fig 2. Fig. 2. IoT Based Expert System The rest of the working of the proposed ES is similar to the traditional Expert Systems. It is implemented using CLIPS (C Language Integrated Production System) developed by NASA [31]. CLIPS is a C based instead of LISP and supports three programming approaches: rule-based, object oriented and procedural. It is portable, extendable, can be easily integrated and supports interactive development. It also has features for verification and validation of expert systems. The proposed expert system consists of the following main components. • Knowledge Base • Inference Engine Agenda • Working Memory • Explanation Facility • User Interface The structure of the expert system has shown in Fig 3. Fig. 3. Expert System in CLIP The first step in the development of expert system is the knowledge acquisition of the domain. The most important thing in the knowledge acquisition is what type of knowledge we require for the expert system. In the proposed system, we need data about pests, insects, diseases, weeds and growth environment required for cotton crops. The knowledge acquisition can have accomplished in three ways. • Experts of the domain are interviewed. 344 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 • Research articles about the domain are reviewed. • Information is obtained by field observations. Field observation may not be much stronger, as being computer scientist we are unaware about the diseases. Another thing which makes the field observation weak is that, all types of diseases, pests, weeds and insects are not spotted on the cotton crop at the same time. We gather the majority of the information by conducting interviews with experts. We get information about the type of diseases, causes of diseases, symptoms of diseases, the insects which attack on the cotton crop, the weeds which destroy the cotton crop, and the insects which spread the disease from one plant to another. Different sensors collect the data, soil sensors collect the data about soil condition, soil moisture and soil content, while weather sensors collect the data about humidity and temperature. Sensors send the data to the server, the server decides about the diseases on the basis of the fact list which is used for the training. In CLIPS, fact list, rule list, and agenda with the activation list kept in memory. All the facts have based on simple if then else logic. The sensors collect data and send to the server, on the server side, we deploy the expert system which processes the data and analyzes the data and send the recommendation to the farmer about crops. C. Server Send Recommendation to Farmer The server processes the sensor data, after processing it send the recommendation to the farmer cell phone. So, for farmer convenience, we develop an android app for farmers. Farmers install the android app on their phone. The server sends the recommendation to the farmer cell phone. The server sends the recommendation in English; farmer can can covert the recommendation to the Urdu or Punjab in his convenient language. IV. IMPLEMENTATION AND EVALUATION This section describes the implementation and evaluation of the expert system for SA. The sensors collect the real-time data and send to the server. On the server side, ES have deployed for extracting the information from sensor data. A. Expert System for CLIPS In this section, we implement different fact list of insect diagnosis, pest diagnosis, weeds diagnosis and irrigation scheduling. Table 1 describes the different fact lists of insects, pest symptoms and recommendation for the attack of insect pest. Table 2 describes the different fact lists of weed symptoms and recommendations for the attack of weeds. Table 3 describes the different fact lists of sucking insects and recommendation for the attack of sucking insects. The irrigation of crops depends on several factors like soil moisture, soil type, depth of root zone and the environment. Every soil has different physical properties and textures like coarse soil, medium texture, soil and heavy fine textured soil. The water capacity of every soul is different, so the amount of irrigation also varies according to the texture of the soil. Environmental fluctuations are also important factors for scheduling of irrigation. The cotton crop has a specific limit of water depletion, if the water gets depleted more than the specified threshold limit, then irrigation should be applied. To calculate the minimum flow of irrigation q (in cubic meters per hour), we used Equation 1. In this equation, Dg is gross application dose, A is an area, I is interval of day, T is operating hour per day and 10 is a constant for hectare. In equation 1 we are presenting a formula for calculating the irrigation dose. q=10𝐴𝐴∗𝐷𝐷𝑔𝑔/𝐼𝐼+𝑇𝑇…………………………..(1) Table 4 describes the irrigation scheduling in the cotton crops. V. RESULTS We conduct this experiment in the cotton fields Sahiwal from June 2015 to December 2015. The server processes the data and sends the recommendation to the farmer. The proposed ES provide diagnose diseases, attack of weeds and attack of pests, it provides the pesticide recommendation for weeds, diseases, and pests. It provides predication of diseases based on sensor data. It provides irrigation scheduling based on temperature and soil contents; it can also provide the dose of irrigation. After that sensors send data to the server, on the server side, we deploy the expert system which processes the sensor data and sends the recommendation to the farmers. Table 5 describes the comparison of the expert system, which has presented during different eras. In the previous expert system, the user manually inputs the symbol of diseases, and they use web-based, ontology based and expert system tool for the development of expert systems. In previous literature, they did not use the concept of IoT for the collection of data. In our proposed solution, we develop an expert system based on the concept of IoT. Different kind of sensors is deployed in the fields which monitor the crops, and send the data to the server, server process the data and send a recommendation to the farmer. Fig 4 represents the relationship between temperature and humidity sensor data, temperature, and humidity, inversely proportional to each other. If temperature increase then humidity level decrease. We are just representing the 117 recording of sensor data. By taking these sensor data, we are scheduling the automatic irrigation. 345 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 TABLE I. INSECT SYMPTOMS AND INSECTICIDES RECOMMENDATION Insects Symptoms Insect Diagnosed Insecticides Recommendation If location=underside of leaves and body IF ?insectpest= Whitefly Then color=yellowish and wing color=white then (?insecticides= Polo500SC ˄ ?dose= 250ml) ˄ insectpest diagnosed. (?insecticides= Confidor200SL˄ ?dose= 250ml)˄ (?insecticides= Mospilan200SP ˄ ?dose= 5gm) ˄ (?insecticides= Danitol30EC ˄ ?dose= 200ml And Use Neem Leaf Extract If temp=warm and shape=spindle shaped and IF ?insectpest=Thrips Then wings=elongated then Thrips diagnosed. (?insecticides= Confidor200SL ˄ ?dose= 80ml) ˄ (?insecticides= Confidor70WS ˄ ?dose= 5gm/kg seed) ˄ (?insecticides= Mospilan 20SP ˄ ?dose= 5gm)˄ (?insecticides= Thiodan 35EC ˄ ?dose= 600ml) If leaves curl downward=yes and IF ?insectpest=Jassid Then color=yellowish then Jassid diagnosed (?insecticides= Baythroid TM 525EC ˄ ?dose= 100ml) ˄ (?insecticides= Nurelle D 505EC ˄ ?dose= 500ml) TABLE II. WEEDS SYMPTOMS AND PESTICIDES RECOMMENDATION Weeds Symptoms Weeds Diagnosed Herbicide Recommendation If stem type=slender and structure=smooth If ?weeds= sedges Then and height=24 inch ?herbicide=Stomp 330EC and ?dose= 1000ml -50ml leafcolor=yellowgreenandleaf ?time = In drilling method structure=flat then sedge weed diagnosed TABLE III. WORMS SYMPTOMS AND INSECTICIDES RECOMMENDAITON Worms Symptoms Worms Diagnosed Insecticides Recommendation If symptoms =chewed holes and caterpillars=white and larvae= yellow then Cotton bollworm diagnosed. If ?insect=American BollWorm Then (?insecticides= Procalim 019EC ˄?dose= 200ml) ˄ (?insecticides= Larvin 80DF ˄?dose= 450 gm) ˄ (?insecticides= Tracer 240SC ˄ ?dose= 80ml) ˄ (?insecticides= Shogan1.8EC ˄ ?dose= 250ml) ˄ (?insecticides= Deltaphos 360EC ˄?dose= 700 ml) (?insecticides= Match 050EC ˄?dose= 800 ml) If color= light brown and rain in August= Yes and Rain in September= Yes rhen Pink BollWorm diagnosed. If ?insect=Pink BollWorm Then (?insecticides= Deltaphos 360EC ˄ ?dose= 600ml) ˄2.5EC ˄ ?dose= 400ml) ˄ (?insecticides= Talstar10 EC ˄ ?dose= 250ml) (?insecticides= Karate if rainfall= high and time >=July and time <=September and wings= four and streak =one white then Spotted Boll Worms Diagnosed. If ?insect=Spotted bollworms Then (?insecticides= Deltaphos 36EC ˄ ?dose= 600ml) ˄ (?insecticides= Karate 2.5EC ˄ ?dose= 400ml) ˄ (?insecticides= Match 50EC ˄ ?dose= 800ml) ˄ (?insecticides= Talstar 10EC ˄ ?dose= 250ml) ˄ (?insecticides= Sumi Alpha 110EC ˄ ?dose= 200ml) TABLE IV. IRRGATION SCHEDULING IN COTTON CROP BASED ON EXPERT SYSTEM Fact Regarding Scheduling Irrigation Scheduling 1 If crop=cotton and area=1.5ha and growing season > =August and growing season <= December and Soil Texture=99mm/m Then Then Irrigation method= Pressured piped surface method 2 If month=August and Time=beginning of August and pre sowing irrigation=0.6m Then Irrigation After=2 days and irrigation=crop establishment 3 If month=August and Time=8 August Then Irrigation After=2 days and irrigation=425m3 4 If month=August and Time=16 August Then Irrigation After=2 days and irrigation=425m3 5 If month=August and Time=24 August Then Irrigation After=2 days and irrigation=425m3 6 If month= September and Time=1 September Then Irrigation After=5days and irrigation=891m3 7 If month= September and Time=1 September Then Irrigation After=5days and irrigation=891m3 8 If month= September and Time=11 September Then 346 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 Irrigation After=5days and irrigation=891m3 9 If month= September and Time=22 September Then Irrigation After=5days and irrigation=891m3 10 If month= October and Time=2 October Then Irrigation After=5days and irrigation=75m3 11 If month= October and Time=11 October Then Irrigation After=5days and irrigation=75m3 12 If month= October and Time=21 October Then Irrigation After=5days and irrigation=75m3 13 If month= October and Time=31 October Then Irrigation After=5days and irrigation=75m3 14 If month= November and Time=13 November Then Irrigation After=5days and irrigation=75m3 15 If month= November and Time=26 November Then Irrigation After=5days and irrigation=75m3 TABLE V. COMPARISON OF PROPOSED IOT BASED EXPERT SYSTEM WITH OTHER ES A uthors C rops T echnology U sed D iagnosis R ecom m end ation T esting O utput 1.Sharm a et al.[32] Soybean W eb based expert system and Identifying Pests based on m orphology of insects Providing C ontrol M easures T ested by E xperts U ser input sym ptom s 2. Pradesh et al.[33] T om ato W eb based expert system , ID 3 algorithm , optim ization algorithm Provide inform ation about diseases, sym ptom s, chem ical control and Provide diagnosis about diseases, sym ptom s, chem ical control and prevention N il U ser input different sym ptom s 3. Q irui et al.[34] Fram ew ork for A griculture O ntology based expert system , C SS, PH P D iagnosis D iseases in agriculture N il N il N il 4.Fahad et al.[35] W heat W eb B ased E xpert System D iagnosis of diseases in w heat R ecom m enda tion for w heat diseases N il U ser input different sym ptom s 5. E drees et al.[36] W heat Sm all talk U sed for variety selection, land preparation, irrigation, planting, Provide recom m endat ion about variety selection, land preparation, irrigation T ested by agriculture engineers, researchers and extension officers. U ser input different sym ptom s. 6. A ndujar et al.[37] C ereals M icrosoft V isual B asic U sed for diagnosis and identification of diseases. It provides recom m endat ion for the diseases. System is validated and verified. U ser Input different sym ptom s. 7. K olhe et al.[9] O ilseeds W eb based and fuzzy logic based approach. U sed for diagnosis of diseases. It provides picture based diseases recom m endat ion. System is validated and verified. U ser input different sym ptom of diseases. 347 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 8. M archal et al.[38] O liveoil C om puter vision based technique for the diagnosis of diseases. U sed for the diagnosis of diseases based on im ages It classifies the diseases into different categories. System is tested and verified in chem ical lab. C am eras capture im age under the static environm ent in lab. 9. K aloudis et al.[39] Identification of insects in forests. E xpert system developed in E X SY S Professional, V er. 5.1.0, Predicting attack of insects, It diagnosis different insects categories. System is tested ad verified by m ultiple users. U ser input different characteristic and sym ptom s of insects. 10. M ansinghet all.[40] C offee E xpert system developed D iagnosis diseases and pests in coffee It diagnosis diseases provides appropriate solutions. System is tested and verified by m ultiple users. U ser Input different characteristic s and sym ptom s of diseases and pests. 11. A ndujar et al.[9] O live C rops A standalone expert system Identification of harm ful organism s like insects , diseases and w eeds It provides im ages after the identification of insects. System has verified and validated by technician and students. U ser m anually input different characteristics and sym ptom s of diseases. Fig. 4. Comparison of Temperarture Sensor and Humidity Sensors Data Fig 5 represents the relationship between soil moisture and leaf wetness data; these are directly proportional to each other. If the soil moisture increases then leaf wetness will increase automatically. Fig. 5. Comparison of Soil Moisture Sensor and Leaf Wetness Sensor Data VI. VALIDITY OF PROPOSED SOLUTION Our proposed solution is not so much costly as one may think due to the costly deployments of the sensors and actuators in the field. It is a fact that farmers invest heavily on electricity, fertilizers, insecticides, and pesticides. They are using these resources efficiently because they are unaware of the actual needs of the cotton crop. This investment can minimized by one-time investment on sensors. The sensors and cameras can monitor the cotton crop 24/7 and provide input for 348 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 proactive actions and optimal use of resources like water, fertilizer, insecticides and pesticides. Before deploying the IoT- based ES, we conduct the survey from users either they are willing to accept the IoT-based ES for agriculture. In our survey 100 different respondents like farmers, experts of agriculture participate in it. Farmers in Pakistan are illiterate, so we conduct surveys to ensure framework feasibility of its implementation, to evaluate the need of the system, after effects and their tradeoff while using IoT-based ES. VII. CONCLUSION In this paper, we have presented an ES for Cotton crop based on the concept of IoT. We tried to develop and initial frame for IoT-based agriculture. We developed an IoT-based ES. It's based ES consists of three modules; the first part consists of the deployment of WSN in the cotton fields. WSN has used for the monitoring of the cotton crop condition. The Waspmote agriculture sensor board has used for the monitoring of the cotton crop condition. It consists of temperature sensors, humidity sensors, leaf wetness sensors and soil sensors. In the concept of the IoT, the server should send the commands to the actuators of the fields, so the actuators of fields can take appropriate decisions. The sever should be intelligent enough to take decisions appropriately. For this purpose, we deploy an ES so that it can make decisions automatically. In Table 5 describes that different ES developed during a different era, but in this paper, we combine the IoT and ES. The sensors send the data to the server on the server side; we deployed the ES, which process and analyzes the sensor data. The data is fed to the ES that analyses it using the knowledge base and produces findings and recommendations. The ES consists of user interface, knowledge base and inference engine. On the server side, we deploy the concept of smart irrigation. Sensors monitor the soil moisture, leaf wetness, temperature, and humidity level in the environment and send the recommendation to the farmer about the irrigation in the cotton crop. In this paper, we ES for identification of different weeds, pests and different insects which attack on the cotton crop. These findings are sent to the farmer’s mobile phone for taking necessary actions in the field. We proposed an initial framework for the working of Smart Agriculture (SA). Before developing the concept of SA we conduct the survey and ask form user either the proposed system will be accepted by users or not. In this survey, we also ask from farmers and experts the flaws of the current system and whether they are satisfied the working of current agriculture or not. After that proposed system was evaluated by 100 different users like farmers and experts of Agri domain and 65 percent of respondents are satisfied with the working of SA and they are willing to accept the concept of SA. As we know farmers are illiterate in Pakistan so that we get 65 percent results. By deploying the IoT-based ES the productivity rate of the cash crops can be increased and problem of farmers also be reduced The proposed system was evaluated by 100 experts from the field and was found helpful for the farmers. VIII. FUTURE WORK In this paper, we present an initial framework for the diagnosis weeds, insects and different pests in cotton crop. We also deployed the concept of Smart Irrigation in cotton crop. For this purpose, we deploy the ES. In the future we will try to deploy the actuators in the fields and we enhance the functionality of server by deploying genetic algorithm, artificial neural network and digital image processing techniques on the server. We can diagnose the diseases in a better way if we deploy the cameras in the fields. REFERENCE [1] Morais, Raul, A. Valente, and C. Serôdio. "A wireless sensor network for smart irrigation and environmental monitoring: A position article." In 5th European federation for information technology in agriculture, food and environement and 3rd world congress on computers in agriculture and natural resources (EFITA/WCCA), pp.45-850. 2005. [2] Agrawal, Sarita, and Manik Lal Das. "Internet of Things—A paradigm shift of future Internet applications." In Engineering (NUiCONE), 2011 Nirma University International Conference on, pp.1-7. IEEE, 2011. [3] Hu, Xiangyu, and Songrong Qian. "IoT application system with crop growth models in facility agriculture." In 2011 6th International Conference on Computer Sciences and Convergence Information Technology ICCIT. 2011. [4] Li, Li, Hu Xiaoguang, Chen Ke, and He Ketai. "The applications of WiFi-based wireless sensor network in internet of things and smart grid." In Industrial Electronics and Applications ICIEA, 2011 6th IEEE Conference on, pp. 789-793. IEEE, 2011. [5] Tuli, Anupriya, Nitasha Hasteer, Mukesh Sharma, and Ankur Bansal. "Framework to leverage cloud for the modernization of the Indian agriculture system." In Electro/Information Technology (EIT), 2014 IEEE International Conference on, pp. 109-115. IEEE, 2014. [6] Liu, Yuxi, and Guohui Zhou. "Key technologies and applications of internet of things." In Intelligent Computation Technology and Automation (ICICTA), 2012 Fifth International Conference on, pp. 197- 200. IEEE, 2012. [7] Zhao, Ji-chun, Jun-feng Zhang, Yu Feng, and Jian-xin Guo. "The study and application of the IOT technology in agriculture." In Computer Science and Information Technology ICCSIT, 2010 3rd IEEE International Conference on, vol. 2, pp. 462-465. IEEE, 2010. [8] Jhuria, Manoj, Ajit Kumar, and Rushikesh Borse. "Image processing for smart farming: Detection of disease and fruit grading." In Image Information Processing (ICIIP), 2013 IEEE Second International Conference on, pp.21-526. IEEE, 2013. [9] González-Andújar, José Luis. "Expert system for pests, diseases and weeds identification in olive crops." Expert Systems with Applications 36, no. 2,pp 3278-3283 ,2009. [10] TongKe, Fan. "Smart Agriculture Based on Cloud Computing and IOT." Journal of Convergence Information Technology 8, no. 2 ,2013. [11] Chen, Xian-Yi, and Zhi-Gang Jin. "Research on key technology and applications for internet of things." Physics Procedia 33, pp. 561-566, 2011. [12] Talpur, Mir Sajjad Hussain, Murtaza Hussain Shaikh, and Hira Sajjad Talpur. "Relevance of Internet of Things in Animal Stocks Chain Management in Pakistan's Perspectives." International Journal of Information and Education Technology 2, no. 1 ,2012. [13] Evangelos A, Kosmatos, Tselikas Nikolaos D, and Boucouvalas Anthony C. "Integrating RFIDs and smart objects into a UnifiedInternet of Things architecture." Advances in Internet of Things 2011 ,2011. [14] Carvin, Denis, Philippe Owezarski, and Pascal Berthou. "Managing the upcoming ubiquitous computing." In Proceedings of the 8th International Conference on Network and Service Management, pp. 1276-280. International Federation for Information Processing, 2012. [15] Prasad, Rajkishore, Kumar Rajeev Ranjan, and A. K. Sinha. "AMRAPALIKA: An expert system for the diagnosis of pests, diseases, and disorders in Indian mango." Knowledge-Based Systems, Vol. 19, no. 1,pp. 9-21,2006. [16] Sarma, Shikhar Kr, Kh Robindro Singh, and Abhijeet Singh. "An Expert System for diagnosis of diseases in Rice Plant." International Journal of Artificial Intelligence, Vol. 1, no. 1 ,pp. 26-31,2010, , 349 | P a g e www.ijacsa.thesai.org (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 7, No. 9, 2016 [17] Balleda, Kaliuday, D. Satyanvesh, N. V. S. S. P. Sampath, K. T. N. Varma, and P. K. Baruah. "Agpest: An efficient rule-based expert system to prevent pest diseases of rice and wheat crops." In Intelligent Systems and Control ISCO, 2014 IEEE 8th International Conference on, pp. 262-268. IEEE, 2014. [18] Negid, N. K. "Expert System for Wheat Yields Protection in Egypt ESWYP." International Journal of Innovative Technology and Exploring Engineering IJITEE, ISSN ,pp. 2278-3075,2014. [19] Kaura, Ramanjeet, Salam Dina, and P. P. S. Pannub. "Expert System to Detect and Diagnose the Leaf Diseases of Cereals." Int J of Current Engineering and Technology , Vol. 3, no. 4 , 2013. [20] Kortuem, Gerd, Fahim Kawsar, Daniel Fitton, and Vasughi Sundramoorthy. "Smart objects as building blocks for the internet of things." Internet Computing, IEEE 14, no. 1 ,pp.~ 44-51,2010. [21] Coetzee, Louis, and Johan Eksteen. "The Internet of Things-promise for the future? An introduction." In IST-Africa Conference Proceedings, 2011, pp. 1-9. IEEE, 2011. [22] Ashton, Kevin. "That ‘internet of things’ thing." RFiD Journal, Vol. 22, no. 7 ,pp.97-114,2009. [23] Chase, Jim. "The evolution of the internet of things." Texas Instruments ,2013. [24] Travis, J. W., E. Rajotte, R. Bankert, K. D. Hickey, L. A. Hull, V. Eby, P. H. Heinemann, R. Crassweller, and J. McClure. "Penn State apple orchard consultant: design and function of the pest management module." In III International Symposium on Computer Modelling in Fruit Research and Orchard Management 313, pp. 209-214. 1992. [25] Salah, A., H. Hassan, K. Tawfik, I. Ibrahim, and H. Farahat. "CITEX: an expert system for citrus crop management." In Proceedings of the Second National Expert Systems and Development Workshop ESADW- 93. 1993. [26] El-Dessouki, A., S. Edrees, and S. El-Azhari. "CUPTEX: An integrated expert system for crop management of cucumber." ESADW-93, May ,1993. [27] Mohammed, M., K. EI-Arby, and A. Rafea. "LIMEX: an integrated expert system for lime crop management." In Second IFAC/IFIP Enr AGENG Workshop on AI in Agriculture, Netherlands. 1995. [28] Plant, R. E., F. G. Zalom, J. A. Young, and R. E. Rice. "CALEX/Peaches, an expert system for the diagnosis of peach and nectarine disorders." HortScience ,Vol. 24, no. 4 ,1989. [29] Ferguson, James J., Fedro S. Zazueta, and Juan I. Valiente. "Citpath: diagnostic and hypertext software for fungal diseases of citrus foliage and fruit." HortScience,Vol. 30, no. 4 ,pp.~ 899-899,1995. [30] Rani, Mercy Nesa, and Thangaswamy Rajesh. "Comparative Analysis on Software’s used in Expert System with Special Reference to Agriculture." MANAGEMENT, Vol.2, no. 8. [31] Saini, Harvinder S., Raj Kamal, and A. N. Sharma. "Web based fuzzy expert system for integrated pest management in soybean." International Journal of Information Technology ,Vol. 8, no. 1 ,pp.55-74, 2002. [32] Babu, MS Prasad. "A web based tomato crop expert information system based on artificial intelligence and machine learning algorithms." ,2010. [33] Qirui, Yin. "Kaas-based intelligent service model in agricultural expert system." In Consumer Electronics, Communications and Networks CECNet, 2012 2nd International Conference on, pp. 2678-2680. IEEE, 2012. [34] Khan, Fahad Shahbaz, Saad Razzaq, Kashif Irfan, Fahad Maqbool, Ahmad Farid, Inam Illahi, and T. Ul Amin. "Dr. Wheat: a Web-based expert system for diagnosis of diseases and pests in Pakistani wheat." In Proceedings of the World Congress on Engineering, Vol. 1, pp. 2-4. 2008. [35] Edrees, Soliman A., Ahmed Rafea, Ibrahim Fathy, and Mohamed Yahia. "NEPER: a multiple strategy wheat expert system." Computers and electronics in agriculture , Vol. 40, no. 1,pp. 27-43,2007. [36] Gonzalez-Andujar, J. L., C. Fernandez-Quintanilla, J. Izquierdo, and J. M. Urbano. "SIMCE: An expert system for seedling weed identification in cereals." Computers and electronics in agriculture,Vol. 54, no. 2 ,pp.115-123,2006. [37] Marchal, P. Cano, D. Martínez Gila, J. Gámez García, and J. Gómez Ortega. "Expert system based on computer vision to estimate the content of impurities in olive oil samples." Journal of Food Engineering , Vol.119, no. 2 ,pp. 220-228,2013. [38] Kaloudis, S., D. Anastopoulos, Constantine P. Yialouris, Nikos A. Lorentzos, and Alexander B. Sideridis. "Insect identification expert system for forest protection." Expert Systems with Applications, Vol. 28, no. 3, pp. 445-452,2005. [39] Mansingh, Gunjan, Han Reichgelt, and Kweku-Muata Osei Bryson. "CPEST: An expert system for the management of pests and diseases in the Jamaican coffee industry." Expert systems with Applications ,Vol.32, no. 1,pp.~ 184-192,2007. 350 | P a g e www.ijacsa.thesai.org I. Introduction II. Related Work A. Expert System in Agriculture III. Iot bASED Expert System A. Deployment of Sensors B. IoT- Based Expert System for Cotton Crop C. Server Send Recommendation to Farmer IV. Implementation and Evaluation A. Expert System for CLIPS V. Results VI. Validity of Proposed Solution VII. Conclusion VIII. Future Work 
work_rntucm3cifeqvf5jj5dyekbyte	----	This may be the author’s version of a work that was submitted/accepted for publication in the following source: Pitchforth, Jay & Mengersen, Kerrie (2013) A proposed validation framework for expert elicited Bayesian Networks. Expert Systems with Applications, 40(1), pp. 162-167. This file was downloaded from: https://eprints.qut.edu.au/52041/ c© Consult author(s) regarding copyright matters This work is covered by copyright. Unless the document is being made available under a Creative Commons Licence, you must assume that re-use is limited to personal use and that permission from the copyright owner must be obtained for all other uses. If the docu- ment is available under a Creative Commons License (or other specified license) then refer to the Licence for details of permitted re-use. It is a condition of access that users recog- nise and abide by the legal requirements associated with these rights. If you believe that this work infringes copyright please provide details by email to qut.copyright@qut.edu.au License: Creative Commons: Attribution-Noncommercial-No Derivative Works 2.5 Notice: Please note that this document may not be the Version of Record (i.e. published version) of the work. Author manuscript versions (as Sub- mitted for peer review or as Accepted for publication after peer review) can be identified by an absence of publisher branding and/or typeset appear- ance. If there is any doubt, please refer to the published source. https://doi.org/10.1016/j.eswa.2012.07.026 https://eprints.qut.edu.au/view/person/Pitchforth,_Jay.html https://eprints.qut.edu.au/view/person/Mengersen,_Kerrie.html https://eprints.qut.edu.au/52041/ https://doi.org/10.1016/j.eswa.2012.07.026 A proposed validation framework for expert elicited Bayesian Networks Jegar PitchforthI, Kerrie Mengersen Queensland University of Technology Abstract The popularity of Bayesian Network modelling of complex domains using ex- pert elicitation has raised questions of how one might validate such a model given that no objective dataset exists for the model. Past attempts at delin- eating a set of tests for establishing confidence in an entirely expert-elicited model have focused on single types of validity stemming from individual sources of uncertainty within the model. This paper seeks to extend the frameworks proposed by earlier researchers by drawing upon other disciplines where measuring latent variables is also an issue. We demonstrate that even in cases where no data exist at all there is a broad range of validity tests that can be used to establish confidence in the validity of a Bayesian Belief Network. Keywords: expert, validation, bayesian network, sensitivity IPhone: +61 403 961 878 Email address: jegar.pitchforth@qut.edu.au () 1Lvl 11, 126 Margaret St. , Brisbane 4000 Preprint submitted to Expert Systems with Applications May 26, 2012 1. Introduction1 Bayesian Networks (BNs) are an increasingly popular tool for modelling2 complex systems, particularly in the absence of easily accessed data. A BN3 describes the joint probability distribution of a network of factors using a4 Directed Acyclic Graph (Pearl, 1988). Factors that influence the likelihood5 of the outcome node being in any given state are represented as nodes on6 the graph. If the state of one model factor influences the state of another a7 directional arc is drawn between the two nodes representing these factors in8 the model. The combination of the nodes and their relationships is the BN9 structure. Each node in the graph can adopt any one of a finite set of states.10 For example, a factor representing magnitude could be classified as ’high’ or11 ’low’. While nodes do not strictly have to be discretised the practice is by far12 more commonly undertaken than not due to its computational convenience,13 and as such we do not discuss models that include non-discretised nodes in14 this paper. Finally, each node and relationship between nodes is quantified15 according to the likelihood of the node adopting a given state. In the case16 of input nodes these probabilities are seen as unconditional, whereas nodes17 internal to the model are dependent upon the states of the preceding nodes.18 The strength and direction of the relationship between model factors is de-19 fined in the conditional probability table associated with the child node.20 BNs are often created through a process of expert elicitation, in which ex-21 perts are asked to create a complex systems model by giving their opinions22 on the model structure, discretisation, and parameterisation. The validity23 of these models is generally tested through one of two procedures: by com-24 paring the model predictions to data available for the subject matter, or by25 2 asking the experts who contributed to the model creation to comment on its26 accuracy. This paper argues that these tests are limited in their ability to27 accurately test the validity of BNs, and presents a framework for more thor-28 ough validity testing. The work presented here stems from questions raised29 during the creation of a BN from expert elicitation to model the inbound30 passenger processing time at Australian airports. The network was elicited31 in collaboration with managerial and operational experts from Australian32 Customs and Border Protection Service (ACBPS) for the purpose of gaining33 more informative reporting of key performance indicators. In particular, the34 modelling of critical infrastructure underlined the importance of establishing35 that both experts and modellers have confidence in the final model produced.36 The paper is structured as follows. First, the concept of validation as it ap-37 plies to BNs is introduced in section 1.1. Second, the sources of confidence38 in BN validity are discussed, including network structure, discretisation, and39 parameterisation in section 1.2. Third, prior approaches to validating latent40 and expert elicited scales and models are introduced, drawing from psycho-41 metrics, system dynamics and other BN research in section sec:prevapproach.42 These principles are then applied to BNs with examples from the airport in-43 bound passenger processing model in section 3.44 1.1. Confidence in Bayesian Belief Network validity45 Model validity is often conceptualised as a simple test of a model’s fit46 with a set of data. However validity is a much broader construct: in essence,47 validity is the ability of a model to describe the system that it is intended48 to describe both in the output and in the mechanism by which that output49 is generated. In this paper we consider this broader definition of validity.50 3 The need for an explicit set of validity tests for BNs over and above com-51 parisons with data is clear. In current practice, where data are available on52 the phenomenon of interest, these data may be used to validate model pre-53 dictions. Several tests of this nature exist, such as a variety of Normal Max-54 imum Likelihood model selection criteria (Silander et al., 2009). However, a55 common reason for using BN models is a lack of available data. Examples56 of phenomena for which data are scarce include population characteristics57 in many developing countries (Shakoor et al., 1997), global epidemiological58 phenomena (Masoli et al., 2004), organised crime (Sobel and Osoba, 2009),59 conservation (Johnson, 2009) and biosecurity risk analysis (Barrett et al.,60 2010). In such cases, expert opinion can be elicited to create a Bayesian61 Belief Network (BBN). A common technique for validating BBNs based on62 expert opinion in the absence of data, is simply to ask the experts whether63 they agree with the model structure, discretisation, and parameterisation64 (see Korb and Nicholson (2010) for an excellent overview of BN applications65 and methods). This simple test is necessary, but not sufficient, to indepen-66 dently verify the validity of a complex model. Even where data are available,67 model fit is only a part of the model’s overall validity. These considerations68 lead to this paper’s proposition of a general validity framework for BNs.69 1.2. Sources of confidence in Bayesian Network validity70 In order to approach a validation framework for BNs, a short discussion of71 the background assumptions of this framework is required. First, we assume72 there exists a latent, unobservable ’true’ model (or set of acceptable ’true’73 models) for the phenomenon of interest against which the expert elicited74 model can be compared. Second, for the purposes of the validity framework75 4 presented in this paper, we consider a BN model to consist of four elements:76 model structure (section 1.2.1), node discretisation(section 1.2.2), and dis-77 crete state parameterisation(section 1.2.3). Each of these elements has been78 raised as a source of uncertainty in BN modelling. We provide a discussion of79 each element and consider the importance of validity within each model ele-80 ment, and within the model as a whole. The model elements are summarised81 in figure 1. Figure 1: Sources of confidence in Bayesian Network validity 82 5 1.2.1. Structure83 There are a number of questions when creating the structure of a BN. The84 first is the appropriate number of nodes to include which is a question of the85 modelling domain, level and scope. It is widely acknowledged that networks86 with a large number of nodes can easily become computationally intractable,87 as can networks with a large number of arcs between nodes (Koller and88 Pfeffer, 1997). The BN creator should ensure that the model is neither too89 simple nor too complex in its explanation of the system.90 1.2.2. Discretisation91 The discretisation process allows us to model systems probabilistically92 by taking continuous factors and assigning them intervals, ordinal states or93 categories, then modelling over the discrete domain. In more recent research,94 Uusitalo (2007) pointed out that such discretisation is a major disadvantage95 of BN modelling if it is necessary for the model, and Myllymaki et al. (2002)96 outlines how the process has the potential to destroy useful information.97 Given the information loss inherent in the discretisation process, ensuring98 that the states are a valid interpretation of the state space of the node is99 critical for a defensible network.100 1.2.3. Parameterisation101 Parameterisation refers to adding the values elicited from experts to the102 belief network (Woodberry et al., 2005). Much work has been conducted103 on controlling this stage of the process (Renooij, 2001), but little has been104 written about how to validate expert responses post-elicitation.105 6 1.2.4. Model Behaviour106 Finally, the behaviour of the model can be seen as the joint likelihood of107 the entire network as well as its sub-networks and relationships, hence con-108 fidence in model behaviour is founded upon the validity of the other three109 dimensions of the model. It is important to note that in the case of BNs,110 we are not only interested in whether the model can tell us what a system111 is doing under certain conditions, but also the factors and relationships that112 bring about this behaviour. This makes the problem of validating the model113 incredibly complex when attempted wholesale and justifies the need for par-114 titioning the dimensions of uncertainty for BNs. As such it is recommended115 that the structure, discretisation and parameterisation are tested for validity116 before any model behaviour tests can be run.117 2. Previous approaches to validity118 2.1. Psychometrics119 The discipline of psychometrics arose as a counterpart to the field of psy-120 chology, which at its foundation attempts to measure latent, unobserved,121 ’true’ variables such as intelligence. Due to this rich tradition, the founda-122 tions of measurement validation in psychometry are particularly solid, and123 serve as a useful base to begin discussion of a similar framework for BNs.124 Psychometrics first identified four types of validity (Cronbach and Meehl,125 1955); more recent research has reclassified and added dimensions of valid-126 ity to establish a full validation framework (Trochim, 2001). Based on the127 framework depicted in figure 2, a psychometric test can pass all these tests of128 validity to varying degrees, providing a multidimensional measure of how well129 7 a particular test measures a latent variable. In psychometric testing there130 are seven commonly tested dimensions of validity: nomological validity, face131 validity, content validity, concurrent validity, predictive validity, convergent132 validity, and discriminant validity. In psychometrics, before any other tests Figure 2: The psychometric validity testing framework adapted from Trochim (2001). 133 of validity can be undertaken, the nomological validity of the validity domain134 should be established. High nomological validity indicates that the measure-135 ment sits well within current academic thought on the subject. Face validity136 refers to the heuristic interpretation of a measure as a valid representation of137 the underlying psychometric construct. Content validity describes both the138 inclusion of all variables believed to be within a domain and the relevance of139 the factors included in the scale. Concurrent validity refers to the behaviour140 8 of a measurement scale; specifically, that the measure varies at the same point141 in time as another theoretically related measure taken on the same sample.142 Convergent validity refers to the criterion that scores on the measure to be143 validated (e.g. intelligence) should match scores on another, theoretically re-144 lated measure (e.g. school grades) in the same sample. Finally, discriminant145 validity refers to the criterion that scores on the measure to be validated146 should be different from scores on tests that measure constructs that are147 theoretically unrelated. While this is a useful paradigm upon which to base148 our exploration, the differences between judging the validity of a complex149 model and the validity of a score of a single construct are significant enough150 to necessitate further exploration into other approaches.151 The parameterisation process is the most similar to the psychometric dis-152 cipline, as the parameters can be treated as scores denoting a given belief153 about the behaviour of that node. Using this approach, we can use the ex-154 tensive literature on psychometrics and group behaviour to help validate the155 parameters we elicit from our experts.156 2.2. System Dynamics157 In his review of system dynamics validation tests Barlas (1996) describes a158 series of eight tests to validate system dynamics models; parameter confirma-159 tion, dimensional consistency, modified behaviour prediction, Turing tests,160 Qualitative Features analysis, extreme conditions testing, behaviour sensi-161 tivity tests and structure confirmation. Each of the tests can be classified in162 terms of the psychometric validity framework but can also be directly applied163 to specific sources of BN model uncertainty. For example, parameter confir-164 mation can be seen as a special test of concurrent validity applied specifically165 9 to model parameterisation. The tests introduced in the Barlas (1996) paper166 are described in more depth in the following section with specific reference167 to BN modelling.168 2.3. Machine Learning169 It is worth mentioning the significant research that has been conducted170 in the field of machine learning, particularly regarding content validity of the171 network structure. Machine learning researchers often use BNs and Bayesian172 Belief Networks to discover true networks using full datasets ( Heckerman173 et al. (1995) is a strong and widely cited example of this method). While174 this work is outside the scope of this paper, it is worth mentioning due to175 the minimalist approach used by machine learning researchers. In particular,176 the discipline is concerned with finding methods of excluding as many nodes177 and relationships from a BN as possible without losing explanatory power.178 2.4. Bayesian Network specific tests179 There are very few validity tests specific to BN modelling, but the few180 that are present are used commonly. Pollino et al. (2007) refers to the con-181 cepts of ’sensitivity to findings’ and ’sensitivity to parameters’ as methods of182 testing the predictive validity of expert-elicited networks. Other tests that183 have been introduced, such as d-separation analysis (Geiger et al., 1990) and184 causal independence-based tests (Cheng et al., 1997) are structural tests only,185 and are often used to establish internal consistency which is more elegantly186 defined as a reliability criterion.187 10 2.5. Problem Statement188 Unlike areas in which objective data are available, BNs built from expert189 elicitation cannot be validated using complete test datasets. As such, the190 concept of validity is not absolute but a question of additive strength. Often191 we cannot say whether a test has been conclusively passed or not, only take192 the weight of evidence over all the tests that have been applied. With this in193 mind we can begin to move toward a framework for validating all sources of194 uncertainty within the BN. While there are some tests introduced in previous195 research, these only test individual aspects of the network and can often only196 reflect the reliability rather than the validity of the model. For BN’s based197 either entirely upon expert elicitation, or a combination of data and expert198 elicitation, to be judged as valid assessments of the knowledge around a199 domain, a more comprehensive and robust framework of validity measures200 needs to be established.201 3. A validity testing framework for expert-elicited Bayesian Net-202 works203 The prior approaches to test and model validation are discussed and re-204 lated to BNs in the following section, with examples from the airports in-205 bound passenger processing network. When applying this validity testing206 framework to BNs, model structure, node discretisation, and overall model207 behaviour must be considered in addition to parameterisation. For this rea-208 son, in the following framework we consider the seven types of validity from209 psychometrics (including their special tests from system dynamics and BN210 modelling disciplines), and their application to the four sources of BN model211 11 uncertainty.212 213 3.1. Nomological validity214 In terms of an expert elicited BN, building nomological validity means215 establishing confidence that the model domain fits within a wider domain216 as established by the literature. For example, the passenger processing BN217 for ACBPS should sit within literature on airport terminals, way finding and218 security as well as other types of complex systems models and spatio-temporal219 model methods. If this test cannot be passed by the network, an argument220 must be made for why this model sits outside all current known research. This221 is very unusual, but may occur in fields such as advanced physics, where new222 information is shifting the entire paradigm of the discipline regularly. If this223 is the case, there may be an argument for a network having low nomological224 validity. Nomological validity is generally applied to the whole domain, but225 the nomological map serves as a reference for finding appropriate comparison226 models in later tests of specific sources of uncertainty. Given the power of227 nomological validity to place the research in a wider context, we begin the228 validation process with the questions:229 • Can we establish that the BN model fits within an appropriate context230 in the literature?231 • Which themes and ideas are nomologically adjacent to the BN model,232 and which are nomologically distant?233 12 3.2. Face validity234 Face validity is one of the most commonly used tests for expert-elicited235 BNs. For example, we can look at our passenger processing BN and check236 that baggage delivery time is part of the model and that it is related to the237 time spent picking up baggage to approximately the right level. However,238 despite the ease of establishing face validity it is considered the weakest form239 of validity within the psychometric framework. One of the primary dangers240 in establishing face validity is criterion contamination an issue that arises241 when the test dataset is the same as the validation set (Darkes et al., 1998).242 In our case, we might ask our set of experts whether they think the network243 looks the same as expected. Unsurprisingly, there are very few cases where244 the experts disagree with their own judgment. A more robust way of estab-245 lishing face validity would be to split the population of experts into test and246 validation groups, and ask the validation group only about the face validity of247 the network (Johnson et al., 2010). In cases where few experts are available,248 we can undertake a number of other strategies normally used for elicitation,249 such as using different experts for different parts of the BN, asking experts250 to assess their answers from a rival’s perspective, asking experts whether the251 model is applicable outside their domain and many others(Low Choy et al.,252 2009; James et al., 2010). In addition, often the entire model is tested at253 once (Korb and Nicholson, 2010). In order to learn as much as possible about254 the model through the validation process it is worthwhile to assess the face255 validity of the structure (including sub-networks), discretisation and param-256 eterisation independently. We therefore suggest the second set of questions257 in this validation stage:258 13 • Does the model structure (the number of nodes, node labels and arcs259 between them) look the same as the experts and/or literature predict?260 • Is each node of the network discretised into sets that reflect expert261 knowledge?262 • Are the parameters of each node similar to what the experts would263 expect?264 3.3. Content Validity265 To test for content validity of the structure we can check that all noted266 factors and relationships from the literature are included in the model, and267 discover which relationships are novel to the BN model. For example, in268 the passenger processing BN we could ensure that all the factors considered269 to important by the regulating bodies are included. To check the content270 validity of the discretisation of nodes within the model, we can ensure that271 all intervals implicated in the literature are included in the network. For272 example, if we were to discover that a node is generally classified at three273 levels in the literature, then a node with binary states would have low content274 validity. From a systems dynamics perspective, Barlas (1996) describes a275 dimensional consistency test which when applied to a BN paradigm could276 be defined as ensuring that all possible states of the node are included in277 the discrete states. For example, if a node were to include binary states278 of above twelve people and below twelve people, then the node would lack279 dimensional consistency as the possibility of there being exactly twelve people280 has been excluded. Finally, the content validity of the parameterisation can281 be checked through comparing expert elicited probabilities and relationships282 14 to analogous relationships in the literature. If parameters in the expert283 elicited model are significantly different, an argument should be made for284 the difference. To assess the content validity of a BN model, the following285 questions are suggested:286 • Does the model structure contain all and only the factors and relation-287 ships relevant to the model output?288 • Does each node of the network contain all and only the relevant states289 the node can possibly adopt?290 • Are the discrete states of the nodes dimensionally consistent?291 • Do the parameters of the input nodes and CPT reflect all the known292 possibilities from expert knowledge and domain literature?293 3.4. Concurrent Validity294 In the context of BNs, concurrent validity can refer to the possibility that295 a network or section of a network behaves identically to a section of another296 network, preferably driven by data. While this seems improbable, the na-297 ture of BN modelling seems to lend well to concurrent validity. For example,298 the passenger processing BN shares some sub networks and nodes with the299 customer satisfaction model for the same airport. In her introduction to Ob-300 ject Oriented Bayesian Networking, Koller and Pfeffer (1997) describes the301 technique as a way of capitalising on this high concurrent validity by build-302 ing networks from instances, or nodes representing sub-networks that can be303 easily transposed to other networks. This method allows large and highly304 complex BNs to be built without the researcher repeating modelling work305 15 performed by other researchers in the same domain. To test the concurrent306 validity of the structure of a BN, we can check other networks in related307 domains for sub-networks that are similar to sub-networks in the network.308 A model with high concurrent validity would have sub-networks in common309 with networks that are theoretically related, with the same number of nodes310 and relationships, with the relationships in the same direction. Similarly,311 when similar sub-networks from theoretically related networks are identified,312 we can judge the validity of the discretisation of nodes and their param-313 eterisation against the intervals of nodes and probabilities supplied in the314 comparison network. In the Barlas (1996) review of system dynamics tests,315 the application of concurrent validity criteria specifically to the parameters316 of the model factors is known as ’parameter confirmation’. Given these ap-317 proaches, the following questions are suggested as tests of a BN’s concurrent318 validity:319 • Does the model structure or sub-networks act identically to a network320 or sub network modelling a theoretically related construct?321 • In identical sub networks, are the included factors discretised in the322 same way as the comparison model?323 • Do the parameters of the input nodes and CPTs in networks of interest324 match the parameters of the sub network in the comparison model?325 3.5. Convergent Validity326 Convergent and discriminant validity are usually considered together, as327 they both reflect the relationship the BN has with other models. Convergent328 16 validity in BNs refers to how similar the model structure, discretisation,329 and parameterisation are to other models that are intended to describe a330 similar system. For example, we would expect our passenger processing BN331 to look similar to a network describing the processing of cargo at a seaport.332 The selection of comparison models is dependent upon the literature and333 knowledge of the domain at hand, but the original nomological map created334 in the first step of validation can be used as a reference for which sources may335 be of use. In particular, the comparison model for establishing convergent336 validity should be taken from an area as nomologically proximal as possible.337 In practise this could mean using a comparison model drawn from another338 complex systems discipline applied to the same domain, or alternatively using339 a BN drawn from a theoretically similar domain. As with the other types340 of validity, we can test the expert elicited BN regarding the convergent and341 discriminant validity of the structure, discretisation and parameterisation in342 isolation using the following questions:343 • How similar is the model structure to other models that are nomologi-344 cally proximal?345 • How similar is the discretisation of each node to the discretisation of346 nodes that are nomologically proximal independent of their network347 domain.348 • Are the parameters of nodes that have analogues in comparison models349 assigned similar conditional probabilities?350 17 3.6. Discriminant Validity351 The counterpart to convergent validity is discriminant validity, defined in352 this framework as the degree to which a model is different to models that353 should be describing a different system. For example, we would expect our354 passenger processing BN to look different to a model describing students’355 progression through school. As in the case of convergent validity, the com-356 parison model can be chosen using the nomological map as a reference guide357 for useful sources. The ideal method for establishing good discriminant valid-358 ity would be to select models from nomologically distal disciplines and work359 toward the construct of interest. Given that convergent validity has already360 been established, the ideal model would be one that is similar in most re-361 spects to the convergent comparison model, but dissimilar in all respects to362 the discriminant comparison model, which would be drawn from an area of363 research very close to the convergent validity comparison model.364 A system dynamics test of experts’ judgement of the discriminant validity of365 any source of uncertainty in a BN model is known as a Simulation Turing test366 (Schruben, 1980). The test requires many versions of the model to be shown367 to the researcher, only one of which is the expert-elicited model in every368 respect. Experts can be asked to choose the correct structure, discretisation369 or parameterisation from either a set of models of through binary choice ex-370 periments in which every model is compared to every other model. As in371 the case of face validity, the Turing test is ideally carried out on a separate372 set of experts to the set that originally created the model to avoid crite-373 rion contamination. The fewer differences in the final model chosen to the374 expert-elicited network, the higher the discriminant validity of that source375 18 of uncertainty. For this framework, the following questions are suggested as376 tests of the discriminant validity of the BN model:377 • How different is the model structure to other models that are nomo-378 logically distal?379 • How different is the discretisation of each node to the discretisation380 of nodes that are nomologically distal independent of their network381 domain?382 • Are the parameters of nodes in the comparison models that have oppo-383 sitional definitions to the node in question parameterised differently?384 • When presented with a range of plausible models, can experts choose385 the ’correct’ model or set of models?386 3.7. Predictive Validity387 In BNs, predictive validity can be considered to encompass both the388 model behaviour and the model output. This is the type of validity cov-389 ered by traditional model and data fitting techniques.390 When applying predictive validity tests within a complex systems and specif-391 ically a BN paradigm, the comparison model can be an alternative hypoth-392 esised model rather than a data-driven model. Such hypothesised models393 could be elicited using a number of techniques, such as case studies or for-394 mal walkthroughs (Barlas, 1996; Pollino et al., 2007). Luu et al. (2009) used395 case studies to formulate alternative hypothetical networks against which396 to compare the predictive validity of their BN model. While they did not397 specifically apply the tests presented in this paper, their work represents one398 19 of few papers to attempt to establish confidence in the predictive validity of399 an expert-elicited BN. Half of the special tests of system dynamics model400 validity presented by Barlas (1996) refer to the predictive validity of the401 model in that they test the model behaviour specifically. Of particular rele-402 vance to establishing confidence in the predictive validity of BN are behaviour403 sensitivity tests, Qualitative Features Analysis and the extreme conditions404 tests. When applied within a BN paradigm, the behaviour sensitivity test405 can be applied to the model structure and parameters by determining to406 which factors and relationships the model is sensitive, and comparing this to407 hypothetical models or alternative empirical models. The terms ’sensitivity408 to parameters’ and ’sensitivity to findings’ are used by Pollino et al. (2007) to409 describe the application of behaviour sensitivity tests to the parameters and410 model behaviour specifically, however it should be noted that this test can411 be just as easily applied to the structure and discretisation of nodes in the412 model as well. These tests are commonly used, and various versions of them413 can be executed using the GeNiE 2.0 (DSL, 2007), Hugin Expert (Andersen414 et al., 1989) or Netica (Norsys, 2007) software packages among others.415 Qualitative features analysis (Carson and Flood, 1990) is a case of predic-416 tive validity testing where behaviour in a hypothetical model is compared417 to the behaviour of individual pairs of nodes, sub-networks and the entire418 model. As in the case of predictive validity, the hypothetical models can be419 achieved through a number of formal strategies; however in this case, we are420 interested in the comparison of simulation output rather than comparison of421 model features directly. It is for this reason that model behaviour is outlined422 as the fourth source of model uncertainty. While this area is the product of423 20 the uncertainty of its component features, predictive validity requires that424 model behaviour be simulated from the model for tests to occur. For this425 reason, predictive validity should be the final type of validity to be tested.426 Finally, the extreme conditions test can be seen as a special case of qualita-427 tive features analysis, as it sets the hypothetical model to extreme conditions428 where the behaviour of the model is more predictable (Forrester and Senge,429 1980). For example, if the number of passengers is set to 0 then the model430 should reflect that there is a probability of 1 that 0 passengers are processed431 within the time range of interest. The direct extreme conditions test ex-432 amines the behaviour of individual pairs of nodes and sub-networks under433 such extreme conditions, while the indirect extreme conditions test examines434 the behaviour of the entire network against such hypotheses. The range of435 tests to establish confidence in the predictive validity of a model is notable436 considering the issue at hand that true objective data on the model are not437 available, and suggests that the lack of data available does not preclude pre-438 dictive validity testing, as hypothesis-driven models can be used in place of439 data-driven models. From examination of the various techniques associated440 with assessing predictive validity, we arrive at the following set of questions:441 • Is the model behaviour predictive of the behaviour of the system being442 modelled?443 • Once simulations have been run, are the output states of individual444 nodes predictive of aspects in the comparison models?445 • Is the model sensitive to any particular findings or parameters to which446 the system would also be sensitive?447 21 • Are there qualitative features of the model behaviour that can be ob-448 served in the system being modelled?449 • Does the model including its component relationships predict extreme450 model behaviour under extreme conditions?451 4. Conclusions and Recommendations452 In this paper we have outlined a broad range of conceptual tests that can453 be applied to validate BNs. These validity tests incorporate standard model-454 data fit comparisons, but expand the construct of validity to the broader455 definition of whether or not a model describes the system it is intended to456 describe, and produces output it is intended to produce. Many of these va-457 lidity tests can be used where no objective data exist.458 By combining existing research from BN validation with validation tests from459 psychometrics as well alternative complex systems disciplines, this paper in-460 troduces a starting point for discussing a framework for building confidence461 in the validity of BNs. The presented framework is not intended to be com-462 prehensive; instead, the aim is to establish that the validity of a BN can be463 tested, and should be tested, independent of the model fit to available data464 or expert confirmation. Disciplines such as psychometrics, with a history of465 measuring latent constructs, can provide a useful perspective on the problem.466 The framework presents a sequence of steps that can be followed to establish467 confidence in model validity, beginning with creating a nomological map of468 the literature surrounding the domain, then gradually building confidence in469 six types of model validity, using both general and specific tests.470 The application of this framework to the BN developed in conjunction with471 22 ACBPS will to our knowledge be a novel practical demonstration of such an472 approach to BN validation. The framework presented in this paper is in-473 tended to be domain-general, and there would be great value in establishing474 the versatility of the tests by applying them to complex models in other do-475 mains. Future work will extend to formalising and quantifying many of the476 tests in the context of BN modelling, and obtaining perspectives on model va-477 lidity from other disciplines that deal with unobserved variables and complex478 systems.479 5. References480 S.K. Andersen, K.G. Olesen, F.V. Jensen, and F. Jensen. Hugin - a shell481 for building bayesian belief universes for expert systems. In Proceedings482 of the Eleventh International Joint Conference on Artificial Intelligence,483 pages 1080–1085, United States of America, 1989. MIT press.484 Y. Barlas. Formal aspects of model validity and validation in system dynam-485 ics. System Dynamics Review, 12(3):183–210, 1996.486 S. Barrett, P. Whittle, K. Mengersen, and R. Stoklosa. Biosecurity threats:487 the design of surveillance systems, based on power and risk. Environmental488 and ecological statistics, 17:503–519, 2010.489 E.R. Carson and R.L. Flood. Model validation: philosophy, methodology490 and examples. Transactions of the Institute of Measurement and Control,491 12:178–185, 1990.492 J. Cheng, D.A. Bell, and W. Liu. An algorithm for bayesian belief network493 23 construction from data. In Proceedings of Conference on Artificial Intelli-494 gence and Statistics, pages 83–90, United States, 1997.495 L.J. Cronbach and P.E. Meehl. Construct validity in psychological tests.496 Psychological Bulletin, 52(4):281–302, 1955.497 J. Darkes, P.E. Greenbaum, and M.S. Goldman. Sensation seekingdisinhibi-498 tion and alcohol use: Exploring issues of criterion contamination. Psycho-499 logical Assessment, 10:71–76, 1998.500 DSL. Genie and smile, 2007. Bayesian Network Modelling software package501 and decision platform.502 J.W. Forrester and P.M. Senge. Tests for building confidence in system503 dynamics models. TIMS studies in the management sciences, 14:209–228,504 1980.505 D. Geiger, T. Verma, and J. Pearl. d-separation: From theorems to algo-506 rithms. In Proceedings of the Fifth Annual Conference on Uncertainty in507 Artificial Intelligence, UAI ’89, pages 139–148, The Netherlands, 1990.508 North-Holland Publishing Co.509 D. Heckerman, D. Geiger, and D.M. Chickering. Learning bayesian networks:510 The combination of knowledge and statistical data. Machine Learning, 20:511 197–243, 1995.512 A. James, S. Low Choy, and K. Mengersen. Elicitator: An expert elicitation513 tool for regression in ecology. Environmental modelling and Software, 25:514 129–145, 2010.515 24 S. Johnson. Integrated Bayesian network frameworks for modelling complex516 ecological issuesy. PhD thesis, Queensland University of Technology, Aus-517 tralia, 2009.518 S. Johnson, F. Harding, G. Hamilton, and K. Mengersen. An integrated519 bayesian network approach to lyngbya majuscula bloom initiation. Marine520 Environmental Research, 69:27–37, 2010.521 D. Koller and A. Pfeffer. Object-oriented bayesian networks. In Proceedings522 of the Thirteenth Annual Conference on Uncertainty in Artificial Intelli-523 gence, JMLR workshop and conference proceedings, pages 302–313, United524 States, 1997.525 K.B. Korb and A.E. Nicholson. Bayesian Artificial Intelligence, page 462 pp.526 CRC Press, United Kingdom, 2010.527 S. Low Choy, R. O’Leary, and K. Mengersen. Elicitation by design in ecol-528 ogy: using expert opinion to inform priors for bayesian statistical models.529 Ecology, 90:265–277, 2009.530 V. Luu, S. Kim, N. Tuan, and S. Ogunlana. Quantifying schedule risk in531 construction projects using bayesian belief networks. International Journal532 of Project Management, 27:39–50, 2009.533 M. Masoli, D. Fabian, S. Holt, and R. Beasley. The global burden of asthma:534 executive summary of the gina dissemination committee report. Allergy,535 59(5):469–478, 2004.536 P. Myllymaki, T. Silander, H. Tirri, and P. Uronen. B-course: A web-based537 25 tool for bayesian and causal data analysis. International Journal on Arti-538 ficial Intelligence, 11(3):369–387, 2002.539 Norsys. Netica, 2007. Proprietary Bayesian Network Modelling software540 package.541 J. Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausi-542 ble Inference. Morgan Kaufmann, 1988.543 C.A. Pollino, O. Woodberry, A. Nicholson, K. Korb, and B.T. Hart. Parame-544 terisation and evaluation of a bayesian network for use in an ecological risk545 assessment. Environmental Modelling and Software, 22:1140–1152, 2007.546 S. Renooij. Probability elicitation for belief networks: issues to consider. The547 Knowledge Engineering Review, 16(3):255–269, 2001.548 L.W. Schruben. Establishing the credibility of simulations. Simulation, 34:549 101–105, 1980.550 O. Shakoor, R.B. Taylor, and R.H. Behrens. Assessment of the incidence of551 substandard drugs in developing countries. Tropical Medicine and Inter-552 national Health, 2(9):839–845, 1997.553 T. Silander, T. Ross, and P. Myllymaki. Locally minimax optimal predic-554 tive modeling with bayesian networks. In Proceedings of the 12th Inter-555 national Conference on Artificial Intelligence and Statistics (AISTATS)556 2009, JMLR workshop and conference proceedings, pages 504–511, United557 States, 2009.558 26 R.S. Sobel and B.J. Osoba. Youth gangs as pseudo-governments implications559 for violent crime. Southern Economic Journal, 75(4):996–1018, 2009.560 W.M. Trochim. Research methods knowledge base, 2001. URL561 http://www.socialresearchmethods.net/kb/index.htm.562 L. Uusitalo. Advantages and challenges of bayesian networks in environmen-563 tal modelling. Ecological Modelling, 203(3), 2007.564 O. Woodberry, A. Nicholson, K. Korb, and C Pollino. Parameterising565 bayesian networks. In Geoffrey Webb and Xinghuo Yu, editors, AI 2004:566 Advances in Artificial Intelligence, volume 3339 of Lecture Notes in Com-567 puter Science, pages 711–745. Springer Berlin / Heidelberg, 2005.568 27 
work_rot6zrcqi5gj5ejyx4xttfge3m	----	wp-p1m-39.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-39.ebi.ac.uk no 220987049 Params is empty 220987049 exception Params is empty 2021/04/06-03:05:30 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220987049 (wp-p1m-39.ebi.ac.uk) Time: 2021/04/06 03:05:30 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_rqoiakpfrbfbvc3r66rojuqd7q	----	PII: 0888-613X(91)90010-J An Expert System Prototype for Inventory Capacity Planning: An Approximate Reasoning Approach I. B. Turksen Department of lndustrial Engineering, University of Toronto, Toronto, Ontario M. Berg Department of Statistics, Haifa University, Haifa, Israel A B S T R A C T A n approximate reasoning f r a m e w o r k is suggested f o r the d e v e l o p m e n t o f an expert system p r o t o t y p e to aid m a n a g e m e n t in planning inventory capacities. The development is considered to be a stage that comes after the analysis o f a stochastic m o d e l Such a m o d e l would p r o v i d e the requisite insight and knowledge about the inventory s y s t e m under specific assumptions. A s a consequence, the model builder(s) would act as expert(s). The restructuring process f r o m the stochastic m o d e l into the approximate reasoning f r a m e w o r k is described in a case s t u d y analysis f o r a M a r k o v i a n p r o d u c t i o n model. The stochastic m o d e l considers a relatively simplified p r o d u c t i o n process: one machine, constant production rate, a c o m p o u n d Poisson d e m a n d process f o r the p r o d u c t together with the reliability f e a t u r e comprising the machine f a i l u r e process and the ensuing repair action. In this context, the authors p r o p o s e an approximate reasoning f r a m e w o r k and describe (1) the identification o f the managerial decision-making rules, which usually contain uncertain (vague, ambiguous, f u z z y ) linguistic terms; and (2) the specification o f m e m b e r s h i p f u n c t i o n s that represent the meaning o f such linguistic terms within context-dependent domains o f concern. They then define a new Address correspondence to L B. Turksen, Department o f Industrial Engineering, University o f Toronto, Toronto, Ontario, Canada M5S 1A4. International Journal of Approximate Reasoning 1991; 5:223-250 © 1991 Elsevier Science Publishing Co., Inc. 655 Avenue of the Americas, New York, NY 10010 0888-613X/91/$3.50 2 2 3 224 I.B. Turksen and M. Berg universal logic incorporating these rules and functions and apply it to inventory capacity planning. Two case examples and a simulation experiment consisting o f 21 cases are summarized with a discussion o f results. KEYWORDS: e x p e r t s y s t e m , k n o w l e d g e acquisition, a p p r o x i m a t e reason- ing, i n v e n t o r y capacity p l a n n i n g , s i m u l a t i o n e x p e r i m e n t s 1. I N T R O D U C T I O N The purpose o f this paper is to suggest a possible approach to bridging the communication gap between stochastic model builders and the managers o f inventory capacity planners. We use an approximate reasoning framework based on fuzzy logic for the development o f an expert system prototype that can serve as an aid to managerial decision making in manufacturing. It is generally known that although stochastic models provide valuable insights into a system's behavior under specific assumptions and identify useful guidelines, more often than not they are not implemented, either because managers do not understand the basic assumptions o f the model, or because they are too complex, or because the precise information required to determine the values o f the model parameters cannot be obtained or are not available. Hence, such models are, b y default, inadequate to help managers to cope with the natural behavior o f many real-life systems. However, stochastic model builders are often experts who can interpret the results o f their models and can express their insightful expert knowledge about such system behavior in a natural language that usually contains vague, ambiguous, uncertain, fuzzy linguistic terms. Such linguistic terms provide (1) a flexible expression o f the system behavior subject to various uncertainties and imprecisions and at the same time (2) help model builders communicate their complex results to management in an appropriate natural language context. Such concerns lead us to suggest the development o f expert system proto- types via an approximate reasoning framework based on fuzzy logic but relying on insights obtained from stochastic models with the model builders acting as the experts. In order to explain in detail how such a development would and could take place, we first summarize briefly both the approximate reasoning framework and the stochastic model under consideration in this paper. Approximate Reasoning Informally, approximate reasoning is the process or processes by which a possible imprecise conclusion is deduced from a collection o f imprecise premises (Zadeh [1]). A number o f alternative approaches are available for reasoning in the design o f knowledge-based systems (Turksen [2], Turksen and Inventory Capacity Planning 225 Zhong [3]). Before we discuss the details o f an alternative approach, let us first review the structure of a possible knowledge-based system that may be used in prototype studies. (Clearly, the following is just a skeleton; other relevant aspects of the development of such a system are omitted in this paper.) 1. Factual knowledge (observed system state) ("facts"): X t is A* AND X 2 is A T A N D . . - AND X n is A* (1) for short: A* = A* AND A* A N D . . - AND A* The user o f an expert system inputs the factual knowledge, choosing the appropriate linguistic terms A T, A* . . . . A* that describe the observed states o f the system. Users may be given the option to state the meaning o f these linguistic terms by specifying membership values a n d / o r functions. Alternatively, they may be given the option to rely on the definitions provided by the domain experts. 2. Rule base (expert rules) ("rules"): I F X 1 is A l A N D X 2 is A 2 A N D . . . A N D X n is A n, T H E N Y is B (2) for short: A ~ B , where A = A l A N D A 2 A N D . . . A N D A n Domain experts (for this discussion, stochastic model builders) pro- vide such rules during the knowledge acquisition phase, where A~, A 2 . . . . . A n and B are the appropriate set o f linguistic terms specified by the domain experts. The domain experts must also provide the meaning o f these linguistic terms by specifying their membership values a n d / o r functions over the domain o f discourse. Rules are then encoded into the knowledge base by a knowledge engineer during the design and development phase. 3. Expert system advice ("response"): Y is (should be) B* (3) The expert system " r e s p o n s e " is provided to a user after the inference subsystem operates on the " r u l e s " and the given " f a c t s " in accordance with an inference scheme. In the current literature o f fuzzy sets, it may be observed that the meanings o f linguistic terms and their logical combinations are generally represented by " p o i n t - v a l u e d " membership assignments. However, it has been observed that 226 I.B. Turksen and M. Berg membership values obtained from domain experts and their logical combina- tions usually turn out to be " i n t e r v a l - v a l u e d " in most measurement experi- ments (Turksen [ 4 - 6 ] , Chameau and Santamarina [7]). In this paper we start out with a knowledge representation scheme based on point-valued membership functions obtained from experts (the stochastic model builders). That is, A 1, A 2 . . . . . A n, B; A T, A* . . . . . A* will all be repre- sented by point-valued membership functions. However, the representation o f the logical combination o f AND and the logical implication I F . . . T H E N will be computed with the disjunctive and conjunctive normal forms, DNF and CNF, respectively. These representations and computations will be further discussed in appropriate sections in the remainder o f the paper. FRAMEWORK The basic approximate reasoning framework used in this paper consists essentially of the following: (i) An interval-valued fuzzy set representation o f " r u l e s " and " f a c t s " (ii) A search for the " r u l e s " closest to the " f a c t s " based on a similarity measure (iii) An inference based on the implication o f the " r u l e s " found in (ii) and the " f a c t s " (iv) An advice based on the linguistic approximation o f the result(s) found in (iii) The details o f approximate reasoning methodology and its framework may be found in Turksen [2, 4 - 6 , 8, 9]. S t o c h a s t i c M o d e l Production/inventory/reliability models have been at the focus o f produc- tion modeling for a long time now. The rule o f inventory is to accommodate fluctuations from the demand side, and the fundamental problem is to relate the production process to the demand process so that, on the one hand, shortages are kept at a desirable low level but, on the other hand, no excessive inventory is built up. An important but frequently neglected element in such systems is the imperfection o f production systems. Posner and Berg [10] studied a model that incorporates reliability (or, indeed, unreliability) factors into the analysis and obtained closed-form analytical results under certain assumptions on the nature of the randomness in the production inventory system and the reliability factors. MODEL ASSUMPTIONS We consider a Markovian production model in which a single machine produces items at a constant production rate 1 (without loss o f generality) up to a level N , the inventory capacity. The production is halted whenever the inventory level reaches N and is resumed at the next demand epoch. Inventory Capacity Planning 22"/ The demand process is compound Poisson: Demands arrive according to a Poisson process at rate X, and order sizes are i.i.d, random variables exponen- tially distributed with a mean /x -~. There is no backlogging o f demand, so excess demand beyond the available inventory is lost. The operating time o f the machine is exponentially distributed with parameter 0. Thus 0 is the failure rate of the operating machine. The repair time o f the failed machine is exponentially distributed with mean a -~. ANALYTICAL RESULTS Let X ( t ) be the inventory level at time t, and set W ( t ) -= N - X ( t ) to be the slack storage capacity at time t. Clearly, W ( t ) [O, N ] . We introduce the variable W ( t ) because to deal with X ( t ) directly is less convenient than to deal with W ( t ) , but at the same time these two variables are equivalent to our understanding o f the behavior o f this system. The limiting density and distribution function of W = l i m t ~ o W ( t ) are f ( . ) and F ( - ) , respectively. Since the knowledge of W alone does not indicate whether or not the machine is in repair, a supplementary variable approach is implemented through the generalized technique o f " s y s t e m p o i n t " in level-crossing analysis (Brill and Posner [11, 12]). W should be partitioned into two parts, W o and W 1, where W 0 indicates the portion o f W while the machine is not under repair, and W~ is the portion of W while the machine is under repair. Correspondingly, both f ( . ) and F ( . ) should be partitioned into two parts, denoted by f o ( ' ) , f l ( ' ) and F0(.), F I ( . ) , respectively. Naturally we have the following two equations: f ( ' ) = f o ( ' ) + f , ( ' ) F ( ' ) = Fo(- ) -t- F1(" ) It is to be observed that aside from the partial densities with respect to the state of the machine, there is also a probability mass f o - Pr(W0 = 0) associated with a full inventory accompanied by production shutoff, and f ~ = P r ( W 1 = N ) associated with an empty inventory and no active produc- tion. The result of the mathematical analysis due to Posner and Berg [10] is summarized in the Appendix. The inventory capacity level N is the decision variable through which a desirable level of " t h e fraction o f satisfied customer d e m a n d " p can be achieved. For computational convenience we define a surrogate decision variable. N - N o v - (4) No 228 I . B . Turksen and M. Berg where N O is the existing inventory capacity. V is thus the fraction o f increase in the existing inventory capacity needed to achieve P0, the desirable level o f p. F o r specificity, we shall set Po = 0.95 and let N O -- 1.5. Setting N O --- 1.5 means an inventory capacity o f 1.5 " l o t s , " where a lot corresponds to a prespecified number o f items. This unit is also used in the definition o f # - t, the mean size o f demand. Another basic unit here is the unit o f time that is used in the definition o f the parameters X, a ~, and 0. The production rate is based on both units and their standardized form, which corresponds to produc- ing 1 lot per unit o f time. K N O W L E D G E A C Q U I S I T I O N The analytical result provides us with a thorough understanding o f this production/inventory/reliability system. It is, however, not very practical from the point o f view o f management. The exactness o f the system parameters is unnecessary to the managers, and furthermore it does not provide managers with insights into the operation o f this system, which is o f vital importance to the effective management o f such a complex system. Many analytical models bear the same language barriers. Even though these models clearly depict the characteristics o f a system for the well-trained model builders and their results could be effectively interpreted b y their builders, they are usually inappropriate for managerial implementation and use. An important, and at times critical, step in the design and development o f an expert system prototype is knowledge acquisition in the f o r m o f rules and their components. Preliminary Considerations In accordance with the analytical model, we have decided to use the fraction o f satisfied customer demand p as our performance criterion and the inventory capacity N as the decision variable, with other parameters tacitly assumed to be outside our control. Furthermore, in agreement with the approximate reasoning rule f o r m (2), we opt for the following rule structure: I F X i s A I A N D / ~ - 1 i s A2 A N D 0 i s A3 A N D a - i is Z 4 A N D p is high T H E N take action B (with respect to N or, equivalently, V ) The Ai, i = 1 . . . . . 4, and B stand for sets o f relevant linguistic descrip- tors. A basic set o f descriptors for each o f the A i in this case is low, Inventory Capacity Planning 229 medium, high (Turksen [2]). (If necessary, this categorization can be made finer by adding linguistic descriptors such as very low o r moderately high, as shown in the simulation experiments.) Thus, an e x a m p l e o f the rules w e aim to construct is: I f the d e m a n d arrival rate )x is low, and the m e a n d e m a n d size /x i is medium, and the failure rate o f the m a c h i n e 0 is low and its repair rate a - ~ is high, then in order to have a high fraction o f satisfied d e m a n d p, increase N moderately (or, equivalently, set a moderate value f o r V). The linguistic descriptors such as low, medium, and high are imprecise, but these are terms that a m a n a g e r not only understands but is also generally willing to give a m e a n i n g representation to by specifying a m e m b e r s h i p function ( Z y s n o [13], T u r k s e n [ 4 - 6 ] , N o r w i c h and T u r k s e n [14], Z i m m e r m a n and Z y s n o [15]). I n contrast, to use the exact model (i.e., the stochastic model) as is, the m a n a g e r w o u l d be required to substitute exact figures f o r the parameters - - a responsibility he o r she m a y be reluctant to assume because o f the ever-acute shortage o f data and the often encountered statistical inference difficulties, and so on. W e thus sacrifice some o f the exactness o f the original model in order to turn it into an implementable d e c i s i o n - m a k i n g tool w h e n e v e r w e are c o n f r o n t e d with either a lack o f data o r a limited a m o u n t o f data. Indeed, the exact analysis can be v i e w e d in this f r a m e w o r k as a building b l o c k in the construc- tion o f this m a n a g e r a l aid (the expert system) w h e r e the other building blocks relate to the qualitative/quantitative interpretations o f the linguistic descriptors as represented b y the c o r r e s p o n d i n g m e m b e r s h i p functions and inferencing methods, and so on. T h e detailed d e v e l o p m e n t o f this p r o c e d u r e f o r the p r o d u c t i o n / i n v e n t o r y / r e l i a b i l i t y case u n d e r consideration is described in the next section. INTERPRETATION B e f o r e w e turn to the actual construction o f the rules, it is useful to review s o m e o f o u r interpretations regarding the regions o f their applicability. T h e p a r a m e t e r N , the i n v e n t o r y capacity, is the sole decision variable b y means o f w h i c h w e want to achieve a desirable level Po for the fraction o f satisfied c u s t o m e r d e m a n d . While p = p(N) is an increasing function o f N , it is not at all clear, o r indeed true, that any desirable P0 can be achieved by m e r e l y increasing N o r V [because w e m a y very well have p(oo) < Po]- This is likely to h a p p e n w h e n the effective production rate is small c o m p a r e d to the d e m a n d rate )~/x-l, b e c a u s e in this case i n v e n t o r y does not build up fast e n o u g h to allow c h a n g e s on N relevant to our goal o f reaching Po- (By effective p r o d u c t i o n rate w e mean the theoretical production rate 1 minus the lost p r o d u c t i o n rate due to machine unreliability.) At the other extreme, w e have the cases w h e r e even a m u c h smaller level o f 230 I . B . Turksen and M. Berg N than the present N O is good enough to achieve P0. Generally speaking, a very low demand rate relative to the effective production rate makes a relatively small N sufficient to generate a high enough p. When the demand rate is close to the production rate 1, the objective V and the action on N are very sensitive to the failure rate, particularly i f the mean repair time o - ~ is high, because a greater or smaller failure rate can make all the difference between adequate supply and the occurrence o f a shortage. Note that this statement refers not only to the expectation o f the shortage but also to its variability. Thus, in the Poisson process, which in our theoretical model characterizes the failure process, the mean number o f failure events per unit o f time is proportional to the variance o f this random quantity, and therefore a high failure rate can significantly increase the range o f required changes on N , thereby detrimentally affecting the stability o f the expert s y s t e m ' s responses. This undesirable effect can be avoided b y limiting the set o f rules to low failure rates. Such a restriction is justified b y the fact that machines with medium or high failure rates are, with a high possibility, not likely to be retained by the management. Consequently, rules involving medium or high failure rates will be eliminated f r o m the system development. Linguistic Descriptors In order to design and develop an expert system that operates with the principles o f approximate reasoning and p e r f o r m s a task heuristically equiva- lent to the analytical model considered earlier, we require that the model builders not only provide us with the linguistic descriptors that identify the aggregate patterns o f behavior but also provide us with " m e a n i n g representa- t i o n " for these linguistic descriptors. As suggested in the previous section these linguistic terms were identified as L , low; M , medium; H , high for each o f the system (independent) parameters; that is, Demand rate k A 1 Demand size # - 1 A 2 Failure rate 0 A 3 Repair rate o - l A 4 Turning to the set o f actions B on N (or, alternatively, on V) that will yield the desirable performance level p, we consider the following four possible Inventory Capacity Planning 231 actions: Low (L) Increase N a bit (which includes the case o f no increase at all) or, equivalently, set a low value for V. Medium (M) Increase N moderately o r , equivalently, set a moderate value for V. High (H) Increase N a lot or, equivalently, set a high value for V. Very high ( V H ) Increase N quite a lot or, equivalently, set a very high value for N . (The term " i n c r e a s e " is interpreted in these actions in percentage terms.) In practice, the meaning representations o f these membership functions are specified either by an expert o f the production system under consideration or by its manager, in accordance with universally accepted forms and on the basis o f relevant considerations (as delineated above for the failure rate). For this exercise, we have determined the membership functions in a cooperative effort between the operations research (OR) specialist and the knowledge engineer. Due to the context-dependent nature o f the curves, these functions need to be adjusted and modified for a given production system. For all parameters, we have kept the convention o f scaling the p a r a m e t e r ' s ranges into the [0, 1] interval by normalizing each base variable with its m a x i m u m in the following manner: Normalized demand rate: D R = k / k m a x , (5a) Normalized demand size: DS = # - l//Zmalx, (5b) Normalized failure rate: FR = 0/0ma x (5c) Normalized repair rate: RR -- a - J/Om-alx (5d) Normalized decision variable: DV = V~ Vma x ( 5 e ) where the m a x i m u m values are set to )kma x 1.0 - l = , ~tma x = 1.0, 0ma x = 0 . 0 5 , Un~alx --- 5 . 0 , Vma x = 7 . 0 MEMBERSHIP FUNCTIONS The notion o f crossover point (Zadeh [16]) with the membership value o f 1/2 plays an important role in determining the membership functions. F o r each variable, we must first identify the points that are significant enough to divide the interval o f the universal set [0, 1] into corresponding linguistic subintervals o f low (L), medium (M), and high (H) or very high (VH), varying according to different system states, in order to 232 I.B. Turksen and M. Berg T a b l e 1. Subintervals of Linguistic Terms Variable Low Medium High Very high DR [0, 0.3] (0.3, 0.7] (0.7, 1.0] DS [0, 0.3] (0.3, 0.7] (0.7, 1.01 RR [0, 0.3] (0.3, 0.8] (0.8, 1.0] FR [0, 0.1] (0.1, 0.6] (0.6, 1.0] DV [0,0.15] (0.15,0.35] (0.35,0.54] (0.54, 1.0] acquire meaningful membership functions for every linguistic term. Again these subintervals are identified by experts as shown in Table 1. It should be noted that the separation point between regions is the crossover point o f the two curves representing the membershp functions o f the two corresponding linguistic terms, and the membership grade (MG) o f this point has the largest uncertainty (Kosko [17]) for both o f these fuzzy sets. For example, consider DR. The point separating the low and medium subintervals is 0.3: MGLow(0.3 ) = MGMedium(0.3 ) = 0.5 An element o f a fuzzy set with a membership grade greater than 0.5 is more likely to belong to this set than not, while a membership grade less than 0.5 indicates that an element is less likely to belong to this set. For example, the interval [0, 0.3] o f DR is intended to be regarded as a low-level demand rate rather than a medium-level demand rate by experts. Therefore 0.3 is chosen as the point separating the low-level and medium-level intervals o f DR, and accordingly a membership grade o f 0.5 is assigned for the fuzzy sets o f both low DR and medium DR. The membership curves are shown in Figures 1 -5 . A different case is the fuzzy set o f Very High for the service level, as shown in Figure 5. Even though the VH curve does not intersect the H curve except at the extreme right side o f the figure, it still has a membership grade o f 0.5 at MG l L M H 1 0.0 ' - ~ 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Figure 1. Four membership functions of VL, L, M, H for DR. Inventory Capacity Planning 233 MG L M H 0.5 0.1 0.2 0.3 0.4 0.5 0,6 0.7 0.8 0.9 1.0 Figure 2. Four membership functions of L, M, AAM, H for DS. MG ' L M H 1 0.0 c Vc~.~ 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 9 1.0 Figure 3. Four membership functions of L, M, LBH, H for RR. MG L M H 1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Figure 4. Three membership functions of L, M, H for FR. point 0.54, which separates intervals o f H and VH, following the same philosophy as we discussed above. Each o f these meaning representations o f the membership curves needs to be further justified by the expert in an appropriate manner. For example, note the steep slope near FR = 0. This is due to the sensitivity o f the notion o f ' " l o w failure rate" to even moderate (absolute) increases; thus, whereas a 0.05 failure rate may look to the manager to be low with a degree o f determination 234 I . B . Turksen and M. Berg MG M H 1 0 . 5 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 J' V/Vm~ Figure 5. Four membership functions of L, M, H, VH for DV. 0.95, a 0.1 failure rate is assigned a degree o f determination 0.5, indicating the highest level o f entropy in the assessment (Kosko [17]). However, with a parameter like the demand rate DR, such sensitivity is unlikely, because demand volatility is anticipated anyway because o f its dependence on human choices (as opposed to machine performance, for which a relatively high degree o f precision is expected). Similar considerations help determine the shapes o f all the membership curves (Figures 1 - 5 ) . Some figures include the membership curves for additional linguistic terms that will be needed for the case study examples discussed later (after we present the inference procedure). Structure o f t h e R u l e B a s e Let us now proceed to the actual construction o f the basic set o f rules o f the expert system. In principle, we should have 81 basic rules, since each o f the A i is assigned three linguistic terms. However, as mentioned earlier, rules involving " a b o v e - l o w " failure rates are impractical because the machine is not likely to be used. Hence, fixing the failure rate input state to " l o w " reduces the number o f input state vectors to the manageable size o f 27. Suppose a manager assesses all parameter values to be low. This will correspond to an input state vector I = ( L , L , L , L ) where the first, second, third, and fourth components o f I correspond to the normalized parameters DR, DS, FR, and RR, respectively. The question is what (minimal) action on N , or equivalently on V - - L , M, H, or V H - - i s required to achieve the desired performance level Po (0.95 in the illustration here). To answer this question we again require expertise. The expertise needed now is o f a m o r e general nature. Such expertise should be valid for an entire class o f production/inventory systems (with machine imperfection incorpo- Inventory Capacity Planning 235 T a b l e 2. Midpoints o f Linguistic Subintervals (Normalized) Variable Low Medium High DR 0.15 0.5 0.85 DS 0.15 0.5 0.85 RR 0.15 0.55 0.9 FR 0.05 rated), and it could be obtained, for example, f r o m production/inventory/reli- ability system managers. In our case, however, the source o f this expertise is again the O R experts who derived the mathematical results and interpreted them with the following simulated analysis. For each independent variable DR, DS, RR, FR, choose the midpoint o f the three linguistic subintervals (L, M, H) as a representation o f this linguistic term. These points are shown in Table 2. It should be noted that a decision to consider only the Low failure rates means that FR is set at 0.05 for L. First, the expert chooses these midpoints as a representation o f the linguistic terms in deciding the formation o f the left-hand side o f the rules. Next, the corresponding values o f these midpoints are determined by (5a)-(5d) and substituted into the analytical solution procedure that is shown in the Appendix to compute the corresponding inventory level and hence by Eq. (4) the service level. The service level in turn is converted to the corresponding subinterval by (5e), which identifies the value o f the base variable in Figure 5 and determines the linguistic term for the response. Thus, the linguistic term is determined for the right-hand side o f the rule. All 27 rules are determined in this manner; the results are presented in Table 3. R E P R E S E N T A T I O N A N D I N F E R E N C E Observe from Table 3 that each rule is a fuzzy relation, with the left-hand side o f the rule being an A N D combination among three variables, connected with the right-hand side o f the rule by an I F . . . T H E N (implication) relation. In order to carry out the operation o f A N D combination and I F . . . T H E N relation, a number o f finite support points for each linguistic term must be identified in Figures 1 - 5 . The values o f the support points for each term are a compromise between the efficiency o f computation and the accuracy o f the inference results. In this study we choose seven points for each t e r m as shown in Table 4. Recall f r o m the Introduction that in this paper the linguistic combinations are 236 I . B . Turksen and M. Berg T a b l e 3. T h e Rule Base for T h r e e I n d e p e n d e n t Variables a Rule Variable Independent Variables Decision No. DR DS RR DV 1 L L L L 2 L L M L 3 L L H L 4 L M L M 5 L M M M 6 L M H M 7 L H L M 8 L H M M 9 L H H M 10 M L L L 11 M L M L 12 M L H L 13 M M L M 14 M M M M 15 M M H M 16 M H L H 17 M H M H 18 M H H H 19 H L L L 20 H L M L 21 H L H L 22 H M L M 23 H M M M 24 H M H M 25 H H L VH 26 H H M VH 27 H H H VH aRecall that the failure rate FR is set to L only. based on the d i s j u n c t i v e a n d c o n j u n c t i v e n o r m a l forms, D N F a n d C N F , respectively, which are e x t e n s i o n s o f the c a n o n i c a l forms in Boolean logic. A N D C o m p o s i t i o n T h e D N F a n d C N F for A N D c o m b i n a t i o n are O N F ( A A N D B ) = A I"1 B (6) C N F ( A A N D B ) = ( A U B ) CI ( A U B c) ¢3 (A ~UB) (7) Inventory Capacity Planning 237 Table 4. Typical Linguistic Descriptors and Their Membership Values for Each Fuzzy Set DR: VL = (1.0/0.0, 0.4/0.25, 0.32/0.3, 0.0/0.5, 0.0/0.7, 0.0/0.8, 0.0/1.0) SL = ~1.0/0.0, 0.57/0.25, 0.25/0.3, 0.0O64/0.5, 0 . 0 / 0 . 7 , 0.0/0.8, 0.0/1.0) LBL = ~1.0/0.0, 0.87/0.25, 0.71/0.3, 0.28/0.5, 0 . 0 / 0 . 7 , 0 . 0 / 0 . 8 , 0.0/1.0) L = (1.0/0.0, 0.76/0.25, 0.5/0.3, 0.08/0.5, 0.0/0.7, 0.0/0.8, 0.0/1.0) M = (0.0/0.0, 0.24/0.25, 0.5/0.3, 1.0/0.5, 0.5/0.7, 0.2/0.8, 0.0/1.0) QBM = (0.0/0.0, 0.057/0.25, 0.25/0.3, 1.0/0.5, 0.25/0.7, 0.04/0.8, 0.0/1.0) H = ~0.0/0.0, 0 . 0 / 0 . 2 5 , 0 . 0 / 0 . 3 , 0.08/0.5, 0 . 5 / 0 . 7 , 0 . 8 / 0 . 8 , 1.0/1.0) DS: L = (1.0/0.0, 0.76/0.25, 0.5/0.3, 0.08/0.5, 0.0/0.7, 0./0.8, 0.0/1.0) M = (0.0/0.0, 0.24/0.25, 0.5/0.3, 1.0/0.5, 0.5/0.7, 0.2/0.8, 0.0/1.0) QBM = (0.0/0.0, 0.1/0.25, 0.24/0.3, 1.0/0.5, 0.3/0.7, 0.15/0.8, 0.0/1.0) H = (0.0/0.0, 0.0/0.25, 0.0/0.3, 0.08/0.5, 0.5/0.7, 0.8/0.8, 1.0/1.0) VH = (0.0/0.0, 0.0/0.25, 0.0/0.3, 0.0/0.5, 0.25/0.7, 0.64/0.8, 1.0/1.0) RR: VL = (1.0/0.0, 0.64/0.25, 0.25/0.3, 0.0/0.55, 0.0/0.8, 0.0/0.86, 0.0/1.0) L = (1.0/0.0, 0.8/0.25, 0.5/0.3, 0.01/0.55, 0.0/0.8, 0.0/0.86, 0.0/1.0) NM = (1.0/0.0, 0.73/0.25, 0.5/0.3, 0.0/0.55, 0.5/0.8, 0.74/0.86, 1.0/1.0) M = (0.0/0.0, 0.27/0.25, 0.5/0.3, 1.0/0.55, 0.5/0.8, 0.26/0.86, 0.0/1.0) H = (0.0/0.0, 0.0/0.25, 0.0/0.3, 0.01/0.55, 0.5/0.8, 0.8/0.86, 1.0/1.0) LBH = (0.0/0.0, 0.0/0.25, 0.0/0.3, 0.17/0.55, 0.8/0.8, 1.0/0.86, 1.0/1.0) VH = (0.0/0.0, 0.0/0.25, 0.0/0.3, 0.0/0.55, 0.25/0.8, 0.64/0.86, 1.0/1.0) DV: L = (1.0/0.0, 0.5/0.15, 0.2/0.25, 0.09/0.35, 0.0/0.5, 0.0/0.54, 0.0/1.0) M = (0.0/0.0, 0.5/0.15, 1.0/0.25, 0.5/0.35, 0.13/0.5, 0.04/0.54, 0.0/1.0) H = (0.0/0.0, 0.06/0.15, 0.18/0.25, 0.5/0.35, 0.95/0.5, 1.0/0.54, 1.0/1.0) VH = (0.0/0.0, 0.0/0.15, 0.02/0.25, 0.1/0.35, 0.32/0.5, 0.5/0.54, 1.0/1.0) where f3, U , and superscript c are used in the set notation to correspond to the intersection, union, and complementation operators, respectively. All the fuzzy sets we have defined so far are point-valued. However, by applying Eqs. (6) and (7) for AND combination, interval-valued fuzzy sets are generated, with CNF(.) and DNF(.) being the upper and lower bounds, respectively. This is certainly not a surprise, as it is argued in Turksen [4-6] that most of the linguistic combinations would be interval-valued when impre- cise knowledge is extracted from experts. That means that our four indepen- dent variables and one decision variable could most probably have been interval-valued fuzzy sets. For ease of computation we adopt a point-valued approach and choose the midpoint of an interval for fuzzy relations whenever such an interval is given or is generated by DNF and CNF operations. Let us choose rule 6 from Table 3 as an example to explain these calcula- 238 I . B . Turksen and M. Berg tions. The left-hand side o f this rule is RLM H = D R L A N D DS M A N D RR H (8) A finite support membership value is chosen f r o m each o f these three fuzzy sets, say a i ~ D R L , b j E DS M , c k E R R H, 1 _< i, j , k _< 7. Let us also choose Max and Min operations, denoted by v and A, to correspond to the U and f) in the set notation o f C N F and D N F in Eqs. (6) and (7) and throughout this paper. This choice is needed because there are a large collection o f operators that correspond to U and O (Turksen [ 4 - 6 ] ) . Hence the membership grades (MGs) for the first two variables are com- puted as M G ( D N F ( D R L A N D DSM) ) = a i A b j (9) M G ( C N F ( D R L A N D D S M ) ) = ( a i v b j ) A ( a i V b f ) A ( a C v b j ) (10) where the complementation is chosen to be the pseudocomplement: c = 1 - a i, b f = 1 - b j , c ~ = 1 - cg a i The midpoint G j o f the interval-valued fuzzy set R L M -~ D R L A N D DS M is computed as rii = ( 1 / 2 ) [ M G ( D N F ( . ) ) + M G ( C N F ( . ) ) ] (11) Next rij is used to compute the C N F and D N F o f A N D with c k similarly determining rijk, which is the midpoint o f the interval-valued fuzzy set o f RLM H in Eq. (8). RLM is a 7 × 7 matrix, and RLM n is a 7 × 7 × 7 matrix. Altogether we have 27 such matrices, since we have 27 rules in our rule base. An c~-cut for each matrix or each rule needs to be computed by a = V ( r i j k A r / ~ , ) , 1 _< i, j , k _< 7, (12) to satisfy the sufficiency condition required for the generalized modus ponens (Turksen [2]). Thus, we have 27 ot's in all, that is, a , , 1 _< n < 27, which are then used to truncate the decision space as discussed next. I F . . . THEN Composition The 27 three-dimensional matrices are then combined with the right-hand side o f each rule, the decision space, b y I F . . . T H E N composition to construct the fuzzy relation o f each rule in Table 3. But we must first truncate the decision space b y c~-cuts, where the o~'s are given by Eq. (12). For example, Inventory Capacity Planning 2 3 9 suppose we have t~ 6 = 0.5. Rule 6 shows that DV should be M, with the following finite support: DV M : ( 0 . 5 , 1 , 0 . 5 , 0 . 1 3 , 0 . 0 4 ) After truncation, we get D v(T) ( 0 . 5 , 1 , 0 . 5 ) v M Those elements that are less than c¢ 6 are discarded, because they are insignifi- cant in carrying the information through this inference procedure (Turksen [2-61). Now we are ready to do DNF and CNF evaluation o f the I F . . . T H E N relation between R and DV. The DNF and CNF expressions o f I F . . . T H E N composition are D N F ( A ~ B ) = ( A A B ) U ( A C A B ) tO ( A c A B ¢) (13) C N F ( A --} B ) = A c O B (14) In the membership domain, again with the application o f Max and Min operators, we get rij~, = (1/2)[MG(DNF(RLM~I ~ D V ~ ) ) ) + MG(CNF(RLM n ~ D V ~ )))] (15) where riik t is the midpoint o f the interval-valued fuzzy sets in matrix RLMHM , that is, R LMHM = R LMH ~ ~--Ml')~ur ( T ) Equation (15) will reduce interval-valued membership grades to point-valued ones. The 27 matrices generated in this manner are represented in four-dimen- sional matrices with membership values. So far we have constructed a set o f rules and the database that is the set o f these matrices, which are required for the inferences in approximate reasoning. I n f e r e n c e P r o c e d u r e A number of alternative procedures are available for the approximate reasoning approach (Turksen [2, 9], Turksen and Zhong [3]). In this paper, we discuss Zadeh's composition rule o f inference [16]. In addition, from among the many approximate reasoning modes, we choose the modus ponens known as the generalized modus ponens (GMP) in our discussion. For GMP, compositional inference is written as R* o ( R ~ DV) = DV* (16) where o is the compositional rule o f inference; R* is an AND combination of I. B. Turksen and M. Berg observed states o f the system, DR*, DS*, RR*; and R ~ DV is a rule to be selected from the rule base with a suitable distance or similarity measure. DV* is the expert system advice (response) to be provided to a user for the combined observed system state R*. Such expert system responses are deter- mined in the following manner. Suppose a user inputs DR*, DS*, and RR* (the observed states); then the inference engine first computes R'LOWER = D N F ( D R * AND DS* AND RR*) R'UPPER ~--- C N F ( D R * AND DS* AND RR*) M G ( R * ) = (1/2)[MG(R~.oWER) + MG(R~ppER) ] Next the closeness o f R* is checked against R , the left-hand side o f each o f the 27 rules. Among the many distance measures (Zwick et al. [18]), we choose the Hamming distance as the closeness measure for this paper, which is d H ( R - R*) = Z rijk-- r*'kl (17) ijk The rule whose left-hand side R has the minimum distance among the 27 rules to the observed system state R* is chosen for G M P computations. The compositional rule o f inference is applied for the rule so chosen. For example, suppose the rule so chosen has left-hand side and right-hand side denoted by R * and DVd', respectively; then R* o ( R ~ --} DV*) = DV* (18) By substituting DNF and CNF operations o f R* ~ DV* into Eq. (18), we can obtain the DNF and CNF o f DV* and the midpoint o f this interval as M G ( D V * ) = ( 1 / 2 ) [ M G ( D N F ( D V * ) ) + MG(CNF(DV*)I (19) Since DVg' in Eq. (18) is already truncated in the implication composition by a0 given by Eq. (12), DV* has only those elements corresponding to the truncated DV*. Finally, DV* is compared to all the linguistic terms in the decision space, that is, D e { L , M , H , VH} to find the closest linguistic term based on the distance measure, in this study the Hamming distance d n given by Eq. (17), that is, d H ( D V * , D V ) = ~ I M G ( D V * ) - M G ( D V ) I Thus, the closest linguistic term is taken as the linguistic approximation to D V * . The output o f the fuzzy inference engine is this linguistic term so chosen. Inventory Capacity Planning 241 Please note that when we make the distance comparison between DV* and the linguistic terms in decision space, D V d' is truncated by the a0-cut [Eq. (12)]. Thus, we should truncate the linguistic terms in the decision space accordingly in order to calculate the H a m m i n g distance properly. To indicate the closeness between D V * and D V we m a y choose another measure, besides the distance measure, which is somewhat more general in the sense that it is in the interval [0, l] and is independent o f any particular situation. This measure is called the s i m i l a r i t y m e a s u r e (Turksen and Zhong [3]) and is defined as 1 S - l + d where d is a distance measure. In our c a s e , d H is the H a m m i n g distance; therefore, 1 S - (20) 1 + d n When d n = O, S = 1, indicating that the two linguistic terms are exactly the same; and when d H = oo, S = O, meaning that the two terms are very far apart. It should be noted that d n and S are equivalent measures o f closeness. S i m u l a t i o n E x p e r i m e n t The proposed approximate reasoning approach was p r o g r a m m e d in LISP and was run on an Apollo workstation in a U N I X environment for two simulation experiments with 21 hypothetical case data shown in Table 6. The membership values o f the linguistic variables are shown in Table 4. The first experiment was run with the application o f max-min operators, the second with the application o f bold union/intersection operators. The results o f these experiments are shown in Tables 7 and 8. Before we discuss the outcome o f these experiments, let us illustrate the inference procedure described in the previous subsection with two examples. EXAMPLE 1 I f the state o f the system we have observed matches exactly to one o f the left-hand sides o f the 27 r u l e s - - f o r example; DR* is L, DS* is M, and RR* is M, that is, rule 5 in Table 3 a n d / o r case 11 in Table 6 - - t h e n R * = R L M M and d H ( R * , R L M M ) ----- 0. I f we input R * into the fuzzy inference engine, we get DV* = DVd', which is M. Thus, M is the linguistic t e r m we should choose according to this observed state o f the system. This process is illustrated in Table 5. The output is M with a similarity measure o f 1. The result is what we should have expected, because the observed state is exactly the left-hand side o f rule 5. The 242 I . B . Turksen and M. Berg T a b l e 5. I n f e r e n c e P r o c e d u r e Observed states DR* = (1.0, 0.76, 0.50, 0.08, 0.0, 0.0, 0.0) DS* = (0.0, 0.24, 0.50, 1.0, 0.50, 0.20, 0.0) RR* = (0.0, 0.27, 0.50, 1.0, 0.50, 0.26, 0.0) R* = DR* AND DS* AND RR* min[ R* - R I = 0 identifies R~ o f rule 5, with a 5 = 0.5. Hence, DV~' is M. DV* = R* o(R~ --* DV~) DV* = (0.5, 1., 0.5) and dH(DV ~, DV*) = 0 1 S = - - = I 1 + d H fuzzy inference engine does certainly p r o d u c e , as it should, an output that is the right-hand side o f rule 5. This e x a m p l e illustrates the fact that an expert system based o n o u r inference p r o c e d u r e p r o d u c e s the expected result w h e n there is an exact match between the o b s e r v e d s y s t e m state and the left-hand side o f a rule in the rule base. H o w e v e r , the real p o w e r o f o u r approximate reasoning has a bit m o r e intelligence built into it in the following sense. Suppose there is no exact match between the o b s e r v e d s y s t e m state and the left-hand side o f a n y rule in the rule base. T h e n the question is, W h a t can o u r a p p r o x i m a t e reasoning p r o c e d u r e p r o v i d e f o r the user? W e s h o w in the next e x a m p l e that even w h e n there is not an exact m a t c h , o u r approximate reasoning has the p o w e r to generate appropri- ate advice f o r the users. EXAMPLE 2 I f the state o f the s y s t e m w e have o b s e r v e d is an arbitrary o n e - - s a y , D R * is very low, DS* is quite a bit medium, and R R * is a little bit high--then the inference p r o c e d u r e is as follows. I f they are not available in the s y s t e m already, the expert (the model builder) should give us m e a n i n g representations o f all the allowable linguistic terms Inventory Capacity Planning 243 T a b l e 6. Simulation Inputs: O b s e r v e d S y s t e m States Case Linguistic Descriptors of Input Parameters No. DR* DS* RR* 1 L M H 2 VL QBM LBH 3 VL M VH 4 VL QBM VL 5 SL QBM VL 6 SL QBM VH 7 VL M NM 8 L L L 9 M M M 10 H H H 11 L M M 12 VL L VH 13 VL L VL 14 H L VH 15 H L LBH 16 LBL L H 17 M H VL 18 M H LBH 19 M VH VL 20 QBM H VH 21 QBM H VL such as v e r y low (VL), quite a bit m e d i u m ( Q B M ) , a little bit high (LBH), specifying their m e m b e r s h i p functions. ( F o r the case study, o u r expert identi- fied these m e m b e r s h i p functions as s h o w n in Figures 1 - 3 and as defined in Table 4.) Thus, D R * is V L , D S * is Q B M , R R * is L B H , that is, case 2 in Table 6. O b s e r v e that this does not m a t c h the left hand-side o f any o f the rules in Table 3. First R * = D R * A N D D S * A N D R R * is c o m p u t e d b y D N F and C N F expressions o f A N D combination. T h e n the nearest R , denoted b y R * , is c h o s e n f r o m a m o n g the 27 left-hand sides o f the rules in the rule base. I n this case, rule 6 is selected; hence, D V d' is M. W i t h the G M P and an c~-cut o f 0.5, w e get D V * = R * o ( R * --* DV~') T h e result is D V * = ( 0 . 5 / 0 . 1 5 , 1 . 0 / 0 . 2 5 , 0 . 5 / 0 . 3 5 ) . T h e c o m p a r i s o n o f distance b e t w e e n D V * and other linguistic terms in 244 I . B . Turksen and M. Berg d e c i s i o n s p a c e D c a n b e s u m m a r i z e d as f o l l o w s : D e c i s i o n S p a c e D H a m m i n g D i s t a n c e M e a s u r e [DVd' - D V * [ S i m i l a r i t y L 1.21 0 . 4 5 M 0 . 0 1.0 H 1.26 0 . 4 4 V H 3 . 1 4 0 . 2 4 T h e r e f o r e , the l i n g u i s t i c t e r m M is the r e s p o n s e o f the e x p e r t s y s t e m for the o b s e r v e d s y s t e m state. T h a t is, w h e n the d e m a n d rate is v e r y l o w , the d e m a n d s i z e is q u i t e a b i t m e d i u m , a n d the r e p a i r r a t e is a little bit h i g h , w e s h o u l d e x p e c t the s e r v i c e l e v e l to b e m e d i u m . SIMULATION EXPERIMENTS L e t us n o w l o o k at the t w o s i m u l a t i o n e x p e r i - T a b l e 7. S i m u l a t i o n E x p e r i m e n t s R e s u l t s - - ( a ) S i m u l a t i o n w i t h M a x - M i n O p e r a t o r s ; (b) S i m u l a t i o n w i t h B o l d U n i o n / I n t e r s e c t i o n O p e r a t o r s Selected Rule Linguistic Descriptor Number of System Response Case No. (a) Co) (a) (b) 1 6 6 M M 2 6 6 M M 3 6 6 M M 4 4 4 M M 5 4 4 M M 6 6 6 M M 7 5 6 M M 8 1 1 L L 9 14 14 M M 10 27 27 VH VH 11 5 5 M M 12 3 3 L L 13 1 1 L L 14 12 12 L L 15 12 12 L L 16 3 3 L L 17 16 16 H H 18 18 18 H H 19 16 16 H H 20 18 18 H H 21 16 16 H H Inventory Capacity Planning 245 T a b l e 8. Simulation E x p e r i m e n t s - - C o m p a r i s o n o f DV* for Max-Min and Bold Operator with or-Cut o f 0.5 Case No. DV*, Max-Min DV*, Bold 2 (0.50, 1.00, 0.50) (0.65, 1.00, 0.65) 3 (0.50, 1.00, 0.50) (0.50, 1.00, 0.50) 4 (0.50, 1.00, 0.50) (0.506, 1.00, 0.506) 5 (0.50, 1.00, 0.50) (0.646, 1.00, 0.646) 6 (0.50, 1.00, 0.50) (0.646, 1.00, 0.646) 7 (0.75, 1.00, 0.75) (1.00, 1.00, 1.00) 12 (1.00, 0.50, 0.00) (1.00, 1.00, 0.00) 13 (1.00, 0.50, 0.00) (1.00, 1.00, 0.00) 14 (1.00, 0.75, 0.00) (1.00, 1.00, 0.00) 15 (1.00, 0.75, 0.00) (1.00, 1.00, 0.00) 16 (1.00, 0.50, 0.00) (1.00, 1.00, 0.00) 17 (0.95, 1.00, 1.00) (1.00, 1.00, 1.00) 18 (0.95, 1.00, 1.00) (1.00, 1.00, 1.00) 19 (0.95, 1.00, 1.00) (1.00, 1.00, 1.00) 20 (0.95, 1.00, 1.00) (1.00, 1.00, 1.00) 21 (0.95, 1.00, 1.00) (1.00, 1.00, 1.00) ments based on the 21 cases shown in Table 6. Both o f these experiments were run using the H a m m i n g distance measure and the associated similarity measure as defined by Eq. (20) in identifying the rule to be selected and in identifying the linguistic descriptor to be displayed as the system response. It is observed in Table 7 that, with the exception o f case 7, there were no differences in the rule selection. In particular, for case 7, we observe that max-min operators selected rule 5 but bold union/intersection selected rule 6. Furthermore, there were no differences in the system response in terms o f the linguistic descriptors on the surface. H o w e v e r , if we look at the internal value o f DV* vectors before it is approximated to a linguistic descriptor via the use o f H a m m i n g distance and similarity measures, we observe some differences (see Table 8). Hence, we realize that the similarity measures have a smoothing effect on the system response in terms o f the linguistic descriptors. Whether such a smooth- ing effect is desirable or not m a y be context- and domain-dependent. There- fore, a system designer in cooperation with the users can decide whether to output the system response in terms o f just the linguistic descriptors or just DV* or both. It should be noted that in Table 8 we list only cases 2 - 7 and 1 2 - 2 1 , where the observed system states do not match the left-hand side o f any rule in the rule base. Since cases 1, 8, 9, 10, and 11 have an exact match, we have no reason to list those cases. As explained in Example 1, the system response corresponds to the right-hand side o f the rule as expected. 246 I.B. Turksen and M. Berg Let us now reconsider cases 2 - 7 and 12-21. They were selected randomly, and case 2 is explained in detail in Example 2. It is rather interesting to note that cases 2 - 7 describe an observed system behavior around " m e d i u m " and the system response is " m e d i u m , " corresponding to our expectations. Simi- larly, when the system behavior is around " l o w , " as in cases 12-16, the system response is " l o w , " and when it is around " h i g h " as in cases 17-21, the system response is " h i g h , " again corresponding to expectations. This is an indication o f robust system behavior. Clearly these results support the hypothe- sis that approximate-reasoning-based expert systems would better serve the operations managers o f such robust systems. Finally, these simulation experiments appear to suggest that once the rule is identified with the help o f a similarity measure we could directly fire the rule, that is, give the right-hand side o f the rule as a response linguistic variable. This needs to be validated theoretically and experimentally in the future. I f this finding is true, then there would be no need for a compositional rule o f inference. This would be analogous to modus ponens in two-valued logic. Could this be true for the case of robust systems? C O N C L U S I O N S In this paper, we have described our approximate reasoning approach for an expert system design and development as an aid to management confronted with a production/inventory capacity problem. We have attempted to show how operations research and approximate reasoning can be synthesized for the solution o f real-life problems in the era of knowledge-based systems. Opera- tions research methodologies can generate valuable insights into the under- standing o f problem domains. Approximate reasoning provides a framework where the insight gained from OR models could be restructured for real-life problems either where there is insufficient information to identify parameters o f the system at hand or where such data are not available owing to various factors and uncertain environmental conditions. Hence, the best we can do is rely on OR experts' assessments and the interpretation o f real system behavior via model analysis. Since such assessments can best be expressed in a natural language setting with linguistic terms providing flexibility o f expression in human judgments, and since approximate reasoning based on fuzzy logic can handle such linguistic uncertainties and imprecision, it appears that the mar- riage o f operations research and approximate reasoning is inevitable with the advance o f expert systems. Approximate reasoning with linguistic variables and their terms provides aggregation o f human knowledge as well as user friendliness. Thus, approxi- mate reasoning is a reasonably good analog to human reasoning. This was illustrated with the second example in the last section. Furthermore, the Inventory Capacity Planning 247 simulated cases 2 - 7 and 12-21 discussed in the previous section indicate and support the commonsense reasoning that managers in charge o f robust systems would get the appropriate support without detailed precision with the aid o f approximate-reasoning-based expert systems. The approximate reasoning shows an advantage o v e r analytical models in that it allows a certain degree o f freedom f r o m accuracy in the information and knowledge acquisition. This freedom and flexibility combined with the p o w e r o f the aggregation due to linguistic variables allows expert system designers and developers to summa- rize the available knowledge in terms o f far fewer rules than if one were to design expert systems that require precise information leading to rule explo- sion. On the other hand, the extra intelligence provided in the inference procedure gives us a way to cope with situations not explicitly included in the rule base and hence with unforeseen future conditions in the observed system behavior. This means that the use o f an expert system is not confined to the set o f rules provided in the knowledge base. There are certain issues that need m o r e elaboration. Some o f these are: 1. Throughout our discussion, point-valued fuzzy sets are employed instead o f interval-valued ones, with the exception o f C N F and D N F expres- sions, which creates interval-valued results for the logical combinations. H o w e v e r , by taking the average o f the upper and lower bounds, respec- tively, we reduce the intervals to points. H o w e v e r , several experimental results obtained so far suggest that interval-valued fuzzy sets represent experts' assessment m o r e naturally. Even though we have shown that interval-valued inference is quite possible, computations required for interval-to-interval inference are a lot more complex and costly at this point in our research (Turksen [9], Turksen and Zhong [3]). 2. In the inference procedure, we have used G M P and the H a m m i n g distance and a similarity measure based on this distance measure. This choice needs justification. It is known that there are other distance measures and other similarity measures (Zwick et al. [18], Turksen and Zhong [3]). H o w e v e r , the choice o f a distance and a similarity measure is still an open question. Until we identify the context-dependent effects o f these distance and similarity measures, we cannot know how to choose an appropriate measure. These issues are left for future research. A C K N O W L E D G M E N T S This investigation was supported in part b y the Natural Science and Engi- neering Council o f Canada and in part by the Manufacturing Research Corpo- ration o f Ontario. The software developments and experiments were carried 248 I . B . Turksen and M. Berg out by graduate students H. Zhao, I. Wilson, and P. Grant. We are grateful for all the support we have received. A P P E N D I X f o ( W ) = A ( e - ~ W - re-~2~,) A[ r ] A ( ~ o ) -- --ff ( . - x - ~ 1 ) ~ - ~ , ~ - - ( . - x - ~ ) e - ~ x A ( 1 - r) f o - X and f N = --~[ # - ~ k - /3' ( e - u N e - # ' N ) - r ( # - ~ k - /32) ( e - " N - - I a /32 -- I z where X A - I = ~1-1 1 - /32 ~'l /31] [/32-/3,_ + ( . - x - / 3 , ) ( . ~ , ~ ; ~ / 3 2 ) e " N e - 8 1 N e - B 2 N + - - - ~ ~ - / 3 2 B = - A r and ( ~ - X - / 3 , ) / ( ~ - / 3 , ) r = (# - x - / 3 2 ) / ( ~ , - / 3 2 ) where A , B , /3~, /3, and r are intermediate parameters used in the analytical solution procedure. A main concern o f management is the fraction o f satisfied customer de- Inventory Capacity Planning 249 mands. This can be expressed in terms o f the model parameters as # [ 1 - e - ~ l N r ] p= ~ A fl, ~ ( I - e -~2N) R e f e r e n c e s 1. Zadeh, U A., A theory of approximate reasoning, Memorandum No. UCB/ERL M77/58, 1977. 2. Turksen, I. B., Approximate reasoning for production planning, Fuzzy Sets Syst. 26, 23-37, 1988. 3. Turksen, I. B., and Zhong, Z., An analogical approximate reasoning based on similarity measure and interval valued fuzzy sets, Fuzzy Sets Syst. 34, 323-346, 1990. 4. Turksen, I. B., Interval valued fuzzy sets based on normal forms, Fuzzy Sets Syst. 191-210, 1986. 5. Turksen, I. B., Interval-valued fuzzy implication, in Recent Development in the Theory and Applications o f Fuzzy Sets (W. Bandler and A. Kandel, Eds.), NAFIPS'86, 576-592 (NAFIPS'86 Outstanding Paper Award), 1986. 6. Turksen, I. B., Measurement o f membership functions, in Applications o f Fuzzy Set Theory in Human Factors (W. Karwowski and A. Mital, Eds.), Elsevier, New York, 55-67, 1986. 7. Chameau, J.-J., and Santamarina, J. C., Membership function, Int. J. Approxi- mate Reasoning 1(3), 287-313, 1987. 8. Turksen, I. B., Fuzzy representation and inference with normal forms, in Man- agement Decision Support Systems Using Fuzzy Sets and Possibility Theory (J. Kacprzyk and R. R. Yager, Eds.), 213-228, 1985. 9. Turksen, I. B., Four methods of approximate reasoning, Int. J. Approximate Reasoning, 3 (2), 121-142, 1989. 10. Posner, M. J. M., and Berg, M., Analysis of a production-inventory system with unreliable production facility, Working Paper 88-02, Dept. of Industrial Engineer- ing, Univ. Toronto, Toronto, Ontario, June 1988. 11. Brill, P. H., and Posner, M. J. M., Level crossings in point processes applied to queues: single-server case, Op. Res. 2 5 , 6 6 2 - 6 7 4 , 1977. 12. Brill, P. H., and Posner, M. J. M., The system point method in exponential queues: a level crossing approach, Math. Op. Res. 6, 31-49, 1981. 13. Zysno, P., Representing human concepts by fuzzy sets: systematical conditions and empirical results, Reprints of Second IFSA Congress, Tokyo, July 20-25, 1987, 618-621. 250 I . B . Turksen and M. Berg 14. Norwich, M. A., and Turksen, I. B., A model for the measurement of membership systems and the consequences of its empirical implementation, Fuzzy Sets Syst. 12, 1-25, 1984. 15. Zimmermann, H.-H., and Zysno, P., Latent connectives in human decision making, Fuzzy Sets Syst. 4, 37-51, 1980. 16. Zadeh, L. A., The concept of a linguistic variable and its application to approxi- mate reasoning. I, II, III, Inf. Sci. 8, 199-249, 301-357; 9, 43-80, 1975. 17. Kosko, B., Fuzzy entropy and conditioning, Inf. Sci. 40, 165-174, 1986. 18. Zwick, R., Carlstein, E., and Budescu, D. V., Measure of similarity between fuzzy concepts: a comparative analysis, Int. J. Approximate Reasoning 1(2), 221-241, 1987. 
work_rqxvspxoird6bl7trrujddufx4	----	ON THE PLAUSIBILITY AND SCOPE OF EXPERT SYSTEMS IN MANAGEMENT Vasant Dhar July 1 9 8 5 Center for Research on Information Systems Information Systems Area Graduate School of Business Administration New York University Working Paper Series CRIS f 9 8 GBA 1 8 5 - 6 1 This paper appears in the J o u r n a l o f M a n a g e m e n t I n f o r m a t i o n S y s t e m s , Vol. 3, No. 5 , 1987 Center for Digital Economy Research Stem School of Business Working Paper IS-85-61 1 Abstract Over t h e last decade-there have been several efforts a t building knowledge based "expert systemsH , mostly in the scientific and medical arenas. Despite the fact t h a t almost all such systems are in their experimental stages, designers are optimistic about their e ~ e n t u a l success. In the last few years, there have been many references t o the possibility of expert systems in the management literature. However, what is lacking is a clear theoretical perspective on how various management problems differ in nature from problems i n other domains, and the in~plications of these differences for knowledge based decision support systems for management. In this paper, I examine some of these differences, what they suggest in t4erms of the functionality t h a t a computer based system must have in order t o support organizational decision making, and the scope of such a system as a decision aid. The discussion is grounded in t h e context of a computer based system called PLANET t h a t exhibits some of the desired functionality. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 1. Introduction Over the last several years, computer-based modeling systems have made i t relatively easy for end users t o develop powerful decision support systems in many application areas. Yet, there is a growing recognition t h a t unless such systems are augmented with representational frameworks and inference mechanisms t h a t take explicit cognizance of the intellectual component of managerial decision making, their utility as decision aids is limited. In parallel efforts in the field of Artificial Intelligence (AT), researchers have been concerned with similar issues, although in problem areas t h a t would probably be regarded as more "structuredn than those encountered in management. Some of the programs t h a t have resulted from this research, commonly referred t o as "expert systems", have received considerable attention because of their ability t o engage in judgmental reasoning similar t o t h a t of domain experts, and exhibit comparable levels of performance. It seems natural to ask whether similar systems might be built t o support decision making in the management arena where many of the more challenging problems tend t o be fairly open-ended, non- repetitive, a n d not amenable t o analytical solutions. Answering this question requires addressing four, more fundamental questions: 1 . what is the nat,ure of expertise in domains where knowledge based support systems1 have heretofore been developed, 2. what is t h e nature of complex managerial problems t h a t distinguishes them from the above class of problems, 3 . given these differences, what problematic aspects of management problems might. knowledge based systems be used t o support, 4. what system functionality and architecture are needed in order t o alleviate such problems. An answer t o the first of these is based on a summary of existing literature in cognitive science and expert systems. In order t o keep the discussion on the last three questions in focus, I restrict the discussion t o managerial problems t h a t require modeling problem situations for decision-making purposes. Specifically, I shall ground the discussion in the context of a resource planning problem for which I have attempted t o '1 use t h e t e r m Bknowledge based system' t o refer t o a program where domain-specific knowledge plays a m a j o r role in inference. T h i s is in c o n t r a s t t o p r o g r a m s t h a t use 'weak methodss (Newell a n d Simon, 1972), t h a t is, s y n t a c t i c , domain-independent m e t h o d s t o guide inference. Center for Digital Economy Research Stem School of Business Working Paper IS-85-61 develop a "planner's assistantH (called PLANET) t o help planning managers with t h e formulation and maintenance of planning models t o support decision making. T h e investigation was initiated by planning managers in a large computer manufacturing company ( W C ) who expressed concern over the inadequacies of existing computer based support tools and the need for a knowledge based tool t o support the planning function. This effort has brought into focus some of the problematic aspects of managerial problems such a s planning, sharpened the distinction between such problems and those encountered in other domains, and the implications of the differences for knowledge based support system architectures. 2. The Relation Between Expertise and Problem Type T h e type of knowledge required t o solve a problem is influenced by the degree t o which the task has been formalized [37]. As a domain becomes better understood, formal theories or normative models are articulated. These provide a basis for understanding and solving problems within t h a t domain. In the absence of this formalization, problem solving and understanding are more likely t o depend on informal, intuitive, possibly unarticulated models. In this section, I consider t h e nature of problem solving in domains t h a t lie a t three different points of this "structurednessH spectrum: highly formalized domains where clearly identifiable bodies of knowledge exist, less structured domains where expertise is more implicit but nevertheless identifiable, and unstructured problems where the knowledge brought t o bear in solving problems, is evolutionary and often "distributedH across several individuals. T h e last of these is characteristic of managerial planning, where information t h a t is used t o construct models for decision-making, is continually changing. 2.1. Expertise in Structured Problem Domains There have been many psychological studies of human problem solving mostly in problem domains t h a t would generally be considered "well structured". Broadly speaking, t h e problems studied have either involved "common sense" reasoning pertaining to everyday physical phenomena 116, 24, 8, 20, 1 2 1 , ~ or specialized knowledge from highly formalized domains such a s physics o r algebra [22, 38, 30, 6, 51. 'sometimes referred t o as #naive physicsm. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 Although humans are generally competent with naive physics problems, competence in solving real physics o r other scientific problems is less common. Larkin [22] explains this phenomenon as follows: *...the process of mentally simulating events so as t o predict their outcome, a facility possessed by most people for common contexts, is extended and refined in a skilled scientist to become a sharp and crucial intuition t h a t can be used in solving difficult, complex or extraordinary problems. Novices, lacking this extended intuition, find such problems difficult" (Larkin, 1983, p.75). Several studies of problem solving in these areas have contrasted expert and novice behavior in order to understand the nature of this extended intuition. A common finding has been t h a t the quality and speed of solution is influenced by the n a t u r e of the representation adopted. Experts appear t o possess the functional equivalent of a large set of perceptual patterns and an "indexing scheme" t h a t enables them t o perceive the important features of a problem. If t h e problem is not exceptionally difficult, they often work "forward" without trial and error (i.e. without the need for backtracking) from general principles toward results t h a t "include" the solution. Chi. et.al [6] explain this in terms of the ability of the expert t o rapidly categorize the problem into an appropriate "principle-oriented" schema. Once correctly classified, axiomatic knowledge can be used t o solve the problem in a primarily top-down manner. Many studies of human problem solving behavior have involved the design of simulation programs. Several of these programs have been used for theory development and validation in domains such as statics [30], dynamics [26], and electronics 141. Evidence gained from observations of human problem solving is typically used to judge the validity of these computational models. An understanding and measurement of the "quality" of expertise is facilitated considerably because of t h e existence of a stable, clearly identifiable body of knowledge in the form of theoretical principles o r normative models. Not surprisingly, expertise in these areas appears t o be highly correlated with individuals' abilities t o recognize and apply the appropriate physical principles involved. The major usefulness of computer based systems as support tools in these domains appears t o be as intelligent tutoring systems t h a t can take cognizance of students' naive concepts about scientific domains, and facilitate the transfer of a principled body of knowledge t o novices. Several experimental systems along these lines have been built for symbolic integration (Kimball, 1983), electronic troubleshooting (41, axiomatically based mathematics [39], probability theory [I], and a consultative system for MACSYMA Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 2.2. Expertise in Expert Systems Expert systems research has been influenced by a growing recognition t h a t high performance programs are not likely to emerge through the clever use of a few powerful domain-independent techniques, but through a systematic formalization and use of large amounts of domain-specific knowledge. T h e implications of this shift toward a "knowledge basedU approach are well summarized by Goldstein and P a p e r t [18]: "The fundamental problem of understanding intelligence is not the identification of a few powerful techniques, but rather t h e question of how to represent large amounts of knowledge in a fashion t h a t permits their effective use and interaction. The current view is t h a t the problem solver (whether man or machine) must know explicitly how t o use its knowledge - with general techniques supplemented by domain-specific pragmatic know-how. Thus we see AJ as having shifted from a power based strategy for achieving intelligence t o a knowledge based approach" [18]. Most A1 research in expert systems has involved development of large knowledge based systems in problem areas where consultative decision support is a practical necessity for solving difficult problems. Major efforts have been in medicine [31, 36, 4 1 , 1 , ~ geological exploration [15, lo,] mass spectroscopy interpretation [23], and computer layout 1251. In contrast t o physics-like domains, these areas are less well understood. Because of this, i t is much harder to measure expertise against a formal, axiomatized body of knowledge. Rather, expertise tends to be implicit, manifested by consistently high performance with difficult problems. These problems typically involve uncertain, ambiguous, and fragmentary data. An expert must therefore judge the reliability of facts in order t o clarify the problem, and acquire additional evidence in such a way so a s t o discriminate among competing conceptualizations of a situation. In affect, "noisy" d a t a coupled with a n inherently large search space requires the use of intelligent heuristics, typically refined through experience, in order t o impose pragmatic constraints on complex, open-ended problems. A major reason for the impressive performance levels of expert systems has been t h e extensive efforts by 3~~~~~~ 1411 specializes in glaucoma assessment a n d t h e r a p y , MYCIN 1361 in antimicrobial t h e r a p y , whereas CADUCEUS 1311 deals with t h e whole of internal medicine. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 system designers a t formalizing this mostly experiential, often subjective knowledge extracted systematically through expert^.^ In fact, an important benefit of this knowledge extraction exercise is the systematization of previously unrecorded or unexpressed knowledge. Some researchers consider the primary contribution of constructing expert systems in such domains as being one of theory formation, thereby moving such problems into the category of Ustructured" problems. 2.3. Expertise in Managerial Problems In the types of problems discussed above, the expertise involved is typically i n d i v i d u a l . F o r managerial problems however, i t is useful t o distinguish among problems where individual expertise or normative models are involved, and organizational level problems involving inputs from multiple individuals. Many attempts a t developing models of expertise for administrative problems have focused on the individual. Such models, some of which are embodied in computer based systems, have been designed in domains such as loan assessment and trust management 171, portfolio management 191, financial diagnosis [3], capital budgeting [2], and welfare eligibility [40]. In contrast to individual problem solving, organizational level problems introduce several types of complexity into the modeling process. These complexities are well chronicled in articles describing the early attempts a t building large corporate simulation models. In these efforts, detailed mathematical models of organizations were constructed compIex problems where closed form solutions were infeasible. [19,29,34,35]. A major motivation for developing such systems was they made i t possible t o evaluate the impacts of alternative policies, opportunities, and external events (all operationalized as parameters of the simulation model) a t the level of the firm. The major knowledge inputs into such models consisted of assumptions about the organization and its external conditions, obtained from multiple sources in the organization. These were then translated into detailed mathematical models for decision making. T h e essential features underlying this type of modeling activity are summarized a s follows: 1. Model Formulation as Assumption Synthesis: formulating models is a n inherently underconstrained exercise involving generation of alternatives for various parts of the task 4 ~ y t h e s a m e token, a continuing problem with such systems has been t h a t - t h e i r carefully crafted knowledge bases t e n d t o be extremely fragile - system behavior often changes in unforseen a n d undesirable ways when knowledge is added. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 environment and the making of choices from among them. Since these choices are often tentative, they can be viewed as assumptions or premises on which expectations and projections a r e based. Any quantitative model must be understood to be conditioned on one such set of symbolic assumptions. Model formulation as assumption synthesis is discussed more formally in section 3 . 2 . 2. Distributed Expertise: formulating models for decision-making involves many individuals from different levels and functional areas of an organization. There are seldom individual experts for broad-based organizational modeling; instead, knowledge about the alternatives in various p a r t s of t h e task environment is contributed by several individuals. A t higher levels, policy issues shape top level decisions. These provide the context for lower level strategies and decisions which can be expressed in terms of an algebraic/mathematical model. T h e form and implications of distributed expertise are discussed more formally in section 3.1. 3. The evolutionary nature of models: decisions are not "one shot" affairs. This contrasts with problem solving in expert systems and instructional systems in structured problem domains where solutions are typically " one-shot * , t h a t is, the decision maker obtains case data, engages in a consultative dialogue (with colleagues or a system), and a solution is obtained. Rather, in an ongoing enterprise, decisions are made in a context established by previous choices. New information is evaluated in light of existing assumptions and expectations. In some cases, the new information may be assimilated cleanly into the existing conceptual framework, perhaps resolving certain ambiguities or uncertainties in the prior assessment. In many cases, however, the new information can be accommodated only if prior assumptions a r e appropriately modified, perhaps leading t o radical restructuring of all or part of the situation model. Mechanisms for managing evolutionary models are described in section 3.3. I t should be noted t h a t our use of the term "distributed expertise" is qualitatively different from expertise in other domains in t h a t i t is neither anchored by a stable body of knowledge a s in physics, nor based on consistent virtuoso performance in some area such as medicine. Rather, i t is a consequence of the necessary diffusion of responsibility across multiple departments or individuals in an organization. The discussion so far is summarized in table 1 which draws o u t the essential features among t h e problem types in terms of five key features. In the following subsection, we discuss the implications of these differences for knowledge based decision support for management. 2.3.1. The Role and Scope of Knowledge Based Support Unfortunately, a fundamental problem with large scale organizational modeling is t h a t the richness of the modeling activity - t h e problem solving involved in formulating t h e algebraic model itself -- is n o t preserved systematically. T h e formulation or synthesis activity is in fact t h e most challenging and creative p a r t of modeling exercise t h a t shapes the structure of t h e mathematical model. Yet, if t h e Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 $ S f n r r , z m m c n X E F I ID a 5 X ' b d v P r r (Dert. & rt I-r. * W ' b VI PI r P O s m 0 "0 1 m OP n P, r- 3 m r* C. N n ~t R g :: c G v r n r * B 0 t n :X 3 m n 2 1 0 m o r. b*Cgn: 0 m a 3 1 p g q 0 a a o m P, m m m s a - m a r m P, ,re+ r . r - o m i ~ r ~ ' 0 3 I 1 r g F m n r . a m x 0 3 1 C 0 3 O O m n ~ v r n m w r C r X m m n C C n n g v . 0 0 1 m 3 C. 0 PI n r 1 t S c C 3 09 U W M O W X i5 S z n r m % B e cn I V) I I O C ( 3 n s s e R 0 1 1 * P , e , r . o 2 - 0 m 3 0 0 3 a C . K ) w n l a r ( a m m m C. P, a m m n 0 n X V. 3 1 b a o m 3 m 3 o n 1 1 r n v 1 ' 4 m sC w 0 b y m 3 n I.'. m Ca 0 cn n P, v r m 3 r: X b o r . r m CTlD P, n r d m r ~ D R W Y s m r. a tn a : F. o n r * R * v, g 8 a e n x 1 0 1 0 0 3 0 Dp v m m R w l m r m m 1 ' 4 r: m r t n 0 v n - g 2 3 3 0 00 r U W V ) 0 W e 5 s z n r ' n 2 !i! 3 ail u I 5.3 ID m o a 1 0 cC b 0 C m u 0 3 7 m 3 n I-. m F a n x C' 00 3 r CC n En P, v r W O - 1 1 6 o r. r v m P, r 3 m l D R m a m r. a m m : C. n * R 5 g 09 3 a - n 1 0 1 0 3 0 a m m r I m m a l 0": m n n 0 w n c1 v. 0 ;4. s 3 0, OP r !4 Z 9 ." o rn M c r j U u rn a M 6 x g ; 0 g E g r' H s 3 V) a 2 /ij f2 5 e rC m $;: m W 2 2 f! cn m 0 Z c n W 3c n r f!R!i! 8 relationships between a large algebraic model and the symbolic assumptions underlying i t are not faithfully preserved, the interpretation of the model becomes difficult, and modification of such models in light of a changing reality can be time consuming, ad hoc, and error prone. Their pragmatic problems notwithstanding, the basic objectives of such modeling efforts were reasonable and are still worth pursuing. If we recognize t h a t i t is actually the formulation/reformulation of the model based on changing assumptions t h a t is most problematic for a manager and his support staff, i t is in this activity where knowledge based support is m o s t needed. Conceptually, this can be achieved by representing the ancillary symbolic knowledge about models, which includes knowledge about the assumptions underlying the various model components. With such knowledge, t h e system can become an active partner in reasoning about changes to the model instead of burdening t h e user with the complete responsibility of maintaining and exploring models. F r o m a design standpoint, what is needed are structures and mechanisms for representing and manipulating the qualitative d a t a t h a t forms the basis for the quantitative model. This emphasis on d e s i g n of the quantitative model from fragmentary qualitative data, (as opposed to the selection from a predefined set of models) requires a computer-based architecture t h a t is capable of representing knowledge t h a t lies outside the scope of current day modeling systems. In the following section, I describe such an architecture t h a t has been shaped by the concerns articulated above. I limit the discussion to synthesis and maintenance of quantitative models only; i t is assumed t h a t if such a model is maintained, a n algebraic model corresponding t o i t can be formulated. 3. Knowledge Based Decision Support for Planning Planning is an important activity in most large organizations. Considerable time and effort of individuals from different parts of the organizations can go into building and maintaining models for planning. Several types of qualitative knowledge are involved in developing such models. However, most current day modeling systems d o not adequately represent such knowledge, thereby placing a heavy burden on the decision maker t o maintain the correspondence between t h e knowledge t h a t can be represented within the system and t h a t which cannot. In this section, I describe a system designed t o Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 represent and manipulate the diversity of knowledge involved in problems such a s planning. With this functionality, the system can play a more complete role in supporting decision makers. 3.1. Knowledge About Alternatives/Assumptions -- Distributed Expertise Conceptually, a n existing model can be viewed a s being the end result of a process involving consideration of a range of alternatives (assumptions) from various parts of the task environment. These alternatives may pertain t o decisions a t various levels of abstraction. For example, in the CMC manufacturing environment, these assumptions pertain t o computer technology t o be used in the product and t h e processes t o be employed in manufacturing it. Figure 1 shows a small set of alternatives about technology and testing processes t h a t might be considered in such a context. In P L m T , knowledge about these different parts of the task environment has been partitioned across a "society of agents* designed to represent standard areas of the planning activity or individual specialists in the different functional areas of the organization who have responsibility in the planning process. These specialists are represented as *objectsw in HOUSE [32], a Franz Lisp object oriented programming environment t h a t is similar in spirit to the FLAVORS package [27]. T h e objects correspond t o the real world entities in the domain under consideration. Referring t o figure 1 , each of the alternatives corresponds to a n object t h a t contains knowledge about a local p a r t of the task environment. Responsibilities of an object (which corresponds to a domain specialist) include responding to decisions being taken in other parts of the manufacturing environment and communicating its decisions so t h a t other specialists may also make appropriate adjustments t o their parts of t h e task environment. These "adjustments" are carried using "action oriented knowledge" which we describe shortly. Other, book- keeping oriented responsibilities of a specialist include keeping track of its current choice (with respect t o whatever decision(s) for which i t is responsible), reasons for it, and possible alternatives t o the existing choice. T h e implementation details of this are described in Dhar [13]. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 CMC Corporate D i v i s i o n s : A l t e r n a t i v e Methods : P o l i c y A l t e r n a t i v e s : A l t e r n a t i v e d i a g n o s t i c : p r o c e s s e s \ \ : e l l l p s l s \ \ \ Figure 1 A small set of alternatives considered in the course of formulating a plan in CMC's timbaktoo division. Choices a t the lower end of a line indicate alternative ways of accomplishing those indicated a t the top end of the line. Center for Digital Economy Research Stem School of Business \Vorking Paper IS-85-6 1 3.2. Assumption Synthesis as State Space Search There are t w o sources of "action orientedn knowledge t h a t are important in assumption synthesis. First, the problem domain itself provides constraints t h a t reflect certain relationships among different parts of the task environment t h a t must be realized. F o r example, in the computer manufacturing environment, two such domain-specific constraints (which we illustrate via a n example shortly) are: 1. "A decision to employ embedded etch board technology rules out using test processes designed for surface etch technologyn5 2. "Using Hitech's etch process requires using Hitech's heatsink technology t o o U Both these constraints are indicated in the search space shown in figure 2. As long as such constraints are applicable, the problem solver is in a "constrained mode.n Thus, a choice on what technology t o use would rule o u t certain testing processes. This could in t u r n trigger other similar rules, setting off a chain of choices. As long as there are such choices t o be made -- either due to a constraint or because there is only a single alternative with respect t o some decision -- the program is in a "constrained mode." There is also a second, quite different way by which choices are made. This is when all possible ramifications of a choice have been propagated and the problem is not yet fully solved, leaving the program in a "quiescent" state. In such situations, a "forced choice" is necessary in order t o continue with the formulation process. This is a characteristic of problems t h a t are inherently ~rnderconstrained, t h a t is, t h e constraint relationships alone are not sufficient t o make choices in all the required parts of the task environment. This requires the program to focus on some area of the task, and evaluate the set of alternatives available there. PLANET assesses t h e desirability of available alternatives on the basis of how they contribute toward the goals and objectives of the organization. This is operationalized as a pairwise comparison of alternatives on an "objectives vector" consisting of resources such as capital, space, and labor.6 The choice is determined on the basis of the resources required by the alternatives in ' ~ m b e d d e d etch boards technology refers t o boards where signals travel through t h e body of t h e board a s opposed t o i t s surface only. F o r a computer manufacturing company, the decision t o use such a technology is a strategic one and has i m p o r t a n t ramifications for decisions in related parts of the task environment. ' ~ e c a u s e t h e program must also sometimes compare high level alternatives for which detailed resource tradeoffs a r e impossible to assess before t h e details about these alternatives have been specified, * m a c r o level' knowledge is used in such situations. Basically, this heuristic knowledge consists of high-level associations a b o u t how t h e various alternatives typically compare across t h e various resources. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 M o t o r o l a ' s p r o c e s s t e s t e r s \ I \ / LEGEND : A -*-*-> B: c h o i c e of A r u l e s o u t B A ----- > B: B is a p o s s i b l e c h o l c e A -..+C> B: c h o i c e of A i m p l i e s B ----- >I): intermediate c h o i c e s Figure 2 A small section of a search space indicating a sequence of choices. In this space, a terminal node would represent a fully formulated plan incorporating a trajectory of choices indicated by the nodes Cl,C2,C3 & C4. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 light of the organization's resource availability picture. A "choice point" involving such a comparison is shown in figure 2. The process of successively making decisions pertaining t o various aspects of the task as shown in figure 2 can be viewed as a state space search. States toward the right represent choices being made in increasingly complete plans. T o summarize, two types of "action oriented knowledge" are brought t o bear in assumption synthesis. First domain-specific constraint relationships among different parts of the task environment must be taken into account. Once the choices resulting from these constraints have been exhausted, i t is necessary t o make forced choices. This requires the program t o focus on a critical p a r t of the task, and make a choice based on a heuristic evaluation function t h a t compares alternatives based on their resource requirements and the existing availability of resources. This can in t u r n lead t o further choices based on constraint relationships. This cycle continues until selections have been made from all parts of the task environment. 3.3. Preserved Process Knowledge The formulation process described above can be viewed as the result of a trajectory of choices in a state space, with the terminal nodes, if generated, representing "complete plans" from which algebraic models can be derived. This includes choices made by the program in its constrained mode, and the forced choices where alternatives are compared across the vector of objectives. Since some of these choices may have the effect of influencing others, the complete plan consists of "clusters of dependencies" in the state- space. One such cluster is shown in figure 3. Comparing figures 2 and 3, we can see t h a t a choice is not necessarily dependent on all chronologically earlier decisions, b u t only on those t h a t directly or indirectly led t o it. This "dependency informationW can play an important role in t h e incremental modification of a large plan. Referring to figure 3, we can see t h a t if the choice " 3 dimensional testers" is retracted, choices dependent on i t and all their dependents if any, need t o be undone. Revised choices in the affected areas are made from the available alternatives. In this example, retracting t h e embedded etch boards decision would also bring into contention, previously eliminated alternatives pertaining to t h e surface etch Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 3 d i m e n s i o n a l t e s t e r s NOT M o t o r o l a ' s TI'S s u r f a c e e t c h s u r f a c e e t c h p r o c e s s p r o c e s s H i t e c h ' s embedded e t c h ( p r o c e s s ) H i t e c h ' s h e a t s i n k embedded e t c t e c h n o l o g y LEGEND A + B c h o i c e of B d e p e n d s on c h o i c e of A Figure 3 A dependency network corresponding t o t h e choices indicated in figure 2 Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 processes. T h e revised choice for test processes would then be made among the previously passed over alternatives (indicated in figure 2), plus others t h a t might have become available since a choice was iast made in t h a t p a r t of the plan. For plans containing thousands of choices, this process of i n c r e m e n t a l plan evolution can serve a n important attention focusing role by highlighting only the affected areas of a plan, and suggesting revised choices in these areas. Maintaining t h e state-space associated with a n existing plan can also be useful for carrying o u t qualitative "what if" analyses of choices. F o r example, a query of the form "what if I use the surface etch board technology" boils down t o undoing the dependencies of the existing assumption (namely, the embedded etch technology), making revised choices for these parts of the task environment, and generating the resource requirements for the hypothesized scenario. This elevates t h e what-if analysis from t h e level of a quantitative model t o one allowing for perturbations of the symbolic assumptions underlying such a model. More generally, this functionality is a statement about the rote of a u s e r in such man-machine interactions. Most DSS literature uses t h e ambiguous term "judgement" t o account for the gap between the symbolic reasoning process of a decision maker and the outputs from a model underlying a system. Unfortunately, this view of decision support does not address issues about whether i t is reasonable t o expect t h e user to make all the right "adjustments" in translating qualitative reasoning into a form expressable for the quantitative model. In contrast, elevating t h e system functionality t o a level where the symbolic real-world assumptions can be manipulated relieves the user from making possibly unrealistic transitions between the two levels. 3.4. Summary of the Main Points It is worth summarizing the discussion so far in light of the four questions raised a t the beginning of this paper, in particular, the iast three. A fundamental characteristic of much of managerial problem solving is t h a t t h e symbolic knowledge about a problem domain is distributed and evolutionary, and must be maintained. A major problem facing planning managers is one of orchestrating t h e synthesis of assumptions into a coherent model, and Center for Digital Economy Research Stem School of Business \Vorking Paper IS-85-6 1 because of the instability of such assumptions, one of maintaining the integrity of this model over time. My approach t o decision support emphasize the design or maintenance of a qualitative model on which a quantitative model can be based. In contrast, most DSS approaches have viewed support via t h e selection from a predesigned set of models. Similarly, the design emphasis here also contrasts with most knowledge based systems to date which have been concerned mainly with classification problems t h a t involve mapping "factsu t o "conclusionsw, given a stable model of the domain. In contrast, I have argued t h a t i t is the formulation and maintenance of t h e model of the domain itself t h a t is a particularly problematic reality in organizations t h a t knowledge based systems can support. F r o m a functionality standpoint, modeling systems o r "DSS generators" form one component of such a support system. They are appropriate for representing algebraic models and performing parametric explorations within a given algebraic model structure. However, much of t h e problematic aspects of managerial decision making involve "getting the model right", a n exercise t h a t must make use of symbolic knowledge not expressable within modeling systems. F o r a system t o be sensitive t o the context of the decision making process, i t must be able to explicitly maintain and reason in terms of this process knowledge, and tie the outputs of this process with a modeling system. Basically, this requires a level of intelligence over and above the knowledge expressed in a n algebraic modeling system. In order for a system to maintain the context surrounding its models, a decision support system must therefore maintain knowledge about alternatives, general domain-specific constraints, resource availability information, and dependency among prior decisions. These constitute the qualitative knowledge components required in order to synthesize and maintain evolving models. Equipped with this functionality, knowledge based systems can play an important role in facilitating a n incremental evolution of models, and provide a continuity perspective t h a t is crucial to managerial decision making, but lacking in the support provided by current day systems. Center for Digital Economy Research Stem School of Business \Vorking Paper IS-85-6 1 4. Summary Much of the power of the PLANET architecture derives from its ability t o collect, preserve, and manipulate a store of domain specific knowledge in order t o reason about a problem situation. This knowledge must be provided to the system by the user. However, a n important part of a manager's job is to create the alternatives and recognize their interrelationship. Reitman [33] suggests t h a t t h e process of generating good moves or actions, particularly in the game playing context, is similar in spirit t o heuristic search. While this approach may be reasonable for domains where the entire set of alternatives, however large, can be generated a priori (i.e. the search space has a definite size), i t is of little value in a managerial planning situation where actions are not defined a priori, but continually generated or "recognized". In fact, a n important function of a h u m a n support staff is one of creating a set of "good" actions t o be examined by a decision maker [33]. The PLANET formalism is limited from this standpoint in t h a t i t is a reactive support tool; the inputs t h a t enable i t t o modify a plan must always come from the user. T h e realization of good actions also must come from the user. It is probably accurate t o say t h a t these creative aspects of decision making are likely t o remain outside the scope of computer based support in the near future. In conclusion, while computer based decision support systems will continue t o have certain limitations a s decision aids, there is nevertheless considerable support potential above and beyond what i s available with current day systems. In this paper, I have attempted t o address what I consider t o be important issues t h a t must be addressed if we are t o develop knowledge based systems t h a t exhibit some of t h e intelligence t h a t is associated with managerial decision-making. Specifically, I have argued t h a t since models used to support decision-making are based on evolving knowledge, such systems must be able t o represent and maintain such knowledge. Although such systems are not "expert systemsD such as those in the scientific and medical arenas, they can nevertheless use knowledge about a problem situation in supporting a decision maker with the formulation and maintenance of assumption-based models relevant t o t h e problem situation. The architecture described in this paper has been designed t o support this activity. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 Acknowledgements I am grateful to Herb Simon, Michael Ginzberg, Barry D. Floyd, and Joseph Davis for helpful comments on earlier drafts of this paper. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 REFERENCES 1 . Barzilay, Amos., SPIRIT: An Intelligent Tutoring System for Probability Theory, Ph.D Thesis, University of Pittsburgh, 1984. 2. Bohanek,M., Bratko, I., and Rajkovic, V., An Expert System for Decision Making, in Processes a n d Tools for Decision Support, Henk Sol (ed.), North-Holland, 1983. 3. Bouwman, M., Human Diagnostic Reasoning by Computer: An Illustration From Financial Analysis, Management Science, vol. 29, no. 6, June, 1983. 4. Brown, J.S., Burton, R., & de Kleer, J., Pedagogical, Natural Language and Knowledge Engineering Techniques in SOPHIE I, I1 and III, in Intelligent Tutoring Systems, Sleeman a n d Brown (eds), Academic Press, 1983. 5. Bundy, A., & Bird, L., Using the Method of Fibres in MECHO t o Calculate Radii of Gyration, in Intelligent Tutoring Systems, Sleeman and Brown (eds), Academic Press, 1983. 6. Chi, M.T.H., Feltovich, P., & Glasser, J., Categorization and Representation of Physics Problems by Experts and Novices, Cognitive Science, 5, 1981. 7. Clarkeson, G.P.E., A Model of the Trust Investment Process, in Computers a n d Thought, Feigenbaum and Feldman (eds.), McGraw-Hill, 1963. 8. Clement, J., Students' Preconceptions in Introductory Mechanics, American J o u r n a l of Physics, 50, 1982. 9. Cohen, P., & Lieberman, M., A Report on FOLIO: A n Expert Assistant for Portfolio Managers, Proceedings of the Eighth International Joint Conference on Artificial Intelligence, 1983. 10. Davis, R., Austin, H., Carlbom, I., Frawley, B., Pruchnik, P . , Sneiderman, R., & Gilreath, A., The Dipmeter Advisor: Interpretation of Geological Signals, IJCAI, 1981. 11. Dearborn, D.C. and Simon, H.A., Selective Perception: A Note on the Departmental Identifications of Executives, Sociometry, volume 21, 1958. 12. de Kleer, J., and Brown, J.S., Assumptions and Ambiguities in Mechanistic Mental Models, i n Intelligent Tutoring Systems, Sleeman and Brown (eds), Academic Press, 1983. 13. Dhar, Vasant., PLANET: An Intelligent Decision Support System for the Formulation and Investigation of Formal Planning Models, Ph.D Thesis, University of Pittsburgh, 1984. 14. Dhar, Vasant., and Quayle, Casey ., An Approach t o Dependency Directed Backtracking Using Domain Specific Knowledge, Proceedings of the Ninth International Joint Conference on Artificial Intelligence (IJCAI), 1985. 15. Duda, R.O., Gashnig, J., & Hart, P . , A Computer-Based System for Mineral Exploration in Experts Systems i n the Microelectronic Age by Michie (ed), Edinburgh Press, 1979. 16. Forbus, K., Qualitative Reasoning About Space and Motion., in Intelligent Tutoring Systems, Sleeman and Brown (eds), Academic Press, 1983. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 17. Genereseth, M. R., The Role of Plans in Intelligent Tutoring Systems, in Intelligent Tutoring Systems, Sleeman and Brown (eds), Academic Press, 1983. 18. Goldstein, I., & Papert, S., Artificial Intelligence, Language, and the Study of &owledge, Cognitive Science,l, 1977. 19. Gordon, R.J., Financial Modeling on Small Systems, I B M Systems Journal, volume 23, May- June, 1973. 20. Kuipers, B., Modeling Spatial Knowledge, Cognitive Science, 2, 1978. 21. Larkin, J., Problem Representation in Physics, in Intelligent Tutoring Systems, Sleeman and Brown (eds), Academic Press, 1983. 22. Larkin, J., McDermott, J., Simon, D.P, Simon, H.A., Models of Competence in Solving Physics Problems, Cognitive Science 1980. 23. Lindsay, R., Buchanan, B., Feigenbaum, E.A., & Lederberg, J., Applications of Artificial Intelligence for Chemical Inference: The DENDRAL Fkoject, McGraw-Hill, 1980. 24. McCloskey, M., Naive Theories of Motion, in Intelligent Tutoring Systems, Sleeman and Brown (eds), Academic Press, 1983. 25. McDermott, J., R1: A Rule-Based Configurer of Computer Systems, Artificial Intelligence, vol 19, no. 1, 1982. 26. McDermott, J., & Larkin, J., in Proceedings of the 2nd Conference of the Canadian Society for Computational Studies of Intelligence, 1978. 27. Moon, David. and Weinreb, Daniel., Lisp Machine Manual, MIT, 1981. 28. Naylor, T., The Politics of Corporate Model Building, Planning Review, no. 13, January 1975. 29. Naylor, T., and Schauland, H., A Survey of Users of Corporate Planning Models, Management Science, May 1976. 30. Novak, G . , Computer Understanding of Physics Problems Stated in Natural Language, Tech. Report NL-30, Department of Computer Science, University of Texas a t Austin, 1976. 31. Pople, Harry, E., Heuristic Methods for Imposing Structure on Ill-Structured Problems: The Structuring of Medical Diagnostics, Artificial Intelligence i n Medicine, Peter Szolovits (ed), Westview Press, Boulder, Colorado, 1982. 32. Quayle, Casey., Object Oriented Programming in Franz Lisp, Working Paper, Decision Systems Laboratory, University of Pittsburgh, 1983. 33. Reitman, Walter., Applying Artificial Intelligence to Decision Support: Where do Good Alternatives Come From? in Decision Support Systems, Ginzberg, Reitman and Stohr (eds), North-Holland, 1982. 34. Rosenkranz, F., An Introduction to Corporate Modeling, Unpublished Ph.D Dissertation, University of Basel, Switzerland, 1975. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 35. Shannon, R., Systems Simulation: The A r t and Science, Prentice-Hall, 1975. 36. Shortliffe, E., MYCIN: Computer Based Medical Consultation, American Elsevier, 1976. 37. Simon, Herbert A., The New Science of Management Decision, Prentice Hall, 1960. 38. Simon, H.A., & Simon, D.P., Individual Differences in Solving Physics Problems, in Children's Thinking: What Develops?, Siegler (ed), Erlbaum, 1978. 39. Smith, R.L., Graves, H., Blaine, L.H., & Marinov, V.G., Computer-assisted Axiomatic Mathematics: Informal Rigor, in Compute Education Learne and Lewis (eds), North Holland, 1975. 40. Subrahmanian, Eswaran., WELDS: A Welfare Eligibility Determination System - An Expert System in an Administrative Context. Expert Systems in Government Symposium, Washington D.C. 1985. 41. Weiss, S., Kulikowski, C.A., Amarel, S., & Safir, A., A Model-Based Method for Computer- Aided Medical Decision Making, Art4 ficial Intelligence, vol. 11, no. 1832, 1978. Center for Digital Economy Research Stem School of Business IVorking Paper IS-85-6 1 
work_rxbr5e34ffcrrfeasj5wwzpuwy	----	International Journal of Scientific Research in Computer Science, Engineering and Information Technology CSEIT206118 | Accepted : 01 Feb 2020 | Published : 06 Feb 2020 | January-February-2020 [ 6 (1) : 91-104 ] International Journal of Scientific Research in Computer Science, Engineering and Information Technology © 2020 IJSRCSEIT | Volume 6 | Issue 1 | ISSN : 2456-3307 DOI : https://doi.org/10.32628/CSEIT206118 91 Expert System for Poultry Management Dr. Oye, N. D.*, Jemimah, N. Department of Computer Science, School of Physical Sciences (SPS), Modibbo Adama University Technology (MAUTECH), P.M. B. 2076, Yola, Nigeria ABSTRACT This research is aimed at developing a Software Framework for poultry management with the objectives: To provide poultry farmers with expert knowledge from the comfort of their homes. The system is developed using generic iterative waterfall model as methodology, HTML, CSS, JavaScript and Bootstrap for graphical user interface. PHP is used as programming language and MySQL as DBMS and normalization technique. The system is hosted locally for validation and testing on Xampp local host. The researchers recommended Poultry farmers to adopt and use the software framework because it can give knowledge on how to manage poultry expert system of different categories in poultry farming. Keywords : Expert, System, Poultry, Management, Generic. I. INTRODUCTION Background of the Study Globally, the demand for the poultry source food has been growing exponentially, particularly in developing countries due to urbanization, income and population growth (Meng-Qiang 2009). However, despite the growing demand for poultry products, poultry farmers worldwide face numerous problems. In Nigeria and Ghana, for example, poultry farmers have suffered setbacks in poultry production due to lack of sources for farmers to access information from experts about farm inputs management and because of that mysterious disease have wiped out the farmer’s birds. The high costs of inputs and the mysterious diseases have significantly reduced the returns of the poultry farmers businesses (Olarinde & Kuponiyi, 2010). In Uganda, the relative importance of poultry industry, particularly the art of keeping traditional birds cannot be under-rated in terms of the provision of the livelihoods of the low income families in the rural and semi-urban areas. This is because local poultry meat, especially chicken meat, has increasingly dominated the diets of many Ugandan rich families (Olarinde & Kuponiyi, 2010). However, the consumption of chicken in poor rural families has been declining due to increasingly high prices. Thus, 80% of the total poultry meat in the markets is consumed by effluent urban and semi-urban dwellers due to their income growth and high purchasing power (Mugga 2009). However, in low-income rural family’s poultry products consumption has declined to approximately 7% only of the country’s total consumption of poultry’s products. Poultry production and farming system has been evolving into a complex business system requiring the accumulation and integration of knowledge and information from many diverse sources. In order to remain competitive, the modern farmer often relies on poultry specialists and advisors to get information for decision making. Unfortunately, assistance of the poultry expert is not always available when the farmer needs it. In order to alleviate this problem, expert systems were identified as a powerful tool with http://ijsrcseit.com/ http://ijsrcseit.com/ http://ijsrcseit.com/ https://doi.org/10.32628/CSEIT206118 Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 92 extensive potential in agriculture (Yelapure & Kulkarni, 2012). Expert system for poultry management is a web based system that will offer a technology based solution that will allow farmers keep up rapidly with information on poultry production records and weigh their income and expenses been made on the farm. Management is the process of working with and through others to archive organizational objectives in a changing environment, is the set of principles relating to the function of planning, organizing, directing, and controlling and the application of these principles in in harnessing informational resources efficiently. Central to this process is the effective and efficient use of limited resources (Kreitner, 2010). Poultry management is concerned with how can the individual farmer so organize the factors of production land, labour and capital on his poultry farm, so adapt practice to his particular environment and so dispose of his product, as to yield him the largest net return, while still maintaining the integrity of his land and equipment ( Butterfield , 2009). Information and Communication Technologies (ICTs) are bringing significant changes to India, as elsewhere. Such ICTs can be exploited to design cost effective systems to provide expert advice particularly to rural communities, helping increase productivity and livelihoods (Adeyemo, 2013). An Expert System is an intelligent computer program that uses knowledge and inference procedures to solve problems difficult enough to require significant human expertise to solve (Balakrishnan, 2013). Expert Systems provide a framework for presenting the latest scientific knowledge and decision making tools. Although Expert Systems have been developed in many agricultural science disciplines, such systems do not always adequately address the end user. The issues and challenges in the development of such systems include involving the user in development, building consensus between developer and user on the definition of usefulness, delivering computer based technologies and the challenges of implementing the system rather than its parts (Halimat & Amosa, 2012). Poultry Expert System (PES) attempts to address these issues and was developed using hypertext markup language (Html), cascading style sheet (Css), and JavaScript on four dimensions of poultry farming i.e. diseases, biosecurity, summer management and drugs used. The present study was carried out to test its applicability among end users’ veterinarians and veterinary students. An Expert System (ES), also called a Knowledge Based System (KBS), is a computer program designed to simulate the problem-solving behavior of an expert in a narrow do main or discipline. The expert system could be developed for decision-making and location specific technology dissemination process. An expert system is software that attempts to reproduce the performance of one or more human experts, most commonly in a specific problem domain, and is a traditional application and/or subfield of artificial intelligence. Expert systems help in selection of crop or variety, diagnosis or identification of pests, diseases and disorders and taking valuable decisions on its management. The expert system which developed earlier were more of text based and could be utilized only by the extension officials and scientists. Keeping the importance of ICT enabled interventions in agriculture and providing timely expert advice to farmers, the expert system on agriculture and animal husbandry was proposed and obtained as network project from Indian Council of Agricultural Research. Statement of the Problem The recent ban on importation of poultry and its products by federal government has really encouraged local production and high return on investment by poultry farmers. In other words, this has led many crop farmers, civil servants and other people in other ventures, to incorporate poultry farming into their http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 93 businesses. However, the vital role of poultry production to both farmers and the nation in general cannot be achieved, if there is no effective knowledge to the farmers and other stakeholders through expert system (Adeyomo, 2013). Poultry farmers however, often depend excessively on the interpersonal sources such as friends, fellow poultry farmers, middlemen, feed sellers and health care providers as their major sources of poultry farm knowledge about poultry production, despite the availability of the needed knowledge through expert system. Consequently, very little of the needed knowledge reach rural communities and Nigeria is clustered within urban areas where more of the farmers live and the actual poultry farming takes place. Small-scale poultry farming which is new to many farmers in Nigeria unlike crop production; it is a knowledge dependent agricultural venture, which involves a lot of risk, thereby leading to loss or low profit, which discourages incoming poultry farmers, breakthrough of diseases in poultry farms and inflation in poultry products. This study sought to ascertain the impact of expert system in poultry farming and develop a framework of knowledge base poultry expert system to solve the above stated problems. This study will help in a good number of ways that include reduced cost in running poultry farms which will make farming more efficient and profitable and encourage poultry farming, prevent epidemic in poultry farms, reduced cost of poultry product to customers, to enhance performance of Agricultural extension personnel and farmer, and to reduce the time required in solving the problems. The scope of this research work is to be able to design and develop a software framework for knowledge base of poultry based expert system that runs on a Personal Computer (PC). This research will not cover aspect of software that runs on an Internet-based or on a network. Therefore, it is restricted to those operations that involve an off-line mode. II. LITERATURE REVIEW Expert system is a computer program that contains expert knowledge about a particular problem domain, often in the form of online or we-based application is able to solve the problems at a level equivalent to or greater than human expert. Knowledge Engineer collects knowledge from domain expert and transfers it into production rules and creates Knowledge base. Inference engine then apply different knowledge acquisition techniques and catch the knowledge and deliver it in the form of advice to solve problem. Because of its explanation facility, it can be used as training tool to agricultural persons. High yield is aim of a poultry product is achieved by acquiring expert knowledge, so that depending on that knowledge, Farmer can take decision related to different factors like the type disease affecting poultry, the cure for the disease and the and the necessary measures to take to prevent future occurrence of such disease. Expert system can be used to make decision at different levels in poultry farming, i.e. at operational and planning level. On operation level, the extension workers in the villages, districts and /or government can use the system to support him in making his decision in giving appropriate advice to growers. On the planning level, the decision makers can use expert system who establishes need of water, vaccines and feeds of the poultry. (Rafea, 2009). Oluby & yerulu (2018) Expert system is one of the tools of AI that could be used to solve various types of real life problems. In expert system ES, the knowledge of expert in a particular domain is transferred to computer machine using the appropriate computer code for the machine to apply the knowledge to solve http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 94 problems in such domain in human-like manner. The machine therefore follows the footstep of the expert to arrive at reasonable solutions. Based on the acquired knowledge, the system is capable of explaining the reason or reasons for arriving at a particular solution. Therefore inane attempt to develop an expert system, a domain expert in a specific field has to be interrogated and the acquired knowledge is stored in a suitable form for solving specific problem using simple reasoning. Poultry refers to bird that people keep for human consumption. It generally includes chicken, turkey, duck, goose, quail, pleasant, pigeon, guinea fowl, peafowl, ostrich, emu and rhea. Thammi & Sudhakar (2011) presented a research work titled “An Information Technology Enabled Poultry Expert System: Perception of Veterinarian and Veterinary Students, University, India”. The major aim of the work was not to carry out diagnosis but rather to collect relevant information from students and veterinarians of the institution on the effectiveness of using expert system for diagnosis and management of poultry diseases. “Animal knowledge based system” was developed by Maryam and Alaa (2013). The article present different methodologies of developing knowledge base system and developed an animal knowledge base system in Egypt. The system was not particular about poultry diseases. Andino and Mahmud (2012) developed poultry disease warning system using Dumpster-Shafer theory and web mapping. Obviously, the method used is more of Statistics and Geographical Information System. The principle adopted looks different from that of Expert System. It was built to visualize map so as to identify the existence of poultry disease in a region by the district, regencies or municipalities. Andino and Hassan (2012) present a research work titled “Avian Influenza Expert System using Demeter- Shafer. The system was able to identify Avian Influenza disease and display the result of identification process. The system only made use of basic probability assignments of symptoms and not Expert System approach to build the knowledge base. The system was also restricted to Avian Influenza as the only type of disease and other common poultry diseases were not considered Going by the literature, it can be observed that each of the existing system have one weakness or the other. This research work is expected to come up with a better system that works on the principle of AI (Artificial Intelligence). Again, Most of the existing systems were not evaluated. In the very few ones that were evaluated, experts in the field were not directly involved. The new system will come up with better evaluation method that that will involve expert from the field. Expert system combines a lot of knowledge of so many experts at one point. By helping people to consider all of the relevant information and by assimilating this information into an understandable format, Expert system assists people in making of environmentally sound and economically viable for farm management decision (Robinson, 2018). A poultry Management Information System is a system for ‘‘collecting, processing, storing and disseminating of data in the form of information needed to carry out the operations functions of the farm’’ (Salami, Ahmadi, 2010). Poultry Management Information System tools helps poultry farmers to identify areas of their operation that needs improvements and guidance on how to address the issues focusing on the potential risk areas and opportunities for improvement, the specific modules include poultry farm (husbandry). The Poultry Management System approach is designed to help poultry farmers better plan their management process, access management performance and effectiveness of management practices, identify for opportunities for improvement and efficiencies. The Poultry http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 95 Management System is a step by step approach; through the implementation of a poultry management system, farmers can improve competitiveness by highlighting where efficiencies can be made and maintaining procedures that streamline management and record keeping. The accountability of poultry management system process helps a poultry farmer to document and demonstrate how the property of farming activities are being managed to minimize impacts on the environment and to optimize efficiency of natural resource use. Poultry Farm Management System provides a process enabling targeted investment to improve practices for better business, using a systematic approach to agricultural business management; farmers can identify and manage risks and opportunities arising from their farming activity. Management attention is focused on implementing recommended practices to address identified risks, and then reviewing progress made against plans and desired outcomes. Poultry management deals with the organization and operation of a farm with the objective of making a livelihood whilst dealing with global trade, traceability and consumer requirements, agricultural policies, environmental requirements, and the multi- functionality of agricultural enterprise as a whole. These functions include strategic, tactical and operational planning, implementation, and documentation, assessment and optimization of the performed work on the fields or on the poultry farms. To improve the execution of these functions, various management systems, database network structures and software architecture have been proposed to serve these purposes. The Use of Information Technology (IT) in Poultry Management Information System Information Technology is the application of computers and telecommunication equipment to store, retrieve, transmit and manipulate data. Between 1920 and 1970, the total inputs used in agriculture increased 20%, while total output increased 179%. (Duncan, Harshbarger, 2009) A few decades ago it was already noted that the output increase was clearly not just an increase in the amount of inputs used but rather the technology knowhow for efficient agricultural inputs utilization which poultry production is of a large percentage. Recently, it was concluded in their research that the use of information from poultry experts, biological innovations, harvesting and hashing machines, and mechanical technology used in poultry farming mainly caused the increase in productivity per poultry worker three folds between 1970 and the 2000s. Over the past 15 years however, poultry farmers started using computers and software systems to organize their financial data and keep track of their transactions with third parties (Batte, 2011) and also monitor their sales more effectively. In the Internet era, where information plays a key role in people’s lives, poultry farming is rapidly becoming a very data intensive industry where farmers need to collect and evaluate a huge amount of information from a diverse number of devices (e.g. sensors, farming machinery, meteorological sensors, etc.) in order to become more efficient in production and communicating appropriate information (Csótó, 2010). These efforts deal with a number of factors such as ecological footprint, product safety, labour welfare, nutritional responsibility, plants’ and animals’ health and welfare, economic responsibility and local market presence. The efforts cover almost all steps in the production chain concerning the daily agricultural tasks, the transactional activities for all involved stakeholders and the support of information transparency in the food chain. Information brought to farmers originates from systems installed by third parties such as meteorological stations or specialized infrastructure, e.g. sensors for measuring temperature, humidity and soil moisture. Farmers need to combine all these data effortlessly and take precise decisions to produce qualitative products, improve their income http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 96 and adhere to governmental regulations and principles. All this information should also be combined with the ‘‘farmer’s internal system of practical knowing and learning’’, building thus a real cognitive system (McCown, 2012). Nowadays, a number of proprietary solutions have been developed to help poultry farmers manage their farms in an effective way. More sophisticated systems track geographical areas, weather patterns and perform numerous advanced predictions on whether it will be conducive for a kind of breed to be ground in a particular geographical location and most of the latter mentioned systems, known as Poultry Farm Management Information Systems (Lewis, 2009), focus on specific tasks and use their own specifications to implement the functionality provided. Currently, these systems are slowly moving into the Internet era and are starting to use some of the well-established networking solutions to improve what they offer to the end users. However, it is widely accepted that the Internet faces a number of shortcomings, especially in handling vast numbers of networked devices (i.e., Internet of Things) or stakeholders. Moreover, there is still no standardized solution to enable a simple and cohesive interoperability among services and stakeholders. The Future Internet (FI) infrastructures are expected to handle these shortcomings. The aim is to propose a functional architecture of a poultry expert system (PES) utilizing Future Internet capabilities. Our goal was not to build a complete management system but rather to focus on those functionalities that can be improved with the use of the innovative FI’s capabilities. Using these capabilities the farmer should be able to perform a number of tasks that are not possible today (e.g., advertise his poultry products effortlessly, discover trustable stakeholders, information and services, combine functionalities from different management systems and services, cope automatically with unstable data network links, etc.). Poultry Expert System for Diagnosis of Broilers Broilers are defined as chickens of meat-type strains raised specifically for meat production. Based on data production from the Ministry of the Republic of Indonesia raised 3.76% from 2015 – 2016. But in reality the price of chicken is expensive, because the amount of market demand is more than the amount of production. Harvest failure due to chicken disease is on the increase because of the lack of information for the poultry farmers to manage and cultivate their poultry to mature for sale. Detecting diseases at early stage can enable to overcome and treat them appropriately. Identifying the treatment accurately depends on the method that is used in diagnosing the diseases. A Diagnosis expert system (DE) can help a great deal in identifying those diseases and describing methods of treatment to be carried out taking into account the user capability in order to deal and interact with expert system easily and clearly. This system has 25 symptoms and 6 diseases using certainty factor method to solve the problem of uncertainty. The result of the research is that Broiler Expert System has been successfully identifying diseases that can solve the problem with 90% accuracy (Setyohadi, 2018). Concept of Poultry Production and Expert System Poultry production as an aspect of livestock production is important to the economic and social development and biological needs of the people of any nation because it assists in alleviating food security, creates employment opportunities for the people who are engaged and creates incomes to the people who are engaged in the projects. It is a process that involves rearing of chicks from day one to the time they mature by using some farm inputs, capital, labour and entrepreneurial talent (Oladeebo 2010). Al-Hassan (2008) points out that, inefficiency in poultry production can result from socio-economic, demographic or environmental factors. However, some of the environmental/exogenous factors such as http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 97 weather, government policies among others are outside the scope or the control of the farmers, and hence their impacts cannot be considered as the causes of the farmers‟ inefficiency. In view of this, Oladeebo (2010) note that farm specific efficiency can be influenced by farmers‟ characteristics (Socioeconomic-factors) which impact on the managerial skills of the farmer. Such socioeconomic characteristics include: the age of the farmer, his/her level of education, number of years of poultry farming experience, access to credit and extension services, contacts and networks, farm size, gender, and engagement in other income generating activities other than farming activities. Coelli & Battese (2011) identified age and schooling (level of education) as factors influencing efficiency in poultry farming. The result of their study indicated that the younger farmers were found to be more efficient than their older counterparts. The situation was made worse by the devaluation of the currency combined with very poor government policy. Nhemachama & Hassan (2010) who also found out that farming experience enhanced a farmer’s knowledge and information and high skills in farming techniques and management, which improve the technical efficiency of the 16 farmer. Farming experience also enables a farmer to adapt to climatic change, new agricultural practices and ability to spread risk. Kaur, Sakhon, & Kingra (2010) conducted their study on technical efficiency of poultry production in Punjab state, India. They used stochastic frontier production to estimate the technical efficiency of poultry production and they found that the mean technical efficiency of poultry production as 87 per cent, 94 per cent, 86 per cent and 87 per cent in semi-hilly, central, south-western and Punjab state as a whole, respectively. The result of their model showed that the technical efficiencies are positively and significantly related to age, education and experience of a farmer. Kaur, sakhon (2010) found that age and the level of education were significant and positively related to productivity of the farmer. Younger and more educated farmers attained high levels of TE, AE and EE than their older and less educated counterparts. Ongundari & Ojo (2009) conducted a study on Economic Efficiency of Small Scale Food Crop Production in Nigeria using the stochastic frontier approach. The study revealed that farm size was one of the contributing factors of the economic inefficiency of the farmers. Small-sized farms exhibited high levels of technical inefficiency due to diseconomies of scale they experienced. The mean total cost of production was Naira 48,712.95 per hectare with a standard deviation of 43,358.81 Naira. The large variability in the standard deviation implied that the farmers included in the sample operated at different levels of farm sizes, which tended to affect their cost levels. According to Oji & Chukwuma (2012) carried a study on technical efficiency of small scale poultry egg production in Imo State of Nigeria and found out that farm size has a significant 17 positive effects on efficiency. They noted that farmers who were not operating at full capacity and would increase output by increasing the number of birds reared. The two researchers also noted that extension contacts and the farmers‟ level of education do have positive impacts on the farmer‟s efficiency. Furthermore, a farmer‟s access to credit also increases his efficiency’s ability. They noted that farmers who had access to credit were found to be more efficient than those who did not access credit. This could be due to the fact that those who accessed credit were able to increase their level of production and benefit from cost advantage that are associated with economies of large scale production. Availability of Market and poultry production Galeboem, Isaac & Mmattio (2009) notes that availability of markets and market information encourages farmers to produce goods that are demanded and hence their confidence that there exists a ready market. A market that is deficient in information and exhibits inconsistency is likely to be http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 98 attractive to the investors. Like any other business, poultry farmers also prefer to invest in poultry farming where there is adequate information exist. Makhura & Delgado (2012) found that distance to the market negatively influences both the decision to participate in markets and the proportion of output sold. Thus, the variable transport costs per unit of distance increases with the potential marketable load size. For farmers in very remote rural areas, geographic isolation through distance creates a wedge between farm gate and market prices. This leads to a shift from production of profitable but highly perishable commodities such as fruits and vegetables to relatively storable low-value cereals (Stifel & Minten, 2008) points out that poultry farmer also wish to invest in markets which are not faced with unfair and un-regulated competition and also in an environment of free trade barriers that have been implemented in other countries. For example, Kolare (2012) points out that in Botswana poultry marketing is highly disorganized. Despite the fact that the buying price of the major poultry input product, corn, is mainly determined by the poultry farmers where they pay low prices, the Botswana poultry industry has not experienced growth over time. This is because word of mouth is the most relied upon method of marketing of the 60% the poultry products. And buyers are the ones who dictate the price to pay for such products. This due to the imperfection of the market as a result of the Ugandan government not focused in establishing a system that can ensure a smooth flow of products from the source to the end-user, thus guaranteeing smooth utilization and higher profit margins for the poultry farming business. Kariuki (2010) In Kenya, points out that the poultry industry also suffers from poor organization and marketing due to little effort on the side of the Kenya government focus on the provision of information in order to facilitate the smooth flow of the poultry products from farmers to the consumers. Another problem that is experienced by poultry farmers in Kenya is the issue of low prices of poultry products in Uganda. This has caused poultry product buyers to illegally cross the border to Uganda to buy cheaper poultry products and then come and resell in Kenya, thus earn better returns. This has affected poultry farming in Kenya due to lack of market. Expert system for Decision Making The conventional decision support has predefined set of input data, after that they begin analysis. They precede the data, step by step as directed by algorithm, to reach conclusion. Here the algorithm plays important role since knowledge is represented in the form of Algorithm. If knowledge of problem got changed then Algorithms need to be changed or rebuilt. Also solving new problem in same domain needs to develop new system. Against this, human expert tends to follow cognitive approach rather than algorithm. They rely on extensive knowledge base (in their mind) which may contain facts, assertions, past mistakes, trial – by – error method. The machine equivalent human experts are expert systems. The expert system works with cognitive approach and stress the knowledge in knowledge base which is separate component. So that changes in knowledge do not change whole structure of expert system. Another advantage is reasoning capability. They can explain reasons for arriving at particular decision. Importance of Poultry Expert System The complexity of problems faced by poultry farmers are lack of information for in poultry farming, disease and parasite of the poultry products, feeds of the poultry products, location and water related problems, drug related problem of the poultry farm product. All the above listed poultry problem can be solved with the use of an expert system. Expert System are computer program that are different from conventional computer programs as they solve http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 99 problems by mimicking human reasoning process, relying on logic, belief, rules of thumb opinion and experience. In agriculture Expert System are capable of integrating the perspectives of individual disciplines such as plant pathology, entomology, horticulture and agricultural meteorology into a framework that best address the type of ad hoc decision making required of modern farmers. Expert system can be one of the most useful tools for accomplishing the task of providing growers with day to day integrated decision support needed to grow their crops.The studies reviewed under this section clearly indicates that, various software were used by the researchers to develop computer-based Expert System and used as an effective tool in various fields of agriculture. The above observations suggest the need to develop a user friendly computer based Expert System considering the flexibility, simplicity, nature of problem and familiarity of the software to the student researcher. Experiences in Using Poultry Expert System for Agricultural Development Rafea , (2014), stated that the application of Expert System generally falls under three classes, namely, Expert System proper, intelligent front-ends, and hybrid systems. An Expert System proper is a purely rule based system, relying on a sizable knowledge base. It is based on a qualitative, causal understanding of how things work. Such a system is more suitable under situation wherein not quantitative data are used. It is essentially conceptual and heuristic rule-based system. An intelligent front-end is a user-friendly interface to a software package, enables the user to interact with the computer using his/her terminology. It minimizes or avoids misuse of complex models by less experienced users. A hybrid system represents the integration of algorithmic techniques with Expert System concepts. Thammi & Rao (2009), Developed a microcomputer-based, graphics-oriented Expert System for use in the design of parallel terrace systems. It divides the design process into manageable activities: digitization of a contour map, input of field and machinery characteristics, definition of "watersheds" to be terraced, definition of the outlet system and waterway divides placement of conventional terraces and placement of parallel terraces based upon a key terrace. The system is able to make design suggestions based on accepted practices and the programmed knowledge of recognized terrace system design experts. Barnabas (2013) observed that an expert's knowledge is the central and key component of developing an Expert System. Furthermore, it is more difficult component. At the end the knowledge acquisition must be. Regarded as much as an art as a methodical and scientific procedure. One approach, however, that often seems to be ignored is the collection, integration and use of research results, while an expert should clearly build research results into their expertise, it is also possible to bypass the expert and use the published results in formulating rules in cases where the research provides a complete and logical answer. Ikenna, (2012), revealed that Expert Systems methodology has shown considerable promise as an information technology. However, limited knowledge of how current information technologies relate to the decision process impedes the adoption of Expert Systems. The significance of developing an economic theory of Expert Systems is substantiated with an empirical application investigating a soybean pest management decision process (SMARTSOY) based on experience with four insect pests causing damage to soya beans in the southeastern USA. SMARTSOY is combined with SOYGRO (soya bean crop growth simulation model). In Poultry Expert system, diseases such as Newcastle Disease or Ranikhet Disease, Mareks Disease, Infectious Bursal Disease, Infectious Bronchitis, Avian Influenza, Colibacillosis, Infectious Coryza, Fowl pox, Ascariasis, Coccidiosis, Gout are included in the key visual symptoms of the Health Adviser component. http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 100 III. METHODOLOGY Research methodology simply involves the procedures for collecting and analyzing information or data, necessary to define or solve the problem for which the research is embarked upon. The system design phase considers the methods of designing the new system and the actual steps involved, system requirement and flowcharts and programs that implement software. This chapter also focuses on the method of collecting data, the steps in the new system, design, user documentation, the tools used and possibly the file organization of the system. Fact Finding Methods Used The data used in the study were collected from two sources; primary and secondary sources. • Primary source: This involves the oral interviews conducted from various poultry farmers at their respective poultry farms. The difficulties they undergo in getting knowledge about their poultry farms are revealed. • Secondary sources: This includes the use of textbooks, dictionary, journals, newspapers and the internet for downloads of necessary materials required to aid in data collection. The Existing System Poultry farmers seeking for knowledge of species, new breeds, or knowledge of poultry diseases, their preventive measures, i.e. how they spread, how to immunize and provide medication, have to seek for experts who are not readily available in the rural community to teach them, and as a result, many poultry farms collapse which will lead to scarcity in poultry products. The proposed system will provide cost effective expert knowledge to both the rural and urban poultry farmers which will help them to overcome most of their challenges such as (high cost of obtaining knowledge, breakthrough of diseases etc.) Software Development Life Cycle (SDLC) The process of building computer software and information systems has been always dictated by different development methodologies. A software development methodology refers to the framework that is used to plan, manage, and control the process of developing an information system. A software development methodology is known as software development life cycle (SDLC). It is used in several engineering and industrial fields such as systems engineering, software engineering, mechanical engineering, computer science, computational sciences, and applied engineering Richard & Barry study as cited in (Bassil, 2012). IV. RESULTS AND DISCUSSION Implementation The prototype of software framework for developing knowledge base of agro based expert system is implemented using HTML for markup with CSS and bootstrap framework to make the design welcoming and responsive. The system is design with star UML and MySQL is custom to implement the design to store and retrieve records while PHP is used as programming language to connect to database, forward data and retrieve. XAMPP localhost server is used to test the framework software and NetBeans IDE is used in writing source codes. The computer system used is installed with windows 10 operating system. http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 101 Figure 1 : workflow chart of expert system for poultry management Figure 2 : Use Case Diagram for Poultry Expert System System Interface The page bellow is the homepage interface with links in form of click-buttons to other pages which enables a visitor to access information. Figure 3: poultry Expert System Interface http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 102 Figure 4 : login Interface for Poultry Expert System V. CONCLUSION The research was carried out to develop a software framework for knowledge base of agro based expert system easier and more effective. The framework and prototype has been developed using PHP, MySQL, CSS and JavaScript, Bootstrap framework in windows environment hosted on XAMPP local host. The following conclusion can be made about the framework: it’s efficient, fast, secure, user friendly, approachable, responsive, cost effective and web- based. The researchers recommended Poultry farmers to adopt and use the software framework because it can give knowledge on how to manage poultry expert system of different categories in poultry farming. VI. REFERENCES [1]. Adeyemo, A. B. (2013). An e-farming framework for sustainable Agricultural developmentin Nigeria. Jurnal of Internet and Information System, 3(1), pp 1-9. [2]. Al - hassan, S. (2008). Technical Efficiency of Rice Farmers in Northern Ghana. Research Paper 178. Nairobi: African Economic Research Consortium (AERC). [3]. Andino M., and Hasan, M. (2012).Avian influenza (H5N1) expert system using dempster- shafer theory. International Conference on Informatics for Development (ICID), pp.92- 93. Andino, M. and Mahmud, H. (2012).Poultry diseases warning system using dempster-shafer theory and web mapping. International Journal of Advanced Research in Artificial Intelligence, 3(5) pp 1-3. [4]. Balakrishnan, M. (2013). Development of an Expert System for Agricultural Commodities. The International Journal of Computer Science & Applications, 2(7), pp 74 - 90. [5]. Barnabas, A. A. (2013). An e-farming framework for sustainable agricultural development in Nigeria. Journal of Internet and Information System , 3(1), pp 1 - 9. [6]. Bassil, Y. (2012). A Simulation Model for the Waterfall Software Development Life Cycle. International Journal of Engineering & Technology (iJET), 2(5), pp 1-6. [7]. Butterfield, K. L. (2009). Report of the first annual meeting, Analysis of therural problem, America Farm Management Association, (AMES). America. 3(8), pp 4-8. [8]. Batte, M. (2011). Changing Computer use in Agriculture. Computers and Electronic in Agriculture, 1(8), pp 1-13. [9]. Coelli, T. and G. Battese, (2011). Identification of factors which influence the Technical inefficiency of Indian farmers. Australian Journal for Agricultural Economic development, 40(2) pp.103-128. [10]. Csótó, M. (2010). Information flow in Agriculture-through new channels for improved effectiveness. Agricultural Information, 2(6), pp 25-34. [11]. Duncan, M., Harshbarger,E. (2009). Agricultural Productivity. Trends and implications for the future economy, 3(4) pp 1-12. [12]. Galeboem, Isaac Mbaiwa and Mmatli .O, (2009). Market study on pig products in Botswana Research and development Division. Local Enterprise Authority Journal 3(2) Pp.11-15. [13]. Halimat, A., and Amosa, B. (2012). An Expert System for Management of Poultry Diseases. International Conference on Computer Technology and Science, 47(22), pp.113 – 117. http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 103 [14]. Ikenna, U. M. (2012). The issue of food feed competition and chicken for demand of food of animal origin, Assianjounal of poultry science 6(2) pp.31-43. [15]. Kariuki, W. (2010), Poultry Diseases in Kenya, Macmillan Publishers, Nairobi Kenya. [16]. Kaur, M., A. K. Mahal, M.K. Shekhon and H. S. Kingra (2010).Technical Efficiency of Wheat Production in Punjab: A Regional Analysis, “Agricultural Economic Research Review Vol.23 pp. 173-179. [17]. Kolare, O.B. (2012). Piggery section Annual Conference for the department of Animal Health and Production, National Veterinary Laboratory. Sebele, Gaborone, Gaborone, Botswana. vol5. pp.98-100 [18]. Kreitner, R. (2010): Houghton Miffin Boston, Management 4th edition.pp.11-15 [19]. Lewis, T. (2009). Evolution of Farm Management Information Systems. Computer Electronic Agriculture, vol. 1(23) pp 3-248. [20]. Makhura, M, Kirsten, J and Delgado, C, (2012). Transaction costs and smallholder participation in the maize market in the Northern Province of South Africa. Seventh Eastern and Southern Africa Regional Maize Conference, 2(4) pp.11– 15. [21]. Maryam, H. and Alaa E. A. Y.(2013). Animal knowledge based system in Egypt. International Journal of Computer Applications, 71(21), pp. 0975-8887. [22]. McCown, .R. (2012). Learning poultry sciences and human development. African journal of education research. 3(2), pp. 166-172. [23]. Meng-qiang, and MA Jin-su (2009) “Poultry disease diagnostic Computer Simulation Research” Project fund Henan Province Department of Education, high yield and efficiency of animal husbandry computer applications Animal Nutrition Optimization System". [24]. Mugga, R. (2009). Modern breeds find their way to Uganda. Journal for work Poultry, 25(2) pp.8- 12. [25]. Nhemachena .C. and R. Hassan (2010). Micro- level analysis of farmers ‟adaptation to climate Change in Southern Africa. IFPRI Discussion paper No.00714, International Food Policy Research Institute, Washington, D.C pp.24-35. [26]. Ogundari, K., Ojo S.O. and Ajibefun, I.A. (2009): Economies of Scale and Cost Efficiency in Small Scale Maize Production: Empirical Evidence from Nigeria Journal of Social Science, 13(2) pp.131-136. [27]. Oji Ukoha Obasi and Chukwuma, Augustine Anyanwu (2012. Technical Efficiency of small scale poultry–Egg production in Nigeria: Empirical study of poultry farmers in Imo state, Nigeria. Research Journal of poultry sciences,1(4) pp.16-22. [28]. Oladeebo O. Adebayo (2010), Socio-Economic Factors Affecting Poultry Farmers. Profitability Analysis of Poultry Production Imo State, 4(6) PP. 39-41 [29]. Oluby D.O. and Yerulu O.M. (2018). Design and implementation of an expert system for Diagnosis, treatment and management of poultry Diseases FULafia Journal of Science & Technology for expert system for Diagnosis, treatment and management of Poultry Diseases Federal College of Education (Technical), Asaba 4(2), pp.(15-20). [30]. Olarinde L.O. and F.A. Kuponiyi (2010), Resource Productivity among Poultry Farmers in Oyo State, Nigeria: Implications for Agricultural Development. J. Sustainable Development. pp. 20-26. [31]. Rafea, A. (2014). Expert System Applications: Agriculture. Central Laboratory for Agricultural Expert Systems, 3(1), pp. 13-28. [32]. Rafea, S. T. (2009). Agriculture: a key Actor in Boosting Economy of a Nation. International http://www.ijsrcseit.com/ Volume 6, Issue 1, January-February-2020 | http://ijsrcseit.com Dr. Oye, N. D. et al Int J Sci Res CSE & IT, January-February-2020 ; 6 (1) : 91-104 104 Research Journal of finance and economics, 24(91) pp8-15. [33]. Robinson, Y. (2018). International Journal of Agriculture and research development. Impact of Expert system in Agricultural development in Nigeria, 6(2), pp. 90-93. [34]. Thammi, R. D. and Sudhakar, R. B. (2011).An information technology enabled poultry expert system: perceptions of veterinarians and veterinary students. International Journal of [35]. Education and Development using Information and Communication Technology (IJEDICT), 2(2), pp. 16-20. [36]. Salamani .P. and Ahmadi Hojat (2010). Review of farm management information system. Journal of farm evolution and the development of product marketing. 7(4), pp 25-30. [37]. Stifel, .D. and Minten, B, (2008). Isolation and agricultural productivity. Journal for Agricultural Economics 3(9), pp. 1–15. [38]. Setyohadi, D.S.P (2018). An Expert System for Diagnosis of Broiler Diseases using Certainty Factor Journal of Physics Conference Series on poultry, 12(18), 2-6. [39]. Thammi, D., and B. Rao, S. ( 2009). An information technology enabled Poultry Expert System. International Journal of Education and Development using Information and Communication Technology, 2(2), pp. 100 - 107. [40]. Yelapure, S. J., and Kulkarni, R. V. (2012). Literature Review on Expert System in Agriculture. International Journal of Computer Science and Information Technologies, 3 (5), 86- 89. Cite this article as : Dr. Oye, N. D., Jemimah, N., "Expert System for Poultry Management ", International Journal of Scientific Research in Computer Science, Engineering and Information Technology (IJSRCSEIT), ISSN : 2456-3307, Volume 6 Issue 1, pp. 91-104, January- February 2020. Available at doi : https://doi.org/10.32628/CSEIT206118 Journal URL : http://ijsrcseit.com/CSEIT206118 http://www.ijsrcseit.com/ https://doi.org/10.32628/CSEIT206118 %20Journal%20URL%20:%20 %20Journal%20URL%20:%20 http://ijsrcseit.com/CSEIT206118 
work_s2a3uuwwqjcfnbzd6piss22kpm	----	Parallel electro-optical rule-based system for fast execution of expert systems Ahmed Louri and Jongwhoa Na The slow execution speed of current rule-based systems has restricted their application areas. Multiprocessor architectures have been proposed to overcome this limitation. However, as the number of processors in a multiprocessor system grows, so does the cost of communication between processors or between processor and memory units. The use of optics for a fast and parallel implementation of rule-based systems is proposed. The proposed optical system is hybrid in nature, using electronics for the user interface and optics for the rule-based inference engine. The proposed system uses two- dimensional planes as basic computational entities and is therefore able to provide concurrent rule processing. Furthermore, it provides highly efficient implementation of the basic operations needed in rule-based systems; namely, matching, selection, and rule firing. The execution speed of the proposed system is theoretically estimated and is shown to be potentially of orders of magnitude faster than current electronic systems. Key words: Expert system, electro-optical rule-based system, inference engine, optical interconnec- tion networks. 1. Introduction A key feature of artificial intelligence computing systems is the large amount of irregular communica- tions required between the large number of process- ing elements'- 3 (PE's). Unfortunately, current elec- tronics technology is unable to provide adequate bandwidth and connectivity for the large number of PE's involved. Consequently, current symbolic com- puting tasks, including rule-based systems (RBS's) require an excessive amount of time to produce results. Optics offers the possibility of eliminating the performance bottlenecks associated with commu- nications. Optics has the advantages of higher spa- tial and temporal bandwidth, no cross talk, larger fan-in and fan-out, and two-dimensional (2-D) data- processing capability.4 5 These advantages are use- ful in exploiting massive parallelism and in achieving higher throughput.6 Optical systems can provide not only adequate communication support, but also the necessary parallelism to perform the most fre- quently used and time-consuming operations in RBS's, namely, pattern matching and comparison opera- tions. The authors are with the Department of Electrical and Com- puter Engineering, University of Arizona, Tucson, Arizona 85721. Received 21 October 1991. 0003-6935/93/111863-13$05.00/0. © 1993 Optical Society of America. This has resulted in major research efforts for exploiting the use of optics in symbolic artificial intelligence computing. 7 -16 Warde and Kottas7 have proposed two different hybrid optical inference ma- chines. The matched-filter inference machine per- forms inferencing using a classical Vander-Lugt matched-filter system.8 The mapped template infer- ence machine matches a template representing the relationship between input and output data to pro- duce related output data with the optical correlo- graph system of Willshaw and Longuet-Higgins."1 Jau et al.9 used a semantic network as a knowledge- representation scheme. Jau showed that a simple query can be answered by using an optical matrix- multiplication technique. Schmidt and Cathey'0 pro- posed an optical mathematical resolution system, which can be used as an inferencing mechanism in an expert system. We propose a parallel electro-optical rule-based system (called EORBS), in which rules are used to represent knowledge and an inference engine is used for reasoning. To take advantage of optics proper- ties, we represent rules in 2-D space, and the infer- ence engine is logically transformed into an optical interconnection network that can carry out inferenc- ing optically and hence in a highly parallel fashion. Section 2 provides a brief background of RBS's. Section 3 describes fundamental concepts of the proposed EORBS. Section 4 discusses the implemen- 10 April 1993 / Vol. 32, No. 11 / APPLIED OPTICS 1863 tation of EORBS. Section 5 describes the operation of EORBS, and Section 6 provides projected perfor- mance data for the EORBS. Section 7 concludes. 2. Background An expert system can be defined as an intelligent system that can mimic some part of human intelli- gence. For example, MYCIN is a rule-based expert system that diagnoses bacterial infections.17 MYCIN uses 500 rules to diagnose approximately 100 causes of infections. An expert system is composed of (1) a RBS and a knowledge-acquisition facility, (2) an explanation facility, and (3) a user interface, as shown in Fig. 1.18 The RBS performs inferencing by using the knowledge base and the inference engine. The knowledge-acquisition facility and the user interface act as an interface unit between the RBS and the user. The explanation facility explains to the user how the results have been obtained. The knowledge base is composed of rules representing universal knowledge and facts that represent current knowl- edge. A rule conjoins condition elements C and action elements Ai with the following format: condition elements action elements if (C A C 2 A C3 ... A Cn) then (Al, A 2 , . . ., Am). The condition elements and action elements are represented by facts, which are the basic units of knowledge in an RBS. The inference engine uses the modus ponens theo- rem as an underlying principle. The modus ponens theorem states that if there is an axiom of the form El E2, and if there is another axiom of the form E, then E2 logically follows.1 9 The inference engine uses modus ponens as an underlying principle as follows: { current fact (A is TRUEE) A (if A, rule then B) J Expert System inferred fact - (B is TRUE) Fig. 1. Logical block diagram of an expert system. Thus, by using known facts and rules, the inference engine can produce new facts in a forward-chaining system, a backward-chaining system, or both.19 In the forward-chaining system the system tries to find final (goal) states by comparing the known facts, which describe the initial states, with the condition- specifying "if" parts of the rules. In the backward- chaining system the system tries to find initial hypoth- eses by comparing the known facts, which here describe the final (goal) states, with the action- specifying "then" parts of the rules. The basic operations of an RBS are as follows: (1) For the match operation (for each rule), deter- mine whether the condition part of the rule matches the current facts. If a rule satisfies the condition part, the rule is added to the conflict set. Otherwise, the rule is discarded. A conflict set is a set of triggered rules that have satisfied condition parts. (2) For the select operation, if the conflict set is empty, the inference engine stops inferencing and reports the failure to the user. If the conflict set is not empty and contains more than one rule, the inference engine selects one rule from the conflict set by applying one or more of the following conflict- resolution strategies.2 0 (a) For specificity ordering, if there are two rules and the condition part of one rule is more specific than the other, discard the latter. (b) For rule ordering, if there are many rules, choose the first rule. (c) For recency ordering, if there are many rules, choose the most (or least) recently used rule. (d) For data ordering, if there are many rules, choose the rule that has the most important condition elements. (3) For the act (rule-firing) operation, the infer- ence engine fires the selected rule by executing its action. Since the current facts are changed by the fired rule, it is important to check whether the changes agree with predefined goals. If the goals are satisfied, the inference engine stops inferencing, and it reports results to the user. Otherwise, the infer- ence engine must continue until the goal is satisfied or until there are no more matching rules. In an RBS, facts and rules can be represented as an inference net or as an AND/OR tree. Figure 2 shows an inference net and an AND/OR tree for the following set of rules (R1-R4): If (a A b) then i, If (a A c) thenj, If (d A i) then u, If (e Aj) then v, where A is a logical AND operator. Considering that the typical number of rules in an RBS is large and that each rule has several 1864 APPLIED OPTICS / Vol. 32, No. 11 / 10 April 1993 raw fact deducible fact Inference net deducible fact raw fact /< AND/OR tree Fig. 2. Graphical interpretation of a knowledge base. condition/action elements, the inference net can be extremely complicated. To traverse this increasing search space effectively, researchers have proposed parallel-processor systems that incorporate sophisti- cated parallel algorithms.2 ' For example, DAD02 (Ref. 1) is a tree-structured multiprocessor system that has 1023 PE's. Each PE is composed of an 8-bit microprocessor, 64 kbytes of random-access memory, and a switch. Owing to the large number of PE's, communication time is a dominating factor in DADO2.2 - Furthermore, since each rule of the knowledge base is unique and requires a different processing time, there is a synchronization problem between rules with a small number of condition elements and rules with a large number of condition elements. As an alternative to the electronic parallel-process- ing systems, optical systems such as the matched- filter inference engine (MFIE)7 and the matrix- multiplication inference engine (MMIE) 9 have been proposed. These systems can represent a large num- ber of related facts, called relations, owing to the multi-dimensionality of optical systems. One charac- teristics of these systems is that many facts within a relation can be processed concurrently (fact-level parallelism). In the MFIE a relation (set of facts) is compared with a condition element of a rule; there- fore, if a problem consists of r rules, condition elements per rule, and a condition elements per relation, the MFIE must execute r/a matched- filtering operations to answer a query. In the MMIE a relation is represented in a 2-D matrix. Then, this relation is loaded into an optical matrix-multiplica- tion unit so that multiple facts in a relation can be compared with a condition element of a rule. Thus, for a problem consisting of r rules, I condition ele- ments per rule, and a condition elements per relation, the MMIE requires rl/a matrix multiplications to generate a final matrix that can be used to answer a query. Note that the MFIE and MMIE require a multiple optical matched filter or a multiple optical matrix multiplier, respectively, to exploit a condition element-level parallelism and a rule-level parallelism available in the rule-based system. In the following sections we discuss an optical RBS architecture that can exploit a condition element-level parallelism and a rule-level parallelism as well as a fact-level parallel- ism. 3. Electro-Optical Rule-Based System: Overview The main objective of the EORBS is the exploitation of maximum parallelism in RBS's. This is achieved by using a 2-D representation of knowledge and multidimensional optical interconnects to speed up the match, select, and act operations of an RBS. EORBS offers a major advantage over previous propos- als (including optical ones): simultaneous process- ing of several rules as opposed to the sequential processing of rules. In order to process many rules in parallel, we have to ensure that multiple rule firing has no side effects. 2 2 EORBS guarantees this by employing a monotonic- reasoning methodology2 3 and by identifying a set of mutually exclusive rules. The monotonic-reasoning methodology ensures that rule-firing operations only augment the facts without deleting or changing any facts that can be sources of side effects. Therefore, once a fact is asserted, the fact remains true forever. Thus the number of true facts keeps growing until the end of an inference operation. To apply the monotonic-reasoning scheme, we have to restrict a variable to be bound to some value during compila- tion time so that the data presented at the run time can be represented by propositions. Consider the following examples F, and F 2 : 72 < 3.832, a b. For F, the value of the expression can be evaluated directly so that the host can assign FALSE to Fl. For F 2 , first, the variables a and b must be bound to some values. Then, the host can evaluate F 2 and assign TRUE or FALSE to F 2. Therefore, if we can collect all the data in advance and we can infer something out of the given knowledge, the monotonic-reasoning meth- odology can be applied. The application areas for the monotonic-reasoning methodology include theorem proving, problem-solving systems, diagnostics, and consultation problems. In order to evaluate rules in parallel, the host must maintain consistency. To maintain consistency, the host must compare a fact element with a group of 10 April 1993 / Vol. 32, No. 11 / APPLIED OPTICS 1865 related rules and facts. Thus the size of the group can be important in that the comparison time de- pends on the number of data in the group. One of the major characteristics of the rule-based program- ming environment is modularity and hierarchy.2 0 These stem from the fact that the structure of human knowledge can be modeled in a modular and hierarchi- cal structure such as a tree, in which each node represents a module that is composed of a small number of rules and facts. In this tree (from a high-level view point), checking whether an incoming fact belongs to some mutually exclusive group of facts or not implies that the incoming fact must be com- pared with a mutually exclusive group. Thus the fact is compared with a specific group that is small in size. Also, note that this checking operation is per- formed at the compilation time. Therefore the check- ing time for a given fact for the group should not cause any excessive overhead at run time. Using a monotonic-reasoning scheme is advanta- geous in that EORBS does not require conflict resolu- tion. Since EORBS explores all the possible paths in the search space without any side effect, it does not have to select which path to follow. An example of parallel rule firing can be found in the fuzzy RBS used in control.2 4 A. Knowledge Representation in the Host Computer In general, to run any high-level programming lan- guage on a computer, one must translate the source code into an object code that can be run by a central processing unit. Likewise, RBS's require the rules to be in human-readable form, which is also required for the translated code that is optimized for the hardware. In the RBS the human-readable rules act as an interface between the user and the computer. These rules are used in editing and debugging the knowledge base of a given problem. As a result, RBS's keep their own version of the translated knowl- edge base. As an example, MYCIN uses a rule com- piler for generating a decision tree that is used by the inference engine. For EORBS we developed a trans- lation algorithm that translates the knowledge base into a 2-D array called a proposition table (PT). Figure 3(b) shows a PT that results from the applica- tion of the translation algorithm introduced below to the knowledge base of Fig. 3(a). An entry of the PT is composed of a condition/action element and a proposition. The condition element slot can be divided further into a left-element slot, an operator slot, and a right-element slot. The left- element slot has the name of an object or variable, the right-element slot has a variable or some constant, and the operator slot contains the operation to be performed with the left-element slot and right- element slot. Also, the proposition slot can be di- vided further into two slots: one for the name (integer identification number) of the proposition and the other for the binary value of the proposition. Along with the PT, the host computer generates a condition array (CA) and an action array (AA), which R1. If (weight < 110 Ib) A (frame = small) A (gender= female) Then (relative-weight = normal) R2. If (weight < 110 lb) A (frame = large) A (gender = male) Then (relative-weight = under-weight) (a) Example rules (a) Condition / Action element Propsition Left-element Operator Rht-element s Vc weight < 110 (lb) F 001 T frame smell F 002 F gender female F 003 T relative-weight = normal F 004 F frame large F 005 F gender male F 006 T relative-weiht = underweight P007 T (b) 001 002 0031004 005 0061X7 007 __- _ (c) T FT F F T T I..... (d) 001 _001 002 .004 003 005 (e) 004 007 K Fig. 3. Representation of the rules and facts in the host computer: (a) example rules, (b) PT, (c) RFA, (d) DFA, (e) CA, (f) AA. are shown in Figs. 3(e) and 3(f), respectively. The CA and the AA represent the condition parts and action parts of all the rules in the knowledge base. The CA has a dimension of r x I and the AA has a dimension of r x m, where r is the number of rules, I is the number of condition elements per rule, and m is the number of action elements per rule. The host also generates two copies of the 2-D fact array [Figs. 3(c) and 3(d)]. In the first copy, called the reference fact array (RFA), each element of the array represents the name of the fact; therefore the RFA is a name template of the optical fact plane of EORBS. In the second copy, called the data fact array (DFA), each element of the array is the value of the element that is designated by the RFA; therefore the DFA is an actual data plane for EORBS. An 1866 APPLIED OPTICS / Vol. 32, No. 11 / 10 April 1993 __ ___ algorithm that interprets a human-readable knowl- edge base accessed by users to generate the PT, CA, AA, RFA, and DFA is described below. (1) Initialize the PT, RFA, DFA, CA, and AA. (2) For i = 1 to (the number of rules), find current rule. (3) Forj = 1 to (number of condition elements of the current rule), do the following: (a) Find the current condition element. (b) Find the symbol for the current condition element. (If the current condition element has al- ready used the symbol, then find the used symbol, else create a new symbol.) (4) With the symbol (current condition element), do the following: (a) Evaluate the value of the symbol. (Fill in a row of the PT.) (b) Calculate PT[symbol = (current element, symbol, symbol value). (c) Calculate CA[i, j] = symbol. (d) Calculatej = j + 1; go to step (3). (5) For k = 1 to (number of action elements of the current rule), do the following: (a) Find the current action element. (b) Find the symbol for the current action ele- ment. (If the current action element has already used the symbol, then find the used symbol or create a new symbol.) (6) With the symbol (current action element), do the following: (a) Evaluate the value of the symbol. (Fill in a row of the PT.) (b) Calculate PT[symbol] = (current element, symbol, symbol value). (c) Calculate AA[i, k] = symbol. (d) Calculate k = k + 1; go to step (5). (7) Calculate i = i + 1; go to step (2). (Find the next rule). (8) For I = 1 to (number of row of the RFA), do step (9). (9) For m = 1 to (number of column of the RFA), do step (10). (10) For n = 1 to (number of row of the PT), do the following: (a) Calculate RFA[1, m] = symbol slot of PT [n]. (b) Calculate DFA[1, m] = symbol-value slot of PT[n]. (c) Calculate m = m + 1 and n = n + 1; go to step propositional logic is used. In a propositional logic system, each fact, goal, condition, or action is repre- sented by a pixel in the corresponding 2-D plane. Each element of the plane is expressed in a proposi- tion that is position encoded in 2-D space, as shown in Fig. 4. The value of a fact F, for example, is represented by light being on/off at the specific location of the fact plane. Similar considerations take place for other planes. In the fact plane the host allocates a small portion of the fact plane for the goal facts. These goal facts occupy the last row of the fact plane and are used at the end of the inference operation. Thus the host computer must generate the RFA such that the last row of the fact plane has goal facts. If the number of goal facts exceeds one row, the host can add rows in the RFA to accommodate a large number of goal facts. The optimal size for the number of goal facts should be determined after a thorough simulation study. C. Description of Electro-optical Rule-Based System The overall role of EORBS is to infer new facts by using the current facts and rules and by comparing the inferred facts to the given goals. Figure 5 shows a block diagram of EORBS. It consists of an elec- tronic host computer, a goal-checking unit (GCU), and an inferencing unit (IU). EORBS receives the necessary facts and rules from the host computer and returns the processed data to the host. The host computer uses a 2-D fact plane to convert electronic data to optical data for optical inferencing. The fact plane is encoded as explained above. The host com- putes and sets the interconnection pattern between the facts and rules in the IU (explained below). The IU performs the match and the act operations. The match operation consists of comparing the cur- rent facts with the condition elements of the rules. During the act operation, the action parts of the rules with satisfied condition elements are executed. A detailed explanation of the match and the act opera- tions is presented in subsection 3.D. The role of the GCU is to determine whether or not there is any assertion of goal facts during the inferenc- ing process. If the output from the GCU indicates a match, EORBS stops inferencing, and the host starts its own operation with the processed results from EORBS; otherwise, the inference engine starts the inference operation. (11) Calculate = 1 +1; go to step (8). B. Knowledge Representation in Electro-optical Rule-Based System In EORBS, knowledge is represented in a 2-D form. Facts are represented in a 2-D plane called the fact plane. Rules have two parts, a condition plane to represent condition elements and an action plane to represent action elements. These planes represent the basic computational entity in EORBS. Also, to increase the utilization ratio of limited 2-D space, FI F21 F3F4 F5 Fact Plane Cli C2.1 Cr1 C1,2 C2,2 C r,2 C2,c .a A1 A2 Ar Condition Plane Action Plane Fig. 4. Representation of the rules and facts for an optical system: F, fact element; Cij, condition element; Ai, action ele- ment. 10 April 1993 / Vol. 32, No. 11 / APPLIED OPTICS 1867 * b * . : - *-. Crc -..... .--- Electrnic signal Optical signal Fig. 5. EORBS. D. Fact-to-Rule and Rule-to-Fact Mapping Performing the match operation means that the condition elements of rules are evaluated with the current values of facts. This is usually a time- consuming operation in conventional RBS's, since it is done sequentially. Here we introduce a new method for performing rule matching that can be carried out in parallel, owing to the parallel nature of optical interconnects. The method consists of first providing a mapping called fact-to-rule (F-R) mapping from the facts to the rules. This mapping specifies the interconnection patterns required from the fact plane to the condition plane. Next, the condition templates of each rule are logically AND'ed into a single element whose value indicates whether the corresponding rule is to be selected or not. For example, consider the following two rules R, and R2 and the general rule format Ri, given in sequence: if (Fl A F2 ), then (Flo); if (F1 A F3 ), then (F20 ); if(Cil A Ci, 2 A Ci, 3 A Ci,4), then (Ai). Cij represents a condition element in the ith row and jth column of the condition plane, and Ai represents an action element of rule Ri. To determine the required interconnection pat- tern, we first compare rule R, and R2 with the given rule format Ri. It can be easily seen that F is mapped into Cll and that F2 is mapped into C1,2, while C1,3 and C1 ,4 are not used in Rl. Also, compar- ing the next rule R2 with Ri, we find that F1 is mapped into C2,1 and that F3 is mapped into C2,2, while C2,3 and C2,4 are not used in R2 . Thus, to perform the match operation, the routing network routes facts F, to C1,1, F1 to C2 ,1, F2 to C1,2 and F3 to C2,2, respectively. An algorithm is described below that generates automat- ically a fact-to-rule mapping table (FRMT) from RFA and CA inputs. (1) For i = 1 to (number of rows on the CA), do step (2). (2) Forj 1 to (number of columns on the CA), do step (3). (3) If CA[i, j] X null, do the following: (a) Find the available slot in the row (content of CA[i, j]) of the FRMT (content of CA[i, j] = symbol identification number in integer). (b) Calculate FRMT(CA[i, j]) = (i, j) [save destination (i, J) at the FRMT]. (c) Calculatej = j + 1; go to (2). (d) Calculate i = i + 1; go to (1). Since rules are defined as the conjunction of condi- tion elements, we can set the unused elements (e.g., C1,3 and C1,4 in Rl) to a logical 1. Next, all the elements of each row of this condition plane are logically AND'ed together to form a one-dimensional (i-D) column vector representing the selected rules for the current iteration. This column vector repre- sents the action plane. In general, the size and the dimension of the interconnection network and the AND-gate array representing the action plane depend on the configuration of the condition plane and the number of action elements per rule. In this exam- ple, we assume that each rule has only one action, hence the action plane is one dimensional. In gen- eral, the action part of the rule may contain more than one action element. In this case the AND array and the action plane are both of dimensions r x m, where r is the number of rules and m is the number of action elements per rule. In a similar fashion, the act operation must assert the action elements of the triggered rules into the current fact plane to form a new fact plane. Again, this operation is carried out by mapping from the action plane to the fact plane. This mapping, called the rule-to-fact (R-F) mapping, asserts new facts from the newly generated action plane into the previous fact plane. Using the above two rules and the algo- rithm described below, we can simply find the re- quired interconnection patterns Al-Flo and A 2-F2 0 for the RFMT from the RFA and AA inputs. (1) For i = 1 to (number of elements on the AA), do (2). (2) If AA[i] ;• null, do the following: (a) Find the available slot in the row (content of AA[i]) of the RFMT (content of AA[i] = symbol identi- fication number in integer). (b) Calculate RFMT(AA[i]) = i ( save destina- tion i at the RFMT). (c) Calculate i = i + 1; go to 1. The new fact plane contains previous facts as well as newly inferred facts. These facts are then sent to the GCU for further processing. A detailed example of this mapping is described in Section 5. 4. Architecture of Electro-Optical Rule-Based System A detailed optical architecture for EORBS is shown in Fig. 6. As described in Section 3, the system consists of a GCU, an IU, and 2-D optical source (fact) and detector planes for input-output (I/O) interfacing with the host. A 2-D optical source such as an electrically addressed spatial light modulator (SLM) 1868 APPLIED OPTICS / Vol. 32, No. 11 / 10 April 1993 1/0 UN1T INFERENCING UNIT Fig. 6. Implementation of the optical RBS: n, number of facts; r, arrows, electrical signal; solid arrow, optical signal. can be used for the fact planes.2 5 For the electrically addressed SLM, a ferro-electric liquid-crystal SLM, an electrically addressed microchannel SLM, or a silicon/lead lanthanum zirconate titonate SLM can be used. Since EORBS must send results to an electronic host, an optical detector array is needed to convert optical signals into electronic signals. EORBS also needs 2-D optical logic-gate arrays to perform AND and OR logic functions. A possible candidate for the optical logic-gate array is the self- electro-optic-effect-device family of devices.2 6 In what follows, we examine the optical implementation of each unit needed in EORBS. A. Input-Output Unit The input-output unit is the interface unit between the electronic host and the optical rule-based system (ORBS). The input-output unit consists of an SLM, an optical OR-gate array 01, and a 2-D detector array. The SLM is used to load the input fact plane (DFA) to the ORBS. The OR-gate array 01 receives two input planes, an initial fact plane from the host and a processed fact plane from the ORBS. Thus, by superimposing the two planes, 01 can generate an updated input fact plane at every cycle for inference operations in the IU. The detector array is used to send the processed fact plane to the host. B. Inferencing Unit The IU is composed of an optical AND-gate array Al and two interconnection networks, IN, and IN 2. The interconnection network IN, performs the F-R mapping, and IN2 performs the R-F mapping, as described in subsection 3.D. IN, uses the fact plane number of rules; , number of condition elements per rule; dashed as an input and generates the condition plane. The AND-gate array A1 receives the condition plane from IN, and produces a column vector that is used as the action plane. This vector is sent to the IN2 for asserting the new facts into the fact plane for the next iteration and to the GCU for goal-checking purposes. Optical Implementation of Fact-to-Rule and Rule-to-Fact Mapping The optical implementation of the F-R mapping requires a 2-D to 2-D network since we are mapping the 2-D fact plane to the 2-D condition plane. The R-F mapping requires a 1-D to 2-D network that maps an element from the 1-D action plane to some elements of the 2-D fact plane. The required inter- connection patterns in the F-R/R-F mappings are any-to-any connections. This means that any ele- ment of the input plane should be able to access any element of the output plane. This can be achieved by several techniques.2 7 -3 3 A possible optical imple- mentation that uses space-invariant holograms is proposed below. Holographic Interconnection Network As an example, let us consider the following facts and rules: facts F 1, F2 , F3 , F4 ; rule 1 is, if (F1 A F2 ), then F 3; rule 2 is, if (F1 A F3 ), then F4 ; rule 3 is, if (F2 A F3), then F4 . 10 April 1993 / Vol. 32, No. 1 1 / APPLIED OPTICS 1869 HOST Also, assume that we have a 2 x 2 fact plane and a 3 x 3 condition plane. Figures 7(a), 7(b), and 7(c) describe the RFA, CA, and AA, respectively. Then, using the FRMT algorithm, we can obtain the FRMT shown in Fig. 7(d). The FRMT indicates the follow- ing interconnection patterns: (F1-C, 1)(F1-C2,1 ), (F2-C2,1)(F2-C3,1), (F3 -C2 ,2 )(F3 -C 3 ,2 ). The holographic-interconnection-network imple- mentation of this mapping is shown in Fig. 8. The major components of this interconnection network are hologram H1, an SLM, and demagnifying lens Li,. The hologram is space invariant, such that each facet acts as a one-to-nine beam splitter. Since all the connections are not necessary, an SLM is used after H1 to block the unnecessary connections. Lens Li, is used for imaging and demagnification purposes. It reduces the 6 x 6 intermediate plane to a 3 x 3 output plane conforming to the size of the condition plane. Figure 8 shows the permitted and the blocked connections as transparent and dark pixels, respec- tively, on the SLM. In general, the space-band- width product (SBWP) of H1 and Li, depend on the given facts and rules. With a i/; x i/; fact plane and a r x I condition plane, the SBWP of H1 is 0(rl) and that of Li, is O(nrl), where n is the number of facts, r is the number of rules, and is the number of condition elements. As described in subsection 3.D, the R-F mapping must compare the action elements with the facts. Using the same example as above and the RFMT algorithm, we generate the RFMT shown in Fig. 7(e). The RFMT indicates the following interconnection 01 02 3 01 03 - - 4 0 1 04102 3 - 4 (a) (b) (C) Source Destination (Fact Element) (Condition Element) 01 C(ll)I(2J) _ 02 Q2,1) , 03 Q3,2) _ (d) Fig. 7. RFMT. Source Destination (Action Element) (Fact Element) 01 3 02 4 03 4 (e) Example of the (a) RFA, (b) CA, (c) AA, (d) FRMT, and (e) patterns between the action plane and the fact plane: (Al-F3 ), (A2 -F4 ), (A3 -F4 )- Figure 9 depicts a holographic-interconnection- network implementation of IN2 that provides such a mapping. IN 2 is implemented by a hologram H 2 , a lens Li2, and an SLM. Hologram H 2 in IN2 is also a multifacet hologram and has the same size as the action plane. Therefore H2 contains three facets, and each facet acts as a one-to-four beam splitter, as shown in Fig. 9. As in IN1 , an SLM is needed to block unwanted connections. Lens L 2 is used to demagnify the image generated by H 2 and the SLM. The SBWP of hologram H 2 is 0(rm) and that of Li2 is 0(nrm), where m is the number of action elements per rule. Note that there are many different ways of achieving these mappings optically. C. Goal-Checking Unit The GCU consists of lens L1 , spatial filter F1, shutter S1, three beam splitters (BS,, BS2 , and BS 3), and detector D1 . At the output stage of the interconnec- tion network IN 2 , the fact plane is divided into two fact planes by beam splitter BS,. One is sent to shutter S1 and the other to mirror M3. The fact plane routed to M3 is filtered spatially by F1 so that only the goal fact portion of the fact plane arrives at lens element L1 . Then the goal fact plane is colli- mated by L1 upon detector D1 and by the control input to shutter S1. When the input to D1 is high, D1 signals to the host computer that the goal has been found. At the same time, the high control signal disables the IU from further processing by disabling S1. On a low signal, the low control signal to D1 enables S1, which in turn permits further inferencing in the IU. 5. Detailed Operation of Electro-optical Rule-Based System For an illustration of the operation of EORBS, con- sider the example of the family tree of Fig. 10(a). The current facts consist of two relationships: is- married-to and is-father-of, as shown in Fig. 10(b). The rules for this system are shown in Fig. 10(c). The goal of this example is to find who is the grandmother of whom (e.g., is-grandmother-of). Using the rules shown in Fig. 10(c) and the transla- tion algorithm presented in Subsection 3.A, the host computer generates the PT, RFA, CA, and AA [Figs. 10(e) and 10(f)]. Next, the DFA is generated using the RFA of Fig. 10(f). If a fact is currently known, then its corresponding location in the fact plane is set to 1, while others are set to 0. Note that, among the elements of the DFA, the host reserves the last row for the goal facts. The goal of this example is to find a grandmother. Therefore the host assigns the goal facts (F18 and F19 ) in the last row of the RFA. The resulting DFA is shown in the table of Fig. 11(a). Then, to provide the means for the F-R mapping and the R-F mapping described in Subsection 3.D, the 1870 APPLIED OPTICS / Vol. 32, No. 11 / 10 April 1993 Multiple Imaging system 2 x 2 input Hologram Hl- fact plane (each subhologram isa lto(3x3) fan-out element) Fl- F2 c '1.1 F3- C2,1.1 Fl F2- F3- F2- C3.1 F3- - Fl- czs ES-Fl- C2,2 E- F3- C2,2 I- Fl- F3- F2- . .2 F4- ca F2- C2,2 F4- l F2-C1.3 C,3 .F3 F4- C1.3 ZI,3 Fl- F2- F3- F4- c2.3 CS,3 Intermediate image SLM Fig. 8. Holographic interconnection network for F-R mapping. host computes the F-R/R-F mapping and sends it to the IU. Using the RFA and the CA of Fig. 10(f) as inputs, the FRMT algorithm computes the required interconnection patterns. For example, fact F is used as Cl and C2,1 of the condition plane. Thus EORBS must be able to connect F to C 1 and C2,1. The complete F-R mapping is shown in Fig. 12(a). To generate the R-F mapping, we use the AA and the RFA of Fig. 10(f) as inputs to the RFMT algorithm. For example, the first action element A1 of the action plane is used as the element Fo of the fact plane in Fig. 10(e). Hence we need a connection for the first action element A to the fact element Fo. In a similar fashion, the complete R-F mapping is gener- ated [Fig. 12(b)]. Once the DFA and the R-F/F-R mapping informa- tion is ready, this complete information is loaded into the optical fact plane and the interconnection net- works IN1 and IN2 . Then, EORBS starts its opera- tion. Interferencing unit (F-R mapping). Initially, the fact plane (DFA) is supplied by the host. However, in subsequent iteractions, the fact plane includes newly asserted facts from previous iterations by using the optical OR-gate array 01. In this example, each Imaging system 21 0 -- .J Ati Jpln action plane Multiple Imaging system m Al- F1 A - A2-Fl2 F1l _ Al- A l - F4F element of the DFA is transferred into a predeter- mined location of the condition plane by using the FRMT described in Fig. 12(a). Figure 11(b) shows the resulting condition plane. The rightmost two unused columns are set to a logical 1. IN1 produces the column vector of Fig. 11(c). In the column vector, a logical 1 in the ith location represents a triggered rule. Inferencing unit (R-F mapping). If there are rules to be fired, the IU executes the action part of the triggered rules by using the RFMT. The triggered rules are mapped onto the fact plane with IN2. In this example, each triggered rule represented in Fig. 11(c) is mapped onto the new fact plane of Fig. 11(d) with the RFMT shown in Fig. 12(b). Goal checking. The GCU receives the DFA from the IU. Upon receiving the DFA [see Fig. 11(d)], the goal fact portion of the DFA (shaded pixels of the DFA) is routed to detector D and shutter S by lens L1 . A high signal to D1 and S implies that a goal has been found, while a low signal implies that no goal has been found at this time. In this example, the control signal to D and S is low, which means that no goal fact has been asserted. This ends the first iteration of EORBS. Imaging system (demagnification) Hologram HI; (each subhologram isa: to(2x2) fan-out element) Itmeie Intermediate ~ image SLM Lens Li2 f' Fig. 9. Holographic interconnection network for R-F mapping. 2 Z x 2 output act plane 10 April 1993 / Vol. 32, No. 11 / APPLIED OPTICS 1871 Imaging system 21 Imaging system (demagnification) 3 3 x 3 output condition plane Lens Lil DAVID ba-maied-to ANN BOB Bs-mnuied-to BEITH JOHN ia.-mird-to JEAN DAVID la-father-of BETH DAVID la-fathe-of JOHN BOB ia-father-of TOM BOB ia-father-of FRED JOHN la-father-of MARY JOHN la-father-f JACK (b) 11I a1 0 0 I11 1 1 0 0 0 0 (a) 1 0 0 1 1 1 1 1 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 01 (b) 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 (C) 0 0 0 O O O Io I i I 111 1 111 I I 1 000 (d) Fig. 11. First run of EORBS: (a) fact plane (DFA with initial facts), (b) output of IN 1 (CA), (c) output of AND-gate array of the IU (AA), (d) new fact plane (DFA at end of first iteration). - did- e U - Popitio Lc Sonditindmcnu Re a V. DAVID is-eerried-to ANN 01 T ROR i -nnien RETHL 0w .T lOHN is-insrted-to JEAN 03 T DAVID iu-father-of BETH 04 T DAVID i-father-of JOHN 05 T BOB is-father-of TOM 06 T BOB i-father-of FRED 0`7 T JDHN i-father-of MAR 08 T JOHN i-father-of JACK 09 T ANN is-mther-of BETH 10 It ANN i-mother-of JOHN 1 BETH is-mother-of TOM 12 F BETH is-oher-of FRED 13 JEAN is-other-of MARY 14 F JEAN is-mnother-of JACK 15 F DAVID is-grdfather-of MARY 16 F DAVID is-grandfather-of JACK 17 F ANN is-grandmother-of TOM 18 F ANN is-gndmother-of FRED 19 F (e) 01 04-- JO 02 06 -- 02 - 13 l01~o 020 103 -1 4. 05 061; 070 0310 rq 1 09 -L I 1 2 1051 08 - - 16 13 14 1 - 51 o4 -1 1 7 1 6 17 - I lxl1 1 8 Refrne Fart Array Condition Ary Acdon Array (f) Fig. 10. Rules and facts for the family tree of David and Ann: (a) diagram of family-tree example, (b) facts known to users, (c) rules from users, (d) translated rules, (e) PT, (f) rules and facts in table form. Second iteration. In the second iteration, the DFA [Fig. 11(d)] is combined with the original facts [Fig. 11(a)], which are given by the host. As in the first iteration, the new DFA is routed to the IU to perform F-R mapping. Then, the results of the F-R mapping of Fig. 13(b) are AND'ed row-wise. The result is shown in the column vector of Fig. 13(c). Next, using the AA, the IU performs an R-F mapping and produces the DFA shown in Fig. 13(d). The GCU routes the goal facts (shaded pixels) of the DFA [Fig. 13(d)] to detector D1. Since there is one pixel (F18) that has a value of 1, the GCU sends a high Source Desination (Fact Element) (Condition Element) 01 C(,1) C(2,1) - 02 C(3,1) C(4,1) - 03 C(5,1) C(6,1) - 04 C(1,2) _ _ 05 C(2,2) C(7,1) C(8,1) 06 C(3,2) - _ 07 C(4,2) _ _ 08 C(5,2) C(7,2) 09 C(6,2) C(8,2) - 10 C(9,1) C(10,1) _ 11 C(9,2) - _ 12 C(10,2) - (a) Source Destination (Action Element) (Fact Element) 01 10 02 11 03 12 04 13 05 14 06 15 07 16 08 17 09 18 10 19 (b) Fig. 12. F-R mapping and R-F mapping: RFMT for family tree of David and Ann. (a) FRMT and (b) 1872 APPLIED OPTICS / Vol. 32, No. 11 / 10 April 1993 DAVITAN BOTBii JOH T 1 N TOM FRED MARY JACK (a) RI IF (DAVID is-manied-to ANN) AND (DAVID ia-father-of BBTIi) THEN (ANN is-.,Iher-of BETH). R2 IP (DAVIDis-manied-to ANN) AND (DAVIDia-father-of JOiN) THEN (ANN ia-mather-of JOHN) R3F (BOBis-marid-toBElH)AND(BOBis-fahr-ofTOM) THEN (BETH is-mother-of TOM). R4 IF (BOB i-matted-to BETH) AND (BOB is-fh-of FRED) THEN (BETH l-moher-of FRED). R5: IP (OHN is-maried-to JEAN) AND (JOiN ia-fhr-of MARY) THEN (JEAN is-moahe-of MARY). R6 IF (JOHN a-manelto JAN) AND (JON la-fther-of JACK) TIHEN (JEAN u-mot-of JACK). R7 IF (DAVID is-faof JOHN) AND (JOHN la-faher-of MARY) THEN (DAVID a-g rrof MARY). R8: IP (DAVIDi-fh-of JOHN) AND (JOHN i-faher-of JACK) THEN (DAVID tdfof JACK). R9 IF (ANNis-oth-of BETH) AND(BfiTHia-moher-of TOM) THEN (ANN a TOM). RIO: IF (ANN is-othr-of BETH) AND (BETHia-mnher-of FRED) THEN (ANN la-g- hoh-a FRED). (C) RI: IFAI ANDBI TilENCI. R2: IF Al AND B2 TIIEN C2. R3: IF A2 AND B3 TIEN C3. R4: IF A2 AND B4 THEN C4 R5: IF A3 AND B5 THEN C5 R6: A3 AND B6 THEN C6. R7: IF B2 AND B5 THEN DI. R8 IF B2 AND B6 THEN D2. R9 I Cl AND C3 THEN El R10: IF C AND C4 TIIENE2. (d) 1 1 1 _ 1 1 1 1 1 _ _ O O (a) 1 1 1 1 0 1111 II 1 I 1 1 1 1 1 1 1 1111 ll i1 1 1 1 111 ll l 1 1. 1 1 1 1 .1 I (b) 1 1 1 1 1 1 1 1 1 1 (C) 0 0 0 1 1 1221o 0 1 0 I 1 (d) 0 1 0 11 Fig. 13. Same as Fig. 11 but for second run. control signal to the host and terminates the opera- tion of EORBS by disabling shutter S1. As soon as the host receives the high control signal (goal-found signal), it reads in the last fact plane and converts it to electronic signals by using the RFA shown in Fig. 10(f) to answer the query. 6. Performance Analysis of Electro-optical Rule-Based System Here we estimate the theoretical performance of the proposed architecture by estimating its execution speed and by comparing it with that of conventional RBS's. In what follows we use the following terms and assumptions for estimating the execution speed: (1) We assume that the response time of optical gate arrays used for logic functions (e.g., optical AND, OR, and optical shutter) is Tr. (2) The setup time of the SLM of size n is Ts. (3) The processing time of transferring elements of one DFA to another is Tp. (4) The access time of the detector plane of the I/O unit is Td. (5) The propagation delay of optical passive de- vices such as lenses, mirrors, beam splitters, and holograms is negligible. (6) We assume that the number of elements in the fact plane is n and that the number of condition elements per rule is 1. (7) We ignore the reconfiguration time of IN1 and IN2 , since these are reconfigured only at the start of each query. While solving a query, the interconnec- tion networks can be considered fixed. A. Execution Speed The execution speed of EORBS can be thought of as being divided into three parts: initialization time Tinit, execution time Tex, and output fetch time Tfetch. In order to estimate the execution time, we consider the longest signal path through EORBS. To complete one iteration, the optical signal must go through the optical OR-gate array (Of), the IU (one AND-gate array and two SLM's), and one shutter S1. Such a path contains five active devices that require switching; therefore the total time required for one iteration is 5Tr. The initialization time includes the setup times for the fact plane and the SLM's in the interconnection networks IN1 and IN 2 . However, these units can be accessed simultaneously, and hence Tinit becomes T, (the setup time for a single SLM). The output fetch time includes the time to read in the detector of the GCU. Thus Tfetch is equal to Td. The data for the input and output planes of EORBS can be formatted in a 2-D array data structure during the knowledge-base compilation time. Since the com- piled rules and facts are used throughout the query processing, the sum of Tinit and Tfetch becomes the actual time for I/O operations required between the host computer and EORBS. Therefore the overall response time of EORBS, Tresponse, becomes Tresponse = Tinit + Tex + Tfetch = Ts + x x 5Trx + Td, (1) where x represents the number of iterations required to solve a given query. Here we evaluate the number of rules that can be processed in EORBS. Given the number of condi- tion elements of a rule, the number of rules that can be represented in EORBS becomes n/. Since the cycle time of EORBS is 5Tr, the maximum number of rules R that can be processed per second is n R = (2) 5lTr However, if the size of the problem (the number of rules) is greater than the size of the condition plane of EORBS, we must partition the problem into a set of subproblems whose size fits into the condition plane of EORBS. In this case the cycle time must include the time taken to read in the DFA from EORBS (detector access time Td), the time taken to relocate some elements of the DFA into a new DFA' (host processing time Tp), and the time to load the input DFA' to the I/O unit of EORBS (SLM setup time T,). In this study, we assume that the SBWP of the space-invariant hologram of the interconnection net- work IN, is 0(rl). With this hologram, any fact element in a DFA can access any location in the condition plane. However, there is a high price to pay for such flexibility, namely, a large fan-out factor. This flexibility can be compromised without degrad- ing the performance of EORBS. One method is limiting the fan-out rl to a small integer a. This is possible because a single fact element is not used as the condition element of all rules in practical RBS's. As an example, the sample RBS used here requires a maximum fan-out of only 3. Even if there exists an RBS that has a fact F that is greater than the permitted fan-out a, we can easily duplicate F into as many copies as necessary to cover the fan-out require- ment. Limiting the fan-out implies that the number of pixels used of the SLM of the interconnection net- works, IN1 and IN 2, can be reduced. That is, given n, 10 April 1993 / Vol. 32, No. 11 / APPLIED OPTICS 1873 I I 1 1 1 the number of facts in the fact plane, and given rl, the number of condition elements in the condition plane, we must use an SLM with a SBWP of nrl in IN1 . Also, note that the maximum number of rules and facts that can be represented in EORBS is limited by the product nrl. Therefore, if we can reduce the required SBWP of the SLM, we can increase the number of rules and facts that can be represented in EORBS by using the unused parts of the SLM. To evaluate R, we assume y represents the ratio of the area that the hologram with the reduced fan-out can cover to the area of the SLM of the interconnec- tion networks. Then, the number of rules that can be represented on EORBS becomes n /yl. Considering this factor, if the size of the problem is greater than the size of EORBS, we see that the maximum number of rules that can be processed per second becomes n R yl(5T + Ts + Tp + Td) (3) As an example, if we assume n = 103, = 5, y = 0.1, Tr = 10-6 s, and T, = Tp = Td = 10-3 s, EORBS can execute 6.65 x 105 rules. In order to better appre- ciate this speed advantage of EORBS, we compare the execution speed of EORBS with the electronic systems. The DAD02 multiprocessor system is estimated to process 70 rules/s.3 The NON-VON is a multiproces- sor system that is composed of 32 powerful large processing elements and 16,000 small processing elements, and it is estimated to process 850 rules/s. 3 4 With processing capability of 6.65 x 105 rules/s, EORBS is expected to achieve a system throughput far better than any electronic RBS can achieve. B. Discussion The performance of EORBS is primary limited by the availability of optical devices with large SBWP and fast response time. The large SBWP permits the optical RBS to represent an increased number of rules, and the fast response time decreases the time taken for a single iteration in the optical RBS. The limited fan-out concept is introduced to increase the number of rules that EORBS can represent with a given SBWP of an SLM. Along with these require- ments, the optical power is also an important problem owing to the fan-out hologram used in the interconnec- tion network. Because of the beam-spreading opera- tion of the fan-out hologram, the fan-out hologram is the major power-consumption factor. However, the limited fan-out concept can also reduce the problem incurred by the beam-spreading operation by control- ling the number of necessary output destinations. The limited fan-out can also relieve the device require- ments. For example, if we reduce the required tar- get area of a beam, the beam-uniformity condition is greatly relieved. Also, since a small number of pixels of the SLM are routed to a pixel of the condition plane (IN,), the contrast ratio of the SLM can be decreased accordingly. 7. Conclusions While rule-based programming and RBS's are popu- lar and advantageous for solving problems in the artificial intelligence domain, particularly in expert systems, their slow execution speed has limited their extensive use. Conventional electronic parallel ma- chines contain fast processors, but have limited com- munication bandwidth, which is critical in parallel processing. Optics can provide the large communica- tion bandwidths and faster execution speeds required for RBS's. We have explored an electro-optical rule-based sys- tem architecture. The architecture exploits the nat- ural parallelism of optics and the advantages of optical interconnects. A distinctive feature of EORBS is the concurrent rule execution, a feature absent in electronic RBS's. In addition, EORBS provides an efficient implementation of the basic operations required for a RBS; namely, selection, matching, and rule firing. The concurrent rule exe- cution nature of EORBS, combined with the speed with which the basic operations are implemented in EORBS, can potentially result in a hybrid RBS with significant performance improvements over any exist- ing RBS's. This research was supported by National Science Foundation grant MIP-9113688. References 1. S. J. Stolfo and D. P. Miranker, "DADO: a parallel processor for expert systems," in Proceedings of the International Confer- ence on Parallel Processing, R. M. Keller, ed. (Institute of Electrical and Electronics Engineers, New York, 1984), pp. 74-82. 2. S. J. Stolfo, "Initial performance of the DADO2 prototype," Computer 20, 75-83 (1987). 3. A. Gupta, "Implementing OPS5 production system on DADO," in Proceedings of the International Conference on Parallel Processing, R. M. Keller, ed. (Institute of Electrical and Electronics Engineers, New York, 1984), pp. 83-91. 4. A. Louri, "3-D optical architecture and data-parallel algo- rithms for massively parallel computing," IEEE Micro 11, 25-84 (1991). 5. P. B. Berra and A. Ghafoor, "The impact of optics on data and knowledge base systems," IEEE Trans. Knowl. Data Eng. 1, 111-131 (1989). 6. A. Guha and M. W. Derstine, "Designing massively parallel optical computers: a case study," Appl. Opt. 29, 2187-2200 (1990). 7. C. Warde and J. Kottas, "Hybrid optical inference machine: architectural considerations," Appl. Opt. 25, 940-947 (1986). 8. A. B. VanderLugt, "Signal detection by complex spatial filtering," IEEE Trans. Inf. Theory IT-10, 2 (1964). 9. J. Y. Jau, F. Kiamilev, Y. Fainman and S. H. Lee, "Optical expert system based on matrix-algebra formulation," Appl. Opt. 27,5170-5175 (1988). 10. R. A. Schmidt and T. Cathey, "Optical implementations of mathematical resolution," Appl. Opt. 26, 1852-1858 (1987). 11. D. J. Willshaw and H. C. Longuet-Higgins, "Associative mem- ory models," Machine Intell. 5, 351 (1970). 12. G. Eichmann and H. J. Caulfield, "Optical learning (inference) machines," Appl. Opt. 24, 2051-2054 (1985). 13. H. J. Caulfield, "Optical inference machines," Opt. Commun. 55,259-260 (1985). 1874 APPLIED OPTICS / Vol. 32, No. 11 / 10 April 1993 14. H. H. Szu and H. J. Caulfield, "Optical expert systems," Appl. Opt. 26, 1943-1947 (1987). 15. A. D. McAuley, "Real-time optical expert systems," Appl. Opt. 26, 1927-1934(1987). 16. D. Casasent and E. Botha, "Optical correlator production system neural net," Appl. Opt. 31, 1030-1040 (1992). 17. F. Hayes-Roth, "The knowledge-based expert system: a tutorial," Computer 17, 11-28 (1984). 18. P. Harmon and D. King, Expert Systems: Artificial Intelli- gence in Business (Wiley, New York, 1985), Sec. 1. 19. P. H. Winston, Artificial Intelligence (Addison-Wesley, Read- ing, Mass., 1984), Chap. 6. 20. L. Brownston, R. Farrell, E. Kant, and N. Martin, Program- ming Expert Systems in OPS5 (Addison-Wesley, Reading, Mass., 1985), Chap. 1. 21. K. Hwang and D. Degroot, Parallel Processing for Supercom- puters and Artificial Intelligence (McGraw-Hill, New York, 1989), Chap. 7. 22. G. E. Blelloch, "CIS: a massively concurrent rule-based system," in Fifth National Conference on Artificial Intelli- gence, T. Kehler and S. Rosenschein, eds. (American Associa- tion for Artificial Intelligence, Philadelphia, Pa., 1986), pp. 735-741. 23. E. Rich, Artificial Intelligence (McGraw-Hill, New York, 1983), Chap. 3. 24. L. J. Huang and M. Tomizuka, "A self-paced fuzzy tracking controller for two-dimensional motion control," IEEE Trans. System. Man and Cybern. 20, 1115-1124 (1990). 25. J. A. Neff, R. A. Athale, and S. H. Lee, "Two-dimensional spatial light modulators: a tutorial," Proc. IEEE 78, 826- 855 (1990). 26. A. L. Lentine, H. S. Hinton, D. A. B. Miller, J. E. Henry, J. E. Cunningham, and L. M. F. Chirovsky, "Symmetric self- electrooptic effect device: optical set-reset latch differential logic gate, and differential modulator/detector," IEEE J. Quantum Electron. 25, 1928-1936 (1989). 27. B. K. Jenkins, P. Chavel, R. Forchheimer, A. A. Sawchuk, and T. C. Strand, "Architectural implications of a digital optical processor," Appl. Opt. 23, 3465-3474 (1984). 28. J. Taboury, J. M. Wang, P. Chavel, F. Devos, and P. Garda, "Optical cellular processor architecture. 1: Principles," Appl. Opt. 27, 1642-1650 (1990). 29. P. Yeh, A. E. T. Chiou, and J. Hong, "Optical interconnection using photorefractive dynamic holograms," Appl. Opt. 27, 2093-2095 (1988). 30. S. D. Smith, A. C. Walker, M. R. Taghizadeh, I. R. Redmond, and B. Robertson, "The optical wiring of photonic digital processing array for optical computing," in Optical Computing '88, P. Chavel, J. W. Goodman, and G. Roblin, eds., Proc. Soc. Photo-Opt. Instrum. Eng. 963, 414-418 (1988). 31. A. Hartmann and S. Redfield, "Design sketches for optical crossbar switches intended for large-scale parallel processing applications," Opt. Eng. 28, 315-327 (1989). 32. A. A. Sawchuk, B. K. Jenkins, C. S. Raghavendra, and A. Varma, "Optical crossbar networks," Computer 20, 50-60 (1987). 33. A. A. Sawchuk and B. K. Jenkins, "Dynamic optical intercon- nections for parallel processors," in Optical Computing, J. A. Neff, ed., Proc. Soc. Photo-Opt. Instrum. Eng. 625, 143-153 (1986). 34. B. W. Wah and G. J. Li, "Design issues of multiprocessors for artificial intelligence," inParallelProcessingforSupercomput- ers, K. Hwang and D. Degroot, eds. (McGraw-Hill, New York, 1989), Chap. 4, pp. 107-159. 35. S. K. Case, P. R. Haugen, and 0. J. Lokberg, "Multifacet holographic optical elements for wave front transformations," Appl. Opt. 20, 2670-2675 (1981). 10 April 1993 / Vol. 32, No. 11 / APPLIED OPTICS 1875 
work_s6wqhsfem5an7fmdkfetereenu	----	Carnegie Mellon University research repository - Browse Browse Discover research from Carnegie Mellon University Follow AllCategoriesgroupsSEARCH Hide footerAboutHow to DepositPreparing DataDMP ToolDeposit SupportContactDeposit AgreementTermsPrivacyToolsFAQsDisclaimerSitemap figshare. credit for all your research. 
work_s7h4q3jntjhiferlkk4f67aqoq	----	doi:10.1016/j.eswa.2005.10.011 Modeling consumer acceptance probabilities L.C. Thomas *, Ki Mun Jung b , Steve D. Thomas a , Y. Wu a a School of Management, University of Southampton, Southampton SO17 1BJ, UK b Department of Informational Statistics, Kyungsung University, Busan 608-736, South Korea Abstract This paper investigates how to estimate the likelihood of a customer accepting a loan offer as a function of the offer parameters and how to choose the optimal set of parameters for the offer to the applicant in real time. There is no publicly available data set on whether customers accept the offer of a financial product, whose features are changing from offer to offer. Thus, we develop our own data set using a fantasy student current account. In this paper, we suggest three approaches to determine the probability that an applicant with characteristics will accept offer characteristics using the fantasy student current account data. Firstly, a logistic regression model is applied to obtain the acceptance probability. Secondly, linear programming is adapted to obtain the acceptance probability model in the case where there is a dominant offer characteristic, whose attractiveness increases (or decreases) monotonically as the characteristic’s value increases. Finally, an accelerated life model is applied to obtain the probability of acceptance in the case where there is a dominant offer characteristic. q 2005 Elsevier Ltd. All rights reserved. Keywords: Student bank account; Acceptance probability; Coarse classifying; Logistic regression model; Linear programming; Accelerated life model 1. Introduction Forecasting financial risk in consumer lending has over the last thirty years become a major growth areas (Rosenberg & Gleit, 1999; Thomas, 2000; Thomas, Edelman, & Crook, 2002). The main approaches are credit scoring and behavioural scoring which are based on statistical or operational research methods of classification. The statistical methods include discriminant analysis, logistic regression, classification trees and survival analysis, while the operational research techniques include linear programming. Discriminant analysis (which is equivalent to linear regression) was proposed by Fisher (1936) as a discrimination and classification tool and was one of the first methods applied to building credit scoring models by discriminating between those loans which in the past had defaulted and those which had not defaulted. Logistic regression is a related statistical modeling method which is now widely used and Wiginton (1980) was one of the first to describe the results of using it in credit scoring. The application of survival analysis for building credit scoring models was introduced by Narain (2004) and developed further by Stepanova & Thomas (2002). Mangasarian (1965) was the first to recognize that linear programming could also be used in classification problems 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.10.011 * Corresponding author. where there are two groups (defaulters and non-defaulters in this case), and this work was extended by Freed & Glover (1981); Hand (1981). In all cases, the approach is to take a sample of previous applicants and to find the best way of separating those who default from those who do not default. In the same way as assessing default risk, statistical and operational research method can be used to determine the probability that an applicant with certain characteristics will accept the offer of a loan (or respond to some direct marketing literature). Here, the two groups are those who take up the offer and those who do not or in the case of direct marketing those who respond to the mailing and those who do not. Response scorecards are widely used in marketing but the acceptance scorecards have had less visibility. However, with the lenders slowly changing their objectives from minimising the risk of default to maximizing the profitability of the loans they offer, estimating this probability of acceptance is becoming more important. Currently, there are two significant changes in the consumer lending process which have added to the need to estimate probability of acceptance but for a more complex problem. The first is that instead of a lender having a specific product to offer— a credit card with a given interest rate for example,— increasingly they can offer a generic product but where the features of the product can vary from consumer to consumer. It has always been the case that the overdraft limit varied but with the advent of risk based pricing, the interest rate offered is also beginning to be varied from consumer to consumer. Similarly in Expert Systems with Applications 30 (2006) 499–506 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa L.C. Thomas et al. / Expert Systems with Applications 30 (2006) 499–506500 credit cards, other features that can be varied would be an annual fee, an initial discount offer, a points scheme and free travel insurance. Other financial products also have variable features. For bank accounts, these would be the overdraft limit, an ATM card, interest paid when in credit, no fee foreign exchange while for mortgages one can vary the loan amount, the interest rate, the connection if any to the official interest rate, a initial discount on the rate, and whether there is cashback. Customer relationship management implies that the lender wants to tailor these features to the customers’ requirements and so make it more likely that they will accept and stay with the product. The second change is that the newer communication and marketing channels, which can be used for applying for loans,—the internet and the telephone for example—allow for the features of the offer to be adjusted during application process. A lender offers the generic product, the credit card say, but during the application process when obtaining the applicants’ information in order to check their default risk, the lender has the opportunity to tailor the features of the product so as to make the applicant more likely to accept the product, if they are offered it. This contrasts with the traditional form filling approach where there is really no interaction in the application process between potential borrower and lender. This means it is now important for a lender to be able to assess the likelihoods of a particular consumer accepting each of the variants of the particular product that he is offering. Moreover, this has to be done in real time during the application process so that the most ‘suitable’, product can be offered at the end of the process. One may ask why does the lender not show all the variants of the product to the consumer and let him choose which he wants. In many retail environments, e.g. clothes, furniture, this is a sensible strategy but in the financial sector there are three reasons why this is not the way things are developing. Firstly, customer are likely to chose the most ‘unprofitable’ product for the lender. Secondly, the prestige of a financial organisaton is lowered if it implies it does not know what is the best product for its customer. Thirdly, with so many combinations, the customers will be spoiled for choice, which unsettles some consumers especially in the financial area. So in this paper, we investigate how to estimate the likelihood of a customer accepting a loan offer as a function of the offer features. Then the lender can decide which acceptable combination of features is optimal to offer the applicant, and both these calculations must be done during the application process. Here, an acceptable combination may be one that is profitable to the lender and the optimality is judged in terms of maximizing the probability of the offer being accepted. Alternatively, an acceptable combination might be one that meets the customer’s minimum requirement and the optimality is to maximize the lender’s profit. There has been a significant amount of work recently on how management science approaches can support marketing through the internet, for example the special issues edited by Geoffrion & Krishnan (2003); Kannan and Rao (2001). Montgomery (2001) took as one of his examples of real internet applications the use of conjoint analysis and logistic regression to identify the importance of the separate components of a product in their utility to the customer. He applied it to identifying the importance of the item price, the shipping price, the sales tax and the delivery time for an internet bookseller. This has some connection with the first approach in this paper but whereas our objective is to decide in real time on a suitable price for the product, his work was more on identifying the importance of the features. A second example in that review was the idea of pricing using versioning in which variants of a product are sold at different prices. Rossi, McCulloch and Allenby (1996) sought to estimate the price sensitivity of a customer to the different alternative versions of the item using that customer’s previous purchase history and used that to target the distribution of a coupon which discounted the price of the item. The internet is also leading to personalization as a way of using data on customers to give them a better choice. Personalization is the process of using this information to develop a targeted solution to the customer, which we are suggesting is the way in which financial institutions will move. Murthi & Sarkar (2003) survey the work in this area looking at the strategic implications of personalization and the standard techniques that can be used to model the information. These include the classification techniques which were previously developed in the credit scoring area and consumer choice models. They make the point that in the taxonomy of price discrimination the ability to charge different prices for the same product to different customers is first degree price discrimi- nation while offering different products at different prices is second degree price discrimination. They point out that in traditional markets first degree discrimination is not practical but in the internet age it becomes a possibility What we are proposing here is a tailoring both of product and price which is neither first nor second degree discrimination. Karuga, Khraban, Nair, and Rice (2001) develop a genetic algorithm approach to deciding which advertisements to target to a customer using the internet and how to schedule the advertisements. They use a linear programme to identify the ‘effectiveness’ of the features that make up the advertisement but it differs from the way we use linear programming and their objective is to identify which advertisements and which sequence of advertisements produces the highest response rate, not how to adjust price and product features so as to maximize the probability of acceptance. Raghu, Kannan, Rao, and Whinstom (2001) consider the next step in the customization problem—namely how to dynamically update the model. They consider the use of questionnaires to dynamically update the consumers’ preferences. Since, the use of these techniques is so new and so little has been done on jointly adjusting price and product features, one problem is the data. There is no publicly available data set on whether customers accept the offer of a financial product - the features of which are changing from offer to offer. Thus, we have developed our own data set using a fantasy student current account (FSCA) which is a web-based application form targeted at students applying for a bank account. Although it is not argued that this data reflects what will happen if students are applying Table 1 Application and offer characteristics used in the model Characteristics Description Application characteristics Age Age of applicant Sex Sex of applicant Status Marital status of applicant Num_children Number of children Num_cards Number of credit card Wage Some income from wage Loan Some income from loan Contribution Some income from parental contribution Travel Interest in travel, true/false Music Interest in music, true/false Cars Interest in cars, true/false Cinema Interest in cinema, true/false Sports Interest in sports, true/false Clubbing Interest in clubbing, true/false Beer Interest in beer, true/false Country western Interest in C and W music, true/false DIY Interest in DIY, true/false Gardening Interest in gardening, true/false Offer characteristics Overdraft Overdraft limit, five choices Creditcard Credit card included with account, four choices TM No fees on ordering foreign currency for travel, yes/no Insurance Discounts on insurance, four choices Interest Interest paid when account in surplus, four choices Introductory Introductory free gift, 10 choice L.C. Thomas et al. / Expert Systems with Applications 30 (2006) 499–506 501 for a real bank account, it does enable us to investigate different approaches to building acceptance functions. This paper suggests three approaches to determine the probability that an applicant with characteristics x will accept an offer with characteristics o using the FSCA data. Section 2 describes the FSCA data, which is used to build the model of consumer acceptance probabilities. In Section 3, a logistic regression (LR) model is used to obtain an acceptance probability estimate. In Section 4, a linear programming approach is used to build an acceptance probability model. To build such a model, we assume there is a dominant offer characteristic, where the attractiveness of the offer increases (or decreases) monotonically as this characteristic’s value increases. The idea is that given the other offer and applicant characteristics one can identify the value of this dominant characteristic at which this applicant would accept the offer. Section 5 extends this idea by saying this cut-off value between accepting and rejecting of the dominant characteristic need not be fixed but can be described by a probability distribution. One can then use an accelerated life (AL) approach to obtain this distribution. This probability distribution of acceptance reflects the uncertainty consumers have in choosing between closely matched offers as well as the changes in the individual’s mood, the environment where the offer is made and the knowledge of competing offers which occur between identical offers being made to customers with identical applicant characteristics. In applying the accelerated life model, one has double censoring, since if a customer accepts an offer with a particular overdraft limit one only knows the minimum acceptance value is below this value, while if he rejects an offer, one only knows the minimum acceptance value is above this offer value. 2. Fantasy student current account data In order to model the likelihood of consumers accepting the offer of a financial product-the features of which are changing from offer to offer, we developed our own data set using a fantasy student current account (FSCA). The FSCA consists of a website which closely follows the application forms for student bank accounts which are used by the major UK banks. The website consists of three pages. The first is an application form for a FSCA, which is similar to the bank account most students in the UK use for their money transactions and their borrowing. The questions are created by looking at the application forms of ten UK lenders including the four major UK retail banks. The questions asked in the form are described by the application characteristics in Table 1, where the first eight concerned demographic and financial information, and the remaining ten addressed interests (some banks did have one or two such marketing oriented questions in their application forms). The second web page makes the offer of an account and outlines the features that were part of the offer. There were six features of the account that could be changed from offer to offer—given by the offer characteristics of Table 1. In order to obtain a reasonable spread of offer combinations each applicant was randomly put into one of four offer categories. In three of these categories, everyone in that category received the same fixed offer varying from £1250 to £1800 overdraft limit. In the fourth category which had the largest probability the offer was given by one of 42 nodes of a decision tree arrived at by splitting on the applicants’ characteristics. The decision tree was constructed subjectively, using obvious associations and a desire to produce a wide spectrum of offers. When the offer was made on the second page, the applicant had to submit whether they accepted or rejected the offer. This is the outcome that the subsequent models seek to estimate. The final page asks applicants to rate the importance of the offer characteristics in their decision and also how they feel about a bank making different offers to different people. This web page (www.management.soton.ac.uk/staff/fair- isaacs/win.asp) is available for all to use but it was publicized widely to the first year students at the University of Southampton in the UK, with regular prize winning draws for those who had completed the application process. Although there was no guarantee the acceptance/rejection decision was the one they would have made to a real account they had all recently opened such accounts and so were well aware of the product and the features offered by the various banks. A data set of the application and offer characteristics for 331 applicants was obtained from the website. There are 18 applicant characteristics and six offer characteristics. In Sections 3–5, we will deal with three approaches to determine the probability that an applicant with applicant characteristics x will accept an offer with features o using this FSCA data. In each case, we build the model on 265 of the cases in the sample http://www.management.soton.ac.uk/staff/fairisaacs/win.asp http://www.management.soton.ac.uk/staff/fairisaacs/win.asp L.C. Thomas et al. / Expert Systems with Applications 30 (2006) 499–506502 and test it on the remaining 66 cases (so this holdout sample is 20% of the population). 3. Logistic regression based acceptance probability approach Logistic regression is a widely used statistical modeling method in which the probability of a dichotomous outcome is estimated. In general, the logistic regression model has the form log p 1Kp � � Z b0 C b1x1 C b2x2 C/C bk xk Z xb; (1) where p is the probability of the outcome of interest, b0 is an intercept term, bi is the coefficient associated with the corresponding explanatory variable xi, xZ(1,x1, x2, ., xk) and bZ(b0, b1, ., bk) 0. So, in logistic regression, one estimates the log of the probability odds by a linear combination of the characteristic variables. Since p/(1Kp) takes values between 0 and N, log [p/(1Kp)] takes values between KN and CN. Taking exponentials on both sides of (1) leads to p Z expðxbÞ 1 C expðxbÞ : Logistic regression can be used to obtain acceptance probability for FSCA data. In the simplest model let the applicant characteristics be xZ(x1, ., xn) and let oZ(o1, ., om) be the offer characteristics. Then the basic logistic regression approach assume that the probability that an applicant with characteristics x will accept offer o satisfies log½pð1KpÞ� Z b0 C b1x1 C/C bnxn C bnC1o1 C/C bnCmom Z yb; (2) where yZ(1, x1, ., xn, o1, ., om) and bZ(b0, b1, ., bnCm) 0. An obvious extension is to allow interaction characteristics iZ(i1,.,ip) which are combined offer and application characteristics, e.g. take value 1 if customer likes travel and offer gives no fee foreign exchange purchase; 0 otherwise. The first step in building a credit scorecard using logistic regression on a sample of past customers is to coarse classifying the characteristics and we will do this for all versions of our Table 2 Result of logistic regression on offer and applicant characteristics Variable Coefficient S.E. Wald Age1 1.163 0.757 2.359 Age2 1.341 0.685 3.829 Age3 0.484 0.810 357 Sex 0.775 0.460 2.833 Children 1.789 1.237 2.092 Cinema K0.704 0.465 2.298 Diy 2.540 1.244 4.164 Overdraft K2.465 0.846 8.483 Travelmoney 0.920 0.495 3.459 Insurance 1.579 0.845 3.494 Constant K0.878 0.873 1.012 acceptance scorecard. In general, the coarse classifying procedure splits the values of a continuous characteristic into bands and the values of a discrete characteristic into groups of values, where the values in each band or group tend to have roughly the same odds of the to outcomes in the original data sample. The binary variables corresponding to the indicator variable for each of the bands and groups chosen are then the ones used in the logistic regression. Coarse classifying improves the robustness of the scorecard being developed, since it increase the size of the group with a particular regression coefficient. More importantly for continuous variables, it allows for non-monotonicity between the characteristic values and the probability of the outcomes. There are some additional difficulties in using coarse classifying in this context since there are substantial differences in the offers being made to the different consumers in the data set. Ways of dealing with this were investigated in Jung and Thomas (2003) but here, we will use the standard approach (Thomas et al., 2004) where one starts with a very fine classification (every decile of a continuous variable and all values of a categorical variable say) and use the chi-square and information statistics to combine some of these classes. The coarse classifying led to the following bands and groups being constructed Age: 20 or less (Age 1); 21–30(Age 2); 31C(Age 3); missing value is the reference group (remember this is a student population) Children: 1 or more; 0 is the reference group Overdraft: £1250 or less; reference group is more than £1250 Current account interest: more than 1% (interest); reference group is 1% or less Insurance; discount on any form of insurance except musical insturments (insurance); insurance on musical instru- ment (music); reference group is no discount on any insurance. The logistic model is then built on the application and offer characteristics only using the 265 cases in the training sample. The offer variables which are significant in the regression and hence have a major impact in the acceptance scorecard were overdraft, interest and insurance. Table 2 gives the result of the logistic regression. The classification power and robustness of the scorecard developed was tested by using the holdout sample and the results DF Sig Exp (B) 1 0.125 3.199 1 0.050 3.822 1 0.550 1.623 1 0.092 2.170 1 0.148 5.981 1 0.130 0.494 1 0.041 12.674 1 0.004 0.085 1 0.063 2.510 1 0.062 4.850 1 0.315 0.416 Table 3 Classification results using the logistic regression model Training data Holdout data Whole data Actual numbers LR Actual numbers LR Actual numbers LR Y-predicted Y 155 121 39 29 194 150 Y-predicted N 0 34 0 10 0 44 N-predicted N 110 57 27 15 137 72 N-predicted Y 0 53 0 12 0 65 L.C. Thomas et al. / Expert Systems with Applications 30 (2006) 499–506 503 given in Table 3 where Y means the applicant said Yes to the offer they were given and N means that they said No, while predicted Y and predicted N is what the model predicted they would say. Since, the sample is relatively small the analysis was repeated using the leave one out approach, where the regression is in turn built on all but one of the sample and tested on the remaining sample point. The results were very similar to those in Table 3. In the previous scorecard, there were no applicant-offer interaction variables present. The model was extended by including such variables concentrating on interactions between the seven application variables and the three offer variables that appeared in the ‘top ten’ variables of the original scorecard. Chi-squared statistics, the F information statistic and D concordance statistics were used to identify the important interactions and three interactions were identified as relevant. However, when these were introduced into the scorecard there was only a minor improvement with one extra case being correctly classified. This is slightly disappointing in that it says the relevant ranking of the probability of accepting the different offers would be the same for all applicants. One suspects this is because of the limited number of applicants in the sample and the artificiality of the situation. In real situations, these interaction terms would be expected to play more of a role. Subject to o i C ei R c1y i 1 C c o i Kej % c1y j 1 C c ei R 0; 4. Exact cut-off on dominant characteristic using linear programming The second approach to developing a model to determine which offer to make to an applicant, assumes that there is a dominant offer characteristic, where the attractiveness of the offer increases (or decreases) monotonically as this character- istic’s value increases. The idea is that, given the other offer and applicant characteristics, one can identify the value of this dominant characteristic at which this applicant would accept the offer. One could then use this value in profit calculations to see whether it is profitable to make such an offer. In the credit card context, this dominant characteristic could be the interest rate charged for borrowers or the credit limit for transactors, but for student bank accounts, where the overdraft is interest free, the overdraft limit is clearly the most important offer feature to most students. So again assume applicant characteristics xZ(x1, x2, ., xn), offer characteristics oZ(o2, ., on) with OZo1 being the dominant offer characteristic and the interaction characteristics i. We are interested in determining the accept/reject level of O, O * , as a linear function of x, o and i. Hence, we assume O � Z c0 C c1x C c2o C c3i Z cy; (3) where yZ(x, o, i). Taking a sample of previous applicants, if applicant i (with characteristics y i ) accepted an offer of o i then o iRcyi while if applicant j (with characteristics y j ) rejected on offer of o j then o j%cyj, where we are assuming that the likelihood of acceptance increases as O increases. Hence, we can use linear programming to determine the coefficients c as follows. Let the sample of previous customers be labeled 1 to n(a)C n(r) where iZ1, 2, ., n(a) accepted the offer and jZ nðaÞC 1; nðaÞC 2; .; nðaÞC nðrÞ rejected the offer. Let applicant i have applicant/offer characteristics yi Zðyi1; .; y i pÞ and be made an offer o i . Then to find the coefficients c that give the best estimate of the overdraft accept/reject indifference level we want to solve the following linear programme. Minimise e1 C . C enðaÞCnðrÞ 2y i 2 C/C cpy i p; i Z 1; 2; .; nðaÞ; 2y j 2 C/C cpy j p; j Z nðaÞ C 1; .; nðaÞ C nðrÞ; i Z 1; 2; .; nðaÞ C nðrÞ: This has some relationship with the linear programming formulation of how to build a default/not default credit scorecard, but the introduction of different cut-off levels for each individual leads to a somewhat different formulation. This model was applied to the fantasy account data set, using the application and offer variables and their character- istics identified in Section 3. All the offer and application variables were used and the coarse classifying developed in the previous section gave 32 binary variables to put into the linear programme, but because of the results in Section 3 no interaction variables were included. The results were as follows (Table 4). One could imagine that linear programming does particu- larly well here because of the number of variables compared with the size of the training sample. As the sample sizes ð4Þ Table 4 Classification results using linear programming model Training data Holdout data Whole data Actual numbers LP Actual numbers LP Actual numbers LP Y-predicted Y 155 138 39 33 194 171 Y-predicted N 0 17 0 6 0 23 N-predicted N 110 102 27 26 137 128 N-predicted Y 0 8 0 1 0 9 Table 5 Coefficients in Linear programming approach Attribute All age 0 Credit card 1 Credit card 2 Credit card 3C Credit card Musical instr. insurance Score 450 650 0 400 900 K250 Attribute All other or no insurance Intro offer-CD player or miniTV £40 No into offer/brewing kit/£40 CDs and CD vouchers Rail card All other or no insurance Score 0 550 150 650 700 0 L.C. Thomas et al. / Expert Systems with Applications 30 (2006) 499–506504 increases it is likely the improvement in classification accuracy over the logistic regression approach may diminish a little. The variables which had significant coefficients were given in Table 5. Clearly, these values should not be taken as definitive given the fantasy nature of the data but they do show how this approach could lead to useful information as part of the scorecard building process. The scorecard in Table 5 says those with 1 credit card are the ones who will accept the lowest overdraft while those with 0 credit cards want an overdraft of £600 more before they accept, those with 2 credit cards want £400 more while the worldly wise with 3C credit cards want £900 more on their overdraft limit before they accept. Similarly, offering insurance on musical instruments means students would accept a loan with £250 less on the overdraft but other forms of free insurance make no difference. The results suggest that initial offers are counterproductive with offering a mini TV means the overdraft limit being looked for goes up by £550 while offering railcards mean it goes up by £700. For a student aged 21 with no credit cards an offer which includes no insurance and free CDs would mean the credit limit needs to be 450C 650C 0C 650Z £1750 before the student will accept. 5. Cut off distribution on dominant characteristic using accelerated life approach In the previous approach, we assumed there was an exact value of the dominant characteristic at which an applicant would change from rejecting the offer to accepting it. A weaker assumption would be to say there is a probability distribution over the values of the dominant characteristic of where the cut- off point between accepting and rejecting lies. Arguments for assuming a probability distribution of acceptance rather than an exact cut-off point as in the previous section is that the changes in the individual’s desires, economic circumstances and the environment in which the offer is made, all of which can fluctuate rapidly, could mean the same person makes different decisions to the same offer at different times. It might also reflect an applicant’s inability to make decisive judgments between incrementally different offers. In this section we show how using accelerated life model, which is one of the models used in survival analysis, can help one determine such a distribution. Survival analysis is the area of statistics that deals with analysis of lifetime data, but recently (Narain, 2004; Thomas, Banasik, & Crook, 1999; Stepanova & Thomas, 2002), one has been able to use survival analysis approaches, especially the proportional hazards model and the accelerated life models to build consumer credit assessment systems. Here, we present a new application of accelerated life models in the consumer acceptance estimation area. To apply the accelerated life model for FSCA data, we again assume that there is an important offer characteristic and the probability of accepting the offer increases or decreases monotonically as this dominant offer characteristic’s value increases. The idea is that given the other offer and applicant characteristics one can identify the value of this dominant characteristic, or at least the distribution function of the value, at which the applicant would accept the offer. To estimate this distribution function assume O1 is the dominant offer characteristic. If one has applicant character- istics x and offer characteristic (t, o) where t is the value of the dominant monotone characteristic O1, then we are interested in the probability of an applicant with characteristics x accepting offer (t, o). Thus, if T is the lowest value of O1 at which the offer will be acceptable, then Probfindividual with characteristic x accepts offer ðt; oÞgZ ProbfT % tjyZðx; oÞgZ FðtjyZðx; oÞÞ where yZ(x1, ., xl, o1, . om) and hence that the probability of an individual with characteristic x rejecting offer (t,o) is given by Sðtjy Z ðx; oÞÞ Z 1KFððtjy Z ðx; oÞÞ: This is known in survival analysis as the survival function. The accelerated life model can be applied to estimate the reject probability for FSCA Data where Table 1 shows the applicant and offer characteristic used in analysis and the overdraft limit is taken –0.5 –0.4 –0.3 –0.2 –0.1 0 0.1 0.2 0.3 0.4 1 2 3 4 5 6 7 8 9 10 P a ra m e te r e st im a te Fig. 1. AL parameter estimates for introductory gift characteristic. Table 6 Coarse classifying of the introductory gift characteristic Binary variable Type of offer No. of observations Accept Reject Introductory_1 1, 10 36 45 Introductory_2 2, 7 107 47 Introductory_3 3, 4, 5, 6, 8, 9 53 45 L.C. Thomas et al. / Expert Systems with Applications 30 (2006) 499–506 505 as the dominant characteristic. Notice that all the data will be either right or left censored. If applicant i with characteristics xi accepts offer (t, oi) then all we can say is T%t. It applicant j with characteristics xj rejects offer (t, oj) then all we can say is TRt. Thus, we never observe uncensored data. In the accelerated life model, one define the survival (reject) function by SðtjyÞ Z S0ðe yb tÞ; (5) where yZ(1, x1, ., xl, o1, ., om), bZ(1, b0, b1, ., blCm) 0 and S0($) is a baseline survival function. One common accelerated life model in survival analysis is to take S0($)to be the Weibull distribution. This gives a baseline survival function of S0ðtÞ Z expfKðltÞ k g; (6) where l and k are the scale and shape parameters of the Weibull distribution. From (5) and (6), the survival function in accelerated life model can be expressed as follows. SðtjyÞ Z expfKðl expðb0 C b1x1 C b2x2 C/C blxl C blC1ol C/C blCmomÞtÞ k g Z expfKðl expðybÞtÞkg: (7) By applying the accelerated life model of doubly censored data, we can obtain the likelihood function as follows. LðqÞ Z YnðaÞ iZ1 ð1KSðtijyiÞÞ YnðaÞCnðrÞ jZnðaÞC1 SðtjjyjÞ Z YnðaÞ iZ1 ½1KexpfKðl expðyibÞtiÞ k g� ! YnðaÞCnðrÞ jZnðaÞC1 expfKðl expðyjbÞtjÞ k g: The maximum likelihood estimates of l, k and b in the Weibull based accelerated life model are then obtained using Newton– Raphson methods. It is still necessary to coarse classify the variables used but we can use the accelerated life approach not just to build the model but also to do the coarse classifying. This overcomes the problem that in the traditional coarse classifying approach we used in the logistic regression and linear programming methods described earlier we do not consider the actual offer made even though there is a strong interaction between the offer level and the accept–reject decision. The coarse classifying method used in this AL model follows closely the approach used for applying proportional hazards models in credit scoring (Stepanova & Thomas, 2002). It consists of the following steps Step 1. (Continuous characteristic) Split the characteristic into n binary variables with approximately equal number of observations in each variable. (Discrete characteristic) A binary variable is created for each attribute of the characteristic. Step 2. Apply AL model with these binary variables. Step 3. Chart parameter estimates. Step 4. Choose the splits based on similarity of parameter estimates. As an illustration of coarse classifying using the AL model, consider the introductory gift characteristic. Fig. 1 shows the histogram of the parameter estimates and consideration of this leads to the classification given in Table 6. Using the binary variables obtained by this coarse classifying the Accelerated life model is built on the same training sample as previously and tested on the holdout sample. The coefficients are estimated using (7) and the results give an accelerated life model of the form SðtjyÞ Z expfKðl expðb0 C b1Var1 C/C bqVarqÞtÞ k g; where Var1, Var2, ., Varq are the indicator variables obtained by the coarse classifying, b0, b1, ., bq are the parameters to be estimated and t is the overdraft likely to be accepted by an applicant with application and other offer characteristics y. To compare the AL model with the LR and the LP models, all the variables are included in the AL model. Table 7 shows the results on a training, holdout and whole sample for all three methods. Thus, on this data set the accelerated life model classifies the least well but is close to the results for the logistic regression approach. Part of this is due to the smallness of the set compared with the number of variables available and part may be due to the fact there were only a few overdraft levels being offered. If there were only one level of overdraft offered, then estimating a distribution of what level an applicant would accept must be less precise than estimating if they would accept the offer at that given level which in a sense is what the logistic regression is strongest at Table 7 Classification results using AR and the other models Training data Holdout data Whole data Actuals LR LP AL Actuals LR LP AL Actuals LR LP AL Y predicted Y 155 121 138 124 39 29 33 28 194 150 171 152 Y predicted N 0 34 17 31 0 10 6 11 0 44 23 42 N predicted N 110 57 102 50 27 15 22 26 137 72 128 62 N predicted Y 0 53 8 60 0 12 5 1 0 65 9 75 L.C. Thomas et al. / Expert Systems with Applications 30 (2006) 499–506506 6. Conclusions This paper introduces three techniques which can be used to build models of the probabilities that a particular consumer will accept different variants of a generic borrowing product like a credit card or account with an overdraft facility. It derives a data set based on students acceptance or rejection of a fantasy student account offer. Thus, one must include many of the caveats when one uses data which is essentially obtained from gaming experiments rather than from real experience. However, the paper does show that it is possible to build acceptance probability models using such data and makes a preliminary investigation of three different approaches. We believe that two of these—linear programming and accelerated life models— have not been tried before in this context. All three approaches are technically feasible and can result in real time decisions about which variant of the product to offer the current applicant in order to maximize profit, though two of them do require the notion of a dominant offer characteristic. For many products, it does seem reasonable to assume that one characteristic has the necessary monotone properties to apply such procedures. Although the results suggest the linear programming approach to estimate a cut-off level on the dominant characteristic does clearly make the best predictions on this data set both on the training sample and on the holdout sample, we feel these results must be taken with a great deal of caution because of the relatively large number of binary variables available compared with the size of the sample. We believe these probability acceptance models will become increasingly important as the consumer lending market matures and it becomes a buyers rather than a sellers market. They are ideally suited to the interactive application processes that modern telecommunication technology is supporting. They also satisfy the customer relationship marketing credo of tailoring the product to the customer. Acknowledgements This work was supported by a research grant from Fair Isaac. References Fisher, R. A. (1936). The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7, 179–188. Freed, N., & Glover, F. (1981). A linear programming approach to the discriminant problem. Decision Sciences, 12, 68–74. Geoffrion, A. M., & Krishnan, R. (2003). E-business and management science: Mutual impacts. Management Science, 49, 1275–1286. Hand, D. J. (1981). Discrimination and classification. Chichester, UK: Wiley. Jung, K. M., & Thomas, L. C. (2004). A note on coarse classifying in acceptance scorecards, Working paper M04-16, School of Management, University of Southampton. Kannan, P. K., & Raghav Rao, H. (2001). Decision support issues in customer relationship management and interactive marketing for e-commerce. Decision Support Systems, 32, 83–94. Karuga, G. G., Khraban, A. M., Nair, S. K., & Rice, D. O. (2001). AdPalette: An algorithm for customizing online advertisements on the fly. Decision Support Systems, 32, 85–106. Mangasarian, O. L. (1965). Linear and nonlinear separation of patterns by linear programming. Operations Research, 13, 444–452. Montgomery, A. L. (2001). Applying quantitative marketing techniques to the Internet. Interfaces, 31, 90–108. Murthi, B. P. S., & Sarkar, S. (2003). The role of the management sciences in research on personalization. Management Science, 49, 1344–1362. Narain, B. (2004). Survival analysis and the credit granting decision. In L. C. Thomas, J. N. Crook, & D. B. Edelman (Eds.), Readings in credit scoring (pp. 235–245). Oxford: Oxford University Press (Reprinted). Raghu, T. S., Kannan, P. K., Rao, H. R., & Whinstom, A. B. (2001). Dynamic profiling of consumers for customized offerings over the Internet; a model and analysis. Decision Support Systems, 32, 117–134. Rosenberg, E., & Gleit, A. (1999). Quantitative methods in credit management: A survey. Operations Research, 42, 589–613. Rossi, P. E., McCulloch, R. E., & Allenby, G. M. (1996). The value of purchase history data in target marketing. Marketing Science, 15, 321–340. Stepanova, M., & Thomas, L. C. (2002). Survival analysis method for personal loan data. Operations Research, 50, 277–289. Thomas, L. C. (2000). A survey of credit and behavioural scoring: Forecasting financial risk of lending to consumers. International Journal of Forecasting, 16, 149–172. Thomas, L. C., Banasik, J., & Crook, J. N. (1999). Not if but when loans default. Journal of the Operational Research Society, 50, 1185–1190. Thomas, L. C., Edelman, D. B., & Crook, J. N. (2002). Credit scoring and its applications. Philadelphia, PA: SIAM. Wiginton, J. C. (1980). A note on the comparison of logit and discriminant models of consumer credit behaviour. Journal of Financial and Quantitative Analysis, 15, 757–770. Modeling consumer acceptance probabilities Introduction Fantasy student current account data Logistic regression based acceptance probability approach Exact cut-off on dominant characteristic using linear programming Cut off distribution on dominant characteristic using accelerated life approach Conclusions Acknowledgements References 
work_s7s7uwa7qfbffcltbrcznpprve	----	doi:10.1016/j.eswa.2006.01.028 www.elsevier.com/locate/eswa Expert Systems with Applications 32 (2007) 674–686 Expert Systems with Applications Expert system for low frequency adaptive image watermarking: Using psychological experiments on human image perception Ho Seok Moon a, Taewoo You b, Myung Ho Sohn b, Hye Soo Kim c, Dong Sik Jang a,* a Industrial Systems and Information Engineering, Korea University, 1, 5-ka, Anam-Dong, Sungbuk-Ku, Seoul 136-701, South Korea b Business Administration, Myongji College, 356-1, Hongeun 3-Dong, Seodaemun-Ku, Seoul 120-776, South Korea c Department of Electronics and Computer Engineering, Korea University, 1, 5-ka, Anam-Dong, Sungbuk-Ku, Seoul 136-701, South Korea Abstract This paper proposes a new system for low frequency adaptive image watermarking based on the statistical data from psychological experiments on human image perception. The new approach can lead to a reduction of degrading the subjective image quality that often occurs when watermark is embedded into low frequency area. In order to reduce the degrading of image quality, the new approach deter- mines the strength of watermark according to local image characteristics such as brightness and contrast. By conducting a behavioral experiment on human image fidelity based on the psycho-visual image association technique, we were able to infer the relationship between the watermark strength and the different levels of image brightness and contrast information. The exact watermark is extracted according to edge characteristics by adopting a so-called edge mask that exploits the coefficients of subbands in the subsampled discrete wavelet transform images. Thus, our new approach does not require original images for watermark. We also show the new approach is practically validated against some standard images. � 2006 Elsevier Ltd. All rights reserved. Keywords: Blind image watermarking; Discrete wavelet transform; Psychological experiments; Human image perception; Edge mask 1. Introduction One of the salient issues in this internet era is that digital contents, such as audio, image, and video, may be copied in an unauthorized manner. This stimulated the development of copy protection technologies like digital watermarking. The digital watermarking is now receiving a widespread attention as a possible solution to the protection applica- tions of intellectual property rights (Cox & Miller, 1997; Dai, Zhang, & Yang, 2003; Lin & Chen, 2000; Swanson, Kobayashi, & Tewfik, 1998). Digital watermarking is defined as a technique of embedding some hidden informa- tion called watermark into digital multimedia without loss of perceptual quality of watermarked data. Once created, 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.01.028 * Corresponding author. Tel.: +82 2 925 3655; fax: +82 2 3290 3776. E-mail addresses: bawooi@korea.ac.kr (H.S. Moon), taewooyou@ mail.mjc.ac.kr (T. You), totalsol@mail.mjc.ac.kr (M.H. Sohn), hyesoo@ korea.ac.kr (H.S. Kim), jang@korea.ac.kr (D.S. Jang). the watermark can be detected or extracted for the purpose of owner identification or/and integrity verification of tested data (Levicky & Foris, 2004). In this paper, we pro- pose a new watermarking technique that is not only appli- cable to images, but also easily expandable to other host media. It is generally argued that a watermark should at least meet the following conditions (Chen & Lin, 2003): (1) It should be perceptually invisible (or transparent). (2) It should be difficult to remove without seriously affecting the image quality. (3) It should robustly resist to image distortions caused by attacks such as common image processing opera- tions and lossy image compression. The major objective of this paper is to propose a new approach to low frequency adaptive image watermarking for security applications that can be distinct from others mailto:bawooi@korea.ac.kr mailto:taewooyou@mail.mjc.ac.kr mailto:taewooyou@mail.mjc.ac.kr mailto:totalsol@mail.mjc.ac.kr mailto:hyesoo@korea.ac.kr mailto:hyesoo@korea.ac.kr mailto:jang@korea.ac.kr Fig. 1. Two-level wavelet decomposition. H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 675 in utilizing the statistical data from psychological experi- ments on human image. It is worthwhile to point out three main contributions that can be drawn from the current study. First, the fidelity of images was enhanced through controlling the watermark strength with the data associ- ated with brightness and contrast as local image character- istics, without sacrificing the robustness of images. To our best knowledge, we considered people’s perception on image that can vary by visual characteristics. The key infor- mation on the strength of image brightness and contrast was acquired through an experiment on human being’s image fidelity. Second, we devised a technique that does not require original images by considering the characteris- tics of edge. The embedding and extracting of watermark according to the edges of images was implemented through the so-called edge mask. The term ‘edge mask’ is defined as a mask representing the edge direction of images, say hor- izontal or vertical. The edge mask plays a role as the best predictor of the original image in extracting the watermark. The edge mask is determined by choosing the direction that takes the lower pixel value between the watermark position and its neighborhood pixels in each direction. Third, the experimental results demonstrate that the proposed method is resistant to some image processing operations and to Joint Photographic Experts Group (JPEG) lossy compression, which usually do not degrade the image qual- ity. Fig. 1 illustrates the overview of the proposed system. The organization of the paper is as follows: A survey of existing studies is made in Section 2. Section 3 discusses the embedding method. Section 4 describes the extracting method. Section 5 contains experimental results and appli- cation. Finally, Section 6 draws a conclusion. 2. Literature review Existing literature reveals two techniques for the water- marking of images: frequency domain and spatial domain (Wu & Shih, 2004). Most of the recent watermarking schemes employ mainly the frequency domain approach because it is superior to the spatial domain approach in robustness and stability (Hsu & Wu, 1998; Lin & Chen, 2000; Miu, Lu, & Sun, 2000; Shih & Wu, 2005). However, there is a crucial question that should be answered: which frequency band in frequency domain can be robust and imperceptible to various attacks? According to Weber’s rule, the low frequency area is more robust than high and middle frequency areas (Shi & Sun, 1999). There have been various methods to embed the watermark into the low fre- quency area (Huang, Shi, & Shi, 2000; Joo, Suh, Shin, & Kikuchi, 2002). It is known that these prior approaches are robust to various attacks. Despite their robustness, the existing methods still leave a problem unanswered. The key concern is that, if the watermark is embedded arbi- trarily in the low frequencies, without adapting to local image characteristics, the image quality could be visually degraded. Thus, of importance is the amount of modifica- tions that can be made by watermark embedding in low frequency pixels. This problem can be solved by using the Human Visual System (HVS) model in digital watermark- ing. The HVS model determines the maximal amount of watermark signal that can be tolerated at every location without affecting the visual quality of the image (Barni, Bartolini, & Piva, 2001; Chen & Lin, 2003; Levicky & Foris, 2004). To embed the watermark with a minimal loss in image fidelity, the watermark strength modulation should reflect sufficiently the local image characteristics such as texture, edge presence, and luminance (Taskovski, Bogdanova, & Bogdanov, 2002). In order to determine the optimal strength of water- mark, we should be not only well aware of local image characteristics, but also be knowledgeable on how human beings feel the fidelity of image visually. The sensitivity of human being to noise strength does not remain constant, but changes according to the surround modulation charac- teristics of noise. The brightness difference required to detect a noise in a particular area increases with the bright- ness of the area near the noise (i.e., watermark embedded) (Stiles, 1978; Wyszecki & Stiles, 1982). In other words, the probability of not distinguishing noise increases as the average brightness rises. Contrast also affects the possibil- ity of detecting noise. Even in the case that the mean brightness is fixed at noise surroundings, the perceived con- trast of the noise embedded at center tend to weaken (Chubb, Sperling, & Solomon, 1989; Olzak & Laurinen, 1999). Thus, we can predict that the strength of watermark is less likely to be perceived by human beings as the level of brightness and contrast increases. Such a prediction indi- cates that the impact of embedded watermark on the fidel- ity of images can vary by the level of brightness and contrast. Therefore, it is less likely that the images with higher level of average brightness and contrast are dis- cerned even at the same strength of watermark embedded. Also, watermarking techniques can be divided into two distinct categories depending on the necessity of original images for the watermark extraction. Although the exis- tence of the original image may facilitate the watermark extraction to a certain extent, two problems can come out: (1) At the risk of insecurity the owners of original images may be compelled to share their work with anyone who wants to check the existence of the watermark and (2) 676 H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 it is time-consuming and cumbersome to search out the originals within the database that correspond to a given watermark. Thus, in order to overcome these problems we need a method for extracting the embedded watermark without requiring the original image (Chen & Lin, 2003). This method is called a blind watermarking technique. The blind watermarking techniques appear far more useful since the availability of an original image is usually unwar- ranted in real-world scenarios (Wang & Pearmain, 2004). As a blind watermarking technique, Lin and Chen (2000) proposed a method in which binary watermark is embedded directly into the Lease Significant Bit (LSB) of Discrete Cosine Transform (DCT) coefficients in an image. This method is not robust because the LSB value of coeffi- cients can be easily forged from various attacks of image processing. Chu (2003) and Wang and Pearmain (2004) suggested a blind watermarking technique that can estimate the pixels of an original image on the supposition that adjacent pixels in the image were similar in a gray level. Chu (2003) intro- duced a method using a random perturbation to subimages obtained by subsampling the image. To extract the water- mark, this method selects randomly a horizontal or vertical subimage and exploits it to estimate an original image. Wang and Pearmain (2004) obtained the estimated pixel value by computing the average of adjacent values (for example, the pixel values in the mask size of 3 · 3, 5 · 5, 7 · 7) of an original image pixel. Though adjacent pixels at gray levels are generally alike in image natures, some regions that can be subject to abrupt changes of pixel values such as edge areas cannot be estimated correctly because of the discrepancy between the value of an original pixel and an estimated pixel. In Chu’s method (2003), when the original image pixel is esti- mated with a pixel of a horizontal (or vertical) subimage, regions with horizontal (or vertical) edges can still be a problem. Also, the method of Wang and Pearmain (2004) Fig. 2. The overall watermarkin that uses average information in the mask can cause false estimations in the edge regions. In addition, their methods do not account for the characteristics of images in deter- mining the strength of watermark. The results of overall literature review can be summa- rized in two-fold: First, the fidelity of images should be enhanced through controlling the watermark strength with the data associated with local image characteristics, without sacrificing the robustness of images. Second, the blind watermarking is more effective than non-blind watermarking. But without considering the edge of image, the extracting process will not be exact. So, the edge of image is important to estimate the original image in blind watermarking. This paper attempted to solve these problems. 3. The embedding method Our embedding strategy is based on a Discrete Wavelet Transform (DWT). In the DWT, an image is first decomposed into four sub- bands, LL1, LH1,HL1 and HH1(L: i.e., low, H: i.e., high) by cascading horizontal and vertical two-channel critically subsampled filter banks (Chen & Lin, 2003). Each entry in subbands LH1, HL1 and HH1 represents the finest scale wavelet coefficients. To obtain the next coarser scale of wavelet coefficients, the subband LL1 is further decom- posed and critically subsampled. The process continues until some final scale is reached. Fig. 2 shows the image decomposed into seven subbands for two scales. The lowest frequency subband is at the top left corner and the highest frequency subband is at the bottom right corner. Thus, in our implementation, we modified the wavelet coefficients selected from subband LL2. In the proposed watermarking method, the original image is a gray-level image of size N1 · N2 and the digital watermark is a binary image of size M1 · M2. g procedure of the system. H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 677 Given the image X[n1, n2], n1 = 0, . . . ,N1 � 1, n2 = 0, . . . , N2 � 1, then X 1½k1; k2� ¼ X ½2n1; 2n2�; X 2½k1; k2� ¼ X ½2n1 þ 1; 2n2�; X 3½k1; k2� ¼ X ½2n1; 2n2 þ 1�; X 4½k1; k2� ¼ X ½2n1 þ 1; 2n2 þ 1� ð1Þ for k1 = 0, . . . ,N1/2 � 1, k2 = 0, . . . ,N2/2 � 1 are four subimages obtained by subsampling. These are trans- formed via DWT to obtain sets of coefficients Yi[k1, k2], i = 1, 2, 3, 4. Since the subimages Xi’s are highly correlated, it is expected that Yi � Yj, for i 5 j except edge (Chu, 2003). The embedding and extracting methods are based on comparisons among the DWT coefficients of the four DWT subimages obtained by subsampling. By making dif- ferent modifications to the DWT coefficients pertaining to different subimages, the watermark without comparison with the original image is extracted. Our overall embedding steps are described in Fig. 3. The algorithm works according to the following steps. 3.1. DWT of the subimage generated by decomposing the original image The original image is decomposed into four subimages according to Eq. (1). Each of the four subimages is trans- Fig. 3. Watermark e formed via two-level DWT to obtain four LL2 subbands, one of which is selected for embedding a watermark. 3.2. Making edge mask The edge mask is used to consider the edge of local image characteristics. The clone mask is constructed according to the pixel-wise relationship between the selected subband and neighboring two subbands. If the LL2 subband of Y1 is selected to embed the watermark, the edge mask is constructed based on the pixel-wise rela- tionship between the selected band and the neighboring two LL2 subbands of Y2 and Y3. In general, DWT subimages have approximately same coefficients at the same spatial location except edge areas. Thus, after embedding a watermark into one of the sub- images, we can easily extract the watermark later by com- paring the watermarked subimage with the rest of the subimages. Thus this method does not need an original image for extracting the watermark. Once we select a subimage to embed the watermark, we have to select a subimage to compare for later watermark extraction. Selecting just a horizontal or vertical neighbor- ing subimage for comparison purpose will be problematic because after DWT there are always some differences between the images around the edges. Thus, to deal with the problem, the proposed method uses an edge mask as in Fig. 4(a). For example, Fig. 4(a)–(d) shows how to select mbedding steps. Fig. 4. Selecting neighbor subimages and an edge mask: (a)–(d) subimage selection (a gray block is a subimage to be watermark-embedded and the two white blocks enclosed in dashed line are neighboring subimages), (e) the example of an edge mask. 678 H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 neighboring subimages (white blocks enclosed in the dashed line) after selecting a subimage to be embedded (gray block) and Fig. 4(e) shows one example of an edge mask. In Fig. 4, H (i.e., horizontal) means the coefficient of a selected subband Y1 at that location is more similar to that of Y2 than to that of Y3, whereas V (i.e., vertical) means the coefficient at that location is more similar to that of Y3 than to that of Y2. In the procedure, it should be noted that, to strengthen ownership, we can have the degree of freedom of which the subimage is selected for embedding the watermark. Thus, selecting a subimage can be used as an author key to preventing a hacker’s attack. 3.3. Pixel-based pseudo-permutation of the watermark In this paper, the binary image of size M1 · M2 of our laboratory logo was used as a watermark. In order to dis- perse the binary pattern on a space, pseudo-random per- mutation is performed as follows: (1) number each pixel from zero to M1 · M2, (2) generate each number in random order, and (3) generate the coordinate pairs by mapping the random sequence number into a 2-D sequence. 3.4. Determining the watermark strength according to image characteristics As mentioned earlier, this paper exploits subjective psy- chological factors when determining watermark strength, based on the psychological presumption that human beings tend to exhibit different sensitivity to noise by levels of brightness and contrast. In order to prove this proposition, we conducted an experiment under quasi-environments. The aim of the experiment is to see how human being’s sen- sitivity to noise, i.e., the strength of watermark that can be detected by human beings, changes by different levels of brightness and contrast. We provided eight graduate stu- dents with different combinations of watermark strength, brightness, and contrast. And we recorded the level of watermark strength they can marginally recognize the dif- ference between original images and watermarked images. In our experiment, we analyzed the brightness and con- trast for about 1200 images in the image database of Corel Draw. Each image was transformed into gray images with 256 intensity levels, and then the mean brightness and mean contrast were calculated by applying Eqs. (2) and (3) to each pixel value. As a result, the average brightness and contrast were 113.21(±42.56) and 56.82(±18.03). The levels of image brightness were classified as three: average level (115), average �1 standard deviation (72), and aver- age +1 standard deviation (158). The levels of contrast were grouped in the same way: average (57), average �1 standard deviation (39), and average +1 standard deviation (75). Mean ðlÞ¼ 1 N 1N 2 XN 1�1 i¼0 XN 2�1 j¼0 X ij ð2Þ Contrast ðSÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPN 1�1 i¼0 PN 2�1 j¼0 ðX ij � lÞ 2 N 1N 2 � 1 s ð3Þ Moon, Sohn, and Jang’s (2004) proposed method is used as watermark embedding method in this experi- ment. The watermark is embedded into lowest resolu- tion representation (LL2) of the original image obtained after applying two-level wavelet decomposition in this method. In this experiment, we used conventional images such as Airplane, Lena, and Baboon. We generated nine combina- tions of images by the three levels of brightness and con- trast. Watermark image was generated by embedding different strength-of-noise (T). By increasing T from 1 to 15 by one, we made watermark images with differing strength. We limited the number of images to three, because the objects of experiments, if they are exposed to too many images, may be distracted, thus result in a distortion of the experimental results. The choice of three images in experiment is better applicable than others as we can verify in Section 5 where the experiment were applied to various 12 images. Fig. 5 shows the images of Lena varying accord- ing to various combinations of contrast and brightness in this experiment. All images were presented on 21 inch monitors at 512 · 512 pixels. Students were allowed to watch the mon- itors at the distance of 50 cm. The participants were unable to recognize what kinds of noise are embedded in images since they did not have any prior experience in the water- marked images. Participants were asked to tell water- marked images from non-watermarked images. Each participant estimated the strength-of-noise 18 times (6 times per an image) under nine combinational con- ditions derived from brightness and contrast. The average of the values was used as the strength measure of noise. The analysis of variance in Fig. 6, demonstrates that the strength-of-noise increases as the mean brightness rises (F(2, 14) = 87.86, p < 0.001) and as the mean contrast rises (F(2, 14) = 103.38, p < 0.001). However, there was no sig- Fig. 5. Lena images used in psychological experiments. Fig. 6. Predicted watermark strength according to the mean brightness and contrast of image. H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 679 nificant interaction effect between brightness and contrast (F(4, 28) = 1.01, p > 0.05). The experiment shows that the sensitivity of human being to noise could be the same though levels of strength-of-noise are physically different. In other words, though the values of peak signal-to-noise ratio (PSNR) based on physical difference of images are the same, the perceived watermark strength can be different across the brightness and contrast of image. Therefore, the water- mark strength should be adjusted properly by the bright- ness and contrast of image. Using the ordinary least square (OLS) method we regressed the mean brightness (l) and the mean contrast (S) on the perceived watermark strength (T). The result of the OLS estimation is summarized in Eq. (4). We present the estimation results using the whole sample correspond- ing to our three raw images. It appears that both the mean brightness and the mean contrast are significantly related to watermark strength. The unit change in brightness leads to 0.025 change in watermark strength, while one unit change in contrast results in 0.114 changes in watermark strength. In this paper we used the estimated regression Eq. (4) to determine the watermark strength. The numbers in parenthesis are t statistics: Watermark strength ðTÞ¼ 2:880 þ 0:025l þ 0:114S ð3:10Þ ð4:54Þ ð7:41Þ R2 ¼ 0:35 ð4Þ 3.5. Modification of coefficients in the selected subimage using edge mask By modifying the coefficients of LL2 based on the edge mask and watermark, we embed the watermark into the LL2 subband of the selected subimage. Fig. 7 shows exam- ples of the embedding procedure, where (a)–(c) represent the LL2 subband of Y1 and two neighboring LL2 subbands of Y2 and Y3, respectively. Fig. 7(d) is the edge mask obtained from (a) to (c). Fig. 7(e) is a permuted binary watermark with size i = 0, . . . , M1 � 1, j = 0, . . . , M2 � 1. If we change the coefficients of LL2 subband of Y1 in accordance with the edge mask and watermark, we come up with four cases: (1) the associated pixels of edge mask and watermark have H and 1, (2) the associated pixels of edge mask and watermark have H and 0, (3) the associated pixels of edge mask and watermark have V and 1, and (4) the associated pixels of edge mask and watermark have V and 0. In each case, the coefficients of LL2 subband of Y1 are modified by Eqs. (5) and (6). Here, Y01 is the LL2 sub- band after the watermark W is embedded, and T is water- mark strength to be determined according to Eq. (4). Fig. 7. Embedding example: (a) The LL2 subband of Y1(Y1(i, j)). (b) The LL2 subband of Y2(Y2(i, j)). (c) The LL2 subband of Y3(Y3(i, j)). (d) Edge mask (S(i,j)), (e) Watermark (W(i, j)). (f) Modified coefficients (Y 01(i, j)). 680 H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 If Sði; jÞ¼ H and W ði; jÞ¼ 1 Y 01ði; jÞ¼ Y 1ði; jÞ if Y 1ði; jÞ > Y 2ði; jÞ Y 2ði; jÞþ T otherwise � Else if Sði; jÞ¼ H and W ði; jÞ¼ 0 Y 01ði; jÞ¼ Y 2ði; jÞ� T if Y 1ði; jÞ > Y 2ði; jÞ Y 1ði; jÞ otherwise � ð5Þ Else if Sði; jÞ¼ V and W ði; jÞ¼ 1 Y 01ði; jÞ¼ Y 1ði; jÞ if Y 1ði; jÞ > Y 2ði; jÞ Y 3ði; jÞþ T otherwise � Else if Sði; jÞ¼ V and W ði; jÞ¼ 0 Y 01ði; jÞ¼ Y 3ði; jÞ� T if Y 1ði; jÞ > Y 3ði; jÞ Y 1ði; jÞ otherwise � ð6Þ The basic rule of Eqs. (5) and (6) is that, if W(i, j) = 1, then Y1(i, j) should be higher than a subband to be compared. Otherwise, Y1(i, j) should be modified to have a higher value. If W(i, j) = 0, Y1(i, j) should be lower than a com- pared subband. If higher, it should be modified to have a lower value. For example, consider the circled pixel in Fig. 7. The edge mask takes H and the watermark takes Fig. 8. Watermark e 1. In this case, the horizontal subimage Y2 is similar to Y1. Then Y1(0, 0) and Y2(0, 0) are compared. Because Y1(0, 0) is lower than Y2(0, 0),Y1(0, 0) should be changed to have higher value than Y2(0, 0) by the amount of strength T according to Eq. (5). In order to explain embed- ding method easily, we set T value to ‘6’ in this example. 4. The extraction method Fig. 8 shows the overall extraction steps. The detailed steps are as follows. 4.1. Decomposing watermarked image and DWT The watermarked image is decomposed into four sub- images by subsampling. then each subimage is transformed via two-level DWT. 4.2. Extracting the watermark using the edge mask There are two steps in this process. First, we select a watermarked subimage and two neighboring subimages. Second, we extract a watermark by performing a pixel-wise xtraction steps. Fig. 9. Extraction example: (a) The LL2 subband of a watermarked subimage Y 0 1 (i, j). (b) The LL2 subband of Y2(i, j). (c) The LL2 subband of Y3(i, j). (d) Edge mask S(i, j). (e) Extracted watermark W*(i, j). H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 681 comparison among the images. More specifically, if the coefficient of a watermarked subimage is higher than that of a neighboring subimage, the watermarked pixel at the same location is set to ‘1’, otherwise its value is set to ‘0’ (see Eqs. (7) and (8)). This is the reverse procedure of the embedding scheme described in the previous section. For example, Fig. 9 shows one extraction procedure where (a) is the LL2 subband (Y 0 1) of the watermarked subimage. For the circled pixel at (0, 0) in (a), the associated pixel of the edge mask has H, see (d). Thus, the coefficient of a horizontal neighboring subimage (i.e., Y2(0, 0) = 213) is selected for comparison. Because the coefficient of Y 01(0, 0) is higher than that of Y2(0, 0), the watermark pixel at (0, 0) is set to ‘1’. If Sði; jÞ¼ H W �ði; jÞ¼ 1 if Y 01ði; jÞ > Y 2ði; jÞ 0 otherwise � ð7Þ Else if Sði; jÞ¼ V W �ði; jÞ¼ 1 if Y 01ði; jÞ > Y 3ði; jÞ 0 otherwise � ð8Þ 4.3. Reversing the permutation and similarity measurement Next process is to reverse the pseudo-random permuta- tion for extracting a visually recognizable watermark according to the predefined pseudo-random order. The extracted watermark is a visually recognizable pat- tern. Subjectively, the viewer compares the result with the referenced watermark. However, the subjective measure- ment is dependent on factors such as the expertise of the viewers and the experimental conditions, etc. Thus, as a proxy for similarity measure we use the normalized corre- lation (NC) between the extracted watermark, W*[i, j] and the referenced watermark W [i, j] given in Eq. (9), where M1 and M2 are the horizontal and vertical sizes of the watermark image, respectively. NC ¼ PM 1�1 i¼0 PM 2�1 j¼0 W ½i; j�W �½i; j�PM 1�1 i¼0 PM 2�1 j¼0 W ½i; j�W ½i; j� ð9Þ To establish a more quantitative measure of impercepti- bility, we make use of the PSNR metric. Although this measure is generally not very accurate, it serves as a good rule of thumb of the invisibility of the watermark. The PSNR is defined as Eq. (10) in units of dB, where X is the original image, X* is the watermarked image, and NT is the number of pixels in X. In general, if the PSNR value is greater than 35 dB then the perceptual quality is accept- able, i.e., the watermark is almost invisible to human eyes (Chen & Lin, 2003). PSNR ¼ 10 log 255 � 255 1 N T PN T �1 i¼0 ðX i � X � i Þ ð10Þ 5. Experimental results and application Fig. 10 shows an example of embedding and extracting results for images of ‘Lena’. In our experiment, 12 standard gray images (Airplane, Bridge, Baboon, Cameraman, Church, Girl, Girl2, House, Lena, Man, Man2, Peppers), as shown in Fig. 11, are used as original images. All of them were sized 512 · 512. Watermarks are embedded in these images. The pattern with ‘I I S LAB’, sized 64 · 64, was used as a watermark. The performance of the proposed method was evaluated in terms of the imperceptibility and robustness against various attacks: (1) the decision of watermark strength regarding image local characteristic, (2) the validity of edge mask, (3) image processing opera- tion (image contrast and histogram equalization), and (4) JPEG compression. Fig. 10. Example of the proposed watermarking approach: (a) test image, (b) watermark, (c) watermarked image (with PSNR = 40.35 dB), (d) extracted watermark with NC = 1.0. Fig. 11. Standard images used in our experiment. 682 H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 5.1. Decision of watermark strength regarding image local characteristic We conduct an experiment on how the estimates from Eq. (4) for various levels of brightness (mean) and contrast (standard deviation) affect the imperceptibility of images. We compute the strength of watermark by partitioning an image into equal slices, calculating the mean and standard deviation of each partition, and plugging the data into (4). The partitioning of image is made on various sizes of partition, as shown in Table 1. Image should not be over- lapped by partitioning. Table 1 reports the PSNR of water- marked images calculated from different strength of watermark according to local image characteristics. That is, the strength of watermark does not remain con- stant, but it is changed based on Eq. (4). The image mean and standard deviation entries in Table 1 represent those for un-partitioned 512 · 512 images. Table 1 shows that, after our proposed method is applied on various mean and standard deviations of images, the subjective perceived image quality was enhanced, and the PSNR stayed high at 40 dB. We also changed the size of partitions, but the results did not change significantly. However, we should note that the quality of image can be degraded, if the size of partition is too small. Thus, it is unnecessary that we partition an image into too small pieces and apply the mean and standard deviations of the partition to the equation. Table 1 Imperceptibility results (PSNR) according to different segmented image size Image PSNR according to segmented size Image mean Standard deviation 64 · 64 32 · 32 16 · 16 8 · 8 4 · 4 1 38.73 38.64 38.54 38.36 38.13 178.4 46.5 2 39.22 39.11 39.04 38.92 38.82 132.8 41.7 3 40.52 40.14 40.07 39.79 39.52 113.8 52.88 4 40.55 40.12 40.11 40.45 39.79 118.7 62.2 5 40.41 40.19 40.08 39.91 39.58 121.4 56.9 6 39.4 39.06 38.92 38.71 38.56 132.3 67.9 7 42.28 42.07 41.7 41.49 41.27 73.6 42.6 8 40.38 40.09 40.14 39.97 39.75 138.1 46.7 9 40.41 40.21 40.08 40.02 39.68 125.2 45.2 10 40.88 40.39 40.13 40.01 39.75 117.8 69.8 11 40.61 40.47 40.33 40.04 39.78 111.4 47.4 12 39.78 39.64 39.39 39.18 39.04 129.9 51.9 H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 683 5.2. The validity of edge mask In watermarking method considering the relationship of neighboring pixels, if the watermark is embedded into area that the values of neighboring pixels are considerably dif- ferent (edge area), it is obvious to degrade image quality because of embedding more strong watermark into that area. In order to prevent degrading image quality, water- mark should not be embedded into that area generally. But this method resulted in inexact watermark extraction. In this paper, we solve the problem through edge mask. Because we embed and extract watermark according to edge mask, we can select neighboring pixel that is not edge Fig. 12. The validity of the edge mask, (a), (d), (g) and (j) are watermarked ima (c), (f), (i) and (l) are extracted watermarks using edge mask. for embedding and extracting. Also, we can prevent severely degrading image quality and extract exact watermark. Fig. 12 shows the validity of edge mask. Fig. 12(a), (d), (g), and (j) are watermarked images with PSNR 40 dB. Fig. 12(b), (e), (h), and (k) are extracted watermarks with- out the edge mask. Here, original pixels were uni-direction- ally estimated with a horizontal or vertical direction. As expected, the quality of extracted watermarks is poor, because estimation cannot be accurate around image edges. Contrary, Fig. 12(c), (f), (i), and (l) are extracted water- marks with edge mask. Since original pixels can be esti- mated with edge mask that reflects image characteristics, ges, (b), (e), (h) and (k) are extracted watermarks without using edge mask, 684 H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 the quality of extracted watermarks is better than the pre- vious cases. Fig. 14. The extracted watermarks of the JPEG compressed version of image number 6: (a) NC = 96.17 for compression ratio 2.83, (b) NC = 95.07 for compression ratio 4.68, (c) NC = 92.85 for compression ratio 5.33, (d) NC = 86.85 for compression ratio 6.85. Table 3 5.3. Image processing operation Fig. 13 and Table 2 show the results of applying the con- trast and histogram equalization to the watermarked image. Table 2 contains the results for the partitioned area by different size of partitions. It is shown that the results are robust to the contrast and histogram equalization attacks regardless of the size of partitioning. Watermark extraction results (NC) after applying JPEG compression Image number NC results according to JPEG compression ratios 2.83 4.68 5.33 6.85 1 88.31 86.5 83.5 59.99 2 85.44 74.51 69.53 62.74 3 95.61 94.53 91.72 83.22 4 91.58 90.7 85.18 76.29 5 94.55 93.9 91.97 83.84 6 96.17 95.07 92.85 86.85 7 93.43 91.41 81.84 67.26 8 89.58 89.33 86.82 79.54 9 87.72 85.91 82.71 77.37 10 91.7 91.36 86.89 77.81 11 94.53 93.7 90.45 80.98 12 86.43 81.05 78.9 74.02 5.4. JPEG lossy compression and application Fig. 14 and Table 3 show the extracted watermarks from the JPEG compressed version at various compression ratios and the corresponding NC values of the extracted watermarks. In this experiment we employed the mean and standard deviation from 16 · 16 partition. Obviously our proposed method also appears robust to the high level of compression. Fig. 15 shows the Graphic User Interface (GUI) which was used in the proposed system. The programs have been implemented in Visual C++. Fig. 13. Watermark extraction results after applying contrast and histogram equalization: (a) contrast enhanced version of Fig. 7(c), (b) extracted watermark from (a) with NC = 84.28, (c) histogram equalized version of Fig. 7(c), (d) extracted watermark from (c) with NC = 90.50. Table 2 Watermark extraction results (NC) after applying contrast and histogram equalization according to different segmented image size Image number NC results according to watermark extraction Contrast attack Histogram equalization attack 64 · 64 16 · 16 4 · 4 64 · 64 16 · 16 4 · 4 1 88.5 85.16 85.21 88.4 88.4 88.23 2 93.09 92.97 92.85 92.26 91.99 92.46 3 85.22 85.35 85.28 93.55 93.68 93.53 4 70.95 70.95 70.95 86.5 86.55 86.25 5 75.39 75.42 75.07 91.6 91.72 91.65 6 69.14 69.53 69.58 93.12 93.04 92.9 7 79.42 79.35 79.35 88.6 88.75 88.99 8 80.86 80.88 80.66 85.45 85.6 85.55 9 84.28 84.18 84.35 90.5 90.58 91.11 10 60.57 60.52 60.52 86.11 86.13 86.79 11 83.13 83.2 83.23 91.4 91.09 91.55 12 83.37 83.35 83.33 93.41 93.41 93.27 Fig. 15. The GUI of the proposed system. H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 685 6. Conclusion This paper proposes a new approach to low frequency adaptive image watermarking based on the statistical data from psychological experiments on human image percep- tion. The subjective image quality tends to be degraded when watermark is embedded into low frequency area. In order to reduce the degrading of image quality, we devised a new approach that can determine the strength of water- mark according to local image characteristics such as brightness and contrast. The relationship between the watermark strength and the different levels of image bright- ness and contrast was statistically established based on a behavioral experiment on human image fidelity using the psycho-visual image association technique. Also, we extracted the exact watermark according to edge character- istics by adopting a so-called edge mask that exploits the coefficients of subbands in the subsampled DWT images. Our new approach is distinct in that the original images are not required for watermark. By experiments to some standard images, we showed the new approach is practi- cally validated. The potential contributions of the paper are summarized as follows. First, the fidelity of images was remarkably enhanced through controlling the watermark strength with the data associated with brightness and contrast as local image characteristics, without sacrificing the robustness of images. In this scheme we took into account people’s perception on image that can vary by visual characteristics. The key information on the strength of image brightness and contrast was acquired through an experiment on human being’s image fidelity. Second, we devised a technique that does not require the original images by con- sidering the characteristics of the edge. Third, the experi- mental results demonstrate that the proposed method is robust to various image processing operations and to JPEG lossy compression and does not degrade the image quality. References Barni, M., Bartolini, F., & Piva, A. (2001). Improved wavelet-based watermarking through pixel-wisa masking. IEEE Transaction on Image Processing, 10(5), 783–791. Chen, L. H., & Lin, J. J. (2003). Mean quantization based image watermarking. Image and Vision Computing, 21, 717–727. Chu, W. C. (2003). DCT-based image watermarking using subsampling. IEEE Transaction on Multimedia, 5(1), 34–38. Chubb, C., Sperling, G., & Solomon, J. A. (1989). Texture interactions determine perceived contrast. Proceeding of the National Academy of Sciences, 86, 9631–9635. Cox, I., & Miller, M. L. (1997). A review of watermarking and the importance of perceptual modeling. Proceedings of Electronic Imaging, 3016, 92–99. Dai, Y. J., Zhang, L., & Yang, Y. X. (2003). A new method of MPEG video watermarking technology. Proceeding of IEEE ICCT, 2, 1845–1847. Hsu, C. T., & Wu, J. L. (1998). Multiresolution watermarking for digital images. IEEE Transaction on Circuits and Systems, 45(8), 1097–1101. Huang, J., Shi, Y. Q., & Shi, Y. (2000). Embedding image watermarks in DC components. IEEE Transaction on Circuits and Systems for Video Technology, 10(6), 974–979. Joo, S. H., Suh, Y. H., Shin, J. H., & Kikuchi, H. (2002). A new robust watermark embedding into wavelet DC component. ETRI Journal, 24(5), 401–404. Levicky, D., & Foris, P. (2004). Human visual system models in digital image watermarking. Radio Engineering, 13(4), 38–43. Lin, S. D., & Chen, C. F. (2000). A robust DCT-based watermarking for copyright protection. IEEE Transaction on Consumer Electronics, 46(3), 415–421. Miu, X. M., Lu, Z. M., & Sun, S. H. (2000). Digital watermarking of still images with gray-level digital watermarks. IEEE Transaction on Consumer Electronics, 46(1), 137–145. Moon, H. S., Sohn, M. H., & Jang, D. S. (2004). DWT-based image watermarking for copyright protection. In AIS 2004 (pp. 490–497). 686 H.S. Moon et al. / Expert Systems with Applications 32 (2007) 674–686 Olzak, L. A., & Laurinen, P. I. (1999). Multiple gain control processes in contrast–contrast phenomena. Vision Research, 39, 231–256. Shi, Y. Q., & Sun, H. (1999). Image and video compression for multimedia engineering: Fundamentals, algorithms, and standards. Boca Raton, FL: CRC Press. Shih, F. Y., & Wu, Y. T. (2005). Enhancement of image watermark retrieval based on genetic algorithms. Journal of Visual Communication and Image Representation, 16, 115–133. Stiles, W. S. (1978). Mechanisms of color vision. London: Academic Press. Swanson, M. D., Kobayashi, M., & Tewfik, A. H. (1998). Multimedia data embedding and watermarking technologies. Proceeding of the IEEE, 86(6), 1064–1087. Taskovski, D., Bogdanova, S., & Bogdanov, M. (2002). Improved low frequency image adaptive watermarking scheme. In ICSP’02 (pp. 26– 30). Wang, W., & Pearmain, A. (2004). Blind image data hiding based on self reference. Pattern Recognition Letters, 25, 1681–1689. Wu, Y. T., & Shih, F. Y. (2004). An adjusted-purpose digital watermark- ing technique. Pattern Recognition, 37(12), 2349–2359. Wyszecki, G., & Stiles, W. (1982). Color science: Concepts and methods, quantitative data and formulae. New York: Wiley. Expert system for low frequency adaptive image watermarking: Using psychological experiments on human image perception Introduction Literature review The embedding method DWT of the subimage generated by decomposing the original image Making edge mask Pixel-based pseudo-permutation of the watermark Determining the watermark strength according to image characteristics Modification of coefficients in the selected subimage using edge mask The extraction method Decomposing watermarked image and DWT Extracting the watermark using the edge mask Reversing the permutation and similarity measurement Experimental results and application Decision of watermark strength regarding image local characteristic The validity of edge mask Image processing operation JPEG lossy compression and application Conclusion References 
work_sais73suafgozanfdqttn5js4i	----	Fast Nonnegative Matrix Tri-Factorization for Large-Scale Data Co-Clustering Fast Nonnegative Matrix Tri-Factorization for Large-Scale Data Co-Clustering Hua Wang, Feiping Nie, Heng Huang, Fillia Makedon Department of Computer Science and Engineering University of Texas at Arlington, Arlington, Texas 76019, USA huawangcs@gmail.com, feipingnie@gmail.com, heng@uta.edu, makedon@uta.edu Abstract Nonnegative Matrix Factorization (NMF) based co- clustering methods have attracted increasing atten- tion in recent years because of their mathematical elegance and encouraging empirical results. How- ever, the algorithms to solve NMF problems usu- ally involve intensive matrix multiplications, which make them computationally inefficient. In this paper, instead of constraining the factor matrices of NMF to be nonnegative as existing methods, we propose a novel Fast Nonnegative Matrix Tri- factorization (FNMTF) approach to constrain them to be cluster indicator matrices, a special type of nonnegative matrices. As a result, the optimiza- tion problem of our approach can be decoupled, which results in much smaller size subproblems requiring much less matrix multiplications, such that our approach works well for large-scale input data. Moreover, the resulted factor matrices can di- rectly assign cluster labels to data points and fea- tures due to the nature of indicator matrices. In ad- dition, through exploiting the manifold structures in both data and feature spaces, we further intro- duce the Locality Preserved FNMTF (LP-FNMTF) approach, by which the clustering performance is improved. The promising results in extensive ex- perimental evaluations validate the effectiveness of the proposed methods. 1 Introduction Clustering, which partitions a data set into different groups unsupervisedly, is one of the most fundamental topic in statis- tical learning. Most traditional clustering algorithms are de- signed for one-side clustering [Nie et al., 2009], i.e. cluster ei- ther data points or features. However, in many real world ap- plications, the clustering based analysis is interested in two- side clustering results, i.e. group the data points and features simultaneously, e.g., “documents” and “words” in document analysis, “users” and “items” in collaborative filtering, “sam- ples” and “genes” in microarray data analysis, etc. Typically, instead of being independent, the different clustering tasks on data and features are closely correlated, and it is challeng- ing for traditional clustering algorithms to utilize the data and features interdependence efficiently. Consequently, co- clustering techniques, which aim to cluster both data and fea- tures simultaneously by leveraging the interrelations between them, have been proposed in recent researches. To name a few, Dhillon [Dhillon, 2001] introduced a bipartite spectral graph partition approach to co-cluster words and documents; Cho et al. [Cho et al., 2004] suggested to co-cluster the exper- imental conditions and genes for microarray data by minimiz- ing the sum-squared-residue; Long et al. [Long et al., 2006] presented a relation summary network model to co-cluster the heterogeneous data on a k-partite graph, and so on. More recently, Ding et al. [Ding et al., 2005] explored the relationships between Nonnegative Matrix Factorization (NMF) [Lee and Seung, 1999; 2001] and K-means/spectral clustering, and proposed to use Nonnegative Matrix Tri- factorization (NMTF) [Ding et al., 2006] to co-cluster words and documents at the same time. Due to its mathematical elegance and encouraging empirical results, NMTF method has been further developed to address various aspects of co- clustering [Wang et al., 2008; Li et al., 2010; Gu and Zhou, 2009; Ding et al., 2010]. However, a notorious bottleneck of NMTF based co-clustering approaches is the slow com- putational speed because of intensive matrix multiplications involved in each iteration step of the solution algorithms, which makes these approaches hard to be applied to large- scale data in real world applications. In this paper, we pro- pose a novel Fast Nonnegative Matrix Tri-factorization (FN- MTF) approach to efficiently conduct co-clustering on large- scale data. Our new algorithms are interesting from the fol- lowing perspectives: • Instead of enforcing traditional nonnegative constraints on the factor matrices of NMTF, we constrain them to be cluster indicator matrices, a special type of nonnega- tive matrices. As a result, the clustering results of our approach are readily stored in the resulted factor ma- trices. However, existing NMF based methods require an extra post-processing step to extract cluster struc- tures from the factor matrices, which often leads to non- unique clustering results. • Due to the nature of indicator matrices, the optimization problems of our approach can be decoupled into sub- problems with much smaller sizes, and the decoupled subproblems involve much less matrix multiplications. Therefore, our approach is computationally efficient and 1553 Proceedings of the Twenty-Second International Joint Conference on Artificial Intelligence scale well to large-scale input data. • Taking into account the manifold structures in both data and feature spaces, we further develop a Locality Pre- served FNMTF (LP-FNMTF) approach to incorporate manifold regularizations. Efficient algorithm to opti- mize the objective with quick convergence is presented. • Promising experimental results on five benchmark data sets show that our approaches not only are faster than state-of-the-art co-clustering methods but also have competitive clustering performance. Notations and problem formalization. Throughout this pa- per, we write matrices as boldface uppercase letters and vec- tors as boldface lowercase letters. Given a matrix M = (mij), its i-th row and j-th column are denoted as mi· and m·j respectively. Traditional clustering methods focus on one-side cluster- ing, i.e., clustering the data side based on the similarities along the feature side. In the co-clustering problem, we clus- ter data points based on the distributions of features, mean- while cluster features based on the distributions of the data points. Formally, given a data set X = { x·i ∈ R d }n i=1 , we write X = [x·1, . . . , x·n] = [ xT1·, . . . , x T d· ]T . Our goal is to group the data points {x·1, . . . , x·n} into c clusters {Cj} c j=1 , and simultaneously group the features {x1·, . . . , xd·} into m clusters {Wj} m j=1 . We use a partition matrix G = [ gT 1· , . . . , gTn· ]T ∈ {0, 1}n×c to represent clustering result of data points, such that gij = 1 if x·i belongs to cluster Cj and gij = 0 oth- erwise. Similarly, we use another partition matrix F =[ fT 1· , . . . , fTd· ]T ∈ {0, 1}d×m to represent the clustering re- sults of features. Here, we call F and G as cluster indicator matrices, because each row of them, i.e., fi·(1 ≤ i ≤ d) or gi·(1 ≤ i ≤ n), has one and only one element equal to 1 to indicate the cluster membership, while the rest elements are 0. We denote the set of all cluster indicator matrices as Ψ. 2 Fast nonnegative matrix tri-factorization (FNMTF) for co-clustering In this section, we first briefly review the background of co- clustering using NMTF, which motivates the optimization ob- jective of our approach. After that, an efficient algorithm to solve our objective will be introduced. 2.1 Objective of FNMTF approach K-means clustering is a standard clustering method in statis- tical learning, which minimizes the following objective: J1 = c∑ j=1 ∑ x ·i∈Cj ‖x·i − c·j‖ 2 = c∑ j=1 n∑ i=1 gij ‖x·i − c·j‖ 2 , s.t. G ∈ Ψn×c, (1) where ‖·‖ denotes the Frobenius norm of a matrix, c·j is the j-th centroid of the data set. Because G ∈ Ψ is a cluster indicator matrix, minimizing J1 is a combinato- rial optimization problem, which is hard to be resolved in general. Therefore, the minimization of J1 is often re- laxed to maximize the following objective [Zha et al., 2001; Ding and He, 2004]: J2 = tr ( G T X T XG ) , s.t. GT G = I . (2) Note that, G in J2 is no longer an indicator matrix, but an arbitrary orthonormal matrix. Recently, Ding et al. [Ding et al., 2005] proved the equiva- lence between the relaxed objective of K-means clustering in Eq. (2) and the NMF objective when orthonormal constraints are enforced on the factor matrices, which minimizes: J3 = ‖X − FG T ‖2, s.t. F ≥ 0, G ≥ 0, FT F = I, GT G = I, (3) where X ∈ Rd×n+ , F ∈ R d×c + and G ∈ R n×c + , and J3 aims to approximate the nonnegative data matrix X by the prod- uct of F and G. The orthonormal constraints here ensure the uniqueness (up to a permutation) of the solution, and together with the nonnegative constraints make the resulted F and G approximate the K-means clustering results on both features and data points (called as “soft labels”) [Ding et al., 2005; 2006]. The latter, simultaneously clustering the rows (fea- tures) and the columns (data points) of an input data matrix, is one of the main strength of NMF defined in Eq. (3). Because the two-factor NMF in Eq. (3) is restrictive, which often gives a rather poor low-rank matrix approximation, one more factor S ∈ Rm×c+ was introduced to absorb the different scales of X, F and G. This leads to NMTF [Ding et al., 2006]: X ≈ FSGT , where F ∈ Rd×m + and G ∈ Rn×c + . S provides increased degrees of freedom such that the low- rank matrix representation remains accurate, while F gives row clusters and G gives column clusters. In order to achieve additional flexibility, in clustering scenarios, the nonnegative constraint on X (thereby the nonnegative constraint on S) can be relaxed [Ding et al., 2010], which leads to the semi-NMTF problem minimizing the following objective: J4 = ‖X − FSG T ‖2, s.t. F ≥ 0, G ≥ 0, FT F = I, GT G = I . (4) Despite its mathematical elegance, Eq. (4) suffers from two problems that impede its practical use. First, similar to Eq. (2), the relaxations on F and G make the immediate out- puts of Eq. (4) are not cluster labels, which require an addi- tional post-processing step and often lead to non-unique solu- tions. Second, and more important, Eq. (4) is usually solved by alternately iterative algorithms, and in each iteration step the intensive matrix multiplications are involved [Ding et al., 2005; 2006; 2010; Wang et al., 2008; Li et al., 2010; Gu and Zhou, 2009]. As a result, it is infeasible to apply such algorithms to large-scale real world data due to the expensive computational cost. In order to tackle these difficulties, instead of solving the relaxed clustering problems as in Eqs. (2–4), we solve the original clustering problem similar to Eq. (1). Specifically, we constrain the factor matrices of NMTF to be cluster indi- cator matrices and minimize the following objective: J5 = ‖X−FSG T ‖2 s.t. F ∈ Ψd×m, G ∈ Ψn×c . (5) 1554 We call Eq. (5) as the proposed Fast Nonnegative Matrix Tri- factorization (FNMTF) approach. Note that, the orthonormal constraints on F and G are re- moved in our objective, as their purposes (unique solution and labeling approximation) are automatically accomplished by the new constraints. Surprisingly, with these new constraints, though more stringent, as shown theoretically shortly in this section and empirically later in Section 4, the computational speed of our approach can be significantly improved. 2.2 An efficient optimization algorithm Following the standard optimization procedures, we alter- nately solve the three variables F, S and G of J5 in Eq. (5). First, fixing F, G, and setting the derivative of J5 with respect to S as zero, we have S = ( F T F )−1 F T XG ( G T G )−1 . (6) Second, when F and S are fixed, the optimization problem to obtain G can be decoupled and we solve the following simpler problem for each i (1 ≤ i ≤ n): min G∈Ψ ‖x·i − FSg T i·‖ 2 . (7) Because gi· (1 ≤ i ≤ n) ∈ Ψ 1×c is a cluster indicator vector in which one and only one element is 1 and the rest are zeros, the solution to Eq. (7) can be easily obtained by: gij = { 1 j = arg mink ‖x·i − f̃·k‖ 2, 0 otherwise, (8) where F̃ = FS and f̃·k is the k-th column of F̃. Note that, Eq. (8) simply enumerates the c vector norms and seeks the maximum one, without involving any matrix multiplication. Finally, when G and S are fixed, the optimization prob- lem to obtain F can be similarly decoupled and we solve the following simpler problem for each j (1 ≤ j ≤ d): min F∈Ψ ‖xj· − fj·SG T ‖2 . (9) Again, since fj· (1 ≤ i ≤ d) ∈ Ψ 1×m is a cluster indicator vector for feature side, the solution to Eq. (9) is: fij = { 1 i = arg minl ‖xj· − g̃l·‖ 2, 0 otherwise, (10) where G̃T = SGT and g̃l· is the l-th row of G̃ T . The procedures to solve J5 are summarized in Algorithm 1. Due to the nature of alternating optimization, Algorithm 1 is guaranteed to converge to a local minima (existing NMF al- gorithms [Ding et al., 2005; 2006; 2010] also converges to a local minima because the objectives J3 and J4 are not convex in both variables F and G), and the proof is skipped due to space limit. As can be seen, in step 2 and step 3, the solutions are obtained by enumerating the vector norms, which is def- initely much faster than the matrix multiplication used in the existing NMF methods. Upon solution, G gives the cluster- ing results of data points, and F gives the clustering results of features directly. Algorithm 1: Algorithm to solve J5 in Eq. (5). Input: Data matrix X = [x·1, . . . , x·n] ∈ R d×n. Initialize G ∈ Ψn×c and F ∈ Ψd×m with arbitrary class indicator matrices; repeat 1. calculate S by Eq. (6) ; 2. calculate G by Eq. (8) ; 3. calculate F by Eq. (10) ; until converges; Output: Indicator matrices G for data point clustering and F for feature clustering. 3 Locality preserved FNMTF (LP-FNMTF) Recent researches showed that many real world data are sam- pled from the nonlinear manifolds which are embedded in the high dimensional ambient space [Belkin and Niyogi, 2002]. However, similar to traditional NMF and NMTF, the pro- posed FNMTF approach assumes that the data points and features are sampled from Euclidean spaces, and fails to dis- cover the intrinsic geometrical and discriminative data and feature structures. Therefore, we further develop our FN- MTF approach and propose the Locality Preserved FNMTF (LP-FNMTF) approach to enforce two geometrically based regularizers from both data and feature sides. 3.1 Manifold regularization and the optimization objective of LP-FNMTF method Because we co-cluster an input data matrix on both data and feature dimensions, we consider two undirected graphs, one constructed from data points, denoted as Gd, and the other one from features, denoted as Gf . The corresponding affin- ity matrices Wd and Wf could be either computed from the input data matrix X (e.g., as in [Gu and Zhou, 2009]), or obtained from prior knowledge. According to manifold as- sumption [Belkin and Niyogi, 2002], the regularization terms to measure the smoothness with respect to the intrinsic mani- folds of data points and features are given by [Cai et al., 2008; Gu and Zhou, 2009]: min G∈Ψ tr ( G T LdG ) , and min F∈Ψ tr ( F T Lf F ) , (11) where Ld = I − D − 1 2 d WdD − 1 2 d is the the normalized graph Laplacian of Gd, and Dd is the degree matrix of Gd; similarly, Lf = I − D − 1 2 f Wf D − 1 2 f , and Df is the degree matrix of Gf . Incorporating Eq. (11) into Eq. (5), the objective of LP- FNMTF approach is to minimize: J6 = ‖X − FSG T ‖2 + α tr ( G T LdG ) + β tr ( F T Lf F ) , s.t. F ∈ Ψd×m, G ∈ Ψn×c, (12) where α and β are regularization parameters to balance the reconstruction error of co-clustering in the first term and la- beling smoothness in the data point space and feature space in the second and third terms, respectively. Because F and G are constrained to be cluster indicator matrices, it is difficult to solve Eq. (12) in general. Hence we simplify this problem using the following proposition. 1555 Proposition 1 Given a symmetric matrix A and its eigen- decomposition A = PΣPT , where Σ ∈ Rc×c is a diagonal matrix with diagonal elements as the c largest eigenvalues, and P is the corresponding eigenvector matrix, the following two optimization problems are equivalent: (P1) : min C∈Ψ tr [ C T (I − A) C ] , (13) (P2) : min C∈Ψ, QT Q=I ‖C − BQ‖2, (14) where Q is an arbitrary orthonormal matrix and B = PΣ1/2 . (15) Proof. (P1) is equivalent to maxC∈Ψ tr ( CT AC ) that is further equivalent to minC∈Ψ ‖CC T − A‖2. By defini- tion the low-rank approximation of A is given by A = BQ (BQ) T , thus (P1) becomes minC∈Ψ,QT Q=I ‖CC T − BQ (BQ) T ‖2. C approximating BQ is equivalent to CCT approximating BQ (BQ) T . Hence, solving (P1) in Eq. (13) can be reasonably transformed to solve (P2) in Eq. (14), which completes the proof of Proposition 1. � Applying Proposition 1 in Eq. (12), the objective of our LP-FNMTF approach is transformed to minimize: J7 = ‖X − FSG T ‖2 + α‖G − BdQd‖ 2 + β‖F − BfQf ‖ 2, s.t. F ∈ Ψd×m, G ∈ Ψn×c, QTd Qd = I, Q T f Qf = I, (16) where Bd and Bf are computed from Ld and Lf following the procedures described in Proposition 1. 3.2 Optimization algorithm Again, we use the alternating iterative method to solve Eq. (16). We first introduce the following theorem. Theorem 1 Let H = BT C and the Singular Value Decom- position (SVD) of H be given by H = UΛVT . Fixing C and B, the optimum Q to problem (P2) defined in Eq. (14) is given by Q = UVT . Proof. When C is fixed, problem (P2) in Eq. (14) is equiva- lent to maxQT Q=I tr ( QT H ) . We have tr ( QT H ) = tr ( QUΛVT ) = tr ( ΛVT QU ) = tr (ΛZ) = ∑ i λiizii, where Z = V T QU and λii and zii are the (i, i)-th entry of Λ and Z, respectively. Note that Z is orthonormal, i.e., ZT Z = I, thus zii ≤ 1. On the other hand, λii ≥ 0 as λii is singular value of H. Therefore, tr ( QT H ) = ∑ i λiizii ≤ ∑ i λii, and when zii = 1 (1 ≤ i ≤ c), the equality holds. That is to say, tr ( QT H ) reaches its maximum when Z = I. Recall that Z = VT QU, the solution to maxQT Q=I tr ( QT H ) or (P2) is Q = UZT VT = UVT . Theorem 1 is proved. � Now, we solve Eq. (16). First, fixing F and G, by setting the derivative of J7 with respect to S as 0, we obtain S = ( F T F )−1 F T XG ( G T G )−1 . (17) Second, fixing F, G and S, we can decouple Eq. (16) into two following subproblems: min QT d Qd=I ‖G − BdQd‖ 2, min QT f Qf =I ‖F − Bf Qf‖ 2 . (18) Applying Theorem 1, Qd = UdV T d where Ud and Vd are obtained by SVD on BTd G; Qf = Uf V T f where Uf and Vf are obtained by SVD on BTf F. Third, we fix S, F and Qd to update G. Because G is a cluster indicator matrix, Eq. (16) is decoupled to the follow- ing simpler problems for each 1 ≤ i ≤ n: min G∈Ψ ‖x·i − F̃g T i·‖ 2 + α‖gi· − ( b̃d ) i· ‖2, (19) where F̃ = FS, ( b̃d ) i· denotes the i-th row of B̃d = BdQd. Thus, the solution can be obtained by gij = { 1 j = arg mink ( ‖x·i − f̃·k‖ 2 − 2α ( B̃d ) ik ) , 0 otherwise . (20) Finally, when fixing S, G and Qf , let B̃f = Bf Qf , G̃ T = SGT and g̃l· is the l-th row of G̃ T , we similarly obtain F as: fij = ⎧⎨ ⎩1 i = arg minl ( ‖xj· − g̃l·‖ 2 − 2β ( B̃f ) jl ) , 0 otherwise. (21) The procedures to solve J7 are summarized in Algorithm 2. Again, because step 4 and step 5 only involve vector norm enumeration without matrix multiplication, our algorithm is more computationally efficient. Empirical results show that the convergence of our algorithm is fast, which make it fea- sible to solve the large-scale real world problems via our ap- proach in practice. Algorithm 2: Algorithm to solve J7 in Eq. (16). Input: Data matrix X = [x·1, . . . , x·n] ∈ R d×n. 1. Initialize G ∈ Ψn×c and F ∈ Ψd×m with arbitrary class indicator matrices; 2. Calculate Bd and Bf from Ld and Lf following the description of Proposition 1; repeat 1. Calculate S by Eq. (17) ; 2. Calculate Qd = UdV T d where Ud and Vd are obtained by SVD on BTd G ; 3. Calculate Qf = Uf V T f where Uf and Vf are obtained by SVD on BTf F ; 4. Calculate G by Eq. (20) ; 5. Calculate F by Eq. (21) ; until Converges; Output: Class indicator matrices G and F for data and feature clustering tasks, respectively. 4 Experiments In this section, we evaluate the proposed FNMTF and LP- FNMTF approaches, and compare them against state-of-the- art (co-)clustering methods, including Semi-NMF (SNMF) [Ding et al., 2010], Orthogonal NMTF (ONMTF) [Ding et al., 2006], Graph regularized NMF (GNMF) [Cai et al., 2008] 1556 Table 1: Description of experimental data sets Data sets # sample # feature # classes Coil20 1140 1024 20 WebKB 4199 1000 4 WebACE 2340 1000 20 CSTR 476 1000 4 RCV1 193844 1979 103 and Dual Regularized Co-clustering (DRCC) [Gu and Zhou, 2009] methods. We also report the clustering results by K- means and NMF [Lee and Seung, 2001] methods as baselines. In our experiments, we use 5 data sets to evaluate the com- pared methods, which are summarized in Table 1. The first four are widely used as benchmarks in clustering literatures [Ding et al., 2006; Cai et al., 2008; Gu and Zhou, 2009; Ding et al., 2010]. The last one has very large sample size and feature size [Chen et al., 2010]. In order to run the exper- iments on contemporary computers, for RCV1 data set, fol- lowing previous studies, we remove the keywords (features) appearing less than 100 times in the corpus, which results in 1979 (out of 47236) keywords in our experiments. To evaluate the clustering results, we adopt three widely used standard metrics: clustering accuracy [Cai et al., 2008], normalized mutual information (NMI) [Cai et al., 2008] and cluster purity [Ding et al., 2006]. 4.1 Clustering results Experiments setup. Because each clustering algorithm has one or more parameters to be tuned, in order to compare fairly, we run these algorithms under different parameter set- tings, and select the best average result for each one. We set the number of clusters as the true number of classes for all clustering algorithms on all data sets. For co-clustering methods, including ONMTF, DRCC and our two methods, the number of feature clusters is set to be the same as that of data clusters, i.e., m = c. For manifold regularized methods, including GNMF, DRCC and our LP-FNMTF methods, we construct nearest- neighbor graph following [Gu and Zhou, 2009], where the neighborhood size for graph construction is set by searching the grid of {1, 2, . . . , 10}, and the regularization parameters (i.e., α and β in Eq. (16)) are set by searching the grid of {0.1, 1, 10, 100, 500, 1000}. Given the number of clusters, no parameter selection is needed for K-means, NMF and SNMF methods. Results. Under each parameter setting of every method in comparison, we repeat clustering 50 times, and the average result is computed. We report the best average result for each method on five data sets in Table 2. From the results in Table 2, we can see that the proposed FNMTF and LP-FNMTF methods consistently outperform the other compared methods, sometimes very significantly, which demonstrate the advantage of our approaches in terms of clustering performance. A more careful examination on the results shows that, the co-clustering methods, including ONMTF, DDRC and our proposed FNMTF and LP-FNMTF methods, generally achieve better clustering results, which Table 3: Average iteration numbers to converge of compared clustering methods. Data Coil20 WebKB WebACE CSTR RCV1 Kmeans 26.4 29.2 28.3 24.2 87.1 NMF 30.3 40.2 38.5 30.1 92.5 SNMF 37.7 45.5 44.1 36.3 98.6 ONMTF 41.2 50.3 49.2 40.3 104.4 GNMF 50.2 62.1 61.1 46.8 112.5 DRCC 51.6 64.4 62.3 47.9 129.1 FNMTF 14.1 16.5 15.9 14.3 45.2 LP-FNMTF 15.2 17.2 16.3 15.6 48.1 0 2 4 6 8 x 10 4 Km ea ns NM F SN MF ON MT F GN MF DR CC FN MT F LP −F NM TF (a) WebKb data set. 0 0.5 1 1.5 2 2.5 x 10 6 Km ea ns NM F SN MF ON MT F GN MF DR CC FN MT F LP −F NM TF (b) RCV1 data set. Figure 1: Convergence time (ms) of compared clustering methods on WebKB and RCV1 data sets. are consistent with the widely accepted hypothesis that clus- tering of features can help clustering of data points. Finally, LP-FNMTF method is superior to FNMTF method on all the data sets except CSTR. This indicates that exploiting the ge- ometric structures in data and feature spaces indeed can im- prove the cluster performance, which verifies manifold as- sumption and confirms the correctness of our algorithms. 4.2 Studies of computational speeds In this subsection, we evaluate the computational speeds of the compared (co-)clustering methods. All our experiments are performed on a Dell PowerEdge 2900 server, which has two quad-core Intel Xeon 5300 sequence CPU processors at 3.0 GHz and 48G bytes memory. We first examine the convergence rates of the compared methods. We repeat clustering 50 times by each method with its optimal parameters on each data set. The average itera- tion numbers of each method on each data set are reported in Table 3. The results show that the proposed FNMTF and LP- FNMTF approaches require much less iterations to converge, therefore they are more computationally efficient. In addition, we also report the average convergence time of the compared methods on WebKB data and RCV1 data sets as in Figure 1. From the results, we can see that our FNMTF and LP-FNMTF methods are only slower than K-means method while much faster than all other state-of-the-art clustering methods. These results are consistent with our theoretical analysis that our methods are implemented on subproblems with much smaller sizes and use much less matrix multipli- cations. Therefore, our approaches are suitable for clustering on large-scale data. The results on three other smaller data sets are not shown due to space limit, from which the same observations can be seen. 1557 Table 2: Clustering results measured by accuracy/NMI/purity of the compared methods. Data Metrics Kmeans NMF SNMF ONMTF GNMF DRCC FNMTF LP-FNMTF Coil20 Accuracy 0.495 0.487 0.527 0.635 0.665 0.680 0.696 0.725 NMI 0.489 0.479 0.511 0.561 0.549 0.566 0.584 0.621 Purity 0.441 0.437 0.468 0.478 0.481 0.483 0.497 0.512 WebKB Accuracy 0.698 0.668 0.621 0.685 0.717 0.725 0.774 0.792 NMI 0.467 0.427 0.418 0.455 0.458 0.487 0.501 0.522 Purity 0.601 0.595 0.604 0.664 0.672 0.674 0.696 0.712 WebACE Accuracy 0.526 0.514 0.527 0.635 0.665 0.680 0.696 0.714 NMI 0.519 0.512 0.538 0.587 0.556 0.571 0.604 0.611 Purity 0.479 0.481 0.491 0.487 0.511 0.493 0.517 0.532 CSRT Accuracy 0.763 0.759 0.699 0.771 0.742 0.812 0.894 0.847 NMI 0.654 0.668 0.614 0.673 0.635 0.681 0.753 0.722 Purity 0.612 0.587 0.614 0.645 0.637 0.656 0.701 0.682 RCV1 Accuracy 0.168 0.156 0.173 0.196 0.201 0.213 0.238 0.241 NMI 0.274 0.274 0.283 0.301 0.312 0.318 0.339 0.342 Purity 0.134 0.121 0.139 0.146 0.152 0.159 0.172 0.179 5 Conclusions In this work, we proposed a novel Fast Nonnegative Matrix Tri-factorization (FNMTF) method to simultaneously cluster both data side and feature side of an input data matrix. We adopt the idea of NMF/NMTF based co-clustering methods, but constrain the factor matrices to be cluster indicator matri- ces, a special type of nonnegative matrices. Through the new constraints, the optimization problem of our method is decou- pled into a number of much smaller subproblems that require much less matrix multiplication than the existing NMF based co-clustering algorithms, which makes our approaches of par- ticular use for real world large-scale data. We further de- veloped our method to incorporate manifold information and proposed Locality Preserved FNMTF (LP-FNMTF) method. We conducted extensive experiments on benchmark data sets that demonstrate promising results of our methods, which is consistent with our theoretical analysis. Acknowledgments This research was supported by NSF-CNS 0923494, NSF-IIS 1041637, NSF-CNS 1035913. References [Belkin and Niyogi, 2002] M. Belkin and P. Niyogi. Lapla- cian eigenmaps and spectral techniques for embedding and clustering. In NIPS, 2002. [Cai et al., 2008] D. Cai, X. He, X. Wu, and J. Han. Non- negative matrix factorization on manifold. In ICDM, 2008. [Chen et al., 2010] W.Y. Chen, Y. Song, H. Bai, C.J. Lin, and E.Y. Chang. Parallel spectral clustering in distributed sys- tems. IEEE TPAMI, 2010. [Cho et al., 2004] H. Cho, I.S. Dhillon, Y. Guan, and S. Sra. Minimum sum-squared residue co-clustering of gene ex- pression data. In SDM, 2004. [Dhillon, 2001] I.S. Dhillon. Co-clustering documents and words using bipartite spectral graph partitioning. In SIGKDD, 2001. [Ding and He, 2004] C. Ding and X. He. K-means clustering via principal component analysis. In ICML, 2004. [Ding et al., 2005] C. Ding, X. He, and H.D. Simon. On the equivalence of nonnegative matrix factorization and spec- tral clustering. In SDM, 2005. [Ding et al., 2006] C. Ding, T. Li, W. Peng, and H. Park. Or- thogonal nonnegative matrix tri-factorizations for cluster- ing. In SIGKDD, 2006. [Ding et al., 2010] C. Ding, T. Li, and M.I. Jordan. Convex and semi-nonnegative matrix factorizations. IEEE TPAMI, 32(1):45–55, 2010. [Gu and Zhou, 2009] Q. Gu and J. Zhou. Co-clustering on manifolds. In SIGKDD, 2009. [Lee and Seung, 1999] D.D. Lee and H.S. Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401(6755):788–791, 1999. [Lee and Seung, 2001] D.D. Lee and H.S. Seung. Algo- rithms for non-negative matrix factorization. In NIPS, 2001. [Li et al., 2010] T. Li, V. Sindhwani, C. Ding, and Y. Zhang. Bridging Domains with Words: Opinion Analysis with Matrix Tri-factorizations. In SDM, 2010. [Long et al., 2006] B. Long, X. Wu, Z.M. Zhang, and P.S. Yu. Unsupervised learning on k-partite graphs. In SIGKDD, 2006. [Nie et al., 2009] Feiping Nie, Dong Xu, Ivor W. Tsang, and Changshui Zhang. Spectral embedded clustering. In IJ- CAI, pages 1181–1186, 2009. [Wang et al., 2008] F. Wang, T. Li, and C. Zhang. Semi- supervised clustering via matrix factorization. In SDM, 2008. [Zha et al., 2001] Hongyuan Zha, Xiaofeng He, Chris Ding, Horst Simon, and Ming Gu. Spectral relaxation for k- means clustering. In NIPS, 2001. 1558 
work_sb2bp5b7yndx7gh2eyj26kewbe	----	Neural networks in process life cycle profit modelling Neural networks in process life cycle profit modelling Teemu Räsänen a,*, Risto Soukka b, Sami Kokki b, Yrjö Hiltunen a a Department of Environmental Science, University of Kuopio, P.O. Box 1627, FIN-70211, Kuopio, Finland b Department of Energy and Environmental Technology, Lappeenranta University of Technology, P.O. Box 20, FIN-53851, Lappeenranta, Finland Abstract Changes in operational environment of the process industry such as decreasing selling prices, increased competition between compa- nies and new legislation, set requirements for performance and effectiveness of the industrial production lines and processes. For the basis of this study, a life cycle profit (LCP) model of a pulp process was constructed using different kind of process information including chemical consumptions and production levels of material and energy flows in unit processes. However, all the information needed in the creation of relevant LCP model was not directly provided by information systems of the plant. In this study, neural networks was used to model pulp bleaching process and fill out missing information and furthermore to create estimators for the alkaline chemical consumption. A data-based modelling approach was applied using an example, where factors affecting the sodium hydroxide consump- tion in the bleaching stage were solved. The results showed that raw process data can be refined into new valuable information using computational methods and moreover to improve the accuracy of life cycle profit models. � 2007 Elsevier Ltd. All rights reserved. Keywords: Life cycle profit modelling; Neural network; Multi-layer perceptron; Variable selection; Pulp industry 1. Introduction In the process industry the development of competitive- ness of an existing production line has become a lifeline for many production units, because competition between companies has increased, new legislation has been pub- lished and the consequences of disturbances are bigger than before. Thus, it is essential to find process developing possibilities to increase profits and maximize returns and at the same time eliminate risks. These new demands set requirements not only for performance and effectiveness of processes but also for optimizing process life cycle profits. However, the construction of large economical models of processes is too complicated using only process models which are based on natural laws. The connections between the units of the processes in large scale plants are usually so complex that modelling based on natural laws is unable to produce all the needed information for a decision-making process. On the other hand, modelling on the basis of nat- ural laws requires lots of resources, like time and staff resources as well as economic and computational resources, and due to the these reasons modelling on sche- dule is often not possible (Soukka, 2007). The applicability of life cycle modelling principles to the modelling of increasingly complex processes has been examined. Life cycle modelling (life cycle assessment and life cycle costing) collects data on the entire life cycle. For this reason it is obvious that everything cannot be modelled on the basis of natural laws. Life cycle modelling requires the easy collection of an immense amount of infor- mation on the different stages of the process life cycle, so that it is impossible to produce vast amounts of measure- ment data due to the restrictions of the resource reserved for the report. Thus life cycle assessment also uses esti- mated, calculated and literature-based information (Sou- kka, 2007). 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.07.006 * Corresponding author. Tel.: +358 44 7162337; fax: +358 17 163191. E-mail addresses: teemu.rasanen@uku.fi (T. Räsänen), risto.soukka@ lut.fi (R. Soukka), sami.kokki@lut.fi (S. Kokki), yrjo.hiltunen@uku.fi (Y. Hiltunen). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 35 (2008) 604–610 Expert Systems with Applications mailto:teemu.rasanen@uku.fi mailto:risto.soukka@ lut.fi mailto:risto.soukka@ lut.fi mailto:sami.kokki@lut.fi mailto:yrjo.hiltunen@uku.fi The life cycle profit (LCP) analysis has been developed for the purpose of evaluating the economic impacts of process changes with a model based on the production principles of information applicable in life cycle modelling. A life cycle profit model is based on several kinds of infor- mation. The measurement data of one point can be used to constitute a statistic distribution. The estimated informa- tion and data from literature are in turn information that can be regarded as unique (Soukka, 2007). Among the other things, data-based computational methods can be used in the production of calculated information and, par- ticularly, in the identification of consequence-related effects including solving material and energy flows of the process (Räsänen et al., 2006). It is essential that life cycle profit model has been con- structed so that it reveals how changes are affecting reve- nues and costs of a production line. On the one hand, life cycle revenues can be increased by removing bottlenecks and production failures. These problems contain a high risk to cause production losses. On the other hand, costs can be reduced for example by affecting energy, chemical or water consumption of processes and the amount of efflu- ents from processes. Raw process data is one of the LCP models information sources and it can be used various ways. Industrial pro- cesses produce typically a huge amount of measured data describing each process performance. Several studies (Hei- kkinen, Kettunen, Niemitalo, Kuivalainen, & Hiltunen, 2005; Hiltunen et al., 2006) have shown that data-driven approaches are a fruitful way of developing for example the process state monitoring. Archived process data is also an important resource for the knowledge management of the process and it can be used for the optimization and improvement of productivity. In this study, we combine data-driven modelling and process life-cycle profit modelling into more intelligent sys- tem to produce valuable and more accurate new informa- tion to decision making. The proposed methods were applied and tested using the data set of an industrial scale pulp bleaching process. 2. The process and the data In this study life-cycle profit model (Soukka, 2007) was constructed for a paper mill, which produces newsprint and directory paper, coated and uncoated fine paper, core- board and spruce timber. The paper mill contains also pulp production and refined mechanical and ground wood pulp are produced for newsprint paper production. Also ECF- pulp is produced for the fine paper production. The life cycle profit model covers the whole production line of the pulp. However, in this study we concentrated on solving the missing or unknown correlations between pro- cess variables in the pulp bleaching process which is shortly described in Fig. 1. The objective of bleaching is to improve the brightness and the cleanliness of the pulp. The raw process data was originated from measuring systems and it was collected from pulp mill databases. The used data set contained 35 variables and 1500 rows. The data set consists of the time period where pine was used as pulp raw material and the time resolution was one hour. Variables were selected by process experts. 3. Methods 3.1. Life-cycle profit modelling The main purpose of the plants life cycle profit (LCP) model is to recognize the development possibilities, which can be achieved by process changes allocated to different stages of the process. The LCP model shows how life cycle profit varies in consequence of the process changes and owing to this it is possible to prioritize different possibilities to develop the process. The LCP analysis is originally used for modelling existing production lines but the model can also be used for strategic planning of new production lines. Therefore, it is possible to evaluate risks that can be actu- alized if big changes in the proportion of prices happen. Hunkeler (1999) has been presented that Life Cycle Profit- ability is an evolutionary means to conduct business prac- tice under scenarios where envirotechnical imperatives Fig. 1. Pulp bleaching stage contains several stages because target brightness of the end product cannot be achieved in only one bleaching step. Pulp is also washed between each stage. T. Räsänen et al. / Expert Systems with Applications 35 (2008) 604–610 605 https://isiarticles.com/article/1515 
work_sbd5jdz4wjbu5e5uvurw7sz2pe	----	doi:10.1016/j.eswa.2006.02.025 www.elsevier.com/locate/eswa Expert Systems with Applications 32 (2007) 1004–1010 Expert Systems with Applications Automatic timbre quality evaluation in Chinese traditional flute industry Hui-Jen Yang a,*, Yun-Long Lay b, Chern-Sheng Lin c a Department of Information Management, National Chinyi Institute of Technology, No. 35, Lane 215, Sec. 1, Chung-San Road, Taiping City, Taichung Hsien 411, Taiwan b Department of Electronic Engineering, National Chinyi Institute of Technology, No. 35, Lane 215, Sec. 1, Chung-San Road, Taiping City, Taichung Hsien 411, Taiwan c Department of Automatic Control Engineering, Feng Chia University, No. 100 Wenhwa Road, Seatwen, Taichung, Taiwan 40724, ROC Abstract Since the opening of Chinese economic, a great amount of Taiwanese merchants enter Mainland China for investment. The annual production of bamboo flutes in China is over 4 millions to provide domestic and international consumers. Owing to the great amount of demand, folk instrument industry is one of the attractive investment items. In order to increase the product manufacturing process and promote the product quality, the implementation and application of an automatic timbre evaluation system with spectrum analysis and cluster algorithm to increase the flute quality becomes an important method for Taiwanese businessmen to earn more profits. This research is to implement an automatic timbre quality system to evaluate the timbre quality of Chinese flute in flute businesses. From research data displays that the automatic timbre evaluation system does help the traditional handicraft manufacture classify the timbre quality of flute more effectively. � 2006 Elsevier Ltd. All rights reserved. Keywords: Automatic quality evaluation system; Timbre; Flute; Spectrum analysis; Cluster algorithm 1. Introduction The bamboo industry is one of 10 big emerging forestry development industries in mainland China. Since 1996, the total annual productivity of bamboo industry has amounted to 20 billion Yuan, is the head of 10 big emerg- ing industrials and also becomes the new benchmark in the Chinese forestry development. Currently, the annual pro- duction of bamboo flutes in China is over 4 millions to pro- vide domestic and international consumers need. The sudden appearance of bamboo business has demonstrated the bright future for Chinese bamboo industry. China is the earliest country to use the bamboo as material to make 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.02.025 * Corresponding author. Tel.: +011 886 0423924505x7923/7999; fax: +011 886 0423923725. E-mail addresses: yanghj@chinyi.ncit.edu.tw (H.-J. Yang), layyl@ chinyi.ncit.edu.tw (Y.-L. Lay), lincs@auto.fcu.edu.tw (C.-S. Lin). bamboo artwork. The distribution of bamboo in mainland China is over 20 provinces, the cities and the autonomous regions. Among them, there are 20 provinces and 141 coun- ties as main areas to produce the bamboo. The number and area of bamboo occupies around one-fourth all over the world. The types of variety and the broad usages in bam- boo have an incomparable superiority with other countries. The bamboo handicraft is the China traditional product, which consumes few resources and has high product eco- nomic value. Bamboo handicraft is a potential industry in China at present. The economic value of the bamboo product is generally above 10 times and sometimes even may reach 100 times. The Taiwanese businessman under- stood the economical pulsation of mainland China, there- fore a great amount of Taiwanese businessmen enter the mainland China for investment. In order to increase the product manufacturing process and promote the product quality, the development and application of an automatic mailto:yanghj@chinyi.ncit.edu.tw mailto:layyl@ chinyi.ncit.edu.tw mailto:layyl@ chinyi.ncit.edu.tw mailto:lincs@auto.fcu.edu.tw H.-J. Yang et al. / Expert Systems with Applications 32 (2007) 1004–1010 1005 timbre evaluation system becomes an important way to enhance the bamboo flute quality. In recent years, with the rapid development in computer and information technology, automatic quality manage- ment systems have been introduced in many manufacturing industries to improve or guarantee the quality of processes and products of organization (Abbott, 1999; Carvalho, Araujo, & Dourado, 1999; Kuo & Wu, 2002; Swanepoel, 2004; Van Harten, Casparie, & Fisscher, 2000). More recently, manufacturing industries have been applying information and communications technology for both information processing and automatic control. However, little evidence is shown in literature for automatic timbre quality evaluation in the traditional Chinese bamboo flute manufacturing industry. The systematic implementing of an automatic quality control system of Chinese bamboo flute can be useful in guaranteeing the timbre quality of flute to reduce the defect rate. Normally, some kinds of quality evaluation work need well-experienced experts to judge the quality of the products, such as wine taste, jew- elry authentication, painting evaluation and so on. By the same token, to evaluate the timbre quality of Chinese bam- boo flute relies heavily on the experience of the technician to judge the timbre quality of flute subjectively. To deter- mine the timbre quality of flute is a time-consuming pro- cess that requires manual subjective judgment, both by flute musicians and by experienced technicians in flute pro- duction factories. In most flute production factories, there are technicians specifically devoted to classifying the type of timbre of flute. The classification process is not uniform, as performing this task depends on the technician’s experi- ence, skill, or even on personal emotion or circumstance. The term ‘‘quality’’ implies a degree of excellence or superiority, value, conforming to specifications, meeting or exceeding customer expectations, suitability and usabil- ity of a product (Abbott, 1999; Verdu Jover, Llorens Mon- tes, & Fuentes, 2004). People like to use all of their senses to evaluate quality, such as smell, taste, touch, hearing and so on. Thus, quality is not a single, well-defined attribute but comprises many characteristics or properties relying on the different context, thus, there is no consistent defini- tion of quality in previous studying (Abbott, 1999; Neu- feldt, 1988). Some focus on the view of product orientation to define quality and some focus on the view of customer orientation to define quality. Some encompass both concepts; this is the definition applied in this study. In general, quality comprises many properties or features, such as sensory properties (appearance, texture, taste and aroma), chemical constituents, mechanical properties, functional properties, defects and so on (Abbott, 1999). Applying this concept to Chinese flute, a good quality flute should possess the feature not only of accuracy of pitch within the whole key, but also producing a clear note with little effort and having a beautiful sound timbre. From the perspective of manufacturer and performer, testifying the accuracy of pitch in flute is probably not a difficult thing. Whereas choosing or making a satisfactory timbre of flute is probably not an easy task for manufacturers or custom- ers. The attributes of sound timbre are: bright, dark, clear, soft, etc., which are too imprecise to justify. Timbre as assessed by people can not be precisely defined, leaving a high degree of uncertainty depending on an individual’s preferences or the unknown influences of individual values on the timbre quality of musical sounds (Kostek, 1999). Timbre is an auditory attribute considered as a percep- tional parameter of sound that is complex and multidimen- sional. Timbre is the feature of sound that allows people to identify it and distinguish it from other sounds with the same pitch and loudness (Houtsma, 1997). ‘‘Timbre’’ is defined as the quality, which distinguishes two sounds with the same pitch, loudness, and duration (Kostek, 1999). Often timbre is used to refer to the ‘‘tone quality’’ or ‘‘tone color’’ (Ilmoniemi, Valimaki, & Huotilainen, 2004). Thus, the lack of a standard automatic quality control or evalu- ation system to distinguish the timbre quality may lead to severe economic losses or wrong assessment of products due to the possible misuse of an inaccurate method or unin- tentional human error. For decades, the timbre quality evaluation of a flute has been subject to the judgment of experts or professional technicians; experts or professional technicians would blow each flute one by one to test its timbre. This process is labor-intensive and time-consuming and the results are very subjective and non-standardized depending on the preferences of each expert or professional technician. In order to objectively assess timbre quality of flute, there is a need to find a method to make it possible to objectively judge the subjective notions of timbre in Chinese Bamboo flute. With the advent of manufacturing technology toward automation, an automatic timbre evaluation system is nec- essary to classify the quality of flute. Essentially, the auto- matic timbre system was used to objectively classify and standardize the timbre of flute in order to be used as an automated evaluation process. The conceptual design of the system in this paper is composed of audio spectrum sys- tem, cluster analysis and decision making scheme to judge the timbre types of flute. Initially, a spectrum module is used to collect the features of each timbre as the resources of classification. Subsequently, for exploring the classifica- tion of flute, a cluster analysis with K-means algorithm was used to infer the class of timbre feature distribution. Finally, a condition decision algorithm is presented to judge the classification of timbre. The results reveal that an automatic-timbre-quality eval- uation system can not only be employed to support the flute technician in objectively judging the timbre quality with a minimal time and effort, but can simultaneously pro- vide valuable information to prove the timbre quality of the flute and increase its economic benefits. 2. Background and characteristics of flute As has been pointed out, the aim of the application shown in this paper is to classify the timbre quality of flute Fig. 1. Framework of the ADQTE system. 1006 H.-J. Yang et al. / Expert Systems with Applications 32 (2007) 1004–1010 automatically. Generally speaking, to make a good quality flute is not an easy task because of the multidimensional and complex nature of the flute itself. Western musical instrument are often made of metal. Metal is a very versatile material, so it can meet different individual requirements. Therefore, the made-up instru- ment has a very standard production process. For instance, a western flute has a standard mechanical production pro- cess. Thus, its inner tube can be well made with a pure cyl- inder. However, flute uses bamboo as a material and bamboo is a natural-born material. It is hard for us to make a standardized and unified raw material. The type and age of bamboo material has a large effect on timbre and intensity level (Di-Yin Net, 2005). As to the timbre evaluations of hard or soft, thin or thick, bright or dark, everyone has different evaluation scales or interests. Thus, the evaluation of timbre in flute is a subjective judgment. However, the attributes of tim- bre on flute should focus on ‘‘sweet, thick, bright, trans- parent, and soft’’ (Chen, 2002). Notice that the quality of timbre of flute is based on the musicians’ or techni- cians’ subjective judgment. In this paper, four indexes related to timbre in flute are used to estimate the quality of timbre, including sweet, transparent, thick and bright. Although the study of musical instrument acoustics holds that different musical instruments have different principles of vocal, there are still some common methods. From the perspective of musical instruments, the vocal part and the resonance part should be separated into dif- ferent system initially, and then their coupling effect should be discussed. The physical characteristics of reso- nance are usually described by input impedance curve. The impedance measurement was originally designed for measurement of the human vocal tract (Epps, Smith, & Wolfe, 1997). It was used for studies on the Boehm flute, the classical flute, and the baroque flute (Wolfe, Smith, Tann, & Fletcher, 2001; Wolfe & Smith, 2001). The vocal part of tube musical instruments can be separated into two sections with air splitting on the edge of plane sur- face. The resonance part of a tube musical instrument is the air in its inner tube. The way to measure its input impedance is to place a steady sine wave sound source at the opening of the inner tube. This transmits a sound into the inner tube to test its intensity level. We improve the impedance measurement by adding the spectrum anal- ysis and clustering algorithm to do the automatic timbre judgment. The blowing hole, vocal part, and body of a Chinese flute is a same piece of a bamboo, thus the method to test the timbre is to blow the flute and check the timbre by ear. Before the flute leaves the factory, the timbre quality eval- uation of the flute depends on the manufacturers blowing test. This study is to implement an automatic flute timbre evaluation system. Through this evaluation system, the classification of timbre of flute can be recognized to achieve the goal of automatic measurement. 3. The automatic flute timbre quality evaluation system (ADTQE) An automatic system for evaluating the timbre quality of flute was designed, implemented and tested. Designing and implementing the system, the spectrum of musical instrument sound, discrete Fourier transform, cluster anal- ysis, and condition decision algorithm were included in this section. 3.1. Framework of ADTQE system The system, shown in Fig. 1, is composed of personal computer, microphone which was placed approximately 1 cm above the membrane, air pump, air valve, air jet throttle, open–close control valves of the flute finger hole and the fixed mount for the flute. The computer software packages include spectral analysis module, air flow pres- sure control module to control volume of blown air, open–close control valve module to change the testing tone, and microphone as a musical input device. The ADTQE system has been embedded with a user-friendly software tool using Borland C++ Builder (Deitel & Deitel, 2004) to develop a Windows-based interface system. 3.2. The spectrum of musical instrument sound The airflow vibration of a tube instrument is similar to the vibration of a harmonic, string instrument. As for flute, there are six finger holes on the bamboo. Pressing closed or releasing among these finger holes will change the wave length of the resonant to produce a different tone. Due to the limited number of finger holes on the flute and its per- forming limitation, it is necessary to widen the sound domain using the method of changing the blowing pressure. As for a tube musical instrument, the timbre in the high octave area is harder to control than the low octave area. From a physics perspective, because the tube of a flute in the high octave area is becoming shorter, the sound H.-J. Yang et al. / Expert Systems with Applications 32 (2007) 1004–1010 1007 vibration will be small. Additionally, after the tube of a flute becomes shorter, it is difficult to vibrate and the num- ber of harmonics obviously reduces. Compared with the vibration of the low octave on the same flute, the timbre of the high octave is tight and dark. In addition, owing to the membrane, the timbre of a flute becomes harmonic and sweet (McAulay & Quatiere, 1986; Serra, 1997). The signal of a time series could denote vs(t) the sum of different frequency signals of a sine wave through Fourier transformation. Fourier transformation can process a cyc- lic signal to produce the spectrum. Including the distrib- uted frequency named as x0 fundamental frequency and its harmonics. The critical part of actual signal spectrum is usually restricted within a small range of frequency x, which is very useful from the perspective of the signal process. The musical sound signal of Fig. 2 is a wave shape of music tone with fundamental frequency 293.7 Hz. It can be represented as the type of spectrum shown as Fig. 3. These two different kinds of representation form are called time-domain and frequency-domain. The frequency- domain of a signal, denoted as va(t), can be represented as va(x) after the Fourier transformation. Sound signal through the conversion of A/D (analog to digital circuit) to get the sampled signal becomes the dis- crete type. Subsequently, using the Discrete Fourier Trans- form transfers the time-domain into the frequency-domain. Discrete Fourier Transform is the application of the Fast Discrete Fourier Transform. It assumes that there are N sampling points Xn as the transform data to calculate using Eq. (1). Through Discrete Fourier Transform, the spec- trum of the sound signal or Periodogram was then acquired. That means the signal of the time-domain transfers into the frequency-domain with frequency components. Fig. 2. A wave shape of music tone with fundamental frequency 293.7 Hz. Fig. 3. A spectrum of music tone with fundamental frequency 293.7 Hz. X n ¼ 1 N XN�1 k¼0 xk e �2pjkn=N; n ¼ 0; 1; 2...N � 1 ð1Þ 3.3. Clustering analysis A Cluster Analysis with K-means algorithm (Kaufman & Rousseeuw, 1989) was applied to investigate the timbre distribution along the whole pitch of a flute. In the K-means algorithm, giving a set of N M-dimension points, denoted as Pn, was classified into K clusters. The centroid point of the cluster was represented by CK. The variance of each cluster can be defined as VARCðLkÞ¼ X i2Lk ðP i � lKÞ 2 ð2Þ where lK is the M-dimensional coordinates of cluster LK and Pi � lK is any distance metric set by the user. In order to reduce the variance of each cluster in Eq. (2), it is neces- sary to move any single point to any other cluster. The to- tal variance is defined by Eq. (3). TVARC ¼ XK i¼1 VARCðLKÞ ð3Þ To avoid the characteristics of converging to a local optimum in K-means calculation, selecting the initial cen- troid points must be done very cautiously. The Squared Euclidean Distance was used to calculate the variance of a cluster in this study. In our system, 75 iteration times were calculated as the best classification solution. 4. Experiment method and results 4.1. Materials and experiment operations Initially, paste the membrane on the flute and then affix the flute on the rack precisely. Adjust the jet throttle to blow in the hole while using a computer to control the air flow volume to the jet mouth. The blowing air intensity will affect the timbre of a flute (Sandell & Martens, 1995). Therefore, the blowing air into the flute must keep a steady speed in each pitch to assure the timbre analysis is under the same blowing pressure. In the example of D tone flute, the open–close control valves of the sound holes slowly goes from completely closed to open from bottom to up. The testing range of sound is from A4 (440 Hz) to B6 (1979 Hz) in D note flute. In order to simplify the represen- tation of the wave data, the intensity of harmonics which were less than 30 dB of the base format were discarded. Each sound within the whole testing domain is extracted by the analog to digital circuit. Then a spectral component is used to analyze the intensity of each harmonic. Fig. 4 shows the experiment operations of the automatic flute timbre quality evaluation system. It is composed of an air jet mechanism to inject the air into the testing flute while controlling the open or close function of the finger Timbre Characteristic: Bright, Thick, Sweet, Transparent Testing Dizi Timbre Judgement ClusteringSpectrum Analysis Finger hole open close controller Air jet mechanism Fig. 4. The experiment operations of the timber quality evaluation system. 1008 H.-J. Yang et al. / Expert Systems with Applications 32 (2007) 1004–1010 hole. A spectrum mechanism was used to extract the timbre features of the flute. A cluster analysis with K-means algo- rithm was used to classify the timbre feature distribution. At the last step, a decision making classifier was applied to judge the characteristics of the timbre features of the flute. These include sweet, bright, transparent, and thick. 4.2. Data acquisition Initially, in order to get the attributes of the timbre data- base, 300 good quality flutes were collected by three profes- sional musicians and then a serial number was put on each flute as training samples. Secondly, the automatic testing system gradually blows the musical tone and records the spectral data of each flute using K-means algorithm to clas- sify into four groups, shown in Fig. 5, where C1, C2, C3, and C4 individually represented the centroid of four groups. Finally, these flutes were arranged by their serial numbers and three professional musicians blew each type of flute and subjectively gave the suitable timbre classifica- tion, where C1 is bright, C2 is sweet, C3 is thick, and C4 is transparent. Timbre( testing flute feature)= Excellent, distance(c0)<0.4 and near(c0); Bright, distance(c1)<1.4 or ( distance>= 1.4 and near(c1)); Sweet, distance(c2)<1.4 or ( distance>= 1.4 and near(c2)); Thick, distance(c1)<1.6 or ( distance>= 1.6 and near(c3)); Transparent, distance(c1)<1.6 or ( distance>= 1.6 and near(c4)); Unqualify, distance(c0)>3; Fig. 5. The rules for defining the class of timbre. 4.3. Timbre range searching Past research shows that the decision condition algo- rithm is frequently used to match data corresponding to training data (Holland, 1992; Lee, 1996). As was pointed out in the last Section 4.2, four timbre classifications (C1–C4) were generated to specify each timbre quality respectively. C0(x*, y*), calculated by Eqs. (4) and (5), is the center of the 300 original samples. The timbre near C0 is considered to have ‘‘excellent’’ timbre quality. x� ¼ P all y P all xx � fðx; yÞP all y P all xfðx; yÞ ð4Þ y� ¼ P all y P all xy � fðx; yÞP all y P all xfðx; yÞ ð5Þ In order to define the radius of timbre distribution, C0, C1, C2, C3, and C4, the center of each timbre index is judged by three flute musicians who decide the radius of a cluster with the same timbre. The results found that the distance of C0 is 0.4, the coverage radius of C0; the dis- tance of C1 is 1.4, the coverage radius of C1; the distance of C2 is 1.4, the coverage radius of C2; the distance of C3 is 1.6, the coverage radius of C3; and the distance of C4 is 1.7, the coverage radius of C4. Furthermore, no tim- bre distribution was found when the distance of C0 is over 3. This indicated that if the distance from the testing sample to C0 is greater than 3, the timbre quality is classi- fied as ‘‘bad’’. The decision condition algorithm is used to classify the type of timbre of flute. The following set of rules defining the system is used to specify the class of tim- bre. If the distance from a testing sample to C0 is less than 0.4 then the timbre of flute is classified as ‘‘excellent’’. Indi- cating this kind of flute has the quality of sweet, bright, transparent, and thick simultaneously. If the distance from a testing sample to C1 is less than 1.4 then the timbre of Table 1 The results of timbre by ADTQE system and three musicians Timbre Classifier Excellent Bright Sweet Thick Transparent Bad ADTQE 78 459 228 315 294 126 Musician 1 81 454 230 309 292 134 Musician 2 77 460 225 311 295 132 Musician 3 82 453 223 315 298 129 Table 2 t-Test for timbre classification by ADTQE and three musicians Independent variables Mean SD t-Value ADTQE system 2.79 1.33 .125 Musicians 2.76 1.27 ***P < 0.001; **P < 0.01; *P < 0.05. H.-J. Yang et al. / Expert Systems with Applications 32 (2007) 1004–1010 1009 flute is classified as ‘‘bright’’. Indicating this kind of flute has the quality of bright. If the distance from a testing sam- ple to C2 is less than 1.4 then the timbre of flute is classified as ‘‘sweet’’. Implying this kind of flute has the quality of sweet. If the distance from a testing sample to C3 is less than 1.6 then the timbre of flute is classified as ‘‘thick’’. Meaning this kind of flute has the quality of thick. If the distance from a testing sample to C4 is less than 1.7 then the timbre of flute is classified as ‘‘transparent’’. Indicating this kind of flute has the quality of transparent. If the dis- tance from the testing sample to all four centers (C1, C2, C3, C4) is larger than 1.7 and the distance from the testing sample to center C0 is less than 3 then the timbre of flute is classified by the nearest distance cluster. If the distance from a testing sample to C0 is larger than 3 then the timbre of flute is classified as ‘‘bad’’. Indicating this kind of flute is unqualified. The pseudo-code of each rule in the decision condition algorithm is shown in Fig. 5. 4.4. Experiment results After the system has implemented, Min-Chu Musical Instrument Company and Chia-Yin Musical Instrument Company provided 1500 flutes as a testing sample to clas- sify its timbre quality. The experiment period is lasted for three months. Before beginning to test its timbre quality, each flute was assigned an identification number. After testing each one by one with the ADTQE system, the fre- quency of each index was counted and recorded. The test- ing results are displayed in Fig. 6, among which 78 are excellent; 459 are bright; 228 are sweet; 315 are thick; 294 are transparent and 126 are bad. Different tests have been performed with the indexes of sweet, transparent, thick, bright, excellent, and bad to clas- sify the timbre quality of flute with the ADTQE system and the technical support of three musicians. After the ADTQE system classification of 1500 flutes, three musicians individ- ually tested the 1500 flutes one by one to classify the type of timbre. The frequency of each index the musicians classi- fied was counted and recorded by identification number. Due to the large amount of data only summary results are presented in Table 1. As seen in Table 1, the discrep- ancy between the system and the musicians is minimal. In order to test the difference between ADTQE system and musicians, a t-test was applied and the results are presented 1500 flutes test result 0 100 200 300 400 500 ex ce lle nt br ig ht sw ee t th ic k tra ns pa re nt ba d timbre classification am ou nt o f ea ch t im br e ty pe timbre Fig. 6. The testing results of 1500 flutes for ADTQE system. in Table 2. As expected, the comparison between the sys- tem and musicians leads to less significant differences. In other words, there are no significant differences between the system classification and musicians’ classification. This result indicated that the classification of timbre quality in the automatic system is valid and accurate. Additionally, an ANOVA test was used to test the classification differ- ences among musicians. The results displayed that there are no significant classification differences among these three musicians (F = 0.004, P = .996 > .000). This result indicated that the evaluation among three musicians is almost identical. 5. Discussion and conclusions The function of the ADTQE system in this study was to provide timbre classification of flute that is accurate, pre- cise, and time saving. Several conclusions can be drawn from the obtained results. The ADTQE system has shown to be a useful and attractive tool to classify the timbre qual- ity of flute. However, there are only 1500 flute testing sam- ples, which is not enough to generalize about a large pool of flute. A further study should be pursued using samples of a larger number of varieties coming from different flute factories to increase the usability and validity of ADTQE system. Furthermore, in order to get a more precise classi- fication on the timbre quality of flute, more sample sizes need to be used in the training stage to further improve the accuracy of these experiment results for advanced study. Additionally, the measurement of indexes on sweet, transparent, bright, and thick is very subjective. It is hard to use a scale or instrument to measure these indexes. This study is only in the preliminary stage. Further studies are needed to clearly define the concepts of these indexes. Due to the characteristics and features of bamboo, it is very difficult to objectively evaluate the timbre quality of flute. Problems are encountered not only within the mate- rial of natural-born bamboo, but also in the implementing process, where technicians possibly resist using an auto- matic evaluation system to judge the timbre quality of flute. 1010 H.-J. Yang et al. / Expert Systems with Applications 32 (2007) 1004–1010 Two possible arguments can depict this. One possible argument is that technicians may always trust that their judgment is better than the system. They may think experi- ence is a key factor in obtaining adequate classification results. The second argument is that technicians would feel they may lose their jobs. Their jobs will be replaced by a computer system. So management should consider how to limit the resistance of technicians to encourage them having intention to use the system as a supportive tool to make their works more efficient and reduce their workload. Moreover, classifying the timbre quality of flute using traditional human-methods is insufficient to classify the timbre quality that customers want or look for in flute products. Customers differ dramatically in their timbre preferences for flute. Thus, the proposed automatic timbre quality evaluation system was developed to enable the pre- diction of total timbre quality of flute products. In short, the ADTQE system can be used as a quality decision tool to classify the timbre quality in Chinese flute industry. At the classifying levels, the system can be used by different technicians who don’t need to possess the tim- bre skill or experience of flute to test the timbre for decid- ing the timbre quality of flute. This automatic system can improve the speed of the process in the timbre quality test- ing task to increase productivities and economic benefits for Chinese flute industry. In addition, the automatic sys- tem has a uniform standard to classify the timbre quality for a long period of time even 24 hours a day, not like tech- nicians who is easily influenced by the environment, mood, psychological factors, limited energy or other factors dur- ing their classification process. Acknowledgements The experiments were under the support of Pin-Chung Lin musician in Chia-Yin Musical Instrument Company and Fa-Chang Lin and Shang-Hwa Su musicians in Min- Chu Musical Instrument Company. The author would like to thank and acknowledge these people for their great important support and help. The author would also like to thank that this work was supported by the National Science Council of Taiwan, ROC under Grant No. NSC-94-2416-H-167-007 and NSC-94-2614-E-167-001. References Abbott, J. A. (1999). Quality measurement of fruits and vegetables. Postharvest Biology and Technology, 15, 207–225. Carvalho, P., Araujo, H., & Dourado, A. (1999). An automatic optical sensor for vessels and fibers quality inspection in pulp production. Computers & Industrial Engineering, 37, 355–358. Chen, C.-S. (2002). The summary of timbre of dizi. Shanghai Art Graduate School. Available from http://suona.com/instr.htm.03-17. Deitel, H. M., & Deitel, P. J. (2004). C How to program (4th ed.). Pearson Education, Inc., Pearson Prentice Hall. Di-Yin Net (2005). How to choose a good quality of dizi. <http:// www.diin.cn/index.asp> Accessed 25.06.05. Epps, J., Smith, J., & Wolfe, J. (1997). A novel instrument to measure acoustic resonances of the vocal tract during speech. Measurement Science and Technology, 8, 1112–1121. Holland, J. M. (1992). Adaptation in natural and artificial systems, 1975 (2nd ed.). University of Michigan Press, MIT Press. Houtsma, A. (1997). Pitch and timbre: definition, meaning and use. Journal of New Music Research, 26, 104–115. Ilmoniemi, M., Valimaki, V., & Huotilainen, M. (2004). Subjective evaluation of musical instrument timbre modifications. Joint Baltic- Nordic Acoustic Meeting, 8–10 June, Mariehamn, Aland, pp. 5–11. Kaufman, L., & Rousseeuw, P. J. (1989). Finding groups in data: An introduction to cluster analysis. New York: John Wiley. Kostek, B. (1999). Soft computing in acoustics, applications of neural networks, fuzzy logic and rough sets to musical acoustics, studies in fuzziness and soft computing. Heidelberg, New York: Physical Verlag. Kuo, H. C., & Wu, L. J. (2002). An image tracking system for welded steams using fuzzy logic. Journal or Materials Processing Technology, 120, 169–185. Lee, H. M. (1996). Group decision making using fuzzy set theory for evaluating the rate of aggregative risk in software development. Fuzzy Sets and Systems, 80(3), 261–271. McAulay, R. J., & Quatiere, T. F. (1986). Speech analysis/synthesis based on a sinusoidal representation. IEEE Transaction on Acoustics, Speech and Signal Processing, 34, 744–754. Neufeldt, V. E. (1988). Webster’s New World Dictionary (3rd college ed.). New York: Simon and Schuster. Sandell, G. J., & Martens, W. (1995). Perceptual evaluation of principal- component-based synthesis of musical timbres. Journal of The Audio Engineering Society, 1013–1028. Serra, X. (1997). Musical sound modeling with sinusoids plus noise. In A. Piccialli, C. Roads, & S. T. Pope (Eds.), Musical signal processing. Swets Zeitlinger Publisher. Swanepoel, K. T. (2004). Decision support system: Real-time control of manufacturing processes. Journal of Manufacturing Technology Man- agement, 15(1), 68–75. Van Harten, W. H., Casparie, T. F., & Fisscher, O. A. M. (2000). Methodological considerations on the assessment of the imple- mentation of quality management systems. Health Policy, 54, 187–200. Verdu Jover, A. J., Llorens Montes, F. J., & Fuentes, M. del M. F. (2004). Measuring perceptions of quality in food products: The case of red wine. Food Quality and Preference, 15, 453–467. Wolfe, J., & Smith, J. (2001). Embouchures and end effects in air-jet instruments. In International congress on acoustics, Rome, Session 8. Wolfe, J., Smith, J., Tann, J., & Fletcher, N. H. (2001). Acoustic impedance of classical and modern flutes. Journal of Sound and Vibration, 243, 127–144. http://suona.com/instr.htm.03-17 http://Http://www.diin.cn/index.asp http://Http://www.diin.cn/index.asp Automatic timbre quality evaluation in Chinese traditional flute industry Introduction Background and characteristics of flute The automatic flute timbre quality evaluation system (ADTQE) Framework of ADTQE system The spectrum of musical instrument sound Clustering analysis Experiment method and results Materials and experiment operations Data acquisition Timbre range searching Experiment results Discussion and conclusions Acknowledgements References 
work_sdrrc56bk5bkfior2rbas4dyty	----	The TOPOSCOUT expert system for stroke localization. 1195 The TOPOSCOUT Expert System for Stroke Localization Klaus Spitzer, PhD, MD, Andreas Thie, MD, Louis R. Caplan, MD, and Klaus Kunze, MD Clinically, strokes are localized by the findings on neurologic examination, TOPOSCOUT is an expert system designed to diagnose the anatomic location and the corresponding vascular territory of strokes based on the clinical signs and symptoms. The inference engine of TOPOSCOUT uses a backtracking algorithm and a rule-based data base that includes associations of neurologic signs with vascular and anatomic areas, TOPOSCOUT is capable of detecting typical stroke patterns, for example, "top-of-the-basilar" or Wallenberg's syndromes. The accuracy of TOPOSCOUT'S diagnoses has been tested for conformity with the final diagnoses of 129 patients in the Hamburg Stroke Data Bank, and a high level of agreement was found for hemispheric lesions. The program runs on microcomputers with MS-DOS and is intended as a practical aid for physicians not fully familiar with topologic stroke diagnosis and as an interactive teaching device. (Stroke 1989;20:1195-1201) The topologic localization of stroke is basedon detailed knowledge of the associationsof clinical signs and symptoms with lesions of particular neuroanatomic structures. Frequently, these associations can be formulated as a rule: If there is hemilateral impairment of sensation and no motor weakness or cognitive defects, then assume a lesion in the contralateral thalamus. The data are subjected to multiple pattern-matching rules to arrive at the probability of the final localization. Rule- based expert systems have been used to analyze gait abnormalities, drug interactions, headache and facial pain, and other medical problems. 1~5 We present TOPOSCOUT, a prototype microcomputer- based expert system for stroke localization, TOPO- SCOUT may be used independently or as a supple- ment to MICROSTROKE, which is an expert system for the diagnosis of stroke type.6 Materials and Methods TOPOSCOUT'S knowledge data base is arranged entirely in the form of rules, presently 171. All data are derived from anatomic charts and textbooks7-15 From the Neurologische Universitatsklinik Hamburg- Eppendorf, Hamburg, Federal Republic of Germany (K.S., A.T., K.K.) and the Department of Neurology, Tufts-New England Medical Center, Boston, Massachusetts (L.R.C.). Presented in part at the 40th Annual Meeting of the American Academy of Neurology, Cincinnati, Ohio, April 1988. Address for correspondence: Dr. Spitzer, Neurologische Uni- versitatsklinik Hamburg-Eppendorf, MartinistraBe 52, D-2000 Hamburg, Federal Republic of Germany. Received June 21, 1988; accepted March 9, 1989. and from 63 clinical publications. Rules are classi- fied into six groups: 1. Signs/symptoms rules (n=42; an example is Rule S37 in Figure 1). TOPOSCOUT is able to find a syndrome by logical deduction from clinical signs provided by the physician-user. Thus, formulation of more complicated diagnostic rules using complex terms substituted for signs is feasible. Questions to the physician-user are avoided if their answers can be logically deduced. 2. Hemisphere/brainstem rules (n=30; an exam- ple is Rule H29 in Figure 1). To estimate the odds for a hemispheric vs. a brainstem lesion, TOPOSCOUT maintains one account for each location. A priori odds were taken from stroke textbooks. 13~1S Depend- ing on the signs in a given patient, TOPOSCOUT multiplies each account by a factor listed in associ- ated rules. Relevance is an additional multiplication factor that takes into account the relative weight of a clinical item in the diagnostic decision. 3. Right/left rules («=21; an example is Rule R14 in Figure 1). These rules comprise logical deduc- tions for detection of the side of a stroke. Right- and left-sided strokes can be detected in the same patient; thus, discovery of bilateral hemispheric or brainstem lesions is possible. 4. Vascular pattern rules (n=26; an example is Rule V08 in Figure 1). Vascular patterns are coded into logical rules. As for the laterality of a stroke, conclusions from several rules can be considered at the same time, corresponding to the involvement of more than one vascular territory. At present, TOPO- SCOUT'S knowledge data base includes only patterns D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1196 Stroke Vol 20, No 9, September 1989 RULE S3 7: IF AID AID AID THEI thsre is weakness in light arm there is no weakness in left arm there is weakness in right lag there is no weakness in left lag there is right heaiparesis; RULE H29 IF thara is dysarthria OB there is impairment of lower cranial narvas TKEI relevance = 2, multiply hemispheric account with 0.3, multiply brainstea account with 2.0, RULE R14: IF the patiant omits tha laft side of a drawn figure AID ha displays abnormal angles and proportions in a drawn figure THEI assuae lesion on tha right (hemispheric) sida; RULE V08: IF thare are no reliable signs of a lesion in tha territory of tha middle cerabral artary AID thare is dyslaxia AID thara is no dysgraphia THEI assuae a lesion in the territory of the posterior cerebral artary; RULE A20 IF thara is aydriasis on tha left AID there is paresis of laft aye auscles AID thare is ptosis on tha laft AID thara is right facial weakness AID thara is right heaiparesis THEI assume WEBER's syndrome on the l e f t , RULE HO9 IF tha c h o l a s t a r o l l e v e l i s a l e v a t a d THEI r e l e v a n c e = 1, m u l t i p l y HCA account w i th 1 . 3 ; m u l t i p l y ICA account w i th S 3 , FIGURE 1. Excerpt from TOPOSCOUT'S knowledge data base. for the main stems of the hemispheric cerebral arteries, the anterior cerebral artery (ACA), the middle cerebral artery (MCA), and the posterior cerebral artery (PCA). Although in 90% of patients the PCA arises from the basilar artery,13 the PCA is analyzed with the ACA and MCA as a cortical supply vessel. So far, TOPOSCOUT cannot distinguish segments and branches of these main arteries. The knowledge data base includes rules that take into account the age-dependent frequency distribution of vascular territories involved in stroke samples.16 5. Anatomic pattern rules («=28; an example is Rule A20 in Figure 1). These rules comprise pat- terns of neurologic signs associated with particular anatomic areas. For example, TOPOSCOUT can detect thalamic stroke and specific brainstem lesions such as Wallenberg's syndrome. We have included rules that provide a tolerance in diagnosis in case a single sign usually associated with a particular syndrome is missing, but the presence of other signs yields a high probability for this syndrome. 6. MCA/ICA rules (n=24; an example is Rule M09 in Figure 1). If stroke in the territory of the MCA is assumed and if there are no signs of an embolic source in the heart or proximal vascular system, a particular set of rules is activated, TOPO- SCOUT tries to assign the vascular lesion to either the internal carotid artery (ICA) or the MCA; data were derived from an investigation of occlusive disease of the MCA by Caplan et al.17 Similar to the estimation of hemispheric vs. brainstem lesion, accounts for both vessels are maintained, resulting in final odds for an ICA vs. an MCA lesion. TOPOSCOUT'S inference engine is based on a back- tracking algorithm and is the section of the com- puter program that includes the specific problem- solving capabilities of the expert system. A backtracking mechanism18-20 is a specific method of problem solving in which the expert system starts with the final goal (G). For TOPOSCOUT, G is to detect stroke laterality, hemispheric vs. brainstem location, and vascular and anatomic patterns. The inference engine searches the data base for rules for which the conclusion is R1-»G. To establish these rules, the inference engine looks for other rules that yield the premises of the first ones (R2->R1). In this D ow nloaded from http://ahajournals.org by on A pril 5, 2021 R/L deteotlon ACA assumed Spitzer et al TOPOSCOUT Expert System 1197 basic I examination! hsm/brtt dataotlon ham lesion assumed brtt latlon aaaumad vase tarr detaotlon oattarn detaotlon R/L detaotlon I pattarn Ideteotlon I MCA/ICA assumed PCA assumed MCA/ICA dateotlon FIGURE 2. Flowchart of TOPOSCOUT'sprogram architecture, hem, hemispheric; brst, brainstem; R/L, right vs. left; vase terr, vascular territory; ACA, anterior cerebral artery; MCA, middle cerebral artery; ICA, internal carotid artery; PCA, posterior cerebral artery; thromb, thrombosis. way, TOPOSCOUT backtracks to rules the premises of which are facts (F) about the current patient obtained by the physician-user. G is then concluded by chaining the conclusions F-». . . R2-»R1-»G. The inference engine includes 52 rules strictly separated from the knowledge data base. As described above, the heuristic character of TOPOSCOUT allows none, one, or more than one solution for the problems right/left, vascular pattern, and anatomic pattern. Mathematical calculations are restricted to the solu- tion of the problems hemispheric vs. brainstem and MCA vs. ICA lesion. The inference engine has access to an ASCII-text file that contains informa- tion used for explanatory and educational purposes. TOPOSCOUT asks the physician-user interactively for details of the neurologic examination, usually in a multiple-choice format. Some questions are omit- ted if the answer can be deduced logically. We included clinical items from the questionnaire of Ropper et al21 for computer-assisted data acquisi- tion in a neurologic intensive care unit, TOPOSCOUT does not use laboratory findings. TOPOSCOUT is linked to a data base in which all personal cases, including the physician-user's final diagnoses confirmed by laboratory studies, are stored. To assess the validity of TOPOSCOUT'S diag- noses, control programs calculating success rates have been implemented. In addition, special sub- routines are capable of retrieving data on authentic cases from large stroke registries, passing them to TOPOSCOUT, and driving the expert system in an automatic fashion. TOPOSCOUT'S programs are written with the VP-Expert22 expert system development shell. An IBM-compatible microcomputer with a floppy disk drive is required. Results Operating TOPOSCOUT is easy; no knowledge of computers is required. Figure 2 shows a flowchart of the program architecture, TOPOSCOUT starts with acquisition of data about a patient's neurologic signs (Figure 3). The physician-user selects one of the options presented in a multiple-choice format. Numerical input is required for some questions such as age. A summary of the patient's findings is displayed and simultaneously written to a disk file for later review. Initially, the system estimates odds for stroke localization in the cerebral hemispheres vs. in the brainstem. In both instances, TOPOSCOUT tries to determine the side of the stroke. For hemispheric lesions, TOPOSCOUT looks for patterns of signs asso- ciated with territories of the main cerebral arteries. If involvement of the MCA is assumed, TOPOSCOUT attempts to distinguish between a vascular lesion in the ICA or the MCA. For this purpose, historic and clinical data about the patient is requested. In addition, the patient's data profile is reviewed for patterns of signs associated with specific anatomic areas. If the presence of a particular syndrome is suspected and TOPOSCOUT needs more information to prove or refute it, the physician-user is consulted again. It takes approximately 3-4 minutes to enter data for a patient. At the end of the session, TOPOSCOUT has summed up all entered information to present its final diag- nosis (Figure 4), which comprises odds for hemi- spheric vs. brainstem lesions and TOPOSCOUT'S pre- diction including the supposed side of the stroke, an anatomic pattern if present, and TOPOSCOUT'S esti- mation of the involvement of large arteries, com- bined with their age-matched odds. D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1198 Stroke Vol 20, No 9, September 1989 How about f a c i a l weakness 7 no weakness L weakness < b i l a t e r a l weakness R weakness How about f a c i a l sensation ? normal < L impairment b i l a t e r a l impairment How about the lower cranial nerves ? no d e f i c i t < L impairment b i l a t e r a l impairment Do nystagmus or vertigo e x i s t ? no < yes Is there dysarthria ? no < yes E impairment R impairment FIGURE 3. Excerpt from sample session: TOPOSCOUT'S data acquisition. L, left; R, right. Enter to select EID to conplete /Q to Quit ' for Unknown Before leaving TOPOSCOUT, the physician-user is prompted to enter the clinical diagnosis based on results of laboratory studies, if available. This authen- tic information is also stored in the disk file. TOPOSCOUT is able to explain its conclusions. Comments about particular keywords include infor- mation about the pertinent literature taken into account at different phases of diagnosis. To test the quality of TOPOSCOUT'S diagnoses, we have transferred data on 129 patients from the Hamburg Stroke Data Bank23 to TOPOSCOUT'S con- trol unit, TOPOSCOUT'S diagnoses were compared with the final diagnoses in the stroke registry based on the results of laboratory examinations, TOPO- SCOUT recognized 87% of all hemispheric strokes regardless of side (76% of right hemispheric strokes and 74% of left hemispheric strokes) and 56% of all brainstem strokes regardless of side (56% of right brainstem strokes and 50% of left brainstem strokes). TOPOSCOUT identified 86% of all MCA-territory strokes regardless of side (76% of right MCA strokes and 73% of left MCA strokes) and 41% of all PCA-territory strokes regardless of side (50% of right PCA strokes and 36% of left PCA strokes). There were no ACA-territory strokes. Discussion TOPOSCOUT is a microcomputer-based expert sys- tem designed to assist physicians in diagnosis of the topology of stroke using information entirely avail- able at the bedside. Simulating the human diagnos- tic approach, TOPOSCOUT'S pattern-matching algo- rithms can direct the physician-user's attention to syndromes that may occasionally be overlooked. TOPOSCOUT'S topologic knowledge data base repre- sents a fast and easily accessible reference for physicians involved in the management of stroke patients. Like many other expert systems in medicine,24 TOPOSCOUT serves as a tool for clinical teaching and may be useful for planning further diagnostic procedures. In addition, since all data of personal cases, including the final diagnoses con- firmed by laboratory studies, are stored, TOPOSCOUT provides a stroke registry. Since the computer file encompasses the com- plete neurologic status of a patient, the physician- user wastes less time with handwritten charts. Tak- ing into account the fast retrieval of the personal stroke registry, we believe that the few minutes spent for data entry are well invested and do not slow down the physician-user in the setting of a busy practice. TOPOSCOUT'S inference engine has been devel- oped applying artificial intelligence techniques in a medical diagnostic process.25-26 The inference engine comprises goal-directed programming, pattern matching abilities, and explanatory capabilities. These methods must be distinguished from decision-support programs based on statistical pro- cedures.27-29 We consider statistically based meth- Odds for hemispheric lesion: 88.3 Odds for brainstem lesion 11 7 Predicting lesion in right hemisphere. Pattern: Pure motor stroke (basis pontis/internal capsule). Predicting lesion in AC* t e r r i t o r y . Predicting lesion in HCA territory: yes Predicting lesion in PCA t e r r i t o r y (age-matched odds: 6X) (age-matched odds: 8lX> (age-matched odds 13X) FIGURE 4. Excerpt from sample session: TOPOSCOUT'S diagnosis. ACA, anterior cere- bral artery; MCA, middle cerebral artery; PCA, posterior cerebral artery. D ow nloaded from http://ahajournals.org by on A pril 5, 2021 Spitzer et al TOPOSCOUT Expert System 1199 ods, especially Bayesian techniques,30-31 to be inad- equate tools for a topologic diagnosis-support system; the set of topologic hypotheses is generally not exhaustive and mutually exclusive as assumed by Bayesian statistics.29-32 Thus, the frequent involvement of two or more cerebral territories by a stroke contradicts the assumption of only one cor- rect diagnosis in each case as required by most statistical procedures. In TOPOSCOUT, mathematical calculation is limited to a few minor problems in which general objections to statistical procedures do not have to be considered. Adversely, the MICRO- STROKE expert system for the diagnosis of stroke type uses modified Bayesian inference techniques.6 The backtracking algorithm was selected to restrict data acquisition to those questions that are needed to pursue relevant goals, TOPOSCOUT skips queries if pertinent information is supplied by other answers or if questions seem to be of limited value in the current diagnostic process; for example, not all neuropsychological phenomena33 are requested for each patient. However, the inference engine simu- lates forward-chaining features by requiring the physician-user to enter complete results of a neuro- logic examination, providing the stroke registry capability. In this way, later reassessment of TOPO- SCOUT'S decision-making process is feasible. Published programs for computer-assisted topo- logic diagnosis in neurology can be classified into two groups based on their neuroanatomic expertise.34 In one group, knowledge data bases interpret the spatial representation of neuroana- tomic items and their relations.34-39 In the second group, knowledge data bases are constructed by associations of neuroanatomic items with clinical signs without considering the structural or func- tional relations of anatomy. Localization is pre- dicted by recognizing patterns of clinical findings.40-43 TOPOSCOUT belongs to this second group. The lack of total knowledge of neural con- nections in the central nervous system precludes storage of all structural and functional relations in a data base. Pattern recognition is an appropriate approach for this representation of rudimentary knowledge, and pattern matching provides the capa- bility to detect lesions at multiple locations. How- ever, TOPOSCOUT shares the disadvantage of all expert systems based on associations of neuroana- tomic items with clinical signs: it is limited to recognizing common clinical patterns and is unable to diagnose patients whose syndromes have not been previously described or have not been included in the rule-based system. At present, there are still other drawbacks to our expert system. One serious handicap is that validity control has shown poor detection of brainstem lesions and involvement of the PCA. Too few rules (31) have been implemented concerning topologic diagnosis of brainstem lesions (compared with 72 rules for hemispheric lesions). If a symptom could be caused by a hemispheric or a brainstem lesion, the former location is often falsely favored by TOPOSCOUT because of probability assignment and disregard of pertinent cosymptoms. Infra ten torial topologic diagnosis forms an obstacle not only for TOPOSCOUT but for other systems as well. For exam- ple, PAL was correct in detecting posterior fossa strokes in only 53% of cases.42-43 Posterior circula- tion stroke is more heterogeneous than anterior circulation stroke. A larger registry of patients with vertebrobasilar stroke will be needed to amplify TOPOSCOUT'S rule system. Most diagnostic failures were caused by TOPO- SCOUT'S insufficient tolerance. Stroke location was often not correctly detected because an appropriate minor symptom was missing, prompting the definite rejection of this location despite other evidence. Some rules adhere too closely to logical links of clinical and anatomic items, a fact that uncovers a general disadvantage of rule-based systems, TOPO- SCOUT was helpless in identifying lesions associated with a broad spectrum of clinical symptoms, such as caudate infarcts.44 We must improve our knowl- edge data base to include more rules that take into account variability of clinical syndromes. Furthermore, the knowledge data base should be expanded to include clinical patterns of stroke in cerebral arteries supplying brainstem structures and patterns associated with stroke in terminal parts of hemispheric arteries. Lacunar infarcts should be identified by TOPOSCOUT. Additional improvement will be achieved by using TOPOSCOUT'S self-learning characteristics when data from past consultations are included in current diagnostic decisions. The rules implemented in this prototype expert system should be regarded as preliminary. They are still inadequate concerning consistency, complete- ness, and independence. We will attempt to over- come these shortcomings by applying knowledge acquisition and verification tools.3'45-46 Since 1979, numerous expert systems have been successfully developed with expert system shells (software tools for developing expert systems).47-49 In contrast, such a shell seems to be too inflexible for this specific stroke localization problem. In particular, inclusion of tolerance in the mechanism of pattern recognition seems to pose severe prob- lems for a shell. To provide TOPOSCOUT with full deductive reasoning capabilities, the expert system is translated into PROLOG.50 In the future, a prospec- tive study will be carried out to verify the validity of the program. This prototype expert system suggests that computer-assisted stroke localization is feasible at the bedside and that, with further improvement, TOPOSCOUT holds promise as a tool in the clinical routine of stroke diagnosis. Finally, we clearly do not deny that skillful and experienced neurologists, with their present 80-86% accuracy rate,51 have legitimate reasons to criticize the competence of the newborn TOPOSCOUT. But we also want to empha- size that these authorities decided for good reasons D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1200 Stroke Vol 20, No 9, September 1989 to spend a considerable number of years as medical students, residents and, possibly, stroke fellows. Therefore, we believe that it is tolerable to accept a training period of several years for the development of an expert system's software and expansion of its knowledge data base before it is compared with the most competent in the field. References 1. Davis R, Buchanan B, Shortliffe E: Production rules as a representation for a knowledge based consultation program. Artif Intelligence 1977;7:15-45 2. Dzierzanowski JM, Bourne JR, Shiavi R, Sandell HS, Guy D: GAITSPERT: An expert system for the evaluation of abnor- mal human locomotion arising from stroke. IEEE Trans Biomed Eng 1985;32:935-942 3. Hudson DL, Estrin T: Derivation of rule-based knowledge from established medical outlines. Comput Biol Med 1984; 14:3-13 4. Matsumura Y: RHINOS: A consultation system for diagnosis of headache and facial pain. Comput Methods Programs Biomed 1986;23:65-71 5. Roach J, Lee S, Wilcke J, Ehrich M: An expert system for information on pharmacology and drug interactions. Comput Biol Med 1985;15:ll-23 6. Spitzer K, Thie A, Caplan LR, Kunze K: MICROSTROKE—A microcomputer-based expert system for stroke type diagno- sis. Stroke (in press) 7. Brazis PW, Masdeu JC, Biller J: Localization in Clinical Neurology. Boston, Little, Brown & Co, 1985 8. Brodal A: Neurological Anatomy. New York, Oxford Uni- versity Press, 1981 9. Chusid J: Correlative Neuroanatomy, Functional Neurol- ogy. Los Altos, Calif, Lange, 1985 10. Duus P: Topical Diagnosis in Neurology. Stuttgart, Thieme, 1983 11. Niuewenhuys R, Voogd J, v Huijzen C: The Human Central Nervous System. A Synopsis and Atlas. Berlin, Springer, 1981 12. Poritsky R: Neuroanatomical Pathways. Philadelphia, WB Saunders Co, 1984 13. Barnett HJM, Mohr JP, Stein BM, Yatsu FM (eds): Stroke. II1I. New York, Churchill Livingstone Inc, 1986 14. Caplan LR, Stein RW: Stroke. Boston, Butterworths, 1986 15. Millikan CH, McDowell FM, Easton JD: Stroke. Philadel- phia, Lea & Febiger, 1987 16. Spitzer K, Thie A, Fuchs B, Kunze K: Ischemic strokes in young adults (abstract). J Neurol 1988;235(suppl 1):5 17. Caplan LR, Babikian V, Helgason C, Hier DB, DeWitt D, Patel D, Stein R: Occlusive disease of the middle cerebral artery. Neurology 1985;35:975-982 18. Hayes-Roth F, Waterman DA, Lenat DB (eds): Building Expert Systems. Reading, Mass, Addison-Wesley, 1983 19. Waterman DA: ,4 Guide to Expert Systems. Reading, Mass, Addison-Wesley, 1986 20. Nilsson NJ: Principles of Artificial Intelligence. Berlin, Springer, 1982 21. Ropper AH, Griswold K, McKenna D, Souder D: Computer- guided neurologic assessment in the neurologic intensive care unit. Heart Lung 1981;10:54-60 22. VP-Expert. Ruled-Based Expert System Development Tool. Berkeley, Calif, Paperback Software International, 1987 23. Spitzer K, Becker V, Thie A, Kunze K: The Hamburg Stroke Data Bank: Goals, design and preliminary results. J Neurol 1989;236:139-144 24. Szolovits P: Using artificial intelligence models of medical decision-making in medical education, in Pages JC, Levy AH, Gremy F, Anderson J (eds): Meeting the Challenge: Informatics and Medical Education. New York, Elsevier, 1983, pp 271-281 25. Szolovits P: Artificial Intelligence in Medicine. Boulder, Colo, Westview Press, 1982 26. Torasso P: Knowledge based expert systems for medical diagnosis. Stat Med 1985;4:317-325 27. Fox J, Barber D, Bardhan KD: Alternatives to Bayes?—A quantitative comparison with rule-based diagnostic infer- ence. Methods Inf Med 1980;19:210-215 28. Schwartz B, Patil RS, Szolovits P: Artificial intelligence in medicine. Where do we stand? N Engl J Med 1987; 316:685-688 29. Bouckaert A: Medical diagnosis: Are expert systems needed? IntJ Biomed Comput 1987;20:123-133 30. Salamon R, Bernadet M, Samson M, Derouesne C, Gremy F: Bayesian method applied to decision making in neurology— Methodological considerations. Methods Inf Med 1976; 15:174-179 31. Duda RO, Hart PE, Nilsson NJ: Subjective Bayesian Meth- ods for Rule-Based Inference Systems. Stanford, Calif, National Computer Conference, 1976, pp 1075-1082 32. Szolovits P, Pauker SG: Categorical and probabilistic rea- soning in medical diagnosis. Artif Intell 1978;11:115-144 33. Pinskaya LB: Neuropsychological phenomena of secondary stem dysfunction in patients with strokes of hemispheric localization. Neurosci Behav Physiol 1986;16:101-104 34. First MB, Weimer BJ, McLinden S, Miller RA: LOCALIZE: Computer-assisted localization of peripheral nervous system lesions. Comput Biomed Res 1982;15:525-543 35. Meyer AU, Weissman WK: Computer analysis of the clini- cal neurological examination. Comput Biol Med 1973; 3:111-117 36. Meyer AU, Weissman WK: The localization of human nervous system lesions by computer analysis. Trans Am Neurol Assoc 1973;98:327 37. Catanzarite VA, Greenburg AG, Bremermann HJ: Com- puter consultation in neurology: Subjective and objective evaluations of the "NEUROLOGIST" system. Comput Biol Med 1982;12:343-355 38. Catanzarite VA, Greenburg AG, Bremermann HJ: Com- puter assisted diagnosis and computer consultation in neu- rology: Preliminary testing of diagnostic accuracy for the NEUROLOGIST system. Int J Neurosci 1981;13:43-54 39. Banks G, Weimer B: Symbolic coordinate anatomy for neurology (SCAN). J Med Syst 1984;8:157-162 40. Reggia JA: A production rule system for neurological local- ization, in Reggia JA, Tuhrim S (eds): Computer-Assisted Medical Decision Making. II. New York, Springer, 1985, pp 49-62 41. Stewart A, Cala L: Mathematical method to utilize a com- puter for diagnosis of site and type of intracerebral mass lesions. BrJ Radiol 1975;48:97-100 42. Hier DB, Caplan LR, Hill H, Evans M, Sinha A: A microcomputer-based expert system to assist in localization of anatomic damage after stroke (abstract). Neurology 1984; 34(suppl 1):83 43. Hier DB, Atkinson GD, Perline R, Hill H, Evans M, Desai B, McCormick WC, Caplan LR: Can a patient data base help build a stroke diagnostic expert system? Med Inf 1986; 11:75-81 44. Caplan LR, Schahmann JD, Baquis GD, Kase CS, Feld- mann E, Greenberg JP, Gorelick PB, Helgason C, Hier DB: Caudate infarcts (abstract). Neurology 1988;38(suppl 1):262 45. Suwa M, Scott AC, Shortliffe EH: An approach to verifying completeness and consistency in a rule-based expert system. Artif Intell Mag 1982;2:16-21 46. Mars NJ, Miller PL: Knowledge acquisition and verification tools for medical expert systems. Med Decis Making 1987; 7:6-11 47. van Melle W: A domain-independent production rule system for consultation programs, in Proceedings of the Sixth International Joint Conference on Artificial Intelligence. Tokyo, 1979. Los Altos, Calif, Morgan Kaufmann, 1979, pp 923-925 48. Weiss S, Kulikowski C: EXPERT: A system for developing consultation models, in Proceedings of the Sixth Interna- tional Joint Conference on Artificial Intelligence. Tokyo, 1979. Los Altos, Calif, Morgan Kaufmann, 1979, pp 942-960 D ow nloaded from http://ahajournals.org by on A pril 5, 2021 Spitzer et al TOPOSCOUT Expert System 1201 49. Blomberg DJ, Guth JL, Fattu JM, Patrick EA: Evaluation of a new classification system for anemias using Consult Learn- ing System. Comput Methods Programs Biomed 1986; 22:119-125 50. Townsend C: Mastering Expert Systems With Turbo Prolog. Indianapolis, Sams, 1986 51. Nadeau SE, Jordan JE, Mishra SK: A prospective evalua- tion of the localizing value of clinical data in acute stroke (abstract). Neurology 1988;38(suppl 1):344 KEY WORDS • cerebrovascular disorders • expert systems D ow nloaded from http://ahajournals.org by on A pril 5, 2021 
work_se473dceh5atrhc37ydulr7ns4	----	Increasing sample efficiency in deep reinforcement learning using generative environment modelling S P E C I A L I S S U E P A P E R Increasing sample efficiency in deep reinforcement learning using generative environment modelling Per-Arne Andersen | Morten Goodwin | Ole-Christoffer Granmo Department of ICT, University of Agder, Grimstad, Norway Correspondence Per-Arne Andersen, Jon Lilletunsvei 9, 4879 Grimstad, Norway. Email: per.andersen@uia.no Abstract Reinforcement learning is a broad scheme of learning algorithms that, in recent times, has shown astonishing performance in controlling agents in environments pres- ented as Markov decision processes. There are several unsolved problems in current state-of-the-art that causes algorithms to learn suboptimal policies, or even diverge and collapse completely. Parts of the solution to address these issues may be related to short- and long-term planning, memory management and exploration for reinforcement learning algorithms. Games are frequently used to benchmark reinforcement learning algorithms as they provide a flexible, reproducible and easy to control environments. Regardless, few games feature the ability to perceive how the algorithm performs exploration, memorization and planning. This article presents The Dreaming Variational Autoencoder with Stochastic Weight Averaging and Generative Adversarial Networks (DVAE- SWAGAN), a neural network-based generative modelling architecture for explo- ration in environments with sparse feedback. We present deep maze, a novel and flexible maze game-engine that challenges DVAE-SWAGAN in partial and fully observable state-spaces, long-horizon tasks and deterministic and stochas- tic problems. We show results between different variants of the algorithm and encourage future study in reinforcement learning driven by generative exploration. K E Y W O R D S artificial experience-replay, deep reinforcement learning, environment Modelling, exploration, generative adversarial networks, generative Modelling, Markov decision processes, model- based RL, neural networks, variational autoencoder 1 | INTRODUCTION Reinforcement learning (RL) is a field of research that has quickly become one of the most promising branches of machine learning algorithms to solve artificial general intelligence (Arulkumaran, Deisenroth, Brundage and Bharath 2017; Kaelbling, Littman and Moore 1996; Li 2017 and Mousavi, Schukat and Howley 2018). There have been several breakthroughs in reinforcement learning research in recent years for Extension of prior study found in Andersen, Goodwin and Granmo (2017, 2018). Received: 16 May 2019 Revised: 10 October 2019 Accepted: 24 January 2020 DOI: 10.1111/exsy.12537 This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2020 The Authors. Expert Systems published by John Wiley & Sons Ltd. Expert Systems. 2020;e12537. wileyonlinelibrary.com/journal/exsy 1 of 15 https://doi.org/10.1111/exsy.12537 https://orcid.org/0000-0002-7742-4907 https://orcid.org/0000-0001-6331-702X https://orcid.org/0000-0002-7287-030X mailto:per.andersen@uia.no http://creativecommons.org/licenses/by/4.0/ http://wileyonlinelibrary.com/journal/exsy https://doi.org/10.1111/exsy.12537 http://crossmark.crossref.org/dialog/?doi=10.1111%2Fexsy.12537&domain=pdf&date_stamp=2020-03-01 environments such as the Atari Arcade (Mnih et al. 2013, 2015), AlphaZero (Silver et al. 2017), OpenAI Five and AlphaStar (Arulkumaran, Cully and Togelius 2019), but no algorithms are capable of human performance without extensive hardware requirements that are accessi- ble to only a few institutions, such as Deep Mind and Open AI. Due to this, reinforcement learning still has several open research questions to address before the general public can deploy these algorithms successfully. It is possible that many of the issues in the current state of the art can be resolved with algorithms that adequately accounts for planning, exploration and memory efficiency at different time- horizons. In current state-of-the-art RL algorithms, long-horizon RL tasks are difficult to master because there is as of yet no optimal exploration algorithm that is capable of proper state-space pruning. Exploration strategies such as ϵ-greedy are widely used in RL, but cannot find an adequate exploration/exploitation balance without exhaustive hyperparameter-tuning. Environment modelling is a promising exploration method where the goal is to imitate the behaviour of a target environment. By constructing an artificial model of the environment, the need to interact with the ground-truth environment is reduced significantly. This enables the RL-algorithm to explore large parts of the state space without the cost of exhausting the ground-truth environment. A balance between exploration and exploitation must be accounted for to learn the underlying dynamics of the environment. The algorithm must, therefore, select observations carefully to only learn essential fea- tures during the exploration phase. By combining generative modelling with deep reinforcement learning, we find that it is possible for agents to learn optimal policies using only generated training data samples. The approach that we present is the dreaming variational autoencoder with two extensions, based on generative adversarial networks and stochastic weight averaging. We also present a new learning environment, deep maze, that aims to bring a vast set of challenges for reinforcement learning algorithms and is the environment used for testing the presented algorithms. This article is organized as follows. Section 2 surveys recent study in the field. Section 3 briefly introduces the reader to preliminaries. Section 4 proposes The Dreaming Variational Autoencoder for environment modelling to improve exploration in RL. Section 5 introduces the deep maze learning environment for exploration, planning and memory management research for reinforcement learning. Section 6 shows results in the deep line wars environment and that RL agents can be trained to navigate through the deep maze environment using only artificial training data. Finally, we summarize our findings and outlines future study in Section 7. 2 | LITERATURE REVIEW In machine learning, the goal is to create an algorithm that is capable of accurately representing a target environment. There is, however, little literature on generative modelling for game environments on the scale we propose in this article. The primary focus of recent RL research has been through improvements in on and off policy algorithms, while less attention has been put into generative exploration. This section introduces a thorough literature review of reinforcement-based generative modelling. Bangaru, Suhas and Ravindran (2016) proposed a method of deducing the Markov Decision Process (MDP) by introducing an adaptive explo- ration signal (pseudo-reward), which was obtained using deep generative model. Their approach was to compute the Jacobian of each state and used it as the pseudo-reward when using deep neural networks to learn the state-generalization. Xiao and Kesineni (2016) proposed the use of generative adversarial networks (GAN) for model-based reinforcement learning. The goal was to utilize GAN for learning dynamics of the environment in a short-horizon timespan and combine this with the strength of far-horizon value iteration RL algorithms. The GAN architecture proposed illustrated near authentic generated images giving comparable results to Mnih et al. (2013). Higgins et al. (2017) proposed DARLA, an architecture for modelling the environment using β-VAE (Higgins et al. 2016). The trained model was used to learn the optimal policy of the environment using algorithms such as Deep Q-Networks (DQN) (Mnih et al. 2015), Asynchronous Actor-Critic Agents (A3C) (Mnih et al. 2016) and Episodic Control (Blundell et al. 2016). DARLA is to the best of our knowledge, the first algorithm to introduce learning without access to the ground-truth environment during training. Buesing et al. (2018) recently compared several methods of environment modelling, showing that it is far better to model the state-space then to utilize Monte-Carlo rollouts (RAR). The proposed architecture, state-space models (SSM) was significantly faster and produced acceptable results compared to auto-regressive (AR) methods. The algorithm VMAV-C is a combination of VAE and attention-based value function (AVF), and mixture density network recurrent neural network (MDN-RNN) from Ha and Schmidhuber (2018). This modification to the original World Models algorithm improved performance in the Cart Pole environment. They used the on-policy algorithm PPO to learn the optimal policy from the latent representation of the state-space (Liang, Wang, Feng, Liu and Huang 2018). Deep Planning Network (PlaNet) is a model-based agent that interpret the pixels of a state to learn the dynamics of an environment. The environment dynamics are stored into latent-space, where the agent sample actions based on the learned representation. The proposed algorithm showed significantly better sample efficiency compared to algorithms such as A3C (Hafner et al. 2018). 2 of 15 ANDERSEN ET AL. Chua, Calandra, McAllister and Levine (2018) recently proposed Probabilistic Ensembles with Trajectory Sampling (PETS). The algorithm uses an ensemble of bootstrap neural networks to learn a dynamics model of the environment over future states. The algorithm then uses this model to predict the best action for future states. The authors show that the algorithm significantly lowers sampling requirements for environments such as half-cheetah compared to SAC and PPO. Stochastic optimal control with latent representations (SOLAR) is an algorithm that learns the dynamics of the environment by exploiting the knowledge from a reinforcement learning policy. This enables the algorithm to learn local models which are used in policy learning for complex systems. SOLAR is built around a probabilistic graphical model (PGM) structure that allows efficient learning of the environment model. By exploi- ting the locality of the model, the authors show that the gradients give good direction for policy improvements during training. The algorithm was compared to model-free methods and showed significantly better performance and data efficient compared to algorithms such as LQR-FLM. (Zhang, Patras and Haddadi 2018) Ha and Schmidhuber (2018) proposed in Recurrent World Models Facilitate Policy Evolution, a novel architecture for training RL algorithms using variational autoencoders. This article showed that agents could successfully learn the environment dynamics and use this as an exploration technique requiring no interaction with the target domain. The architecture is mainly three components; vision, controller and model, the vision model is a variational autoencoder that outputs a latent-space variable of an observation. The latent-space variable is processed in the model and is fed into the controller for action decisions. Their algorithms show state-of-the-art performance in self-supervised generative modelling for rein- forcement learning agents. One of the most recent advancements in generative modelling for reinforcement learning is the Neural Differential Information Gain Optimisation (NDIGO) algorithm by Azar et al. (2019), a self-supervised exploration model that learns a world model representation from noisy data. The primary features of NDIGO are its robustness to noise due to their method to cancel out negative loss and to give positive learning more value. The authors show in their maze environment that the model successfully converges towards an optimal world model even when introducing noise. The author claims that the algorithm outperforms previous state-of-the-art, being the Recurrent World Model from. 3 | BACKGROUND We base our work on the well-established theory of reinforcement learning originally formulated in Sutton, Precup and Singh (1999), defining the problem as a MDP as 〈S,A,ℛ,T 〉 pairs. The state-space, S represents all possible states while the action-space, A represents all available actions the agent can perform in the environment. ℛ is the reward function while T denotes the transition function T :S×A!Sð Þ, which is a mapping from state st∈S and action at∈A to the future state st + 1. After each performed action, the environment dispatches a reward signal, ℛ :S!r. We call a sequence of states and actions a trajectory denoted as τ = (s0, a0, …, st, at) and the sequence is sampled through the use of a stochas- tic policy that predicts the optimal action in any state: πθ(at| st), where π is the policy and θ are the parameters. The primary goal of the reinforce- ment learning is to reinforce good behaviour. The algorithm should try to learn the policy that maximizes the total expected discounted reward given by, J πð Þ=  st,atð Þ~p πð Þ PT i = 0 γiℛ sið Þ � � Mnih et al. (2015). Several algorithms try to address the problem of reinforcement learning, commonly divided into three categories; Value Iteration, Policy Itera- tion and Actor-Critic Algorithms, with variations of on-policy and off-policy learning. Autoencoders are commonly used in supervised learning to encode arbitrary input to a compact representation, and using a decoder to recon- struct the original data from the encoding. The purpose of autoencoders is to store redundant data into a densely packed vector form. In its sim- plest form, an autoencoder consists of feed-forward neural network where the input and output layer is of equal neuron capacity and the hidden layer smaller, used to compress the data. Such autoencoder can be defined as: ϕ :X!Z, ψ :Z!X, where ϕ,ψ : argminϕ,ψ X− ϕ × ψð ÞX̂ �� ��2. In this TABLE 1 A comparison of features in general reinforcement learning, variational autoencoders and generative adversarial networks Algorithm branch Sample efficient Sequential data Generative modelling Labelled data Reinforcement learning No Yes No No Variational autoencoders Yes Yes Yes Yes Generative adversarial networks Yes No Yes Yes DVAE-GAN Yes Yes Yes Yes Note: The purpose of this illustration is to show that the proposed algorithm uses generative modelling to inherit sample efficient generative modelling for sequential data for use in reinforcement learning algorithms. ANDERSEN ET AL. 3 of 15 notation, X defines the input, Z , the encoding and X̂ as the reconstructed data. There are, however several issues with autoencoders, as they are not generative, in the sense that they can transition between features in the input data. To remedy this, Kingma and Welling (2013) proposed the Variational Autoencoder (VAE) algorithm that enables interpolation between features in the latent space. The interpolation is done by intro- ducing a vector of means μ and a vector of standard deviations σ. These vectors, along with the additional KL-Loss gives the algorithm the ability to learn variations in the input data, enabling the output to be vastly more diverse. Our background for choosing the following branch of algorithms are described in Table 1. Algorithms based on reinforcement learning learns by hands-on interaction. Experience is attained by exploring the environment, but in some cases, the environment may be exhausted before the agent can learn to behave optimally. To remedy this, we introduce a model-based exploration approach using VAE to model the environment with sequential data as input. This increase the sampling efficiency of the RL algorithm by a significant amount. To increase the performance of the environment model, GAN is used to strengthen the quality of the environment-model, where the VAE and GAN continually reinforce each other's performance. 4 | THE DREAMING VARIATIONAL AUTOENCODER The Dreaming Variational Autoencoder (DVAE) is an end-to-end solution for generating probable future states ŝt + n from an arbitrary state-space S using state-action pairs explored prior to st + n and at + n. Figure 1 illustrates the DVAE model, where Algorithm 1 works as follows. First, the agent collects experiences for utilizing experience-replay in the Run-Agent function. At this stage, the agent explores the state-space guided by a Gaussian distributed policy. The agent acts, observes, and stores the observations into the experience-replay buffer D. After the agent reaches terminal state, the DVAE algorithm encodes state-action pairs from the replay-buffer D into probable future states. This is stored in the replay- buffer for artificial future-states D̂. Algorithm 1 The Dreaming Variational Autoencoder 1: Initialize replay memory D and D̂ to capacity N 2: Initialize policy πθ 3: function Run-Agent(T , D) 4: for i = 0 to N_EPOCHS do 5: Observe starting state, s0 ~N 0,1ð Þ 6: while st not TERMINAL do 7: at πθ(st = s) FIGURE 1 Illustration of the DVAE model. The model consumes state and action pairs, yielding the input encoded in latent-space. Latent-space can then be decoded to a probable future state. Q zjXð Þ is the encoder, zt is latent-space and P Xjzð Þ is the decoder. DVAE can also use LSTM to better learn longer sequences in continuous state- spaces 4 of 15 ANDERSEN ET AL. 8: stþ1,rt,terminalt T st,atð Þ 9: Store experience into replay buffer D st,at,rt,stþ1,terminaltð Þ 10: st st + 1 11: end while 12: end for 13: end function 14: Initialize encoder Q zjXð Þ 15: Initialize decoder P Xjzð Þ 16: Initialize DVAE model T θ ¼P XjQ zjXð Þð Þ 17: function DVAE. 18: for di in D do 19: st, at, rt, st + 1 di ⊳ Expand replay buffer pair 20: Xt st, at 21: zt Q Xtð Þ ⊳ Encode Xt into latent-space 22: ŝtþ1 P ztð Þ ⊳ Decode zt into probable future state 23: Store experience into artificial replay buffer D̂(ŝt,at,rt, ŝtþ1,terminalt) 24: ŝt ¼ ŝtþ1 25: end for 26: return D̂ 27: end function Figure 2 illustrates how the algorithm can generate sequences of artificial trajectories using T θ =P XjQ zjXð Þð Þ, where z =Q zjXð Þ is the encoder, and T θ =P Xjzð Þ is the decoder. With state s0 and action Aright as input, the algorithm generates state ŝ1 which in the table can be observed is simi- lar to the real state s1. With the next input, Adown, the DVAE algorithm generates the next state ŝ2 which again can be observed to be equal to s2. Note that this is without ever observing state s1. Hence, the DVAE algorithm needs to be initiated with a state, for example, s0, and actions follows. It then generates (dreams) next states. The requirement is that the environment must be partially discovered so that the algorithm can learn to behave similarly to the target environment. To predict a trajectory of three timesteps, the algorithm does nesting to generate the whole sequence: τ = ŝ1,a1, ŝ2,a2, ŝ3,a3 =T θ T θ T θ s0,Arndð Þ,Arndð Þ,Arndð Þ. The algorithm does this well early on, but have difficulties with longer state sequences in continuous state-spaces. TABLE 2 Feature overview of VAE, GAN, DVAE-GAN and DVAE-SWAGAN Model Features Image quality Data generalization VAE �Converge to local minimum�Stable �Few artefacts�Blurry �Low GAN �Converge to saddle points�Unstable �Artefacts�Sharp �Medium DVAE-GAN �Converge to saddle points�Unstable �Few artefacts�Sharp �Medium DVAE-SWAGAN �Converge to saddle points�Stable �Few artefacts�Sharp �High Note: The highlighted text outlines the benefits of the model. FIGURE 2 DVAE algorithm for generating states using T θ versus the real transition function T . First, a real state is collected from the replay- memory. DVAE can then produce new states from current the trajectory τ using the state-action pairs. θ represent the trainable model parameters ANDERSEN ET AL. 5 of 15 4.1 | Generative adversarial network approach (DVAE-GAN) To combat the divergence behaviour in continuous state-space, we extend the model to use generative adversarial networks. The most common cases of divergence are found where there are (a) complex state-transitions, (b) many state-transitions and (c) sparse state-transitions. In general, we find these properties in continuous and stochastic environments. By using an adversarial approach, we see that DVAE-GAN better generalize for such problems and is far more stable due to increased diversity compared to DVAE (See Table 2 for a detailed comparison). Figure 3 illustrates the proposed DVAE-GAN extension to the original DVAE architecture. In this model, two new components; the generator G and discriminator D is introduced. This is an adversarial approach previously proposed by Makhzani, Shlens, Jaitly, Goodfellow and Frey (2015). The generator G(n| St, At) samples from a Gaussian distribution, while being conditioned on the current state and action, to predict the latent-space distribution zgan. The discriminator D(zvae, zgan) is a neural network that tries to predict the validity of the input, in this case, if the latent-space variable is from the ground-truth distribution. A min-max game between the generator and the discriminator fuels learning where the generator minimizes its error towards the real latent-space and the discriminator learns to distinguish between the real and fake latent distribution. In the VAE model, the latent-space distribution is sampled from zvae = αt + (μt × N(0, 1)) as proposed by Kingma and Welling (2013). The discriminator should then evaluate zvae to be genuine, and its parameters are updated according to the confidence of the prediction. To increase the stability of the VAE, we add the loss of the discriminator to the VAE loss, where we use the original loss function first proposed by Kingma and Welling (2013). 4.2 | Stochastic weight averaging approach (DVAE-SWAGAN) We introduced DVAE-GAN which tries to combat divergence for complex environments. The GAN architecture increases the diversity of the latent- space, but model collapse and instability become a problem. Table 2 outlines a quick overview of the benefits of each of the introduced generative models. VAE is known for its stability but has limited capabilities in its latent-space capacity. The quality of VAE is good but suffers from a severe blur- ring of the decoded latent-space. Compared to other algorithms, the VAE architecture does not generalize well to a diverse data set. The generative adversarial networks have gotten increased attention due to their diverse latent-space. The images of GANs are known for its sharpness, but they suffer from artefacts in the produced output. These networks are state-of-the-art in the sense that they generalize well to the data set. DVAE-GAN as proposed by Makhzani et al. (2015) have few artefacts, which is an improvement from the regular GAN architecture. There is, however, high variance in the model which makes it unstable beyond what is seen in vanilla VAE and GAN. To improve the model instability, we introduce Stochastic weight averaging (SWA), first proposed by Izmailov, Podoprikhin, Garipov, Vetrov and Wilson (2018). DVAE-SWAGAN is a FIGURE 3 The proposed DVAE-GAN architecture. While the original VAE architecture persist, as described in Algorithm 1, a new generative adversarial networks component is added for increased generalization across the latent-space 6 of 15 ANDERSEN ET AL. combination of VAE, SWA and GAN to significantly reduce the variance of the predictions. The algorithm is in principle the averaged ensemble of DVAE-GAN along the trajectory of SGD. We use a cyclical learning rate (Smith 2015) and average the weights each training iteration creat- ing the DVAE-SWAGAN model, seen in Figure 4. 5 | ENVIRONMENTS The DVAE algorithm was tested in two environments. The first environment is the deep line wars that were introduced by Andersen et al. (2017), a simplified Real-Time Strategy game. We present deep maze, a flexible environment with a wide range of challenges suited for reinforcement learning research. The deep line wars environments feature a continuous environment where complex strategies must be planned. The deep maze environment provides simpler rules and a non-continuous action and state-space that in-comparison is far simpler then deep line wars. 5.1 | The deep maze environment The deep maze is a flexible learning environment for controlled research in exploration, planning and memory for reinforcement learning algo- rithms. Maze solving is a well-known problem, and is used heavily throughout the RL literature Sutton et al. (1999), but is often limited to small FIGURE 5 Overview of four distinct MDP scenarios using deep maze. (a) A small, fully observable MDP. (b) A large, fully observable MDP. (c) Partially observable MDP having a vision distance of three tiles. (d) Partially observable MDP having ray-traced vision DVAE-GAN1 DVAE-GAN2 ....... DAVAE-GANn-1 DAVAE-GANn DVAE-SWAGAN OutputInput FIGURE 4 By using ensembles of DVAE-GAN models, the training is significantly more stable and prediction across sparse states yield better results for sequential input ANDERSEN ET AL. 7 of 15 and fully observable scenarios. The deep maze environment extends the maze problem to over 540 unique scenarios including Partially Observ- able Markov Decision Processes (POMDP). Figure 5 illustrates a small subset of the available environments in the deep maze, ranging from small- scale MDP's to large-scale POMDP's. The deep maze further features custom game mechanics such as relocated exits and dynamically changing mazes. RL agents depend on sensory input to evaluate and predict the best action at the current timestep. Preprocessing of data is essential so that agents can extract features from the input data. For this reason, deep maze has built-in state representation for imaging and raw state matri- ces. The game engine is modularized and has an SDK that enables the development of third-party scenarios. This extends the capabilities of deep maze to support nearly all possible scenario combination in the realm of maze solving.1 The deep maze learning environment presents the following scenarios. (a) Normal, (b) POMDP, (c) Limited POMDP and (d) Timed Limited POMDP. The first mode exposes a seed-based randomly generated maze where the state-space an MDP. The second mode narrows the state-space observation to a configurable area around the player. In addition to radius-based vision, the POMDP mode also features ray- tracing vision that better mimic the sight of a physical agent. The third and fourth mode is intended for memory research where the agent must find the goal in a limited number of time-steps. In addition to this, the agent is presented with the solution but fades after a few initial time steps. The objective is for the agent to remember the solution to find the goal. All scenario setups have a variable map-size ranging between 2 × 2 and 56 × 56 tiles. FIGURE 7 The training loss for DVAE in the 2 × 2 No-Wall and 8 × 8 deep maze scenario. The experiment is run for a total of 1000 (5000 for 8 × 8) epochs. The algorithm only trains on 50% of the state-space to the model for the 2 × 2 environment while the whole state-space is trainable in the 8 × 8 environment FIGURE 8 For the 2 × 2 scenario, only 50% of the environment is explored, leaving artefacts on states where the model is uncertain of the transition function. In more extensive examples, the player disappears, teleports or gets stuck in unexplored areas FIGURE 6 The graphical user interface of the deep line wars environment 8 of 15 ANDERSEN ET AL. 5.2 | The deep line wars environment The deep line wars environment was originally introduced in Andersen et al. (2017). Deep line wars is a real-time strategy environment that makes an extensive state-space reduction to enable swift research in reinforcement learning for RTS games. The game objective of deep line wars is to invade the enemy player with mercenary units until all health points are depleted, see Figure 6. For every friendly unit that enters the far edge of the enemy base, the enemy health pool is reduced by one. When a player purchases a merce- nary unit, it spawns at a random location inside the edge area of the buyers base. Mercenary units automatically move towards the enemy base. To protect the base, players can construct towers that shoot projectiles at the opponent's mercenaries. When a mercenary dies, a fair percentage of its gold value is awarded to the opponent. When a player sends a unit, the income is increased by a portion of the units gold value. As a part of the income system, players gain gold at fixed intervals. 6 | EXPERIMENTS We centre our experiments around the DVAE, DVAE-GAN, DVAE-SWA and DVAE-SWAGAN. The goal of the proposed extensions is to improve the model stability so that the DVAE algorithm can produce better quality output for continuous and sparse state-spaces. In this section, we show results of model-based reinforcement learning using DVAE in the deep-maze and deep-line wars environment. We show the performance of encoding raw pixel input to a compact representation and to decode this representation to probable future states. FIGURE 9 Results of 8 × 8 Deep Maze modelling using the DVAE algorithm. To simplify the environment, no reward signal is received per iteration. The left caption describes current state, st, while the right caption is the action performed to compute, st + 1 =T st,atð Þ TABLE 3 Results of the deep maze 11 × 11 and 21 × 21 environment, comparing DQN Mnih et al. (2015), TRPO Schulman et al. (2015) and PPO Schulman et al. (2017) Algorithm 11 × 11 Average performance (%) Converged epoch 21 × 21 Average performance (%) Converged epoch DQN-D̂ 94.56 9314 64.36 N/A TRPO-D̂ 96.32 5320 78.91 7401 PPO-D̂ 98.71 3151 89.33 7195 DQN-D 98.26 4314 84.63 8241 TRPO-D 99.32 3320 92.11 4120 PPO-D 99.35 2453 96.41 2904 Note: The optimal path yields performance of 100% while no solution yields 0%. Each of the algorithms ran 10000 epochs for both map-sizes. Converged Epoch represents at which epoch the algorithm converged during training. ANDERSEN ET AL. 9 of 15 FIGURE 10 A typical deep maze of size 11 × 11. The lower-right square indicates the goal state, the dotted-line is a retrace of the predicted optimal path for that maze, while the final square represents the player's current position in the state-space. The controller agent is DQN, TRPO and PPO (from left to right) FIGURE 11 The DVAE algorithm applied to the deep line wars environment. Each epoch illustrates the quality of generated states in the game, where the left image is real state s and the right image is the generated state ŝ 10 of 15 ANDERSEN ET AL. 6.1 | The dreaming variational autoencoder 6.1.1 | Deep maze For the deep maze environment, the algorithm must be able to generalize over many similar states to model a large state-space. DVAE aims to learn the transition function T st,atð Þ, bringing the state from st to st + 1. We use the deep maze environment because it provides simple rules, with a controllable state-space complexity. Also, we can omit the importance of reward for some scenarios. We trained the DVAE model on two No-Wall deep maze scenarios of size 2 × 2 and 8 × 8. For the encoder and decoder, we used the same convolution architecture as proposed by Pu et al. (2016) and trained for 5000 epochs2 in the 8 × 8 scenario and 1000 epochs for 2 × 2, respec- tively. The reasoning behind different epoch lengths is because we expect simpler environments to converge faster. For the encoding of actions and states, we concatenate the flattened state-space and action-space, having a fully connected layer with ELU activation before calculating the latent-space. We used the Adam optimizer Kingma and Ba (2015) with a learning-rate of 3e-05 to update the parameters. To calculate the loss we used the same loss function as proposed by Kingma and Welling (2013). Figure 7 illustrates the loss during the training phase for the DVAE algorithm in the No-Wall Deep Maze scenario. In the 2 × 2 scenario, DVAE is trained on only 50% of the state space, which results in noticeable graphics artefacts in the prediction of future states, seen in Figure 8. In the 8 × 8 environment, the algorithm is allowed to train on all possible states, and we observe in Figure 9 that there is significantly better image quality throughout the sampling process. (a) (b) (c) FIGURE 12 Training of DVAE compared to the DVAE-SWA, DVAE-GAN and DVAE-SWAGAN extension. (a) Total Loss for the tested algorithms. (b) Autoencoder loss (L2) for the algorithms. (c) Variational Autoencoder Loss (KL) for the tested algorithms ANDERSEN ET AL. 11 of 15 6.2 | Training RL-agents using D̂ replay-buffer The goal of this experiment is to observe the performance of RL-agents using the generated experience-replay D̂ from Algorithm 1 in Deep Maze environments of size 11 × 11 and 21 × 21. In Table 3, we compare the performance of DQN Mnih et al. (2013), TRPO Schulman, Levine, Abbeel, Jordan and Moritz (2015) and PPO Schulman, Wolski, Dhariwal, Radford and Klimov (2017) using the DVAE generated D̂ to tune the parameters. Because TRPO and PPO are on-policy algorithms, the generated states must be generated on-the-fly so that the algorithm remains on-policy. Figure 10 illustrates three maze variations of size 11 × 11, where the agent has learned the optimal path. We see that the best performing algorithm, PPO Schulman et al. (2017) beats DQN and TRPO using eitherD̂ or D. The DQN-D̂ agent did not converge in the 21 × 21 environment, but it is likely that value-based algorithms could struggle to map inaccurate states with graphical artefacts generated from the DVAE algorithm. These artefacts significantly increase the state-space significantly, but empirical data suggest that on-policy algorithms perform better on noisy state-spaces. 6.3 | Deep line wars The DVAE algorithm works well in more complex environments, such as the deep line wars game environment Andersen et al. (2017). Here, we expand the DVAE algorithm with LSTM to improve the capability of generating time-bound data, such as animations seen in Figure 1. Figure 11 illustrates the state quality during training of DVAE in a total of 6000 epochs. Both players draw actions from a Gaussian dis- tributed policy. The algorithm understands that the player units can be located in any tiles after only 50 epochs, and at 1000 epochs we observe the algorithm makes significantly better predictions of the probability of unit locations (i.e. some units show more densely in the output state). At the end of the training, the DVAE algorithm is to some degree capable of determining both towers, and unit locations at any given time-step during the game epoch. 6.4 | Extending the dreaming variational autoencoder The goal of DVAE-SWA, DVAE-GAN and DVAE-SWAGAN is to perform better in continuous state-spaces, such as the deep line wars environ- ment. The experiments were performed using a map size of 11 × 11 sampling actions from a PPO policy. FIGURE 13 The first row represents the ground truth future state, while the second row is the predicted future state. The DVAE-SWAGAN algorithm sampled states after training for 1000 epochs, see Figure 12. Notice that the quality is notably better compared to Figure 11. We found that states were generalized too much producing some inaccurate predictions. Despite this inaccuracy, the RL agents learned a policy capable of beating random agents. We believe this is because similar states often represent similar value functions 12 of 15 ANDERSEN ET AL. Figure 12 shows the training loss of the algorithms DVAE, DVAE-SWA, DVAE-GAN and DVAE-SWAGAN for 1000 epochs (x-axis), where the y-axis describes the loss value. The new architectures perform significantly better than DVAE across all loss components of the architecture. The consequence of lower loss is better image quality (autoencoder loss), and better transitions (variational loss). For these experiments, we tried to model the deep line wars environment using RGB input. Figure 13 illustrates the resulting images for each of the algorithms. Here, we see that DVAE-SWAGAN perform significantly best, in terms of quality and accuracy. 7 | CONCLUSION AND FUTURE STUDY This article introduces The Dreaming Variational Autoencoder along with its extensions DVAE-SWA, DVAE-GAN and DVAE-SWAGAN as a neural network-based generative modelling architecture to enable exploration in environments with sparse reward. The DVAE algorithm successfully generates authentic world models in non-continuous state-spaces where the dynamics of the environment is simple. It works well for small environments but is limited when the state-sequence become too large. The algorithm performs marginally better when using LSTM for sequence prediction, but is a significant performance drop due to the increased model complexity. For most environments, such as deep line wars and deep maze, it is sufficient to run DVAE using only fully connected nodes. The DVAE-SWAGAN improves the original model significantly and enables the algorithm to imitate environment models with continuous state-space better. DVAE-SWAGAN performs better in all environments including deep line wars and most environments found in the GYM reinforcement learning environment. There are, however, several fundamental issues that limit DVAE, and DVAE-SWAGAN from fully modelling environments. In some situations, exploration may be a costly act that makes it impossible to explore all parts of the environment in its entirety. The algorithms cannot accurately predict the outcome of unexplored areas of the state-space, making the prediction blurry or incorrect. To combat this, the model should be improved further to include some sense of logic, and understanding of the environment dynamics. In current state-of- the-art, this frequently introduced as domain knowledge that is manually crafted by the programmer, but the hope is that future research will find a method for self-supervised domain knowledge modelling. Reinforcement learning has many unresolved problems, and the hope is that the deep maze and the deep line wars learning environment can be a useful tool for future research. For future study, we plan to introduce an inverse reinforcement learning component to learn the reward function ℛ̂. We also plan to explore non-parametric variants. DVAE and environment modelling is an ongoing research question, and the goal is that reinforcement learning algorithms could utilize this form of dreaming to make the algorithm far more sample efficient. ORCID Per-Arne Andersen https://orcid.org/0000-0002-7742-4907 Morten Goodwin https://orcid.org/0000-0001-6331-702X Ole-Christoffer Granmo https://orcid.org/0000-0002-7287-030X ENDNOTES 1 The deep maze is open-source and publicly available at https://github.com/CAIR/deep-maze. 2 We regard the term epoch as an episode, which is frequently in literature. REFERENCES Andersen, P.-A., Goodwin, M., & Granmo, O.-C. (2017). Towards a deep reinforcement learning approach for tower line wars. In M. Bramer & M. Petridis (Eds.), Lecture notes in computer science (including subseries lecture notes in artificial intelligence and lecture notes in bioinformatics) (Vol. 10630 LNAI, pp. 101–114) NY, USA: Springer, Cham. https://doi.org/10.1007/978-3-319-71078-5_8 Andersen, P.-A., Goodwin, M., & Granmo, O.-C. (2018, Dec). The dreaming variational autoencoder for reinforcement learning environments. In Max Bramer & M. Petridis (Eds.), Artificial intelligence (XXXV ed 11311, pp. 143–155). Springer, Cham. Retrieved from http://link.springer.com/10. 1007/978-3-030-04191-5_11 doi: https://doi.org/10.1007/978-3-030-04191-5_11 Arulkumaran, K., Cully, A., & Togelius, J. (2019). AlphaStar: An evolutionary computation perspective (Tech. Rep.). Retrieved from https://deepmind.com/ blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/ Arulkumaran, K., Deisenroth, M. P., Brundage, M., & Bharath, A. A. (2017). Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine, 34 (6), 26–38. Retrieved from https://ieeexplore.ieee.org/document/8103164/. doi: https://doi.org/10.1109/MSP.2017.2743240 Azar, M. G., Piot, B., Pires, B. A., Grill, J.-B., Altché, F., & Munos, R. (2019, feb). World discovery models. arxiv preprint arXiv:1902.07685. Retrieved from http://arxiv.org/abs/1902.07685 Bangaru, S. P., Suhas, J., & Ravindran, B. (2016, nov). Exploration for multi-task reinforcement learning with deep generative models. arxiv preprint arXiv: 1611.09894. Retrieved from http://arxiv.org/abs/1611.09894 Blundell, C., Uria, B., Pritzel, A., Li, Y., Ruderman, A., Leibo, J. Z., … Hassabis, D. (2016, jun). Model-free episodic control. arxiv preprint arXiv:1606.04460. Retrieved from http://arxiv.org/abs/1606.04460 ANDERSEN ET AL. 13 of 15 https://orcid.org/0000-0002-7742-4907 https://orcid.org/0000-0002-7742-4907 https://orcid.org/0000-0001-6331-702X https://orcid.org/0000-0001-6331-702X https://orcid.org/0000-0002-7287-030X https://orcid.org/0000-0002-7287-030X https://github.com/CAIR/deep-maze https://doi.org/10.1007/978-3-319-71078-5_8 http://link.springer.com/10.1007/978-3-030-04191-5_11 http://link.springer.com/10.1007/978-3-030-04191-5_11 https://doi.org/10.1007/978-3-030-04191-5_11 https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/ https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/ https://ieeexplore.ieee.org/document/8103164/ https://doi.org/10.1109/MSP.2017.2743240 http://arxiv.org/abs/1902.07685 http://arxiv.org/abs/1611.09894 http://arxiv.org/abs/1606.04460 Buesing, L., Weber, T., Racaniere, S., Eslami, S. M. A., Rezende, D., Reichert, D. P., … Wierstra, D. (2018, feb). Learning and querying fast generative models for reinforcement learning. Arxiv Preprint arXiv:1802.03006. Retrieved from http://arxiv.org/abs/1802.03006 Chua, K., Calandra, R., McAllister, R., & Levine, S. (2018, may). Deep reinforcement learning in a handful of trials using probabilistic dynamics models. Advances in Neural Information Processing Systems 31, 4759–4770. Retrieved from http://arxiv.org/abs/1805.12114 Ha, D., & Schmidhuber, J. (2018, sep). Recurrent world models facilitate policy evolution. Advances in Neural Information Processing Systems 31, 2455– 2467. Retrieved from http://arxiv.org/abs/1809.01999 Hafner, D., Lillicrap, T., Fischer, I., Villegas, R., Ha, D., Lee, H., & Davidson, J. (2018, nov). Learning latent dynamics for planning from pixels. in Proceedings of the 36th International Conference on Machine Learning. Retrieved from http://arxiv.org/abs/1811.04551 Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., … Lerchner, A. (2016, nov). beta-VAE: Learning basic visual concepts with a con- strained variational framework. in International Conference on Learning Representations. Retrieved from https://openreview.net/forum?id= Sy2fzU9gl Higgins, I., Pal, A., Rusu, A., Matthey, L., Burgess, C., Pritzel, A., … Lerchner, A. (2017). DARLA: Improving zero-shot transfer in reinforcement learning. In D. Precup & Y. W. Teh (Eds.), in Proceedings of the 34th International Conference on Machine Learning (Vol. 70, pp. 1480–1490). International Convention Centre, Sydney, Australia: PMLR. Retrieved from http://proceedings.mlr.press/v70/higgins17a.html Izmailov, P., Podoprikhin, D., Garipov, T., Vetrov, D., & Wilson, A. G. (2018, mar). Averaging weights leads to wider optima and better generalization. Retrieved from http://arxiv.org/abs/1803.05407 Kaelbling, L. P., Littman, M. L., & Moore, A. W. (1996, apr). Reinforcement learning: A survey. Journal of Artificial Intelligence Research, 4, 237–285. Retrieved from http://arxiv.org/abs/cs/9605103 doi: 10.1.1.68.466 Kingma, D. P., & Ba, J. L. (2015). Adam: A method for stochastic optimization. in Proceedings, International Conference on Learning Representations 2015. doi: https://doi.org/10.1145/1830483.1830503 Kingma, D. P., & Welling, M. (2013, dec). Auto-encoding variational Bayes. arxiv preprint arXiv:1312.6114. Retrieved from http://arxiv.org/abs/1312.6114 doi: https://doi.org/10.1051/0004-6361/201527329 Li, Y. (2017, jan). Deep reinforcement learning: An overview. Arxiv preprint arXiv:1701.07274. Retrieved from http://arxiv.org/abs/1701.07274 Liang, X., Wang, Q., Feng, Y., Liu, Z., & Huang, J. (2018, dec). VMAV-C: A deep attention-based reinforcement learning algorithm for model-based control. arxiv preprint arXiv:1812.09968. Retrieved from http://arxiv.org/abs/1812.09968 Makhzani, A., Shlens, J., Jaitly, N., Goodfellow, I., & Frey, B. (2015, nov). Adversarial autoencoders. Retrieved from http://arxiv.org/abs/1511.05644 Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap, T., Harley, T., … Kavukcuoglu, K. (2016). Asynchronous methods for deep reinforcement learning. In M. F. Balcan & K. Q. Weinberger (Eds.), Proceedings of the 33rd International Conference on Machine Learning (Vol. 48, pp. 1928–1937). New York, USA: PMLR. Retrieved from http://proceedings.mlr.press/v48/mniha16.html Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., & Riedmiller, M. (2013, dec). Playing atari with deep reinforcement learning. Neural Information Processing Systems. Retrieved from http://arxiv.org/abs/1312.5602 Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves A., Riedmiller M., Fidjeland A. K., Ostrovski G., Petersen S., Beattie C., Sadik A., Antonoglou I., King H., Kumaran D., Wierstra D., Legg S. Hassabis, D. (2015, feb). Human-level control through deep reinforcement learning. Nature, 518(7540), 529–533. Retrieved from https://doi.org/10.1038/nature14236 Mousavi, S. S., Schukat, M., & Howley, E. (2018). Deep reinforcement learning: An overview. In Y. Bi, S. Kapoor, & R. Bhatia (Eds.), Proceedings of Sai Intelli- gent Systems Conference (Intellisys) 2016 (pp. 426–440). Cham: Springer International Publishing. Retrieved from http://link.springer.com/10.1007/ 978-3-319-56991-8_32 Pu, Y., Gan, Z., Henao, R., Yuan, X., Li, C., Stevens, A., & Carin, L. (2016). Variational autoencoder for deep learning of images, labels and captions. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, & R. Garnett (Eds.), Advances in neural information processing systems (pp. 2352–2360). NY, USA: Cur- ran Associates, Inc. Retrieved from http://papers.nips.cc/paper/6528-variational-autoencoder-for-deep-learning-of-images-labels-and-captions.pdf Schulman, J., Levine, S., Abbeel, P., Jordan, M., & Moritz, P. (2015). Trust region policy optimization. In F. Bach & D. Blei (Eds.), Proceedings of the 32nd International Conference on Machine Learning (Vol. 37, pp. 1889–1897). Lille, France: PMLR. Retrieved from http://proceedings.mlr.press/ v37/schulman15.html Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017, jul). Proximal policy optimization algorithms. arxiv preprint arXiv:1707.06347. Retrieved from http://arxiv.org/abs/1707.06347 Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., … Hassabis, D. (2017). Mastering the game of go without human knowledge. Nature, 550, 354–359. https://doi.org/10.1038/nature24270 Smith, L. N. (2015, jun). Cyclical learning rates for training neural networks. Retrieved from http://arxiv.org/abs/1506.01186 Sutton, R. S., Precup, D., & Singh, S. (1999). Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning (Vol. 112; Tech. Rep.). Xiao, T., & Kesineni, G. (2016). Generative adversarial networks for model based reinforcement learning with tree search (Tech. Rep.). Berkeley: University of California. Retrieved from http://tedxiao.me/pdf/gans_drl.pdf Zhang, C., Patras, P., & Haddadi, H. (2018, mar). Deep learning in mobile and wireless networking: A survey. IEEE Communications Surveys and Tutorials, 21, 2224–2287. Retrieved from http://arxiv.org/abs/1803.04311 AUTHOR BIOGRAPHIES Per-Arne Andersen is a Ph.D. candidate in Artificial Intelligence at the University of Agder, Norway where he also com- pleted his B.Sc. and M.Sc. in 2015 and 2017, respectively. He has expertise in several domains in computer science, including security, software development and machine learning. His primary field of research is reinforcement learning in environments with incomplete information, including real-time games and industry-near systems. Towards the future, Per-Arne hopes to have an active role in the scientific community for reinforcement learning and to contribute towards the understanding of artificial general intelligence. 14 of 15 ANDERSEN ET AL. http://arxiv.org/abs/1802.03006 http://arxiv.org/abs/1805.12114 http://arxiv.org/abs/1809.01999 http://arxiv.org/abs/1811.04551 https://openreview.net/forum?id=Sy2fzU9gl https://openreview.net/forum?id=Sy2fzU9gl http://proceedings.mlr.press/v70/higgins17a.html http://arxiv.org/abs/1803.05407 http://arxiv.org/abs/cs/9605103 https://doi.org/10.1145/1830483.1830503 http://arxiv.org/abs/1312.6114 https://doi.org/10.1051/0004-6361/201527329 http://arxiv.org/abs/1701.07274 http://arxiv.org/abs/1812.09968 http://arxiv.org/abs/1511.05644 http://proceedings.mlr.press/v48/mniha16.html http://arxiv.org/abs/1312.5602 https://doi.org/10.1038/nature14236 http://link.springer.com/10.1007/978-3-319-56991-8_32 http://link.springer.com/10.1007/978-3-319-56991-8_32 http://papers.nips.cc/paper/6528-variational-autoencoder-for-deep-learning-of-images-labels-and-captions.pdf http://proceedings.mlr.press/v37/schulman15.html http://proceedings.mlr.press/v37/schulman15.html http://arxiv.org/abs/1707.06347 https://doi.org/10.1038/nature24270 http://arxiv.org/abs/1506.01186 http://tedxiao.me/pdf/gans_drl.pdf http://arxiv.org/abs/1803.04311 Morten Goodwin received the B.Sc. and M.Sc. degrees from University of Agder, Norway, in 2003 and 2005, respec- tively, and the Ph.D. degree from Aalborg University Department of Computer Science, Denmark, in 2011, on applying machine learning algorithms on eGovernment indicators which are difficult to measure automatically. He is an Associate Professor with the Department of ICT, University of Agder, deputy director for Centre for Artificial Intelligence Research, coordinator for the International Master's Programme in Artificial Intelligence, a public speaker and an active researcher. His main interests include swarm intelligence, deep learning and adaptive learning in the fields of medicine, games and chatbots. He has more than 70 peer reviews scientific publications, and has supervised more than 110 student projects including Master's and Ph.D. theses. Prof. Ole-Christoffer Granmo is director and founder of the Centre for Artificial Intelligence Research (CAIR) at the Uni- versity of Agder, Norway. He obtained his master's degree in 1999 and the Ph.D. degree in 2004, both from the Univer- sity of Oslo, Norway. Granmo develops theory and algorithms for systems that explore, experiment and learn in complex real-world environments. His research interests include artificial intelligence, machine learning, learning automata, bandit algorithms, deep reinforcement learning, Bayesian reasoning and computational linguistics. Within these areas of research, Dr. Granmo has written more than 125 refereed journal and conference publications. He is co-founder of the Norwegian Artificial Intelligence Consortium (NORA). Apart from his academic endeavours, Granmo is also co-founder of the company Anzyz Technologies AS. How to cite this article: Andersen P-A, Goodwin M, Granmo O-C. Increasing sample efficiency in deep reinforcement learning using generative environment modelling. Expert Systems. 2020;e12537. https://doi.org/10.1111/exsy.12537 ANDERSEN ET AL. 15 of 15 https://doi.org/10.1111/exsy.12537 Increasing sample efficiency in deep reinforcement learning using generative environment modelling 1 INTRODUCTION 2 LITERATURE REVIEW 3 BACKGROUND 4 THE DREAMING VARIATIONAL AUTOENCODER 4.1 Generative adversarial network approach (DVAE-GAN) 4.2 Stochastic weight averaging approach (DVAE-SWAGAN) 5 ENVIRONMENTS 5.1 The deep maze environment 5.2 The deep line wars environment 6 EXPERIMENTS 6.1 The dreaming variational autoencoder 6.1.1 Deep maze 6.2 Training RL-agents using D replay-buffer 6.3 Deep line wars 6.4 Extending the dreaming variational autoencoder 7 CONCLUSION AND FUTURE STUDY Endnotes REFERENCES 
work_sfc3mik6snaw7buodbobenrsra	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/authorsrights http://www.elsevier.com/authorsrights Author's personal copy Cluster center initialization algorithm for K-modes clustering Shehroz S. Khan a,⇑, Amir Ahmad b a David R. Cheriton School of Computer Science, University of Waterloo, Canada b Faculty of Computing and Information Technology, King Abdulaziz University, Rabigh, Saudi Arabia a r t i c l e i n f o Keywords: K-modes clustering Cluster center initialization Prominent attributes Significant attributes a b s t r a c t Partitional clustering of categorical data is normally performed by using K-modes clustering algorithm, which works well for large datasets. Even though the design and implementation of K-modes algorithm is simple and efficient, it has the pitfall of randomly choosing the initial cluster centers for invoking every new execution that may lead to non-repeatable clustering results. This paper addresses the randomized center initialization problem of K-modes algorithm by proposing a cluster center initialization algorithm. The proposed algorithm performs multiple clustering of the data based on attribute values in different attributes and yields deterministic modes that are to be used as initial cluster centers. In the paper, we propose a new method for selecting the most relevant attributes, namely Prominent attributes, com- pare it with another existing method to find Significant attributes for unsupervised learning, and perform multiple clustering of data to find initial cluster centers. The proposed algorithm ensures fixed initial cluster centers and thus repeatable clustering results. The worst-case time complexity of the proposed algorithm is log-linear to the number of data objects. We evaluate the proposed algorithm on several cat- egorical datasets and compared it against random initialization and two other initialization methods, and show that the proposed method performs better in terms of accuracy and time complexity. The initial cluster centers computed by the proposed approach are close to the actual cluster centers of the different data we tested, which leads to faster convergence of K-modes clustering algorithm in conjunction to bet- ter clustering results. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction Cluster analysis is a form of unsupervised learning that is aimed at finding underlying structures in the unlabeled data. The objective of a clustering algorithm is to partition a multi- attribute dataset into homogeneous groups (or clusters) such that the data objects in one cluster are more similar to each other (based on some similarity measure) than those in other clusters. Clustering is an active research topic in pattern recognition, data mining, statistics and machine learning with diverse application such as in image analysis (Matas & Kittler, 1995), medical appli- cations (Petrakis & Faloutsos, 1997) and web documentation (Boley et al., 1999). The partitional clustering algorithms such as K-means (Anderberg, 1973) are very efficient for processing large numeric datasets. Data mining applications require handling and explora- tion of data that contains numeric, categorical or both types attributes. The K-means clustering algorithm fails to handle data- sets with categorical attributes because it minimizes the cost function by calculating means and distances. The traditional way to treat categorical attributes as numeric does not always produce meaningful results because generally categorical do- mains are not ordered. Several approaches have been reported for clustering categorical datasets that are based on K-means paradigm. Ralambondrainy (1995) presents an approach by using K-means algorithm to cluster categorical data by converting mul- tiple category attributes into binary attributes (using 0 and 1 to represent either a category absent or present) and treats the bin- ary attributes as numeric in the K-means algorithm. Gowda and Diday (1991) use a similarity coefficient and other dissimilarity measures to process data with categorical attributes. CLARA (Clustering LARge Application) (Kaufman & Rousseeuw, 1990) is a combination of a sampling procedure and the clustering pro- gram Partitioning Around Medoids (PAM). Guha, Rastogi, and Shim (1999) present a robust hierarchical clustering algorithm, ROCK, that uses links to measure the similarity/proximity be- tween a pair of data objects with categorical attributes that are used to merge clusters. However, this algorithm has worst-case quadratic time complexity. Huang (1997) presents the K-modes clustering algorithm by introducing a new dissimilarity measure to cluster categorical data. The algorithm replaces means of clusters with modes (most 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.07.002 ⇑ Corresponding author. E-mail addresses: s255khan@uwaterloo.ca (S.S. Khan), amirahmad01@gmail. com (A. Ahmad). Expert Systems with Applications 40 (2013) 7444–7456 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy frequent attribute value in a attribute), and uses a frequency based method to update modes in the clustering process to minimize the cost function. The algorithm is shown to achieve convergence with linear time complexity with respect to the number of data objects. Huang (1998) also points out that in general, the K-modes algo- rithm is faster than the K-means algorithm because it needs less iterations to converge. In principle, K-modes clustering algorithm functions similar to K-means clustering algorithm except for the cost function it minimizes, and hence suffers from the same draw- backs. Similar to K-means clustering algorithm, the K-modes clus- tering algorithm assumes that the number of clusters, K, is known in advance. Fixed number of K clusters can make it difficult to pre- dict the actual number of clusters in the data that may mislead the interpretations of the results. The K-means/K-modes clustering algorithm falls into problems when clusters are of differing sizes, density and non-globular shapes. The K-means clustering algo- rithm does not guarantee unique clustering due to random choice of initial cluster centers that may yield different groupings for dif- ferent runs (Jain & Dubes, 1988). Similarly, the K-modes algorithm is also very sensitive to the choice of initial cluster centers and an improper choice may result in highly undesirable cluster struc- tures. Random initialization is widely used as a seed for K-modes algorithm for its simplicity, however, this may lead to non-repeat- able clustering results. Machine learning practitioners find it diffi- cult to rely on the results thus obtained and several re-runs of K- modes algorithm may be required to arrive at a meaningful conclusion. In this paper, we extend the work of Khan and Ahmad (2012) and present a multiple clustering approach that infers cluster structure information from multiple attributes by using the attri- bute values present in the data for computing initial cluster cen- ters. This approach focus only on Prominent attributes (discussed in Section 4.2) that are important for finding cluster structures. We also use another unsupervised learning method to find Signif- icant attributes (Ahmad & Dey, 2007a, 2007b) and compare it with the proposed approach. The proposed algorithm performs multiple clustering based on distinct attribute values present in different attributes to generate multiple clustering views of the data that are utilized to obtain fixed initial cluster centers (modes) for K-modes clustering algorithm. The proposed algo- rithm has worst-case log-linear time complexity with respect to the number of data objects. The present paper extends the previ- ous work in terms of: � Using a unsupervised method to compute significant attributes and compare their clustering performance and quality of initial cluster centers against the centers computed by prominent attributes. � Comparing the quality of initial cluster centers by using all attri- butes and prominent attributes. � Analyzing the closeness of initial cluster centers to the actual centers by using prominent, significant and all attributes. � Performing comprehensive experiments, presentation of results, time scalability analysis, inclusion of more datasets, and extended discussions on the multiple clustering challenges from the perspective of the proposed approach. The rest of the paper is organized as follows. In Section 2, we present a short survey of the research work on cluster center ini- tialization for K-modes algorithm. Section 3 briefly discusses the K-modes clustering algorithm. In Section 4, we present the pro- posed multiple attribute clustering approach to compute initial cluster centers along with three different approaches to choose dif- ferent number of attributes to generate multiple clustering views. Section 5 shows the detailed experimental analysis of the proposed method on various categorical datasets and compare it with other cluster center initialization methods. Section 6 concludes the paper with pointers to future work. 2. Related work The K-modes algorithm (Huang, 1997) extends the K-means paradigm to cluster categorical data and requires random selection of initial cluster centers or modes. As discussed earlier, a random choice of initial cluster centers leads to non-repeatable clustering results that may be difficult to comprehend. The random initializa- tion of cluster centers may only work well when one or more ran- domly selected initial cluster centers are similar to the actual cluster centers present in the data. In the most trivial case, the K-modes algorithm keeps no control over the choice of initial clus- ter centers and therefore repeatability of clustering results is diffi- cult to achieve. Moreover, an inappropriate choice of initial cluster centers can lead to undesirable clustering results. The results of partitional clustering algorithms are better when the initial parti- tions are close to the final solution (Jain & Dubes, 1988). Hence, it is important to invoke K-modes clustering with fixed initial clus- ter centers that are similar to the true representative centers of the actual clusters to get better results. There are several research papers reported for computing initial cluster centers for K-modes algorithm, however, most of these methods suffer from either of the following two drawbacks: (a) The initial cluster center computation methods are quadratic in time complexity with respect to the number of data objects – these type of methods mitigate the advantage of linear time complexity of K-modes algorithm and are not scalable for large datasets. (b) The initial cluster centers are not fixed and have randomness in the computation steps – these type of methods fare as good as random initialization methods. We present a short review of the research work done to com- pute initial cluster centers for K-modes clustering algorithm and discuss their associated problems. Khan and Ahmad (2003) use density-based multiscale data con- densation (Mitra, Murthy, & Pal, 2002) approach with Hamming distance to extract K initial cluster centers from the datasets, how- ever, their method has quadratic complexity with respect to the number of data objects. Huang (1998) proposes two approaches for initializing the cluster centers for K-modes algorithm. In the first method, the first K distinct data objects are chosen as initial K-modes, whereas the second method calculates the frequencies of all categories for all attributes and assign the most frequent cat- egories equally to the initial K-modes. The first method may only work if the top K data objects come from disjoint K clusters. The second method is aimed at choosing diverse cluster center that may improve clustering results, however a uniform criteria for selecting K-initial cluster centers is not provided. Sun, Zhu, and Chen (2002) present an experimental study on applying Bradley and Fayyad’s iterative initial-point refinement algorithm (Bradley & Fayyad, 1998) to the K-modes clustering to improve the accuracy and repetitiveness of the clustering results. Their experiments show that the K-modes clustering algorithm using refined initial cluster centers leads to higher precision results that are more reliable than the random selection method without refinement. This method is dependent on the number of cases with refinements and the accuracy value varies. Khan and Kant (2007) propose a method that is based on the idea of evidence accumula- tion for combining the results of multiple clusterings (Fred & Jain, 2002) and only focus on those data objects that are less vulnerable to the choice of random selection of modes and to choose the most S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 7445 Author's personal copy diverse set of modes among them. Their experiments suggest that the computed initial cluster centers outperform the random choice, however the method does not guarantee fixed choice of initial cluster centers. He (2006) presents two farthest point heu- ristic for computing initial cluster centers for K-modes algorithm. The first heuristic is equivalent to random selection of initial clus- ter centers and the second uses a deterministic method based on a scoring function that sums the frequency count of attribute values of all data objects. This heuristic does not explain how to choose a point when several data objects have same scores, and if it ran- domly breaks ties, then fixed centers cannot be guaranteed. The method only considers the distance between the data points, due to which outliers can be selected as cluster centers. Wu, Jiang, and Huang (2007) develop a density based method to compute the K initial cluster centers which has quadratic complex- ity. To reduce the worst case complexity to O(n1.5), they randomly select square root of the total points as a sub-sample of the data, however, this step introduces randomness in the final results and repeatability of clustering results may not be achieved. Cao, Liang, and Bai (2009) present an initialization method that consider dis- tance between objects and the density of the objects. Their method selects the object with the maximum average density as the first initial cluster center. For computing other cluster centers, the dis- tance between the object and the already known clusters, and the average density of the object are considered. A shortcoming of this method is that a boundary point may be selected as the first center that can affect the quality of selection of subsequent initial cluster centers. Bai, Liang, Dang, and Cao (2012) propose a method to com- pute initial cluster centers on the basis of a density function (de- fined by using the average distance of all the other points from a point) and a distance function. The first cluster center is decided by the density function. The remaining cluster centers are com- puted by using the density function and the distance between the already calculated cluster centers and the probable new cluster center. In order to calculate the density of a point they calculate the summary of all the other points. Hence, there is information loss that may lead to improper density calculation, which can affect the results. A major problem with this research paper lies in the evaluation of results. For at least two datasets, the accuracy, preci- sion and recall values are computed incorrectly. From the confu- sion matrix presented in the paper the accuracy, precision and recall values for � Dermatology data should be 0.6584, 0.6969, 0.6841 and not 0.7760, 0.8527, 0.7482 � Zoo data should be 0.7425, 0.7703, 0.8654 and not 0.9208, 0.8985 and 0.8143. The confusion matrix mis-classify almost half of the data objects of first cluster and thus accuracy cannot reach the value indicated in the paper. In comparison to the above stated research works, the proposed algorithm (see Section 4 for details) for finding initial clusters cen- ters for categorical datasets circumvents both the drawbacks dis- cussed earlier i.e. its worst-case time complexity is log-linear in the number of data objects and it provides deterministic (fixed) initial cluster centers. 3. K-modes algorithm for clustering categorical data Due to the limitation of the dissimilarity measure used by tra- ditional K-means algorithm, it cannot be used to cluster categorical dataset. The K-modes clustering algorithm is based on K-means paradigm, but removes the numeric data limitation whilst preserv- ing its efficiency. The K-modes algorithm (Huang, 1998) extends the K-means paradigm to cluster categorical data by removing the barrier imposed by K-means through following modifications: 1. Using a simple matching dissimilarity measure or the Hamming distance for categorical data objects. 2. Replacing means of clusters by their modes (cluster centers). The simple matching dissimilarity measure (Hamming dis- tance) can be defined as following. Let X and Y be two categorical data objects described by m categorical attributes. The dissimilar- ity measure d(X,Y) between X and Y can be defined by the total mis- matches of the corresponding attribute categories of two objects. Smaller the number of mismatches, more similar the two objects are. Mathematically, we can say dðX; YÞ¼ Xm j¼1 dðxj; yjÞ ð1Þ where dðxj; yjÞ¼ 0 ðxj ¼ yjÞ 1 ðxj – yjÞ � , and d(X,Y) gives equal importance to each category of an attribute. Let N be a set of n categorical data objects described by m cat- egorical attributes, M1, M2 , . . . , Mm. When the distance function de- fined in Eq. (1) is used as the dissimilarity measure for categorical data objects, the cost function becomes CðQÞ¼ Xn i¼1 dðNi; Q iÞ ð2Þ where Ni is the ith element and Qi is the nearest cluster center of Ni. The K-modes algorithm minimizes the cost function defined in Eq. (2). The K-modes algorithm assumes that the knowledge of number of natural grouping of data (i.e. K) is available and consists of the following steps (taken from Huang (1997)): 1. Select K initial cluster centers, one for each of the cluster. 2. Allocate data objects to the cluster whose cluster center is near- est to it according to Eq. (2). Update the K clusters based on allo- cation of data objects and compute K new modes of all clusters. 3. Retest the dissimilarity of objects against the current modes. If an object is found such that its nearest mode belongs to another cluster rather than its current one, reallocate the object to that cluster and update the modes of both clusters. 4. Repeat step 3 until no data object has changed cluster membership. 4. Proposed approach for computing initial cluster centers using multiple attribute clustering Khan and Ahmad (2004) show that for partitional clustering algorithms, such as K-Means, � Some of the data objects are very similar to each other, that is why they share same cluster membership irrespective of the choice of initial cluster centers, and � An individual attribute may also provide some information about initial cluster centers He, Xu, and Deng (2005) present a unified view on categorical data clustering and cluster ensemble for the creation of new clus- tering algorithms for categorical data. Their intuition is that the attributes present in a categorical data contribute to the final cluster structure. They consider the distinct attribute values of an 7446 S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 Author's personal copy attribute as cluster labels giving ‘‘best clustering’’ without consid- ering other attributes and create a cluster ensemble. Müller, Günnemann, Färber, and Seidl (2010) defines multiple clusterings as setting up multiple set of clusters for every data object in a dataset with respect to multiple views on the data. The basic objective of multiple clustering is to represent different perspectives on the data and utilize the variation among the clustering results to gain additional knowledge about the structure in the data. They discuss several challenges that arise due to multi- ple clustering of data and merging of their results. One of the major challenges is related to the detection of different clusterings re- vealed by multiple views on the data. This problem of multiple views has been studied in the original data space (Caruana, Elha- wary, Nguyen, & Smith, 2006), orthogonal space (Davidson & Qi, 2008) and subspace projections (Agrawal, Gehrke, Gunopulos, & Raghavan, 1998). Other challenges include given knowledge about known clusterings, processing schemes for clustering, the number of multiple clusterings and flexibility. We take motivation from these research works and propose a new cluster initialization algorithm for categorical datasets that perform multiple clustering on different attributes (in the original data space) and uses distinct attribute values in an attribute as cluster labels. These multiple views provide new insights into the hidden structures of the data that serve as a cue to find consistent cluster structure and aid in computing better initial cluster centers. In the following subsections, we will present three approaches to select different attribute spaces that can help in generating differ- ent clustering views from the data. It is to be noted that all the pro- posed approaches assume that the desired number of clusters, K, is known in advance. 4.1. Vanilla approach A Vanilla approach is to consider all the attributes (m) present in the data and generate M clustering views that can be used for fur- ther analysis (see details for the follow up steps in Section 4.4). 4.2. Prominent attributes Khan and Ahmad (2012) present that only few attributes may be useful to generate multiple clustering views that can help in computing initial cluster centers for K-modes algorithm. These rel- evant attributes are extracted based on the following experimental observations: 1. There may be some attributes in the dataset whose number of attribute values are less than or equal to K. Due to fewer attri- bute values per cluster, these attributes possess higher discrim- inatory power and will play a significant role in deciding the initial cluster centers as well as the cluster structures. The set of these relevant attributes are called Prominent attributes (P). 2. For the other attributes in the dataset whose number of attri- bute values are greater than K, the numerous attribute values in these attributes will be spread out per cluster. These attri- butes add little to determine proper cluster structure and con- tribute less in deciding the initial representative modes of the clusters. Algorithm 1 shows the steps to compute Prominent attributes from a dataset. The number of attributes in the set P is defined as p = jPj. In the algorithm, p = 0 refers to a situation when there are no prominent attributes in the data and p = m means that all attributes are prominent attributes. In both of these scenarios, all the attributes in the data are considered prominent or else a re- duced set, P, of prominent attributes (equals to p) is chosen. Algorithm 1. Computation of Prominent attributes Input: N = data objects, M = Set of attributes in the data, m = jMj = Number of attributes in the data, p = 0 Output: P = Set of Prominent attributes P = / for i = 1 ? m do if Number of distinct attribute values in Mi > 1 && Mi 6 K then Add Mi to P increment p end if end for if p = 0 k p = m then use all attribute and call computeInitialModes(Attributes M) else use reduced prominent attributes and call computeInitialModes(Attributes P) end if 4.3. Significant attributes As discussed in the previous section, we select prominent attri- butes as we expect that these attributes play important role in clustering. The following section is taken from the work of Ahmad and Dey (2007a) that discusses an approach to rank important attributes in a dataset. We use their method to find significant attributes from the dataset. Ahmad and Dey (2007a, 2007b) propose an unsupervised learning method to compute the significance of attributes. On the basis of their significance, important attributes can be se- lected. In this method the most important step is to find out the distance between any two categorical values of an attri- bute. The distance between two distinct attribute values is computed as a function of their overall distribution and co- occurrence with other attributes. The distance between the pair of values x and y of attribute Mi with respect to the attribute Mj, for a particular subset w of attribute Mj values, is defined as follows: /wij ðx; yÞ¼ pðwjxÞþ pð� wjyÞ� 1 ð3Þ where p(wjx) denotes the probability that elements of the dataset with attribute Mi equal to x have attribute Mj value such that it is contained in w and, p(�wjy) denotes the probability that elements of the dataset with attribute Mi equal to y have attribute Mj value such that it is not contained in w. The distance between attribute values x and y for Mi with re- spect to attribute Mj is denoted by /j(x,y) and is given by /jðx; yÞ¼ pðWjxÞþ pð� WjyÞ� 1 ð4Þ where W is the subset of values of Mj that maximizes the quantity p(wjx) + p(�wjy). The distance between x and y is computed with respect to every other attribute. The average value of distances will be the distance /j(x,y) between x and y in the dataset. The average value of all the attribute values pair distances is taken as the signif- S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 7447 Author's personal copy icance of the attribute. Algorithm 2 shows the steps to compute the significance of attributes in the data. Algorithm 2. Computation of significance of attributes Input: D = Categorical Dataset, N = data objects, M = Set of Attributes in the data, m = jMj = Number of attributes in the data Output: S = Set of attributes sorted in order of their significance for every attribute Mi do for every pair of categorical attribute values ðx; yÞ do Sum = 0 for every other attributes Mj do /jðx; yÞ¼ maxðpðwjxÞþ pyi ðwjyÞ� 1 where w is subset of jth attribute values Sum = Sum + /j(x, y) end for Distance /(x, y) between categorical values ðx; yÞ¼ Sumðm�1Þ end for The average value of all the pair distances is taken as the significance of the attribute. end for We provide an example below to illustrate Algorithm 2. Con- sider a pure categorical dataset with three attributes M1, M2 and M3 as shown in Table 1. We compute the significance of attribute M1 by calculating the distance of each pair of attribute value with respect to every other attribute. In this case there is only one pair (L, T), therefore; The distance between L and T with respect to M2 is: maxðpðWjLÞþpð�WjTÞ�1Þ¼pðCjLÞþpð�CjTÞ�1¼1þ 2 3 �1¼ 2 3 where W is the subset of values of M2 Similarly, the distance between L and T with respect to M3 is: maxðpðWjLÞþpð�WjTÞ�1Þ¼pðEjLÞþpð�EjTÞ�1¼1þ 1 2 �1¼ 1 2 The average distance between L and T is: /ðL; TÞ¼ 1 2 2 3 þ 1 2 � � ¼ 0:58 As there is only one pair of values in the attribute M1, the signif- icance of attribute M1 (i.e. the average of distances of all pairs) = 0.58. This method to compute significance of attribute has been used in various K-means type clustering algorithms for mixed numeric and categorical datasets (Ahmad & Dey, 2007a, 2007b, 2011). Gen- erally the cost function of K-means type algorithm give equal importance to all the attributes. Ahmad and Dey (2007a, 2007b, 2011) show that with incorporating these significance of attributes in the cost function, better clustering results can be achieved. Ji, Pang, Zhou, Han, and Wang (2012) show that this approach is also useful for fuzzy clustering of categorical datasets. In this paper, we will use this approach to select the significant attributes from the datasets. 4.4. Computation of initial cluster centers In the preceding sections, we discussed three methods of choos- ing attributes that can be used for computation of initial cluster centers. The Vanilla approach chooses all the attributes whereas the prominent attribute approach (see Section 4.2) has the ability to choose fewer number of attributes depending upon the distribu- tion of attribute values in different attributes in the data. We will discuss the potential problem of choosing all the attributes in Sec- tion 5.3. The method to compute significant attributes (see Sec- tion 4.3) provides a ranking of all the attributes in order of their significance in the dataset. However, there is no straight-forward way to choose the most significant attributes in the data except to use a arbitrary cut-off value. For the experimentation, we choose the number of significant attributes to be the same as prominent attributes and discard rest of them, if all the attributes in the data- set are prominent then all of them are considered significant. The main idea of the proposed algorithm is to partition the data into clusters that corresponds to the number of distinct attribute values for Vanilla/Prominent/Significant attributes, and generate a cluster label for every data object present in the dataset. This clus- ter labeling is essentially a clustering view of the original data in the original space. Repeating this process for all the Vanilla/Promi- nent/Significant attributes yield a number of cluster labels that rep- resent multiple partition views of every data object. The cluster labels that are assigned to a data object over these multiple clus- terings is termed as cluster string and the number of total cluster strings is equal to the number of data objects present in the dataset. As noted in Section 4, some data objects will not be affected by choosing different initial cluster centers and their cluster strings will remain same. The distinct number of cluster strings represents the number of distinguishable clusters in the data. The algorithm assumes that the knowledge of the natural clusters in the data i.e. K is available and if the number of distinct cluster strings are more than K then it merges them into K clusters of cluster strings, such that the cluster strings within a cluster are more similar than others. Lastly, the cluster strings within each K clusters are re- placed by their corresponding data objects and modes of every K cluster is computed that serves as the initial cluster centers for the K-modes algorithm. In summary, the proposed method finds the dense localized regions in the dataset in the form of distin- guishable clusters. If their count is greater than K then it merges them to K clusters (and has the ability to ignore the infrequent clusters) and finds their group modes to be used as initial cluster centers. This process helps in avoiding the outliers contributing to the computation of initial cluster centers. Table 1 Categorical dataset. Attributes M1 M2 M3 L C E L C F T C F T K F T D F Table 2 Cluster strings of different data objects. Data point Cluster string D1 1-1-3-2 D2 2-2-1-1 D3 1-1-3-2 D4 2-2-1-1 D5 1-2-4-2 D6 2-1-4-1 D7 2-2-2-1 D8 1-1-3-2 D9 2-2-2-1 D10 1-2-3-1 7448 S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 Author's personal copy Algorithm 3. computeInitialModes(Attributes A) Input- Dataset N, n = jNj = the number of data objects, and A is the set of categorical attributes with a = jAj = number of attributes. If all the attributes are considered, then A = M and a = m, If prominent/significant attributes are considered, a 6 m i.e. a = p. jaij is the cardinality of the ai attribute and K is a user defined number that represent the number of clusters in the data. Output- K cluster centers Generation of cluster strings for i = 1. . .a do 1. Divide the dataset into jaij clusters on the basis of these jaij attribute values such that data objects with different val- ues (of this attribute ai) fall into different clusters. Com- pute cluster centers of these jaij clusters. 2. Partition the data by performing K-modes clustering that uses the cluster centers computed in above step as initial cluster centers. 3. Assign cluster label to every data object. Sti defines the cluster label of tth data object computed by using ai attri- bute, where t = 1,2. . .n. end for The cluster labels assigned to a data object is considered as a cluster string, resulting in the generation of n clustering strings. 4. Find distinct cluster strings from n strings, count their fre- quency, and sort them in descending order. Their count, K0, is the number of distinguishable clusters. 5. if K0 = K Get the data objects corresponding to these K clus- ter strings, and compute cluster centers of these K clusters. These will be the required initial cluster centers. 6. if K0 > K Merge similar distinct cluster string of K0 strings into K clusters (more details in Section 4.4.1) and compute the cluster centers. These cluster centers will be the required cluster centers. 7. if K0 < K Reduce the value of K and repeat the complete process. The steps to find initial cluster centers by using the proposed approach are presented in Algorithm 3. The computational effi- ciency of step 4 of the proposed algorithm can be improved by using other approaches such as on-line suffix trees (Ukkonen, 1995) that can perform string comparisons in time linear in the length of the string. To illustrate Algorithm 3, we present a descriptive example. Suppose we have 10 data objects D1, D2 , . . . , D10, defined by 4 cate- gorical attributes with K = 2. Let the cardinality of M1, M2, M3 and M4 are 2, 2, 4, 2. For the Vanilla approach, we consider all the attri- butes and first divide the data objects on the basis of attribute M1 and calculate 2 cluster centers because cardinality of M1 is 2. We run K-modes algorithm by using these initial cluster centers. Every data object is assigned a cluster label (either 1 or 2) and the same process is repeated to all other attributes. As there are 4 attributes, each data object will have a cluster string that consists of 4 labels. For example data object D1 has 1-2-2-1 as the cluster string. This means that in the first run (using M1 to create initial clusters) the data object D1 is placed in cluster 1, in the second run (using M2 to create initial clusters) the data object D1 is placed in cluster 2 and so on. We will get 10 different cluster strings corresponding to every data object. Suppose we get the following clustering strings for different data objects as shown in Table 2. We calculate the frequency of all the distinct strings as shown in Table 3. We take 100.5 � 3 most frequent cluster strings (details in Sec- tion 4.4.1 on this step) and cluster them by using hierarchical clus- tering with K = 2. The similar strings 2-2-1-1 and 2-2-2-1 are merged in one cluster. This leads to two clusters containing the cluster strings 1-1-3-2 and 2-2-1-1, 2-2-2-1 with their correspond- ing data objects, i.e. Cluster1 = {D1, D3, D8} Cluster2 = {D2, D4, D7, D9} The data objects belonging to these clusters are to be used to com- pute the required 2 cluster centers as K = 2. The other infrequent cluster strings and their corresponding data objects are assumed to be outliers that do not contribute in computing the initial cluster centers. The centers of these clusters serve as the initial cluster centers for the K-Modes algorithm. For prominent features approach, the attributes with attribute values less than or equal to the number of clusters are selected. In the example shown, attri- butes M1, M2 and M4 attributes are selected and the same proce- dure is followed with these three attributes to compute the initial cluster center. In the significant attributes approach, firstly the significant attributes are calculated, then they are used to cal- culate the initial cluster centers. For example if M1, M2 and M3 are the most significant attributes, then they are used to calculate the initial cluster centers following the above procedure. Algorithm 3 may give rise to an obscure case where the number of distinct cluster strings are less than the chosen K (assumed to represent the natural clusters in the data). This case can happen when the partitions created based on the attribute values of A attri- butes group the data in almost the same clusters every time. An- other possible scenario is when the attribute values of all attributes follow almost same distribution, which is normally not the case in real data. This case also suggests that probably the cho- sen K does not resemble with the natural grouping and it should be changed to a different value. The role of attributes with attribute values greater than K has to be investigated in this case. Generally, in K-modes clustering, the number of desired clusters (K) is se- lected without the knowledge of natural clusters in the data. The number of natural clusters may be less than the number of the de- sired clusters. If the number of the cluster strings (K0) obtained is less than K, a viable solution is to reduce the value of K and then apply the proposed algorithm to calculate the initial cluster cen- ters. However, this particular case is out of the scope of the present paper. 4.4.1. Merging clusters As discussed in step 6 of Algorithm 3, there may arise a case when K0 > K, which means that the number of distinguishable clus- ters obtained by the algorithm are more than the desired number of clusters in the data. Therefore, K0 clusters must be merged to arrive at K clusters. As these K0 clusters represent distinguishable clusters, a trivial approach could be to sort them in order of cluster string frequency and pick the top K cluster strings. A problem with this method is that it cannot be ensured that the top K most fre- quent cluster strings are representative of K clusters. If more than Table 3 Example of frequency computation of distinct cluster strings. String Frequency Data objects 1-1-3-2 3 D1,D3,D8 2-2-1-1 2 D2,D4 2-2-2-1 2 D7,D9 1-2-4-2 1 D5 2-1-4-1 1 D6 1-2-3-1 1 D10 S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 7449 Author's personal copy one cluster string comes from same cluster then the K-modes algo- rithm will give undesirable clustering results. This fact is also ver- ified experimentally and holds to be true. Keeping this issue in mind, we propose to use the hierarchical clustering method (Hall et al., 2009) to merge K0 distinct cluster strings into K clusters. The hierarchical clustering generates more informative cluster structures than the unstructured set of clusters returned by non-hierarchical clustering methods (Jain & Dubes, 1988). Most hierarchical clustering algorithms are deterministic and stable in comparison to their partitional counterparts. How- ever, hierarchical clustering has the disadvantage of having qua- dratic time complexity with respect to the number of data objects. In general, K0 cluster strings will be less than n. However, to avoid extreme case such as when K0 � n, we only choose the most frequent n0.5 distinct cluster strings. This will make the hier- archical algorithm log-linear with the number of data objects (K0 or n0.5 distinct cluster strings here). The Hamming distance (defined in Section 3) is used to compare the cluster strings. The proposed algorithm is based on the observation that some data objects al- ways belong to same clusters irrespective of the initial cluster cen- ters. The proposed algorithm attempts to capture those data objects that are represented by most frequent strings. The infre- quent cluster strings can be considered as outliers or boundary cases and their exclusion does not affect the computation of initial cluster centers. In the best case, when K0 � n0.5, the time complex- ity effect of log-linear hierarchical clustering will be minimal. This process generates K M-dimensional modes that are to be used as initial cluster centers for K-modes clustering algorithm. For merg- ing cluster strings (in Section 5), we use ‘single-linkage’ hierarchi- cal clustering, however other options such as average-linkage, complete-linkage, etc. can also be used. Continuing with the example shown in Section 4.4, we start with n0.5 strings as the lowest level of the tree for the bottom up approach of hierarchical clustering. Similar strings are merged up to the level where the number of clusters is equal to the number of desired cluster, K. The data objects belonging to strings in a clus- ter are used to compute initial cluster center. In the example shown in the previous section, there are three strings, 1-1-3-2, 2- 2-1-1 and 2-2-2-1, that are to be used for computing initial cluster centers. The number of designated clusters is 2. The similar strings 2-2-1-1 and 2-2-2-1 are merged, resulting in two clusters. All the data objects corresponding to these strings within a cluster is used to compute the cluster centers. 4.4.2. Choice of attributes The proposed algorithm starts with the assumption that there exists prominent attributes in the data that can help in obtaining distinguishable cluster structures that can either be used as is or be merged to obtain initial cluster centers. In the absence of any prominent attributes (or if all attributes are prominent), the Vanilla approach, all the attributes are selected to find initial cluster cen- ters. Since attributes other than prominent attributes contain attri- bute values more than K, a possible repercussion is the increased number of distinct cluster strings due to the availability of more cluster allotment labels. This implies an overall reduction in the individual count of distinct cluster strings and many small clusters may be generated. In our formulation, the hierarchical clusterer imposes a limit of n0.5 on the top cluster strings to be merged, therefore some relevant clusters could lay outside the bound dur- ing merging. This may lead to some loss of information while com- puting the initial cluster centers. The best case occurs when the number of distinct cluster strings are less than or equal to n0.5. 4.4.3. Evaluating time complexity The proposed algorithm to compute initial cluster centers has three parts, 1. Compute Vanilla/Prominent/Significant attributes 2. Compute initial cluster centers 3. If needed, merge clusters The time complexity of computation of Vanilla/Prominent attri- butes is O(nm), whereas for computing significant attributes is O(nm2T2) (Ahmad & Dey, 2007a), where T is the average number of distinct attribute values per attribute and T � n. Computation of initial cluster centers (from Algorithm 3) needs the basic K- modes algorithm to run P times (in the worst-case m times). As the K-modes algorithm is linear with respect to the size of the dataset (Huang, 1997), the worst-case time complexity will be O(rKm2n), where r is the number of iterations needed for conver- gence and r � n. For merging the distinct cluster strings into K clusters, computeInitialModes (Attributes A) uses hierarchical clus- tering. The worst-case complexity of the hierarchical clustering is O(n2 log n), however the proposed approach chooses only n0.5 most frequent distinct cluster string (see Section 4.4.1), therefore the worst-case complexity for merging cluster strings become O(n log n). Combining all the parts together, we get two worst-case time complexities: � Using All/Prominent attributes – O(nm + rKm2n + n log n) � Using Significant attributes – O(nm2T2 + rKm2n + n log n) It is to be noted that the worst-case time complexity using signif- icant attributes is higher than using all/prominent attributes due to the additional computation time spent in finding out the signifi- cance of attributes. However, the worst-case time complexity of using both the methods is log-linear in the number of data objects. With prominent attributes approach, the proposed method is advantageous for the datasets when n � rKm2. For significant attri- butes approach, the method may be useful for those datasets when n � m2T2 + rKm2. 5. Experimental analysis 5.1. Datasets To evaluate the performance of the proposed initialization method, we use several pure categorical datasets from the UCI Ma- chine Learning Repository (Batche & Lichman, 2013). A short description for each dataset is given below. Soybean Small. This dataset consists of 47 cases of soybean dis- ease each characterized by 35 multi-valued categorical variables. These cases are drawn from four populations, each one of them representing one of the following soybean diseases: D1-Diaporthe stem canker, D2-Charcoat rot, D3-Rhizoctonia root rot and D4-Phy- tophthorat rot. Ideally, a clustering algorithm should partition these given cases into four groups (clusters) corresponding to the diseases. The clustering results on Soybean Small data are shown in Table 5. Breast Cancer Data. This data has 699 instances with 9 attri- butes. Each data object is labeled as benign (458% or 65.5%) or malignant (241% or 34.5%). There are 9 instances in attribute 6 and 9 that contain a missing (i.e. unavailable) attribute value. The clustering results of breast cancer data are shown in Table 6. Zoo Data. It has 101 instances described by 16 attributes and distributed into 7 categories. The first attribute contains a unique animal name for each instance and is removed because it is non- informative. All other characteristics attributes are Boolean except for the character attribute corresponds to the number of legs that lies in the set 0, 2, 4, 5, 6, 8. The clustering results of Zoo data are shown in Table 7. 7450 S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 Author's personal copy Lung Cancer Data. This dataset contains 32 instances described by 56 attributes distributed over 3 classes with missing values in attributes 5 and 39. The clustering results for lung cancer data are shown in Table 8. Mushroom Data. Mushroom dataset consists of 8124 data ob- jects described by 22 categorical attributes distributed over 2 clas- ses. The two classes are edible (4208 objects) and poisonous (3916 objects). It has missing values in attribute 11. The clustering results for mushroom data are shown in Table 9. Congressional Vote Data. This data set includes votes for each of the U.S. House of Representatives Congressmen on the 16 key votes. Each of the votes can either be a yes, no or an unknown dis- position. The data has 2 classes with 267 democrats and 168 republicans instances. The clustering results for Vote data are shown in Table 10. Dermatology Data. This dataset contains six types of skin dis- eases for 366 patients that are evaluated using 34 clinical attri- butes, 33 of them are categorical and one is numerical. The categorical attribute values signify degrees in terms of whether the feature is present, contain largest possible amount or relative intermediate values. In our experiment, we discretize the numeri- cal attribute (representing the age of the patient) to contain 10 cat- egories. The clustering results for Dermatology data are presented in Table 11. We used the WEKA framework (Hall et al., 2009) for the data pre-processing and implementing the proposed algorithm.1 5.2. Comparison and performance evaluation metric To evaluate the quality of clustering results and their fair com- parison, we used the performance metrics used by Wu et al. (2007) that are derived from information retrieval. Assuming that a data- set contains K classes, for any given clustering method, let ei be the number of data objects that are correctly assigned to class Ci, let bi be the number of data objects that are incorrectly assigned to class Ci, and let ci be the data objects that are incorrectly rejected from class Ci, then precision, recall and accuracy are defined as follows: PR ¼ PK i¼1 ei eiþbi � � K ð5Þ RE ¼ PK i¼1 ei eiþci � � K ð6Þ AC ¼ PK i¼1ei N ð7Þ Jain and Dubes (1988) noted that the results of partitional clus- tering algorithms improve when the initial cluster centers are close to the actual cluster centers. To measure the closeness between initial cluster centers computed by the proposed method and the actual modes of the clusters in the data, we define a match metric, matchMetric ¼ 1 K � m XK i¼1 Xn j¼1 dðinitialij; actualijÞ ð8Þ where initialij is the jth value of the initial mode for the ith cluster, actualij is the corresponding jth value of the actual mode for the ith cluster and d is defined same as in Section 3. The matchMetric will give degree of closeness between initial and actual modes with a value of 0 means no match and 1 means exact match between them. 5.3. Effect of number of attributes In Section 4.4.2, we discussed that choosing all the attributes can lead to the generation of large number of cluster strings, spe- cially if the attributes have many attribute values. To test this intu- ition, we performed a comparative analysis of the effect of the number of selected attributes on the number of distinct cluster strings (generated in Step 6 of Algorithm 3). In Table 4, m is the to- tal number of attributes in the data, p is the number of prominent attributes, s is the number of significant attributes (s = jSj and s = p), CSM, CSP and CSS are the number of distinct cluster string obtained using Vanilla, Prominent and Significant attributes, and n0.5 is the limit on the number of top cluster strings to be merged using hier- archical clustering. The table shows that choosing a Vanilla ap- proach (all attributes) leads to larger number of cluster strings, whereas with the proposed approach (either prominent or signifi- cant attributes) they are relatively smaller. This fact can be seen for the Soybean Small, Mushroom and Dermatology data. For Lung Cancer data p � m therefore the number of cluster strings are equivalent. For Soybean Small and Mushroom datasets, for same number of Prominent and Significant attributes, the corresponding number of cluster strings are different (CSP – CSS). This is due to the fact that the set of p prominent and significant attributes is dif- ferent because both methods select attributes by using different approaches. While the choice of prominent attributes (when they are less than m) should reduce the overall cluster strings as it se- lects attributes with fewer attribute values, the attributes selected by significance method may contain attributes with more attribute values that results in more cluster strings. For Zoo, Vote and Breast Cancer datasets, all the attributes were prominent therefore p and m are same and hence CSP = CSA. It is to be noted that the number of distinct cluster strings using proposed approach for Zoo and Mush- room datasets are within the bounds of n0.5 limit. 5.4. Clustering results In this section, we present the K-Modes clustering results that use the initial cluster centers computed with the proposed method. We conducted five set of experiments, their details are as follows: � Experiment1: We compared the clustering results obtained by using prominent attributes to find initial cluster centers with the method of random selection of initial cluster centers and the methods described by Cao et al. (2009) and Wu et al. (2007). As mentioned in Section 2, there are some computa- tional inaccuracies in the work of Bai et al. (2012), therefore we exclude their method from comparison with our work. For random initialization, we randomly group data objects into K clusters and compute their modes to be used as initial cluster centers. � Experiment2: We compared the clustering results obtained by using prominent and significant attributes to find initial cluster centers. Table 4 Effect of choosing different number of attributes. Dataset Vanilla Prominent Significant n0.5 m CSA p CSP s CSS Soybean 35 25 20 21 20 23 7 Mushroom 22 683 5 16 5 44 91 Dermatology 34 357 33 352 33 357 20 Lung-Cancer 56 32 54 32 54 32 6 Zoo 16 7 16 7 16 7 11 Vote 16 126 16 126 16 126 21 Breast-Cancer 9 355 9 355 9 355 27 1 The Java source code is publicly available at http://www.cs.uwaterloo.ca/ �s255khan/code/kmodes-init.zip. S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 7451 Author's personal copy � Experiment3: For some datasets, the Vanilla attributes are differ- ent from prominent attributes, for those cases we compared their clustering results. � Experiment4: For all the three approaches to compute initial cluster centers i.e. Vanilla, Prominent and Significant, we com- puted the matchMetric to measure the quality of initial cluster centers in terms of their closeness to the actual modes or cluster centers of the data. � Experiment5: We performed a scalability test by increasing the number of data objects � 100,000 and recording the time spent in computing initial cluster centers. We also compared the order of time complexities of the proposed method with two other initialization methods. Experiment1. Tables 5–11 show the clustering results, with con- fusion matrix representing the cluster structures obtained by seed- ing K-modes algorithm with the initial cluster centers computed using the proposed method. It can be seen that the proposed initialization method outperforms random cluster initialization when used as seed for K-modes clustering algorithm for the cate- gorical data in accuracy, precision and recall. The random initiali- zation method gives non-repeatable results, whereas the proposed method gives fixed clustering results. Therefore, repeat- able and better cluster structures can be obtained by using the pro- posed method. In comparison to the initialization methods of Cao et al. and Wu et al., we evaluate our results in terms of: � Accuracy – The proposed method outperforms or equals other methods in 4 cases and perform worse in one case. � Precision – The proposed method performs well or equals other methods in 2 cases and performs worse in 3 cases. � Recall – The proposed method outperforms or equals other methods in 4 cases and performs worse in 1 case. The results for Congressional Vote and Dermatology data are not available from the papers from Cao et al. and Wu et al., there- fore we compared the clustering accuracy of the proposed method against the random initialization method. The clustering results for Table 5 Clustering results for Soybean Small data. Cluster Class D1 D2 D3 D4 (a) Confusion matrix D1 10 0 0 0 D2 0 10 0 0 D3 0 0 10 2 D4 0 0 0 15 Random Wu Cao Proposed (b) Performance comparison AC 0.8644 1 1 0.9574 PR 0.8999 1 1 0.9583 RE 0.8342 1 1 0.9705 Table 6 Clustering results for Breast Cancer data. Cluster Class Benign Malignant (a) Confusion matrix Benign 453 56 Malignant 5 185 Random Wu Cao Proposed (b) Performance comparison AC 0.8364 0.9113 0.9113 0.9127 PR 0.8699 0.9292 0.9292 0.9318 RE 0.7743 0.8773 0.8773 0.8783 Table 7 Clustering results for Zoo data. Cluster Class a b c d e f g (a) Confusion matrix a 39 0 0 0 0 0 0 b 0 19 0 0 0 0 0 c 0 1 4 0 4 0 0 d 2 0 1 13 0 0 0 e 0 0 0 0 0 0 1 f 0 0 0 0 0 8 2 g 0 0 0 0 0 0 7 Random Wu Cao Proposed (b) Performance comparison AC 0.8356 0.8812 0.8812 0.8911 PR 0.8072 0.8702 0.8702 0.7224 RE 0.6012 0.6714 0.6714 0.7716 Table 8 Clustering results for Lung Cancer data. Cluster Class a b c (a) Confusion matrix a 8 7 0 b 1 6 8 c 0 0 2 Random Wu Cao Proposed (b) Performance comparison AC 0.5210 0.5000 0.5000 0.5000 PR 0.5766 0.5584 0.5584 0.6444 RE 0.5123 0.5014 0.5014 0.5168 Table 9 Clustering results for Mushroom data. Cluster Class Poisonous Edible (a) Confusion matrix Poisonous 3052 98 Edible 864 4110 Random Wu Cao Proposed (b) Performance comparison AC 0.7231 0.8754 0.8754 0.8815 PR 0.7614 0.9019 0.9019 0.8975 RE 0.7174 0.8709 0.8709 0.8780 Table 10 Clustering results for Congressional Vote data. Cluster Class Republican Democrat (a) Confusion matrix Republican 158 55 Democrat 10 212 Random Proposed (b) Performance comparison AC 0.4972 0.8506 PR 0.5030 0.8484 RE 0.5031 0.8672 7452 S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 Author's personal copy Dermatology data with random initialization are worse due to mixing up of data objects among various clusters. The above results are very encouraging due to the fact that the proposed method attempts to find dense localized regions and dis- cards the boundary cases, thus ensuring the selection of better ini- tial cluster centers with log-linear worst-case time complexity. The method of Wu et al. induces random selection of data objects and Cao et al. can select boundary cases as initial cluster centers which can be detrimental to the clustering results. The accuracy values of proposed method are better than or equal to other methods. The only case where the proposed method perform worse in all three performance metric is the Soybean Small dataset. This dataset has only 47 data objects, our algorithm could not cluster only 2 data objects correctly. However, due to the small size of the data- set, the clustering error appears to be large. We observe that on some datasets the proposed method gives worse values for precision, which implies that in those cases some data objects from non-classes are getting clustered in given classes. The recall values of proposed method are better than the other methods, which suggests that the proposed approach tightly con- trols the data objects from given classes to be not clustered to non-classes. Breast Cancer data has no prominent attribute in the data and uses all the attributes and produces comparable results to other methods. Lung Cancer data, though smaller in size has high dimension and the proposed method is able to produce better precision and recall rates than other methods. It is also observed that the proposed method performs well on large dataset such as Mushroom data with more than 8000 data objects. In our experi- ments we did not encounter a scenario where the distinct cluster strings are less than the desired number of clusters (step 7 of Algo- rithm 3). Experiment2. Table 12 shows that for Zoo, Vote and Breast Can- cer datasets, the clustering results using prominent and significant attributes are same. This is because all the attributes are consid- ered in these datasets for computing the initial cluster centers. For other datasets, we observe that using prominent attributes is a better choice than significant attributes. Although we choose the same number of prominent and significant attributes (specially when all the attributes are not prominent), their clustering results varies because both of the attribute spaces may contain different set of attributes. The reason is that by definition (see Section 4), prominent and significant attributes use different criteria to choose relevant attributes for computing initial cluster centers. Moreover, generating the ranking of significant attributes is costlier in terms of time complexity than computing prominent attributes (see Section 4.4.3 for details). Experiment3. As per Algorithm 1, for Zoo, Vote and Breast Cancer data all the attributes are prominent. For the rest of the datasets, this is not the case and the prominent attributes are less than the total number of attributes. We performed an experiment to analyze the scenario when there are fewer prominent attributes and its impact on overall clustering results. Table 13 shows that for all the datasets except Soybean Small, choosing prominent attributes less than the total number of attributes improve the clustering performance. Choosing all attributes in comparison to Table 11 Clustering results for Dermatology data. Cluster Class Seboreic dermatitis Psoriasis Lichen planus Cronic dermatitis Pityriasis rosea Pityriasis rubra pilaris (a) Confusion matrix Seboreic dermatitis 53 7 5 2 36 0 Psoriasis 0 96 0 0 0 0 Lichen planus 0 0 66 0 0 0 Cronic dermatitis 2 1 0 39 0 0 Pityriasis rosea 6 8 1 11 13 4 Pityriasis rubra pilaris 0 0 0 0 0 16 Random Proposed (b) Performance comparison AC 0.2523 0.7732 PR 0.2697 0.7909 RE 0.2954 0.7570 Table 12 Comparison of clustering results using Prominent and Significant attributes. Dataset Prominent Significant AC PR RE AC PR RE Soybean 0.9574 0.9583 0.9705 0.6809 0.7549 0.7176 Mushroom 0.8815 0.8975 0.8780 0.5086 0.7303 0.5256 Dermatology 0.7732 0.7909 0.7570 0.6502 0.5601 0.5512 Lung-Cancer 0.5000 0.6444 0.5168 0.46875 0.5079 0.4838 Zoo 0.8911 0.7224 0.7716 0.8911 0.7224 0.7716 Vote 0.8506 0.8484 0.8672 0.8506 0.8484 0.8672 Breast-Cancer 0.9127 0.9318 0.8783 0.9127 0.9318 0.8783 Table 13 Comparison of clustering results using Vanilla and Prominent attributes. Dataset Vanilla Prominent AC PR RE AC PR RE Soybean 0.9787 0.9772 0.9853 0.9574 0.9583 0.9705 Mushroom 0.6745 0.7970 0.6627 0.8816 0.8976 0.87803 Dermatology 0.4180 0.3889 0.3420 0.7372 0.7909 0.7570 Lung-Cancer 0.5000 6317 0.5017 0.5 0.6444 0.5168 Table 14 Comparison of matchMetric and its effect on K-modes algorithm convergence. Dataset Vanilla Prominent Significant matchMetric #Itr matchMetric #Itr matchMetric #Itr (a) p < m Soybean 0.7357 2 0.9643 2 0.8428 2 Mushroom 0.6136 2 0.8863 2 0.5681 1 Dermatology 0.6176 5 0.6813 6 0.6274 6 Lug-Cancer 0.6726 6 0.7261 5 0.7321 8 Dataset Vanilla matchMetric #Itr (b) p = m Zoo 0.8661 2 Vote 0.7500 2 Breast-Cancer 0.8333 3 S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 7453 Author's personal copy fewer prominent attributes generates more cluster strings (see Table 4). If these cluster strings are more than n0.5, then many relevant cluster strings may be not be chosen, which if included could have contributed in computation of initial cluster centers. Experiment4. For all the datasets using the three methods of computing initial cluster centers, we computed the matchMetric (see Eq. (8)), which measures the degree of closeness of initial and actual modes. We also studied the impact of quality of initial cluster centers on the convergence of the K-modes algorithm (in terms of number of iterations, #Itr). Table 14(a) shows the case when prominent attributes are less than total number of attributes. The initial cluster centers selected by prominent/significant attri- butes are always closer to the actual modes of the datasets in terms of the matchMetric and therefore the K-modes algorithm converges in very few iterations with good cluster structures (see discussion of clustering results in Experiment1). Similar results were obtained when all the attributes are chosen as prominent attributes and used to compute initial cluster centers (see Table 14(b)). The high values of matchMetric show that the initial cluster centers are close to the actual cluster centers and the K-modes clustering algorithm with these initial cluster centers converges fast with good cluster- ing performance. The reason for the initial cluster centers to be close to the actual cluster centers is that the proposed method finds dense localized clusters, merges them if needed and discards the insignificant clusters. Experiment5. Time Scalability of the Proposed Algorithm. We per- formed an experiment to test the scalability of the proposed meth- od for computing initial cluster centers for large datasets. We used the Mushroom dataset (see Section 5.1) that contains 8124 data objects described by 22 categorical attributes and 2 clusters. We made copies of this dataset in multiples of 2, 4, 6, 8, 10 and 12 such that the data size varies from 8124 to 113,736. We execute the pro- posed algorithm for computing initial cluster centers on each of these copies separately. We ran the experiment on a HP Touch- Smart tm2 machine with Intel Pentium™ U4100 1.3 GHz proces- sor, L2 cache 2048 KB and 4 GB RAM. Fig. 1 shows the plot between the different sizes of the data and the corresponding time consumed in computing the initial cluster centers. It can be ob- served that the time cost of the proposed method grows almost lin- early with the increase in the number of data objects. The experimental results suggest that the proposed cluster center ini- tialization method scales linearly and can be implemented for large datasets. Table 15 compares the time complexities of the proposed clus- ter initialization algorithms with the two competing initialization methods of Cao et al. (2009) and Wu et al. (2007). In the proposed algorithm (with prominent attributes), if rKm2 is larger than logn (which is more likely to be true for high dimensional dataset with large number of clusters), the complexity is decided by the second term, rKm2n, which is linear in number of data objects and similar to Cao’s method and better than Wu’s method (with respect to the number of data objects). This linear time complexity behavior is also observed in our scalability experiments. Multiple Attributes Clustering Challenges In Section 4, we men- tioned some of the challenges in employing multiple clustering ap- proaches (as defined by Müller et al. (2010)). The proposed method uses the multiple clustering approach to find initial cluster centers for a partitional clustering algorithm. The proposed approach suc- cessfully generated and detected the multiple clustering views from the data and can process different distinguishable clusters into relevant number of clusters by a modified hierarchical cluster- ing approach or use them unaltered, whichever the case may be (as 0 2 4 6 8 10 12 x 10 4 5 10 15 20 25 30 35 40 45 50 Number of data objects Ti m e (s ec on ds ) Fig. 1. Time consumption in computing initial cluster center for variable data size. Table 15 Comparison of time complexities. Initialization method Order of complexity Cao et al. O(nmK2) Wu et al. O(cn), where c can be between 2 to n0.5 Proposed method O(nm + rKm2n + n log n), for all/prominent attributes 7454 S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 Author's personal copy discussed in Algorithm 3). In the worst-case, the proposed algo- rithm will generate clustering views equal to the number of total attributes in the data. This is a significant improvement over other approaches such as by Khan and Kant (2007), which can run arbi- trary number of times for evidence accumulation. The proposed method is flexible and tested on various categorical datasets, how- ever a known limitation is the advance knowledge of number of natural clusters in the data. 6. Conclusions K-modes clustering algorithm is employed to partition the cat- egorical data into pre-defined K clusters, however the clustering results intrinsically depend on the choice of random initial cluster centers, that can cause non-repeatable results and produce impro- per cluster structures. In this paper, we propose an algorithm to compute the initial cluster centers for the categorical data by per- forming multiple clustering of data based on the attribute values present in different attributes. The present algorithm is based on the experimental fact that similar data objects form the core of the clusters and are not affected by the selection of initial cluster centers, and that individual attribute also provides useful informa- tion in generating cluster structures. The proposed algorithm is composed of two parts – relevant attributes selection and comput- ing initial cluster centers. For choosing the relevant attributes from the data, we presented two competitive methods. The first method chooses Prominent attributes on the basis of the attribute values present in an attribute and the second method computes the rank- ing of Significant attributes by an unsupervised learning method. Based on the selected attributes, the proposed algorithm partitions the data multiple times to generate multiple clustering views of the data. The multiplicity of clustering views is captured in the form of cluster strings, which produces distinct distinguishable clusters in the data that may be greater than, equal to or less than the desired number of clusters (K). If it is greater than K, then a modified hierarchical clustering is used to merge similar cluster strings into K clusters, if it is equal to K then the data objects cor- responding to cluster strings can be directly used as initial cluster centers. An obscure possibility may arise when the cluster strings are less than K, in this case, it is assumed that the current value of K is not the true representative of the desired number of clus- ters. In our experiments we did not get such situation, largely be- cause it can happen in a rare occurrence, when all the attribute values of different attributes cluster the data in the same way. These initial cluster centers when used as seed to K-modes cluster- ing algorithm, improves the accuracy of the traditional K-modes clustering algorithm that uses random cluster centers as starting point. Since the proposed method provides a definitive choice of initial cluster centers (zero standard deviation), consistent and repetitive clustering results can be obtained. We also show that the initial cluster centers computed by using prominent attributes performs better than significant attribute selection approach and has the advantage of lower computational complexity. The initial cluster centers computed by the proposed approach are found to be very similar to the actual cluster centers of the data that leads to faster convergence of K-modes clustering algorithm and better clustering results. The performance of the proposed method is bet- ter than random initialization and better than or equal to the other two methods compared on all datasets except one case. The biggest advantage of the proposed method is the worst-case log-linear time complexity of computation and fixed choice of initial cluster centers from dense localized regions, whereas the other two meth- ods lack one of them. When the number of desired clusters is not available in ad- vance, we would like to extend the proposed multi-clustering ap- proach for the categorical data for finding out the natural number of clusters present in the data, in addition to computing the initial cluster centers for such cases. The present algorithm to compute Prominent attributes sometimes select all the attributes in the data, however our experiments indicate that considering fewer most relevant attributes is a better choice than choosing all attributes. We would like to further investigate such cases in fu- ture. We would like to extend the Significant attributes approach by ranking them according to their significance in the final consensus building instead of taking a fixed number of attributes. In other words, while computing the similarity of cluster strings in the merging algorithm, more weights will be given to the clustering re- sults computed by using more significant attributes. References Agrawal, R., Gehrke, J., Gunopulos, D., & Raghavan, P. (1998). Automatic subspace clustering of high dimensional data for data mining applications. In L. M. Haas & A. Tiwary (Eds.), SIGMOD conference (pp. 94–105). ACM Press. Ahmad, A., & Dey, L. (2007a). A k-mean clustering algorithm for mixed numeric and categorical data. Data Knowledge Engineering, 63. Ahmad, A., & Dey, L. (2007b). A method to compute distance between two categorical values of same attribute in unsupervised learning for categorical data set. Pattern Recognition Letters, 28, 110–118. Ahmad, A., & Dey, L. (2011). A k-means type clustering algorithm for subspace clustering of mixed numeric and categorical datasets. Pattern Recognition Letters, 32, 1062–1069. Anderberg, M. R. (1973). Cluster analysis for applications. New York: Academic Press. Bai, L., Liang, J., Dang, C., & Cao, F. (2012). A cluster centers initialization method for clustering categorical data. Expert Systems with Applications, 39, 8022–8029. Boley, D., Gini, M., Gross, R., Han, E.-H., Karypis, G., Kumar, V., et al. (1999). Partitioning-based clustering for web document categorization. Decision Support Systems, 27, 329–341. Bradley, P. S., & Fayyad, U. M. (1998). Refining initial points for k-means clustering. In J. W. Shavlik (Ed.), ICML (pp. 91–99). Morgan Kaufman. Cao, F., Liang, J., & Bai, L. (2009). A new initialization method for categorical data clustering. Expert Systems and Applications, 36, 10223–10228. Caruana, R., Elhawary, M. F., Nguyen, N., & Smith, C. (2006). Meta clustering. In ICDM (pp. 107–118). IEEE Computer Society. Davidson, I., & Qi, Z. (2008). Finding alternative clusterings using constraints. In ICDM (pp. 773–778). IEEE Computer Society. Bache, K., & Lichman, M., (2013). UCI machine learning repository, http:// archive.ics.uci.edu/ml. Fred, A. L. N., & Jain, A. K. (2002). Data clustering using evidence accumulation. In ICPR (4) (pp. 276–280). Gowda, K. C., & Diday, E. (1991). Symbolic clustering using a new dissimilarity measure. Pattern Recognition, 24, 567–578. Guha, S., Rastogi, R., & Shim, K. (1999). Rock: a robust clustering algorithm for categorical attributes. In Proceedings of the 15th international conference on data engineering, 23–26 March 1999, Sydney, Austrialia (pp. 512–521). IEEE Computer Society. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., & Witten, I. H. (2009). The weka data mining software: an update. In SIGKDD Explorations: Vol. 11 of 1. He, Z. (2006). Farthest-point heuristic based initialization methods for k-modes clustering. CoRR, abs/cs/0610043. He, Z., Xu, X., & Deng, S. (2005). A cluster ensemble method for clustering categorical data. Information Fusion, 6, 143–151. Huang, Z. (1997). A fast clustering algorithm to cluster very large categorical data sets in data mining. In Research issues on data mining and knowledge discovery. Huang, Z. (1998). Extensions to the k-means algorithm for clustering large data sets with categorical values. Data Mining and Knowledge Discovery, 2, 283–304. Jain, A. K., & Dubes, R. C. (1988). Algorithms for clustering data. Upper Saddle River, NJ, USA: Prentice-Hall, Inc.. Ji, J., Pang, W., Zhou, C., Han, X., & Wang, Z. (2012). A fuzzy k-prototype clustering algorithm for mixed numeric and categorical data. Knowledge-Based Systems, 30, 129–135. Kaufman, L., & Rousseeuw, P. J. (1990). Finding groups in data: an introduction to cluster analysis. John Wiley. Khan, S. S., & Ahmad, A. (2004). Cluster center initialization algorithm for k-means clustering. Pattern Recognition Letters, 25, 1293–1302. Khan, S. S., & Ahmad, A. (2003). Computing initial points using density based multiscale data condensation for clustering categorical data. In Proceedings of 2nd international conference on applied artificial intelligence. Khan, S. S., & Ahmad, A. (2012). Cluster center initialization for categorical data using multiple attribute clustering. In E. Mülle, T. Seidl, S. Venkatasubramanian, & A. Zimek (Vol. Eds.), Workshop proceedings of the 3rd multiclust workshop: discovering, summarizing and using multiple clusterings, USA (pp. 3–10). Khan, S. S., & Kant, S. (2007). Computation of initial modes for k-modes clustering algorithm using evidence accumulation. In Proceedings of the 20th international joint conference on artificial intelligence (IJCAI) (pp. 2784–2789). S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 7455 Author's personal copy Matas, J., & Kittler, J. (1995). Spatial and feature space clustering: applications in image analysis. In CAIP (pp. 162–173). Mitra, P., Murthy, C. A., & Pal, S. K. (2002). Density-based multiscale data condensation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24, 734–747. Müller, E., Günnemann, S., Färber, I., & Seidl, T. (2010). Discovering multiple clustering solutions: grouping objects in different views of the data. In G. I. Webb, B. Liu, C. Zhang, D. Gunopulos, & X. Wu (Eds.), ICDM (pp. 1220). IEEE Computer Society. Petrakis, E. G. M., & Faloutsos, C. (1997). Similarity searching in medical image databases. IEEE Transactions on Knowledge Data Engineering, 9, 435–447. Ralambondrainy, H. (1995). A conceptual version of the k-means algorithm. Pattern Recognition Letters, 16, 1147–1157. Sun, Y., Zhu, Q., & Chen, Z. (2002). An iterative initial-points refinement algorithm for categorical data clustering. Pattern Recognition Letters, 23, 875–884. Ukkonen, E. (1995). On-line construction of suffix trees. Algorithmica, 14, 249–260. Wu, S., Jiang, Q., & Huang, J. Z. (2007). A new initialization method for clustering categorical data. In Proceedings of the 11th Pacific-Asia conference on advances in knowledge discovery and data mining PAKDD’07 (pp. 972–980). Berlin, Heidelberg: Springer-Verlag. 7456 S.S. Khan, A. Ahmad / Expert Systems with Applications 40 (2013) 7444–7456 
work_sfdi3u5lxrfd5bs6kvpimltvbu	----	Path relinking for the vertex separator problem Path relinking for the vertex separator problem Submitted by Jin-Kao Hao on Sun, 06/04/2017 - 09:49 Titre Path relinking for the vertex separator problem Type de publication Article de revue Auteur Ma, Fuda [1], Wang, Yang [2], Hao, Jin-Kao [3] Editeur Elsevier Type Article scientifique dans une revue à comité de lecture Année 2017 Langue Anglais Date 1er Oct. 2017 Pagination 332–343 Volume 82 Titre de la revue Expert Systems with Applications ISSN 09574174 Mots-clés graph partitioning [4], Path relinking [5], Population-based heuristics [6], Vertexseparator [7] Résumé en anglais This paper presents the first population-based path relinking algorithm for solving the NP-hard vertex separator problem in graphs. The proposed algorithm employs a dedicated relinking procedure to generate intermediate solutions between an initiating solution and a guiding solution taken from a reference set of elite solutions (population) and uses a fast tabu search procedure to improve some selected intermediate solutions. Special care is taken to ensure the diversity of the reference set. Dedicated data structures based on bucket sorting are employed to ensure a high computational efficiency. The proposed algorithm is assessed on four sets of 365 benchmark instances with up to 20,000 vertices, and shows highly comparative results compared to the state-of-the-art methods in the literature. Specifically, we report improved best solutions (new upper bounds) for 67 instances which can serve as reference values for assessment of other algorithms for the problem. URL de la notice http://okina.univ-angers.fr/publications/ua15972 [8] DOI 10.1016/j.eswa.2017.03.064 [9] Lien vers le document http://www.sciencedirect.com/science/article/pii/S0957417417302270 [10] Titre abrégé Expert Systems with Applications Liens [1] http://okina.univ-angers.fr/publications?f%5Bauthor%5D=25615 [2] http://okina.univ-angers.fr/publications?f%5Bauthor%5D=26865 [3] http://okina.univ-angers.fr/jinkao.hao/publications [4] http://okina.univ-angers.fr/publications?f%5Bkeyword%5D=8649 [5] http://okina.univ-angers.fr/publications?f%5Bkeyword%5D=8833 http://okina.univ-angers.fr/publications?f%5Bauthor%5D=25615 http://okina.univ-angers.fr/publications?f%5Bauthor%5D=26865 http://okina.univ-angers.fr/jinkao.hao/publications http://okina.univ-angers.fr/publications?f%5Bkeyword%5D=8649 http://okina.univ-angers.fr/publications?f%5Bkeyword%5D=8833 http://okina.univ-angers.fr/publications?f%5Bkeyword%5D=23000 http://okina.univ-angers.fr/publications?f%5Bkeyword%5D=22999 http://okina.univ-angers.fr/publications?f%5Bkeyword%5D=22999 http://okina.univ-angers.fr/publications/ua15972 http://dx.doi.org/10.1016/j.eswa.2017.03.064 http://www.sciencedirect.com/science/article/pii/S0957417417302270 [6] http://okina.univ-angers.fr/publications?f%5Bkeyword%5D=23000 [7] http://okina.univ-angers.fr/publications?f%5Bkeyword%5D=22999 [8] http://okina.univ-angers.fr/publications/ua15972 [9] http://dx.doi.org/10.1016/j.eswa.2017.03.064 [10] http://www.sciencedirect.com/science/article/pii/S0957417417302270 Publié sur Okina (http://okina.univ-angers.fr) http://okina.univ-angers.fr 
work_sgdejolmezfsfeseqbroy6peum	----	tai100bGO.eps Memetic search for the quadratic assignment problem Una Benlica, Jin-Kao Haob,∗ aCHORDS, University of Stirling, Stirling FK9 4LA, United Kingdom bLERIA, Université d’Angers, 2 bd Lavoisier, 49045 Angers, Cedex 01, France Accepted to Expert Systems with Applications, 12 August 2014 Abstract The quadratic assignment problem (QAP) is one of the most studied NP- hard problems with various practical applications. In this work, we propose a powerful population-based memetic algorithm (called BMA) for QAP. BMA in- tegrates an effective local optimization algorithm called Breakout Local Search (BLS) within the evolutionary computing framework which itself is based on a uniform crossover, a fitness-based pool updating strategy and an adaptive mu- tation procedure. Extensive computational studies on the set of 135 well-known benchmark instances from the QAPLIB revealed that the proposed algorithm is able to attain the best-known results for 133 instances and thus competes very favorably with the current most effective QAP approaches. A study of the search landscape and crossover operators is also proposed to shed light on the behavior of the algorithm. Keywords: Memetic algorithm; local search; landscape analysis; quadratic assignment; combinatorial optimization. 1. Introduction The quadratic assignment problem (QAP) is a classic NP-hard combinatorial optimization problem with a number of applications (Cheng, Ho, & Kwan, 2012; Garey & Johnson, 1979; Li, Xu, Jin, & Wang, 2012; Miao, Cai, & Xu, 2014; Nikolić & Teodorović, 2014; Pardalos, Rendl, & Wolkowicz, 1994). QAP is to determine a minimal cost assignment of n facilities to n locations, given a flow aij from facility i to facility j for all i, j ∈ {1, ..., n} and a distance bqp between locations q and p for all q, p ∈ {1, ..., n}. Let Π denote the set of the permutation functions π : {1, ..., n} → {1, ..., n}, then QAP can mathematically be formulated as follows: minπ∈Πf(π) = n∑ i=1 n∑ j=1 aijbπiπj (1) ∗Corresponding author. Email addresses: ube@cs.stir.ac.uk (Una Benlic), hao@info.univ-angers.fr (Jin-Kao Hao) Preprint submitted to Elsevier August 13, 2014 where a and b are the flow and distance matrices respectively, and π ∈ Π is a solution where πi represents the location chosen for facility i. The problem objective is then to find a permutation π∗ in Π that minimizes the sum of the products of the flow and distance matrices, i.e., f(π∗) ≤ f(π), ∀π ∈ Π. Besides the facility location problem, QAP is notable for its ability to for- mulate a number of other practical problems such as backboard wiring in elec- tronics, design of typewriter keyboards, campus planning, analysis of chemical reactions for organic compounds, balancing turbine runners, and many others. QAP can equally formulate some classic combinatorial optimization problems such as the traveling salesman, maximum clique and graph partitioning prob- lems. Reviews on some significant applications of QAP can be found in Burkard (1991); Duman & Or (2007) and Pardalos, Rendl, & Wolkowicz (1994), while many solution methods are reviewed in Anstreicher (2003). QAP is among the most studied and the hardest combinatorial optimization problems. In fact, from a theoretical point of view, QAP is NP-hard (Garey & Johnson, 1979). Consequently, no exact algorithm is expected to solve the problem in a polynomial time and even small instances may require considerable computation time. This hardness is confirmed in practice since the existing ex- act algorithms can solve to optimality only small instances from the QAP bench- mark library with up to 36 locations. Even approximation of the problem with a guaranteed performance is known to be very hard (Hassin, Levin, & Sviridenko, 2009). For these reasons, heuristic and metaheuristic methods constitute a nat- ural and useful approach for tackling this problem (Blum, Puchinger, Raidl, & Roli, 2011). Such algorithms aim to provide satisfactory sub-optimal solutions in acceptable computing time, but with no theoretically provable guarantee that the attained solutions are the optimal ones. Performance of these heuristic al- gorithms is typically assessed using a set of benchmark instances. Among the numerous heuristic algorithms reported for QAP in the literature, local search methods are very popular approaches, including simulated annealing (Wilhelm & Ward, 1987), tabu search (Battiti & Tecchiolli, 1994; James, Rego, & Glover, 2009a,b; Misevicius & Kilda, 2006; Skorin-Kapov, 1990; Taillard, 1991), and iterated local search (Benlic & Hao, 2013c; Stützle, 2006). Population-based approaches constitute another class of popular tools for finding high quality near-optimal solutions for QAP (Ahuja, Orlin, & Tiwari, 2000; Drezner, 2003, 2008; Fleurent & Ferland, 1993; Merz & Freisleben, 2000; Misevicius, 2004; Stützle, 2006). In this work, we are interested in solving QAP with heuristic algorithms. We introduce a powerful memetic algorithm (BMA) for QAP which combines an effective local search algorithm (BLS - Breakout Local Search), a crossover operator, a pool updating strategy, and an adaptive mutation mechanism. BLS is a general local search method which has shown very good results for several NP-hard problems including maximum clique (Benlic & Hao, 2013a), maximum cut (Benlic & Hao, 2013b), QAP (Benlic & Hao, 2013c) and vertex separator (Benlic & Hao, 2013d). Its basic idea is to use a descent procedure to discover local optima and employ dedicated perturbations to continually move from one attractor to another in the search space. In this paper, we integrate BLS into the 2 memetic framework, thus extending our work of Benlic & Hao (2013c). As we show in this paper, the proposed memetic algorithm BMA exhibits an excellent performance on the whole set of 135 well-known QAP benchmark instances. Indeed, BMA attains the best-known solution for all the instances except for only two cases. Furthermore, it outperforms its local search procedure BLS which confirms the usefulness of the memetic framework. In order to gain some insight into the functioning of the proposed memetic algorithm, we perform a landscape analysis and justify the choice for the used crossover operator (see Section 5). The paper is organized as follows. In the next section, we briefly review the most effective QAP approaches and highlight the contributions of this work. In Section 3, we present our proposed memetic algorithm (BMA) and detail its main components. Moreover, we highlight the differences and similarities between BMA and the reviewed state-of-art approaches. Computational results and comparisons with the top-performing QAP algorithms are presented in Section 4. Section 5 provides a landscape study that we use to justify the choices for the crossover operator of our memetic algorithm, and additionally shows a comparison of several crossover operators. Conclusions are provided in the last section. 2. State-of-art approaches for QAP and main contributions In this section, we provide a literature review of the most popular heuristic approaches for QAP, including four population-based algorithms (Drezner, 2008; Merz & Freisleben, 2000; Misevicius, 2004; Stützle, 2006) and two local search algorithms (James, Rego, & Glover, 2009a; Misevicius & Kilda, 2006), followed by a summary of the main contributions of this study. In Section 3.5, we discuss in more detail the relationships between these approaches and the proposed BMA. Among the reviewed approaches, four algorithms from Drezner (2008); James, Rego, & Glover (2009a); Misevicius (2004) and Misevicius & Kilda (2006) report the best results on some particular class of QAP benchmark instances. For this reason, these four algorithms will be used as reference algorithms for our comparative study. Nevertheless, it is important to mention that none of the existing QAP approaches can be considered as the most effective for all the different types of QAP instances, due to significant differences in structure of these instance (see Section 5). A popular memetic approach for QAP (MA-QAP) is introduced in Merz & Freisleben (2000) which incorporates the 2-opt local search procedure and an adaptation of the standard uniform crossover UX that does not perform any implicit mutation. The selection for reproduction is performed on a purely random basis, while the selection for survival is achieved by choosing the best individuals from the pool of parents and children. To overcome premature convergence, the restart technique proposed in Eshelman (1991) is employed. During the run, it is checked whether the average distance of the population has dropped below a certain threshold or whether the average fitness of the population did not change after a certain number of consecutive generations. If 3 one of these conditions holds, the whole population is mutated except the best individual, and each mutated individual is improved by the 2-opt local search to obtain a local optimum. Afterwards, the algorithm proceeds with the usual recombination process. We mention this work since it is one of the first memetic algorithms applied to QAP and achieved remarkable results at the time it was published. The Improved Hybrid Genetic Algorithm (IHGA) (Misevicius, 2004) incor- porates a robust local improvement procedure as well as an effective restart mechanism based on shift mutations. The author slightly improved the clas- sic scheme of a uniform like crossover (ULX) to get a new optimized crossover (OX). The optimized crossover is a crossover that (a) is ULX and (b) produces a child that has the best fitness value among the children created by M runs of ULX. The offspring is then improved with a local search mechanism, based on the swap neighborhood, which contains a tabu search procedure and a so- lution reconstruction procedure. The reconstruction is achieved by performing a number µ of random swaps, where µ is varied according to the instance size. Once convergence of the algorithm is observed, all the individuals but the best undergo the shift mutation (SM), which simply consists in shifting all the items in a wrap-around fashion by a predefined number of positions. IHGA is one of the best-performing algorithms for the unstructured instances and real-life like instances and is used as one of the references in our comparative study. Drezner (2008) shows extensive computational experiments on QAP using various variants of a hybrid genetic algorithm. The author compared the modi- fied robust tabu search (MRT) and the simple tabu search as local optimization algorithms combined with a crossover operator. Moreover, different parent selec- tion (distance modification, gender modification) and pool updating strategies were tested. The best version of the memetic algorithm is MRT60 which inte- grates the modified robust tabu search MRT for offspring improvement. MRT is identical to the robust tabu search (RTS) (Taillard, 1991) except that the tabu tenure is generated in [0.2n, 1.8n] rather than in [0.9n, 1.1n]. MRT60 is the best-performing memetic algorithm for the grid-based instances and is used as another reference algorithm in our comparative study. Population-based Iterated Local Search (PILS) (Stützle, 2006) is another highly effective algorithm. The underlying iterated local search (ILS) algorithm starts from a random assignment, and applies a first-improvement local search procedure based on the 2-opt neighborhood. To speed up the search process, the algorithm uses the don’t look bit strategy, previously proposed to accelerate local search algorithms for TSP. Once a local optimum is reached, ILS applies a perturbation that consists of exchanging k randomly chosen items, where k is varied between kmin and kmax. Stützle extends the described ILS to a population-based algorithm where no interaction between solutions takes place, and each single solution is improved by the standard ILS. In the proposed PILS, the population consists of µ solutions and in each iteration λ new solutions are generated. A selection strategy, based both on quality and distance between solutions, is then employed to form a new population of µ solutions from the set of µ + λ solutions. 4 Cooperative parallel tabu search algorithm (CPTS) (James, Rego, & Glover, 2009a) executes in parallel several tabu search (TS) operators on multiple pro- cessors. The TS operator is a modified version of Taillard’s RTS (Taillard, 1991) obtained by changing the stopping criterion and the tabu tenure parameters for each processor participating in the algorithm. In order to accomplish coopera- tion between TS processes, CPTS maintains a global reference set which uses information exchange to promote both intensification and diversification in a parallel environment. CPTS globally obtains excellent results on the whole set of QAP instances and is used as another reference algorithm in our comparative study. Iterated tabu search (ITS) by Misevicius & Kilda (2006) follows the general scheme of the iterated local search metaheuristic. It employs a traditional tabu search to reach local optima and triggers a perturbation (reconstruction) phase in order to escape the attained local optimum. The “ruined” solution becomes the new starting point for the basic TS procedure. ITS uses a perturbation mechanism which varies adaptively the number of random perturbation moves in some interval. ITS obtains excellent results on the unstructured instances and real-life like instances and is used as another reference in our comparison. Compared to the existing studies on QAP, this work has the following main contributions: First, the proposed BMA algorithm is based on a new local optimization pro- cedure (i.e., BLS) which adopts an adaptive perturbation mechanism to better escape local optima. The computational study discloses that this algorithm performs very well on the set of 135 very popular QAP benchmark instances by attaining the best-known result in 133 cases. This work thus confirms the usefulness of the memetic framework for QAP. Second, we provide an empirical justification to explain why the uniform crossover is the best operator for QAP in comparison with the other crossover operators studied in the literature. Third, ideas of the proposed algorithm could help in designing effective heuristics for other related permutation problems and applications such as those mentioned in the introductory section. 3. An effective memetic algorithm for QAP The term memetic algorithm (MA) is employed to designate a general heuris- tic approach which typically combines local optimization with a population- based paradigm (Moscato, 1989; Moscato & Cotta, 2003). The purpose of such a combination is to take advantages of both crossover that discovers unexplored promising regions of the search space, and local optimization that finds good solutions by concentrating the search around these regions. Since memetic al- gorithm is a problem-independent framework, it needs to be properly adapted to the specific problem at hand to ensure the best performance (Hao, 2012; Krasnogor & Smith, 2005). In particular, local optimization operator and re- combination operator are the two key components to consider. Finally, as for 5 any population-based method, a healthy diversity of population must be main- tained to avoid premature convergence. Previous studies show that memetic algorithms are able to achieve excellent performances for a number of optimiza- tion problems (Hart, Krasnogor, & Smith, 2004; Moscato & Cotta, 2003; Neri, Cotta, & Moscato, 2012). Given an initial population which consists of locally optimal solutions, a memetic approach generates new solutions by applying crossover and/or mu- tation, followed by a phase of local search to improve each offspring solution. The right choice for a crossover operator depends on problem structure and landscape properties (in Section 5, we show a study on the relationship between the performance of a crossover operator and the structural properties of a given problem). Moreover, the success of a memetic approach is conditioned by the ef- fectiveness of the local search procedure. While the main role of the crossover is to discover unexplored promising regions of the search space (i.e., exploration), local search basically aims to find good solutions by concentrating the search around these regions (i.e., exploitation). The proposed memetic algorithm for QAP (BMA) employs the uniform crossover operator (Section 3.1) (In Section 5.2, we justify why we chose this particular crossover), a Breakout Local Search (BLS) procedure (Section 3.2), a fitness-based population replacement strategy (Section 3.3), and an adaptive mutation mechanism (Section 3.4). Each offspring solution, generated with the uniform crossover, is improved with the BLS procedure. Our memetic approach then applies a pool updating strategy to possibly replace the worst individual from the population with the improved offspring solution. To avoid premature convergence, BMA triggers an adaptive mutation mechanism to the entire popu- lation if the best solution found during the search has not been improved during a fixed number of generations. This displaces the search to distant regions each time a search stagnation is detected. The general architecture of our memetic approach is described in Algorithm 1. The main components are detailed in the following sections. 3.1. Parents selection and recombination To determine a subset p ⊂ P of |p| parent individuals, we employ the tour- nament selection strategy. Let λ be the size of the tournament pool. We select each individual πi ∈ p in the following way: randomly choose λ individuals from P ; among the λ chosen individuals, place the best one into p if it is not already present in p. The time complexity of this operation is O(|P |). An advantage of the tournament selection is that the selection pressure can easily be adjusted by changing the size of the tournament pool λ. The larger the tournament pool is the less is the chance to select weaker individuals. An experimental evaluation of our QAP approach has revealed that putting a higher pressure on better individuals gives better results than using a random selection. For the crossover process, we employ a standard uniform operator (UX) that recombines two parent individuals from subset p. Elements of the parents are scanned from left to right and each element in the offspring keeps with equal probability the value j ∈ [1...n] found in either of the two parents, under the 6 Algorithm 1 General scheme of the proposed BMA algorithm 1: Initialize the number of BLS iterations for short and long runs ts and tl respectively, the minimum mutation degree µmin and the increment m of mutation degree 2: Randomly generate initial population P 3: P ← BLS(P, ts) /* Improve each individual with ts iterations of BLS, Sect. 3.2 */ 4: π best ← BestIndividual(P) /* Initialize the best individual */ 5: µ ← µmin /* Initialize the current mutation degree */ 6: for i := 0 to number of generations φ do 7: Select a subset of parent individuals p from P /* Sect. 3.1 8: π 0 ← Crossover(p) /* Generate an offspring, Sect. 3.1 */ 9: π 0 ← BLS(π0, tl) /* Improve offspring π 0 with long BLS run, Sect. 3.2 */ 10: P ← ReplacementStrategy(P, π0) /* Sect. 3.3 */ 11: if (πbest not improved after ν generations) then 12: P ← Mutate(P, µ) /* Sect. 3.4 */ 13: P ← BLS(P, ts) 14: Update the best solution πbest if necessary 15: µ ← µ + m /* Increase mutation degree */ 16: end if 17: if (f(πbest) > f(π0) or µ > n) then 18: µ ← µmin /* Reset mutation degree to default */ 19: end if 20: if (f(πbest) > f(π0)) then 21: π best ← π 0 22: end if 23: end for constraint that j has not been assigned before to any element in the offspring. Any unassigned element is given a random value j such that j does not appear twice in the offspring. An example of this recombination process is illustrated in Figure 1. The complexity of this crossover is O(n). Since the uniform crossover permits great flexibility, different variations of its basic procedure have been proposed and applied to QAP (Merz & Freisleben, 2000; Misevicius, 2004). Despite its simplicity, the UX has shown to provide very good results on different combinatorial optimization problems including QAP. A comparison of the simple UX with several other crossover operators used for QAP is reported in Section 5.2. 3.2. Breakout local search (BLS) The local optimization procedure has a significant impact to the overall performance of a memetic algorithm. In our case, we adopt an existing local search algorithm BLS-QAP (Benlic & Hao, 2013c) (call it BLS for simplicity). As most local search algorithms for QAP, BLS employs the swap operator which consists in exchanging the values from two different positions in π, i.e., permuting the locations of two facilities. It uses the best improvement descent procedure to exploit the whole swap neighborhood N(π), which is evaluated in 7 Figure 1: An example of the uniform crossover (UX) O(n2) time thanks to an effective neighborhood evaluation strategy proposed in Taillard (1991). Once a local optimum is returned by the best improvement descent pro- cedure, BLS triggers a multi-type diversification mechanism which adaptively determines the type T of perturbation moves and the number L of perturbations (called jump magnitude) by considering some information related to the search state. This mechanism combines two complementary types of perturbation: a guided perturbation (using a tabu list) and a random perturbation. The tabu- based perturbation uses a selection rule that favors swap moves that minimize the cost degradation, under the constraint that the move has not been applied during the last γ iterations (where γ is the tabu tenure that takes a random value from a given range), while the random perturbation performs moves se- lected uniformly at random. In order to insure a good balance between an intensified and a diversified search, BLS-QAP alternates probabilistically be- tween the two types of perturbations. The probability of applying a particular perturbation is determined dynamically depending on the current number ω of consecutive non-improving attractors visited (see Benlic & Hao (2013c) for more details). We limit the probability of applying the tabu-based perturbation over the random perturbation to take values not less than Q. To determine the jump magnitude L for the following perturbation phase, the proposed algorithm uses a basic strategy which increments L if the local search procedure returned to the immediate previous local optimum, and otherwise resets L to its initial value L0. Once the type T and the number L of perturbations are determined, we apply accordingly L moves of type T to the current solution. The resulting solution is used by the next round of the best improvement descent procedure as its new starting point. Once an offspring solution is created in the crossover phase (line 8 of Alg. 1), it is improved with tl iterations of the BLS procedure (line 9 of Alg. 1). 8 As previously mentioned, the complexity of each iteration of the BLS descent procedure is O(n2). The number of the descent iterations performed in each iteration of BLS depends on the size sbasin of the basin of attraction of a local optimum. This, in turn, depends on the type of perturbation used for the previous perturbation phase. More precisely, after a phase of the directed (tabu- based) perturbation, the descent-based local search requires, on average, less steps to attain a local optimum than after the random perturbation. However, the computational complexity of one iteration of the tabu-based perturbation is O(n2), while the complexity of the random perturbation is O(1). Therefore, the time complexity of one BLS iteration is bounded within O(sbasin·n 2)+O(L·n2). 3.3. Pool updating strategy For each offspring solution π0 created by the crossover operator and im- proved with the BLS procedure (Section 3.2), we decide whether π0 should be inserted into the population pool (lines 8-10 of Alg. 1). To base this decision, our algorithm uses a classic replacement strategy which inserts π0 into P if there is no individual in P identical to π0 (i.e., ∀πi ∈ P, πi 6= π0), and if the fitness of π0 is better than the fitness of the worst individual πworst from P (i.e., f(π0) < f(πworst)). If this condition holds, P is updated by replacing πworst with π0. To determine whether an identical solution to π0 is already present in P , we compute the hamming distance between π0 and each πi ∈ P . The ham- ming distance is defined as the number of positions at which the corresponding elements are different. If two solutions are identical, the corresponding ham- ming distance is 0. The overall time complexity of this pool update strategy is thus O(|P | · n). One potential risk of this replacement strategy is the premature loss of popu- lation diversity, since offspring is being inserted into the population regardless of its distance to other individuals in the population. To avoid this problem, more sophisticated pool updating strategies have been proposed in the literature that maintain population diversity by considering the similarity (distance) between the offspring and other individuals from the population (see for instance Benlic & Hao (2011); Lü & Hao (2010); Porumbel, Hao, & Kuntz (2010)). In our case, the diversity is maintained by an adaptive mutation mechanism described in Section 3.4. As a result, the proposed memetic approach is not really sensitive to the pool replacement strategy employed. 3.4. Adaptive mutation procedure As soon as a search stagnation is detected, i.e., the best solution πbest found during the search has not been updated for a certain number of generations, our BMA algorithm mutates the entire population with an adaptive mutation procedure that adjusts the diversification strength µ ∈ [µmin, n] depending on the previous search progress (see lines 11-16, Algorithm 1). At the beginning, µ is set to µmin and is gradually augmented by an increment m if a solution better than πbest has not been attained during ν generations. Once a solution better than πbest is obtained, or µ has reached the maximum possible value n (n being the problem size), µ is reset to µmin (see lines 17-19, Algorithm 1). 9 Figure 2: An example of mutation with 3 exchanges for µ = 3 The mutation procedure of our memetic algorithm, which was previously used in Merz & Freisleben (2000) for QAP, exchanges a sequence of locations in the solution π to create an offspring that has a distance of µ to its parent π. The distance measure used for this mutation operator is the well-known hamming distance. To ensure that the distance between the offspring and its parent is µ, in each step, the second selected location is swapped again in the subsequent step, such that the resulting distance is equal to one plus the number of swaps. An example of the mutation procedure is illustrated in Figure 2. The complexity of the mutation procedure is O(|P | · µ). 3.5. Discussions In this Section, we discuss similarities and differences between BMA and the previously mentioned state-of-art QAP approaches, in particular those based on the memetic framework. In Section 2, we reviewed three popular memetic QAP algorithms, MA-QAP (Merz & Freisleben, 2000), IHGA (Misevicius, 2004), and MRT60 (Drezner, 2008), the two latter ones being among the best performing algorithms cur- rently available in the literature for this problem. The genetic operators em- ployed by BMA are quite similar to those of MA-QAP and IHGA. Indeed, these approaches create offspring solution by means of different adaptations of the classic uniform crossover operator which provides surprisingly good results for the QAP instances. One of the contributions of our work is an explanation for the efficiency of these random crossovers obtained by analyzing the structures of the QAP instances (see Section 5). Moreover, as MA-QAP and IHGA, BMA applies a mutation mechanism to the entire population in order to avoid premature convergence. However, unlike the previously used mutation strategies, the mutation mechanism of BMA adaptively adjusts the diversification strength depending on the previous search progress. 10 Yet, the most important difference between BMA and the three reference memetic algorithms is the local search procedure. As stated earlier, BLS, em- ployed by BMA for offspring improvement, constitutes the key component of BMA. It combines the steepest descent with a dedicated and adaptive diver- sification mechanism. On the other hand, the local search phase performed by MA-QAP is the basic steepest descent algorithm, while IHGA and MRT60 improve solutions by means of a simple tabu search procedure. PILS (Stützle, 2006) is a population-based approach which uses an adap- tation of the iterated local search framework to improve each newly generated solution. Although the perturbation mechanism of PILS determines the number of perturbation moves in an adaptive way, it is only based on random moves which constitutes the main difference with our BLS procedure. Two local search algorithms, CPTS (James, Rego, & Glover, 2009a) and ITS (Misevicius & Kilda, 2006) (see Section 2), are among the top performing QAP methods. Even though BLS’s directed perturbation is based on the use of a tabu list, CPTS is not very much related to BLS since it consists of a parallel execution of several tabu search (TS) operators on multiple processors. The proposed BLS is more related to ITS since both algorithms are variants of the iterated local search method. Nevertheless, there are two major differences between ITS and BLS. Firstly, BLS does not consider the tabu list during its local search (descent) phases, unlike ITS which constraints each iteration of its local search phase with a tabu list. As such, BLS and ITS explore different trajectories during their respective search, leading to different local optima. In fact, one of the keys to the effectiveness of BLS on QAP is that it completely excludes diversification during local search, unlike tabu search for which the intensification and diversification are always intertwined. Secondly, unlike ITS which applies solely random moves to perturb the current local optimum, BLS adaptively choses between two types of moves (random and directed) according to the search status, leading to variable levels of diversification. Finally, we mention the approach proposed by Kelly, Laguna, & Glover (1994). In their work, the authors present a method which also starts with the steepest descent to find local optima, and then applies a diversification strategy to incrementally restrict the set of allowed swap moves in order to escape local optima. The diversification procedure of this approach exploits the search history as well, but unlike BLS, performs perturbations in a purely deterministic way. It consists of imposing a “maximum tabu tenure” to moves performed during each descent of local search, and applying the best move among the non prohibited ones to perturb a local optimum. 4. Experimental evaluation 4.1. Benchmark instances It is worthy to recall that given the very nature of heuristic algorithms, it is a common practice to evaluate the performance of a new algorithm by testing it on well-established benchmark instances of the problem and comparing its results 11 Table 1: Settings of important parameters. Para. Description Value |P | Population size 15 ts number of iterations for short BLS run 5000 tl number of iterations for long BLS run 10000 µmin minimal mutation degree 0.5n m increment of mutation degree 0.1n λ the size of the tournament pool 4 ν number of generations without improvement before mu- tation |P | L0 initial jump magnitude of BLS 0.05n (T. I & II), 0.15n (T. III & IV) γ tabu tenure for directed perturb. with BLS random[0.9n, 1.1n] Q smallest probability for applying BLS directed perturba- tion 0.75 with those of the state-of-art methods. In our case, we evaluate the performance of our BMA algorithm on the set of 135 instances from the QAPLIB1, whose size n varies from 12 to 150 and is indicated in the instance name. These instances are very popular and largely used in the literature. They are typically classified into four types covering various real applications and random problems: Type I. Real-life instances obtained from practical applications of QAP; Type II. Unstructured, randomly generated instances for which the distance and flow matrices are randomly generated based on a uniform dis- tribution; Type III. Randomly generated instances with structure that is similar to that of real-life instances; Type IV. Instances in which distances are based on the Manhattan distance on a grid. Among the 135 instances from the QAPLIB, we focus on the set of 21 selected instances. The remaining 114 instances (including all the real-life instance of Type I) are omitted from our experimental comparisons since BLS and BMA (and many other state-of-art QAP methods) can solve them to optimality in every single trial within a very short computation time (often less than a second). However, to be exhaustive, we include the results of our BMA algorithm for these 114 instances in the Appendix (Table 8). 4.2. Experimental protocol The proposed memetic algorithm BMA2 is programmed in C++, and com- piled with GNU g++ on a Xeon E5440 with 2.83GHz and 2GB. Like other QAP heuristic algorithms, BMA requires a number of parameters to be tuned (see 1urlhttp://www.seas.upenn.edu/qaplib/inst.html 2The code of our memetic algorithm, used to obtained the reported results, will be made available online at http://www.info.univ-angers.fr/pub/hao/BMA.html 12 Table 1), most of which are related to its BLS local optimizer and adopt the values used in Benlic & Hao (2013c). According to the parameter analyses per- formed in Benlic & Hao (2013b,c), the most relevant BLS parameters are the initial jump magnitude L0 and the smallest probability for applying directed over random perturbation Q. Moreover, as discussed in Benlic & Hao (2013c), the optimal setting of these parameters depends on the landscape properties of the QAP instance at hand and may vary for different instances. Additionally, following the existing MA literature on discrete optimization (Benlic & Hao, 2011; Bontouxa, Artigues, & Feillet, 2010; Dorne & Hao, 1998; Hao, 2012; Lü & Hao, 2010; Merz & Freisleben, 2000), BMA maintains a population of fairly limited size. This, as well as the other parameters related to the genetic algo- rithm components, is determined with a preliminary experiment. Even though it is possible to find a configuration of parameters better than the one used in this paper, the computational experiments reported in this section show that the adopted setting performs globally well on the tested benchmark instances. More generally, since the source code of our BMA algorithm will be made avail- able online, the potential user can possibly apply any preferred method to tune these parameters. According to the practice in the QAP literature, the stopping condition is the elapsed time. In our case, we set the maximum time limit to 2 hours for all the instances, except for the largest instance tho150 for which we set the maximum time allowed to 10 hours. Notice that some reference QAP algo- rithms (Drezner, 2008; James, Rego, & Glover, 2009a) use an even higher time limit for tho150 to reach the reported result. As shown below, the best-known solutions are very often attained long before these time limits. Furthermore, we provide computational results of our BMA algorithm with the time limit reduced to 30 minutes. The reported results for BMA and BLS are obtained over 10 independent executions. We focus primarily on the comparisons in terms of the solution quality with respect to the best-known results (BKR) reported in the literature, which were obtained with different QAP algorithms under various conditions. For indicative purposes, we also compare the results produced by our memetic approach and those obtained with the 4 state-of-art QAP approaches which are the current best performing methods: 1. Cooperative Parallel Tabu Search (CPTS) algorithm by James, Rego, & Glover (2009a). The reported results are obtained using ten (1.3GHz) Intel Intanium processors; 2. Iterated Tabu Search (ITS) by Misevicius & Kilda (2006). The results are obtained using a 900 MHz Pentium computer; 3. Improved Hybrid Genetic Algorithm (IHGA) by Misevicius (2004). The reported results are obtained on a x86 Family 6 processor; 4. A variant of a Hybrid Genetic Tabu Search Algorithm (MRT60) by Drezner (2008). The reported results are obtained on a Pentium IV 2.8 GHz com- puter. 13 An exhaustive comparative analysis with the reference approaches is not a straightforward task because of the differences in computing hardware, re- sult reporting methodology, termination criterion, etc. This comparison is thus presented only for indicative purposes. Nevertheless, this experiment provides interesting indications on the performance of the proposed algorithm relative to the state-of-art algorithms. In addition, we evaluate the contribution of the proposed memetic approach by comparing (under the same conditions) BMA with its BLS procedure (Benlic & Hao, 2013c) which itself shows to be highly effective on the tested QAP instances. This allows us to highlight the usefulness of the memetic framework. According to the literature (Drezner, 2008; Misevicius, 2004; Stützle, 2006), we employ several criteria for evaluation and comparison of our memetic algo- rithm with other QAP approaches. 1. The number of instances for which an optimal or best-known solution is reached within a reasonable computing time. This constitutes an indicator on the effectiveness of an algorithm in terms of the solution quality. 2. The success rate of reaching an optimal or best-known solution which provides information about the robustness of an algorithm. 3. The percentage deviation δ̄avg of the average solution from the published best-known result over a certain number of runs. The percentage deviation between solutions is computed as δ̄ = 100(z − z̄)/z̄[%], where z is the average result over a given number of runs and z̄ the best-known objective value. This indicator provides additional information about the robustness of an algorithm. Beside the aforementioned criteria, we also mention the amount of compu- tational time required by each approach to reach the reported results. This pro- vides an indication about the computational efficiency of an algorithm. More- over, Table 7 in the Appendix provides a detailed summary of the computational results for 21 selected QAPLIB instances obtained by BMA within a time limit of 2 hours (10 hours for instance tho150) and 30 minutes respectively. 4.3. Computational results and comparisons Table 2 reports comparative results with BLS (Section 3.2), CPTS (James, Rego, & Glover, 2009a), ITS (Misevicius & Kilda, 2006) and IHGA (Misevicius, 2004), for unstructured instances (Type II) and real-life like instances (Type III). Since MRT60 (Drezner, 2008) does not report results on these instances, it is excluded from this comparison (in fact, MRT60 only reports results for instances of Type IV). The second column ‘BKS’ shows for each instance the best-known objective value ever reported in the literature. For each algorithm, column δ̄avg indicates the percentage deviation between an average solution, obtained with the given approach over 10 independent trials, and the best-known solution. The success rate for reaching the best-known solution over 10 trials is given in 14 Table 2: Comparative results between the proposed memetic algorithm (BMA), BLS (Benlic & Hao, 2013c), and three best performing QAP approaches on unstructured instances (Type II) and real-life like instances (Type III): CPTS (James, Rego, & Glover, 2009a), ITS (Misevicius & Kilda, 2006) and IHGA (Misevicius, 2004). The success rate of reaching the best-known result over 10 executions is indicated between parentheses. Computing times are given in minutes for indicative purposes. Problem BKS BMA BLS CPTS ITS IHGA % δ̄avg t(m) % δ̄avg t(m) % δ̄avg t(m) % δ̄avg t(m) % δ̄avg t(m) Random instances (Type II) tai40a 3139370 0.059(2) 8.1 0.022(7) 38.9 0.148(1) 3.5 0.210(1) 0.8 0.209(1) 1.4 tai50a 4938796 0.131(2) 42.0 0.157(2) 45.1 0.440(0) 10.3 0.373(0) 3.0 0.262(0) 5.0 tai60a 7205962 0.144(2) 67.5 0.251(1) 47.9 0.476(0) 26.4 0.330(1) 9.7 0.583(0) 12 tai80a 13499184 0.426(0) 65.8 0.517(0) 47.3 0.691(0) 94.8 0.494(0) 25.0 0.756(0) 53.3 tai100a 21052466 0.405(0) 44.1 0.430(0) 39.0 0.589(0) 261.2 0.427(0) 60.0 0.606(0) 200.0 Average 0.233 45.5 0.275 43.6 0.469 79.2 0.367 19.2 0.483 54.3 Real-life like instances (Type III) tai50b 458821517 0.000(10) 1.2 0.000(10) 2.8 0.000(10) 13.8 0.000(10) 0.9 0.000(10) 0.3 tai60b 608215054 0.000(10) 5.2 0.000(10) 5.6 0.000(10) 30.4 0.000(10) 2.2 0.000(10) 0.7 tai80b 818415043 0.000(10) 31.3 0.000(10) 11.4 0.000(10) 110.9 0.000(10) 5.8 0.000(10) 2.5 tai100b 1185996137 0.000(10) 13.6 0.000(10) 16.0 0.001(8) 241.0 0.000(9) 23.3 0.000(10) 7.3 tai150b 498896643 0.060(1) 78.1 0.100(0) 80.5 0.076(0) 7377.8 0.100(1) 60.0 0.111(2) 38.3 Average 0.012 25.9 0.020 23.3 0.015 1554.8 0.020 18.4 0.022 9.8 1 5 Table 3: Comparative results between the proposed memetic algorithm (BMA), BLS (Benlic & Hao, 2013c), CPTS (James, Rego, & Glover, 2009a) and MRT60 (Drezner, 2008) on grid-based (Type IV) instances. The success rate of reaching the best-known result over 10 executions is indicated between parentheses. Computing times are given in minutes for indicative purposes. Problem BKS BMA BLS CPTS MRT60 % δ̄avg t(m) % δ̄avg t(m) % δ̄avg t(m) % δ̄avg t(m) sko72 48498 0.000(10) 3.5 0.000(10) 4.1 0.000(10) 69.6 0.000(10) 19.9 sko81 66256 0.000(10) 4.3 0.000(10) 13.9 0.000(10) 121.4 0.000(10) 31.9 sko90 90998 0.000(10) 15.3 0.000(10) 16.6 0.000(10) 193.7 0.000(10) 48.5 sko100a 115534 0.000(10) 22.3 0.001(9) 20.8 0.000(10) 304.8 0.000(10) 73.6 sko100b 152002 0.000(10) 6.5 0.000(10) 10.8 0.000(10) 309.6 0.000(10) 73.6 sko100c 147862 0.000(10) 12.0 0.000(10) 15.5 0.000(10) 316.1 0.000(10) 73.6 sko100d 149576 0.006(9) 20.9 0.001(5) 38.9 0.000(10) 309.8 0.000(10) 73.6 sko100e 149150 0.000(10) 11.9 0.000(10) 42.5 0.000(10) 309.1 0.000(10) 73.6 sko100f 149036 0.000(10) 23.0 0.000(10) 17.3 0.003(4) 310.3 0.000(9) 43.5 wil100 273038 0.000(10) 14.5 0.000(10) 18.9 0.000(10) 316.6 0.000(10) 73.6 tho150 8133398 0.008(3) 416.4 0.023(1) 268.8 0.013(0) 1991.7 0.003(3) 1223.6 Average 0.001 50.1 0.002 42.6 0.001 413.9 0.000 164.5 parentheses next to the value of δ̄avg. The CPU time (in minutes) is only given for indicative purposes. From Table 2, we can make the following observations. For the unstructured instances (Type II), BMA finds the best-known solution for 3 out of the 5 instances, with an average deviation δ̄avg of 0.233 over the 5 instances. For three hard instances tai40a, tai50a and tai60a, it reaches the best-known solution in 20% of the trials. Compared to its local search procedure BLS, BMA statistically outperforms (with p-value< 0.05) BLS for two hard instances, tai60a and tai80a, and reduces the average percentage deviation from 0.275 to 0.233 over the 5 instances. The current most effective approach for instances of Type II is probably ITS (Misevicius & Kilda, 2006), which attains the best-known result for 2 out of the 5 instances with an average deviation δ̄avg of 0.367. One notices that ITS shows worse results than both BMA and BLS, but it requires shorter computing time. To verify the performance of BMA under short computing budgets, we show in Appendix (Table 7) the results of BMA within a significantly shorter cutoff limit (30 minutes). From Table 7, we observe that BMA does remain very competitive with ITS, with an average deviation δ̄avg of 0.356 over the 5 instances. Finally, the two other reference algorithms CPTS and IHGA achieve results which are slightly worse than BMA, BLS and ITS. For the real-life like instances (Type III), BMA is able to attain the best- known solution for all the instances in every single trial, except for the largest instance tai150b where the success rate is 10%. It reports an average deviation δ̄avg of 0.012 over the 5 instances. Unlike BMA, BLS is unable to reach the best-known solution for tai150b with the given computing conditions. When the running time of BMA is greatly reduced, it is still able to attain the best- known solution for all the five instances but with an average deviation δ̄avg of 0.051 (see Table 7). Two reference approaches, ITS (Misevicius & Kilda, 2006) and IHGA (Misevicius, 2004), are also able to attain the best-known result for all the 5 real-life like instances with an average deviation δ̄avg of 0.020 and 0.022 respectively. 16 We now turn our attention to the instances with grid distances (Type IV, Table 3). Among the 4 reference methods, only CPTS (James, Rego, & Glover, 2009a) and MRT60 (Drezner, 2008) report results for these instances. In Table 3, we show the results of our BMA and BLS algorithms together with those of CPTS and MRT60 for these Type IV instances. From Table 3, we observe that BMA is able to reach the best-known result for all the instances of this type. For 9 out of 11 instances, it has a success rate of 100%. For the two remaining instances (sko100d and tho150), the success rate is 90% and 30% respectively. BLS also attains the best-known result for the hardest instance tho150 with a success rate of 10%. MRT60 has a slightly better success rate than BMA on instance sko100d (10/10 v.s. 9/10) and reports a slightly worse success rate than BMA on sko100f (9/10 v.s. 10/10). On the other instances, BMA and MRT60 have the same success rates with a very slight advantage for MRT60 in terms of the average gap over all the instances. On the other hand, CPTS performs globally well except for the hardest instance tho150 for which CPTS fails to attain the best-known solution. We conclude that BMA competes favorably with MRT60 and CPTS on the Type IV instances. 5. Analysis and discussion The performance of a memetic algorithm may be influenced by the character- istics of the search landscape like the average distance between local optima and the relative distance of local optima to the nearest global optimum. Analyses of correlation between solution fitness and distance to global optimum have shown to be particularly useful for a better understanding of algorithm behavior and for designing suitable operators for a more effective search. Landscape studies have been previously reported for QAP in Merz & Freisleben (2000) and Stützle (2006) and for other well-known problems like the TSP problem (Boese, 1995) and the flow-shop scheduling problem (Reeves, 1997). In this section, we report a similar analysis, based on solutions sampled by our BLS procedure which are probably quite different from the solutions used in the previous studies on QAP. We perform the landscape analysis on 16 QAPLIB instances, based on a set of distinct solutions obtained after 2000 independent runs of our BLS approach. The number of iterations per run is set to 200000. As the distance between solutions, we calculate the number of facilities that are allocated to distinct locations in two solutions π and π′, i.e., d(π, π′) = |{i|πi 6= π ′ i}|. Since global optima for the analyzed instances are not known, we use instead the best-known local optima to compute fitness-distance correlation and refer to them as global optima. 5.1. Landscape analysis - FDC and distribution of local optima The fitness distance correlation (FDC) coefficient ρ (Jones & Forrest, 1995) captures the correlation between the solution fitness and its distance to the nearest global optimum (or best-known solution if global optimum is not avail- able). For a minimization problem, if the fitness of a solution decreases with 17 the decrease of distance from the optimum, then it should be easy to reach the target optimum for an algorithm that concentrates around the best candidate solutions found so far, since there is a “path” to the optimum via solutions with decreasing (better) fitness. A value of ρ = 1 indicates perfect correlation between fitness and distance to the optimum. For correlation of ρ = −1, the fitness function is completely misleading. FDC can also be visualized with a FD plot, where the same data used for estimating ρ is displayed graphically. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 20 40 60 80 100 % d e vi a tio n t o t h e b e st k n o w n s o lu tio n Distance to optimum tai100a 0 0.05 0.1 0.15 0.2 0.25 0.3 0 20 40 60 80 100 % d e vi a tio n t o t h e b e st k n o w n s o lu tio n Distance to optimum sko100a 0 0.5 1 1.5 2 2.5 0 20 40 60 80 100 % d e vi a tio n t o t h e b e st k n o w n s o lu tio n Distance to optimum tai100b Figure 3: Distances of local optima to the best-known solution based on solutions sampled by the BLS algorithm for instances tai100a (Type II), sko100a (Type III) and tai100b (Type IV). In column ‘ρ’ of Table 4, we report FDC coefficients for the 16 selected QAP instances. For illustrative purpose, FD plots for three instances (tai100a, sko100a and tai100b) are given in Figure 3. As it can be seen from the FDC coefficients in Table 4, there is a clear difference in correlation among instances of different types. For randomly generated instances (Type II), the FDC coefficient ρ is negative except in one case (tai50a) where ρ is close to zero. Indeed, from the FD plots in Figure 3 it is clear that there is no correlation between fitness 18 Table 4: Analytical results for 16 QAP instances. Column ‘#dlo’ indicates the number of distinct local optima over 2000 independent runs of BLS; Columns ‘avg dlo’, ‘avg dgo’ and ‘avg dhq’ report respectively the average distance between local optima, the average distance between local and global optima, and the average distance between the highest quality local optima; Column ‘ρ’ shows the value of the fitness-distance correlation coefficient. Instance #dlo avg dlo avg dgo avg dhq ρ Type tai40a 349 38.5 39.2 38.4 -0.028 II tai50a 1760 48.6 49.1 48.7 0.029 II tai60a 1995 58.6 59.2 58.6 -0.024 II tai80a 2000 78.8 79.0 78.8 -0.058 II tai100a 2000 98.8 99.0 98.9 -0.012 II tai80b 730 61.7 78.7 68.7 0.040 III tai100b 1007 80.7 97.7 89.1 0.11 III sko72 228 66.8 68.3 65.7 0.287 IV sko81 388 75.4 74.5 76.5 -0.048 IV sko90 487 85.08 85.7 82.8 0.348 IV sko100a 968 95.4 94.2 95.4 0.366 IV sko100b 628 95.9 92.5 95.5 0.280 IV sko100c 559 91.0 91.3 93.9 0.233 IV sko100d 1212 96.3 93.9 96.3 0.433 IV sko100e 545 94.3 92.4 96.2 0.569 IV sko100f 1230 96.6 93.5 97.2 0.317 IV and distance for the random instance tai100a. For grid-based instances (Type IV), significant FDC exists except in one case (sko81, ρ < 0), while for two real-life like instances tai80b and tai100b the FDC is also low ρ < 0.15 but higher than for random instances. The FD plots in Figure 3 for instances sko100a and tai100b also confirm this observation. Table 4 additionally reports the average distance between local optima avg dlo, the average distance between local optima and the nearest global optimum avg dgo, and the average distance between the highest quality local optima avg dhq. It can be observed that, regardless of the instance type, the values of avg dlo, avg dgo, and avg dhq are very large, close to the maximal possible value n. This implies that local optima are scattered in the search space. Even high quality local optima are distant from each other and share no common structure. These observations will allow us to understand why the uniform crossover operator is appropriate for QAP. 5.2. Comparison of recombination operators The objective of this section is to justify, based on the landscape analysis provided in Section 5.1, the choices for the crossover operator used by our BMA algorithm. We compare three versions of our BMA algorithm incorporating the three recombination operators of the literature for permutation problems: the standard uniform crossover (UX) described in Section 3.1, the block crossover (BX) and the distance preserving crossover (DPX). The block crossover (BX), also called the multi-point crossover, is one of the classic recombination operators. Firstly, it chooses randomly a certain number of crossing points. Once the points have been chosen, starting from left to right, all the elements of the first parent are simply copied over to the offspring π0 until the first crossing point is reached. Then all the elements of the second parent are copied over to π0 until the second crossing point is reached. This procedure 19 continues until reaching the last position in π0. During the crossover process, we constrain an element to take on a value j ∈ [1...n] that has not already been assigned to some element in π0. The rest of unassigned elements are randomly assigned a value j ∈ [1...n] such that solution feasibility is maintained. The distance preserving crossover (DPX) has previously been applied to QAP in Merz & Freisleben (2000). The basic idea of DPX is to create an offspring that has the same distance to each of its parents, and that distance is equal to the distance between the parents themselves. Elements with the same values in both parents are copied to the offspring. The values of all the other elements change. Generally, we can classify recombination operators into two classes, those that try to exploit an existing structure in a problem (e.g., the DPX crossover) and those that perform recombination in a purely random way (e.g., UX and BX crossovers). The former operators usually perform well on the class of problems with exploitable global structure and landscapes with highly correlated local optima (e.g., the traveling salesman problem (Bontouxa, Artigues, & Feillet, 2010) and the graph partitioning problem (Benlic & Hao, 2011)). Otherwise, these operators become destructive introducing a too strong perturbation. Table 5 shows the average percentage deviation from the best-known result δ̄avg for the three versions of our BMA over 10 runs on 16 QAPLIB instances. The stopping condition used is the number of generations φ = 1, 500 with 10, 000 BLS iterations per generation. We indicate in parentheses the number of times each version of BMA reached the best-known result over the 10 executions. Moreover, we study the degree of perturbation pstr induced by each crossover, and report in Table 6 the average pstr caused by the three operators over the 1,500 generations. Here, we define the perturbation degree pstr as the minimum distance between the created offspring and one of its parents expressed as a percentage of n. From Table 5, we observe that the best performance is obtained by the BMA version integrating the standard UX operator, with an average δ̄avg of 0.073 over the 16 instances. The second best performing BMA version inte- grates the block crossover (BX) with an average δ̄avg of 0.079. On the other hand, the results indicate that DPX is less effective with an average δ̄avg of 0.100 over the 16 instances. We further observe that the difference in performance between operators UX & BX and DPX is particularly notable on unstructured instances (Type II) since local optima are the most uncorrelated for these in- stances. Indeed, on the other instances with slightly higher FDC coefficient ρ, this difference is much less obvious. As expected, we observe from Table 6 that on random, unstructured in- stances, the amount of perturbation induced with DPX is much stronger than with UX and BX, since DPX is designed to exploit the problem structure that is missing for instances of Type II. Indeed, the degree of perturbation pstr with DPX crossover on unstructured instances is almost 100%. On the other hand, the amount of perturbation caused by DPX is greatly smaller when applied to more structured instances of Type IV. 20 Table 5: Percentage deviations δ̄avg of the average solution (obtained after 10 runs) from the published best-known result for the three versions of our BMA integrating respectively the uniform (UX), the block (BX), and the distance preserving (DPX) crossover. Instance UX BX DPX Instance UX BX DPX tai40a 0.059(2) 0.067(1) 0.074(0) sko100a 0.000(10) 0.000(10) 0.001(9) tai50a 0.135(2) 0.250(1) 0.286(0) sko100b 0.000(10) 0.000(10) 0.000(10) tai60a 0.220(2) 0.249(1) 0.270(0) sko100c 0.000(10) 0.000(10) 0.000(10) tai80a 0.410(0) 0.420(0) 0.536(0) sko100d 0.000(10) 0.000(10) 0.000(10) tai100a 0.340(0) 0.270(0) 0.431(0) sko100e 0.000(10) 0.000(10) 0.000(10) sko72 0.000(10) 0.000(10) 0.000(10) sko100f 0.000(10) 0.000(10) 0.000(10) sko81 0.000(10) 0.000(10) 0.000(10) tai80b 0.000(10) 0.000(10) 0.000(10) sko90 0.000(10) 0.010(8) 0.003(9) tai100b 0.000(10) 0.000(10) 0.000(10) Table 6: The average perturbation degree pstr over 1500 generations introduced by three versions of our BMA integrating respectively the uniform (UX), the block (BX), and the distance preserving (DPX) crossover. pstr is expressed as a percentage of n. Instance UX BX DPX Instance UX BX DPX tai40a 45.8 36.7 95.9 sko100a 21.8 13.7 95.4 tai50a 46.9 39.0 97.1 sko100b 16.9 14.0 47.6 tai60a 45.2 38.6 97.5 sko100c 18.9 15.9 41.6 tai80a 48.4 36.9 98.4 sko100d 18.5 11.8 40.3 tai100a 46.9 36.5 98.6 sko100e 29.5 22.7 43.1 sko72 22.3 32.4 76.8 sko100f 28.3 20.6 63.9 sko81 25.4 21.1 67.1 tai80b 13.7 10.6 35.6 sko90 27.9 24.6 49.0 tai100b 17.3 11.9 27.5 6. Conclusion In this article, we presented a simple and effective memetic algorithm (BMA) for the well-known quadratic assignment problem. BMA combines a dedicated local search named Breakout Local Search (BLS) with the standard uniform crossover (UX), a simple pool updating strategy and an adaptive mutation mechanism. The BLS procedure, which is the key component of BMA, alternates between a local search phase (to reach local optima) and a dedicated perturbation phase (to discover new promising regions). The perturbation mechanism of BLS dy- namically determines the number of perturbation moves and adaptively chooses between two types of perturbation moves of different intensities depending on the current search state. Genetic operators (crossover and mutation) are integrated to further enforce the search capacity of BMA. The choice of these operators is based on observa- tions made from a landscape analysis which investigates the structure of QAP instances. The analysis revealed a very low fitness-distance correlation between local optima for randomly generated instances, and a medium correlation for instances of other types. These observations provide a basis to explain why the standard uniform crossover generally leads to better results for unstructured instances (Type II) than a crossover that tries to exploit the problem structure. We evaluated the proposed algorithm on the set of 135 benchmark instances from the QAPLIB. Computational results revealed that BMA performs very well on these instances. Indeed, BMA outperforms its local search component (BLS) and is able to reach the current best-known solution for 133 instances. 21 In particular, BMA performs particularly well on unstructured instances (Type II) which are considered to be the hardest for the existing QAP methods. For real-life like instances (Type III) and instances with grid distances (Type IV), BMA remains competitive with respect to the best performing approaches as well. The computing time needed for our proposed algorithm to reach its best solution varies on average from several seconds to about one hour, except for the largest problem for which 7 hours are needed. When the computing time is limited to 30 minutes, BMA is still able to attain the best known result for 130 out of the 135 benchmark instances. Acknowledgment We are grateful to the anonymous referees for valuable suggestions and com- ments which helped us improve the paper. This work was carried out when the first author was with the LERIA, Université d’Angers (France) and was partially supported by the Region of ‘Pays de la Loire’ (France) within the RADAPOP and LigeRO Projects. References Ahuja, R.K., Orlin J.B., & Tiwari, A. (2000). A greedy genetic algorithm for the quadratic assignment problem. Computers & Operations Research, 27(10), 917–934. Anstreicher, K. M. (2003). Recent advances in the solution of quadratic assign- ment problems. Mathematical Programming, Series B 97, 27–42. Battiti, R., & Tecchiolli, G. (1994). The reactive tabu search. ORSA Journal of Computing, 6(2), 126–140. Benlic, U., & Hao, J.K. (2011). A multilevel memetic approach for improving graph k-partitions. IEEE Transactions on Evolutionary Computation, 15(5), 624–642, 2011. Benlic, U., & Hao, J.K. (2013a). Breakout local search for maximum clique problems. Computers & Operations Research, 40(1), 192–206. Benlic, U., & Hao, J.K. (2013b). Breakout local search for the max-cut problem. Engineering Applications of Artificial Intelligence, 26(3), 1162–1173. Benlic, U., & Hao, J.K. (2013c). Breakout local search for the quadratic assign- ment problem. Applied Mathematics and Computation 219(9), 4800-4815. Benlic, U., & Hao, J.K. (2013d). Breakout local search for the vertex separator problem. In F. Rossi (Ed.) Proceedings of the 23th Intl. Joint Conference on Artificial Intelligence (IJCAI-13), IJCAI/AAAI Press, pages 461–467, Bei- jing, China, August 2013. 22 Blum, C., Puchinger, J., Raidl, G.R., & Roli, A. (2011). Hybrid metaheuristics in combinatorial optimization: A survey. Applied Soft Computing 11(6), 4135– 4151. Boese, K.D. (1995). Cost versus distance in the traveling salesman problem. Technical Report TR-950018, UCLA CS Department. Bontouxa, B., Artigues, C., & Feillet, D. (2010). A memetic algorithm with a large neighborhood crossover operator for the generalized traveling salesman problem. Computers & Operations Research, 37, 1844–1852. Burkard, R.E. (1991). Locations with spatial interactions: The quadratic assign- ment problem. In P. Mirchandani and L. Francis (Eds.), Discrete Location Theory, New York. Cheng, C.H. Ho, S.C., & Kwan, C.L. (2012). The use of meta-heuristics for airport gate assignment. Expert Systems with Applications, 39(16), 12430– 12437. Dorne, R., & Hao, J.K. (1998). A New Genetic Local Search Algorithm for Graph Coloring. In A. E. Eiben, T. Bck, M. Schoenauer, H.P. Schwefel (Eds.): PPSN 1998, Lecture Notes in Computer Science, 1498 745–754. Drezner, Z. (2003). A new genetic algorithm for the quadratic assignment prob- lem. INFORMS Journal on Computing, 15(3), 320–330. Drezner, Z. (2008). Extensive experiments with hybrid genetic algorithms for the solution of the quadratic assignment problem. Computers & Operations Research, 35(3), 717–736. Duman, E., & Or, I. (2007). The quadratic assignment problem in the con- text of the printed circuit board assembly process. Computers & Operations Research, 34(1), 163–179. Eshelman, L.J. (1991). The CHC adaptive search algorithm: How to have safe search when engaging in nontraditional genetic recombination. Proceedings of the First Workshop on Foundations of Genetic Algorithms, pages 265–283, Morgan Kaufmann. Fleurent, C., & Ferland, J.A. (1993). Genetic hybrids for the quadratic assign- ment problem. DIMACS Series in Mathematics and Theoretical Computer Science, 173–187. Garey. M., & Johnson, D. (1979). Computers and Intractability. Freeman. N.Y.. Hao, J.K. (2012). Memetic algorithms in discrete optimization. In F. Neri, C. Cotta and P. Moscato (Eds.) Handbook of Memetic Algorithms. Studies in Computational Intelligence 379, Chapter 6, pages 73–94. Hart, W.E., Krasnogor, N., & Smith, J.E. (2004). Recent advances in memetic algorithms. Studies in Fuzziness and Soft Computing 166, Springer. 23 Hassin, M., Levin, A., & Sviridenko, M. (2009). Approximating the minimum quadratic assignment problems. ACM Transactions on Algorithms, 6(1), ar- ticle number 18. James, T. Rego, C., & Glover, F. (2009a). A cooperative parallel tabu search algorithm for the quadratic assignment problem. European Journal of Oper- ational Research, 195(3), 810–826. James, T. Rego, C., & Glover, F. (2009b). Multistart tabu search and diversi- fication strategies for the quadratic assignment problem. IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and Humans, 39(3), 579–596. Jones, T., & S. Forrest, S. (1995). Fitness distance correlation as a measure of problem difficulty for genetic algorithms. Proceedings of the 6th International Conference on Genetic Algorithms, Morgan Kaufmann, 184–192. Kelly, J.P., Laguna, M., & Glover, F. (1994). A study of diversification strategies for the quadratic assignment problem. Computers & Operations Research, 21(8), 885–893. Krasnogor, N., & Smith, J. (2005). A tutorial for competent memetic algo- rithms: model, taxonomy and design issues. IEEE Transactions on Evolu- tionary Computing, 9(5), 474–488. Li, F., Xu, L.D., Jin, C., & Wang, H. (2012). Random assignment method based on genetic algorithms and its application in resource allocation. Expert Systems with Applications, 39(15), 12213–12219. Lü, Z., & Hao, J.K. (2010). A memetic algorithm for graph coloring. European Journal of Operational Research, 203(1), 241–250. Merz, P., & Freisleben, B. (2000). Fitness landscape analysis and memetic algorithms for the quadratic assignment problem. IEEE Transactions on Evolutionary Computation, 4(4), 337–352. Miao, Z., Cai, S., & Xu. D. (2014). Applying an adaptive tabu search algorithm to optimize truck-dock assignment in the crossdock management system. Ex- pert Systems with Applications, 41(1), 16–22. Misevicius, A. (2004). An improved hybrid genetic algorithm: New results for the quadratic assignment problem. Knowledge Based Systems, 17(2-4), 65–73. Misevicius, A., & Kilda, B. (2006). Iterated tabu search: An improvement to standard tabu search. Information Technology and Control, 35(3), 187–197. Moscato, P. (1989). On evolution, search, optimization, genetic algorithms and martial arts: Towards memetic algorithms. Caltech Concurrent Computation Program, Report 826, California Institute of Technology, Pasadena, Califor- nia, USA. 24 Moscato, P., & Cotta, C. (2003). A Gentle Introduction to memetic algorithms. In F. Glover and G. A. Kochenberger (Eds.), Handbook of Metaheuristic. Kluwer, Norwell, Massachusetts, USA. Neri, F., Cotta, C., & Moscato, P. (Eds.) (2012). Handbook of Memetic Algo- rithms. Studies in Computational Intelligence 379, Springer. Nikolić, M., & Teodorović, D. (2014). A simultaneous transit network design and frequency setting: computing with bees. Expert Systems with Applications, 41(16), 7200–7209. Pardalos, P.M., Rendl, F., & Wolkowicz, H. (1994). The quadratic assignment problem: A survey and recent developments. In Proceedings of the DIMACS Workshop on Quadratic Assignment Problems, volume 16 of DIMACS Series in Discrete Mathematics and Theoretical Computer Science, 1–42. Porumbel, D.C., Hao, J.K., & Kuntz, P. (2010). An evolutionary approach with diversity guarantee and well-informed grouping recombination for graph coloring. Computers & Operations Research, 37(10), 1822–1832. Reeves, C.R. (1997). Landscapes, operators and heuristic search. Annals of Operations Research, 86, 473–490. Skorin-Kapov, J. (1990). Tabu search applied to the quadratic assignment prob- lem. ORSA Journal on Computing, 2, 33–45. Stützle, T. (2006). Iterated local search for the quadratic assignment problem. European Journal of Operational Research, 174(3), 1519–1539. Taillard, E. (1991). Robust taboo search for the quadratic assignment problem. Parallel Computing, 17, 443-455. Wilhelm, M.R., & Ward, T.L. (1987). Solving quadratic assignment problems by simulated annealing. IIE Transactions, 19(1), 107–119. Appendix This appendix includes two tables (Tables 7 and 8). Table 7 shows the best, average and worst result obtained by BMA after 10 independent runs for 21 selected QAPLIB instances, within a time limit of 2 hours (10 hours for instance tho150) and 30 minutes respectively. Table 8 shows the computational results of our BMA algorithm on the set of 114 easy QAP instances of the QAPLIB. For all these instances, each run of BMA can attain the best known result with a computing time ranging from 0 second to at most 2.5 minutes. 25 Table 7: Computational results of the proposed BMA algorithm with a time limit of 2 hours and 30 minutes respectively on the set of 21 hard instances from the QAPLIB. Columns ‘% δ̄best’, ‘% δ̄avg’ and ‘% δ̄worst’ show respectively the percentage deviation of our best, average, and worst solution from the best-known solution (BKR) over all the trials; Column ‘tavg(min)’ indicates the average computing time in minutes when the best solution in each trial was attained. Problem Time limit of 2 hours Time limit of 30 minutes Name BKS % δ̄best % δ̄avg % δ̄worst tavg(m) % δ̄best % δ̄avg % δ̄worst tavg(m) tai40a 3139370 0.000(2) 0.059 0.074 8.1 0.000(1) 0.067 0.074 3.7 tai50a 4938796 0.000(2) 0.131 0.291 42.0 0.053(0) 0.264 0.409 3.8 tai60a 7205962 0.000(2) 0.144 0.259 67.5 0.165(0) 0.307 0.369 13.9 tai80a 13499184 0.246(0) 0.426 0.561 65.8 0.430(0) 0.563 0.670 16.4 tai100a 21052466 0.269(0) 0.405 0.550 44.1 0.340(0) 0.582 0.715 11.2 tai50b 458821517 0.000(10) 0.000 0.000 1.2 0.000(10) 0.000 0.000 1.2 tai60b 608215054 0.000(10) 0.000 0.000 5.2 0.000(10) 0.000 0.000 5.2 tai80b 818415043 0.000(10) 0.000 0.000 31.3 0.000(9) 0.028 0.283 3.1 tai100b 1185996137 0.000(10) 0.000 0.000 13.6 0.000(8) 0.028 0.100 7.9 tai150b 498896643 0.000(1) 0.060 0.365 78.1 0.000(1) 0.198 0.225 15.0 sko72 66256 0.000(10) 0.000 0.000 3.5 0.000(8) 0.012 0.063 1.3 sko81 90998 0.000(10) 0.000 0.000 4.3 0.000(10) 0.000 0.000 4.5 sko90 115534 0.000(10) 0.000 0.000 15.3 0.000(7) 0.025 0.123 4.43 sko100a 152002 0.000(10) 0.000 0.000 22.3 0.000(7) 0.005 0.016 9.8 sko100b 153890 0.000(10) 0.000 0.000 56.5 0.000(7) 0.001 0.004 6.75 sko100c 147862 0.000(10) 0.000 0.000 12.0 0.000(10) 0.000 0.000 8.9 sko100d 149576 0.000(9) 0.006 0.065 20.9 0.000(6) 0.024 0.176 11.45 sko100e 149150 0.000(10) 0.000 0.000 11.9 0.000(10) 0.000 0.000 9.93 sko100f 149036 0.000(10) 0.000 0.000 23.0 0.000(7) 0.001 0.005 10.6 wil100 273038 0.000(10) 0.000 0.000 14.5 0.000(10) 0.000 0.000 7.7 tho150 8133398 0.000(3) 0.007 0.064 416.4 0.019(0) 0.057 0.096 22.4 26 Table 8: Computational results of our BMA algorithm for the set of 114 easy instances of the QAPLIB. Column ‘BKR’ indicates the optimal/best-known result ever reported in the literature. Column ‘#best′ indicates the number of times a best-known solution was found over 100 executions; Column ‘% δ̄avg’ shows the percentage deviation of our average solution from the best-known solution (BKR) over all the trials; Column ‘tavg(sec)’ indicates the average computing time in seconds when the best solution in each trial was attained. Instance BKR #best % δ̄avg tavg(sec) Instance BKR #best % δ̄avg tavg(sec) bur26a 5426670 100/100 0.000 0.8 tai20a 703482 100/100 0.000 0.0 bur26b 3817852 100/100 0.000 0.8 tai25a 1167256 100/100 0.000 0.0 bur26c 5426795 100/100 0.000 0.5 tai30a 1818146 100/100 0.000 0.0 bur26d 3821225 100/100 0.000 0.2 tai35a 242002 100/100 0.000 0.1 bur26e 5386879 100/100 0.000 0.2 rou15 354210 100/100 0.000 0.4 bur26f 3782044 100/100 0.000 0.2 rou20 725522 100/100 0.000 0.5 bur26g 10117172 100/100 0.000 0.3 lipa20a 3683 100/100 0.000 0.1 bur26h 7098658 100/100 0.000 0.2 lipa20b 27076 100/100 0.000 0.0 tai64c 1855928 100/100 0.000 2.4 lipa30a 13178 100/100 0.000 0.2 chr12a 9552 100/100 0.000 0.2 lipa30b 151426 100/100 0.000 0.1 chr12b 9742 100/100 0.000 0.2 lipa40a 31538 100/100 0.000 1.1 chr12c 11156 100/100 0.000 0.2 lipa40b 476581 100/100 0.000 0.3 chr15a 9896 100/100 0.000 0.3 lipa50a 62093 100/100 0.000 1.5 chr15b 7990 100/100 0.000 0.4 lipa50b 1210244 100/100 0.000 0.4 chr15c 9504 100/100 0.000 0.3 lipa60a 107218 100/100 0.000 8.2 chr18a 11098 100/100 0.000 0.4 lipa60b 2520135 100/100 0.000 0.4 chr18b 1534 100/100 0.000 0.5 lipa70a 169755 100/100 0.000 30.7 chr20a 2192 100/100 0.000 2.1 lipa70b 4603200 100/100 0.000 1.5 chr20b 2298 100/100 0.000 5.7 lipa80a 253195 100/100 0.000 51.2 chr20c 14142 100/100 0.000 0.8 lipa80b 7763962 100/100 0.000 4.1 chr22a 6156 100/100 0.000 1.6 lipa90a 360630 100/100 0.000 149.1 chr22b 6194 100/100 0.000 1.8 lipa90b 12490441 100/100 0.000 4.2 chr25a 3796 100/100 0.000 9.5 tai10a 135028 100/100 0.000 0.0 els19 17212548 100/100 0.000 0.4 tai12a 224416 100/100 0.000 0.1 esc16a 68 100/100 0.000 0.0 tai15a 388214 100/100 0.000 0.2 esc16b 292 100/100 0.000 0.0 tai17a 491812 100/100 0.000 0.4 esc16c 160 100/100 0.000 0.0 nug12 578 100/100 0.000 0.1 esc16d 16 100/100 0.000 0.0 nug14 1014 100/100 0.000 0.1 esc16e 28 100/100 0.000 0.0 nug15 1150 100/100 0.000 0.2 esc16f 0 100/100 0.000 0.1 nug16a 1610 100/100 0.000 0.2 esc16g 26 100/100 0.000 0.1 nug16b 1240 100/100 0.000 0.3 esc16h 996 100/100 0.000 0.0 nug17 1732 100/100 0.000 0.3 esc16i 14 100/100 0.000 0.0 nug18 1930 100/100 0.000 0.3 esc16j 8 100/100 0.000 0.0 nug20 2570 100/100 0.000 0.4 esc32a 130 100/100 0.000 0.0 nug21 2438 100/100 0.000 0.4 esc32b 168 100/100 0.000 0.1 nug22 3596 100/100 0.000 0.4 esc32c 642 100/100 0.000 0.1 nug24 3488 100/100 0.000 0.6 esc32d 200 100/100 0.000 0.1 nug25 3744 100/100 0.000 0.6 esc32e 2 100/100 0.000 0.1 nug27 5234 100/100 0.000 0.5 esc32g 6 100/100 0.000 0.0 nug28 5166 100/100 0.000 0.7 esc32h 438 100/100 0.000 0.1 nug30 6124 100/100 0.000 1.8 esc64a 116 100/100 0.000 0.2 scr12 31410 100/100 0.000 0.1 esc128 64 100/100 0.000 1.3 scr15 51140 100/100 0.000 0.3 had12 1652 100/100 0.000 0.2 scr20 110030 100/100 0.000 0.5 had14 2724 100/100 0.000 0.1 tho30 149936 100/100 0.000 0.3 had16 3720 100/100 0.000 0.3 tho40 240516 100/100 0.000 0.4 had18 5358 100/100 0.000 0.2 wil50 48816 100/100 0.000 23.6 had20 6922 100/100 0.000 0.3 tai10b 1183760 100/100 0.000 0.0 kra30a 88900 100/100 0.000 2.4 tai12b 39464925 100/100 0.000 0.0 kra30b 91420 100/100 0.000 1.2 tai15b 51765268 100/100 0.000 0.1 kra32 88700 100/100 0.000 1.1 tai30b 637117113 100/100 0.000 0.0 ste36a 9526 100/100 0.000 4.9 tai35b 283315445 100/100 0.000 0.0 ste36b 15852 100/100 0.000 1.3 tai40b 637250948 100/100 0.000 0.2 ste36c 8239110 100/100 0.000 2.6 sko42 15812 100/100 0.000 1.3 rou12 235528 100/100 0.000 0.2 sko49 23386 100/100 0.000 0.5 tai20b 122455319 100/100 0.000 0.0 sko56 34458 100/100 0.000 1.0 tai25b 344355646 100/100 0.000 0.0 sko64 48498 100/100 0.000 1.5 27 
work_shveq3cihjehzguqdy6ovwfmcu	----	This is the author’s version of a work that was submitted/accepted for pub- lication in the following source: Yigitcanlar, Tan (2014) Empirical approaches in knowledge city research. Expert Systems with Applications, 41(12), pp. 5547-5548. This file was downloaded from: http://eprints.qut.edu.au/67390/ c© Copyright 2014 Elsevier Ltd. NOTICE: this is the author’s version of a work that was accepted for pub- lication in Expert Systems with Applications. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Expert Systems with Applications, Volume 41, Issue 12, (15 September 2014) DOI: 10.1016/j.eswa.2014.02.005 Notice: Changes introduced as a result of publishing processes such as copy-editing and formatting may not be reflected in this document. For a definitive version of this work, please refer to the published source: http://doi.org/10.1016/j.eswa.2014.02.005 http://eprints.qut.edu.au/view/person/Yigitcanlar,_Tan.html http://eprints.qut.edu.au/67390/ http://doi.org/10.1016/j.eswa.2014.02.005 1 Empirical Approaches in Knowledge City Research In the 21st century, it has become apparent that ‘knowledge’ is a major factor of postmodern production (Yigitcanlar et al., 2007). Beyond this, in today’s rapidly globalizing world, knowledge, along with the social and technological settings, is seen as a key to secure economic prosperity and quality of life (Yigitcanlar et al., 2008a). However, limiting the benefits of a ‘knowledge-based development’ to only economic gains—and to a degree to social ones—is quite a narrow sighted view (Yigitcanlar et al., 2008b). Thus, the concept of ‘knowledge-based urban development’ is coined to bring economic prosperity, environmental sustainability, a just socio-spatial order and good governance to cities, and as a result producing a purposefully designed city—i.e., ‘knowledge city’—generating positive environmental and governance outcomes as well as economic and societal ones (Yigitcanlar, 2011; Carrillo et al., 2014). The competitive nature of the global knowledge economy pushes cities to invest on their knowledge edges. This ultimately means almost all cities face the challenge of developing as knowledge cities (Yigitcanlar, 2010). Hence, many cities worldwide have already undergone major transformations in the 21st century, an era in which the role of knowledge in economic, social, environmental and governance issues becomes highly critical for city competitiveness (Yigitcanlar, 2009). Presently, urban administrations, policymakers and planners are in need of new approaches to harness the considerable opportunities of knowledge-based urban development for knowledge city transformation (Lonnqvist et al., 2014). Within cities, concepts of knowledge city and knowledge-based urban development have become emerging areas of research interest, which links interests of planners, economists, geographers, and environmental, social and other scientists (Yigitcanlar et al., 2012). Despite this growing interest, there has been limited empirical research conducted in the field (Yigitcanlar & Lonnqvist, 2013). Therefore, there is an urgent need for comprehensive research in this area. For this reason, the Special Issue focuses on investigating knowledge cities. It aims at exemplifying empirical approaches in knowledge city research to generate new insights, and potential avenues and directions for building prosperous knowledge cities across the globe. This Special Issue of the Expert Systems with Applications contains the full versions of a selection of best papers that were presented at the 6th Knowledge Cities World Summit (KCWS-2013) held from September 9-12, 2013 in Istanbul, Turkey, which is co-organized by the World Capital Institute and Istanbul Sehir University. This Special Issue consists of the following ten papers focusing on complementary aspects of the empirical knowledge city investigation. The guest editor has selectively identified and invited the authors to re-write and extend their short articles for re-submission to the Special Issue. These extended versions underwent a second round of independent double-blind and editorial review before final decision was made. The papers contained in this Special Issue represent the latest research in the field. Collectively I hope this collection of papers, which provide rich and diverse perspectives on the topic, will function as an acknowledgment of the centrality of empirical knowledge city research for our cities to achieve a thriving knowledge-based urban development, bridge the research gap, and shed light on the new empirical research directions in the field. 1) Benchmarking the performance of global and emerging knowledge cities 2) Knowledge-city index construction: an intellectual capital perspective 3) Regional evolution and waves of growth: a knowledge-based perspective 2 4) Knowledge patterns and spatial dynamics of industrial districts in knowledge cities: Hsinchu, Taiwan 5) Innovation quality in knowledge cities: empirical evidence of innovation award competitions in Finland 6) Migrant knowledge workers: an empirical study of global Sydney as a knowledge city 7) Capital system, creative economy and knowledge city transformation: insights from Bento Gonçalves, Brazil 8) Algorithm-embedded IT applications for an emerging knowledge city: Istanbul, Turkey 9) Network-based innovation systems: a capital base for the Monterrey city-region, Mexico 10) Knowledge spaces and places: from the perspective of a “born-global” start-up in the field of urban technology Following this brief editorial introduction, the Special Issue starts with a Position Paper by Tan Yigitcanlar (Paper 1: Benchmarking the performance of global and emerging knowledge cities). This paper advocates the essential role of city benchmarking to achieve a prosperous knowledge-based urban development and knowledge city transformation. The paper first introduces the methodology of a novel performance assessment model—the Knowledge- Based Urban Development Assessment Model (KBUD/AM). It then highlights lessons from the application of the model in an international knowledge city performance analysis study conducted with 11 global and emerging knowledge cities. Next in Paper 2, Víctor-Raúl López-Ruiz, José-Luis Alfaro-Navarro and Domingo Nevado-Peña (Knowledge-city index construction: an intellectual capital perspective) proceed on the same topic of knowledge city performance assessment, and first classify and review popular city indices, and then propose a new indexing methodology—the Knowledge City Index (KCI)—to measure intangible capital as the growth capacity of knowledge cities. Authors discuss the results of the application of the methodology in the European context. Paper 3 by Robert Huggins, Hiro Izushi, Daniel Prokop and Piers Thompson (Regional evolution and waves of growth: a knowledge-based perspective) offers an analysis of the major world regions according to a set of key economic evolution trends—i.e., the fifth wave growth, the third & fourth wave growth, and government-led third wave growth—by presenting a novel empirical approach. In doing so, the authors aim to uncover the underlying structure of the changes in knowledge-based resources, capabilities and outputs across the major knowledge city-regions of the world. In Paper 4 (Knowledge patterns and spatial dynamics of industrial districts in knowledge cities: Hsinchu, Taiwan), Hung-Nien Hsieh, Tai-Shan Hu, Ping-Ching Chia and Chieh-Chung Liu focus on the development of clusters of knowledge-based corporations that has become an important strategic factor in increasing the competitiveness of knowledge cities. This paper scrutinizes a knowledge city-region—Hsinchu, Taiwan—by focusing on the correlation between the spatial dynamics of knowledge in major industries and innovation based on empirical data. In turn, Teemu Makkonen and Tommi Inkinen in Paper 5 (Innovation quality in knowledge cities: empirical evidence of innovation award competitions in Finland) deal with the innovation awards as a supporting tool in innovation policy of knowledge cities. The paper searches for firm-level evidence by surveying innovation award winning companies from Finland. The research indicates that innovation awards are an additional tool for innovation promotion, alongside innovation inducement policies and an additional indicator of innovation quality in the context of knowledge cities. 3 Richard Hu (Paper 6: Migrant knowledge workers: an empirical study of global Sydney as a knowledge city) looks at knowledge city formation from the angle of creative class of knowledge workers. The paper reveals insightful findings of a study on migrant knowledge workers employed in the knowledge-intensive industries and highly skilled occupations in Sydney, Australia. The research provides a better understanding of global Sydney as a knowledge city and its relationship with migration. Ana Cristina Fachinelli, Francisco Javier Carrillo and Anelise D’Arisbo, with Paper 7 (Capital system, creative economy and knowledge city transformation: insights from Bento Gonçalves, Brazil), focus on the points of conceptual convergence between the generic capital system taxonomy and creative economy, and the potential value of this convergence for the development of an emerging knowledge city. Authors, within this perspective, investigate the case of Bento Gonçalves, Brazil. Next, Melih Bulu, Muhammed Ali Önder and Vural Aksakalli offer Paper 8 (Algorithm- embedded IT applications for an emerging knowledge city: Istanbul, Turkey). This paper advocates the role of infrastructural resources in the knowledge city transformation— particularly algorithm-embedded information technology applications. In the case of an emerging knowledge city—i.e., Istanbul, Turkey—authors classify and investigate algorithm- embedded information technology application areas related to management of infrastructural resources of the city in a systematic way. Then, in Paper 9 (Network-based innovation systems: a capital base for the Monterrey city-region, Mexico) Blanca Garcia and Danilo Chavez aim to shed light on the research question of how knowledge-intensive clustering communities, such as higher education institutions and other local actors, are building their communities in the knowledge-based arena. In order to do so, they investigate the City of Monterrey as the case study in the context of Mexico-Texas borderland with the help of regional innovation systems perspective. Lastly, Paper 10, (Knowledge spaces and places: from the perspective of a “born-global” start-up in the field of urban technology) by Luís Carvalho, Inês Plácido Santos and Willem van Winden, analyzes how a global start-up firm, namely Living PlanIT, is linked to different types of places, and how it explores and exploits territorial innovation potentials. By doing so, it illustrates how the interaction with different knowledge milieus provides unique resources for the technology development, commercialization and societal legitimation for the case study. I wish to thank the authors of the abovementioned papers for accepting the invitation and submitting and revising their manuscripts within a short time frame, and thank the referees for their thorough and timely reviews. Finally, I want to thank Founding Editor Professor Jay Liebowitz and Editor-in-Chief Professor Binshan Lin of the Expert Systems with Applications for inviting me to edit this Special Issue. Guest Editor Tan Yigitcanlar School of Civil Engineering and Built Environment Queensland University of Technology 2 George Street, Brisbane, QLD 4001, Australia Tel: +61.7.3138.2418 E-mail: tan.yigitcanlar@qut.edu.au mailto:tan.yigitcanlar@qut.edu.au 4 References Carrillo, J., Yigitcanlar, T., Garcia, B., & Lonnqvist, A. (2014). Knowledge and the city: concepts, applications and trends of knowledge-based urban development. Washington, DC: Routledge. Lonnqvist, A., Kapyla, J., Salonius, H., & Yigitcanlar, T. (2014). Knowledge that matters: identifying regional knowledge assets of Tampere Region, European Planning Studies, http://dx.doi.org/10.1080/09654313.2013.814621. Yigitcanlar, T., Baum, S., & Horton, S. (2007). Attracting and retaining knowledge workers in knowledge cities, J. Knowledge Management 11(5): 6–17. Yigitcanlar, T., O’Connor, K., & Westerman, C. (2008a). The making of knowledge cities: Melbourne’s knowledge-based urban development experience, Cities 25(2): 63-72. Yigitcanlar, T., Velibeyoglu, K., & Martinez-Fernandez, C. (2008b). Rising knowledge cities: the role of knowledge precincts. J. Knowledge Management 12(5), 8-20. Yigitcanlar, T. (2009). Planning for knowledge-based development: global perspectives, J. Knowledge Management 13(5): 228-242. Yigitcanlar, T. (2010). Making space end place for the knowledge economy: knowledge- based development of Australian cities, European Planning Studies 18(11): 1769- 1786. Yigitcanlar, T. (2011). Position paper: redefining knowledge-based urban development, Int. J. Knowledge Based Development 2(4): 340-356. Yigitcanlar, T., Metaxiotis, K., & Carrillo, J. (2012). Building prosperous knowledge cities: policies, plans and metrics. Cheltenham, UK: Edward Elgar. Yigitcanlar, T., & Lonnqvist, A. (2013). Benchmarking knowledge-based urban development performance: results from the international comparison of Helsinki, Cities 31(1): 357- 369. 
work_skrvzqpikvdw5f7sksbktmh32u	----	doi:10.1016/j.eswa.2005.10.007 Artificial neural networks with evolutionary instance selection for financial forecasting Kyoung-jae Kim * Department of Information Systems, Dongguk University, 3-26, Pil-dong, Chung-gu, Seoul 100-715, South Korea Abstract In this paper, I propose a genetic algorithm (GA) approach to instance selection in artificial neural networks (ANNs) for financial data mining. ANN has preeminent learning ability, but often exhibit inconsistent and unpredictable performance for noisy data. In addition, it may not be possible to train ANN or the training task cannot be effectively carried out without data reduction when the amount of data is so large. In this paper, the GA optimizes simultaneously the connection weights between layers and a selection task for relevant instances. The globally evolved weights mitigate the well-known limitations of gradient descent algorithm. In addition, genetically selected instances shorten the learning time and enhance prediction performance. This study applies the proposed model to stock market analysis. Experimental results show that the GA approach is a promising method for instance selection in ANN. q 2005 Elsevier Ltd. All rights reserved. Keywords: Instance selection; Genetic algorithms; Artificial neural networks; Financial forecasting 1. Introduction In general, artificial neural networks (ANNs) can produce robust performance when a large amount of data is available. However, ANN often exhibits inconsistent and unpredictable performance on noisy data. In addition, it may not be possible to train ANN or the training task cannot be effectively carried out without data reduction when a data set is too huge. Data reduction can be achieved in many ways such as feature selection or feature discretization (Blum & Langley, 1997; Kim & Han, 2000; Liu & Motoda, 1998). One facet of data mining concerns the selection of relevant instances for this reason. Instances are a collection of training examples in supervised learning and instance selection chooses a part of the data that is representative and relevant to the characteristics of all the data. Instance selection is one of popular methods for dimensionality reduction and is directly related to data reduction. Although instance selection is the most complex form of data reduction because the computa- tionally expensive prediction methods must be invoked more often to determine the effectiveness of instance selection, we can usually remove irrelevant instances as well as noise and 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.10.007 * Tel.: C82 2 2260 3324; fax: C82 2 2260 3684. E-mail address: kjkim@dongguk.edu redundant data (Liu & Motoda, 2001; Weiss & Indurkhya, 1998). Many researchers have suggested instance selection methods such as squashed data, critical points, prototype construction, in addition to many forms of sampling (Liu & Motoda, 2001). The efforts to select relevant instances from an initial data set have stemmed from the need to reduce immense storage requirements and computational loads (Kuncheva, 1995). The other perspective on this subject, as pointed out in Dasarathy (1990), is to achieve enhanced performance from the learning algorithm through instance selection. In addition, training time may be shortened by use of the proper instance selection algorithm. This paper proposes a new hybrid model of ANN and genetic algorithms (GAs) for instance selection. An evolution- ary instance selection algorithm reduces the dimensionality of data and may eliminate noisy and irrelevant instances. In addition, this study simultaneously searches the connection weights between layers in ANN through an evolutionary search. The genetically evolved connection weights mitigate the well-known limitations of gradient descent algorithm. The rest of this paper is organized as follows: Section 2 presents the research background. Section 3 proposes the evolutionary instance selection algorithm and describes the benefits of the proposed algorithm. Section 4 describes the application of the proposed algorithm. Conclusions and the limitations of this study are presented in Section 5. Expert Systems with Applications 30 (2006) 519–526 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa K.-j. Kim / Expert Systems with Applications 30 (2006) 519–526520 2. Research background For some applications, quality of data mining is improved with additional instances. However, the number of instances may tend to increase the complexity of induced solution. Increased complexity is not desirable, but may be the price to pay for better performance. In addition, increased complexity decreases the interpretability of the result (Weiss & Indurkhya, 1998). In this sense, many researchers have suggested instance selection methods. The following sections present some instance selection methods as described by prior research. 2.1. Instance selection methods Instance-based learning algorithms often faced the problem of deciding which instances to store for use during generalization in order to avoid excessive storage and time complexity, and to improve generalizability by avoiding noise and overfitting (Wilson & Martinez, 2000). Many researchers have addressed the problem of training data reduction and have presented algorithms for maintaining an instance base or case base in instance-based learning algorithms. Kuncheva (1993) classified instance selection techniques (or editing techniques) into the following three categories: Condensed Nearest Neighbor rule, Generated or Modified Prototypes, and Two-Level Classifiers. The following presents some basic concepts of each category as described by prior research. A detailed explanation may be found in the references of this paper. 2.1.1. Condensed nearest neighbor rule Hart (1968) made one of the first attempts to develop an instance selection rule. Hart’s algorithm, the Condensed Nearest Neighbor rule, finds a subset S of the training set T such that every member of T is closer to a member of S of the same class than to a member of S of a different class. Subsequent work extended Hart’s algorithm, specifically the Selective Nearest Neighbor rule (Ritter, Woodruff, Lowry, & Isenhour, 1975) and the Reduced Nearest Neighbor rule (Gates, 1972). In addition, Wilson (1972) introduced the Edited Nearest Neighbor algorithm and Tomek (1976) proposed the All k-NN method of editing. 2.1.2. Generated or modified prototypes This category is composed of techniques that establish new prototypes or adjust a limited number of instances. A large group of studies within this category are implemented by ANN including feature-map classifiers, learning vector quantiziers (Kuncheva, 1995). 2.1.3. Two-level classifiers This category employs two or more classifiers and allocates a part of all instances to the classifier, which appears most appropriate. Tetko and Villa (1997) proposed the Efficient Partition algorithm, which is used to obtain an efficient partition of noisy instances, whose distribution is proportional to the complexity of the analyzed function. This is to focus the training of ANN on the most complex and informative domains of the data set and accelerate the learning phase. They concluded that the efficiently partitioned instances enhance the predictability of ANN in comparison with a random selection of instances. Oh and Han (2000) proposed the integrated neural network model using change-point detection. They partitioned instances according to each detected change- point, and then applied each partitioned instance to each ANN of multiple ANN. Instance selection in instance-based learning algorithms may be considered as a method of knowledge refinement and it maintains the instance-base. In this sense, some researchers proposed many instance selection algorithms for maintaining the case-base in case-based reasoning (CBR) systems. Smyth (1998) presented an approach to maintenance, which is based on the deletion of harmful and redundant cases from the case- base. In addition, McSherry (2000) suggested an instance selection method in the construction of a case library in which evaluation of the coverage contributions of candidate instances are based on an algorithm called disCover. This algorithm reverses the direction of CBR to discover all cases that can be solved with a given case-base. Although many different approaches have been used to address the problem of case authoring and data explosion for instance-based algorithms, there is little research on instance selection in ANN. Reeves and Taylor (1998) suggested that a GA is a promising approach to finding ‘better’ training data set for classification problems in radial basis function (RBF) nets. Reeves and Bush (2001) reported that the GA can also be used effectively to find a smaller subset of a ‘good’ training set in RBF nets for both classification and regression problems. Although, the GA has been shown to be a promising instance selection method for RBF nets, its performances on other neural network models are untested. 2.2. Genetic algorithms The GA has been investigated recently and shown to be effective in exploring a complex space in an adaptive way, guided by the biological evolution mechanisms of selection, crossover, and mutation (Adeli & Hung, 1995). The GA simulates the mechanics of population genetics by maintaining a population of knowledge structure, which is made to evolve (Odetayo, 1995). The problems must be represented in a suitable form to be handled by the GA. The GA often works with a form of binary coding. If the problems are coded as chromosomes, the population is initialized. Each chromosome within the population is gradually evolved by biological operations. Once the population size is chosen, the initial population is randomly generated (Bauer, 1994). After the initialization step, each chromosome is evaluated by the fitness function. According to the value of the fitness function, the chromo- somes associated with the fittest individuals will be reproduced more often than those associated unfit individuals (Davis, 1994). K.-j. Kim / Expert Systems with Applications 30 (2006) 519–526 521 The GA works with three operators that are iteratively used. The selection operator determines which individuals may survive (Hertz & Kobler, 2000). The crossover operator allows the search to fan out in diverse directions looking for attractive solutions and permits chromosomal material from different parents to be combined in a single child. In addition, the mutation operator arbitrarily alters one or more components of a selected chromosome. It provides the means for introducing new information into the population. Finally, the GA tends to converge on optimal or near-optimal solutions (Wong & Tan, 1994). The GA is usually employed to improve the performance of artificial intelligence techniques. For ANN, the GA was applied to the selection of neural network topology including optimizing a relevant feature subset, determining the optimal number of hidden layers and processing elements. In addition, some researchers searched the connection weights of ANN using the GA instead of local search algorithms including a gradient descent algorithm. They suggested that global search techniques including the GA might prevent ANN from falling into a local optimum (Gupta & Sexton, 1999; Kim & Han, 2000; Sexton, Dorsey, & Johnson, 1998). 2.3. Prior research on stock market prediction using ANN Many studies on stock market prediction using artificial intelligence (AI) techniques have been performed during the past decade. The early days of these studies focused on estimating the level of the return on stock price index. One of the earliest studies, Kimoto, Asakawa, Yoda, and Takeoka (1990) used several learning algorithms and prediction methods for developing a prediction system for the Tokyo Stock Exchange Prices Index. They used the modular neural network to learn the relationships among various market factors. They concluded that the correlation coefficient produced by their model is much higher than that produced by multiple regression. However, the correlation coefficient may not be a proper measure for prediction performance. Kamijo and Tanikawa (1990) used the recurrent neural network for analyzing candlestick charts. A candlestick chart is a Japanese style chart used to visualize stock price patterns. In these studies, they did not perform any statistical test for the significance of the empirical results. Some researchers investigated the issue of predicting the stock index futures market. Choi, Lee and Lee (1995) and Trippi and DeSieno (1992) predicted the daily direction of change in the S&P 500 index futures using ANN. Trippi and DeSieno (1992) combined the outputs of individual networks using logical (Boolean) operators to produce a set of composite rules. They suggested that their best composite synthesized rule set system achieved a higher gain than previous research. Choi et al. (1995) compared their approach with previous study and suggested that they earned a higher annualized gain than the previous study. However, the annualized gain may not be an appropriate measure for prediction performance because it varies according to the fee for trade and the trading strategy. Duke and Long (1993) predicted German government daily bond futures using backpropagation (BP) neural networks. They reported that the 53.94% of the patterns are accurately predicted through the moving simulation method. Most of the above studies simply applied ANN to stock market prediction. Recent research tends to hybridize several AI techniques. Nikolopoulos and Fellrath (1994) developed a hybrid expert system for investment advising. In their study, genetic algorithms were used to train and configure the architecture of investor’s neural network component. Hiemstra (1995) proposed fuzzy expert systems to predict stock market returns. He suggested that ANN and fuzzy logic could capture the complexities of functional mapping because they do not require the specification of the function to approximate. Some researchers tend to include novel factors for the learning process. Kohara, Ishikawa, Fukuhara and Nakamura (1997) incorporated prior knowledge to improve the performance of stock market prediction. Prior knowledge in their study included non-numerical factors such as political and inter- national events. They made use of prior knowledge of stock price predictions and newspaper information on domestic and foreign events. A more recent study of Lee and Jo (1999) developed an expert system, which uses knowledge in a candlestick chart analysis. The expert system had patterns and rules, which could predict future stock price movements. The experimental results revealed that a developed knowledge-base could provide excellent indicators. In addition, Tsaih, Hsu and Lai (1998) integrated a rule-based technique and ANN to predict the direction of change of the S&P 500 stock index futures on a daily basis. Stock market data, however, includes tremendous noise and non-stationary characteristics; thus, the training process for ANN tends to be difficult. In addition, the possibility of local convergence of the gradient search techniques may be another difficulty for learning patterns. 3. A GA approach to instance selection for ANN As mentioned earlier, there are many studies on instance selection for the instance-based learning algorithm. However, there are few studies on instance selection for ANN. Thus, there are few relevant theories concerning instance selection for ANN. This paper proposes the GA approach to instance selection for ANN (GAIS). The overall framework of GAIS is shown in Fig. 1. In this study, the GA supports the simultaneous optimization of connection weights and selection of relevant instances. The algorithm of GAIS consists of the following three phases: GA search phase, feed-forward computation phase, and validation phase. 3.1. GA search phase In the GA search phase, the GA searches the search space to find optimal or near-optimal connection weights and relevant instances for ANN. The populations, the connection weights and the codes for instance selection, are initialized into random values before the search process. The parameters for searching X1 X2 X3 X4 I1 I2 I3 I4 I5 I6 I7 I8 I9 I10 X1 X2 X3 X4 I3 I6 I7 I10 GA Fitness function ANN Assign input connection weights Assign hidden connection weights Evaluation S el ec t re le va nt i ns ta nc es Fig. 1. Overall framework of GAIS. K.-j. Kim / Expert Systems with Applications 30 (2006) 519–526522 must be encoded on chromosomes. This study needs three sets of parameters. The first set is the set of connection weights between the input layer and the hidden layer of the network. The second set is the set of connection weights between the hidden layer and the output layer. As mentioned earlier, the above two sets may mitigate the limitation of the gradient descent algorithm. The third set represents the codes for instance selection. The strings have the following encoding: each processing element in the hidden layer receives signals from the input layer. The first set of bits represents the connection weights between the input layer and the hidden layer. Each processing element in the output layer receives signals from the hidden layer. The next set of bits indicates the connection weights between the hidden layer and the output layer. The following bits are instance selection codes for the training data. The parameters to be searched use only the information about the selected instances within the training data. In this phase, the GA operates the process of crossover and mutation on initial chromosomes and iterates until the stopping conditions are satisfied. 3.2. Feed-forward computation phase This phase is the process of feed-forward computation in ANN. Proper activation function is required to facilitate the learning process. However, there are no clear criteria regarding which activation function to use. Some researchers rec- ommended the sigmoid function for classification problems and the hyperbolic tangent function for forecasting problems because of the difference between the sigmoid and the hyperbolic tangent function for the value range of delta weights with the SSE error function (Coakley & Brown, 2000). In addition, the majority of back-propagation applications used the sigmoid activation function (Hansen, McDonald, & Nelson, 1999). There are few comparative studies between the sigmoid function and other activation functions in ANN. This study uses the sigmoid function as the activation function because this study is performed to classify the accurate direction of change in the daily stock price index. The linear function is used as a combination function for the feed-forward computation with the derived connection weights from the first phase. 3.3. Validation phase The derived connection weights are applied to the holdout data. This phase is indispensable to validate the general- izability because ANN has the eminent ability of learning the known data. Table 1 summarizes the algorithms of GAIS. 4. Application: analysis of the stock market data This section applies GAIS to stock market prediction. The efficiency and effectiveness of GAIS may be properly tested because the stock market data is very noisy and complex. Many studies on stock market prediction using artificial intelligence techniques were performed in the past decade. Some of them, however, did not produce outstanding prediction accuracy partly because of the tremendous noise and non-stationary Table 1 The algorithms of GAIS Step 0 Initialize the populations (the connection weights between layers and the codes for instance selection). (Set to small random values between 0.0 and 1.0) Step 1 If stopping condition is false, do Step 2. Otherwise, stop the process Step 2 Do Steps 3–9 Step 3 Each processing element in the input layer receives an input signal and forwards this signal to all processing elements in the hidden layer Step 4 Each processing element in the hidden layer sums its weighted input signals and applies the sigmoid activation function to compute its output signal of the hidden processing element and forwards it to all processing elements in the output layer Step 5 Each processing element in the output layer sums its weighted signals from the hidden layer and applies the sigmoid activation function to compute its output signal of the output processing element and computes the difference between the output signal and the target value Step 6 Calculate fitness. (Fitness function: Average predictive accuracy on the selected instances within the training data) Step 7 Select individuals to become parents of the next generation Step 8 Create a second generation from the parent pool. (Perform crossover and mutation) Step 9 Test stopping condition and go back to Step 1 K.-j. Kim / Expert Systems with Applications 30 (2006) 519–526 523 characteristics in stock market data. If these factors are not appropriately controlled, the prediction system does not produce significant performance. When the prediction is executed using long-term data, this is more important to manage the consistency of prediction. 4.1. Application data The application data used in this study consists of technical indicators and the direction of change in the daily Korea stock price index (KOSPI). The total number of samples is 2348 trading days, from January 1991 to December 1998. This study divides the samples into eight data sets according to the trading year. Experiments are repeated eight times for each data set to reflect specific knowledge as time passes. The direction of daily change in the stock price index is categorized as ‘0’ or ‘1’. ‘0’ means that the next day’s index is lower than the today’s index, and ‘1’ means that the next day’s index is higher than today’s index. I select 12 technical indicators as feature subsets by the review of domain experts and prior research. Table 2 gives selected features and their formulas. Table 2 Selected features and their formulas Names of feature Stochastic %K Stochastic %D Stochastic slow %D Momentum ROC (rate of change) LW %R (Larry William’s %R) A/D Oscillator (accumulation/distribution oscillator) Disparity 5 days Disparity 10 days OSCP (price oscillator) CCI (commodity channel index) RSI (relative strength index) C, closing price; L, low price; H, high price; LLn, lowest low price in the last n days; LtCCt)/3; SMt, Pn iZ1 MtKiC1=n; Dt, Pn iZ1 jMtKiC1 KSMtj=n; Up, upward price change; Dw 4.2. Experiments The following experiments are carried out: 4.2.1. Whole training data The whole training samples are used as the training data. This is the conventional method of data analysis. 4.2.2. Selected instances with GAIS Experiments on stock market data are implemented using GAIS. The procedure of the experiment is as follows. The GA searches for optimal or near-optimal connection weights and relevant instances for ANN. As mentioned earlier, this study needs three sets of parameters: The connection weights between the input and the hidden layer, the connection weights between the hidden and the output layer, and the codes for instance selection. This study uses the following encoding for the strings: 12 input features are used and 12 processing elements in the hidden layer are employed. Each processing element in the hidden layer receives 12 signals from the input layer. The first 144 bits represent the connection weights between the input Formula (CtKLLtK5)/(HHtK5KLLtK5)!100PnK1 iZ0 %KtKi=n PnK1 iZ0 %DtKi=n CtKCtK4 (Ct/CtKn)!100 (HnKCt)/(HnKLn)!100 (HtKCtK1)/(HtKLt) (Ct/MA5)!100 (Ct/MA10)!100 (MA5KMA10)/MA5 (MtKSMt)/(0.015!Dt) 100K100= 1C PnK1 iZ0 UptKi=n � � = PnK1 iZ0 DwtKi=n � �� � HHn, highest high price in the last n days; M, moving average of price; Mt, (HtC , downward price change. Table 3 Number of instances Set Year Total 1991 1992 1993 1994 1995 1996 1997 1998 Training instances for GANN 234 236 237 237 235 235 234 234 1882 Selected instances for GAIS 74 71 87 66 93 86 93 85 655 Holdout instances for GANN & GAIS 58 58 59 59 58 58 58 58 466 Table 4 Average predictive performance (hit ratio: %) Year GANN GAIS Training Holdout Training Holdout 1991 63.68 53.45 74.32 72.41 1992 64.83 56.90 77.46 58.62 1993 61.18 59.32 70.11 59.32 1994 62.87 57.63 74.24 61.02 1995 69.36 65.52 81.72 67.24 1996 65.11 65.52 76.74 77.59 1997 64.96 58.62 65.59 58.62 1998 61.11 56.90 78.82 68.97 Total 64.13% 59.23% 74.87% 65.45% K.-j. Kim / Expert Systems with Applications 30 (2006) 519–526524 layer and the hidden layer. These bits are searched from K5 to 5. Each processing element in the output layer receives signals from the hidden layer. The next 12 bits indicate the connection weights between the hidden layer and the output layer. These bits also varied between K5 and 5. The following bits are instance selection codes for the training data. The chromosome of these bits consists of n genes (where n is the number of initial training instances), each one with two possible states: 0 or 1. ‘1’ means the associated instance is selected into the analysis and ‘0’ means the associated instance is not chosen. The encoded chromosomes are searched to maximize the fitness function. The fitness function is specific to applications. In this study, the objectives of the model are to approximate connection weights and to select relevant instances for the correct solutions. These objectives can be represented by the average prediction accuracy of the selected instances within the training data. Thus, this study applies the average prediction accuracy of the selected instances within the training data to the fitness function. Mathematically, the fitness function is represented as Eq. (1): Fitness Z 1 n Xn iZ1 CRi ði Z 1; 2; .; nÞ if POi Z AOi CRi Z 1 otherwise CRi Z 0 ( (1) where CRi is the prediction result for the ith trading day which is denoted by 0 or 1, POi is the predicted output from the model for the ith trading day, and AOi is the actual output for the ith trading day. For the controlling parameters of the GA search, the population size is set at 100 organisms and the crossover and mutation rates are varied to prevent ANN from falling into a local minimum. The value of the crossover rate is set at 0.7 while the mutation rate is 0.1. For the crossover method, the uniform crossover method is considered better at preserving the schema, and can generate any schema from the two parents, while single-point and two-point crossover methods may bias the search with the irrelevant position of the variables. Thus, this study performs crossover using the uniform crossover routine. For the mutation method, this study generates a random number between 0 and 1 for each of the variables in the organism. If a variable gets a number that is less than or equal to the mutation rate, then that variable is mutated. As the stopping condition, only 100 generations are permitted. 4.3. Experimental results and discussions This study compares GAIS to the conventional ANN with the GA. The conventional ANN with the GA, named GANN, denotes the ANN model with the connection weights, which are determined by the GA. This model does not use the gradient descent algorithm but uses the GA to determine the connection weights between layers. However, this model analyzes all available training data to learn. On the other hand, GAIS also uses the GA to determine the connection weights, but learns the patterns of the stock market data from the selected instances through an evolutionary search. For the GANN model, about 20% of the data is used for holdout and 80% for training. The training data is used to search for the optimal or near-optimal parameters and is employed to evaluate the fitness function. The holdout data is used to test the results with the data that is not utilized to develop the model. The number of the training instances in GANN and the number of the selected instances within the training instances in GAIS for each year are presented in Table 3. Table 4 describes the average prediction accuracy of each model. In Table 4, GAIS outperforms GANN by 6.22% for the holdout data. In addition, GAIS has higher accuracy than GANN by 10.74% for the training data. This result may be caused by the benefits of the instance selection through evolutionary search techniques. The McNemar tests are used to examine whether GAIS significantly outperforms GANN. This test may be used with nominal data and is particularly useful with a before–after measurement of the same subjects (Cooper & Emory, 1995). The McNemar value and its p values of the holdout data are K.-j. Kim / Expert Systems with Applications 30 (2006) 519–526 525 5.262 and 0.022, respectively. This means that GAIS performs better than GANN at the 5% statistical significance level. 5. Concluding remarks Prior studies tried to optimize the controlling parameters of ANN using global search algorithms. Some of them only focused on the optimization of the connection weights of ANN. Others placed little emphasis on the optimization of the learning algorithm itself, but most studies focused little on instance selection for ANN. In this paper, I use the GA for ANN in two ways. I first use the GA to determine the connection weights between layers. This may mitigate the well-known limitations of the gradient descent algorithm. In addition, I adopt the evolutionary instance selection algorithm for ANN. This directly removes irrelevant and redundant instances from the training data. I conclude that GA-based learning and the instance selection algorithm (GAIS) signifi- cantly outperforms the conventional GA-based learning algorithm (GANN). The prediction performance may be more enhanced if the GA is employed not only for instance selection but also for relevant feature selection, and this remains a very interesting topic for further study. Although instance selection is a direct method of noise and dimensionality reduction, feature selection effectively reduces the dimensions of feature space. In addition, while ANN performed well with GA-based learning and instance selection, other instance-based learning algorithms including CBR may also prove effective in place of ANN. Of course, there are still many tasks to be done for GAIS. The generalizability of GAIS should be further tested by applying it to other problem domains. References Adeli, H., & Hung, S. (1995). Machine learning: Neural networks, genetic algorithms, and fuzzy systems. New York: Wiley. Bauer, R. J. (1994). Genetic algorithms and investment strategies. New York: Wiley. Blum, A., & Langley, P. (1997). Selection of relevant features and examples in machine learning. Artificial Intelligence, 97(1–2), 245–271. Choi, J.H., Lee, M.K., & Lee, M.W. (1995). Trading S&P 500 stock index futures using a neural network. Proceedings of the third annual international conference on artificial intelligence applications on wall street (pp. 63–72). New York. Coakley, J. R., & Brown, C. E. (2000). Artificial neural networks in accounting and finance: Modeling issues. International Journal of Intelligent Systems in Accounting, Finance and Management, 9, 119–144. Cooper, D. R., & Emory, C. W. (1995). Business research methods. Chicago: Irwin. Dasarathy, B. V. (1990). Nearest neighbor (NN) norms: NN pattern classification techniques. California: IEEE Computer Society Press. Davis, L. (1994). Genetic algorithms and financial applications. In G. J. Deboeck (Ed.), Trading on the edge (pp. 133–147). New York: Wiley. Duke, L. S., & Long, J. A. (1993). Neural network futures trading—a feasibility study. In: Society for worldwide interbank financial telecomunications, adaptive intelligent systems (pp. 121–132). Amsterdam: Elsevier. Gates, G. W. (1972). The reduced nearest neighbor rule. IEEE Transactions on Information Theory, 18(3), 431–433. Gupta, J. N. D., & Sexton, R. S. (1999). Comparing backpropagation with a genetic algorithm for neural network training. Omega, 27(6), 679–684. Hansen, J. V., McDonald, J. B., & Nelson, R. D. (1999). Time series prediction with genetic-algorithm designed neural networks: An empirical comparison with modern statistical models. Computational Intelligence, 15(3), 171– 184. Hart, P. E. (1968). The condensed nearest neighbor rule. IEEE Transactions on Information Theory, 14, 515–516. Hertz, A., & Kobler, D. (2000). A framework for the description of evolutionary algorithms. European Journal of Operational Research, 126(1), 1–12. Hiemstra, Y. (1995). Modeling structured nonlinear knowledge to predict stock market returns. In R. R. Trippi (Ed.), Chaos & nonlinear dynamics in the financial markets: Theory, evidence and applications (pp. 163–175). Chicago, IL: Irwin. Kamijo, K., & Tanigawa, T. (1990). Stock price pattern recognition: A recurrent neural network approach. Proceedings of the international joint conference on neural networks (pp. 215–221). San Diego, CA. Kim, K., & Han, I. (2000). Genetic algorithms approach to feature discretization in artificial neural networks for the prediction of stock price index. Expert Systems with Applications, 19(2), 125–132. Kimoto, T., Asakawa, K., Yoda, M., & Takeoka, M. (1990). Stock market prediction system with modular neural network. Proceedings of the International Joint Conference on Neural Networks (pp. 1–6). San Diego, CA. Kohara, K., Ishikawa, T., Fukuhara, Y., & Nakamura, Y. (1997). Stock price prediction using prior knowledge and neural networks. International Journal of Intelligent Systems in Accounting, Finance and Management, 6(1), 11–22. Kuncheva, L. I. (1993). ‘Change-glasses’ approach in pattern recognition. Pattern Recognition Letters, 14, 619–623. Kuncheva, L. I. (1995). Editing for the k-nearest neighbors rule by a genetic algorithm. Pattern Recognition Letters, 16(8), 809–814. Lee, K. H., & Jo, G. S. (1999). Expert systems for predicting stock market timing using a candlestick chart. Expert Systems with Applications, 16(4), 357–364. Liu, H., & Motoda, H. (1998). Feature transformation and subset selection. IEEE Intelligent Systems and Their Applications, 13(2), 26–28. Liu, H., & Motoda, H. (2001). Data reduction via instance selection. In H. Liu, & H. Motoda (Eds.), Instance selection and construction for data mining (pp. 3–20). Massachusetts: Kluwer Academic Publishers. McSherry, D. (2000). Automating case selection in the construction of a case library. Knowledge Based Systems, 13(2–3), 133–140. Nikolopoulos, C., & Fellrath, P. (1994). A hybrid expert system for investment advising. Expert Systems, 11(4), 245–250. Odetayo, M. O. (1995). Knowledge acquisition and adaptation: A genetic approach. Expert Systems, 12(1), 3–13. Oh, K. J., & Han, I. (2000). Using change-point detection to support artificial neural networks for interest rates forecasting. Expert Systems with Applications, 19(2), 105–115. Reeves, C. R., & Bush, D. R. (2001). Using genetic algorithms for training data selection in RBF networks. In H. Liu, & H. Motoda (Eds.), Instance selection and construction for data mining (pp. 339–356). Massachusetts: Kluwer Academic Publishers. Reeves, C. R., & Taylor, S. J. (1998). Selection of training sets for neural networks by a genetic algorithm. In A. E. Eiden, T. Bäck, M. Schoenauer, & H.-P. Schwefel (Eds.), Parallel problem-solving from nature-PPSN V. Berlin: Springer. Ritter, G. L., Woodruff, H. B., Lowry, S. R., & Isenhour, T. L. (1975). An algorithm for a selective nearest neighbor decision rule. IEEE Transactions on Information Theory, 21(6), 665–669. Sexton, R. S., Dorsey, R. E., & Johnson, J. D. (1998). Toward global optimization of neural networks: A comparison of the genetic algorithm and backpropagation. Decision Support Systems, 22(2), 171–185. Smyth, B. (1998). Case-base maintenance. Proceedings of the 11th international conference on industrial & engineering applications of artificial intelligence & expert systems (pp. 507–516). K.-j. Kim / Expert Systems with Applications 30 (2006) 519–526526 Tetko, I. V., & Villa, A. E. P. (1997). Efficient partition of learning data sets for neural network training. Neural Networks, 10(8), 1361–1374. Tomek, I. (1976). An experiment with the edited nearest neighbor rule. IEEE Transactions on Systems, Man, and Cybernetics, 6(6), 448–452. Trippi, R. R., & DeSieno, D. (1992). Trading equity index futures with a neural network. Journal of Portfolio Management, 19, 27–33. Tsaih, R., Hsu, Y., & Lai, C. C. (1998). Forecasting S&P 500 stock index futures with a hybrid AI system. Decision Support Systems, 23(2), 161–174. Weiss, S. M., & Indurkhya, N. (1998). Predictive data mining: A practical guide. California: Morgan Kaufmann Publishers. Wilson, D. L. (1972). Asymptotic properties of nearest neighbor rules using edited data. IEEE Transactions on Systems, Man, and Cybernetics, 2(3), 408–421. Wilson, D. R., & Martinez, T. R. (2000). Reduction techniques for instance- based learning algorithms. Machine Learning, 38, 257–286. Wong, F., & Tan, C. (1994). Hybrid neural, genetic, and fuzzy systems. In G. J. Deboek (Ed.), Trading on the edge (pp. 243–261). New York: Wiley. Artificial neural networks with evolutionary instance selection for financial forecasting Introduction Research background Instance selection methods Genetic algorithms Prior research on stock market prediction using ANN A GA approach to instance selection for ANN GA search phase Feed-forward computation phase Validation phase Application: analysis of the stock market data Application data Experiments Experimental results and discussions Concluding remarks References 
work_slbvbdfjyzhnfkl3t2tfdiqp7a	----	06 April 2021 POLITECNICO DI TORINO Repository ISTITUZIONALE Analysis of diabetic patients through their examination history / Antonelli, D.; Baralis, E.; Bruno, G.; Cerquitelli, T.; Chiusano, S.; Mahoto, N.. - In: EXPERT SYSTEMS WITH APPLICATIONS. - ISSN 0957-4174. - STAMPA. - 40:11(2013), pp. 4672-4678. Original Analysis of diabetic patients through their examination history Publisher: Published DOI:10.1016/j.eswa.2013.02.006 Terms of use: openAccess Publisher copyright (Article begins on next page) This article is made available under terms and conditions as specified in the corresponding bibliographic description in the repository Availability: This version is available at: 11583/2518542 since: 2016-11-23T14:12:11Z Elsevier Ltd Analysis of diabetic patients through their examination history Dario Antonellia, Elena Baralisb, Giulia Brunoa, Tania Cerquitellib,∗, Silvia Chiusanob, Naeem Mahotob aDepartment of Production Systems and Economics, Politecnico di Torino, Turin, Italy bDepartment of Control and Computer Engineering, Politecnico di Torino, Turin, Italy Abstract The analysis of medical data is a challenging task for health care systems since a huge amount of interesting knowledge can be automatically mined to effectively support both physicians and health care organizations. This paper proposes a data analysis framework based on a multiple-level cluster- ing technique to identify the examination pathways commonly followed by patients with a given disease. This knowledge can support health care or- ganizations in evaluating the medical treatments usually adopted, and thus the incurred costs. The proposed multiple-level strategy allows clustering patient examination datasets with a variable distribution. To measure the relevance of specific examinations for a given disease complication, patient examination data has been represented in the Vector Space Model using the TF-IDF method. As a case study, the proposed approach has been applied to the diabetic care scenario. The experimental validation, performed on a real collection of diabetic patients, demonstrates the effectiveness of the ap- proach in identifying groups of patients with a similar examination history and increasing severity in diabetes complications. Keywords: data mining, cluster analysis, patient examination history, diabetes ∗Corresponding author Email addresses: dario.antonelli@polito.it (Dario Antonelli), elena.baralis@polito.it (Elena Baralis), giulia.bruno@polito.it (Giulia Bruno), tania.cerquitelli@polito.it (Tania Cerquitelli), silvia.chiusano@polito.it (Silvia Chiusano), naeem.mahoto@polito.it (Naeem Mahoto) Preprint submitted to 1. Introduction Nowadays, large amount of medical data, storing the medical patient his- tory, is collected by health care organizations. Data mining, which focuses on studying effective and efficient algorithms to transform large amounts of data into useful knowledge (Pang-Ning T. and Steinbach M. and Kumar V., 2006), may provide valuable insight into these huge data collections. For example, data mining techniques can be used to extract a variety of infor- mation on the patient history such as the medical protocols usually adopted for patients with a given disease. Healthcare organizations can exploit this knowledge to improve their current processes, assess new medical guidelines, or enrich the existing ones. Medical guidelines represent standard medical pathways specifying the actions necessary to treat, with optimal effectiveness and efficiency, patients with a given disease. This study addresses the problem of analysing patients’ examination data to identify the examination pathways commonly followed by patients. This issue is crucial for health care organizations, because it can significantly impact on the effectiveness of the medical treatments as well as on the costs incurred by the organizations. The data analysis framework proposed in this paper exploits a multiple- level clustering approach to discover, in a data collection with a variable distribution, cohesive and well-separated groups of patients with a similar ex- amination history. Cluster analysis is an exploratory data mining technique that partitions a data object collection into groups based on object proper- ties, without the support of additional a priori knowledge (in contrast with classification algorithms using class label information) (Pang-Ning T. and Steinbach M. and Kumar V., 2006). To cluster data collections with a vari- able distribution, the proposed multiple-level clustering strategy iteratively focuses on disjoint dataset portions and locally identifies clusters. Among the state-of-the-art clustering techniques, the density-based DBSCAN algo- rithm (Ester et al., 1996) has been adopted due to its properties. DBSCAN allows the identification of arbitrarily shaped clusters, is less susceptible to noise and outliers, and does not require the specification of the number of expected clusters in the data. To highlight the relevance of specific exam- inations for a given clinical condition, in the proposed framework patient examination data has been represented in the Vector Space Model (VSM) (Salton G., 1971) using the TF-IDF method (Pang-Ning T. and Steinbach M. and Kumar V., 2006). 2 As a reference case study, the proposed framework has been applied to a real dataset of diabetic patients provided by the National Health Center (NHC) of the Asti province (Italy). The results showed that, starting from a large collection of raw examination data, the framework allows the iden- tification of clusters containing patients with a similar examination history. More specifically, clusters contain patients with increasing disease severity, as patients are tested with more and more specific examinations to diagnose diabetes complications. The results were discussed with the support of clini- cal domain experts, showing a fairly good correlation among the examination pathways suggested by the clusters and the guidelines for diabetes disease (ICD-9-CM, 2011). The paper is organized as follows. Section 2 describes the state-of-the- art clustering methods and explains the selection of the DBSCAN algorithm for this study. Section 3 analyses previous related work. Section 4 presents the proposed framework and describes its building blocks, while the results obtained for the real diabetic patient dataset are discussed in Section 5. Finally, Section 6 draws conclusions and future work. 2. Selection of the clustering algorithm Cluster analysis partitions objects into groups (clusters) so that objects within the same group are more similar to each other than those objects assigned to different groups (Pang-Ning T. and Steinbach M. and Kumar V., 2006). Clustering algorithms require the definition of a metric to evaluate the similarity (or distance) between objects based on the features describing objects. Clustering algorithms can be classified into the following four categories (Pang-Ning T. and Steinbach M. and Kumar V., 2006): (i) center, (ii) density, (iii) model, and (iv) hierarchical-based methods. In center-based methods (e.g., K-means (Juang & Rabiner, 1990)), a clus- ter is a set of data objects in which each object is closer (more similar) to the prototype that defines the cluster than to the prototype of any other cluster. The prototype is the most representative point in the cluster. These meth- ods find spherical-shaped clusters, unless clusters are well separated, and are sensitive to outliers. In density-based methods (e.g., DBSCAN (Ester et al., 1996)), a cluster is a dense area of data objects surrounded by an area of low density. These 3 approaches are less sensitive to the presence of outliers than center-based techniques and can identify non-spherical shaped clusters. Model-based methods (e.g. EM (G. McLachlan and T. Krishnan, 1997), COBWEB (Fisher, 1987a,b)) hypothesize a mathematical model for each cluster, and then determine the best fit between the model and the object collection. These methods can deal with outliers and noise. However, sim- ilarly to K-means, EM requires the specification of the number of expected clusters. Hierarchical-based methods exploit an agglomerative or divisive approach to generate a hierarchical collection of clusters. The (most common) agglom- erative approach (Pang-Ning T. and Steinbach M. and Kumar V., 2006) ini- tially assigns each data object to a singleton cluster. The two closest clusters are then iteratively merged using a cluster proximity measure (e.g., single- link, complete-link, or group average average) (Pang-Ning T. and Steinbach M. and Kumar V., 2006). These methods are often used when the underlying application requires the creation of a taxonomy, which is not the case for our application scenario. Density-based methods showed remarkable properties for clustering pa- tients based on their examination history. More specifically, the very effective density-based algorithm DBSCAN (Ester et al., 1996) has been selected in this study. Differently from other algorithms (e.g., K-means), DBSCAN is less sensitive to outliers and can find arbitrarily shaped clusters. Outliers, when not identified and isolated as in DBSCAN, are clustered together with the other data objects, thus decreasing cluster cohesion. In addition, DB- SCAN does not require an a priori specification of the number of clusters in the data, as opposed to K-means and EM. Medical datasets can include outliers as specific examination pathways for some disease conditions and clusters can be non-spherical shaped. Besides, since our aim is discovering the examination pathways usually adopted for a given disease through an explorative data analysis, the expected number of clusters can be hardly guessed a priori. For these reasons, the DBSCAN algorithm has been adopted for the cluster analysis in this study. The main characteristics of DBSCAN are reported in the following Section 2.1. To discover clusters in datasets with a variable distribution, state-of-the- art clustering algorithms can be applied in a multiple-level fashion to focus on different dataset portions and locally identify cluster (e.g., the bisecting K-means algorithm (Pang-Ning T. and Steinbach M. and Kumar V., 2006)). 4 The multiple-level strategy adopted in this study exploits the DBSCAN algo- rithm to select the dataset part analyzed at each iteration and locally cluster it. 2.1. The DBSCAN algorithm The DBSCAN algorithm (Ester et al., 1996) relies on two input parame- ters, named Eps and MinPts, to define a density threshold in the data space. A dense region in the data space is a n-dimensional sphere with radius Eps and containing at least MinPts objects. The DBSCAN algorithm iterates over the data objects in the collection by analyzing their neighborhood. It classifies objects as being (i) in the interior of a dense region (a core point), (ii) on the edge of a dense region (a border point), or (iii) in a sparsly occupied region (a noise or outlier point). Any two core points that are close enough (within a distance Eps of one another) are put in the same cluster. Any border point close enough to a core point is put in the same cluster as the core point. Outlier points (i.e., points far from any core point) are isolated. DBSCAN can discover arbitrarily shaped clusters and identify outliers as objects in a low density area in the data space. The effectiveness of DBSCAN is affected by the selection of the Eps and MinPts values. Section 4.2 discusses how this issue has been addressed in this study. 3. Related work Several works exploited data mining techniques to analyze medical data by addressing different pathologies and facets of various diseases. Clustering algorithms have been widely exploited to analyze medical data for patients affected by different diseases (Ahmed & Funk, 2011; Buczak et al., 2009; Choong et al., 2000; Mulroy et al., 2003; Van Rooden et al., 2010; Santamaria et al., 2003). (Van Rooden et al., 2010) reviewed the cluster methods used to identify Parkinson’s disease subtypes. It showed that the K-means algorithm (Juang & Rabiner, 1990) was mostly adopted, but also highlited its two major limitations, i.e., the sensitiveness to outliers and the need of defining the expected number of clusters. To overcome these issues, the DBSCAN algorithm has been used in this study, being able to internally evaluate the optimal number of clusters and automatically identify outliers. In (Buczak et al., 2009), patients were represented by tracking the number of occurrences for each clinical event (e.g., hospital visit, lab order). Patients 5 were then grouped using a hierarchical agglomerative clustering method with Ward’s linkage as proximity measure (Pang-Ning T. and Steinbach M. and Kumar V., 2006). Differently from (Buczak et al., 2009), in this study we ex- ploit the TF-IDF scheme to weight the relevance of specific examinations for each diabetes condition. In addition, a multiple-level cluster approach is used to identify groups of patients in datasets with a variable data distribution. Concerning the analysis of medical data for diabetic patients, several works, mainly exploiting classification techniques, have been proposed (Kare- gowda et al., 2012; Meng et al., 2012; Mohamudally & Khan, 2011; Zhong et al., 2012). Classification is a supervised data mining approach that assigns new unlabeled data to a class label by means of a model built from data with known class label (Pang-Ning T. and Steinbach M. and Kumar V., 2006). In (Karegowda et al., 2012), diabetic patients were categorized with a K-nearest neighbor (KNN) classifier (Pang-Ning T. and Steinbach M. and Kumar V., 2006). To ehnance classification accuracy, in a pre-processing step a genetic algorithm and a feature selection technique (i.e., the correlation method (Pang-Ning T. and Steinbach M. and Kumar V., 2006)) are used to identify the relevant features for classification. In (Meng et al., 2012) the authors compared three prediction models (i.e., logistic regression, decision tree, and artificial neural networks) for diabetic patients classification. The comparison was based on common risk factors collected from both diabetic and pre-diabetic patients with a standard questionnaire. This work showed that the decision tree model (C5.0 (Pang-Ning T. and Steinbach M. and Kumar V., 2006)) yielded the best accuracy, followed by logistic regression, and artificial neural networks. The work in (Zhong et al., 2012) proposed a multi-level support vector machine approach to classify and predict clinical charge profiles as well as the length of hospital stay for patients affected by heart, diabetes, and cancer diseases. In (Mohamudally & Khan, 2011), different data mining algorithms (the K-means algorithm, C4.5 decision tree classifier, artificial neural networks, and 2D graphs for data visualization) are integrated to predict, classify, and visualize a medical diabetic dataset. A parallel effort has been devoted to exploit clustering techniques for diabetic patients by addressing different issues as food analysis (Phanich et al., 2010), gait patterns (Sawacha et al., 2010), and relationships among diabetes and risk factors (Chaturvedi, 2003). Food clustering analysis for diabetic patients has been proposed in (Phanich et al., 2010). Using Self- Organizing Map (SOM) and the K-means algorithm, this work provided a Food Recommendation System suggesting proper substituted foods. The 6 work in (Sawacha et al., 2010) proposed a cluster analysis of biomechanical data to group patients with similar diabetic gait patterns. In (Chaturvedi, 2003) the spatial clusters of diabetes prevalence in Texas has been analyzed. The relationship of risk factors (i.e., age and obesity) associated with diabetes have also been analyzed. Differently from these works, this study exploits clustering techniques to identify groups of patients with similar examinations history. 4. Methodology The proposed framework to analyse the patient examination history con- tains four main steps: (i) data collection, (ii) data transformation, (iii) cluster analysis, and (iv) cluster evaluation. The building blocks of the framework are shown in Figure 4 and detailed in the following subsections. The patient examination log data is first collected and then transformed using the Vector Space Model (VSM) representation (Salton G., 1971). Ex- amination frequencies are weighted through the Term Frequency (TF) - In- verse Document Frequency (IDF) scheme (Pang-Ning T. and Steinbach M. and Kumar V., 2006). The multiple-level clustering approach is applied to identify, in a dataset with a variable distribution, groups of patients with a similar examination history. The DBSCAN algorithm has been exploited for the cluster analysis. Finally, clustering results are evaluated through a quality index balanc- ing both intra-cluster homogeneity and inter-cluster separation. Silhouette (Rousseeuw, 1987) has been considered as reference index. Cluster sets are also analysed together with a domain expert to assess their significance. To analyse the examination pathways represented by the cluster set, each cluster has been characterized with the examinations appearing in it. 4.1. Representation of the patient examination history The data recording the patient examination history is represented using the Vector Space Model (VSM) (Salton G., 1971). Each patient is a vector in the examination space. Each vector element corresponds to a different exam- ination and is associated with a weight describing the examination relevance for the patient. More specifically, it reports the weighted number of times the examination was repeated by the patient. The Term Frequency (TF) - Inverse Document Frequency (IDF) scheme (Pang-Ning T. and Steinbach M. and Kumar V., 2006) has been adopted to weight examination frequency. 7 W eight ed examination frequency vect ors Data transformation Cluster analysis Cluster evaluation Data collection VSM representation using TF- I DF scheme Pat ient examination log dat a Mult iple- level clust ering Cluster set Clust er evaluat ion Figure 1: The proposed framework for the cluster analysis of patient examination data Both the VSM representation and the TF-IDF scheme have been applied in previous works to represent text documents. The adopted data representation allows highlighting the relevance of spe- cific examinations for a given patient condition. The TF-IDF value increases proportionally to the number of times an examination appears in the pa- tient history, but is offset by the frequency of the examination in the patient collection, which helps to control the fact that some examinations are gen- erally more common than others. Unweighted examination frequencies do not properly characterize the patient condition, since standard routine tests usually appear with high frequency, while more specific tests may appear with lower frequency. More formally, let D be a collection of patient records and Σ = {e1, . . . , ek} the set of examinations done by at least one patient in D. Each patient pi in D is represented by a weighted examination frequency vector vpi of |Σ| cells. Each element vpi[j] of vector vpi reports the weighted frequency wpi,ej of examination ej for patient pi, i.e., 8 vpi = [wpi,e1, . . . , wpi,e|Σ|]. (1) The TF-IDF weight wpi,ej for the pair (pi, ej) is computed as the prod- uct of two terms, called Term Frequency (TFpi,ej) and Inverse Document Frequency (IDFej), wpi,ej = TFpi,ej ∗ IDFej. (2) The Term Frequency TFpi,ej for the pair (pi, ej) represents the relative frequency of examination ej for patient pi. It is given by TFpi,ej = fpi,ej/ ∑ 1≤k≤|Σ| fpi,ek, (3) where fpi,ej is the number of times patient pi underwent examination ej and ∑ 1≤k≤|Σ| fpi,ek is the total number of examinations done by patient pi. The Inverse Document Frequency IDFej for examination ej represents the frequency of ej in the patient collection. It is computed as IDFej = Log[|D|/|pk ∈ D : fpk,ej 6= 0|], (4) where |D| is the number of patients in the collection D and |pk ∈ D : fpk,ej 6= 0| is the number of patients in D who underwent (at least once) examina- tion ej. Mathematically, the base of the log function does not matter and constitutes a constant multiplicative factor towards the overall result. The TF-IDF weight wpi,ej for the pair (pi, ej) is high when examination ej appears with high frequency in patient pi and low frequency in patients in the collection D. When examination ej appears in more patients, the ratio inside the IDF’s log function approaches 1, and the IDFej value and TF- IDF weight wpi,ej become close to 0. Hence, the approach tends to filter out common examinations. 4.2. The multiple-level DBSCAN approach for cluster analysis Density-based algorithms can effectively discover clusters of arbitary shape and filter out outliers, thus increasing cluster homogeneity. Clusters are iden- tified as dense areas of data objects surrounded by an area of low density. In the DBSCAN algorithm, density is evaluated based on the user-specified parameters Eps and MinPts. One single execution of DBSCAN discovers dense groups of patients according to one specific setting for these parame- ters. Patients in lower density areas are labeled as outliers and not assigned 9 to any cluster. Hence, different parameter settings are needed to discover clusters in datasets with a variable data distribution as the one considered in this study. Groups of patients with close examination histories may have both different cardinalities and densities, i.e., groups may contain examina- tion histories with different degrees of similarity. The proposed multiple-level clustering approach allows clustering datasets with a variable distribution by iteratively applying the DBSCAN algorithm on different (disjoint) dataset portions. The whole original dataset is clus- tered at the first level. Then, at each subsequent level, patients labeled as outliers in the previous level are re-clustered. The DBSCAN parameters Eps and MinPts are properly set at each level. In the following, Section 4.2.1 presents the measure used to evaluate the similarity between patient examination histories, while Section 4.2.2 de- scribes how the DBSCAN parameters and the number of clustering levels have been selected. 4.2.1. Similarity between patient examination histories The cosine similarity measure has been adopted to evaluate the simi- larity between the weighted examination frequency vectors representing the patient examination histories. This measure has been often used to compare documents in text mining (Steinbach et al., 2000). Let pi and pj be two arbitrary patients in the collection D. Let vi and vj be the corresponding weighted examination frequency vectors as in Equation 1. The cosine similarity between vi and vj is computed as cos(vi, vj) = vi • vj ‖vi‖ ‖vj‖ = ∑ 1≤k≤|Σ| vi[k]vj[k] √ ∑ 1≤k≤|Σ| vi[k] 2 √ ∑ 1≤k≤|Σ| vj[k] 2 , (5) where cos(vi, vj) is in the range [0,1]. cos(vi, vj) equal to 1 describes the exact similarity of examination histories for patients pi and pj, while cos(vi, vj) equal to 0 points out that patients have complementary histories (i.e., the sets of their examinations are disjoint). 4.2.2. Number of clustering levels and DBSCAN parameters The dataset density is preliminarly analysed using the k-dist graph (Pang- Ning T. and Steinbach M. and Kumar V., 2006) to select both the number of 10 iterations for the multiple-level clustering approach and the Eps and MinPts values for each iteration. For each patient in the collection, the k-dist graph plots the distance to its kth nearest neighbor according to their examination history. On the x- axis patients are sorted by the distance to the kth nearest neighbor, while on the y-axis distances to the kth nearest neighbor are reported. The k value corresponds to the MinPts parameter. The y-axis represents possible values of the Eps parameter. By cutting the graph at a given value on the y-axis, the corresponding px value on the x-axis partitions the patient collection into the following two subsets. Patients placed on the left hand side of px are labeled by DBSCAN as core points, and those on the rigth side of px as outlier or border points. Sharp changes in the k-dist graph identify dataset portions with a dif- ferent density (Pang-Ning T. and Steinbach M. and Kumar V., 2006). The multiple-level strategy analyzes these dataset portions in different iterations. The Eps value to cluster each dataset part is selected in correspondence with the sharp change appearing in the graph. Section 5.2 discusses the selection of DBSCAN parameters and number of iteration levels for the diabetes dataset considered in this study. 4.3. Cluster evaluation The discovered cluster set is evaluated using the Silhouette index (Rousseeuw, 1987). Silhouette allows evaluating the appropriateness of the assignment of a data object to a cluster rather than to another by measuring both intra- cluster cohesion and inter-cluster separation. The silhouette value for a given patient pi in a cluster C is computed as s(pi) = b(pi) − a(pi) max{a(pi), b(pi)} , s(pi) ∈ [−1, 1], (6) where a(pi) is the average distance of patient pi from all other patients in the cluster C, and b(pi) is the smallest of average distances from its neighbour clusters. The silhouette value for a cluster C is the average silhouette value on all its patients. Negative silhouette values represent wrong patient place- ments, while positive silhouette values a better patient assignments. Clusters with silhouette values in the range [0.51,0.70] and [0.71,1] respectively show that a reasonable and a strong structure have been found (Kaufman, L. and 11 Rousseeuw, P. J., 1990). The cosine similarity metric has been used for sil- houette evaluation, since this measure was used to evaluate patient similarity in the cluster analysis (see Section 4.2.1). 4.4. Data mining tool The DBSCAN algorithm available in the RapidMiner system has been used for the cluster analysis within the proposed framework. The RapidMiner toolkit (Rapid Miner Project, 2013) is an open-source system consisting of a number of data mining algorithms to automatically analyze a large data collection and extract useful knowledge. The procedures for data transformation and cluster evaluation have been developed in the Phyton programming language (Python Software Founda- tion, 2013). These procedures transform the patient examination log data into the VSM representation using the TF-IDF scheme and compute the silhouette values for the cluster set provided by the cluster analysis. 5. Results and discussion This section presents and discusses the results obtained when analysing a real collection of examination log data for diabetic patients with the proposed framework. In the following, Section 5.1 describes the dataset considered in this study, while Section 5.2 specifies the framework configuration for the cluster analysis. The cluster results are presented and discussed in Section 5.3. Finally, Section 5.4 evaluates the performance of the multiple-level clus- tering approach in terms of execution time. 5.1. Diabetic patient dataset The dataset considered in this study was collected by the Local Health Center (LHC) of the Asti province in Italy. The LHC database stores all the accesses to the health care system in the Asti province in the year 2007. From this database the examination log data of all patients with overt di- abetes were extracted. Raw data contain 95,788 records with examinations performed by 6,380 patients. They contain both (i) routine and (ii) more specific examinations to analyze diabetes complications on various degrees of severity. The dataset includes both male and female patients in a wide age range (between 4 and 95 years). The diagnostic and therapeutic procedures are defined using the ICD 9-CM (International Classification of Diseases, 9th revision, Clinical Modification) (ICD-9-CM, 2011). 12 5.2. Selection of parameters for clustering diabetic patients To select the number of iterations for the multiple-level clustering strategy and the DBSCAN parameter for each level, we relied on the k-dist graph as discussed in Section 4.2.2. In selecting these parameters, we addressed the following issues. To discover representative examination pathways for the diabetes, we aim at avoiding clusters including few patients. In addition, to consider all different patient examination histories, we aim at limiting the number of patients labeled as outliers and thus unclustered. We plotted the k-dist graph, and analysed the cluster results, by varying the MinPts (i.e., k) value. Low MinPts values were not suitable for our purpose, because DBSCAN identifies small groups of patients. At the same time, large MinPts values may increase the number of outlier points included in the clusters. The experimental results showed that MinPts in the range [25,35] provided similar results, that were compliant with the above issues. We selected MinPts=30. From the k-dist graph for k=30, we observed three main dataset portions with a different density for Eps in the range [0.2-0.3], [0.4-0.5], and [0.6-0.7], respectively. Consequently, we adopted a three-level clustering approach, with each level focusing on one among the three dataset parts. The selected Eps values were 0.3, 0.4, and 0.6 for the first, second, and third clustering level, respectively. 5.3. Analysis of the clustering results Starting from a large collection of raw examination data, the proposed framework allows the discovery of a set of clusters containing patients with a similar examination history. The multiple-level DBSCAN approach, iterated for three levels, computed clusters that progressively contain patients with increasing severity in diabetes, because patients are tested using more and more specialized examinations. More specifically, first-level clusters contain patients mainly undergoing routine tests to monitor diabetes conditions, or some basic tests to diagnose disease complications. Second-level clusters col- lect patients that are tested using an increasing number of examinations to diagnose some diabetes complications. Examinations become progressively more numerous and specific in third-level clusters, indicating patients that can have diabetes complications of increasing severity. Since at each level clusters contain more specific examinations, a lower number of patients is contained in each cluster and the cluster size tends to reduce progressively. Clusters show good cohesion and separation as they are characterized by high 13 silhouette values. The results, discussed with the support of a clinical do- main expert, show a fairly good correlation among the examination pathways suggested by the clusters and the guidelines for diabetes disease (ICD-9-CM, 2011). Cluster properties are discussed in detail in the following subsections. Tables 1, 2, and 3 report, for each first- and second-level cluster, the ex- amination frequencies computed as the percentage of patients in the cluster tested by each examination. Clusters are named as Cij in the tables, where j denotes the level of the multiple-level DBSCAN approach providing the cluster and i locally identifies the cluster at each level j. 5.3.1. First-level cluster set First-level clusters are reported in Tables 1 and 2. Clusters can be par- titioned into the following two main groups: clusters containing patients (i) with standard examinations to monitor diabetes conditions (clusters C11- C51, in Table 1), (ii) coupled with basic examinations to diagnose disease complications (clusters C61-C111 in Table 2). The two largest clusters (C11 and C21) contain patients who mostly per- formed standard routine tests. Besides routine examinations, all patients in cluster C31 had a specialistic visit and were tested with usual basic exam- inations to diagnose the most frequent diabetes complications, as risks for cardiovascular disease and eye problems. All patients in clusters C41 and C51 only had a checkup visit, together with the glucose level test in cluster C51. These two clusters may include patients usually tested in private structures and periodically reporting test results to NHC. Patients in clusters C61-C111 were additionally tested to diagnose diabetes complications in the (a) eye (C61), (b) cardiovascular system (C71), (c) both eye and cardiovascular system (C81), (d) carotid (C91), and (e) limb (C101). Finally, (f) cluster C111 includes tests for the liver, kidneys, and in particular cardiovascular system. Differently from second- and third-level clusters, di- abetes complications were monitored in clusters C61-C111 using a few (quite standard) tests which showed a limited degree of severity. Only the cardio- vascular system has been thoroughly tested in cluster C111. Standard routine examinations still appear in clusters C61-C111 even though (usually) with a lower frequency than in clusters C11-C51. Patients with an examination history significantly dissimilar from all the others are labeled as outliers and not included in any cluster. The DBSCAN 14 Table 1: Examination frequencies (%) in first-level clusters containing patients undergoing routine tests (Eps=0.3, MinPts=30) Category Examination C11 C21 C31 C41 C51 Routine Checkup visit 78 96 58 100 100 Glucose level 78 98 63 - 100 Urine test 72 97 58 - - Venous blood 96 75 35 - - Capillary blood 72 97 58 - - Haemoglobin 100 - - - - Specialistic visit - 13 100 - - Cardiovascular Electrocardiogram - - 100 - - Eye Fundus Oculi - - 28 - - Number of patients 223 1,764 43 110 41 Silhouette 0.67 0.55 0.85 0.99 1.0 algorithm was re-applied on these patients only (3,509 patients), with differ- ent parameter values. The results are discussed in Section 5.3.2. 5.3.2. Second-level cluster set Second-level clusters contain patients with more diversified examination histories. More specifically, the following two main categories of clusters can be identified: (i) clusters containing patients tested with specific examina- tions to diagnose a given diabetes complication (clusters C12-C22); (ii) clus- ters with patients who underwent various examinations to diagnose different diabetes complications (clusters C32-C52). These two categories respectively indicate patients that can be seriously affected by a particular disease com- plication or by more than one disease complication at the same time. Second- level clusters are reported in Table 3. Clusters C12 and C22 include specific examination to diagnose eye compli- cations. More specifically, all patients in cluster C12 underwent a battery of tests to assess vision and ability to focus on objects (called ”Eye examination” in Table 3). Instead, all patients in cluster C22 had Retinal photocoagula- tion, a laser operation done in cases of long-term eye complications, such as proliferative retinopathy. All patients in clusters C32-C52 may suffer complications on the cardio- vascular, liver, and kidneys systems, but with different degrees of severity. 15 Table 2: Examination frequencies (%) in first-level clusters including basic tests to diagnose diabetes complications (Eps=0.3, MinPts=30) Category Examination C61 C71 C81 C91 C101 C111 Routine Checkup visit 77 78 66 62 68 97 Glucose level 74 74 64 62 59 97 Urine test 74 74 64 57 56 92 Venous blood 57 60 53 48 44 97 Capillary blood 74 73 63 55 56 92 Haemoglobin - - - 14 12 100 Specialistic visit - - - - - - Complete blood count - - - - 3 3 Cardiovascular Electrocardiogram - 100 100 - - 42 Cholesterol - - - - - 100 HDL Cholesterol - - - - - 100 Triglycerides - - - - - 100 Eye Fundus Oculi 100 - 100 - - 39 Liver ALT - - - - - 100 AST - - - - - 100 Kidney Creatinine - - - - - 3 Creatinine clearance - - - - - 100 Culture urine - - - - - 100 Microscopic urine analysis - - - - - 100 Uric Acid - - - - - 100 Carotid ECO doppler carotid - - - 100 - - Limb ECO doppler limb - - - - 100 - Number of patients 294 144 140 42 34 36 Silhouette 0.65 0.74 0.79 0.95 0.97 0.90 16 Patients in cluster C52 are at risk of more severe liver complications than those in clusters C32 and C42, since their examination history includes more tests in this category. Analogously, cluster C32 is characterized by more severe renal complications than clusters C42 and C52. Cardiovascular com- plications have similar severity in clusters C32-C52, because examinations in this category appear with similar frequency. At this stage, 2,939 patients are classified as outliers and not assigned to any cluster. We further applied the DBSCAN algorithm on them by modifying the parameter setting. The results are analyzed in Section 5.3.3. 5.3.3. Third-level cluster set The results collected at this stage (with DBSCAN parameter Eps=0.6 and MinPts=30) show a similar trend to second-level clusters. Third-level clusters contain patients that may suffer more complications at the same time (e.g., complications on carotid, liver, and cardiovascular systems) and are tested with more specific examinations (e.g., transcutaneous oxygen and carbon dioxide monitor or upper abdominal ultrasound). By stopping the multiple-level DBSCAN approach at this level, only 1,239 patients labeled as outliers remain unclustered, with respect to the initial collection of 6,380 patients considered at the first level (i.e., about 19% of patients). By further applying the DBSCAN algorithm on this outlier set, fragmented groups of patients can be identified. These clusters can represent patients affected by more rare diabetes complications, and thus with examinations different from those done by most patients. 5.4. Execution time The run time of DBSCAN at the first, second, and third level is 9 min 10 sec, 2 min 25 sec, and 1 min 45 sec, respectively. The run time progressively reduces because less patients are considered at each subsequent level. 6. Conclusion To get the most out of large and complex medical databases, innovative data analysis techniques are needed to extract useful knowledge in a timely fashion. In this paper a data analysis framework based on a multiple-level clustering approach has been proposed to identify groups of patients with a similar examination history in a dataset with a variable data distribution. The TF-IDF method has been exploited to represent patient examination 17 Table 3: Examination frequencies (%) in second-level clusters, including examinations to diagnose more severe diabetes complications (Eps=0.4, MinPts=30) Category Examination C12 C22 C32 C42 C52 Routine Checkup visit 65 90 22 98 71 Glucose level 52 92 22 100 95 Urine test 37 90 19 98 68 Venous blood 35 84 100 100 98 Capillary blood 37 85 17 98 63 Haemoglobin 13 34 100 98 83 Specialistic visit - 13 - - - Complete blood count 3 15 45 7 93 Cardiovascular Electrocardiogram 17 21 70 47 20 Cholesterol 7 25 100 97 93 HDL Cholesterol 4 26 100 99 92 Triglycerides 7 26 100 97 90 Eye Fundus Oculi 50 38 53 49 27 Angioscopy - 30 - - - Eye examination 100 5 - - - Renital photocoagulation - 100 - - - Liver ALT - 21 10 98 98 AST - 21 8 97 98 Bilirubin - 2 - - 100 Gamma GT - - 100 - 95 Kidney Creatinine 4 20 99 11 78 Creatinine clearance 2 10 - 99 - Culture urine 2 11 67 97 44 Microscopic urine analysis 4 16 2 82 73 Uric acid - 13 3 97 68 Microalbuminuria - 7 100 60 37 Carotid ECO doppler carotid - 5 - - 2 Limb ECO doppler limb - - - - - Number of patients 46 61 139 283 41 Silhouette 0.92 0.88 0.72 0.69 0.87 18 data, thus highlighting the relevance of the different examinations. The pro- posed methodology has been validated in the diabetic care scenario on a real dataset of diabetic patients. The analysis identified cohesive and well- separated groups of patients with standard or more specific examinations for diabetes, showing a good correlation with medical guidelines for dia- betes (ICD-9-CM, 2011). Future developments of the proposed approach will explore the correlation between patient examination data and additional aspects of the medical treatments such as pharmaceutical drug therapies. Furthermore, we plan to apply the proposed approach to different medical contexts (e.g., cardiac patients). 7. Acknowledgments The authors wish to thank Dr. Baudolino Mussa and Dr. Dario Bellomo for their advice and fruitfull discussions. References Ahmed, M. U., & Funk, P. (2011). Mining rare cases in post-operative pain by means of outlier detection. In IEEE Symposium on Signal Processing and Information Technology (ISSPIT) (pp. 35–41). Buczak, A. L., Moniz, L. J., Feighner, B. H., & Lombardo, J. S. (2009). Mining electronic medical records for patient care patterns. In IEEE Sym- posium on Computational Intelligence and Data Mining (CIDM) (pp. 146– 153). Chaturvedi, K. (2003). Geographic concentrations of diabetes preva- lence clusters in texas and their relationship to age and obesity. http://www.ucgis.org/summer03/studentpapers/kshitijchaturvedi.pdf. Re- trieved, 9 , 2010. Choong, P. F. M., Langford, A. K., Dowsey, M. M., & Santamaria, N. M. (2000). Clinical pathway for fractured neck of femur: a prospective, con- trolled study. Med J Aust., 172 , 423–426. Ester, M., Kriegel, H.-P., Sander, J., & Xu, X. (1996). A density-based algorithm for discovering clusters in large spatial databases with noise. In Knowledge Discovery and Data Mining (KDD) (pp. 226–231). 19 Fisher, D. H. (1987a). Improving inference through conceptual clustering. In National conference on Artificial intelligence (AAAI) (pp. 461–465). Fisher, D. H. (1987b). Knowledge acquisition via incremental conceptual clustering. Machine Learning, 2 , 139–172. G. McLachlan and T. Krishnan (1997). The EM algorithm and extensions. Wiley series in probability and statistics. John Wiley and Sons. ICD-9-CM, I. (2011). International Classification of Diseases, 9th revision, Clinical Modification. Available: http://icd9cm.chrisendres.com. Last ac- cess on March 2011, . Juang, B.-H., & Rabiner, L. (1990). The segmental k-means algorithm for estimating parameters of hidden markov models. IEEE Transactions on Acoustics, Speech and Signal Processing, 38 , 1639–1641. Karegowda, A., Jayaram, M., & Manjunath, A. (2012). Cascading k-means clustering and k-nearest neighbor classifier for categorization of diabetic patients. International Journal of Engineering and Advanced Technology (IJEAT), (pp. 147 – 151). Kaufman, L. and Rousseeuw, P. J. (1990). Finding groups in data: An introduction to cluster analysis. Wiley. Meng, X.-H., Huang, Y.-X., Rao, D.-P., Zhang, Q., & Liu, Q. (2012). Com- parison of three data mining models for predicting diabetes or prediabetes by risk factors. The Kaohsiung Journal of Medical Sciences, (pp. 93–99). Mohamudally, N., & Khan, D. M. (2011). Application of a unified med- ical data miner (umdm) for prediction, classification, interpretation and visualization on medical datasets: the diabetes dataset case. In Interna- tional conference on Advances in data mining: applications and theoretical aspects (ICDM) (pp. 78–95). Berlin, Heidelberg: Springer-Verlag. Mulroy, S., Gronley, J., Weiss, W., Newsam, C., & Perry, J. (2003). Use of cluster analysis for gait pattern classification of patients in the early and late recovery phases following stroke. Gait Posture, 18 , 114–25. Pang-Ning T. and Steinbach M. and Kumar V. (2006). Introduction to Data Mining. Addison-Wesley. 20 Phanich, M., Pholkul, P., & Phimoltares, S. (2010). Food recommendation system using clustering analysis for diabetic patients. In IEEE Interna- tional Conference on Information Science and Applications (ICISA) (pp. 1–8). Python Software Foundation, P. S. (2013). Python Programming Language Official Website. Available: http://www.python.org/ Last access on Jan- uary 2013, . Rapid Miner Project, R. M. (2013). The Rapid Miner Project for Machine Learning. Available: http://rapid-i.com/ Last access on January 2013, . Rousseeuw, P. J. (1987). Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Computational and Applied Mathemat- ics, (pp. 53–65). Salton G. (1971). The SMART retrieval system: experiments in automatic document processing. Prentice-Hall. Santamaria, N., Houghton, L., & Kimmel, A., L. Graham (2003). Clinical pathways for fractured neck of femur: A cohort study of health related quality of life, patient satisfaction and clinical outcome. Australian Journal of Advanced Nursing, 20 , 24–29. Sawacha, Z., Guarneri, G., Avogaro, A., & Cobelli, C. (2010). A new classi- fication of diabetic gait pattern based on cluster analysis of biomechanical data. Journal of Diabetes Science and Technology, 4 , 1127–38. Steinbach, M., Karypis, G., & Kumar, V. (2000). A comparison of document clustering techniques. In KDD Workshop on Text Mining. Van Rooden, S. M., Heiser, W. J., Kok, J. N., Verbaan, D., van Hilten, J. J., & Marinus, J. (2010). The identification of parkinson’s disease subtypes using cluster analysis: A systematic review. Movement Disorders, 25 , 969–978. Zhong, W., Chow, R., & He, J. (2012). Clinical charge profiles prediction for patients diagnosed with chronic diseases using multi-level support vector machine. Expert Systems with Applications, 39 , 1474 – 1483. 21 
work_snfq2jntf5aybdqtyr56uz5n64	----	Microsoft Word - ESWA-D-16-03854.docx 1 Effective Features to Classify Skin Lesions in Dermoscopic images Zhen Maa, João Manuel R. S. Tavaresb a Instituto de Ciência e Inovação em Engenharia Mecânica e Engenharia Industrial, Faculdade de Engenharia, Universidade do Porto, Rua Dr. Roberto Frias, s/n 4200- 465 Porto, Portugal; email: zhen.ma@fe.up.pt b Instituto de Ciência e Inovação em Engenharia Mecânica e Engenharia Industrial, Departamento de Engenharia Mecânica, Faculdade de Engenharia, Universidade do Porto, Rua Dr. Roberto Frias, 4200-465 Porto, Portugal; email: tavaresfe.up.pt Corresponding Author: Prof. João Manuel R. S. Tavares Faculdade de Engenharia da Universidade do Porto (FEUP) Departamento de Engenharia Mecânica (DEMec) Rua Dr. Roberto Frias, s/n, 4200-465 PORTO - PORTUGAL Tel: +315 22 5081487, Fax: +315 22 5081445 Email: tavares@fe.up.pt, Url: www.fe.up.pt/~tavares 2 Effective Features to Classify Skin Lesions in Dermoscopic images Abstract Features such as shape and color are indispensable to determine whether a skin lesion is a melanoma or not. However, there are no fixed guidelines to define which features are effective and how to combine them for classification. This lack of definition impedes the development of the automatic analyses of dermoscopic images. In this work, a search for effective features was carried out using a support vector machine. Three image databases were used to verify the feasibility and sensitivity of the automatic classification used. The results showed which features had a major influence on the classification performance, and confirmed the need to use various types of features in this process. Keywords Skin lesion; melanoma; ABCD rule; shape features; color features; feature selection. 3 1. Introduction Dermoscopy has been widely used for the diagnosis of skin lesions. The additional details provided by a dermoscope, compared to the inspection by the naked-eye, contribute to a significant improvement in the detection rate of melanoma (Binder et al., 1995; Argenziano et al. 2002; Kittler et al., 2002; Heymann, 2005). However, the experience of the dermatologist can significantly affect the diagnostic accuracy (Binder et al., 1995; Kittler et al., 2002); therefore, an effective automatic computer-aided model to analyze dermoscopic images is urgently needed. A scheme for such a model includes three major steps: first, the border of the skin lesion is identified on the image under study (Ma & Tavares, 2016; Silveira et al., 2009); then, features are extracted based on the segmentation, and with these features the lesion is classified into an appropriate category. Many mature algorithms have been proposed for the classification step, such as support vector machines (SVMs) and artificial neural networks (ANNs) (Kulkarni et al., 1998; Hastie et al., 2009); however, their performances rely on the features from the second step. Unfortunately, there is no recommended set of features for detecting melanomas in the current literature. Although many novel features have been proposed to improve the classification accuracy of these algorithms, the validation is normally carried out using different methods with different feature sets. The use of redundant or irrelevant features for classification not only decreases the computational efficiency but also affects the performance. Moreover, in many cases, cross validation is used to test a classification scheme, which can present over-optimistic results since images in the same database are prone to have similar imaging conditions and alike structures of skin lesions. Feature selection algorithms have been proposed to find the significant features, especially in the 4 area of bioinformatics (Guyon et al., 2002; Saeys, Inza, & Larranaga, 2007). Nevertheless, the key benefit of using such an algorithm is to reduce the dimensionality of features without loss of significant information (Bermingham et al., 2015). Limitations such as the choice of the evaluation metric and image database can affect the decision of whether a feature should be kept or not. In this work, effective features are explored without focusing on factors such as the accuracy of segmentation, the choice of the classification algorithm, and the representativeness of the training set. Three dermoscopic image databases were used in this study and the feasibility of automatic analysis on dermoscopic images was investigated based on the inter-database validation. This article is structured as follows: in the next section, the features selected for classification are introduced; then, in Section 3, the experiments are described and the results are presented and discussed; in the last section, the conclusions and perspectives for future work are pointed out. 2. Features The ABCD rule or ABCDE rule (Friedman, Rigel, & Kopf, 1985) is a clinical guideline to determine whether a skin lesion is a melanoma or a common nevus (Abbasi, Shaw, & Rigel, 2004; Friedman et al., 1985). This evaluation criteria is composed of the asymmetry of the lesion region (A), the geometric properties of the lesion border (B), the color of the skin lesion (C), the diameter of the lesion region (D), and the elevation/evolution of the skin lesion (E). The E rule requires follow-up inspections, and the other four criteria can be grouped into two categories: measures that reflect the geometric properties of a skin lesion (A, B, and D rules), and measures that are related with the lightness and chromatic information (C rule). The ABCD rule requires a 5 subjective evaluation of the different aspects of a skin lesion; hence, in order to automize the process in an expert system, these rules need to be quantified. 2.1 Shape features To measure the shape of a skin lesion, the following six features were adopted to reflect the B and D rules: perimeter of the lesion; area of the lesion; perimeter to area ratio; aspect ratio - defined as the width divided by height of the minimum rectangle that bounds the lesion region; extent ratio - defined as the area of the lesion region divided by the area of the minimum bounding rectangle; and solidity ratio - defined as the area of skin lesion divided by the area of the convex hull of the lesion boundary (Haidekker, 2010; Olsen, 2011). The first two measures describe the size of the skin lesion, while the remaining ones indicate its regularity. Besides these commonly used features, the total variation of the lesion border is adopted as the seventh feature to represent the smoothness: 𝑉 = 𝜅 𝑑𝑠& , (1) where the lesion border 𝐶 is parameterized as 𝑢 𝑠 , and 𝜅 = ∇ ∙ ∇+ ∇+ is the mean curvature with ∇ denoting the divergence operator. For the symmetry property - A rule, an ideal measure should be irrelevant to the position, size and orientation of the contour. Invariant image moments are well suited to these requirements (Flusser & Suk, 2006), and Hu’s invariant set has been shown to be effective in various applications of image analysis (Hu, 1962; Flusser & Suk, 2006). Therefore, the seven image moments in Hu’s invariant set were selected as the eighth to the fourteenth shape features for classification. 6 2.2 Color features Although the appearance of a skin lesion can vary considerably, information concerning its color is always the main visual feature distinguishing it from the neighboring skin. The changes of pigment inside a skin lesion are important for both segmentation and classification steps. An ideal color feature should capture the differences between a melanoma and a non-melanoma nevus. Additionally, the imaging conditions can appreciably affect the appearance of a skin lesion on dermoscopic images; therefore, the ideal color features should be able to handle the diverse lighting conditions in different image databases. The color information of a dermoscopic image is stored as a triplet 𝑟,𝑔,𝑏 for each image pixel. The three color channels in the RGB color space are correlated and consequently they are hard to be used to evaluate the differences between colors. To overcome this problem, the CIE L*a*b* and CIE L*u*v* color spaces were adopted (Gonzalez and Woods, 2007). The benefit of separating lightness information from chromaticity can considerably decrease the distortions caused by lightness when defining color features. Accordingly, the first ten color features were chosen as the means and standard deviations of the 𝐿∗,𝑎∗,𝑏∗,𝑢∗ and 𝑣∗ values of the skin lesion. In addition, given that 𝑎∗,𝑏∗ and 𝑢∗,𝑣∗ are the coordinates representing the position of a chromaticity in the color coordination system, the difference between a chromaticity and the mean chromaticity of the skin lesion is calculated as the Euclidean distances: 𝑑8 = 𝑎∗ − 𝑎8 ∗ : + 𝑏∗ − 𝑏8 ∗ :, (2) 𝑑: = 𝑢∗ − 𝑢8 ∗ : + 𝑣∗ − 𝑣8 ∗ :, (3) 7 where 𝑎8∗ , 𝑏8∗ , 𝑢8∗ and 𝑣8∗ are the means of 𝑎∗,𝑏∗,𝑢∗ and 𝑣∗ of the skin lesion and correspondingly 𝑎8∗,𝑏8∗ and 𝑢8∗,𝑣8∗ are the chromatic geometric centroids of the skin lesion. Then, the next four color features were defined as the mean and standard deviation of 𝑑8 and 𝑑: inside the skin lesion: 𝑑8, 𝑑8_𝜎, 𝑑:, and 𝑑:_𝜎. Moreover, the same 14 features of the neighboring healthy skin (the means and standard deviations of L∗,a∗,b∗,u∗,v∗ and the means and standard deviations of the two Euclidean distances to the geometric color centroids of the healthy skin) next to the skin lesion were added to the color features; these features not only provide the information of color transitions from normal skin to skin lesion, but also compensate the possible inaccuracy of the lesion border. Also, the lightness difference and the Euclidean distance between the geometric centroids of healthy skin and skin lesion were adopted as three additional color features: 𝑑B = 𝑎C ∗ − 𝑎8 ∗ : + 𝑏C ∗ − 𝑏8 ∗ :, (4) 𝑑D = 𝑢C ∗ − 𝑢8 ∗ : + 𝑣C ∗ − 𝑣8 ∗ :, (5) 𝑑E = 𝐿C − 𝐿8 , (6) where 𝐿C and 𝐿8 are the mean lightness of the neighboring skin and skin lesion, and 𝑎C∗, bC∗ , uC∗ , and vC∗ are the means of a∗, b∗, u∗, and v∗ of the neighboring skin. Additionally, color saturation was used to correlate the lightness with chromaticity: S = 0 if R + G + B = 0 1 − LMN O,P,Q ORPRQ B otherwise , (7) and the 32nd to 35th color features were defined as the mean and standard deviation of the saturation of the skin lesion (𝑠8, 𝑠8_𝜎) and the neighboring skin (𝑠C, 𝑠C_𝜎), plus the 36th feature as the ratio of the means of the two regions: 𝑑S = 𝑠8 𝑠C. (8) 8 Table 1 lists all the shape and color features used for selection. 3. Experiments 3.1 Training and testing Although an unbiased classification requires the skin lesions in the training set to be representative, different types of skin lesions can have large variations in shape and appearance, so it is difficult to define how typical a skin lesion is. Also, due to the size of the available image databases, a routine validation of a classifier or new feature is through the leave-one-out strategy. Consequently, the evaluation of the performance may be biased and the high sensitivity and specificity achieved with one image database may not hold true for another. In order to make the evaluation objective, three databases of dermoscopic images were used. Details of these databases are given in Table 2, with samples of melanoma and non- melanoma skin lesions shown in Figure 1. As can be seen from the Figs 1a to f, the images were acquired under different conditions and from patients of diverse origins. In these databases, the skin lesions were actually classified into more accurate sub-categories; but for simplicity, this additional information was ignored and each skin lesion was classified as either melanoma or non-melanoma. The ground truth of lesion borders were provided in all the databases, which were manually segmented by qualified technicians. In order to obtain the statistics of neighboring healthy skin referred to in Section 2.2, a 10-pixel-wide band next to the lesion border was used; this outside band covers a moderate region that is sufficient to reflect the transition of healthy skin to skin lesion in the images. Given the sizes of the databases, we adopted the strategy to choose one database as the testing set and combined the other two as the training set. However, if the training set was 9 formed by the 1st and 2nd image databases shown in Table 2, the total number of melanomas (69) would not be large enough to be used for extracting the differential information between melanoma and non-melanoma nevi. Consequently, the classification would not be able to achieve a good performance on the 3rd database, particularly with the fact that the 3rd database contains Spitz and Reed nevi (Lyon, 2010; Yoradjian et al., 2012). Thus, the other two cases were adopted: the combination (Comb 1) that used the 1st and 3rd databases as the training set and the 2nd database as the testing set; and the combination (Comb 2) that used the 2nd and 3rd databases as the training set and the 1st database as the testing set. The next problem to solve was related to the different magnitudes of the 50 features chosen for classification; for example, the value of lightness ranges from 0 to 100, while saturation only varies from 0 to 1. A feature scaling step was necessary to balance such differences. Among the features to be selected, 31 were related to the L, a∗, b∗, u∗, and v∗ channels, and since their region-based values were already at similar intervals. These features were kept unchanged and then the remaining 19 features were scaled as follows: x = 50 ∗ VWVXYZ VX[\WVXYZ , if xLMN ≠ xL`V 0, if xLMN = xL`V , (9) where xLMN and xL`V are the minimum and maximum values of feature x in the training set. The support vector machine was chosen for this two-class classification, due to its effectiveness already shown in diverse areas of studies (Chu & Wang, 2005; Filho et al., 2015), and the radial basis function kernel was adopted for its flexibility, stability and general popularity (Celebi et al., 2007). Accordingly, three parameters had to be defined: 𝐶, 𝜀, and 𝛾. Parameter 𝐶 controls the trade-off between the error of a classifier on the 10 training set and the margin between classes; parameter 𝜀 determines the accuracy of the approximation; and parameter 𝛾 is related to the values of feature vectors and determines the behaviors of kernel functions for approximation. Values of these parameters can appreciably affect the performance of classification; therefore, to assure a stable evaluation of a feature combination, the values of 𝛾, 𝐶 and 𝜀 of the SVM were fixed as 1𝑒 − 3, 500 and 1 in the experiments. 3.2 Pre-selection criteria In order to find the best combination among the 14 shape features and 36 color features with a training set and a testing set, an exhaustive search requiring 2EC classifications would have to be made. This however could lead to an unfeasibly long computation time. To overcome this problem, the standard deviation of a feature would not be used if the mean of that feature was not used; for example, if the mean of a∗ of the skin lesion (a8∗ ) was not used for classification then the standard deviation of a∗ of the skin lesion (a8∗_σ) would not be used either. This strategy was adopted because there are 15 measures with both their means and standard deviations chosen as the features for selection. While the standard deviation indicates the variations of a measure, it can also be reflected by other features; for example, the variation of a* channel inside the skin lesion is partly reflected by the mean of distance d1, This strategy decreases the total number of classification to 2EC ∙ B D 8E . In addition, since chromaticity is described by two-dimensional coordinates in the color spaces, the following strategies were implanted: features of a∗ channel are bound with the features of b∗ channel; for example, if a8∗ was used, b8∗ would be used; if a8∗_σ was used, then b8∗_σ would be. Likewise, features of u∗ and v∗ were bound for use; and features related to d8 and d: were bound; and the same for dB and dD. 11 Furthermore, the following procedure was used to implement separate searches among the shape features and color features: In the first phase, all the color (shape) features were used for classification, and the combinations of shape (color) features were searched to find the optimal set; then, with the optimal set of shape (color) features, the combinations of color (shape) features were reversely searched to find the best match. With the color (shape) features selected from the second phase, we can again search the combinations of shape (color) features to find the best match. Hence, the two-phase procedure can be iterated, and the best performance of classification would be improved after each iteration. Thus, if the search is to find the best color features for a set of shape features, the number of classifications would be 2:B ∙ B D 8C ; and if the searching is to find the best shape features for a set of color features, the number would be 28D. In the experiments, the performance of classification was evaluated based on three measures: overall accuracy (OA), which is the percentage of the skin lesions that are correctly classified; sensitivity of a classification (S1) that is defined as the percentage of correctly classified melanomas among all the melanomas; and specificity (S2), which is defined as the percentage of correctly classified non-melanoma nevi among all the non- melanoma nevi. Higher values of these indices indicate a better classification. 3.3 First phase Following the procedure defined above, all the shape features were used for classification then the search was among the combinations of color features. For Comb 1, the highest overall accuracy was 83.5% and was achieved by two sets of color features. Both of them contained the mean lightness of skin lesion (𝐿8), lightness difference (𝑑E), standard deviation of lightness inside the skin lesion (𝐿8_𝜎), mean saturation of the skin lesion (𝑠8), 12 mean of 𝑑8 and mean of 𝑑: inside the skin lesion (𝑑8 and 𝑑:); while one set included two extra features: mean lightness of neighboring skin (𝐿C) and ratio of mean saturations (𝑑S). With all the shape features and the six mutual color features, the sensitivity was 87.5% and the specificity was 82.5%; while with the additional two color features, these two indices were 77.5% and 85.0%, respectively. The highest sensitivity achieved in this phase was 97.5%, but the corresponding specificity was only 40.0%, indicating that many non-melanoma nevi were wrongly classified as melanomas. If both sensitivity and specificity were considered, a ranking on the average of these two values gave the best combination that achieved 92.5% of sensitivity and 80.0% of specificity; the set of color features has 82.5% of overall accuracy and was composed of six elements: 𝐿8, 𝑠8, 𝑑8, 𝑑:, 𝑑E and 𝑑S. However, when we performed the classification for Comb 2 with all the shape features and the aforementioned three sets of color features, the overall accuracy of classification was only 55.0%, 54.0% and 61.0%, respectively. For Comb 2, the best overall accuracy of classification was 72.0% with 62.1% sensitivity and 76.1% specificity, achieved by a set of three color features: mean of 𝑎∗ and mean of 𝑏∗ inside the skin lesion (𝑎8∗ and 𝑏8∗), and ratio of mean saturations (𝑑S). Similarly, if this feature set is used for Comb 1, the overall accuracy was only 52.5% with 45.0% of sensitivity and 54.4% of specificity. On the other hand, if all the color features were used for classification, the best results after searching among the combinations of shape features gave 84.5% of overall accuracy for Comb 1 with 60.0% of sensitivity and 90.6% of specificity, and 70% of overall accuracy for Comb 2 with 34.5% of sensitivity and 84.5% of specificity. Like the findings above, the set of shape features that achieved good results for Comb 1 did not perform well for Comb 2, and vice versa. 13 The results show that the sets of features that achieved the highest overall accuracy were not the ones with the highest sensitivity or specificity; consequently, a measure is needed to decide which set of features should be used in the second phase of the search. Given the composition of the training sets of Comb 1 and Comb 2, the total number of wrong classifications (WN) in the two tests was chosen as the index to rank the features. Accordingly, seven sets of color features (found with all the shape features) and nine sets of shape features (found with all the color features) were chosen for the second phase of the search. Table 3 lists these 16 feature sets and the indices of the corresponding classification; the performance of these feature sets was moderate, but more balanced. The minimum of WN (Wrong Number) equals 70 when all the shape features were used, the set of color features included: mean lightness of skin lesion (L8), lightness difference (dE), mean saturation of neighboring skin and skin lesion (sC and s8), mean of d8 and mean of d: inside the skin lesion (d8 and d:). The minimum of WN found with all the color features was 72, achieved by five sets of shape features. All of them contained the following three features: aspect ratio, smoothness of lesion border, and area of skin lesion; the difference between them was the choice of the Hu's seven invariant image moments. The frequency of occurrence (FO) of each feature appearing in the feature sets with 𝑊𝑁 ≤ 80 and 𝑊𝑁 ≤ 90 were calculated and are illustrated in Figure 2. For the color features, mean lightness of skin lesion (𝐿8), mean lightness of neighboring skin (𝐿C), lightness difference (𝑑E), standard deviation of lightness inside the skin lesion (𝐿8_𝜎), mean saturation of the neighboring skin and skin lesion (𝑠C and 𝑠8), mean of 𝑑8 and mean of 𝑑: inside the skin lesion (𝑑8 and 𝑑:), and means of 𝑎∗,𝑏∗, 𝑢∗ and 𝑣∗ inside the skin lesion (𝑎8∗, 𝑏8∗, 𝑢8∗, 𝑣8∗) were features that had high frequencies. For the shape features, aspect ratio, smoothness of lesion border, and area of skin lesion had the highest 14 frequencies. Additionally, the difference found among the frequencies of Hu’s seven image moments was small, which implies that the effectiveness of these moments may vary when paired with other features, but generally they are of equal importance. 3.4 Second phase This phase included 16,384 classifications for each of the seven sets of color features indicated in Table 3, and 472,392 classifications for each of the nine sets of shape features. With the color features fixed, the optimal set of shape features found in this phase improved the performance for both Comb 1 and Comb 2. For example, for the first set of color features in Table 3, the optimal set of shape features achieved 78.0% of overall accuracy for Comb 2; and for the second set of color features referred to in Table 3, the highest overall accuracy for Comb 1 increased to 90.5% with 92.5% of sensitivity and 90.0% of specificity. Table 4 lists the best matches of shape and color features found in this phase. Like in the first phase, the effective shape features were different for each of the seven sets of color features. Figure 3 shows the FO of each shape feature in the feature sets whose WN ≤ 50,60,70,80,90, and Figure 4 illustrates the FO of each color feature based on the results for the nine sets of shape features. Aspect ratio, smoothness of lesion border, and area of the lesion region were the three shape features that always had high frequencies, especially among the feature sets with overall accuracy above 80.0%. This agrees with the finding of the first phase and indicates that these three shape features are effective and have a large influence on the performance of the classification. Figure 3 shows that the ranking of frequencies is generally the same, and when the WN increases, the difference between frequencies becomes smaller since more and more combinations 15 of shape features can achieve the same lower performance. For the color features, mean saturation of neighboring skin and skin lesion (𝑠C and 𝑠8), mean of 𝑑8 and mean of 𝑑: inside the skin lesion (𝑑8 and 𝑑:), and lightness difference (𝑑E) were the ones with the highest frequencies, especially when the overall accuracy was greater than 80.0%. Similarly, the difference of frequencies among color features decreases when the WN increases. These five color features were also included in the ones with high frequencies of the first phase, which confirms their effectiveness; the missing features from the first phase, such as 𝐿8, 𝑎8∗ and 𝑏8∗, are the ones that directly describe the color information; this phenomenon was probably caused by the large variations in the appearances of skin lesions. 3.5 Further selection Since iterating the two-phase procedure generates a no-worse result, a further search was carried out based on the results of the second phase. The sets of features used in this phase are listed in Table 4. After this search, we confirmed that most of the combinations in Table 4 were already the optimal matches, however, we found two extra combinations that were able to achieve equal performance, and are listed in Table 5; the first one is a set of color features that achieved the same WN as the one presented in Table 4, and the second one is a set of shape features with which the classification can achieve an even better performance. Nevertheless, another search with these two sets of features confirmed that the matches of shape features and color features in Table 5 were already the best and no further sets of features would able to achieve an equal or better performance. Hence, it is safe to conclude that with the search procedures performed the sets of features in Tables 4 and 5 were the optimal ones for classification. 16 Then, the frequencies of the shape features in the feature sets whose 𝑊𝑁 ≤ 50,60,70,80,90 were calculated. In the calculation we excluded the feature sets that had one of the following conditions: 𝑆1 ≤ 60.0% for Comb 1, or 𝑆1 ≤ 40.0% for Comb 2, or 𝑆2 ≤ 70.0% for either Comb 1 or Comb 2. These exclusions guaranteed that only the ones that had a good performance in all the indices of classification were taken into account. Similarly, the frequencies of color features were calculated based on the results of the second phase and the extra search; the results are shown in Figs. 5 and 6. The images show that the shape and color features with the highest frequencies in this phase are in line with the findings of the first and the second phase searches; hence, the effectiveness of these features was once again demonstrated. 3.6 Discussion Although the training set of Comb 2 contains more samples of skin lesions, the classification achieved a better performance for Comb 1. One possible reason is due to the low imaging resolution of the 1st database, which considerably affects the calculation of color features. However, the 2nd database contains images with incomplete profiles of skin lesions, and so the lesion borders on these images are inevitably inaccurate and can lead to wrong values of shape features. The inaccuracy of shape features and color features both have negative impacts on the classification; a pertinent question is which type of feature has the greater influence on the classification. Following up this idea, classification was carried out using features of just one type. The results showed that with only the color features, the best overall accuracy was around 83% for Comb 2, with sensitivity ranging from 48.3% to 72.4% and specificity from 87.3% to 97.2%, respectively. On the other hand, with only the shape features, the best overall accuracy 17 was about 59% with 31.0% sensitivity (also the highest value found for this index) and 70.4% specificity, respectively. Table 6 indicates the best sets of shape features and color features found according to the WN. The Table shows that the minimum of WN with features of only one type was not much different to the minimum found using both types. However, the sensitivity and the specificity with both types were higher, and the performance of classification was more stable for different image databases. The results also indicated that the color features are more robust for classification, which confirms the fact that these features are region-based. The empirical search with Comb 1 and Comb 2 identified the features that are effective to achieve a satisfactory classification. However, it is worth pointing out that only using these features does not guarantee the best performance. In fact, classification using the five color features and three shape features that have the highest frequencies only achieved a moderate result with 83.0% of overall accuracy for Comb 1 and 73.0% of overall accuracy for Comb 2; remembering that there are features of low frequencies appearing in the set of features with the best performance. Nevertheless, despite this randomness, these eight features are most likely to achieve a good classification performance, and so should be prioritized in selection. 4. Conclusions The features used in the automatic analysis of dermoscopic images have a critical influence on the performance of classification. However, a set of features that achieves good results on one database may not have an equal performance on another database, and a “redundant” feature may become indispensable to correctly detect a specific type 18 of skin lesion. In this work, we aimed to find the effective features for detecting melanomas experimentally. The three image databases used in the study were from different origins and acquired under diverse imaging conditions, which provides an objective basis for evaluation. The results obtained confirmed the effective use of both shape and color features and suggested the need to combine them to acquire high classification accuracy and robustness. A comprehensive study on features for classification has many practical constraints, because image databases from diverse studies are always different and the diagnostic accuracy of dermoscopy itself is not 100% even under the optimal exam conditions (Kittler et al., 2002). Nonetheless, the findings in our work showed that the performance of the automatic analysis of skin lesions in dermoscopic images is comparable to experienced dermatologists. Furthermore, the ABCD rule may not always be able to detect a small-sized melanoma or a melanoma with a regular shape and homogeneous color (Grin et al., 1990). In addition, there are some specific features that can be effective in classifying a particular type of skin lesion; for example, ridges and furrows were shown to be effective in detecting acral lentiginous melanoma (Iyatomi et al., 2008; Bradford et al., 2009; Yang et al., 2017). These subjects are challenges to be explored, and future work will continue to focus on solving these issues and finding more effective features. Acknowledgements This work was funded by European Regional Development Funds (ERDF), through the Operational Program ‘Thematic Factors of Competitiveness (COMPETE), and Portuguese Funds, through “Fundação para a Ciência e a Tecnologia” (FCT), under the 19 project: FCOMP-01-0124-FEDER-028160/PTDC/BBB-BMD/3088/2012. The first author also thanks FCT for the post-doc grant: SFRH/BPD/97844/2013. Authors gratefully acknowledge the funding of Project NORTE-01-0145-FEDER- 000022 - SciTech - Science and Technology for Competitive and Sustainable Industries, co-financed by “Programa Operacional Regional do Norte” (NORTE2020), through “Fundo Europeu de Desenvolvimento Regional” (FEDER). References Abbasi, N., Shaw, H., & Rigel, D. (2004). Early Diagnosis of Cutaneous Melanoma. JAMA: The Journal of Medical Association, 292, 2771–2776. Argenziano G., Soyer, H. P., De Giorgi, V., Piccolo, D., Carli, P., Delfino, M., Ferrari, A., Hofmann-Wellenhof, R., Massi, D., Mazzocchetti, G., Scalvenzi, M., & Wolf, I. H. (2002). Dermoscopy: a tutorial. EDRA Medical Publishing & New Media. Bermingham, M. L., Pong-Wong, R., Spiliopoulou, A., Hayward, C., Rudan, I., Campbell, H., Wright, A. F., Wilson, J. F., Agakov, F., Navarro, P., & Haley, C. S. (2015). Application of high-dimensional feature selection: evaluation for genomic prediction in man. Scientific Reports, 5, 10312. Binder, M., Schwarz, M., Winkler, A., Steiner, A., Kaider, A., Wolff, K., & Pehamberger, H. (1995). Epiluminescence microscopy. A useful tool for the diagnosis of pigmented skin lesions for formally trained dermatologists. Archives of Dermatology, 131, 286–291. Bradford, P. T., Goldstein, A. M., McMaster, M. L., Tucker, M. A. (2009). Acral lentiginous melanomaincidence and survival patterns in the United States, 1986- 2005. JAMA Dermatology, 145, 427-434. 20 Celebi, M. E., Kingravi, H. A., Uddin, B., Iyatomi, H., Aslandogan, Y. A., Stoecker, W. V., Moss, R. H. (2007). A methodological approach to the classification of dermoscopy images. Computerized Medical Imaging and Graphics, 31, 362-373. Celebi, M. E., Wen, Q., Hwang, S., Iyatomi, H., & Schaefer G. (2013) Lesion border detection in dermoscopy images using ensembles of thresholding methods. Skin Research and Technology, 19, e252-e258. Chu, F., & Wang, L. (2005). Applications of support vector machines to cancer classification with microarray data. International Journal of Neural Systems, 15, 475-484. Filho, M., Ma, Z. & Tavares, J. M. R. S. (2015). A review of the quantification and classification of pigmented skin lesions: from dedicated to hand-held devices. Journal of Medical Systems, 39: 177. Flusser, J., & Suk, T. (2006). Rotation moment invariants for recognition of symmetric objects. IEEE Transactions on Image Processing, 15, 3784–3790. Friedman, R. J., Rigel, D. S., & Kopf, A. W. (1985). Early detection of malignant melanoma: the role of physician examination and self-examination of the skin. CA: A Cancer Journal for Clinicians, 35, 130–151. Gonzalez, R. C., & Woods, R. E. (2007). Digital image processing (3rd ed.). Prentice Hall. Grin, C. M., Kopf, A. W., Welkovich, B., Bart, R. S., & Levenstein, M. J. (1990). Accuracy in the clinical diagnosis of malignant melanoma, Archives of Dermatology, 126, 763-766. Guyon, I., Weston, J., Barnhill, S., & Vapnik, V. (2002). Gene selection for cancer classification using support vector machines. Machine Learning, 46, 389-422. 21 Haidekker, M. (2010). Advanced biomedical image analysis. (1st ed.) New Jersy: John Wiley & Sons, (Chapter 9). Hastie, T., Tibshirani, R., & Friedman, J. (2009). The elements of statistical learning: data mining, inference, and prediction. (2nd ed.). Springer Science & Business Media. Heymann, W. R. (2005). Clinical and microscopic diagnosis of melanoma. Journal of the American Academy of Dermatology, 52, 133–134. Hu, M. K. (1962). Visual Pattern Recognition by Moment Invariants. IRE Transactions on Information Theory, 8, 179–187. Iyatomi, H., Oka, H., Celebi, M. E., Ogawa, K., Argenziano, G., Soyer, H. P., Koga, H., & Saida, T. (2008). Computer-based classification of dermoscopy images of melanocytic lesions on acral volar skin. Journal of Investigative Dermatology, 128, 2049-2054. Kittler, H., Pehamberger, H., Wolff, K., & Binder, M. (2002). Diagnostic accuracy of dermoscopy. Lancet Oncology, 3, 159–165. Kulkarni, S. R., Lugosi, G., & Venkatesh, S. S. (1998). Learning pattern classification-A survey. IEEE Transactions on Information Theory, 44, 2178–2206. Lyon, V. B. (2010). The Spitz Nevus: Review and Update. Clinics in Plastic Surgery, 37, 21–33. Ma, Z., & Tavares, J. M. R. S. (2016). A Novel Approach to Segment Skin Lesions in Dermoscopic Images Based on a Deformable Model. IEEE Journal of Biomedical and Health Informatics, 20, 615–623. Mendonca, T., Ferreira, P. M., Marques, J. S., Marcal, & A. R., Rozeira, J. (2013). PH² - a dermoscopic image database for research and benchmarking. Conf Proc IEEE Eng 22 Med Biol Soc. 5437-5440. Olson, E. (2011). Shape factors and their use in image analysis–part 1: theory. J GXP Compliance, 15, 85–96. Saeys, Y., Inza, I., & Larranaga, P. (2007). A review of feature selection techniques in bioinformatics. Bioinformatics, 23, 2507–2517. Silveira, M., Nascimento, J. C., Marques, J. S., Marçal, A. R. S., Mendonça, T., Yamauchi, S., Maeda, & J., Rozeira, J. (2009). Comparison of segmentation methods for melanoma diagnosis in dermoscopy images. IEEE Journal on Selected Topics in Signal Processing, 3, 35–45. Yang, S., Oh B., Hahm S., Chung, K. Y., & Lee, B. U. (2017). Ridge and furrow pattern classification for acral lentiginous melanoma using dermoscopic images. Biomedical Signal Processing and Control, 32, 90-96. Yoradjian, A., Simoes, M. M., Enokihara, S., & Paschoal, F. M. (2012). Nevo de spitz e nevo de reed. Anais Brasileiros de Dermatologia, 87, 349–359. 23 FIGURE CAPTIONS Fig. 1 Examples of dermoscopic images from the three databases overlapped with the lesion borders (blue contours): (a) non-melanoma in the 1st database, (b) melanoma in the 1st database, (c) non-melanoma in the 2nd database, (d) melanoma in the 2nd database, (e) non-melanoma in the 3rd database and (f) melanoma in the 3rd database. Fig. 2 Illustration of the frequencies of features based on the first search phase: (a) color features and (b) shape features. Fig. 3 Illustration of the frequencies of shape features based on the second search phase. Fig. 4 Illustration of the frequencies of color features based on the second search phase. Fig. 5 Illustration of the frequencies of shape features based on the second search phase and the extra searches. Fig. 6 Illustration of the frequencies of color features based on the second search phase of searching and the extra searches. 24 TABLES Table 1 Features for selection Type Number Details Shape features 14 Ratio of perimeter to area of the skin lesion; smoothness of the lesion border; the seven Hu’s invariant image moments; perimeter of the lesion border; area of the skin lesion; aspect ratio; extent ratio; solidity ratio. Color features 36 𝐿8, 𝑎8∗, 𝑏8∗, 𝑢8∗, 𝑣8∗, 𝐿C, 𝑎C∗, 𝑏C∗, 𝑢C∗, 𝑣C∗, 𝑑E, 𝑑B, 𝑑D, 𝑠8, 𝑠C, 𝑑S, 𝐿8_𝜎, 𝑎8∗_𝜎, 𝑏8∗_𝜎, 𝑢8∗_𝜎, 𝑣8∗_𝜎, 𝐿C_𝜎, 𝑎C∗_𝜎, 𝑏C∗_𝜎, 𝑢C∗_𝜎, 𝑣C∗_𝜎, 𝑠8_𝜎, 𝑠C_𝜎, 𝑑8, 𝑑:, 𝑑w (the mean distance to the geometric centroid 𝑎C∗,𝑏C∗ of neighboring healthy skin), 𝑑x (the mean distance to the geometric centroid 𝑢C∗,𝑣C∗ of neighboring healthy skin), 𝑑8_𝜎, 𝑑:_𝜎, 𝑑w_𝜎, 𝑑x_𝜎. 25 Table 2 Image databases used in the study Database Size Melanoma Origin 1 100 29 Private clinics in US and Australia (Ma and Tavares, 2016; Celebi et al., 2013) 2 200 40 PH2 image database (Mendonca et al., 2013) 3 404 105 Interactive atlas of dermoscopy (Argenziano et al. 2002) 26 Table 3 Feature sets chosen for the second search phase Features WN3 Comb 1 Comb 2 OA3 S13 S23 OA3 S13 S23 1000000000100110000000000000110000001 70 82.5% 70.0% 85.6% 65.0% 41.4% 74.6% 1000000000100101000000000000110000001 74 82.5% 92.5% 80.0% 61.0% 44.8% 67.6% 0000010000100111000000000000110000001 75 81.5% 70.0% 84.4% 62.0% 44.8% 69.0% 1000010000100110100000000000110000001 75 81.5% 60.0% 86.9% 62.0% 51.7% 66.2% 1111100000000110100000000010110000001 75 78.5% 25.0% 91.9% 68.0% 44.8% 77.5% 1000000000100110100000000010110000001 76 80.0% 45.0% 88.8% 64.0% 51.7% 69.0% 1000010000000110000000000000110000001 76 80.5% 60.0% 85.6% 63.0% 41.4% 71.8% 010100011011002 72 84.5% 60.0% 90.6% 59.0% 34.5% 69.0% 010100101011002 72 84.5% 60.0% 90.6% 59.0% 34.5% 69.0% 010101001011002 72 84.5% 60.0% 90.6% 59.0% 34.5% 69.0% 010101010011002 72 84.5% 60.0% 90.6% 59.0% 34.5% 69.0% 010101011011002 72 84.5% 60.0% 90.6% 59.0% 34.5% 69.0% 110111100111002 73 81.5% 57.5% 87.5% 64.0% 31.0% 77.5% 110111101111002 73 81.5% 57.5% 87.5% 64.0% 31.0% 77.5% 110111110111002 73 81.5% 57.5% 87.5% 64.0% 31.0% 77.5% 110111111111002 73 81.5% 57.5% 87.5% 64.0% 31.0% 77.5% 1 Combination of color features represented by a binary string, ‘0’ – unused and ‘1’ – used. From left to right: the sequence of features is listed in Table 1. 2 Combination of shape features represented by a binary strings, ‘0’ – not used and ‘1’ – used. From left to right: the sequence of features is listed in Table 1. 3 WN – Wrong number; OA – Overall accuracy; S1 – Sensitivity; S2 – Specificity. 27 Table 4 Feature sets selected for another search Features W N Comb 1 Comb 2 OA S1 S2 OA S1 S2 1000000000100110000 000000000110000001 011001000011102 45 90.0% 75.0% 93.8% 75.0% 44.8% 87.3% 011001001011102 90.0% 75.0% 93.8% 75.0% 44.8% 87.3% 110100011011002 90.5% 72.5% 95.0% 74.0% 44.8% 85.9% 111100011011002 46 90.0% 72.5% 94.4% 74.0% 44.8% 85.9% 111101101011002 89.5% 72.5% 93.8% 75.0% 44.8% 87.3% 110101000011002 90.0% 67.5% 95.6% 74.0% 44.8% 85.9% 1000000000100101000 000000000110000001 111001001011002 45 90.0% 95.0% 88.8% 75.0% 34.5% 91.5% 111000001011002 46 89.0% 95.0% 87.5% 76.0% 37.9% 91.5% 010100011011002 0000000000000101000000000010000000001 46 88.5% 67.5% 93.8% 77.0% 41.4% 91.5% 010100101011002 0000000000000000000000000000110011001 45 89.5% 62.5% 96.3% 76.0% 65.5% 80.3% 010101010011002 0000000000000000000000000000111100001 46 89.5% 62.5% 96.3% 75.0% 62.1% 80.3% 1 Combination of color features with the sequence defined in Table 1. 2 Combination of shape features with the sequence defined in Table 1. 28 Table 5 Feature sets selected for an extra search Features WN Comb 1 Comb 2 OA S1 S2 OA S1 S2 110101000011002 1000000000100110000000000000110000001 46 90.0% 67.5% 95.6% 74.0% 44.8% 85.9% 0000000000000001000000000000110000001 88.5% 60.0% 95.6% 77.0% 65.5% 81.7% 0000000000000101000000000010000000001 010000010011002 43 90.0% 72.5% 94.4% 77.0% 44.8% 90.1% 010100011011002 46 88.5% 67.5% 93.8% 77.0% 41.4% 91.5% 1 Combination of color features with the sequence defined in Table 1. 2 Combination of shape features with the sequence defined in Table 1. 29 Table 6 Best feature sets based on the classification with features of one type Shape features / Color features Comb 1 Comb 2 OA S1 S2 OA S1 S2 1 000101000110102 90.0% 65.0% 96.3% 72.0% 6.9% 98.6% 0000010011000001000001000000000000001 86.5% 57.5% 93.8% 80.0% 31.0% 100.0% 2 100011100110102 89.5% 60.0% 96.9% 72.0% 3.4% 100.0% 0000011100000011000001000000000000001 87.0% 57.5% 94.4% 79.0% 37.9% 95.8% 3 000100111110102 89.5% 70.0% 94.4% 71.0% 3.4% 98.6% 0110010000000001000001000000000000001 90.5% 70.0% 95.6% 72.0% 10.3% 97.2% 4 000110010110102 89.0% 70.0% 93.8% 72.0% 3.4% 100.0% 0000011100000010000001000000000000001 87.5% 52.5% 96.3% 77.0% 27.6% 97.2% 5 000111001011102 89.5% 55.0% 98.1% 71.0% 6.9% 97.2% 0000010011000000000001000000000000001 87.0% 60.0% 93.8% 77.0% 20.7% 100.0% 6 001011011101102 89.5% 55.0% 98.1% 71.0% 3.4% 98.6% 0110010000100000000001000000110011001 88.0% 52.5% 96.9% 74.0% 37.9% 88.7% 1 Combination of color features with the sequence defined in Table 1. 2 Combination of shape features with the sequence defined in Table 1. 30 FIGURES Figure 1a Figure 1b 31 Figure 1c Figure 1d 32 Figure 1e Figure 1f 33 Figure 2a Figure 2b 34 Figure 3 35 Figure 4 36 Figure 5 37 Figure 6 
work_so3kemvffbawlapikcxbap5khe	----	An expert system to identify co-regulated gene groups from time-lagged gene clusters using cell cycle expression data Expert Systems with Applications 37 (2010) 2202–2213 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a An expert system to identify co-regulated gene groups from time-lagged gene clusters using cell cycle expression data Li-Ching Wu a, Jhih-Long Huang b, Jorng-Tzong Horng b,d,*, Hsien-Da Huang c a Institute of System Biology and Bioinformatics, National Central University, Taiwan b Department of Computer Science and Information Engineering, National Central University, Taiwan c Institute of Bioinformatics, National Chiao-Tung University, Taiwan d Department of Bioinformatics, Asia University, Taiwan a r t i c l e i n f o Keywords: Expert system Data mining Gene expression Bioinformatics 0957-4174/$ - see front matter � 2009 Elsevier Ltd. A doi:10.1016/j.eswa.2009.07.053 * Corresponding author. Address: Department of Bi Taiwan. E-mail address: horng@db.csie.ncu.edu.tw (J.-T. Ho a b s t r a c t Motivation: The analysis of time series gene expression data can provide us with the opportunity to find co-regulated genes that show a similar expression patterns under a contiguous subset of experimental conditions. However, these co-regulated genes may behave almost independently under other condi- tions. Furthermore, the similarity in the expression pattern might be time-shifted. In that case, we need to be concerned with grouping genes that share similar expression patterns under a contiguous subset of conditions and where the similarity in expression pattern might have time lags. In addition, to be consid- ered a time-shifted similar pattern, because co-regulated genes in a biological process may show a peri- odic pattern in their cell cycle expression, we also should group genes with periodic similar patterns over multiple cell cycles. If this is carried out, we can regard such grouped genes as cell-cycle regulated genes. Results: We propose a method that follows the q-cluster concept [Ji, L., & Tan, K. L. (2005). Identifying time-lagged gene clusters using gene expression data. Bioinformatics, 21(4), 509–516] and further advances this approach towards the identification of cell-cycle regulated genes using cell cycle micro- array data. We used our method to cluster a microarray time series of yeast genes to assess the statisti- cally biological significance of the obtained clusters we used the p-value obtained from the hypergeometric distribution. We found that several clusters provided findings suggesting a TF–target relationship. In order to test whether our method could group related genes that other methods have found difficult to group, we compared our method with other measures such as Spearman Rank Correla- tion and Pearson Correlation. The results of the comparison demonstrate that our method indeed could group known related genes that these measures regard as having only a weak association. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction DNA microarray technology enables the simultaneous study of gene expression levels on a large scale. Expression level is the log- arithm of the abundance of the mRNA of a gene under a specific condition. The gene expression data of a microarray is arranged as a data matrix. Each gene corresponds to one row and each condition to one column. Each element of this matrix represents the expression level of a gene under a specific condition. The conditions of a microarray may be different time points, different environmental conditions or different organs. The analysis of microarray data has facilitated the study of genetic regulatory networks. The correlation patterns of genes with experimental conditions can be used to identify the networks that are comprised ll rights reserved. oinformatics, Asia University, rng). of correlated genes and thus how the correlated genes interact with each other. Clustering methods can be applied to either the genes or the conditions of the microarray matrix separately. However, some problems occur when applying clustering to the analysis of gene expression data. A set of genes may simultaneously activate a par- ticular biological process over certain contiguous conditions but behave independently under other condition. Therefore, we need to group genes that have similar behavior under a specific subset of the conditions. Clustering can not satisfy this requirement. The biclustering method is a technique that makes this possible and al- lows the grouping of genes and conditions simultaneously within a data matrix. The goal of biclustering is to find a bicluster that is a subset of genes that show similar behavior under a specific subset of the conditions (Cheng & Church, 2000). Thus, genes in the same bicluster are co-expressed and further are likely to be co-regulated. In the analysis of time-series expression data, a set of genes may activates a particular biological process over a certain contiguous http://dx.doi.org/10.1016/j.eswa.2009.07.053 mailto:horng@db.csie.ncu.edu.tw http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 2203 set of conditions instead under a discrete condition. In such a case, we should find biclusters for a contiguous subset of conditions. However, in fact, co-expressed genes do not regulate each other simultaneously but only after a certain time lag. That is to say, there is a transcriptional time lag whereby the regulator gene takes time to express its protein product and a further delay occurs as the target gene responds to the regulator protein. Hence, because of the transcriptional time lag of co-regulated genes, we need to take time-lagged co-regulated genes into consideration when forming biclusters. In addition, when considering time-lagged sim- ilarity of expression patterns between genes, it is necessary to con- sider biclusters with coherent values for both the rows and columns of the expression matrix. This is because their expression relies on a promoter that is a structural regulatory sequence recog- nized by a TF of the RNA polymerase holoenzyme. The reason that co-expressed genes share a common sequence within their pro- moter will therefore result in shared expression. However, the rec- ognition efficiency of this TF is not the same for every gene having the same promoter. This condition leads to biclusters having a vari- ety of coherence values. There are two kinds of coherence values for a bicluster. The first is a shifted similarity pattern that can be viewed as based on an additive model. The second is a scaled sim- ilarity pattern that can be viewed as based on a multiplicative model. In a mathematical sense, scaling and vertical shifting of the expression level can be referred to as linear transformations. Consider two time-series x and y. In this case y is a linear transfor- mation of x if it can be expressed as y ¼ mx þ b. Many biological processes show periodic pattern such as the cell cycle process, therefore it is useful to find periodically regu- lated genes with similar periodic patterns of expression. We can then use these cell-cycle regulated genes to map the transcrip- tional regulatory network that controls the cell cycle. A suffix tree is a data structure that contains all suffixes of a string s. It has been widely used for string matching and exact se- quence comparison (Ukkonen, 1995). This approach was used to develop an algorithm for building a suffix tree that runs in time OðnÞ. Once a suffix tree is built, most problems can be solved in lin- ear-time using it. The suffix tree built for a set of strings is called a generalized suffix tree (Gusfield, 1997). In order to avoid creating empty suffixes, we usually append to s an extra character $ before the building of the suffix tree. The key feature of the generalized suffix tree is that any leaves in this tree contain two pieces of infor- mation: The first is the string number and the second is the start- ing position of a suffix that makes up this string. 2. Related work A great many approaches have been developed for the identifi- cation of co-regulated genes from microarrays. The correlation method is one that determines whether two variables have a strong global association, but this approach does not take time lag issue into consideration. However, another correlation method, the Cross-Correlation Method (Kato et al., 2001), is different from the traditional Pearson Correlation approach. It takes into account time-lagged co-regulations when testing the expression levels of two genes and identifies if there is a significant linear relationship. However, this approach still only determines whether two genes have strong global similarity but does not determine local similar- ity. Yet another approach is the edge detection method (Filkov et al., 2001), which scans through each gene’s expression curve to find where big changes in expression level (edges) occur, and then sums up the number of edges making up the expression curves of the two genes that have the same direction in order to generate a similarity score. Gene pairs are likely to have a posi- tive-regulated relationship and if this is true, they will be given a high score. On the other hand, edges that are farther apart are gi- ven a lower score. Although the edge method strongly focuses on local similarity between two gene expression curves, there are al- ways too few edges that match between gene pairs and this gives rise to relatively low similarity scores. Yet another approach is the Event Method (Kwon et al., 2003). This first transforms the direc- tional changes in expression level into a directional event (Rising, Constant or Falling) by calculating the slope values at each time point; these are then converted into events. When the transforma- tion process is complete, global sequence alignment by the Needle- man–Wunsch algorithm is used to find the best match between the two event strings taking noise and time lag into account. Gene pairs that have an activation relationship are likely to have a high similarity score for the event strings. Although the event method is efficient at finding gene TF–target relationships, it is unable to pro- duce a high score for related genes when the event strings are sim- ilar for periodical short matches because the method uses global alignment. Moreover, it is computationally inefficient because it needs n2 pairwise global alignments if there are n genes. There is one further limitation: this is that it provides putative TF–target relationships for users by using gene pairs instead of gene clusters. It is clear that clusters are more efficient than pairs when finding TF–target relationships. The next available approach is CLARITY (Balasubramaniyan et al., 2005). This can find locally similar re- gions in gene expression profiles by measuring similarity based on Spearman rank correlation. In order to discovering local time- shifted relationships between two profiles, the program enumer- ates all possible alignments in a systematic way. In order to reduce the complexity of computation for Oðn2Þ, it use an approximate algorithm. Although CLARITY tests similarity between genes by measuring shape (the qualitative behavior) of the expression pro- files, which is very useful, there are a few limitations to this pro- gram. One exception is where cell cycle co-expressed genes are highly related to the cell cycle but their qualitative behavior differ- ent over multiple cycles. The result is a horizontal shift and scale problem with the expression profiles and CLARITY finds it difficult to identify such genes as having a strong association. A further ap- proach is q-cluster (Ji & Tan, 2005). Here, the profile is transformed into three type of change denoted by 1, �1 and 0. In this method, the pattern of a q-cluster is indicated as an event string of length ðq � 1Þ to show the changes that occur and this reflects how expression level changes from condition i to condition i þ 1 under the q conditions. As the data is transformed into three distinct clas- ses, there are 3q�1 q-clusters in total and each q-cluster has a un- ique q-value, where 0 6 q-value 6 3q�1 (Ji & Tan, 2005). Although q-cluster provides users with detailed information that allows the detection of periodically co-regulated genes, users must search for these genes themselves. That is to say, q-cluster does not fur- ther group genes together according the periodic similar patterns relative to the cell cycle expression data. Spellman et al. (1998) first performed a genome-wide transcrip- tional analysis of the mitotic cell cycle of Saccharomyces cerevisiae using microarrays and found about 800 periodically expressed genes that peak each cycle. We can regard these genes as cell-cycle regulated genes. In another similar investigation (Whitfield et al., 2002) it proved possible to identify genes periodically expressed over the human cell cycle using microarray analysis. Although a large number of studies (Cho et al., 1998; Spellman et al., 1998) have revealed over 800 genes that are cell-cycle regulated in yeast, there has been a lacking of a good method to further group these periodically expressed genes into clusters by biclustering in order to get an understanding of transcription regulatory networks with- in cell cycle. When biclustering (Cheng & Church, 2000) was first applied to gene expression data, it used the measure, mean squared residue, to find the biclusters. In later years, many approaches to 2204 L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 biclustering of expression data have been proposed. However, if these approaches focus on finding exclusive biclusters in a time series of expression data, there are a number of problems that are of concern. A major problem is that they ignore many addi- tional relationships across the time series such as time-lagged relationships between transcription factors and thus genes that are activated by this transcription factor. Another problem is that the process groups genes that show similar activity patterns un- der a subset of discrete conditions instead of under contiguous conditions. In the analysis of time series dataset, it is reasonable to restrict the analysis to biclusters with contiguous columns to reduce the time complexity of computation. 3. Methods In order to take into account the bicluster’s coherent values when biclustering, we firstly transform the expression data into event strings. We then use the set of event strings to construct a generalized tree. When the generalized tree has been built, we can use it to form biclusters quickly while taking time lags into consideration. Furthermore, during the analysis of cell cycle expression data, we need to focus on finding cell-cycle regulated genes. Therefore, we further transform the event strings into trans- actions representing what event substrings (items) are included in genes’ event strings (transactions). Next, we use Apriori’s concept to obtain the set of similar patterns (items) that occur simulta- neously with the same genes (transactions). Finally, we use the positions at which the similar patterns occur for the same genes to further group the genes taking periodically similar patterns and time lags into consideration. if a00i;j�1 ¼ U;a 00 i;j ¼ N;Ai;j > s; a 00 i;jþ1 ¼ N;Ai;jþ1 > s; . . . ; a 00 i;jþk ¼ N;Ai;jþk > s; Ai;jþkþ1 > s; a 00 i;jþkþ1 ¼ D; then a00i;j ¼ H;a 00 i;jþ1 ¼ H; . . . ; a 00 i;jþk ¼ H; if a00i;j�1 ¼ D;a 00 i;j ¼ N;Ai;j < �s; a 00 i;jþ1 ¼ N;Ai;jþ1 < �s; . . . ; a 00 i;jþk ¼ N;Ai;jþk < �s; Ai;jþkþ1 < �s; a 00 i;jþkþ1 ¼ U; then a00i;j ¼ L;a 00 i;jþ1 ¼ L; . . . ; a 00 i;jþk ¼ L: 8>>>>< >>>>: 3.1. Phase 1: Transforming expression data into event strings When we have n genes and m conditions in the form of time- series expression data, the time-series expression data can be represented as a A ¼ n � m matrix, where Ai;j represents the Table 1 The original expression matrix. C1 C2 C3 C4 C5 C6 G1 �0.42 �0.01 0.19 �0.09 �0.10 �0.68 G2 �0.43 0.15 0.3 0.05 0.07 �0.02 G3 �0.11 0.27 0.38 0.46 �0.06 �0.29 G4 �0.27 �0.13 0.2 0.17 0.09 0.14 C5 �0.25 �0.28 �0.35 0.82 0.58 0.02 Table 2 The matrix showing the changes from the original matrix. C1,2 C2,3 C3,4 C4,5 C5,6 C G1 0.98 20 �1.47 �0.11 �5.8 G2 1.35 1 �0.83 0.4 �1.29 G3 3.45 0.41 0.21 �1.13 �3.83 G4 0.52 2.54 �0.15 �0.47 0.56 G5 �0.12 �0.25 3.34 �0.29 �0.96 � expression level of gene i under condition j. First, Matrix A is con- verted into a A0 ¼ n �ðm � 1Þ matrix such that A0i;j ¼ Ai;jþ1�Ai;j jAi;jj if Ai;j–0; Infinity if Ai;j ¼ 0 and Ai;jþ1 > 0; �Infinity if Ai;j ¼ 0 and Ai;jþ1 < 0; 0 if Ai;j ¼ 0 and Ai;jþ1 ¼ 0: 8>>>>< >>>>: to reflect the change of each gene expression level between two neighboring time points. The transformation function we used comes from q-cluster. Each entry A0i;j reveals the directional change (the rate of change across time) from the expression level of gene i at time point j to the expression level of gene i at time point j þ 1. After A has transformed into A0 matrix, we are interested in how the change in the gene expression level can be converted into a set of symbols, R. R contains five symbols, fD; U; N; H; Lg meaning down-regulated, up-regulated, unchanged, unchanged with high expression, and unchanged low expression, respectively. We use a threshold t (same with q-cluster) to transform the changes into event symbols such that A00i;j ¼ U if A0i;j P t; D if A0i;j 6�t; N otherwise: 8>< >: Further, we also use a threshold s to transform the symbol, N into H or L to describe what happens with the expression profiles in more depth. For every sub-matrix a00 ¼ 1 � k of A00 whose start- ing row is i and starting column is j, where i 2 1 . . . :n; j 2 1 . . . m � 1; j þ k 6 m � 1, and After the conversion phase, matrix A0 is transformed into matrix A00 and A00i;j represents the converted symbol of the change from the expression level of gene iat time point j to the expression level of gene i at time point j þ 1. As an example, Tables 1–3 show the process of converting the gene expression levels. The original matrix A is shown in Table 1. In Table 2, the matrix A0 reflects the changes in C7 C8 C9 C10 C11 C12 �0.16 0.33 0.29 0.11 0.35 0.21 0.33 0.26 �0.41 0 �0.4 0.03 �0.49 �0.11 0.03 �0.16 0.43 �0.03 �0.34 0.1 0.23 �0.34 0.13 0.19 �0.51 �0.22 �0.4 �0.22 1.07 0.11 6,7 C7,8 C8,9 C9,10 C10,11 C11,12 0.76 3.06 �0.12 �0.62 2.18 �0.4 17.5 �0.21 �2.58 1 �1 1.08 �0.69 0.78 1.27 �6.33 3.69 �1.07 �3.43 1.29 1.3 �2.48 1.38 0.46 26.5 0.57 �0.82 0.45 5.86 �0.9 Table 3 The result of conversion of the original matrix. Some event symbols N are have transformed into H to describe what is happening within the gene profiles in more depth. C1,2 C2,3 C3,4 C4,5 C5,6 C6,7 C7,8 C8,9 C9,10 C10,11 C11,12 G1 U U D N D U U N ? H D U D G2 U U D U D U N ? H D U D U G3 U U N ? H D D D U U D U D G4 U U N ? H D U D U U D U U G5 N N U N ? H D D U D U U D Table 5 The candidate 2-itemsets after pruning. G2[5, 3] and G5[2, 4] can not be counted as one because their positions overlap. The items (similar patterns) in the itemset are not limited to be different from each other, see itemset, UDU, UDU. The itemsets {UUD, UUD}, {UHD, UHD} and {UHD, DDU} have a support of less than 2 and therefore L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 2205 gene expression levels on conversion from matrixA. Finally, the re- sult of conversion process ðt ¼ 0:3; s ¼ 0:1Þ forms matrix A00, which is shown in Table 3. In the latter case, some symbols N have now been transformed into H, because, despite the expression not varying greatly, the expression is still high and therefore important in itself. 3.2. Phase 2: Using converted matrix to construct the generalized tree Let the set of event strings S ¼fS1; S2; . . . ; Sng be obtained from each row in converted matrix A00, which comes from the first phase. Each of these strings has m � 1 event symbols and corresponds to the symbols in a row of the converted matrix. To build a suffix tree, add a special end of string character,, to each event string. Then, use the set of event strings S to construct the generalized tree T. After the generalized suffix tree T has been built, we can use it to find the common subsequences shared among more than two sequences. Based on such common subsequences, the similar sequences (event strings) are grouped together. Use of a generali- zed suffix tree allows a pattern of length n to be found in kstrings in Oðn þ kÞ time. This is known as Exact String Matching. We can use Exact String Matching to find co-regulated genes in a similar way to q-cluster. As shown in Table 4, genes which have an arbi- trarily similar pattern of length 3 can be grouped together. Like q-cluster, users could then choose an interesting pattern that is re- peated to pinpoint TF–target relationships. For example, G3 and G5 have a similar pattern, UUD, and G5 activates this pattern after G3 by two time lags. A TF–target relationship is likely between G3 and G5. So far, our method is able to achieve the same objectives as q- cluster, but our approach gives more. If users want to see other lengths among similar patters, users just query through the gener- alized tree T that has been built by transforming the event strings. This is not true for q-cluster and the q-clusters need to be calcu- lated again. Furthermore, although q-cluster provides users with detailed information that helps the detection of periodically co- regulated genes, users still have to search manually. That is to say, q-cluster does not further group genes together according to the periodic similar patterns for cell cycle expression data. In our method, we take into account periodically similar patterns to group periodically co-regulated genes as can be seen in the next phase. 3.3. Phase 3: Identifying similar patterns (items) that occur simultaneously among genes (transactions) Let T ¼fT 1; T 2 . . . T ng be the set of transaction, where n is the number of genes in expression data. A transaction can be thought Table 4 Grouping genes that have similar activation patterns. Pattern Gene [start position] UUD G1[0], G2[0], G3[6], G4[6], G5[8] UHD G1[6], G2[5], G3[1], G4[1], G5[2] UDU G2[1], G2[3], G2[8], G3[7], G4[4], G4[7], G5[6] � � � � � � DDD G3[3] DDU G3[4], G5[4] DND G1[2] of as a basket (gene) that contains a set of items (event substrings of a fixed length), which are bought together. A generalized suffix tree allows us to quickly find common subsequences that are shared among gene event strings; therefore we can use it to built transactions representing the event substrings that are included in gene event strings. When we have a set of transactions, T, we can use these to find the frequent h-itemsets (h similar patterns), and count the number of times they appear simultaneously in transactions (genes). The number of times that a transaction is found is defined by us as the support S, which is different to the Apriori approach (Agrawal and Srikant, 1994). There are several steps involved in finding frequent h-itemsets that are similar to the Apriori algorithm, but there are also some differences. First, find all frequent 1-itemsets. Then, do the next steps for each iteration until the frequent h-itemsets are found. The candidate k-itemsets are next generated by merging frequent ðk � 1Þ-itemsets. As the next step, prune the candidate k-itemsets whose subsets are infre- quent. Take the candidate k-itemsets that survive the pruning pro- cess and count the number of times they appear simultaneously among transactions. If the k items (similar patterns) among every k-itemsets do not overlap (similar pattern positions should not overlap each other), then we count this as one. Furthermore, we can limit the distance of items’ starting positions to be larger than dand then count this as one. Finally, extract the frequent k-itemsets, a subset of candidate k-itemsets, with counts that are no less than the support S. For example, if we use the transactions shown in Ta- ble 3, they can be transformed into frequent 2-itemsets and the re- sult is shown in Table 5. It should be noted that in our method, the items in every itemsets are not limited to being different from each other. The reason that this is done is so we can take into account in- stances when the periodic similar patterns among genes are the same. If G2 and G4 are examined, it will be found that they have the periodic similar pattern, UDU, which occur at position [3, 8] in G2 and at position [4, 7] in G4. Although we have collected the transactions (genes) that have period similar patterns, we still need to further group these genes according to the periodically similar patterns’ positions to find related transactions; this is carried out in the next phase. If the number of grouped genes is less than threshold S, it is clear that the number of grouped genes can never greater than the threshold S. Thus, we can regard the threshold S as the final frequent 2-itemsets are {UHD, DDD}, {UHD, UDU}, {UDU, UDU}, etc. Patterns Gene [start positions] Support UUD, UUD 0 UUD, UHD G1[0, 6], G2[0, 5], G3[6, 1], G4[6, 1], G5[8, 2] 5 � � � � � �, � � � UHD, UHD 0 UHD, UDU G2[5, 1], G2[5, 3], G2[5, 8], G3[1, 7], G4[1, 4], G5[2, 6] 5 � � � � � � � � � UHD, DDU G3[1, 4] G5[2, 4] 1 UDU, UDU G2[3, 8], G4[4, 7] 2 � � � � � � � � � 2206 L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 a first cutoff that is used to limit the size of the grouped genes and avoid unnecessary computations. The detailed pseudo code is below: Input: The number of similar patterns,h. Output: frequent h-itemsets. Scan the transactions to find frequent h-itemsets 1: Scan the transactions t 2 T to find frequent 1-itemsets; 2: for ðk ¼ 2; k <¼ h; k þþÞ{ 3: Generate candidate k-itemsets; // By merging a frequent ðk � 1Þ-itemsets. 4: Prune candidate k-itemsets whose subsets are infrequent; 5: Scan the transactions t 2 T to count the occurrences of itemsets in candidate k-itemsets; // kitems’ starting position do not overlap. 6: Extract the frequent k-itemsets, a subset of candidate k- itemsets with counts no less than support S; 7: } 8: Return h-itemsets; 3.4. Phase 4: Further grouping of genes according to their similar pattern positions When we have obtained frequent h-itemsets from the above phases, we need to further group the genes according the positions at which these similar patterns (items) occur simultaneously for the genes (transactions). Using Table 5 as an example, it is possible to see how further clustering is carried out. We already know that the itemset, UUD, UHD, in Table 5 occurs at positions, [0, 6] in G1, [0, 5] in G2, [6, 1] in G3, [6, 1] in G4 and [8, 2] in G5. Suppose that a node is a set of positions for itemset, UUD, UHD, among the same gene (transaction). Hence, this set of nodes is {G1[0,6], G2[0, 5], G3[6, 1],G4[6, 1],G5[8, 2]}. The edge is the Euclidian Distance be- tween two nodes. For example, the distance between node, G1[0, 6], and node, G2[0, 5] is ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð0 � 0Þ2 þð6 � 5Þ2 q ¼ 1. We create a graph that contains these nodes and computed the edges as Fig. 1. The process of grouping. (a) The original graph has five node and 10 edges. The dis has the minimal distance 0 within whole graph. We merge node G3 and G4, and then u merged node and the other nodes. (c) We merge nodes, G1 and G2, and update this me compute the distances between the merged node and the other nodes. (d) The final res shown in Fig. 1a. When the graph is complete, the edge containing the minimum distance is first selected. If the selected edge’s dis- tance is less than the cutoff threshold d we set it to3 and the two nodes that are connected by this edge are merged (see Fig. 1b). The merged node’s position is re-defined as the average of the ori- ginal two nodes’ positions (see Fig. 1b) and the distances along the edges that the merged node is connected with are re-computed for next merging. Next, we similarly select another edge that has the minimal distance for the whole graph and then merge two nodes if the minimal distance between them is less than the cutoff threshold d. Of course, we then should update the merged nodes and the edges that connect the new node with the other nodes. This merging process is continued until we can no longer find a minimal distance that is less than the cutoff threshold d. The final result is shown in Fig. 1d. We find that we have indeed grouped these genes into two groups, [G1, G2] and [G3, G4, G5] according to the similar pattern positions. The pseudo code is presented below: Input: The frequent h-itemset and these transactions (genes) that items (similar patterns) among this frequent h-itemset occur simultaneously at. Output: The result of grouping according the positions. 1: Build the initial graph; 2: Compute Euclid Distance between two nodes; 3: repeat 4: minDistance = Select the minimal distance; 5: if(minDistance < cutoff threshold dÞ { 6: Merge two nodes and then update merged node and dis- tances between merged node and other nodes; 7: }else{ 8: Break; 9: } 10: end 11: Return grouping result; tances are already computed by Euclidian Distance. (b) The edge between G3 and G4 pdate the merged node’s positions. We also re-compute the distances between the rged node’s positions by averaging the merging nodes’ positions and then also re- ult of grouping according to the original positions is two clusters. L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 2207 3.5. Phase 5: Functional enrichment Based on the hypothesis that co-expressed genes are likely to have a similar function, we can use this to help analyze the large number of clusters generated by our procedure. In order to show that the clusters we generated have biological significance, we use p-values that are obtained from the hypergeometric distribu- tion. The probability (p-value) of observing at least m genes from a specific class within a study cluster of size n is given by p-value ¼ 1 � Xm�1 i¼0 f i � � � N � f n � i � � N n � � where f is the total number of genes from the population within a specific class and N is the total number of genes within the total population (defined as all genes represented on the microarray). Thus, the p-value corresponds to the probability of obtaining at least m gene of the cluster in a random set of size n. A low p-value indicates that the genes annotated belong to an enriched class that is statistically biologically significant. Our system has been coupled to Ontologizer (Robinson et al., 2004) to calculate p-values to detect statistically significant enrichment within one or more GO catego- ries. Furthermore, we also calculate p-value for the statistically sig- nificant enrichment for some TFs’ known target-gene groups. Moreover, we have modified some open source code from Ontolo- gizer so that these subroutines are completely integrated into our system. 4. Experimental results 4.1. Clustering with the aim of finding time-lagged co-regulated genes In this experimental system, we applied Spellman’s yeast cell cycle dataset that includes 6331 open reading frames. The full dataset contains all the expression data for the alpha factor, cdc 15 and elutriation time course experiments. We used only the al- pha factor dataset to validate our approach. We examined the re- sults obtained with this dataset by our method for the detection of time-lag co-regulated genes over the cell cycle. We applied the approach to the cell-cycle regulated genes that were identified by Spellman’s group (Spellman et al., 1998) in our experiments by using the identified cell-cycle regulated genes as the population of genes tested using our method. These identified cell-cycle regu- lated genes involves 800 genes and include 14 well-known cell-cy- cle TFs, SWI4, YOX1, SWI5, FKH1, RAP1, YHP1, HCM1, ACE2, STP2, GAT3, PHD1, GCR1, TEC1 and MET28. Furthermore, we add some other known cell-cycle TFs to the original dataset including ARR1, MBP1, RPN4, YAP1, FKH2, MCM1, STE12, YAP5, STP1 making Fig. 2. Fkh2 (black bold line) really regulates the genes displayed here from Cluster 1 a total of 809 genes. The Spellman’s yeast cell cycle dataset covers two cell cycles and for this reason we formed clusters based on whether the gene profiles had two similar patterns of length 3 units over two cell cycles. The threshold s was set to 0.1 and the threshold t was set to 0.3. The distance d between two similar pat- terns was set to 2. The threshold D was 5 to take into account time lag. The support S for cluster size was set at 2 to avoid missing any significant clusters. Our method is designed to fit the special needs of users; hence, users are able to pick out clusters of interest according to what they think are significant similar patterns within these clusters. For example, someone may believe that the patterns, UUH, UUD, are significant whereas the patterns, DDD, UUU, are insignificant. In order to filter some clusters that are insignificant, we can pick up clusters whose similar patterns both start with the character, U, and contain at least one TF. To realize the biological relevance of these chosen clusters, the p-value is calculated from the hyper- geometric distribution in order to model the probability of observ- ing at least mgenes, from a cluster with n genes that may by chance contain a TF that regulates f known (documented) genes from the 809 genes in the dataset. We use YEASTRACT (Teixeira et al., 2006) to search for documented TF–target relationships for every cluster and then calculate the p-value for each TF that is grouped with its known regulated genes. We present only the top few clusters (p-value < 0.01), which are summarized in Table 9 in the Appendix. Although most of the clusters we have picked up do not show a statistically significant p-value, the result is really quite good. If Cluster 1 is examined, it is found to contain the TF, Fkh2, and this TF would seem to regulate the genes found in Cluster 1. In order to validate whether Fkh2 really regulates these Cluster 1 genes, we plot Fkh2 and the genes that the result is coordinated regulation (see Fig. 2). It can be seen that Fkh2 over-expresses simultaneously with or before the genes in this cluster and this supports the idea Fkh2 regulates these genes. Similarly, in chosen Cluster 2, Swi4 would seem to regulate the genes that form this cluster. In order to confirm regulation they were also plotted in a similar way to above (see Fig. 3). The Edge Detection and Event Method is one method of finding a regulated relationship. However, there are some drawbacks. Firstly, the result is regulator–target pairs instead of clusters, but a TF does not regulate only one gene but a group of genes with similar function; thus cluster results are more useful than pairs. Secondly, the regulator–target pairs that are identified by the Event Method only say there is a have high correlation and do not provide more detailed information such as the exact lag time. We can also pick up clusters whose similar patterns both start with the character U but do not contain any known TFs. Then, we used YEASTRACT to obtain TFs that regulate some of the genes from the chosen clusters. We calculated the p-value obtained from the hypergeometric distribution to model the probability of because it over-expresses simultaneously or before than known regulated genes. Fig. 3. Swi4 (black bold line) really regulates the genes displayed here from Cluster 2 because it over-expresses simultaneously or before than known regulated genes. 2208 L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 observing at least m genes from a chosen cluster with n genes by chance if a TF (not be included in this chosen cluster) regulates f known genes from the 808 genes in the dataset. We again present only the top clusters (p-value < 0.01), which are summarized in Ta- ble 9. The statistically most significant clusters (p-value < 0.01) that contain a TF and also contain this TF’s regulated genes. It is clear that if the percentage of a cluster’s genes that are regulated by a TF is much more larger than the percentage of the population genes that are regulated by the same TF, this cluster is more statis- tically significant. Table 10 in Appendix the results show that although the chosen clusters do not contain any TFs, most of the genes among them seem to be regulated by a TF. Moreover, based on the fact that genes with a similar function are always regulated by same TF, the results reveal that the clusters generated by our method have some biological significance. We also plotted the TF for the known regulated genes and the gene expression profiles of the cluster to Fig. 5. Although Ace2 (black bold line) does not group with its known regulated genes expresses simultaneously or before than known regulated genes. Fig. 4. Although Swi5 (black bold line) does not group with its known regulated genes expresses simultaneously or before than known regulated genes. validate whether the group created by our method occurred by chance or had real biological significance. If we plot the TF Swi5 and genes’ expression profiles from Cluster 21 (see Fig. 4), we find that Swi5 is over-expressed simultaneously or before the regulated genes and this supports the idea that Swi5 regulates these genes. Similarly, the TF Ace2 seems to regulate the genes from Cluster 16 (see Fig. 5). 5. Comparison of other approaches In order to test whether our method could group co-regulated genes that other method find hard to group, we compared our method with some correlation measures such as Spearman Rank Correlation and Pearson Correlation. Initially, we picked two genes, MCM7 and MCM4, which are components of the MCM complex (Davey et al., 2003). Fig. 6 shows the expression levels of these , we could show that it really regulated the genes displayed here because it over- , we could show that it really regulated the genes displayed here because it over- Fig. 6. The profiles of two genes, MCM7 and MCM4, which are components of the MCM complex. Table 6 The pairwise similarity between MCM7 and MCM4 as calculated by Pearson Correlation and Spearman Rank Correlation. Pearson Correlation Spearman Rank Correlation 0.6 0.66 Table 7 The pairwise similarity matrix for CDC68, SWI4 and CLN1 as calculated by Pearson Correlation. CDC68 SWI4 CLN1 CDC68 0 0.68 0.69 SWI4 0 0 0.83 CLN1 0 0 0 Table 8 The pairwise similarity matrix for CDC68, SWI4 and CLN1 as calculated by Spearman Rank Correlation. CDC68 SWI4 CLN1 CDC68 0 0.70 0.70 SWI4 0 0 0.91 CLN1 0 0 0 L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 2209 two genes and similar expression patterns that were detected by our method. Then, we calculated the pairwise similarity between the expression profiles of the individual genes according to Spear- man Rank Correlation and Pearson Correlation (see Table 6). We find that the pairwise similar scores for MCM7 and MCM4 are actu- ally quite low (60.7). The reason for this is presence of scaled and shifted similarity patterns between these co-regulated genes and this causes these co-regulated genes to be difficult to group to- gether by numerical measures. In contrast, our method transforms the expression profiles into event strings to take into account scal- ing and shift factors in the expression level between related genes. We picked out three genes, CDC68, SWI4 and CLN1 that have a regulated relationship between each other. CDC68 is a required activator of SWI4 and SWI4 is a required activator of the G1 cyclin genes CLN1 and CLN2. CDC68’s role at the CLN promoters may be indirect (Lycan et al., 1994). Fig. 7 presents the expression levels of these three genes and we also calculated the pairwise similarity matrix between the expres- sion profiles of these three genes according to Spearman Rank Cor- relation and Pearson Correlation (see Tables 7 and 8, respectively). In a similar way to the previous result, we found that the pairwise similar scores are not high with half of scores being relatively low (60.7). Again, Spearman Rank Correlation and Pearson Correlation find it hard to group these genes using similarity matrixes. In real- ity, there are some drawbacks to these two numerical measures. Firstly, they do not provide detailed information on co-regulated gene pairs including time lag between them and the pattern simi- larity between two genes. The information on similar patterns that we are interested in includes the starting time point and what hap- pens over time to the patterns such as the change between two neighborly time points. Fig. 7. The profiles of three genes, CDC68, SWI4 and CLN1, w 6. Conclusion By converting the gene expression values, we present a local method to find potential TF–target relationships allowing the detection of time-lagged gene clusters based on periodically sim- ilar patterns over multiple cycles. Genes that have the same peri- odic patterns are grouped together and these genes are likely to be cell-cycle regulated genes that are controlled simultaneously by a TF. Generally, these TFs are included in same group as the regulated genes. The similarity of the two expression subprofiles is according to the specific changes in expression level by the genes. Our approach provides users with more detailed informa- tion about the detected similar patterns and users can use this information to choose the more significance clusters themselves and thus decrease the number of candidate clusters. We have experimented with our method on a time-series expression dataset (Spellman et al., 1998). Genes with similar hich have regulated relationships between each other. 2210 L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 patterns are found to be co-expressed genes and further, they are likely to have a TF–target relationship. In order to validate our method, we used p-value to show that the clusters we generated were able to help find TF–target relationships. Our method was compared with some other measures such as Pearson correlation and Spearman rank correlation and we found that our method is able to group highly related genes together that other measures found hard to do. Furthermore, unlike the other methods, our approach can provide more detailed information on Table 9 The statistically most significant clusters (p-value < 0.01) that contain a TF and also contain regulated by a TF is much more larger than the percentage of the population genes that a Cluster number Similar patterns Number of genes (n) Cluster with this TF Number of genes be regulated by this TF (m) % o gen by 1 UUH, UUH 25 Swi4 18 72 2 UUN, UUU 25 Fkh2 12 60 3 UUU, UUU 77 Fkh2 16 33. 4 UHD, UUH 26 Swi4 14 53. 5 UHD, UUH 31 Swi4 15 50 6 UUH, UUN 10 Fkh2 6 60 7 UHD, UHD 33 Swi4 15 46. 8 UDD, UUU 18 Yhp1 5 31. the TF–target relationships. Finally, the results from the Edge Detection and Event Method are pairs of genes and the clusters ob- tained by our method are clearly more useful than pairs for finding a TF–target relationship. Appendix A See Tables 9 and 10. this TF’s regulated genes. It is clear that if the percentage of a cluster’s genes that are re regulated by the same TF, this cluster is more statistically significant. f this cluster’s es be regulated this TF % of population genes be regulated by this TF (f) p-Value Genes are included in this cluster are this TF’s known regulated genes 15.9 2.74E�10 YPL127C, YJL187C YMR199W, YPL267W YPR202W, YER189W YHL049C, YEL075C YDR224C, GR109C YBL003C, YHR218W YPL163C, YER111C YCR065W, YMR307W YDR225W, YFL064C 10.1 3.51E�8 YNL068C, YJR092W YPL141C, YNL058C YJL051W,YPR119W YML119W, YML034W YMR032W, YLR190W YGR108W,YHR152W 3 10.1 4.81E�6 YMR001C, YPL141C YJL051W, YHR023W YPR119W,YIL106W YML034W, YNL068C YJR092W, YKL096W YPL242C, YNL058C YPL155C, YLR131C YPR156C, YHR152W 8 15.9 6.25E�6 YJL187C, YGR014W YPL267W, YPR202W YER189W ,YHL049C YEL075C, YGR109C YHR218W, YPL163C YER111C, YCL025C YCR065W, YFL064C 15.9 9.42E�6 YJL187C, YMR199W YPL267W, YPR202W YER189W, YHL049C YEL075C, YOR074C YGR109C, YHR218W YPL163C, YPR120C YER111C, YCR065W YFL064C 10.1 1.37E�4 YNL068C, YJR092W YPR119W, YML034W YLR190W, YGR108W 8 15.9 2.57E�4 YJL187C, YPL267W YPR202W, YER189W YHL049C, YEL075C YOR074C, YGR109C YHR218W, YPL163C YPR120C, YER111C YCL025C, YCR065W YFL064C 2 5.3 9.48E�4 YJL194W, YPR019W YLR274W, YIL106W YBR202W Table 9 (continued) Cluster number Similar patterns Number of genes (n) Cluster with this TF Number of genes be regulated by this TF (m) % of this cluster’s genes be regulated by this TF % of population genes be regulated by this TF (f) p-Value Genes are included in this cluster are this TF’s known regulated genes 9 UDD, UDD 129 Met28 7 6.1 1.7 1.28E�3 YIR017C, YPR167C YER091C, YJL078C YGR055W, YFR030W YGL184C 10 UDD, UUD 177 Tec1 7 4.6 1.7 7.64E�3 YNL283C, YDR055W YNL166C, YER070W YCR018C, YML100W YBR083W 11 UHD, UUU 34 Fkh1 6 22.2 6.9 7.97E�3 YNL068C, YPR119W YPL155C, YCL063W YAR071W, YOR315W Table 10 The statistically most significant clusters (p-value < 0.01) do not contain any TFs but in which most of the genes are regulated by a TF. It is clear that if the percentage of a cluster’s genes that are regulated by a TF is much more larger than the percentage of the population genes that are regulated by the same TF, this cluster is more statistically significant. Cluster number Similar patterns Number of genes (n) Cluster without this TF Number of genes be regulated by this TF (m) % of this cluster’s genes be regulated by this TF % of population genes be regulated by this TF (f) p-Value Genes are included in this cluster are this TF’s known regulated genes 1 UHD, UUU 62 Mbp1 18 38.2 12.1 1.71E�6 YDR309C, YOR230W YGR014W, YNL339C YIL177C, YHR219W YHL049C, YDR113C YDL018C, YKL096W YMR179W, YJL074C YHR218W, YHR153C YLR462W, YML027W YGR296W, YJL225C 2 UHH, UHH 7 Swi4 7 100 15.9 2.28E�6 YPL127C, YDR224C YLR183C, YBL003C YKR013W, YMR307W YDR225W 3 UHH, UUH 13 Swi4 9 69.2 15.9 2.1E�5 YPL127C, YMR199W YDR224C, YLR183C YBL003C, YKR013W YGR189C, YMR307W YDR225W 4 UUH, UUU 53 Swi4 17 41.4 15.9 5.18E�5 YMR199W, YNL339C YKL008C, YPR202W YHR219W, YHL049C YEL075C, YNL300W YDL018C, YBL003C YMR179W, YHR218W YGL038C, YLR462W YML027W, YGR296W YMR307W 5 UDD, UHD 87 Mbp1 22 27.5 12.1 5.85E�5 YDR309C, YJL073W YJL187C, YBL111C YAR008W, YJR030C YNL339C, YIL177C YNL030W, YHR219W YDL018C, YKL096W YMR179W, YDR545W YDL003W, YHR153C YLL067C, YLR103C YLR462W, YML027W YGR296W, YJL225C 6 UUN, UUU 9 Fkh1 5 62.5 6.9 6.37E�5 YMR215W, YMR001C YJL051W, YCL063W YPR156C 7 UDD, UUH 51 Mbp1 16 32.6 12.1 7.47E�5 YJL187C, YAR008W YNL339C, YNL030W YHR219W, YOR075W YDL018C, YMR179W YDR545W, YDL003W YHR153C, YLL067C YLR103C, YLR462W (continued on next page) L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 2211 Table 10 (continued) Cluster number Similar patterns Number of genes (n) Cluster without this TF Number of genes be regulated by this TF (m) % of this cluster’s genes be regulated by this TF % of population genes be regulated by this TF (f) p-Value Genes are included in this cluster are this TF’s known regulated genes YML027W, YGR296W 8 UDU, UHD 36 Gat3 7 19.4 3.3 8.86E�5 YCR041W, YEL076C YHL049C, YEL075C YLR463C, YFL065C YFL064C 9 UDU, UUH 31 Gat3 6 20 3.3 2.18E�4 YLR467W, YEL076C YHL049C, YEL075C YLR463C, YFL064C 10 UUH, UUU 30 Swi4 13 43.3 15.9 2.55E�4 YPL127C, YJL187C YDR507C, YGR086C YIL141W, YKL113C YPL256C, YBL003C YDL055C, YMR307W YER001W, YDR225W YNL031C 11 UDD, UUN 11 Fkh2 6 54.5 10.1 2.77E�4 YPL141C, YNL058C YJL051W, YML119W YMR032W, YHR152W 12 UDL, UUU 9 Mcm1 4 66.6 8.8 7.57E�4 YJR092W, YPR119W YGL021W, YLR131C 13 UUH, UUU 28 Fkh1 7 31.8 6.9 3.95E�4 YJR092W, YPR119W YGL021W, YPL155C YLR131C, YML034W YAR071W 14 UDD, UUU 60 Swi5 11 21.1 7.1 6.01E�4 YDR055W, YKL164C YPL283C, YGR086C YBR083W, YLR464W YLR467W, YGR041W YKL163W, YNL328C YLR463C 15 UDD, UUD 20 STP1 5 26.3 4 6.4E�4 YER070W, YJR154W YDR501W, YML100W YLR121C 16 UDN, UUH 22 Ace2 5 22.7 4 1.33E�3 YER124C, YNL327W YKL185W, YHR143W YLR079W 17 UUD, UUN 15 Yap5 4 28.5 5.9 7.04E�3 YIL129C, YPL141C YNL058C, YIL158W 18 UDU, UUU 76 Gat3 7 12.5 3.3 1.56E�3 YLR467W, YHL049C YEL075C, YLR462W YLR463C, YBL112C YPR203W 19 UDN, UHD 24 Ace2 5 20.8 4 2.02E�3 YER124C, YNL327W YKL185W, YHR143W YLR079W 20 UDN, UUD 13 Yhp1 4 30.7 5.3 3.49E�3 YJL115W, YNL339C YLR413W, YLL067C 21 UDD, UUH 48 Swi5 9 19.1 7.1 4.22E�3 YLR464W, YLR467W YKL164C, YLR049C YPL283C, YLR463C YJL078C, YGR086C YPL158C 22 UDD, UDN 41 Gcr1 4 10.8 1.9 4.55E�3 YGR240C, YMR055C YGR143W, YIL162W 23 UDD, UUU 39 Mcm1 8 22.8 8.8 8.77E�3 YMR001C, YLR040C YJL157C, YJL194W YNL058C, YLR274W YGR143W, YHR152W 24 UDD, UDN 17 Rpn4 6 35.2 10.7 6.08E�3 YOL016C, YDR055W YBL023C, YKR012C YLR013W, YOR273C 25 UDD, UdN 39 Yap5 6 18.7 5.9 8.83E�3 YLL066C, YPL208W YLL067C, YML050W YBR007C, YPL283C 26 UHH, UUU 25 Mcm1 6 27.2 8.8 9.53E�3 YJR092W, YGL021W YLR131C, YIL158W YBR202W, YDR451C 2212 L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 L.-C. Wu et al. / Expert Systems with Applications 37 (2010) 2202–2213 2213 References Agrawal, R., & Srikant, R. (1994). Fast algorithms for mining association rules, Proc. of the 20th VLDB Conf., Santiago, Chile. Balasubramaniyan, R., Hullermeier, E., et al. (2005). Clustering of gene expression data using a local shape-based similarity measure. Bioinformatics, 21(7), 1069–1077. Cheng, Y., & Church, G. M. (2000). Biclustering of expression data. In Proceedings of the International Conference on Intelligent Systems of Molecular Biology (Vol. 8, pp. 93–103). Cho, R. J., Campbell, M. J., et al. (1998). A genome-wide transcriptional analysis of the mitotic cell cycle. Molecular Cell, 2(1), 65–73. Davey, M., Indiani, C., et al. (2003). Reconstitution of the Mcm2-7p heterohexamer, subunit arrangement, and ATP site architecture. Journal of Biological Chemistry, 278(7), 4491–4499. Filkov, V., Skiena, S., et al. (2001). Identifying gene regulatory networks from experimental data. In Proceedings of RECOMB. Gusfield, D. (1997). Algorithms on strings, trees and sequences: Computer science and computational biology. London: Cambridge University Press. Ji, L., & Tan, K. L. (2005). Identifying time-lagged gene clusters using gene expression data. Bioinformatics, 21(4), 509–516. Kato, M., Tsunoda, T., et al. (2001). Lag analysis of genetic networks in the cell cycle of budding yeast. Genome Informatics, 12, 266–267. Kwon, A. T., Hoos, H. H., et al. (2003). Inference of transcriptional regulation relationships from gene expression data. Bioinformatics, 19(8), 905–912. Lycan, D., Mikesell, G., et al. (1994). Differential effects of Cdc68 on cell cycle- regulated promoters in Saccharomyces cerevisiae. Molecular and Cellular Biology, 14(11), 7455–7465. Robinson, P. N., Wollstein, A., et al. (2004). Ontologizing gene-expression microarray data: Characterizing clusters with Gene Ontology. Bioinformatics, 20(6), 979–981. Spellman, P. T., Sherlock, G., et al. (1998). Comprehensive identification of cell cycle- regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Molecular Biology of the Cell, 9(12), 3273–3297. Teixeira, M. C., Monteiro, P., et al. (2006). The YEASTRACT database: A tool for the analysis of transcription regulatory associations in Saccharomyces cerevisiae. Nucleic Acids Research, 34(Database issue), D446–D451. Ukkonen, E. (1995). On-line construction of suffix trees. Algorithmica, 14, 249–260. Whitfield, M. L., Sherlock, G., et al. (2002). Identification of genes periodically expressed in the human cell cycle and their expression in tumors. Molecular Biology of the Cell, 13(6), 1977–2000. An expert system to identify co-regulated gene groups from time-lagged gene clusters using cell cycle expression data Introduction Related work Methods Phase 1: Transforming expression data into event strings Phase 2: Using converted matrix to construct the generalized tree Phase 3: Identifying similar patterns (items) that occur simultaneously among genes (transactions) Phase 4: Further grouping of genes according to their similar pattern positions Phase 5: Functional enrichment Experimental results Clustering with the aim of finding time-lagged co-regulated genes Comparison of other approaches Conclusion Appendix A References 
work_spcnckclmje4doh62z7nmj5udq	----	Development of an expert system for selection of multiphase flow correlations Vol.:(0123456789)1 3 Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 https://doi.org/10.1007/s13202-018-0442-7 O R I G I N A L PA P E R - P R O D U C T I O N E N G I N E E R I N G Development of an expert system for selection of multiphase flow correlations Mohamed A. Abd El‑Moniem1 · Ahmed H. El‑Banbi2 Received: 17 September 2017 / Accepted: 6 February 2018 / Published online: 15 February 2018 © The Author(s) 2018. This article is an open access publication Abstract Our target was to develop an expert system to help petroleum engineers in selecting the most suitable multiphase flow cor- relation in the absence of measured flowing pressure. A large database of pressure points was collected and analyzed using many multiphase flow correlations. The expert system was developed with a set of rules to identify the best correlation for variety of well, flow, and PVT conditions. The expert system is based on the idea of clustering the data and finding the best multiphase flow correlation(s) for each cluster. The error associated with the selected correlation is also quantified for every correlation in each data sub-cluster so the engineer would expect the accuracy of pressure drop prediction when utilizing this approach. Over the entire database, if one multiphase flow correlation is selected, the overall mean absolute percent error ranges between 12.7 and 57.5%, while the range of errors for best correlation(s) in different data sub-clusters range from 0.01 to 3% for most cases with accurate PVT. The expert system was validated by a new set of data. It succeeded in identify- ing the best correlation(s) 70% of the times, and the calculated pressure was more accurate than using one correlation by a factor of 2. Use of the expert system in the validation database gave a mean absolute percent error of 8.8%. This represents approximately one-third of the error value when any single correlation is used over the entire validation dataset. (Error for using single correlation ranges from more than 21 to 29%.) Keywords Multiphase flow · Out flow performance · Multiphase correlations · Production engineering · Nodal analysis List of symbols API American Petroleum Institute (density measurement) At Actual flowing pressure data bbl Barrel Begg Beggs and Brill correlation BHT Bottom-hole temperature d Diameter D Day DRM Duns and Ros modified correlation DRO Duns and Ros original correlation ESP Electrical submersible pump FB Fancher and Brown correlation FT Foot Ft Forecasted flowing pressure data from different 14 correlations °F Degree Fahrenheit GOR Gas/oil ratio GRE GRE correlation HB Hagedorn and Brown correlation HYDRO HYDRO correlation in Inch ID Inner diameter MAPE Mean absolute percent error MPE Mean percent error MB Mukherjee and Brill correlation n Number of data points OD Outer diameter ORK Orkiszewski correlation PET Petroleum experts correlation PET 2 Petroleum experts 2 correlation PET 3 Petroleum experts 3 correlation PET 4 Petroleum experts 4 correlation PET 5 Petroleum experts 5 correlation qL Liquid flow rate Electronic supplementary material The online version of this article (https ://doi.org/10.1007/s1320 2-018-0442-7) contains supplementary material, which is available to authorized users. * Mohamed A. Abd El-Moniem eng_mohamedali1986@yahoo.com 1 Amal Petroleum Company, Cairo, Egypt 2 Cairo University, Giza, Egypt http://crossmark.crossref.org/dialog/?doi=10.1007/s13202-018-0442-7&domain=pdf https://doi.org/10.1007/s13202-018-0442-7 1474 Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 qo Oil flow rate scf Standard cubic feet T.S. Tubing size W.C. Water cut Ɣg Gas specific gravity ε Pipe roughness % Percentage Introduction Multiphase flow correlations are routinely used to calculate pressure (and temperature) drop in the tubing and flow lines for production and injection wells. The pressure drop cal- culations have many applications in artificial lift design and optimization, well deliverability forecasting, and flow assur- ance applications. Over several decades, many investigators developed multiphase flow correlations to calculate pressure drop in tubing at variety of conditions. Each correlation was derived for certain conditions. Several comparison studies between different correlations for specific sets of data are also available in the literature. Predicting multiphase flow pressure drop can sometimes present a challenge for practic- ing engineers. Multiphase flow occurs due to the variations of phases that flow through one pipe. These phases may be oil, water, and gas. Due to the different properties of gas compared with oil and water for viscosities and other properties, the gas tends to flow faster than oil and past the liquid in the upward flow (gas slippage). The in situ liquid volume in the pipe is defined as the liquid holdup. The pressure drop calculations depend on the properties of the mixture of liquid including oil and water in addition to the gas. The total pressure drop is equal to the summation of the hydrostatic, friction, and acceleration pressure drops. The hydrostatic term depends on the mixture density, which is a function of the liquid holdup. The friction term depends on the friction factor for the mixture. The acceleration term appears due to the change in velocity resulting from the change in pressure. The friction calculations depend on the flow regime, which in turn depends on the superficial veloc- ity of liquid and gas phases. The most common flow regimes for vertical wells include single liquid flow, bubble, slug, churn, annular, and mist flow. The possible flow regimes for horizontal wells usually include dispersed bubble, annular, slug, stratified wavy, and stratified smooth. Over the years, both empirical correlations and mecha- nistic models were derived to calculate the pressure drop in pipes. Empirical correlations can be classified into three categories (Ansari et al. 1990): 1. Correlations that assume no slip and do not consider flow patterns. Examples include Poettmann and Carpen- ter (1952) and Fancher and Brown (1963). 2. Correlations that consider the slip but not the flow pat- terns. Examples include Gray (1978) and Hagedorn and Brown (1964 and 1965), etc. 3. Correlations that consider both the slip and the flow pat- terns. Examples include correlations by Duns and Ros (1963) and Beggs and Brill (1973). The mechanistic models, however, usually include some physical phenomena with analytical relations to predict the flow patterns. In these models, mostly analytical expressions are used to calculate the different parameters for each flow regime. Examples for mechanistic models include Aziz et al. (1972) and Ansari et al. (1990). Although many correlations to calculate the pressure drop in pipes are available, clear rules on how to select appropri- ate correlations do not exist in the absence of actual pressure measurements. In this work, an expert system is developed to select the best correlation(s) for specific well and flow conditions. Expert system development The first step in developing the expert system was to collect a large database of pressure drop points covering a wide variety of well and flow conditions. Abd El-Moniem (2016) and Abd El-Moniem and El- Banbi (2015) collected bottom-hole flowing pressure data points from both the literature from Ashiem (1986), Bax- endell and Thomas (1961), Chierici et al. (1974), Hill and Wood (1994), Peffer et al. (1988) and Reinicke et al. (1987) and actual oil wells covering different well and flow con- ditions. The current study used around 2730 oil well data points for developing the expert system and 144 different Table 1 Range of data used in developing and validating the expert system Data for development of the expert system (2730 points) Data for validation of the expert system (144 points) Property Range Property Range qo, bbl/D 5–36,800 qo, bbl/D 15–31,427 GOR, scf/bbl 0–40,000 GOR, scf/bbl 223–6081 W.C., % 0–98 W.C., % 0–76 BHT, °F 82–370 BHT, °F 121–264 API, ° 8–60 API, ° 13–46.5 ɣg 0.57–1.81 ɣg 0.57–0.93 Depth, ft. 250–14,358 Depth, ft. 650–12,862 1475Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 data points for its validation. The range of the data is listed in Table 1. To cover all the points in the database, more than 560 well models were developed using a commercial pipe mod- eling software. We performed around 1042 simulation runs and calculated the predicted bottom-hole flowing pressure using 14 different correlations for each point. We, then, cal- culated the error from each correlation for each point. Table 2 shows the average and mean absolute percent error for each correlation through the entire database. The minimum and maximum absolute percent errors are also shown. The mean absolute percent error ranges from 12.67 to 57.52 for the best and worst correlation over the entire dataset. More importantly, the table shows large range of error between minimum and maximum errors for all correlations. As shown in Figs. 1 and 2 as examples, we show the error for two selected correlations for all the data points in the database to show the correlation performance for different points. The large variations in the error (some points show very small error, while other points show very large error); we can conclude that no correlation is good for all the points within the database. However, each correlation has reason- able prediction capabilities for some range of well and flow Table 2 Average error and average absolute error of the different fourteen correlations for the entire database Correlation Prosper Software Help Manual, Petroleum Experts (2013) Database for development of the expert system (2730 points) Mean percent error % Mean absolute percent error % Minimum absolute error % Maximum absolute error % Duns and Ros modified 11.09 19.42 0.02 729.68 Hagedorn and Brown − 4.74 13.68 0.00 466.54 Fancher and Brown − 10.38 16.53 0.01 408.25 Mukherjee and Brill (1983) 3.85 14.54 0.02 481.03 Beggs and Brill 8.38 13.93 0.00 712.03 Petroleum experts − 1.61 13.31 0.01 605.45 Orkiszewski (1967) − 5.53 15.23 0.00 737.27 Petroleum experts 2 − 0.80 12.88 0.00 568.50 Duns and Ros original 2.82 12.81 0.00 321.14 Petroleum experts 3 − 4.82 13.46 0.00 511.42 GRE − 0.75 12.69 0.00 582.91 Petroleum experts 4 − 2.73 13.34 0.00 320.65 Hydro 3P 49.16 57.52 0.00 954.75 Petroleum experts 5 − 0.75 12.67 0.00 586.94 Fig. 1 Hagedorn and Brown correlation performance through the entire database 0.00 20.00 40.00 60.00 80.00 100.00 1 60 50 82 15 8 38 4 39 1 23 1 24 2 27 7 37 9 88 4 81 0 93 7 10 36 89 1 88 5 83 4 97 3 81 8 87 3 10 10 87 2 91 0 94 7 95 8 88 3 10 57 10 65 10 73 10 82 10 89 13 11 13 88 14 91 15 49 A bs E ro r, % Run # HB Correla�on Performance Fig. 2 Mukherjee and Brill cor- relation performance through the entire database 0.00 20.00 40.00 60.00 80.00 100.00 1 60 50 82 15 8 38 4 39 1 23 1 24 2 27 7 37 9 88 4 81 0 93 7 10 36 89 1 88 5 83 4 97 3 81 8 87 3 10 10 87 2 91 0 94 7 95 8 88 3 10 57 10 65 10 73 10 82 10 89 13 11 13 88 14 91 15 49 A bs E ro r, % Run # MB Correla�on Performance 1476 Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 conditions. Our objective was to divide the data into clusters (groups) and to identify the best correlation(s) within each data cluster. The database was divided into clusters according to well depth, well geometry and tubing size as shown in Fig. 3. Shallow depth is for points that are less than 2000  ft. TVD, deep is for pressure points that are more than 9000 ft. in depth, and the moderate depth is between. We, then, divided the data into sub-clusters according to production criteria. Table 3 shows the selected criteria based on oil rate, water cut, and GOR. Therefore, each cluster of the major ones in Fig. 3 was further divided into smaller segments (sub-cluster) representing the different production conditions given in Table 3. The selection of cluster and sub- cluster ranges was essentially arbitrary. However, we wanted to make sure enough data points were available for every cluster and sub-cluster. We also wanted to have the cluster and sub-clusters bounded by easy values to remember. These sub-clusters within the clusters were given differ- ent code numbers (reference numbers) as listed in Table 4. Each segment of data is then defined by which cluster it lies in and its reference number (sub-cluster number). Table 4 also contains the number of data points available for every sub-cluster. The number of data points is given in the green cells. The sub-clusters represented by the red cells did not contain data points. A Microsoft Access program was developed for the entire dataset to store the actual values of pressure, calcu- lated values from all 14 correlations, and the error for every point. The Access Database was then used to extract a total of 131 sub-clusters for tubular flow and identify the best correlation(s) in each sub-cluster. The clusters (13) are divided based on geometry (well depth, well deviation, and ID). The sub-clusters division is based on flow conditions (oil rate, WC, and GOR). Accord- ing to these divisions, we should theoretically have 13 × 36 sub-cluster. However, many sub-clusters did not contain data (e.g., it is not possible to have high oil rate wells with small ID, and it is not possible to have high GOR wells with high oil rate and high WC). The sub-clusters that contained enough data points to draw conclusions were 131. As listed in Table 4, some sub-cluster contained large number of data points, and others contained fewer data points. We further classified the sub-clusters according to how many points they had into three categories: 1. Strong sub-cluster containing more than 20 points. 2. Moderate sub-cluster containing between 5 and 20 points. 3. Weak sub-cluster containing less than 5 points. The identification of different levels of correlation strength was covered by defining a correlation “strength fac- tor” and covered in Abd El-Moniem and El-Banbi (2015). They also reported the error for every correlation in every sub-cluster. Results The detailed results of sub-clusters can be found in Abd El- Moniem (2016). A sample of the results is listed in Table 5. The tables given in the paper and the appendix are sufficient to fully utilize the expert system. Application of the expert system The use of the expert system is based on selecting the appro- priate cluster as shown in Fig. 3. The selection of the clus- ter is based on (1) well type (vertical, deviated, or horizon- tal), (2) tubing size, and (3) depth. The following step is to select the sub-cluster based on (4) oil production rate, (5) gas/oil ratio (GOR), and (6) water cut as listed in Table 4. The sub-cluster is determined from Fig. 3 and Table 4. The detailed sub-cluster tables to determine the best correlation, its strength, and expected error range are given in “Appendix Fig. 3 Clusters map Table 3 Sub-clusters production criteria for oil wells [after Abd El- Moniem and El-Banbi (2015)] qo, bbl/D W.C., % GOR, scf/bbl < 2500 < 5 < 1000 2500–10,000 5–30 1000–5000 > 10,000 30–70 > 5000 > 70 1477Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 A.” An example on how to use the expert system is also given in “Appendix B.” We also developed a VBA (Visual Basic code) for Micro- soft Excel to facilitate the selection of the appropriate cor- relation. The user can input the six inputs required by the expert system, and the VBA code will automatically find the following information: (1) the cluster and sub-cluster where the user data point is located, (2) the multiphase correlation(s) that should be used to predict the pressure drop for this particular data point, (3) the expected error of this best correlation(s) in the database, and (4) the strength factor for the selected correlation(s) within the sub-cluster. The authors can share the VBA code with interested readers. An example run of the VBA code is shown in Fig.  4. The expert system run for this example case came up with two best correlations (not one). The strength factor for the first correlation (MB) is 43%, and the strength factor for the second correlation (HYDRO) is 35%. It is the same for the mean absolute percent error (1.54% for MB and 2.22% for HYDRO). Validation of the expert system A new dataset with 144 points from 62 wells was used in validating the expert system. All these data points were not used in developing the expert system. The validation data- set was distributed among 48 sub-clusters of data. All 14 Table 4 Sub-clusters reference number and data points for oil wells Criteria Reference Number Cluster Number, Data Availability and Number of Points qo GOR W.C. “bbl/D” “scf/bbl” “%” A B C D E F G H I J K L M < 2,500 < 1,000 < 5 1 29 73 99 154 25 9 23 23 84 2 2 6 10 5-30 2 2 42 30 137 1 6 19 7 52 1 30-70 3 6 168 28 414 9 5 17 7 108 4 6 >70 4 13 9 60 3 8 4 6 25 1, 00 0- 5, 00 0 <5 5 7 17 16 19 4 1 69 4 5-30 6 2 1 10 2 14 1 7 30-70 7 2 24 1 84 2 18 2 14 >70 8 4 2 2 7 6 >5,000 < 5 9 1 2 2 7 5-30 10 1 1 4 2 30-70 11 1 2 6 >70 12 8 3 25 1 19 2, 50 0- 10 ,0 00 < 1,000 < 5 13 22 69 36 118 4 10 7 3 24 8 8 16 5-30 14 14 3 37 3 5 2 7 30-70 15 5 1 12 5 >70 16 1, 00 0- 5, 00 0 <5 17 1 2 4 2 3 1 16 5-30 18 2 30-70 19 1 >70 20 >5,000 < 5 21 1 5-30 22 30-70 23 >70 24 1478 Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 correlations were run for all 144 points, and the best correla- tion (and its error) was determined for each point. The expert system was also run for all 144 points, and the best correla- tion selected by the expert system was compared to the best correlation based on the error calculation. The expert system validation was done on two levels. First, Table 6 shows a summary for all 144 points, and it shows that the expert system was able to predict the best correlation in 70% of the cases. Secondly, the predicted correlation by the expert sys- tem was used to compute the pressure drop for every point in the 144 points. Equations (1) and (2) are used to calculate the mean absolute percent error and the mean percent error. Table 7 shows the error results for the expert system predic- tion over the 144 points (whether the predicted correlation from the expert system was the best one or not) in compari- son with each correlation error over the same data (144 data (1)MAPE = 100 n n∑ i=1 | | | | A t − F t A t | | | | (2)MPE = 100 n n∑ i=1 A t − F t A t Table 4 (continued) Criteria Reference Number Cluster Number and Data Availability qo GOR W.C. “bbl/D” “scf/bbl” “%” A B C D E F G H I J K L M >10,000 < 1,000 < 5 25 1 2 7 23 5 2 6 5-30 26 4 30-70 27 >70 28 1,000-5,000 <5 29 1 2 18 5-30 30 30-70 31 >70 32 >5,000 < 5 33 1 5-30 34 30-70 35 >70 36 Table 5 Sample of complete results for one sub-cluster [after Abd El-Moniem and El-Banbi (2015)] Correlation No. of Points Strength (%) Avg. Error (%) Avg. Abs. Error (%) DRO 15 65 0.17 1.24 ORK 12 52 0.39 1.80 PET2 10 43 -0.84 1.23 PET4 9 39 0.03 0.86 DRM 9 39 0.16 1.61 PET5 8 35 -0.20 0.77 HYDRO 8 35 -0.51 0.82 GRE 8 35 -0.13 0.88 PET 6 26 -0.10 0.75 HB 6 26 -0.10 0.75 PET3 6 26 -0.10 0.75 BEGG 6 26 -0.97 1.46 MB 4 17 -0.28 1.08 Total Points 23 1479Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 points). The mean absolute percent error is around 8.8% using the expert system compared to more than 21% for the best single correlation over the validation dataset. The vali- dation results also show that even in the cases where the best correlation was not selected by the expert system (30% of the cases), the selected correlation always produced reasonably good results compared with the actual values (i.e., when the expert system fails to select the absolute best correlation, it will select one of the second best correlations). Sensitivity of the expert system to uncertain input parameters In this section, we wanted to test the strength of the expert system and how it reacts to inaccurate input data. The meas- ured GOR is usually among the most uncertain parameters in oil well operations. Other parameters that may carry some uncertainty include API gravity of the oil, specific gravity of the gas, and tubing roughness. In each one of these four parameters, a random subset of the database was used to test the effect of the parameter sensitivity on the expert system accuracy. In the sensitivity analysis study, we used a ran- domly selected dataset out of the database. We made sure the selected dataset covered points from most sub-clusters. The sensitivity analysis work involves changing a parameter (e.g., GOR or API) and running all correlations again with the new value for GOR or API, then extracting the data, calculating errors, and comparing correlations with actual data. Enough points were tested to validate and generalize the conclusions regarding the effect of these parameters. For each point within the subset of the database, the parameter was perturbed by 10–20% and the expert system was run to predict the best correlation(s). The suggested correlation by the expert system was also used (with the perturbed input parameter) to predict the pressure drop and compared with the actual value of the pressure. The follow- ing paragraphs summarize our findings from the sensitivity study. GOR sensitivity GOR affects both the input to the expert system and the input to correlation calculations. Four groups of wells were selected: (1) low GOR and high oil rate wells, (2) high GOR Fig. 4 Expert system best correlation/strength factor/mean absolute percent error output window Table 6 Expert system validated points status Table 7 Performance results of the expert system versus all correla- tions for validation dataset Correlation Mean percent error % Mean absolute percent error % Expert system − 0.22 8.82 Duns and Ros modified 17.16 28.06 Hagedorn and Brown 5.81 24.14 Fancher and Brown − 2.04 29.18 Mukherjee and Brill 14.55 21.20 Beggs and Brill 17.53 22.16 Petroleum experts 8.55 23.16 Orkiszewski 2.83 25.79 Petroleum experts 2 9.47 23.30 Duns and Ros original 13.77 21.59 Petroleum experts 3 5.18 24.88 GRE 8.70 22.65 Petroleum experts 4 9.76 21.96 Hydro 3P 11.07 21.74 Petroleum experts 5 9.27 22.33 1480 Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 and high oil rate wells, (3) low GOR and low oil rate wells, and (4) high GOR and low oil rate wells. For low GOR and high rate wells, the expert system could predict the best correlation with a success rate of 56%, when the GOR input (one of the sensitive input parameters) was changed by ± 20%. This reflects that changing GOR for low GOR high rate wells has a moderate effect in changing the order of the best correlation. For high GOR and high rate wells, the expert system could predict the best correlation by 75%. It reflects that changing GOR for high GOR high rate wells by a percent- age of ± 20% has a slight effect in changing the order of the best correlation. For low GOR and low rate wells, the expert system could predict the best correlation by 37.5%. It reflects that changing GOR for low GOR low rate wells by a percentage of ± 20% has a significant effect in changing the order of the best correlation. For high GOR and low rate wells, the expert system could predict the best correlation 95% of the cases. We can safely conclude that a ± 20% error in GOR for those wells had minor effect in changing the order of the best correlation. GOR was found to have considerable effect on the calcu- lated pressure drop from all correlations, especially for low GOR wells (in both high oil rate and low oil rate wells). This is probably due to changing flow regimes in the calculation of pressure drop. Although GOR here is the producing GOR, solution gas/oil ratio (Rs) is responsible for part of the pro- ducing GOR, especially in low GOR wells. Therefore, for low GOR wells, flow regimes may change rapidly due to the release of gas with the decline of pressure and temperature in the tubing and consequently change the order of the best correlation as listed in Table 8. Other parameters sensitivity Other parameters were investigated to check their uncer- tainty on the expert system and the error of the calculated pressure by the selected correlation(s). These parameters include API, gas specific gravity and tubing roughness. Changing the API of the oil by ± 10% had slight effect on both the expert system selection and the calculated pressure drop by the selected correlation. In more than 90% of the cases, the expert system was successful in picking up the best correlation even with the ± 10% error in API. For gas specific gravity, changing gas specific gravity with ± 10% had minor effect on changing the order of the best correlation as the expert system predicted the best cor- relation with 90%. The exception was for low GOR low rate wells, as the expert system could obtain the best correlation with only 37.5%. For tubing roughness, the expert system could obtain the best correlations 95% of the time. The exception was for high GOR high rate wells, where the expert system could predict the best correlation with only 37.5%. The selection of the range of sensitivity analysis param- eters was mainly based on measurement errors in operations of oil fields. GOR measurement error may reach ± 20%, while API and specific gravity of gas are usually estimated with much more certainty. Table 8 summarizes the results of the sensitivity analysis of these parameters and their effect on changing the order of the best correlation and conse- quently the capability of the expert system to predict the best correlation even with some measurements errors. Conclusions We utilized a large dataset to test the accuracy of 14 cor- relations among the most widely used multiphase flow cor- relations. We also used an approach to take advantage of the fact that each correlation is strong under certain conditions of well geometry, flow conditions, and PVT properties. The approach is based on data clustering. The following conclu- sions can be made: Table 8 Effect of different sensitive parameters on correlation prediction Well flow condition GOR effect API effect Gas specific gravity effect Tubing roughness effect Low GOR and high production rate Medium Low Low Low Low GOR and low production rate High Low High Low High GOR and high production rate Low Low Medium Medium High GOR and low production rate Low Low Low Low 1481Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 1. The overall mean absolute percent error is around 12.7% for the best multiphase flow correlation in the entire dataset, while the range of errors for the best correlation(s) in different data sub-clusters is from 0.01 to 3% for most cases with accurate PVT. 2. In the validation dataset, the mean absolute percent error was 8.8% when the expert system was used. This repre- sents approximately one-third of the error value when any single correlation is used over the entire dataset. (Error for using single correlation ranges from more than 21 to 29%.) 3. From the sensitivity runs, we found that GOR, API, and gas specific gravity may have different degrees of effectiveness on the correlation selection. PVT data are probably one of the most important multiphase flow cor- relations input data affecting the accuracy of the correla- tion. 4. In practice, an accurate gas rate should be determined through production test separator especially for low rate oil wells as GOR has important effect on pressure drop results. GOR was found to have impact on changing the best correlations for different clusters of oil data in case of low GOR wells. However, GOR has minor effect on selection of best correlations for high GOR wells. 5. Tubing roughness was found to have small impact on changing the best correlations for different clusters of oil well data (except for high GOR and high oil rate wells). Open Access This article is distributed under the terms of the Crea- tive Commons Attribution 4.0 International License (http://creat iveco mmons .org/licen ses/by/4.0/), which permits unrestricted use, distribu- tion, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Appendix (A) To use the expert system, the user needs to collect the well information and flow conditions. Required well information is: (1) well depth, (2) well type (vertical, deviated, or hori- zontal), and (3) tubing ID. Required flow conditions are: (4) oil rate, (5) water cut, and (6) producing GOR. The well information is used with the clusters map presented in Fig. 3 to define the data cluster. Then flow conditions are used to define the “sub-cluster” reference number from Table 4. The reference number is then used with the set of tables in this appendix (Tables 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, Table 9 Best correlations for cluster (A) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” A1 29 PET 5/GRE 31 1.9 MB 28 2.1 A2 2 ORK 100 1.0 A3 6 GRE/PET 5 33 0.9 PET 2 33 1.0 PET 4 33 1.4 HYDRO 33 2.4 BEGG 33 2.5 ORK 33 15.7 MB 33 25.9 A8 4 BEGG 75 3.2 A13 22 GRE 82 1.9 PET 5 82 2.3 Table 10 Best correlations for cluster (B) Sub-cluster Total points Correlation Strength factor “%”  Mean absolute percent error “%” B1 73 PET 5 37 0.9 PET 4 36 0.9 GRE 34 0.8 B2 42 ORK 31 2.5 B3 168 MB 42 2.9 GRE 39 2.7 PET 5 39 2.7 B4 13 DRM 38 1.9 PET 2 31 0.4 PET 5 31 0.6 B5 7 PET 3 57 1.8 PET 5 43 1.8 B6 2 DRO 100 2.3 B7 24 BEGG 75 27.2 DRO 67 29.9 B13 69 PET 5 64 1.2 GRE 62 1.1 B14 14 BEGG 50 1.9 PET 50 2.4 B15 5 GRE 40 1.2 PET 2 40 1.8 BEGG 40 2.0 http://creativecommons.org/licenses/by/4.0/ http://creativecommons.org/licenses/by/4.0/ 1482 Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 21) to define the best correlation, its strength factor, and its mean absolute percent error. An example for the use of the expert system is given in “Appendix B.” Tables through 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, and 21 show detailed results of the best correlation(s) for the different data clusters. Appendix (B) Example A vertical well is completed with a 2.375 in ID tubing, and gauge depth of 12,024 feet is flowing oil at a rate of 282 bbl/D with 0% water cut and 699 scf/bbl GOR. Use the expert system to find the best correlation and its expected accuracy. Using Fig. 3 in the paper, the well lies in data cluster E. With the flow conditions, we find from Table 4 the sub-clus- ter reference number to be 1. Therefore, the data given in this example lie in sub-cluster E1. We use Table 13 because it represents the data cluster for this point; then, we go to row 1, which represents the sub-cluster. We find from this table that the best correlation is MB with expected mean abso- lute percent error of 5.76%. Notice that this mean absolute percent error was the error for this best correlation over the Table 11 Best correlations for cluster (C) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” C1 99 PET 3 29 5.8 DRO 26 8.4 C2 30 FB 47 6.3 ORK 47 6.6 C3 28 BEGG 54 4.6 MB 54 14.3 C4 9 HYDRO 56 19.6 C5 17 MB 53 4.5 BEGG 47 5.7 C6 1 DRO 100 28.61 C8 2 DRM 50 1.0 PET 2 50 1.6 BEGG 50 2.0 ORK 50 18.9 C10 1 DRO 100 1.59 C11 1 GRE 100 0.05 PET 5 100 0.1 C12 8 PET 2 63 8.7 BEGG 63 11.3 C13 36 GRE 58 2.1 PET 5 58 2.5 C15 1 DRO 100 1.0 C25 1 DRM 100 1.3 Table 12 Best correlations for cluster (D) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” D1 154 GRE 34 1.7 PET 5 30 1.5 PET 4 29 1.4 D2 137 FB 30 3.1 HB 28 3.2 ORK 26 6.3 GRE 24 2.2 D3 414 BEGG 35 1.4 DRO 34 1.7 PET 4 33 1.3 GRE 32 1.3 D4 60 MB 63 5.1 HB 62 5.3 D5 16 PET 3 38 4.3 ORK 31 2.7 D6 10 DRO 60 6.2 D7 84 ORK 80 18.9 D10 1 ORK 100 8.4 HYDRO 100 8.7 D13 118 PET 31 0.9 PET 2 29 1.3 GRE 28 1.0 D14 37 BEGG 57 2.1 D15 12 DRO 50 0.7 BEGG 42 0.9 D17 1 FB 100 0.8 D25 2 ORK 100 2.2 1483Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 Table 13 Best correlations for cluster (E) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” E1 25 MB 72 5.8 BEGG 56 6.6 E2 1 DRM 100 1.53 PET 4 100 1.88 PET 5 100 1.95 E3 9 DRM 56 0.7 ORK 44 1.43 HYDRO 44 1.81 E5 19 DRM 42 2.0 E6 2 ORK 100 1.6 E9 2 HYDRO 100 2.4 E13 4 PET 4 100 4.5 FB 100 4.5 PET 3 100 4.5 E14 3 FB 100 46.08 ORK 100 47.41 E15 5 HB 100 131.75 E18 2 ORK 100 5.1 E25 7 HYDRO 100 143.03 E26 4 HB 100 122.64 E29 1 PET 3 100 14.4 PET 2 100 13.3 BEGG 100 11.3 Table 14 Best correlations for cluster (F) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” F1 9 FB 89 2.7 GRE 89 3.5 PET 4 89 3.6 F2 6 ORK 100 10.3 F3 5 FB 100 0.8 HB 100 1.3 F4 8 FB 100 0.1 BEGG 100 0.4 PET4 100 0.5 F5 4 PET 4 100 0.6 PET 3 100 0.6 F6 1 PET 4 100 0.2 PET 5 100 0.5 F6 14 ORK 71 15.6 F13 10 DRM 50 4.6 F17 2 DRM 100 1.1 F29 2 DRM 100 1.1 BEGG 100 2.0 Table 15 Best correlations for cluster (G) Sub-Cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” G1 23 FB 65 1.4 G2 19 PET 58 1.5 HB 53 1.5 PET 3 53 1.6 G3 17 MB 53 1.5 G4 4 DRM 50 1.7 G7 2 FB 100 23.3 G8 2 DRO 100 5.0 G11 2 HB 100 47.0 FB 100 47.0 PET 3 100 47.3 G12 3 DRO 67 35.1 G13 7 HB 100 0.6 GRE 71 0.4 PET 5 71 0.5 G17 4 HB 100 1.2 PET 4 75 1.4 Table 16 Best correlations for cluster (H) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” H1 23 DRO 65 1.2 ORK 52 1.8 H2 7 DRO 57 1.3 DRM 57 3.9 H3 7 DRO 100 5.6 DRM 86 5.7 H4 6 DRM 50 1.4 DRO 50 2.3 GRE 33 1.4 PET 5 33 1.5 H5 1 MB/BEGG 100 1.94 H13 3 ORK 100 2.3 H17 2 BEGG 50 0.5 DRM 50 1.2 1484 Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 Table 17 Best correlations for cluster (I) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” I1 84 MB 43 1.5 HYDRO 35 2.2 I2 52 PET 3 48 1.1 FB 44 1.7 HB 42 1.6 I3 108 BEGG 35 2.4 DRM 34 5.2 I4 25 MB 64 6.4 GRE 60 6.5 I5 69 DRM 67 2.3 I6 1 FB 100 0.3 I7 18 FB 67 18.7 I8 7 DRM 43 2.1 I9 7 DRO 100 0.4 FB 71 0.8 PET 4 71 1.2 I10 4 GRE 100 3.5 PET 4 50 2.6 I11 6 FB 83 7.7 HB 83 7.7 PET 3 83 8.3 I12 25 PET 2 40 7.2 I13 24 FB 50 1.5 I14 5 MB 80 0.6 PET 4 80 0.9 PET 80 1.4 I17 3 PET 4 100 0.6 PET 5 67 0.7 I19 1 ORK 100 1.5 FB 100 2.3 I25 23 FB 74 1.7 HB 74 1.8 ORK 57 1.6 Table 18 Best correlations for cluster (J) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” J1 2 ORK/DRM 50 2.0 J5 4 PET 3 50 2.5 DRM 50 31.4 J7 2 PET 3 50 0.0 BEGG 50 3.1 J12 1 BEGG 100 24.0 J17 1 BEGG 100 3.4 DRM 100 3.5 Table 19 Best correlations for cluster (K) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” K1 2 ORK 100 8.4 K2 1 ORK 100 0.6 K6 7 PET 2 43 1.6 PET 3 29 0.4 K7 14 FB 57 11.9 K8 6 MB 50 2.3 BEGG 50 4.8 K10 2 GRE 100 0.7 PET 5 100 1.0 K12 19 BEGG 53 22.7 DRM 42 30.7 K13 8 ORK 88 14.3 HB 63 13.7 PET 3 63 14.4 K17 16 HB 67 2.2 PET 2 63 2.3 K21 1 PET 2 100 0.4 MB 100 0.9 K25 5 ORK 100 2.6 FB 100 2.7 HB 100 2.8 K29 18 PET 2 56 0.9 PET 3 56 1.1 K33 1 BEGG 100 0.2 ORK 100 0.6 Table 20 Best correlations for cluster (L) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” L1 6 DRM 50 2.5 FB 50 10.2 L3 4 BEGG 100 1.0 DRO 75 1.1 MB 75 1.2 L13 8 PET 3 75 0.3 HB 75 0.7 L14 2 HYDRO 100 1.0 BEGG 50 0.0 PET 2 50 0.6 L25 2 BEGG 100 2.4 DRO 50 3.1 1485Journal of Petroleum Exploration and Production Technology (2018) 8:1473–1485 1 3 available data points within this sub-cluster. The available data points are 25 (which are also given in Table 13). We used the suggested correlation by the expert system to calculate the bottom-hole flowing pressure for this well from the tubing head pressure and PVT data. The calculated pres- sure by the chosen correlation was 3066.57 psi with absolute percent error 2.7%. Table 22 gives the calculated pressure and absolute percent error by the expert system chosen cor- relation (MB) compared with the measured pressure and the calculated pressure by other correlations. References Abd El-Moniem MA (2016) Evaluation and proper selection of mul- tiphase flow correlations. MS Thesis, Cairo University, Egypt Abd El-Moniem MA, El-Banbi AH (2015) Proper selection of mul- tiphase flow correlations. In: SPE 175805 presented at North Africa technical conference and exhibition, Cairo, 14–16 Sept Ansari AM, Sylvester ND, Shoham O, Brill JP (1990) A comprehen- sive mechanistic model for upward two phase flow in wellbores. In: SPE 20630 presented at the 65th annual technical conference and exhibition, New Orleans, 23–26 Sept, pp 151–165 Ashiem H (1986) MONA, an accurate two phase well flow model based on phase slippage. In: SPE 12989, SPE production engineering, May, pp 221–230 Aziz K, Govier GW, Fogarasi M (1972) Pressure drop in wells produc- ing oil and gas. J Can Pet 11:38–48 Baxendell PB, Thomas R (1961) The calculation of pressure gradients in high rate flowing wells. J Pet Technol 13:1023–1028 Beggs HD, Brill JP (1973) A study of two phase flow in inclined pipes. J Pet Technol 25:607–617 Chierici GL, Ciucci GM, Sclocchi G (1974) Two phase vertical flow in oil wells prediction of pressure drop. J Pet Technol 26:927–938 Duns H Jr., Ros NCJ (1963) Vertical flow of gas and liquid mixtures in wells. In: Section II, Paper 22-PD 6, proceedings of the sixth world petroleum congress, Frankfurt, 19–26 June, pp 451–465 Fancher GH Jr, Brown KE (1963) Prediction of pressure gradients for multiphase flow in tubing. Soc Pet Eng J 3:59–69 Gray HE (1978) Vertical flow correlation in gas wells, user’s manual for API 14B surface controlled subsurface safety valve sizing computer program, 2nd edn. American Petroleum Institute, Dal- las (Appendix B) Hagedorn AR, Brown KE (1964) The effect of liquid viscosity in two phase vertical flow. J Pet Technol 16:203–210 Hagedorn AR, Brown KE (1965) Experimental study of pressure gradi- ents occurring during continuous two phase flow in small diameter vertical conduits. J Pet Technol 17:475–484 Hill TJ, Wood DG (1994) Slug flow: occurrence, consequences, and prediction. In: SPE 27960 presented at the University of Tulsa Petroleum Engineering symposium, OK, USA, 29–31 Aug, pp 53–62 Mukherjee H, Brill JP (1983) Liquid holdup correlations for inclined two phase flow. J Pet Technol 35:1003–1008 Orkiszewski J (1967) Predicting two phase pressure drops in vertical pipe. J Pet Technol 19:829–838 Peffer JW, Miller MA, Hill AD (1988) An improved method for cal- culating bottomhole pressures in flowing gas wells with liquid present. In: SPE production engineering, November, pp 643–655 Poettmann FH, Carpenter PJ (1952) The multiphase flow of gas, oil and water through vertical flow string with the application of the design of gas lift installation. Drilling and Production Practice. American Petroleum Institute, New York Prosper Software Help Manual, Petroleum Experts (2013) Reinicke KM, Remer RJ, Hueni G (1987) Comparison of measured and predicted pressure drops in tubing for high water cut gas wells. In: SPE production engineering, August, pp 165–177 Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Table 21 Best correlations for cluster (M) Sub-cluster Total points Correlation Strength factor “%” Mean absolute percent error “%” M1 10 ORK 60 21.8 M3 6 DRO 100 0.7 M13 16 PET 3 69 0.2 MB 69 0.7 M14 7 BEGG 100 0.2 HYDRO 100 0.3 DRO 86 0.9 M25 6 DRO 100 1.9 BEGG 100 2.7 Table 22 Calculated pressure and absolute percent error from differ- ent correlations compared to actual data Correlation Calculated pressure, psi Absolute percent error, % Measured (actual) 2985 – MB (predicted by expert system) 3066.57 2.7 DRM 3210.15 7.5 HB 2734.28 8.4 FB 2673.61 10.4 BEGG 3317.26 11.1 PET 2806.17 6.0 ORK 3256.22 9.1 PET 2 2897.19 2.9 DRO 3354.18 12.4 PET 3 2796.87 6.3 GRE 3297.77 10.5 PET 4 3313.33 11.0 HYDRO 3389.8 13.6 PET 5 3281 9.9 Development of an expert system for selection of multiphase flow correlations Abstract Introduction Expert system development Results Application of the expert system Validation of the expert system Sensitivity of the expert system to uncertain input parameters GOR sensitivity Other parameters sensitivity Conclusions References 
work_spfqxtkldncvxmsrx42a4dmkwy	----	PII: 0888-613X(88)90096-5 108 Abstracts "window": The similarity between input and output is a measure of the compatibility between test results and standard. A window is a subset of a segment and represents the range of values that are of particular interest. In most cases the boundaries of the windows cannot be defined with certainty. Most decision processes involve many alternatives that may depend on several dimensions. For each alternative, a "stack of windows" is defined by specifying a window for each dimension. Then decision making is equivalent to filtering the characteristics of a given situation through the stacks. The acceptability of each alternative is determined by an evaluation function that assesses the extent of individual matches. Available knowledge-acquisition techniques support this form of knowledge represen- tation. Fuzzy windows facilitate the establishment of consensus among experts and the analysis of composite solutions. Classification systems built with this approach allow the use of different evaluation functions and support the development of a database that can be searched. Knowledge representation by fuzzy windows is substantiated by findings in decision making and creative thinking. Address correspondence to J. L. Chameau. Backward Chaining with Fuzzy Goals and Rules B r i a n S c h o t t a n d T h o m a s W h a l e n Decision Science L a b o r a t o r y , Georgia State University, B o x 316, A t l a n t a , Georgia, 30303-3083 This paper describes a knowledge engine that employs a backward-chaining control structure that is generalized to fuzzy variables and rules. The engine extends earlier fuzzy linguistic variable processing systems developed by the authors that employ fuzzy linguistic variables and fuzzy logic in their data and rules. The backward control mechanism permits more sophisticated, efficient use of diagnostic knowledge-centered decision support systems. Design criteria, implementation notes, and possible extensions are discussed. This is a publicly owned, research-oriented system; the procedures, dubbed BACHFUGUE (written in the APL programming language), are available from the authors. Related systems exist or are under development. BACHFUGUE is being interfaced with a synthetic grammar system that manages the linguistic terms in SAPIR. Although originally written in LISP, SAPIR has been translated into APL for compatibility with BACHFUGUE. Relations, as well as if-then rules, are powerful expressive mechanisms in symbolic knowledge systems. Fuzzy relations in the present system must be decomposed (edited) into if-then rules. A more automatic means of dealing with such relations is being explored. Address correspondence to Brian Schott. A New Concept of Fuzzy Rule-Based Expert Systems S h a h r a m S e y e d a n d P . A . R a m a m o o r t h y SyStelligence Consulting Group, 805 R h o d e s Hall, D e p a r t m e n t o f Abstracts 109 Electrical and Computer Engineering, University o f Cincinnati, Cincinnati, Ohio 45221-0030 The concept of fuzzy rule-based expert systems based on soft (interactive) rather than hard (noninteractive) fuzzy operators is developed. A methodology is provided for the inclusion of certainty factors (CFs) in the expert systems associated with the belief attributed to the data and production rules. A new fuzzy implication operator for implementing soft production rules and a new fuzzy inference system for fuzzy firing the soft rules are proposed. The advantages of this system are its simplicity and modularity (more complex forms of fuzzy rules can easily be represented using the suggested structure). Being highly modular, the system lends itself to VLSI implementation. This base system is used to develop a forward fuzzy-fired rule-based expert system shell, called FUZZTRAN (FUZZy TRANslation), in which each forward fuzzy rule is translated into a prolog clause. By executing this translated fuzzy PROLOG program, forward reasoning can be done without a rule interpreter. A system prototype using a military situation assessment fuzzy knowledge base is under investigation. Address correspondence to Shahram Seyed. On Fuzzy Graph Searching A d n a n K . S h a o u t Department o f Electrical and Computer Engineering, Syracuse University, Syracuse, New York 13244-1240 Graph searching finds applications in many areas. Fuzzy graphs are the most natural means to incorporate the uncertainties inherent in certain problems. Pattern recognition and artificial intelligence are just two examples of many areas where uncertainties should be addressed in solving realistic problems. The objective of this paper is to introduce a searching algorithm for fuzzy graphs that finds the shortest path. The complexity of the algorithm is O ( n 2) for a graph with n nodes. The complexity of the algorithm can be reduced to O(n) ifparallelism in the algorithm is considered. Fuzzy Process Modeling M a r i a n S. S t a c h o w i c z a n d M a r i a E . K o c h a f i s k a Institute o f Electrotechnics and Electronics, Technical University o f Cracow, ul. Warszawska 24, 31-155 Krak6w, Poland The problem of the influence of constructional parameters in a fuzzy model on the accuracy of process description in terms of fuzzy relations is considered. A method of constructing this relation on the basis of a linguistic model of the object is described. Particular attention is given to those constructional parameters that do not have an explicit mathematical interpretation, i.e., the definition of a fuzzy implication and interpretation of the sentence connective ALSO. The influence of those parameters on the accuracy of the fuzzy model is considered. The dependence of the effect upon character, among others, of which relations are the best models as a function of the shape of the performance curve of the described model is also investigated. Address correspondence to Marian S. Stachowicz. 
work_spr26jzvmfeyxlyrt3lp6um3ly	----	Empirical Findings on Ontology Metrics D. Rodŕıgueza,b∗, M.A. Siciliaa, E. Garćıaa aDepartment of Computer Science University of Alcalá, Ctra. Barcelona, Km. 31.6 28871 Alcalá de Henares, Madrid, Spain bDepartment of Computing and Communication Technologies Oxford Brookes University Wheatley Campus, Oxford OX33 1HX, UK {daniel.rodriguezg,msicilia,elena.garciab}@uah.es Abstract Ontologies are becoming the preferred way of representing, dealing and reasoning with large volumes of information in several domains. In consequence, the creation, evaluation and maintenance of ontolo- gies has become an engineering process that needs to be managed and measured using sound and reliable methods. As part of any ontol- ogy engineering or revision process, metrics can play a role helping in identifying possible problems or incorrect use of ontology elements, along with providing a kind of quality assessment that complements reviews requiring expert inspection. However, in spite of the fact that there are a number of ontology metric proposals described in the liter- ature, there is a lack of empirical studies that provide a basis for their interpretation. This paper reports a systematic exploration of exist- ing ontology metrics from a large set of ontologies extracted from the Swoogle search engine. The Ontorank value used inside the Swoogle to rank search results is used as a contrast for some existing metrics. The results show that the existing proposed metrics evaluated in general do not seem to be indicators of that ranking, but they can be helpful to identify particular kinds of ontologies. Keywords. Ontology, metrics, Swoogle, Ontorank ∗Corresponding author 1 1 Introduction Ontologies [4] are explicit representations of domain concepts and their re- lationships. More formally, an ontology defines the conceptualisations of a problem domain and a set of constraints on how terms can be combined to model the domain, using a formal language. Ontologies are nowadays used for different purposes in different systems, and as a consequence, they have become artifacts in the development of software. In the Software Engineering domain, metrics play an important role in design, development and identifying software defects of future maintenance problems. Like it happens in the software engineering tasks, generally the construction of ontologies follows an iterative and incremental approach and a number of metrics has been proposed to measure different aspects such as reliability, reusability, cohesion, etc. (see for example [14]). However, un- like it happens in software engineering domain, there is a lack of empirical studies that validate how the different metrics are capable of judging qual- ity properties and their interpretation. Inquiry on the relation of quality indicators and metrics is in consequence an important open challenge for a more systematic approach to developing ontology-based software systems. The problem with ontology metrics is that ontologies are very heterogeneous in their structure, objectives and level of formality. In consequence, there is a need to carry out exploratory studies that could eventually result in in- sights about the metrics applicable to different kinds of ontologies and their interpretation for each of them. This paper reports on on such an exploratory study, carried out by mea- suring a large set of OWL ontologies obtained from the Swoogle reposi- tory together with a measure of their popularity using Ontorank [2]. The study has been carried out by implementing and using an open source soft- ware framework for computing ontology metrics expressed in the Ontology Web Language (OWL)1, that was already used in a previous preliminary study [9]. OWL has been chosen as the language is a W3C recommendation and has rapidly become the standard language for expressing formal ontolo- gies. The results of the study point out that current metrics are not directly interpretable in a general way, but there seems to be different categories of ontologies that could be discernible from the value of the metrics. The rest of the paper is structured as follows. Section 2 covers the background on ontology metrics and some related work. Next, Section 3 describes the statistical analyses performed together with some findings. 1 http://www.w3.org/TR/owl-guide/ 2 Finally, Section 4 concludes the paper and outlines future research work. 2 Background Ontologies have been a resource in the field of artificial intelligence since the 70’s. However since the inception of the Semantic Web, ontologies have become the principal recourse to integrate and deal with online information, and a new set of standards to formally represent them has been developed and gained widespread adoption. The Ontology Web Language (OWL) is one of such standards that belongs to a family of knowledge representa- tion languages created for the Semantic Web [1]. The OWL has reached the status of W3C (World Wide Web Consortium) recommendation. From a technical point of view OWL extends the RDF (Resource Description Framework) and RDF-S (RDF Schema) allowing us to integrate a variety of applications using XML as interchange syntax. The OWL has evolved since its inception in 2004 to a new specification in 2009 (OWL2) with different sub-languages depending on the expressiveness and reasoning capabilities provided. There are however common elements and OWL ontologies are composed of (i) classes that can be nested as sets of individuals, (ii) individuals as instances of classes, i.e., objects of the domain and (iii) properties as binary relations between individuals. It is also possible to specify property domains, cardinality ranges and reasoning on ontologies. From these elements a number of authors have proposed metrics to measure the quality of ontologies. Several authors have proposed metrics related to ontologies or adapted from Object-oriented software. For example, Yao et al. [17] proposed a set on metrics to measure the cohesion in ontologies such as the number of root classes, number of leaf classes and average depth of inheritance tree of all leaf nodes. Authors validated the metrics both theoretically and empirically. Their empirical validation consisted of comparing statistically the results of the metrics and subjective evaluation of eighteen evaluators. Tatir et al. [13] propose three orthogonal dimensions to evaluate the quality of an ontology. The first dimension is quality of the ontology schema representing the real world. The second dimension evaluates the content of the ontology, i.e, how well populated is the ontology and if it reflects the real world. Finally, the third dimension evaluates if the of instances and relations agree with the schema. Schema metrics proposed by the authors include relationship, attribute and inheritance richness. Instance metrics are divided into (i) knowledge base metrics such as class richness, average 3 population and cohesion and (ii) class metrics which include importance, fullness, inheritance richness, relationship richness connectivity and read- ability. Authors analysed two general purpose ontologies SWETO and TAP as well as the GlycO (Glycomics Ontology) ontology. Vrandečić and Sure [15] propose a framework for designing ontologies though a process of normalisation and considering stable metrics, i.e., met- rics considering the open world assumption. Metrics and metric frameworks can help with processes or methodologies for the design of ontologies or ontological tools. A well-known methodology for the evaluation of ontological adequacy of taxonomic relationships is On- toClean defined by Guarino and Welty [5]. Another process is the OntoMet- ric process defined by Lozano-Tello and Gómez-Pérez [7]. OntoMetric is a process to select ontologies using a broad range of metrics, not only those that can be collected form the actual ontologies. A suite a of ontology metrics to measure design complexity has been re- cently proposed [18], reporting an analysis of desirable properties for metrics described by Weyuker, and providing results of an empirical evaluation in which the metrics proposed are shown to be discriminating ontologies that are considered by the authors to be of different complexity. In another direction, Ding et al. [3, 2] have developed Swoogle, a se- mantic search engine in which the popularity of a semantic web document is defined using the Ontorank measure adapted from the well-known PageRank algorithm [12]. That ranking metric is an external criteria that can be used to contrast ontology metrics. The problem with the ranking is that its inter- pretation as a quality measure is not evident and has not been empirically studied. However, as other quality measures are not available for ontologies, it represents an interesting opportunity to contrast existing metrics. We also developed an open source framework to extract the OWL on- tology metrics called OntoMetrics. The OntoMetrics is an open source Java implementation that makes use the Java libraries of Protégé2 an ontology editor that provides an application programmers interface (API) for loading, saving and manipulating OWL and RDF files. In this paper we have use this framework to collect metrics from a large number of ontologies together with their popularity measured using Ontorank. 2 http://protege.stanford.edu/ 4 3 Empirical Analysis of Ontology Data In this section, the metrics gathered are described and then the results of the statistical analysis is reported. Only basic metrics related to the main elements of ontologies have been implemented and used. Particularly, metrics dealing with classes inferred (i.e., defined classes) or axioms were not considered as a majority of the ontologies available for the study lacked these kinds of constructs. 3.1 Dataset In order to empirically analyse the relationships between the previously de- fined metrics (Section 2), we downloaded around 1,500 ontologies defined in OWL using Swoogle3, which is a search engine for the Semantic Web. A crawler was developed ad hoc that submits a blank query to Swoogle and then navigates the pages downloading automatically the OWL files. Only ontologies expressed in OWL were considered as they represent the majority of the available ones, and allows for an homogeneous analysis. Using the previously commented Ontometrics framework, we collected the following metrics: • No. of Classes (noc) is a simple count of the number of classes (|C|) contained in the ontology. • No. of Instances (noi) is a simple count of the number of instances (|I|) contained in the ontology. • No. of Properties (nop) is a simple count of the number of properties (|p|) contained in the ontology. • No. of Root Classes metric (norc) [17] corresponds to the number of root classes (without superclasses) explicitly defined. Let C the classes in an ontology: norc = |Ci|, ¬∃Cj | Ci * Cj. The range of this metric is from 1 to |C|. There is at least one root and the larger the number of root classes, the more diverse the ontology will be. • No. of Leaf Classes metric (nolc) [17] is the sum of all leaf classes, i.e., those without subclasses, in an ontology. Let C the set of classes in an ontology: nolc = |Ci|, ¬∃Cj | Cj * Ci. A a root class without out inheritance can also be considered a leaf class, the range of this metrics is from 1 to |C|. 3 http://swoogle.umbc.edu/ 5 • Average Population metric (ap) [13] measures the average distribution of instances across all classes. Formally, it is defined as: ap = |I| |C| According the authors, this metric is supposed to be used in conjunc- tion with the Class Richness metric as an indication if there is enough information in the ontology. • Class Richness metric (cr) [13] is the ratio between the number of classes that have instances (|C′|)divided by the total number of classes. This metric provides an indication of how many instances are really related to classes defined in the schema. • Explicit Depth of Subsumption Hierarchy (dosh). This metric cor- responds to the maximum length of hierarchical relationships in the ontology explicitly defined, i.e., longer path from a root class. • Relationship Richness metric (rr) [13] is defined as the ratio of the number of relationships divided by the sum of the number of subclasses plus the number of relationships: rr = |P| |SC|+|P|, where |P | is the number of relationships and SC is the number of subclasses (or the number of inheritance relationships). According to the authors, this metric reflects the diversity of relations and placement of relations in the ontology. An ontology that contains many relationships other than class-subclass relations (values close to 1) is richer than a taxonomy with only class-subclass relationships (values close to zero). • Inheritance Richness metric (ir) [13] is defined as the average number of subclasses per class. Formally, ir is defined as: ir = ∑ Ci∈C |H c(C1, Ci)| |C| where Hc(C1, Ci)| is the number of subclasses (C1) for a class (Ci) and the divisor (C) is the total number of classes. According to the authors, this metric represents the distribution of information across different levels of the ontology inheritance tree and serves as an indi- cator of how well knowledge is grouped into different categories and subcategories in the ontology. Values close to zero indicate flat or hor- izontal ontologies representing perhaps more general knowledge while large values represent vertical ontologies describing detailed knowledge of a domain. 6 • Ontorank [2, 3] is an adaptation from Google’s PageRank [12] to em- ulate user’s navigation behaviour at document level of granularity. OntoRank for a semantic document a is defined as OntoRank(a) = wPR(a) + ∑ x∈OTC(a) wPR(x) where OTC(x) is the set of semantic documents that transitively a imports or extends, and wPR(a) is defined as: and wPR(a) = (1 − d) + d ∑ linkto(x,a) wPR(x) · f(x, a)∑ link(x,−,y) f(x, y) f(x, a) = ∑ link(x,l,a) weight(l) where link(a, l, b) represents a link from a to b semantic documents with semantic tag l, linkto(x, a) set of documents directly linked to a, weight(l) is a user specified navigation preference on semantic links with tag l, d is the probability of navigating from one document to another. The Ontorank was directly provided by Swoogle, and represents an in- dicator of quality analogous to the PageRank [11] algorithm used in the Google search engine. Table 1 summarizes the main metrics reported on the literature, classified by the kind of elements considered. Table 5 in Appendix A shows an excerpt from the data collected. 3.2 Statistical Analysis As stated previously, we downloaded a large number of ontologies from Swoogle from which we could extract the metrics previously defined. The total number of ontologies from which we could collect all metrics was 1,413. Table 2 shows the descriptive statistics for the metrics collected. The first interesting observation is that the average, median and third quartile values for most metrics are very low (see Table 2 and Figure 1), i.e., the vast majority of ontologies are small and quite flat (the ontology 7 Table 1: Summary of Ontology Metrics used in this Work Metric Acronym No. of Classes noc No. of Instances noi No. of Properties nop No. of Root Classes norc Number of Leaf Classes nolc Average Population ap Class Richness cr Explicit Depth of Subsumption Hierarchy dosh Inheritance Richness ir Relationship Richness Metric rr Ontorank or Table 2: Descriptive Statistics Avg Med V ar StdDev Min Max Range Q3 StdSkw StdKrt or 2.74 0.85 968.5 31.12 0.64 962.31 961.67 1 379 5300.1 ap 1.34 1 8.83 2.97 0 67 67 1.20 170.44 1469.95 cr 0.54 0.75 0.20 0.44 0 1 1 1 -1.84 -13.9 dosh 2.54 1 5.75 2.4 1 30 29 3 40.13 113.57 ir 0.34 0 0.17 0.42 0 1.83 1.83 0.75 9.97 -9.31 noc 36.11 6 6166.3 78.53 1 945 944 34 80.44 293.25 noi 28.13 6 9522.8 97.59 0 991 991 9 94.85 328.43 nolc 27.46 5 3815.8 61.77 0 823 823 27 90.87 367.80 nop 24 0 2997.8 54.75 0 952 952 23 107.11 690.53 norc 6.69 5 127.37 11.29 0 194 194 6 131.7 857.18 rr 2.78 1 13.57 3.68 0 71 71 5 156.8 1326.26 8 0 33.5 67 508 75 69 28 11 143 579 0 0.5 1 1 15.5 30 ap cr dosh 4 3 0 0 1 0 0.92 1.83 1 473 945 0 35.5 71 ir noc rr Figure 1: Histograms hierarchy is in average very small). An exception is the ClassRichness (cr) metric with a ’u’ shape which may be a reflection of two different ways of designing or using an ontology, either the instances are kept separately of the classes that define them or both classes and instances are stored together with the ontology. Also in general, the distributions of the metrics are not following any identifiable distribution. Values of skewness and kurtosis in Table 2 allow us to discard normal distributions for all of them. The few large values are statistically outliers as shown in Figure 2 by using the red colour. When analysing actual ontologies with large values, these correspond to special ontologies such as those in the Open Biomedi- cal Ontologies (OBO) repository. For example, the NCI Thesaurus (ncithe- saurus.owl) and the disease ontology.owl are both extremely large ontologies with a special purpose. There are reasons for this that could be hypothesized to come from some differences on the languages and approaches to engineer ontologies in the biomedical domain [6]. Table 3 shows the Spearman rank correlation coefficient between the different metrics (also graphically in Figure 3). It can be observed that the OntoRank metric is not correlated with any other metric, and therefore, it seems that the popularity of an ontology cannot be identified by the set of metrics used in this study. We have analysed this relationship with 9 Figure 2: Box-and-whisker Plots Table 3: Spearman Rank Correlation Coefficients cr dosh ir noc noi nolc nop norc rr ap 0.6878 -0.4387 -0.4735 -0.3157 0.6261 -0.2800 -0.2754 0.1959 0.4837 cr -0.6995 -0.7151 -0.5607 0.2789 -0.5105 -0.6500 0.0787 0.7250 dosh 0.9776 0.8073 0.1370 0.7460 0.7959 0.0998 -0.7443 ir 0.7853 0.0848 0.7148 0.7709 0.0288 -0.7526 noc 0.3173 0.9812 0.6583 0.4275 -0.4580 noi 0.3369 0.2525 0.5574 0.1828 nolc 0.6154 0.4613 -0.3822 nop 0.2487 -0.5948 norc 0.2400 other data mining approaches using several Weka’s [16] classifiers such as regression or classification trees to confirm this effect. The relationship between the rest of the metrics were contrasted using factor analysis. Table 4 shows the resulting factor loadings using princi- pal component analysis with standarised values as an extraction method in which 3 components were extracted. The first component is characterized by high values in dosh, ir, noc that roughly measure depth and breadth of the subsumption hierarchy, i.e., metrics related with the schema of the ontology. The second component is characterized by a high correlation with the ap and noi so it is more related with the content of the ontology rather than the schema. and properties and also with relationship richness that relates both of them. In the third component, the largest values are also related to the ap and noi. There are some threats to the validity of this study. First of all, we analysed single files, but Semantic Web documents can import or extend other documents in order to be able of working together. Coupling [10] or cohesion [8] metrics should be also be included. Furthermore, the metrics 10 Figure 3: Graphical representation of the Spearman Rank Correlations Table 4: Factorial Analysis of the Ontological Metrics Factor 1 Factor 2 Factor 3 ap -0.0786494 0.551809 -0.681197 cr -0.727543 0.33865 0.167987 dosh 0.855587 -0.205325 -0.164372 ir 0.825037 -0.29855 -0.300246 noc 0.835051 0.129089 0.290564 noi 0.344157 0.653781 -0.458953 nolc 0.797677 0.165676 0.337331 nop 0.676127 0.259735 0.0893104 norc 0.437119 0.61869 0.338713 rr -0.351663 0.566569 0.2992 11 computed were extracted from the OWL files, with no associated reasoning process. This entails that the metrics do not consider elements that would be potentially inferred from reasoners, e.g. automatic classifications or sub- sumption relationships that are not explicit. While this may seem a major limitation, the fact that most ontologies do not contain axioms or defined classes (classes expressed through logical necessary and sufficient conditions) suggests that explicit elements are enough for the empirical assessment of metrics in the current state of development of ontologies. 4 Conclusions and Future Work This paper addressed the empirical analysis of a large number of ontologies through a collection of metrics in order to propose interpretations for the metrics that can be subject to hypothesis contrast and eventually lead to a practical applicability of these metrics. The OntoRank was used as a con- trast as the only available assessment mechanism that can be automatically computed without requiring expert assessment of the ontologies. It should be noted that such kind of indicators cannot substitute but complement ex- pert of logical assessments as those that have been already reported in the literature. To do so, we developed an open source tool, called Ontometrics, used to collect the set of metrics presented in this paper together with their Ontoranking value provided by Swoogle. The results showed that there is no correlation between the popularity of an ontology and the metrics used in this work. There are some peculiarities for example most ontologies are quite flat and small, perhaps reflecting that most of the ontologies are not mature. Some exceptions can be found in the OBO foundry4 repository which contains mature ontologies from the biomedical domain. In this case, although it is difficult to establish concrete thresholds to define when an ontology needs to be improved, such metrics helped in identifying ontologies for special purposes such as thesaurus. Future work will be directed to further develop our metrics framework with new metrics and update the framework to the new OWL API5. There is also room for further analysis of the metrics, for example, analysing how ontologies evolve by carrying out longitudinal studies accounting for the evolution of ontologies. 4 http://www.obofoundry.org/ 5 http://owlapi.sourceforge.net/ 12 References [1] Tim Berners-Lee, James Hendler, and Ora Lassila. The semantic web. Scientific American, 2001. [2] Li Ding, T. Finin, A. Joshi, Yun Peng, Rong Pan, and P. Reddivari. Search on the semantic web. Computer, 38(10):62–69, Oct. 2005. [3] Li Ding, Tim Finin, Anupam Joshi, Rong Pan, R. Scott Cost, Yun Peng, Pavan Reddivari, Vishal Doshi, and Joel Sachs. Swoogle: a search and metadata engine for the semantic web. In Proceedings of the thirteenth ACM international conference on Information and knowledge manage- ment (CIKM’04), pages 652–659, New York, NY, USA, 2004. ACM. [4] Thomas R. Gruber. Toward principles for the design of ontologies used for knowledge sharing. International Journal of Human-Computer Studies, 43(5-6):907–928, 1995. [5] Nicola Guarino and Christopher Welty. Evaluating ontological decisions with ontoclean. Communications of the ACM, 45(2):61–65, 2002. [6] Robert Hoehndorf, Anika Oellrich, Michel Dumontier, Janet Kelso, Dietrich Rebholz-Schuhmann, and Heinrich Herre. Relations as pat- terns: bridging the gap between OBO and OWL. BMC Bioinformatics, 11:441, 2010. [7] A. Lozano-Tello and A. Gómez-Pérez. ONTOMETRIC: A method to choose the appropriate ontology. Journal of Database Management, 15(2):1–18, 2004. [8] Yinglong Ma, Beihong Jin, and Yulin Feng. Semantic oriented ontology cohesion metrics for ontology-based systems. Journal of Systems and Software, 83(1):143–152, 2010. [9] N. Manouselis, M.A. Sicilia, and D. Rodŕıguez. Exploring ontology metrics in the biomedical domain. In Workshop on Knowledge Repre- sentation and Decision Making (KREAM), International Conference in Computational Science (ICCS), pages 2313–2322. Elsevier, May 2010. [10] A.M. Orme, H. Tao, and L.H. Etzkorn. Coupling metrics for ontology- based system. IEEE Software, 23(2):102–108, mar. 2006. [11] Larry Page, Sergey Brin, R. Motwani, and T. Winograd. The pagerank citation ranking: Bringing order to the web, 1998. 13 [12] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical Report 1999-66, Stanford InfoLab, November 1999. [13] Samir Tartir, I. Budak Arpinar, Michael Moore, Amit P. Sheth, and Boanerges Aleman-Meza. Ontoqa: Metric-based ontology qual- ity analysis. In IEEE Workshop on Knowledge Acquisition from Dis- tributed, Autonomous, Semantically Heterogeneous Data and Knowl- edge Sources, 2005. [14] Mike Uschold and Michael Grninger. Ontologies: principles, methods, and applications. Knowledge Engineering Review, 11(2):93–155, 1996. [15] Denny Vrandec̆ić and York Sure. How to design better ontology metrics. In Enrico Franconi, Michael Kifer, and Wolfgang May, editors, The Se- mantic Web: Research and Applications, volume 4519 of Lecture Notes in Computer Science, pages 311–325. Springer Berlin/Heidelberg, 2007. [16] I.H. Witten and E. Frank. Data Mining: Practical machine learning tools and techniques. Morgan Kaufmann, San Francisco, 2005. [17] Haining Yao, Anthony Mark Orme, and Letha Etzkorn. Cohesion met- rics for ontology design and application. Journal of Computer Science, 1(1):107–113, 2005. [18] Hongyu Zhang, Yuan-Fang Li, and Hee Beng Kuan Tan. Measuring design complexity of semantic web ontologies. Journal of Systems and Software, 83(5):803–814, 2010. A Data Collected Example 14 T a b le 5 : E x ce rp t o f D a ta C o ll ec te d O W L o r a p m cr m d o s h m ir m n o cm n o im n o lc m n o p m n o r cm r r m h tt p :/ / in fe re n c e w e b .s ta n fo rd .e d u / 2 0 0 4 / 0 7 / iw .o w l 9 6 2 .3 0 6 0 .0 6 0 .0 6 6 0 .8 2 3 3 2 2 2 7 6 6 0 .9 3 h tt p :/ / in fe re n c e -w e b .o rg / 2 .0 / p m l- ju st ifi c a ti o n .o w l 4 5 5 .8 8 3 1 0 .0 3 0 .0 3 5 0 .8 7 3 9 1 2 7 7 0 5 0 .9 h tt p :/ / in fe re n c e -w e b .o rg / 2 .0 / p m l- p ro v e n a n c e .o w l 3 8 2 .8 4 2 8 0 .0 3 0 .0 3 5 0 .9 3 1 1 2 0 4 9 3 0 .9 h tt p :/ / w w w .m in d sw a p .o rg / g la p iz c o / te ch n ic a l. o w l 1 2 5 .4 1 8 5 0 .1 2 0 5 0 .8 8 1 7 2 1 0 1 6 4 4 h tt p :/ / e b iq u it y .u m b c .e d u / o n to lo g y / p u b li c a ti o n .o w l 7 0 .0 3 2 0 7 0 .2 3 0 3 0 .8 5 1 3 3 1 1 3 2 2 0 .9 7 h tt p :/ / in fe re n c e -w e b .o rg / 2 .0 / d s. o w l 6 7 .2 6 8 7 8 1 1 1 0 1 1 1 2 1 1 h tt p :/ / e b iq u it y .u m b c .e d u / o n to lo g y / p e rs o n .o w l 6 5 .4 6 8 1 2 0 .1 4 0 3 0 .9 5 2 1 3 1 7 1 0 1 0 .9 1 h tt p :/ / w w w .d a m l. o rg / se rv ic e s/ o w l- s/ 1 .1 / P ro c e ss .o w l 2 9 .5 5 3 5 1 0 .1 8 0 .1 3 0 .6 8 5 0 9 3 7 6 5 1 6 0 .8 2 h tt p :/ / e b iq u it y .u m b c .e d u / o n to lo g y / p ro je c t. o w l 1 4 .4 9 5 6 9 1 0 2 0 .7 5 4 4 3 1 0 1 0 .9 3 h tt p :/ / w w w .d a m l. o rg / se rv ic e s/ o w l- s/ 1 .1 / G ro u n d in g .o w l 1 4 .0 0 8 4 6 0 .1 6 0 .0 9 3 0 .6 7 5 8 9 4 3 9 0 1 9 0 .8 3 h tt p :/ / w w w .d a m l. o rg / se rv ic e s/ o w l- s/ 1 .1 / P ro fi le .o w l 1 2 .6 1 8 9 8 0 .2 3 0 .0 9 3 0 .6 6 5 3 1 2 3 9 8 4 1 8 0 .8 3 h tt p :/ / e b iq u it y .u m b c .e d u / o n to lo g y / p h o to .o w l 1 1 .5 6 0 0 8 1 0 1 0 1 1 1 7 1 1 h tt p :/ / w w w .d a m l. o rg / se rv ic e s/ o w l- s/ 1 .1 / S e rv ic e .o w l 9 .6 4 4 8 3 5 0 0 1 0 4 0 4 1 1 4 1 h tt p :/ / e b iq u it y .u m b c .e d u / o n to lo g y / a ss o c ia ti o n .o w l 9 .0 0 2 4 6 6 3 0 1 0 1 3 1 4 1 1 h tt p :/ / w w w .d a m l. o rg / se rv ic e s/ o w l- s/ 1 .0 / S e rv ic e .o w l 7 .8 3 1 4 5 9 2 0 1 0 4 8 4 1 1 4 1 h tt p :/ / tr u st .m in d sw a p .o rg / o n t/ tr u st .o w l 7 .3 7 7 7 0 8 0 0 1 0 1 0 1 1 6 1 1 h tt p :/ / a n n o ta ti o n .s e m a n ti c w e b .o rg / is w c / is w c .o w l 6 .8 0 1 3 3 9 1 .5 2 0 .1 8 3 0 .6 4 3 3 5 0 2 7 3 5 1 2 0 .9 2 h tt p :/ / w w w .b io p a x .o rg / re le a se / b io p a x -l e v e l1 .o w l 6 .7 6 5 5 6 1 0 .1 5 0 5 0 .9 6 2 7 4 1 8 4 9 2 0 .9 h tt p :/ / w w w .m in d sw a p .o rg / 2 0 0 4 / w w w 0 4 p h o to .o w l 6 .7 0 7 6 1 6 0 0 4 0 .7 5 2 0 0 1 3 1 8 4 0 .8 6 h tt p :/ / w w w .m in d sw a p .o rg / 2 0 0 4 / te rr o rO n t. o w l 6 .1 1 4 2 0 7 0 0 3 0 .6 5 3 1 0 2 5 8 4 1 1 0 .9 2 h tt p :/ / e b iq u it y .u m b c .e d u / o n to lo g y / n e w s. o w l 5 .0 7 4 0 9 2 1 0 1 0 1 1 1 4 1 1 h tt p :/ / w w w .w 3 .o rg / 2 0 0 4 / 0 8 / P re se n ta ti o n s. o w l 4 .8 1 8 7 8 5 0 .1 2 0 1 0 8 1 8 8 5 1 h tt p :/ / e b iq u it y .u m b c .e d u / o n to lo g y / c o n fe re n c e .o w l 4 .0 2 1 8 5 1 0 1 0 1 1 1 1 0 1 1 h tt p :/ / w w w .d a m l. o rg / se rv ic e s/ o w l- s/ 1 .1 / A c to rD e fa u lt .o w l 3 .9 8 1 4 3 0 .2 2 0 .0 9 3 0 .6 5 5 4 1 2 4 0 9 1 1 9 0 .8 4 h tt p :/ / w w w .o p e n m e ta d ir .o rg / o m 2 / p ri m -3 .o w l 3 .9 6 4 2 8 0 .0 7 0 3 0 .8 2 2 8 2 2 0 8 9 1 0 .9 3 h tt p :/ / w w w .l e h ig h .e d u / z h p 2 / 2 0 0 4 / 0 4 0 1 / u n iv -b e n ch .o w l 3 .5 6 6 3 8 7 0 0 5 0 .9 3 4 3 0 3 2 3 2 5 0 .8 9 h tt p :/ / w w w .x sp a n .o rg / o b o .o w l 3 .2 5 3 4 2 1 0 0 .1 7 2 0 .3 3 6 0 5 1 2 2 0 .9 7 h tt p :/ / w w w 4 .n c su .e d u / n v d e sa i/ o w l/ P ro to c o l. o w l 2 .9 7 8 3 3 4 0 .0 3 0 .0 3 3 0 .5 3 3 2 1 2 8 4 6 1 4 0 .7 9 h tt p :/ / d a ta p o rt a l. u c a r. e d u / sc h e m a s/ e sg .o w l 2 .9 0 7 6 3 5 0 0 5 0 .5 8 9 8 0 8 4 1 7 4 4 1 0 .9 5 .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . 15 
work_spv2h432ejh57fu6rif6mlqq34	----	Comparisons of machine learning techniques for detecting malicious webpages Expert Systems with Applications 42 (2015) 1166–1177 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Comparisons of machine learning techniques for detecting malicious webpages http://dx.doi.org/10.1016/j.eswa.2014.08.046 0957-4174/� 2014 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. H.B. Kazemian ⇑, S. Ahmed Intelligent Systems Research Centre, School of Computing, London Metropolitan University, United Kingdom a r t i c l e i n f o a b s t r a c t Article history: Available online 16 September 2014 Keywords: K-Nearest Neighbor Support Vector Machine Naive Bayes Affinity Propagation K-Means Supervised and unsupervised learning This paper compares machine learning techniques for detecting malicious webpages. The conventional method of detecting malicious webpages is going through the black list and checking whether the web- pages are listed. Black list is a list of webpages which are classified as malicious from a user’s point of view. These black lists are created by trusted organizations and volunteers. They are then used by modern web browsers such as Chrome, Firefox, Internet Explorer, etc. However, black list is ineffective because of the frequent-changing nature of webpages, growing numbers of webpages that pose scalability issues and the crawlers’ inability to visit intranet webpages that require computer operators to log in as authen- ticated users. In this paper therefore alternative and novel approaches are used by applying machine learning algorithms to detect malicious webpages. In this paper three supervised machine learning techniques such as K-Nearest Neighbor, Support Vector Machine and Naive Bayes Classifier, and two unsupervised machine learning techniques such as K-Means and Affinity Propagation are employed. Please note that K-Means and Affinity Propagation have not been applied to detection of malicious web- pages by other researchers. All these machine learning techniques have been used to build predictive models to analyze large number of malicious and safe webpages. These webpages were downloaded by a concurrent crawler taking advantage of gevent. The webpages were parsed and various features such as content, URL and screenshot of webpages were extracted to feed into the machine learning models. Computer simulation results have produced an accuracy of up to 98% for the supervised techniques and silhouette coefficient of close to 0.96 for the unsupervised techniques. These predictive models have been applied in a practical context whereby Google Chrome can harness the predictive capabilities of the classifiers that have the advantages of both the lightweight and the heavyweight classifiers. � 2014 Elsevier Ltd. All rights reserved. 1. Introduction Web security threats are increasing day by day (Facebook, 2010; Malware, 2011; Sood & Enbody, 2011). The open nature of the Internet allows malicious webpages to pose as ‘safe webpages’ and consequently some users are misled to think that these webpages are safe. As the use and the speed of the Internet increased over the last two decades, web developers have increased the usage of images, JavaScript and other elements. The Google search engine is a clear example. At the beginning, it had very few elements. There are now more elements, graphics, stylesheets and the HTML specifications which have been added as time went on. Initially, the only way to create a webpage was by static HTML. JavaScript was then added for user interactivity. ActiveX, Silverlight, Java Applets, etc. were further added to include features. For example, ActiveX allowed browsers to host various executables which enabled users to read PDF and various other file formats such as Flash, DivX, etc. Web developers started using the integrated development environ- ments that generated a considerable HTML markup language and this increased the HTML payload. The number of browsers increased and some of these browsers, especially Internet Explorer had their own quirks and needed more work from the developers. These factors raised the complexity of the webpages that led to potential increase in how webpages are ‘adversely affected’ and have become malicious. Cross Site Scripting (XSS) injects malicious code from an unexpected source. These malicious codes can get hold of the cookies, browsing history and then send them over to the mali- cious webpage. Thus the user’s privacy is jeopardized. There have been many attempts to prevent this sort of attacks (Lucca, Fasolino, Mastoianni, & Tramontana, 2004). XSS not only affects the user but also it affects the server. The webpage is used as the vehicle to transfer infections to multiple users. The malicious http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.08.046&domain=pdf http://dx.doi.org/10.1016/j.eswa.2014.08.046 http://dx.doi.org/10.1016/j.eswa.2014.08.046 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 1167 code then executes in the user’s browser. The problem has been intensified with the addition of scripting capabilities that did not exist at the beginning of the history of web browsing. With the addition of scripting capabilities, the users are benefitting with a better user experience but have become prone to these additional security problems. These scripts run on users’ brows- ers. The web developer may build a webpage only using HTML, but an attacker can still inject scripts to make it susceptible to scripts. These scripts can then access the cookies that are used for authentication. The XSS vulnerability therefore affects the users and the webpages. Take for example, a user visits a web- page and decides to purchase a product. The user adds the items to the basket and would like to checkout. Then he fills in a form to register. Each of these users is uniquely identifiable by the webpage through the use of cookies. The criminal will be able to look at the cookies and impersonate the user and buy the prod- ucts, without the knowledge of the user. By the time the user has realized the problem, the money has already been dispatched from the user’s account. Almost all HTML tags are wrapped by ‘greater than’ and ‘less than’. To write the script tag, these two characters are needed. There are several combinations of characters that can be generated (Braganza, 2006). The combinations are quite vast and will likely to increase. The combinations of letters that generate the letters are dependent on browser version and the default language. Braganza (2006) states that the browser cannot be trusted because of these extensive possibilities and some precautions are required. To cleanse, data entered are encoded and data displayed are both decoded, this process is known as ‘sanitization’. In terms of how the webpage is deployed to the user, the operations team have to make sure that the firewall or any other forms of preventative measures are kept up to date. Another security threat that is very difficult to detect is clickjacking. This is a relatively new threat that has become more prevalent through the advancement of modern browsers. The interesting thing about clickjacking is that it does not use security vulnerabilities, rather uses the browsers most common feature such as hyperlinks. The user is encouraged to click a link to a webpage. But this webpage has two webpages one is dis- played to the user and the other one the malicious webpage which is hidden from the user (Hansen & Grossman, 2008). The hidden webpage executes the malicious code even though the user thinks that the information is on the right webpage. This technique is very hard to detect by inspecting the source code and there have not been many successful ways to prevent it from happening. Drive-by-download occurs without the knowledge of a user and the downloaded file is used for malicious purposes. This malicious executable installs itself on users’ computer. This is a very popular method that has been used by Harley and Bureau (2008) to spread malware infection on the Internet. There are three components in the attack, the web server, the browser and the malware. An attacker finds a web server to serve the malware. The user who vis- its a webpage hosted in this web server is then exploited by the webpage, and some code utilizes software loopholes to execute commands on the user’s browser are injected. The web server sub- sequently provides the malware that is downloaded by the brow- ser. As a result, the browser that is targeted will have a known vulnerability that the attacker will try to exploit. Internet Explorer had many instances of ActiveX loopholes that the attackers had used and are still using and Harley and Bureau (2008) have pro- vided potential solutions. The first solution is to completely isolate the browser from the operating system so that the arbitrary codes are not at all executed on the browser. Another solution is for web crawlers to visit webpage and see whether they are hosting any malware content. But the attackers can avoid by using a URL that does not have a corresponding hyperlink. Crawlers by its nature only visit URLs that have a corresponding hyperlink. Browsers these days use publicly available blacklists of mali- cious webpages. These blacklists are updated after a few days or even a month. These gaps allow for webpages to be affected while being unnoticed to the crawler. At this point, the users will also get affected, because the browser thinks the webpage to be secure, as it has never been in the blacklist. Take another scenario where a regular webpage may be hacked and injected with malicious code visible only to some users or a group of users of an organization or a country. The blacklists will not be able to ‘blacklist’ those either. Some crawlers do not validate the JavaScript code because the code is executed on the server and not in a browser. This allows client vulnerabilities to pass through easily. Even though some of the scripts which are assumed to be safe, these scripts can load mali- cious scripts remotely and then execute them on the computer. Some scripts create iFrames and then load external webpages that are malicious (Provos, Mavrommatis, Rajab, & Monrose, 2008). These external webpages then gets hold of the cookies and steal the identity. The users then browse this malicious webpage and get infected and are then easily tracked by remote users from somewhere else. The users also may run malicious executables without even knowing that the executables have already access to the system and are monitored from somewhere else. Webpages are the common victims to all these threats that have been described above. The features in a webpage can indicate whether it is malicious or not. Researchers have studied and analyzed a large number of features with or without machine learning tech- niques described below. Kan and Thi (2005) carried out one of the first research work that utilized machine learning to detect malicious webpages. This work ignored webpage content and looked at URLs using a bag-of-words representation of tokens with annotations about the tokens’ posi- tions within the URL. A noteworthy result from Kan and Thi’s research is that lexical features can achieve 95% accuracy of the page content features. Garera, Provos, Chew, and Rubin (2007)s work used logistic regression over 18 hand selected features to clas- sify phishing URLs. The features include the presence of red flag key words in the URL, features based on Google’s page ranking, and Google’s webpage quality guidelines. Garera et al. achieved a classi- fication accuracy of 97.3% over a set of 2500 URLs. Although this paper has similar motivation and methodology, it differs by trying to detect all types of malicious activities. This paper also uses more data for training and testing, as described in the subsequent sec- tions. Spertus (1997) suggested an alternative approach and endeavored to identify malicious webpages, Cohen (1996) employed the decision trees for detection and Dumais, Platt, Heckerman, and Sahami (1998) utilized inductive learning algo- rithms and representations for text categorization. Guan, Chen, and Lin (2009) focused on classifying URLs that appear in webpages. Several URL-based features were used such as webpage timing and content. This paper has used similar techniques but applied them to webpages which have much more complex structures with better accuracies. Mcgrath and Gupta (2008) did not construct a classifier but performed a comparative analysis of phishing and non-phish- ing URLs. With respect to data sets, they compared non-phishing URLs drawn from the DMOZ Open Directory Project to phishing URLs from Phishtank (2013) and a non-public source. The features they analyzed included IP addresses, WHOIS thin records (contain- ing date and registrar provided information only), geographic infor- mation, and lexical features of the URL (length, character distribution and presence of predefined brand names). The differ- ence is that this paper utilizes different types of features to add to the novelty. Provos et al. (2008) carried out a study of drive-by exploit URLs and used a patented machine learning algorithm as a pre-filter for virtual machine (VM) based analysis. This approach is based on heavyweight classifiers and is time consuming. Provos et al. (2008) used the following features in computer simulation, 1168 H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 content based features from the page, whether inline frames are ‘out of place’, the presence of obfuscated JavaScript, and finally whether iFrames point to known malicious sites. Please note, an ‘iFrame’ is a window within a page that can contain another page. In their evaluations, the machine learning based pre filter can achieve 0.1% false positives and 10% false negatives. Provos et al.’s approach is very different to this paper as the features are primarily focused on iFrames. Bannur, Saul, and Savage (2011)s research has some similarities to this paper but it uses a very small dataset and furthermore this paper utilizes various other types of features. Abbasi, Zhang, Zimbra, Chen, and Nunamaker (2010) and Abbasi, Zahedi, and Kaza (2012) ran some classification to detect fake med- ical sites but the size of dataset was very small, whereas this paper focuses on detecting any type of malicious webpages. Fu, Wenyin, and Deng (2006) and Liu, Deng, Huang, and Fu (2006)s research considered the visual aspects of a webpage to determine whether the page is malicious or not. Le, Markopoulou, and Faloutsos (2010) detected phishing webpages only using the URLs. And, Ma, Saul, Savage, and Voelker (2009b) looked at online learning to detect malicious webpages from URL features. As the security improves, the attackers will devise new ways to avoid barriers raised by administrators and web developers. To fur- ther improve security, an automated tool is required in order to detect the vulnerabilities. One approach to automation is for web developers to secure and enhance their webpages. But there are limits to the extent that developers can work to secure webpages. Web developers are bound by the web frameworks they use (Okanovic, Mateljan, & May, 2011). If the web frameworks fail to take preventative measures, the users’ machines get infected and the webpages become vulnerable. This paper takes the research in malicious webpages described above further by applying three supervised machine learning techniques such as Naive Bayes Clas- sifier, K-Nearest Neighbor and Support Vector Machine, and two unsupervised machine learning techniques like K-Means and Affin- ity Propagation, and compares the results. The novel unsupervised techniques of K-Means and Affinity Propagation have not been applied to detection of malicious webpages by any other research- ers in the past. Moreover, to add to the novelty, the research utilizes more complex structures for better accuracies, more data for train- ing and testing, and employs both the lightweight and the heavy- weight classifiers for detection of all types of malicious activities. 2. Machine learning models This section provides a brief overview of the five machine learning techniques that have been used in this paper; they are Naive Bayes Classifier, K-Nearest Neighbor, Support Vector Machine, K-Means and Affinity Propagation, briefly described below. The computer simulation results of these machine learning methods are discussed in Section 3. 2.1. Naive Bayes Classifier Naive Bayes Classifier uses the Bayes’ theorem. The classifier assumes that all the features are independent of each other. It learns pðCkjxÞ by modeling pðxjCkÞ and pðCkÞ, using Bayes’ rule to infer the class conditional probability (Bayes & Price, 1763). The model is yðxÞ¼ arg max x pðCkjxÞ¼ arg max x pðxjCkÞ� pðCkÞ ¼ arg max x YD i¼1 pðxijCkÞ� pðCkÞ ¼ arg max x XD i¼1 log pðxijCkÞþ log pðCkÞ ð1Þ The training in this paper was carried out using Gaussian likelihood (alternative options for training include multivariate likelihood and multinomial likelihood). pðxjCkÞ¼ YD i¼1 Nðlik; rikÞ ð2Þ where: � Ck are the classes where C = {C1, C2, . . . , Ck}. � Nðlik; rikÞ is the normal distribution. � l is the mean of the Gaussian distribution. � r is the standard deviation of the Gaussian distribution. The complexity of the model OðNMÞ is as such that each training instance must be visited and each of its features ought to be counted. For non-linear problems, it can only learn linear boundaries for multivariate/multinomial attributes. With Gaussian attributes, quadratic boundaries can be learnt with unimodal distributions (Jordan, 2002). Naive Bayes is generally used in text classification and is one of the most widely used classification algo- rithm in machine learning because it is fast and space efficient, which is also noticed in the simulation results described in Section 3. 2.2. K-Nearest Neighbor K-Nearest Neighbor works in such a way that the label of a new point x̂ is classified with the most frequent label t̂ of the k nearest training instances. It is modeled as t̂ ¼ arg max C X i:xi2Nkðx;x̂Þ dðti;CÞ ð3Þ where: � Nkðx; x̂Þ k points in x closest to x̂. � Euclidean distance formula: ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPD i¼1ðxi � x̂iÞ 2 q . � d(a, b) 1 if a = b; 0o/w. The model does not require any optimization and it trains itself using cross validation to learn the appropriate k. k regularizes the classifier, as k ? N the boundary becomes smoother. OðNMÞ is used as space complexity, since all training instances and all their features need to be kept in memory. K-Nearest Neighbor uses a very simple technique for classification, and cannot handle large training dataset as shown in the results section. It can handle non-linear boundaries easily (Altman, 1992). 2.3. Support Vector Machines SVM was developed by Cortes and Vapnik (1995) and it is widely regarded as one of the most effective models for binary classification of high dimensional data. SVM and indeed any other supervised classifiers use a similar technique to classify webpages shown in Fig. 1. SVM is modeled as hðxÞ¼ b þ XN n¼1 yiai Kðx; xiÞ ð4Þ where � h(x) is the distance of the decision boundary. � b is the bias weight. � a is a coefficient that maximize the margin of correct classifica- tion on the training set. � N is the number of features. � K is the kernel function. � x is a feature vector. Fig. 1. Supervised machine learning architecture for classifying webpages. H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 1169 The positive or negative sign of this distance indicates the side of the decision boundary on which the example lies. The value of h(x) is limited to predict a binary label for the feature vector x. The model is trained by initially specifying a kernel function K(x, x0) and then computing the coefficients ai that maximize the margin of correct classification on the training set. The required optimization can be formulated as an instance of quadratic pro- gramming, a problem for which many efficient solutions have been developed (Chang & Lin, 2012). 2.4. K-Means K-Means is a geometric clustering, hard-margin algorithm, where each data point is assigned to its closest centroid. It is mod- eled using hard assignments rnk 2 {0, 1} so that "n P krnk = 1, i.e. each data point is assigned to one cluster k (MacQueen, 1967). The geometric distance is calculated using the Euclidean distance, l2 norm: jjxn � lkjj 2 ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi XD i¼1 ðxni � lkiÞ 2 vuut ð5Þ where � lk is cluster centroid. � D is the no of points. � x is one of the points. The mini-batch K-Means algorithm uses mini batches to reduce the computation time while still attempting to optimize the same objective function. Mini-batches are subsets of the input data, ran- domly sampled in each training iteration. These mini-batches dras- tically reduce the amount of computation required to converge to a local solution. Mini batch k-means converges faster than K-Means, but the quality of the results is reduced. In practice, the difference in quality can be quite small, as shown in the results section. 2.5. Affinity Propagation Affinity Propagation is an unsupervised algorithm created by Dueck (2009) where each data point acts as centroids and these data points choose the number of clusters. The following is an example of how to represent the centroid for datapoint i ci 2f1; . . . ; Ng ð6Þ The algorithm maximizes the following function SðcÞ¼ XN i¼1 sði; ciÞþ XN k¼1 dkðcÞ ð7Þ The similarity of each point to the centroid is measured by the first term in Eq. (7) and the second term is a penalty term denoted as -1. If some data point i has chosen k as its exemplar that is ck – k, but k has not chosen itself as an exemplar i.e. ck = k, then the following constraints could be presented in Eq. (8). dkðcÞ¼ �1 if ck–k but 9i : ci ¼ k 0 otherwise � ð8Þ A factor graph can represent the objective function and it is possible to use N nodes, each with N possible values or with N2 binary nodes. Each variable node ci sends a message to each feature node dk and each factor node dk sends a message to each variable node ci. The number of clusters is controllable by scaling the diagonal term S(i, i), which shows how much each data point would like to be an exemplar. Although Affinity Propagation was developed very recently it has a very good performance as shown later in the results section. The overall architecture of the unsupervised machine learn- ing is outlined in Fig. 2. 3. Results The architecture of the computer simulation carried out for this paper is presented in Fig. 3. The experiment was conducted on an Intel Xeon E3-1220 4 Cores � 3.1 GHz computer with 12 GB of RAM. First, 100,000 webpages were downloaded using a crawler Fig. 3. The architecture of Web Application Classifier (WAC). Fig. 2. Unsupervised machine learning architecture for classifying webpages. 1170 H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 and then converted into feature vectors. Then a tool named Web Application Classifier (WAC) took these vectors as inputs, and applied the machine learning algorithms described in the previous section to create the predictive models. These predictive models read vectors of the new webpages to produce an output that indi- cated whether a webpage is safe or not. The Sections 3.1–3.5 describe which features were used and how the webpages were pre- processed and cleansed before placing inside the predictive models. 3.1. Data sources The downloaded webpages were divided into two sets, malicious and safe. The source for the list of safe webpages was gathered primarily from Alexa (2013), the malicious ones were collated primarily from Phishtank (2013) and two types of repre- sentation of each webpage were created. One contained all the HTML codes and other only had English characters. 3.2. Features 3.2.1. Semantic This paper utilizes vector space representation to extract the semantic features, as it is commonly used in classification and clustering. Salton, Wong, and Yang (1975) carried out research using a vector space model for automatic indexing of webpages, Cohen and Singer (1999) used a context-sensitive learning method H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 1171 for categorizing text, and Sahami, Dumais, Heckerman, and Horvitz (1998) utilized a Bayesian approach to classify emails for detecting spam. Vector space representation denotes webpages as vectors in a very high dimensional space. Each webpage is represented as a Boolean or numerical vector; in this paper, it is represented as a numerical vector. In this very high dimensional space, each term is a sequence of alphanumeric characters. A given webpage has in each component a numerical value specifying some function f of often a term corresponding to the dimension appearing in the webpage. Salton and Buckley (1988) used an alternative term called weighting but it is not used in this simulation. In a vector space representation of webpages, n(ti, d) is the number of occur- rences of term ti in webpage d; n(ti, d) is a random variable. The function f is applied to n(ti, d) and produces a value for the ith com- ponent of the vector for webpage d. The identity function f(a) = a is applied to the term counts. Other common functions that are applied to the term frequencies are outlined below. fðaÞ¼ logða þ 1Þ ð9Þ fðaÞ¼ ffiffiffi a p ð10Þ fðaÞ¼ a a þ const ð11Þ Eq. (9) was defined by Robertson and Jones (1976). Eq. (10) was used in the scatter/gather system (Cutting, Karger, Pedersen, & Tukey, 1992) for webpage clustering and was found to outperform Eq. (9). Robertson and Walker (1994) proposed Eq. (11) and found this general form to be useful for webpage retrieval by making use of various instantiations of the constant value. Fig. 4. Visual representation of supervised learning models demo Term frequency – inverse document frequency (TFIDF) weighting is the most used function for the webpage term frequen- cies. The term frequencies (TF) in each webpage and the inverse document (webpage) frequency (IDF) of each term in the entire collection are part of the weighting function. IDF is defined as IDFðtÞ¼ log N nt � � ð12Þ where N is the total number of webpages in the collection and nt is the number of webpages in which term t appears at least once. The TFIDF weight for a term t in a webpage d is the product of the term frequency and the inverse webpages frequency for that term, returning: TFIDFðt; dÞ¼ nðt; dÞ:IDFðtÞ ð13Þ This paper uses a simple Boolean representation of webpages that records whether or not a given term appears in a webpage. Most rule-based methods (Apt´e, Damerau, & Weiss 1994; Cohen, 1996) use an underlying Boolean model and Boolean vector representation has been used in probabilistic classification models. Another way to incorporate word frequency information into probabilistic classification models is by using a parametric distribution, such as bounded Gaussian or Poisson distribution to capture the probability of words appearing number of times in webpages. Yang and Chute (1994) provide further evidence that Boolean representation is adequate by comparing Boolean and frequency-based representations. nstrating clear separations of malicious and safe webpages. Table 1 Results of comparisons of supervised machine learning models that detect malicious webpages (key: KNN: K-Nearest Neighbor, LS: Linear SVM, RS: RBF SVM, NB: Naive Bayes). No. of webpages KNN (%) LS (%) RS (%) NB (%) 50 74 80 79 77 100 75 82 83 78 500 79 86 92 78 5000 91 93 97 84 100,000 95 93 98 89 1172 H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 Word stemming reduces the words in the webpages to their root forms, known as word stems. Porter (1997) described a simu- lation model that utilizes the word-stemming algorithm, which is able to combine similar and dissimilar items into one. But Frakes and Baeza-Yates (1992) compared various stemming methods to unstemmed representations and showed that in many cases the performances of both representations are very close to each other. 3.2.2. URLs URLs identify webpages and are used as unique identifiers. Many malicious webpages have suspiciously looking characters in their URLs and in their contents. Sometimes URLs have spelling mistakes too. The lexical features of URLs were fed into the machine learning models. If there were spelling mistakes or suspi- cious characters in the URL, then they were regarded as suspicious. 3.2.3. Page links Webpages have many links in order to provide further informa- tion. The webpages that link to malicious webpages are likely to be malicious themselves. The computer simulations extracted all the Fig. 5. ROC curves for supervise links from each webpage and they were also fed into the machine learning models. 3.2.4. Visual features All the features that have been mentioned so far are text based such as source codes, stripped HTML, domain names and URLs. The visual features are based on images. First, screenshots of webpages were downloaded by passing the URLs to PhantomJS (a headless webkit browser with JavaScript API). PhantomJS took each URL, saved a screenshot of the webpage and converted them to Portable Network Graphics (PNG) file format. Images were then converted to a format understandable by the proposed models. There are two popular techniques that are generally used such as Speeded Up Robust Features (SURF) and Scale Invariant Feature Transform (SIFT) (Lowe, 2004). The simulation used SURF, as it has less strin- gent licensing options compared to SIFT. The idea behind using the visual features is that malicious webpages look simpler because they are likely to have less input from designers whereas safe web- pages are designed better. An exciting feature was that the unsupervised machine learning models were used in conjunction with supervised machine learn- ing models and each screenshot of the webpages were analyzed using SURF. The values from the SURF were clustered using the unsupervised machine learning models. These clusters were then fed into the supervised machine learning models to further improve the classification. 3.3. Evaluation of the machine learning models Machine learning methods discussed previously were used on different combinations of features. First the webpages were classi- fied from just content features, then other types of features were d machine learning models. Table 2 Results based on datasets provided by Ma et al. (2009a). Classifier Accuracy (%) K-Nearest Neighbor 91 RBF Support Vector Machine 97 Linear Support Vector Machine 92 Naive Bayes 85 H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 1173 added and the gain was reexamined. It was found in this research that the highest accuracy is obtained by combining URL, page-link, semantic TFIDF and SURF features, which adds to the paper’s nov- elty. This combination of features was used as the optimal feature configuration. The accuracy of the unsupervised models had to be done differently because the truth labels are not known and the evaluation had to be performed using the model itself. Finally the machine learning techniques were trained on data sets with varying ratios of malicious and safe webpages. 70% of the labeled webpages were used for training and 30% for testing. The ratio of malicious to safe webpages is the same in testing as well as train- ing for the supervised machine learning models. The supervised classification performances were evaluated in terms of precision and recall, while the silhouette coefficient was used to evaluate the unsupervised techniques. The silhouette coefficient outputs a score relating to a model with better-defined clusters. The silhouette coefficient is defined for each sample and composed of two extremes, bounded between �1 for incorrect clustering and +1 for highly dense clustering. 3.3.1. Supervised techniques Fig. 4 shows a visual representation of webpages as safe or mali- cious, using supervised classification. K-Nearest Neighbor, Radial Fig. 6. Confusion matrices for supervi Basis Function (RBF) SVM, Linear SVM and Naive Bayes clearly sep- arate the safe and malicious webpages. The webpages used in Fig. 4 is somewhat smaller in numbers in order to demonstrate pictorial representations of the outcomes. Table 1 further demonstrates the overall results of the supervised techniques. The accuracy for all the supervised models improved as the number of webpages increased. In general, SVM outperformed the rest. In the case of SVM, the accuracy values are remarkably low even when a small number of webpages are applied. As soon as the number of webpages exceeds 500, the accuracy increases. In the case of other models, a similar trend is observed at the beginning, as the number of webpages increases the accuracy also increases. The results suggest that the models were able to generalize better as more patterns emerged from various sources. Receiver Operating sed machine learning techniques. Fig. 7. Visual representation of unsupervised learning models demonstrating clear separations of malicious and safe webpages. Table 3 Results of comparisons of unsupervised machine learning models that detect malicious webpages. Classifier Silhouette coefficient Mini Batch K-Means 0.877 Affinity Propagation 0.963 K-Means 0.877 1174 H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 Characteristic (ROC) curve and Confusion Matrix are used to analyze the efficiencies and performances of the supervised algo- rithms, as outlined in Figs. 5 and 6. ROC curve is a graph which demonstrates the efficiency and performance of a classifier system by plotting true positive rate against false positive rate at various threshold settings. Fig. 5 illustrates that Linear SVM and RBF SVM perform the best out of the four classifiers with 0.93 and 0.91 respectively. K-Nearest Neighbor performs the worst, because it had less access to training data due to memory constraints. Confusion Matrix is a contingency table enables visualization of the efficiency and performance of a supervised learning algorithm, which makes it easy to understand if the system is confusing or mislabeling two classes. In Fig. 6, the confusion matrices have four sections, True Positive, True Negative, False Positive and False Negative. True Positive denotes that a malicious webpage is cor- rectly identified as malicious. True Negative represents that a safe Fig. 8. The Chrome extension uses 4 steps to dec webpage is correctly labeled as safe. False Positive means that a safe webpage is incorrectly identified as malicious. False Negative indicates that a malicious webpage is incorrectly labeled as safe. The proposed machine learning techniques are designed as such where the priority is given to detect the malicious websites that is the true positives. With this is mind, the outstanding True Posi- tives from highest to lowest are in the following order RBF SVM (97.8%), Linear SVM (92.4%), Naïve Bayes (76.4%) and K-Nearest Neighbor (9.9%). The supervised models were also run against one of the most popular datasets provided by Ma, Saul, Savage, and Voelker (2009a). The data file had based on SVM and therefore the data was converted into recognizable formats to be fed into the machine learning models. Table 2 shows the results of the simula- tions using Ma et al. datasets. All the supervised algorithms scored over 85% and RBF SVM performed the best with 97% accuracy. 3.3.2. Unsupervised techniques Fig. 7 shows that the three unsupervised algorithms do clearly separate the malicious and safe webpages. In Fig. 7, a smaller number of webpages were used in order to demonstrate pictorial representations of the outcomes. The detailed numerical results are outlined in Table 3 using 100,000 webpages. Affinity Propaga- tion performs the best among the unsupervised machine learning ide whether a webpage is malicious or not. H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 1175 algorithms in Table 3 because the silhouette coefficient is closest to 1. Furthermore, Affinity Propagation algorithm identifies three clusters. The red cluster is grouped as outlier, whereas the other two Mini Batch K-Means and K-Means find only two clusters. 3.4. Online learning The conventional batch processing machine learning models cannot learn incrementally. Online learning needs to employ a different approach than the traditional batch processing to accom- modate for the new incoming data. This problem is addressed in this paper by using streaming data in the form of a list of malicious Fig. 9. The Chrome extension shows that t Fig. 10. The Chrome extension shows that the webpages and safe webpages, which are compiled from various sources. This allows the supervised predictive models inside WAC to train automatically as new the data is coming in. This has been found to be very useful with the use of Chrome extension, which is described in Section 3.5. 3.5. Chrome extension The chrome extension has been built using the architecture shown in Fig. 8. The Content Script looks at the loading document and sends the loaded source code to the predictive classifier. The classifier then parses, creates the features and then responds with he website loaded by the user is safe. website loaded by the user is malicious. 1176 H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 whether it thinks that the webpage is safe or malicious. The Back- ground Script receives the response and displays whether it is safe as demonstrated in Fig. 9 or malicious as shown in Fig. 10. The four steps in the Chrome Extension are outlined below: Step 1: Grab the webpage. Step 2: Extract the features and send them to the predictive model. The predictive model then sends a response. Step 3: Content Script notifies the Background Script with the response. Step 4: Background Script displays the response. ‘Heavyweight’ classifiers are more accurate but they have poor prediction time. Heavyweight classifiers use more features and so have a higher accuracy. ‘Lightweight’ classifiers does the opposite, it uses less features and consumes the features from the browser. This paper utilizes Google Chrome which exploits the predictive capabilities of the proposed supervised and unsupervised machine learning models by taking into account the advantages of both the heavyweight and the lightweight classifiers. The Chrome extension obtains all the features from the browser and sends them to the clas- sifier which uses more features, thus have a quick prediction time and yet higher accuracy. The Chrome extension used the supervised models for accurate classification and the unsupervised models were used to ascertain the clusters of the screenshots of webpages. 4. Conclusion This paper presents the evaluation of three supervised machine learning models and two unsupervised machine learning models for text classification, to detect webpages as either malicious or not. All the supervised techniques were trained and applied to a large number of webpages and manually split into two classes. In a nutshell, all of the machine learning techniques show encourag- ing performance with accuracies above 89% for supervised tech- niques and a silhouette coefficient of 0.87 for unsupervised techniques using 100,000 webpages. As the number of webpages increased the accuracy rates for each type of machine learning model were improved. The accuracies of the unsupervised models are not better than the supervised one, but came close. This is interesting because the unsupervised models were not aware of the two classes of malicious and safe. The major contribution of this paper is to explore a range of machine learning algorithms that use a wide range of features including the use of unsupervised algorithms. The computer simu- lation results show that the classifiers can obtain information from the URLs, page links, semantics and visual features of webpages. Different sets of data on ratios of malicious and safe webpages do not significantly affect the error rates of the classifiers. The content features of webpages contributed the most to reduce the error rates, and then the other significant ones were the URLs followed by the SURF visual features. The visual features take more computational resources and they are time consuming to compute. However, the SURF features can capture the visual information faster than the SIFT. Another significant contribution of this paper is the use of Chrome extension in the browser. This allowed to detect whether a webpage is malicious or not in a very short time. The lightweight classifiers are very fast but are slightly less accu- rate, whereas heavyweight classifiers are more accurate but take more time. This extension enabled the SVM classifier to be ‘middle- weight’ with very high accuracy and yet less prediction time. The final contribution is the use of online learning to detect malicious webpages, which allows for the training to take place without going through the ‘old training data’. This is very significant for real world applications. There is one limitation in this paper, that is malicious webpages will not be detected if the harmful elements are outside of the features that have been used in the proposed algorithms. Despite this limitation, the paper’s contribution is encouraging as the accuracy rates are improved through the use of visual features. There are possibilities that this research work could be taken further, some of which are described below. (i) Changing the implementation of the whole system to JavaScript which will combine the separate processes into one. This approach will allow the application to run on the browser as a standalone extension without external dependencies. This method requires that the machine learning models to be rewritten. The question to be answered is whether the prediction time will still remain the same. (ii) Using deep learning to extract features automatically. With the current scheme all the features have to be specified in the imple- mentation and it is difficult to incorporate new features because the system has to be pre-processed and it is time consuming. Deep learning may be able to extract features automatically and probably will adapt to the future threats that could use new fea- tures which have not been discovered as yet. The question remains the same whether deep learning will be as effective as the current successful systems. (iii) Using multiple processing units of a Graphical Processing Unit. Using distributed processing in general will improve the time to build the machine learning models and also will improve the prediction time. However, the main challenge will be to rewrite the algorithms so that they will make an efficient use of hardware especially on mobile devices. Acknowledgement The authors would like to thank Technology Strategy Board - KTP, UK for their generous support for part of the research [Grant No. ktp006367]. References Abbasi, A., Zahedi, F. M., & Kaza, S. (2012). Detecting fake medical web sites using recursive trust labeling. ACM Transactions of Information Systems, 30(4), 22:1–22:36. URL http://dx.doi.org/10.1145/2382438.2382441. Abbasi, A., Zhang, Z., Zimbra, D., Chen, H., & Nunamaker, J. F. (2010). Detecting fake websites: The contribution of statistical learning theory. MIS Quarterly, 34(3), 435–461. URL <http://dl.acm.org/citation.cfm?id=2017470.2017473>. Alexa (2013). Alexa. URL <http://s3.amazonaws.com/alexa-static/top-1m.csv.zip>. Altman, N. (1992). An introduction to kernel and nearest-neighbor nonparametric regression. The American Statistician, 46(3), 175–185. Apt´e, C., Damerau, F., & Weiss, S. M. (1994). Automated learning of decision rules for text categorization. ACM Transactions of Information Systems, 12(3), 233–251. URL http://dx.doi.org/10.1145/183422.183423. Bannur, S. N., Saul, L. K., & Savage, S. (2011). Judging a site by its content: learning the textual, structural, and visual features of malicious web pages. In Proceedings of the 4th ACM workshop on security and artificial intelligence. AISec ’11 (pp. 1–10). New York, NY, USA: ACM. URL http://dx.doi.org/10.1145/ 2046684.2046686. Bayes, M., & Price, M. (1763). An essay towards solving a problem in the doctrine of chances. By the late rev. Mr. Bayes, F. R. S. communicated by Mr. Price, in a letter to john canton, A. M. F. R. S.. Philosophical Transactions, 53, 370–418. URL <http://rstl.royalsocietypublishing.org/content/53/370.short>. Braganza, R. (2006). Cross-site scripting: Cross-site scripting an alternative view. Network Security, 2006(9), 17–20. URL http://dx.doi.org/10.1016/S1353- 4858(06)70425-1. Chang, C. -C., & Lin, C. -J. (2012). Libsvm: A library for support vector machines. URL <http://www.csie.ntu.edu.tw/cjlin/libsvm/>. Cohen, W. W. (1996). Learning trees and rules with set-valued features. Proceedings of the thirteenth national conference on artificial intelligence. AAAI’96 (Vol. 1, pp. 709–716). Springer. Cohen, W., & Singer, Y. (1999). Context-sensitive learning methods for text categorization. ACM Transactions on Information Systems (TOIS), 17(2), 141–173. Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 273–297. Cutting, D. R., Karger, D. R., Pedersen, J. O., & Tukey, J. W. (1992). Scatter/gather: A cluster-based approach to browsing large document collections. In Proceedings of the 15th annual international ACM SIGIR conference on research and development in information retrieval. SIGIR ’92 (pp. 318–329). New York, NY, USA: ACM. URL http://dx.doi.org/10.1145/133160.133214. Dueck, D. (2009). Affinity propagation: Clustering data by passing messages. http://dx.doi.org/10.1145/2382438.2382441 http://dl.acm.org/citation.cfm?id=2017470.2017473 http://s3.amazonaws.com/alexa-static/top-1m.csv.zip http://refhub.elsevier.com/S0957-4174(14)00528-4/h0020 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0020 http://dx.doi.org/10.1145/183422.183423 http://dx.doi.org/10.1145/2046684.2046686 http://dx.doi.org/10.1145/2046684.2046686 http://www.rstl.royalsocietypublishing.org/content/53/370.short http://dx.doi.org/10.1016/S1353-4858(06)70425-1 http://dx.doi.org/10.1016/S1353-4858(06)70425-1 http://www.csie.ntu.edu.tw/cjlin/libsvm/ http://refhub.elsevier.com/S0957-4174(14)00528-4/h0050 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0050 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0050 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0055 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0055 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0060 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0060 http://dx.doi.org/10.1145/133160.133214 H.B. Kazemian, S. Ahmed / Expert Systems with Applications 42 (2015) 1166–1177 1177 Dumais, S., Platt, J., Heckerman, D., & Sahami, M. (1998). Inductive learning algorithms and representations for text categorization. In Proceedings of the seventh international conference on information and knowledge management. CIKM ’98 (pp. 148–155). New York, NY, USA: ACM. URL http://dx.doi.org/ 10.1145/288627.288651. Facebook (2010). Facebook suffers from rash of clickjacking. Network Security 2010 (6), 20. URL <http://www.sciencedirect.com/science/article/pii/ S1353485810700866>. Frakes, W. B., & Baeza-Yates, R. (Eds.). (1992). Information retrieval: Data structures and algorithms. Upper Saddle River, NJ, USA: Prentice-Hall Inc. Fu, A. Y., Wenyin, L., & Deng, X. (2006). Detecting phishing web pages with visual similarity assessment based on earth mover’s distance (EMD). IEEE Transactions of Dependable and Secure Computing, 3(4), 301–311. URL http://dx.doi.org/ 10.1109/TDSC.2006.50. Garera, S., Provos, N., Chew, M., & Rubin, A. D. (2007). A framework for detection and measurement of phishing attacks. In Proceedings of the 2007 ACM workshop on recurring malcode. WORM ’07 (pp. 1–8). New York, NY, USA: ACM. URL http:// dx.doi.org/10.1145/1314389.1314391. Guan, D., Chen, C., & Lin, J. (2009). Anomaly based malicious URL detection in instant messaging. In Proceedings of the joint workshop on information security (JWIS). Hansen, R., & Grossman, J. (2008). Clickjacking, URL <http://www.sectheory.com/ clickjacking.htm>. Harley, D., & Bureau, P. -M. (2008). Drive-by downloads from the trenches. In 3rd International Conference on Malicious and Unwanted Software, MALWARE 2008 (pp. 98–103). Jordan, A. (2002). On discriminative vs. generative classifiers: A comparison of logistic regression and Naive Bayes. Advances in Neural Information Processing Systems, 14, 841. Kan, M., & Thi, H. (2005). Fast webpage classification using URL features. In Proceedings of the 14th ACM international conference on information and knowledge management (pp. 325–326). ACM. Le, A., Markopoulou, A., & Faloutsos, M. (2010). Phishdef: URL names say it all. CoRR abs/1009.2275. Liu, W., Deng, X., Huang, G., & Fu, A. Y. (2006). An antiphishing strategy based on visual similarity assessment. IEEE Internet Computing, 10(2), 58–65. URL http:// dx.doi.org/10.1109/MIC.2006.23. Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2), 91–110. Lucca, G. A. D., Fasolino, A. R., Mastoianni, M., & Tramontana, P. (2004). Identifying cross site scripting vulnerabilities in web applications. In Proceedings of the web site evolution, sixth IEEE international workshop. WSE ’04 (pp. 71–80). Washington, DC, USA: IEEE Computer Society. URL <http://dl.acm.org/ citation.cfm?id=1025133.1026460>. Ma, J., Saul, L. K., Savage, S., & Voelker, G. M. (2009a). Beyond blacklists: Learning to detect malicious web sites from suspicious URLs. In Proceedings of the 15th ACM SIGKDD international conference on knowledge discovery and data mining. KDD ’09 (pp. 1245–1254). New York, NY, USA: ACM. URL http://dx.doi.org/10.1145/ 1557019.1557153. Ma, J., Saul, L. K., Savage, S., & Voelker, G. M. (2009b). Identifying suspicious URLs: An application of large-scale online learning. In Proceedings of the 26th annual international conference on machine learning. ICML ’09 (pp. 681–688). New York, NY, USA: ACM. URL http://dx.doi.org/10.1145/1553374.1553462. MacQueen, J. B. (1967). Some methods for classification and analysis of multivariate observations. In L. M. L. Cam & J. Neyman (Eds.). Proceedings of the fifth Berkeley symposium on mathematical statistics and probability (Vol. 1, pp. 281–297). University of California Press. Malware (2011). A great year for malware. Computer fraud and security 2011 (1), 20. URL <http://www.sciencedirect.com/science/article/pii/ S1361372311700082>. Mcgrath, D. K., & Gupta, M. (2008). Behind phishing: An examination of phisher modi operandi. In Proceedings of the USENIX workshop on large-scale exploits and emergent threats. Okanovic, V., & Mateljan, T. (2011). Designing a new web application framework. In MIPRO, 2011 proceedings of the 34th international convention (pp. 1315–1318). Phishtank (2013). Phistank. URL <http://www.phishtank.com>. Porter, M. F. (1997). Readings in information retrieval. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, Ch. An algorithm for suffix stripping (pp. 313–316). URL <http://dl.acm.org/citation.cfm?id=275537.275705>. Provos, N., Mavrommatis, P., Rajab, M. A., & Monrose, F. (2008). All your iFrames point to us. In Proceedings of the 17th conference on security symposium. SS’08 (pp. 1–15). Berkeley, CA, USA: USENIX Association. URL <http:// www.dl.acm.org/citation.cfm?id=1496711.1496712>. Robertson, S., & Jones, K. (1976). Relevance weighting of search terms. Journal of the American Society for Information Science, 27(3), 129–146. Robertson, S. E., & Walker, S. (1994). Some simple effective approximations to the 2- poisson model for probabilistic weighted retrieval. In Proceedings of the 17th annual international ACM SIGIR conference on research and development in information retrieval. SIGIR ’94 (pp. 232–241). New York, NY, USA: Springer- Verlag New York Inc.. URL <http://dl.acm.org/citation.cfm?id=188490.188561>. Sahami, M., Dumais, S., Heckerman, D., & Horvitz, E. (1998). A bayesian approach to filtering junk e-mail. In Learning for text categorization: Papers from the 1998 workshop (Vol. 62). Salton, G., & Buckley, C. (1988). Term-weighting approaches in automatic text retrieval. Information Processing and Management, 24(5), 513–523. URL http:// dx.doi.org/10.1016/0306-4573(88)90021-0. Salton, G., Wong, A., & Yang, C. S. (1975). A vector space model for automatic indexing. Communication ACM, 18(11), 613–620. URL http://dx.doi.org/10.1145/ 361219.361220. Sood, A. K., & Enbody, R. J. (2011). Frame trapping the frame busting defence. Network Security, 2011(10), 8–12. URL <http://www.sciencedirect.com/science/ article/pii/S1353485811701052>. Spertus, E. (1997). Smokey: Automatic recognition of hostile messages. In Proceedings of the fourteenth national conference on artificial intelligence and ninth conference on innovative applications of artificial intelligence. AAAI’97/ IAAI’97 (pp. 1058–1065). AAAI Press. URL <http://dl.acm.org/ citation.cfm?id=1867406.1867616>. Yang, Y., & Chute, C. G. (1994). An example-based mapping method for text categorization and retrieval. ACM Transactions of Information Systems, 12(3), 252–277. URL http://dx.doi.org/10.1145/183422.183424. http://dx.doi.org/10.1145/288627.288651 http://dx.doi.org/10.1145/288627.288651 http://www.sciencedirect.com/science/article/pii/S1353485810700866 http://www.sciencedirect.com/science/article/pii/S1353485810700866 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0085 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0085 http://dx.doi.org/10.1109/TDSC.2006.50 http://dx.doi.org/10.1109/TDSC.2006.50 http://dx.doi.org/10.1145/1314389.1314391 http://dx.doi.org/10.1145/1314389.1314391 http://www.sectheory.com/clickjacking.htm http://www.sectheory.com/clickjacking.htm http://refhub.elsevier.com/S0957-4174(14)00528-4/h0115 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0115 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0115 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0120 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0120 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0120 http://dx.doi.org/10.1109/MIC.2006.23 http://dx.doi.org/10.1109/MIC.2006.23 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0135 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0135 http://dl.acm.org/citation.cfm?id=1025133.1026460 http://dl.acm.org/citation.cfm?id=1025133.1026460 http://dx.doi.org/10.1145/1557019.1557153 http://dx.doi.org/10.1145/1557019.1557153 http://dx.doi.org/10.1145/1553374.1553462 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0155 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0155 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0155 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0155 http://www.sciencedirect.com/science/article/pii/S1361372311700082 http://www.sciencedirect.com/science/article/pii/S1361372311700082 http://www.phishtank.com http://dl.acm.org/citation.cfm?id=275537.275705 http://www.dl.acm.org/citation.cfm?id=1496711.1496712 http://www.dl.acm.org/citation.cfm?id=1496711.1496712 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0190 http://refhub.elsevier.com/S0957-4174(14)00528-4/h0190 http://dl.acm.org/citation.cfm?id=188490.188561 http://dx.doi.org/10.1016/0306-4573(88)90021-0 http://dx.doi.org/10.1016/0306-4573(88)90021-0 http://dx.doi.org/10.1145/361219.361220 http://dx.doi.org/10.1145/361219.361220 http://www.sciencedirect.com/science/article/pii/S1353485811701052 http://www.sciencedirect.com/science/article/pii/S1353485811701052 http://www.dl.acm.org/citation.cfm?id=1867406.1867616 http://www.dl.acm.org/citation.cfm?id=1867406.1867616 http://dx.doi.org/10.1145/183422.183424 Comparisons of machine learning techniques for detecting malicious webpages 1 Introduction 2 Machine learning models 2.1 Naive Bayes Classifier 2.2 K-Nearest Neighbor 2.3 Support Vector Machines 2.4 K-Means 2.5 Affinity Propagation 3 Results 3.1 Data sources 3.2 Features 3.2.1 Semantic 3.2.2 URLs 3.2.3 Page links 3.2.4 Visual features 3.3 Evaluation of the machine learning models 3.3.1 Supervised techniques 3.3.2 Unsupervised techniques 3.4 Online learning 3.5 Chrome extension 4 Conclusion Acknowledgement References 
work_stictr4ivbbi5b6acnnworse3y	----	Aalborg Universitet When To Test? Troubleshooting with Postponed System Test Ottosen, Thorsten Jørgen; Jensen, Finn V. Publication date: 2010 Document Version Early version, also known as pre-print Link to publication from Aalborg University Citation for published version (APA): Ottosen, T. J., & Jensen, F. V. (2010). When To Test? Troubleshooting with Postponed System Test. Department of Computer Science, Aalborg University. General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. ? Users may download and print one copy of any publication from the public portal for the purpose of private study or research. ? You may not further distribute the material or use it for any profit-making activity or commercial gain ? You may freely distribute the URL identifying the publication in the public portal ? Take down policy If you believe that this document breaches copyright please contact us at vbn@aub.aau.dk providing details, and we will remove access to the work immediately and investigate your claim. Downloaded from vbn.aau.dk on: April 06, 2021 https://vbn.aau.dk/en/publications/f8ecc587-23fd-4361-a0d4-60934b9f477f When To Test? Troubleshooting with Postponed System Test. Thorsten J. Ottosen and Finn V. Jensen Technical Report 10-001 Department of Computer Science Aalborg University, Denmark October 19, 2010 Abstract. Troubleshooting is about computing a cost-efficient way to repair a faulty device. In this paper we investigate the most simple action-based trou- bleshooting scenario with a single extension: the overall system test is not required to be performed after each action, but it may be postponed until several actions have been performed. As shown by Kadane and Simon, the simple troubleshooting scenario is easily solvable in polynomial time. However, we conjecture that the new troubleshooting scenario is NP-hard and therefore describe an Θ(n3) time heuristic. The new heuristic guar- antees optimality for a class of models, and for other classes of models we benchmark it against the optimal solution and other simpler greedy heuristics. The benchmark is performed on artificial models as well as a real-world troubleshooting model, and they suggest that the heuristics are close to optimal. Introduction Imagine that you have a device which has been running well up to now, but suddenly it is malfunctioning. A set of faults F describes the possible causes to the problem. To fix the problem you have a set A of actions, which may fix the problem by repairing one or more faults. An action, α, has a positive cost Cα and it has two possible outcomes, namely "α = yes" (the problem was fixed) and "α = no" (the action failed to fix the problem). Informally, our task is to determine the best sequence of actions until the problem is fixed or until no more actions remains to be carried out. For an overview of state of the art of decision theoretic troubleshoot- ing we refer the reader to (Jensen et al., 2001), (Langseth and Jensen, 2001), (Skaanning and Vomlel, 2001), (Langseth and Jensen, 2003), (Koca and Bil- giç, 2004a), (Koca and Bilgiç, 2004b), and (Warnquist et al., 2008). Basically all non-trivial troubleshooting domains are NP-hard (Vomlelová, 2003). The scenario investigated in this article is similar to troubleshooting with con- ditional costs, but the proof for NP-hardness does not carry through in this simple scenario. Nevertheless we conjucture that our scenario is NP-hard. In traditional troubleshooting it has been assumed that each action is followed by a test that determines whether the entire system D is working. Furthermore, it has generally been assumed that the cost of this test, CD, is so low that it should not even be modelled. We show that this assumption is incorrect in general and develop terminology and algorithms for this new scenario. In this paper we shall examine troubleshooting problems where the fol- lowing model parameters are given: 1. An overall problem D that we need to solve (e.g. "my printer does not work"). 2. A set of faults F = {f1, . . . , fm} describing the possible causes to the problem (e.g. "toner cartridge empty"). 3. For each fault f ∈F, a probability P(f) describing how likely it is that f is present when troubleshooting begins. 4. A set of actions A = {α1, . . . ,αn} that can potentially solve D (e.g. "change the toner cartridge"). 5. For each action α ∈A, a cost Cα describing the resources required to perform the action. 6. For each action α ∈ A, a success probability P(α = yes|f) describ- ing the likelihood of solving D by performing the action when f is present. 2 7. A cost of testing if the problem D remains, CD > 0. Furthermore, we have the following simplifying assumptions about the model: Assumption 1. The single fault assumption: Initially the problem D is known to exist and it is due to the presence of a single fault from F. Assumption 2. The action result assumption: repeating a failed action will not fix the problem. Assumption 3. The carefulness assumption: by performing an action or testing the system, we never introduce new faults. Assumption 4. The independent actions assumption: Each action repairs a dis- tinct set of faults that may be causing the problem D. We uphold these assumptions throughout this paper. Remark. Assumption 1 implies∑ f∈F P(f) = 1 and by Assumption 4 the repair probability of an action can be computed as P(α = yes) = ∑ f∈F P(α = yes|f) · P(f) . However, since actions may be imperfect (P(α = yes|f) < 1), Assumption 4 does not imply ∑ α∈A P(α = yes) = 1 . Furthermore, the above equation can also be violated by the fact that not all faults may be addressed by an action. Before we can define our optimization problem formally, we need some notation and terminology. We write εi to denote that the first i actions have failed (ε0 ⊆ εi), and we have by assumption P ( ε0 ) = 1 because the device is faulty. We call P(α = yes) for the repair probability of α, and we shall abbreviate it with Pα. To make it clear that the cost of an action includes the cost of testing the system CD, we write CA+D for CA + CD and so CA never includes the cost of testing the system. For each V ⊆A we can form a compound action A with the following properties PA = ∑ α∈V Pα, CA = ∑ α∈V Cα, |A| = |V | . 3 To make the distinction clear, we call actions α ∈ A for atomic actions. We can see that a partition of A into k subsets {V1, . . . , Vk} induces a new model with independent actions {A1, . . . , Ak}. We shall also say that {A1, . . . , Ak} is a partition of A whenever the corresponding sets of atomic actions form a partition of A. From now on we shall not distinguish be- tween the set of actions and the compound action; that is, we use A to mean a set of atomic actions and write A = {α1, . . . ,α|A|}. A troubleshooting se- quence is a sequence of compound actions s = 〈A1, . . . , A`〉 prescribing the process of repeatedly performing the next action until the problem is fixed or the last action has been performed and such that {A1,. . . ,A`} is a partition of A. When each compound action contains only one action, we say that the sequence is atomic. A subsequence of s, s[k,m] = 〈Ak+1, . . . , Am〉 with k ≤ m, is called a partial troubleshooting sequence if k ≥ 1 or m < `. If k = 0, we may simply write s[m]. Definition 1. Let s = 〈A1, . . . , A`〉 be a troubleshooting sequence. Then the expected cost of repair (ECR) of s is the mean of the cost until an action succeeds or all actions have been performed: ECR (s) = ∑̀ i=1 CAi+D · P ( εi−1 ) . Formally, the problem is then to find a troubleshooting sequence with min- imal ECR (·) over all possible sequences. We call such a sequence for opti- mal. An important definition is that of efficiency: Definition 2. The efficiency of an action A is ef(A) = PA CA+D . Its importance stems from the following results. Theorem 1 (Kadane and Simon (1977)). Let s = 〈α1, . . . ,αn〉 be an atomic troubleshooting sequences, and let CD = 0. Then s is optimal if and only if ef(αi) ≥ ef(αi+1) for i ∈{1, . . . ,n− 1} . Theorem 2 (Jensen et al. (2001)). Let s = 〈α1, . . . ,αn〉 be an atomic trou- bleshooting sequence. Then a necessary condition for s to be optimal, even when CD > 0, is that ef(αi) ≥ ef(αi+1) for i ∈{1, . . . ,n− 1} . Corollary 1. Theorem 2 holds for arbitrary troubleshooting sequences. 4 Proof. Because a troubleshooting sequence partitions A, the compound ac- tions can be seen as virtual actions which are also independent, thus satis- fying Assumption 3. All other assumptions cannot be invalidated. We may also compute the ECR of any atomic troubleshooting sequence: Theorem 3 (Jensen et al. (2001)). Let s = 〈α1, . . . ,αn〉 be an atomic trou- bleshooting sequence. Then the ECR of s may be computed as ECR (s) = n∑ i=1 Cαi+D ·  1 − i−1∑ j=1 Pαj   , where 1 − ∑i−1 j=1 Pαj = P ( εi−1 ) . Corollary 2. Theorem 3 holds for arbitrary troubleshooting sequences. Proof. Same argument as for Corollary 1. Thus, due to Assumption 1-4, we may completely ignore the F, P(f), and P(α = yes|f) once the repair probabilities have been computed. Therefore we only use Pα in the rest of this paper. The following example shows that Theorem 1 does not extend to arbi- trary troubleshooting sequences when CD > 0. Example 1. Consider a model with four atomic actions α1, α2, α3 and α4. The other properties of the model are as follows: Pα Cα Cα+D ef(α) α1 0.24 1 2 0.12 α2 0.42 3 4 0.105 α3 0.2 1 2 0.1 α4 0.14 19 20 0.007 So CD = 1. We have ECR (〈α1,α2,α3,α4〉) = 2 + 0.76 · 4 + 0.34 · 2 + 0.14 · 20 = 8.52 ECR (〈{α1,α2},α3,α4〉) = 5 + 0.34 · 2 + 0.14 · 20 = 8.48 thus proving that it is more efficient to form the compound action A = {α1,α2}. To be able to test heuristic algorithms it is necessary with an algorithm that is guaranteed to find an optimal value. Exhaustive searches are al- ways able to do this, albeit the algorithm might take exponential or super- exponential time. Algorithm 1 shows an exhaustive search with pruning (line 11, applying Theorem 2) and memoization (line 5-8). 5 Algorithm 1 Exhaustive Search With Compound Actions 1: function GENERATEALLSEQUENCES(A, &S) 2: Let all = ∅ 3: for i = 1 to |A| do 4: for all (|A| i ) ways to construct the first compound action A do 5: if A 6∈ S then 6: Compute A from scratch 7: Set S = S∪{A} 8: end if 9: Let after = GENERATEALLSEQUENCES(A\ A,S) 10: for s = 〈A1, . . . , A`〉 ∈ after do 11: if ef(A1) ≤ ef(A) then 12: Let s′ = 〈A, A1, . . . , A`〉 13: Set all = all∪{s′} 14: end if 15: end for 16: end for 17: end for 18: return all 19: end function 20: function OPTIMALCOMPOUNDACTIONSEQUENCE(A) 21: Let S = ∅ 22: Let all = GENERATEALLSEQUENCES(A,S) 23: return arg min s∈all ECR (s) 24: end function 6 Let us briefly discuss the complexity of this algorithm. The algorithm re- cursively generates all possible sequences of compound actions. The num- ber of such sequences is n! ·2n−1 because for each of the n! permutations of the n atomic actions we can form 2n−1 different partitions. Thus the run- ning time and memory requirement of the algorithm is O(n! · 2n). Note, however, that this is not a tight upper bound since a troubleshooting se- quence s = 〈A1, . . . , A`〉 can be constructed by ∏` i=1 |Ai|! different order- ings. Later we show that the complexity of an exhaustive search can be reduced to Θ(n3 · n!). Two Simple Greedy Heuristics In this section we shall develop two simple greedy algorithms. Further- more, we also give examples showing that none of the algorithms are guar- anteed to find an optimal troubleshooting sequence. For the first algorithm we consider how the ECR of a troubleshooting sequence can be improved by merging two neighbouring actions into one compound action. We have the following result. Theorem 4. Let the two troubleshooting sequences s and s′ be given as s = 〈. . . , Ax, Ax+1, . . .〉 s′ = 〈. . . , A, . . .〉 with A = {Ax, Ax+1} and everything else being the same. Then s′ has a strictly lower ECR than s if and only if CD > CAx+1 · Pαx 1 − ∑x j=1 Pαj . Proof. When the sequence with the new compound action is better, we must have ECR ( s′ ) − ECR (s) ≤ 0 . We observe that most of the terms cancel each other out and we are left 7 with CA+D · P ( εx−1 ) − [ CAx+D · P ( εx−1 ) + CAx+1+D · P(ε x) ] < 0 m CA+D < CAx + CD + (CAx+1 + CD) · P(εx) P(εx−1) m CAx+1 · ( 1 − P(εx) P(εx−1) ) < CD · P(εx) P(εx−1) m CD > CAx+1 · ( P ( εx−1 ) P(εx) − 1 ) m CD > CAx+1 · ( Pαx 1 − ∑x j=1 Pαj ) Theorem 4 immediately suggests a procedure where we greedily merge ad- jacent actions until the ECR is no longer improved. We start with an atomic troubleshooting sequence and form compound actions by repeated appli- cation of Theorem 4. Justified by Theorem 1, there are two obvious choices of the initial sorting: with respect to descending efficiency or descending P C -value. We have summarized this procedure in Algorithm 2. (The special function 1st arg(·) returns the first index for which the boolean argument returns true.) As specified, the algorithm runs in Θ(n2) time. However, if the sum in line 5 in CompoundAction(·) are precomputed , the total cost of all calls to CompoundAction(·) can be made linear. The running time of the algorithm is then dominated by the initial sorting of the actions leading to an Θ(n · lg n) worst case running time. Is Algorithm 2 guaranteed to find an optimal troubleshooting sequence? The example below show it is not. Example 2. We consider the following model: Pα Cα Cα+D ef(α) Pα Cα α1 0.61 1 11 0.055 0.61 α2 0.21 5 15 0.014 0.042 α3 0.18 3 13 0.013 0.06 So CD = 10. Using the ef(·) order Algorithm 2 first computes 10 < 5 · 0.61 1 − 0.61 ≈ 7.8 . 8 Algorithm 2 Computing sequence with compound actions greedily 1: function COMPOUNDACTION(〈α1, . . . ,αn〉 ,x) 2: if x = n then 3: return {αn} 4: end if 5: Let m = n 1st arg i=x ( CD ≤ Cαi+1 · Pαi 1 − ∑i j=1 Pαj ) 6: return {αx, . . . ,αm} 7: end function 8: function GREEDYCOMPOUNDACTIONSEQUENCE(A) 9: Let s′ = 〈α1, . . . ,αn〉 such that ef(αi) ≥ ef(αi+1) or Pαi Cαi ≥ Pαi+1 Cαi+1 . 10: Let s = 〈〉 11: Let i = 1 12: while i ≤ n do 13: Let A = COMPOUNDACTION(s′, i) 14: Set s = s + A 15: Set i = i + |A| 16: end while 17: return s 18: end function Since this is false, it continues with the second test 10 < 3 · 0.21 1 − 0.61 − 0.21 = 3.5 which means that the algorithm suggest the sequence consisting of a single compound action {α1,α2,α3}. If we repeat the calculations using the PC - order, we get the same result. The ECR is easily seen to be 19, but ECR (〈α1,{α2,α3}〉) = 11 + 0.39 · 18 = 18.02 so the algorithm is not optimal. Let us now consider another greedy heuristic. The algorithm is based on the idea that we should repeatedly form the most efficient compound action until all atomic actions have been assigned a compound action. To do this, we need a simple way of computing the most efficient compound action from a set of atomic actions. We first need a lemma. Lemma 1. Let a,b,c, and d be strictly positive numbers. Then a + b c + d ≥ a c if and only if b d ≥ a c 9 Proof. a + b c + d ≥ a c ⇔ a · c + b · c ≥ a · c + a ·d ⇔ b d ≥ a c Theorem 5. Let V ⊆ A be given. Let A be a compound action consisting of ac- tions in V with ef(A) maximal and with |A| minimal over those actions with max- imal efficiency. Then A can be found by the following procedure: define ef(A) = 0 when A = ∅; then repeatedly add an action α ∈ V to A with the highest Pα Cα value until ef(A) does not increase. Proof. Let A consist of actions in V such that ef(A) is maximized and |A| is minimal. Then ef(A) equals∑ α∈A Pα CD + ∑ α∈A Cα = ∑ α∈A\{β} Pα + Pβ CD + ∑ α∈A\{β} Cα + Cβ = SP + Pβ SC + Cβ = P C where β is chosen arbitrarily. Let furthermore α ∈ V \ A. We shall prove Pβ Cβ ≥ P C ≥ Pα Cα which implies the theorem. We first prove Pβ Cβ ≥ P C (i). Because ef(A) is maximal and A has minimal size, Lemma 1 gives SP + Pβ SC + Cβ > SP SC ⇔ Pβ Cβ > SP SC (ii). Then assume Equation (i) is false; then by Lemma 1 SP + Pβ SC + Cβ ≥ Pβ Cβ ⇔ SP SC ≥ Pβ Cβ which is a contradiction with Equation (ii). We then consider the second inequality. Since ef(A) is maximized, Lemma 1 implies P C ≥ P + Pα C + Cα ⇔ P C ≥ Pα Cα which completes the proof. Remark. In (Langseth and Jensen, 2001) a similar heuristic procedure is reached albeit for a slightly different problem. 10 Algorithm 3 Greedy algorithm with max-efficient actions 1: function GREEDYCOMPOUNDACTIONSEQUENCE2(A) 2: Let s′ = 〈α1, . . . ,αn〉 such that Pαi Cαi ≥ Pαi+1 Cαi+1 3: Let s = 〈〉 4: Repeatedly add the most efficient compound action to s 5: return s 6: end function Algorithm 3 shows a greedy algorithm based on Theorem 5. Like Al- gorithm 2 we can precompute the sums used in the efficiency and get a running time of Θ(n · lg n). The following example shows that Algorithm 3 is not guaranteed to find an optimal troubleshooting sequence. Example 3. Consider the following model Pα Cα Cα+D Pα Cα α1 0.15 1 2 0.150 α2 0.35 2 3 0.175 α3 0.50 3 4 0.167 where the ordering considered is 〈α2,α3,α1〉. Algorithm 3 computes the following values to determine the first compound action: ef({α2}) = 0.35 3 = 0.117 ef({α2,α3}) = 0.35 + 0.5 6 = 0.1416 ef({α2,α3,α1}) = 1 7 = 0.1429 Thus the suggested sequence is one single compound action with ECR 7. However, ECR (〈{α2,α3},α1〉) = 6 + 0.15 · 2 = 6.3 which shows the algorithm is not optimal. Having established the heuristic nature of both greedy algorithms, we would like to know if they at least find a local optimum. We use the following def- inition to define a local optimum. Definition 3. Let s = 〈α1, . . . ,αn〉 be a sequence of atomic actions, and let sc = 〈B1, . . . , Bm〉 be a compound action sequence. Then sc is said to be consistent with s if ∀ k < l αi ∈ Bk and αj ∈ Bl =⇒ i < j. 11 In other words, sc can be seen as an ordered partition of s. If an atomic action sequence s is given, we call the set of all compound action sequences consistent with s for C(s). A local optimum is then a com- pound action sequence sc ∈ C(s) with minimal ECR. We then say the (or- dered) partition induced by sc is optimal. From Example 2 and 3 it follows that neither Algorithm 2 nor Algorithm 3 guarantees an optimal partition. An Algorithm For Optimal Partitioning We have now seen how the greedy algorithms fail to provide optimal se- quences even under the restriction that the solution must be consistent with a single seed-sequence s of atomic actions. However, since there are 2n−1 ways to partition the actions of s, a brute force approach is quite impracti- cal even for models of moderate size. By exploiting Theorem 6 below we can construct an O(n3) dynamic programming algorithm. Lemma 2. Let k ≥ 0 and let sc = 〈A1, . . . , A`〉 be a partial troubleshooting sequence that is performed after the first k actions have been performed. Then the contribution of sc to the total ECR is given by ECRk(s c) = ∑̀ i=1 P ( εk+i−1 ) · CAi+D where P ( εk+i−1 ) = 1 − k∑ j=1 Pαj − i−1∑ j=1 PAj . Proof. Follows directly from Corollary 2. The Lemma ensures that the following definition is sound. Definition 4. Let s = 〈α1, . . . ,αn〉 be an atomic troubleshooting sequence. For any partial sequence s[k,m] with k ≤ m the ECR of an optimal com- pound action sequence consistent with s[k,m] is defined by ECR∗(s[k,m]) = { min sc∈C(s[k,m]) ECRk(s c) if k < m 0 if k = m . Theorem 6. Let s = 〈α1, . . . ,αn〉 be an atomic troubleshooting sequence. Then for k < m, ECR∗(s[k,m]) can be calculate as ECR∗(s[k,m]) = m min i=k+1 ( ECRk(〈{αk+1, . . . ,αi}〉) + ECR∗(s[i,m]) ) 12 Proof. We use induction in m−k. Basis step: m−k = 1. Since there is only one compound sequence consistent with 〈αm〉, we get ECR∗(s[k,m]) = min(ECRk(〈{αm}〉) + 0) = ECRk(〈{αm}〉) so the theorem holds in this case. Induction step: assume the theorem holds for m−k ≤ x and let m−k = x + 1. Then let s∗ = arg min sc∈C(s[k,m]) ECRk(s c) for which we must prove ECR∗(s∗) = m min i=k+1 ( ECRk(〈{αk+1, . . . ,αi}〉) + ECR∗(s[i,m]) ) . By Lemma 2 it is correct to split the ECR into two terms corresponding to an ordered partition of s∗. Furthermore, by the induction hypothesis we can compute ECR∗(s[i,m]) for any i since m−i ≤ x. Then consider that s∗ must start with a compound action consisting of y ∈ {1, 2, . . . ,x + 1} actions. These choices of the first action correspond exactly to the expressions that we minimize over, and the result follows. Theorem 6 shows that the problem of finding an optimal partition of a seed sequence s requires optimal partitions of smaller partial sequences (the optimal-substructure property) and that these smaller partial sequences are needed by the computation of several bigger sequences (the overlap- ping subproblems property). These two properties enables a dynamic pro- gramming algorithm, that is, an algorithm that only computes solutions to subproblems once by storing the results in a table. Algorithm 4 uses this approach to find the optimal partition of a seed-sequence s. The algorithm runs in Θ(n3) time and uses Θ(n2) space. The correctness of the algorithm follows from the above theorem. Algorithm 4 now also means that we can make a more efficient exhaustive search for the best troubleshooting se- quence. The algorithm runs in Θ(n3 · n!) time and is shown in Algorithm 6. Unfortunately, Algorithm 5 is not guaranteed to find an optimal trou- bleshooting sequence, as the following example shows. Example 4. Consider the model from Example 3, but take CD = 2: Pα Cα Cα+D ef(α) Pα Cα α1 0.15 1 3 0.05 0.150 α2 0.35 2 4 0.0875 0.175 α3 0.50 3 5 0.1 0.167 13 Algorithm 4 Optimal Partition into Compound Actions 1: procedure FILLCELL(row,col, s, CD, &ecrMatrix, &prevMatrix) Require: row > col 2: Let N = row − col 3: Let A = 〈αcol, . . . ,αcol+N〉 from s 4: Let ecr = P ( εcol−1 ) · CA 5: Set ecrMatrix[row,col] = ecr 6: Set prevMatrix[row,col] = col 7: for prev = col + 1 to row − 1 do 8: Let ecr′ = ecrMatrix[prev,col] + ecrMatrix[row,prev] 9: if ecr′ < ecr then 10: Set ecrMatrix[row,col] = ecr′ 11: Set prevMatrix[row,col] = prevMatrix[row,prev] 12: end if 13: end for 14: end procedure 15: function OPTIMALPARTITION(s = 〈α1, . . . ,αn〉, CD) 16: Let N = |s| 17: Let ecrMatrix = {0} 18: Let prevMatrix = {0} 19: for row = 1 to N do 20: for col = row − 1 down to 1 do 21: FILLCELL(row,col, s, CD,ecrMatrix,prevMatrix) 22: end for 23: end for 24: Let s∗ = 〈〉 25: Let prev = N 26: repeat 27: Let i = prevMatrix[prev, 1] 28: Let A = 〈αi+1, . . . ,αprev〉 29: Set s∗ = 〈A, s∗〉 30: until prev < 1 31: ASSERT(ECR (s∗) = ecrMatrix[N, 1] ) 32: return s∗ 33: end function Algorithm 5 Compound Action Sequence with Optimal Partition 1: function COMPOUNDACTIONSEQUENCE(A) 2: Let s = 〈α1, . . . ,αn〉 such that ef(αi) ≥ ef(αi+1) or Pαi Cαi ≥ Pαi+1 Cαi+1 3: Let s = OPTIMALPARTITION(s) 4: return s 5: end function 14 Algorithm 6 Optimized Exhaustive Search With Compound Actions 1: function OPTIMALCOMPOUNDACTIONSEQUENCE2(A) 2: Let bestEcr = ∞ 3: Let bestSeq = 〈〉 4: for all n! permutations of A s do 5: Let s′ = OPTIMALPARTITIONING(s) 6: if ECR (s′) < bestEcr then 7: Set bestEcr = ECR (s′) 8: Set bestSeq = s′ 9: end if 10: end for 11: return s′ 12: end function For the ef(·) ordering, the algorithm constructs the following tables—here visualized as one table with a pair of values (ECR∗(s[k,m]), i) where i− 1 is the split that minimized the (partial) ECR:   (0, 0) (0, 0) (0, 0) (0, 0) (5, 0) (0, 0) (0, 0) (0, 0) (7, 0) (2, 1) (0, 0) (0, 0) (7.45, 2) (2.45, 2) (0.45, 2) (0, 0)   Looking at the bottom left entry, we get that the last action is {α1} and then looking at the entry above it, that the first action is {α3,α2} and 〈{α3,α2},α1〉 has ECR 7.45. Even though the initial P C ordering is different, the algorithm returns the same troubleshooting sequence. However, ECR (〈{α1,α3},α2〉) = 6 + 0.35 · 4 = 7.4 which shows that the algorithm does not lead to an optimal sequence. A Heuristic for The General Problem We have seen that by changing the ordering of the atomic actions, we could improve the ECR compared to keeping the ordering fixed. We shall there- fore consider the orthogonal task of optimizing the ordering given that the partitioning is fixed. Definition 5. Let sc = 〈A1, . . . , A`〉 and sd = 〈B1, . . . , B`〉 be two compound action sequences of the same size. Then sc and sd are partition equivalent if |Ai| = |Bi| ∀i ∈{1, 2, . . . ,`}. 15 Lemma 3. Let sc = 〈. . . , Ax, . . . , Ay, . . .〉 and sd = 〈. . . , Bx, . . . , By, . . .〉 be two partition equivalent compound action sequences. Furthermore, let the only difference between sc and sd be that two actions α ∈ Ax and β ∈ Ay have been swapped s.t. β ∈ Bx and α ∈ By. Then ECR (sc) − ECR(sd) = (Cα − Cβ) · [ P(β) − P(α) + y−1∑ i=x P(Ai) ] + [P(β) − P(α)] · y∑ i=x+1 CAi+D . Proof. Let sc and sd be given as above. We then have ECR(sc) = ∑̀ i=1 CAi+D · (1 − i−1∑ j=1 P(Aj)) = ∑̀ i=1 CAi ·ai ECR(sd) = ∑̀ i=1 CBi+D · (1 − i−1∑ j=1 P(Bj)) = ∑̀ i=1 CBi · bi . The difference between these two numbers can be expressed as ECR(sc) − ECR(sd) = ∆X + ∆I + ∆Y (i) where • ∆X is the difference caused by changes to CAx, • ∆I is the difference caused by ai changing into bi for all compound actions between Ax and Ay, and • ∆Y is the difference causes by changes to CAy and ay. For i ∈{1, . . . ,x− 1,y + 1, . . . ,`} the terms cancel out. We have ∆X = ax · (CAx+D − CBx+D) = ax · (Cα − Cβ) . Furthermore, since bi = ai + P(α) − P(β)∀ i ∈{x + 1, . . . ,y}: ∆I = y−1∑ i=x+1 (ai − bi) · CAi+D = [P(β) − P(α)] · y−1∑ i=x+1 CAi+D Finally we have ∆Y = ay · CAy+D − by · CBy+D = ay · CAy+D − by · (CAy+D − Cβ + Cα) = (ay − by) · CAy+D − by · (Cα − Cβ) = [P(β) − P(α)] · CAy+D − by · (Cα − Cβ) . 16 Now observe that by = ay − [P(β) − P(α)] = ( ax − y−1∑ i=x P(Ai) ) − [P(β) − P(α)] which implies that ∆Y = [P(β) − P(α)] · CAy+D − [ ax − P(β) + P(α) − y−1∑ i=x P(Ai) ] · (Cα − Cβ) . Inserting these three results into Equation (i) yields ECR(sc) − ECR(sd) = (Cα − Cβ) · [ P(β) − P(α) + y−1∑ i=x P(Ai) ] + [P(β) − P(α)] · y∑ i=x+1 CAi+D as required. Theorem 7. Let sc, α and β be given as in Lemma 3. Then the ECR of sc can be improved by swapping α and β if and only if (Cα − Cβ) · [ P(β) − P(α) + y−1∑ i=x P(Ai) ] > [P(α) − P(β)] · y∑ i=x+1 CAi+D Proof. We should swap the two actions if ECR(sc) − ECR(sd) > 0 m (Cα − Cβ) · [ P(β) − P(α) + y−1∑ i=x P(Ai) ] + [P(β) − P(α)] · y∑ i=x+1 CAi+D > 0 m (Cα − Cβ) · [ P(β) − P(α) + y−1∑ i=x P(Ai) ] > [P(α) − P(β)] · y∑ i=x+1 CAi+D . It turns out that in certain cases we can avoid the full analysis of Theorem 7. 17 Proposition 1. Let sc, α and β be given as in Lemma 3. If P(β) > P(α) and Cα ≥ Cβ, or P(β) = P(α) and Cα > Cβ then ECR(sc) can be improved by swapping α and β. Proof. If P(β) > P(α) and Cα ≥ Cβ, then P(β)−P(α) < 0, but Cα−Cβ ≥ 0. In Theorem 7 the two other factors are strictly positive, so the left hand side becomes non-negative while the right hand side becomes negative making the inequality true. If P(β) = P(α) and Cα > Cβ, then the left hand side is strictly positive while the right hand side is zero which also makes the inequality true. Based on the above results, we can easily search for a better partition equivalent troubleshooting sequence. The procedure is given in Algorithm 7. Arguably, after the algorithm has terminated, it could be that more ac- tions could be swapped to improve the ECR. However, we have been un- successful in proving the number of possible swaps. Therefore we prefer an algorithm that terminates deterministically. Furthermore, we tried to run the procedure until no more swaps could be made on our test models, and we found that it did not improve the ECR compared to Algorithm 7. Algorithm 7 1: procedure OPTIMIZEPARTITION(& sc=〈A1, . . . , An〉) 2: for x = 1 to n do 3: for α ∈ Ax do 4: for y = x + 1 to n do 5: for β ∈ Ay do 6: if (Cα − Cβ) · [ P(β) − P(α) + ∑y−1 i=x P(Ai) ] > 7: [P(α) − P(β)] · ∑y i=x+1 CAi+D 8: then 9: SWAP(α,β) 10: end if 11: end for 12: end for 13: end for 14: end for 15: end procedure It should be clear that the algorithm runs in Θ(n3) where n is the number of atomic actions. The reason it cannot run in Θ(n2) time is that checking Theorem 7 takes linear time on average, albeit when Proposition 1 applies it only takes constant time. Because any swapping invalidates the involved 18 sums, it is not possible to precompute these as with the greedy approaches we saw earlier. We can also formulate the improved heuristic described in Algorithm 8. Algorithm 8 Compound Action Sequence with Optimized Partition 1: function COMPOUNDACTIONSEQUENCE(A) 2: Let s = 〈α1, . . . ,αn〉 such that ef(αi) ≥ ef(αi+1) or Pαi Cαi ≥ Pαi+1 Cαi+1 3: Let s = OPTIMALPARTITION(s) 4: OPTIMIZEPARTITION(s) 5: return s 6: end function We have by means of a computer constructed examples that show that the algorithm is not guaranteed to find optimal troubleshooting sequences. However, for certain models we can prove that optimality is ensured. Con- sider a model where the actions can be ordered into the sequence s = 〈α1, . . . ,αn〉 such that Pαi ≥ Pαi+1 and Cαi ≤ Cαi+1 for i ∈{1, 2, . . . , n − 1}; we call such models for non-reordable models. Theorem 8. Let s = 〈α1, . . . ,αn〉 be an atomic troubleshooting sequence in a non-reordable model such that ef(αi) ≥ ef(αi+1). Then there exists an optimal troubleshooting sequence sc consistent with s. Proof. Let sd = 〈B1, . . . , B`〉 and sc = 〈A1, . . . , A`〉 be partition equivalent compound troubleshooting sequences where sd is optimal and sc is consis- tent with s. Now assume sd is not consistent with s. Since sc 6= sd, we can find the index of the first pair of compound actions Ax and Bx where the two sequences differ. Then at least one action β ∈ Ax, but β 6∈ Bx and at least one action α ∈ Bx, but α 6∈ Ax. Furthermore, α must be found in Ay with y > x. Since sc is consistent with s and since the model is non- reordable Pβ ≥ Pα and Cβ ≤ Cα (i). We now consider two cases. Case 1: at least one of the inequalities in (i) is strict. Then if we apply Theorem 7 on sd we find that we can improve the ECR of sd by swapping α and β. However, this is a contradiction to sd being optimal. Hence in this case any optimal sequence must be consistent with s. Case 2: we have equality in Equation (i). In that case we can safely swap α and β in sd without altering the ECR. If we now have sd = sc, then we are done. If not, we can repeat the procedure until sd = sc. Since each swap puts at least one atomic action into its rightful compound action, the procedure terminates in a finite number of steps, thus proving that there exists an optimal sequence consistent with s. 19 Corollary 3. Algorithm 5 and 8 find an optimal troubleshooting sequence in non- reorderable models. Proof. Both algorithms calls OptimalPartition(·) after having sorted the actions with respect to descending efficiency. Results In this section we shall investigate the performance of the heuristic algo- rithms. In particular, we investigate how the following two model param- eters affect the precision: 1. The closeness of the initial efficiency of actions. 2. The closeness of the cost of the actions. Due to the difficulty of finding an optimal solution, we could not inves- tigate models with more than 8 actions. In the models below the repair probabilities have not been normalized to make the specification accurate (but they are of course normalized at runtime) and the efficiency stated is the efficiency rounded to three decimals when CD = 0. Model 1: close efficiencies + close costs α1 α2 α3 α4 α5 α6 α7 α8 Pα 0.21 0.17 0.19 0.155 0.185 0.139 0.17 0.135 Cα 1.8 1.4 1.7 1.3 1.6 1.2 1.5 1.1 ef(α) 0.086 0.090 0.083 0.088 0.085 0.085 0.084 0.091 Model 2: close efficiencies + non-close costs α1 α2 α3 α4 α5 α6 α7 α8 Pα 0.36 0.205 0.32 0.155 0.28 0.11 0.25 0.05 Cα 8 4 7 3 6 2 5 1 ef(α) 0.026 0.030 0.026 0.030 0.027 0.031 0.029 0.029 Model 3: non-close efficiencies + close costs α1 α2 α3 α4 α5 α6 α7 α8 Pα 0.4 0.105 0.52 0.055 0.78 0.13 0.6 0.02 Cα 1.8 1.4 1.7 1.3 1.6 1.2 1.5 1.1 ef(α) 0.085 0.029 0.117 0.016 0.187 0.042 0.153 0.007 Model 4: non-close efficiencies + non-close costs α1 α2 α3 α4 α5 α6 α7 α8 Pα 0.16 0.15 0.05 0.165 0.165 0.159 0.07 0.165 Cα 8 4 7 3 6 2 5 1 ef(α) 0.018 0.035 0.007 0.051 0.025 0.073 0.013 0.152 20 For each of the four initial models we run all the algorithms for increas- ing values of CD. The following observation shows when it does not make sense to increase CD anymore. Proposition 2. Once CD is made so high that the optimal sequence consists of one compound action, then increasing CD can never change the optimal sequence. The increment in CD was determined as one permille of the largest cost of any action in the model. Starting from CD = 0 we kept sampling until the optimal sequence consists of only one compound action. The tables below summarize the result for the four models where each column describes the results of a given algorithm. The first four rows summarize the percent- wise deviation from the optimal ECR, and the last row shows the percent of cases where the algorithm was optimal. The number of increments to CD is given in the parenthesis after the model name. The first algorithm column is the simple ef(αi) ≥ ef(αi+1) sorting. For Algorithm 2, 5, and 8 we give two numbers. The first corresponds to an initial ef(·) ordering and the second corresponds to an initial P C ordering. Model 1: close efficiencies + close costs (5828) ef(αi) ord. Alg. 3 Alg. 2 Alg. 5 Alg. 8 min 0 0 0/0 0/0 0/0 max 128.26 45.56 4.28/2.83 1.47/0.68 1.08/0.63 mean 73.14 10.05 1.66/0.66 0.77/0.07 0.15/0.02 median 79.37 5.97 1.52/0.37 0.79/0 0.11/0 % opt. 1.48 0.05 1.49/26.56 1.49/62.06 39.88/84.80 Model 2: close efficiencies + non-close costs (4201) ef(αi) ord. Alg. 3 Alg. 2 Alg. 5 Alg. 8 min 0 0 0/0 0 /0 0/0 max 97.16 38.90 6.17/2.34 5.05/0.48 4.17/0.48 mean 54.71 9.01 2.75/0.33 1.87/0.06 1.09/0.03 median 58.87 5.71 3.15/0.07 1.77/0.01 1.17/0 % opt. 0.33 0.05 8.40/32.85 9.50/46.04 18.64/63.91 Model 3: non-close efficiencies + close costs (79145) ef(αi) ord. Alg. 3 Alg. 2 Alg. 5 Alg. 8 min 0 0 0/0 0/0 0/0 max 149.50 3.93 2.80/2.80 0/0 0/0 mean 124.27 0.08 0.16/0.16 0/0 0/0 median 137.14 0 0/0 0/0 0/0 % opt. 0.45 76.27 73.89/73.89 100/100 100/100 21 Model 4: non-close efficiencies + non-close costs (18085) ef(αi) ord. Alg. 3 Alg. 2 Alg. 5 Alg. 8 min 0 0 0/0 0/0 0/0 max 219.13 4.87 2.62/2.62 0/0 0/0 mean 162.80 0.35 0.24/0.24 0/0 0/0 median 180.95 0 0/0 0/0 0/0 % opt. 0.25 53.08 76.08/76.08 100/100 100/100 We can summarize the results as follows: 1. Algorithm 8 is by far the best heuristic. Algorithm 3 is the worst of the new heuristics. Algorithm 2 performs surprisingly well, almost as good as Algorithm 5 and would thus be considered the preferred choice for a system under real-time constraints. 2. Models with non-close efficiencies are much easier to solve than mod- els with close efficiencies. Models with non-close costs seem to induce larger deviations from the optimal ECR than models with close costs. We consider none of these findings surprising. We have also tested the algorithms on a real-world model with 32 actions. The results are summarized in the table below where the statistics are com- puted as the percent-wise deviation from the overall best algorithm (Algo- rithm 8 with P C ordering). The −0 indicates that a small percentage of the cases where better than Algorithm 8. Also notice that the apparent conflict between "median" and "% best" in column two is due to rounding of the median. Model 5: 32 actions (33145) ef(αi) ord. Alg. 3 Alg. 2 Alg. 5 Alg. 8 min 0 0 0/0 −0/0 −0 max 667.30 3.79 25.30/19.05 10.68/0.05 6.07 mean 555.34 0.14 5.94/3.14 3.25/0 1.51 median 598.70 0 3.73/1.32 3.59/0 1.30 % best 0.01 48.15 0.01/0 0.01/99.43 0.01 The conclusion here is the same with one noticeable difference: Algorithm 3 is now the third best algorithm, only beaten by Algorithm 5 and 8 with P C ordering and far better than any other heuristic. The fact that Algorithm 5 and 8 are very close to each other also suggests that we are quite close to the optimal value. 22 Conclusion And Further Research We have extended a classical polynomial troubleshooting scenario with postponed system test. Albeit this is a somewhat simple extension, we conjecture that the new scenario is NP-hard. We have identified a class of non-reordable models where the best of the described heuristics are guar- anteed optimal. For reorable models, we found that the suggested heuris- tics provide quite close approximations to an optimal solution—even for heuristics with an O(n·lg n) complexity. Apart from proving or disproving NP-hardness, future research includes finding guarantees on the perfor- mance of the heuristics as well as guarantees about local optimality when optimizing a given partition. Acknowledgements We would like to thank Sven Skyum for the discussions that lead to The- orem 5. We are also grateful to Dezide Aps for providing the real-world model used in the experiments. Finally, we would like to thank our col- legues of the Machine Intelligence group at Aalborg University for provid- ing such a fruitful research environment. References F. V. Jensen, U. Kjærulff, B. Kristiansen, C. Skaanning, J. Vomlel, and M. Vomlelová. The sacso methodology for troubleshooting complex sys- tems. Artificial Intelligence for Engineering Design, Analysis and Manufac- turing, 15:321–333, 2001. J. Kadane and H. Simon. Optimal strategies for a class of constrained se- quential problems. The Annals of Statistics, 5:237–255, 1977. E. Koca and T. Bilgiç. Troubleshooting with questions and conditional costs. Proceedings of the 13th Turkish Symposium on Artificial Intelligence and Arti- ficial Neural Networks, pages 271–280, 2004a. E. Koca and T. Bilgiç. A troubleshooting approach with dependent actions. In R. L. de Mántaras and L. Saitta, editors, ECAI 2004: 16th European Con- ference on Artificial Intelligence, pages 1043–1044. IOS Press, 2004b. ISBN 1-58603-452-9. H. Langseth and F. V. Jensen. Decision theoretic troubleshooting of coherent systems . Reliability Engineering & System Safety, pages 49–62, 2003. 23 H. Langseth and F. V. Jensen. Heuristics for two extensions of basic trou- bleshooting. In Proceedings of the Seventh Scandinavian Conference on Arti- ficial Intelligence, pages 80–89. IOS Press, 2001. ISBN 1-58603-161-9. C. Skaanning and J. Vomlel. Troubleshooting with simultaneous models. Symbolic and Quantitative Approaches to Reasoning with Uncertainty, 6th, 2001. M. Vomlelová. Complexity of decision-theoretic troubleshooting. Int. J. Intell. Syst., 18(2):267–277, 2003. H. Warnquist, M. Nyberg, and P. Säby. Troubleshooting when action costs are dependent with application to a truck engine. In Proceeding of the 2008 conference on Tenth Scandinavian Conference on Artificial Intelligence, pages 68–75, Amsterdam, The Netherlands, The Netherlands, 2008. IOS Press. ISBN 978-1-58603-867-0. 24 
work_suomniwtjnb2nj44i7xi7rkjou	----	[PDF] Ergonomic risk assessment based on a Bayesian-optimised expert system | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1504/IJHFMS.2013.055782 Corpus ID: 12309252Ergonomic risk assessment based on a Bayesian-optimised expert system @article{Dansie2013ErgonomicRA, title={Ergonomic risk assessment based on a Bayesian-optimised expert system}, author={Chris Dansie and R. Sesek and D. Bloswick}, journal={International Journal of Human Factors Modelling and Simulation}, year={2013}, volume={4}, pages={23} } Chris Dansie, R. Sesek, D. Bloswick Published 2013 Engineering International Journal of Human Factors Modelling and Simulation Improper application of an ergonomic analysis tool increases the likelihood of high-risk jobs not being detected, thus jeopardising worker’s health. Likewise, significant time and cost may be incurred by redesigning jobs improperly identified as high risk. Utah Intelligent Data Driven Ergonomic Assessment System (IDDEAS) is an expert system that aggregates the outputs of multiple analysis tools to create a more predictive ergonomic analysis tool. Rules in the expert system were optimised by… Expand View via Publisher inderscience.com Save to Library Create Alert Cite Launch Research Feed Share This Paper 1 Citations View All Figures and Tables from this paper table 1 figure 1 figure 2 table 2 table 3 table 4 table 6 table 8 table 9 View All 9 Figures & Tables One Citation Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency An Application of Bayesian Regression in Ergonomics Ashutosh Kumar Singh, R. Dalpatadu Computer Science 2020 Save Alert Research Feed References SHOWING 1-5 OF 5 REFERENCES RULA: a survey method for the investigation of work-related upper limb disorders. L. Mcatamney, E. Nigel Corlett Engineering, Medicine Applied ergonomics 1993 2,124 PDF View 1 excerpt, references methods Save Alert Research Feed The Strain Index: a proposed method to analyze jobs for risk of distal upper extremity disorders. J. S. Moore, A. Garg Medicine American Industrial Hygiene Association journal 1995 615 View 1 excerpt, references methods Save Alert Research Feed A functional job evaluation technique, in ergonomics Occupational Medicine: State of the Art Reviews, 1992 Nonfatal Occupational Injuries and Illnesses Requiring Days Away From Work , 2009 – Table 4 [ online ] 2010 Evaluation and Refinement of Ergonomic Survey Tools to Evaluate Worker Risk of Cumulative Trauma Disorders [dissertation], University of Utah, Salt Lake City, UT 1999 Related Papers Abstract Figures and Tables 1 Citations 5 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_swz3plmhirdarmagm3xemu65iu	----	NSF Workshop, Washington, DC May 9-10, 2019 - About About Agenda Workshop Organizers Keynote Speakers Recommended Reading About About Welcome to the webpage of the NSF Operations Engineering (OE) Program Workshop on Decision Analytics for Dynamic Policing. This workshop is funded by NSF (NSF CMMI-1917624) to explore research directions in decision analytics and operations research (DA/OR) for enabling improved dynamic policing strategies from a variety of aspects. NSF program directors will be welcome to attend at any time during the workshop. In these pages, you will find general information about the workshop (below) and information on workshop agenda, workshop organizers, keynote speakers, and recommended reading. All workshop attendees should review The University of Texas at Arlington’s code-of-conduct policy. The workshop code of conduct with follow this policy. All workshop attendees should review The University of Texas at Arlington’s code-of-conduct policy. The workshop code of conduct with follow this policy. https://www.uta.edu/policy/hop/5-513 Date and Location Date and Location Date: May 9-10, 2019 Location: Renaissance Arlington Capital View Hotel, Arlington, VA Project Summary and Report NSF Workshop Report (May 22, 2020) The traditional approach in law enforcement studies utilizes randomized experimental design. However, this approach does not take advantage of today’s data collection capabilities. Within the last decade, the concept of predictive policing was proposed by a RAND report, funded by the National Institute of Justice (NIJ), to categorize and encourage academic research developing machine learning and statistical modeling tools to accurately predict crime. Unfortunately, existing predictive policing methods are flawed in four critical aspects: (1) They omit the human perspective, including human behavior, political biases, and community perception.  (2) They assume that the data include the necessary information for both prediction and action. (3) They assume that historical data on the policing system will be representative for the future. (4) They assume unlimited police resources. A key element in the RAND report is the concept of prediction-led policing. For this concept, it is recognized that prediction alone is inadequate without connection to a policing strategy that identifies appropriate actions. This proposed workshop on decision analytics for dynamic policing organized local teams of police departments (PDs) and academics with expertise in DA/OR and criminal justice (CJ) to build a cross-disciplinary understanding of policing challenges that may be addressable by DA/OR. Discussions took place within and across these populations, so as to examine all three perspectives. DA/OR as a discipline is not well known within the law enforcement context. Towards filling this gap, the workshop identified and recommended new directions for research in DA/OR to develop methods for policing that are cost-effective, consider social influences, identify the appropriate data, address the dynamic nature of policing, and utilize existing knowledge from PD/CJ expertise. Overall, the potential impact of DA/OR research in law enforcement is extremely high, but progress will require care, collaboration, persistence, and an open mind. Objectives The objectives of the proposed workshop are: Build an understanding of the dynamics of policing systems. Build an understanding of the controllable decisions and uncertain elements in dynamic policing systems, for example, intervention teams, criminal responses to police actions, community outreach, interactions between PDs and disciplines related to policing (such as forensic science divisions), etc. Create processes that facilitate research partnerships between DA/OR, CJ, and PDs to encourage the development of DA/OR methods for dynamic policing systems. Identify critical research themes in dynamic policing that may be studied using DA/OR. Identify methodological advancements needed in DA/OR to address the critical research themes in dynamic policing. Program 2019 NSF Workshop on Decision Analysis for Dynamic Policing Program (PDF) Department of Industrial, Manufacturing, and Systems Engineering Box 19017 | 500 West First Street | 420 Woolf Hall | Arlington, TX 76019 ©2019 The University of Texas at Arlington | College of Engineering 
work_sxuiagnoyfa5zptllsawoqgiu4	----	36 BY A. TERRY BAHILL, PAT N. HARRIS, AND ERICH SENN Sometimes what you learn building an expert system is more important than whether or not it succeeds LESSONS LEARNED T^% * 7 7 * T—i Building Expert Systems W e have made many expert sys- tems. Two of our best are Co- gito, which gives installation advice for bringing up the U N I X (Bell Laboratories) 4.2BSD operat- ing system on a VAX (Digitial Equipment Corp.), and Data Communication Diagnosis (DCD), which helps a person connect a ter- minal to a computer. This article w i l l com- pare and contrast the development, tool se- lection, and testing of these two systems. BACKGROUND Before Cogito was completed, two detailed reference manuals were used for U N I X 4.2BSD i n s t a l l a t i o n instructions. 1 2 These vo- luminous manuals were difficult to use. In contrast, Cogito is easy to use because it fil- ters the information and presents only rel- evant advice about the user's computer sys- tem. Cogito remembers data the user has previously entered and uses it to customize its response. Cogito is w r i t t e n in M . I (Tek- nowledge Inc.), runs on a personal com- puter, and uses if-then production rules to encode the task knowledge. DCD is a knowledge-based system de- signed to help engineers connect a periph- eral device (terminal or printer) to a com- puter, modem, or local area network. Most of DCD was w r i t t e n in OPS5, although parts were written in C and M . I . Instructions for connecting peripheral de- vices to computers are contained in the ref- erence and user manuals of the devices. Un- fortunately, their usage is fraught with difficulty. It is not easy to collect all the rel- evant information for the two devices in question since this information is spread over several chapters (for example, pin con- nections are discussed in the hardware sec- tion and transfer parameters in the set-up section) in several manuals written by differ- ent companies. DCD had three specific requirements: • Provide enough advice to people who have some previous computing experience but little experience in data communica- AI EXPERT • SEPTEMBER 1988 tions so they can design a cable with a 25- pin connector on each end that can be used to connect a peripheral device to a host device • Read in a description (name) of the de- vices and ask the user for values of param- eters not stored in the expert system's data base, provide a help feature to support the process of answering questions, and produce a sequence of instructions to help the user successfully interconnect the devices • Derive an approximate solution even if some information is missing or the answers to questions are unknown. Unfortunately, modern computer users are often overwhelmed with information. Expert systems can help reverse this trend. Voluminous data with many interrelation- ships presented over a short time span can overload the user. Data become useful in- formation when the relationships between items are cohesive. Both Cogito and DCD eliminate extraneous information and relate relevant information, reducing the chances of user overload. CREATING THE KNOWLEDGE BASE Our first expert system,3 made in 1984, had a flat tree: every premise was tied directly to the conclusion. However, to emulate human input-output behavior, we believed that an expert system should have a bushy tree. We discovered that we had to invent interme- diary conclusions if we were to mimic a con- sultation with a human. Furthermore, hu- mans have a limited short-term memory, thus a solution going from premises to con- clusions should follow a path of simple knowledge intermediates. Listing 1 shows the hierarchy we used to make a bushy tree for DCD. Of course, even with this hierarchy, the rules do not wrrite themselves; the expert still had to add knowledge to get the rules. For example, the expert knew that pins 2 and 3 are used to transmit and receive data. So if pin 2 transmits data, pin 3 receives data. The if-then production rules shown in Fig- ure 1 were derived from the hierarchy of Listing 1. The hierarchy shown in Listing 1 helped make sure rules were not over- looked. For example, it was easy to check that each pin was contained in at least one rule. This hierarchy was not very structured, but it still helped in writing the rules. Rules are easier to write in problem domains that can be structured more systematically with a technique that divides the problem space into a set of objects where every object has attributes and every attribute has values. This object-attribute-value ordered set is called an O-A-V triplet. Typical usage of O-A-V triplets is: the ob- ject has an attribute and a value is an a t t r i - bute. Figure 2 shows a knowledge base re- presented in the O-A-V schema. The lower left triplet in this knowledge base can be written as "coat has markings" and "striped is a marking." Another way of writing the O-A-V triplet is as a sentence ("attribute of object is val- ue"); for example, "marking of coat is striped." This usage makes the construction of facts easy: 37 AI EXPERT • SEPTEMBER 1988 description-parameters name [A..Z], [a..z], [-] sytek-unit-number XXXX.X --> [0..9] general-parameters baud rate 110 to 19200, unknown standard RS-232C, RS-422, CENTRONICS, unknown synchronization asynchronous, synchronous, unknown duplexity full-duplex, half-duplex, simplex, unknown hardware-protocol not-required, required, unknown software-protocol none, Xon/Xoff, ACK/NAK, unknown echo on, off, unknown function mapping 2-2. 2-3. unknown pinl frame-ground, not-available, unknown pin2 transmit-data, receive-data. unknown pin3 receive-data. transmit-data, unknown pin4 request-to-send, not-available, unknown pin.5 clear-to-send, not-available, unknown pin6 data-set-ready, not-available, unknown pin7 signal-ground, not-available, unknown pin8 data-carrier-detect, not-available, unknown pin20 data-terminal-ready, not-available, unknown character-format data-bits 7, 8. unknown stop-bits 1, 2. unknown parity none, even. odd. unknown LISTING 1. Parameter hierarchy. FIGURE 1. If-then production rules derived from the hierarchy in Listing 1. 38 type of animal is mammal category of mammal is ungulate extremities of ungulate is hooves. W i t h such facts, it is easy to write if-then pro- duction rules, for example: if composition of coat is hair goal = advice. question(computer-pin2) = 'What is the function of pin 2 on the computer side?'. Iegalvals(computer-pin2) = [transmit-data. receive-data. unknown]. if computer-pin2 = transmit-data then computer-pin3 = receive-data. if computer-pin3 = receive-data and terminal-pin3 = receive-data then sketch = diagraml. if sketch = diagraml and display(['Interconnect the data signals like this ± 7'.nl .nl]) then advice then type of animal is mammal. if type of animal is mammal and extremities of animal is hooves then category of animal is ungulate. if category of animal is ungulate and marking of coat is stripes then identity of animal is zebra. A tree is seldom built all at one time; in- stead, branches are added incrementally. When a branch is added, sometimes some- thing that is a value at one level becomes an object at a lower level ("mammal" is a value in "type of animal is mammal," but becomes an object in "category of mammal is ungu- late.") This O-A-V = O-A-V linkage is shown in the right column of Figure 2. In contrast, sometimes O-A-V triplets are linked together without an intervening attribute: type of animal is mammal marking of coat is stripes. This O-A-V-O-A-V linkage is shown in the left column of Figure 2. Teknowledge distributes a tutorial expert system, SACON, with their PC-based shell, M . I . The company has been refining SA- CON for 10 years. It is composed exclusive- ly of O-A-V = O-A-V linkages like those in the right column of Figure 2. This implies that perhaps these are the best linkages for an expert system. We have subsequently tried several other methods for constructing if-then production rules without forcing human experts to dis- tort their reasoning processes to produce if- then-type rules. We have found the analytic hierarchy process4-' helpful in constructing O-A-V triplets. SELECTING THE PROPER TOOL Consider the following analogy. A sculptor wishes to create an image of a dove, and has both marble and ice in which to carve. Al- though the mediums of ice and marble re- quire different tools and speed of construc- tion, the marble and ice doves are the same as the image within the sculptor's mind. The difference in mediums does not affect the result. Now suppose the sculptor wants to sculpt with iron rebar and a torch. This dove will not appear the same. In this case, the medium does affect the result. Current expert system shells force a struc- ture on the knowledge base that may be dif- ferent from the expert's structure; for ex- ample, requiring the knowledge to appear as production rules, semantic networks, or frames. When an expert has trouble devel- oping a rule that covers a certain situation, AI EXPERT • SEPTEMBER 1988 it is often said that the expert's knowledge is nonverbal and internalized. Alternatively, perhaps the expert cannot develop a rule that fits the situation because he or she does not t h i n k in terms of if-then production rules. Cogito provided m a n y examples of how humans are forced to fit their knowledge into the constraints of a tool. Cogito has many rules that m i g h t otherwise be simple lists. Furthermore, Cogito's problem do- main is probably most suitable for a for- ward-chaining strategy. If the inference en- gine had been capable of extensive forward- chaining, the knowledge base could have been shortened. Thus, the tool used defi- nitely affected the coding of the knowledge. However, an engineer's job is to find and apply the best available tool to solve a prob- lem. The tools available for Cogito were M . I and OPS5. The decision to use M . I was primarily based on the differences in rule implementations and ease of i n p u t and out- put. Cogito's rules are English-like phrases. OPS5 rules are reminiscent of LISP code; experts who are not f a m i l i a r with OPS5 or LISP have trouble reading and understand- ing OPS5 rules. Rule readability is impor- tant so the knowledge engineer can verify with the expert that the intent of the rule is the same as the coded rule. With M . I , the knowledge engineer can concentrate on problem semantics and not on the syntax of OPS5 or LISP. We selected OPS5 for DCD primarily be- cause it used forward chaining, and this seemed the best type of inferencing when all the data are available at the beginning of the inferencing process. However, a subset of the problem (creating a cable plan) was also written in M . I , which gave us the op- portunity to compare these two tools. The q u a l i t y of a system's user interface is crucial for a system's survival. M . I has an advantage in this matter since it provides le- gal value checking, which is d i f f i c u l t and cumbersome to implement in OPS5. The human-computer interface of M . I is better than t h a t of OPS5, but this did not matter in our case since all the data were collected from the human expert before the infer- ence process started using a module written in C. The knowledge base contains the rules to solve the problem at hand. Rules written in M . I are easier to read than those in OPS5. However, the easier rule syntax of M . I is only an advantage for novice knowledge en- gineers. Based on our experiences, we think experienced programmers can write either M . I or OPS5 rules with equal ease. On the other hand, M . I is at a disadvantage be- cause its conclusions cannot have values for several different expressions. This handicap is not shared by OPS5, where several differ- ent expressions can be in the conclusion of a animal type composition carnivoremarking striped spotted ungulate extremities hooves V-0 V-O FIGURE 2. Object-attribute- value hierarchy. 39 AI EXPERT • SEPTEMBER 1988 TABLE 1. Criteria and attributes used by the Multi-Attribute Utility Technique to help select an expert system shell; w— weight, r—rating, ra—ranking. rule, reducing the number of rules. M . I automatically handles the answer "unknown." In OPS5, we had to specify in each rule how "unknown" should be han- dled. The inference engine of M . I uses cer- tainty factors. OPS5, on the other hand, does not provide a mechanism that deals w i t h uncertainty. Probability calculations can be done in OPS5, but this is cumber- some. However, DCD, like most expert sys- tems, did not need certainty factors. In fact, in one of our expert system classes at the University of Arizona, Tucson, we found that only six of the 25 student-generated systems used certainty factors. QUANTITATIVE SELECTION We have helped build over 80 expert sys- tems for class projects, master's theses, and commercial ventures. We are often asked why we chose a particular expert system shell. Often our qualitative answers do not satisfy the questioner. Therefore, for the DCD project, we provided a quantitative an- swer to this question. No one-to-one match exists between problems and software tools. However, con- ceptions of particular types of problems are the same for several different domains. These conceptions are commonly referred to as consultation paradigms. Examples of consultation paradigms are diagnosis/ad- vice, planning, and design. Diagnosis/ad- vice is the consultation paradigm used by DCD. According to P. Harmon and D. King, 6 the shells the best suited for diagnosis/ad- vice are M . I , S . I , and OPS5. Hundreds of commercially available expert system shells CRITERIA and attributes w r ra REPRESENTATION Facts 0.064 78 3 if-then rules 0.065 80 3 Inference strategy 0.062 76 3 Uncertainty 0.059 72 3 IMPLEMENTATION Software available 0.080 98 1 Hardware available 0.820 100 1 Data base access 0.078 96 1 INTERFACING User display 0.074 90 2 Explanation facilities 0.069 84 2 Help functions 0.070 86 2 Editor 0.072 88 2 Debugging aid 0.067 82 2 SUPPORT Documentation 0.057 70 4 Courses 0.050 62 4 Applications 0.049 60 4 SUMMATION — 1222 — M.1 S.1 OPS5 0.9 0.9 0.9 0.9 0.9 0.9 0.1 0.1 1.0 0.9 0.9 0.7 1.0 1.0 0.5 0.8 0.8 0.5 0.5 0.8 0.7 1.0 0.8 0.0 0.5 0.9 0.8 0.8 0.5 1.0 0.8 0.6 1.0 0.8 1.0 1.0 0.9 0.8 0.5 0.5 1.0 0.7 0.8 1.0 0.8 0.689 0.685 0.834 40 exist; we did not consider every one for DCD, but restricted ourselves to these three. The following discussion illustrates a technique and is not meant to advocate a particular shell. To select the shell for our data-communi- cation problem, we concentrated on the fol- lowing criteria (Table 1): • Representation: The knowledge that en- ables us to solve data-communication prob- lems must be formulated with if-then pro- duction rules. The inference strategy used should be forward-chaining since all neces- sary data are given at the beginning of the inference process. The tool to be selected must provide an approximate solution when some information is missing or the answers to questions are unknown. • Implementation: When this project be- gan, available hardware included several personal computers, many PDF l l s , and a VAX. Available software included OPS5, M . I , a n d ROSIE. • Interfacing: Questions asked by the sys- tem should be displayed on a screen and an- swers entered on a keyboard. Solutions to a problem must be written to a data file. The knowledge base must be built and modified with a tool-independent text editor. • Support: Documentation should be rea- dable with easy-to-understand examples. Re- ferences to past tool applications and their level of success should be available. Many software packages that are com- mercially available (such as Expert Choice4) aid multiobjective decision analysis in the face of nebulous and uncertain data. We used the Multi-Attribute U t i l i t y Technique (MAUT),7 sometimes called SMART, to choose among M . I , S . I , and OPS5. The re- sults are shown in Table 1. In 'Table 1, the importance of a particular description criterion is expressed with a rank (ra) from 1 to 4, which is then scaled into the rating (r) with a range of 0 to 100. For example, rank 1 was assigned to the cri- terion implementation since the required software and hardware had to be available to finish the project on time. The weight (w) of an attribute is its indi- vidual rating divided by the sum of all the ratings. When the weights have been de- rived for the attributes, it is time to assign values (between 0 and 1) for the attributes to the three software systems being com- pared. For example, the attribute software available has an assigned value of 1.0 for OPS5 since all the necessary software was up and running. S.I has a value of 0.0 be- cause we did not have it and could not af- ford to buy it. Next the weights are multiplied by the as- signed values and the sums placed at the bottom of the columns for M . I , S . I , and OPS5. The best tool for the job was the one AI E X P E R T • SEPTEMBER 1988 4.4.1. Initializing /etc/fstab Change into the directory /etc and copy the appropriate file from: fstab.rm03 fstab.rmOS fstab.rmSO fstab.raGO fstab.raSO fstab.ra81 fstab.rpOS fstab.rp07 fstab.rkO? fstab.up160m (160MB up drives) fstab.up300m (300MB up drives) fstab.hp400m (400Mb hp drives) fstab.up (other up drives) fstab.hp (other hp drives) to the file /etc/fstab, i.e.: #cd /etc #cp fstab.xxx fstab This will set up the initial information about the usage of disk partitions. LISTING 2. Instructions for installing fstab.2 LISTING 3. Cogito's instructions for installing fstab. 42 with the largest sum, in this case OPS5. (Please note that the ratings, weights, and assigned values are subjective; they were chosen by Senn. The other authors would have assigned different values and therefore might have drawn different conclusions.) USER SATISFACTION We found it easier to install a U N I X operat- ing system using Cogito t h a n the U N I X ref- erence manuals because with Cogito the in- formation presented to the user was tailored to the specific application. For ex- ample, in the disk-definition state, the file '/etc/fstab' must be created. Assume that the boot disk is a 'AMPEX 3 C O M ' connect- ed to ubaO at drive 1. This implies that the disk address is 1. Also assume the user's name is 'Pat' and that Pat is ready to install 'fstab'. Since the instructions given in the manual are generic, the installer must relate the general instructions to the specific appli- cation. Listing 2 is a copy of a relevant sec- tion of the manual. Listing 3 is a copy of Co- gito's instructions. Cogito has remembered information the user gave it a half-hour ago: that the user's name is Pat and the disk is an AMPEX # cd /etc * cp fstab.up300m junk * vi junk (Edit the file. Pat. a. Add the line Vdev/upOb::sw::'. b. Give the global substitute command ':g/upO/s//up1/'. c. Save the new contents and quit the editor.) # cat junk » fstab 300M. Cogito has used this information to make its instructions specific. Cogito's instructions are personalized, relevant to the user's task, clear and complete. This example illustrates that the manual's instructions are incomplete and rely on pre- vious and implicit knowledge. The instruc- tions are incomplete because the swap parti- tion '/dev/upOb', must be added to the file. They rely on previous knowledge because the user must recall that the U N I X stand- alone disk name for a 'AMPEX 300M' is 'up'. They rely on implicit knowledge be- cause the user must somehow know that the disk address of the partitions must be changed from '0' to T. TESTING Testing Cogito was difficult. It was impossi- ble to present the system with every possible combination of inputs and evaluate its out- puts. The best test we devised was to let in- tended users try Cogito in many hypotheti- cal situations. If the knowledge base was incomplete, then in some situations the ad- vice given to the users should be incorrect. Cogito was tested by the computer sys- tems administrator and a professor of the Dept. of Systems and Industrial Engineering and the computer systems administrator for the Dept. of Computer Science at the Uni- versity. All found Cogito's advice to be com- plete and correct. In an effort to bolster our testing, we asked several nonexpert computer users to try the system. Their evaluations empha- sized the difficulties they had using Cogito and making sense of its queries or output, and the extent to which they were able to fool the system with plausible (but nonsensi- cal) inputs. Recently, we gave Cogito a real test. Be- cause of inadequate glue on the heads of our RA81 disk, we had to rebuild our oper- ating system from the distribution tapes. Cogito helped us. We found a few omissions in Cogito's advice, but no mistakes, and completed the task in about 12 hours. It took us three months the first time we built the system, but we were way down on the learning curve then. This experience has reinforced our belief that all expert systems are inadequately test- ed. No quantitative procedures exist for testing expert systems. Most tests merely in- volve running a few case studies; they do not exhaust all possibilities. For example, we are confident that Cogito works well for small VAX 750 systems, but we cannot be sure it will work as well for 730s or 780s. We carefully planned our tests and evalua- tions for DCD. Because the DCD system was designed for users with little previous expe- rience in data communication, we allowed juniors from our microcomputer class to AI EXPERT • SEPTEMBER 1988 test and evaluate it. Thirty-three student teams (two members per team) participated in testing and evaluation. It is impossible to test all the combina- tions of problems likely to occur in data communication. Therefore, only two typical test cases were given. System users were asked to use DCD to: • Design a cable that would connect a WYSE terminal to a VAX computer and to give advice about setting up the WYSE ter- minal to make the interconnection work • Design a cable that would connect a WYSE terminal to the Sytek local area net- work and to give advice about setting up the WYSE terminal and the Sytek port. We specified two major subtasks in the evaluation process. First, we focused on the accuracy of the system. Was the advice given by the system sufficient to make the inter- connection work? This assessment was triv- ial since the advice given by the system re- sulted in go/no-go situations. Second, the students evaluated the quality of the human- computer interaction. Was the system easy to use? These evaluation criteria should be directly proportional to future frequency of use of the system. All the undergraduate students accom- plished the tasks with an average time of 20 minutes, suggesting that the accuracy of the system was 100%. The students rated the quality of human-computer interaction as "user friendly." As a further test of the DCD system, a selected graduate student tried to connect a WYSE terminal to the VAX computer and a WYSE terminal to the Sytek net using only the manufactures' user and reference manuals. He succeeded after 1.5 hours—about three times longer than the undergraduate students who established the same connections by consulting the DCD system. After this initial test, the DCD system was used by juniors at the University for five consecutive semesters. This continued us- age proves its usefulness. However, DCD's usefulness is diminishing as our equipment changes, and none of our knowledge engi- neers have experience with OPS5 to update the knowledge base. GENERAL-PURPOSE TESTING The most difficult aspect of expert system design is testing. The traditional testing method is to have a human expert run many sample cases on the expert system. This consumes a lot of the expert's time and does not guarantee finding all mistakes. Con- versely, brute-force enumeration of all in- puts is impossible for most systems. There- fore we have developed a general-purpose tool to help debug knowledge bases without the intervention of human experts. We have tie-Based Technology for nalltalk-8O. .ly there's an expert system shell for the Smalltalk-80™ m. For those wanting a powerful way to combine rule- 1 technology with true object-oriented programming, 1BLE™ provides the right tools for the task. iUMBLE is an integrated expert system tool that runs tly in the Smalltalk-80 environment. It features both #ard and forward chaining, a modular certainty m, an advanced user interface including a graphical ser, and a complete programmer interface. > Bang for the Buck! HUMBLE costs far less than >arable expert systems. Before buying, ask yourself: / expand this tool to fit my particular needs? Are the 'es available if I need to change them? Will this tool rate with my own programs? Can I port the results of my to end user machines easily? at all?" If you answered to any of these questions, then take a good look at IBLE. IUMBLE is available for the Apple Macintosh, Sun-2 iun-3 Series, Tektronix 4400 and 4310 Series, Apollo [AIN Series, and Xerox 1100 and 6085 Series work- ins. [UMBLE runs in any License Version of the Smalltalk- >tem. Knowledge bases can be transferred between any ltalk-80 version without modification. [HUMBLEf • .u* i«xl«r.' For more information, contact the Smalltalk-80 Marketing Manager Xerox Special Information Systems P.O. Box 5608 Pasadena, CA 91107 (818) 351-2351 CIRCLE 17 ON READER SERVICE CARD XEROX®, Smalltalk-80, and HUMBLE are trademarks of Xerox Corporation. * ACM Conference on Object-Oriented Programming Systems, Languages, and Applications, September 25-30,1988. San Diego, California 44 A. Terry Bahill is a professor of systems engineering at the University of Arizona in Tucson. He is also a vice president of the IEEE Systems, Man, and Cybernetics Society and an associate editor of IEEE Expert. Pat Harris is a systems engineer. Erich Senn is head of technical staff with Telekuhrs, Schweiz, Zurich, Switzerland. tried to make this tool generic so it can work on any knowledge base, no matter which expert system shell is used. The first component of the tool is a sim- ple spelling checker with a preprocessor that replaces hyphens, underbars, and par- entheses with spaces and removes terms spe- cific to the shell being used, such as legalvals and automaticmenu. The second component of this tool is designed for backward-chain- ing s h e l l s . It collects all terms that appear in the premises (the //parts) of the rules, and flags as potential mistakes those that do not also appear in the conclusions (the then parts) of the rules or in questions or goal statements. It also traces every term appear- ing in a premise to the conclusion of that rule, then to the premise of another rule, and so on u n t i l a goal statement is reached. If it does not reach a goal statement, then a rule is missing. The third component of this tool, which is not yet completed, acts like Tieresias8 in pointing out rules that do not fit the pattern established by the rest of the rules in the knowledge base. For example, when we ran this component on our animal-discrimina- tion knowledge base, it said, "Rule 15 is sus- picious, because most rules that mention the animal's fur also ask if it has spots. Rule 15 does not." The fourth component comes into play at run time. The firing of each rule is record- ed. Rules that never fire and rules that fire for all test cases are probably mistakes and are brought to the attention of the human expert. SIZE Cogito contains 800 rules, 200KB, and took 800 hours to build. DCD contains 60 rules, 40KB (plus an interface program and a large help file), and took 120 hours to build. To better understand these numbers, we performed statistical analysis of 25 simple expert systems written by students in partial f u l f i l l m e n t of the requirements of a course in expert systems in 1987. (We call these sys- tems simple because they did not use exter- nal functions or special hardware.) The average student spent 100 hours on the project: 15% of this time was spent learning the subject, 10% interviewing the expert, 60% developing and debugging the knowledge base, and 15% testing the sys- tem. The average knowledge base con- tained 150 rules and required 32KB. From these systems and from a less formal evalua- t i o n of 60 other student-generated expert systems constructed over a three-year time span, we concluded that the number of rules (or knowledge base entries) is not a good in- dication of the complexity of a system. However, we also concluded that in rou- tine expert systems, novice knowledge engi- neers consume three to four hours per kilo- byte of knowledge base. Sophisticated knowledge engineers may expend as much as five to 10 hours per kilobyte, because they know how to collapse similar rules into more general expressions and might be us- ing external functions. APPROPRIATENESS We used M.I and OPS5. One is primarily a backward chainer; the other primarily a for- ward chainer. We do not think much differ- ence exists between the two. For our prob- lems, we could have made either type of chaining work. After learning each of these expert system shells, Senn remarked, "It's just another language." This remark has another implication about the knowledge-transfer process. Ex- pert systems have made one step forward accompanied, unfortunately, by one step backward. While expert systems permit a higher level of abstraction, they also de- mand that the knowledge base remains tightly coupled to the inference engine. Even though the knowledge has already been captured in Cogito or DCD, convert- ing this knowledge to another inference en- gine would be difficult. A fundamental breakthrough occurred when traditional programming languages became indepen- dent of the processor; a similar break- through must be effected within the expert system realm. How7 can one identify a task appropriate for a PC-based expert system? First, there must be a human who performs that task better than most other people. Second, the task's solution must be explainable by the human expert in words without relying on pictures. Third, the problem must be solv- able a 20-minute or even one-hour tele- phone conversation with the expert. If it would take a human two days to solve the problem, it is far too complicated for an ex- pert system; if the human gives the answer in two seconds, it is too simple. Fourth, problems that involve determining one of many possible solutions are ideal candidates for expert systems, such as answering the question, "What disease does the person have?" Fifth, the problem must encompass uncertainty or inexactness in the data, or a diagram on paper would be superior. The first four criteria help predict wheth- er the expert system will be successful, whereas the fifth helps decide whether a complex expert system shell is needed or if some simpler tool would suffice. The animal classification system (Figure 2) is an example of a problem domain that has no uncertain- ty or inexactness. This task could be per- formed better using a diagram such as Fig- ure 2 than an expert system on a PC. Given these criteria, giving advice for AI EXPERT • SEPTEMBER 1988 bringing up U N I X on a VAX computer was an inappropriate task for a PC-based expert system. It cannot be done in a one-hour conversation with an expert. We estimate that it would take an expert one or two days to do this task (it took us three months to do it the first time!). Cogito's 800 rules filled two floppy disks. We succeeded in making an expert system that worked, but it was hard work. A more powerful tool (such as KEE [IntelliCorp], S.I [Teknowlege], ART [Inference Corp.], or Knowledge Craft [Carnegie Group Inc.]) would have been more appropriate. However, helping a person connect a ter- minal to a computer is an appropriate task for an expert system. In the future, we think most computer equipment manufacturers will have expert systems (interactive text- books) to help their sales personnel install their equipment. This w i l l insulate their re- search engineers from mundane questions and build confidence in the sales force, which will then be able to answer difficult questions about their product. The two real-world expert systems dis- cussed here were designed to perform cer- tain tasks, and they did them; so they were successful. However, the more important as- pect of these endeavors was that we learned a great deal about making expert systems. We learned that an expert system shell is not Merlin's long-lost magic wand. You can- not do things with expert system shells that you could not do in any other language; but you can develop a prototype a lot faster. We learned that expert systems can be made friendlier than conventional computer pro- grams. But we also learned that testing and validating expert systems is d i f f i c u l t . Q] This research was partially supported by grant no. AFOSR-88-0076from 'the Air Force Office of Scientific Research. REFERENCES 1. i'XIX Programmer's Manual Reference Guide. Berke- ley, C a l i f . : U S E N I X Assoc., 1985. 2. U\IX Systems Manager's Manual. Berkeley, Calif.: U S E N I X Assoc., 1985. 3. Moller, R., A.T. Bahill, and L. Swisher. "The De- velopment of ESI AC, an Expert System for I d e n t i f y i n g A u t i s t i c Children," in Proceedings of the International Conference on Systems, Man, and Cybernetics. New York, N.Y.: IEEE, 1985, pp. 875-878. 4. Forman, E.H., and T.L. Saaty. Expert Choice. McLean, Va.: Decision Support Software, 1986. 5. Saaty, T.L., and K.P. Kerns. Analytic Planning: Or- ganization of Systems. New York, N.Y.: Pergamon Press, 1985. 6. Harmon, P., and D. King. Expert Systems: Artificial Intelligence in Business. New York, N.Y.: Wiley, 1985. 7. Edwards, W. "How to Use M u l t i a t t r i b u t e U t i l i t y Measurement for Social Decision Making." IEEE Tran- scripts on Systems, Man, and Cybernetics SMC-7: 326-340, 1977. 8. Buchanan, B.C., and E.H. S h o r t l i f f e . Rule-Based Expert Systems. Reading, Mass.: Addison-Wesley, 1984. Now available in Turbo C,- Microsoft C,® JPI Modula 2,* and Logitech Modula 2/ Turbo Expert. Now it doesn't take a genius to plug into Expert Systems. 7or only $99.95, you can incorporate the power of a full-fledged Expert System into your TURBO PASCAL programs. Seamlessly. A f f o r d a b l y . Anally. Actual Expert Systems, developed for simple use by any Turbo Pascal 4.0 programmer. Take a look at all the features you suddenly have available with this single Turbo Pascal 4.0 Unit: The ability to create large Expert Systems, or even l i n k m u l t i p l e Expert Systems together. A powerful backward-chaining inference engine. Easy flow of both data and program :ontrol between Turbo Expert and the other parts of your program, to provide Expert System analysis of any database, spreadsheet, file >r data structure. The ability to add new rules in the middle of a consultation, so your Expert Systems can l e a r n — r e a l l y learn—and )ecome even more intelligent. You also have the ability to create large rule bases and still have plenty of room left for your program, thanks to conservative memory ise. You can link multiple rule bases, you'll be compatible with our Turbo Companion units, and you have available advanced features like late and time arithmetic, confidence factors, windowing, demons, agendas, blackboards, and more. Imagine a single "EXE" file containing your user interface and data handling, and a f u l l Expert System. For a limited time, get a FREE copy of our Turbo SnapShot graphics package worth $79.95. We'll give one away with every copy of urbo Expert sold between now and September 30. This package will let you capture graphics images from other programs nd use them in any Turbo Pascal program. You can even convert images from any CGA or EGA format to any other. * On top of all that, Turbo Snapshot has routines for graphic gauges and dials as well as mouse support. You'll have all ou need for a sophisticated graphics f r o n t - e n d for your Expert Systems — f r e e . B.-.^^.I,.B Call for more i n f o r m a t i o n or to order, (317) 876-3042. Software Artistry Inc., 3500 Depauw Blvd., Suite 2021, Indianapolis, VT.VJJ 45 V 46268. I n c l u d e $5.00 for shipping and handling. CIRCLE 18 ON READER SERVICE CARD 
work_szd4dqbzvvhetfefbpab6uyo5y	----	The development of an intelligent tutoring system on design of liquid retaining structures 1 Expert Systems with Applications, Vol. 22, No. 2, 2002, pp. 169-178 EXPERT SYSTEM APPLACTION ON PRELIMINARY DESIGN OF WATER RETAINING STRUCTURES K.W. Chau Department of Civil & Structural Engineering The Hong Kong Polytechnic University, Hong Kong F. Albermani Department of Civil Engineering University of Queensland, Australia ABSTRACT Design of liquid retaining structures involves many decisions to be made by the designer based on rules of thumb, heuristics, judgment, code of practice and previous experience. Various design parameters to be chosen include configuration, material, loading, etc. A novice engineer may face many difficulties in the design process. Recent developments in artificial intelligence and emerging field of knowledge-based system have made widespread applications in different fields. However, no attempt has been made to apply this intelligent system to the design of liquid retaining structures. The objective of this study is, thus, to develop a knowledge-based system (KBS) that has the ability to assist engineers in the preliminary design of liquid retaining structures. Moreover, it can provide expert advice to the user in selection of design criteria, design parameters and optimum configuration based on minimum cost. The development of a prototype KBS for the design of liquid retaining structures (LIQUID), using blackboard architecture with hybrid knowledge representation techniques including production rule system and object-oriented approach, is presented in this paper. An expert system shell, Visual Rule Studio, is employed to facilitate the development of this prototype system. INTRODUCTION In the past decade, the potential of artificial intelligence (AI) techniques for providing assistance in the solution of engineering problems has been recognized. Knowledge-based systems (KBS) are considered suitable for solving problems that demand considerable expertise, judgment or rules of thumb. As a result of years of research in artificial intelligence, KBS have emerged as a most promising application covering a wide range of applications (Adeli & Hawkins 1991, Chau 1992, Chau & Chen 2001, Chau & Ng 1996, Chau & Yang 1992, Chau & Yang 1994, Chau & Zhang 1995, Lin & Albermani 2001, Tah & Carr 2001, Waheed & Adeli 2000). KBS have developed into practical problem solving tools that can reach a level of performance comparable to that of a human expert in some specific problem domains. All these applications can be broadly classified into the following categories: diagnosis; design; data interpretation; planning; and education. Areas of early applications of KBS technology include medical diagnosis, mineral exploration and chemical spectroscopy. Many researchers are developing programs that borrow AI concepts to automate common engineering analyses. AI has made widespread applications in different fields. In structural engineering field, there is a need to develop programming environments that can incorporate engineering judgment along with algorithmic tools. (Kitzmiller & Kowalik 1987) However, no attempt has been made to apply this intelligent system to the design of liquid retaining structures. The objective of this work is, thus, to develop a KBS prototype that has the ability to assist engineers in the design of liquid retaining structures and to provide expert advice to the user in selection of design criteria, design parameters and This is the Pre-Published Version. 2 optimum structural section based on minimum cost. The development of a KBS in design of liquid retaining structures (LIQUID), using blackboard architecture with hybrid knowledge representation techniques including production rule system and object-oriented approach, is presented. An expert system shell, Visual Rule Studio, is employed to facilitate the development of this prototype system. It is a coupled system in which AI-based symbolic processing is combined with the traditional numerical processing. The KBS developed is based on British Standards Code of Practice BS8007: 1987: Design of concrete structures for retaining aqueous liquids (British Standards Institution 1987). KBS are interactive computer programs that mimic the decision making and reasoning processes of human experts in solving a specific complex problem, by providing expert advice, answering questions, and justifying their conclusions. Figure 1 shows the schematic view of a KBS. Three basic components are knowledge base, context and inference mechanism. The knowledge base is a collection of general facts, rules of thumb and causal models of the behavior specific to the problem domain. The context contains facts that reflect the current state of the problem, constructed dynamically by the inference mechanism from the information provided by the user and the knowledge base. The inference mechanism guides the decision making process by using the knowledge base to manipulate the context. In addition, three other desirable components are a user interface, an explanation facility and a knowledge acquisition module. The interface is responsible for translating the interactive input as specified by the user to the form used by the KBS. The explanation module provides explanations of the inferences used by the KBS: why a certain fact is requested and how a conclusion was reached. The knowledge acquisition module serves as an interface between the experts and the KBS and provides a means for entering domain specific knowledge into the knowledge base. For a conventional program, the order of execution is predetermined. Updates need considerable effort other than by the programmer. The programmer must ensure completeness and uniqueness of the solution. The user, perceiving the program as a blackbox, has no idea to why certain results have been produced. On the contrary, KBS eliminate the above impediments by partitioning between the knowledge base and the control strategy. This allows for incremental addition of knowledge without manipulation of the overall program structure and hence the programmer need not guarantee completeness. By ranking several alternatives with inexact inference methods, several solutions with different confidence factors can be provided for a particular input condition. The user can also question the results through the explanation module. Several approaches for declarative representation of knowledge are available in the AI literature, i.e. rule-based production system, frames and object-oriented programming. A production system is a collection of rules and is believed to be good at describing heuristic knowledge. A frame system is suitable for a complex and rich representation of knowledge. Object-oriented programming concept is used, in which a computer program consists of a number of independent objects that process jobs by exchanging information they need via messages. It seems a good idea to utilize a hybrid approach combining two representations to solve structural design problems. Several tools are available for developing a knowledge-based system, i.e. traditional programming languages, high-level tools or expert system shells. The programming language designed for AI (LISP or PROLOG) can be used, but the programming effort required will be tremendous. High-level tools can provide an integrated knowledge engineering environment combining features of the AI languages appropriately and efficiently. Typical examples of this type of approach are ART (Clayton 1985) and KEE (Intellicorp 1986). An expert system shell provides the skeleton of a knowledge-based system incorporating inference engine, user interface and knowledge storage medium. Once the domain knowledge is filled in, it can perform the desired function of a knowledge-based system with the specified inference strategy and representation method. 3 KNOWLEDGE ON DESIGN OF LIQUID RETAINING STRUCTURES Different Configurations and Design Parameters Liquid retaining structure is a structure which is designed and constructed to retain aqueous liquid. It is subject to lateral liquid pressure and earth pressure when it is located underground. In Hong Kong, most of these structures are constructed using reinforced concrete with design life of 50 years. Crack widths need to be checked to ensure impermeability of concrete and prevention from corrosion of reinforcement. Normally, two kinds of classification are used regarding liquid retaining structures, i.e. according to the shape or the location. Based on the shape, it is classified as rectangular, circular or polygon. Based on its location, it is classified as underground or above ground. Compared with circular tank structure with the same width, rectangular liquid retaining structure has larger volume. However, because of stress concentration at corners, rectangular structures will be more vulnerable to failure. Since a circular structure can be constructed monolithically without any construction joints, it has better strength quality. With precise structural analysis, circular structure has a better control in deflection, crack width, bending moment resistance, axial compression resistance, and shear resistance than rectangular structure. Polygon liquid retaining structure is usually used for aesthetic purposes, such as a fountain in a garden and the retaining height is usually not very high. Its major design consideration is on crack width to ensure its impermeability and is seldom used in industrial or domestic fields. Underground liquid retaining structure usually has larger base area, which cannot easily be supported by structure above the ground. In some cases, the tank is connected to underground pipe network system, in order to reduce maintenance cost, it will be constructed underground to suit the invert level of the pipe network system. The underground structure is mainly subjected to lateral earth pressure and lateral water pressure, which is due to the underground water table. Besides checking structural failure mode, bearing capacity of soil and settlement of structure also need to be checked. If the soil bearing capacity does not satisfy the requirement, pile or raft foundation will be required. Liquid retaining structure above the ground is mainly subject to liquid pressure due to its own retaining liquid. The structure can either rest on ground concrete slab immediately or rest on supports. Earthquake loading is another concern especially for fire prevention tanks. However this type of loading will not be considered in this paper. There are a great variety of factors affecting the decision in selecting design criteria and design method. These factors are: dimensions, location, ground condition, support condition, groundwater conditions, aesthetic properties, design life, exposure condition, usage, roofing, availability of construction materials. Liquid retaining structures, like any other engineering structures, should not fail to satisfy any of its performance criteria. According to the Code of Practice BS8007, the two main classes of limit state to be considered are serviceability limit state and ultimate limit state. Serviceability limit state is design against deflection and crack width. Serviceability Limit State Cracks in reinforced concrete structures cannot be avoided. Moreover, concrete expands and contracts with the increase in temperature during hydration of cement and the subsequent fall in temperature. The effects of thermal contraction and drying shrinkage may be controlled by the provision of reinforcements and movement joints. The design calculations of both the serviceability limit state of cracking due to thermal and moisture effects and flexural effects can be tedious and time consuming. 4 In the design of liquid-retaining structures it is essential to restrict the width of cracks in the structure for direct tension and flexure or restrained temperature and moisture effects. This assumes that, if cracks do not exceed maximum design surface crack widths, they can heal autogeneously. Normal crack width control is 0.2mm while, for severe cases, allowable crack width is 0.1mm. Appendix A of BS 8007 provides calculations of minimum reinforcement, crack spacing and crack widths in relation to temperature and moisture effects. Two criteria should be met. First of all, minimum reinforcement required to control the early thermal and shrinkage cracking (within three days) is related to the ratio between the direct tensile strength of immature concrete and the characteristic strength of reinforcement. Moreover, it should also be greater than the reinforcement ratio , based on area of surface zone, determined from the following equation:  = (fct/fb)( /2 Wmax)(/2)(T1+T2) (1) where fct = direct tensile strength of immature concrete fb = average bond strength between concrete and steel  = bar diameter Wmax = maximum design surface crack width  = coefficient of thermal expansion of mature concrete T1 = temperature rise due to hydration of cement T2 = temperature fall due to seasonal variation Appendix B of BS 8007 provides calculations of crack widths in mature concrete under structural loading. The service flexural capacity of different slab thickness and reinforcement arrangements under differing conditions of crack width limitation, concrete strength and cover differ. The design is based on the elastic theory for cracked section. Based on the equilibrium of forces, the neutral axis position is determined to be a function of the modular ratio and area of reinforcement of a given section and is independent of the applied moment. From BS 8007 Appendix B, the crack width W is determined from the following equation: where m is the average strain for calculation of crack width allowing for concrete stiffening effect acr is the distance from the point considered to the surface of the nearest longitudinal bar c is the minimum cover to the tension reinforcement x is the neutral axis position h is the slab thickness In calculating crack width, BS 8007 gives allowance to the stiffening effect of concrete between cracks m = 1 - 2 where 1 is the average strain for calculation of crack width 2 is the strain due to the stiffening effect of concrete between cracks For crack width W of 0.2 mm )(21 3 xh ca a W cr mcr      (2) )(3 )( 2 2 xdAE xhb ss    (4) (3) 5 For crack width W of 0.1 mm where d is the effective depth of section b is the width of section Es is moduli of elasticity of reinforcement As is area of reinforcement. Ultimate Limit State The ultimate flexural capacity of a range of slab thickness and reinforcement arrangements under differing conditions of service and concrete strength and cover is calculated and tabulated in the KBS based on British Standards Code of Practice BS8110: 1985: Structural use of concrete (British Standards Institution 1985). Concrete stress is limited to 0.45 of the characteristic strength whilst reinforcement stress is limited to 0.87 of the yield strength. Besides, the neutral axis position and the lever arm are limited to 0.5 and 0.95 of the effective depth respectively. Part of a sample table showing ultimate moment capacity and unit cost of section (for concrete slab thickness 200-225mm, grade 40/20 concrete, high yield reinforcement, concrete cover 40mm, unit cost of concrete $500/m3, unit cost of reinforcement $3000/tonne) is shown in Table 1. Since all these design parameters can vary at different design situations and an enormous amount of data are involved, a Microsoft Access database file Section is used to store the table. Of course, in order to minimize the size of this database, some heuristics are used to limit the choice of some design parameters to only practical values. For examples: the concrete slab thickness is limited to 200, 225, 250, 275, 300, 350,400, 450, 500, 600, 700, 800, 900 and 1000 mm; allowable reinforcement sizes are 10, 12, 16, 20, 25, 32 and 40 mm; concrete cover is limited to 40, 50, 60 and 75 mm. Structural Optimization The unit cost per meter length of the concrete section (consisting of concrete area and reinforcement area) can be calculated and tabulated in the KBS. Each concrete section will have its service moment capacity, ultimate moment capacity, ultimate shear capacity, crack width limitation. Structural optimization can be effected so as to achieve the minimum unit cost per meter length of concrete section. KBS FOR DESIGN OF LIQUID RETAINING STRUCTURES (LIQUID) Blackboard Architecture The blackboard architecture is intended to support development of systems in domains characterized by interaction between diverse sources of knowledge and hence provides a framework for integrating knowledge from several sources. The blackboard serves as a global data structure, which facilitates this interaction. A common analogy may be made to problem-solving in domains where a number of experts in different areas of specialties co-operate over the solution which any one of them could never achieve alone. In order to facilitate this process, they agree to use a blackboard to post (or write) any partial result they can contribute separately. Each expert takes turns to write on the blackboard and, in case more experts wish to write simultaneously, the conflict is resolved by some pre-defined strategy. The blackboard architecture has been successfully used in solving a wide range of tasks, such as speech recognition, signal processing, and planning. (Engelmore & Morgan 1988) A blackboard system consists of a number of knowledge sources that communicate through a blackboard and are )(3 )(5.1 2 2 xdAE xhb ss    (5) 6 controlled by an inference mechanism. Figure 2 shows the architecture of the current blackboard system. The main components of a typical blackboard system are knowledge sources, entries, blackboard, and inference mechanism. The knowledge base consists of a number of knowledge sources (KSs), which contain the knowledge. These knowledge sources are independent chunks of knowledge and do not directly communicate with each other. Instead, they participate in the problem solving process by creating entries in a global database – the blackboard. Knowledge modules look at the blackboard to see if suitable data is present to trigger their execution. If they are selected, execution results in new or altered data on the blackboard, which will then trigger other knowledge modules. Solving a problem using blackboard architecture is based on cooperation of the knowledge modules present. Each knowledge source consists of a condition-action pair. Actions are executed whenever the conditions are satisfied in the blackboard. Entries are the immediate results produced by the system. In a typical system, each entry has a certainty factor as well as a specification. The blackboard or context consists of the information or entries generated by the knowledge sources during the problem solving process. It is organized into a number of levels each representing different aspects or stages of the solution process. Each level contains objects and attributes that are important to the representation of the problem. Normally, knowledge sources are specific to certain levels in the blackboard, i.e., the activation of a certain knowledge sources depends on the entries generated at certain levels in the blackboard, while the actions of the knowledge source modify entries at some other level. The main units in the blackboard are hypotheses, which at various levels are related through structural relationships. Design context and the processes in the design, both represented as objects, are organized separately. Declarative knowledge, divided into two groups (Design Status and Design Concept), is stored in the blackboard. Design Status contains attributes that represent indicators used to keep information of the current stage of every design entities. The processes in the design are represented in the knowledge source level. Multi-formalism approach is employed consisting of objects, rules, procedural methods, extensive numerical algorithm and databases. Usually, in a typical blackboard model, the inference mechanism consists of the agenda (or scheduler) and the monitor. The agenda keeps track of all the events in the blackboard and calculates the priority of execution for knowledge sources that were generated as a result of the activation of other knowledge sources. The monitor takes the element with the highest priority and executes it. However, there is no fixed agenda and monitor in the current blackboard model. Since a scheduler is not applied, control of design process is performed by Process Control Knowledge modules in knowledge sources. These modules act opportunistically upon being triggered by user or situation during the design process. Since design steps in this system are explicitly seen on the main screen display, the sequence of design processes are primarily selected by the user. However, validity of the sequencing is checked by Process Control Knowledge modules in the knowledge source. Without fixed agenda, the user is free to change input parameters and check intermediate results given by the system during the design session. Because of its modularity, the blackboard architecture enables easy incremental development of a software system. Developers can integrate different methods of knowledge representation in a single system because of the modularity of knowledge sources. Features of LIQUID To facilitate development of the knowledge-based system, expert system shell containing specific representation methods and inference mechanisms is employed. The knowledge base and explanation 7 facility of the system have been developed using a commercially available expert system shell called Visual Rule Studio which is a hybrid application development tool that integrates object-oriented techniques and expert system technology with traditional, procedural programming. (RuleMachine Corporation 1998) Visual Rule Studio installs as an integral part of Microsoft Visual Basic 5.0 as an ActiveX Designer. By isolating rules as component objects, separate from objects and application logic, the shell allows developers to leverage the proven productivity of today’s component oriented development tools, such as Visual Basic. Rule development becomes a natural part of the component architecture development process. It produces objects that can interact with virtually any modern development product. Objects are used to encapsulate knowledge structure, procedures, and values. An object’s structure is defined by its class and attribute declarations within a RuleSet. Object behavior is tightly bound to attributes in the form of facets, methods, rules, and demons. Figure 3 shows the structure of Visual Rule Studio components. Each attribute of a class has a specific attribute type. The attribute types are compound, multicompound, instance reference, numeric, simple, string, interval, and time. Facets provide control over how the inference engines process and use attributes. Methods establish developer-defined procedures associated with each attribute. The set of backward-chaining rules that conclude the same attribute is called the attribute’s rule group. The set of forward-chaining demons that reference the same attribute in their antecedents is called the attribute’s demon group. LIQUID combines expert systems technologies, object-oriented programming, relational database models and hypertext/graphics in Microsoft Windows environment. By defining various types of windows as different classes, such as Check Box, Option Button, List Box, Command Button, Text Box, etc., they can inherit common characteristics or/and possess their own special properties. The knowledge used has been acquired mostly from written documents such as code of practice, textbooks and design manuals and complemented by experienced engineers involved with the design of liquid retaining structures. The domain knowledge is translated into procedures and methods using object-oriented representation. The system can be compiled and encrypted to create a run-only system. This run-only system can be installed on a microcomputer for office use. The user can always overrule any design options and recommendations provided by the system. In other words, it plays the role of a knowledgeable assistant only. The input data provided by the user will be rejected if it is not within the range specified. It can explain its line of reasoning for obtaining an answer. It provides in multi-window graphics text display where graphic images are combined with valuable textual information. This kind of intelligent graphics is extremely valuable to structural designers because it enhances their confidence in the design provided by the knowledge-based system. The system offers a friendly user interface. Mouse, keyboard, or both can be used to navigate the application. The use of a mouse or other pointing device makes the data entry a simple task even for novice computer users. As such, users simply point and click their way through the process to appreciate the dynamic behavior of the system. Input data entry is kept at minimum. Input data are provided by the user mostly through selection of appropriate values of parameters from the menus and providing answers to the queries made by the system. The inference engines control the strategies that determine how, from where, and in what order a knowledge base draws its conclusions. These inference strategies model the reasoning processes an expert uses when solving a problem. The Process Control Knowledge module involves metalevel knowledge which establishes the problem solving strategy and controls the execution of the Design Knowledge modules. It evaluates the Design Status and decides what action should be performed mainly in data-driven forward chaining mechanism. The knowledge representations of the Design 8 Knowledge modules, however, need both forward and backward chaining inference mechanism to arrive at the solution. EXAMPLE OF CONSULTATION SESSION In order to demonstrate how LIQUID assist engineer in the preliminary design of liquid retaining structure, a sample run is demonstrated in this section. The purpose of the design is to find a feasible optimum structural section in term of minimum unit cost per meter length, based on the standard slab thickness and reinforcement size as well as bar spacing. Little explanation facility is required since each screen display was designed to be user-friendly to follow. Upon execution of the package LIQUID, the main menu screen is displayed as shown in Figure 4. A number of command buttons representing different design functions, which are grouped into three groups: Preliminary Design, Detailed Design, and Design Summary, can be activated by a mouse. Alternatively, a pull-down menu can also be employed for the same purpose. Tool-tip-text, giving detailed description on the function of the command button, is also available when the mouse pointer is dragged and placed over the button. It should be noted that only a few buttons, including Structural specification Button, Help Button, Exit Button, Restart Button, are enabled and all the others requiring the completion of some pre-requisite processes are disabled. To start the preliminary design of liquid retaining structure in this example, Structural specification in Preliminary design has to be selected. Structural specification display is then popped up as shown in Figure 5. The user is required to input the basic spatial requirements, which are volume, height, shape, location, exposure condition, density of liquid and, if the shape is rectangular, the width/breadth ratio. The validity of the data input will be checked and, if any data are not within normal limits, warning message will be displayed to guide the correct input. Simple structural analysis program will be evoked to determine the maximum service and ultimate moments, shear force, deflection, etc. By selecting Search for configuration Button, LIQUID will search for 15 feasible alternative sections from the database Section which will be shown in Figure 6. Back to the main menu, the next process is to select Alternative Evaluation where the system will evaluate each of these 15 recorded feasible configuration by using engineering heuristic based on span/depth ratio, unit cost and stability. As shown in Figure 6, 15 best alternative sections will be tabled in order of minimum unit cost per meter length which can satisfy all the criteria including crack width and strength. The computation of unit cost per meter length includes only the material cost of concrete and reinforcement and excludes cost of formwork and excavation. Here, it must be noted that it is not necessary for the user to adopt the best alternative proposed by the system, which has the minimum unit cost per meter length. The user can override the system’s decision by clicking the User’s section Option Button and choose any among the 15 proposed alternatives by pointing it in the table. The user can also view the database of sectional properties as well as make changes to the default parameters by selecting Sectional property Button. Screen showing details of sectional properties with default material properties, concrete cover and unit cost of materials is shown in Figure 7. The user can change any of these default parameters from a list of available choices. The database will then be updated, reflecting the corresponding changes, to give the revised service moment capacity, ultimate moment capacity, ultimate shear capacity and unit cost per meter length. When the Crack width checking Button is selected from the main menu, details of the chosen parameters, the actual crack width and the statement on whether or not the crack width criterion is satisfied will be shown in Figure 8. In this example, the computed actual crack width under the current 9 configuration is 0.17 mm which is less than the designed crack width of 0.2 mm and hence the crack width criterion is satisfied under the provision of BS 8007. Further Development In order to improve the accuracy of structural analysis results, a rigorous commercially available finite element analysis package can be evoked as an external program in the detailed design stage and integrated into knowledge-based system. A model generation program will be required to generate files containing information such as nodal coordinates, member connectivity, support condition and loading specification which are ready to be analyzed by the analysis program. CONCLUSIONS A great variety of factors affect decisions in selecting design criteria and design parameters. An integrated microcomputer knowledge-based system (LIQUID), which provides much information necessary to make decisions, was developed to combine expert knowledge with conventional algorithmic programming and relational databases. Advice of structural optimization can be effected in the preliminary design stage so as to achieve the minimum unit cost per meter length of concrete section. It has been demonstrated that the hybrid knowledge representation approach combining production rule system and object-oriented programming technique to the design of liquid retaining structures is possible with the implementation of blackboard system architecture. It is appropriate to integrate algorithmic and symbolic programming on structural design into a single computer-aided environment running under a Windows platform. The educational spin-off of knowledge-based systems in training novice engineers or in transferring knowledge cannot be overemphasized. REFERENCES Adeli, H. and Hawkins, D.W. (1991) A Hierarchical Expert System for Design of Floors in Highrise Buildings, Computers & Structures, 41(4): 773-778. Engelmore, R. and Morgan, T. (1988) Blackboard Systems, Addison_Wesley, Wokingham. British Standards Institution (1987) BS 8007:Design of Concrete Structures for Retaining Aqueous Liquids, British Standards Institution, London. British Standards Institution (1985) BS 8110:Structural Use of Concrete, British Standards Institution, London. Chau, K.W. (1992) An Expert System for the Design of Gravity-type Vertical Seawalls, Engineering Applications of Artificial Intelligence, 5(4): 363-367. Chau, K.W. and Chen, W. (2001) An Example of Expert System on Numerical Modelling System in Coastal Processes, Advances in Engineering Software, 32(9): 695-703. Chau, K.W. and Ng, Vitus (1996) A Knowledge-Based Expert System for Design of Thrust Blocks for Water Pipelines in Hong Kong, Journal of Water Supply Research and Industry - Aqua, 45(2): 96-99. Chau, K.W. and Yang, Wen-wu (1992) A Knowledge-Based Expert System for Unsteady Open Channel Flow, Engineering Applications of Artificial Intelligence, 5(5): 425-430. Chau, K.W. and Yang, Wen-wu (1994) Structuring and Evaluation of VP-Expert Based Knowledge Bases, Engineering Applications of Artificial Intelligence, 7(4): 447-454. 10 Chau, K.W. and Zhang, X.N. (1995) An Expert System for Flow Routing in a River Network, Advances in Engineering Software, 22(3): 139-146. Clayton, B.D. (1985) “ART-Programming Tutorial, Vols 1-3”, Inference Corporation, Los Angeles, California, U.S.A. IntelliCorp (1986) “KEE Software Development User Manual”, 3rd ed. Kitzmiller, C.T. and Kowalik, J. S. (1987) Coupling Symbolic and Numeric Computing in Knowledge-Based Systems, AI Magazine, Summer: 5-90. Lin, S. and Albermani, F. (2001) Lattice-Dome Design Using a Knowledge-based System Approach, Computer-aided Civil and Infrastructure Engineering, 16(4): 268-286. RuleMachines Corporation (1998) Visual Rule Studio Developer’s Guide, Indialantic. Tah, J.H.M. and Carr, V. (2001) Knowledge-based Approach to Construction Project Risk Management, Journal of Computing in Civil Engineering, ASCE, 15(3): 170-177. Waheed, A. and Adeli, H. (2000) A Knowledge-based System for Evaluation of Superload Permit Applications, Expert Systems with Applications, 18(1): 51-58. 11 Slab thickness (mm) Bar size (mm) Bar spacing (mm) Concrete cover (mm) Bar area percentage (%) Ultimate moment (kNm/m) Ultimate shear (kN/m) Unit cost ($/m) 200 10 125 40 0.41 37 107 130 200 12 150 40 0.49 44 114 136 200 10 100 40 0.51 46 116 137 225 10 125 40 0.35 43 114 142 200 16 225 40 0.59 51 120 142 200 12 125 40 0.59 52 121 143 200 16 200 40 0.66 57 125 147 225 12 150 40 0.42 51 121 148 225 10 100 40 0.44 54 123 149 200 12 100 40 0.73 64 130 153 200 16 175 40 0.76 64 130 154 Table 1. Part of sample table showing ultimate moment capacity and unit cost of section (for concrete slab thickness 200-225mm, grade 40/20 concrete, high yield reinforcement, concrete cover 40mm, unit cost of concrete $50/m3, unit cost of reinforcement $3000/tonne) 12 Figure 1 Schematic View of a Knowledge-based System User interface Explanation facility Knowledge acquisition facility Context Inference mechanism Knowledge base User Expert 13 Legends: Figure 2 Blackboard Architecture of the Current KBS Agenda triggered by user or situation Knowledge base composing of knowledge sources Blackboard with hierarchical levels Design Status 1st Design Concept Level (k-1)th Design Concept Level kth Design Concept Level DKS DKS DKS DKS PCKS DKS Inference Mechanism entry Design knowledge source Process Control knowledge source 14 Figure 3 Structure of Visual Studio Components Class Name Properties Attributes Name Type Facets Methods Rules Demons Instance Attribute Value 15 Figure 4 Main menu of the knowledge-based system 16 Figure 5 Screen showing structural specification in preliminary design 17 Figure 6 Screen showing 15 alternative sections with unit cost 18 Figure 7 Screen showing sectional properties allowing changes to design parameters 19 Figure 8 Screen showing crack width checking in accordance to BS 8007 
work_t34rnly4tjhlhedbnoeq53nhwe	----	COMEX: A Cost Management Expert System COMEX: A Cost Management Expert System Danijela Grahovac1, Vladan Devedzic2 FON, School of Business Administration, University of Belgrade, POB 52, Jove Ilica 154, 11000 Belgrade, Serbia 1Email: danielle.grahovac@gmail.com 2Email: devedzic@fon.rs Abstract -Understanding and using cost accounting is a key to success in any type of organization, because it helps managers to collect important data about the costs assigned to products, services and customers and use it for planning and control function. This paper discusses development and application of an expert system for reasonable decision making and continuous processes improvement. The system is called COMEX (COst Management EXpert system) and is implemented in JavaDON, an open-source expert systems shell based on the OBOA framework for developing intelligent systems. The paper also shows how COMEX can be used for improvement of the Billing Department business processes. COMEX relies on information generated using the Activity Based Costing method, accounting method based on activities, created to improve cost management system by focusing on the driver costs of the single activity. Keywords: Expert system, Cost Management System, ABC method I. INTRODUCTION A Cost Management System (CMS) consists of a set of formal methods developed for planning and controlling the organization’s cost-generating activities related to its short-term objectives and long- term strategies. Business entities face two major challenges: achieving profitability in the short run, and maintaining a competitive position in the long run. An effective CMS must provide to the managers with the information needed to meet both of these challenges. The information generated from a CMS should benefit all functional areas of the organization entity. Thus, the system should integrate the areas shown in Fig. 1 [1], and should “improve the quality, content, relevance, and timing of cost information that managers use for short-term and long-term decision making” [9]. Fig. 1. An Integrated Cost Management System The information generated from a CMS should help managers measure and evaluate performance. The measurements may be used to evaluate human or equipment performance or to evaluate future investment opportunities. Finally, to maintain a competitive position in an industry, a company must generate the information necessary to define and implement its organizational strategies. As the strategy represents the link between an organization’s goals and objectives and the operational activities executed by the organization, companies have to be certain that such a connection exists. Information provided by the CMS enables managers to perform strategic analyses on issues such as: a) determining core competencies and organizational constraints from a cost-benefit perspective, and b) assessing the positive and negative financial and nonfinancial factors of strategic and operational plans. Since the world of business competition is dynamic, and creative managers are constantly devising new business practices and innovative approaches to competition, a CMS must be dynamic to. Managers in many companies were concerned about the product costing information being provided by the traditional cost accounting systems. The general consensus is that product costs currently being developed are useful in preparing financial statements, but are often of limited use for management decision making. At the same time, Activity-based costing (ABC) helps companies to work more efficiently, determine costs more accurately, and control and evaluate performance more effectively. The rest of the paper is organized as follows. In Section 2 we describe the Activity-Based Costing (ABC) method. Section 3 is the central section in the paper. It discusses the architecture and design of COMEX in detail. Section 4 shows some discussions. Finally, Section 5 contains some conclusion remarks and further directions. II. ACTIVITY-BASED COSTING In a business context, an activity is defined as a repetitive action performed in fulfillment of business functions. Management should first investigate activities that reflect the major and most significant processes conducted by the company. These activities normally overlap several functional areas and occur horizontally across the firm’s departments lines. ABC is a new method of cost assignment, more compatible with the increased high technology environment in which business operates. ABC is a powerful tool for measuring company’s performances. It recognizes cause-effect relationships between assigned objects costs. Once activity costs are identified, ABC assigns costs to products/services based on the resources they consume. ABC assigns costs to products on the basis of the types and quantities of activities that must be performed to create those ICIT 2011 The 5th International Conference on Information Technology products. This costing system accumulates costs for activity centers in multiple cost groups at a variety of levels (unit, batch, product, and organizational) and then allocates these costs using multiple cost drivers. Thus, costs are assigned more accurately, and managers can focus on controlling activities that cause costs rather than trying to control the costs that result from the activities. The use of ABC should provide a more realistic picture of actual production costs than has traditionally been available. Managers have more success in understanding how organization is using its own capital by assigning costs for defined activities using the ABC system. As a more accurate CMS than traditional cost accounting, ABC identifies opportunities to improve business process effectiveness and efficiency by determining the "true" cost of a product or a service. It is implemented in many companies around the world. The way of applying ABC systems differs from one organization to another. Some organizations use ABC system as the main CMS, but many are selective and can be applied on subunits or separate functions of organizations. Note, however, that although ABC typically provides better cost estimates than the traditional overhead allocation process, it is not a panacea for all managerial concerns. This is the point where an expert system can takes a part. III. COMEX COMEX is a Cost Management Expert system developed to support the work of billing departments of utility companies. COMEX can be applied to any kind of the billing (e.g., electricity, water or waste). Previous research has shown that good conception of the bill influences customer relationship, which is a very important part of this business branch [3]. The purpose of COMEX is to satisfy the customer’s needs and stay concurrent with lower costs. For example, the customer inquiries are consuming a large percentage of the company's resources, and reducing costs for this activity brings savings and customer’s satisfaction. This is applicable to all other activities. COMEX recommends how to achieve these savings. COMEX interprets the data resulting from applying the ABC method. ABC method is the process of studying activities to classify them into activity centers, than trace costs to those activities and determine the cost drivers. The map of the activities supplied by the ABC method is essential for a knowledge engineer during the design phase of development. COMEX is implemented using the JavaDON ES shell, developed at FON, School of Business Administration, University of Belgrade, Serbia, as an academic project. JavaDON is an open-source shell with a graphical user interface for Windows platform (see [Tomic et al., 2006] for details). A. Knowledge Representation COMEX is essentially a frame-based system. There are nine frames in its knowledge base: a) Customer b) Activity c) BillInquiry d) WebSite e) AccountVerification f) CustomerInquiry g) Correspodence h) Expert i) ExpertSolution The frames are inter-related. The UML class diagram in Fig. 2 explains the relations among the frames. Fig. 2. – System COMEX Class Diagram Figure 3 shows all nine frames; then the frame Customer is expanded and its slots are shown. Fig. 3. –The Customer frame and the typeOfCustomer slot Each frame can have subframes and parent frames. There are two slots in the Customer frame but no subframes and parent frames. SubFrames and ParentFrames have no children nodes in Fig. 3. All frames branch directly from the root of the knowledge base. These frames have slots that represent the corresponding frame's attributes. Each slot contains the definition of the attribute it represents, a question to be presented to the end-user in dialogs with COMEX whenever the system asks the user to supply a value for the attribute, possible answers, and a description (canned text used for providing explanations in the user's consultations with COMEX). A slot's attribute is shown in its node in the expanded tree. For example, the typeOfCustomer slot from the Customer frame, Fig. 4, is defined to ask the user "What is the type of the customer?", and the user is supposed to select between two possible types of customers: Commercial and Residential. ICIT 2011 The 5th International Conference on Information Technology Fig. 4. – The typeOfCustomer slot B. Dialogs Another very important piece of knowledge in the COMEX knowledge base is the typeOfActivity slot in the Activity frame. The related question that COMEX will ask the user during his/her session with the system is «Which activity do you want to improve? » There are four possible answers: Bill Inquiry, Customer Inquiry, Account Verification and Correspondence (Fig. 5). Fig. 5. Selecting an activity to improve After the user selects an activity, COMEX will typically ask a question about the average percentage of the total cost allocated to the chosen activity. The average slot in the Activity frame stores the value of the total cost allocated in the activity of the Billing Department (BD). Depending on activity selected, the value entered is compared to the actual market average percentage. For example, the market average percentage for the Account Verification activity is 16%. Anything under this value doesn’t need improvement; in case of a greater value, COMEX goes forward into improvement and asks the question Q11: Question: Q10 What percent of the total cost is allocated for this activity? The related rules in the knowledge base are: Rule17: IF A=“Commercial“ AND 16 < average THEN Solution1 End Rule18: IF A=“Commercial“ AND 16 >= average THEN ask-question Q11. Fig. 6 shows the activity diagram for one of the four branches, i.e. Account Verification improvement. There are five different endings that can happen, depending on the user’s answers. Select type of customer Enter bill costs Select type of activity Enter activity participation in % Enter allContractsVerification Enter contractPercentage Enter serviceCosts [billCosts > market average] Check costs Show solution1 [billCosts <= market average] Check participation Compare costs Show solution9 [billCosts>serviceCosts] Show solution8 Compare percentages Show solution7 Check Type of Activity [typeOfActivity<>"Verifikacija"] Check Type of Customer [typeOfActivity="Verifikacija"] [typeOfCustomer="Privredni subjekti"] Show warning [typeOfCustomer="Domacinstva"] M [billCosts<=serviceCosts] [True] [False] [contractPercentage <=10] M [contractPercentage >10] [ average < market average ] [ average >= market average ] ExpertUser Fig. 6. – Activity diagram for Account Verification ICIT 2011 The 5th International Conference on Information Technology C. User Modeling COMEX users are knowledge engineers and end users. A knowledge engineer is the one who creates the knowledge base and tests the system. Speaking in terms of UML, the use cases related to knowledge engineers are illustrated in Fig. 7. They include building the knowledge base, creating a new project, opening and updating the knowledge base, and the like. During the design of the COMEX system, the knowledge engineer communicated with a cost management expert, an ABC expert, and accounting and system quality experts. All the important knowledge was entered in the knowledge base through multiple iterations. Edit COMEX Save COMEX Close COMEX <<extend>> Knowledge Engineer Create Project COMEX <<extend>> <<extend>> <<include>> Build Knowledge Base <<extend>> Open Project COMEX <<extend>> <<extend>> <<include>><<extend>> Fig. 7. Knowledge engineer's use cases with COMEX COMEX end users are ABC experts, accountants, managers, business analysts, and students. End users do not have the permission to update knowledge base elements. He/she opens the system (Open Project COMEX), enters the session data (Enter Data) and starts the forward chaining inferencing (Forward Chaining), Fig. 8. Open Project COMEX Us er Inference <<include>> Enter DataForward Chaining <<include>> <<include>> Fig. 8. – End-user use cases D. Reasoning and Conflict Resolution COMEX reasons using forward chaining with its rules. Fig. 9 shows some of the COMEX' rules. IF and THEN clauses use frame slots to create logical expressions (IF clauses) and assertions (THEN clauses). In addition to IF and THEN parts, each rule has its TYPE (which can be either AND or OR) and IMPORTANCE (the rule's priority, used to resolve conflicts when multiple rules can fire). For example, the BillCostComGreater rule (Fig. 9) states that if the slot typeOfCustomer from the Customers frame has the value Commercial and the billCost slot from the same frame has a value greater than or equal to 8.5 (IF clause), then the system will ask the question (the ask_question action) related to the typeOfActivity slot of the Activity frame. That means if the selected customer class is Commercial and bill is greater than a certain limit, COMEX will proceed with the process improvement. Fig. 9. The structure of rules in COMEX Since COMEX is a forward-chaining system, in each cycle of its reasoning process a set of candidate rules to fire is created (the rules that have their IF clauses satisfied). Conflict resolution and selection of the rule to fire is done by selecting the highest priority rule. Each rule can be fired only once. The reasoning process is initiated by having three rules automatically satisfied when the session starts: a) INIT (Fig. 10), b) BillCostCommRes (Fig. 12) and c) ShowSolution (Fig. 15). The INIT rule's importance is set to 10 and thus it will be fired first. Fig. 10. The INIT rule Firing the INIT rule means that the action ask_question(Customer:typeOfCustomer) will be executed. As a result, the user is presented the question about the customer type as in Fig. 11. ICIT 2011 The 5th International Conference on Information Technology Fig. 11. The start of a session Upon selecting the customer type and clicking Next, the next cycle of forward chaining is executed and the BillCommRes rule (with the IMPORTANCE set to 9.0, Fig. 12) is fired. Its THEN clause executes the action ask_question(Customer:billCost) and the user is asked to enter the appropriate bill cost, Fig. 13. In this example, it is assumed that the user is a Commercial customer and that he has entered the value 6.5 that corresponds to this customer type. Fig. 12. The BillCostCommRes rule Fig. 13. The result of executing the ask_question(Customer:billCost) action This result in COMEX activates the BillCostComLess rule (Fig. 14), because the bill cost entered is less than market average. The rule's THEN clauses are added to the working memory and the slots Expert: finalSolution and ExpertSolution:solution are assigned the appropriate values. The ShowSolution rule (Fig.15) will use these values when presenting the solution that COMEX displays to the user at the end of the session, Fig. 16. Note that ShowSolution has an empty IF clause, so it will certainly fire. However, it has the lowest IMPORTANCE (-10), hence it will fire only at the end of the session. Fig. 14. The BillCostComLess rule Fig. 15. The ShowSolution rule Fig. 16. – The solution presented by COMEX This simple example demonstrates how COMEX works. In real- world examples, its reasoning takes many more steps and can get quite complex. IV. DISCUSSION Why expert system? Is this technology “in” or took off in 1990’s? The answers to these questions are reflecting the authors' inspiration to research the subject. If we called COMEX a business rule management system (BRMS), it would certainly be “in”. And what is a BRMS? The little secret is that a BRMS was originally called "expert system" and its rule engine was called "inference engine". Then it was called "business rule engine", and today this kind of system is known as "business rule management system”. Hence COMEX may change its name/categorization over time. Expert systems deserve all due respect. Expert systems didn't really go away. They went undercover [8]. This statement can be justified by, e.g., AMEX's Authorizer's Assistant (ES for credit cards transactions management) [2]. This magnificent ES is working perfectly since 1988. ICIT 2011 The 5th International Conference on Information Technology V. CONCLUSIONS COMEX demonstrates how business effectiveness can be improved by applying a new method for cost management and expert system technology to interpret the data resulting from applying the method. In other words, the ABC method calculates the real costs and COMEX reasons over the data to provide useful advice and recommendations. Together, they bring significant improvement to cost management. Our current and future work is oriented towards putting the COMEX demo online and commercialization of the developed full version. REFERENCES [1] Barfield, Jesse T.; Raiborn, Cecily A.; Kinney, Michael R. Student Solutions Manual: Cost Accounting: Traditions and Innovations, Cincinnati, Ohio, U.S.A. South-Western Pub. 1998, Third Ed. [2] American Express Authorizer's Assistant, [Online]. Available: http://bizrules.info/page/, last visited, Jan, 2009 [3] Charles T. Horngren, Gary L. Sundem, Management Accounting, Fourth Canadian Edition, 2001 [4] Cotton, W., Activity-based costing in New Zealand, Working paper, SUNY [5] Devedzic, V., Inteligentni informacioni sistemi, Digit i FON, Beograd, 2000 [6] D.Grahovac, R.Kučinar, ABC (Activity Based Costing) system in the function of measuring continuous processes improvement, 2 and Internacional Conference ICQME, 2007., Milocer [7] Paul M. Collier, Accounting for Managers: Interpreting accounting information for decision-making, Aston Business School, Aston University, 2003 John Wiley & Sons Ltd. [8] Ask Rolo, Why business rules? Why not expert systems? [Online]. Available:http://bizrules.info/weblog/2007/04/why_business_rules_why _not_exp.html, last visited, Dec, 2008 [9] Steven C. Schnoebelen, Integrating an Advanced Cost Management System into Operating Systems, Journal of Cost Management , 1993 [10] G. Shimic, V. Devedzic, Building an intelligent system using modern Internet technologies, Expert Systems With Applications, Vol.25, No.3, 2003, pp. 231-246. [11] Bojan Tomic, JavaDON: Shell za razvoj ekspertnih sistema zasnovan na modelu OBOA, BSc thesis, FON, University of Belgrade, 2005 ICIT 2011 The 5th International Conference on Information Technology 
work_t3lalctekbfwxp26fbp3znzryi	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220995238 Params is empty 220995238 exception Params is empty 2021/04/06-03:05:39 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220995238 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:39 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_t3owubjwgbfqbksdvigm3wafcy	----	Microsoft Word - pajhoheshnameh46.docx )ايراني -رويكرد اسالمي(فصلنامه پژوهشنامه اقتصادي 65-84 ، صفحات1391پاييز ، 46، شماره دوازدهمسال هاي بورس اوراق بهادار تهران با بندي شركتارايه يك رويكرد جديد براي رتبه هاي فازي پوششي دادهاستفاده از تحليل و مهران نعمتي ، هما صمدي، اميرنظري شيركوهيلقمان حاتمي شيركوهي 6/12/1390 :تاريخ پذيرش 10/5/1390 :تاريخ دريافت چكيده است كه از جمله نهادهاي عمده و اساسي در بازار سازمان يافته ينهاد ،بورس هاي پذيرفته شده در بورس اوراق شركتارزيابي عملكرد . شود سرمايه محسوب مي - بهادار و انتخاب نوع واحد تجاري، موضوعي مهم براي مديران مالي و سرمايه گذاري را تا حد توان ريسك سرمايهگذاران است، زيرا با يك ارزيابي جامع، مي - بندي شركتهدف از انجام اين پژوهش ارزيابي و اولويت. قابل قبولي كاهش داد ها در شرايط عدم قطعيت مي با استفاده از رويكرد تحليل پوششي دادههاي سها شاخص تأثيرگذار بر ارزيابي 10در راستاي هدف، اين پژوهش با استفاده از . است فهرست ارايه شده از سوي سازمان بورس اوراق بهادار شركت موجود در 50عملكرد، سال چهارم ماهه سه - تهران بهادار بورس اوراق ترفعال شركت 50«را با عنوان دهد كه روش تحليل نتايج اين طرح نشان مي. دهدمورد بررسي قرار مي» 1388 - سازي نوسانات و عدم قطعيتهاي فازي، رويكرد مناسبي براي مدل پوششي داده ها در هاي مالي مختلف و ارزيابي عملكرد شركت هاي سال هاي موجود در داده . هاي تأثيرگذار است صشرايط عدم قطعيت شاخ .JEL :C02, B41, G11 بنديطبقه .، بورس اوراق بهادار منطق فازيها، ارزيابي عملكرد، بندي، تحليل پوششي دادههاي رتبه شاخص: هاكليدواژه  ،رونيكپست الكت دانشگاه آزاد اسالمي، واحد رودبار، مربي گروه حسابداري :hatami@iauroudbar.ac.ir .  دانشجوي مهندسي صنايع، دانشگاه آزاد اسالمي، دانشكده مهندسي صنايع و مكانيك، واحد قزوين، پست الكترونيكي :nazari.aminsh@gmail.com.  ،كارشناسي ارشد مهندسي صنايع، دانشگاه آزاد اسالمي، دانشكده صنايع، واحد تهران جنوب .homasamadi@gmail.com: پست الكترونيكي  مربي گروه رياضي، دانشگاه آزاد اسالمي، واحد رودبار، پست الكترونيكي :mehran.nemati53@yahoo.com. 46دوازدهم شماره سال )ايراني - رويكرد اسالمي( اقتصاديفصلنامه پژوهشنامه 66 مقدمه -1 گيري در ارتباط با هر موضوعي نياز به اطالعات و دانش مرتبط در دنياي رقابتي امروز، تصميم گذاري در بازار بورس كه يك بازار تخصصي و نياز به داشتن چنين اطالعاتي براي سرمايه. دارد گذاري بوده و همچنين از حجم اطالعاتي هاي سرمايه داراي ريسك بيشتري نسبت به ديگر حوزه ه و گذار بايد دانش فني مربوط به تجزيبدين منظور سرمايه. شودبااليي برخوردار است، تشديد مي . هاي سهامي و ارزيابي عملكرد آنها را داشته باشد تحليل اطالعات مربوط به شركت گذار در بازار سهام، انتخاب سهام مناسب و معقول در زمان گيري سرمايهمسأله كليدي براي تصميم ق هاي سهامي يكي از مباحثي است كه پس از رون بندي شركترتبه. گذاري استدرست براي سرمايه هاي ها و معيارها همواره مورد تحقيق و پژوهش واقع شده و براي آن روشگونه شركتگرفتن اين ها، معيار مطمئن و متداولي از آنجا كه ميزان سوددهي و وضعيت مالي شركت. اندزيادي تعريف كرده مالي منتشر شده هاي ها از نظر وضعيت مالي و صورتبراي ارزيابي آنهاست، در اين مطالعه نيز شركت بندي با استفاده از بندي ارايه شده است، اما رتبههاي زيادي براي رتبه اگرچه روش. شوندبندي ميرده ها روش مطلوبي به نظر ها و عدم قطعيت در دادهدليل كمبود دادهها و منطق فازي بهتحليل پوششي داده آن را به قطعات مساوي به نام سهم تقسيم شود كه سرمايهشركت سهامي به شركتي گفته مي. رسدمي براساس قانون تجارت ايران، . آيدها به شمار ميكنند كه منبع اصلي تأمين مالي اين شركتمي و ) دارنفر سهام 3با حداقل (شركت سهامي خاص : شوندهاي سهامي به دو نوع تقسيم مي شركت در . ها تابع ضوابط خاص خود هستنداين شركتهر يك از ). دارسهام 5با حداقل (شركت سهامي عام هاي سهامي عامي است كه در بازار بورس و سهام هاي سهامي، شركت اين مطالعه منظور از شركت صورت سود و زيان، ترازنامه و صورت : اند ازهاي مالي اساسي عبارت صورت. ايران ثبت شده باشد . هاي نقدي جريان گيري براي گذار به يك روش و چهارچوب تصميمان شده، نياز سرمايهبا در نظر گرفتن مطالب بي هاي هاي برتر در بازار بورس كه بتواند عدم قطعيت گيري در مورد انتخاب شركتتسهيل فرآيند تصميم با توجه به مطالب بيان شده در ارتباط . شودروشني احساس ميسازي كند، بهموجود در اطالعات را مدل هاي ارزيابي بندي برمبناي ارزيابي عملكرد، در اين پژوهش پس از شناسايي شاخصتبهبا لزوم ر ها عملكرد و تهيه فهرست جامعي از آنها در بازار بورس ايران و با استفاده از روش تحليل پوششي داده 67 بورس اوراق بهادار تهران هاي بندي شركتارايه يك رويكرد جديد براي رتبه ند و گيربرتر سهامي موجود در بازار بورس، مورد ارزيابي قرار مي شركت 50در محيط عدم قطعيت، .شوندبندي ميرتبه پيشينه تحقيق -2 كند كه ارزيابي يك شركت نبايد تنها از ديدگاه مالي باشد، بلكه بايد به، تأكيد مي)1386(نبوتي 4رو، وي با تكيه بر مدل كارت ارزيابي متوازن كه ازاين. جانبه مورد ارزيابي قرار گيردصورت همه هاي دهد و با استفاده از روشداخلي و رشد و يادگيري را پوشش ميهاي جنبه مالي، مشتري، فرآيند .بندي صنايع ايران پرداخته استبه رتبه SAWگيري چندمعياره از جمله تاپسيس و تصميم آنها . اندگيري در خريد و فروش سهام پرداخته، در تحقيقي به مسأله تصميم)2007( 1البدوي و همكاران ضريب شاخص مختلف مانند نسبت جاري، 23و با در نظر گرفتن PROMETHEEبا استفاده از روش گيري براي هاي سهامي و تصميم بندي شركتبه رتبه... ، ارزش بازار، نسبت نفدينگي وقدرت پرداخت بندي براي ارزيابي و رتبهرا دلي كمي ، م)1385(پور العليبدآذر و ع .خريد و فروش سهام پرداختند ، SAW ،TOPSIS گيري چندشاخصه جبراني مانندهاي تصميم ازرگاني استاني با روشهاي ب سازمان هاي در نظر گرفته شده توسط بيشتر معيار. ارايه كردند تاكسونومي كالسيك و تاكسونومي غيركالسيك، . گرفتهاي كيفي بودند كه سه بعد اصلي نيروي انساني، عمليات سازمان و مديريت را دربر مي آنها، معيار هاي داد كه روش تاكسونومي غيركالسيك سازگارترين روش ارزيابي سازماننتايج تحقيق آنها نشان مي .بازرگاني است هاي شركت كار خريد سهامهب كهرا الملليهاي بين ، پنج پورتفوليو از شركت)2000(و همكاران 2ردمن بندي قرار ارپ، ترينر و جنسن مورد ارزيابي و رتبههاي ش كردند با استفاده از شاخصمي ديگر مبادرت پورتفوليو، چهار براي ترينر و معيار شارپ دو از حاصل بنديداد كه رتبهنتايج پژوهش آنها نشان مي. دادند ريسك با و كل ريسك با آمده دستهب بازده بود كهآن دهنده امر نشان بودند؛ اين يكي دقيقاً .است نزديك سيستماتيك ريسك به كل ريسك ن،بنابراي و دارد يكساني سيستماتيك نتايج شركت به پنجاه مربوط گزارش شود، مي منتشر تهران بهادار اوراق بورس در كه هايي گزارش از يكي تهران بهادار اوراق بورس در برتر هاي شركت ارزيابي. است بهادار اوراق بورس در شده پذيرفته برتر . شودبازار انجام مي بر هاشركت ثيرگذاريأت ميزان و سهام نقدشوندگي قدرت از تركيبيبرمبناي 1- Albadvi etal, 2007, PP. 673–683. 2- Redman etal, 2000, PP. 75–85. 46دوازدهم شماره سال )ايراني - رويكرد اسالمي( اقتصاديفصلنامه پژوهشنامه 68 شامل تعداد كه سهام دادوستد ميزان: اند ازشود، عبارتهايي كه با بورس در نظر گرفته مي شاخص هاي روز كه شامل تعداد سهام دادوستد تناوب. شده است دادوستد سهام شده و ارزش دادوستد سهام سهام تعداد بازار كه شامل ميانگين بر تأثيرگذاري شده است و ميزان انجام هاي دادوستدهدادوستد و دفع هاي شاخص و هاها، مدلاي از روشخالصه. شركت است سهام جاري ارزش منتشر شده و ميانگين .است شده داده نشان ،1 جدول شماره در هاشركت بنديرتبه براي استفاده مورد هاشركت بنديرتبه براي استفاده مورد هاي شاخص و هامدل ها،روش از ايخالصه - 1 جدول روش شاخص/معيار حوزه سال سازمان/نويسنده رديف هرسال بورس اوراق بهادار تهران 1 پذيرفته شده در (ها تمام شركت )بورس اوراق بهادار تهران - ميزان فروش، حجم معامالت هاي بازرگاني سازمان 1385 پورعبدالعليآذر و 2 نيروي انساني، عمليات سازمان و مديريت TOPSIS ،SAW ، تاكسونومي كالسيك و تاكسونومي غيركالسيك 1386 نبوتي 3 پذيرفته شده در (ها تمام شركت )بورس اوراق بهادار تهران هاي داخلي و مالي، مشتري، فرآيند رشد و يادگيري كارت ارزيابي تاپسيس، + متوازن SAW 1387 زادهدربندي و عيوض 4 پذيرفته (هاي خودرو و توليد قطعات شركت )شده در بورس اوراق بهادار تهران تاپسيس و تاكسونومي وري اجزاي هزينه كلبهره 2007 البدوي و همكاران 5 پذيرفته شده در (ها تمام شركت )بورس اوراق بهادار تهران ، ضريب قدرت پرداختنسبت جاري، ...ارزش بازار، نسبت نفدينگي و PROMETHEE 2010 فسنگري و منتظر 6 پذيرفته شده در (ها تمام شركت )بورس اوراق بهادار تهران ها،هاي فروش، پروژه هاي سياست معيار - داران، و شاخصهاي حسابرسي، سهامگزارش و سهام شناور سود هر سهمهاي ي فازيخبرهسيستم 2000 ردمن و همكاران 7 كار هب كه الملليهاي بين شركت ديگر هايشركت خريد سهام .كردندمي مبادرت - شارپ، ترينر و جنسن هر ساله مجله فورچيون 8 فعاليت مختلف صنعتي و 53 شركت برتر 500( بازرگاني )آمريكايي - ميزان درآمد و سودآوري مؤسسه 9 Value Line هر ساله شركت مختلف در تمام 1700 صنايع در سطح جهان - عملكرد قيمت مؤسسه 10 Standard and Poor’s - اوراق بهادار ريسك تمام صنايع در سطح جهان هر ساله هاي حاضر در بورس آتن شركت 2008 ساماراس و همكاران 11 نسبت قيمت سهم به P/E شاخص شاخص ،ترپنجاه شركت فعال(درآمد )قيمت و بازده نقدي تحليل بنيادي هاي هلدينگ مالي تايوان شركت 2009 لو و لو 12 دارايي، حقوق صاحبان سهام، تعداد كاركنان، درآمد، سود، سود هر سهم و ارزش بازار هاتحليل پوششي داده در تايوان دالل سهامهاي شركت 2008 هو و اُه 13 كارمزدهزينه عملياتي، كارمندان، سود، دالالن، سود هر سهم، درآمد خالص هاتحليل پوششي داده 69 بورس اوراق بهادار تهران هاي بندي شركتارايه يك رويكرد جديد براي رتبه هاانتخاب مدل و شاخص -3 يك بر تنها بنديرتبه هاي روش بيشتر ، تأكيددهددست آمده از مرور ادبيات نشان مينتايج به انجام آنها عملكرد براساس هاشركت بنديرتبه كه شودمي موجب امر اين و است اصلي شاخص عملكرد نه است، اصلي شاخص يك براساس بنديرتبه هايي روش چنين حاصل عبارتي،به. 1نشود را شركت عملكرد هاداده پوششي تحليل روش اما است، متنوعي هاي شاخص برآيند كه آنها گيري استفاده در اندازههاي مورد بيشتر مدل همچنين .سنجدمي مختلف هاي برپايه شاخص هاي هاي رگرسيوني، روش دليل استفاده از توابع اقتصادي، مدلهاي كارايي و عملكرد، به متغير هاي مختلف پارامتريك، آماري بوده و همبستگي، تحليل واريانس، تابع جانشيني و آزمون ها را دارا هستند، اما بين نهادهها و شكاف ها و ستادههايي مانند محدوديت در انتخاب نهاده اشكال ها و ايرادها را ندارد، زيرا روشي رياضي و ناپارامتريك ها اين محدوديتروش تحليل پوششي داده ارزيابي سنجش و بر ها عالوهروش تحليل پوششي داده. 2بر فنون آماري نيستاست كه مبتني به داده ستاده نسبت از استفاده با كيكيتف طوربه نيز را افزايش آنها هايراه عملكرد، و كارايي . 3دهدمي سطوح ارايه تمام در را ورينحوه افزايش بهره و كندمي پيشنهاد جداگانه سطح هر براي اندازه الزم و كافي وجود ندارد و در بازار بورس تهران، التزام قانوني براي افشاي اطالعات به طور همان. 4ايجاد انگيزش و التزام قانوني را ندارندهاي عملياتي آن به حد كافي قدرت سازوكار هاي عملياتي بورس اوراق بهادار تهران، اعالم يكي از سازوكاركه در قسمت پيشين اشاره شد، شود و بندي براساس شاخصي خاص انجام ميشركت برتر است، اما اين رتبه 50فهرست اسامي ها برترين شركت را مشخص ت كه اين روشها آن اسگونه روشايراد اساسي وارد بر اين عبارتي، فهرست به. 5ترين شركت را تعيين كندترين يا حجيمكند، بلكه ممكن است بزرگ نمي .هاي برتر نيست دهنده شركتارايه شده از سوي بورس نشان دست آمده از مرور ادبيات براي طراحي روش تحقيق به اين صوررت است كه رهنمود نتايج به صورت هاي پذيرفته شده در بورس اوراق بهادار تهران، عملكرد آنها بايد به بندي شركترتبهاي بر .1383قدرتيان و انواري رستمي، -1 .119 -138، صص 1386آذر و همكاران، - 2 .1386ميرغفوري و شفيعي رودپشتي، -3 .25 -43، صص 1385لو، انواري رستمي و ختن -4 .1383قدرتيان و انواري رستمي، -5 46دوازدهم شماره سال )ايراني - رويكرد اسالمي( اقتصاديفصلنامه پژوهشنامه 70 از طرفي، مدل . هاي متنوع مورد ارزيابي قرار داده شود و با توجه به معيارها و شاخص جانبههمه از آنجا كه . شدها را داشته باارايه شده بايد قابليت پوشش دادن شرايط عدم قطعيت و ابهام در داده صورت تركيبي با منطق فازي ابزاري براي مدل كردن عدم قطعيت و ابهام است، از اين منطق به . مدل ارزيابي عملكرد استفاده خواهد شد بنابراين، با توجه به مطالب بيان شده، مدل پيشنهادي اين پژوهش براي ارزيابي عملكرد . هاي فازي استدار تهران، مدل تحليل پوششي دادههاي پذيرفته شده در بورس اوراق بها شركت هاي مربوط به يك مقطع زماني خاص از مزاياي عالوه بر مزاياي ساده شده، استفاده نكردن از داده اين امتياز به اين معناست كه در صورت استفاده از . هاي فازي استديگر مدل تحليل پوششي داده هاي ابليت ارزيابي عملكرد را با استفاده از طيفي از دادههاي فازي، قمدل تحليل پوششي داده هاي مربوط به چند مقطع زماني است كننده دادهها، بياناين طيف از داده. ها داردمربوط به شاخص . شودها ميتر شدن دادهكه موجب واقعي ها در نظر شاخص اند كه در ادبيات بيش از سايرهايي در نظر گرفته شده در اين پژوهش، شاخص ها در نظر هاي روش تحليل پوششي داده ها و ستادهعنوان نهادهشاخصي كه به 10. گرفته شده است ريسك (سود هر سهم، نسبت جاري، نسبت آني، شاخص بتا : اند ازاند، عبارتگرفته شده قيمت ها، نسبت ، جمع دارايي)ريسك غيرسيستماتيك(، سود خالص، شاخص سيگما )سيستماتيك .به سود، رشد فروش، نسبت بازده حقوق صاحبان سهام هاآوري دادهروش جمع -4 گذاري در بورس و همچنين توان از نتايج تحقيق حاضر براي ارايه رهنمود سرمايهاز آنجا كه مي توان آن را از نظر هدف و ماهيت تحقيق، در رده ها استفاده كرد، ميجانبه شركتارزيابي همه شمار آورد، زيرا هدف تحقيقات كاربردي، توسعه دانش كاربردي در يك ت كاربردي بهتحقيقا به عبارت ديگر، تحقيقات كاربردي به سمت كاربرد عملي دانش هدايت . زمينه خاص است در ها، تحقيق حاضر جزء تحقيقات توصيفي است، زيرا از نظر نحوه گردآوري داده. شود مي متغيرها توسط طور معمول بهگيرند و شوند و مورد مشاهده قرار مي اب ميمتغيرها انتختحقيق توصيفي .آورند وجود نميها شرايطي را به شوند و براي وقوع پديده نمي اريكمحقق دست 71 بورس اوراق بهادار تهران هاي بندي شركتارايه يك رويكرد جديد براي رتبه هاي مورد نياز براي محاسبه هاي مورد نياز تحقيق كه همان داده پس از شناسايي كامل داده هاي با توجه به حجم نمونه، داده. هاي اوليه پرداختيم آوري دادهجمعگانه هستند، به ده هاي شاخص چهارم ماهه سه( تهران بهادار بورس اوراق ترفعال شركت 50متغير مورد مطالعه، براي 10اوليه ، در چند مرحله مراجعه به سازمان بورس اوراق بهادار مركزي تهران، سازمان بورس و )1388سال ش مديريت پژوهش، توسعه و مطالعات اسالمي، دانشكده مديريت دانشگاه تهران، اوراق بهادار بخ هاي داده. آوري شدجمع» آورد نوينره«افزار دانشكده اقتصاد دانشگاه تهران و استفاده از نرم هاي مربوط به گزارش ترازنامه، گزارش صورت سود و زيان، گزارش آوري شده شامل دادهجمع آوري هاي جمع همچنين اعتبار داده. و گزارش ريسك و بازدهي بود EPS، روند هاي مالي نسبت . شده با نظر كارشناسان و خبرگان مسايل مالي و بورس مورد تأييد قرار گرفت تهران در بهادار بورس اوراق ترفعال شركت 50از آنجاكه هدف اين تحقيق، ارزيابي عملكرد ماهه 36هاي وري شده در اين مرحله از تحقيق، دادهآهاي جمع شرايط فازي است، داده .هاي فازي تبديل خواهند شد ها، به دادهپردازش دادههاي يادشده است كه در مرحله پيش شركت بورس هستند كه سه ماهه هر سال توسط سازمان بورس برتر شركت 50 جامعه آماري اين تحقيق شركت برتر بورس اوراق بهادار سه 50اري مربوط به شوند؛ در اين پژوهش، جامعه آممعرفي مي .است 1388ماه آخر سال هاي فازي بندي تحليل پوششي دادهمدل رتبه -5 ارايه 0- 1يك مقدار كارايي بين DMU، به هر DEAاستاندارد دانيم، مدلطور كه ميهمان بندي كرد و را رتبه) 1(هاي غير كارا DMUتوان ميزان كارايي، تنها مي براساس. كند مي اين تحقيق از روش پيشنهادي . توان مشخص ساخترا نمي) 1(هاي كارا DMUتفاوت بين ها در شرايط فازي ارايه شده است، DMUبندي كامل كه براي رتبه) 2002( 1ساعتي و همكاران .نشان داده شده است) 1( توسط آنها در مدلمدل ارايه شده . استفاده خواهد كرد 1- Saati etal, 2002, PP.255-267. 46دوازدهم شماره سال )ايراني - رويكرد اسالمي( اقتصاديفصلنامه پژوهشنامه 72 j;0 j r; n 1j ~ rjyj ~ rpy i; n 1j ~ ijxj ~ ipx.t.s pZmin         .دهداعداد فازي را نشان مي "~"در اين مدل، x,x,x(x( حال اگر براساس تعريف اعداد فازي مثلثي، قرار دهيم uijmijlijij ~  و )y,y,y(y uij m ij l ijij ~ ، بازنويسي كرد) 2(صورت مدل توان بهرا مي) 1(مدل . كه (ها ها و سطح باالتر وروديتر وروديسطح پايين DMUها، براي هر DMUبندي منظور رتبهبه اگر در اين حالت، . شودمقايسه مي 1با بخش داخلي مرز كارايي) است DMUبهترين بخش يك بيش از يك خواهد شد و DMU، در خارج مرز كارايي قرار گيرد، كارايي DMUبهترين بخش 2.شوندبندي ميرتبه هاDMUبا استفاده از اين ايده، j;0 j r;)y)1(y,y)1(y(j)y)1(y,y)1(y( i;)x)1(x,x)1(x(j)x)1(x,x)1(x(.t.s pZmin n 1j u rj m ij l rj m rj u rp m ip l rp m rp n 1j u ij m ij l ij m ij u ip m ip l ip m ip         . نشان داد 3توان در مدل ها را ميDMUبندي با توجه به اين توضيحات، مدل نهايي رتبه j;0 j r;)y)1(y(j)y)1(y( i;)x)1(x(j)x)1(x(.t.s Zmin n 1j u rp m rp l rp m rp n 1j u ip m ip l ip m ip         1- Efficiency Frontier 2- Saati, Ibid. )1( )2( )3( 73 بورس اوراق بهادار تهران هاي بندي شركتارايه يك رويكرد جديد براي رتبه مطالعه موردي -6 بورس اوراق بهادار تهران را كه فهرست آنها از سوي سازمان ترشركت فعال 50اين تحقيق، شركت كه 50فهرست اين . دهدشود، مورد بررسي قرار ميبورس اوراق بهادار تهران اعالم مي .ارايه شده است 2بوده، در جدول شماره 1388مربوط به سه ماهه چهارم سال 1388 سال چهارم ماهه سه - تهران بهادار بورس اوراق ترفعال شركت 50 فهرست - 2 جدول شركت شماره شركت شماره اميد گذاريسرمايه 26 سايپا 1 خودرو پارس 27 گذاري بهمنسرمايه 2 ساختمان ايران گذاريسرمايه 28 گذاري بوعليسرمايه 3 سبحان دارويي گروه 29 ملت بانك 4 )پتروشيمي شازند(* اراك پتروشيمي 30 ملي ايران گذاريسرمايه 5 خودرو ديزل ايران 31 كارآفرين بانك 6 ايران مس صنايع ملي 32 ترانسفو ايران 7 حيانجابربن داروسازي 33 )مپنا(* هاي نيروگاهيمديريت پروژه 8 سايپا رايان ليزينگ 34 )هلدينگ(بهشهر صنايع توسعه 9 اصفهان مباركه فوالد 35 سينا بانك 10 البرز دارو 36 آذينسايپا 11 )توليدي بهمن(* گروه بهمن 37 )هلدينگ(گذاري غدير سرمايه 12 خوزستان و فارس سيمان 38 )هلدينگ(گذاري توكافوالد سرمايه 13 شاهد گذاريسرمايه 39 ليزينگ ايران 14 نوسازي و ساختمان تهران 40 صنايع جوشكاب يزد 15 ومعدن صنعت گذاريسرمايه 41 پتروشيمي گذاريسرمايه 16 ايران روي معادن توسعه 42 توسعه معادن و فلزات 17 )صنعتي بهشهر(* گروه صنايع بهشهر ايران 43 )هلدينگ(رنا گذاريسرمايه 18 تجارت بانك 44 ليزينگ صنعت و معدن 19 شهيد باهنر مس 45 معدني و صنعتي چادرملو 20 نيروكلر 46 اراك سازيماشين 21 سيمان سپاهان 47 گذاري صندوق بازنشستگيسرمايه 22 آذرآب صنايع 48 لعابيران 23 كوثر داروسازي 49 بانك پارسيان 24 لقمان دارويي 50 ايران اتومبيل قطعات 25 .آورد نوين استافزار رهها در نرماسامي داخل پرانتز نام جديد شركت* 46دوازدهم شماره سال )ايراني - رويكرد اسالمي( اقتصاديفصلنامه پژوهشنامه 74 FDEAهاي ورودي و خروجي مدل شاخص -7 شاخص شناسايي شد 47ها، تعداد بندي و ارزيابي عملكرد شركتبا مرور ادبيات مربوط به رتبه عنوان شاخصهايي است كه در ادبيات بيش از يك بار به دهنده شاخص، نشان3جدول شماره هاي ورودي و خروجي از نظرات شود در انتخاب شاخصيادآوري مي. اندمؤثر در نظر گرفته شده .ناسان بورس و استادان دانشگاه نيز استفاده شده استكارش هاي برتر به ترتيب تكرار شاخص - 3 جدول تكرار نماد شاخص رديف Earnings Per Share EPS 8 سود هر سهم 1 (Liquidity Ratio (Current Ratio ) نسبت جاري(هاي نقدينگي نسبت 2 CR 7 Liquidity Ratio (Quick Ratio) QR 6 )نسبت آني( هاي نقدينگي نسبت 3 Beta Index (Systematic Risk) β 6 )ريسك سيستماتيك(شاخص بتا 4 Net Income NI 5 سود خالص 5 6 ريسك (شاخص سيگما )غيرسيستماتيك Sigma Index (Unsystematic Risk) σ 4 Total Assets TA 4 هاجمع دارايي 7 Price/Earnings Ratio P/E 4 سود به قيمت نسبت 8 Sales Growth SG 3 رشد فروش 9 Return On Owners’ Equity ROE 2 نسبت بازده حقوق صاحبان سهام 10 بندي بيشتر به آنها اشاره شده است، در هايي كه در زمينه ارزيابي عملكرد و رتبه با توجه به شاخص هاي بندي شركتهاي مؤثر در ارزيابي و رتبه عنوان شاخصهاي زير را به اين پژوهش شاخص .گيريمپذيرفته شده در بورس اوراق بهادار تهران در نظر مي )EPS( سود هر سهم - 1 گذاران و تحليلگران مالي مورد توجه هاي مهمي است كه توسط سرمايه سود هر سهم يكي از معيار .گيردبسياري قرار مي سود هر سهم سود خالص سود تقسيمي سهام ممتاز خالص )4( متوسط تعداد سهام در دست صاحبان سهام 75 بورس اوراق بهادار تهران هاي بندي شركتارايه يك رويكرد جديد براي رتبه )CR(نسبت جاري - 2 اين نسبت . هاي جاري هاي جاري به بدهي نسبت جاري عبارت است از حاصل تقسيم دارايي اين نسبت . هاي جاري مؤسسه است هاي جاري از محل دارايي توانايي بازپرداخت بدهي دهنده نشان :قابل محاسبه است از رابطه زير نسبت جاري هاي دارايي جاري هاي بدهي جاري )QR( نسبت آني - 3 به )دريافتني وجوه نقد، موجودي بانك و اسنــاد(تقسيم نقدترين اقالم دارايي جاري از آنينسبت .آيددست ميههاي جــاري ب بدهي نسبت آني هاي دارايي جاري موجودي كاال هاپرداخت پيش هاي بدهي جاري )NI(سود خالص - 4 .آيددست ميسود خالص هر مؤسسه از روابط زير به سود خالص = ماليات و ديگر كسور قانوني ـ سود قبل از پرداخت ماليات ) 7( سود قبل از پرداخت ماليات = هاي عملياتي ـ سود ناخالصهزينه )8( سود ناخالص = صوالت و خدماتبهاي تمام شده كاالي فروش رفته ـ فروش مح )9( )TA(ها جمع دارايي - 5 .آيد دست ميهاي غيرجاري به هاي جاري و دارايي ها از جمع داراييجمع دارايي هاجمع دارايي =هاي جاري دارايي +هاي غيرجاري دارايي )10( )P/E(سود به قيمت نسبت - 6 انداز آينده پردازد و چشم مي گذار براي سهم هدف اين نسبت بيان رابطه قيمتي كه يك سرمايه .بيني شده آن است شركت و سود پيش )11 ( قيمت سهم )SG(رشد فروش - 7 :درصد رشد فروش از فرمول زير قابل محاسبه است درصد رشد فروش )12( بينيپيش فروش فروش واقعي فروش واقعي 100 )5( )6( 46دوازدهم شماره سال )ايراني - رويكرد اسالمي( اقتصاديفصلنامه پژوهشنامه 76 )ROE(نسبت بازده حقوق صاحبان سهام - 8 گذاري شده توسط وابسته به مبلغ سرمايه ،كند كه بازده حاصل از درآمدكيد ميأتاين نسبت معيار سنجش موفقيت يا عدم موفقيت مديريت شركت در همچنين اين نسبت، . است نسهامدارا .ن استداراسهام سوددستيابي به افزايش )β) (ريسك سيستماتيك(ضريب بتا - 9 در صورتي كه ضريب بتا براي يك دارايي بيشتر . است سيستماتيكضريب بتا معياري براي محاسبه ريسك آن سهم بيشتر از نوسانات بازار خواهد بود و به آن دارايي پرريسك گفته بازدهي از يك باشد، نوسانات .شود مي )13( بازدهي بازار, بازرهي سهم بازدهي بازار )σ) (ريسك غير سيستماتيك(ضريب سيگما -10 .كردتوان به دو دسته كلي ريسك سيستماتيك و ريسك غيرسيستماتيك تقسيم كل ريسك بازار را مي از جمله نوع محصول ،ات خاص شركتيريسك غيرسيستماتيك ريسكي است كه ناشي از خصوص .استداران عمده و غيره ساختار سرمايه سهام )14 ( 2 0 1 ( ) 1 n i i i r E r n       ir بازده سهام در دوره i اين ضريب ميزان انحراف معيار بازده پنج سال است. ام است. 10تر بورس اوراق بهادار تهران است، شركت فعال 50در راستاي هدف اصلي اين تحقيق كه ارزيابي عملكرد FDEA ،10شركت از طريق روش 50در ادامه، براي ارزيابي عمكرد اين . شاخص مختلف در نظر گرفته شد مبناي اين تفكيك توجه به مقاالت . كنيمشاخص شناسايي شده را به دو گروه ورودي و خروجي تفكيك مي شود، جزء آل محسوب ميكه كمينه آن حالت ايده موجود در ادبيات و اين موضوع است كه شاخصي هاي خروجي آيد، جزء شاخصحساب ميآل بههاي ورودي و شاخصي كه حداكثر بودن آن حالت ايده شاخص .در نظر گرفته شده است 77 بورس اوراق بهادار تهران هاي بندي شركتارايه يك رويكرد جديد براي رتبه FDEAهاي ورودي و خروجي در مدل شاخص - 1 جدول نماد شاخص خروجي نماد شاخص ورودي EPS( O1(درآمد هر سهم I1 شاخص بتا O2 رشد فروش I2 شاخص سيگما O3 نسبت جاري P/E( I3(نسبت قيمت به درآمد O4 نسبت آني I4 نسبت حقوق صاحبان سهام O5 سود خالص I5 جمع دارايي هاپردازش دادهپيش -8 ماهه 36هاي سازي دادهها به فازيپردازش دادههاي اوليه مورد نياز، در فرآيند پيش پس از استخراج داده سازي با توجه به نظر خبرگان شود كه فرآيند فازييادآوري مي. گانه پرداختيمهاي ده براي متغير 50ماهه مربوط به 36هاي سازي دادهدر فازي. فعال در بورس و مديران مسايل مالي انجام شده است :ه استاز سه روش زير استفاده شد شركت عنوان حد برابر ميانگين به 2/1عنوان حد وسط، به 1388هاي سه ماهه سال ميانگين داده - . 1عنوان حد پايين عدد فازي در نظر گرفته شده استبرابر ميانگين به 8/0باال و ),,(): 2005( 2فرمول با كلي -   كه در آن و ميانگين و ، هاي طي دوره 1388هاي نسبت آني و نسبت جاري در سال انحراف استاندارد شاخص . هاي مورد بررسي است ماهه شركتماهه و دوازدهماهه، نهماهه، ششسه  ،)(minكه در آن 15استفاده از فرمول - ix و)(max ix به ترتيب برابر ميانگين ، ماهه، هاي سه طي دوره 1388ها در سال حداقل و حداكثر شاخص جمع كل داراييو . هاي مورد بررسي است ماهه شركتماهه و دوازدهنه ماهه، شش )15( ))(max,),(min( ii xx   گانه، براي 10هاي هاي مطلوب مربوط به متغير دادههاي اوليه، در پايان فرآيند تجزيه و تحليل داده .شركت مورد مطالعه، استخراج شد 50 1- Azadeh etal, 2010,PP.1-15. 2- Buckley, 2005. 46دوازدهم شماره سال )ايراني - رويكرد اسالمي( اقتصاديفصلنامه پژوهشنامه 78 نتايج و پيشنهادها - 9 ها در فهرست دهنده امتياز شركتكه نشان) 3مدل (بندي نتايج عددي مربوط به اجراي مدل رتبه همچنين در جدول . نشان داده شده است α، در سطوح مختلف 6بندي بوده، در جدول شماره رتبه بندي ارايه شده از طريق روش بندي ارايه شده از سوي سازمان بورس و رتبه، رتبه5شماره .اندپيشنهادي، با يكديگر مقايسه شده FDEAبندي سازمان بورس و روش مقايسه رتبه - 5 جدول شركت رتبه شركت رتبه گزارش بورس FDEAروش (α = 0.5) گزارش بورس FDEAروش (α = 0.5) 28 26 اميد گذاري سرمايه 17 1 سايپا 39 27 خودرو پارس 1 2 گذاري بهمنسرمايه 38 28 ساختمان ايران گذاريسرمايه 9 3 گذاري بوعليسرمايه 4 29 سبحان دارويي گروه 11 4 ملت بانك 40 30 )پتروشيمي شازند( پتروشيمي اراك 36 5 ملي ايران گذاريسرمايه 50 31 خودرو ديزل ايران 8 6 كارآفرين بانك 6 32 ايران مس ملي صنايع 10 7 ترانسفو ايران هاي نيروگاهيمديريت پروژه )مپنا( 26 33 حيانجابربن داروسازي 18 8 41 34 سايپا رايان ليزينگ 48 9 )هلدينگ(بهشهر صنايع توسعه 20 35 اصفهان مباركه فوالد 13 10 سينا بانك 16 36 البرز دارو 35 11 سايپا آذين 2 37 )بهمنتوليدي( گروه بهمن 19 12 )هلدينگ(گذاري غدير سرمايه گذاري توكافوالد سرمايه )هلدينگ( 30 38 خوزستان و فارس سيمان 15 13 3 39 شاهد گذاريسرمايه 32 14 ليزينگ ايران 34 40 نوسازي و ساختمان تهران 5 15 صنايع جوشكاب يزد 49 41 و معدن صنعت گذاريسرمايه 47 16 پتروشيمي گذاريسرمايه 43 42 ايران روي معادن توسعه 25 17 توسعه معادن و فلزات 33 43 )صنعتي بهشهر( گروه صنايع بهشهر ايران 42 18 )هلدينگ(رنا گذاريسرمايه 12 44 تجارت بانك 7 19 ليزينگ صنعت و معدن 45 45 شهيد باهنر مس 37 20 معدني و صنعتي چادرملو 23 46 نيروكلر 44 21 اراك سازيماشي گذاري صندوق سرمايه بازنشستگي 46 47 سيمان سپاهان 29 22 21 48 آذرآب صنايع 24 23 لعابيران 31 49 كوثر داروسازي 14 24 بانك پارسيان 27 50 لقمان دارويي 22 25 ايران اتومبيل قطعات 79 بورس اوراق بهادار تهران هاي بندي شركتارايه يك رويكرد جديد براي رتبه خوبي نشان داده نشده است؛ براي ها بهبندي شركت، رتبهα= 1شود در سطح طور كه مشاهده ميهمان اند، اما در خود تخصيص دادهرا به 1رتبه ) ايران ترانسفوو البرز دارو، آذرابسايپا، (شركت اول 4مثال، هر توان معيار را نمي α= 1دست آمده در سطح تايج بهرو، نازاين. شودساير سطوح چنين اتفاقي مشاهده نمي هاي فازي، نتايج مربوط به روش تحليل پوششي داده. ها قرار دادبندي و سنجش كارايي شركترتبه كه دوره مورد بررسي اين تحقيق است، شركت ايران خودرو 1388رسد؛ زيرا در سال نظر مي تر بهمنطقي ، شود ايران محسوب ميفناوري نوعي نماد صنعت و خاورميانه و بهترين شركت خودروسازي بزرگكه مجمع ايران خودرو برگزار شد و ، 1388آمد، به طوري كه در شهريور شركتي ورشكسته به شمار مي ايران خودرو مجمعبايد گفت كه . ميليارد توماني اين گروه صنعتي اعتراف كردند 123به زيان آن مديران بودرسيده ،يعني نزديك به قيمت اسمي ،تومان 138كه ارزش سهام ايران خودرو به دشدر حالي برگزار ميليارد 4در حالي كه معادل ،داد و مروري بر شرايط سهام اين شركت در بورس اوراق بهادار نشان مي خبر (نداشت وجود ي براي خريد آن مطرح بود، هيچ تقاضاي واگذاري سهام به كارگزاران عرضهتومان به . گرددنيز بازمي 1387، يعني سال 1388وضعيت آشفته شركت ايران خودرو، به سال قبل از ). 1آنالين براساس صورت وضعيت ارسالي ايران خودرو به سازمان ، )2بورس نيوز(پايگاه خبري بورس گزارش بوده صفر ،1387اسفند 30داران براي سال مالي منتهي به حسابرسي، سود قابل تقسيم اين شركت به سهام .است گذاري، بسيار منطبق با شرايط هاي سرمايه شود، چگونگي فعاليت شركتطور كه مشاهده ميهمان هاي وكار شركتنحوه كسب. وكار استگذاري و كسبكنوني جامعه و وضعيت بازار سرمايه بازار رقابتي سازي درست، ظرفيت موفقيت دراي است كه در صورت پيادهگونهگذاري بهسرمايه در راستاي نتايج . كنددست آمده از اين تحقيق نيز اين واقعيت را تأييد مينتايج به. سرمايه را دارد شركت برتر دارد؛ شركت 50گذاري بهمن رتبه اول را در فهرست دست آمده، شركت سرمايهبه ريد وهاي خ زمينهاش در ها است كه براساس اساسنامهگذاري بهمن يكي از اين شركتسرمايه ا، ههسسؤها، مانواع شركت سرمايه سيس، تشكيل و مشاركت درأفروش سهام، اوراق مشاركت، ت خصوصي عرضه عمومي و نويسي وها از طريق پذيرهمين سرمايه شركتأت ها،و پروژه هاطرح و خارجي، ارايه خدمات حقوقي معتبر داخلي و اوراق مشاركت به اشخاص حقيقي و سهام و ها، استفاده از تسهيالت مالي واعتباري اداره امور شركت ها وگذاريزمينه سرمايه مشاوره در 1- http: www.khabaronline.ir/news-15791.aspx 2- http: www.boursenews.ir/fa/pages/?cid=21692 46دوازدهم شماره سال )ايراني - رويكرد اسالمي( اقتصاديفصلنامه پژوهشنامه 80 تمام اخذ نمايندگي از اعتباري داخلي وخارجي، مالي و ياههسسؤبيمه، م هايها، شركت بانك انتشار هاي دولتي،شركت ها وها، سازمانوزارتخانهتمام حقوقي، عقد قرارداد با اشخاص حقيقي و . ، فعاليت داردخارج از كشور داخل و اعطاي نمايندگي، ايجاد شعبه در اوراق مشاركت، اخذ و 2شود كه در بين طيف وسيعي از صنايع مختلف موجود، شركت برتر مشاهده مي 10با نگاهي به جايگاه از 2عبارتي، هاي اول و نهم قرار دارند و به گذاري بهمن و بوعلي در رتبهسرمايه شركت هاي اين امر مؤيد آن است كه شركت. اندگذاري كردههاي سرمايه اول را نصيب شركتجايگاه 10 شركت برتر 50توان در فهرست همين موضوع را مي. تري هستندهاي موفق ، شركتگذاريسرمايه . شركت است 12گذاري حاضر در اين فهرست، هاي سرمايه تعداد شركت. نيز مشاهده كرد .گذاري هستندشركت برتر، از نوع سرمايه 50هاي موجود در فهرست درصد شركت24 عبارتي، به بندي شركت برتر بورس را رتبه 50، 1388ماهه چهارم سال هاي سه اين پژوهش با استفاده از داده ماه سال 12در ها براي مثال داده(يافت ها گسترش ميآوري دادهكرده است كه اگر طيف جمع تري را توانست نتايج جامع، مي)گرفتشد و نتايج آن مورد ارزيابي قرار ميآوري ميجمع 1388 هاي در راستاي توسعه تحقيقات و پژوهش). هاآوري دادهمحدوديت در جمع(منعكس كند :شرح زير است توانند در اين زمينه انجام شوند بهراستا با اين تحقيق، پيشنهادهاي آتي كه مي هم .ها و معيارهاتوسعه شاخص - هايي كه مسأله ارزيابي عملكرد و كارگيري روش پيشنهادي در ساير زمينهبه - .مطرح است بندي رتبه .گذارياستفاده از اين روش براي ايجاد يك سيستم پشتيبان تصميم براي سرمايه - .مربوط ها به تفكيك صنعتبندي شركتاستفاده از اين روش براي رتبه - دست آمده از اين تحقيق، طراحي رويكردي نوين و جامع براي ارزيابي عالوه بر نتايج عددي به تواند در آينده مورد استفاده هاي مهم اين پژوهش است كه مي ها، از دستاوردعملكرد شركت .هاي مختلف قرار گيرد محققان زمينه بنديشركت فعال تر بورس اوراق بهادار تهران در رتبه 50امتياز -6 جدول 0 =α 0.25 =α 0.50 =α 0.75 =α 1 =α امتياز شركت امتياز شركت امتياز شركت امتياز شركت امتياز شركت امتياز شركت امتياز شركت امتياز شركت امتياز شركت امتياز شركت 1 2.414 26 0.992 1 1.926 26 0.812 1 1.543 26 0.667 1 1.241 26 0.548 1 1 26 0.451 2 2.25 27 2.781 2 1.83 27 2.135 2 1.494 27 1.642 2 1.222 27 1.276 2 1 27 1 3 2.276 28 1.083 3 1.934 28 0.895 3 1.566 28 0.773 3 1.256 28 0.667 3 1 28 0.559 4 3.545 29 1.132 4 2.53 29 0.962 4 1.839 29 0.806 4 1.352 29 0.675 4 1 29 0.564 5 0.515 30 0.882 5 0.419 30 0.735 5 0.352 30 0.613 5 0.307 30 0.521 5 0.275 30 0.462 6 2.722 31 2.769 6 2.121 31 2.039 6 1.654 31 1.519 6 1.289 31 1.138 6 1 31 0.875 7 2.472 32 2.019 7 2.074 32 1.605 7 1.74 32 1.268 7 1.378 32 0.99 7 1 32 0.814 8 2.842 33 1.498 8 2.22 33 1.101 8 1.715 33 0.823 8 1.316 33 0.668 8 1 33 0.543 9 3.572 34 0.996 9 2.711 34 0.823 9 1.955 34 0.68 9 1.407 34 0.568 9 1 34 0.482 10 2.437 35 1.552 10 2.049 35 1.293 10 1.769 35 1.078 10 1.385 35 0.897 10 1 35 0.744 11 1.193 36 2.997 11 0.982 36 2.674 11 0.801 36 2.166 11 0.654 36 1.505 11 0.543 36 1 12 1.176 37 1.398 12 0.956 37 1.173 12 0.793 37 0.981 12 0.657 37 0.819 12 0.544 37 0.681 13 0.978 38 2.284 13 0.775 38 1.851 13 0.616 38 1.505 13 0.507 38 1.226 13 0.437 38 1 14 9.405 39 2.35 14 5.078 39 1.858 14 2.946 39 1.491 14 1.739 39 1.226 14 1 39 1 15 2.893 40 2.091 15 2.469 40 1.742 15 1.994 40 1.452 15 1.44 40 1.207 15 1 40 1 16 1.123 41 1.398 16 0.954 41 1.165 16 0.81 41 0.99 16 0.688 41 0.812 16 0.595 41 0.652 17 1.758 42 1.017 17 1.452 42 0.931 17 1.199 42 0.782 17 0.989 42 0.647 17 0.816 42 0.534 18 1.677 43 3.679 18 1.44 43 2.296 18 1.239 43 1.967 18 0.987 43 1.417 18 0.799 43 1 19 2.867 44 1.161 19 2.51 44 0.934 19 2.119 44 0.757 19 1.6 44 0.617 19 1 44 0.506 20 1.534 45 2.509 20 1.258 45 1.948 20 1.038 45 1.526 20 0.872 45 1.229 20 0.735 45 1 21 1.161 46 0.997 21 1.017 46 0.858 21 0.851 46 0.735 21 0.716 46 0.627 21 0.597 46 0.533 22 1.769 47 1.137 22 1.417 47 0.923 22 1.161 47 0.76 22 1.039 47 0.627 22 0.893 47 0.519 23 1.34 48 3.784 23 1.227 48 2.746 23 1.087 48 1.972 23 0.828 48 1.407 23 0.651 48 1 24 3.601 49 1.175 24 5.142 49 1.065 24 4.853 49 0.867 24 2.795 49 0.708 24 1 49 0.579 25 3.858 50 2.235 25 2.274 50 1.822 25 1.945 50 1.489 25 1.395 50 1.22 25 1 50 1 46دوازدهم شماره سال )ايراني - رويكرد اسالمي( اقتصاديفصلنامه پژوهشنامه 82 منابع فارسي -الف گيري كارايي نسبي ، اندازه)1386بهار (آذر، عادل، علي اصغرانواري رستمي و محمدرضا رستمي هاي شاخص(ها هاي حاضر در بورس اوراق بهادار با رويكرد تحليل پوششي داده شركت . 50 ، شماره14حسابرسي، سال و حسابداري هاي ، بررسي)تكنولوژي اطالعات ها با رويكرد هاي بازرگاني استان ارزيابي سازمان، )1385( پوراميرحسين عبدالعلي، عادل و آذر MADM ،39 شماره، زرگانيمجله پژوهشنامه با. هاي بندي شركتاي رتبه، بررسي مقايسه)1385 بهار(لو انواري رستمي، علي اصغر و محسن ختن هاي هاي بورس اوراق بهادار تهران، بررسي هاي سودآوري و شاخص برتر براساس نسبت .43 حسابرسي، شماره و حسابداري بورس در حاضر هاي شركت بندي، رتبه)1387 اسفند و بهمن( زادهعلي عيوض و شكوفه دربندي، كل هزينه اجزاي وريبهره هاي شاخص برحسب آن، قطعات توليد و خودرو گروه در فعال و .4شماره وابسته، صنايع و خودرو مهندسي ماهنامه ،1382 تا 1386 هاي سال طي توليد جامع ارزيابي عملكرد و ، طراحي مدل )1383(كاشان، قدرتيان و علي اصغرانواري رستمي .المللي مديريتدومين كنفرانس بينها، بندي شركت رتبه دانشگاهي هايكتابخانه بندي، رتبه)1386(اهللا و ميثم شفيعي رودپشتي ميرغفوري، سيدحبيب :مورد(بردا و هاداده پوششي تحليل هايتكنيك از با استفاده عملكرد براساس سطح .3، شماره 10اطالع رساني، جلد و ، فصلنامه كتابداري)يزد دانشگاه هاي كتابخانه هاي گيري با معيارهاي تصميم بندي صنايع ايران براساس تكنيك، رتبه)1386(نبوتي، حجت .عيصنا يمهندس يالملل نيكنفرانس ب نيپنجمچندگانه، رسانيهاي اطالعو پايگاه التين - ب Albadvi, A., S. K Chaharsooghi, and A. Esfahanipour (2007), Decision Making in Stock Trading: An Application of PROMETHEE. European Journal of Operational Research, Vol.177, NO.2. Azadeh, A., Asadzadeh, Bukhari S. M., A., Izadbakhsh, H. R (2010), An Integrated Fuzzy DEA Algorithm for Efficiency Assessment and Optimization of Wireless Communication Sectors With Ambiguous Data, The International Journal of Advanced Manufacturing Technology, Vol.52, NO.(5-8),: 805-819. 83 …شده در بورس اوراق بهادار پذيرفته هايبررسي عوامل مؤثر بر ساختار مالي شركت Buckley J (2005), Simulating Fuzzy Systems (Studies in Fuzziness and Soft Computing), Springer, Berlin. Redman, A. L, Gullett N. S., and Manakyan H (2000), The Performance of Global and International Mutual Funds, Journal of Financial and Strategic Decisions, Vol.13, NO.1. Fasanghari, M., Montazer, G. A (2010), Design and Implementation of Fuzzy Expert System for Tehran Stock Exchange Portfolio Recommendation, Expert Systems with Applications, Vol. 37, NO. 9. Ho, C. T.B, Oh, K. B (2008), Measuring Online Stock Broking Performance, Industrial Management & Data Systems, Vol.108, NO.7. Lo, S. F, Lu, W. M (2009), An Integrated Performance Evaluation of Financial Holding Companies in Taiwan, European Journal of Operational Research, Vol.198, NO.1. Samaras, G. D , Matsatsinis, N. F., Zopounidis, C (2008), A Multicriteria DSS for Stock Evaluation Using Fundamental Analysis. European Journal of Operational Research, Vol.187, NO. 3. Saati, S. M., Memariani A., and Jahanshahloo G. R (2002), Efficiency Analysis and Ranking of DMUs with Fuzzy Data, Fuzzy Optimization and Decision Making, Vol.1, NO.3. http:// www.iranbourse.com http://www.boursenews.ir http://www.khabaronline.ir http://money.cnn.com/magazines/fortune/fortune500/ http://www.valueline.com/ http://www.standardandpoors.com/home/en/us 
work_t5napb5vujdv7i5alyacc3dtoy	----	PII: 0925-5273(93)90112-X International Journal of Production Economics, 30-31 (1993) 453-464 Elsevier 453 An expert system for a local planning environment G.J. Meester Laboratory of Production Management and Logistics, Department of Mechanical Engineering, University of Twente, P.O. Box 217, 7500 AE Enschede. The Netherlands Abstract In this paper, we discuss the design of an Expert System (ES) that supports decision making in a Local Plannmg System (LPS) environment. The LPS provides the hnk between a high level factory planning system (rough cut capacity planning and material coordination) and the actual execution of Jobs on the shopfloor. by specifymg a detailed workplan. It is divided m two hterarchtcal layers: planmng and scheduling. At each level, a set of different algorithms and heuristics is available to anticipate different situations. The Expert System (which is a part of the LPS) supports decision making at each of the two LPS layers by evaluatmg the planning and scheduling conditions and, based on this evaluation, advising the use of a specific algorithm, and evaluatmg the results of using the proposed algorithm. The Expert System is rule-based while knowledge (structure) and data are separated (which makes the ES more flexible m terms of fine-tuning and adding new knowledge). Knowledge is furthermore separated in algorithmic knowledge and company specific knowledge. In this paper we discuss backgrounds of the expert system in more detail. An evaluation of the Expert system is also presented 1. Introduction Production control has become increasingly complex during the last few decades. On the one hand market requirements tend towards more diversity, a higher quality and reliable due dates. On the other hand the rapid ad- vancements in information technology have increased the possibilities on the shopfloor, thereby leading to more complex planning and control problems. In this paper we concentrate on discrete parts manufacturing environments. It is com- monly accepted that the complexity of the overall planning problems should lead to a hierarchical planning and control framework, where demand management and long term capacity planning are high level decisions, Correspondence to: G.J. Meester, Laboratory of Produc- tion Management and Logistics, Department of Mech- anical Engineering, University of Twente, P.O. Box 217, 7500 AE Enschede, The Netherlands. materials coordination is at the medium level, while detailed shopfloor planning and control can be found at the lower levels. For the high and medium level decisions we assume that a Global Planning System (GPS) has been implemented. In practice we often may find an MRP II system (Manufacturing Resource Planning system) at this place. However, it is well known that Global Planning Systems such as e.g. an MRP II sys- tem are not adequate in translating a planned set of orders into detailed work instructions at the shopfloor level. To this end, a Local Plann- ing System (LPS) has recently been developed at the University of Twente (cf. Refs [l-3]). Hence, this LPS acts as shopfloor planning and control system and as such bridges the gap between the Global Planning System and the actual execution of jobs on the shopfloor. To highlight the functioning of the LPS in some more detail, we assume (reflecting common industrial practice) that the factory shopfloor can be decomposed into relatively 0925-5273/93/$06.00 0 1993 Elsevier Science Publishers B.V. All rights reserved. CORE Metadata, citation and similar papers at core.ac.uk Provided by Universiteit Twente Repository https://core.ac.uk/display/11456198?utm_source=pdf&utm_medium=banner&utm_campaign=pdf-decoration-v1 454 G. J. MeesteriLocal planning tmmnn~e~~t autonomous groups of machines (this may be a functional arrangement but also a cell in a cellular manufacturing environment). Now, the Global Planning System organizes the materials flow between these machine groups (to be called Factory Systems (FS) in the se- quence). Subsequently, the Local Planning Sys- tem attempts to match this flow with the critical resources within each FS, such as tools, fixtures, machine capacities and operators. Hence, the FS is the local planning environment for the LPS. In performing these tasks, a large number of decisions have to be made while facing several, sometimes conflicting, objectives. To support this decision making an expert system may be helpful, in particular to evaluate specific work- ing conditions, to suggest planning and sched- uling algorithms and to judge the final results. The objective of this paper is to describe what such an expert system should look like, and to demonstrate that it can support an operator to substantially improve the performance of the LPS. In other words: it is demonstrated how we can formalize knowledge about planning and scheduling and how we can use this knowl- edge operationally. Other research on expert systems in the type of production situations discussed here can be found in Refs. [4-73. For general discussions on expert systems the reader is referred to Refs. [S-12]. In the remainder of this article we will focus on the expert system. First the expert system concept in general will be outlined and some properties of the concept of knowledge are discussed. Next the architecture, its tasks and functions are described for the particular LPS situation. The LPS-Expert system combi- nation has been evaluated in four different situ- ations. The evaluation of the expert system is presented in Section 5. Conclusions and sugges- tions for further research are given in Section 6. 2. The expert system: principles, tasks and knowledge The basic idea behind an expert system for an industrial company is that the company should save its investments in knowledge, even if skilled employees should leave the company. Expert systems can be used when we have to deal with problems in such fields as [13] : (1) interpretation and understanding of a complex data structure, (2) ordering/scheduling, (3) the evaluation of situations, (4) diagnostics, (5) the tracing of failures. There are different types of expert systems which can be used to solve these problems [9] such as: (1) rule based expert systems, (2) expert systems based on first order predi- cate logic, (3) expert systems based on structured objects (e.g. frames and trees). Looking at the characteristics of the tasks that an expert system should perform in the particular environment, studied in this paper, we have selected a rule based system. Rules can represent a unit piece of knowledge. Advan- tages of a rule based ES are: ~ it is easy to work with in test situations, _ it is easy to implement in a control concept (as will be explained in the next sections). A more elaborate discussion of arguments leading to the choice of a rule based ES is given in Ref. [14]. The structure of such an expert system is described in detail in Ref. [15]. Knowledge is represented in what we call “knowledge rules” (IF-THEN-ELSE). These rules are mutually independent. One of the properties of an expert system in general is that the knowledge (structure) and the data are strictly separated. The advantage of an expert system in contrast with a tradi- tional computer program is that if something changes in the problem environment, it will not cause much problems in the ES. One should simply add a new rule or new data. In a traditional program, the entire source code may have to be changed. The structure of an expert system will be briefly explained now. An expert system is con- structed from three basic elements: 1) the database, 2) the knowledge base, 3) the infer- ence engine. In our research framework we G.J. MeesterJLocal planning encironment 455 have added some slightly more advanced con- cepts like a statistics module (and in the future a simulation module). See Fig. 1. The inference engine matches patterns from the “rule base” or knowledge base, with recom- mended data from the database. The combina- tion represents the knowledge. In Fig. 1 we have also included the module “statistics”. By changing data in the database, it implicitly changes the knowledge. To clarify what is meant by knowledge, how it can be used opti- mally and how it should be represented in the rule based expert system, some techniques are illustrated by using an example. Example: Suppose the “expert” knows that computer time will increase exponentially fast when the algorithm “complete enumeration” (this algorithm calculates every possible sched- ule and gives always the optimal solution) will be used by a set of jobs larger than 20. We can implement this knowledge as follows: IF ord# ~20 (condition] THEN advise algorithm = complete enumeration (action}, certainty factor = 100 {certainty}. I : simulation It__. statistics :______: I I_____________, Fig. 1. Expert system, structure. Ii KNOWLEDGEBASE !+ IF a > b THEN action. I One can distinguish three main parts in the knowledge rules, which are explained in short: (1) (2) (3) The condition part: The condition part can be true or false. If it is true, action is taken, if it is false, no action will be taken. The condition part may find the values of the corresponding parameters (constants, variables, etc.) in the database or from other rules in the rule base (tree search). For instance in Fig. 2 action will be taken if a > b. The value of the parameter a is found in the database. However, b is not and therefore the inference engine now searches in the knowledge base and finds b in the second rule. Since now d is needed, the inference engine will search for it. In this way a tree of rules can be checked. The action part: This part of the rule de- scribes the action that has to be taken. An action can be: _ changing data in the database (for con- trol applications), - placing at inferences disposal of data for tree search activities, - proposing an algorithm. The certainty part: This part specifies the value of the so-called certaintyfactor. Basi- cally the value of the certainty factor indi- cates the probability that a recommended action will lead to satisfactory results. We will come back to the construction and role of the certainty factor in more detail below. Under complex environmental conditions it is not always obvious what may be expected from the performance of a given rule or algo- rithm (e.g. priority dispatching rules). Statis- tical knowledge about the performance of such a rule in the past, under similar conditions, may therefore be helpful to predict its behavior IF d 2 2 THEN b = 2, IF c <> 1 THEN d = 10 Fig. 2. Example knowledge representation. 456 G.J. Meester/Local planmng encironrnent in the future. In this way a sort of learning capability is included in the expert system. This idea has motivated the introduction of the certaintyfactor. Basically its value specifies the probability that an algorithm performs sat- isfactorily. Since any knowledge about recent performances should be included, a reasonable up-date procedure for the value of the cer- tainty factor is as follows: CF,,, = (CFold*N + 6 x 100) N+l . (1) Here CFold and CF,,, represent the initial and the updated values of the certainty factor, respectively. Furthermore, we set 6 = 1 if the solution determined by the relevant algorithm in the last cycle is satisfactory, and 6 = 0 other- wise. N is the number of previously recorded tests upon which the certainty factor has been based so far. For practical reasons the value of N should neither be too low nor too high. When N is very low the certainty factor will be strongly influenced by each additional test. When N is high the influence becomes negli- gible. When the rule is tested often, N is high and the value will become very stable. As cir- cumstances may change, considering all for- mer tests has its disadvantages. Therefore, both a lower bound and an upper bound are given to (in some cases even a constant value will do). Having thus motivated the use of the cer- tainty factor (basically to incorporate learning effects) we next explain its use. Given a particu- lar planning or scheduling problem faced by the LPS, we first check the environmental con- ditions and objectives to find what algorithms or rules might be applicable. Next, we simply choose the rule with the highest CF-value, where ties are broken arbitrarily. Another way to incorporate knowledge about the performance of certain rules or algo- rithms is by adjusting the conditions under which it may be applied. These conditions, such as the condition in our example, may be changed as a result of recent performance evaluations. E.g. in our example let us replace the factor 20 by a factor WF. The condition hence becomes ord d WF. In this case, as a re- sult of statistical analysis one may adapt the value of the factor WF and write this new value in the database. Complex statistical pro- cedures may be required to find good values for such factors. An expert system will only perform well if the knowledge added to the data and knowl- edge base is well defined. This means: the pro- cess of knowledge acquisition must be carried out thoroughly and accurate. A formal de- scription of the knowledge acquisition process is given in Ref. [16]. For our particular situ- ation we list a few important steps. Note that knowledge acquisition basically concerns the representation of knowledge rules (Fig. 3). (1) First the relevant relations in combination with the relevant indicators have to be traced. To get these relations one may use a knowledge relation chart (see Fig. 3). In such a chart the relations between the LPS algorithms and the potential environment indicators are given. After an interview with an expert or analyses done on the LPS algorithms, the relations between the indicators and the LPS-algorithms found can be added into the chart. (2) After finding the important relations, the effective area of the indicators must be found (i.e. in our example the effective area was the interval [0, 201. Values of 20 or more are not relevant). Possibilities here are to test the LPS algorithms for different input values. The condition part can be defined in combination with the action part (LPS-algorithm). Fig. 3. Knowledge acquisition chart. G.J. MeesterlLocal planning environment 457 (3) Finally a certainty factor must be set. When there is no information about the performance of the LPS-algorithm an ar- bitrary value must be added. In the long run the certainty factor converges to a stable value (in a stable environment). With this information the condition part, action part and certainty part can be projected on the knowledge representation structure (IF-THEN-ELSE structure) which yields the knowledge rules that can be implemented. 3. An expert system for the LPS A Local Planning System contains several modules, structured in a hierarchical frame- work. The two most important ones are the “Planning module” and the “Scheduling module”. The planning module performs the follow- ing functions: _ It receives a set of orders (see Fig. 4) from the GPS that have to be produced in a certain time period of one to three weeks, say. - It checks capacity and handles utilization planning on a high aggregation level. This Fig. 4. LPS. means in particular: _ moving orders in time, _ selecting alternative process plans, avail- able in “technical data” (cf. Fig. 4). _ It selects a subset of orders for the next day from the above described order set. This sub- set forms the input for the scheduling module. The scheduling module performs the follow- ing functions: _ It creates the schedule for the next day by assigning all operations from the selected subset of orders to machines and by sequencing all operations assigned to that specific machine. - It creates task lists for the resources (e.g. tools, machines, fixtures, operators, etc.). Both levels contain a range of different algo- rithms and heuristics. For most particular situations there is an algorithm or heuristic available that performs best under these circumstances. This means that an LPS oper- ator (planner) has to interpret the situation (given by the GPS and the actual situation on the shopfloor) and has to select an algorithm that yields a satisfactory planning/scheduling solution. The algorithms and heuristics available on “planning” level are: (1) For utilization planning: _ moving orders in time: _ sorting algorithms to select jobs on specific properties, _ clustering algorithms on technical constraints (for tools and fixtures), _ clustering algorithms with precedence relations, release and due dates. - process plan selection: - complete enumeration (workload bal- ancing), _ mixed integer programming (work- load balancing), - linear programming, continuous rounded (workload balancing), _ two-opt heuristic (pair-wise inter- change) (workload balancing), _ sorting on resource load (workload balancing), (2) For workloading: - linear programming. The module scheduling contains algorithms and heuristics to generate schedules which focus on the utilization of resources, short term due-date performance, throughput times, and schedule robustness. (1) Before using actual dispatching and sched- uling algorithms, it is possible to use a preparation tool to let the dispatching and scheduling algorithms perform better. For example: _ add all operator relations, _ add relations for a specific fixture type, _ add all machine-tool relations for a spe- cific machine. (2) The dispatching may be used to accom- modate e.g.: _ order due dates, _ operations due dates, - priorities. The actual scheduling procedure is the shifting bottleneck procedure developed by Adams et. al. [17]. (3) After having generated the schedule, there is a possibility to improve the schedule somewhat further by one of the secondary scheduling procedures: _ make feasible, _ adapt for gaps. The heuristics and algorithms are described more extensively in Ref. [3]. The expert system should support the deci- sion making in selecting algorithms on differ- ent levels. It may either support the operator or operate as an autonomous program with- out intervention of the operator. Let us return to the specific expert system structure de- veloped for the LPS environment. The LPS basically consists of two layers: a planning level and a scheduling level. An important rea- son to split the system in such a hierarchical structure is to be able to deal with the com- plexity of the problems encountered. Also the ES developed here is used in an experimental environment, with a data and knowledge base on each separate level it seems easier to test. Having just one level where all decisions should be made it may become harder to evalu- ate the effects of any new rule or alteration. As mentioned earlier, the expert system has to perform the following tasks. On both, the planning and the scheduling level it has to: _ evaluate the situation in the planning and scheduling environment, _ support the operator by choosing the right algorithm or heuristic in a specific situation (generate), _ evaluate the result of one algorithm or heu- ristic (are we satisfied or not?) (evaluate). The generate-evaluation concept (depicted in Fig. 5) is implemented on both LPS levels. On each level the expert system has to advise. All relevant data according to this advice and to the result of this advice is temporarily saved (as best). The evaluation part checks this data. If “evaluation” is satisfied, the expert system permits to move to a lower level of the LPS. If “evaluation” is not satisfied, the expert system has to generate a new solution by sug- gesting a new algorithm. Remark I. When “evaluation” is not satis- fied, it will explain to the advice-part what’s wrong. So the advice part is able to re-advise “intelligently”. We call this control loop the short control loop. Remark 2. When “evaluation” continues to return “not satisfied” it may advise to go to ____________ evalua :ion part advise part j w o on with ‘best’ to next stag & Fig. 5. Control mechanism. G.J. MeesterlLocal planning environment 459 a lower LPS-level after a certain number of loops. This is to protect against infinite loops. (Temporary solution “best” is given to the lower level). Remark 3. The evaluation system on the LPS-scheduling level has the possibility to send information to the LPS-planning level. Since we deal with a hierarchic system, it may possibly happen that the planning module presents an order subset to the scheduling module which does not allow to meet a specific criterion under any schedule. Both advice and evaluation part have to be fitted into the expert system (see Section 4). 4. Implementation The expert system is implemented on a HP 9000 workstation. It operates together with the LPS interface -1 , I LPS under an X-window representation man- ager on a UNIX operating system. Both are written in Standard Pascal, see Ref. [15]. The choice for Pascal is motivated by the fact that in this type of research one has to deal with many new concepts and changes. In Pascal it is easy to develop new ideas. We also have investigated different software packages, like the IBM Expert Shell [9] and the Delfi II Expert Environment. Both of the packages are professional expert environments but the link between these Expert Systems and the LPS is difficult to make. Also extensions on the Ex- pert System framework the hardly to realize. Once having discussed the control mechan- ism and the knowledge structure we are now ready to present the complete expert system structure. The structure is depicted in Fig. 6. Observe the knowledge rules embedded in this concept. expert system I GPS I i I IF_TKN-ELS.. IFJKNELSE.. lzzza= tnference engine Fig. 6. Control chart. 460 G.J. Meester/Lod plunnmg ewirorwnent Not mentioned till now but important for the ES-concept is the interface unit (see Fig. 6). The interfaces are mechanisms which are able to compress files of information in a special way. In particular, a file given by the GPS can be transformed into a vector of indicators. In that way it is easier for expert system to interpret the GPS situation. For a more elab- orate treatment of the interfaces we refer to Ref. [14]. To further highlight Fig. 6 one, step of the advisory and evaluation process of the expert system will be described. Step I. The GPS file is translated into an indicator vector. The mass of available data will be compressed to a set of indicators which represent the important characteristics of the file. The vector is stored in the database. Step 2. After a request for advice is received, the inference engine matches the data and the rules and presents the advice. It also reports to the module statistics. (1) (2) (3) (4) Step 3. The result of the advice is imple- mented. Step 4. The planning file (containing the results of following the advice) is transformed in a way similar to the one described in step 1. The combination tested in an operation oriented FS situation. This case concerns an FS which contains 5 rather similar milling/drilling machines (the only differ- ence between the machines is the number of driven rotation and translation axes). The second case concerns a part oriented FS consisting of 5 machines of three differ- ent types. We may think about families of products visiting this machining centre (part flow). For details see below. The third case is taken from Ref. [IS]. Actually this FMS situation concerns a pilot plant in a Dutch research institute (IPL-TNO, Apeldoorn). It consists of a turning centre and a FS. The machines and a tool preparation cell are integrated in a CIM environment (the TN0 super- visory control system). The final case is a comparison made be- tween OPIS/ISIS, a scheduling system de- veloped by Fox et al. [4], and the LPS-ES. The test situation used in this case is de- scribed in Ref. [19]. Special in this case is that the FS consists of a number of cells, where each cell contains a number of sim- ilar machines. Step 5. There are two possibilities. The re- sult of the evaluation is: “ok”: -+ We continue with scheduling; go to step 1, “not ok”: --t a new advice is given. The differ- ence is that the evaluation changes the data in the database on the advice level (see Fig. 6). Go to step 1. In our expert system we have implemented the knowledge evaluation by means of the cer- tainty factor. The module “statistics” takes care of the handling of the factor N. For every rule which generates an advice a value of N is maintained. How the LPS-ES combination works in a particular environment (recall that the ES implementation is situation-specific) is ex- plained in more detail, Although the LPS-ES combination is extensively tested for each of the four cases described above, at this point, for the sake of brevity only the knowledge acquisition and the evaluation of the second case are described. 5.1. Outline of the second case 5. Evaluation of the expert system 5.1.1. The factory system The FS consists of 5 machines: two of type A, two of type B and one of type C. Each machine has a tool magazine with a capacity of forty tools. To evaluate the power and usefulness of the Each job follows a predetermined route LPS-ES (Local Planning System-Expert Sys- along the machine types: A + B + C. The tem) combination, it is implemented in four maximum number of different operations different environments. within one job is three on machine A and B G.J. MeesterlLocal planning environment 461 and two on machine C. The total number of operations within one job is at most five. Two shifts (16 hours) per day are assumed. 5.1.2. Jobs 80 jobs have to be produced in one week. Every job has to be produced in 1 to 5 days, with a mean of 3. Jobs also can have a priority index ranging from 1 to 3 where 3 means a rush job. Every job consists of one or more products of one available type. For every product type there are two process plans avail- able which can be selected for production. Other job characteristics are: the batch size: 1 to 5 with a mean of 3, the number of operations: 1 to 5 with a mean of 3, the process time: 10 to 50 minutes with a mean of 30, special tools per operation: 1 to 5 with a mean of 3. A file generator has been developed to gen- erate a GPS file with the preferred factory and job characteristics. 5.2. Knowledge acquisition for the second case rithm, based on the situation given by the GPS and depending on the status of the FS. For this information is required (cf. Kl . . . Kn in Fig. 2). This information must be compressed into a limited set of indicators (see also Fig. 6). Based on the results of step 1 and given the information stored in the GPS data file, we have chosen the following set of indicators for the selection of planning algorithms: (1) the total workload per machine, (2) total workload released/total workload, (3) the number of different parts, (4) the average number of alternative process plans, At this point the knowledge acquisition pro- (5) the total number of operations, cess (in analogy of Section 2) for the expert (6) the average order priority, system for the second case is described. We (7) the average release date, confine with the description of the acquisition (8) the average due date, process for the LPS-“planning level”. (9) the number of orders. 5.2.1. Step I: dejinition of useful algorithms, heuristics and solution rules First step in the knowledge acquisition pro- cess is to select a subset of useful algorithms (cf. Pl . . . Pn in Fig. 2) from the nine algorithms available on the planning level. In this case the objective on the planning level is to balance the workload among the machines. According to this objective we have, based on the knowledge and experience of the LPS-builder, selected the following algo- rithms: For the process plan selection: - full enumeration, _ a heuristic on workload balancing, _ a mixed integer programming formulation, _ a heuristic for sorting on the workload of one specific machine. For work loading: - linear programming with a utilization of 95%, - a heuristic with a tool capacity bound and a utilization of 95%. 5.2.2. Step 2: dejinition of indicators for the advice part The expert system has to advise an algo- 5.2.3. Step 3: definition of indicators for the evaluation part The evaluation is an important part of the expert system framework. When an algo- rithm is selected and executed, the results may be judged to be poor or even not accept- able by the evaluation part of the expert sys- tem, according to some specified criteria. These criteria may be simple operational cri- teria such as the average or maximum lateness but may also concern more “fuzzy” conditions such as the capabilities of operators to deal with the resulting solution. To this end some control indicators are needed. We have 462 G.J. Merster/Local platmng twwontntwt selected: (10) (11) (12) c A parameter indicator whether or not the result is almost satisfactory, what exactly is meant by “satisfactory” is discussed below. The number of attempts so far. The algorithms evaluated earlier. These parameters are written into the database. They provide sufficient information to yield a good advice. Generally, the pro- cess to find the proper parameters for practical situations will be iterative. _ J.4. Step 4: dejinition sf the knowledge rules Now that we have defined a set of important indicators in combination with a set of useful algorithms in the previous steps, we may con- struct the knowledge rules by using a chart as depicted in Fig. 2. After the determination of the indicator intervals and the certainty factors our knowledge acquisition process is com- pleted. We implemented two types of knowledge. We can distinguish: (1) (2) theoretical- knowledge: this is knowledge about the actual LPS-algorithms. This means that the relation between the input data characteristics and the output perfor- mance of each algorithm and its properties in different situations are stored in knowl- edge rules. This knowledge is generally usable, situation specific knowledge which may concern specific company knowledge or FS-configuration constraints (e.g.: tool X is not available for job of type Y). This knowl- edge is specific for each of the four cases and therefore different in each of the cases. 5.3. Evaluation of the espert sqxtern in the second case 5.3.1. HON. to evaluate an expert system The expert system is implemented to sup- port the operator in decision making among the different algorithms, rules and heuristics available on the different levels in the LPS. So, what we are interested in, is what the quality is of the advice to select a specific algorithm, as given by the expert system. Did the algorithm, suggested by the expert system, perform well? Before we can give judgement about a satisfac- tory performance of an advised algorithm, we first must know what exactly is meant by satis- factory. In other words: how can one measure the quality of an advice. 5.3.2. The qualit), of an advice To establish the quality of an advice one has to evaluate the actual output of the LPS, which is actually a complete production sched- ule. A schedule can be evaluated by perfor- mance indicators like maximum lateness, tardiness, utilization of machines or through- put times. These performance indicators can be compared with lower bounds. Lower bounds can be found by using e.g the shifting bottleneck procedure (cf. Ref. [17] ) for the lateness criterium and a mathematical programming formulation for the utilization criterium. Deviation of the performance indi- cators from the lower bound gives a measure for the quality of a production schedule. In this research, the quality of the produc- tion schedules is measured for every advice generated by the expert system. The reason for not presenting the data to match, is that the data gives more information about the quality of the used algorithms themselves instead of given straight the information about the qual- ity of the decision of choosing the algorithm. So, in our evaluation we have used the follow- ing performance measure: did the advised al- gorithm lead to a solution (i.e. a workload balancing or a production schedule) which could not be improved by applying any of the other available algorithms according to the performance measures described above (in particular the lateness criterium). 5.3.3. Evaluation results of’ the second case The LPS Expert system is tested on 20 dif- ferent GPS order sets. G.J. Meester,iLocul planthg etwironment 463 In 10 cases the expert system advised these algorithms which lead the LPS to an “opti- mal” production schedule at once. In 7 cases the expert system used its evalu- ation optimization mechanism (cf. Figs. 5 and 6) on planning level only (4 times) on schedul- ing level only (3 times) and on both levels at the same time (2 times) to find the optimal combination of algorithms leading to the “op- timal” production schedule. In 2 cases the expert system advised to re- turn to the LPS-planning level for a complete new LPS session, after recognizing infeasibility at the LPS-scheduling level. These cases show that the control mechanism as depicted in Figs. 5 and 6, works well. In one case the expert system did not advise the algorithm that finally led to the “optimal” production schedule. 5.3.4. Remarks (1) (2) (3) The results of these tests reflects the quality and the completeness of the know- ledge gathered during the acquisition phase. Although only the second case is outlined here, equivalent results have been found in the other three test environments. The fact that most of the times an optimal production schedule has been found indi- cates that the situation presented by the GPS is interpreted well by the expert system. 5.3.5. Conclusions extracted from the test cases include (1) (2) In all four test cases, the LPS performs well on the advices given by the ES. We only once could find a better production sched- ule by using a different algorithm then the one advised by the expert system in the second case. In the various, often complex, situations and circumstances the expert system appears to be a useful tool in supporting decision making to control the LPS. (3) The theoretical knowledge (not presented here but described in Ref. [14]) added till now is functioning well. (4) It appears to be very easy to add knowl- edge, and make the universal LPS situ- ation specific (e.g. by adding practical constraints). 6. Conclusions and suggestions for further research In this paper we have outlined the structure and functions of a rule-based Expert System (ES), which has been developed to assist deci- sion making by a Local Planning and Schedul- ing System (LPS), operating at a production cell level. In particular, we have underlined the need to recognize and interpret complex situ- ations in the Local Planning Environment, as well as the fact that any advice from the LPS should be based on both theoretical knowl- edge (algorithms and heuristics) and factory specific practical knowledge. A prototype ver- sion of the Expert System has been pro- grammed. Finally, we have evaluated the performance to the ES by considering some typical cases; this evaluation in particular ad- dresses the quality of the advisory function, not primarily the quality of the underlying algorithms (which is quite a different topic). Further research might be concentrated on number of aspects, including: the influence of alterations in the condition part of the underlying rules. The knowledge evaluation has been based primarily on the values of the certainty factors. Changing the range of values for which a condition of a specific rule should hold may severely in- fluence the ultimate performance of that rule, the extension of the ES with a simulation part (cf. Fig. 1). Such an extension yields the possibility to generate additional knowledge by simulating several alternatives. See Ref. [14] for some more elaborate ideas, knowledge acquisition in a real manufactur- ing environment. This will be the next topic in forthcoming research. 464 G.J. MeesterlLocal planning environment Acknowledgement ISBN 90-6789-050-2. The author is grateful to Professor W.H.M. Zijm for detailed comments on earlier versions of this paper which have substantially im- proved the presentation of the material. c91 Cl01 References iIll1 Jackson, P., 1987. Expert Systems. Addison- Wesley, Reading. Horvitz, E.J., Heckerman, D.E. and Langlotz, C.P., 1985. A Framework for Comparing Alternative Formalisms for Plausible Reasoning. In: Proceed- ings Fifth National Conference on Artificial Intel- ligence. Vol. 1: Science. Morgan Kaufmann Publishers, Los Altos, pp. 210, 214. Turksen, I.B., 1988. An approximate reasoning framework for aggregate production planning. NATO AS1 series, Vol. 49, C.I.M, pp. 243, 266. Levesque, H.J. and Brachman, R.J., 1985. A Funda- mental Tradeoff in Knowledge Representation and Reasoning, pp. 42, 70. Buchanan, B.C. and Duna, R., 1981. Principles of Rulebased Expert Systems. Stanford University Memo, HPP-81-4. Meester, G.J., 1990. Een Expertsysteem voor Planning- en Schedulingsalgoritmen, Master thesis University of Twente. Saywer, B. and Foster, D., 1986. An Expert System in Pascal. VI c21 c31 c41 c51 C61 c71 PI Bossink, G.J., 1990. A Framework for Planning and Scheduling Manufacturing Systems. In: Proceed- ings Workshop on Decisional Architectures in Automated Manufacturing, Genova, Italy. Bossink, G.J., 1990. Planning and Scheduling (Flex- ible) Manufacturing Systems. In: ISCIE Proceed- ings, 1990 Japan-U.S.A. symposium on Flexible Automation, Vol. III, pp. 1109. 1112. Bossink, G.J., 1992. Planning and Scheduling for Flexible Discrete Parts Manufacturing. PhD. Thesis. ISBN 90-9004508-2. University of Twente. Enschede, The Netherlands. Fox, B.P.. Allen, M.S., Smith, S.F. and Strohm. G.A., 1983. ISIS: A Constraint Directed Reasoning Approach to Job Shop Scheduling. Carnegie Mel- lon Univerity, CMU-RI-TR-83-8, Pitsburg, USA. Bensana, E., Bel, G. and Dubois, D., 1988. OPAL. A knowledge based system for industrial jobshop scheduling. NATO AS1 series, Vol. 49 C.I.M, pp. 295, 330. Wu, SD., 1987. An Expert System approach for the control and scheduling of flexible manufacturing cells. Worden, B. and Weber, E., 1989. Erfurungen emes Expertsystems in Konstructions berreich. In: V.D.I. Berichte 775: Expertsysteme in Entwicklung und Konstruktion, pp. 215, 268. Bonnet, A., 1987. L’In~elligence artificielle; Promesses et R&alit&s. InterEditions, Paris, France, WI Cl31 Cl41 Cl51 Cl61 Cl71 Cl81 Cl91 c201 Buchanan, B.C., Bartlow, D. and Benet. J., 1983. Constructing an Expert System. (Eds. Hayes-Roth, Waterman and Lenat), chapter 5. Adams, J., Balas, E. and Zwawack, D., 1988. The shifting Bottleneck Procedure for Job Shop Sched- uling. Manage. Sci., 34(3). Aanen, E., 1988. Planning and Scheduling in Flex- ible Manufacturing Systems. PhD Thesis, ISBN 90-9002349-6, University of Twente, Enschede, The Netherlands. Fox, M.S., Chang, W.Y. and Peng. S.O.. 1990. Fac- tory Model and Test Data Descriptions: OPIS Experiments. Internal Report. Carnegie Mellon University. CMU-RI-TR-90-6, Pitsburg, USA. IBM, Knowledge Processing Series, Expert System Development Environment, Expert System Con- sultation Environment, and Expert System refer- ence manual. 
work_t67ptrnx25bcdnze7t352dhoqm	----	�� The Expert System as a Proposal for Creating Innovative Strategy Lendel Viliam, Varmus Michal Abstract The process of creating an innovation strategy is a complex one. The formulation of innova- tion strategy requires a more intensive calculation that makes it possible to select the optimal variant of innovation strategy for enterprises. Similarly, the creation of different innovation strategies requires the use of information technology. A place has to be allocated to hold in- termediate results. Also, work with larger amounts of data and knowledge must be stored in transparent database, to avoid loss, confusion and difficulty in searching for information. This paper examines the use of an expert system as an appropriate means of meeting the require- ments of creating an innovation strategy. The paper examines in detail the various modules of the proposed expert system, as well as the preconditions for a successful performance. Key words: innovation strateg y, expert system, innovation, knowledge, expert, management 1 INTRODUCTION The current period is marked by many changes. Competition is constantly sharpening, busi- nesses feel effects of global economic crisis, constantly developing new products, customers, primarily through the Internet have a lot of information on which the decision to buy. Cus- tomer focus has become part of a corporate approach of many companies (Strišš et al., 2009). Efforts to identify customer requirements and transferred to the new product is the task of today’s marketers. Businesses are trying to exploit these changes and turn them into their own competitive advan- tage. Innovation activities are essential to the survival and growth of the company. They sig- nificantly affect to its performance, market position and market power. They are a key business process, because through these companies are trying to achieve some competitive advantage. The basic precondition for the creation and use of innovation in the enterprise is well worded innovation strategy. 2 INNOVATION STRATEGY AND ITS MAKING PROCEDURE Innovation strategy is understood as innovative business approach to the choice of objectives, methods and ways to fully utilize and develop the innovative potential of the company (Len- del, Varmus, 2010, p. 49). This is the direction of its boundary, which determines the potential of innovative strategies. Innovation strategy is closely linked with corporate strategy, therefore, must reflect the basic features of this strategy. Corporate strategy defines the scope of the enterprise within the meaning of the industry and markets in which the enterprise operates (Kislingerová, Nový, 2005, p. 106). joc_3-2010_v1c.indd 47 17.12.2010 19:05:25 Journal of Competitiveness | 2/2010�� Innovative potential of the strateg y can be defined as a measure of innovation strategy, which would have been achieved in the optimal utilization of all sources of innovation strategy (Lendel, Varmus, 2010, p. 49). The level of innovation potential of the strategy depends on the level and quality of the components of the innovative resources strategy. Innovation strategy, we understand the sources of innovative opportunities, skills, knowledge, experi- ence, invention and innovation, the firm available, or is unable to obtain in time. Innovative resources of strategy consist of four basic, interrelated modules, namely: (Lendel, Varmus, 2010, p. 49) Bank of inventions, Bank of innovative opportunities, Knowledge base, Bank of innovation. The process of innovation strategy is a complex process that contains six main parts. This is a defining vision and mission of the company, identifying strategic objectives, detailed analysis of the enterprise environment (internal and external), strategy formulation itself, its implemen- tation and subsequent evaluation associated with the control. The most important process of creating an innovation strategy we considered the formulation of strategy. The process to generate the different variants of innovation strategy, its analysis and evaluation according to established rules and criteria specified. Based on past performance of activities may be to select an appropriate solution to an optimal variant innovation strategy for the company. Strategy formulation process is marked by more intensive computations oc- curring mainly in the selection of appropriate solutions. Even generating different options strategies requires innovative use of information technology. For subsequent evaluation of the various options it is necessary to interim results be imposed in an area for the purpose of subsequent confrontation and comparison with current outputs. It should also be based on more data and knowledge, which must be stored in transparent database. This will avoid the emergence of common situations where they are being confused, searching, lost and resting mainly due to absence of awareness of their existence within the enterprise. All of these prereq- uisites and requirements for successful development of innovative strategies can be achieved by introducing an expert system which will provide to senior managers detailed information necessary for decision-making. For these reasons, we consider for purposes of creating a successful innovation strategy for the introduction of appropriate expert system that will ensure effective work with knowledge and innovation-related data. Knowledge-based systems using artificial intelligence solution allows arbitrarily complex problems. Used knowledge is considered crucial for high-efficiency innova- tion strategy. These systems are used in solving the problem of expert knowledge. 3 PROPOSAL OF EXPERT SYSTEM FOR WORK WITH INNOVATION AND CREATION INNOVATION STRATEGY Expert systems can be understood as a knowledge system in which expert knowledge is used in a very specific problem area. The main objective of the proposed expert system will achieve the best response to the real data on innovation, thereby ensuring high quality decision-making     joc_3-2010_v1c.indd 48 17.12.2010 19:05:25 �� innovation strategy. Creating an expert system is a complex process as the project site, as well as for programming. Based on the analysis of the literature on the creation of knowledge and expert systems (Návrat et al. (2007); Spalek et al. (2005); Kelemen & Liday (1996)) and after careful examination of the issue of innovative strategies (Horňáková & Zaušková (2008); Dupaľ & Molnár (2002); Kováč (2007); Tidd et al. (2007); Dupaľ et al. (1997); Zaušková (2006); Zaušková & Loučanová (2008)) proposes that we an expert system to work with the knowledge needed to create an innovation strategy consisted of the following basic parts: The core system (knowledge base, data base, working memory, and stack mechanism infer- ence appropriate solutions), Input / output module, Explanatory module, Protocol, Other components of the system (knowledge base editor, editor of the database module learning outcomes generator module external sources. Proposal of an expert system for work with innovation creation and innovation strategy (Fig. 1) was, in addition to learning foreign and domestic literature, created (processed) on the basis of previous research. The purpose of the research was to identify readiness of selected Slovak companies for the deployment and use of innovative marketing strategies on the base of the identification of key elements, work with innovative ideas, opportunities, innovation and ap- plication of lateral thinking. Research was conducted on the sample 236 senior managers of medium and large enterprises operating in the Slovak Republic. Most managers were contacted through an electronic questionnaire (93.2%). 6.3% of top managers have personally reached out through semi-structured interview. Fig. 1 – Proposal of expert system architecture to work with innovations. Source: own elaboration      Editor of data base Editor of knowledge base Module of learning Explanatory module Protocol Generator of results In-put/Out-out module User (top managers, marketers,...) Expert (work with knowledge) External data External programs Module of external sources Expert system Inference mechanism Selection operator Generator Restriction generated solutions Testing of data compliance Knowledge base Data base Bank of inventions Bank of innovative opportunities Bank of innovations Tray of appropriate solutions Working memory CORE OF SYSTEM joc_3-2010_v1c.indd 49 17.12.2010 19:05:26 Journal of Competitiveness | 2/201050 The results obtained from research formed the basis for the development content of the vari- ous elements of the proposed system. In particular, interviews with top managers helped to gain a more comprehensive view of the implementation of innovation strategy in the company. Figure 1 shows the architecture of the proposed expert system for dealing with innovation in the development of innovative strategies. This is a complex system whose components must interact with each other and provide the necessary knowledge in real time. The basic precondition for the successful operation of the proposed expert system is the exist- ence of actual knowledge base module and the module data base. Kelemen and Liday (1996) emphasize the need to strictly distinguish the data structure representing generally applicable and accepted evidence from the data structure. This is due to the different requirements of access and manipulation. The proposed expert system will perform two basic actors: the user and the expert. User is a person who, in practice the expert system uses the capabilities of working with innova- tion and creation of innovative strategies. These are the top managers and marketers. Expert knowledge is a source of innovation and innovation strategies. 4 CHARACTERISTICS OF FUNDAMENTAL ELEMENTS OF THE PROPOSED EXPERT SYSTEM The proposed expert system consists of modules, which provide its functionality. Each mod- ule performs a specific task. Outputs from one module are inputs to the second module. The successful performance of the proposed expert system is essential to ensure coherence and seamless communication between modules. 4.1 Knowledge Base Knowledge Base focused on expert knowledge gained. It provides a space for the collection of all knowledge that can be used in the innovation process. Tidd (2007) points out that innova- tion is closely linked with knowledge. It states that “the issue is to create new opportunities by combining different sets of knowledge” (Tidd, 2007, p. 16). These are the knowledge of technological possibilities and the appropriate configurations to meet the needs of business and customers. This knowledge may take the form of experience accumulated by the company during its existence or be the result of the review process (technology, market competition ...). The main purpose of knowledge base is to provide space for an appropriate mix of skills into a successful innovation. In addressing the knowledge seeking knowledge manager needs to proceed further in solv- ing the problem. The knowledge base must be designed to allow efficient access to required knowledge, while allowing the largest store of knowledge. It must correspond to the choice of implementing data structures. 4.2 Data Base Data Base contains all the unique information relating to innovation. It consists of Bank of inventions, the Bank of innovative opportunities and the Bank of innovations. Bank of inventions is the search space, creation, evaluation and storage of inventions that may joc_3-2010_v1c.indd 50 17.12.2010 19:05:26 51 participate in the next phase in creating an innovation strategy. Invention is new knowledge that results in a change in the structure and level of knowledge and this is a new idea of pos- sible change, respectively new approach (Zaušková, 2006, p. 37). It is about suggestions, ideas and thoughts of a solution. The process of inventions is a complex process that requires a large number of variants. Innovative ideas are the product of information coming from the sur- roundings and the business system itself. This information is identified to the needs and con- cept solutions. Important processes are the recognition of status and choice of appropriate means. The result of this process is an innovative idea (invention), which is ready for their review and decision on its need and value for the purposes of innovation strategy. Bank of innovative opportunities is a space to store and work with the identified innovation op- portunities. An innovative idea (invention) is a necessary but insufficient condition. The idea must be feasible and satisfy the conditions for potential success. We are talking about find of opportunities - the process of change an idea into an opportunity. Review inventions oppor- tunity for innovation is a complex process that requires some assessment of up to 100 ideas for one successful product (Zaušková, Loučanová, 2008, p. 40). Here, the enterprise can create its own rating system. During evaluating, the managers must reckon with feedback in the form of corrections and improvement ideas, combining them, and breaking down under (Zaušková, Loučanová, 2008, p. 40). The Bank also serves innovative opportunities for the conservation of innovative opportunities for the company are not immediate importance. Bank to gradually accumulating innovation opportunities for peer evaluation can create the necessary synergies. Bank keeps all the innovation created by innovation and creates the environment for their effective management and their translation into successful innovation strategy. Innovation is understood the practical transfer of new ideas into people’s products, services, processes, systems and social relations. 4.3 Working memory Working memory provides space for the solution of the problem. It will include three compo- nents: a solution, plan and agenda. Solution provides to the space in working memory, which requires the solution to the problem created. It is about hypothesis, proposed structure and so on. The plan contains a sequence of actions to be taken to solve the problem. The plan may, for example a sequence of such steps: design the scheme-specific variants of innovation strategy and then optimize. The priorities of each decision module plans - scheduler, which selects the plan and the work to be carried out within the plan developed. Agenda in the form of data structure stores a sequence of actions to be taken to implement the selected activities planner. The priorities assigned to interpret these actions to maintain and ensure the activation of broken shares in the event that other actions have higher priority. Working memory can be seen as a place where solutions are stored intermediate data and arrange the activities of the solutions. 4.4 Input-Output Module The main task of I / O module is to create an interface between knowledge-based system and its surroundings, which is represented by end-users. These are the senior managers and staff involved in the process of creating an innovation strategy. Other end users are business mar- joc_3-2010_v1c.indd 51 17.12.2010 19:05:26 Journal of Competitiveness | 2/2010�� keters who use the knowledge gained in the evaluation of inventions and innovative opportu- nities in the implementation analysis to create an innovation strategy. Through this module, the user sees an expert system. The main objective in this area is to create user interfaces that allow communication affordable for solving the problem, i.e. while creating an innovative business strategy and interpretation of appropriate solutions found. 4.5 Explanatory module The basic module is a function of clarifying clarification, explanation and justification of the decision, which is the output of an expert system. Managers and marketers obtain the neces- sary justification for the final solution in the form of the chosen innovation strategy. Explana- tory module offers managers and marketers the opportunity to browse the knowledge base, view and modify working memory. They get invaluable help when debugging a knowledge base to guide the course of the solution. The importance of clarifying the module and see the possibilities of development of various innovative options strategies. Managers would be able to choose alternatives to develop innovative strategies that expert system still has not been examined. Explanatory module would allow examining the impact of certain decisions on innovation located by the innovation strategy. History of the solution is charged in a protocol module then uses to explain. 4.6 Inference mechanism It is allows finding the required knowledge in knowledge base, data base and using them to develop innovative strategies. It can derive from these bases for further information and knowledge. Its work is based on a knowledge base and data base on which influences the choice available to operators, limiting the number of tested solutions proposed generator and controls compliance testing solutions generated with the actual data. One of the important outcomes of inference mechanism is tray of appropriate solutions. This module contains the appropriate solu- tions, which are rated according to their fitness level. They then enter into explanatory module, through input / output module which gives to user (the senior manager or marketers). 4.7 Other components of the proposed expert system Other important components of the proposed expert system are: Editor of knowledge base, Editor of data base, Module of learning, Generator of results, Module of external sources. Editor of knowledge base provides a constant update, completion and dissemination of knowledge base. This change occurs not only during its development, but often also during operation. The reasons leading to these changes may be different. The most common is the acquisition of new knowledge that may help in the process of innovation strategy of company. Second, the manager can identify errors that must be removed (such as rules for generating variants of an innovation strategy, importance of amending the innovation process ...). The same principle is based on the editor of data base, which, unlike the editor working from knowledge base informa-      joc_3-2010_v1c.indd 52 17.12.2010 19:05:26 �� tion relating to inventions, innovative opportunities and innovations themselves. There may be reason for changes such as incorrect identification of ideas and their subsequent translation into innovative opportunities... An important part of the proposed expert system is a module of learning. Its main objective is to promote acquisition of knowledge. It ensures that is always based on the current situation. The obtained knowledge is stored back into a knowledge base and used for future proposal innovative business strategy. Generator summarizes the results of partial results in a reasonably integrated whole, without extra information requested in the form of a comprehensible form. His contribution is in pro- viding effective, efficient, differentiated, comprehensive, current, serving mainly commercial information for decision-making in innovation. Communicates with input / output module to provide senior managers with the required information and understandable. Module of external sources provides communication of the expert system with their environ- ment. The main activity of this model is to work with external data and work with external programs. Inference mechanism in case of request certain data will search in the data base. If there in not find the required information, the management is submitted to module external sources. He begins to search external data source. In case of found of required fact it is allowed insert it into the data base and send back the management to inference mechanism. Likewise it does even in case of necessary expertise. As inference mechanism doesn’t appear requisite knowledge in the knowledge base, submit the management to module of external sources. It begins to search external data source. If successful, will embed the acquired knowledge into a knowledge base, if fail then it turns through the input / output module for expert, who knowledge supplemented by the necessary knowledge base editor. Then handed back control to inference mechanism. 5 CONDITIONS FOR SUCCESSFUL WORK OF THE PROPOSED EXPERT SYSTEM That the proposed expert system to work effectively, it is necessary to ensure that the follow- ing conditions: Efficient data acquisition, Implementation of quality input / output module, A detailed and careful analysis of the enterprise environment. The quality of the implementation of input / output module can substantially multiply the performance generated by the system (Návrat et al., 2007, p. 259). These are the creation of executive comfort interactions with expert systems. The aim should be to create an acceptable user interface allowing the creation of innovative communication strategies and appropriate business located as a result the interpretation of the innovation strategy. The most important prerequisite for the successful operation of an expert system we consider a detailed and thorough analysis of the enterprise environment. In a first step, the undertaking must determine its innovation capacity. It consists of the sum of knowledge, experience, resources, assets and managerial capabilities and skills in business available, or is unable to obtain in time.    joc_3-2010_v1c.indd 53 17.12.2010 19:05:26 Journal of Competitiveness | 2/2010�� This is the basis for creating an innovation strategy. Then there is a mapping of innovation potential, the rate of innovation means business, it can reach the optimal utilization of all components of innovative capacity. The next step, the enterprise must assess and identify the current level of use of innovative capacity. This analysis will provide a realistic picture of the possibilities of innovation, which in turn translate into specification of innovative requirements. These are the selec- tion of the main operators, i.e. areas in which are interesting for the enterprise in terms of its vision and mission and will form the essence of innovation strategy. It can be the innovative area in which a company makes in terms of innovative capacity using the best results. Another area that needs to be addressed is the establishment of rules making innovation strategy. The rules will operate the proposed expert system. An important part of innovation is to define require- ments, system of evaluation. The company must have clear criteria by which consider innovation strategy chosen, respectively according to which attributes to monitor its implementation over time. These attributes will form the basis for continuous evaluation of innovative strategies that will indicate the timeliness of an innovation strategy and the measures for its upgrade. 6 FUTURE WORK – REAL APPLICATIONS OF EXPERT SYSTEM FOR STRATEGIC PLANNING IN PRACTICE Expert systems are now using modern technology such as fuzz y logic, neural networks or genetic algorithms. Have their advantage in the field of strategic management, knowledge management, marketing, quality management and logistics. Are developed and used for quality assessment of the supplier (Kwong et al., 2002), to develop and formulation of marketing strategies (Dav- ies et al., 2002). Liebowitz (1998) emphasizes that expert systems must be an integral part of knowledge management. Only then can the senior management to obtain the necessary output for its decisions. Azadeh et al. (2009) deal with the creation of an expert system for strategic planning. Their proposed expert system allows the assessment of strengths, weaknesses, opportunities and threats. It is based on information inputs consisting of external and internal factors. The output of system is the proposal of strategies to eliminate bottlenecks and improve the system itself. Han (2003) dealt with a neural expert system approach to designing an intelligent strategic planning system. The proposed neural expert system could Providence ‘goal-seeking “functions, which prove to be very useful for unstructured decision-making problems, specifically in strategic planning. He created a prototype of this system called StratPlanner, which was experimentally tested on data from the Korean automobile industry. The results confirmed that the neural expert sys- tems approach is useful for performing competitive analyzes. Neural expert system prototype was designed to diagnose strategic problems and design appropriate strategic alternatives for the current competitive situation. Han (2003) dealt with using neural networks and expert systems techniques. In practice, it may meet with several applications of expert systems, which are designed to sup- port strategic planning. In most cases, however, they are focused solely on creating a marketing strategy. For example, Quick Insight is an expert system for the assessment of market opportuni- ties and business ideas. It evaluates the product (service) in comparison with its competitors and sets the probability of success of the product on the market. In evaluating and assessing us- joc_3-2010_v1c.indd 54 17.12.2010 19:05:26 �� ing the expertise on which a decision can be made in relation to the product (innovation, with- drawal, replacement ...). Quick Insight is used by more than 500 companies such as 3M, AT & T, Caterpillar, Equitable Life, General Electric, IBM, Pillsbury and many others [13]. The main benefit of this expert system is the possibility of rapid and accurate analysis of market potential for products and services of the company. The basic objects of analysis include product, serv- ice, price, market, competition and environment. Quick Insight helps measure the company’s ability to achieve its goals and plans. It allowed make the experiment when entering different input values, which result is in the generation of alternative strategies. As an output the system provides arguments. Each claim contains a list of key factors to be considered in the analysis. Expert system interprets the results in terms of good business models, including the Boston Consulting Group Matrix, GE Business Screen, Competitive Advantage, and over 30 more in a comprehensive evaluation report [13]. Business Insight is an expert system designed to assess and develop marketing and development of marketing strategies for larger businesses. It provides a unique insight into complex relation- ships between multiple business concepts. It examines individual allowable strategies that can be modified and improved before making a decision. An important part of expert system is knowledge base for product and services. Business Insight offers cooperation with experts in areas of strategic planning. It also allows comparison of alternative strategies based on their strengths and weaknesses, explore market opportunities and critical assessment of key success factors. It uses several methods such as strategic management (Michael Porter’s Five Competi- tive Forces, GE Business Strategy Matrix, Boston Consulting Group Matrix, SWOT Analysis, Product Life Cycle Analysis, Pricing Strategy and dozens of other models) [14]. The results are compared with a knowledge base of best business strategies and evaluated several marketing and business concepts relating to the problem. The results are reflected in the form of analysis to include a statement about the problems of the proposed marketing strategy. Business Insight generates a comprehensive report based on which managers can plan and justify their strategic choices [14]. Those cases mapped develop problems using expert systems in terms of strategic planning. Most applications are focuses exclusively on issues of competition, marketing and product portfolio. Innovation and innovative strategies has not yet been charted in detail. Therefore, our design expert system for creating an innovation strategy can be considered as a first step towards the use of expert systems in this area. The proposed expert system includes several new features which will allow the necessary information for policy-making. Our aim was to develop a proposal for a comprehensive expert system for creating an innova- tion strategy. Show a comprehensive proposal to site the elemental and relational. Name and define the basic building blocks of this system in detail and describe relationships outputs of the system. In the next steps in our future research, we want to focus on the compilation of this prototype system using appropriate programming language in collaboration with our col- leagues at the Faculty of Management and Informatics, Faculty of Electrical Engineering and the University of Žilina. We plan to set up a prototype to test in real conditions of medium or large company operating in the region of Žilina. joc_3-2010_v1c.indd 55 17.12.2010 19:05:26 Journal of Competitiveness | 2/2010�� 7 CONLUSION Innovations are currently a prerequisite for competitiveness. The economic crisis forced most businesses to savings in all business areas. On the other hand, it should be noted that the economic crisis for some time gone and come again to revive the economy if re-distribution markets. Successful companies are the ones that have implemented an innovative strategy to invest in R & D and innovation. The proposed expert system is currently focused on facili- tating the process of making innovation strategy. It provides senior managers with a tool to obtain information necessary for decision making. The proposed expert system is an extended architecture for a complete solution of innovative strategies. However, there is scope for its further expansion as well as its reduction. It depends on the type of business, entering the quantity of data and knowledge, number of possible sce- narios for the innovation strategy. Each firm is characterized by specific processes and areas. The author seeks to create a relatively universal expert system for creating an innovation strat- egy, which it is possible to modify (adding or reducing) the individual modules. References DUPAĽ, A., BARÁNEK, I., FÜZYOVÁ, Ľ. 1997. Manažment inovácií podniku. Bratislava: Ekonóm. 1997. ISBN 80-225-0841-1. HORŇÁKOVÁ, R., ZAUŠKOVÁ, A. 2008. Vyhodnotenie inovačného potenciálu a inovatívnosti vo vybraných malých a stredných podnikoch drevospracujúceho priemyslu. Zvolen: TU vo Zvolene. 2008. ISBN 978-80-228-1889-6. KELEMEN, J., LIDAY, M. 1996. Expertné systémy. Bratislava: vydavateľstvo SOFA. 1996. 201 s. ISBN 80-85752-32-8. KISLINGEROVÁ, E., NOVÝ, I. et al. 2005. Chování podniku v globalizujícím se prostředí. Praha: C. H. Beck 2005. ISBN 80-7179-847-9. KOVÁČ, M. 2007. Tvorba a riadenie inovácií. Košice: Technická univerzita v Košiciach. 2007. 121 s. ISBN 80-8073-690-1. MOLNÁR, P., DUPAĽ, A. 2002. Manažment inovácií podniku. Bratislava: Ekonóm, 2002. ISBN 80-225-1642-2. NÁVRAT, P. a kol. 2007. Umelá inteligencia. Bratislava: Slovenská technická univerzita v Bra- tislave. ISBN 978-80-227-2629-0. SPALEK, J., JANOTA, A., BALAŽOVIČOVÁ, M., PŘIBYL, P. 2005. Rozhodovanie a riadenie s podporou umelej inteligencie. Žilina: EDIS – vydavateľstvo ŽU. 2005. ISBN 80-8070-354-X. STRIŠŠ, J., VODÁK, J., KUBINA, M., JANKAL, R., SOVIAR, J. 2009. Marketingové riadenie. 2009. Žilina: EDIS. ISBN 80-8070-680-7. TIDD, J., BESSANT, J., PAVITT, K. 2007. Řízení inovací. Zavádění technologických, tržních a organizačních změn. Brno: COMPUTER PRESS, 2007. 549 s. ISBN 978-80-251-1466-7. ZAUŠKOVÁ, A., 2006. Riadenie inovácií. Zvolen: Technická univerzita vo Zvolene. 2006. 220 s. ISBN 80-228-1634-5. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. joc_3-2010_v1c.indd 56 17.12.2010 19:05:26 �� ZAUŠKOVÁ, A., LOUČANOVÁ, E. 2008. Inovačný manažment. Zvolen: Technická univer- zita. 2008. 91 s. ISBN 978-80-228-1953-4. Strategic Planning Software: Quick Insight - Business Idea Assessor. [online]. [2010-09-05]. Available on the Internet: http://www.planware.org/saleqi.htm. Strategic Planning Software: Business Insight - Strategic Plan Evaluator. [online]. [2010-09-05]. Avail- able on the Internet: http://www.planware.org/salebi.htm. AZADEH, A., SHARIFI, S., SABERI, M. 2009. Design and Implementation of a Human Centred Expert for Improvement of Strategic Planning in a Manufacturer of Construction Products. In: Australian Journal of Basic and Applied Sciences. 3 (3): 2447-2458. ISSN 1991-8178. LI, H., LI, Z., LI, X. L., HU, B. 2000. A production rescheduling expert simulation system. European Journal of Operational Research, 124(4), 283-297. HAN, J. H. 2003. A Neural Expert System with Goal Seeking Functions for Strategic Plan- ning. In: Proceedings of the Academy of Information and Management Sciences. Volume 7. Number 2. 5-10. Las Vegas. 2003. KWONG, C.K., IP, W.H., CHAN, J.W.K. 2002. Combining scoring method and fuzzy ex- pert systems approach to supplier assessment: a case study, Integrated Manufacturing Systems, 13(7): 512-519. LIEBOWITZ, J., 1998. Expert systems: an integral part of knowledge management, Kyber- netes: The International Journal of Systems & Cybernetics, 27(2): 170-175. LI, S., DAVIES, B., EDWARDS, J., KINMAN, R., DUAN, Y. 2002. Integrating group Del- phi, fuzzy logic and expert systems for marketing strategy development: the hybridization and its effectiveness, Marketing Intelligence and Planning, 20(5): 273-284. LENDEL, V., VARMUS, M. 2010. Innovative Potencial of Strategy. In: Vedecko-odborný časopis Ekonomika – Management – Inovace. Moravská vysoká škola Olomouc. Ročník II. 1/2010. str. 47 – 53. ISSN 1804-1299 Authors contacts: Ing. Viliam Lendel, PhD., Ing. Michal Varmus University of Žilina Faculty of Management Science and Informatics Department of Management Theories Univerzitná 8215/1, 010 26 Žilina Slovak Republic E-mail: viliam.lendel@fri.uniza.sk, e-mail: michal.varmus@fri.uniza.sk 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. joc_3-2010_v1c.indd 57 17.12.2010 19:05:26 
work_t2jhwiqqivhbpcpzotx2ebh3xm	----	Time-to-Event Prediction with Neural Networks Håvard Kvamme Time-to-Event Prediction with Neural Networks Thesis submitted for the degree of Philosophiae Doctor Department of Mathematics Faculty of Mathematics and Natural Sciences 2019 © Håvard Kvamme, 2019 Series of dissertations submitted to the Faculty of Mathematics and Natural Sciences, University of Oslo No. 2346 ISSN 1501-7710 All rights reserved. No part of this publication may be reproduced or transmitted, in any form or by any means, without permission. Cover: Hanne Baadsgaard Utigard. Print production: Reprosentralen, University of Oslo. Preface This thesis is submitted in partial fulfillment of the requirements for the degree of Philosophiae Doctor at the University of Oslo. The thesis is a collection of four papers with the common theme of time-to-event prediction with neural networks. Their chronological order is representative of my understanding of the field. Although I have a background in statistics from the Norwegian University of Science and Technology, I have always found the empirical mindset of machine learning to be closer to my way of thinking. It has, therefore, been interesting to work at the intersection of statistical survival analysis and a part of machine learning that originated in computer science. I am deeply grateful to my three supervisors Ørnulf Borgan, Ida Scheel, and Kjersti Aas, as they allowed me the freedom to explore my ideas, but managed to always be available with knowledge, assistance, and encouragement. In particular, I want to thank my main supervisor and main collaborator Ørnulf. His incredible knowledge of survival analysis has been invaluable for our work, and his rigor and attention to detail has caught more of my mistakes than I care to admit. I want to thank Ida for always having time for my questions, no matter how little time she had available. I want to thank Kjersti for having multiple interesting projects available for me when I started in 2016, and for getting us through the revisions and publishing process of our first paper. I am also grateful to my other co-authors Nikolai Sellereite and Steffen Sjursen. Additionally, I want to thank Nikolai for showing me the KKBox churn prediction data set which ended up being an important part of two other papers he was not a part of. During my time as a research fellow, I got to visit Trevor Hastie at Stanford University for six months, and I am very grateful to him for including me in his research group. His students, in addition to the rest of the Stanford statistics students and staff, made my stay very enjoyable. I am grateful to BigInsight for providing me with necessary funding for the stay, in addition to the funding for all other research-related trips. I want to thank all my colleagues at the Section of Statistics and Data Science at UiO and at the Norwegian Computing Center, in addition to Gudmund Horn Hermansen and Erlend Aune for inspiring conversations and conference trips. Finally, I want to thank my friends and family for making my time outside the university so great. In particular, I am grateful to Amalie for her patience, understanding, and for never questioning the time I spent on my research, even when it came at the expense of our time together. Håvard Kvamme Oslo, December 2019 i List of Papers Paper I Håvard Kvamme, Nikolai Sellereite, Kjersti Aas, and Steffen Sjursen. Predicting Mortgage Default Using Convolutional Neural Networks. Expert Systems with Applications, 102: 207–217, 2018. Paper II Håvard Kvamme, Ørnulf Borgan, and Ida Scheel. Time-to-Event Prediction with Neural Networks and Cox Regression. Journal of Machine Learning Research, 20(129): 1–30, 2019. Paper III Håvard Kvamme and Ørnulf Borgan. Continuous and Discrete-Time Survival Prediction with Neural Networks. Submitted for publication. Paper IV Håvard Kvamme and Ørnulf Borgan. The Brier Score under Administrative Censoring: Problems and Solutions. Submitted for publication. iii Contents Preface i List of Papers iii Contents v 1 Introduction 1 1.1 Neural Networks and Statistical Models . . . . . . . . . . . 2 2 Neural Networks 5 2.1 A Brief Historical Perspective . . . . . . . . . . . . . . . . 5 2.2 The Multilayer Perceptron . . . . . . . . . . . . . . . . . . 7 2.3 Training Networks . . . . . . . . . . . . . . . . . . . . . . . 8 2.4 Overfitting and Regularization . . . . . . . . . . . . . . . . 10 2.5 Convolutional Networks . . . . . . . . . . . . . . . . . . . . 12 3 Time-to-Event Prediction 17 3.1 Event-Time Modeling . . . . . . . . . . . . . . . . . . . . . 17 3.2 Regression Models . . . . . . . . . . . . . . . . . . . . . . . 22 3.3 Evaluation of Survival Estimates . . . . . . . . . . . . . . . 26 4 Summary of Papers 29 4.1 Paper I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 4.2 Paper II . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 4.3 Paper III . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 4.4 Paper VI . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 5 Discussion 33 5.1 The Survival Methods . . . . . . . . . . . . . . . . . . . . . 33 5.2 Can We Trust Individual Predictions? . . . . . . . . . . . . 34 5.3 Extensions of the Methods . . . . . . . . . . . . . . . . . . 35 5.4 Evaluation of Survival Estimates . . . . . . . . . . . . . . . 43 5.5 Time-to-Event Prediction and Machine Learning . . . . . . 45 Bibliography 47 Papers 56 v Chapter 1 Introduction As the title of this thesis suggests, we investigate how machine learning, and in particular neural networks, can be applied for time-to-event prediction. The topic of time-to-event prediction is a subfield of survival analysis predominately concerned with when in the future an event will occur. So, contrary to the rest of survival analysis research, we make little effort to understand why the event occurred. The event in question can be a mortgage default, as in Paper I; customers stopping their subscription to a service, as in Paper II; the failure time of a mechanical system; an actual death from a disease; or any other suitable event. In the context of this thesis, the term neural networks refers to a set of machine learning algorithms, also known as artificial neural networks, connectionist systems, or deep learning, and are not to be confused with the eponymous biological networks. So, when are we willing to trade some of the understanding of the event- time process for improved predictive performance? Often, we are interested in both. Take customer relationship management as an example. One would, of course, like to know why customers are churning (leaving the service). But to retain current customers, a viable strategy can be to simply identify customers that are likely to churn and provide them some discount or offer. Even if we are primarily interested in understanding the effect of the variables, valuable insights might be obtained by studying methods purely focused on prediction. If such a “prediction-focused” model has much better predictive performance than our “understandable” model, the latter is clearly not using the available information to its full potential and might require some modifications. This way a “prediction-focused” model can be used to benchmark our “understandable” model. The thesis essentially consists of three parts: the first is mainly about finding a good neural network model for a specific problem, the second part explores how methods from survival analysis can be combined with neural networks for improved predictive performance, and the third part is concerned with the evaluation of such predictions. The three topics represent the chronological progress of my research, as each of them was motivated by experiences encountered in the previous topic. In the first project, we created a neural network model for predicting mortgage defaults for Norway’s largest commercial bank. We later understood that such time-to-event problems could benefit from combining methodology from survival analysis and machine learning. Then, while developing such methodology, we found that some of the evaluation criteria did not behave as expected and could, possibly, give misleading results. So, naturally, we continued by investigating the performance metrics. To make one’s research accessible to others, it is import to provide the 1 1. Introduction necessary details in a usable format. Code has, therefore, been an essential addition to the papers. This has resulted in a python package that contains implementations of the proposed methods, data sets, simulations, and evaluation metrics in Papers II, III, and IV. The package is built on the popular deep learning framework PyTorch (Paszke et al., 2017) and is available at github.com/havakv/pycox. The thesis is structured as follows: In Chapter 2, I will give an introduction to neural networks in the context of our papers, as all the predictive methodology is built on neural networks. This chapter is most relevant for Paper I, as it is concerned with finding a network model that is suitable for time-series modeling. The three remaining papers, on the other hand, consider the networks as general function approximators that minimize some objective function and is more concerned with the survival methodology. In Chapter 3, I will give a brief introduction to the field of survival analysis while focusing on predictive modeling. I will introduce the statistical models that are the foundations of Papers II and III, in addition to similar methods in the literature. In short, most of the approaches are classical statistical models where a linear predictor function is replaced by a neural network. This chapter also addresses some of the common criteria for evaluating the predictive performance of time-to-event models. Here I will briefly discuss some of the problems that led us to the investigations in Paper IV. A summary of the four papers is given in Chapter 4. In Chapter 5, I will discuss some natural extensions of the proposed methods. Also, I will use the opportunity to address some shortcomings of the papers and proposed methods. 1.1 Neural Networks and Statistical Models A substantial part of this thesis is about improving the predictive performance of classical statistical models with methodology from machine learning. To illustrate the general approach, I will start by explaining how the Logistic Regression model for binary classification can be extended with neural networks. This should also provide the reader with the necessary terminology for understanding Chapter 2 on neural networks. Consider an individual i with covariates xi ∈ Rp and response yi ∈ {0, 1}. We assume that yi is an observation of the random variable Yi ∼ Bernoulli(πi), where the parameter πi depends on the covariates, πi = P(Yi = 1 |xi). Logistic Regression assumes that the relationship between xi and πi is given by the logistic function πi = σ(xi) = 1 1 + exp[−φ(xi)] , (1.1) where φ(xi), typically, is the linear predictor function φ(xi) = wT xi with parameters w ∈ Rp. For a data set of n individuals, we can estimate the 2 https://github.com/havakv/pycox Neural Networks and Statistical Models parameter values w by minimizing the negative log-likelihood of Bernoulli data loss = − n∑ i=1 ( yi log[σ(xi)] + (1 −yi) log [1 −σ(xi)] ) , (1.2) with respect to the parameters w. For a new individual j, π̂j = σ̂(xj) is the estimated probability of Yj = 1, and π̂j can be used to make a prediction of the response yj. The terminology in machine learning differs slightly from that of classical statistics. The response yi is often called a label or target, and the parameters w are called weights. The logistic function in (1.1) is referred to as the sigmoid function, and we rarely address its inverse; the link function that is so common in statistics. The data set is typically split into multiple subsets that serve different purposes, and the subset used to minimize the loss function (1.2) is called the training set. A model generalizes well if it has accurate predictions and we often refer to the topic as predictive performance. Consequently, we are typically not interested in minimizing the loss (1.2) with respect to the training set, but instead, we want the loss to be small for data we have not yet seen. This can be approached by considering the loss of a held-out subset, called a validation set, which essentially serves the same purpose as in cross-validation. Finally, we use a held-out test set to quantify the predictive performance of the model. If we now refer to the linear predictor φ(xi) = wT xi as a neural network, we have left the field of classical statistics and are instead doing machine learning. This is because a neural network is simply a parametric model such as φ(xi), though we are typically interested in more complex functions than wT xi. A neural network consists of a set of transformations of the covariates that are applied in sequence. Consider three such transformations denoted h1, h2, and h3. A network can then be defined as φ(xi) = h3(h2(h1(xi))). The transformations are typically referred to as layers, and each layer’s outputs are the values obtained from the transformation. We refer to the number of layers as the depth of the network, and the term deep learning refers to deep networks. There is, however, no formal depth requirement for a network to be considered deep. The structure of the network is often called the network architecture and refers to the general design of the network. Returning to the Logistic Regression, but with neural network terminology, we find that the linear predictor φ(xi) is a single-layer network (depth of one), consisting of a single dense, or fully-connected, layer. If we include the sigmoid function (1.1) as a part of the network, we can call σ(xi) a two-layer network where φ(xi) is a hidden layer. It is, however, more common to refer to transformations without parameters, such as the sigmoid, as activation functions, and we typically consider them part of a parametric layer. One can also consider the loss function (1.2) a part of the network, as it is conceptually no different from the other layers. This interpretation is mostly useful for illustrative purposes. The most standard neural network structure is the multilayer perceptron (MLP). It consists of multiple dense layers on the form h(z) = g(Wz), where g is 3 1. Introduction an activation function, W is a matrix of parameters or weights, and z represents the output of the previous layer. For layers h1, h2, and h3, with parameters W1 ∈ Rq×p, W2 ∈ Rr×q, w3 ∈ Rr, and sigmoid activation functions (applied element-wise), we have the network φ(xi) = wT3 σ ( W2 σ [ W1xi ]) . (1.3) If we replace the linear predictor in the Logistic Regression with this network, we will have a much more flexible model. We do, however, also have many more parameters. To train the model, we apply some version of the gradient descent algorithm. The gradients are obtained with the back-propagation algorithm which is, essentially, application of the chain rule of differentiation. It should now be clear that neural networks can be applied to many statistical models simply by replacing the linear predictor wT x of the classical statistical model with a network such as in (1.3). The network architecture of modern MLP’s differs slightly from the example in (1.3). This is one of the topics of Chapter 2, together with gradient descent, back-propagation, other layer types, and more. 4 Chapter 2 Neural Networks The field of neural networks is a branch of machine learning concerned with gradient-based optimization of models composed of sequences of transformations. It is commonly referred to as deep learning, artificial neural networks, or even artificial intelligence. One might argue that these terms actually represent different research areas or subcategories, a distinction that will be ignored in this thesis as I will simply use the term neural networks. Recently, neural networks have received much attention, both from academia and the industry. This has resulted in impressive results in areas such as image recognition (Krizhevsky et al., 2012; Szegedy et al., 2015; He et al., 2015a), image segmentation (Long et al., 2015; He et al., 2017), language modeling and language translation (Wu et al., 2016; Gehring et al., 2017; Vaswani et al., 2017; Radford et al., 2018, 2019), and image and audio generation (Goodfellow et al., 2014; Oord et al., 2016; Brock et al., 2018; Razavi et al., 2019). In part, the successes of neural networks are owed to the modularity of the back-propagation algorithm for automatic gradient computation (Kelley, 1960; Rumelhart et al., 1986). While back-propagation is essentially the chain rule of differentiation, in practice it allows for modular implementations of the desired transformations. This makes the networks very applicable, as it is easy to extend or change existing network structures. In the following, I will first give a brief overview of the development of neural networks, starting from the late 1950s. I will then present multilayer perceptrons (MLP’s), which are the most basic networks, and I will cover how they are trained and regularized. At the end of the chapter, I will describe convolutional neural networks, which are an important part of Paper I. 2.1 A Brief Historical Perspective The development of neural networks dates back to the perceptron by Rosenblatt (1958) and ADALINE by Widrow and Hoff (1960), both of which are binary classifiers with linear decision boundaries. Both methods were heavily criticized, especially by Minsky and Papert (1969), for their limited ability to express richer functions, and received limited academic attention. More than a decade later, Rumelhart et al. (1985, 1986) reestablished interest in neural networks by presenting the benefits of applying multiple perceptrons sequentially with non-linear functions in between them. This directly addressed the concerns raised by Minsky and Papert (1969). Furthermore, Hornik et al. (1989) established the notion that these multilayer perceptrons (MLP) are universal approximators in the sense that they can “approximate any Borel measurable function from one finite dimensional space to another to any desired degree of accuracy, provided 5 2. Neural Networks sufficiently many hidden units are available.” At the same time, LeCun (1989) introduced convolutional neural networks, or CNN’s, that are essentially a sparse version of the MLP with some parameters in the network forced to be identical (called weight sharing). This idea was tailored to the grid structure of images represented by pixels and was found to be quite successful in recognizing handwritten digits for the US Postal Service (LeCun et al., 1989). The networks were, however, very computationally expensive and somewhat unpractical to train. This period also includes the development of the generative models: Boltzmann machines (Ackley et al., 1985), Belief networks (Neal, 1992), and the Helmholtz machine (Dayan et al., 1995), in addition to Recurrent neural networks (Rumelhart et al., 1986) and Long Short-Term Memory networks (Hochreiter and Schmidhuber, 1997) for processing sequential data. At the time, the methods were quite computationally expensive, and it was problematic to train deeper networks. As a result, other methods such as Random Forests and Support-Vector Machines received more attention. Renewed interest in the field came with the breakthrough of Hinton et al. (2006) who were able to efficiently train deep Belief networks by greedy layer-wise pre-training. Shortly after, Bengio et al. (2007) and Poultney et al. (2007) showed that other deep network structures could be trained with the same strategy. Furthermore, they were able to show that both unsupervised pre-training and deeper networks improved generalization. With this new focus on depth, neural networks were rebranded as “deep learning” (Allen, 2015; Goodfellow et al., 2016), which marks the beginning of modern neural network research. The following period was marked by three important discoveries. The first was that graphical processing units (GPUs) could massively speed up the training process of the networks (Mohamed et al., 2009; Raina et al., 2009), which enabled researchers to faster explore new networks and allowed for more parameters in the networks. The second discovery was that better non-linear activation functions and better parameter initialization made the optimization problem much nicer (Jarrett et al., 2009; Glorot and Bengio, 2010; Nair and Hinton, 2010; Glorot et al., 2011). This had a large impact on both training time and generalization error. The third discovery was the importance of large labeled data sets. While this might seem obvious, researchers had, to some extent, been focusing on marginal improvements on smaller data sets. Faster computers and much “nicer” networks, enabled networks to be fitted to much larger data sets. The most prominent example of the importance of large data sets is likely the 2012 computer vision contest ILSVRC (Russakovsky et al., 2015). Here the convolutional network by Krizhevsky et al. (2012) achieved a top-5 error rate of 15.3% (compared to the second place of 26.1%). Their network architecture was actually quite similar to the architectures proposed a decade earlier by LeCun et al. (1998), but deeper and with the advancements discussed above. Although there had been increasing interest in neural network research since 2006, the entry by Krizhevsky et al. (2012) was, in fact, the only neural network entry to the competition that year. Their success, however, marks a turning point in the popularity of neural networks and the competition has since only been won 6 The Multilayer Perceptron by neural network entries. In the years after 2012, there have been numerous success stories of neural networks for a variety of applications, some of which were mentioned in the introduction to this chapter. For a more detailed historical overview, see, e.g., the blog series by Andrey Kurenkov1, or the book by Goodfellow et al. (2016). 2.2 The Multilayer Perceptron Multilayer perceptrons, or MLP’s, are the simplest and, arguably, the most fundamental network structures. MLP’s were briefly covered in the context of Logistic Regression in Section 1.1, but we will here provide a more general overview of the topic. To fit a standard MLP, we need input covariates x ∈ Rp, targets y ∈ Rm, and a differentiable loss function loss(φ(x), y), where φ(x) denotes the output of the MLP. The loss function can, for instance, be the mean squared error, or the negative log-likelihood of Bernoulli data (binary cross-entropy). An MLP consists of one or more dense, or fully-connected, layers of the form h(x) = Wx + b, where W ∈ Rq×p and b ∈ Rq are learnable parameters. It is common to refer to the parameters W as weights and b as biases. The biases b are sometimes excluded from the model, as we did in Section 1.1. For φ(x) to be a non-linear function of x, we need to add non-linear transformations to the network. These transformations are called activation functions and are commonly applied to the output of a dense layer. In fact, activation functions are so common that we typically consider them part of the layer, meaning we have the layer h(x) = g (Wx + b). The most commonly used activation function is the rectified linear unit, or ReLU, defined as g(z) = max{0,z}. There are, however, many proposed alternatives to this function, such as the ELU (Clevert et al., 2015), PReLU (He et al., 2015b), and SELU (Klambauer et al., 2017). Up till the late 2000s, it was more common to use the sigmoid function g(z) = 1/(1 + exp[−z]) or g(z) = tanh(z), as these more closely mimic the on/off behavior of a biological neural network. Both were, however, found to work rather poorly as the resulting gradients become very small when z is not close to 0. Nevertheless, they are still useful in other settings; the sigmoid is, for instance, commonly used to scale the final output of a network to be in [0, 1] for classification tasks. This brings us to the final layer of the MLP, the output layer. This is often a dense layer, but with an output activation that is suited to the loss function. For instance, for binary classification we can use the sigmoid activation, for multi-class classification we can use the softmax gj(z) = exp(zj)/ ∑m k=1 exp(zk), for j = 1, . . . ,m, and for regression one can simply use the identity g(z) = z. Conventions here vary, but sometimes the output activation is implicitly included in the loss function to ensure numerical stability. This is, however, most relevant for the implementation and usage of a method, and often less relevant for descriptive purposes. 1andreykurenkov.com/writing/ai/a-brief-history-of-neural-nets-and-deep-learning 7 http://www.andreykurenkov.com/writing/ai/a-brief-history-of-neural-nets-and-deep-learning/ 2. Neural Networks �(�) z3z2 z4z1 z5 x2x1 x3Input Hidden Output � = (�)ℎ1 � �(�) = (�)ℎ2 Figure 2.1: Simple MLP illustration with three covariates, a hidden layer with five units, and a single output. Figure 2.1 shows the typical illustration of a one-hidden-layer MLP. Although the figure does not provide much additional insights compared to simply writing the model as φ(x) = h2(h1(x)), it illustrates that it is common to think of a neural network as a computational graph. In particular, frameworks for working with neural networks, such as Theano, Tensorflow, and PyTorch, are typically built on this graph representation. The nodes represent values and the arrows represent the matrix multiplication with the parameters. As the activation functions are considered part of the layers, the hidden layer is actually h1(x) = g1 (Wx + b), and the output layer is h2(z) = g2 ( wT z ) . It can be shown that with a sufficiently “wide” hidden layer h(x), an MLP can approximate any function (Hornik et al., 1989). In practice, however, it is typically better to stack multiple hidden layers sequentially, giving networks such as φ(x) = h3(h2(h1(x))). 2.3 Training Networks The parameter values of a neural network φ(x) are generally found by minimizing the loss function with some version of gradient descent. In the example of the Logistic Regression in Section 1.1, the loss function was the negative log-likelihood of Bernoulli data, and we will generally consider negative log-likelihoods as loss functions in this thesis. Consider the network φ(x) with parameters denoted by w for simplicity, i.e., w contains all the parameters of all the layers. For a training set of n individuals, each with covariates xi, we want to obtain the minimizers w∗ = argmin w n∑ i=1 loss (φ(xi), yi) , (2.1) 8 Training Networks which, for a negative log-likelihood loss, corresponds to the maximum likelihood estimates. In practice, we find the parameter estimates by some version of gradient descent, meaning we calculate the partial derivatives of the loss with respect to the parameters, and move the parameter values in the negative direction of these partial derivatives. With α denoting a step size, a somewhat informal notation for this is w ← w −α∇w ( n∑ i=1 loss (φ(xi), yi) ) . By repeating this gradient step multiple times, we will eventually reach a minimum of the loss function. Usually, we perform each gradient step by only considering a small subset of the data set at a time. This subset is called a batch and is changed for every gradient update. In the neural networks literature, we refer to this procedure as stochastic gradient descent or just SGD, and it comes in many flavors such as Adagrad (Duchi et al., 2011), RMSprop (Hinton et al., 2012), and Adam (Kingma and Ba, 2014). Their distinction is not important for the understanding of this thesis. If we have a model with many parameters compared to the size of the data set, the optimal parameters in (2.1) are probably overfitted to the training set, meaning that they generalize poorly. It is, therefore, common to monitor the loss on a held-out validation set, and stop the optimization procedure when this validation loss stops improving. This is known as early stopping and is just one of many regularization techniques, a topic we will further investigate in Section 2.4. 2.3.1 Back-propagation For all but the simplest neural networks, it can be cumbersome to derive the gradients of all the parameters. To this end, back-propagation, an application of the chain rule of differentiation, was developed to efficiently obtain the gradients of the parameters in neural networks. Consider the neural network φ(x) = hm(hm−1(. . . (h1(x)))), consisting of the transformations z1 = h1(x), z2 = h2(z1), . . . , zm = hm(zm−1), each with parameters denoted w1, w2, . . . , wm. We require that each transformation in the network can calculate the partial derivatives of its output zk with respect to its parameters wk and its input zk−1. Informally written, this means that layer k needs to calculate ∂zk/∂wk and ∂zk/∂zk−1. By the chain rule of differentiation, the gradients of the parameters with respect to the loss can be obtained by ∂loss(φ(x), y) ∂wk = ∂loss(φ(x), y) ∂zk ∂zk ∂wk = ∂loss(φ(x), y) ∂zm ∂zm ∂zm−1 · · · ∂zk+1 ∂zk ∂zk ∂wk . 9 2. Neural Networks Obtaining the gradients of all parameters in the network is called a backward pass while calculating the loss is called a forward pass. It should now be clear that modifying or replacing a layer zk = hk(zk−1) only requires redefining ∂zk/∂wk and ∂zk/∂zk−1, and the rest of the implementation is left untouched. This modularity is rather powerful as it greatly simplifies experimentation with new layers and network structures. In fact, in Paper II and III, we propose methods based on standard neural network structures, but with new loss functions, which can be viewed as the final layer of the computational graph. The development of neural network methodology is further simplified by frameworks such as Tensorflow (Abadi et al., 2015), Keras (Chollet et al., 2015), PyTorch (Paszke et al., 2017), Torch (Collobert et al., 2002), Caffe (Jia et al., 2014), and Theano (Bergstra et al., 2010). Some of these also perform automatic differentiation, meaning one only needs to implement the transformations and not the partial derivatives. For a more in-depth discussion of back-propagation, see, e.g., Chapter 6.5 of the book by Goodfellow et al. (2016). 2.4 Overfitting and Regularization Neural networks are generally overparameterized. Rumelhart et al. (1986) found that networks with just enough connections to perform a given task tend to get stuck in local minima that are far worse than the global minima, and that “adding a few more connections creates extra dimensions in weight-space and these dimensions provide paths around the barriers that create poor local minima in the lower dimensional subspaces”. So, creating a neural network is not simply about finding the right number of parameters in the network. Typically, the best model, in terms of generalization error, is an appropriately regularized large model (Goodfellow et al., 2016). There is, of course, much literature on regularization, and in the following, we will investigate some of the most relevant. 2.4.1 Weight Decay Neural networks aside, regularization in statistics and machine learning is typically associated with penalization of the loss with respect to the size of the parameter values. Consider a model φ(x) with parameters w. The loss penalized by the L2 norm is then lossγ = loss(φ(x), y) + γ 1 2 ‖w‖22, where γ is a hyperparameter. The gradient updates with vanilla gradient descent will then take the form w ← w −α∇lossγ, or, equivalently, as ∇12‖w‖ 2 2 = w, w ← (1 −αγ)w −α∇loss(φ(x), y). (2.2) 10 Overfitting and Regularization For expression (2.2), we can interpret the gradient update as performing the regular non-penalized update w ← w − α∇loss(φ(x), y), but we shrink the weights w by a constant factor (1 − αγ) at every iteration. As a result, L2- regularization is referred to as weight decay in the neural network literature. In fact, generally, we do not calculate the penalty 12‖w‖ 2 2 directly but, instead, consider the weight decay a part of the optimization scheme. 2.4.2 Ensemble Learning and Dropout A common approach to improving the generalization of a model is to train an ensemble of models and combine their predictions. Bagging (Breiman, 1996) and Random Forest (Breiman, 2001) are two well-known methods built on this principle, and both average the predictions of many tree models. Averaging of multiple predictions reduces the variance and, consequently, improves the generalization error of methods with high variance and low bias. As neural networks are high variance, such ensemble strategies are applicable. There are many different approaches to ensemble learning for neural networks as one can bootstrap or subsample the training set (such as in bagging), use subsamples of the covariates, combine randomly initialized homogeneous network structures, etc. In Paper I, we found that averaging the predictions of networks fitted to subsets of the covariates (individual time series) substantially improved the performance over combining all covariates in a single network. We also averaged predictions from multiple randomly initialized networks for further improvements. Similar approaches are commonly used by the winning entries in prediction competitions such as the ILSVRC (Russakovsky et al., 2015). Dropout (Srivastava et al., 2014) can be thought of as a computationally efficient alternative to training an ensemble of neural networks and is one of the most commonly applied regularization strategies for deep neural networks. The idea is to randomly “drop” units from the network (represented by nodes in Figure 2.1), and thus train an ensemble of all possible subnetworks. With φγ(x) denoting a subnetwork γ obtainable with dropout, we want to minimize the expected loss over all possible subnetworks Eγ [loss(φγ(x), y)]. This is, however, typically intractable, so we instead minimize an unbiased estimate by sampling γ. In practice, dropout is applied by multiplying the output of a layer with a mask of iid 0/1 variables drawn from a Bernoulli(p) distribution (at every batch iteration), where p is a hyperparameter. The ensemble prediction can then be obtained by computing the average over a sample of masks φensemble(x) = 1 m m∑ k=1 φγk (x). A more common, less computationally expensive, alternative is to scale the output of each dropout layer by the dropout-probability p, meaning we compute an estimate of the expected output of that layer. This approach, called weight scaling, has limited theoretical justifications but works well in practice. 11 2. Neural Networks 2.4.3 Data Augmentation Possibly the simplest way to improve the generalization of a machine learning model is by increasing the size of the training set (assuming the quality of the data is reasonable). Obtaining more data if often hard or expensive, so data augmentation techniques instead attempt to artificially increase the size of the training set. The most basic form of data augmentation is to introduce noise to the covariates (Sietsma and Dow, 1991), in which case dropout can be viewed as an augmentation scheme. For image recognition, data augmentation has become an important consideration when training networks. There has been great success in artificially augmenting data sets by subjecting images to label-preserving transformations such as rotation, translation, scaling, and mirroring. Krizhevsky et al. (2012) emphasize this approach as an important part of their winning entry to the LSVRC-2012 competition (Russakovsky et al., 2015). In Paper I, we explore how data augmentation can be applied to time-series data by considering multiple overlapping windows of the same time series per individual. This is found to have a substantial impact on the performance of our models. 2.4.4 Other Approaches to Regularization Three other common regularization techniques are batch normalization (Ioffe and Szegedy, 2015), though its main objective is to improve optimization; cyclical learning rates (Loshchilov and Hutter, 2016), to find better local minima; and early stopping of the optimization routine based on a held-out validation set. Finally, an alternative approach to regularization is to change the structure of the neural networks. If one tailor the network architecture to a specific problem, one can potentially reduce the number of parameters without sacrificing the network’s ability to model the objective. Two such approaches are the use of sparse layers, meaning that many weights are set to zero; and the use of shared parameter values, meaning multiple parameters are fixed to be identical. The most prominent application of these two techniques is the development of the convolutional networks by LeCun (1989). 2.5 Convolutional Networks The name neural networks is from biology, as some of the first learning algorithms were modeled after the brain. It is, however, hard to argue that modern networks are more than loosely inspired by biological networks. Still, convolutional networks claim to capture some of the same properties as the primary visual cortex (LeCun et al., 2010; Goodfellow et al., 2016). Convolutional networks, convolutional neural networks, ConvNets, and CNN’s are all terms we use to describe neural networks where convolutional layers are a crucial part of the network architecture. LeCun (1989) is considered the inventor of CNN’s, though he was heavily inspired by the works of Fukushima (1980) and Lang (1988). The 12 Convolutional Networks Figure 2.2: The LeNet-5 architecture from LeCun et al. (1998) main motivation of CNN’s was to reduce the number of free parameters in a network by tailoring its architecture to the grid structure of an image represented by pixels. By enforcing sparsity and weight sharing, LeCun could accomplish this reduction without necessarily reducing the size of the network. In Figure 2.2, you can see the CNN proposed by LeCun et al. (1998) for recognizing digits (even though the figure use the character A as an example). In the figure, convolutions are represented by the small square on the input image. By applying this square to different parts of the input image one obtains the outputs in “C1: feature maps”. The small squares represent convolutional kernels and are often referred to as filters. This is because each filter is “looking” for a pattern in the image, and its output (C1: feature maps) is a map describing which parts of the input image that contain the filter pattern. For example, a filter might look for horizontal lines in the image, while a second filter might look for vertical lines. Figure 2.2 also contains other layers than the convolutional layers, but for now, we only note that the last layers are an MLP as described in Section 2.2. Images are not really part of this thesis, but in Paper I, we used CNN’s for time-series classification. Therefore, when describing the convolutions in detail, I will phrase them in terms of time series. The reader should, however, note that the following generalizes to images. Convolutional layers are, essentially, the discrete convolution operation known from mathematics (see, e.g., Goodfellow et al., 2016). We do, however, prefer to think of a convolution as a filter w applied to a patch x̃ of the time series x, where w is a vector of parameters. In this case, the convolution takes the form of the cross-correlation c = wT x̃. The same filter w is then applied to numerous patches from the time series, denoted x̃i, which results in a vector of outputs c. We can alternatively express this operation as c = Ax, where x is the full time-series and A is a sparse diagonal-constant matrix (Toeplitz matrix) with w descending along the diagonal. This means that the convolution can be viewed as a fully-connected layer, but with sparse weights and weight sharing. The new feature series c tells us something about the correlation between the time series and the filter w. In other words, c tells us where in the time series we find patterns close to that of the filter w. As w can only express a single 13 2. Neural Networks Figure 2.3: Illustration from Paper I of a convolutional layer applied to a time series. Only the first part of the full network is included. From left to right the figure shows a time series, a convolution operation with the resulting features, and pooling operation with the resulting features. pattern, we repeat the operation with q different filters {w1, w2, . . . wq}, giving the corresponding feature series {c1, c2, . . . cq}. This is illustrated in Figure 2.3 where we have a time series of length 365 and apply 32 different filters of size 9. Each filter w consists of randomly initialized parameters and is fitted in the same manner as the dense layers in the MLP’s. After the convolutional layers, we typically add an MLP that performs the final transformations. When trained, the different filters can extract various information that is useful for minimizing our loss function. When we build a neural network consisting of multiple convolutional layers applied in sequence (with non-linear activation functions in between), the network can learn a hierarchy of representations. This is hard to illustrate for time series, but it is quite common to view for images (see, e.g., Simonyan et al., 2013; Zeiler and Fergus, 2014). For images, we generally find that the first layers only learn to recognize simple patterns such as edges with different orientations, while the subsequent layers are able to detect more complex patterns such as corners and textures. The last layers can find whole objects such as faces, text, and animals. 2.5.1 Controlling the Feature Space Consider a time-series input of length m and a filter of length v. If we slide the filter across the time series one index at a time, the resulting output will be of length m−v + 1. For deeper networks (many convolutional layers), the output might become very short. Zero padding can be applied to prevent this shrinkage from occurring and works by adding zeros to the beginning and end of the time series, thus making the time series longer. The typical argument for applying zero padding is, however, that it preserves the information at the borders of the series. 14 Convolutional Networks On the other hand, it can be beneficial to reduce the length of the convolved features, as larger feature spaces require more computations and are more prone to overfitting. We can, therefore, use downsampled convolutions, or strided convolutions, where we move the filter by s indices at a time. This will reduce the size of the output by a factor 1/s. An alternative to downsampled convolutions is to use some version of pooling. Pooling summarizes an area of the input (i.e., input to the pooling operation) by computing a simple function such as the average or maximum. This creates a representation that is approximately invariant to small translations in the input and is, therefore, particularly useful when the existence of a pattern is more important than the exact location of that pattern. If we apply pooling to non-overlapping patches, we reduce the length of the output. This is illustrated in Figure 2.3, where we perform non-overlapping max pooling over 4 features (days) at a time and, consequently, reduce the feature length by a factor of 4. Correspondingly, an example of pooling for images is shown in Figure 2.2 where pooling is referred to as subsampling. 2.5.2 CNN’s for Time Series In Paper I, we applied CNN’s to the time-series classification problem of mortgage default. At that time, there were limited literature on CNN’s for time series, as recurrent neural networks (RNN’s), such as the long short-term memory networks (Hochreiter and Schmidhuber, 1997), were the preferred architectures for processing sequential data. This is, in a sense, surprising as the predecessor of the CNN’s, the time-delay neural networks (Lang, 1988) can be considered convolutions applied to time series. Although versions of RNN’s continue to be the most popular architecture for time series, CNN’s have the benefit of being more computationally efficient. Therefore, CNN’s for time series is an active research area, and there have been multiple examples CNN’s claiming to have better performance than RNN’s (Zhang et al., 2015; van den Oord et al., 2016; Bai et al., 2018; Elbayad et al., 2018). 15 Chapter 3 Time-to-Event Prediction Time-to-event prediction is a subfield of survival analysis concerned with prediction of future events. In classical statistics, survival analysis tends to focus on understanding how variables affect the event-time distribution. We here distinguish us from the majority of the statistical survival literature, as we will make no claim of understanding the survival process. Instead, we model the event-times as a “black box” that produces estimates of the event-time distribution. In the following, we will start with a basic introduction to survival analysis to make it clear how we can model the event-time distributions. Next, we will look at some classical regression models and see how they can be extended with neural networks. At the end of the chapter, we will discuss some of the typical evaluation criteria used for evaluating survival predictions. 3.1 Event-Time Modeling In this chapter, individuals are generally denoted by i and their covariates by xi. But in the following basic introduction to survival analysis, the covariates are disregarded for simplicity of notation and, instead, only a single individual is considered. Let f(t) be the probability density function (PDF) of the continuous-time event time T∗. In survival analysis, it is common to study the survival function S(t), which gives the probability of survival beyond time t S(t) = P(T∗ > t). The hazard rate h(t) is also quite commonly studied and, for continuous time, it is defined by the conditional probability h(t) = lim ∆t→0 P(t ≤ T∗ < t + ∆t |T∗ ≥ t) ∆t . This means that the hazard tells us something about the immediate likelihood of an event, given that the event has yet to occur, h(t) ∆t ≈ P(t ≤ T∗ < t + ∆t |T∗ ≥ t). Note that the density, the survival function, and the hazard rate are all representations of the same distribution, so knowing one is sufficient to obtain the other two. So, with F(t) and H(t) denoting the cumulative distribution 17 3. Time-to-Event Prediction 0 2 4 6 8 10 Time 0.000 0.025 0.050 0.075 0.100 0.125 0.150 Density: f(t) 0 2 4 6 8 10 Time 0.0 0.2 0.4 0.6 0.8 1.0 Survival: S(t) 0 2 4 6 8 10 Time 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Hazard: h(t) Figure 3.1: Illustration of the density, survival, and hazard for two individuals. function and cumulative hazard rate F(t) = ∫ t 0 f(u) du, H(t) = ∫ t 0 h(u) du, we have that f(t) = −S′(t) = h(t) S(t), S(t) = 1 −F(t) = exp[−H(t)], h(t) = f(t) S(t) = − d log [S(t)] dt . An illustration of these three representations of the event-time distribution is shown in Figure 3.1 for two individuals (blue and orange). Clearly, the three functions give contrasting views of the survival distribution, and it is reasonable to study all three. In the four papers of this thesis, the objective was to obtain estimates of the survival function and we will, therefore, consider estimation of the survival function the objective of this thesis as well. The hazard and density are, however, very useful means to this end. 3.1.1 Censoring and Truncation In Figure 3.1, one might notice that the survival functions S(t) do not drop below approximately 0.2, but it looks like both would continue to decrease if we extend the time axis beyond 10. If we are only able to make observations in the interval between the start and 10, we say that individuals are censored at time 10. Censoring, and the related topic truncation, do, however, extend far beyond this illustrative case, and is a fundamental part of survival analysis. Let C∗ denote a right-censoring time. C∗ is typically considered a random variable, though that may not always be the case. I will only use the notation 18 Event-Time Modeling Time Calendar Time End of Study Event Censoring    Figure 3.2: Example of administrative censoring (progressive type I) for five individuals (two censored and three observed events). The left figure represents the time scale in which we model the event times. The right figure represents the same individuals, but in the calendar time in which the data was collected. C∗ for the censoring time, and the reader is expected to understand from context whether or not it is a random variable. Let us first consider random censoring, where C∗ is a random variable with density fC∗ (t) and survival function SC∗ (t) defined in the same manner as for the event time T∗. A typical example of random censoring is when an individual i decides to no longer be part of a study and, consequently, we only know that the individual had still to experience the event at time C∗i . As both C ∗ i and T ∗ i are random variables, we are only able to observe the shorter of the two. We, therefore, define the observed time Ti and event indicator Di as follows, Ti = min{T∗i ,C ∗ i } (3.1) Di = 1{T∗i ≤ C ∗ i }. (3.2) In Papers II and III, we consider observations on this form. Administrative censoring, or progressive type I censoring, is an alternative where the right-censoring times C∗i are known for all individuals. So we know C∗i even when T ∗ i < C ∗ i . Administrative censoring often occurs in studies where individuals have different times of entry to the study (in calendar time) with a defined end date of the study. This is illustrated in Figure 3.2, where the left figure shows the event and censoring times in the study, while the right figure shows the calendar times at which the individuals were recorded. As an example, the figure could illustrate the time to mortgage defaults, such as in Paper I, where the mortgages are started on different dates and the end date of the study is today. The censoring time C∗i will, in this case, be the difference between the entry date and the end-of-study date. The observations are still given by (3.1) and (3.2), but we now also know all C∗i ’s. As the time of entry can be considered random, it is common to consider C∗i random in these data sets too. Also, researchers tend to disregard the additional information provided by C∗i and treat the problem as we did for non-administrative censoring. 19 3. Time-to-Event Prediction Only right-censored observations were considered in the papers that constitute this thesis. Left-truncation is, however, also quite common in survival data. As an example, for a left-truncation time Vi, we will only observe individual i if T∗i > Vi. This means that if T ∗ i < Vi we might not even know that the individual ever existing. Consequently, we are limited to model the distribution conditioned on the truncation time, or we can make sufficiently strong assumptions on the distribution of Vi to model the full event times. As an example of left-truncation, Klein and Moeschberger (2003) consider the age at death of residents at a retirement center. Individuals need to survive to a sufficient age to enter the retirement center, meaning all individuals that died before this time never entered the study. The topic of left-truncation will be revisited in Chapter 5. For a more extensive review of other types of censoring and truncation, see, e.g., Chapter 3 of the book by Klein and Moeschberger (2003). 3.1.2 Discrete-Time Survival Thus far, we have considered the time scale to be continuous. Sometimes, it can, however, be more appropriate with a discrete time scale. As an example of discrete-time survival data, we can consider the event of leaving a subscription service where each subscription payment lasts for a full month at a time. This means that the events can only occur after month number 1, 2, 3, etc., even though the decision to leave might happen sometime between the discrete time points. It is also quite common to have events that occur in continuous-time, but we are only able to make observations at discrete points in time. Let 0 = τ0 < τ1 < ... denote the discrete time scale. The probability mass function (PMF), the survival function, and the discrete hazard are defined as f(τj) = P(T∗ = τj), S(τj) = P(T∗ > τj) = ∑ k>j f(τk), h(τj) = P(T∗ = τj |T∗ > τj−1) = f(τj) S(τj−1) . Consequently, we also have that f(τj) = S(τj−1) −S(τj), S(τj) = [1 −h(τj)] S(τj−1) = j∏ k=1 [1 −h(τk)]. Assuming the same time scale for the censoring distribution, we denote the corresponding PMF and survival fC∗ (τj) and SC∗ (τj). In Paper III, we are concerned with discrete-time models, and we also address how the discrete-time models can be applied to continuous-time data. Both discretization schemes for continuous-time data and interpolation schemes for obtaining continuous-time survival estimates are considered. 20 Event-Time Modeling 3.1.3 The Likelihood for Right-Censored Event Times Continuing with the discrete time-scale and right-censored observations of the form Ti = min{T∗i ,C ∗ i }, Di = 1{T∗i ≤ C ∗ i }, we will derive the likelihood. For observations t and d, the probability of the observations is P(T = t,D = d) = P(T∗ = t,C∗ ≥ t)d P(T∗ > t,C∗ = t)1−d. If the observed censoring time c is deterministic, then d = 1 implies that t ≤ c, and d = 0 implies that t = c. This gives us P(T = t,D = d) = P(T∗ = t)d P(T∗ > t)1−d = f(t)d S(t)1−d. (3.3) It is, however, more common to consider random censoring C∗. If we assume that T∗ and C∗ are independent, we obtain the probability P(T = t,D = d) = [P(T∗ = t) P(C∗ ≥ t)]d [P(T∗ > t) P(C∗ = t)]1−d = [f(t) (SC∗ (t) + fC∗ (t))] d [S(t)fC∗ (t)] 1−d = [ f(t)d S(t)1−d ][ fC∗ (t) 1−d (SC∗ (t) + fC∗ (t)) d ] . Now, assuming f(t) does not depend on the parameters of fC∗ (t), we can view [fC∗ (t) 1−d (SC∗ (t) + fC∗ (t)) d] as a constant and only consider the proportional expression P(T = t,D = d) ∝ f(t)d S(t)1−d, which is identical to the expression with deterministic censoring in (3.3). So, in both cases, for individuals denoted by i, with covariates xi, observed times ti, and event indicators di, we consider the likelihood L = ∏ i f(ti |xi) di S(ti |xi) 1−di. (3.4) If we substitute the PMF with the PDF in (3.4), we obtain the corresponding likelihood for continuous-time data. As both the PMF and PDF are denoted f(ti |xi), the expression is unchanged. The derivations are also very close to that of (3.4). The likelihood for right-censored event times in (3.4) is the foundation for many survival methods, some of which will be addressed in the next sections. In fact, the survival methods presented Papers II and III are built on this likelihood. However, if we have, e.g., truncated or left-censored data, some alterations need to be made. This is somewhat outside the scope of this thesis, so I will refer the interested reader to Chapter 3.5 of the book by Klein and Moeschberger (2003). 21 3. Time-to-Event Prediction 3.2 Regression Models The application of regression models in survival analysis typically involves maximization of the survival likelihood (3.4). It is, however, more common to phrase the likelihood in terms of the hazard rate h(ti |xi), rather than the PMF/PDF f(ti |xi). Considering continuous-time data, with individuals denoted by i with covariates xi, observed times ti, and event indicators di, we can rewrite (3.4) to L = ∏ i h(ti |xi) di exp [−H(ti |xi)] . (3.5) Correspondingly, for discrete-time we have L = ∏ i  h(ti |xi)di [1 −h(ti |xi)]1−di ∏ j: τj<ti [1 −h(τj |xi)]   . (3.6) Many survival models have been proposed by defining a parameterization of h(t |x), but we will here only consider the most relevant for the work in Papers II and III. A more general overview is given in the book by Klein and Moeschberger (2003), and Tutz and Schmid (2016) provide an overview of discrete-time regression models. 3.2.1 Cox Regression The Cox proportional hazards model assumes that the continuous-time hazard rate is defined as h(t |x) = h0(t) exp[g(x)], (3.7) where h0(t) is a non-parametric function and g(x) is a parametric function, typically g(x) = βT x. Note that the hazard ratio between two individuals i and j is constant with respect to time h(t |xi) h(t |xj) = exp[g(xi) −g(xj)], which explains why we call it a proportional hazards model. We want to fit this model with the likelihood (3.5), but the likelihood can be made arbitrarily large by letting h0(t) be zero except for close to the observed event times {ti : di = 1} where we let it peek higher and higher. We, therefore, assume a cumulative baseline hazard H0(t) in the form of a step function with steps at the event times. With ηi denoting the increase in H0(t) at time ti, we have H0(t) = ∑ i: ti≤t ηi. 22 Regression Models Under this extended model, the likelihood (3.5) takes the form L = ∏ i (ηi exp[g(xi)]) di exp (−H0(ti) exp[g(xi)]) . (3.8) The maximizer of (3.8) with respect to ηi can be shown to be η̂i = di∑ j∈Ri exp[g(xj)] , where Ri = {j : tj ≥ ti} is the set of individuals still at risk at time ti. Correspondingly, we have Ĥ0(t) = ∑ i: ti≤t di∑ j∈Ri exp[g(xj)] , (3.9) which is known as the Breslow estimator. By inserting these maximizers into (3.8) we obtain the profile likelihood L =  ∏ i ( exp[g(xi)]∑ j∈Ri exp[g(xj)] )di exp [ − ∑ i di ] ∝ ∏ i: di=1 exp[g(xi)]∑ j∈Ri exp[g(xj)] . (3.10) The expression in (3.10) is known as the Cox partial likelihood and can be maximized to obtain parameter estimates of g(x). With such estimates ĝ(x), survival estimates can be obtained with Ŝ(t |x) = exp[−Ĥ(t |x)] = exp[−Ĥ0(t) ĝ(x)]. In summary, we have shown that Cox regression maximizes the survival likelihood for right-censored event-times (3.5), by considering a cumulative baseline hazard H0(t) in the form of a step function. Note that the estimates above do not directly provide us with estimates of the baseline hazard h0(t). However, we are rarely interested in these estimates anyway. For a more rigorous derivation of the partial likelihood as a profile likelihood, see Johansen (1983) or the book by Aalen et al. (2008, p. 204). The connection between the partial likelihood and the survival likelihood emphasizes the relationship of Cox regression to the other methods discussed in this chapter. It is, however, an uncommon way to introduce Cox regression, and a more standard approach is given in Paper II. The Cox model can be made more flexible by considering other versions of the parametric function g(x) in (3.7). For instance, we can let g(x) be parameterized by a neural network. Faraggi and Simon (1995) were the first to propose this approach, though with limited improvements over regular Cox regression. However, with modern deep neural networks, numerous papers have 23 3. Time-to-Event Prediction shown improved predictive performance of this approach (Katzman et al., 2018; Ching et al., 2018; Yousefi et al., 2017; Zhu et al., 2016; Zhu et al., 2017). In Paper II, we propose an even more flexible relative risk model without the proportionality constraint of the Cox model. In short, this proposed Cox-Time method is made possible by allowing the parametric function to depend on time, meaning we have g(t, x). This time-dependence makes the partial log- likelihood (3.10) quite computationally expensive, so the partial log-likelihood is approximated with techniques from nested case-control studies. A proportional hazards version of this method is also proposed which is referred to as CoxCC and Cox-MLP (CC) in Papers II and III. 3.2.2 Discrete-Time Regression What we today know as Cox regression was originally proposed in the paper by Cox (1972), one of the most cited statistical papers of all time. It is, however, less known that the same paper also proposed a discrete-time hazard parameterization by considering the sigmoid (inverse logit) of the linear predictor φj(x) = αj + βT x, h(τj |x) = 1 1 + exp[−φj(x)] . (3.11) Cox considered this an ad hoc modification of his continuous-time model and, therefore, proposed an estimation procedure analogous to the partial likelihood (3.10). Brown (1975) later showed that we can estimate the parameters of h(τj |x) in the same manner as a regular Logistic Regression. By defining the labels yij = 1{τj = ti,di = 1}, the likelihood (3.6) can be written as L = ∏ i ∏ j: τj≤ti h(τj |xi) yij [1 −h(τj |xi)] 1−yij , which we recognize as the Bernoulli likelihood. Brown (1975), therefore, named this approach the Logistic-Hazard model, but it is now also referred to as Logistic Discrete Hazard and Partial Logistic Regression. The logistic function (3.11) is still very commonly used, though there are alternatives such as the probit link, log link, and clog-log link (Tutz and Schmid, 2016, Chapter 3). There have also been proposed different versions of φj(x) and, to the best of my knowledge, Biganzoli et al. (1998) were the first to let φj(x) be parameterized by a neural network. Gensheimer and Narasimhan (2019) later parameterized φj(x) with modern deep neural networks. An alternative to the Logistic-Hazard is to instead parameterize the PMF f(τj |xi) in (3.4). This was the approach of the DeepHit method proposed by Lee et al. (2018). However, Lee et al. (2018) assume all individuals experience an event within a finite time scale τ1, . . . ,τm. This is an assumption we find unnecessary, so in Paper III we propose a similar version that allows for survival past time τm. 24 Regression Models In Paper III, we compare the Logistic-Hazard and PMF approach to parameterizing the survival likelihood. Furthermore, we investigate how they can be applied to continuous-time data by discretization of the time scale and interpolation for continuous-time survival estimates. 3.2.3 The Piecewise Exponential Model The piecewise exponential model was first proposed by Holford (1976) and later extended by Friedman (1982). For a defined set of times 0 = τ0 < τ1 < · · · < τm = τ, assume that the hazard is constant in each interval, meaning h(t |x) = ηj(x) for t ∈ (τj−1, τj]. Now, defining yij = 1{ti ∈ (τj−1, τj], di = 1} and ∆t̃ij =   τj − τj−1, if ti > τj ti − τj−1, if τj−1 < ti ≤ τj 0, if ti ≤ τj−1, the likelihood in (3.5) can be written as L = ∏ i ∏ j: τj≤ti [∆t̃ij ηj(xi)] yij exp [ −∆t̃ij ηj(xi) ] . This is proportional to the likelihood of independent Poisson-distributed variables yij with expectations µij = ∆t̃ij ηj(xi). So if we define ηj(x) = exp[φj(x)], we can use GLM software for the optimization procedure. Fornili et al. (2014) extended this methodology by parameterizing φj(x) with a neural network. In Paper III we revisit this method, but with modern frameworks for neural networks and some modifications to the link function. 3.2.4 Binary Classifiers Outside the survival literature, a common approach to event-time prediction is to frame the problem as binary classification. For a time τ, we want to estimate the probability of experiencing an event before this time, meaning we want to estimate P(T∗i ≤ τ). By disregarding individuals censored before time τ, meaning that we remove individuals with C∗i < min{T ∗ i ,τ}, we can fit a binary classifier to the remaining individuals with labels yi = 1{ti > τ}. This was our approach in Paper I for predicting mortgage defaults. If we want to obtain predictions for more than a single point in time, we can either repeat the steps above for multiple τj’s, or we can create a model that does this implicitly. The latter can be achieved with a neural network with an output for every τj. This approach was applied in Paper IV. Clearly, estimates obtained from such methods are dependent on the censoring distribution. However, for a negligible amount of censoring, the estimates of the survival function will be close to those of a method that accounts for censoring. Furthermore, if we are only interested in discrimination, meaning well-calibrated estimates is not important, the binary classifier can be a reasonable approach. 25 3. Time-to-Event Prediction There are of course some benefits to using a survival method in this case too, as the ranking can be affected by the censoring bias. But, at least in my experience, these differences are smaller than those of the evaluation metrics that also consider the calibration of the estimates. 3.3 Evaluation of Survival Estimates A substantial part of the machine learning literature is based on empirically verifying that a new method performs better than competing methodology for some evaluation criteria. So, when we combine survival methodology with methods from machine learning, it is crucial that we have suitable evaluation criteria for quantifying the performance of the proposed methods. A good evaluation metric for survival methodology needs to account for censored observations. This is quite problematic as popular metrics typically depend on the censoring distribution. If the scores are affected by the researcher’s assumptions on the censoring distribution, it is not obvious to what degree we can trust the results. To stress this point, consider the analogous situation in classification where there are missing labels (responses) in our test set. An evaluation metric, such as accuracy, will need to make some assumptions on the distribution of the missing labels. For example, we can assume that the labels are missing at random and simply disregard the missing labels. If, in fact, the probability of a label being missing is correlated with the label itself, the scores would be biased as the class proportions would be wrong. Returning to survival data, a seemingly obvious choice for evaluating survival predictions is to use the likelihood for right-censored event times in (3.4) and (3.5). The likelihood is, however, not very well suited for comparing estimates from different approaches, such as discrete and continuous time, and is, therefore, rarely used. An alternative metric, that we use in Papers II, III, and IV, is the Brier score. If there is no censoring, the Brier score is essentially the mean squared prediction error between our survival estimates Ŝ(t |xi) and the current state of the individuals 1{T∗i > t}, BS(t) = 1 n n∑ i=1 [ 1{T∗i > t}− Ŝ(t |xi) ]2 . (3.12) Therefore, the Brier score is a function of the time t. In practice, to obtain a single number for the score, we can calculate the integrated Brier score by numerically approximating the integral IBS = 1 t2 − t1 ∫ t2 t1 BS(u) du, for a suitable time interval [t1, t2]. 26 Evaluation of Survival Estimates 3.3.1 Inverse Probability of Censoring Weighting For right-censored event times, the Brier score can be approximated by a weighting scheme that preserves the expectation of the score. Graf et al. (1999) proposed to weight the Brier score (3.12) by the inverse probability of censoring, BSIPCW(t) = 1 n n∑ i=1 [ Ŝ(t |xi) 2 1{ti ≤ t, di = 1} ŜC∗ (ti − |xi) + [1 − Ŝ(t |xi)] 2 1{ti > t,} ŜC∗ (t |xi) , ] where ti− denotes the time right before ti. They proposed to use the Kaplan- Meier estimator to obtain censoring estimates ŜC∗ (t |x), meaning one assumes that the censoring distribution is independent of the covariates. Gerds and Schumacher (2006) extended the idea to covariate-dependent censoring estimates. The Brier score is obtained directly from the survival estimates. This makes the Brier score attractive for comparing any methods that produce survival estimates, which is the case for most survival methods. A drawback is that the scores are dependent on our estimate of the censoring distributions, which can be as difficult to estimate as the event-time distributions. If the censoring distributions substantially affect the scores, it can be hard to make any strong claims based on the results. In Paper IV, we discuss some of the issues when this scheme is applied to administratively censored times, in particular when covariates are very informative of the censoring times. The IPCW scheme is more general than the Brier score and can be applied to any suitable function. For example, in Paper II, we use the same weighting scheme for the binomial log-likelihood, as proposed by Graf et al. (1999). 3.3.2 Concordance The Harrell Jr et al. (1982) concordance index, or C-index, is a commonly used evaluation metric in the predictive survival literature. For uncensored data, it can be interpreted as the relative frequency of concordant pairs, meaning, for a given time t, it provides an estimate of C(t) = P(Ŝ(t |xi) < Ŝ(t |xj) |T∗i < T ∗ j ). (3.13) This is particularly useful for “one-to-one” corresponding models, defined by the property S(t |xi) < S(t |xj) ⇐⇒ S(s |xi) < S(s |xj) for all s and t, as the concordance is not dependent on the time, i.e., C(t) = C(s). The Cox proportional hazards model is an example of a “one-to-one” corresponding model, as the survival functions are given by S(t |x) = S0(t) exp[g(x)]. For models that are not “one-to-one”, it is somewhat common to replace the survival functions in (3.13) with summary statistics M(x), as this results in a single score that is not a function of the time. In this case, M(xi) is meant to represent the ranking of individual i’s “predicted event time”, and can, for example, be the integral of the survival function, M(x) = ∫ τ 0 S(t |x) dt. This approach might suppress some drawbacks of model restrictions, such as proportional hazards, but the interpretability of a single number for the 27 3. Time-to-Event Prediction concordance might be worth it. Alternatively, one can, of course, integrate the concordance scores in (3.13) with respect to the time t by some numerical approximation. For censored data, the concordance is dependent of the censoring distribution, making the scores harder to interpret. It is, however, still quite common to use it. To retain the original interpretation, Uno et al. (2011) and Gerds et al. (2013) proposed an IPCW concordance in a similar manner to that of the Brier score in Section 3.3.1. Antolini et al. (2005) proposed a version of the concordance that instead estimates Ctd = P(Ŝ(Ti |xi) < Ŝ(Ti |xj) |Ti < Tj, Di = 1), which they showed is equivalent to a weighted ROC AUC over time. This metric is attractive because it provides a single score (without integrating over time) while still considering the time dependence of the estimates. Furthermore, for “one-to-one” corresponding models it is equivalent to the Harrell Jr et al. (1982) concordance index. In Paper II, we show that by using the concordance of Antolini et al. (2005) for hyperparameter tuning of Random Survival Forests (Ishwaran et al., 2008), we obtain better predictive performance than that obtained by the Harrell Jr et al. (1982) concordance, both in terms of discrimination and calibration. 28 Chapter 4 Summary of Papers 4.1 Paper I Håvard Kvamme, Nikolai Sellereite, Kjersti Aas, and Steffen Sjursen. Predicting Mortgage Default Using Convolutional Neural Networks. Expert Systems With Applications, 102: 207–217, 2018. This paper is concerned with predicting defaults of private mortgages at Norway’s largest bank DNB. In contrast to similar work in the literature, we only use the state of each customer’s checking account, savings account, and credit card, while disregarding all other information about the customers. Hence, the aim of the paper is twofold: to show that the transaction data holds information useful for estimating the risk of a mortgage, and how to extract this information. No survival methodology was considered in this paper. Instead, the problem was approached as time-series classification. As was argued in Section 3.2.4, it is reasonable to apply a binary classifier to survival data if there are relatively few censored observations. Although this was not addressed in the paper, the binary classifier only considers mortgage defaults in the span of one year, meaning there was very little censoring. The data set consists of the daily balances of the checking accounts, the savings accounts, and the credit cards, in addition to the daily number of checking account transactions and the total amount transferred to the checking account. For each customer, a full year of data was used as covariates. Convolutional networks were applied to the times series and it was found better to fit individual networks for each of the time series, rather than fitting a single network with all the time series as input. At the time of this project, it was not very common to apply convolutional networks to time series, and most literature instead considered recurrent networks for this purpose. We found, however, that the convolutions performed better than such recurrent networks, and in the years since, we have seen an increase in the use of convolutions for these types of problems. The training set only consisted of 13,000 individuals. Hence, an augmentation technique was proposed. The scheme uses multiple “windows” of a customer’s time series, and was shown to improve predictive accuracy substantially. In relation to this, empirical studies found that increasing the size of the training set would likely improve the performance of the classifier. This suggests that we were not able to take advantage of all the information available in the time series. In conclusion, the proposed methodology was able to extract valuable 29 4. Summary of Papers information from the transaction time series. We do, however, believe that by combining this research with other data sources, further improvements would be possible. 4.2 Paper II Håvard Kvamme, Ørnulf Borgan, and Ida Scheel. Time-to-Event Prediction with Neural Networks and Cox Regression. Journal of Machine Learning Research, 20(129): 1–30, 2019 In this paper, we propose new ways to apply neural networks to survival data. The task is approached by extending the well-known Cox regression. There are numerous papers on parameterizing a Cox regression with neural networks, but we distinguish us from them by proposing a model with non-proportional hazards. This is achieved by allowing for interactions between the covariates and the time. In particular, the hazard model is h(t |x) = h0(t) exp [g(t, x)], where h0(t) is a non-parametric function and g(t, x) is parameterized by a neural network. This model is actually a relative risk model rather than a Cox model. Regardless, we refer to the method as Cox-Time, as the name is likely more familiar to most readers. Training the proposed Cox-Time method with the Cox partial log-likelihood would be very computationally expensive. By instead using an approximation of the partial log-likelihood from nested case-control studies, we are able to fit the model in reasonable time. Prediction with the Cox-Time method can, however, still be quite computationally expensive. Some suggestions were proposed to alleviate the cost, but further research is warranted. A proportional Cox model, fitted with the same approximation, is also proposed. The method is referred to as CoxCC or Cox-MLP (CC), where CC refers to the case-control approximation of the partial log-likelihood. The validity of the approximation was verified through simulation studies, and we found that, in general, very few control samples were needed. The proposed methodology was compared with other methods from the literature in a study of five data sets. Four of the data sets were obtained from the survival literature, each consisting of 1,000 to 10,000 individuals. The fifth was created with data made available by the Kaggle KKBox Churn Prediction Challenge. Although not originally intended for time-to-event prediction, we were able to create a data set of more than 2 million individuals. As neural networks typically excel with larger data sets, we wanted to have at least one large data set in our experiments. Our proposed Cox-Time method was found to perform well compared to competing methods, in particular in terms of the Brier score and binomial log-likelihood, both weighted by the inverse probability of censoring. 30 Paper III 4.3 Paper III Håvard Kvamme and Ørnulf Borgan. Continuous and Discrete-Time Survival Prediction with Neural Networks. Submitted for publication, 2019 This paper is, in some ways, a continuation of Paper II. However, the semi- parametric Cox model is abandoned, and we instead direct our attention to fully parametric approaches that directly optimize the likelihood of right-censored event-times. The explored methods either assume a discrete time-scale or consider the time-scale to be divided into intervals. Consequently, we need to address the issue of how discrete-time models can be applied to continuous-time data sets. Two discretization schemes are considered; one that splits the time scale into equidistant intervals, and one that uses the quantiles of the estimated event-time distribution to split the time scale. To obtain continuous-time predictions from the discrete-time models, we consider two interpolation schemes; one that assumes a constant density function in each interval (linear interpolation of survival estimates), and one that assumes constant hazard rates (exponential survival function in each interval). For smaller data sets, a coarse discretization grid can be beneficial as it reduces the number of parameters in the neural networks. This was verified in a simulation study. It was also found that the discretization grid based on quantiles was especially useful for the small data sets. Larger data sets, on the other hand, allows for finer discretization grids, meaning the differences between the two discretization and interpolation schemes decrease. It was, generally, found that the granularity of the discretization grid had a much larger impact on the predictive performance than the discretization and interpolation schemes. The paper compares the two approaches to discrete-time modeling discussed in Section 3.2.2, i.e., the Logistic-Hazard and the PMF method. In terms of predictive performance, the two methods are very close, but we find that the Logistic-Hazard generally gives more stable performance, while the PMF method is more sensitive to hyperparameter configurations. In this paper, we also propose a continuous-time method that assumes constant hazards in each interval. This is essentially the piecewise exponential model described in Section 3.2.3, but with modern neural networks and a softplus function replacing the regular exponential function in h(τj |x) = exp[φj(x)]. We find that its performance is competitive with the two other methods discussed in this paper, in addition to other methods found in the literature. In general, the methods described in this paper seem to perform slightly better than the Cox-Time methods proposed in Paper II. 31 4. Summary of Papers 4.4 Paper VI Håvard Kvamme and Ørnulf Borgan. The Brier Score under Ad- ministrative Censoring: Problems and Solutions. Submitted for publication, 2019 This paper is an investigation of the Brier score, which is commonly used to evaluate survival predictions. In short, the Brier score is the mean squared difference between the survival estimates and the state of the individuals at a given time denoted 1{T∗i > t}. For right-censored event-times, the Brier score can be weighted by the inverse probability of censoring, called IPCW. The paper shows that if the censoring times can be identified from the covariates, then the IPCW Brier score might be biased. Furthermore, this bias is equivalent to the bias of the binary classifiers in Section 3.2.4 and, therefore, benefits the binary classifiers. This is also verified through a simulation study where we show that a binary classifier can obtain better IPCW Brier scores than the true survival function. The paper also addresses that estimation of the censoring distribution can be problematic, especially if the distribution is dependent on the covariates. Under administrative censoring, meaning all censoring times are observed, we propose the administrative Brier score which does not require estimation of the censoring distribution and does not have the bias of IPCW scores when the censoring times can be identified from the covariates. This is verified through a simulation study. The KKBox churn prediction data set from Paper II is administratively censored and is, therefore, used as a case study in the paper. We find that the data set exhibits the traits we produced with simulations, meaning the covariates are likely to hold information about the censoring times of a subset of the customers. We show that the IPCW Brier score is substantially affected by the estimated censoring distribution. For the highest times, meaning the highest proportion of censored customers, binary classifiers outperform the Logistic- Hazard method. However, for the administrative Brier score, we get that the Logistic-Hazard performs better than the binary classifiers, as we would expect. 32 Chapter 5 Discussion In this thesis, we have investigated how neural networks can be used to improve the predictive performance of existing survival methods. As this topic is still somewhat unexplored, there are almost limitless extensions that could be addressed here. I have tried to choose a relevant subset and instead provide some more details. I will, also, use the opportunity to discuss some of the weaknesses of the discussed methodology, in addition to my experiences working with each of the proposed methods. Instead of addressing each of the four papers individually, the following is a general discussion of the machine learning methodology for time-to-event prediction. 5.1 The Survival Methods In Papers II and III, the predictions by various survival methods are compared using metrics such as the concordance and the Brier score. After working with numerous approaches to survival prediction, I would like to take the opportunity to address some of my experiences that are not as easily quantifiable and share some of my thoughts on the methods. Random Survival Forests (RSF) by Ishwaran et al. (2008) is by far the simplest method to use. As with other Random Forest methods, very little preprocessing of the data is required (if any at all), it is easy to find a good set of hyperparameters, and as it is fully non-parametric, there are no numerical issues. This means that it is easy to get good and stable performance. However, in my experience, RSF is very slow to train on larger data set and, in general, it took many times longer to train than the neural networks. I do not know if this is because the method is very computationally expensive or if it was just the implementation I used that was slow. The Cox-Time method generally performs better than the Cox methods with proportional hazards, but obtaining predictions is much more computationally expensive. This was discussed in Paper II. Comparing the proportional CoxPH method and CoxCC method, where the first minimizes the negative partial log-likelihood as described in Section 3.2.1 and the latter minimizes an approximation of this negative partial log-likelihood, we found that CoxCC has slightly less stable performance than CoxPH. It is also worth mentioning that the implementation of CoxCC is more complicated than that of CoxPH. However, the CoxCC method has a validation loss that is more comparable with the training loss which gives better insights to the training process. In Paper III, we found that the Logistic-Hazard and PC-Hazard were generally the two best-performing methods. Compared to Cox-Time, predictions are 33 5. Discussion much less computationally expensive and, compared to all the semi-parametric Cox methods, the implementation is much simpler. In fact, considering the implementation, the Logistic-Hazard method is probably the simplest, as it requires only a few lines of code to implement it in a deep learning framework. Also, it is likely one of the more intuitive methods for those without a background in survival analysis. The downside of the Logistic-Hazard and PC-Hazard is that they require discretization of the time scale, which introduces additional hyperparameters. RSF and the Cox-Methods, on the other hand, all handle the continuous time non-parametrically. 5.2 Can We Trust Individual Predictions? The methods discussed in Papers II and III give predictions in the form of survival estimates. These estimates were evaluated with multiple performance metrics such as the concordance index and the Brier score. However, these scores only verify that the survival estimates work well on average and do not address the individual predictions. As none of the methods produce uncertainty estimates, can we really trust the survival estimate for a single person? To illustrate this concern, I have created a simple simulation study with constant and covariate independent hazard, meaning that all individuals have the same survival function. I then include a single covariate that is a monotone function of the administrative censoring time. This is supposed to illustrate a scenario where we have covariates that contain a lot of information about the censoring times, just as in Paper IV. If we fit a Cox model with proportional hazards to this data set, it would simply learn to disregard this covariate and give survival estimates close to the true survival function. The non-proportional methods discussed in Papers II and III, on the other hand, have more difficulties with this estimation. In Figure 5.1, I have plotted the survival estimates of various methods, obtained by training on 1,000 samples from the simulation study. The vertical dotted red line in each plot gives the censoring time. Note that this censoring time represents the covariate used to obtain the survival predictions in each plot. The blue line in each plot represents the true survival function we want to estimate and, as it is independent of the censoring time, it is the same in all the six plots. As expected, for times smaller than the censoring time (left of the vertical red line), all methods are able to estimate the survival function reasonably well. However, the methods clearly struggle with understanding that all individuals have the same survival function. As a result, the methods are not concerned with the survival estimation after the censoring time, as this is never considered in the likelihood. At least this is my interpretation of the results. In a sense, these results are perfectly reasonable, as we should not predict for longer time horizons than what is available in the training set for a given set of covariates. However, usage of such methods clearly requires the practitioner to be aware of what time horizon that is applicable for each individual. This could 34 Extensions of the Methods 0.00 0.25 0.50 0.75 1.00 S( t) True Cox-Time Logistic-Hazard PC-Hazard PMF Censor-time 0 20 40 60 80 100 Time 0.00 0.25 0.50 0.75 1.00 S( t) 0 20 40 60 80 100 Time 0 20 40 60 80 100 Time Figure 5.1: Comparison of survival estimates by various survival methods. The only covariate is a monotone function of the censoring time, which is represented by the vertical red dotted lines. The blue line is the true survival function, which is independent of the covariate. Hence, the true survival function is identical in all six plots. be handled by also estimating the censoring distribution and only use survival estimates for times t that satisfies ŜC∗ (t |xi) > ξ, where ξ is some positive threshold. Estimation of the censoring distribution is, in itself, a survival problem censored by the event times. This means that we need to know the time horizons that are reasonable for our censoring estimates. The point I am trying to make is that one needs to be careful when using the individual survival estimates of these methods. Especially, as the methods can give survival predictions for much longer time horizons than what is reasonable for a particular individual. Uncertainty estimates would highly benefit these methods, though they may be hard to obtain. A reasonable starting point could be to consider estimates of the censoring distribution. 5.3 Extensions of the Methods The methods discussed in Papers II and III are somewhat basic, in the sense that they only consider right-censored event times. DeepHit (Lee et al., 2018) is the exception as it also handles multiple event types. Natural extensions of the methods would be to account for other types of censoring and truncation (left, right, interval), competing risks, dynamic prediction (time-dependent covariates), recurrent events, and multistate models. These are all topics addressed in the statistical survival literature. In the following, we will investigate how a few of these extensions can be approached for the methods in Papers II and III. We 35 5. Discussion will here use the same notation as presented in Chapter 3, meaning that, for individual i with covariates xi, the event time is T∗i , the censoring time is C ∗ i , and the potentially right-censored time is Ti with Di denoting if it is an event or censored observation. 5.3.1 Left Truncation Section 3.5 of the book by Klein and Moeschberger (2003) addresses how the survival likelihood changes with censoring and truncation of the types left, right, and interval. This is a good starting point for extending the methods that directly optimize the survival likelihood, such as Logistic-Hazard, the PMF method, and PC-Hazard. For the corresponding extension of Cox regression, see, e.g., Pan and Chappell (2002) and Rennert and Xie (2018). In the following, I will shortly address the case of left truncation, as it the most common of the truncation types. If we have left-truncated data, our estimates will be conditioned on these left-truncation times. So, for ν denoting a left-truncation time, we can estimate S(t |ν) = P(T∗ > t |T∗ > ν). Note that this conditioning is similar to the hazard specification. This makes the hazard suitable for working with left-truncated data. Consider a discrete time scale. In the same manner as in Section 3.1.3, we can derive the probability distribution of T and D, but now conditioned on the left-truncation time ν < t, P(T = t,D = d |T > ν) = P(T = t,D = d,T > ν) P(T > ν) = P(T = t,D = d) P(T∗ > ν) P(C∗ > ν) = [ f(t)d S(t)1−d S(ν) ][ fC∗ (t) 1−d (SC∗ (t) + fC∗ (t)) d SC∗ (ν) ] ∝ f(t)d S(t)1−d S(ν) . This means that the “regular” likelihood contribution of an individual (3.4), is scaled by the inverse of the survival probability at the left-truncation time. For an individual i with left truncation time νi, the discrete-time likelihood contribution is then Li = f(ti |xi) di S(ti |xi) 1−di S(νi |xi) . (5.1) Written in terms of the hazard, this is equivalent to Li = h(ti |xi) di [1 −h(ti |xi)] 1−di κ(ti)−1∏ j=κ(νi) [1 −h(τj |xi)] , 36 Extensions of the Methods where the τj’s denote the discrete time scale, and κ(ti) is the index corresponding to time ti, meaning ti = τκ(ti). These two ways of expressing the likelihood contribution can be used to extend the Logistic-Hazard and PMF method in Paper III to account for left-truncated event-times. For continuous time, the likelihood contribution can still be expressed by (5.1), but with f(ti |xi) denoting the event-time density function instead of the probability mass function. Written in terms of the hazards, the continuous-time likelihood contribution is Li = h(ti |xi) di exp [H(νi |xi) −H(ti |xi)] . This can be used to extend the PC-Hazard method in Paper III to account for left-truncated event times. For Cox regression, the left-truncated event times can be handled by modifying the risk set in Section 3.2.1 to Ri = {j : νj < ti ≤ tj}, as explained in the book by Aalen et al. (2008, p. 32). This should also work for the proposed Cox-Time and CoxCC in Paper II. For all the methods discussed here, the survival functions are monotone. Hence, the estimates of S(t |x) are dependent on the estimates up till time t. Left truncation can cause a data set to contain very few individuals that are followed from time zero, so if we estimate the survival from time zero, poor estimation at the lower times might ruin the later survival estimates. It is, therefore, more reasonable to estimate the conditional survival S(t |ν0, x) = P(T∗ > t |T∗ > ν0, x), (5.2) for a lower time ν0 at which the set of individuals we follow, {i : νi ≤ ν0}, is of reasonable size. 5.3.2 Dynamic Prediction The KKBox data set explored in Papers II and IV contains the subscription histories of KKBox’s customers. We created a simple data set for predicting churn, with stationary covariates obtained from each customer’s first subscription. This means that we considered the starting point t = 0 to be the time of the first subscription for each customer. In practice, however, it is probably more interesting to make dynamic predictions in calendar time, meaning that from a given calendar date t we want to predict the probability of churn for each time t + v. This is more in line with the original KKBox Kaggle competition, as they wanted churn predictions from a given date, and not from the time of each customer’s first subscription. Explicitly, for time-dependent covariates x(t) and a given date t, we want to predict time v into the future, conditioned on all information up till time t. Using the notation of (5.2), we can express the conditional survival as S(t + v |t, {x(u)}u≤t) = P(T ∗ > t + v |T∗ > t, {x(u)}u≤t). 37 5. Discussion One approach to this objective is to jointly model the event-time distribution and the distribution of the time-dependent covariates. While there is an extensive statistical literature on this topic (e.g., Ibrahim et al., 2010; Lawrence Gould et al., 2015), it requires estimation of the covariate process. A more straightforward approach is that of landmarking (Van Houwelingen, 2007; van Houwelingen and Putter, 2011). In its simplest form, one creates K data sets, one for each landmarking time sk, by truncating individuals with Ti < sk, and fit individual models to each of the K data sets. This way, model k can predict S(sk + v |sk, x(sk)) = P(T∗ > sk + v |T∗ > sk, x(sk)), where x(sk) is the covariates at time sk. This approach requires one model for each landmarking time sk, which is problematic for a large K. It is, therefore, desirable to combine all these K models into a single model. This can be approached by conditioning on the landmarking time in the sense that we consider sk a covariate. Also, it is probably better to let the covariates be some function of all previous information, ensuring that the covariate space is of constant dimensions p. We denote these covariates x̃(sk) ∈ Rp, as they are a function of the landmarking times and all previous information, giving x̃(sk) = g ( sk,{x(u)}u≤sk ) . Partly conditional modeling (Zheng and Heagerty, 2005) closely resembles landmarking, but considers the residual time v instead of t + v. This means that we reset the time at every landmark sk. For a set of landmarking times {sk} K k=1, we substitute individual i by a set of “fake” individual ik, one for each sk, with covariates x̃i(sk) and observed time vik = ti−sk. This means that we can fit the methods in Papers II and III with no alterations and their survival predictions will represent S(v |sk, x̃(sk)). This approach was used by Martinsson (2016) who considered v |sk, x̃(sk) to be Weibull distributed, and constructed x̃(sk) with a recurrent neural network over {x(u)}u≤sk . To my understanding, we are, technically, not restricted to use the same landmarking times for all individuals, and one could have a different set for each individual. However, I have not seen anyone address this in the literature. By following the reasoning of landmarking or partially conditional modeling, it should be possible to extend the methods in Papers II and III to work for dynamic prediction. Note, however, that this section should be considered a starting point, as I have not tested any of the approaches, and there may be issues I have failed to address. In Paper I, we proposed an augmentation scheme for artificially increasing the size of the mortgage default data set. In short, for each individual, we extracted data from multiple time periods in a similar manner to that of partial conditional modeling. However, the objective was to classify mortgage defaults within one year, and we only used covariates from one year of data. This means that, for each landmarking time sk, we estimated S(365 |sk, {x(u)} sk u=sk−365), where 38 Extensions of the Methods the data {x(u)}sku=sk−365 was processed by a convolutional network. Paper I could benefit from a partial conditional modeling approach, as the specific time of each mortgage default holds more information than the binary labels, and building on an established approach is likely more reasonable than our ad hoc data augmentation scheme. 5.3.3 Competing Risks In this thesis, I have only considered situations where there is only one event type for each data set. In some situations, there can, however, be multiple competing event types. For example, there are multiple diseases a patient can die from. We typically think of these types of events as competing processes where we only observe the event type with the shortest event time. The survival literature, therefore, refers to these situations as competing risks. In the following, we will investigate possible extensions of the methods in Papers II and III to enable them to handle multiple event types. To this end, we need the random variable R ∈{0, . . . ,nr} for denoting the event type with nr possible event types and R = 0 denoting a censored time. As before, the lower case letters represent the corresponding observations, meaning an observation consist of the tuple (ti,ri), and we let di be the event indicator di = 1{ri > 0}. The cumulative incidence functions Fr(t |x) = P(T∗ ≤ t,R = r |x). are commonly reported in competing risk studies and we will consider the estimation of these functions our objective. 5.3.3.1 Logistic-Hazard From the statistical literature on discrete-time survival analysis, we know that a natural extension of the Logistic-Hazard model is the Multinomial Logit Hazard model. The following is a derivation of the Multinomial Logit Hazard model, based on the presentation by Tutz and Schmid (2016). The event-specific hazard at the discrete time τj is hr(τj |x) = P(T∗ = τj, R = r |T∗ ≥ τj, x), and the overall hazard is h(τj |x) = P(T∗ = τj |T∗ ≥ τj, x) = nr∑ r=1 hr(τj |x). This gives the survival function and event probability S(τj |x) = P(T∗ > τj |x) = j∏ k=1 [1 −h(τk |x)] , fr(τj |x) = P(T∗ = τj,R = r |x) = hr(τj |x) S(τj−1 |x). 39 5. Discussion The cumulative incidence functions can be expressed as Fr(τj |x) = j∑ k=1 fr(τk |x). (5.3) For a function φ(x) ∈ Rm×nr , such as a neural network with an output node for each combination of the time and the event type, we define the event-specific hazards as hr(τj |x) = σjr(x) = exp[φjr(x)] 1 + ∑nr k=1 exp[φjk(x)] , and we let σj0(x) = P(T∗ > τj |T∗ ≥ τj, x) = 1 1 + ∑nr k=1 exp[φjk(x)] , denote the conditional survival. Note that σjr(x) is the softmax function with the statistical convention of a reference class φj0(x) = 0. Recall that di = 1{ri > 0}. Disregarding x for simpler notation and assuming that the censoring time is independent of the event time (given covariates), the probability of T and R is P(T = t, R = r) = [P(T∗ = t,R = r) P(C∗ ≥ t)]d [P(T∗ > t) P(C∗ = t)]1−d ∝ P(T∗ = t,R = r)d P(T∗ > t)1−d. This means the likelihood contribution of individual i is Li = fri (ti |xi) di S(ti |xi) 1−di (5.4) = hri (ti |xi) di [1 −h(ti |xi)] 1−di κ(ti)−1∏ j=1 [1 −h(τj |xi)], where κ(ti) denotes the index corresponding to time ti, meaning that ti = τκ(ti). By introducing labels yijr = { 1{ti = τj,ri = r}, if r > 0 1{ti > τj} di, if r = 0, such that ∑nr r=0 yijr = 1, we can write the negative log-likelihood as loss = − n∑ i=1 κ(ti)∑ j=1 nr∑ r=0 yijr log[σjr(xi)]. We recognizer this as the negative of the multinomial log-likelihood, also known as the categorical cross-entropy. This is one of the most commonly used loss functions for neural networks (e.g., for image classification) and should, therefore, be available in any deep learning framework. 40 Extensions of the Methods 5.3.3.2 PMF For the discrete-time PMF method, Lee et al. (2018) have already proposed a competing risks version expressed by the event probabilities fr(τj |x) = P(T∗ = τj,R = r |x). To my understanding, however, they assume S(τm |x) = 0. We will, therefore, investigate a version without this restriction. By considering observations on the discrete time scale τ1, . . . ,τm, we simply let m∑ j=1 nr∑ r=1 fr(τj |x) + S(τm |x) = 1, with S(τm |x) ∈ [0, 1]. In the same manner as in Paper III, for j ∈{1, . . . ,m} and r ∈{1, . . . ,nr}, we let the risk-specific PMF be defined by the softmax fr(τj |x) = exp[φjr(x)] 1 + ∑m k=1 ∑nr q=1 exp[φkq(x)] , S(τm |x) = 1 1 + ∑m k=1 ∑nr q=1 exp[φkq(x)] . The survival function can then be written as S(τj |x) = m∑ k=j+1 nr∑ r=1 fr(τk |x) + S(τm |x). The negative log-likelihood corresponding to (5.4) can now be written as loss = − n∑ i=1 di log[fri (ti |xi)] − n∑ i=1 (1 −di) log   m∑ k=κ(ti)+1 nr∑ r=1 fr(τk |xi) + S(τm |xi)   . This does not correspond to any commonly used loss function, meaning implementation requires some more work than for the competing risks version of the Logistic-Hazard. The cumulative incidence functions can be obtained from the risk-specific PMF fr(τj |x) with (5.3). 5.3.3.3 PC-Hazard For continuous time, the cumulative incidence functions take the form of the integral Fr(t |x) = ∫ t 0 fr(u |x) du = ∫ t 0 hr(u |x) S(u |x) du. (5.5) 41 5. Discussion The continuous-time PC-Hazard method in Paper III can be extended to competing risks by considering the cause-specific hazards. With κ(t) now denoting the interval of time t, meaning that t ∈ (τκ(t)−1, τκ(t)], the cause- specific hazard is constant in each interval hr(t |x) = ηκ(t) r(x). The continuous-time survival function can be expressed as (Prentice et al., 1978) S(t |x) = exp [ − nr∑ r=1 ∫ t 0 hr(u |x) du ] , and, by following the same line of reasoning as in Paper III, we can write the survival function as S(t |x) = nr∏ r=1  exp[−ηκ(t) r(x) (t− τκ(t)−1)] κ(t)−1∏ j=1 exp[−ηjr(x)∆τj]   , where ∆τj = τj − τj−1. The continuous-time likelihood contribution, corresponding to (5.4), is then Li = hri (ti |xi) di S(ti |xi) = ηκ(ti) ri (xi) di nr∏ r=1  exp[−ηκ(ti) r(xi) (t− τκ(ti)−1)] κ(ti)−1∏ j=1 exp[−ηjr(xi)∆τj]   , and the negative log-likelihood follows. We can, alternatively, represent the likelihood as a Poisson-likelihood. This is achieved by considering independent Poisson-distributed variables with expectation µijr = ∆t̃ijηjr, as described in Appendix C of Paper III. 5.3.3.4 Cox-Time The Cox proportional hazards model can be extended to competing risks by considering the cause-specific hazard hr(t |x) = h0r(t) exp[φr(x)], and the partial likelihood (Holt, 1978; Prentice et al., 1978) PL = nr∏ r=1 n∏ i=1 ( exp[φr(xi)]∑ j∈Ri exp[φr(xj)] ) 1{ri=r} , where Ri = {j : tj ≥ ti} and ri = 0 for censored times. Here φ(x) ∈ Rnr can be parameterized by a neural network with an output node for each event type. The cause-specific baseline hazards can be obtained with the Breslow 42 Evaluation of Survival Estimates estimator (3.9), and the cumulative incidence functions can be expressed in terms of the risk specific hazard increments (Cheng et al., 1998). The Cox-Time method proposed in Paper II can be extended to competing risks in the same manner by considering PLCC = nr∏ r=1 n∏ i=1 ( exp[φr(ti, xi)]∑ j∈R̃i exp[φr(ti, xj)] ) 1{ri=r} , where R̃i ⊂Ri. In summary, it should be possible to make competing risks version of the methods in Papers II and III. 5.4 Evaluation of Survival Estimates Neural networks can represent much more flexible models than classical statistical models, but this added flexibility can introduce new issues not really relevant for the simpler models. As an example, in Paper IV we discuss how the IPCW Brier score is biased for the KKBox churn data set. For simpler models than ours, this bias is likely not an issue, but by introducing neural networks, we find that these biases needed to be addressed. A substantial part of machine learning research is concerned with empirical evaluation of methods on various data sets. To some degree, for prediction, it does not matter if your method is reasonable as long as it achieves good scores. Combining machine learning methodology with survival analysis is somewhat problematic, as censoring and truncation cause the test set to have incomplete observations. We typically account for censoring and truncation by making assumptions about their distributions and relationship to the event- time distribution. When these assumptions are essential for the evaluation metric, the choices made by the researcher may substantially affect the results. Phrased differently, we have missing data in our test set and the score accounts for the missing data by assuming something about its distribution. Different assumptions may lead to different results. The IPCW Brier score in Paper IV can be biased if the covariates contain information that can identify some of the censoring times. Interestingly, the bias is in favor of the bias of a binary classifier that disregards individuals as they are censored. By choosing a method based on “best performance” in terms of the IPCW Brier score, we can end up choosing a binary classifier over a method that accounts for censored observations in a reasonable manner. For someone not familiar with survival analysis, the aforementioned binary classifier can seem to be a reasonable approach. To convince this audience to move to survival methodology, we need to be able to show the benefits of survival approaches. Rather than arguing that survival methodology is better suited for time-to-event prediction than binary classifiers, we need to provide evaluation metrics that verify this empirically. 43 5. Discussion 5.4.1 The Concordance Index Recall the that the concordance index is a metric for evaluating the discriminatory performance of survival estimates Ŝ(t |x). It aims at estimating the probability C(t) = P(Ŝ(t |xi) < Ŝ(t |xj) |T∗i < T ∗ j ), which can be somewhat problematic for right-censored observations. Uno et al. (2011) and Gerds et al. (2013) handle right censoring by inverse probability of censoring weighting (IPCW) of the concordance in a similar manner to that of the Brier score in Section 3.3.1. For administrative censoring with identifiable censoring times, this IPCW score might suffer from the same problems as the Brier score addressed in Paper IV. Therefore, a similar study to that of Paper IV is warranted and, depending on the findings, one could create an administrative concordance index by following the same line of reasoning as for the Brier score. The concordance by Antolini et al. (2005), discussed in Section 3.3.2, estimates Ctd = P(Ŝ(Ti |xi) < Ŝ(Ti |xj) |Ti < Tj, Di = 1), which has the benefit of giving the score as a single number instead of a function of t. Unfortunately, the Ctd is dependent on the censoring distribution, leaving the interpretation somewhat unintuitive. It would be desirable to find a scaling, such as IPCW, that accounts for the censoring distribution in a manner that lets us instead estimate P(Ŝ(T∗i |xi) < Ŝ(T ∗ i |xj) |T ∗ i < T ∗ j ). If one could manage to create such a score, it would also be reasonable to investigate the impact of administrative censoring with identifiable censoring times. 5.4.2 On the Administrative Brier Score The administrative Brier score, presented in Paper IV, is useful when the censoring times C∗i can be identified from the covariates xi. This is because the IPCW Brier score requires a censoring distribution with SC∗ (t |xi) > 0, but with sufficient information about the censoring time, we approach SC∗ (t |xi) = 1{C∗i > t}. In short, the administrative Brier score BSA(t) disregards all individuals with a censoring time C∗i < t. Note that we observe the censoring time for all individuals, meaning we remove both individuals that are censored and individuals that have experienced the event prior to time t. For future survival prediction, it is problematic to use covariates that can predict the censoring time, as this essentially means that the covariates of future individuals are different from the covariates in our training set. Extrapolation is generally hard, and we want to avoid it if possible. It is, therefore, reasonable to remove covariates that are closely connected to the censoring times. If this substantially changes the predictions, one has at least learned something about the non-stationary behavior of the data set. Note that if we are able to remove covariates closely related to the censoring time, we can use the IPCW Brier score instead, as we now have SC∗ (t |xi) > 0. 44 Time-to-Event Prediction and Machine Learning This will require estimation of the censoring distribution, but the interpretability of the IPCW score can be worth this disadvantage. If one prefers the IPCW scores, the administrative score can still be useful for understanding the data set and could be applied to verify that the IPCW scores are reasonable. 5.5 Time-to-Event Prediction and Machine Learning I believe there is great potential in combining methodology from machine learning with that of survival analysis. This research field is relatively new, and the approaches discussed in this thesis are only concerned with a minor subfield of survival analysis. There are still parts of the survival literature that could benefit from machine learning methodology and at the point of writing there are plenty of “low-hanging fruits”. For example, in this chapter, we have briefly addressed the topics of left-truncation, competing risks, dynamic prediction, and extensions of the concordance index. Returning to the mortgage defaults in Paper I, this problem could benefit from a survival approach to the default modeling. Interestingly, there are actually multiple levels of delinquencies before a customer defaults on their mortgage. By considering a multistate model appropriate for the delinquencies, one could provide the model with more information than the simple binary responses of the defaults. Such multistate models have not been addressed in this thesis and would be an interesting direction for future research. For the customer churn data sets in Papers II and IV, the customers can start and stop their subscription to KKBox multiple times. This means that individuals can move between two states (customer and non-customer) multiple times. This type of data is reasonable to consider with recurrent events models or multiple-spell modeling (Tutz and Schmid, 2016, Chapter 10). This is also a widely studied topic in survival analysis that has received limited attention with machine learning methodology. A substantial part of machine learning research compares methods by empirical evaluation on a held-out test set. For event-time prediction, there are, however, very few large data sets and the current literature generally relies on somewhat small data sets. The resulting variance in the evaluation scores makes it hard to compare methods. In Paper II, we contribute to this issue by modifying the existing KKBox Kaggle data set, resulting in an event-time data set with more than 2 million individuals. We do, however, need more large data sets for the further advancement of the field. I hope that the research I have been a part of can help to draw some attention to this field, and maybe inspire someone to further improve on the methodology for time-to-event prediction. 45 Bibliography Odd Aalen, Ørnulf Borgan, and Håkon Gjessing. Survival and Event History Analysis: A Process Point of View. Springer Science & Business Media, 2008. Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. David H Ackley, Geoffrey E Hinton, and Terrence J Sejnowski. A learning algorithm for Boltzmann machines. Cognitive Science, 9(1):147–169, 1985. Kate Allen. How a Toronto professor’s research revolutionized artificial intelligence. The Star, Apr 2015. Laura Antolini, Patrizia Boracchi, and Elia Biganzoli. A time-dependent discrimination index for survival data. Statistics in Medicine, 24(24):3927–3944, 2005. Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271, 2018. Yoshua Bengio, Pascal Lamblin, Dan Popovici, and Hugo Larochelle. Greedy layer-wise training of deep networks. In Advances in Neural Information Processing Systems, pages 153–160, 2007. James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: a cpu and gpu math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), volume 4. Austin, TX, 2010. Elia Biganzoli, Patrizia Boracchi, Luigi Mariani, and Ettore Marubini. Feed forward neural networks for the analysis of censored survival data: a partial logistic regression approach. Statistics in Medicine, 17(10):1169–1186, 1998. Leo Breiman. Bagging predictors. Machine Learning, 24(2):123–140, 1996. Leo Breiman. Random forests. Machine Learning, 45(1):5–32, 2001. Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 47 Bibliography Charles C Brown. On the use of indicator variables for studying the time- dependence of parameters in a response-time model. Biometrics, 31(4):863–872, 1975. S C Cheng, Jason P Fine, and L J Wei. Prediction of cumulative incidence function under the proportional hazards model. Biometrics, 54:219–228, 1998. Travers Ching, Xun Zhu, and Lana X Garmire. Cox-nnet: An artificial neural network method for prognosis prediction of high-throughput omics data. PLoS Computational Biology, 14(4):e1006076, 2018. François Chollet et al. Keras. https://keras.io, 2015. Djork-Arné Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015. Ronan Collobert, Samy Bengio, and Johnny Marithoz. Torch: A modular machine learning software library, 2002. David R Cox. Regression models and life-tables. Journal of the Royal Statistical Society. Series B (Methodological), 34(2):187–220, 1972. Peter Dayan, Geoffrey E Hinton, Radford M Neal, and Richard S Zemel. The Helmholtz machine. Neural Computation, 7(5):889–904, 1995. John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011. Maha Elbayad, Laurent Besacier, and Jakob Verbeek. Pervasive attention: 2d convolutional neural networks for sequence-to-sequence prediction. arXiv preprint arXiv:1808.03867, 2018. David Faraggi and Richard Simon. A neural network model for survival data. Statistics in Medicine, 14(1):73–82, 1995. Marco Fornili, Federico Ambrogi, Patrizia Boracchi, and Elia Biganzoli. Piecewise exponential artificial neural networks (PEANN) for modeling hazard function with right censored data. In Computational Intelligence Methods for Bioinformatics and Biostatistics, pages 125–136, 2014. Michael Friedman. Piecewise exponential models for survival data with covariates. The Annals of Statistics, 10(1):101–113, 1982. Kunihiko Fukushima. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. Biological Cybernetics, 36(4):193–202, 1980. 48 https://keras.io Bibliography Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1243–1252. JMLR. org, 2017. Michael F Gensheimer and Balasubramanian Narasimhan. A scalable discrete- time survival model for neural networks. PeerJ, 7:e6257, 2019. Thomas A Gerds and Martin Schumacher. Consistent estimation of the expected Brier score in general survival models with right-censored event times. Biometrical Journal, 48(6):1029–1040, 2006. Thomas A Gerds, Michael W Kattan, Martin Schumacher, and Changhong Yu. Estimating a time-dependent concordance index for survival prediction models with covariate dependent censoring. Statistics in Medicine, 32(13):2173–2184, 2013. Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pages 249–256, 2010. Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Geoffrey Gordon, David Dunson, and Miroslav Dudík, editors, Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15, pages 315–323, 2011. Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z Ghahramani, M Welling, C Cortes, N D Lawrence, and K Q Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 2672–2680. Curran Associates, Inc., 2014. Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. Erika Graf, Claudia Schmoor, Willi Sauerbrei, and Martin Schumacher. Assessment and comparison of prognostic classification schemes for survival data. Statistics in Medicine, 18(17-18):2529–2545, 1999. Frank E Harrell Jr, Robert M Califf, David B Pryor, Kerry L Lee, and Robert A Rosati. Evaluating the yield of medical tests. Journal of the American Medical Association, 247(18):2543–2546, 1982. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2015a. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision, pages 1026–1034, 2015b. 49 Bibliography Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961–2969, 2017. Geoffrey E Hinton, Simon Osindero, and Yee-Whye Teh. A fast learning algorithm for deep belief nets. Neural Computation, 18(7):1527–1554, 2006. Geoffrey E Hinton, Nitish Srivastava, and Kevin Swersky. Neural networks for machine learning: Lecture 6a: Overview of mini-batch gradient descent, 2012. Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997. Theodore R Holford. Life tables with concomitant information. Biometrics, 32 (3):587–597, 1976. John D Holt. Competing risk analyses with special reference to matched pair experiments. Biometrika, 65(1):159–165, 1978. Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural Networks, 2(5):359–366, 1989. Joseph G Ibrahim, Haitao Chu, and Liddy M Chen. Basic concepts and methods for joint models of longitudinal and survival data. Journal of Clinical Oncology, 28(16):2796, 2010. Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of The 32nd International Conference on Machine Learning, pages 448–456, 2015. Hemant Ishwaran, Udaya B Kogalur, Eugene H Blackstone, and Michael S Lauer. Random survival forests. Annals of Applied Statistics, 2(3):841–860, 2008. Kevin Jarrett, Koray Kavukcuoglu, Marc’Aurelio Ranzato, and Yann LeCun. What is the best multi-stage architecture for object recognition? In 2009 IEEE 12th International Conference on Computer Vision, pages 2146–2153, Sep. 2009. Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM International Conference on Multimedia, pages 675–678. ACM, 2014. Søren Johansen. An extension of Cox’s regression model. International Statistical Review, 51(2):165–174, 1983. Jared L Katzman, Uri Shaham, Alexander Cloninger, Jonathan Bates, Tingting Jiang, and Yuval Kluger. Deepsurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology, 18(1):24, Feb 2018. 50 Bibliography Henry J Kelley. Gradient theory of optimal flight paths. Ars Journal, 30(10): 947–954, 1960. Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. In Advances in Neural Information Processing Systems, pages 971–980, 2017. John P Klein and Melvin L Moeschberger. Survival Analysis: Techniques for Censored and Truncated Data. Springer, New York, 2. edition, 2003. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, pages 1097–1105, 2012. Kevin J Lang. The development of the time-delay neural network architecture for speech recognition. Technical Report CMU-CS-88-152, 1988. A Lawrence Gould, Mark Ernest Boye, Michael J Crowther, Joseph G Ibrahim, George Quartey, Sandrine Micallef, and Frederic Y Bois. Joint modeling of survival and longitudinal non-survival data: current methods and issues. report of the dia bayesian joint modeling working group. Statistics in Medicine, 34 (14):2181–2195, 2015. Yann LeCun. Generalization and network design strategies. Technical Report CRG-TR-89-4, University of Toronto, 1989. Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and Lawrence D Jackel. Backpropagation applied to handwritten zip code recognition. Neural Computation, 1(4):541–551, 1989. Yann LeCun, Léon Bottou, Yoshua Bengio, Patrick Haffner, et al. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11): 2278–2324, 1998. Yann LeCun, Koray Kavukcuoglu, and Clément Farabet. Convolutional networks and applications in vision. In Proceedings of 2010 IEEE International Symposium on Circuits and Systems, pages 253–256. IEEE, 2010. Changhee Lee, William R Zame, Jinsung Yoon, and Mihaela van der Schaar. Deephit: A deep learning approach to survival analysis with competing risks. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431–3440, 2015. Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 51 Bibliography Egil Martinsson. WTTE-RNN : Weibull Time To Event Recurrent Neural Network. Master’s thesis, Chalmers University Of Technology, 2016. Marvin Minsky and Seymour Papert. Perceptrons: An introduction to computational geometry. MIT press, 1969. Abdelrahman Mohamed, George Dahl, and Geoffrey Hinton. Deep belief networks for phone recognition. In Nips Workshop on Deep Learning for Speech Recognition and Related Applications, volume 1, page 39. Vancouver, Canada, 2009. Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pages 807–814, 2010. Radford M Neal. Connectionist learning of belief networks. Artificial Intelligence, 56(1):71–113, 1992. Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. Wei Pan and Rick Chappell. Estimation in the Cox proportional hazards model with left-truncated and interval-censored data. Biometrics, 58(1):64–70, 2002. Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. In NIPS Autodiff Workshop, 2017. Christopher Poultney, Sumit Chopra, Yann L Cun, et al. Efficient learning of sparse representations with an energy-based model. In Advances in Neural Information Processing Systems, pages 1137–1144, 2007. Ross L Prentice, John D Kalbfleisch, Arthur V Peterson Jr, Nancy Flournoy, Vern T Farewell, and Norman E Breslow. The analysis of failure times in the presence of competing risks. Biometrics, 34(4):541–554, 1978. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. URL https://s3-us-west-2. amazonaws. com/openai-assets/researchcovers/languageunsupervised/language understanding paper. pdf, 2018. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, 1(8), 2019. Rajat Raina, Anand Madhavan, and Andrew Y Ng. Large-scale deep unsupervised learning using graphics processors. In Proceedings of the 26th Annual International Conference on Machine Learning, pages 873–880. ACM, 2009. 52 Bibliography Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high- fidelity images with vq-vae-2. arXiv preprint arXiv:1906.00446, 2019. Lior Rennert and Sharon X Xie. Cox regression model with doubly truncated data. Biometrics, 74(2):725–733, 2018. Frank Rosenblatt. The perceptron: a probabilistic model for information storage and organization in the brain. Psychological Review, 65(6):386, 1958. David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning internal representations by error propagation. Technical report, California Univ San Diego La Jolla Inst for Cognitive Science, 1985. David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. Nature, 323:533–536, 1986. Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211–252, 2015. Jocelyn Sietsma and Robert JF Dow. Creating artificial neural networks that generalize. Neural Networks, 4(1):67–79, 1991. Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15:1929–1958, 2014. Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Computer Vision and Pattern Recognition (CVPR), 2015. Gerhard Tutz and Matthias Schmid. Modeling Discrete Time-to-Event Data. Springer, 2016. Hajime Uno, Tianxi Cai, Michael J Pencina, Ralph B D’Agostino, and L J Wei. On the c-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data. Statistics in Medicine, 30(10):1105–1117, 2011. Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alexander Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. 53 Bibliography Hans van Houwelingen and Hein Putter. Dynamic Prediction in Clinical Survival Analysis. CRC Press, 1st edition, 2011. Hans C Van Houwelingen. Dynamic prediction by landmarking in event history analysis. Scandinavian Journal of Statistics, 34(1):70–85, 2007. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998–6008, 2017. Bernard Widrow and Marcian E Hoff. Adaptive switching circuits. Technical report, Stanford Univ Ca Stanford Electronics Labs, 1960. Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016. Safoora Yousefi, Fatemeh Amrollahi, Mohamed Amgad, Chengliang Dong, Joshua E Lewis, Congzheng Song, David A Gutman, Sameer H Halani, Jose Enrique Velazquez Vega, Daniel J Brat, et al. Predicting clinical outcomes from large scale cancer genomic profiles with deep survival models. Scientific Reports, 7(11707), 2017. Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, pages 818–833, 2014. Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, pages 649–657, 2015. Yingye Zheng and Patrick J Heagerty. Partly conditional survival models for longitudinal data. Biometrics, 61(2):379–391, 2005. Xinliang Zhu, Jiawen Yao, and Junzhou Huang. Deep convolutional neural network for survival analysis with pathological images. In 2016 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 544–547, 2016. Xinliang Zhu, Jiawen Yao, Feiyun Zhu, and Junzhou Huang. Wsisa: Making survival prediction from whole slide histopathological images. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6855–6863, 2017. 54 Papers I 57 Expert Systems With Applications 102 (2018) 207–217 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Predicting mortgage default using convolutional neural networks Håvard Kvamme a , ∗, Nikolai Sellereite b , Kjersti Aas b , Steffen Sjursen c a Department of Mathematics, University of Oslo, Niels Henrik Abels hus Moltke Moes vei 35, Oslo 0851, Norway b Statistical Analysis, Machine Learning and Image Analysis, Norwegian Computing Center, Gaustadalleen 23a, Oslo 0373, Norway c Group Risk Modelling, DNB ASA, Dronning Eufemias gate 30, Oslo 0191, Norway a r t i c l e i n f o Article history: Received 15 August 2017 Revised 17 February 2018 Accepted 18 February 2018 Available online 19 February 2018 Keywords: Consumer credit risk Machine learning Deep learning Mortgage default model Time series a b s t r a c t We predict mortgage default by applying convolutional neural networks to consumer transaction data. For each consumer we have the balances of the checking account, savings account, and the credit card, in addition to the daily number of transactions on the checking account, and amount transferred into the checking account. With no other information about each consumer we are able to achieve a ROC AUC of 0.918 for the networks, and 0.926 for the networks in combination with a random forests classifier. © 2018 Elsevier Ltd. All rights reserved. 1. Introduction The ability to discriminate bad customers from good ones is im- portant for banks and other lending companies. A small improve- ment in prediction accuracy may result in a large gain in profitabil- ity. Early identification of high risk consumers may aid the pre- vention of loan defaults and help the consumers to better manage their personal economy. In credit scoring, one builds a model for the correspondence be- tween default and various loan obligor characteristics based on a relevant sample of people, and use this model to predict the prob- ability that a person will repay his debts. There is an extensive literature on credit scoring, both for as- sessing private loans ( Butaru et al., 2016; Chi & Hsu, 2012; Khan- dani, Kim, & Lo, 2010; Sousa, Gama, & Brandão, 2016 ) and corpo- rate loans ( Jones, Johnstone, & Wilson, 2015; Ravi Kumar & Ravi, 2007 ). Some recent work include Abellán and Castellano (2017) ; Chen, Zhou, Wang, and Li (2017) ; Xia, Liu, Li, and Liu (2017) , and Barboza, Kimura, and Altman (2017) . For an overview and com- parison of papers, see García, Marqués, and Sánchez (2014) and Lessmann, Baesens, Seow, and Thomas (2015) . All the papers above attempt to model delinquencies and de- faults by applying machine learning algorithms to a set of ex- planatory variables. While there are variations in the information ∗ Corresponding author. E-mail addresses: haavakva@math.uio.no (H. Kvamme), nikolai.sellereite@nr.no (N. Sellereite), kjersti.aas@nr.no (K. Aas), sasjurse@gmail.com (S. Sjursen). that is available to researchers (see e.g. Lessmann et al., 2015 ), the constructed explanatory variables tend to be quite similar. Papers typically use information from credit bureaus, such as number of outstanding accounts, delinquent accounts, and balance on other loans; individual account characteristics, such as current balance of the individuals accounts and monthly income; and demographic data, such as age and marital status. Butaru et al. (2016) also in- clude macroeconomic variables, such as interest rates and unem- ployment statistics, as an attempt to make the delinquency model generalize better over longer periods of time. As all these papers use similar explanatory variables, the re- searchers commonly explore differences between scoring models rather than the benefit of adding new explanatory variables. There are however some exceptions: Khandani et al. (2010) explore the benefit of adding information from detailed purchase volumes to their models. This includes travel expenses, gas station expenses, bar expenses, etc. Chi and Hsu (2012) also introduce consumer transaction data through an aggregated measure called average uti- lization ratio of credit. In this paper we further investigate how transaction data can be used for credit scoring. In a joint research with Norway’s largest financial service group, DNB, we use transaction data to predict mortgage defaults. In 2012, the average Norwegian made 323 card transactions, where 71% of the value transferred was through debit payments ( Norges-Bank, 2012 ). Hence, transactional data may pro- vide a useful description of user behavior, and subsequently con- sumer credit risk. The transaction information consists of the daily balances on consumers’ credit, checking, and savings accounts, in addition to https://doi.org/10.1016/j.eswa.2018.02.029 0957-4174/© 2018 Elsevier Ltd. All rights reserved. 208 H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 the daily number of transactions on the checking account, and the amount transferred into the checking account. Our dataset is thus a collection of time series for each customer. We do not use any of the common covariates mentioned above, as our goal is to investi- gate the value of the information available in the time series data. In a production setting, our model can be combined with existing risk models to improve the overall performance. The idea behind this paper is to use deep learning, or deep neural networks ( LeCun, Bengio, & Hinton, 2015 ), to predict mort- gage defaults. Deep learning have had a dramatic impact in fields like image classification, text mining, and speech recognition. Com- mon to these three fields is that the data is unstructured. Thus, the original data (pixels, letters, words, frequencies) needs to be trans- formed into meaningful covariates before they can be passed to a classical algorithm. Until recently, such tasks have been solved by the manual creation of informative covariates from the data. Deep neural nets, on the other hand, process the data in a sequential manner, learning multiple levels of abstraction. Hence, a deep net use data to “learn” how to create good explanatory variables for the problem at hand. In this paper we apply a type of deep neural networks denoted convolutional neural networks (CNNs). To the extent of our knowl- edge, we are the first to apply CNNs to consumers’ account bal- ances to predict mortgage defaults. Transaction data has previously been used for credit scoring, see e.g. Khandani et al. (2010) . How- ever, most existing work relies on heuristic hand-crafted feature design. It is often hard for us humans to figure out appropriate features or covariates for credit scoring. Hence, the purpose of this study is to use a convolutional neural network to automate fea- ture engineering from the raw time series, in a systematic way. The learned features resulting from such a model may be viewed as the higher level abstract representation of the low level raw time se- ries signals. There have been some successful attempts applying deep neural nets to time series data in other application ar- eas. In the field of human activity recognition, Ordóñez and Roggen (2016) ; Yang, Nguyen, San, Li, and Krishnaswamy (2015) , and Hammerla, Halloran, and Ploetz (2016) use sensor data to recognize activities improving the state-of-the-art. Cui, Chen, and Chen (2016) ; Prasad and Prasad (2014) ; Zheng, Liu, Chen, Ge, and Zhao (2014) , and Le Guennec, Malinowski, and Tave- nard (2016) present network structures for time series classifi- cation across multiple domains. Also, Sirignano, Sadhwani, and Giesecke (2016) assess mortgage risk using a deep net on 294 ex- planatory variables. However, on the contrary to our work, they use multilayer perceptrons instead of CNNs, and the majority of the variables are loan-level feature and performance variables. The deep net is mainly used to identify complex interactions between the input variables. Much work in credit scoring is related to regulatory frame- works such as the Basel Accord. The Basel II regulations require transparency in the loan-granting process. Due to their non-linear structure, CNNs are usually considered as black boxes. That is, it is usually not obvious what exactly makes them arrive at their predictions. However, since this lack of transparency in many ap- plications is considered a major drawback, the development of methods for explaining and interpreting deep learning models has recently attracted increased attention, see e.g. Lundberg and Lee (2016) ; Samek, Wiegand, and Müller (2017) ; Shrikumar, Green- side, Shcherbina, and Kundaje (2016) , and Ribeiro, Singh, and Guestrin (2016) . Hence, it is not obvious that CNNs will remain inappropriate for building regulatory models in the future. Mean- while, a credit scoring system based on a CNN may be useful in many other applications. Financial institutions may e.g. use them in their own internal risk estimation or as part of a validation ex- ercise (Pillar 2 of the Basel framework). A lender commonly makes two types of credit decisions: first, whether to grant credit to a new applicant, and second, how to deal with existing applicants, including whether to increase their credit limits. The first problem is denoted application scor- ing and the latter behavioral scoring. The majority of previous work focus on application scoring, while prior work on behav- ioral scoring is much less developed, see e.g. Thomas (20 0 0) and Kennedy, Mac Namee, Delany, O’Sullivan, and Watson (2013) . The methodology proposed in this paper may be used both for appli- cation and behavioral scoring. Consumer transaction histories con- tain implicit repayment behavior, making them suitable for behav- ior scoring. However, as the new Revised Payment Service Direc- tive (PSD2) becomes implemented across the EU and the European Economic Area during 2018, the banks (and other lending institu- tions) will have access to transaction data even for new customers. Hence, this kind of data may also be used for application scoring. The remainder of the paper is organized as follows. In Section 2 we describe our dataset and how it was processed be- fore fitting the neural networks. In Section 3 we introduce convo- lutional neural networks, and show how they were applied to our problem. In Section 4 we evaluate the performance of the proposed algorithms. Finally, in Section 5 , we summarize our findings. 2. Data This study is a collaboration with the largest Norwegian finan- cial service group, DNB, using data from their banking services. The dataset consists of a sample of 20,989 customers who either pre- viously had a mortgage, or got approved for a mortgage at some point between January 2012 and April 2016. For every customer we have the daily balances on their checking account, savings account, and credit card, in addition to the daily number of transactions on the checking account. We also know the daily amounts that are transferred into the checking account (e.g. from salary, payments from friends, etc.). Hence, we have a set of five time series for each customer. Finally, adding the sum of checking account, savings ac- count, and credit card as a new time series, we end up with a total of six series. A summary of the six series can be found in Table 1 . For a customer to be granted a mortgage by DNB, he or she is required to open a checking account at the bank, given that this criterion is not already satisfied. There are however no require- ments as far as savings accounts and credit cards are concerned. As a result, there are many missing series in the dataset. Our definition of default is the one used in Basel II, i.e. that the obligor is past due for more than 90 days on the obligation to the bank. Housing prices in Norway have generally increased steadily since 2003, and thus, the mortgage market has seen few defaults ( Finanstilsynet, 2016 ). The lack of defaults poses a problem due to the fact that our algorithms need large datasets to generalize well. In addition to this, the large class imbalance (ratio between non- defaults and defaults) is commonly regarded as a problem for clas- sification algorithms. Both issues are addressed in later sections. 2.1. Training and test set We divide our data into a training set and a test set. For train- ing, we extract transaction histories from the time period Decem- ber 31, 2012 to December 31, 2013 and use the outcome (default / non-default) in the period from January 1, 2014 to January 1, 2015 as response variables. A subset of the training set was held out from training and used as a validation set for tuning of our algo- rithms. For testing, we extract transaction histories from the time period February 28, 2014 to February 28, 2015, and response vari- ables from the period March 1, 2015 to March 1, 2016. There are no customers who appear in both the training and test set, mean- ing that our study is both “out-of-time” and “out-of-sample”. Cus- H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 209 Table 1 Time series. Series Abbrev. Explanation Checking account ch Balance on the checking account. Savings account sa Balance on the savings account. Credit card cc Balance on the bank issued credit card. Checking transactions trch Number of transaction on the checking account. Into checking in Sum of transactions into the checking account. Sum sum Sum of series ch, sa, and cc. The time series used in this paper. Each time series consists of values aggregated to a daily level. Table 2 Data proportions for training, validation, and test sets. Dataset Defaults Non-defaults Total Default type Training 1298 11,398 12,696 90 or 60 days past Training augmented 95,647 841,247 936,894 90 or 60 days past Validation 329 6043 6372 90 or 60 days past Test 96 1825 1921 90 days past tomers for whom we did not have a full year of transaction data, were removed both from the training and the test set. However, we evaluate the performance of our methods on customers with miss- ing data in Section 4.5 . For a further specification of the datasets, see Table 2 . 2.2. Preprocessing time series Convolutional neural networks require the input to be scaled. Hence, for images it is common to normalize the pixels to lie in the range [0, 1] ( Goodfellow, Bengio, & Courville, 2016 ). Pixels how- ever have the benefit of all being in [0, 255] and somewhat evenly distributed in this interval. On the contrary, there are great dif- ferences in the magnitude of the accounts. If we scale all series based on the most extreme account values, most of the series will have very small variations, making it hard for our neural net to learn from the data. Therefore, the time series were scaled such that each individual series is in the range [0, 1] through x = x − min (x ) max (x ) − min (x ) , (1) where x denotes one time series. With this scaling, the neural net receives inputs of similar magnitude. However, this comes at the cost of removing the magnitude of the individual series. Hence, the network can only find information in the relative patterns of the series, and has no knowledge about the actual size of the dif- ferent accounts. Missing series were treated as accounts without any movements. 3. Methods While there is an extensive literature on time series classifica- tion, we have chosen to focus only on the subset of algorithms that falls under deep learning. For a review of other state-of-the-art methods, see e.g. Bagnall, Lines, Bostrom, Large, and Keogh (2016) . In the following, we will first, in Section 3.1 , introduce convolu- tional neural networks and how they were applied to our credit scoring problem. Then, in Section 3.2 , we will discuss methods used to avoid overfitting and how we approached the class imbal- ance problem. 3.1. Convolutional neural networks Deep learning arguably owes much of its success to its extent of modularity. The framework is based on combining differentiable transformations, where each such transformation is commonly re- ferred to as a layer . The way one organizes the layers is called the network architecture . When fitting such a network to a dataset, or performing the “training” as it is denoted in machine learning, it is necessary to define a loss function that can be optimized. The loss function does not necessarily need to be the objective we actually want to optimize, but rather a function that indirectly optimizes our true goal ( Goodfellow et al., 2016 ). For binary classification it is common to use the binary cross entropy Loss = − ∑ i { y i log p i + (1 − y i ) log (1 − p i ) } , (2) where y i denotes the true class label as {0, 1}, and p i denotes our prediction for customer i in [0, 1]. Eq. (2) is the same as the neg- ative log likelihood used in logistic regression. Note that the loss decreases as the p i ’s move closer to the y i ’s, hence reducing the loss should improve our objective. During training the data is passed through the network, the loss is calculated, the gradients are computed with respect to all the parameters in the network, and the parameters are optimized through some version of gradient descent. The gradients can be ob- tained in a nice way using backpropagation ( LeCun, Bottou, Orr, & Müller, 1998 ), which basically is application of the chain rule for differentiation. Backpropagation calculates the gradients in a se- quential manner, such that the gradients of each transformation can be derived solely from the error passed back from the suc- ceeding layer. The main takeaway here is that layers can be im- plemented independent of each other, simplifying the process of combining them in an architecture. Convolutional neural networks (CNNs) constitute the subset of neural networks that contain convolutional layers. However, they typically contain other layers as well. The following sections de- scribe the different types of layers that have been used in our neu- ral networks. 3.1.1. Convolutional layers The first two layers of our network are presented in Fig. 1 . The left block shows an input time series that is 365 days long. For now, assume that the dataset only consist of a single time series per customer (instead of six). A convolutional layer consists of multiple filters that are applied to the input series. A filter is simply a vector (or matrix) consisting of weights that need to be optimized. The first convolutional layer has filters of length 9, as shown in the figure. The application of such a filter to the input series is just a sum over the element- wise product of the filter weights and the time series, resulting in 210 H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 Fig. 1. The first convolutional layer and first pooling layer of our model. The left block is the input series. The gray area marks the part of the series for which a convolutional filter is currently being applied. The middle block shows the convolved features (or activations) resulting from the 32 different convolutional filters. The right block shows the result of applying max pooling to the middle block. a convolved feature (or activation). A convolved feature is there- fore a measure of the correlation between the filter and the rele- vant part of the input. The middle block of Fig. 1 displays all the convolved features from the input, using 32 different filters. Each filter is slided across the input to generate as many convolved fea- tures as the length of the time series 1 . Due to the fact that the 32 filters are randomly initialized, they end up extracting different information from the input. After applying the filters to a time series, the convolved features are passed through a ReLU transform, which is simply max {0, x }. Such activation functions are applied to make the transformation non-linear, and they are commonly considered a part of the convo- lutional layer rather than a separate layer. See Gu et al. (2017) for a discussion of other activation functions. 3.1.2. Max pooling layers A convolutional layer preserves the resolution in the temporal dimension. There can however be some benefits of reducing the temporal resolution, such as reducing the number of parameters in the next layer, and introducing some shift-invariance ( Gu et al., 2017 ). As a result, many CNNs include so-called max pooling lay- ers, which simply replace a set of values with their maximum. Note that this layer does not introduce any additional parameters. We use max pooling after each convolutional layer. An example is shown to the right in Fig. 1 , where four values of each column of the convolved features are pooled into one. As shown in the fig- ure, pooling is applied separately for each feature series. Thus, the pooling layer preserves the number of feature series, while reduc- ing the temporal resolution. See Gu et al. (2017) for a discussion of other pooling layers. 3.1.3. Fully connected layers To perform the actual classification, it is necessary to combine the convolved features into a binary classifier. This is typically done through one or more fully connected (or dense ) layers, which form 1 Zero-padding is used to preserve the temporal dimension ( Goodfellow et al., 2016 ). a multilayer perceptron (a classical neural net). A dense layer re- quires a one-dimensional input, so the columns of the last layer are concatenated into a single vector x . The transformation is then simply the inner product with a weight matrix, z = x T W, with some non-linear activation function (typically ReLU). The number of columns in W determines the dimension of the output. Our fi- nal layer has only two outputs, one for each class, and therefore a softmax activation function is applied to ensure that the predic- tions are in the interval [0, 1]. Softmax (z) c = e z c ∑ j e z j , (3) where c and j refers to classes. Note that the softmax can be ap- plied for an arbitrary number of classes, but for binary classifica- tion it it equivalent to the logit function. As a result, in combina- tion with our choice of loss function, this final layer is equivalent to a logistic regression. 3.1.4. Network architecture and training In our network architecture we alternate between a convolu- tional layer and a max pooling layer two times, followed by two fully connected layers. For an input of length 365 days, this results in 199,234 weights that need to be optimized. The whole structure is displayed in Fig. 2 , were we also show the length of the filters and number of filters for the convolutions, the size of the pooling kernels, and the number of output nodes in the dense layers. This architecture was developed manually by evaluating on a validation set 2 , and the hyper parameter search was therefore not as rigor- ous as with e.g. a grid search. However, it was found that a variety of model configurations resulted in quite similar performance. In particular, further increasing the number of nodes, filters, or layers did not really seem to affect the performance. We train one network for each of our six time series, com- pletely independently, using the Adam optimization algorithm ( Kingma & Ba, 2014 ) with default parameters and batch size of 512. The resulting six predictions are averaged to give a final prediction 2 Note that the validation set is not the same as the test set. H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 211 Fig. 2. Our convolutional neural net. The blue layers are convolutional layers and the yellow layers are fully connected layers. All activations are ReLU, with the ex- ception of softmax in the final layer. Dropout, with rate 0.5, is performed between the last two layers. (For interpretation of the references to color in this figure leg- end, the reader is referred to the web version of this article.) for each customer. When computing the averages, predictions cor- responding to missing series are discarded. We also experimented with averaging the log-odds instead of the predictions, which re- sulted in a slight performance decrease. We also tried an alternative model where we train a network on all six series simultaneously. Thus, the input is 365 × 6 rather than 365 × 1, meaning that the first convolutional filter will have size 9 × 6, rather than 9 × 1. Apart from this, the architecture is identical to the architecture for the individual series. By combin- ing all series in the input, the network has the potential to learn interactions between a customer’s time series. However, as this model is more complex, it will require more data than the pre- vious model. In addition to the two models presented here, we have also tried deeper and more shallow models, as we will get back to in Section 4.2 . In early stages of the project, we even experimented with recurrent neural networks in the form of Long Short-Term Memory networks (LSTM). They were found to be outperformed by the CNNs, in addition to being much more computationally ex- pensive. Hence, they were dropped from further investigation. 3.2. Overfitting and class imbalance In the following sections we explain the techniques we have used to prevent the models from overfitting and how we handled the large class imbalance between defaulting and non-defaulting customers. 3.2.1. Regularization Deep neural networks are commonly overparameterized, mak- ing them prone to overfitting ( Gu et al., 2017 ). Subsequently, regu- larizing techniques have been introduced to cope with these prob- lems. Among the more common is the use of a validation set for monitoring the training process. We used this set to stop the gradi- ent descent iterations when the net starts to overfit. This technique is commonly referred to as early stopping . Another common form of regularization is dropout ( Srivastava, Hinton, Krizhevsky, Sutskever, & Salakhutdinov, 2014 ), which simply sets activations to zero with a given probability dur- ing training. This prevents units in the network from co-adapting too much. At test time, the dropout layer scales the activations according to the dropout rate. We use dropout with rate 0.5 between the two last layers, but it can in practice be applied between any two layers. See Goodfellow et al. (2016) and Gu et al. (2017) for a further review of regularization. 3.2.2. Data augmentation The problem of predicting mortgage defaults is highly imbal- anced in that the number of defaults is very small compared to the number of non-defaults. As a classifier’s performance commonly deteriorates under imbalance, we under-sample the non-defaulting customers to make the two classes more similar in size. More- over, to increase the number of defaults in the training set, we use delinquencies of 60 days in addition to those of 90 days. Hence, we ended up with a training set consisting of 12,696 customers, 10% of whom where characterized with default. Deep neural nets commonly require more data than classi- cal machine learning algorithms to perform well. As a result, data augmentation schemes have been explored in the literature ( Goodfellow et al., 2016; Krizhevsky, Sutskever, & Hinton, 2012 ). Most of this work has been on images, and does not necessar- ily generalize well to other data sources. However, recently some attempts of performing data augmentation on time series have been proposed. Le Guennec et al. (2016) explore the possibilities of training a CNN in a semi-supervised manner using time series from other datasets. Cui et al. (2016) propose a sliding window technique where they split the time series in many overlapping slices, and train a classifier using each slice as an individual series. This is in some sense similar to other time series prediction tasks where the dataset originally consists of many overlapping series of data (e.g. Ordóñez & Roggen (2016) ). In this paper, we use a slight variation of the sliding windows approach of Cui et al. (2016) . Our approach can be described as fol- lows. For each defaulting customer we first define a training period which ends one month before the default, and starts two years ear- lier. If the customer did not exist in the dataset at the start of this period, the start was set to the earliest date the customer existed. We then extract the time series corresponding to the first 365 days of the training period and use that as an observation. By jumping a fixed number of days, called stride in the machine learning lit- erature, we can get a new observation that is partially overlapping with the previous observation. This process is repeated until the end of our defined training period is reached. In this way, we get a training set consisting of multiple overlapping series from each de- faulting customer. This augmentation scheme is illustrated in Fig. 3 . The gray area shows the defined training periods for three different customers, and the red dots mark the time of their defaults. For id: 2 we show three of the observations (windows) that are added to the training set. Note that the first of the corresponding periods ends one year before the default occurs, while the last ends one month before. Thus both these observations, and all in between, fit our objective of predicting a default in the following 365 days. We also need to perform a similar augmentation for the non- defaulting individuals. Our first experiments showed that sampling non-defaulting and defaulting customers from different time peri- ods made our net learn different characteristics of these periods instead of the actual objective. Hence, to avoid this sampling bias, we had to sample the start and end dates for the non-defaulting customers from the ones of the defaulting customers. Ideally one would use strides of 1 as this would give most diversity in the training data. However, our preliminary analysis 212 H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 Fig. 3. Illustration of our data augmentation scheme. The gray area identifies the time period from which we extract data, and the red dots show the time of default. The sliding window approach is illustrated for id: 2. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) showed that strides of 1 did not give better results than strides of 5. Hence, since the memory usage of the latter is only a fifth of the first, we decided to use strides of 5 in our final analysis. In Table 2 we have shown the size of our training set, the train- ing set augmented with the sliding windows, the validation set, and the test set. The financial group DNB does not want to dis- close its true default ratio. Hence, to generate the test set, 20 0 0 customers were randomly sampled in such a way that the fraction of defaulting customers was 5%. After removing customers with missing history, we ended up with a test set of 1921 customers. The size of the test set is quite small, which results in larger vari- ance in our results. Note however that we have experimented with different test sets from different time periods, and the results seem to be quite consistent. 4. Experiments In the following, we first present the measures used to evaluate the performance of our models. We then describe our experiments and their results. 4.1. Evaluation measures The Receiver Operating Characteristic (ROC) curve is a com- mon measure for evaluating a classifier’s ability to separate two classes. The ROC curve is a plot of the true positive rate (for de- fault) against the false positive rate, for all thresholds. A threshold is the chosen cut-off in the estimated scores from the neural net. Since the ROC curve is based on the true positive rates and the false positive rates, it is invariant to class proportions. Moreover, it is not dependent on the quality of the predictions (probability estimates), only on the classifier’s ranking capabilities. ROC curves are often summarized through the area under the curve, or AUC. With perfect classification, the AUC will be 1, while random predictions result in an AUC of 0.5 on average. AUC is a general measure of binary classification, and is widely used across disciplines. There are however more specialized mea- sures for evaluating a credit scoring model. One such measure is the Expected Maximum Profit (EMP) ( Verbraken, Bravo, Weber, & Baesens, 2014 ), which is a profit based performance measure ac- counting for the benefits generated by healthy loans and the costs caused by loan defaults. We have chosen not to use EMP in this paper. The main reason is that this measure is heavily dependent Table 3 AUC for different CNN architectures. Model Mean AUC Std AUC AUC for average pred. Full, 1 conv layer 0.8884 0.0055 0.8979 Full, 2 conv layers 0.8879 0.0041 0.8982 Full, 3 conv layers 0.8897 0.0048 0.9015 Indiv, 1 conv layer 0.9022 0.0022 0.9043 Indiv, 2 conv layers 0.9146 0.0025 0.9184 Indiv, 3 conv layers 0.9106 0.0033 0.9148 AUC mean and standard deviation for 10 experiments for each of six differ- ent CNNs. “Full” refers to the approach with all time series as input, and “Indiv” to the approach where we compute the average of the predictions from six individually trained models. The rightmost column gives the AUC of the predictions averaged over the 10 trained models. on the class proportions. As stated earlier, we have used a fictive default rate, since the bank did not want to disclose its true ratio. Hence, it does not make any sense to use a measure that signifi- cantly changes when the class proportions are altered. Our convolutional neural network is initialized with ran- dom weights. Moreover, we use dropout by multiplying random Bernoulli variables with the input to the last layer, and the net is trained with stochastic gradient descent. As a result, the network will be slightly different every time it is trained. Hence, to control how our results are influenced by randomness, we repeat all ex- periments 10 times and report the average and standard deviation of the 10 resulting AUC values. 4.2. Competing architectures In the following section, the two CNNs introduced in Section 3.1.4 are compared and evaluated. The architectures of the two models are identical, except that the first takes individual time series as input ( Indiv model ), while all six time series are input to the latter ( Full model ). Note that all training sets are augmented as described in Section 3.2.2 . Summary statistics for the AUCs are shown in Table 3 under the names Indiv, 2 conv layes and Full, 2 conv layers . We see that the individual model is clearly better than the full model. This might indicate that the full model is not able to take advantage of interactions between the series, or at least that the added benefit is smaller than the downside of the added complexity of the model. H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 213 Table 4 AUC for individual series. Series Mean AUC Std NM Mean AUC NM Std Prop Missing ch 0.8632 0.0040 0.8646 0.0036 0.0068 in 0.7116 0.0108 0.7140 0.0107 0.0068 cc 0.7636 0.0080 0.8099 0.0101 0.2405 sa 0.7902 0.0056 0.8178 0.0039 0.1463 sum 0.8752 0.0078 0.8752 0.0078 0 trch 0.8472 0.0064 0.8480 0.0063 0.0068 All 0.9146 0.0025 0.9146 0.0025 0 In the NM columns (Not Missing) we have removed customers having missing series or series without any activity. “Prop missing” gives the proportion of series removed from the test set in NM. The AUC is averaged over 10 experiments. The “All” series gives the score of the averaged predictions (Indiv model). Table 5 AUC for different subsets of the training data. Proportion Mean AUC Std AUC 1.00 0.9146 0.0026 0.90 0.9109 0.0048 0.75 0.9086 0.0054 0.50 0.9028 0.0033 0.25 0.8954 0.0082 0.10 0.8827 0.0062 Non-augmented 0.8880 0.0042 For each subset, we sample the customers from the training set, and use sliding window for data augmentation. The left- most column shows the proportion of customers sampled from the training set, while the remaining columns show the average and standard deviation of AUC for the 10 repetitions. 1.00 gives the results for the full training set (Indiv model), while “Non- augmented” refers to using all customers, but with no data aug- mentation. Table 6 AUC for different lengths of time series. Model Mean AUC Std AUC AUC for average pred. 365 days 0.9146 0.0026 0.9184 180 days 0.8945 0.0051 0.8981 91 days 0.8744 0.0051 0.8791 31 days 0.8425 0.0043 0.8460 The “Mean AUC” is averaged over 10 repetitions. The “AUC for average pred.” gives the AUCs for the averaged predictions of the 10 repetitions of the same model. All models have the same ar- chitecture as the 365 days CNN. Table 3 also includes four other models. The only difference be- tween the models is the number of convolutional layers (and pool- ing layers), and whether the input consist of individual series or all six series simultaneously. For the full specification of the models, see Fig. A.1 in the Appendix. From the table we see that all the individual models have higher average AUC than the full models. Additionally, it seems that there is no reason to have more than two convolutional layers in the networks. Averaging the predictions from six individually trained models has a regularizing effect. To check whether the differences in per- formance between the full models and the individual models are caused by this averaging, we also, for each model, compute the AUC for the average of the 10 predictions for each customer. The results are shown in the rightmost column of Table 3 (“AUC for average pred.”) We see that this improves the AUC for all models. The individual models are still better than the full models, but the difference seems to be smaller. In the remainder of this section we concentrate on the Indiv model, 2 conv layers , since this is the one that shows the best per- formance in Table 3 . Table 7 AUC for different length of series. Model Mean AUC Std AUC AUC for average pred. RF 365 days 0.9130 – – RF 31 days 0.8907 – – Combined 365 days 0.9254 0.0011 0.9260 Combined 31 days 0.8941 0.0 0 06 0.8946 RF is a random forests classifier with covariates extracted from the unscaled time series. Combined gives the averaged prediction of the CNN and the random forests classifier. The results for the combined classifier are aver- aged over 10 repetitions of the CNN. The “AUC for average pred.” gives the AUC for the averaged predictions of the 10 repetitions of the CNN. 4.3. Importance of different series Our model creates predictions based on six different time se- ries: checking account balance (ch), amount transferred into check- ing account (in), credit card balance (cc), savings account balance (sa), the sum of checking, saving and credit card balances (sum), and number of transactions on the checking account (trch). To in- vestigate how much information there is in each series, we report their individual AUC values in Table 4 . As before, we report the re- sults averaged over 10 experiments. The checking account (ch and trch) and the sum seem to be most informative, while the amount transferred into the checking account (in) provides less informa- tion. In Section 2 it was stated that missing savings accounts and credit card accounts are regarded as accounts without any move- ments. This means that the comparison between the different se- ries in the two leftmost columns of Table 4 might not be com- pletely fair. Hence, we did a new experiment in which we re- moved all customers missing the relevant time series from the test set. The results are shown in columns 3 and 4 of the table (NM Mean and NM Std), while the proportions of removed customers is shown in column 5. From the table we see that the AUC for credit card accounts (cc) and savings accounts (sa) significantly increase, while there are small differences for the other series. This is rea- sonable, as the proportion of missing data is largest for the credit cards and savings accounts. 4.4. Training data In this section we investigate how the size of the training set affects the performance of our classifier. This can help us un- derstand to what extent more training data would improve our discriminator. We also evaluate the effect of our augmentation scheme. Table 5 shows AUC for models trained on different subsets of the training data. The subsets were created by randomly sampling customers. For a given proportion of the full training set we train 10 models in the same way as previously for the full training set. Investigating the table, the performance seems to be improved with the size of the training data. This suggests that with more data we might be able to obtain even better results than those pre- sented in this paper. The last row of the leftmost column in Table 5 shows the re- sults of fitting the same model to the original (not augmented) training set. We see that using 10% to 25% of the original customers with augmented data results in the same performance as using all customers with non-augmented data. 4.5. Length of time series To investigate to what extent lack of information affects the performance of our model, we have also fitted our CNN to series of different lengths, More specifically, we used the last month (31 214 H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 Table 8 Evaluation metrics. Model Threshold Accuracy Sensitivity Specificity AUC Brier Score H-measure CNN 0.3 0.954 0.374 0.985 0.915 0.0381 0.564 Combined 0.3 0.947 0.527 0.970 0.925 0.0364 0.592 LR 0.3 0.910 0.490 0.935 0.864 0.0458 0.455 MLP (512) 0.3 0.899 0.590 0.918 0.875 0.0550 0.484 RF 0.3 0.920 0.602 0.940 0.913 0.0386 0.557 CNN 0.5 0.953 0.064 0.999 0.915 0.0381 0.564 Combined 0.5 0.956 0.177 0.997 0.925 0.0364 0.592 LR 0.5 0.943 0.219 0.981 0.864 0.0458 0.455 MLP (512) 0.5 0.927 0.455 0.953 0.875 0.0550 0.484 RF 0.5 0.957 0.332 0.991 0.913 0.0386 0.557 Different evaluation metrics for two different threshold values. All results are based on the time series with length 365 days. For each model we have averaged the metrics over 10 trained models. CNN is the best- performing CNN from Table 3 , LR is a logistic regression model, MLP is a Multilayer perceptron with one hidden layer and 512 nodes, RF is the random forest model, and Combined is the combined model in which we averaged the predictions from the CNN and the RF. Fig. 4. ROC curves for the CNN, the RF, and their averaged predictions model. The time series are the full 365 day long series. The CNN comes from the averaged predictions over the 10 experiments. Fig. 5. ROC curves for the CNN, the RF, and their averaged predictions model. The time series used are contain only the last 31 days of data. The CNN comes from the averaged predictions over the 10 experiments. H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 215 days), the last quarter (91 days), and the two last quarters (182 days) of the original dataset to train the same model as previously. The results are shown in Table 6 . As can be seen, there is a clear drop in performance when the time series get shorter. On the downside, this makes it harder to evaluate customers with a short relationship to the bank. However, the results also suggest that it might be possible to increase the performance of our models by applying them to time series longer than 365 days. 4.6. Classical covariates As described in Section 2.2 , all time series are scaled to be in [0, 1] before fitting the CNN. This kind of scaling means that in- formation about the magnitude of the series is lost. To assess the potential decrease in results connected to this removal of relevant information, we fit a random forests classifer (RF), consisting of 800 trees, to explanatory variables created from the original, un- scaled time series. RF is often credited as a very strong classifier outperforming several alternative methods (including Support Vec- tor Machines and Neural networks) when applied to credit scor- ing ( Brown & Mues, 2012; Kruppa, Schwarz, Arminger, & Ziegler, 2013; Lessmann et al., 2015 ). For each of the six time series, we computed the mean, max, minimum, standard deviation, and the standard deviation scaled by the mean. These covariates were com- puted both for the full series and for the last month (31 days). In addition, we divided each of the covariates for the full series by the equivalent covariate calculated for the last month, producing a total of 15 features for each series. The resulting AUC is shown in the upper row of Table 7 . We also fitted a logistic regression model and a Multilayer perceptron (with one hidden layer and 512 nodes) using the same explanatory variables. The AUC values for these models were significantly worse (0.86 and 0.88 see Table 8 ) than that for the RF, confirming the observations from previous studies. Hence, we decided not to include these models in the fur- ther comparisons. Comparing the numbers in the upper rows of Tables 6 and 7 we see that for the 365 days period, the AUC for the CNN is slightly better than that for the RF. We also used a combined model in which we averaged the predictions from the CNN and the RF. From the third row of Table 6 we see that there is a small in- crease in performance compared to that obtained using one of the models only. The ROC curves for the CNN, the RF and the com- bined model are shown in Fig. 4 . Running the diagnostic test for comparing ROC curves described in DeLong, DeLong, and Clarke- Pearson (1988) verifies that there are no significant differences be- tween the AUCs for any pair of ROC-curves. Table 8 gives the values of 6 evaluation metrics for each of the 5 models. In addition to Accuray, Sensitivity, Specificity and AUC, the Brier Score ( Brier, 1950 ) and the H-measure ( Hand, 2009; 2010 ) are reported. The AUC is sometimes criticised for treating the relative severities of misclassifications differently when different classifiers are used. The H-measure, on the other hand, may accommodate expert knowledge regarding misclassification costs, whenever that is available. We want to treat misclassifications of the smaller class as more serious than those of the larger class, since otherwise very little loss would be made by assigning everything to the larger class. Hence, we set the cost of misclassifying a bad customer as healthy to 0.95 and the cost of misclassifying a good customer as defaulter to 0.05. We have also compared the three methods using data only from the last 31 days. The AUC-values are given in the last row of Table 6 and rows two and four in Table 7 , respectively, while the corresponding ROC-curves are shown in Fig. 5 . As can be seen from the tables and the figure, the performance of the RF does not degrade as much as that of the CNN when decreasing the length of the time series to 31 days. The p -value of the test for comparing the two ROC-curves is 0.005, meaning that the AUC for the RF is significantly higher than that for the CNN in this case. 5. Discussion The ability to discriminate bad customers from good ones is very important for banks and other lending companies. There is a large literature on methods used to predict defaults of consumer loans. Status quo in both the industry and the academic litera- ture is to use models with handcrafted explanatory variables. In contrast, we use a highly nonlinear convolutional neural network (CNN), where the input is raw account transactional data. More specifically, we have used daily balances of customers’ checking ac- counts, savings accounts, and credit card accounts, in addition to daily number of transaction on the checking accounts, and amount transferred into the checking accounts to predict mortgage delin- quency. A CNN has a number of hyperparameters which need to be cho- sen, such as the number of layers, type of layers, and the type of regularization. Our best performing CNN with two convolutional and two fully connected layers achieved an AUC of 0.915. Consider- ing the fact that our training set only consists of 12,696 customers, and that we do not use any other information than the account data, the results are quite promising. In our analysis, we found that the AUC for our CNN was in- creasing with the size of the training set, suggesting that a larger dataset might result in even higher performance. In addition, the performance increased with the length of the time series, indicat- ing that using a time series that is longer than 365 days may give better results. Comparing the predictions from a single CNN with ensemble predictions from 10 randomly initialized models showed that the increase in AUC using several model was limited. By using transaction data only, one throws away a lot of other informative data that is typically used for credit scoring, e.g. the loan balance, socioeconomic data, payment history, and credit bu- reau data. Further, all time series are scaled to be in the interval [0,1] before fitting the CNN, meaning that information about the magnitude of the time series is lost. To account for the scaling, a random forests classifier was fitted to covariates extracted from the unscaled series, resulting in an AUC of 0.913. By combining the predictions of the CNN and the random forests, we achieved an AUC of 0.925. Future research directions could include the combination of our model with other credit scoring models that use more clas- sical credit information; applying our model on a larger dataset, as CNNs tend to perform better with more data; and using more granular data in the form of labeled transactions (e.g. groceries, gas, rent, vacation). Acknowledgments This work was supported by The Norwegian Research Council 237718 through the Big Insight Center for research-driven innova- tion. We thank Ørnulf Borgan for comments and suggestions on the paper. We also thank Sven Haadem for his help on creating a dataset from DNB’s mortgage portfolio. Appendix A Fig. A.1 . 216 H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 Fig. A.1. Our three convolutional neural net architectures for single time series. We also have a version of each model with 365 × 6 input. The blue layers are convolutional layers and the yellow layers are fully connected layers. All activations are ReLU, with the exception of softmax in the final layer. Dropout, with rate 0.5, is performed between the last two layers. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) References Abellán, J., & Castellano, J. G. (2017). A comparative study on base classifiers in en- semble methods for credit scoring. Expert Systems with Applications, 73 , 1–10. doi: 10.1016/j.eswa.2016.12.020 . Bagnall, A., Lines, J., Bostrom, A., Large, J., & Keogh, E. (2016). The great time series classification bake off: A review and experimental evaluation of re- cent algorithmic advances. Data Mining and Knowledge Discovery . doi: 10.1007/ s10618- 016- 0483- 9 . Barboza, F., Kimura, H., & Altman, E. (2017). Machine learning models and bankruptcy prediction. Expert Systems with Applications, 83 , 405–417. doi: 10. 1016/j.eswa.2017.04.006 . Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78 (1), 1–3 . Brown, I. , & Mues, C. (2012). An experimental comparison of classification algo- rithms for imbalanced credit scoring data sets. Expert Systems with Applications, 39 (3), 3446–3453 . Butaru, F., Chen, Q., Clark, B., Das, S., Lo, A. W., & Siddique, A. (2016). Risk and risk management in the credit card industry. Journal of Banking & Finance, 72 , 218– 239. doi: 10.1016/j.jbankfin.2016.07.015 . Chen, X. , Zhou, C. , Wang, X. , & Li, Y. (2017). The credit scoring model based on lo- gistic-bp-adaboost algorithm and its application in p2p credit platform. In X. Li, & X. Xu (Eds.), Proceedings of the fourth international forum on decision sciences (pp. 119–130). Singapore: Springer Singapore . Chi, B.-W., & Hsu, C.-C. (2012). A hybrid approach to integrate genetic algorithm into dual scoring model in enhancing the performance of credit scoring model. Expert Systems with Applications, 39 (3), 2650–2661. doi: 10.1016/j.eswa.2011.08. 120 . Cui, Z., Chen, W., & Chen, Y. (2016). Multi-Scale Convolutional Neural Networks for Time Series Classification. CoRR http://arxiv.org/abs/1603.06995 . DeLong, E. R. , DeLong, D. M. , & Clarke-Pearson, D. L. (1988). Comparing the areas under two or more correlated receiver operating characteristic curves: A non- parametric approach. Biometrics, 44 (3), 837–845 . Finanstilsynet (2016). Finansielt utsyn 2016. http://www.finanstilsynet.no/Global/ Venstremeny/Rapport/2016/Finansielt _ utsyn _ 2016.pdf . Accessed: 2017-02-06. García, V., Marqués, A. I., & Sánchez, J. S. (2014). An insight into the experimental design for credit risk and corporate bankruptcy prediction systems. Journal of Intelligent Information Systems, 44 (1), 159–189. doi: 10.1007/s10844- 014- 0333-4 . Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning . MIT Press . http:// www.deeplearningbook.org . Gu, J., Wang, Z., Kuen, J., Ma, L., Shahroudy, A., Shuai, B., Liu, T., Wang, X., & Wang, G. (2017). Recent Advances in Convolutional Neural Networks. CoRR http: //arxiv.org/abs/1512.07108 . Hammerla, N. Y., Halloran, S., & Ploetz, T. (2016). Deep, Convolutional, and Recur- rent Models for Human Activity Recognition Using Wearables. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence (pp. 1533– 1540). New York, New York, USA: AAAI Press. http://dl.acm.org/citation.cfm?id= 3060832.3060835 . Hand, D. J. (2009). Measuring classifier performance: A coherent alternative to the area under the ROC curve. Machine Learning, 77 (1), 103–123. doi: 10.1007/ s10994- 009- 5119- 5 . Hand, D. J. (2010). Evaluating diagnostic tests: The area under the ROC curve and the balance of errors. Statistics in Medicine, 29 (14), 1502–1510. doi: 10.1002/sim. 3859 . Jones, S., Johnstone, D., & Wilson, R. (2015). An empirical evaluation of the perfor- mance of binary classifiers in the prediction of credit ratings changes. Journal of Banking & Finance, 56 , 72–85. doi: 10.1016/j.jbankfin.2015.02.006 . Kennedy, K. , Mac Namee, B. , Delany, S. J. , O’Sullivan, M. , & Watson, N. (2013). A window of opportunity: Assessing behavioural scoring. Expert Systems with Ap- plications, 40 (4), 1372–1380 . Khandani, A. E., Kim, A. J., & Lo, A. W. (2010). Consumer credit-risk models via machine-learning algorithms. Journal of Banking & Finance, 34 (11), 2767–2787. doi: 10.1016/j.jbankfin.2010.06.001 . Kingma, D. P., & Ba, J. (2014). Adam: A Method for Stochastic Optimization. CoRR http://arxiv.org/abs/1412.6980 . Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bot- tou, & K. Q. Weinberger (Eds.), Advances in neural information processing sys- tems 25 (pp. 1097–1105). Curran Associates, Inc. http://papers.nips.cc/paper/ 4824- imagenet- classification- with- deep- convolutional- neural- networks.pdf . Kruppa, J. , Schwarz, A. , Arminger, G. , & Ziegler, A. (2013). Consumer credit risk: In- dividual probability estimates using machine learning. Expert Systems with Ap- plications, 40 (13), 5125–5131 . Le Guennec, A. , Malinowski, S. , & Tavenard, R. (2016). Data augmentation for time series classification using convolutional neural networks. In Proceedings of the ECML/PKDD Workshop on Advanced Analytics and Learning on Temporal Data ,Riva Del Garda, Italy . LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521 (7553), 436– 4 4 4. doi: 10.1038/nature14539 . LeCun, Y., Bottou, L., Orr, G. B., & Müller, K. R. (1998). Efficient backprop. Neural Networks: Tricks of the Trade , 9–50. doi: 10.1007/3- 540- 49430- 8 _ 2 . Lessmann, S., Baesens, B., Seow, H.-V., & Thomas, L. C. (2015). Benchmarking state- of-the-art classification algorithms for credit scoring: An update of research. Eu- ropean Journal of Operational Research, 247 (1), 124–136. doi: 10.1016/j.ejor.2015. 05.030 . H. Kvamme et al. / Expert Systems With Applications 102 (2018) 207–217 217 Lundberg, S., & Lee, S. (2016). An unexpected unity among methods for interpreting model predictions CoRR abs/1611.07478 . Norges-Bank (2012). Årsrapport om betalingssystem 2012. http://static.norges-bank. no/pages/94894/Betalingssystem _ 2012 _ o.pdf?v=27052013101201&ft=.pdf . Ac- cessed: 2017-02-09. Ordóñez, F., & Roggen, D. (2016). Deep convolutional and LStm recurrent neu- ral networks for multimodal wearable activity recognition. Sensors, 16 (1), 115. doi: 10.3390/s16010115 . Prasad, S. C., & Prasad, P. (2014). Deep Recurrent Neural Networks for Time Series Prediction. CoRR http://arxiv.org/abs/1407.5949 . Ravi Kumar, P., & Ravi, V. (2007). Bankruptcy prediction in banks and firms via sta- tistical and intelligent techniques – A review. European Journal of Operational Research, 180 (1), 1–28. doi: 10.1016/j.ejor.2006.08.043 . Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). Why should i trust you?: Explaining the predictions of any classifier CoRR abs/1602.04938 . Samek, W., Wiegand, T., & Müller, K. (2017). Explainable artificial intelligence: Un- derstanding, visualizing and interpreting deep learning models CoRR abs/1708. 08296 . Shrikumar, A., Greenside, P., Shcherbina, A., & Kundaje, A. (2016). Not just a black box: Learning important features through propagating activation differences CoRR abs/1605.01713 . Sirignano, J., Sadhwani, A., & Giesecke, K. (2016). Deep Learning for Mortgage Risk. ArXiv e-prints http://adsabs.harvard.edu/abs/2016arXiv160702470S . Sousa, M. R., Gama, J., & Brandão, E. (2016). A new dynamic modeling framework for credit risk assessment. Expert Systems with Applications, 45 , 341–351. doi: 10. 1016/j.eswa.2015.09.055 . Srivastava, N. , Hinton, G. , Krizhevsky, A. , Sutskever, I. , & Salakhutdinov, R. (2014). Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15 , 1929–1958 . Thomas, L. C. (20 0 0). A survey of credit and behavioural scoring: forecasting fi- nancial risk of lending to consumers. International journal of forecasting, 16 (2), 149–172 . Verbraken, T., Bravo, C., Weber, R., & Baesens, B. (2014). Development and applica- tion of consumer credit scoring models using profit-based classification mea- sures. European Journal of Operational Research, 238 (2), 505–513. doi: 10.1016/j. ejor.2014.04.001 . Xia, Y., Liu, C., Li, Y., & Liu, N. (2017). A boosted decision tree approach using Bayesian hyper-parameter optimization for credit scoring. Expert Systems with Applications, 78 , 225–241. doi: 10.1016/j.eswa.2017.02.017 . Yang, J. B. , Nguyen, M. N. , San, P. P. , Li, X. L. , & Krishnaswamy, S. (2015). Deep convo- lutional neural networks on multichannel time series for human activity recog- nition. In Proceedings of the twenty-fourth international conference on artificial intelligence . In IJCAI’15 (pp. 3995–4001). AAAI Press . Zheng, Y., Liu, Q., Chen, E., Ge, Y., & Zhao, J. L. (2014). Time series classification using multi-channels deep convolutional neural networks. Lecture Notes in Computer Science , 298–310. doi: 10.1007/978- 3- 319- 08010- 9 _ 33 . II 71 Journal of Machine Learning Research 20 (2019) 1-30 Submitted 6/18; Revised 7/19; Published 8/19 Time-to-Event Prediction with Neural Networks and Cox Regression H̊avard Kvamme haavakva@math.uio.no Ørnulf Borgan borgan@math.uio.no Ida Scheel idasch@math.uio.no Department of Mathematics University of Oslo P.O. Box 1053 Blindern 0316 Oslo, Norway Editor: Jon McAuliffe Abstract New methods for time-to-event prediction are proposed by extending the Cox proportional hazards model with neural networks. Building on methodology from nested case-control studies, we propose a loss function that scales well to large data sets and enables fitting of both proportional and non-proportional extensions of the Cox model. Through simulation studies, the proposed loss function is verified to be a good approximation for the Cox partial log-likelihood. The proposed methodology is compared to existing methodologies on real-world data sets and is found to be highly competitive, typically yielding the best performance in terms of Brier score and binomial log-likelihood. A python package for the proposed methods is available at https://github.com/havakv/pycox. Keywords: Cox regression, customer churn, neural networks, non-proportional hazards, survival prediction 1. Introduction In this paper, we consider methodology for time-to-event prediction, a part of survival analysis that reasons about when a future event will occur. Applications of time-to-event predictions can be found in a variety of settings such as survival prediction of cancer patients (e.g., Vigan et al., 2000), customer churn (e.g., Van den Poel and Lariviere, 2004), credit scoring (e.g., Dirick et al., 2017), and failure times of mechanical systems (e.g., Susto et al., 2015). Arguably, the field of survival analysis has predominantly focused on interpretability, potentially at some cost of predictive accuracy. This is perhaps the reason why binary classifiers from machine learning are commonly used in industrial applications where survival methodology is applicable. However, while the binary classifiers can provide predictions for one predetermined duration, one loses the interpretability and flexibility provided by modeling the event probabilities as a function of time. Furthermore, in time-to-event data, it is common that some individuals are not followed all the way to their event time, resulting in censored times rather than event times. While binary classifiers typically ignore these observations, one of the main objectives in survival analysis is to account for them. Hence, c©2019 H̊avard Kvamme, Ørnulf Borgan, and Ida Scheel. License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/. Attribution requirements are provided at http://jmlr.org/papers/v20/18-424.html. Kvamme, Borgan, and Scheel in applications with a substantial amount of censoring, the use of survival models tends to be advantageous. In our work, we propose an approach for combining machine learning methodology with survival models. We do this by extending the Cox proportional hazards model with neural networks, and further remove the proportionality constraint of the Cox model. Building on methodology from nested case-control studies (e.g., Langholz and Goldstein, 1996) we are able to do this in a scalable manner. The resulting methods have the flexibility of neural networks while modeling event times continuously. Building on the PyTorch frame- work (Paszke et al., 2017), we provide a python package for our methodology, along with all the simulations and data sets presented in this paper.1 The paper is organized as follows. Section 2 contains a summary of related work. In Section 3, we review some basic concepts from survival analysis and introduce the Cox proportional hazards model with our extensions. In Section 4 we discuss some evaluation criteria of methods for time-to-event prediction. In Section 5, we conduct a simulation study, verifying that the methods we propose behave as expected. In Section 6 we evaluate our methods on five real-world data sets and compare their performances with existing methodology. We conclude in Section 7. 2. Related Work The extension of Cox regression with neural networks was first proposed by Faraggi and Simon (1995), who replaced the linear predictor of the Cox regression model, cf. formula (3) below, by a one hidden layer multilayer perceptron (MLP). It was, however, found that the model generally failed to outperform regular Cox models (Xiang et al., 2000; Sargent, 2001). Katzman et al. (2018) revisited these models in the framework of deep learning and showed that novel networks were able to outperform classical Cox models in terms of the C-index (Harrell Jr et al., 1982). Our work distinguishes itself from this in the following way: The method by Katzman et al. (2018), denoted DeepSurv, is constrained by the proportionality assumption of the Cox model while we propose an extension of the Cox model where proportionality is no longer a restriction. In this regard, we propose an alternative loss function that scales well for both the proportional and the non-proportional cases. Similar works based on Cox regression include SurvivalNet (Yousefi et al., 2017), a frame- work for fitting proportional Cox models with neural networks and Bayesian optimization of the hyperparameters, and Zhu et al. (2016) and Zhu et al. (2017) which extended the Cox methodology to images. Both Zhu et al. (2016) and Zhu et al. (2017) replace the MLP of DeepSurv with a convolutional neural network and applied these methods to pathological images of lung cancer and to whole slide histopathological images. An alternative approach to time-to-event prediction is to discretize the duration and compute the hazard or survival function on this predetermined time grid. Luck et al. (2017) proposed methods similar to DeepSurv, but with an additional set of discrete out- puts for survival predictions and computed an isotonic regression loss over this time grid. Fotso (2018) parameterized a multi-task logistic regression with a neural net that directly computes the survival probabilities on the time grid. Lee et al. (2018) proposed a method, 1. Implementations of methods and the data sets are available at https://github.com/havakv/pycox. 2 Time-to-Event Prediction with Neural Networks and Cox Regression denoted DeepHit, that estimates the probability mass function with a neural net and com- bine the log-likelihood with a ranking loss; see Appendix D for details. Furthermore, the method has the added benefit of being applicable for competing risks. The majority of the papers mentioned benchmark their methods against the random survival forests (RSF) by Ishwaran et al. (2008). RSF computes a random forest using the log-rank test as the splitting criterion. It computes the cumulative hazards of the leaf nodes and averages them over the ensemble. Hence, RSF is a very flexible continuous-time method that is not constrained by the proportionality assumption. 3. Methodology In the following, we give a brief review of some concepts in survival analysis. For a more in-depth introduction to the field, see, for example, Klein and Moeschberger (2003). Our objective is to model the event distribution as a continuous function of time. So with f(t) and F(t) denoting the probability density function and the cumulative distribution function of an event time T∗, we want to model P(T∗ ≤ t) = ∫ t 0 f(s) ds = F(t). As alternatives to F(t), it is common to study the survival function S(t) and the hazard rate h(t). The survival function is defined as S(t) = P(T∗ > t) = 1 −F(t), and is commonly used for visualizing event probabilities over time. For specifying models, however, it is rather common to use the hazard rate h(t) = f(t) S(t) = lim ∆t→0 1 ∆t P(t ≤ T∗ < t + ∆t | T∗ ≥ t). If we have the hazard rate, the survival function can be retrieved through the cumulative hazard, H(t) = ∫ t 0 h(s)ds, by S(t) = exp[−H(t)]. (1) The survival function and the hazard rate therefore provide contrasting views of the same quantities, and it may be useful to study both. Working with real data, the true event times are typically not known for all individuals. This can occur when the follow-up time for an individual is not long enough for the event to happen, or an individual may leave the study before its termination. Instead of observing the true event time T∗, we then observe a possibly right-censored event time T = min{T∗,C∗}, where C∗ is the censoring time. In addition, we observe the indicator D = 1{T = T∗} labeling the observed event time T as an event or a censored observation. Now, denoting individuals by i, with covariates xi and observed duration Ti, the likelihood for censored survival times is given by L = ∏ i f(Ti |xi)DiS(Ti |xi)1−Di = ∏ i h(Ti |xi)Di exp[−H(Ti |xi)]. (2) We will later refer to this as the full likelihood. 3 Kvamme, Borgan, and Scheel 3.1. Cox Regression The Cox proportional hazards model (Cox, 1972) is one of the most used models in survival analysis. It provides a semi-parametric specification of the hazard rate h(t |x) = h0(t) exp[g(x)], g(x) = βT x, (3) where h0(t) is a non-parametric baseline hazard, and exp[g(x)] is the relative risk function. Here, x is a covariate vector and β is a parameter vector. Note that the linear predictor g(x) = βT x does not contain an intercept term (bias weight). This is because the intercept would simply scale the baseline hazard, and therefore not contribute to the relative risk function, h0(t) exp[g(x) + b] = h0(t) exp[b] exp[g(x)] = h̃0(t) exp[g(x)]. The Cox model in (3) is fitted in two steps. First, the parametric part is fitted by max- imizing the Cox partial likelihood, which does not depend on the baseline hazard, then the non-parametric baseline hazard is estimated based on the parametric results. For individual i, let Ti denote the possibly censored event time and Ri denote the set of all individuals at risk at time Ti (not censored and have not experienced the event before time Ti). Note that Ri includes individuals with event times at Ti, so i is part of Ri. The Cox partial likelihood, with Breslow’s method for handling tied event times, is given by Lcox = ∏ i ( exp[g(xi)]∑ j∈Ri exp[g(xj)] )Di , (4) and the negative partial log-likelihood can then be used as a loss function loss = ∑ i Di log   ∑ j∈Ri exp[g(xj) −g(xi)]   . (5) Let β̂ be the value of β that maximizes (4), or equivalently, minimizes (5). Then the cumulative baseline hazard function can be estimated by the Breslow estimator Ĥ0(t) = ∑ Ti≤t ∆Ĥ0(Ti) (6) ∆Ĥ0(Ti) = Di∑ j∈Ri exp[ĝ(xj)] , where ĝ(x) = β̂ T x. If desired, the baseline hazard h0(t) can be estimated by smoothing the increments, ∆Ĥ0(Ti), of the Breslow estimate, but the cumulative baseline hazard typically provides the information we are interested in. 3.2. Cox with SGD The Cox partial likelihood is usually minimized using Newton-Raphson’s method. In our work, we instead want to fit the Cox model with mini-batch stochastic gradient descent 4 Time-to-Event Prediction with Neural Networks and Cox Regression (SGD), to better scale to large data sets. As the loss in (5) sums over risk sets Ri, which can be as large as the full data set, it cannot be computed in batches. Nevertheless, it is possible to do batched iterations by subsampling the data set (to a batch) and restrict the set Ri to only contain individuals in the current batch. This scales well for proportional methods such as DeepSurv (Katzman et al., 2018), but would be very computationally expensive for our non-proportional extension presented in Section 3.4. Hence, we propose an approximation of the loss that is easily batched. Intuitively, we can approximate the risk set Ri with a sufficiently large subset R̃i, and weight the likelihood accordingly with weights wi, L = ∏ i ( exp[g(xi)] wi ∑ j∈R̃i exp[g(xj)] )Di . (7) The weights should ensure that the weighted sum over the subset R̃i in (7) is a reasonable approximation of the full sum over Ri in (4). By choosing a fixed sample size of the sampled risk sets R̃i, we can now optimize the objective by batched gradient descent. The individual i is always included in the sampled risk set R̃i to ensure that each of the products in (7) is bounded above by 1. As the weights wi do not contribute to the gradients of the logarithm of (7) (as can be seen by differentiating with respect to the model parameters), we can simply drop them from the loss function. Also, in practice we do not compute the loss for Di = 0 as these entries do not contribute to (7). Finally, if we average the loss to make it independent of the data set size, we obtain loss = 1 n ∑ i:Di=1 log   ∑ j∈R̃i exp[g(xj) −g(xi)]   , (8) where n denotes the number of events in the data set. In our experiments in Sections 5 and 6, we find that it is often sufficient to sample only one individual j from the risk set, which gives us the loss loss = 1 n ∑ i:Di=1 log (1 + exp[g(xj) −g(xi)]) , j ∈Ri\{i}. (9) One benefit of (8) is that it is, in a sense, more interpretable than the negative partial log- likelihood in (5). Due to the sample dependence in the mean partial log-likelihood (MPLL), i.e., the expression in (5) divided by n, the magnitude of the MPLL is dependent on the size of the risk sets. Hence, for a change of batch size, the mean partial log-likelihood changes. This prohibits a comparison of losses across different batch sizes. Comparably, the loss in (8) is not affected by the choice of batch size, as the size of R̃i is fixed. As a result, we can derive the range of values we expect the loss to be in. Using (9) as an example, we know that it is typically in the range (0, 0.693], as a trivial g(x) = const, gives a loss = log(2) ≈ 0.693, and the minimum is obtained by letting g(xi) →∞, g(xj) →−∞, which results in a loss that tends towards 0. Sampling of the risk sets in Cox’s partial likelihood is commonly done in epidemiology and formalized through the nested case-control design, originally suggested by Thomas 5 Kvamme, Borgan, and Scheel (1977). In (8), case refers to the i’s, while the controls are the j’s sampled from Ri\{i}. Goldstein and Langholz (1992) show that for the Cox partial likelihood, the sampled risk sets produce consistent parameter estimators. While their results do not extend to non- linear models, it is still an indication that the loss function in (8) is reasonable. Our sampling strategy deviates from that of the nested case-control literature in two ways. Firstly, we sample a new set of controls for every iteration, instead of keeping control samples fixed. Secondly, we sample controls with replacement, as this requires less com- putation than sampling without replacement. Note, however, that we typically sample a single control, in which case it does not matter if we sample with or without replacement. 3.3. Non-Linear Cox Having established the simple loss function in (8), which can be computed with SGD, the generalization of the relative risk function exp[g(x)] is rather straightforward. In this paper, we replace the linear predictor g(x) = βT x by a g(x) parameterized by a neural network. While our proposed loss function is not a requirement for the adaptation of a neural network (see, e.g., DeepSurv by Katzman et al., 2018), it really helps for the further extensions in Section 3.4. Also, it has been found that batched iterations can improve predictive performance (Keskar et al., 2016; Hoffer et al., 2017). Our generalization of g(x) leaves the presented theory in Sections 3.1 and 3.2 essentially unchanged, so we do not repeat the likelihoods and loss functions for this model. We will later refer to the Cox proportional hazards model parameterized with a neu- ral network as Cox-MLP. To differentiate between minimizing the negative partial log- likelihood in (5), as done by DeepSurv, and our case-control approximation in (8), we will denote the corresponding methods by Cox-MLP (DeepSurv) and Cox-MLP (CC), respec- tively. For the non-linear Cox models, the loss does not necessarily have a unique minimizer for g(x). Therefore, we add a penalty to the loss function to encourage g(x) to not deviate too far from zero penalty = λ ∑ i:Di=1 ∑ j∈R̃i |g(xj)|. (10) Here λ is a tuning parameter, and note that i is included in R̃i. 3.4. Non-Proportional Cox-Time The proportionality assumption of the Cox model can be rather restrictive, and parame- terizing the relative risk function with a neural net does not affect this constraint. Ap- proaches for circumventing this restriction are typically based on grouping the data based on a categorical covariate and applying a stratified version of the Cox model (Klein and Moeschberger, 2003, chap. 9). We propose a parametric approach that does not require stratification. Continuing with the semi-parametric form of the Cox model, we now let the relative risk function depend on time, h(t |x) = h0(t) exp[g(t, x)]. (11) 6 Time-to-Event Prediction with Neural Networks and Cox Regression In practice, we let g(t, x) handle the time as a regular covariate, which enables g(t, x) to model interactions between time and the other covariates. This is similar to the approach taken in classical survival analysis, where the non-proportional effect of a covariate x may be modeled by including time-dependent covariates like x · t and x · log t. The model (11) is no longer a proportional hazards model. However, it is still a relative risk model with the same partial likelihood as previously, only now with an additional covariate. Following the approach from Section 3.2, we have the loss function loss = 1 n ∑ i:Di=1 log   ∑ j∈R̃i exp[g(Ti, xj) −g(Ti, xi)]   , (12) and we include the penalty in (10), with g(Ti, xj) replacing g(xj). We will later refer to models fitted by (12) as Cox-Time. Note that the loss has the same Ti for both xi and the xj’s. Consequently, if we had used the full risk set Ri instead of the subset R̃i, as is the case for the loss in (5), the loss would become very computationally expensive. In fact, for the full risk set, the time complexity of the loss would be O(n · |Ri|) = O(n2), where |Ri| denotes the size of the risk set. But for (12) we get O(n · |R̃i|) = O(n), as |R̃i| is fixed and small. In the proportional case, to compute the loss in (5) one only needs to compute g(xj) once (per iteration) for each j, and reuse that value in all other risk sets. This ensures the linear time complexity for the classical Cox models. We can find the Breslow estimate for the cumulative baseline hazard H0(t) using (6) with ĝ(xj) replaced by ĝ(Ti, xj). Note that we still need the non-parametric baseline, as g(t, x) is restricted to model interactions between the time and the covariates. To see this, consider g(t, x) = a(t, x) + b(t), and observe that b(t) cancels out in the loss. 3.5. Prediction We can obtain predictions from the relative risk models by estimating the survival function in (1), Ŝ(t |x) = exp[−Ĥ(t |x)]. For the proportional hazards models, the relative risk function does not depend on time, enabling us to integrate only over the baseline hazard and compute the relative risk separately, H(t |x) = ∫ t 0 h0(s) exp[g(x)] ds = H0(t) exp[g(x)]. By first estimating H0(t) on the training data with (6), we only need to compute exp[g(x)] to obtain predictions. Computation of the estimate Ĥ0(t) requires a single pass over the whole training set, in addition to sorting the training set by time. In the case of models with non-proportional hazards, such as models fitted by Cox-Time in Section 3.4, predictions are much more computationally expensive. As the relative risk is time-dependent, we now need to integrate over both the baseline hazard and g(t, x), H(t |x) = ∫ t 0 h0(s) exp[g(s, x)] ds. 7 Kvamme, Borgan, and Scheel In practice, we estimate the cumulative hazards by Ĥ(t |x) = ∑ Ti≤t ∆Ĥ0(Ti) exp[ĝ(Ti, x)], where ∆Ĥ0(Ti) is an increment of the Breslow estimate and ĝ(Ti, x) is the estimate of g(Ti, x) obtained from the neural network. This is clearly rather computationally expensive as we need to compute ĝ(Ti, x) for all distinct event times Ti ≤ t. Furthermore, for continuous- time data, computation of the cumulative baseline hazard through the Breslow estimate, ∆Ĥ0(Ti) = Di∑ j∈Ri exp[ĝ(Ti, xj)] , (13) scales quadratically. To alleviate the computational cost, one can compute the cumulative hazards over a reduced number of distinct time points. Hence, Cox-Time is trained on continuous-time data but produces discrete-time predictions, with the benefit of the discretization happening after the network is fitted. In practice, we perform this discretization by computing the baseline on a random subset of the training data and subsequently control the resolution of the time grid through the sample size. 4. Evaluation Criteria Metrics for evaluating the performance of methods for time-to-event prediction should ac- count for the censored individuals. In the following, we describe the metrics used in the experimental sections of this paper. 4.1. Concordance Index In survival analysis, the concordance index, or C-index (Harrell Jr et al., 1982), is arguably one of the most commonly applied discriminative evaluation metrics. This is likely a result of its interpretability, as it has a close relationship to classification accuracy (Ishwaran et al., 2008) and ROC AUC (Heagerty and Zheng, 2005). In short, the C-index estimates the probability that, for a random pair of individuals, the predicted survival times of the two individuals have the same ordering as their true survival times. See Ishwaran et al. (2008) for a detailed description. As the C-index only depends on the ordering of the predictions, it is very useful for evaluating proportional hazards models. This is because the ordering of proportional haz- ards models does not change over time, which enables us to use the relative risk function instead of a metric for predicted survival time. It is, however, not obvious how the C-index should be applied for non-proportional hazards models (Gerds et al., 2012; Ishwaran et al., 2008). We will use a metric based on the time-dependent C-index by Antolini et al. (2005), which estimates the probability that observations i and j are concordant given that they are comparable, Ctd = P{Ŝ(Ti |xi) < Ŝ(Ti |xj) | Ti < Tj, Di = 1}. (14) 8 Time-to-Event Prediction with Neural Networks and Cox Regression However, to account for tied event times and survival estimates, we make the modifications listed by Ishwaran et al. (2008, Section 5.1, step 3). This is to ensure that predictions independent of x, Ŝ(t |x) = Ŝ(t), yields Ctd = 0.5 for unique event times. Note that for proportional hazards models, our metric is equivalent to the regular C-index. 4.2. Brier Score The Brier score (BS) for binary classification is a metric of both discrimination and cal- ibration of a model’s estimates. In short, for N binary labels yi ∈ {0, 1} with probabil- ities pi of yi = 1, the BS is the mean squared error of the probability estimates p̂i, i.e., BS = 1 N ∑ i (yi − p̂i) 2 . To get binary outcomes from time-to-event data, we choose a fixed time t and label data according to whether or not an individual’s event time is shorter or longer than t. Graf et al. (1999) generalize the Brier score to account for censoring by weighting the scores by the inverse censoring distribution, BS(t) = 1 N N∑ i=1 [ Ŝ(t |xi) 2 1{Ti ≤ t,Di = 1} Ĝ(Ti) + (1 − Ŝ(t |xi)) 2 1{Ti > t} Ĝ(t) ] . (15) Here N is the number of observations, Ĝ(t) is the Kaplan-Meier estimate of the censoring survival function, P(C∗ > t), and it is assumed that the censoring times and survival times are independent. The BS can be extended from a single duration t to an interval by computing the integrated Brier score IBS = 1 t2 − t1 ∫ t2 t1 BS(s) ds. (16) In practice, we approximate this integral by numerical integration, and we let the time span be the duration of the test set. In our experiments we found that using 100 grid points were sufficient to obtain stable scores. 4.3. Binomial Log-Likelihood The mean binomial log-likelihood is a commonly used metric in binary classification that measures both discrimination and calibration of the estimates of a method. Using the same inverse censoring weighting as for the Brier score, we can apply this metric to censored duration time data, BLL(t) = 1 N N∑ i=1 [ log[1 − Ŝ(t |xi)] 1{Ti ≤ t,Di = 1} Ĝ(Ti) + log[Ŝ(t |xi)] 1{Ti > t} Ĝ(t) ] . (17) The binomial log-likelihood can also be integrated in the same manner as (16) IBLL = 1 t2 − t1 ∫ t2 t1 BLL(s) ds. 9 Kvamme, Borgan, and Scheel 0 10 20 30 40 50 60 Epochs 8.7 8.6 8.5 8.4 8.3 8.2 M PL L 1 2 4 8 Figure 1: Box plots giving the mean partial log-likelihood (MPLL) of the test sets for different training epochs. The colors show how many controls were sampled during training (in addition to the case). 5. Simulations To empirically investigate our proposed methodology, we conducted a series of simulations. These experiments are by no means exhaustive but are rather meant to verify that the methods behave as expected. In the following, we let classical Cox refer to a Cox regression with g(x) = βT x obtained with the Lifelines python package (Davidson-Pilon et al., 2018). For experimental details exempt from the main article, we refer the reader to Appendix C.3. We first investigate the behavior of our proposed loss (8). In particular, we want to examine the impact the number of sampled controls has on the fitted models, in addition to how well the results from using our loss agree with those from using the Cox partial log- likelihood, i.e., the loss (5). To this end, we simulated survival times from a proportional hazards model h(t |x) = h0(t) exp[g(x)], g(x) = βT x, (18) with constant baseline hazard h0(t) = 0.1, and β T = [0.44, 0.66, 0.88]. The covariates were sampled uniformly from [−1, 1]. We drew censoring times independent of the covariates with constant hazard c(t) = 1 30 , and, in addition, we censored all individuals that were still under observation at time 30. This resulted in approximately 30 % censored individuals. We sampled 10,000 individuals for training and 10,000 for testing, and fitted our Cox model by SGD as described in Section 3.2. This method will be referred to as Cox-SGD. Four models were fitted by sampling 1, 2, 4, and 8 controls (in addition to the case). The whole experiment was repeated 100 times, and the mean partial log-likelihood (MPLL) of the test sets are visualized in Figure 1. The figure indicates that the number of sampled controls does not affect the rate of convergence, but we note that the computational complexity increases with the number of sampled controls. 10 Time-to-Event Prediction with Neural Networks and Cox Regression 100 1000 10000 Size data set 0.3 0.2 0.1 0.0 0.1 D iff . C oe f 0.44 100 1000 10000 Size data set 0.66 1 2 4 8 100 1000 10000 Size data set 0.88 Figure 2: Differences between Cox-SGD and Classical Cox parameter estimates for different data set sizes. The numbers above the plots give the value of the true coefficient. Each box plot is based on 100 observations. The legend above the plots states the number of sampled controls for Cox-SGD. The experiment was repeated with training sets of 1,000 individuals to verify that the results were not simply due to the size of the training data. The same patterns were found in this setting, so the figure is exempted from the paper. Next, we compare the parameter estimates obtained from our proposed loss (Cox-SGD) with the estimates obtained with classical Cox regression. For data sets of sizes 100, 1,000, and 10,000, we fitted models to 100 sampled data sets. The differences between the Cox- SGD parameter estimates and the classical Cox estimates are displayed in Figure 2, where the legend above the plots gives the number of controls sampled for the Cox-SGD method. For the data sets of size 100, we observe that the Cox-SGD estimates seem to be slightly smaller than the Cox estimates, and this difference is larger for fewer sampled controls. However, as the data sets increase in size, the estimates for the two methods agree well. Finally, we want to compare the likelihoods obtained by the two methods. As the mean partial log-likelihood depends on data set size, it is really not comparable across data sets. We will therefore instead use the full likelihood (2), which also depends on the baseline hazard, to compare the methods. Note that the partial likelihood may be interpreted as a profile likelihood, so the full likelihood and the partial likelihood are closely related (see, e.g., Klein and Moeschberger, 2003, p. 258). We obtain cumulative baseline hazard estimates by (6), and baseline hazard estimates by numerical differentiation of the cumulative baseline hazard estimates. We report the mean log-likelihood (MLL) of (2) to compare results for different data set sizes. In Figure 3, we show the difference in the MLL between the Cox- SGD method and the classical method for Cox regression (for each sampled data set). We used the training MLL as we are here only interested in the losses’ abilities to optimize the objective, and not the generalization to a test set. From the figure, we observe that, for smaller data sets, more sampled controls in Cox-SGD seems to give likelihood estimates closer to those of a classical Cox regression. As the data sets increase in size, the MLL of the Cox-SGD seems to converge to that of the classical Cox, regardless of control sample 11 Kvamme, Borgan, and Scheel 100 1000 10000 Size data set 0.010 0.008 0.006 0.004 0.002 0.000 D iff . M LL 1 2 4 8 Figure 3: Differences in mean log-likelihood between Cox-SGD and classical Cox using (2). The figure gives results for different data set sizes, and the legend gives the number of sampled controls. Each box plot is based on 100 observations. size. The MLL’s of the classical Cox regression is approximately -2.2, meaning that even for the smallest data sets the differences in MLL for 1 sampled control is around 0.1 %. Hence, it seems that our loss (8) approximates the negative partial log-likelihood rather well. Furthermore, while a higher number of sampled controls can give lower training error, the effect of the number of sampled controls decreases with the size of the data sets. In the further experiments, we will for simplicity only use one sampled control, as this was found sufficient. Furthermore, in Appendix C we have included simulations for the methods using neural networks for non-linear and non-proportional models, to verify that our proposed methods behave as expected. In summary, these simulations verify that for observations drawn from a proportional hazards model with a non-linear relative risk function, Cox models parameterized by neural networks provide better estimates of the partial log-likelihood than classical Cox models. Further, by drawing observations from a non-proportional relative risk model, the simulations verify that our Cox-Time method (Section 3.4) is able to obtain better estimates of the survival functions than methods which assume proportional hazards. 6. Experiments In the following, we will evaluate our proposed methods on real data sets and compare their performance to existing methods from the literature. In total, we use five data sets. One large data set for a more in-depth analysis (see Section 6.2), and four smaller data sets commonly used in the survival analysis literature. 6.1. Four Common Survival Data Sets We base this experimental section on the data sets provided by Katzman et al. (2018), as they are made available through the DeepSurv python package, and need no further prepro- 12 Time-to-Event Prediction with Neural Networks and Cox Regression Data set Size Covariates Unique Durations Prop. Censored SUPPORT 8,873 14 1,714 0.32 METABRIC 1,904 9 1,686 0.42 Rot. & GBSG 2,232 7 1,230 0.43 FLCHAIN 6,524 8 2,715 0.70 Table 1: Summary of the four data sets used in the experiments in Section 6.1. cessing. The data sets include the Study to Understand Prognoses Preferences Outcomes and Risks of Treatment (SUPPORT), the Molecular Taxonomy of Breast Cancer Inter- national Consortium (METABRIC), and the Rotterdam tumor bank and German Breast Cancer Study Group (Rot. & GBSG). Katzman et al. (2018) also provide the Worcester Heart Attack Study (WHAS) data set. However, their version of the data set is actually a case-control data set, meaning it contains multiple replications of some individuals, some- thing the authors seem to have overlooked. We replace it with the Assay Of Serum Free Light Chain (FLCHAIN) made available in the survival packages of R (Therneau, 2015). For FLCHAIN, we remove individuals with missing values. Further, we remove the “chap- ter” covariate, which gives the cause of death. Table 1 provides a summary of the data sets. For a more detailed description, we refer to the original sources (Therneau, 2015; Katzman et al., 2018). 6.1.1. Methods and Hyperparameter Tuning In Sections 3.3 and 3.4 we presented two new survival methods based on case-control sam- pling and neural networks: a proportional Cox method and a non-proportional Cox method, which we will refer to as Cox-MLP (CC) and Cox-Time respectively. We will compare our methods to a classical linear Cox regression referred to as Classical Cox (Linear), DeepHit (Lee et al., 2018), and Random Survival Forests (RSF) (Ishwaran et al., 2008). We will also compare to a proportional Cox method similar to DeepSurv (Katzman et al., 2018), but our version performs batched SGD by computing the negative partial log-likelihood in (5) on a subset of the data set. Furthermore, we choose not to restrict the network structure and optimization scheme to that of Katzman et al. (2018). Hence, this method is identical to our proportional Cox method in Section 3.3, except that it computes the negative partial log-likelihood of a batch, while we use case-control sampling in the loss function. We will refer to these two methods as Cox-MLP (DeepSurv) and Cox-MLP (CC) respectively. We do not compare with Luck et al. (2017) as their method is another proportional hazard method and is therefore restricted in all the same ways as our other proportional methods. As we will show in Section 6.1.2, the proportional hazards assumption is very restrictive, and methods based on this assumption are therefore not able to compete with other methods such as DeepHit, RSF, and Cox-Time. For the neural networks, we standardize the numerical covariates and encode the cat- egorical covariates by entity embeddings (Guo and Berkhahn, 2016) half the size of the number of categories. For the Classical Cox (Linear) regression, we one-hot encoded the categorical covariates (dummy variables), and in RSF we simply passed the covariates with- out any transformations. The networks are standard multi-layer perceptrons with the same 13 Kvamme, Borgan, and Scheel Method SUPPORT METABRIC Rot. & GBSG FLCHAIN Classical Cox (Linear) 0.598 0.628 0.666 0.790 Cox-MLP (DeepSurv) 0.611 0.636 0.674 0.790 Cox-MLP (CC) 0.613 0.643 0.669 0.793 Cox-Time 0.629 0.662 0.677 0.790 DeepHit 0.642 0.675 0.670 0.792 RSF (Mortality) 0.628 0.649 0.667 0.784 RSF (Ctd) 0.634 0.652 0.669 0.786 Table 2: Concordance (Ctd) for the experiments in Section 6.1. number of nodes in every layer, ReLU activations, and batch normalization between layers. We used dropout, normalized decoupled weight decay (Loshchilov and Hutter, 2019), and early stopping for regularization. SGD was performed by AdamWR (Loshchilov and Hut- ter, 2019) with an initial cycle length of one epoch, and we double the cycle length after each cycle. Learning rates were found using the methods proposed by Smith (2017). As the four data sets are somewhat small, we scored our fitted models using 5-fold cross- validation, where the hyperparameter search was performed individually for each fold. For all the neural networks, we performed a random hyperparameter search over 300 parameter configurations and chose the model with the best score on a held-out validation set. We scored the proportional Cox methods by the partial log-likelihood, and for the Cox-Time method, we used the loss (12). Model hyperparameter tuning for DeepHit and RSF was performed with the time-dependent concordance index (Antolini et al., 2005). However, we also include an RSF tuned in the manner proposed by the authors, i.e., by computing the concordance of the mortality (Ishwaran et al., 2008, Section 4.1). The two versions of RSF are in the following denoted RSF (Ctd) and RSF (Mortality), respectively. A list of the hyperparameter search spaces can be found in Appendix A.1. 6.1.2. Results We compare the methods using the metrics presented in Section 4, i.e., the time-dependent concordance, the integrated Brier score, and the integrated binomial log-likelihood. While the concordance solely evaluates a method’s discriminative performance, the Brier score and binomial log-likelihood also evaluate the calibration of the survival estimates. Table 2 shows the time-dependent concordance, or Ctd, averaged over the five cross- validation folds. As expected, the methods that assume proportional hazards, in general, perform worse than the less restrictive methods, and Cox-MLP (DeepSurv) and Cox-MLP (CC) are very close to each other. We see that RSF (Ctd) has better concordance than RSF (Mortality), which is expected as RSF (Ctd) uses the same metric for hyperparameter tuning as for evaluation. Cox-Time seems to do slightly better than the RSF methods, which is impressive as we have not used concordance for hyperparameter tuning. DeepHit seems to have the best discriminative performance overall, but, as we will see next, this comes at the cost of poorly calibrated survival estimates. 14 Time-to-Event Prediction with Neural Networks and Cox Regression Method SUPPORT METABRIC Rot. & GBSG FLCHAIN Classical Cox (Linear) 0.217 0.183 0.180 0.096 Cox-MLP (DeepSurv) 0.214 0.176 0.170 0.092 Cox-MLP (CC) 0.213 0.174 0.171 0.093 Cox-Time 0.212 0.172 0.169 0.102 DeepHit 0.223 0.184 0.178 0.120 RSF (Mortality) 0.215 0.175 0.171 0.093 RSF (Ctd) 0.212 0.176 0.171 0.093 Table 3: Integrated Brier score weighted by estimates of the censoring distribution for the experiments in Section 6.1. Method SUPPORT METABRIC Rot. & GBSG FLCHAIN Classical Cox (Linear) -0.623 -0.538 -0.529 -0.322 Cox-MLP (DeepSurv) -0.619 -0.532 -0.514 -0.309 Cox-MLP (CC) -0.615 -0.515 -0.509 -0.314 Cox-Time -0.613 -0.511 -0.502 -0.432 DeepHit -0.637 -0.539 -0.524 -0.487 RSF (Mortality) -0.619 -0.515 -0.507 -0.311 RSF (Ctd) -0.610 -0.517 -0.507 -0.311 Table 4: Integrated binomial log-likelihood weighted by estimates of the censoring distri- bution for the experiments in Section 6.1. Tables 3 and 4 show the integrated Brier score and integrated binomial log-likelihood, both weighted with Kaplan-Meier estimates of the censoring distribution. Here, for both metrics, closer to zero is better. We find that both metrics yield very similar results, as the orderings of the methods are almost identical. First, we see that Cox-Time seems to generally perform the best, but it struggles with the FLCHAIN data set. However, we note that, for FLCHAIN, Cox-MLP (DeepSurv) has the best integrated Brier score and binomial log-likelihood while Cox-MLP (CC) has the best concordance. This indicates that the proportionality assumption is quite reasonable for this data set. Again, we find that there is very little difference between Cox-MLP (DeepSurv) and Cox-MLP (CC), which is as expected. The RSF methods generally perform well. Note that the two RSF methods perform equally well here, even though RSF (Ctd) had the best concordance. Classical Cox (Linear) does rather poorly, which was expected as it has very restrictive model assumptions. Even though DeepHit, in general, had the best concordance, it has the worst integrated Brier score and binomial log-likelihood in three out of the four data sets. The loss function in DeepHit, given by formula (D.2) in Appendix D, is a convex combination of the negative log-likelihood and a ranking loss, determined by a parameter α. For α = 1 we obtain only the negative log-likelihood and for α = 0 we obtain only the ranking loss. As we 15 Kvamme, Borgan, and Scheel Data set Size Churned Censored Prop. Censor Unique users Train 1,786,333 1,279,358 506,975 0.28 1,582,202 Test 661,748 473489 188,259 0.28 586,001 Validation 198,665 142,104 56,561 0.28 175,801 Table 5: Summary of the KKBox churn data set. do hyperparameter tuning based on the concordance, we see that α tends towards smaller values, which results in excellent discriminative performance at the cost of poorly calibrated survival estimates. 6.2. KKBox Churn Case Study Thus far we have compared competing survival methodologies on fairly small data sets. We now perform a case study on a much larger data set, as this is more interesting in the context of neural networks. The WSDM KKBox’s churn prediction challenge was proposed in preparation for the 11th ACM International Conference on Web Search and Data Mining.2 The competition was hosted by Kaggle in 2017, with the goal of predicting customer churn on a data set donated by KKBox, the leading music streaming service in Asia. The available data provide us the opportunity to create a survival data set with event times and censoring indicators. We stick with the churn definition given by KKBox, were a customer is considered churned if he or she fails to resubscribe within 30 days after the previous subscription expired. Note, however, that our use of the data is not comparable to the Kaggle competition, as we work with survival times from the start of a subscription period, while they consider durations from a fixed calendar date. KKBox provides multiple data sources, but as we are primarily interested in evaluating our methods, we spend less time on feature engineering and only use a subset of covariates with general customer information (e.g., city, age, price of subscription). Furthermore, a customer that has previously churned and later resubscribed, is treated as a new customer with some covariate information describing the previous subscription history. This gives us a total of 15 covariates. We split the data into a training, a testing, and a validation set, and some information about these subsets are listed in Table 5. A more in-depth description of the KKBox data set can be found in Appendix B.1. 6.2.1. Methods and Hyperparameter Tuning We use the same methods as in Section 6.1. However, as this data set is very large, we replace the classical Cox regression with our Cox-SGD (Linear) method from Section 3.2. We standardize and encode the covariates in the same manner as for the smaller data sets. The networks and training are also the same as earlier, but we multiply the learning rate by 0.8 at the start of every cycle, as we found this to give more stable training. The KKBox data set is quite large, so we are not able to explore as large a hyperparameter space as in the previous experiments. Hence, we do not include weight decay, and perform 2. https://www.kaggle.com/c/kkbox-churn-prediction-challenge 16 Time-to-Event Prediction with Neural Networks and Cox Regression Method Layers Nodes Dropout α σ Cox-MLP (DeepSurv) 6 256 0.1 - - Cox-MLP (CC) 6 128 0 - - Cox-Time 8 256 0 - - DeepHit 6 512 0.1 0.001 0.5 Table 6: KKBox model architectures. α and σ only applies the DeepHit (see Appendix D). Method Ctd IBS IBLL Cox-SGD (Linear) 0.816 0.127 -0.406 Cox-MLP (DeepSurv) 0.841 0.111 -0.349 Cox-MLP (CC) 0.844 0.119 -0.379 Cox-Time 0.861 0.107 -0.334 DeepHit 0.888 0.147 -0.489 RSF (Mortality) 0.855 0.112 -0.352 RSF (Ctd) 0.870 0.111 -0.352 Table 7: Evaluation metrics for the KKBox data. a grid search over a small number of suitable parameters. The hyperparameter search is described in detail in Appendix A.2. The best configurations are given in Table 6. For RSF, the hyperparameter search based on Ctd yielded 8 covariates sampled for each split and a minimum of 50 observations in each leaf node. With concordance of mortality as the validation criterion, the best fitted model used 2 covariates for splitting, and a minimum leaf node size of 10. Furthermore, we found that 500 trees were sufficient, as there was little improvement compared to 250 trees. 6.2.2. Results For evaluation, we fitted each of the methods five times and computed the time-dependent concordance index (Ctd), the integrated Brier score (IBS), and the integrated binomial log- likelihood (IBLL) of the fitted models. The median scores are presented in Table 7. We use the median because the two proportional Cox-MLP methods yielded rather unstable results, where some of the fitted models performed very badly. From the table, we see that DeepHit continues to outperform the other methods in terms of concordance while having the worst performance in terms of IBS and IBLL. Furthermore, Cox-Time has the best IBS and IBLL, while still providing a decent concordance. RSF continues to do well across all metrics, while again, the tuning based on Ctd seems to yield better results than tuning base on the concordance of the mortality. Cox-SGD (Linear) does rather poorly, as it is very restricted, and serves more as a baseline in this context. The Cox-MLP methods seem to again perform reasonably close to each other, at least when taking into account that we found both methods to yield rather unstable results. We are 17 Kvamme, Borgan, and Scheel 0 100 200 300 400 500 600 700 800 Time 0.000 0.025 0.050 0.075 0.100 0.125 0.150 0.175 0.200 Br ie r s co re Cox-SGD (Linear) Cox-MLP (DeepSurv) Cox-MLP (CC) Cox-Time DeepHit RSF (Mortality) RSF (Ctd) Figure 4: Brier score on KKBox data set. The methods are the same as in Table 7. not sure why this was the case, but note that the combination of the flexible neural net and the proportionality constraint might be problematic for large data sets. In Figure 4, we display the Brier scores used to obtain the IBS. Again, we see that DeepHit does poorly for all durations. The instability of the Cox-MLP (CC) is also very apparent for shorter durations. Cox-Time is clearly doing very well for all durations, but interestingly, we observe the Cox-MLP (DeepSurv) provides the best fitted model for larger durations. We could make a corresponding figure for the binomial log-likelihood, but as it is very similar to the Brier score plot, and provide us with no new insights, we have not included it. 6.2.3. Survival Curves One of the benefits survival models have over binary classifiers is their ability to produce survival curves. In Figure B.1 in Appendix B, we show nine examples of estimated survival curves from Cox-Time on the test data. The curves nicely illustrate the extent of detail the method has learned. To obtain a more general view of the predictions, we cluster the estimated survival curves of the test set. For an equidistant grid 0 = τ0 < τ1 < · · · < τm, the survival curve of individual i is represented by a vector [Ŝ(τ0 |xi), Ŝ(τ1 |xi), . . . , Ŝ(τm |xi)], and by considering these as feature vectors, we apply K-means clustering to the test set with 10 clusters. In Figure 5, we display the cluster centers and the proportions of the test set assigned to each of the clusters. This is a reasonable approach for segmenting customers based on their churning behavior. A natural next step would be to further investigate the clusters, but as we consider this somewhat outside the scope of this paper, we only make 18 Time-to-Event Prediction with Neural Networks and Cox Regression 0 100 200 300 400 500 600 700 800 Time 0.0 0.2 0.4 0.6 0.8 1.0 S( t) 19 % 18 % 13 % 12 % 10 % 7 % 7 % 7 % 4 % 3 % Figure 5: Cluster centers of survival curves from KKBox test data. The centers were gener- ated by K-means clustering on survival curves from Cox-Time. The legend gives the proportion of test data assigned to each cluster (rounded to nearest integer). a few observations. First, we see that 19 % of the customers are assigned to a cluster that does not provide much detailed information about their behavior, but instead provides a survival curve with a rather constant slope. In sharp contrast, the second largest group (18 %) is at high risk of churning immediately. Furthermore, we observe that many of the curves seem to have higher hazards at the end of each month (drops in the survival curves around 30, 60, 90 days), and we hypothesize that this is a result of customers paying for a full month at a time. The smallest cluster, constituting only 3 % of the test data, has a sharp drop around day 400. Investigating the covariates of the assigned customers reveals that most of them had prepaid for a 411 days long subscription. However, the large drop after 400 days indicates that only around 25 % of them were interested in continuing their subscription. Our choice of 10 clusters is mainly motivated by how many curves that can be visualized in a single plot, without being too crowded. Further increasing the number of clusters would likely reveal more detailed behavior. 7. Discussion In this paper, we propose extensions of the Cox proportional hazards model. By parame- terizing the relative risk function of a Cox model with neural networks, we can model rich relationships between the covariates and event times. Furthermore, we allow the networks to model interactions between the covariates and time, resulting in models that are no longer constrained by the proportionality assumption of the Cox model. Building on methods for nested case-control studies, we propose a loss function that can be computed in batches, enabling the models to scale to large data sets. We verify through simulation studies that fitting a Cox model using our proposed loss function gives results close to those obtained using the full Cox partial likelihood. 19 Kvamme, Borgan, and Scheel We compare our suggested methodology with classical Cox regression, random survival forests (RSF), DeepHit, and DeepSurv on 5 real-world data sets, and find that our proposed Cox-Time method performs very well, and has the best overall performance in terms of integrated Brier score (IBS) and integrated binomial log-likelihood (IBLL). DeepHit has, in general, the best discriminative performance, but this comes at the cost of poorly calibrated survival estimates. Finally, we show how estimated survival curves (event probabilities as functions of time), can be used as a descriptive tool to better understand event-time data sets. This is illus- trated by an example where we cluster the survival estimates of a customer churn data set, and show that this customer segmentation provides a useful view of the churning process. Interesting expansions of our methodology include the extension to handle multiple competing events, time-dependent covariates, dynamic predictions, and recurrent events. Furthermore, it would be of interest to explore other data sources that require more ad- vanced network structures such as convolutions and recurrent neural networks. Finally, less computationally expensive alternatives for creating survival estimates for the Cox-Time method should be explored. Acknowledgments This work was supported by The Norwegian Research Council 237718 through the Big Insight Center for research-driven innovation. Appendix A. Details on Hyperparameter Tuning In the following, we provide further details about the experiments conducted in Section 6. Here we list hyperparameter configurations and details about the model fitting procedures. A.1. Four Common Survival Data Sets Tuning Table A.1 gives the hyperparameter search space used for Rot. & GBSG, SUPPORT, METABRIC, and FLCHAIN. The square brackets describe continuous variables. In the ex- periments in Section 6.1, we sample 300 random parameter configurations for each method, for each fold of each data set. In the table, “α” and “σ” are given in (D.2) in Appendix D, “Num. durations” are the number of discrete durations (equidistant) used in DeepHit, “λ” is the penalty in (10), “Log duration” refers to a log-transform of the durations passed to Cox-Time, “Ridge” is a ridge penalty used in classical Cox regression, “Split covariates” and “Size leaf” are the number of covariates used for splitting, and the minimum node size of RSF. A.2. KKBox Tuning Hyperparameters in the KKBox study were found by a grid search over the relevant pa- rameters in Table A.2. The table consists of three sections, where the top represents the networks, the bottom represents RSF, and the middle contains network parameters that were found on a smaller network with two layers and 128 nodes. “α” and “σ” controls the 20 Time-to-Event Prediction with Neural Networks and Cox Regression Hyperparameter Values Layers {1, 2, 4} Nodes per layer {64, 128, 256, 512} Dropout [0, 0.7] Weigh decay {0.4, 0.2, 0.1, 0.05, 0.02, 0.01, 0} Batch size {64, 128, 256, 512, 1024} α (DeepHit) [0, 1] σ (DeepHit) {0.1, 0.25, 0.5, 1, 2.5, 5, 10, 100} Num. durations (DeepHit) {50, 100, 200, 400} λ (Cox-Time and Cox-MLP (CC)) {0.1, 0.01, 0.001, 0} Log durations (Cox-Time) {True, False} Ridge (Cox (Linear)) {1000, 100, 10, 1, 0.1, 0.01, 10−3, 10−4, 10−5} Split covariates (RSF) {2, 4, 6, 8} Size leaf (RSF) {2, 8, 32, 64, 128} Table A.1: Hyperparameter search space for experiments on Rot. & GBSG, SUPPORT, METABRIC, and FLCHAIN. Hyperparameter Values Layers {4, 6, 8} Nodes per layer {128, 256, 512} Dropout {0, 0.1, 0.5} α(*) {0, 0.001, 0.1, 0.2, 0.5, 0.8, 0.9, 0.99, 0.999, 1} σ(*) {0.01, 0.1, 0.25, 0.5, 1, 10, 100} Log durations(*) {True, False} Split covariates {2, 4, 6, 8} Size leaf {8, 10, 20, 50} Table A.2: KKBox hyperparameter configurations. (*) denotes parameters found with a two layer network with 128 nodes. 21 Kvamme, Borgan, and Scheel 0 50 100 150 200 250 300 350 Time (days) 0.0 0.2 0.4 0.6 0.8 1.0 S( t) Figure B.1: Survival curves from the case study in Section 6.2, which models the times at which customers of a streaming service churn. Each curve gives the estimated probabilities for a customer not having churned (probabilities of still being a customer at any given time). The time axis shows the number of days since first subscription. The curves are generated by the Cox-Time method from Section 3.4. loss function of DeepHit, and we assumed it should generalize well across network struc- tures. The same goes for whether or not we should log-transform the durations passed to Cox-Time. Hence, to reduce the hyperparameter search, we found suitable values with a smaller network. For the proposed Cox-MLP (CC) and Cox-Time, we used a fixed penalty λ = 0.001 in (10). All networks were trained with batch size of 1028, and the best performing archi- tectures can be found in Table 6. Appendix B. Details from KKBox Churn Case Study In the following, we provide some details of the KKBox case study that were exempt from the main article. In Figure B.1, we show an example of nine survival curves estimated by Cox-Time on the KKBox data set. Each line represents an individual from the test set. It is clear that the Cox-Time method has learned to represent a variety of survival curves. B.1. KKBox Data Set KKBox provides data consisting of general customer information (city, age, gender, initial registration), transaction logs listing how customers manage their subscription plans, and user logs containing some aggregate information of the customers’ usage of the streaming 22 Time-to-Event Prediction with Neural Networks and Cox Regression service. KKBox defines a customer as churned if he or she has failed to resubscribe to their service more than 30 days after their last subscription expired. Through the transaction information, we can create a data set with survival times and censoring indicators. We keep KKBox’s original churn definition of 30 days absence and calculate the survival time from the date of the first subscription or the earliest record of the customer in the transaction logs. Customers that have previously churned but resubscribed, are treated as new customers with some covariate information describing their previous subscription history. All covariates are calculated at the time of subscription of each customer, and they do not change with time. This is a simplification, as the covariates of a customer are typically not stationary. However, as our objective is to evaluate the proposed methodology, we refrain from doing extensive feature engineering. In this regard, we further disregard the user logs, as we get a reasonable set of covariates from the customer and transaction information. The data sets are summarized in Table 5. As some of the customers have churned multiple times, and are therefore included multiple times, the table also includes the number of unique customers in the respective data sets. We have a total of 15 covariates, where 7 are numeric and 8 are categorical. The numerical covariates give the time between subscriptions (for customers that have previously churned), number of previous churns, time since first registration, number of days until the current subscription expires, listed price of the current subscription, the amount paid for the subscription, and age. All numerical covariates, except the number of previous churns, were log-transformed. The categorical covariates are gender (3 categories), city (22 categories), how the customer registered (6 categories), and 5 indicator variables specifying if an individual has previously churned, if the subscription is canceled, if the age is unrealistic, if the subscription is automatically renewed, and if we do not know when the customer first registered. Appendix C. Additional Simulations In the following, we provide some additional simulations used to further investigate the proposed methods, and in Appendix C.3 we explain how the simulated data were generated. We again stress that the aim of the simulations is only to verify the expected behavior of our methods, and should not be interpreted as a general evaluation of them. C.1. Non-Linear Hazards Continuing from the simulations in Section 5, we do a simple study of the increased flexibility provided by replacing the linear predictor in a Cox regression by a neural network. To evaluate this, we extend the simulations in Section 5 by replacing g(x) = βT x with the non-linear function g(x) = βT x + 2 3 (x21 + x 2 3 + x1x2 + x1x3 + x2x3), (C.1) where xi denotes element i of x. The simulations are otherwise unchanged from the linear case in Section 5. We sample 10,000 training samples, 10,000 test samples, and 1,000 samples used for validation (validation data is used for early stopping), and fit the Cox-SGD and 23 Kvamme, Borgan, and Scheel 11 10 9 8 7 6 5 4 True PLL 11 10 9 8 7 6 5 4 Es tim at ed P LL Cox-SGD Cox-MLP Figure C.1: Partial log-likelihood estimates of Cox-SGD and Cox-MLP plotted against the true partial log-likelihoods, for 2,000 samples of the test set. classical Cox regression from Section 5. Additionally, we fit the Cox model in Section 3.3, where g(x) is parameterized by a one-hidden layer MLP (multilayer perceptron) with 64 hidden nodes and ReLU activations. This model will be referred to as Cox-MLP (we drop (CC) as we do not consider DeepSurv here). In Figure C.1 we have, for 2,000 individuals of the test set, plotted the individual partial log-likelihood estimates of Cox-SGD and Cox-MLP against the true individual partial log- likelihoods. That is, for each i with Di = 1, we plot ˆ̀ i = − log   ∑ j∈Ri exp[ĝ(xj) − ĝ(xi)]   against the true `i. The closer a method estimates g(x) to the true predictor in (C.1), the closer the scatter plot should be to the identity function (straight line with slope 1 through the origin). As expected, we see that Cox-MLP produces likelihood estimates very close to the true likelihoods, while Cox-SGD struggles to represent this non-linear function. C.2. Non-Proportional Hazards In our final simulations, we investigate the effect of removing the proportionality constraint in the Cox models. Building on the previous simulations, we add a time dependent term to the risk function. The hazard function is now given by h(t |x) = h0 exp[g(t, x)], with h0 = 0.02 and g(t, x) = a(x) + b(x) t, a(x) = gph(x) + sign(x3), b(x) = |0.2 (x0 + x1) + 0.5 x0x1|, 24 Time-to-Event Prediction with Neural Networks and Cox Regression 12 10 8 6 4 True PLL. 12 10 8 6 4 Es tim at ed P LL Cox-MLP Cox-Time Figure C.2: Partial log-likelihood estimates of Cox-MLP and Cox-Time plotted against the true partial log-likelihoods, for 2,000 samples of the test set. were gph(x) is the function in (C.1). The simulations are otherwise unchanged from the previous experiments. We require the term b(x) to be non-negative to ensure that the haz- ards increase with time. Furthermore, we have added sign(x3) to a(x) as this is essentially equivalent to having two different baselines, h̃0 = h0 exp[sign(x3)] ∈ {0.0074, 0.054}. Both of these choices are motivated by our attempt to produce reasonable looking survival curves. We sample 10,000 training samples, 10,000 test samples, and 2,000 samples for valida- tion, and fit a Cox-MLP model and a Cox-Time model. Both Cox-MLP and Cox-Time parameterize g with a 4 hidden layer MLP with ReLU activations, 128 nodes in each layer, and dropout between layers with rate 0.1. In Figure C.2, we have plotted the estimated partial log-likelihoods of Cox-MLP and Cox-Time against the true partial log-likelihoods for 2,000 samples of the test set. Though the difference between the methods is not very large, Cox-Time appears to capture the true values better than Cox-MLP. To further illustrate some of the differences between Cox-MLP and Cox-Time, we have in Figure C.3 plotted nine survival curves Ŝ(t |x) from the test set. The figure shows both the true curves and the curves estimated by the two methods. It is seen that Cox-Time is able to estimate the true survival curves better than those produced by Cox-MLP. C.3. Simulation Details In the following, we explain in detail how we generated our simulated data sets in Section 5 and Appendices C.1 and C.2. We want to generate survival times T∗ from relative risk models of the form h(t |x) = h0(t) exp[g(t, x)]. (C.2) 25 Kvamme, Borgan, and Scheel 0.00 0.25 0.50 0.75 1.00 S( t) 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 S( t) 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0 10 20 30 Time 0.00 0.25 0.50 0.75 1.00 S( t) 0 10 20 30 Time 0.00 0.25 0.50 0.75 1.00 0 10 20 30 Time 0.00 0.25 0.50 0.75 1.00 True Cox-MLP Cox-Time Figure C.3: Examples of survival curves Ŝ(t |x) from the test set in the non-proportional simulation study. If the function g does not depend on t, we have a proportional hazards model, which is just a special case of the relative risk models. Let H(t |x) denote the continuous (increasing) cumulative hazard H(t |x) = ∫ t 0 h(s |x) ds, and let V be exponentially distributed with parameter 1. Then we can obtain survival times through the inverse cumulative hazard T∗ = H−1(V |x). This can be shown to hold through the hazard’s relationship to the survival function in (1), S(t |x) = P(T∗ > t |x) = P(H−1(V |x) > t) = P(V > H(t |x)) = exp[−H(t |x)], as V is exponentially distributed with P(V > v) = exp(−v). Hence, we can obtain survival times by transforming generated samples from an exponential distribution. To obtain an analytical expression for the inverse cumulative hazard H−1(v |x), we restrict the models in (C.2) to have constant baseline h0, in addition to a simple time dependence in g(t, x). For the linear predictor in Section 5 and the non-linear predictor in Appendix C.1 we simulate with proportional hazards, meaning g(t, x) = g(x). Hence, we get cumulative hazards and inverse cumulative hazards of the form H(t |x) = th0 exp[g(x)], H−1(v |x) = v h0 exp[g(x)] . 26 Time-to-Event Prediction with Neural Networks and Cox Regression In the non-proportional simulations in Appendix C.2, we add a time dependent term g(t, x) = a(x) + b(x) t, which gives us H(t |x) = h0 exp[a(x)] (exp[b(x) t] − 1) b(x) , H−1(v |x) = 1 b(x) log ( 1 + v b(x) h0 exp[a(x)] ) . Appendix D. DeepHit DeepHit by Lee et al. (2018) is a discrete survival model for competing risks. However, as we only consider one type of event, we will express the method in terms of a single cause. DeepHit considers time to be discrete, so to fit it to the continuous-time data sets in Section 6, we discretize the event times with an equidistant grid between the smallest and largest duration in the training set. The number of discrete time-points is considered a hyperparameter, given by “Num. durations” in Table A.1. Now, assume time is discrete with 0 = τ0 < τ1 < τ2 < · · · < τm. Let y(x) = [y0(x),y1(x), . . . ,ym(x)] T denote the output of a neural net with covariates x and a soft- max output activation. Then, y(x) can be interpreted as the estimated probability mass function of the duration times yj(xi) = P̂(T ∗ i = τj |xi). The estimated survival function is then given by Ŝ(τj |x) = 1 − j∑ k=1 yk(x), and the discrete negative log-likelihood corresponding to the continuous version in (2), is lossL = − N∑ i=1 [ Di log(yei (xi)) + (1 −Di) log(Ŝ[Ti |xi]) ] . Here, ei denotes the index of the event time for observation i, i.e., Ti = τei . Furthermore, DeepHit adds a second loss that attempts to improve its ranking capabilities, lossrank = ∑ i,j Di 1{Ti < Tj}exp ( Ŝ(Ti |xi) − Ŝ(Ti |xj) σ ) . (D.1) The loss of DeepHit is a combination of these two losses, where the ranking loss is scaled by a constant. We deviate slightly from the original implementation here and instead use a convex combination of the two, loss = α lossL + (1 −α) lossrank, (D.2) where α and σ from (D.1) are considered hyperparameters. 27 Kvamme, Borgan, and Scheel References Laura Antolini, Patrizia Boracchi, and Elia Biganzoli. A time-dependent discrimination index for survival data. Statistics in Medicine, 24(24):3927–3944, 2005. David R. Cox. Regression models and life-tables. Journal of the Royal Statistical Society. Series B (Methodological), 34(2):187–220, 1972. Cameron Davidson-Pilon, Jonas Kalderstam, Ben Kuhn, Andrew Fiore-Gartland, Luis Moneda, Paul Zivich, Alex Parij, Kyle Stark, Steven Anton, Lilian Besson, et al. Cam- davidsonpilon/lifelines: v0.14.1, 2018. Lore Dirick, Gerda Claeskens, and Bart Baesens. Time to default in credit scoring using survival analysis: a benchmark study. Journal of the Operational Research Society, 68 (6):652–665, 2017. David Faraggi and Richard Simon. A neural network model for survival data. Statistics in Medicine, 14(1):73–82, 1995. Stephane Fotso. Deep neural networks for survival analysis based on a multi-task framework. arXiv preprints arXiv:1801.05512, 2018. Thomas A. Gerds, Michael W. Kattan, Martin Schumacher, and Changhong Yu. Estimat- ing a time-dependent concordance index for survival prediction models with covariate dependent censoring. Statistics in Medicine, 32(13):2173–2184, 2012. Larry Goldstein and Bryan Langholz. Asymptotic theory for nested case-control sampling in the Cox regression model. Annals of Statistics, 20(4):1903–1928, 1992. Erika Graf, Claudia Schmoor, Willi Sauerbrei, and Martin Schumacher. Assessment and comparison of prognostic classification schemes for survival data. Statistics in Medicine, 18(17-18):2529–2545, 1999. Cheng Guo and Felix Berkhahn. Entity embeddings of categorical variables. arXiv preprint arXiv:1604.06737, 2016. Frank E. Harrell Jr, Robert M. Califf, David B. Pryor, Kerry L. Lee, and Robert A. Rosati. Evaluating the yield of medical tests. Journal of the American Medical Association, 247 (18):2543–2546, 1982. Patrick J. Heagerty and Yingye Zheng. Survival model predictive accuracy and ROC curves. Biometrics, 61(1):92–105, 2005. Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: clos- ing the generalization gap in large batch training of neural networks. arXiv preprints arXiv:1609.04836, 2017. Hemant Ishwaran, Udaya B. Kogalur, Eugene H. Blackstone, and Michael S. Lauer. Random survival forests. Annals of Applied Statistics, 2(3):841–860, 2008. 28 Time-to-Event Prediction with Neural Networks and Cox Regression Jared L. Katzman, Uri Shaham, Alexander Cloninger, Jonathan Bates, Tingting Jiang, and Yuval Kluger. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology, 18(1), 2018. Nitish S. Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping T. P. Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016. John P. Klein and Melvin L. Moeschberger. Survival Analysis: Techniques for Censored and Truncated Data. Springer, New York, 2. edition, 2003. Bryan Langholz and Larry Goldstein. Risk set sampling in epidemiologic cohort studies. Statistical Science, 11(1):35–53, 1996. Changhee Lee, William R. Zame, Jinsung Yoon, and Mihaela van der Schaar. Deephit: A deep learning approach to survival analysis with competing risks. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. Margaux Luck, Tristan Sylvain, Hélöıse Cardinal, Andrea Lodi, and Yoshua Ben- gio. Deep learning for patient-specific kidney graft survival analysis. arXiv preprint arXiv:1705.10245, 2017. Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic dif- ferentiation in PyTorch. In NIPS Autodiff Workshop, 2017. Daniel J. Sargent. Comparison of artificial neural networks with other statistical approaches. Cancer, 91(8):1636–1642, 2001. L. N. Smith. Cyclical learning rates for training neural networks. In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 464–472, 2017. Gian A. Susto, Andrea Schirru, Simone Pampuri, Seán McLoone, and Alessandro Beghi. Machine learning for predictive maintenance: A multiple classifier approach. IEEE Trans- actions on Industrial Informatics, 11(3):812–820, 2015. Terry M. Therneau. A package for survival analysis in S. Version 2.38, 2015. Duncan C. Thomas. Addendum to: Methods of cohort analysis: appraisal by application to asbestos mining, by F. D. K. Liddell, J, C. McDonald and D. C. Thomas. Journal of the Royal Statistical Society: Series A (General), 140(4):469–491, 1977. Dirk Van den Poel and Bart Lariviere. Customer attrition analysis for financial services using proportional hazard models. European Journal of Operational Research, 157(1): 196–217, 2004. 29 Kvamme, Borgan, and Scheel Antonio Vigan, Marlene Dorgan, Jeanette Buckingham, Eduardo Bruera, and Mari E. Suarez-Almazor. Survival prediction in terminal cancer patients: a systematic review of the medical literature. Palliative Medicine, 14(5):363–374, 2000. Anny Xiang, Pablo Lapuerta, Alex Ryutov, Jonathan Buckley, and Stanley Azen. Com- parison of the performance of neural network methods and Cox regression for censored survival data. Computational Statistics & Data Analysis, 34:243–257, 2000. Safoora Yousefi, Fatemeh Amrollahi, Mohamed Amgad, Chengliang Dong, Joshua E. Lewis, Congzheng Song, David A. Gutman, Sameer H. Halani, Jose E. V. Vega, Daniel J. Brat, et al. Predicting clinical outcomes from large scale cancer genomic profiles with deep survival models. Scientific Reports, 7(1):11707, 2017. Xinliang Zhu, Jiawen Yao, and Junzhou Huang. Deep convolutional neural network for survival analysis with pathological images. In 2016 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 544–547, 2016. Xinliang Zhu, Jiawen Yao, Feiyun Zhu, and Junzhou Huang. WSISA: Making survival pre- diction from whole slide histopathological images. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6855–6863, July 2017. 30 III 103 IV 133 Preface List of Papers Contents Introduction Neural Networks and Statistical Models Neural Networks A Brief Historical Perspective The Multilayer Perceptron Training Networks Overfitting and Regularization Convolutional Networks Time-to-Event Prediction Event-Time Modeling Regression Models Evaluation of Survival Estimates Summary of Papers Paper I Paper II Paper III Paper VI Discussion The Survival Methods Can We Trust Individual Predictions? Extensions of the Methods Evaluation of Survival Estimates Time-to-Event Prediction and Machine Learning Bibliography Papers 
work_t6sksff4qjacvcnrfjtxa7ai2e	----	WSRC-MS-90-246 AN EXPERT SAMPLE ANALYSIS PLANNER (U) WSRC-MS--90-246 by DEgl 005124 W. A. SpencerandW. S.Parks Westinghouse Savannah River Company Savmmah River Site Aiken, SC 29808 A paper proposed for presentation at Department of Energy Complex Conference "AI irt the DOE Complex" Idaho National Engineering Laboratory Idaho Falls, ID October 9-11, 1990 and for publication in the proceedings of the meeting !..itI',! / iggI:i DISCLAIMER This report was prepared as an account of work sponsored b? an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsi- bility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Refer- ence herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recom- mendation, or favoring by the United States Government or an_ agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. This article was prepared in connection with work done under Contract No. DE-AC09-89SR18035 with the U. S. Department of Energy. By acceptance of this article, the publisher and/or recipient acknowledges the U. S. Government's right to retain a nonexclusive, royalty-free license in and to any copyright covering this article, along with the right to reproduce and to authorize others to reproduce all or part of the copyrighted article. DISTRIBUTION OF THllS _.X_,,-',UMIE:A!'F_SUNI.IMITEID I WSRC.MS-90.246 AN EXPERT SAMPLE ANALYSES PLANNER (U) by W. A. Spencer and W. S. Parks Westinghouse Savannah River Company Savannah River Site Aiken, SC 29808 AI_TRACT Analytical chemists are faced with the problem of choosing an appropriate analytical technique for a particular sample and weiglfing the options as they affect precision, time, and cost. This paper describes a computer technique to assist managers in reviewing the alternatives and to match needs with the resources available. This paper proposes an expert system, knowledgeable of analytical chemistry techniques, to create sample plans. Sample planning is an appropriate topic for expert systems because scarce human expertise is required to complete sample plans. A sample plan is the description of how samples received at the Savannah River Laboratory are handled, controlled, megua_'ed, and dispositioned. Sample planning is difficult because multiple experts are needed, planning is not a static function, and planning is time consuming. An Expert Sample Analyses Planner (XSAP) is proposed to create sample plans for laboratory managers. XSAP supplements the scarce knowledge of analytical techniques creating sample plans based on analysis constraints, methods available, and time requirements. XSAP interacts with the chemist to suggest sample plans. XSAP considers equipment available locally, at other Savannah River laboratories, at other Department of Energy facilities, and at other commercial laboratories. XSAP allows options on scheduling: best solution, cheapest solution, best local solution, and fastest solution. i INTRODUCTION The task of The Analytical SerJices Group (ASG) personnel is to develop consistent plans to assure quality-analyses/cost-savings to the customer. Assurance of quality analyses requires attention to analysis scheduling and methods. Assurance of cost-savings depends on the development of unambiguous sample plans for the technicians of ASG. A sample plan must be established before any analysis can be conducted by ASG. The sample plan is the description of how samples received at the ASG laboratory are handled, controlled, measured, and dispositioned. The development of sample plans is not a static function; periodic reviews are conducted for established customers because analytical methods change. ASG assigns a chemist or "planner" to create sample plans. The planner is familiar with the routine analytical methods and instruments offered by ASG. And the planner is familiar with offsite laboratories to supplement ASG services. The planner programs the ASG Laboratory Information Management System (LIMS) to support the customer. The planner sets default instructions for ASG to safely handle samples and notes instructions for assigned methods. Samples logged into the LIMS data base become permanent records to identify the sample, determine the hazards, perform the analyses, and dispose of the sample. The planner must know the Customer, the Study, the Material, the Profile, and the Parameters to create the LIMS record. The info_mation required for LIMS: CUS2DMER (e.g., Actinide Tech, DWPT TNX, REACTORS, M AREA, etc.) STUDY (e.g., Pu Scrap, IDMS, J. Bibler) MATERIAL (Glass, Sludge, Pu Scrap, etc.) PROFILE (lists of analytical methods, e.g., SEM]XRD) PARAMETERS ASG (Receiver Initials) ......................... used to activate ASG Description ................................................ brief description Radioactivity ........................................ how much and what Fissionable ............................. no or yes, what and how much Chemical Hazards ............................... warn ASG personnel Submitter ............................................. who gets the results Disposal ..................................... return or dispose of sample Sample Size ........................................ helps identify sample Heterogeneous ........................ if multiphase give instruction Requester .................................... technician or other contact Analyses ....... specific instructions (e.g., IC: chloride, nitrate) Comments ...................................... special message to ASG Planners are expert in one or two areas of analysis but cannot be expert in all analy_es. The planner relies on existing plans and other analytical experts to create customer sample plans. Relying or_existing plans requires that the plans be accurate. Relying on other experts requires the customer to wait until experts are available. Overlap of analysis techniques requires multiple experts to convene to develop a sample plan. New computing techniques are required to assist sample planners _4. The Environmental Protection Agency (EPA) is studying new computing strategies to assist ls boratory managers 8. Knowledge base techniques to automate portions of the analysis process have been created3,4,1°,_l,l_,_8,195°. These computing techniques are known as expert systems. The dynamic nature of analysis, the growing number of work orders, and the lack of comprehensive planning expertise dictates that ASG consider other methods for accurate sample planning. The EPA has developed expert systems in the areas of quality assurance, sampling techniques, selection of laboratories, selection of methods, review of data, diagnosis of sampling techniques, and the evaluation of laboratory analyses. DISCUSSION The ASG requires an Expert Sample Analyses Planner (XSAP). The best solution method for this problem is an intelligent reasoning system. This type of system is referred to as an expert system. Problems appropriate for this solution can be identified using seven criteria15: 1) The need for the solution justifies the cost and effort of building an expert system, 2) Human expertise is not available in all situations where it is needed, 3) The problem may be solved using symbolic reasoning techniques, 4) The problem domain is well structured, 5) The problem may not be solved using traditional computing methods, 6) Cooperative and articulate experts exist, 7) The problem is of sufficient size and scope. The need for the solution justifies the cost and effort of building an expert system 24. The cost savings is in excess of the development costs. A feasibility study for XSAP shows that the system will take 2-4 man-years to develop, and should pay for itself in 1-3 years. Human expertise is not available in all situations when needed. ASG requires a tool that will allow the customers access to expert advise when the expert is not available. The customer can select sample plans that best suit their specific requirements25, 26. The problem may be solved using symbolic reasoning techniques. This problem is solved algorithmically using tables, empiricals, and "rules-of-thumb". User queries may be required, but real-time device inputs are not required 26. The problem domain is well structured. Sample plans are derived from "first-principles" or rules from experts. These decisions can be represented in logical sequences, tables of values, or heuristic measures. Cooperative and articulate experts exist. We have the planners that are currently performing the function and experts at other DOE laboratories. These individuals will be used for the acquisition of information to program this system. This system will not displace experts - but will be a tool to enhance their productivity. Because most sample plans are taken from existing plans or created from empiricals, the size of the XSAP problem is well constrained. This program is proposed for the purpose of sample plan production and table displays. Constraints The planner may constrain the sample plan and XSAP must allow the planner to override the suggested plan. Planning with XSAP is interactive and iterative. The selections by XSAP cannot bearbitrary, so XSAP must be able to describe its rationale for plan development. XSAP describes criterion for plan development, and displays tables, charts, graphs and other empiricals available to the planner for inspection. XSAP includes detailed descriptions for the reasoning process. XSAP optimizes solutions on best, best local, cheapest, and fastest. XSAP weights sample plans based on cost, time, and quality. These constraints require XSAP to optimize resource allocation. Equipment availability/proximity/cost are factors. XSAP considers equipment available at Savannah River Laboratory, at other SRS labs, at other DOE facilities, and at other commercial labs. LIMS users/planners can access XSAP for online analytical reference. Charts, graphs, tables, or other depictions of empiricals are available to LI1VS users and planners. The r:_r interface is a critical design componentS, _°. Presenting data in a format that is familiar to the user is important for the success of the expert system. The interface must be simple so that users will not be required to learn extensive commands and routines, and the data/display must be presented in a manner that is meaningful. Access to XSAP is partitioned. Initially customers will only have access to features of the online documentation. Customer access to all the features of XSAP will be a future endeavor and must be accounted for in the design. LIMS compatibility is maintained. XSAP will be developed on the VAX platform so that it can be distributed over the local area network. Other platforms may be considered a_ technology and markets develop for XSAI_,I$4. S_pe i XSAP architecture will have four functional parts: 1) the knowledge base or rules; 2) the user interface; 3) the ORACLE data base; 4) the tables and analytical data. i XSAP Architecture QQO ? - I__ Knowledge Base 1 Proto_,t_ H_lr, o_ The hardware for the prototype is a MICROVAX II/GPX. The prototype is developed on a stand-alone system running VMS version 5.3. VAX/LISP version 3.0 is the executive controlling the inference engine and frame structure in KEE. KEE (Knowledge Engineering Environment) version 3.1, from Intellicorp, maintains the inference engine and the object system. The LIMS data base is written in ORACLE version 5.0. Knowledge base The knowledge base (KB) is the executive that directs program function 2o. The KB contains the rules to combine empirical evidence from the analytical tables, customer information from the ORACLE data base, and user input from the interface to create sample plans. The KB is divided into control and data/knowledge. Knowledge can be i represented as procedural, qualitative, and semantic 1_. Procedural knowledge is maintained in the KEE TELL/ASK facility. TELL/ASK is the inference mechanism for the system. Qualitative/Semantic knowledge is maintained in the KEE frames. The KB control is modeled in the ACTOR paradigm1, 6,15. Each of the functions of the XSAP system is provided by an ACTOR in KEE. The ACTORS are members of the ACTOR knowledge base. Each ACTOR is modeled as a KEE object called a unit. The behavior of the ACTOR is modeled as methods for the KEE objects. The ACTOR methods are written in LISP. The functional description of the code has produced the following ACTORS: Interface Manager, Resource Manager, ORACLE Manager, Table Manager, Knowledge Manager, Executive Manager. The data/knowledge is maintained in KEE. Static and dynamic entities are modeled in the system. The active entity in the system is called a "request". Each request is modeled as a dynamic object in KEE. Each request has behaviors which are modeled as methods in the dynamic unit. Static knowledge is modeled in KEE objects. Static information includes methods/machines available, and customer information. IIIII IIII I IIII II I i II ..... Control Rules Data , _, ACTORS TELL/ASK UNITS I IIIH KEE KNOWLEDGEBASE IIII I • Interface The user interface is based on interactive graphics. The interface is iconic and mouseable. The user will direct sample plan constructionwith mouse and keyboard instructions. The interface is the most visible part of _he software, and it must be robust and easy to use6,1s,24. The interS'ace is controlled by the Interface Manager. The Interface Manager is an ACTOR in the KEE system. The behavior of the interface is controlled by the ACTOR methods. The methods are written iri LISP code and DECWindows graphics. The Interface Manager has the responsibility for user control of XSAP. The Interface Manager reads mouse clicks that dictate program execution. The interface is menu/mouse driven, but may require user data input. The Interface Manager must read data/character input from the keyboard. The Interface Manager assimilates the direction from the user and displays information to the terminal. ORACLE Data Base The ORACLE data base is ACCESS*LIMS. ACCESS*LIMS is proprietary software developed by PE NELSON. The LIMS system resides on a Digital Equipment Corporation VAX 3800 running VMS 21. Access to LIMS is distributed over local area nodes and access to data is controlled. Access to LIMS is controlled by the VAX System Manager and the ORACLE Data Base Administrator 22. Figure2-1. TypicalACCESS*LIMShardwareconfiguration. ,i ACCESS*LIMS is an object-oriented :implementation. A template of the required information is created at installation. The planner fills in the template to create an instance. Each instance is called a submission. XSAP will assist the planner in filling in the template object by supplying the required information. IEMPLATE INSTANf:_._E.E J i -,L I 1 , Fromlemplaleto Instance. Analytical tables The analytical tables must be developed. These tables contain empirical data. No protocols exist for compilation of these tables. AMES laboratory will assist in the creation of these tables 4. Process The first step in creation of XSAP is completion of Knowledge Acquisition (KA) 15. Knowledge systems differ from traditional coding regimes in the type of information modeled. KA is the procedure used in the development of knowledge based systems that generates the actual functions of the code. We have stated the general requirement of the system: to be an automated assistant to produce sample plans for the ASG "planner'25. The KA function explicitly ferrets out the information that is required by the system. Once planning information has been identified it can then be codified. A prototype of the system will be created after KA. The prototype is used for proof-of-correctness._, 17. This assures both the veracity of the software and the logic processing. Experts will be consulted for the procedures that must be emulated in XSAP. Each planner has an individual planning process; differences will be encountered in the procedures used by the various experts. The prototype is used to identify/rectify the differences before implementation of the system. The prototype is primitive and is intended for incremental development. Software/hardware is designated for the project. Software/hardware is selected to assure compatibility with the existing LIMS system 6. XSAP will be developed on a Microvax II/GPX 21, manufactured by Digital Equipment Corporation (DEC). The development platform will be a stand-alone system. Once development, testing and configuration are complete, the code will be ported to the VAX running the PE NELSON LIMS system. VAX/LISP implements the software executive _1. VAX LISP is the DEC implementation of Common LISP (the ANSI standard) and is compatible with any ANSI standarc LISP implementation 23. The XSAP executive will be written in LISP and connect to KEE (Knowledge Engineering Environment) from Intellicorp. KEE and VAX/LISP are compatible with the ORACLE software. VAX/LISP can interface to ORACLE with a programmed interface 2. VAX/LISP interfaces with existing software. KEE implements the knowledge base. KEE is a powerfultool for developing large expert systems2_. Three major components of KEE are utilized: the inference engine, the truth maintenance system, and the object system. These components comprise most of the software required by an expert system. i CONCLUSION The ASG is implementing XSAP to create sample plans. XSAP is an iategrated system that interfaces with graphics, ORACLE, and analytical tables. XSAP will take 2 to 4 man-years to develop. The system will provide a considerable cost/benefit savings to ASG. Knowledge acquisition for XSAP is in progress and a prototype is in development. The prototype will be completed in the first quarter of FY91. ACKNOWI_DGMENT The information contained in this article was developed during the course of work done under Contract No. DE-AC09-89SR18035 with the U. S. Department of Energy. R_'ERENCES 1. G.A. Agha. ACTORS: A Model of Concurrent Computation in Distributed Systems. MIT Press, Cambridge MA, (1986). 2. ICP. Berkbigler, G.J. Papcun, N.L. Marusak, J.E. Hutson. _Applying Expertise to Data in the Geologist's Expert System ". Proceedings of the. Conference on Artificial Intelligence in the Department of Energy Complex, '89, Los Alamos, NM, p5-95. 3. T2. Bridge and M.H. Williams. _Expert Systems Design in High-performance Liquid Chromatography". Analytical Proceedings, February 1988, Vol.25, p43-44. 4. S.D. Brown, T,Q. Barker, R_I. Larivee, S.L. Monfre, H.R. Wilk. _Chemometrics". Analytical Chemistry, 60, p252 (1988). 5. W.G. De Ruig, G. Dijkstra, R.W. Stephany. "Chemometric Criteria for Assessing the Certainty of Qualitative Analytical Methods _. Analytica Chimica Acta, 223, p277-282 (1989). 6. G.A. Finn. "Rules of Thumb for Implementing Expert Systems in Engineering _. InTech, Vol. 4, No. 4 (1987). 7. I. Grief, C. Hewitt. "Actor Semantics of PLANNER-73". _ Conference Record of the Second ACM Symposium on Principles of Programming Languages, Palo Alto CA, p67-86, (1975). 8. J.M. Hushon. _Expert Systems in the Environmental Lab _. Environmental Lab, Feb/Mar, pl6, (1990). 9. T.A.H.M. Janse, G. Kateman. _Enhancement of Performance of Analytical Laboratories". Analytica Chimica Aeta, 150, p219 (1983). 10. S.R. Jordan, L.E. Hopper, R.C. Driver, L.F. Arrowood. "Bringing Knowledge-based Systems to a Large Computing Organization _. Proceedings of the Conference on Artificial Intelligence in the Department of Energy Complex, Alken SC, (1988). 11. J. K]aessens, G. Kateman, _Problem Solving by Expert Systems in Analytical Chemistry". Atmlyticsche Chemie, Vol. 326, No. 3 (1987). 12. tCJ. Lindsay. '_Expert Systems in the CIM Environment _. Manufacturing Technology International, (1987). 13. C. Liteanu, I.Rica. _LTtilization of the Amount of Information in Evaluation of Analytical Methods". Analytical Chemistry, Vol. 51, No. 12 (1979). 14. C. Loehle, R.E. Osteen. "IMPACT: An Expert System for Environmental Impact Assessment", Proceedings of the Conference on Artificial Intelligence in the Department of Energy Complex,, Aiken SC, (1988). 15. G.F. Luger, W.A. Stubblefield. Artificial Intelligence and the Design of Expert Systems. The Benjamin/Cummings Publishing Co., Redwood City CA, (1989). 16. M.J. Manthey. "An Operator Semantics for the Actor Model". University of New Mexico Technical Report, CSS85-9, (1985). 17. J. Melaragno, M.IL Allen. "A Plan for the Application of Artificial Intelligence to DoD Logistics _. Report PL816R1, (1989). 18. F.A. Settle. _he Application of Expert Systems in the General Chemistry Laboratory". Journal of Chemical Education, Vol. 64, No. 4 (1987). 19. FA. Settle, M_ Pleva. _rhe Expert System: A teel for the Analytical Chemist". American Laboratory, 20, p64-78 (1988). 20. F_A.Settle, M. Pleva, C. Booker, H. Iams, D. McClintock, T. Moore. "An Information System for Selecting Methods of Chemical Analysis". American Laboratory, Vol. 19, No. 4, (1987). 21. A. Sherwood. _Workstations The Emerging Alternative for AF. PCA/, July/August, (1990). 22. D.M. Smith. _Work Load Model _. Prodeedings of the Conference on Artificial Intelligence in the Department of Energy Complex, Los Alamos NM, p4-17, (1989). 23. G.L. Steele. Common LISP The Language. 2hd Edition, Digital Press, Bedford MA, 1990. 24. /LP. Wade, S.R. Crouch, D. Betteridge. _Development of Microcomputer.based Expert Systems for Analytical Chemists _. Trends in Analytical Chemistry, Vol. 7, No. 10, (1988). 25. A.P. Wade, S.R. Crouch. "Intelligent Automation: Looking toward the 1990's". Spectroscopy, Vol. 3, No. 10, (1984). 26. C.M Wong, S. Lanning. "Artificial intelligence in Chemical Analysis". Energy and Technology Review, Lawrence Livermore National Laboratory, (1984). { l,,,ll : = l_fltl -_,-_ tliii ,-- z Illlilll T < ' (dj_j VS " ,.,.-,'_ ,y. Iii -= liq _" fulfill .T. ° 'li'") _ -¢j _- II II I I ,,,,,, Z _ 00Lh i ___ II I _ III _ , , II ___ II _ tPllltll r_ 7- IlUl III I 'ld_ "_ | 6 \ fUlfil I "r.C ! i • i 
work_t7so4ijv25bp7oqfwv76o4tfdy	----	MediExpert: An Expert System based on Differential Diagnosis focusing on Educational Purposes 1 MediExpert: An Expert System based on Differential Diagnosis focusing on Educational Purposes Aris Papakonstantinou1, Haridimos Kondylakis1,2,* and Emmanouil Marakakis1 1 Department of Electrical and Computer Engineering, Hellenic Mediterranean University, Heraklion, Crete, Greece 2 Institute of Computer Science, FORTH, Heraklion, Crete, Greece Abstract The early and accurate identification of a disease is important for its effective treatment. However, medical errors represent a serious problem and pose a threat to patient safety. To this direction, appropriate and continuous education of the medical personnel has been widely recognized as an important mean to reduce medical errors and increase the quality of the health system. In this paper, we present MediExpert, an expert system targeting on continuous education of health personnel, providing also guidelines to persons that either cannot easily move due to age related comorbidities, or because they are away from healthcare units, further recommending users to talk with their doctors. It is based on differential diagnosis, employs ontologies for effective classification of health related problems and intelligent algorithms to enhance continuous education. We present the various components of the system and we elaborate on the benefits gained when using it for education. Keywords: Expert Systems, Differential Diagnosis, Educational Expert Systems. Received on 13 February 2020, accepted on 24 March 2020, published on 26 March 2020 Copyright © 2020 Aris Papakonstantinou et al., licensed to EAI. This is an open access article distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/3.0/), which permits unlimited use, distribution and reproduction in any medium so long as the original work is properly cited. doi: 10.4108/eai.13-7-2018.163844 *Corresponding author. Email: kondylak@ics.forth.gr 1. Introduction Diagnosis is one of the most important tasks performed by health providers and the impact of diagnostic errors on patient safety has been highly recognized [10]. However, although quantifiable (for example in the Harvard Medical Practice Study, diagnostic errors accounted for 17% of preventable errors [6]) limited attention has been shown at improving diagnostic errors and thousands of patients die or suffer every year due to them. To this direction, expert systems could benefit both the diagnostic procedure and the education of health providers. An expert system is a computer system that emulates the decision-making ability of a human expert. The development of such a system requires the relevant knowledge to be extracted from an expert and then to be represented in knowledge base for reasoning. Based on this knowledge, it emulates the decision-making ability of a human expert and can aid human experts in decision making and in education. In this paper, we present MediExpert, an expert system for enabling decision support and education on diagnosis. The final objective of MediExpert is primarily to provide remotely, preliminary medical diagnosis in remote areas, away from healthcare units mainly in rural areas, or in cases that aged people cannot easily move, further recommending users to talk with their doctors. In addition, it can be used to enhance physicians in decision support without being limited to a specific medical specialty. To this direction, students of medical schools can use it for educational purposes. The students can practice diagnosis on several hypothetical scenarios by presenting a set of symptoms and asking them to find potential diseases associated with the specific set of symptoms and vice versa. The novelty of the system lies in the fact that it can be widely deployed independently of the specific domain, it is simple to be understood and be EAI Endorsed Transactions on e-Learning Research Article EAI Endorsed Transactions on e-Learning 10 2018 - 07 2020 | Volume 6 | Issue 19 | e3 http://creativecommons.org/licenses/by/3.0/ Aris Papakonstantinou, Haridimos Kondylakis and Emmanouil Marakakis 2 used by health experts and that it uses ontologies to annotate and link available knowledge identifying partial matches or generalizations using this ontology. The remaining of this paper is structured as follows: In Section 2, we elaborate on related work. Then, in Section 3, we present system architecture and we elaborate on the various components and algorithms used. Section 4 presents a demonstration scenario using the system and finally Section 5 concludes this paper and presents directions for further work. 2. Related work Within the years, several expert systems have been developed trying to support decision making in diagnostic procedures. APACHE III for example is a system [5] trying to predict the person's risk of dying in a hospital. This prediction is based on a comparison of the medical history of 18,000 cases, stored in the system database. It has an average of 95% predictive accuracy and it uses a score based mechanism. Although we also use a scoring mechanism, our system is mostly targeted on educational purposes and does not rely on available history of cases rather than already existing knowledge. LISA on the other hand, is a Clinical Information and Decision Support System for co-operative care in childhood acute lymphoblastic leukaemia [3]. It is primarily concerned with providing support during the patient's treatment period, where weekly decisions on drug dosing should be made. Dose adjustment rules are applied using the Guideline Modelling Language PROforma and the Recommendations are provided in clinical setting by the TALLIS PROforma approval mechanism. Our system however, does not try to implement existing clinical guidelines but only to educate and to offer help in diagnostic procedures. ISABEL is a web-based clinical decision support system providing support for paediatric diagnostic decisions [11]. Isabel uses Autonomy's natural language processing software and consists of a proprietary medical database with over 11,000 diagnoses and 4,000 drugs. It supports templates for querying and HL7 for interoperability. Although the system is really promising with nice features like the natural language processing it is not targeting on educating health personnel and as a commercial product is difficult to be modified to do so. PUFF is an expert system introduced to interpret pulmonary function test data [1]. Its reasoning is based on a backward chaining and it uses about 400 rules in knowledge base. Although promising, it is dedicated to pulmonary function tests. Therapy Edge HIV is a web-based clinical decision support system introduced in 2005 that deals with HIV treatment [2]. Its reasoning is based on temporal guidelines to assess the patient’s current state and create alternative treatment options. Therapy Edge HIV implements an API to communicate with external systems, providing information in XML format. Again, this is a system, which is limited to a specific domain of knowledge whereas MediExpert has been developed to be a generic tool. The expert system in [7] derives differential diagnosis of epilepsy in childhood. This system as our system uses a meta-rule. The rules in [7] as in MediExpert that are fired are instances of this meta-rule. The expert system in [7] uses effectively the knowledge of experts for diagnosis. It does not run on the web and it does not have any student teaching component. Docs ‘n Drugs [8] is another intelligent tutoring system for web-based and case-oriented training in medicine, however is was identified to have a poor user experience. There, the development of a training case influences the correctness of the learner's answers, whereas ICD-10 and other ontologies are also used. We believe using ontologies is to the right direction. However, user experience with the system was low, and as such, we would like to investigate a simpler, cleaner approach. The training that is provided by Docs ‘n Drugs is case oriented while MediExpert provides a more general medical training. Finally, COSMOS [4] is another web-based expert system trying to offer decision support in diagnosis to interdisciplinary experts. As such, it can be used for diagnostic purposes in health. However, due to the complexity of the temporal rules that need to be defined, the system is really difficult to be used by domain experts. In addition, it has not been exploited for educational purposes. 2. MediExpert architecture and components The MediExpert employs a three-layered architecture, shown in Figure 1. It consists of the web interface, the diagnostic sub-system and the knowledge base. In the sequel, we describe in details each one of those layers. Figure 1. MediExpert architecture 3.1 The graphical user interface EAI Endorsed Transactions on e-Learning 10 2018 - 07 2020 | Volume 6 | Issue 19 | e3 MediExpert: An Expert System based on Differential Diagnosis focusing on Educational Purposes 3 In the top layer, two modules, the expert and the student module enable user interaction. This layer has been implemented using CSS/Javascript and HTML. The expert module is implemented as a web application available by using a web-browser. This web interface enables experts to define the rules for differential diagnosis. The user interface is simple, yet powerful, enabling the uninterrupted addition of new rules. Those rules can come from medical books or existing knowledge from medical experts. In our experiments, we used rules from a standard differential diagnosis book [9] that are used for educational and instructional purpose describing the diagnostic procedure. Different books might contain similar procedure and methods for the medical diagnosis and our knowledge base is flexible enough to be extended with additional knowledge. An example table indicating the differential diagnosis of cough is shown in Table 1. Knowledge rules are formulated based on this tabular information by a medical expert with the help of a Knowledge Engineer, as it will be described in the sequel. The student module is implemented as a mobile application and it can be used either in class or at home. It is able to select randomly cases with specific symptoms, enabling students to suggest a diagnosis. Then the suggested diagnosis is compared to the diagnosis derived by our system and the results are returned to the user. In addition, the system enables students to select a disease and then to select from a list the symptoms related to this disease. The system proceeds to the tutoring process by comparing the selected symptoms to the ones recorded by the bibliography. As such, the tool can be used for training health personnel and for helping with the diagnostic process. Table 1. Table of Differential Diagnosis of Cough Condition Nature of Patient Nature of Symptoms Associa ted Sympto ms Precipitating and Aggravating Factors Ameliora ting Factors Physical Findings Diagnostic Studies Chronic or Recurrent Cough Postnasal drip May not be aware of condition Frequent throat clearing and hawking. Cough worse in morning Recumbency Chronic sinutis Vasomotor rhinitis Allergic rhinitis Nonallergic rhinitis with eosinophilia Mucoid secretions in posterior pharynx Palpation, percusssion, and transillumination of sinuses reveal sinusitis Mucosa of nose/oropharynx:cobbmest one Asthma May have family history of allergies, atopy, or asthma Recurrent cough Minimally or not productive (if productive, secretions clear and mucoid) Shortne ss of breath Exercise May be worse during seasonal allergies Bilateral wheezing Pulmonary function tests Response to isoproterenol and methaclorine Chronic bronchitis Most common cause of chronic cough in adults (especiall y smokers) May be minimally productive Often worse in morning Smoking cessation Scattered rhonchi Chronic obstructiv e pulmonar y disease Elderly patients Shortne ss of breath Lungs hyperresonant to percussion Auscultation reveals distant breath sounds, scattered rhonchi wheezed, or prolonged expiration Pulmonary function tests EAI Endorsed Transactions on e-Learning 10 2018 - 07 2020 | Volume 6 | Issue 19 | e3 Aris Papakonstantinou, Haridimos Kondylakis and Emmanouil Marakakis 4 3.2 The diagnostic sub-system We selected Prolog for the implementation of application logic, as Prolog is a programming language based in logic. When some symptoms are provided and the disease should be identified, the preconditions in the rules, available in the Knowledge Base (KB), are unified and the most similar disease/problem is returned. On the other hand, when a disease is selected, the symptoms are retrieved from the available rules. We have to note that our algorithms are intelligent enough thus they process partial matches and they return a matching percentage, enabling student to understand their success rate. Differential Diagnosis (DD) of a disease is the process of distinguishing between two or more diseases with similar symptoms the one which the patient is suffering. This differentiation is based on systemic comparison of symptoms, signs and laboratory findings. Our Expert System follows a clinical approach of DD as in Seller & Symons 2012 [9] and follows the International Classification of Diseases (ICD-10). Seller’s and Symons’ Proverb: “If you don’t think about it, you will never diagnose it.”. The reasoning in this approach is as follows. Step 1: Initially, check that the Complaint (symptom) is member of the driving complaint list. Step 2: Then, follow diagnosis by processing the derived rules based on tables as the one shown in Table 1.. The construction of the rules is based on the following conditions and characteristics. Step 2a: Nature of patient. Identifies those conditions that are most prevalent within a particular subgroup (e.g. children, the elderly and premenopausal, diabetic, hypertensive, and immunocompromised individuals). Step 2b: Nature of symptoms. Further identifies conditions by amplifying additional characteristics of the symptoms (how, when, where, radiation, acute/chronic, and others). Step 2c: Associated symptoms. Any additional complaint (e.g., headache) could contradict or ensure the diagnosis Step 2d: Precipitating and aggravating factors. For example, the pain of gastritis is worsened by the ingestion of most foods, particularly alcoholic beverages. Peptic ulcer pain usually begins an hour or so after eating and it is generally relieved by eating. If epigastric pain occurs primarily or it is worsened in the recumbent position, peptic esophagitis should be suspected. Step 2e: Ameliorating factors. For example, if the patient experiences relief after eating or taking antacids, peptic ulcer or peptic esophagitis is the probable cause of pain. The pain of gastritis, though worsened by the ingestion of food and alcoholic beverages, may be relieved by antacids. Step 2f : Physical findings. Physical examination can often provide major clues to the diagnosis. Step 2g: Diagnostic studies. For example, ultrasound, mammography, radiographs, CT Scans, magnetic resonance imaging (MRI). Step 3: Derivation of diagnosis A novel feature of our system is that the constructed KB is complemented with terms from the ICD-10† ontology, in order to ensure interoperability and a common understanding of the various health problems. ICD10 is a standard medical classification ontology, which we exploit to record and identify similarities between health problems. The ICD10 taxonomy can be represented as a tree, with health problems as its nodes. In the 2017 version of ICD10, there are four levels in the tree, in addition to the root level. Sibling nodes that belong to lower levels share greater similarity than siblings that belong to upper levels. Our diagnostic sub-system is able to perform reasoning on the various layers identifying for example that “S27.4 Injury of bronchus 2” is a subcategory of “S27 Injury of other and unspecified intrathoracic organs” applying this to the decision support process. 3.3 Meta-rule and form of rules The reasoning process is driven using a meta-rule. The next meta-rule encodes the reasoning of using the DD tables. This meta-rule is general enough to cover all cases of health problems with some initial symptoms. All domain-specific rules are instances of this meta-rule. This meta-rule is as follows where ∈, ⊆, ∩ and ≠ are set operators. If (Complaint ∈ Driving_complaint_list) and ( (Nature_of_patient ⊆ Nature_of_patients_list) or ({Nature_of_symptoms ∩ Nature_of_symptoms_list} ≠ Ø) or ({Associated_symptoms ∩ Associated_symptoms_list} ≠ Ø) or ({Precipitating_and_aggravating_factors ∩ Precipitating_and_aggravating_factors list} ≠ Ø) or ({Ameliorating factors ∩Ameliorating_factors_list} ≠ Ø) or ({Physical_finding ∩ Physical_finding_list} ≠ Ø) ) Then Diagnosis = (Condition, ICD-10_code, Diagnostic Studies) In the aforementioned rule, Condition is the diagnostic result, ICD-10_code is the ICD-10 term and the † http://www.icd10data.com/ EAI Endorsed Transactions on e-Learning 10 2018 - 07 2020 | Volume 6 | Issue 19 | e3 MediExpert: An Expert System based on Differential Diagnosis focusing on Educational Purposes 5 Diagnostic Studies are the studies to ensure the diagnostic result. An instance of the aforementioned rule, which encodes part of the knowledge shown in Figure 2, is the following: If ( (chronic_or_Recurrent_Cough ∈ {acute_Cough, chronic_or_Recurrent_Cough} and ({may_have_family_history_of_allergies, atopy} ∩ {may_have_family_history_of_allergies, atopy, asthma} ≠ Ø) or ({minimally_or_not_productive} ∩ {recurrent_cough, minimally_or_not_productive } ≠ Ø) or ({Shortness of breath} ∩ {Shortness of breath} ≠ Ø) or ({may_be_worse_during_seasonal_allergies} ∩ {exercise, may_be_worse_during_seasonal_allergies} ≠ Ø) or ({bilateral_wheezing} ∩ {bilateral_wheezing } ≠ Ø) ) Then Diagnosis = (asthma, J45, pulmonary_function_tests) For this example, the diagnostic elements are the following: • The Condition is asthma, • The ICD-10_code is J45 • The Diagnostic Studies are the pulmonary_function_tests 3.4 The Knowledge Base The sentences in the knowledge base are similar to normal text and they are easily understood by any health scientist without computer programming expertise. The rules that are stored in our knowledge base are of the following form, which directly corresponds to the table shown in Table 1: [Id, Complaint, Nature_of_patient, Nature_of_symptoms, Associated_sympotms, Precipitating_and_aggravating_factors, Ameliorating_factors, Physical_findings, Condition, ICD-10_related_problems]. Note that each rule is represented as a list. The reasoning engine of MediExpert processes this rule representation as “if-then” rule during diagnosis. We have to note that the domain experts are able to generate the aforementioned rules through a nice GUI, which are then stored internally as Prolog clauses. In addition, the constructed KB is complemented with terms from the ICD-10 in order to ensure interoperability and a common understanding of the various health problems. 4. System Demonstration Scenario and Preliminary Results In the sequel, we demonstrate the sub-system for the students, as shown in Figure 2. More specifically, we presented the system to 30 student nurses at the Technological Educational Institute of Crete, asking the class to identify symptoms about a specific complaint (e.g. “Chest Pain”). If a student successfully selects the symptoms, s/he proceeds to the next problem. A different approach is just to present the symptoms and then to ask for the proper disease. As we exploit the ICD-10 terminology we are able to exploit generalizations and specializations of the terms used, however with appropriate warnings (e.g. throat pain is also a pain and can replace it as a more generic symptom). More specifically the performed scenario is the following: Students’ scenario: In the first step of this, shown in Figure 2a, the mobile version of the MediExpert, presents to the student a list of symptoms of a specific condition. Then the program invites the student to choose from a list of possible answers the correct one as a diagnosis (shown in Figure 2b). Finally, the program reveals to the student the correct answer, as shown in Figure 2c. Preliminary results from students: All students using the system found it interesting, usable and recognized its potential to be used as an educational tool, helping them in the difficult process of linking symptoms and other factors with diagnosis. Figure 2. Selecting symptoms for Chest Pain (a), selecting diagnosis for Chest Pain (b), and checking the given answer for Chest Pain (c). Experts’ scenario: Next, we demonstrate the expert’s web interface focusing on the diagnosis process. As such, the first page asks about the complaint of the patient (shown in Figure 3). EAI Endorsed Transactions on e-Learning 10 2018 - 07 2020 | Volume 6 | Issue 19 | e3 Aris Papakonstantinou, Haridimos Kondylakis and Emmanouil Marakakis 6 Figure 3. Complaint Selection The next step presents the symptoms and factors in checkbox form and the user should check the ones that are present in the specific case, as shown in Figure 4. Figure 4. Symptoms Selection Figure 5. Potential Conditions Finally, the last step is to present the diagnostic result, as shown in Figure 5. Preliminary results from experts. We made a short demonstration to two medical practitioners in order to get an indication of the usability of the system. The practitioners also recognized its usefulness, provided us with additional books to extract rules and proposed instead of letting the doctors to write the rules, to provide a search/select functionality where by simple clicks will be able to generate those rules almost automatically. Currently we are investigating those aspects, whereas we have to note that the next version of the system with a new updated interface is about to be released in the following weeks. 5. Conclusions and Discussion This paper presents an expert system based on differential diagnosis for educational purposes. The summary of the main features of MediExpert are as follows. The system is extended easily with the addition of new differential diagnosis rules, through a nice user interface and it can be used to educate health students on their difficult task of diagnosis. The system employs ontologies to identify and classify the diseases and implements intelligent algorithms for reasoning and relating symptoms and diseases. A web interface and a mobile application support the communication of the system with its potential users. The system employs an intelligent reasoning and in case of partial matching of diagnostic rules, it derives alternative diagnoses by using ICD-10 thus exploiting the ICD-10 tree for identifying close diagnoses. In the envisioned deployment of the system, it will be available online for all, enabling both students of medical schools to test their knowledge and on individuals to get more information about the process of differential diagnosis and what could be the cause of their symptoms. We have to note that always the system will recommend to the individuals to further discuss their findings with their doctor as by no means we intend to replace their doctor. As next step, we intend to proceed on an extensive usability evaluation of the system using students from the medical university of Crete, and to implement the EAI Endorsed Transactions on e-Learning 10 2018 - 07 2020 | Volume 6 | Issue 19 | e3 MediExpert: An Expert System based on Differential Diagnosis focusing on Educational Purposes 7 collected requests for changes. In addition we intend to investigate both a more sophisticated inference engine be added to this system and also shifting from a rule based system to one that uses learning algorithm that can help the students see why they failed (gave the wrong diagnosis) and maybe provide more information. References [1] Aikins, J.S, Kunz, J.C., Shortliffe, E.H., Fallat, R.J. (2018) PUFF: An expert system for interpretation of pulmonary function data. Comp. Biom. Res., 16(3), 199-208. [2] Boulme, R., Gonzalez, D., Schmit, J.C. (2004) Storing genotypic resistance data and linking to other clinical information. XV International AIDS Conference. [3] Bury, J., Hurt, C., Bateman, C. (2002) LISA: A Clinical Information and Decision Support System for Childhood Acute Lymphoblastic Leukaemia. AMIA. [4] Giannoulis, M., Kondylakis, H., Marakakis, E.I. (2018) COSMOS: A Web-Based, Collaborative Knowledge System Using Ontologies and Managing Uncertainty. PETRA, pp. 441-448. [5] Knaus, W., Wagner, D., Draper, E., et al. (1991) The APACHE III prognostic system. Risk prediction of hospital mortality for critically ill-hospitalized adults. Chest, 100(6), 1619-1636. [6] Leape, L.L., Brennan, T.A., Nan Laird, M.P.H., et al. (1991) The Nature of Adverse Events in Hospitalized Patients - Results of the Harvard Medical Practice Study II. N Engl J Med., 324(6):377-84. [7] Marakakis, E., Vassilakis, K., Papadakis, N. (2010) Meta- rules and uncertain reasoning for diagnosis of epilepsy in childhood. Expert Systems with Applications, 37, 6979- 6985. [8] Martens, A., Bernauer, J., Illmann, T., Seitz, A. (2001) Docs 'n drugs-the virtual polyclinic: an intelligent tutoring system for web-based and case-oriented training in medicine. Proceedings AMIA, pp 433-437. [9] Seller, R.H., Symons, A.B. (2012) Differential Diagnosis of Common Complaints, 6th Edition, Saunders. [10] Singh, H., Petersen, L.A., Thomas, E.J. (2006) Understanding diagnostic errors in medicine: a lesson from aviation. Quality & Safety in Health Care, 5(3), 159-164. [11] Vardell, E., Moore, M. (2011) Isabel, a clinical decision support system. Medical Reference Services Quarterly, 30(2), 158-66. EAI Endorsed Transactions on e-Learning 10 2018 - 07 2020 | Volume 6 | Issue 19 | e3 https://dblp.uni-trier.de/pers/hd/i/Illmann:Torsten https://dblp.uni-trier.de/pers/hd/s/Seitz:Alexander 
work_t7v3a7qghfb3da5qeckzj34ji4	----	PII: 0923-4748(91)90019-N 363 References Miles, R.E., and Snow, C.C., 1978. Organizational Strategy, Structure and Process. McGraw-Hill, New York. Porter, M.E., 1985. Competitive Advantage: Creating and Sustaining Superior Performance. Free Press, New York. Porter, M.E., 1980. Competitive Strategy: Techniques for Analyzing Industries and Competitors. Free Press, New York. The Prentice-Hall Guide to Expert Systems, by Robert A. Edmunds, Prentice-Hall, Englewood Cliffs, NJ, 1988,440 pp. Mr. Edmunds has produced a broad introduction to expert systems, designed to allow traditional data processing professionals and even computer novices to make an informed judgement about whether to invest time and money ex- ploring expert systems further. Its coverage is sufficiently comprehensive that Artificial Intelligence practitioners may find nuggets of information about as- pects of the field with which they are unfamiliar. The book searches for a mid- dle ground between deep technical exposition and gee-wiz fluff. Considering the scope and difficulty of the task and the bewildering rapidity with which the field is changing, Edmunds has succeeded to a creditable extent; the book should be useful to its intended audience. Edmunds begins in Chapter 1 with some examples to whet the readers’ ap- petite, and a brief overview of the subject. Chapter 2 argues that knowledge and knowledge workers have become key components of modern corporations, and that expert systems can provide a competitive edge in this area; it also discusses some pitfalls associated with the technology. Chapters 3 through 7 provide the technical heart of the book, discussing the components of an expert system, the process of “knowledge engineering”, the fundamental knowledge representation paradigms, inference engines, and the concept of a programming environment. Chapters 8 through 11 provide some insight into the problem of buying an expert system, discussing programming languages, hardware, the development process, and finally, a case history. Chapters 12 through 14 discuss tools and their vendors, while Chapters 15 through 19 profile some examples companies and programs. Finally, Chapter 20 concludes with a discussion of some fundamental questions potential users need to address in going further. The index is useful, but perhaps a little brief in a book that may be used for reference. For so broad an undertaking, the book is reasonably well organized and tech- nically sound. The book assumes only a general familiarity with computers, but will be most valuable to those with some experience in writing or contract- ing for traditional software systems. For example, Edmunds discusses the de- velopment process by laying out the steps in a conventional software proj- 364 ect,contrasting traditional practice with expert system approaches. He repeatedly discusses the pros and cons of various approaches, including the entire field, in an objective and informative way; managers considering con- tracting for an expert system will no doubt find these sections particularly useful. Examples are used extensively, and seem fairly well chosen to provide as much technical insight as possible without excessively burdening the novice. Inevitably, the treatment is somewhat uneven. For example, the discussion of LISP includes property lists, little used in most modern programs, and omits structures (a heavily used feature), as well as the more advanced and less standardized object-oriented features of this rapidly changing language. The hardware discussion omits the Apple Macintosh, which has recently achieved considerable favor as a low-end AI workstation. The book also exhibits the difficulty of writing about so dynamic a field in less technical ways. For example, it includes charts of the growth and distri- bution of the Artificial Intelligence market for periods beginning as early as 1989 and ending in 1990, without clearly labeling the reliability of the data. The 1990 figures are of course projections; they show continued exponential growth quite different from the actual path of the industry. The discussion of Symbolics, Inc., omits completely the company’s more recent layoffs, stock slump, and struggling status. These problems contribute to a more important difficulty, inherent in the book’s intended mission. Artificial intelligence is a complex, ill-defined sub- ject. A great deal has been accomplished, but these accomplishments are partly obscured by layers of hype and reaction. Determining what can be automated remains a subtle, difficult task, as is determining an appropriate approach. The term expert system means many different things to different people; Edmunds has presented a grab bag of technologies (omitting many more ) each of which is a substantial subject in itself. The expert systems presented in the book are for the most part the old standbys, with XCON as the major example of a com- mercial success. The degree to which these systems have succeeded or failed is not made clear, let alone the reasons why. Thus, while Edmunds makes a serious effort to cut through the hype and present a balanced picture, readers should not expect firm and reliable guid- ance sufficient to allow an informed decision on a major project. They will find a great deal of useful information, presented at a simple level, about an impor- tant and useful set of capabilities. The book should at least substantially en- hance their ability to talk to professionals in the field, or to read further. For management professionals interested in the subject, it will be a very useful first source. Reviewed by Allen C. WARD Department of Mechanical Engineering and Applied Mechanics University of Michigan Ann Arbor, MI 48109, USA 
work_ta5im6f32vgy3kyfnqnenznrry	----	Texture-based descriptors for writer identification and verification Expert Systems with Applications 40 (2013) 2069–2080 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Texture-based descriptors for writer identification and verification D. Bertolini a, L.S. Oliveira a,⇑, E. Justino b, R. Sabourin c a Federal University of Parana (UFPR), R. Rua Cel. Francisco H. dos Santos, 100, Curitiba, PR 81531-990, Brazil b Pontifical Catholic University of Parana (PUCPR), R. Imaculada Conceição, 1155, Curitiba, PR 80215-901, Brazil c Ecole de Technologie Superieure, 1100 rue Notre Dame Ouest, Montreal, Quebec, Canada a r t i c l e i n f o a b s t r a c t Keywords: Writer identification Writer verification Pattern recognition Texture 0957-4174/$ - see front matter � 2012 Elsevier Ltd. A http://dx.doi.org/10.1016/j.eswa.2012.10.016 ⇑ Corresponding author. Tel./fax: +55 41 33613655 E-mail address: lesoliveira@inf.ufpr.br (L.S. Oliveir In this work, we discuss the use of texture descriptors to perform writer verification and identification. We use a classification scheme based on dissimilarity representation, which has been successfully applied to verification problems. Besides assessing two texture descriptors (local binary patterns and local phase quantization), we also address important issues related to the dissimilarity representation, such as the impact of the number of references used for verification and identification, how the frame- work performs on the problem of writer identification, and how the dissimilarity-based approach com- pares to other feature-based strategies. In order to meet these objectives, we carry out experiments on two different datasets, the Brazilian forensic letters database and the IAM database. Through a series of comprehensive experiments, we show that both LBP- and LPQ-based classifiers are able to surpass pre- vious results reported in the literature for the verification problem by about 5 percentage points. For the identification problem, the proposed approach using LPQ features is able to achieve accuracies of 96.7% and 99.2% on the BFL and IAM and databases respectively. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction In the last decade, several researchers have dedicated a consid- erable amount of effort to solving the problems of writer identifi- cation and verification. The former task concerns the retrieval of handwritten samples from a database using the handwritten sam- ple under study as a graphical query. It provides a subset of rele- vant candidate documents on which complementary analysis will be performed by the expert. The latter task must, on its own, yield a decision as to whether or not two samples of handwriting were produced by the same writer (Bensefia, Paquet, & Heutte, 2005). To cope with these problems, two different groups of ap- proaches have been proposed in the literature. The first groups consists of local approaches, as they are based on specific features of the writing, and, in general, involve a segmentation process. Very often these features are localized, extracted from characters or allographs. Some examples of this strategy can be found in Bensefia et al. (2005), Bulacu, Schomaker, and Vuurpijl (2003), Kirli and Gulmezoglu (2011), Marti, Messerli, and Bunke (2001), Siddiqi and Vincent (2010), Srihari, Cha, Arora, and Lee (2002). One of the main bottlenecks encountered in the local ap- proaches occurs at the segmentation stage. As in many other pat- tern recognition problems, when the segmentation fails, most of the subsequent tasks, such as feature extraction and classification, ll rights reserved. . a). are compromised. In order to avoid this outcome, some authors proposed a second group of approaches, referred to as global ap- proaches. In this case, the methods identify the writer of a docu- ment based on the overall look and feel of the writing. In other words, a good way of avoiding segmentation is to look at handwrit- ing as a texture. This strategy was first tried by Said, Tan, and Baker (2000), where the texture was represented by Gabor filters and the gray level co-occurrence matrix (GLCM). A similar strategy was employed by Bush, Boles, and Sridharan (2005) for script identifi- cation. More recently, Hanusiak, Oliveira, Justino, and Sabourin (2011) discussed the use of GLCM for author verification. They demonstrate in their experiments that texture-based features are a good alternative to author verification. To the best of our knowledge, the best results using the global approach have been achieved using the well-known GLCM and its descriptors, which were proposed by Haralick, Shanmugan, and Dunstein (1973) almost 40 years ago. Since then, other texture descriptors have been developed and successfully applied in vari- ous areas. Two of them, local binary patterns (LBP) and local phase quantization (LPQ), have attracted a great deal of attention because of their performance in a number of applications. The concept of the LBP was first proposed by Ojala, Pietikäinen, and Harwook (1996) as a simple approach, robust in terms of grayscale varia- tions, which proved its ability to efficiently discriminate among a wide range of rotated textures. Later, they extended their work (Ojala, Pietikäinen, & Mäenpää, 2002) to produce a grayscale and rotation invariant texture operator. The concept of LPQ was http://dx.doi.org/10.1016/j.eswa.2012.10.016 mailto:lesoliveira@inf.ufpr.br http://dx.doi.org/10.1016/j.eswa.2012.10.016 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 2070 D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 originally proposed by Ojansivu and Heikkila (2008), and has been shown to be robust in terms of blur, and to outperform LBP in tex- ture classification (Ojansivu, Rahtu, & Heikkila, 2008). With this in mind, we consider here both LBP and LPQ as texture descriptors to perform writer verification and identification. We apply the same classification scheme used in Bertolini, Oliveira, Justino, and Sabourin (2010), Hanusiak et al. (2011), Rivard, Granger, and Sabourin (2011), based on the dissimilarity represen- tation. This scheme, also known as the writer-independent (WI) approach, takes into account a dichotomy transformation which makes it possible to transform any n-class pattern recognition problem into a 2-class problem. This property allows us to design a verification/identification system, even when a limited number of samples from a large number of users is available. The underly- ing hypothesis of the dissimilarity-based approach is that the learning set is representative of the entire population of legitimate users enrolled in the verification/identification system (Rivard et al., 2011; Srihari et al., 2002). This means that it is not necessary to retrain the dissimilarity model each time an unknown writer is presented to the system. In addition to exploiting various texture descriptors, we address the following issues that the work of Hanusiak et al. (2011) has left as open questions: (i) what is the impact of the number of refer- ences and the fusion rules used for verification and identification? (ii) how does the dissimilarity-based approach perform on the problem of writer identification? (iii) how does the dissimilarity- based approach compare to the feature-based approaches? To answer these questions, we have carried out experiments on two different databases, the Brazilian forensic letter (BFL) database (Freitas, Oliveira, Sabourin, & Bortolozzi, 2008) and the IAM data- base (Marti & Bunke, 2002). First, we addressed the verification problem. Our results using LBP and LPQ surpass, by a considerable margin, those of the GLCM-based classifier introduced in Hanusiak et al. (2011) on the BFL database. In the case of the IAM database, the two texture descriptors also achieved remarkable results with error rates below 0.5%. For the identification problem, we highlight, through a series of detailed experiments, the importance of the number of references available to correctly identify a given writer using the dissimilarity framework along with the texture features. To show the efficiency of this approach, we compare it to other two feature-based classi- fication strategies. Our results show that the proposed approach is able to achieve accuracies of 94.7% on the BFL and 94.5% on the IAM using LPB features, and 99.2% on the BFL database and 96.7% on the IAM database using LPQ features. These results compare favorably with those of the state of the art. This work is structured as follows: Section 2 describes the dat- abases considered in our experiments. Section 3 describes the dis- similarity framework and how the dissimilarity feature vectors are created. Section 4 introduces the proposed texture-based feature sets and classification methods, and our experimental results and a discussion are presented in Section 5. Finally, Section 6 concludes this work and indicates some future directions for our work. 2. Databases Two databases were considered in this work, the Brazilian forensic letter (BFL) database (Freitas et al., 2008) and the IAM database (Marti & Bunke, 2002). Both are described in the follow- ing subsections. 2.1. BFL database This database is composed of 315 writers, with three samples per writer, for a total of 945 images. The samples were provided by undergraduate students in three different sessions over a one month period. The texts were collected on a sheet of white A4 pa- per with no pen-draw baseline, and then scanned in gray levels at 300 dpi (3760 � 2448 pixels). Each writer was allowed to use his/ her own pen, which means that numerous different pens were used. The text is concise (131 words in Portuguese), and complete in the sense that it contains all the characters (letters and numer- als) and certain character combinations of interest. This makes it suitable for text-dependent writer identification as well. Fig. 1a shows the sample content. In order to validate the main hypothesis of the dissimilarity- based approach, i.e. that the writers used for training are represen- tative of the entire population, we divided the database into two corpora: one composed of 200 writers for training, and the other of 115 writers for testing. Four different partitions for training were considered: 25, 50, 100, and 200 writers, the idea being to analyze the impact of the number of writers used for training on the overall performance. The remaining 115 writers were used for testing. Fig. 1 depicts a sample of the BFL database. Using the algorithm described in Section 4, we extracted nine blocks of texture (256 � 256 pixels) from each image. Since we have three samples per writer, this adds up to 8505 texture images. Nine is the maximum number of fragments that we could extract from the writers for whom we had less handwriting information: although the number of words was the same, the characters were very small. In most cases, we could have extracted more fragments, but we decided to fix the number of fragments to 9, so that all the authors would be equally represented. 2.2. IAM database The IAM dataset (Marti & Bunke, 2002) is one of the best known and widely used databases in problems such as handwriting recog- nition and writer identification. It comprises forms with handwrit- ten English text of variable content. The images have been scanned at 300 dpi, 8 bits/pixel, in gray-scale. A total of 650 writers have contributed to the dataset, with 350 writers having only one page, 300 writers with at least two pages, and 125 writers with at least four pages. Fig. 2a shows the distribution of the IAM database. This database was divided in the same proportions as the BFL database, which means that the specimens contributed by the 650 writers represented in the database were divided into training and testing. As with the BFL database, four different partitions for training were considered: 50, 100, 205, and 410 writers, again, the idea being to analyze the impact of the number of writers used for training on the overall performance. The remaining 240 writers were used for testing. Fig. 2b shows an example of the IAM database. Since some samples in this database contain only a few lines of handwriting, it was not possible to create 9 fragments of 256 � 256 pixels, which was desirable to facilitate the analysis of the experi- mental results. However, we were able to create 9 blocks of 256 � 128 pixels. 3. The dissimilarity framework In this work, we have adopted the framework used by Hanusiak et al. (2011), which is based on a dichotomy transformation (Cha & Srihari, 2002) that makes it possible to reduce any insurmountable pattern recognition problem to a 2-class problem. Writer identifi- cation is an example of such a problem. Given a queried handwrit- ten document and a reference handwritten document, the aim is to determine whether or not the two documents were produced by the same writer. Let Vi and Qi be two vectors in the feature space, labeled lV and lQ respectively. Let Zi be the dissimilarity feature Fig. 1. BFL database: (a) the contents of the Brazilian forensic letter and (b) a sample of the database. Fig. 2. IAM database (Marti & Bunke, 2002): (a) distribution and (b) sample of the database. D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 2071 Fig. 3. Dichotomy transformation: (a) samples in the feature space and (b) samples in the dissimilarity space where (+) stands for the vectors associated to the within class and (�) stands for the vectors associated to the between class. 2072 D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 vector resulting from the dichotomy transformation Zi = jVi � Qij, where j�j is the absolute value. This dissimilarity feature vector has the same dimensionality as Vi and Qi. In the dissimilarity space, there are two classes that are inde- pendent of the number of writers: the within class (+) and the be- tween class (�). The dissimilarity vector Zi is assigned the label lZ, according to Rivard et al. (2011): lZ ¼ þ if lV ¼ lQ ; � otherwise; � ð1Þ Fig. 3 illustrates this transformation. Suppose there are four writers, {x1, . . . , x4}, and each one of them provides three samples. The feature extraction process extracts a vector (X1, X2) from each sample, and these are shown in Fig. 3a. Then, a dichotomy transfor- mation takes place and computes the dissimilarity between the features of each pair of samples to form vectors (Z1, Z2). These vec- tors, which we call dissimilarity feature vectors, are shown in Fig. 3b. We can see in Fig. 3 that the dichotomy transformation affects the geometry of the distribution. In the feature space, multiple boundaries are needed to separate all the writers. In the dissimilar- ity space, by contrast, only one boundary is necessary, since the problem is reduced to a 2-class classification problem. The number of samples in the dissimilarity space is larger, because these sam- ples are made up of every pair of feature vectors. We can also see in Fig. 3 that, if both samples come from the same writer (genuine), then all the components of such a vector should be close to 0, otherwise they come from different writers (a forgery), in which case the components should be far from 0. This is true under favor- able conditions. However, as in any other feature representation, the dissimilarity feature vector can be affected by intra-writer var- iability. This variability could generate values that are far from zero, even when the dissimilarity between the samples produced by the same writer is measured. As mentioned earlier, one advantage of this approach is that even writers whose specimens were not used for training can be identified by the system. This characteristic is quite attractive, since it obviates the need to train a new model every time a new Fig. 4. The dissimilarity framework used for writer i writer is introduced. In our experiments, we emphasize this fea- ture by using disjoint sets of writers for training and testing. The framework underpinning this work is depicted in Fig. 4. Ini- tially, a handwritten document is converted to a texture image. Then, the texture is split into n equal parts, Ri (i = 1, 2, . . . , n), which are sent to the feature extraction module. The resulting feature vectors, Vi, are stored in a database. The actual feature extraction process is discussed in Section 4. When a queried handwritten doc- ument is presented to the system, it is also converted to a texture and split into m equal parts, Si(i = 1, 2, . . . , m). These m textures un- dergo the same feature extraction process, and so creating the fea- ture vectors Qi. Then, the dissimilarity feature vectors Zi = jVi � Qij are computed and sent to the SVM classifier, which yields a deci- sion on each dissimilarity feature vector. The final decision, D, is based on combining these partial decisions, and is obtained by means of a fusion rule. Section 5 discusses how this combination is achieved. 3.1. Dissimilarity feature vectors The dissimilarity framework requires the classifiers to discrim- inate between genuine (positive) samples and forgeries (negative). To generate the positive samples, we computed the dissimilarity vectors among R genuine samples (references) of each writer (one segment of texture extracted from each letter), which resulted in R 2 � � different combinations. The same number of negative sam- ples is generated by computing the dissimilarity between one ref- erence of one author against one reference of other authors picked at random. Considering, for example, 50 writers and three texture seg- ments (R = 3) for training, we would have 150 (3 � 50) positive samples and 150 (3 � 50) negative samples. Fig. 5 exemplifies this process. In Fig. 5a, Va, Vb, and Vc are the reference feature vectors ex- tracted from the reference images (e.g. texture segments) for a gi- ven writer. Based on these three vectors, three dissimilarity vectors (Z1, Z2, and Z3) are computed. These are positive (genuine) dissim- ilarity vectors, which are expected to have components close to 0. A similar process is depicted in Fig. 5b to create the negative (forg- ery) dissimilarity vectors. In this case, the reference feature vectors are compared with the feature vectors of other authors picked at random, and it is expected that they will have components that are far from 0. 4. Building textures and extracting features The texture blocks were built using the algorithm described in Hanusiak et al. (2011). To make this work self-contained, we de- scribe the main steps of such an algorithm below. First, the image is binarized using the Otsu algorithm and then scanned, top-down and left–right, to detect all the connected components of the image. The 8-adjacency was considered in this work. Small components, such as periods, commas, strokes, and noise, are discarded at this time. The bounding box of the dentification/verification (Hanusiak et al., 2011). Fig. 5. Dissimilarities (a) among genuine samples of the same writer to generate the positive samples and (b) among genuine samples from different writers to generate the negative samples. Fig. 6. The texture generation process (a) filling a line and (b) spaced texture. D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 2073 remaining components is then used to extract the original compo- nents of the gray level image. The components in gray levels are then aligned with the new image using the center of mass of the bounding box. This process is depicted in Fig. 6a. After filling the first line we compute the average height of all the connected components used in such a process. This value is used to define the y-coordinate of the next line, which is given by new y ¼ previous y þ h 2 ð2Þ where the previous_y is the y-coordinate used to fill the previous line (in the case of the first line, a constant, k = 150, was used for both databases, BFL and IAM) and h is the average height of all the connected components used to fill the previous line. Reducing the gap between the lines by dividing h by two allows us to build more representative textures, otherwise the texture will contain too many blank spots, as in Fig. 6b. This denominator was found empirically. Figs. 7 and 8 show an example of the texture created from the original gray level letter for BFL and IAM respectively. The final texture image representing the writer’s handwriting is fi- nally segmented into blocks. In Section 5, we discuss the impact of using different block sizes. This segmentation scheme differs from those presented in the literature (Bush et al., 2005; Said et al., 2000), in the sense that no preprocessing step, such as slant correction, is needed for line segmentation. Besides making the segmentation simpler, the pro- posed texture generation method also keeps some features, such as skew and slant. 4.1. Local binary patterns Ojala et al. (2002) present a model to describe texture, called lo- cal binary patterns (LBP). In this model, each pixel C contains a set of neighbors P, equally spaced at a distance of R and C. A histogram h is defined by the texture intensity differences be- tween C and its neighbors, P. When the neighbors do not corre- spond to an image pixel integer value, that value is obtained by interpolation. An important characteristic of this descriptor is its invariance to changes in the value of the average intensity of the central pixels, when comparing it with its neighbors. Considering the resulting sign of the difference between C and each neighbor P, by definition, we assign a result of 1 to a positive sign, and 0 otherwise. This makes it possible to obtain the invari- ance of the intensity value of pixels in gray scale format. With this information, the LBP value can be obtained by multiplying the bin- ary elements for a binomial coefficient. So, a value 0 6 C0 6 2P is generated, which is the value of the feature vector. Observing the non uniformity of the vector obtained, Ojala et al. (2002) introduced a concept based on the transition between 0 s and 1 s in the LBP image. They explained that a binary LBP code is considered uniform if the number of transitions is less than or equal to 2, also considering that the code is seen as a circular list. That is, the code 00100100 is not considered uniform, because it contains four transitions, while the code 00100000, because it only has two transitions, is characterized as uniform. Fig. 9 illustrates this idea. So, instead of using the whole histogram, the size of which is 2P, we can use only the uniform values, which constitute a smaller fea- ture vector. This version of the descriptor is called ‘‘u2’’, a label that accompanies the values of the radius R and the neighborhood size P, and so the definition of LBP is as follows: LBPlabelP;R . A version of the descriptor rotation invariant is also defined, called ‘‘riu2’. But, in this work, the best results were provided by the uniform descriptor ‘‘u’’. Furthermore, we observed during the experiments that the feature extraction with LBPu28;2 is fast and accurate enough for the proposed application. Then, we set P = 8 and R = 2 for the experiments described in this paper. This pro- duces a feature vector of 59 components, which was normalized using the min–max rule. 4.2. Local phase quantization The local phase quantization (LPQ) (Ojansivu & Heikkila, 2008) is based on the blur invariance property of the Fourier Fig. 7. Original letter and texture blocks – sample from the BFL database. Fig. 8. Original letter and texture blocks – sample from the IAM database. 2074 D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 phase spectrum. The local phase information of an N � N image f(x) is extracted by the 2D DFT (short-term Fourier transform (STFT)) f̂ uiðxÞ¼ ðf � UuiÞx ð3Þ The filter Uui is a complex valued m � m mask, defined in the discrete domain by Uui ¼ e �j2puT i yjy 2 Z2;kyk16 r n o ð4Þ where r = (m � 1)/2, and ui is a 2D frequency vector. In LPQ only four complex coefficients are considered, corresponding to 2D fre- quencies u1 = [a, 0] T, u2 = [0, a] T, u3 = [a, a] T, and u4 = [a, �a]T, where a = 1/m. For the sake of convenience, the STFT presented in Eq. (3) is expressed using the vector notation presented in Eq. (5) Fig. 9. LBP uniform pattern (Ojala et al., 2002). (a) The two transitions showed identifies the pattern as uniform, and (b) with four transitions, it is not considered a uniform pattern. Fig. 10. 2 � 2 Confusion matrix. D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 2075 f̂ uiðxÞ¼ w T ui fðxÞ ð5Þ where wu is the basis vector of the STFT at frequency u and f(x) is a vector of length m2 containing the image pixel values from the m � m neighborhood of x. Let F ¼ ½fðx1Þ; fðx2Þ; . . . ; fðxN2Þ� ð6Þ denote an m2 � N2 matrix that comprises the neighborhoods for all the pixels in the image and let w ¼ ½wR; wI� T ð7Þ where wR ¼ Re½wu1 ; wu2 ; wu3 ; wu4� and wI ¼ Im½wu1 ; wu2 ; wu3 ; wu4�. In this case, Re{�} and Im{�} return the real and imaginary parts of a complex number, respectively. The corresponding 8 � N2 transformation matrix is given by bF ¼ wF ð8Þ In Ojansivu and Heikkila (2008), the authors assume that the image function f(x) is a result of a first order Markov process, where the correlation coefficient between two pixels xi and xj is exponentially related to their L2 distance. Without a loss of gener- ality, they define each pixel to have unit variance. For the vector f, this leads to a m2 � m2 covariance matrix C with elements given by ci;j ¼ rkxi�xjk ð9Þ where k�k stands for the L2 norm. The covariance matrix of the Fou- rier coefficients can be obtained from D ¼ wCwT ð10Þ Since D is not a diagonal matrix, i.e., the coefficients are corre- lated, they can be decorrelated by using the whitening transforma- tion E ¼ V T bF where V is an orthogonal matrix derived from the singular value decomposition (SVD) of the matrix D that is D0 ¼ V T DV ð11Þ The whitened coefficients are then quantized using qi;j ¼ 1 if ei;j P 0; 0 otherwise � ð12Þ where ei,j are the components of E. The quantized coefficients are represented as integer values from 0 to 255 using binary coding bj ¼ X7 i¼0 qi;j2 i ð13Þ Finally, a histogram of these integer values from all the image positions is composed and used as a 256-dimensional feature vec- tor in classification. 5. Experiments and discussion Our experiments are divided into two parts. Those in the first part deal with verification, and those in the second part focus on the problem of writer identification. In all the experiments, support vector machines (SVM) were used as classifiers. The free parame- ters of the system and for SVM training were chosen using 5-fold cross validation. Various kernels were tried, and the best results were achieved using a Gaussian kernel. Parameters C and c were determined through a grid search. The overall error rate that we used for evaluation purposes in this work is given by Eq. (14). This rate is always computed on the testing set. Overall error rate ¼ FP þ FN TP þ TN þ FP þ FN ð14Þ where FP, FN, TP, and TN stand for False Positive, False Negative, True Positive, and True Negative, respectively. These statistics are defined in the 2 � 2 confusion matrix depicted in Fig. 10. One of the limitations of SVMs is that they do not work in a probabilistic framework. There are several situations where it would be very useful to have a classifier which produces a poster- ior probability P(classjinput). In our case, as depicted in Fig. 4, we are interested in estimating probabilities because we want to try different fusion strategies, like Sum, Max, Min, Average, and Med- ian. Due to the benefits of having classifiers estimating probabili- ties, many researchers have been working on the problem of estimating probabilities with SVM classifiers (Milgram, Cheriet, & Sabourin, 2005; Platt, 1999; Sollich, 2002). In this work, we have adopted the strategy proposed by Platt (1999). 5.1. Writer verification Writer verification is the task of determining whether or not a handwritten text has been written by a certain person. It is, by nat- ure, a binary problem. Given an input feature vector x extracted from a text S and a claimed identity I, determine whether or not (I, x) belongs to class x1 or x2. If it belongs to x1, the claim is true, i.e. the text has been written by author I, and if it belongs to x2, the claim is false, i.e. the text was produced by an impostor. Unlike the identification problem, where the task consists of identifying I among all the writers enrolled in the system, the verification task performs a 1:1 comparison. To establish some basis for comparison, we reproduced the best experiment reported in Hanusiak et al. (2011), but replaced the GLCM descriptors by LBP and LPQ. To this end, we used the BFL database with 100 and 115 writers for training and testing respec- tively. Five texture images per author were used as references (R = 5) to generate positive and negative samples, and five texture images (S = 5) were used for verification. The fusion rule applied to combine the classifier’s output was the Sum rule, which performs best, according to Hanusiak et al. (2011). Now, from Table 1, it is easy to see that both LBP and LPQ produce significantly better re- sults than the best results reported in Hanusiak et al. (2011). As mentioned earlier, we divided the IAM database in the same way that we divided the BFL database, i.e. we fixed a subset for testing (240 writers) and different subsets for training (50, 100, Table 1 Comparison between GLCM (Hanusiak et al., 2011), LBP, and LPQ (R = S = 5). Descriptor Overall error rate (%) GLCM (entropy) (Hanusiak et al., 2011) 5.0 LBP 1.3 LPQ 1.3 Table 2 Performance of the texture descriptor on the IAM database for different number of writers in the training set. Descriptor Fusion rule Overall error rate (%) Number of writers 25 100 205 410 LBP Max 4.8 2.5 1.3 0.4 Majority Vote 2.7 1.5 1.5 0.6 Sum 1.5 0.4 0.4 0.4 LPQ Max 1.4 0.4 2.5 1.2 Majority Vote 0.8 1.2 1.5 1.5 Sum 0.4 0.6 1.2 0.4 Table 3 Performance of the texture descriptors on writer verification for different texture image sizes – BFL database (R = S = 5). Size (W � H) Overall error rate (%) LBP LPQ 64 � 64 12.6 5.6 182 � 128 1.7 1.3 209 � 128 2.2 1.3 256 � 128 1.3 2.6 329 � 128 0.4 2.1 460 � 128 1.7 2.1 768 � 128 2.6 1.7 209 � 256 1.7 3.0 256 � 256 1.3 1.3 329 � 256 0.4 1.3 460 � 256 0.8 1.7 768 � 256 0.8 1.3 Table 4 Error rates (%) of the texture descriptors on the BFL database (R = S = 5). Fusion rule LBP LPQ Writers for training Writers for training 25 50 100 200 25 50 100 200 Sum 30.4 27.0 17.9 13.1 11.3 13.0 7.8 6.0 Max 61.7 52.5 46.6 33.1 20.0 6.9 5.3 9.5 Product 31.3 27.0 18.3 13.1 12.1 15.6 10.4 7.8 Median 37.8 24.8 20.0 12.8 13.0 8.7 5.2 6.9 Table 5 Error rates (%) of the texture descriptors on the IAM database (R = S = 5). Fusion rule LBP LPQ Writers for training Writers for training 50 100 205 410 50 100 205 410 Sum 80.4 77.5 31.3 20.0 25.4 20.8 19.5 17.5 Max 85.0 82.1 62.9 31.3 44.1 30.0 21.6 20.4 Product 80.2 79.2 33.0 22.5 27.9 22.0 21.6 20.0 Median 82.1 80.5 31.5 15.5 26.2 18.3 15.4 12.0 2076 D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 205, and 410 writers). As described in Section 2.2, some letters in the IAM database contain only a few lines of text, which does not allow us to generate fragments of 256 � 256 pixels. In order to have the same number of texture images considered in the BFL database, we used images of 256 � 128 pixels. The results for wri- ter verification on the IAM database using different fusion rules and varying the number of writers on the training set are reported in Table 2. In this experiment, we note that the behavior of LBP and LPQ differ slightly. While LBP achieves lower error rates as the number of writers in the training set increases, LPQ is able to achieve lower error rates using fewer writers for training. Table 2 shows that the classifier trained with LPQ is capable of achieving an error rate of 0.4% using no more than 25 writers for training. In the light of the remarkable results on the verification prob- lem for the two databases using texture images of 256 � 256 pixels (BFL) and 256 � 128 pixels (IAM), we wondered if we could reduce the size of the texture images without sacrificing performance. Smaller images have some advantages, such as processing smaller samples (e.g. with few lines of text), faster feature extraction, and the capacity to generate more pieces of texture, and hence more references. As we discuss later in this paper, a larger number of ref- erences is important when dealing with writer identification. These experiments were performed on the BFL database, which al- lows us to create different fragment sizes. In this case, 100 and 115 writers were considered for training and testing respectively. Ta- ble 3 shows the impact of using different sizes of texture image. We can draw two conclusions from Table 3. First, very small texture images, such as those 64 � 64 pixels, are not suitable for either LBP or LPQ. Second, larger images, those larger than 256 � 128 pixels, can yield some reduction in the overall error rates; however, it appears that, beyond a certain point, larger images contain more variability, which does not help in reducing the error rates. This to some extent corroborates the findings pre- sented in Hanusiak et al. (2011), where the authors show that tex- ture images larger than 256 � 256 pixels do not bring about any improvement. It is worth noting, though, that the descriptors con- sidered in that case were based on GLCM. 5.2. Writer identification According to the definition provided in Jain, Ross, and Prabhakar (2004), the identification problem consists of identifying writer I among all the writers enrolled in the system. Given an input fea- ture vector x from a texture image S, we determine the identity Ic, c 2 1, 2, . . . , N, where N is the number of writers enrolled in the system. Hence, S 2 Ic, if maxc{Dmodel(x, Rc)}, where Dmodel is the dis- similarity model trained to return an estimation of posterior prob- ability, which indicates that S and the reference Rc belong to the same writer. However, the identification system can also provide a list of documents that are similar to the queried document. The size of this list, also known as the hit list, can vary, e.g. 1, 5, or 10. The re- sults are then expressed in terms of TOP-1, TOP-5, or TOP-10 writer identification performance. This means that a hit list will be con- sidered correct if at least one version of the queried specimen ap- pears on it. Our baseline experiment for identification used the same proto- col as we applied for verification, i.e. R = S = 5, and texture images of 256 � 256 pixels for the BFL database and 256 � 128 pixels for the IAM database. Tables 4 and 5 show the TOP-1 performance for the BFL and the IAM database respectively, using different fu- sion rules and four different training sets. A look at the first part of Tables 4 and 5 reveals that the LBP- based classifier produces lower error rates as we increase the num- ber of writers in the training set. For the IAM database, this reduc- tion is more compelling. The error rate drops from 82% to 15.5% when the Median is used as the fusion rule. The second part of Ta- bles 4 and 5 present the performance of the LPQ-based classifier. Table 6 Error rates (%) for different number of writes and training references (R) – BFL database (S = 5). Number of references R LBP LPQ Writers for training Writers for training 25 50 100 200 25 50 100 200 3 52.8 18.3 18.5 18.3 11.3 9.5 5.2 6.0 5 37.8 24.8 20.8 12.8 13.0 8.7 5.2 6.9 7 52.2 32.2 18.3 8.7 7.8 7.8 7.8 6.9 9 31.3 19.2 13.9 7.9 12.1 4.3 7.8 6.5 Table 7 Error rates (%) for different number of writes and training references (R) – IAM database (S = 5). Number of references R LBP LPQ Writers for training Writers for training 50 100 205 410 50 100 205 410 3 42.5 74.1 30.8 20.8 14.1 17.5 17.0 13.75 5 82.1 80.5 31.5 15.5 26.2 18.3 15.4 12.0 7 45.8 48.3 26.3 15.5 20.4 19.5 13.3 11.6 9 57.5 52.9 22.9 12.0 12.5 12.0 13.7 10.4 Table 8 Evolution of the number of texture images for identification – BFL database (R = 9). Fusion rule LBP LPQ Number of references Number of references S = 3 S = 5 S = 7 S = 9 S = 3 S = 5 S = 7 S = 9 Sum 21.8 7.8 3.5 6.1 11.3 7.8 0.8 0.8 Max 28.7 25.2 28.7 27.0 10.4 7.8 5.2 6.9 Product 22.6 10.5 6.1 9.6 11.3 9.5 4.3 3.4 Median 23.5 9.6 5.4 5.3 13.9 9.5 2.6 0.8 TOP-5 2.7 0.8 0.8 0.8 4.3 3.4 0.0 0.8 TOP-10 0.8 0.8 0.8 0.8 1.7 0.1 0.0 0.0 D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 2077 For the IAM database, we observe the same behavior, i.e. the larger the number of writers in the training set, the smaller the error, but, for the BLF database the LPQ-based classifier is able to achieve the best performance without using the largest training set. In this case, the lower error rate (5.2%) was achieved with 100 writers in the training set. So far, we have used five texture images (R = 5) to generate the dissimilarity feature vectors and five texture images (S = 5) for identification. The fusion rules are then used to produce a final decision. One aspect worth investigating is the impact of the num- ber of references per writer used for training and identification. By increasing R, we increase the number of positive and negative sam- ples in the training set. By increasing S, we can rely on more data to produce a final decision. Since, in the previous experiments, the Median was the best fu- sion rule, we decided to adopt it for the subsequent experiments. Tables 6 and 7 show the evolution of the number of training refer- ences (R) for both the BFL and the IAM databases. It is easy to see from these tables that increasing R reduces the overall error rates. In spite of the size of R, Tables 6 and 7 are similar to Tables 4 and 5 respectively, in the sense that the LPQ-based classifier is able to achieve lower error rates with fewer writers in the training set. By analyzing the errors and the hit lists produced by the classi- fiers, we note that, in most cases, the correct writer was not very far from the classifier’s TOP-1 choice. With this in mind, we pro- pose increasing the number of texture images for identification (S). The rationale behind this is that, if we could count on more data to make a decision, we would profit from the information available on the hit list. Tables 8 and 9 show the evolution of S for the BFL and the IAM database respectively. In both cases, we used the largest training set available. TOP-5 and TOP-10 are re- lated to the Median rule. By adding more texture images for identification, we are able to reduce the overall error rates considerably. Our best results were achieved with the LPQ-based classifier in both databases, 0.8% and 3.3% for BFL and IAM respectively. Compared with our baseline results reported in Tables 4 and 5, the error rates were reduced by 4.4 and 8.7 percentage points for BFL and IAM respectively. Fig. 11 shows the cumulative match characteristic (CMC) curve Bolle, Connell, Pankanti, Ratha, and Senior (2005), which plots the probability of identification against the 1:N candidate list size re- turned. It shows the probability that a given user will appear on any of the candidate lists. The faster the CMC curve approaches 1, which indicates that the user always appears on a particular size of candidate list, the better the matching algorithm. In these fig- ures, we have plotted the results achieved with the median fusion rule. It is important to bear in mind that, in general, in this kind of system, a recognition rate of 100% on a TOP-1 hit list is not neces- sary, since a domain specialist can make a final decision based on a TOP-5 or a TOP-10 hit list. However, to be able to use the TOP-5 or the TOP-10 list efficiently, it is very important that these lists achieve a high performance. In our case, we can see from the CMC curves that we were able to reach performance above 99% in both databases, i.e. the correct answer is always on the TOP-5 hit list. Table 10 summarizes several works on writer verification/iden- tification reported in the literature. Comparing these results is not a straightforward task, since in some cases the database used is not publicly available. In the case of IAM, a more direct comparison is possible, since the dataset is publicly available. However, in several cases, the authors used only a subset of it. In our case, we divided the original dataset into two corpora of different writers, in order to better assess the dissimilarity framework. We believe that it is Table 9 Evolution of the number of texture images for identification – IAM database (R = 9). Fusion rule LBP LPQ Number of references Number of references S = 3 S = 5 S = 7 S = 9 S = 3 S = 5 S = 7 S = 9 Sum 30.8 16.7 9.6 7.1 31.2 16.2 8.7 9.1 Max 27.5 30.8 24.2 22.5 23.7 19.5 14.5 12.9 Product 31.3 19.8 12.5 9.6 31.6 19.1 12.5 10.4 Median 31.3 16.3 11.7 5.4 32.0 10.4 6.2 3.3 TOP-5 5.9 1.7 0.0 0.4 8.7 1.6 0.8 0.0 TOP-10 2.1 0.8 0.0 0.0 2.5 1.2 0.8 0.0 2078 D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 not fair to have the same writers in both the training and testing groups when using a dissimilarity-based system. In spite of the dif- ferent databases, Table 10 still gives us a good basis for comparison. 5.3. Impacts of the number of texture images used for identification From the experiments reported so far, we can assert the impor- tance of the number of texture images (S and R), as well as the fu- sion rules. In our experiments, we observed that, in general, the Median rule provides the best results. However, we wonder what Fig. 11. The CMC curves for (a) LPB-based classifier on BFL database, (b) LPB-based clas classifier on IAM database. could be gained by combining several decisions instead of relying on a single one. In other words, would it be better to use just one large texture image for identification, instead of m small ones? The idea behind using several small texture images is to better represent the writer’s variability, hoping that these images will provide a certain degree of complementarity which can be exploited by the fusion rules. In spite of the fact that the texture images are extracted from the same handwritten letter, it has been observed that one writer may use different writing styles in differ- ent parts of a single sample. To show the importance of using multiple texture images for identification in the proposed approach, we designed an experi- ment on the BFL database using an LPQ-based classifier. Instead of splitting the large texture area created from the questionable handwritten sample into m pieces, we used it to perform an iden- tification, i.e. S = 1 with a texture image of 2304 � 256 pixels. Since S = 1, the fusion rule is no longer necessary. Such a strategy pro- duces an elevated error rate of about 44%. This experiment can be compared with the second part of Table 4. As we can see, in that case, our worst result, an error rate of 13.9%, was achieved using three references (S = 3). The weak performance of this experiment can be explained by the fact that the large texture area contains all the variability of sifier on IAM database, (c) LPQ-based classifier on BFL database, and (d) LPQ-based Table 10 Summary of the state of the art on author verification and identification. Ref. Data Year Features Writers Classifier Performance (%) Verification Identification Hanusiak et al. (2011) BFL 2010 Texture 315 SVM 96.1 – Marti et al. (2001) IAM 2001 Structural 20 k-NN – 90.7 Schlapbach and Bunke (2004) IAM 2004 Geometric 120 HMM 97.5 96.5 Bensefia et al. (2005) IAM 2005 Graphemes 150 VSM 96.0 86.0 Schomaker and Bulacu (2007) IAM 2007 Graphenes 650 Dist. Hamming 97.2 89.0 Imdad et al. (2007) IAM 2007 Directional 30 SVM – 83.0 Schlapbach and Bunke (2007) IAM 2007 Geometric 100 HMM 97.5 96.0 Schlapbach and Bunke (2008) IAM 2008 Geometric 100 GMM – 97.8 Siddiqi and Vincent (2010) IAM 2010 Global and Local 650 Dist. X2 97.7 91.0 Kirli and Gulmezoglu (2011) IAM 2011 Global and Local 93 NDDF – 98.7 Said et al. (2000) – 1998 Gabor e GLCM 40 WED – 96.0 Zois and Anastassopoulos (2000) – 1999 Morphological 50 MLP – 96.5 Cha and Srihari (2000) – 2002 Micro and Macro 1000 k-NN – 81.0 Shen et al. (2002) – 2002 Texture 50 k-NN – 97.6 He and Tang (2004) – 2004 Gabor 50 WED – 97.0 Ubul et al. (2009) – 2009 Gabor and ICA 55 k-NN – 92.5 Ours BFL 2012 Texture (LPQ) 315 SVM 99.4 99.2 Ours IAM 2012 Texture (LPQ) 650 SVM 99.6 96.7 Table 11 Error rates (%) of different strategies of classification using the LPQ features. Strategy BFL IAM Dissimilarity 0.80 3.3 Pairwise 1.74 14.8 One-against-others 2.61 11.7 D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 2079 the queried document, but, after feature extraction, all that vari- ability is lumped into the same feature vector. 5.4. Comparing with writer-dependent approaches At this point, we may ask if the good performance reported in this work is due to: (i) the texture descriptors we used, (ii) the dis- similarity framework we used, or (iii) combining the dissimilarity framework with texture features. To address these points, we trained two other writer-dependent (WD) approaches. The wri- ter-dependent or personal model is a feature-based approach that considers one model per author. Usually, it yields good results, but its drawback is that, for each new author, a new model should be built. Another important issue in this strategy is that a consider- able amount of data is generally necessary to train a reliable model. In our case, the number of samples available for learning is small (9 fragments of texture per writer). The first WD model we implemented was a multi-class SVM using a pairwise approach. In this strategy, the number of classifi- ers that should be trained is q (q � 1)/2, where q is the number of classes (writers in our case). This approach shows its limitations as the number of writers increases. The second strategy was one-against-others decomposition, which works by constructing an SVM xi for each class q which first separates that class from all the others. Compared to the pairwise approach, the one-against-others strategy is more suitable for our application, because only one new model must be trained each time a new writer is enrolled in the system. In order to keep the same protocol, the number of classes (q) used in these experiments is the number of writers in the testing set, i.e. 115 for BFL and 240 for IAM. Unlike the dissimilarity proto- col, both approaches, pairwise and one-against-others, need sam- ples of the same writer in the training and testing sets. In the case of the BFL database, all the authors have three handwritten letters, and so we were able to split them into two samples (18 tex- ture images of 256 � 256 pixels) for training and one letter for test- ing (9 texture images of 256 � 256 pixels). In the IAM, by contrast, some authors have contributed only two samples, which allows us to divide them into one sample for training (9 texture images of 256 � 128 pixels) and the other for testing (9 texture images of 256 � 128 pixels). Table 11 summarizes the results. Regarding the questions raised at the beginning of this section, these results show us that the dissimilarity-based approach com- bined with the texture feature offer a robust framework for writer identification. In the case of the BFL database, where more samples are available for training, the dissimilarity approach achieved slightly better results. A considerably larger difference, though, can be observed for the IAM database, where the training set is smaller. The possibility of generating positive and negative sam- ples using the dichotomy transformation makes the dissimilarity approach suitable, even when only a few samples per writer are available. 6. Conclusion In this work, we have addressed the problems of writer verifica- tion and identification using the same framework as proposed by Hanusiak et al. (2011), in which explicit segmentation is avoided by generating a texture using the writer’s handwriting. Thereafter, features are extracted from the texture, and dissimilarity feature vectors are used to train an SVM classifier. We have demonstrated that both LBP and LPQ are interesting alternatives for describing this kind of texture. As in Ojansivu et al. (2008), we have observed in our work that the classification accuracy of LPQ is higher than with the well-known LBP. However, both LPQ and LBP surpass by a considerable margin the results achieved by GLCM descriptors in writer verification. Our experimental results show that the dissimilarity-based ap- proach that we have successfully applied to verification problems is also a viable strategy for identification problems, in that it achieves a performance comparable to the state of the art. We have shown the importance of a larger number of references for testing in this approach, and of limiting that number of references (S and R) to nine, so that all the writers are equally represented. We also show that the dissimilarity approach compares favorably with classic classification approaches, such as the pairwise and one- against-others methods. 2080 D. Bertolini et al. / Expert Systems with Applications 40 (2013) 2069–2080 On aspect worth investigating is the upper limit to the number of references used for testing. We plan to extend the current pro- tocol to use more references for testing when they are available. Based on the results reported in this work, we believe that we could reduce the errors for those writers even more if more hand- writing were available. In future work, we also plan to investigate whether or not all the available writers in the training set are re- quired, in order to build a good dissimilarity model. Acknowledgement This research has been supported by The National Council for Scientific and Technological Development (CNPq) Grant 301653/ 2011-9. References Bensefia, A., Paquet, T., & Heutte, L. (2005). A writer identification and verification system. Pattern Recognition Letters, 26(13), 2080–2092. Bertolini, D., Oliveira, L. S., Justino, E., & Sabourin, R. (2010). Reducing forgeries in writer-independent off-line signature verification through ensemble of classifiers. Pattern Recognition, 43(1), 387–396. Bolle, R. M., Connell, J. H., Pankanti, S., Ratha, N. K., & Senior, A. W. (2005). The relation between the ROC curve and the CMC. In 4th Workshop automatic identification advanced technologies (pp. 15–20). Bulacu, M., Schomaker, L., & Vuurpijl, L. (2003). Writer identification using edge- based directional features. In 8th International conference on document analysis and recognition, Edinburgh, Scotland (pp. 937–941). Bush, A., Boles, W., & Sridharan, S. (2005). On measuring the distance between histograms. IEEE Transations on Pattern Analysis and Machine Inteligence, 27, 1721–1732. Cha, S.-H., & Srihari, S. (2000). Multiple feature integration for writer verification. In 7th International workshop on frontiers on handwriting recognition (pp. 333–342). Cha, S.-H., & Srihari, S. N. (2002). On measuring the distance between histograms. Pattern Recognition, 35, 1355–1370. Freitas, C., Oliveira, L. S., Sabourin, R., & Bortolozzi, F. (2008). Brazilian forensic letter database. In 11th International workshop on frontiers on handwriting recognition, Montreal, Canada. Hanusiak, R., Oliveira, L. S., Justino, E., & Sabourin, R. (2011). Writer verification using texture-based features. International Journal on Document Analysis and Recognition, 15(3), 213–226. Haralick, R. M., Shanmugan, K. S., & Dunstein, I. (1973). Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, 3(6), 610–621. He, Z. Y., & Tang, Y. Y. (2004). Chinese handwriting-based writer identification by texture analysis. In 2004 International conference on machine learning and cybernetics (Vol. 6, pp. 3488–3491). Imdad, A., Bres, S., Eglin, V., Rivero-Moreno, C., & Emptoz, H. (2007). Writer identification using steered hermite features and svm. In 9th International conference on document analysis and recognition (pp. 839–843). Jain, A. K., Ross, A., & Prabhakar, S. (2004). An introduction to biometric recognition. IEEE Transactions on Circuits and Systems for Video Technology, 14, 4–20. Kirli, O., & Gulmezoglu, M. (2011). Automatic writer identification from text line images. International Journal on Document Analysis and Recognition, 1–15. Marti, U. V., & Bunke, H. (2002). The IAM-database: An english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5(1), 39–46. Marti, U. V., Messerli, R., & Bunke, H. (2001). Writer identification using text line based features. In 8th International conference on document analysis and recognition, Seattle, USA (pp. 101–105). Milgram, J., Cheriet, M., & Sabourin, R. (2005). Estimating accurate multi-class probabilities with support vector machines. In International joint conference on neural networks, Montreal, Canada (pp. 1906–1911). Ojala, T., Pietikäinen, M., & Harwook, D. (1996). Comparative study of texture measures with classification based on feature distributions. Pattern Recognition, 29, 51–59. Ojala, T., Pietikäinen, M., & Mäenpää, T. (2002). Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(7). Ojansivu, V., & Heikkila, J. (2008). Blur insensitive texture classification using local phase quantization. In Proceedings of image and signal processing (ICISP 2008) (pp. 236–243). Ojansivu, V., Rahtu, E., & Heikkila, J. (2008). Rotation invariant local phase quantization for blur insensitive texture analysis. In International conference on pattern recognition. Platt, J. (1999). Probabilistic outputs for support vector machines and comparison to regularized likelihood methods. In A. Smola et al. (Eds.), Advances in large margin classifiers (pp. 61–74). MIT Press. Rivard, D., Granger, E., & Sabourin, R. (2011). Multi-feature extraction and selection in writer-independent offline signature verification. International Journal on Document Analysis and Recognition. Said, H. E. S., Tan, T. N., & Baker, K. D. (2000). Personal identification based on handwriting. Pattern Recognition, 33, 149–160. Schlapbach, A., & Bunke, H. (2004). Using hmm-based recognizers for writer identification and verification. In Proceedings of 9th international workshop on frontiers in handwriting recognition (pp. 167–172). Schlapbach, A., & Bunke, H. (2007). A writer identification and verification system using hmm based recognizers. Pattern Analysis and Applications, 10, 33–43. Schlapbach, A., & Bunke, H. (2008). Off-line writer identification and verification using gaussian mixture models. In S. Marinai & H. Fujisawa (Eds.), Machine learning in document analysis and recognition. Studies in computational intelligence (Vol. 90, pp. 409–428). Berlin/Heidelberg: Springer. Schomaker, L., & Bulacu, M. (2007). Text-independent writer identification and verification using textural and allographic features. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(4), 701–717. Shen, C., Ruan, X. G., & Mao, T. L. (2002). Writer identification using gabor wavelet. In 4th World congress on intelligent control and automation (Vol. 3, pp. 2061– 2064). Siddiqi, I., & Vincent, N. (2010). Text independent writer recognition using redundant writing patterns with contour-based orientation and curvature features. Pattern Recognition, 43, 3853–3865. Sollich, P. (2002). Bayesian methods for support vecotr machines: Evidence and predictive class probabilities. Machine Learning, 46(1–3), 21–52. Srihari, S. N., Cha, S.-H., Arora, H., & Lee, S. (2002). Individuality of handwriting. Journal of Forensic Sciences, 47. Ubul, K., Tursun, D., Hamdulla, A., & Aysa, A. (2009). A feature selection and extraction method for uyghur handwriting-based writer identification. In 2009 International conference on computational intelligence and natural computing (pp. 345–348). Zois, E. N., & Anastassopoulos, V. (2000). Morphological waveform coding for writer identification. Pattern Recognition, 33(3), 385–398. Texture-based descriptors for writer identification and verification 1 Introduction 2 Databases 2.1 BFL database 2.2 IAM database 3 The dissimilarity framework 3.1 Dissimilarity feature vectors 4 Building textures and extracting features 4.1 Local binary patterns 4.2 Local phase quantization 5 Experiments and discussion 5.1 Writer verification 5.2 Writer identification 5.3 Impacts of the number of texture images used for identification 5.4 Comparing with writer-dependent approaches 6 Conclusion Acknowledgement References 
work_tam6owliffd2bjszksxq5j5cvy	----	Role and Performance of Different Traditional Classification and Nature-Inspired Computing Techniques in Major Research Areas 1 Diagnosis Heart Disease Using Mamdani Fuzzy Inference Expert System Iftikhar Naseer1,*, Bilal Shoaib Khan1, Shazia Saqib2, Syed Nadeem Tahir3, Sheraz Tariq1, Muhammad Saleem Akhter1 1Department of Computer Science, Minhaj University, Lahore, Pakistan. 2Department of Computer Science, Lahore Garrison University, Lahore, Pakistan. 3Ph.D. Scholar, IAS Department, University of the Punjab, Lahore, Pakistan. Abstract The death ratio caused by heart diseases is threating around the world. Efficient and accurate diagnosis through information technology can turn over this picture. This article proposed Diagnosis Heart Disease using Mamdani Fuzzy Inference (DHD-MFI) based expert system which intelligently diagnoses heart disease. In an explorative pattern, the current research has taken six conducive variables for the purpose of fuzzy logic technical enhancement in the diagnosis of heart disease. The input fields comprise of age, chest pain, electrocardiography, blood pressure systolic, diabetic and cholesterol are transmitted with the help of Fuzzy rules which are framed in the light of low, normal, high and very high intensity among the input variations. The single output is obtained as a clinical decision support system for the heart diagnosis by using the Mamdani Inference method. The proposed DHD-MFI based expert system gives 94% overall accuracy. Keywords: DHD-MFI, CAD, ECG, BP, CP, Mamdani Inference. Received on 02 August 2019, accepted on 05 January 2020, published on 15 January 2020 Copyright © 2020 Iftikhar Naseer et al., licensed to EAI. This is an open access article distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/3.0/), which permits unlimited use, distribution and reproduction in any medium so long as the original work is properly cited. doi: 10.4108/eai.15-1-2020.162736 *Corresponding author. Email: iftikharnaseer@gmail.com 1. Introduction Information Technology plays an important role in every aspect of life in this era. Decision support systems based on knowledge, proficiency and logical reasoning ability [1]. Fuzzy logic has the ability to handle the uncertainty of data in a proper approach [1, 2]. Heart disease is the main cause of death all over the world [3]. The proper diagnosis of heart disease is very important to survive from death [4]. Medical practitioners use their expertise to predict heart disease. Sometimes, doctors may get confused about the diagnosis of heart disease due to many risk factors involved in it [3, 4, 5]. According to Mehdi et. al., heart disease diagnosis is a very difficult task that needs accurate results [6]. Computer-Aided Diagnostic (CAD) system is very helpful in diagnosing heart disease condition of the patients. They pointed out that the consideration of accuracy always demands acute diagnosis of heart disease [6]. In this regard, the CAD system is very helpful in the diagnosis of heart disease and the suffering level of the patients. In this context, CAD is frequently used for the diagnosis of cardiovascular diseases all over the world but still, medical data faces problems to properly implement it [4, 6]. To handle this problem the researchers presented a system that is based on fuzzy logic for the heart disease diagnosis. The fuzzy logic system based on CAD with five input variables that consume a short time for the medical process to decide the diagnosis of heart disease. The performance of the proposed system was compared with the previous rough- fuzzy classifier system which used an open-source cardiovascular disease dataset. The results have shown 72.6% accuracy which is better than the past work [6]. Kasbe and Pippal designed a system with the help of fuzzy logic for the diagnosis of heart diseases [4]. This system has used 13 input variables and 1 output variable. The database has taken from V.A. Medical Center, Long EAI Endorsed Transactions on Scalable Information Systems Research Article EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e3 2 Beach, and Cleveland Clinic Foundation. The centroid method was applied for defuzzification. The medical experts, as well as the patients, can use this system with ease. The researchers have observed its accuracy 93.33% which is better than previous systems [4]. Aliz Adehsania et. al., observed many existing models for diagnosis heart diseases [5]. According to researchers, angiography is the best accurate system for the diagnosis of cardiovascular diseases but this method faces problems and cost issues. So, the researchers found a novelty technique for the diagnosis of heart disease by using data mining. The researchers proposed a model based on the classification to predict heart disease. The results of their model were better than the previous method [5]. Adeli and Neshat designed a fuzzy expert system for cardiovascular disease diagnosis based on V.A. Medical Centre, Long Beach and Cleveland Clinic Foundation database [7]. The system used 13 parameters as input and one parameter as output. The accuracy of the system was 94% which is better than the existing fuzzy expert systems [7]. 2. Related Work Anooj has proposed a system based on weight fuzzy rule for the diagnosis of cardiovascular disease [8]. In the first step of the system, the weighted fuzzy rules generated from the mining techniques and variable selections and in the second step, a fuzzy rule-based system is developed with the help of weighted fuzzy rules. The selected datasets Cleveland, Hungary, and Switzerland have a large amount of information about heart disease. This clinical decision support system has been tested and experimental results are found much accurate and precise [8]. Hassan et. al., stated that heart disease is the main reason for death all over the world [9]. The early prediction of cardiovascular disease can help to reduce the death rate. Many systems are available for early prediction of heart disease and still, there is a need to explore more. The researchers proposed a fuzzy soft expert system for the early detection of heart disease based on the fuzzy soft set theory. Their proposed system based on four parts i.e. fuzzy values; fuzzy soft set, fuzzy soft set parameter, and an algorithm. The study based on systolic blood pressure, low-density lipoprotein cholesterol, blood sugar, maximum heart rate, old pack and age variables. It is perceived that the suggested model is supportive for the medical experts in taking decisions regarding heart diseases [9]. Akinyokun et. al., designed a model based on the advancement of a Web and Fuzzy Logic-based Expert System for heart disease failure [10]. The proposed system comprises a Knowledge Base (which is comprised of a Database), a Fuzzy Logic part, a Fuzzy inference engine, and a decision support engine. All these elements of the system converted into the cognitive and emotional filters as well as tele-prescription facilities. Their proposed model was implemented using JavaScript, Hypertext Mark-up Language and Hypertext Preprocessor with My Structured Query Language as a database. The results shared with heart experts and satisfied with the performance of the system [10]. Bhatla et. al., introduced a new technique for coronary illness by using fuzzy logic and data mining [11]. This proposed methodology depended on Decision Tree and Naïve Bayes. The six parameters were diminished into four parameters i.e. Chest pain, blood pressure, No. of vessels colored and treadmill test which had less expense of lab test. In the proposed system, the output based on Normal, Low Risk, Medium Risk, and High Risk. The results stored in the database that connected with this model. In the database, the patient's record was maintained. The correctness and efficiency were measured through Naïve Bayes and Decision Tree using fuzzy logic. The results showed 0 for Decision Tree and 0.0072 for Naïve Bayes Classifier [11]. Computational intelligence techniques: Fuzzy System [12, 13], Swarm Intelligence [14], Evolutionary Computing [15, 16] and Neural Network [17] are mostly used in different fields of sciences like Intelligent Diagnose System [12], Smart Health [13], Smart City [18], Cloud Computing [19], Robotics [20], Wireless communication etc. Nowadays, hybrid structures of these four branches are very popular due to their performance [13, 14, 18]. In last decades predictions of the medical disease using machine learning approaches are one of the hot research areas. 3. Research Methodology The basic aim of this system is to present an applicable and reliable expert system for the acute diagnosis of heart disease. Figure 1 shows the proposed system research methodology. Iftikhar Naseer et al. EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e3 3 In the first phase, relevant variables are selected with the consultation of Medical Practitioners. If variables are less important, then they are discarded. In the second phase, the fuzzification is applied to crisp input values for the transformation of input sets. After that fuzzy rules are applied on fuzzy input sets in fuzzy inference. After fuzzy rules, defuzzification is applied on fuzzy output set which finally converts fuzzy output set into crisp output. 3.1. Proposed DHD-MFI Expert System In proposed DHD-MFI expert system based on six input parameters like Age (A), Blood Pressure systolic (BP), Cholesterol (C), Diabetic (D), Chest Pain (CP) and Electrocardiography (ECG) and one output Diagnosis Heart Disease (DHD) is used as shown in figure 2. Fig. 2. Proposed DHD-MFI Expert System Proposed Diagnosis Heart Disease-MFI expert system mathematically can be written as: 𝜇𝜇𝐴𝐴∩𝐵𝐵𝐵𝐵∩𝐶𝐶∩𝐷𝐷∩𝐶𝐶𝐵𝐵∩𝐸𝐸𝐶𝐶𝐸𝐸(𝑎𝑎, 𝑏𝑏𝑏𝑏, 𝑐𝑐, 𝑑𝑑, 𝑐𝑐𝑏𝑏, 𝑒𝑒𝑐𝑐𝑒𝑒) = min�𝜇𝜇𝐴𝐴(𝑎𝑎), 𝜇𝜇𝐵𝐵𝐵𝐵(𝑏𝑏𝑏𝑏), 𝜇𝜇𝐶𝐶(𝑐𝑐), 𝜇𝜇𝐷𝐷(𝑑𝑑), 𝜇𝜇𝐶𝐶𝐵𝐵(𝑐𝑐𝑏𝑏), 𝜇𝜇𝑒𝑒𝑒𝑒𝑒𝑒(𝑒𝑒𝑐𝑐𝑒𝑒)� (1) 3.2. Input Fuzzy Sets: Statistical values of input fuzzy parameters are used for the diagnosis of heart disease as shown in table 1. In this research, six input parameters are used with the following description. Table 1. Proposed DHD-MFIS Input Fuzzy Sets Sr. No. Input Variables Range Description 1 Age (Year) 0 – 18 14 – 38 >36 Child Young Old 2 Blood Pressure BP Systolic (mm Hg) < 90 80 – 140 >138 Low Normal High 3 Cholesterol mg/dL (mg per < 95 90-120 Optimal Desirable Fig. 1. Proposed DHD-MFI Expert System Research Methodology Age MFI Expert System DHD Diabetic Cholesterol Blood Pressure Chest Pain ECG Diagnosis Heart Disease Using Mamdani Fuzzy Inference Expert System EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e3 4 deciliter) >125 Above Normal 4 Diabetic (HbA1c) 0 – 0.6 0.55 – 8.0 >7.5 Controlled Well Controlled Uncontrolled 5 Chest Pain 0 – 0.35 0.3 – 0.70 0.65 – 1 Non-Cardiac Atypical Typical 6 ECG(Electro- cardiography) (mV) 0 – 0.35 0.3 – 0.70 0.65 – 1 Non-Significant ST-Segment Depression ST-Segment 3.3. Output Fuzzy Sets: The Diagnosis Heart Diseases (DHD) is the output of the proposed DHD-MFI expert system which is shown in table 2. Table 2. Output Fuzzy Sets Sr. No. Output Variables Range Description 1 Diagnosis Heart Diseases (DHD) 0 – 0.28 0.25 – 0.58 0.50 – 0.82 0.75 – 1 Negative Border Line Positive Strongly Positive 3.4. Membership Functions: Membership function provides values between 0 and 1. Tram membership function is used in the proposed DHD- MFI based expert system shown in table 3. The Proposed DHD-MFI expert system Input/Output (I/O) variables mathematically Membership Functions (MF) are shown in table 3. Table 3. Proposed DHD-MFI Mathematical MF representation of I/O variable S r. N o Input Variabl es Mathematical Representation of Membership Functions 1 Age (Year) µAge (a) 𝜇𝜇𝐴𝐴,𝐶𝐶 (𝑎𝑎) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚�1, 18 − 𝑎𝑎 4 � , 0�� 𝜇𝜇 𝐴𝐴,𝑌𝑌(𝑎𝑎) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑎𝑎 − 14 4 , 1, 42 − 𝑎𝑎 4 � , 0�� 𝜇𝜇𝐴𝐴,𝑂𝑂 (𝑠𝑠) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑎𝑎 − 38 6 , 1 � , 0�� 2 Blood Pressure BP Systolic (mm Hg) µBS (bs) 𝜇𝜇𝐵𝐵𝐵𝐵,𝐿𝐿 (𝑏𝑏𝑠𝑠) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚�1, 90 − 𝑏𝑏𝑠𝑠 10 � , 0�� 𝜇𝜇 BS,𝑁𝑁(𝑏𝑏𝑠𝑠) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑏𝑏𝑠𝑠 − 80 10 , 1, 140 − 𝑏𝑏𝑠𝑠 5 � , 0�� 𝜇𝜇𝐵𝐵𝐵𝐵,𝐻𝐻 (𝑏𝑏𝑠𝑠) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑏𝑏𝑠𝑠 − 138 2 , 1 � , 0�� 3 Cholesterol mg/dL (milligrams per deciliter) µLDL (ch) 𝜇𝜇𝐶𝐶,𝑂𝑂 (𝑐𝑐) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚�1, 90 − 𝑐𝑐 5 � , 0�� 𝜇𝜇 C,𝐷𝐷(𝑐𝑐) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑐𝑐 − 95 5 , 1, 125 − 𝑐𝑐 5 � , 0�� 𝜇𝜇𝐶𝐶,𝐴𝐴𝑁𝑁 (𝑐𝑐) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� c − 120 5 , 1 � , 0�� 4 Diabetic (HbA1c) µD (d) 𝜇𝜇𝐷𝐷,𝐶𝐶 (𝑑𝑑) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚�1, 0.6 − 𝑑𝑑 . 05 � , 0�� 𝜇𝜇 D,WC(𝑑𝑑) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑑𝑑 − 0.55 0.5 , 1, 0.80 − 𝑑𝑑 0.5 � , 0�� 𝜇𝜇𝐷𝐷,𝑈𝑈𝑁𝑁 (𝑑𝑑) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� d − 0.75 0.05 , 1 � , 0�� 5 Chest Pain µCP(cp) 𝜇𝜇𝐶𝐶𝐵𝐵,NC (𝑐𝑐𝑏𝑏) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚�1, 0.35 − 𝑐𝑐𝑏𝑏 . 05 � , 0�� 𝜇𝜇 CP,AT(𝑐𝑐𝑏𝑏) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑐𝑐𝑏𝑏 − 0.3 0.5 , 1, 0.70 − 𝑐𝑐𝑏𝑏 0.5 � , 0�� 𝜇𝜇𝐶𝐶𝐵𝐵,𝑇𝑇 (𝑐𝑐𝑏𝑏) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� cp − 0.65 0.05 , 1 � , 0�� 6 ECG (Electrocar dio graph) (mV) µECG(ecg) 𝜇𝜇𝐸𝐸𝐶𝐶𝐸𝐸,NC (𝑒𝑒𝑐𝑐𝑒𝑒) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚�1, 0.35 − 𝑒𝑒𝑐𝑐𝑒𝑒 . 05 � , 0�� 𝜇𝜇 ECG,ST−SD(𝑒𝑒𝑐𝑐𝑒𝑒) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑒𝑒𝑐𝑐𝑒𝑒 − 0.3 0.5 , 1, 0.70 − 𝑒𝑒𝑐𝑐𝑒𝑒 0.5 � , 0�� 𝜇𝜇𝐸𝐸𝐶𝐶𝐸𝐸,ST−S (𝑒𝑒𝑐𝑐𝑒𝑒) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� ecg − 0.65 . 05 , 1� , 0�� 7 Diagnosis Heart Disease (DHD) 𝜇𝜇𝐷𝐷𝐻𝐻𝐷𝐷,Negative (𝑑𝑑ℎ𝑑𝑑) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚�1, 0.28 − 𝑑𝑑ℎ𝑑𝑑 . 03 � , 0�� 𝜇𝜇 DHD,BorderLine(𝑑𝑑ℎ𝑑𝑑) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑑𝑑ℎ𝑑𝑑 − 0.25 0.3 , 1, 0.58 − 𝑑𝑑ℎ𝑑𝑑 0.8 � , 0�� 𝜇𝜇 DHD,Positive(𝑑𝑑ℎ𝑑𝑑) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� 𝑑𝑑ℎ𝑑𝑑 − 0.50 0.8 , 1, 0.82 − 𝑑𝑑ℎ𝑑𝑑 0.7 � , 0�� 𝜇𝜇𝐷𝐷𝐻𝐻𝐷𝐷,StronglyPositive (𝑑𝑑ℎ𝑑𝑑) = �𝑚𝑚𝑎𝑎𝑚𝑚 �𝑚𝑚𝑚𝑚𝑚𝑚� dhd − 0.75 . 07 , 1� , 0�� 3.5. Graphical Representation of Membership Function: The Proposed DHD-MFI expert system I/O variables graphical membership functions are shown in table 4. Table 4. Proposed DHD-MFI Graphical MF representation of I/O variable Sr. No. Input Variables Graphical Representation of Membership Functions 1 Age (Year) µAge (a) 2 Blood Pressure BP Systolic Iftikhar Naseer et al. EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e3 5 (mm Hg) µBPS (bp) 3 Cholesterol mg/dL (milligrams per deciliter) µLDL (ch) 4 Diabetic (HbA1c) µDiabetic (d) 5 Chest Pain µChest Pain(cp) 6 ECG (Electro cardio graphy) (mV) µECG(ecg) 7 Diagnosis Heart Disease 3.6. Fuzzy Proposition: Fuzzy proposition for the proposed Diagnosis Heart Disease-MFI expert System is: t: a x bp x c x d x cp x ecg → DHD (2) where a, bp, c, d, cp, ecg & dhd represents Age, Blood Pressure Systolic, Cholesterol, Diabetic, Chest Pain, Electrocardiography and Diagnosis Heart Disease respectively. All input and output parameter values are mapped from real range to probability ranges because the fuzzy expert system works on probability range 0-1 [20, 21]. Hence, the function t-Norm in equation (2) is defined as: t: [0,1]x [0,1]x [0,1]x [0,1]x [0,1]x [0,1] → [0,1] (3) Membership function of Diagnosis Heart Disease can be written as: 𝜇𝜇𝐷𝐷𝐻𝐻𝐷𝐷(𝑑𝑑ℎ𝑑𝑑) = 𝑡𝑡[𝜇𝜇𝐴𝐴(𝑎𝑎), 𝜇𝜇𝐵𝐵𝐵𝐵(𝑏𝑏𝑏𝑏), 𝜇𝜇𝐶𝐶(𝑐𝑐), 𝜇𝜇𝐷𝐷(𝑑𝑑), 𝜇𝜇𝐶𝐶𝐵𝐵(𝑐𝑐𝑏𝑏), 𝜇𝜇𝑒𝑒𝑒𝑒𝑒𝑒(𝑒𝑒𝑐𝑐𝑒𝑒)] = min�𝜇𝜇𝐴𝐴(𝑎𝑎), 𝜇𝜇𝐵𝐵𝐵𝐵(𝑏𝑏𝑏𝑏), 𝜇𝜇𝐶𝐶(𝑐𝑐), 𝜇𝜇𝐷𝐷(𝑑𝑑), 𝜇𝜇𝐶𝐶𝐵𝐵(𝑐𝑐𝑏𝑏), 𝜇𝜇𝑒𝑒𝑒𝑒𝑒𝑒(𝑒𝑒𝑐𝑐𝑒𝑒)� (4) 3.7 Rule Base: The proposed Diagnosis Heart Disease-Mamdani Fuzzy Inference expert system consists of 6 inputs and 1 output which makes 729 all possible fuzzy rules. The fuzzy rules are developed with the help of medical experts/practitioners. Some fuzzy rules of the proposed system are shown below. Conditional statement IF-THEN fuzzy rules are applied to the membership functions. These IF-THEN rules are part of the fuzzy rule base. Rules surface, rules viewer, etc. are depending upon fuzzy rule base. The fuzzy rule base of our expert system has 729 rules. Rules are denoted by Rdhd n , where 1 ≤ n ≤ 729. 𝑹𝑹𝒅𝒅𝒅𝒅𝒅𝒅 𝟏𝟏 = IF Age is Young AND Blood Pressure is Normal AND Cholesterol is optimal AND Diabetic is controlled AND Chest Pain is Non-Cardiac AND ECG is Non- Significant THEN Diagnosis Heart Disease is Negative. 𝑹𝑹𝒅𝒅𝒅𝒅𝒅𝒅 𝟐𝟐 = IF Age is Young AND Blood Pressure is High AND Cholesterol is optimal AND Diabetic is controlled AND Chest Pain is Non-Cardiac AND ECG is ST- Segment Depression THEN Diagnosis Heart Disease is Positive. 𝑹𝑹𝒅𝒅𝒅𝒅𝒅𝒅 𝟑𝟑 = IF Age is Young AND Blood Pressure is low AND Cholesterol is optimal AND Diabetic is controlled AND Chest Pain is Non-Cardiac AND ECG is ST- Segment Elevation THEN Diagnosis Heart Disease is Strongly Positive. 𝑹𝑹𝒅𝒅𝒅𝒅𝒅𝒅 𝟒𝟒 = IF Age is Old AND Blood Pressure is high AND Cholesterol is optimal AND Diabetic is well controlled AND Chest Pain is Non-Cardiac AND ECG is ST- Diagnosis Heart Disease Using Mamdani Fuzzy Inference Expert System EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e3 6 Segment Depression THEN Diagnosis Heart Disease is Positive. . . . 𝑹𝑹𝒅𝒅𝒅𝒅𝒅𝒅 𝟕𝟕𝟐𝟐𝟕𝟕 = IF Age is old And Blood Pressure is Very High And Cholesterol is above normal AND Diabetic is uncontrolled AND Chest Pain is typical AND ECG is ST-Segment Elevation THEN Diagnosis Heart Disease is Strongly Positive. 3.8 Fuzzy Inference Engine: The fuzzy inference engine is the method towards representing an offered contribution to a yield utilizing fuzzy logic [20, 21]. Membership functions, fuzzy logic operators and IF-Then rules are an important part of the Fuzzy Inference Engine [21]. All rules in the fuzzy rule base are combined into a single fuzzy relation that lies under the inner product on the input universe of discourse, which is then viewed as an only fuzzy IF-THEN rule. A suitable operator for combining the rules is union. Let RDHD n be a fuzzy relation that represents fuzzy IF- THEN rule of the proposed expert system which is, 𝑅𝑅𝐷𝐷𝐻𝐻𝐷𝐷 𝑛𝑛 = 𝐴𝐴𝑛𝑛 x 𝐵𝐵𝐵𝐵𝑛𝑛 x 𝐶𝐶𝑛𝑛 x 𝐷𝐷𝑛𝑛 x 𝐶𝐶𝐵𝐵𝑛𝑛 x 𝐸𝐸𝐶𝐶𝐸𝐸𝑛𝑛 → 𝐷𝐷𝐷𝐷𝐷𝐷𝑛𝑛 The above equation can be written as 𝜇𝜇𝐴𝐴∩𝐵𝐵𝐵𝐵∩𝐶𝐶∩𝐷𝐷∩𝐶𝐶𝐵𝐵∩𝐸𝐸𝐶𝐶𝐸𝐸(𝑎𝑎, 𝑏𝑏𝑏𝑏, 𝑐𝑐, 𝑑𝑑, 𝑐𝑐𝑏𝑏, 𝑒𝑒𝑐𝑐𝑒𝑒) = 𝜇𝜇𝐴𝐴(𝑎𝑎) ∩ 𝜇𝜇𝐵𝐵𝐵𝐵(𝑏𝑏𝑏𝑏) ∩ 𝜇𝜇𝐶𝐶(𝑐𝑐)) ∩ 𝜇𝜇𝐷𝐷(𝑑𝑑)) ∩ 𝜇𝜇𝐶𝐶𝐵𝐵(𝑐𝑐𝑏𝑏)) ∩ 𝜇𝜇𝐸𝐸𝐶𝐶𝐸𝐸(𝑒𝑒𝑐𝑐𝑒𝑒) The rules of the proposed system are interpreted as a single fuzzy relation defined by 𝑅𝑅729 = ⋃ 𝑅𝑅𝐻𝐻 𝑛𝑛729 𝑛𝑛=1 (5) This combination of rules is called a Mamdani combination. Assume iφε and Ψ be input and output fuzzy sets respectively. We obtain the output of the fuzzy inference engine as 𝜇𝜇𝑁𝑁𝑒𝑒𝑒𝑒𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑒𝑒 ∩𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝑒𝑒𝐵𝐵𝐵𝐵𝑁𝑁𝑛𝑛𝑒𝑒 ∩𝐵𝐵𝐵𝐵𝑃𝑃𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑒𝑒 ∩𝐵𝐵𝑁𝑁𝐵𝐵𝐵𝐵𝑛𝑛𝑒𝑒𝐵𝐵𝑆𝑆𝐵𝐵𝐵𝐵𝑃𝑃𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑒𝑒(𝜃𝜃) = 𝑆𝑆𝑆𝑆𝐵𝐵𝑁𝑁 𝜖𝜖 (𝐴𝐴,𝐵𝐵𝐵𝐵,𝐶𝐶,𝐷𝐷,𝐶𝐶𝐵𝐵,𝐸𝐸𝐶𝐶𝐸𝐸)𝑡𝑡[𝑏𝑏𝑁𝑁, 𝑙𝑙] (6) Where, 𝑏𝑏𝑁𝑁 = µi(a, bp, ch, d, cp, ecg) And 𝑙𝑙 = µ729(a, bp, ch, d, cp, ecg, DHD) 3.9 Product Inference Engine: The Product Inference Engine (PIE) of the proposed Diagnosis Heart Disease-MFI expert system can be written as 𝜇𝜇𝜃𝜃(Diagnosis Heart Disease) = max 1≤𝑛𝑛≤729 �𝑆𝑆𝑆𝑆𝐵𝐵𝑁𝑁 𝜖𝜖 (𝐴𝐴,𝐵𝐵𝐵𝐵,𝐶𝐶,𝐷𝐷,𝐶𝐶𝐵𝐵,𝐸𝐸𝐶𝐶𝐸𝐸) ��(𝐸𝐸) 729 𝑗𝑗=1 , 𝐽𝐽𝑁𝑁 �� (7) Where, 𝐸𝐸 = µ𝐴𝐴,𝐵𝐵𝐵𝐵,𝐶𝐶,𝐷𝐷,𝐶𝐶𝐵𝐵,𝐸𝐸𝐶𝐶𝐸𝐸(𝑎𝑎, 𝑏𝑏𝑏𝑏, 𝑐𝑐, 𝑑𝑑, 𝑐𝑐𝑏𝑏, 𝑒𝑒𝑐𝑐𝑒𝑒) And 𝐽𝐽𝑁𝑁 = µ𝐴𝐴1𝑁𝑁,𝐴𝐴2𝑁𝑁,𝐴𝐴3𝑁𝑁,𝐴𝐴4𝑁𝑁�𝑎𝑎1,𝑎𝑎2,𝑎𝑎3,𝑎𝑎4,� 3.10 De-fuzzifier: Defuzzifier is known for generating a quantifiable inference in crisp logic on the basis of fuzzy sets and corresponding membership degree [20, 21]. It is supposed to be the vital need that is fulfilled for the proper fuzzy control system. Defuzzifier is one of the definitive crucial parts of a fuzzy expert system. In defuzzifier, fuzzy sets converted into the crisp output. Centre of gravity defuzzifier, the center of average defuzzifier and maximum defuzzifier are three sorts of defuzzifier [12, 13]. The important defuzzifier amongst them is the center of gravity defuzzifier is used in the proposed DHD-MFI expert system. The center of gravity is taken as a useful defuzzifier technique. The center of gravity defuzzifier specifies the δ∗ as the center of the area covered by the membership function of δ that is δ∗ = ∫𝜃𝜃𝜇𝜇𝜃𝜃𝜃𝜃𝑑𝑑𝜃𝜃 ∫𝜇𝜇𝜃𝜃 𝜃𝜃𝑑𝑑𝜃𝜃 (8) The graphical representation of defuzzifier of the proposed Diagnosis Heart Disease-MFI expert system is shown in fig 3 & 4. Iftikhar Naseer et al. EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e3 7 Fig. 3. Proposed DHD-MFI expert system Rule surface of w.r.t. ECG & Diabetic. Fig. 3 demonstrates the output of the diagnosis of heart disease based on ECG (X-axis) and Diabetic (Y-axis). It shows that if ECG having values between 0-0.3 and Diabetic values 0-0.5 then there is no chance of Heart Disease. If ECG values greater than 0.3-0.65 and Diabetic values 0.5-0.75 then heart disease is positive. If ECG values 0.65-1 and Diabetic values 0.75 – 1 then Diagnosis Heart Disease is strongly positive. Fig. 4. Proposed DHD-MFI expert system Rule surface of w.r.t. Chest Pain & ECG. Fig. 4 displays the output of the diagnosis of heart disease consists of Chest Pain and ECG. If the Chest Pain is non- cardiac (0- 0.3) and ECG is non-significant (0-0.3) then there is no chance of a diagnosis of heart disease. If Chest Pain is atypical or typical and ECG is non-significant then heart disease diagnoses borderline. All values (non- cardiac, atypical, typical) of Chest Pain and ECG ST- Segment Depression then heart disease level on positive. All values (non-cardiac, atypical, typical) of Chest Pain and ECG ST-Segment Elevation then heart disease level on strongly positive. 4. Outcomes Figure 5 shows the accuracy of the proposed DHD-MFI expert system which is plotted after the comparison of predicted proposed system outcomes with human experts. The proposed DHD-MFI expert system has shown 93% accuracy along with the minor probability of error 7% in picking Negative cases. In the same way, this system has selected 94% borderline cases accurately with the probability of error 6%. Furthermore, it is seen that 96% positive cases are accurately identified with the probability of error 4%. Whereas, 94% strongly positive cases are chosen accurately by the system with the probability of error 6%. Finally, the overall efficiency of the system is found 94% and the probability of error is 6%. Fig. 5. Accuracy Chart of Proposed DHD-MFI expert system 5. Discussion The primary objective of the present study to design a comprehensive expert system for heart disease diagnosis by test reports taken from the Cardiac Department of Punjab Institute of Cardiology Hospital, Lahore. Pakistan. The efficiency of the proposed method is randomly checked on 280 records. Negative results obtained from 70 samples in which 5 were wrong. The probability of correctness and error was 0.93 and 0.07 respectively. The probability of correctness and error was 0.94 and 0.06 Diagnosis Heart Disease Using Mamdani Fuzzy Inference Expert System EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e3 8 respectively while taking 70 samples of Borderline in which 4 are incorrect. From 70 Positive results out of which 3 are incorrect providing probability of correctness and error 0.96 and 0.04 respectively and 70 from Strong Positive 0.94 and 0.06 was the probability of correctness and error. Cumulative probability of correctness was 0.94 obtained from the proposed system and probability of error was 0.06.The proposed expert system is basic and simple to use for both medical experts and non- professional. The main goal of this explorative research is to analyze the various degrees of Heart illness. 6. Conclusion The precision of the proposed Diagnosis Heart Disease- Mamdani Fuzzy Inference expert system is 94%. The efficacy of DHD-MFI expert system in the Pakistani context has provided a unique perception of scale in terms of the impact of six variables in driving heart complexity. When the results of the DHD-MFI expert system are shared with the heart experts, they found them very much helpful in making a decision regarding the level of heart disease. On these bases, the proposed system has provided a novelty of strength and support in the treatment and diagnosis of cardiovascular ailments. 7. Future Work The present research work has opened new avenues for future researchers in the field of heart diagnosis. This can be possible by improving the more correctness and accuracy of the proposed system with the help of the neural network, neuro-fuzzy, swarm intelligence and evolutionary computing. The present work can likewise be reached out to another better and reliable treatment (arrangement) of heart diseases like Echocardiogram, Angiograph, and Bypass Surgery. References 1. Atta, A., Abbas, S., Khan, M. A., Ahmed, G., & Farooq, U. (2018). An adaptive approach: Smart traffic congestion control system. Journal of King Saud University-Computer and Information Sciences. 2. Iqbal, K., Khan, M. A., Abbas, S., Hasan, Z., & Fatima, A. (2018). Intelligent Transportation System (ITS) for Smart-Cities using Mamdani Fuzzy Inference System. INTERNATIONAL JOURNAL OF ADVANCED COMPUTER SCIENCE AND APPLICATIONS, 9(2), 94-105. 3. Hassan, A., Bilal, H. M., Khan, M. A., Khan, M. F., Hassan, R., & Farooq, M. S. (2018, October). Enhanced Fuzzy Resolution Appliance for Identification of Heart Disease in Teenagers. In International Conference on Intelligent Technologies and Applications, Springer, Singapore, 28-37. 4. Kasbe, T., & Pippal, R. S. (2017, August). Design of heart disease diagnosis system using fuzzy logic. In 2017 International Conference on Energy, Communication, Data Analytics and Soft Computing (ICECDS) (pp. 3183-3187). IEEE. 5. Aliza dehsani, R., Zangooei, M. H., Hosseini, M. J., Habibi, J., Khosravi, A., Roshanzamir, M & Nahavandi, S. (2016). Coronary artery disease detection using computational intelligence methods. Knowledge-Based Systems, 109, 187- 197. 6. Mahdi, M., Faisal, I., Shakir, M.M., Praveen, S., & Aqilah, B. H. (2019). Fuzzy Logic System for Diagnosing Coronary Heart Disease. International Journal of Engineering & Technology, 8 (1.7), 119-125. 7. Adeli, A., & Neshat, M. (2010, March). A fuzzy expert system for heart disease diagnosis. In Proceedings of International Multi Conference of Engineers and Computer Scientists, Hong Kong (Vol. 1, pp. 28-30). 8. Anooj, P. K. (2012). Clinical decision support system: Risk level prediction of heart disease using weighted fuzzy rules. Journal of King Saud University-Computer and Information Sciences, 24(1), 27-40. 9. Hassan, N., Sayed, O. R., Khalil, A. M., & Ghany, M. A. (2017). Fuzzy soft expert system in prediction of coronary artery disease. International Journal of Fuzzy Systems, 19(5), 1546-1559. 10. Akinyokun, O. C., Babatunde, I. G., Arekete, S. A., & Samuel, R. W. (2015). Fuzzy logic-driven expert system for the diagnosis of heart failure disease. Artif. Intell. Research, 4(1), 12-21. 11. Bhatla, N., & Jyoti, K. (2012). A Novel Approach for heart disease diagnosis using Data Mining and Fuzzy logic. International Journal of Computer Applications, 54(17). 12. Ahmad, G., Khan, M. A., Abbas, S., Athar, A., Khan, B. S., & Aslam, M. S. (2019). Automated Diagnosis of Hepatitis B Using Multilayer Mamdani Fuzzy Inference System. Journal of healthcare engineering, 2019. 13. Zahra, S. B., Athar, A., Khan, M. A., Abbas, S., & Ahmad, G. (2019). Automated diagnosis of liver disorder using multilayer neuro- fuzzy. International Journal Of Advanced And Applied Sciences, 6(2), 23-32. 14. AsadUllah, M., Khan, M. A., Abbas, S., Athar, A., Raza, S. S., & Ahmad, G. (2018). Blind Channel and Data Estimation Using Fuzzy Logic-Empowered Opposite Learning-Based Iftikhar Naseer et al. EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e3 9 Mutant Particle Swarm Optimization. Computational Intelligence and Neuroscience, 2018. 15. Khan, M. A., Umair, M., & Choudhry, M. A. S. (2015). GA based adaptive receiver for MC- CDMA system. Turkish Journal of Electrical Engineering & Computer Sciences, 23(Sup. 1), 2267-2277. 16. Umair, M., Khan, M. A., &Choudry, M. A. S. (2013, January). GA backing to STBC based MC-CDMA systems. In Intelligent Systems Modelling & Simulation (ISMS), 2013 4th International Conference on (pp. 503-506). IEEE 17. Anwar Saeed, Muhammad Yousif, Areej Fatima, Sagheer Abbas, Muhammad Adnan Khan, Leena Anum & Ali Akram. (2019). An Optimal Utilization of Cloud Resources using Adaptive Back Propagation Neural Network and Multi-Level Priority Queue Scheduling. The ISC International Journal of Information Security (ISeCure). 11(special issue), 145-151. doi: 10.22042/isecure.2019.11.0.19. 18. Sagheer Abbas, Tahir Alyas, Atifa Athar, Muhammad Adnan Khan, Areej Fatima & Waseem Ahmed Khan. (2019). Ranking of Cloud Services by measuring multiple Parameters using Adaptive Neuro-Fuzzy Inference System, EAI Endorsed Transactions on Scalable Information Systems, 18: e5, P.1-7, doi: 10.4108/eai.13-7-2018.159354. 19. Areej Fatima, Muhammad Adnan Khan, Sagheer Abbas, Muhammad Waqas, Leena Anum & Muhammad Asif. (2019). Evaluation of Planet Factors of Smart City through Multi- layer Fuzzy Logic (MFL). Isecure-Isc International Journal Of Information Security. Volume 11. Special Issue Summer and autumn, pp.51-58, DOI: 10.22042/ISECURE.2019.11. 0.7. 20. Sadaf Hussain, Sagheer Abbas, Tanweer Sohail, Muhammad Adnan Khan & Atifa Athar. (2019). Estimating Virtual Trust of Cognitive Agents using Multi-layered Socio-Fuzzy Inference System” Journal of Intelligent & Fuzzy Systems, vol. Pre-press, no. Pre-press, pp. 1-16. DOI: 10.3233/JIFS-18760. 21. Ayesha Atta, Muhammad Adnan Khan, Sagheer Abbas, Gulzar Ahmad & Areej Fatima,(2019). Modeling Smart Road Traffic Congestion Control System Using Machine Learning Techniques, Neural Network World, 19(2), 99- 110. DOI: 10.14311/NNW.2019.29.008. Diagnosis Heart Disease Using Mamdani Fuzzy Inference Expert System EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e3 http://dx.doi.org/10.4108/eai.13-7-2018.159354 Abstract 1. Introduction 2. Related Work 3. Research Methodology 4. Outcomes 5. Discussion 6. Conclusion 7. Future Work References 
work_tdnnq4nh5ncl5ivyaa77h3ysoq	----	Title: A new approach to particle swarm optimization algorithm Author: Ireneusz Gościniak Citation style: Gościniak Ireneusz. (2015). A new approach to particle swarm optimization algorithm. "Expert Systems with Applications" (Vol. 42, iss. 2 (2015), s. 844-854), doi 10.1016/j.eswa.2014.07.034 [postprint] A New Approach to Particle Swarm Optimization Algorithm Ireneusz Gosciniak∗ Institute of Computer Science, University of Silesia, Bȩdzińska 39, 41-200 Sosnowiec, Poland Abstract Particularly interesting group consists of algorithms that implement co-evolution or co-operation in natural environments, giving much more powerful implemen- tations. The main aim is to obtain the algorithm which operation is not influ- enced by the environment. An unusual look at optimization algorithms made it possible to develop a new algorithm and its metaphors define for two groups of algorithms. These studies concern the particle swarm optimization algorithm as a model of predator and prey. New properties of the algorithm resulting from the co-operation mechanism that determines the operation of algorithm and significantly reduces environmental influence have been shown. Definitions of functions of behaviour scenarios give new feature of the algorithm. This feature allows self controlling the optimization process. This approach can be success- fully used in computer games. Properties of the new algorithm make it worth of interest, practical application and further research on its development. This study can be also an inspiration to search other solutions that implementing co-operation or co-evolution. Keywords: Co-evolutionary systems, PSO algorithm, predator-prey algorithm, immune algorithm, optimization method, games, artificial intelligence, entropy, multifractal analysis. ∗Corresponding author Email address: ireneusz.gosciniak@gmail.com (Ireneusz Gosciniak) Preprint submitted to Expert Systems with Applications 3.6.2014 1. Introduction Observations of systems of living organisms are inspiration for the creation of modern computational techniques. Evolution algorithms are like metaphors of biological organisms, adopting from them terminology and mechanisms of operation as well. Adaption mechanisms borrowed from biology decide on the distribution of individuals in the environment. These operators perform the functions responsible for the exploration of environment and the exploitation of areas of local extrema. Adaptation mechanisms make these algorithms more ef- ficient than a completely random search of solution space. Creating an artificial system as a metaphor or set of metaphors connected with the functioning of liv- ing organisms removes associated with it restrictions. Unfortunately, for such a system greater limits associated with its implementation are imposed. From the No Free Lunch Theorem [1] one can result that there is no universal optimiza- tion algorithm for all classes of tasks. This is a consequence of relation between the behaviour of algorithm and the problem being solved. However, it gives the inspiration to create new solutions and conducts investigations on the behaviour of the algorithm and its suitability for solving the problems of particular class. In most cases it leads to the attempts to increase the calculation efficiency by modifying the existing algorithms. Particularly interesting group consists of algorithms that implement co-evolution in natural environments because the NFL theorem cannot be applied to them. Evolutionary algorithms differ from stochastic algorithms in very efficient adaptive mechanism for searching the solution space. That is why stochastic algorithms require greater number of it- erations in the optimization process but are less likely to stop the optimization process in the local optimum. The usefulness of the algorithm is determined by the rules that are well-developed for stochastic algorithms. However, defining metaphors of natural environments for the algorithms – that is, de facto, the creation of completely new algorithms is not trivial. New algorithm must therefore be searched in the group of algorithms that implement co-evolution (cooperation) in natural environments basing on rules 2 developed for stochastic algorithms. The main aim is to obtain an algorithm on which operation the environment would have a very small impact. This feature allows controlling the optimization process and not only tuning the algorithm to the problem being solved as it is nowadays. Many problems are treated as unchangeable – they are represented by the stationary environment. However, the change in resources, tasks or other elements of the system results in the changes of problem from stationary into non-stationary – these problems are represented by the non-stationary environment. The majority of the algorithms used in non-stationary environments are adapted from algorithms applied in the stationary environments. The presented algorithm has been designed to use it in the non-stationary environments. Thus, the article presents the situation which is opposite to the mostly discussed. An unusual look at the optimization algo- rithms made it possible to develop the new algorithm and define its metaphors in two groups of algorithms. This algorithms can be used to describe artificial life. The resulting algorithms are effective optimization algorithms and the proposed approach introduces new features in their operation. The new algorithm and its metaphors in the group of immune algorithms and particle swarm optimization algorithms are presented in the article. Functions of behaviour scenarios are defined in the particle swarm optimization algorithm. New properties of the algorithm resulting from the co-operation mechanism have been shown. It de- termines the algorithm behaviour and reduces the environmental impact. This is the original and unpublished achievement of work. Immune algorithm is not widely discussed because research results were partially presented in [2] and the results of research carried out on its development require a new study. Modern PSO algorithms of high-efficiency should be classified as the hybrid algorithms. This proposed algorithm is presented rather as a base one – the base for fu- ture modifications. It was compared with different algorithms, also older (there are the references to the older literature) because presented there solutions can be considered as a base form, which is subject to further improvements. On the basis of the description it is being easy to notice the necessity to make modifications that will create a hybrid algorithm of high efficiency. 3 2. The comparison of selected algorithms The analysis of algorithms operation becomes more complicated. Modifica- tions affect many aspects of algorithm’s operation. There are lots of terms that are closely dependent on each other. Exploration and exploitation of the solu- tion space are contradictory goals. It becomes extremely difficult to maintain an appropriate balance between exploration and exploitation of the solution space during the work of an algorithm. Convergence realises the exploitation and reduces the diversity of the population. Reducing of the population diversity causes the loss of information on the solution space – the memory loss. Con- centrated individuals form a cluster. The excessive closeness of particles does not increase the information on the solution space. So in this place it should be reminded that the aim of this algorithm operation is to search for solution space to designate the global extreme or set of local extremes. There is also the effect of modifications on the algorithm operation. Frequently used modification is a mutation – but it may have the character improving the exploration or exploita- tion. Co-evolutionary systems are the most interesting. The co-evoluton can be different in each group of algorithms and the functions creating co-evolution can be diffrent in each cooprating system There is also the group of modifications basing on applying of other methods known from mathematics – implement- ing as a local method. Presentation of the structure of base algorithm will be preceded by the discussion on the selected algorithms. This discussion is very general, and it has only been used to introduce existing solutions. However, this would help to understand the concept of the new algorithm. During dislocation, population forms a compact group of individuals that exploits one part of the solution area; however, exploration is carried out by the movement of population. Phases of movement can be distinguished – in each phase the dislocation effectiveness of a population is not the same. Swarm adapts to the environment during subsequent iterations. The swarm leader rep- resents the position of the best adaptation (the best solution). The assignment of swarm particle neighbours is performed usually once at the beginning of cal- 4 culations – it makes the designation of the best adapted neighbour easier. The change in behaviour of particles swarm is a function of changes in the leader be- haviour. There are many modifications of the above-mentioned PSO algorithm – the majority of them can be found in [3], [4], [5] and [6]. The behavior of the PSO algorithm depends on the internal weights. The exploration or exploitation nature of the algorithm work depends on the inertia weight. Appropriate change in this coefficient during the algorithm work will have a significant impact on the efficiency of its work. Linear decrease in the weight factors was proposed in the work [7]. In the paper [8] the inertia weight is reducing in the course of the algorithm work. In [9] the decrease in inertia weight using fuzzy methods was proposed. In paper [10] is proposed an improved self-adaptive particle swarm optimization algorithm (ISAPSO). In the process algorithm of evolution param- eters are changed dynamically: cognitive and social learning rates parameters. It allows to maintain the diversity of the population. Control strategy has a random character which permits to take into account the various constraints of solved the problem. In [11] a local neighbourhood version is used – only the behaviour of neighbouring particles is taken into account. Keeping diversity of populations during the algorithm work is also important for PSO algorithms. The diversity of the population increases the chances of local extreme leaving. For keeping diversity of population PSO algorithm with self-organized critical- ity was introduced in [12]. In order to achieve a greater variety of particles the ”critical value” is created when two particles are too close to each other. Negative entropy was used in [13] to discourage premature convergence. The neighborhood of other particles has the impact on the behavior of par- ticles in the swarm. The neighborhood’s analysis is the basis for separation of species or the use of multi-swarm. In [14] a dynamically changing neighborhood was used. The influence of neighbors, which depends on the fitness function and the position in relation to particles, is presented in the study [15]. To achieve it, in [16] some collision-avoiding mechanisms were applied, the individuals moving was used in the study [12], whereas mutation was applied in the paper [17]. It should be here mentioned that operators typical of genetic algorithms, such as 5 mutation, crossover or selection were used in PSO [18]. In [19] the introduction of an additional repellor was suggested. It influences the swarm behaviour by directing it into the areas of the environment which have better adaptation. The use of multi-swarm allows to maintain the diversity of the particles. Algorithms creating clusters are particularly noteworthy. There are very inter- esting groups of algorithms. In multi-swarm and clusters creating systems there are key issues such as: how to define a promising area in the solution space and how to implement motion of the particles in the direction of various sub-areas, how to determine the required number of sub-swarms or clusters, and how to generate the sub- swarms or clusters. In [20] NbestPSO algorithm was proposed, which is designed to locate many solutions. Particle’s neighborhood in NbestPSO algorithm is defined as the closest particles in the population. The best neighborhood is determined on the base of the average distance of the nearest particles . In [21] NichePSO algorithm was proposed – the main swarm can create sub- swarm when the niche is identified. The criterion for a sub-swarm creation is the lack of significant changes in subsequent iterations, while the sub-swarm can absorb particles or other sub-swarms in dependance on the distance. In [22] the adaptive niching PSO algorithm (ANPSO) is proposed; the radius of the species is here determined on the base of population statistics. PSO algorithm basing on species (SPSO) is presented in [23, 24]. This algo- rithm dynamically adjusts the number and size of swarms through the creation of ordered list of particles sorted according to their fitness and grouping parti- cles of a particular species. In every generation, SPSO aims to identify many spores of species into the swarm. The particles within a radius of the spore are assigned to the same species. In the improved version of SPSO the mechanism to remove duplicates of species particles was introduced [25]. Another improved version of SPSO us- ing regression (rSPSO) was also introduced [26]. In [27] the use of k-means clustering algorithm grouping particles in the pre-defined number of clusters 6 was proposed. In order to obtain clusters stability the algorithm iterations are executed three times. To determine the number of clusters in the ”k-means PSO” algorithm in [28] particle distribution generated by a combination of sev- eral probabilistic distributions is proposed – each cluster corresponds to other distribution. Then, finding the optimal number of clusters is equivalent to find- ing the best-fit models. This algorithm gives better results for the problems described in the stationary environments than SPSO algorithms and ANPSO. Co-evolution of swarms (CESO) was proposed in [29]. In the CESO two swarms, one of which uses the differential evolution (CDE) [30], and the second one model of the PSO, co-operate with each other. The swarm, which uses the CDE is responsible for the diversity while a swarm of PSO tracks the global optimum. In Clustering PSO (CPSO) [31] every particle learns - knows its own, his- torically the best position and the best position of the nearest neighbour, which affects the motion of the particles. Using hierarchical clustering method the whole swarm in CPSO can be divided into smaller sub-swarms. It enables the adaptive detection of sub-areas. In the CPSO, hierarchical clustering method is realized by two steps: rough grouping and clusters refining. The strategy for the best global particle was also introduced in CPSO [31]. In [32] some simplifications in comparison to the original CPSO were ap- plied: the training process has been removed, hierarchical clustering method is simplified to only one phase. In [33] co-evolution of two swarms was suggested; one of them optimizes the penalty factors and the second looks for the optimum solution (swarms move in the spaces of optimum solutions for penalty factors for the contrary swarm). Another example of the cooperation system is the predator-prey model [34] which uses the model of games theory implementing war strategy. This approach was used in optimization of fuzzy clustering [35]. Article [36] describes the cooperation (co-evolution), which uses the strategy of sardines being attacked by sea wolfs (orcas). Many common features influencing the work of algorithms results from the above discussion. Mechanisms of exploitation of local extremes 7 and mechanism of exploration of space solutions depend on each other. Solutions that make these mechanisms independent on each other should be searched in cooperating or co-evolutionary systems. The operation of these sub-systems is dependent on each other, but their functions may be different. And so, the algorithm implements co-operation of two systems designating local extremes – the idea is found in [11] and [19]. The particles of both systems are directed to the same local extremes – differently than in the work [33]. The algorithm implements a game strategy based on the round-up strategy, which seems to be similar to the strategy described in [34]. The main difference in the proposed algorithm is a way of elimination of the particles, which results from the excessive proximity of particles in the same group. Examples of prevention of excessive particles closeness are above described. In many studies e.g. [5], [37] the mutation is considered to be the important coefficient in the progress of the algorithm. The proposed method reduces the range of mutations and causes adaptation based on the location of the particles. A strong mutation was replaced by the creation of random particles instead of eliminated ones. The algorithm presented below seems to be similar to the algorithm pre- sented in the work [38]. However, it is characterised by different operation mechanics. It is a predator-prey algorithm, that illustrates the example of the behavior of sardine shoals and hunting orcas. Predators are heading in cen- ter of shoals of sardines – it looks like dispersing of preys, which escape from predators. This contributes to the fact that the particles avoid of local solu- tion optimum to find the global and optimal solution. Predators play the role of exploatation (they realize convergence), while preys escaping from predators realize exploration of the solution space (they play the role of algorithm diversi- fication). The algorithm implements a scenario in which the nearest predator is elected and it is determined whether the particle being the prey escapes (which depends on the distance between predator and prey and basing on it the escape velocity is calculated). This description indicates significant differences between the algorithm [38] and the algorithm proposed in this paper. The algorithm tuning is also important. Tuning algorithm and also control 8 the operating parameters of the algorithm is also becoming an important issue. Depends on a compromise between speed and accuracy of the algorithm work - or even find a global solution. Makes it necessary to analyze the behavior of particles, their dynamics and trajectory. In [39] presented a new strategy for analyzing the behavior of the algorithm and new parameters for increasing the efficiency of the algorithm. Due to the mechanism of co-evolution and self- adaptation only large changes in parameters give a significant effect. A tuning of the algorithm is not widely discussed because it describes the behavior of the algorithm – so the importance of parameters becomes obvious. 3. The base algorithm The new algorithm implements cooperation (in evolutionary systems it is termed co-evolution) of the two systems, hereinafter referred to as elements of sample points and seed points (similarly to stochastic systems). The algorithm description is presented below: 1. Random creation of initial trial’s sample points and seed points. 2. Activation of a seed point – random choice of the seed point 3. Reduction of seed points – checking whether other seed points are present in the defined neighbourhood of the active seed point. The active seed point is the best of them. Seed points with poorer adaptation are replaced by the randomly created ones. 4. Activation of the sample points (defining the cluster) – sample points that are present in the neighbourhood of the seed point become active. 5. Reduction of the active sample points – among the active sample points, the elements which are closer than the defined distance are removed. They are replaced by randomly created sample points. 6. Processing – the replacement of active sample points elements and active seed point with the results of their processing according to the predefined rules (the execution of local method). The rule of processing is described below. 9 7. Iteration from step 2 until reaching the final stop. Behaviour of the algorithm still requires further clarification. Sample points and seed points of the initial trial are randomly created. Set of sample points is greater than set of seed points. Sample points create a kind of net, nodes of which are moving in the direction of local extrema under the influence of processing methods. The density of sample points in the areas of local extrema will be higher than in the other areas. On the basis of information about active sample points, seed points define their new positions moving also to local extrema. Speed of movement (of both active elements of sample points and seed point) depends on the density of the sample points. The speed is greater when the density is less. Thanks to this, active seed point by using the information from active trial points moves much faster outside the local extrema than inside them. It is the movement of seed point which is similar to the eye tracking. Generally, sample points are moving slower than the seed points. Sample points are responsible for the exploration of solution space, and the seed points for the exploitation of local extreme – close correlation is between these ele- ments. The reduction of seed points demonstrates the identification of local extrema as well as helps in the exploration of the environment. The reduction of sample points is the indicator of local extrema exploitation as well as it in- dicates the exploration of solution space. Sample points perform some kind of approximation of the environment function. Seed points initiate the process of optimization in the location of these points. This cooperation significantly re- duces the impact of the environment to the algorithm operation. Stop criterion is based on the monitoring of solutions generated by the algorithm in connection with information about the reduction parameters. 4. The construction of metaphors The processing method determines the behaviour of two groups of particles: one group is called sample points with function of exploration and the second one is called seed points with function of exploitation over the search space. 10 This method has been implemented as a metaphor for the immune system and the PSO system. Due to the fact that metaphors are not implemented as a canon of these algorithms, they will hereafter be referred to as Semi-Immune and Semi-PSO. The exception from the standard approach to the immune algorithm is that the antigen is represented by the seed points, not by the environment. Further- more, antigens change their position under the influence of antibodies, which are represented by the sample points. A similar situation happens in the nature when viruses mutate to defend themselves against the immune system. How- ever, the mechanism of autoimmunity is the basis for the removal of antibodies. This mechanism results in autoaggression to its own cells. On the other hand, the removal of the weakest of the antibodies that are grouped and surrounded by antigens is an interpretation of the mechanism of their reduction. The Semi-PSO algorithm implements a strategy of round-up – it seems to be a kind of predator-prey strategy – as the cooperation of two particles sys- tems: predators represented by sample points and prey – which are the seed points. Prey is encirclemented by a group of predators. Predators move in the direction of their group leader and prey. The group leader tries to cut off the escape route of prey. Prey, on the basis of both the leader and the weakest predator observations, runs towards the safe place – a local extremum. Encir- clement of more than one prey results in the elimination of the weakest of them. However, excessive approch of predators causes the elimination of the weakest of them. These mechanisms seem to be the natural elements of a struggle for survival. The principle of predator-prey algorithms described so far in other papers was different. The adaptation of weighting coefficients that create the behavioral scenarios seems to be the violation of these algorithms canon. A detailed description of the Semi-PSO algorithm is to be presented in later chap- ters. This algorithm has a high efficiency of both exploration and exploitation of the solution space. 11 5. Semi-PSO – round-up algorithm Lyrics in a playful way show the strategy of the algorithm: It is not the point to catch the bunny But to chase him ... Zieliński A., Osiecka A., Bunny, Skaldowie, Muza 1969 (in Polish, translation by author). The strategy of the algorithm supports the view that the most enjoyable thing about playing is not catching the ”bunny”, but the process of tracking it down. It demonstrates the practical importance of this strategy. The proposed algorithm shows cooperation of two systems of particles: preda- tors and their prey. Predators use the strategy of round-up – they outnumber the prey. Encirclemented by the group of predators, prey moves faster than they are. The prey by escaping from one encirclement falls into another one. Prey, based on the observation of the leader of predators and the weakest of them, is run towards the safe place that is to the local extreme. In this algorithm, neighbourhood of predators and prey must be determined in each iteration. Predators move in the direction towards the prey and the group leader who tries to cut off escape route of prey. Predators and prey are concentrated in areas of local extrema. Encirclement of more than one prey causes the elimina- tion of the weakest of them. On the other hand, the excessive rapprochement of predators causes an elimination of the weakest of them. These mechanisms seem to be the natural elements of a struggle for survival. Randomly created in- dividuals replace the eliminated ones. The discussed algorithm can be described as follows: 1. Random creation of predators (Ei = [ei1, ei2, ..., eid]) and prey (Zi = [zi1, zi2, ..., zid]) – cooperating system. 2. Activation of the prey – random choice of a particle. 3. Prey elimination – searching whether the given neighbourhood of the ac- tive prey for other individuals from their group are present; worse adapted 12 ones are replaced by randomly created new ones; prey that is left becomes the active one. 4. Activation of the predators – predators in the neighbourhood of the prey become active. 5. Predators elimination – the predators that are the weakest and are too close to the strong ones are eliminated and replaced by non-active ran- domly created ones. 6. Processing – replacement of active elements, as follows: Assessing the value of fitness function for the active predators (f (Ei)) and the selection of the best (Ei,best) and the worst (Ei,worst) of them. Calculation of the velocity vectors for predators (V Ei ) and prey (V Z i ) using the following equations: V Ei (t + 1) = w EV Ei (t) + c E 1 ϕ E 1 (Ei,best −Ei) +cE2 ϕ E 2 (Zi −Ei) , (1) V Zi (t + 1) = w ZV Zi (t) + c Z 1 ϕ Z 1 (Ei,best −Zi) +cZ2 ϕ Z 2 (Zi −Ei,worst) . (2) where: ϕE1 , ϕ E 2 , ϕ Z 1 , ϕ Z 2 – are the random values from the range [0, 1]; cE1 , c E 2 , c Z 1 , c Z 2 – are the learning rates; w E, wZ – are the particle’s inertia weights. The values of these coefficients are adapted in order to obtain the relevant scenarios of behaviour. The maximum values of the weighting coefficients are only given. Weighting coefficients have strong influence on the be- haviour of particles, so their adaptation has a significant impact on the efficiency of the algorithm work. New positions of the predators and po- sition of the prey are determined according to the following equations: Ei (t + 1) = Ei (t) + V E i (t + 1) , (3) 13 Figure 1: The algorithm diagram. 14 Zi (t + 1) = Zi (t) + V Z i (t + 1) . (4) 7. Iteration from step no. 2 until termination criterion is reached. The movement of predators depends on the position of the best predator and prey location. However, the movement of the prey depends on the position of the best and worst of predators – the movement is carried out in the direc- tion of the best predator and in the direction opposite to the worst one. This processing allows to detect the local extrema. Predators are responsible for the exploration of solution space, and the prey is responsible for exploitation of local extrema – there is strong interdependence between them. Elimination of prey demonstrates the identification of local extremum; it is also an indicator of the environment exploration. The reduction of predators is an indicator of local ex- trema exploitation and the mechanism of the solution space exploration (cluster identification). The weighting coefficients are responsible for creating scenar- ios of behaviour that are dependent on the mutual position of active elements. Predator being in close proximity to the prey tries to cut off its escape route (prey escapes in the direction of local extreme). The direction in which it moved so far has much less influence on its behaviour. Also, predator movement in the direction of prey may facilitate its escape. Therefore, in this case the weighting coefficients wE and cE2 should have less influence on predator behaviour. The cE1 weighting coefficient responsible for the movement in the direction of the leader of predators should have the dominant influence on the behaviour of the predator. In this particular situation weighting coefficients wE and cE2 have dominant influence on the movement of the leader of predators. In the case of long distance, predator moves fast in the direction of prey. Therefore, weighting coefficients wE and cE2 have the dominant influence on his behaviour. In this case the weighting coefficient cE1 would be less important in the behaviour of the predator. Scenarios of prey behaviour would also be dependent on its position in the relation to predators. In the case of bad position of prey, that is a short distance from the weakest of the predators, the weighting coefficients wZ and cZ2 15 determining the rapid escape from this location would have a significant impact on its behaviour. However, in the case of small distance from leader of predators to prey, providing also of the approach to a local extreme, weighting coefficient cZ1 would have a significant impact on its behaviour. Examples of carrying out the functions of the weighting coefficients are presented below. Example I: cE1 = ‖Ei,best −Ei‖ ‖Ei,best −Ei‖ + ‖Zi −Ei‖ · cE1,max, (5) cE2 = ‖Zi −Ei‖ ‖Ei,best −Ei‖ + ‖Zi −Ei‖ · cE2,max, (6) wE = cE2 (7) and cZ1 = ‖Ei,best −Zi‖ ‖Ei,best −Zi‖ + ‖Zi −Ei,worst‖ · cZ1,max, (8) cZ2 = ‖Zi −Ei,worst‖ ‖Ei,best −Zi‖ + ‖Zi −Ei,worst‖ · cZ2,max. (9) wZ = cZ2 (10) Example II: cE1 = min{‖Ei,best −Ei‖ ,‖Zi −Ei‖} ‖Ei,best −Ei‖ · cE1,max, (11) cE2 = min{‖Ei,best −Ei‖ ,‖Zi −Ei‖} ‖Zi −Ei‖ · cE2,max, (12) wE = min{‖Ei,best −Ei‖ ,‖Zi −Ei‖} ‖Zi −Ei‖ ·wEmax, (13) or 16 wE = ( 1 − min{‖Ei,best −Ei‖ ,‖Zi −Ei‖} ‖Ei,best −Ei‖ ) ·wEmax, (14) where wE can be 0, and cZ1 = min{‖Ei,best −Zi‖ ,‖Zi −Ei,worst‖} ‖Ei,best −Zi‖ · cZ1,max, (15) cZ2 = min{‖Ei,best −Zi‖ ,‖Zi −Ei,worst‖} ‖Zi −Ei,worst‖ · cZ2,max, (16) wZ = min{‖Ei,best −Zi‖ ,‖Zi −Ei,worst‖} ‖Zi −Ei,worst‖ ·wZmax, (17) or wZ = ( 1 − min{‖Ei,best −Zi‖ ,‖Zi −Ei,worst‖} ‖Ei,best −Zi‖ ) ·wZmax, (18) where wZ can be 0. In these expressions cE1,max, c E 2,max, w E max, c Z 1,max, c Z 2,max, w Z max are the maxi- mum values of weighting coefficients. These coefficients have the same meaning as coefficients of a typical PSO algorithm. However, the functions described above are responsible for adaptation. Consequently, small changes in the values of these coefficients will not influence visibly on the behavior of the algorithm. The principles to determine their values is the same as in PSO algorithm [4]. The algorithm can take into account the situation in which the predators are on neighbouring hunting areas, while prey can move throughout all the space. This algorithm, as the PSO algorithm, belongs to the group of stochastic algorithms. The present algorithm in a very effective way combines stochastic search of solution space with the search described by behaviour of the particles. This algorithm also has a property of observation similar to behaviour of the eye. 17 6. The analysis of Semi-PSO algorithm The evaluation of the algorithm is performed by comparing the effectiveness of the designation of global extreme of test functions in relation to certain al- gorithms of optimization. Test functions describe the stationary environment. Acronyms formed for compared algorithms are as follows: S-PSO – Semi Particle Swarm Optimization (discussed algorithm), PSO – Particle Swarm Optimiza- tion [40], RCGA – Real-Coding Genetic Algorithm [41], CGA – Continuous Genetic Algorithm [42], CHA – Continuous Hybrid Algorithm [43], ECTS – Enhanced Continuous Taboo Search [44], CTSS – Continuous Taboo Simplex Search [45], SEA – Simplex and Evolution Algorithm [46]. Acronyms formed for test functions are as follows: EM – Easom, FR – Fichier, SH – Shubert, G&P – Goldstein and Price, ZV – Zakharov (for 2, 5 and 10 Dimensions), RK – Rosenbrock (for 2, 5 and 10 Dimensions), BN RC – Branin RCOS, DJ – De Joung, S4,n – Shekel (for n = 7, 10). The tolerance radius of the preys for functions: DJ , S4,7 and S4,10 is twice times higher and for functions: ZV and RK for 10D it is five times higher than the others. The result of tolerance area increase is the decrease in precision of extreme determining. Predators’ tolerance areas set in the algorithm are the same except functions of the 10D spaces. For these functions, tolerance areas of predators and preys are the same. For functions DJ, S4,7 and S4,10 the tolerance radius of the preys is twice smaller than the tolerance radius of predators, while in the remaining cases it is four times smaller. For the data presented in tables and figures 2 to 5 the number of predators amounts to 100 and the number of preys – 10, whereas for figures 7 to 9 the number of predators amounts to 20 and the number of preys – 6. This allows to compare behavior of the algorithm with the same settings in different test environments. The consequence of this approach is suboptimal efficiency of the algorithm – despite it, the efficiency of the algorithm is satisfactory, as shown by the results contained in table 2. Operating parameters of the algorithm for these functions were set at the same value and it influenced on the results obtained. But, it allows the identification 18 of interesting properties illustrated by the results collected in table 4. Table 1 summarizes the number of iterations needed to achieve 100% success rate. Only for SEA algorithm and test function Easom, Zakharov and Rosen- brock succes rate is 97%. The presented algorithm is compared with seven optimization algorithms for fourteen test functions. Such type of analysis is the basis for the majority of articles and it allows the comparison of the algorithm effectiveness, yet its behavior is not shown. The mechanism of behavior seems to be important when choosing an algorithm for the problem. In an indirect way such information is obtained using various test functions. Knowledge of the behavior of the algorithm may be important in solving untypical problems. The success rate determines the ability to solve the problem. The number of iterations of the algorithm allowing to achieve success determines its cost. The study (table 1) shows that the S-PSO is better than PSO, RCG and CGA, but worse than SHI, CTS and SEA PSO. The proposed algorithm is better realized for the functions: DJ, S4,7, S4,10, ZV and RK for 5D and 10D. In order to illustrate the algorithm better, table 1 can be supplemented by data contained in table 2. Figures in table 2 show a large variation – it results from the algorithm operation. The initial distribution will decide of the number of iterations required to achieve the success. Figure 2 shows the history of global extreme designation for the G&P func- tion. The tolerance radius is lower by 5% in the case of characteristic b. These characteristics show the distinct effect of the tolerance area on the precision of extreme designation. The extreme designation with less precision is significantly faster as well. It is obvious because the time and precision of global extreme designation are conflicting goals. The tolerance area allows to determine which of these goals will be realised. The analogous graphs are presented for function RK for 2D (figure 3) – they are the typical graphs illustrating the convergence of the algorithm. The history of changes in the standard deviation of global extreme designa- tion is shown in figure 4. This graph shows that for all tests the error of local extreme designation decreases in time. However, this increase in precision of ex- 19 Table 1: The comparison of the number of iterations of the algorithm (S-PSO) with other algorithms for the selected test function. Algorithm S-PSO∗ PSO RCG CGA ECTS CHI CTS SEA Test fun. EM 653 740 642 1504 1284 952 325 197∗∗ FR 307 580 – 430 – 132 98 258 SH 721 800 946 575 370 345 283 420 G&P 425 480 270 410 231 259 119 – ZV 2D 613 380 437 620 195 215 78 90∗∗ RK 2D 497 1660 596 960 480 459 369 266∗∗ BN RC 433 740 490 620 254 295 125 272 DJ 358 500 449 750 338 371 155 – S4,7 621 29180 1143 680 910 620 590 – S4,10 768 30160 1235 650 898 635 555 – ZV 5D 1066 1530 1115 1350 2254 950 – – ZV 10D 2764 7440 2190 6991 4630 100 – – RK 5D 2423 33100 4150 3990 2142 3290 – – RK 10D 6340 106960 8100 21563 15720 14563 – – * – proposed algorithm, ** – success rate reaches 97%. Figure 2: The history of global extreme designation for the G&P function. 20 Table 2: Statistical data of numbers of iterations Param. numbers of iterations of the algorithm Test fun. min. avg. max. med. std. dev. EM 260 653 985 698 260 FR 108 307 684 263 195 SH 306 721 984 778 191 G&P 149 425 683 447 207 ZV 2D 254 613 987 582 273 RK 2D 110 497 754 640 263 BN RC 75 433 732 492 218 DJ 115 358 736 243 242 S4,7 380 621 853 612 169 S4,10 474 768 953 852 191 ZV 5D 746 1066 1437 1017 269 ZV 10D 1588 2764 3806 2805 899 RK 5D 1762 2423 3487 2439 562 RK 10D 4527 6340 7848 6159 1172 Figure 3: The history of global extreme designation for the RK 2D function. 21 Figure 4: The history of changes in the standard deviation of global extreme designation for FR function. Figure 5: The history of global extreme designation for the ZV 2D function. treme designation is limited by the existence of the tolerance area. It illustrates the convergence of the algorithm in the global extreme designation. Figure 5 shows the history of the global extreme designation. The tolerance radius decreases by 5% for the following characteristics. Characteristics ”c” has the smallest tolerance area and simultaneously – the smallest error. It is worth noticing that the extreme designation in all cases is very fast. Because the error value of extreme designation depends on the selection of the value of tolerance area, the implementation of the adaptation mechanism will reduce the error significantly, as it was confirmed by tests carried out (table 3). Fitness function is always designated for active predators and prey, as well as for newly created particles which replace the eliminated predators and prey. In table 4 average values of prey and predators elimination are given, as well as 22 Table 3: Statistical data of error of extremes designation Param. Error of extremes designation Test fun. min. avg. max. med. std. dev. EM 2,0E-05 2,5E-04 5,4E-04 2,1E-04 1,56E-04 FR 7,34E-03 8,29E-02 1,88E-01 8,3E-02 5,56E-02 SH 1,31E-02 6,07E-01 1,84E+00 5,01E-01 6,21E-01 G & P 1,55E-03 2,52E-02 5,17E-02 2,51E-02 1,8E-02 ZV 2D 6,0E-05 2,59E-04 7,7E-04 1,3E-04 2,56E-04 RK 2D 1,8E-04 2,54E-02 5,92E-02 1,7E-02 2,43E-02 BN RC 2,0E-04 3,7E-03 8,85E-03 1,99E-03 3,72E-03 DJ 7,0E-04 2,95E-02 5,64E-02 3,23E-02 1,63E-02 S4,7 5,04E-03 2,44E-02 4,84E-02 2,43E-02 1,65E-02 S4,10 1,88E-03 2,82E-02 6,41E-02 2,43E-02 2,0E-02 ZV 5D 2,16E-01 5,31E-01 9,38E-01 4,71E-01 2,68E-01 ZV 10D 1,0E+00 2,49E+00 3,99E+00 2,51E+00 1,07E+00 RK 5D 6,23E-01 9,95E-01 3,99E+00 9,32E-01 2,6E-01 RK 10D 8,61E+00 1,48E+01 1,95E+01 1,57E+01 3,97E+00 23 Table 4: Data of particles’ activity. Param. Prey Predators Active Test fun. elimination∗ elimination∗ predators∗ EM 0,03 6,12 4,76 FR 0,05 3,78 4,70 SH 0,06 10,02 5,31 G&P 0,03 3,06 4,24 ZV 2D 0,05 3,44 4,40 RK 2D 0,03 3,08 4,70 BN RC 0,04 3,51 4,23 DJ 0,02 1,18 3,65 S4,7 0,01 1,03 4,26 S4,10 0,01 1,09 4,27 ZV 5D 0,02 1,35 3,89 ZV 10D 0,02 1,14 3,73 RK 5D 0,04 3,15 6,00 RK 10D 0,004 0,47 4,21 * – per one itreration the average valuess of active predators referenced to one cycle of the algorithm. It is easy to notice that the average values of predators and prey elimination and active predators computed for one work cycle of algorithm are dependent on the parameters of the algorithm work and the environment has little effect on them. This feature gives new possibilities in the application of the algorithm. To evaluate the efficiency of the algorithm better, the data of algorithm operation are given in table 5. Using data from tables 1 and 4, the average cost of work per one algorithm iteration was determined and then the average cost of the designation of extreme was calculated, which is expressed as the average number of the fitness function computing. The cost of the algorithm work is clearly dependent on the environment and 24 Table 5: Cost of extremes designation. Param. avg.∗ min. avg. max. Test fun. EM 11,91 3095 7772 11727 FR 9,53 1030 2927 6521 SH 16,40 5017 11813 16133 G&P 8,33 1241 3543 5690 ZV 2D 8,89 2257 5443 8771 RK 2D 8,81 969 4378 6643 BN RC 8,77 658 3800 6420 DJ 5,86 674 2099 4312 S4,7 6,31 2397 3915 5380 S4,10 6,38 3023 4895 6078 ZV 5D 6,26 4672 6675 9000 ZV 10D 5,90 9361 16293 22437 RK 5D 10,19 17963 24706 35549 RK 10D 5,68 25730 36034 44605 * – per one itreration 25 Table 6: Data for comparation of extremes designation error. Algorithm SPSO Modified PSO Uniform Design PSO Test fun. SH 6,2059E-01 5,8247E-14 2,3437E-14 BN RC 3,7222E-03 0,0E+00 0,0E+00 RK 10D 3,9692E+00 6,2436E-01 1,1308E+00 it increases with the complexity of the environment. The number of particles equals to 110, but only (on average) less than 8% is involved in the calculation of one cycle and only these particles are subject to adaptation mechanism. The remaining particles fulfill also a very important function – they contain information on the solution space. It is of great importance for the proper operation of the algorithm (see the description of the algorithm). Discussions on precision of extreme designation can be extended of the ex- emplary comparison with the results obtained in [4] (table 6). The proposed algorithm compared with algorithms tuned for high precision of extreme desig- nation obtains worse results. In the article [47] the test environments were divided into four groups with the following test functions: Group A: Unimodal and Simple Multimodal Prob- lems: 1) Sphere function, 2) Rosenbrock’s function, Group B: Unrotated Multi- modal Problems: 3) Ackley’s function, 4) Griewank’s function, 5) Weierstras’s function, 6) Rastrigin’s function 7) Non-continuous Rastrigin’s function, 8) Schwefel’s function. Group C: Rotated Multimodal Problems: 9) Rotated Ack- ley’s function, 10) Rotated Griewank’s function, 11) Rotated Weierstras’s func- tion, 12) Rotated Rastrigin’s function, 14) Rotated Schwefel’s function Group D: Composition Problems: 15) Composition function 1 (CF1) in [48]: (CF1), 16) Composition function 5 (CF5) in [48]: (CF2). By means of the above- mentioned test environments the efficiency of the following algorithms group was verified : PSO with inertia weight (PSO-w) [7]; PSO with constriction factor (PSO-cf) [49]; Local version of PSO with inertia weight (PSO-w-local); 26 Figure 6: The global extreme designation error by S-PSO and the standard deviation for groups of test functions. Local version of PSO with constriction factor (PSO-cflocal) [50]; Unified Par- ticle Swarm Optimization (UPSO) [51]; Fully informed particle swarm (FIPS) [15]; Fitness Distance Ratio based on Particle Swarm Optimization (FDR-PSO) [10]; Cooperative PSO (CPSO-H) [39]; Cellular PSO (CLPSO) [47]. For the same main settings, the effectiveness efficiency of the proposed algo- rithm operation was referred to the range of values obtained in the particular groups of test functions. As it results from the presented graphs the obtained values are located within the ranges presented in the graphs. For a better graphic presentation ranges in the figures are limited – the minimum values for the figure 6a, b reach 9.84E − 118 ± 3.56E − 117 , whereas for figure 6c, d – 1.16E − 113 ± 2.92E − 113. Effective criterion for stopping the algorithm work is difficult to achieve. It is shown among others in the paper [6]. In the proposed algorithm the removal of preys is responsible for the extreme identifying (see the description of the algorithm). These removals are repeated in the same areas of the solution space. These eliminations combined with slight changes in fitness function are certain indicators of the extreme. The obtained values of algorithm cost comply with such criterion (table 7). These results can be compared with the average values 27 Figure 7: The distribution of predators in the environment (a, b) and the multifractal spectrum of the presented distributions(c, d). calculated on the base of data for the various ”stop criteria” and presented in the study [6]. The numbers in parentheses denote the fraction of runs that located the global extreme. On this basis, it is possible to conclude that the computation cost of the proposed algorithm is higher in the majority of cases – but it gives the certainty of global extreme designation. The multifractal analysis may be used for evaluating the algorithm work. Obtained multifractal spectrum gives information about the efficiency of search- ing the solution space by predators as well as the behaviour of the prey. Figure 7a and b shows the distribution of predators in the environment. They perform a uniform exploration of the solution space in the search of prey. It should be noted that the distribution of predators in a single iteration is not uniform, because the predators encircle the prey. A comparison of figures 7a and b indicate that the distribution of predators in figure 7a is more uniform. It is confirmed by the graph of multfractal analysis where the multifractal spectrum from the figure 7d is relevant to the distribution of predators in figure 7b. A similar analysis can be performed by observing the behavior of prey. Figure 8 clearly shows prey grouped in local extremes. Clustering of prey is also illustrated by the multifractal spectrum. The graph of multifractal anal- 28 Table 7: Data of the algorithm work. Param. SPSO LDW CENTER SIMPLE DYNAMIC Test fun. PSO PSO PSO PSO min. 3095 807 813 793 806 EM avg. 7772 5141 5196 4245 3985 max. 11727 17999 18205 14478 13446 min. 5017 2512 2535 1794 1805 (0,94) (0,94) (0,98) (0,99) SH avg. 11813 5778 5890 3528 2940 (0,96) (0,96) (0,99) max. 16133 9958 10224 5838 4510 min. 658 942 949 856 902 (0,96) (0,96) (0,99) (0,94) BN RC avg. 3800 4002 4043 1766 1658 (0,97) (0,97) (0,98) max. 6420 8496 8698 3088 2664 min. 2397 3516 3318 1613 1885 (0,51) (0,52) (0,48) (0,45) S4,7 avg. 3915 8170 8379 2648 2380 (0,62) (0,63) (0,54) (0,54) max. 5380 12845 13337 5283 3703 (0,82) (0,81) (0,72) (0,73) min. 3023 4342 4239 1726 1768 (0,62) (0,63) (0,49) (0,45) S4,10 avg. 4895 9107 9319 2729 2313 (0,71) (0,71) (0,56) (0,54) max. 6078 14685 15186 5328 3703 (0,93) (0,92) (0,78) (0,79) 29 Figure 8: The distribution of prey in the environment (a, b) and multifractal spectrum of the presented distribution (c, d). ysis 8d and distribution in figure 8b confirm the clustering. What is more, the multifractal analysis may be particularly useful in the research on multi- dimensional problems. Data from this analysis can be useful in the tuning of the algorithm and also for the automatization of algorithm tuning process while working. Since the experiments on the automatic tuning of algorithm work have not been conducted, it is suggested to start further research. Figure 9 illustrates the behavior of encircled prey. The increase of prey speed movement caused by the increase of maximum value of the weighting coefficients is clearly visible in the following figures: 9a, b, c, d. In figure 9a the prey cannot escape from the encirclement, as evidences the strong exploitation nature of the algorithm (does not mean algorithm stag- nation due to the mechanism of elimination). Whereas figures 9b, c, d illustrate the successive encirclements and escapes of prey. Figures 9b and c show a good balance between exploration and exploitation. Figure 9b shows more intensive exploitation, while the figure 9c shows more intensive exploration. Strong ex- ploratory character of algorithm is shown in figure 9d. The behavior of prey in figure 9c is the best. It moves very fast outside the area of extreme and it follows to the next extreme. Prey moves much slower inside extreme performing 30 Figure 9: Sample path of encircled prey movement. its exploitation. This behavior is also an illustration of the process of observa- tion. To obtain information about the image, the human eye moves in a way passing through various points of the image focusing on the details. This sug- gests that eye movement has an exploration phase and an exploitation phase. Moving prey creates an association with this movement. For figure 9a it can be concluded that gaze is fixed on one point and for figure 9d that gaze is restless (glassy eyed). While on the figures 9b and c, gaze is moving through the image by focusing on some detail. We experience the power of the sense of sight daily. Thanks to these properties, the algorithm should use high efficiency also in the non-stationary environments. 7. Future work On the base of above-presented discussion it is easy to indicate the poten- tial modifications of algorithm that should improve its efficiency and retain its operation mechanics. The first modification can realised as a local method for a designated cluster by predators. This method can be executed after the elim- ination of the prey. When such a cluster determines the local extreme, another modification can be proposed – particles are discouraged to move in direction 31 of this area. There are also possible other modifications as the adaptive change of tolerance area or changes in the behavior of scenarios. One should assume that the presented algorithm in the hybrid form meets the high requirements of efficiency. However, the properties and the possibility to apply this algorithm in non-stationary environments are currently the subject of a separate study. 8. Conclusion The new algorithm is proposed in the article. Its principle of operation is based on the cooperation of two systems of particles. Semi-PSO algorithm de- scribes a situation taken from life – a round-up game strategy. The presented algorithm is not described so far in the world literature and they have features similar to the mechanism of observation. They have high efficiency of random search of solutions space combined with searches resulting from the motion of particles. The first particle system has the exploration function, and the other one – exploitation function. This classification is conventional and results from analysis of the behavior of particle systems. As it has been shown, coopera- tion of the particles have a strong influence on the behavior of the algorithm – stronger than the impact of the environment. In the S-PSO algorithm behav- ioral scenarios are applied, that are implemented by functions of the weighting coefficients. Parameters of the algorithm, discussed in the article, give interest- ing ability to control its operation in comparison with other algorithms. This algorithm has the self-control ability between the process of exploitation and exploration of the solution space. This also applies to the stop criterion, which for this group of algorithms is determined on the basis of their behavior – the extreme is indicated by a prey elimination. The comparison S-PSO algorithm with other algorithms shows its good qualities. The present form of the algo- rithm has very interesting properties. However, the algorithm in its present form cannot be of practical importance. The reason for this is a rather big error of extreme designation. The future works should be focused on the modifica- tion introducing the effective exploitation method activated by the occurrence 32 of prey elimination. Creating of such a hybrid algorithm can significantly accel- erate and improve the accuracy of extreme designation. Major weakness of the proposed algorithm for the stationary environments is a repeated designation of the same extreme – it can be also eliminated. Properties of the new algorithm makes it worthy of interest, the practical ap- plication and work on its development. This study can be also as an inspiration to search other solutions that implement co-operation or co-evolution. References [1] D. Wolpert, W. Macready, No free lunch theorems for optimization, IEEE Transaction on Evolutionary Computation 1 (1) (1997) 67–82. [2] I. Gosciniak, Immune algorithm in non-stationary optimization task, in: Proceedings of the 2008 International Conference on Computational Intel- ligence for Modelling Control & Automation, CIMCA ’08, IEEE Computer Society, Washington, DC, USA, 2008, pp. 750–755. [3] D. Sedighizadeh, E. Masehian, Particle swarm optimization methods, tax- onomy and applications, International Journal of Computer Theory and Engineering 1 (5) (2009) 1793–8201. [4] T. Chen, T. Chi, On the improvements of the particle swarm optimization algorithm, Advances in Engineering Software 41 (2) (2010) 229–239. [5] H. Gao, W. Xu, Particle swarm algorithm with hybrid mutation strategy, Applied Soft Computing 11 (8) (2011) 5129–5142. [6] I. Tsoulos, A. Stavrakoudis, Enhancing PSO methods for global optimiza- tion, Applied Mathematics and Computation 216 (10) (2010) 2988–3001. [7] Y. Shi, R. Eberhart, A Modified Particle Swarm Optimizer, in: Proceed- ings of IEEE International Conference on Evolutionary Computation, IEEE Computer Society, Washington, DC, USA, 1998, pp. 69–73. 33 [8] H. Fan, Y. Shi, Study on vmax of particle swarm optimization, in: Proc. Workshop Particle Swarm Optimization, Indianapolis, 2001. [9] Y. Shi, R. Eberhart, Particle swarm optimization with fuzzy adaptive iner- tia weight, in: Proc. Workshop Particle Swarm Optimization, Indianapolis, 2001, pp. 101–106. [10] T. Peram, K. Veeramachaneni, C. Mohan, Fitness-distance-ratio based par- ticle swarm optimization, in: Swarm Intelligence Symp., 2003, pp. 174–181. [11] X. Hu, R. Eberhart, Y. Shi, Engineering optimization with particle swarm, in: IEEE Swarm Intelligence Symposium, SIS 2003, IEEE Neural Networks Society, Indianapolis, 2003, pp. 53–57. [12] M. Lovbjerg, T. Krink, Extending particle swarm optimizers with self- organized criticality, in: Proc. Congr. Evol. Comput., Honolulu, 2002, pp. 1588–1593. [13] X. Xie, W. Zhang, Z. Yang, Dissipative particle swarm optimization, in: Proc. Congr. Evol. Comput., 2002, pp. 1456–1461. [14] X. Hu, R. Eberhart, Multiobjective optimization using dynamic neighbor- hood particle swarm optimization, in: Proceedings of the Evolutionary Computation on 2002. CEC ’02. Proceedings of the 2002 Congress - Vol- ume 02, CEC ’02, IEEE Computer Society, Washington, DC, USA, 2002, pp. 1677–1681. [15] R. Mendes, J. Kennedy, J. Neves, The fully informed particle swarm: Sim- pler, maybe better., IEEE Transaction on Evolutionary Computation 8 (3) (2004) 204–210. [16] T. Blackwell, P. Bentley, Dont push me! collision-avoiding swarms, in: Proc. IEEE Congr. Evol. Comput., Honolulu, 2002, pp. 1691–1696. [17] V. Miranda, N. Fonseca, New evolutionary particle swarm algo- rithm (epso) applied to voltage/var control, in: Proc. 14th Power 34 Syst. Comput. Conf., Seville, Spain, 2002, p. ”[Online] Available: http://www.pscc02.org/papers/s21pos.pdf”. [18] P. Angeline, Using selection to improve particle swarm optimization, in: Proc. IEEE Congr. Evol. Comput., Anchorage, 1998, pp. 84–89. [19] A. Leontitsis, D. Kontogiorgos, J. Pange, Repel the swarm to the optimum, Applied Mathematics and Computation 173 (1) (2006) 265–272. [20] R. Brits, F. Engelbrecht, A.P. andvan den Bergh, Solving systems of un- constrained equations using particle swarm optimization, in: Proc. 2002 IEEE Conf. Syst., Man, Cybern., 2002, pp. 102–107. [21] R. Brits, A. Engelbrecht, F. van den Bergh, A niching particle swarm opti- mizer, in: Proc. 4th Asia-Pacific Conf. Simulated Evolution and Learning, 2002, pp. 692–696. [22] S. Bird, X. Li, Adaptively choosing niching parameters in a pso, in: Proc. 2006 Genetic Evol. Comput. Conf., 2006, pp. 3–10. [23] X. Li, Adaptively choosing neighborhood bests using species in a parti- cle swarm optimizer for multimodal function optimization, in: Proc. 2004 Genetic Evol. Comput. Conf., 2004, pp. 105–116. [24] D. Parrott, X. Li, A particle swarm model for tracking multiple peaks in a dynamic environment using speciation, in: Proc. 2004 Congr. Evol. Comput., 2004, pp. 98–103. [25] D. Parrott, X. Li, Locating and tracking multiple dynamic optima by a particle swarm model using speciation, in: IEEE Trans. on Evol. Comput., Vol. 10, 2006, pp. 440–458. [26] S. Bird, X. Li, Using regression to improve local convergence, in: Proc. 2007 IEEE Congr. Evol. Comput., 2007, pp. 592–599. 35 [27] J. Kennedy, Stereotyping: Improving particle swarm performance with cluster analysis, in: Proc. 2000 Congr. Evol. Comput., 2000, pp. 1507– 1512. [28] A. Passaroand, A. Starita, Particle swarm optimization for multimodal functions: A clustering approach, Journal of Artificial Evolution and Ap- plications 2008 (Article ID 482032) (2008) 15. [29] R. Lung, D. Dumitrescu, A collaborative model for tracking optima in dynamic environments, in: Proc. 2007 Congr. Evol. Comput., 2007, pp. 564–567. [30] R. Thomsen, Multimodal optimization using crowding-based differential evolution, in: Proc. 2004 Congr. Evol. Comput., 2004, pp. 1382–1389. [31] C. Li, S. Yang, A clustering particle swarm optimizer for dynamic opti- mization, in: Proc. 2009 Congr. Evol. Comput., 2009, pp. 439–446. [32] S. Yang, C. Li, A clustering particle swarm optimizer for locating and tracking multiple optima in dynamic environments, in: IEEE Trans. on Evol. Comput., Vol. 14, 2010, pp. 959–974. [33] Q. He, L. Wang, An Effective Co-evolutionary Particle Swarm Optimization For Constrained Engineering Design Problems, Engineering Applications of Artificial Intelligence 20 (1) (2007) 89–99. [34] J. Arquilla, D. Ronfeldt, Swarming and the Future of Conflict, RAND National Defense Research Institute, Santa Monica, CA, U. S., 2000. [35] W. Jang, H. Kang, B. Lee, K. Kim, D. Shin, S. Kim, Optimized fuzzy clustering by predator prey particle swarm optimization, in: IEEE Congress on Evolutionary Computation, CEC2007, 2007, pp. 3232–3238. [36] M. Higashitani, A. Ishigame, K. Yasuda, Pursuit-escape particle swarm optimization, IEEJ Transactions on Electrical and Electronic Engineering 3 (1) (2008) 136–142. 36 [37] K. Trojanowski, S. Wierzchoń, Immune-based algorithms for dynamic op- timization, Information Sciences 179 (10) (2009) 1495–1515. [38] M. Higashitani, A. Ishigame, K. Yasuda, Particle swarm optimization con- sidering the concept of predator-prey behavior, in: 2006 IEEE Congress on Evolutionary Computation, 2006, pp. 434–437. [39] F. van den Bergh, A. Engelbrecht, A cooperative approach to particle swarm optimization, IEEE Transactions on Evolutionary Computation 8 (2004) 225–239. [40] H. Kuo, J. Chang, C. Liu, Particle swarm optimization for global optimiza- tion problems, Journal of Marine Science and Technology 14 (3) (2006) 170–181. [41] M. Bessaou, P. Siarry, A genetic algorithm with real-value coding to opti- mize multimodal continuous functions, Structural and Multidiscipline Op- timization 23 (2001) 63–74. [42] R. Chelouah, P. Siarry, A continuous genetic algorithm designed for the global optimization of multimodal functions, Journal of Heuristics 6 (2) (2000) 191–213. [43] R. Chelouah, P. Siarry, Genetic and nelder-mead algorithms hybridized for a more accurate global optimization of continuous multiminima function, European Journal of Operational Research 148 (2) (2003) 335–348. [44] R. Chelouah, P. Siarry, Tabu search applied to global optimization, Euro- pean Journal of Operational Research 123 (2000) 256–270. [45] R. Chelouah, P. Siarry, A hybrid method combining continuous taboo search and nelder-mead simplex algorithms for the global optimization of multiminima functions, European Journal of Operational Research 161 (2005) 636–654. 37 [46] H. Kuo, J. Chang, K. Shyu, A hybrid algorithm of evolution and simplex methods applied to global optimization, Journal of Marine Science and Technology 12 (4) (2004) 280–289. [47] J. Liang, A. Qin, P. Suganthan, S. Baskar, Comprehensive learning particle swarm optimizer for global optimization of multimodal functions, IEEE Transactions on Evolutionary Computation 10 (3) (2006) 281–295. [48] J. Liang, P. Suganthan, K. Deb, Novel composition test functions for nu- merical global optimization, in: Proc. Swarm Intell. Symp., 2005, pp. ”[On- line], Available: http://www.ntu.edu.sg/home/EP–NSugan”. [49] M. Clerc, J. Kennedy, The particle swarm-explosion, stability, and conver- gence in a multidimensional complex space, IEEE Transactions on Evolu- tionary Computation 6 (1) (2002) 58–73. [50] J. Kennedy, R. Mendes, Population structure and particle swarm perfor- mance, in: IEEE Congr. Evol. Comput., 2002, pp. 1671–1676. [51] K. Parsopoulos, M. Vrahatis, Upsoa unified particle swarm optimization scheme, Lect. Ser. on Computational Sciences (2004) 868–873. 38 Gosciniak3 Gosciniak3 
work_tdsdaz6vbfaqblkxi3toiegswq	----	A genetic fuzzy expert system for automatic question classification in a competitive learning environment Expert Systems with Applications 39 (2012) 7471–7478 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A genetic fuzzy expert system for automatic question classification in a competitive learning environment Elena Verdú, María J. Verdú ⇑, Luisa M. Regueras, Juan P. de Castro, Ricardo García School of Telecommunications Engineering, University of Valladolid, Paseo Belén, 15, 47011 Valladolid, Spain a r t i c l e i n f o a b s t r a c t Keywords: Intelligent tutoring systems Educational technology Automatic question classification Competitive learning Genetic algorithms Fuzzy systems 0957-4174/$ - see front matter � 2012 Elsevier Ltd. A doi:10.1016/j.eswa.2012.01.115 ⇑ Corresponding author. Address: ETSI Telecomunic Valladolid, Spain. Tel.: +34 983423707; fax: +34 9834 E-mail addresses: elever@tel.uva.es (E. Verdú), ma luireg@tel.uva.es (L.M. Regueras), jpdecastro@tel.uva. uva.es (R. García). Intelligent tutoring systems are efficient tools to automatically adapt the learning process to the student’s progress and needs. One of the possible adaptations is to apply an adaptive question sequencing system, which matches the difficulty of the questions to the student’s knowledge level. In this context, it is impor- tant to correctly classify the questions to be presented to students according to their difficulty level. Many systems have been developed for estimating the difficulty of questions. However the variety in the appli- cation environments makes difficult to apply the existing solutions directly to other applications. There- fore, a specific solution has been designed in order to determine the difficulty level of open questions in an automatic and objective way. This solution can be applied to activities with special temporal and run- ning features, as the contests developed through QUESTOURnament, which is a tool integrated into the e- learning platform Moodle. The proposed solution is a fuzzy expert system that uses a genetic algorithm in order to characterize each difficulty level. From the output of the algorithm, it defines the fuzzy rules that are used to classify the questions. Data registered from a competitive activity in a Telecommunications Engineering course have been used in order to validate the system against a group of experts. Results show that the system performs successfully. Therefore, it can be concluded that the system is able to do the questions classification labour in a competitive learning environment. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction interest, motivation and engagement by arousing their competitive in- During the last years, the learning process is changing substan- tially in order to be centred on the students and adapted to their needs and features. Different studies have shown the effectiveness of the new adaptive learning systems (Verdú, Regueras, Verdú, de Castro, & Pérez, 2008). Many of these systems attempt to be more adaptive by offering students questions with difficulty levels according to their skills and capabilities. The aim is to increase the efficiency and the level of interaction and motivation of stu- dents (Lilley, Barker, & Britton, 2004). Too difficult or too easy questions can frustrate and decrease students’ motivation, while adaptive question sequencing provides a more efficient and effec- tive learning (Wauters, Desmet, & Van den Noortgate, 2010). More- over, according to (Lee & Heyworth, 2000), students should be able to score higher if the items or problems are arranged according to their difficulty level, since after solving easier problems, they feel more motivated to solve the harder ones. On the other hand, the competitive learning systems, as the QUES- TOURnament system, are an effective technique to capture students’ ll rights reserved. ación, Paseo Belén 15, 47011 23667. rver@tel.uva.es (M.J. Verdú), es (J.P. de Castro), ricgar@tel. stincts (Anderson, 2006; Philpot, Hall, Hubing & Flori, 2005). Moreover, competitive learning reduces procrastination, a common cause for stu- dents failing to complete assignments (Lawrence, 2004) and improves the learning process (Regueras et al., 2009). QUESTOURnament is a telematic tool integrated into the e- learning platform Moodle that allows teachers to organize dynamic contests in any knowledge domain (Regueras et al., 2009). Students compete for getting the highest marks and being at the top in the ranking. They must solve exercises (known as challenges in QUES- TOURnament) within a time limit and as soon as possible, since the scoring function varies with time. The competitive nature of QUESTOURnament motivates stu- dents but also can provoke stress and discouragement in the worst classified students. To assign the adequate opponents and ques- tions to a student may be an effective strategy to reduce these neg- ative effects (Wu et al., 2007). Therefore the system should group students by knowledge level so that students with similar skills compete together and answer questions with a difficulty level suit- able for them. In this context, it is very important to correctly classify ques- tions by difficulty level. However, it is difficult for teachers to accu- rately estimate the difficulty level according to the students’ level of competence (Watering & Rijt, 2006). Experience helps teachers to better estimate the difficulty level of the questions, but even senior teachers sometimes fail and have to rectify when they http://dx.doi.org/10.1016/j.eswa.2012.01.115 mailto:<xml_chg_old>elever.@tel.uva.es</xml_chg_old><xml_chg_new>elever@tel.uva.es</xml_chg_new> mailto:<xml_chg_old>marver.@tel.uva.es</xml_chg_old><xml_chg_new>marver@tel.uva.es</xml_chg_new> mailto:<xml_chg_old>luireg.@tel.uva.es</xml_chg_old><xml_chg_new>luireg@tel.uva.es</xml_chg_new> mailto:<xml_chg_old>jpdecastro.@tel.uva.es</xml_chg_old><xml_chg_new>jpdecastro@tel.uva.es</xml_chg_new> mailto:<xml_chg_old>ricgar.@tel.uva.es</xml_chg_old><xml_chg_new>ricgar@tel. uva.es</xml_chg_new> mailto:<xml_chg_old>ricgar.@tel.uva.es</xml_chg_old><xml_chg_new>ricgar@tel. uva.es</xml_chg_new> http://dx.doi.org/10.1016/j.eswa.2012.01.115 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 7472 E. Verdú et al. / Expert Systems with Applications 39 (2012) 7471–7478 analyze the answers given by their students. An automatic estima- tion system could be the basis for an effective adaptation process. A lot of systems that automatically estimate the difficulty level of items can be found in the literature (Burghof, 2001; Cheng, Shen, & Basu, 2008; Jong, Chan, Wu, & Lin, 2006; Lee, 1996; Wauters et al., 2010). However, the variety in the nature of the application environments makes difficult to apply the existing solutions di- rectly to other applications. Therefore, a specific solution has been designed in order to turn the competitive e-learning system QUES- TOURnament into an intelligent system. The objective is to make learning more effective and to mitigate some of the practical draw- backs of competitive learning. This paper discusses the validity of an expert system that auto- matically estimates the difficulty level of the questions posed in the QUESTOURnament competitive learning system. Section 2 introduces the major issues about teachers’ perception of difficulty and summarizes the search towards the solution. The expert sys- tem is described in Section 3. Section 4 starts with a description of the experiment developed in order to validate the system. Next, a study that analyzes the accuracy of the estimations of difficulty obtained by the intelligent system is presented. Finally, the main conclusions are stated. 2. Background 2.1. Teachers’ perception of difficulty The correct estimation of the difficulty level of learning material (questions, items. . .) is very important in the design and definition of assessment processes, adaptive learning systems or standard setting methods. However, there are not too many studies about the perception and estimation of difficulty level by teachers. Estimating the difficulty level of questions is not an easy job. Several studies (Alexandrou-Leonidou & Philippou, 2005; Hadjidemetriou & Williams 2002; Lee & Heyworth, 2000; Watering & Rijt, 2006) question the ability of teachers to make accurate diffi- culty level estimations of learning material since teachers usually fail to identify the correct difficulty level according to the students’ ability. In general terms, students’ performance tends to be overestimated by teachers (Goodwin, 1999; Impara & Plake, 1998; Verhoeven, Verwijnen, Muijtjens, Scherpbier, & Van der Vleuten, 2002). Moreover, according to Watering and Rijt (2006), if the accu- racy of teachers’ perception of difficulty is analysed by categories, teachers tend to overestimate the difficulty of easy items and under- estimate the difficulty of hard items. Impara and Plake (1998) also suggest that estimating item difficulty accurately is quite difficult; however, they do not think that teachers systematically underesti- mate the difficulty of hard items and overestimate the difficulty of easy items. In this respect, other contradictory results are found too. For example, Mattar (2000) states that teachers are less success- ful at rating very difficult or very easy items, while Zhou (2009) indicates that teachers classify better the hardest items. In short, although there are not conclusive studies about the tendency of teachers when they classify questions by difficulty le- vel, all researchers agree on the difficulty of doing this classifica- tion. Therefore, an automatic system that adjusts the difficulty level of questions according to the students’ behaviour would be a very useful support tool and a key component for a truly adaptive learning environment. 2.2. In search of an intelligent solution for a competitive tool There are many domain-dependent intelligent tutoring systems (ITSs) that provide students an adequate learning path through the different topics of a subject, according to the previously learnt topics. These systems are based on techniques such as Bayesian Networks (Hibou & Labat, 2004; Nouh, Karthikeyani, & Nadarajan, 2006; Vomlel, 2004) and require the previous definition of knowl- edge domains by using, for example, domain-specific ontologies (Colace & De Santo, 2006). Modelling these networks of knowledge components and their dependencies, generalizing them for every student, is not an easy task (Noguez, Sucar, & Ramos, 2005), espe- cially for domain-independent systems like QUESTOURnament, which can be used for diverse subjects and levels of education. Many domain-independent ITSs focus on presenting questions and problems adapted to the students’ knowledge level. They often apply the Item Response Theory (IRT) to estimate both the charac- teristics of the questions, such as difficulty or guessing probability, and the knowledge level of students (Chen, Lee, & Chen, 2005; Lilley et al., 2004), independently of the knowledge domain. However, the correct application of traditional theories for tests implies some assumptions, which are not met by many examination contexts, especially when telematic tools are used for distance learning. Moreover, some of the characteristics of more specific tools, such as the competitive nature of QUESTOURnament, make the applica- tion of these theories difficult for the environment under study. The typically used IRT models are one-dimensional, that is, they assume that the response to a question depends on a single trait, usually the knowledge level. Besides, it is also supposed that the response a student gives to a specific question does not depend on the responses given to other questions (Embretson & Reise, 2000). Therefore, using IRT entails carefully designing the tests so that these both conditions are fulfilled. Moreover, conventional IRT models only the response accuracy, ignoring response time; since it was thought to be used in pure power tests (Roskam, 1997), which assume that students have unlimited time to solve a question. Even if limited time could be assumed, at least the requirement should be that time is not a factor that affects the stu- dents’ response. However, in a competitive environment as QUES- TOURnament, time is very important, since only the first student who answers a challenge correctly will be able to obtain the high- est score for that challenge. Therefore, there are different factors that could distort the results obtained by the IRT methods when applied to the QUESTOURnament system. Students can apply different strategies during competition and even different personality factors can determine the students’ final response to an item. Several challenges can be posed at the same time and students have to select one of them to be solved first. Many students tend to read all the different questions and select the one that seems the easiest to be solved first. Difficult chal- lenges are usually read several times and solved after the easiest ones have been answered. On the other hand, two students with exactly the same knowledge level could respond to a same ques- tion differently, as one can be more persistent and devote more time to solve the question while another one can be more anxious with the competition and quickly respond to be the first one. Con- sequently, time and number of readings are important factors that should be taken into account in the model, but its modelling is dependent on the actual students’ behaviour. Moreover, when teachers pose challenges to QUESTOURna- ment, they do not have any restriction related to time, type of questions or skills needed to solve them. They are free to use any configuration of the system in any context. Then, there are some important factors that can vary: � Maximum time available to submit an answer to a challenge. � Type of questions (open questions, multiple choice questions, true/false questions, short response questions, problems, etc). � Context surrounding students when they solve the questions: a contest may be developed in classroom or on distance during one or several days. E. Verdú et al. / Expert Systems with Applications 39 (2012) 7471–7478 7473 � Personality of the students (e.g. the stress of the student faced with a competitive situation can influence on the response). There are different models adapted from the classic IRT that cover different partial aspects of the searched solution but there is not a model that covers all aspects required by the specific fea- tures of the QUESTOURnament system. Roskam (1997) presents a model based on IRT for speed tests with time limit where correct- ness and response time are integrated. Van der Linden (2007) pro- poses a flexible hierarchical solution that basically has an IRT model, a time-response distribution model and a higher level structure that has into account the dependencies between the items and the students’ parameters in those models. For each of these components, the most suitable model can be used. Anyway, a model based on IRT, which took into account all possi- ble factors that influence the response a student gives to a challenge within QUESTOURnament, according to the so many different contexts of application, would be vastly complex. There are other solutions to determine the difficulty level of the learning material. However, most of these proposals are too simplistic – like the solu- tion used in Jong et al. (2006), where the difficulty is estimated as the ratio between the number of times that a question is incorrectly answered and the total number of answers – or are too focused on the target subject – such as the solution described in Kunichika, Urushima, Hirashima, and Takeuchi (2002), which estimates the difficulty level of questions about English language sentences. After analysing classical and specific solutions, it was decided to design an ad-hoc solution for the system, whose fundamentals could be applied to other systems used in open contexts. This solu- tion is based on the definition of a fuzzy genetic expert system, which classifies the questions in several difficulty levels. There are examples of successful application of this kind of systems to e-learning environments such as the one described by Romero, Gonzalez, Ventura, del Jesus, and Herrera (2009). They use an evolu- tionary algorithm to learn fuzzy rules, which describe relationships be- tween the students’ interactions with the e-learning system Moodle and the final marks obtained in the course. Typically, genetic learning of rules assumes a predefined set of fuzzy membership functions gen- erated by human domain experts (Cordón, 2004). However, as afore- mentioned, the different nature of the challenges that can be posed through QUESTOURnament, as well as the varied students’ profiles, makes it very difficult to define and generalize fuzzy sets and rules. Teachers can use QUESTOURnament for multiple-choice questions or for laborious exercises or problems. Since, for example, ten minutes can be a very short time for a complex problem but a long time for a true/false question, it is very difficult to predefine the fuzzy member- ship function for the time parameter. Moreover, the contests with QUESTOURnament can be developed in very different contexts, for example, during face-to-face classes or on distance, even lasting sev- eral weeks. All these elements (nature of the questions, application contexts of the system, profiles and behaviours of the students. . .) make it necessary to define fuzzy sets and fuzzy rules each time a group of questions are classified. Doing it by hand should be very laborious and impractical, so an automatic system is needed. Besides, according to Nebot, Mugica, Castro, and Acosta (2010), learning the fuzzification process parameters by genetic algorithms instead of using the expert’s criteria provides better results. Then, the proposed system starts from scratch. Taking some data about the interaction of the students with QUESTOURnament and the initial difficulty level estimated by the teacher, it learns both the adequate membership functions with their linguistic values as well as the fuzzy rules. In the next section, the complete system is detailed. Along this description of the genetic fuzzy expert system some real case examples are included to facilitate comprehension. The details of the real case and the corresponding experiment results are set out in Section 4. 3. The expert system A genetic fuzzy expert system has been designed that generates fuzzy sets and rules appropriate for each specific case. The knowl- edge base is provided by a Fuzzy Model Generator that includes a genetic system capable of identifying the characteristics of the questions for each difficulty level. The estimation of the difficulty level then takes place in two phases. During a first phase the Fuzzy Model Generator learns from the Facts Base (formed by the students’ response patterns) and dynamically creates the classification rules and the fuzzy sets of the input variables for the specific data. During a second phase, the fuzzy expert system infers the difficulty level of each question. The components of all the system are shown in Fig. 1. From Moodle and QUESTOURnament logs, three parameters are consid- ered in the response patterns: time in minutes from the last read- ing of the question until the submission of the answer, grade obtained for that answer and number of accesses or readings be- fore submitting the answer. All these factors depend on the stu- dents’ behaviour when answering a question and are related to the difficulty level of each challenge (as aforementioned). All these data make up a set of context-dependent and noisy usage patterns that are stored in the Facts Base and feed the intelligent system. For each difficulty level, the genetic system uses the response patterns of all the questions belonging to that level (according to the initial classification made by the teacher) and obtains a charac- terization of their responses as crisp sets. From these crisp sets the Fuzzy Model Generator creates the fuzzy sets and rules of the Knowledge Base. Once the fuzzy sets and the rules of a group of questions have been generated, the Inference Engine can infer the difficulty level of the patterns in the Facts Base. Finally, the dif- ficulty level of each question is calculated as the median of the dif- ficulty level of its response patterns and the challenges repository is updated with the new difficulty level. Thus, the system combines the students’ behaviour and the teachers’ perception in order to objectively estimate the real diffi- culty level of each challenge. 3.1. The genetic system The objective of the genetic algorithm for the proposed system is to generate groups of crisp sets that characterize the students’ responses for three difficulty levels: easy, moderate and hard. The system groups challenges by difficulty level according to the initial classification made by the teacher. The genetic algorithm then uses the responses for all the questions belonging to a specific diffi- culty level in order to obtain its characterization. As above men- tioned, the input of the genetic algorithm is a set of response patterns with the structure <time, grade, accesses>. The crisp sets then are ranges of time, grade and number of accesses that together include the highest number of response patterns for a specific diffi- culty level. Therefore, a possible solution to the problem is repre- sented with the following coded chromosome or individual [t1, t2, g1, g2, a1, a2], being t1 and t2 the lower and upper limits of a time range, g1 and g2, the lower and upper limits of a grade range, and a1 and a2, the lower and upper values of a number of accesses range. The genetic algorithm implements the crossover operator BLX-a, the uniform mutation operator, the roulette wheel as selection method, and a fitness function based on the support measure typi- cally used to evaluate inferred rules. In addition, due to the fact that the response patterns for a question do not depend only on the ques- tion itself but also on the behaviour of the students answering (e.g. knowledge level, persistence, etc.), the algorithm also incorporates niching methods, such as sharing, in order to promote the diversity and to be able to characterize each difficulty level by several groups Fig. 1. Architecture of the genetic fuzzy expert system. Table 1 Groups of ranges delivered by the genetic system (where Acc. = number of accesses). Hard level Moderate level Easy level Time Grade Acc. Fitness Time Grade Acc. Fitness Time Grade Acc. Fitness [1, 48] [0, 5] [3, 4] 0.143 [4, 28] [30, 50] [1, 2] 0.301 [13, 36] [44, 60] [1, 2] 0.107 [7, 35] [0, 18] [1, 2] 0.251 [0, 25] [80, 100] [1, 2] 0.541 [27, 74] [84, 100] [1, 2] 0.167 7474 E. Verdú et al. / Expert Systems with Applications 39 (2012) 7471–7478 of ranges. More details about the fitness function, selection method, crossover and mutation operators and diversity methods are available in Verdú, Regueras, Verdú, and de Castro (2010a,2010b). 3.2. Generation of the Fuzzy Model For each difficulty level, the genetic algorithm obtains groups of ranges of the input variables that characterize the response pat- terns of that difficulty level. Table 1 shows the ranges obtained by the genetic algorithm in the experiment (which is described in detail in Section 4.1). Later, from these groups of ranges, the Fuz- zy Model Generator obtains the membership functions and the classification rules of the Fuzzy Model. The fuzzy sets for the input variables grade, time and number of accesses corresponding to the data of Table 1 are represented in Fig. 2. In principle, a fuzzy set is defined for each range found by the genetic algorithm. For example the fuzzy set of the input vari- able grade with linguistic value ‘‘Very Low (VL)’’ corresponds to the grade range [0, 5], found by the genetic system for the hard level questions. However, when two ranges are very similar, such as the grade range [80, 100] found in easy questions and the grade range [84, 100] found in moderate ones, the algorithm assigns only one fuzzy set for both ranges, ‘‘Very High (VH)’’ in this specific example. On the other hand, when the system finds a range that is not similar to another range but includes or overlaps this range, the system cre- ates several fuzzy sets. For example, the grade range [0, 18] includes the grade range [0, 5]. In this case the system generates two differ- ent linguistic values, ‘‘Very Low (VL)’’ and ‘‘Low (L)’’. The number of linguistic values of the input variables (Very Low, Low, Medium, High, etc.) is not fixed as it depends on the number of fuzzy sets determined by the Fuzzy Model Generator each time a group of questions are classified. In the given example, five fuzzy sets have been defined for the input variables grade and time and then, they take the linguistic values ‘‘Very Low (VL)’’, ‘‘Low (L)’’, ‘‘Medium (M)’’, ‘‘High (H)’’ and ‘‘Very High (VH)’’. How- ever, only two fuzzy sets have been defined for the input variable number of accesses and then, two linguistic values are used: ‘‘Low (L)’’ and ‘‘High (H)’’. The membership functions of the output variable Difficulty are not dynamically set, unlike those of the input variables, and always take the trapezoidal shapes shown in Fig. 2. Once the fuzzy sets have been automatically created, the Fuzzy Model Generator defines the fuzzy rules from these fuzzy sets and the results of the Genetic Algorithm (see Table 1). For example, for the easy difficulty level the following fuzzy rules have been auto- matically defined: IF GRADE IS VH AND TIME IS VL AND ACCESSES IS L THEN DIF- FICULTY IS EASY IF GRADE IS VH AND TIME IS L AND ACCESSES IS L THEN DIFFI- CULTY IS EASY Two rules have been created from an only group of ranges found by the genetic system since the time range was split during the fuzzy sets creation phase. This same procedure is followed in order to define all the fuzzy rules from each group of ranges deliv- ered by the genetic algorithm. 0 10 20 30 40 50 60 70 80 90 100 0 0.2 0.4 0.6 0.8 1 Grade M H VHLVL 0 10 20 30 40 50 60 70 80 90 100 0 0.2 0.4 0.6 0.8 1 Time VL L HM VH 0 1 2 3 4 5 6 7 8 0 0.2 0.4 0.6 0.8 1 Accesses L H 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 Difficulty Easy Moderate Hard D eg re es of m em be rs hi p Fig. 2. Fuzzy sets of the input and the output variables. Fig. 3. Sliced-cube FAM representation of the set of rules. E. Verdú et al. / Expert Systems with Applications 39 (2012) 7471–7478 7475 The fuzzy rules can be graphically represented by using a cube form called Fuzzy Associative Memory (FAM) as shown in Fig. 3, where the 16 rules that describe the given data are represented. At the end of this process, a set of fuzzy rules as well as the membership functions of the input and output variables describing the problem have been defined and stored in the Knowl- edge Base. 3.3. The inference engine The inference engine uses the Mamdani method to infer the difficulty level corresponding to a response pattern from its three crisp input variables: time, grade and number of accesses. The Matlab Fuzzy Logic Toolbox has been used to simulate the operation of this component. Again, a concrete example is used to show this operation. Fig. 4 shows the fuzzy inference of the Fig. 4. Fuzzy inference of the difficulty of the input pattern <31, 67, 3>. 7476 E. Verdú et al. / Expert Systems with Applications 39 (2012) 7471–7478 difficulty for the response pattern with time, grade and number of accesses equal to 31, 67 and 3, respectively. Fuzzy inference takes place in four steps. First, the crisp input variables are fuzzified during the fuzzification phase. The rule eval- uation phase then takes place. In the example, 16 rules describe the behaviour of the system but this input pattern only fulfils the three antecedent conditions of two rules. The classical Min method has been applied for the fuzzy operator AND. During the third step, the rule consequents are aggregated into a single fuzzy set using the Max composition method. Last, defuzzification obtains a crisp output using the MOM (Mean of Maximum) method. The pattern of the example has been assigned a difficulty value equal to 1.27, which corresponds to a difficulty level halfway between moderate and hard, but closer to moderate, as it can be seen in Fig. 2. 4. Results The hypothesis to be tested in this paper is that the designed ex- pert system performs as a human expert, that is, an expert teacher that is able to reclassify questions by difficulty by means of a thor- ough analysis of the students’ behaviour while answering. In order to analyze and validate the performance of this expert system, this one has been tested with real data from a contest developed with the QUESTOURnament tool in an undergraduate course of Diploma in Telecommunications Engineering at the Uni- versity of Valladolid (Spain). 4.1. The experiment The study was carried out from February until June 2010 (dur- ing three weeks with a 2-hour laboratory session in each week), with 38 enrolled students. All these students participated in the contest and 12 challenges (exercises on IP addressing and routing), were posed by the teacher. The system asked the teacher, upon creating the challenges, to classify them according to the estimated difficulty level: easy, moderate or hard. According to the teacher’s initial estimation; 4 challenges were classified as easy, 3 as moderate and 5 as hard. The total number of available answers for the easy, moderate and hard levels was 134, 103 and 169, respectively. For each answer, one response pattern is recorded with the grade obtained, the number of accesses and the time from reading to answering. All these records made up the in- put patterns to the system. In the previous sections, it has been explained the design and operation of the genetic fuzzy expert system with the help of some examples related to the real case presented now. Therefore, Table 1 corresponds to the output given by the genetic algorithm with the input data corresponding to the 12 challenges posed in this course. These groups of ranges found by the genetic system were used by the Fuzzy Model Generator to create the membership functions of the input variables shown in Fig. 2 as explained in Section 3.2. In the same way, the output of the genetic system and the generated fuzzy sets were used to define the 16 classification rules shown in Fig. 3, which describe the behaviour of the students when answer- ing these challenges of different difficulty levels. For example, the system found that answers with Very High grade and Low number of accesses correspond to the easy difficulty level if the time is Low or Very Low, or to the moderate difficulty level if the time is Med- ium or Higher. On the other hand, as it was expected, easy questions are char- acterized by Very High grades and Very Low or low time. There are also moderate questions with answers graded Very High but the time in this case is higher than the time required for the easy ques- tions before mentioned, indicating that time and difficulty are in- versely related, which is consistent with the study presented in Mason, Zollman, Bramble, and O’Brien (1992). Applying the rules to the students’ response patterns, the expert system obtains the new difficulty level for each challenge. Thus, next step is to update the challenges repository consequently. According to the classification done by the system (see the third column in Table 2), three challenges should be reclassified (ques- tions number 2, 8 and 10) as their initial difficulty does not match the difficulty assigned by the system. Once the questions have been reclassified, it is necessary to val- idate the intelligent system. 4.2. Validation of the intelligent system Since expert systems are addressed to perform at close to hu- man expert levels and to solve problems without a defined correct solution, they should typically be validated against human experts (O’Keefe, Balci & Smith, 1987). Therefore, the chosen method for the system validation has been a ‘‘validation against a group of Table 2 Data for validation of the expert system. Id Initial classification by teacher Expert system (crisp value) Expert system (linguistic value) Human expert 1 Human expert 2 Human expert 3 1 Moderate 1.08 Moderate Moderate Moderate Moderate 2 Moderate 1.51 Hard bordering on moderate Hard bordering on moderate Hard Hard 3 Easy 0.38 Easy Easy Easy Easy 4 Easy 0.52 Easy bordering on moderate Easy Easy Easy bordering on moderate 5 Moderate 1.07 Moderate Moderate Moderate Moderate 6 Hard 1.48 Hard bordering on moderate Hard Hard Hard bordering on moderate 7 Easy 0.37 Easy Easy Easy Easy 8 Easy 1.08 Moderate Moderate Moderate Moderate 9 Hard 1.68 Hard Hard Hard Hard 10 Hard 1.32 Moderate bordering on hard Moderate Moderate Moderate bordering on hard 11 Hard 1.62 Hard Hard Hard Hard 12 Hard 1.55 Hard bordering on moderate Moderate bordering on hard Hard Moderate bordering on hard Table 3 Values of weighted kappa. Human expert 1 Human expert 2 Human expert 3 Expert system Human expert 1 – 0.901 0.805 0.837 Human expert 2 0.901 – 0.813 0.747 Human expert 3 0.805 0.813 – 0.947 Expert system 0.837 0.747 0.947 – E. Verdú et al. / Expert Systems with Applications 39 (2012) 7471–7478 7477 experts’’ based on the method described in Mosqueira-Rey, Moret- Bonillo, and Fernández-Leal (2008). This method provides a mea- sure of agreement between the human experts and verifies if the expert system performs as one of them and, therefore, it can be then incorporated into the group of experts without making the agreement level worse. In the experiment described the group of experts consisted of the teacher of the course and other two teachers who are also ex- perts on the subject. Table 2 shows the difficulty levels estimated by the genetic fuzzy expert system and the group of experts for each of the 12 challenges. First column is simply a challenge iden- tifier. Second column shows the initial classification done by the teacher. Third column shows a crisp number that represents the difficulty of a challenge obtained by the system, which ranges from 0 to 2 (see Fig. 2). Fourth column represents the corresponding lin- guistic value for this crisp output. Last three columns show the classification done by the human experts, who carefully assigned a difficulty level to all the questions after analyzing the actual re- sults and behaviour of students who answered them. The level of agreement between each pair of human experts has been measured through the weighted kappa (Mosqueira-Rey et al., 2008). The results of the measure of kappa (see Table 3) show a strong agreement between pairs of experts; since values of kappa higher than 0.80 indicate an almost perfect agreement whereas values in the range 0.61–0.80 indicate a significant agreement (Viera & Garrett, 2005). The level of agreement between the expert system and each human expert varies from a significant agreement (kappa = 0.747) to an almost perfect agreement (kappa = 0.947). Next, Williams index (Mosqueira-Rey et al., 2008) has been used as a measure to verify that introducing the system into the group of experts does not decrease the agreement level of the group. Values equal or higher than 1 indicate a good agreement whereas values lower than 1 imply that the agreement in the group of experts with the expert system included is worse than the agreement among only human experts. The Williams index has been calculated from the values for kappa of Table 3, and a value of 1.005 has been ob- tained. Therefore, it can be concluded that the system has per- formed successfully and is able to do the reclassification labour on behalf of the teacher satisfactorily. 5. Discussion and conclusion An expert system that satisfactorily classifies the challenges posed in the competitive learning system QUESTOURnament according to their real difficulty level has been designed. Teachers can insert challenges into the QUESTOURnament tool and the genetic fuzzy expert system will readjust the initial difficulty level estimated by the teacher according to the real behaviour of stu- dents when facing up to the challenges. The system has been tested with real data and the results have been successfully validated against human experts. Once the system has been validated, its results can also be used to study and analyze the accuracy of estimations done by teachers. When compared with the difficulty level obtained by the expert system, the teacher’s estimation (Initial classification by teacher in Table 2) is quite accurate, having into account that the teacher uses a three-scale method, whereas the values calculated from the data of the system include intermediate levels. The teacher’s estimation does not match the difficulty assigned by the system in three cases. Specifically, the teacher overestimates the difficulty of one question and underestimates the difficulty of other two questions without following any rule. Thus, no conclusion can be obtained about the tendency of the teacher. The results of Table 2 only show that teachers estimate better the difficulty of hard questions than of easy or moderate questions. For future work it is intended to study the possible inclusion of more parameters, related to students’ profile and behaviour, in the response patterns. Besides, the system assumes that the initial classification of questions done by the teacher is good enough. It could be interesting to study in depth how dependant is the ap- proach on the initial difficulty levels assigned by teachers. Finally, the tests show that a same difficulty level is character- ized by different fuzzy sets. This is probably due to the different behaviour of students when solving a challenge (for example, some of them can be more resolute while other ones can be more persis- tent) and, of course, to their different knowledge level. Then, the output of the system could also be used for students clustering be- cause typical behaviours of students can be detected from the rules shown in the FAM representation of Fig. 3. For example, there are students who, when they do not know how to answer a challenge, read it several times and, finally, submit a quick response just in case they get it right by chance. This corresponds to the hard level questions characterized by a high number of accesses, Very Low to medium time for solving and Very Low grade. There are other rules for hard level questions that correspond to those more promising students who get a high grade even in hard questions. Therefore, each rule for the same difficulty level may correspond to students with similar knowledge level and/or behaviour when answering 7478 E. Verdú et al. / Expert Systems with Applications 39 (2012) 7471–7478 questions; this result can be used to effectively classify students and to refine the model by taking into account their competences and profiles. Thus, it is also planned to study the possibility of using the different fuzzy sets obtained by the Fuzzy Model Gener- ator to detect and classify groups of students according to their knowledge level and behaviour profile. References Alexandrou-Leonidou, V., & Philippou, G. N. (2005). Teachers’ beliefs about students’ development of the pre-algebraic concepto of equation. In Proceedings of the 29th Conference of the International Group for the Psychology of Mathematics Education (pp. 41–48). Melbourne: University of Melbourne. Anderson, J. R. (2006). On cooperative and competitive learning in the Management Classroom. Mountain Plains Journal of Business and Economics - Pedagogy, 7, 1–10. Burghof, K. L. (2001). Assembling an item-bank for computerised linear and adaptive testing in Geography. International Education Journal, 2(4), 74–83. Chen, C.-M., Lee, H.-M., & Chen, Y.-H. (2005). Personalized e-learning system using item response theory. Computers & Education, 44(3), 237–255. Cheng, I., Shen, R., & Basu, A. (2008). An algorithm for automatic difficulty level estimation of multimedia mathematical test items. In Proceedings of the 8th IEEE International Conference on Advanced Learning Technologies (pp. 175–179). Los Alamitos, CA: IEEE Computer Society. Colace, F., & De Santo, M. A. (2006). A tutoring tool based on bayesian approach. In Proceedings of Sixth International Conference on Advanced Learning Technologies (pp. 109–113). Washington DC: IEEE Computer Society. Cordón, O. (2004). Ten years of genetic fuzzy systems: Current framework and new trends. Fuzzy Sets and Systems, 141(1), 5–31. Embretson, S. E., & Reise, S. P. (2000). Item response theory for psychologists. Mahwah, NJ: Lawrence Erlbaum Associates. Goodwin, L. D. (1999). Relations between observed item difficulty levels and Angoff minimum passing levels for a group of borderline examinees. Applied Measurement in Education, 12, 13–28. Hadjidemetriou, C., & Williams, J. S. (2002) Teachers’ pedagogical content knowledge: graphs, from a cognitivist to a situated perspective. In Proceedings of the 26th Conference of the International Group for the Psychology of Mathematics Education (pp. 57–64). Norwich: University of East Anglia. Hibou, M., & Labat, J.-M. (2004). Embedded Bayesian network student models. In Proceedings of the Fifth International Conference on Information Technology Based Higher Education and Training (pp 468–472). Istanbul: IEEE. Impara, J. C., & Plake, B. S. (1998). Teachers’ ability to estimate item difficulty: A test of the assumptions in the Angoff standard setting method. Journal of Educational Measurement, 35(1), 69–81. Jong, B.-S., Chan, T.-Y., Wu, Y.-L., & Lin, T.-W. (2006). Applying the adaptive learning material producing strategy to group learning. Lecture Notes in Computer Science, 3942, 39–49. Kunichika, H., Urushima, M., Hirashima, T., & Takeuchi, A. (2002). A computational method of complexity of questions on contents of English sentences and its evaluation. In Proceedings of the International Conference on Computers in Education (pp. 97–101). Auckland, New Zealand: IEEE Computer Society. Lawrence, R. (2004). Teaching Data Structures Using Competitive Games. IEEE Transactions on Education, 47(4), 459–466. Lee, F. L. (1996). Electronic Homework: An Intelligent Tutoring System in Mathematics. PhD Thesis. The Chinese University of Hong Kong. Lee, F. L., & Heyworth, R. M. (2000). Problem complexity: A measure of problem difficulty in algebra by using computer. Education Journal, 28(1), 85–107. Lilley, M., Barker, T., & Britton, C. (2004). The development and evaluation of a software prototype for computer-adaptive testing. Computers & Education, 43(1), 109–123. Mason, E., Zollman, A., Bramble, W. J., & O’Brien, J. (1992). Response time and item difficulty in a computer-based high school mathematics course. Focus on Learning Problems in Mathematics, 14(3), 41–51. Mattar, J. D. (2000). Investigation of the validity of the Angoff standard setting procedure for multiple-choice items. Ph.D. dissertation. University of Massachusetts. Mosqueira-Rey, E., Moret-Bonillo, V., & Fernández-Leal, Á. (2008). An expert system to achieve fuzzy interpretations of validation data. Expert Systems with Applications, 35(4), 2089–2106. Nebot, A., Mugica, F., Castro, F. & Acosta, J. (2010). Genetic fuzzy system for predictive and decision support modelling in e-learning. In Proceedings of the 2010 IEEE International Conference on Fuzzy Systems (pp. 1804–1811). IEEE Computer Society. Noguez, J., Sucar, E., & Ramos, F. (2005). A probabilistic relational student model for virtual laboratories. In Proceedings of the Sixth Mexican International Conference on Computer Science (pp. 2–9). Puebla, Mexico. doi:10.1109/ENC.2005.7. <http:// ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=1592194>. Nouh, Y., Karthikeyani, P., & Nadarajan, R. (2006). Intelligent tutoring system- Bayesian student model. In Proceedings of the 1st International Conference on Digital Information Management (pp 257–262). Bangalore: IEEE. O’Keefe, R., Balci, O., & Smith, E. P. (1987). Validation of expert system performance. IEEE Transactions on Expert Systems, 2(4), 81–90. Philpot, T. A., Hall, R. H., Hubing, N., & Flori, R. E. (2005). Using games to teach statics calculation procedures: Application and assessment. Computer Applications in Engineering Education, 13(3), 222–232. Regueras, L. M., Verdú, E., Muñoz, M. F., Pérez, M. A., de Castro, J. P., & Verdú, M. J. (2009). Effects of competitive e-learning tools on higher education students: A case study. IEEE Transactions on Education, 52(2), 279–285. Romero, C., Gonzalez, P., Ventura, S., del Jesus, M. J., & Herrera, F. (2009). Evolutionary algorithms for subgroup discovery in e-learning: A practical application using Moodle data. Expert Systems with Applications, 36(2), 1632–1644. Roskam, E. E. (1997). Models for speed and time-limit tests. In W. J. van der Linden & R. K. Hambleton (Eds.), Handbook of Modern Item Response Theory (pp. 188–207). New York: Springer-Verlag. Van der Linden, W. J. (2007). A hierarchical framework for modeling speed and accuracy on test items. Psychometrika, 72(3), 287–308. Verhoeven, B. H., Verwijnen, G. M., Muijtjens, A. M. M., Scherpbier, A. J. J. A., & Van der Vleuten, C. P. M. (2002). Panel expertise for an Angoff standard setting procedure in progress testing: Item writers compared to recently graduated students. Medical Education, 36, 860–867. Verdú, E., Regueras, L. M., Verdú, M. J., de Castro, J. P., & Pérez, M. A. (2008). An analysis of the research on adaptive learning: The next generation of e-learning. WSEAS Transactions on Information Science and Applications, 5(6), 859–868. Verdú, E., Regueras, L. M., Verdú, M. J., & de Castro, J. P. (2010a). Estimating the difficulty level of the challenges proposed in a competitive e-learning environment. Lecture Notes in Artificial Intelligence, 6096, 225–234. Verdú, E., Verdú, M. J., Regueras, L. M., & de Castro, J. P. (2010b). A diversity- enhanced genetic algorithm to characterize the questions of a competitive e- learning system. In Proceedings of IEEE International Conference on Advanced Learning Technologies (pp. 25–29). Los Alamitos, CA: IEEE Computer Society. Viera, A. J., & Garrett, J. M. (2005). Understanding interobserver agreement: The Kappa statistic. Family Medicine, 37(5), 360–363. Vomlel, J. (2004). Building adaptive tests using Bayesian Networks. Kybernetika, 40, 333–348. Watering, G. V. D., & Rijt, J. V. D. (2006). Teachers’ and students’ perceptions of assessments: A review and a study into the ability and accuracy of estimating the difficulty levels of assessment items. Educational Research Review, 1, 133–147. Wauters, K., Desmet, P., & Van den Noortgate, W. (2010). Adaptive item-based learning environments based on the item response theory: Possibilities and challenges. Journal of Computer Assisted Learning, 26, 549–562. Wu, W. M. C., Cheng, H. N. H., Chiang, M.-C., Deng, Y.-C., Chou, C.-Y., Tsai, C.-C., & Chan, T.-W. (2007). Answer matching: A competitive learning game with uneven chance tactic. In Proceedings of the First IEEE International Workshop on Digital Game and Intelligent Toy Enhanced Learning (pp.89–96). Los Alamitos, CA: IEEE Computer Society. Zhou, W. (2009). Teachers’ estimation of item difficulty: What contributes to their accuracy? In Proceedings of the 31st Annual Meeting of the North American Chapter of the International Group for the Psychology of Mathematics Education (pp. 261–264). Atlanta: Georgia State University. http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=1592194 http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=1592194 A genetic fuzzy expert system for automatic question classification in a competitive learning environment 1 Introduction 2 Background 2.1 Teachers’ perception of difficulty 2.2 In search of an intelligent solution for a competitive tool 3 The expert system 3.1 The genetic system 3.2 Generation of the Fuzzy Model 3.3 The inference engine 4 Results 4.1 The experiment 4.2 Validation of the intelligent system 5 Discussion and conclusion References 
work_te23qoqhl5dx5nrldisibugyha	----	A hybrid model for business process event and outcome prediction Article DOI: 10.1111/exsy.12079 A hybrid model for business process event and outcome prediction Mai Le,1,2 Bogdan Gabrys1 and Detlef Nauck2 (1) School of Design, Engineering and Computing, Bournemouth University, Bournemouth, UK E-mail: mai.phuong@bt.com; bgabrys@bournemouth.ac.uk (2) BT Research and Technology, Ipswich, UK E-mail: detlef.nauck@bt.com Abstract: Large service companies run complex customer service processes to provide communication services to their customers. The flawless execution of these processes is essential because customer service is an important differentiator. They must also be able to predict if processes will complete successfully or run into exceptions in order to intervene at the right time, preempt problems and maintain customer service. Business process data are sequential in nature and can be very diverse. Thus, there is a need for an efficient sequential forecasting methodology that can cope with this diversity. This paper proposes two approaches, a sequential k nearest neighbour and an extension of Markov models both with an added component based on sequence alignment. The proposed approaches exploit temporal categorical features of the data to predict the process next steps using higher order Markov models and the process outcomes using sequence alignment technique. The diversity aspect of the data is also added by considering subsets of similar process sequences based on k nearest neighbours. We have shown, via a set of experiments, that our sequential k nearest neighbour offers better results when compared with the original ones; our extension Markov model outperforms random guess, Markov models and hidden Markov models. Keywords: artificial intelligence, method, system 1. Introduction Very often, processes are poorly documented because they have evolved over time or the legacy information technology systems that were used to implement them may have changed. This adds the additional challenge to identify the actual process model. Process mining (van der Aalst, 2011) can reconstruct process models from workflow data to some extent by searching for a model that explains the observed workflow. This process model contains the list of all unique process prototypes (workflow sequences). Data from this mining process can help to address the problem of proactive process monitoring and process event prediction. Firstly, in order to predict process events, we consider approaches from data mining (Berry & Linoff, 2004). Because of the nature of the data, approaches that can deal with temporal sequence data are the most relevant. One example from this class of approaches is Markov models (MMs) that have been used to study stochastic processes. In the work of Pitkow and Pirolli (1999), for example, it has been shown that Markov models are well suited to the study of Web-users’ browsing behaviour. Event sequences are used to train a Markov model that encodes the transition probabilities between subsequent events. Event sequences can represent customer behaviour, customer transactions, workflow in business processes and so on. The prediction of the continuation of a given sequence is based on the transition probabilities encoded in the Markov model. The order of a Markov model represents the number of past events that are taken into account for predicting the subsequent event. Intuitively, higher order Markov models are more accurate than lower order Markov models because they use more information present in the data. This has been shown to be true in the prediction of Web browsing behaviour. Lower order models may not be able to discriminate the difference between two subsequences, especially when the data are very diverse, that is if there are a large number of unique sequences. In the work of Ruta and Majeed (2011), it has been shown that higher order Markov models can be more accurate than lower order models in the context of customer service data. However, higher order models can suffer from weak coverage in particular when the data are diverse and not clustered. For sequences that are not covered by the model, a default prediction is required. Default predictions typically reduce the accuracy of the model. It is obvious that with increasing order of the model, the computational complexity also grows. In the case of plain Markov models, there are already approaches to deal with the trade-off between coverage and accuracy. These approaches have been introduced in the studies of Deshpande and Karypis (2004), Eirinaki et al. (2005) and Pitkow and Pirolli (1999). The general idea is to merge the transition states of different order Markov models, which is followed by pruning ‘redundant’ states. A particular method is the selective © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 Markov model, an extension of All Kth order Markov models (Deshpande & Karypis, 2004). Secondly, to predict the process outcomes, bearing in mind that the data can be very diverse, our idea is to allocate a number of sequences from the historical data that are most similar to the given sequence expecting that they would have similar outcome. K nearest neighbour (KNN) is chosen as the predictive model here. In our area of application, we consider a number of sequences in which we want to find K sequences that are most similar to a given sequence S s jð Þ 1 ; …; s jð Þ nj � � , where j is the index of the sequence. The similarity is determined using a distance function, and the prediction should be the majority class among the classes of K nearest sequences. This method is used by Ruta et al. (2006) to predict churn using data from a telecommunication company. The authors combined KNNs and the theory of survival. In their work, the authors used Euclidean distances to calculate the distance between the given sequence and all sequences in the data sample. In this study, as our data consist of event sequences, we use the edit distance in the process of comparing two sequences. In order to exploit the temporal characteristics of process data, we combine KNN with sequence alignment from biology to form a sequential KNN. In this paper, we first present a hybrid approach to address the lack of coverage in higher order Markov models. This approach is a combination of pure Markov models and sequence alignment technique (Waterman, 1994). The sequence alignment technique is applied when the default prediction is required. This is the case when the given sequence cannot be found in the transition matrix. The matching procedure is applied in order to extract those sequences (patterns) from the transition matrix that are most similar to the given sequence. We determine the prediction for the given sequence based on the predictions for the sequences we obtained. We secondly present a sequential KNN approach that is used to predict process outcomes. Our sequential KNN matches sequences using local alignment. The output of the matching is used to rank the similarities of sequences. We are testing our approaches on workflow data from a telecommunication business process, but they are likely to be applicable in a variety of other sequential prediction problems – especially in the case of our extended Markov model if Markov models have already been used and shown to be relevant. The rest of this paper is organised as follows. Sequence alignment and distance measurement are presented in Section 2. Section 3 introduces the predictive models that are used in this study. Section 4 then presents the experimental results and evaluation. Finally, in Section 5, our conclusions and directions for future work are discussed. 2. Sequence alignment To determine sequence similarity, both distance measures and similarity measures can be used because distance and similarity are dual concepts. For numeric variables, well- known distance measures exist and can be easily applied. However, in sequence analysis, sometimes, we have to work with sequences that are constructed from symbols, for example, categories (churn prediction), phonemes (speech recognition) or characters (hand-writing recognition). In this case, we need specialised functions that have the ability of measuring the similarity of symbolic sequences. Sequence alignment is very common in bio-informatics and has a relatively long history in this domain. The target entities of sequence alignment in bio-informatics are amino acid sequences of proteins, DNA sequences and so on. Sequence alignment is used for a number of purposes (Needleman & Wunsch, 1970); for example, to compare new DNA sequences with DNA databases. Transactions of DNA sequences into amino acid sequences are compared with protein databases in order to verify if the relationships that are found between them could have occurred by chance and so on. Algorithms used in sequence alignment are mainly divided into global alignment and local alignment. Global alignment provides a global optimisation solution, which spans the entire length of all query sequences. In contrast, local alignment aims to find the most similar segments from two query sequences. We use the idea of sequence alignment in Markov models but do not use existing sequence alignment algorithms. Instead, we build a simple alignment procedure for short subsequences. We also use the local alignment algorithm and combine it with KNNs. The local algorithm is presented in the following: Given two sequences, we have to consider the order of the events and compute the score of matching the ith event in one sequence with the jth event in the other sequence, i = {1, …, len1}, j = {1, …, len2} where len1 and len2 are the lengths of the two given sequences. The aim of the local algorithm (Smith & Waterman, 1981; Waterman, 1994) is to find a pair of most similar segments in the given sequences. There are two key matrices used in local alignment: the substitution matrix and the score matrix. 1. Substitution matrix: In biology, a substitution matrix describes the rate at which one amino acid in a sequence transforms to another amino acid over time. The entries of this matrix present the probabilities of one amino acid transforming (mutating) into another. There are different ways of generating the substitution matrix. The simplest way is to not take into account amino acid mutations and instead just give a score of 1 to the same amino acids and use a score of 0 to a pair of different amino acids. s i; jð Þ ¼ 0 if event i ≠ event j 1 otherwise � In this case, the substitution matrix is an identity matrix, the elements of the main diagonal are 1 and all the others are 0. To present mutations, a more complicated form of substitution matrix is used. The elements of the matrix off the main diagonal can have © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 a value different from 0, depending on the likelihood and frequency of transformation between the two amino acids. 2. Score matrix: A score matrix is built based on the following formula: hi0 ¼ h0j ¼ h00 ¼ 0 (1) where hi0, h0j and h00 values are the initial values for the recursive formula that is used to compute hij. hij ¼ max hi�1; j � δ; hi�1; j�1 þ s xi; yj � � ; hi; j�1 � δ; 0 � � (2) where xi and yj are events at positions i and j from the given sequences. s(xi, yj) is the score from the substitution matrix corresponding to events xi and yj. Example: Given two sequences ABCDE and EBCAD, the corresponding score matrix is as follows: A B C D E E 0 0 0 0 1 B 0 1 0 0 0 C 0 0 2 0 0 A 1 0 0 1 0 D 0 0 0 1 0 0 BBBBBBBBB@ 1 CCCCCCCCCA The ith event in a sequence can be aligned to the jth event in another sequence or can be aligned to nothing (deletion). This leads to a number of possible matchings of the two sequences. The optimal pair of aligned segments is identified by first finding the highest score in the matrix. This element is the end of the optimal aligned path. Then, the path is filled by tracking back from that optimal highest score diagonally up towards the left corner until 0 is reached. In the example, the best match is BC. 3. Predictive models This section presents Markov models, KNNs and their extensions that can help enhance the current approaches to process management in a telecommunication business context by predicting future events and process outcomes based on the current and past data. Here, a business process instance (Sj) is a composition of discrete events (or tasks) in a time-ordered sequence, Sj = s jð Þ 1 ; s jð Þ 2 …s jð Þ nj n o where sj takes values from a finite set of event types E = {e1, …, eL}. Apart from its starting time t jð Þ i and duration of T jð Þ i , each of these events has its own attributes. For simplicity, we assume that a process does not contain any overlapping events that means there are no parallel structures. The goals of our predictive models are to predict the next event s Nþ1ð Þ iþ1 following event s Nþ1ð Þ i in a given process instance SNþ1 ¼ s Nþ1ð Þ1 ; s Nþ1ð Þ 2 ; …; s Nþ1ð Þ i�1 n o based on the data from completed process instances S = {S1, S2 … SN} or to predict the process outcomes (success/failure). A simple predictive model is illustrated in Figure 1. In the following subsections, we will discuss some examples of process prediction models, namely Markov models, hybrid Markov models (MSAs), hidden Markov models (HMMs), KNNs and sequential KNNs (KnsSA). 3.1. Markov models Markov models are a kind of stochastic model. They are based on the mathematical theory of Markov chains. Markov models (Papoulis & Pillai, 1991) have been used in studying stochastic processes in terms of modelling and predicting customer behaviour. The idea of the model is to use k > 0 steps to predict the following (k + 1)th step. Given a sequence of random variables {Xn}, a first order Markov model uses the current step xi � 1 to predict the next step xi Figure 1: A simple predictive model. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 and to generalise; a kth order Markov model uses the last k steps to predict the next one: p xi xi�1; …; xi�nÞ ¼ p xi xi�1; …; xi�kÞ:jðjð (3) To construct a Markov model a matrix M, M = (mij) is created containing the probabilities for moving from the previous k steps to the next. The columns of the matrix correspond to the unique tasks, i = (1, …, I), and the rows of the matrix correspond to the unique patterns of length k that are present in the training data set that is used to construct the model. These unique patterns are called states j = (1, …, J). Elements mij of matrix M are the probabilities pij: pij ¼ nij ni : (4) where ni is the number of times pattern i occurs in the data set, and nij is the number of times pattern i is followed by step j in the data set. Higher order Markov models can be expected to be more accurate. However, we have to take into account their weak coverage property. For example, consider using a Markov model to predict Web page access and consider a sequence of Web pages {P1, P2, P1, P3, P2, P1} that is used to create the model. In a second order Markov model, there would be no sample {P2, P3}. Consequently, the prediction for this pattern would have to be the default prediction of the model, that is, the most frequent step (P1 in this example). One approach to overcome the problem is to merge the transition states of different order Markov models, after pruning the ‘redundant’ states. A particular method is the selective Markov model, an extension of all Kth order Markov models (Deshpande & Karypis, 2004). Markov models are often not dynamic, and once the model is trained, the obtained matrix is fixed and used for predictions. It is important that the training set data are large enough to be representative for the encountered patterns. In order to accommodate changes in data, the matrix can be rebuilt from scratch after a certain period. It is straightforward to create a dynamic Markov model that adapts to new data by storing the counts nij and ni and updating M with additional rows and/or columns if new patterns and/or steps are encountered in the new data. We can also apply a discounting factor to the current counts to give more weight to the pattern frequencies in new data. 3.2. Hidden Markov models In the field of sequential data, HMMs are a very powerful tool. Similar to Markov models, HMMs are a type of stochastic model. They are also based on the mathematical theory of Markov processes that further developed into the theory of HMMs by Baum in the 1960s (Baum et al., 1970). HMMs received much attention because of their application in speech recognition. Moreover, there is a wide range of applications ranging from telecommunication, recognition (gesture, speech and so on) to finance, biology and so on. Many approaches in temporal data series only deal with observed variables. We can consider how the performance of the model changes after adding an unobservable or hidden variable that we assume to have an influence on the observed variables. HMMs assume that there is a latent variable that influences the values of the observed variable. The benefits of this approach are shown, for example, in the work of Cox and Popken (2002). As an example, for an application domain, consider modelling customer behaviour based on customer event sequences (e.g. historic purchases and contacts with customer service). A latent variable that influences customer behaviour could be the level of customer satisfaction. Let O = {o1, …, oN} be the sample of observed variables and Q a sequence of hidden states {q1, …, qN}. The observed variable can be continuous and follow a (usually Gaussian) distribution, or it can be discrete and take values from a finite set V = {v1, …, vK}. The hidden states are discrete, and their value space is Ω, Ω = {s1, …, sM}. The expression of the joint probability of observed sequence is as follows: p o1; o2; …; onjq1; …; qnð Þ ¼ p q1ð Þ Yn i¼2 p qijqi�1ð Þ Yn i¼1 p oijqið Þ (5) Because of the complexity of the computation, lower order hidden Markov models are more popular. The most popular one is the first order hidden Markov model that assumes that the current state depends only on the preceding state and is independent from the earlier states. p qn qn�1; on�1; …; q1; o1Þ ¼ p qn qn�1Þjðjð (6) Intuitively, higher order HMMs can be expected to be more accurate (Thede & Happer, 1999). However, the lack of coverage problem needs to be considered, as mentioned in Section 3.2. It is obvious that with increasing order of the model, we are facing more complex computation. An HMM is defined by a set of three parameters λ = (A, B, π): A ¼ aij � � ¼ p qijqj � �� � (7) where A is the transition matrix and an element aij is the probability to move from state j to state i. It is necessary to point out that only homogeneous HMMs are considered here, implying the time independence of the transitions matrix. The non-homogeneous type is discussed by Netzer et al. (2007). Furthermore, B ¼ bij � � ¼ p oijqj � �� � (8) where B is the matrix containing probabilities of having ot = vi, denoted as oi, if the hidden state is qj at time t. Finally, πj ¼ p q1 ¼ sj � � (9) is the probability that sj is the initial state at time t = 1. There are three fundamental problems in HMMs: estimation, evaluation and decoding. The estimation problem © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 is about finding the optimal set of parameters λ = (A, B, π) that maximise p(O|λ′) given a sequence O of observed states and some initial model λ′. The Baum–Welch algorithm is used for this. The evaluation problem means finding p(O|λ) given the sequence of observations O and the model λ using the forward (backward) algorithm. To solve the decoding problem, the Viterbi algorithm is applied for finding the most realisable, suitable sequence of hidden states q1, …, qk for the associated sequence O and model λ. Researchers have been working on improvements of the performance of HMMs. Similar to Markov models, there are approaches to deal with the trade-off between coverage and accuracy. However, merging the transition states of higher order hidden Markov models is much more complicated than for Markov models. 3.3. Hybrid Markov models The work presented in this paper investigates the idea of combining higher order Markov models with a sequence alignment technique (Waterman, 1994) in order to maintain the high accuracy of the higher order Markov models and, at the same time, compensate for their lack of coverage. When a higher order Markov model is employed to predict process events and the given sequence has no match in the transition matrix M, we would be forced to produce a default prediction that would be detrimental to the model accuracy. To this end, we present our approach, MSA for Markov sequence alignment, which is based on the assumption that similar sequences are likely to produce the same outcome. Basically, the steps that MSA took to identify the next step for a previously unseen sequence Snew have no match in the transition matrix as follows: • Compare Snew with all sequences (patterns and states) in the transition matrix, and extract l states that are most similar. • Generate prediction based on these l selected states. For each state i in the transition matrix M = (mij), there is a corresponding predictive row vector, (mi1, …, miJ). The predicted next step of state i is the task j with mij = max (mi1, …, miJ). Thus, the predictions for the l selected states contain l* unique events, l* < = l, E� ¼ e1; …; ; el�f g. Then, the final prediction for Snew is the most frequent event in E *. If, however, there is no single most frequent event in E*, then the one with highest probability is selected. Alternatively, we go back to the l most similar states and select the sequence that is most similar to Snew and use its predicted step as the prediction for Snew. If there is no single most similar sequence, we select the one with the highest frequency from this set. In order to compare Snew with the sequences in the transition matrix, we need to be able to define and measure their similarities or distances. Here, the sequence alignment technique is applied (Waterman, 1994). In particular, we adopt the edit (deletion and insertion) weighted distance function for its flexibility in determining the degree of similarity. Given two sequences, a weight is assigned to each pair of events at the same position in the two sequences. The weight is 0 if the two events are identical, 1 or δ (a specified value) if they are different. In case both events of the pair must be deleted in order to keep the longest possible similar subsequences, the corresponding weight is 1. If only one of the two events must be deleted, the weight is δ. The sum of weights of the comparison is then used to determine the similarity of sequences. The lower the weight is, the more similar two sequences are. This algorithm starts by constructing a matrix where the elements represent similarity between any pair of events from the two sequences to be matched. This matrix is then used to find the most suitable editing of the given sequences where deletions and insertions are penalised. The optimal editing is chosen based on the total score computed. For example, given two sequences ABCDE and EBCAD, the matrix looks as follows: A B C D E E 0 0 0 0 1 B 0 1 0 0 0 C 0 0 1 0 0 A 1 0 0 0 0 D 0 0 0 1 0 0 BBBBBBBBB@ 1 CCCCCCCCCA Next, the first events A and E from both sequences are deleted that results in a weight of 1. The second event in both sequences is B and has weight 0. The fourth event A is deleted from the sequence EBCAD in order to match two Ds in fourth and fifth positions in the two sequences. We add δ to the sum of weights. Finally, we delete the last event E from the sequence ABCDE, adding another δ. The total weight for matching the two sequences is w = 1 + 0 + 0 + δ + δ = 1 + 2δ. The following example will illustrate our method for the case of the transition matrix of a third order Markov model A B C ABC 0:1 0:2 0:7 CBC 0:1 0:5 0:4 BAA 0:2 0:5 0:3 0 BBB@ 1 CCCA and a given sequence BCC. This sequence has not occurred before and is not stored in the matrix. The aims are to find the most similar sequences from the matrix and use their predictions to generate the prediction for the given sequence BCC. We first match BCC to ABC: A B C B 0 1 0 C 0 0 1 C 0 0 1 0 BBB@ 1 CCCA © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 The weight of this comparison is w = δ + 0 + δ = 2δ. Similarly, the resulting weights of matching CBC and BAA against BCC are w = 2δ and w = 2, respectively. Let δ = 0.4 and l = 2, the two chosen sequences in this case are then ABC and CBC and their predictive vectors are (0.1, 0.2, 0.7) and (0.1, 0.5, 0.4), respectively. Based on these vectors, the predictive next steps are {B, C}. The weights and frequency of occurrence of these two events are equal; hence, the next step of the given sequence BAC is predicted to be C because of higher transition probability, m13 = 0.7 that is greater than m22 = 0.5. Algorithms 1 and 2 illustrate the procedure of matching two sequences by deletion and insertion of symbols with penalties. 3.4. KNNs combined with sequence alignment KNN is one of the classical approaches in data mining (Berry & Linoff, 2004). It is essentially a non-parametric approach; hence, one of its advantages is that there is no training procedure required. The model automatically selects sequences from the continuously updated data. The only configuration possible is a choice of different values for K ≥ 1, that is, how many nearest neighbours we want to use. KNNs can be used in its original non-sequential form (Eastwood & Gabrys, 2009), or it can be extended into a sequential approach (Ruta et al., 2006). The core idea is to find similar sequences expecting that these sequences have a common behaviour and outcome. This is a reasonable expectation, for example, in biology where similar DNA or protein sequences are expected to have the same shape function. Algorithm 2 Function calculateScore(n, maxCommonIndexes1, maxCommonIndexes2, maxCommonLength): computes the score of matching two sequences 1: totalPoints=0 2: totalDeltas = 0 3: maxCommonIndexes1(maxCommonLength + 1) = n + 1 4: maxCommonIndexes2(maxCommonLength + 1) = n + 1 5: last1 = 0 6: last2 = 0 7: for i = 1 : maxCommonLength + 1 do 8: dist1 = maxCommonIndexes1(i) – last1 – 1 9: dist2 = maxCommonIndexes2(i) – last2 – 1 10: if (dist1 > dist2) then 11: noPoints = dist2 12: noDeltas = dist1 – dist2 13: else 14: noPoints = dist1 15: noDeltas = dist2 – dist1 16: end if 17: totalPoints = totalPoints + noPoints 18: totalDeltas = totalDeltas + noDeltas 19: last1 = maxCommonIndexes1(i) 20: last2 = maxCommonIndexes2(i) 21: end for 22: return totalPoints + totalDeltas*deltaValue Because of the sequential nature of business processes, we want to extract K similar sequences in terms of their temporal characteristics and not in terms of numerical quantities. This can be seen as a string comparison problem. Our idea is to adopt the sequence alignment approach from biology and combine it with KNNs. The resulting approach is named K nearest sequence with sequence alignment (KnsSA). By combining with sequence alignment, KNNs allow us to sequentially compare symbolic sequences. Given N sequences, our approach builds a distance matrix (dij)N * N Algorithm 1 Function scan(row, col, length): identifies the optimal editing process to match two sequences via deletion/ insertion of symbols 1: commonIndexes1(length) = row 2: commonIndexes2(length) = col 3: if (length > maxCommonLength) then ▹ store the longest path 4: maxCommonLength = length 5: for i = 1 : maxCommonLength do 6: maxCommonIndexes1(i) = commonIndexes1(i) 7: maxCommonIndexes2(i) = commonIndexes2(i) 8: end for 9: maxCommonScore = 10: calculateScore(n,maxCommonIndexes1,maxCommonIndexes2,maxCommonLength) 11: else 12: if (length == maxCommonLength) then 13: score = calculateScore(n,commonIndexes1,commonIndexes2, length) 14: if (score < maxCommonScore) then ▹ if the paths are equal, store the smallest score 15: for i = 1 : maxCommonLength do 16: maxCommonIndexes1(i) = commonIndexes1(i) 17: maxCommonIndexes2(i) = commonIndexes2(i) 18: end for 19: maxCommonScore = score 20: end if 21: end if 22: end if © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 storing the alignment scores for matching any two sequences. The elements of the distance matrix are then used to rank the sequences. The elements of the distance matrix are obtained by matching every pair of sequences using local alignment. 4. Evaluation Having defined MSA and KnsSA, we now present a discussion of our empirical evaluation. We assess the efficiency of our approaches and explore potential future research directions for employing our Markov extension approach in business process event forecasting and our sequential symbolic KNN approach in business process outcome forecasting. Our models have been implemented in MATLAB and run against four data sets of process instances from the records of a major telecommunication company. The first data set (DS1) consists of telecommunication line fault repair records for a 9-month period. The second (DS2) covers a 1-month period and represents a process for fixing broadband faults. The third data set DS3 is from a different fault repair process. Data sets DS1 and DS2 are used in the experiments of our extended Markov model. Data sets DS3 and DS4 are used in the KnsSA experiments. 4.1. Process analysis and data preprocessing In our data sets, the population of process instances turned out to be very diverse and not straightforward to work with. For instance, DS1 is very diverse and contains 28963 entries, 2763 unique process instances and 285 unique tasks. The process instance length varies from 1 to 78. On the other hand, DS2 represents a significantly smaller scale set with only 5194 entries, 794 process instances and 10 unique tasks. The process instance length varies from 2 to 32. DS3 represents a small scale set with only 10000 entries, 633 process instances and hundreds of unique tasks. DS4 is also a real process with available labels with 11839 entries, 571 process instances with different lengths and hundreds of unique tasks. The length of process instances in DS3 and DS4 varies considerably. Figure 2 shows a visualisation of 10% of set DS1. The diagram has been created by BT’s process mining tool Aperture (Taylor et al., 2012) that uses workflow data to create a process diagram. For this process, the diagram illustrates the complexity of the workflow data. It is basically impossible to visually analyse or understand the process from this figure. This is a typical scenario for processes that have evolved over time and are poorly documented. 4.2. Extension Markov model experiments To evaluate MSA, we benchmarked our model with three other approaches: • RM – random model: In order to find the next task following the current one, we randomly select from the set of potential next tasks. For example, if from the Figure 2: Process model obtained by using Aperture for DS1 visualising a highly complex process. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 historical data, we know that tasks following A belong to the set {C, D, E}, we randomly select a value from that set as the predicted next step. • All Kth Markov models: A number of different order Markov models are generated from first order to kth order. Given a sequence, we start with the highest order Markov model. If the given sequence cannot be found in the transition matrix, we create a new shorter sequence by removing the first event in the sequence. We then continue the procedure with the next lower order Markov model until we either find a match in a transition matrix or after trying the first order Markov model, a default prediction is required. • HMM – hidden Markov models: We tested several first order HMMs with different lengths for the input sequences and different number of hidden states and picked the best one. For each comparison, we used 90% entries of the data sets as training data and the remaining 10% were used as test data. Basically, for each sequence in the test data, we checked if our model can correctly predict the next step of a particular task (i.e. we measured the number of successful predictions). The results were evaluated using 10-fold cross-validation. The results are displayed as a percentage. After the first order to seventh order MSAs were built, we investigated different values for l, the number of sequences (patterns) in the transition matrix that is most similar to the given sequence. l was varied from 1 to 7, and the results show that l = 5 is optimal for DS1 and l = 3 is optimal for DS2. We now look at the specific results. Figures 3 and 4 illustrate MSAs accuracy against the two data sets, DS1 and DS2, respectively. The results show that by incorporating the default prediction improvement module, MSA outperforms other comparable models, especially when the order of the Markov model increases. This is because the relevance of the comparison between sequences is directly proportionate to their lengths. As can be seen, the second order MSA performs best among a range of different orders of MSA. For the case of DS1, it is about 27% correct. The third order MSA works best in the case of DS2, it provides correct predictions in about 70% of the cases. Nonetheless, the highest performance order depends on the data set. For the data set that provides a good performance with a relatively high Markov order (i.e. fifth and above), the role of our default prediction improvement module becomes highly significant. Figure 5 shows the performances of all models against the two data sets. As can be seen, RM performs worst with only around 10% success for DS2. When applied to the bigger data set DS1, the result goes down to nearly 2% (14 times worse than that of MSAs’ average performance). This can be explained by the fact that with DS2, each task can have an average of seven potential next tasks, whereas this figure is nearly 30 for DS1. Thus, the probability of picking the right next task reduces as the set size increases. The results highlight the difficulty of handling complex data sets. When the data is not too diverse (i.e. DS2), MSA (fifth order) obtains the highest number of correct predictions with a result of 63%. Compared with other benchmarks, MSA results are better than a fifth-order Markov model that achieves 57%, an All Kth (K = 5) Markov model that achieves 60% and an HMM that achieves 43%. Please note that we did not pick the most Figure 3: Percentage of correct predictions before and after applying the default prediction module in Markov models using data set DS1. Figure 4: Percentage of correct predictions before and after applying the default prediction module in Markov models using data set DS2. Figure 5: Percentage of correct predictions of different models on DS1 and DS2 data sets. © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 accurate MSA and All Kth Markov model to compare with the HMM and the RM. The reason is that All Kth Markov models and MSAs only have an advantage with higher orders. In order to see how well they cope with default predictions, we chose a fifth order MSA and an All fifth Markov model. When the data is more fragmented as in the case of DS1, the effectiveness of our model is reduced. However, even though the performance on DS1 is not as good as the performance on DS2, our model still performs an important role considering the requirement of selecting a single correct task from 30 potential candidates on average. Even in this case, we still manage to outperform the HMM result by 20% and also outperform the All fifth Markov model result by 3%. 4.3. Sequential KNN experiments 4.3.1. Data preparation We carried out a number of experiments based on records from two real processes (DS3 and DS4). As we aim to predict the process outcome, it is necessary to label the outcome as either success or failure, and this needs to be carried out before the proposed method can be applied. There is, however, an issue in our available process data. Records that are labelled with outcomes are not available for many processes. For some processes, it is not always trivial to define process success or failure. To this end, we have to look for different suitable criteria to classify process instances into success/failure for different data sets. One such criterion is the duration used to complete a process instance. That is, to label the process instance as a success if the time duration is under a chosen threshold, otherwise it is labelled as a failure. In the case of DS4, however, the difference between the actual delivered date and the delivered date promised to the customer is used as the criterion to determine the success and failure. Particularly, if the actual delivered date is before the agreed date, then that process instance is classified as success, otherwise it is classified as failure. 4.3.2. Results To evaluate KnsSA, we benchmarked our models with two other approaches: • RM – random model: In order to find the outcome of the process, we randomly generate a number between 0 and 1; if the generated number is greater than 0.5, the outcome is success (1) and vice versa; if the generated number is smaller than 0.5, the outcome is failure (0). • Original KNN: We chose K nearest sequences in terms of having common unique tasks. For example, given two sequences A, B, D and A, A, C, there are one A, one B and one D in the first sequence; there are two As and one C in the second one. Each unique task can be considered as one category. The distance for each category is computed as the difference in the number of occurrences of the corresponding task. Then, the sum over all categories is taken in order to obtain the total distance between any two given sequences. For instance, the two sequences given previously consist of four categories A, B, C and D. The distance for category A is dA = 1 and those of categories B, C, D are dB = 1, dC = 1 and dD = 1, respectively. The resulting total distance of these two sequences is d = dA + dB + dC + dD = 4. For the proposed models, we investigate the effect of K as it is important to obtain the reasonable number of similar sequences. As the labels are 0 and 1, we decide to select odd values for K so we can always extract the outcome/ label of the given sequence based on the K obtained sequences. The data sets DS3 and DS4 have a large number of unique tasks and the difference between the lengths of the sequences is substantial. Intuitively, the value of K should be small taking into account the diversity of the data. The results of the local KnsSA applied to data sets DS3 and DS4 are presented in the Figure 6. The performances of the original KNN, random guess and the proposed models, local KnsSAs, applied to the two data sets are presented in Figure 7. The results show that the proposed model outperforms both benchmarking models original KNN and random guess. This also implies that the temporal characteristics of the data are important for predicting the process outcome. 5. Conclusions In this paper, we have introduced several models for predicting the next step in business processes in a telecommunication services context where no formal description of the processes exists or the real process significantly deviates from the Figure 6: Percentage of correct predictions of local KnsSA using data sets DS3 and DS4. Figure 7: Percentage of correct predictions of different models on data sets DS3 and DS4. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 designed path prototype when being applied in real environments and under realistic constraints. Specifically, we first applied Markov and hidden Markov models to generate probability matrices for moving from one event to the next and then using them to provide predictions. Next, we have proposed a novel predictive model (MSA) that is an extension of Markov models employing sequence alignment. In order to analyse its effectiveness, we have empirically evaluated MSA against two data sets from a major telecommunication company with varying degrees of diversification. The results have clearly demonstrated that MSAs outperform both Markov models and HMMs by at least 10%. In addition, we have also shown that higher order Markov models outperform a first order Markov model. The default prediction module we introduced significantly improves the corresponding original Markov model result when the order of the model is high (starting from the fifth order Markov model). Nonetheless, although the improvement provided by the new default prediction module is quite significant, higher order MSAs do not outperform second order MSA in the case of the data set DS1 and third order MSA in the case of the data set DS2. Another contribution of the paper is that we propose a new sequential KNN for symbolic sequences. This proposed method outperforms the benchmarking models and proves that sequential characteristics of the data play an important role in predicting the process outcome. These experimental results motivate us to use sequence alignment technique as a similarity measure to group sequences that have similar characteristics and then tackle them separately. Our future research will look into using sequence alignment and K-means clustering to cluster data into K groups and then treat each group with suitable approaches. References VAN DER AALST, W. (2011) Process Mining: Discovery, Conformance and Enhancement of Business Processes, New York Icn: Springer-Verlag. BAUM, L.E., T. PETRIE, G. SOULES and N. WEISS (1970) A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains, The Annals of Mathematical Statistics, 41, 164–171. BERRY, M.J.A. and G.S. LINOFF (2004) Data Mining Techniques: For Marketing, Sales, and Customer Relationship Management, New York: Wiley. COX, L.A. JR. and D.A. POPKEN (2002) A hybrid system- identification method for forecasting telecommunications product demands, International Journal of Forecasting, 18, 647–671. DESHPANDE, M. and G. KARYPIS (2004) Selective Markov models for predicting web page accesses, ACM Transactions on Internet Technology (TOIT), 4, 163–184. EASTWOOD, M. and B. GABRYS (2009) A non-sequential representation of sequential data for churn prediction, in Proceedings of the KES2009 Conference, Santiago, Chile. EIRINAKI, M., M. VAZIRGIANNIS and D. KAPOGIANNIS (2005) Web path recommendations based on page ranking and Markov models, WIDM’05, 2–9. NEEDLEMAN, S.D. and C.D. WUNSCH (1970) A general method applicable to the search for similarities in the amino acid sequence of two proteins, Journal of Molecular Biology, 48, 443–453. NETZER, O., J. LATTIN and V.S. SRINIVASAN (2007) A hidden Markov model of customer relationship dynamics, Marketing Science, 27, 185–204. PAPOULIS, A. and S.U. PILLAI (1991) Probability, Random Variables and Stochastic Processes, New York: McGraw-Hill. PITKOW, J. and P. PIROLLI (1999) Mining longest repeating subsequences to predict world wide web surfing, in The 2nd USENIX Symposium on Internet Technologies & Systems, Colorado, USA, 139–150. RUTA, D. and B. MAJEED (2011) Business process forecasting in telecommunication industry, in GCC Conference and Exhibition (GCC), 2011 IEEE, Dubai, United Arab Emirates, 389–392. RUTA, D., D. NAUCK and B. AZVINE (2006) K nearest sequence method and its application to churn prediction, in Intelligent Data Engineering and Automated Learning IDEAL 2006, volume 4224 of Lecture Notes in Computer Science, Corchado, E., Yin, H., Botti, V. and Fyfe, C. (eds). Berlin: Springer, 207–215. SMITH, T.F. and M.S. WATERMAN (1981) Identification of common molecular subsequences, Journal of Molecular Biology, 147, 195–197. TAYLOR, P., M. LEIDA and B.A. MAJEED (2012) Case study in process mining in a multinational enterprise, in Lecture Notes in Business Information Processing, Berlin: Springer. THEDE, S. and M. HAPPER (1999) A second-order hidden Markov model for part-of-speech tagging, in The 37th Annual Meeting of the ACL, Maryland, USA, 175–182. WATERMAN, M. (1994) Estimating statistical significance of sequence alignments, Philosophical Transactions of the Royal Society of London, Series B: Biological Sciences, 344, 383–390. The authors Mai Le Mai Le is currently a research assistant at Bristol University. She is working on a collaborative project between Bristol University and BT’s Research and Innovation Division. Her research focuses on data analytics and visual analytics that enable autonomous systems to make decisions in unusual situations based on human judgment and on data from previous similar situations. Mai obtained her MSc in optimisation and modelling from the University of Paris 6, in 2007. She is completing her PhD at Bournemouth University in mathematical modelling and predictive analysis of temporal sequential data. Bogdan Gabrys Bogdan Gabrys received an MSc in electronics and telecommunication from Silesian Technical University, Poland, in 1994, and a PhD in computer science from Nottingham Trent University, UK, in 1998. After many years of working at different Universities, he moved to the Bournemouth University in UK in January 2003, where he holds a Chair in Computational Intelligence position and acts as the founding director of an interdisciplinary Data Science Institute. His current research interests lay in a broad area of intelligent, complex adaptive systems and include a wide range of machine learning, biologically/nature inspired learning and hybrid intelligent techniques encompassing data and © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 information fusion, learning and adaptation methods, multiple classifier and prediction systems, processing and modelling of uncertainty in predictive modelling, pattern recognition, diagnostic analysis and decision support systems. Detlef Nauck Detlef Nauck is chief research scientist for Data Science with BT’s Research and Innovation Division located at Adastral Park, Ipswich, UK. He is leading a group of international scientists working on Intelligent Data Analysis and Autonomic Systems. Detlef has been working on data analytics projects across a number of areas such as customer service analytics, real-time business intelligence, business process analytics and network analytics. His current work focusses on introducing autonomic principles like self-learning, self-healing and self-optimisation into business processes and network management. Detlef is a visiting professor at Bournemouth University and a private docent at the Otto-von-Guericke University of Magdeburg, Germany. Detlef holds an MSc (1990) and a PhD (1994) in computer science both from the University of Braunschweig, Germany. He also holds a Habilitation (postdoctoral degree) in computer science from the Otto- von-Guericke University of Magdeburg, Germany (2000). © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 
work_te3lxb5txberbfhhmt7funqfuy	----	2013_11_01_Gonciarz.pub Management Systems in Production Engineering   2013, No 3 (11), pp 5‐8       Abstract:  Expert systems can be defined as computer programs, whose main task is to simulate a human expert, usually in a nar‐ row field of exper se. Possible applica ons of modern informa on technology are very extensive, ranging from medici‐ ne, geology and technology to applica ons in the field of economic and financial decision support.  The purpose of this paper is to present the prac cal applica on of an expert system that supports the process of mana‐ ging the produc on of yachts and has a high suitability for use in this applica on.  Using the expert system described in the paper reduces the  me during the design and produc on prepara on process.    AN EXPERT SYSTEM FOR SUPPORTING THE PRODUCTION OF PLEASURE BOATS  INTRODUCTION  In the highly compe ve situa on of the yacht market,  it is obvious that success will reward  those companies that  are be er able than the compe on to make the right deci‐ sions in a  mely and efficient manner. Large companies can  afford to employ a numbers of highly qualified profession‐ als to respond properly at any  me and take the right deci‐ sions.  In addi on, a financially strong company can afford  to purchase custom‐wri en and highly advanced so ware  that constantly analyzes the main func ons of the compa‐ ny.  Small  and  medium‐sized  enterprises  are  in  a  different  situa on where budgets are much slimmer and it is difficult  to  compete  with  larger  corpora ons.  It  makes  sense  for  these  small  and  medium  sized  en es  to,  look  for  other  ways to improve their compe veness. Among the alterna‐ ves are expert systems that can successfully support work  in even the most esoteric of yacht disciplines, such as hull  design and construc on.  Expert  systems  are  capable  of  computa onal,  qualita‐ ve, descrip ve and explanatory func ons.  An addi onal advantage of expert systems is the ease of  use, in a process which boils down to a series of ques ons  and answers between the computer program and the user,  in which the system receives relevant informa on, not only  from the user  but also from external sources of knowledge,  such as spreadsheets, and other calcula on tools.  In yacht building, the  me required for the produc on  of an individual boat is o en less important than a focus on  extremely high quality, providing excep onal comfort and  luxury to the owner and his or her guests with state of the  art furnishings and equipment. This equipment will include  electronic  and  hydraulic  systems  providing  the  ability  to  control and maneuver the vessel in adverse condi ons with  a li le or no stress and with a very small crew.   Very o en the reputa on of the boat building company  is the determining factor in se ng the boat’s final price and  producing the order to proceed with design comple on and  construc on.  A common prac ce of boat manufacturers  is to select  individual  items  of  equipment  from  recommenda ons  in  catalogs, which may well be weighted to specify all items to  be “safe” in terms of strength and power output, but which  may  also  be  unnecessarily  heavy  and  expensive.  This  has  the poten al for increasing the final cost of the vessel, and  in  some  cases  such  as  diesel  engines  for  bringing  about  premature  failure  as  a  result  of  excessive  low  power  and  low temperature opera on.  Conversely,  selec ng  under‐sized  or  under‐powered  components  will  lead  to  frequent  breakdowns,  early  re‐ placements, and  in extreme situa ons, may endanger the  lives and safety of the people on the vessel.   Yacht manufacturing costs are also greatly affected by  excessive  expenditures  for  labor    to  install  over‐  sized  equipment, well intended to ensure the safety of the yacht  and ensure the comfort of its use. A fully developed expert  system will op mally choose, customize and calculate the  required materials and equipment for safety, comfort and  cost control.  In other words, the  informa on provided by  such an expert system will significantly affect all costs, caus‐ ing the final price of the to be much lower while maintain‐ ing  the  desired  comfort  and  capabili es  required  in  the  completed yacht. In addi on, the use of this expert system  will mean that produc on will be smooth and efficient.  THE PRODUCTION PROCESS OF THE EXPERT SYSTEM  An  expert  system  is  a  computer  program,  which  con‐ tains knowledge about a specific and usually narrow field.  The system has a capability to solve problems comparable  with that of a human expert in the same filed of knowledge  Tomasz GONCIARZ  Parker Poland Sp. z o.o.  Marcin PERZYK  Warsaw University of Technology  Key words: expert systems, informa on technologies, produc on of yacht     6                                                                                                               Management Systems in Produc on Engineering 3(11)/2013                               T. GONCIARZ, M. PERZYK ‐ An expert system for suppor ng the produc on of pleasure boats    – it is a computer so ware designed to solve problems that  require  specialized  knowledge.  Expert  systems  belong  to  the field of ar ficial intelligence, which is the study of issues  including fuzzy logic, evolu onary computa on, neural net‐ works, ar ficial  life and robo cs. Ar ficial  intelligence  is a  branch of computer science, the object of which is to study  the  rules  of    human  behavior  and  intelligence,  to  create  formal  models  of  that  behavior  and  human  thought  pro‐ cesses and, as a result, to create computer programs that  simulate the behavior and intellect of humans [4, 7].  The star ng point to development of the expert system  is analysis of the process of prepara on of produc on [5].   The main idea is to create programs, when a knowledge  and  reasoning  techniques  are  introduced,  can  generate  answers similar to those that would be provided by a highly  experienced human being. In effect, you will use the system  to  access  the  human  expert’s  knowledge  and  experience  through the user interface of the computer or other device  running the program. The user of this service asks ques ons  and receives answers and explana ons presented in various  forms, such as:  text,  video,  sound,  photo,  figure,  scheme.  The benefits of using expert systems include [6]:   saving money. In the long run they are much less ex‐ pensive than human experts, especially  in helping to  solve problems that require the most exper se  a lack of experienced people in the field;  they work faster, do not get  red, they are more relia‐ ble than people;  they are consistent, objec ve, accurate;  they  analyze  large  amounts  of  data,  using  computer  capabili es of requires a computer;  they are useful for solving complex problems in areas  where there is accumulated empirical informa on ;  they are able to answer ques ons by presen ng infor‐ ma on  in  an  intui vely  understandable  way.  Expert  system  users  do  not  need  to  understand  the  inner  working of the systems;  they save  me;  they are always available;  they are easy to use.  Model of an expert system for supporting the production of recreational crafts Recrea onal cra , to which expert systems can be ap‐ plied, include sailing yachts, displacement‐hull power boats  and planing power boats. These vessels have a  long tradi‐ on, a variety of applica ons and a large variety of design  solu ons.   The presented expert system can address, among other  things, issues such as the calcula on of the maximum speed  of  the  boat  and  the  required  engine  power  to  a ain  that  speed, the choice of power train components, exhaust sys‐ tems, the selec on of steering and thruster devices to con‐ trol and maneuver the boat, and help to resolve many other  ques ons and problems. The conclusions and explana ons  are  provided  in  narra ve  and  quan ta ve  form  with  dia‐ grams, drawings, photographs, films and sound. The expert  system  is designed  in such a way that  it can be easily ex‐ panded  by  the  input  of  new  informa on  and  knowledge.  For example, as new technology develops, the system engi‐ neer  or  programmer  can  update  or  extend  the  data  base  and programming of the expert system.   This  par cular  system  has  been  wri en  in  such  a  way  that it can be easily reprogrammed for applica on in fields  other than yacht construc on because all of the substan ve  knowledge  is  contained  in  discrete,  and  therefore  inter‐ changeable,  data  files,  assembled  in  applica on  related  data bases. Each database corresponds to a specific chapter  of  the Main Menu,  to be used  in  solving  problems  in  the  business or other field to which the data base applies. With  this  architecture,  the  expert  system  design  can  easily  be  applied  to  other  disciplines  by  switching  data  bases  (sources  of  knowledge), using  appropriate access paths  in  the main control unit of the system and changing the con‐ tents of the menu control block [1, 2, 3, 8]. Of course, new  data bases must be adequate for the scale of the problems  they are intended to address, and must be forma ed to be  compa ble with the central programs of the expert system.   The expert system is built basing on:  the first author's experience,  consulta ons  with  other  programmers,  systems  ana‐ lysts and engineers,    informa on from the marine design offices and their  naval architects,   interviews  with  people  working  directly  in  manufac‐ turing,  observing and interviewing the produc on workers,  conduc ng interviews with experts,  conduc ng interviews with sales personnel ,  conduc ng  interviews  with  customers  –  end  users/ owner of the yacht or other product,  search in the materials sector, including catalogs and  websites.  Fig. 1 Schema c diagram of the expert system presented    Described expert system has been built as an applica on  of the PC‐Shell computer tool, an independent tool for con‐ struc on of expert systems [6]. It combines various meth‐ ods  of  solving  problems  and  knowledge  representa on.  It  can be used in any field, hence the range of its applica ons  is very wide. The PC‐Shell is mainly predisposed to solve the  problems of the diagnos c and classifica on and interpreta‐ on of data (Fig. 1) [6].      Management Systems in Produc on Engineering 3(11)/2013                                                                                                               7                      T. GONCIARZ, M. PERZYK ‐ An expert system for suppor ng the produc on of pleasure boats  Program’s presentation (Fig. 2) Fig. 2 The expert system welcome screen  In Fig. 3 a view of the func on which enables the calcu‐ la on of the required engine power for the boat is shown.  The  user  selects  from  the  Main  Menu  func ons  ,,Engine  power calculator” and enters values such as the length and  width of the hull at the water line, dra  and number of en‐ gines.  A er clicking  the  Calculate  bu on  the  required  en‐ gine power for the boat Ne is returned.  User  can  get  addi onal  explana ons  from  the  expert  system  which  may  help  him/her  even  at  the  stage  of  the  applica on  process.  Those  func on  is  called  ,,WHAT  IS  IT ?” (Fig. 4).      Fig. 3 Main Menu and the Engine Power Calculator dialog box  Fig. 4 Menu and addi onal explana on window ,,WHAT IS IT ?”    8                                                                                                               Management Systems in Produc on Engineering 3(11)/2013                               T. GONCIARZ, M. PERZYK ‐ An expert system for suppor ng the produc on of pleasure boats    mgr inż. Tomasz Gonciarz  Parker Poland Sp. z o.o.  ul. Reniferowa 88, 03‐289 Warszawa, POLAND  e‐mail: tomasz.gonciarz@parker.com.pl    prof. dr hab. inż. Marcin Perzyk  Warsaw University of Technology,  Faculty of Produc on Engineering   Ins tute of Manufacturing Processes  ul. Narbu a 85, 02‐524 Warszawa, POLAND  e‐mail: M.Perzyk@wip.pw.edu.pl  Assessment of the expert system in terms of its usefulness An evalua on of the expert system in terms of its con‐ cept, design and usability was made by tes ng  it  in three  independent  companies  in  the  marine  industry.  The  first  two  are  leading  Polish  manufacturers  of  boats,  which  for  many years have engaged in the commercial produc on of  recrea onal cra s. They make mass market products and  also custom, one of a kind, products. These products go to  the Polish market as well as for export. The third company,  which  tested  the  expert  system,  is  a  marine  engineering  firm providing boat building services, maintenance and the  installa on  of  equipment  to  yacht  owners  and  operators.  This very special enterprise enjoys a great reputa on in the  yach ng industry.  The staff of these three companies, the expert system  was  tested  by  a  selected group  of  experts  and  specialists  who,  on  a  daily  basis,  work  in  manufacturing,  in  pre‐ produc on,  and  as  designers  and  technicians  responsible  for the assembly process.  In the selec on process of the experts, the competence  of each candidate was carefully ve ed. The experts invited  to par cipate in the study did not know each other and had  not previously interacted. Each expert was required to de‐ vote  minimum  of  twenty  hours  to  tes ng.  In  these  inde‐ pendent tests, the expert system received very high marks  and  posi ve  recommenda ons  from  all  the  companies  in  which the product was tested.   Inter alia, the expert system was used in the design and  produc on of a new model displacement‐hull motor yacht,  which  is  intended  for  recrea onal  boa ng  on  inland  and  coastal waters.   Responses  obtained  from  the  expert  system  were  checked  and  compared  with  the  professional  litera‐ ture by the people who tested the program.  The  system  was  found  to  be  a  highly  useful  program  due to its following characteris cs:  easy to find informa on,  interface transparency,  speed of responses,  the program includes extensive knowledge.  The tested expert system has helped to shorten the pre‐ produc on phase of the manufacturing process and result‐ ed in a drama c reduc on in final vessel costs by op miz‐ ing the selected equipment package. It has made a unique  contribu on to manufacturing efficiency.   CONCLUSIONS  The knowledge of a single expert may not be as exten‐ sive  as  the  informa on  contained  in  a  tradi onal  wri en  book. However, it is easier to search for the informa on in  an expert system than in literature and, as a consequence,  a person can work more efficiently. An expert system is an  alterna ve  to  wri en  instruc ons  and  other  informa on  sources. The expert system discussed  in this paper  is well  developed  and  highly  prac cal.  It  streamlines  the  pre‐ produc on  processes,  as  has  been  demonstrated  by  the  manufacturers who have tested the program.  REFERENCES  [1]  Buchalski  Z.:  Computer  Advisory‐Decision  System  for  the  Logis cs  Services  Support.  Polish  Journal  of  Envi‐ ronmental Studies, Vol. 18, No 3B, 2009, s 53‐57.  [2]  Buchalski Z.: Knowledge Management of Expert System  Based on the Symbolic Representa on of Natural Lan‐ guage Sentences. [w:]: Borzemski L., Grzech A., Świątek  J., Wilimowska Z. (red.): Informa on Systems Architec‐ ture and Technology.  Oficyna Wydawnicza Politechniki  Wrocławskiej. Wrocław 2006, s. 75‐85.  [3]  Buchalski Z.: Zarządzanie wiedzą w podejmowaniu de‐ cyzji  przy  wykorzystaniu  systemu  ekspertowego.  [w:]  Bazy danych. Struktury, algorytmy, metody. WKŁ. War‐ szawa 2006, s. 471‐478.  [4]  Koch C., Tononi G.: Test for consciousness. The World  of Science. No 7 (239), 2011, pages 32‐35.  [5]  Knosala  R.  i  in.:  Zastosowania  metod  sztucznej  inteli‐ gencji w  inżynierii produkcji. WNT. Warszawa 2002, s.  2.  [6]  Michalik K.: Integrated package of ar ficial intelligence  Aitech  Sphinx 4.5. Aitech  Intelligent  Laboratory.  Kato‐ wice 2006, pages 1‐8.  [7]  Mulawka  J.:  Systemy  ekspertowe.  WNT.  Warszawa  1996, s. 20.  [8]  Tkaczyk  W.:  Inżynieria  Wiedzy.  Akademicka  Oficyna  Wydawnicza EXIT. Warszawa 2010, s. 113.  << /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Dot Gain 20%) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Error /CompatibilityLevel 1.4 /CompressObjects /Tags /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.1000 /ColorConversionStrategy /CMYK /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams false /MaxSubsetPct 100 /Optimize true /OPM 1 /ParseDSCComments true /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo true /PreserveFlatness true /PreserveHalftoneInfo false /PreserveOPIComments true /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Apply /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true ] /NeverEmbed [ true ] /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 300 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages false /ColorImageDownsampleType /Bicubic /ColorImageResolution 450 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.50000 /EncodeColorImages false /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /ColorImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 300 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages false /GrayImageDownsampleType /Bicubic /GrayImageResolution 540 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages false /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /GrayImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 1200 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages false /MonoImageDownsampleType /Bicubic /MonoImageResolution 1200 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages false /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 >> /AllowPSXObjects false /CheckCompliance [ /None ] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile () /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description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> /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000410064006f006200650020005000440046002065876863900275284e8e9ad88d2891cf76845370524d53705237300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef69069752865bc9ad854c18cea76845370524d5370523786557406300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /CZE <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> /DAN <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> /DEU <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> /ESP <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> /ETI <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> /FRA <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> /GRE <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a stvaranje Adobe PDF dokumenata najpogodnijih za visokokvalitetni ispis prije tiskanja koristite ove postavke. Stvoreni PDF dokumenti mogu se otvoriti Acrobat i Adobe Reader 5.0 i kasnijim verzijama.) /HUN <FEFF004b0069007600e1006c00f30020006d0069006e0151007300e9006701710020006e0079006f006d00640061006900200065006c0151006b00e90073007a00ed007401510020006e0079006f006d00740061007400e100730068006f007a0020006c006500670069006e006b00e1006200620020006d0065006700660065006c0065006c0151002000410064006f00620065002000500044004600200064006f006b0075006d0065006e00740075006d006f006b0061007400200065007a0065006b006b0065006c0020006100200062006500e1006c006c00ed007400e10073006f006b006b0061006c0020006b00e90073007a00ed0074006800650074002e0020002000410020006c00e90074007200650068006f007a006f00740074002000500044004600200064006f006b0075006d0065006e00740075006d006f006b00200061007a0020004100630072006f006200610074002000e9007300200061007a002000410064006f00620065002000520065006100640065007200200035002e0030002c0020007600610067007900200061007a002000610074007400f3006c0020006b00e9007301510062006200690020007600650072007a006900f3006b006b0061006c0020006e00790069007400680061007400f3006b0020006d00650067002e> /ITA <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> /JPN <FEFF9ad854c18cea306a30d730ea30d730ec30b951fa529b7528002000410064006f0062006500200050004400460020658766f8306e4f5c6210306b4f7f75283057307e305930023053306e8a2d5b9a30674f5c62103055308c305f0020005000440046002030d530a130a430eb306f3001004100630072006f0062006100740020304a30883073002000410064006f00620065002000520065006100640065007200200035002e003000204ee5964d3067958b304f30533068304c3067304d307e305930023053306e8a2d5b9a306b306f30d530a930f330c8306e57cb30818fbc307f304c5fc59808306730593002> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020ace0d488c9c80020c2dcd5d80020c778c1c4c5d00020ac00c7a50020c801d569d55c002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /LTH <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> /LVI <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> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken die zijn geoptimaliseerd voor prepress-afdrukken van hoge kwaliteit. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <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> /POL <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> /PTB <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> /RUM <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> /RUS <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> /SKY <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> /SLV <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> /SUO <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> /SVE <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> /TUR <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> /UKR <FEFF04120438043a043e0440043804410442043e043204430439044204350020044604560020043f043004400430043c043504420440043800200434043b044f0020044104420432043e04400435043d043d044f00200434043e043a0443043c0435043d044204560432002000410064006f006200650020005000440046002c0020044f043a04560020043d04300439043a04400430044904350020043f045604340445043e0434044f0442044c00200434043b044f0020043204380441043e043a043e044f043a04560441043d043e0433043e0020043f0435044004350434043404400443043a043e0432043e0433043e0020043404400443043a0443002e00200020042104420432043e04400435043d045600200434043e043a0443043c0435043d0442043800200050004400460020043c043e0436043d04300020043204560434043a0440043804420438002004430020004100630072006f006200610074002004420430002000410064006f00620065002000520065006100640065007200200035002e0030002004300431043e0020043f04560437043d04560448043e04570020043204350440044104560457002e> /ENU (Use these settings to create Adobe PDF documents best suited for high-quality prepress printing. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.) >> /Namespace [ (Adobe) (Common) (1.0) ] /OtherNamespaces [ << /AsReaderSpreads false /CropImagesToFrames true /ErrorControl /WarnAndContinue /FlattenerIgnoreSpreadOverrides false /IncludeGuidesGrids false /IncludeNonPrinting false /IncludeSlug false /Namespace [ (Adobe) (InDesign) (4.0) ] /OmitPlacedBitmaps false /OmitPlacedEPS false /OmitPlacedPDF false /SimulateOverprint /Legacy >> << /AddBleedMarks false /AddColorBars false /AddCropMarks false /AddPageInfo false /AddRegMarks false /ConvertColors /ConvertToCMYK /DestinationProfileName () /DestinationProfileSelector /DocumentCMYK /Downsample16BitImages true /FlattenerPreset << /PresetSelector /MediumResolution >> /FormElements false /GenerateStructure false /IncludeBookmarks false /IncludeHyperlinks false /IncludeInteractive false /IncludeLayers false /IncludeProfiles false /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe) (CreativeSuite) (2.0) ] /PDFXOutputIntentProfileSelector /DocumentCMYK /PreserveEditing true /UntaggedCMYKHandling /LeaveUntagged /UntaggedRGBHandling /UseDocumentProfile /UseDocumentBleed false >> ] >> setdistillerparams << /HWResolution [2400 2400] /PageSize [612.000 792.000] >> setpagedevice 
work_tgelto4n6zhcrhbncjs6tjthfa	----	Giuseppina Gini Giuseppina Gini Associate professor - Dipartimento Elettronica, Informazione e Bioingegneria , Politecnico di Milano , piazza L. da Vinci 32, I-20133 MILANO Biographical sketch to the same page in Italian List of publications (a few to download) Research Robotics and AI Focus on humanoid robotics Focus on e_science Teaching and tutoring Software download: SARpy for chemoinformatics Software download: poliGMDH neural net Software download: IReNNS structure learning networks   EU projects: http://www.vega-qsar.eu http://www.life-prosil.eu http://www.caleidos-life.eu http://www.antares-life.eu Member EUCOG II © G. Gini - L. Quartarone please send your comments to giuseppina.gini@polimi.it 
work_tgesidkd55afvnhcxpgtueo52e	----	Intelligent impact assessment of HRM to the shareholder value Intelligent impact assessment of HRM to the shareholder value George Xirogiannis a,*, Panagiotis Chytas b,1, Michael Glykas c,2, George Valiris b,1 a University of Piraeus, Department of Informatics, 80, Karaoli and Dimitriou Street, 185 34 Piraeus, Greece b University of Aegean, Department of Business Administration, 8, Michalon Street, Chios 82 100, Greece c University of Aegean, Department of Financial and Management Engineering, 31, Fostini Street, Chios 82 100, Greece Abstract Despite the extensive research in human capital management and performance measurement, intelligent treasoning mechanisms, which integrate human resource (HR) practices into strategic-level shareholder decisions, are still emerging. This paper discusses a novel approach of designing a decision-modeling tool, which assesses the impact of contemporary human resource management (HRM) prac- tices to the shareholder value and satisfaction. The underlying research addresses the problem of establishing HRM interrelationships in order to drive the overall business performance from the shareholder value perspective. The proposed methodology tool utilizes the fuzzy causal characteristics of fuzzy cognitive maps (FCMs) to generate a hierarchical and dynamic network of interconnected HR perfor- mance drivers. The intelligent computing characteristics of FCMs are also employed to establish a dynamic feedback and bi-directional alignment of HRM practices and strategic improvement. Finally, this research provides a practical insight on the applicability of soft approaches in capturing and illustrating the effect of HRM practices. � 2007 Elsevier Ltd. All rights reserved. Keywords: Fuzzy cognitive maps; HR performance; Decision support; Strategic planning 1. Introduction Enterprises are evolving in turbulent and equivocal envi- ronments (e.g. Drucker, 1993; Grove, 1999; Kellys, 1998). This requires enterprises to be alert and watchful for the detection of weak signals (e.g. Ansoff, 1975) or discontinu- ities of emerging threats and to initiate further probing based on such detection (Walls et al., 1992). Enterprises today face critical business challenges (Ulrich, 1998) like globalization, profitability through growth, technology integration, intellectual capital management, continuous change, etc. Such challenges require organizations to build new capabilities, but it is not always apparent who should be responsible for developing those capabilities. Perhaps, everyone and no one, but in any case this is a unique HR’s opportunity to play a leadership role in enabling organizations to meet such competitive challenges. Ensur- ing that human resource (HR) strategies are in place to deal with these challenges is increasingly recognized as critical to success (Leopold et al., 1999). Human resource management (HRM) in the literature has been considered a second- or third-order strategy, lar- gely related to implementation rather than shareholder level decision-making. The process of HR strategy formu- lation and evaluation had not been widely conceptualized until recently. Moreover, the impact of HRM practices to the shareholder strategic value is not modeled adequately, despite the utilization of sophisticated performance evalua- tion mechanisms at the employee level. The evidence that HR issues are fundamental to business is compelling at the level of unit labor costs, but whether they are funda- mental to the strategy process remained highly question- able until recent years (Ritson, 1999). This can be attributed to the fact that contemporary performance eval- uation mechanisms focus on analyzing the operational 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.08.103 * Corresponding author. Tel.: +30 210 4142000; fax: +30 210 4142328. E-mail addresses: georgex@unipi.gr (G. Xirogiannis), p.chytas@chios. aegean.gr (P. Chytas), mglikas@aegean.gr (M. Glykas), gval@aegean.gr (G. Valiris). 1 Tel.: +30 22710 35000; fax: +30 22710 35099. 2 Tel.: +30 22710 35400; fax: +30 22710 35499. www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 35 (2008) 2017–2031 Expert Systems with Applications mailto:georgex@unipi.gr mailto:p.chytas@chios. aegean.gr mailto:p.chytas@chios. aegean.gr mailto:mglikas@aegean.gr mailto:gval@aegean.gr effectiveness of the human capital rather than addressing the issue of strategic alignment. This paper addresses the problem of designing a novel modeling methodology tool to act as an intelligent decision support mechanism for evaluating the impact of HRM practices to the shareholder value and satisfaction. This attempt focuses on bridging the gap between the opera- tional characteristics of HRM practices and the strategic decisions at the shareholder level. The proposed methodol- ogy tool utilizes the fuzzy causal characteristics of fuzzy cognitive maps (FCMs) to generate a hierarchical and dynamic network of interconnected HR performance driv- ers. By using FCMs, the proposed novel mechanism simu- lates the operational efficiency of HR models with imprecise relationships and quantifies the impact of HRM activities to the overall shareholder satisfaction model. This research builds on contemporary HRM experience to establish a ‘‘soft computing’’ approach on how to inter- relate HRM activities and shareholder value. The FCM approach does not pose as a substitute of traditional HRM operations nor does it offer an alternative to HR performance evaluation. It presents an intelligent deci- sion-making framework for strategic-level HRM based on scenario building and ex ante impact assessment. Primarily, the proposed model targets the principle directors and stakeholders of HRM projects (e.g. HR department, change management leaders, business strategy leaders, etc) assisting them to reason effectively about the status of strategic-level performance metrics, given the (actual or hypothetical) implementation of a set of HR practice changes. However, the holistic nature of the pro- posed model may couple effectively with other strategic performance evaluation systems. Given the demand for effective shareholder positioning, such a succinct mecha- nism of conveying the essential dynamics of HR practices is believed to be useful for anyone contemplating or under- taking a strategy formulation exercise. Nevertheless, the explanatory nature of the mechanism can prove to be use- ful in a wider educational setting. As far as the contribution to decision-making is con- cerned, the application of FCMs as an intelligent modeler of HR knowledge is believed to be novel. Moreover, this paper extends typical FCM algorithms in order to adapt to the distributed nature of typical HR activities. Also, this research adopts a new qualitative approach to interpret fuzzy linguistic variables to weight and concept values in order to support further the soft computing characteristics of the tool. It is the belief of this paper that the fuzzy rea- soning capabilities enhance considerably the usefulness of the proposed mechanism while reducing the effort of iden- tifying quantitative impact measurements. As far as the added value of this research is concerned, the proposed methodology offers an alternative approach to HRM based on shareholder value analysis and modeling. Preliminary experiments indicate that the proposed approach can be effective and realistic, without employing detailed quantita- tive calculations. This paper consists of six sections. Section 2 presents a short literature overview. Section 3 presents an overview of the proposed system, Section 4 discusses the new approach to HRM modeling and Section 5 discusses the major advan- tages of the proposed tool. Finally, Section 6 concludes this paper and briefly discuses future research activities. 2. Literature overview 2.1. Contemporary HRM As firms become increasingly aware that people are among their most valuable strategic assets, they are reap- praising the way in which they manage their human capital. The emphasis is shifting from personnel management to the wider, strategic concept of human resource manage- ment in which human resource policies and activities, including training and development, are linked closely to business strategy. HR specialists who wish to develop a strategic approach to people management must establish credibility with top management as the key figures to achieve successful results (Handy et al., 1989). Strategic human resource management addresses a num- ber of key issues. Typical examples (loosely adapted from Baker, 1999) include internal integration of personnel pol- icies, their external integration with overall strategy, line management responsibility for HR implementation, indi- vidual rather than collective employee relations, HR com- mitment, HR initiatives, the managerial role of ‘‘enabler’’, ‘‘empowerer’’, and ‘‘facilitator’’, etc. Following a resource-based view of an enterprise, firms can develop sustainable competitive advantage by creating value in a way that is rare and difficult for competitors to imitate (Baker, 1999). Although traditional sources of com- petitive advantage such as natural resources, technology, economies of scale, value creation, and so forth, the resource-based argument is that these sources are increas- ingly easy to imitate, especially in comparison to a complex social structure such as an employment system. If that is so, HR strategies may be an important source of sustainable competitive advantage (Lado & Wilson, 1994). But one may ask how HR management can deliver orga- nizational excellence and competitive advantage. Accord- ing to Ulrich (1997) there are four ways to do so: • HR management could become a partner with senior and line managers in strategy execution, helping to move planning from the conference room to the marketplace. • HR management could become an expert in the way work is organized and executed, delivering administra- tive efficiency to ensure that costs are reduced while quality is maintained. • HR management could become a champion for employ- ees, vigorously representing their concerns to senior management and at the same time working to increase employee contribution; that is, employees’ commitment to the organization and their ability to deliver results. 2018 G. Xirogiannis et al. / Expert Systems with Applications 35 (2008) 2017–2031 https://isiarticles.com/article/23187 
work_tglq5qrusrcgjpwp4hvipqaeqa	----	CMSY7 Kent Academic Repository Full text document (pdf) Copyright & reuse Content in the Kent Academic Repository is made available for research purposes. Unless otherwise stated all content is protected by copyright and in the absence of an open licence (eg Creative Commons), permissions for further reuse of content should be sought from the publisher, author or other copyright holder. Versions of research The version in the Kent Academic Repository may differ from the final published version. Users are advised to check http://kar.kent.ac.uk for the status of the paper. Users should always cite the published version of record. Enquiries For any further enquiries regarding the licence status of this document, please contact: researchsupport@kent.ac.uk If you believe this document infringes copyright then please contact the KAR admin team with the take-down information provided at http://kar.kent.ac.uk/contact.html Citation for published version Kafal� , Özgür and Torroni, Paolo (2018) Comodo: Collaborative Monitoring of Commitment Delegations. Expert Systems with Applications, 105 . pp. 144-158. ISSN 0957-4174. DOI https://doi.org/10.1016/j.eswa.2018.03.057 Link to record in KAR http://kar.kent.ac.uk/66566/ Document Version Author's Accepted Manuscript COMODO: Collaborative Monitoring of Commitment Delegations Özgür Kafalıa,∗, Paolo Torronib aSchool of Computing, University of Kent Canterbury, CT2 7NF, UK bDISI, University of Bologna V.le Risorgimento, 2, 40136, Bologna, Italy Abstract Understanding accountability in contract violations, e.g., whom is accountable for what, is a tedious, time-consuming, and costly task for human decision-making, especially when contractual responsibilities are delegated among parties. Intelligent software agents equipped with expert capabilities such as monitoring and diagnosis help save time and improve accuracy of diagnosis by formal reasoning upon electronic contracts. Such contracts are represented as com- mitment norms, a well studied artifact in multi-agent systems, which provide semantics for agent interactions. Due to the open and heterogeneous nature of multi-agent systems, commitments are often violated. When a commitment is violated, e.g., an exception occurs, agents need to collaborate to understand what went wrong and which agent is responsible. We propose COMODO: a framework for monitoring commitment delegations and detecting violations. We define a complete set of possible rational delegation schemes for commitments, identifying for each combination of delegations what critical situations may lead to an improper delegation and potentially to a commitment violation. COMODO provides a sound and complete distributed reasoning procedure that is able to find all improper delegations of a given commitment. We provide the complete implementation of COMODO using the Reactive Event Calculus, and present an e-commerce case study to demonstrate its workings. Due to its generic nature, we discuss the appli- cation of our approach to other distributed diagnosis problems in emergency healthcare, Internet of Things and smart environments, and security, privacy, and accountability in the context of socio-technical systems. Keywords: Agent-based commerce, Norms, Commitment delegation, Diagnosis, Computational logic 1. Introduction A commitment describes a contract between two agents: the debtor commits to bringing about a property for the creditor (Singh, 1999). Commitments are used to give agent interaction a social semantics (Torroni et al., 2009). The idea of a social semantics is to abstract away from the agent internals and provide a social meaning to agent message exchanges. In a contract-based multi-agent system, several such commitments are in effect. Consider Amazon Prime’s ∗Corresponding author Email addresses: R.O.Kafali@kent.ac.uk (Özgür Kafalı), paolo.torroni@unibo.it (Paolo Torroni) Preprint submitted to Elsevier April 3, 2018 next-day delivery scheme: Amazon is committed to provide a one day delivery as long as the customer pays until noon. This can be represented by a conditional commitment: C(amazon, customer, paid(customer, item) ∧ prime(customer), a[1, 12], delivered(item), a[12, 36]) where paid(customer, item)∧prime(customer) is the condition enabling the commitment (antecedent), and a[1, 12] specifies the absolute time window for the payment (in this example, each unit represents one hour). The conditional commitment states that, if the customer pays until noon and is an Amazon Prime customer, then amazon (the debtor) commits to the delivery of the item (delivered(item) is the consequent of this commitment) in the absolute time interval a[12, 36]. Now, let’s consider another commitment where the bank is committed to verify the customer’s payment in three days after the customer initiates the payment. We can represent this commitment with a relative deadline for the consequent as follows: C(bank, customer, paid, a[∞, ∞], verified, r[0, 3]). The use of commitments to model agent interaction has been advocated especially in heterogeneous and open settings where autonomous agents must interact flexibly so as to handle exceptions and seize opportunities (Yolum & Singh, 2002). In these settings, traditional representations of protocols, such as finite state machines or AUML interaction protocol diagrams, are inadequate. One reason why commitment-based approaches are more flexible than traditional approaches is that they enable the stakeholders to delegate their commitments. Delegation, the act of giving control or authority (e.g., a job, a duty, etc.) to another agent (Castelfranchi & Fal- cone, 1998; Norman & Reed, 2010), may be desirable for several reasons. For instance, agents may not be capable of satisfying the properties they are committed to bringing about. This is a very common case in e-commerce sce- narios. Amazon delegates its deliveries to a courier (e.g., UPS). In our example, a delegation of delivery by Amazon (delegator) creates a new commitment: C(ups, amazon, ⊤, a[∞, ∞], delivered, a[31, 45]) whereby a new agent, in this case UPS (delegatee), commits to Amazon to carrying out the delivery (delegandum). Note that the absence of a time reference in the antecedent is represented by a[∞, ∞]. Here, the commitment for delivery between Amazon and the customer is extended with UPS. The customer may not be aware of this extension until the delivery is completed, or something goes wrong (e.g., the deadline passes). In that case, this connection should be revealed so that if the problem is related to UPS, it can be identified. We call such problems exceptions in the sense that they do not account for the expected outcome of a commitment. Some causes of exceptions have been identified by related work as (i) a violation where a commitment is violated by its debtor (Singh, 1999), (ii) a misalignment where two commitments are not aligned with each other due to different observations of the participating agents (Kafalı & Torroni, 2012), or (iii) an improper delegation where a commitment is delegated without respecting the delegandum’s deadline (Kafalı & Torroni, 2011). This paper significantly extends our previous work on commitment delegation (Kafalı & Torroni, 2011), and 2 provides a distributed framework, COMODO, for detecting exceptions caused by inconsistencies regarding such dele- gations. When there are many commitments in the system at hand, in order to identify an exception we need effective ways to explore the space of commitments. In particular, we need to identify links between commitments and exclude irrelevant instances from our search. The process of tracking individual commitment states is called commitment monitoring (Chesani et al., 2009). We extend monitoring to enable tracking down exceptions by following the links between commitments of different agents. To this end, we define a language for commitments and deadlines, which enables modeling a variety of interesting e-commerce situations. The language allows us to define properties as con- junctions of atomic propositions. Properties are associated with absolute or relative deadlines. We define a complete set of possible rational delegation schemes, identifying for each combination of delegations what critical situations may lead to improper delegations and possibly to commitment violations. A naive way of diagnosing exceptions would be to scan the set of commitments relative to a given transaction, and to look for the property that has been violated. However, this does not solve the problem if the property changes because of delegations, and if some knowl- edge is local to agents. Thus, we need to identify types of delegations and define a distributed algorithm that only makes use of the knowledge that is locally available to the agent. COMODO provides such an algorithm. We prove that our algorithm is sound and complete. Alongside with the theoretical results, we also provide an implementation for COMODO based on the Reactive Event Calculus. Finally, we present a case study to demonstrate its workings. Contributions and Implications: Our core contribution is a fully distributed monitoring and diagnosis procedure for intelligent agents that mimics the reasoning of a human expert (Jackson, 1986) in the context of e-commerce exceptions. The knowledge base of an agent contains stateful commitments that keep track of its interactions with other agents. Using the facts contained in its knowledge base, the agent makes inference using temporal reasoning via the Reactive Event Calculus engine. Additional contributions include: (i) the extension of the commitment language presented in (Kafalı & Torroni, 2011) with deadlines for antecedents of commitments; (ii) an exhaustive list of multi- agent commitment delegation schemes; and (iii) a complete implementation of the diagnosis procedure in the Reactive Event Calculus. The proposed diagnosis procedure can be extended with an explanation capability since commitments provide semantics for agents’ interactions. Such explanations would enable human users to understand what has transcribed during a transaction and help introspection about the exception situation. Intelligent agents equipped with such expert capabilities have potential implications on several important domains regarding all phases of distributed exception handling (Kafalı et al., 2017b; Soeanu et al., 2016; Sun et al., 2012; Vasconcelos et al., 2009; Xu et al., 2011), e.g., planning, monitoring, conflict resolution, semantic reasoning, and diagnosis. While we have applied the diagnosis procedure on an e-commerce case study, its generic nature enables deployment in application areas including emergency healthcare, Internet of Things and smart environments, and security and privacy in the context of sociotechnical systems. The rest of the paper is structured as follows. Section 2 describes our extensions to commitments. Section 3 discusses commitment delegation with a temporal analysis. Section 4 introduces COMODO’s distributed monitoring 3 procedure and its formal properties. Section 5 presents a case study. Section 6 reviews the relevant literature and places our contribution in the context of expert and intelligent systems. Section 7 discusses the limitations of COMODO and presents potential future directions. The Appendix provides proofs of formal properties as well as COMODO’s implementation. 2. Commitments Commitments represent contracts or protocols (Yolum & Singh, 2002; Chopra et al., 2010). A commitment, as originally defined by Singh (Singh, 1999), is a directed obligation between two agents: the debtor commits to the creditor to bringing about a given property. Our definition of commitments extends Singh’s definition with the notion of a deadline, following a recent line of research on reasoning with commitments in time (Chesani et al., 2009; Kafalı & Torroni, 2012). Definition 1. C(X, Y , Q, a[t1, t2], P , γ[t3, t4]) represents a commitment, where the debtor X commits to the creditor Y to satisfying the consequent P when the antecedent Q holds. If Q is ⊤, then X is committed to Y unconditionally1. When Q is not ⊤, it is associated with the temporal constraint a[t1, t2]. Similarly, P is associated with the temporal constraint γ[t3, t4], where γ[t3, t4] can be one of the following: • a[t3, t4] defines an absolute deadline, where P has to be brought about at some point between t3 and t4, • r[t3, t4] defines a relative deadline, where P has to be satisfied between t3 and t4 time units as of the time t Q gets satisfied, i.e., P has to be brought about at some point between t + t3 and t + t4. In Definition 1, X, Y are agents, and Q, P are atomic (or conjunctions of) propositions. Note that the antecedent can only have an absolute deadline, and a relative deadline is only defined for the consequent when the antecedent is not ⊤. In the remainder of the paper, we sometimes call commitments whose antecedent is ⊤ base-level, and whose antecedent is not ⊤ conditional (Yolum & Singh, 2002). When we discuss commitments independently of the temporal constraints, we use the simplified notation C(X, Y , Q, P ). Our commitment language currently does not support negation or disjunction of propositions, nor nested commitments. The reason for this is purely pragmatic, most realistic e-commerce scenarios can be represented with conjunction, and adding negation or disjunction will reduce efficiency. When P is a conjunction of propositions, we assume that all the atomic propositions in P have the same deadline. A commitment is a live object and changes state through its life-cycle (Yolum & Singh, 2002). We make use of the following five commitment states: • conditional, when Q is not yet satisfied, • expired, when a[t1, t2] has expired and Q is not satisfied, 1⊤ is a constant symbol indicating a fictitious property that does not need to be satisfied because it is already true. 4 • fulfilled, when P is satisfied with respect to γ[t3, t4], • violated, when γ[t3, t4] has expired and P is not satisfied, and • active, otherwise. Example 1. Consider C(amazon, customer, paid(customer, item) ∧ clothing(item), a[∞, ∞], discount(customer, 20, clothing), r[0, 1]). This conditional commitment states that if the customer purchases a clothing item, then Ama- zon will issue a discount for the next clothing purchase in an hour from the time the payment is made. For example, if the customer pays at time 3, this commitment will become the active base-level commitment C(amazon, customer, ⊤, a[∞, ∞], discount(customer, 20, clothing), a[3, 4]). 3. Delegation When an agent X is bound to a commitment C, it may decide to carry out the consequent (if X is the debtor) or the antecedent (if X is the creditor of a conditional commitment) only by itself, or by delegating C in part, or in full, to other agents, which will act as subcontractors. Multiple commitments may then originate from C. These will be, directly or indirectly, related to C. Previous work has looked at commitments and their relations from different angles. In particular, Chopra and Singh (2009) compare commitments via a strength relation using the commitments’ properties, whereas Kafalı et al. (2012) focus on the temporal aspects of commitments and provide similarity relations based on the commitments’ deadlines. We combine both approaches, propose direct and indirect delegation relations, and show which cases are relevant to monitoring. Definition 2. A delegation of a commitment C(X, Y , Q, P ), called primary, is a new commitment where either X or Y plays the role of the creditor or debtor (delegator), and a new agent Z (delegatee) is responsible for bringing about the antecedent Q or part of Q (for conditional commitments only), or the consequent P , or part of P . The common property between primary and delegation is called delegandum. In the sequel, we use the notation debtor(C, X) to indicate that C’s debtor is X, and delegatee(C, Cj , Y ) to indicate that the delegatee of C’s delegation Cj is Y . Next, we show how different kinds of delegations are defined and combined, considering variations of C(amazon, customer, paid, delivered) as our primary (Cprim). 3.1. Basic Delegations Basic delegations are instances of Definition 2 involving two commitments. They can be of six types: explicit, implicit, antecedent, and their duals (inverse delegations). 5 Amazon Customer UPS primary C(amazon, customer, paid, delivered) explicit delegation C(ups, customer, deliveryFee, delivered) inverse explicit delegation C(customer, ups, delivered, deliveryFee) Figure 1: Explicit delegation. The inverse case is shown with a dashed arrow. Definition 3. Commitment C(Z, Y , Q′, P ′) is an explicit delegation of commitment C(X, Y , Q, P ) iff P |= P ′.2 This type of delegation was proposed by Yolum and Singh (2002) as the result of a “Delegate” operation. A new commitment is created, whereby the new debtor is committed to the same creditor, and if P = P ′, the primary is canceled following a “Cancel” operation (Yolum & Singh, 2002). An explicit delegation is shown in Figure 1. The new debtor UPS replaces the former debtor Amazon. Definition 4. Commitment C(Y , Z, Q′, P ′) is an inverse explicit delegation of commitment C(X, Y , Q, P ) iff P |= Q′. The creditor Y of the primary is now the debtor of the new commitment, and Y wishes to achieve P (or part of it) via a new creditor Z. This is an inverse delegation to achieve P since there is no obligation for Z to satisfy P, still, Z is motivated to satisfy P if he wishes Q to be satisfied. The concept of inverse delegation was introduced by Kafalı and Torroni (2011), inspired by the work of Chopra et al. (2010). An inverse explicit delegation is shown in Figure 1. Note that the roles of creditor and debtor are reversed, and accordingly the antecedent and the consequent are reversed as well. Definition 5. Commitment C(Z, X, Q′, P ′) is an implicit delegation of commitment C(X, Y , Q, P ) iff P |= P ′. The debtor of the primary is now the creditor of a new commitment for (part of) the consequent P. The primary becomes dependent on the delegation, as proposed by Kafalı et al. (2012). An implicit delegation is shown in Figure 2. Note that the creditor is Amazon, which is the primary’s debtor. The primary is not cancelled, because a commitment must be kept to the former creditor (the customer). Definition 6. Commitment C(X, Z, Q′, P ′) is an inverse implicit delegation of commitment C(X, Y , Q, P ) iff P |= Q′. 2The semantics of |= will depend on the language of the antecedent/consequent properties. In this paper for simplicity we consider properties to be conjunctions of propositions, therefore P |= P ′ ⇔ (P = P ′) ∨ (P = P ′ ∧ P ′′), where P, P ′, P ′′ are all (conjunctions of) propositions. 6 Amazon Customer UPS primary C(amazon, customer, paid, delivered) implicit delegation C(ups, amazon, deliveryFee, delivered) inverse implicit delegation C(amazon, ups, delivered, deliveryFee) Figure 2: Implicit delegation. The inverse case is shown with a dashed arrow. The debtor of the primary also becomes the debtor of a new commitment where the antecedent is (part of) the primary’s consequent. This type of delegation, as well as the next two ones (antecedent and inverse antecedent delegation), were introduced by Kafalı and Torroni (2011). An inverse implicit delegation is shown in Figure 2. Definition 7. Commitment C(Z, Y , Q′, P ′) is an antecedent delegation of commitment C(X, Y , Q, P ) iff Q is not ⊤and Q |= P ′. The creditor of the primary also becomes the creditor of a new commitment for (part of) the antecedent of the primary. An antecedent delegation is shown in Figure 3. Since the former consequent (delivered) does not appear in the antecedent delegation, in order to maintain a commitment about the former consequent, the primary is not cancelled. Antecedent delegations and implicit delegations can be combined together and bind multiple commitments into causal relations (see Section 3.2). The last type of delegation we consider is the inverse variant of antecedent delegation. Definition 8. Commitment C(Y , Z, Q′, P ′) is an inverse antecedent delegation of commitment C(X, Y , Q, P ) iff Q is not ⊤and Q |= Q′. The creditor of the primary is now the debtor of a new commitment whose antecedent is (part of) the antecedent of the primary. As in the previous case, the primary is not canceled. An inverse antecedent delegation of the primary is shown in Figure 3. Definitions 3 - 8 give an exhaustive account of how a commitment can be rationally delegated, i.e., so as to preserve the responsibilities of roles in relation with the primary’s properties (Kafalı & Torroni, 2011). We denote via dlg(Ci, C), that Ci is a delegation of C. 3.2. Causal Delegations We will now shift the focus to commitments that are linked to each other via other commitments. To this end, in (Kafalı & Torroni, 2011) we introduced the concept of commitment similarity. Here we extend similarity, in order to 7 Amazon Customer Bank primary C(amazon, customer, paid, delivered) antecedent delegation C(bank, customer, account, paid) inverse antecedent delegation C(customer, bank, paid, account) Figure 3: Antecedent delegation. The inverse case is shown with a dashed arrow. capture the notion of chains of delegations. We are interested in cases where two seemingly unrelated commitments are connected to each other via a third commitment. Let us review all possible combinations of delegations described previously: I. Explicit delegations: There are no possible chains with explicit delegations since the primary is cancelled in the process. II. Implicit delegations: Consider the case where an implicit delegation is followed by another implicit delegation as depicted below. C1 = C(X, Y, Q, P) C2i = C(Z, X, R, P) C3i = C(W, Z, O, P) Now, assume that X respects the deadline for P when delegating to Z, but Z does not when delegating to W . Eventually, this would lead to the violation C1. But, it would also violate C2i. Thus, this can be identified by only looking at the individual delegation of C2i to C3i. That is, we do not need to consider more than one implicit delegation at a time. III. Antecedent delegations: Consider the case where an antecedent delegation is followed by another antecedent delegation as depicted below. C1 = C(X, Y, Q, P) C2a = C(Z, Y, R, Q) C3a = C(W, Y, O, R) Assume that Y delegates Q to Z with a deadline relative to R, and Y further delegates R to W with an absolute deadline. Now, there is a no way of knowing C1 would be violated due to C2a (since it has a relative deadline) unless we take into account C3a. Thus, all three commitments are connected. Accordingly, we may need to consider more than one antecedent delegation at a time, in order to identify the source of an exception. 8 IV. Implicit delegation followed by antecedent delegation: C1 = C(X, Y, Q, P) C2i = C(Z, X, R, P) C3a = C(W, X, O, R) Now, assume that X delegates P to Z with a deadline relative to R, and X further delegates R to W with an absolute deadline. Similar to the antecedent delegations case, all three commitments should be considered in order to identify a problem. V. Antecedent delegation followed by implicit delegation: C1 = C(X, Y, Q, P) C2a = C(Z, Y, R, Q) C3i = C(W, Z, O, Q) Again, a problem with this case can be identified by looking at two commitments at a time as in the implicit delegations case. Definition 9. Commitment C1 = C(X1, Y1, Q1, P1) is a causal delegation of commitment C2 = C(X2, Y2, Q2, P2) via commitment C3 = C(X3, Y3, Q3, P3) if (a) P2 |= P3 and X2 = Y3 (implicit delegation), and Q3 |= P1 and Y3 = Y1 (antecedent delegation), or (b) Q2 |= P3 and Y2 = Y3 (antecedent delegation), and Q3 |= P1 and Y3 = Y1 (antecedent delegation). We call C1 cause, C2 outcome, and C3 connective. The sequence of delegations from the outcome to the cause forms a causal delegation chain. Note that the number of connectives might increase making the delegation chain longer, e.g., a series of implicit delegations followed by a series of antecedent delegations. Note that the first part (series of implicit delegations) is Case II, and can be tackled by looking at commitments pairwise. Definition 9 connects three commitments through two delegations; either (i) an implicit delegation followed by an antecedent delegation, or (ii) an antecedent delegation followed by another antecedent delegation. This allows us to trace chained commitments where the delegandum changes along the delegation chain. We account for all relations between a given commitment (primary) and its direct and indirect delegations, within the scope of a single agent. Now, let us demonstrate a typical case for Definition 9. Example 2. Consider the following set of commitments as depicted in Figure 4: C2.1 = C(amazon, customer, paid, delivered) C2.2 = C(ups, amazon, deliveryFee, delivered) C2.3 = C(bank, amazon, account, deliveryFee) 9 Amazon Customer UPS Bank outcome C(amazon, customer, paid, delivered) connective C(ups, amazon, deliveryFee, delivered) cause C(bank, amazon, account, deliveryFee) Figure 4: Causal delegation. Cause and outcome are connected via connective. That is, Amazon’s commitment for delivery depends on UPS, which in turn depends on the bank for the payment of delivery. According to C2.1 (Cprim), once the customer pays, Amazon will have the goods delivered. Now, Amazon delegates the delivery to UPS via C2.2. However, in order to deliver, UPS needs payment for delivery. Amazon makes another delegation to the bank for payment via C2.3. Thus, Amazon’s delivery (C2.1, outcome) is now connected via C2.2 (connective) to bank’s payment (C2.3, cause), which is an example of case (a) in Definition 9. For each commitment whose antecedent or consequent is a conjunction of properties, there may be more than one delegation. We thus obtain a delegation tree. We can trace all delegations of a given commitment by exhaustive search of the delegation tree. 3.3. Delegation Tree A delegation tree is a set of connected delegations. This is formally described in the next two definitions. Definition 10. A delegation chain σ = C1, . . . , Cj ⊂ C is a set of commitments such that ∀i, 1 < i ≤ j, Ci is a direct3 delegation of Ci−1. The first element is called the root of the chain. Definition 11. A delegation tree τ = (V, A) is a tree, whose nodes are commitments, V ⊆ C, such that for every edge (Ci, Cj) ∈ A, Cj is a direct delegation of Ci. 3.4. Temporal Analysis We will now enrich the relations we have defined so far, by taking into account temporal constraints. We seek to identify and understand the reasons behind exceptional situations that can lead to faulty behaviour. To this end we will define cases of delegations where the deadline of the primary is not properly propagated onto the delegation. We will use the term improper to label a delegation whose deadline exceeds the primary’s deadline, and that can possibly 3A direct delegation is one of the cases described in Definitions 3 - 8. 10 cause a mismatch between the satisfaction conditions believed by the delegator and delegatee. We will consider the overall system as it is observed at a specific time, the time of observation. First, we describe how two time intervals are compared. Definition 12. Let t be the time of observation. An interval I1=γ1[t1, t2] exceeds an interval I2=γ2[t3, t4], iff either of the following holds: • I1 is absolute, I2 is absolute, and t1 < t3 or t2 > t4, • I1 is absolute, I2 is relative, and t1 < t3 + t or t2 > t4 + t, • I1 is relative, I2 is absolute, and t + t1 < t3 or t + t2 > t4, • I1 is relative, I2 is relative, and t1 < t3 or t2 > t4. Next, we use the notion of exceeding intervals to define improper delegations4. Definition 13. Let Cid = C(Z, X, Q ′, , P ′, I′) be an implicit delegation of Cprim = C(X, Y , Q, , P , I). Cid is an improper consequent delegation of Cprim iff I ′ exceeds I. The inverse case is defined similarly. Example 3. Consider the following set of commitments: C3.1 = C(amazon, customer, paid, a[1, 12], discount ∧ delivered, a[31, 45]) C3.2 = C(office, amazon, ⊤, a[∞, ∞], discount, [31, 31]) C3.3 = C(ups, amazon, ⊤, a[∞, ∞], delivered, a[35, 50]) Now, discount has been delegated correctly, since C3.2’s deadline does not exceed that of C3.1. C3.3 instead is an improper delegation, whose deadline exceeds that of C3.1. Note that the occurrence of an exception is not inevitable, since UPS may still complete delivery before time 45. However, C3.3 creates a vulnerability, and may be the root of future exceptions. Definition 14. Let Cad = C(Z, X, Q ′, , P ′, I′) be an antecedent delegation of Cprim = C(X, Y , Q, I, P , ). Cad is an improper antecedent delegation of Cprim iff I ′ exceeds I. The inverse case is defined similarly. Definition 15. Let σ be a delegation chain rooted in C, and let Cj ∈ σ be a (direct or causal) delegation of Ci ∈ σ. Let I=γ[ts, te] be C’s interval and Ii=γi[ti,s, ti,e] be Ci’s interval, for each Ci in σ. Let t be the time of observation. Cj is an improper causal delegation of C iff either of the following holds: • I and all Ii are relative, and Σ j i=0ti,s < ts or Σ j i=0ti,e > te, • assuming Ik is the last absolute deadline in σ, and tk is the time Ck’s property is satisfied 4We omit the deadline interval in the commitment with ‘ ’ when it is not relevant to the discussion. 11 Symbol Description δpro set of proper delegations of a given commitment δimp set of improper delegations of a given commitment C = {〈C, δpro, δimp〉, . . . } for each known commitment C, a triplet consisting of: C, its proper delegations δpro, and its improper delegations, δimp X ▽REC C C contains all (locally) known information about X’s commitments, extracted by a REC reasoner such as ComMon δexc commitments to be excluded from a monitoring process δout, δj , δk output of monitoring processes (sets of improper delegations) X �δexc δout C the result of a monitoring process issued by X about C and excluding δexc, is the set of improper delegations δout X �δexc δout Y ≫ C the result of a monitoring process requested by X to agent Y about C and excluding δexc, is the set of improper delegations δout Table 1: Notation for COMODO’s monitoring process. – I is absolute and tk + Σ j i=k+1 ti,s < ts or tk + Σ j i=k+1 ti,e > te, or – I is relative and tk + Σ j i=k+1 ti,s < t + ts or tk + Σ j i=k+1 ti,e > t + te. Definition 16. Let C, Cj ∈ C. Cj is an improper delegation of C, denoted dlgimp(Cj , C), iff • Cj is an improper consequent delegation of C, or • Cj is an improper antecedent delegation of C, or • Cj is an improper causal delegation of C, or • ∃ Ck ∈ C such that dlg(Ck, C) and dlgimp(Cj , Ck). A delegation which is not improper is called a proper delegation and denoted by dlgprop(Ci, C) (meaning that Ci is a proper delegation of C). 4. Monitoring In this section, we describe COMODO’s distributed monitoring procedure. At a given time of observation t, a monitoring process M records all the improper delegations Mt that occurred up to t. Definition 17. A monitoring process M is a process whose inputs are 12 • a set A of agents, • a set C of commitments among agents in A, • a narrative Tt of events up to a given time of observation t, and • a commitment model and domain knowledge defining the states of commitments in C based on Tt, and whose output is Mt = {(Ci, Cj)|dlgimp(Ci, Cj)}, where Ci, Cj ∈ C. The purpose of monitoring is to identify all the exceptions (e.g., improper delegations) among agents’ commit- ments at a given time of observation, considering the available knowledge. Note that M is an abstract concept, as we cannot assume that there is always an agent who has complete global knowledge. We use this to demonstrate that our distributed monitoring procedure produces the same output as the global monitoring process. Typically, a monitoring process starts from a specific commitment Cm whose delegations need to be analysed. For example, Cm’s creditor X might want to check the situation with Cm some time before P ’s (the property of Cm) deadline expires, in order to prevent potential problems. So, X will ask Y ’s collaboration (as the debtor of Cm). Accordingly, Y will run a local monitoring process about P , and report back to X. The initial commitment about P may in turn be linked to a number of other commitments, thus originating a chain of commitments, possibly involving additional agents, other than X and Y . We refer to a narrative Tt of events to trace a protocol execution. In particular, the successful completion of a given action by a given agent will be represented by a particular event in Tt. We do not model action duration, but only completion. Tt contains all the elements that describe a specific protocol execution up to time point t. 4.1. Distributed Monitoring We will now describe the distributed monitoring procedure that agents follow to detect improper delegations. The monitoring procedure is a derivation process, described by the local rules L1 and L2 (intra-agent reasoning) and the social rules S1 - S3 (inter-agent reasoning). Table 1 summarizes the notation. Given an agent X ∈ A and a commitment Cm ∈ C, a derivation X � ∅ δout Cm starts when X decides to monitor one of its commitments Cm. The ∅ symbol (which is an input to the derivation) signifies that no commitment is initially excluded from the monitoring process, because no commitment has been analysed yet. The output δout is a set of improper delegations, which might be empty in some cases. The monitoring procedure may propagate from agent to agent, as described by the social rules. As commitments get analysed by the agents involved in the monitoring process, they are included in the set δexc when performing further derivation. In this way, we prevent agents from analysing the same commitment more than once. In a concrete implementation, answering to a monitoring request could be implemented as a background agent behaviour, whereas issuing a monitoring request could be implemented by a communicative act from an agent X to an agent Y , that implements the “answering to a monitoring request” behaviour. A possible architecture for distributed monitoring is described in (Kafalı & Torroni, 2012), where observations are 13 local to the agent, and commitment-based contract specifications are instead shared, i.e., accessible to both the debtor and the creditor of each commitment. Each agent involved will only use its local knowledge of commitments to contribute to the derivation by applying local and social monitoring rules. Part of the local reasoning amounts to checking which commitments are linked to the subject commitment, via proper or improper delegations. This is defined in the REC5 language, assuming that for commitment tracking purposes each agent relies upon tools such as ComMon. However, in the general case, the delegation check could be done by using any procedure that queries a local database of commitments. Local monitoring: These are the rules used for monitoring the agent’s commitments locally. They describe intra- agent reasoning, which is based on the agent’s local knowledge base (i.e., own commitments and fluents). This is performed via the agent’s internal REC engine, for which the details will be given in the implementation part (see Section 4.2). L1) X ▽REC C ∧ 〈Cm, δpro, δimp〉 ∈ C ∧ δout = δimp \ δexc ∧ δout 6= ∅ X �δexc δout Cm By rule L1, if the agent identifies any improper delegations of the currently monitored commitment Cm via querying its REC engine locally, and these commitments are not already contained in δexc (the set containing the commitments that are already processed during monitoring), then they are added to the output (δout) of the monitoring process. Consider the commitments in Example 3: let X be Amazon, Cm be C3.1, and δexc = ∅. Now, when Amazon queries their REC engine, he will find out that δpro = {C3.2} and δimp = {C3.3}. Thus, δout = {C3.3} which contains the only improper delegation of C3.1. L2) X ▽REC C ∧ 〈Cm, δpro, δimp〉 ∈ C ∧ (δpro ∪ δimp) \ δexc = ∅ ∧ debtor(Cm, X) X �δexc ∅ Cm By rule L2, if there are no locally known delegations of the monitored commitment Cm, and X is Cm’s debtor, the result is an empty set. This rule complements L1, and is a termination condition for some branches of the distributed monitoring process, when there are no more delegations left in the corresponding delegation chain for the subject commitment. Social monitoring: These rules describe how the derivation process propagates from one agent to another agent, and how the results are combined. They describe the inter-agent reasoning, which is based on the monitoring interactions (e.g., requests and responses) among the agents. In the following rules, we use the notation ≫, whose semantics is given in Table 1, to indicate a request for monitoring. 5REC (Reactive Event Calculus) is an event calculus-based language and reasoning framework (Chesani et al., 2009). ComMon is a REC- based monitoring engine that can be downloaded from http://ai.unibo.it/projects/comMon. 14 S1) X ▽REC C ∧ 〈Cm, δpro, δimp〉 ∈ C ∧ Cj ∈ δpro \ δexc ∧ delegatee(Cm, Cj, Y ) ∧ X �δexc δj Y ≫ Cj ∧ X � δexc∪{Cj} δk Cm ∧ δout = δj ∪ δk X �δexc δout Cm By rule S1, if there is a locally known proper delegation Cj which is not to be excluded (Cj ∈ δpro \ δexc), X delegates monitoring to Cj ’s delegatee Y , thereby obtaining a result δj . X will then continue monitoring its other delegations, excluding Cj from the process, thereby obtaining a result δk. The final result δout is the union of the two partial results, δj ∪ δk. S2) X ▽REC C ∧ 〈Cm, δpro, δimp〉 ∈ C ∧ (δpro ∪ δimp) \ δexc = ∅ ∧ debtor(Cm, Y ) ∧ X 6= Y ∧ X �δexc δout Y ≫ Cm X �δexc δout Cm By rule S2, if there is no locally known delegation of the monitored commitment Cm, Cm’s creditor X makes a monitoring request to Cm’s debtor Y , and the result δout is provided by Y as the response. S3) Y �δexc δout Cm X �δexc δout Y ≫ Cm By rule S3, an agent Y answers to X’s request for monitoring concerning a given commitment Cm by executing a monitoring process about Cm and propagating the result back to X. This procedure relies on local reasoning and collaboration among agents to produce monitoring results that, ide- ally, should be equivalent to the global results produced by the abstract monitoring process M. Under the assumption that the REC reasoner provides sound and complete results, we can prove the following theorems (see Appendix A): Theorem 1 (Soundness). Given a commitment Cm ∈ CT , and an agent X ∈ A, if X � ∅ δout Cm and Ci ∈ δout, then (Ci, Cm) ∈ MT . By Theorem 1, if the distributed monitoring process identifies an exception in the form of an improper delegation Ci of a given commitment Cm, then (Ci, Cm) is an outcome of the global monitoring (see Definition 17). Theorem 2 (Completeness). ∀ (Ci, Cm) ∈ MT , ∃ an agent X ∈ A and a derivation X � ∅ δout Cm such that Ci ∈ δout. By Theorem 2, for any two given commitments Ci, Cm ∈ CT , if Ci is an improper delegation of Cm, then there is a possible run of the distributed monitoring process starting from some agent X that identifies it as such. 4.2. Implementation We have provided an implementation for COMODO. We have written specifications in the REC language, and utilised ComMon for monitoring of commitments. More specifically, the input to ComMon is the following: 15 • a commitment theory that contains the rules for manipulation of commitments, • a domain model that contains the protocol rules that describe the agents’ domain, • an event trace that contains the actions of the agents throughout time. Given these inputs, ComMon produces an outcome that displays the agents’ fluents through time. This is used to monitor the individual states of the commitments at run-time. Moreover, we have defined a subset of the commitment relations introduced in this paper in the REC language, thus extending the the commitment model with a delegation model and an exception model, in order to accommodate local reasoning. Additional details on the implementation can be found in Appendix B. 5. Case Study Let us now use Amazon’s Prime next-day delivery scheme6 to demonstrate how COMODO works. We have the following three commitments to represent the process for the customer to order an item from Amazon: • C1 = C(amazon, customer, paid ∧ prime, a[1, 12], delivered, r[12, 36]): Amazon must deliver the client’s order within the following day, • C2 = C(ups, amazon, packaged, a[1, ∞], delivered, r[6, 24]): When the item is packaged, UPS can deliver it in the next 24 hours, • C3 = C(office, amazon, confirmed, a[1, ∞], packaged, r[6, 24]): Confirmed orders are packaged in the next 24 hours. Note that the customer only knows about the first commitment C1. In addition, the following two actions are known to the customer: • pay(customer, amazon) → paid. • deliver(ups, customer) → delivered. The semantics of the actions given by the above rules is that when the action on the left-hand side is executed, then the fluent on the right-hand side holds. For the sake of simplicity, we assume that action executions are successful and their effects are independent of the context. Now, consider the following trace of events: 12 pay(customer, amazon) 17 confirm(amazon, office) 30 package(office, amazon) That is, customer pays for the item at noon. Amazon confirms the order at 5 pm, and the item is packaged next morning at 6 am. The following commitments are in place at time 30: 6www.amazon.com/prime 16 C1 = C(amazon, customer, ⊤, a[∞, ∞], delivered, a[24, 48]) C2 = C(ups, amazon, ⊤, a[∞, ∞], delivered, a[36, 54]) C3 = C(office, amazon, ⊤, a[∞, ∞], packaged, a[23, 41]) Notice the pattern among these three commitments; C2 is an implicit delegation of C1 (Definition 5), and C3 is an antecedent delegation of C2 (Definition 7). Then, C3 is a causal delegation of C1 via C2 (Definition 9). First, we look at the global monitoring result considering all the commitments in the system. Assume that no delivery has occurred until time 48. C1 is indeed violated since its deadline has passed. Because of the causal delegation, C2 and C3’s deadlines together affect that of C1. Even though the packaging of the item is completed at time 30, UPS has 24 hours for delivery, which will eventually exceed C1’s deadline. If the delivery is completed at time 54, C2 is fulfilled. However, C1 is still violated. Here, Amazon should have confirmed customer’s order earlier, or set a tighter deadline for C2. Next, we look at the agents’ local reasoning: • Customer: C = {〈C1, {}, {}〉} – Rules L1, L2, and S1 do not apply, – Rule S2 delegates to Amazon. • Amazon: C = {〈C1, {}, {C2, C3}〉, ...} – Rule L1 applies, and finds an improper delegation, – Rule S3 propagates the result to the customer. Now, let us change the trace of events so that the protocol will not lead to any improper delegations. Consider the following trace: 12 pay(customer, amazon) 13 confirm(amazon, office) 16 package(office, amazon) Notice that the confirmation of the order is performed earlier in this case, which leads to an earlier deadline for packaging (C3). This also affects the delivery of the item which depends on packaging C2. Figure 5 shows the output of the ComMon tool for Amazon and the customer’s reasoning. The horizontal axis shows the timeline of events that have occurred during execution. Alongside such events, we inserted additional tick events, whose only purpose is to force ComMon to update and display the state of commitments and fluents at every time point. The commitments and fluents are shown alongside the vertical axis together with how their states change over time. Note that we omit antecedent deadlines for brevity. 17 Customer (exception) Customer (no exception) Amazon (exception) Amazon (no exception) Figure 5: ComMon output for Amazon’s Prime next-day delivery. 6. Related Work We have presented a commitment-based approach to support automated reasoning in expert systems, where intel- ligent agents employ distributed monitoring and diagnosis to resolve exceptions caused by commitment delegations. Commitments are a specific type of norm to regulate interactions among various stakeholders in sociotechnical sys- tems (Barth et al., 2006; Singh, 2013; Hao et al., 2016; Kafalı et al., 2016, 2017a,b; Vasconcelos et al., 2009). In the rest of this section, we first summarize the strengths and limitations of relevant approaches from the literature (including COMODO) in Table 2, and then we review each approach in more depth. A good deal of related work on commitments investigates formal properties of commitments (Lorini, 2010; Singh, 2008; Khan & Lespérance, 2006; Verdicchio & Colombetti, 2003), temporal extensions of commitment languages 18 Approach Pros Cons Chesani et al. (2013); Kafalı & Tor- roni (2012) Temporal commitments Centralised approach Sun et al. (2012); Xu et al. (2011) Domain ontologies; Automated ne- gotiation Focused on e-commerce domain Ajmeri et al. (2016); Günay & Yolum (2013); Vasconcelos et al. (2009) Conflict resolution; Generalizable to all norm types Centralised approach Al-Saqqar et al. (2016); El- Menshawy et al. (2013); Herd et al. (2015) Formal verification of protocol properties Design-time approach Castelfranchi & Falcone (1998); Norman & Reed (2010); Falcone & Castelfranchi (2001); Santos et al. (1997) Attribution of responsibility and accountability; Foundational con- cepts; Theoretical analysis No practical application explored Gao & Singh (2014) Automated extraction of norms from business contracts Both prac- tical and dialectical commitments No exception handling or diagno- sis; Design-time approach COMODO Distributed monitoring and diagno- sis; Generic and generalizable pro- cess; Run-time approach; Flexible agent execution Single application domain; Single norm type Table 2: Comparison of COMODO with relevant literature. (Fornara & Colombetti, 2004; Mallya et al., 2004; Venkatraman & Singh, 1999), normative models (Antoniou et al., 2009; Kafalı et al., 2017a; Vasconcelos et al., 2012), and norm monitoring and planning (Alechina et al., 2016; Fornara & Colombetti, 2010; Gasparini et al., 2016; Meneguzzi et al., 2015; Spoletini & Verdicchio, 2009). The temporal extensions to commitment languages and monitoring procedures proposed in (Chesani et al., 2009, 2013) are used in this paper for local reasoning. Temporal constraints are also used in (Kafalı & Torroni, 2012) to compare commitments. This approach, as well as others, mainly focuses on the relations between pairs of commitments in two-agent interactions. In (Chopra & Singh, 2015), delegation is also taken into account from the perspective of 19 commitment alignment, but no temporal aspects are considered. Agent-based approaches have been utilized for monitoring, detection, and handling of exceptions. Kafalı and Yolum (2016) propose an approach for monitoring an agent’s interactions to determine whether the agent is pro- gressing as expected. In particular, they verify whether the agent’s expectations (represented by a set of propositions and commitments) are satisfiable by its current state. Sun et al. (2012) propose a distributed environment for e- procurement processes, where each agent is assigned to different a task such as search, negotiation, monitoring, or exception handling. Xu et al. (2011) propose a taxonomy of logistics exceptions based on previous work in the lit- erature. They differentiate among potential problems related to deliveries such as late or partial delivery, and explore dependencies between all logistics components. It would be interesting to apply COMODO on these domains as well as extend our delivery scheme with a domain ontology. Günay and Yolum (2013) discuss the feasibility of a set of commitments, i.e., whether it is possible for an agent to honor all its (existing and prospective) commitments. They formulate feasibility as a constraint satisfaction prob- lem. Vasconcelos et al. (2009) propose methods for resolving conflicts among norms. Their resolution method, norm curtailment, manipulates the constraints associated with norms, e.g., reduce the scope of a prohibition to avoid conflict with an obligation. Ajmeri et al. (2016) propose Coco, a formalism to express and reason about conflicting commitment instances at runtime, and dominance among them. Coco employs Answer Set Programming to compute nondominated commitment instances and uses Alechina et al.’s (2013) framework to determine compliance of actions with nondominated commitment instances. Compared to Coco, while we do not explicitly deal with conflicts among commitment delegations, a detected violation during monitoring indicates a potential conflict. In the general context of commitment frameworks, not addressing monitoring, many authors considered the use of commitments to represent, model and verify protocols. Among them, El-Menshawy et al. (2013) propose several methods for commitment-based protocol verification using model checking. Similar verification based approaches (Al-Saqqar et al., 2016; Herd et al., 2015) are performed at design time. We focus instead on run-time verification. We do not elaborate here on how to use our analysis at design time, although that may be a possible application. The concept of commitment delegation has been proposed by Singh and Yolum (2002). The delegation mechanism gives great flexibility to commitment-based protocols. However, it also lays itself open to misuse and may induce possible mismatches among agent beliefs about deadlines associated with properties. Improper delegations eventually drive the system into a state of violation, where some agents believe that there has been no violation at all. In this work, we presented an in-depth analysis of improper delegations, and proposed an effective distributed reasoning procedure for finding all improper delegations of a given commitment. Santos et al. (1997) investigate aspects of organised interaction and propose a characterisation of “transmission of agency” (a concept related to delegation and attribution of responsibility) and the analysis of the conditions under which a given organisation recognises that an agent has fulfilled his responsibilities. Castelfranchi and Falcone (1998; 2001) develop a theory of delegation and adoption, using a plan-based approach. The authors’ perspective is more general than that of Santos et al. It does not necessarily target an institutionalised environment, but it considers task 20 and delegation to be foundational concepts of agency and autonomy in the broader perspective. In any case, when an agent delegates a task to another agent, the latter takes care of the interests of goals of the former remotely, i.e., far from it, and without its monitoring and intervention (control). According to Castelfranchi and Falcone, in delegation, an agent A tries to achieve some of its goals through another agent B’s actions, thus A has the goal that B performs a given action. Delegation and adoption are thus characterised in terms of mental states of the agents involved in the interaction. The authors distinguish between “weak” adoption, i.e., based on spontaneous initiative, and “strict” adoption, i.e., accompanied by a formal agreement (contract). Both Santos et al. and Castelfranchi and Falcone stress the fundamental importance of expressing agent behaviour without referring to concrete actions when delegating, as a basis for flexibility (“open” delegation in (Castelfranchi & Falcone, 1998)). The relationship between openness and control is further explored in (Falcone & Castelfranchi, 2001), where Falcone and Castelfranchi propose a theory of adjustable autonomy, where trust is the cognitive basis for adjusting autonomy. Gelati et al. (2004) provide a formal analysis of the idea of normative coordination, in the belief that the adoption of a normative perspective would allow a substantial progress in the creation of agent societies. Agents can achieve flexible co-ordination by conferring normative positions to other agents. The building blocks of their analysis are declarative power (the capacity of the power-holder of creating normative positions by proclaiming such positions), representation (the representative’s capacity of acting in the name of its principal) and mandate (the mandate’s duty to act as the mandator has requested). Norms are also investigated in agent-based supply-chain environments (Vasconce- los et al., 2012), and conflicts are discussed in the form of exceptions. Moreover, agents are used to resolve exceptions in e-procurement systems, where several agents enact different roles (Sun et al., 2012). Another use for agents is discussed in (Chen & Nof, 2012), where agents detect and prevent errors in sequential production lines. Governa- tori (2013) proposes a conceptual abstract framework to model normative requirements, formalizes different types of obligations, and verifies whether a business process is compliant with requirements (set of obligations). Integrating delegations into the above approaches would be an interesting directions to pursue. Several authors, including Lorini et al. (2007; 2009) and Norman and Reed (2010), proposed a semantic charac- terisation of delegation in relation with agent mental states such as beliefs and intentions (Lorini et al., 2007), with the semantics of speech acts (Longin et al., 2009), and with the concept of responsibility (Norman & Reed, 2010). In contrast to these approaches, we do not follow a normative perspective, and we do not make any formal refer- ence to the foundational concepts above. We refer instead to the intuitive notion of responsibility and accountability, to justify the links that may bind two commitments together. In particular, we will say that when a commitment be- tween two agents X and Y is delegated, a third agent (the delegatee), will be responsible for bringing about a property derived from the initial commitment (the delegandum). To the best of our knowledge, there are no other works in the literature that propose ways to reason about chains of commitment delegations and identify (potentially) problematic situations. Our work also has a practical interest. We show how the monitoring process can be implemented using efficient, off-the-shelf tools. We are not aware of other works that cover both theoretical and practical aspects of the problem we address. 21 Verifying agent executions against commitment specifications (or interaction protocols in general) has been the focus of recent research, both at design-time as well as run-time. Gao and Singh (2014) propose a method for extracting contracts from actual business relationships. Contracts are represented as both practical and dialectical commitments. In this paper, we only focus on practical commitments since our focus is on e-commerce protocols. Kafalı et al. (2014) propose a distributed algorithm to verify at run-time whether the goals of the agent will be satisfied via its commitments. They make use of temporal constraints and implement their work with REC like we do here. However, their commitment relations are basic compared to our extensive study of commitment delegations. Abushark et al. (2014) also focus on detecting exceptions in the form of defects. However, this is not designed for run-rime detection as we do here. They compare agent designs with protocol specifications, and aim to help agent developers to reduce defects in design. 7. Discussion In this paper, we have built upon previous work (Kafalı & Torroni, 2011), where we discuss a systematic clas- sification of commitment delegation types using a simple commitment language. There, we have used motivating examples inspired from an e-commerce scenario, to show that delegation can follow meaningful patterns, other than the traditional way of delegating commitments proposed in the literature (Yolum & Singh, 2002). Moreover, we have introduced the concept of similarity, improper delegation, and monitoring process. In COMODO, we have extended the language for commitments by introducing relative and absolute deadlines represented as time intervals. We have further explored the concepts of similarity and improper delegation by giving an exhaustive account of all possible improper delegations, and we have provided a sound and complete distributed reasoning procedure that is able to find all improper delegations of a given commitment. Limitations • In this work, we have focused on a specific type of norm, commitment, which is a dominant artifact in e- commerce contracts. However, in other domains such as healthcare, authorization and prohibition norms are commonly used to represent regulations. • We have not evaluated the expressiveness of our commitment language in multiple contract domains. While the deadline conditions we introduced are helpful in representing e-commerce contracts, extending the language with disjunction and maintenance properties would further increase expressiveness. Maintenance properties can be related with a maintenance goal (Chesani et al., 2009), where a certain property should hold at all times during a specified interval. Moreover, temporal constraints such as those proposed by several languages for temporal representation and reasoning (e.g., before, after, until) may indeed be useful in some applications (Marengo et al., 2011). However, this would significantly increase the complexity of our temporal reasoning agents. 22 • We have demonstrated the working of COMODO on one case study from e-commerce, which constitutes a threat to external validity. Other works on normative models explore emergency healthcare (Kafalı et al., 2016, 2017a), and security and privacy (Barth et al., 2006; Kafalı et al., 2017b). Exploring norm delegations in such settings would provide valuable insight to our distributed diagnosis procedure. Implications • All phases of distributed exception handling (Soeanu et al., 2016; Sun et al., 2012; Vasconcelos et al., 2009; Xu et al., 2011), e.g., planning, monitoring, conflict resolution, and diagnosis, are crucial capabilities for any expert system that deals with private and confidential information. Intelligent agents should detect and resolve incon- sistencies that arise from their interactions using partial information and without violating their users’ privacy. In this work, we focused on an e-commerce application by representing electronic contracts with commitments, which provide flexible execution for software agents. Apart from the e-commerce domain, our distributed mon- itoring and diagnosis process can be adopted in safety-critical domains such as intrusion detection (Geib & Goldman, 2001) and other crime detection (Jarvis et al., 2005) by integrating it with additional AI-based meth- ods. Commitments and additional normative representations can be combined with ontologies and semantic reasoning to provide additional expert capabilities to agents (Kafalı et al., 2017b; Xu et al., 2011). • Explainable AI is a great concern in existing machine learning applications. The diagnosis process we have proposed can be extended with additional explanation capabilities since commitments add semantics to agents’ interactions, and help human users understand exception situations. Intelligent agents equipped with such expert capabilities have potential implications on several other important domains including emergency healthcare, Internet of Things and smart environments, and security and privacy in the context of sociotechnical systems. Future Work • In principle, some of the notions that we introduced for the purpose of run-time monitoring could also be used for auditing or at design-time. For example, it may be useful to introduce design constraints, or guarantee mech- anisms to prevent agents from causing improper delegations. We do not deal with design issues here, but as a future work it would be interesting to study the application of the improper delegation notion in contexts other than monitoring, or monitoring in the planning domain (Soeanu et al., 2016). Moreover, extending commit- ments with sanctions (Nardin et al., 2016) would add another dimension to COMODO’s delegation monitoring procedure. Sanctions provide compensation for commitment violations, therefore act as deterrence against violating commitments. • It would be interesting to elaborate on how agents use the outcome of monitoring, e.g., in order to detect inconsistencies, contradictory facts or inappropriate situations, as well as to investigate the integration of the agent’s monitoring capability with other reasoning capabilities in concrete agent architectures. 23 • Our diagnosis process can be extended with other AI-based approaches such as plan recognition (Kautz, 1987), goal recognition (Lesh & Etzioni, 1995), and intention recognition (Sadri, 2012) to extend the application domain beyond e-commerce. Such recognition approaches often have implementations in the Event Calculus (EC), therefore we can seamlessly integrate those into the COMODO framework. References Abushark, Y., Thangarajah, J., Miller, T., & Harland, J. (2014). Checking consistency of agent designs against interaction protocols for early-phase defect location. In Proceedings of the 13th International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 933–940). Ajmeri, N., Jiang, J., Chirkova, R. Y., Doyle, J., & Singh, M. P. (2016). Coco: Runtime reasoning about conflicting commitments. In Proceedings of the 25th International Joint Conference on Artificial Intelligence (IJCAI) (pp. 17–23). New York: IJCAI. Al-Saqqar, F., Bentahar, J., & Sultan, K. (2016). On the soundness, completeness and applicability of the logic of knowledge and communicative commitments in multiagent systems. Expert Systems with Applications, 43, 223–236. Alechina, N., Dastani, M., & Logan, B. (2013). Reasoning about normative update. In Proceedings of the 23rd International Joint Conference on Artificial Intelligence (IJCAI) (pp. 20–26). AAAI Press. Alechina, N., Halpern, J. Y., Kash, I. A., & Logan, B. (2016). Decentralised norm monitoring in open multiagent systems: (extended abstract). In Proceedings of the 2016 International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 1399–1400). Antoniou, G., Dimaresis, N., & Governatori, G. (2009). A modal and deontic defeasible reasoning system for modelling policies and multiagent systems. Expert Systems with Applications, 36, 4125–4134. Barth, A., Datta, A., Mitchell, J. C., & Nissenbaum, H. (2006). Privacy and contextual integrity: Framework and applications. In Proceedings of the IEEE Symposium on Security and Privacy (SP) (pp. 184–198). Washington, DC: IEEE Computer Society. Castelfranchi, C., & Falcone, R. (1998). Towards a theory of delegation for agent-based systems. Robotics and Autonomous Systems, 24, 141–157. Chen, X. W., & Nof, S. Y. (2012). Agent-based error prevention algorithms. Expert Systems with Applications, 39, 280–287. Chesani, F., Mello, P., Montali, M., & Torroni, P. (2009). Commitment tracking via the reactive event calculus. In Proceedings of the 21st International Joint Conference on Artificial Intelligence (IJCAI) (pp. 91–96). Chesani, F., Mello, P., Montali, M., & Torroni, P. (2013). Representing and monitoring social commitments using the event calculus. Autonomous Agents and Multiagent Systems, 27, 85–130. Chopra, A. K., Dalpiaz, F., Giorgini, P., & Mylopoulos, J. (2010). Reasoning about agents and protocols via goals and commitments. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 457–464). Chopra, A. K., & Singh, M. P. (2009). Multiagent commitment alignment. In Proceedings of the 8th International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 937–944). Chopra, A. K., & Singh, M. P. (2015). Generalized commitment alignment. In Proceedings of the 2015 International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 453–461). El-Menshawy, M., Bentahar, J., El Kholy, W., & Dssouli, R. (2013). Verifying conformance of multiagent commitment-based protocols. Expert Systems with Applications, 40, 122–138. Falcone, R., & Castelfranchi, C. (2001). The human in the loop of a delegated agent: the theory of adjustable social autonomy. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 31, 406–418. Fornara, N., & Colombetti, M. (2004). A commitment-based approach to agent communication. Applied Artificial Intelligence, 18, 853–866. Fornara, N., & Colombetti, M. (2010). Representation and monitoring of commitments and norms using owl. AI Commun., 23, 341–356. Gao, X., & Singh, M. P. (2014). Extracting normative relationships from business contracts. In Proceedings of the 13th International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 101–108). Gasparini, L., Norman, T. J., Kollingbaum, M. J., & Chen, L. (2016). Decision theoretic norm-governed planning: (extended abstract). In Proceedings of the 2016 International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 1265–1266). 24 Geib, C. W., & Goldman, R. P. (2001). Plan recognition in intrusion detection systems. In DARPA Information Survivability Conference & Exposition II, 2001. DISCEX’01. Proceedings (pp. 46–55). IEEE volume 1. Gelati, J., Rotolo, A., Sartor, G., & Governatori, G. (2004). Normative autonomy and normative co-ordination: Declarative power, representation, and mandate. Artificial Intelligence & Law, 12, 53–81. Governatori, G. (2013). Business process compliance: An abstract normative framework. Information Technology, 55, 231–238. Günay, A., & Yolum, P. (2013). Constraint satisfaction as a tool for modeling and checking feasibility of multiagent commitments. Applied Intelligence, 39, 489–509. Hao, J., Kang, E., Sun, J., & Jackson, D. (2016). Designing minimal effective normative systems with the help of lightweight formal methods. In Proceedings of the 24th ACM SIGSOFT International Symposium on the Foundations of Software Engineering (FSE) (pp. 50–60). Herd, B., Miles, S., McBurney, P., & Luck, M. (2015). Monitoring hierarchical agent-based simulation traces. In Proceedings of the 2015 International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 463–471). Jackson, P. (1986). Introduction to expert systems. Addison-Wesley Pub. Co., Reading, MA. Jarvis, P. A., Lunt, T. F., & Myers, K. L. (2005). Identifying terrorist activity with ai plan recognition technology. AI Magazine, 26, 73. Kafalı, Ö., Ajmeri, N., & Singh, M. P. (2016). Revani: Revising and verifying normative specifications for privacy. IEEE Intelligent Systems, 31, 8–15. Kafalı, Ö., Ajmeri, N., & Singh, M. P. (2017a). Kont: Computing tradeoffs in normative multiagent systems. In Proceedings of the 31st Conference on Artificial Intelligence (AAAI) (pp. 3006–3012). Kafalı, Ö., Günay, A., & Yolum, P. (2014). Gosu: computing goal support with commitments in multiagent systems. In Proceedings of the 21st European Conference on Artificial Intelligence (ECAI) (pp. 477–482). Kafalı, Ö., Jones, J., Petruso, M., Williams, L., & Singh, M. P. (2017b). How good is a security policy against real breaches? a HIPAA case study. In Proceedings of the 39th International Conference on Software Engineering (ICSE) (pp. 530–540). Buenos Aires: IEEE Computer Society. Kafalı, Ö., & Torroni, P. (2011). Social commitment delegation and monitoring. In Computational Logic in Multiagent Systems (CLIMA XII) (pp. 171–189). volume 6814 of LNCS. Kafalı, Ö., & Torroni, P. (2012). Exception diagnosis in multiagent contract executions. Annals of Mathematics and Artificial Intelligence, 64, 73–107. Kafalı, Ö., & Yolum, P. (2016). Pisagor: A proactive software agent for monitoring interactions. Knowledge and Information Systems, 47, 215–239. Kautz, H. A. (1987). A formal theory of plan recognition. PhD thesis, University of Rochester. Khan, S. M., & Lespérance, Y. (2006). On the semantics of conditional commitment. In Proceedings of the 5th International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 1337–1344). Lesh, N., & Etzioni, O. (1995). A sound and fast goal recognizer. In Proceedings of the 14th International Joint Conference on Artificial Intelligence (IJCAI) (pp. 1704–1710). Longin, D., Lorini, E., & Nguyen, M. H. (2009). Delegation as a communicative act: a logical analysis. In Proceedings of the Workshop on Coordination, Organizations, Institutions, and Norms in Agent Systems (COIN). Lorini, E. (2010). A logical analysis of commitment dynamics. In Proceedings of the 10th International Conference on Deontic Logic in Computer Science (DEON) (pp. 288–305). Springer volume 6181 of LNCS. Lorini, E., Troquard, N., Herzig, A., & Castelfranchi, C. (2007). Delegation and mental states. In Proceedings of the 6th international joint conference on Autonomous agents and Multiagent systems (AAMAS) (pp. 153:1–153:3). Mallya, A. U., Yolum, P., & Singh, M. P. (2004). Resolving commitments among autonomous agents. In In International Workshop on Agent Communication Languages and Conversation Policies (pp. 166–182). Springer volume 2922 of Lecture Notes in Computer Science. Marengo, E., Baldoni, M., Baroglio, C., Chopra, A. K., Patti, V., & Singh, M. P. (2011). Commitments with regulations: reasoning about safety and control in regula. In Proceedings of the 10th International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 467–474). Meneguzzi, F., Telang, P. R., & Yorke-Smith, N. (2015). Towards planning uncertain commitment protocols. In Proceedings of the 2015 Interna- tional Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 1681–1682). 25 Nardin, L. G., Balke-Visser, T., Ajmeri, N., Kalia, A. K., Sichman, J. S., & Singh, M. P. (2016). Classifying sanctions and designing a conceptual sanctioning process for socio-technical systems. The Knowledge Engineering Review, 31, 142–166. Norman, T. J., & Reed, C. (2010). A logic of delegation. Artif. Intell., 174, 51–71. Sadri, F. (2012). Intention recognition in agents for ambient intelligence: Logic-based approaches. In Agents and Ambient Intelligence (pp. 197–236). IOS Press volume 12 of Ambient Intelligence and Smart Environments. Santos, F. A. A., Jones, A. J. I., & Carmo, J. (1997). Action concepts for describing organised interaction. In Proceedings of the 47th Hawaii International Conference on System Sciences (HICSS) (pp. 373–382). Singh, M. P. (1999). An ontology for commitments in multiagent systems: Toward a unification of normative concepts. Artificial Intelligence and Law, 7, 97–113. Singh, M. P. (2008). Semantical considerations on dialectical and practical commitments. In Proceedings of the 23rd National Conference on Artificial Intelligence (AAAI) (pp. 176–181). Singh, M. P. (2013). Norms as a basis for governing sociotechnical systems. ACM Transactions on Intelligent Systems and Technology (TIST), 5, 21:1–21:23. Soeanu, A., Debbabi, M., Allouche, M., Bélanger, M., & Léchevin, N. (2016). Hierarchy aware distributed plan execution monitoring. Expert Systems with Applications, 43, 66–81. Spoletini, P., & Verdicchio, M. (2009). An automata-based monitoring technique for commitment-based multiagent systems. In Coordination, Organizations, Institutions and Norms in Agent Systems (COIN) (pp. 172–187). Springer volume 5428 of Lecture Notes in Computer Science. Sun, S. X., Zhao, J., & Wang, H. (2012). An agent based approach for exception handling in e-procurement management. Expert Systems with Applications, 39, 1174–1182. Torroni, P., Yolum, P., Singh, M. P., Alberti, M., Chesani, F., Gavanelli, M., Lamma, E., & Mello, P. (2009). Modelling interactions via commitments and expectations. In Handbook of Research on Multiagent Systems: Semantics and Dynamics of Organizational Models (pp. 263–284). Vasconcelos, W. W., Garcı́a-Camino, A., Gaertner, D., Rodrı́guez-Aguilar, J. A., & Noriega, P. (2012). Distributed norm management for multiagent systems. Expert Systems with Applications, 39, 5990–5999. Vasconcelos, W. W., Kollingbaum, M. J., & Norman, T. J. (2009). Normative conflict resolution in multiagent systems. Autonomous Agents and Multiagent Systems, 19, 124–152. Venkatraman, M., & Singh, M. P. (1999). Verifying compliance with commitment protocols. Autonomous Agents and Multiagent Systems, 2, 217–236. Verdicchio, M., & Colombetti, M. (2003). A logical model of social commitment for agent communication. In Proceedings of the 2nd International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 528–535). Xu, D., Wijesooriya, C., Wang, Y.-G., & Beydoun, G. (2011). Outbound logistics exception monitoring: A multi-perspective ontologies’ approach with intelligent agents. Expert Systems with Applications, 38, 13604–13611. Yolum, P., & Singh, M. P. (2002). Flexible protocol specification and execution: applying event calculus planning using commitments. In Proceedings of the 1st International Conference on Autonomous Agents and Multiagent Systems (AAMAS) (pp. 527–534). 26 Appendix A. Proofs Proof of Theorem 1 Given a commitment Cm ∈ CT , and an agent X ∈ A, if X � ∅ δout Cm and Ci ∈ δout, then (Ci, Cm) ∈ MT . Proof 1. We prove soundness by contradiction. Given an agent X and two commitments Ci, Cm, assume that X � ∅ δout Cm and Ci ∈ δout, and (Ci, Cm) /∈ MT . We have the following possibilities: • Rule L1 applies, and Ci is an improper delegation of Cm. MT finds it since both Cm and Ci are members of CT . Thus, (Ci, Cm) ∈ MT . We reach a contradiction. • Rule S1 applies. There is no improper delegation of Cm, but other (proper) delegations exist. • Rule S2 applies. There are no delegations of Cm. Note that Rule L2 does not hold initially since the monitoring result is not empty. The first case immediately ends with a contradiction while the second and third cases propagate monitoring to other agents. At some point, delegations of Cm cease to exist 7. Let Z be the last agent that delegates Cm. Let us review each monitoring case: • Rule L2 does not apply since δout cannot be empty. • Rule S2 does not apply since Z delegated Cm. • Assume rule L1 applies, and Cj is an improper delegation of Ci (which is a delegation of Cm). MT finds it since Cm, Ci and Cj are all members of CT . Thus, (Cj , Cm) ∈ MT . We reach a contradiction. • Assume rule S1 applies, and let the delegatee be W . For W , let us review each monitoring case: – Rule L1 does not apply since δimp is empty. – Rule S1 does not apply since δpro is empty. – Rule S2 does not apply since the debtor of the commitment Cm is W itself. – Assume rule L2 applies. This results in an empty set for the monitoring result. However, δout cannot be empty. We reach a contradiction. This demonstrates that every possible case of local monitoring, that identifies an exception, leads to a contradic- tion against the global monitoring process not identifying that exception, thus proving soundness. 7Since protocol trace time is fixed, infinite delegations cannot occur. 27 Proof of Theorem 2 ∀ (Ci, Cm) ∈ MT , ∃ an agent X ∈ A and a derivation X � ∅ δout Cm such that Ci ∈ δout. Proof 2. We prove completeness by contradiction. Let us consider two commitments Ci and Cm, such that (Ci, Cm) ∈ MT and ∀X, δout X � ∅ δout Cm and Ci /∈ δout. Then, by Definition 17 dlgimp(Ci, Cm). Let debtor(Cm, X). The following cases are possible (Definition 16): • Ci is an improper consequent delegation of Cm, or • Ci is an improper antecedent delegation of Cm, or • Ci is an improper causal delegation of Cm, or • ∃ Cj ∈ C such that dlg(Cj, Cm) and dlgimp(Ci, Cj ). In the first three cases, X ▽REC C, 〈Cm, δpro, δimp〉 ∈ C and Ci ∈ δimp. Then by rule L1, X � ∅ δout Cm and Ci ∈ δout. Now, assume X � ∅ δout Cm where Ci /∈ δout. If X is the debtor of Cm, then Ci is also a commitment of X. When X queries the REC reasoner, < Cm, δpro, {Ci, ...}> ∈ C. Thus, rule L1 applies with Ci ∈ δout. We reach a contradiction. In the last case where there is a delegation chain rooted in Cm and that includes Cj , let Y be the delegatee or debtor of Cj and Y 6= X. When Y is the delegatee, then rule S1 applies. When Y is the debtor, then rule S2 applies. Both cases lead to rule S3. Rule S3 recursively starts a new derivation starting from Y . By iterating the same reasoning, we eventually reach the case where Cj is a direct delegation of Cm (since the delegation chain is finite). Thus, rule L1 applies as above. We reach a contradiction. This demonstrates that every possible case of global monitoring, that identifies an exception, leads to a contradic- tion against the local monitoring process not identifying that exception, thus proving completeness. 28 Appendix B. Implementation We have provided an implementation for COMODO using the ComMon tool. Below we explain some important code segments8 from the case study presented in Section 5. The ComMon tool only needs Java. The simplest way to run the example is to execute java -jar ComMon.jar (or double-click on the ComMon.jar file icon) on a selected agent folder. To run tests such as this one, select tab (Model) from the left-hand side menu and copy-paste the KB of your agent of choice. Then hit the Run, and copy-paste on the right-hand box called trace the desired evolution of events. Once the events are in place, select Start and then Log from the bottom. Use Stop to restart and Export to save the output on a file. Now, we describe parts of the REC code. Listing 1 shows the commitment theory that is shared by all the agents. First, the states of the commitments are described. Note that, in addition to the four states described in Section 2, we have detached to describe a conditional commitment that has become active. This is for implementation purposes so that we do not lose track of origin of the active commitment (i.e., the original condition commitment). Then, the rules that describe the state transitions are defined. Following the Event Calculus, in REC, we can express that an event initiates (or terminates) a temporal fluent, by way of initiates(Event, Fluent, Time) relations. A commitment with its state is considered a temporal fluent. Listing 2 shows the rules that describe the domain model of Amazon. This covers most of the process, and the domain models for other agents are described similarly. First, an exception is described either as a direct improper delegation, or an indirect improper delegation. Then, the rules for fluent manipulation are given in terms of action- consequence relations. For example, a payment from the customer to Amazon initiates the fluent paid at the time of the event. The rules for contract execution are given in terms of commitment create operations. For example an offer from Amazon to the customer creates a conditional commitment between the two agents regarding the Prime delivery scheme. Note that this exact rule is also contained in the customer’s domain model as they share this commitment. However, not all such rules are in the customer’s domain model, e.g., the details of the transaction between Amazon and the office is omitted from the customer. Here, we also support conjunction of fluents for the consequents of commitments. If the consequent of a commit- ment is a conjunction of fluents, then we represent it as a Prolog list, which contains all the fluents that are elements of the conjunction. We describe how delegations with conjunctions are handled below. Listing 3 shows the rules that describe explicit delegation, which is based on the discussion in Section 3.2. Other delegation types are described similarly. Delegations with conjunction of fluents is handled by parsing the list of fluents that make up the conjunction. Note that the deadline intervals are not taken into consideration while describing the delegation similarity relations. The description for improper (causal) delegation is given in Listing 4 by taking into consideration the deadline intervals of the commitments (see Section 3.4). 8The complete implementation can be downloaded from http://mas.cmpe.boun.edu.tr/ozgur/code.html, Section 3. 29 Listing 1: Commitment theory. % commitment states conditional(C, T):- holds_at(status(C, conditional), T). detached(C, T):- holds_at(status(C, detached), T). active(C, T):- holds_at(status(C, active), T). fulfilled(C, T):- holds_at(status(C, fulfilled), T). violated(C, T):- holds_at(status(C, violated), T). % create as conditional or active initiates(E, status(C, conditional), T):- ccreate(E, C, T). initiates(E, status(C, active), T):- create(E, C, T). % conditional to active terminates(E, status(C1, conditional), T):- detach(E, C1, C2, T). initiates(E, status(C1, detached), T):- detach(E, C1, _, T). initiates(E, status(C2, active), T):- detach(E, _, C2, T). detach(E, c(Tc, X, Y, Q, _, P, r(T1,T2)), c(Tc, X, Y, true, _, P, a(T3,T4)), T):- conditional(c(Tc, X, Y, Q, _, P, r(T1,T2)), T), initiates(E, Q, T), T3 is T + T1, T4 is T + T2. % active to fulfilled terminates(E, status(C, active), T):- discharge(E, C, T). initiates(E, status(C, fulfilled), T):- discharge(E, C, T). discharge(E, c(Tc, X, Y, true, _, P, a(T1,T2)), T):- active(c(Tc, X, Y, true, _, P, a(T1,T2)), T), T >= T1, T =< T2, initiates(E, P, T). % active to fulfilled terminates(E, status(C, active), T):- violate(E, C, T). initiates(E, status(C, violated), T):- violate(E, C, T). violate(_, c(Tc, X, Y, true, _, P, a(T1, T2)), T):- active(c(Tc, X, Y, true, _, P, a(T1, T2)), T), T > T2. 30 Listing 2: Domain model (Amazon). % exception model initiates(_, exception(C1, C2), T):- holds_at(improperDelegation(C1, C2), T). initiates(_, exception(C1, C2), T):- holds_at(improperDelegation(C1, C), T), active(C2, T), delegation(C, C2). % fluent manipulation initiates(exec(pay(customer, amazon, Item)), paid(Item), _). initiates(exec(confirm(amazon, office, Item)), confirmed(Item), _). initiates(exec(package(office, amazon, Item)), packaged(Item), _). initiates(exec(deliver(ups, customer, Item)), delivered(Item), _). % contract execution ccreate(exec(offer(amazon, customer, Item)), c(T, amazon, customer, paid(Item), delivered(Item), r(12,36)), T):- prime(customer). ccreate(exec(offer(ups, amazon, Item)), c(T, ups, amazon, packaged(Item), delivered(Item), r(6,24)), T). create(exec(confirm(amazon, office, Item)), c(T, office, amazon, true, packaged(Item), rel(6,24)), T). 31 Listing 3: Delegation model. % cases of explicit delegation explicitDelegation(c(Tc1, Z, Y, true, _, P1, _), c(Tc2, X, Y, true, _, P2, _)):- partOf(P1, P2), Tc1 > Tc2, X \= Z. explicitDelegation(c(Tc1, Z, Y, _, _, P1, _), c(Tc2, X, Y, true, _, P2, _)):- partOf(P1, P2), Tc1 > Tc2, X \= Z. explicitDelegation(c(Tc1, Z, Y, true, _, P1, _), c(Tc2, X, Y, _, _, P2, _)):- partOf(P1, P2), Tc1 > Tc2, X \= Z. explicitDelegation(c(Tc1, Z, Y, _, _, P1, _), c(Tc2, X, Y, _, _, P2, _)):- partOf(P1, P2), Tc1 > Tc2, X \= Z. % conjunction partOf(P, P). partOf(P, [P|_]). partOf(P, [_|L]):- partOf(P, L). partOf([P|L1], L2):- partOf(P, L2), partOf(L1, L2). Listing 4: Improper delegation. % improper causal delegation initiates(_, improperDelegation( c(Tc3, X3, Y3, true, _, P3, a(T5, T6)), c(Tc1, X1, Y1, true, _, P1, a(T1, T2))), T):- active(c(Tc1, X1, Y1, true, _, P1, a(T1, T2)), T), conditional(c(Tc2, X2, Y2, Q2, _, P2, r(T3, T4)), T), active(c(Tc3, X3, Y3, true, _, P3, a(T5, T6)), T), implicitDelegation(c(Tc2, X2, Y2, Q2, _, P2, r(T3, T4)), c(Tc1, X1, Y1, true, _, P1, a(T1, T2))), antecedentDelegation(c(Tc3, X3, Y3, true, _, P3, a(T5, T6)), c(Tc2, X2, Y2, Q2, _, P2, r(T3, T4))), (T4 + T6) > T2. 32 
work_tiiluctv7ndclmyslqke6472ay	----	Path relinking for the fixed spectrum frequency assignment problem Path relinking for the fixed spectrum frequency assignment problem Submitted by Jin-Kao Hao on Mon, 01/26/2015 - 10:31 Titre Path relinking for the fixed spectrum frequency assignment problem Type de publication Article de revue Auteur Lai, Xiangjing [1], Hao, Jin-Kao [2] Editeur Elsevier Type Article scientifique dans une revue à comité de lecture Année 2015 Langue Anglais Pagination 4755-4767 Volume 42(10) Titre de la revue Expert Systems with Applications ISSN 0957-4174 Résumé en anglais The fixed spectrum frequency assignment problem (FS-FAP) is a highly relevant application in modern wireless systems. This paper presents the first path relinking (PR) approach for solving FS-FAP. We devise four relinking operators to generate intermediate solutions (or paths) and a tabu search procedure for local optimization. We also adopt a diversity-and-quality technique to maintain population diversity. To show the effectiveness of the proposed approach, we present computational results on the set of 42 benchmark instances commonly used in the literature and compare them with the current best results obtained by any other existing methods. By showing improved best results (new upper bounds) for 19 instances, we demonstrate the effectiveness of the proposed PR approach. We investigate the impact of the relinking operators and the population updating strategy. The ideas of the proposed could be applicable to other frequency assignment problems and search problems. URL de la notice http://okina.univ-angers.fr/publications/ua7075 [3] DOI 10.1016/j.eswa.2015.01.025 [4] Lien vers le document http://dx.doi.org/10.1016/j.eswa.2015.01.025 [4] Liens [1] http://okina.univ-angers.fr/publications?f[author]=25597 [2] http://okina.univ-angers.fr/jinkao.hao/publications [3] http://okina.univ-angers.fr/publications/ua7075 [4] http://dx.doi.org/10.1016/j.eswa.2015.01.025 Publié sur Okina (http://okina.univ-angers.fr) http://okina.univ-angers.fr/publications?f[author]=25597 http://okina.univ-angers.fr/jinkao.hao/publications http://okina.univ-angers.fr/publications/ua7075 http://dx.doi.org/10.1016/j.eswa.2015.01.025 http://dx.doi.org/10.1016/j.eswa.2015.01.025 http://okina.univ-angers.fr 
work_tj5danyxxfc6fdabrx7mnz56fe	----	NASA Technical Reports Server (NTRS) 
work_tlnnymyub5cjpn4mlig4sao3fu	----	Secure communication for electronic business applications in mobile agent networks Secure communication for electronic business applications in mobile agent networks Woei-Jiunn Tsaur Department of Information Management, Da-Yeh University, Changhua 515, Taiwan a r t i c l e i n f o Keywords: Mobile agent Electronic business applications Smart card Proxy signature Network security a b s t r a c t The mobile agent plays an increasingly important role in electronic business applications, because it can provide the essential properties of personalization, automation and intelligence, etc. This paper proposes several appropriate security schemes for protecting mobile agent networks in electronic business appli- cations. As far as mobile agent security is concerned, we develop a proxy signature scheme for protecting mobile agents against malicious agent hosts. The proposed proxy signature scheme can protect users’ pri- vate keys stored in smart cards, and provide the fairness of contracts signed by agents. In addition, we also design a proxy authenticated encryption scheme so that the signature of the contracts will satisfy users’ constraints, and the non-repudiation of servers can be achieved. On the other hand, as far as agent host security is concerned, we apply the idea of proxy signature to construct an authentication scheme for protecting agent hosts. This scheme is to achieve the requirements of authentication and authoriza- tion. Furthermore, we also implement the proposed security schemes to achieve security requirements of confidentiality, integrity, authenticity, and non-repudiation for protecting Linux-based mobile agents and hosts in an electronic auction application. Hence, we affirm that the proposed security schemes are suit- able for practical electronic business applications in mobile-agent-based network environments. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction In recent years, there are many business applications based on mobile agent on a variety of networks (Benouhiba & Nigro, 2006; Kang, Lee, & Choi, 2008; Kim, Kwon, & Kwak, 2010; Park, Kang, & Kim, 2006; Wu et al., 2010; Yun, Lee, Yu, & Choi, 2009). The agents of the business applications usually provide personalization, auto- mation and intelligence, etc. However, it also results in many secu- rity threats such as stealing data from hosts by agents and tampering constraints of agents by hosts. For instance, when a mo- bile agent carrying a user’s private key roams among servers on the Internet, the agent may find a bid satisfies the user’s constraints, and then sign the bid (Chess et al., 1995; White, 1994). However, users will not wish to equip agents with their private signature keys when the agents may execute on untrusted agent hosts (Maes, Guttman, & Moukas, 1999; Sander & Tschudin, 1998; Takeda, Iino, & Nishida, 1995). On the other hand, a problem specific to mobile agents is the protection of the agent platforms running the agents. A hostile agent can destroy the hard drive, steal data, or do all sorts of undesirable operations to agent platforms. In this paper we will develop efficient security schemes based on cryptographic solutions (Mambo, Usuda, & Okamoto, 1996; Sander & Tschudin, 1998) for prevention of both agents and hosts tampering. This paper develops a proxy signature scheme and a proxy authenticated encryption scheme for protecting mobile agents against malicious agent hosts using the proposed ECC-based self- certified public key cryptosystem. The proposed proxy signature scheme can protect users’ private keys stored in smart cards, and provide the fairness of contracts signed by agents. The proposed cryptosystem is constructed using the ECC, and it also integrates the identity-based public key cryptosystem with the self-certified public key cryptosystem (Girault, 1992; Petersen & Horster, 1997; Saeednia, 1997, 2003) to provide higher security strength. Furthermore, based on the proposed cryptosystem, we employ the proposed proxy signature scheme to further design a proxy authenticated encryption scheme so that the signature of the con- tracts will satisfy users’ constraints, and the non-repudiation of servers can be achieved. In summary, these proposed schemes are able to accomplish the security requirements of confidentiality, integrity, authenticity, and non-repudiation for protecting mobile agents in electronic business applications. On the other hand, this paper also presents an authentication scheme for protecting mo- bile agent hosts against unauthorized mobile agents. In such a scheme, a mobile agent can register once to the system authority for several services in the mobile-agent-based networks. Finally, we implement the proposed security schemes for protecting Li- nux-based mobile agent networks in an electronic auction application. 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.07.105 E-mail address: wjtsaur@yahoo.com.tw Expert Systems with Applications 39 (2012) 1046–1054 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.07.105 mailto:wjtsaur@yahoo.com.tw http://dx.doi.org/10.1016/j.eswa.2011.07.105 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa The rest of this paper is organized as follows. In Section 2, we briefly describe the elliptic curve cryptosystems. Section 3 first develops an efficient public key cryptosystems, and then several security schemes constructed using it are designed for protecting mobile-agent-based electronic business applications. In Section 4, security analyses about attacks on the proposed schemes consoli- date the feasibility of the schemes. Performance evaluation of the proposed schemes, which is measured by the required computa- tional effort and communicational cost, is given in Section 5. In Section 6, we present the implementation of the proposed schemes on an electronic auction application. Finally, some concluding re- marks are presented in Section 7. 2. Elliptic curve cryptosystems (ECC) Assume that P is a point with order n on an elliptic curve E: y2 = x3 + ax + b (mod p), where 4a3 + 27b2 – 0 (mod p), and Q is some other point on the same curve. Let P = (xP, yP), Q = (xQ, yQ), and P + Q = R = (xR, yR). If xP – xQ, set k ¼ yQ�yP xQ�xP ; if xP = xQ, set k ¼ 3x 2 P þa 2yP . Then the point R = (xR, yR) can be defined by using the fol- lowing formulae: xR ¼ k2 � xP � xQ yR ¼ðxP � xRÞk � yP In ECC, the elliptic curve discrete logarithm problem (ECDLP) is to determine an integer x (0 6 x 6 n � 1) such that Q = x � P if such an x exists. As long as n and p are large enough, it is computationally intractable to find x with knowing E, Q, and P. Koblitz (1987) and Miller (1986) implemented this characteristic to elliptic curve cryp- tosystems. We need 1024-bit keys when using modular exponenti- ation schemes, like RSA or ElGamal cryptosystems, but we can get the same security level only using 160 bits in ECC. In addition, RSA needs to generate Ni = pi � qi for each user, respectively; however, ECC generates only fixed public information stored at SA, and can afterwards uses it repeatedly. Therefore, the storage cost required by ECC is less than that required by RSA. 3. Security schemes for protecting mobile agent networks In this section, we first develop an efficient public key crypto- systems, and then several security schemes constructed using it are designed for protecting mobile-agent-based electronic busi- ness applications. 3.1. Initialization The entities in the system are a system authority (SA), users (Ui), hosts (Hi), and mobile agents (MA) generated by specific users. We assume that SA is responsible for key generation and user registra- tion. We then define the notations used in the proposed schemes as follows: � p: a field size, where p is typically either an odd prime or a power of 2 in general applications, and its length is about 160 bits. � An elliptic curve E over Fp: E: y2 = x3 + ax + b, where the two field elements a, b 2 Fp and 4a3 + 27b2 – 0 (mod p), and all the points (x, y), x 2 Fp, y 2 Fp, on E form the set of E(Fp) containing a point O called the point at infinity. � B: a base point of order n over E(Fp), where n is a large prime (160 bits) and the number of Fp-rational points on E, denoted by # E(Fp), is divisible by n. � sSA: SA’s private key, where sSA 2 [2, n � 2]. � PSA: SA’s public key, where PSA = sSA � B (‘‘�’’ means the multiplica- tion of a number and an elliptic curve point.). � h( ): a one-way hash function that accepts a variable length input and produces a fixed length output value j, where j 2 [2, n � 2] and its length is 160 bits. The one-way hash func- tion h( ) should satisfy the properties (Harn, 1994) that given h(x), it is computationally infeasible to find x0 – x such that h(x0) = h(x), meanwhile, h(x0) – h(x) if and only if x0 – x. � X(P): output the x-coordinate of point P. After that, SA publishes E, B, p, n, PSA and h, while keeping sSA secret. 3.2. The proposed public key cryptosystems The operations of the proposed public key cryptosystems are di- vided into two phases: the system setup phase and the key gener- ation phase. 3.2.1. The system setup phase SA creates a system public key and some public parameters in this phase, and then SA releases these parameters. SA randomly chooses a numbersSA and keeps it secret. Then SA computes the system public key PSA ¼ sSA � B 3.2.2. The key generation phase User Ui and host Hi perform the following steps to register to SA, and obtain the corresponding public key, respectively. They also compute their private keys in this phase. Step 1. Ui and Hi execute the following tasks, respectively: (1-1) Select an identity information, denoted by Ii. (1-2) Randomly choose an integer xi 2 [2, n � 2] as the master key. (1-3) Compute Zi = h(xikIi) � B (1-4) Submit {Ii, Zi} to SA. Step 2. SA executes the following tasks for Ui and Hi, respectively: (2-1) Randomly choose a time-variant integer ki 2 [2, n � 2]. (2-2) Compute a public key Pi and its witness wi, where Pi ¼ Zi þðki � hðIiÞÞ � B ¼ðPix; PiyÞ wi ¼ ki þ sSA � ðPix þ hðIiÞÞ ðmod nÞ (2-3) Return {Pi, wi} to Ui and Hi, respectively. Step 3. Ui and Hi then execute the following tasks, respectively: (3-1) Calculate their own private keys as si ¼ wi þ hðxikIiÞ ðmod nÞ (3-2) Verify the authenticity of Pi by testing if si � B ¼ Pi þ hðIiÞ � B þ½ðPix þ hðIiÞÞ mod n� � PSA ð1Þ If the verification result of Eq. (1) is correct, then the partici- pant’s public key is Pi and the corresponding private key is si; otherwise, it means that the public key Pi is altered in the transmission. For the consideration of security, the private key si is stored in a smart card for subsequent electronic business applications. In the following, we show that the private key si and the corre- sponding public key Pi satisfy Eq. (1). Theorem 1. User Ui and host Hi can utilize Eq. (1) to verify his/her public key Pi by himself/herself. W.-J. Tsaur / Expert Systems with Applications 39 (2012) 1046–1054 1047 https://isiarticles.com/article/3791 
work_tmca7pcswbd2zjvxcp6yg3azbi	----	VODKA: Variant objects discovering knowledge acquisition Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 36 (2009) 2433–2450 Expert Systems with Applications VODKA: Variant objects discovering knowledge acquisition Shian-Shyong Tseng a,b,*, Shun-Chieh Lin a a Department of Computer and Information Science, National Chiao Tung University, No. 1001, Ta Hsueh Road, Hsinchu 300, Taiwan, ROC b Department of Information Science and Applications, Asia University, No. 500, Lioufong Road, Wufong Shiang, Taichung 413, Taiwan, ROC Abstract Many knowledge acquisition methodologies have been proposed to elicit rules systematically with embedded meaning from domain experts. But, none of these methods discusses the issue of discovering new modified objects in a traditional classification knowledge based system. For experts to sense the occurrence of new variants and revise the original rule base, to collect sufficient relevant information becomes increasingly important in the knowledge acquisition field. In this paper, the method variant objects discovering knowledge acquisition (VODKA) we proposed includes three stages (log collecting stage, knowledge learning stage, and knowledge polishing stage) to facilitate the acquisition of new inference rules for a classification knowledge based system. The originality of VODKA is to identify these new modified objects, the variants, from the way that the existing knowledge based system fails in applying some rules with low certainty degree. In this method, we try to classify the current new evolving object identified according to its attributes and their corre- sponding values. According to the analysis of the collected inference logs, one of the three recommendations (including adding a new attribute-value of an attribute, modifying the data type of an attribute, or adding a new attribute) will be suggested to help experts observe and characterize the new confirmed variants. VODKA requires E-EMCUD (extended embedded meaning capturing and uncertainty deciding). EMCUD is a knowledge acquisition system which relies on the repertory grids technique to manage object/attri- bute-values tables and to produce inferences rules from these tables. The E-EMCUD we used here is a new version of EMCUD to update existing tables by adding new objects or new attributes and to adapt the original embedded rules. Here, a computer worm detection prototype is implemented to evaluate the effectiveness of VODKA. The experimental results show that new worm variants could be discovered from inference logs to customize the corresponding detection rules for computer worms. Moreover, VODKA can be applied to the e-learning area to learn the variant learning behaviors of students and to reconstruct the teaching materials in improving the performance of e-learners. � 2007 Elsevier Ltd. All rights reserved. Keywords: Knowledge acquisition; Variant discovering; EMCUD; VODKA; Computer worm 1. Introduction As time goes on, new objects in many domains are incre- mentally evolving due to the environment changes and the knowledge explosion. Besides unchangeable objects in the course of time, some previously created objects may 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.12.055 * Corresponding author. Address: Department of Computer and Infor- mation Science, National Chiao Tung University, No. 1001, Ta Hsueh Road, Hsinchu 300, Taiwan, ROC. Tel.: +886 3 5712121x56658; fax: +886 3 5721490. E-mail addresses: sstseng@cis.nctu.edu.tw (S.-S. Tseng), jielin@cis. nctu.edu.tw (S.-C. Lin). become obsolete and disappear after a long period of time, and some new objects might be added or modified for adapting to the dynamic environmental changes with the time, which is the phenomenon of object evolution. In this paper, knowledge can be classified as static knowledge and dynamic knowledge according to the stability of knowledge in a dynamic environment. Static knowledge remains the same while dynamic knowledge will be updated when environment changes. For example, due to the nature of evolution, some well-known virus could evolve into a new virus which, suggested by the experts, should be sin- gled out as a new variant to classify all kinds of viruses definitely. mailto:sstseng@cis.nctu.edu.tw mailto:jielin@cis. nctu.edu.tw mailto:jielin@cis. nctu.edu.tw 2434 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 As we know, a knowledge based system (KBS) is an intelligent computer program using knowledge and infer- ence procedures to solve problems which are difficult enough to require significant human expertise for the solu- tions, such as disease diagnosis, investment prediction, or computer science problems. Since the phase of knowledge acquisition is a bottleneck when constructing a KBS, many methodologies and related tools, e.g., NeoETS (Boose & Bradshaw, 1986), AQUINAS (Boose & Bradshaw, 1987), KITTEN (Shaw & Gaines, 1987), EMCUD (Hwang & Tseng, 1990), KADS (Wielinga, Schreiber, & Breuker, 1992), KANAL (Kim & Gil, 2001), OMCS-2 (Singh et al., 2002), INDUS (Caragea et al., 2005) have been pro- posed to rapidly build prototypes in the past twenty years. Most of the existing knowledge acquisition systems employ the repertory grid test, originally developed by per- sonal construct theory (PCT) for eliciting static knowledge, to identify different objects and distinguishing these objects in a selected domain. PCT can help domain experts formu- late the quality of static knowledge with/without knowl- edge engineers. But only the static knowledge used for classifying and distinguishing well-known objects by a finite set of training cases in most the systems, such as ID3 (Quinlan, 1986), KAISER (Tsujino, Takegaki, & Nishida, 1990; Tsujino, Dabija, & Nishida, 1992), C- KAT (Zacklad & Fontaine, 1995), version space (Hong & Tseng, 1997), AVT-DTL (Zhang & Honavar, 2003). And few of them attack the topic of acquiring dynamic knowledge, in fact, the incremental learning of dynamic knowledge by continuously being aware of the variant objects as time goes on is important Consequently, the lack of sufficient information about variants may result in a problem of observing the occur- rence of evolving objects for human experts. Now, how to collect sufficient relevant information to help experts sense the occurrence of variants and reuse the original rule base becomes one of the important issues. Embedded meaning capturing and uncertainty deciding (EMCUD), one of a repertory grid based knowledge acqui- sition tools, was proposed to elicit the embedded meanings of knowledge (embedded rules bearing on objects and object attributes) to classify objects. Based upon an attri- bute ordering table (AOT) to record the relative impor- tance of each attribute to each object, EMCUD guides the experts to determine the certainty degree of each embedded rule, in order to extend the coverage of original rules generated by traditional repertory grid. With different certainty degree, some variant objects modified from original objects may not be classified by the original rules but may be classified by the other embed- ded rules; even these embedded rules may have margin val- ues of certainty factor (CF) due to the weak suggestions of domain experts. Since the higher the CF value, the more reliable the results are, this kind of objects could be singled out as a variant object class because the new characteristics of these objects are emerging. Therefore, if the variants could be detected and recommended as the new objects by experts, the related ambiguous attributes (minor attri- butes) could be refined or new attributes could be added to improve the classification ability in a KBS. These attri- butes might result in the marginally acceptable CF values of weak embedded rules. In this paper, a new iterative methodology, variant objects discovering knowledge acquisition (VODKA) is proposed to learn new variant objects based upon monitor- ing the inference behaviors of weak embedded rules. We focus our attention on incrementally collecting new cases of modified objects and learning the dynamic knowledge to modify the original classification KBS. In order to clearly single a new variant object out, the previous meth- odologies need to be applied recurrently, and therefore the originally elicited knowledge might not be easily reused. The goal of VODKA is to facilitate the acquisition of new inference rules of the new modified objects for a clas- sification KBS, from which new modified objects will be identified from the attribute values. The new rules, gener- ated by VODKA, will be able to cope with the new variant objects. They are similar to previously known embedded rules in the KBS. Since the relative importance of each attribute can be represented by using AOT, some minor attributes will be relaxed or ignored in order to capture the embedded mean- ings of embedded rules with acceptable CF values in EMCUD. In other words, these kinds of attributes are not used to classify the object. New variants classified by weak embedded rules should be singled out of original objects. Obtained from the collection of minor attribute- value pairs, the fired frequency of weak embedded rules is a big help for the experts to discover new information of variant objects. These attribute-value pairs are then to be provided to help experts determine whether to single variant objects out or not. According to the analysis of the collected inference logs, one of the three different recommendations: adding a new attribute-value of a minor attribute, modifying the data type of a minor attribute, or adding a new attribute, will be suggested to help experts characterize the corresponding rules to refine the KBS if a new variant object is confirmed to be singled out. To evaluate the performance of VODKA, a simple com- puter worm detection prototype system with VODKA and worm classification embedded rule base based upon DRAMA (Lin, Tseng, & Tsai, 2003) has been implemented to discover the new variant worms generated by an attack- ing traffic generator. Experts can be helped by VODKA to quickly and easily single out the new variant objects and customize the detection rules for computer worms. Also, VODKA has been applied on an e-learning domain to learn the variant learning behaviors of e-learners, which shows that VODKA is a big help for teachers to prepare or to reconstruct the teaching materials, the variant course materials, for these e-learners to easily understand the learning materials. In a word, VODKA can enhance the classification ability for new objects of a KBS since the S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 2435 new evolving objects can be incrementally learned and dis- covered by collecting the sufficient information. 2. Related work Several knowledge acquisition methodologies and related systems are introduced in this section first. Then repertory grid, one of the popular indirect knowledge acquisition techniques, is also discussed. Finally, the elicitation of embedded meaning and some problems of traditional knowledge acquisition methodologies are discussed. 2.1. Knowledge acquisition systems As we know, knowledge in many domains and the expe- rience of domain experts is continuously growing. Many knowledge acquisition methodologies have been proposed to help knowledge engineers acquire the useful knowledge and then to transfer this knowledge into a knowledge base or other computerized representation forms. In general, there are three approaches for knowledge acquisition (Crowther & Hartnett, 1996; Hwang & Tseng, 1990; Mcgraw & Harbison-Briggs, 1989): (1) Interviewing experts by experienced knowledge engineers: formalizing the elicited knowledge after interviewing experts. However, it is usually time-con- suming if the communication between domain experts and knowledge engineers is insufficient. (2) Machine learning: learning the knowledge by collect- ing many useful cases and instances with/without the involvement of domain experts. However, the quality of the results usually relies on the selected training cases. (3) Knowledge acquisition systems: assisting domain experts in generating useful rules using knowledge acquisition systems with/without the help of knowl- edge engineers. These tools could reduce the effort of communication between knowledge engineers and domain experts and could reduce the risk and difficulty of selecting the suitable training cases. In the past decades, many knowledge acquisition sys- tems, e.g., NeoETS (Boose & Bradshaw, 1986), AQUI- NAS (Boose & Bradshaw, 1987), KITTEN (Shaw & Gaines, 1987), EMCUD (Hwang & Tseng, 1990), KADS (Wielinga et al., 1992), MCRDR (Kang, 1996), KAMET (Cairo, 1998), MedFrame/CADIAG-IV (Boegl, 1997; Kolousek, 1997; Leitich et al., 2001), KANAL (Kim & Gil, 2001), OMCS-2 (Singh et al., 2002), INDUS (Caragea et al., 2005) have been developed to build prototypes and to iteratively elicit the knowledge from domain experts. However, most of them can not be used to construct the dynamic knowledge to classify the variant objects in a dynamic environment using the previous obtained attri- butes. For acquiring the dynamic knowledge, the tradi- tional approach needs to rerun their knowledge acquisition process. Ontology, which defines concepts and the relationships between concepts, was indicated to be useful in a knowl- edge acquisition process in recent years. In computer sci- ence area, the ontology is a conceptualized data structure to be used in knowledge systems. Based on the same ontol- ogy, different systems can communicate with each other, or the knowledge inside computer systems may be structured and presented more accurately. ‘‘An ontology may take a variety of forms, but necessarily it will include a vocabulary of terms, and some specification of their meaning. This includes definitions and an indication of how concepts are inter-related which collectively impose a structure on the domain and constrain the possible interpretations of terms” (Uschold, King, Moralee, & Zorgios, 1998). To build up the domain ontology, identifying the domain concepts is the first step. Moreover, a systematical approach to acquire the relations of concepts from exper- tise is also a key factor for ontology construction. In recent years, due to the increasing requirement for inducing domain knowledge into computer systems (Kitamura & Mizoguchi, 2003; Noy & Sintek, 2001), many researches (Alani & Kim, 2003; Eriksson, 2003; Fensel & Angele, 1988; Frank & Farquhar, 1999; Leibbink, Witteman, & Mayer, 2002; Maedche & Staab, 2001) were proposed to discover, represent, and use of ontologies. Especially in KBSs, ontology becomes a key factor to build a successful knowledge base with more meaningful knowledge content for users. Moreover, the software agents technique is used in knowledge acquisition systems to acquire knowledge from autonomous, distributed, and semantically heteroge- neous data sources with a domain specific ontology (Cara- gea, Silvescu, & Honavar, 2001; Caragea et al., 2005; Rosaci, 2005; Rosaci, Terracina, & Ursino, 2004). However, ontology used in these researches was usually assumed to be existent or be able to be obtained directly from experts. However, in real cases, when a knowledge engineer starts at designing a knowledge acquisition pro- cess, the ontology of the domain concepts is usually not available yet. It is usually a time-consuming process to build an ontology only by interviewing a domain expert. 2.2. Repertory grid methodology and relevant systems Repertory grid, based on Kelly’s personal construct theory (Kelly, 1955) which reports how people make sense of the world, could be used as an efficient knowledge acquisition technique in identifying different objects and distinguishing these objects in a domain. It is the basis of several computer assisted knowledge acquisition tools, such as ETS (Boose, 1984, 1985), AQUINAS (Boose & Bradshaw, 1987) and KSSO (Gaines, 1987). A single repertory grid represented as a matrix whose columns have element objects (labels) and whose rows have construct’s attributes (labels) can classify a class of objects, or individuals. The value assigned to an element-construct 2436 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 pair need not be Boolean. Grid values have numeric rat- ings, probabilities, and other characteristics, where each value reflects a degree. Then, the expert is asked to fill the grid with 5-scale ratings, where ‘‘1” represents the most relevant attribute to the object; ‘‘2” represents that the attribute may be relevant to the object; ‘‘3” represents ‘‘unknown” or ‘‘no relevance”; ‘‘4” represents that the object may have the opposite characteristic; ‘‘5” represents the most relevant opposite characteristic to the object. The whole concept of repertory grid technique can be described with the following steps: (1) Elicit all of the element objects, e.g., E1, E2, E3, E4, E5 from the expert. (2) Elicit the construct attributes (and their opposites), e.g., C1; C2; C3; C4ðC01; C 0 2; C 0 3; C 0 4Þ, from the expert. Each time three elements are chosen to ask for a con- struct to distinguish one element from the other two. (3) Rate all of the [element, construct] entries of the grid with value range from 1 to 5. An illustrative example is given in Table 1. As repertory grid technique has been widely used by researchers, some extensions have been made to enrich its representative ability for covering more knowledge, the value assigned to an element-construct pair may be Bool- ean, numeric ratings, probabilities, etc. For example, Dixit and Pindyck (1994), and Hwang (1995) extended the reper- tory grid technique to the fuzzy table, in which constructs were fuzzy attributes that could be rated by means of fuzzy linguistic terms from a finite set. Castro-Schez, Jennings, Luo, and Shadbolt (2004) developed a technique using a fuzzy repertory grid for acquiring the finite set of attributes or variables that the expert uses in a classification problem to characterize and discriminate a set of elements. Moreover, several models have been proposed for han- dling uncertainties in expert systems through generating more meaningful rules from the repertory grid oriented approaches. EMYCIN certainty factor (CF) model was first used to determine the degree of the belief of a rule for uncertain reasoning (Shortliffe & Buchanan, 1975). Embedded meaning capturing and uncertainty deciding (EMCUD) knowledge acquisition system was proposed to extract rules with embedded meaning from hierarchical repertory girds by defining the impacts of the constructs of each element (Hwang & Tseng, 1990) and was success- fully applied in a medical diagnostic system of acute exan- thema in Taiwan (Hwang & Tseng, 1991). WebGrid (Shaw Table 1 The illustrative example of a repertory grid with ratings Element construct E1 E2 E3 E4 E5 C1 5 1 5 1 1 C 0 1 C2 4 4 4 1 4 C 0 2 C3 1 4 5 1 4 C 0 3 C4 1 4 4 5 5 C 0 4 & Gaines, 1996), Calgary’s web-based knowledge modeling and inference tool, is based on repertory grid elicitation and analysis. Blythe, Kim, Ramachandran, and Gil (2001) proposed an acquisition interface that integrates previously developed techniques to guide users to set con- straints in different aspects of knowledge acquisition. Although these methodologies are proposed to extend the ability of uncertain reasoning to classify the well- known objects, none of them discusses the issue of discov- ering and classifying new variant objects. It is also difficult for experts to sense the occurrence of variant knowledge, which is modified in the dynamic environment as time goes on. Therefore, a new knowledge acquisition system based upon EMCUD is proposed in this paper to guide domain experts to create additional attributes for classifying new variant objects through the observations of the interested inference results. 2.3. Elicitation of embedded meanings The embedded meanings referred to here represent the information that domain experts take for granted but which is implicit to the people who are not familiar with the application domain. The lack of embedded meaning will probably make an expert system fail to infer some cases being trivial to experts. SEEK (Politakis & Weiss, 1984) and SEEK2 (Ginsberg, Weiss, & Politakis, 1988) have been proposed to obtain embedded meanings by some efficient refinement processes. However, the major problem of SEEK and SEEK2 is that the case database is assumed to be available although it is difficult to collect sufficient cases in some applications. Moreover, it would be also time-consuming and boring for experts to offer a conclu- sion for each case in the database before starting the refine- ment procedure. Thus, EMCUD (Hwang & Tseng, 1990) is proposed to elicit the embedded meanings of knowledge from the existing hierarchical repertory grids given by experts. Additionally, it will also guide experts to determine the certainty degree of each rule with embedded meaning for extending the coverage of generated original rules. To capture the embedded meanings of the resulting grids, the Attribute Ordering Table (AOT), which is used to record the relative importance of each attribute to each object, is employed. The values in each AOT entry, a pair of attribute and object, may be labeled ‘‘X”, ‘‘D” or an integer number. ‘‘X” means no relationship existing between the attribute and the object. ‘‘D” means that the attribute dominates the object, i.e., if the attribute is not equal to the entry value, it is impossible for the object to be implied. Integer numbers are used to represent the rela- tive importance degree of the attribute of the object instead of dominating the corresponding object. If the attribute does not equal the attribute-value, it is still for the object to be implied. A larger integer number implies that attri- butes must be more important to the object. Using AOT, the original rules generate some rules with embedded meaning, and the CF value of each rule, which is Table 2 The acquisition table of four computer worms Attribute Object Nimda CodeRed Blaster Welchia 100-thread (A1) X True X X System reboot (A2) X True True True DoS type (A3) Email flood TCP flood Windows Update flood ICMP flood Email attached file (A4) {sample.exe; puta!!scr} X X X TCP port (A5) X {80} {135; 4444} {80; 135} S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 2437 between �1 and 1, could be determined to indicate the degree of supporting the inference result. The higher the CF value, the more reliable the results are. The EMCUD algorithm is listed as follows. Algorithm 1. EMCUD algorithm Input: The hierarchical grids. Output: The guiding rules with embedded meaning. Step 1: Build the corresponding AOT with each grid of the hierarchical multiple grids. Step 2: Generate the possible rules with embedded meaning. Step 3: Select the accepted rules with embedded meaning through the interaction with experts. Step 4: Generate automatically the CF of each rule with embedded meaning. All rules generated by EMCUD can be categorized into two classes: original and embedded rules with acceptable CF value, and discarded rules with unacceptable CF value, according to the confidence degree of domain experts. To determine the CF value of each embedded rule, we have to firstly determine the upper and the lower bounds of CF values of accepted embedded rules. CF value of each rule can be automatically determined by a fuzzy mapping function. Thus, the useful embedded rules with acceptable CF values could be used to cover more uncertainty cases. Since embedded rules with weak acceptable CF values (the CF values below a user defined threshold) usually mean that domain experts might lack strong confidence, objects matching weak embedded rules derived from origi- nal objects may be the candidates for new variants. For example, the object satisfying the conditions (attribute- value pairs) of the embedded rules with CF = 0.5 means the expert might suggest that it would be marginally classi- fied into the object class and the minor attributes of the embedded rule might be not clearly defined. Therefore, the fired frequencies of this kind of weak embedded rules should be used to discover the occurrence of new variant objects. 2.4. Problems with conventional knowledge acquisition methods With the changing environment, the adaptation of the acquired rules should be required to cope with new vari- ants. However, experts may not be aware of the occurrence of new variant candidates and may have insufficient evi- dence to construct the knowledge of the variants using con- ventional repertory grid approaches. Although EMCUD could be used to generate more useful embedded rules for covering more similar objects in extended object classes, it still lacks the ability of grid evolution for singling these new variants out; e.g., EMCUD should manually regener- ate the original and embedded rules to classify these variant objects by interacting with domain experts after collecting sufficient information about these variants. Therefore, enhancing the adaptation ability of embedded rules becomes increasingly important to achieve the ability of grid evolution in a classification KBS. In this paper, the embedded rules from EMCUD are categorized into three classes: the original rules with strong CF values, the embedded rules with marginally acceptable CF values, and the discarded rules with low CF values. Hence, a new knowledge acquisition methodology is pro- posed to discover the occurrence of new variant objects using the fired frequency of embedded rules with margin- ally acceptable CF values. A simple computer worm detec- tion prototype in Example 1 is used to illustrate the inability for discovering variants using EMCUD. Example 1. The example of classifying four computer worms. In recent years, the number of computer worms is dramatically increasing to threaten the reliability of Inter- net. Table 2 shows the acquisition table of four computer worms (Kienzle & Elder, 2003; Mirkovic, Martin, & Reiher, 2002; Moore, Shannon, Voelker, & Savage, 2003; Weaver, Paxson, Staniford, Cunningham, & Maulik, 2003) including Nimda, CodeRed, Blaster, and Welchia using five attributes including 100-thread, System reboot, DoS type, Email attached file, and TCP port. The 100-thread means 100-threads with Boolean are simultaneously exe- cuted by one program. The system reboot Boolean attribute will be set to True if the system has been automatically rebooted. The attacking methodologies of worms could be classified into one kind of DoS type with String attribute (Mirkovic et al., 2002). The email attached file attribute with Set data type is also a useful attribute to classify these worms. Most of the worms could communi- cate with each other using different TCP port with Set data type. An example of constructing an AOT table from the acquisition table shown in Table 2 is given as follows: EMCUD: If DoS type is not equal to Email flood, is it possible for Nimda to be implied? EXPERT: No. 2438 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 The answer means the DoS type dominate Nimda, and hence AOT [Nimda, DoS type] = ‘‘D”. EMCUD: If Email attached file is not equal to any ele- ment of {sample.exe, puta!!scr}, is it possible for Nimda to be implied? EXPERT: YES. The answer means that Email attached file does not dominate Nimda. The questions for 100-thread and Nimda will not be asked, since the entry [Nimda, 100-thread] is labeled ‘‘X”. Therefore, the entry AOT [Nimda, 100- thread] is labeled ‘‘X”, too. This is the same for AOT [Nimda, System reboot] and AOT [Nimda, TCP port] entries. The entry AOT [Nimda, Email attached file] is set to be 1, since the Email attached file is the only attribute that does not dominate Nimda. If there are more than one attributes do not dominate the object, e.g., the System reboot, the DoS type, and the TCP port do not dominate Blaster, the following questions will be asked by EMCUD. (1) Is System reboot more important than DoS type? (2) Is System reboot less important than DoS type? (3) Is System reboot as important as DoS type? The expert indicates that System reboot is as important as DoS type to Blaster. Moreover, the expert also indicates Table 3 The AOT table of four computer worms Attributes Object Nimda CodeRed Blaster Welchia A1 X 2 X X A2 X 1 2 2 A3 D 1 2 1 A4 1 X X X A5 X X 1 2 Table 4 Partial detection rules generated by EMCUD Rule # Conditions A1 A2 A3 A R1,0 – – Email flood ( R1,1 – – Email flood : R2,0 True True TCP flood – R2,1 True False TCP flood – R2,2 False True TCP flood – R2,3 True False : (TCP flood) – R3,0 – True Windows update flood – R3,1 – True Windows update flood – R3,2 – False Windows update flood – R3,3 – True : (Windows update flood) – R3,4 – False Windows update flood – R4,0 – True ICMP flood – R4,1 – True : (ICMP flood) – R4,2 – True ICMP flood – R4,3 – True : (ICMP flood) – that System reboot is more important than TCP port to Blaster; and hence the entries AOT [Blaster, System reboot] = AOT [Blaster, DoS type] = 2 and AOT [Blaster, TCP port] = 1. After each entry value of AOT is deter- mined as shown in Table 3, the embedded meaning implied by the AOT can be extracted. Now we use the first column of Table 2 to show the information implied by an AOT. The column expresses the following meanings: (1) A3 dominates Nimda: If A3 is not equal to Email flood, it is impossible for Nimda to be implied. (2) A4 does not dominate Nimda: If A4 is neither equal to sample.exe nor puta!!scr, it is still possible for Nimda to be implied. In practice, the hierarchy rules can be generated while hierarchical grids are given. To simplify the discussion, Table 4 shows partial detection rules (simple rules) of a classification KBS based upon the Tables 2 and 3 to classify these worms using five attributes in single grids. Ri,j represents the jth highest rank of CF in object i, and the highest rank is 0. The R1,0 is the original rule of Nimda to classify the original Nimda objects and R1,1 is the embed- ded rule of Nimda to classify the extended Nimda objects. The Mask Table of minor attributes shown in Table 5 indicates the minor attributes for all embedded rules. Each row in Mask Table is a bit vector of attributes, where the ith bit is set to 1 representing the ith minor attribute that is negated or ignored. For example, the M2,3 (0, 1, 1, 0, 0) means the 2nd and 3rd minor attributes in R2,3 are ignored. On the Internet, each worm can be represented as a set of attribute-value pairs. We can automatically collect such attribute-value pairs and feed them into the classification KBS to classify them in the suitable category. Since new worms might have been derived from old worms, the difference between their attribute values seems to be slight. As mentioned above, EMCUD could generate lots of Conclusion CF 4 A5 Object sample.exe; puta!!scr) – Nimda 0.8 (sample.exe; puta!!scr) – Nimda 0.4 – CodeRed 0.8 – CodeRed 0.6 – CodeRed 0.4 – CodeRed 0.4 {135; 4444} Blaster 0.7 :{135; 4444} Blaster 0.57 {135; 4444} Blaster 0.43 {135; 4444} Blaster 0.43 :{135; 4444} Blaster 0.3 {80; 135} Welchia 0.8 {80; 135} Welchia 0.67 :{80; 135} Welchia 0.53 :{80; 135} Welchia 0.4 Table 5 The Mask Table of ignored attributes Mask # A1 A2 A3 A4 A5 M1,0 0 0 0 0 0 M1,1 0 0 0 1 0 M2,0 0 0 0 0 0 M2,1 0 1 0 0 0 M2,2 1 0 0 0 0 M2,3 0 1 1 0 0 M3,0 0 0 0 0 0 M3,1 0 0 0 0 1 M3,2 0 1 0 0 0 M3,3 0 0 1 0 0 M3,4 0 1 0 0 1 M4,0 0 0 0 0 0 M4,1 0 1 0 0 0 M4,2 0 0 0 0 1 M4,3 0 1 0 0 1 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 2439 embedded rules with different CF values for accommodat- ing the knowledge of the changed worms due to the property of minor attributes; e.g., R1,1 ‘‘IF (DoS type = E- mail flood) AND : (Email attached file = (sample.exe; puta!!scr)) THEN Nimda”, a marginally acceptable embed- ded rule with CF = 0.4, may be fired by a new Nimda variant which is treated as a member of original Nimda class. If this rule has been fired frequently due to a specific value of the attribute ‘‘email attached file”: ‘‘readme.exe” (more evidence of the occurrence of the candidates of Nimda variants have been gathered), a new original rule and an embedded rule could be generated: ‘‘IF (DoS type = Email flood) AND (Email attached file = read- me.exe) THEN Nimda.B”, a subset of extended Nimda object class namely Nimda.B, with CF = 0.8 and ‘‘IF (DoS Embedded Rule Base Inference Engine (DRAMA) Inference Log Experts New Variants Acquisition Meta Knowledge New Varian Acquisition Table Adjust Extended EMCUD Main Acquisition Table Update Fig. 1. The framewo type = Email flood) AND : (Email attached file = read- me.exe) THEN Nimda.B” with CF = 0.5. These rules help to single the Nimda.B class out of the extended Nimda object class. 3. Knowledge acquisition by discovering variant objects Although EMCUD and other similar approaches could be manually rerun to acquire variant knowledge from domain experts to classify new variant objects, it might be costly and hard to obtain the knowledge due to the insufficient information about variants. To simplify our discussion, let’s assume some objects in O1 class belong to the original object class (OO1) of O1, which can be clas- sified by original rules of O1. The other objects in O1 class classified by embedded rules of O1 belong to the extended object class (EO1) of O1, where OO1 � EO1. In the EO1, some modified objects can be classified by the embedded rules of O1 with weak CF values, which are singled out to be a variant object class (VO1) of O1 with the significant attributes emerged from minor attributes. That is, VO1 � EO1 and VO1 \ Oi = /, where 1 6 i 6 m, and m is the number of distinct object classes. Because the embed- ded rules with diverse CF values represent different sup- ports to classify objects, the ones with marginally acceptable CF values might be triggered by some candi- dates of new variant classes. Therefore, our idea is to ana- lyze the behaviors of weak embedded rules (the weak suggestions by experts) to construct the new variants acqui- sition table for extracting new variant knowledge. In the following, we will propose a new iterative knowl- edge acquisition methodology, variant objects discovering Real Instances Stage III: Knowledge Polishing Stage I: Log Collecting Stage II: Knowledge Learning ts User Inter- face User rk of VODKA. 2440 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 knowledge acquisition (VODKA) which is an iterative pro- cess as shown in Fig. 1, to provide the ability of grid evo- lution, where each iteration consists of three stages, log collecting stage, knowledge learning stage, and knowledge polishing stage, to analyze the inference behaviors of the embedded rules in a KBS. Firstly, the embedded rule base will be created according to the original main acquisition table using EMCUD or VODKA. Then the inference behaviors (facts/attribute- value pairs) will be collected iteratively to discover the can- didates of the variants during Stage I according to the meta knowledge. The ignored attribute-value pair of the minor attribute will be treated as an item and a set of ignored attribute-value pairs will be treated as a transaction to dis- cover the association between interesting attribute-value pairs in Stage II. Consequently, the new variants acquisi- tion table based on the discovered knowledge could be gen- erated by interacting with domain experts through the new variants acquisition procedure. Finally, the rules applicable to new variants will be incrementally generated and the main acquisition table will be iteratively adjusted using E-EMCUD in Stage III. The algorithm of VODKA is shown as follows. Algorithm 2. The algorithm of VODKA Input: The original main acquisition table T and embed- ded rule base RB. Output: The rules with embedded meaning about variants. Stage I: Collect all facts of the weak embedded rules as real inference log of the RB. Stage II: Generate the new variants acquisition table T0. Step 1: Discover large itemsets L using the inference log. Step 2: Generate T0 using L and additional attributes provided by experts. Stage III: Use E-EMCUD to generate rules of new variants. Step 1: Generate rules according to T0. Step 2: Merge T0 into original main acquisition table T. 3.1. Meta knowledge in log collecting stage Without loss of generality, let’s assume that there are k attributes to classify m objects in the main acquisition table. Thus, the total number of embedded rules is limited. In order to assist domain experts in noticing and analyzing the occurrence of the candidates of variant objects, the fol- lowing four meta rules are used to collect the inference log (fact/raw data) of weak embedded rules to help experts sense the occurrence of new variants. MR1: IF Ri,j is fired THEN Increase Ci,j by one. MR2: IF CF(Ri,j) 6 THCF, THEN Log Ri,j. MR3: IF Ci,j P THcnt AND CF(Ri,j) 6 THCF THEN Run New Variants Acquisition Algorithm to acquire the new variants acquisition table and Reset TimeOut. MR4: IF TimeOut = THPeriod THEN Run New Variants Acquisition Algorithm and Reset TimeOut. The meta rule MR1 is used to count the fired frequency of each embedded rule (Ci,j). The meta rule MR2 means that all facts (attribute-value pairs) of the embedded rules with marginally acceptable CF lower than strong CF bound threshold (THCF) are logged as a record, (Ri,j, A1, A2, . . . , Ak, CF(Ri,j)). The meta rule MR3 means that if there exists one weak embedded rule with fired frequency exceeding the warning line threshold (THCNT), new vari- ants may be discovered iteratively using VODKA. The meta rule MR4 means that VODKA will be executed peri- odically to refresh the new variants acquisition table. The TimeOut will be reset when MR3 or MR4 is triggered. 3.2. New variants discovering methodology at the knowledge learning stage At the knowledge learning stage, the new variants acqui- sition table will be generated through interacting with domain experts based upon the observation of inference log. An ignored attribute-value pair, i.e., (DoS type = TCP flood), is treated as an item and the transaction is repre- sented as a set of ignored attribute-value pairs, i.e., {(DoS type = TCP flood), (TCP port = {135; 4444})}. The inference log could be automatically transformed into the transaction database (D) and the item set (I) using the Mask Table of ignored attributes. In order to obtain the candidates of new variants, we apply Apriori algorithm (Agrawal, Imielinksi, & Swami, 1993) to discover large itemsets (L) that will provide more useful information. After generating the large itemsets, new variants acqui- sition table might be elicited following the new variants acquisition algorithm. The new objects using unclear attri- butes would be singled out accordingly, if the experts reconfirm the addition of the new variant object. Thus, one of three recommendations including adding a new attribute value to a minor attribute, modifying the data type of a minor attribute, adding a new attribute, will be further given to adjust the main acquisition table. If a new changing object is singled out, the new value of the minor attribute could be added to characterize of new objects. If the initial data type of a certain attribute is too rough to describe the object, a superior data type is rec- ommended and the values of the attribute in both original object and variant should be modified. For example, the BOOLEAN data type may be refined to SINGLE VALUE data type (Hwang & Tseng, 1990). If changing the data type still can not discriminate the new variants from origi- nal objects, acquiring a new attribute from domain experts will be suggested in VODKA. Thus, the new variants acquisition table will be created iteratively using the discov- ered large itemsets in Algorithm 3. However, adding a new S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 2441 attribute, which is very time-consuming because it results in creating a new row in new variant acquisition table, is the last choice for classifying variant objects. Algorithm 3. New variants acquisition algorithm Input: Inference log and the main acquisition table T, the minimal support d. Output: The new variant object class VO, new attribute set AN, and new variants acquisition table T0. Step 1: Transform inference log into the transaction data set D. Step 2: Discover large itemsets L by d using D. Step 3: For each large itemset, ask experts to determine whether it belongs to a new variant or not. Step 4: If a new variant is confirmed, ask experts to acquire the related information about this new variant. Tab Th Ru R3, R3, R3, R3, R3, R3, R3, R3, R3, R3, Store VOnew in VO. Add a new column to represent the new variant VOnew in T 0. Ask experts to confirm whether changing the data type of attribute Ai is needed or not, where 1 6 i 6 k, and k is the number of attributes. Step 4.1: If no data type needs to be changed, Suggest Recommendation I. Add a new value of Ai of the VOnew. Else, ask experts to confirm whether adding a new attribute is needed or not. Step 4.2: If no new attribute needs to be added, Sug- gest Recommendation II to modify the data type of Ai. Ask experts to acquire the mapping function of values between original and new data types. Add a column in T0 to represent the original object with new mapping values. Else, Suggest Recommendation III. Ask experts to acquire the values of the new attri- bute Anew. Add a new row in T0 to represent the new variant Anew. Store Anew in AN. Add a column in T0 to represent the original object with new attribute value if needed. Step 5: Return VO, AN, T0. le 6 e partial inference logs of Blaster le # A1 A2 A3 4 17 False Windows Update fl 2 100 False Windows Update fl 4 8 False Windows Update fl 2 119 False Windows Update fl 4 17 False Windows Update fl 2 100 False Windows Update fl 4 11 False Windows Update fl 2 76 False Windows Update fl 4 66 False Windows Update fl 2 100 False Windows Update fl Besides interaction with domain experts, the computa- tional cost of the algorithm is dependent on Step 2. The size of collecting inference log and minimal support threshold setting will affect the computational cost. For different types of inference logs, different learning algorithms can be selected to learn and discover the candidate behaviors of variants. Example 2. The variant learning example of a Blaster worm In this example, let’s assume that the fired sequence of some embedded rules of Blaster worms with marginally acceptable CF values is given in Table 6. Let’s assume minimal support is set to 30%; the large itemsets L including L1 = {(A2 = False); (A5 = X)} and L2 = {(A2 = False, A5 = X)} will be provided to experts for further recommendation; i.e., L will be used to generate the new variants acquisition table T0 according to the recom- mendations suggested by VODKA. VODKA: Does the attribute-value pair (A2 = False) belong to any new variant object? EXPERT: Yes. /*It means that a new variant contains the selected attribute-value pair (A2 = False). Otherwise, the large itemset is discarded and another large itemset is chosen to be examined. */ VODKA: What is the name of the new variant object? EXPERT: VOnew. A new column will be added in T0 to represent the variant, VOnew, separated from the original object. VODKA: Is the data type of A2 required to be changed? EXPERT: No. /* It means the data type does not need to change after adding new variant (Recommenda- tion I). Otherwise, VODKA will ask the following ques- tions. */ The row representing A2 with new attribute value and one column representing the variant object will be created in T0. VODKA: Is any new attribute required to be added? A4 A5 Object CF ood – X Blaster 0.3 ood – {135} Blaster 0.43 ood – X Blaster 0.3 ood – {4444} Blaster 0.43 ood – X Blaster 0.3 ood – {135} Blaster 0.43 ood – X Blaster 0.3 ood – {4444} Blaster 0.43 ood – X Blaster 0.3 ood – {135} Blaster 0.43 2442 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 EXPERT: No. /* It means no need to add new attribute and Recommendation II is then suggested. Otherwise, VODKA will suggest Recommendation III. */ Recommendation II. VODKA: What is the new name and new value set of the attribute A2? EXPERT: NDTnew, VDTnew. / * For each old value in A2, VODKA will ask experts to define the mapping between old and new value sets. */ The row representing A2 with new mapping values and two columns representing the original and variant objects will be created in T0. Recommendation III VODKA: What is the name and value set of the new attribute-value pair? EXPERT: Anew, AVnew. / * VODKA will ask experts to provide a set of values (AVnew) of the new attribute Anew. */ A new row representing the useful attribute namely Anew with a set of value (AVnew) and two columns representing original object and new variant will be added in T0. If all large itemsets are confirmed, the new variant acquisition table T0 can be used to generate embedded rules of discov- ered variants in the knowledge polishing stage. 3.3. Knowledge polishing using extended EMCUD Based upon the new variants acquisition table, we pro- pose extended EMCUD (E-EMCUD) algorithm as shown in Algorithm 4 to generate new embedded rules and adjust the original embedded rule base. Therefore, we can grace- fully update the embedded rule base using the small new variants acquisition table instead of using the whole large main acquisition table. Algorithm 4. The E-EMCUD algorithm Input: The original main acquisition table T, rule base RB, the new variants acquisition table T0, variant objects VO, and new attributes AN, and THCF. Output: The adjusted RB including rules with embedded meaning of new variants. Step 1: Build the corresponding AOT of T0. Step 2: Generate the embedded rules with embedded meaning of T0. Step 3: Generate the embedded rules with CF values including the original rules with strong CF and embed- ded rules with marginally acceptable CF. Step 4: Merge the new variants acquisition table T0 into main acquisition table T. Step 4.1: Append VO and AN as new columns and rows in T, respectively. Step 4.2: Ask experts to fill the values of the modified attributes of other objects in T if necessary. Step 4.3: Ask experts to examine the values of the new attributes of other objects in T if necessary. Step 5: Reset the new variants acquisition table T0. Based upon the AOT table generated in Step 1, the embedded rules generated in Step 2 will be classified into original and embedded rules in Step 3. Then, the new vari- ants acquisition table will be merged with the main acqui- sition table in Step 4. Finally, the new variants acquisition table will be reset in Step 5 to learn other variants. 3.4. The analysis of VODKA The cost of running VODKA can be divided into two categories: computational cost and interaction cost. Assume there are k attributes to classify m objects in the original main acquisition table, where the grid size is k*m. For simplifying our discussion, we use ERk,m to rep- resent the total number of embedded rules in the classifica- tion KBS, where ERk,m < m *2k. During the log collection stage, assume n instances are matched by the classification KBS. For each instance, it has Pe probability to be classified by weak embedded rules; hence the size of interesting inference log database in VODKA is n*Pe. During the knowledge learning stage, the computational cost is dominated by the learning algorithm we selected. In this paper, Apriori algorithm is used to learn the candi- dates of variant worms. Hence, the computational cost is O(Aproori). For example, if the size of database has n transactions (each transaction has k attributes) and the maximal length of large itemsets is len, then the time com- plexity of traditional Apriori algorithm is O(n*k*len). Assume L large itemsets are discovered and used to notify experts to determine the existence of the variants. For each embedded rule, assume it has Pv probability to evolve a variant; hence P�v ERk;m, denoted V, variants might be discovered. Therefore, the order of interaction with experts is V, where V < L. During the knowledge polishing stage, the E-EMCUD integrates new acquisition table into original acquisition table. The computational cost of E-EMCUD for generat- ing rules is dependent on the size of new acquisition table (GRID), denoted O(GRID). For example, using our E- EMCUD to generate embedded rules, it costs 0.05 ms �0.15 ms to generate one rule. Fig. 2 shows that the time for generating rules using different grid sizes. The computa- tional time is approximately linear growing when setting a fixed attribute number with different numbers of objects in Fig. 2a, and the growth rate is exponential when setting a fixed object number with different numbers of attributes in Fig. 2b. (a) Various number of objects (b) Various number of attributes 2000 1500 1000 500 0 2000 2500 3000 1500 1000 500 0 5 10 15 20 Objects T im e (m s) T im e (m s) Attributes = 10 Objects = 10 Attributes 25 30 5 7 9 10 11 12 Fig. 2. The time of generating rules using different grid size. S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 2443 In short conclusion, the cost of VODKA consists two parts: computational cost and interaction cost. The size of interesting inference log database: n*Pe. � Computational cost: O(Apriori) + O(GRID). � Interaction cost: V. 4. VODKA implementation and case studies Two case studies including computer worms and e- learning domain are used to evaluate the performance of VODKA. 4.1. VODKA implementation VODKA is implemented by DRAMA (Lin et al., 2003), a new object-oriented rule based system platform imple- mented using pure Java language, to refine the embedded rule base by observing the inference behaviors of weak embedded rules. It includes DRAMA server, console, knowledge extractor, and rule editor. Also, it provides an application programming interface (API) to access DRAMA server in DRAMA integrated systems. There are four basic relations between knowledge concepts defined in DRAMA: Reference, Extension-of, Trigger and Acquire. The Reference relation represents the associ- ation of two different knowledge classes (KCs) if the KCs have common piece of knowledge, which is useful for using original knowledge to construct new knowledge. Exten- sion-of relation is used to extend or modify the KC con- structed by other people, which is useful for knowledge sharing and exchanging. The Trigger and Acquire relations are used to represent the interaction of different KCs. The log collecting stage is encoded by four meta-rules described in Section 3.1 in DRAMA; the knowledge learning stage and E-EMCUD are implemented using the JSP to make a communication channel using the API provided by DRAMA. 4.2. The case study of computer worms With the rapid development of network technology, the network security becomes one of the most important issues today. To prevent network environment from intrusions, lots of researches and different systems are proposed to detect, filter, or prevent intrusions properly. Recently, the computer worms become more complicated owing to com- bining several signatures of previous known worms. The number of computer worms has grown dramatically to influence the wide computer networks due to the property of easily modifying the source codes of original computer worms to create new variants for escaping the detection of related systems, e.g., Symantec Norton (Symantec, 2005), Network Viruswall (Trend Micro, 2005), etc. They are very difficult for experts to get and analyze the signa- tures because they have incredibly sophisticated character- istics (Kienzle & Elder, 2003; Moore et al., 2003; Weaver et al., 2003). Some researchers have proposed machine learning, data mining, and clustering approaches to dis- cover and learn outliers or abnormalities (the new types of objects) (Lin, Tseng, & Lin, 2002). Generally speaking, a computer worm usually self-prop- agates via a network to acutely generate a crisis in our sys- tems through the following four stages: target selection, exploitation, infection, and propagation (Weaver et al., 2003). At the Target Selection Stage, a worm performs reconnaissance and simply probes potential victims to see if it’s running a service on a particular port. If the service is running, the worm goes to exploitation stage, in which it compromises the target by exploiting a particular vulner- ability and published exploits. If it succeeds, the worm goes to infection stage, in which it sets up on the newly infected machine. Finally, in propagation stage, the worm starts to spread by choosing new targets. And another victim will enter the next four stages cycle. In our worm detection prototype system, the knowledge of computer worms can be divided into several KCs, including the service provided by the host may be infected by certain worms and then produced some symptoms in host or network. The related attributes of various com- puter worms can be collected by some probe tools and used to evaluate the ability of VODKA, which deployed in the prototype system. In order to evaluate VODKA system, an experimental environment shown in Fig. 3 for detecting various com- puter worms is built. In this environment, the victim could receive both the normal traffic and the attacking traffic Fig. 3. The experimental environment for detecting computer worms. 2444 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 (various worm behaviors). All received traffic can be trea- ted as normal or attacking behavior, which can be trans- formed as attribute-value pairs. The network traffic collected from Internet is assumed as normal traffic since most attacking behaviors with significant signatures will be filtered by a firewall. The attacking traffic generator is designed to randomly generate various worms attacking traffic to infect the victim. Besides the attacking traffic, some signatures, e.g., the system status, host vulnerability information, and large e-mailing behavior, of the victim infected by worms can be also collected. The probe, such as Nessus (Tenable, 2005), is also used to automatically collect these worm related attributes (symptoms). The clas- sification rules of 15 kinds of worm families, including ori- ginal worms and some variant worms, are extracted using EMCUD and then these worm classification rules are stored into the worm KB. Then, these attributes are used to trigger the corresponding classification rules in the worm KB. If variant worms occurred frequently in a period of time, some candidate worm variants may be discovered by VODKA. Finally, the corresponding embedded rules of variant worms confirmed by experts will be generated to update the worm KB. To evaluate the effectiveness of VODKA, we generated 20 kinds of test samples including the behaviors of 15 kinds of original worm families and five kinds of new worm fam- ilies to randomly attack the victims. The experimental result shows that VODKA can detect 100% of original worms since the classification rules are stored in the worm KB. For detecting the occurrence of variant worms, VODKA can learn the 85% of variant worms. However, VODKA can detect 80% of new worm families in our experiments because the significant difference of new family can not be discovered easily. Thus, deploying the most complementary configurations in a collaborative framework could be effi- cient in discovering the new created worms. The following example shows that the variant objects in this domain can be discovered by VODKA, where THCNT is set to 4, THCF is set to 0.7, and the minimal support is set to 30%. Recommendation I. Elicitation of new variants by adding a new attribute value of a minor attribute. Nimda worm, a famous Email flooding worm, can be propagated to victims through the attached files in email. By monitoring the attached filename in email, we can dis- cover the large itemset L = (A4 = readme.exe) shown in Table 7 according to the embedded rule R1,1 in Table 4. Based upon the large itemsets, VODKA will ask the fol- lowing questions. VODKA: Does the attribute-value pair (A4 = read- me.exe) belong to any new variant object? EXPERT: Yes. VODKA: What is the name of the new variant object? EXPERT: Nimda.B. VODKA: Is the data type of A4 required to be changed? EXPERT: No. Consequently, the new variant acquisition table of Nim- da.B shown in Table 8 will be generated after interviewing the experts in this iteration. Hence, an original rule ‘‘IF (DoS type = Email flood) AND (Email attached file = (readme.exe)) THEN Nim- da.B, CF = 0.8” and an embedded rule ‘‘IF (DoS type = Email flood) AND : (Email attached file = (readme.exe)) THEN Nimda.B, CF = 0.5” of the Nimda.B will be gener- ated using E-EMCUD based upon the Nimda.B acquisi- tion table. Recommendation II. Elicitation of new variants by chang- ing the data types of attributes. In the priori generation of CodeRed worm, generating numerous threads to attack the victim through launching TCP flooding is one of the famous characteristics. Hence, it is useful to detect the CodeRed by analyzing the gener- ated anomaly threads in the protected system. For the par- tial detection rules of CodeRed shown in Table 9, the following shows how the values in Boolean data type will be logged as the integer value of the attribute A1 instead of true/false value. Using the above inference log, the large itemset L = (A1 = 600) can be discovered based upon the embed- ded rule R2,2 since the minimal support 30% is satisfied. Then, VODKA will ask the following questions. VODKA: Does the attribute-value pair (A1 = 600) belong to any new variant object? EXPERT: Yes. VODKA: What is the name of the new variant object? Table 7 The partial inference logs of Nimda Rule # A1 A2 A3 A4 A5 Object CF R1,1 6 True Email flood readme.exe {137} Nimda 0.4 R1,1 100 False Email flood sample1.exe {25} Nimda 0.4 R1,1 17 True Email flood readme.exe {137} Nimda 0.4 R1,1 14 False Email flood readme.exe {25} Nimda 0.4 R1,1 4 False Email flood readme.exe {445} Nimda 0.4 R1,1 19 False Email flood readme.exe {80} Nimda 0.4 R1,1 44 False Email flood hash.exe {25} Nimda 0.4 R1,1 38 True Email flood readme.exe {138} Nimda 0.4 R1,1 100 False Email flood inter.exe {25} Nimda 0.4 R1,1 28 False Email flood readme.exe {25} Nimda 0.4 Table 10 The new variant acquisition table of CodeRed.II Attributes Objects CodeRed CodeRed.II Threads 100 600 System reboot True True DoS type TCP flood TCP flood Email attached file X X TCP port X X S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 2445 EXPERT: CodeRed.II. VODIKA: Is the data type of A1 required to be changed? EXPERT: Yes. VODKA: Is any new attribute required to be added? EXPERT: No. VODKA: Can the Single Value data type be used to change the original Boolean data type of A1? EXPERT: Yes. VODKA: What is the new name and new value set of the attribute A1? Expert: Threads, (100, infinite). VODKA: What is the new value of the original True value of the attribute A1 in CodeRed? EXPERT: 100. Therefore, the new variant acquisition table of CodeRe- d.II shown in Table 10 will be generated. Consequently, an original rule ‘‘IF (Threads = 600) AND (System reboot = True) AND (DoS type = TCP Table 8 The new variant acquisition table of Nimda.B Attributes Object Nimda.B Threads X System reboot X DoS type Email flood Email attached file {readme.exe} TCP port X Table 9 The partial inference logs of CodeRed Rule # A1 A2 A3 R2,2 150 True TCP flood R2,2 600 True TCP flood R2,1 100 False TCP flood R2,2 600 True TCP flood R2,2 150 True TCP flood R2,1 100 False TCP flood R2,2 600 True TCP flood R2,2 600 True TCP flood R2,2 600 True TCP flood R2,2 300 True TCP flood flood) THEN CodeRed.II, CF=0.9” and an embedded rule ‘‘IF : (Threads = 600) AND (System reboot = True) AND (DoS type = TCP flood) THEN CodeRed.II, CF = 0.3” will be generated to classify the CodeRed.II. Recommendation III. Elicitation of new object by adding new attributes. As mentioned in Example 2, we can obtain the large itemsets L = {(A2 = False); (A5 = X); (A2 = False, A5 = X)}, which will be used to elicit the embedded rules of the new variant. The symbol ‘‘X” means no attribute value of A5 is logged, similar to ‘‘Do not care” attribute, and (A2 = False, A5 = X) will also be pruned too. There- fore, VODKA system will ask the following questions. VODKA: Does the attribute-value pair (A2 = False) belong to the new variant object? EXPERT: Yes. VODKA: What is the name of the new variant object? A4 A5 Object CF – {80} CodeRed 0.4 – {80} CodeRed 0.4 – {80} CodeRed 0.6 – {80} CodeRed 0.4 – {80} CodeRed 0.4 – {80} CodeRed 0.6 – {80} CodeRed 0.4 – {80} CodeRed 0.4 – {80} CodeRed 0.4 – {80} CodeRed 0.4 2446 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 EXPERT: Blaster.B. VODKA: Is the data type of A2 required to be changed? EXPERT: Yes. VODKA: Is any new attribute required to be added? EXPERT: Yes. VODKA: What is the name and value set of the new attribute? EXPERT: UDP port, (0, 65535). VODKA: What is the value of the UDP port attribute in Blaster.B and Blaster? EXPERT: 69, X (means Do not care). Hence, the Blaster.B acquisition table shown in Table 11 is generated. Consequently, a new original rule ‘‘IF (System reboot = Flase) AND (DoS type = Windows update flood) AND (UDP port = {69}) THEN Blaster.B, CF = 0.8” and an embedded rule ‘‘IF (System reboot = Flase) AND (DoS Table 11 The new variant acquisition table of Blaster.B Attributes Objects Blaster Blaster.B 100-thread (A1) X X System reboot (A2) True False DoS type (A3) Windows Update flood Windows Update flood Email attached file (A4) X X TCP port (A5) {135; 4444} {135; 4444} UDP port (A6) X 69 Table 12 The adjusted main acquisition table of simple computer worms Attributes Objects Nimda Nimda.B CodeRed CodeR Threads X X 100 600 System reboot X X True True DoS type Email flood Email flood TCP flood TCP fl Email attached file {samle.exe; puta!!scr} {readme, exe} X X TCP port X X X X UDP port X X X X Table 13 AOT table of simple computer worms Attributes Objects Nimda Nimda.B CodeRed Co Threads X X 2 2 System reboot X X 1 1 DoS type D D 1 1 Email attached file 1 1 X X TCP port X X X X UDP port X X X X type = Windows update flood) AND : (UDP port = {69}) THEN Blaster.B, CF = 0.5” of new variant Blaster.B will be generated to classify Blaster.B. As shown in Table 12, four variants (Nimda.B, CodeR- ed.II, Blaster.B, and Welchia.II) have been successfully sin- gled out using VODKA after several iterations. Table 13 shows the AOT table after interacting with domain experts using E-EMCUD, and Table 14 shows the new embedded rule base of the discovered variants and original worms. If more real instances can be used, the embedded rule base will evolve and become more precise for classifying the computer worms. 4.3. The case study of e-learning With the vigorous development of the Internet, e-learn- ing systems including online learning, employee training courses, and e-book, have become more and more popular over the world in the past ten years (Beishuizen & Stout- jesdijk, 1999; Hwang, 1999; Yoshikawa, Shintani, & Ohba, 2000). As we know, if the same teaching materials (the knowledge of teachers) are provided to all e-learners based on the predefined strategies or the predefined learning maps, the learning efficiency will be diminished. Therefore, teachers want to apply appropriate teaching strategy to provide personalized learning content and learning sequence for e-learners to improve their learning efficiency. Thus, adaptive learning environments (Shang, Shi, & Chen, 2001; Sheremetov & Arenas, 2002; Triantafllou, Poportsis, & Demetriadis, 2003; Tsai & Tseng, 2002) have been proposed to offer different teaching materials for dif- ed.II Blaster Blaster.B Welchia Welchia.II X X X X True False True True ood Windows Update flood Windows Update flood ICMP flood ICMP flood X X X X {135; 4444} {135; 4444} {80;135} {80; 135; 445; 3127} X 69 X X deRed.II Blaster Blaster.B Welchia Welchia.II X X X X 2 2 2 2 2 2 1 1 X X X X 1 1 2 2 X 1 X X Table 14 The rules generated from Tables 12 and 13 R1,0: IF (DoS type = Email flood) AND (Email attached file = (sample.exe; puta!!scr)) THEN Nimda, CF = 0.8 R1,1: IF (DoS type = Email flood) AND (Email attached file = : (sample.exe; puta!!scr)) THEN Nimda, CF = 0.4 R2,0: IF (Threads = 100) AND (System reboot = True) AND (DoS type = TCP flood) THEN CodeRed, CF = 0.8 R2,1: IF (Threads = 100) AND : (System reboot = True) AND (DoS type = TCP flood) THEN CodeRed, CF = 0.6 R2,2: IF : (Threads = 100) AND (System reboot = True) AND (DoS type = TCP flood) THEN CodeRed, CF = 0.4 R2,3: IF (Threads = 100) AND : (System reboot = True) AND : (DoS type = TCP flood) THEN CodeRed, CF = 0.4 R3,0: IF (System reboot = True) AND (DoS type = Windows update flood) AND (TCP port = {135; 4444}) THEN Blaster, CF = 0.7 R3,1: IF (System reboot = True) AND (DoS type = Windows update flood) AND : (TCP port = {135; 4444}) THEN Blaster, CF = 0.57 R3,2: IF (System reboot = False) AND (DoS type = Windows update flood) AND (TCP port = {135; 4444}) THEN Blaster, CF = 0.43 R3,3: IF (System reboot = True) AND : (DoS type = Windows update flood) AND (TCP port = {135; 4444}) THEN Blaster, CF = 0.43 R3,2: IF (System reboot = False) AND (DoS type = Windows update flood) AND : (TCP port = {135; 4444}) THEN Blaster, CF = 0.3 R4,0: IF (System reboot = True) AND (DoS type = ICMP flood) AND (TCP port = {80; 135}) THEN Welchia, CF = 0.8 R4,1: IF (System reboot = True) AND : (DoS type = ICMP flood) AND (TCP port = {80; 135}) THEN Welchia, CF = 0.67 R4,2: IF (System reboot = True) AND (DoS type = ICMP flood) AND : (TCP port = {80; 135}) THEN Welchia, CF = 0.53 R4,3: IF (System reboot = True) AND : (DoS type = ICMP flood) AND : (TCP port = {80; 135}) THEN Welchia, CF = 0.4 R5,0: IF (DoS type = Email flood) AND (Email attached file = readme.exe) THEN Nimda.B, CF = 0.8 R5,1 IF (DoS type = Email flood) AND : (Email attached file= readme.exe) THEN Nimda.B, CF = 0.5 R6,0: IF : (Threads = 600) AND (System reboot = True) AND (DoS type = TCP flood) THEN CodeRed II, CF = 0.9 R6,1: IF (Threads = 600) AND (System reboot = True) AND (DoS type = TCP flood) THEN CodeRed II, CF = 0.3 R7,0: IF (System reboot = False) AND (DoS type = Windows update flood) AND (UDP port = {69}) THEN Blaster.B, CF = 0.8 R7,1: IF (System reboot = False) AND (DoS type = Windows update flood) AND : (UDP port = {69}) THEN Blaster.B, CF = 0.5 R8,0: IF (System reboot = True) AND (DoS type = ICMP flood) AND (TCP port = {80; 135; 445; 3127}) THEN Welchia II, CF = 0.8 R8,1: IF (System reboot = True) AND (DoS type = ICMP flood) AND : (TCP port = {80; 135; 445; 3127}) THEN Welchia II, CF = 0.5 Table 15 The learning sequence of students Student ID Learning sequence 1 hB, C, A, D, E, F, G, H, I, Ji 2 hA, B, H, D, E, F, C, G, I, Ji 3 hA, D, F, G, H, B, C, I, Ji 4 hA, B, D, E, C, F, G, Hi 5 hA, C, J, F, B, H, D, E, I, Gi 6 hB, H, F, D, E, A, G, C, Ii 7 hA, J, E, H, B, C, I, D, Gi 8 hB, C, G, E, A, H, D, I, J, Fi 9 hC, E, G, F, J, B, H, A, Di 10 hB, C, A, J, D, E, G, H, F, Ii S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 2447 ferent learners in accordance with their aptitudes and eval- uation results. After learners learned the teaching materials through the adaptive learning environment, the teachers can further analyze the historical learning records and then refine or reorganize the teaching materials and tests if needed. Therefore, more and more attention has been paid to the research of personalized instruction in computer education environment. However, it is difficult to monitor the change of learner’s behaviors for teachers quickly. As we know, the quiz for learners is useful to evaluate their learning achievement. For example, teachers should provide easier learning content or learning sequence for the learners with lower learning achievement. Hence, VODKA provides a good idea to assist teachers in observ- ing the occurrence of variant learning behaviors of e-learn- ers through a sequence of grades of online quiz. In this case, the objects to be classified are defined as learning behaviors of learners, where each behavior con- sists of profiles, learning sequence, and a quiz grade of the learner, where learning sequence can be used to gener- ate approximate teaching materials for matching the lear- ner’s needs. The learners can be firstly clustered into several groups according to the similarity of the learning behaviors, and teachers can provide appropriate learning content for each group in advance. However, e-learners might change their learning sequences due to the different learning situation, learning equipments (desktop, PDA, etc.), course content (text, video, etc.), and learning time (day or night). This causes the evolution of learning behav- iors of e-learners and results in various learning achievements. Assume that the on-line testing system is implemented in an intelligent e-learning system. VODKA can monitor the variant learning behaviors to evaluate the learning perfor- mance of each e-learner, where each learner has different learning sequences. The teachers will be then notified to generate a suitable learning sequence and apply these mate- rials on those e-learners with variant learning behaviors. Here, each learning sequence deviated from one of prede- fined learning sequences will be treated as a variant learn- ing behavior. In e-learning, it is important for e-learners to gain a good grade after learning some materials with a specific learning sequence. Hence, the grade of quiz is trea- ted as a CF value for collecting these good variant learning sequences. If many learners gained grades higher than a threshold with similar or same learning sequence (high fre- quency), then some good variant learning sequences will be discovered to notify teachers to determine these new learn- ing sequences. Therefore, the log is collected as the pair of hLSi, CFii, where LSi is the learning sequence of the e-lear- ner i; and the CFi is the grade of this learner. Example 3 illustrates the concept of e-learning using VODKA. Table 16 The maximal frequent learning patterns of good students Large itemset Maximal frequent learning patterns L2 A ? F A ? H A ? J B ? H C ? D C ? F C ? H E ? F F ? G G ? H L3 A ? D ? G B ? C ? G L4 B ? D ? E ? G 2448 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 Example 3. The concept of e-learning using VODKA. To simplify our discussion, we assume VODKA collects several good learning behaviors of e-learners and their grades of quiz are larger than a threshold. For the learning sequence log shown in Table 15, LS1 = hB, C, A, D, E, F, G, H, I, Ji denotes that Learner 1 studies the learning content B first and then studies the learning contents C, A, D, E, F, G, H, I, J sequentially. In this example, the sequential pattern mining algorithm instead of the original Apriori algorithm is applied (Aggrawal & Srikant, 1995; Srikant & Aggrawal, 1996). Therefore, we use the Modified GSP algorithm to discover the maximal frequent learning patterns as shown in Table 16. The details can be found in (Su et al., 2006). For example, in L4, we have discovered that one candidate of new learning sequence of good e-learners is to learn B course content first and then to study the learning contents D, E, G sequentially. Hence, these candidates of various learning sequences will be suggested by VODKA for teachers to generate new variant learning sequences. In this case study, we illustrate that VODKA can collect all interesting learning behaviors (learning sequences) of e- learners whose testing grade from online quiz system is good; hence, then the maximal frequent learning sequence, a part of the whole learning sequence, will be used to recommend teachers to adapt their course for variant learning behaviors if necessary. 5. Conclusion In this paper, we proposed VODKA methodology to iteratively monitor the frequent inference behaviors of weak embedded rules with margin values of CF to assist human experts in discovering the new variant objects and singling them out of original objects. The frequent ignored attribute-value pairs of minor attributes can be learned by applying the Apriori algorithm and be treated as the new characteristics of variant objects. The new variants acquisi- tion algorithm is proposed to interact with experts to acquire the relationships between new variant objects and object attributes. Three recommendations, including add- ing a new attribute value of an attribute, changing the data type of an attribute, or adding a new attribute, are pro- posed to help experts confirm the new variants according to the frequently ignored attribute-value pairs. Addition- ally, we proposed E-EMCUD to integrate the new variant acquisition table into the main acquisition table for extracting the embedded rules of new variants. A computer worm detection prototype based upon DRAMA has been implemented and deployed in an experimental environ- ment. This environment includes a firewall to filter computer worm traffic from Internet (normal traffic) and an attacking traffic generator to randomly generate various worms to infect a victim. By this means, we evaluate the performance of VODKA. The results show that new worm variants can be singled out of the corresponding extended worm object classes after observing the occurrence of worm instances in the inference logs. Then, the detection rules for computer worms can be customized. Also, an e-learning case is shown that VODKA can help teachers quickly and easily single out the variant learning behaviors of e-learners. Then teachers can be notified to prepare new learning con- tent and learning sequence for e-learners with similar vari- ant behaviors. Therefore, these kinds of learners could easily understand the new learning materials. VODKA method can enhance the classification ability for new objects of a classification KBS since the new evolving objects can be incrementally learned and discovered by col- lecting the sufficient information. We are going to extend VODKA into a collaborative framework to help experts discover more evolving objects. Acknowledgement This work was partially supported by National Science Council of the Republic of China under Grant Nos. NSC93-2752-E-009-006-PAE, NSC95-2752-E-009-015-PAE and NSC96-2752-E-009-006-PAE. References Aggrawal, R., Srikant, R. (1995). Mining sequential patterns. In Proceed- ings of 11th international conference on data engineering (pp. 3–14). Agrawal, R., Imielinksi, T., Swami, A. (1993). Mining association rules between sets of items in large database. In Proceedings of the ACM SIGMOD conference (pp. 207–216). Alani, H., & Kim, S. (2003). Automatic ontology-based knowledge extraction from web documents. IEEE Intelligent Systems, 18(1), 14–21. Beishuizen, J. J., & Stoutjesdijk, E. T. (1999). Study strategies in a computer assisted study environment. International Journal of Learn- ing and Instruction, 9, 281–301. Blythe, J., Kim, J., Ramachandran, S., Gil, Y. (2001). An integrated environment for knowledge acquisition. In Proceedings of the interna- tional conference on intelligent user interfaces (pp. 13–20). Boegl, K. (1997). Design and implementation of a web-based knowledge acquisition toolkit for medical expert consultation systems. Doctorial thesis, Technical University of Vienna, Austria. S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 2449 Boose, J. H. (1984). Personal construct theory and the transfer of human expertise. In Proceedings of AAAI-84 conference, California (pp. 27– 33). Boose, J. H. (1985). A knowledge acquisition program for expert systems based on personal construct psychology. International Journal of Man- Machine Studies, 23(5), 495–525. Boose, J. H., Bradshaw, J. M. (1986). NeoETS: Capturing expert system knowledge in hierarchical rating grids. In IEEE expert system in government symposium. Boose, J. H., & Bradshaw, J. M. (1987). Expertise transfer and complex problems: Using AQUINAS as a knowledge-acquisition workbench for knowledge-based systems. International Journal of Man-Machine Studies, 26(1), 3–28. Cairo, O. (1998). KAMET: A comprehensive methodology for knowledge acquisition from multiple knowledge sources. Expert Systems with Applications, 14(1), 1–16. Caragea, D., Silvescu, A., Pathak, J., Bao, J., Andorf, C., Yan, C., et al. (2005). Knowledge acquisition from autonomous, distributed, seman- tically heterogeneous data sources. In Proceedings of the annual meeting of the international society for computational biology (ISMB 2005), Poster Program, Detroit, Michigan. Caragea, D., Silvescu, A., & Honavar, V. (2001). Towards a theoretical framework for analysis and synthesis of agents that learn from distributed dynamic data sources. Emerging neural architectures based on neuroscience. Springer-Verlag (pp. 547–559). Castro-Schez, J. J., Jennings, N. R., Luo, X. D., & Shadbolt, N. R. (2004). Acquiring domain knowledge for negotiating agents: A case of study. International Journal of Human-Computer Studies, 61(1), 3–31. Crowther, P., Hartnett, J. (1996). Using repertory grids for knowledge acquisition for spatial expert system. In Proceedings of Australia and New Zealand conference on intelligent information systems, Adelaide, SA, Australia, November 18–20 (pp. 14–17). Dixit, A. K., & Pindyck, R. S. (1994). Investment under uncertainty. Princeton University Press. Eriksson, H. (2003). Using JessTab to integrate Protege and Jess. IEEE Intelligent Systems, 18(2), 43–50. Fensel, D., & Angele, J. (1988). The knowledge acquisition and represen- tation language, KARL. IEEE Transaction on Knowledge and Data Engineering, 10(4), 527–550. Frank, G., & Farquhar, A. (1999). Building a large knowledge base from a structured source (KA and ontology). IEEE Intelligent Systems, 14(1), 47–54. Gaines, B. R. (1987). An overview of knowledge-acquisition and transfer. International Journal of Man-Machine Studies, 26, 453–472. Ginsberg, A., Weiss, S. M., & Politakis, P. (1988). Automatic knowledge base refinement for classification systems. Artificial Intelligence, 35(2), 197–226. Hong, T. P., & Tseng, S. S. (1997). A generalized version space learning algorithm for noisy and uncertain data. IEEE Transaction on Knowl- edge and Data Engineering, 9(2), 336–340. Hwang, G. J. (1995). Knowledge acquisition for fuzzy expert systems. International Journal of Intelligent Systems, 10, 541–560. Hwang, G. J. (1999). A knowledge-based system as an intelligent learning advisor on computer networks. Proceedings of the IEEE Transactions on Systems Man and Cybernetics, 2, 153–158. Hwang, G. J., & Tseng, S. S. (1990). EMCUD: A knowledge acquisition method which captures embedded meanings under uncertainty. Inter- national Journal of Man-Machine Studies, 33, 431–451. Hwang, G. J., & Tseng, S. S. (1991). On building a medical diagnostic system of acute exanthema. Journal of Chinese Institute of Engineers, 14(2), 185–195. Kang, B. (1996). Multiple classification ripple down rules. Ph.D Thesis, University of New South Wales. Kelly, G. A. (1955). The psychology of personal constructs. New York: Norton. Kienzle, D. M., Elder, M. C. (2003). Recent worms: A survey and trends. In Proceedings of the WORM’03, October 27, Washington DC, USA 2003. Kim, J., Gil, Y. (2001). Knowledge analysis on process models. In Proceedings of the international joint conference on artificial intelligent, Seattle, Washington, USA. Kitamura, Y., & Mizoguchi, R. (2003). Ootology-based description of functional design knowledge and its use in a functional way server. Expert Systems with Applications, 24(2), 153–166. Kolousek, G. (1997). The system architecture of an integrated medical consultation system and its implementation based on fuzzy technology. Doctoral thesis, Technical University of Vienna, Austria. Leibbink, H. J., Witteman, C. L. M., Mayer, J. J. C. (2002). Ontology- based knowledge acquisition for knowledge systems. In Proceedings of the 14th Dutch–Belgian artificial intelligence conference (pp. 195–202). Leitich, H., Kiener, H. P., Kolarz, G., Schuh, C., Graninger, W., & Adlassnig, K. P. (2001). A prospective evaluation of the medical consultation system CADIAG-II/RHEUMA in a rheumatological outpatient clinic. Methods of Information in Medicine, 40, 213–220. Lin, S. C., Tseng, S. S., & Lin, Y. T. (2002). A new mechanism of mining network behavior. Lecture Notes in Artificial Intelligence, 2336, 218–223. Lin, Y. T., Tseng, S. S., & Tsai, C. F. (2003). Design and implementation of new object-oriented rule base management system. Journal of Expert Systems with Applications, 25(3), 369–385. Maedche, A., & Staab, S. (2001). Ontology learning for the Semantic Web. IEEE Intelligent Systems, 16(2), 72–79. Mcgraw, K. L., & Harbison-Briggs, K. (1989). Knowledge acquisition: Principles and guidelines. Prentice-Hill International Editions, 1–27. Mirkovic, J., Martin, J., Reiher, P. (2002). A taxonomy of DDoS attacks and DDoS defense mechanisms. TR. 020018, Computer Science Department, University of California, Los Angeles. Moore, D., Shannon, C., Voelker, G. M., Savage, S., 2003. Internet quarantine: Requirements for containing self-propagating code. In Proceedings of INFOCOM 2003, March 30–April 3, San Francisco, USA, 2003. Noy, N. F., & Sintek, M. (2001). Creating semantic web contents with Protege-2000. IEEE Intelligent Systems, 16(2), 60–71. Politakis, P., & Weiss, S. M. (1984). Using empirical analysis to refine expert system knowledge bases. Artificial Intelligence, 22, 673–680. Quinlan, J. R. (1986). Induction of decision trees. Machine Learning, 1, 81–106. Rosaci, D. (2005). Exploiting agent ontologies in B2C virtual market- places. Journal of Universal Computer Science, 11(6), 1011–1040. Rosaci, D., Terracina, G., & Ursino, D. (2004). A framework for abstracting data sources having heterogeneous representation formats. Data and Knowledge Engineering, 48(4), 1–38. Shang, Y., Shi, H. C., & Chen, S. S. (2001). An intelligent distributed environment for active learning. ACM Journal of Education Resources in Computing, 1(2), 4–17. Shaw, M. L. G., Gaines, B. R. (1996). Web grid: Knowledge modeling and inference through the world wide web. In Proceedings of the 10th knowledge acquisition workshop (pp. 65-1–65-14). Shaw, M. L. G., & Gaines, B. R. (1987). KITTEN: Knowledge initiation and transfer tools for experts and novices. International Journal of Man-Machine Studies, 27, 251–280. Sheremetov, L., & Arenas, A. G. (2002). EVA: Am interactive web-based collaborative learning environment. Computers & Education, 39(2), 161–182. Shortliffe, E. H., & Buchanan, B. G. (1975). A model of inexact reasoning in medicine. Mathematical Bioscience, 23, 351–379. Singh, P., Lin, T., Mueller, E. T., Lim, Grace, Perkins, T., et al. 2002. Open mind common sense: Knowledge acquisition from general public. In Proceedings of the first international conference on ontologies, databases, and applications of semantics for large scale information systems (pp. 1223–1237). Srikant, R., Aggrawal, R. (1996). Mining sequential patterns: General- izations and performance improvements. In The Fifth International Conference on Extending Database Technology, 1996. Su, J. M., Tseng, S. S., Wang, W., Weng, J. F., Yang, David J. T., & Tsai, W. N. (2006). Learning portfolio analysis and mining for SCORM 2450 S.-S. Tseng, S.-C. Lin / Expert Systems with Applications 36 (2009) 2433–2450 compliant environment. Journal of Educational Technology & Society, 9(1), 262–275. Symantec Co., 2005. Symantec Products and Services. <http:// www.symantec.com/product/>. Tenable Network SecurityTM, 2005. Nessus Open Source Vulnerability Scanner Project. <http://www.nessus.org/>. Trend Micro Co., 2005. Trend Mirco Network Viruswall. <http:// www.trendmicro.com/tw/products/network/overview.htm>. Triantafllou, E., Poportsis, A., & Demetriadis, S. (2003). The design and the formative evaluation of an adaptive educational system based on cognitive styles. Computers & Education, 41(1), 87–103. Tsai, C. J., & Tseng, S. S. (2002). Building a CAL expert system based upon two-phase knowledge acquisition. Expert Systems with Applica- tions, 22, 235–248. Tsujino, K., Dabija, V., & Nishida, S. (1992). Knowledge acquisition driven by constructive and interactive induction. Lecture Notes in Artificial Intelligence, 599, 153–170. Tsujino, K ., Takegaki, M ., Nishida, S., 1990. A knowledge acquisition system that aims at integrating inductive learning and explanation- based reasoning. In Proceedings of the first Japanese knowledge acquisition for knowledge – based systems workshop, Tokyo, Japan (pp. 175–190). Uschold, M., King, M., Moralee, S., & Zorgios, Y. (1998). The enterprise ontology. The Knowledge Engineering Review, 13(1), 31–89. Weaver, N., Paxson, V., Staniford, S., Cunningham, R., Maulik, U. (2003). A taxonomy of computer worms. In Proceedings of WORM’03, October 27, Washington DC, USA, 2003. Wielinga, B., Schreiber, A., & Breuker, J. (1992). KADS: A modeling approach to knowledge engineering. Journal of Knowledge Acquisition, 4(1), 5–53. Yoshikawa, A., Shintani, M., & Ohba, Y. (2000). Intelligent tutoring system for electric circuit exercising. IEEE Transactions on Eduction, 35, 222–225. Zacklad, M., & Fontaine, D. (1995). Systematic building of conceptual classification systems with C-KAT. International Journal of Human- Computer Studies, 44(5), 603–627. Zhang, J., Honavar, V. (2003). Learning decision tree classifiers from attribute value taxonomies and partially specified data. In Proceedings of the international conference on machine learning (ICML-03), Washington, DC. http://www.symantec.com/product/ http://www.symantec.com/product/ http://www.nessus.org/ http://www.trendmicro.com/tw/products/network/overview.htm http://www.trendmicro.com/tw/products/network/overview.htm VODKA: Variant objects discovering knowledge acquisition Introduction Related work Knowledge acquisition systems Repertory grid methodology and relevant systems Elicitation of embedded meanings Problems with conventional knowledge acquisition methods Knowledge acquisition by discovering variant objects Meta knowledge in log collecting stage New variants discovering methodology at the knowledge learning stage Knowledge polishing using extended EMCUD The analysis of VODKA VODKA implementation and case studies VODKA implementation The case study of computer worms The case study of e-learning Conclusion Acknowledgement References 
work_tmfrcdzitzaspbglkw5e3ppzcy	----	Recommender system architecture for adaptive green marketing Recommender system architecture for adaptive green marketing Ying-Lien Lee a,⇑, Fei-Hui Huang b a Department of Industrial Engineering and Management, Chaoyang University of Technology, Taichung County 413, Taiwan b Department of Marketing and Distribution Management, Oriental Institute of Technology, Taipei County 220, Taiwan a r t i c l e i n f o Keywords: Green marketing Green consumerism Recommender system Fuzzy inference system a b s t r a c t Green marketing has become an important method for companies to remain profitable and competitive as the public and governments are more concerned about environmental issues. However, most online shopping environments do not consider product greenness in their recommender systems or other shop- ping tools. This paper aims to propose the use of recommender systems to aid the green shopping process and to promote green consumerism basing upon the benefits of recommender systems and a compliance technique called foot-in-the-door (FITD). In this study, the architecture of a recommender system for green consumer electronics is proposed. Customers’ decision making process is modeled with an adaptive fuzzy inference system in which the input variables are the degrees of price, feature, and greenness and output variables are the estimated rating data. The architecture has three types of recommendation: information filtering, candidate expansion, and crowd recommendation. Ad hoc customization can be applied to tune the recommendation results. The findings are reported in two parts. The first part describes the potentials of using recommender systems in green marketing and the promotion of green consumerism; the second part describes the proposed recommender system architecture using green consumer electronics as the context. Discussion of the proposed architecture and comparison with other systems are also included in this part. The proposed architecture provides a capable platform for person- alized green marketing by offering customers shopping advices tailored to their preferences and for the promotion of green consumerism. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Recommender systems have become an important technology for electronic commerce on many fronts (Bose, 2009; Kauffman & Walden, 2001). It can filter for online shoppers the vast amount of information, saving the customers from the information over- load problem (Chen, Shang, & Kao, 2009). It can be a decision aid for customers who are challenged when they are in the market for unfamiliar products. It can be a strategic marketing platform on which online venders can personalize promotions and sales for each customer (Chen, 2008; Shih, Chiu, Hsu, & Lin, 2002). Recommender systems have been vigorously researched and developed in the fields of academia and business. Some notable examples include Apple Inc.’s Genius of iTunes that make music recommendations, University of Minnesota’s MovieLens and Netflix’s Cinematch that recommend movie titles, Amazon.com’s recommender system that generates recommendations of an assortment of products, and Outbrain.com’s blog rating widget that recommends blogs a rater might be interested in. The domain of recommender systems is not limited to the famous instances men- tioned above. Recommender systems for news, web pages, jokes, academic articles, consumer electronics, restaurants, and a pleth- ora of other subject matters, have been researched and imple- mented (Adomavicius & Tuzhilin, 2005; Iijima & Ho, 2007). However, to our knowledge, few researches have dealt with rec- ommender system of green product. Green product is increasingly important in our global village as the general public is becoming more concerned of our impact on the planet. Driven by this trend, companies have been trying to de- sign and manufacture greener products, and have been trying to promote their products and brand images by communicating their greenness to the customers via a variety of channels. Yet, eco- labeling remains one of the fundamental ways to inform the cus- tomers how green their products are and in what respect their products are green. Eco-labels, usually issued by third-party orga- nizations, are textual or graphical presentations of the environ- mental characteristics of a product, which can be found on the product itself, on the packaging, or in the manual. Examples of eco-labels include Green Seal, Energy Star, and WEEE (Waste Elec- trical and Electronic Equipment Directive). Studies have shown that public education campaign is one of the key determinants of 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.01.164 ⇑ Corresponding author. Address: No. 168, Jifong E. Rd., Wufong Township, Taichung County 413, Taiwan. Tel.: +886 4 23323000; fax: +886 4 23742327. E-mail address: yinglienlee@gmail.com (Y.-L. Lee). Expert Systems with Applications 38 (2011) 9696–9703 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.01.164 mailto:yinglienlee@gmail.com http://dx.doi.org/10.1016/j.eswa.2011.01.164 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa successful eco-labeling programs (Malcohn, Paulos, Stoeckle, & Wang, 1994). Public education campaign of eco-labeling programs can be done via methods such as media coverage, regulation, promotion, school curriculum, and so on. This research proposes using recom- mender systems, in addition to these methods, as a means to edu- cate and inform on-line customers. The justification of such proposal relies on two of the primary functions of a recommender system: information filtering and candidate expansion. When cus- tomers are confronted with a flood of products or with unfamiliar products, they may have difficulty in making a shopping decision. Based upon what they have purchased before, a recommender sys- tem can help the customers by filtering out items that are unlikely to be preferred. For example, Amazon.com’s recommender system generates a personalized list of recommended products each time a customer visits their web site. As to candidate expansion, when a customer is evaluating the decision to buy a product, a recom- mender system can ensure that other good candidates are included in the consideration set by finding related products based upon the product under consideration. Take Amazon.com’s recommender system for example again. When a customer is looking at the cat- alog page of a product, the recommender system recommends items similar to the current item. Tapping into the capabilities of information filtering and candidate expansion, recommender sys- tems can be transformed into a green product advocate informing the customers of available choices that are greener. A recom- mender system of green products can also sieve through a set of products to retrieve only the items matching an implicitly or explicitly degree of greenness designated by a customer. Such sys- tem can also find other products whose greenness and other as- pects are comparable based upon a product under consideration. The reduced effort in the decision making process may enhance the quality and users’ satisfaction of the decision (Häubl & Trifts, 2000), which in turn will make green shopping a more enjoyable experience. The adaptability of a recommender system can also contribute to the promotion of green consumerism by using a technique called foot-in-the-door (FITD) technique (Freedman & Fraser, 1966). FITD is a compliance technique in which a person is more likely to accept a larger request if this request is preceded by a smaller request. The technique is also found to be effective in com- puter-mediated communication (CMC) in addition to face-to-face or telephone communications (Guéguen, 2002). In a recommender system of green products, items with higher degree of greenness and with comparable or equal degrees of price and feature can be first recommended to a user who is reluctant to buy green prod- ucts. Appropriate feedback should be given to the user about the environmental contribution of the purchase one has made. The de- gree of greenness of the recommended items in the future can be adjusted accordingly if the users’ purchasing transactions reflect acceptance or rejection of the items. The goal of this paper is to develop a recommender system architecture for green consumer electronics. Instead of simply add- ing an additional green attribute to the conventional recommender systems, the architecture uses an adaptive behavioral agent to find the products of a certain degree of greenness according to users’ behaviors. The agent uses an adaptive fuzzy inference system to learn users’ behavior over time with a basic assumption that a bilateral relationship of either symbiosis or antibiosis exists be- tween the pairs of price vs. feature, price vs. greenness, and feature vs. greenness. The rest of this paper is organized as follows. The next section gives brief review of recommender systems and fuzzy inference systems. The proposed architecture is presented and discussed in Section 3. Conclusions and future research directions are presented in the final section. 2. Related work 2.1. Recommender system Recommender systems have a variety of forms with different functions (Manouselis & Costopoulou, 2007; Wan, Menon, & Rama- prasad, 2007). Therefore, it warrants a clear definition of the kind of recommender systems this paper is dealing with. Schein, Pope- scul, Ungar, and Pennock (2005) define recommender systems as the following: ‘‘Recommender systems suggest items of interest to users based on their explicit and implicit preferences, the pref- erences of other users, and user and item attributes’’. This defini- tion points out the fundamental parts and necessary input and output data of a recommender system. First, a recommender sys- tem needs data of preferences from single user or multiple users. The system can explicitly elicit preferences from users by asking them to rate some items, or implicitly by inferring their prefer- ences from past transactions (Resnick & Varian, 1997). Second, a recommender system requires attributes of users and items. Manouselis and Costopoulou (2007) refer to these two sets of attri- butes as ‘‘user model’’ and ‘‘domain model’’, respectively. Several representations can be used as user models, such as per user prod- uct ratings, demographic attributes, transaction histories, and so on. On the other hand, domain models can be represented as char- acteristics of products and as derived attributes such as taxono- mies, hierarchies, and ontologies. Both models may utilize the acquired user preferences to derive their own data. The core of a recommender system is the mechanism of sugges- tion generation based upon the user model and domain model. The mechanism can be formulated as follows (Adomavicius & Tuzhilin, 2005): Let C be the set of all customers and P be the set of all prod- ucts that a recommender knows of. In addition, let U(c, p) be the utility function that associates (c, p) pairs with utility values which can be ratings, profits, or some other measurements. The objective of a recommender system is to find a set of items p0 e P such that U(c, p) is maximized for a customer. The mathematical formulation is as follows: 8c 2 C; p0 ¼ arg max p2P Uðc; pÞ In the formulation, ‘‘arg max’’ means ‘‘the argument of the maximum’’. Recommender systems can be generally classified into three categories according to the mechanism of recommendation gener- ation (Adomavicius & Tuzhilin, 2005; Schein et al., 2005): (1) Con- tent-based systems recommend items that are similar to the ones a user preferred in the past. (2) Collaborative systems recommend items that other like-minded users preferred in the past. (3) Hybrid systems recommend items by combining content-based and col- laborative methods in recommendation generation. As Adomavicius and Tuzhilin (2005) point out, content-based and collaborative systems have some challenges to be dealt with. For content-based systems, the first problem is ‘‘limited content analysis’’, in which case the recommendation is limited by the fea- tures associated with the items. However, some features are hard- er to extract than others are. For example, extracting features from textual information is easier than from multimedia data. Also, items that are identical in terms of features are indistinguishable. The second problem is overspecialization, in which case the sys- tem can only recommend items that are similar to items a user liked in the past. In other words, the lack of diversity may jeopar- dize the practicality of a recommender system. The third problem is ‘‘new user problem’’, in which case a user is unable to get reli- able recommendations until a sufficient amount of transactions are present for the recommender system to learn about the users’ preferences. Y.-L. Lee, F.-H. Huang / Expert Systems with Applications 38 (2011) 9696–9703 9697 https://isiarticles.com/article/22904 
work_tnbeadapefeuvbnzlzj3nzxv5a	----	Penalty1vs0.tex Data Quality Assessment of Maintenance Reporting Procedures Manik Madhikermia, Sylvain Kublerb,∗, Jérémy Robertb, Andrea Budaa, Kary Främlinga aAalto University, School of Science P.O. Box 15400, FI-00076 Aalto, Finland bUniversity of Luxembourg, Interdisciplinary Centre for Security, Reliability & Trust 4 rue Alphonse Weicker L-2721 Luxembourg Abstract Today’s largest and fastest growing companies’ assets are no longer physical, but rather digital (software, al- gorithms. . . ). This is all the more true in the manufacturing, and particularly in the maintenance sector where quality of enterprise maintenance services are closely linked to the quality of maintenance data reporting pro- cedures. If quality of the reported data is too low, it can results in wrong decision-making and loss of money. Furthermore, various maintenance experts are involved and directly concerned about the quality of enterprises’ daily maintenance data reporting (e.g., maintenance planners, plant managers. . . ), each one having specific needs and responsibilities. To address this Multi-Criteria Decision Making (MCDM) problem, and since data quality is hardly considered in existing expert maintenance systems, this paper develops a Maintenance Reporting Quality Assessment (MRQA) dashboard that enables any company stakeholder to easily – and in real-time – assess/rank company branch offices in terms of maintenance reporting quality. From a theoretical standpoint, AHP is used to integrate various data quality dimensions as well as expert preferences. A use case describes how the pro- posed MRQA dashboard is being used by a Finnish multinational equipment manufacturer to assess and enhance reporting practices in a specific or a group of branch offices. Keywords: Data Quality, Information Quality, Multi-Criteria Decision Making, Analytic Hierarchy Process, Decision Support Systems, Maintenance. 1. Introduction Data and Information quality is one of the most competitive advantages for an organization in today’s digital age, for example, with the rapid evolution of Internet of Things, Industry 4.0, Big Data and Cloud Computing (Xu et al., 2010; Chen et al., 2014). Com- panies are trying hard to find out relevant strategies to make their products (physical or virtual) standout with respect to their competitors. Quality improvement of products, processes and services requires the collec- tion and analysis of data to solve quality-related prob- lems (Li et al., 2015; Köksal et al., 2011). Companies need to provide after-sales services such as mainte- nance and warranty services to ensure that the deliv- ered product is reliable and in full accordance with the customer requirements. Nonetheless, providing such services inevitably generate costs for businesses (Fang & Huang, 2008). As indicated by Mobley (2002), one third of all maintenance costs is wasted as the ∗Corresponding author Email addresses: manik.madhikermi@aalto.fi (Manik Madhikermi), sylvain.kubler@uni.lu (Sylvain Kubler), jeremy.robert@uni.lu (Jérémy Robert), andrea.buda@aalto.fi (Andrea Buda), kary.framling@aalto.fi (Kary Främling) result of unnecessary or improper maintenance prac- tices. More recent studies have confirmed that mainte- nance is a major cost issue, with a ratio between main- tenance costs and added-value higher than 25% in some sectors (Sophie et al., 2014). In fact, data qual- ity practices – including maintenance reports – have a considerable impact on maintenance tasks, risks and business performance since poor data quality results in losses across a number of fronts (Arputhamary & Arockiam., 2015), and reciprocally, high data qual- ity fosters enhanced business activities and decision- making. A successful maintenance program often relies on a detailed planning and intelligent decision-making support systems. This is all the more true given that planning maintenance involves managing a set of complex tasks and resources to guarantee the max- imum possible operational availability of equipment (Palma, 2010). Various stakeholders with different re- sponsibilities are involved in this management, such as (i) Maintenance planners who are responsible for scheduling planned maintenance activities; (ii) Plant managers who are responsible for cost reporting and savings; (iii) Maintenance managers who are respon- sible for the execution of planned/unplanned mainte- nance activities, and so on. All these experts have Preprint submitted to Elsevier June 28, 2016 a common goal: reducing maintenance downtime to increase productivity. In this respect, they usually make use of maintenance reports as decision support tools, which contain useful record information such as technical maintenance logs, asset location, descrip- tion of defect location codes, scheduled maintenance date, etc. It is thus of importance to develop and im- plement strategies for enhanced reporting practices, data quality control and management (Jones-Farmer et al., 2014). Nonetheless, requirements related to the data and associated quality attributes are tightly cou- pled with the stakeholder’s needs and responsibilities. For example, maintenance managers pay more atten- tion to technical log records for their daily decision- making, whereas plant managers rather use defect and asset location-related information to manage their in- ventory. All this provides irrefutable evidence of the complexity of developing a flexible, intelligent and integrated decision-making support system for data quality assessment and maintenance management; it implies to take into consideration various stakeholder roles, needs, quality dimensions, and other techni- cal and organizational aspects (Vujanovićet al., 2012; Shafiee , 2015). Given the Multi-Criteria Decision Making (MCDM) nature of the problem, this paper investigates and develops a Maintenance Reporting Quality Assessment (MRQA) tool, whose underly- ing framework relies on AHP. The primary goal of this tool is to help companies to dynamically assess quality of daily maintenance data reporting activities, while taking into account specific needs or role of the end-user (i.e., a company stakeholder). The summary of the paper is as follows: Sec- tion 2 conducts a thorough literature review of both (i) existing expert maintenance systems making use of MCDM techniques, and (ii) existing data qual- ity frameworks, against which our research is moti- vated. Section 3 provides insight into the research methodology underlying the MRQA framework/tool development. Section 4 thoroughly details the MRQA framework and underlying mathematical theory. Sec- tion 5 describes a use case that shows how the pro- posed MRQA decision-making support dashboard is being used by a Finnish multinational Original Equip- ment Manufacturer (OEM) company to assess and rank company branch offices in terms of maintenance reporting quality. Conclusions, implications, limita- tions and future research are discussed in Section 6. 2. Data quality in expert maintenance systems To understand how crucial and complex it is to properly address data quality in maintenance settings, section 2.1 discusses the key maintenance business levels, along with previous research works that have used MCDM techniques to address challenges at each of these levels. Section 2.2 discusses existing frame- works for data quality analysis and management in maintenance processes. 2.1. Expert maintenance systems Maintenance is a complex process that is usually triggered by an equipment failure or planned repair. This process requires planning, scheduling, control- ling as well as deploying maintenance resources to perform the necessary maintenance actions (Duffuaa et al., 2001). Adopting an efficient approach to or- ganize maintenance management (MM) activities is a prerequisite to its success. Several MM frameworks have been developed and applied for this purpose, one of the earliest being put forward by Pintelon & Gelders (1992) who pointed out three important busi- ness levels in the decision-making process, including the (i) Operational level: decision regarding market- ing and finance; ii) Planning & control level: deci- sions regarding resource and scheduling management, and performance reporting; iii) Managerial level: de- cisions regarding how to optimize actions and poli- cies to be performed on-site. Later on, Levrat et al. (2008) proposed a similar three business level-based MM framework, namely: • Strategic level: strategic axis are expressed in quantitative and qualitative terms, and organi- zational maintenance strategies are defined such as corrective and preventive maintenance, risk- based or condition-based maintenance, etc.; • Tactical level: maintenance actions such as scheduling and resource planning are planned; • Operational Level: actual work is carried out in addition to access performance and future equip- ment conditions. Making decisions at each of these three levels im- plies dealing with multiple, conflicting, and incom- mensurate criteria and/or objectives, as well as hu- man judgments. Research on human judgements and decision making shows that the human brain is able to consider only a limited amount of information at any one time (Simpson, 1996), which makes it unreli- able to take decisions when facing complex problems. MCDM techniques, such as AHP, TOPSIS, ELEC- TRE, PROMETHEE, Fuzzy MCDM, etc., have been proven to be of great value in supporting decision- makers at each MM level, as summarized in Table 1. At the “Strategic level”, MCDM techniques are considered for various purposes, including (i) main- tenance policy selection, (ii) tool/contractor selection, and (iii) cost estimation. Table 1 provides an “at a glance” overview of scientific papers that have made use of MCDM techniques for each of these purposes. Bevilacqua and Braglia (2000); Wang et al. (2007); 2 Table 1: MCDM Techniques Applied in Maintenance Industry AHP FAHP ANP DEA VIKOR ELECTRE PROMOTHEE MAUT S tr a te g ic L e v e l Maintenance Policy Se- lection (Bertolini and Bevilacqua, 2006; Bevilacqua and Braglia, 2000; Goossens et al., 2015; Labib et al., 1998; Pramod et al., 2007; Shyjith et al., 2008; Tan et al., 2011; Zaim et al., 2012) (Azizi et al., 2014; Ilangkumaran & Ku- manan, 2009; Hos- seini et al., 2015; Ferdousmakan et al., 2014; Wang et al., 2007; Fouladgar et al., 2012) (Kumar & Maiti, 2012; Shahin et al., 2012; Pourjavad et al., 2013; Zaim et al., 2012) Azadeh et al. (2014); Sheikhalishahi (2014) (Ilangkumaran & Kumanan, 2012; Ahmadi et al., 2010) (Zhangqiong & Guozheng , 1999; Li et al., 2007) (Emovon et al. , 2015; de et al., 2015c; Monte et al., 2015) Tools and companies Se- lection (Bertolini et al., 2004; Garcı́a- Cascales et al., 2009; Ha et al., 2008; Triantaphyllou et al., 1997) (Durán, 2011) (Ha et al., 2008) (dGonçalves et al., 2014; Gomez et al., 2011) (Kuo et al., 2012; de et al., 2015a) Cost Estimation (Chou , 2009) (Chen et al., 2005) T a c ti c a l L e v e l Maintenance prioritiza- tion (Farhan & Fwa, 2009; Moazami et al., 2011; Taghipour et al., 2011) (Ouma et al., 2015) (Wakchaure & Jha, 2011) (Liu et al., 2012; Cafiso et al., 2002; Hankach et al., 2011; Trojan & Morais, 2012b) (Monte & de Almeida-Filho, 2016; de et al., 2015b) Resource Planning (Azadeh et al., 2013) (Cavalcante et al., 2010; Almeida et al., 2013) (Garmabaki et al., 2016; de Almeida., 2001; Liu & Fran- gopol, 2006) Maintenance Scheduling (Coulter et al., 2006; Eslami et al., 2014) (Van et al., 2013) (Certa et al., 2013) (Almeida, A. T., 2012) O p e ra ti o n a l L e v e l Critical Component Iden- tification (Dehghania et al., 2012) (Cavalcante et al. , 2007, 2010) Measuring/Assessment Efficiency (Wang et al., 2010) (Muchiri et al., 2011; Vujanovićet al., 2012; Van & Pintelon, 2014) Sun (2004); Peck et al. (1998); Ozbek et al. (2010a,b); Hjalmarsson et al. (1996); Liu & Yu (2004); Rouse et al. (2002); Fallah-Fini et al. (2015); Jeon et al. (2011); Roll et al. (1989); Charnes et al. (1984) (de un Caso, 2008) (e Costa et al., 2012) Maintenance Action Se- lection (Kumar & Maiti, 2012) (Alarcón et al., 2007; Thor et al., 2013) 3 Tan et al. (2011); Fouladgar et al. (2012) developed MCDM-based maintenance policy selection frame- works taking into account maintenance cost, added- value and safety dimensions. Shahin et al. (2012) rather focused on the selection of appropriate (opti- mum) maintenance strategies, paying special attention to reliability, availability and maintainability criteria and potential interdependencies (via ANP). Gomez et al. (2011); Durán (2011) developed a similar ap- proach, considering the same criteria, but rather ap- plying ELECTRE II and FAHP respectively. Select- ing appropriate tools and/or contractors for outsourc- ing activities plays also an important role at the strate- gic level, as it affects the whole maintenance manage- ment process. In this respect, Bertolini et al. (2004) developed an AHP-based outsourcing service selec- tion model considering maintenance-related criteria. Maintenance budgeting and cost estimation are other important strategic decisions that need to be properly managed. To this end, Chou (2009) and Chen et al. (2005) develop two distinct utility-based assessment approaches, respectively relying on AHP and ELEC- TRE II, which enable decision-makers to estimate – based on historical data of similar projects – pave- ment and pipeline maintenance costs. Looking at the “Tactical level” now, MCDM tech- niques are mainly applied for maintenance work plan- ning purposes, which includes (i) task prioritization, (ii) task scheduling, and (iii) resource planning. Ta- ble 1 reports some scientific papers that have made use of MCDM techniques for each of these purposes. Cafiso et al. (2002); Farhan & Fwa (2009); Moazami et al. (2011); Ouma et al. (2015) and Babashamsi et al. (2016) have all studied prioritization of road mainte- nance with the objective to reduce the overall cost (cri- teria considered in this studies being traffic volume, road safety, pavement width. . . ). Other studies such as (Trojan & Morais, 2012a,b; Monte & de Almeida- Filho, 2016) developed MCDM-based frameworks for maintenance prioritization in the context of water sup- ply networks, looking at strategies for reducing costs and water losses. Taghipour et al. (2011) devel- oped a framework in the context of healthcare main- tenance management for medical equipment prioriti- zation, considering mission criticality, age, risk, re- call and hazard alerts as main prioritization criteria. Resource planning is also a very critical aspect to be tackled at the tactical level, as resources can be either human or non-human in nature. For example, Van et al. (2013) develop a three-stage approach of personnel rostering (i.e., human scheduling) for aircraft mainte- nance, whereas de Almeida. (2001) are seeking to op- timize spare-provisioning (i.e., non-human resource allocation). Finally, at the “Operational level”, MCDM tech- niques are often applied for (i) task efficiency assess- ment, (ii) Critical component identification, and (iii) Maintenance action selection. Table 1 shows that most of the papers implement a data envelopment analysis (DEA) for task efficiency assessment, which is a well-known tool for benchmarking in operations management. The identification of critical compo- nents is also very important to addressed to offer en- hanced predictive maintenance services (Cavalcante et al. , 2010; Dehghania et al., 2012). Along with crit- ical component identification comes the challenge of making the right decisions and actions on the field to avoid causing any disruption, delay or monetary loss. A few studies have been using MCDM techniques to select the best maintenance action(s) on-site, such as Nyström & Söderholm (2010) who are seeking to improve railway track maintenance practices, or still Alarcón et al. (2007) who apply ELECTRE for mini- mizing telecommunication network disruption during maintenance activities. Given the significant number of papers discussed above and classified in Table 1 (40 papers at the Strategic level, 19 at the Tactical level, and 20 at the Operational level), we feel it is appropriate to analyze and identify criteria that are the most commonly used at each MM level, which will therefore help us to state whether or not existing studies takes into account data quality-related criteria. The outcome of our analysis shows that the three most commonly used criteria at each MM level are (see Appendix B for a complete overview of the analysis outcome and criteria lists): • Strategic level: (i) Cost (22.7%), (ii) Resource Availability & Utilization (10.1%), Added value (7.6%); • Tactical level: Cost (21.9%), Environmental/Op- erational Conditions (8.7%) and Safety (9.4%); • Operational Level: Cost (25.8%), Resource Availability & Utilization (19.4%) and Added Value (4.8%). These results clearly show that data quality is hardly considered in the reviewed papers, whereas it can have a major impact on expert decisions, as previously dis- cussed. The only paper that integrates a data qual- ity criterion is (Van & Pintelon, 2014), where the au- thors measure the “accuracy” of maintenance report records. Given the fact that existing expert mainte- nance systems fail, or have no specific interest, to take into account various data quality dimensions, as well as in providing experts with the possibility to spec- ify – in real-time – their own preferences regarding each of these dimensions, our research aims to fulfill this gap by adapting existing data quality frameworks to the maintenance sector (the next section discussing such frameworks). From the MM’s viewpoint, our re- search primarily addresses the “Tactical” and “Oper- ational” levels with the objective to assess quality of enterprises’ daily maintenance reporting activities. 4 2.2. Data quality frameworks Although first data or information quality frame- works were introduced back in the 90′ (Krogstie et al., 1995; Wang & Strong, 1996; Jarke & Vassiliou, 1997), research efforts has recently gained momentum in an increasingly number of sectors due to the digi- talization of almost every industry, e.g. in the context of i) smart cities for open data portal quality assess- ment, e.g. in (Umbrich et al., 2015) whose metadata items are assessed in terms of quality; ii) product life- cycle management, e.g. in (Wellsandt et al., 2015) where authors separate ‘fitting’ from ‘unfitting’ in- formation from a manufacturing/design decision pro- cess perspective; iii) query processing, e.g. in (Sam- paio et al., 2015) for incorporating data quality pro- filing dimensions in the processing of queries involv- ing quality-aware query language extensions; or more recently (iv) Google’s Analytics Advocate shared his own framework “TITE” (time, interactions, trends, and events) to help marketers gain context and get ac- tionable insights from their data (Waisberg, 2015). Although some data quality frameworks are generic enough to be applied in different contexts and sectors (Krogstie et al., 1995; Kahn et al., 2002; Maurino & Batini, 2009), they often need to be tuned/adapted to each application case. It is, nonetheless, difficult to state in what respects one framework is better than another since data quality is commonly thought of as a multi-dimensional concept with varying attributed characteristics, which depend on the author’s philo- sophical viewpoint, past experience, application do- mains, and so forth (Ofner et al., 2013). In our re- search, we decided to consider the framework intro- duced by Krogstie et al. (1995) which, even if it dates back to 1995, provides a very detailed and complete overview of data quality concepts and relationships. Section 3 provides greater detail on the Krogstie’s framework, and to what extent it is instanciated to cope with maintenance reporting procedures. 3. Research Methodology: Data quality frame- work instantiation to MRQA purposes The research methodology used in this study for de- termining what data quality dimensions must be in- tegrated to our model (in light of the MRQA prob- lem) is presented in this section. To this end, sec- tion 3.1 presents the high-level concepts and relation- ships covered by the Krogstie’s framework, while sec- tion 3.2 describes both what concepts/relationships from that framework are relevant to our problem and how they are integrated based on AHP. 3.1. Krogstie framework concepts and definitions Concepts and relationships underlying the Krogstie’s data quality framework are depicted in Figure 1 and described hereinafter: Modeling Domain Knowledge Quality Model Externalization Semantic Quality Syntactic Quality Physical Quality Language Extension Language Quality Participant Knowledge Audience Interpretation Social Technical Language Quality Social Quality Perceived Semantic Quality Pragmatic Quality Pragmatic Quality Figure 1: Krogstie’s data quality framework • Physical Quality: about externalizability (i.e., the knowledge of some social actors has been externalized by the use of a conceptual model- ing language) and internalizability (i.e., the ex- ternalized model is persistent and available, thus enabling participants to make sense of it); • Syntactic Quality: correspondence between the model and the language extension of the lan- guage in which the model is written; • Semantic Quality: correspondence between the model and domain, where domain is considered as the ideal knowledge about the situation to be modeled. Krogstie’s framework contains two se- mantic goals: Validity and Completeness; • Perceived Semantic Quality: correspondence be- tween the actor interpretation of a model and his/her current knowledge of the domain. In line with the semantic quality, two goals are defined by the authors: Perceived Validity and Perceived Completeness; • Pragmatic Quality: correspondence between the model and the “Audience Interpretation” of it; • Social Quality: about people “agreement”; • Knowledge Quality: from a pure standpoint of social construction, it is difficult to talk about the quality of explicit knowledge. On the other hand, within certain areas such as mathematics, what is regarded as ‘true’ is comparatively stable, and it is inter-subjectively agreed that certain peo- ple have more valid knowledge of an area than others. The ‘quality’ of the participant knowl- edge can thus be expressed by the relationship between the audience knowledge and domain; • Language Quality: appears as means for model quality in the framework. The authors have re- grouped factors from earlier discussions on lan- 5 Table 2: Criteria and its sub-criteria description related to the data quality dimensions Criteria Sub-Criteria Description Type Believability (CB) Length of Work Description (CB1) Length of the work description related to a work order. I CB1 avg Work Log Conflict (CB2) Work description conflict among different form fields related to a same report. I CB2 var Technician Log Variation (CB3) Technical log variation among a set of reports related to a same maintenance work order. I CB3 var Completeness (CC ) Asset Location reported (CC1 ) Asset location (in the product) where maintenance was performed. I CC1 fill Description reported (CC2 ) Description of work to be done in particular maintenance work. I CC2 fill Start & Finish Date reported (CC3 ) Actual Start and Finish dates and times of work completed. I CC3 fill Target Start Date reported (CC4 ) Targeted start date of the maintenance work. I CC4 fill Target Finish Date reported (CC5 ) Targeted finish date of the maintenance work. I CC5 fill DLC Code reported (CC6 ) Actual location of the defect within product (DLC standing for “Defect Location Code”). I CC6 fill Schedule Start Date reported (CC7 ) Scheduled start date of the maintenance work. I CC7 fill Schedule Finish Date reported (CC8 ) Scheduled Finish date of the maintenance work. I CC8 fill Timeliness (CT ) This is average delay of reporting on individual site I CT avg guage quality as follows: i) Domain appropriate- ness; ii) Participant Knowledge Appropriateness; iii) Technical actor interpretation enhancement. 3.2. MCDM-based Krogstie framework instanciation Given the above definitions, and based on the Finnish OEM company’s requirements, three key con- cepts/relationships and a working assumption form the foundation of our framework. First, we assume that the Physical Quality (cf. Figure 1), and particu- larly the externalized model, is 100% persistent and available, thus enabling participants to make sense of it. A potential study assessing how persistent their implementations are compared with the initial ex- pert statements/knowledge will be achieved in future work. The OEM company then expressed require- ments regarding three of the Krogstie’s framework concepts/relationships, as highlighted in red/bold in Figure 1, namely: 1. Semantic Quality: the OEM company wants to know to which extent the service data reported by each operator (on each site) can be trusted, or more exactly can be considered as “true”, “real” and “credible”, in order to carry out the plan- ning activities. This is referred to as the “Be- lievability” criterion (CB), whose various facets of CB are formalized in the form of sub-criteria, i.e. as Believability quality indicators denoted by {CB1..CB3} (see Table 2 for more information); 2. Language Quality: the OEM company wants to know to which extent the service data reported by each operator is complete, or is of sufficient depth and breadth for the task at hand. To put it another way, this criterion, referred to as Com- pleteness (CC ), reflects the level of details re- ported by each operator with regard to each re- port/form field that needs to be entered (in ac- cordance with the company’s business logic). Similarly to CB, several Completeness quality indicators are defined, respectively denoted by {CC1 . . .CC8} (see Table 2); 3. Knowledge Quality: the OEM company wants to know to which extent the service data reported by each operator is sufficiently “up to date”, which is depending on the time difference between the maintenance work achievement and the task re- porting. This criterion, referred to as Timeliness CT , is based on the assumption that the longer the time spent to submit the report, the lesser the quality of the reporting (operator are likely to for- get details over time). As emphasized in Table 2, no sub-criterion is defined for this dimension. To ease the understanding of those data quality di- mensions and sub-dimensions, an illustration of the overall maintenance reporting quality assessment is presented in Figure 2. First, maintenance opera- tors carry out maintenance work orders/tasks on each OEM site (denoted by Site 1. . . Site z), thus generating multiple reports. Figure 2 presents a simplified view of (i) the report’s content, and (ii) the comparison pro- cess based on the criteria introduced in Table 2. We emphasize, using “smileys”, how and why a report’s content (or form field content) can positively or nega- tively impact on the maintenance reporting quality. In light of the MCDM problem – integration of various quality dimensions, expert preferences, report contents. . . – AHP has been chosen to help organiz- ing critical aspects of the problem in a manner sim- ilar to that used by the human brain in structuring the knowledge (Saaty, 1980). As highlighted in our literature review (cf. section 2), there are a number of MCDM techniques such as AHP/ANP, TOPSIS, ELECTRE, PROMETHEE, MAUT, and much more (e.g., hybrid MCDM combining these techniques to- gether or even with other theories such as fuzzy logic) (Behzadian et al., 2012; Mardani et al., 2015). There 6 SITE 1 SITE 2 . . . SITE z x x x x y xx x x y x x xy y OEM Database Maintenance Operator (Site 1) Report ID : 1D Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) ID : 1389706 Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) ID : 1389706 Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Power Controller 4v Done 28/08/2014 23/08/2014 24/08/2014 27/08/2014 Front axle 34.8YH Done 02/05/2014 07/06/2014 Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 2) Report ID : 2A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zD Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) ID : 1389706 Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) ID : 1389706 Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site z) Report ID : zA Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Fuel System 01X.2 System changed by... 28/08/2014 23/08/2014 24/08/2014 27/08/2014 Chassis has been re... 28/08/2014 23/08/2014 24/08/2014 27/08/2014 Maintenance Operators per OEM site Example when comparing operator reports between Sites 1 and z Maintenance Reporting Quality Assessment of the OME company CB1 : Length of Work Description One world (”Done”) is too short to properly describe the maintenance opration The description seems to be long enough in reports zA & zD CB2 : Work Log Conflict High number of conflict/variation in the set of reports available on site 1 Conflicts/Variations in content of the reports rarely occur CC1 : Asset Location Reported Field “Asset Location” filled out in report 1A as well as in report 1D Field “Asset Location” filled out in report zA but not in report zD. . . . . .. . . . . . CT : Average Delay of Reporting Reports 1A was made 1h after the task, while report 1D was made with a delay of 3 weeks Both Reports zA and zD have been made with a delay inferior to 2h MCDM technique Site ranking considering all reports available on the different OEM sites : {Site 1, Site 2, Site 3. . . Site z} 2 3 SITE 1 1 SITE 2 SITE z Figure 2: Stages composing the maintenance reporting quality assessment framework are no better or worse techniques but some techniques are better suited to particular decision-making prob- lems than others. For example, AHP only deals with linear preferences and not with contextual preferences where values of one or several criteria may affect the importance or utility of other criteria. In this study, AHP is used (and combined with TOPSIS) for two reasons: i) we only deal with linear preferences and ii) AHP provides a powerful impartial, logical, and easy-to-use grading system, thus reducing personal biases and allowing for comparing dissimilar alterna- tives. Those characteristics are probably the main rea- sons for its success. According to a recent survey on MCDM techniques (Mardani et al., 2015), AHP is the second most used technique (applied in 16% of the reviewed literature1) after Hybrid MCDM (19.89%). 1 In total, 150 scientific journal papers were reviewed. 4. AHP-based MRQA framework The hierarchical structure defined using AHP con- sists consists of four levels, as depicted in Figure 3, namely: • Level 1: the overall goal of the study is to rank the different OEM company sites in terms of maintenance reporting quality; • Levels 2 and 3: the set of data quality dimensions (criteria) and sub-criteria defined in Table 2; • Level 4 the OEM company sites representing the alternatives. Given this hierarchy, AHP does perform the follow- ing computation steps for identifying the final ranking of the alternatives with respect to the overall goal: 1. Compare each element in the corresponding level and calibrate them on the numerical scale. This 7 Site 1 Site 2 Site 3 Site 4 . . . . . . . . . . . . . . . Site 54 CB1 CB2 CB3 CC1 CC2 CC3 CC4 CC5 CC6 CC7 CC8 CT Believability Completeness Timeliness Reporting Quality Assessment and Ranking of OEM Sites Level 1 Level 2 Level 3 Level 4 Figure 3: AHP structure of the maintenance reporting quality assessment process requires n(n−1) 2 pairwise comparisons, where n is the number of elements (diagonal elements be- ing equal to “1” and the other elements being the reciprocal of the earlier comparisons); 2. Perform calculation to find the maximum eigen- value, consistency index (CI), consistency ratio (CR), and normalized values; 3. If the computed eigenvalue, CI and CR are sat- isfactory, then decision/ranking is done based on the normalized values. These three stages 1, 2 and 3 are detailed in the following sections. In an effort to facilitate the un- derstanding, a scenario is considered in the following, whose parts are preceded by the symbol “➫”. 4.1. Pairwise comparison based preference measure- ment According to Blumenthal (1977), two types of judgment exist, namely: i) “Comparative judgment, which is the identifica- tion of some relations between two stimuli both present to the observer”; ii) “Absolute judgment, which involves the relations between a single stimuli and some information held in short term memory about some former comparison stimuli, or about some previously ex- perienced measurement scale using which the ob- server rates the single stimulus.” In a comparative/relative measurement, each alterna- tive is compared with many other alternatives, that is why this is also referred in the AHP literature to as “pairwise comparisons as ratio measurement” (Mumpower et al., 2012). In an absolute measure- ment, each alternative is compared with an ideal alter- native the expert knows of or can imagine, that is why this is referred to as “pairwise comparison based pref- erence measurement”. This section details the “pair- wise comparison based preference measurement” ap- proach, which is used at level 2 and 3 of the AHP structure (cf. Figure 3), while section 4.2 details the “pairwise comparisons as ratios” approach, which is used at level 4. Regarding pairwise comparison based preference measurement, decision makers have to evaluate the importance of one criterion (or sub-criterion) with re- spect to the others. To this end, OEM stakehold- ers perform pairwise comparisons among the iden- tified criteria as formalized in Eq. 1, with m the number of criteria to be compared (e.g., at level 2: m = |{CB,CC,CT}| = 3). The expert evaluation is carried out based on the 1- to 9-point Saaty’s scale: {1,3,5,7,9}; wi j = 1 meaning that Ci and C j are of equal importance and wi j = 9 meaning that Ci is strongly favored over C j. Note that all variables used in this paper are summarized in Table 3. P =             C1 . . . Cm C1 w11 . . . w1m .. . .. . ... .. . Cm wm1 . . . wmm             (1) The computation of the normalized eigenvector of P then enables to turn qualitative data into crisp ra- tios. Although several approaches exist in the lit- erature for normalized eigenvector computation, the Simple Additive Weighting (SAW) method (Hwang & Yoon, 1981) is used in this study, namely: WCi = ∑m j=1 wi j ∑m k=1 ∑m j=1 wk j , w ji =        1 i = j 1 wi j i , j (2) WC = [WC1, ..,WCi, ..,WCm ] P is characterized as consistent if, and only if Eq. 3 is respected. However, it is not that simple to fulfill this prerequisite when dealing with real expert prefer- ences, or when the number of criteria increases. Saaty (1980) proved that for consistent reciprocal matrix, the largest eigenvalue is equal to the size of the com- parison matrix (or λmax = m) and, accordingly, intro- duced CI as the deviation or degree of consistency2 2RI corresponds to the consistency index of a pairwise matrix generated randomly. 8 Table 3: Variable definitions Variables Description Cx abbreviation for criterion x with x = {1,2, ..,m}. Three criteria are defined at level 2 of the hierarchy structure, namely: CB, CC , CT (cf. Table 2). Cxh abbreviation for a sub-criterion of criterion x with h = {1,2, ..,y}. In this study, h = {1..3} for x = B (i.e., three sub-criteria of CB), h = {1..8} for x = C (i.e., three sub-criteria of CC ) and h = ∅ for x = T (i.e., CT does not have sub-criteria). P abbreviation for “Pairwise Comparison matrix”, whether at level 2, 3 or 4 of the AHP structure. wi j crisp value of a pairwise comparison matrix located at row i, column j of P. Al represents an alternative l in the AHP structure, with l = {1,2, ..,z}. In this study, l is the set of OEM sites to be assessed/ranked in terms of maintenance reporting quality. WCx , WCxh represents the eigenvalue of criterion Cx or sub-criterion Cxh (the eigenvector results from the P’s weight derivation process). In practice, it indicates the importance of one (sub)criterion over the others. I Cxh φ (Al) represents a digital indicator used for computing pairwise comparisons as ratios (i.e., measurable elements). Two indica- tors are defined, namely φ= {fill,avg,var} (see Eqs. 10, 14 and 16). W Al Cxh represents the eigenvalue of alternative Al with respect to sub-criterion Cxh. In practice, it indicates how good (or bad) the data quality is, regarding alternative/site l, with respect to Cxh . Rep Total(Al) function returning the total number of reports available on Site l. Size Delay(r,Al) function returning either (i) the reporting delay value or (ii) description length value of a given report (denoted by r) available on Site l. Rep filled(Al) function returning the number of reports (on Site l) for which a given “form field” has been filled out. Rep var(Al) function returning the number of reports/work orders (on Site l) containing content variation/conflicts. (see Eq. 4). If CR is smaller or equal to 10%, the in- consistency is regarded as acceptable. wi j = wik × wk j ∀i,k∈N|i,k; j∈N−{i,k} (3) CI = λmax − m m − 1 CR = CI RI (4) ➫ In this scenario, pairwise comparisons are filled out by the OEM’s executive officer. Eq. 5 shows the officer preference specifications regarding criteria at Level 2 of the AHP structure. The computed normal- ized eigenvector shows that the officer judges all cri- teria (at this level) of equal importance. Eq. 6 shows the pairwise comparisons carried out at Level 3 of the AHP structure, with regard to sub-criteria CBx | x = {1,2,3} (see Eq. 7 for the details of WCB1 computa- tion). The eigenvector (cf. Eq. 6) emphasizes that the officer judges “Length of Work Description” (i.e., CB1) slightly more important than “Work Log Con- flict” (i.e., CB2) for his/her own task, and highly more important than “Technician Log Variation” (i.e., CB3).              CB CC CT CB 1 1 1 CC 1 1 1 CT 1 1 1              ➠              WCB 0.33 WCC 0.33 WCT 0.33              (5) CR=0               CB1 CB2 CB3 CB1 1 3 5 CB2 1 3 1 3 CB3 1 5 1 3 1               ➠              WCB1 0.61 WCB2 0.29 WCB3 0.10              (6) CR=0.040 WCB1 = 1 + 3 + 5 1 + 3 + 5 + 1 3 + 1 + 3 + 1 5 + 1 3 + 1 (7) = 0.61 Similarly, the OEM officer carries out pair- wise comparisons between sub-criteria CCx | x = {1,2, ..,8}, as detailed in Eq. 8. According to the re- sulting eigenvector, CC1 is deemed as the most impor- tant sub-criterion (WCC1 ), followed by CC3 and CC2 re- spectively. Since CT does not have any sub-criterion, no pairwise comparison is required. 4.2. Pairwise comparisons as ratio measurement As previously mentioned, pairwise comparisons as ratio measurement is used at level 4 of the AHP struc- ture in order to compare alternatives with respect to each criterion, based upon measurable/supervised sys- tem parameters, e.g. how many times the field “DLC Code” (CC6) has been reported on Site l compared with the other sites (i.e., how many times such as form field has been left empty by maintenance operators). Eq. 9 gives insight into the pairwise comparisons as ratio matrix of the set of alternatives/sites Al with re- spect to the monitored system parameter Cxh. The ra- tio value is denoted by I Cxh φ (Al), as described in Ta- ble 3. The normalized eigenvector values of the pair- wise comparisons as ratio matrix with respect to cri- terion Cxh are denoted by W Al Cxh in Eq. 9. 9 CR=0.096                                               CC1 CC2 CC3 CC4 CC5 CC6 CC7 CC8 CC1 1 3 1 3 7 3 9 3 CC2 1/3 1 1/3 3 5 3 5 3 CC3 1 3 1 3 5 3 5 3 CC4 1/3 1/3 1/3 1 3 3 5 1 CC5 1/7 1/5 1/5 1/3 1 1 3 3 CC6 1/3 1/3 1/3 1/3 1 1 5 1/3 CC7 1/9 1/5 1/5 1/5 1/3 1/5 1 1/5 CC8 1/3 1/3 1/3 1 1/3 3 5 1                                               ➠                                               WCC1 0.266 WCC2 0.167 WCC3 0.242 WCC4 0.099 WCC5 0.065 WCC6 0.058 WCC7 0.023 WCC8 0.080                                               (8)                                 A1 A2 . . . Az A1 1 I Cxh φ (A1 ) I Cxh φ (A2 ) . . . I Cxh φ (A1 ) I Cxh φ (Az ) A2 I Cxh φ (A2 ) I Cxh φ (A1 ) 1 . . . I Cxh φ (A1 ) I Cxh φ (Az ) .. . .. . .. . ... .. . Az I Cxh φ (Az ) I Cxh φ (A1 ) I Cxh φ (Az ) I Cxh φ (A2 ) . . . 1                                 ➠                     W A1 Cxh W A2 Cxh .. . W Az Cxh                     (9) Three digital indicators I Cxh φ (Al) are defined (i.e., φ = {fill,avg,var}), which are described below. Ta- ble 2 highlights what indicators is used with regard to each criterion (see column named “Type”): • I Cxh fill (Al) (Filled Indicator – Eq. 10): used to cal- culate the proportion of reports where a given “field” was filled out on Site l ; Rep filled(Al) returning the number of reports that have been filled out, and Rep Total(Al) returning the total number of reports available on Site l: I Cxh fill (Al) = Rep filled(Al) Rep Total(Al) (10) ➫ Let us consider pairwise comparisons as ra- tio measurements between Sites 1 and 2, with re- spect to CC6. On Site 1, 76 maintenance reports have been carried out and 45 of them contain the DLC code (meaning that 59% of the available reports contain the requested information, see Eq. 11), while on Site 2 only 44% of the avail- able reports contain this information (see Eq. 12). The resulting pairwise comparisons as ratio ma- trix with respect to CC6 is given in Eq. 13, in which the above computed I CC6 fill (A1) and I CC6 fill (A2) are considered for the pairwise comparison as ratio measurement between Sites 1 and 2 (see row 1/column 2 of the matrix, and vice-versa). The resulting eingevector indicates how good (or bad) the reporting quality with respect to CC6 in this case – is regarding each site. I CC6 fill (A1) = 45 76 = 59% (11) I CC6 fill (A2) = 49 88 = 44% (12)                    A1 A2 . . . A54 A1 1 59 44 . . . 0.15 A2 44 59 1 . . . 0.67 . . . . . . . . . ... . . . A54 6.64 1.50 . . . 1                    ➠                        W A1 CC6 0.187 W A2 CC6 0.002 . . . . . . W A54 CC6 3E-06                        (13) Pairwise comparison as ratios can lead to com- putational issues when dividing I Cxh φ (Ai ) I Cxh φ (A j ) since the denominator may be null. For example, consid- ering the above scenario, if I CC6 fill (A2) = 0 (mean- ing that the DLC Code was never reported by any operator on Site 2), then I CC6 fill (A1 ) I CC6 fill (A2 ) = 59 0 , which pre- vents from performing the division. To bypass this problem, a penalty score θ is assigned to the corresponding site (i.e., A j) with respect to cri- terion Cxh, i.e. W A j Cxh = θ. However, a more in- depth study must be conducted to identify what penalty score should be assigned, how it affects the overall results, and so on. This study is pre- sented in Appendix B to avoid overloading the paper. • I Cxh avg(i) (Average Indicator – Eq. 14): used to cal- culate the average delays for maintenance re- porting per site (i.e., regarding CT ), or the av- erage length of work description (i.e., CB1) per site. Mathematically, I Cxh avg (Al) is computed based on Eq. 14, where Size Delay(k,Al) is either (i) the reporting delay value or (ii) the description length value, of a given report (denoted by r) available on Site l. I Cxh avg(Al) = Rep Total(Al ) ∑ r=1 SizeDelay(r,Al) Rep Total(Al) (14) 10 WCB1 0.610 WCB2 0.290 WCB3 0.100 WCC1 0.266 WCC2 0.167 WCC3 0.242 WCC4 0.099 WCC5 0.065 WCC6 0.058 WCC7 0.023 WCC8 0.080 WCB = 0.33 WCC = 0.33 WCT = 0.33 Reporting Quality Assessment and Ranking of OEM Sites 1.00 Site 1 : W A1 C B1 Site 2 : W A2 C B1 Site 54 : W A54 C B1 Site 1 : W A1 C B3 Site 2 : W A2 C B3 Site 54 : W A54 C B3 Site 1 : W A1 C C6 = 0.187 Site 2 : W A2 C C6 = 0.002 Site 54 : W A54 C C6 = 3E-06 Site 1 : W A1 C T Site 2 : W A2 C T Site 54 : W A54 C T Level 1 Level 2 Level 3 Level 4 S e e E q . 5 S e e E q . 6 & 8 S e e E q . 1 3 Figure 4: AHP structure and associated weights ➫ Let us assume that 4 reports are available on Site 1 (i.e. Rep Total(A1) = 4) and that the work description length is respectively equal to 44, 5, 13 and 101. The average indicator with regard to CB1 (on Site 1) is therefore equal to 40.75, as detailed in Eq. 15. The resulting pairwise com- parisons as ratios matrix is not presented due to similarities with the matrix detailed in Eq. 13. I CB1 avg (Al) = 44 + 5 + 13 + 101 4 = 40.75 (15) • I Cxh var (i) (Variation Indicator – Eq. 14): used to cal- culate the number of reports and/or work orders that contain variations or conflicts. One possi- ble conflict could be that the operator indicates a DLC code related to the car’s wheel (see CC6), while indicating in the Work Description (see CC2 ) that the car’s pump has been fixed. I Cxh var (Al) is computed based on Eq. 16, with Rep var(Al) the number of reports that contain content varia- tion (or conflicts) on Site l. I Cxh var (Al) = Rep var(Al) Rep Total(Al) (16) In an effort to summarize all the variables and weights computed in this section, we provide an “at a glance” representation of the AHP hierarchy in Fig- ure 4, which refers to the different equations consid- ered to compute the variable weights. In the next sec- tion, we present how those different weights are ag- gregated to obtain the final site ranking. 4.3. Alternative ranking using TOPSIS The different weights must now be aggregated in or- der to obtain a global weight of each alternative with respect to all criteria, which is computed based on Eq. 17. All these global weights are summarized in the form of a matrix in Eq. 18. GW Al Cxh = W Al Cxh × WCxh × WCx (17)                                  CB1 . . . CB3 CC1 . . . CC8 CT A1 GW A1 CB1 . . . GW A1 CB3 GW A1 CC1 . . . GW A1 CC8 GW A1 CT A2 GW A2 CB1 . . . GW A2 CB3 GW A2 CC1 . . . GW A2 CC8 GW A2 CT . . . . . . . . . . . . . . . . . . . . . . . . Az GW Az CB1 . . . GW Az CB3 GW Az CC1 . . . GW Az CC8 GW Az CT                                  (18) ➫ For illustration purposes, Eq. 19 details the global weight computation for A1 (i.e., Site 1) with respect to criterion CC6, which implies to takes into account WCC -related weight. GW A1 CC6 = W A1 CC6 × WCC6 × WCC (19) = 0.187 × 0.058 × 0.333 = 0.0036 The global weights (cf. Eq. 17) must now be aggre- gated for each alternative in order to obtain the final quality score, based on which the final site ranking is generated. To this end, the TOPSIS method is em- ployed or, to be more accurate, combined with AHP. TOPSIS introduces for each alternative the closeness coefficient denoted by R(Al), which implies comput- ing for each criterion xh the positive ideal solution (PIS) denoted by d+ xh and negative ideal solution (NIS) denoted by d− xh , as formalized in Eq. 20 and 21 respec- tively. The distances measuring the separation from PIS and NIS are then computed in Eq. 22 and 23, re- spectively denoted D+ Al and D− Al ). d + xh = max l=1..z ( GW Al Cxh ) (20) d − xh = min l=1..z ( GW Al Cxh ) (21) D + (Al) = √ ∑ xh ( GW Al Cxh − d+ xh )2 l = 1, ..,z (22) D − (Al) = √ ∑ xh ( GW Al Cxh − d− xh )2 l = 1, ..,z (23) A prior alternative has a longer distance to NIS and a shorter distance to PIS. Consequently, the closeness 11 Table 4: Alternative ranking illustration One ranking per quality dimension Overall Ranking Believability Completeness Timeliness Site 1 30th 3rd 2nd 12th Site 2 4th 15th 27th 15th . . . . . . . . . . . . . . . Site 11 34rd 7th 1st 14th . . . . . . . . . . . . . . . Site 32 1nd 2th 18th 7th . . . . . . . . . . . . . . . Site 47 19th 35th 31th 28th . . . . . . . . . . . . . . . coefficient to the ideal solution for each alternative can be formulated as in Eq. 24, where R(Al) denotes the final performance score of Site l. The larger the R(Al) score, the better the maintenance reporting quality on the corresponding site. R(Al) = D−(Al) D+(Al) + D −(Al) l = 1, ..,z (24) The overall site ranking can therefore be generated based on the R(Al) performance scores. Nonetheless, let us note that in Eq. 22 and 23, if: • xh = {CB1, ..,CB3,CC1, ..,CC8,CT}: a single and overall ranking of the sites is generated (i.e., all criteria/sub-criteria are aggregated), as shown in Table 4 (see column “Overall Ranking”); • xh = {CB1, ..,CB3} or xh = {CC1, ..,CC8} or xh = {CT}: one ranking per quality dimension (i.e., CC , CB or CT ) is generated, as shown in Ta- ble 4 (see column named “One ranking per qual- ity dimension”). Having indicators per dimen- sion enable e.g. a site manager to further inves- tigate i) what dimension(s) must be enhanced in the short, medium or long term, ii) to track the evolution over time of the reporting quality of one or a group of sites with respect to specific dimensions, etc. ➫ Figure 5 gives insight into how a company’s stakeholder can use the “ranking per quality dimen- sion” to better understand how a site behaves (i.e., how good/bad it is) with respect to one or more di- mensions3: the larger the surface areas in Figure 5, the better the site ranking and, as a consequence, the better the reporting quality on this site. 5. OEM use case Two distinct use cases, defined at the tactical and operational levels, are presented in this section and show how the Finnish OEM company takes advan- tage of the MRQA dashboard. Figure 6 provides an 3 The four sites highlighted in bold in Table 4 are considered and displayed in this example. Timeliness 1 st 12th 23 th 34 th 45 th 1 st 12th 23 th 34 th 45 th 54 th 45 th 34 th 23 th 12th 1 st Site 11 Site 32 Site 47 Completeness Believability Figure 5: Comparison of sites 11, 32 and 47 (cf. Table 4) overview of the architecture and associated tools that have been developed/set up in the company: main- tainers on the different sites report maintenance work order-related information using the company’s online form (see “Reporting Service” in Figure 6). A screen- shot of the company’s online form is provided in Fig- ure 7(a), which has been annotated to help the reader to understand what form fields correspond to what AHP (sub-)criteria. In total, by summing all reports from all sites, 275 585 reports have been processed and analyzed. The MRQA dashboard thereby enables any site stakeholder (e.g., plant manager, head officer. . . ) to assess – at a given point in time and based on the his/her own preferences – the quality of reporting of the 54 branch offices. As highlighted in Figure 6, when a stakeholder requests for the site ranking ser- vice, the overall ranking is computed at the head of- fice (i.e., in Finland). In practice, a set of SQL queries is performed against the different database systems – spread over the 54 sites – that contain the maintenance reports. The retrieved information is then used as in- puts of the pairwise comparisons as ratio measure- ment process. A screenshot of the MRQA dashboard is given in Figure 7(b), which provides the stakeholder with the possibility to: • access it through a web browser; • vizualize in a user-friendly way both the loca- tions of the OEM sites (see the dashboard ele- ment named “Map of Company Sites”) and their corresponding quality/ranking (see “Final As- sessment & Ranking”); • deeper investigate the maintenance reporting quality of one or more sites with respect to one or more quality dimensions (see the “Disintegrated Quality View” dashboard element); 12 Internet Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date MRQA algorithm(s) { { { { { { Head Office (OEM) b ➠ b ➠ b ➠ ✓ T ask 1 Tas k 2 ✓ T ask 3 ✓ T ask 1 Tas k 2 ✓ T ask 3 ✓ T ask 1 Tas k 2 ✓ T ask 3 Reporting Service Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date MRQA dashboard Site Manager or other b ➠ b ➠ b ➠ Reporting Service ✓ T ask 1 Tas k 2 ✓ T ask 3 ✓ T ask 1 Tas k 2 ✓ T ask 3 ✓ T ask 1 Tas k 2 ✓ T ask 3 Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Maintenance Operator (Site 1) Report ID : 1A Asset Location Description Actual End-Date Target Start Date . . . Scheduled Start Date Scheduled End Date Site Manager or other MRQA dashboard Other sites Figure 6: Overall infrastructure underlying the MRQA dashboard • adjust the assessment period, e.g., to assess/com- pare sites over the last few weeks, months or years (see the “Assessment period” dashboard el- ement); • modify his/her preferences related to the crite- ria importance, e.g. if he/she wants to give – at a specific point in time – further importance to one or more dimensions (e.g., Completeness over Believability) or sub-dimensions (e.g., CC4 over CC1 ). This is discussed further in this section, but the reader can already refers to Figure 10 to have an overview of the dashboard functionality. Two distinct scenarios are presented in sections 5.1 and 5.2 respectively. The first one provides in- sight into the reporting quality assessment results when considering the importance between criteria as roughly equivalent, while the second scenario high- lights how the MRQA dashboard can be used for a specific purpose, namely to set up a cost-reduction ac- tion plan in the proposed scenario. 5.1. Scenario 1: Equivalence between criteria At the operational level, the head officer wants to have an overview of the maintenance reporting qual- ity regarding all sites, without prioritizing any qual- ity dimension. To this end, the officer does perform pairwise comparisons by specifying that all criteria are equal in importance (as carried out in Eq. 5), and similarly for the sub-criteria. Figure 8(a) gives insight – in the form of a histogram – into the quality assess- ment results, where the x-axis refers to the 54 sites and the y-axis to the quality score obtained after applying AHP. It can be observed that three sites stand out (hav- ing the highest quality scores), namely sites 46, 12 and 18 respectively. The officer wants to further investigate the reasons behind the low quality score of Sites 6 and 54 (i.e., sites having the poorest quality). To this end, the of- ficer selects – in the “Disintegrated Quality View” dashboard element in Figure 7 – these two sites, thus having a deeper insight into the site level quality with respect to each level 2 quality dimension. It can be observed that Site 6 has a particularly poor ranking regarding both the “Timeliness” and “Completeness” dimensions, while Site 54 has a poor ranking regard- ing “Timeliness” and “Believability”. The officer can even go a step further in the analysis in order to under- stand the reasons behind the low quality score of a site regarding one of these dimensions. For example, in the dashboard’s screenshot (cf. “Disintegrated quality view (level 3)” in Figure 7), the officer has selected the ‘Completeness’ dimension and can visualize the per- centage of form fields (which have been turned into sub-criteria CC1 to CC8) that have been or not reported with respect to the total number of reports on the se- lected sites (i.e., on Sites 6 and 54). It can be observed that maintainer operators on Site 6 report less often information than on Site 54, which is the main reason why Site 6 has a lower quality rank than Site 54, as pointed out above. Nonetheless, an interesting point in this graph is that CC6 reports more CC6 -related in- formation (i.e., the ‘DLC Code’), namely 55% against 18% for Site 54. In summary, this first scenario offered an overview of the different MRQA dashboard functionalities, and how the associated views can be used as decision 13 CB1 ➙ ➙ CC2 CC3 ➙ CC4➙ CC5➙ CC7➙ CC8 Legend Form fields assessed in terms of “Believability”: CB = {CB1,CB2,CB3} Form fields assessed in terms of “Timeliness”: CT Form fields assessed in terms of “Completeness”: CC = {CC1,CC2, . . . ,CC8} (a) Online form used/filled out by maintainer operators on each OEM site Disintegrated Quality View (level 3) (b) MRQA dashboard interface Figure 7: Screenshots of the “Maintenance Work Order System” & the “MRQA dashboard” 14 R e la ti v e Q u a li ty S c o re 0 0.2 0.4 0.6 0.8 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 Site number (i.e., set of alternatives) (a) Scenario 1: equivalence between criteria 0 0.024 0.048 0.072 0.096 0.12 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 Site number (i.e., set of alternatives) (b) Scenario 2: Cost-reduction action plan Figure 8: Overall site ranking: Scenarios 1 & 2 Failure ✹ Create Work Order Spare Parts Arrived Start repairing at scheduled date b Start date End date Repair work completion Report after customer checking & Validation y✔ T I M E L I N E S S B Maintenance work order tracking system (see Figure 7(a)) T I M E L I N E S S B Figure 9: Problem to be addressed in the maintenance reporting process to avoid financial losses support tools by various maintenance stakeholders. The second scenario, presented in the next section, puts further emphasis on how a site stakeholder can take advantage of the dashboard for improving cost- reduction action plan. 5.2. Scenario 2: Cost-reduction action plan Over the past few years, the OEM company has been facing a major problem in the maintenance re- porting process at the tactical level, leading to sub- stantial financial losses. The traditional maintenance process4 is depicted in Figure 9, where a work order is created by the maintenance planner as soon as a problem/failure is reported by the customer. Once the spare parts are supplied on site, the site manager re- ports it into the system, and the maintainer can start to repair the defective equipment at the scheduled time. As emphasized in Figure 9, the maintainer has to spec- ify – into the maintenance work order tracking sys- tem – the “Start date” and “End date” (both dates be- ing taken into account in our AHP under the “Com- pleteness” dimension, and particularly with CC3 ). Fol- lowing the “Repair work completion” (i.e., End date), it is necessary to wait until the customer checks and validates the maintenance service as depicted in Fig- ure 9. Such a time interval actually corresponds to 4 Only the Scheduled Maintenance Process is described (not the Unscheduled process) to ease the understanding. the “Timeliness” criterion. This interval is very crit- ical for the OEM company because the company has to pay a penalty fee that depends on the time inter- val (obviously under the condition that the customer does not validate, or complain against the service, and obtains a favorable ruling). Given the above-mentioned problem, the head of- ficer wants to draw up an assessment of the overall situation (i.e., with regard to each site), and to shape a proper action plan for reducing financial losses. This plan consists first in identifying and managing sites that have the poorest timeliness quality, since they have the highest financial risk. To this end, the offi- cer does specify that Timeliness (CT ) is strongly more important than Believability (CB) and Completeness (CC ) in order to bring to light the sites with the poor- est Timeliness quality. Figure 10 shows how this can be specified through the MRQA dashboard (see red/- dashed frame). The final ranking is generated and given in Figure 8(b), showing that Sites 18, 42, 12, 6 and 5 are respectively the branch offices that have the highest risk for monetary losses. Although the action plan to be set up by the OEM company to reduce such risks on the identified sites is out of scope of this paper, it is worth noting that IoT messaging protocols will likely be implemented in the future to address part of the problem. At a more concrete level, such protocols will be used to gener- ate ‘event-based’ notifications to the customer as soon 15 Pairwise comparison at level 1 of the AHP structure by the OEM head officer Figure 10: Dashboard/User Interface to adjust the pairwise comparison based preference measurement as the repair work is completed (i.e., when the “End date” is entered in the system), e.g. by explicitly stat- ing that the customer has a n-day deadline to check and validate the maintenance work. From a practical viewpoint, the recent IoT standards published by The Open Group (Främling et al., 2014) will first be im- plemented on the riskiest sites. 6. Conclusions, implications, limitations and fu- ture research 6.1. Conclusions Data, information and knowledge are the “new oil” of the digital era, and are at the heart of all busi- ness operations. It is therefore crucial for companies to implement the right infrastructure to monitor and improve the quality of data generated throughout the lifecycle of the company’s assets (either physical or virtual assets). It is a fact that the data quality has a significant impact on the overall incomes and ex- penditures of companies: poor data quality impacting the downstream processes, and reciprocally, high data quality fostering enhanced business activities and de- cision making. However, it remains challenging to as- sess information quality, as information is not as tan- gible as physical assets. The literature review carried out in this paper brings to light the fact that current expert maintenance sys- tems fail, or have no specific interest, to take into ac- count data quality dimensions in the maintenance re- porting quality assessment process. To fulfill this gap, this paper develops a Maintenance Reporting Qual- ity Assessment (MRQA) dashboard that enables any company stakeholder to easily – and in real-time – assess/rank company branch offices in terms of main- tenance reporting quality. In this respect, AHP is used to integrate various data quality dimensions as well as expert preferences. The paper presents two scenar- ios showing how the MRQA dashboard is being used by a Finnish multinational equipment manufacturer to assess and enhance maintenance reporting practices in one or more branch offices. This should contribute to enhance other organization activities such as: • after-sales services: the quality of maintenance reports makes it possible to assess the mainte- nance work, thus helping to reach a higher qual- ity after-sales services; • on the design of future generations of products: processing and analyzing relevant maintenance reports help to better understand how company assets behave throughout their lifecycle which, in turn, help to enhance the design of future gen- erations of products (Främling et al., 2013); • predictive maintenance strategies: providing real-time and remote predictive maintenance is becoming a very promising area in the so-called IoT, whose objective is to provide systems with the capability to discover and process real-time 16 data and contexts so as to make pro-active deci- sions (e.g., to self-adapt the system before a pos- sible failure) (Främling et al., 2014). Although real-time data is of the utmost importance in the predictive maintenance process, combining such data with historical maintenance reporting data (regarding a specific product item) has the poten- tial to generate new knowledge and lead to more effective and product-centric decisions; • government regulation compliance: in some do- mains, it is mandatory to comply with govern- ment regulations (e.g., in automotive, avionics, or healthcare domains). In this respect, assess- ing the quality of maintenance reporting can pre- vent the company from having regulation non- compliance issues, e.g. by carefully following the data quality on each branch office and identi- fying as soon as possible quality issues regarding one or more dimensions; 6.2. Implications This research presents three main theoretical im- plications. First, it contributes to the literature on maintenance management (MM) by proposing a thor- ough state-of-the-art on the use of MCDM techniques at each MM level (Strategic, Tactical, Operational), which helps identifying criteria at each of these levels (cf. Table 5). This list of criteria can be of potential value to future researchers working in MM. Second, this research contains an approach to identify relevant data quality dimensions (based on existing data qual- ity frameworks), and to turn them into a hierarchical AHP structure. A theoretical framework is then pro- posed, enabling the assessment and ranking of differ- ent company branch offices in terms of maintenance reporting quality. To the best of our knowledge, and as evidenced through our state-of-the-art, this is the first research work that addresses this specific goal. Finally, the research also contributes three main managerial implications. First, it enables organization stakeholders to realize how important it is to moni- tor and assess maintenance reporting practices, as it can impact downstream but also upstream activities of the organization. Second, the proposed dashboard helps practitioners to quickly identify, based on their needs and preferences, how one or a group of sites behave (i.e., how good/bad they are) with respect to one or more dimensions. This is helpful to estab- lish their strategic plans to improve current practices, which may result in savings of both money and time. 6.3. Limitations The theoretical implications discussed above rely both on the Krogstie’s data quality framework to iden- tify key data quality dimensions, and AHP as MCDM technique to structure these quality dimensions in the form of a hierarchy that makes easier for maintenance stakeholders to specify their needs. However, this re- search has several limitations. First, only a few con- cepts and relationships from the Krogstie’s framework were considered (see red/bold elements in Figure 1), which is due to the data that has been made available by the Finnish OEM company, as well as to their own expectations/needs. In future research work, the pro- posed AHP framework and underlying criteria should be extended to take into consideration the other con- cepts/relationships such as Language Quality (e.g., for domain appropriateness, participant knowledge ap- propriateness. . . ), Syntactic Quality, etc.. Such an extension might potentially require to combine AHP with other tools and techniques for semantic pro- cessing and matching purposes for example, or still for handling uncertainty and vagueness in the expert judgments/preferences (e.g., using fuzzy logic). Furthermore, although it is already a great achieve- ment for the Finnish company to be able to iden- tify how good/bad their branch offices are in report- ing maintenance data, we would have liked to carry out a post-analysis to evaluate benefits of the post- action plans carried out in the different branch of- fices. For example, we are aware that the company have developed on-site training programs, which have been customized according to the quality results re- lated to each site (e.g., if a site fails in addressing one or more data quality dimensions, the training program is customized accordingly). However, such a post- analysis and insightful implications cannot be pro- vided because of contractual obligations and project constraints. 6.4. Future research From a research perspective, we developed a frame- work that makes possible the ranking of maintenance sites based on their respective reporting quality. Cur- rently, the proposed MRQA dashboard is limited to the visualization of the site’s data quality score and as- sociated rank. In future research work, we would like to extend this tool, and particularly to re-use the final site’s (AHP) data quality score as an input parameter of a more advanced framework (e.g., that would in- tegrate live sensor data from manufacture equipment) that would make it possible to decide – in real-time – what predictive failure model for machine is better suited. This is based on two working assumptions: • a weak assumption: the higher the maintenance reporting quality score on-site (denoted by R(Al) in this study), the higher the confidence of the failure prediction; • a strong assumption: the confidence of one or more predictive models (e.g., binary logic model, cox regression model, regression trees model. . . ) 17 Table 5: Percentage of Criteria used in the Maintenance Management (MM) literature Strategic Level Tactical Level Operational Level Cost 22.7 Cost 21.9 Cost 25.8 Resource Availability & Utilization 10.1 Environment./Operation. Condition 14.1 Resource Availability & Utilization 19.4 Added Value 7.6 Safety 9.4 Added Value 4.8 Safety 7.6 Resource Availability & Utilization 9.4 DownTime & Time to repair 4.8 Reliability 7.6 Risk/Severity 9.4 Quality 4.8 Environment./Operation. Condition 5.0 DownTime & Time to repair 7.8 Risk/Severity 4.8 Quality 5.0 Reliability 6.3 Organizational Process 4.8 Risk/Severity 5.0 Added Value 3.1 Failure Frequency 3.2 Failure Frequency 4.2 Knowledge 3.1 Knowledge 3.2 Feasibility (implementation) 4.2 Resources Age 3.1 Environment./Operation. Condition 3.2 Repairability 3.4 Failure Frequency 1.6 Reliability 3.2 DownTime & Time to repair 3.4 Detectability 1.6 Safety 1.6 Flexibility 3.4 Feasibility (implementation) 1.6 Repairability 1.6 Knowledge 1.7 Maintenance Frequency 1.6 Abilities & development 1.6 Geographical Location 2.5 Comfort 1.6 Collaboration with Stakeholders 1.6 Component Failed 1.7 Automation 1.6 Operational Time 1.6 Detectability 0.8 Laws and Regulation 1.6 Maintenance Impact 1.6 Difficulty and Challenges 0.8 Number of affected people 1.6 Performance 1.6 Support and Services 0.8 Customer Category 1.6 Management & Organization 0.8 Number of sorties flown 1.6 Machine Uses 1.6 might evolve according to the reporting quality score, whose evolution might even (potentially) differ from one model to another. To put it an- other way, we might assume that according to the on-site maintenance reporting quality score (e.g., if R(Al) < 60%), a binary logic model might pro- vide more confident predictions than a regression trees model, or vice-versa. The objective of future research will be to validate (or invalidate) these two working assumptions and, if val- idated, to propose a more advanced framework that is able to switch between two or more predictive models and react accordingly. 7. Acknowlegement The research leading to this publication is sup- ported by the Finnish Metals and Engineering Compe- tence Cluster (FIMECC) S4Fleet program and the Na- tional Research Fund Luxembourg (grant 9095399). Appendix A. Percentage of criteria considered in the Maintenance literature Based on the summary matrix given in Table 1, re- porting what MCDM techniques is commonly used at each MM level, we carried out an in-depth analysis to identify the most commonly used criteria at each of these level in order to see whether data quality is prop- erly addressed in the maintenance literature, and par- ticularly regarding maintenance reporting activities. The analysis outcome, regarding each MM level, is given in the form of tabular in Table 5, which high- light that data quality is hardly considered in the re- viewed papers knowing that “Quality” refers, in most of the reviewed papers, to other quality aspects than Data Quality, except in (Van & Pintelon, 2014). Appendix B. Penalty score selection The methodology defined to tune the penalty score consists in studying whether the introduced penalty has a significant impact on the overall ranking. Let us consider, in Eq. B.1, the pairwise comparisons as ratio matrix introduced as example in section 4.2, where it is assumed now that the form field(s) related to crite- rion CC6 has/have been left empty in all reports car- ried out on Site 1 (i.e., in 100% of the reports). Con- sequently, I CC6 fill (A1) = 0, as highlighted in the first column and row of the matrix in Eq. B.1. Our strat- egy is to give out a penalty score (denoted by θ) as eigenvalue to the corresponding site (i.e., to Site 1) as shown in Eq. B.1.                                   A1 A2 . . . A54 A1 1 ✓✓ 0 44 . . . ✚ ✚ ✚0 I CB1 fill (Al ) A2 ✓✓ 44 0 1 . . . 0.67 . . . . . . . . . ... . . . A54 ✚ ✚ ✚ICB1 fill (Al ) 0 1.50 . . . 1                                   ➠                            θ W A2 CC6 . . . W A54 CC6                                                    A1 A2 . . . A54 A1 — — — — A2 — 1 . . . 0.67 . . . . . . . . . ... . . . A54 — 1.50 . . . 1                         ➠                            θ W A2 CC6 . . . . . . W A54 CC6                            (B.1) In order to select the penalty score θ, we propose to carry out an analysis to determine whether the intro- duced score impacts substantially or not on the overall 18 %corr Siteranking 1 0 0 54 {CB = 0.6, CC = 0.2, CT = 0.2} Jacc. {CB = 0.4, CC = 0.4, CT = 0.2} Jacc. {CB = 0.4, CC = 0.2, CT = 0.4} Jacc. {CB = 0.2, CC = 0.6, CT = 0.2} Jacc. {CB = 0.2, CC = 0.4, CT = 0.4} Jacc. MRQA tool {CB = 0.2, CC = 0.2, CT = 0.6} θ = 0 Jacc. {CB = 0.6, CC = 0.2, CT = 0.2} {CB = 0.4, CC = 0.4, CT = 0.2} {CB = 0.4, CC = 0.2, CT = 0.4} {CB = 0.2, CC = 0.6, CT = 0.2} {CB = 0.2, CC = 0.4, CT = 0.4} {CB = 0.2, CC = 0.2, CT = 0.6} θ = −1 Figure A.11: Comparison process – based on the Jaccard similarity coefficients – set up for similarity measurements between distinct site rankings (i.e., considering various criteria preferences and penalty scores) ranking. To this end, two distinct penalty scores are considered: • θ = 0: the site is penalized compared with the other sites since any site that does not get a penalty has automatically an eignevalue greater than zero, or to be more precise 0 < W Al Cxh < 1; • θ = −p | p ∈ R−: the site is penalized compared to the other sites, whose effect (unlikeθ= 0) is to bring down the overall ranking when aggregating all AHP dimensions/criteria. To identify whether the penalty scores impact (in a substantial manner) the overall ranking, we propose – as depicted in Figure A.11 – to generate/compute the alternative ranking for each penalty score (con- sidering a given set of criteria weights/preferences) and to compare whether the two rankings vary from each other. This process has been tested for six com- binations of criteria weights, as emphasized in Fig- ure A.11. The similarity measure between distinct rankings is based on the Jaccard similarity coeffi- cients, whose principle is described in section Ap- pendix B.1. Results and concluding remarks about the penalty score selection are presented in section Ap- pendix B.2. Appendix B.1. Jaccard-based similarity measure The Jaccard similarity coefficients (Tan et al., 2006) can be used to measure a similarity between two dis- tinct lists A and B, as formalized in Eq. B.2 (i.e., the size of the list intersection divided by the size of the list union). In our study, the union size is equal to the number of alternatives/sites z. A Jaccard similarity coefficient goes from 0 (no common list) to 1 (identi- cal lists). J(A, B) = |A ∩ B| |A ∪ B| = |A ∩ B| z (B.2) Let A, B and C be three distinct lists consisting of five sites {S1, ..,S5}, where each site receives a final rank as presented in Figure B.12. In this example, two Jaccard similarity coefficients J(A, B) and J(A,C) are calculated. Both coefficients are equal because the in- tersections |A∩B|and |A∩C| have the same cardinality. A B C S1 1 1 7 S2 2 2 6 S3 3 3 1 S4 4 6 2 S5 5 7 3 J(A, B) = |A∩B| z = |1,2,3| |1,2,3,4,5| = 3 5 J(A,C) = |A∩C| z = |1,2,3| |1,2,3,4,5| = 3 5 Figure B.12: Computation of Jaccard similarity coefficients In our study, sites are ordered according to their data quality score. It could be worthwhile to define a similarity coefficient that would take into account the rank. To this end, let us define Lq to be a sub- list of L, where Lq consists of sites from rank 1 to q (q ≤ z). A progressive similarity coefficient Jq(A, B) can therefore be computed as in Eq. B.3. Jq(A, B) = J(Aq, Bq) (B.3) Figure B.13 details the evolution of the Jaccard progressive coefficients Jq(A, B) and Jq(A,C) with q = 1,2, ...,5 (see lists A, B, and C given in Figure B.12). Appendix B.2. Penalty score impact and selection As previously stated and summarized in Fig- ure A.11, the alternative ranking for each penalty 19 J1(A, B) = |A1∩B1| 1 = 1.00 J1(A,C) = |A1∩C1| 1 = 0.00 J2(A, B) = |A2∩B2| 2 = 1.00 J2(A,C) = |A2∩C2| 2 = 0.00 J3(A, B) = |A3∩B3| 3 = 1.00 J3(A,C) = |A3∩C3| 3 = 0.33 J4(A, B) = |A4∩B4| 4 = 0.75 J4(A,C) = |A4∩C4| 4 = 0.50 J5(A, B) = |A5∩B5| 5 = 0.60 J5(A,C) = |A5∩C5| 5 = 0.60 Figure B.13: Computation of Jaccard progressive coefficients 0 0.2 0.4 0.6 0.8 1 0 9 18 27 36 45 54 S im il a ri ty c o e ffi c ie n ts Site ranking Criteria Preferences: {CB = 0.6,CC = 0.2,CT = 0.2} Criteria Preferences: {CB = 0.4,CC = 0.4,CT = 0.2} Criteria Preferences: {CB = 0.4,CC = 0.2,CT = 0.4} Criteria Preferences: {CB = 0.2,CC = 0.6,CT = 0.2} Criteria Preferences: {CB = 0.2,CC = 0.4,CT = 0.4} Criteria Preferences: {CB = 0.2,CC = 0.2,CT = 0.6} Figure B.14: Penalty score impact on site ranking: θ= 0 vs. θ=−1 score (i.e., for θ = 0 and θ = −1) is generated, where the two resulting rankings are compared based on the Jaccard similarity measure. In total, six sim- ilarity comparisons are performed (cf. Figure A.11), whose results are displayed in Figure B.14. These re- sults show that the choice of the penalty score does not lead to significant changes in the final ranking (al- though a few sites move up or down between the 27th and 36th positions), and is not dependent on the crite- ria weights. Given this observation, the penalty score θ = 0 has been chosen in this study. An additional reason for choosing this score is that the sum of the eigenvector values are equal to 1 (thus respecting the eigenvector property/axiom), which is not true when choosing θ=−1. References Ahmadi, A., Gupta, S., Karim, R. and Kumar, U. (2010). Selection of maintenance strategy for aircraft systems using multi-criteria decision making methodologies. International Journal of Relia- bility, Quality and Safety Engineering, 17, 223–243. Alarcón, M. J., Grau, J. B. and Torres, J. (2007). Application of ELECTRE I method to restoration actions in telecommunication network maintenance. IEEE. Almeida, A.T. (2012). Multicriteria model for selection of preven- tive maintenance intervals. Quality and Reliability Engineering International, 28, 585–593. Almeida-Filho, A., Ferreira, R.J. and Almeida, A. (2013). A DSS based on multiple criteria decision making for maintenance planning in an electrical power distributor. Springer. Arputhamary, B. and Arockiam, L. (2015). Data Integration in Big Data Environment. Bonfring International Journal of Data Min- ing, 5, 1. Azadeh, A., Sheikhalishahi, M., Firoozi, M. and Khalili, S.M. (2013). An integrated multi-criteria Taguchi computer simulation-DEA approach for optimum maintenance policy and planning by incorporating learning effects. International Journal of Production Research, 51, 5374–5385. Azadeh, A., Sheikhalishahi, M., Khalili, S. M. and Firoozi, M. (2014). An integrated fuzzy simulation–fuzzy data envelopment analysis approach for optimum maintenance planning. Interna- tional Journal of Computer Integrated Manufacturing, 27,181– 199. Azizi, A. and Fathi, K. (2014). Selection of optimum maintenance strategies based on a fuzzy analytic hierarchy process. Manage- ment Science Letters, 4, 893–898. Babashamsi, P., Golzadfar, A., Yusoff, N. I. M., Ceylan, H. and Nor, N. G. M. (2016). Integrated fuzzy analytic hierarchy process and VIKOR method in the prioritization of pavement maintenance activities. nternational Journal of Pavement Research and Tech- nology, 9, 112–120. Behzadian, M., Khanmohammadi Otaghsara, S., Yazdani, M. and Ignatius, J. (2012). A state-of the-art survey of TOPSIS applica- tions. Expert Systems with Applications, 39, 13051–13069. Bertolini, M., Bevilacqua, M., Braglia, M. and Frosolini, M. (2004). An analytical method for maintenance outsourcing service se- lection. International Journal of Quality & Reliability Manage- ment, 21, 772–788. Bertolini, M. and Bevilacqua, M. (2006). A combined goal pro- gramming – AHP approach to maintenance selection problem. Reliability Engineering & System Safety, 91, 839–848. Bevilacqua, M. and Braglia, M. (2000). The analytic hierarchy pro- cess applied to maintenance strategy selection. Reliability Engi- neering & System Safety, 70, 71–83. Blumenthal, A. L. (1977). The process of cognition. Prentice Hall/- Pearson Education. Cafiso, S., Di, G. A., Kerali, H. and Odoki, J. (2002). Multicri- teria analysis method for pavement maintenance management. Transportation Research Record: Journal of the Transportation Research Board, 1816, 73–84. Cavalcante, C. A. V. and Costa, A. P. C. S. (2010).Multicriteria Model of Preventive Maintenance. Brazilian Journal of Oper- ations & Production Management, 3,71–86. Cavalcante, C. A. V., and Ferreira, R. J. P. and de Almeida, A. T. (2010). A preventive maintenance decision model based on multicriteria method PROMETHEE II integrated with Bayesian approach. IMA Journal of Management Mathematics, 21, 333– 348. Cavalcante, C.A.V. and De Almeida, A.T. (2007). A multi-criteria decision-aiding model using PROMETHEE III for preventive maintenance planning under uncertain conditions. Journal of Quality in Maintenance Engineering, 13, 385–397. Certa, A., Enea, M. and Lupo, T. (2013). ELECTRE III to dynami- cally support the decision maker about the periodic replacements configurations for a multi-component system. Decision support systems, 55,126–134. Charnes, A., Clark, C. T., and Cooper, W. W. and Golany, B. (1984). A developmental study of data envelopment analysis in measur- ing the efficiency of maintenance units in the US air forces. An- nals of Operations Research, 2,95–112. Chen, M., Mao, S., & Liu, Y. (2014). Big data: A survey. Mobile Networks and Applications, 19, 171–209. Chen, L., Weng, M. and Zhang, G. (2005). Utility Optimality Method for Pipeline Integrity Maintenance Costs. Petroleum En- gineering Construction, 2,004. Chou, J. (2009). Web-based CBR system applied to early cost bud- geting for pavement maintenance project. Expert Systems with Applications, 36, 2947–2960. Coulter, E. D., Sessions, J. and Wing, M. G. (2006). Scheduling forest road maintenance using the analytic hierarchy process and heuristics. Silva Fennica, 40, 143. 20 de Almeida, A. T. (2001). Multicriteria decision making on mainte- nance: spares and contracts planning. European Journal of Op- erational Research, 129,235–241. de Almeida, A. T., Cavalcante, C. A. V., Alencar, M. H., Ferreira, R. J. P., de Almeida-Filho, A. T and Garcez, T. V. (2015). Decision on Maintenance Outsourcing. Springer. de Almeida, A. T., Cavalcante, C. A. V., Alencar, M. H., Ferreira, R. J. P., de Almeida-Filho, A. T. and Garcez, T. V. (2015). Decisions on Priority Assignment for Maintenance Planning. Springer. de Almeida, A. T., Cavalcante, C. A. V., Alencar, M. H., Ferreira, R. J. P., and de Almeida-Filho, A. T. and Garcez, T. V. (2015). Preventive Maintenance Decisions. Springer. de un Caso, E. (2008). The Efficiency of Preventive Maintenance Planning and the Multicriteria Methods: A Case Study. Com- putación y Sistemas, 12,208–215. Dehghanian, P., Fotuhi-Firuzabad, M., Bagheri-Shouraki, S. and Kazemi, A. A. R. (2012). Critical component identification in reliability centered asset management of power distribution sys- tems via fuzzy AHP. Systems Journal, 4, 593–602. dGonçalves, C. D. F. and Dias, J. A. M. and Cruz-Machado, V. A. (2014). Decision Methodology for Maintenance KPI Selection: Based on ELECTRE I. Springer. Duffuaa, S. O., Ben-Daya, M., Al-Sultan, K. S. and Andijani, A. A. (2001). A generic conceptual simulation model for mainte- nance systems. Journal of Quality in Maintenance Engineering, 7, 207–219. Durán, O. (2011). Computer-aided maintenance management sys- tems selection based on a fuzzy AHP approach. Advances in En- gineering Software, 42, 821–829. e Costa, C. A .B., Carnero, M. C. and Oliveira, M. D. (2012). A multi-criteria model for auditing a Predictive Maintenance Pro- gramme. European Journal of Operational Research, 217,381– 393. Emovon, I., Norman, R. A. and Murphy, A. J.(2010). Hybrid MCDM based methodology for selecting the optimum mainte- nance strategy for ship machinery systems. Journal of Intelligent Manufacturing, ,1–13. Eslami, S., Sajadi, S. M. and Kashan, A. H. (2014). Selecting a preventive maintenance scheduling method by using simulation and multi criteria decision making. International Journal of Lo- gistics Systems and Management, 18, 250–269. Fallah-Fini, S., Triantis, K., Rahmandad, H. and Jesus, M. (2015). Measuring dynamic efficiency of highway maintenance opera- tions. Omega, 50,18–28. Fang, C.-C., & Huang, Y.-S. (2008). A Bayesian decision analysis in determining the optimal policy for pricing, production, and warranty of repairable products. Expert Systems with Applica- tions, 35, 1858–1872. Farhan, J. and Fwa, T. (2009). Pavement maintenance prioritiza- tion using analytic hierarchy process. Transportation Research Record: Journal of the Transportation Research Board, 12–24. Ferdousmakan, M., Vasili, M., Vasili, M., Tang, S.H. and Lim, N.T. (2014). Selection of Appropriate Risk-based Maintenance Strat- egy by Using Fuzzy Analytical Hierarchy Process. In 4rd Euro- pean Seminar on Computing, Pilsen, Czech Republic (pp. 77). Främling, K., Holmström, J., Loukkola, J., Nyman, J., and Kaustell, A. (2013). Sustainable PLM through Intelligent Products. Engi- neering Applications of Artificial Intelligence, 26, 789–799. Fouladgar, M.M., Yazdani-Chamzini, A., Lashgari, A., Zavadskas, E. K. and Turskis, Z. (2012). Maintenance strategy selection us- ing AHP and COPRAS under fuzzy environment. International journal of strategic property management, 16, 85–104. Främling, K., Kubler, S., and Buda, A. (2014). Universal Messag- ing Standards for the IoT from a Lifecycle Management Per- spective. IEEE Internet of Things Journal, 1, 319–327. Garcı́a-Cascales, M. S., and Lamata, M. T. (2009). Selection of a cleaning system for engine maintenance based on the ana- lytic hierarchy process. Computers & Industrial Engineering, 56, 1442–1451. Garmabaki, A. H. S., Ahmadi, A. and Ahmadi, M. (2016). Maintenance Optimization Using Multi-attribute Utility Theory. Springer. Gomez, A. and Carnero, M. C. (2011). Selection of a Computerised Maintenance Management System: a case study in a regional health service. Production Planning and Control, 22,426–436. Goossens, A. J. M. and Basten, R. J. I. (2015). Exploring main- tenance policy selection using the Analytic Hierarchy Process; an application for naval ships. Reliability Engineering & System Safety, 142, 31–41. Ha, S. H. and Krishnan, R. (2008). A hybrid approach to supplier selection for the maintenance of a competitive supply chain. Ex- pert Systems with Applications, 34, 1303–1311. Hankach, P. and Lepert, P. (2011). Multicriteria Decision Analysis for Prioritizing Road Network Maintenance Interventions. Inter- national Journal of Pavements, 10,. Hjalmarsson, L. and Odeck, J. (1996). Efficiency of trucks in road construction and maintenance: an evaluation with data envelop- ment analysis. Computers & operations research, 23,393–404. Hosseini Firouz, M. and Ghadimi, N. (2015). Optimal preven- tive maintenance policy for electric power distribution systems based on the fuzzy AHP methods. Complexity,, DOI 10.1002/c- plx.21668. Hwang, C. L., and Yoon, K. (1981). Multiple Attribute Decision- Making Methods and Applications. Springer Verlag, Berlin, Hei- delberg, New York. Ilangkumaran, M. and Kumanan, S. (2009). Selection of mainte- nance policy for textile industry using hybrid multi-criteria de- cision making approach. Journal of Manufacturing Technology Management, 20, 1009–1022. Ilangkumaran, M. and Kumanan, S. (2012). Application of hybrid VIKOR model in selection of maintenance strategy. Interna- tional Journal of Information Systems and Supply Chain Man- agement (IJISSCM), 5,59–81. Jarke, M. and Vassiliou, Y. (1997). Data Warehouse Quality: A Re- view of the DWQ Project. In Proceedings of the Conference on Information Quality (IQ1997), Cambridge, MA (pp. 299–313). Jeon, J., Kim, C. and Lee, H. (2011). Measuring efficiency of total productive maintenance (TPM): a three-stage data envelopment analysis (DEA) approach. Total Quality Management & Busi- ness Excellence, 22, 911–924. Jones-Farmer, L. A., Ezell, J. D. and Hazen, B. T. (2014). Applying control chart methods to enhance data quality. Technometrics, 56, 29–41. Kahn, B. K., Strong, D. M., and Wang, R. Y. (2002). Information quality benchmarks: product and service performance. Commu- nications of the ACM, 45, 184–192. Köksal, G., Batmaz, l., and Testik, M. C. (2011). A review of data mining applications for quality improvement in manufacturing industry. Expert systems with Applications, 38, 13448–13467. Krogstie, J., Lindland, O. I., and Sindre, G. (1995). Defining quality aspects for conceptual models. In Proceedings of the IFIP8.1 Working Conference on Information Systems Concepts: Towards a Consolidation of Views (ISCO), Marburg, Germany (pp. 216– 231). Kumar, G. and Maiti, J. (2012). Modeling risk based maintenance using fuzzy analytic network process. Expert Systems with Ap- plications, 39, 9946–9954. Kuo, T. C. and Wang, M. L. (2012). The optimisation of mainte- nance service levels to support the product service system. In- ternational Journal of Production Research, 50, 6691–6708. Labib, A. W., O’Connor, R. F. and Williams, G. B. (1998). An ef- fective maintenance system using the analytic hierarchy process. Integrated Manufacturing Systems, 9, 87–98. Levrat, E., and Iung, B. and Crespo Marquez, A. (2008). E- maintenance: review and conceptual framework. Production Planning & Control, 19, 408–429. Li, C., and Xu, M. and Guo, S. (2007). ELECTRE III based on ranking fuzzy numbers for deterministic and fuzzy maintenance strategy decision problems. IEEE. Li, J., Tao, F., Cheng, Y. and Zhao, L.. (2015). Big data in product lifecycle management. The International Journal of Advanced 21 Manufacturing Technology, 81, 667–684. Liu, J. and Yu, D. (2004). Evaluation of plant maintenance based on data envelopment analysis. Journal of Quality in Maintenance Engineering, 10,203–209. Liu, M. and Frangopol, D. M. (2006). Decision support system for bridge network maintenance planning. Springer. Liu, J., LU, X. and QU, C. (2012). A Priority Sorting Approach of Maintenance Task During Mission Based on ELECTRE TRI. Fire Control & Command Control, ,S1. Mardani, A., Jusoh, A. and Zavadskas, E. K. (2015). Fuzzy multi- ple criteria decision-making techniques and applications – Two decades review from 1994 to 2014. Expert systems with Appli- cations, 42, 4126–4148. Maurino, A. and Batini, C. (2009). Methodologies for data quality assessment and improvement. ACM Computing Surveys, 41, 1– 52. Moazami, D., Behbahani, H. and Muniandy, R. (2011). Pavement rehabilitation and maintenance prioritization of urban roads us- ing fuzzy logic. Expert Systems with Applications, 38, 12869– 12879. Mobley, R. K. (2002). An introduction to predictive maintenance. Butterworth-Heinemann. Monte, M. B. S. and others. (2015). A MCDM Model for Preventive Maintenance on Wells for Water Distribution. In IEEE Interna- tional Conference on QSystems, Man, and Cybernetics (SMC), 2015 (pp.268–272). Monte, M. B.S and de Almeida-Filho, A. T. (2016). A Multicrite- ria Approach Using MAUT to Assist the Maintenance of a Wa- ter Supply System Located in a Low-Income Community. Water Resources Management, , 1–14. Muchiri, P., Pintelon, L., Gelders, L. and Martin, H. (2011). De- velopment of maintenance function performance measurement framework and indicators. International Journal of Production Economics, 131, 295–302. Mumpower, J. L., Phillips, L. D., Renn, O. and Uppuluri, V. R. R. (2012). Expert Judgment and Expert Systems. Springer Science & Business Media. Nyström, B. and Söderholm, P. (2010). Selection of maintenance actions using the analytic hierarchy process (AHP): decision- making in railway infrastructure. Structure and Infrastructure Engineering, 6, 467–479. Ofner, M., Otto, B. and Österle, H. (2013). A Maturity Model for Enterprise Data Quality Management. Enterprise Modelling and Information Systems Architectures, 8, 4–24. Ouma, Y. O., Opudo, J. and Nyambenya, S.(2015). Comparison of Fuzzy AHP and Fuzzy TOPSIS for Road Pavement Maintenance Prioritization: Methodological Exposition and Case Study. Ad- vances in Civil Engineering, ,DOI 10.1155/2015/140189. Ozbek, M. E., de la Garza, J. M. and Triantis, K. (2010). Data and modeling issues faced during the efficiency measurement of road maintenance using data envelopment analysis. Journal of Infras- tructure Systems, 16, 21–30. Ozbek, M. E., de la Garza, J. M. and Triantis, K. (2010). Effi- ciency measurement of bridge maintenance using data envelop- ment analysis. Journal of Infrastructure Systems, 16, 31–39. Palma, J., de León Hijes, F. C. G., Martı́nez, M. C. and Cárceles, L. G. (2010). Scheduling of maintenance work: A constraint-based approach. Expert Systems with Applications, 37, 2963–2973. Peck, M. W., Scheraga, C. A. and Boisjoly, R. P. (1998). Assess- ing the relative efficiency of aircraft maintenance technologies: an application of data envelopment analysis. Transportation Re- search Part A: Policy and Practice, 32, 261–269. Pintelon, L. M. and Gelders, L.F. (1992). Maintenance management decision making. European journal of operational research, 58, 301–317. Pourjavad, E., Shirouyehzad, H. and Shahin, A. (2013). Selecting maintenance strategy in mining industry by analytic network process and TOPSIS. International Journal of Industrial and Systems Engineering, 15, 171–192. Pramod, V. R., Sampath, K., Devadasan, S.R., Jagathy Raj, V.P. and Moorthy, G. D. (2007). Multicriteria decision making in main- tenance quality function deployment through the analytical hier- archy process. International Journal of Industrial and Systems Engineering, 2, 454–478. Roll, Y., Golany, B. and Seroussy, D.(1989). Measuring the effi- ciency of maintenance units in the Israeli Air Force. European Journal of Operational Research, 43, 136–142. Rouse, P., Putterill, M. and Ryan, D. (2002). Integrated perfor- mance measurement design: insights from an application in air- craft maintenance. Management Accounting Research, 13, 229– 248. Saaty, T. L. (1980). The Analytic Hierarchy Process. New York: McGraw-Hill. Sampaio, S. F. M., Dong, C. and Sampaio, P. (2015). DQ 2 S – A framework for data quality-aware information management. Expert Systems with Applications, 42, 8304–8326. Shafiee, M. (2015). Maintenance strategy selection problem: an MCDM overview. Journal of Quality in Maintenance Engineer- ing, 21, 378–402. Shahin, A., Pourjavad, E. and Shirouyehzad, H. (2012). Selecting optimum maintenance strategy by analytic network process with a case study in the mining industry. International Journal of Pro- ductivity and Quality Management, 10, 464–483. Sheikhalishahi, M. (2014). An integrated simulation-data envelop- ment analysis approach for maintenance activities planning. In- ternational Journal of Computer Integrated Manufacturing, 27, 858–868. Shyjith, K., Ilangkumaran, M. and Kumanan, S. (2008). Multi- criteria decision-making approach to evaluate optimum mainte- nance strategy in textile industry. Journal of Quality in Mainte- nance Engineering, 14, 375–386. Simpson, L. (1996). Do decision makers know what they prefer?: MAVT and ELECTRE II. Journal of the Operational Research Society, 47, 919–929. Sophie, S.-Z., Thomas, A., Dominik, L., Ralf, H., Pierrick, B. and Javier, G.-S. (2014). Supreme sustainable predictive main- tenance for manufacturing equipment. In European Congress & Expo on Maintenance and Asset Management (EuroMainte- nance), Helsinki, Finland, (pp. 1–6). Sun, S. (2004). Assessing joint maintenance shops in the Taiwanese Army using data envelopment analysis. Journal of Operations Management, 22, 233–245. Taghipour, S., Banjevic, D. and Jardine, A. K. S. (2011). Prioriti- zation of medical equipment for maintenance decisions. Journal of the Operational Research Society, 62, 1666–1687. Tan, P.-N., Steinbach, M. and Kumar, V. (2006). Introduction to data mining. Addison-Wesley Longman Publishing Co., Inc. Tan, Z., Li, J., Wu, Z., Zheng, J. and He, W. (2011). An evalua- tion of maintenance strategy using risk based inspection. Safety science, 49, 852–860. Thor, J., Ding, S. and Kamaruddin, S. (2013). Comparison of multi criteria decision making methods from the maintenance alter- native selection perspective. The International Journal of Engi- neering and Science, 2, 27–34. Triantaphyllou, E., Kovalerchuk, B., Mann, L. and Knapp, G. M. (1997). Determining the most important criteria in maintenance decision making. Journal of Quality in Maintenance Engineer- ing, 3, 16–28. Trojan, F. and Morais, D. C. (2012). Using ELECTRE TRI to sup- port maintenance of water distribution networks. Pesquisa Op- eracional, 32, 423–442. Trojan, F. and Morais, D. C. (2012). Prioritising alternatives for maintenance of water distribution networks: a group decision approach. Water Sa, 38, 555–564. Umbrich, J., Neumaier, S. and Polleres, A. (2015). Quality assess- ment & evolution of Open Data portals. In 3rd International Conference on Future Internet of Things and Cloud (FiCloud), Roma, Italy (pp. 404–411). Van den Bergh, J., De Bruecker, P., Beliën, J., De Boeck, L. and De- meulemeester, E. (2013). A three-stage approach for aircraft line maintenance personnel rostering using MIP, discrete event sim- ulation and DEA. Expert Systems with Applications, 40, 2659– 22 2668. Van Horenbeek, A. and Pintelon, L. (2014). Development of a maintenance performance measurement frameworkusing the an- alytic network process (ANP) for maintenance performance in- dicator selection. Omega, 42, 33–46. Vujanović, D., Momčilović, V., Bojović, N. and Papić, V. (2012). Evaluation of vehicle fleet maintenance management indicators by application of DEMATEL and ANP. Expert Systems with Ap- plications, 39, 10552–10563. Waisberg, D. (September 2015). Data Analytics: A Matrix for Better Decision Making. https://www.thinkwithgoogle. com/articles/data-analysis-a-matrix-for-better- decision-making.html#utm_source=LinkedIn&utm_ medium=social&utm_campaign=Think Wakchaure, S. S. and Jha, K. N. (2011). Prioritization of bridges for maintenance planning using data envelopment analysis. Con- struction Management and Economics, 29, 957–968. Wang, R. Y., and Strong, D. M. (1996). Beyond accuracy: What data quality means to data consumers. Journal of management information systems, 12, 5–33. Wang, J., Fan, K. and Wang, W. (2010). Integration of fuzzy AHP and FPP with TOPSIS methodology for aeroengine health as- sessment. Expert Systems with Applications, 37, 8516–8526. Wang, L., Chu, J. and Wu, J. (2007). Selection of optimum main- tenance strategies based on a fuzzy analytic hierarchy process. International Journal of Production Economics, 107, 151–163. Wellsandt, S., Wuest, T., Hribernik, K., and Thoben, K.-D. (2015). Information Quality in PLM: A Product Design Perspective. In Advances in Production Management Systems: Innovative Production Management Towards Sustainable Growth (AMPS), Tokyo, Japan (pp. 515–523). Xu, L. and He, W. and Li, S. (2010). Internet of Things in indus- tries: A survey. IEEE Transactions on Industrial Informatics, 10, 2233–2243. Zaim, S., Turkyilmaz, A., Acar, M. F., Al-Turki, U. and Demirel, O. F. (2012). Maintenance strategy selection using AHP and ANP algorithms: a case study. Journal of Quality in Maintenance En- gineering, 18, 16–29. Zhangqiong, W. and Guozheng, S. (1999). A Study of Appraisal and Decision-making Support System for Maintenance Scheme of Metal Structure of Crane. Journal of Wuhan University, 5, 007. 23 https://www.thinkwithgoogle.com/articles/data-analysis-a-matrix-for-better-decision-making.html#utm_ source=LinkedIn&utm_medium=social&utm_campaign=Think https://www.thinkwithgoogle.com/articles/data-analysis-a-matrix-for-better-decision-making.html#utm_ source=LinkedIn&utm_medium=social&utm_campaign=Think https://www.thinkwithgoogle.com/articles/data-analysis-a-matrix-for-better-decision-making.html#utm_ source=LinkedIn&utm_medium=social&utm_campaign=Think https://www.thinkwithgoogle.com/articles/data-analysis-a-matrix-for-better-decision-making.html#utm_ source=LinkedIn&utm_medium=social&utm_campaign=Think Introduction Data quality in expert maintenance systems Expert maintenance systems Data quality frameworks Research Methodology: Data quality framework instantiation to MRQA purposes Krogstie framework concepts and definitions MCDM-based Krogstie framework instanciation AHP-based MRQA framework Pairwise comparison based preference measurement Pairwise comparisons as ratio measurement Alternative ranking using TOPSIS OEM use case Scenario 1: Equivalence between criteria Scenario 2: Cost-reduction action plan Conclusions, implications, limitations and future research Conclusions Implications Limitations Future research Acknowlegement Percentage of criteria considered in the Maintenance literature Penalty score selection Jaccard-based similarity measure Penalty score impact and selection 
work_torjayyuybdkzi4lzfzfope36m	----	Expert system for the optimisation of melt extruded net structures A. Rawal*1 and P. J. Davies2 Net structures were produced by replacing the static die (spinneret) with two concentric dies (consisting of slots) rotating in opposite directions in a melt extrusion process. A series of net structures and filaments were produced from a square die slot shape. Similarly, a series of filaments were produced from the corresponding static square die slot using a laboratory melt extruder. The effect of die rotation on the filament geometries was investigated. Finally, an expert system was developed based on empirical observations and theoretical modelling, which allows the filament cross-sectional area and its shape to be characterised with regard to the die slot dimensions and extrusion parameters. Moreover, the net geometry has also been predicted and visualised using the same expert system. Keywords: Die slot, Expert system, Filament, Land length, Melt extrusion, Net Introduction Polymers are the most rapidly growing materials in terms of use and innovations in processing technology, one of the main applications of polymers is in the production of net structures. These may be used for reinforcement, drainage, filtration or a combination of systems. The net structures are generally classified according to their production method. Punching tech- niques such as Tensar 1 are used for large-scale nets and grids and the production method consists of two or three stages. Here a polymer sheet has a regular series of holes punched out followed by orientation in either just the machine direction or both machine and cross directions producing the final grid. Weaving techniques, for example leno weave, may also be employed to produce the larger holes in the net structures. However, the simplest way of producing the net structure is by replacing the static die or a spinneret normally used for filament production by two concentric counter- rotating dies 2 in a melt extrusion process. In this process, polymer chips are fed to a hopper from which they pass into a single screw extruder. The single screw extruder screw has two tasks: first to melt the polymer and second to pump the molten polymer through the die head at a constant rate. The die head consists of a concentric disk and an annulus consisting of slots, rotating in opposite directions. When the slots in the disk and the annulus are offset from each other, the filaments of the net are formed. However, when the slots are adjacent, a joint is formed between the two filaments. Finally, the net structure is drawn under water and given some stretch and orientation by means of drawing rollers to produce the required net structure. This technique is mainly employed for the net structures required in geotechnical applications. This single-stage process can result in an increase in the production rate 3 and a reduction in power consumption, 4 and an improved quality of net can be achieved with the rotation of the die as the molecules of polymer align themselves in the direction of flow, thus enhancing the mechanical properties of net structure. In a melt extrusion process, the filament shape and dimensions are mainly dependent upon the correspond- ing die slot shape and dimensions. However, the pressure drop and flow rate of the polymer inside the die slot also characterises the filament geometry. In the case of the rotating die system, the filament shape is also influenced by the rotational speed of the die. Furthermore, in rotating die systems, product devel- opment has required an iterative process in the devel- opment of new dies. Generally, the die requires several modification steps or complete remanufacture before the required net structures are produced. This involves significant cost, time and labour. The development of a computer model defining the effect of die shape and rotation on the final filament properties will allow prediction of net structure from a given die slot and process parameters. Therefore, the work reported here is the development of an expert system for computing the filament geometries under a given set of process parameters. Moreover, the net geometry was also predicted and visualised using the same expert system. Experimental Aim The aim of this investigation was to develop a process model based on empirical observations and theoretical modelling, which will allow the filament cross-sectional 1Current address, CSIR, National Fibre, Textile and Clothing Centre, Gomery Avenue, Summerstrand, Port Elizabeth 6000, South Africa 2Centre for Materials Research and Innovations, Bolton Institute, Bolton, BL3 5AB, UK *Corresponding author, email arawal@csir.co.za � 2005 Institute of Materials, Minerals and Mining Published by Maney on behalf of the Institute Received 15 September 2004; accepted 30 March 2005 DOI 10.1179/174328905X48513 Plastics, Rubbers and Composites 2005 VOL 34 NO 2 47 area and its shape to be characterised with regard to the slot dimensions, extrusion parameters and polymer rheology. This has allowed the development of an expert system, which predicts the area and shape of the filament for square die-slot structures. Materials and methods For the extrusion, two polymers were used. The first was a low density polyethylene (LDPE) with a MFI of 0.7 g min 21 at 190uC and the second was a high density polyethylene (HDPE) with a MFI of 0.035 g min 21 at 190uC. Both were extruded with a carbon black masterbatch such that the final extruded polymer contained 1% carbon black by weight. Extrusion equipment For the extrusion of the net structures, an extrusion system was used at Tensar International, this is a single screw melt extruder with 80 mm diameter screw. A series of static die extrusions were also carried out within the authors’ own laboratories at Bolton Institute. These were produced using a melt extruder with a screw length to diameter ratio (L/D) of 21 : 1. Extrusion matrix The experimental matrices were defined using MaxSuite (software) based on Taguchi’s technique 5 using a two- level full factorial design. Table 1 shows the matrix for the net extrusions in which the five main process Table 1 Experimental design for rotating dies Sample number Die speed, rpm Extruder speed, rpm Take-up speed, m min21 Temperature, uC MFI, g min21 1 1.5 25.5 2.48 195 0.035 2 4.5 25.5 2.48 195 0.035 3 1.5 50.9 2.48 195 0.035 4 4.5 50.9 2.48 195 0.035 5 1.5 25.5 5.0 195 0.035 6 4.5 25.5 5.0 195 0.035 7 1.5 50.9 5.0 195 0.035 8 4.5 50.9 5.0 195 0.035 9 1.5 25.5 2.48 225 0.035 10 4.5 25.5 2.48 225 0.035 11 1.5 50.9 2.48 225 0.035 12 4.5 50.9 2.48 225 0.035 13 1.5 25.5 5.0 225 0.035 14 4.5 25.5 5.0 225 0.035 15 1.5 50.9 5.0 225 0.035 16 4.5 50.9 5.0 225 0.035 17 1.5 25.5 2.48 165 0.7 18 4.5 25.5 2.48 165 0.7 19 1.5 50.9 2.48 165 0.7 20 4.5 50.9 2.48 165 0.7 21 1.5 25.5 5.0 165 0.7 22 4.5 25.5 5.0 165 0.7 23 1.5 50.9 5.0 165 0.7 24 4.5 50.9 5.0 165 0.7 25 1.5 25.5 2.48 195 0.7 26 4.5 25.5 2.48 195 0.7 27 1.5 50.9 2.48 195 0.7 28 4.5 50.9 2.48 195 0.7 29 1.5 25.5 5.0 195 0.7 30 4.5 25.5 5.0 195 0.7 31 1.5 50.9 5.0 195 0.7 32 4.5 50.9 5.0 195 0.7 Table 2 Experimental design for static dies Sample number Extruder speed, rpm Take-up speed, m min21 Temperature, uC MFI, g min21 1 17.42 5 195 0.035 2 34.84 5 195 0.035 3 17.42 10 195 0.035 4 34.84 10 195 0.035 5 17.42 5 225 0.035 6 34.84 5 225 0.035 7 17.42 10 225 0.035 8 34.84 10 225 0.035 9 17.42 5 165 0.7 10 34.84 5 165 0.7 11 17.42 10 165 0.7 12 34.84 10 165 0.7 13 17.42 5 195 0.7 14 34.84 5 195 0.7 15 17.42 10 195 0.7 16 34.84 10 195 0.7 Rawal Expert system for the optimisation of melt extruded net structures 48 Plastics, Rubbers and Composites 2005 VOL 34 NO 2 parameters, i.e. melt flow index (MFI), temperature, die speed, extruder speed and take-up speed, are considered. For the static die extrusion, a series of sixteen samples were produced based on a static single-square slot die (2 mm deep and 2 mm wide) using four main process parameters, namely, MFI, temperature, extruder speed and take-up speed in order to investigate the effect of die rotation on the filament shape within the net structure. This experimental matrix is shown in Table 2. Analysis technique for area and shape of the filament In order to analyse the extruded samples, a methodology was required which would allow filament cross-sections to be obtained without damage. Previous work at Bolton 6 demonstrated the usefulness of embedding samples in resin to minimise sample damage during sectioning before analysis and a similar technique has been followed here. The resin used was Araldite CY212 which is a typical resin used in the sectioning of samples before microscopy. The plasticity of the resin was modified using dibutyl phthalate to allow a clear section to be obtained. Images of the sections were obtained using a low magnification optical microscope with a CCD camera attached. The signal from the CCD was fed into a computer and the resultant image saved as a bitmap. Analysis of the images was carried out using the University of Texas Health Science Center, San Antonio (UTHSCSA) Image Tool program 7 providing data for both the area and shape of the filaments, which could then be compared with the respective die slot dimensions and shape. The shape of the filaments was specifically defined using the in-built functions of Image Tool program such as background separation and filtration. These functions separate the background from the filament image and, apparently, identify the exact shape of the filament. The data for the area was analysed statistically using MaxSuite to provide empirical models for the area of the filaments. Visual observations were also made for the shape of the filaments. Results and discussion Geometry of net structures Within a melt extrusion net production, the die head consists of a disk and an annulus, which will usually be counter rotating. Both components have slots cut on the bearing surface and each slot results in the formation of a filament. Because of the rotation of the die, the slots in the disk and the annulus will meet and, at this point, the two molten polymer streams form a single stream thus forming a joint. The resultant structure will be a cylindrical net. The geometry of the net structure can be expressed by mesh angle and mesh count. Mesh angle is defined as the half of the angle formed between the filaments in the vertical direction and mesh count is defined as the number of joints formed within the net structure in the machine direction in a specified distance, as shown in Fig. 1. Consider a single filament emerging from a die of radius R with a rotational velocity w and a linear velocity or take-up speed v, which forms a helix with angle a (i.e. mesh angle) due to the rotational effect and the vectorial representation of a typical filament is shown in Fig. 2. From this representation, a is expressed as tan a~ wR v (1) Assuming that the inner and outer parts of the die are counter rotating with angular velocities (rpm) wi and wo, respectively, then the relative angular velocity wr of the inner part to the outer is wr~(wi{({wo))~(wizwo) (2) If there are ni and no slots on the inner and outer parts of the die, respectively, then, for a single slot on the inner die, the number of joints made per minute by slots on the outer die passing the slots on the inner die ji can be represented by ji~(wizwo)no (3) Similarly, for a single slot on the outer die, the number of joints made per minute by slots on the inner die passing the slots on the outer die jo can be represented by jo~(wizwo) ni (4) Therefore, mesh count in the machine direction (i.e. AB as shown in Fig. 2) per unit length jimd formed during each revolution of inner or outer die can be expressed by either jimd~(wizw0) no=(v sec a) (5) 1 Mesh angle and mesh count defined in a net structure 2 Velocity vector diagram Rawal Expert system for the optimisation of melt extruded net structures Plastics, Rubbers and Composites 2005 VOL 34 NO 2 49 or jimd~(wizw0)ni=(v sec a) (6) Analysis of cross-sectional area of the filaments Area measurements were carried out for both the rotating and static die extrusions. In the case of rotating dies, the filament sectioning was carried out at four distinct positions by dividing the circumference of the net into four equal parts. Furthermore, in order to examine the effect of direction of sectioning, the net structures were sectioned either parallel or perpendicular to the machine directions. The sectioning was carried out before the joint, after the joint and at the midpoint of the filaments between the joints. It was expected that the filament sections obtained before and after the joint should have greater cross-sectional area, as shown in Fig. 3a and b. Since the two streams of the polymer emerging from the die slot attempt to form a single stream during the process, they leave the trailing and leading edges before and after the joints, respectively. 8 The average area for the three views was found to be more uniform by cutting the cross-section parallel to the machine direction, as shown in Fig. 3a. Therefore, in the current research work, the sections of the filaments at the midpoint between the two joints were obtained at four positions in the machine direction. Use of multivariate linear regression has allowed a statistical analysis of the data to be carried out using MaxStat (a commercial software package). A compar- ison between the modelled area and experimental observations from static and rotating die slots are shown in Figs. 4 and 5, respectively, and gives excellent correlations. Therefore, in order to ensure that these results have real significance, the t-test was performed on cross-sectional area values. The t-test (as shown in Table 3) for area values has revealed that the correlation coefficients based upon theoretical models and experi- mental data are extremely high and found to be significant at the 0.1% level. Hence, the models can be used in developing the expert system. Moreover, Table 4 shows the comparison of percentage contribution of the 3 Filament section a parallel to the machine direction b perpendicular to the machine direction 4 Area model for static square die slot Table 3 Significance of correlation coefficient Property Area of the filament in rotating square die slot, mm2 Area of the filament in static single square die slot, mm2 Degree of freedom 30 14 Correlation coefficient 0.96 0.99 Expected t-value for 0.1% significance level 3.9 4.1 Calculated t-value 18.59 37.04 5 Area model for rotating square die slot Table 4 Comparison of percentage contribution of each process parameter in rotating and static square die slots Process parameter Rotating square die-slot, % Static square die-slot, % Die speed 0.7 – Extruder speed 35.8 32.0 Take-up speed 42 44.5 Temperature 0.2 0.0 MFI 6.7 17.1 Die speed and extruder speed 0.1 – Die speed and take-up speed 0.1 – Die speed and temperature 0.3 – Die speed and MFI 0.4 – Extruder speed and take-up speed 1.8 4.5 Extruder speed and temperature 0.4 0.0 Extruder speed and MFI 0.0 0.8 Take-up speed and temperature 2.1 0.0 Take-up speed and MFI 0.2 0.7 Temperature and MFI 0.0 0.1 Rawal Expert system for the optimisation of melt extruded net structures 50 Plastics, Rubbers and Composites 2005 VOL 34 NO 2 coefficients and the interaction coefficients to the area models. Here, it can be seen that the area is mainly dependent on two factors: extruder speed and take-up speed for both static and rotating dies. This may be expected as it can be assumed that the higher the extruder speed, the greater the throughput rate, resulting in an increase of the filament area. Similarly, increase in the take-up speed will result in greater attenuation of the filament, therefore, lowering the filament area. Whilst the MFI has a significant effect, as both the polymers have different swelling tendencies and the processing temperature and die speed has a negligible effect on the cross-sectional area of the filament. Effect of die rotation on filament shape For extrusion under static and rotating die conditions, the extruder was set up such that the polymer throughputs were matched. As the die slots for each had a similar dimension, the change in area should be similar. More importantly, the effect of die rotation on the filament shape may be compared by direct observa- tion. Fig. 6 shows the comparison of filament shapes produced from square die slots, using static and rotating dies whilst keeping all other parameters constant. This has revealed the effect of die rotation on the filament geometry. In general, the polymer emerging from the die slot tends to deviate from its corresponding die slot shape due to higher surface tension. However, the effect of surface tension can be modified by changing the shear rate of the polymer. For polyethylene, the apparent viscosity decreases by increasing the shear rate. 9 Therefore, filament approaching a circular shape can be produced from a square die slot by increasing the shear rate. Empirical observations have been made which show that these stresses vary depending on the speed of die rotation. Further theoretical modelling is underway to more clearly define this tangential stress effect and the results will be reported later. 10 Computer simulation of filament shapes in melt extruded net structures The empirical and theoretical models developed here have formed the basis of an expert system with the primary aim of defining an undrawn net structure from the process parameters. A computer model in the form of an expert system has been developed using Delphi 11 (a computer programming language). The basic form for data entry is shown in Fig. 7. It is intended that this will also provide the calculated data. In addition to providing the data in the form of expert system, it will also create an output text file. This text file is structured such that it can be read by a Virtual Reality Modelling Language (VRML) browser. From the text file created, a computer visualisation of the net structure may be observed and manipulated. These computer models are generated based upon analogies that have been used in previous models. 12 Here, a filament is created by sweeping a two-dimensional (2D) cross-sectional area along a three-dimensional (3D) path. 13 An example of the semi-circular cross-section net structure obtained from a square die slot is shown in Fig. 8. Testing of expert system The expert system developed was tested at the collabor- ating company. The eight randomly selected net structures, i.e. sample numbers 3, 4, 9, 12, 15, 19, 24 and 32 were produced from a square slot (1.78 mm wide, 1.78 mm deep and 5 mm land length), based upon the experimental design shown in Table 1. It was observed that the extent of circularity in the shape of the filament increases with the die rotational speed as shown in 6 Shape of the filament produced from a static square die slot b rotating square die slot 7 Data entry form for process parameters Rawal Expert system for the optimisation of melt extruded net structures Plastics, Rubbers and Composites 2005 VOL 34 NO 2 51 Fig. 9. Furthermore, the mesh count, mesh angle and the filament cross-sectional areas were also predicted and compared with the corresponding measured values. The predicted and measured values of filament cross- sectional area, mesh angle and mesh count are shown in Table 5. Furthermore, the t-test has shown that correlation coefficients based on expert system models and experimental data are significant at the 0.1% level as shown in Table 6. Thus, the expert system can be highly useful in predicting the filament and net geometries under a given set of process parameters, obtained from a square die slot. Conclusions A series of empirical models have been developed, which, in combination with theoretical geometrical data, allow the structure of an extruded net to be defined as a consequence of the process parameters. The empirical models for the cross-sectional area of the filaments have demonstrated that the structure is mainly dependent on two parameters: extruder speed and take-up speed. These empirical models of the filament cross-sectional area in combination with geometrical data have formed the basis of an expert system in the form of a computer model. The area models show good correlation with the measured area values. However, the shape of the filament in the net structure is complicated by the effect of the die rotation, resulting from a tangential shear on the polymer. It is accepted that controlling the filament cross-section in both area and shape will have a significant effect on the final properties of the net structure and is an important step in understanding the extrusion behaviour. 8 Semi-circular filament shape in a net structure 9 Effect of die rotation on the filament shape Table 5 Testing of an expert system Mesh angle, degrees Mesh count per 10 cm Filament area, mm2 Sample number Measured Predicted Measured Predicted Measured Predicted 3 5.2 5.51 3.33 3.25 3.55 3.97 4 15.1 16.15 10.5 9.4 2.65 3.12 9 5.5 5.51 3 3.25 1.89 1.54 12 15.2 16.15 11 9.4 3.72 3.12 15 3.1 2.74 1.8 1.6 1.28 1.57 19 5.3 5.51 3.33 3.25 4.78 4.28 24 8.2 8.17 6.3 4.8 1.97 2.30 32 7.5 8.17 6.5 4.8 2.45 2.29 Table 6 Significance of correlation coefficient between measured and modelled filament area (expert system) Property Filament area, mm2 Mesh angle, degrees Mesh count per 10 cm Degree of freedom 6 6 6 Correlation coefficient 0.93 0.99 0.99 Expected t-value for 0.1% significant level 5.21 5.21 5.21 Calculated t-value 6.19 17.19 17.19 Rawal Expert system for the optimisation of melt extruded net structures 52 Plastics, Rubbers and Composites 2005 VOL 34 NO 2 Acknowledgements The authors are highly thankful to Department of Trade & Industry (DTI) and Tensar International Ltd, U.K for their support provided to this research work. References 1. T. S. Ingold and K.S. Miller: ‘Geotextiles handbook’, 1988, London, Thomas Telford. 2. F. B. Mercer: ‘Improvements relating to the production of net or netlike fabrics by extrusion methods’, GB Patent No. 836,555, June 1, 1960. 3. S. Bilgin: ‘Extrusion of pipes using a rotating die system’, M.Sc. Thesis, University of Manchester Institute of Science & Technology, 1981. 4. N. V. Tyabin, V. C. Bortnikov, E. M. Tsentovskii and K. D. Vachagin: Mekh. Polim., 1968, 4, 531. 5. R. K. Roy: ‘A Primer on the Taguchi method’; 1990, New York, Van Nostrand Reinhold. 6. P. J. Davies and A. R. Horrocks: J. Coated Fabrics, 1999, 28(2), 118–138. 7. http://ddsdx.uthscsa.edu/dig/itdesc.html 8. A. Rawal: ‘Generation of an expert system for the optimisation of net extrusion processes’, 2002, Ph.D Thesis, Bolton Institute, Bolton, UK. 9. A. Ziabicki: ‘Physical fundamentals of the fibre spinning process’, (ed. H. F. Mark, S. M. Atlas and E. Cernia), Vol.1, 1967; New York, Interscience Publishers, John Wiley Sons, Inc. 10. A. Rawal and P. J. Davies: in preparation. 11. ‘Borland Delphi 3 Guide: Users Guide’; 1997, Scotts Valley, CA, Borland International Inc. 12. P. Potluri, A. Rawal, M. Rivaldi and I. Porat: Composites: Part A, 2003, 34, 481–492. 13. A. L. Ames: ‘VRML 2.0 Sourcebook’, 2nd edn, 1997; New York, John Wiley & Sons. Rawal Expert system for the optimisation of melt extruded net structures Plastics, Rubbers and Composites 2005 VOL 34 NO 2 53 
work_tpfrk7xqvngs7h5t7mqmquu5eu	----	Bank Credit Authorization Using Fuzzy Expert System Procedia Computer Science 102 ( 2016 ) 659 – 662 Available online at www.sciencedirect.com 1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Organizing Committee of ICAFS 2016 doi: 10.1016/j.procs.2016.09.458 ScienceDirect 12th International Conference on Application of Fuzzy Systems and Soft Computing, ICAFS 2016, 29-30 August 2016, Vienna, Austria Bank credit authorization using fuzzy expert system Mustafa Menekay* * Department of International Business, Near East University, P.O.BOX:99138, Nicosia, North Cyprus, Mersin 10, Turkey Abstract Companies Financial and economic performance is determined by analyzing their financial statements, that includes such important information as growth in sales, return to stockholders, profit margins, return on equity and credit risks. This information is useful for investors, creditors and other external users. To analyze financial condition of company and make decision need more professionalism and good economical and financial background from managers. In many cases financial manager make decision in condition of uncertainty. Economical instability, political situation needs many companies to work in condition of uncertainty, fuzziness of information. In this paper we take a closer look how information in the financial statement can be combined, analyzed and used by fuzzy expert systems to support many important financial decisions in credit analysis and authorization. This is achieved by the development of knowledge base using knowledge of experienced specialists, human experts in this field. In expert system knowledge base is combined with inference engine to make final decision. © 2016 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the Organizing Committee of ICAFS 2016. Keywords: Neural network; retina recognition; backpropagation algorithm;neural network. 1. Introduction Company’s financial measures are often used to rank corporate performance. Growth in sales, return to stockholders, profit margins, return on equity and credit risk are measures of economic and financial performance that can be determined by analyzing company’s financial statements. Financial statements include too much important information that is useful to investors, creditors and other external users. In today’s global economy, investments capital is always on the move. Through organized capital markets, investors each day shift billions of investment dollars among different companies, industries and nations. Capital flows to those areas in which investors expect to earn the greatest returns with least risk. * Mustafa Menekay.: +90-392-392-223-6464/378. E-mail address: mustafa.menekay@neu.edu.tr © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Organizing Committee of ICAFS 2016 http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2016.09.458&domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2016.09.458&domain=pdf 660 Mustafa Menekay / Procedia Computer Science 102 ( 2016 ) 659 – 662 Financial services industry was viewed as an area of high potential for artificial intelligence applications, especially expert systems. To the disillusionment of both the developer and user communities these early expectations have not been met. A few expert systems have been implemented for selected functions in banking, brokerage, and insurance environments. The financial services industry has been a vigorous user of expert system techniques. Advisory programs have been created to assist bankers in determining whether to make loans to businesses and individuals. Some of these research works about application of ESs to finance is given in3,4. One of important step in this study is bank credit analysis and authorization. In this paper using ES the bank credit authorization problem is considered. The benefits of using expert systems for credit analysis are speed and accuracy, both which far exceed human capacity. 2. Expert System for Credit Analysis There are number of financial institutions for credit analysis and authorization in the world. Decision making in this area need credit investigation, which allow exposing the main characteristics of the problem. Investigation includes the determining of the input factors and their influence to the final decision. These input factors are: experience, trustworthy and stability of clients, total assets, total liabilities, type of production, amount of equity, investment, state of economy, financial standing and detailed analysis of client’s financial information, capacity and ability to manage cash, number of business owned, credit history, business stability, collateral factors, etc. As is shown the credit analysis and authorization depend on many factors and it is characterized by risky and uncertain environment. Many of these factors are characterized linguistic interpretability, fuzzy values. In this condition, credit analyzer to make decision, must possess professionalism, wide knowledge about problem domain. For this reason to collect and combine experts knowledge in this field and create automated system that will make decision in credit authorization acquire actuality. To solve this problem the development of expert system for bank credit analysis and authorization problem is considered. An expert system is a problem solving and decision making system based on knowledge of it’s task and logical rules or procedures for using knowledge. Knowledge is typically represented as “if then“ rules that reside in the expert system’s knowledge base. The knowledge base is combined with a problem-solving mechanism or inference engine to reach conclusions, and explain this reasoning via a user interface. Employees is in organization demonstrate expertise through the skillful application of knowledge in a particular domain. Expertise differentiates one company’s products and services from another’s. Credit authorization is one of area in the financial services industry where expert systems are in use and providing benefit. In these areas it is the combination of computer technology and expertise that creates value for customers. Expert System accumulates the knowledge of multiple human experts that give system more breadth that single person is likely to achieve. This reduces risk of doing business. Bank credit analysis and authorization depend on many factors and most of decisions are making in risky and uncertain environment. Such as, many of the banking structures are operated in macro economic instability, political uncertainty. They are working with different kinds of clients. For this reason during development of the system it is necessary to take into account the factors related to economical and political condition and the factors related to the clients. In many cases the values of these factors are characterized by linguistic interpretability, uncertainty, fuzziness. Fuzziness of input information needs to apply fuzzy technology for information processing. Expert System main step in development is the development of knowledge base. In most cases knowledge base has if-then structure that describes the relation between premise and conclusion parts. KB includes analysis of enterprise with relevant benchmarks, measure of liquidity and credit risk, measure of profitability. As a result of analysis the main factors that are dominant in credit assignment have been determined. On the base of these factors the suitability of credit assignment is determined. Constructed KB for credit assignment has two level structures. In the first layer the repayment ability, ability to manage, and the risk of operation are determined. Then these parameters are used to determine the suitability of the credit assignment in second level. At the same time repayment, ability to manage, and credit risk depends on many factors. Repayment ability characterizes the capability of the borrower to pay the loan back. Ability to manage evaluates the managerial skills of the applicant. This is depend on experience, education, number of businesses managed previously. The credit risk is determined by 661 Mustafa Menekay / Procedia Computer Science 102 ( 2016 ) 659 – 662 business stability of client, number of missed credit payments, asset/debt ratio, bond rating. The inputs for determining these factors are total assets, total liabilities, type of production, amount of equity, investment, experience, last year profit, last year sales, number of business owned, credit history, asset/debt ratio, business stability, number of payment missed. If credit operation is done by Turkish lira, then one of important parameter in credit operation is dollar-Turkish lira ratio. The oscillation the value of this factor can change the level of risk, considerable. The values some of these factors are characterized by linguistic terms, fuzzy values. The fragment of KB for the first layer of credit assignment is given in fig. 1. IF TOTAL_ASSETS TWO TIMES IS MORE THAN CREDIT_REQUEST THEN REPAYMENT IS HIGH; IF TOTAL_ASSETS TWO TIMES IS LESS THAN CREDIT_REQUEST AND TOTAL_ASSETS ONE TIME IS MORE THAN CREDIT_REQUEST THEN REPAYMENT IS MEDIUM ELSE REPAYMENT IS LOW; IF CREDIT_REQUEST IS LESS THAN 18% OF EARNINGS THEN ABILITY_MANAGE IS YES IF BUSINESS_STABILITY IS HIGH AND CREDIT_MISSED IS LESS THAN TWO TIMES AND ASSET/DEBT_RATIO IS HIGH AND BOND_RATING IS HIGH AND DOLLAR_TL_RATIO IS NORMAL THEN RISK IS LOW Fig. 1. The fragment of KB for first layer of credit assignment. Here HIGH, MEDIUM, LOW are linguistic terms. In the work the form of the linguistic terms are taken in triangle form. EARNINGS=LAST_YEAR_PROFIT* LAST_YEAR_SALES/100 ; The fragment of KB for second layer of credit assignment is given in fig. 2. IF REPAYMENT IS HIGH OR REPAYMENT IS MEDIUM AND MANAGE IS YES AND RISK IS MEDIUM OR RISK IS LOW THEN CREDIT IS APPROVED; IF REPAYMENT IS HIGH OR REPAYMENT IS MEDIUM AND MANAGE IS YES AND RISK IS HIGH THEN CREDIT IS NOT_APPROVED; Fig. 2. The fragment of KB for second layer of credit assignment. As shown using the input factors of KB in the first layer the repayment, ability to manage and risk are determined. Using these in the second layer the approving of credit assignment is determined. Outputs of the rules describe how decision of credit manager. 662 Mustafa Menekay / Procedia Computer Science 102 ( 2016 ) 659 – 662 Using knowledge of experienced specialists the knowledge base for bank credit authorization is constructed. Realization of expert system for bank credit authorization is performed in the ES shell of ESPLAN. During consultation session the input data that describes current conditions of client is given to the ES input. Using these inputs ES make decision about credit authorization. Because of KB is constructed on the base of knowledge of many expert specialists the described approach allows more accurately make decision of credit authorization. CONCLUSION Company’s decision making in financial field is very complicated and ill-structured task. ES is one of the methodologies that can be adapted to this kind of problem and in this research it is used for credit authorization problem. In the developed ES using results of credit analysis the knowledge base is created. Many of input parameters in this KB are characterized by fuzzy values. For this reason the fuzzy technology is used to make decision. It is realized using fuzzy ES logic inference mechanism ESPLAN. References 1. Stuart JR, Peter N. Artificial Intelligence. A Modern Approach. Prentice-Hall International Inc; 1995. 2. Aliev RA, Shachnazarov MM, Gulko DE. Scheduling Expert Systems. News of Academy of Sciences of USSR Tech. Cybernetics 1988; 5: 118-128. 3. Heuer S, Koch U, Cryer C. Invest: An expert system for financial investment management Decision Support and Expert systems, Thomson information /Publishing Group; 1990: 176_187. 4. Allahverdi N, Tutuncu K. Design of ES in financial field. The third international conference on mathematical & computational applications. Konya: Turkey; September 4-6 2002. 
work_tqp5gog4djcedc2l5b34qfaawa	----	TEHNIKA KVALITET IMS, STANDARDIZACIJA I METROLOGIJA 18 (2018) 6 875 (EMEXS) Prikaz kompjuterske aplikacije BOJANA V. , Iritel a.d. Beograd UDC: 005:620.9]:004.4 DOI: 10.5937/tehnika1806875J perspektiva. Mali broj studija U ovom radu je EMEXS u formi kompjuterske aplikacije, zasnovan primenu informacionih i komunikacionih tehnologija. U radu je opisan EMEXS ekspertski sistem, merenje efekata primene ovog ekspertskog sistema u privredi. ekspertski sistem, kompjuterska aplikacija 1. UVOD plek- snim primene novih energetskih politika. Evropska unija je usvojila energetsku strategiju do 2030. godine, po od 27% i smanjenje emisije CO2 za 40% [1]. m, [2]. Jedno od najperspektivnijih sredstava za postizanje po- [3]. Pobo- ati ogroman uti- - [4]. J komponenti - [5], a prikaz jednog od njih, u formi ekspertskog sistema, dat je kroz ovaj rad. U ovom radu, dat je prikaz ekspertskog sistema za EMEXS (eng. Energy Management Expert System), koji organizacije mogu da koriste kako bi ocenile svoj nivo zrelosti za i primenile potrebna po- put 23 e-mail: bojana.jovanovic.123@gmail.com Rad primljen: 06.07.2018. 07.09.2018. . Za kreiranje eks- autora Jo- standarda ISO 50001, CMMI (eng. Capability Matu- rity Model Integration) nivoima zrelosti i PDCA (eng. Plan-Do-Check-Act) ciklusu. Postoji podatak da je do sada izdato ukupno 20.216 sertifikata ISO 50001 ( je 69% 2015. godinu) [7], pa je ovo . , EMEXS je u skladu sa jednim od glavnih [8], usmeren na Savremeni trendovi na- energije [9]. Neki od primera su real-time sistemi [10], sistemi za modelovanje i optimizaciju [11], sis- [12], - menom cloud computing -a [13], itd. Industrijske ko- - kasnost u njihovim proizvodnim postrojenjima [14]. DNV GL [15] web ka- tore. Kolokotsa i saradnici [16] su prikazali sistem zasnovan na web tehnologiji za zgrade univerzitetskog naselja. Faia i saradnici [17] energije u stambenim zgradama. Sucic i saradnici [18] EKSPERTSKI SISTEM ZA 876 TEHNIKA KVALITET IMS, STANDARDIZACIJA I METROLOGIJA 18 (2018) 6 na prikup - ljenje za energiju SAD-a [19] je razvilo online apli- standarda ISO 50001. Mnogi radovi razmatraju kako se - [3]. Mali broj studija raz- . 2. PRIKAZ EKSPERTSKOG SISTEMA ZA - EMEXS Ekspertski sistem prikazan u ovom radu ima za cilj da obuhvati sve oblasti koje razmatra standard ISO 50001, kako bi pomogao organizacijama da sertifikuju - . Oblast u eks- pertskom sistemu , a pre- uzete su iz modela zrelosti autora i [6]. Ekspertski sistem je strukturiran kroz nivoe zre- losti, trendovi. 2.1. EMM50001, kao osnova za EMEXS U ovom radu je prikazan ekspertski sistem za EMEXS. Model zrelosti autora izabran je kao osnova si- stema. Na osnovu strukture pomenutog modela zrelosti [6], 21 modul: 1) , 2) Pri- - , 3) Imenovanje energetskog mena- , 4) Definisanje energetske politike, 5) Energe- tsko planiranje, 6) Identifikacija i vrednovanje zakon- skih i drugih zahteva koji se odnose na energiju, 7) Energetsko preispitivanje, 8) Uspostavljanje energe- tske poredbene vrednosti, 9) Definisanje indikatora energetske performanse, - bnih energetskih ciljeva i akcionih planova, 11) Pri- mena energetskih planova, sle- , 13) Interna i eksterna komunikacija, 14) Dokumentacija i zapisi energetskog menad , energetsku performansu, 16) Energetski efikasno projektovanje i modifikovanje zgrada, opreme, sistema i procesa, 17) Energetski efikasna nabavka, 18) Pra- , 19) In- , 20) Pri- mena korektivnih i preventivnih mera koje se odnose na energiju, od strane rukovodstva. preuzeti iz pomenutog modela zrelosti [6], tako da se nivo zrelosti dobija odgovaranjem na pitanja u okviru svakog modula, na svakom od nivoa zrelosti. U skladu pet nivoa zrelosti: , 2) Upravljan, 3) Defi- nisan, 4) Kvantitativno upravljan, i 5 ) Optimizovan. 2.2. EMEXS struktura i logika Ekspertski sistem EMEXS se sastoji od tri - sobno povezane jedinice (slika 1): Jedinica modula Ovde se nalazi 21 modul, gde svaki, kroz grupu pitanja, - ocenjivanje trenutnog nivoa zrelosti energije. - tnik i sistem bodovanja, preuzet iz pomenutog modela zrelosti [6]. Kao rezultat, korisnik dobija informaciju o sveukupnom nivou zrelosti organizacije, ali i o nivou zrelosti po svakom modulu. , korisnik , , kao i da odredi jake strane . Jedinica preporuka Na osnovu rezultata samo- ocenjivanja , kori- , u od- nosu na nivo koji je identifikovan samoocenjivanjem. Preporuke su prikazane sveukupno i po modulima, ta izabere samo one koja smatra korisnim. , ako primeni sve preporuke koje dobije u odnosu na ocenjen trenutni nivo. iz [6]. Jedinica alata Za svaku od preporuka koje daje - - varijante me- nata, kao i obrasce koji se mogu primeniti. Postoji sistem. Slika 1 Struktura ekspertskog sistema EMEXS TEHNIKA KVALITET IMS, STANDARDIZACIJA I METROLOGIJA 18 (2018) 6 877 2.3. Alat za samoocenjivanje, kao osnova za izradu preporuka Kada korisnik pokrene aplikaciju, odmah ulazi u upitnik za samoocenjivanje. Upitnik za samoocenji- vanje se sastoji od 21 pitanja. Za svako pitanje postoji pet odgovora, a korisnik bira samo jedan od odgovora. Svaki odgovor, zapravo, predstavlja nivo zrelosti. Odgovori koji se koriste u sistemu su osnova za kreiranje preporuka koje organizacija treba da primeni - rgije za jedan nivo. Slika 2 Primer dela upitnika za samoocenjivanje - navigacioni meni. 2.4. EMEXS Na osnovu rezultata samoocenjivanja, sistem omo- nivo. Preporuke su date sveobuhvatno i po modulima. Slika 3 Primer preporuka po modulima Slika 4 Primer sveobuhvatne preporuke Slika 5 zrelosti 2.5. alata ovog ekspertskog sistema. Na osnovu pretrage koncepti, . primeri iz prakse kako metodu. Gotovi obrasci da koristi u svojim procesima i praksama. Predlog strukture dokumenata sugestije kako . EKSPERTSKI SISTEM ZA 878 TEHNIKA KVALITET IMS, STANDARDIZACIJA I METROLOGIJA 18 (2018) 6 Slika 6 Slika 7 Primer alata s Slika 8 Primer alata sa gotovim obrascem Slika 9 Primer alata sa predlogom strukture doku- menta Alati su prikazani sveukupno, kao i po modulima, tako da korisnik ima pregled svih alata koji mu se nude, ali i onih koji se odnose na konkretni modul. (tabela 1). Vrste alata po oblas- tima Literatura Certification Europe [Internet]. 2017 [citirano 22.06.2018]. Dostupno na: http://certificationeurope.com/benefits-of-iso- 50001/ Leonardo Energy [Internet]. 2017 [citirano 11.12.2017]. Do- stupno na: https://help.leonardo-energy.org/hc/en-us/articles/- 201941411-What-are-the-benefits-of-energy-management- and-ISO-50001- BSI Group [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: https://www.bsigroup.com/en-GB/iso-50001-energy- management/ Office of energy efficiency & renewable energy [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: https://www.energy.gov/- ISO50001 Western Michigan University [Internet]. 2012 [citirano 11.12.2017]. Dostupno na: https://wmich.edu/mfe/mrc/gre- enmanufacturing/pdf/Posters/ISO50001%20Poster%- 20May%202012%20(1).pdf Quality Systems Enhancement [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://www.enhancequality.com/- iso-standards/iso-50001-energy-management-system/ ISO Quality Services [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: https://www.isoqsltd.com/iso-certification/iso- 50001-energy-management-certification/ Caffal C. Energy management in industry, Centre for the Analysis and Dissemination of Demonstrated Energy Tec- hnologies, Analysis Series, Vol. 17.C, 1995. Kannan R, Boie W. Energy management practices in SME-case study of a bakery in Germany, Energy Conversion and Man- agement, Vol. 44, No. 6, pp. 945-959, 2003. Tam KW, Leung CW, Probert SD. Energy management in a dairy-products plant. Applied Energy, Vol. 32, No. 2, pp. 83- 100, 1989. Network for Business Sustainability [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: https://nbs.net/p/executive-report- measuring-amp-valuing-environmental-381ab81b-0c4e-4dfe- be03-5638dbe73308 Unilever [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: https://www.unilever.com/sustainable-living/reducing- environmental-impact/ IRIS [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: https://iris.thegiin.org/metric/4.0/OD4108 National Science Foundation (NSF) [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: https://www.nsf.gov/- bfa/dias/policy/papp/pappg17_1/environimpacts_checklist.pdf UNESCO [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://www.unesco.org/new/en/culture/themes/underwater- cultural-heritage/protection/threats/environmental-impact-and- climate-change/ Manpower Group [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://www.manpower.com.hk/environmental- impact.aspx World Health Organization (WHO) [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://www.who.int/globalchange/environment/ecosystem_ass essment_large.jpg?ua=1 ISO 50001:2011. Energy Management Systems - Requirements with Guidance for Use. Schneider Electric. Energy Efficiency Projects Ensure Healthy Financial Performance for Healthcare Facilities, 2010. TEHNIKA KVALITET IMS, STANDARDIZACIJA I METROLOGIJA 18 (2018) 6 879 Vrste alata po oblas- tima Literatura ISO 50001:2011. Energy Management Systems - Requirements with Guidance for Use. Association of Energy Engineers [Internet]. 2015 [citirano 19.11.2015]. Dostupno na: https://www.aeecenter.org/i4a/pages/index.cfm?pageID=3351 Ministry of Mining and Energy of the Republic of Serbia [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://www.mre.gov.rs/latinica/sistem-energetskog- menadzmenta.php ELearning Industry [Internet]. 2014 [citirano 11.12.2017]. Dostupno na: https://elearningindustry.com/3-ways-measure- training-effectiveness Tremblay DG. Towards the learning organization: collective knowledge development in the multimedia sector. Direction de la recherche, Télé-université, Université du Québec. 2003. University of South Florida [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://www.usf.edu/hr/ Zendesk [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: https://support.zendesk.com/hc/en-us/articles/203664076-Best- practices-Developing-content-for-your-knowledge-base Harmsworth S, Turpin S. Creating an effective dissemination strategy. TQEF National Co-ordination Team. Vol. 5, 2000. WISER project [Internet]. 2012 [citirano 11.12.2017]. Dostupno na: http://www.wiser.eu/programme/internal-and- external-information-sharing/ Business Academy [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: https://www.biznis-akademija.com/motivacija- ucinak-i-nagradjivanje-zaposlenih Moj Posao [Internet]. 2006 [citirano 11.12.2017]. Dostupno na: https://www.moj-posao.net/Savjet/60808/Motivacija-i- nagradjivanje-zaposlenika/6/ MCB blog [Internet]. 2012[citirano 11.12.2017]. Dostupno na: http://www.mcb.rs/blog/kako-motivisati-zaposlene/ T. Nematerijalne nagrade i njihov uticaj na motivaciju zaposlenih. EMC review- , Vol. 3, No. 1, 2012. S G. Atraktivnost stimulacija , 2011. University of Belgrade [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://www.bg.ac.rs/files/sr/studije/studije- uni/Uvod_statisticke_metode_istrazivanja.pdf University of Novi Sad, Technical Faculty "Mihajlo Pupin" Zrenjanin [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://www.tfzr.uns.ac.rs/Content/files/0/Nedelja%203.pdf European IPPC Bureau [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://eippcb.jrc.ec.europa.eu/reference/ Nancy R. , Second Edition, ASQ Quality Press, pp. 96-99, 2004. Andersen B, Fagerhaug T, Beltz M. Root Cause Analysis and Improvement in the Healthcare Sector: A Step-by-Step Guide, Milwaukee, WI: ASQ Quality Press, pp. 124-125, 2010. Bialek R, Duffy GL, Moran JW. The Public Health Quality Improvement Handbook, Milwaukee, WI: ASQ Quality Press, pp. 168 170, 2009. Parnaby J, Towill DR, Seamless healthcare delivery systems, International Journal of Health Care Quality Assurance, Vol. 21, No. 3, pp. 249-273, 2008. Pallant J. -a, Prevod 4. izdanja, Mikro knjiga, Beograd, 2011. U.S. Environmental Protection Agency, Health care guide to pollution prevention implementation through environmental management systems, U.S. Environmental Protection Agency, Cincinnati, Ohio, 2005. ISO 50004:2014 Energy management systems - Guidance for the implementation, maintenance and improvement of an energy management system. Schneider Electric. ISO 50001: Recommendations for Compliance, Schneider Electric, 2012. Vrste alata po oblas- tima Literatura Hystead M. Healthcare Marketing Fix: Internal Communications [Internet]. 2011 [citirano 05.07.2013]. Dostupno na: http://www.htkmarketing.com/healthcare- marketing-fix-internal-communications J M. , Fakultet organizacionih nauka, Beograd, 2010. EN 15224:2012 Health care services - Quality management systems - Requirements based on EN ISO 9001:2008. IWA 1:2005 Quality Management Systems - Guidelines for process improvements in health service organizations. ISO 1828:2012 Health informatics - Categorial structure for terminological systems of surgical procedures. ISO 10159:2011 Health informatics - Messages and commu- nication - Web access reference manifest. Smallbusiness Chron [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://smallbusiness.chron.com/competencybased- performance-reviews-11504.html Apple [Internet]. 2016 [citirano 11.12.2017]. Dostupno na: https://images.apple.com/environment/pdf/Apple_Environmen tal_Responsibility_Report_2016.pdf International Civil Aviation Organization [Internet]. 2016 [citi- rano 11.12.2017]. Dostupno na: https://www.icao.int/enviro- nmental-protection/- Documents/ICAO%20Environmental%20Report%202016.pdf Boeing [Internet]. 2016 [citirano 11.12.2017]. Dostupno na: http://www.boeing.com/resources/boeingdotcom/principles/en vironment/pdf/2016_environment_report.pdf Hungarian Central Statistical Office [Internet]. 2013 [citirano 11.12.2017]. Dostupno na: https://www.ksh.hu/docs/eng/xftp/- idoszaki/ekornyhelyzetkep13.pdf TUI Group [Internet]. 2014 [citirano 01.12.2017]. Dostupno na: https://www.tuigroup.com/damfiles/default/tuigroup- 15/en/sustainability/Reporting/tui_cruises_environmental_repo rt_2014-2.pdf-b42e43124890a28fc88c74e968ee47a9.pdf Ikea [Internet]. 2016 [citirano 11.12.2017]. Dostupno na: http://www.ikea.com/ms/en_US/img/ad_content/IKEA_Group _Sustainability_Report_FY16.pdf Samsung [Internet]. 2016 [citirano 11.12.2017]. Dostupno na: http://www.samsung.com/us/aboutsamsung/sustainability/sust ainabilityreports/download/2016/2016-samsung-sustainability- report-eng.pdf ConocoPhillips [Internet]. 2015 [citirano 11.12.2017]. Dostupno na: http://www.conocophillips.com/sustainable- development/SD% 20Report%20Download/16-0310%20SD%20Book.pdf Canon [Internet]. 2016 [citirano 11.12.2017]. Dostupno na: http://www.canon.com/csr/report/pdf/canon-sus-2016-e.pdf Banaeian N, Mobli H, Nielsen IE, Omid M. Criteria definition and approaches in green supplier selection a case study for raw material and packaging of food industry, Production & Manufacturing Research, Vol. 3, No. 1, pp. 149-168, 2015. AIIM [Internet]. 2017 [citirano 11.12.2017]. Dostupno na: http://www.aiim.org/What-Is-Document-Imaging# DNN Extension [Internet]. 2017 [citirano 23.01.2018]. Dostupno na: http://dnnextension.com/Services/DNN- Document-Management-System V G. Akreditovane laboratorije, Fakultet organizacionih nauka, Beograd, 2010. Kolka JW. ISO 9001 and Health Care, Quality Progress, 1999. EKSPERTSKI SISTEM ZA 880 TEHNIKA KVALITET IMS, STANDARDIZACIJA I METROLOGIJA 18 (2018) 6 Vrste alata po oblas- tima Literatura ISO 50015:2014 Energy management systems - Measurement and verification of energy performance of organizations - General principles and guidance. ISO 50001:2011 Energy Management Systems - Requirements with Guidance for Use. Tehnis [Internet]. 2017 [citirano 23.01.2018]. Dostupno na: http://www.tehnis.privreda.gov.rs/sw4i/download/files/cms/att ach?id=429 Galankashi MR, Chegeni A, Soleimanynanadegany A, Memari A, Anjomshoae A, Helmi SA, Dargi A. Prioritizing green supplier selection criteria using fuzzy analytical network process, Procedia CIRP, Vol. 26, pp. 689-694, 2015. Gurel O, Acar AZ, Onden I, Gumus I. Determinants of the green supplier selection, Procedia - Social and Behavioral Sciences, Vol. 181, pp. 131-139, 2015. 2.6. Kreiranje EMEXS-a u formi kompjuterske aplikacije EMEXS Java programskog jezika i JavaFX razvojnog okru- interfejsa. Java programskog jezika je bio obe- funkcionisanja aplikacije na - formama. Aplikacija koristi odvojene niti za pokre- tanje upitnika i obradu odgovora, projektni - [20]. - - , minimizira se vre- me nakon datog odgovora na poslednje pitanje potrebno obraditi samo poslednji odgovor, dok su svi prethodni odgovori . Svi podaci faj- lovima, ikovanje pitanja, alata i preporuka bez menjanja same aplikacije. Da bi . 3. I SMERNICE ZA DALJI RAD - ment energije EMEXS. Prednost EMEXS- [6], koji se zasniva na standardu ISO 50001, CMMI nivo- ima zrelosti i PDCA ciklusu. mo- , pa je samim tim pod kontro- lom korisnika. Jedinstveni je alat koji kombinuje sa- moocenjivanje, - . Ekspertski si- , nu- , u skladu sa zahtevima standarda ISO 50001 i nivoima zrelosti [6]. EMEXS-a od - vedeno da bi se prikazalo da je EMEXS univerzalno biti dodati novi alati za koje se utvrdi da su neophodni. Plan je da se testiranje sprovede u dve faze. U prvoj fazi, ekspertski eventualno sistem izmenio i uskladio sa preporukama EMEXS - za koji vremenski period, - mena EMEXS- , itd. U dru EMEXS-a. 4. ZAHVALNICE Ministarstva prosvete, nauke og razvoja, pod oznakom III 43008 Razvoj metoda, . - LITERATURA [1] Azuma H, Magnani S, Compared analysis of the economic and environmental benefits by using an energy management system in different European countries, Energy Procedia, Vol. 126, pp. 266-273, 2017. [2] International Energy Agency. World Energy Outlook 2012 - Executive Summary [Internet]. 2012 [citirano 15.05.2015]. Dostupno na: www.iea.org/publica- tions/freepublications/publication/English.pdf [3] Schulze M, Nehler H, Ottosson M, Thollander P, Energy management in industry a systematic review of previous findings and an integrative conceptual framework, Journal of cleaner production, Vol. 112, pp. 3692-3708, 2016. [4] Johansson MT, Thollander P. A review of barriers to and driving forces for improved energy efficiency in Swedish industry recommendations for successful in-house energy management, Renewable & Sustai- nable Energy Reviews, Vol. 82, pp. 618-628, 2018. [5] Finnerty N, Sterling R, Coakley D, Contreras S, Coffey R, Keane MM. Development of a Global Ene- rgy Management System for non-energy intensive multi-site industrial organisations: A methodology, Energy, Vol. 136, pp. 16-31, 2017. [6] B J, ISO 50001 standard-based energy management maturity model proposal and validation in industry. Journal of cleaner production, Vol. 112, pp. 2744-2755, 2016. TEHNIKA KVALITET IMS, STANDARDIZACIJA I METROLOGIJA 18 (2018) 6 881 [7] ISO Survey 2016. [Internet]. 2016 [citirano 12.02.2018]. Dostupno na: https://www.iso.org/the- iso-survey.html [8] May G, Stahl B, Taisch M, Kiritsis D. Energy mana- gement in manufacturing: From literature review to a conceptual framework. Journal of cleaner producti- on, Vol. 167, pp. 1464-1489, 2017. [9] Pusnik M, Al-Mansour F, Sucic B, Gubina AF. Gap analysis of industrial energy management systems in Slovenia, Energy, Vol. 108, pp. 41-49, 2016. [10] Lajevardi B, Haapala KR, Junker JF. Real-time mo- nitoring and evaluation of energy efficiency and ther- mal management of data centers, Journal of Manu- facturing Systems, Vol. 37, pp. 511-516, 2015. [11] Meike D, Pellicciari M, Berselli G. Energy efficient use of multirobot production lines in the automotive industry: Detailed system modeling and optimiza- tion, EEE Transactions on Automation Science and Engineering, Vol. 11, No. 3, pp. 798-809, 2014. [12] Porzio G. F, Fornai B, Amato A, Matarese N, Van- nucci M, Chiappelli L, Colla V. Reducing the energy consumption and CO2 emissions of energy intensive industries through decision support systems an exa- mple of application to the steel industry, Applied Energy, Vol. 112, pp. 818-833, 2013. [13] Ren L, Zhang L, An efficient it energy-saving ap- proach based on cloud computing for networked gre- en manufacturing, Advanced Materials Research, Vol. 139, pp. 1374-1377, 2010. [14] Bunse K, Vodicka M, Schönsleben P, Brülhart M, Ernst FO. Integrating energy efficiency performance in production management gap analysis between industrial needs and scientific literature. Journal of cleaner production, Vol. 19, No. 6-7, pp. 667-679, 2011. [15] DNV G. L, Lumina - performance benchmarking [Internet]. 2017 [citirano 23.01.2018]. Dostupno na: https://www.dnvgl.com/assurance/lumina/index.htm l [16] Kolokotsa D, Gobakis K, Papantoniou S, Georgatou C, Kampelis N, Kalaitzakis K, ... Santamouris M. Development of a web based energy management system for University Campuses: The CAMP-IT pla- tform, Energy and Buildings, Vol. 123, pp. 119-135, 2016. [17] Faia R, Pinto T, Abrishambaf O, Fernandes F, Vale Z, Corchado J. M, Case based reasoning with expert system and swarm intelligence to determine energy reduction in buildings energy management, Energy and Buildings, Vol. 155, pp. 269-281, 2017. [18] Sucic B, Al-Mansour F, Pusnik M, Vuk T, Context sensitive production planning and energy mana- gement approach in energy intensive industries, Ene- rgy. Vol. 108, pp. 63-73, 2016. [19] U. S. Department of Energy, The 50001 Ready Navi- gator [Internet]. [citirano 04.06.2018]. Dostupno na: https://navigator.industrialenergytools.com/ [20] Cornell University - Department of Computer Sci- ence [Internet]. 2018 [citirano 16.05.2018]. Dostu- pno na: https://www.cs.cornell.edu/courses/cs- 3110/2010fa/lectures/lec18.html SUMMARY ENERGY MANAGEMENT EXPERT SYSTEM (EMEXS) PROPOSAL OF COMPUTER APPLICATION Today's organizations face complex expectations for energy savings. Many papers address aspects of how energy efficiency can be improved, but dominated by a technical perspective. The minority of studies picked up organizational aspect of energy management. This paper proposes an energy management expert system EMEXS in form of a computer application, based on the maturity model known in the literature, as one specific energy strategy model with innovative focus, based on contemporary trends which suggest the implementation of information and communication technology. The paper describes EMEXS system and its implementation challenges and opportunities. Also, the paper proposes some directions for future testing and measuring effects of applying. Key words: energy management, expert system, energy management software, computer application 
work_tsqqejrexfc3jaosibziha4c5i	----	Microsoft Word - MEC1 - powder4.doc JKAU: Eng. Sci., vol. 14 no. 2, pp. 87-99 (1423 A.H./ 2002 A.D.) 87 Towards an Expert System for Powder Production Processes : (I) the Atomization Process MAHIR H. ES-SAHEB and ASHRAF RADWAN Mechanical Engineering Department, King Saud University P.O. Box 800, Riyadh 11421, Saudi Arabia E-mail : essaheb@ksu.edu.sa ABSTRACT. The practical definition for the expert system, as a system simulating the expert methods and knowledge in dealing with or manipulating special tasks is presented. The application of such a system in powder technology is attempted. The main features of the proposed system design are documented. It constitutes, actually, another part of an effort towards building an expert system for powder technology. It deals with the parameters of each powder production process, such as atomization, in relation to certain required powder’s characteristics. It extracts, and finds out the effect of each individual process parameter on the powder characteristics, and models and formulates these relations. The relational model, and the related system software modules are designed. The Structured Query Language (SQL) is used as a standard database management language, to build up the major block programs of the system software modules. The suggested expert system is meant to be flexible, easy to implement, modify and extend. The powder production by the atomization process is given as an example of how one can control the process parameters to get certain specific powder characteristics. KEYWORDS: Powder technology, manufacturing systems, process selection, expert systems, software design. 1. Introduction The powder technology field is one of the most suitable fields for expert systems implementation. The area contains a huge amount of variables, information and data, which are mostly empirical. Analytical analysis is rare and extremely difficult to obtain due to the complex nature of this field. Mahir H. Es-Saheb and Ashraf Radwan 88 The area of powder production processes is considered to be one of many important parts in the field of powder technology and constitutes a major part of its 'science'. In the literature, vast information is available on the numerous powder production processes [1-4]. However, most of the research activities, in this area, tend to concentrate on the individual process parameters with the main intention of producing successfully a powder for a specific application[5,6]. Consequently, a systematic investigation of the effect of the different process parameters and their inter-relations on the process as a whole and on the final powder characteristics are limited and scattered. Therefore, each company has its own practice in achieving these goals. This is actually an 'art', which depends heavily on the individual experience (i.e. of the technologist) and huge experimental trials. Furthermore, each production process has its own parameters, and its environmental factors. Not only that, some powder characteristics are interdependent on each other for the same powder type[1-3, 6, 7]. This, therefore, makes the relations among powder characteristics and the production process parameters very complicated and extremely difficult to obtain. Hence, most relational parameters available are empirical. Also, it is clear that this field involves high experience content. Thus, this field is an ideal candidate for expert systems. Also, as stated above, each powder producer or company has its own individual expertise and knowledge. These experiences and knowledge should be documented (i.e. knowledge based) and manipulated, and aided by an expert system, so that, the work can be carried out independent of the expert. Thus, the experts may be utilized for research and development work, rather than supervising and carrying out daily routine activities. Though the work in this paper is mainly concerned with software design, as will be shown in the next sections, it presents a basic portion of an expert system. In fact, it constitutes the first step in the realization of any system (i.e. the design step). This step, is usually followed by the design implementation step needed towards building the expert system. Usually, this can be carried-out by the software engineer, where he can employ any programming language and any suitable computer system to construct the needed package or utilizing an expert shell. Thus at this stage, for benefit completion, it is important to give a brief account of the expert system. Basically, an “expert system” is a computer program that represents and reasons with knowledge of some specialist subject with a view to solving problems or giving advice[8]. It consists, mainly, of a knowledge base and the inference engine. Also, it contains in addition a few sub-systems, namely: (i) User interface is the vehicle through which the user views and interacts with the system. (ii) Developer interface is the vehicle Towards an Expert System for Powder Production….. 89 through which the knowledge engineer develops the system. (iii) Explanation facility provides explanations on the reasoning of the system. (iv) System interface links the expert system to external programs such as databases, spreadsheets, algorithms, tables, empirical relations, etc., which work in a support role for the system. An expert system may completely fulfill a function that normally requires human expertise, or it may play the role of an assistant to a human decision maker. The term knowledge-based system is sometimes used as synonym for ‘expert system’ although, strictly speaking the former is more general. However, the process of constructing an expert system is often called knowledge engineering, and is considered to be ‘applied artificial intelligence’[9- 13]. In recent years, expert systems witnessed intense research activities and are widely applied in many fields and industries[14,15]. Consequently, the work described here falls in this category and presents a contribution in this direction where the application of expert system in powder technology field is attempted. The work presented in this paper tends to address these issues in relation to the atomization process. It constitutes another part of an effort towards building a complete expert system for powder technology, described by Radwan and Es- Saheb[16]. The first part, proposed by the authors[17], has dealt with the materials selection, and processes selection in general with respect to the powder characteristics. Meanwhile, the work presented here deals with the parameters of the process in relation to certain required powder characteristics. The suggested expert system is meant to be flexible, easy to implement, modify and extend. The main objectives, however, are to extract and find out the effect of each of the production process parameters on the powder characteristics, and to model and formulate these relations. Meanwhile, the secondary objective is to design and formulate the rules, which may be applied to the selected process. 2. System Structure The production of powder by the atomization process is one of the most versatile and widely used processes[4, 6, 7]. Thus, the production of powder by the atomization process has been chosen in this work as a typical example to illustrate how to relate available process parameters to the powder characteristics and how it is implemented in our suggested expert system. The system assumed that, each powder characteristics-process parameters relation is documented in a table. Thus, each table is believed to be a knowledge base, contains the individual expertise knowledge, or intensive experimental work, which relates each powder characteristics and the proper process parameters. Mahir H. Es-Saheb and Ashraf Radwan 90 As reported in the literature[3, 6, 18], the powder characteristics can be classified into primary and secondary powder characteristics. The primary characteristics are mostly process dependent. However, in the atomization process in particular and in the other powder production processes generally, the shape characteristics which is considered as process inherited characteristics, are excluded from the powder characteristics-process parameters relation, as shown in Fig. 1. Meanwhile, the relation between the primary and secondary characteristics is displayed in Fig. 2. Based on this, the relational model following and the software modules for the system are designed. The details of these are presented in the following sections. Now, since the atomization process is selected to demonstrate the implementation of our suggested 'expert system', it is important to state that, each process of powder production has its own parameters. Thus, the atomization process, particularly, has many parameters related to; (i) The melt, such as the melt temperature, pressure and speed; (ii) The nozzle arrangements and conditions, such as, the nozzle cross-section area, length, friction, number of nozzles and the geometric arrangements; (iii) The jet parameters, such as temperature, pressure and flow. (iv) The cooling media used, such as, water, air, steam, argon, and nitrogen, etc. (v) In addition to several environmental factors. These are displayed in Fig. 1. FIG. 1. The relational process parameters affecting the primary characteristics. Towards an Expert System for Powder Production….. 91 However, as stated above there are some inherited process characteristics for each production process, such as the particle shapes in the case of atomization process. This process is known to produce only rounded particle (i.e. drop like) shapes of various sizes. These particle sizes are heavily affected by the melt, nozzle and jet conditions and characteristics. Meanwhile, the other powder characteristics related to the surface particle properties such as the amount of surface oxide, the absorbed gas layer and gas content are affected by these factors as well as the cooling media and some environmental conditions. These parameters can be characterized as high or low. Ranges can be set for high and low values according to experience or experimental work. The proposed system is flexible enough, so that it can accommodate any modification or updating activities. In recent works by Radwan and Es-Saheb[16, 17], they showed how the powder production process can be selected. Now, the next task in the completion of the system would be the selection of the production process parameters which affect the related powder characteristics. Again for the atomization process; The shape is considered to be constant, while the particle size may be controlled by the melt, nozzle and jet parameters. For example, the melt parameters, which contribute to the size variations, and may be considered are: the temperature of the melt, the pressure and the speed (i.e. the flow). Furthermore, for the nozzle we may consider the nozzle's cross-section area, length, friction condition, geometric arrangement and the number of nozzles. Also, for the jet parameters we may consider the jet's temperature, pressure and its flow type. Therefore, these three major parameters; the melt, nozzle and jet FIG. 2. The relationship between primary and secondary characteristics. Mahir H. Es-Saheb and Ashraf Radwan 92 may be considered as composite attributes[19] to the size of the particles. Consequently, these may constitute a relation with the following headings: {PCHAR, MELT, NOZZLE, JET} where: PCHAR : powder characteristics MELT : composite attribute made of these headings; {MTEMP, MPRES, MSPEED} MTEMP : melt temperature; MPRES : melt pressure; MSPEED: melt speed. NOZZLE : the composite attribute of the nozzle parameters, which includes the following headings; {NAREA, NLEN, NFRIC, NNUM, NGEOM} NAREA : nozzle cross-sectional area; NLEN : nozzle length; NFRIC: nozzle friction conditions; NNUM: number of nozzles; NGEOM : the geometric arrangement of nozzles. JET : the jet composite attributes, which in turn has the following headings; {JTEM, JPRES, JFLOW} JTEM : jet temperature; JPRES: jet pressure; JFLOW : the jet flow condition (laminar, turbulent, mixed, fast, slow, etc.). This relation which relates the particle size to the melt, nozzle and jet parameters may be called MNJ (Melt-Nozzle-Jet) relation. As mentioned above, the other powder characteristics involved are the amount of surface oxide, adsorbed gas layer and gas content. These characteristics are mainly affected by the process cooling media and other environmental conditions. The cooling media may be water, air, steam, or any noble gas (e.g. argon, nitrogen, etc.). However, as far as the environmental conditions are concerned, there are several factors contributing in this aspect. For example, the furnace lining and structure, the crucible material and geometry (open/close type), etc., as seen in Fig. 1. Thus, the cooling media, and the environmental factors attributes, which are related to the powder characteristics, (such as the amount of surface oxide, adsorbed gas layer, moisture, and gas content), all may be included in a relation named CMEF (Cooling-Media-Environmental-Factors). This relation will include the following headings: Towards an Expert System for Powder Production….. 93 {PCHAR, CM, EF} where: CM : composite attribute for the cooling media parameters, and consists of: {WATER, AIR, STEAM, ARGON, NITROGEN,...} EF : composite attribute for the concerned environmental factors, such as: {OXLIN, ALKLIN, OPENFS, CLOSFS, OPENCR, CLOSCR,...} where: OXLIN : acidic furnace lining; ALKLIN : alkaline furnace lining; OPENFS : open furnace structure; CLOSFS : closed furnace structure; OPENCR : open crucible; CLOSCR : closed crucible; ………………………….. The MNJ and CMEF relations can be joined together as a natural join[20] due to the common attribute (i.e. heading) PCHAR. The result, thus, is a relation called MNJCMEF. As far as the other powder characteristics, such as the specific surface area, apparent density, tap density, friction conditions, and flow, etc., are concerned, it is clear that, all of these properties are mainly related to the fundamental and primary particle size and shape characteristics. In addition, the powder flow and friction conditions are also related to the adsorbed gas layer or/and the amount of surface oxide (i.e. the particle surface conditions). The higher these parameters are, the higher is the friction condition and consequently, the lower is the flow of the powder. Again, the rules for θ - restriction[20] may be applied to extract the proper parameters, which satisfy certain characteristics required or recommended from the user side. These rules have been handled previously by the authors[17]. Thus, it is known by now that, the "θ" stands for any chosen simple scalar comparison operator (=, ≠, ≥, ≤ etc.). This can be seen if one considers the relation MNJCMEF which displays the merits of this operation. Therefore, to illustrate this for a given particle size, for example, there is only one recommended value, or Null for each attribute presented in the relation such that: MNJCMEF WHERE PCHAR θ Pchar. 1 AND PCHAR θ Pchar. 2 AND ------------------------------------ PCHAR θ Pchar. n ; Mahir H. Es-Saheb and Ashraf Radwan 94 - Null for unrelated parameter. While, the attributes of MNJ relation have integer values, the attributes of CMEF relation have binary [1/0] or boolean [true/false] values. So that, the final result will be a table or a relation with tuples t, which include certain required characteristics and their related parameters. 3. System Software Modules It is well known that expert systems are usually built using Rule-based programming languages like Prolog, or LISP. However, our ultimate objective, as stated above, is to achieve such a complete system for powder technology. But, due to the complex nature of this problem, a complete expert system will be very huge, diversified, and difficult to achieve. Thus the strategy adopted, as suggested by the authors in earlier works[16,17], is to divide the system into parts (i.e. sub expert systems), later to be integrated to form a complete expert system. Therefore, in this work a part of this system (i.e. sub system) is presented, which deals with the powder production processes. It concentrates, as a first step, on the software design and the related modules. Obviously, at this stage of design and for these type of relational systems described above, the Structured Query Language (SQL) is usually used to build up the system's software modules. Particularly, it is known also to be a suitable language for searching databases for retrievable information. It is known that, according to SQL/92[21,22], a relation is represented as a table and each attribute type (i.e. heading) is presented as a domain[23], which contains all the attributes values. Thus, for each attribute a domain is created carrying the name of its attribute. For example, the syntax for creating a domain for the temperature of the melt (MTEMP) attribute is: CREATE DOMAIN MTEMP INTEGER (4); which is a simple integer data type. However, the powder characteristics domain may include letters and numbers, this may be created as: CREATE DOMAIN PCHAR CHAR (10); Also, as mentioned above, the domains of the cooling media and environmental factors are binary or boolean. Therefore, these may be formulated as: CREATE DOMAIN WATER BOOLEAN; The domains of the other attributes can be created using the same SQL syntax. Thus, when all the domains are declared a base table can be created to hold and relate all these domains to a primary domain such as PCHAR domain. Also, there is a base table for each process to relate its parameters to the Towards an Expert System for Powder Production….. 95 concerned powder characteristics. Therefore, for example the base table of the atomization process can be created using SQL/92 as: CREATE TABLE MNJCMEF (PCHAR, MELT, NOZZLE, JET, CM, EF), PRIMARY KEY (PCHAR); As noticed earlier the parametric attributes are presented as composite attributes. This is done only for the sake of simplifying the presentation, otherwise it is not accepted by SQL syntax. Hence, the domains which are defined or declared in the previous section would be presented instead. Also, the values of the primary key PCHAR are not null by default. On the other side, null values are accepted by the other domains. Thus, for example; the melt, nozzle and jet attributes may have values null against the other attributes of the powder characteristics such as; the amount of surface oxide, adsorbed gas layer and gas content, and vice versa. The user interface with the MNJCMEF base table may be thought of as a snapshot table[19, 24] named USINPUT. This includes two domains; the PCHAR domain and a user input domain called INUSER2. The INUSER2 is in fact a binary or a boolean domain, which is filled by the user, for each powder characteristic value. It is assumed that, the PCHAR attributes are displayed on the screen, and the user has to give 1 for a true value, otherwise null. The creation of the snapshot is as follows: CREATE SNAPSHOT USINPUT AS (PCHAR, INUSER2), REFRESH EVERY 10 MINUTES; When joining together the two tables USINPUT and MNJCMEF, on the condition that the matching is considered under the true values for INUSER2, the result will be by a view table as a window showing the restricted true tuples only, from the MNJCMEF table. Hence, this will high light the related parameters for the required specific powder characteristics. Finally, as an example for a data manipulation operation, a retrieval operation to extract some parameters to control the atomization process, in order to get certain powder characteristics concerning the particle size, amount of surface oxide, and its gas content, is given. This expressed by SQL, will be as follows: SELECT MNJCMEF.* FROM MNJCMEF Mahir H. Es-Saheb and Ashraf Radwan 96 WHERE MNJCMEF. PCHAR = ------------- particle size AND MNJCMEF. PCHAR = ------------- amount of surface oxide AND MNJCMEF. PCHAR = ------------- gas content; As mentioned above, the PCHAR domain is the domain within the MNJCMEF relation, which holds the values of the powder characteristics which are affected by the related parameters of the powder production process. Also, the SELECT here will reveal full details of all the parameters which satisfy the required powder characteristics. As stated above the expert system shells can be used. In fact, these shells are intended to allow non-programmers to take advantage of the efforts of programmers who have solved a problem similar to their own[12]. It is clear that all shells are not suited to all tasks. Unfortunately, it is difficult to be rigorous in one’s recommendations concerning what shell should be used for what problem. This is because we do not have very clear ideas concerning how the broad range of expert system tasks should be classified. Though the modern shells are more flexible than earlier shells, the general problem of selecting the appropriate expert system tools is still an important issue and causes some uncertainty. However, Hayes-Roth, et al.[13] proposed some very general issues to consider when selecting an expert system building tool. High level languages give programmers a fast prototyping tool, so that more flexible designs can be explored and evaluated at relatively low cost in terms of time and effort. The user interface will typically not be as ’friendly’ as that provided by a shell, but a fluent programmer will nonetheless be able to make rapid progress. Production rule languages, object-oriented programming languages, and procedural deduction systems usually provide the expert system builder with a few more degrees of freedom than a shell with regard to the details of such things as the specification of control and the handling of uncertainty[8]. These issues become essential in constructing the final intended expert system for powder technology. However, for the level of design involved and presented at this stage of the work, the designed system has to be, simple, tested and verified. This, in fact, is conducted by implementing the proposed design utilizing ‘Visual-Basic’ programming language. This is an important step before going any further in building the final system. Thus, for this purpose, a complete computer program package (using Visual-Basic) is developed, constructed and tested. The preliminary results obtained from this package confirm the success of the software design presented. The discussion of the full Towards an Expert System for Powder Production….. 97 results, the system flexibility, capability and extension as well as the description of the program including the flow chart, data-base creation, modifications and some display aspects together with other future features and the general future extensions of the work, such as the inclusion of, empirical formulas, characteristic charts, statistical particle size distribution, other powder production processes, the automation of knowledge acquisition, the possible use of expert shells, etc. will be the subjects of future publications. 4. Conclusion The practical definition for the expert system, as a system simulating the expert methods and knowledge in dealing with or manipulating special tasks is presented. In this regard, first, the relation between the primary powder characteristics and the powder production process parameters is defined. Thus, the relational model and software modules are designed. The Structured Query Language (SQL/92) is implemented in constructing the system software modules. This language (i.e. SQL) displayed good capability, and flexibility in this type of applications. Therefore, these concepts are realized in the system so that it is possible to find out the powder production process’s parameters, which satisfy certain powder characteristics of size, shape, and surface conditions. This suggested system is capable of providing the user with the required parameters in order to control the process and achieve the desired powder characteristics, similar to a human expert. The designed system is successfully tested using a computer package developed in Visual-Basic. The preliminary results obtained, thus far, are very positive and encouraging. However, more work and effort is still needed in order to achieve the final complete expert system for powder technology. References [1] Kipphut, C.M and German, R.M., Powder Selection for Shape Relation in Powder Injection Molding, Int. J. of Powder Metallurgy, 27, 2, 117-124, (1991). [2] Goetzel, C.G., Treatise on Powder Metallurgy, I., Inter-Science Publs. Inc., New York, (1949). [3] Hausner, H. and Kumar Mal, M. (editors)., Handbook of Powder Metallurgy, 2nd Edition, Chemical Publishing Co. Inc., New York, (1982). [4] Lawley, A., Powder Atomization Processes and Fundamentals, Int. J. of Powder Met. and Powder Techn., 13, 2, (1977). [5] Es-Saheb, M.H., The Cold Compaction of Polyvinyl Chloride (PVC) Powders, J. King Saud Univ.-Eng. Sci. 1, 6, Riyadh, 99-113, (1994). [6] Dowson, G., Powder Metallurgy: The Process and its Products, Adam Hilger, New York, USA, (1990). [7] Alting, L., Manufacturing Engineering Processes, 2nd Ed., Marcel Dekker, Inc., New York, USA, 281-300, (1994). Mahir H. Es-Saheb and Ashraf Radwan 98 [8] Jackson, P., Introduction To Expert Systems, 3rd edition, Addison-Wesley Longman Limited, Harlow, England, (1999). [9] Waterman, D.A., A Guide to Expert Systems, Reading, Addison-Wesley, (1986). [10] Giarratano, J. and Riley, G., Expert Systems: Principles and Programming, 2nd edn., Boston, MA: PWS Publishing, (1994). [11] Stefik, M., Introduction To Knowledge Systems, San Francisco, CA: Morgan Kaufmann, (1995). [12] van Melle, W.J., System Aids in Constructing Consultation Programs, Ann Arbor, MI: UMI Research Press, (1981). [13] Hayes-Roth, F., Waterman, D.A. and Lenat, D., Building Expert Systems, Reading, MA: Addison-Wesley, (1983). [14] Quinlan, J.R., Applications of Expert Systems, Sydney, Addison-Wesley, (1987). [15] Harmon, P. and Sawyer, B., Creating Expert Systems for Business and Industry, New York, Wiley, (1990). [16] Radwan, A. and Es-Saheb, M.H., Knowledge Based Powder Selection System, Proc. of the Internl. Conf. Advances in Materials and Processing Technologies AMPT'95, Dublin, August 8-12, 839-847, (1995). [17] Radwan, A. and Es-Saheb, M.H., Computerized Relational Models for Powder Technology: Powder and Process Selection, J. Intel. Manufacturing JIM, 9, (1998). [18] Es-Saheb, M.H. and Cogun, C., A Statistical Investigation of Selected Loose Cast Iron Powder Properties, Proc. of the 4th Saudi Eng. Conf., Development of Technical and Industrial Base in Saudi Arabia, King Abdul Aziz University, Jeddah, Nov. 5-8, Saudi Arabia, 217-225, (1995). [19] Date, C.J., An Introduction to Database Systems, Addison Wesley, (1995). [20] Mittra, S.S., Principles of Relational Database Systems, Prentice Hall Int. Inc., (1991). [21] Cannon, S. and Otten, G., SQL - The Standard Handbook, Maidenhead, UK: McGraw Hill Int., (1993). [22] Date, C.J. and Darwen, H., A Guide to the SQL Standard, 3rd Edition, Reading, Mass., Addison Wesley, (1993). [23] Date, C.J., What Is a Domain?, In C.J. Date, Relational Database Writings 1985-1989, Reading, Mass., Addison Wesley, (1990). [24] Codd, E.F., The Relational Model for Database Management, Version 2, Reading, Mass., Addison Wesley, (1990). Towards an Expert System for Powder Production….. 99 :ساليب إنتاج المساحيق نظام خبرة ألنحو أسلوب التذرير) 1( أشرف حافظ رضوان. د ، ماهر حمدي الصاحب. د الرياض - جامعة الملك سعود - الهندسة الميكانيكية قسم التعريف العملي لنظام الخبرة ، كنظام يحاكي تقديم تم . صلخستالم وتمت . تناول المهام الخاصة أو التعاملطرق الخبرة والمعرفة في حيث تم توثيق ، المساحيق تطبيق هذا النظام في مجال تقنية ةمحاول ا جزًءالواقعوالذي يمثل في المالمح الرئيسة لتصميم النظام المقترح آخر ضمن الجهود الرامية إلى بناء نظام خبرة متكامل لتقنية إنتاج كل مسحوق أسلوب هذا النظام مع عوامل ويتعامل. المساحيق وبالتالي . للمسحوق بذاتها خصائصبهدف الحصول على على حدة عوامل من تأثير كل عامل وإيجادن النظام يقوم باستخالص إف المسحوق ، ومن ثم يقوم بنمذجة بمفرده على خصائصاألسلوب وقد تم تصميم نموذج العالقات ووحدات . العالقاتوصياغة هذه قياسية كلغة" سكويل " لغةاستخدمتوقد . برامج النظام ذات العالقة البرامج الرئيسة في قطاعاتإلدارة قواعد البيانات ، في بناء المقترح يكون نظام الخبرة أنوقد روعي . وحدات برامج النظام ولتوضيح . ، وسهل التطبيق ، والتعديل وكذلك يمكن توسعته امرنً إنتاج المساحيق أسلوب المقدم ، فقد استخدم الخبرةعمل نظام التحكم في عوامل األسلوب كيفيةريقة التذرير كمثال لبيان بط .خصائص معينة بذاتها ي للحصول على مسحوق ذ 
work_ttlctb3rwvgmhijg7efaku343i	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 221016381 Params is empty 221016381 exception Params is empty 2021/04/06-03:06:05 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 221016381 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:06:05 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_tu23zpv7fjcfnaq36nzqczn22i	----	PII: 0888-613X(87)90008-9 A Reduction Methodology for a Differential Diagnosis Expert System Akram Salah D e p a r t m e n t o f C o m p u t e r S c i e n c e , C a i r o U n i v e r s i t y Kevin D. Reilly D e p a r t m e n t s o f C o m p u t e r a n d I n f o r m a t i o n S c i e n c e s a n d B i o s t a t i s t i c s a n d B i o m a t h e m a t i c s , U n i v e r s i t y o f A l a b a m a a t B i r m i n g h a m A B S T R A C T The simple production rule representation is generalized by adding p r o g r a m s to a management s y s t e m that manipulate rules in a rule-based system. B y adapting this methodology, a single generalized rule can represent a group o f simple ones. Then programs are e m p l o y e d to satisfy the general rule in a partial way while recursively reducing a decision p r o b l e m into smaller ones o f the s a m e nature until a decision is made. It is s h o w n that the reduction m e t h o d is m o r e efficient than the simple rule approach and that it m i n i m i z e s the n u m b e r o f rules used to express a problem. The concept o f using a m a n a g e m e n t program to manipulate a set o f rules is emphasized through solving a p r o b l e m in a differential diagnosis expert system. A comparison between the n u m b e r o f rules e m p l o y e d to express a p r o b l e m is made to s h o w advantages o f the reduction m e t h o d o l o g y over the simple rule representation. K E Y W O R D S : production systems, expert systems, reduction algorithm, prolog, decision tables, relational databases, medical diagnosis I N T R O D U C T I O N Much recent research focuses on computer systems that facilitate methodolo- gies simulating experts' knowledge-based decision-making strategies (see, for example, [1-4]). The most popular current way to create such systems is to incorporate large amounts of "domain-dependent" knowledge acquired from experts. Because experts often express their decision-making processes in sets of Address correspondence to Kevin D. Reilly, Departments ~ o f Computer and Information Sciences and Biostatistics and Biomathematics, University o f Alabama at Birmingham, Birmingham, Alabama 35294. International Journal of Approximate Reasoning 1987; 1 : 131 - 139 © 1987 Elsevier Science Publishing C o . , Inc. 52 Vanderbilt A v e . , N e w Y o r k , N Y 10017 0 8 8 8 - 6 1 3 X / 8 7 / $ 3 . 5 0 131 CORE Metadata, citation and similar papers at core.ac.uk Provided by Elsevier - Publisher Connector https://core.ac.uk/display/81989388?utm_source=pdf&utm_medium=banner&utm_campaign=pdf-decoration-v1 132 Akram Salah and Kevin D. Reilly if-then rules, rule bases are devised ( H a y e s - R o t h [5]). Such systems are usually called knowledge-intensive rule-based systems. A knowledge-intensive rule-based system consists mainly o f a set o f rules describing different decision situations in the p r o b l e m under question, together with actions to be taken in each case. A rule in such a system is represented as a structure typically in the f o r m A l & A 2 & • • • & A n - ' C Such a rule is interpreted as follows: if Ax and A 2 and . . . and A n are true simultaneously, then consequently C is true. The left side o f a rule contains a conjunction o f atoms called c o n d i t i o n s and the right side is called a c o n c l u s i o n o r a c o n s e q u e n c e (Salah and Y a n g [4]). I f an expert expresses his o r her decision-making process as a collection o f simple if-tben rules, each o f them can be represented directly as stated above. A p r o b l e m arises if an expert expresses rules in a less explicit w a y o r in s o m e f o r m such as a function o v e r a set o f rules. T h e n it is the responsibility o f the rule acquirer, w h e t h e r p e r s o n o r machine, to decide upon a representation o r to p r o v i d e s o m e kind o f control on processing such rules. In this article we s h o w an a p p r o a c h that can facilitate a generalization o f simple rule representation. This article is part o f a larger study that has b o r r o w e d concepts f r o m relational database systems (RDBSs), such that rules are stored in a rule base and then a m a n a g e m e n t s y s t e m retrieves the rule u n d e r question. T h e system exploits a n u m b e r o f features studied previously (Reilly et al. [6], Yang [7], Reilly et al. [8]), w h e r e key c o n c e r n s have been incorporation into a P r o l o g f r a m e w o r k (Kowalski [9], C l o c k s i n and Mellish [10]) o f k n o w l e d g e representa- tions and R D B S s ( B r u y n o o g h e [11]). It is shown that the a p p r o a c h increases the efficiency o f a rule-based system. THE PROBLEM T h e p r o b l e m arises in a differential diagnosis expert s y s t e m where conditions are either s y m p t o m s , observations, o r test results gathered by a physician to be used to derive a conclusion, which in this case is a disease o r a class o f diseases. Rules used to derive such a conclusion would typically be in the f o r m O l & O 2 & . • . & O n - ' D where each O i f o r 1 _< i _< n is an observation, and D is a disease o r class o f diseases. ( F r o m here on, we refer to any a t o m on the left side o f a rule as an observation (an observation can be a test result o r a s y m p t o m ) and on the right side as a disease. Thus, this rule is read as follows: if all observations 1 t h r o u g h n exist simultaneously in a patient, then this case can be diagnosed as D . ) O u r p r o b l e m arises when a g r o u p o f rules is expressed within a single if-then Reduction Algorithm in Expert Diagnosis 133 statement. Our particular concern is this: given a set o f n observations and a conclusion D such that if any k-subset o f observations o f n holds, k _< n, then D can be concluded. That is, any k observations o f {O1 & 0 2 & . . . & O , } - - , D A simple example o f such a case is the c o m m o n cold, where there are about 12 observations and any 3 o f them (existing simultaneously) establish the diagnosis. Examples o f a similar nature occur in rheumatic diseases (more discussion is provided below). Actually, this is a generalization o f a rule application. The special case in which k = n defines the " n o r m a l " rule structure, that is, the case in which all the conditions have to be satisfied to derive the consequence. A more formal view o f this problem is as follows. In a rule system there is a set o f conditions for each decision situation, each condition having a domain o f values. The left side o f any rule represents an element in the Cartesian product o f the domains o f these conditions. The case here m a y be conceptualized as having one condition with one domain o f observations, say, O with length n, such that i f any k-subset o f O with k < n occurs simultaneously, then the diagnosis is established. This expresses a set o f rules, each one having a condition part c E O k, where k < n, and the same consequence D , which is the disease under consideration. T o represent this situation within an expert system, we examine two alternatives to set the stage for subsequent comparison. 1. The single if-then statement is re-expressed as a set o f simple rules. Each such simple r u l e c o n t a i n s k observations on its left side and D on its right side. Needless to say, the resulting number o f rules consumes much m e m o r y space, complicates the search when the system is applied, and reduces the efficiency o f the system. 2. The production system is extended such that if the " a n y k out o f n " formulation is expressed, it can be handled automatically. Note that this problem differs f r o m those representations o f " i n e x a c t r e a s o n i n g " or uncertainty (Prade [12], Rosenbloom et al. [13]) in which subsets o f conditions are used to derive a c o n s e q u e n c e - - f o r instance, probabilistically, fuzzily, or using weighting schemes. R E D U C T I O N M E T H O D The reduction method is based on viewing rule-base systems as a set o f rules together with p r o g r a m s that manage such rules. This view enables us to add program code to the management system such that it can extend the simple representation o f production rules. Here, we apply this methodology to enable a direct representation o f the generalized f o n n discussed above. We denote a p r o b l e m as " a k / n diagnostic p r o b l e m " when the diagnosis is 134 Akram Salah and Kevin D. Reilly dependent on n observations such that if any k o f them are found to exist in a case, then the diagnosis is established. F o r example, the case discussed by Weiss and Kulikowski ([14], p. 119 ff.) in diagnosis o f rheumatic diseases such as mixed connective tissue disease ( M C T D ) involves 10 observations. I f any 4 o f these 10 exist in a patient, then he or she definitely has a rheumatic disease. Using our terminology, we say that this is a 4/10 diagnostic problem. A Reduction Algorithm T o solve a k/n diagnostic problem using the reduction methodology, we p e r f o r m the following: 1. Pick any k symptoms. 2. Name them temporarily T1 . . . . . Tk. 3. Check decision table (k) with results for the k symptoms. 4. The output o f the decision table is tS. 5. The problem now is tS/R, where R = n - k. 6. For any 6/R diagnostic problem: (a) if 6 = 0 diagnosis is POSITIVE; terminate. (b) if 6 > R diagnosis is N E G A T I V E ; terminate. (c) i f 6 _< R go to step 1 (with k = 6 , n = R ) for further reduction. The set o f tables in Table 1 depicts the situation in a simplified form to make it easier to focus on the steps o f the reduction method. In realistic cases, actions may involve reports back to the user on the rules that are fired, auxiliary calculations (for instance, o f a statistical nature), or other options. In such cases, a table action portion would include additional information along with the number o f remaining tests that are depicted in this set o f tables. It should be noted that the use o f tables to describe the algorithm does not necessarily imply that implementation by tables is mandated. I f tables are used in the implementa- tion, they need not always be stored; that is, there are cases in which they can be generated. An Example To illustrate the method, we use a specific example o f a 4/7 diagnostic problem. To solve the diagnostic problem: 1. Pick any 4 observations. 2. N a m e them temporarily T 1 , T 2 , T 3 , and T4. 3. Check the first table in Table 1 with the results o f these 4 observations: (P = positive or N = negative) 4. The possible cases are as follows: (a) I f the results o f all 4 are P, then the diagnosis is definitely established. (b) I f only 3 are P, then we need to check 1 more o f the remaining 3. Reduction Algorithm in Expert Diagnosis 13S T a b l e 1. D e c i s i o n T a b l e s U s e d for 4/n, 3/n, 2/n, and 1/n P r o b l e m s T I T2 T3 T4 P P P P P P P P N N N N N N N N F P P P N N N N P P P P N N N N P P N N P P N N P P N N P P N N P N P N P N P N P N P N P N P N ~i 0 1 1 2 1 2 2 3 1 2 2 3 2 3 3 4 F o r a n y 4/n p r o b l e m T1 P P P P N N N N T2 P P N N P P N N T3 P N P N P N P N 6 0 1 1 2 1 2 2 3 F o r a n y 3/n p r o b l e m T I P P N N T2 P N P N 6 0 1 1 2 F o r a n y 2/n p r o b l e m T I P N 6 0 1 F o r a n y 1/n p r o b l e m (c) I f o n l y 2 a r e P, t h e n w e n e e d to c h e c k 2 m o r e o f the r e m a i n i n g 3. (d) I f o n l y 1 is P, t h e n w e n e e d to c h e c k 3 m o r e o f the r e m a i n i n g 3. (e) I f all a r e N , t h e n h y p o t h e t i c a l l y w e n e e d to c h e c k 4 o f the r e m a i n i n g 3, w h i c h is i m p o s s i b l e ; thus, w e r e j e c t the d i a g n o s i s . In c a s e (a) t h e r e a r e 4 o b s e r v a t i o n s ; all o f t h e m h o l d , and the d i a g n o s i s is e s t a b l i s h e d ( f u r t h e r c h e c k s a r e 0 o f 3). In c a s e s (b), (c), o r (d), a d i a g n o s i s is not 136 Akram Salah and Kevin D. Reilly established because there is insufficient input information. Instead o f restarting the problem, we can define a new reduced problem such that we check only the remaining observations. N o w we need to check either 1/3 in case (b), 2/3 in case (c), or 3/3 in case (d). In case (e) we can reject the diagnosis because the total number o f observations is 7, 4 o f them have already been checked and failed, and the remaining observations are 3 in number. To establish a diagnosis, 4 observations need to exist; therefore, it is impossible to establish a diagnosis from this situation (4/3). Thus, the method either establishes a diagnosis from the information provided or uses the information to reduce the problem to a smaller problem o f the same nature. The new problem can be solved recursively by the same methodology. Commentary We can cite several advantages o f the reduction methodology: (1) there is guaranteed recursive reduction until a solution is reached; (2) the number o f rules to be checked is less than using the simple rule approach (see Table 2); (3) the tables given in Table 1 can be used for any diagnostic problem k/n, regardless o f the value for n; and (4) the growth o f the number o f rules is limited, as all the decision tables are complete (Welland [15]), and thus there is no possibility o f adding rules to any o f them. As can be seen, the n u m b e r o f rules in the reduction method depends on the length o f the subset that establishes the diagnosis, k, rather than the length o f the domain o f observations, n. This is an important property o f the reduction methodology, as in simple rule approaches the number o f rules grows exponentially with the length o f the set o f observations, assuming that simple rule generation uses either combinations or permutations. According to Weiss and Kulikowski ([14], pp. 118-119), a problem similar to what we have been intimating was detected while an expert system for diagnosis for rheumatic diseases was being implemented. As the expert system evolved, the number o f its (physician) users increased; consequently, the number o f observations known to the system increased. The expert system started with a 4/ 10 diagnostic problem and was extended to 4/18 and eventually to 4/35. A T a b l e 2. N u m b e r o f Rules Used to Build a Knowledge Base Using Using Using Problem ID Permutation Combination Reduction 4/10 5020 210 31 4/18 73440 3060 31 4/35 1256640 52360 31 Reduction Algorithm in Expert Diagnosis 137 simple treatment o f the p r o b l e m (see Table 2) would make the number o f rules increase exponentially with any increase in the number o f observations. A final point to be noted about the reduction method is that it does not depend on any particular application. The algorithm was developed for an expert system for differential diagnosis o f rheumatic diseases, but it can be used in any other rule representation o f the same nature. E N V I R O N M E N T The system that we e m p l o y for representing the methodology o f this article is based on an extension o f a previously defined system called E X P R D (EXtended Prolog Rule Data system), an integration o f a Prolog, a relational database, and a decision table system (Salah [16]). This system is used to store or generate decision tables such as those appearing here. Prolog programs expressing the reduction algorithm are added as a part o f the management program. Sets o f observations are stored in the system as database relations. An interactive dialogue prompts the user to provide the proper information for the diagnostic problem and invokes the reduction algorithm. I f a decision is reached, the program advises the user whether the diagnosis is established or rejected. I f the information is not sufficient to establish the diagnosis, the program prompts the user to provide m o r e information. An example dialogue in a session is as follows: Give me a test you performed: Arthralgia What is the result o f arthralgia ( p = positive, n = negative): p **Diagnosis is P O S I T I V E * * **Chronic p o l y a r t h r i t i s > 6 weeks is a significant factor** What is the result o f synovial fluid inflammatory (p = positive, n = negative): p These results are not sufficient to establish a diagnosis 138 Akram Salah and Kevin D. Reilly What is the result o f subcutaneous nodules (p = positive, n = negative): p • • • • *Diagnosis is NEGATIVE** • *Two positive symptoms are noted** Do you wish a trace o f this dialog? No • • • A general philosophy in dealing with rule systems emerges from our methodology: a management system is employed in which rules are dealt with as one o f the components. Such a management system can be viewed as a meta-rule program that helps a user (or an expert-system administrator) to build, manipulate, query, and analyze a rule system. C O N C L U S I O N Although the reduction methodology for differential diagnosis expert systems is self-contained in the sense that it solves a well-defined problem, if we take a broader view o f the situation, we see this method as part o f the overall rule- management environment. The environment conceptualization emphasizes use of meta-level processing to manipulate rule-like representations. Given a k / n diagnostic problem, an extended form o f rule, the management program is designed to generate a set o f simple rules or employ the reduction methodology to reduce the problem to a smaller problem of the same nature. Employing management programs on the meta-level facilitates a global .view for expert systems, allowing operations such as generation, reduction, or analysis o f rules. R e f e r e n c e s 1. CODASYL Decision Table Task Group, A Modern Appraisal o f Decision Tables, Association for Computing Machinery, New York, 1982. 2. Dahl, V., Logic programming as a representation of knowledge, IEEE Computer, 16(10), 106-111, 1983. 3. Fikes, R., and Kehler, T., The role of frame-based representation in reasoning, Comm. o f the Assn. f o r Computing Machinery 28, 904-920, 1985. Reduction Algorithm in Expert Diagnosis 139 4. Salah, A., and Yang, C. C., Rule-based systems: A set-theoretic approach, Proceedings o f the 3rd Annual Computer Science Symposium on Knowledge- Based Systems: Theory and Application, Columbia, SC, 1986. 5. Hayes-Roth, F., Rule-based systems, Comm. o f the Assn. f o r Computing Machinery 28,921-932, 1985. 6. Reilly, K., Salah, A., Morgan, P., and Rowe, P., Multiple representations in a language-driven memory model, in Papers on Computational and Cognitive Science (E. Battistella, Ed.), Indiana Univ. Linguistic Club, Bloomington, Ind., 87- 94, 1984. 7. Yang, C. C., Relational Databases, Prentice-Hall, Englewood Cliffs, N.J., 1986. 8. Reilly, K. D., Salah, A., and Yang, C. C., A Logic Programming Perspective on Decision Table Theory and Practice, University of Alabama at Birmingham Tech. Report, 1986. 9. Kowalski, R., Logic f o r Problem Solving, Elsevier-North Holland, New York, 1979. 10. Clocksin, W., and Mellish, C., Programming in Prolog, Springer-Verlag, New York, 1981. 11. Bruynooghe, M., Prolog in C f o r Unix Version 7: A Reference Manual, Katholieke Univ., Leuven, Belgium, 1980. 12. Prade, H., A computational approach to approximate and plausible reasoning with applications to expert systems, IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-7, 260-283, 1985. 13. Rosenbloom, P., Laird, J., McDermott, J., Newell, A., and Orciuch, E., R1-Soar: An experiment in knowledge-intensive programming in a problem-solving architec- ture, IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-7, 561-569, 1985. 14. Weiss, S., and Kulikowski, C., A Practical Guide to Designing Expert Systems, Rowman & Allanheld, Philadelphia, 1984. 15. Welland, R., Decision Tables and Computer Programming, Heyden & Son, London, 1981. 16. Salah, A., A n Integration o f Decision Tables and a Relational Database System into a Prolog Environment, PhD Thesis, Univ. of Alabama at Birmingham, Birmingham, Ala., 1986. 
work_tuincs53g5djxid6hkzpbfseya	----	Microsoft Word - P_1856 Sensors & Transducers, Vol. 164, Issue 2, February 2014, pp. 227-232 227 SSSeeennnsssooorrrsss &&& TTTrrraaannnsssddduuuccceeerrrsss © 2014 by IFSA Publishing, S. L. http://www.sensorsportal.com Expert System Development on On-line Measurement of Sewage Treatment Based Process 1 Jianjun QIN, 2 Huajun GUO 1 Beijing University of Civil Engineering and Architecture, No. 1 Zhanlanguan Rd., Xicheng District, Beijing 100044, P. R. China 2 Hefei Xinan Patent agent co., LTD, No. 1137 East Changjiang Rd., Heifei 230011, P. R. China 1 Tel.: 86-10-68322420, fax: 86-10-68322420 1 E-mail: Chinaqjj@163.com Received: 19 October 2013 /Accepted: 28 November 2013 /Published: 28 February 2014 Abstract: This article puts forward a solution in which an instrument on-line automatic measurement and expert system process are optimized according to the complexity and great process dynamics of sewage treatment process. Firstly modeling has been set up with configuration sewage treatment process in which the process has been integrated into the computer software environment. Secondly certain number of water quality automatic monitoring instruments and sensor probes are set in the reaction tanks according to the needs of process changes and management. The data information acquired can be displayed and recorded at the real time. A human- machine integration expert system featuring computer automation management is developed for the base by one-off method thus to realize the intelligent and unmanned management. The advantages brought about from it can fill up the inexperience of the on-site management personnel and solve the contradiction between the water quality dynamics and difficulty in the process adjustment. Copyright © 2014 IFSA Publishing, S. L. Keywords: Data acquisition, Expert system, On-line measurement, Process management, Sewage treatment process. 1. Introduction There are lots of sewage treatment methods. Generally the ideal treatment results can be achieved only when they can operate properly with the processing steps of the several intercoupling according to the proper parameters [1]. The treatment results will be affected by the sewage quality condition and flow in the specific process implementation [2]. Therefore the operating parameters of process should be adjusted properly according to the water quality and flow so it is necessary to obtain the indicators such as water quality and flow parameters [3-5]. In the ideal condition, although the measurement of various sewage quality indicators can be completed with various monitoring instruments and tools, the measurement principle and acquisition periods of the detecting instruments can be different and the water quality parameters can be relatively great even to the location of same treatment tank, so it is so hard to obtain more real time information, which results in the difficulty in implementing the process that directs the sewage treatment accurately [5, 6]. Automatic monitoring and control equipment can be applied in practical operation of lots of sewage treatment works, for instance the comprehensive monitoring can be conducted on the sewage treatment process with Article number P_1856 http://www.sensorsportal.com/ Sensors & Transducers, Vol. 164, Issue 2, February 2014, pp. 227-232 228 monitoring system consisting of PLC, and moreover the acquisition of real time data distribution at several points can be conducted through filed control bus and the real time control commands can be sent out through remote computer. However such links cannot be isolated. Actually the implementation of the sewage treatment process relies on the experience of the operators to a great extent. The real data acquisition cannot timely be reflected to the process adjustment, thus the greater deviation can be avoided but the existing problems is the precision is not high and excessive reliance on the experience on the field operators. 2. Design on On-line Monitoring System 2.1. Selection and Layout of On-line Measurement Device The on-line detecting instruments that are used in this article can be shown as follows: total nitrogen/total phosphorus/COD analyzer, Amtax™ Compact ammonia nitrogen analyzer. Sensor electrodes can be used to monitor the two parameters: ORP and PH and on-line monitor can be used to monitor the parameters such as NO3-, NH4+, PO43-, COD, TP, TN and DO. Output signal of the instruments and electrodes is 4-20 mA direct current. There are 24 data acquisition points and output signal is analog quantity signal, which can be shown in Table 1. Table 1. Statistics of sensors and executing elements. No. Name Type or specification Qty. 1. DO sensor Sensor electrode MV3030 2 Transmitter MF41 2. ORP sensor Sensor electrode MV3015 3 Transmitter EMC173- K030-F-P 3. PH sensor Sensor electrode MV3010 1 Transmitter EGA173- K030-F-P 4. NO3- sensor Sensor electrode MV3016 3 Transmitter NO31508 5. NH4+ sensor Sensor electrode MV3016 3 Transmitter NH41508 6. PO4 3- monitor Phosphax Compact PO4 6 7. TN monitor NPW-150 2 8. TP monitor 2 9. COD monitor 2 10. Flow sensor LWGY-40A 7 11. Solenoid valve 2 12. Electromagnetic switch 5 13. Pump BT300-1F 7 14. Agitator 4 15. Air compressor 1 The real time continuous survey and remote monitoring can be conducted on the water quality through the data acquisition on the sewage tanks so as to master the water quality situation, warn and forecast the pollution accident on the water quality and reach the emission standards. Fig. 1 is the relevant relation figures for the data acquisition points and control points. The information of the data acquisition is the water quality information obtained through the on-line measurement equipment. The main control variables include water inflow, extraction volume of clear liquid, aeration rate, internal reflux A, internal reflux B, sludge reflux and the remaining sludge displacement. The different types of data acquired by the various water quality sensors from the different sewage tanks can be transmitted in different ways to control system consisting of IPC and PLC. Its main application is to conduct the real time monitoring the sewage treatment according to the working cycles and feedback to the system. PLC can acquire the data from the water quality sensor in a current form. In the sewage treatment process, the BCFS°,R process can play an auxiliary role in the control system, assisting the two systems in working together (distributed closed cycle control system and human-machine interactive intelligent control platform). Fig. 1. Relation figure between data acquisition and control points. 2.2. Layouts of the Data Acquisition Points Through the research on BCFS°,R process, it is preliminarily confirmed that the monitoring parameters on the inflow, effluent and each reaction tank can be shown as follows: Water inlet R0: COD, TP, TN, NH4+, PO43-; Anaerobic R1: PO43-, ORP; Anoxia tank R3: NO3-, PO43-, ORP; Anoxia/aerobiotic tank R4: DO, NO3-, ORP; Aerobiotic tank R5: DO, NO3-, PO43-; Flow tank (water outlet) R6: COD, TP, TN, NH4+, PO43-; inflow water quality  inflow water volume  effluent quality  water quality in the reaction tank （ORP、NO2+- —N、NO3+—N）  clear liquid extraction volume  sludge reflux quantity  DO concentration, temperature and sludge concentration in aeration tank  the remaining sludge displacement  internal reflux A  internal reflux B  aeration quantity Sensors & Transducers, Vol. 164, Issue 2, February 2014, pp. 227-232 229 Phosphorus sedimentation tank R7: pH, NH4+, PO43-; To clearly express the location of each water quality sensor, the layout of the main measurement instruments have been drawn out in combination with the sewage treatment process as shown in Fig. 2. The volume and types that are needed in the actual water quality monitoring in each tank can be different. Fig. 2. Layout of data acquisition points. 2.3. Ammonia Nitrogen On-line Measuring System In this section, ammonia nitrogen on-line measurement is taken as an example to explain the conversion relation between the direct current data acquired and concentration. Ammonia nitrogen sensors can be connected the built-in connection port. Its signal can be outputted in direct current forms (0 or 4~20 mA) with the function of port number as shown in Table 2. The relation between ammonia nitrogen and output current can be obtained through the several measurements. Each concentration corresponds to stable current value. An approximately linear relation will be formed by connecting these points as shown in Fig. 3. Thus we can educe the relation of formula 1 and the functions of ammonia nitrogen sensor to output current has been maintained until the new concentration value has been read out. Fig. 3. The linear relation diagram between current acquired and concentration. Table 2. Port number and functions. Connection port number Design function 1. Wrong information –NC connection 2. Wrong information 3. Wrong information-NO connection 4. Minimum 1-NC connection 5. Minimum 1 6. Minimum 1-NO connection 7. Maximum 1-NC connection 8. Maximum 1 9. Maximum 1 NO connection 10. +current 1 11. -current 2 12. Screen current 1 13. (operational) Minimum 2-NC connection 14. (operational) Minimum 2 15. (operational) Minimum 2-NO connection 16. (operational) Minimum 2-NC connection 17. (operational) Maximum 2 18. (operational) Maximum 2 NO Connection 19. (operational) +current 2 20. (operational) -current 2 21. (operational) Screen current 22. Not used 23. G DIN measurement bus 24. RB DIN measurement bus 25. RA DIN measurement bus 26. TB DIN measurement bus 27. TA DIN measurement bus 28. DIN shield measurement bus Input current (20 0) Concentration (120 2) / mA mg L    (1) The selected measuring range of the ammonia nitrogen sensor is 2-120 mg/L and the output signal is Current (0-20 mA) C o n ce n tr at io n ( 2 -1 2 0 m g /L ) Sensors & Transducers, Vol. 164, Issue 2, February 2014, pp. 227-232 230 0-20 mA. When the corresponding physical quantity is calculated according to the output value of the analog quantity output modules, the consideration of the input/output range and range of analog quantity output module should be taken to find out the proportional relation of the figures after A/D conversion of the measured physical quantities. In the corresponding relations among the analog value of analog output modules and analog quantity expressed in percentage, the most important one is upper limit and lower limit (100 % and -100 %) of bipolar analog quantity range, which correspond to the analog values of 27648 and -27648. the upper and lower limits of monopolar analog quantity range (100 % and 0 %) correspond to the analog value of 27648 and 0 respectively. 2.4. Real Time Water Quality Parameter Display The acquisition of other water quality parameters is similar to the data acquisition principle of ammonia nitrogen sensor. The real time water quality parameter information of each sensor corresponding to each sewage treatment tank can be obtained as shown in Fig. 4. We can also inquiry the historical information on water quality through WinCC configuration interface; moreover it can be displayed in the forms of curve and statement as shown in Fig. 5. Fig. 4. Display of real time sewage quality parameters 3. Design and Realization of Expert System 3.1. Overall Structure of Expert System The process adjustment implemented in the sewage treatment works is more dependent on the artificial expert, namely the experience of skilled technicians. The data information for reference is so limited. Therefore the data acquired through on-line measurement and acquisition system can meet the needs of accurate management to the sewage treatment process, thus to design the expert system with higher reliability. The structure of the expert system used in this article can be shown as Fig. 6. The data on sewage treatment is from two parts, one is the data automatically acquired from the on-line monitoring instruments, which is the main data, the other is other data temporarily inputted by the management personnel, for example, some data inputted when the deviation occurs or some data that is easily to acquire artificially but not easy to acquire by the instruments. Fig. 5. Inquiry on historical water quality parameters. Fig. 6. Structure of expert system for BCFS°,R process. 3.2. Design on Knowledge Database After making a comparison on the similarities and differences between database and knowledge database, this article analyzes the composition and features of the expert system related to BCFS°,R process, believing that relation database based knowledge can better manage the knowledge database with design plan given as well. Knowledge management system established with the relational data base can make the whole knowledge database have the ability to better store, maintain and extend knowledge. In this article, the rules are divided into the rule elements then the rule elements can be combined together again to from a production rule Human-machine interface (including the explanation subsystem) Human-machine interface Other data of sewage (inputted by the management personnel) Dynamic database (sensor data acquisition and analysis result) Static database (knowledge base and sets of rules) Sensors & Transducers, Vol. 164, Issue 2, February 2014, pp. 227-232 231 by using the relation between primary key and external key. In BCFS°,R process, DO and ORP are taken as the main control parameters because currently the other parameters have been not allocated the real time acquisition instruments. There is some main knowledge on the controls of the aeration valves and pump in reflow process. 1. When DO is too low in aerobic tank (low than 1.5 mg/L, carry out aeration. 2. ORP is too low in Anoxic/Aerobic pool and anaerobic tank (generally above -150 mV), carry out aeration. 3. DO is too high in aerobic tank (DO in Anoxic/Aerobic pool is generally not higher than 0.5 mg/L), stop aeration. 4. The control on the cycle A, B and c of reflux pump can be conducted according to ORP reference values. For example, to the knowledge that if the DO in the aerobic tank should generally above 1.5 mg/L, otherwise the aeration valve of Anoxic/Aerobic pool should be opened so as to improve the oxygen dissolution in the sewage, it can he expressed in the production rules as if “DO in the aerobic tank is lower than 1.5 mg/L, aeration is needed in Anoxic/Aerobic pool”. The control suggestion of the system on executive components is to open the aeration valve of Anoxic/Aerobic pool. To the two water quality parameters of DO and ORP, there are 10 pieces of control rules that have been confirmed currently as shown in Table 3 through the discussion with the field operators and relevant experts. 3.3. Inference Engine The composition and realization of the inference engine depend on the nature of the field related problems and expression methods and structure of the knowledge. In this project the rule based means of knowledge expression are used through the discussion with the relevant experts. As the system makes the judgment according to the data acquired, the forward reasoning driven by data is more suitable for this system. CLIPS can not only support the rule based means of knowledge expression and its reasoning mechanism is also based on forward reasoning control strategy. This can be the main reason why CLIPS language is used by the expert system in the design. CLIPS development system itself is an inference engine. In this article, the method of insertion of CLIPS into VC++6.0 is adopted so as to realize the expert system design with a good human –machine interface. The reasoning process of the expert system in the system can be shown in Fig. 7. 3.4. Design on Human-Machine Interface Human-machine interface is generated under VC++6.0 development environment. Enter the system and log on the interface as Fig. 8. Input the code users’ name and password and then enter expert system operation interface as Fig. 9. Table 3. Control Rules about DO and ORP. No. Precondition Conclusion Suggestions BQ01 DO is less than 1.5 mg/L in the aerobiotic tank R5. Aeration is needed in the Anoxic/Aerobic pool Open the aeration valve of Anoxic/Aerobic pool. BQ02 ORP is less than -150 mV in Anoxic/Aerobic pool R4 Aeration is needed in the Anoxic/Aerobic pool Open the aeration valve of Anoxic/Aerobic pool. BQ03 ORP is less than - 150 mV in anoxia tank R3 Aeration is needed in the Anoxic/Aerobic pool Open the aeration valve of Anoxic/Aerobic pool. BQ04 DO is greater than 0.5 mg/L in Anoxic/Aerobic pool R4 Aeration quantity is too great in Anoxic/Aerobic pool Close the aeration valve of Anoxic/Aerobic pool. HLA01 ORP is less than - 200 mV in anaerobic tank R1. Insufficient reflux volume of cycle A Increase the pump flow of reflux A HLA02 ORP is greater than -100 mV in anaerobic tank R1. Too large reflux volume of cycle A Reduce the pump flow of reflux A HLB01 ORP is less than -100 mV in anoxia tank R3. Insufficient reflux volume of cycle B Increase the pump flow of reflux B HLB02 ORP is greater than -100 mV in anoxia tank R3. Too large reflux volume of cycle B Reduce the pump flow of reflux B HLC01 ORP is less than +100 Mv in Anoxic/Aerobic pool R4. Insufficient reflux volume of cycle C Increase the pump flow of reflux C HLC02 ORP is greater than +300 mV in Anoxic/Aerobic pool R4. Insufficient reflux volume of cycle C Reduce the pump flow of reflux C Fig. 7. Flow chart of inferring process of expert system. Start Read the water quality data Is there a proper rules? If more than one rule According to the rules of the control strategy and deposited in the database Search rules Select top rule End Sensors & Transducers, Vol. 164, Issue 2, February 2014, pp. 227-232 232 Fig. 8. Interface of logging in expert system for BCFS°,R. Fig. 9. Main menu of expert system. 4. Process Human-Machine Interface and Data Interaction of Expert System Human-machine interface designed with Configuration software can input the data acquired into the database with data format of SQL 2000. In this system, WinCC and expert system can be located in the same computer. Direct access can be conducted to the filed database on water quality in WinCC project through ODBA or OLE-DB and then the water quality data can be read in the expert system program. The control results acquired from the expert system can be inputted into the database of human-machine interface. WinCC human –machine interface and data conversion process of expert system can be shown in the Fig. 10. 5. Conclusions and Prospect With the popularization of sewage treatment process developing toward miniaturization and decentralization, some advanced sewage treatment process has been promoted to the small towns. The first and foremost problems we face in the process are the operation maintenance and daily management but there are not enough experienced technicians. This article put forward the sewage treatment process expert system based on the data automatic measurement can solve this bottle neck problem. At the first stage of the process operation, the problems such as poor data accuracy and lots of artificial intervention needed in the process adjustment. However with the data acquisition quantity increased and process operation stabilized, the system can realize the automatic operation furthest and can be directly used in the similar sewage treatment process with remarkable scaled profit. Fig. 10. Data exchange between WinCC and expert system. References [1]. J. Baeza, D. Gabriel, et al., In-line fast OUR (oxygen uptake rate) measurements for monitoring and control of WWTP, Water Science & Technology, Vol. 45, Issue 4, 2002, pp. 19-28. [2]. D. Vrecko, N. Hvala, et al., Wastewater treatment benchmark: what can be achieved with simple control?, Instrumentation, Control and Automation, Vol. 45, Issue 4-5, 2002, pp. 127-13. [3]. J. Comas, I. Rodríguez-Roda, et al., Demonstration of a tool for automatic learning and re-use of knowledge in the activated sludge process, Instrumentation, Control and Automation for Water and Wastewater Treatment and Transport Systems IX, Vol. 53, Issue 4-5, 2006, pp. 303-311. [4]. P. Vanrolleghem, D. Lee, On-line monitoring equipment for wastewater treatment processes: State of the art, Automation in Water Quality Monitoring, Vol. 47, Issue 2, 2003, pp. 1-34. [5]. D. Choi, H. Park, A hybrid artificial neural network as a software sensor for optimal control of a wastewater treatment process, Water Research, Vol. 35, Issue 16, 2001, pp. 3959-3967. [6]. K. Villez, K. Steppe, et al., Use of Unfold PCA for on-line plant stress monitoring and sensor failure detection, Biosystems Engineering, Vol. 103, Issue 1, 2009, pp. 23-34. [7]. Y. Liu, P. Pei, Asymptotic analysis on autoignition and explosion limits of hydrogen–oxygen mixtures in homogeneous systems, International Journal of Hydrogen Energy, Vol. 31, Issue 5, 2006, pp. 639-647. ___________________ 2014 Copyright ©, International Frequency Sensor Association (IFSA). All rights reserved. (http://www.sensorsportal.com) SQL Server 2000 database ODBC data source VC6.0 Wincc Monitor ing system Expert system ODBC data source << /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Dot Gain 20%) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Error /CompatibilityLevel 1.4 /CompressObjects /Tags /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /CMYK /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams false /MaxSubsetPct 100 /Optimize true /OPM 1 /ParseDSCComments true /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo true /PreserveFlatness true /PreserveHalftoneInfo false /PreserveOPIComments true /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Apply /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true ] /NeverEmbed [ true ] /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 300 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 300 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.50000 /EncodeColorImages true /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /ColorImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 300 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 300 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /GrayImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 1200 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 1200 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 >> /AllowPSXObjects false /CheckCompliance [ /None ] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile () /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description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> /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000410064006f006200650020005000440046002065876863900275284e8e9ad88d2891cf76845370524d53705237300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef69069752865bc9ad854c18cea76845370524d5370523786557406300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /CZE <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> /DAN <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> /DEU <FEFF00560065007200770065006e00640065006e0020005300690065002000640069006500730065002000450069006e007300740065006c006c0075006e00670065006e0020007a0075006d002000450072007300740065006c006c0065006e00200076006f006e002000410064006f006200650020005000440046002d0044006f006b0075006d0065006e00740065006e002c00200076006f006e002000640065006e0065006e002000530069006500200068006f006300680077006500720074006900670065002000500072006500700072006500730073002d0044007200750063006b0065002000650072007a0065007500670065006e0020006d00f60063006800740065006e002e002000450072007300740065006c006c007400650020005000440046002d0044006f006b0075006d0065006e007400650020006b00f6006e006e0065006e0020006d006900740020004100630072006f00620061007400200075006e0064002000410064006f00620065002000520065006100640065007200200035002e00300020006f0064006500720020006800f600680065007200200067006500f600660066006e00650074002000770065007200640065006e002e> /ESP <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> /ETI <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> /FRA <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> /GRE <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a stvaranje Adobe PDF dokumenata najpogodnijih za visokokvalitetni ispis prije tiskanja koristite ove postavke. Stvoreni PDF dokumenti mogu se otvoriti Acrobat i Adobe Reader 5.0 i kasnijim verzijama.) /HUN <FEFF004b0069007600e1006c00f30020006d0069006e0151007300e9006701710020006e0079006f006d00640061006900200065006c0151006b00e90073007a00ed007401510020006e0079006f006d00740061007400e100730068006f007a0020006c006500670069006e006b00e1006200620020006d0065006700660065006c0065006c0151002000410064006f00620065002000500044004600200064006f006b0075006d0065006e00740075006d006f006b0061007400200065007a0065006b006b0065006c0020006100200062006500e1006c006c00ed007400e10073006f006b006b0061006c0020006b00e90073007a00ed0074006800650074002e0020002000410020006c00e90074007200650068006f007a006f00740074002000500044004600200064006f006b0075006d0065006e00740075006d006f006b00200061007a0020004100630072006f006200610074002000e9007300200061007a002000410064006f00620065002000520065006100640065007200200035002e0030002c0020007600610067007900200061007a002000610074007400f3006c0020006b00e9007301510062006200690020007600650072007a006900f3006b006b0061006c0020006e00790069007400680061007400f3006b0020006d00650067002e> /ITA <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> /JPN <FEFF9ad854c18cea306a30d730ea30d730ec30b951fa529b7528002000410064006f0062006500200050004400460020658766f8306e4f5c6210306b4f7f75283057307e305930023053306e8a2d5b9a30674f5c62103055308c305f0020005000440046002030d530a130a430eb306f3001004100630072006f0062006100740020304a30883073002000410064006f00620065002000520065006100640065007200200035002e003000204ee5964d3067958b304f30533068304c3067304d307e305930023053306e8a2d5b9a306b306f30d530a930f330c8306e57cb30818fbc307f304c5fc59808306730593002> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020ace0d488c9c80020c2dcd5d80020c778c1c4c5d00020ac00c7a50020c801d569d55c002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /LTH <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> /LVI <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> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken die zijn geoptimaliseerd voor prepress-afdrukken van hoge kwaliteit. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <FEFF004200720075006b00200064006900730073006500200069006e006e007300740069006c006c0069006e00670065006e0065002000740069006c002000e50020006f0070007000720065007400740065002000410064006f006200650020005000440046002d0064006f006b0075006d0065006e00740065007200200073006f006d00200065007200200062006500730074002000650067006e0065007400200066006f00720020006600f80072007400720079006b006b0073007500740073006b00720069006600740020006100760020006800f800790020006b00760061006c0069007400650074002e0020005000440046002d0064006f006b0075006d0065006e00740065006e00650020006b0061006e002000e50070006e00650073002000690020004100630072006f00620061007400200065006c006c00650072002000410064006f00620065002000520065006100640065007200200035002e003000200065006c006c00650072002000730065006e006500720065002e> /POL <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> /PTB <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> /RUM <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> /RUS <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> /SKY <FEFF0054006900650074006f0020006e006100730074006100760065006e0069006100200070006f0075017e0069007400650020006e00610020007600790074007600e100720061006e0069006500200064006f006b0075006d0065006e0074006f0076002000410064006f006200650020005000440046002c0020006b0074006f007200e90020007300610020006e0061006a006c0065007001610069006500200068006f0064006900610020006e00610020006b00760061006c00690074006e00fa00200074006c0061010d00200061002000700072006500700072006500730073002e00200056007900740076006f00720065006e00e900200064006f006b0075006d0065006e007400790020005000440046002000620075006400650020006d006f017e006e00e90020006f00740076006f00720069016500200076002000700072006f006700720061006d006f006300680020004100630072006f00620061007400200061002000410064006f00620065002000520065006100640065007200200035002e0030002000610020006e006f0076016100ed00630068002e> /SLV <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> /SUO <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> /SVE <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> /TUR <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> /UKR <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> /ENU (Use these settings to create Adobe PDF documents best suited for high-quality prepress printing. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.) >> /Namespace [ (Adobe) (Common) (1.0) ] /OtherNamespaces [ << /AsReaderSpreads false /CropImagesToFrames true /ErrorControl /WarnAndContinue /FlattenerIgnoreSpreadOverrides false /IncludeGuidesGrids false /IncludeNonPrinting false /IncludeSlug false /Namespace [ (Adobe) (InDesign) (4.0) ] /OmitPlacedBitmaps false /OmitPlacedEPS false /OmitPlacedPDF false /SimulateOverprint /Legacy >> << /AddBleedMarks false /AddColorBars false /AddCropMarks false /AddPageInfo false /AddRegMarks false /ConvertColors /ConvertToCMYK /DestinationProfileName () /DestinationProfileSelector /DocumentCMYK /Downsample16BitImages true /FlattenerPreset << /PresetSelector /MediumResolution >> /FormElements false /GenerateStructure false /IncludeBookmarks false /IncludeHyperlinks false /IncludeInteractive false /IncludeLayers false /IncludeProfiles false /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe) (CreativeSuite) (2.0) ] /PDFXOutputIntentProfileSelector /DocumentCMYK /PreserveEditing true /UntaggedCMYKHandling /LeaveUntagged /UntaggedRGBHandling /UseDocumentProfile /UseDocumentBleed false >> ] >> setdistillerparams << /HWResolution [2400 2400] /PageSize [612.000 792.000] >> setpagedevice 
work_tuo4bsb5qnfzfmkqa3xrh6hvwu	----	The Human Behaviour Indicator; A Measure of Behavioural Evolution Abubaker Elbayoudi1,2, Ahmad Lotfi1,∗, Caroline Langensiepen1 School of Science and Technology, Nottingham Trent University, Clifton Lane, NG11 8NS, Nottingham, United Kingdom Abstract Activities of daily living (ADL) or activities of daily working (ADW) may be affected by changes in a person’s health or well-being. Measuring pro- gressive changes in one activity or multiple activities is representative of behavioural variations. By inspecting the trends in multiple activities, it is possible to identify and predict human behavioural changes. We refer to the trends in people’s behaviour as behavioural evolution. In this paper, we propose a novel indicator to measure the progressive changes representing a participant’s behavioural evolution. The proposed indicator presents ac- tivities as a holistic measure, which first combine multi-activities and then measure the progressive changes in the combined activities for each single day. Real data sets were collected from a wireless sensor network and used to examine our proposed technique. As part of this process, we were able to quantify progressive changes for individual and aggregated activities. Our experimental results demonstrated that: (1) the proposed approach can identify and distinguish normal and abnormal behaviours; (2) large data sets gathered from sensors in an intelligent environment represented in various time series can be visualised in a simple and more understandable format; (3) identifying trends in ADLs or ADWs is a relevant means of sharing information with carers or supervisors. ∗Corresponding author: Email: ahmad.lotfi@ntu.ac.uk 1All authors are with School of Science and Technology, Nottingham Trent Univer- sity, Clifton Lane, NG11 8NS, Nottingham, United Kingdom. Ahmad Lotfi (Email: ah- mad.lotfi@ntu.ac.uk ); Abubaker Elbayoudi (Email: abubaker.elbayoudi@ntu.ac.uk ) and Caroline Langensiepen (Email: caroline.langensiepen@ntu.ac.uk ) 2PhD research student who has conducted the research as part of his thesis. Preprint submitted to Expert Systems with Applications October 23, 2018 The following abbreviations are used in this manuscript: ADL Activities of Daily Living ADW Activities of Daily Working AmI Ambient Intelligence EWMA Exponentially Weighted Moving Average EMD Empirical Mode Decomposition HBI Human Behaviour Indicator MA Moving Average MCI Mild Cognitive Impairment PD Parkinson’s Disease SCI Synthetic Composite Indicators WHBI Weighted Human Behaviour Indicator EWHBI Exponential Weighted Human Behaviour Indicator Keywords: Activities of daily living; activities of daily working; trend analysis; moving average; behaviour indicator; behavioural evolution 1. Introduction Measuring progressive change in people’s activities is a subject of re- search interest in the machine learning community and has recently at- tracted more attention (Mahmoud et al., 2014) (Wang et al., 2018) (Nweke et al., 2018). These could be activities of daily living (ADL) or activities of daily working (ADW). A number of medical conditions and their treat- ments are associated with activities disorders such as reduced movement over time. Parkinson’s Disease (PD), for example, is characterised by slowness of movement (Zhan et al., 2016), and some Mild Cognitive Impairment (MCI) causes a slight but noticeable and measurable decline in ADL (Wallace et al., 2017)(Kearns et al., 2017). Understanding progressive changes in human behaviour is the key chal- lenge in formulating an effective intelligent environment. Many researchers have investigated human activity recognition in order to solve the prob- lem of activity extraction and prediction. These studies have used different statistical as well as Computational Intelligence methods and techniques to identify the behaviour of occupants based on temporal data gathered from sensor networks. There are many challenges in using such methods to 2 find the relationship between the gathered sensory data from an intelligent environment and identifying the actual behaviour of the participant. Un- derstanding human behaviour from low level sensory data and interpreting them in a meaningful manner are the main challenges that the research here face. The research aims to identify trends in ADL or ADW and interpret them in a presentable format to the user. Change of behaviour of a participant in an Ambient Intelligence environments (AmI) is an indicator of a person’s social and health status. The paper is primarily concerned with the inter- pretation of progressive changes in the participant’s behaviour. The project is aiming to assess the ADLs or ADWs of a person who lives or works in- dependently in his/her own home or office and present it as a single value for each day. This will provide a holistic view of the monitored person’s behaviour and it will be used to forecast the behavioural changes of the par- ticipant and it will raise an early warning of an abnormal behaviour when it is observed or expected to happen in the near future. This paper, therefore, will employ and assess such approaches to discover suitable methods for constructing a synthetic composite indicator (SCI). In particular, it aims to deliver an intelligent technique as a framework to build a human behaviour indicator (HBI) that can help to enable consistent and transparent assessments and forecasting of the progressive changes in human behaviour. The HBI represents a holistic report based on multi- ple sensors/activities representing progressive changes in the participant’s behaviour. An alternative means of identifying the behavioural changes is by clas- sifying the activities into normal and abnormal groups. The underlying technique is to use supervised or unsupervised classifiers to learn the pat- tern within the activities. There are some challenges to be addressed by this approach. Firstly, human activities are a combination of many interrelated activities and classifying activities individually will not provide a holistic view of the changes. Secondly, the classification would require suitable tech- niques to deal with temporal data classification. To be able to have a measure of progressive changes, the first step is to integrate some pervasive measurements into home and work environments to collect the data representing ADL or ADW. There are already many off- the-shelf products available to collect such data. Once individual activities are recognised, identifying trends in users’ behaviour over a period of time will provide useful information. Trends in people’s behaviour are used to identify progressive changes and predict behavioural abnormalities. We refer to the trends in people’s behaviour as behavioural evolution. For instance, 3 identifying the progressive changes in the behaviour of a person suffering from MCI will allow the caregiver to monitor and intervene when abnormal behaviour is predicted. The ADL pattern will change over time and this is a consequence of the individual’s condition and progression of the disease. Identifying an evolving behavioural pattern will help to predict the trend in the ADL before any abnormalities are identified. As a different example, ob- serving and analysing a person’s work activity in an office environment will provide managers with an analytical tool to measure employee’s productiv- ity. This could be a holistic measure of many activities including computer work, desk work, meetings etc. In this paper, we seek answers to the following questions: • How to extract behavioural patterns of a person’s ADL or ADW by analysing sensory data collected from different sensors in an intelligent environment. • How the sensory data is processed to become a time series and then to identify trends within the data. • How the data is forecast to extract important daily patterns from them and predict the direction of the trends in the data. • How the multiple data that represent multiple activities or events in every individual day are combined in one single datum to represent the participant’s overall behaviour. • How unexpected patterns and anomalies within the combined data can be identified. This paper is organised as follows: related work is presented in Sec- tion 2 followed by the new human behaviour indicator and its constructions introduced in Section 3; three case studies using different data sets are dis- cussed in Section 4 and the conclusions and suggestions for future work are presented in Section 5. 2. Related Work In this section a brief summary of the related work in the context of trend analysis, abnormality detection and activities recognition is presented. 4 2.1. Trend Analysis Trend analysis techniques are used to understand the participants’ gen- eral health variation and behaviour evolution. Analysing the activities of a participant can only be fully understood if the daily activities are examined in terms of person, place and time. Numerous researchers have investigated trend analysis techniques that can be applied to human behaviour activi- ties extracted from intelligent environments. For example, Mahmoud et al. (2011) presented the importance of trends in the ADL of a single elderly per- son. Researchers on the BackHome (Rafael-Palou et al., 2015) and iCarer (Lotfi et al., 2017) projects introduced monitoring systems to send feed- back on any changes in the participant’s behaviour and habits to the carer. Trends in activities of daily living disability in a large sample of community- dwelling are presented in (Yu et al., 2016). A very similar study of trends in disability of instrumental activities of daily living among older Chinese adults is presented (Liang et al., 2017). Trends in activities of daily living among stroke survivors are investigated by Chattopadhyay et al. (2013). Forkan et al. (2015b) used Holt’s linear trend method to forecast the time series data of vital signs such as heart rate or blood pressure. In another effort, Forkan et al. (2015a) used trend analysis to detect trends in their data, but it is not clear which trend analysis technique they used. A linguistic summarisation for describing long-term trends in changes in human behaviour is presented by Ros et al. (2011); the procedure is used to provide information to older adults and their carers or family in language that is easy to understand by applying a measure of similarity to compare behaviours that are adapted over time. There idea is based on clustering and comparing the similarities between the activities in a specific period in order to detect trends. However, the study concentrated on linguistic summarisation more than trend analysis itself. Financial time series are considerably difficult to analyse and predict. Furlaneto et al. (2017) have applied a bias effect on predicting market trends with empirical mode decomposition (EMD) to predict market indices. Bagher et al. (2017) introduced the concept of trend to capture dynamics in user interests for a content-based recommender system. The authors pro- posed a Bayesian nonparametric model to construct the trend distributions. 2.2. Abnormality Detection Suryadevara et al. (2013) applied a double exponential smoothing strat- egy to the activity duration time series in models that detects changes in the behavioural and health-related status of a monitored patient who lives 5 in a smart home. Saives et al. (2015) present a model to improve the auton- omy of medically monitored behaviour changes for patients in a smart home. Smart home environments can perceive long-term changes that may cause health concerns. Such systems alert the carers and family of any important changes in the occupant’s behaviour, such as diet, daily tasks or health. AmI offers such solutions, for example, by using human behaviour recognition to monitor the person’s activities and alert the carer if something abnormal is detected. Human behaviour recognition has been demonstrated to be a valuable key to understand people’s needs (Lara and Labrador, 2013). Probabilistic models such as the Hidden Markov Model (HMM) and Bayesian belief networks can be used to model human behaviour (Liu et al., 2010)(Xie and Wu, 2012). Soft computing and machine-learning techniques such as artificial neural networks (ANNs) and support vector machines (SVM) can also be used in this way (Yin et al., 2008)(He and Jin, 2009). However, all mentioned techniques face difficulties in processing large amounts of low-level sensory data; therefore, it is essential to transform the sensory data into a suitable format that can be processed (Cook et al., 2009). Aran et al. (2016) have investigated anomaly detection in elderly daily behaviour based on a probabilistic spatio-temporal model. They have sum- marized daily behaviour to be able to identify the anomaly. A Bayesian formulation is provided for anomaly detection in (Ordóñez et al., 2015). Ac- tivities of residents were extracted using Bayesian statistics and behaviours were estimated based on three probabilistic features. The proposed ap- proach has been validated using 14-25 days of real data. The research in (Suryadevara and Mukhopadhyay, 2012) proposes to compute a wellness in- dex to capture abnormal behaviour using several weeks of observations from elderly living alone. Their proposed approach is based on fusing of informa- tion gathered from sensors in the environment. The wellness functions are calculated during the runtime of the system as background process taking the activity durations from the respective files. The proposed approach is rather simplistic with no experimental data from a real environment. In a very recent publications, Kim and Cho (2018) have applied a hy- brid of convolutional neural network (CNN) and long short-term memory recurrent neural network (LSTM) known as C-LSTM to web traffic anomaly detection. Similar work is reported in video surveillance systems and recog- nition of abnormal behaviour. Readers are refer to (Mabrouk and Zagrouba, 2018) for a review of the subject area. Based on our survey of the literature, the proposed technique presented in the next section to measure the behavioural evolution is novel and it has not been investigated by other researchers. 6 3. The Human Behaviour Indicator To represent human behaviour evolution through a simple measure re- quires some method of combining different types of data. This will require the construction of a composite indicator which can be viewed as a means of reducing multivariable inputs into a single and meaningful output that can be interpreted by non-experts. Aggregating multivariable data seems a simple process, however “the methods for aggregating vast amounts of empirical data remain rather crude” (Cherchye et al., 2002). After giving a short overview of composite indicators below, the Human Behaviour Indi- cator (HBI) will be explained. 3.1. Composite Indicators The aggregated output usually represents a holistic view using a single numerical score and/or an ordinal rank. To achieve a single output value, the index must go through accurate development steps. These steps involve consideration of included or excluded variables, handling missing data, the given weight for each variable etc. Issues related to generating a composite indicator include: • The similarity process of the composite index elements: There are many steps that may be considered to start the measurement of the index, including making sure that the available variables are appropri- ate, that the variables are sufficient and well defined to characterise and explain all elements, and suitable to develop a new measurement index to examine progressive changes in a certain domain. • The clustering process: Clustering can be applied based on the com- parison of the similarity between different elements involved in the composite index using different clustering techniques. The cluster- ing process could use the distance measurement methods such as Eu- clidean, Squared Euclidean or City Block etc. Researchers have developed many techniques to handle the measurement of changes in different domains; however, in this research, only statistical aggregation techniques are used. There are two major types of aggrega- tion technique: additive (linear) and multiplicative (geometric or nonlinear) (Zhou et al., 2010). Additive aggregation technique can offer compensability between aggregated indicators; if an indicator has poor performance, then an indicator that has a significant value can cover the poor one, which may result in the creation of a biased composite indicator. 7 Figure 1: Proposed framework to create the Human Behaviour Indicator. Using weighted methods to develop a statistical composite indicator is crucial to illustrate the results in a single variable outcome, and choosing the best strategies for weighting variables is still the focus of researchers. Equal Weights (EW) is recommended as the standard technique for constructing composite indicators and (Hopkins, 1991) has used EW in building compos- ite indicators. However, some authors disagree with this approach, such as Cherchye et al. (2007), who raised many issues with using EW. They ar- gued that this could result in composite indicators using a “fundamentally flawed” method. 3.2. The Construction of the Human Behaviour Indicator The schematic diagram of the proposed framework to create the Human Behaviour Indicator (HBI) is shown in Figure 1, and comprises three phases: • In the first phase, the ADL or ADW data is collected and pre-processed. This includes the handling of missing data and the extraction of the features that will be involved in representing the holistic view of the activity. Each feature represents specific information from a single activity; for example, the start time and the duration of each event. 8 • In the second phase of the proposed framework, the data are analysed using trend analysis techniques. • Finally, in the third phase of development, the data created from the trend analysis is standardised to a uniform unit of measurement and then aggregated. The boundaries of the three levels of the holistic view of the monitored person are also generated. These levels are the level of normal activity behaviour, the warning level to represent small changes in activity behaviour, and the abnormal level to detect the outlier activity behaviour. 3.2.1. Date Pre-Processing Data preparation (pre-processing) is an important step because inade- quately controlled methods of data-gathering can result in out-of-range val- ues, impossible data combinations or missing values. Analysing data that is not carefully screened can produce misleading results. Thus, the represen- tation and quality of data comes first, before running the analysis. There are many statistical methods to deal with missing and outlier data; we need to consider which method is suitable to our data. The common methods in- clude: listwise deletion, pairwise deletion, mean/mode substitution, dummy variable adjustment, regression imputation, maximum likelihood, and mul- tiple imputation (Peugh and Enders, 2004) Data examination and exploration before performing and interpreting analysis are very important; therefore, discussing ways to evaluate and un- derstand missing data is the first step that should be taken after formatting the data. The following steps are the principal options for analysing missing data: • Analyse only available data (i.e. ignore the missing data) • Impute the missing data by replacing values, and treat these as if they were observed (e.g. impute the mean, impute based on predicted values using analysis methods) • Impute the missing data and consider that these were imputed with uncertainty (e.g. multiple imputations, simple imputation methods with adjustment to the standard error) • Assume by using statistical models to process the missing data based on their relationships with the available data. In this research, the pairwise deletion method was used, as it was suitable to process our data. It was used to improve the validity of the research 9 results and to reduce the waste of resources caused by missing data. The basic idea for this method is to analyse all cases in which the variables of interest are present, and it attempts to minimise the loss that occurs in listwise deletion. This method keeps as many cases as possible for each analysis and uses all possible information with each analysis. On the other hand, it cannot compare analyses because the sample is different each time (Peugh and Enders, 2004). Table 1 shows samples of missing data in our data sets; therefore, we did not discard the whole data, and only the missing value was discarded from the analysis of the data set. 3.2.2. Trend Analysis In this research, we investigated whether the use of trend analysis tech- niques to smooth data gives better results when interpreting human be- haviour. It is found that the moving average (MA) is a technique that can be used to gain an overall idea of the trends in a data set; it is useful for forecasting long-term trends (Droke, 2001). The MA creates a series of av- erages of different sub-sets of the full data set and can be calculated for any period. For example, if we have several years’ worth of data from monitor- ing an elderly person in his/her smart home, then a moving average of days, months or years can be calculated. Moreover, using long-term monitoring data to calculate a moving average will help us to see trends in this person’s behaviour and forecast his/her well-being situation. Table 1: A sample of real data including missing data. Smart Home Date and Time Sensor Type Location Day1 , 09 : 04 : 17.656 PIR Living Room Missing Data PIR Master Bedroom Day1 , 11 : 06 : 03.796 PIR Kitchen Day2 , 10 : 04 : 17.656 Missing Data Living Room Day2 , 10 : 14 : 10.050 PIR Living Room Day3 , 11 : 45 : 10.766 PIR Master Bedroom Day4 , 12 : 06 : 03.996 PIR Kitchen Smart Office Date and Time Sensor Type Location Day1 , 09 : 04 : 17.656 PIR Chair Day2 , 09 : 05 : 03.766 Missing Data PC Day3 , 10 : 30 : 03.796 Door Out 10 However, the MA has different versions; in this research, it is found the exponentially weighted moving average (EWMA) to be the best technique applied to detect trends in ADL/ADW data. EWMA, also known as an ex- ponential moving average (EMA), is an infinite impulse response filter that applies a weighting factors (for each older datum) that decreases exponen- tially and never reaches zero (Holt, 2004)(Gardner, 2006). The EMA for a series P may be calculated recursively: EMA1 = P1 (1) for t > 1 EMAt = β.Pt + (1 − β).EMAt−1 (2) where: • β represents the degree of weighting decrease, which is a constant smoothing factor between 0 and 1. A higher β discounts older obser- vations faster, • Pt is the datum point value at time t, • EMAt is the value of the EMA at time t. EMA1 can be initialised in different ways, most commonly by setting EMA1 to P1. It can be initialised by setting EMA1 to an average of the first four or five observations. It is very important to initialise EMA1 because it affects the resultant MA depending on the β values. Choosing small values of β make the choice of EMA1 relatively more important than large β values. Higher β values will discount older observations faster (Handbook, 2013). EMAt as a weighted sum of datum points Pt is presented in Equation 3: EMAt = β × (Pt−1 + (1 − β) × Pt−2 + (1 − β)2 (3) ... × Pt−3 + (1 − β)k × Pt−(k+1)) + ...(1 − β)k+1 × EMAt−(k−1) for any suitable k = 0, 1, 2, ..., the weight of the general datum point Pt−i is β(1 − β)(i−1). This can be expressed as Equation 4 to show the steps of EMA towards to the latest datum point, using a proportion of the difference (each time). EMAnew = EMAold + β × (Pcurrent − EMAold) (4) 11 3.2.3. Multivariate Aggregation To identify the overall changes and, ultimately, to forecast changes in the ADL or ADW, a new indicator, the Human Behaviour Indicator (HBI), is introduced. This stage has two main tasks; the first task is to build the composite indicator, which will compute the progressive changes in the behaviour based on the events that are performed in the whole day. The events are the actions that the participant performs during his/her activity. Each event has a start time and a duration. For example, when the person works at the computer and sits on the chair, we will compute the start time of using the computer until he/she stops using it as an event, and, at the same time, the start time of using the chair until he/she leaves it as an event. This idea of using the events gave us the opportunity to work with overlapped events and activities. Therefore, we were able to compare or measure the changes of events on different bases such as daily or weekly. In this research, we used the percentage of changes between the events when compared to each other as the best way to measure the changes. However, this task had three steps, as follows: • There are different ways in which we could aggregate human behaviour events and to calculate the percentage of change of the compared events. As part of our investigation, it is found that Weighted Human Behaviour Indicator (WHBI) expressed in Equation 5 and the Expo- nential Weighted Human Behaviour Indicator (EWHBI) expressed in Equation 6 are the best candidates for data represented in this paper. Each equation has its way to reach the final results but both of them gave the same final results. Therefore, either of them could be used as a start point of the calculations. • Calculate Pri = ∑ (Pi+n) ∗ 2∑ (Pi) ∗ (Pi+n) . (5) • Calculate Pri = (α) ∑ (Pi+n) ∗ 2 (1 − α) ∑ (Pi) ∗ (Pi+n) . (6) where: P is current data point value after using EWMA (i.e. start time, duration) of the activity for i=1,2,3,.. and n is the number of data points used to calculate EWMA. • Calculate the summation of changes in events based on all daily events. 12 totald = ∑ Pri. (7) where: d is the number of each day and i=1,2,3,.. represents the percentage of change of each compared event of that day. • Calculate the percentage of total changes of all daily events. Indexi = (totald/totald−n) × 100. (8) where: i,d=1,2,3,.. and n is the number of data points that are used to calculate EWMA. The second task is to calculate the boundaries of the changes. In other words, the index is centralised to 100, because the indicator will present the percentage of the changes in the overall of daily activities after calculating the percentage of change between the values of compared events’ features (i.e start time and duration of each event). The system has three boundaries on each side of the centre value (100). These boundaries represent three main categories: normal, warning and abnormal. The following steps show the process of calculating the indicator and the boundaries: • Calculate the average of all Indexes’ values. avg = ∑ Index/N. (9) where: N is the total number of data points. • Calculate the normal boundary. NB = 100 ± Avg. (10) where: NB is the normal boundary. • Calculate the warning boundary. WB = AV ± NB. (11) where: WB is the warning boundary and AV is the accepted value to represent the warning situation. • Calculate the abnormal boundary. AB = WB ± Pt. (12) where: AB is the abnormal boundary and Pt is a data value at time t. 13 4. Case Studies and Discussion In this section a brief description of data used in this research is provided before the proposed method of formulating the indicator is introduced. The data sets used for this research are collected from within intelligent environ- ments equipped with some sensors to measure indoor activities of an occu- pant. The indoor movements are represented by the sequences of movement from one place to another. Monitoring the movements of a person occupying the environment and collecting the data representing his/her mobility can detect this person’s transitions. Table 1 shows the format of the data summarisation, which represents two people, each occupying an intelligent environment. The first person occupies a smart home and includes examples of his/her activates based on his/her sequential movements during the occupancy of a specific location. Each behaviour can be represented in two order levels. The first level is the activity sequence, which shows the person’s movements. The second level is the sequence of actions, which shows the occupancy time of the activity that the person has performed. For example, walking to the kitchen is a moving activity and sitting in the living room is an occupancy activity. Activities in a smart home are presented by five main categories: “Bed- room”, “Living Room”, “Kitchen”, “Bathroom” and “Hall” activities, rep- resenting sleeping, socialising, eating, cleaning and moving between places, respectively. Activities in a smart office are presented by four main cate- gories: “Chair”, “PC”, “PIR” and “Door” activities, representing sitting, working, occupancy and going in/out, respectively. In this section, three separate case studies are reported. Two of these case studies have real data sets collected from intelligent environments. The first data set represent the ADL of an elderly person who lives in a smart home and the second data set represents the ADW of a person who works in a smart office. In the third case study, a simulated data set is used; this data set simulates an elderly person who lives in a smart home. Real data were collected using wireless motion sensors, PIR and door sensors. The output values of these sensors are discrete. In the smart home these values represent the occupancy of one area at a time, whereas in the smart office, the values may overlap because of the possibility of using different equipment at the same time; for example, the participant may use the PC while sitting on the chair and occupying the desk. However, the results and discussion presented below are based on one participant in each intelligent environment. 14 4.1. Case Study 1: Combine all events of ADLs extracted from real data of a monitored person who lives in a smart home This study aims to establish whether the proposed technique can show the overview of a person’s status and detect abnormality within the be- haviour of a participant who lives in a real environment. In this case study, the data represents all events of ADLs that are completed in different areas of a smart house. To demonstrate the holistic view of the monitored per- son’s activity status based on the sensor network measurement, the data sets are represented in separate groups. Each group shows the start time and duration of each event in one area. These events represent the occupancy activities in each area of the smart home. For example, Figure 2 and Figure 3 show the start time and duration time of events that represent the occupancy activities for the bedroom and kitchen, respectively. In both figures, a sample of only three days is depicted. These days are the same weekday of three weeks. It is very difficult to see each area’s events in one graph if we are hoping to understand the overall behaviour changes or to interpret the behaviour; however our technique can display the holistic view of the person’s behaviour in one graph. For instance, Figure 4 shows the status of a person based on the collected data for 140 days. As can be seen from Figure 4 in the upper graph, the participant is often in a good situation. Frequently, the percentage of changes in this person’s behaviour are around the average of his/her daily routine. However, on some days we can see abnormality in the percentage of changes, which goes beyond the second boundaries that represent the red area in this graph. To explain what led to this result, the lower graph in Figure 4 shows the number of times the participant visited the areas and the total duration spent in each area, alongside the mean of number of visits and the total duration for each area. We thus have at least an idea of the differences in the patterns of normal and abnormal behaviour, as result of using our technique to interpret the overview of the participant’s behaviour in an appropriate graph of all events that occurred in all areas in the home. In addition, this technique makes it possible to identify the abnormality from low-level sensory signals even if we do not have detailed knowledge of the subject. Table 2 shows a sample of the real data recorded in a smart home, which represents the three level that are used in our technique. This information shows that our technique can measure the changes of human behaviour. 15 Figure 2: Sample of events in the Bedroom over three days. Figure 3: Sample of events in the Kitchen over three days. 16 Figure 4: Holistic view of an elderly person’s behaviour in a smart home. 4.2. Case Study 2: Combine all events of ADWs extracted from real data of a monitored person who works in a smart office The aim of this case study is to investigate whether the proposed tech- nique can show an overview of the person’s status and detect abnormality within the behaviour of a participant who works in a real environment. In this case study, the data represents all events of ADWs that are performed in different areas in a smart office. To demonstrate the holistic view of the monitored person’s activity status based on the sensor network measure- ments, the data set is represented in separate groups. Each group shows the Table 2: A sample of the results of using HBI with the smart home’s data. Places Normal Warning Abnormal T.D N.V.P T.D N.V.P T.D N.V.P Bedroom 561 34 488 35 300 30 Kitchen 203 51 217 56 175 55 Bathroom 97 15 154 11 278 13 Hall 102 69 121 82 113 74 Living Room 477 56 458 72 572 73 T.D: Total duration, N.V.P: Number of visiting a place 17 start time and duration of each event in one area. These events represent the occupancy activities in each area in the smart office. It is very important to notice that the data sets overlapped in their times. For example, when the person is using his/her PC, the PC’s sensor will trigger; at the same time, the chair’s sensor will trigger. Figure 5 and Figure 6 show the start times and durations of events that represent the occupancy activities for the desk and the chair, respectively. In both figures, only a sample of three days is depicted. These days are the same weekday in three weeks (e.g. Mondays). Applying our technique to such data can help to understand the overall behaviour changes or the behaviour interpretation by displaying the holistic view of the person’s behaviour in one graph. For instance, Figure 7 shows the status of a person based on the collected data for three months. As can be seen in the upper graph of Figure 7, the participant is using his/her office normally. However, on some days an abnormality occurs and we can see the change in the percentage of changes levels, which goes beyond the second boundaries that represent the red area in this graph. To explain what led to this result, the lower graph in Figure 7 shows the number of times the participants used the equipment in the office and the total duration of each use, alongside the mean number of usage and the total duration of use of each piece of equipment. Therefore, we have at least an idea of the differences in the patterns in the normal and abnormal behaviour. Table 3 shows that we gained important results from using our technique with data from multiple sensors that sometimes overlapped. The HBI shows the progressive changes in the participant’s behaviour and interprets it in one relevant graph of all events using all sensors in the office. In addition, this technique makes it possible to identify the abnormality from low-level sensory signals, even if we do not have detailed knowledge of the subject. Table 3: A sample of the results of using HBI with the smart office’s data. Places Normal Warning Abnormal T.D N.T.S T.D N.T.S T.D N.T.S Chair 401 83 338 54 303 9 Duration 404 56 513 60 595 67 PC 401 449 291 213 44 48 T.D: Total duration, N.V.P: Number of triggering a sensor 18 Figure 5: Sample of events using the desk over three days. Figure 6: Sample of events using the chair over three days. 19 Figure 7: Holistic view of a person’s behaviour in a smart office. Figure 8: Sample of simulated events in the Bedroom over three days. 20 Figure 9: Sample of simulated events in the Kitchen over three days. Figure 10: Holistic view of simulated activities of an elderly person in a smart home. 21 4.3. Case Study 3: Combine all events of ADLs extracted from simulated data of an elderly person In this case study, simulated data prepared in our previous work (El- bayoudi et al., 2015) were used to demonstrate that the proposed technique can detect abnormalities in simulated data, as well to show the overview of the person’s status. In this case study, the data represents all events of a less mobile person’s ADLs that are simulated to be carried out in differ- ent areas in a smart house. The simulated data are grouped into events to demonstrate the holistic view of the monitored person’s activities. Each group shows the start time and duration of each event in one area. These events representing the participant’s activities in each area in the simulated smart home. For example, Figure 8 and Figure 9 respectively show the start time and duration time of events that represent the occupancy activities for the bedroom and the kitchen. In both figures, a sample of only three days is depicted. These days are the same weekday in three weeks (e.g. Mondays). However, it is very difficult to understand the overall situation of the monitored person by observing each area’s events graph, but we can easily understand the overall behaviour changes or the behaviour interpretation if we have a single graph to show the changes. For instance, Figure 10 shows the progressive changes of a person’s behaviour based on the simulated data for around 150 days. As can be seen from Figure 10 in the upper graph, the participant has some abnormal days based on the percentage of changes in this person’s behaviour. On the other hand, we can see that this person has a more stable situation because the frequency of the percentage of changes in this person’s behaviour are around the average of his/her daily routine. To explain what led to this result, the lower graph in Figure 10 shows the number of visits to the areas and the total duration in each area, alongside the mean of the number of visits and the total duration of each area used as a scale of Table 4: A sample of the results of using HBI with the simulated smart home’s data. Places Normal Warning Abnormal T.D N.V.P T.D N.V.P T.D N.V.P Bedroom 640 13 673 14 701 14 Kitchen 96 19 102 18 104 18 Bathroom 64 10 48 9 48 10 Go Out 79 1 65 1 49 1 Living Room 560 21 550 19 537 22 T.D: Total duration, N.V.P: Number of visiting a place 22 changes. Therefore, we have at least an idea of the differences in the patterns of normal and abnormal behaviour. Again, in this case study, we can see from the results from using our technique, that it can be used to interpret the overview of the participant’s behaviour in an appropriate graph of all events that occurred in all areas in the home. In addition, this technique makes it possible to identify the activity levels that are proposed in our technique (normal, warning and abnormal) from low-level sensory signals as it is shown in Table 4. This study has par- ticularly measured the similarity in binary data sets which represent daily activities. The investigation presented here aims to assess the monitoring and measurements that can identify the changes in the pattern of human behaviour, and to distinguish between normal and abnormal behaviour pat- terns of participants. The HBI technique is proposed and applied to binary time series to measure progressive changes in human behaviour and to de- tect abnormal behaviour. The HBI allows the supervisor/carer to observe any changes to the regular pattern on a daily basis and it can interpret and explain the observed changes. The results from HBI are tested and applied to real data sets in two case studies and they confirm the suitability of the recommended indicator. The major findings of this research in terms of measuring progressive changes in human behaviour activities using binary time series data are listed below: 1. Very limited research is reported in the literature on measurement of changes in human behaviour and analysis of binary time series, in particular for analysing real binary data sets collected from intelligent environments. This research has helped with better interpretation and understanding of binary time series using the investigated techniques. 2. Real data sets collected from sensors in intelligent environments could lead to a large amount of data. This binary data is hard to work with and needs to be converted into time series using proper techniques that must not be misleading about the results. 3. Using EWMA has helped to be able to detect trends in the partici- pant’s behaviour and predict the future of his/her behaviour. 4. The demonstrated results indicate that the combination of prediction and trend analysis techniques produced a better understanding of time series. 23 5. The importance of the extraction of daily events from the raw data (time sequence) as inputs to the used techniques is investigated. 6. The impact of the aggregation of daily events and observation of the behaviour changes using HBI technique is investigated. This technique gives clear results of behavioural changes and interprets them in a suitable graph. The HBI gives superior results even when data are overlapped such as data collected from a smart office. 7. Overall, this research could play a vital role in the fields of observing human behaviour changes with more development. 5. Conclusions and Future Work In this paper, a novel technique for measuring progressive changes in a person’s behaviour presented in the form of ADL/ADW is introduced. The proposed approach generate an indicator referred to as HBI which represent multiple activities of a person. The overall progressive changes and HBI could also be depicted in a single graph. Furthermore, the results presented here show that HBI is a very promising approach to interpret binary data sets collected from intelligent environments. The data sets investigated in this paper were based on single participants who occupied a house equipped with appropriate sensors. The results presented here show that data gath- ered from ambient sensors can provide important information about the patterns and behaviour of the participant. A direct comparison of our proposed approach with other techniques e.g. clustering techniques may not be possible. Most of the existing ap- proaches deal with each activity individually. For example, if we classify the data gathered from sleeping activity only, it may represent abnormality in the sleeping activity. However, when this is studied with other activities including eating (kitchen activity) the overall activity may not be abnormal. Further investigation, in which further work could be undertaken, is the implementation of the approach presented in this paper for multiple users in intelligent environments. The current approach is not tested using data sets representing multiple occupants (when visitors or carers are visiting) or when pet animals are involved. This study will serve as a base for future studies in which a combination of data sets from several sensors can be used. The current study has only examined human behaviour changes for a single user without any disruption to the normal routines. Building an easily interpreted user interface for 24 our system to train and test other techniques is part of our future plan. The progressive changes techniques used in this research could be extended to measure more complex human behaviour. The indicator presented here could also be used as a tool for predicting an activity or the overall behaviour of the user. References Aran, O., Sanchez-Cortes, D., Do, M.-T., Gatica-Perez, D., 2016. Anomaly detection in elderly daily behavior in ambient sensing environments. In: Chetouani, M., Cohn, J., Salah, A. A. (Eds.), Human Behavior Under- standing. Springer International Publishing, Cham, pp. 51–67. Bagher, R. C., Hassanpour, H., Mashayekhi, H., 2017. User trends modeling for a content-based recommender system. Expert Systems with Applica- tions 87, 209 – 219. Chattopadhyay, K., Douiri, A., Sheldenkar, A., Wolfe, C., Rudd, A., Chen, R., 2013. Trends in activities of daily living among stroke survivors: Anal- ysis from the south london stroke register. International Journal of Reha- bilitation Research 2 (2), 6–21. Cherchye, L., Kuosmanen, T., et al., 2002. Benchmarking sustainable devel- opment: A synthetic meta-index approach. Documento preparado en el contexto del programa de investigación en Métodos No paramétricos en Enomomı́a de la Producción, los recursos Naturales y el Medio Ambiente. Cherchye, L., Lovell, C. K., Moesen, W., Van Puyenbroeck, T., 2007. One market, one number? a composite indicator assessment of eu internal market dynamics. European Economic Review 51 (3), 749–779. Cook, D. J., Augusto, J. C., Jakkula, V. R., 2009. Ambient intelligence: Technologies, applications, and opportunities. Pervasive and Mobile Com- puting 5 (4), 277–298. Droke, C., 2001. Moving averages simplified. Marketplace Books. Elbayoudi, A., Lotfi, A., Langensiepen, C., Appiah, K., 2015. Modelling and simulation of activities of daily living representing an older adult’s behaviour. In: Proceedings of the 8th ACM International Conference on PErvasive Technologies Related to Assistive Environments. ACM, p. 67. 25 Forkan, A., Khalil, I., Ibaida, A., Tari, Z., 2015a. Bdcam: Big data for context-aware monitoring-a personalized knowledge discovery framework for assisted healthcare. IEEE transactions on cloud computing. Forkan, A. R. M., Khalil, I., Tari, Z., Foufou, S., Bouras, A., 2015b. A context-aware approach for long-term behavioural change detection and abnormality prediction in ambient assisted living. Pattern Recognition 48 (3), 628–641. Furlaneto, D. C., Oliveira, L. S., Menotti, D., Cavalcanti, G. D., 2017. Bias effect on predicting market trends with emd. Expert Systems with Applications 82, 19 – 26. Gardner, E. S., 2006. Exponential smoothing: The state of the artpart ii. International journal of forecasting 22 (4), 637–666. Handbook, E. S., 2013. Single exponential smoothing. URL http://www.itl.nist.gov/div898/handbook/pmc/section4/pmc431.htm He, Z., Jin, L., 2009. Activity recognition from acceleration data based on discrete consine transform and svm. In: Systems, Man and Cybernetics, 2009. SMC 2009. IEEE International Conference on. IEEE, pp. 5041–5044. Holt, C. C., 2004. Forecasting seasonals and trends by exponentially weighted moving averages. International journal of forecasting 20 (1), 5– 10. Hopkins, M., 1991. Human development revisited: A new undp report. World Development 19 (10), 1469–1473. Kearns, W. D., Fozard, J. L., Nams, V. O., March 2017. Movement path tortuosity in free ambulation: Relationships to age and brain disease. IEEE Journal of Biomedical and Health Informatics 21 (2), 539–548. Kim, T.-Y., Cho, S.-B., 2018. Web traffic anomaly detection using c-lstm neural networks. Expert Systems with Applications 106, 66 – 76. Lara, O. D., Labrador, M. A., 2013. A survey on human activity recogni- tion using wearable sensors. Communications Surveys & Tutorials, IEEE 15 (3), 1192–1209. Liang, Y., Welmer, A.-K., Möller, J., Qiu, C., 2017. Trends in disability of instrumental activities of daily living among older chinese adults, 1997- 2006: population based study. BMJ Open 7 (8). 26 Liu, C.-D., Chung, Y.-N., Chung, P.-C., 2010. An interaction-embedded hmm framework for human behavior understanding: With nursing en- vironments as examples. Information Technology in Biomedicine, IEEE Transactions on 14 (5), 1236–1246. Lotfi, A., Langensiepen, C., Moreno, P. A., Gómez, E. J., Chernbumroong, S., 2017. An ambient assisted living technology platform for informal car- ers of the elderly-icarer. EAI Endorsed Transactions on Pervasive Health and Technology 3 (9), 152393. Mabrouk, A. B., Zagrouba, E., 2018. Abnormal behavior recognition for intelligent video surveillance systems: A review. Expert Systems with Applications 91, 480 – 491. Mahmoud, S., Lotfi, A., Langensiepen, C., 2014. User activities outliers detection; integration of statistical and computational intelligence tech- niques. Computational Intelligence. Mahmoud, S. M., Akhlaghinia, M. J., Lotfi, A., Langensiepen, C., 2011. Trend modelling of elderly lifestyle within an occupancy simulator. In: Computer Modelling and Simulation (UKSim), 2011 UkSim 13th Inter- national Conference on. IEEE, pp. 156–161. Nweke, H. F., Teh, Y. W., Al-garadi, M. A., Alo, U. R., September 2018. Deep learning algorithms for human activity recognition using mobile and wearable sensor networks: State of the art and research challenges. Expert Systems with Applications. Ordóñez, F. J., de Toledo, P., Sanchis, A., Feb 2015. Sensor-based bayesian detection of anomalous living patterns in a home setting. Personal and Ubiquitous Computing 19 (2), 259–270. Peugh, J. L., Enders, C. K., 2004. Missing data in educational research: A review of reporting practices and suggestions for improvement. Review of educational research 74 (4), 525–556. Rafael-Palou, X., Vargiu, E., Miralles, F., 2015. Monitoring people that need assistance through a sensor-based system: Evaluation and first results. In: Proceedings of AI-AM/NetMed 2015 Artificial Intelligence and Assistive Medicine (AIME 2015), CEUR Workshop Proceedings. Vol. 1389. pp. 22– 31. 27 Ros, M., Pegalajar, M., Delgado, M., Vila, A., Anderson, D. T., Keller, J. M., Popescu, M., 2011. Linguistic summarization of long-term trends for understanding change in human behavior. In: Fuzzy Systems (FUZZ), 2011 IEEE International Conference on. IEEE, pp. 2080–2087. Saives, J., Pianon, C., Faraut, G., 2015. Activity discovery and detection of behavioral deviations of an inhabitant from binary sensors. Automation Science and Engineering, IEEE Transactions on 12 (4), 1211–1224. Suryadevara, N., Mukhopadhyay, S. C., Wang, R., Rayudu, R., 2013. Fore- casting the behavior of an elderly using wireless sensors data in a smart home. Engineering Applications of Artificial Intelligence 26 (10), 2641– 2652. Suryadevara, N. K., Mukhopadhyay, S. C., June 2012. Wireless sensor net- work based home monitoring system for wellness determination of elderly. IEEE Sensors Journal 12 (6), 1965–1972. Wallace, B., Knoefel, F., Goubran, R., Masson, P., Baker, A., Allard, B., Guana, V., Stroulia, E., 2017. Detecting cognitive ability changes in pa- tients with moderate dementia using a modified ’whack-a-mole’ game. IEEE Transactions on Instrumentation and Measurement PP (99), 1–14. Wang, X., Xu, Y., Hu, H., Liu, M., Li, G., 2018. Feedback-based metric learning for activity recognition. Expert Systems with Applications. Xie, B., Wu, Q., 2012. Hmm-based tri-training algorithm in human activity recognition with smartphone. In: Cloud Computing and Intelligent Sys- tems (CCIS), 2012 IEEE 2nd International Conference on. Vol. 01. pp. 109–113. Yin, J., Yang, Q., Pan, J. J., 2008. Sensor-based abnormal human-activity detection. Knowledge and Data Engineering, IEEE Transactions on 20 (8), 1082–1090. Yu, R., Wong, M., Chang, B., Lai, X., Lum, C. M., Auyeung, T. W., Lee, J., Tsoi, K., Lee, R., Woo, J., 2016. Trends in activities of daily living disability in a large sample of community-dwelling chinese older adults in hong kong: an age-period-cohort analysis. BMJ Open 6 (12). Zhan, A., Little, M. A., Harris, D. A., Abiola, S. O., Dorsey, E., Saria, S., Terzis, A., 2016. High frequency remote monitoring of parkinson’s disease via smartphone: Platform overview and medication response detection. arXiv preprint arXiv:1601.00960. 28 Zhou, P., Ang, B., Zhou, D., 2010. Weighting and aggregation in composite indicator construction: A multiplicative optimization approach. Social Indicators Research 96 (1), 169–181. 29 
work_twk2q7jfhngdtbzulmfas2oc7e	----	MANAGING EXPERT SYSTEMS : A Framework and Case Study Rob R. Weitz Department of Information, Operations, and Management Sciences Leonard N. Stern School of Business, New York University 44 West 4th Street, New York, NY 10012 Center for Digital Economy Research Stern School of Business Working Paper IS-90-06 Abstract ?his paper addresses the problem of managing the development and implementation of a large expert system in an organization. A traditional systems analysis and design methodology is used as a framework to highlight similarities and differences in managing large scale traditional computer based projects and large expert systems. As a non-technical, prescriptive guide, this article focusses on defining at each stage in the project, the tasks to be accomplished, resouras required, impact on the organization, likely benefits and potential problems. The case of a large expert system ' implemented by a multi-national corporation'across several Ehropean sites is used to clarify and expand upon the mgement guidelines pmvided. Introduction Research in the field of Artificial Intelligence (AI) signals great p h s e for the next generation of hardware and software. At the present time, expert systems are arguably the most commercially successful product of Artificial Intelligence research; they have crossed the threshold of the laboratory and are beginning to mke their presence felt in real world applications. While others have described applications, suggested opportunities for the technology, or emphasized organizational considerations, (Wnard-Barton and Sviokla 1988, Luconi et al. 1986, Leonard-Barton 1987), to date the management of their introduction into the workplace and their inpact once there still remain largely unexplored. This paper presents a prescriptive guide for managing large scale expert systems from ineption 4Arough implementation and maintenance. It is motivated by experience with the developmeslt of a sizable expert system by a large, multinational ampany, the cgrcwing literature of cases describing expert systems in practice, and the belief that management and organizational considerations (as opposed to m i c a 1 wizardry) must remain paranmunt for Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 the success of such systems to be achieved. The thrust of this paper is to focus awareness on 1) the processes and resouras required for an expert system project, 2) the costs and benefits of such an undertaking, and 3) organizational and task changes likely to result from the intrcduction of the system. This paper is not a technical guide for building expert systems; technical concerns are expressed here only insofar as they are inextricably linked to the management of such systems. A great deal of experience has been gathered regarding the design, implementation and maintenance of sttraditional" computer based systems, from both technicdl and management viewpoints. The pitfalls, players and positive practices have been identified in a large bcdy of existing research, and are well described in the information systems literature. (See Senn 1984, and Burch and Grudnitski 1986, for example.) The management of expert systems does not lie completely independent from this previous computer based project management qrience. Indeed, one trend is towards whisiblen expert systems - that is, expert systems which are integrated into conventional data processing hardware and software (Kozlov 1988) . This paper builds on the existing groundwork by taking as a framework a traditional systems analysis and design (SA&D) methodology and adapting it for application to expert systems. Ekpert systems (ES) are computer prqmms for solving difficult, ssfuzzyss problems in dcnnains where human expertise is nonrdlly associated with a great deal of training and experience. Application areas to date include such areas as fault diagnosis, tax planning, credit evaluation, geological prospecting, chemical analysis and medical diagnosis. (See Kupfer 1987 and Kneale 1986 for K>pular-press descriptions of expert systems qperating in business environments; Waterman 1986 provides an extensive Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 werview of ES applications.) Expert systems are typically characterized by: The utilization of larye amounts of damain specific knowledge. The ability to use incomplete or uncertain information. The capacity to explain their behavior (a kind of self-knowledge). Symbol manipulation, that is ttreasoning" about obj&, as opposed to numerical dpulation ( w h i c h typifies traditional computer programs). Perfor'mne levels at, or exceeding, those of experts in the problem domain. Typically, sizeable e x p e r t system are built in an iterative, incremental fashion via repeated interviews between one or more damain experts and a wknowlPdqe engineertt. Briefly, the task of the knowledge engineer is to elicit the e x p e r t knowledge, map the knowledge into a suitable structure and actually code the knowledge using appropriate software and hardware. The process tends to be tedious and the consuming, and has in fact been referred to as the ttknawledge engineering bottleneckw. A good overview of experts systems in general, and of the building process is provided by Hayes-Roth et al. (1983) and Hanmn and King (1985) . For E3 the i m p o m of early development of working prototypes is stressed. The prototype is haementally improved and its capabilities expanded by repeated trials with the damain expert and actual use in a test environment. In fact, several versions of the prototype are typically successively developed, until a sufficiently evolved version is realized for possible release. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Wzilding and implerrrenting a larye expert system is both time consuming and resaurce intensive. While improved software envirommts have helped speed developnat, and recent experience has suggested some guidelines for easing the overall process samewhat, existing verifiable ES applications suggest that, for large systems, the time r e q x b d to go from prototype to implementation is typically on the order of person-years, with costs measured in at least the tens of thousands of dollars. For a case in point, see Linden (1982). Clearly, larger systems (as measured by the amount of knmledge acquired and by the number of users) require more effort than smaller ones. In any case, it seems that under the appropriate circumstances the value of autmted expertise supports the magnitude of these efforts. 3. A mditioml Systems Analysis and Design FYamework Creating a larye computer based system for more than personal use is a campla task requiring technical skills, creativity and good management of resources. A guide for this process has been established through experience and while details vary f m source to source, the overall thrust is fairly standardized in the information systems literature. One such systems analysis and design framwork is provided in Figure 1, and is due to Iucas (1982). As indicated previously, this paper adapts the traditional SA&D procedures for use with expert systems. Figure 1. about here. The next section traces this outline as it applies to expert systems; fumhmntal variations of the traditional SA&D process are noted, and described or referenced. The resultant SA&D process for expert systems is summarized in Figure 2. Figure 2. about here.. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 After detailing the SA&D framework for expert systems, a '*real worldn ample is presented. It should be stressed that the case description serves to illustrate the framework and not define it. The Comet eqert system was developed without benefit of the mmtive model outlined here - in fact the project wBs in large part a motivating influence in developnent of the framework. It should be clear that the case description not only clarifies the model, but also serves to demonstrate the benefits of follcxnling the proposed SA&D framework, arid the disadvantages of deviating fmm it. 4. Framework 4.1 Inception Inception refers to the realization that a ccmputer based technology or application can be of value to an organization. The idea may refer to investigating a technology in an exploratory sense, or more specifically applying a technology to an existing or proposed procedure within the organization. At this stage the envisioned system is naturally somewhat murky in its details, though the overall goals - to reduce costs, ease a production bottleneck, maintain a competitive position, or better manage a process, for vie, are more clear. Several variations of possible systems are likely entertained at this point. The question proposed here is, why should the envisioned system (or one of the envisioned systems) be an expert system? It should be kept in mind that e x p e r t systems are not cheaper or easier to build than conventional systems; it is more realistic to assums the contrary, Therefore, there should be strong, positive reasons for proposing an expert system solution. Selecting the right tool (or technology) is always a question of matching the characteristics of the problem with the capabilities of the tool. Criteria for tasks well suited for expert systems Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 have been en-ted elsewhere (Bobrow et 61. 1986, Qm 1987 for example). These criteria are briefly summarized below. Ekprt system are used to replace or assist e x p e r t s . (A rough definition of an expert is someone who can solve difficult problems more quickly and with less effort than a novice, or tlaveragetv person. Typically, expertise is acquired through lengthy training and experience, and is limited for an individual to a particular field of knowledge.) Insofar as technical criteria are concern&, for applying an expert system to a problem, the task should clearly Have semi-structured solution processes. E k p r t systems were created in an attempt to model problems that do not have algorithmic or "step- by-stepw solutions. (Traditional, mathematical models or decision tree programs are appropriate for these type of problems. ) Expert systems are well suite3 for tasks where a b d y of loosely structured knowledge exists; this technology was designed for capturing the heuristics or yules of thwibn for arriving at good solutions. Involve a well c-ibed, limited field of knawledge. Ekpert systems are not general problem solvers. (As a case in point, each ES application in medicine focusses on one specialty; for example, MYCINts damain is bacterial infections, while INTERNIm covers the domain of internal medicine. ) Moreover, problems requiring bmrnon sensem are not well suited for an ES solution. Be difficult, but not too difficult. As a general guideline the task should take a human expert samewhere between minutes and hours. Given the above necessary conditions, an expert system is more strongly remmm-ded for tasks that Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Require explanation to the user of the system's reasoning in solving the problem. Involve reasoning with uncertain or inccarq?lete information. Entail resoning about symbols, or ~tobjectsll, as opposed to mathemtical manipulations. The above criteria are purely technical requirements, and define the minimum considerations for f u r t h e r pursuing the process of tkinking about applying expert system technology. Related to the above technical CO- are more organizational requirements for an expert system to be envisioned as a solution. These include that Human experts exist and are willing and able to participate in the project. Human expertise is scarce and/or expensive. Management is likely to commit the required resources to the project. (This implies that the task is viewed as an important one.) At the inception stage, some indication that these non-technical concerns are satisfied must be sensed, or at least their negation must not be evident. For both traditional and expert systems, the inception stage produces a proposdl outlining the reasons for the project, its goals, expect& benefits, possible risks, alternative approaches, time frame, and required resources. The entire inception process serves to better define the project, and to garner organizational support. The inception phase ends with a decision to either stop consideration of the project, or to continue f u r t h e r investigation and evaluation. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 In traditional SA&D, the next step is a full-fledged feasibility study, (The prototype proposal may in fact be viewed as a "first draftw feasiblity study.) For expert systems, successful campletion of the inception stage means approving the development of a demonstration prototype, and concurrently und- the feasibility study. Typically, the inception report includes an estimate of the resources which will be required, and an approximate schedule for the entire project. However, since approval at the end of the inception phase means cormnitment to the project solely through the feasibility study (for traditional SA&D) or t.hroqh prototyping and feasibility study (for expert SA&D) , the inception report should specify in detail a timetable and the resources required for these next steps. A decision to go ahead with the full system will be made when the feasibility study, or in the case of expert SA&D, the demonstration prototype and feasiblity study are ready. Finally, the inception report for an ES may include evaluation criteria for the prototype. While criteria may be difficult to specify, what constitutes lfsuccessw for the prototype should be defined in advance. (De- * . evaluation criteria is discussed in the next section.) For many organizations, a process for gaining a basic understakiing of what ES technology has to offer, in general or with respect to the application domain, is required in order to campile the inception proposal. lhis may entail formal training, attendance at semhmrs, transfer within the organization of individuals experid with ES technology, contact with ES software firms, or the hiring of consultants. The objective is to acquire the capability of fully assessing the potential project, and to begin to codlesce the resources for its undertakhg. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 4.2. Feasibility Study As with traditional SA&D, the feasibility study is caprid of a comparative evaluation of all conceivable alternative systems. ( ~ t minimum a single new system is proposed and evaluated with respect to the existing procedures.) The ultimate purpose of the study is to select the best of all possible alternatives, but this part of the analysis serves many other pmpses. Among these, it helps solidify ideas, provides a common source of reference, and serves as a focus for gaining camnitment from users and management. The feasibility study is a fuller, more detailed version of the inception report. It should be ready for release when the demonstration prototype is ready, or shortly thereafter. The contents of a traditional SA&D study report are outlined below. Follawing each directive is a description of haw that directive applies to an expert system feasibility study. The report should: Describe the current system including a rationale for dhanging it. As the saying goes, I1If it ain't broke, don't fix it!" Shortcomhgs in the current system and improvements provided by the suggested system should be clearly specified. Ekpert system proposals based on rationales of selecting a "high tech, state-of- the-art solutiont1 should be rejected. Explain why the particular solution is proposed, as opposed to some other alternative; already existing, similar systems both inside and outside the organization should be referenced if possible. Why is an expert system appropriate? The previous section describes problems and envimnments wkich suggest an expert system solution. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Lay out the goals, scope and objectives of the envisioned system. It is important to set realistic expectations for the proposed system. Terms such as ttartificial intelligencett and expert systemtt tend to spur the imagination. Bear in mind that this technology typically cannot do what a human expert cannot do and that, in any case, the breadth of knawledge enampassed by the system will be limited. Specify the bounds of the proposed system; i.e., what are not the goals, objectives and expected capabilities of the system. Describe what the proposed system will do. Include in this description haw the system will work, who will use it, haw the task itself and related tasks will change. The follawing questions should be addressed. Will ttexpertstt themselves be using the system (i.e., doctors using a medical diagnosis system) or will wtechniciansH use the system? What skills will be required of the users? Will the users be new employees or current workers already doing the task under the present system? If current workers, how will their input into system design, and their goodwill in general be solicited? Will the system ttde-skilltt their task, and if so how do they feel about it? Haw will job responsibilities and lines of cxmmmication change? The averall technical specifications should be outlined; i.e., how the system will interface with existing information systems, data needs, input/output mechanisms, performance -ts, etc. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Specify a timetable for the project, including periodic reviews and expected performance of the system at each review. As mentioned above, expert systems are developed incrementally. The first review should occur at the unveiling of the demonstration prototype system. (This review should take place two to eight months from the start of the developtent work.) Feedback from this review should direct M e r technical work on the system, while any organizational difficulties which have surfaced should be aired and attended to. Attendees at the meeting should include those involved with developrent, representatives of the ultimate users of the system, and those responsible for the direction of the project. At minimum, one more review will be required (on the order of 6 months after the first meeting) to decide to either test release, delay or cancel the project. 'I'ypically another review is held after test release, and just prior to full release. Detail the resoumas required, when they will be required and from whom they will be required. For the proposed project, allocations should be included for knowledge engineers, domain experts, programmers, hardware, software, training, and travel cqemes. Note that the participation of domain experts is crucial, while their time is typically at a premium. (Hence the value of automating their expertise.) E?rwisions for re-specifying their job description to include work on the new system should be detailed. Cbmmnly, prqpmms are required to write user, and system to system interfaces. A fuller description of all the players required for such a project and how they differ from those involved in developing a traditional system is given at the end of this section. As with traditional systems, the choice of hardware and Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 software is one of making trade-offs. Indeed, much of the criteria (price, speed, caqatibility, reputation of supplier, for ample) are the same. Hardware selections fall into three categories: micmcmputers, minis/mainframes, and workstations. Special purpose expert system developxnent software (typically called "shellsf* or fterrvirorrments**) exist for each category of hardware and ease the prqramhg task considerably. It is typically the krwwledge engineer's responsibility to choose appropriate h a m h a r e and software for the project. (Gilmore and Huward 1986 and Mettrey 1987 discuss ES software selection. For details on ES hardware and software, and more on the specifics of the knowledge engineering process see Harmon and King 1985, Holsapple and Whinston 1987, Waterman 1986 and Hayes-Roth et al., 1983. ) The reswsces required to achieve the demonstration prototype should be made explicit, as typically after the initial approval, the next Ifgo - no gott decision point is at the prototype review. Provide a cost-benefit analysis. This analysis should include both tangibles and intangibles; typically some estimates are required, but the process should be supported with a sensitivity analysis. Benefits commonly ascribed to expert systems include: preservation and dissemination of scarce w i s e , relieving experts of tedious tasks and thereby allowing them to concentrate on more diff icult/mre interesting problems, speedier solutions and more consistent problem solving. Huber (1984) identifies knowledge as a key, strategic resaurce in the wpost-industrialf* organization. Important costs include: personnel ( i . e. , knowledge engineer and expert) , software and hax&are (perhaps specifically for expert system develapment and use), and those expnses associated with training, operations and updating. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Define, as best as possible, evaluation criteria and agree on how the success of the system will be measured. The evaluation criteria should be utilized during the periodic reviews. Expert system will typically be evaluated on a host of criteria. These include the quality of the solution, the speed of solution, the manner in which the solution is reached (transpaency), breadth of knowledge, explanation and help facilities, user satisfaction, and the ease with which the system can be transported, modified and up%ted. The relative importance placed upon each of these criteria is detemLined by the projectf s goals ' and objectives. The criteria should be established by the ultimate users of the system and the managers of the task in question, in conjunction with the developers of the system. Assign mnership of the system over the course of its lifetime. In many applications, w t i n g of the knowlpdqe encampassed by the system is a large, ongoing job due to the nature of the task in question. Consider, for example, a diagnosis expert system for saw large piece of machinery. Frequent updates will be required if new models of the machine are produced, parts (or part numbers) change, and new faults (and procedures for finding them) appear as the machine ages. Responsibility after release of the system may or m y not rest with the developers, but in any case this responsibility and the mechanisms for und- the maintenance task should be made explicit. Many of these points will be elaborated upon in the subsequent sections. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 4.2.1 Fersonnel Same c a n r e n t s abut the players irrvolved i n expert systems development and hclw they ccglp>are with those for traditional system are i n order. In traditional systems, three major players are identif ied: users, the information system (IS) department, and management. Much has been written of the responsibilities of each of these agents aver the course of the life of the information system project (see Iucas 1982, for example). Briefly stated, ideally the user participates i n the design process and may in fact originate the project idea. The user is best equiped to understand the workings of the present system and therefore to provide input for system specifications, and good test examples. Clearly user satisfaction with the system is an important criterion for success. The IS department reviews the feasibility study, designs the system, specifies possible alternatives and the trade-offs they imply (languages, "off-the-shelf1# or special purpose programs, use of service bureaus, batch versus time sharing mode, hardware, etc. ) , handles the required programming, documentation, training, conversion and maintenance. Management responsibilities are werall approval and direction of the project, and providing ccmnnitment in terms of resources and recognition. Expert system projects include additional players. First, the domain expert is not necessarily a typical user of the system. If the expert is in fact just an llaverage userM, then his/her ir~~olvement in the E S development will, in any case, be much more intimate than that required of a user in traditional systems. Second, the position of knaJledge engineer requires skills that are not necessarily found i n IS departments. Tfiis role may be filled by groomhg in-house personnel, or through external services. These services may be obtained via independent consultants, and expert system software or hardware conp>anies. In traditional system develoyxnent, the IS department has the responsibility for choosing outside vendors and services. The situation is analagous here, with many of the trade-offs (such as reputation of the Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 campany, cost, type of service provided) being similar. Hmever, the IS department m u s t be knowledgeable about ES technology and the market in order to evalulate these trade-offs. F'urther, the IS department will have to work with the external organization to prcanote a smooth interface between &sting systems and pmcdures and the new system. If an outside organization is contracted for developing the system, anmqements for transfering awnersbip to the IS department (and what that crwnership entails) should be specified. 4.3. Prototype The use of prototypes has been referred to several times. Prototype in this sense means a scaled-dawn (in scope, p e r or both) version of the fully errvisioned system. While a working prototype phase may exist in the development of traditional information systems (Alavi 1984, Jamon 1986), it is not an inescapable part of the stamlard system building methodology; further, for traditional systems prototyping may refer simply to creating screen tho& pups" in order to improve the user interface. The formal prototype stage in ES development is suggested k c a u s e 1) it m y be han3 to know what is feasible when it comes to automating expertise without attempting a working, trial version, 2) given the newness of the technology (and the hype that surrounds it) a prototype can help set realistic expectations, 3) a demonstration can serve to garner and solidify enthusiasm for the project, and 4) a phased approach involves less risk. Finally, due to the inaxmental nature of building expest systems and the software developent tools created to support this developent, prototypes may be built rather quickly. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 4.4. Analysis, Design, Specifications, and In developing traditional informtion systems, each of the steps of analysis, design/specification and pmgmmhg is ideally separate and distinct; one does not prcx=eed to a subsexpent step until finished with the previous one. In the analysis phase, the current system is studied; transactions, data volumes, decisions made, etc., are detailed. Design/specification of the new system involves enumerating hardware and software, input and output files, media, pxoxdures for use, security and error control. Writing and testing the program follows the design phase. For expert systems, some analysis, design and detailing of specifications is usually required for the inception phase, and a l q e part of these tasks plus some pragmmbq is completed as part of the prototype system and feasibility report. These processes however are far less differentiated in ES developmnt than for traditional systems. In fact, these processes more closely resembles a paradigm for decision support system development. (See Spmgue and Carlson, 1982, and Keen and Scott Morton, 1978 for example.) The inmemental, iterative prooedure for building ES has already been described. 7his is a function of the fact that experts simply cannot sit d m and fully and completely specify their m problem solving behavior. Ekpert systems are construcked as a collaborative and iterative effort between one or more knowledge engineers and one or more damain experts. The knowledge engineer is experienced in eliciting knowledge from an expert, structuring knawledge such that it can be pzmpmned, and then coding the knowledge. Working frcan the first prototype, the knowledge engineer can improve, refine and expand the capabilities of the system throu* further interaction with the expert, This loop: eliciting, structuring and coding knowledge is trav- many times before a suitable version for release is produced. Here, pragmmbg may be considered to be a function of the analysis, design and specification of the system and visa versa. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 At the beginning of the development of the system, infrequent contacts behem the knmledge engineer and the expert will suffice; as development proceeds, more frequent and more lengthy meetings will be required. In any case, it is crucial that the expert has sufficient time and W t i v e to work with the knowledge engineer on the system. 4.5. Testing For both traditional and expert SA&D, the feasibility study should specify a timetable for release of the system, first as one (or more) trial or test versions, and later as the ggproductgg version. The test version receives limited distribution and is used to fine tune the system for general release. For each release, the system should pass the minimum requbments set up in admnce in the feasibility report. Disagreement and mertahty concerning the evaluation pmedure is likely to exist. Developers m y be most concerned with technical criteria, and users with quality interfaces, while managers worry about how the system will help solve ggbusinessw problems. (For ES even a purely technical evaluation based on the quality of solutions provided by the system may be difficult to conduct. For example, e x p e r t s themselves may have differing opinions as to what constitutes a "rightgg answer.) While disagreement may exist as to haw to evaluate the system, the time to resolve such differences is prior to development. (See Hayes-Roth et al. 1983, for a g o d discussion of evaluatimj expert systems.) In any event, expert systems nust be evaluated under ~rnfltiple criteria. Possible evaluation criteria include whether use of the system: Results in better solutions. Results in faster solutions. Results in more consistent solutions. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Prarides greater worker satisfaction. Is easy to use. Reduces training time (for the humn users). Reduces deperdence on scarce individuals. Has resulted in greater insight into the problem. Reduces the number of extremely bad solutions. These criteria, should be easy to derive fram the goals of the system as stated early in the project. The more difficult part is measuring how the system fulfills these criteria. While same data collection procedures may already exist (the or quality standards for diagnosis tasks, as an example), data collection may have to be initiated as part of the project to allow for a "before and afterw study. While the level at which the system satisfies each criterion m y be flexible, other requirements may be both easy to detemnhe and rather fixed, for example: Time to repair must be reduced by half. Relatively novice technicians must be able to do the task (with the aid of the system) as opposed to the experts who do it now. Cdlls to the resident expert for help on difficult problems must be all but eliminated. The number of people doing the task should be reduced by 10%. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 rXrring testing, the system should be pushed to discover its limits, in terms of the range of its knmledge, but also on very traditional dimensions: security, back-up procedures, and error handling, for example. User ard/or mgement satisfaction with the system can be assessed via formdl questionaires or interviews. Acceptance will be predicated on not just h m well the system does what may be mmmly defined as its task, but also on ease of use, appropriate help facilities, speed, and on how well it supports the user in doing the task as (s)he sees it. A more f?m%mental question regards not the system per se, but rather the individuals for whom the system is designed. Rmwledge workers, may or may not want a machine . looking wer their shoulders while they do kheir jobs, particularly if they view the system as skill or prestige reducing. (Doetors are the classic case in point.) This possibility should be addressed early in system develoyxnent, so that disaffection due to resentment does not show up among users at the evaluation stage. As with traditional systems, both tangible and intangible outcomes will likely need to be measured; release of the system is, of course, contingent on the benefits outweighing the costs. Training, conversion and installation are similar to that for traditional systems. =ropriate doamentation should be developed by the krwwledge engineer in conjunction with the IS department if necessary. Training procedures are set up in cooperation with the users and management. Hopefully, conversion may be phased in, with the date for conversion agreed upon by all those involved. As with traditional systems, the question of ccanpatibility of hardware and software is important. With expert systems, ccanpatibility may focus on the use of AI languages or specialized workstations which need to interface with more conventional hardware and software. Methods for accessing ampmy data, and generally interfacing Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 with existing systems are frequently a necessity. These issues should have been planned for in the design phase. Problen-ts can occur involving the conversion and installation phase of ES if develcpent of the system was performed by personel outside the oryanizationts MIS function. This outside developmt may happen, given that the skills and training requi_rrrl to develop an ES are rather specialized. Cooperation of the MIS group is essential for installing the ES as part of the oryanizationts overall network; their :!buy inff should be managed from the start of the project. 4.7. Operations The operations phase for traditional SA&D involves fixing errors, and when necessary rraking changes to the system. For both traditional informition systems and expert systems, ass* an artside agent has been involved in the developmmt phase, the arrangement should include provisions for ongoing maintenance of the system, in the traditional sense (i . e. , information ffhot-lineff, software revisions, etc. ) . Maintenance duties w h i c h are the responsibility of the IS department should be clarif ied. For ES particular attention must be paid to the maintenance task as these systems work in knowledge intensive areas, and problem solving knowledge in practical applications frequently changes. 'Updating may be required for knowledge proper, for data the system uses, or both. Some asessmmt of the extent of updating should be made early in the project, based on the nature of the task. For the Xa3N system, Digital Ekpipment Corporation's ES for configuring their Vax family of cmpters, saw eight individuals are employed full time on the updating, that is, collecting and em=oding task. This huge effort is due, at least in part, to the fact that the system must constantly be modified to work on new products (McDemtt Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 1984). An estimate of the frequency of knowledge updating and the assignment of updating responsibilities should be part of the project report New knuwl&ge (and alterations to the system independent of the knowledge base) will likely be suggested by users of the system, particularly if they are e x p r t s . Same formal mechanism should be established to capture these suggestions, and system performance in general. Unfortunately the knowledge encapsulated in expert systems cannot be updated simply by adding knowledge. (Autmted updating, that is, clpdating performed entirely by users suggesting new knawledge to the system, is an important area of research in XI. Of course, systems w h i c h learn from their awn mistakes, another research area, would solve this problem.) Aside from having to put the new knowledge into a form the system can understand, the new knowledge must be tested for its effects on the existing knowledge. This is to say that individuals familiar with the details of structure and operation of the system must be involved in the maintenance task, in addition to domain experts. If the system was developed by knowledge engineers outside the organization, updating will have to be performe3 with the assistance of these knowledge engineas. Alternatively, as part of the project sufficient expertise must be brought in-house (i.e., to the IS department) . Again, depending on the nature of the task, its projected future rqmhemmts, and the data involved, maintenance may be a considerable job. Finally, at sane point after the system has been in regular operation, same thought should be given to haw well the system performs. Are the projected cost savings, or productivity i m p m e m e n t s being realized? Does the system satisfy current and projected demands? How well does the project contribute to the overall strategy of the organization? Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Digital Ekpipent Corporation, one of world's largest computer mmufacturers, is perhaps best knuwn for its Vax series of minicomputers. An interml rep* describing expert system technology was circulated within Digital's manufachurhg/repair engineering group in Nijmegen, Holland in early 1985. The report defined the technology and its capabilities, and suggested a range of possible application areas within the group's purview. An ES approach for these applications was w r t e d by the follawing rationales : Knowledge, currently, in and of itself is a vital camrcdity for high technology campanies. Expert systems are a means of managing knaJledge. In particular, the knawlt-dge required for the manufacture and repair of computer hardware will likely increase in scope and complexity as the products themselves becane more complex and product life cycles shorten. A forecasted high demand and increasing cost for human experts. A long term need to increase productivity and to reduce costs. A competitive incentive to maintain and develop expertise in the area of AI. ?he document recarmnends one application in particular, an ES for assisting in the task of haKtware module (tfboardw) diagnosis and repair. ?his damain was suggested firstly on economic grounds, and secondly on the rationale that relatively well defined, diagnosis type problems are Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 typically well handled by FS tdmolcgy. A pilot project is called for in order to both m o r e fully explore the application area, and to prwide support tcmxds a long tesm strategy. The creation of a project team is stiplated, ccanposed of a -ledge engineer, domain expert and local process manager, who muld receive overall direction f m a management steering committee. ?he report appointed individuals to the steering committee. Internal consulting support ( f m Digital's AI group) and external consulting support ( f m local university and government projects) were suggest&. Finally, the required funding, and a timetable for forming the project team, evaluating resouw=es, and developing the pilot (prototype) system(s) was delineated. Upon approval of the report, the two individuals whose responsibilities were to include the knowledge engineering function were sent to internal training courses covering ES technology, and its application within Digital. 'Ibis education phase was necessary in order for the individuals, who work in the manufacturing function and whose professional training is in engineering, to 1) beccrme fully acguainted with the opportunities and methodologies associated with ES, 2) provide expertise in the selection of an appropriate application domain, 3) contribute to the feasibility study and 4) ultimately, act as krwwledge engineers during developrent. Lastly, at this point, liason with Digital's internal AI consulting group was established for maintaining ongoing assistance. The timetable pravided by the project report specified approhtely six months for the process of education of the -ledge engineers, further specification of the domain application, and developtent of a demonstration prototype. The fXLl feasibility study would be recpired within a month of the pmtatype* Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 5.2 Feasibility Study The feasibility study (or 9?roject Plan" as it is called within Digital), was prepared by the manager of the strategic engineering function at the Nijmegen facility, the two laxxlJledge engineers assigned to the project, and a representative of Digital's internal AI consulting group. The authors did not include a domain e x p e r t . As all cmputer vendors, Digital is concerned with assuring reliability of its products and, in the case of hardware failure, minimizing customer dmtime. When a harctware fault occurs at a client site in Europe or the Middle East, the offending board (or boards) are isolated and removed from the computer installation by a field technician who then replaces them with functioning equipment. ?fie faulty boards are then sent to one of 17 field service sites located throughout E b m p e , or to the one in Israel. At the field sites, the boards are either repaired, scrapped, or sent to Digital's central repair facility (CRF) in Nijmegen, Holland for mre extensive testing. The CRF has mre sophisticated test equipment and more experienced personnel than the field service sites. Major costs are incurred in this process due to the inventory of boards required to support such a system, skipping fees, expensive test equipment and the training of repair personnel. Training costs are particularly important as diagnosing procedures may be different for different boards, and hardware is constantly changing. Additional costs arise when technicians replace sets of components on a board unnecessarily. This occurs when a problem area is isolated, but detmmmmg . , exactly the compmt responsible is especially difficult and time consuming. There are currently over 500 different boards, each one containing same 150 - 250 electronic ccmpnents. Repair times range from 30 minutes to several hours, depending on the board itself, the equipent available at the repair site, and the expertise of the technician. Test equipment varies Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 fram sinpsle voltage and current meters, to -built devices which automatically run a series of tests on the suspect M e , to full-fledged, high-end computers. Many thousands of modules pass through the system per month, each M e typically costing in the humlreds of dollars. Digital's repair function can be considered quite a sizeable business in its own right. The system will be available at each repair site, in the form of terminals at the workbench of the repair technician. 'lkmugh interactive dialogue with the technician, the system will direct the repair process by suggesting appropriate test pmcedwes and intexpreting the results of the. tests. An ES solution is proposed because there is a good match between the characteristics of the problem and the capabilities of expert systems. F'rom a technical point of view, the problem is difficult but limited in breadth, experts exist, and the task involves symbolic, not mathematical manipulation. Morec~er, this is a standard diagnosis problem and as such tends to be well suited for an ES application. F'rom a practical/business point of view, e x p e r t s are available, and the possible tangible benefits are sizable. FWber, Digital, as a major computer capmy and one irnro1ve.d in AI researc;h and products, has a practical interest in learning about this technology through its own experience. A c c o r d i n g to the report, the Knowledge-Based Board Diagnosis System (KEBIE) will reduce costs by increasing local repair, thereby minimizing **in-pipeline inventoryw, speeding repair, distributing scarce expertise, lowering test &pent -ts, cutting training time, and decreasing cmponent usage. F'urther, the -ledge and records developed through the system may provide fault infomtion useful for the design of future haxdmre pKxtucts. Diagnosing all conceivable faults or incorporating all possible repair informtion are not goals of the project. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Digital's motives in this project are, firstly, to implement this particular expert system. However, their goals extend beyond this immediate task tnmiis twr, others: 1) the development of a rrred.lanism for evaluating the feasibility of this technology in other applications, and 2) the creation of a set of generic software modules that can sewe as a basis for future expert systerrrs in the manufacturing/repair ernriromt. The feasibility study begins with a description of the current system and a discussion of the opportunities for improvement. A single alternative, the e x p A system is proposed; therefore the analysis compares current procedures to the proposed one. 5.2.1 33enefits Cost reduction in inventory was estimated via a simple model which estimates the savings in boards "in the pipelinen e x p e d A via use of the . The following input to the model is required; input estimates were varied in order to provide a sensitivity analysis. The total volume of boards in the system. The fraction of boards sent for repair to the central repair facility. The carrying cost of inventory. The Yurn around timew of boards sent to the central facility. The wturn arcxlnd timew of boards in local repair. The cost of a module. Additional expect& benefits were approximated by providing estimates of the impact of the system on each of the following categories. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Decrease in average time to diagnose. Reduction in learning curve cycle. Lrrwering of depreciation costs. Reduction in component usage. Actual percentages were proposed for each category, based upon judgments of the project team. No dollar value was provided for these savings, though such a transfomtion could be made. 5.2.2 Costs Costs for the system are broken d m into three categories: hardware, software and personnel. The system will be developed on a micro-Vax (a stand-alone Vax workstation), and will run on any Vax under the VMS operating system. At the repair workbenches, access to the system is to be assured through lxxfhxe and software connections to Vax's already on site, or if a Vax is not already available on site, use of the system will require the purchase of a micro-Vax. S o f t w a r e required includes the O X - 5 shell, the knowledge base, the knowledge crpdating mechanism, and the interface code, all of which is considered part of the KBBSX;. (See the next sections on prototype develoysment, and systems analysis and design for more on hardware and software.) Fersonnel requhmmts include the project manager, two knowledge engineers, internal consultants, domain experts, and users, While specific individuals were identified and assigned the task of manager, knawledge engineer and consultant, only the eventual necessity of recruiting articulate, qualified, and willing domain experts and users was recognized. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 A quarterly budget e x k n d h g over the length of the project (a period of dlmost two years) included costs for personnel, travel, training, hardware and software. 5.2.3 Project Planning and Specifications The folluw-ing time frame for the project was suggested. I: Quarters 1 and 2. Select a computer and several of its boards on wkich to focus. The application should be direct& taiJards a stable, well known product. The V a x 11/750 (Ccxnet) was suggested. Develop prototypes. The first demonstration prototype is to be tested at Nijmegen. A second prototype is to be field tested at a local repair facility. Evdluations of each prototype are to be conducted. Revision and improvement of the system should be e>rpected at each evaluation. The system is to be distributed to those local facilities wkich request copies. 111. Quarters 3, 4, 5 and 6 Select and develop an ES for a new and lesser knawn project. Create additional module repair systems for additional pmducts. Develop, based on experience gained thus far, procedurdl guidelhes and generic software tools for M e r use in the repair process developnent envirorIment. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 The project plan clarifies the limits of the project: the system should not be exptxted to diagnose all conceivable faults, or to incorporate all possible information CO- board repair. System specifications include: 1) capability to aid in the diagnosis of 30% - 50% of the fault., 2) a response the varying from 10-40 seconds, 3) a context sensitive ffhelpff function, and 4) a Nkistory file" option for recording the diagnosis procedures used by the technicians for problems not solvable via use of the K 6 B . Detailed criteria by which the system would be evaluated were not cited in the report. Technicians currently doing the repair task wili still do the task under the new system; that is, no change in the individuals doing the task is The task itself will change in the sense that a new piece of test equipment (the ES) will be added to the tools available to the technician. It is expected that fewer boards will have to be sent to central repair locations, less boards will be scrapped, fewer consultations with other technicians (on difficult problems) will be required, and the learning curve for repair will shorten. No special training is expected to be needed, other than minimal on the job training. Scnne concern was expressed about the KBEDS acting to de-skill the task of board diagnosis, and a concommitant resentment of the system on the part of the technicians. Solutions to this possible eventuality were to a) design the system to avoid de-skilling, b) encourage technicians to move on to more complicated products, and c) encourage technicians to take on updating and eKMnding the KBBlXj as part of their jds. 5.3 Prototype A nurmber of e x p e r t system shells were evaluated with the assistance of Digital's internal AI group. ?he first prototype was developed with a cxmmercially available PC-based product. Rris initial prototype contained Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 lrnawledge about a subset of possible faults for a single Comet board. A decision was then made to m e to the OPS-5 language on a V a x - b a s e d system. It was felt that this platform allawed for the greatest potential in terms of cammication and networking within the Digital manufacturing and repair e n v h m t . Digital policy directs only the use of already "provent1 in applications ES software. While ES within Digital have employed a variety of software tools, OK-5 was ultimately chosen for its flexibility, and longv history of use. The prototype senred to: 1) deepen the project team's understanding of the technical and organizational issues surounding the board diagnosis problem, 2) gather technical support from the internal AI consulting group, 3) gather project support from the repair facilities, and 4) averall, dem3nstrate the technical feasibility of the application. 5.4 Analysis, Design, Specification, and Rqxmmhg Developrrrent of the system progressed in stages, from prototype to test- release version to full-release version. While the prototype contained limited knclwledge about one board, the test release version could reason about faults on four. Knuwledge was i.nmamtally added to the system, broadening and deeping its capabilities through full release. Aside from some ccmmnly encountered problems (i. e. , the e x p r t s had difficulty expressing their knowledge, the process was ~ ~ l y time consuming), the project team acknowledged their frustration in accessing the experts. The team members emphasized the importance of gaining ccnnmitmmt frcxn verifiable ex.prts earlier in the project. Each use of the system during test and full release autmtically sent a Ituse reportf1 electronically back to the project team at the Nijmegen facility. The report recorded the type of board under test, and a rating on a seven-point scale of h w well the KBJ3DS was able to diagnose the fault. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 (This scale, described in the next section was used for evaluation purposes.) The technician could also enter free form amments regardhg performnce. Thus, if the system could not find the fault, information was collected r e q a d h g the type of fault which occurred, and h m the technician proceeded to diagnose the problem. !Ms knowledge, pramptly passed back to the project team in Nijmegen, could then be incorporated into the system by lneers. the knowledge eng' 5.5 Testing 5.5.1 Test Blease Several months after the demonstration prototype was completed, an updated system with a more extensive knowledge base was ready for evaluation. A successful review at this stage would mean an official test release at one repair facility. (?his facility was select& based on the willingness of the plant manager to participate in the test. ) In order to allaw for more complete evaluation, the system was installed at the proposed test site, in Ewy, F'rance, where local staff could samewhat informally examine it. This feedback could then be utilized in the review for test release. Members of the test release review team included the EVry facility plant managers, a repair engineer from Evry who had had access to the system, a representative from Digital's intenadl AI group, and two representatives from the advanced mmufacturing group who had experience with another ES project. The day long meeting's agenda began with an historicdl overview of the goals and objectives of the project, an outline of its current and p l m status, and a demonstration, including "hands-onH usage of the system. Overall the project was on schedule, the designem having created a system containing krmwledge mgardhg four Carnet modules. In the process, they had gained significant understanding of the board diagnosis task, and solid experience in knowledge engineering. Problems Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 regarding accessing damain experts and the task of gleaning expert knowledge were discussed. Aftes the presentation by the developerkt team and an opportunity to try the system themselves, the review members broke into groups, spendirq an hour and a half discussing the project. The team then regrouped to discuss their conclusions with each other and the project team, and subsequently provide feedback to the project team concaning further developent and irrprzlementation. From both the response of the technicians who had tried the XBBIE on site, and from the demonstration at the review meeting, it was felt that the system was ready for field testing. While no technical redirection was suggested, several proposals (made by the Evry test site managemnt) concesning organizational issues were enterbitled and ultimately incorpomted into further develcqrme~t plans. The first of these concerned better orientation of the personnel at the test facilities concerning the nature, scope and -ts of the project. It was suggested that the technicians, engineers and managers at the repair sites needed to be better informed prior to release of the system as to how the system works, the h a r d w a ~ mired, who will use it, and what if any, training would be required. It was agreed that a steering curtunittee should be set up at the repair facilities, capsed of project members, and facility managers and technicians; the objective of these steering groups would be to better manage the transition and maintenance of the KBBE at each site. It was also suggested that several technicians be brouFplt more closely into collaboration with the devel-t team in their work tclwards enhancing the system. FWther, the nature of a formdl evaluation mechanism was debated. The project team proposed that the system be evaluated during test release phase according to the following mechanism. Each use of the system could produce one of seven possible results, listed belm. The project team Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 had specified the nwdmm or minimum percentage of boards to fall in each category; these are listed in the right hand column. 0) No Response < 30% 1) 'lbtally Incorrect Response < 5% 2) Unclear and Misleading Response < 5% 3) Neutral Response < 10% 4) Samewhat Helpful R e s p o n s e > 15% 5) Very Helpful -rise > 20% 6) System Found Fault > 15% The review team, in particular the manager of the E k q test site, proposed alternative criteria by which the system should be measwed for successful exit from the test release phase. As opm to specifying minimum or m x h m percentages for each category, it was suggested that at least 60% of the trials produce results greater than 3, with an average performme of at least 4.5. Moreuver, it was pointed out that of major concern to mgement was that average time to repair (A!ITR) be reduced; a goal for success would be a reduction in this measure by 33%. Finally, it was noted that no measure for opmtor satisfaction with the user interface had been established. It was proposed that feedback from the technicians ragarding satisfaction with the user interface be obtained. These proposals were approved by the review and project teams as suitable criteria for evaluating the system during test release, in anticipation of full release. 5.5.2 Full F&lea~e The system was in test release for sane three months. At the end of this period, an evaluation was held to determine if and how a full release should be und-. ?his evaluation concluded that: Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Overall, the system performance gene.rally met the technical standards n q u i m d for full release. (Reliable ATTR data were not obtainable hawever; it was agreed that these data would be appropriately determhed and made available in the near future. ) ?he system, both f m the technicianst and manager's viewpoints, was extremely valuable as a learning tool; that is, as a means for novice technicians, or more experienced technicians unfamiliar with Comet boards, to "get up to speedH quickly and with minimum frustration. Novice technicians working with the KF3BaS could maintain output at about the same level as very experiencd technicians. sane concern was expressed by the manager of the repair facility concerning the time spent by his technicians in working with the knowledge engineers in qdating the knowledge base. It was suggested that this be considered formally as an investmmt by Digital, as this time could not, under present corporate guidelines be counted nproductive timew vis-a-vis the established repair metrics. Finally, the experience during the test phase reinforced the belief that a careful preparation be made at each repair facility prior to intmducing the KBBaS. Fersonnel at the facilities should be made aware of the capabilities, reqyirements, limitations, and benefits of the system before it comes on site. 5.5.3 m-Release Full, inmemental release to the other test facilities follwed. Six months fram the start of this full diffusion of the system, a final evaluation was performed based both on the criteria established at the review for test release, and a cost savings analysis. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Performnce data w e r e collected for the start, mid-point, and end of the six month long, full release evaluation period. Continuous i m p r o v m t vis-a-vis the seven point rating scale w a s observed, though the level and improvement of perforx-ane varied between the boards. W i t h respect t o the average time to repair metric, as ccanpared to process stamiards defined corporate wide, the reduction i n ATI!R ranged from 20% to 6%, depending on the board. More impmssive were the %ealtt ATI'R improvements for novice technicians; using the system reduced the ATI!R by approximately 50% for two of the boards. 77- feedback from both technicians and managers was very positive. Again, the big advantage of the system was perceived' to be the sharp rise it induced i n the learning curve, and the improved overall performance of novice technicians. 5.6 Training, Comersion/Installation After s u m f u l evaluation of the test release system, the K B B E was released t o the other Carnet repair f a c i l i t i e s , one a t a time. The project team in each case was responsible for presenting the system t o the f a c i l i t y staff. Training r e q u k i l was muumdl. . . 5.7 Operations The automatic, el-nic fommiihq of diagnostic reports assured a mechanism for keeping track of needed updates. These could be incorporated into the system by the knawledge engineers in the Nijmegen f a c i l i t i e s , and the updated knawledge bases tested. A t regular intervals, the revised system can be transmitted to the lccal repair sites. A s the inception phase described, the Comet KBBDS was seen as an early stage of an overall kwwledge engineering pracess a t Digital. W i t h the Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 future development of additional ES in mind, a mugh analysis was performed canparing the estimated cost savings of the Comt system, with the estimated costs of developing and maintaining a similar system. Based on the Comt project experience, the estbted total costs for developtent arsd maintenance (i.e., knwledge amsition, validation, hardware, etc. ) of another similar system would amunt to between $25,000 - and $50,000. Estimated annual savings due to the Comt system, based solely on the reductions in AlITR (based on process standards) and in the learning curve, more than offset tkis cost. (These savings do not include those due to inventoxy or scrap reduction, nor those due to reductions of boards Itin the pipelinew. ) It should be noted hwever, that the managers of the repair process find that the greatest value of the system is in the increased flexibility it provides. In bringing novices "up to speeY1 quickly, and in general allwing less exprierx'Rd technicians to perform mre proficiently, the Carnet ES has allowed for 1) peak work periods to be handled with relative ease, 2) decreased dependence on the constant availability of expert technicians and 3) greater freedom for the e x p e r t s to work on the more difficult problems. Overall, the system was seen to have achieved its original objectives, Since the implementation of the Comet ES, the system has been expanded to include diagnostic krxrwledge about mdules on Micro-Vax 2000 workstations. Cummt discussions focus on future directions of expert system technology as part of the overall strategy in Digitalf s repair and manufacturing environment. Figure 3. presents an overview of milestones of the Comet expert system project. Figure 3. about here Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 ?he task of managing the develcpmt and implmtation of a large scale expert system is in many ways similar, but in substantive ways different than that for large, traditional ccmputer based systems. This paper has served to highlight the similarities, describe the differences and provide a rationale for the contnsks. Figure 4 summarizes the differences in the SA&D process for the two systems. Figure 4. about here Among the differ- with important managerial implications, are that an Es: Captures and manimates knowl&ge, as opposed to information. Includes a working prototype phase. Is developed in an incremental, iterative style. R q u i r e s additional players; in particular, at least one knowledge engineer and one damain expert. May involve non-traditional software and hanNare. Will likely involve significant revisions (q@ating) of the system once in operation. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Successful management of an ES project requires: First, clearly, recognizing when an ES solution is appropriate to the problem at hand, and when it's not. U n d e r s t a n d j n g the likely financial resources required. Identifying attainable goals for the system, and the associated benefits. Specifying evaluation criteria at several phases of the project. Setting an appropriate timetable. Identifying an e x p e r t , amt ensuring his participation. Enlisting or training a knowledge engineer. Caution concernhg the use of ES hardware and software, vis a vis maintenance and interfacing with existing systems. Understanding that conventional pmcpmmhg resources will no doubt be necessary* Ekpecting knowledge engineering to be tedious, time co&, iterative and incremental. Managing expectations and skepticism. Considering the organizational W o r task changes w h i c h are likely to result from implementation of the system. Allotting resouras and mechanisms for ongoing updating. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 This paper has addressed in detail each of these issues, by describing the life cycle of an ape& system. Each phase has been defined, and a prescriptive guide tcrward managing the resources and responsibilities required over the course of tkis life cycle was presented. An example of the developent, implementation and maintenance of a large-scale, multi-site ape& system served to illustrate the conceptual fmmamrk. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 References Alavi, M., "An Asesmmt of the Prototyping Approach to Information Systems Developimentw, Communications of the Am, vol. 27, no. 6, June 1984. Bdbrm, D.G., Mittal, S. and Stefik, M.J., t@EhpaA Systems: Perils and pranise@@, Cammications of the Aail, vol. 29, no. 9, 1986. Wlrch, J, and Grudnitski, G., Information Svstems: Theow and Practice (4th ed.) , Wiley, 1986. Dym, C.L., l@Issues in the Design and Implementation of Expert Systemst1, Artificial Intelliqence for Ensineerinq Desicm, Analvsis Manufacturing lAIEaAM), vol. 1, no. 1, 1987. Gilmore, J. and Howard, C., "EkperIz System Tool Evaluationtt, P r d i n q of the Second Annual Conference on Ekpert Systems Tools and iPs,~lications, - Avignon, FraTx=e, 1986. (Authorsr address: Artificial Intelligence Branch, Georgia Tech Research Institute, Atlanta, Georgia 30332, USA.) Harmon, P. and King, D., Expert Systems: Artificial Intelliuence in E3usiness, John Wiley and Sons, 1985. Hayes-Roth, F., Watemm, D. and Iemt, D., Build- E % x z t Svstems, Addison-Wesley, 1983. Holsapple, C. and Whinston, A. , Wzsiness E3mer-t Systems, Irwin, 1987. Hukr, G., ItTfie Nature and Design of Post-Industrial OrganizationsH, Manauement Science, vol. 30, no. 8, August 1984. Janson, M., @IApplying a Pilot System and Prototyping Approach to Systems Development and Implementation, Information Mznaqement, vol. 10, no. 4, 1986. Keen, P. and Scott Morton, M., Decision Suport Systems: Organizational Perspec=tive, Addision-Wesley, 1978. M e , D., "HOW Coqpers and Lybrand Put Ekprtise Into Its Computers1@, Wall Street Journal, p. 33, Nwember 14, 1986. Kozlw, A., I1Rethhkhg Artificial Intelligencew, Technoloqv Wzsiness, May 1988. Kupfer, A. , "NOW, Live Experts on a Floppy Disktt, Fortune, October 12, 1987. Leonard-Barton, D. , l@lhe Case for Integrative Inrxnmtion: An Expert System at Digital1', Slcan Manauemat Review, Fall 1987. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Ir?Onard-E3artonI D. and Sviokla, J., wPutting EYpert Systems to Workw, Il[arvard B u s i n e s s Review, March-April 1988. Linden, E., lvIntellicorp: The Selling of Artificial Intelligencew, Hiqh Technolay, W s x h 1985. Luconi, F., Malone, T., and Scott Morton, M.S., vvE3rpest Systems: The Next Challenge for Managersn, Sloan Manacfement Review, Summer 1986. Lucas, H., Information Systems Concepts For Management, McGraw-Hill, 1982. McDemott, J., lvR1 Revisited: Four Years in the Trenchesvv, a Maqazine, Fall 1984. Mettxey, W., "An Assessrm?nt of Tools for Building Larye Knowledge-Based Systems1v, Maqazine, vol. 8, no. 4, Winter 1987. Senn, J., Analysis and Desiqn of Information Systems, McGraw-Hill, 1984. Sprague, R. and Carlson, E., Wri1di.w Effective Decision Support Systems, Prentice-Wl, 1982. Waterman, D. , _A Guide to EamA Systems, Addison-Wesley, 1986. Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Inception Feasibility Study systems Analysis Design Specifications lksting Training ~onversion/Installation Operations Figure 1. A Traditional Systems Analysis and Design F'rarnework ( f m meas 1982) Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Inception + Prototype Propcrsal Appmval of Prototype Development Workkj Versions (Incremental Improvements) Demonstration Prototype + Feasibility Report Approval of Project Developrent of System for Test Release Approval of Test Release System (Evaluation Criteria Passed) Test System Released Evaluation of Test System Perfonnance Improvement of Test System For General Release General Release (pefhaps phased in) Evaluation of General Release System Performance Maintmarle~ting Figure 2. Life Cycle of An Expert System Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Knuwledge Wineer brought on board; started coordination with external consultants; selection of Comet application. April 1986: Evaluated ES shells; coordination with internal AI group. July 1986 Developed first prototype (for one Comet module) on a PC using Wrsonal Consultant Plus fram Texas Instruments; decision to go to OPS5 based system on Vax; t m m s p r t prototype to Vax. Sep- - October 1986 Prototype reviewed; project approved; direction defined to include four Comet M e s , second Knmledge Ehgineer brought on board. N o v e ~ n b e r 1986 - May 1987 System capabilities expanded, review for test release; evaluation criteria better defined; problems/successes isolated. June - September 1987 Trial implementation; approval for phased in full release. 1987 - March 1988 Phased release to Comet field repair sites; remote update acquisition; project evaluation. Figure 3. Milestone!3 of the Corrret ES project Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 Feasibility study Traditional System Approval for: feasibility study. Task Selected: iniormat ion-based Not Applicable Analysis Consecutive, Design well defined stages. ~pecifications Testing Well established P-• operations Fixing errors. Occasional updates. A p p r o v a l for: feasibility study and demonstration prototype. Task Selected: laxrwledge-based Working, limited version. Inmementally developed. Includes 'evaluation of prototype. ES specific costs and benefits. Iterative, incremental process. Additional players required: knowledge engineer and domain expest. Wk#-ittt answer may not exist. EsIperts may disagree. Likelihood that users are skilled ttknowledge workerstt. Fossibility of non- standard hard/software. Likelihood of frequent updates* Figure 4. Systems Analysis and Design: Contrasting Traditional and EXpert Sy- Center for Digital Economy Research Stem School of Business IVorking Paper IS-90-06 
work_tx2wszrrlzbyhcn3vjiblxod4y	----	Sliding window-based support vector regression for predicting micrometeorological data Expert Systems With Applications 59 (2016) 217–225 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Sliding window-based support vector regression for predicting micrometeorological data Yukimasa Kaneda a , ∗, Hiroshi Mineno b , c a Graduate School of Integrated Science and Technology, Shizuoka University, 3-5-1 Johoku, Naka-ku, Hamamatsu, Shizuoka 432-8011, Japan b College of Informatics, Academic Institute, Shizuoka University, 3-5-1 Johoku, Naka-ku, Hamamatsu, Shizuoka 432-8011, Japan c JST, PRESTO, 4-1-8 Honcho, Kawaguchi, Saitama, 332-0012, Japan a r t i c l e i n f o Article history: Received 4 February 2016 Revised 29 March 2016 Accepted 13 April 2016 Available online 23 April 2016 Keywords: Predicting micrometeorological data Data extraction Dynamic aggregation Support vector regression Ensemble learning a b s t r a c t Sensor network technology is becoming more widespread and sophisticated, and devices with many sen- sors, such as smartphones and sensor nodes, have been used extensively. Since these devices have more easily accumulated various kinds of micrometeorological data, such as temperature, humidity, and wind speed, an enormous amount of micrometeorological data has been accumulated. In recent years, it has been expected that such an enormous amount of data, called big data, will produce novel knowledge and value. Accordingly, many current applications have used data mining technology or machine learn- ing to exploit big data. However, micrometeorological data has a complicated correlation among different features, and its characteristics change variously with time. Therefore, it is difficult to predict microme- teorological data accurately with low computational complexity even if state-of-the-art machine learning algorithms are used. In this paper, we propose a new methodology for predicting micrometeorological data, sliding window-based support vector regression (SW-SVR) that involves a novel combination of sup- port vector regression (SVR) and ensemble learning. To represent complicated micrometeorological data easily, SW-SVR builds several SVRs specialized for each representative data group in various natural envi- ronments, such as different seasons and climates, and changes weights to aggregate the SVRs dynamically depending on the characteristics of test data. In our experiment, we predicted the temperature after 1 h and 6 h by using large-scale micrometeorological data in Tokyo. As a result, regardless of testing periods, training periods, and prediction horizons, the prediction performance of SW-SVR was always greater than or equal to other general methods such as SVR, random forest, and gradient boosting. At the same time, SW-SVR reduced the building time remarkably compared with those of complicated models that have high prediction performance. © 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). 1 s t c s o c e a t a m a o s M f a m e a i l h 0 . Introduction Sensor network technology is becoming more widespread and ophisticated, and devices with many sensors have been used ex- ensively. The devices can very easily obtain various kinds of mi- rometeorological data such as temperature, humidity, and wind peed. Micrometeorological data is affected strongly by the surface f the earth and is related to our lives and industrial activity. Ac- ordingly, the data has been used by many applications such as nvironmental control systems for greenhouses ( Othman & Shaz- li, 2012; Park & Park, 2011 ). Moreover, more advanced applica- ions exploit the data to a greater extent by using machine learning nd data mining technology. Furthermore, an enormous amount of ∗ Corresponding author. E-mail address: kaneda@minelab.jp (Y. Kaneda). e c i ttp://dx.doi.org/10.1016/j.eswa.2016.04.012 957-4174/© 2016 The Authors. Published by Elsevier Ltd. This is an open access article u icrometeorological data has been accumulated by many devices, nd it has been expected that analyzing such an enormous amount f data, called big data, will produce novel knowledge and value. To predict micrometeorological data effectively, a number of re- earchers have studied machine learning ( Smith, Hoogenboom, & cClendon, 2009 ). These researchers described prediction methods or micrometeorological data; particularly, prediction performance nd computational complexity were often mentioned. Meanwhile, icrometeorological data has a complex correlation among differ- nt features such as temperature and humidity. Moreover, its char- cteristics change variously with time. Therefore, even if big data s given as training data, it is not easy to predict micrometeoro- ogical data accurately. Furthermore, in many cases, so that mod- ls can have high prediction performance, they have to become omplicated, and the computational complexity increases. Accord- ngly, some models probably cannot be built from big data in a nder the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). http://dx.doi.org/10.1016/j.eswa.2016.04.012 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2016.04.012&domain=pdf http://creativecommons.org/licenses/by/4.0/ mailto:kaneda@minelab.jp http://dx.doi.org/10.1016/j.eswa.2016.04.012 http://creativecommons.org/licenses/by/4.0/ 218 Y. Kaneda, H. Mineno / Expert Systems With Applications 59 (2016) 217–225 2 l a m c g n e s t o e t 2 i o g l a r k l t 2 v h m v c p a m s s v t v m s w m m A a practical amount of computing time. In other words, there is a trade-off relationship between high prediction performance and low computational complexity. However, compatibility is required in some practical use. As the prediction performance in applica- tions becomes higher, the quality provided by the applications be- comes better. For example, in the case of environmental control systems based on prediction ( Kolokotsa, Pouliezos, Stavrakakis, & Lazos, 2009 ), the higher prediction performance enables the sys- tems to provide precise control, precise management, and better environments. On the other hand, models that need a long time for training are worthless in practical use. In current situations where the amount of usable data has increased remarkably, this trade-off relationship has become a more critical issue. Recently, one type of machine learning algorithm, support vec- tor machines (SVMs), have been used successfully in various fields. The basic theory is a more efficient learning method based on probably approximately correct (PAC) learning. Moreover, SVMs can separate non-linear data with low computational complex- ity. Since most data observed in the real world is likely to have non-linear relationships, SVMs have also been applied to mi- crometeorological data prediction ( Antonanzas, Urraca, Martinez- de-Pison, & Antonanzas-Torres, 2015; Mohammadi, Shamshirband, Anisi, Alam, & Petkovi ́c, 2015; Urraca, Antonanzas, Martinez-de- Pison, & Antonanzas-Torres, 2015 ). Moreover, SVMs led to better prediction performance than other algorithms such as artificial neural networks (ANNs) and the autoregressive integrated mov- ing average (ARIMA) model ( Chevalier, Hoogenboom, McClendon, & Paz, 2011; Maity, Bhagwat, & Bhatnagar, 2010 ). However, when SVMs learn big data, the computational complexity is still a matter of concern. Another alternative learning method, ensemble learn- ing, has also been used more widely for predicting micrometeo- rological data ( Singh, Gupta, & Rai, 2013 ). The prediction perfor- mance of ensemble learning is greater than or equal to that of SVMs. The basic methodology is a combination of weak learners built from different kinds of training data. The combination yields a higher generalizing capability that a single model cannot rep- resent. In particular, some researchers proposed improved meth- ods that could be applied to micrometeorological data prediction ( Wang & Japkowicz, 2009; Xie, Li, Ngai, & Ying, 2009 ). However, it is difficult to apply the methods to regression, and it is possi- ble that the models will not be able to follow micrometeorological data whose characteristics always change with time. In this paper, we propose a new methodology for predicting micrometeorological data, sliding window-based support vector re- gression (SW-SVR). SW-SVR involves a novel combination of sup- port vector regression (SVR) and ensemble learning. To represent complicated micrometeorological data easily, SW-SVR builds sev- eral SVRs specialized for each representative data group in vari- ous natural environments, such as different seasons and climates. The specialized SVRs are built based on our previous proposed method, dynamic short-distance data collection (D-SDC) that ex- tracts effective data for specific data prediction by taking account of movements: changes in data during prediction horizons. Each weak learner built from each extracted data specializes on spe- cific data and predicts accurately the data similar to the special- ized data. Then, SW-SVR aggregates all the predicted values based on weights decided by the similarity between test data and each data specialized by weak learners. This new ensemble learning methodology that changes weights dynamically enables following micrometeorological data whose characteristics hardly change with time. Our results demonstrated that the prediction performance of SW-SVR was always greater than or equal to that of other general methods such as SVR, random forest, and gradient boosting. At the same time, SW-SVR reduced the building time remarkably com- pared with that of complicated models that have high prediction performance. . Related work As mentioned in the introduction, to predict micrometeoro- ogical data effectively, SVMs and ensemble learning have gener- lly been used. These algorithms have higher prediction perfor- ance for micrometeorological data than traditional methods be- ause SVMs use not only a margin maximizing algorithm whose reat performance was proved by PAC learning but also the ker- el trick that enables non-linear separation. On the other hand, nsemble learning provides higher generalizing capability that a ingle model cannot represent. In this section, a brief summary of hese algorithms and some improved algorithms are given. More- ver, so that SW-SVR can draw advantages from both SVMs and nsemble learning, several problems of these algorithms for prac- ical use are discussed. .1. Support vector regression SVMs, introduced by Vapnik,(1995 ), have been used successfully n various fields. In the simplest case, binary classification, SVMs btain a separating hyperplane decided by maximizing the mar- in. The margin means the norms between different classes. PAC earning proved that maximizing the margin produces high gener- lization ability. Moreover, the kernel trick enables SVMs to sepa- ate data non-linearly with low computational complexity. Various inds of data observed in the real world are likely to have non- inear relationships. Accordingly, SVMs are used in many applica- ions such as micrometeorological data prediction ( Kisi & Cimen, 012; Maity et al., 2010 ). Meanwhile, SVMs for regression, support ector regression (SVR), uses the same methodology as SVMs that ave the highest generalization ability. In this section, a brief sum- ary of SVR is given as follows. First, the linear function for regression is given as follows: f ( x ) = w T x + b. Then, as with SVMs, SVR also minimizes the norm of the weight ector w ; the L 2 norm ‖ w ‖ 2 is often used, and minimizing ‖ w ‖ 2 orresponds to maximizing the margin. Meanwhile, SVR tolerates rediction error �. Therefore, the primal problem of SVR is shown s follows: inimize ‖ w ‖ 2 2 ubject to { y i − ( w T x i + b ) ≤ �( w T x i + b ) − y i ≤ �. Moreover, to take some errors into account further, the same lack variables ξ as soft margin SVMs are introduced. The slack ariables mean penalties and increase in proportion to errors be- ween true values and predicted values. The problem that the slack ariables are introduced into is shown as follows: inimize ‖ w ‖ 2 2 + C ∑ i ( ξi + ξ ∗i ) ubject to ⎧ ⎨ ⎩ y i − ( w T x i + b ) ≤ � + ξi ( w T x i + b ) − y i ≤ � + ξ ∗i ξi , ξ ∗ i ≥ 0 . here the constant C means the balance between the effect of aximizing the margin and penalties. To minimize the above for- ula, the slack variables in the formula must also be minimized. ccordingly, the slack variables depending on the errors are shown s follows: ξi = { 0 ( y i − ( w T x i + b ) ≤ � ) y i − ( w T x i + b ) − � otherwise Y. Kaneda, H. Mineno / Expert Systems With Applications 59 (2016) 217–225 219 ξ t t w t r c A s m m m s l l ϕ d h o t a c w p ( i p d c t p & b s t d r n p v p f l p u c 2 i t T g Algorithm 1 Bagging for regression. Input: Training data: D = { ( x 1 , y 1 ) , . . . , ( x N , y N ) } where x i ∈ X , y i ∈ Y Number of weak learners: n For t = 1 to n do 1. D t ← generate sample from D with replacement 2. H t ( X ) ← build a weak learner from D t Output: H(X ) = 1 n n ∑ t=1 H t (X ) Algorithm 2 Boosting for regression. Input: Training data: D = { ( x 1 , y 1 ) , . . . , ( x N , y N ) } where x i ∈ X , y i ∈ Y Number of weak learners: n Weights: w i = 1/ N For t = 1 to n do 1. H t ( X ) ← build a weak learner from D by using weights w t 2. �t ← compute error rate of H t ( X ) 3. αt ← compute reliability of prediction result of H t ( X ) based on �t 4. w t+1 ← update weights w t based on αt Output: H(X ) = n ∑ t=1 ( αt H t (X ) ) / n ∑ t=1 αt r i p b w p A a T n m ( a c i s i e e l s d i b g A t e i i i ( f b b t a p ∗ i = { 0 (( w T x i + b ) − y i ≤ � ) ( w T x i + b ) − y i − � otherwise . The above formulas mean that a penalty is not given when he error is lower than �, but the error is regarded as a penalty hat cannot be tolerated when the error is higher than �. In other ords, SVR tolerates errors less than �, but errors over � are solely aken into account as penalties. Finally, the dual problem is de- ived from the above primal problem by Lagrange multiplier and orresponds to a quadratic programming problem as with SVMs. s a result, since a unique global optimal solution is solved, SVR is uperior to traditional algorithms that might fall into a local opti- al solution, such as ANNs. The dual problem derived by Lagrange ultiplier is shown as follows: aximize − 1 2 ∑ i, j ( αi + α∗i ) ( α j + α∗j ) x T i x j − � ∑ i ( αi + α∗i ) + ∑ i y i ( αi − α∗i ) ubject to {∑ i ( αi − α∗i ) = 0 αi , α ∗ i ∈ [ 0 , C ] . Moreover, the above dual problem can easily involve non- inear map ϕ to consider a higher dimension. To introduce non- inear map ϕ in the above problem, kernel function K(x i , x j ) = t (x i ) ϕ(x j ) is defined and used instead of x T i x j . Then ϕ t (x i ) ϕ (x j ) is etermined based on K( x i , x j ) without calculation on a mapped igher dimension; this method is called the kernel trick. SVR based n maximizing the margin and the kernel trick yields high predic- ion performance. Meanwhile, conventional quadratic programming solvers, such s the steepest descent method, have very high computational omplexity; the computational complexity is approximately O ( N 3 ) here N is the number of training data. Accordingly, a quadratic rogramming solver for SVMs, sequential minimal optimization SMO), has become de facto standard ( Platt, 1998 ). SMO special- zed for SVM reduce the computational complexity of SVM to ap- roximately O ( N 2 ). Nevertheless, when an enormous amount of ata is inputted as training data, the computational complexity in- reases substantially. To solve the problem, a theory that regards he quadratic programming problem as a computational geometry roblem, core vector machine (CVM), was proposed ( Tsang, Kwok, Cheung, 2005 ). The prediction performance of CVM is compara- le to that of SVMs, and the computational complexity decreases ubstantially. However, according to a paper ( Loosli, 2007 ), predic- ion performance and computational complexity of CVM strongly epend on the values of parameters. Therefore, when essential pa- ameter tuning for practical use is taken into account, CVM does ot always satisfy both high prediction performance and low com- utational complexity. SVR is one of the best algorithms in machine learning from the iewpoint of prediction performance. In particular, it has been ex- ected that the kernel trick used in the dual problem is effective or predicting micrometeorological data that has a complex corre- ation among different f eatures. However, the com putational com- lexity to solve the dual problem is often still long for practical se. Thus, it is difficult to apply conventional SVR directly to mi- rometeorological data prediction. .2. Ensemble learning Ensemble learning has been studied recently and used increas- ngly. The basic methodology of ensemble learning is a combina- ion of weak learners built from different kinds of training data. he combination yields a higher generalizing capability that a sin- le model cannot represent. As with SVMs, ensemble learning can epresent non-linear relationships and has been used for predict- ng micrometeorological data. In particular, the two kinds of ap- roaches, bagging and boosting, have often been used in ensem- le learning. The approaches differ greatly on the method to build eak learners and aggregate them. Bagging uses several training data generated by bootstrap sam- ling. The algorithm of basic bagging for regression is shown in lgorithm. 1 . In bagging, different kinds of training data are cre- ted by sampling inputted original training data with replacement. hen, weak learners are built from each sampled training data. Fi- ally, each predicted value is aggregated by majority vote or arith- etic average. In particular, random forest, introduced by Breiman Breiman, 2001 ), to which randomness in feature selection is also pplied, often demonstrates better prediction performance than onventional models such as SVMs. Random forest is used in var- ous applications and has been extended to other improved ver- ions. For example, to predict imbalanced data observed frequently n the real world more accurately, improved balanced random for- st (IBRF) has been proposed ( Xie et al., 2009 ). IBRF involves an fficient sampling method for imbalanced data and cost-sensitive earning that penalizes misclassification of minority class more trongly. The authors showed that IBRF was more effective to pre- ict imbalanced data than class-weighted SVMs and conventional mproved random forest for imbalanced data prediction. Boosting builds repeatedly weak learners by using weights ased on the error rate. The algorithm of basic boosting for re- ression such as Adaboost ( Freund & Schapire, 1997 ) is shown in lgorithm. 2 . Unlike bagging, almost all boosting algorithms use he same training data, but the training data is weighted repeat- dly. Boosting alternates between building weak learners by us- ng weights and updating weights. Finally, each predicted value s aggregated by weighted average. Various kinds of algorithms n boosting have been studied and proposed; gradient boosting Friedman, 2001 ) in particular has shown the best prediction per- ormance in many competitions. Meanwhile, as with IBRF, the oosting algorithm for imbalanced data, boosting-SVM, has also een proposed ( Wang & Japkowicz, 2009 ). The main characteris- ic of boosting-SVM is using asymmetric misclassification cost. The uthors demonstrated that boosting-SVM enabled more accurate rediction of both the majority class and minority class. 220 Y. Kaneda, H. Mineno / Expert Systems With Applications 59 (2016) 217–225 Training data Extracted data Center of cluster Test data Number of weak learners: 3 Threshold of extraction Training data Specialized object Movement of data Training data at end of prediction horizon (a) Extraction of training data by D-SDC. (b) Weighted ensemble learning in SW-SVR. Fig. 1. Processing overview of SW-SVR. Algorithm 3 Sliding window-based support vector regression. Input: Training data set: S = { ( x 1 , y 1 , x ′ 1 ) , . . . , ( x N , y N , x ′ N ) } where x i ∈ X, y i ∈ Y, x ′ i ∈ X ′ Test data: P Number of weak learners: n Weight parameters: p , q Preprocessing: 1. apply normalization to X and X ′ 2. fit kernel approximation and PLS regression to X and X ′ 3. M i = || x i − x ′ i || , i = 1 . . . N 4. G t ← each center of kmenas (X ) , t = 1 . . . n For t = 1 to n do 1. D ti = || G t − x i || , i = 1 . . . N 2. r t = N ∑ i =1 ( w i M i ) / N ∑ i =1 ( w i ) where w i = 1 /D p ti 3. S t = { ( x i , y i ) | D ti < r t } , i = 1 . . . N 4. H t ( X ) ← train LinearSVR ( S t ) Output: H ( P ) = n ∑ t=1 ( w t H t (P) ) / n ∑ t=1 ( w t ) where w t = 1 / || G t − P|| q l 2 a r t a p t t i d r p f c t e e a i o f m w m When micrometeorological data including many unusual natu- ral environments is regarded as imbalanced data, the above meth- ods are likely to classify micrometeorological data more accurately. However, these approaches cannot be applied to regression. More- over, according to our previous research ( Suzuki, Kaneda, & Mi- neno, 2015 ), there is proper training data depending on test data. In other words, weights to aggregate weak learners built from dif- ferent kinds of training data should depend on test data. 3. SW-SVR: Sliding window-based support vector regression We propose a new methodology for predicting micrometeoro- logical data, sliding window-based support vector regression, com- bining methodologies of SVR and ensemble learning. The basic the- ories are based on D-SDC, our previous proposed method to extract effective data for specific data prediction, and novel weighted en- semble learning as shown in Fig. 1 . First, to represent complicated micrometeorological data easily, SW-SVR builds several SVRs spe- cialized for each representative data group in various natural envi- ronments, such as different seasons and climates. The specialized SVRs are built based on D-SDC that extracts effective data for spe- cific data prediction by taking account of movements: changes of data during prediction horizons ( Fig. 1 (a)). Each weak learner built from each extracted data specializes on specific data and accurately predicts the data similar to the specialized data. Afterward, each weak learner is aggregated with weights determined dynamically at the time of prediction so as to maintain the prediction perfor- mance of micrometeorological data whose characteristics always change with time ( Fig. 1 (b)). The weights are decided by the simi- larity between test data and each data specialized by weak learn- ers. Even if the characteristics of micrometeorological data always change with time, SW-SVR always gives priority to weak learners that are more suitable for predicting test data. The details of the SW-SVR algorithm are shown in Algorithm. 3 . The procedure for training consists of two kinds of preprocessing, iterated learning, and dynamic aggregation. The procedures of each part are shown as follows. The below-mentioned algorithms in SW-SVR use the L 2 norm: the Euclid distance, and the performance is related to feature space. For example, if feature space includes noisy features or non-linear relationships between features, the performance will probably be reduced substantially. In particular, micrometeoro- logical data has a complex correlation among different features such as temperature and humidity. Accordingly, feature space must be mapped into other feature space that takes into account the presence of noise and non-linear relationships. In our approach, we use kernel approximation ( Rahimi & Recht, 2007 ) and partial east squares (PLS) regression ( Tenenhaus, Vinzi, Chatelin, & Lauro, 005 ) to map into new feature space. Kernel approximation gener- tes new feature space and involves higher dimensions that rep- esent non-linear data as linear data with a very low computa- ional complexity. Actually, a combination of kernel approximation nd linear SVMs led to faster prediction performance that is com- arable to that of exact SVM ( Cao, Naito, & Ninomiya, 2008 ). On he other hand, PLS regression is a supervised dimension reduc- ion methodology. This method can reduce dimensions by extract- ng latent variables that have a strong relationship with a depen- ent variable. If feature space includes noisy features, the effect is educed because of PLS regression. The combination of kernel ap- roximation and PLS regression enables SW-SVR to use effective eature space for calculation of the L 2 norm in micrometeorologi- al data. According to our previous research, to accurately predict par- icular specific data in micrometeorological data, it is necessary to xtract effective training data for specific data prediction ( Suzuki t al., 2015 ). In SW-SVR, these several specific data is selected in dvance, and weak learners are built from extracted effective train- ng data for predicting each specific data. Meanwhile, micromete- rological data involves various natural environments such as dif- erent seasons and climates. Therefore, each selected specific data ust represent more varied natural environments that probably ill appear so as to represent micrometeorological data by several odels. In SW-SVR, each specific data is selected by a clustering Y. Kaneda, H. Mineno / Expert Systems With Applications 59 (2016) 217–225 221 a m d I n T s t t K t s D a o f w p a e w v c d m i m d o d l a d a t s m t a r o n m w g t b l a b t i t l p i m d o p t S w d s H a s c d T u p S i i N r i t w c t O 4 4 m g S p k b b t s p m p m n y w p t a i a a t c i lgorithm, k-means ( Macqueen, 1967 ). The k-means is one of the ost famous non-hierarchical clustering algorithms and classifies ata faster under several clusters than other clustering algorithms. n SW-SVR, the k-means classifies all training data into the same umber of clusters as the number of weak learners given by users. hen, each center of clusters is used as specific data that repre- ents various natural environments. After selecting several specific data, SW-SVR iterates data ex- raction and building a model. First, SW-SVR extracts effective raining data for predicting each specific data by D-SDC ( Suzuki, aneda, & Mineno, 2014 ). The theory of D-SDC is similar to that of he k-nearest neighbor (k-NN) algorithm, and D-SDC also extracts ome training data similar to a specialized object. However, in our -SDC, the amount of extracted data depends on the movement of specialized object with time. The movement r means the change f a specialized object during prediction horizons as shown in the ollowing equation: r t = ‖G t − G ′ t ‖ here G is a specialized object, and G ′ is a specialized object after rediction horizons. D-SDC extracts training data whose norm from specialized object is shorter than the movement r . Accordingly, xtracted training data S by D-SDC is given as follows: S t = { ( x i , y i ) | ‖G t − x i ‖ < ‖G t − G ′ t ‖ } here x is the feature of training data and y is the dependent ariable of training data. D-SDC is based on the movement r be- ause the movement r is strongly related to autocorrelation of ata surrounding a specialized object. In micrometeorological data, ovements in specific natural environments are mutually sim- lar, and the autocorrelation becomes lower when these move- ents are bigger. For example, in Japan, the change of weather is rastic every spring, and the natural environments change various ther natural environments with time. Meanwhile, when we pre- ict time series data such as micrometeorological data, autocorre- ation means correlation between features and a dependent vari- ble, and more training data is required for highly accurate pre- iction when autocorrelation is lower. Since D-SDC extracts the mount of data surrounding a specialized object in proportion to he movement r , extraction that considers autocorrelation of data urrounding a specialized object is achieved. However, the move- ent r is unknown because G ′ is not observed. Meanwhile, as men- ioned above, movements of data surrounding a specialized object re mutually similar. Therefore, D-SDC estimates the movement based on movements of training data similar to a specialized bject by weighted average, where the weights are reciprocals of orms between a specialized object and each training data. Move- ents of training data can be calculated by referring to the time hen each training data is observed. The estimated movement r is iven as follows: r t = ‖G t − G ′ t ‖≈ ∑ N i =1 w i ‖x i − x ′ i ‖ ∑ N i =1 w i where w i = 1 ‖G t − x i ‖ p , N is the number of training data, and p is a weighted parame- er. Afterward, SW-SVR builds several linear SVRs as weak learners ased on the extracted data. As described above, a combination of inear SVR and kernel approximation is comparable to SVR using kernel method. Moreover, linear SVR can be built much faster y using liblinear ( Fan, Chang, Hsieh, Wang, & Lin, 2008 ), an op- imized implementation for linear SVMs, instead of other general mplementations of SVMs such as libSVM ( Chang & Lin, 2011 ). Al- hough a usable kernel in liblinear is restricted to the linear kernel, iblinear can build the model much faster by solving the primal roblem instead of the dual problem. Furthermore, since all train- ng data is divided into smaller amounts of extracted data, each odel can be built faster, and it is easier to learn each extracted ata by parallel processing. The predicted values of SW-SVR take into account the change f natural environments with time. In general ensemble learning, rediction for regression depends on the weighted average, and he weights are determined at the time of training. However, SW- VR determines weights dynamically at the time of prediction. The eights are determined by the norm between test data and each ata specialized by weak learners. A final hypothesis of SW-SVR is hown as follows: ( P ) = ∑ n t=1 w t H t ( P ) ∑ n t=1 w t where w t = 1 ‖ G t − P ‖ q , P is the test data, n is the number of weak learners, H ( X ) is hypothesis, and q is a weighted parameter. In our approach, ince the weights of ensemble learning are determined dynami- ally for every prediction, SW-SVR can follow micrometeorological ata whose characteristics always change with time. Finally, we describe the computational complexity of SW-SVR. o represent complicated micrometeorological data easily, SW-SVR ses the various conventional methods besides D-SDC we pro- osed: kernel approximation, PLS regression, k-means, and linear VR. The computational complexity of these methods in general ncreases linearly; in other words, the computational complexity s approximately equal to O ( N ) where the number of training data is even bigger than the number of the dimensions and each pa- ameter of these methods. Moreover, the computational complex- ty of D-SDC corresponds to O ( nN ) because D-SDC just iterates N imes of distance calculation n + 1 times where n is the number of eak learners in SW-SVR. Therefore, if N is even bigger than n , the omputational complexity of D-SDC also increases linearly. The to- al computational complexity of SW-SVR is approximately equal to ( N ) that is even less than that of SVR. . Evaluation .1. Experiment We compared the performance of SW-SVR with other standard ethods for regression: k-NN, decision tree (DT), Adaboost, bag- ing, random forest (RF), gradient boosting (GB), linear SVR, and VR using a radial basis function (RBF) kernel that shows higher erformance in various fields (RBF-SVR). Note that the kernel of ernel approximation in SW-SVR is also the RBF kernel, and the ase learner in Adaboost and bagging is the decision tree that has een used generally. Moreover, to evaluate SW-SVR in more de- ail, we evaluated the performance of linear SVR with mapping: tandard linear SVR to which the same mapping as SW-SVR is ap- lied (“mapped SVR”). Mapped SVR clarifies each performance of apping feature space and ensemble learning based on D-SDC. All arameters of the used models were adjusted by the grid search ethod. Baseline for this evaluation was the performance of the aivest persistent model as shown in the following formula: ˆ i + �t = y i here ˆ y is the predicted value, y is the true value, and �t is the rediction horizons. We evaluated the performance by two ways: hold-out valida- ion and 10-fold cross-validation. We predicted the temperature fter 1 h and 6 h by using large-scale micrometeorological data n Tokyo ( Japan Meteorological Agency, n.d. ). The data consists of tmospheric pressure, temperature, relative humidity, wind speed, nd irradiance. In hold-out validation, training periods are limited o the earlier periods than testing periods so as to assume practi- al use; test data is always predicted based on past training data n practical use. The training periods were from 3 months to 5 222 Y. Kaneda, H. Mineno / Expert Systems With Applications 59 (2016) 217–225 (a) Testing periods: 1 month. (b) Testing periods: 6 months. (c) Testing periods: 12 months. 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 3 6 12 24 36 60 M A P E [ % ] (l o g s ca le ) Training periods [months] Persistent k-NN DT Adaboost Bagging RF GB Linear SVR mapped SVR RBF-SVR SWSVR 8.0E+00 1.6E+01 3.2E+01 6.4E+01 1.3E+02 2.6E+02 5.1E+02 3 6 12 24 36 60 M A P E [ % ] (l o g s ca le ) Training periods [months] Persistent k-NN DT Adaboost Bagging RF GB Linear SVR mapped SVR RBF-SVR SWSVR 5.0E+00 1.0E+01 2.0E+01 4.0E+01 8.0E+01 1.6E+02 3.2E+02 3 6 12 24 36 60 M A P E [ % ] (l o g s ca le ) Training periods [months] Persistent k-NN DT Adaboost Bagging RF GB Linear SVR mapped SVR RBF-SVR SWSVR Fig. 2. MAPE for prediction after 1 h for each algorithm. Note that (b) and (c) are shown with log scale. S e h t i b p p r c t o d t d a l o e d d m 1 t S t d l p c t c a p z b s i w t p d S years before September 1, 2014, and testing periods were from 1 month to 1 year later the same day. By varying the training pe- riods and the testing periods, the performance under the various usage scenarios is evaluated. On the other hand, the periods for 10-fold cross-validation were 6 years from September 1, 2009 to September 1, 2015. Note that the amount of data per month was approximately 40 0 0 because the data was accumulated every 10 minutes. In this evaluation, we used the mean absolute percentage error (MAPE) as the index of prediction error and building time calculated based on the CPU clock time as the index of computa- tional complexity. MAPE is shown as follows: MAP E = 100 N N ∑ i =1 ∣∣∣∣ y i − ˆ y i y i ∣∣∣∣ where N is the number of test data, y is the true value, and ˆ y is the predicted value. Moreover, we evaluated the average of each extraction rate by D-SDC in each experimental condition so as to analyze the performance of SW-SVR and D-SDC further. All imple- mentations for this evaluation are in Python, and implementations in scikit-learn ( Pedregosa et al., 2012 ) were used for all methods except SW-SVR. This evaluation was performed on a single core of a machine with an Intel Core i5-2500 K Processor and 12GB of RAM; even though several methods, such random forest and SW- SVR, can be performed on parallel processing, the methods were performed on a single core so as to evaluate the building time of all methods fairly. 4.2. Results and discussion Fig. 2 and 3 show the prediction error in the prediction hori- zons of 1 h and 6 h, respectively. Note that a log scale is used in Figs. 2 (b), (c), 3 (b), and (c). The results indicate that SW-SVR produced the best average performance in all models during the whole testing periods, training periods, and prediction horizons. In particular, the effect occurs noticeably when testing periods are longer than training periods. On the other hand, in this situation, almost all methods except SW-SVR have often lower performance than the naivest persistent model as baseline. The results demon- strate that the conventional superior methods do not always dis- play the great performance for micrometeorological data predic- tion depending on difficulty of the prediction caused by training periods and testing periods and prediction horizons. Moreover, in algorithms based on SVR, the prediction performance of SW-SVR is almost the best, followed in order by those of RBF-SVR, mapped VR, and linear SVR. The difference between mapped SVR and lin- ar SVR is due to the effect of mapping feature space. On the other and, the difference between SW-SVR and mapped SVR is due to he effect of ensemble learning based on D-SDC. These compar- sons demonstrated that both mapping feature space and ensem- le learning based on D-SDC are effective for improving prediction erformance. Meanwhile, mapped SVR also tended to have lower rediction performance than that of SW-SVR when the testing pe- iods are longer than the training periods. Accordingly, under this ondition, ensemble learning based on D-SDC is particularly effec- ive. When the testing periods are longer than the training peri- ds, the effective training data for predicting the test data is re- uced. We considered that a little training data that D-SDC ex- racted for building models corresponded to the effective training ata for predicting the test data. Actually, Fig. 4 indicates the aver- ge of each extraction rate by D-SDC and demonstrates that weak earners of SW-SVR are always built from a very small proportion f the whole training data. SW-SVR that always predicts microm- teorological data accurately regardless of the amount of training ata is very practical and useful. Table 1 shows the results of 10-fold cross-validation in the pre- iction horizons of 1 h and 6 h. SW-SVR was often superior to all ethods including RBF-SVM in hold-out validation. However, in 0-fold cross-validation, although SW-SVR had higher the predic- ion performance than that of all methods except RBF-SVR, RBF- VR was superior to SW-SVR slightly. The results demonstrate that he prediction performance of SW-SVR is affected by temporal or- er between training data and test data, and SW-SVR is particu- arly suited to be used for practical use in which test data is always redicted based on past training data. Meanwhile, even in 10-fold ross-validation, the magnitude relation of the prediction error be- ween mapped SVR and linear SVR and SW-SVR was same as the ase of hold-out validation. Therefore, both mapping feature space nd ensemble learning based on D-SDC are effective for improving rediction performance in cross-validation. Fig. 5 and 6 show the building time in the prediction hori- ons of 1 h and 6 h, respectively. Figs. 5 (a) and 6 (a) show the uilding time of models that have high prediction performance as hown in Figs. 2 and 3 , RF, GB, RBF-SVR, and SW-SVR, when train- ng periods were varied. Note that the number of weak learners as 10 0 0 in the ensemble learning series, cost parameter was 1 in he SVR series, and σ of SW-SVR was 0.0 0 0 01; σ of SW-SVR was a arameter of the RBF kernel in kernel approximation. These results emonstrated that the building time of ensemble learning, such as W-SVR, increases more gently than that of SVR. In particular, the Y. Kaneda, H. Mineno / Expert Systems With Applications 59 (2016) 217–225 223 (a) Testing periods: 1 month. (b) Testing periods: 6 months. (c) Testing periods: 12 months. 7.5 8.0 8.5 9.0 9.5 10.0 10.5 11.0 11.5 12.0 12.5 3 6 12 24 36 60 M A P E [ % ] (l o g s c a le ) Training periods [months] Persistent k-NN DT Adaboost Bagging RF GB Linear SVR mapped SVR RBF-SVR SWSVR 3.0E+01 6.0E+01 1.2E+02 2.4E+02 4.8E+02 3 6 12 24 36 60 M A P E [ % ] (l o g s c a le ) Training periods [months] Persistent k-NN DT Adaboost Bagging RF GB Linear SVR mapped SVR RBF-SVR SWSVR 2.0E+01 4.0E+01 8.0E+01 1.6E+02 3.2E+02 3 6 12 24 36 60 M A P E [ % ] (l o g s c a le ) Training periods [months] Persistent k-NN DT Adaboost Bagging RF GB Linear SVR mapped SVR RBF-SVR SWSVR Fig. 3. MAPE for prediction after 6 h for each algorithm. Note that (b) and (c) are shown with log scale. 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 3 6 12 24 36 60 E xt ra ct io n r a te [ % ] Training periods [months] Number of weak learners: 10 Number of weak learners: 100 Number of weak learners: 1000 0.0 2.0 4.0 6.0 8.0 10.0 12.0 14.0 16.0 18.0 3 6 12 24 36 60 E xt ra ct io n r a te [ % ] Training periods [months] Number of weak learners: 10 Number of weak learners: 100 Number of weak learners: 1000 (b) Prediction horizons: 6 hours.(a) Prediction horizons: 1 hour. Fig. 4. Average of each extraction rate by D-SDC in SW-SVR. Table 1 MAPE of 10-fold cross-validation for each algorithm. Methods Prediction horizons SW-SVR k-NN DT Adaboost Bagging RF GB Linear SVR mapped SVR RBF-SVR Persistent 1h 5 .18608 8 .59929 5 .81042 11 .10375 10 .24014 5 .57213 5 .27190 5 .43892 5 .25274 5 .16985 5 .96816 6h 23 .49826 26 .52433 25 .99290 29 .93160 29 .58125 25 .55044 24 .14987 24 .68383 24 .26108 20 .94132 24 .86800 1.E-02 1.E-01 1.E+00 1.E+01 1.E+02 1.E+03 1.E+04 1.E+05 1.E+06 T im e [ s e c ] (l o g s c a le ) 1.E-01 1.E+00 1.E+01 1.E+02 1.E+03 T im e [ s e c ] (l o g s c a le ) RF: Depth = 5 RF: Depth = 10 GB: Depth = 5 GB: Depth = 10 SW-SVR RF: Depth = 5 RF: Depth = 10 GB: Depth = 5 GB: Depth = 10 RBF SVR: σ = 0.1 RBF-SVR: σ = 0.00001 SW-SVR - 1.E+00 1.E+01 1.E+02 1.E+03 1.E+04 1.E+05 T im e [ s e c ] (l o g s c a le ) (b) Ensemble learning series. (c) SVR series.(a) Different training periods. 3 6 12 24 36 60 Training periods [months] 1 5 10 50 100 Cost 10 50 100 500 1000 Number of weak learners Linear SVR RBF-SVR: σ = 0.1 RBF-SVR: σ = 0.001 RBF-SVR: σ = 0.00001 SW SVR: σ = 0.1 SW SVR: σ = 0.001 SW SVR: σ = 0.00001 - - - Fig. 5. Building time for prediction after 1 h for each model. Note that all figures are shown with log scale. 224 Y. Kaneda, H. Mineno / Expert Systems With Applications 59 (2016) 217–225 (b) Ensemble learning series. (c) SVR series.(a) Different training periods. 1.E+00 1.E+01 1.E+02 1.E+03 1.E+04 1.E+05 3 6 12 24 36 60 T im e [ s e c ] (l o g s c a le ) Training periods [months] 1.E-01 1.E+00 1.E+01 1.E+02 1.E+03 10 50 100 500 1000 T im e [ s e c ] (l o g s c a le ) Number of weak learners 1.E-02 1.E-01 1.E+00 1.E+01 1.E+02 1.E+03 1.E+04 1.E+05 1.E+06 1 5 10 50 100 T im e [ s e c ] (l o g s c a le ) Cost RF: Depth = 5 RF: Depth = 10 GB: Depth = 5 GB: Depth = 10 RBF-SVR: σ = 0.1 RBF-SVR: σ = 0.00001 SW-SVR RF: Depth = 5 RF: Depth = 10 GB: Depth = 5 GB: Depth = 10 SW-SVR Linear SVR RBF-SVR: σ = 0.1 RBF-SVR: σ = 0.001 RBF-SVR: σ = 0.00001 SW-SVR: σ = 0.1 SW-SVR: σ = 0.001 SW-SVR: σ = 0.00001 Fig. 6. Building time for prediction after 6 h for each model. Note that all figures are shown with log scale. p t m a a m f t S d p a S t t n w s S a M S t s s i s 5 m t a e u p c b t l m building time of SW-SVR is shortest when the training periods be- come longer. In other words, the rate of building time increase of SW-SVR is the gentlest in all the methods when training data in- creases. These results indicate that, as mentioned above, the com- putational complexity of SW-SVR is less than that of conventional methods including random forest and gradient boosting. SW-SVR is effective for training of an enormous amount of data in terms of building time. Next, Figs. 5 (b) and 6 (b) show the building time of the mod- els with better performance in ensemble learning, RF, GB, and SW- SVR, when the number of weak learners was varied. Note that the cost parameter of SW-SVR was 1, σ of SW-SVR was 0.0 0 0 01, and training periods were 12 months. SW-SVR needs a longer build- ing time than RF and GB using shallow DT when the number of weak learners is lower. However, when the depth of DT becomes deeper or the number of weak learners becomes higher, SW-SVR can build the model faster than or at the same speed as RF and GB. Moreover, SW-SVR, as with RF, can be run easily in parallel environments, and it is expected that the building time of SW-SVR will become even shorter. Finally, Figs. 5 (c) and 6 (c) show the building time of the mod- els based on SVR when the parameters of SVR were varied. Note that the number of weak learners was 100, and the training peri- ods were 12 months. These results indicate that the building time of SW-SVR is significantly shorter than that of RBF-SVR but longer than linear SVR. Meanwhile, Fig. 4 demonstrates that weak learn- ers of SW-SVR are always built from a very small proportion of the whole training data. In particular, when prediction horizons were 1 h, the average of each extraction rate was 0.47 percent at best and 1.82 percent at worst. On the other hand, when prediction horizons were 6 h, the average of each extraction rate was 7.57 percent at best and 16.25 percent at worst. Nevertheless, the rea- son the computational complexity of SW-SVR is larger than linear SVR is that the increase of computational complexity due to build- ing several models is larger. However, since the amount of training data of each weak learner reduces substantially, the computational complexity to build one model in SW-SVR reduces also. Accord- ingly, when the number of models one CPU builds reduces by us- ing parallel processing, the computational complexity of the overall SW-SVR is lower than or equal to that of linear SVR. Meanwhile, as with linear SVR, SW-SVR never depends on the change of param- eters related to SVR, and the building time is always a constant. As mentioned in the above discussion, the building time of SW- SVR solely depends on the number of weak learners and training i eriods. Therefore, SW-SVR can avoid an unexpected long building ime in parameter tuning that changes each parameter variously. These results demonstrate that SW-SVR predicts complicated icrometeorological data with the best prediction performance nd the lowest computational complexity compared with standard lgorithms. In particular, we found that dynamic aggregation of odels built from very little extracted data by D-SDC is effective or compatibility of high prediction performance and low compu- ational complexity. However, there are problems to be solved in W-SVR. Firstly, the prediction performance of SW-SVR sometimes eteriorates despite an increase of training data. In particular, this roblem occurred under the conditions that prediction horizons re 6 h as shown in Fig. 3 . This is because data extracted by D- DC involves unnecessary training data for highly accurate predic- ion. If D-SDC extracts the same data as the extracted data when raining periods are shorter, the prediction performance of SW-SVR ever deteriorates due to an increase of training data. Therefore, e must review both feature mapping and algorithms of D-SDC o as to avoid extracting unnecessary training data. Meanwhile, W-SVR is based on a combination of several algorithms: kernel pproximation, PLS regression, k-means, D-SDC, and linear SVR. oreover, each algorithm has several parameters. Therefore, SW- VR has more varied parameters, and it takes more time to tune he parameters. In this experiment, we used a grid search roughly o as to decide the parameters in a certain time. However, there is till room for improvement in the prediction performance by us- ng other approaches such as a genetic algorithm instead of a grid earch ( Huang & Wang, 2006 ). . Conclusion and future work In this paper, we proposed a new methodology for predicting icrometeorological data, SW-SVR that involves a novel combina- ion of SVR and ensemble learning. To take the advantages of SVR nd ensemble learning, SW-SVR builds several SVRs specialized for ach representative data group in various natural environments by sing D-SDC that extracts effective training data for specific data rediction. Moreover, to follow micrometeorological data whose haracteristics always change with time, prediction of SW-SVR is ased on dynamically weighted ensemble learning depending on he similarity between test data and each data specialized by weak earners. As a result of evaluation experiments using large-scale icrometeorological data, the prediction performance of SW-SVR s greater than or equal to other general methods such as SVR, RF, Y. Kaneda, H. Mineno / Expert Systems With Applications 59 (2016) 217–225 225 a c f f m u w F v t c m e S A K R A B C C C F F F H J K K L M M M O P P P R S S S S T T U V W X nd GB. Moreover, SW-SVR reduces the building time substantially ompared with complicated models that have high prediction per- ormance. We anticipate that dynamic aggregation of models built rom various kinds of extracted data by D-SDC can contribute to ore sophisticated studies of micrometeorological data prediction. In future work, we should evaluate SW-SVR in more varied sit- ations to show that SW-SVR works effectively. In particular, we ill use more complicated data that consists of many features. urthermore, when SW-SVR is applied to applications such as en- ironmental control systems, the performance of overall applica- ions should be evaluated. Currently, we have developed an agri- ultural support system using SW-SVR, which controls environ- ents in greenhouses depending on the activity of the plants. The valuation of the applications will describe the superiority of SW- VR in practical use. cknowledgements This study was partially supported by JST, PRESTO , and JSPS AKENHI ( 26 6 60198 ), Japan. eferences ntonanzas, J. , Urraca, R. , Martinez-de-Pison, F. J. , & Antonanzas-Torres, F. (2015). Solar irradiation mapping with exogenous data from support vector regression machines estimations. Energy Conversion and Management, 100 , 380–390 . reiman, L. (2001). Random forests. Machine Learning, 45 (1), 5–32 http://doi.org/10. 1023/A:1010933404324 . ao, H., Naito, T., & Ninomiya, Y. (2008). Approximate RBF kernel SVM and its ap- plications in pedestrian classification. The 1st International Workshop on Ma- chine Learning for Visionbased Motion Analysis - MLVMA’08 , 1–9 http://hal. archives- ouvertes.fr/inria- 00325810/ . hang, C., & Lin, C. (2011). LIBSVM : A library for support vector machines. ACM Transactions on Intelligent Systems and Technology (TIST), 2 , 1–39 http://doi.org/ 10.1145/1961189.1961199 . hevalier, R. F. , Hoogenboom, G. , McClendon, R. W. , & Paz, J. A. (2011). Support vec- tor regression with reduced training sets for air temperature prediction: A com- parison with artificial neural networks. Neural Computing & Applications, 20 (1), 151–159 Retrieved from < Go to ISI>://WOS:0 0 028667480 0 015 . an, R.-E., Chang, K.-W., Hsieh, C.-J., Wang, X.-R., & Lin, C.-J. (2008). LIBLINEAR: A library for large linear classification. The Journal of Machine Learning, 9 (2008), 1871–1874 http://doi.org/10.1038/oby.2011.351 . reund, Y., & Schapire, R. (1997). A desicion-theoretic generalization of on-line learning and an application to boosting. Computational Learning Theory, 55 (1), 119–139 http://doi.org/10.1006/jcss.1997.1504 . riedman, J. H. (2001). Greedy function approximation: A gradient boosting ma- chine. Annals of Statistics, 29 (5), 1189–1232 . uang, C. L., & Wang, C. J. (2006). A GA-based feature selection and parameters optimizationfor support vector machines. Expert Systems with Applications, 31 (2), 231–240 http://doi.org/10.1016/j.eswa.2005.09.024 . apan Meteorological Agency. (n.d.).. Japan meteorological agency http://www.jma.go. jp/jma/indexe.html . isi, O., & Cimen, M. (2012). Precipitation forecasting by using wavelet-support vec- tor machine conjunction model. Engineering Applications of Artificial Intelligence, 25 (4), 783–792 http://doi.org/10.1016/j.engappai.2011.11.003 . olokotsa, D., Pouliezos, A., Stavrakakis, G., & Lazos, C. (2009). Predictive con- trol techniques for energy and indoor environmental quality management in buildings. Building and Environment, 44 (9), 1850–1863 http://doi.org/10.1016/j. buildenv.20 08.12.0 07 . oosli, G. (2007). Comments on the core vector machines : fast SVM training on very large data sets. The Journal of Machine Learning Research, 8 , 291–301 . acqueen, J. (1967). Some methods for classification and analysis of multivari- ate observations. In Proceedings of the fifth berkeley symposium on mathemati- cal statistics and probability: 1 (pp. 281–297). http://doi.org/citeulike- article- id: 6083430 . aity, R. , Bhagwat, P. , & Bhatnagar, A. (2010). Potential of support vector regression for prediction of monthly streamflow using endogenous property. Hydrological Processes, 24 (7), 917–923 . ohammadi, K. , Shamshirband, S. , Anisi, M. H. , Alam, K. A. , & Petkovi ́c, D. (2015). Support vector regression based prediction of global solar radiation on a hori- zontal surface. Energy Conversion and Management, 91 , 433–441 . thman, M. F., & Shazali, K. (2012). Wireless sensor network applications: A study in environment monitoring system. In Procedia Engineering: 41 (pp. 1204–1210). http://doi.org/10.1016/j.proeng.2012.07.302 . ark, D. H., & Park, J. W. (2011). Wireless sensor network-based greenhouse envi- ronment monitoring and automatic control system for dew condensation pre- vention. Sensors, 11 (4), 3640–3651 http://doi.org/10.3390/s110403640 . edregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Courna- peau, D., Brucher, M., Perrot, M. & Duchesnay, É. (2011). Scikit-learn: machine learning in python. The Journal of Machine Learning Research 12, 2825–2830. http://doi.org/10.1007/s13398- 014- 0173- 7.2 latt, J. C. (1998). Fast training of support vector machines using sequential minimal optimization. Advances in Kernel Methods , 185–208 http://doi.org/10.1109/ISKE. 2008.4731075 . ahimi, A., & Recht, B. (2007). Random features for large-scale kernel machines. Advances in Neural Information Processing Systems, 20 , 1177–1184 http://doi.org/ 10.1.1.145.8736 . ingh, K. P., Gupta, S., & Rai, P. (2013). Identifying pollution sources and predicting urban air quality using ensemble learning methods. Atmospheric Environment, 80 , 426–437 http://doi.org/10.1016/j.atmosenv.2013.08.023 . mith, B. A., Hoogenboom, G., & McClendon, R. W. (2009). Artificial neural networks for automated year-round temperature prediction. Computers and Electronics in Agriculture, 68 (1), 52–61 http://doi.org/10.1016/j.compag.20 09.04.0 03 . uzuki, Y. , Kaneda, Y. , & Mineno, H. (2014). SW-SVR improved by short-distance data collection method (pp. 1–8) IPSJ SIG Technical Report, 2014-MBL-73(9) . uzuki, Y., Kaneda, Y., & Mineno, H. (2015). Analysis of support vector regression model for micrometeorological data prediction. Computer Science and Informa- tion Technology, 3 (2), 37–48 http://doi.org/10.13189/csit.2015.030202 . enenhaus, M., Vinzi, V. E., Chatelin, Y. M., & Lauro, C. (2005). PLS path modeling. Computational Statistics and Data Analysis, 48 (1), 159–205 http://doi.org/10.1016/ j.csda.20 04.03.0 05 . sang, I. W., Kwok, J. T., & Cheung, P.-M. (2005). Core vector machines: Fast SVM training on very large data sets. Journal of Machine Learning Research, 6 , 363– 392 http://doi.org/10.1111/j.1442- 9993.2007.01810.x . rraca, R. , Antonanzas, J. , Martinez-de-Pison, F. J. , & Antonanzas-Torres, F. (2015). Estimation of solar global irradiation in remote areas. Journal of Renewable and Sustainable Energy, 7 (2), 023136 . apnik, V. N. (1995). The Nature of Statistical Learning Theory : Vol. 8. Springer http: //doi.org/10.1109/TNN.1997.641482 . ang, B. X., & Japkowicz, N. (2009). Boosting support vector machines for imbal- anced data sets. Knowledge and Information Systems, 25 (1), 1–20 http://doi.org/ 10.1007/s10115- 009- 0198- y . ie, Y., Li, X., Ngai, E. W. T., & Ying, W. (2009). Customer churn prediction using improved balanced random forests. Expert Systems with Applications, 36 (3 PART 1), 5445–5449 http://doi.org/10.1016/j.eswa.2008.06.121 . http://dx.doi.org/10.13039/100005197 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0001 http://doi.org/10.1023/A:1010933404324 http://hal.archives-ouvertes.fr/inria-00325810/ http://doi.org/10.1145/1961189.1961199 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0005 http://doi.org/10.1038/oby.2011.351 http://doi.org/10.1006/jcss.1997.1504 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0008 http://doi.org/10.1016/j.eswa.2005.09.024 http://www.jma.go.jp/jma/indexe.html http://doi.org/10.1016/j.engappai.2011.11.003 http://doi.org/10.1016/j.buildenv.2008.12.007 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0012a http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0012a http://doi.org/citeulike-article-id:6083430 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0015 http://doi.org/10.1016/j.proeng.2012.07.302 http://doi.org/10.3390/s110403640 http://doi.org/10.1007/s13398-014-0173-7.2 http://doi.org/10.1109/ISKE.2008.4731075 http://doi.org/10.1.1.145.8736 http://doi.org/10.1016/j.atmosenv.2013.08.023 http://doi.org/10.1016/j.compag.2009.04.003 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0022 http://doi.org/10.13189/csit.2015.030202 http://doi.org/10.1016/j.csda.2004.03.005 http://doi.org/10.1111/j.1442-9993.2007.01810.x http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30178-6/sbref0026 http://doi.org/10.1109/TNN.1997.641482 http://doi.org/10.1007/s10115-009-0198-y http://doi.org/10.1016/j.eswa.2008.06.121 Sliding window-based support vector regression for predicting micrometeorological data 1 Introduction 2 Related work 2.1 Support vector regression 2.2 Ensemble learning 3 SW-SVR: Sliding window-based support vector regression 4 Evaluation 4.1 Experiment 4.2 Results and discussion 5 Conclusion and future work Acknowledgements References 
work_txjggjbpvjgfvcf7orekdbf6o4	----	CWI Scanprofile/PDF/300 Centrum voor Wiskunde en lnformatica Centr~ for Mathematics and Computer Science M. Bezem Semantics and consistency of rule-based expert systems ~ Computer Science/Department of Software Technology Report CS-R8824 July The Centre for Mathematics and Computer Science is a research institute of the Stichting Mathematisch Centrum, which was founded on February 11, 1946, as a nonprofit institution aim- ing at the promotion of mathematics. computer science, and their applications. It is sponsored by the Dutch Government through the Netherlands Organization for the Advancement of Pure Research (Z.W.0.). Copyright ~ Stichting Mathematisch Centrum, Amsterdam ,. &q !.( l ~I"°:} I< 14 1bl3 \~ 4 { Semantics and Consistency of Rule-based Expert Systems Marc Bezem Centre for Mathematics and Computer Science P.O. Box 4079, 1009 AB Amsterdam, The Netherlands Consistency of a knowledge-based system has become a topic of growing concern. Every notion of con- sistency presupposes a notion of semantics. We present a theoretical framework in which both the seman- tics and the consistency of a knowledge base can be studied. This framework is based on first order flat many-sorted predicate logic and is sufficiently rich to capture an interesting class of rule-based expert sys- tems and deductive databases. We analyse the feasibility of the consistency test and prove that this test is feasible for knowledge bases in Horn format without quantification. This report is a revised and extended version of report CS-R8736. The extension concerns the use of expressions in rules. 1980 Mathematics Subject Classification: 68T30, 68T15. 1987 CR Categories: 1.2.3, 1.2.4, F.4.1. Key Words & Phrases: knowledge-based systems, rule-based expert systems, knowledge representation, consistency. Note: The material for this article has been taken from the PRISMA documents P176 and P297. The research is part of the PRISMA project (PaRallel Inference and Storage MAchine), a joint effort with Philips Research Eindhoven, partially supported by the Dutch "Stimulerings-projectteam lnformatica-onderzoek" (SPIN). INTRODUCTION The plan of this paper (a revised and extended version of [BJ) is as follows. First the reader is intro- duced to the knowledge representation used in rule-based expert systems. We shall indicate some semantical problems in relation to this knowledge representation. Then we explain in an informal way how many-sorted predicate logic comes in. In the next section we describe syntax and semantics of many-sorted predicate logic. We assume some knowledge of first order logic. Thereafter we shall be able to characterize rule-based expert systems as first order theories. The Tarski semantics solves the semantical problems mentioned above. Furthermore we shall derive several results on decidability and consistency of rule-based expert systems. Unfortunately, some natural equality and ordering axioms are not in Hom format (see [Re] for a discussion on the domain closure axiom). Hence test- ing consistency with a standard theorem prover would be very inefficient. In the fourth section we describe a technical device, a certain kind of null value, which allows feasible consistency testing in the presence of equality and ordering axioms, which are not in Hom format. The underlying idea is partiality of functions, thus avoiding the search for consistent function values in a (possibly gigantic) product space. This kind of null value, being quite different from null values as described in [IL], appears to be new. In the last section we show how to extend our results to knowledge bases with arithmetical (and other) expressions in the conditions of the rules. We shall focus our attention on rule-based expert systems, but the techniques can also be applied to deductive databases with func- tional dependencies. Report CS-R8824 Centre for Mathematics and Computer Science P.O. Box 4079, 1009 AB Amsterdam, The Netherlands 2 1. RULE-BASED EXPERT SYSTEMS l. L In rule-based expert systems shells such as EMYCIN [BS] or DELFI2 [L], knowledge about some specific domain can be expressed in facts and in rules of the form IF antecedent THEN consequent. Facts are so-called object-attribute-value triples, or <o,a, v > triples for short. The antecedent of a rule is a conjunction of disjunctions of conditions, and conditions are definite statements, such as same, notsame and less than, about <o,a,v) triples. We restrict ourselves to rules having as consequent a conjunction of conclusions of the form conclude <o,a, v >. In most cases so-called certainty factors are associated with the facts and the rules. Certainty factors range from LOO (definitely true) to - LOO (definitely false). ~e certainty factor of a fact expresses a measure of certainty about that fact, whereas the certainty factor of a rule scales the measure of certainty about the consequent with respect to the measure of certainty about the antecedent. In DELFI2 an object tree (called context tree in MYCIN) is used to state properties of and relations between different objects, which cannot be expressed by the rules. The nodes of this tree are objects, labeled by their attributes and respective domains of values. The path from a node to the root of this tree constitutes the context of that node. In other words: the objects occurring in the subtree of a node are sub-objects of the object belonging to that node. Furthermore it is stated in the object tree whether an attribute is singlevalued or mul- tivalued. The interpretation of the knowledge in rule-based expert systems is more operational than declara- tive: same <o,a,v> is true if and only if <o,a,v> occurs as fact (with certainty factor > 0.2), conclude <o,a, v > has the effect that <o,a, v > is added as fact (with appropriate certainty factor), and if the antecedent of a rule evaluates to true, then that rule may be fired, i.e. all conclusions occurring in the consequent are executed. We remark that same <o,a,v> and conclude <o,a,v> have the same declarative meaning as the fact <o,a,v>, i.e. attribute a of object o has value v. L2. Consider the following real-life example extracted from HEPAR, an expert system for the diag- nosis of liver and biliary disease, built with DELFI2. IF same <patient,complaint,colicky yain> THEN conclude <patient,pain,colicky> (LOO) IF same <patient,abdyain,yes) AND same <pain, character, continuous) THEN conclude <patient,pain,colicky > (-LOO) IF same <patient,complaint,abdominalyain> OR same <patient,pain,colicky> THEN conclude <patient,abdyain,yes) (LOO) These three rules (from a rule base consisting of over 400 rules) show two objects, patient and pain, four attributes, namely complaint, pain and abd _pain of patient and character of pain, as well as several values. 1.3. A first observation which can be made is the following. Five items in the rules above refer to pain: the values colicky _pain and abdominal _,.Pain, the attributes pain and abd _pain of the object patient, and the object pain, which is a sub-object of patient, as stated by the object tree of HEP AR. 3 The interrelations between these items do not seem to be expressible by the formalism. A second observation on the three rules above is their inconsistency in the presence of the facts <patient,complaint,colicky _pain> (1.00) and <pain,character,continuous> (LOO). The inference engine reasons backwards, using a simple loop check to prevent infinite looping. Depending on the presence of the fact <patient,complaint,abdominal_pain> (1.00), the inference engine did or did not hit upon the contradiction <patient,pain,colicky> (1.00 and -1.00). In both cases the contradiction was completely ignored. These observations show a defect of the knowledge representation used in rule-based expert sys- tems, namely the lack of a dear semantics. 1.4. Basically our approach amounts to interpreting (o,a,v> by a(o,v) in the multivalued, and by a( o) = v in the singlevalued case. Here o and v are constants for elements of a domain 0 of objects and a domain V of values. In the multivalued case a denotes a relation, i.e. a subset of 0 X V, and in the singlevalued case a function from 0 to V. If o is a sub-object of o', then o' is added as argument of a. 1.5. The following examples show how to extend our interpretation of <o,a, v > triples to atoms. conclude <patient, complaint, abdominal _pain> becomes complaint (patient,abdominal _pain) less _than <patient,temperature,36.8> becomes temperature (patient) < 36.8 same <pain, character, continuous> becomes character (patient,pain) = continuous Note that the fact that pain is a sub-object of patient is expressed by adding patient as an argument of the function character. Another example, based on the rules of 1.2, is to be found in 4.6. We feel that the existing formalisms for handling uncertainty are unsatisfactory. As we do not have a good alternative we leave the subject aside. 1.6. Under the interpretation described above, a rule-based expert system becomes a theory in first order many-sorted predicate logic, in short: a many-sorted theory. To keep this paper self-contained we give a short, introductory description of the syntax and semantics of many-sorted predicate logic in the next section. In Section 3 we shall characterize expert systems as many-sorted theories of a cer- tain type. This approach has the following advantages: - The declarative semantics of the expert system becomes perfectly clear, being the Tarski semantics of the associated many-sorted theory. - Logical concepts such as consequence, consistency and decidability get a clear meaning in relation to the expert system. - Theorem proving techniques for testing consistency, such as resolution, become available for the expert system. 4 The choice for many-sorted instead of one-sorted logic is motivated as follows: - It is natural to make a distinction between numerical data (ordered by <)and symbolic data (usu- ally unordered). - Subdividing a set of constants into sorts and typing the predicate and function symbols considerably reduces the number of well-formed formulas, which is important for debugging large knowledge bases. - In the presence of variables the many-sorted approach imposes restrictions on unification, which considerably reduces the resolution search space (see for example [W]). We shall not make use of this advantage in the present paper. 2. MANY-SoRTED PREDICATE LoGIC 2.1. The syntax of many-sorted predicate logic extends the syntax of ordinary, one-sorted, predicate logic by having a set of sorts ~. instead of just one sort. Moreover we have variables xr and constants er for all sorts GE~. Furthermore we have function symbols fl' x ... Xam--><Jo (m>O), where the notion of type G1 X · · · XGm~Go replaces the notion of arity from the one-sorted case. We also have proposi- a X ••· Xa f f tion symbols p; and predicate symbols P;' m (m>O) 0 type G1 x ... XGm· Terms are ormed from variables and constants by function application (respecting the sorts). Atoms are either proposi- tion symbols or the application of a predicate symbol to terms of appropriate sorts. With the help of propositional connectives and quantifiers, atoms are combined into formulas. The sets ~. CONS, FUNC, PROP and PRED of, respectively, sorts, constants, function symbols, proposition symbols and predicate symbols, form together the similarity type of a specific many-sorted predicate calculus. For practical reasons we assume that the similarity type is finite. 2.2. A many-sorted structure 01L consists of: (a) A non-empty set Aa for each GE~, called the domain of sort G. (b) For each constant er an element Cj E Aa. (c) For each function symbol fl'x · · · xam-->ao a mapping]; : Aa, X · · · XAam ~Aa.· ( d) For each proposition symbol p; a truth value p;. _ (e) For each predicate symbol Pf'x ··· xam a mapping P; : Aa, X · · · XAam ~{TRUE,FALSE}. 2.3. An assignment in 01Lis a mapping a assigning to each variable xr an element a(x7) of Aa. 2.4. The interpretation in 01L of a term t under an assignment a, denoted by ./'ff(t) or t for short, is inductively defined as follows: (a) xr = a(x7) (b) er = Cj () j'OX···Xa-->a0 ( ) f-(- -) c J ;' m t,,. .. ,tm = ; ti. ... ,Im The truth value in 01L of an atom Pf'x · · · xam(t 1,. • .,tm) under an assignment a is given by P;(ti. ... , tm). 2.5. The truth value in 01L of a formula F under an assignment a, denoted by ~(F), is inductively defined as follows: (a) If Fis an atom, then ~(F) is given by 2.4. (b) ~ respects the truth tables of the propositional connectives. (c) ~<:rfxr F) = TRUE if and only if for all assignments a', which differ at most on xr from a, we have Cl,f/}(F) = TRUE. (d) ~(3xr F) = TRUE if and only if there exists an assignment a', which differs at most on xr from a, such that Cl,f}(F) = TRUE. 5 2.6. A formula Fis true in~ denoted by l='!JR,F, if ci.ff(F) = TRUE for all assignments a. 2.7. A sentence (or closed formula) is a formula without free variables (i.e. variables which are not bound by a quantifier). It will be clear that for sentences S the truth value ci.ff(S) does not depend on the assignment a. As a consequence we have either l='!Jf(,S or l='!JR,-,S, for every sentence S. Let SENT denote the set of sentences. · 2.8. Let f C SENT. ~is called a model for f, denoted by l='!JR,f, if l='!JR,S for all S ef. 2.9. S E SENT is called a (semantical) consequence of (or implied by) r c SENT if for all many- sorted structures~ we have: if l='!JR,f, then l='!JR,S. This will be denoted by f 1= S (or 1= S if r is empty). Furthermore we define the theory off as the set Th(f) = {S E SENT I f 1= S}. 2.10. f c SENT is called consistent if f has a model. Th(f) is called decidable if there exists a mechanical decision procedure to decide whether a given sentence S is a semantical consequence of r or not. 2.11. Two many-sorted structures are called elementarily equivalent if exactly the same sentences are true in both structures. 2.12. REMARKS. 2.12.1. We refrain from giving an axiomatization of many-sorted predicate logic since our main con- cern will be model theory. Most textbooks on mathematical logic provide a complete axiomatization of ordinary (one-sorted) predicate logic. It suffices to generalize the quantifier rules in order to obtain an axiomatization of flat many-sorted predicate logic. 2.12.2. Of course, one-sorted predicate logic is a special case of many-sorted predicate logic. As a consequence, the latter is as undecidable as the former. More precisely: 1= S is undecidable, provided that the similarity type is rich enough ( CHuRcH, TuruNG, 1936, see also [M, 16.58)). 2.12.3. Conversely, many-sorted predicate logic can be embedded in one-sorted predicate logic by adding unary predicate symbols S" (x ), expressing that x is of sort o, and replacing inductively in for- mulas VxY F (resp. 3x7 F) by Vx(S"(x)~F) (resp. 3x(S"(x)AF)). Let A' be the one-sorted sentence obtained from A E SENT in this way. It can be proved (see [M]) that I= A if and only if r I= A', where r = {3xS"(x)loe~} expresses the fact that the domains are non-empty. This embedding allows us to generalize immediately many results on one-sorted predicate calculus to the many-sorted case (e.g. the compactness theorem). We shall not make use of this possibility in the present paper. 3. RULE-BASED ExPERT SYSTEMS AS MANY-SORTED THEORIES 3.0. We propose the following terminology for certain kinds of many-sorted theories: - Indexed propositional expert systems. - Universally quantified expert systems. 3.1. An indexed propositional expert system is a many-sorted theory axiomatized by: (a) Explicit axioms (the rule base and the fact base), which are boolean combinations of atoms of the f P11,X ••• Xu.,( ') f th f 1<11,x ••• X11.,)-->l10( ') - II ( < II > II) orm c,. .. ,c or o e orm c, ... ,c - 110 c resp. 110 c , 110C , with constants c, .. .,c1,c11 of appropriate sorts. Such atoms (here and below called constant-atoms, or c-atoms for short) may be viewed as indexed propositions, which explains the name. Note that we conform to the convention to write =, < and > as infix predicates. 6 (b) Implicit axioms for equality of each sort and ordering of each sort for which an ordering is appropriate. The axioms for =0 , equality of sort 11, are (loosely omitting sort super- and sub- scripts): 'r:/x(x =x), Vxi,x 2(x 1 =x2~X2 =x 1), 'r:/xi,x2,x 3((x1 =x2/\x2 =x3)~x 1 =x 3 ), 'r:/xi,x2((x 1 =x2/\F(x1))~F(x2)) for formulas F, -,ci=cj for O~i=/=j~n, 'r:/x(x =c0 v · · · Vx =en), the domain closure axiom. These axioms express that = is a congruence relation on a finite domain, where every element has exactly one name. Let EQ denote the set of equality axioms for all sorts 11, then we have by definition that either EQ 1= c;=cj or EQ I= -,c;=cj for all i,j. The axioms for< and> are: 'r:/xi,x2,X3((x1 <x2/\x2<x3)~x1 <x3), 'r:/x(-,x<x), Vxi,x 2(x 1 <x2 Vx 1 =x 2 Vx2 <x 1) for sorts which are totally ordered, Vxi,x2(X1 <x2 B x2>x1), a subset A of { c; <cjl O~i=/= j ~n} U {-,ci <cjl O~iJ ~n} (see below). These axioms express that < is a transitive, irreflexive and (possibly) total ordering with inverse >. Let 0 denote the set of ordering axioms. We require that A is such that either 0 1= ci<cj or 0 I= -,c; <cj, for all i,j. It follows in particular that EQ U 0 is consistent. . The idea behind the implicit axioms is that = and < are provided by the system and have a fixed meaning, whereas the other predicates are user-defined. We put IA = EQ U 0, so IA denotes the set of implicit axioms. 3.2. A universally quantified expert system differs from an indexed propositional one by allowing not only constants, but also variables in the explicit axioms. All explicit axioms are assumed to be univer- sally closed. 3.3. Among the theories that do not fall under 3.1 and 3.2 are many theorem provers. A theorem prover might be an undecidable theory. The theoretical observations below show that, from a logical point of view, expert systems as defined in 3.1 and 3.2 are very simple, decidable theories. 3.4. LEMMA. For every S E SENT there exists a boolean combination S' of closed atoms such that EQ l=S BS'. PROOF. Replace inductively every subformula 'r:/xF(x) by .f\c0)/\ · · · /\F(cn) and 3x.f\x) by F(c0)v · · · v F(cn) where 'r:/x(x =c0 v · · · Vx =en) E EQ. D 3.5. LEMMA. For every closed atom A there exists a boolean combination A' of c-atoms such that EQ l=A BA'. PROOF by giving a typical example. Let A be for instance P(f(f_'(c 00 ),c0 '),c 02 ) with f of type 112 X 111 ~u0 and f' of type 110~112 . Let A 3 be the sentence 3x02 3y 00 lf'(c 00 )=x 02 /\f(x 0 2,c 0 ' )= y 00 /\P(y 00 ,c02 )]. Now apply Lemma 3.4 to A 3 and obtain A'. D / 7 3.6. Both lemmas above can cause combinatorial explosions. Therefore they are only of theoretical use. They tell us that, in the presence of EQ, boolean combinations of c-atoms have the same expres- sive power as full many-sorted predicate logic. 3.7. LEMMA. If EQ C f C SENT, then every model off is elementarily equivalent to a model whose domains consist of exactly the interpretations of all constants. PRooF. Let~ be a model off with EQ C f C SENT. By EQ the interpretations of the equality predicates in ~ are equivalence relations which are congruences with respect to the interpretation of all other predicate and function symbols in ~ It follows that ~ and ~ I=, the quotient structure of ~ modulo equality, are elementarily equivalent. By the axioms Vx(x =c0 v · · · vx =en) and -.c; =cj (i=/=j) from EQ, it follows that the domains of~ I= consist of exactly the interpretations of all constants. D 3.8. LEMMA. If EQ C f C SENT, then Th(f) is decidable. PRooF. Let f be such that EQ C f C SENT. Then we can apply Lemma 3.7 and observe that there are just finitely many non-isomorphic ~ = 's. In other words: up to dividing out = and identifying individuals with the same name there are just finitely many different models of r. Moreover all models of r are finite. Hence we can, for any given S E SENT, test in finite time whether S holds in all models of r or not. D 4. TESTING CONSISTENCY 4.1. Definition 2.10 and Lemma 3.7 suggest the following procedure for testing the consistency of theories f with EQ c f: generate all many-sorted structures whose domains consist of exactly the interpretations of all constants, and test each time whether the many-sorted structure is a model of f or not. This procedure is in general not feasible. A first step towards a f easihle consistency test is the alternative characterization of consistency for indexed propositional expert systems, described in the following paragraphs. Let r be an axiomatization of an indexed propositional expert system (see 3.1). Let Pi, ... ,Pm be a list of all c-atoms of the form P(c, ... ,c') occurring in r, and let ti. ... ,tn list all terms f(c,. . .,c') occur- ring in r. The idea behind the following construction is that it is not necessary to have an entire many-sorted structure to be able to interpret r. Let A 0 = {c3, .. .,c~} be the set of all constants of sort aE~. Equality of sort a is interpreted by syntactical identity on A 0 • Then all equality axioms from EQ are satisfied. The ordering (if any) on A 0 is induced by 0, i.e. c;<cj if and only if 0 I= e;<c·. We need the following notions: - A truth valuation of P 1,. .. ,Pm is an assignment of either TRUE or FALSE to each P; (1-s;;;ios;;;;m). - A valuation of tl>···•tn is an assignment of a unique element of the appropriate domain A 0 to each tj (1-s;;;jos;;;;n). It will be clear that a truth valuation of P1> .. .,Pm and a valuation of t1> .. .,tn suffice for an interpretation of f. Each model off yields a truth valuation of P1,. . .,Pm and a valuation of IJ> .. .,tn. Conversely, each truth valuation of PJ> .. .,Pm and valuation of t 1,. • .,tn for which every explicit axiom of r is true, can be extended in an arbitrary way to a many-sorted structure~ 1= f. Thus we have the following THEOREM. Let conditions be as above. Then we have: r is consistent if and only if there exists a truth valuation of P 1>··· ,Pm and a valuation oft 1,. . ., tn for which every explicit axiom of r is true. 4.2. Theorem 4.1 suggests a simple algorithm for testing the consistency of an indexed propositional expert system f: generate all valuations and truth valuations and test each time whether the explicit 8 axioms of f are satisfied or not. This clearly leads to combinatorial explosions. A second step towards a feasible consistency test is imposing language restrictions on our expert systems. A common language restriction is Horn format. A clause is a finite disjunction of atoms and nega- tions of atoms (so called positive and negative literals). A conjunctive normal form is a finite conjunc- tion of clauses. A Horn clause is a clause containing at most one positive literal. A unit clause is a clause consisting of one positive literal. The connection between Horn clauses and production rules is easily seen by the equivalence of (A 1 /\ • • • /\Ak) ~ B and A 1 v · · · v A;; v B+, where the super- scripts + and - denote whether a literal occurs positively or negatively. However, the implicit axioms 'Vx(x =coV · · · Vx =en) and 'Vxi,x2(x1 <x2Vx1 =x2Vx2<x1) are not Horn clauses. As a consequence we can only require the explicit axioms to be Horn clauses. This is not sufficient for feasible consistency checking, as will be demonstrated in the next paragraph. The idea is to reduce the satisfiability problem for propositional logic, which is known to be NP-complete (see [GJ]), to the consistency problem of indexed propositional expert systems, all whose explicit axioms are Hom clauses. Let a be a sort with exactly two constants: c8 and cY. Then we have 'Vx(x =co Vx =c 1)eEQ. For any propositional atom A, let f A be a function symbol of type a~a. It is not difficult to see that {-.fA(co)=coVA,fA(co)=coV-.A}UEQ 1= hfA(co)=ci) BA. So the positive literal A is equivalent to the negative literal •fA(co)=c1 in any f containing the two Horn clauses -.fA(c0 )=co VA and fA(co)=coV-.A and the equality axioms EQ. Let C be any propo- sitional conjunctive normal form. Let f c be the indexed propositional expert system with explicit axioms -.fA(c0 )=c0 vA and fA(c 0 )=coV-.A for all positive literals A occurring in C. Then f c is consistent and satisfies f c 1= CBHc, where He is the set (conjunction) of Horn clauses obtained from C by replacing all positive literals A by their equivalent negative literal. Moreover we have that C is satisfiable if and only if f c UH c is consistent. This yields a polynomial reduction of the satisfiability problem for propositional logic to the consistency problem of indexed propositional expert systems, all whose explicit axioms are Horn clauses. A similar reduction could be established using 'Vxi,x2(x 1 <x2 Vx1 =x2 Vx2<x1) instead of 'Vx(x =co Vx =c1). 4.3. In the previous subsection we showed that testing consistency is NP-hard without further language restrictions. The problem is to specify language restrictions, which are strong enough to guarantee a feasible consistency test, and still allow enough expressivity for some given application. This third and final step towards a feasible consistency test will be achieved in the following theorem by the introduction of constants undefined, which are different from (wrt. =) and incomparable with (wrt. <) any other constant. These constants are used for the simulation of partiality of functions, thus avoiding the search for consistent function values in a (possibly gigantic) product space of sorts. Further intuition will be given in 5.4.1. THEOREM. Let f be the axiomatization of an indexed propositional expert system satisfying the following conditions: (I) All explicit axioms off are Horn clauses. (2) For every sort ae~ there exists at least one constant, which does not occur in the explicit axioms of f. Such a constant, say c8, will be denoted by undefined0 • So we have: 'Vx(x =undefinedvx =c1 V · · · Vx =cn)eEQ and -.undefined=cieEQfor all I:o;;;;;;i:o;;;;;;n. (3) For every sort ae~ the ordering (if any) of sort a is partial with respect to undefined0 • More precisely: we require that 0 I= -.undefined<c; and 0 I= -.undefined>c;for a/l I:o;;;;;;i:o;;;;;;n. (4) The orderings < and> do not occur in the positive literals occurring in the explicit axioms off. Then the consistency off can be tested in polynomial time. PRooF. List the explicit axioms of r as follows (the super- and subscripted capitals denote c-atoms): At Aj CI,"1 V · · · VCI,"n, vDt Ci,1 V • · • VCi,n, V Di BI,"1 V • • • VBI,"m, B;:1 v · · · v B;:m, 9 Now apply the following algorithm, which is in fact a special case of hyper-resolution (see [R]). Under the conditions of the theorem, positive literals are either of the form P(c, ... ,c'), or of the form f(c, ... ,c')=c''. The current set of unit clauses is denoted by Us. Anticipating the generalization in Sec- tion 5 the formulation of the algorithm is slightly more general than necessary: "EQ U Us is consistent'' simply says that we do not have f(c, ... ,c')=c; E Us and f(c, ... ,c')=cj E Us with i=/::j, and "c-atom X is implied by IA U lfs" is the case if either XeU80 or X is of the form f(c, ... ,c')<c; (resp. >c;) and we have f(c, ... ,c')=cjEUs with 0 ~ cj<c; (resp. 0 ~ cj>c;). (* generation of unit clauses *) lfs :={At,·· -,Aj}; Unew := {}; REPEAT Us:= UsU Unew; Unew := {}; IF EQ U Us is consistent THEN FOR each clause c-:-1 v · · · vc-:- vD-:t-1, 1,n; 1 DO cancel all C;"J such that Cij is implied by IA U Us; IF all negative literals of the clause are cancelled THEN Unew := UnewU{Dt} OD UNTIL 10 (* test if the unit clauses form a model *) IF EQ U Us is consistent AND there exists no clause Bi:1 V · · • V Bi:m; such that all B"/j are implied by IA U Us THEN f is consistent ELSE f is inconsistent The algorithm terminates since the number of literals is strictly decreasing. By the cancellations the explicit axioms of f are transformed into an equivalent set of Hom clauses. The consistency in the THEN-part can be seen as follows. First note that EQ U Us is consistent. Secondly, every clause not in Us contains a negative literal L - whose complement L + is not implied by IA U Us. It can happen that IA UUs implies L - (for example f(c)=2EUs, L - =-,f(c)<l, 1<2EdC0). Then every clause containing L - holds in any model of IA U Us. If neither L - nor L + is implied by IA U Us, then we have either L + =P(c, ... ,c')fl.Us, or L + contains a function term f(c, ... ,c') such that f(c, ... ,c')=c"EUs for no constant c". We define the following (truth-) valuation (see 4.1): { TRUE if P;EUs P; = FALSE otherwise t·= J s { c if t·=c EU. 1 undefined otherwise The valuation is well-defined since EQ U Us is consistent. By the properties of the constants undefined (conditions (2) and (3) of the theorem), the literals L - mentioned above are TRUE in the (truth-) valuation. Hence every clause containing L - is TRUE in the above (truth-) valuation, which consti- tutes (in the sense of Theorem 4.1) a model of the entire set of clauses. The algorithm is clearly qua- dratic in the number of occurrences of literals. To get the consistency off in complexity class P we tacitly assumed that the primitive operations used in the algorithm above are in P. D REMARKS. 4.3.1. As follows by close inspection of the proof above, it would suffice to require the following weakening of condition (2): for every sort CJ for which a function symbol f of type ... ~CJ occurs in a negative literal occurring in the explicit axioms off, there exists at least one constant which does not occur on the right-hand side of an equation occurring in such a literal. However, we think it is more systematic to require condition (2) as it stands. In practice the constants undefined are simply added to every domain of constants occurring in the knowledge base. 4.3.2. Condition (4) can not be missed, which can be seen as follows. Assume for some sort CJ we have exactly three constants different from undefined, which are totally ordered by c 1 <c 2 <c3 • Then the unit clause /(c)>c 1 is equivalent to /(c)=c 2 V f(c)=c 3 , which enables a similar construction as in the third paragraph of 4.2. 4.3.3. In view of condition (1), occurrences of notsame (o,a,v> in the antecedent of a rule can be problematic, since negative conditions in a rule result in positive literals in the clausal form. This problem can be overcome by postponing the consistency test until these occurrences evaluate to either TRUE or FALSE (e.g. after querying the user). Then the knowledge base can be transformed into an equivalent knowledge base satisfying (1). An alternative would be to allow clauses with more than one positive literal and to apply the following lemma. Of course the consistency test then becomes in gen- eral NP-hard, the worst-case time complexity doubles with every application of the lemma and we can only hope that practical cases are not the worst. 11 LEMMA. Let r be an axiomatization of an indexed propositional expert system. Let furthermore, for some c-atom A, r 1 be obtained from r by omitting (an arbitrary number of) explicit axioms which are in clausal form (not necessarily Horn format) and contain A as positive literal, and f 2 by deleting A from (an arbitrary number of) such axioms. Then we have: r is consistent if and only if {A } Ur 1 is consistent or r 2 is consistent. PROOF. As to the if-part we remark that {A}Uf1 1= rand f 2 1= r. As to the only-if-part, note that for every many sorted structure~ either ~A or 1=~.A. So if~ is a model of r then~ is either a model of {A} Uf1 or of f 2 , by the definition of fi. f 2 • 0 4.3.4. In our opinion the domain closure axiom is realistic in many applications (apart from the fact that every computer is a finite automaton). The lemmas 3.4-3.8 essentially depend on it. However, Theorem 4.3 remains valid when the domain closure axioms are removed from EQ. For, detected con- tradictions do not depend on the domain closure axioms and, conversely, consistency with the domain closure axiom trivially implies consistency without it. 4.4. Let us briefly discuss the semantical consequences of the conditions (2) and (3) from the previous theorem, since they may slightly deviate from the intended meaning of the knowledge base. It is possi- ble that the consistency of r essentially depends on the valuation f(c, ... ,c')=undefined, i.e. that any valuation f(c, ... ,c')=c; (l:s;;;i:s;;;n) would not yield a model for r. One could say that j(c, ... ,c')=undefined possibly saves the expert system from inconsistencies by preventing production rules with occurrences of f(c, ... ,c') in the antecedent from firing. Since such rµles have obviously not been used in the inference, this may be considered an advantage. On the other hand, this may be con- sidered a disadvantage in cases where f(c, ... ,c')=undefined is not realistic (e.g. temperature (patient) = undefined). In these cases we suggest to add the appropriate unit clause f(c, ... ,c')=c; and to test consistency again. 4.5. Conceptually speaking it is not difficult to generalize the algorithm from 4.3 to universally quantified expert systems, although some care has to be taken in quantifying x in clauses containing literals of the form -,f(c, ... ,c')=x. In these cases only restricted quantification of the form 'r/x=j=undefined is allowed. Of course, the algorithm from 4.3 can become very inefficient (from P(c0), P(c 1), -,P(xi)V · · · V-,P(xn)VQ(xi, ... ,xn), for instance, 2n instances of Q are generated), so we suggest limited use of variables (or, preferably, the use of a more efficient algorithm). 4.6. Ex.AMPLE. We demonstrate the consistency test in action on the case of 1.2. Taking into account the object tree (sub-objects, single/multivalued attributes) the three rules yield the following clauses in shorthand: -,C(p,cp )V P(p,c ), -,f ahdp(p)= yesV-,f char(p,pa)=coV-,P(p,c), -,C(p,ap)V f abdp(p)= yes,-,P(p,c)v f ahdp(p)= yes. When we add the unit clauses C(p,cp) and fchar(p,pa)=co, the empty clause is easily derived and the algorithm decides to inconsistency. If we only add C(p,cp ), then the algorithm decides to consistency on the basis of the following (truth-)valuation: C(p,cp)=P(p,c)=TRUE, C(p,ap)=FALSE, f aMp(P )=yes, fchar(p,pa)=undefined 12 5. ExPRESSIONS 5.1. The extension of the notion of indexed propositional expert system we discuss in this section aims at supporting arithmetic. We start with a real-life example, again taken from HEPAR. A simi- lar example can be found in [BS], pg. 297. greater _than <biochemistry, ((total_ bili - direct_ bili)I total_ bili)* I 00, 60 > This example shows an extension of the object-attribute-value representation by allowing attribute expressions instead of only attributes. A straightforward extension of the interpretation of <o,a, v > tri- ples into many-sorted predicate logic yields the following atom in shorthand: ((tb(b )-db(b)))!tb(b ))* 100>60 In order to capture attribute expressions, the notion of indexed propositional expert system has to be extended. To this end we add new function symbols, called operators and denoted by op,·+, ... , to our language. Moreover the implicit axioms are extended with the positive diagram of each operator op, i.e. with all defining equations op(c;, ... ,cj) = CJ(i, ... ,j) , where the constants are assumed to be of the sort that is required by the type of op. Here f does not belong to the language but denotes a function of indices describing the definition of op. Like = and <, operators are provided by the system. Let OP denote the set of axioms for operators. A typical example of an operator is addition of integers, denoted by + and written as an infix operator. From now on IA will denote EQ U 0 U OP. 5.2. We now extend the notion of c-atom as defined in Section 3 in the following way: a c-atom is a closed atom of the form P(c, ... ,c') or of the form: expression relation symbol expression Here a relation symbol is one of the symbols < , = , > and an expression is a term built up from constants and terms of the form f(c, ... ,c') by applying the operators (respecting types and sorts). 5.3. From now on indexed propositional expert systems as defined in 3.1 are understood to be extended in the sense of 5.1 and 5.2 above. In order to be able to carry over Theorem 4.3 we intro- duce the following definition. DEFINITION. Assume every sort has a constant undefined as stated above in the conditions (2) and (3) of Theorem 4.3. An operator op is called strict if the implicit axioms for op satisfy the following condi- tion: op(c, ... ,c')= undefined if one or more of the arguments c, ... ,c' equals undefined. Note that for a strict multiplication * of integers we have O* undefined = undefined and not O* undefined = 0. 5.4. THEOREM. Let r be an axiomatization of an indexed propositional expert system satisfying the con- ditions ( 1)-(4) of Theorem 4.3 above. Assume moreover that the following conditions hold: (5) Every operator is strict and takes polynomial time. ( 6) No operators occur in the positive literals occu"ing in the explicit axioms of r. (7) All equations occu"ing in negative literals in the explicit axioms of r are of the form expression = c. Then the consistency of r can be tested in polynomial time. 13 PROOF. Due to its general formulation, the algorithm from 4.3 carries immediately over to the present case. However, the argument demonstrating the correctness of this algorithm is slightly more compli- cated, and needs to be restated in some detail. We first note that the evaluation of expressions (such as the example in 5.1) may require more than one unit clause (both the values of tb(_b) and db(_b) are necessary). Furthermore, under the conditions (4) and (6) of the theorem, positive literals are of the form P(c, ... ,c'), f (c, ... ,c')=c", f (c, ... ,c')= g(d, ... ,d'), or c =c'. The latter kind of literals evaluate immediately to TRUE or FALSE (by using EQ), so we may assume without loss of generality that they do not occur in the explicit axioms of r. Moreover, under condition (7) of the theorem, negative literals are essentially of the form -, P(c, ... ,c'), .., expression I < expression2 or .., expression =c. By this more general form of literals involved, as well as the fact that IA = EQ U 0 U OP instead of IA = EQ U 0, the statements "EQ U Us is consistent" and "c-atom X is implied by IA U Us'' have a more general meaning than in 4.3. For example, the former statement now also excludes f(c)= l,f(c)= g(d),g(d)=2E'Us, whereas the latter holds for X the c-atom from 5.1, with -, I, *and > having their usual meaning axiomatized in IA, and tb(b )= 100,db(b )= 39E Us. The consistency in the THEN-part follows from the following (truth-) valuation: { TRUE if P;EUs P; = FALSE otherwise { c if EQU Us t: tj=c tj = undefined otherwise Again we have that EQ U Us is consistent and that every clause not in Us contains a negative literal L - whose complement is not implied by IA U Us. It can happen that IA U Us implies L - (for exam- ple L - = -,f(c)+ f(c')<O, and f(c)= f(c'),f(c')= I E Un with + and < their usual meaning). Then every clause containing L - holds in any model of IA U Us. If neither L - nor L + is implied by IA U Us, then we have either L + =P(c, ... ,c')l1. Us, or L + contains a function term f(c, ... ,c') such that EQ U Us does not imply f(c, ... ,c')=c'' for any constant c". By the properties of the constants undefined (conditions (2) and (3) of the theorem), the fact that every operator is strict (condition (5) of the theorem), as well as the syntactic restriction on negative literals (condition (7) of the theorem), the literals L - mentioned above are TRUE in the (truth-) valuation. Hence every clause containing L - is TRUE in the above (truth-) valuation, which constitutes (in the sense of 4.1) a model of the entire set of clauses. D REMARKS. 5.4.1. The intuitive idea behind the conditions (1), (4) and (6) is that the positive information, which can be seen as the driving force of the algorithm, should be definite. Omission of (1), (4) or (6) would allow the introduction of incomplete positive information of the form, respectively, A+ v B+, f(c)<2, or f(c)+ f(c')=3. The other conditions (2), (3), (5) and (7) concern the simula- tion of partiality by means of the constants undefined, used for the circumvention of combinatorial explosions that would be caused by the search for consistent function values in a (possibly gigantic) product space of sorts. Thus all conditions are devoted to avoiding combinatorial explosions which would occur in the presence, implicit or explicit, of disjunctions of positive literals. 5.4.2. In Section 4 it is shown that the conditions (1)-(4) cannot be missed. It will be clear that condi- tion (5) of Theorem 2.5, being quite natural, can also not be missed. We shall demonstrate further- more that neither (6) nor (7) can be missed, by showing that the consistency test becomes NP-hard (see (GJ]) in cases in which (6), respectively (7), do not hold. This will be done in 5.4.3, respectively 5.4.4, by showing that certain disjunctions of positive literals are semantical consequences of indexed propositional expert systems with explicit axioms in Hom format. Then a similar argument as developed in 4.2, immediately yields the desired result. 14 5.4.3. As to (6), let fJ be a sort with exactly two constants, TRUE and FALSE, different from undefined/J. Let eq be a strict equality operator of type pxp~p, i.e. eq yields undefined if one or both of its arguments equals undefined and otherwise eq yields TRUE if its arguments are equal, and FALSE else. It is easily seen that (with f of type ... ~ft) EQUOPU{eq(f(c), f(c))=TRUE} I= f(c)=TRUE v f(c)=FALSE. 5.4.4. As to (7), let o be a sort with exactly one constant 0 different from undefineda. We trivially have (with f and f' of type ... ~o) EQU{-,f(c)= f'(c')} I= f(c)=O V j'(c')=O. At first sight condition (7) seems to be a nasty restriction, since it excludes conditions of the form f(c)= f'(c') in the rules. The reason is that =,being reflexive, does not propagate partiality properly (undefined =undefined evaluates to TRUE). There is, however, nothing against using strict equality operators in conditions, i.e. eqaxa-.fJ(f(c ), f'(c')) = /J TRUE instead of f(c) =a f'(c'). CONCLUSION We argued how to interpret a rule-based expert system as a many-sorted theory. Then the Tarski semantics yields a well-defined notion of consistency, and theorem proving techniques such as resolu- tion can be used for testing consistency. The relevance for expert systems lies in the fact that incon- sistencies make the conclusions of a knowledge-based system highly unreliable. For, what reason do we have to believe a conclusion if its negation could also be derived? We provided means for detecting inconsistencies, but did not discuss how to deal with them. Note that we considered an expert system as a fixed theory, whereas in practice the theory grows during the interaction with a user answering questions posed by the system (in this way the inconsistency of 1.2 was obtained). In our opinion only relatively consistent answers are to be accepted by the system. Starting with a consistent knowledge base we thus maintain consistency as an invariant of the interaction. Recent experiments with a PROLOG implementation of the consistency test and a medical knowledge base show that the models, produced by the consistency test at several moments· during the interaction between the user and the system, are very informative. REFERENCES [B] M. BEZEM, Consistency of rule-based expert systems. Proceedings of the 9-th Conference on Automated Deduction, Springer Lecture Notes in Computer Science 310, pp. 151-161, Springer- Verlag, Berlin (1987). [BS] B.G. BUCHANAN, E.H. SHORTLIFFE, Rule-based expert systems: the Mycin experiments of the Stan- ford Heuristic Programming Project. Addison-Wesley, Reading, Massachusetts (1984). [GJ] M.R. GAREY, D.S. JOHNSON, Computers and intractability: a guide to the theory of NP- completeness. Freeman, San Francisco, California (1979). [IL] T. IMIELINSKI, W. LIPSKI JR., Incomplete information in relational databases. Journal of the ACM 31, 4, pp. 761-791 (1984). [L] P.J.F. LUCAS, Knowledge representation and inference in rule-based expert systems. Report CS- R8613, Centre for Mathematics and Computer Science, Amsterdam (1986). [M] J.D. MONK, Mathematica/ Logic, Springer-Verlag, Berlin (1976). [R] J.A. ROBINSON, Automatic deduction with hyper-resolution. International Journal of Computer Mathematics 1, pp. 227-234 (1965). [Re] R. REITER, Equality and domain closure in first order databases. Journal of the ACM 27, 2, pp. 235-249 (1980). [W] C. WALTHER, A many-sorted calculus based on resolution and paramodulation. Pitman, London (1987). 
work_txvr4bsdxfcjpe4icgjngqo7la	----	Microsoft Word - 26_BankoZoltan.doc Magyar Kutatók 8. Nemzetközi Szimpóziuma 8th International Symposium of Hungarian Researchers on Computational Intelligence and Informatics 295 Correlation Based Dynamic Time Warping Zoltán Bankó, János Abonyi Department of Process Engineering, University of Pannonia, Veszprém, Hungary www.fmt.uni-pannon.hu/softcomp, e-mail: abonyij@fmt.uni-pannon.hu Abstract: Dynamic Time Warping (DTW) is a widely used technique for univariate time series comparison. This paper proposes a new algorithm for the comparison of multivariate time series which generalize DTW for the needs of correlated multivariate time series. Keywords: multivariate time series, principal component analysis, dynamic time warping, segmentation 1 Introduction A time series is a sequence of values measured as a function of time. These kinds of data are widely used in the field of process engineering, medicine, bioinformatics, chemistry and finance. The increasing popularity of knowledge discovery and data mining tasks for discrete data such as clustering (partitioning the data into coherent, but not predefined subsets), classification (placing the data into predefined, labeled groups) and kNN (k-nearest neighbor) search has indicated the growing need for similarly efficient methods for time series databases. These methods share a common requirement: a similarity measure has to be defined between the elements of a given database. The similarity of multivariate time series can be approached from two different perspectives. The first way is the application of metrics based warping measures such as Dynamic Time Warping (DTW) and Longest Common SubSequence (LCSS). These techniques are perfectly suitable for univariate tasks like speech recognition, where the analyzed process represented by one variable only. In most cases these methods can be easily generalized for the needs of the multivariate time series, where the process depends on two or more variables. However, their application for comparing correlated multivariate time series is often not as effective as it is expected. The direct comparison of the variables used by these approaches ignores the hidden process, i.e. the correlation between process variables and often this hidden process carries the real information. Hence, Principal Component Analysis (PCA) based similarity measures are used to overcome this problem. PCA considers the Z. Bankó et al. Correlation Based Dynamic Time Warping 296 time series as a whole but does not take into account the alterations in the relationship between the variables. Our main goal is to construct a similarity measure which not only compares the variables directly, but the process behind them and is able to deal with the changes in the correlation structure of the variables. 1.1 Contribution In this paper the problem of creating similarity measures for multivariate time series is investigated. We present a new and intuitive multivariate dynamic time warping method which is based on the correlation of variables and uses principal component analysis. This makes our similarity measure invariant to phase shifts of the time axis and it is capable to compensate the differences in sampling rates and ‘locally elastic’ shifts (local time warping) of the time series. We do not examine the data validity and we omit the preprocessing steps for brevity. We treat that the time series are sampled (or can be resampled) at equispaced time points. The rest of the paper is organized as follows. Section 2 details the theoretical background of the proposed similarity measure, i.e. dynamic time warping, principal component analysis and segmentation. In Section 3 we introduce our novel similarity measure which allows to combine the DTW and PCA, while Section 4 conducts a detailed empirical comparison of our method with basic techniques on verification databases widely used by the time series data mining community [1]. Finally in Section 6 we present our conclusions and suggestions for future work. 2 Background An n-variable, m-element time series, [ ]nn xxxX ,,, 21 K= , is an m-by-n element matrix, where [ ]Tiiii mxxxx )(,),2(),1( K= , is the ith variable and )( jxi its jth element. [ ])(,),(),()( 21 jxjxjxjX nn K= is the jth sample of nX . The similarity between nX and nY is denoted by ),( nn YXs , where 1),(0 ≤≤ nn YXs , ),(),( nnnn XYsYXs = and 1),( =nn XXs . Obviously, the similarity is nothing more than a real number between zero and one which expresses the tightness of connection between the processes behind the time series. The closer is the number to one the more similar the processes are. In Magyar Kutatók 8. Nemzetközi Szimpóziuma 8th International Symposium of Hungarian Researchers on Computational Intelligence and Informatics 297 practice, the term distance or dissimilarity ( d ) is used instead of similarity. The value of the distance is given by a number ranged from zero to one. This can be associated with similarity: ),(1),( nnnn YXsYXd −= . Note that is not expected from an applied distance to satisfy the triangle inequality when the distance is not used as the basic distance of an indexing method. 2.1 Dynamic Time Warping Getting valuable information from time series data by data mining methods require a good similarity measure; even so, the traditional approaches are rarely precise enough for the most applications. This is caused by the brittleness of the conventional comparison methods. Distance measures such as Euclidean distance, are unable to handle the distortions in time axis, so these distortions almost randomly affect the distance between time series. The solution for this problem is the application of dynamic time warping which can ‘warp’ the original time series (nonlinearly dilate or compress their time axes) to be similar in shape to the query series as much as possible. In the following the DTW algorithm is reviewed for univariate time series x and y . To align these sequences, firstly we define a grid D with size n x m where each cell represents a distance between the appropriate indices of the two time series. In this step we can choose any application-dependent distance like L1 and L∞ norms, but generally Euclidean distance is suggested because it allows of the efficient indexing of dynamic time warping [2]. Considering this we have: 2 2))()((),( jyixjiD −= (1) Using grid D, we can create arbitrary mappings – called warping paths – between x and y. The construction of a warping path [ ])(,),2(),1( lppp K has to be restricted with the following constraints: - Boundary condition: The path has to start in )1,1(D and end in ),( mnD . - Monotonicity: The path has to be monotonous, i.e. always heading from )1;1(D to ),( mnD . If ),()( jiDkp = and )','()1( jiDkp =+ then 0' ≥−ii and 0' ≥− jj . - Continuity: The path has to be continuous. If ),()( jiDkp = and )','()1( jiDkp =+ then 1' ≤−ii and 1' ≤− jj . Z. Bankó et al. Correlation Based Dynamic Time Warping 298 Figure 1 The cumulative distance matrix and the optimal warping path on it To find the optimal warping path (the DTW distance of the two time series), every warping path has an assigned cost which is the sum of values of the affected cells divided by the normalization constant K: ⎟ ⎟ ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎜ ⎜ ⎝ ⎛ = ∑ = K ip yxd l i DTW 1 )( min),( (2) The value of K depends on the application and in most cases this is the length of the path, but it can also be omitted. More information about the method of defining K and its significance can be found in Reference [3]. Note that the Euclidean distance is a special case of DTW, i.e. we choose the path that is located on the diagonal of grid D and K = 1. Obviously, the number of paths exponentially grows by the size of time series. Fortunately, the optimal path can be found in O(nm) time with the help of dynamic programming using cumulative distance matrix D̂ . A cell of the cumulative matrix contains the sum of the appropriate cell value in matrix D and the minimum of the three cells from where the cell can be reached: ⎥ ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎢ ⎣ ⎡ +−− +− +− = ),()1,1(ˆ ),()1,(ˆ ),(),1(ˆ min),(ˆ jiDjiD jiDjiD jiDjiD jiD (3) The DTW distance between the two time series can be found in ),(ˆ mnD . Magyar Kutatók 8. Nemzetközi Szimpóziuma 8th International Symposium of Hungarian Researchers on Computational Intelligence and Informatics 299 2.2 PCA Similarity Factor PCA is a well known and widely used multivariate statistical technique for reducing the dimensionality of the data. Fundamentally, this is a linear transformation that projects the original (Gaussian distributed) data to a new coordinate system with minimal loss of information. In multivariate cases the information is the structure of the original data, i.e. the correlation between the variables and the alteration of the correlation structure among them. To create a projection based on this, PCA selects the coordinate axes of the new coordinate system one by one according to the greatest variance of any projection. In projected space the first coordinate (first principal component) is the vector which has maximal variance in the original data space, the second coordinate (second principal component) has maximal variance in the original data space among all vectors which orthogonal to the first coordinate, the third coordinate orthogonal to the first two and has maximal variance in original data space, etc. Technically speaking, we select the eigenvectors of the covariance matrix of the original data (according to the descending order of corresponding eigenvalues) for axes in the new dimensional space. Figure 2 Dimension reduction with the help of PCA from 3 (filled dots) to 2 (patched dots) dimensions. Notice that the correlation between the dots is maximally preserved. In most real multivariate processes the high-order principal components do not add any significant information to the low-order components so we can omit them to perform dimensionality reduction. Krzanowski defined [4] the PCA similarity factor to measure the similarity between different data by comparing the hyperplanes (the dimensionality reduced, new coordinate systems). Z. Bankó et al. Correlation Based Dynamic Time Warping 300 Let nX and nY two multivariate time series with n variable. pX nU , and pYnU , indicates the matrices of eigenvectors which belong the most important p eigenvalues of nX 's and nY 's covariance matrix, i.e. the two hyperplanes. The similarity between them: p UUUUtr YXs pX T pYpY T pX nnPCA nnnn )( ),( ,,,,= (4) The similarity factor has a geometrical explanation, because it measures the similarity between the two hyperplanes by computing the squared cosine values between all the combinations of the first p principal components from pX nU , and pYn U , : ∑∑ = = Θ= p i p j jinnPCA p YXs 1 1 , 2cos 1 ),( (5) where ji,Θ is the angle between the ith principal component of nX and the jth principal component of nY . 2.3 Segmentation The ith segment of nX is a set of consecutive time points, [ ])(;);1();(),( bXaXaXbaS nnni K+= . The c-segmentation of time series nX is a partition of nX to c non-overlapping segments, [ ]),1(;);,1();,1( 21 mkSbaSaSS ccX n ++= K . In other words, a c- segmentation splits nX to c disjoint time intervals, where a≤1 and mk ≤ . The main goal of segmentation is to find homogenous segments by the definition of a cost function. This function can be any arbitrary function which projects from the space of multivariate time series to the space of the positive real numbers and can be framed in several ways [5]. Abonyi et al. presented a PCA based segmentation in Reference [6], where the authors used the Hotelling’s T2 statistics and the Q reconstruction error as the measure of the homogeneity of the segments, i.e. the cost function. Magyar Kutatók 8. Nemzetközi Szimpóziuma 8th International Symposium of Hungarian Researchers on Computational Intelligence and Informatics 301 Figure 3 The distance measures used in PCA based segmentation 2.4 Related Work The data mining community has been investigating different similarity measures for univariate time series for a long time. The majority of these works have focused on the Minkowski metrics (Lp norms), especially on the Euclidean distance. This popularity does not only arise from its simple geometric representation and fast computability, but also from the fact that the Euclidean distance is preserved under orthogonal transformations such as Fourier and Wavelet Transform and satisfies the triangle inequality. From the viewpoint of data mining the main advantage of these transforms is that the original time series – in most cases – are well approximated using the first few coefficients of their transformed forms and the relations of the distances between the elements are also preserved. Technically speaking, the Euclidean distance allows the efficient indexing of large time series databases. The GEMINI framework generalized the creation of the indexing methods, thus, it provided facilities for new approaches which held out higher accuracy. In parallel with this, the increasing and less and less costly computational power made it possible to create similarity measures without indexing capabilities for tasks where the quality of the result is much more important than the speed of the calculation. Hence, new Z. Bankó et al. Correlation Based Dynamic Time Warping 302 techniques such as Dynamic Time Warping, Longest Common Subsequences and Symbolic Aggregate approXimation has been studied extensively from the requirements of ECG signal analysis through handwriting recognition and verification to tornado prediction. On the other hand, the PCA based similarity factor also gains in popularity thanks to its outstanding property which makes possible to recognize the direction of the changes in the distance between the variables. Another advantage of this method is its optimal variable reduction behavior which makes it ideal for tasks with numerous variables such as process engineering jobs. Johannesmayer [6] modified the PCA similarity factor by weighting each principal component with the square root of its eigenvalue: , )( cos)( ),( 1 , 2 1 1 ∑ ∑∑ = = = Θ = p i Y i X i ji p i p j Y j X i nnPCA nn nn YXs λλ λλ λ (6) where nXiλ and n Y iλ are the corresponding eigenvalues for the ith and jth principal component of nX and nY . This principle was developed by K. Yang [7] who presented the logical extension of PCA similarity factor and λPCAs called Eros (Extended Frobenius Norm): ∑∑ == Θ== n i ii n i T YXinnEros wiUiUwYXs nn 11 ,cos)()(),( (7) where iΘcos is the angle between the two corresponding principal components and iw is the weighting vector which is based on the eigenvalues of the sequences in the data set. 3 Correlation Based Dynamic Time Warping of Multivariate Time Series In the previous chapters, all of the required techniques have been reviewed for correlation based dynamic time warping of multivariate time series. Magyar Kutatók 8. Nemzetközi Szimpóziuma 8th International Symposium of Hungarian Researchers on Computational Intelligence and Informatics 303 3.1 Proposed Algorithm The following steps have to be performed for the realization of correlation based multivariate time warping: • Segment the time series of the given database by correlation to create the most homogenous segments. The projection error or Hotelling’s statistic can be used as the cost function. This segmentation can be done off-line in most cases. • Segment the query time series when it arrives similarly. • Calculate the distance between the query and the time series of the database with dynamic time warping. The base distance of DTW can be any distance which is based on the previously presented covariance based similarities (e.g. 1− PCAs ). The size of matrix D̂ (the computation time of DTW) depends on the number of the segments. 4 Evolution Method and Experimental Results A modified leave-one-out k-nearest neighbor search algorithm [8] was used to perform the validation of multivariate dynamic time warping. After the algorithm executes the recall-precision graph, which has been frequently used to measure the performance of Content Based Image Retrieval (CBIR) systems as well as Information Retrieval (IR) systems, can be plotted. The precision expresses the proportion of the relevant sequences from all the sequences are retrieved. Similarly, the recall is the percentage of the total relevant documents in a database retrieved by the k-nearest neighbor search. 4.1 Comparison of Similarity Measures According to Reference [9], two dataset were used for the validation and both of them are available from the Internet. The first one is the database of SVC2004 (Signature Verification Conference 2004) which contains 40 sets of signature data1. This is an ideal dataset for correlation based multivariate time warping because the correlation alternates many times between the variables. 1 Each set contains 20 genuine signatures from one signature contributor and 20 skilled forgeries from five other contributors. Although both genuine signatures and skilled forgeries are available, obviously the 800 genuine signatures were used for the validation. Z. Bankó et al. Correlation Based Dynamic Time Warping 304 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Recall P re ci si on CbDTW SegPCA SPCA Figure 4 The Recall-Precision graph of SVC2004 dataset Firstly, the best projection was selected for similarity factor based comparison. The best result was the one which compares the 2 dimensional hyperplanes. The first two eigenvalues describe more than the 99% of the total variance. This case is signed by the red line. After this, the sequences were segmented into 20 parts and the distances between the corresponding segments were calculated with the help of the Krzanowski similarity factor and the segmentation cost was the Hotelling’s T2 statistics. The graph of SegPCA shows that this technique really outperforms the segmentation free version. For fairness to this method, the correlation based DTW (CbDTW) used the same segmentation while performed the non-linear pairing between the segments. The superiority of this method is obvious for this dataset. Magyar Kutatók 8. Nemzetközi Szimpóziuma 8th International Symposium of Hungarian Researchers on Computational Intelligence and Informatics 305 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Recall P re ci si on CbDTW SegPCA SPCA Figure 5 The Recall-Precision graph of AUSLAN dataset The second validation was processed on the Australian Language Sign dataset (AUSLAN) collected by Mohammed Waleed Kadous [10]. This contains 95 signs and each of them has 27 examples, so the total number of sings is 2565. The average length of the time series in the database is only 57. In Figure [5], the high values show that the correlation between the variables is much stronger than in SVC2004 dataset and this characterize the classes much more. The data was projected into 2 dimensional hyperplanes similarly to the first validation. These principal components describe more than the 95% of the total variance. For the last two methods, interpolation has to be executed and segmentation was applied on these interpolated versions of the original data. The segmentation cost was selected to Hotelling’s T2 statistic and the number of segments was 20. Conclusions The paper presented a new similarity measure for multivariate time series by combining dynamic time warping and principal component analysis. The results of the validations are more then promising, however, there is still a lot of work that has to be done and there are lots of questions that have to be answered. The other similarity measures (such as Eross and λ PCAs ) do not provide as good results Z. Bankó et al. Correlation Based Dynamic Time Warping 306 as the PCA similarity factor. Presumably the reason for this is the segmentation. It has to be examined how the best segmentation result can be achieved for a given similarity measure. Acknowledgement The authors acknowledge the financial support of the Cooperative Research Centre (VIKKK, project 2004-I), the Hungarian Research Found (OTKA 49534), the Bolyai János fellowship of the Hungarian Academy of Science, and the Öveges fellowship. Thanks to Prof. Eamonn J. Keogh for providing datasets from Time Series Data Mining Archive. References [1] S. Hettich, S. D. Bay: The UCI KDD Archive, http://kdd.ics.uci.edu, University of California, Department of Information and Computer Science, 1999 [2] E. Keogh: Exact Indexing of Dynamic Time Warping, in Proceedings of 28th VLDB Conference, Hong Kong, China, 2002, pp. 406-417 [3] H. Sakoe, S. Chiba: Dynamic Programming Algorithm Optimization for Spoken Word Recognition, IEEE Transactions on Acoustics, Speech, and Signal Processing, Vol. 27(1), 1978, pp. 43-49 [4] W. J. Krzanowski: Between-Groups Comparison of Principal Components, in Journal of the American Statistical Society, pp. 74:703-707, 1979 [5] E. Keogh, S. Chu, D. Hart, M. J. Pazzani: An Online Algorithm for Segmenting Time Series, in Proceedings of ICDM, 2001, pp. 289-296 [6] J. Abonyi, B. Feil, S. Németh, P. Árva: Principal Component Analysis- based Time Series Segmentation, in Proceedings of the IEEE International Conference on Computational Cybernetics, 2003 [7] M. C. Johannesmeyer: Abnormal Situation Analysis Using Pattern Recognition Techniques and Historical Data, M.sc. thesis, University of California, Santa Barbara, CA, 1999 [8] K. Yang, C. Shahabi: A PCA-based Similarity Measure for Multivariate Time Series, in Proceedings of the 2nd ACM International Workshop on Multimedia Databases, 2004, pp. 65-74 [9] E. Keogh, S. Kasetty: On the Need for Time Series Data Mining Benchmarks: a Survey and Empirical Demonstration, in Proceedings of the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2002, pp. 102-111 [10] Mohammed Waleed Kadous: Temporal Classification: Extending the Classification Paradigm to Multivariate Time Series, PhD thesis, School of Computer Science & Engineering, University of New South Wales, 2002 
work_tz77goofobcfnnvng4gkkwvpjq	----	Abstract—Quay crane scheduling problem (QCSP) is the problem of the allocation of quay cranes to handle the unloading and loading of containers at seaport container terminals and defining the service sequence of vessel bays of each quay crane. The treatment of crane interference constraints and the increased in vessel size make the problem difficult to solve. Due to the growing interest in applied research for this problem, many researchers have used different algorithms and methods to obtain some solutions. This paper will propose a modified genetic algorithm combined with priority rules to deal with it. The advantage of the proposed algorithm comes from the fact that crane interference constraints, non-simultaneous constraints and precedence constraints can all be handled in simple ways and is to provide practical solutions. The effectiveness and reliability of the approach are tested using several benchmark instances proposed by Meisel and Bierwirth (2011). A comparison with the current best-known solutions reveals that the proposed method is capable of finding the optimal or near-optimal solution in reasonable computing time. Index Terms—Crane scheduling, genetic algorithm, optimization, priority rules. I. INTRODUCTION Containers are large boxes that are used to transport goods from one destination to another by several transportation systems. Transport containers oversea are carried out by vessels. Container terminals are the place at which vessels are charged and discharged. To perform container discharging and charging operations from a vessel when it is berthed, quay cranes (QCs) are critical resources in container terminals. The QCs move along the quay on rails to take/put containers off/on the deck and holds. They are the most expensive equipment at terminals and can be used both at an automated and a manned terminal. Nowadays due to the large growth rates of sea transportation, terminals are faced with more and more containers to be handled. Moreover, because of the high competition between container terminals, improving QCs operating efficiency is very important to achieve high productivity (short completion time at low cost) and customer satisfaction. Fig. 1 shows a picture of quay cranes operating to load or unload containers to and from the vessel. QCSP refers to finding an optimal schedule for QCs serving a vessel in which the allocation of quay cranes to handle the unloading and loading tasks, sequencing these Manuscript received September 9, 2018; revised January 15, 2019. Vi H. Nguyen is with Faculty of Engineering, Vietnamese-German University, Binh Duong, Vietnam (e-mail: vi.nh@vgu.edu.vn). Duc T. Nguyen is with Civil and Environmental Engineering (CEE) Department, Old Dominion University, Virginia, USA (e-mail: dnguyen@odu.edu). tasks on each QC, and task’s operation start time is specified to get the scheduling targets. It was firstly described by Daganzo (1989). He stated that the QCSP seems related to the machine scheduling field, parallel machine scheduling problem, and a resource-constrained project scheduling problem with precedence constraints but more complicated. His paper examined crane scheduling for ports, studied the problem of QC assignment among different holds on multiple vessels to minimize the ships’ delay at berth. As a result, crane idle time is minimized and berth throughput maximized, and this reduces queuing delay as well [2]. In a related study, Peterkofsky and Daganzo (1990) proposed a branch and bound algorithm to deal with the QCSP problem with the objective of minimizing total weighted tardiness. However, in these two studies, QCs interference and safety distance between two adjacent QCs which are important features of QC operation are not considered [3]. Fig. 1. A drawing of quay cranes working on vessel [1]. Lim et al. (2004) further considered spatial constraints which related to the relative positions of cranes and jobs in QCSP. Three types of spatial constraints are identified which are non-crossing constraints (or inference constraint), neighborhood constraint (or safety distance constraint) and job-separation constraint (or non-simultaneous constraint). They adopted dynamic programming, tabu search, as well as squeaky wheel optimization heuristics to solve the problems with both separately and simultaneously constraints cases [4]. Kim and Park (2004) formulated the quay crane scheduling problem and proposed a branch and bound method and a heuristic algorithm, Greedy randomized adaptive search procedure for solving the QCSP. In the formulation, the QCSP with multiple tasks involved in a ship-bay is considered with crane interference, safety distance constraints as well as non-simultaneous and precedence relations among tasks [5]. Moccia et al. (2006) revised Kim and Park’s model by incorporating travel times for the quay cranes that subsequently process tasks in the same bay. They proposed a branch-and-cut algorithm as the Marine Quay Crane Scheduling Using a Combined Modified Genetic Algorithm and Priority Rules Approach Vi H. Nguyen and Duc T. Nguyen International Journal of Materials, Mechanics and Manufacturing, Vol. 7, No. 1, February 2019 doi: 10.18178/ijmmm.2019.7.1.422 21 solution methodology and applied to solve the same benchmark instances used by Kim and Park (2004). Their approach yielded better solutions than Kim and Park (2004) and is able to handle the bigger size of practical problems [6]. Later on, Sammarra et al. (2007) solved the QCSP, which have been studied by Kim and Park (2004) and Moccia et al. (2006) under the same assumption made, by separating the problem into two parts, routing problem and the scheduling problem. They proposed a local search algorithm for the scheduling sub-problem and a Tabu Search (TS) algorithm for routing problem. Their methods can obtain almost optimal solutions in the medium-size problem with reasonable computation time [7]. Bierwirth and Meisel (2009) revised the QCSP model proposed by Sammarra et al. (2007) and developed a heuristic based on the branch-and- bound algorithm. In comparison with existing methods, their proposed heuristic outperformed in terms of both the objective function value and the computation time [8], [9]. Meisel and Bierwirth (2011) proposed a unified approach for evaluating the performance of different model classes and solution procedures. In their paper, they also proposed a set of benchmark instances with a computational results document for each instance [1]. S.H. Chung and K.L. Choy proposed a modified genetic algorithm to deal with the QCSP. Their model is based on the model developed by Kim and Park (2004) and the modification made by Moccia et al. (2006). They used a new approach for defining the chromosomes: a chromosome to have two parts. In the first part, from left to right, genes represent the tasks and their performance sequence whereas in the second part, they represent the quay cranes assigned to each task of the first part. They adopted order crossover proposed by Gen and Cheng (1996) and swapping mutation in their proposed algorithm. Then set of well-known benchmarking data obtained from Kim and Park (2004) for testing and comparing. They yielded that the proposed algorithm performs as good as many existing algorithms and it is much faster than the existing approaches [10]. Narges Kaveshgar et al. (2012) extends the research in the QCSP area by utilizing the GA that is available in the latest version of Global Optimization Toolbox in MATLAB 7.13 to facilitate development. The mathematical formulation used for the QCSP in their paper is based on the one developed by Kim and Park (2004). By using an initial solution based on the S-LOAD rule developed by Sammarra, Cordeau, Laporte, and Monaco (2007), using a new approach for defining the chromosomes (i.e., solution representation) to reduce the number of decision variables, and using new procedures for calculating tighter lower and upper bounds for the decision variables, they improved the efficiency of the GA search. Their proposed GA is capable of finding the optimal or near-optimal solution in a significantly shorter time when comparing to the current best-known solutions of benchmark instances proposed by Meisel and Bierwirth (2011) [11]. All of the above algorithms require the users and readers to have a strong background in mathematics and programming. With the goal to develop a simple and effective method to solve the QCSP, this paper will propose a modified generic algorithm combined with simple priority rules which do not require mathematical formulation. By testing benchmark instances which are downloaded from [15], and comparing our results with the current best-known solutions, it can be shown that the proposed method can get to the optimal or near-optimal solutions in reasonable computing time. II. PROBLEM DESCRIPTION In the QCSP for container groups, there is a set of tasks Ω = {1, 2… n} and a set of QCs Q = {1, 2. . . q}. Each task i ∈ Ω represents a loading or discharging operation of a certain container group. The tasks have individual processing times and bay positions. They must be processed by a QC without preemption. Pairs of tasks which are located within the same bay may have precedence relation among tasks: when discharging and loading operations must be performed at the same ship-bay, the discharging operation must precede the loading operation. When discharging operation is performed in a ship-bay, tasks on the deck must be performed before tasks in the hull of the same ship-bay are performed. Also, the loading operation in a hold must precede the loading operation on the deck of the same ship-bay [3]. Crane interference constraints are involved in the QCSP because QCs are rail mounted: All QCs can move between two adjacent bays in an identical travel time t> 0. They are not allowed to cross each other and have to keep a safety margin δ, measured in units of bays at any time. Certain tasks cannot be performed simultaneously because, at the same time, no two QCs can operate at the same bay or at two bays that do not satisfy the safety distance requirement. The objective of QCSP is to determine the allocation of quay cranes to handle the unloading and loading tasks, sequencing these tasks on each QC, and task’s operation start time so that the completion time of a ship operation is minimized. TABLE I: A SMALL QCSP INSTANCE [9]. Task index i 1 2 3 4 5 6 7 8 9 Processing time 22 46 8 70 10 38 40 16 12 Bay position 1 1 2 3 5 5 7 9 11 Precedence- constrained task ( ) ( ){ }6,5,2,1=φ Non- simultaneous tasks ( ) ( ) ( ) ( ) ( ){ }6,5,4,3,3,2,3,1,2,1=ψ QC1 0 1 ,1 1 0 == rl QC2 0 2 ,4 2 0 == rl QC travel time t=1 Safety margin 1=δ Table 1 shows the data for a small QCSP instance which is used in the paper “A fast heuristic for quay crane scheduling with interference constraints” by Christian Bierwith and Frank Meisel to illustrate the proposed solution procedure. The problem contains nine container groups placed in a vessel with eleven bays. Two QCs are assigned to this vessel. The interpretation of the above restriction is as follow ( ) ( ){ }6,5,2,1=φ : Task 1 must precede task 2; Task 5 must International Journal of Materials, Mechanics and Manufacturing, Vol. 7, No. 1, February 2019 22 precede task 6. ( ) ( ) ( ) ( ) ( ){ }6,5,4,3,3,2,3,1,2,1=ψ : Task 1 and 2 cannot be performed simultaneously, etc. QC1 01,110 == rl : Location of QC1 at time 0 is 1, i.e. bay 1. Ready time of QC1 is 0 QC2 02,420 == rl : Location of QC2 at time 0 is 4, i.e. bay 4. Ready time of QC2 is 0 t=1: 1 time unit to travel from bay to bay 1=δ : Safety margin between any 2QC’s is 1 bay III. COMBINE MODIFIED GENERIC ALGORITHM AND PRIORITY RULES Genetic algorithm (GA) is a stochastic search method that mimics the metaphor of natural biological evolution. Genetic algorithm starts with no knowledge of the correct solution and depends completely on responses from its environment and evolution operators (i.e. reproduction, crossover, and mutation) to arrive at the best solution [12]. In this paper, traditional GA will be modified then combine with the priority rule to solve the QCSP. A. Chromosome Representation GA starts with a random population of individuals which are called chromosomes. Each chromosome represents a possible solution. A chromosome of the modified GA represents a sequence handling of tasks in which a gene is a task number. Chromosome 2 3 1 6 7 8 4 9 5 Handling Sequence 1 2 3 4 5 6 7 8 9 Fig. 2. A sample chromosome. B. Validation of Chromosome with Precedence- Constrained Check Chromosome to make sure that it satisfies the precedence constraint. If it does not satisfy, the constraint then swap the task positions to make it satisfactory. In the example, ( ) ( ){ }6,5,2,1=φ means task 1 must be processed before task 2 and task 5 must be handled before task 6. Therefore, the sample chromosome is modified as follow Fig. 4. Non-interference constraints and non-simultaneously constraints will be handled when calculating objective function. C. Assigning QCs and Calculating Objective Function by Applying Priority Rules Based on the handling sequence of tasks represented by the chromosome, the QCs objective function, which is makespan, can be calculated by using the following procedure. Step 1: Determine the ready time and current location of each quay crane, the checking time is the minimum time of all the ready times. Fig. 3. Flowchart of methodology using modified GA combined priority rules. Modified chromosome 1 3 2 5 7 8 4 9 6 Sequence handling 1 2 3 4 5 6 7 8 9 Fig. 4. Modified the sample chromosome in Fig. 3. Step 2: At the checking time, determine which quay cranes can handle the first unassigned task indicated in the chromosome that does not interfere with the other quay cranes and also satisfy the safety constraint between adjacent cranes as well satisfy the non-simultaneous clause. No interference with the other cranes means that the crane in question in trying to reach a target location cannot cross over the other cranes. The safety constraint implies that the minimum physical distance between any two cranes must equal the safety distance value. The non-simultaneous clause requires some tasks designated to be placed in some bays not be executed at the same time. There are three possible cases at each checking time. Case 1: There is one or more cranes available at checking-time and can handle target job: move to step 3: Apply the assignment priority rules. Tournament Selection Stop? Order Crossover Random initialization of parent population Optimum solution Y N Input: ship stowage plan (with all constraints), Time required to carry each task, QCs travel time, QCs ready time Calculate the fitness value: use priority rules to assign tasks for QCs that must satisfy the constraints. If all assignment ways violate the constraints, give penalty Y N Satisfy precedence relationship Swapping positions of tasks Mutation Select new generation International Journal of Materials, Mechanics and Manufacturing, Vol. 7, No. 1, February 2019 23 Fig. 5. Illustration of priority rules for assigning QC’s to task. Case 2: There is one or more cranes can handle target job but they are not available at checking time. They should remain in place and continue to work on their assignment. For cranes which have finished their jobs will remain idle. The checking time is increased to the next minimum cranes ‘ready time. The longer idle time of cranes makes the makespan value worse. Therefore, an optimal makespan will have the minimum quay crane's idle time. Update the current position, the completion time, and the ready time of the quay crane at the new checking time (To track the current locations of quay cranes exactly with the checking time). Repeat step 1 and step 2. Case 3: There is no crane that is able to handle the target job: add a penalty to the fitness value of the solution. Step 3: Priority rules for assigning QC’s to task Two assumptions are made in this paper: 1 - The bays are identified in increasing order from left to right, i.e. bay 1, bay 2, bay 3, etc. 2 - The QC’s are identified in increasing order from left to right, i.e. QC1, QC2, QC3, etc. Case 1: Task i is at a location non-equidistant from either crane, then it will be assigned to the closest crane (rule 1). Case 2: Task i is at a location equidistant from either crane. The allocation of task I follows from the following rules: Rule 2: If the location ID of the next task, task i+1, is higher than or equal to the location ID of the crane with higher ID, i.e. QC2, then assign task i to the crane with the lower ID, i.e. QC1. Rule 3: If the location ID of the next task, task i+1, is lower than or equal to the location ID of task i and that of the crane with higher ID, i.e. QC2, then assign task I to the crane with the higher ID, i.e. QC1. TABLE II: AN ILLUSTRATION OF THE APPLICATION OF THE PRIORITY RULES Bay location 1 2 3 4 5 6 7 8 9 10 11 Initial QC location QC1 QC2 Task-Bay requirements 1&2 3 4 5&6 7 8 9 Feasible sequence of task (chromosome) 1 3 2 5 7 8 4 9 6 Bay requirement 1 2 1 5 7 9 3 11 5 QC1 ready time QC1 location QC2 ready time QC2 location Checking time Initial state 0 1 0 4 0 Iter 1 Rule 1: Task 1 to QC1 22 1 0 4 0 Iter 2 Case 2: only crane 1 can handle task 3 but not available 22 1 22 4 22 Iter 3 Rule 1: Task 3 to QC1 31 2 22 4 22 Iter 4 Case 2: only crane 1 can handle task 2 but not available 31 2 31 4 31 Iter 5 Rule 1: Task 2 to QC1 78 1 31 4 31 Iter 6 Rule 1: Task 5 to QC2 78 1 42 5 42 Iter 7 Rule 1: Task 7 to QC2 78 1 84 7 78 Iter 8 Case 2: only crane 2 can handle task 8 but not available 84 1 84 7 84 Iter 9 Rule 1: Task 8 to QC2 84 1 102 9 84 Iter 10 Rule 1: Task 4 to QC1 156 3 102 9 102 Iter 11 Rule 1: Task 9 to QC2 156 3 116 11 116 Iter 12 Only crane 2 available at checking time and can handle task 6, then it to QC2 156 3 160 5 156 International Journal of Materials, Mechanics and Manufacturing, Vol. 7, No. 1, February 2019 24 Rule 4: If the location ID of the next task, task i+1, is between the location ID of task i and that of the crane with higher ID, i.e. QC2, then assign task i to the crane with the lower ID, i.e. QC1. Rule 5: If the location ID of the next task, task i+1, is between the location ID of task i and that of the crane with lower ID, i.e. QC1, then assign task i to the crane with the higher ID, i.e. QC2. The start time of the target task is made equal to the ready time of the assigned quay crane. Update the current crane position, its completion time, and its ready time. Step 4: Step 1-3 are repeated for the next task in the sequence. The following illustration is to show how to calculate the objective function value for the example shown in Table 1. Table 2 provides an illustration of the application of the priority rules elaborated above. Finally, makespan = 160. IV. TOURNAMENT SELECTION From the population, random choose S competitors, with S being the tournament size, then figure out the winner of the tournament which is the individual with the highest fitness of the S tournament competitors. The winner is then inserted into the mating pool. Tournament winners make the mating pool have higher average fitness than the average population fitness. Then from the mating pool, the parents are selected for mating to create the succeeding generation [13]. V. ORDER CROSSOVER (OX) Then from the mating pool, the parents are randomly selected for mating by crossover to create the succeeding generation. Order crossover proposed by Gen and Cheng (1997) [14] is applied. 1) From the mating pool, two parent chromosomes are given. Two random crossover points are selected from one parent, partitioning it into a left, middle and right substring. 2) Produce the first offspring chromosome by copying the middle substring of the first parent into the corresponding positions. 3) Delete the tasks which exist in the substring of the offspring from the second parent. The remaining tasks in the second parent are copied to the child from left to right according to the sequence in the second parent. 4) The second offspring copies the middle substring of the second parent into the corresponding positions. Delete the tasks which are already in the middle substring from the first parent. The remaining tasks of the first parent are placed into the second child from left to right to the order of the sequence. VI. REVERSE SEQUENCE MUTATION (RSM) Reverse sequence mutation is used to mutate offspring chromosomes. In the reverse sequence mutation operator, we take a sequence S limited by two random positions. The gene order in this sequence will be reversed. Parent 7 9 3 4 5 6 1 2 8 Child 7 9 1 6 5 4 3 2 8 Fig. 6. Illustration of inverse mutation. The number of offspring chromosomes that have mutations in a population is determined by the mutation rate parameter. TABLE III: COMPARISON OF THE PROPOSED METHODOLOGY WITH CURRENT BEST SOLUTION AS REPORTED IN MEISEL AND BIERWIRTH (2011) AND IN NARGES KAVESHGAR (2012) Ex. no. Size (cranes x tasks) Meisel and Bierwirth (2011) Our modified GA combined priority rule Developed GA by Narges Kaveshgar - 2012 Gap between our GA and Meisel 1 2x10 520 520 520 0 2 508 508 508 0 3 513 513 513 0 4 510 510 510 0 5 515 515 515 0 6 513 513 513 0 7 511 511 511 0 8 513 513 513 0 9 512 512 512 0 10 549 549 549 0 1 2x15 514 514 514 0 2 507 507 507 0 3 515 515 515 0 4 513 514 516 0.01 5 507 507 507 0 6 508 511 513 0.03 7 507 507 508 0 8 508 508 513 0 9 507 507 507 0 10 513 513 514 0 1 2x20 508 508 509 0 2 509 509 514 0 3 509 509 509 0 4 509 509 513 0 5 506 507 507 0.01 6 508 508 508 0 7 507 507 507 0 8 510 510 510 0 9 508 508 508 0 10 507 507 511 0 1 2x25 508 510 513 0.02 2 507 507 513 0 3 507 507 507 0 4 507 507 507 0 5 507 507 507 0 6 507 507 507 0 7 508 508 508 0 8 507 507 507 0 9 506 506 507 0 10 506 510 513 0.04 1 4x50 774 786 784 0.12 2 771 790 782 0.19 3 772 798 784 0.26 4 765 808 803 0.43 5 762 786 792 0.24 6 765 798 769 0.33 7 782 800 782 0.18 8 761 810 781 0.49 9 798 810 860 0.12 10 759 802 792 0.43 International Journal of Materials, Mechanics and Manufacturing, Vol. 7, No. 1, February 2019 25 VII. NEW GENERATION SELECTION In this paper, we apply the Steady-State method for the new generation selection: n worst old parents are deleted and are replaced them with n best new offspring. n is a parameter to be experimented with. VIII. EXPERIMENTAL RESULTS AND DISCUSSION Fig. 7 shows the grant chart of the small QCSP instance in Table 1. It specifies the allocation of quay cranes to handle tasks, sequencing these tasks on each QC, and the operation start-time and complete time of tasks. Optimum makespan of this instance is 145 (days). Fig. 7. Grant chart for optimum solution of the small QCSP instance in Table 1. To evaluate the effectiveness of the proposed methodology, the experiments used benchmark data provided by Meisel and Bierwirth (2011) which can be downloaded from [15] and compare its performance with the best-known solutions as reported in Meisel and Bierwirth (2011). The parameters used in the proposed algorithm are: population size = 2x number of crane x number of tasks (individuals), crossover rate = 0.9, the mutation probability = 0.1, and the tournament size = 4. The new generation is created by 50% of best parents combined with 50% of best offsprings. The number of evolution = 10 x population size. The algorithm is programmed by using Matlab Language and performed on an Intel ® Core ™2, CPU 2.13 GHz, RAM 4.00 GB, 64-bit system. Based on the compared results in Table 3, the modified GA combined priority rule is capable to find the optimum or very close optimum solution for all the benchmark instances although it is very simple to apply. The precedence constraints are handled effectively by the method to build chromosome and to validate them. The priority rules help to assign QCs to do jobs efficiently and avoid interface. The computation time of this method is affected by the size of the problem. It is increased when the size of the problem increases. The processing time to solve these instances ranges from 3 to 60 seconds which is reasonable and acceptable. Overall, the modified GA combined priority rule proposed a simple, effective and efficient way to solve the difficult quay crane scheduling problem with crane interference constraints. Comparing with other algorithms or methods which require the users and readers to have a strong background in mathematics and programming, the proposed modified generic algorithm combined with simple priority rules does not require mathematical formulation but still is capable of finding the optimal or near-optimal solution in reasonable computing time. REFERENCES [1] F. Meisel and C. Bierwirth, “A unified approach for the evaluation of quay crane scheduling models and algorithms,” Computers & Operations Research, vol. 38, pp. 683-693, 2011. [2] C. F. Daganzo, “The crane scheduling problem,” Transportation Research Part B, vol. 23, no. 3, pp. 159-175, 1989. [3] R. I. Peterkofsky and C. F. Daganzo, “A branch and bound solution method for the crane scheduling problem,” Transportation Research Part B: Methodological, vol. 24, no. 3, pp. 159-172, 1990. [4] A. Lim, B. Rodrigues, F. Xiao, and Y. Zhu, “Crane scheduling with spatial constraints,” Naval Research Logistics, vol. 51, pp. 386-406, 2004. [5] K. H. Kim and Y. M. Park, “A crane scheduling method for port container terminals,” European Journal of Operational Research, vol. 156, no. 3, pp. 752-768, 2004. [6] L. Moccia, J. F. Cordeau, M. Gaudioso, and G. Laporte, “A branch- and-cut algorithm for the quay crane scheduling problem in a container terminal,” Naval Research Logistics, vol. 53, no. 1, pp. 45- 59, 2006. [7] Sammarra et al., “A tabu search heuristic for the quay crane scheduling problem,” Journal of Scheduling, vol. 10, no. 4-5, pp. 327- 336, 2007. [8] C. Bierwirth and F. Meisel, “A fast heuristic for quay crane scheduling with interference constraint,” Journal of Scheduling, vol. 12, pp. 345-360, 2009. [9] C. Bierwirth and F. Meisel, “A survey of berth allocation and quay crane scheduling problems in container terminals,” European Journal of Operational Research, vol. 202, pp. 615-627, 2010. [10] S. H. Chung and K. L. Choy, “A modified genetic algorithm for quay crane scheduling operations,” Expert System with Applications, vol. 39, pp. 4213-4221, 2012. [11] N. Kaveshgar, N. Huynh, and S. K. Rahimian, “An efficient genetic algorithm for solving the quay crane scheduling problem,” Expert System with Applications, vol. 39, pp. 13108-13117, 2012. [12] K.-H. Choi, T. T. Tran, and D.-S. Kim, “A new approach for intelligent control system design using the modified genetic algorithm,” International Journal Intelligent Systems Technologies and Applications, vol. 9, no. 3-4, 2010. [13] V. Nguyen and H. P. Bao, “An efficient solution to the mixed shop scheduling problem using a modified genetic algorithm,” Journal of Procedia Computer Science, vol. 95, pp. 475-482, 2016. [14] M. Gen and R. Cheng, Genetic Algorithms and Engineering, 1st ed. New Jersey: John Wiley & Sons, 1997. [15] Martin-Luther-Universität Halle-Wittenberg. QCSPgen-A benchmark generator for quay crane scheduling problems. [Online]. Available: http://prodlog.wiwi.uni-halle.de/forschung/research_data/qcspgen/ Vi H. Nguyen is currently a lecturer in Sustainable Product Development Global Production Engineering Management (GPEM), Faculty of Engineering, Vietnamese-German University. She got the full- scholarship from Vietnam Education Foundation for 5-Year Combined Masters-PhD Program in the USA. She obtained her out-standing student award and her Ph.D. degree in design and manufacturing from Mechanical Engineering Department, Old Dominion University in 2017. Her research interests are sustainable manufacturing, engineering design optimization, scheduling and planning optimization. Duc T. Nguyen obtained his B.S. (Northeastern University, 1974), M.S. (U.C. Berkeley, 1976), and Ph.D (University of Iowa, 1982) degrees in civil/structural engineering. He has been with Civil Engineering Faculty at ODU since 1985. His teaching activities (including 3 published textbooks, 1 co-edited book), research works (with 53 journal papers, 81 conference proceeding papers, 53 technical reports, 22 abstracts, and 62 seminars) and 36 funded projects (Totally = over $4,000,000) in large- scale parallel computational mechanics have led to several international/national/regional awards. International Journal of Materials, Mechanics and Manufacturing, Vol. 7, No. 1, February 2019 26 
work_u3dpwvocwvdsfhjouzguzxxfy4	----	Editorial: Third special issue on Knowledge Discovery and Business Intelligence Paulo Cortez1 Manuel Filipe Santos1 1 ALGORITMI Research Centre, Department of Information Systems, University of Minho, 4800-058 Guimarães, Portugal Email: pcortez@dsi.uminho.pt, mfs@dsi.uminho.pt 1 Introduction Expert Systems (ES) were proposed in the mid 1970s [Arnott and Pervan, 2014] with the goal of building computerized systems that mimic human behavior to solve real-world tasks. Such systems were based on Artificial Intelligence (AI) techniques, typically by adopting explicit (human understandable) knowledge, extracted from domain experts (e.g. by using interviews), and that was stored in a knowledge base [Buchanan, 1986]. In the last decades, the world as changed due to advances in information and communication technology (e.g. massive usage of computers and personal mobile devices, Internet and social media, usage of digital cameras and other sensors). In effect, we are now in the age of data, where a large portion of organizational, societal or personal activities is captured digitally [Mojsilovic, 2014]. Following this change, ES have evolved to include data-driven models, either solely or complemented by expert-driven knowledge. Such change is reflected in the ES journal, which currently publishes several articles related with data analysis fields (e.g. analytics and business intelligence, data mining and knowledge discovery, big data, data science). In this special issue, we highlight two of data related terms: Knowledge Discovery (KD) and Business Intelligence (BI). KD is often used as a synonym of data mining and it consists in an AI subfield that uses machine learning algorithms to extract high-level interesting knowledge from raw data [Fayyad et al., 1996]. BI is a popular management term [Arnott and Pervan, 2014] and it represents several technologies (e.g., data warehouses, KD, dashboards) that store and process organizational data in order to support managerial decision-making [Delen et al., 2014]. The ‘Knowledge Discovery and Business Intelligence’ (KDBI) thematic track was proposed for the EPIA conference on AI in 2009 with the goal of promoting the interaction between the KD and BI areas. Since then, the track has been in- cluded in all EPIA biennial conferences. After 2011, the KDBI track has been associated with special ES journal issues, which included extended versions of the best KDBI papers. The first special issue was published in 2013 and it included the best KDBI 2011 track papers [Cortez and Santos, 2013], while the second spe- cial issue appeared in 2015 and it encompassed the best KDBI 2013 track papers [Cortez and Santos, 2015]. 1 This issue, entitled ‘Third special issue on Knowledge Discovery and Business Intelligence’, contains recent KD and BI contributions that can be used in ES to produce a valuable impact in real-world applications. It includes extended versions of papers from the 4th KDBI thematic track, of the 17th EPIA conference on AI (EPIA 2015), held in Coimbra, Portugal. The track received 18 paper submissions and the authors of the best papers were invited to extend their works for this special issue. After two rounds of reviews, which included reviewers from the KDBI track and also ES journal, the best six papers were accepted, corresponding to an overall acceptance rate of 33%. Due to the interest in data-driven models, in the last years there has been sev- eral interesting developments in the KDBI area. Despite this progress, there are still many challenges and opportunities. For instance, most KD algorithms were targeted for single label classification tasks and often these algorithms cannot deal adequately with label ranking, which is useful in several real-world applications (e.g. modeling user preferences). Also, the financial domain still rises many challenges: it is not clear what is the best approach (regression or classification) to forecast trading actions; and easy to interpret tools are needed to better disclose the rela- tionships among price variations from distinct financial products. Moreover, most of the real-world data has a temporal dimension and there is still room for proposing specialized algorithms that use this dimension in the KD or BI process (e.g. time series retrieval, visual representation of temporal changes). Furthermore, the ‘Ex- tract, Transform, and Load’ (ETL) is a vital component of BI systems but it often requires a substantial manual effort in terms of its design and implementation, even when there are several common ETL processes that are repeated through distinct BI projects. All these challenges and opportunities are addressed in the six papers accepted in this special issue, which we will briefly detail in the next section. 2 Contents of the special issue In the first paper ‘Label Ranking Forests’, de Sá et al (2016) propose a novel KD algorithm for Label Ranking (LR), which is a classification variant task. The goal is to learn the implicit function that performs a mapping between a set of inputs, which characterize an item, and a ranking of labels (instead of a single label, as in standard classification). LR is used in several real-world applications, such as mi- croarray analysis, image categorization or modeling user preferences. The proposed Label Ranking Forests (LRF) algorithm uses an ensemble of decision tree methods for LR and it is considered a natural adaptation of the popular random forest algo- rithm for LR. Several experiments were held using 16 datasets from the KEBI Data Repository. Overall, competitive LR results were achieved by the LRF algorithm when compared with LR decision tree methods. In ‘A comparative study of approaches to forecast the correct trading actions’, Báıa and Torgo (2016) perform an extensive comparison of two main approaches to forecast trading actions of financial markets. The first approach uses standard regression models to predict the daily variation on prices and then uses pre-defined decision rules in order to transform the numeric predictions into trading actions. The second approach uses classification models that directly forecast the trading decision 2 (hold, buy, sell). The data analyzed included assets prices of 12 companies, ranging from 7 to 30 years of closing prices daily data. Several machine learning models were tested (e.g. neural networks, support vector machines, random forests). The main conclusion of the paper is that there is no significant difference between the two main approaches: numeric price variation prediction and and direct classification of trading actions. The study contains two additional recommendations: re-sampling strategies (for regression or classification) are not recommended in this financial domain, even if the data is imbalanced; also, the usage of cost-benefit matrices is promising to enhance the classification models. Also approaching the financial context, in ‘Ramex-Forum: a tool for displaying and analysing complex sequential patterns of financial products’, Tiple et al (2016) present an improved version of the Ramex-Forum algorithm, leading to a KD method that is capable of extracting knowledge from multivariate time series in a visual way and that is easy to interpret. The algorithm was applied in two real-world applications: petroleum production prices and risk analysis of European financial institutions. The obtained results attest the algorithm capability to show relevant price variations in financial markets. In ‘Aggressive pruning strategy for time series retrieval using a multi-resolution representation based on vector quantization coupled with discrete wavelet trans- form’, Muhammad Fuad (2016) proposes a novel KD algorithm for time series in- dexing and retrieval. The algorithm adopts a multi-resolution representation of the series based on Haar wavelets and vector quantization. The experiments were con- ducted using 10 time series with distinct sizes and from different repositories. The proposed algorithm obtained competitive results when compared with a two other representation methods (single-resolution and other multi-resolution algorithm). Temporal data was also studied in the paper of Gérk (2016), entitled ‘Visual analytics of educational time-dependent data using interactive dynamic visualiza- tion’. The paper presents a new web-based visualization framework for BI and that is targeted for an interactive exploration by the decision maker. In particular, the framework is based on motion charts and clustering techniques, allowing to reveal changes over time in terms of two-dimensional space animations. The framework was demonstrated using real-world educational data. A total of 16 human partici- pants evaluated the quality of the framework, confirming the proposed animations as an interesting contribution to the analytic process. In the last paper, Oliveira and Belo (2016) approach ETL processes, which play a key role in BI by extracting data from several sources into a data warehousing system. The paper, entitled ‘On the specification of ETL patterns behavior, a domain-specific language approach’, proposes the use of patterns to represent com- mon ETL tasks (e.g. surrogate key pipelining or intensive data loading). Using the business process modeling notation (BPMN), it is shown how a typical and crucial ETL process (data quality enhancement) can be improved in terms of its design and implementation. 3 Acknowledgments We would like to thank the other KDBI 2015 track (of EPIA) co-organizers, Lúıs Cavique, João Gama and Nuno Marques. Also, we thank the authors, who con- tributed with their papers, and the reviewers (from the KDBI 2015 program com- mittee and the ES journal). This work has been supported by COMPETE: POCI- 01-0145-FEDER-007043 and FCT - Fundação para a Ciência e Tecnologia within the Project Scope: UID/CEC/00319/2013. References [Arnott and Pervan, 2014] Arnott, D. and Pervan, G. (2014). A critical analysis of decision support systems research revisited: the rise of design science. Journal of Information Technology, 29(4):269–293. [Báıa and Torgo, 2016] Báıa, L. and Torgo, L. (2016). A comparative study of approaches to forecast the correct trading actions. Expert Systems, pages –. [Buchanan, 1986] Buchanan, B. G. (1986). Expert systems: working systems and the research literature. Expert Systems, 3(1):32–50. [Cortez and Santos, 2013] Cortez, P. and Santos, M. F. (2013). Knowledge Discov- ery and Business Intelligence. Expert Systems, 30(4):283–284. [Cortez and Santos, 2015] Cortez, P. and Santos, M. F. (2015). Recent advances on knowledge discovery and business intelligence. Expert Systems, 32(3):433–434. [Delen et al., 2014] Delen, D., Turban, E. and Sharda, D. R. (2014). Business In- telligence: A Managerial Perspective on Analytics. Pearson, 3rd edition. [de Sá et al, 2016] de Sá, C. R., Soares, C., Knobbe, A. and Cortez, P. (2016). Label Ranking Forests. Expert Systems, pages –. [Fayyad et al., 1996] Fayyad, U., Piatetsky-Shapiro, G., and Smyth, P. (1996). Ad- vances in Knowledge Discovery and Data Mining. MIT Press. [Muhammad Fuad, 2016] Muhammad Fuad, M. M. (2016). Aggressive pruning strategy for time series retrieval using a multi-resolution representation based on vector quantization coupled with discrete wavelet transform. Expert Systems, pages –. [Géryk, 2016] Géryk, J. (2016). Visual analytics of educational time-dependent data using interactive dynamic visualization. Expert Systems, pages –. [Mojsilovic, 2014] Mojsilovic, A. (2014). The age of data and opportuni- ties. IBM Big Data & Analytics Hub, http://www.ibmbigdatahub.com/blog/ age-data-and-opportunities. 4 [Oliveira and Belo, 2016] Oliveira, B. and Belo, O. (2016). On the specification of ETL patterns behavior, a domain-specific language approach. Expert Systems, pages –. [Tiple et al, 2016] Tiple, P., Cavique, L. and Cavalheiro Marques, N. (2016). Ramex-Forum: a tool for displaying and analysing complex sequential patterns of financial products. Expert Systems, pages –. The authors Paulo Cortez Paulo Cortez is Associate Professor with Habilitation at the Department of Informa- tion Systems, University of Minho, Portugal. He is also coordinator of Information Systems and Technologies (IST) research group of ALGORITMI Centre with 48 PhD researchers. His research interests include: Business Intelligence (Decision Support, Data Mining and Forecasting); and Artificial Intelligence (Computational Intelli- gence, Neural Networks, Evolutionary Computation and Applications). Currently, he is associate editor of the Expert Systems journal and participated in 13 R&D projects (principal investigator in 3). He is co-author of more than one hundred indexed (ISI, Scopus) publications in international journals (e.g., Decision Support Systems, Applied Soft Computing) and conferences (e.g., IEEE IJCNN). Web-page: http://www3.dsi.uminho.pt/pcortez Manuel Filipe Santos Manuel Filipe Santos is Associate Professor at the Department of Information Sys- tems, University of Minho, Portugal, and leader of the Intelligent Data Systems group of Centro Algoritmi, with interests in the fields of: Business Intelligence, Data Mining and Learning Classifier Systems. He participated in several R&D projects (principal investigator in 3). He is also co-author of several publications in international journals (e.g., Artificial Intelligence in Medicine) and conferences (e.g., GECCO). 5 
work_u4xfxpixhzd43lprth3rqw7zce	----	Microsoft Word - feldolgozasalatt.doc Computational predictive programs (expert systems) in toxicology Emilio Benfenatia, and Giuseppina Gini, b a Istituto di Ricerche Farmacologiche ‘Mario Negri’, Via Eritrea 62, 20157, Milano, Italy b Dipartimento di Elettronica e Informazione, Politecnico di Milano, 20133, Milano, Italy „...certain reactive 6roups. For instance, Ashby and Paton (1993) hsted many toxic residues responsi- ble ir adverse activity. CompuDrug starkd Eom thi approach and encoded into its ES, caHed HazardExpert the behaviour of selected residues based on a report by...” Abstract The increasing number of pollutants in the environment raises the problem of the toxicological risk evaluation of these chemicals. Several so called expert systems (ES) have been claimed to be able to predict toxicity of certain chemical structures. Different approaches are currently used for these ES, based on explicit rules derived from the knowledge of human experts that compiled lists of toxic moieties — for instance in the case of programs called HazardExpert and DEREK — or relying on statistical approaches, as in the CASE and TOPKAT programs. Here we describe and compare these and other intelligent computer programs because of their utility in obtaining at least a first rough indication of the potential toxic activity of chemicals. Author Keywords: Toxicity; Expert systems; Artificial intelligence; Computer systems; QSAR Abbreviations: AI, artificial intelligence; CASE, Computer Automated Structure Evaluation (a research program); COM-PACT, Computer Optimized Molecular Parametric Analysis of Chemical Toxicity (a research program); DEREK, Deductive Estimation of Risk from Existing Knowledge (a commercial program); ES, expert systems; HazardExpert, a commercial program; MULTICASE, the improved version of CASE; NMR, nuclear magnetic resonance; NTP, National Toxicology Program (US); QSAR, quantitative structure-activity relationship; TOPKAT, Toxicity Prediction by Komputer Assisted Technology (a commercial program) Toxicology Volume 119, Issue 3 , 16 May 1997, Pages 213-225 
work_u6foxozv3bft5ee3mg2kjkj444	----	None 
work_uagodgrks5asbextfki5ymoq4a	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220997691 Params is empty 220997691 exception Params is empty 2021/04/06-03:05:42 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220997691 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:42 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_uayzfd3dtrgszgwrvrlzbzwrme	----	An intelligent questionnaire analysis expert system Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 36 (2009) 2699–2710 Expert Systems with Applications An intelligent questionnaire analysis expert system Yian-Shu Chu a, Shian-Shyong Tseng a,b,*, Yu-Jie Tsai a, Ren-Jei Luo a a Department of Computer Science, National Chiao Tung University, 1001 Ta Hsueh Road, Hsinchu, Taiwan 300, ROC b Department of Information Science and Applications, Asia University, 500 Liufeng Road, Wufeng, Taichung, Taiwan 413, ROC Abstract In the questionnaire analysis, how to find a statistically significant difference between two or more groups in a continuous measure is one of the major problems in researches. However, it is difficult for researchers to solve the issue of finding possible statistically significant difference. There are two causes of this issue. The one is that the process of finding the statistically significant differences is highly depen- dent on researchers’ intuition and experience, and the other is that the original questionnaire data may not be good enough to find the statistically significant differences. In this paper, we build a data warehouse and a forward-chaining rule-base expert system with three kinds of indicators, Increase, StepDown, and Dice, for drilling down the data warehouse to assist researchers in exploring the data to select appropriate statistics meth- ods to find possible significant differences. The prototype of this expert system has been implemented, and the results of experiment about satisfaction survey showed finding the significant difference becomes easier, and users were interested in the idea of this system. � 2008 Elsevier Ltd. All rights reserved. Keywords: Statistically significant difference; Expert system; Data warehousing; On-line analytic processing (OLAP) 1. Introduction Social science research is the use of scientific methods to investigate human behavior and social phenomenon (Black, 1999; Punch, 2005; Schutt, 2004). Since it is impos- sible for researchers to thoroughly observe the huge popu- lation for a behavior or a phenomenon, they usually use questionnaire survey instead of investigating the whole population. Questionnaire survey is usually done by selecting some representative samples from population according to the sampling methods. For analyzing the questionnaire data, researchers can use not only descriptive statistics methods to describe the basic distribution of the samples, but also inferential statistics methods to infer the real human behav- ior and social phenomenon. 0957-4174/$ - see front matter � 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2008.01.076 * Corresponding author. Address: Department of Computer Science, National Chiao Tung University, 1001 Ta Hsueh Road, Hsinchu, Taiwan 300, ROC. E-mail address: sstseng@cis.nctu.edu.tw (S.-S. Tseng). Significance of group differences is one of the four types of the research questions (Tabachnick & Fidell, 1996). For example, in a questionnaire survey for elementary school students, ‘‘Is there a significant difference between boys and girls about the hours they access the Internet every week?” is an interesting phenomenon that researchers want to know. In the questionnaire analysis, how to find a statis- tically significant difference between two or more groups in a continuous measure is one of the major research issues. In the traditional quantitative research, the researchers will firstly propose several hypotheses of the subject according to their experiences, and then determine whether there exists a statistical significance in some hypothesis in a try-and-error manner. The quality of the result using the manner, of course, is in accordance with the quality of the hypotheses, and the process of finding the statistically significant differences is highly dependent on researchers’ intuition and experience. For example, in a questionnaire survey of elementary school students’ Internet usage behavior, a researcher might make a hypothesis, ‘‘There is a statistically significant difference between different gen- ders about the hours they access the Internet every week,” mailto:sstseng@cis.nctu.edu.tw 2700 Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 and then use appropriate inferential statistics method according to his/her knowledge of statistics to test this hypothesis. Without making the hypothesis, the difference can not be found even if it really exists. Therefore, how to acquire the knowledge and experience of senior researchers might be helpful for the junior researchers. Besides, granularity of original questionnaire data may not be good enough to find the statistically significant dif- ferences. For example, in a questionnaire survey of elemen- tary school students’ Internet usage behavior, if there is no significant difference between different resident regions about the hours students access the Internet every week, the researcher might conclude there is no significant differ- ence between different resident regions. However, each region may contain several counties in geography. By dril- ling down the student’s resident dimension to lower levels of granularity, it is still possible to find a significant differ- ence between different counties. When researchers analyze a questionnaire data, they have to make some hypotheses of the possible statistically significant differences from the data and use their knowl- edge like rules to select the appropriate statistics methods to test the chosen hypothesis according to the facts about the characteristics of data, for example data scale of attri- butes. Therefore, we acquire the knowledge and experience of senior researchers to construct a forward-chaining rule- base expert system, which can assist researches in making hypotheses of the possible statistically significant differences from the data and making selections of inferential statistic methods to test researchers’ hypotheses. Since the granular- ity of original data warehouse may not be good enough to help researchers finding the possible statistically significant differences, three indicators, Increase, StepDown, and Dice, are designed and metarules are included to determine the degree of indicators. Hence, in the expert system, a signifi- cant difference viewer is constructed based on the indicators and metarules to assist junior researchers in exploring data. For the need of selecting appropriate inferential statistic methods to test researchers’ hypothesis, the expert system is also designed to give suggestions for appropriate statistic methods to test hypotheses. Since researchers may be inter- ested in understanding why these methods are suggested by the expert system and how to use them, the expert system not only gives explanations about why these methods are appropriate to analyze the data, but also provides a learning platform for researchers who want to learn these methods in more detail. Finally, a satisfaction survey has been done to get user’ satisfaction of using this expert system. The results showed finding the significant difference becomes easier, and users were interested in the idea of this system. 2. Preliminary 2.1. Quantitative research Quantitative research techniques (Black, 1999; Punch, 2005; Schutt, 2004; Zhang, Padmanabhan, & Tuzhilin, 2004) are part of primary research and the quantified data can be collected through structured interviews, experiments, or surveys. Quantitative research is all about quantifying relationships between variables. Variables are things like weight, performance, time, and treatment. Researchers mea- sure variables on a sample of subjects, which can be tissues, cells, animals, or humans, and express the relationship between variables using effect statistics, such as correlations, relative frequencies, or differences between means. ‘‘Hypothesis Testing” is the most popular statistics method to analyze the relationship between variables. And the researchers are most concerned about the differences between means, because it can immediately and effectively indicate the causality of the subject. If some variable like the gender of students versus the hours that students access the internet per week has a difference which is a statistical sig- nificance at a given 1 � a confident level, we said that there is a significant difference between different gender of students. Furthermore, the quantitative research researches aimed to determine the relationship between one thing and another in a population. Quantitative research designs are either descriptive (subjects usually measured once) or experimental (subjects measured before and after a treat- ment). An experiment establishes causality. For an accu- rate estimate of the relationship between variables, a descriptive study usually needs a sample of hundreds or even thousands of subjects; an experiment, especially a crossover, may need only tens of subjects. 2.2. Data warehouse and indicator Data warehouse and OLAP (Berndt, Hevner, & Studn- icki, 2003; Chang, 2006; Chaudhuri & Dayal, 1997; Inmon, 1996) model the data as multidimensional database structure (Agarwal, Agrawal, & Deshpande, 1996; Sarawagi, 2000), and multidimensional database structure views data as data cube (Chen, 2003; Datta & Thomas, 1999; Gray, Chaudhuri, Bosworth, et al., 1997; Palpanas, Koudas, & Mendelzon, 2005). The data structure of data cube is generally repre- sented in a form of star schema, snowflake schema, or fact constellation schema (Han & Kamber, 2001), where the star schema is the most basic one, and the other two can be derived from it. To simplify our discussion, in this paper, a data cube is represented in the form of star schema. The data warehouse supports the analysis tool OLAP, and users can use some OLAP operations, like roll-up, drill-down, dice, etc., to explore the data cubes. However, the exploring process is not automated, and users still need to explore the data cube by her/his intuition and experience. Sarawagi (Sarawagi, Agarwal & Megiddo, 1998) proposed a Discovery-driven Exploration of OLAP Data Cubes approach, which provides the following three kinds of pre- computed indicators to assist users to explore the data cubes. � SelfExp: This indicates the degree of surprise of the cell value, relative to other cells at the same level of aggregation. Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 2701 � InExp: This indicates the degree of surprise somewhere beneath the cell, if we drill down from it. � PathExp: This indicates the degree of surprise for each drill-down path from the cell. However, these indicators are focused on finding the exceptions in data cubes, not statistically significant difference. 2.3. NORM In traditional forward-chaining rule-base expert system, the rule base consists of all rules and facts. The system needs to go through every matching rule when conducting inference for the proper result. This might become ineffi- cient when the number of rules and facts become large. Therefore, many researches aim to improve the mainte- nance of rule-based expert system by incorporating the objected-oriented approach. We apply the knowledge model, NORM, to represent knowledge according to object-oriented concept. It is based on the principles about how people ponder and learn to acquire knowledge and contains knowledge classes and the relations between knowledge classes, including refer- ence, extension-of, trigger, and acquire. For example, in Trigger relationship, it triggers another knowledge class with current facts as knowledge transfer. In other words, the remnant knowledge in original knowledge class should not be considered. During inferring, present inference pro- cess of the fact collection aborts, and a new inference pro- cess will start with the triggered knowledge class. 3. Idea There are two main issues in assisting users in analyzing questionnaire data. When users start to analyze question- naire data, they explore the data first. However the granu- larity of original questionnaire data may not be good enough to explore the data. For instance, there is no signif- icant difference between different values in higher levels of granularity. But it is still possible to find a significant dif- ference in lower levels of granularity. Therefore, the first issue is how to easily explore the full data to make hypoth- eses of the possible statistically significant differences. In addition, after users make hypotheses of the possible statis- tically significant differences, they make a hypothesis and use the inferential statistics method to test this hypothesis. Since there are several methods can be used to test hypoth- eses, the methods user used to test hypotheses may not be appropriate. Therefore, the second issue is how to select appropriate inferential statistics methods to test hypothe- ses according to the data. The expert system providing assistants for these two issues can ease users finding the statistically significant differences from their questionnaire data. For the first issue, we interview the experts of data anal- ysis. According to their experience and knowledge how they find the possible statistically significant differences from the original questionnaire data, they may use data warehouse and OLAP to explore the data. However, these two tools only support general viewer of data. The experts still have to operate these tools to find the possible statisti- cally significant difference according to their experiments and experience. Therefore, three kinds of indicator, Increase, StepDown, and Dice which can be used to help compute the degree of statistically significant difference are designed according to their exploring data experience using data warehouse and OLAP. Some metarules for these indicators are also defined to determine which indicator most likely has more degree of statistically significant dif- ference. Therefore, a significant difference viewer using these indicators is designed to assist users in exploring the data cubes in data warehouse to make hypotheses of possible statistically significant differences. As we know, ontology can be used to describe the con- cepts and the relations between these concepts. For the sec- ond issue, we interview the experts of data analysis to construct ontology to represent the domain experts’ knowl- edge about statistics methods selection. With the ontology, the knowledge components are fur- ther acquired and transformed into NORM-based rule classes of the knowledge base. A basic course scheme is also transformed from a domain ontology using the Ontol- ogy-based Learning Sequences Construction Scheme pro- posed by Chen (Chen, Tseng, Liu, Chang, & Chen, 2005) and adaptive learning sequences are generated from domain ontology, students’ learning portfolios and infer- ence chains using the Ontology-based adaptive learning sequences construction algorithm proposed by Chang (2006). These two algorithms are used to generate the same learning courses as master teachers. 4. The knowledge in intelligent questionnaire analysis expert system 4.1. Statistically significant difference indicator Nowadays, researchers generally firstly collect the required data, and then find and learn the appropriate functions to explore and analyze questionnaire by utilizing database, excel, or SPSS softwares, but the process is very hard. Therefore, we apply data warehousing technology and use OLAP to explore the data on-line from various views. Although OLAP is easy to explore data, the previ- ous indicators are not suitable on finding statistically sig- nificant difference. After interviewing, three kinds of indicator are proposed to assist researchers explore the data cubes in data warehouse to find possible statistically significant differences. � Increase: This shows the degree of statistically significant difference by changing the view of an additional dimension. 2702 Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 � StepDown: This shows the degree of statistically signifi- cant difference by changing the view of the stepping down dimension. � Dice: This shows the degree of statistically significant difference by changing the view of dicing for some value. Furthermore, thresholds of these indicators acquired from the experts are used to determine if the data has the possible statistically significant differences or not. If the degree of the indicator is over the thresholds, it means there exists possible statistically significant differences. In addi- tion, the experts also define some rules to determine which indicator should be the most possible statistically signifi- cant differences if there are two or more indicators over the thresholds. Two examples are given below. IF degree of statistically significant difference in Increase > = 0.8 THEN the Increase indicator signals for having statistically significant difference. IF an Increase indicator signals and a Dice indicator sig- nals THEN the Dice indicator cancels the signal. 4.2. Domain ontology In the system, a domain ontology is utilized to represent domain experts’ knowledge. As shown in Fig. 1, the example ontology describes three parts of concepts about statistics. The upper part describes the basic concepts and relations of statistics, the middle part gives the rules of selection about appropriate statistic methods to analyze significance of group differ- ences, and the bottom part is information which contains learning materials about each statistic method. The mean- ings of the relations are explained as follows: Fig. 1. A simple example � A Kind of: Concept class A is a kind of concept class B means that A is a kind of B. For example, data with one continuous dependent variable is a kind of significance of group differences. � A Part of: Concept class A is a part of concept class B means that A is a part of B. For example, Descriptive Statistics is a part of Statistics. � A Strategy of: Concept class A is a strategy of concept class B means that strategy B is appropriate to analyze data with B. For example, One-way ANOVA and T- Test are appropriate to analyze data with one discrete independent variable. � Contents of: Concept class A is contents of concept class B means that A is contents of B. For example, Contents of One-way ANOVA has learning materials about sta- tistic method, One-way ANOVA. � Pre-requisite constraint: If concept class A is a pre-requi- site to concept class B, then A should be learned before B. For example, Descriptive Statistics have Pre-requisite relation to Probability, so someone had better learn Descriptive Statistics before she/he wants to learn Probability. As shown in Fig. 1, the part of the ontology which is cir- cumscribed describes experts used factorial ANOVA to analyze the data with multiple discrete independent variables. 4.3. Rules transformation The middle part of the ontology shown in Fig. 1 is trans- formed into seven rule classes as shown in Fig. 2 used to infer appropriate methods. In each rule class, it verifies of domain ontology. Fig. 2. Rule classes in the knowledge base of the system. Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 2703 some conditions and triggers to another rule class. The conditions used for inferring appropriate methods are the number of dimensions in data which is corresponding to the dependent variable in the middle part of ontology in Fig. 1, the number of values in each dimension which is corresponding to independent variable in the middle part of ontology in Fig. 1, and the data type of each dimension. Some examples of rules are listed below: Rule 1: IF dimension number = 4 THEN data dimen- sion = Three or above. Rule 2: IF data dimension = Three or above THEN class trigger = Class Three Dimensions or above. Rule 3: IF data scale = Nominal or data scale = Ordinal THEN strategy = T-Test. Rule 1 and Rule 2 are rules in the Statistics Methods rule class, and Rule 3 is a rule in the Two Values rule class. 5. Architecture of intelligent questionnaire analysis expert system 5.1. Data preprocessing Before importing the questionnaire data, some related legacy database about geographical data, population data, etc. are required to be integrated firstly into a data warehouse. In Fig. 3, there are two processes to integrate the imported data. First, the system import user’ data based on the metadata user imported with the questionnaire raw data. The metadata describes the formats of the ques- tionnaire data, for example, data type, scale, and hierarchy of attributes. Second, the system integrates the imported data and the other legacy data. This process is accom- Import Data Data Integration Data Warehouse Questionnaire Raw Data Metadata Legacy Data Questionnaire Data Integrated Data Fig. 3. Data preprocessing module of system. plished through the tool, Microsoft SQL Server 2005. Finally, the integrated data is stored in the data warehouse. 5.2. Knowledge transformation To build up the experts’ knowledge in our expert sys- tem, the ontology needs to be transformed and stored into the knowledge base and content package repository through three processes. As shown in Fig. 4, the Knowl- edge Acquisition process transforms the ontology to the rule classes. There is an example in Sections 4.2 and 4.3. We transform the example ontology in Fig. 1 to the rule classes in Fig. 2. In order to provide learning sequences for each statistics method in the system, we refined the OALSC algorithm (Chang, 2006). The Ontology-based adaptive learning sequences construction algorithm is used to generate adap- tive learning sequences. The inputs of the algorithm are domain ontology, students’ learning portfolios and infer- ence chains. The outputs of the algorithm are adaptive learning sequences. The main idea of OALSC algorithm is to integrate inference chains containing the suggested information of statistics methods with specific ontology consisting of statistics methods nodes and the related learn- ing concept nodes, and the students’ learning portfolios are used to make the learning sequences more adaptive through the OALSC algorithm. The refined OALSC process generates adaptive learn- ing sequences for each statistics method from the ontol- ogy consisting of statistics methods nodes and the related learning concept nodes. These learning sequences are stored in content package repository. As shown in Fig. 5, an adaptive learning sequence for One-way ANOVA can be generated from the upper part of the example ontology in Fig. 1. We also retrieve the contents of each statistics method in the ontology to get the materials of learning courses and the contents are stored in content package repository. In Fig. 6, we retrieve the contents of One-way ANOVA from a rectangle node in the ontology in Fig. 1. 5.3. System architecture The architecture of the expert system shown in Fig. 7 is detailedly introduced as follows. Knowledge Base (NORM) Content Package Repository Rule Classes Adaptive Learning Sequences Courses Ontology of Data Analysis Experts Knowledge Acquisition OALSC Algorithm Construct Fig. 4. Knowledge transformation module of system. Content Package Repository Adaptive Learning Sequences OALSC Algorithm Student Learning Profiles Knowledge Base (NORM) Inference Engine (DRAMA) Facts Inference Results Learning Users Data Warehouse Ontology of Data Analysis Learning Platform (SCORM RTE) Metarules Rule Base Significant Difference Viewer U ser Interface M odule D ata W arehouse A P 1 Fig. 7. System architecture. Statistics Descriptive Statistics Inferential Statistics ANalysis Of Variance One-way ANOVA Fig. 5. The leaning sequence about One-way ANOVA. 2704 Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 In order to assist users in exploring their questionnaire data, some results of descriptive statistics about the data, e.g., mean, standard deviation, can be provided via the data warehouse API, and OLAP is also provided to assist users in viewing the distributions of the data, to obtain the analysis reports. Significant Difference Viewer is constructed to assist users in finding the possible statistically significant differ- ences. It generates three kinds of indicators based on the metarules and uses different degree of color to represent different degree of the possible statistically significant dif- ferences. It also provides functions corresponding to the three kinds of indicators to change view of the data. Because users need suggestions to determine which sta- tistics methods are appropriate to test the hypotheses from experts, an inference engine is constructed to use the knowledge in the knowledge base and to infer the appropri- ate statistics methods like experts. The explanations about these results are also provided according to the rules used in the inference process. One-Way ANOVA -------------------------------------------------------------------------------- A One-Way Analysis of Variance is a way to test the equality of thr Assumptions The populations from which the samples were obtained mu The samples must be independent. The variances of the populations must be equal. Hypotheses The null hypothesis will be that all population means are eq different. In the following, lower case letters apply to the individual s co ... llectively. That is, n is one of many sample sizes, but N is the tota Fig. 6. The learning course a Since users may be interested in understanding how to use them, a learning platform is also constructed to provide learning courses. In order to provide courses adapted to each user, the OALSC algorithm is used to generate the adaptive learning sequences based on the students’ profile, ontology, and the method of inference results. ee or more means at one time by using variances. st be normally or approximately normally distributed. ual, the alternative hypothesis is that at least one mean is amples and capital letters apply to the entire set l sample size. bout One-way ANOVA. Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 2705 6. Implementation We have implemented a prototype of the system using the following tools: � Data warehouse and OLAP Microsoft SQL Server 2005 is used to help us construct data warehouse and OLAP, because it can provide friendly user interface for building data warehouse and OLAP. � Ontology Protégé is used to build the ontology for the system, because it can provide complete functions to build ontol- ogy, and can generate an XML file about ontology. � Knowledge base and inference engine DRAMA is used to construct knowledge base and infer- ence engine in the system. DRAMA is a rule-based, client- server tool/environment for KBS development. It can assist knowledge engineers in building up an expert system. Briefly, DRAMA contains lots of innovative techniques including object-oriented technology, knowledge inheri- tance, etc. It also contains useful tools, like rule verification tool, knowledge acquisition assistant tool and the inference server. � Learning platform As we know, SCORM standard is most popular stan- dard of learning materials. Hence, learning platform is con- structed by SCORM RTE 2004. � User interface Microsoft Visual Studio 2005 ASP.NET with C# is used to construct user interface. The following figures are given to illustrate the opera- tions of the system. Fig. 8. Main page of After log-in the system, it shows the main page to users in Fig. 8. Users can use functions by choosing the menu at the left part of the page. First of all, we choose the Descriptive Statistic node to get descriptive statistic of our data here. We use the drop down list on the top to select the Health improve life style and health education course requirements questionnaire survey to analyze, and then select a basic data, Health Sit- uation, and a question, Stress Management, to analyze. We press the Observe button to get the results shown in Fig. 9. Second, we choose Explore Questionnaire Data node to get statistical tables or diagrams about our data. Similar to Descriptive Statistic function, we select the questionnaire first, and add an attribute, Health Situation, and a question in our data, Stress Management, to the table and the dia- gram respectively to get the analysis results as shown in Fig. 10. Third, as shown in Fig. 11, we choose Significant Differ- ence node to get assistants for analyzing significant differ- ences in our data. It is similar to the two functions above; we choose our questionnaire and the question, Stress Management, to analyze first, and then we press the Observe button to begin analyzing. The system displays the analyzed results using different color. We compare the color of each cell in the table to see where might have sig- nificant differences. We can also click one cell to get further analysis. We press Suggestion for Appropriate Analysis Methods button to get suggestion from the system to ana- lyze these data here. Fig. 12 suggests the appropriate statistics methods. We can get information about why system suggests these meth- ods by pressing the Detail button or the learning materials about the method by pressing the Learn button. Figs. 13 and 14 show the explanation and learning materials about One-way ANOVA, respectively. 7. Experiments Two experiments are designed to evaluate if the system really provides users assistants for questionnaire analysis or not. The first experiment is a case study to verify if the expert system. Fig. 9. Some descriptive statistic of data computed by the expert system. Fig. 10. Using OLAP to explore data. 2706 Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 the analysis results and suggestions from the system are correct and appropriate, and a satisfaction questionnaire survey is designed in the second experiment to get satisfac- tion of users. 7.1. Case study The questionnaire of the case study is the Health improve life style and health education course requirements questionnaire survey, which was proposed by a Sanitary and Health Care Center, to analyze students’ life style for planning sanitary and health activities in the future. It has two parts in the questionnaire, basic data part and life style part. There are 19 attributes in the basic data part, e.g., gender, department, and score, and 24 questions in the life style part. Each question belongs to one of the items designed to be analyzed and the items are self-actualiza- tion, health responsibility, sports, nutrition, interpersonal relationship, and Stress Management. There are 1203 records in the database. In Table 1, the analyzed results by the system shows schoolwork pressure, health situation, and comparison with same ages may have significant differences in the self-actualization item. To analyze the attribute, school- work pressure, the system suggested One-way ANOVA and provided a learning sequence and learning course Fig. 11. Possible significant differences analyzed by the expert system. Fig. 12. Suggestion of appropriate methods from the expert system. Fig. 13. Explanation of the suggestion from the expert system. Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 2707 Fig. 14. Adaptive learning materials provided by the expert system. Table 1 The analysis results of the case study Item Possible attributes have significant difference Self-actualization School pressure Health situation Compare with same ages Health responsibility Resident* Native place* Sports Resident Compare with same ages Health situation Nutrition Resident* Health situation School pressure Native place* Interpersonal relationship Health situation School pressure Compare with same ages 2708 Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 about One-way ANOVA for the junior researchers, Statis- tics -> Descriptive Statistics -> Inferential Statistics -> Analysis Of Variance -> One-way ANOVA as shown in Fig. 5 and (Fig. 6) respectively. We compared these results with the results using T-test and One-way ANOVA. Most attributes really have signifi- cant differences. Some attributes, which are verified by T-test or One-way ANOVA, do not have significant differ- ences in Table 1, and we marked them a star sign after them. After analyzing the data, the reason why the results are dif- ferent is because the numbers of records have significant dif- ference between different groups, and this can be improved by remodeled the indicators in the future. The researchers in the Sanitary and Health Care Center agreed those suggested methods are appropriate to analyze these data. 7.2. Satisfaction survey A questionnaire survey was designed to evaluate the user satisfaction of this expert system. There are 3 parts total 11 questions in this questionnaire: system usage related 5 ques- tions (Q01–Q05), expected functions 4 questions (Q06– Q09), and total evaluation 2 questions (Q10–Q11). Each question was measured by the 5-point Likert scale ranging from 1 (strongly disagree) to 5 (strongly agree). Two groups of people were involved in the experiment. The first group was composed of senior quantitative researchers, including 2 teachers who taught quantitative research class and 3 senior social science researchers. The second group was composed of 17 students who were studying the quantitative research class. The process of the experiment contains two steps. First, they used the expert system by a given data warehouse which will be described later to find all the significant differences in the data warehouse. Second, after using the system, they did the questionnaire. A data warehouse was built for this project using the data in the first experiment, and the questionnaire data and some legacy data were integrated into a data cube. The data warehouse was given for these 28 persons, and hoped them find significant differences from this data ware- house. After using the system, a questionnaire survey was taken to measure the satisfying degree of this system. The experiment results are shown in Table 2. The mean points of each question in the first group are lower than the mean points in the second group. The questions with lower mean values (smaller than 3.000) in the first group are Q01 (easy to use), Q02 (friendly user interface), and Q05 (effi- ciency), and, on the contrary, the questions with higher mean values (greater than 4.000) in the first group are Q03 (speed) Table 2 The results of the questionnaire No. Question First group Second group Mean SD Mean SD t Sig. (p) Q01 This system is easy to use 2.800 0.837 3.647 1.115 �1.563 0.134 Q02 The user interface of this system is friendly 2.400 0.548 3.529 1.281 �2.855* 0.011 Q03 The speed of this system is fast 4.200 0.447 4.059 0.827 0.362 0.721 Q04 This system can help me find significant difference correctly 3.000 0.707 4.177 1.185 �2.753* 0.018 Q05 This system can help me find significant difference efficiently 2.800 0.447 4.412 1.064 �3.257** 0.004 Q06 I hope this system can suggest me the appropriate inferentially statistical methods to test significant difference 3.400 0.548 4.529 0.943 �2.527* 0.020 Q07 I hope this system can test significant difference automatically 4.200 0.837 4.647 0.606 �1.334 0.197 Q08 I hope this system can provide the teaching material about the inferentially statistical methods 3.200 0.837 4.294 1.160 �1.950 0.065 Q09 I hope this system can add different inferentially statistical methods 3.600 0.548 4.118 0.928 �1.556 0.147 Q10 I will use this system again in the future 3.000 1.000 4.059 0.966 �2.139* 0.045 Q11 Generally speaking, I am satisfied with this system 3.400 0.548 4.000 0.791 �1.576 0.131 * p < .05. ** p < .01. Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 2709 and Q07 (hope to automate). On the other hand, the ques- tions with lower mean values (smaller than 3.700) in the sec- ond group are Q01 (easy to use) and Q02 (friendly user interface), and the questions with higher mean values (greater than 4.400) in the second group are Q05 (efficiency), Q06 (hope to suggest), and Q07 (hope to automate). Besides, in the comparison of the two groups, there are statistically significant differences in Q02 (friendly user interface), Q04 (correctness), Q05 (efficiency), Q06 (hope to suggest), and Q10 (will use this system again) between two groups. The mean points of each question in the first group are lower than the mean points in the second group, which means the acceptability of senior quantitative researchers to the system was lower than the acceptability of junior quantitative researchers. Since the system is a prototype, Q01 (easy to use) and Q02 (friendly user interface) got the lower points in the two groups: however, this can be improved in the future. There are significant differences in Q04 (correctness), Q05 (efficiency), and Q10 (will use this system again), and the p value of Q05 is smaller than 0.01, especially. That also means the acceptability of senior researchers to the system was lower than the acceptability of junior researchers. Q06 (hope to suggest) got higher points in the second group, and Q07 (hope to automate) got higher points in both groups, which means they are interested in the idea of this system and they hope more useful functions can be added into the system. Generally speaking, they are satisfied with the system. 8. Conclusion In the questionnaire analysis, how to find a significant difference between two or more groups in one measure is one of the major problems which social science researchers are concerned about. However, finding possible significant differences is difficult for social science researchers. In order to assist junior researchers in finding possible significant differences, in this paper, we build an expert system to help users find the possible significant differences from the data cube. The expert system provides the Significant Difference Viewer to assist users in exploring data. The viewer uses three kinds of indicators designed according to the experts’ experiments and the metarules defined by experts. The expert system also provides suggestion to select appropri- ate statistics methods to test users’ hypotheses. The system not only gives explanations about the suggestions but also provides learning courses for these methods. The prototype of the proposed expert system was also implemented, and the experiment about a case study and the satisfaction of the system was done. The case study shows the analysis results are almost correct and suggestions are appropriate by the system. The satisfaction survey shows the acceptability of senior quantitative researchers to the sys- tem was lower than that of junior quantitative researchers. They were interested in the idea of this system and they hope more useful functions can be added into the system. Gener- ally speaking, they were satisfied with the system. There are some future works about this research. First, the Indicators proposed in this paper can be defined more deeply to increase the precision. Second, in addition to sig- nificant difference, this expert system may consider some different kinds of questionnaire analysis methods, like regression and correlation, to help social science research- ers make questionnaire analysis more easily. Acknowledgments This research was partially supported by National Science Council of Republic of China under the number of NSC95- 2520-S009-007-MY3 and NSC95-2520-S009-008-MY3. References Agarwal, S., Agrawal, R., Deshpande, P. M. et al. (1996). On the computation of multidimensional aggregates. In Proceedings of the 22nd international conference on VLDB (pp. 506–521), Mumbai, India. 2710 Y.-S. Chu et al. / Expert Systems with Applications 36 (2009) 2699–2710 Berndt, D. J., Hevner, A. R., & Studnicki, J. (2003). The Catch data warehouse: Support for community health care decision-making. Decision Support Systems, 35(3), 367–384. Black, T. R. (1999). Doing quantitative research in the social sciences: An integrated approach to research design, measurement and statistics. Sage Publications Ltd.. Chang, C. H. (2006). Building a learning-by-doing remedial tutoring system for DNS management. Master Thesis, National Chiao Tung University, Taiwan. Chaudhuri, S., & Dayal, U. (1997). An overview of data warehousing and OLAP technology. ACM SIGMOD Record, 26(1), 65–74. Chen, C. S. (2003). Design and implementation of a unifying intelligent DNS management system. PhD Dissertation, National Chiao Tung University, Taiwan. Chen, R. Y., Tseng, S. S., Liu, C.L., Chang, C. S. & Chen, C. S. (2005). Learning sequences construction using ontology and rules. In Pro- ceedings of ICCE2005 (13th international conference on computers in education), November 28–December 2, 2005. Datta, A., & Thomas, H. (1999). The cube data model: A conceptual model and algebra for on-line analytical processing in data ware- houses. Decision Support Systems, 27(3), 289–301. Gray, J., Chaudhuri, S., Bosworth, A., et al. (1997). Data cube: A relational aggregation operator generalizing group-by, cross-tab, and sub-totals. Data Mining and Knowledge Discover, 1(1), 29–53. Han, J., & Kamber, M. (2001). Data mining: Concepts and techniques. Morgan Kaufmann. Inmon, W. H. (1996). Building the data warehouse (2nd ed.). John Wiley & Sons Inc.. Palpanas, T., Koudas, N., & Mendelzon, A. (2005). Using datacube aggregates for approximate querying and deviation detection. IEEE Transactions on Knowledge and Data Engineering, 17(11), 1465–1477. Punch, K. F. (2005). Introduction to social research (2nd ed.). Sage Publications Ltd.. Sarawagi, S. (2000). User-adaptive exploration of multidimensional data. In Proceedings of the 26th international conference on VLDB conference (pp. 307–316), Cairo, Egypt. Sarawagi, S., Agrawal, R., & Megiddo, N. (1998). Discovery-driven exploration of OLAP data cubes. Research Report, IBM Almaden Research Center. Schutt, R. K. (2004). Investigating the social world: the process and practice of research (4th ed.). Pine Forge Press. Tabachnick, Barbara G., & Fidell, Linda S. (1996). Using multivariate statistics (3rd ed.). HarperCollins Publishers. Zhang, H., Padmanabhan, B., & Tuzhilin, A. (2004). On the discovery of significant statistical quantitative rules. In Proceedings of the 10th ACM SIGKDD international conference on knowledge discovery and data mining (pp. 374–383). An intelligent questionnaire analysis expert system Introduction Preliminary Quantitative research Data warehouse and indicator NORM Idea The knowledge in intelligent questionnaire analysis expert system Statistically significant difference indicator Domain ontology Rules transformation Architecture of intelligent questionnaire analysis expert system Data preprocessing Knowledge transformation System architecture Implementation Experiments Case study Satisfaction survey Conclusion Acknowledgments References 
work_ubfy5jeaqnd6fdba2wjezdziky	----	A customer satisfaction inventory model for supply chain integration A customer satisfaction inventory model for supply chain integration C.Y. Lam *, W.H. Ip Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong a r t i c l e i n f o Keywords: Supply chain management Simulation a b s t r a c t Customer relationship is increasingly influencing the performance of inventory management in a supply chain, and customer retention or migration directly affect the number of customer orders and demands as well as the inventory level. Therefore, the aim of this paper is to propose a customer satisfaction inven- tory (CSI) model that incorporates customer relationship management into an inventory model, where the probabilistic concepts of Markov chains of uncertainties in customer relationships of retention or migration are adopted. This enables us to determine both a CSI level for replenishing the inventory level to best fit future customer demand and a customer CSI value of the net profit or loss of an organization from a customer over its purchasing life against the inventory cost of an organization. The modeling and development of the proposed customer satisfaction inventory CSI model are discussed in this paper. Sim- ulation and analysis are conducted for the proposed model, and satisfactory results are achieved, such that the proposed model can help to determine a CSI level that can fully meet orders and demand as well as minimizing inventory cost and maximizing CSI value. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction Inventory management and customer relationship manage- ment are two co-related aspects in a supply chain, where customer relationship increasingly influences the performance of inventory management (Silver, 1981; Swink, Narasimhan, & Wang, 2007). The tendency of a customer’s retention or migration can directly affect the number of customer orders and demands as well as the inventory level in a supply chain (Sramek, Mentzer, & Stank, 2008; Zhu & Thonemann, 2004). Information, including marketing strategy or sales demand etc. (Wang, Lau, & Lau, 2008) should also be shared within a supply chain. Among various expenditures in a supply chain, inventory costs generally account for the majority. Therefore, minimizing the inventory cost is a major objective in a supply chain, and it is widely accepted that keeping excessive stock is costly and ineffective (Aggawal, 1974). During the past few decades, various literatures and models have been published and proposed for effective inventory manage- ment (Axsater, 2006; Bartmann & Beckmann, 1992; Lewis, 1997; Lu, Song, & Zhu, 2005; Pfohl, Cullmann, & Stolzle, 1999; Watts, Hahn, & Sohn, 1994; Xiong & Helo, 2006; Yung, Ip, & Wang, 2007), for example the Economic Order Quantity Model, RM- System, Newsvendor Problem, AHM Model, Bisection Method, etc. However, most of these inventory management models only account independently for various demand patterns, quantity dis- count, stockout costs, lead time variations, multi-stage/multi-item situations, etc., but few concepts or models are suggested for incor- porating the customer relationship into the inventory management model. Therefore, in this paper, a customer satisfaction inventory CSI model is proposed for the supply chain. This is an integrated model that incorporates the probabilistic concepts of Markov chains of uncertainties in customer relationships of retention or migration into an inventory model, such that the customer loyalty towards a supply chain becomes a decision variable in the inven- tory management, and customer relationship management and inventory management can then be integrated in the supply chain. Supply chain integration can enhance the overall supply chain per- formance (Fuentea, Rosa, and Cardo, 2008; Lam, Chan, Ip, & Lau, 2008; Lam, Ip, & Lau, 2009). More importantly, customer loyalty and retention play a vital role in a supply chain, as acquisition costs considerably exceed retention costs (Bloomberg, 2001; Dyche, 2002), and loyal customers with a long-standing relationship with the supply chain will make regular repeat purchases, thereby facil- itating a gain in long-term profits (Chan & Ip, 2008; Ip, Chen, Lau, Choy, & Chan, 2008). The Markov chains model is a probabilistic model accounting for the uncertainties in customer relationships. It uses probability and expected value to measure future relationships with an indi- vidual customer. The concepts of Markov chains are capable of modeling customer relationships and supporting business decision making for business growth, features that are in high demand in the disciplines of marketing and customer relationships (Bornnen- berg, 1998; Chen, Ip, & Sheng, 2005; Lu & Jiang, 2004). It has been claimed that the major advantages of the Markov chain model are 0957-4174/$ - see front matter � 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.07.063 * Corresponding author. E-mail addresses: cy.lam@polyu.edu.hk (C.Y. Lam), mfwhip@polyu.edu.hk (W.H. Ip). Expert Systems with Applications 38 (2011) 875–883 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2010.07.063 mailto:cy.lam@polyu.edu.hk mailto:mfwhip@polyu.edu.hk http://dx.doi.org/10.1016/j.eswa.2010.07.063 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa its flexibility and sophistication (Isaacson and Madsen, 1976; Pu- terman, 1994). Markov chains make use of a series of states of a system or process to identify all possible conditions; a Markov chains model of purchases states for a customer relationship (Pfe- ifer and Carraway, 2000) is shown in Fig. 1. It has a wide range of Markov properties, including reducibility, periodicity, recurrence, ergodicity, steady-state analysis, and limited distribution for pre- cise prediction. Markov chains modeling can handle customer retention and customer migration situations, and can apply either to an existing customer or to a prospective customer. Therefore, by incorporating Markov chains into the inventory model, customer relationships of retention and migration can be analyzed and eval- uated in the inventory management of the supply chain. The structure of this paper is as follows: in Section 2, the pro- posed customer satisfaction inventory CSI model is defined and modeled, and the assumptions and algorithms for the modeling are presented. In Section 3, a decision support system is developed to illustrate the effect of the proposed CSI model, datasets are collected and analyzed, and the empirical analysis results are pre- sented afterward. Finally, a conclusion with further developments is given in Section 4. 2. Development of the customer satisfaction inventory CSI model The proposed customer satisfaction inventory CSI model is an integrated model that incorporates the probabilistic concepts of Markov chains of uncertainties in customer relationships of reten- tion or migration into an inventory model, such that the customer loyalty towards a supply chain becomes a decision variable in the inventory management as well as in the determination of a cus- tomer satisfaction inventory level (CSI level) and a customer satis- faction inventory value (CSI value) for the model. Therefore, the proposed CSI model is an integrated model that is a function of CSI level and CSI value. In the CSI model, the CSI level is a utility function that represents a deterministic inventory level based on the probabilistic customer relationship of retention or migration, such that the replenishment inventory level can be better tailored to future customer demand without keeping excessive inventory. The CSI value is another util- ity function that represents a value of the net profit or loss of an organization from a customer over its purchasing life against the inventory cost of an organization, such that this value is also gov- erned by the probabilistic customer relationship of retention or migration as well as the inventory cost of setup cost, backorder cost, and inventory holding cost. For the development of the CSI model, customer satisfaction di- rectly affects the customer relationship of retention or migration, therefore, a set of Markov chains stating state = {s1, s2, . . ., st} is de- fined for the state of customer retention/migration, with transition probability p at time t = 1, . . ., t, such that if chain state si is cur- rently at t = i, then it will have a transition probability P = pij that the chain state moves to state sj at t = j, where the transition prob- ability is a customer retention probability (pij) or customer migra- tion probability (1 � pij). The customer retention probability determines the degree of likelihood that a customer will repur- chase in a supply chain, while the customer migration probability determines the degree of likelihood that a customer will terminate the order in the next time state. A schematic diagram for the prob- abilistic concepts of Markov chains in inventory management is illustrated in Fig. 2. Considering the diagram in Fig. 2, if a customer at state s1 pur- chases from a supply chain, then the customer retention probabil- ity for repurchase in the next time stage s2 is p12, while the customer migration probability for purchase termination is p2: P ¼ p1 1 � p1 0 0 0 0 0 0 p2 0 1 � p2 0 0 0 0 0 .. . .. . .. . .. . .. . .. . .. . .. . pi 0 0 0 1 � pi 0 0 0 pj 0 0 0 0 1 � pj 0 0 .. . .. . .. . .. . .. . .. . .. . .. . pt�1 0 0 0 0 0 0 1 � pt�1 0 0 0 0 0 0 0 1 2 66666666666666664 3 77777777777777775 : The above transition probability matrix summarizes the transition probabilities shown in Fig. 2. Matrix P is the one-step transition matrix. To evaluate customer purchase probability over t period, a t-step transition matrix is used. The t-step transition matrix is sim- ply the matrix multiplication product of t one-step transition matri- ces. With the transition probability pij of the customer retention in a set of Markov chain states, the CSI level in the proposed model with respect to its maximum inventory level MAX and minimum inven- tory level MIN at time t = j is defined as CSI levelj ¼ MAX � CSI leveli þ Sj � Dj � �� � � pij for CSI leveli þ Sj � Dj � � < MIN; ð1Þ where S is the inventory supply while D is the customer demand, and pij is the transition probability, which is further defined as pij ¼ 1; PðretentionÞ P PðmigrationÞ; pij; otherwise ( ð2Þ Therefore, with the CSI level, the setup cost, backorder cost, and inventory holding cost can be determined. The setup cost is the cost for setting up an order to fulfill a demand or to replenish the inven- tory, and it is directly related to the order frequency, so the setup cost A at time t = j with respect to its previous reference setup cost, and the expected setup rate is defined as Aj ¼ Ai � D1 þ D2 þ���þ Dj � � CSI levelj : ð3Þ Backorder cost is the cost for each demand unfilled and for its back- ordered time, i.e. the supply from the inventory level cannot fulfill 1 2 3 4 5 1-P1 1-P2 1-P3 1-P4 1.0P1 P2 P3 P4 Fig. 1. Markov chains model of purchasing states for customer relationship. 876 C.Y. Lam, W.H. Ip / Expert Systems with Applications 38 (2011) 875–883 https://isiarticles.com/article/2514 
work_ubze5xpme5eohenrp77llb54aq	----	A multilingual ontology framework for R&D project management systems A multilingual ontology framework for R&D project management systems Ou Liu a,*, Jian Ma b a School of Accounting and Finance, The Hong Kong Polytechnic University, Hong Kong b Department of Information Systems, City University of Hong Kong, Hong Kong a r t i c l e i n f o Keywords: R&D project management Multilingual ontology User preference a b s t r a c t R&D project management systems are developed to facilitate the efficient management of R&D projects in government-funding agencies and universities. They should support R&D information sharing among researchers and research administrators from different nationalities using different languages. This paper presents an ontology-based methodology to the development of R&D project management systems with multilingual supports. In particular, a multilingual ontology is proposed, which consists of domain model, application model, user model and linguistic model. The methodology is used in the development of an R&D project management system for the Innovation and Technology Commission of the Hong Kong SAR Government. The system supports three languages and facilitates R&D information sharing among users with different culture backgrounds and usage preferences. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction Research and development (R&D) project management is an important task for government-funding agencies and research institutions. It usually involves several phases, i.e., proposal sub- mission, project selection, approved project administration and project deliverable dissemination. To support the critical decision making tasks in R&D project management, project related informa- tion needs to be shared among different parties with different knowledge backgrounds at organizational levels, e.g. project applicants and reviewers with different research interests or speaking different languages. Especially, for some countries/re- gions that use multiple official languages (such as Hong Kong and Switzerland), project related information and documentation must be stored, processed, and generated in different languages. Thus there is a need to provide R&D project management systems with multilingual supports and organizational knowledge sharing capabilities. Current research on R&D project management shows increasing interests in developing decision models and methods to support components of the project management process (such as project selection) (Klapka & Pinos, 2002; Lee, Park, & Shin, 2009; Sun & Ma, 2005), or the whole process (Chen, Hsu, & Chang, 2008; Józe- fowska, Bła _zewicz, & Słowiński, 2009; Valls & Weglarz, 2005). But many methods are not being utilized or have limited impacts on real world R&D project management (Sun, Ma, Fan, & Wang, 2008; Tian, Ma, Liang, Kwok, & Liu, 2005). To facilitate the use of decision models in real world applications, decision support systems are developed (Klapka & Pinos, 2002; Lin & Hsieh, 2004; Turban, Aronson, Liang, & Sharda, 2007). However, these systems can hardly support multilingual information processing and knowledge sharing among parties with different backgrounds at the organizational level. Information systems with multilingual support have been stud- ied in several application domains, such as in medicine (Boyer, Baujard, Griesser, & Scherrer, 2001; Dejean, Gaussier, Renders, & Sadat, 2005; Lu et al., in press) and in astronautics (Bensaid & Astor, 2002; Walker, Bayer, & Monjar Bayer, 2002). From these studies, it can be seen that a domain-specific multilingual termi- nology (terminology is regarded as a collection of terms in a spe- cific domain) is very important in supporting the multilingual information processing and sharing. However, no multilingual ter- minology is currently available for R&D project management. Fur- thermore, ‘‘terminology” is only a ‘‘vocabulary of a domain/field/ discipline/specialized subject” (Temmerman & Knops, 2004), with more emphasis on language-specific ‘‘terms” instead of the ‘‘con- cepts” behind it. With the current evolution in the artificial intelligence research, ontology engineering is becoming a popular technology to enhance terminology. An ontology is similar to a dictionary or glossary, but with a greater detail and structure that enables computers to pro- cess the ontology’s content (Fensel, 2004; Plisson, Ljubic, Mozetic, & Lavrac, 2007). It consists of a set of concepts, axioms, and relation- ships that describe a domain of interests, and represents an agreed- upon conceptualization of the domain’s ‘‘real world” setting. Thus, ontology can facilitate knowledge sharing within a specific domain (such as R&D project management) at the organizational level. In this paper, a multilingual ontology framework is proposed for the development of R&D project management systems, where 0957-4174/$ - see front matter � 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2009.12.046 * Corresponding author. Tel.: +852 27662071; fax: +852 23569550. E-mail addresses: afliuou@inet.polyu.edu.hk (O. Liu), isjian@cityu.edu.hk (J. Ma). Expert Systems with Applications 37 (2010) 4626–4631 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2009.12.046 mailto:afliuou@inet.polyu.edu.hk mailto:isjian@cityu.edu.hk http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa domain model, application model, user model and linguistic model are designed to handle related knowledge in R&D project manage- ment. The framework and corresponding methodology is applied in the development of a R&D project management system in the Innovation and Technology Commission (ITC) of the Hong Kong SAR Government. The system supports three languages (i.e. Eng- lish, Traditional Chinese and Simplified Chinese as they are the offi- cial languages in Hong Kong), as well as the information sharing among different user groups (e.g. project applicants, project reviewers and administrators) from all over the world. Section 2 of the paper presents the multilingual ontology frame- work and corresponding methodology. The application to ITC is re- ported in Section 3, including a project management workflow and the system architecture. Finally, Section 4 summarizes and dis- cusses the results of our work. 2. An multilingual ontology framework for R&D project management An ontology is a knowledge repository in which concepts and terms are defined as well as relationships between these concepts (Fensel, 2004). Implicit knowledge for humans is made explicit for computers by ontology. A multilingual ontology framework is designed in this paper to support the representation of multilingual information in R&D pro- ject management. The knowledge in the multilingual ontology in- cludes language-independent knowledge and language-dependent knowledge. Inside, the language-dependent knowledge refers to the knowledge that relies on a specific language, forming a set of linguistic models. And the language-independent knowledge refers to the knowledge that does not rely on any specific language. It is categorized into a domain model, an application model and a set of user models (refer to Fig. 1): (1) Domain model. It contains domain knowledge, which is the general structural knowledge about the application domain. In our case, it is the knowledge about R&D project management. (2) Application model. It contains knowledge about the software application, including functions, menu items, etc. (3) User model. A set of user models are used to capture knowl- edge about users, including their profiles, behaviors and preferences. They can be used to generate personalized interface. (4) Linguistic model. For each language supported by the multi- lingual ontology, there is a correspondent linguistic model. It defines the linguistic terminology of the concepts in the domain model and application model for a specific language. Details of the models are given next. (1) Domain model can be regarded as a domain ontology. The knowledge inside is presented as concepts in the domain and the relations among them. Domain concepts are basic building blocks of the domain model. Each concept is charac- terized by a set of attributes. Concepts are organized accord- ing to a special relation, i.e., concept–subconcept relation. Thus they form a tree-like concept hierarchy. According to previous work (Tian et al., 2005), certain key concepts can be identified for R&D project selection, and a domain model is as shown in Fig. 2. At the panel on the left, there is a con- cept-tree with a node root as the starting point of all con- cepts. Each of the other nodes is the subconcept of its parent (upper-level node). For example, University is a sub- concept of Organization. Subconcepts represent the relation- ship of inheritance. Besides concept–subconcept relation, there are other normal relations. For example, in the central dialog box in Fig. 2, the relations of the concept ‘‘investiga- tor” (project investigator) are shown on the right when the concept is being edited. Attributes of a concept are also regarded as special relations but related to a concrete range rather than another concept. In summary, domain model conceptualizes the application domain and organizes the domain knowledge explicitly in a computer-readable man- ner; it forms a basis for information sharing among people with different knowledge backgrounds at the organizational level. (2) Application model contains knowledge about the system functions organized by function–subfunction hierarchy. The knowledge is also related to the user interface elements (such as menu items) correspondent to specific functions. (3) User models are utilized to capture knowledge about users, including their profiles, behaviors and preferences. A user model is a knowledge resource which contains explicit assumptions regarding all aspects of the user in interaction with the system; and the knowledge is useful when generat- ing user-tailored interface or receiving information from the user. For example, the language of the system interface is chosen according to the user’s language preference. Also, users with different roles or positions in the R&D project management lifecycle can use different authorized system functions with different system menus; other human-com- puter interface elements (like labels, messages and online helps) are also affected. Even more, user models can contain other information about users, such as usage preferences and behavior patterns. (4) Linguistic models define the multilingual terminology of the concepts in the domain model and the application model. For each supported language, the linguistic model defines the mapping from the concepts to the terms in that lan- guage. Suppose that in a given context (such as R&D project management) there is one term expressed in a specific lan- guage corresponding to a concept, we can then define this mapping as: concept ðiÞ!m term ðiÞ; for each language m: A simplified data schema of linguistic models used in the Hong Kong ITC case is shown in Fig. 3. Note that linguistic models for the Language-independent knowledge User model 1 Domain model: domain concepts, relations, … Language-dependent knowledge Linguistic model 1: concept(i)-> term(1,i) User model n Linguistic model m: concept(j)-> term(m,j) Application model Fig. 1. High level structure of the multilingual ontology. O. Liu, J. Ma / Expert Systems with Applications 37 (2010) 4626–4631 4627 https://isiarticles.com/article/3289 
work_uced2zdroneerhjtivs57rihci	----	A hybrid model for learning from failures Expert Systems With Applications 93 (2018) 212–222 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa A hybrid model for learning from failures Calvin Stephen a , Ashraf Labib b , ∗ a University of Manchester, School of Mechanical Aerospace and Civil Engineering, M13 9PL, UK b University of Portsmouth, Operations and Systems Management Group, Portsmouth PO1 3DE, UK a r t i c l e i n f o Article history: Received 4 August 2017 Revised 22 September 2017 Accepted 12 October 2017 Available online 14 October 2017 Keywords: Decision analysis Multiple criteria Decision making a b s t r a c t In this paper we propose the usage of a hybrid of techniques as complementary tools in decision analysis for learning from failures and the reason behind systems failure. We demonstrate the applicability of these tools through an aviation case study, where an accident investigation report was obtained from the Directorate of Accident Investigation in the Ministry of Transport and Communications in Botswana to provide as a basis for the application of the model. The report included all the factual information required to carry out the investigation using the hybrid of FTA, RBD, AHP, HoQ and the DMG tools. We discuss the steps followed in applying the tools in the process of learning from failure. It also shows the importance of such tools in accident investigations by showing the importance of prioritising the available options in order of their importance to the accident under investigation. Most of the available research in learning from failure focuses mostly on the direct causal factors of the failure event. Here we provide a holistic approach to learning from failure by focusing on both direct and indirect causes of a failure event through the use of Reliability Engineering tools, Multi Criteria Decision Making tools and House of Quality. © 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license. ( http://creativecommons.org/licenses/by/4.0/ ) b c i t e a e r a n w a s a r o 1. Introduction In many organisations failure is always the cause of conflicts as they have inherited a blame and lack of trust culture ( Cox, Jones, & Collinson, 2006 ; and Jefcott, Pidgeon, Weyman, & Walls, 2006 ). Even though this is the case, some organisations view failure as an opportunity to obtain lessons for continual improvement hence a chance of gaining competitive advantage over their nearest rivals. Failure can be defined in many different ways of which the use is influenced by the context it is used on. Torell and Ave- lar (2010) described failure in two distinct ways as the inability of a product or system to perform its required function and also as the inability of a component to perform its required function without hindering the function of the product as a whole. The ability to learn from failures helps organizations, engineers and designers to put in place measures to avert the same inade- quacies from re-occurring. Labib (2015) explains that for clear un- derstanding of the causes of a failure, there is a major need to analyse four factors, which are; human, design, organizational and socio-cultural factors. ∗ Corresponding author. E-mail addresses: calvin.stephen@postgrad.manchester.ac.uk (C. Stephen), ashraf.labib@port.ac.uk (A. Labib). i s l c i https://doi.org/10.1016/j.eswa.2017.10.031 0957-4174/© 2017 The Authors. Published by Elsevier Ltd. This is an open access article u By doing so Labib and Read (2015) suggested that four main enefits could be obtained that include easy identification of root auses of the failure and the associated reasons. The other benefit s that such analysis of failure can help to institute long term plans o prevent similar events from re-occurring and can also act as an arly warning signal just prior to the event in order for defensive ctions to be taken. They also suggested that it helps decision mak- rs with information on priorities for resource allocation for both ecovery and prevention. Labib and Read (2015) proposed categorising of causal factors s either direct cause or contributing factors when dealing with atural disasters. This approach can also be useful when dealing ith failures associated with multi-disciplinary environments such s in aviation where there is an interaction of many specialties uch as operations, maintenance, air traffic control, meteorology, irport services, fire fighting etc. When dealing with failure engineers tend to tackle only the di- ect causes of a failure event hence putting less or almost no effort n averting indirect causes of a failure incident. As a result these ndirect causes remain unsolved hence continuing hidden in the ystem, with a chance of causing further failure in the future. It is the purpose of this paper to present a hybrid model for earning from failures where both the direct causes and indirect auses of failure are investigated. This model utilises the reliabil- ty engineering tools of Fault Tree Analysis (FTA), Reliability Block nder the CC BY license. ( http://creativecommons.org/licenses/by/4.0/ ) https://doi.org/10.1016/j.eswa.2017.10.031 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2017.10.031&domain=pdf http://creativecommons.org/licenses/by/4.0/ mailto:calvin.stephen@postgrad.manchester.ac.uk mailto:ashraf.labib@port.ac.uk https://doi.org/10.1016/j.eswa.2017.10.031 http://creativecommons.org/licenses/by/4.0/ C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 213 D ( H H o d f e t f t n 2 i f r p s S i l o ‘ e a I S i o t o s i b c n c e l t n o m m l t f a l v t b i a w f w c o o u o b l T h s o v a w a s t t f t ( d m T o u e h a e d t h m c m o A m p i c r c t u c t p t f p e s d b w m r m iagrams (RBD) and Fault Modes, Effects and Criticality Analysis FMECA); Multi-criteria Decision Analysis techniques of Analytic ierarchical Process (AHP) and Decision Making Grid (DMG); and ouse of Quality (HoQ). To explain the usefulness and application f the model a case study is used. The next section provides a detailed literature review on how ifferent researchers use the above-mentioned technique to learn rom failure. This is followed by a brief summary of the failure vent that will be used as the case study for the application of he proposed model with the subsequent section focusing on the ramework itself and its application on the case study. Finally sec- ion five gives the conclusion of the report underlining the weak- ess and the strength of the proposed approach. . Theoretical frameworks There is a number of research work carried out by scholars and ndustry experts in order to come up with models of learning from ailures. These literature works act as a starting point for further esearch in this important area and also as a guide for the model roposed in this paper. Classification of hybrid models and modelling of operational re- earch tools can be traced back to the work of Shanthikumar and argent (1983) , who suggested that hybrid approaches can man- fest itself in two ways; either through the models and their so- ution procedures, or through the use of the solution procedure f independent types of models. The former option they called it hybrid model’, whereas the latter they termed it as ‘hybrid mod- lling’. In our approach we will focus on the former option where n output of one type of modelling can be an input to the other. n terms of types and usage of operational research (OR) models, hanthikumar and Sargent (1983) suggested that modelling is used n five ways (i) in analysis, where modelling is used to obtain an utput for a given system and input, (ii) in optimization, where he model and its solution procedure are used to find the values f the decision variables to optimize an objective function, (iii) in ynthesis, where a model is developed to convert a set of inputs nto a set of desired outputs, (iv) in gaining insight into a system’s ehaviour by developing a model of it and using its solution pro- edure to explore its behaviour, and (v) in the comparison of alter- ative systems, where modelling of various alternative systems are arried out to determine the "best" one. In our work we are inter- sted here in two types of synthesis, and gaining insight through earning lessons from failures. Morgan, Belton, and Howick (2016) and Morgan, Howick, & Bel- on, 2017 developed a good review about use of hybrid OR tech- iques, where they concluded that mixing OR modelling meth- ds raises many philosophical issues and that there are argu- ents that suggest benefits and potential problems of mixing OR ethods in general. However, they argue that real-world prob- em situations are highly complex and multidimensional, and po- entially may benefit from different paradigms to focus on dif- erent aspects of a situation. Howick, Ackermann, Walls, Quigley, nd Houghton (2017) used a case study to illustrate how one can earn from mixing OR methods and specifically they focused on the alue or impact of such integration of methods. However, most of he survey literature about case studies of mixing methods tend to e applied to a hybrid f two or maximum three methods, whereas n our case we develop a framework that utilises multiple methods nd we highlight the benefit of using each one. Love, Lopez, and Edwards (2013 ) developed a learning frame- ork that can be used to mitigate design errors and potential ailures and accidents in the construction industry. Their frame- ork acknowledges the fundamental pathogenic influences that ontribute to errors and failures. As such it suggests that a group f approaches should be implemented simultaneously at a project, rganisational and people level in order to lessen errors and fail- res. Failure to do this, according to Love et al (2013 ) would depend n time until the next major failure is experienced. They continue y explaining that reviewing past experiences is the first step in earning from failures but the much bigger step is taking action. his is because taking action involves a major change in both be- aviour and culture. When analysing the Fukushima accident, Zubair, Park, Heo, Has- an, and Aamir (2015) noticed that there exist basic precursors f nuclear accidents that are inherently difficult to quantify with ague priorities. So, to overcome these shortfalls they proposed model, which combined the AHP and the Bayesian Belief Net- ork (BBN). These helped them to accomplish sensitivity analysis nd prior probabilities into posterior probabilities of precursors. As uch they found out that design is the most important precursor hough the chance of an accident is also dependent on other fac- ors such as culture and plant specific conditions, which can af- ect the distribution of prior probability. For a review of AHP in erms of its methodological variation, please see Ishizaka and Labib 2011a, b) . In their research, Ishizaka and Labib (2014) studied the Bhopal isaster and proposed a model for learning from failure. In their odel they demonstrated that the FTA can be improved in Crisis ree Analysis (CTA) in order to map a crisis with the introduction f the revolving gate as opposed to the AND and OR gate that are sed in an FTA. The CTA caters for amplified impact of the input vent to the final event. They also suggested that the RBD could also map crisis with yper-blocks as the complement of the revolving gate. Their model lso utilises the AHP method to measure the criticality of the basic vents. Through the use of their model more realistic and sound ecisions can be made unlike when using each technique in isola- ion. In a bid to show that the use of FTA and RBD can systematically elp in solving complex industrial failures, Yunusa-Kaltungo, Ker- ani, and Labib (2017) applied these techniques to investigate a hronic rotary kiln refractory brick failure in a fully integrated ce- ent plant. They compared the efficiency of these methods to the ne that was being used in the plant that is based on Root Cause nalysis (RCA). The results obtained indicated that the investigative ethod that was used in the plant that is based on RCA failed to revent future occurrences. Through the usage of FTA and RBD the nvestigative team obtained a holistic understanding of the failure ausing factors and their interrelations hence helping in avoiding epetition in the future. Both FTA and RBD have been used in a omplimentary manner ( Bhattacharjya & Deleris, 2012 ). Labib (2015) emphasized the importance of the FTA and RBD echniques in creating a framework for learning from failures. He sed these techniques to analyse the Bhopal disaster and he con- luded that they could be used to serve as both knowledge reten- ion and decision support tools. According to Labib (2015) they can rovide practitioners with guidelines to follow the root cause of he problem, equips them with the tool box leading to more ef- ective decision making practices, process safety and environment rotection. Morgan et al. (2016) presented insights on using hybrid mod- ls by mixing OR methods of system dynamics and discrete-event imulation within a real world project. They presented the model evelopment process, the role of each modelling method and the enefits of using such hybrid models in project design. In their ork, they have shown that by using hybrid models in comple- entary, each model add value to the other resulting in an all- ound solution to the problem. On the other hand, Labib and Read (2015) proposed a hybrid odel for learning from failure that utilises both the reliability en- 214 C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 t e i p d e p t i o e p p t o l a r t a t t t f c e i p t i w c a a w A t h n C a w t t w a e w 4 b t e p o p L gineering tools and the multi criteria decision analysis tools. They used the reliability tools of FMECA, FTA and RBD in their model. They used this model to study the Hurricane Katrina disaster. The FTA is the starting point of their model creating inputs for FMECA and RBD. The output for the FMECA, FTA and RBD act as inputs to the MCDA method of AHP which produces the outputs helping the user to make either selection decisions or resource allocation decisions. All the models proposed in the above studied literature have been applied on major disasters as such one can wonder if they can be of ultimate importance in minor failures. They also concen- trate more on the direct cause of the failure with less emphasis on indirect factors. These outlines the importance of the proposed model as it will focus on both direct and indirect causes of a fail- ure with the use of a case study in which no lives were lost. It also tries to appreciate the benefits of using hybrid models by us- ing techniques in complementary. 3. ZS-CME serious incident ZS-CME is a bombardier CRJ-100 series aircraft that is registered by the South African civil Aviation Authority under the ownership of CemAir. This aircraft suffered main landing gear wheel disinte- gration upon landing during its scheduled flight from Cape Town to Gaborone. This incident was investigated as a means to derive lessons and gather facts as to what happened, how it happened, when it hap- pened, where and why it happened by the directorate of accident investigation in the ministry of transport and communications in Botswana. The purpose of this investigation was to obtain facts in order to prevent similar incidents from occurring in the future. According to ICAO annex 13, aircraft wheel disintegration are not classified as accidents but circumstances surrounding the ZS- CME incident made it to be have classified as a serious incident rather than just an incident ( Moakofi, 2016 ). The below sections give a clear synopsis of what really happened. 3.1. Synopsis of the incident On August 31, 2015 ZS-CME operated by Air Botswana under a lease agreement on a scheduled service as BOT 332 experienced starboard outer wheel disintegration upon landing at Sir Seretse Khama International Airport (SSKIA) in Gaborone. The aircraft de- parted Cape Town International Airport (CTIA) earlier that day where it was reported to have experienced excessive vibration dur- ing the take-off roll but the flight crew misjudged as minor and continued with the flight. Two seconds after touch down it was reported that the crew, air traffic control (ATC) and even the fire fighters heard a loud bang sound. Even though the crew had no idea what was the problem they taxied the aircraft for almost 1.3 km in order to clear the run- way for the service that was landing behind them while increasing the inherent risk to passengers. All passengers and crew disem- barked safely as it was noticed the aircraft was tilting to the right as the wheel has disintegrated destroying the right main landing gear and ripping off the inner flap hence creating a fire hazard as the wings contained fuel. 3.2. Investigation findings The worrying issues were the prolonged period the fire and res- cue services (FRS) took to heed help to the occupants in the air- craft and the unavailability of aircraft engineers to attend the in- cident. These prompted the investigators to dig deep to see what could have caused this delays and the unavailability of engineers. As a result, they found out that the Public Address (PA) sys- em that would have made it possible for the ATC to communicate ffectively with FRS was unserviceable which meant the relay of nformation from the ATC to the FRS was ineffective as it has to ass through a third person before the information can reach its estination. The other startling discovery was that since the aircraft was op- rated by Air Botswana under a lease agreement, CemAir did not rovide the ground support engineers at the airport even though he lease agreement stated that, “the lessor shall supply duly qual- fied ground engineers/technicians who shall be available in the perations area to carry out daily line maintenance and minor ngine and airframe inspections on the aircraft as required, as er included costs” ( Moakofi, 2016 ). It was evident the lessee had aid for such services because they had no engineers appropriately rained on the type of the aircraft because they had no such type n their fleet. It is reported that the maintenance crew arrived ater on from Johannesburg. Since the incident resulted in debris all over the runway it was lso noticed that the airport had no serviceable runway sweeper egardless of the threat foreign object debris on the runway pose o safe air travel. The measures put in place to act as an alternative re not only time consuming but also ineffective as compared to he use of a runway sweeper. See Davidson and Labib (2003) on he impact of debris of rubber from the wheel on the runway of he Concorde accident. The direct causes of the incident upon investigation where ound to be originated from a major maintenance work that was arried out some two and half months prior to the incident. It was vident that during maintenance work there occurred an incorrect nstallation of brake lines to the inboard/outboard swivel assembly orts. This would result in a faulty operation of the anti-skid sys- em producing a pro-skid condition which when activated would ncrease load on the landing gear instead of reducing it. This cross iring of brake lines can be attributed to design errors in the air- raft landing gear system which made it possible. Investigators found out that the aircraft manufacturer became ware of the design error and offered a service bulletin (SB), which ccording to Moakofi (2016) it was not evident whether the SB as affected, as the aircraft records from their previous owner, an merican company, were inadequate to tell. The effective date of he SB was 26 December 2014 and operators where required to ave complied within 6600 flying hours from the effective date but ot later than March 2017. Moakofi (2016) also found out that during the take-off roll in TIA the aircraft experienced severe vibration that could have been n indication of fault initiation. The crew then decided to continue ith the flight as they considered the vibration to be minor. Even hough Moakofi (2016) did not mention anything about the vibra- ion limits of the aircraft it can be argued that the pilot decision as informed by what the indicators told them or maybe they cted in negligence. The indicators might have given them inad- quate information that could have left the crew indecisive about hat measure to take. . Proposed model The proposed model is an extension of the model proposed y Labib and Read (2015) that encompasses reliability engineering ools of FTA, FMECA and RBD and the MCDM tool of AHP. As an xtension to this model the new model will make use of a sim- lified HoQ matrix and a modified MCDM technique of DMG. The verview of the proposed model is shown in Fig. (1) . The FMECA tool is not explained in this paper but its ap- lication in the proposed model is the same as explained in abib and Read (2015) , and for a review of its variations please see C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 215 Fig. 1. The model overview and interface of techniques. L F i o l a r h t s 4 ( f t i s u l s i t m t t i o h r w d t t f T f g f i a t o e O h m w d t f s iu, Liu, and Liu (2013) . It is the authors’ view that the use of MECA will provide risk priority numbers that will show the crit- cality of the basic events to the AHP. This combat the importance f other basic events to the failure hence resulting in devising a so- ution to the critical causal factors only. As a result the events that re not deemed critical remain unsolved and hidden in the system esulting in them causing failures in the future. In the following, the criteria for application of the proposed ybrid model for learning from failure are summarised. The de- ails about each of the tools used are given in the subsequent sub- ections. Step 1 : Develop an FTA for the root causes of the failure event, expand by mapping the FTA into an RBD and use them to derive an equivalent RBD model. Step 2 : With the information on root causes from the FTA formu- late a FMECA study. Step 3 : Using the risk priority numbers from the FMECA, hier- archical model from the FTA and information on parallel and series structures from the RBD, complete the AHP. Step 4 : Use causes of failure and ways of eliminating them from the FMECA, and the relative importance and priority numbers from the AHP to formulate a House of Quality matrix. Step 5 : Obtain comparison parameters from the AHP model to create a DMG and use relative importance weights from the HoQ matrix in case of events belonging to the same cell of the grid to make decisions on which to prioritise more within the same cell. .1. Fault Tree Analysis (FTA) and the Reliability Block Diagrams RBD) The FTA shows how basic events interact leading to the overall ailure under study. At the top of the FTA is the undesirable failure hat is under study, which in our case study is the ZS-CME serious ncident, with different failures connected underneath until the ba- ic events are reached. Basic events are the root causes of such fail- res that lead to the overall failure under investigation. The use of ogic AND- and OR- gates show the relationship between the ba- ic events and the failures. The AND- gate shows that the system s parallel and the OR-gate indicates series systems. Fig. (2) shows he FTA for the ZS-CME serious incident. It is from the FMECA that we obtain information on failure odes to be used on the formulation of the FTA. We also ob- ain information on how basic events are related towards causing he failure under investigation helping in formulating the reliabil- ty block diagram of the system. The FTA also shows the hierarchy f events that took place towards the failure under investigation ence giving input information for the AHP tool. In other words, the hierarchies in both FTA and AHP are al- eady considering every element (contributing factor). However, e group these factors under the two categories of direct and in- irect causes. The justification for the usage of the two AND- gates that leads o the ZS-CME serious incident is because the analysis is made af- er the incident has taken place, which means all the events that all under those branches had a part to play towards the incident. he occurrence of either event 1 or event 2 would have resulted in ailure on the operational side hence the justification for the OR- ate. Occurrence of unclear maintenance records had a negative ef- ect on the maintenance works carried out on the aircraft lead- ng to maintenance faults. The same applies to the failure by the ircraft design team to avoid interconnectivity of components in heir design, an aspect that played a vital role in the occurrence f the incident. These events on their own would have resulted in rrors in the maintenance of the aircraft hence the usage of the R- gate. As for the indirect causes of the incident, each event would ave had an effect without the influence of another event. This eans that event 6 would result in the outcome of the incident ithout event 7 or event 8. The same applies to the two other in- irect causes hence the reasoning behind the OR- gate. The RBD designed from the interdependency information ob- ained from the FTA is shown in Fig. (3) . It can be noted that rom this diagram the indirect causes of the incident form a series ystem, a finding that should be a main cause for worry. Failure 216 C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 Fig. 2. The Fault Tree analysis (FTA) of the ZS-CME serious incident. Fig. 3. The Reliability Block Diagram (RBD) of the ZS-CME serious incident. 1 o 4 w m of either event 6, 7 or 8 results in a complete breakdown of the whole branch of the RBD leaving reliance only on the direct causes branch. This shows that the indirect causes should never be taken lightly when analysing such a failure. The minimum cut set for this system would be a failure of one event in the indirect causes branch of the RBD (6, 7 or 8) and fail- ure of three events from the direct causes. The three events from the direct causes branch could be event 3 and either event 1 or event 2 and either event 4 or event 5. Shown below is the deriva- tion of the minimum cut set of this system. Cut set = (6 + 7 + 8). (4 + 5). (3). (1 + 2) which implies that the minimum cut sets are; t Minimum cut sets are 1.3.4.6; 1.3.4.7; 1.3.4.8; 1.3.5.6; 1.3.5.7; .3.5.8; 2.3.4.6; 2.3.4.7; 2.3.4.8; 2.3.5.6; 2.3.5.7 and 2.3.5.8 From the minimum cut sets we can also notice the importance f event 3. Its occurrence weakens all other sets. .2. Analytical Hierarchy Process (AHP) The MCDM technique of AHP helps in assessing the relative eights of multiple options against given criteria in an intuitive anner ( Parthiban, Zubar, & Garge, 2012 ). From the FTA we get he hierarchical information to be used on the AHP with the sec- C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 217 Fig. 4. AHP structure for the direct causes of the ZS-CME serious Incident. o c t b s c t w p r c p a u a b 1 t s t t v e t r a a t a i a P a a c v n t m m o t v h r p c o t i j t s w r p a 4 a c m t o c t o T s ( l nd level being the criteria and the basic events being the sub riteria. The alternatives will be other common factors to consider when rying to solve the basic events and these will include the proba- ility of re-occurrence when the basic event remains unsolved, the afety impact (severity) caused by the basic event and the cost in- urred when trying to devise a solution. These three common fac- ors are considered as decision variables since any decision taken ill tend to focus on mitigation against risk in the form of both robability of re-occurrence and severity, as well as the required esource allocation in terms of cost incurred. Using the AHP, to create the evaluation criteria, basic events are ompared and given weights that will give the information on their riorities when trying to put in place measures that will help to void the similar failure from occurring in the future. This is done sing the methodology that was explained by Saaty (2008) where matrix is formed by comparing each basic event against the other asic events from the FTA and giving a score between 1 and 9 with indicating that the events are equally important and 9 indicating hat one event is absolutely more important than the other. The cores depend on the authors’ judgement of the criteria. However he authors were informed by secondary data based on informa- ion included in the investigations reports of the accident. The formed matrix is then normalised by making the sum of all alues in a column equal to one. We then get the weight of each valuation criteria by adding the values in each row and obtaining he average. To ensure that both the direct causes and the indi- ect causes of the failure are considered as seen from the RBD they re both important, two AHP models are created one for the direct nd the other for the indirect causes. Fig. (4) shows the AHP for he direct causes of the failure with weights of both criteria and lternatives indicated. As for indirect causes, the model is shown n Fig. (5) . In order to come up with the weights for alternatives, we create matrix by comparing the alternatives (safety impact of the basic, robability of reoccurrence if it remains unsolved, cost of devising solution) to each other with respect to each basic event and give score of 1– 9. Then this matrix will undergo a normalisation pro- ess described above for the evaluation criteria and the average of alues in a row obtained. Finally to obtain the weights of alternatives the matrix of alter- atives with respect to the basic event is multiplied with the ma- c rix of criteria. Table 1 and 2 shows the pairwise comparison of the ain criteria with respect to the goal and the pairwise comparison atrix of the alternatives with respect to event 4 respectively. In order to explain how the priorities (weights) are derived nce the comparisons matrices are completed, we use the tradi- ional AHP eigenvalue method as described in Appendix A . From the AHP structure for direct causes we can notice that de- ising a solution for event 4(design errors) have to be given the ighest priority followed by event 3, event 5, event 2 and event 1 espectively. Also from the alternatives we can deduce that it is im- ortant to consider the safety impacts of each criterion before we an consider the probability of re-occurrence and the cost of devel- ping a solution. Probability of re-occurrence has a higher priority han the cost of devising a solution. This information will be very mportant in the formulation of the modified decision making grid. The weakness of the AHP include too much dependency on udgement of the person who is using it as such it can be subjec- ive an aspect that can be eliminated by having a group of experts tating their views on what a score to give to a certain criteria ith respect to the goal or an alternative with respect to a crite- ia. The strength includes simplifying of the users decision-making rocess by expertly comparing criteria with respect to the goal and lternatives with respect to criteria. .3. Simple House of Quality (HoQ) matrix According to Kuei (2002) HoQ is a structured and systematic pproach designed to translate customer needs into appropriate ompany business objectives. The HoQ matrix is made up of six ajor sections that show how the customer specifications are ranslated into the designer’s language. A full HoQ matrix is shown n Fig. (6) . As shown in Fig. (6) the formulation of a HoQ matrix start with ustomer requirements being defined and given the relative impor- ance weights as suggested by the customers, which are the rows, r the ‘Whats’ (i.e. what the customers wants) in the HoQ matrix). he next step is to come up with what designers can achieve to atisfy such requirements, which are the columns, or the ‘Hows’ ie how can the customer requirements be fulfilled). This is fol- owed by the customer’s perception of where the company is as ompared to competitors. At the centre of the matrix is the inter- 218 C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 Fig. 5. AHP structure for the indirect causes of the failure. Table 1 Pairwise comparison of the main criteria with respect to the direct causes of the failure. Design errors Unclear maintenance records FOD Pilot indecision Pilot negligence Weights Design errors 1 7 3 8 8 0.507 Unclear maintenance records 1/7 1 1/5 4 3 0.108 FOD 1/3 5 1 7 6 0.283 Pilot indecision 1/8 1/4 1/7 1 3 0.063 Pilot negligence 1/8 1/3 1/6 1/3 1 0.040 Table 2 Pairwise comparison of alternatives with respect to Event 4 (design error). Safety impact of the event Probability of Re -occurrence if unsolved Cost of devising a solution Weights Safety impact of the event 1 5 9 0.723 Probability of Re -occurrence if unsolved 1 \ 5 1 5 0.216 Cost of devising a solution 1 \ 9 1 \ 5 1 0.061 ‘ t e r v a 4 v a c b ( f m w o t f D h S relationship matrix, which shows how the customer needs relate to the engineering characteristics. At the roof of the matrix is a depiction of how the engineer- ing characteristics affect each other. As such the roof of the matrix presents an opportunity for engineers to specify the various en- gineering features that have to be improved collateral ( Hauser & Clausing, 1988 ). The final aspect of the HoQ matrix is the technical assessment and target for each engineering characteristic for the betterment of the product. Hauser and Clausing (1988) described the HoQ as a kind of con- ceptual map that provides the means for inter functional planning; which means it can be used as a diagrammatic representation. It is as such, that in the proposed model a simple HoQ matrix is employed in order to provide a visual representation of the causal factors of the failure and the ways of eliminating or reducing the severity of such factors. The causal factors will occupy the part of the matrix where cus- tomer needs are defined and the engineering characteristics sec- tion will be occupied with ways of eliminating the causal factors to the failure under investigation. The weights obtained for the cri- teria’s in the AHP will be transferred to the relative importance section. It must be noted that both direct and indirect causes to the fail- ure are treated as having equal importance as such each have its section in the matrix. A simplified HoQ matrix for the failure be- ing investigated is shown in Fig. (7) . In Fig. (7) , we simply map the Whats’ (rows) in the HoQ model against the ‘Hows’ (columns). Not hat the ‘Hows’ are potential solutions that in our view can address ach of the rows in varying degrees as captured by the X’s in the elationship matrix in the middle of the grid, which is a simplified ersion of HoQ model. Note that the top of the matrix was gener- ted using the information obtained from Moakofi (2016) . .4. Decision making grid (DMG) The decision making grid is an MCDM technique which pro- ides means of identifying which maintenance actions are vi- ble for a system In order to provide optimised balance between ost and performance risks. Labib developed this concept in 1996 y combining the rule-based approach with the AHP for MCDM Labib, Williams, & Connor, 1998 ). It acts as a map where using multiple criteria the worst per- orming machines can be classed ( Labib, 1998 ). This grouping of achines aid in the implementation of appropriate actions that ill improve their performance, hence moving them to the region f low downtime and low frequency. The objective is to improve he performance of the machines so that they move to the low requency and low downtime cell of the DMG. Fig. (8) shows the MG as proposed by Labib (1998) . The machines that are classed in the cell for high frequency and igh downtime are considered to be the worst performing ones. o, to ensure the improvement in performance for such machines C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 219 Fig. 6. A full HoQ Matrix ( Hauser & Clausing, 1988 ). a c e t g m t h n o l F c ( h s d d c i t t a o s t r m a v p d n design out maintenance strategy is used. For machines that are lassed in the low frequency high downtime cell, the rightful strat- gy to implement is the condition-based maintenance. If a machine is put in the low downtime and high frequency, he rule that applies is autonomous maintenance or Skill Level Up- rade (SLU). This implies that operators are trained to perform the aintenance associated with such machine, as the tasks are rela- ively easy. Whereas, if a machine is put in the low frequency and igh downtime, the rule that applies is Condition-Based Mainte- ance (CBM). This implies that there is need to monitor condition f a major type of problem that seldom occurs. As for the ones al- ocated to a low frequency and low downtime cell, an ‘Operate To ailure - OTF’ strategy has to be implemented. For the remaining ells of the DMG, the usage of the Total Preventive Maintenance TPM) strategy needs to be continued. Finally a high frequency, igh downtime implies a Design Out Maintenance (DOM) strategy ince the whole machine needs to be reconfigured. Traditionally, the DMG model that compares frequency and owntime have been used in helping decision makers and policy evelopers in selecting the rightful maintenance strategies for their ritical assets. This technique has recently been extended to help n learning from failures. Aslam-Zainudeen and Labib (2011) stated hat the technique has also been used in crisis management. In this paper we modify the DMG to use it in ensuring that all he causal factors of a failure are solved and preventive measures re put in place as to avoid them to aid in the formulation of an- ther failure in the future. From the AHP developed in the earlier ection we obtain information on which two alternatives we need o pay attention to in order to ensure that the basic events don’t esult in another failure in the future. The two alternatives that received higher rankings in the AHP odels provided earlier are the safety impact of the basic events nd the Probability of re-occurrence. As such we are going to de- elop a DMG with increasing safety impact on the x-axis and the robability of re-occurrence in the y-axis. Each axis would then be ivided into three levels (low, medium and High) to form a grid of ine sections as shown in Fig. (9) . 220 C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 Fig. 7. The simplified HoQ matrix. Fig. 9. The modified DMG model. p c s m o c a h c l s a Therefore, in order to summarise how in Fig. (9) , the results of previous methods can feed into this grid, we do this in two ways, one is to determine the new set of axes used (compare original grid in Fig. (8) to the modified one in Fig. (9) ). The second way is through plotting the different basic events from Fig. (2) and their ranking with respect to the two axes as shown in Figs. (4 and 5 ) into the grid. As for how the borders have been set, this judgement is based in the variation in the values, but can also be formalised using different methods. For more details about different methods to set the borders in DMG, please see Seecharan, Labib, and Jar- dine (2018) . Using personal judgement each basic event is allocated to the most appropriate cell of the grid. Each of these cells indicates the priority that should be given to the allocated event. For example the basic event that is allocated to the high safety impact high Fig. 8. DMG ( Lab robability of re-occurrence cell must be given higher priority as ompared to the events in other cells. The objective of this DMG is to ensure that preventive mea- ures for the basic events that are in the High-high cell in the atrix are put in place so that they move to the cells that are f lower safety impact and lower probability of re-occurrence as ompared to their initial allocated cell. The order of ensuring that ll basic events are tackled is to start with event allocated to the igh safety impact high probability of re-occurrence (high-high) ell then high-medium, medium-high, medium-medium, high-low, ow-high, low-medium and low-low respectively. As the output of the proposed model each basic event that re- ulted in the occurrence of the incident under study are prioritized nd preventive measure put in place to ensure that their influences ib, 1998 ). C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 221 e l a i d r t i i e f a t a a l r f o s h e u t c e b Z 5 L v f i h s g u n d h i a w a t r t p o i t o p t l u a t t e s a & l h I A f T f ( A m p ( a b ⎡ ⎢⎣ o A w c i a R A B C C D H H I ven on the future are eliminated. These priorities indicates the evel of response, at which the solutions need to be implemented, s such could be compared to the maintenance strategies proposed n the original DMG model by Labib (1998) . In the proposed model, DOM has the same meaning as ‘imme- iate response’. This means that event 3 and 4 would require being esolved immediately. The level of response required in resolving he causal factors reduces with the decrease in either the safety mpact or the probability of re-occurrence, with the factors located n OTF cell being the last one to resolve. Event 7 is solved before vent 6 because from the AHP model for indirect causes the rating or Probability for reoccurrence is higher than that of safety impact nd this is the opposite for direct causal factors. The strengths of this feature of the proposed model is that all he causal factors to the failure are taken and put in one place nd solutions developed one at a time ensuring that the ones that re of high safety impact and have high chance of re-occurring if eft unattended are given attention first before dealing with the emaining ones. This helps to ensure that no causal factors to a ailure are left hidden in the system an issue that can spark a re- ccurrence in the future. The weakness of this feature is that basic events that lie in the ame cell of the grid but at different extremes are treated the same ence in actual facts they have different states. The example being vent 3 and 4. As a solution to this weakness, the proposed model tilises the HoQ matrix that feeds information on relative impor- ance of events to the simplified DMG hence priorities. As such, we an tell that event 4 have to be given higher priority than event 3 ven though the two are in the same cell of the DMG. This can also e solved by the application of fuzzy logic as proposed by Aslam- ainudeen and Labib (2011) . . Conclusions The proposed hybrid model is an extension of the model by abib and Read (2015) . The ZS-CME serious incident has been in- estigated using the proposed model which showed that the causal actors number 3 and 4 should be given high priority when devis- ng preventive measures as they fall in the high safety Impact and igh probability of re-occurrence cell in the modified DMG. The re- ults from the HoQ indicates that even though event 3 and 4 are iven the same priority in the DMG event 4 should be given the tmost priority as it has a high value of relative importance. For Indirect causes, the modified DMG shows that event 7 eeds more priority as compared to event 6 because it falls un- er the high probability of re-occurrence region as compared to igh safety Impact region. This is so because the AHP model for ndirect causes of the failure have awarded a high weight to prob- bility of re-occurrence rather than safety impact, as it is the case ith direct causes. The novelty of the proposed model came from the fact that HoQ nd DMG are used to show the priorities that need to be given o each of the causal factors of the failure and comparing their esults. This comparison cancels out the weakness of the DMG hat is associated with events in the same cell but on the op- osite extremes hence eliminating the need for fuzzy logic. The ther strength that is associated with the proposed model is that t leaves no stone unturned in the event of a failure as was seen by aking consideration of both the direct and indirect causal factors f the failure. The weakness of the proposed model comes with too much de- endency on personal judgement. This does not only bring bias to he failure investigation but also inconsistencies. The authors be- ieve that the inconsistencies and bias can be eliminated by the se of a group of people in situations where personal judgements re required hence resulting in the use of collective judgement of he group. However, group decision making has its own assump- ions and challenges which are beyond the scope of this paper. For xample, individual decision makers in the group can be either as- umed to be equally weighted, or a method needs to be derived to llocate weights to each decision maker ( Chakhar, Ishizaka, Labib, Saad, 2016; Ishizaka & Labib, 2011b ). In addition, there is a chal- enge in finding a suitable group that can represent all the stake- olders in the decision making process ( Poplawska, Labib, Reed, & shizaka, 2015 ). cknowledgements The authors would like to thank the reviewers for their insight- ul comments and suggestions to improve the quality of the paper. he second author would like to acknowledge the funding received rom the Centre for Research and Evidence on Security Threats ESRC Award: ES/N009614/1 ). ppendix A In order to briefly describe the traditional AHP eigenvalue ethod , we start from the case of a consistent matrix with known riorities p i . If the matrix is perfectly consistent, the transitivity rule 1) holds for all comparisons a ij : i j = a ik · a k j (1) In this case, the comparisons of the alternative i and j is given y p i / p j . If, we multiply it with the priority vector � p , we obtain: p 1 / p 1 p 1 / p 2 … p 1 / p n p 2 / p 1 p 2 / p 2 … p 2 / p n … … … … p n / p 1 p n / p 2 … p n / p n p 1 p 2 ... p n ⎤ ⎥ ⎦ = n ⎡ ⎢ ⎣ p 1 p 2 ... p n ⎤ ⎥ ⎦ r grouped: � p = n � p here � p : vector of the priorities n : dimension of the matrix A : comparison matrix (2) Eq. (2) is the formulation of an eigenvector problem. The cal- ulated priorities are exact for a consistent matrix. When slight nconsistencies are introduced, priorities should vary only slightly ccording to the perturbation theory ( Saaty, 2003 ). eferences slam-Zainudeen, N. , & Labib, A. (2011). Practical application of the Decision Making Grid (DMG). Journal of Quality in Maintenance Engineering, 17 (2), 138–149 . hattacharjya, D. , & Deleris, L. A. (2012). From reliability block diagrams to fault tree circuits. Decision Analysis, 9 , 128-137 . hakhar, S. , Ishizaka, A. , Labib, A. , & Saad, I. (2016). Dominance-based rough set approach for group decisions. European Journal of Operational Research, 251 , 206–224 . ox, S. , Jones, B. , & Collinson, D. (2006). Trust relations in high-reliability organiza- tions. Risk Analysis, 26 (5), 1123–1138 . avidson, G. , & Labib, A. W. (2003). Learning from failures: Design improvements using a multiple criteria decision making process. Journal of Aerospace Engineer- ing, Proceedings of the Institution of Mechanical Engineers Part G, 217 (4), 207–216 . auser, J. R., & Clausing, D. (1988). The house of quality. Havard Business Review , May-June. pp. 63–73. owick, S. , Ackermann, F. , Walls, L. , Quigley, J. , & Houghton, T. (2017). Learning from mixed OR method practice: The NINES case study. Omega, 69 , 70–81 . shizaka, A. , & Labib, A. (2014). A hybrid and integrated approach to evaluate and prevent disasters. Journal of the Operational Research Society, 65 , 1475–1489 . http://dx.doi.org/10.13039/501100000613 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0001 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0002 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0003 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0004 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0005 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0006 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0007 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0007 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0007 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0007 222 C. Stephen, A. Labib / Expert Systems With Applications 93 (2018) 212–222 M P S S Y Z Ishizaka, A. , & Labib, A. W. (2011a). Review of the main developments of AHP. Expert Systems With Applications, 38 (2011), 14336–14345 . Ishizaka, A. , & Labib, A. W. (2011b). Selection of new production facilities with the group analytic hierarchy process ordering method. Expert Systems With Applica- tions, 38 , 7317–7325 . Jefcott, S. , Pidgeon, N. , Weyman, A. , & Walls, J. (2006). Risk, trust and safety culture in the UK train operating companies. Risk Analysis, 26 , 1105–1121 . Kuei, C.-H. (2002). House of quality (QFD) in a minute. The International Journal of Quality and Reliability Management, 19 (4), 4 87–4 88 . Labib, A. W. , Williams, G. B. , & O’Connor, R. F. (1998). An intelligent maintenance model (System): An application of A.H.P. and a fuzzy logic rule-based controller. Journal of Operational Research Society, 9 (7), 745–757 . Labib, A. (1998). World-class maintenance using a computerised maintenance man- agement system. Journal of Quality in Maintenance Engineering, 4 (1), 66–75 . Labib, A. (2015). Learning (and unlearning) from Failures: 30 yearson from Bhopal to Fukushima an analysis through reliability engineering techniques. Process Safety and Environmental Protection, 547 , 1–11 . Labib, A. , & Read, M. (2015). A hybrid model for learning from failures: The Hurri- cane Katrina disaster. Expert Systems with Applications, 42 , 7869–7881 . Liu, H. C. , Liu, L. , & Liu, N. (2013). Risk evaluation approaches in failure mode and effects analysis: A literature review. Expert systems with applications, 40 (2), 828–838 . Love, P. E. D. , Lopez, R. , & Edwards, D. J. (2013). Reviewing the past to learn in the future: Making sense of design errors and failures in construction. Structure and Infrastructure Engineering, 9 (7), 675–688 . Moakofi, O. B. (2016). MTC/AIG/43/15 Report on the serious incident to Bombardier CRJ-100, ZS-CME on Runway 08 at SSKIA on 31 August 2015 . Aircraft Incident Re- port. Gaborone: Ministry of Transport and Communication Directorate of Acci- dent Investigation. organ, J. S. , Belton, V. , & Howick, S. (2016). Lessons from mixing OR methods in practice: Using DES and SD to explore a radiotherapy treatment planning pro- cess. Health Systems, 5 , 166–177 . Morgan, J. , Howick, S. , & Belton, V. (2017). A toolkit of designs for mixing discrete event simulation and system dynamics. Forthcoming in European Journal of Op- erational Research, 257 , 907–918 . arthiban, P. , Zubar, H. A. , & Garge, C. P. (2012). A multi criteria decision making approach for supplier selection. Procedia Engineering, 38 , 2312–2328 . Poplawska, J. , Labib, A. , Reed, D. , & Ishizaka, A. (2015). A dynamic framework for stakeholder identification and salience measurement using fuzzy logic method- ology applied to a corporate social responsibility case study. Journal of Cleaner Production, 105 , 103–115 . aaty, T. (2003). Decision-making with the AHP: Why is the Principal Eigenvector necessary? European Journal of Operational Research, 145 (1), 85–91 . aaty, T. L. (2008). Decision making with the analytical hierarchy process. Interna- tional Journal of Services Sciences, 1 (1), 83–98 . Seecharan, T. S. , Labib, A. , & Jardine, A. K. S. (2018). Approaches to criticality and se- lection of maintenance strategies: Decision making grid vs. jack-knife diagram. Journal of Quality in Maintenance Engineering (in press) . Shanthikumar, J. G. , & Sargent, R. G. (1983). A unifying view of hybrid simulation/ analytic models and modeling. Operations Research, 31 , 1030 . Torrel, W. , & Avelar, V. (2010). Mean time between failure: Explanation and stan- dards. APC by Schneider Electric, 1 (78), 1–10 . unusa-Kaltungo, A. , Kermani, M. M. , & Labib, A. (2017). Investigation of critical fail- ures using root cause analysis methods: Case study of ASH cement PLC. Engi- neering Failure Analysis, 75 , 25–45 . ubair, M. , Park, S. , Heo, G. , Hassan, M. , & Aamir, M. (2015). Study on nuclear acci- dent precursors using AHP and BBN, a case study of Fukushima accident. Inter- national Journal of Energy Research, 39 , 98–110 . http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0008 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0008 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0008 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0008 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0009 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0010 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0011 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0012 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0013 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0013 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0014 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0014 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0015 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0016 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0017 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0018 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0019 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0020 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0021 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0022 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0023 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0023 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0024 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0024 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0024 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0024 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0024 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0025 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0026 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0026 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0026 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0026 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0027 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0028 http://refhub.elsevier.com/S0957-4174(17)30713-3/sbref0028 A hybrid model for learning from failures 1 Introduction 2 Theoretical frameworks 3 ZS-CME serious incident 3.1 Synopsis of the incident 3.2 Investigation findings 4 Proposed model 4.1 Fault Tree Analysis (FTA) and the Reliability Block Diagrams (RBD) 4.2 Analytical Hierarchy Process (AHP) 4.3 Simple House of Quality (HoQ) matrix 4.4 Decision making grid (DMG) 5 Conclusions Acknowledgements Appendix A References 
work_ucw2yjhblnglzeroqfl72wyxny	----	Development of a QFD-based expert system for CNC turning centre selection econstor Make Your Publications Visible. A Service of zbw Leibniz-Informationszentrum Wirtschaft Leibniz Information Centre for Economics Prasad, Kanika; Chakraborty, Shankar Article Development of a QFD-based expert system for CNC turning centre selection Journal of Industrial Engineering International Provided in Cooperation with: Islamic Azad University (IAU), Tehran Suggested Citation: Prasad, Kanika; Chakraborty, Shankar (2015) : Development of a QFD- based expert system for CNC turning centre selection, Journal of Industrial Engineering International, ISSN 2251-712X, Springer, Heidelberg, Vol. 11, pp. 575-594, http://dx.doi.org/10.1007/s40092-015-0122-x This Version is available at: http://hdl.handle.net/10419/157461 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by/4.0/ www.econstor.eu ORIGINAL RESEARCH Development of a QFD-based expert system for CNC turning centre selection Kanika Prasad1 • Shankar Chakraborty1 Received: 9 March 2015 / Accepted: 17 August 2015 / Published online: 20 October 2015 � The Author(s) 2015. This article is published with open access at Springerlink.com Abstract Computer numerical control (CNC) machine tools are automated devices capable of generating com- plicated and intricate product shapes in shorter time. Selection of the best CNC machine tool is a critical, complex and time-consuming task due to availability of a wide range of alternatives and conflicting nature of several evaluation criteria. Although, the past researchers had attempted to select the appropriate machining centres using different knowledge-based systems, mathematical models and multi-criteria decision-making methods, none of those approaches has given due importance to the voice of cus- tomers. The aforesaid limitation can be overcome using quality function deployment (QFD) technique, which is a systematic approach for integrating customers’ needs and designing the product to meet those needs first time and every time. In this paper, the adopted QFD-based methodology helps in selecting CNC turning centres for a manufacturing organization, providing due importance to the voice of customers to meet their requirements. An expert system based on QFD technique is developed in Visual BASIC 6.0 to automate the CNC turning centre selection procedure for different production plans. Three illustrative examples are demonstrated to explain the real- time applicability of the developed expert system. Keywords CNC turning centre � Expert system � Multi- criteria decision-making � Quality function deployment Introduction Machine tools have been around since the industrial rev- olution, extensively used to manufacture parts/components of machines, which is a process of selectively removing material to create a desired shape. They are capable of producing parts/components of different shapes and sizes, having simple to complex contours. These days’ products are becoming much more complex, and difficult to design and manufacture. Hence, the manufacturing organizations are forced to develop and adopt new technologies to avoid long design and machining time for complex products. So, machine tools have been gradually evolved out over the past few decades to meet the increasing demand of man- ufacturing complicated components with high degree of accuracy in large volume. The computer numerical control (CNC) machine tools are now being extensively applied for automated machining operations to help in achieving faster production rate with decreased human involvement and effort. Development of CNC machine tools is a great contribution to the manufacturing domain as automation of the machining process with flexibility to handle small to medium batch quantities in part production now becomes possible. This CNC technology can be applied to milling, turning, grinding, boring, drilling machines, flame cutters, etc. The CNC as the name suggests is equipped with computers that help in organizing and restoring informa- tion to attain high accuracy and speed in part production. Its basic aim is to achieve the desired objectives of the manufacturing organizations within the limited available budget. In an industrial setting, CNC machine tools can be combined into entire cells of tooling machines that can operate independently of each other. They are often driven by completely digital designs, which eliminate the need for design blueprints to be physically drawn up. All the CNC & Shankar Chakraborty s_chakraborty00@yahoo.co.in Kanika Prasad pdkanika@gmail.com 1 Department of Production Engineering, Jadavpur University, Kolkata, West Bengal 700 032, India 123 J Ind Eng Int (2015) 11:575–594 DOI 10.1007/s40092-015-0122-x http://crossmark.crossref.org/dialog/?doi=10.1007/s40092-015-0122-x&amp;domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1007/s40092-015-0122-x&amp;domain=pdf machines are able to accurately control motions in multiple directions, and hence, can generate complex contours and shapes on the workpieces with higher dimensional accu- racy and precision. Many of them are capable of running for several days without human intervention. These auto- mated features of CNC machine tools make it possible to produce thousands of identical parts/components with minimal supervision, and allow the operators to perform other tasks thus saving a lot of time. Besides this, they can also produce parts/components with a level of accuracy that can be nearly impossible to attain using the older tools. This improved accuracy can help eliminate waste due to production of less defective parts. Features, like automatic tool and job change, etc. substantially reduce the machin- ing time, thereby trimming down the total production cost at the end. The CNC machine tools, capable of performing repetitive complicated and unsafe machining operations, become highly productive and cost effective, thus gaining wide acceptance in manufacturing industries. A manufac- turing organizations’ productivity is directly related to the proper choice of its CNC machine tools as their proper use can increase the overall production while effectively uti- lizing the available resources and reducing the chance of human injury. Although CNC machine tools are highly productive and flexible, they are also quite expensive to procure, install and maintain. However, their ability to enhance productivity can easily offset the huge initial investment if they are properly evaluated and selected. Hence, it is an important decision for the production planners to choose the most appropriate CNC machine tool among various available alternatives to fulfil the organi- zational requirements. The CNC machine tool selection process is focussed on fulfilment of two basic requirements, i.e. (a) boundary (fixed) conditions, and (b) performance expectations (de- sired results). For a CNC machine tool, the boundary conditions include spatial constraints, range of spindle speed, etc. whereas, performance expectations comprise positional accuracy, repeatability, capability to generate complicated parts, etc. Once these conditions and expec- tations are identified and prioritized, the most appropriate CNC machine tool that meets the basic application requirements can be easily selected. Choosing the best suited CNC machine tool from a wide range of similar alternatives is a complex and time-consuming task as it involves consideration of a large number of qualitative and quantitative factors, such as capital cost, table area, three axes movement, power, spindle speed range, machining diameter and length, tool capacity, flexibility, safety, etc. which are sometimes interrelated to each other. Technical brochures can effectively convey various machine features/ specifications and some level of technical performance data, but they do not provide true comparisons with the other competing machines. So, lack of accurate informa- tion is another problem faced by the production planners while selecting a CNC machine tool for a specific appli- cation. The production planners also need to analyse huge amount of raw data consisting of many interrelated factors for proper and effective evaluation of available CNC machine tool alternatives, which involves human expertise in a particular domain. But, human expertise is scarce, and it may not be always possible to analyse a large amount of data and crucial details of a problem. Moreover, it has limited working memory, and hence, cannot comprehend the data quickly. These shortcomings can be overcome while developing a database containing the technical specifications of various CNC machine tools, which can be updated from time to time. The database can then be integrated with an expert system to ease out and automate the CNC machine tool selection process. A selection pro- cedure is not a simple task, but rather a sequence of interdependent activities that must take into consideration the customers’ requirements, manufacturing economics, design expectations and above all, human safety. Thus, there is a need for a systematic and rational approach that can aid in solving the CNC machine tool selection prob- lem, avoiding human intervention and expertise. In this paper, an expert system based on quality function deployment (QFD) technique is developed in Visual BASIC 6.0 to automate the CNC turning centre selection procedure for three different production plans, i.e. flexible, mass and tailor made (customer specific). The database containing the technical specifications of more than 200 CNC turning centres is developed in MS-Access. The QFD technique is augmented to systematically integrate the needs of customers with various engineering or technical characteristics and derive the priority weights for different technical requirements while evaluating the feasible CNC turning centres. Three numerical examples are illustrated to demonstrate the applicability of the developed QFD-based expert system for CNC turning centre selection. The rest of the paper is organized as follows. A literature review on the past researches is provided in Sect. 2 and QFD methodology is explained in Sect. 3. Section 4 describes the development procedure of QFD-based expert system for CNC turning centre selection. Three illustrative examples are provided in Sect. 5, and in Sect. 6, final conclusions are drawn. Literature review in machine tool selection Till date, many studies have been reported in the literature on solving machine tool selection problems using diverse multi-criteria decision-making (MCDM) methods, mathe- matical models and knowledge-based systems. Sun (2002) 576 J Ind Eng Int (2015) 11:575–594 123 applied data envelopment analysis (DEA) to assess 21 CNC machines with respect to their technical and cost factors. Sensitivity analysis with variable variation and weight restrictions identified six CNC lathes as ‘good buys’ to be subsequently recommended for further consideration. Yurdakul (2004) presented a new strategic justification tool employing analytic hierarchy process (AHP) and analytic network process (ANP) which were applied to calculate the contribution of machine tool alternatives to manufacturing strategy and rank them based on the developed hierarchical structures. Ayağ and Özdemir (2006) introduced triangular fuzzy numbers in pair-wise comparison of the AHP matrix to overcome the vagueness and uncertainties in judgments of the decision makers in the conventional AHP method while evaluating machine tool alternatives. Ayağ (2007) integrated AHP method with simulation technique to determine the best machine tool satisfying the needs and expectations of a manufacturing organization. The AHP method was adopted to narrow down the list of feasible machine tool alternatives and a simulation generator was then used to automatically model the manufacturing organization. Durán and Aguilo (2008) developed a fuzzy AHP-based software for evaluation and justification of machine tools. Önüt et al. (2008) proposed a combined fuzzy AHP and fuzzy technique for order preference by similarity to ideal solution (TOPSIS) for machine tool selection, while incorporating triangular fuzzy numbers in the traditional AHP and TOPSIS methods. Dağdeviren (2008) integrated AHP and preference ranking organization method for enrichment evaluation (PROMETHEE) for equipment selection. The AHP method was applied to analyse the structure of the selection problem and deter- mine the criteria weights, whereas, PROMETHEE method was employed to obtain the final ranking of the alterna- tives. Yurdakul and İç (2009) discussed about the benefits of using fuzzy numbers instead of crisp numbers in a TOPSIS method-based machine tool selection model. İç and Yurdakul (2009) developed a decision support system (DSS) using extended versions of MCDM approaches, like fuzzy AHP and fuzzy TOPSIS to help the decision makers in machining centre selection decisions. Qi (2010) devel- oped a comprehensive evaluation model for machine tool selection and applied fuzzy integral approach for aggre- gation of performance scores of the alternatives with respect to different criteria. Ayağ and Özdemir (2011) proposed an intelligent approach for machine tool selection using fuzzy ANP to consider the vagueness and uncertainty existing in the importance attributed to judgment of the decision maker. Özgen et al. (2011) presented a combined application of modified DELPHI method, AHP and PRO- METHEE approaches with fuzzy set theory for solving machine tool selection problems. İç (2012) applied an integrated TOPSIS and design of experiments approach to solve a CNC machine tool selection problem in a real-time industrial environment. İç et al. (2012) developed a com- ponent-based machining centre selection model based on AHP, which would use only the technical specifications while evaluating the machining centre components. Ilangkumaran et al. (2012) developed an evaluation model based on AHP and VIKOR (VlseKriterijumska Opti- mizacija I Kompromisno Resenje) methods under fuzzy environment for selection of the best machine tool among various alternatives. Taha and Rostam (2012) developed a DSS to select the best alternative machine tool using a hybrid approach of fuzzy AHP and PROMETHEE meth- ods. Ayağ and Özdemir (2012) applied ANP together with the modified TOPSIS method for performance analysis on machine tools. Furthermore, a fuzzy ANP approach was adopted to deal with the imprecise and uncertain human comparison judgments. Samvedi et al. (2012) integrated fuzzy AHP and grey relational analysis (GRA) approaches for selection of a machine tool from a given set of candi- date alternatives. Fuzzy AHP was applied to calculate the criteria weights followed by GRA method to rank the alternatives. Aghdaie et al. (2013) applied step-wise weight assessment ratio analysis (SWARA) and complex propor- tional assessment of alternatives with grey relations (COPRAS-G) methods for machine tool evaluation and selection. Dawal et al. (2013) presented a simple approach for multi-attribute-based selection of machine tools using fuzzy AHP and TOPSIS methods. Tho et al. (2013) inte- grated intuitionistic fuzzy entropy and TOPSIS method to deal with the vague information in the decision-making process for machine tool selection. Nguyen et al. (2014) presented a hybrid approach integrating fuzzy ANP and COPRAS-G methods for evaluating machine tools with consideration of interactions between the considered attri- butes. Xin et al. (2014) proposed an optimal machine tool selection approach based on interval-valued fuzzy C-means clustering algorithm. Sahu et al. (2015) applied VIKOR method for determining a compromise ranking list of five alternative CNC machine tools while considering 21 sub- jective evaluation criteria. A comparative study between various MCDM tech- niques that had been used in the past and the developed QFD-based model to solve the machine tool selection problems is provided in Table 1. It can be observed from the literature that the earlier MCDM methods as applied for machine tool selection are unable to take into account the voice of customers in the evaluation process. But, the customers’ requirements have an immense importance in the present-day manufacturing scenario where there is an enormous competition to capture every single percentage of market share available, which is primarily driven by an organization’s ability to satisfy its customers. Further, it is also realized that till date, no J Ind Eng Int (2015) 11:575–594 577 123 attempt has been made to interrelate the technical requirements of various machine tools with the corre- sponding customers’ requirements. Moreover, the MCDM methods that had been previously applied to select machine tools have one or more inherent draw- backs. Like, in AHP and ANP methods, huge mathe- matical calculations are involved, and if there is any ambiguity or uncertainty in the pair-wise comparison matrix, the reliability of the derived solutions is itself questionable. Similarly, TOPSIS method introduces two reference points, i.e. positive ideal and negative ideal solutions, but it does not consider the relative importance of the distances of the alternatives from those two points. It signifies that the selected alternative may not always be the best solution (i.e. closest to the ideal solution). PROMETHEE method is based on some preference functions and in most of the real-time situations, the decision maker faces a problem in selecting the most appropriate preference function. VIKOR method is more time consuming as the final decision is to be compro- mised taking into consideration two other factors as acceptable advantage and acceptable stability of the decision. On the other hand, DEA comprises of too many lengthy computations and cannot often be solved manually. The decision maker needs to have some soft skills in computer programming for performing such calculations. These drawbacks of MCDM methods pre- viously adopted for machine tools selection can be effectively addressed while developing a QFD-based expert system, which can not only incorporate the cus- tomers’ requirements into the selection process but also interrelate them with the technical requirements. It is also supposed to be superior to other MCDM methods on its ability to deal with the dynamic nature of the decision-making problems. Hence, in this paper, a user- friendly software prototype with graphical interface in Visual BASIC 6.0 is developed to help the production planners in selecting the best CNC turning centre to meet the dynamic requirements of a specific production system and accomplish the managerial benefits of an automated selection procedure. QFD methodology In this era of global competition, the success of any organization depends on its ability to understand and meet the ever changing needs of the customers. QFD method- ology is now being recognized as an efficient technique to deal with the voice of customers which includes the cus- tomers’ needs for a product, customers’ perceptions on the relative importance of those needs, and the relative per- formance of the manufacturing organization and its main competitors on those needs. QFD basically consists of two components, i.e. quality and function, which are deployed in the design process. The quality deployment brings cus- tomer’s voice into the design phase, whereas, function deployment links different organizational functions and technical requirements in the design to manufacturing. It is a focused methodology for carefully listening to the voice of customers, and then effectively responding to those needs and expectations, thereby attaining the highest cus- tomer satisfaction. It is employed to translate the cus- tomers’ requirements, in terms of engineering or technical Table 1 Comparative study between different MCDM methods used for machine tool selection Method Flexibility Operational approach Computational time Complexity Decision maker’s involvement Compensatory nature Type of data AHP High Pair-wise comparison Very high Very high Very high Yes Ordinal ANP High Pair-wise comparison Very high Very high Very high Yes Ordinal PROMETHEE Moderate Pair-wise comparison High High Very high No Mixed GRA High Grey system Moderate Moderate Moderate Yes Mixed DEA Very low Efficiency measurement Very high High No No Cardinal TOPSIS Low Euclidean distance Moderate Moderate Moderate No Cardinal VIKOR Low Euclidean distance High Moderate Moderate No Cardinal QFD-based expert system High Pair-wise comparison Very low Moderate Low Yes Mixed 578 J Ind Eng Int (2015) 11:575–594 123 characteristics, that can be deployed through product planning, process planning, service design and part development. This quality improvement tool was developed in late 1960s in Japan by Akao who is regarded as the father of QFD and was first implemented at the Mitsubishi Heavy Industries Kobe Shipyard in 1972 under the guidance of Shigeru Mizuno and Yasushi Furukawa (Akao 1990). Ford Motor Company, Toyota, Procter and Gamble, Mitsubishi, Campbell’s soup, Hewlett-Packard, Kodak, IBM, Xerox and 3M Corporation were among the early adopters of QFD methodology. Chan and Wu (2002) reviewed the implementation of QFD technique in different organiza- tions, such as shipbuilding, automobiles, electronics, soft- ware, banking and accounting, health care, education and research, retail outlets, apartment layouts, airline services, office equipment, consumer products, financial services, telephone services, gas and electrical services, distribution networks, traffic management, food industry, fishing industry, chocolate industry, online gaming, hot bar sol- dering, etc. It is observed that the above-mentioned appli- cations are for product and process development, but lately, QFD has also been applied for selection of suppliers (Bevilacquaa et al. 2006; Bhattacharya et al. 2010; Shad et al. 2014), non-traditional machining processes (Chak- roborty and Dey 2007), industrial robots (Karsak 2008), product design and development (Liu 2011; Soota et al. 2011) and materials (Mayyas et al. 2011; Prasad and Chakraborty 2013). QFD is such a systematic, robust and practical quality improvement tool that apart from the above-cited domains, it has also been applied in some unconventional fields, like game of soccer (Partovi and Corredoira 2002). QFD understands how the customers or users become interested and satisfied with the end products. Customers’ requirements and their relationships with the design char- acteristics are the driving force behind QFD methodology (Dursan and Karsak 2013). It can be used to translate subjective quality criteria into objective ones that can be quantified and measured. It is a complimentary method for determining how and where priorities are to be assigned in product or process development, and intelligently links the needs of the customers with the design and development of a product. There are three basic steps in implementing QFD methodology. At first, the spoken and unspoken wants or needs of the customers are prioritized. In the next step, those needs are translated into technical characteris- tics and specifications, and in the final step, a quality product or service is developed and delivered focusing everybody towards customer satisfaction. QFD methodology can be adopted to process both qualitative and quantitative data. Its main merit over the other MCDM approaches is that it provides flexibility to the decision makers to correlate both customer needs and engineering metrics through assigning scores and weights to them, and at the same time, it defines the direction of improvement for each metric which may be directly or inversely proportional to each other (Mayyas et al. 2011). QFD for a product is developed through brainstorming of a team, which comprises of six to eight persons, consisting of representatives from various cross-functional departments, like marketing, design, production, quality assurance, testing, purchasing, vendor, etc. The representatives from those cross-functional departments are collectively known as the QFD team, as shown in Fig. 1. The main advantage of this cross-functional group decision-making approach is that it takes opinions from the representatives of various departments and incorporates them into the product, lead- ing to a better quality product and higher customer satis- faction. This also avoids any biasness and partiality in the decision-making process. QFD employs a matrix format to capture a number of issues vital for the planning process. According to Dai and Blackhurst (2012), the overall process of QFD is based on its core matrix framework, called house of quality (HoQ), which is used to intertwine customers’ needs, service design or management requirements, target design goals and competitive product or service evaluations (Sharma and Rawani 2008). Although, HoQ is the primary tool in QFD method, the statements or demands of the customers may not always be clear or comprehensible, therefore, some other tools are also required which can interpret and explain the voice of customers clearly. Thus, along with HoQ, seven other management and planning tools are used to identify and prioritize customers’ expectations quickly Marketing/ Sales Quality assurance Design Purchase Vendor Production Testing QFD team Fig. 1 A QFD team J Ind Eng Int (2015) 11:575–594 579 123 and effectively. Figure 2 presents those seven management and planning tools, while the basic structure of HoQ matrix is shown in Fig. 3. HoQ translates the customers’ requirements, based on marketing research and benchmarking data, into an appropriate number of engi- neering targets to be met by a new product design. Basi- cally, HoQ is the nerve centre and the engine that drives the entire QFD process (Raissi et al. 2012). The procedural steps involved in the development of HoQ matrix are described as below: Step I In the first step, various market segments are determined and subsequently analysed to identify the potential customers. The QFD team then conducts customer surveys to accumulate the relevant information about customers’ requirements or expectations from the product/service. The seven management and quality tools, i.e. affinity diagrams, tree diagrams, relations diagrams, matrices and tables, process decision program charts, AHP, and blueprinting are employed to analyse and categorise those information as primary, sec- ondary and tertiary. The customers’ require- ments are then prioritized based on customers’ choice and its relative importance to them using a 1–5 rating scale, where 1 having the least priority and 5 having the maximum priority. These customers’ Affinity diagrams Matrices and tables Relations diagrams AHP Process decision program chart Hierarchy trees Blueprinting Management and planning tools Fig. 2 Management and planning tools C us to m er s’ r eq ui re m en ts Technical requirements Primary Secondary T er tia ry S ec on da ry Tertiary P rim ar y Interrelationship matrix Technical correlation matrix Planning matrix Prioritized technical requirements Fig. 3 House of quality matrix 580 J Ind Eng Int (2015) 11:575–594 123 requirements are placed on the left wall of HoQ matrix along the rows. Step II In this step, the key performance indicators in an organization to achieve customer satisfac- tion are recognized. Additionally, regulatory standards and technical requirements dictated by the management are also identified. Fur- ther, these technical requirements are orga- nized as primary, secondary and tertiary using different management and quality tools, and are arranged at the top of HoQ matrix along the columns. Step III The interrelationship matrix, which shows the relationship between customers’ requirements and technical requirements is now established and positioned at the centre of HoQ matrix. The relationship between pairs of customers’ requirement and technical requirement is portrayed using symbols or numbers, termed as correlation index, as shown in Table 2. Step IV The performance measures in the existing designs usually conflict with each other. The technical correlation matrix, which is more often referred to as roof matrix, shows the relationship between various technical requirements. This roof matrix also helps in establishing the interrelationship matrix, and identifies where the technical requirements must work together to avoid any design conflict. Step V Once the roof matrix is constructed, the planning matrix measuring the performance of the organization with respect to its bench- marked competitive organization is devel- oped. The performance of the organization is then ranked on a scale of 1–5, 1 being the least satisfying and 5 being the excellent performance. This matrix is set on the right side of the interrelationship matrix. Step VI Finally, the priorities assigned to the technical requirements are recorded and compared to those of the benchmarked competitor accord- ing to the relative weight of each relationship. Then, these values are linked back to the customers’ requirements to meet the new design requirements and are positioned at the bottom of HoQ matrix as prioritized technical requirements. Development of a QFD-based expert system for CNC turning centre selection The developed QFD-based expert system relates the dynamic requirements of the customers with technical specifications of CNC turning centres and then selects the most suitable machine based on the considered evaluation criteria. The basic framework for design and development of the QFD- based expert system is exhibited in Fig. 4. It has five basic modules, e.g. recognition of customers’ requirements, iden- tification of technical requirements, creation of the database, development of the expert system and evaluation of the alternatives to select the best CNC turning centre. Based on a market survey using questionnaires and customers’ feedback, the wants and needs of the customers related to CNC turning centres are first accumulated. These customers’ voices are then prioritized using different management and planning tools. The 12 most important customers’ requirements associated with the selection of CNC turning centres are detailed out as follows: (a) Allocated fund—It is associated with the initial acquisition cost and investment needed to procure and set up a CNC turning centre in a manufacturing organization. It also includes the expenditure made on installation of the said machine tool. An organi- zation’s objective is always to reduce the allocated fund and keep it as minimum as possible. (b) Availability of space—It is related to the total space occupied by a CNC turning centre with respect to its length, width and height dimensions. There are always some spatial constraints within the shop floor because of which machine tools having smaller overall dimensions are always preferred. (c) Capacity—This characteristic of a CNC turning centre deals with the maximum dimension, i.e. length, diameter and weight of the workpiece that can be machined. It is always better to have a CNC machining centre with higher capacity. (d) Productivity—It can be defined as the volume of workpieces machined by a CNC turning centre per Table 2 Correlation index for interrelationship matrix Number Symbol Relation 1 Very weak relation 3 Weak relation 5 Moderate relation 7 Strong relation 9 Very strong relation J Ind Eng Int (2015) 11:575–594 581 123 unit time. As productivity is directly proportional to a manufacturing organization’s profit, it is always preferred to have highly productive CNC turning centres. (e) Machining time—It is the time taken by a CNC turning centre to machine a workpiece having the desired dimensions and shape. Less machining time reduces the lead time to market a product. So, a CNC turning centre which has less machining time is more beneficial for a manufacturing organization. (f) Power requirement—It deals with the power rating of a CNC turning centre, i.e. amount of power required to operate it. It is always beneficial to have machines which consume less power. (g) Flexibility—Flexibility of a CNC turning centre relates to its ability to deal with the changing part configurations to allow variations in part assembly and process sequence, and changes in production volume and product design. A manufacturing orga- nization must be proficient to produce reasonably priced customized parts/components of higher dimensional accuracy that can be quickly delivered to the customers. (h) Ease of machine tool handling—The CNC machine tool handling and changing task should be easy as it requires frequent human intervention. Easy machine tool handling makes it simple to respond to any change in the production plan. (i) Auxiliary attachments—These are the additional attachments provided with CNC turning centres to enhance their overall machining performance. For example, guards are added to increase operator safety, automatic tool and job changers are often provided to reduce human interference and total machining time, and turret may be equipped with multiple tool holders so that a range of different machining operations can be performed simultaneously. (j) Surface finish—It relates to the dimensional accu- racy and smoothness of the workpiece surface machined by a CNC turning centre. This is also crucial from the aesthetics viewpoint of the product. QFD team discussion table Quality assurance Design Vendor Purchase Production Testing Marketing/ Sales QFD-BASED EXPERT SYSTEMWebsite Database Collection of data Identification of technical requireemnts Obtain customers’ requirements Retrieval of alternatives Evaluation of alternatives Best CNC turning centre Customers Decision maker Fig. 4 Basic framework for development of QFD-based expert system 582 J Ind Eng Int (2015) 11:575–594 123 (k) Generation of complicated parts—The CNC turning centres are able to accurately control motions in multiple directions, and hence, can generate complex shape contours on the workpieces with higher precision and accuracy. Therefore, a CNC turning centre which can easily machine complicated part geometries has an added advantage over the others. (l) Adaptability—It is associated with a CNC turning centre’s ability to adapt to new machining condi- tions. The CNC machines having this feature are favoured over the others to satisfy diverse customer requirements. Next, an expert panel comprising of members from various departments of a manufacturing organization, like purchasing, production, testing, quality assurance, mar- keting, design, etc. is interviewed to obtain their opinions regarding the technical specifications of CNC turning centres. After critically analysing those specifications, 13 technical requirements are finally shortlisted based on their impact on the selection procedure, as enlisted below: (a) Area (mm 2 )—This specification of a CNC turning centre denotes the total space occupied by it on the shop floor. It is an important consideration in case of any spatial constraint within the shop floor. (b) Cost—It relates to the capital investment required to procure and install a CNC turning centre. A CNC turning centre having less initial capital investment is always preferred to that requiring higher capital outflow. (c) Height (mm)—It is one of the dimensions of a CNC turning centre. It actually measures the spread of the machine in vertical direction. (d) Maximum length (mm)—It refers to the maximum length of a workpiece that can be accommodated and machined on a CNC turning centre. This specifica- tion is related to the capacity of a CNC turning centre. (e) Maximum diameter (mm)—It is related to the maximum diameter of a workpiece that can be machined on a CNC turning centre. Like maximum length, it is also related to the capacity of the machine. (f) Maximum spindle speed (min -1 )—Rotations of the spindle of a CNC turning centre per minute is its spindle speed. It can be correlated to overall productivity and attainable surface finish. (g) Travel X-axis (mm)—It signifies the maximum length that the tool can move in X-direction. (h) Travel Z-axis (mm)—It indicates the maximum length that the tool can travel in Z-direction. (i) Rapid traverse X-axis (m/min)—It expresses the movement of the tool turret at the fastest rate in X- axis, requiring only an end point for this movement. (j) Rapid traverse Z-axis (m/min)—Similarly, rapid traverse Z-axis is the movement of the tool turret at the fastest rate in Z-axis direction. (k) Spindle motor power (kW)—It is the power rating of a CNC turning centre. It is directly related to the consumption of electrical power while running the machine. (l) Number of tools—It is the maximum number of tools that can be accommodated in a tool holding device of a CNC turning centre. More number of tools facilitates in generating complicated shapes on the workpieces and also reduces the total machining time. (m) Weight (kg)—It is another specification of a CNC turning centre indicating its overall weight. Machine weight is quite critical when a load limit exists on the shop floor. Among all these technical specifications, cost of a CNC turning centre is expressed using a qualitative scale of 1–9. The actual range of cost (in USD) along with its scale values and interpretations is given in Table 3. The detailed information and relevant data regarding the technical specifications of various CNC turning centres are accumulated from the brochures of different manufacturers available online. Then, these collected data for CNC turning centres are stored into MS-Access option of Visual BASIC 6.0. An exhaustive database containing technical specifica- tions of more than 200 CNC turning centres is thus created. The next stage involves in the development of the QFD- based expert system, which can be broadly divided into two phases. At first, the related HoQ matrix is constructed and the feasible alternatives satisfying the set criteria values are extracted from the database. The HoQ matrix developed here is a simplified one where technical correlation and Table 3 Scale indicating range of cost for CNC turning centres Cost (in USD) Scale Interpretation 25,000–30,000 1 Lowest 30,001–45,000 2 Very very low 45,001–60,000 3 Very low 60,001–70,000 4 Low 70,001–85,000 5 Medium 85,001–105,000 6 High 105,001–130,000 7 Very high 130,001–155,000 8 Very very high 155,001–180,000 9 Highest J Ind Eng Int (2015) 11:575–594 583 123 planning matrices are not considered. Only the prioritized technical requirements are taken into account at the bottom of HoQ matrix. The customers’ requirements and technical requirements are already identified, and they are placed on the left wall of HoQ matrix along the rows and at the top of HoQ matrix along the columns, respectively. Once the customers’ requirements and technical requirements are placed in HoQ matrix, the corresponding interrelationship matrix is developed. Three production plans, i.e. ‘Flexible Production’, ‘Mass Production’ and ‘Custom’ are provided to the production planners to choose the kind of interrelationship matrix to be developed to accommodate the dynamic demands of the customers. Flexibility is the measure of how fast a setup can convert its process(es) from making an old line of products to a new set of items. A flexible production plan is often required to permit low cost switching from one product line to another. This type of production plan can efficiently produce highly customized and unique products in varying volumes while utilizing CNC technology. Its use can reduce the need for human intervention and provide an infrastructure which can react quickly to deviations in the production plan. Consistent and better quality products, with lower cost can be produced using the same manpower while adopting a flexible production plan. While, mass production is the manufacturing of large volumes of stan- dardized products, making many copies of the products very quickly, using assembly line techniques. It is often characterized by mechanization to achieve high volume output, efficient flow of materials through various stages of manufacturing, careful supervision of quality standards and minute division of labour. As there is a continuous flow of materials, there is no queuing at any stage of the production process. Supervision is also easy in case of mass produc- tion because only few instructions are necessary. If any of the first two options is chosen, i.e. ‘Flexible Production’ and ‘Mass Production’, an automatically filled up interrelationship matrix with default values appears. On the other hand, if ‘Custom’ option is selected, the end user needs to fill up the interrelationship matrix based on the subjective judgments. The customers’ requirements can be either beneficial (higher the better) or non-beneficial (lower the better), and are attributed by the value of the corre- sponding improvement driver (?1 for beneficial criteria and -1 for non-beneficial criteria). The next stage com- prises of assigning priority weights to the requirements of the customers. For assigning priority values to customers’ requirements, a scale of 1–5 is set, where 1—not important, 2—important, 3—much more important, 4—very impor- tant and 5—most important. After critically analyzing the relationship between customers’ requirements and techni- cal specifications of CNC turning centres, it is observed that productivity is highly correlated to spindle speed, motor power and rapid traverse speed; whereas, area, height and weight of CNC turning centre have the least relationship with it. Similarly, it is also found that flexi- bility is strongly related to spindle speed and number of tools, while availability of space is highly interrelated with area and height. Furthermore, it is revealed that the allo- cated fund is greatly associated with cost of the CNC turning centre; whereas, machining time is strongly related to spindle speed, and rapid traverse in X- and Z-axes. Moreover, it is also noticed that power requirement is positively related to spindle motor power, but it is least influenced by maximum length and diameter of the work- piece, and travel in X- and Z-axes. Additionally, it is observed that number of tools considerably influences the capability of a CNC turning centre to generate complicated parts. The interrelationships between the remaining cus- tomers’ requirements and technical requirements are sub- sequently developed similarly with values from an appropriate scale of 1–9, where 1—very very weak, 2— very weak 3—weaker, 4—weak, 5—moderate, 6—strong, 7—stronger, 8—very strong and 9—very very strong. Once the HoQ matrix is filled up with all the necessary data, the ‘Weight’ functional key is pressed to obtain the priority weights of all the technical requirements, using the following equation: wj ¼ Xn i ¼ 1 Pri � IDi � correlation index ð1Þ where wj is the weight for jth technical requirement, n is the number of customers’ requirements, IDi is the value of improvement driver for ith customer requirement, Pri is the priority assigned to ith customer requirement and correla- tion index is the relative importance of jth technical requirement with respect to ith customer requirement. These weights are subsequently used for calculation of the performance scores of the feasible CNC turning centres. The second phase of QFD-based expert system embarks with identifying the most important technical requirements based on which the end user wants to evaluate the candi- date CNC turning centres. Once the desirable technical requirements are shortlisted by the end user, the range for each selected technical requirement needs to be specified. Depending on the given ranges of values for the shortlisted technical requirements, a list of feasible CNC turning centres is then extracted. The final stage of the expert system consists of evalu- ating the feasible candidate alternatives to select the best CNC turning centre. In this stage, a final set of CNC turning centres that needs to be evaluated based on the set criteria is chosen from the list of feasible alternatives. A decision matrix comprising of the selected technical 584 J Ind Eng Int (2015) 11:575–594 123 specifications with respect to each chosen CNC turning centre is then developed. This decision matrix now needs to be normalized to make it dimensionless so that the performance of all the alternatives can be compared with respect to the set criteria. The following linear normaliza- tion procedure is adopted here. For beneficial criteria (technical requirements for which weights are positive) Normalized value ¼ Property value � Smallest value Highest value � Smallest value ð2Þ For non-beneficial criteria (technical requirements for which weights are negative) Normalized value ¼ Smallest value � Property value Highest value � Smallest value þ 1 ð3Þ The performance score for each CNC turning centre is now computed using the following equation: Performance score PSið Þ ¼ Xn j¼1 wj � Normalized valueð Þij i ¼ 1; 2; . . .; m; j ¼ 1; 2; . . .; nð Þ ð4Þ where m is the number of alternatives and n is the number of technical requirements. The weights for the selected technical requirements are automatically retrieved from the HoQ matrix. Based on these performance scores, the fea- sible CNC turning centres are ranked and a graphical representation showing the performance score of each CNC turning centre is automatically generated. The best per- forming CNC turning centre is finally identified and its retailed technical specifications are displayed along with its actual photograph. Figure 5 exhibits a flowchart for the developed QFD- based expert system to help the end user to navigate through it properly. The procedural steps for running this software prototype are enlisted as below: Step I An opening window containing the guidelines to be followed by the end user for running the QFD-based expert system appears at first. Step II Once the guidelines are understood, the end user selects an option from the type of production plans and presses the ‘Next’ key. A HoQ matrix depending on the type of the production plan appears in a new window. Step III The customers’ requirements are identified as beneficial or non-beneficial while assign- ing appropriate improvement driver values. Step IV A priority value between 1 and 5 is assigned to each customer requirement. Step V Based on the type of production plan, an interrelationship matrix, either filled up or blank, appears. If it is blank, it needs to be filled up with necessary values, else proceed to the next step. Step VI The ‘Weight’ functional key is pressed to obtain the priority weights of all the technical requirements in the HoQ matrix. Step VII A set of evaluation criteria is chosen from the list of available technical requirements to finalize the selection decision. Step VIII The ‘Input Range’ functional key is pressed to generate empty cells to capture ranges of the selected criteria values within which the spec- ifications of CNC turning centres should lie. Step IX All the feasible alternatives satisfying the given ranges of specifications are extracted by pressing ‘Feasible Alternatives’ key. Step X The end user then shortlists the final set of alternatives to be evaluated. Step XI The ‘Next’ key is pressed to automatically develop the corresponding decision matrix with the technical specifications of the finally selected alternatives with respect to the set evaluation criteria in a new window. Step XII The performance scores and ranks of the finally selected alternatives are computed after pressing the ‘Calculate Rank’ key. Step XIII The ‘Rank Analysis’ key is pressed to graph- ically display the performance score of each alternative. Step XIV The ‘Machine Details’ key is pressed to display the detailed technical specifications of the most appropriate CNC turning centre along with its actual photograph. The interrelationship matrices for flexible and mass production plans are slightly different from each other. For example, presence of auxiliary attachments in a CNC turning centre makes it more flexible. This affects the overall cost of the machine more strongly in flexible pro- duction than mass production. In flexible production plan, the correlation indices between flexibility and all other technical requirements are high due to greater impact of these requirements on flexibility. Moreover, a large volume of finished products is required in mass production and hence, if the number of tools is more in a CNC turning centre, its productivity will also be high. Therefore, there exists a stronger relationship between productivity and number of tools in mass production than flexible J Ind Eng Int (2015) 11:575–594 585 123 Start the selection process Run the application Go through the guidelines in opening window Select the type of production plan from ‘Type of production system’ module and press ‘Next’ key Assign improvement driver to each customer’s requirement Enter priority values for customers’ requirements Press ‘Weight’ key to obtain the weights of technical requirements Choose the selection criteria from the list of technical requirements Press ‘Input Range’ key and enter ranges for the selected criteria Fill up the interrelationship matrix as per the guidelines Is interrelationship matrix is filled? Press ‘Feasible Alternatives’ key to list all the feasible alternatives Choose final set of alternatives after comparing all the feasible alternatives Press ‘Next’ key to develop decision matrix with technical specifications of the finally selected alternatives Press ‘Calculate Rank’ key to compute performance scores and ranks of the alternatives Press ‘Rank Analysis’ key to display the graph with performance scores Press ‘Machine Details’ key to get details of the best alternative with a photograph End of the selection process Yes No Fig. 5 Flowchart of the QFD-based expert system 586 J Ind Eng Int (2015) 11:575–594 123 production. Similarly, the capacity of a CNC turning centre for mass production should be such that it can accommo- date more number of tools so that several machining operations can be performed simultaneously. Hence, capacity of the machine and number of tools are more strongly related in mass production than flexible production. Illustrative examples An Intel � Core TM i5-2034M CPU with 2.50 GHz, 4.00 GB RAM operating platform is required to run this QFD-based expert system. Three illustrative examples are provided to demonstrate its applicability. Figure 6 displays the instruction sheet, which first appears on the screen when this software prototype is run to assist the end user to get acquainted with it. Example 1 In this example, the end user desires to select an appro- priate CNC turning centre for a manufacturing system which is subjected to change in the type and volume of parts produced. A flexible production plan can only offer a variety of parts/products with rapidly changing production level as demanded by the customers. Taking this feature of flexible production into account, ‘Flexible Production’ option is chosen from ‘Type of Production System’ mod- ule. On pressing the ‘Next’ key, the corresponding HoQ matrix with the filled up interrelationship matrix according to flexible production plan now appears in a new window, as shown in Fig. 7. The improvement driver and priority values for each customer’s requirement are then entered in their respective columns. The improvement driver value of -1 for ‘Allocated fund’, ‘Availability of space’, ‘Ma- chining time’, ‘Power requirement’ and ‘Surface finish’ reveal that they are the non-beneficial attributes requiring minimum values. A priority value of 5 assigned to ‘Allo- cated fund’, ‘Availability of space’ and ‘Flexibility’ sig- nifies that these customers’ requirements have the highest importance. On the other hand, a priority value of 1 for ‘Ease of machine tool handling’ and ‘Auxiliary attach- ments’ shows that they are not so much important to the customers in this case. The ‘Weight’ key is then pressed to derive the priority weight of each technical requirement present in the HoQ matrix. Next, depending on the end requirements of the final product, cost, maximum diameter, maximum length, spindle motor power and weight are shortlisted to finalize the CNC turning centre selection decision. In the HoQ matrix, among the five criteria chosen for evaluation of CNC turning centres, negative priority weights for cost, spindle motor power and weight identify Fig. 6 Opening window having guidelines for the end user J Ind Eng Int (2015) 11:575–594 587 123 them as non-beneficial technical requirements. Now, pressing of the ‘Input Range’ key generates the necessary empty cells, where the ranges of values for the selected technical requirements are entered. Then, the functional key ‘Feasible Alternatives’ is then pressed to display a list of feasible CNC turning centres satisfying all the set cri- teria values. A final set of 12 alternatives to be evaluated is chosen from this list of feasible alternatives for ultimate selection of the most appropriate CNC turning centre for the given application. In Fig. 8, the final decision matrix for the CNC turning centre selection problem is developed from the database on pressing the ‘Next’ functional key. Pressing of the ‘Cal- culate Rank’ key then normalizes the decision matrix, computes the performance scores and displays the ranking preorder of the considered CNC turning centres. Addi- tionally, their relative performances are graphically dis- played when the end user presses the ‘Rank Analysis’ button. Finally, the ‘Machine Details’ key is pressed to retrieve the detailed technical specifications and an actual photograph of the best suited CNC turning centre. Here, model ST20 manufactured by Haas Automation Inc. is identified as the best CNC turning centre, followed by GENOS L 300E-MY. On the other hand, LOC-500 is indicated as the least preferable choice for the given application. It can be noticed from the specifications of model ST20 that it can simultaneously satisfy all the shortlisted technical requirements, i.e. its cost, weight and spindle motor power are low; and maximum diameter and maximum length are high. Example 2 In this example, a CNC turning centre for producing a large volume of standardized parts/components needs to be selected from a wide range of available alternatives. Mass production is the suitable approach for producing parts in large volume at low unit cost. Keeping this in mind, the end user selects ‘Mass Production’ option from ‘Type of Pro- duction System’ module. A HoQ matrix with already filled up interrelationship data for mass production plan now appears, as shown in Fig. 9. Similar to the earlier example, after identifying the beneficial and non-beneficial attri- butes, and entering the priority for each customer’s requirement, the ‘Weight’ button is pressed to obtain the priority weights for the technical requirements. A priority Fig. 7 HoQ matrix for flexible production plan 588 J Ind Eng Int (2015) 11:575–594 123 value of 5 assigned to ‘Allocated fund’ shows that it is the most important criterion for the customers. ‘Productivity’ and ‘Generation of complicated parts’ are also important for this particular CNC turning centre selection problem and hence, a priority value of 4 is assigned to them. In this case, ‘Availability of space’ and ‘Power requirement’ is not the constraints for the customers; therefore, they are allotted a priority value of 1 signifying their least impor- tance. Four technical requirements, i.e. cost, rapid traverse X-axis, travel X-axis and maximum spindle speed are identified here as the governing requirements. A negative priority weight of cost implies that it is a non-beneficial criterion, whereas, for the other three criteria, it is always better to have their higher values. Then, the ranges of values for these selected criteria are entered in their respective cells and the functional key ‘Feasible Alterna- tives’ is pressed to retrieve all the feasible alternatives fulfilling the set criteria values. From the list of feasible alternatives, the end user selects a set of nine alternatives for final assessment. The decision matrix for this problem with all the selected technical specifications for the finally chosen alternatives is shown in Fig. 10. In the next step, their performance scores and ranking preorder are determined after pressing the ‘Calculate Rank’ key. MACTURN 350-W manufactured by Okuma Corporation is identified as the most appropriate choice for this application, while LB-35II (M) 850-C emerges out as the least suitable CNC turning centre under the given conditions. To retrieve all the technical details of MACTURN 350-W along with its photograph, the ‘Machine Details’ key is then pressed. A close review of the details of MACTURN 350-W reveals that although its cost is slightly high (85001-105000 USD), but it has comparatively higher values for the other three technical parameters, i.e. maximum spindle speed, rapid traverse X-axis and travel X-axis. Hence, its selection can be justified with respect to the considered criteria. Example 3 This example illustrates the selection of a CNC turning centre when the end user has the flexibility to fill up the interrelationship matrix based on the association between customers’ requirements and technical requirements. After selecting the ‘Custom’ option, when the end user presses the ‘Next’ key, a blank HoQ matrix is automatically Fig. 8 Output of the expert system for flexible production plan J Ind Eng Int (2015) 11:575–594 589 123 generated. After identifying the appropriate beneficial and non-beneficial attributes, ?1 or -1 value is assigned to them. For this example, ‘Allocated fund’ and ‘Capacity’ are the most important customers’ requirements, having a priority value of 5 assigned to them. ‘Availability of space’ and ‘Productivity’ have been assigned a priority value of 4. On the other hand, a priority value of 1 implying least importance to the customers is allotted to ‘Ease of machine tool handling’ and ‘Adaptability’. Figure 11 exhibits the HoQ matrix filled up based on the opinions and judgments of the end user showing the relationship between cus- tomers’ requirements and technical requirements. The weight of each technical requirement is then derived. The end user now shortlists area, cost, height, maximum diameter and maximum spindle speed from the technical requirements list, based on which the CNC turning centres are to be evaluated. In the next step, the ‘Feasible Alter- natives’ key is pressed to obtain a list of candidate alter- natives satisfying all the set criteria values. A decision matrix containing the selected technical specifications for the final set of alternatives is developed in a new window, as displayed in Fig. 12, when the user presses the ‘Next’ key. This QFD-based expert system identifies MACTURN 250 manufactured by Okuma Corp. as the best CNC turning centre, whereas, GENOS L 300-W is the least suitable choice for the given application. It is observed that the performance scores of MACTURN 250 and MACTURN 550-W are almost same. A closer look at the technical specifications of these two machines reveals that the former has lower values for area, cost and height which are the non-beneficial attributes. The maximum spindle speed for MACTURN 250 is also higher, which is a beneficial criterion. Therefore, its selection as the best turning centre is justified. Real-time implementation of the developed expert system A specific machine shop of a manufacturing organization is considered here for real-time application of the developed expert system. The identity of this organization is not disclosed for confidentiality and anonymity purpose, and hereafter, it is referred to as XYZ Limited. It is noticed that the said organization used to practice the conventional (manual) technique for selection of CNC turning centres for various machining applications. This manual approach of short listing the appropriate CNC turning centres for a specific application is often a costly and time-consuming task due to various activities involved, like creating and receiving records, record keeping and maintenance, retrieving data for comparison and identification of a suitable CNC turning centre, and disaster prevention and recovery planning of vital records. Those activities require scarce resources, like manpower, office space, operating Fig. 9 HoQ matrix for mass production plan 590 J Ind Eng Int (2015) 11:575–594 123 Fig. 10 Output of the expert system for mass production plan Fig. 11 HoQ matrix for custom option plan J Ind Eng Int (2015) 11:575–594 591 123 supplies, etc. But, it is observed that once this expert sys- tem is implemented in the said organization for the selec- tion of CNC turning centre, many of the activities utilized earlier in the manual method are either removed from the selection process or required in meagre amount. This implies that with the implementation of the developed expert system, there is a considerable reduction in man- power and time for upkeepng of records, and smaller office space is required due to comparatively less number of documents. Additionally, in the current automated envi- ronment, fewer operating supplies are consumed because of computerization of the selection process. A cost-benefit analysis is carried out to find out the monetary benefits of implementing this expert system in the considered orga- nization, which is estimated to be in the range of 3600–4000 USD per month. This derived cost-benefit can be attributed to savings under various cost heads due to automation of the selection procedure, like 3000–3100 USD per month is spared because of reduction in man- power requirement, 400–500 USD per month is saved due to reduced office space utilization, 100–200 USD per month is cut down on account of decreased consumption of operating supplies and 100–200 USD per month is cur- tailed owing to cloud computing-based disaster planning. Although the development and subsequent implementation of the expert system in the said organization require an initial investment, it is often offset by the derived saving. Moreover, once it is successfully installed, there is no recurring maintenance cost. It is also noticed that managers of the said organization can derive a plethora of benefits from application of this expert system, such as (a) reduc- tion in costs related to the selection process, (b) better utilization of workforce associated with decreased number of employees and increased personnel efficiency, (c) en- hancement of managerial efficiency with better decision- making, (d) effective utilization of information in the organizational environment, (e) strengthening of manage- ment information system for timeliness of CNC turning centre selection decision, and (f) accuracy in the decision- making process leading to higher production rate. Conclusions In today’s competitive manufacturing environment, the need for quality products is of utmost importance. There- fore, achieving cost-efficient, consistent machining results through automation is attractive to the manufacturing Fig. 12 Output of the expert system for ‘Custom’ option 592 J Ind Eng Int (2015) 11:575–594 123 organizations. In this paper, an expert system based on QFD methodology is designed and developed to help the process planners in CNC turning centre selection, where a large number of available alternatives need to be evaluated based on several conflicting criteria. It eases out and automates the entire selection procedure, while eliminating rigorous calculations involved, thereby reducing the time taken to arrive at the best decisive action. The level of human intervention is also minimized and the end users do not need to have an in-depth technical knowledge regard- ing various CNC machine tools. It ensures an error-free computation of CNC turning centre selection decision. Moreover, as discussed earlier, it encompasses both cus- tomers’ requirements and technical requirements in the selection procedure, thus providing the end users with a competitive edge over the previously adopted methodolo- gies. It can also work in a group decision-making envi- ronment where opinions from different individuals can be sought. It can be employed for selection of CNC turning centre for any type of production system depending on the requirements of the end users. The database for CNC turning centres can be upgraded from time to time to make it more dynamic. It can also be applied for selection of various other CNC machine tools, such as milling, grind- ing, etc. while creating a new module and database within the same system. Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://crea tivecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. References Aghdaie MH, Zolfani SH, Zavadskas EK (2013) Decision making in machine tool selection: an integrated approach with SWARA and COPRAS-G methods. Inz Ekon Eng Econ 4(1):5–17 Akao Y (1990) Quality function deployment. Productivity Press, Cambridge Ayağ Z (2007) A hybrid approach to machine-tool selection through AHP and simulation. Int J Prod Res 45(9):2029–2050 Ayağ Z, Özdemir RG (2006) A fuzzy AHP approach to evaluating machine tool alternatives. J Intell Manuf 17(2):179–190 Ayağ Z, Özdemir RG (2011) An intelligent approach to machining center selection through fuzzy analytic network process. J Intell Manuf 22(2):163–177 Ayağ Z, Özdemir RG (2012) Evaluating machine tool alternatives through modified TOPSIS and alpha-cut based fuzzy ANP. Int J Prod Econ 140(2):630–636 Bevilacquaa M, Ciarapicab FE, Giacchetta G (2006) A fuzzy-QFD approach to supplier selection. J Purch Supply Manag 12(1):14–27 Bhattacharya A, Geraghty J, Young P (2010) Supplier selection paradigm: an integrated hierarchical QFD methodology under multiple-criteria environment. Appl Soft Comput 10(4):1013–1027 Chakroborty S, Dey S (2007) QFD-based expert system for non- traditional machining process selection. Expert Syst Appl 32(4):1208–1217 Chan L-K, Wu M-L (2002) Quality function deployment: a literature review. Eur J Oper Res 143(3):463–497 Dağdeviren M (2008) Decision making in equipment selection: an integrated approach with AHP and PROMETHEE. J Intell Manuf 19(4):397–406 Dai J, Blackhurst J (2012) A four-phase AHP-QFD approach for supplier assessment: a sustainability perspective. Int J Prod Res 50(19):5474–5490 Dawal SZM, Yusoff N, Nguyen H-T, Aoyama H (2013) Multi- attribute decision-making for CNC machine tool selection in FMC based on the integration of the improved consistent fuzzy AHP and TOPSIS. ASEAN Eng J Part A 3(2):15–31 Durán O, Aguilo J (2008) Computer-aided machine tool selection based on a fuzzy-AHP approach. Expert Syst Appl 34(3):1787–1794 Dursan M, Karsak EE (2013) A QFD-based fuzzy MCDM approach for supplier selection. Appl Math Model 37(8):5864–5875 İç YT (2012) An experimental design approach using TOPSIS method for the selection of computer-integrated manufacturing technologies. Robot Comput-Integr Manuf 28(2):245–256 İç YT, Yurdakul M (2009) Development of a decision support system for machining center selection. Expert SystAppl 36(2):3505–3513 İç YT, Yurdakul M, Eraslan E (2012) Development of a component- based machining centre selection model using AHP. Int J Prod Res 50(22):6489–6498 Ilangkumaran M, Sasirekha V, Anojkumar L, Boopathi Raja M (2012) Machine tool selection using AHP and VIKOR method- ologies under fuzzy environment. Int J Model Oper Manag 2(4):409–436 Karsak EE (2008) Robot selection using an integrated approach based on quality function deployment and fuzzy regression. Int J Prod Res 46(3):723–738 Liu H-T (2011) Product design and selection using fuzzy QFD and fuzzy MCDM approaches. Appl Math Model 35(1):482–496 Mayyas A, Shen Q, Mayyas A, Abdelhamid M, Shan D, Qattawi A, Omar M (2011) Using quality function deployment and analyt- ical hierarchy process for material selection of body-in-white. Mater Des 32(5):2771–2782 Nguyen H-T, Dawal SZM, Nukman Y, Aoyama H (2014) A hybrid approach for fuzzy multi-attribute decision making in machine tool selection with consideration of the interactions of attributes. Expert Syst Appl 41(6):3078–3090 Önüt S, Kara SS, Efendigil T (2008) A hybrid fuzzy MCDM approach to machine tool selection. J Intell Manuf 19(4):443–453 Özgen A, Tuzkaya G, Tuzkaya UR, Özgen D (2011) A multi-criteria decision making approach for machine tool selection problem in a fuzzy environment. Int J Comput Intell Syst 4(4):431–445 Partovi FY, Corredoira RA (2002) Quality function deployment for the good of soccer. Eur J Oper Res 137(3):642–656 Prasad K, Chakraborty S (2013) A quality function deployment-based model for materials selection. Mater Des 49:525–535 Qi J (2010) Machine tool selection model based on fuzzy MCDM approach. In: Proceedings of International conference on intel- ligent control and information processing, Dalian, China, pp 282–285 Raissi S, Izadi M, Saati S (2012) Prioritizing engineering character- istic in QFD using fuzzy common set of weight. Am J Sci Res 49:34–49 Sahu AK, Datta S, Mahapatra SS (2015) GDMP for CNC machine tool selection with a compromise ranking method using J Ind Eng Int (2015) 11:575–594 593 123 http://creativecommons.org/licenses/by/4.0/ http://creativecommons.org/licenses/by/4.0/ generalised fuzzy circumstances. Int J Comput Aided Eng Technol 7(1):92–108 Samvedi A, Jain V, Chan FTS (2012) An integrated approach for machine tool selection using fuzzy analytical hierarchy process and grey relational analysis. Int J Prod Res 50(12):3211–3221 Shad Z, Roghanian E, Mojibian F (2014) Integration of QFD, AHP, and LPP methods in supplier development problems under uncertainty. J Ind Eng Int 10(1):9. doi:10.1007/s40092-014- 0051-0 Sharma JR, Rawani AM (2008) Quality function deployment: integrating comprehensive matrix and SWOT analysis for effective decision making. J Ind Eng Int 4(6):19–31 Soota T, Singh H, Mishra RC (2011) Fostering product development using combination of QFD and ANP: a case study. J Ind Eng Int 7(14):29–40 Sun S (2002) Assessing computer numerical control machines using data envelopment analysis. Int J Prod Res 40(9):2011–2039 Taha Z, Rostam S (2012) A hybrid fuzzy AHP-PROMETHEE decision support system for machine tool selection in flexible manufacturing cell. J Intell Manuf 23(6):2037–2049 Tho NH, Dawal SZM, Yusoff N, Tahriri Aoyama FH (2013) Selecting a CNC machine tool using the intuitionistic fuzzy TOPSIS approach for FMC. Appl Mech Mater 315:196–205 Xin Y, Tian X, Huang L (2014) Optimal machine tools selection using interval-valued data FCM clustering algorithm. Math Probl Eng. doi:10.1155/2014/921647 (Article ID 921647, 10 pages) Yurdakul M (2004) AHP as a strategic decision-making tool to justify machining center selection. J Mater Process Technol 146(3):365–376 Yurdakul M, İç YT (2009) Analysis of the benefit generated by using fuzzy numbers in a TOPSIS model developed for machine tool selection problems. J Mater Process Technol 209(1):310–317 594 J Ind Eng Int (2015) 11:575–594 123 http://dx.doi.org/10.1007/s40092-014-0051-0 http://dx.doi.org/10.1007/s40092-014-0051-0 http://dx.doi.org/10.1155/2014/921647 Development of a QFD-based expert system for CNC turning centre selection Abstract Introduction Literature review in machine tool selection QFD methodology Development of a QFD-based expert system for CNC turning centre selection Illustrative examples Example 1 Example 2 Example 3 Real-time implementation of the developed expert system Conclusions Open Access References 
work_ueiooy5b2nbznmxdh3opjkske4	----	The validation of Expert System Traffic psychological assessment to Romanian Driving Schools Procedia - Social and Behavioral Sciences 30 (2011) 457 – 464 Available online at www.sciencedirect.com doi:10.1016/j.sbspro.2011.10.090 Procedia Social and Behavioral Sciences Procedia - Social and Behavioral Sciences 00 (2011) 000–000 www.elsevier.com/locate/procedia WCPCG-2011 The validation of Expert System Traffic psychological assessment to Romanian Driving Schools Mihai Aniţeia*, Mihaela Chraifb, Gernort Schuhfriedc, Markus Sommerd a Professor, PhD, University of Bucharest, Faculty of Psychology and Educational Sciences / Bd. M. Kogalniceanu 050107, Bucharest, Romania bPostdoctoral fellow, University of Bucharest, Faculty of Psychology and Educational Sciences, Bd. M. Kogalniceanu 050107, Bucharest, Romania c Schuhfried GmbH,Vienna, Austria dAsistent Professor, PhD, University of Vienna, Vienna, Austria Abstract Analyzing the multiple regression model for the composite criterion, the multiple correlation coefficient, evidence a high and statistically significant correlation between the predictors and the criterion (r=0.741, p<0.05). Also, the beta coefficients provide that the variables of the tests are predictors for the performances registered in traffic (p<0.05). This study based on the findings of the previous research highlight that the Romanian driving schools should improve the psychological assessment batteries with modern and validated instruments. The predictive regression validation model emphasizes the importance of using high performance statistical programs in choosing the psychological tests for evaluation. © 2011 Published by Elsevier Ltd. Keywords: criterion validation, psychological assessment battery, time reaction, tachistoscope test, driving behavior.Introduction 1. Theoretical framework In traffic psychology psychologists are mainly interested in predicting driving behavior using abilities tests (Schuhfried, Sommer, Aniţei & Chraif, 2010; Sommer, Schuhfried, Aniţei & Chraif, 2010) abilities and personality tests (Ulleberg & Rundmo, 2003; Herle, 2009) and autoperception of driving behaviour and risk tendencies (Dula & Ballard, 2003; Aniţei, Chraif, Niculae & Vancea, 2009). Many previous studies in traffic psychological assessment highlight the use of standardized tests in predictive validation of driving behavior (Ulleberg and Rundmo, 2003; Karner & Neuwirth 2001; Sommer, Arendasy, Olbrich & Schuhfried, 2004) cited by Risser, Chaloupka, Grundler, Sommer, Häusler & Kaufmann, 2008). Other researchers highlighted the objections a disadvantageous method because of low correlation coefficients obtained in recent validation studies as a consequence of the lack of correspondence between the generality of the predictor variable and the criterion measure (Sommer, Herle, Häusler, Risser, Schützhofer & Chaloupka, 2008). Thus, many authors, underline classical methods of statistical judgment formation (discriminatory analysis, regression analysis) could be vulnerable and offering less precision (Bortz, * Aniţei Mihai. Tel.: +040721232207 E-mail address: anitei_mihai@yahoo.com 1877-0428 © 2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of the 2nd World Conference on Psychology, Counselling and Guidance. Open access under CC BY-NC-ND license. © 2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of the 2nd World Conference on Psychology, Counselling and Guidance. Open access under CC BY-NC-ND license. http://creativecommons.org/licenses/by-nc-nd/3.0/ http://creativecommons.org/licenses/by-nc-nd/3.0/ 458 Mihai Anitei et al. / Procedia - Social and Behavioral Sciences 30 (2011) 457 – 464 Mihai Aniţei/ Procedia – Social and Behavioral Sciences 00 (2011) 000–000 2 1999; Brown and Wicker, 2000). Consequently, many methods of validation have been developed. Sommer & Häusler (2005) emphasized the artificial neural network validation as robust methods for pattern recognition tasks with little prerequisites to data characteristics (Bishop, 1995; Rojas, 2000; Warner & Misra, 1996). Risser, Chaloupka, Grundler, Sommer, Häusler & Kaufmann (2008) presented the validation of two tests batteries of the Expert System Traffic using standardized driving tests artificial neural network in calculating the criterion validity. Other methods of validation are based on simulations and animations. The traffic flow model is a discrete, stochastic, time step based microscopic model, with driver-vehicle-units as single entities. The model contains a psycho-physical car following a model for longitudinal vehicle movement and a rule-based algorithm for lateral movements (Fellendorf & Vortisch, 2001). The model is based on the continued work of (Wiedemann, 1974; Wiedemann, 1991). The authors emphasize that both microscopic calibration and macroscopic validation results show that simulation tools based on the psycho-physical car-following model can reproduce traffic flow very realistically under different real-world conditions. Therefore, it is possible but also necessary to adapt the model to the local traffic situation; at least national traffic regulations and driving styles must be taken into account. Adaptation of the model can be based on microscopic data gathered by probe vehicles equipped with electronic sensors or on macroscopic flow data, as it is normally available from measurement sites. Van Tongeren, Gietelink, De Schutter, & Verhaegen (2007) presented a microscopic traffic model for the validation of advanced driver assistance systems. The model describes single-lane traffic and is calibrated with data from a field operational test. The authors used the Monte Carlo simulation of single-lane traffic scenarios with application to cooperative adaptive cruise control system. Another model of validation using simulation is presented by (Pérez Arias, Kretz, Ehrhardt, Hengst, Vortisch & Hanebeck, 2009). This paper presents a novel framework for combining traffic simulations and extended range telepresence which create the impression of being present in a remote environment. The authors showed that real user's position data can thus be used for validation and calibration of models of pedestrian dynamics, while the user experiences a high degree of immersion by interacting with agents in realistic simulations. Starting from previous experimental studies regarding the validation of psychological assessments for driving schools in Romania using both paper and pencil tests and computer designed tests from Vienna Tests System (Aniţei, Mincu & Chraif, 2008; Anitei & Chraif, 2008; Schuhfried, Sommer, Aniţei & Chraif, 2010; Sommer, Schuhfried, Aniţei & Chraif, 2010) and from the validation methodology required at the Methodology Commission from the Romanian College of Psychologists we provide in this paper a model of criterion validation for the Expert System Traffic psychological assessment on Romanian population. 2. The objectives and hypothesis 2.1 The objective The objective is focused on the validation of the Expert System Traffic psychological assessment taking in into consideration the cross-cultural study (Aniţei, Schuhfried, Chraif, Giehl & Buzea, 2010) to predict the driving behavior of the potential Romanian drivers in traffic. 2.2 The hypothesis The independent variables of The Reaction Test, The Determination test, The Tachistoscope test, The Cognitrone test and The Adaptive Matrix test are predictor variables for driving performances in traffic. 459Mihai Anitei et al. / Procedia - Social and Behavioral Sciences 30 (2011) 457 – 464 Mihai Aniţei/ Procedia – Social and Behavioral Sciences 00 (2011) 000–000 3 3. The Method 3.1. The participants The participants were 352 male and female, aged between 18 and 71 years (M=38.15; S.D.= 11.74) from a few driving schools from in Romania, rural and urban areas, different levels of education. 3.2. The instruments  The Adaptive Matrices Test (figure 1) (Schuhfried, 2007). The test shows that the association between the test length and the measurement precision is optimized. This psychological test is a non-verbal test for assessing general intelligence as revealed in the ability to think inductively. The main area of application is organizational psychology, traffic psychology, aviation psychology, educational psychology. The items are somewhat similar to the classical matrices, but in contrast they are constructed on the basis of explicit psycho-logically-based principles involving detailed analysis of the cognitive processes used in solving problems of this type. The items were analyzed using the Rasch dichotomous probabilistic test model and the corresponding characteristic values were estimated for the items [26]. Figure 1. The Adaptative Matrices Test (Schuhfried, 2007)  The Reaction Test (figure 2) (Schuhfried, 2007). This test measures reaction time both as a simple choice and a multiple choice reaction. Yellow light stimulus modalities are available in the test battery, so that different stimulus constellations for the measurement of reaction time can be created. These can range from individual stimuli to simultaneous or sequentially presented stimulus combinations. Figure 2. Time reaction to yellow (Schuhfried, 2007) The use of a rest key and a reaction key makes it possible to distinguish between reaction and motor time. 460 Mihai Anitei et al. / Procedia - Social and Behavioral Sciences 30 (2011) 457 – 464 Mihai Aniţei/ Procedia – Social and Behavioral Sciences 00 (2011) 000–000 4  The Determination test (DT) (figure 3)(Schuhfried, 2007).The test is used to measure time reaction to different visual and auditory stimuli, stress tolerance and the associated ability to react. The test requires the respondent to use his cognitive skills to distinguish different colors and sounds, to memorize the relevant characteristics of stimulus configurations, response buttons and assignment rules and to select the relevant responses according to the assignment rules laid shown in the instructions and learned in the course during the test. Figure 3. The Determination test (Schuhfried, 2007) The difficulty of the DT arises from the need to sustain continuous, rapid and varying responses to rapidly changing stimuli.  Tachistoscopic test (ATAVT) (figure 4 ) [25] (Schuhfried, 2007). The ATAVT tests observational ability by presenting images of traffic situations. The items are constructed using an explicit, theory-led rationale which is based on detailed analysis of the cognitive processes involved in solving the test. Figure 4. The Tachistoscopic test (Schuhfried, 2007) The design of the ATAVT is based on the principles used in the well-known TAVTMB test. The ATAVT test is based on TAVTMB test but it takes into account recent research and findings on the perception of scenes and objects.  The Cognitrone test (figure 5) [25]. The test measure attention and concentration with the support given by the validity of the Rasch model. Numerous validation studies prove the construct and the criterion of validity. Depending on the test form, it can be applied in the field of attention assessment and concentration through the comparison of figures concerning their congruence. 461Mihai Anitei et al. / Procedia - Social and Behavioral Sciences 30 (2011) 457 – 464 Mihai Aniţei/ Procedia – Social and Behavioral Sciences 00 (2011) 000–000 5 Figure 5. The Cognitrone test (Schuhfried, 2007) The test can be applied in the following areas: traffic psychology, aviation psychology, clinical and health psychology, neuropsychology, organizational psychology, sport psychology.  The questionnaire for evaluating a driver’s performance (CEPCA 2008) The questionnaire was designed using as a model a description for evaluating the driver’s performances while driving on route. This evaluation description has been used in the driving school Ilioara during applying the test battery together with the CEPCA questionnaire. The questionnaire for the evaluation of the drivers’ performance while in traffic, named after the initials CEPCA as it has designed and validated as a converged model [27], [28] at the driving school Ilioara, has ten-item Likert scale from 1 (strong disagreement) to 5 (strong agreement). Applied as a test pilot questionnaire, the value of Cronbach Alpha obtained was of 0.736 for the 10 items and a correlation coefficient of 0.637 (p>0.01) by evaluating each individual examined with both questionnaires. Consequently, a pilot study was published previous to validation [23]. 4. The results and discussions After applying the psychological tests from the test battery Expert System Traffic psychological assessment, the data collected have been analyzed with SPPSS 17 programme. Table 1 shows the descriptive statistics of the independent and dependent variables. Table 1. Descriptive statistics, mean and standard deviation Variable Mean Standard deviation 1. Total performance in driving 63.81 12.14 2. TAHITO correct 46.27 11.26 3. TAHITO incorrect 31.93 13.47 4. DT omitted 23.29 15.21 5. DT correct 47.95 14.18 6. DT incorrect 27.16 10.66 7. TR reaction time 49.94 12.83 8. TR motor time 67.82 18.51 9. S.D. reaction time 41.12 15.27 10. S.D. motor time 38.61 18.35 11. Cognitrone 42.36 28.62 12. AMT 3388..1155 2266..7744 The correlation matrix in table 2 reveals the statistically significant correlations between the criteria and the predictor variables (the battery tests). Analyzing the multiple regression model for the composite criterion, the multiple correlation coefficient shows a high and statistically significant correlation between the predictors and the criterion (r=0.741, p<0.05). Also, the beta coefficients show that the variables of the tests are predictors for the performances registered in traffic (p<0.05). 462 Mihai Anitei et al. / Procedia - Social and Behavioral Sciences 30 (2011) 457 – 464 Mihai Aniţei/ Procedia – Social and Behavioral Sciences 00 (2011) 000–000 6 Table 2. The correlation matrix between the independent and dependent variables Variable 1 2. 3. 4. 5. 6. 7. 8. 9 10 11 12 1. Total performance in driving 1.00 2. tahito correct .48** 1.00 3. tahito incorrect -.36** -.47** 1.00 4. DT omitted -.31** -.48** .32** 1.0 5. DT correct .54** .53** -.18* -.44** 1.00 6. DT incorrect -.36** -.11 .29** .28** -.29* 1.0 7. TR reaction time .43** .10 .05 -.21** .11 .19* 1.0 8. TR motor time .47** .07 .26** -.22* .39** -.38** .28** 1.00 9. S.D. react. time .32** .08 .02 .05 .08 -.08 .38** .09 1.0 10. S.D. motor time .27** .07 .09 -.06 .29** .31** .08 .51** .14 1.0 11. Cognitron .37** .22** .08 -.28** .09 -0.10 -.07 .12 -.09 .05 1.0 12. AMT .28** .25** .24** -.23** .38** -.22** .05 .09 ..1122 .11 0.35** 1.00 *p<0.05 and **p<0.01 The correlation matrix (table 2) reveals the statistically significant correlations between the criteria and the predictors (independent variables of the psychological test). Thus, the total performances in traffic, has a statistically significant positive correlation with the following predictors: tahitoscop_corect (.48**), DT corect (.54**), reaction time (.43**), motor time (.47**), cognitrone (.37**) and AMT (.28**). The same criteria have a statistically significant negative correlation with the following independent variables: Dt omitted (-.31**), DT incorrect (-.36**) and tahito incorrect (-.36**). For the criterion total performances in traffic, the following regression model has been applied (table 3): Table 3. The multiple regression model for the dependent variable: total performances in traffic Dependent variable: total performances in traffic Standardized coefficients t p Independent variables β 1 Constant 22.17** 12.54 0.000 3. TAHITO correcte 0.21* 1.78 0.031 4. TAHITO incorrecte -0.42** -2.71 0.000 5. DT omitted -0.061 -0.47 0.482 6. DT correcte incorrect 0.72** 1.43 0.002 7. DT incorrecte correct -0.57* -1.38 0.042 8. TR reaction time 0.69** 1.72 0.003 9. TR motor time 0.54* 1.92 0.02 10. S.D reaction time 0.0049 0.69 0.35 11. S.D. motor time 0.0081 0.53 0.27 12. Cognitron 0.49* 1.378 0.031 13. AMT 00..3388** 11..5533 00..004411 F 7.629** 00..00000011 R 0.741 R2 0.549 R2 Adjusted 0.536 *p<0.05 and **p<0.01 463Mihai Anitei et al. / Procedia - Social and Behavioral Sciences 30 (2011) 457 – 464 Mihai Aniţei/ Procedia – Social and Behavioral Sciences 00 (2011) 000–000 7 As we can be observe in table 3 the regression model explains 54.9% of the variance (R square value). Also that the model is statistically significant (F=7.629; p=0.001) and the R square value is 0.549. The regression equation, obtained from the multiple regression model (table 3) is the following: total performances in traffic = 22.17+0.21*Tahit_correcte-0.42*tahito_incorrecte+ 0.72*DT_correcte- 0.57*Dt incorrect+0.69* TR reaction timet+0.54*TR_motor_time +0.49*Cognitron+0.38*AMT Figure 6. a)The The Histogram for Criteria; b) The scatterplot: predictors total performances in traffic Figure 6 a and b show that there is a strong positive correlation between the observed variables and the expected ones. The regression model has a predictive value for the chosen criteria: right curve, cross road. 5. Conclusions and recomandations The predictive regression validation model emphasizes the importance of using high performance statistical programs in choosing the psychological tests for evaluation. Thus, Expert System Traffic psychological assessment taking into consideration the cross-cultural study (Aniţei, Schuhfried, Chraif, Giehl & Buzea, 2010) to predict the driving behavior of the potential Romanian drivers in traffic was as a predictive model on a Romanian selected sample with age between 18 and 70 years, different levels of education. In addition to the studies previously validated (Schuhfried, Sommer, Aniţei & Chraif, 2010; Sommer, Schuhfried, Aniţei & Chraif, 2010), the tests of peripheral perception, composed time reaction, viziotest and the concentrated attention test have been removed in order to achieve a predictive validation for the test battery Expert System Traffic psychological assessment used in Europe and all over the world as well. Also the cost of the test battery has been taking into account. The negative correlation between these variables Dt_omitted (-.31**), DT_incorrect (-.36**) and tahito_incorrect (-.36**) highlights the fact that participants obtained low performances for the incorrect and omitted tasks, which takes to a higher level of performance in traffic. So, the test battery Expert System Traffic psychological assessment from the computerised psychological assessment company Vienna Test Sytem, standardized and validated in countries such as Austria, Germany, Saudi Arabia, it is also standardized and validated on Romanian population. Analyzing the statistical significance of the beta standardized coefficients (table 3), it can be observed that the independent variables explain the regression model. Thus, we can conclude that psychologists either apply and experimentally validate the psychological testing batteries or use validated and standardized psychological test batteries for the psychological assessment used in driving schools. This study based on the findings of the previous research show that the Romanian driving schools should improve the psychological assessment batteries with modern and validated instruments. The predictive regression validation model emphasizes the importance of using high performance statistical programs in choosing the psychological tests for evaluation. In addition to the results of the individual tests, the Expert System Traffic generates a highly valid overall assessment of the subject's driving-specific ability. Therefore, we can conclude that the Romanian psychologists should select and validate the psychological tests using performance criteria in experimental studies, before using it in driving schools as psychological assessments. 464 Mihai Anitei et al. / Procedia - Social and Behavioral Sciences 30 (2011) 457 – 464 Mihai Aniţei/ Procedia – Social and Behavioral Sciences 00 (2011) 000–000 8 Also, they have to take into consideration that the Standard Quality Driving Curriculum which needs to be kept continuously updated by considering the new developments that technology brings to vehicles and roads, all of which require the acquisition of new skills by drivers. References Aniţei, M., Mincu, C.L. & Chraif, M. (2008). Correlative study between the processing of stimuli in the peripheral perceptual field and speed and distance appreciation in the central perceptual field, The Journal of Human Resources [Revista Psihologia Resurselor Umane], Cluj- Napoca, no. 1. Anitei, M. & Chraif, M. (2008). Designing a test battery for predicting the driver’s behaviour, in the volume of the International Conference with international participation, “Titu Maiorescu” University, Ed. Univ."Titu Maiorescu". Aniţei, M., Schuhfried, G., Chraif, M., Giehl, M. & Buzea, E. (2010). The cross-cultural validation of the psychological tests, [Interkulturelle validierung von tests], in procedeengs of the 6th International Conference of Applied Psychology, Romania, Timişoara, 93-97. Aniţei, M., Chraif, M., Niculae, L. & Vancea, D. (2009). The gender influence in perception driving agresivity, in Proceedings of “Psychology and Society” [Influenţa genului în agresivitatea autopercepută a conducerii autovehicolului “Psihologie şi societate”- Sesiune anuală de comunicări ştiinţifice cu participare internaţională] Ed. Titu Maiorescu. Bishop, C.M. (1995). Neural networks for pattern recognition. Oxford, Oxford University Press, 1995. Bortz, J. (1999). Statistik für Sozialwissenschaftler [Statistics for social scientists]. Berlin, Springer. Brown, M.T. & Wicker, L.R. (2000). Discriminant analysis. In H. E. A. Tinsley & S. D. Brown (Eds.), Handbook of applied multivariate statistics and mathematical modelling, San Diego, CA: Academic Press, 209-234. Dula, C.S. & Ballard, M.E. (2003). Development and evaluation of a measure of dangerous, aggressive, negative emotional and risky driving. Journal of Applied Social Psychology, issue 33, volume 2, 263-282. Fellendorf, M. & Vortisch, P. (2001). Validation of the microscopic trans. model VISSIM in diferent real-world situations. In Proceedings of the Transportation Research Board, Washington, DC. Herle, M. (2009). Personality measurement in traffic psychological assessment. In: W.R. Nickel, G. Meinhard, I. Born (Hrsg). Proceedings zu Fit to drive. 4th International Traffic Expert Congress, 56-61. Bonn: Kirschbaum. Hornke, L.F., Kuppers, A. & Etzel, S. (2000). Konstruction und evaluation eines Adaptativen Matrizentest. Diagnostica, 46, 182-188. W.M.K. Trokim, Social research methods, http://www.socialresearchmethods.net Karner, T. & Neuwirth, W. (2001). Die Bedeutung der peripheren Wahrnehmung in der verkehrspsychologischen Untersuchung. Psychologie in Österreich, 20, 183-186. Pérez Arias, A., Kretz, T., Ehrhardt, P., Hengst, S., Vortisch, P. & Hanebeck, U.D. (2009). A Framework for Evaluating the VISSIM Traffic Simulation with Extended Range Telepresence. In: Proceedings of the 22nd Annual Conference on Computer Animation and Social Agents, 13-16. Pitariu, H. & Chraif, M. (2009). The validation of scales with behavioural anchors, The Psychology Journal, Ed. Academiei Române, issue 3/4. Risser, R., Chaloupka, Ch., Grundler, W., Sommer, M., Häusler, J. & Kaufmann, C. (2008). Using non linear methods to investigate the criterion validity of traffic-psychological test batteries. Accident Analysis & Prevention, Volume 40, Issue 1, 149-157. Rojas, R. (2000). Neuronal Networks, A systematic introduction, Heidelberg, Springer, 2000. Schuhfried, G., Sommer, M., Aniţei, M. & Chraif, M. (2010). The validation methodology of a psychological assessment battery for drivers’ schools in Romania. Romanian Journal of Applied Experimental Psychology, volume 1, issue 1, 5-14. Schuhfried, G. (2007). Handbook of the tests. Vienna Tests System, www.schuhfried.at Sommer, M., Schuhfried, G., Aniţei, M. & Chraif, M. (2010). A model of driving test battery validation. Journal of Psychology, 56, no. 3–4, 243– 250. Sommer, M., Arendasy, M., Olbrich, A. & Schuhfried, G. (2004). Qualitätsverbesserung in der verkehrspsychologischen Diagnostik mit neuronalen Netzen: Eine Pilotstudie. Zeitschrift für Verkehrssicherheit, 50, 193-198. Sommer, M., Herle, M., Häusler, J., Risser,R., Schützhofer, B. & Chaloupka, Ch. (2008). Cognitive and personality determinants of fitness to drive Transportation Research Part F. Traffic Psychology and Behaviour, Volume 11, Issue 5, 362-375. Sommer, M. & Häusler, J. (2005). Non-Linear Methods for the Identification of Drivers at Risk to Cause Accidents. In L. Dorn (ed), Driving Behaviour and Training. Volume II, 425-436, Hamphire, Ashgate. Ulleberg, P. & Rundmo, T. (2003). Personality, attitudes and risk perception as predictors of risky driving behaviour among young drivers. Safety science, 41, 427-43. Van Tongeren, R., Gietelink, O., De Schutter, B. & Verhaegen, M. (2007). “Traffic modelling validation of advanced driver assistance systems,” Proceedings of the 2007 IEEE Intelligent Vehicles Symposium, Vol IV, Istanbul, Turkey, 1246-1251. Warner, B. & Misra, M. (1996). Understanding neural networks as statistical tools, The American Statistician, Volume 50, 284-293. Wiedemann, R. (1974). Simulation des Straßenverkehrsflusses. Schriftenreihe des Instituts für Verkehrswesen der Universität Karlsruhe, Heft 8. Wiedemann, R. (1991). Modelling of RTI-Elements on multi-lane roads. In: Advanced Telematics in Road Transport edited by the Comission of the European Community, DG XIII, Brussels. 
work_uewjv5lepbhp5nirtc27m6vj7m	----	Business information extraction from semi-structured webpages Business information extraction from semi-structured webpages Nahk Hyun Sung a,*, Yong Sik Chang b a Department of Management Information Systems, Yong-In University, 470 Samga-dong, Yongin, Kyungki 449-714, South Korea b Department of e-Business, Hanshin University, 411 Yangsan-dong, Osan, Kyungki 447-791, South Korea Received 9 October 2003; accepted 1 December 2003 Abstract To protect online consumers, as OECD Guidelines recommend, Internet shopping malls should provide information about their business on their webpages. In Korea, The Consumer Protection Law in Electronic Commerce, forced Internet shopping malls to provide their business information, so that consumers could easily identify them. Since most Korean Internet shopping malls provide consumers with business information in a semi-structured format on their homepages, a software agent can easily identify them. To investigate automatically the provision of the business information with the Internet shopping malls, this article proposes the methods of gathering URLs of Internet shopping malls, of monitoring alterations of webpages, and of extracting business information. Business information extraction in our research is based on synonyms and indicator words of the attributes. We used inductive learning to raise the efficiency of information extraction. With experiments, we showed the potentialities of our agent system. The average extraction accuracy of our agent system was 89.3%. q 2004 Elsevier Ltd. All rights reserved. Keywords: Electronic commerce; Internet shopping mall; Business information; Information extraction; Agent 1. Introduction The factors that affect public confidence in Internet shopping malls are reputations of shopping malls, clearness of business information, protection policies for consumers’ privacy information, and security policies for payment, etc. Among these factors, clearness of business information is a basic factor that can lead to confidence in shopping malls in the electronic commerce environment. OECD announced the Guidelines for Consumer Protec- tion in Electronic Commerce in 1999 (OECD, 1999). The OECD Guidelines and the Guidelines of Membership Nations, created shortly thereafter, specify that Internet shopping malls should provide at least a basic minimum of business information on their webpages, including the name of the business, the name of the representative, geographical address, telephone number, fax number, e-mail address, and business license number. As examples, Fig. 1 depicts two homepages including business information: BEST BUY Co., INC. (www.bestbuy.com) in the US and LGeshop (www. lgeshop.com) in Korea. As we can see in the examples of BEST BUY Co., Inc. and LGeshop, while most Internet shopping malls in the US provide their business information, scattered in several pages, in an unstructured format, most Korean Internet shopping malls provide their business information on the bottom of their homepages in a semi-structured format. In Korea, The Consumer Protection Law in Electronic Commerce, which came into effect in July 2002, forced Internet shopping malls to provide a minimum of seven forms of business information, including the name of the business, the name of the representative, geographical address, telephone number, fax number, e-mail address, and business license number, so that consumers could easily identify them. Therefore, in Korea, Internet shopping malls should provide their business information on their webpages. Since most Korean Internet shopping malls provide their business information in a semi-structured format, an agent can easily identify them when compared with other countries such as the US. If any shopping mall intentionally omits all or a part of the required business information, they can be regarded 0957-4174/$ - see front matter q 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2003.12.008 Expert Systems with Applications 26 (2004) 575–582 www.elsevier.com/locate/eswa * Corresponding author. Tel.: þ82-1620-71809; fax: þ82-3133-02885. E-mail addresses: nhsung@yongin.ac.kr (N.H. Sung); yschang@ hanshin.ac.kr (Y.S. Chang). http://www.bestbuy.com http://www.lgeshop.com http://www.lgeshop.com http://www.elsevier.com/locate/eswa as a suspect of online fraud. If an organization would detect Internet shopping malls which lack business information and admonish them for not providing business information, it would enhance public confidence in electronic commerce; however, it is difficult for a person to visit a large number of Internet shopping malls’ homepages to investigate whether or not they provide business information. Fig. 1. Examples of homepages providing with business information. BEST BUY Co., Inc. in the US provides its business information in an unstructured format and LGeshop in Korea provides its business information, given inside the rounded rectangle made of dotted lines, in a semi-structured format. It was translated into English for readability. N.H. Sung, Y.S. Chang / Expert Systems with Applications 26 (2004) 575–582576 https://isiarticles.com/article/22613 
work_ufcw2a7d5rfhvjyra7j3cl3ejq	----	A waveletbased expert system for digital subscriber line topology identification INTERNATIONAL JOURNAL OF COMMUNICATION SYSTEMS Int. J. Commun. Syst. (2014) Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/dac.2795 A wavelet-based expert system for digital subscriber line topology identification V. D. Lima1,*,† , A. Klautau2, J. Costa1, K. Ericson3, A. Fertner3 and C. Sales1 1Applied Electromagnetism Laboratory, Federal University of Pará (UFPA), 66073-900 Belém-PA, Brazil 2Signal Processing Laboratory, Federal University of Pará (UFPA), 66073-900 Belém-PA, Brazil 3Ericsson AB, 164 80 Stockholm, Sweden SUMMARY This work proposes a new method for automatically identifying topologies of lines with one or more sections in a telephone network. The method is based on the examination of both impulse response and time-domain reflectometry trace of a line under test. They are analyzed using a method based on the wavelet transform that identifies and extracts features that contain information about the line topology. Those features are interpreted by an expert system composed of three sequential modules that estimate, respectively, the type of line makeup (serial or bridge tap), the lengths of the line sections, and the corresponding cable type, which are the parameters that completely identify the topology according to the assumed model. A thorough comparison with two state-of-the-art methods is also presented using several twisted-pair copper cables. The results show that the proposed method provides good accuracy with respect to topology identification at low computational cost. Copyright © 2014 John Wiley & Sons, Ltd. Received 7 February 2013; Revised 13 January 2014; Accepted 21 March 2014 KEY WORDS: DSL; line topology identification; continuous wavelet transform; expert system; double- ended line testing; single-ended line testing 1. INTRODUCTION The wireline broadband access currently represents a large share of the telecommunication industry. In the end of the first quarter of 2012, there were about 612.6 million fixed lines in the world, representing 11.5% growth in relation to the same period in the previous year. The digital subscriber line (DSL) continues to be the predominant technology [1]. In general, the use of the DSL is so attractive because a relatively fast and cheap procedure for service deployment can be used once the service is provided through an already existing infrastructure, the copper access-network. However, the quality of service is strongly dependent of the physical characteristics of the network, which was originally designed for plain old telephone service (POTS). Features such as bridge taps, which might not affect POTS, are impairments for DSL. In addition, the knowledge about the network physical structure and consequently its performance is often incomplete or nonexistent [2]. The topology of a telephone line can be made up of one or more sections with different char- acteristics. The identification of these sections can be performed through the analysis of line measurements that can be categorized as follows: single-ended line testing (SELT), which requires measurement equipment only in one line end and double-ended line testing (DELT), which requires equipment in both terminals [3]. It should be highlighted that, even if DELT is not available, there exist techniques to estimate the frequency response through SELT measurements [4, 5]. *Correspondence to: V. D. Lima, Applied Electromagnetism Laboratory at the Federal University of Pará (UFPA), 66073-900 Belém-PA, Brazil. †E-mail: vinicius@ufpa.br Copyright © 2014 John Wiley & Sons, Ltd. V. D. LIMA ET AL. Having a reasonable line topology estimate, it is possible to evaluate the characteristics of the connection that may impact the quality of service. This gives to the service provider the oppor- tunity of improving the network structure. Further, once the line topology estimate is obtained, line monitoring and fault detection are simplified by comparisons of new tests with the results obtained previously. In [6–9], time domain reflectometry (TDR) is discussed and an iterative method, called here SELT-tdr, based on maximum likelihood estimation is proposed. An identification of the twisted- pair line sections is achieved through the comparison of a time domain signal within selected time intervals with curves generated using a line model. The accurate estimation of these intervals is essential. However, a method to detect the intervals is not clearly defined. In [10–13], the one-port scattering parameter S11.f / is measured in frequency domain. After measurement, calibration and pre-processing, S11.f / is analyzed in time domain, where important extracted features are inter- preted through a system based on Bayesian networks for estimating the line topology. In [14], it is proposed a combined use of complementary code based correlation time domain reflectometry (CTDR) and frequency domain reflectometry (FDR). First, CTDR measurements are used to esti- mate the line discontinuities, as an initial guess, and after a FDR-based optimization method is used to refine the estimate. In [15], the limitations of SELT are theoretically analyzed and it is concluded that the structure of a network cannot be always identified without ambiguity using only SELT. In [16], a method based on a genetic algorithm (GA), called TIMEC, analyzes frequency response H.f / and S11.f / with a multi-objective criterion using a line model as reference for line topology identification. The method permits using a priori knowledge about the line. On the other hand, the GA implies a relatively high computational cost when compared with other methods. In [17], an algorithm that compares each DELT measurement with calculated impulse response of a “canonical” set of line topologies is presented. In this paper, we propose an approach that combines SELT and DELT information for topology identification of lines with one or more sections. It is assumed that a line topology can be completely identified by determining three essential parameters of each one of its sections: length; type of cable; which can be defined by its physical properties; and the type of section that can be serial or bridge tap. In order to estimate these parameters, the impulse response and the TDR trace are estimated from DELT and SELT measurements, respectively. Afterwards, they are processed using a wavelet transform to identify and extract features that contain information about the line topology. It must be highlighted that this phase is crucial to the accuracy of the method. The extracted features are then applied to an expert system that aims at correctly identifying the topology parameters. This paper is organized as follows. An overview about how the signals of interest can be interpreted is presented in Section 2. In Section 3, the wavelet-based method for characterizing and extracting features from the signals measured in a line is described. Section 4 describes the expert system proposed for automatically interpreting the signal features and obtaining an esti- mate of the line topology. Identification results, analysis, and comparison with the state-of-the-art methodologies are presented in Section 5. The final remarks are made in Section 6. 2. CHARACTERISTICS OF THE MEASURED SIGNALS Identifying the topology of a given telephone line means identifying the connected line sections that bind the central office (CO) to the customer-premises equipment (CPE). These sections can be made of different cable types connected in cascade (serial sections) or in parallel (bridge taps). This way, the topology of a telephone line can be represented as a set ‚ D¹�1; �2; : : : ; �nsº ; (1) where the subset �k, with k D 1;2; : : : ;ns, represents the k-th line section from CO to CPE or, respectively, port 1 to port 2, considering the line as a two-port network, and ns is the number of sections that composes the line. Each subset �k is represented by the parameters �k D¹�k; lk;�kº ; (2) Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac A WAVELET-BASED EXPERT SYSTEM FOR DSL TOPOLOGY IDENTIFICATION where �k 2 ¹bridge tap; serialº, lk, and �k are, respectively, the type of section, its length, and the type of cable that constitutes the k-th line section. For the majority of previous works about line topology identification [8–10, 12, 16], finding �k corresponds to estimate the number of American wire gauge that best fits to the measured data for the k-th section. However, a variety of parameters is required to properly describe a cable type [18] and ideally �k should be the set of nominal physical properties that identify a specific cable. The following discussion will suggest how the signal of interest can be interpreted in order to estimate ‚. The VUB0 model [18] is used to simulate these signals because it is a causal model and is defined as a function of physical properties of the line. The configuration of values for the parameters used in the model is based on [19] and �k is assumed here to be represented by the conductor diameter. In this work, the signals of interest are the impulse response and the TDR trace. The impulse response can be obtained from the inverse Fourier transform of the frequency response. The TDR trace can be obtained from the input impedance Zin.f / through R.f / D Zin.f / Zin.f /CZs Vs.f /; (3) where Zs is the output impedance in the signal source, Vs.f / is the signal generated by the source, in general a rectangular pulse, and R.f / is the resulting voltage in the connection between the source and the line. The TDR trace r.t/ can be obtained by the inverse Fourier transform of R.f /. 2.1. Interpreting impulse responses The voltage reflection coefficient � is defined as the ratio of the reflected voltage V� to the incident voltage VC and can be written as � D V� VC D ZA �ZB ZA CZB ; (4) where ZA and ZB are the characteristic impedances of the line sections after and before the discontinuity, respectively. The voltage transmission coefficient T is expressed by T D 1C�: (5) The characteristic impedance of a metallic line at high frequencies is approximately a pure resis- tance [20]. This implies that reflection coefficient is real-valued and 1 > � > �1, whereas T is always positive. Hence, the impulse response is composed of a series of positive (or upward) pulses delayed in time. Each succeeding pulse is resulting from a different ‘path’ taken by the input signal into the line. Let us consider the topology illustrated in Figure 1(a) and assume that the test signal is injected by a source connected to port 1 and measured in port 2. Figure 1(b) depicts four possible paths in which parts of the input signal can follow from port 1 to port 2, namely by NA, NB, NC , and ND. Figure 1(c) shows the resulting simulated impulse response h.t/. The arrival times of the pulses resulting of each path are, respectively, represented by t NA, t NB, t NC , and t ND. The length Li of the i-th path can be calculated by the time of arrival ti of the i-th pulse by Li D tivp, where vp is the velocity of propagation of the transmitted signal, which is assumed constant here. The first pulse to reach port 2 in t NA is resulting from path NA, the straight path from port 1 to port 2 of length LT , which is an important parameter in the topology characterization. The presence of each bridge tap originates a new path and consequently a new pulse in h.t/. In Figure 1(b), the paths NB and NC are composed of bridge taps. The length of the bridge taps, in the third and fifth sections, can be calculated by l3 D � t NC � t NA � 2 vp and l5 D � t NB � t NA � 2 vp: (6) Unfortunately, it is not possible to determine the position of bridge taps using only the impulse response. If the position of a bridge tap is changed, the path length traveled by the signal remains the same, arriving with the same shape in the receiver. Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac V. D. LIMA ET AL. (a) (b) (c) Figure 1. In (a), a topology of six sections with bridge taps in the third and the fifth sections. In (b), four paths are defined by the impedance mismatches. In (c), the simulated impulse response for the topology in (a) is illustrated. Positive pulses can also be generated by multiple reflections. The path ND traveled by the pulse that reaches the receiver in t ND is an example of a pulse generated by multiple reflections. In addition, connections among serial sections with different cable types typically do not generate perceptible modifications in the resulting signal. 2.2. Interpreting TDR traces For interpreting TDR traces, it is assumed here that port 2 is an open-circuit. The essential features in a TDR trace are defined by three main types of line connections: serial to serial, serial or bridge tap to open termination, and serial to bridge tap. Each one is discussed in the sequel. 2.2.1. Serial to serial. The connection between two serial sections of different types of cable is characterized by a subtle change in the signal decay pattern. Figure 2 shows TDR traces simulated from topologies with two serial sections. As showed in [6], when the previous impedance is smaller than the subsequent, � is positive. The opposite happens when larger impedance is connected with a smaller one. The arrival times tA and tB of the signal features generated by the impedance that mismatches A and B, respectively, are represented in Figure 2. The time of arrival tA can be used to estimate the Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac A WAVELET-BASED EXPERT SYSTEM FOR DSL TOPOLOGY IDENTIFICATION Figure 2. Comparison among TDR traces simulated from topologies of two sections. The input signal is a rectangular pulse with 1 �s of duration and 10 V of amplitude. The output impedance of the source is 50 ohm. Figure 3. TDR trace simulated for a line with one bridge tap. The reference TDR trace has l1 D 2000 m and � D 0:5 mm. location of point A, a serial to serial connection, through l1 D tA 2 vp: (7) 2.2.2. Serial or bridge tap to open termination. For an open-circuited termination (� D 1), if the input pulse is positive, the reflected pulse is also positive. In the example depicted in Figure 2, the pulse related to termination B reaches port 1 in tB. 2.2.3. Serial to bridge tap. A bridge tap can be considered in parallel with the equivalent impedance of the following sections which means that � < 0. Consequently, if the input pulse is positive, a reflected pulse generated by a serial to bridge tap connection is negative [6]. Figure 3 shows a TDR trace simulated for a topology with one bridge tap. The position of the bridge tap can be obtained from the arrival time tA of the negative pulse via l1 D tAvp=2. The open-circuited termination of Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac V. D. LIMA ET AL. the bridge tap originates a positive pulse. The interval between this positive pulse and the negative pulse generated by the bridge tap beginning can be used to estimate the bridge tap length using l2 D .tB � tA/ 2 vp: (8) The pulse of arrival time tC is related to the line termination C . However, in this case, it is not possible to differentiate the positive pulses without a priori information. Additionally, spurious pulses can rise because of multiple reflection of input signal between the impedance mismatches. Pulses and decay variations are very important signal events generated by topology mismatches. If properly identified, these events can be used to infer the topology parameters of a given line. In several signal processing applications, the boundaries of events and the transition between states indicating important features may be indicated through detection of edges [21–28]. Because r.t/ and h.t/ are typically continuous and without eccentricities, edges can be detected as isolated sin- gularities, which are points at which a given function is not defined, or it fails to be well-behaved in some particular way, such as differentiability or continuity. This section discussed how the signals can be interpreted, but this requires detecting the events of interest via the associated singularities. Wavelets are used for estimating these singularities, as will be discussed in the next section. 3. WAVELET-BASED FEATURE EXTRACTION This work proposes a method for extracting features of r.t/ and h.t/ based on wavelet transforms, which provide a powerful multiscale representation of signals. The method consists of two steps: (i) detecting the edges in the signals and (ii) classifying these edges as pulses (positive or negative) or decay variations. 3.1. Singularity detection with wavelets The continuous wavelet transform (CWT) of a given continuous and square-integrable function f.t/ can be computed through the convolution of f.t/ with a set of dilated wavelets. A mother wavelet, or simply wavelet, is a function continuous both in time and frequency domain with a zero average, normalized as k k D 1, and centered at t D 0. A family of wavelets is obtained by scaling by s and translating by u the function u;s.t/ D 1 p s N � t �u s � ; (9) where N represents the complex conjugate of . The CWT of f.t/ at time u and scale s, computed with respect to the wavelet .t/, is given by Wf.u;s/ D C1Z �1 f.t/ u;s.t/dt D f.u/� s.u/ (10) with s.t/ D 1 p s N � �t s � : (11) One of the most crucial features of a wavelet in the evaluation of the local regularity of a signal is the number of its vanishing moments. If the wavelet has N vanishing moments defined by C1Z �1 tk .t/dt D 0 for 0 6 k < N; (12) Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac A WAVELET-BASED EXPERT SYSTEM FOR DSL TOPOLOGY IDENTIFICATION and a compact support, then can be written as the N -th order derivative of a function ˇ also of compact support such that [29] .t/ D .�1/N dNˇ.t/ dtN with C1Z �1 ˇ.t/dt ¤ 0: (13) Thus, by using ˇ, the resulting CWT is a multiscale differential operator of order N , and (10) can be rewritten as Wf.u;s/ D sN dN duN Œf.u/�ˇs.u/� (14) with ˇs.t/ D s �1=2 Ň .�t=s/. If has order N D 1, the wavelet modulus maxima are the maxima of the first-order derivative of f smoothed at the scale s by ˇs.u/. The term modulus maximum describes any point .u0;s0/ such that jWf.u;s0/j is a local maximum at u D u0. A maxima line is any connected curve s.u/ in the plane .u;s/ along which all points are modulus maxima. The singularities are detected by following the wavelet transform local maxima from coarser to finer scales and finding the abscissa where the wavelet modulus maxima converge in the form of maxima lines [29]. Thus, the proposed edge detection method is summarized in the following steps. 3.1.1. Calculus of the CWT. It is not guaranteed for any that any modulus maximum located at .u0;s0/ belongs to a maxima line propagating to the finest scale. In [29], it is proved that if ˇ is Gaussian, the modulus maxima of the CWT of any continuous and square-integrable function f.t/ are in maxima lines that are not interrupted when the scale decreases. Then, the wavelet function used here to compute the CWT of the signals under analysis is the normalized first derivative of the Gaussian function. Derivatives of other orders can also be used to detect the singularities; however, the usage of the derivative of the first-order is simpler than higher orders. 3.1.2. Local modulus maxima and thresholding. In the calculus of local modulus maxima, spurious can be generated by numerical errors, mainly in regions where the CWT is close to zero. The strategy adopted here to remove the spurious and, at the same time, to control the detection sensitivity is based on the definition of thresholds at each scale of the CWT. However, we should not define one unique threshold for all CWT scales because each scale contains a different energy content of the signal under analysis. The threshold needs to be proportional to the average amplitudes in the scale, so that the algorithm eliminates small variations. The arithmetic mean and the variance of each scale can be used to assure that the significant maxima will be retained by the algorithm. Thus, the proposed threshold for each scale of the CWT is given by �i D ks 2 6664 nP jD1 ˇ̌ Wf � uj ;si �ˇ̌ n C 2u;si 3 7775 ; (15) where si is the i-th scale, n is the number of the elements in the i-th scale that is equal to the number of samples of the signal f , and 2u;si is the variance of the coefficients at scale si. The constant of sensibility ks is the parameter used to provide control over the sensibility in the search for local extrema. For instance, setting ks D 0:5 duplicates the sensibility and ks D 2 decreases by half the default sensibility. A hard thresholding is applied to each scale of the CWT, that is, zero is set to all coefficients whose absolute value is lower than �i. 3.1.3. Singularity detection. From the ‘thresholded’ CWT, the maxima lines are found. The sin- gularities are detected when the abscissas where these lines converge at finest scales are found. Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac V. D. LIMA ET AL. (a) (b) Figure 4. In (a), a pulse containing two singularities, detached by circles, is showed. In (b), its correspondent CWT with the local maxima converge to abscissas that correspond to singularities. Figure 4(a) shows a rectangular pulse whose singularities are detached by circles. Figure 4(b) shows its correspondent CWT with the local maxima converging to abscissas that correspond to singularities. The local modulus maxima are the extrema that can be both peaks and valleys. Peaks are related to variations from lower amplitudes to higher ones and, conversely, valleys are related to variations from higher to lower amplitudes. Using this information, rise edges can be identified by peaks in the finest scale in CWT and fall edges by valleys. 3.2. Classification of estimated singularities Let P D¹p1; : : : ; pmº be the set of m singularities detected in f.t/, where pi D .ti;f .ti// with i D 1; : : : ;m and ti is the time of detection of the i-th singularity. Now, it is necessary to deter- mine if the detected points are related to pulses, decay variations, or spurious caused by numerical error or noise. Because r.t/ and h.t/ have different characteristics, the proposed method for feature extraction defines different algorithms to classify the features of the two signals. In Section 2, impulse responses are characterized by positive, distinct, and sharp pulses. The algorithm verifies, for each singularity pi, if the pair of successive singularities pi and piC1 are, respectively, the rise and fall edges. Each pair identified positively is associated a positive pulse. Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac A WAVELET-BASED EXPERT SYSTEM FOR DSL TOPOLOGY IDENTIFICATION On the other hand, TDR traces are characterized by positive and negative pulses and decay vari- ations. Besides, pulses are smoother than in impulse response and can be overlapped generating a mixed event. The first issue is to determine if the detected point pi characterizes a pulse or a decay deviation. It is assumed that the slope of the tangent line at a pi related to pulse tends to be sharper than a point related to decay variation. Thus, the proposed algorithm calculates the approximation of derivatives of first-order at pi, through central difference formula, in order to find slope of the tangent line at that point. A threshold is defined. If the derivative is lower than , pi represents a decay variation. Otherwise, it is necessary to verify if pi pertains to a pulse. To identify pulses, the following basic assumptions are adopted: the input signal is a well-known wideband rectangular pulse, and the propagation channel is approximately a time-invariant linear system. Thus, each echo, as well as the input signal, is composed of a start and an end edge whose time interval between them �tip is approximately the same to the width of the input pulse. This way, for pi to be an edge of a pulse, the following conditions must be true. (1) There is a pk, where i < k 6 m that represents a type of edge different from pi into the interval �tip˙�, where � is a margin used to take into account numerical errors, dispersion, and noise. Its value is defined here through simulations as 20% of the value of �tip. If more than one singularity accomplishes this condition, the pk is selected whose interval pk �pi results in a smaller mean squared error in the comparison with �tip. (2) There is only one local maximum or minimum in the interval Œpi;pk�. (3) pi is not an end edge of a previous pulse, and the local extremum in the interval Œpi;pk� is not closer to zero than the other points in the interval. If all the aforementioned conditions are true, the algorithm verifies if pi is a rise edge. If yes, pi and pk correspond to a positive pulse. Otherwise, pi and pk correspond to a negative pulse. If one of the conditions is false, pi is considered spurious. 4. KNOWLEDGE-BASED LINE TOPOLOGY IDENTIFICATION The proposed expert system uses an explicit model of knowledge that interprets the features extracted from r.t/ and h.t/. This knowledge model was developed through the systematization on a set of rules of the relations between the features of a signal transmitted across a multi-section metallic line and its impedance mismatches. In order to validate these rules, systematic observa- tions of several different topologies simulated using causal line models were used. The test set is composed by measurements on real cables and these results are described in Section 5. Thus, the method was structured into three algorithms that estimate sequentially and, respec- tively, the following sets defined from the model established in Section 2 for the topology of a transmission line, T D ® �1;�2; : : : ;�nS ¯ ; L D ® l1; l2; : : : ; lnS ¯ ; and ˆ D ® �1;�2; : : : ;�nS ¯ : (16) Those three algorithms are explained as follows. 4.1. Identifying the set T The identification of T defines how many sections the line under test have and if these sections are serials or bridge taps. In r.t/, the beginning of the line sections is represented by the arrival time of negative pulses and decay variations inside an interval delimited by t D 0 and a reference time tref , which is correspondent to LT . The method estimates LT from the arrival time of the first pulse in h.t/, t h 1C. This way, tref is calculated by using tref D 2t h 1C. Next, all the events detected inside Œ0; tref � in r.t/ are sequentially analyzed. Assuming that the first section is a serial section, the algorithm includes a new serial Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac V. D. LIMA ET AL. section for each detected decay deviation and a new bridge tap for each negative pulse. In the interval between occurrences of two consecutive bridge taps, if any decay deviation is found, one serial section is defined. Thus, an estimate for T is obtained. In general, pulses are detected in h.t/. However, the algorithm has rules for exceptional cases in which pulses are not detected in h.t/ (what usually occurs when the line is very long). These rules are defined in accordance with the features detected in r.t/. � The first detected pulse is positive. tref is estimated as the time of arrival of this first positive pulse. In this case, there are no bridge taps and one or more serial sections can occur. � The first detected pulse is negative and there are one or more positive pulses. It is not possible to obtain a unique estimate for tref . The algorithm considers the arrival time of each positive pulse as a hypothesis for tref . Consequently, more than one estimate for T can be obtained. � Positive pulse was not detected. It is not possible to estimate LT and, consequently, T cannot be fully estimated. However, a partial estimate can be obtained. Decay deviations and negative echoes, if occur, are sequentially analyzed until the last characteristic detected. However, the algorithm cannot estimate bridge tap lengths. 4.2. Estimating the set L Given an estimate for T , the proposed method uses the signal features to obtain an estimate for the length of each estimated line section. The algorithm is divided in two parts: one for serial sections and other for bridge taps. The basic idea is the same for both: finding the time of start ts and the time of end te of the interval in r.t/ that represents the section. This interval is used to estimate the length of the k-th section using lk D .te � ts/ 2 vp: (17) 4.2.1. Algorithm for serial sections. ts and te are estimated from the arrival time of the TDR trace feature correspondent to the element before and after the section under analysis, respectively. Case network element before the section is � The port 1. ts D 0; � A bridge tap. ts D t r B� , where tr B� is the arrival time of the negative pulse that corresponds to the beginning of the bridge tap before the section under analysis; � A serial section. ts D t r Bd , where tr Bd is the arrival time of the decay deviation that corresponds to the beginning of the section under analysis. Case network element after the section under analysis is as follows: � The port 2. te D tref ; � A bridge tap. te D t r A� , where tr A� is the arrival time of the negative pulse that corresponds to the beginning of the bridge tap after the section under analysis; � A serial section. te D t r Ad , where tr Ad is the arrival time of the decay variation that corresponds to the beginning of the serial section after the section under analysis. 4.2.2. Algorithm for bridge taps. Pulses detected in h.t/ after the first one can be related to bridge taps. They can provide length estimates for these sections. The set of ‘candidate lengths’ obtained from h.t/ can be calculated by using lhC.j �1/ D � th jC � th1C � 2 vp; (18) where lh C .j �1/ is the .j �1/-th candidate length, and th jC is the time of arrival of the j -th positive pulse detected in h.t/ considering j > 1. Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac A WAVELET-BASED EXPERT SYSTEM FOR DSL TOPOLOGY IDENTIFICATION Possible lengths also can be obtained in r.t/ from the interval between negative pulse correspond- ing to the bridge tap under analysis and all positive pulses detected after this negative echo. These ‘TDR candidates’ are defined by lrC.m/ D � trmC� t r n� � 2 vp; (19) where lr C .m/ is the m-th TDR candidate length, and trmC is the time of arrival of the m-th positive pulse detected in r.t/ after trn�, the time of arrival of the n-th negative pulse. The length attributed to the section under analysis is the value of lh C that obtains the smallest absolute error in the comparison with the values of lr C . After obtaining the lengths for all of the estimated sections, the total serial length of the line is obtained through LT D t h 1Cvp: (20) Once the length of the serial sections is calculated from features of r.t/ and LT is estimated from h.t/, it is possible to have differences between these values. As the value of the LT obtained from h.t/ tends to be more reliable, the serial sections are adjusted proportionally to the value calculated for LT . 4.3. Classification of the set ˆ Assuming that a telecommunication network is composed of a finite set of types of cable, for a given line topology, the set ˆ would be composed of a combination of elements from this ‘set of cable types’. Once estimates for T and L were obtained, a set of hypotheses for ˆ can be defined from a database of cable types. Using a line model, SELT and DELT signals can be generated for each topology hypothesis. This work proposes to compare these simulated signals with line measurements to obtain an estimate for ˆ. The idea is to select the hypothesis that generates the smallest mean squared error in relation to the measured curves. The cable type database is generated from network samples. For each type of cable, the values of the physical and geometric parameters of the line model are adjusted in agreement with the cable configurations. The parameters can be obtained from a priori information or estimated from line measurements. In [30], an interesting method for identification of parameters of twisted-pairs from input impedance measurements through a combination of analytical approach and optimization process is proposed . However, the larger the database, the larger the number of operations is needed to obtain the solu- tion. For example, consider a database composed of four types of cable. If a topology is estimated with four sections, being the third section, a bridge tap and the others serial sections, 192 hypothesis for ˆ would be defined (4�3�4�4 D 192, once a connection between serial sections cannot be made up of the same type of cable). On the other hand, if six sections are estimated, with bridge taps in the third and the fifth sections, 3072 hypothesis would be generated (4�3�4�4�4�4 D 3072). A method for accelerating the classification process is to compare the first serial sections through TDR trace. It can be accomplished because of the fact that the first serial sections before the first derivation (if occurs) produces consecutive signal segments that can be compared separately. Then, for the second example presented in the previous paragraph, for classifying the first serial section, it is necessary to simulate a TDR trace for a single line with infinite (or very long) length for each type of cable in the database. Each simulated curve is compared with the measured signal in the interval between t D 0 and the time of arrival of the decay deviation that delimits the end of the interval related to the section under test. In this case, only four possibilities are tested. In relation to the second section, that is, also serial, the same procedure is applied except for the cable type that was already selected for the previous section. The third section is a bridge tap and, from this point, the comparison of curves for the possible combinations for the last two sections taking into account the first two sections is identified. As consequence, only 263 combinations would be tested (4C3C4�4�4�4 D 263). Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac V. D. LIMA ET AL. It is important to highlight that the proposed classification methodology (as well as most methods of topology identification) assumes that the physical parameters of the line have stationary behavior in time and space. Or even that in the variation of the mean, these parameters are sufficiently slow. An abrupt change could have a direct impact on the signal shape and, consequently, on the features identified by the method described in Section 3. However, abrupt changes could be considered as faults, which are out of scope of this work. Another consideration is the importance of accuracy of the feature extraction method. The iden- tification of the sets T , L and ˆ is directly dependent on the correct identification of the signal characteristics. In the next section, the robustness of the method is tested from measurements with real cables and comparison with the state-of-the-art techniques. 5. RESULTS AND ANALYSIS 5.1. General conditions of evaluation In order to evaluate the proposed wavelet-based expert system (WES) for line topology identifica- tion, a test set composed of measurements in twisted-pair copper cables was created . Twelve line topologies were used and divided in three subsets: topologies with only one section (lines 1 to 6), lines made up of more than one serial sections (lines 7 to 10), and lines with one bridge tap (lines 11 and 12). All topologies are composed of two cable types: Ericsson TEL 481 02, 0.4-mm diameter wire, and Ericsson TEL 313 000, 0.5-mm diameter wire, named here A and B, respectively. These test lines are described in Table I. The cable type database, used by the algorithms, is composed of the two types mentioned earlier and two other cable types: Furukawa CTP-APL 65, 0.65-mm diam- eter wire and a theoretical cable 0.9-mm diameter wire, whose parameters are defined in [19]. These cable types are named here C and D, respectively. The physical parameters of the cables A, B, and C were obtained using the method defined in [30]. The measurements were performed in the laboratory. For each one of the test lines, measurements of frequency response H.f /, scattering parameter S11, and open-circuit input impedance Zin were carried out. All those measurements were performed using the network analyzer Agilent 4395A, for H.f / and S11 and the impedance analyzer 4394A, for Zin. The frequency band is taken from 4.3125 kHz to 2.208 MHz, where each measurement has 512 points. Each quantity was measured five times for each line and the signal used in the tests is the average value of these measurements in order to average out the noise influence. Baluns were used in all measurements to properly carry out the connection between equipments and cables. The WES inputs, h.t/ and r.t/, were obtained as described in Section 2. All the thresholds and constants present in the method were selected from the best results of a previous training stage, not the just described test set. The training and the validation sets were composed of signals simulated using the VUB0 model [18]. The constant of sensibility ks defined for analysis of h.t/ and r.t/ is defined as equal to 1. The threshold for inclination of the tangent in the detected singularities in r.t/, , is defined as equal to 0.02. The velocity of propagation is set to 68.7% of the speed of light in vacuum. All the tests were processed using a computer with a CPU Intel Core 2 Duo E7400, 2.80 GHz, with 3 GB of RAM. 5.2. Baseline comparison This work provides a comparison of WES results with the ones obtained by two state-of-the-art methodologies presented in Section 1: TIMEC, described in [16] and SELT-tdr, described in [8, 9]. These two methodologies were first compared in [16] using simulated signals. Both presented good results for different topologies; however, TIMEC obtained better results for the simulated data. For the current work, a comparison using measurements obtained using real cables is proposed. The original implementation of the TIMEC presented in [16], kindly provided by the authors, was used. The method SELT-tdr was also implemented in [16] following the process described in [8, 9], Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac A WAVELET-BASED EXPERT SYSTEM FOR DSL TOPOLOGY IDENTIFICATION Table I. Parameters of the test lines and results estimated by WES and state-of-the-art methods. Here, the letter ‘S’ represents the ‘serial section’ and ‘BT’ represents the ‘bridge tap’. The best estimated results for each parameter are highlighted by bold characters. Line parameters WES SELT-tdr TIMEC Line section �k lk (m) �k �k lk (m) �k �k lk (m) �k �k lk (m) �k 1 �1 S 400 A S 409.6 A S 390.6 B S 374.9 A 2 �1 S 500 A S 500.7 A S 477.4 B S 478.0 A 3 �1 S 1000 A S 1001.4 A S 976.4 B S 1016.6 A 4 �1 S 400 B S 409.6 C S 368.9 A S 101.1 A �2 - - - - - - - - - S 312.9 C 5 �1 S 500 B S 500.7 B S 455.7 A S 501.9 A 6 �1 S 1000 B S 1001.4 C S 1084.9 A S 718.6 C �2 - - - - - - - - - S 291.8 A 7 �1 S 500 A S 542.4 A S 976.4 B S 485.6 A �2 S 500 B S 459.0 B - - - S 543.9 C 8 �1 S 500 A S 558.8 A S 1171.7 B S 559.8 A �2 S 1000 B S 988.7 B S 303.8 D S 144.4 B �3 - - - - - - - - - S 853.7 C 9 �1 S 1000 A S 1053.0 A S 1367.0 B S 1210.3 A �2 S 400 B S 358.0 B - - - BT 686.0 B �3 - - - - - - - - - S 3467.2 D �4 - - - - - - - - - S 1069.4 C �5 - - - - - - - - - S 513.1 B 10 �1 S 200 A S 363.0 B S 195.3 A S 571.9 A �2 S 1000 B S 1366.6 A S 86.8 C S 184.5 B �3 S 500 A - - - - - - S 1013.7 C 11 �1 S 500 A S 546.2 A S 477.4 B S 506.0 A �2 BT 200 A BT 204.8 A - - - BT 213.7 B �3 S 1000 B S 1001.4 B - - - S 1052.0 C 12 �1 S 200 A S 733.1 B S 672.6 A S 697.9 A �2 S 500 B - - - - - - - - - �3 BT 500 A BT 523.4 A BT 499.1 D BT 499.8 A �4 S 500 B S 495.9 A S 520.8 D S 503.4 D WES, wavelet-based expert system; SELT-tdr, single-ended line testing-time domain reflectometry. except by the use of derivatives to detect the intervals of analysis once this process is not described in the original papers. This implementation is used here. The inputs to TIMEC are S11.f / and H.f /. Hence, as the proposed method, TIMEC relies on both single-ended and double-ended measurements. The SELT-tdr uses only TDR trace (SELT) as input, which was obtained here from the input impedance as described in Section 2. 5.3. Results and system evaluation Table I summarizes the results obtained using WES and the baseline methods for the topologies in the test set. The best estimated results for each parameter are highlighted by bold characters. The error in the topology identification e.T / is defined as the number of lines where T was not perfectly estimated over the total number of topologies in the test set. The WES estimated correctly 10 of 12 topologies, resulting in e.T / D 16:67%. The two errors correspond to serial sections missed in lines 10 and 12. The parameter e.nSS/ is the error in the estimation of the number of serial sections nSS in the line under test. For instance, line 12 has nSS D 3 and the number of bridge taps nBT is 1. WES identified correctly the number of bridge taps in all the cases resulting Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac V. D. LIMA ET AL. Table II. Error obtained in the estimation of T for each evaluated technique. Parameters WES SELT-tdr TIMEC e.�/ 16.7% 41.7% 41.7% e.nSS/ 16.7% 41.7% 41.7% e.nBT / 0.0% 8.3% 8.3% PT 0.2 s 3.2 s 1823.2 s WES, wavelet-based expert system; SELT-tdr, single-ended line testing-time-domain reflectometry. Figure 5. Comparison among the estimates for LT of each test line for WES and the two baseline methods. in e.nBT/ D 0%. The performance of the WES was better than TIMEC and SELT-tdr in all cases for identification of T . TIMEC and SELT-tdr presented similar results in these parameters, resulting from different lines in different ways. The TIMEC presented the tendency of introducing sections, whereas SELT-tdr missed the sections. The worst case for TIMEC was in line 9, where the method inserted three wrong sections. SELT-tdr missed two sections in line 11. These results are summarized in Table II, where a comparison among the average processing time PT , measured in seconds, that each technique needed to process the measurements is also presented. These figures are only approximate because the software code was not optimized. However, they indicate the complexity of the compared methods. The results showed that TIMEC was the most time-consuming technique. WES and the SELT-tdr were less time-consuming, with WES being an order of magnitude faster than SELT-tdr. Figure 5 presents the results of the estimation of LT for each test line through the comparison among the percentile error in the LT estimation ep.LT/ given by ep .LT/ D �ˇ̌ ˇ OLT �LT ˇ̌ ˇ � L T �100; (21) where OLT is the estimated value of LT . Lines 1 to 3 and 4 to 6 are the same except by the cable types. For WES, in relation to the identification of T and estimation of L, the difference between the two types of cable was irrelevant. The results were exactly the same comparing 1 with 4, 2 with 5, and 3 with 6. Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac A WAVELET-BASED EXPERT SYSTEM FOR DSL TOPOLOGY IDENTIFICATION Table III. Error obtained in the bridge tap identification for each evaluated technique. The best estimated results for each parameter are highlighted by bold characters. Line Parameter WES (%) SELT-tdr TIMEC (%) 11 e.lBT / 2.40 - 6.85 e.pBT / 9.24 - 1.20 12 e.lBT / 4.68 0.18 0.04 e.pBT / 4.73 3.91 0.30 WES, wavelet-based expert system; SELT-tdr, single-ended line testing- time-domain reflectometry. With exception of lines 1, 8, and 12, WES presented the best results for the estimation of LT . The ep.LT/ value for WES was 1.40% (15.79 m), whereas for TIMEC, excluding the result of line 9 was 2.92% (27.89 m), and SELT-tdr, excluding the results of lines 10 and 11 was 4.13% (30.36 m). The bridge tap identification is evaluated using the absolute error in the estimation of the length and the position of the bridge tap, respectively, given by e � OlBT � D ˇ̌ ˇOlBT � lBT ˇ̌ ˇ lBT �100; and e. OpBT/ D j OpBT �pBT j pBT �100; (22) where OlBT and lBT are, respectively, the estimated and the real bridge tap length, and OpBT and pBT are, respectively, the estimated and the real bridge tap position. Table III summarizes the results obtained using WES and the baseline methods for the bridge tap identification. The best estimated results for each parameter are highlighted by bold characters. Wavelet-based expert system and TIMEC were able to successfully identify the bridge taps in lines 11 and 12, whereas SELT-tdr missed the bridge tap in line 11. However, TIMEC found an inexistent bridge tap in line 9. Except by the bridge tap length of line 11, TIMEC presented the best results for bridge tap identification. Overall, the errors for the three methods were smaller than 10%. In relation to the classification of ˆ, WES estimated the correct cable type sequence in eight cases, whereas TIMEC was well succeeded in three cases. SELT-tdr missed the sequence for all lines. However, considering now the 22 line sections present in the 12 test lines, TIMEC classified 13 line sections with the right type of cable, whereas WES obtained 14 classifications well succeeded. SELT-tdr correctly classified two sections. 6. CONCLUSIONS An expert system for line topology identification based on wavelet analysis was proposed in this paper. The method is based on the examination of the impulse response (estimated via H.f /) and TDR trace (that can be obtained directly or from Zin) of a line under test. The signals are segmented by an edge detection method based on wavelets. The detected edges are analyzed and the main features are extracted from the signals using two algorithms specifically designed for r.t/ and h.t/, respectively. The signal features are interpreted by an expert system composed of three sequential modules that estimate, respectively, the type (serial or bridge tap), the length, and the cable type of the line sections. A comparison between WES and two state-of-the-art methods was presented. From these results, advantages of the proposed technique could be noticed such as the simplicity of its implementa- tion and low computational cost. Besides the lower cost, the majority of results achieved by WES were considerably more accurate than the other two methods. It is important to take into account that TIMEC and WES use SELT and DELT measurements, whereas SELT-tdr uses only SELT measurements. For the used data, the WES processing time varied between 0.17 and 0.22 s. Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac V. D. LIMA ET AL. These results demonstrate that WES has the ability to evaluate a large number of lines of a net- work in a relatively short time with good accuracy. This may be important to speed up the user feedback after a query regarding the line under test. Furthermore, because of the simplicity of the technique implementation and low computational cost, WES can be considered to be embedded on distribution points, which would be very useful for VDSL2 and G.fast, where DSL transmission equipment is not in the CO. Although this method has been tested here for twisted-pair copper cables used for DSL communication, its application is not restricted to this type of transmission line. Finally, WES is composed of three separated modules and it can be easily complemented by other techniques. For instance, if the sets T and L were identified, any other methodology can be used to classify the cable types. ACKNOWLEDGEMENTS This work received financial support from the Research and Development Centre, Ericsson Telecomuni- cações S.A., Brazil, and CAPES, Brazilian Ministry of Education. REFERENCES 1. World broadband statistics: a short report from global broadband statistics - Q1 2012. Technical Report , Point Topic Ltd: London, UK, 2012. 2. Golden P, Dedieu H, Jacobsen K. Fundamentals of DSL Technology. Auerbach Publications: Boca Raton, FL, USA, 2006. 3. T’Joens Y, Bostoen T, Bosch SVd, Vandaele P. Managing home and access domains in modern broadband networks. Bell Labs Technical Journal 2008; 13(1):247–262. DOI: 10.1002/bltj.20293. 4. Bostoen T, Boets P, Zekri M, van Biesen L, Pollet T, Rabijns D. Estimation of the transfer function of a subscriber loop by means of a one-port scattering parameter measurement at the central office. IEEE Journal on Selected Areas in Communications 2002; 20(5):936–948. DOI: 10.1109/JS{AC}.2002.1007376. 5. Rodrigues R, Sales C, Klautau A, Ericson K, Costa J. Transfer function estimation of telephone lines from input impedance measurements. IEEE Transactions on Instrumentation and Measurement 2012; 61(1):43–54. DOI: 10.1109/TIM.2011.2157431. 6. Galli S, Waring DL. Loop makeup identification via single ended testing: beyond mere loop qualification. IEEE Journal on Selected Areas in Communications 2002; 20(5):923–935. DOI: 10.1109/JS{AC}.2002.1007375. 7. Galli S, Kerpez KJ. Signal processing for single-ended loop make-up identification. Proceedings of the 6th IEEE Workshop on Signal Processing Advances in Wireless Communications, New York City, New York, USA, 2005; 368– 374. DOI:10.1109/SPAWC.2005.1506049. 8. Galli S, Kerpez KJ. Single-ended loop make-up identification-part I: a method of analyzing tdr measurements. IEEE Transactions on Instrumentation and Measurement 2006; 55(2):528–537. DOI: 10.1109/TIM.2006.870134. 9. Kerpez KJ, Galli S. Single-ended loop-makeup identification - part II: improved algorithms and performance results. IEEE Transactions on Instrumentation and Measurement 2006; 55(2):538–549. DOI: 10.1109/TIM.2006.870136. 10. Vermeiren T, Bostoen T, Boets P, Chebab XO, Louage F. Subscriber loop topology classification by means of time-domain reflectometry. Proceedings of the IEEE International Conference on Communications, ICC’03, Vol. 3, Anchorage, Alaska, USA, 2003; 1998–2002. DOI: 10.1109/ICC.2003.1203949. 11. Boets P, Bostoen T, van Biesen L, Pollet T. Measurement, calibration and pre-processing of signals for single- ended subscriber line identification. Proceedings of the 20th IEEE Instrumentation and Measurement Technology Conference, IMTC ’03, Vol. 1, Vail, Colorado, USA, 2003; 338–343. 12. Boets P, Bostoen T, van Biesen L, Gardan D. Single-ended line testing: a white box approach. Proceedings of the 4th IASTED International Multi-Conference on Wireless and Optical Communications, Banff, AB, Canada, 2004; 393–398. 13. Boets P, Bostoen T, van Biesen L, Pollet T. Preprocessing of signals for single-ended subscriber line testing. IEEE Transactions on Instrumentation and Measurement 2006; 55(5):1509–1518. DOI: 10.1109/TIM.2006.880290. 14. Bharathi M, Ravishankar S. Single ended loop topology estimation using FDR and correlation TDR in a DSL modem. Cyber Journals: Multidisciplinary Journals in Science and Technology, Journal of Selected Areas in Telecommunications (JSAT) 2012; 2(6):40–48. 15. Neus C. Reflectometric analysis of transmission line networks. PhD Thesis, Vrije Universiteit Brussel, 2011. 16. Sales C, Rodrigues RM, Lindqvist F, Costa J, Klautau A, Ericson K, i Riu JR, Brjesson PO. Line topology identifica- tion using multi-objective evolutionary computation. IEEE Transactions on Instrumentation & Measurement 2010; 59(3):715–729. 17. Kerpez KJ. Automated loop identification on DSL lines. International Journal of Communication Systems 2009; 22(12):1479–1493. DOI: 10.1002/dac.1029. Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac A WAVELET-BASED EXPERT SYSTEM FOR DSL TOPOLOGY IDENTIFICATION 18. Boets P, Zekri M, van Biesen L, Bostoen T, Pollet T. On the identification of cables for metallic access networks. Pro- ceedings of the 18th IEEE Instrumentation and Measurement Technology Conference, IMTC’01, Vol. 2, Budapest, Hungary, 2001; 1348–1353. DOI: 10.1109/IMTC.2001.928292. 19. Test procedures for digital subscriber line (DSL) transceivers. International Telecommunication Union (ITU) recommendation G.996.1, 2001. 20. Time Domain Reflectometry Theory. Application Note 1304-2. Agilent Technologies, Inc.: USA, 2006. 21. Neus C, Boets P, van Biesen L. Feature extraction of one port scattering parameters for single ended line testing. XVIII IMEKO World Congress, Rio de Janeiro, Brazil, 2006; 6 pages. 22. Ruinskiy D, Lavner Y. An effective algorithm for automatic detection and exact demarcation of breath sounds in speech and song signals. IEEE Transactions on Audio, Speech, and Language Processing 2007; 15(3):838–850. 23. Ng B, Ab-Rahman MS, Premadi A. Development of monitoring system for FTTH-PON using combined ACS and SANTAD. International Journal of Communication Systems 2010; 23:429–446. DOI: 10.1002/dac.1078. 24. Nes PG. Fast multi-scale edge-detection in medical ultrasound signals. Signal Processing 2012; 92(10):2394–2408. DOI: 10.1016/j.sigpro.2012.02.021. 25. Luo GY. On-line wavelet filtering of narrowband noise in signal detection of spread spectrum system for location tracking. International Journal of Communication Systems 2012; 25(5):598–615. DOI: 10.1002/dac.1278. 26. Lv S, Liu J. A novel signal separation algorithm based on compressed sensing for wideband spectrum sensing in cognitive radio networks. International Journal of Communication Systems 2013. DOI: 10.1002/dac.2495. 27. Bhuyan MK, Chakraborty BK. Motion adaptive video coding scheme for time-varying network. International Journal of Communication Systems 2013. DOI: 10.1002/dac.2712. 28. Huang H, Huang S, Chen J, Wang R, Xiong J. An image information hiding algorithm based on grey system theory. International Journal of Communication Systems 2013. DOI: 10.1002/dac.2551. 29. Mallat S. A Wavelet Tour of Signal Processing (2nd edn). Academic Press: San Diego, CA, USA, 1999. 30. Borges G, Rodrigues RM, Sales C, Ericson K, Costa J. Cable parameters identification for DSL systems. Proceedings of IEEE International Conference on Computer as a Tool, EUROCON’11, Lisbon, Portugal, 2011; 1–4. Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014) DOI: 10.1002/dac A wavelet-based expert system for digital subscriber line topology identification SUMMARY INTRODUCTION CHARACTERISTICS OF THE MEASURED SIGNALS Interpreting impulse responses Interpreting TDR traces Serial to serial Serial or bridge tap to open termination Serial to bridge tap WAVELET-BASED FEATURE EXTRACTION Singularity detection with wavelets Calculus of the CWT Local modulus maxima and thresholding Singularity detection Classification of estimated singularities KNOWLEDGE-BASED LINE TOPOLOGY IDENTIFICATION Identifying the set T Estimating the set L Algorithm for serial sections Algorithm for bridge taps Classification of the set RESULTS AND ANALYSIS General conditions of evaluation Baseline comparison Results and system evaluation CONCLUSIONS REFERENCES 
work_uhbwuojxfrhnjgyzt5yapiczx4	----	doi:10.1016/j.eswa.2003.09.002 Discovering role-relevant process-views for disseminating process knowledge Minxin Shen, Duen-Ren Liu* Institute of Information Management, National Chiao-Tung University, 1001 Ta Hsueh Road, Hsinchu 300, Taiwan, Abstract Workflow technology supports the management of organizational process knowledge. However, workers (representing organizational roles) cannot easily obtain a global view of a complex and large workflow. This work proposes a process-view based dissemination of process knowledge to address this problem. A process-view is an abstracted process derived from a physical process, and can provide adaptable task granularity. The relationships among tasks, roles and operations, as specified in role-based access control systems, are used to evaluate the degrees of relevance between roles and tasks. Next, a novel algorithm is proposed to generate automatically role-relevant process-views based on degrees of relevance. Knowledge management systems can thus disseminate abstracted process knowledge through process-views for those roles in a complex workflow. q 2003 Elsevier Ltd. All rights reserved. Keywords: Process knowledge; Knowledge management; Knowledge dissemination; Workflow management 1. Introduction Knowledge management (KM) helps organizations to utilize their intellectual capital to gain sustainable competitive advantages (Davenport & Prusak, 1998; Edvinsson & Malone, 1997; Liao, 2002). An organization has various types of knowledge, including for example tacit knowledge, explicit knowledge, and cultural knowl- edge. Process knowledge refers to knowledge of business processes. Knowledge Management Systems (KMS) sup- port KM activities by integrating information and communication technologies. As an effective process management tool, workflow management systems (WfMS) allow a business to analyze, simulate, design, enact, control and monitor general business processes (Georgakopoulos, Hornick, & Sheth, 1995; Leymann & Altenhuber, 1994). Therefore, managing process knowl- edge through WfMS or workflow components of KMS is straightforward. Appropriate knowledge representation is crucial for establishing a successful KMS. Despite notational differ- ences, activity-based methodologies (Georgakopoulos et al., 1995) are extensively used process modeling techniques for representing process knowledge. A typical activity-based approach designs a workflow using a top- down decomposition procedure. This stepwise refinement allows knowledge engineers or process modelers to define a process more easily and completely than do one-step approaches. In practice, workflow participants possess different needs and types of authority when obtaining information on business processes. For example, a high-level manager may need aggregated information of a process, and a marketing manager may not have the authority and need to know each specific step of production flow. However, the activity- based model cannot flexibly alter task granularity of a process based on role characteristics. Therefore, WfMS may not provide each organizational level and unit with an appropriate view of that process. This work aims to develop a method that can disseminate process knowledge at customized granularity to workers. Delivering relevant and necessary documents to workers in order to accomplish their tasks in a workflow environ- ment has been addressed by (Abecker, Bernardi, Maus, Sintek, & Wenzel, 2000; Staaba & Schnurr, 2000). However, these studies suggested task-specific information, rather than process-oriented knowledge, such as of the progress status of an entire workflow. Our previous study (Liu & Shen, 2003) described a process-view model that 0957-4174/$ - see front matter q 2003 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2003.09.002 Expert Systems with Applications 26 (2004) 301–310 www.elsevier.com/locate/eswa * Corresponding author. Tel: þ886-35712121; fax: þ886-35723792. E-mail address: dliu@iim.nctu.edu.tw (D.-R. Liu). http://www.elsevier.com/locate/eswa enhances the ability of conventional activity-based process models to abstract processes, and can present a process definition at adaptable task granularity. A process-view, i.e. a virtual process, is abstracted from an actual process. According to the requirements of distinct organizational roles, a process modeler can design various process-views, providing the process information suitable for each participant. However, the preliminary process-view model requires process modelers (domain experts) to define various process-views for distinct roles. Process-view design is complex and time-consuming. Therefore, this work proposes a novel approach to assist the discovery of role-relevant process-views and applies the results to disseminate process knowledge. This work first presents a quantitative method for evaluating the degrees of relevance between tasks and organizational roles. Different roles may perform different workflow operations on tasks, under the restriction of permission rules in role-based access control systems (Ferraiolo, Sandhu, Gavrila, Kuhn, & Chandramouli, 2001). Thus, permission rules, which prescribe the author- ization relationships among roles, tasks and operations, and audit logs of task execution are utilized to measure the degrees of relevance between roles and tasks. Based on the relevance degrees and the granular threshold, novel algorithms are proposed herein to derive role-relevant process-views. Process designers or workers specify the granular threshold to control the granularity of generated process-views. Thus, during workflow enactment, KMS or WfMS can disseminate process knowledge at a suitable granularity to participants; that is, more relevant parts are presented with a finer resolution and less relevant parts are presented more coarsely. The rest of this paper is organized as follows. Section 2 presents the process-view based dissemination of process knowledge. Formal definitions and the means of construct- ing an order-preserving process-view, as presented in (Liu & Shen, 2003), are also summarized. Next, Section 3 presents the procedure for evaluating relevance degrees between roles and tasks. The algorithms for discovering role-relevant process-views are also presented. Section 4 then discusses and compares the process-view based knowledge dissemination with related work. Section 5 draws conclusions. 2. Dissemination of process knowledge A process-view, an abstracted process derived from a physical process, reveals abstracted process information. Process designers can define various role-relevant process- views, according to the characteristics of participating roles, to achieve different levels of information concealment. The following presents process-view based dissemination of process knowledge, and then formally defines base processes, process-views and order-preserving process- views. 2.1. Process-view based knowledge dissemination A process that may have multiple process-views is referred to herein as a base process. A process-view is generated from either base processes or other process-views and is considered a virtual process. From the users’ perspective, a process-view resembles a typical process that consists of activities (tasks) and dependencies although it is an abstracted form of an implemented process. Fig. 1 illustrates process-view based dissemination of process knowledge. Assume that the base process shown in Fig. 1 is a manufacturing process. Marketers do not need to know every step in the process, although they must know the progress of order fulfillment to serve their customers. A process modeler can design an appropriate process-view for the marketing department as follows: a1; a2; and a3 are mapped into va1; a4 and a5 are mapped into va2; a6 and a7 are mapped into va3 (i.e. va1 ¼ {a1; a2; a3}, va2 ¼ {a4; a5}, va3 ¼ {a6; a7}). Thus, WfMS or KMS can disseminate process knowledge at the granularity suitable for marketers to serve their customers. 2.2. Basic definitions: base process and process-view Fig. 2 illustrates how the components of our model are related. To differentiate the terminology used in base process and process-view, this work uses the terms virtual activity/dependency for the process-view while the terms base activity/dependency are used for the base process. A virtual activity is an abstraction of a set of base activities and corresponding base dependencies. A virtual dependency connects two virtual activities in a process-view. Fig. 1. Process-view based dissemination of process knowledge. M. Shen, D.-R. Liu / Expert Systems with Applications 26 (2004) 301–310302 Fig. 3 shows an example of base process, where the split and join structures are defined by the Workflow manage- ment coalition (WfMC) (Workflow management coalition, 1998). AND-split: An activity splits into multiple parallel activities that are all executed. XOR-split: An activity splits into multiple mutually exclusive alternative activities, only one of which is followed. AND-join: Multiple parallel executing activities join into a single activity. XOR-join: Multiple mutually exclusive alternative activities join into a single activity. The following summarizes basic definitions of base process and process-view. Please refer the work of Liu and Shen (2003) for detailed definitions, semantics and examples. Definition 1. (Base process) A base process BP is a 2-tuple kBA, BDl, where 1. BD is a set of dependencies. A dependency is denoted by dep (x; y; C). Condition C represents the constraints, such as time or events, that determine whether routing can proceed from activity x to activity y: 2. BA is a set of activities. An activity is a 3-tuple kSPLIT_flag; JOIN_flag; SCl; where (a) SPLIT_flag=JOIN_flag may be ‘NULL’, ‘AND’, or ‘XOR’. NULL indicates that this activity has a single outgoing/incoming dependency (Sequence). Given mul- tiple outgoing dependencies, AND indicates all succeed- ing branches are followed (AND-split), while XOR indicates only one succeeding branch is followed (XOR- split). Given multiple incoming dependencies, AND indicates that this activity can be started if all incoming dependencies have satisfied condition (AND-join), while XOR indicates that this activity can be started if one of the incoming dependencies has satisfied condition (XOR-join). (b) SC is the starting condition of this activity. A workflow engine evaluates SC to determine whether this activity can be started. If JOIN_flag is NULL, SC equals the condition associated with its incoming dependency. If JOIN_flag is AND/XOR, SC equals Boolean AND/XOR combination of the con- ditions of all incoming dependencies. 3. For x; y [ BA: (a) If there is a path from x to y in BP; then the ordering of x is higher than y; i.e. x precedes y: Their ordering relation in BP is denoted by x . y or y , x: (b) If no path exists from x to y or from y to x in BP, then x and y are ordering independent, i.e. x and y proceed independently. Their ordering relation in BP is denoted by x1y: Definition 2. (Process-view) A process-view is a 2-tuple kVA; VDl, where (1) VA is a set of virtual activities. (2) VD is a set of virtual dependencies. (3) Analogous to base process, ;vai; vaj [ VA; the ordering relation between vai and vaj may be ‘ . ’, ’ , ’, or ‘1’. 2.3. Order-preserving process-views According to the different properties of a base process, various approaches can be developed to derive a process- view. A novel order-preserving approach to derive a process-view from a base process has been presented by Liu and Shen (2003) and is summarized below. Notably, this summary only presents the case that a base process does not contain loops. Please refer Liu and Shen (2003) for the case that a base process contains loops. The order-preserving approach ensures that the original execution order in a base process is preserved. A legal virtual activity in an order- preserving process-view must follow three rules: member- ship, atomicity, and order preservation rules (Liu & Shen, 2003). Simply, a virtual activity/dependency is an aggrega- tion of a set of base activities/dependencies. The following defines virtual activities and virtual dependencies in an order-preserving process-view. Definition 3. (Virtual activity) For a base process BP ¼ kBA, BDl, a virtual activity va is a 5-tuple kA; D; SPLIT_flag; JOIN_flag; SCl, where 1. A is a nonempty set, and its members follow three rules: a. Members of A are base activities that are also members of BA or other previously defined virtual activities that are derived from BP: b. va is started if one member activity is started, and is completed if all member activities are completed. (Note: this is a simplified expression of the original definition by Liu and Shen (2003)) Fig. 2. Process-view model. Fig. 3. Sample process. M. Shen, D.-R. Liu / Expert Systems with Applications 26 (2004) 301–310 303 c. For any x [ BA; x � A; R [ {, , . ,1}: if existing a y [ A such that x R y holds in BP; then x R z holds in BP for all z [ A: That means the ordering relations between x and all members (base activities) of A are identical in BP: 2. D ¼ {depðx; y; Cxy)l dep (x; y; Cxy) [ BD and x; y [ A}. 3. SPLIT_flag may be ‘NULL’ or ‘MIX’. NULL suggests that va has a single outgoing virtual dependency while MIX indicates that va has multiple outgoing virtual dependencies. 4. JOIN_ flag may be ‘NULL’ or ‘MIX’. NULL suggests that va has a single incoming virtual dependency while MIX indicates that va has multiple incoming virtual dependencies. 5. SC is the starting condition of va. The SPLIT_flag and JOIN_flag cannot simply be described as AND or XOR since va is an abstraction of a set of base activities that may be associated with different ordering structures. Therefore, MIX is used to abstract the complicated ordering structures. A WfMS evaluates SC to determine whether va can be started. Members of A are called va’s member activities, and members of D are called va’s member dependencies. To save space, the abbreviated notation va ¼ kA; Dl is employed below to represent a virtual activity. Definition 4. (Virtual dependency) For two virtual activities vai ¼ kAi; Dil and vaj ¼ kAj; Djl that are derived from a base process BP ¼ kBA; BDl, a virtual dependency from vai to vaj is vdep (vai; vaj; VCij) ¼ {dep (ax; ay; Cxy)l dep (ax; ay; Cxy) [ BD; ax [ Ai; ay [ Aj}, where the virtual condition VCij is a Boolean combination of Cxy: 3. Discovering role-relevant process-views A role-based approach is proposed herein to discover process-views suitable for workflow participants. This section first describes the measurement of relevance degrees between roles and tasks, and then proposes three algorithms to generate automatically process-views. 3.1. Role-task relevance The crucial part of process-view design is finding the relevance degree between each activity (task) and each role in a given base process. Based on these relevance degrees, process designers can define appropriate process-views for different organizational roles. Consequently, WfMS or KMS can disseminate process knowledge at a suitable granularity to participants, i.e. more relevant parts are presented with a finer resolution and less relevant parts are coarser. Process designers can directly evaluate the relevance degrees between tasks and roles. However, inconsistency and bias may occur without criteria for evaluating relevance degrees. Moreover, as the number of tasks grows, the evaluation becomes excessively complex. Besides, knowl- edge embedded in access control systems is not utilized. WfMC has defined several workflow operations for controlling and monitoring task execution (Workflow Management Coalition, 1995). For example, WMFetchAc- tivityInstanceState returns an activity state, and WMReas- signWorkItem reassigns an activity from one participant to another. Moreover, some workflow applications may provide extended operations such as ProgressTracking and PerformerSelection. Whether a role is authorized to perform an operation on a task object is determined by the organization’s security systems, or more specifically, by role-based access control systems (Ferraiolo et al., 2001). Hence, this work applies these different operations as the criteria for evaluating degrees of relevance since their number is finite, and the relationships among tasks, roles and operations are specified by security constraints. 3.1.1. Relevance evaluation based on workflow operations The associations between roles and tasks are established based on workflow operations. Each task is associated with a set of operations that participating roles may perform on it. For example, a role may perform ProgressTracking operation on a task. Therefore, if participating roles evaluate the relevance degrees of these operations, then the role-task associations can be quantified. However, security con- straints are such that roles cannot perform operations on tasks arbitrarily. Accordingly, the relevance degrees between roles and tasks are evaluated based on the usage of workflow operations and the access control policies. Fig. 4 depicts the framework for evaluating role-task relevance. For example, if the relevance degree of role r with respect to operation op is 0.4 (role profile); task t supports operation op (task profile), and r is authorized to perform op on t (permission rule), then the relevance degree between role r and task t is assessed as 0.4. The main components are as follows. Task profile. Different tasks may be associated with different operations. A task profile records which operations can be performed on a given task. Role profile. A role has different degrees of relevance to its authorized operations. A role r’s profile includes a set of pairs koperation op; relevance degree degl, which indicates that the relevance degree of role r with respect to operation op is deg: Permission rules. These rules are derived from access control systems, and govern which role is authorized to perform which workflow operation on which task. Namely, a permission rule krole r; task t; operation opl states that role r is authorized to perform operation op on task t: Task execution log. Each record in this database is a 3- tuple kTaskInstNo, RoleInstNo, operation namel, where M. Shen, D.-R. Liu / Expert Systems with Applications 26 (2004) 301–310304 TaskInstNo/RoleInstNo is a task/role instance identifier. A log record states that a role instance has performed a specific operation on a task instance. Construction of role profile. Initially, role profiles are constructed by analyzing these logs. Let M denote the number of tasks on which operation op is performed, and let N denote the number of tasks on which role r performs operation op: The default degree of relevance of r with respect to op is N=M: For example, if the ProgressTracking operation is performed on 100 tasks, and the IT manager executes 20 s of them, then, the relevance degree of IT manager with respect to the ProgressTracking operation is 0.2. Several factors can also be considered to enrich the role profiles. Intuitively, a higher frequency or cost associated with an operation performed by a role corresponds to a higher relevance degree of the operation to the role. Let Q denote the number of tasks executed by role r: The frequency that r performs operation op is N=Q: Cost can be measured by the time and money consumed by roles in performing operations on tasks. In summary, different statistics can be calculated from historical data. Multi- criteria decision making methods can be adopted to derive the best combination of these statistics. Moreover, relevance feedback from workers further adjusts the relevance values to fit their information needs. Generally, the degree of relevance is more easily described qualitatively in linguistic terms than by numerical quantities. The notion of linguistic variables proposed by Zadeh (1965) may be useful in solving such a problem. A linguistic variable is a variable whose values are words or sentences in a natural language. For example, the values of the linguistic variable ‘relevance’ can be described using five linguistic scales: irrelevant, barely relevant, moderately relevant, very relevant and extremely relevant. Fuzzy numbers, i.e. fuzzy sets (Zadeh, 1975) defined on the real line R, can represent these linguistic scales. Linguistic variables and fuzzy numbers allow process designers to use linguistic measures to assess degrees of relevance, and overcome the vagueness of human decision-making. However, the relevance degree is represented by crisp values herein to keep the focus on the evaluation procedure. The interested reader may refer Shen, Tzeng, and Liu (2003), wherein fuzzy linguistic quantifiers are used to evaluate the appropriateness of workers for performing tasks in workflow systems. Evaluation of role-task relevance. The relevance degrees between a given role and tasks of a workflow can be derived from the role/task profiles and permission rules. For example, given a permission rule krole r; task t; operation op1l, role r’s profile {k op1; 0.4l, k op2; 0.8l, …}, and task t’s profile {k op1; Yl, k op1; Yl,…}, the relevance degree between role r and task t is 0.4. However, if another rule k r; t; op2l is added, then the relevance degree is max(0.4,0.8) ¼ 0.8. That is, the maximum relevance degree is selected when a role is authorized to perform multiple operations on a single task. Finally, process designers use the granular threshold (described in Section 3.2.1) to determine the granularity of generated process-views. The process-view definition tool automatically derives process-views according to role-task relevance and threshold value. Section 3.1.2 presents detailed algorithms. 3.1.2. Relevance evaluation based on organizational authority Different types of organizational authority are alternative evaluation criteria. An organization can be viewed statically and dynamically, as shown in Fig. 5. The design of an organizational structure legitimately distributes authority and responsibility to task holders. The resultant job descriptions specify the relationships among roles, tasks, and authority. Fig. 5 also shows the general format and examples of job descriptions. Therefore, using organiz- ational authority as criteria for evaluating the role-task Fig. 4. Evaluating role-task relevance. M. Shen, D.-R. Liu / Expert Systems with Applications 26 (2004) 301–310 305 relevance may be friendly and meaningful for organiz- ational managers and process designers. Basically, authority is a higher level concept than workflow operations. That is, ‘a role has authority over a task’ means that ‘the role is authorized to perform a set of operations on the task’. Each type of authority corresponds to a set of related workflow operations. Therefore, given a mapping table that states the correspondence between different types of authority and workflow operations, process designers can evaluate role-task relevance based on the exertion of authority. In the previous framework depicted in Fig. 4, role-task associations are established based on the performance of operations. With the mapping table, operation-based associations are transformed straight- forwardly to authority-based operations. 3.2. Algorithms for generating process-views Algorithms are proposed to generate automatically process-views based on the relevance degrees derived as described above. This section first presents three algorithms to generate all legal virtual activities (tasks). Then, the generation of virtual dependencies is discussed. 3.2.1. Generating virtual activities of a process-view Algorithm 1 determines the virtual activity set of a process-view definition from a base process. The objective of this algorithm is to discover virtual activities whose relevance degree approximates the granular threshold TH, as specified by the process designer. The process of generating virtual activities begins with the highest ordering activities in the base process. When the total relevance degree of a set of base activities approximates the granular threshold TH, a virtual activity is found. The above steps are repeated against residual base activities until virtual activities cover all base activities of the base process. Consequently, the virtual activity set of the target process- view is obtained. The following explains four important parts of algorithm 1, including the granular threshold, the total relevance degree, the virtual activity generator (algorithm 2), and the verification of order-preserving property (algorithm 3). Algorithm 1. (The generation of virtual activity set) Granular threshold TH. This parameter determines the granularity of generated process-view. A virtual activity is an aggregation of a set of base activities. When the sum of the relevance degrees of some base activities approximates the threshold value, these activities can form a virtual activity, which is seen as relevant enough to the role. Process designers subjectively specify TH according to their personal experience, expertise, and the application domain. A larger TH corresponds to the generation of fewer virtual activities (and more base activities included in a virtual activity). Notably, TH must be larger than or equal to the maximum relevance degree between roles and tasks (line 3). Total relevance degree FRD( ). Simply, if the function fRD (base activity ba) returns the relevance degree of ba and function FRD (activity set A) returns the total relevance degree of A; then FRDðAÞ ¼ P fRDðaiÞ for all ai [ A: However, ordering structures should be considered. Consider an example in which the threshold value equals 1, a base activity a1 splits into a2 and a3; and their relevance Fig. 5. Different views of an organization and job descriptions. M. Shen, D.-R. Liu / Expert Systems with Applications 26 (2004) 301–310306 degrees to a role are 0.6, 0.3, and 0.4, respectively. Clearly, the three activities do not belong to the same virtual activity, according to the above formula ðfRDða1Þ þ fRDða2Þ þ fRDða3Þ . threshold). However, during run-time, all activi- ties that follow AND-split are executed while only one activity is executed after XOR-split. Thus, if the split structure is XOR-split in the example, then the maximum relevance degree of the three activities is expected to be 0.6 þ 0.4, rather than 0.6 þ 0.3 þ 0.4 or 0.6 þ 0.3. Namely, the total relevance degree FRDðAÞ is the maximum expected value of P fRDðaiÞ for all ai [ A: The above discussion also holds for XOR-join. Algorithm 2. (The generation of a virtual activity whose total relevance degree approximate granular threshold) Virtual activity generation. Algorithm 2 discovers a virtual activity that contains a given base activity and has approximately the maximum degree of relevance below the granular threshold TH. For a virtual activity va ¼ kA, Dl, the members of A must be identified first, since D can be derived from A. Initially, A contains only the given activity a (line 3). Algorithm 2 then determines whether the activities adjacent to members of A can be added to maximize its total relevant degree ðFRDðAÞÞ to approximate TH. A is updated during the while loop (lines 7 , 15) by adding the adjacent activities that cause A to satisfy three conditions: A conforms to the order-preserving property (line 10, see Algorithm 3); total relevance degree of A does not exceed the threshold ðFRDðAtmpÞ # THÞ; A does not overlap with previously derived virtual activities ðAtmp # RASÞ: The repeat-until loop (lines 4 , 16) continues until no other adjacent activity is added to A (line 16, A ¼ TAS), i.e. no more adjacent activity can be added to A while still satisfying the threshold limit and maintaining order-preservation. Following the determination of A, the members of D are those dependencies whose succeeding and preceding activities are both members of A (Definition 3.1). Thus, the virtual activity of va ¼ kA; Dl is generated (lines 17 , 18). Algorithm 3. (The generation of a legal virtual activity) Verification of order-preserving property. For a given activity set AS, algorithm 3 is capable of obtaining the member activities of a minimum and order-preserving virtual activity. According to the definition of a virtual activity (Definition 3.1c), a legal virtual activity va ¼ kA; Dl must satisfy the order-preserving condition: the ordering relations between x and all members of A that belong to BP are identical for any base activity x [ BA and x � A: An activity x; x � AS; is included in A for order preservation. AS is obviously a starting point for identifying x: The algorithm begins from AS; by initializing an activity set TAS that equals AS; to check whether AS is a legal (i.e. order-preserving) virtual activity. If AS is not legal, AS is updated by including activities that violate the order- preserving condition. To determine which of the activities should be added into AS to form a legal and minimal virtual activity, the algorithm considers the activities that are adjacent to members of AS: The algorithm determines whether adjacent activities of AS satisfy the order-preser- ving condition (lines 9 , 12). AS is updated during the while loop (lines 6 , 15), by adding adjacent activities that violate the order-preserving condition. If AS is updated, the repeat-until loop is repeated to check the order-preserving condition. The repeat-until loop (lines 3 , 16) continues to repeat until no more adjacent activity is added into AS (line 16, AS ¼ TAS), i.e. all adjacent activities of AS satisfy the order-preserving condition. Finally, AS is a legal virtual activity. Since this virtual activity conforms to Definition 3, it is a legal virtual activity. Moreover, the algorithm checks the ordering relations from adjacent activities, creating a minimal virtual activity. Naming virtual activities. Since virtual activities are aggregations of base activities, the aggregation must be given a meaning. Relevance degrees can be used to interpret the generated virtual activities. A higher degree of a base activity corresponds to its stronger influence on the formation of the virtual activity. Therefore, the relevance degrees can be used to determine which member activities influence the formation of a virtual activity, and a name or M. Shen, D.-R. Liu / Expert Systems with Applications 26 (2004) 301–310 307 label can be assigned to that virtual activity. However, no absolute guidelines determine how high a relevance degree must be before a given member activity is called influential on the formation of a virtual activity. Usually, only member activities with the top N relevance degrees are considered in the naming of a virtual activity. 3.2.2. Generating virtual dependencies Following the generation of all virtual activities, the virtual dependencies are derived based on Definition 4. Given the virtual activity set VA of a process-view VP derived from a base process BP; whether or not a virtual dependency is associated with two virtual activities can be determined. For any two distinct virtual activities vai ¼ kAi; Dil and vaj ¼ kAj; Djl : if ’ax [ Ai and ’ay [ Aj such that depðax; ay; Cxy) exists in BP; then vdepðvai; vaj; VCij) exists in VP and depðax; ay; Cxy) is a member of vdepðvai; vaj; VCijÞ: After checking each base dependency, all virtual dependencies and their members can be derived. The derivation of VC of virtual dependencies and SPLIT/JOIN_flag of virtual activities are omitted herein and can be found in the work of Liu and Shen (2003). 3.3. Prototype systems This section briefs the architecture of our prototype system as presented by Liu and Shen (2003). The prototype is implemented mainly using the Java 2 platform, Enterprise Edition (J2EE) (Kassem, 2000). J2EE offers rich and uniform application programming interfaces (API) to access diverse business systems. For example, the Java Database Connectivity (JDBC) API gives a vendor-independent interface to different DBMSs. Therefore, the prototype is independent of any specific implementation of a DBMS, directory service, or messaging service. The prototype is built on BEA WebLogic (our J2EE server) (Gomez & Zadrozny, 2000), Microsoft SQL Server (our DBMS) (Vieira, 2000), and Microsoft Active Directory (MSAD, our directory service) (Lowe-Norris, 2000). Fig. 6 shows the system’s architecture. The prototype system uses the Java Naming and Directory Interface (JNDI) API to locate participants that registered in MSAD. The role designer maps role profiles onto the organizational units, groups, and users defined in MSAD. Role profiles are consulted during base/virtual process design. A process modeler employs the role designer and the virtual/base process definition tool to specify workflow participants (internal workers and trading partners), base processes and process-views. The role-based procedure for deriving process-views as elucidated in previous sections is implemented in the virtual process definition tool. During run-time, the enactment module consults MSAD to obtain clients’ responsibilities, and then submits/receives messages to clients’ work-lists or progress coordinators. In this prototype, the enactment module communicates with the clients, the interoperation module, and the invoked applications through the Java Message Service (JMS) API. JMS supports the publish/subscribe asynchronous com- munication mechanism in the process-view model. The client-side work-list presents the activities that must be performed by the client. The progress coordinator displays abstracted progress information of the workflow in which the client participates. Role-relevant process knowledge is disseminated to workers through client-side progress coordinators. 4. Related work 4.1. Dissemination of working knowledge Just-in-time knowledge delivery aims to provide workers with the knowledge required to perform a task when they need it (Cole, Fischer, & Saltzman, 1997). That is, working information is delivered automatically to workers according to their working context (situation) when they are actually performing a task. Context-aware information delivery involves similar ideas, e.g. (Brown & Jones, 2001; Fischer & Ye, 2001). Broadly, these studies characterize infor- mation sources and working context by common features, and then deliver relevant information to workers based on the feature matching between context and information sources. In some information retrieval based approaches, each working context is associated with a query predicate which is a Boolean combination of features used to describe information sources. When workers enter a specific context, a relevant query is executed and resultant information is recommended to them. Generally, these approaches must establish comprehensive metadata. Using similar ideas, several approaches have been developed to deliver relevant and necessary documents to workers to enable them to accomplish their tasks in a workflow environment (Abecker et al., 2000; Reimer, Margelisch, & Staudt, 2000; Staaba & Schnurr, 2000). However, such approaches suggest task-specific infor- mation, rather than process-oriented knowledge. Being informed about a whole process is important since workers are responsible for the outcomes of their participating workflow (Hammer & Champy, 1993). Therefore, this work focuses on providing workflow-scope information. WithFig. 6. Architecture of the prototype system. M. Shen, D.-R. Liu / Expert Systems with Applications 26 (2004) 301–310308 the support of process knowledge, workers can, for example, identify process bottlenecks and monitor progress status. Both workflow and task-scope knowledge are necessary for effective KM. 4.2. Process mining Workflow mining aims to discover process definitions from historical sequences of executed tasks, as described for example in Refs. (Agrawal, Gunopulos, & Leymann, 1998; Datta, 1998; Hwang & Yang, 2002; Weijters & Van der Aalst, 2001). These studies focused on evaluating the correlation among tasks to infer task orderings. Meanwhile, Chun et al. proposed a domain knowledge-based approach to automating the generation of workflow definitions (Chun, Atluri, & Adam, 2002). They assumed that given domain ontologies have defined all possible services and corre- sponding implementations (tasks and their orderings). Thus, a workflow definition is automatically generated based on user-selected services. In contrast to the above references, this work discovers appropriate process abstractions from a given base process definition for workflow participants. Therefore, this work focuses on measuring the relevance between organizational roles and tasks. The definitions of role-relevant process-views are generated from a given base process, according to the resulting relevance degrees. 4.3. Procedure of process-view design Our previous work (Liu & Shen, 2003) focused on process-view and order-preservation. In that work, process- views were designed through an interactive procedure. Process designers select some activities as a seed, and then the process-view definition tool discovers a legal virtual activity from the seed. This step is repeated until all virtual activities are identified. Notably, the selection of seeds is based on a designer’s subjective assessment of the relevance between roles and tasks. This work abandons the above approach and proposes a new framework for evaluating role-task relevance to simplify process-view design. Although only operation usage is considered as an evaluation criterion, other attributes can be easily added to the role/task profile to enable advanced evaluation of relevance. This work also proposes new algorithms to generate automatically process-views based on the results of evaluation. The aim of the proposed algorithm is to discover virtual activities whose relevance degree approximates the granular threshold. The algorithm starts from the first activities (start nodes) of a process and merges base activities into virtual activities. However, several possible approaches to finding process-views remain. For example, base activities whose relevance degree exceeds a threshold constitute individual virtual activities. Then, other base activities are merged into those virtual activities. For another example, the number of required virtual activities is given and then base activities are merged. These approaches are worthy of further exploration. 5. Conclusions and future work This work presents a process-view based dissemination of process knowledge. Task granularity of delivered information is adapted to the needs of workflow partici- pants. Workers can thus obtain helpful views of a large and complex workflow. This work proposes a systematic procedure for measuring the degrees of relevance between roles and tasks to support the discovery of role-relevant process-views. Role-based access control systems formally authorize roles to perform workflow operations on tasks. The relationships among roles, tasks and operations are listed in permission rules. Therefore, this work uses workflow operations as criteria to evaluate role-task relevance. Role-relevant process-views are automatically generated based on evaluation results using the proposed algorithms. Accordingly, KMS or WfMS can supply workers with process-oriented knowledge at the granularity suitable for their organizational role. Future work will address three issues. First, a real case should be considered to validate the proposed approach. Second, only role characteristics (role profile) are con- sidered herein. However, workers may have various features, such as working expertise and experience, so future work should construct user profiles to assist WfMS or KMS in providing more personalized knowledge dissemi- nation. Finally, the process-view model was applied to coordinate inter-organizational workflows (Liu & Shen, 2003). The approach proposed herein may be extended to discover business-to-business process-views by analyzing contracts of cooperation. Acknowledgements This research was supported by National Science Council of the Republic of China under the grant NSC 92- 2416-H-009-010. References Abecker, A., Bernardi, A., Maus, H., Sintek, M., & Wenzel, C. (2000). Information supply for business processes: Coupling workflow with document analysis and information retrieval. Knowledge-Based Systems, 13(5), 271 – 284. Agrawal, R., Gunopulos, D., & Leymann, F. (1998). Mining process models from workflow logs. Proceedings of the 6th International Conference on Extending Database Technology (EDBT’98), Valencia, Spain, March23 – 27. Brown, P. J., & Jones, G. J. F. (2001). Context-aware retrieval: Exploring a new environment for information retrieval and information filtering. Personal and Ubiquitous Computing, 5(4), 253 – 263. M. Shen, D.-R. Liu / Expert Systems with Applications 26 (2004) 301–310 309 Chun, S. A., Atluri, V., & Adam, N. R. (2002). Domain knowledge-based automatic workflow generation. Proceedings of 13th International Conference on Database and Expert Systems Applications (DEXA’02) Aix-en-Provence, France, September 2 – 6, 81 – 93. Cole, K., Fischer, O., & Saltzman, P. (1997). Just-in-Time knowledge delivery. Communications of the ACM, 40(7). Datta, A. (1998). Automating the discovery of as-is business process models: Probabilistic and algorithmic approaches. Information Systems Research, 9(3), 275 – 301. Davenport, T. H., & Prusak, L. (1998). Working knowledge: How organizations manage what they know (1st ed). Boston: Harvard Business School Press. Edvinsson, L., & Malone, M. (1997). Intellectual capital: Realizing your company’s true value by finding its hidden brainpower. NY, USA: Harper Business Publishers. Ferraiolo, D. F., Sandhu, R., Gavrila, S., Kuhn, D. R., & Chandramouli, R. (2001). Proposed NIST standard for role-based access control. ACM Transactions on Information and Systems Security, 4(3), 224 – 274. Fischer, G., & Ye, Y. (2001). Exploiting context to make delivered information relevant to tasks and users. Proceedings of Workshop on User Modeling for Context-Aware Applications at the 8th International Conference on User Modelling, Sonthofen, Germany, July 13 – 16. Georgakopoulos, D., Hornick, M., & Sheth, A. (1995). An overview of workflow management-from process modeling to workflow automation infrastructure. Distributed and Parallel Databases, 3(2), 119 – 153. Gomez, P., & Zadrozny, P. (2000). Professional J2EE programming with be a web logic server. Birmingham: Wrox Press. Hammer, M., & Champy, J. (1993). Reengineering the corporation. NY, USA: Harper Business Publishers. Hwang, S.-Y., & Yang, W.-S. (2002). On the discovery of process models from their instances. Decision Support Systems, 34(1), 41 – 57. Kassem, N. (2000). Designing enterprise applications with the Java 2 platform. Boston: Addison-Wesley, Enterprise Edition. Leymann, F., & Altenhuber, W. (1994). Managing business processes as an information resource. IBM Systems Journal, 33(2), 326 – 348. Liao, S.-h. (2002). Knowledge management technologies and applications- literature review from 1995 to 2002. Expert Systems with Applications, 25(2), 155 – 164. Liu, D.-R., & Shen, M. (2003). Business-to-business workflow interopera- tion based on process-views. Decision Support Systems, (in press). Liu, D.-R., & Shen, M. (2003). Workflow modeling for virtual processes: An order-preserving process-view approach. Information Systems, 28(6), 505 – 532. Lowe-Norris, A. G. (2000). Windows 2000 Active Directory. Cambridge, USA: O’Reilly. Reimer, U., Margelisch, A., & Staudt, M. (2000). EULE: A knowledge- based system to support business processes. Knowledge-Based Systems, 13(5), 261 – 269. Shen, M., Tzeng, G.-H., & Liu, D.-R. (2003). Multi-criteria task assignment in workflow management systems. Proceedings of the 36th Hawaii International Conference on System Sciences (HICSS-36’03), Big Island, Hawaii, USA, January 6 – 9, 6 – 9. pp. 202 – 210. Staaba, S., & Schnurr, H.-P. (2000). Smart task support through proactive access to organizational memory. Knowledge-Based Systems, 13(5), 251 – 260. Vieira, R. (2000). Professional SQL server 2000 programming. Birming- ham, UK: Wrox Press. Weijters, A. J. M. M., & Van der Aalst, W. M. P. (2001). Process mining: Discovering workflow models from event-based data. Proceedings of the 13th Belgium – Netherlands Conference on Artificial Intelligence (BNAIC 2001), Amsterdam, Netherlands, October, 25 – 26. Workflow Management Coalition, The Workflow Reference Model, Technical report WfMC TC-1003, Jan. 19, 1995 Workflow Management Coalition, Workflow Management Application Programming Interface (Interface 2 and 3) Specification, Technical report WFMC-TC-1009, 1998 Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338 – 353. Zadeh, L. A. (1975). The concept of a linguistic variable and its application to approximate reasoning. Parts 1, 2, and 3. Information Sciences, 8(3), 199 – 249. see also: 8(4), 301 – 357, 1975, 9(1), 43 – 80, 1976. M. Shen, D.-R. Liu / Expert Systems with Applications 26 (2004) 301–310310 Discovering role-relevant process-views for disseminating process knowledge Introduction Dissemination of process knowledge Process-view based knowledge dissemination Basic definitions: base process and process-view Order-preserving process-views Discovering role-relevant process-views Role-task relevance Algorithms for generating process-views Prototype systems Related work Dissemination of working knowledge Process mining Procedure of process-view design Conclusions and future work Acknowledgements References 
work_uhgw47pq3fdspbt4y27sfjkela	----	NASA Technical Reports Server (NTRS) 
work_uhxq45b5drcezegsqlqotlo2ga	----	<4D6963726F736F667420576F7264202D20E3CCE1C920C7E1D1C7DDCFEDE420C8C7E1DAD1C8ED2E727466> 2005No.3Vol.13Al-Rafidain Engineering Accepted 30 th March2005 Submitted 5 th May 2004 A COMPUTER EXPERT SYSTEM FOR ANALYSIS AND CONTROL OF WATER HAMMER PROBLEMS Dr. Rasul M. Khalaf Prof. Dr. Alaa H. Kadoury College of Engineering College of Engineering University of Tikrit Univerity of Al-Mustansiria Wisam J. Al-Hilo College of Engineering University of Al-Mustansiria Abstract The current system called Expert System for Analysis and Control of Water Hammer Problems “ESACWHP” is developed to help engineers in the design and analysis of water hammer problems using the characteristic method with aid of the programming language called Visual Basic. Results obtained from the developed of “ESACWHP” show good agreement with that solved by traditional lengthy methods, in addition it is capable of handling more variables. The system is recommended to be used for analysis and design. .. . )Characteristic Method ( )Windows ()Visual Basic ( . . 2005No.3Vol.13Al-Rafidain Engineering 1. Introduction Expert systems are computer programs that contain knowledge in a specific domain. The expert system often uses this knowledge to perform tasks a human can do, but with a longer time. Expert system has become commercially important because it provides the way of archiving expertise and making it routinely available where it is needed. Since early eighties, researchers have developed an entirely new kind of tools, a computer that serves as an assistant whose skills include finding reasonable solutions to problems for which there may be no hard and fast “right” answers. The “expert” computer system uses extensive experience-based knowledge of a subject to guess intelligently in the same way as a human expert does. Before computer analysis, the general equations describing water hammer in pipeline system are simplified in some manner to permit solution by arithmetic, graphical… etc. Matching the boundary condition at pumps and turbines is at best, difficult and understood by relatively few engineers. Modern analysis techniques, including numerical methods of solving partial differential equations, are brought within the capability of solving a wide range of water hammer problems. To protect the piping system against the extreme pressure developed during the transient state, it can be designed with a liberal factor of safety to withstand these pressure changes. Such a design will be uneconomical and various devices are employed to control these changes. Two types of devices are available. The first, such as surge tank, air vessels and bypasses, reduces the rate of net change in initial conditions. The second, such as safety valve, pressure regulating valves and air inlet valves, restricts the minimum or maximum pressure developed during the transient state. A brief description of the two types of devices may be found in Ref. [ ]. The designed expert system, which is constructed to control, and analyze water hammer, is called “Expert System for Analysis and Control of Water Hammer Problems” and in abbreviation,can be written as “ESACWHP”. 2. Theoretical Aspects Befor introducing a brief discription of general equation for water hammer problem, a basic architecture of an expert system is viewed. 2.1 The Structure of an Expert System A human expert uses knowledge and reasoning to arrive at conclusions. Similarly, an expert system relies on knowledge and performance reasoning. The reasoning carried out in expert system attempts to mimic human experts in combining pieces of knowledge. Thus, the structure or architecture of an expert system partially resembles how a human expert performs. Thus, there is an analogy between an expert and an expert system. 2005No.3Vol.13Al-Rafidain Engineering The basic architecture of an expert system consists of three parts: the knowledge base, the inference engine (inference mechanism) and the working memory,as shown in Fig.(1) ,[4]. a. The Knowledge Base (KB) The Knowledge base is the medium through which a human expert’s knowledge is made available to the computer. It contains general problem solving knowledge as well as expert knowledge about how to solve problems or how to interact with the user and it is mostly built into the way the inference engine operates. b. The Working Memory It contains information that the system has received about the problem at hand. In addition, any information the expert system derives about the problem is stored in the working memory. c. The Inference Engine (Inference Mechanism) The inference engine combines facts and rules to arrive at conclusions. But if a large number of facts and rules are matched then how can the systems choose the right facts or rules and the wrong choice may set the system on wild goose chase where most of the reasoning is of little assistance in performing the task. Thus, the system needs method of inference that can select which rule can be applied at each step in the reasoning process. In performing inference, the inference engine tries to establish the truth or facility of a statement called a goal. A goal is a fact whose truth-value is to be determined [2]. 2.2 Water Hammer Considerations The characteristic method is utilized a special property of hyperbolic partial differential equations to find their numerical solutions. For a system of hyperbolic partial differential equations, there are two characteristic directions in the s-t plane (See Fig.2) in which the integration of the partial differential equations is reduced to the integration of a system of ordinary differential equations. The advantages of this method are: it is a method of solution which allows the direct inclusion of friction losses, it offers ease in handling the boundary conditions and in the programming of complex systems, it is a general method, i.e. the program once written and it can be used for analyzing different piping systems having the same boundary conditions, and the transient state conditions, obtained by using this method are close to the actual situation. The restrictions in this method are the flow must be one-dimensional, the wave speed is constant during the transient state, and the time increment is chosen accordingly to satisfy the stability conditions [6]. The general equations for water hammer by characteristics method are: 2005No.3Vol.13Al-Rafidain Engineering RRLLRLRLRLP VVVV D tf VVt a g HH a g VVV 2 sin 2 1 ……….(1) RRLLRLRLRLP VVVV D tf g a VVtVV g a HHH 2 sin 2 1 ……....(2) The boundary conditions and control devices can be found in Refs.[3] and [6] 3. Design and Development of ESACWHP The expert system is constructed as a skeleton of expert system and as a base of information about water hammer using a Visual Basic technique.This system is designed to be simple and easy to handle, using two system units (SI & English unit). In order to provide basic information for the user, tables of elasticity, density…etc, for liquids at different temperatures are included in the system. The expert system contains aid instruments and auxillary softwares (wave speed calculations, graphical representation …etc) to expand the information about water hammer. 3.1 ESACWHP Properties The properties of the current expert system are described by the following steps: 1- ESACWHP depends on the information bases that essentially related to the water hammer theory .Its programs are designed initially using Fortran language and changed later to the Basic language in order to be properly interactive with Visual Basic program. 2- ESACWHP displays the questions through out (42) windows or more. Several windows are designed and categorized to fit the most general and partical problems. ESACWHP is written with more than 20000 program steps , as shown in Figs. (3) and (4). 3- The question windows related to the required information are designed to investigate the aims in article (2) above. 4- Conclusion tool of ESACWHP tells the users about the errors that may be committed by mistake in order to overcome them. It gives warnings to avoid the mistakes, as well as it corrects them whensoever happened. 5- The conclusion tool leads the users to achieve the aims followed in articles (2,3) using the bases of information followed in article (1). 6- When all the calculations and operations are achieved throughout the system without mistakes, the required information would display on windows according to their specific locations in the expert system. 2005No.3Vol.13Al-Rafidain Engineering 3.2 ESACWHP Operation When [Vertical Pump] was choosen and clicking [Ok], a related window will be displayed ,as shown in Fig. (11). When Click [Surge Tank], its window will At the start of the ESACWHP, a window is displayed ,as shown in Fig. (5), containing the name of the program, the producer, and its version. There is a clause “Press any key” at the bottom of the window indicating that the program is ready when any key would be pressed. After that, the main window of the system displays ,as shown in Fig. (6), this window contains the menu bar and the boundary conditions of the system, as listed below: Menu Bar: 1- File 2- Wave speed 3- Graph 4- Setting 5- Help Boundary conditions: 1- Pump Station 2- Surge tank 3- Air chamber 4- Surge valve 5- Dead end 6- Network 7- Valve 8- Atmospheric Discharge When Click [Wave Speed] a corresponding window displays, as in Fig. (7), this window is used to calculate wave speed in fluids. When Click [Graph] a related window displays, as in Fig. (8),this window presents a separated software called WHAM linked to ESACWHP to explain the instant variation of H.G.L. and the discharge for a pipe with a valve installed at the downstream and a reservoir at the upstream, this window includes five options for a valve to be selected, as shown in Fig. (8). When Click [Pump Station], its window will be displayed ,as shown in Fig.(9), containing two types of pump: Horizontal Pump Vertical Pump When [Horizontal Pump] was choosen and associating with click [Ok], a specific window will be displayed ,as shown in Fig. (10). be displayed, as shown in Fig.(12). When Click[Air Chamber], its window will be displayed,as shown in Fig. (13). When Click [Surge Valve], its window will be displayed, as shown in Fig. (14). When Click [Dead End] , its window will be displayed in order to analyze the flow from a reservoir in which the downstream pipe contains dead end or closed valve, as shown in Fig. (15). 2005No.3Vol.13Al-Rafidain Engineering When click [Networks], a corresponding window will be displayed in order to analyze the pipe networks (pipes branch) in each position whereas the user selects, as shown in Fig. (16). When Click [valve], its window will be displayed to analyze the flow from a reservoir in which the downstream pipe contains a valve, as shown in Fig.(17). When Click [Atoms. Discharge], its window will be displayed, as shown in Fig.(18). There are many applications for ESACWHP in order to analyze the water hammer problems in hydraulic systems.The results obtained from these applications that can be found in Ref. [1], were verified by traditional lengthy method and the tests show good agreement. 3.3 Capabilities of ESACWHP ESACWHP used in this study, is discribed by the followings: 1- ESACWHP reduces the time and helps the hydraulic engineers to analyze the network. 2- ESACWHP is easy and clear to use by engineers. Besides it warns the user about the faults throughout the input or within the analysis. 3- ESACWHP can save the results in separate file in order to be used later for further analysis. 4- ESACWHP accepts the future development and is capable of further modifications and improvement to solve other problems which are not considered by ESACWHP. 5- ESACWHP contains a branched programs which work as clarification tool for water hammer recycle. 4. Conclusions and Recommendations From this research context and from the development of “ ESACWHP ”, the followings could be concluded: 1- A number of published and solved problems are analyzed using the expert system “ESACWHP”, the obtained results show a good agreement.For more details see Ref.[ 1] . 2- The expert system gives solutions for most of the main problems caused by water hammer phenomenon, such as pumps power failure and sudden closing of valve. 3- ESACHWP is a friendly software and can be handled by an engineer with a medium level of computer knowledge. The following modifications can be recommended for the expert system: 1- The expert system can be modified to include the analysis of a system at the beginning of the pump operation, because of the expected occurrence of changes in the pressure, (before reaching steady state conditions). 2005No.3Vol.13Al-Rafidain Engineering 2- The expert system can be modified to include the analysis of a system in case of partial operation of pumping sets. 3- The software “ Excel ” can be used to support the expert system “ESACWHP”. This makes it able to represent the results in a graphical fashion. References [1]. Al-Hilo ,W. J.“ A Computer System for Analysis and Control of Water Hammer Problems ” M.Sc. Thesis , Depart. of Environmental Engineering, Al- Mustansiria University ,2002. [2]. Beerel, and Annabel, C., “Expert System Strategic Implications and Applications ”, Ellis Horwood Ltd., U.K., 1987. [3]. Chaudhry, M.H., “Applied Hydraulic Transients ”, 1st Edition, Van Nostrand Reinhold company, New York, 1986. [4]. Hart, A., “Knowledge Acquisition for Expert Systems ” Kogan Page Ltd., 2nd Edition, London, 1989. [5]. Tullis, J.P., “Hydraulics of Pipelines, Pumps, Valve, Cavitation, Transients”, 1st Edition, John Wiley & Sons, New York, 1989. [6]. Watters, G.Z., “Analysis and Control of Unsteady Flow in Pipeline”, 2nd Edition, Butter worth, Chicago, California, 1984. 2005No.3Vol.13Al-Rafidain Engineering t S S 1 S 2 S 3 S N S N+1 t t2 P 1 P 2 P 3 P N P N+1 C+ C - Figure (2) The Characteristic Grid for a Typical Pipe. Figure (1) Expert System Architecture. Expert System Program Inference Engine Working Memory User Output Data Input Data Knowledge Engineer 2005No.3Vol.13Al-Rafidain Engineering Figure (3) Windows of the Expert System. Figure (4) Some Programmable Steps of the Expert System. 2005No.3Vol.13Al-Rafidain Engineering Figure (6)Window Contains the Menu Bar and the Boundary Conditions Figure (5) Window Displays When the System Beginning. 2005No.3Vol.13Al-Rafidain Engineering Figure (7) Window Used to Calculate Wave Speed. 2005No.3Vol.13Al-Rafidain Engineering Figure (9) Window to Select Pump Station Type. Figure (10) Window for Analysis Horizontal Pump Power Failure. 2005No.3Vol.13Al-Rafidain Engineering Figure (11) Window for Analysis Vertical Pump Power failure. Figure (12) Window for Analysis of Surge Tank in Network. 2005No.3Vol.13Al-Rafidain Engineering Figure (13) Window for Analysis Air Chamber in Network. Figure (14) Window for Analysis Surge Valve in Network. 2005No.3Vol.13Al-Rafidain Engineering Figure (15) Window for Analysis Dead End in Network. Figure (16) Window for Analysis Networks (Branched Pipes). 2005No.3Vol.13Al-Rafidain Engineering Figure (17) Window for Analysis Network has Valve at Downstream. Figure (18) Window for Analysis Network has Jet at Downstream. 
work_ujqxiiu7tret3mlyqklik47rym	----	International Journal of Engineering & Advanced Technology (IJEAT) International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-6, August 2020 200 Retrieval Number: F1335089620/2020©BEIESP DOI: 10.35940/ijeat.F1335.089620 Published By: Blue Eyes Intelligence Engineering and Sciences Publication  Abstract: Experts are, generally, busy and few in number. Getting them to work any time and in several locations is difficult, if not impossible, and requires time, cost and stress. Soil stabilization and modification represent an excellent domain for the development of an expert system, which contains a knowledge base of heuristic rules about types and remedies of stabilization. Thus, an Expert system program is a suitable solution together with the all the unpublished knowledge of the experts and combine it into a single accessible program, which could come in handy whenever a novice, student or practitioner engineer need it. The knowledge is extracted from literatures, questionnaire form and interviews with senior pavement engineers (expert), who have long expertise in flexible pavement engineering, especially in stabilization. The knowledge is then represented in knowledge base and encodes to be suitably entered in expert system builder shell (ES BUILDER). The system after being developed is have then to be evaluated by the experts and the users, to evaluate its efficiency and to recommend improvements and offer suggestions to improve the user interface. The overall commands after the evaluation stated that the software was very favorable. The users were very impressed by the system and it's efficient and solutions for the soil stabilization. Keywords: Expert System, Soil Stabilization, Flexible Pavement, ES Builder. I. INTRODUCTION Transportation in the third world countries basically depended on road network rather than on another facilities such as railway, marine and airways. It is therefore expected that this type of network is exposed to the damaging effects resulting specially from axle load of commercial trucks, in addition to the weather effects. Soil is the basis for all types of constructions, such as buildings, roads, and airports, so all of these constructions need soil with certain engineering properties that are compatible with it. In order to extend the useful life of the fixable pavement, fragile and weak soils must be improved and stabilized to obtain high serviceability level and to decrease the rate of soil deterioration. The diagnosis of soil problems and applying the appropriate solution require a significant amount of engineering judgment. This engineering judgment accumulates during long practice and training period to form an expert. In Iraq, there is lack of experts of soil stabilization Revised Manuscript Received on July 20, 2020. * Correspondence Author Mahmood Rashid Mahmood, Civil Engineering, University of Technology, Baghdad, Iraq. E-mail: mahmoudal_qaissy@yahoo.com Ammar A. M. Shubber, Civil Engineering, University of Technology, Baghdad, Iraq. E-mail: 40162@uotechnology.edu.iq Asaad Hamdan Maryoosh*, Civil Engineering, University of Technology, Baghdad, Iraq. E-mail: asaad.hamdan@yahoo.com under flexible pavement, and experience is limited to few experts. There is also lack in the distribution of these experts. Therefore, the development of a system that provides expert consultation in the domain of soil under flexible pavement became a real need. The most valuable asset of the system is its availability, as it can be used at any time or place wherever experience is required, minimizing the necessity of consulting a human expert. II. OBJECTIVE OF THE STUDY Based on the formulation of the problem of the current research, the main objective of the study is formulated in a few stages. The first would be the classification of the soil and its properties under flexible pavement according to the available literature and expert's opinion, alongside their suitable remedies. The construction of a knowledge base that incorporates the gathered information in a form of rules, suitable to be implemented in an expert system environment of -a diagnostic- advisory nature. Then the developing of an expert system for soil stabilization under the flexible pavement that can be provided an advice about soil stabilization and can be used by practicing highway engineers to overcome problem in this domain. After that a validation is needed of the developed system through an extensive evaluation process in a real environment by both experienced and practicing engineers. Finally, determining the system after development to be used as an assistant tool to practicing engineers or instruction system to the novice engineers through the repeated use of the program. III. SOIL STABILIZATION The essential idea of providing a soil with the suitable engineering criteria and/or modified materials; to perform as a foundation for a certain function, is not recent. [1] Humanity had to come with an alternative method, which materialized with the “Macadam” roads, rubble or gravel-topped roads which was used in the 1850s for the first time in the American continent. [2] In the 1900s adding infilling materials was employed for seepage control and dam building especially the gravelly and the rockfill types in Egypt, the gaps between the coarse aggregate was filled with a fine soil to reduce permeability. Then experiments took a high curve, producing the first experiment of the sand-clay mixture in the United States in 1906, this experiment was met with a high skeptically from the engineers at that time, but to meet the high increase of the motor in Europe in the 1930s, vehicle traffic, the wide acceptance of the soil stabilization for road Developing Expert System for Soil Stabilization under Flexible Pavement Mahmood R. Mahmood, Ammar A. M. Shubber, Asaad Hamdan Maryoosh mailto:mahmoudal_qaissy@yahoo.com mailto:40162@uotechnology.edu.iq mailto:asaad.hamdan@yahoo.com Developing Expert System for Soil Stabilization under Flexible Pavement 201 Retrieval Number: F1335089620/2020©BEIESP DOI: 10.35940/ijeat.F1335.089620 Published By: Blue Eyes Intelligence Engineering and Sciences Publication construction was widely endorsed. [3] The purpose of soil and ground improvement is essentially to alter the natural properties of soil (and/or rock) and control the behavior of a geotechnical features or earthwork in order to improve the behavior and performance of a project. Among the properties that are usually targeted for improvement are: [1] 1. Reducing compressibility to avoid settlement. Increasing strength to improve stability, bearing capacity, or durability. 2. Reducing permeability to restrict groundwater flow. 3. Increasing permeability to allow drainage. 4. Lowering the potential of (earthquake-induced) liquefaction. according to [4], the first classification type was based on the character of the techniques and, accordingly, three main categories can be distinguished: 1. Mechanical methods which ensure soil stability without the addition of any foreign material. Thus the soil properties can be improved. 2. Physical methods: which is altering the physical properties of the soil. Chemical methods: The chemical reactions such as ion exchange, Precipitation, Polymerization and Oxidation. IV. EXPERT SYSTEM Nowadays, expert systems are increasingly applied in different domains to solve complex real-world problems, which are traditionally solved using expert human judgment and experience. Many industrial and governmental organizations especially in Europe and the United States are now applying this new technology to solve problems and go towards their projected own objectives. [5] A definitions of many, regarding the expert system is “Programs capable of expert-level-performance in specialized fields”. [6] It is known that Expert System is different than other types of programs, due to the knowledge of the application area of the expert systems being separated logically and physically from the reasoning or inference mechanism. [5] The expert system is also depending on knowledge and capable of performing reasoning. The reasoning that is performed in an expert system, is an attempt to imitate human expert in the combining of the pieces of knowledge. The components of the expert system structure according to [7] are: 1. Knowledge Acquisition Facility. 2. Knowledge Base. 3. Inference Engine. 4. User Interface. Expert system techniques are used in various Fields, such as medical, industrial, economic and civil engineering, and the main application of the express systems can be classified as follows: [8] 1. Diagnosis 2. Selection 3. Advisory 4. Monitoring and Control 5. Interpretation of data 6. Prediction 7. Design 8. Planning V. METHODOLOGY A. Theoretical survey This stage deals with the comprehensive review of specialized books, reports, papers, manuals and other written sources in the field of soil stabilization laying under flexible pavement, and it was performed by the researcher to construct the initial form of the required knowledge base. B. Experience acquisition During this stage detailed and thorough interviews with the selected number of experts were made, to obtain the required domestic knowledge in the diagnosis of the properties of the soil and the remedies for its deficiency. C. Model building The information gathered in the first and second stages were used, after representation and encoding process, to build an expert system that can be easily accessed by practicing highway engineers to obtain the required consultation with minimum constraints of time and cost. D. Model evaluation Final stage, the developed system was subjected to extensive consultation sessions in order to test its efficiency and performance in solving different type of domain problems. VI. PROGRAM IN OPERATION These pictures are examples for the program during operation: Fig. 1. Sample of the tree of the program fed with information International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-6, August 2020 202 Retrieval Number: F1335089620/2020©BEIESP DOI: 10.35940/ijeat.F1335.089620 Published By: Blue Eyes Intelligence Engineering and Sciences Publication Fig. 2. The first step of the program while it asks about the soil type Fig. 3. The part where the user chooses either to use “Educational” or “Expert” Fig. 4. The program ask about the preferences of the soil in question Fig. 5. The conclusion that is given by the program after asking a sufficient number of questions VII. RECOMMENDATIONS The following recommendations could be considered for improving the developed system ES3 as well as improving the ways of stabilization under flexible pavement: 1) There is a need to develop an Iraqi stabilization manual compatible with the available material, equipment, Iraqi traffic, climatic factors and soil types. 2) The quality control for the stabilization procedure and the stabilization materials has to improve. 3) Using the developed “ES3” system as an educational system and a training tool. The educational value is significant in using any kind of expert system; the student will gain some familiarity with an important and rapidly growing computer application. The system will assist the student to understand the logic and the technical background of the stabilization. On the other hand the system can transfer the experts’ knowledge, which are often not available to provide the analytical advice to the novice engineer, thus the program works as a training tool. VIII. CONCLUSION The following conclusions can be drawn from the research field work and the development of “ES3” 1) The general benefits of ES3 may be summarized in the following points: a) Providing expert knowledge on soil stabilization under fixable pavement during construction and its planning for less experienced and inexperienced engineers especially when experts are not available or very expensive to be consulted. in addition, the system can assist to excite the expert's interest with new ideas. b) Time will be saved in waiting for expert to find the best suitable stabilizer for the soil. c) The system can be used as an educational aid (or teaching program) to train the novice engineers on how to use the different methods of stabilization with different soil types. d) The system must be used as a tool to assist users, rather than to replace human expert, because no matter how efficient the expert system is, it is still limited in knowledge base and requires updating. e) The updating process must be done by knowledge engineer or an expert, who is familiar with the expert system shell, and has real new information to modify the existing knowledge base. The knowledge is an accumulation of imagination, judgment, and reasoning which are found in human experts and must be elicited from them first, and then encoded as a computer program. f) The development system works on PC computer as minimum requirement, and can be saved on USB flash, therefore, it is easy to handle and can be used in many location to minimize costs and delays. 2) During the knowledge collection from the expert, it was noticed that there is a lack stabilization techniques because there is no direct Iraqi soil stabilization manual contains the suitable remedies according to the Iraqi traffic, climatic factors and soil types. The developed system may provide such stabilization manual information with more flexible way. There is a lack in expert's distribution. Developing a diagnostic expert system, to be used as a guideline, provided a great assistant. REFERENCES 1. P. G. Nicholson, Soil Improvement and Ground Modification Methods. Butterworth Heinemann is an imprint of Elsevier, 2014. 2. A. Kezdi, Stabilized Earth Roads. Budapest: Elsevier Scientific Publishing Company, 1979. 3. J. A. Charles, “Ground Improvement: The Interaction of Engineering Science and Experience-Based Technology,” Geotechnique, vol. 52, no. 7, pp. 527–532, May 2002, doi: 10.1680/geot.2002.52.7.527. Developing Expert System for Soil Stabilization under Flexible Pavement 203 Retrieval Number: F1335089620/2020©BEIESP DOI: 10.35940/ijeat.F1335.089620 Published By: Blue Eyes Intelligence Engineering and Sciences Publication 4. T. S. Amhadi and G. J. Assaf, “Overview of Soil Stabilization Methods in Road Construction,” Springer, Cham, 2019, pp. 21–33. 5. K. Parsaye and M. Chignell, Expert Systems For Experts. Wiley; 1 edition, 1988. 6. R. J. Allwood, Techniques and Applications of Expert Systems in The Construction Industry. Halsted Press, 1989. 7. Y. Oshima, Information Control Problems in Manufacturing Technology. IFAC International Symposium, Tokyo, Japan, 1978. 8. C. S. Battu, “Design of NIPFI: an Advisory Expert System for R•Leasurement and Improvement of Productivity of Foodservice Industries,” 1989. AUTHORS PROFILE Dr. Mahmood R. Mahmood was born in Diyala, Iraq in 1960; he received the B.Sc. degree in Building and Construction Eng., from the University of Technology, Iraq in 1982, and the M.Sc. degree in soil Mechanics from Herriot -Watt University, Edinburgh-Scotland U.K in 1987. He received his Ph.D degree in Geotechnical engineering from Al-Rasheed College, University of Technology, Iraq in 2001. He is currently an Assist. Professor at the Department of Civil, University of Technology, Iraq. Dr. Ammar A.M. Shubber, was born in Baghdad Iraq in 1972, he received his B.Sc. degree in Building and Construction Engineering, from the University of Technology, Iraq in 1996, and the M.Sc. degree in Highway and Airport engineering from the University of Technology, Iraq in 2000, He received his Ph.D. degree in Roads and Railway engineering at Southwest Jiaotong University, Chengdu, China in 2008. He is currently a Lecturer at the Department of Civil engineering, University of Technology, Iraq. Asaad Hamdan Maryoosh*, he was born in 1985 in Baghdad, Iraq. He got his B.Sc. in civil engineering from University of Technology, Baghdad, Iraq, 2011. And applied to the M.Sc. degree in 2017 in the same university. 
work_ujtg6wuf65dkbcqfvnwz2jnys4	----	Microsoft Word - SOPNN_Optimization.doc 1 Optimization of Self-Organizing Polynomial Neural Networks Author: Ivan Marić Rudjer Boskovic Institute Bijenicka 54 10000 Zagreb Croatia Email: ivan.maric@irb.hr Phone: +385-1-4561191 CORE Metadata, citation and similar papers at core.ac.uk https://core.ac.uk/display/33995641?utm_source=pdf&utm_medium=banner&utm_campaign=pdf-decoration-v1 2 Abstract The main disadvantage of self-organizing polynomial neural networks (SOPNN) automatically structured and trained by the group method of data handling (GMDH) algorithm is a partial optimization of model weights as the GMDH algorithm optimizes only the weights of the topmost (output) node. In order to estimate to what extent the approximation accuracy of the obtained model can be improved the particle swarm optimization (PSO) has been used for the optimization of weights of all node-polynomials. Since the PSO is generally computationally expensive and time consuming a more efficient Levenberg-Marquardt (LM) algorithm is adapted for the optimization of the SOPNN. After it has been optimized by the LM algorithm the SOPNN outperformed the corresponding models based on artificial neural networks (ANN) and support vector method (SVM). The research is based on the meta-modeling of the thermodynamic effects in fluid flow measurements with time-constraints. The outstanding characteristics of the optimized SOPNN models are also demonstrated in learning the recurrence relations of multiple superimposed oscillations (MSO). Keywords Polynomial neural networks GMDH Levenberg-Marquardt algorithm Particle swarm optimization Time series modeling 3 1. Introduction Approximation of complex multidimensional systems by SOPNN, also known as the GMDH polynomial neural networks (PNN), was introduced by Ivakhnenko, (1971). The SOPNN are constructed by combining the low order polynomials into multi layered polynomial structures where the coefficients of the low-order polynomials (generally 2-dimensional 2 nd order polynomials) are obtained by polynomial regression (Chapra & Canale, 1998) with the aim to minimize the approximation error. GMDH models may achieve reasonable approximation accuracy at low complexity and are simple to implement in digital computers (Maric & Ivek, 2011). The GMDH is resistant to over-fitting since it uses separate data sets for regression and for model selection. When applied to real time compensation of nonlinear behavior, the self-organizing nature of GMDH may eliminate the complicated structural modeling and parameterization, common to conventional modeling strategies (Iwasaki, Takei & Matsui, 2003). The performance of the SOPNN is generally evaluated by a single parameter measure (Witten & Eibe, 2005), typically by the least square error, which minimizes the model approximation error rather than its complexity. When building the models for time-constrained applications the constraints can be efficiently embedded into the model selection metrics (Maric & Ivek, 2011). It was shown (Maric & Ivek, 2011) that the raw SOPNN models (GMDH PNN) are inferior to multilayer perceptron (MLP) when considering the accuracy with respect to the complexity. It was also concluded (Maric & Ivek, 2011) that SVM is not appropriate for building the low complexity models for time-constrained applications. The SOPNN node-polynomial weights, after calculated by the regression, remain unchanged during the rest of the training process resulting in sub-optimal SOPNN models. The accuracy and the prediction of models may be improved significantly when trained by the genetic programming and back-propagation (BP). It was shown (Nikolaev & Iba, 2003) that the population-based search technique, relying on the genetic programming and the BP algorithm, enables to identify the networks with good training as well as generalization performances. The BP improves the accuracy of the model but it is known to often get stuck in local minima. The idea of this paper is to adapt a more robust procedure for the optimization of the SOPNN relation with respect to its weights. The PSO is a nature inspired algorithm, which enables the optimization of model weights by simulating the flight of a bird flock (Eberhart & Kennedy, 1995). Since the PSO is simple to implement it has been used in our experiments for the estimation of the approximation abilities of raw SOPNN models. After the PSO improved significantly the approximation accuracy of various SOPNN models a more complex Levenberg-Marquardt (LM) algorithm (Levenberg, 1944; Marquardt, 1963) has been adapted for the optimization of model weights. Although widely used for the optimization of ANN, the use of the LM algorithm for the optimization of weights of the SOPNN to the best of my knowledge has not been reported in the literature. The LM algorithm converges many times faster than the PSO and increases the approximation accuracy of the SOPNN model substantially. This paper describes the adaptation of PSO and LM algorithm for the optimization of SOPNN and demonstrates how the approximation accuracy of the original GMDH model can be significantly improved after optimizing its weights. In the following section, the GMDH, PSO and the LM algorithm are described. The PSO and the LM algorithm are adapted for the optimization of SOPNN weights. Section 3 describes the procedure for the estimation of the execution time for SOPNN and MLP in time constrained applications. A procedure for the compensation of thermodynamic effects in flow rate measurements is summarized in section 4 and the results of the simulations of the flow rate error compensation procedure by the surrogate models are given in section 5. Finally in section 6, the outstanding performances of the SOPNN are demonstrated on MSO task that has been widely studied in echo state networks (ESN) literature. 2. GMDH, PSO and LM algorithm 2.1. GMDH algorithm The GMDH algorithm (Ivakhnenko, 1971) constructs the models by combining the low-order polynomials into multi layered polynomial networks. Fig. 1 illustrates a complete two-layer feed-forward SOPNN representing a 3-dimensional system, where pλ,i denotes a low-order and low-dimensional polynomial corresponding to the i th node of the layer λ and xi represents the i th independent variable. 4 First order and second order two-dimensional polynomials are preferred as their cascading does not increase rapidly the order of overall polynomial relation and because they are fast to process. In this paper we will restrict our attention to a complete second-order two-dimensional polynomial 216 2 25 2 1423121 iiiiiiiiiiiii zzazazazazaap λλλλλλλλλλλλλ +++++= , (1) where zλi1 and zλi2 represent the input variables and aλi1,...,aλi6 are the corresponding weights (coefficients) obtained by the polynomial regression. Note that zλi1 and zλi2 can be any combination of two different variables from lower layers including the independent variables (xi) and the derived regression polynomials (pλi). For example, the input variables for the polynomial p2,6 from Fig. 1 are the polynomial p1,1 from the first layer and the independent variable x3. Fig. 1. The GMDH algorithm assumes two independent data sets: a training set of M samples M 1i }),{( == titit yxD , (2) where each sample consists of a data vector ( ) tiKtititi xxx ,...,, 21 =x , K ti x R∈ , and the corresponding dependent variable R∈ ti y , as well a validation data set of N samples N 1i }),{( == viviv yxD (3) with data vector ( ) viKvivivi xxx ,...,, 21 =x , K vi x R∈ , and the corresponding dependent variable R∈ vi y . The algorithm uses the training data set (2) to fit the coefficients of the regression polynomial (1) and the validation set (3) to verify the approximation error of the polynomial. The polynomials are then ranked according to some predefined metrics (Maric & Ivek, 2011). In order to calculate the coefficients, aλi1,...,aλi6, in (1) by the polynomial regression, a set of 6 simultaneous linear equations ( ) 6 11 2 0 ==       =− ∂ ∂ ∑ k M m tm itm ik py a λ λ (4) must be solved, where M is a total number of training samples, tm y is the m th sample value of the dependent variable from the training data set (2) and tm i pλ is the value of the i th polynomial at layer λ corresponding to the m th data vector from the training data set (t). In our implementation of the GMDH algorithm the above set of linear equations is solved by the Gauss elimination method using forward elimination, back substitution, and pivoting (Chapra & Canale,1998). As illustrated in Fig. 1. the total number of possible nodes in each layer is increasing rapidly by the increase of the layer number and can be easily calculated by the following simple iterative equation 5 ( ) ∑ − = + + −⋅ = 1 0 1 2 1 λ λ λλ λ i i NN NN N , where λ=0,1,… denotes the corresponding layer, Nλ is the total number of nodes at layer λ and the second term in the equation equals zero for λ=0. To prevent the combinatorial explosion the maximum number of nodes retained per layer is generally limited. The nodes retained at lower layers are combined multiple times to produce the nodes at higher layers. To speedup the algorithm the retained polynomials may be tabulated. 2.2. Particle Swarm Optimization of SOPNN Weights Let N ℜ∈a be the vector, {ai}, i=1,...,N, in N-dimensional space. Let the SOPNN polynomial relation P(x,a) be the particle whose position in the space is defined by the vector a. Particle swarm optimization (Eberhart & Kennedy, 1995) minimizes the nonlinear fitness function: ( ) ( )[ ]∑ = −= K k kk Pye 1 2 , axa (5) where K is the total number of training vectors, xk denotes the k th M-dimensional input training vector, yk=y(xk) is the corresponding training value and a denotes N-dimensional vector representing the coefficients of all nodes of the SOPNN polynomial P(xk,a). Before starting the PSO the position a and the velocity v of each particle are initialized by: NiKkLUrLa iikiiki ,...,1,,...,1, 0 ==−+= (6) and ( ) NiKkLUsv iikiki ,...,1,,...,1,12 0 ==−⋅−= (7) where 0 ki a and 0 ki v are the corresponding i th component of the k th particle’s initial position and velocity, Li and Ui are the corresponding lower and upper boundaries for the i th dimension of the search space, and rki and ski denote the random numbers from the range [0,1]. For each particle k, the best particle position Ak and the best fitness value Ek are initialized by setting them equal to the initial position and the initial fitness value of the particle, respectively. The pointer to the particle α having the best fitness value, Eα, of all particles is also initialized. According to (Shi & Eberhart, 1998) the particles are manipulated iteratively by using the following equations: ( ) ( ) NiKkaAScaARcvwv j kiiki j kikiki j ki j ki ,...,1,,...,1̧, 21 1 ==−+−⋅⋅+⋅= + α (8) and NiKkvaa j ki j ki j ki ,...,1,,...,1, 11 ==+= ++ , (9) where j ki v and j ki a denote the corresponding i th component of the k th particle’s velocity and position in j th iteration, w denotes the inertia weight (0.8<w<1.2), c1 and c2 are constants and Rki and Ski are random numbers in the range [0,1]. 2.3. LM Optimization of SOPNN Weights 2.3.1. Adaptation of LM algorithm to SOPNN The optimization problem can be formulated in the following way: given the set of K samples each consisting of N–dimensional input data vector x and the dependent variable y, ( ){ }Kkyxxx kNkkk ,...,1,,,...,, 21 = , minimize the nonlinear fitness function (5) by optimizing all the weights {a1,...,aM} of the SOPNN polynomial relation P. In each iteration step the LM algorithm calculates the increment vector δ, δ1,...,δM, and estimates the new weight 6 vector a=a+δ, a1+δ1,...,aM+δM, in order to decrease the value of the fitness function (5). To calculate the increments, i.e. to estimate the improved polynomial weight vector (a=a+δ) for the next iteration, the polynomial P is calculated by ( ) ( ) ( ) ( ) M M kk kk a P a P PP δδ ⋅ ∂ ∂ ++⋅ ∂ ∂ +≈+ axax axδax , ... , ,, 1 1 , (10) where the polynomial P(xk,a+δ) denotes the new approximation of the dependent variable y. After substituting a+δ for a and (10) for P(xk,a) in (5) we obtain: ( ) ( ) ( ) ( ) ∑ =         ⋅ ∂ ∂ −−⋅ ∂ ∂ −−=+ K k M M kk kk a P a P Pye 1 2 1 1 , ... , , δδ axax axδa (11) By setting the partial derivatives of (11) with respect to increments δj, j=1,...,M, equal to zero, a set of M simultaneous linear equations with M unknowns (δ1,...,δM) is obtained ( ) ( ) ( ) ( ) M j K k M M kk kk j k a P a P Py a P 1 1 1 1 0 , ... , , , = =         =                 ⋅ ∂ ∂ −−⋅ ∂ ∂ −−⋅ ∂ ∂ ∑ δδ axax ax ax (12) or in vector notation: ( ) ( )[ ]aPyJδJJ −= TT (13) where J denotes the Jacobian matrix i.e. the first order partial derivatives of P with respect to a. After introducing an adjustable nonnegative damping factor γ and the diagonal matrix of J T J we obtain the Levenberg-Marquardt equation ( )( ) ( )[ ]aPyJδJJJJ −=⋅+ TTT diagγ , (14) summarizing the set of M linear equations with M unknowns ( ) ( ) ( ) ( ) ( ) ( )∑∑∑∑ ∑∑∑∑ ∑∑∑∑ ==== ==== ==== − ∂ ∂ =        ∂ ∂ +++ ∂ ∂ ∂ ∂ + ∂ ∂ ∂ ∂ − ∂ ∂ = ∂ ∂ ∂ ∂ ++        ∂ ∂ ++ ∂ ∂ ∂ ∂ − ∂ ∂ = ∂ ∂ ∂ ∂ ++ ∂ ∂ ∂ ∂ +        ∂ ∂ + K k kk M k K k M k M K k M kk K k M kk K k kk k K k M kk M K k k K k kk K k kk k K k M kk M K k kk K k k Py a P a P a P a P a P a P Py a P a P a P a P a P a P Py a P a P a P a P a P a P 11 2 1 2 2 1 1 1 1 21 21 2 2 2 1 21 1 1 11 11 21 2 1 2 1 1 1... ...1 ...1 δγδδ δδγδ δδδγ M (15) where Pk denotes the SOPNN polynomial relation P (xk,a). 2.3.2. Calculation of partial derivatives As can be seen from Fig. 1, each layer contains the corresponding number of nodes. The i th node in the layer λ, n(λ, i), λ =1,2,..., is characterized by the corresponding weights { } 61 ,..., ii aa λλ of the basic polynomial (1), where the inputs 1i zλ and 2izλ denote either the polynomials from lower layers or the independent variables. In our algorithm the polynomials and the corresponding partial derivatives are calculated recursively. Partial derivatives of the polynomial Pλi with respect to the weights of its topmost node n(λ, i) are: 7 1 1 = ∂ ∂ i i a P λ λ , 1 2 i i i z a P λ λ λ = ∂ ∂ , 2 3 i i i z a P λ λ λ = ∂ ∂ , 2 1 4 i i i z a P λ λ λ = ∂ ∂ , 2 2 5 i i i z a P λ λ λ = ∂ ∂ and 21 6 ii i i zz a P λλ λ λ = ∂ ∂ , where 1i zλ and 2izλ are the inputs to node n(λ, i). Partial derivatives of the polynomial Pλi with respect to j th weight of any lower layer node n(x, y) are calculated by: xyj xy xy i xyj i a P P P a P ∂ ∂ ⋅ ∂ ∂ = ∂ ∂ λλ , where xy P is the polynomial with the topmost node n(x, y) having the corresponding weights axyj, j=1,…,6, and xy i P P ∂ ∂ λ is the partial derivative of the polynomial with topmost node n(λ, i) with respect to the polynomial with topmost node n(x, y). Note that partial derivative xy i P P ∂ ∂ λ is calculated by a chain rule over all existing connections between the nodes n(λ, i) and n(x, y). The calculation of partial derivatives by chain rule is illustrated for the hypothetical example of the SOPNN shown in Fig. 2. For example, the polynomial P34 in Fig. 2 can be calculated by the following concatenated polynomial relations ( ) ( )( ) ( )( )( ) 3425414321425231432141221122334 ,,,,,,,,,,,, aaaaaa xxxPPxxPxxPPP , where Pλi denotes the polynomial with the topmost node n(λ, i) and aλi is the n(λ, i)-th node weight vector, with the corresponding weights aλij, j=1,…,6. The partial derivative of the polynomial P34 with respect to the weights of the topmost node of the polynomial P14 from Fig. 2 is: 6,...,1, 14 14 14 25 25 34 14 23 23 34 14 34 = ∂ ∂ ⋅        ∂ ∂ ⋅ ∂ ∂ + ∂ ∂ ⋅ ∂ ∂ = ∂ ∂ j a P P P P P P P P P a P jj , where 2534623344342 23 34 2 PaPaa P P ++= ∂ ∂ , 2334625345343 25 34 2 PaPaa P P ++= ∂ ∂ , 1223614235233 14 23 2 PaPaa P P ++= ∂ ∂ , 425614254252 14 25 2 xaPaa P P ++= ∂ ∂ , 1 141 14 = ∂ ∂ a P , 2 142 14 x a P = ∂ ∂ , 3 143 14 x a P = ∂ ∂ , 2 2 144 14 x a P = ∂ ∂ , 2 3 145 14 x a P = ∂ ∂ and 32 146 14 xx a P = ∂ ∂ , with j a 14 , j a 23 , j a 25 and j a 34 denoting the j th weight of the corresponding topmost node of the polynomials P14, P23, P25 and P34, respectively. Fig. 2. 8 2.3.3. LM algorithm in pseudo code In order to speedup the minimization of the fitness function the nonnegative damping factor γ is adjusted at the beginning of each optimization cycle. The algorithm starts with the best polynomial obtained from GMDH algorithm P(xk,a) and calculates the corresponding error e0=e(a) by using (5), sets the maximum allowed damping factor γM, maximum number of iterations iM and initializes the damping factor γ=γ0 and the coefficient ν>1. The following pseudo code illustrates the simplified flow diagram of the LM algorithm (Marquardt, 1963): Begin with: i=0, γ= γ0, e0=e(a) Eq. (5) 1. Set γ1=γ 2. Calculate δ1, and e1=e(a+δ1); Eq. (15) and (11) 3. Set γ2= γ/ν 4. Calculate δ2 and e2=e(a+δ2); Eq. (15), and (11) 5. If e1<e0< e2 Then γ=γ1, e0=e1, a=a+δ1 and GoTo 12. 6. If e2<e0< e1 Then γ=γ2, e0=e2, a=a+δ2 and GoTo 12. 7. If e0<e1 and e0< e2 Then γ=γ1·ν, 8. If γ>γM Then GoTo 13. 9. γ1=γ 10. Calculate δ1, and e1=e(a+δ1); Eq. (15) and (11) 11. If e1<e0 Then γ=γ1, e0=e1, a=a+δ1 12. i=i+1, If i≤iM Then GoTo 1. 13. End The idea of the LM algorithm is very simple but its adaptation to SOPNN is somewhat complicated by the recursive calculation of the polynomials and its derivatives. 3. SOPNN and MLP models for time-constrained applications In certain cases the execution of the procedures based on “first principles” may be unfeasible in real time. In these cases, it is reasonable to construct the corresponding meta-model (Jin, Chen & Simpson, 2000) by maximizing the approximation accuracy for the given computational complexity. To derive a meta-model from the original high-complexity model by machine learning technique it is necessary to generate sufficient training and validation examples from the original model. SVM (Vapnik, Golowich & Smola, 1997) and ANN (Ferrari & Stengel, 2005) can be efficiently used for the approximation of multidimensional nonlinear functions. In measurement and control applications (Maric & Ivek, 2011) it is often necessary to maximize the approximation accuracy for the given computational complexity. The computational complexity can be measured by the corresponding execution time (ET). Our models are tailored for the flow computer based on low computing power microcomputer (8-bit/16-MHz) with implemented floating point (FP) subroutines for single precision addition, multiplication, division and the exponential function. In this section we will briefly summarize the procedure for the estimation of the ET of SOPNN and MLP models in low computing power systems (Maric & Ivek, 2011). 3.1. Computational Complexity of SOPNN The ET of the SOPNN is defined by the total number of polynomial nodes and the maximum ET of the basic low order polynomial (1). The maximum ET of the SOPNN (Maric & Ivek, 2011) can be estimated by ( ) mulmuladdaddSOPNN TNTNNT ⋅+⋅⋅≤ (16) where N denotes the total number of polynomial nodes, while Nadd and Nmul denote the corresponding total number of FP additions and FP multiplications, necessary to calculate the basic polynomial (1), and Tadd = 50 µs and Tmul = 150 µs are the corresponding maximum execution times of the FP addition and FP multiplication. If we rewrite the polynomial (1) by Horner’s rule, its calculation is reduced to Nadd = 5 FP additions and Nmul = 5 FP multiplications i.e. ( ) ( ) 25322614211 iiiiiiiiiiii zaazzazaazap λλλλλλλλλλλλ +++++= . 9 3.2. Computational Complexity of MLP We trained the feed-forward MLP consisting of one output neuron and M-1 neurons in the hidden layer, each neuron employing the sigmoid activation function. The MLP scheme is shown in Fig. 3. The maximum ET of the calculation procedure for the simple MLP with N inputs, one output neuron and M-1 neurons in the hidden layer can be estimated by (Maric & Ivek, 2011) ( )( )( ) ( ) divaddmuladdmlp TTTMTTNMT +++++−≤ exp 11 (17) where Texp, and Tdiv denote the maximum execution time for the FP exponential function and FP division, respectively. The maximum ETs for the above FP operations in the flow computer are: Texp = 3470 µs and Tdiv = 430 µs. Fig. 3. 4. Compensation of flow rate error We investigated the combined effect of the Joule-Thomson (JT) coefficient and the isentropic exponent of a natural gas on the accuracy of flow rate measurements based on differential devices (Maric & Ivek, 2010). The measurement of the flow rate of a natural gas (ISO-20765-1, 2005) flowing in a pipeline through the orifice plate with corner taps (ISO-5167-2, 2003) is illustrated in Fig. 4. Fig. 4. A detailed description of the flow rate equation with the corresponding iterative computation scheme is given in (ISO-5167-2, 2003). The calculation of the natural gas flow rate depends on multiple parameters: 10 ( )dDpTPqq uuuuuu ,,,,,,, κγρ∆= (18) where qu, ρu, γu and κu represent the corresponding mass flow rate, density, viscosity and the isentropic exponent calculated at upstream pressure Pu and temperature Tu, while ∆P, D and d denote the differential pressure across the orifice plate, internal diameter of the pipe and the orifice, respectively. In case of upstream pressure and downstream temperature measurement, as suggested by (ISO-5167-2, 2003), the flow rate (18), changes to: ( )dDpTPqq ddddud ,,,,,,, κγρ∆= (19) where qd, ρd, γd and κd denote the corresponding mass flow rate, density, viscosity and the isentropic exponent calculated in “downstream conditions” i.e. at the upstream pressure pu and the downstream temperature Td. For certain natural gas compositions and operating conditions the flow rate qd may differ significantly from qu and the corresponding compensation for the temperature drop effects, due to JT expansion, may be necessary in order to preserve the requested measurement accuracy (Maric & Ivek, 2010). Precise compensation of the flow rate error needs double calculation of the properties of a natural gas and the flow rate what makes the compensation unfeasible in real time in low computing power embedded systems. To reduce the computational burden we aim to derive a low-complexity flow rate correction factor model that will enable direct compensation of the flow rate error caused by the measurement of the downstream temperature. The correction factor model has to be simple enough in order to be executable in real time and accurate enough to ensure the requested measurement accuracy. The flow rate correction factor K can be obtained precisely by dividing the true flow rate qu calculated in the upstream conditions, (18), by the flow rate qd calculated in the “downstream conditions” (19): d u q q K = . (20) For the given correction factor (20), the flow rate at upstream pressure and temperature (18) can be calculated directly from the flow rate computed in “downstream conditions” (19), i.e. du qKq ⋅= . The relative flow rate measurement error Er can be estimated by comparing the uncompensated (qd) and precisely calculated (qu) flow rate i.e. ( ) uudr qqqE −= . Our objective is to derive the surrogate model of the flow rate correction factor. Given the surrogate model (Ksm) for the flow rate correction factor (20), the true flow rate qu can be approximated by: dsmsm qKq ⋅= , where qsm denotes the flow rate corrected by the surrogate model (Ksm) of the real correction factor K. The relative measurement error Esm can be estimated by comparing the approximate (qsm) and the precisely calculated (qu) flow rate i.e. ( ) uusmsm qqqE −= (21) The correction is modeled by SOPNN and MLP, having equivalent computational complexities, and the corresponding approximation errors are compared. Note that natural gas properties, like density, isentropic exponent and viscosity, need to be calculated and are involving the natural gas characterization parameters in the flow rate calculation, like molar fractions of the components, superior calorific value and relative density. The procedures for the correction of the flow rate error and for the estimation of optimal set of input parameters (Table 1) needed for the correction factor modeling are elaborated in (Maric & Ivek, 2010). Table 1 Optimal Set of input Parameters for Natural Gas Flow Rate Correction Factor Modelling Index Parameter description Range of application 0 XCO2 - mole fraction of carbon dioxide 0 ≤ XCO2 ≤ 0.20 1 XH2 - mole fraction of hydrogen 0 ≤ XH2 ≤ 0.10 2 p - absolute pressure in MPa 0 < p ≤ 12 3 T - temperature in K 263 ≤ T ≤ 368 4 ∆p - differential pressure in MPa 0 ≤ ∆p ≤ 0.25p 5 ρ - density in kg/m3 intermediate result 6 ρr - relative density 9.55 ≤ ρr ≤ 0.80 7 HS - superior calorific value in MJ/m 3 30 ≤ HS ≤ 45 8 β - orifice to pipe diameter ratio: d/D 0.1 ≤ β ≤ 0.75 11 The next section describes the correction factor modeling and the results of the simulation of the flow rate error. 5. Flow rate error compensation modeling For the purpose of meta-modeling, the precise calculation of the natural gas flow rate correction factor is implemented in software and run on a digital computer. The training data set, validation data set and 10 test data sets, each consisting of 20000 samples of correction factor, were randomly sampled across the entire space of application (Table 1) in accordance with (Maric & Ivek, 2010). The ranges of application are shown in Table 1. The maximum tolerable ET and the maximum acceptable root relative squared error (RRSE) (Maric & Ivek, 2011) of the correction factor surrogate model in our flow computer prototype are limited to 30 ms and 5%, respectively. For the given maximum ET of the model, maximum ET of the FP operations, and 9 input parameters, the maximum acceptable number of SOPNN nodes and MLP neurons can be calculated by (16) and (17), respectively. In order to build the models with equivalent ET not exceeding 30 ms, the MLP (Fig. 3) is limited to 5 neurons with the maximum ET, Tmlp ≤ 27.75 ms, estimated by (17). The complexity of the corresponding SOPNN model (Fig. 5) is limited to maximum 27 polynomial nodes with the maximum ET, TSOPNN ≤ 27.00 ms, estimated by (16). L 08 L 07 L 06 L 05 L 04 L 03 L 02 L 01 L 00 x 0 x 1 P 0 L 09 L 10 L 11 x 2 x 3 x 4 x 5 x 6 x 7 x 8 P 1 P 3 P 2 P 4 P 5 P 6 P 7 P 8 P 9 P 10 P 11 P 12 P 13 P 14 P 15 P 16 P 17 P 18 P 19 P 20 P 21 P 22 P 23 P 24 P 25 P 26 Fig. 5. The approximation accuracy of the derived models are tested on 10 randomly generated test data sets for the flow rate correction factor, each consisting of 20000 samples, and the results are shown in Fig. 6. The corresponding mean RRSE and standard deviation are given in Table 2. The results obtained from 10 independent test data sets clearly indicate that the models are not over-fitted. Fig. 6. 12 Table 2 Average RRSE of the correction factor for 10 test data sets when approximated by MLP and SOPNN models Root relative squared error: Errs in % MLP-LM SOPNN SOPNN-PSO SOPNN-LM Mean value 3.036 4.558 3.690 1.214 Standard deviation 0.02525 0.04099 0.04158 0.00968 The MLP-LP in Fig. 6 and Table 2 denotes the feed forward MLP (Fig. 3) with 4+1 neurons trained by the LM algorithm. The SOPNN represents the 27-node self-organized model (Fig. 5) trained by the GMDH algorithm using the compound measure ( ) ( ) ( )2 0 2 0 1 exeexewrrsrrswCE TTcEEcE ⋅−+⋅= for model selection (Maric & Ivek, 2011), which combines the constraints upon model approximation error (Errs≤Errs0) and ET (Texe≤Texe0) by the weighting coefficient (0≤cw≤1). SOPNN-PSO and SOPNN-LM represent the same SOPNN model optimized by the PSO and the LM, respectively. From Fig. 6 and Table 2 it can be seen that MLP trained by the LM algorithm (MLP-LM) achieves about 33% lower root relative squared error than the non-optimized SOPNN model having approximately the same complexity. Optimization of the SOPNN model by PSO (SOPNN-PSO) is time consuming. The procedure is computationally intensive due to the high dimensional search space (27*6=162 weights) requesting large number of particles and iterations. The unknown lower and upper boundaries of the search space are complicating the procedure additionally. To overcome the problem the optimization has been divided into epochs each consisting of 2000 generations. At the beginning of each epoch the boundaries are fixed around the position of the best particle from the previous epoch in the following way: Li=Ai(1-µ), Ui=Ai(1+µ), where Ai denotes the position of the best particle in the i th dimension, Li and Ui denote the lower and the upper boundary of the i th dimension of the search space and µ is a positive number. The µ has been varied from 0.01 to 1.5 and the best results, in this particular application, are obtained for µ=1.2. The SOPNN-PSO improves the SOPNN accuracy for about 20%. Since the PSO shows very slow but constant improvement of the model, the LM algorithm for the optimization of the SOPNN weights has been implemented in order to speed-up the convergence. The SOPNN-LM converges rapidly. It decreased the SOPNN error almost four times (73.3%) and outperformed all other models. It achieved 60% better accuracy than the corresponding MLP-LM. In each cycle of the LM algorithm, the damping coefficient has been automatically adjusted for maximum decrease of the RRSE and the model converges rapidly to its optimum weights. The same effect of the LM optimization on model error has been observed for models at all layers. Fig. 7 illustrates the average error in logarithmic scale obtained for the eight best ranked models from each layer, before (SOPNN) and after they have been optimized by the LM algorithm (SOPNN-LM). Fig. 7. The results show that the SOPNN polynomial relation offer substantially better approximation abilities than it can be concluded from the characteristics of raw models obtained directly from the GMDH algorithm. We have obtained similar results by modeling the procedures for the calculation of various thermodynamic properties of natural gas including molar heat capacities, speed of sound, etc. The model generated by the GMDH algorithm is generally far from optimal since the weights of each node- polynomial, after calculated by the regression, remain unchanged throughout the rest of the learning process. The GMDH algorithm is trying to maximize the approximation accuracy of the model by embedding the system internal dependencies into the model structure. To achieve this it performs the regression of the polynomial 13 weights of the uppermost node only. In this way the GMDH algorithm controlled by the imposed metrics produces the model that generally achieves the structure, the weights and the approximation accuracy that can be considered as a very good starting point for further optimization. Next Section demonstrates the outstanding ability of the SOPNN to model the polynomial-type recurrence relations. 6. Multiple Superimposed Oscillations Modeling In this section we discuss the learning of the recurrence relations from time series by using SOPNN. We analyze the performances of SOPNN on multiple superimposed oscillations (MSO) task that has already been attempted with varying degrees of success by using ANN and SVM (Xue, Yang & Haykin, 2007; Schmidhuber, Wierstra, Gagliolo & Gomez, 2007; Holzmann & Hauser, 2010; Ceperic, Gielen & Baric, 2012). The following multiple sinusoids are modeled: ( ) ( )nny 311.0sin2.0sin 2 += , (22) ( )nyy 42.0sin 23 += , (23) ( )nyy 51.0sin 34 += , (24) ( )nyy 74.0sin 45 += , (25) where n=1,…,700. Note that the frequencies of the sinusoids are not integer multiples of each other. As described in (Ceperic, Gielen & Baric, 2012) the first 400 samples (n=1,...,400) are used to train the model, while the rest of data (n=401,...,700) is used to test the model. The data is generated in double floating point (FP) number precision. We have limited the complexity of our SOPNN to maximum 20 nodes (N), where each node- polynomial can be calculated by only 5 FP additions and 5 FP multiplications. In the worst case the maximum total number of 100 FP additions and 100 FP multiplications would be necessary to calculate the SOPNN output, which is far below the computational complexities of the corresponding models described in the above mentioned references. As in (Ceperic, Gielen & Baric, 2012), we limited the maximum dimension (D) of the input space to 50 what corresponds to maximum 50 delayed output samples in the recurrent configuration. 6.1. Comparison with other approaches Table 3 shows the mean square error (MSE) and the normalized root mean square error (NRMSE) (Holzmann & Hauser, 2010) on the test data set, obtained by the non-optimized GMDH model (SOPNN) and by the same model after it has been optimized by the proposed LM algorithm (SOPNN-LM). In Table 3, D denotes the dimensionality, i.e. the corresponding total number of delayed output signals, and N denotes the total number of nodes in SOPNN. From Table 3 it can be seen that the dimension (D) of the input space and the number of nodes (N) of the SOPNN are increasing with the number of sinusoids in the MSO in order to preserve high prediction accuracy. Table 3 Mean square error in the prediction of the next sample of the MSO test data set (y2, y3, y4, y5; n=401,...,700) by the corresponding non-optimized (SOPNN) and optimized (SOPNN-LM) N-node, D-dimensional SOPNN model generated by the GMDH algorithm using the corresponding training data set (y2, y3, y4, y5; n=1,...,400) SOPNN SOPNN-LM MSO D N MSE NRMSE MSE NRMSE y2 4 3 1.28E-04 9.16E-03 3.07E-27 5.53E-14 y3 10 8 1.16E-05 2.75E-03 3.89E-25 5.04E-13 y4 16 13 6.47E-05 5.52E-03 2.28E-24 1.04E-12 y5 50 16 2.66E-04 1.01E-02 6.06E-25 4.84E-13 Xue, Yang & Haykin, 2007, use four reservoirs each with 100 neurons to approximate the two sine problem (22) with “decoupled echo state network with lateral inhibition”. The MSE they obtained on test data set was 14 3· 10 -4 and is almost 23 orders of magnitude higher than the MSE obtained by the corresponding y2 SOPNN-LM model. Also, a huge complexity of their model is incomparable to low complexity of our 3-node, 4-dimensional y2 SOPNN-LM model (Table 3). Schmidhuber, Wierstra, Gagliolo & Gomez, 2007, use PI-EVOLINO and EVOKE networks to model the MSO data sets. They reported the following normalized root mean square errors: y2: NRMSE=4.15E-3, y3: NRMSE=8.04E-3, y4: NRMSE=1.01E-1 y5: NRMSE=1.66E-1. The NRMSE errors they reported are at least 9 orders of magnitude worse than the results obtained by the optimized SOPNN-LM (Table 3). Also, the generalization results they obtained for two sinusoids y2 by the EVOKE networks, which are based on SVM, are even much worse. The huge complexity of their models cannot be compared with the simplicity of SOPNN models. Ceperic, Gielen & Baric, 2012, use recurrent sparse support vector regression machines trained by active learning in time-domain combined with the optimization of SVM hyper parameters to model the MSO data sets. They reported the following errors: • y2: MSE=3.57E-8, NRMSE=1.88E-4, using 13 SV; • y3: MSE=2.33E-6, NRMSE=1.23E-3, using 15 SV; • y4: MSE=1.75E-5, NRMSE=2.86E-3, using 15 SV; • y5: MSE=9.16E-5, NRMSE=5.94E-3, using 15 SV, which are considerably better than the errors reported in (Xue, Yang & Haykin, 2007) and (Schmidhuber, Wierstra, Gagliolo & Gomez, 2007) but still the orders of magnitude worse than the errors obtained by the corresponding optimized SOPNN-LM model shown in Table 3. Also the computational complexity (Maric & Ivek, 2011) of the SVM model with 13 or 15 support vectors is much higher than the complexity of the corresponding SOPNN-LM model with 3, 8, 13 or 16 2-dimensional, 2 nd order node-polynomials (1). Holzmann & Hauser, 2010, achieved very good results using echo state networks (ESN) with arbitrary infinite impulse response filter neurons and a delay&sum readout. They used a reservoir of 100 specialized neurons tuned to different frequencies, which are able to generate a superposition of sinusoids. They did not report the results in numerical form but from graphical presentation of the NRMSE in logarithmic scale it can be seen that their errors for y2 (NRMSE>2E-9), y3 (NRMSE>2E-7), y4 (NRMSE>1E-5) and y5 (NRMSE>5E-5) are more than 4 orders of magnitude higher than the corresponding SOPNN-LM errors (Table 3). Again, the huge complexity of ESN models cannot be compared with the simplicity of the corresponding SOPNN. 6.2. Generalization abilities of SOPNN Table 4 shows the MSE and the NRMSE for the next sample prediction obtained on various combinations of sinusoids from y5 MSO by using the same model (Fig. 8) generated and optimized on y5 training data. Note that the dimension of each test data set in Table 4 is equal to 50 as determined by the model (SOPNN-LM-y5), which is built and optimized by using 50-dimensional training data set obtained from the first 400 samples (n=1,...,400) of y5 MSO (25). The output from the 16 th node-polynomial P15 in Fig. 8 is the next predicted output (x50) calculated from 50 delayed outputs x0,…,x49, where x0 denotes the oldest and x49 the most recent output from 50- samples window. In the next step the outputs x1,...,x50 are used to predict the output x51 and so on. From Fig. 8 it can be seen that the SOPNN model uses only 12 out of 50 delayed outputs to predict the next output. As can be seen from Table 4, the same SOPNN-LM-y5 model is able to accurately predict any subset of sinusoids from y5 MSO (25). It is extremely superior to any model tailored to a specific MSO by various ANN and SVM approaches. In order to demonstrate another superior characteristic of SOPNN, let us consider for example the following MSO signal obtained after changing the amplitudes and phase shifts of all the corresponding sinusoids of y5 MSO (25) e.g.: ( ) ( ) ( ) ( ) ( )n nn nny a 74.032.0sin8.2 51.01sin35.142.01.2sin1.2 311.07.0sin8.12.03.1sin2.1 5 − −+−++ −−−+= . (26) The next sample prediction errors MSE=1.91E-22 and NRMSE=4.45E-12 have been obtained by the same SOPNN-LM-y5 model after predicting the data set (y5a, n=401,...,700) arranged as 50 delayed output samples in a recurrent configuration. Fig. 9 illustrates the modeled values for y5a (26) and the corresponding model error for the first 300 steps calculated by the SOPNN-LM-y5 (Fig. 8). 15 Table 4 MSE and RMSE in the prediction of the next sample of the MSO and single sinusoid test data obtained by the same 50-Dimensional SOPNN generated and optimized on y5 training data Test data set SOPNN-LM-y5 MSO MSE NRMSE y2 5.31E-26 2.30E-13 y3 8.16E-26 2.31E-13 y4 9.76E-26 2.14E-13 y5 6.06E-25 4.84E-13 sin(0.2n) 5.76E-26 3.39E-13 sin(0.311n) 5.09E-26 3.21E-13 sin(0.42n) 4.63E-26 3.04E-13 sin(0.51n) 4.56E-26 3.01E-13 sin(0.74n) 2.56E-26 2.26E-13 L 08 L 07 L 06 L 05 L 04 L 03 L 02 L 01 L 00 x 49x 0 x 1 ... P 15 P 8 P 13 P 12 P 14 P 10 P 11 P 7 P 6 P 5 P 2 P 0 P 3 P 1 P 9 P 4 L 09 L 10 L 11 L 12 L 13 L 14 L 15 Fig. 8. 16 -10 -5 0 5 10 0 50 100 150 200 250 300 Prediction horizon O u tp u t v a lu e (a) -8 -6 -4 -2 0 2 4 6 8 0 50 100 150 200 250 300 Prediction horizon M o d e l e rr o r x 1 0 -1 0 Next sample prediction errors: MSE = 1.91E-22 NRMSE = 4.45E-12 (b) Fig. 9. Fig. 9 shows the results of the prediction of y5a test data by SOPNN-LM-y5 model starting with y5a data vector, which consists of 50 delayed outputs preceding the output sample n=401. Fig. 9(a) shows the prediction of y5a test data by SOPNN-LM-y5 for the prediction horizon from 0 to 300 samples. Fig. 9(b) shows the corresponding prediction error for the same prediction horizon. The model preserves high accuracy even if changing the amplitudes and phases of the sinusoids from which the y5 MSO signal is composed. Note that maximum absolute prediction error did not exceed 8· 10 -10 in the whole prediction horizon. Fig. 10 shows the absolute model error in logarithmic scale when predicting test data for y5 MSO (25) and y5a MSO (26) by the same SOPNN-LM-y5 model in the prediction horizon from 0 to 3000 samples. 17 1.0E-11 1.0E-10 1.0E-09 1.0E-08 1.0E-07 1.0E-06 1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01 0 500 1000 1500 2000 2500 3000 Prediction horizon M o d e l a b s o lu te e rr o r in L o g s c a le y5a y5 Fig. 10. From Fig. 10 it can be seen that the prediction error is increasing exponentially when increasing the prediction horizon. When using SOPNN-LM-y5 to predict the y5a MSO test data (black line), the corresponding maximum errors are more than 50 times higher than the errors obtained on y5 MSO test data (gray line). But, as can be concluded from Fig. 9 and Fig. 10, the y5a MSO prediction error still remains extremely low in a wide prediction horizon. Similar error characteristics are obtained by SOPNN-LM-y5 model when predicting any subset of sinusoids from y5 MSO. From Table 4 and Figs. 9 and 10 it can be seen that unlike any other above mentioned approach the SOPNN, optimized by the LM algorithm, has extraordinary generalization abilities. It can accurately predict in a wide prediction horizon any subset of sinusoids from the MSO signal it has been trained for, even if their amplitudes and phases are significantly changed. The same equations can be also used to accurately predict any sinusoid or a subset of superimposed sinusoids from y5 MSO (25). The equations and the coefficients of the SOPNN-LM-y5 model (Fig. 8) are given in Appendix A for testing purposes. 7. Conclusion The paper presents an efficient adaptation of the Levenberg-Marquardt algorithm for the optimization of the SOPNN. The paper points out that LM algorithm makes the SOPNN very competitive for the approximation of complex systems and procedures, particularly in real-time applications, since high approximation accuracy is generally achieved with low complexity models. In computationally intensive real-time applications the complex procedures may be replaced by simplified SOPNN surrogates and thus become feasible in real-time. The paper particularly emphasizes a high accuracy/complexity ratio of the optimized model and the simplicity of its implementation in software. The LM algorithm has been tested by modeling the computationally intensive procedure for the correction of the flow rate error. It was shown that the approximation characteristics of the original SOPNN can be improved substantially when optimizing the model weights by the LM algorithm. The SOPNN outperformed the corresponding MLP model when both optimized by the LM algorithm and having approximately equal computational complexities. Similar results have been obtained by modeling the procedures for the calculation of various thermodynamic properties of a natural gas. The SOPNN proves to be extremely efficient in learning polynomial-type recurrence relations from time series. When optimized by the LM algorithm the SOPNN outperformed the ANN and the SVM in MSO modeling. When modeling a MSO recurrence relation the optimized SOPNN displays outstanding generalization ability, uncommon to ANN and SVM models, and can accurately predict any subset of superimposed sinusoids from the MSO signal it has been trained for. The model prediction error remains very low even if changing the amplitudes and phases of the superimposed sine waves. SOPNN is also able to learn with extreme accuracy the recurrence 18 relations of multiple products of sine waves, modulated MSO, damped MSO, etc. The ability to learn the recurrence relation and to accurately predict the future values from past known samples opens the possibilities of using the SOPNN in modeling the dynamic behavior of complex systems. Appendix A The concatenated polynomial relation of the SOPNN model from Fig. 8 P = P15(P14(P13(P12(P11(P10(P9(P8(P7(P6(P5(P4(P3(P2(P0(x48,x49),P1(x47,x49)),x41),x0),x33),x48),x15),x2),x41),x33), P1(x47,x49)),x46),x37),x43),x14) can be computed by 16 basic polynomials, P0(x48,x49), P1(x47,x49), P2(P0,P1), P3(P2,P41), P4(P3,x1), P5(P4,x33), P6(P5,x48), P7(P6,x15), P8(P7,x2), P9(P8,x41), P10(P9,x33), P11(P10,P1), P12(P11,x46), P13(P12,x37), P14(P13,x43), P15(P14,x14), where the basic polynomials are calculated by ( ) 215, 2 24, 2 13,22,11,0,21 , zzazazazazaazzP iiiiiii +++++= using the following optimized double precision coefficients: a0,0=2.4089806255330348e-001, a0,1=-2.0609933517875878e-002, a0,2=1.7531109297004226e+000, a0,3=1.8381037377004718e-009, a0,4=4.6419038970178617e-005, a0,5= -8.6782752584535839e-008, a1,0=-6.6581136153327869e-001, a1,1=-2.5039829582659459e-009, a1,2=1.1149843617468462e+000, a1,3=-6.7607152698616724e-012, a1,4=2.9528090425706862e-005, a1,5=-1.5739033560547226e-011, a2,0=-7.8345182851489101e-002, a2,1=3.0222771938049342e+000, a2,2=-1.4659393406572994e+000, a2,3=3.1719446468444557e-001, a2,4=7.8358643543900208e-001, a2,5=-9.9744717755725765e-001, a3,0=-7.1836925226834172e-002, a3,1=8.2254766389549738e-001, a3,2=-5.8661286884162590e-002, a3,3=-9.4478934614005779e-007, a3,4=1.0203172272677782e-004, a3,5=-7.1603355247232822e-012, a4,0=-7.8622868604762711e-002, a4,1=8.3924346238730552e-001, a4,2=-2.3551250558386455e-001, a4,3=5.8200253943802276e-008, a4,4=4.5880465656650697e-009, a4,5=-3.2699243420244513e-008, a5,0=-5.7663794534583354e-002, a5,1=8.5302808485752646e-001, a5,2=-1.7417718227724835e-001, a5,3=-7.0354372488410098e-008, a5,4=5.3723532859500584e-005, a5,5=-8.4611612366864986e-011, a6,0=-1.0373663629806662e-001, a6,1=8.5125981079556934e-001, a6,2=-6.4996174285114006e-001, a6,3=4.8155545589541646e-010, a6,4=-6.7537943733023871e-005, a6,5=-1.1056387926085760e-009, a7,0=-8.2497971373195644e-002, a7,1=8.9229466216913467e-001, a7,2=1.7609360498148038e-002, a7,3=5.2983902284172561e-008, a7,4=2.1005689852310062e-011, a7,5=2.1265022979134083e-009, a8,0=-7.9273328703511364e-002, a8,1=8.9884511364949637e-001, a8,2=6.0847463081777652e-002, a8,3=1.1340108864742532e-007, a8,4=7.9841230450439977e-010, a8,5=2.3588069735832587e-008, a9,0=-6.4817529180249148e-002, a9,1=8.9631687643991975e-001, a9,2=-3.5335719981528518e-001, a9,3=-1.9340241083767072e-007, a9,4=-4.4699594154393525e-005, a9,5=9.3368507730513902e-011, a10,0=-8.6117614268338985e-002, a10,1=9.4290247056516585e-001, a10,2=3.2597051000195628e-003, a10,3=1.0750613301384445e-009, a10,4=-3.0999033115579099e-005, a10,5=6.4343070771435519e-012, a11,0=-3.0294545534514258e-002, a11,1=1.0343523735701383e+000, a11,2=1.5009384616591143e-001, a11,3=-1.5912006880327145e-009, a11,4=1.6831199002889785e-004, a11,5=-1.4626868034716834e-010, a12,0=-4.5010467101454295e-002, a12,1=9.4652096952271103e-001, a12,2=-3.7223918066824557e-001, a12,3=-2.0759958560079309e-009, a12,4=-3.9017659043448758e-010, a12,5=1.9822967708575749e-009, a13,0=-4.3242875776603580e-002, a13,1=9.9731718722116014e-001, a13,2=6.0688857337153193e-002, a13,3=-3.4003929112323072e-009, a13,4=-2.2978021794518529e-011, a13,5=-7.5532778919304302e-010, 19 a14,0=-3.3716883954380848e-002, a14,1=1.0001258995629190e+000, a14,2=3.7500419296080839e-001, a14,3=1.7889822638343193e-009, a14,4=-6.2592651157127744e-010, a14,5=-3.3383808879078198e-009, a15,0=-3.3337185503942837e-002, a15,1=1.0819274725734265e+000, a15,2=3.3496218434058894e-002, a15,3=4.8130435297607735e-009, a15,4=-1.9617688051299804e-014, a15,5=-1.2826297843293008e-013. The variables x0, …,x49 correspond to the delayed output samples in recurrent configuration where x0 denotes the oldest and x49 the most recent sample as described in Section 6. Acknowledgements This work has been supported by the Croatian Ministry of Science, Education, and Sports through the project “Computational Intelligence Methods in Measurement Systems” – No. 098-0982560-2565. References Ceperic, V., Gielen, G., & Baric, A. (2012), Recurrent sparse support vector regression machines trained by active learning in the time-domain, Expert systems with applications, vol. 39, 10933-10942. Chapra, S. C., & and Canale R. P. (1998), Numerical Methods for Engineers, 3 rd ed., McGraw-Hill, New York. Eberhart, R. C., & Kennedy, J. (1995), A new Optimizer Using Particles Swarm Theory, Proceedings of the Sixt International Symposium on Micro Machine and Human Science (Nagoya, Japan), IEEE Service Center, Piscataway, NJ, 39-43. Ferrari, S., & Stengel, R. F. (2005), Smooth function approximation using neural 756 networks, IEEE Transactions on Neural Networks, vol. 16, no. 1, 24–38. Holzmann, G., & Hauser, H. (2010), Echo state networks with filter neurons and a delay&sum readout, Neural Networks, vol. 23, 244-256. ISO-20765-1. (2005), Natural gas – Calculation of thermodynamic properties - Part1: Gas phase properties for transmission and distribution applications, Ref. No. ISO-20765-1:2005. ISO-5167-2. (2003), Measurement of fluid flow by means of pressure differential devices inserted in circular- cross section conduits running full - Part 2: Orifice plates, Ref. No. ISO-5167-2:2003(E). Ivakhnenko, A. G. (1971), Polynomial Theory of Complex Systems, IEEE Transactions on Systems Man, and Cybernetics, Vol. SMC-1, No.4, 364-378. Iwasaki, M., Takei, H., & Matsui, N. (2003), GMDH-Based Modeling and Feedforward Compensation for Nonlinear Friction in Table Drive Systems, IEEE Transactions on Industrial Electronics, Vol. 50, No. 6, 1172-1178. Jin, R., Chen, W., & Simpson, T. W. (2000), Comparative studies of metamodeling techniques under multiple modeling criteria, AIAA-2000-4801, AIAA/USAF/ NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, 8th, Long Beach, CA. Levenberg, K. (1944), A method for the solution of certain problems in least squares, Quarterly of Applied Mathematics, Vol. 2, 164–168. Maric, I., & Ivek, I. (2010), Compensation for Joule-Thomson effect in flow rate measurements by GMDH polynomial, Flow Measurement and Instrumentation, vol. 21, no. 1, 134-142. Maric, I., & Ivek, I. (2011), Self-Organizing Polynomial Networks for Time-Constrained Applications, IEEE Transactions on Industrial Electronics, vol. 58, no. 5, 2019-2029. Marquardt, D. (1963), An algorithm for least-squares estimation of nonlinear parameters, SIAM Journal on Applied Mathematics, Vol. 11, 431–441. Nikolaev, N. Y., & Iba, H. (2003), Learning Polynomial Feedforward Neural Networks by Genetic Programming and Backpropagation, IEEE Transactions on Neural Networks, Vol. 14, No. 2, 337-350. Schmidhuber, J., Wierstra, D., Gagliolo, M., & Gomez, F. (2007), Training recurrent networks by evolino. Neural Computation, 19, 757–779. Shi, Y. H., & Eberhart, R. C. (1998), A Modified Particle Swarm Optimizer, Proceedings of International Conference on Evolutionary Computation, IEEE WCCI, Anchorage, Alaska, 69-73. 20 Vapnik, V., Golowich, S. E., & Smola, A. J. (1997), Support vector method for function approximation, regression estimation and signal processing, Advances in Neural Information Processing Systems 9, Mit Press, Cambridge, 281-287,. Witten, I. H., & Eibe, F. (2005), Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations, ISBN: 9780120884070, Morgan Kaufmann Publishers. Xue, Y., Yang, L., & Haykin, S. (2007), Decoupled echo state networks with lateral inhibition, Neural Networks, vol. 20, 365-376. 21 Figure captions Fig. 1. Illustration of SOPNN construction. Fig. 2. Hypothetical example of SOPNN. Fig. 3. Feed forward MLP scheme. Fig. 4. A schematic diagram of the natural gas flow rate measurement using an orifice plate with corner taps. Fig. 5. SOPNN model of flow-rate correction factor satisfying the constraint on maximum ET. Fig. 6. Root relative squared error for MLP and SOPNN models measured on 10 randomly generated test data sets. Fig. 7. Average RRSE calculated before and after optimization by the LM algorithm of the eight best SOPNN models from each layer Fig. 8. SOPNN modeled and optimized by using (y5, n=1,...,400) 50-dimensional training data set. The output from P15 is calculated from 50 delayed outputs x0,…,x49, and represents the next predicted output x50. Note that x0 is the oldest and x49 is the most recent output from 50-sample window. Fig. 9. Illustration of the modeled values and the model error for the first 300 test data samples (y5a, n=401,…,700) obtained by the model SOPNN-LM-y5 (Fig. 8) with 50 delayed outputs in recurrent configuration: (a) modeled values for y5a, (b) model error for y5a. Note that SOPNN-LM-y5 is learned on y5 (25) training data and used to predict y5a (26) test data. Fig. 10. Illustration of the absolute prediction error in Log scale for the first 3000 test data samples of y5 and y5a, (n=401,…,3400), obtained by the model SOPNN-LM-y5 with 50 delayed outputs in recurrent configuration. 
work_uk4h6l4hanhxvaqxqifjbsltha	----	1 t o d 2 s & b ( M ( p Cooperative controllers for highways based on human experience J. Pérez , V. Milanés , J. Godoy , J. Villagrá , E. Onieva R E S U M E N worki develop uvres h vehicle oopera The AUTOPIA program has been advances have focused on the However, so far, these manœ experiments with autonomous where several vehicles had to c context, the main challenge was to t communication describes the experien system implemented in a Citröen C3. is known as cooperative ACC (CACC) (van Ar 2006). ng on the development of intelligent autonomous vehicles for the last 10 years. Its latest ment of cooperative manœuvres based on communications involving several vehicles. ave been tested only on private tracks that emulate urban environments. The first s on real highways, in the framework of the grand cooperative driving challenge (GCDC) te in order to perform cooperative adaptive cruise control (CACC), are described. In this ranslate, through fuzzy controllers, human driver experience to these scenarios. This ces deriving from this competition, specifically that concerning the controller and the . Introduction During the last few years, many advanced driver assistance sys- ems (ADAS) based on on-board vehicle sensors have been devel- ped for automotive applications. Vision systems for blind angle etection (Lin et al., 2010), lane departure warning (Bansal et al., 008), collision avoidance (Llorca et al., 2011), and lidar-based ystems have been widely applied in vehicle detection (Pauwelussen Feenstra, 2010). In spite of these advances, the latest trends have een focused on wireless communications -both vehicle-to-vehicle V2V) (Hosseini Tabatabaei, Fleury, Qadri, & Ghanbari, 2011; itropoulos et al., 2010) and vehicle-to-infrastructure (V2I) Belanovic et al., 2010; Nguyen, Berder, & Sentieys, 2011) – to erceive the environment as precisely as possible. Adaptive cruise control (ACC), one of the most conventional forms of ADAS, was developed some years ago. It acts on the longi- tudinal control of the vehicle, permitting it to follow a leader – act- ing on the throttle and brake pedals autonomously – and to maintain a predefined headway with the vehicle in front (Naranjo, Gonzalez, Reviejo, Garcia, & de Pedro, 2003). The next step in the evolution of this technology is based on cooperation among differ- ent vehicles in order to reduce this headway between vehicles. This em, van Driel, & Visser, Furthermore, the broad scope and growing popularity of intelli- gent transportation systems (ITS) makes it difficult to include all the recent advances in a single research project. There are many projects and groups around the world working on ITS, but in differ- ent fields (perception, control, communications, actuation, ADAS, and traffic control, among others). For example, the project HAVEit (Flemisch et al., 2010) is developing a control architecture to im- prove safety, efficiency, and comfort in driving through a virtual system that takes partial control of the vehicle according to differ- ent risk situations. It is designed to respond to the increase in traffic density, the ever greater flood of information available to drivers (in both autonomous and manual modes), and the growing population. Its highly automated vehicle illustrates the character- istics of the future of mobility. Another interesting project in the integration of vehicle applications is the strategic platform for ITS (SPITS) (Koenders, 2011), which is developing new concepts for the communication protocols among smart vehicles. Some American research groups have made important advances in the ITS field. Firstly, the Californian PATH of the Institute of Transportation Studies at the University of California, Berkeley, has demonstrated on a real automated highway system that the platooning problem can be solved with the use of magnetic sensors (Tan & Bougler, 2001). This group has accumulated much experi- ence in developing various projects over the last 20 years (Shladover, 2007). Secondly, the Defense Advanced Research Project Agency has organized one of the most important competitions involving autonomous vehicles – the DARPA urban challenge http://dx.doi.org/10.1016/j.eswa.2012.08.011 mailto:joshue.perez_rastelli@inria.fr mailto:vicente.milanes@berkeley.edu mailto:vicente.milanes@berkeley.edu mailto:jorge.godoy@csic.es mailto:jorge.villagra@csic.es mailto:enrique.onieva@decsai.ugr.es http://dx.doi.org/10.1016/j.eswa.2012.08.011 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. AUTOPIA team in the GCDC 2011 competition. (Urmson et al., 2008). Several research groups have participated, trying out their control systems and algorithms in both desert and urban scenarios. It was in this context that the first competition among autono- mous vehicles in Europe was organized last year – the grand coop- erative driving challenge (GCDC). It involves different research groups from around the world who talk to each other over a wire- less network. The aim was to carry out two-lane CACC with differ- ent vehicles. The present communication describes the controller and algorithm used by the AUTOPIA program in this competition. It is based on human experience emulated through fuzzy logic. The AUTOPIA program has experience in cooperative manœ- uvres among vehicles based on wireless communications using the IEEE 802.11g standard. All our manœuvres had been tested be- fore in urban scenarios (under 50 km/h, on our private test tracks). For this reason the GCDC competition was considered to be an excellent scenario in which to test our systems performance on highways (MilanTs, Llorca, Vinagre, González, & Sotelo, 2010; PTrez, MilanTs, & Onieva, 2011). Our main challenges in order to have the capacity to participate in the competition were to adapt our communication systems to the CALM FAST/IEEE 802.11p pro- tocol stack (a GCDC requirement), and to develop new controllers capable of working at any speed range. The rest of the paper is organized as follows. In Section 2, we de- scribe our group’s background. The system architecture, the vehi- cle used in the GCDC competition, and the communication system are presented in Section 3. Section 4 describes the human knowledge based controllers developed for highway scenarios. The different experiments, testing various speed profiles, are described in Section 5. Finally, the conclusions and some remarks are given in Section 4. 2. AUTOPIA program AUTOPIA is a Spanish research program whose initial long-term goal was the automation of vehicles. It forms part of the Centre of Automation and Robotics of the Polytechnic University of Madrid and of the Spanish National Research Council (UPM-CSIC). Its conception arose from the confluence of two trends: fuzzy logic systems and mobile robots. The group had been working in these two fields for several years, and with AUTOPIA combined them to start working with autonomous vehicles. AUTOPIA began at the end of the 1990s with the group’s acqui- sition of two electric vans. Using these, we were capable of performing both autonomous vehicle guidance (Naranjo, Gonzalez, Garcia, de Pedro, & Haber, 2005) and our first cooperative manœ- uvre – an ACC+Stop&Go-based on V2V communications in urban environments (Naranjo et al., 2003). These vehicles were instru- mented with a DC motor connected to the brake pedal, and an analogue card for the throttle. Later, the group acquired two gaso- line-propelled vehicles which were also automated (MilanTs et al., 2010). This allowed to test more complex cooperative manœuvres such as crossing intersections (MilanTs, PTrez, Onieva, & González, 2010), overtaking on two-way roads (PTrez, MilanTs, Onieva, Godoy, & Alonso, 2011), and merging (MilanTs, Godoy, Villagrá, & PTrez, 2011). Presently we are working on a traffic management system for complex manœuvres. This system would transmit to each vehicle involved a unique control value representing a trade-off between safety and traffic flow (MilanTs et al., in press). If the receiving vehicle is manually driven, this value would be shown in a head- up display and treated by an ADAS in order to provide appropriate advice to the driver. If, however, the receiving vehicle is autono- mous, this signal would be used to automatically regulate the vehicle’s speed to fit the specific scenario – intersection traversing, merging, etc. Our research motivations are mainly directed to autonomous vehicle control and cooperative manœuvres through V2X communications. Perception and world modeling has been developed in cooperation with other research groups, e.g., pedes- trian avoidance using vision systems (Llorca et al., 2011) and recognition of RFID signals for intelligent cruise control (PTrez et al., 2010). With these antecedents, we decided to choose one of our gaso- line-propelled vehicles to participate in the GCDC (Fig. 1). The vehi- cle is fully automated, but steering wheel action was deactivated during the competition since the GDCD is focused purely on longi- tudinal control. 2.1. GCDC competition The GCDC is the first international competition in Europe with teams in the field of cooperative driving. It was held on motorway 270, between Helmond and Eindhoven (Holland). The GCDC0s objectives are to accelerate the implementation of these systems and significantly contribute to improving traffic problems. The participating teams developed different longitudinal con- trol strategies for platooning in highway scenarios. The aim of such control systems is to maintain a safe distance with respect to the vehicles in front. The vehicles exchanged information over a wireless network in the different trials. Fig. 2 shows the position of the vehicles at the beginning of each heat. When the traffic light of Platoon B (second light in the figure) turns green,1 both lines of cars of this platoon move to join the cars of Platoon A. The highway test start when the Platoon A’s light (first light in the figure) turns green. The two platoons are preceded by a leader vehicle (TNO) belonging to the organization. More than nine teams, from different countries, participated in this edition. Details of the competitions can be found in Kwakkernaat (2011). Platoon B, second light Platoon A, first light 1 2 3 4 5 68 7 Vehicle TNO 4 3 2 1 Fig. 2. Position of the vehicles at the start of every heat. 3. System architecture The AUTOPIA architecture for autonomous vehicles has been described in previous work (Llorca et al., 2011; MilanTs et al., 2010; PTrez et al., 2011). However, some modifications were made in order to satisfy GCDC’s requirements. The perception stage uses the data from sensors. For the vehi- cle’s positioning, an error of up to 1 m was allowed by the compe- tition’s organizers. Bearing this in mind, a differential global positioning system (DGPS) whose error is always less than 0.5 m was used for location. Since the driving area for the competition is completely open, and there are no building or tree canopy occlu- sions, the DGPS was used as the main sensorial input. Some bridges have to be passed under in the competition. Therefore, information from the controller area network (CAN) bus – wheel speeds and acceleration – was used to handle DGPS outages in these zones. Additionally, wireless communication was used. In the planning stage, the appropriate controller is chosen to drive our vehicles through the different traffic situations that may be encountered – overtaking, crossing intersections, merging, etc. It was slightly modified to include the CACC driving and other situations – such as dealing with traffic lights – considered in the competition (Kwakkernaat, 2011). An additional CC controller for high speeds was therefore tuned for use in the heats in which our vehicle was to lead the second platoon, i.e., the front position at the second traffic light (Fig. 2). 3.1. Hardware implementation The vehicle used to participate in GCDC 2011 was a Citroën C3s, called Platero. This platform is fully automated – steering wheel, throttle, and brake – but only the longitudinal controller was needed for the competition. Therefore, the automatic steering wheel system was deactivated during the competition. Fig. 1 shows the vehicle during the GCDC preparation and competition. Fig. 3 shows the control architecture and instrumentation of our vehicle. The central box corresponds to the on-board PC, and all the threads used in our control program are depicted. As was noted above, the perception stage is responsible for communication with the environment and the positioning system. A communication box was added to exchange information with the infrastructure and the other participants of the GCDC 2011 (Kwakkernaat, 2011). It was provided by the organizers to exchange data among all the vehicles, and is based on the IEEE 802.11p protocol. Moreover, Fig. 3 shows the extra peripherals added in response to the GCDC requirements. These are: throttle and brake pedal sensors, emergency button, green and red flashing lights, and soft- ware control of the brake lights. 3.2. Implementation of wireless communications AUTOPIA had already developed the automation of cooperative manœuvres based on wireless communications (MilanTs et al., 2010; PTrez, Milanes, Onieva, Godoy, & Alonso, 2011), sending similar information to that required by the GCDC organization, but with another custom standard to codify and transmit the data. For this occasion, a new communication module based on Abstract Syntax Notation One (ASN.1) coding was created. GCDC communications were organized by means of an intelli- gent communications box running a CALM (Communications Access for Land Mobiles) server (Kwakkernaat, 2011). The organization provided this box with several communications de- vices for connection to the computer, including two UDP sockets. Hence, the programming modifications were done to send and read the packages through a UDP socket. Each message had its own transmission rate. Some of them were periodic, and some were sent only on demand. Some origi- nated in the infrastructure, and some in the vehicles. The ASN.1 standard was devised for compact data transmission with the goal of sending data in a form as compact as possible. Our software was designed to have four levels. The lowest level has the task of generating bit streams from the data, and conversely of decoding the data out of incoming bit streams. The second level is responsi- ble for coding the single data elements. The third level is to code and decode the structured elements. Finally, the fourth level converts internal AUTOPIA data into GCDC data. 4. Control algorithm The core of the system is the design of the controllers. This is based on fuzzy logic whose purpose is to transfer human experi- ence in highway scenarios to the system. Two different situations have to be considered according to the vehicle’s position in each heat (Fig. 2): � Leading a CACC: This case only occurs when the vehicle is located as leader at the second traffic light, i.e., without the organization’s vehicle in front of it. For these heats, a CC system was implemented to implement comfortable acceleration until rejoining Platoon A. Contro program st Fuzzy logic C Receiving GPS data. Reading CAN-vehicle data Information proc Traffic situation a If platoon A Waiting until the first light turns green If we are the leader Change to leader status If we are not the leader I If we p finis Challenge over, Stop the vehicle Fig. 4. Flow chart of the AUTOPIA program soft � Not leading a CACC: This case is considered when the vehicle has some other vehicle immediately in front of it. Under these circumstances, the priority for the vehicle is to react as soon as possible to the action of its leading vehicle. Before designing the controllers for these two situations, a switching law between them had to be established. The minimum distance between cars (dmin) was fixed by the organization at 10 m. Bearing this in mind, the rule to alternate between the two control- lers was set as follows: CC controller if dc > v c t � 0:5amaxt2 � dmin CACC controller if dc 6 v c t � 0:5amaxt2 � dmin ð1Þ where amax is the maximum deceleration allowed (4.5 m/s 2), vc is the current velocity, dc is the current distance between our vehicle and the preceding one, and t is the time gap. Fig. 4 shows the control algorithm’s flow chart as a function of the evolution of each heat. The following subsections describe the software implementation, and the speed controllers for the CC and CACC scenarios. 4.1. Software implementation Fuzzy logic control has become increasingly popular in recent years (Yusofa, Rahmanb, Khalida, & Ibrahimc, in press). Its ability to control imprecise or vague processes has been used to perform the control of numerous industrial applications (Precup & Hellendoorn, 2011). Their main advantage is that fuzzy logic controllers can be easily and intuitively designed using human reasoning as the If platoon B l arts ACC Fuzzy logic CC Receiving/sending information to the vehicles and infrastructure essing nalysis Waiting until the second light turns green If we are the leader f we are not the leader If we join with the platoon A ass the h line ware operation for the GCDC competition. Table 1 Fuzzy rules for the CC at high speed. Speed error Speed Low Medium High Negative Keep Brk_little Brk Central Keep Accel_little Accel_med Positive Accel_med Accel Accel_high knowledge base in order to manage complex plants without an exact or precise model of the system to be controlled. This intuitive behavior on the part of fuzzy-logic-based control- lers can be used to formally define the control of complex systems, i.e., a linguistic description can be made of how a nonlinear plant can be controlled from only knowing its response to some pre- defined inputs. In this sense, the application to different kinds of systems whose model is difficult to obtain analytically, but where one may know how they work thanks to human experience, per- mits these systems to be controlled. This human experience can be considered as the fuzzy-logic-controller’s knowledge base. Since fuzzy controllers form the core of the control system, an embedded fuzzy processor developed by the AUTOPIA program to execute the control was used (MilanTs et al., 2010; PTrez et al., 2011). In designing the controller, the main goal was to implement a control system capable of managing the vehicle’s actuators as humans do, and as intuitively as possible so that the controller can be easily re-adapted to any other item with similar characteristics. 4.2. CC controller The CC controller was developed for the case of being leader at the second traffic light (Fig. 2). It was designed to be capable of fol- lowing any pre-defined speed from 0 to 80 km/h. It uses two input variables: � Speed in km/h. Three different membership functions were defined for low, medium, and high speed to modulate the action on the throttle and brake pedals (upper part of Fig. 5). � Speed error in km/h. This is defined as the difference between the target speed and the current speed. The central membership function was introduced with the goal of performing as smooth Fig. 5. Membership functions of the CC controller. as possible a control around the target speed. Since the main goal of this controller is to keep the vehicle at around such a tar- get speed, the controller was designed to give priority to com- fort rather than minimum speed error (middle part of Fig. 5). The output of the fuzzy controller is the normalized action on the throttle and brake pedals, defined in the range ½�0:3; 0Þ for the brake, and in the range (0, 0.7] for the throttle. These con- straints were set experimentally, taking into consideration the maximum accelerations and decelerations permitted in the competition. Table 1 gives the rule base used for the CC control. Finally, Fig. 7 shows the control surface generated in which the output of the controller, i.e., the action on the throttle and brake pedals, is de- picted as a function of the input variables. 4.3. CACC controller The second controller implemented was designed to perform platooning in highway scenarios. The main goal was to follow the pre-defined distance with respect to the leading vehicle with an error as small as possible, allowing stronger actions than in the CACC case for both the brake and the throttle. Specifically, the maximum action on the throttle was set at 85%, and the max- imum action on the brake at 40%. The inputs for the fuzzy controller were chosen to maintain the vehicle as close as possible to the reference distance. In this context, the controller inputs were: � Distance error in meters. This is calculated as the difference between the real distance with respect to the leading car and the desired distance calculated following the organization’s formula. Four membership functions were defined, three of them in the interval [�1, 1] in order to keep the vehicle as close as possible to the reference. The other was defined for positive values, and was included to try to follow sharp accelerations on the part of the leading vehicle. Its mission was to increase the action on the throttle in these cases since the physical limita- tions of our vehicle made accelerations close to 2 m/s2 difficult to achieve (upper part of Fig. 6). � Speed error in km/h. This is defined as the difference between the leading vehicle’s velocity and the ego-vehicle’s velocity. The middle part of Fig. 6 shows the membership functions. As was the case for the Distance error, there are four membership functions. One is used in the case of high negative values of the variable, increasing the action on the brake pedal. Table 2 gives the rule base used for the CACC control. A further two rules were added to increase the control action in extreme situations (when the information from the other vehicles is lost; when there are failures in the position systems; etc.): IF Speed error Very neg. THEN Output Brake. IF Distance error Very pos. THEN Output Accel. The control surface is shown in Fig. 8. Changes are allowed close to the zero values of both variables. Smoother actions are defined Fig. 6. Membership functions of the CACC controller. Table 2 Fuzzy rules for the CACC at high speed. Speed error Distance error Negative Central Positive Negative Brk Brk_little Keep Central Brk_little Keep Accel Positive Keep Accel_med Accel_high 0 20 40 60 80 −4 −2 0 2 4−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 Speed Error (km/h)Speed (km/h) P ed al Fig. 7. CC controller surface. −35 −30 −25 −20 −15 −10 −5 0 5 0 5 10 15 20 25 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 Speed Error (km/h) Distance Error (m) P ed al Fig. 8. CACC controller surface. Table 3 Error of the experiments relative to the leading vehicle. Exps. Speed error Distance error Mean Median Mean Median Exp. 1 1.13 0.58 1.52 0.91 Exp. 2 3.07 1.64 2.07 0.90 Exp. 3 1.25 0.72 0.88 0.58 for higher values of the two variables, i.e., for very negative speed errors and very positive distance errors, so as to begin adjusting the vehicle’s speed in advance and avoid sharp actions. Heat 1 5.95 2.79 7.67 2.94 Heat 8 3.46 2.51 1.98 1.09 Heat 11 5.75 2.97 13.61 1.71 Heat 17 3.97 2.59 6.73 1.84 5. Experimental results During the days before the GCDC competition, several tests were performed, together with other competitors, on a private track of one of the sponsors in order to tune and adjust the control- lers. Then, during the GCDC weekend, more than 20 heats were participated in with all the teams (Kwakkernaat, 2011). The follow- ing two subsections describe the performance in the experiments of the days prior to the competition and the results at the GCDC, respectively. 5.1. Prior tests During the first pre-competition testing days, our vehicle was capable of exchanging information with other vehicles in order to tune our fuzzy controllers. Until that time, these controllers had only been tested at low speeds (MilanTs et al., 2011; PTrez et al., 2010). Three of the tests performed will be described in this subsection. The first experiment was carried out with other team vehicle leading. Fig. 9 shows the speeds reached. During this test, several speed profiles were tested. Our control algorithm was capable of following the leading vehicle with really good performance.) As noted above, distance and speed errors are used as inputs for the fuzzy controller. Fig. 12 shows the evolution of these two vari- ables during this test. Table 3 lists the mean and median errors of each variable in each experiment and heat of the GCDC explained in this paper. In the first experiment, both values are low, even though this first test included sharp changes. The second experiment is shown in Fig. 10. This was carried out with an aggressive profile. As can be seen, our controller responds 0 100 200 300 400 500 0 10 20 30 40 50 60 70 80 Time (seconds) S pe ed (k m /h ) Leader vehicle (AnnieWAY) AUTOPIA team Fig. 9. Experiment 1: Vehicle speeds – the leader vehicle and AUTOPIA’s behavior. 0 50 100 150 200 250 0 20 40 60 80 100 Time (seconds) S pe ed (k m /h ) AUTOPIA team Leader Vehicle (TNO) Fig. 10. Experiment 2: Vehicle speeds – the leader vehicle and AUTOPIA’s behavior. 0 50 100 150 200 250 300 350 0 50 100 Time (seconds) S pe ed (k m /h ) AUTOPIA team Leader Vehicle (TNO) Fig. 11. Experiment 3: Vehicle speeds – the leader vehicle and AUTOPIA’s behavior. 0 50 100 150 200 250 300 350 400 −20 0 20 S pe ed e rr or (k m /h ) Experiment 1 Experiment 2 Experiment 3 0 50 100 150 200 250 300 350 400 −10 0 10 D is ta nc e er ro r (m ) Time (seconds) Fig. 12. Speed and distance errors in each experiment. 0 50 100 150 200 250 300 0 20 40 60 80 100 Time (seconds) S pe ed (k m /h ) Reference speed Heat 1 0 50 100 150 200 250 300 0 20 40 60 80 100 Time (seconds) S pe ed (k m /h ) Reference speed Heat 8 0 50 100 150 200 250 300 0 20 40 60 80 100 Time (seconds) S pe ed (k m /h ) Reference speed Heat 11 0 50 100 150 200 250 300 0 20 40 60 80 100 Time (seconds) S pe ed (k m /h ) Reference speed Heat 17 Fig. 13. AUTOPIA’s behavior in different starting positions. 0 50 100 150 200 250 −20 0 20 S pe ed e rr or (k m /h ) Heat 1 Heat 8 Heat 11 Heat 17 0 50 100 150 200 250 −10 0 10 D is ta nc e er ro r (m ) Time (seconds) Fig. 14. Speed error and distance error in the heats of Team 3. without delay at that speed. Both the speed and the distance errors are low (Experiment 2 of Fig. 12). The last experiment before the GCDC weekend was done with the leader vehicle using the speed profile of the competition. Although the mean distance error (Table 3) was greater than in the first experiment (due to greater changes in the reference), the median remained almost identical (0.90 m). The competition speed profile covered the first 250 s (Fig. 11). After that, no longer within the competition profile, the leading car accelerated up to 120 km/h. The error increased at the end of the experiment due to the greater accelerations of the leading vehicle (gray line of Fig. 11). The controller reached the reference distance. The values of the errors are listed in Table 3. 5.2. Results at the GCDC The GCDC competition consisted of several heats with vehicles in different starting positions. Fig. 2 shows possible vehicle locations at the beginning of each heat, i.e.: 1. Behind the leader vehicle. 2. In the second position at the first light. 3. In the first position at the second light. 4. In the second position at the second light. To evaluate how the vehicle behaved differently according to its location at the beginning of the heat, we shall present the results will be presented for some different starting positions. Fig. 13 shows one of the best executed heats of our vehicle. In the first profile (Fig. 13, upper left), our system showed the poorest behavior since we started in the last position. The upper right plot of Fig. 13 shows our vehicle following the speed of the leader vehicle. Some jumps in our speed behavior (at around seconds 110 and 170) were due to failures of our position- ing system (to be explained in Section 6). In Table 3, one observes that the speed and distance errors were greater than in the four prior experiments, but still within reasonable limits. In the second platoon heat (lower left plot of Fig. 13), our full system worked correctly. When the vehicle was in the first position (Heat 11), it accelerated until it reached the first platoon. It changed to a constant speed (40 km/h) and joined the first platoon, avoiding unnecessary braking. At the beginning of this heat, the distance error was high (lower plot of Fig. 14) because our vehicle was the leader of the second platoon, but the distance to the last vehicle of the first platoon was constantly monitored. Once the switching condition from the CC to the CACC controller was met, the vehicle joined the first platoon and the distance error was con- sidered as an input to the control. Some jumps occurred at around second 180. These were due to failures of our positioning system when passing under bridges. The last plot (Fig. 13, lower right) shows the behavior in the sec- ond position of the first platoon. In this heat, there was again a large mean distance error in our performance (Table 3). 6. Limitations, difficulties, and lessons learned GCDC was a good scenario in which to test our algorithms against those of various research teams from different countries. The resources we had available for GCDC were limited – just three members in our team, and we were the last to join the competition – but we were able to finish all the heats. Nonetheless, some important failures were detected that need to be resolved. Most of these failures had never happened before on our own test tacks. They were the following: � Problems receiving data volume: Our communication system had been tested for only 3 days before the competition. In par- ticular, these tests involved only up to four other teams. The system worked properly (see Section 5.1). However, in the com- petition heats, with all the information from the other vehicles and the infrastructure being sent simultaneously, our on-board program ran too slowly, and some threads were damaged. � Problems reading data from the CAN bus: Connected with the preceding problem, the most significant failure consisted of errors in reading the vehicle’s CAN bus. These occurred due to saturation of the threads. � Positioning errors: The same thread was used to read the speed of the vehicle from the CAN bus in order to provide the vehicle’s positioning when it was passing under the bridges. Because of the failures in the thread, when the GPS signal was lost under a bridge we were unable to provide accurate positioning. 7. Conclusions As the first European demonstration of cooperative driving systems, GCDC involved nine vehicles, each from a different insti- tution. It permitted us to test our systems outside our own facili- ties, and to translate human driver know-how to our fuzzy control for highway-type scenarios. We drew the following conclu- sions about the different systems involved from our participation in this competition: � Hardware systems: Different cooperative manœuvres had been tested in our facilities involving three (MilanTs et al., 2011) and four of our own vehicles (MilanTs et al., in press), giving good results. Our theoretical results showed our communication system to be capable of managing more than 50 vehicles without overload (MilanTs et al., in press). In the pre-competition tests at the GCDC too, involving up to five vehicles, the system worked properly. In the heats themselves, however, the inclusion of communications from both more vehicles and the infrastructure overloaded our on-board unit. This problem can be solved by including an auxiliary PC for communications management. � Software systems: Both the CC and the CACC fuzzy-logic-based control algorithms gave good results when no hardware prob- lems occurred. Moreover, adjustments were made to them, based on our driving experiences, in a short time. � Positioning system: Unexpected failures due to massive quan- tities of data coming from other vehicles and the infrastructure blocked some threads – the CAN bus controller in particular – and positioning system failures occurred while passing under the bridges. The DGPS was sufficient to meet the organization’s precision requirements, but the additional system designed to function while passing under the bridges did not work (MilanTs, Naranjo, Gonzalez, Alonso, & de Pedro, 2008). � Environment perception: Finally, it is clear that communica- tion systems will play a key role in the future in traffic safety. However, they cannot be relied on as the only source of infor- mation, but need to be combined with vision and laser/lidar systems. Although we were capable of participating in the competition using just communications, more sensors have to be included in order to be tolerant to failures whether due to the leading vehicle or to our own vehicle’s control architecture. Future work will focus on testing our systems with the inclu- sion of more vehicles in order to reproduce the failures detected during the GCDC. Also, new work has been initiated on environ- ment perception with the aim of including more sensors in the architecture. Acknowledgements The authors are grateful to Citroën España S.A. and Tecnología GPS S.A. since, without their collaboration, AUTOPIA’s participation would have been impossible. They also want to express their grat- itude to the TNO organizers for their support before and during the competition. This work was supported by the CYCIT (Spain) and Plan Nacional (Spain) from the GUIADE (P9/08) and TRANSITO (TRA2008-06602-C03-01) projects, respectively. Bansal, M., Das, A., Kreutzer, G., Eledath, J., Kumar, R., & Sawhney, H. (2008). Vision- based perception for autonomous urban navigation. In Proceedings of 11th international IEEE conference intelligent transportation systems ITSC 2008 (pp. 434–440). Belanovic, P., Valerio, D., Paier, A., Zemen, T., Ricciato, F., & Mecklenbrauker, C. F. (2010). On wireless links for vehicle-to-infrastructure communications. IEEE Transactions on Vehicular Technology, 59(1), 269–282. Flemisch, F., Nashashibi, F., Rauch, N., Schieben, A., Glaser, S., Temme, G., et al. (2010). Towards highly automated driving: Intermediate report on the Haveit- joint system. In 3rd European road transport research arena (pp. 1–12). Hosseini Tabatabaei, S. A., Fleury, M., Qadri, N. N., & Ghanbari, M. (2011). Improving propagation modeling in urban environments for vehicular ad hoc networks. IEEE Transactions on Intelligent Transportation Systems, 12(3), 705–716. Koenders, E. (2011). An open communication platform for cooperative systems. In 8th International automotive congress, NL. Kwakkernaat, M. (2011). GCDC 2011 rules and technology document. TNO, tech. rep. Lin, B.-F., Chan, Y.-M., Fu, L.-C., Hsiao, P.-Y., Chuang, L.-A., & Huang, S.-S. (2010). Incorporating appearance and edge features for vehicle detection in the blind- spot area. In Proceedings of 13th international intelligent transportation systems (ITSC) IEEE conference (pp. 869–874). Llorca, D. F., Milanẃs, V., Alonso, I. P., Gavilán, M., Daza, I. G., PTrez, J., et al. (2011). Autonomous pedestrian collision avoidance using a fuzzy steering controller. IEEE Transactions on Intelligent Transportation Systems, 12(2), 390–401. MilanTs, V., Llorca, D. F., Vinagre, B. M., González, C., & Sotelo, M. A. (2010). Clavileño: Evolution of an autonomous car. In Proceedings of 13th international intelligent transportation systems (ITSC) IEEE conference (pp. 1129–1134). MilanTs, V., Villagrá, J., Godoy, J., Sim, J., PTrez, J., & Onieva, E. (in press). An intelligent V2I-based traffic management system. IEEE Transactions on Intelligent Transportation Systems 1–10. MilanTs, V., Godoy, J., Villagrá, J., & PTrez, J. (2011). Automated on-ramp merging system for congested traffic situations. IEEE Transactions on Intelligent Transportation Systems, 12(2), 500–508. MilanTs, V., Naranjo, J., Gonzalez, C., Alonso, J., & de Pedro, T. (2008). Autonomous vehicle based in cooperative GPS and inertial systems. Robotica, 26(5), 627–633. MilanTs, V., PTrez, J., Onieva, E., & González, C. (2010). Controller for urban intersections based on wireless communications and fuzzy logic. IEEE Transactions on Intelligent Transportation Systems, 11(1), 243–248. Mitropoulos, G. K., Karanasiou, I. S., Hinsberger, A., Aguado-Agelet, F., Wieker, H., Hilt, H.-J., et al. (2010). Wireless local danger warning: Cooperative foresighted driving using intervehicle communication. IEEE Transactions on Intelligent Transportation Systems, 11(3), 539–553. Naranjo, J. E., Gonzalez, C., Garcia, R., de Pedro, T., & Haber, R. E. (2005). Power- steering control architecture for automatic driving. IEEE Transactions on Intelligent Transportation Systems, 6(4), 406–415. Naranjo, J. E., Gonzalez, C., Reviejo, J., Garcia, R., & de Pedro, T. (2003). Adaptive fuzzy control for inter-vehicle gap keeping. IEEE Transactions on Intelligent Transportation Systems, 4(3), 132–142. Nguyen, T.-D., Berder, O., & Sentieys, O. (2011). Energy-efficient cooperative techniques for infrastructure-to-vehicle communications. IEEE Transactions on Intelligent Transportation Systems, 12(3), 659–668. Pauwelussen, J., & Feenstra, P. J. (2010). Driver behavior analysis during ACC activation and deactivation in a real traffic environment. IEEE Transactions on Intelligent Transportation Systems, 11(2), 329–338. Precup, R., & Hellendoorn, H. (2011). A survey on industrial applications of fuzzy control. Computers in Industry, 62(3), 213–226. PTrez, J., MilanTs, V., Onieva, E., Godoy, J., & Alonso, J. (2011). Longitudinal fuzzy control for autonomous overtaking. In Proceedings of IEEE internatonal mechatronics (ICM) conference (pp. 188–193). PTrez, J., Milanes, V., Onieva, E., Godoy, J., & Alonso, J. (2011). Longitudinal fuzzy control for autonomous overtaking. In Proceedings of IEEE international conference on mechatronics ICM 2011 (pp. 1–5). PTrez, J., MilanTs, V., & Onieva, E. (2011). Cascade architecture for lateral control in autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 12(1), 73–82. PTrez, J., Seco, F., MilanTs, V., JimTnez, A., Dfaz, J., & de Pedro, T. (2010). An RFID- based intelligent vehicle speed controller using active traffic signals. Sensors, 10(6), 5872–5887. Shladover, S. E. (2007). Path at 20—History and major milestones, 8(4), 584–592. Tan, H., & Bougler, B. (2001). Vehicle lateral warning, guidance and control based on magnetic markers. PATH program, UC Berkeley, Tech. rep. Urmson, C., Anhalt, J., Bae, H., Bagnell, J., Baker, C., Bittner, R., et al. (2008). Autonomous driving in urban environments: Boss and the urban challenge. Journal of Field Robotics, 25(8), 425–466. van Arem, B., van Driel, C. J. G., & Visser, R. (2006). The impact of cooperative adaptive cruise control on traffic-flow characteristics. IEEE Transactions on Intelligent Transportation Systems, 7(4), 429–436. Yusofa, R., Rahmanb, R., Khalida, M., & Ibrahimc, M. (2011). Optimization of fuzzy model using genetic algorithm for process control application. Journal of the Franklin Institute. 
work_ukcbaezzfjcydourjajrqgk22a	----	A Belief Rule Based Expert System to Assess Tuberculosis under Uncertainty J Med Syst (2017) 41: 43 DOI 10.1007/s10916-017-0685-8 MOBILE & WIRELESS HEALTH A Belief Rule Based Expert System to Assess Tuberculosis under Uncertainty Mohammad Shahadat Hossain1 · Faisal Ahmed1 · Fatema-Tuj-Johora1 · Karl Andersson2 Received: 23 November 2016 / Accepted: 9 January 2017 / Published online: 30 January 2017 © The Author(s) 2017. This article is published with open access at Springerlink.com Abstract The primary diagnosis of Tuberculosis (TB) is usually carried out by looking at the various signs and symptoms of a patient. However, these signs and symptoms cannot be measured with 100 % certainty since they are associated with various types of uncertainties such as vague- ness, imprecision, randomness, ignorance and incomplete- ness. Consequently, traditional primary diagnosis, based on these signs and symptoms, which is carried out by the physicians, cannot deliver reliable results. Therefore, this article presents the design, development and applications of a Belief Rule Based Expert System (BRBES) with the ability to handle various types of uncertainties to diag- nose TB. The knowledge base of this system is constructed by taking experts’ suggestions and by analyzing historical data of TB patients. The experiments, carried out, by tak- ing the data of 100 patients demonstrate that the BRBES’s This article is part of the Topical Collection on Mobile & Wireless Health. � Karl Andersson karl.andersson@ltu.se Mohammad Shahadat Hossain hossain ms@cu.ac.bd Faisal Ahmed faisalcsecubd@gmail.com Fatema-Tuj-Johora inna.johora@gmail.com 1 Department of Computer Science and Engineering, University of Chittagong, Chittagong, Bangladesh 2 Pervasive and Mobile Computing Laboratory, Luleå University of Technology, SE-931 87 Skellefteå, Sweden generated results are more reliable than that of human expert as well as fuzzy rule based expert system. Keywords Expert system · Belief rule base · Uncertainty · Tuberculosis · Signs and symptoms Introduction Tuberculosis (TB) is considered as one of the life threaten- ing infectious diseases all over the world, usually, caused by the bacterium Mycobacterium tuberculosis. It is usually two types, namely Pulmonary TB (PTB) and Extra-pulmonary TB (ETB). PTB affects lungs, while ETB can attack any organ of the body except brain, spine, heart, pancreas, skele- tal striated muscle, and thyroid. The rate of occurrence of PTB is much higher than that of ETB [1–3]. In 2014, about 9.6 million people became ill and 1.5 mil- lion died from TB all over the world. It has been observed that over 95 % death from TB occurs in low and middle income countries. It is considered as one of the top five causes of death for women aged between 15 to 44 [4]. The TB bacteria are usually encapsulated as tiny capsules, called tubercles, in the people with healthy immune system. This stage is known as latent TB. In this stage, the bacteria remain inactive and cannot spread to other people. On the contrary, when people’s immune system becomes weak and hence, it is unable to prevent the growth of bacteria. Eventually, TB becomes active in the human body. Only active pulmonary TB is contagious and the bacteria spread into the air through the cough and sneeze of the affected people. However, ETB is not contagious. In case of PTB nearby people can eas- ily be infected during inhaling. TB can be fatal if it is not treated in time, causing serious complications in the lungs, http://crossmark.crossref.org/dialog/?doi=10.1007/s10916-017-0685-8&domain=pdf http://orcid.org/0000-0003-0244-3561 mailto:karl.andersson@ltu.se mailto:hossain_ms@cu.ac.bd mailto:faisalcsecubd@gmail.com mailto:inna.johora@gmail.com 43 Page 2 of 11 J Med Syst (2017) 41: 43 forming hole between the nearby airways, making breath- ing difficult because of blocked airways. The primary signs and symptoms of TB are coughing more than three weeks, coughing up blood, fatigue, unintentional weight loss, chest pain, prolonged fever, lack of appetite and night sweating [1–3]. A physician generally determines the suspicion of TB based on these signs and symptoms. Signs are measured by physician while symptoms are expressed by the patients [5, 6]. Patients usually express the symptoms by using lin- guistic terms such as high, medium and low, which are imprecise, ambiguous and vague. Therefore, these linguis- tic terms cannot express the level of symptoms with 100 % certainty and hence, it inherits the types of uncertainty mentioned. In some cases, patients may ignore the importance of coughing since they consider it is related to other common diseases, which is an example of uncertainty due to igno- rance. The sputum smear microscopy, which is a method to diagnose the presence of active TB, sometimes it is unable to detect. This is an example of uncertainty due to incompleteness. A comprehensive survey has been car- ried out in consultation with the physicians of the various TB hospitals, located in Chittagong District of Bangladesh, to identify the types of uncertainties, associated with each of the signs and symptoms of TB, which are described in Table 1. Since the traditional way of determining suspicion of TB is usually carried out by the physicians by looking at the signs and symptoms, it does not consider the above uncertain phenomenon. Thus, the method jeopardizes the accuracy of the detection of TB. However, an expert sys- tem which emulates the decision making process of human being can be considered as an appropriate tool to address the uncertain phenomenon to accurately detect the suspicion of TB. An expert system consists of two components, namely knowledge-base and the inference mechanisms. However, such an expert system should have the knowledge repre- sentation schema to acquire uncertain clinical knowledge. At the same time, inference mechanism should have the robust reasoning algorithms with the capability to handle various types of uncertainties as mentioned. Therefore, in this study the development of a belief rule-based expert sys- tem has been considered, where belief rule base used to handle uncertain knowledge and the evidential reasoning is used as the inference mechanism. The remaining of the article is organized as follows. Section “Literature review” presents the literature review, “An overview of belief rule based expert system’s method- ology” gives an overview of the belief rule based expert sys- tems methodology, and “BRBES to diagnose TB” describes the design, architecture, knowledge-base construction along with an overview of the system interface. Section “Results and discussion” includes the results and discussion, while “Conclusion” concludes the article. Literature review Several studies have been conducted with reference to the diagnosis of Tuberculosis (TB). Multilayer neural networks (MLNN) structures were used to facilitate the analysis of tuberculosis [7]. Back-propagation with momentum and Levenberg-Marquardt algorithms were used to perform the task of training in the MLNN. However, in this approach, an absence of explicit relationship between the input and output data noticed, which is necessary to ensure the trans- parent diagnosis of TB. Moreover, this approach has not considered the uncertainty issues, associated with the signs and symptoms, related to the input data. The ensemble classifiers such as Bagging, AdaBoost and Random Forest tree were used to estimate the performance in detecting pulmonary tuberculosis [8]. Coughing sound detection algorithm [9] and lung auscultation software [10] were also used to TB diagnosis. The above approaches con- sidered most of the signs and symptoms as appear in Table 1 in diagnosing TB. However, they lack the procedures to measure these signs and symptoms by taking account of the various uncertainties; rather there measuring approach is Boolean in nature. In [11], K-mean clustering was combined with the dif- ferent classifiers, namely, Support Vector Machine (SVM), Naı̈ve Bayesian Classifier and K-Nearest Neighbor (K-NN) to support the prediction of tuberculosis [5]. Moreover, there are systems developed to accurately classify tumor and epilepsy using Genetic Algorithm by combining multiclass classification method and Support Vector Machine (SVM) [12, 13]. However, SVM lacks the transparency [14, 15] since by using this method the relationship between the signs and symptoms of the patient and the diagnosis cannot be estab- lished in an understandable way. However, in case of TB diagnosis, a continuous generation of scenario to establish the relationship between the signs and symptoms and the treatment as well as with the diagnostic process is essential [16, 17]. Fuzzy Cluster Means (FCM or Fuzzy C-Mean) analysis [18, 19] used a clustered data set for identifying differ- ent types of TB. TUBERDIAG [20] is also a fuzzy rule based system for detecting TB which combines positive and negative knowledge. Although fuzzy based expert system can handle uncertainties due to vagueness, imprecision and ambiguity, they are not equipped to address uncertainties due to ignorance, incompleteness and randomness, which are the cases with the signs and symptoms of TB as shown in Table 1. J Med Syst (2017) 41: 43 Page 3 of 11 43 Table 1 Types of uncertainties associated with tuberculosis Sign Types of uncertainty Description Coughing Ignorance Most patients ignore coughing since they consider it is related to a common disease Imprecision Most patients ignore coughing since they consider it is related to a common disease Incompleteness “Sputum smear microscopy” is the primary method to diagnose pulmonary TB. However, the method is unable to detect half of all active TB cases Coughing up blood Ignorance Sometimes blood remains unnoticed due to thick and larger amount of phlegm Fatigue Imprecision, vagueness, ambiguity People consider fatigue as a normal symptom of overwork- ing and do not consult with doctor Prolonged fever Ignorance, vagueness, randomness Usually patients can not exactly tell the duration and temperature when doctor asks about the fever Night sweating Ignorance, vagueness Most people think that warm weather, humidity or wear- ing heavy cloths are liable for night sweating and hide the symptom from the doctor In addition, the above mentioned approaches [7–11, 14– 20] lack the appropriate inference procedures to address different types of uncertainties such as vagueness, impreci- sion, ambiguity, ignorance, incompleteness and randomness in a single integrated framework. Such an integrated frame- work has an important role to make robust decision to support TB diagnostic decision making process. Therefore, it is necessary to employ an appropriate knowledge representation schema to capture uncertain knowledge that exists with the signs and symptoms of TB. Belief Rule Base (BRB) is widely used to represent this type of knowledge [21–23]. In addition, BRB can be used to demonstrate the explicit non-linear relationship between the input-output data which is necessary to ensure the trans- parent diagnosis of TB. Evidential reasoning in combination with the BRB can be used as the inference mechanism of the expert system which has the capability to handle all types of uncertainties in an integrated framework [20–28]. There- fore, the following section will represent the belief rule based expert systems (BRBESs) methodology, consisting of knowledge-base construction and the inference procedures. An overview of belief rule based expert system’s methodology An expert system is mainly consists of two components, namely, knowledge-base and the inference procedures. In BRBESs, belief rule-base is used to represent the domain knowledge under uncertainty. On the other hand, inference procedures of BRBEs consists of input transformation, rule activation weight calculation, belief update and rule aggre- gation using evidential reasoning [21]. Each of them will be elaborated below. Domain knowledge representation using BRB A belief rule is the extension of traditional IF-THEN rule, where a belief structure is used in the consequent part. Antecedent part of the belief rule consists of one or more antecedent attributes with associated referential values, while consequent part consists of one consequent attribute. Knowledge representation parameters such as rule weight and antecedent attribute weight are used. A belief rule base (BRB) consists of one or more than one belief rules. The reason for adopting IF-THEN rule is that it is considered as an appropriate mechanism to represent human knowledge [21]. In addition, non-linear causal relationship between the antecedent and consequent attributes can be established in a belief-rule-base. Equation 1 represents an example of belief rule. Rk : ⎧ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩ IF (A1 is A k 1) ∧ (A2 is Ak2) ∧ ... ∧ (ATk is AkTk ) THEN ⎧ ⎪⎪⎪⎪⎪⎪⎨ ⎪⎪⎪⎪⎪⎪⎩ {(C1, βk1 ), (C2, βk2 ), ..., (CN , βkN )}, (( N∑ n=1 βkn ≤ 1)), with rule weight 0 ≤ θk ≤ 1, and attribute weight δk1 , δ k 2 , ..., δ k T ≥ 0 satisfying Tk∑ i=1 δk i = 1 (1) where, A1, A2, ..., A k Tk , Tk ∈ {1, 2, ..., T }, are the antecedent attributes used in the kth rule. Ak i ∈ {Ai1 , Ai2 , ..., Aiji } is the referential value of antecedent attribute. C1, C2, ..., CN are the referential values of the con- sequent attribute while βki is the belief degree to which Ci 43 Page 4 of 11 J Med Syst (2017) 41: 43 is believed to be true. If N∑ i=1 βki = 1, the belief rule is said to be complete, otherwise it is incomplete. In this way, BRB addresses the uncertainty resulting from the incompleteness. Equation 2 represents the example of a belief rule from the domain of TB. Here, the consequent attribute is “TB Suspicion” with three referential values consisting of “High”, “Medium”, and “Low” with degree of belief 0.8, 0.2, and 0.0. The rule is said to be complete since the summation of the belief degrees stands at 1. ⎧ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩ IF ⎧ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩ (Coughing more than 3 weeks is medium) and (Coughing up blood is High) and (Chest pain is High) and (Fatigue is Low) and (Prolonged fever is medium) and (Lack of appetite is Low) and (Weight loss is High) and (Night sweating is Low) THEN TB suspicion is (High, 0.8), (Medium, 0.2), (Low, 0.0) (2) Each antecedent attribute of this rule also consists of three referential values and hence, the total number of rules in this BRB can be calculated by applying Eq. 3, which is 6,561. L = T∏ i=1 (Ji ) (3) where L is the total number of rules in a sub rule base and Ji is the referential values of the ith antecedent attribute. BRBESs inference procedures Each of the inference procedures used in a BRBES is described below. Input transformation Input transformation consists of the distribution of the input value of an antecedent attribute over its different referential values by applying Eq. 4. The distributed value with each referential values of the antecedent attribute is called match- ing degree or the degree of belief. It is interesting to note that when the value of this matching degree is calculated for each of the referential values of the antecedent attribute (AA), this value is assigned to only those rules where this referential value exits with the AA. As an example, Eq. 2 consists of eight antecedent attributes, each with three referential values. When the input data for one antecedent attribute is collected from the TB patient, its matching degrees to the corresponding referen- tial values are calculated by applying Eq. 4. Consequently, the matching degree, related to a referential value corre- sponding to the antecedent attribute of a rule is assigned. In this way, input data for the eight antecedent attributes can be collected for a patient and their corresponding match- ing degree is assigned in a rule. Once the rule is assigned with the corresponding matching degree then it is said to be active and the rule is called packet antecedent. This phe- nomenon can be described that the rule is in the RAM while the initial rule base is in the secondary memory. αih = u(Aih+1) − A∗i u(Aih+1) − u(Aih) and αih+1 = A∗ i − u(Aih) u(Aih+1) − u(Aih) (4) where u(Aih+1) and u(Aih) are grade values of Aih+1 and Aih respectively. Table 2 shows the matching degree of the antecedent attribute value into its different referential values. For exam- ple, the input value of the antecedent attribute ”Cough” is identified as ”low”, which is weighted as 10 % by the expert and its corresponding matching degrees associated with the referential values (in this case they are “High”, “Medium” and “Low”) obtained by using Eq. 4. Calculation of activation weights Rule activation weight calculation consists of calculating the combined matching degree (αk ) as well as the weight of a rule in the BRB. Equation 2 consists of eight antecedent attributes and hence, it is important to calculate their com- bined matching degree, which can be calculated by using multiplicative function [as shown in Eq. 5], allowing the inter-relationship between the attributes [21]. αk = Tk∏ i=1 (α k i ) δ k i , δ k i = δk i max i=1,...,Tk (δ k i ) (5) so that 0 ≤ δk i ≤ 1 where δk i (i = 1, ..., Tk) is the rela- tive weight of the ith antecedent attribute in the kth belief rule. Tk is the total number of antecedent attributes in the kth rule. Here, δk i = 0 is meaning that the attribute which has zero importance and hence, no impact on the aggre- gation process, while δk i = 1 demonstrates the significant impact. Moreover, overall belief increases with the incre- ment of the individual belief of the antecedent attributes. The rule activation weight is calculated by using Eq. 6 [21]. Wk = {θkαk} { L∑ i=1 θi αi } (6) J Med Syst (2017) 41: 43 Page 5 of 11 43 where θk is the relative weight of the kth rule and L is the total number of belief rule in the belief rule-base. When the Wk of a rule is zero then it has no impact in the BRB while it is “1” then its important is high. Belief degree update Equation 2 shows the presence of eight antecedent attributes, necessary to assess the suspicion of TB. However, there could be some situation when the data of the some attributes could not be available. In such situation, the initial belief degrees that were assigned to the referential values of the consequent attribute needs to be updated by applying Eq. 7 [21]. This phenomenon is an example uncertainty due to ignorance. βki = βki ∑T k t=1(τ (k, t) ∑Ji j =1 αtj ) ∑Tk t=1 τ (k, t) (7) where τ (k, t) = ⎧ ⎨ ⎩ 1, if used in defining Rk(t = 1, ..., Tk) or 0, otherwise Here, βki is the original belief degree while βki is the updated belief degree. The original belief degree is updated while any ignorance is noticed. For example, if the antecedent “cough” is ignored, then the initial belief degrees are updated as shown in Table 3. Inference using evidential reasoning In order to obtain the aggregated value of the referential val- ues of the consequent attribute, based on the input data of the antecedent attributes, either recursive or analytical evi- dential reasoning (ER) algorithms as shown in Eq. 8 can be applied [22]. To reduce the computational complexity ana- lytical ER algorithm is found effective to calculate the final belief degree βj . βj = μ × [∏Lk=1(ωkβkj + 1 − ωk ∑N j =1 βkj ) − ∏L k=1(1 − ωk ∑N j =1 βkj )] 1 − μ × [∏Lk=1(1 − ωk)] (8) with μ = [∑Nj =1 ∏L k=1(ωkβkj + 1 − ωk ∑N j =1 βkj ) − (N − 1) × ∏Lk=1(1 − ωk ∑N j =1 βkj )]−1. The final combined result or output generated by ER is represented by {(C1, β1), (C2, β2), ..., (CN , βN )} where βj is the final belief degree attached to the j th referential value Cj of the consequent attribute C, which is obtained after all activated rules in the BRB are combined by using ER. This output can be converted into a crisp/numerical value [as shown in Eq. 9] by assigning a utility score to each referential value of the consequent attributes [20]. H (A ∗ ) = N∑ j =1 u(Cj )βj (9) where H (A∗) is the expected score expressed as a numerical value and u(Cj ) is the utility score of the j th referential value. BRBES to diagnose TB This section presents the architecture along with imple- mentation strategy of the belief rule-based expert system (BRBES) to diagnose Tuberculosis (TB). This is followed by the presentation of the knowledge-base construction as well as a description of the BRBES’s interface. Architecture and implementation A system architecture can be defined how its compo- nents are organized. It is also important to know the pattern of system organization, which is known as archi- tectural style. BRBES presented in this article follows three-layer architecture, consisting of web-based interface, application and database management layers as shown in Fig. 1. Web-based interface: Since the BRBES to be used by the physicians, patients and researchers at various hospitals of Bangladesh, especially in the rural areas, a web-based user-friendly interface is necessary. Therefore, to ensure the usability of the system by the mentioned users at any time and at any place, a web-based interface has been devel- oped for the BRBES. This interface has been developed by integrating various web technologies such as Javascript, Jquery, HTML and CSS. The application layer, which facil- ities inference and database access, has been developed by using PHP because of its simplicity, shorter develop- ment cycle and it can be used through online. The data- base management layer, which consists of clinical data and knowledge-base, developed by using MySQL, which is a relational database management system. MySQL is exible, user friendly and ensures security as well as faster data access. 43 Page 6 of 11 J Med Syst (2017) 41: 43 Table 2 Input Transformation Serial No. Antecedent Name Antecedent Value High Medium Low 1 Cough Low, 10 % 0 0.2 0.8 2 Blood with cough High, 60 % 0.2 0.8 0 3 Chest pain High, 80 % 0.6 0.4 0 4 Fatigue Low, 30 % 0 0.6 0.4 5 Fever High, 85 % 0.7 0.3 0 6 Lack of appetite Medium, 50 % 0 1 0 7 Weight loss High, 90 % 0.8 0.2 0 8 Night sweating Low, 15 % 0 0.3 0.7 Knowledge base construction in BRB The knowledge-base construction consists of developing a BRB tree by identifying the necessary antecedent and con- sequent attributes. Figure 2 shows the single level BRB structure to diagnose TB. The leaf nodes represent the antecedent attributes while root node consequent attribute. The eight antecedent attributes, having three referential val- ues each, are identified and they are verified in consultation with the physicians, located at the various hospitals of Chittagong City of Bangladesh. The BRB consists of 6,561 rules since it comprises eight antecedent attributes each with three referential values. The number of rules are calculated by using Eq. 3. A BRB usu- ally can be established [20] by acquiring domain expert knowledge, by collecting historical data, by using earlier rules if they are available and by developing rules in a random way without any prior-knowledge. In this study, the initial BRB has been constructed by acquiring knowledge of physicians or the domain experts as well as by using patients’ historical data. Each rule of the BRB has given rule weight “1” while each antecedent attribute’s weight is considered as “1”. An example of such rule is illustrated in Eq. 2. Table 4 illustrates the initial BRB of the BRBES. BRB interface An interface can be defined as a media, facilitating the users to interact with the system. Figure 3 illustrates the interface of the BRBES to diagnose TB, allowing the acquisition of the input value, associated with each of the eight antecedent Table 3 Belief Degree Update Rule Id High Medium Low 2 Initial 0 0.6 0.4 Update 0 0.48 0.32 Fig. 1 Architecture of BRBES Fig. 2 BRB Framework to Assess TB J Med Syst (2017) 41: 43 Page 7 of 11 43 Table 4 A Sample of Initial Belief Rule-Base for Assessment of Tuberculosis Suspicion Rule No Rule Weight Antecedents Consequent IF Then A1 A2 A3 A4 A5 A6 A7 A8 TB suspicion High Medium Low 1 1 H H H H L H H H 1.0 0.0 0.0 2 1 H L M L H L L L 0.0 0.6 0.4 3 1 L H L M H L L M 0.0 0.1 0.9 4 1 L L L L H L L L 0.0 0.0 1.0 5 1 H M H L H L L H 0.1 0.5 0.4 6 1 M H H L M L H L 0.8 0.2 0.0 7 1 H M M M H L L M 0.1 0.7 0.2 8 1 M L L L H M M L 0.1 0.6 0.3 9 1 H H H H H H H M 0.4 0.6 0.0 10 1 L M H H H M M L 0.0 0.1 0.9 attributes, either from the patients or from the physicians. Figure 3 illustrates the matching degree, associated with the antecedent attributes with reference to the data of the first patient as shown in Table 5. For example, the input value of the antecedent attribute A1 (coughing more than three weeks) is obtained as “98”. This is acquired by asking the patient about the intensity level of the A1, which is in the range of 0-100. This input value is then distributed over the three referential values of A1, which are “(High, 0.98)”, “(Medium, 0.0)” and “(Low, 0.2)”. This distributed value is called the matching degree obtained by using Eq. 4. The suspicion level of the TB of this patient is obtained by using Eqs. 5, 6, 7 and 8. This is measured as degree of belief associated with each of the referential values of the suspicion level of Tuberculosis and found as (High, 90.72), (Medium, 0.06), and (Low, 0.22). These fuzzy values are converted into crisp value by using Eq. 9 and obtained as 95.25 %. Fig. 3 BRBES interface to assess TB 43 Page 8 of 11 J Med Syst (2017) 41: 43 Table 5 TB Suspicision by Fuzzy Logic, BRBES, and Expert SL. NO (1) A1 (2) A2 (3) A3 (4) A4 (5) A5 (6) A6 (7) A7 (8) A8 (9) Fuzzy Result(%) (10) BRBES Result(%) (11) Expert Opinion(%) (12) Bench- mark (13) 1 98 95 90 80 98 87 9.8 85 76.34 95.25 70 1 2 95 15 45 22 103 30 1.0 15 54.72 51.64 52 1 3 25 95 15 55 102.5 20 1.8 60 58.71 42.93 55 0 4 10 30 27 18 101.5 25 3.5 12 32.67 19.47 45 0 5 18 60 87 12 99.5 15 1.5 80 38.81 33.28 41 1 6 90 86 98 18 98 17 9.7 23 65.53 72.45 67 1 7 94 58 67 45 102 16 2.5 67 70.77 74.67 70 1 8 85 32 13 20 104 78 1.2 23 49.23 64.13 46 1 9 92 97 75 87 98 96 8.9 55 72.56 91.63 69 1 10 15 67 60 92 101.5 67 1.2 15 45.61 32.01 50 0 11 90 87 69 60 103 75 4 60 73.03 84.05 75 1 12 87 46 82 77 99 68 4 76 84.45 84.77 82 1 13 75 80 90 74 101.5 90 2 83 82.12 85.18 77 1 14 46 70 82 91 104 40 1 75 56.32 69.86 60 1 15 70 80 85 70 103 66 3 68 78.71 80.26 81 1 Results and discussion To demonstrate the applicability and the reliability of the BRBES to diagnose tuberculosis (TB), the system fed with the input data received from the TB patients of a hospital located in the Chittagong City of Bangladesh (Fig. 4). The input data, associated with the eight attributes, of 77 TB patients have been collected. The real laboratory test results of those patients collected and they were con- sidered as the benchmark data. If TB is found positive in laboratory test result for a patient, then the benchmark is considered as “1”, otherwise it is “0” for that patient. Table. V illustrates the collected data on the eight attributes of a patient (column 2 through 9) along with level of TB suspi- cion generated by the BRBES (Column 11). Table. V also illustrates the expert opinion on the level of TB suspicion Fig. 4 Tuberculosis Hospital, Chittagong, Bangladesh J Med Syst (2017) 41: 43 Page 9 of 11 43 Fig. 5 ROC Curves Comparing the result of BRBES and Expert Data (Column 12) and the benchmark data is recorded in col- umn 13. In order to compare the reliability of the BRBES’s results a Fuzzy Rule Based Expert System (FRBES) to measure the suspicion level of TB developed in MatLab environment and the results generated for the same data by using FRBES are recorded in Column 10. For sim- plicity, Table 5 only presents the data of 15 patients out of 77. Fig. 6 ROC Curves Comparing the Result of BRBES, FRBES and Expert Data 43 Page 10 of 11 J Med Syst (2017) 41: 43 The Receiver Operating Characteristic (ROC) curves are usually used to analyze the accuracy and reliability of the diagnostic tests having ordinal or continuous results. There- fore, the method was considered, to measure the reliability of BRBES in comparison with expert opinion and FRBES. The accuracy of a system to assess the level of suspicion of the TB can be measured by calculating the Area Under Curve (AUC) [29, 30]. For example, if the AUC is found to be one for the results generated by the BRBES then the system can be considered as 100 % reliable. Figure 5 illus- trates the ROC curves plotted for both the BRBES and the Expert Opinion. The ROC curve plotted by the blue line in this figure is associated with the results generated by the BRBES with AUC of 0.910 (95 % confidence intervals 0.848 - 0.972). The ROC curve plotted by the green line in Fig. 5 is obtained against the physician’s opinion, and its AUC is 0.710 (95 % confidence intervals 0.587 - 0.815). However, Fig. 6 illustrates the ROC curves for BRBES, FRBES and Expert Opinion. The ROC curve plotted by the red line in Fig. 6 is obtained against the FRBES and its AUC is 0.777 (95 % confidence intervals 0.680 - 0.873). While the AUC values of Fig. 6 for both the BRBES and the Expert Opinion are same as of Fig. 5. Table 6 summarizes the above results associated with BRBES, FRBES and Expert Opinion. From Figs. 5 and 6 as well as from Table 6 it can be observed that AUC of Expert Opinion is much less than from both the BRBES and FRBES. The reason for this is that dur- ing the interviewing and conversation with the physicians it has been observed that they are not aware of the uncertainty issues related to the signs and symptoms of the TB. The reason for this is that during the interviewing and conversation with the physicians it has been observed that they are not aware of the uncertainty issues related to the signs and symptoms of the TB. Therefore, the reliability of their assessment level of TB suspicion is much lower than that of BRBES and FRBES. From the Table 6 as well as from Figs. 5 and 6, it can also be observed that AUC of BRBES is much larger than that of FRBES. The reason for this is that fuzzy rule based expert system (FRBES) only considers uncertainty due to the vagueness, imprecision and ambiguity. However, BRBES includes uncertainties due to Table 6 Reliability Comparison among Three Systems Test Result Variables Area Asymptotic 95 % Confidence Interval Lower Bound Upper Bound BRBES 0.910 0.848 0.972 Fuzzy System 0.777 0.680 0.873 Expert Data 0.701 0.587 0.815 the ignorance, incompleteness and randomness in addition to the vagueness, imprecision and ambiguity. Further, the inference procedures of the FRBES which uses either Mamdanior T-S methods are not equipped to pro- cess uncertainty issues during the reasoning process. On the contrary, BRBES uses evidential reasoning procedures as the inference engine which is equipped to handle types of uncertainties mentioned before. Conclusion This article described the design, implementation and appli- cations of a Belief Rule Based Expert system (BRBES), allowing the measurement of the level of suspicion of TB by taking account of its various signs and symptoms. The system allows the understanding of the relationship between signs and symptoms and the level of suspicion of TB of a patient in an explicit and transparent way. This will allow the identification of the signs and symptoms those are respon- sible for increasing the suspicion level of TB of a patient. In this way, various scenarios by taking account of the signs and symptoms of a patient from various perspectives can be carried out by using the BRBES. The physicians will eventually be able to select the appropriate medicine for a patient. In addition, all types of uncertainties such as vague- ness, imprecision, ambiguity, ignorance and incompleteness can be handled in an integrated framework which makes the system more robust as evident from the comparative results generated by using both the manual system and the fuzzy rule based expert system as illustrated in Table 6. There- fore, the BRBES can provide a decision support platform to the physicians and would serve a savior by offering primary health care to the people with reduced time and low diagno- sis cost. In future, the BEBES would be extended to support optimal learning by training the knowledge representation parameters such as rule weight, attribute weight, and belief degrees. Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http:// creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Compliance with Ethical Standards Conflict of interests Author Mohammad Shahadat Hossain declares that he has no conflict of interest. Author Faisal Ahmed declares that he has no conflict of interest. Author Fatema-Tuj-Johora declares that she has no conflict of interest. Author Karl Andersson declares that he has no conflict of interest. http://creativecommons.org/licenses/by/4.0/ http://creativecommons.org/licenses/by/4.0/ J Med Syst (2017) 41: 43 Page 11 of 11 43 Funding This study was funded by Swedish Research Council under Grant 2014-4251. Ethical approval All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. Informed consent Informed consent was obtained from all individ- ual participants included in the study. References 1. Walker, B.R., Colledge, N.R., Ralston, S.H., and Penman, I.D. Davidson’s principles and practice of medicine. 22nd ed. United Kingdom: Elsevier Health Sciences, 2014. 2. Goldman, L., and Bennet, J.C. Cecil textbook of medicine. 24th ed. Philadelphia: Elsevier Saunders, 2012. 3. Walker, H.L., Chen-Yuan, C., Gie, R.P., and Enarson, D.A. Crofton’s clinical tuberculosis. 3rd ed. London: Macmillan Edu- cation, 2009. 4. WHOTuberculosis, http://www.who.int/mediacentre/factsheets/ fs104/en/, accessed: January 7, 2017. 5. Karim, R., Hossain, M.S., Khalid, M.S., Mustafa, R., and Bhuiyan, T.A.: A belief rule based expert system to assess bronchiolitis suspicion from signs and symptoms under uncertainty. In: SAI Intelligent systems conference (intellisys). London, UK, 2016. 6. Hossain, M.S., Zander, P.-O., Kamal, M.S., and Chowdhury, L., Belief-Rule-Based Expert systems for evaluation of E- Government: a case study. Expert. Syst. 32(5):563–577, 2015. 7. Er, O., Temurtas, F., and Çetin Tannkulu, A., Tuberculosis disease diagnosis using artificial neural networks. J. Med. Syst. 34(3):299– 302, 2010. 8. Rusdah, E., and Wardoyo, R.: Winarko Preliminary diagnosis of pulmonary tuberculosis using ensemble method. In: Interna- tional conference on data and software engineering (ICoDSE). Yogyakarta, Indonesia, 2015. 9. Tracey, B.H., Comina, G., Larson, S., Bravard, M., López, J.W., and Gilman, R.H.: Cough detection algorithm for monitoring patient recovery from pulmonary tuberculosis. In: 2011 Annual international conference of the IEEE engineering in medicine and biology society. Boston, MA, USA: EMBC, 2011. 10. Lestari, R., Ahmad, M., Alisjahbana, B., and Djatmiko, T.: The Lung Diseases Diagnosis Software: Influenza and Tuberculosis Case Studies in The Cloud Computing environment. In: 2012 International conference on cloud computing and social network- ing (ICCCSN). Bandung, Indonesia, 2012. 11. Asha, T., Natarajan, S., and Murthy, K.N.B.: A data mining approach to the diagnosis of tuberculosis by cascading clustering and classification. J. Comput. 3(4), 2011. 12. Shen, C.P., Lin, J.W., Lin, F.S., Lam, A.Y.Y., Chen, W., Zhou, W., Sung, H.Y., Kao, Y.H., Chiu, M.J., Leu, F.Y., and Lai, F.: GA-SVM Modeling of multiclass seizure detector in epilepsy analysis system using cloud computing. Soft. Comput. 1–11, 2015. doi:10.1007/s00500-015-1917-9. 13. Lin, F.S., Shen, C.P., Liu, C.H., Lin, H., Huang, C.Y.F., Kao, C.Y., Lai, F., and Lin, J.W., A high performance multiclass classifica- tion framework using cloud computing architecture. Journal of Medical and Biological Engineering 35(6):795–802, 2015. 14. Calzada, A., Liu, J., Wang, H., Nugent, C., and Martinez, L., Application of a spatial intelligent decision system on Self-Rated health status estimation. J. Med. Syst. 39(11):1–18, 2015. 15. Huysmans, J., Dejaeger, K., Mues, C., Vanthienen, J., and Baesens, B., An empirical evaluation of the comprehensibility of decision table, tree and rule based predictive models. Decis. Support. Syst. 51(1):141–154, 2011. 16. Ainon, R.N., Bulgiba, A.M., and Lahsasna, A., AMI Screening using linguistic fuzzy rules. J. Med. Syst. 36(2):463–473, 2012. 17. Bojarczuk, C.C., Lopes, H.S., and Freitas, A.A., Genetic program- ming for knowledge discovery in chest pain diagnosis. IEEE Eng. Med. Biol. Mag. 19(4):38–44, 2000. 18. Soundararajan, K., Sureshkumar, S., and Anusuya, C., Diagnos- tics decision support system for tuberculosis using fuzzy logic. International Journal of Computer Science and Information Tech- nology & Security (IJCSITS) 2(3):684–689, 2012. 19. Imianvan, A.A., and Obi, J.: Fuzzy cluster means expert system for the diagnosis of tuberculosis. Global J. Comp. Sci. Technol. 11(6), 2011. 20. Phuong, N.H., Hung, D.H., Tuan, D.T., Co, N.V., Duong, B.D., Thinh, P.T., Hoa, N.P., and Linh, N.N.: Designing an experimen- tal expert system for lung tuberculosis diagnostics using fuzzy set theory. In: 1998 IEEE International conference on systems, man, and cybernetics. San Diego, CA, USA, 1998. 21. Yang, J.-B., Liu, J., Wang, J., Sii, H.-S., and Wang, H.-W., Belief rule-base inference methodology using the evidential rea- soning approach-rimer. IEEE Transactions on Systems Man, and Cybernetics-part A: Systems and Humans 36(2):266–285, 2006. 22. Hossain, M.S., Monrat, A.A., Hasan, M., Karim, R., Buhiyan, A.T., and Khalid, M.S.: A belief rule based expert system to assess mental disorder under uncertainty. In: 2016 5th interna- tional conference on informatics, electronics & vision (ICIEV). Dhaka, Bangladesh, 2016. 23. Hossain, M.S., Andersson, K., and Naznin, S.: A belief rule based expert system to diagnose measles under uncertainty. In: Proceed- ings of the 2015 international conference on health informatics and medical systems (HIMS 2015). Las Vegas, Nevada, USA, 2015. 24. Hossain, M.S., Hasan, M.A., Uddin, M., Islam, M.M., and Mustafa, R.: A belief rule based expert system to assess lung cancer under uncertainty. In: 2015 18Th international confer- ence on computer and information technology (ICCIT). Dhaka, Bangladesh, 2015. 25. Karim, R., Andersson, K., Hossain, M.S., Uddin, M.J., and Meah, M.P.: A belief rule based expert system to assess clinical bron- chopneumonia suspicion. In: Future Technologies Conference 2016 (FTC 2016). San Francisco, CA, USA, 2016. 26. Hossain, M.S., Haque, M.A., Mustafa, R., Karim, R., Dey, H.R., and Yousuf, M.: An expert system to assist the diagnosis of ischemic heart disease. Int J Integr Care, 2016. 27. Hossain, M.S., Khalid, M.S., Akter, S., and Dey, S.: A belief rule-based expert system to diagnose influenza. In: 2014 9Th international forum on strategic technology (IFOST). Cox’s Bazar, Bangladesh, 2014. 28. Rahaman, S., and Hossain, M.S.: A belief rule based clinical deci- sion support system to assess suspicion of heart failure from signs, symptoms and risk factors. In: 2013 International conference on informatics, electronics & vision (ICIEV). Dhaka, Bangladesh, 2013. 29. Metz, C.E., Basic principles of ROC analysis. Semin. Nucl. Med. 8(4):283–298, 1978. 30. Hanley, J.A., The robustness of the binormal assumptions used in fitting ROC curves. Med. Decis. Mak. 8(3):197–203, 1988. http://www.who.int/mediacentre/factsheets/fs104/en/ http://www.who.int/mediacentre/factsheets/fs104/en/ http://dx.doi.org/10.1007/s00500-015-1917-9 A Belief Rule Based Expert System to Assess Tuberculosis under Uncertainty Abstract Introduction Literature review An overview of belief rule based expert system's methodology Domain knowledge representation using BRB BRBESs inference procedures Input transformation Calculation of activation weights Belief degree update Inference using evidential reasoning BRBES to diagnose TB Architecture and implementation*.3pt Knowledge base construction in BRB BRB interface Results and discussion Conclusion Open Access Compliance with Ethical Standards Conflict of interests Funding Ethical approval Informed consent References 
work_ul32nhcuong5tjoo2sljvdc5nm	----	PII: 0888-613X(89)90009-1 A Valuation-Based Language for Expert Systems Prakash P. Shenoy S c h o o l o f Business, Umverstty o f Kansas, Lawrence, Kansas A B S T R A C T A new language based on valuations ts proposed as an alternatwe to rule-based languages f o r constructing knowledge-based systems. Valuation-based languages are superior to rule-based languages f o r maintaining consistency m the knowledge base, f o r cachmg references, f o r managmg uncertainty, and f o r nonmonotomc reasonmg. A n abstract description o f a valuatzon-based language is gwen Two specifw instances o f valuation-based languages are described. The first ts designed to represent categortcal knowledge. The ablhty o f such a language to mamtam conststency and cache references ts demonstrated with an example. The second ts an evidential language--a valuatzon-based language m whwh valuattons are behef functions. The abthty o f ewdenttal languages to perform nonmonotonic reasoning and manage uncertainty is demonstrated with an example. KEYWORDS. valuation-based language, rule-based language, valuation system, k n o w l e d g e - b a s e d system, rule-based system, consistency in k n o w l e d g e bases, caching inferences, truth m a i n t e n a n c e systems, evidential systems, n o n m o n o t o n i c reasoning, m a n a g e m e n t o f uncer- tainty I N T R O D U C T I O N This paper proposes a new language based on " v a l u a t i o n s " as an alternative to rule-based languages for budding knowledge-based systems. This language is inspired by the axlomauc framework for propagation o f probabdmes and b e h e f funcuons (Shenoy and Shafer [1, 2]) and by Rs extension, which includes constraint propagatl,3a and discrete optlrmzataon (Shenoy and Shafer [3, 4]). Since the primary objects in the axiomatic framework are called valuaUons, we refer to this language as being valuation-based, and we call a formal structure created using this language a valuation system. A popular language for building a knowledge-based system is a production or a rule-based language (Brownston et all. [5], Davis and King [6]). While these Address correspondence to P. P. Shenoy, School o f Business, Summerfield Hall, Lawrence, Kansas 66045-2003 lnternauonal Journal of Approxtmate Reasoning 1989, 3 383--41 ! © 1989 Elsevier Sctence Pubhshmg Co , Inc 655 Avenue of the Americas, New York, NY 10010 0888-613X/89/$3 50 Umverstty of Kansas, 383 384 Prakash P Shenoy languages have many attractive features, they also have some serious shortcom- ings. In this paper we will focus on four major shortcomings o f rule-based languages that are not shared by valuauon-based languages. These four shortcormngs are referred to as the problem o f consistency, the problem o f caching, the problem o f nonmonotonic reasoning, and the problem o f managing uncertainty. A special case o f a valuation system is an evidential network (Shenoy and Shafer [1, 2]). The use o f evidenual networks to manage uncertainty is well understood (see, e.g., Shafer et al. [7]). However, the use o f valuaUon systems for representing categorical knowledge, maintaining consistency, and perform- mg intelligent caching and nonmonotonlc reasoning is not widely understood Valuation systems include as a special case b e h e f networks and moral graphs. Belief networks have been proposed by Pearl [8, 9] and moral graphs by Launtzen and Splegelhalter [10] for managing uncertainty using probablhtles (see also Heckerman and Horvltz [11]) The use o f valuation systems m representing and propagating probabilities is descnbed m Shenoy and Shafer [1, 2] and Shafer and Shenoy [12, 13] Valuation languages can also be used to propagate constraints (Seldel [14], Dechter and Pearl [15], Shenoy and Shafer [3]) and to solve discrete optimization problems, both constrained and unconstrained (Bertele and Bnoschi [16], Shenoy and Shafer [4]) Other problems that fit in the framework o f valuauon languages Include solution to systems o f equations (Rose [17]), propagation of Spohman belief functions (Spohn [18], Hunter [19]), retrieval from acychc database schemes (Malvestuto [20], Been et al [21]), and use o f a Kalman filter (Dempster [22], Melnhold and Singpurwala [23]). An outline o f this paper is as follows. In the following secuon we discuss some problems with rule-based languages In the third section we gwe an abstract description o f a valuation-based language, and in the fourth section we describe a specific instance o f a valuation language designed to represent categoncal knowledge We demonstrate, using an example, how such a language can be used to maintain consistency m a knowledge base and how Inferences are cached. In the fifth section we describe an ewdentml language--a valuatmn language that uses b e h e f functions as valuauons. We also briefly describe a truth maintenance system and show the correspondence between concepts m truth maintenance systems and concepts m evidential systems Next, using an example, we show how ewdential languages can be used to reason nonmonotonicaUy and manage uncertmnty. We conclude with a summary and some general comments. SOME PROBLEMS WITH RULE-BASED L A N G U A G E S In this secUon, we look at some of the shortcomings o f pure rule-based languages. In particular, we focus on the problems o f consistency, caching, Valuauon-Based Language for Expert Systems 385 nonmonotomc reasoning, and management o f uncertainty. Since there ~s no universally accepted formal definition o f a rule-based language, we will use the model o f a pure production system given in D a w s and Kang [6] as a representative rule-based language Consistency In large knowledge bases, consistency is an important issue By consistency, we mean the absence o f syntactic contra&ctions. An example o f a syntactic contradiction is a premise A = a and two rules: A = a -~ B = b and A = a --, B = - b . (The symbol --, denotes the truth-functional con&tional.) Rule-based languages lack expressive power to check for contra&ctions. Accordingly, most commercial lmplementaUons o f rule-based languages pro- vide little or no support for checlong for contradlcUons. However, this does not mean that such checking cannot be done outside the formal structure o f rule- based languages. In recent years, there have been several stu&es on e f f l o e n t methods for checlong for contra&ctlons in rule-based languages (see, e . g . , Adams [24], Suwa et al. [25], Nguyen et al. [26], Pearl [27], Touretzky [28], and G m s b e r g [29]). As we shall see, unlike rule-based languages, maintenance o f consistency is an Integral part o f valuation-based languages Caching Regarding caching o f references, typically, backward-chaining, goal-driven rule-based languages do not cache any inferences, whereas forward-chaining production systems cache all Inferences in worlong m e m o r y In either case, caching m rule-based languages is o f little help to the knowledge engineer in understanding the implications o f the knowledge in the knowledge base. As we shall see, valuation languages cache and display certain inferences, and this can be very useful in the knowledge engineering process. Nonmonotonic Reasoning The subject o f nonmonotomc or default reasoning is an important area m artificial intelligence W e often use assumptions or defaults as facts untd we observe something that contradicts the references we have derived. W e then need to retract some assumpuons or defaults to avoid the contradiction. A famous example is that o f Tweety the bird. Most birds fly W e may initially use the rule I f X is a bird then X f l w s as an assumption or a default. Upon learning that Twecty is a bird, we m a y infer that Tweety flies. H o w e v e r , we m a y subsequently learn that Tweety is a penguin and does not fly. At this stage, to keep our knowledge base contra&ction-free, we need to retract the assumption that led to the contra&ction. The construction o f efficient procedures to enable nonmonotonic or default 386 Prakash P. Shenoy reasoning is the subject o f considerable research m artificial intelligence (McCarthy [30], McCarthy and Hayes [31], McDermott and Doyle [32], Moore [33], Reiter [34]). Uncertainty Finally, it is now well known that pure rule-based languages are inadequate both to represent uncertain knowledge and to make references from such knowledge (Shafer [35], Heckerman and Horvltz [36]). F o r example, MYCIN used certainty factors and PROSPECTOR used a pair o f likelihood ratios with each rule to represent uncertainty (Shorthffe and Buchanan [37], Duda et al [38]). However, these systems are brittle. They give the right answers in only the simplest o f cases. One solution to some o f these problems is to couple a truth maintenance system to the knowledge base (Doyle [39], de Kleer [40], Reiter and de Kleer [41]). Truth maintenance systems were dewsed by logicianS In artificial intelligence to reason with incomplete and uncertain information symbohcally without using numerical calcuh such as probability theory or belief functions. Truth maintenance systems are still in the developmental stage and are the subject o f intense research m artificial intelligence. Another solution has been to control the sequence o f inferences so that the correct results are obtained. This approach has been studied, for example, by Laskey and Lehner [42] and by D'AmbrosIo [43]. AN ABSTRACT DESCRIPTION OF A VALUATION-BASED L A N G U A G E This section gives an abstract description o f a valuation-based language. The language consists o f objects, and operators that operate on the objects. The objects are used to represent knowledge The operators are used to make mferences from the knowledge In rule-based languages, the objects are variables and rules and the operator is modus ponens. In valuation-based languages, the objects are called variables and valuations, and the operators are called combination, marginalizatlon, and solution. The level o f abstractness at which this language is described here forces us to omit the computational details o f how precisely the three operators are used to make inferences. This allows us to concentrate on the concepts (For a more formal and less abstract exposition with theorems and proofs, we refer the reader to Shenoy and Shafer [1-4] and Shafer and Shenoy [13].) However, since abstract descriptions can be difficult to comprehend, we describe two specific valuation-based systems in the succeeding sections with concrete examples. Valuataon-Based Language for Expert Systems 387 Variables and Configurations W e use the s y m b o l ~ x f o r the set o f possible values o f a variable X , and w e call ~dTx the f r a m e f o r X . W e wall be c o n c e r n e d with a fimte set 9C o f variables, and w e will a s s u m e that all the variables In 9C have finite frames Given a fimte n o n e m p t y set h o f variables, w e let 'Wh denote the Cartesian p r o d u c t o f '~7x f o r X m h ; ~ h = X { %Vx[X E h }. W e call "~h the f r a m e f o r h. W e will refer to elements o f 'Wh as configurattons o f h. PROJECTION OF CONFIGURATIONS ProJection o f configurations simply means d r o p p i n g extra coordinates; if (w, x , y , z) is a c o n f i g u r a t i o n o f { W, X , Y, Z } , f o r example, then the projection o f (w, x , y , z) to { W, X } is simply (w, x), which is a c o n f i g u r a t i o n o f { W, X } I f g and h are sets o f variables, h c g, and x is a configuration o f g, then we will let x *h denote the projection o f x to h. T h e projection x *h is always a configuration o f h. I f h = g and x is a c o n f i g u r a t i o n o f g, then x *h = x Valuations Given a set h o f variables, there is a set ~ h . T h e elements o f ~ h are called valuattons o f h. W e will let ~ denote the set o f all valuations, that is, ~ = (-J { ~ h [ h c ~E}. Valuations are p r i m m v e s m o u r abstract description and as such require no definition But as w e shall see shortly, they are objects that can be c o m b i n e d , m a r g l n a h z e d , and solved. Intmtively, a valuation o n h represents s o m e k n o w l e d g e about the variables m h Examples o f valuations on h are an array, a function H : ~¢~h ---' ~ + ( ~ + denotes the set o f n o n - n e g a t w e real numbers); a superarray, a function H : 2 v:h -+ ~/+ (2wh denotes the set o f all subsets 0f'Wh); a rule, a function H : %Vh -+ { true, false }, etc PROPER VALUATIONS F o r each h c ~ , there Is a subset (Ph o f ~ h w h o s e elements will be called p r o p e r valuattons on h. Let (P denote I,.) { (Ph I h c ~E }, the set o f all p r o p e r valuations. Intuitively, a p r o p e r valuation represents k n o w l e d g e that is consistent in itself T h e notion o f p r o p e r valuations is important as it enables us to define c o m b m a b l l i t y o f valuaUons, it allows us to define existence o f solutions, and it allows us to constrain the definitions o f combination and marginalization to meaningful operations. E x a m p l e s o f p r o p e r valuaUons are apotenttal, a f u n c u o n P" ~¢7h ~ ~L that is not identically z e r o f o r all configurations; a superpotential, a f u n c n o n m : 2wh ~t+ that is not z e r o f o r all n o n e m p t y subsets o f "~¢h; a satisfiable rule, a function R : 'Wh ~ {true, false} that is not identically false f o r all 388 Prakash P. Shenoy configurataons; etc. Potentials c o r r e s p o n d to unnormalized p r o b a b d l t y distributions, and superpotentials c o r r e s p o n d to u n n o r m a l l z e d basic probability assignment functions. Combination W e assume there is a m a p p i n g ® : ~ × ~ --) ~7, called combination, such that 1. I f G and H are valuations on g, and h, respectively, then G ® H is a valuation o n g LI h . 2. I f either G o r H is not a p r o p e r valuation, then G ® H is not a p r o p e r valuation 3. I f G and H are both p r o p e r v a l u a t i o n s , then G @ H m a y o r m a y not be a p r o p e r valuation I f G @ H is not a p r o p e r valuaUon, then w e shall say that G and H are not combinable. I f G ® H is a p r o p e r valuation, then w e shall say that G and H are combinable and that G ® H is the combmatton oJ G and H. Intuitively, as its name suggests, c o m b i n a t i o n corresponds to a g g r e g a u o n o f k n o w l e d g e I f G and H are p r o p e r valuations on g and h representing k n o w l e d g e about varmbles in g and h , r e s p e c u v e l y , then G ® H represents the a g g r e g a t e d k n o w l e d g e about variables m g U h . F o r potentials, c o m b i n a t i o n c o r r e s p o n d s to polntw~se multiplication, if G and H are potentmls on g and h, respectively, then (G ® H ) ( x ) = G(x~g)H(x~h). F o r basic p r o b a b d l t y assignment f u n c u o n s , combmatxon c o r r e s p o n d s to D e m p - s t e r ' s rule ( D e m p s t e r [44, 45]). F o r rules, ff G and H are rules o n g and h , respectively, then G ® H is a rule on g t3 h such that (G ® H ) ( x ) = true lff G(x ~g) = true and H ( X ~h) = true. Marginalization W e assume that f o r each h c_ 9C, there IS a m a p p i n g ~h: I,.J { ~ g l g ~- h} %9h, called marginalization to h, such that 1. I f G is a valuation o n g and h c g , then G sh is a valuation on h 2. I f G is a p r o p e r valuation, then G ~h is a p r o p e r valuation. 3. I f G IS not a p r o p e r valuation, then G ~h i s not a p r o p e r valuation W e will call G ~h the marginal o f G f o r h. Intuitively, marginalization c o r r e s p o n d s to crystalhzatlon o f knowledge. I f G is a valuation o n g representing s o m e k n o w l e d g e about variables in g, and h _c g, then G ~h represents the k n o w l e d g e about variables in h implied b y G i f w e disregard varmbles in g - h . In the case o f potentials, m a r g m a l i z a t l o n f r o m g to h is summation o v e r the configurations o f g - h . I n the b e h e f - f u n c t l o n case, margmallzation Is explained in the section " A n Evidential L a n g u a g e . . . " F o r rules, i f G is a rule o n g, then ValuaUon-Based Language for Expert Systems 389 G ~h is a rule on h such that GSh(x) = true fff there is a configuration y o f g - h such that G(x, y ) = true. Solution We assume that for each g c_ 9C, there is a mapping if. ~ g ~ 2 ~ g called solutton such that 1. I f G is a proper valuation on g, then i f ( G ) is a nonempty subset o f ~Vg. 2. I f G is a valuauon on g that is not proper, then i f ( G ) = ~ . The configurations m i f ( G ) are called soluttons f o r G. Intmtlvely, the solution operator maps knowledge from the space o f valuations to the space o f configurations. W e encode knowledge as valuaUons so that we can aggregate and crystalhze it. However, we need to decode the result The solution operator simply serves as a decoding mechanism In the case o f probablhUes, solutions may correspond to configurauons w~th the highest probabihty or simply configurations with posltwe probabdmes. For b e h e f functions, solutions may correspond to configurauons with the highest plauslbihty or simply configurations with posltwe plausibihtles F o r rules, solutions may correspond to configurations whose value ~s true Propagation of Valuations. A valuation-based language (VL) makes references by 1. Combining all proper valuations m the system (the resulting valuation, if proper, is called the jomt valuation), 2. Computing the marginal o f the joint valuation for each varxable in the system, 3. Computing the set o f all solutions for the margmals o f the joint valuation for each variable; and 4. Computing the set o f all solutions for the joint valuation The above is only a conceptual descripUon o f the actions o f a valuaUon-based language. It is not an algorithm I f there are n variables in the system, and each variable has two configurations m ~ts frame, then there are 2 n configurations o f all variables. Hence, it will not be feasible to compute the joint valuation when there are a large number o f variables. The V L does not actually compute the joint valuation. It computes the marglnals o f the joint valuation without exphcltly computing the joint valuation, and it does this using only local computations. An algorithm for computing exact marglnals and solutions is described in detail in Shenoy and Shafer [1-4]. An algorithm for computmg approximate margmals is described m Pearl [46], and an algorithm for computing an approximate soluuon to the joint valuauon ~s described in Klrkpatnck et al. [47] and G e m a n and G e m a n [48]. 390 Prakash P Shenoy Valuation System A valuation system (VS) consists o f a fimte set o f variables ~E, a finite frame 'Wx for each variable X in 9C, and a finite collection o f valuauons { V,},eM where each valuation V, is on some subset o f ~E. A valuatton network is a graph whose vertices represent either variables or valuations. I f valuation 1I, is on a subset h o f vertices, then this is represented in the valuation network by Including an edge between the vertex corresponding to V, and all variable vertices Xj such that Xj E h. The valuation network serves as a graphical representation of a valuation system and can be used as a user interface The valuation network is also used by the VL to propagate the valuations The algorithm for computing exact margmals and solution requires that the valuation network be a tree If the valuation network is not a tree, then this algorithm embeds it in a tree by clustering variables Such a tree, called a Markov tree, is then used to compute the marglnals and solutions (Shenoy and Sharer [1-4]). The simulation algorithms for computing marglnals and soluuons use the valuation network directly Capabifities o f a Valuation-Based Language A VL has the following capablllUes: 1. A VS can be extended by adding new vartables and adding new proper valuations. 2. A VS can also be reduced by removing variables and valuations. 3. Each time the VL recewes a new proper valuation, It checks whether or not it is combinable with the proper valuations already present in the system. 4. I f the new proper valuation is combinable with the valuauons already present m the system, then the VL accepts the new valuation. I f the new proper valuation is not combinable, then the VL rejects it and informs the user o f its acUon. 5. Each time the VL accepts a proper valuanon, it finds the marginal o f the joint valuaUon (the valuation obtained by combining all proper valuations in the system) for each variable in the system. This IS accomphshed using local computaUons ff an efficient Markov tree can be found for the valuation network (Shenoy and Shafer [1, 2]) or by stochastm simulation otherwise (Pearl [46]) 6. The VL also computes for each variable the set o f all solutaons for the marginal o f the joint valuation for that variable. Once we have the marginal o f the joint valuauon for a variable, computing the set o f all solutions is simply done by exhaustive enumeration o f the frame for that variable. 7. I f necessary, the VL can compute a configuration o f all variables that is a Valuauon-Based Language for Expert Systems 391 solution for the joint valuation. This can be done using an exact algorithm ff an efficient M a r k o v tree can be found for the valuataon network (Shenoy and Shafer [3, 4]) or by stochastic relaxaUon and annealing (Kirkpatnck et al [47], G e m a n and G e m a n [48]) A V A L U A T I O N L A N G U A G E FOR CATEGORICAL K N O W LED G E In this section, we describe an instance o f a valuation-based language designed to represent categorical knowledge--the lond o f knowledge tradiUon- ally represented b y rules in rule-based systems. Next, we show by means o f a small example how consistency is maintained m the knowledge base and how references are cached Our e x p o s m o n here is informal. A formal treatment (with theorems and proofs) o f the valuation language described m this section is given m Shenoy and Shafer [3]. Suppose we are interested m representing categorical knowledge m a valuauon system. Let us describe what valuations are and what the combination marginalization, and soluUon operations are for such systems VALUATIONS A v a l u a t t o n o n h is a function H : 'Wh ---' {t, f } , where t means true and f means false Thus a rule that relates the values o f variables m set h is represented as a valuation on h F o r example, consider the rule I f A = a t h e n B = b that relates two variables A and B whose frames are, respecuvely, "~7,4 = {a, - a } , and ~ B = {b, - b}. Then the rule can be represented by the valuation V o n {A, B} defined as follows: V(a, b) = t, V(a, - b ) = f , V ( - a , b) = t, V ( - a , - b ) = t. Consider the valuation U on h such that U ( x ) = t for all x E ~¢h. Obwously, such a valuation tells us nothing about the variables m h. W e call such a valuation the v a c u o u s v a l u a t i o n o n h. PROPER VALUATIONS Suppose H is a valuaUon on h. W e shall say that H is a p r o p e r v a l u a t i o n i f there exists a configuration x o f h such that H ( x ) = t. Thus a proper valuation cannot be identically equal to f for all configurations. COMBINATION Suppose G and H are valuauons on g and h, respectwely The valuauon G @ H on g I.) h is defined as follows: ( G ® H ) ( x ) = I ) for all x E q~v~suh. If G ( X ~g) = t and H ( X i h ) = t otherwise 392 Prakash P Shenoy MARGINALIZATION Suppose G is a valuation on g, and suppose h _c g. Then the marginal o f G f o r h, G *h, is defined as follows: for all x E 'Wh. If there IS a y E e~g_ h such that G ( x , y ) = t otherwise SOLUTION Suppose G is a valuaUon on g. The solution f o r G, denoted by ~b(G), is a subset o f %Vg such that y E if(G) if and only if G ( y ) = t. The combination, margmahzatlon, and solution operations are used by the VL to make Inferences from the knowledge Suppose a knowledge base Is built incrementally by adding valuations one at a time Consistency in the knowledge base is maintained by the V L by checking whether the added valuation is proper and combinable with the proper valuations already present in the system. Thus combinablhty o f valuations corresponds to consistency in the knowledge base (Shenoy and Shafer [3]) As valuations are added to the knowledge base, the system propagates all valuations and computes the marginal o f the joint valuation for each variable and the solutions for each o f these marglnals. More precisely, suppose { Rh [ h E 3E } is a collection o f proper combinable valuations in the system. The valuation @{Rh [h E 3~2 } is called the j o m t valuatton. The valuation system computes ( ~ { R h [ h E 3(~})~{x, } for each variable X, and also computes ak((®{Rhih E 3C }),{x,}). In doing so, the VS acts as a cache. At all times, the VS indicates the relevant inferences of the knowledge in the knowledge base AN EXAMPLE The following example is adapted from Ethenngton [49]. The knowledge base consists o f four rules as follows R u l e 1. Gullible citizens are citizens. R u l e 2. Elected crooks are crooks. R u l e 3. Cmzens &slike crooks. R u l e 4. Gullible citizens do not dislike elected crooks First we observe that Fred is a gullible citizen Next we observe that Dick is an elected crook. We would hke to consult our knowledge base to see if Fred dislikes Dick or not. One representaUon o f this knowledge base is as follows Let C = c, G = g, K = k, E = e, and D = d be five variables and their respective configurations representing X is a citizen, X is a gullible Otlzen, Y is a crook, Y is an elected crook, and X dislikes Y, respectively Suppose all five o f these variables are binary variables Valuauon-Based Language for Expert Systems T a b l e 1. T h e V a l u a t i o n s C o r r e s p o n d i n g to the F o u r R u l e s 393 %V{c,a} RI 'W{x,e} R2 c g t k e t c - g t k - e t - c g f - k e f - c - g t - k - e t "~7 { C,K,D } R 3 't~7 { G,E,D } R 4 c k d t g e d f c k - d f g e - d t c - k d t g - e d t c - k - d t g - e - d t - c k d t - g e d t - c k - d t - g e - d t - c - k d t - g - e d t - c - k - d t - g - e - d t R u l e s 1, 2, 3, a n d 4 a r e r e p r e s e n t e d b y p r o p e r v a l u a a o n s on { C , G } , { K , E } , { C , K , D } , a n d { G , E , D } , r e s p e c t i v e l y , as s h o w n in T a b l e 1 S u p p o s e t h e s e v a r i a b l e s a n d valuaUons a r e e n t e r e d m the s y s t e m A n e t w o r k r e p r e s e n t a t i o n o f the s y s t e m is s h o w n in F i g u r e 1 In t h a t f i g u r e , v a r i a b l e s a r e r e p r e s e n t e d b y c i r c l e s a n d v a l u a t i o n s a r e r e p r e s e n t e d b y s q u a r e s . F o r e a c h v a n a b l e , the set o f all s o l u u o n s f o r the m a r g i n a l o f the j o i n t v a l u a t i o n f o r that v a r i a b l e is i n d i c a t e d i n s i d e the v a r i a b l e v e r t e x . A s can b e seen f r o m F i g u r e 1, for e a c h v a r i a b l e the m a r g i n a l o f the j o i n t v a l u a t i o n for t h a t v a r i a b l e is the v a c u o u s valuaUon N o w , s u p p o s e w e e n t e r the o b s e r v a t i o n that F r e d ts a g u l h b l e c m z e n . T h i s is r e p r e s e n t e d in the s y s t e m a s a p r o p e r v a l u a t i o n F1 o n { G } as f o l l o w s F l ( g ) = t, F 1 ( - g ) = f . T h e s y s t e m a c c e p t s this p r o p e r v a l u a t i o n , a n d a f t e r p r o p a g a U o n it d i s p l a y s the r e s u l t s as s h o w n in Fxgure 2 N o t e that the s y s t e m p r o p e r l y c o n c l u d e s t h a t F r e d is a c i t i z e n . H o w e v e r , the s y s t e m a l s o c o n c l u d e s that Y is not an e l e c t e d c r o o k ! T h i s is t h e first hint w e h a v e that s o m e t h i n g is w r o n g w i t h out k n o w l e d g e b a s e . T h e s y s t e m has c o n c l u d e d s o m e t h i n g a b o u t Y w i t h o u t b e i n g t o l d it e x p l i c i t l y , a n d this is not an r e f e r e n c e w e e x p e c t f r o m the k n o w l e d g e b a s e . T h e r e a s o n f o r the r e f e r e n c e Y is n o t a n e l e c t e d c r o o k is the c o n t r a d i c t o r y n a t u r e o f r u l e s 3 a n d 4. F i n a l l y , w e e n t e r the o b s e r v a U o n that D i c k is an e l e c t e d c r o o k . T h i s 394 Prakash P Shenoy @ Figure 1. The valuation network with five variables and four rules Table 2. The Valuation Corresponding to Rule 5 ~dT{G,x,D~ R5 g k d t g k - d t g - k d t g - k - d t - g k d t - g k - d f - g - k d t - g - k - d t Valuation-Based Language for Expert Systems 395 Figure 2. The valuation network after valuation Fl Is included observation is represented as a proper valuation F2 on {E} as follows: F z ( e ) = t, F 2 ( - e ) = f . This time the system refuses to accept the valuation because the system detects that t h e j o i n t valuation R I @ R2 ® R3 ® R4 ® Fl ® F2 Is not a proper valuation. This signals that the knowledge in the system is inconsistent. Suppose we r e m o v e rule 3 f r o m the system and substitute instead rule 5 as follows: R u l e 5 Nongullible citizens dishke crooks. Rule 5 is represented in the system as the valuation R5 on {G, K , D } as shown m Table 2 The valuation system accepts valuation R5 wxth the results shown m Figure 3. Note that the system now concludes nothing about Y. Finally we enter valuation/72 in the system. This lame the system accepts the valuation wxth the results shown in Figure 4 Thus we conclude that Fred does not dmhke Dick. W e have not described the exact process by which the valuation language arrives at the results displayed m Figures 1-4 A computationally efficient procedure m sparse networks that uses only local computation is described m Shenoy and Shafer [3] Prakash P Shenoy Figure 3. added Ciuzen(X) ~ = { c } Citizen(X) V = {g} v = {,:t, ,-,d} The valuation network after valuation R3 is removed and valuation R5 1s ~ Crook(Y) huzen(X) ~ R5 v = { c} I V={ k} 396 Cluzen(X) Figure 4. The valuation network after valuataon F2 is included Valuation-Based Language for Expert Systems 397 A N EVIDENTIAL L A N G U A G E FOR U N C E R T A I N A N D N O N M O N O T O N I C R E A S O N I N G In this section, we describe another valuation language called an evidential language The valuations m this language are belief functions Propagation o f belief functions has been studied by Shafer and Logan [50], Shenoy and Shafer [1, 2, 51], Shenoy et al [52], Kong [53, 54], Dempster and Kong [55], Shafer et al. [56], Mellouh [57], Shafer and Shenoy [13], Dempster [22], and Almond [58] Zarley [59] describes an implementation o f an evidential system on a Symbohcs workstation (see also Zarley et al. [60]) Yen-Teh Hsia has implemented an evidential system called A U D I T O R ' S A S S I S T A N T on a Macintosh microcomputer. Shafer et al [7] describe an application o f A U D I T O R ' S A S S I S T A N T for assisting in audit decisions The use o f probabtlmes or belief functions to p e r f o r m nonmonotonlc reasoning is not new. Such an approach has been suggested, for example, by Baldwin [61], Ginsberg [62], and Rich [63]. The essence o f these approaches IS to relax the binary constraint o f Boolean logic and allow truth values to be measured by a number between 0 and 1 Our approach IS different We do not tack on probabilities or belief functions to logic. Instead, we show that pure behef-function reasoning is mherently nonmonotonic. A similar approach is taken by G r o s o f [64], who discusses how probabfliStlC reasoning is nonmono- tonic. In this section, we will first briefly describe evidential systems. Next, we sketch the basic definitions in a truth maintenance system and describe the correspondence between concepts in an truth maintenance system and concepts in a evidential system Fmally, we study a small example in nonmonotomc reasoning and demonstrate how evidential systems handle such problems This example also serves to illustrate the management o f uncertainty in evidential systems An Evidential System In evidential systems (ES), proper valuations correspond to superpotentlals, which are unnormahzed basic probabihty assignment functions. First we will briefly describe the basics o f the theory o f belief functions (Shafer [65]) Next, we define superpotentials and combination, marginahzatlon, and solution for superpotentlals Suppose XVh is the frame for a subset h o f variables A basic probablhty assignment functton (bpa function) for h is a non-negative, real-valued function m on the set o f all subsets o f 'Wh such that 1. m(fZS) = 0 2. X { m ( , 0 l a c_ ~ h } = 1 Intuitively, r e ( a ) represents the degree o f belief assigned exactly to ~x (the proposition that the true configuration o f h is in the set t~) and to nothing smaller. 398 Prakash P Shenoy A bpa function is the belief f u n c u o n equivalent o f a probability mass assignment function m probability theory. W h e r e a s a probability mass function is restricted to assigning probability masses only to singleton configurations o f variables, a bpa function is allowed to assign probability masses to sets o f configurations without assigning a n y mass to the individual configurations contained m the sets. F o r example, I f w e have absolutely no k n o w l e d g e about the true value o f a variable, w e can represent this situation b y a bpa function as follows" m(cd2h) = 1, m(a) = 0 for all other a E 2 w h Such a function is called a vacuous bpafunction. N o t e that m Bayesian t h e o r y the only w a y to express total ignorance is to assign a mass o f I / n to each value, where n is the total n u m b e r o f possible values Thus, m Bayesian theory, w e are unable to distinguish between equally likely configurations and total ignorance. The theory o f b e h e f functions offers richer semantics. Associated with a bpa function are two related functions called b e l i e f and plausibility. A belieffunctton is a function Bel: 2Wh ~ [0, 1] such that S e l ( a ) = Y ~ { m ( ~ ) l ~ _ a } W h e r e a s m(a) represented the b e h e f assigned exactly to a , B e l ( a ) represents the total b e h e f a s s l g n e d to a . N o t e that B e l ( ~ ) = 0 and Bel('Wh) = 1 f o r any belief function. F o r the v a c u o u s bpa function m , the c o r r e s p o n d i n g b e h e f function Bel is given b y Bel(%qh) = 1, B e l ( a ) = 0 for all other a E 2wh A plaustbilityfunctton is a function PI" 2wh --' [0, 1] such that P l ( a ) = Y ~ { m ( ~ ) l ~ t3 a ~ : ~ } P l ( a ) represents the total d e g r e e o f b e h e f that could be assigned to a . N o t e that P l ( a ) = 1 - Bel( - a ) , where - a represents the c o m p l e m e n t o f a in 'Wh; - a = 'Wh -- a . A l s o note that P l ( a ) _ B e l ( a ) F o r the v a c u o u s bpa function, the c o r r e s p o n d i n g plausibdlty function Is P I ( ~ ) = 0 , P l ( a ) = 1 for all other a E 2 ~ h . I f a bpa function m is also a p r o b a b d i t y mass function 0 . e . , all the probability masses are assigned only to singleton subsets), then Bel(tr) = P l ( a ) = X { m ( { x } ) I x E a } = p r o b a b d l t y o f proposition a SUPERPOTENTIALS Suppose h is a subset o f variables. A superpotentialfor h is a non-negative, real-valued function o n the set o f all subsets o f 'Wh such that the values o f n o n e m p t y subsets are not all zero. Given a superpotential H o n h , Valuatmn-Based Language for Expert Systems 399 we can construct a bpa function H ' for h f r o m H as follows: H ' ( ~ ) = O , H , ( a ) = H ( a ) / y , { H ( ~ ) l ~ c_ 'Wh, ~:/:~} Thus superpotentials can be thought o f as unnormalized bpa functmns SuperpotentlalS correspond to the notion o f proper valuatmns in the general framework. PROJECTION AND EXTENSION OF SUBSETS Before we can define combinatmn and marginahzation for superpotenuals, we need the concepts o f projection and extension for subsets o f configuratmns. I f g and h are sets o f variables, h c g, and $ is a nonempty subset o f 'Wg, then the projection o f ~ to h, denoted by ~*h, IS the subset o f 'Wn given by ~,n = { x * h l x F o r example, ff a is subset o f ~dT{ w,x,r,z}, then the marginal o f a to {X, Y} consists o f the elements o f %V{x, r} that can be obtained by projecting elements o f ,x to %V{x.r}. By extensmn o f a subset o f a frame to a subset o f a larger frame, we mean a cylinder set extensmn. I f g and h are sets o f varmbles, h _c g, h :# g, and ~ is a subset o f ' W h , then the extenston o f ~ to g is J~ x ~d?g_h. IfJ~ is a subset o f %Vn, then the extension o f ~ to h is defined to be ~ . W e will let ~)g denote the extensmn o f ~ to g, For example, ff a is a subset o f 'W{ w.x}, then the vacuous extension o f a to { W, X , Y) Z} IS d, X ¢~{y,z}. COMBINATION For superpotentmls, comblnatmn ~s called D e m p s t e r ' s rule (Dempster [44, 45]) Consider two superpotentials G and H on g and h, respectwely. I f X { G ( a ) n ( ~ ) l ( a *(guh)) N (~t(gUh)):#~} =#0 (1) then their combination, denoted by G @ H , is the superpotential on g U h given by (G ~ H)(c)=Y~{G(,~)H(g)I(a ~uh)) n ( ~ ( * u h ) ) = c } (2) for all c c_ %Vguh. I f ~ { G ( a ) H ( ~ ) l ( a ~(gUh)) n (~ r(guh)) :# ~ } = 0, then we say that G and H are not combmable. Intumvely, i f the bodies o f evidence on which G and H are based are independent, then G @ H is supposed to represent the result o f poohng these two bodies o f evidence. Note that condition (1) ensures that G @ H defined m (2) is a superpotentml. I f c o n d m o n (1) does not hold, this means that the two bodies o f evidence corresponding to G and H contradmt each other completely and it is not possible to combine such evidence MARGINALIZATION Suppose G ~s a superpotential for g, and suppose h _c g. 400 Prakash P Shenoy T h e n the marginal o f G f o r h is the s u p e r p o t e n t i a l G ~h for h d e f i n e d as f o l l o w s : G~h(a.)=~,{G(~)l ~ c_ "~Tg such that ~*h=a.} f o r all subsets ,x o f ~,Vh. SOLUTION T h e r e a r e s e v e r a l d e f i n i t i o n s o f s o l u t i o n p o s s i b l e f o r e v i d e n t i a l s y s t e m s . F o r n o n m o n o t o n i c r e a s o n i n g , w e w i l l d e f i n e a s o l u t i o n for m to be a c o n f i g u r a t i o n w h o s e p l a u s i b i l i t y IS p o s i t i v e . F o r m a l l y , s u p p o s e m is a b p a on h. S u p p o s e Pl is the p l a u s i b i l i t y f u n c t i o n on h c o r r e s p o n d i n g to m T h e n w e s a y that x E 'Wh is a solutton f o r m f f P l ( { x } ) > 0. A Truth Maintenance System A s s u m e a p r o p o s m o n a l l a n g u a g e c o n s i s t i n g o f p r o p o s i t i o n a l s y m b o l s , the l o g i c a l c o n n e c t i v e s A, V, - - , ~ , ~ , f o r m u l a s , a n d the u s u a l s t a n d a r d e n t a i l m e n t r e l a t i o n = " I f S is a set o f f o r m u l a s and w is a f o r m u l a , t h e n S = w i f e v e r y a s s i g n m e n t o f truth v a l u e s to the p r o p o s i t i o n a l s y m b o l s o f the l a n g u a g e that m a k e s e a c h f o r m u l a o f S t r u e a l s o m a k e s w true. A hteral is a p r o p o s i t i o n a l s y m b o l o r the n e g a t i o n o f a p r o p o s i t i o n a l s y m b o l A clause is a finite d i s j u n c t i o n o f h t e r a l s w i t h no h t e r a l s r e p e a t e d w h o s e truth v a l u e is t r u e A premtse is a l i t e r a l w h o s e truth v a l u e ~s t r u e . A categortcal justification is a c o n d i t i o n a l w h o s e truth v a l u e is true. N o t e that a c a t e g o r i c a l j u s t i f i c a t i o n c a n b e r e p r e s e n t e d as a c l a u s e F o r e x a m p l e , the c o n d i t i o n a l A = a --, B = b c a n b e r e p r e s e n t e d as a c l a u s e as f o l l o w s : - ( A = a) v ( B = b). A n assumption is a l i t e r a l w h o s e truth v a l u e is a s s u m e d to b e t r u e in the a b s e n c e o f a c o n t r a d i c t i o n A noncategorwal justzfication is a c o n d i t i o n a l w h o s e truth v a l u e is a s s u m e d to b e t r u e In the a b s e n c e o f a c o n t r a d i c t i o n . A nogood is a c l a u s e w h o s e truth v a l u e is false. A knowledge base is a c o l l e c t i o n o f j u s t i f i c a t i o n s ( r u l e s ) , p r e m i s e s ( o b s e r v a t i o n s ) , and a s s u m p t i o n s ( u n c e r t a i n j u d g m e n t s ) . J u s t i f i c a t i o n s m a y b e c a t e g o r i c a l o r n o n c a t e g o r i c a l . C a t e g o r i c a l j u s t i f i c a t i o n s m a y d e s c r i b e l o g i c a l r e l a t i o n s b e t w e e n p r o p o s i t i o n a l s y m b o l s N o n - c a t e g o r i c a l j u s t i f i c a t i o n s m a y d e s c r i b e facts that a r e u s u a l l y b u t not a l w a y s t r u e T h e f u n c t i o n s o f a truth m a i n t e n a n c e s y s t e m ( T M S ) a r e as f o l l o w s : 1. T h e u s e o f n o n c a t e g o r i c a l j u s t i f i c a t i o n s and a s s u m p t i o n s o r d e f a u l t s is p e r m i t t e d . 2. I n the a b s e n c e o f a c o n t r a d i c t i o n , n o n c a t e g o r l c a l j u s t i f i c a t i o n s a n d a s s u m p t i o n s a r e a s s u m e d to b e t r u e 3. I f t h e r e is a c o n t r a d i c t i o n In the k n o w l e d g e b a s e , then s o m e n o n c a t e g o n c a l j u s t i f i c a t i o n s o r a s s u m p t i o n s o r b o t h n e e d to b e r e t r a c t e d so that c o n s i s t e n c y IS r e s t o r e d W h e n an a s s u m p t i o n o r a n o n c a t e g o r l c a l j u s t i f i c a - t i o n is r e t r a c t e d , all I n f e r e n c e s m a d e u s i n g t h e s e a s s u m p t i o n s and n o n c a t e g o r i c a l j u s t i f i c a t i o n s m u s t a l s o b e r e t r a c t e d . Valuation-Based Language for Expert Systems 401 4. A l l i n f e r e n c e s t h a t a r e c o n s i s t e n t w i t h the k n o w l e d g e in the k n o w l e d g e b a s e s h o u l d b e d i s p l a y e d to the u s e r so t h a t t h e u s e r is a w a r e o f the i m p l i c a t i o n s o f t h e k n o w l e d g e . USING AN EVIDENTIAL L A N G U A G E AS A TMS W e w i l l n o w o u t h n e a c o r r e s p o n d e n c e b e t w e e n t h e c o n c e p t s in a T M S a n d c o n c e p t s in an e v i d e n t i a l s y s t e m (ES). A h t e r a l in a T M S is r e p r e s e n t e d In an E S b y a v a r i a b l e and one o f its v a l u e s . T h u s X = x is an E S r e p r e s e n t a t i o n o f the h t e r a l x w h e r e x b e l o n g s to ' W x , the set o f p o s s i b l e v a l u e s o f v a r i a b l e X . F o r e x a m p l e , s u p p o s e the p r o p o s m o n T W E E T Y IS A B I R D is r e p r e s e n t e d m a T M S as a h t e r a l . In the E S , this c o u l d b e r e p r e s e n t e d b y a v a r i a b l e B I R D w i t h t w o p o s s i b l e v a l u e s y e s a n d n o . T h e n the l i t e r a l T W E E T Y I S A B I R D c o r r e s p o n d s to B I R D - y e s in an E S A p r e m i s e is a h t e r a l w h o s e t r u t h v a l u e is t r u e I n an ES, a p r e m i s e is r e p r e s e n t e d as a c a t e g o r i c a l b e h e f f u n c t i o n . ] : o r e x a m p l e , the p r e m i s e X -- x IS r e p r e s e n t e d b y a b e l i e f f u n c t i o n on XVx g i v e n b y m ( { x } ) = 1. A n a s s u m p t i o n in a T M S is a l i t e r a l w h o s e t r u t h v a l u e is set to t r u e in the a b s e n c e o f a c o n t r a d i c t i o n m the k n o w l e d g e b a s e . I n the E S , an a s s u m p t i o n X = x is r e p r e s e n t e d b y a n o n c a t e g o r i c a l b e l i e f f u n c t i o n Bel (with b a s i c p r o b a b i l i t y a s s i g n m e n t m ) o n %Vx s u c h that m ( { x } ) = p a n d m ( ' W x ) = 1 - p w h e r e 0 < p < 1. T h e a c t u a l v a l u e o f p w i l l d e p e n d on the p a r t i c u l a r a s s u m p t i o n , p c a n b e i n t e r p r e t e d to b e the p r i o r d e g r e e o f b e l i e f in the a s s u m p t i o n A j u s U f i c a t i o n is a c o n & t i o n a l , x l A X2 A " • • A Xn--* y w h e r e x l , x2, • • ", xn, y a r e h t e r a l s . I n an E S , a c a t e g o r i c a l j u s t i f i c a t i o n x~ A x2 A • • • A xn ~ y is r e p r e s e n t e d as a c a t e g o r i c a l b e l i e f funcUon o n the f r a m e %Vh, w h e r e h = { X i , X2, • • ", A n , Y}. F o r e x a m p l e , c o n s i d e r t w o v a r i a b l e s X a n d Y w i t h f r a m e s ~d~x = {x, - x } a n d ~*Vr = { y , - y } . T h e n the c a t e g o r i c a l j u s t i f i c a t i o n x ~ y is r e p r e s e n t e d m the E S as a c a t e g o r i c a l b e l i e f f u n c t i o n on % V ( x , y ) g w e n b y m ( { ( x , y ) , ( - x , y ) , ( - x , - y ) } ) = 1 N o n c a t e g o r l c a l j u s t i f i c a t i o n s a r e r e p r e s e n t e d m the E S as n o n c a t e g o n c a l b e l i e f f u n c t i o n s T h e r e a r e s e v e r a l w a y s in w h i c h this can b e d o n e . T h e m o s t a p p r o p r i a t e w a y wall d e p e n d o n the n a t u r e o f the p a r t i c u l a r j u s U f i c a t l o n . T h e first t y p e o f b e h e f - f u n c t i o n r e p r e s e n t a t i o n o f a n o n c a t e g o r i c a l j u s t i f i c a - t i o n is c a l l e d e x c e p t i o n a l • T h e e x c e p t i o n a l r e p r e s e n t a t i o n o f a n o n c a t e g o n c a l j u s t a f i c a t l o n is i m p l i e d b y M c C a r t h y ' s [66] f o r m u l a t i o n F o r e x a m p l e , c o n s i d e r the n o n c a t e g o n c a l j u s t i f i c a t i o n M O S T B I R D S F L Y T h i s can b e r e p r e s e n t e d in a 402 Prakash P Shenoy TMS by a categorical jusUficatlon and an assumption as follows: BIRD = y e s A E X C E P T I O N A L _ B I R D = n o ~ F L Y = y e s Assume E X C E P T I O N A L _ B I R D = n o Here E X C E P T I O N A L _ B I R D = n o ~s a hteral that captures all the conditions under which birds fly. Let B = b, E = - e , a n d F = f d e n o t e the ES representation o f the literals B I R D = y e s , E X C E P T I O N A L _ B I R D = n o , and F L Y = y e s . Then the justificaUon M O S T BIRDS F L Y can be represented in an ES by two independent basic probability assignment functions, m~ on 'W~B,~,F) and m2 on "~,V E as follows: m l ( e ~ { B , E , F } - - {(b, - e, - f ) } ) = 1 m2({ - e } ) = p , m2(~gTe) = 1 - - p where 0 < p < 1. Note that ml (~) m2 Is a basic probablhty assignment on ~/B.E,F) given by m l G m2({(b, - e , f ) , ( - b , - e , f ) , ( - b , - e , -f)})=p m l G mZ(~C{B.E,F} -- {(b, - e , - f ) } ) = 1 - p m~ • m2 xs then the exceptional representation m an ES o f the jusUficatlon MOST BIRDS F L Y The second type o f behef-function representation o f a noncategorical justification is called a s s o c t a t t o n a l . Consider again the justificaUon M O S T BIRDS F L Y W e can interpret this to mean that birds are associated with flying with a certain degree o f behef. This associaUon m a y just go one way, that is, we may not necessarily assocmte all flying objects wRh birds. Interpreted in this way, we can represent this justification by a basic probability assignment function m3 on 'W{B,F} as follows m3({(b,f), ( - b , f ) , ( - b , -f)})=p, m 3 ( ~ C C { B , F ) ) = l - - p w h e r e 0 < p < 1. Obwously, excepUonal representaUons o f noncategorlcal justifications have greater expressive p o w e r than assocmtional representations. In the bird example, i f the basic probability assignment functaon m l (~) m2 is marginalized by deleting the E varmble, then we obtain precisely the assocmtlonal representation m3, that is, (ml @ m2) ~{B'F} = m3. H o w e v e r , this expressive power comes at a computataonal cost since m o r e variables are required in the exceptional representation than m the assocmtional representation. Consider a knowledge base represented b y a collection o f bpa functions {m, [i = 1, • • -, n} representing premises, rules, and assumptions. Suppose my is an assumption X = x. W e shall say that the assumption m , xs r e t r a c t e d b y t h e k n o w l e d g e b a s e { m ~ l i = 1, . . . , n} i f p l q x ) ( { x } ) = 0 where P l q x} is the plausibility function corresponding to ( ~ {m, [i = 1, - - . , n } ) q x } . W e shall say Valuation-Based Language for Expert Systems 403 that the assumption m~ is c o n f i r m e d b y the k n o w l e d g e base { m , l t = 1, . . . , n} l f m ~ X ) ( { x } ) = 1 where m = ( O { m , lt = 1, - - ' , n}). An Example Consider the following knowledge base: R u l e 1. Most Repubhcans (at least 80%) are not pacifists. R u l e 2. Most Quakers (at least 90%) are paclficists First we observe that Nlxon is a Republican. Then we observe that Nixon is also a Quaker. We would like to consult our knowledge base to find out whether Nixon is a pacifist or not. Next we will add the premise that Nlxon is not a pacifist and see how the evidential system reconciles this premise with rule 2. One representaUon o f this knowledge base is as follows. Let R = r, Q = q, P = p be three variables and their respective configurauons representmg the propositions X Is a Republican, X is a Quaker, and X is a pacifist, respectively. Furthermore, let ER = er and EQ = eq be two more variables and their respectwe configurations representing the proposmons X Is an excepUonal Repubhcan and X is an exceptional Quaker, respectively We will represent rule 1 with categorical rule 1 and assumption 1 as follows. CATEGORICAL RULE 1 I f X lS a Repubhcan and X is not an excepaonal Repubhcan, then X is not a paclficlst ASSUMPTION 1. X IS not an excepUonal Republican The bpa function representaUon o f categorical rule 1 is as follows mt(5~2{e, E R , P } - {(r, --er, p ) } ) = 1 The bpa function representation o f assumption 1 is as follows: m2({ - er}) = 0.8, m2(~CCeR) = 0.2 We will represent rule 2 with categorical rule 2 and assumption 2 as follows: CATEGORICAL RUL~ 2. I f X Is a Quaker and X is not an exceptional Quaker, then X is a pacifist ASSUMZrION 2 X IS not an exceptional Quaker. The bpa funcuon representation o f categorical rule 2 is as follows. m3('W{O, EO, p} -- {(q, -- eq, - - p ) } ) = 1 The bpa function representation o f assumption 2 is as follows: m4({ - e q } ) = 0 . 9 , m4('vdTEO) = 0.1 I f we enter these four bpa functions in the ewdential system, the resultmg 404 Prakash P Shenoy Figure 5. The evidential network with two categorical rules and two assumptmns evidential network is as shown in Figure 5. As before, variable vertices are shown as circles and valuation vertmes are shown as squares. In addition to displaying the set o f all soluuons for the marginal o f the joint valuation, the marginal bpa function is also displayed. I f {x, - x } is the frame for variable X , then the marginal o f the joint valuation for X is shown as a vector ( m ~{x)({x}), ( m t { X ) ( { - x } ) , (mqXI({x, - x } ) ) , where m = (~ {m, ll = 1, . . - , n}. Suppose we now enter the premise that Nlxon is a Republican. This is represented as a bpa function as follows m s ( { r } ) = 1 The evidential system accepts this bpa function with the results as shown m Figure 6. Note that the b e h e f in the proposition that Ntxon is not a pacifist has increased f r o m 0 to 0.8 and the b r i e f in the p r o p o s m o n Nlxon is not a Quaker has increased f r o m 0 to 0 72. Suppose we now enter the premise that Nlxon is a Quaker. This is represented by a bpa functmn as follows m6({ q})= 1 The ewdenUal system accepts this bpa functmn with the results shown m Figure 7. Note that as per the ES, Nlxon could either be a pacifist or not. The plausibility o f Nlxon being a pacifist (0.71) is higher than the plauslbdity that Nixon is not a pacifist (0.35). This is because Quakers have higher belief (0 90) o f being pacificists than Repubhcans have o f not being pacifists (0.80). Valuation-Based Language for Expert Systems 405 Figure 6. The evldenUal network with the prermse that Nlxon is a Repubhcan Figure 7. The evidential network with the prenuse that Nlxon Is a Quaker 406 Prakash P Shenoy Figure 8. The evidential network with the premise that Nlxon is not a paclficst Now suppose we enter the prermse that Nlxon is not a pacifist. This is represented by a bpa function as follows: m7({ - - p } ) = 1 The ES accepts this bpa function, and the results are displayed in Figure 8. Note that the assumption that Nlxon is not an exceptional Quaker has been retracted by the evidence! SUMMARY A N D CONCLUSIONS The main objective o f this article 1s to introduce a new language for budding knowledge-based systems as an alternative to rule-based-languages Whereas rule-based languages use rules as a knowledge representation device and modus ponens as an operation for making references, our language uses proper valuations as a knowledge representation device and three operations-- combination, marginallZatlon, and solution--for making inferences. Combina- tion corresponds to aggregation o f knowledge, marginallzation corresponds to crystalhzation o f knowledge, and solution is a decoding mechanism that maps knowledge from the space o f valuations to the space o f configurations. Conceptually, the language combines all valuations, finds the marginal o f the joint valuation for each variable, and then finds the solution for each marginal. Like rule-based languages, our valuation-based language retains the modular- Valuation-Based Language for Expert Systems 407 ity feature Each valuatmn represents a distract modular chunk o f knowledge. I f the combination operator is commutative and assocmtlve, then, like rule-based languages, valuation-based languages are nonprocedural. These desirable features o f rule-based languages are retained Unlike rule-based languages, our valuation-based language automatically maintains consistency in the knowledge base, caches and displays relevant inferences, reasons nonmonotomcally, and pernuts coherent management o f uncertainty A natural question IS, what is the computational p o w e r o f valuation-based languages? Anderson [67] has formally shown that is is possible to imagine coding any given Turing machine using a pure production system. W e suspect that valuation-based languages have the same computational power, but we do not have a proof. A C K N O W L E D G M E N T This w o r k was supported in part by the National Science Foundation under grant IRI-8610293 and a Research Opportunities in Auditing grant 87-135 f r o m the Peat M a r w l c k FoundaUon. The foundation on which tins paper rests is the result o f research done jointly with Glenn Shafer o v e r the last three years. I owe a lot to Glenn. I would also like to acknowledge the influence o f the A I group m the Business School Finally, this p a p e r has benefitted f r o m useful comments f r o m Bruce D ' A m b r o s i o and three anonymous referees. This paper is a rewsion o f Shenoy [68] References 1 Shenoy, P. P , and Shafer, G , An araomatlc framework for Bayesian and belief- funcuon propagation, Proc 4th Workshop on Uncertainty m AI, Minneapolis, Mlnn, 307-314, 1988 2 Shenoy, P P , and Shafer, G , Axioms for probabdlty and behef-functlon propagation, Workmg Paper No 209, School of Business, Umverslty of Kansas, Lawrence, K a n , 1988 3 Shenoy, P P , and Shafer, G., Constraint propagation, Workang Paper No 208, School of Business, Umverslty of Kansas, Lawrence, Kan., 1988 4 Shenoy, P P , and Shafer, G , Axioms for discrete optlratzatlon using local computauon, School of Business Working Paper No. 207, Umverslty of Kansas, Lawrence, K a n , 1988 5 Brownston, L. S , Farrell, R G , and, Kant, E., and Martin, N , Programming Expert Systems in OPS5: A n Introduction to Rule-Based Programmmg, Addison-Wesley, Reading, Mass., 1985 408 Prakash P. Shenoy 6 Davis, R , and King, J J , The origin o f rule-based systems in AI, m Rule-Based Expert Systems: The M Y C I N Experiments o f the Stanford Heuristw Program- mmg ProJect (B G Buchanan and E H Shortliffe, Eds ), Addison-Wesley, Reading, Mass., 20-52, 1984 7 Shafer, G , Shenoy, P P , and Srlvastava, R P , AUDITOR'S ASSISTANT a knowledge engmeermg tool for audit deoslons, Worlong Paper No 197, School of Business, University of Kansas, Lawrence, K a n , 1988 8 Pearl, J , Fusion, propagaUon and structuring in belief networks, A I 29, 241-288, 1986 9 Pearl, J , Networks o f Behef: Probablhstw Reasomng m Intelhgent Systems, Morgan Kanfmann, Palo Alto, Cal , 1988 10 Launtzen, S L , and Splegelhalter, D J , Local computations with probabflttles on graphical structures and their apphcatlon to expert systems (with discussion), J Roy. Star Soc. Ser B 50(2), 157-224, 1988 11 Heckerman, D E , and Horvltz, E J , On the expressiveness of rule-based systems for reasoning with uncertainty, Proc 6th National Conference on AI (AAAI-87), Seattle, W a s h , 1, 121-126, 1987 12 Sharer, G , and Shenoy, P P , Probability propagation, Working Paper No 200, School o f Business, University of Kansas, Lawrence, K a n , 1988 To appear in Proe 2nd International Workshop on AI and Statistics, Fort Lauderdale, F l a , 1989 13 Shafer, G , and Shenoy, P P , Local computation m hypertrees, Working Paper No 201, School of Business, University of Kansas, Lawrence, K a n , 1988 14 Seldel, R , A new method for solving constraint satisfaction problems, Proc 7th International Joint Conference on AI (IJCAI-81), Vancouver, B C , Canada, 1 , 3 3 8 - 342, 1981 15 Deehter, R , and Pearl, J , Tree-clustering schemes for constraint processing, Proc 7th National Conference on AI (AAAI-88), St Paul, Minn , 1, 150-154, 1988 16 Bertele, U , and Brloschl, F , Nonsertal Dynamw Programming, Academic, New York, 1972. 17 Rose, D J , A graph-theoretic study o f the numerical solution o f sparse posmve definite systems of linear equations, m Graph Theory and Computing (R C Read, Ed ), Acadenuc, New York, 183-217, 1973 18 Spohn, W , Ordinal condmonal functions a dynamic theory of epistemic states, in Causatton m Dectston, B e h e f Change, and Stattstws, Vol 2, (W L Harper and B Skyrms, Eds ), D Reidel, Dordrecht, Holland, 105-134, 1988 19 Hunter, D , Parallel belief revision, Proc 4th Workshop on Uncertainty in AI, Mmneapohs, Minn., 170-176, 1988 20. Malvestuto, F M , Decomposing complex contingency tables to reduce storage requirements, Proc 1986 Conference on Computational Statistics, 66-71, 1986 Valuation-Based Language for Expert Systems 409 21 Been, C , Fagm, R , Maler, D , and Yannakakas, M , On the deslrablhty o f acychc database schemes, J. ,4CM, 30(3), 479-513, 1983 22 Dempster, A. P , Construction and local computation aspects of network behef functions, Research Report S-125, Department of Statistics, Harvard Umverslty, Cambridge, Mass , 1988 23 Melnhold, R J , and Smgpurwalla, N D , Understanding the Kalman filter, A m . S t a t , 37, 241-288, 1982 24 Adams, E , Probablhty and the logic of condmonals, m Aspects o f Inductlve Logw (J Hmtlkka and P Suppes, Eds ,), North-Holland, New York, 1986 25 Suwa, M Scott, A C , and Shorthffe, E H , An approach to verifying completeness and consistency m a rule-based expert system, A I Mag. 3(3), 16-21, 1982 26 Nguyen, T A , Perlons, W A , Laffey, T J , and Pecora, D , Checkang an expert system's knowledge base for consistency and completeness, Proc 9th International Joint Conference on AI (IJCAI-85), Los Angeles, Cal , 1 , 3 7 5 - 3 7 8 , 1985 27 Pearl, J , Deciding consistency m inheritance networks, Tech Report No 870053 (R96), Cogmtlve Systems Laboratory, Umverslty o f Cahforma at Los Angeles, Cal , 1987 28 Touretzky, D S , The Mathematws o f Inherttance Systems, Morgan Kaufmann, Los Altos, C a l , 1986 29 Gmsberg, A , Knowledge-base reduction a new approach to checking knowledge bases for mconslstency and redundancy, Proc 7th NaUonal Conference on AI (AAAI-88), St Paul, M m n , l , 585-589, 1988 30. McCarthy, J , Clrcumscnptlon--a form of non-monotomc reasoning, A113, 27-39, 1980 31 McCarthy, J , and Hayes, P J , Some philosophical problems from the standpoint o f artlficml mtelhgence, m Machine Intelhgence, Vol. 4 (B Meltzer and D Mlchle, Eds ), Edinburgh Umverslty Press, 463-502, 1969 32 McDermott, D , and Doyle, J , Non-monotomc logic I, A I 13, 41-72, 1980 33 Moore, R C , Semantical consideration on nonmonotomc logic, A ! 25. 75-94, 1985 34 Relter, R , A logic for default reasoning, ,41 13, 81-132, 1980 35 Shafer, G , Probabdlty judgment m artlficml mtelhgence and expert systems, Stat. Scl. 2(1), 3--44, 1987 36 Heckerman, D E , and HorvltZ, E J , The myth o f modularity m rule-based systems for reasomng w~th uncertmnty, m Uncertamty zn Artificial Intelhgence, Vol 2 (J F Lemmer and L N Kanal, Eds ), North-Holland, New York, 23-34, 1988 410 Prakash P Shenoy 37 Shorthffe, E., and Buchanan, B. G , A model of inexact reasomng m medicine, Math. Btosct. 23, 351-379, 1975 38 Duda, R , Hart, P , and Ndsson, N , Subjective Bayesian methods for rule-based reference systems, m Readmgs m Arttfictal Intelhgence (B L Webber and N J Ndsson, Eds ), Tioga, Palo Alto, C a l , 192-200, 1981 39 Doyle, J., A truth maintenance system, A I 12(3), 231-272, 1979 40 de Kleer, J , An assumption-based TMS, A ! 28, 127-162, 1986 41 Relter, R , and de Kleer, J , Foundations of assumptaon-based truth maintenance systems" prehnunary report, Proc 6th National Conference on AI (AAAI-87), Seattle, W a s h , 1, 183-188, 1987 42 Laskey, K. B , and Lehner, P E , Behef maintenance an integrated approach to uncertmnty management, Proc 7th National Conference on AI (AAAI-88), St Paul, M m n , 1 , 2 1 0 - 2 1 4 , 1988 43. D'Ambroslo, B., A hybrid approach to reasomng under uncertmnty, Int. J. A p p r o x t m a t e Reasoning 2(1), 29--46, 1988 44 Dempster, A. P , New methods for reasoning toward posterior dlstnbuUons based on sample data, A n n . M a t h . Stat. 37, 355-374, 1966 45. Dempster, A P , Upper and lower probabllmes induced by a multlvalued mapping, A n n . Math. Stat. 38, 325-339, 1967 46 Pearl, J , Ewdentml reasomng using stochastic simulation o f causal models, A I 32, 245-257, 1987 47 garkpatnck, S , Gelatt, C D , J r , and Vecchl, M P , Optamlzatlon by simulated annealing, Science 220, 671-680, 1983 48 Geman, S , and Geman, D , Stochastic relaxation, Gibbs dlstnbuuon, and the Bayesmn restoration o f images, I E E E Trans. P A M I 6, 721-741, 1984 49 Ethenngton, D W , More on inheritance hierarchies with exceptions" default theories and mferentml distance, Proc 6th National Conference on AI (AAAI-87), Seattle, W a s h , 1 , 3 5 2 - 3 5 7 , 1987 50 Sharer, G , and Logan, R , Implementmg Dempster's rule for hierarchical evidence, A I 3 3 , 2 7 1 - 2 9 8 , 1987 51 Shenoy, P P , and Shafer, G., Propagating behef functions using local computa- tions, I E E E E x p e r t 1(3), 43-52, 1986. 52 Shenoy, P P , Shafer, G, and Mellouli, K (1986), Propagation o f b e h e f functions, a distributed approach, Proc 2nd Workshop on Uncertainty m AI, Phdadelphm, P e n n , 249-260, 1986 Also m Uncertamty m A r t O c l a i l n t e l h g e n c e , Vol 2 (J. F Lemmer and L N. Kanal, Eds.), North-Holland, New York, 325-335, 1988. 53 Kong, A , Multivariate behef functions and grapfucal models, Ph.D. Thesis, Department o f StatlsUcs, Harvard Umverslty, Cambridge, Mass., 1986 Valuation-Based Language for Expert Systems 411 54 Kong, A , A behef function generahzatlon o f Gibbs ensembles, Tech Report No 239, Department o f Statmtlcs, Umverslty of Chicago, Chicago, I l l , 1988 55 Dempster, A P , and Kong, A , Uncertain evtdence and artificial analysis, Research Report S-108, Department of Statisncs, Harvard University, Cambridge, M a s s , 1986 56. Shafer, G , Shenoy, P P , and Mellouh, K , Propagating behef functions in qualitative Markov trees, Int. J. Apprommate Reasomng 1(4), 349-400, 1987 57 MeUouh, K , On the propagation of behefs m networks using the Dempster-Shafer theory o f evtdence, Ph D Thesis, School of Business, University o f Kansas, Lawrence, K a n , 1987 58 Almond, R , Fusion and propagation in graphical belief models, Research Report S- 121, Department o f Statastics, Harvard Umverslty, Cambridge, M a s s , 1988 59 Zarley, D K., An evidential reasoning system, Working Paper No. 206, School of Business, University o f Kansas, Lawrence, K a n , 1988 60 Zarley, D K , Hsia, Y T , and Shafer, G , Evidential reasomng using DELIEF, Proc 7th National Conference on AI (AAAI-88), Minneapolis, M i n n , 1,205-209, 1988 61 Baldwin, J F , Evidential support logic programmang, Fuzzy Sets Syst. 24, 1-26, 1987 62 Ginsberg, M. L , Non-monotomc reasoning using Dempster's rule, Proc. 4th National Conference on AI (AAAI-84), Austin, T e x , 126-129, 1984 63 Rich, E., Default reasoning as likelihood reasoning, Proc. 3rd National Conference on AI (AAAI-83), Washington, D C , 348-351, 1983 64 Grosof, B N., Non-monotonicity m probablhstic reasomng, in Uncertamty m Arttftctal Intelhgence Vol 2 (J F Lemmer and L N Kanal, Eds ), North- Holland, New York, 237-249, 1988 65 Shafer, G , A Mathematwal Theory o f Evtdence, Pnnceton Umv Press, Princeton, N J , 1976 66 McCarthy, J , Applications of circumscription to formahzing common-sense knowledge, A I 28(1), 89-116, 1986 67 Anderson, J , Language, Memory and Thought, Erlbaum, Hfllsdale, N J , 1976 68 Shenoy, P P , Valuation systems a language for knowledge-based systems, Working Paper No 203, School of Business, Umverslty of Kansas, Lawrence, K a n , 1988 
work_ul5aon6enngwvleeiivqhyl54m	----	Cooperative Search for Fair Nurse Rosters Simon Martin1, Djamila Ouelhadj1, Pieter Smet2,3, Greet Vanden Berghe2,3, Ender Özcan4 1Logistics and Mathematics Management Group, University of Portsmouth, Department of Mathematics, UK email: simon.martin@port.ac.uk, djamila.ouelhadj@port.ac.uk 2CODeS, KAHO, Computer science, Belgium ITEC, iMinds, KULeuven, Belgium. e-mail: pieter.smet@kahosl.be, greet.vandenberghe@kahosl.be 4Automated Scheduling, Optimisation and Planning Research group, University of Nottingham, Department of Computer Science, UK email: exo@nottingham.ac.uk Abstract The development of decision support systems acceptable for nurse roster- ing practitioners still presents a daunting challenge. Building on an existing nurse rostering problem, a set of fairness-based objective functions recently introduced in the literature has been extended. To this end, a generic agent- based cooperative search framework utilizing new mechanisms is described, aiming to combine the strengths of multiple metaheuristics. These different metaheuristics represent individual planners’ implicit procedures for improv- ing rosters. The framework enables to explore different ways of assessing nurse rosters in terms of fairness objectives. Computational experiments have been conducted across a set of benchmark instances. The overall results indi- cate that the proposed cooperative search for fair nurse rosters outperforms each metaheuristic run individually. Keywords: Cooperative search, agent-based systems, nurse rostering, fairness. 1. Introduction Optimisation problems are usually solved by one or more human planners who developed a mental model of the objective to be optimised and a heuristic procedure for generating solutions. One of the hardest issues to address when Preprint submitted to Expert Systems with Applications May 14, 2013 developing decision support systems is modelling of the problem in a way that the human planners experience as natural, while producing solutions that reveal part of the optimisation process. When multiple decision makers are involved, not necessarily sharing the same models and the same heuristics, the problem turns out to be even more challenging. An obvious example can be found in hospitals, where individual nurses cooperate and have a say in establishing the monthly roster for the entire ward, at least when the process is not fully automatised. An interesting question is whether situations with multiple decision makers can be mimicked into an optimisation approach and whether the obtained results are better than after optimisation from a single planner’s point of view. Nurse rostering is a domain of interest to many researchers and practition- ers across different fields, including computer science, operational research and artificial intelligence (Burke et al., 2004, 2001; Beddoe and Petrovic, 2006; Özcan, 2005, 2007; Petrovic and Vanden Berghe, 2012). A nurse ros- ter is a timetable consisting of shift assignments of nurses at a health-care institution. Creating nurse rosters subject to a set of constraints based on individual preferences and on requirements imposed by administration is a challenging task (Even et al., 1976; Karp, 1972). Cooperative search is defined as a search process performed by agents ca- pable of exchanging information, e.g., models, solutions and/or other search space characteristics during the search process (Clearwater et al., 1992; Crainic and Toulouse, 2008; Blum and Roli, 2003; Hogg and Williams, 1993; Toulouse et al., 1999). It has been successfully used to solve a number of difficult combinatorial optimisation problems, such as multi-commodity lo- cation with balancing requirements (Crainic et al., 1995, 1997), capacitated network design (Crainic and Gendreau, 2002), vehicle routing (Bouthillier and Crainic, 2005), quadratic assignment (James et al., 2009), labour con- straint scheduling (Cavalcante et al., 2001), permutation flow shop scheduling (Ouelhadj and Petrovic, 2010; Vallada and Ruiz, 2009). These studies have shown that the combination of several metaheuristics with different parame- ter settings increases the robustness of the global search relative to variations in problem instance characteristics. Agent-based cooperative metaheuristic approaches are fairly new in nurse rostering. Wang and Wang (2009) developed an agent-based approach to self rostering. Haspeslagh et al. (2009) addressed the problem of exchanging nurses between wards and sorting out personnel shortages. Haspeslagh et al. (2009) presented a Pareto optimal negotiation approach for leveraging the 2 workload across different wards. Most of the work on agent-based fairness centres around game theoretic approaches (De Jong et al., 2008) where the agents are used to model human behaviour and interactions. To the best of our knowledge, there has been no work in the literature on agent-based cooperative search for fairness making use of multiple fairness-based objective functions. Smet et al. (2013b) introduced different different fairness-based objec- tive functions representing individual planners’ impressions of good quality rosters We propose a novel cooperative search agent-based framework that uses this recent work. (Smet et al., 2013a) present different solution methods which may correspond to the planners’ individual and implicit procedures for improving rosters. The cooperative search agent-based framework combines the strengths of different metaheuristics and fairness models. The agents are able to use the same or different fairness objective functions enabling them to find the most appropriate model for a given problem. By the same to- ken, each agent in this framework is autonomous and capable of executing different metaheuristic combinations with different parameter settings. The agents cooperate asynchronously using pattern matching and reinforcement learning to produce high quality solutions with respect to fairness. A brief overview of the nurse rostering problem and a set of alternative fairness objective functions along with measures of fairness for nurse ros- tering are provided in Section 2. In Section 3, we present an agent-based framework for cooperative metaheuristic search for fair nurse rostering. The experimental design is discussed in Section 4, including the implementation details of different metaheuristic agents. The results of the computational experiments are reported in Section 5. Section 6 concludes with a discussion on the contribution and some directions for future research. 2. Problem description and fairness-based objective functions The goal of automated nurse rostering is to assign shifts to a set of nurses so that 1) the minimum staff requirements are fulfilled and 2) the nurses’ contracts are respected (Burke et al., 2004). Nurse rostering is a highly constrained scheduling problem which was proven to be NP-hard (Karp, 1972) in its simplified form. The problem is represented as a constraint optimisation problem using 5-tuples: (i) set of nurses, (ii) set of days (periods) including the relevant bits from the previous 3 Hard constraints Soft constraints Single Assignment Start Per Nurse Per Day Coverage Constraints No Overlap between Assignments Assignment to the Primary Skill Honour Skill Types Rest Times Operations on Defined Assignments Only Requests Schedule Locks Time-related Constraints Table 1: Hard and soft constraints and upcoming schedule, (iii) set of shift types, (iv) set of skill types and (v) constraints. The nurse rostering model addressed in this paper is based on the model of Bilgin et al. (2012). A roster consists of a set of assignments from the discrete space defined by nurses, skill types, shift types, and schedule period subject to given constraints. It defines different types of constraints: hard, soft, time-related and coverage constraints. The hard and soft constraints of the model are described in Table 1 and a full description of this model can be found in Bilgin et al. (2012); Smet et al. (2013a). A roster is feasible if and only if all hard constraints are satisfied. Times constraints are a type of soft constraint concerned with sequences of shift patterns. Soft constraints represent preferences. A solution method aims at resolving as many soft con- straints as possible for a given problem. When a soft constraint is violated, a penalty is incurred proportional to the importance of the constraint and the severity of its violation. The importance of each constraint is expressed using weights. The higher the weight, the more important the constraint is, rela- tive to the other constraints. A frequently used objective function in nurse rostering is referred to as the weighted sum objective function MinWS. Let c ∈ C be the set of soft constraints then, wc is the weight associated with constraint c and nc the number of violations of c. We define: qros(i) = ∑ 0≤c≤|C| wcnc (1) We also define pviol which represents the coverage violations accrued by any over- or under-staffing in a given roster. Therefore MinWS consists of two parts: qros(i) associated with the costs of assigning a nurse to a given shift and pviol the coverage constraints. Thus the objective is to find a roster with 4 the lowest overall penalty, denoted as MinWS. MinWS = min (∑ n∈N qros(n) + pviol ) (2) Smet et al. (2013b) address the issue of fairness in automated nurse ros- tering. The fair distribution of contractual violations among nurses is an important contributor to the satisfaction nurses get based on their work schedules, and thus a large influence on the overall job satisfaction of nurses. Smet et al. (2013b) show that, the commonly used MinWS does not result in fair solutions. They investigated alternative fairness models from the lit- erature (Julian et al., 2002; Vasupongayya and Chiang, 2005) to solve the problem of fairness in nurse rostering and show that by optimising these models, fairer solutions are obtained, generally at the cost of an increased number of constraint violations. The objective functions they introduced will also be used in this study. They are defined below by Equations 4, 5 and 6. All these fairness objective functions try to even the distribution of nurse roster violations so that no nurse is favoured over another. First of all, we define the average individual roster quality of the nurses: µ, calculated as: µ = 1 |N| ∑ n∈N qros(n) (3) Based on (Smet et al., 2013b) we describe the four fairness objective functions used in this study: MinMax = min ( |N| ( max n∈N qros(n) ) + pviol ) (4) MinDev = min ((∑ n∈N (|µ− qros(n)|) + |N|µ ) + pviol ) (5) MinError = min ( |N| (( max n∈N qros(n) − min n∈N qros(n) ) + µ ) + pviol ) (6) 5 We also introduce the objective function MinSS which is similar to MinMax as it tries to reduce the worst roster of a given nurse. However, it does this by emphasising on the worst individual rosters by squaring. The aim is therefore to reduce these poor individual rosters. w1 and w2 are weights used to scale the two parts of equation. MinSS = min ( w1 ∑ n∈N qros(n) 2 + w2p 2 viol ) (7) 3. Agent-based cooperative metaheuristic search 3.1. An agent-based framework for cooperative metaheuristic search Cooperative search provides a class of strategies to design more effective search methodologies by combining metaheuristics for solving combinatorial optimisation problems. This area has been little explored in operational re- search. We introduce a generic agent-based distributed framework where each agent implements a metaheuristic and a fairness objective function. An agent continuously adapts itself during the search process using a coopera- tion protocol based on reinforcement learning and pattern matching. Good patterns which make up improving solutions, are identified and shared by the agents. The agent-based system is composed of a launcher agent and metaheuris- tic agents (Martin et al., 2012). • Launcher agent: The launcher agent reads problem instances from XML benchmark files, executes seed algorithms using one of the fairness objective functions and then sends the seed schedules, one at a time, to each of the metaheuristic agents. After the improving phase, it also collects the final schedules from the metaheuristic agents and selects the fairest rosters. • Metaheuristic agent: The metaheuristic agents receive a problem from the the launcher agent which they try to optimise using their given metaheuristic and local search heuristic combinations. They also cooperate asynchronously with the other agents according to a coop- eration protocol where the agents try to find recurring good schedule patterns. The agents share the good patterns and each try to build new improving rosters which they then optimise according to the given 6 fairness objective functions and metaheuristics. After a number of it- erations, the metaheuristic agents each send their best roster to the launcher. The launcher then chooses the fairest roster. The metaheuristic agents cooperate asynchronously by iterating a com- munication protocol which is a distributed algorithm. At each iteration a conversation initiator (identified in the previous iteration) sends the result of its latest metaheuristic search using its own fairness objective function to the other agents. The agents compare this result with their own and then generate patterns (Section 3.3) which they send back to the initiator. The initiator then identifies good patterns and shares them amongst the meta- heuristic agents. Each agent then generates a new potential solution. The metaheuristic agent with the lowest objective function value in the present iteration is chosen to be the initiator of the next. In the next iteration the new potential solutions are used to further the search in the manner just described. Figure 1 shows how the agents perform a search by cooperating with each other running asynchronously in parallel and executing different meta- heuristic and heuristic combinations with different parameter settings. They use ontologies (Section 3.2) to enable the agent-based system to be adapted easily to new problem domains. The internal structure of a metaheuristic agent is mediated at two levels through the use of ontologies. Ontologies are defined as a set of general representational primitives to model semantic elements of a domain (Gruber, 1993). 3.2. Ontologies The ontology currently used by the framework generalises the notions of: assignments, pairs of assignments, constraints and roster or schedule. In the ontology these are called SolutionElements, HeuristicData, Constraints and SolutionData objects respectively. • SolutionElements: A SolutionElement represents the assignment of a nurse with specific skills to a shift on a certain day as required by a given nurse rostering problem. • HeuristicData: A HeuristicData object contains two SolutionEle- ments objects. These represent pairs of assignments in a roster that 7 Launcher agent Metaheuristic agentN E x c h a n g e g o o d s o lu tio n s a n d p a tte rn s a t Create agent/Solve problem A best-found N-1 Metaheuristic agent N-1 S e n d g o o d s o lu ti o n s a t th e e n d o f e v e ry M Send pattern N-2 Receive s best-found 2 best-found N-2 E x c h a n g e g o o d s o lu tio n s a n d p a tte rn s a t e a c h c o n v e rs a tio n Receive i,pattern i Send s: best-so-far N-2 Send pattern 2 Metaheuristic agent 2 Metaheuristic agent N-20 S e n d g o o d s o lu ti o n s a t th e e n d o f e v e ry M c o n v e rs a ti o n s leader of the jth conversation Receive s best-found 1 Metaheuristic agent 1 E x c h a n g e g o o d s o lu tio n s a n d p a tte rn s a t = Conversations between launcher and metaheuristic agents = Conversations between metaheuristic agents S e n d g o o d s o lu ti o n s a t th e e n d o f e v e ry M Send pattern 1 Receive s Receive s Figure 1: Agent-based framework for cooperative metaheuristic search will be used in the cooperation protocol to identify good patterns of assignments in improving rosters. • Constraints: The Constraints interface is the interface between the high level framework and the concrete constraints used by the nurse rostering problem. These are used to verify a valid roster. • SolutionData: A SolutionData object represents the list of assign- ments that make up a valid nurse roster. The ontology is represented in XML, corresponding to the representa- tion of most of the nurse rostering benchmark problems. This makes the interface between problem definition and ontology seamless in practice. As a consequence the elements defined in a specific problem instance will be used 8 directly by the system. Figure 2 shows the structure of the ontology and how SolutionElements is the interface between the framework and the nurse rostering problem instances. The SolutionElements, HeuristicData, Constraints and SolutionData ob- jects are also used in the agents to facilitate inter-agent cooperation. The next section describes the cooperation protocol and how this is im- plemented with the use of ontologies. Figure 2: The interface between the framework and problem instances 3.3. Asynchronous cooperative search by pattern matching and reinforcement learning for nurse rostering The agents cooperate by exchanging patterns defined as a permutation of assignments. This is a simple distributed metaheuristic. Given such a permutation of assignments, it is always possible to generate n pairs from it. In the following example, we define the permutation of nurse assignments using the following ID’s, where n = 10: <2,4,7,6,5,8,9,0,1,3>. The following n pairs can be generated: (2, 4), (4, 7), (7, 6), (6, 5), (5, 8), (8, 9), (9, 0), (0, 1), (1, 3), (3, 2) 9 . The permutation is broken into patterns of the same length while retain- ing the basic order of the permutation. The agents can then each compare pairs of assignments generated by their own metaheuristic with pairs shared by the other agents. All the pairs from each agent are scored based on how frequently they appear. Only those pairs that have the highest frequency score are shared out amongst the agents. The idea behind this step is that the lists of assignments have already been optimised by each agent’s respec- tive metaheuristic so each list must have elements that make up a good roster. It is argued that these pairs will occur more frequently as the search progresses. Once the good pairs are shared out, the agents use these pairs to generate a new roster by using a greedy heuristic. This is a simple heuristic that makes the new roster by taking the good patterns first and then filling the rest of the roster with the agent’s previous best-so-far permutation. The new permutation of assignments is now ready to be optimised by an agent’s metaheuristic and fairness objective function. This protocol involves diversification and intensification phases. Diversi- fication involves pattern matching and the application of the greedy heuristic to produce a new roster. This roster is then intensified using a metaheuristic. One application of these two phases is called a conversation. It should be noted that the pattern matching results in a controlled diversification of the search. Good patterns are retained and used to build new potential solutions which will diversify the search just enough to enable new areas of the solution space to be explored. The metaheuristics then intensify the search. The asynchronous cooperative search is implemented as follows: 1. A metaheuristic agent taking on the role of conversation initiator starts a conversation. It takes a new roster either generated from a previous conversation or supplied by the launcher agent. The new roster is then improved by the initiator agent. When an improved roster is generated locally, it is sent to the other metaheuristic agents. 2. The metaheuristic agents have also generated their best-so-far rosters using their fairness objective functions. They break up the rosters sent from the initiator and their own into pairs. The pairs are then compared and only those that are common to both rosters are kept. HeuristicData objects are created from these pairs storing the first and second assignments of the pair. These are then sent by the metaheuris- 10 tic agents to the initiator. The metaheuristic agents also send the value of their fairest roster found so far. These rosters will be used by the ini- tiator to determine which metaheuristic agent will be the new initiator in the next conversation. 3. Upon receiving the HeuristicData objects from the other metaheuristic agents, the initiator pools them locally. Each HeuristicData object is scored by counting how frequently it occurs in the pool. The initiator then tries to build a linked list from these high scoring HeuristicData objects. For example, if the pool contains the following HeuristicData objects with first and second assignments expressed here as pairs (4,7) (6,1) (7,2) (2,6) (5,9) (3,8), the linked list generated from the HeuristicData objects will have the following order (4,7) (7,2) (2,6) (6,1). Any Heuris- ticData objects not linked in this way are stored in an unlinked list (5,9) (3,8). 4. The initiator then determines which metaheuristic agent is going to be the initiator in the next conversation. This is done by pooling all the fairness objective function values of the best rosters found so far by the metaheuristic agents and then identifying which metaheuristic agent has the best fairness objective function value. The metaheuristic agent with the best objective value will be the new initiator in the next conversation. The initiator then sends these lists of best-so-far HeuristicData objects to the metaheuristic agents. In the same message it also indicates which metaheuristic agent will be the new initiator in the next conversation. 5. The metaheuristic agents receive the lists of HeuristicData objects. Both initiator and metaheuristic agents then create a new roster from these objects, as well as their current best roster. The new roster is created by trying to build first a list of HeuristicData objects, while the metaheuristic agent’s best-so-far roster provides any missing numbers. In this way a new unique roster for each agent is generated and the objective function value is calculated. 6. The conversations are repeatedly exchanged between the metaheuristic agents for a maximum number of conversations set in the configuration file of the launcher agent. 11 4. Experimental Design Section 4.1 discusses the configuration of the agent-based system. Section 4.2 real-world examples from two Belgian hospitals are described that are used in the experiments. We also describe two different scenarios and the measures used to verify the fairness of a roster. 4.1. Implementation of the agent-based framework The agent-based framework for cooperative search is implemented using the open source FIPA compliant development platform JADE (Bellifemine et al., 2005). We implemented three different metaheuristic agents in the framework: • Tabu Search agents (Glover, 1990): The tabu search agents im- plement a basic tabu search. The search starts from a feasible roster and moves iteratively from the current roster to its best neighbouring roster using neighbourhoods even if that neighbourhood worsens the objective function value. • Simulated Annealing agents (Bertsimas and Tsitsiklis, 1993): The simulated annealing agents implement the basic algorithm with a geometric cooling schedule. • Variable Neighbourhood Search agents (Bilgin et al., 2012): The variable neighbourhood search agents implement a perturbation heuristic which uses local search heuristics to improve an initial solu- tion. The algorithm proceeds by randomly choosing a heuristic from the list of available neighbourhood operators. Each possible move is validated against the hard constraints ensuring that at every stage of the search each new solution remains feasible. Tabu Search, Simulated Annealing and VNS use the local search moves developed by Bilgin et al. (2012). The neighbourhoods are: • Shift: It assigns a nurse to a new shift to the roster. • Delete shift: This move deletes a shift from the roster. • Single shift day: An assignment is removed from a nurse’s schedule and added to another nurse on the same day if the second nurse has no assignment on that day and has the associated skill type. 12 • Change assignment based on compatible shift type: The shift type of an assignment is changed to another compatible shift type de- fined in the coverage constraints for the associated day and skill type. • Change assignment based on skill type: This neighbourhood can be used when the nurses have at least two different skill types. It deletes an assignment and adds another assignment to one of the nurse’s other skill types. • General Assignment Change: An assignment is changed to another shift type where the skill type of the assignment remains the same. The experiments were conducted using a network of 13 agents, configured as follows: • Launcher agent • 4 Tabu Search agents • 4 Simulated Annealing agents • 4 Variable Neighbourhood Search agents The metaheuristic agents are configured from a file which they execute at start-up. This file specifies to the agent which metaheuristic to run; how many iterations to perform; which fairness objective function to use; also, parameters such as the size of tabu tenure or which cooling schedule function to use if configured to run SA. All agents run the same local search heuristics described above. The parameter settings of the individual agents remain unchanged throughout all testing. Each agent evaluates its given metaheuristic for 500 iterations or until no further improving solutions are found. Experimentation shows that it provides a good balance between diversification using pattern matching while the 500 iterations of the metaheuristic intensifies the new ros- ter just enough that the agents tend to converge to good rosters. TS agents have a tabu tenure set to seven. SA agents use a geometric temperature cooling schedule set at 0.9 for a more diverse search. All agents randomly select the next local search heuristic to run. Each agent ran on a HP Compaq 6000 pro with Intel Core 2 duo E8400 processor with 4 GB RAM in a 2x2 GB configuration. The agents were configured to use only 1 GB of memory. 13 4.2. Experimental data and fairness evaluation measures The experiments were conducted on the data used by Smet et al. (2013b)1. The instances are based on four different hospital wards from two Belgian hospitals: emergency, geriatrics, psychiatry and reception. Table 2 provides an overview of the instance characteristics. For each ward, two cases are con- sidered: one where all the nurses have the same contract (referred to as name of instance i), and one where each nurse has an individual contract with both common and personalised constraints (referred to as name of instance d). Instance No. of nurses No. of shifts No. of skills Planning period Emergency 27 27 4 28 days Geriatrics 21 9 2 28 days Psychiatry 19 14 3 31 days Reception 19 19 4 42 days Table 2: The properties of the benchmark instances used during the experi- ments obtained from a Belgian hospital. The fairness objective functions guiding the search process towards fair rosters and the metrics measuring fairness are separated. This is due to the reason that the objective functions take into account different compo- nents, particularly, contractual constraints and coverage constraints during the search process, while the fairness metrics ignore the coverage constraints to better assess how fair the full roster is from the point of view of nurses. In this study, we use the Jains fairness index (Jain et al., 1984; Muhlenthaler and Wanka, 2012), denoted by Jains, to measure the fairness of a roster. Equation 8 shows how this measure is calculated for a given solution with N, the set of nurses. Jains = ( ∑ n∈N qros(n) )2 |N|· ∑ n∈N q2ros(n) (8) 1http://allserv.kahosl.be/~pieter/nurserostering 14 The value of Jains fairness index varies from 1/|N| (worst fairness case) to 1 (best fairness case), which denotes that the violations for each individual nurse is the same. If the assignment of a set of nurses to form a roster generates no violations then the value of the roster is 1. In such as case the roster is deemed to be completely fair. To compare the performance of two different fairness objective functions, A and B over a given set of instances, we use %gap indicating the percentage change that an algorithm using the objective function B generates over A (i.e., gap from A to B): %gap = A−B B × 100 (9) Assuming that A = MinFO, where MinFO ∈ {MinMax, MinDev, MinError, MinSS} and B = MinWS, if an algorithm using the minimis- ing objective function MinFO improves the (average) solution quality when compared to an algorithm using MinWS over a set of runs for an instance, then (average) %gap < 0. If B is 0 the search has found a solution with no violations. This %gap measure is not applicable (N/A) in this case. 4.3. Experimental Scenarios Two experimental scenarios were conducted to evaluate the performance of the proposed cooperative search. In the first scenario, all 12 metaheuristic agents executed the same objective function on all the instances. Each ob- jective function was tested separately in this manner. The same tests were conducted in stand alone mode where metaheuristic agents do not cooper- ate. Here a single agent runs the same experiments evaluating its fairness objective function the same number of times as all the 12 cooperating agents added together. For example, in cooperation mode 12 agents conducted 200 conversations evaluating their objective functions as most 500 times. There- fore, in the stand alone experiments, a metaheuristic evaluates the fairness objective function 12 × 200 × 500 = 1200000 times. In all the experiments, w1 = w2 = 1 for MinSS. The Jains fairness index defined in (Equation 8) is used to evaluate the relative fairness of the solutions produced by each configuration. The average percentage increase (Equation 9) from MinWS for each instance was used to evaluate the efficiency of each fairness objective function in terms of number of constraint violations. 15 In the second scenario, the same instances were used but all four fairness objective functions are used at the same time to generate the nurse rosters. The idea behind this test is to identify which fairness measures are best for each problem instance. Of the twelve metaheuristic agents, a group of three different metaheuristic agents generated nurse rosters using MinMax, three ran MinDev, three ran MinError and finally three ran MinSS. When the search is finished, each agent sends their best roster to the launcher which then selects the roster with the largest Jains fairness measure. All experiments were conducted on real-world benchmark problems with 20 repeated runs for each problem instance. In the first scenario, these ex- periments were conducted five times for each of the five objective functions. The resulting nurse rosters were then each recalculated using the other ob- jective functions not used in the optimisation of the roster so that they could be compared with the original. In the second scenario each experiment was also repeated 20 times for each instance. The agents conducted only 200 conversations to complete each search taking no longer than 6 minutes to produce a new roster for each problem instance. 5. Computational Results In this section, we present the results of the two test scenarios described in Section 4.3. In the first test scenario, the metaheuristic agents use coop- erative search to generate the nurse rosters using the same fairness objective function, while in the second one, the metaheuristic agents use cooperative search and a mixture of the fairness objective functions to generate the nurse rosters. As a statistical test, a Wilcoxon singed-rank test has been performed to compare pairwise performance of two given algorithms. The following no- tation is used: Given A versus B, < (≤) denotes that B is better than A and this performance variance is (not) statistically significant within a 95% confidence level. The software used for generating and evaluating rosters was provided by the authors of (Smet et al., 2013a)2. 5.1. Scenario 1 test results Table 3 summarises the first scenario results using average Jains fairness index over 20 runs for each instance. Considering the average performance 2Personnel rostering software kernel, KAHO, Gent 16 of cooperative search with respect to the Jains fairness index using differ- ent fairness objective functions, MinDev ad MinMax are the best choices under scenario 1, performing well on all instances. Their performance is sig- nificantly better than the rest of the objective functions. MinDev performs better than MinMax on average across five instances including Emergency i, Emergency d, Geriatric i, Geriatric d, Psychiatry d and this performance difference is statistically significant. MinDev and MinMax both deliver a similar performance on Psychiatry i. On the other hand, MinMax performs significantly better than MinDev on the Reception instances. MinError, MinSS and MinWS follows MinMax in the given order when their over- all average performances are compared. Smet et al. (2013b) showed that MinWS produces the least fair rosters and similar phenomena are observed under the cooperative search Scenario 1 setting. Instance MinWS MinMax MinDev MinError MinSS Emergency i 0.5638 0.9983 0.9991 0.9074 0.9609 Geriatric i 0.4451 0.9784 0.9861 0.6023 0.5613 Psychiatry i 0.6069 0.9995 0.9995 0.9054 0.8938 Reception i 0.6273 0.8808 0.8515 0.7191 0.1076 Emergency d 0.6295 0.9973 0.9982 0.8574 0.9228 Geriatric d 0.6013 0.9980 0.9986 0.9491 0.3349 Psychiatry d 0.4446 0.9986 0.9988 0.6882 0.6719 Reception d 0.3321 0.9895 0.9824 0.8052 0.2603 Table 3: The average Jains fairness index over 20 runs of a given fairness- based objective function for each benchmark instance under Scenario 1, where the bold entries indicate the best one for a given instance. Table 4 shows the average Jains fairness index over 20 runs for each in- stance considering the stand-alone tests in which a single metaheuristic is used as a sequential metaheuristic. When a given metaheuristic makes use of a different objective function, its performance changes. In general, MinDev is a better choice than the other fairness-based objective functions to be used by a stand-alone metaheuristic. MinMax is the second best choice as an objective function to produce fair rosters. As expected, MinWS does not generate high quality solutions as well as the other objective functions in terms of fairness in most of the cases. MinWS generates fairer solutions when compared to MinSS for Reception i, regardless of the standalone algo- rithm used. A similar phenomenon is observed for the standalone VNS using MinWS when compared to VNS using MinWS on the Geriatric instances. 17 VNS performs slightly better than the other stand-alone metaheuristics in the overall when MinDev is used as the fairness-based objective function. Cooperative search based on Scenario 1 using MinDev performs significantly better than all stand-alone metaheuristics regardless of the objective function used on all instances, except for Psychiatry i and Reception i. For Psychi- atry i, Scenario 1 using MinDev performs similar to the stand-alone VNS using MinDev. For Reception i, stand-alone VNS using MinMax performs better than Scenario 1 using MinDev. Figure 3 provides the box plots of Jains fairness index for the Emergency i instance over 20 runs considering cooperative and stand-alone metaheuristics using different objective func- tions as an example. The results indicate the success of cooperative fairness approach over the approach without cooperation when MinDev, MinMax or MinSS is used as an objective function. This performance variation is statistically significant. 18 Stand-alone VNS Instance MinWS MinMax MinDev MinError MinSS Emergency i 0.7477 0.9812 0.9966 0.8697 0.9605 Geriatric i 0.5646 0.9714 0.9646 0.5413 0.6101 Psychiatry i 0.8070 0.9874 0.9995 0.9652 0.9809 Reception i 0.7148 0.8893 0.8277 0.8192 0.5766 Emergency d 0.6829 0.9845 0.9955 0.8094 0.9556 Geriatric d 0.6362 0.9747 0.9981 0.9466 0.4337 Psychiatry d 0.6299 0.8929 0.9931 0.7714 0.9158 Reception d 0.4970 0.8654 0.7106 0.6601 0.5618 Stand-alone Simulated Annealing Instance MinWS MinMax MinDev MinError MinSS Emergency i 0.8657 0.9751 0.9927 0.9337 0.9314 Geriatric i 0.6789 0.9718 0.9533 0.8117 0.7530 Psychiatry i 0.8979 0.9542 0.9929 0.9532 0.9809 Reception i 0.7660 0.8380 0.8332 0.8323 0.7792 Emergency d 0.8222 0.9592 0.9706 0.8878 0.8869 Geriatric d 0.7572 0.9460 0.9820 0.8838 0.9175 Psychiatry d 0.6946 0.9160 0.9767 0.7723 0.8998 Reception d 0.6204 0.8451 0.7449 0.7184 0.7157 Stand-alone Tabu Search Instance MinWS MinMax MinDev MinError MinSS Emergency i 0.8398 0.9821 0.9900 0.9274 0.9406 Geriatric i 0.6949 0.9618 0.9660 0.7306 0.7320 Psychiatry i 0.8693 0.9778 0.9979 0.9699 0.9721 Reception i 0.7875 0.8532 0.8071 0.7995 0.7681 Emergency d 0.7562 0.9711 0.9921 0.8784 0.9394 Geriatric d 0.7256 0.9551 0.9391 0.8577 0.9245 Psychiatry d 0.7101 0.8857 0.9650 0.7783 0.9048 Reception d 0.5796 0.8033 0.7680 0.6397 0.7734 Table 4: The average Jains fairness index obtained over 20 runs for a given fairness-based objective function for the stand-alone case, where the bold entries indicate the best one for a given instance. 19 (a) MinMax (b) MinDev (c) MinError (d) MinSS Figure 3: The box plots of Scenario 1 algorithm and stand-alone metaheuris- tics based on Jains fairness index values obtained from 20 runs on Emer- gency i for each fairness-based objective function. We have re-evaluated all rosters produced at the end of each run by the Scenario 1 algorithm using a given fairness-based objective function with respect to the MinWS objective function. For each instance, the average result obtained by Scenario 1 using MinWS as objective function over 20 runs is used as a basis. Table 5 provides the average %gap (Equation 9) from the basis to the average MinWS over 20 runs when a fairness-based objec- tive function is used. The negative entries indicate improvement. Although MinDev and MinMax produce more fair rosters, the quality of these rosters is not as good in terms of the generic objective value, as expected (Smet et al., 2013b). In Scenario 1, it can be observed that MinSS and MinError performs better than the rest of the fairness-based objective functions on average, generating high quality roster with respect to MinWS. Scenario 1 using MinError and MinSS make 5-7% and 80-90% improvement upon the results of Scenario 1 algorithm using MinWS for the Reception instances and Geriatric i, respectively. On the other hand, MinSS is one of the worst performing objective function in terms of fairness (Table 3). 20 Instance MinWS MinMax MinDev MinError MinSS Emergency i 23245 367.55% 37.21% 22.88% 39.55% Geriatric i 3799 316.98% 427.11% 61.66% 85.34% Psychiatry i 29516 331.34% 315.43% 1.46% 8.45% Reception i 66719 39.36% 18.72% -5.21% -81.57% Emergency d 17477 465.85% 406.33% 39.54% 45.72% Geriatric d 38835 18.07% 74.55% -6.99% -90.03% Psychiatry d 11591 642.50% 600.85% 1.66% 41.61% Reception d 19018 71.84% 113.35% -5.40% -79.55% Table 5: The average % gap from MinWS to the MinWS value of each fairness-based objective function for each instance. 5.2. Scenario 2 test results and comparison to Scenario 1 In the second test scenario, all the fairness models are used and tested at once. This is motivated by the idea that people planning a roster will have different concepts/models of fairness. If these models are allowed to cooperate, will the rosters be fair and obtained fast? Furthermore will this approach enable us to identify the best performing fairness-based objective function and guide search process? To this end the experiments are con- ducted in accordance with the first scenario in that 12 metaheuristic agents and 1 launcher employed under the cooperative search framework and 20 runs are performed for each instance. However, in this scenario three agents are grouped to form four different groups. Each group optimises with respect to one of the MinMax, MinDev, MinError and MinSS fairness-based objective functions, separately. Table 6 provides the results for Scenario 2 and compares its performance to the best stand-alone metaheuristic, VNS using MinMax and cooperating agents under Scenario 1 using MinDev based on the average and best Jains fairness index over 20 runs. The Scenario 2 algorithm outperforms all stand- alone algorithms (Table 4) regardless of the objective function generating the most fair rosters for any given instance and this performance variation is statistically significant for all instances. Moreover, cooperative search under Scenario 2 consistently outperforms Scenario 1 using MinDev or any other fairness based objective function on average and this performance difference is statistically significant for each instance. Scenario 1 and 2 produce the best rosters for three and four instances, respectively, with a tie in one instance. Figure 4 provides the box plots of Jains fairness index for the Geriatric i instance over 20 runs considering cooperative and stand-alone metaheuris- tics using different objective functions as an example. VNS, Tabu and SA 21 performs best using MinDev on this instance. The results confirm the suc- cess of Scenario 2 over an approach with no cooperation even if it is used with the best performing objective function and this performance variation is statistically significant, except for Emergency i. Still, Scenario 2 performs slightly better than Scenario 1 using MinDev. Best-of-runs Average Instance VNS Scenario 1 Scenario 2 VNS Scenario 1 vs. Scenario 2 Emergency i 0.9993 0.9998 0.9997 0.9966 0.9991 ≤ 0.9993 Geriatric i 0.9946 0.9978 0.9980 0.9646 0.9861 < 0.9942 Psychiatry i 0.9997 0.9999 0.9999 0.9995 0.9995 < 0.9997 Reception i 0.8439 0.9540 0.9557 0.8277 0.8515 < 0.9528 Emergency d 0.9982 0.9997 0.9995 0.9955 0.9982 < 0.9989 Geriatric d 0.9991 0.9993 0.9998 0.9981 0.9986 < 0.9992 Psychiatry d 0.9994 0.9999 0.9997 0.9931 0.9988 < 0.9992 Reception d 0.7727 0.9949 0.9999 0.7106 0.9824 < 0.9986 Table 6: Performance comparison of stand-alone VNS metaheuristic and scenario one using MinDev, and Scenario 2 based on the average and best- of-runs Jains fairness index values over 20 trials for each instance. Bold entries mark the best algorithm. Figure 4: The box plot for Scenario 2 (S2) and Scenario 1 (S1), Tabu, SA and VNS using MinDev based on Jains fairness index obtained from 20 runs for Geriatric i. Table 7 provides the average %gap (see Equation 9) from the basis (first column) in Table 5 to the average MinWS over 20 runs under Scenario 2. This quantity indicates how MinWS objective value changes while the fairness is optimised under Scenario 2. The results are consistent with the previous findings by Smet et al. (2013b), indicating that optimising MinWS does not help much in terms of fairness. Scenario 2 is performing worse than 22 Scenario 1 in the overall, regardless of the objective function used within the framework in terms of MinWS, yet it is much better at producing fair rosters as compared to Scenario 1. Instance MinWS Scenario 2 Emergency i 23245 340.60% Geriatric i 3799 501.93% Psychiatry i 29516 311.68% Reception i 66719 28.60% Emergency d 17477 456.62% Geriatric d 38835 67.49% Psychiatry d 11591 601.16% Reception d 19018 177.32% Table 7: The average %gap over 20 run for a given objective function mea- sured from the basis (first column of Table 5) to Scenario 2 for each instance. 6. Conclusion This paper describes a flexible generic agent-based cooperative search framework to build fair nurse rosters. The cooperative search framework combines the strength of multiple metaheuristics and several fairness ob- jective functions to generate high quality rosters. The agents improve the quality of local rosters using the same or different metaheuristic solution methods with the same or different fairness objective functions and param- eter settings. The agents cooperate asynchronously using pattern matching and reinforcement learning in order to generate fair nurse rosters. The experiments were conducted to evaluate the effectiveness of using co- operation to generate fair nurse rosters using existing real world benchmark problems. The experimental results demonstrated the success of cooperative search in producing fairer rosters. Furthermore, when all agents cooperate using the same fairness objective function, the function minimising the devi- ation from the average roster turned out to be superior, while the weighted sum objective function produced the least fair rosters. This observation con- firms the results obtained by Smet et al. (2013b). Finally, cooperative search with combined fairness models produced the fairest nurse rosters. Supposedly, this phenomenon can be ascribed to the fact that cooperation of the different fairness models introduced an element of efficiency. This could be in line with the strength of human coopera- tion, where the cooperation of nurses sharing different impressions of fair 23 nurse rosters can lead to better rosters in terms of fairness. The results have confirmed that the cooperation of multiple decision makers who do not necessarily share the same models and the same heuristics produced fairer rosters than optimisation from a single planner’s point of view. Future work will investigate the fairness issues in other problem domains. References G. R. Beddoe and S. Petrovic. Selecting and weighting features using a ge- netic algorithm in a case-based reasoning approach to personnel rostering. European Journal of Operational Research, 10(1):649–671, 2006. F. Bellifemine, F. Bergenti, G. Caire, and A. Poggi. Jadea java agent devel- opment framework. Multi-Agent Programming, pages 125–147, 2005. D. Bertsimas and J. Tsitsiklis. Simulated annealing. Statistical Science, pages 10–15, 1993. B. Bilgin, P. De Causmaecker, B. Rossie, and G. Vanden Berghe. Local search neighbourhoods to deal with a novel nurse rostering model. Annals of Operations Research, 194(1):33–57, 2012. C. Blum and A. Roli. Metaheuristics in combinatorial optimization: Overview and conceptual comparison. ACM Computing Surveys (CSUR), 35(3):268–308, 2003. A. L. Bouthillier and T. G. Crainic. A cooperative parallel meta-heuristic for the vehicle routing problem with time windows. Computers & Operations Research, 32(7):1685–1708, 2005. E. K. Burke, P. De Causmaecker, S. Petrovic, and G. Vanden Berghe. Fitness evaluation for nurse scheduling problems. In Proceedings of the Congress on Evolutionary Computation (CEC2001), pages 1139–1146, Seoul, Korea, May 27-30 2001. IEEE Press. E. K. Burke, P. De Causmaecker, G. Vanden Berghe, and H. Van Landeghem. The state of the art of nurse rostering. Journal of Scheduling, 7(6):441–499, 2004. 24 C. C. B. Cavalcante, C. Carvalho de Souza, M. W. P. Savelsbergh, Y. Wang, and L. A. Wolsey. Scheduling projects with labor constraints. Discrete Applied Mathematics, 112(1):27–52, 2001. S. H. Clearwater, T. Hogg, and B. A. Huberman. Cooperative problem solving. Computation: The Micro and the Macro View, pages 33–70, 1992. T. Crainic and M. Toulouse. Explicit and emergent cooperation schemes for search algorithms. Learning and intelligent optimization, pages 95–109, 2008. T. G. Crainic and M. Gendreau. Cooperative parallel tabu search for capac- itated network design. Journal of Heuristics, 8(6):601–627, 2002. T. G. Crainic, M. Toulouse, and M. Gendreau. Synchronous tabu search parallelization strategies for multicommodity location-allocation with bal- ancing requirements. OR Spectrum, 17(2):113–123, 1995. T. G. Crainic, M. Toulouse, and M. Gendreau. Toward a taxonomy of parallel tabu search heuristics. INFORMS Journal on Computing, 9:61–72, 1997. S. De Jong, K. Tuyls, and K. Verbeeck. Artificial agents learning human fairness. In Proceedings of the 7th international joint conference on Au- tonomous agents and multiagent systems-Volume 2, pages 863–870. In- ternational Foundation for Autonomous Agents and Multiagent Systems, 2008. S. Even, A. Itai, and A. Shamir. On the Complexity of Timetable and Multicommodity Flow Problems. SIAM Journal on Computing, 5(4):691– 703, 1976. F. Glover. Tabu search: a tutorial. Interfaces, pages 74–94, 1990. T. R. Gruber. A translation approach to portable ontology specifications. Knowledge acquisition, 5(2):199–220, 1993. S. Haspeslagh, P. De Causmaecker, and G. Vanden Berghe. A multi-agent system handling personnel shortages in hospitals. In Proceedings of the 4th Multidisciplinary International Conference on Scheduling: Theory and Applications (MISTA 2009), MISTA, pages 693–695, Dublin, August 2009. 25 T. Hogg and C. P. Williams. Solving the really hard problems with coop- erative search. In Proceedings Of The National Conference On Artificial Intelligence, pages 231–231, 1993. R. Jain, Dah-Ming Chiu, and William R. Hawe. A quantitative measure of fairness and discrimination for resource allocation in shared computer system. Eastern Research Laboratory, Digital Equipment Corporation, 1984. T. James, C. Rego, and F. Glover. A cooperative parallel tabu search algo- rithm for the quadratic assignment problem. European Journal of Opera- tional Research, 195(3):810–826, 2009. D. Julian, M. Chiang, D. O’Neill, and S. Boyd. Qos and fairness constrained convex optimization of resource allocation for wireless cellular and ad hoc networks. In INFOCOM 2002. Twenty-First Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, volume 2, pages 477–486. IEEE, 2002. R. M. Karp. Reducibility Among Combinatorial Problems. In R. E. Miller and J. W. Thatcher, editors, Complexity of Computer Computations, pages 85–103. Plenum Press, 1972. S. Martin, D. Ouelhadj, P. Beullens, and E. Özcan. A generic agent-based framework for cooperative search using pattern matching and reinforce- ment learning. Technical report, University of Portsmouth, January 2012. M. Muhlenthaler and R. Wanka. Fairness in academic course timetablling. In Practice and Theory of Automated Timetabling (PATAT), pages 114–130, 2012. D. Ouelhadj and S. Petrovic. A cooperative hyper-heuristic search frame- work. Journal of Heuristics, 16(6):835–857, 2010. E. Özcan. Memetic algorithms for nurse rostering. In Proceedings of the 20th international conference on Computer and Information Sciences, IS- CIS’05, pages 482–492. Springer-Verlag, 2005. ISBN 3-540-29414-7, 978- 3-540-29414-6. doi: 10.1007/11569596 51. URL http://dx.doi.org/10. 1007/11569596_51. 26 E Özcan. Memes, self-generation and nurse rostering. In EdmundK. Burke and Hana Rudov, editors, Practice and Theory of Automated Timetabling VI, volume 3867 of Lecture Notes in Computer Science, pages 85–104. Springer Berlin Heidelberg, 2007. ISBN 978-3-540-77344-3. doi: 10. 1007/978- 3- 540- 77345- 0 6. URL http://dx.doi.org/10.1007/978-_ 3-_540-_77345-_0_6. S. Petrovic and G. Vanden Berghe. A comparison of two approaches to nurse rostering. Annals of Operations Research, 2012. P. Smet, P. De Causmaecker, B. Bilgin, and G. Vanden Berghe. Nurse ros- tering: a complex example of personnel scheduling with perspectives. In Automated Scheduling: Real World Case Studies. Springer, 2013a. P. Smet, S. Martin, D. Ouelhadj, E. Ozcan, and G. Vanden Berghe. Fairness in nurse rostering. Technical report, KU Leuven - KAHO Sint-Lieven and University of Portsmouth, 2013b. M. Toulouse, K. Thulasiraman, and F. Glover. Multi-level cooperative search: A new paradigm for combinatorial optimization and an application to graph partitioning. Euro-Par’99 Parallel Processing, pages 533–542, 1999. E. Vallada and R. Ruiz. Cooperative metaheuristics for the permutation flowshop scheduling problem. European Journal of Operational Research, 193(2):365–376, 2009. Sangsuree Vasupongayya and Su-Hui Chiang. On job fairness in non- preemptive parallel job scheduling. Parallel and Distributed Computing and Systems (PDCS), (17), 2005. Z.G. Wang and C. Wang. Automating nurse self-rostering: A multiagent systems model. In 2009 IEEE International Conference on Systems, Man and Cybernetics, volume 1–9 of SMC 2009, pages 4422–4425, 2009. 27 
work_ulmqtopujvdqpmx3hxgjsvqirm	----	The Development of a Bilingual Fuzzy Expert System Shell J. King Saud University, Vol. 21, Comp. & Info. Sci, pp. 27-43, Riyadh (2009/1430H.) The Development of a Bilingual Fuzzy Expert System Shell Hassan Mathkour, Israa Al-Turaiki and Ameur Touir Department of Computer Science King Saud University Riyadh, Saudi Arabia mathkour@ccis.ksu.edu.sa, binmathkour@yahoo.com, touir@ksu.edu.sa (Received 21/12/2008; accepted for publication 02/03/2009) Abstract. Fuzzy logic has been incorporated in many expert systems to solve real world problems that are inherently ambiguous. With fuzzy logic it is possible to program human intuition through the development of fuzzy expert system shells. A fuzzy expert system shell is a tool that helps build expert systems to manage fuzzy problems. Commercial as well as non-commercial fuzzy expert system shells are available. These shells provide variety of functions to facilitate the development of fuzzy expert systems for real world problems in different application areas such as medicine, engineering, and finance. To the best of our knowledge, none of the available fuzzy shells is natively developed for the Arabic language. This paper describes the development and the experimentation of a bilingual fuzzy expert system shell. This shell is intended to be a research tool for fuzzy expert systems developers in bilingual environments similar to those in the Arab world where users and developers use multi-languages due to their educational backgrounds and working environments. The shell processes fuzzy terms of the Arabic language as well as the English language. The shell is a general purpose shell that provides users with the ability to develop Arabic/English fuzzy expert systems using a simple Graphical User Interface. It applies implication methods that bear resemblance to human intuition. In the process of the development, a comparison of various fuzzy expert system shells has been performed to identify strengths and weaknesses of available shells. Experiments with our shell are reported and its performance is compared to existing shells that use different implication methods. Keywords: Expert system shells, knowledge based systems, Fuzzy Implication methods, Bilingual Systems. 1. Introduction Expert system shells are versatile tools that are used to create expert systems. Fuzzy expert system shells have been developed to allow for reasoning that deals with crisp and fuzzy sets. These shells allow incorporation and manipulation of imprecise information using fuzzy set theory developed by Zadeh (Zadeh, 1965). They are used to create expert systems that can handle imprecise situations effectively. The ability to operate under imprecise environment makes expert systems closely behave like human being and provides a natural representation of people's daily terminologies. The ability of treating ambiguities, in a manner similar to human experts, makes expert systems versatile and adaptable to unforeseen circumstances which are difficult to avoid in real life applications. This has made fuzzy logic a suitable means to deal with the fuzziness of data and knowledge frequently encountered in the terminologies of human experts when developing knowledge based systems (Kelmet and Slany, 1993). There have been attempts to design fuzzy expert system shells for large-scale general-purpose as well as domain specific applications (Philip, 1991; Aly and Vrana, 2006). Over the years, a large number of expert system shells have been developed and several of them are commercially available. JFK (López-Ortega, 2006), FuzzyShell (Pan, 1996), FuzzyJess (Orchard, 2001), FuzzyCLIPS (Orchard, 2004), FLINT (Shalfield, 2005), FLOPS (Siler and Buckley, 2005), Fuzzy Logic (Mathworks, 1999), and FuzzyJ toolkit (Council, 2001; Orchard, 2001) are examples of expert system shells. We have analyzed several of the existing shells in an attempt to indentify a shell having features that natively supports application development in Arabic language while allowing for application development in other languages. We searched for a shell that accommodates for Arabic fuzzy terms naturally and which employ intuitional inference methods. Our unsuccessful endeavor and realizing that making such a shell available will be useful for bilingual developers and users in research and educational environments motivated us to design and implement a fuzzy shell with Arabic/English support. 27 28 Mathkour et al.: The Development of a Bilingual Fuzzy Expert System Shell In the process, we have found it helpful to furnish a comparison for a set of the available fuzzy shells. These shells differ from one another in several aspects. For example, most of the shells implement inference methods that are mentioned in (Zadeh, 1975; Mamdani, 1977) while many (Fukami, 1980; Mizumoto, 1981; Mizumoto, 1982) have advocated that the methods that are based on the interpretation given in (Mizumoto et al., 1979; Mizumoto et al., 1979; Mizumoto et al., 1979) perform better as they induce human intuition. In this research, we have taken the interpretation that is supported in (Mizumoto et al., 1979; Mizumoto et al., 1979; Mizumoto et al., 1979; Mizumoto, 1981). Our work in (Mathkour et al., 2009) introduces an Arabized fuzzy expert system shell. In this paper, we present the development of a bilingual (Arabic/English) fuzzy expert system shell, which is an extension of our work in (Mathkour et al., 2009), to allow for both the Arabic and English languages. We also report on experiments with the shell using real life data to demonstrate and analyze its human-like behavior using the selected inference methods. To measure its effectiveness, we have compared its performance with some of the available shells. We report on the experiments and comparison of our shell with FuzzyClips (Orchard, 1996) and FuzzyJ (Council, 2001; Orchard, 2001). The objective of our extended shell is to provide a comprehensive tool that is intended to be a research tool for fuzzy expert systems developers in multi-lingual environments similar to those in the Arab world where users and developers use multi- languages due to their educational backgrounds and working environments. It is a general purpose shell that is based on the implication methods: Rs, Rg, Rgs, Rgg, Rsg and Rss (Fukami, 1980; Siler and Buckley, 2005; Mizumoto et al., 1979; Mizumoto et al., 1979; Mizumoto, 1981). It is also observed that many shells use dedicated programming languages for the expert system application development. Consequently, application developers are required to learn the programming languages that are supported by these shells. Learning a new programming language is not a desired requisite, especially for those who do not have a programming aptitude. Learning a new programming language distracts developers from their main objective of developing expert systems in their specific domains. In our shell, we have used a visual environment by adopting a simple graphical user interface. The interface supports both Arabic as well as English languages and it can be tailored for other languages by adding the user interface support for the required language. In (Mathkour et al., 2009), we developed comparison criteria to evaluate aspects of available expert system shells. The criteria include evaluation of end-user interface, developer Interface and availability and installation of shell. In this paper, we further discuss these criteria and employ them to formulate comparison tables of a larger number of existing expert system shells. The rest of the paper is organized as follows: Section 2 presents a comparison of twenty expert system shells along with brief description of the comparison criteria. Section 3 presents the developed fuzzy shell, describes the implication methods, and the implementation. Section 4 presents experimentation with the system. Section 5 presents a comparison of our shell to some existing ones. Section 6 concludes the paper. 2. Comparison of Existing Shells We have endeavored to compare the features of twenty shells of those available commercially and otherwise. These include Fuzzy Logic(Mathworks, 1999), JFK (López-Ortega, 2006), FuzzyJess (Orchard, 2001), FuzzyCLIPS (Orchard, 2004), FuzzyShell (Pan, 1996), FLINT (Shalfield, 2005), FLOPS (Siler and Buckley, 2005), CLIPS (Giarratano, 1998), Jess (Friedmann-Hill, 1999), Flex(Vasey, 1996), PSS (Forgy, 1981), ESB (Kent and Denholm, 1990), ESBuilder (Ishihara et al, 1995), and FuzzyJ toolkit (Council, 2001; Orchard, 2001). First we present a discussion of the comparison criteria, then present the results of our comparison in Table 8. 2.1 End-user interface The user interface is an important component of any software development tool as it allows interaction between application developers and the tool. The user interface must be natural in the context applications that are being developed thereby releasing the developers from learning extraneous concepts and focusing on the development issues. J. King Saud University, Vol. 21, Comp. & Info. Sci, Riyadh (2009/1430H.) 29 The quality of the user interface is judged by its ease of use and naturalness. The following features are indicators of the quality of an interface: 1. Explanation facilities: This is used to explain the process through which the system has arrived at a decision. 2. User friendliness: This is judged by the quality of graphical user interface components such as menus, buttons, and usage of a natural language. 3. The ability to change the earlier answers without having to repeat the session from the beginning. 2.2 Developer interface The expert system developers enter their knowledge through the rule editor. The rule editor should support the rule type selection and creation, rule change and update process, mathematical operations to implement the inference engine strategies, built-in member functions, de- fuzzification methods, certainty factor handling, error correction, and fact refinement and documentation. In addition to these, the rule editor must have provisions to interact with external environments like DBMSs, Spread sheets and Programming in modern languages like Java and C#. Features related to the rule editor are shown in Table 1 to Table 5 with their respective weights. 2.3 Procurement and installation The availability of these tools could be problematic in some linguistic regions of the world. Once available, their installation is not always straight forward. Hence we have used it as an evaluation factor. Table 6 and 7 shows the weight assigned to measure the ease of procurement and installation. Table 1. Rule type weight Rule Type Weight Complex IF-THEN-ELSE rule 5 Complex IF-THEN rule (multiple antecedents or/ and multiple consequents) 3 Simple IF-THEN rule (one antecedent one consequent) 1 Table 2. Rule chaining weight Method Symbol Forward F Backward B No built in chaining strategy (user defined) NA Table 3. Math capability weight Supported Math Functions Weight Advanced math functions 5 Basic math functions 3 None 1 Table 4. Inference strategy weight Supported Inference Strategies Weight None 1 One or Two 3 Three 5 Table 5. Documentation weight Documentation Weight Comprehensive & easy to read 5 Brief 1 Table 6. Procurement weight Procurement Method Weight Download from the Internet 5 Order package CD 1 Table 7. Installation weight Installation Method Weight Unpack (run) one file 5 Unpack source and compile 1 3. The Proposed Bilingual Fuzzy Expert System Shell The entry point to the system provides the users with the option of building expert systems using Arabic or English knowledge bases (Fig. 1). Upon selection, Arabic or an English screen portraying the main components of system is displayed (Fig. 2.a and 2.b). The main components of the shell are the variable editor, rule editor, and the inference engine. 3.1 The variable editor The variable editor’s main purpose is to provide functions to create, edit, and delete fuzzy variables, their fuzzy values, membership functions, and universe of discourse. The layout of our variable editor is shown in Fig. 3.a and 3.b. The variable editor can be launched from the menu button of “Variable Editor” “محرر المتغيرات” in Fig. 3. Created variables and their properties can be seen from a dropdown menu. 30 Mathkour et al.: The Development of a Bilingual Fuzzy Expert System Shell ; a - - • - 1-' - - - - - - - 1- - - - "a - - . -1- • - - - - - - - 1- - - - ~ - - - - - - - - - - - - - 1- - - - ~ - - - - - - - - - - - - - 1- - - - :- ........ - - 1- ;; .... - ~ _.... - - I........ - ..... ~- - - - - 1- - - - - - - - - 1- - - - .~ - - - - -1- ' - - - - - - - 1- - - - .! ......... - - I ..... ;; - - - -.... .... ..... I....... .... ..... J - - - - 1- ;; - - - - - - - 1- - - - ~... ..... - - - I ..... ~ ...... - - - .... - I......... .... .... •• - - - - - - - - - - - - - 1- - - - J! --:Ii - - ..... - I..... ... ... ... ... ... - ..... - I.... - .... .... 0- - - - - 1- - - - - - - - - 1- - - - ~ - - - - 1- - - - - - - - 1- - - - ~" - - - - 1- - - - - - - - - 1- - - - ~.ti - - - -1- - - - - - - - - 1- - - - ~ - - - - 1- · - - - - - - - 1- - - - ~ - - - - 1-0 1- - - - - - - 1- - - - LR - - - - !:..:L - - - - - - = != : : : ~ - - - ..... I-:.if-....... ..... .......... ..... 1-' ~. ~ tt· ~ !!j~hJ!,,-=i Hi ---'=-----1 "'''''' ....... " _ _ '<I.O;i J. King Saud University, Vol. 21, Comp. & Info. Sci, Riyadh (2009/1430H.) 31 Fig. 1. System entry point. (a) (b) Fig. 2. The arabic and english components of the system. 3.2 Rules and Rule-base Editor For The rule-base editor permits application developers to create, edit and delete rules. The rules are of the form IF antecedents Then consequents. A rule may have more than one antecedent and one consequent. Also, the editor allows the “Else part” in the rules. We use rules as our knowledge representation scheme because they are natural in representing expert knowledge, and they are easier to understand, modify, and maintain. The rule editor can be launched from the menu entitled “KB Editor” “ المعرفةمحرر قواعد ” on the menu bar of Fig. 2. Figure 4 shows the layout of the Knowledge base editor at the creation of a new knowledge base. A new knowledge base is created using “New KB” “ادخال قاعدة معرفة “ in the menu bar of Fig. 2. (a) (b) Fig. 3. Arabic//english variable editor layout. (a) 32 Mathkour et al.: The Development of a Bilingual Fuzzy Expert System Shell (b) Fig. 4. The arabic/english layout of the knowledge base editor. 3.3 Inference rngine The inference engine uses implication, composition, aggregation and linkage as given in (Leung and Lam, 1988; Aly and Vrana, 2006; Bandler and Kohout, 1980), and briefly described in the following subsections. It is the part of the knowledge based system that is responsible for deriving conclusions from existing data, i.e., deriving new knowledge from existing ones. 3.3.1 The implication methods Our shell has a backward inference engine and uses the implication methods discussed in (Fukami et al. 1980; Mizumoto et al., 1979; Mizumoto et al., 1979; Mizumoto et al., 1979, Mizumoto, 1981), namely, Rs, Rg, Rgs, Rgg,Rsg, and Rss. Details of the inference methods are found in (Mizumoto et al., 1979; Mizumoto et al., 1979). The choice of the inference methods is based on the observation that such methods closely mirror the human intuitions as compared to those in (Zadeh, 1999; Zadeh, 2006; Zadeh, 1975; Mamdani, 1977). This has been advocated in previous work (Mizumoto, 1981; Mizumoto, 1982). The shell allows the user to either use all the implication methods or select one of them. Figure 5 shows a screen shot of the working of the system when conclusion is obtained using the Rs Implication method. The conflict resolution strategy used in our shell is the most specific strategy. A fuzzy inference method needs to satisfy the criteria shown in Table 9, in order to resemble human intuition (Fukami et al, 1980; Mizumoto, 1981; Mizumoto, 1982). The inference methods presented in (Zadeh, 1994) do not satisfy the criteria in Table 9, except CriterionIV-1. The inference methods in (Mamdani, 1977) on the other hand satisfy Criterion I and II-2. Criterion II-2 is applicable when there is no strong relation between “x is A” and “y is B”. In criterion IV-1, information about y cannot be inferred from the conditional inference “if x is A then y is B” when “x is not A”. Details of related issues are found in (Bandler and. Kohout, 1980; Willmott, 1980; Mizumoto, 1981, Mizumoto, 1982). (a) (b) Fig. 5. Conclusion obtained using Rs implication. Following the criteria in Table 9, fuzzy inferences can be classified into the following four types. Illustration of criteria that are satisfied by the implication methods Rs, Rg, Rgs, Rgg, Rsg, and Rss is given in Table 10 (Fukami, 1980; Mizumoto, 1981; Mizumoto, 1982). Type 1: The binary relation between the antecedent A and the consequence B is translated into ),( BARs . In type 1 inference, Criteria I, II-1, III, IV-1 are satisfied. J. King Saud University, Vol. 21, Comp. & Info. Sci, Riyadh (2009/1430H.) 33 Table 9. Fuzzy inference criteria Criterion I Ant 1: if x is A then y is B An t2: x is A Cons: y is B Criterion II-1 Ant 1: if x is A then y is B Ant 2: x is very A Cons: y is very B Criterion II-2 Ant 1: if x is A then y is B Ant 2: x is very A Cons: y is B Criterion III Ant 1: if x is A then y is B Ant 2: x is more or less A Cons: y is more or less B Criterion IV-1 Ant 1: if x is A then y is B Ant 2: x is not A Cons: y is unknown Criterion IV-2 Ant 1: if x is A then y is B Ant 2: x is not A Cons: y is not B Type 2: The binary relation between the antecedent A and the consequence B is translated into ),( BARg . In type 2 inference, Criteria I, II-2, III, IV-1 are satisfied. Type 3: The binary relation between the antecedent A and the consequence B is translated into ),( BARsg . In type 3 inference, Criteria I, II-1, III, IV-2 are satisfied. Type 4: The binary relation between the antecedent A and the consequence B is translated into ),( BARgg . In type 4 inference, Criteria I, II-2, III, IV-2 are satisfied. An Rs implication example: For the rule, If x is small then y is middle, where U=V= 0+1+2+3+4+5+7+8+9+10, A=small= 1/0+0.8/1+0.6/2+0.4/3+0.2/4, and B= middle =0.2/2+0.4/3+0.8/4+1/5+0.8/6+0.4/7+0.2/8, Rs(A,B) is given in the Table 11. 3.3.2 Composition A fuzzy composition relation R(A,B) of R1 and R2 is simply the relation obtained by applying R1 and R2 one after the other. The most frequently used Table 10. Criteria satisfied by each implication method Ant 2 Cons Rs Rg Rgs Rgg Rsg Rss A B + + + + + + Very A Very B + - - - + + Very A B - + + + - - More or less A More or less B + + + + + + Not A Not B - - + + + + Table 11. Rs(A,B) U V 0 1 2 3 4 5 6 7 8 9 10 0 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 1 1 1 0 0 0 0 2 0 0 0 0 1 1 1 0 0 0 0 3 0 0 0 1 1 1 1 1 0 0 0 4 0 0 1 1 1 1 1 1 1 0 0 5 1 1 1 1 1 1 1 1 1 1 1 6 1 1 1 1 1 1 1 1 1 1 1 7 1 1 1 1 1 1 1 1 1 1 1 8 1 1 1 1 1 1 1 1 1 1 1 9 1 1 1 1 1 1 1 1 1 1 1 10 1 1 1 1 1 1 1 1 1 1 1 34 Mathkour et al.: The Development of a Bilingual Fuzzy Expert System Shell composition operator in fuzzy logic is the Max-Min composition operator in (Zadeh, 1999; Zadeh, 1975) and it is the one we used in our shell. Max-Min Composition Let R be a fuzzy relation in X × Y, and S be a fuzzy relation in Y × Z. The Max-Min composition of R and S, RoS, is a fuzzy relation in X × Z such that RoS↔μRoS(x,z) = v {μR(x,y) ^ μS(y,z) } A Max-Min composition example: Suppose we have the following two relations R and S: 6.00 11.0 8.05.0 1.019.0 06.04.0 SR max{min(0.4,0.5), min(0.6, 0.1), min(0, 0)} = max{ 0.4, 0.1, 0} = 0.4 max{min(0.4,0.8), min(0.6, 1), min(0, 0.6)}= max{ 0.4, 0.6, 0} = 0.6 max{min(0.9,0.5), min(1, 0.1), min(0.1, 0)}= max{ 0.5, 0.1, 0} = 0.5 max{min(0.9,0.8), min(1, 1), min(0.1, 0.6)}= max{ 0.8, 1, 0.1} = 1 3.3.3The Aggregation and link operators An aggregation operator is needed when a rule has k conditions. The rule is decomposed into k implications. Each implication is used separately to infer a value by applying the fuzzy implication The values are then aggregated using the aggregation operators used in the rule including OR and AND. The final result is obtained after a MAX operation over the corresponding values inferred by all the rules or fuzzy membership functions (Zadeh, 1975; Fukami, 1980; Mizumoto, 1981; Mizumoto1982; Zadeh, 1999). 3.4 Implementation Issues Similar to that in (Mathkour et al., 2009), the main data structures used in the implementation of the shell are arrays. Since all the implication methods used here depend on the Rs and Rg operations, there was a need to implement Rs and Rg operations as separate methods. Both methods accept two parameters and return the result after performing Rs or Rg implication. Each of the six implications was implemented as methods that accept two matrices and return the result after performing the implication of the whole matrices. 4. Experimental Results The purpose of this section is to illustrate that our fuzzy expert system shell produces correct results and the implication methods used satisfy the criteria mentioned in Table 9. The knowledge base used in this experiment is taken form (Ganoud, et al. 2005). In (Ganoud, et al. 2005), the authors study the influence of random factors on the planning of building works. Deciding the exact period of building projects is a very difficult job due to the fact that building projects are affected by different unpredictable factors. The random factors which are studied include three factors: the cessation of machines, the absence of some professionals, and the influence of weather condition. The rules are given in Table 12 and the membership functions are given in Figures 6 to 9 (Ganoud, et al. 2005). Fig. 6. The membership function of "absence of professionals". Fig. 7. The membership function of "weather changes". Fig. 8. The membership function of machine cessation". R y1 y2 y3 x1 0.4 0.6 0 x2 0.9 1 0.1 S z1 z2 y1 0.5 0.8 y2 0.1 1 Y3 0 0.6 J. King Saud University, Vol. 21, Comp. & Info. Sci, Riyadh (2009/1430H.) 35 Fig. 9. The membership function of the conclusion "increase in the period of building". The knowledge base consists of 27 fuzzy rules. All the rules have the same number of antecedents. The same fuzzy variables are used in all the 27 rules as well as the same conclusion (target). The shell was run several times, each time with different observation. The observations are: (a) (b) Fig. 10. Result of observation 1. Observation 1: X is low, Y is high and Z is high. It is observed (as illustrated in Figure 10) that using Rg,Rgs and Rgg has given the expected results according to the Table 9. On the other hand, using the implications Rs, Rsg, and Rss we found out upon examining the resulting matrices that the results were not exact as expected but were very close to what was expected. Table 12. The rules data used in the experiment (Ganoud, et al. 2005) القاعدة نسبة التعطل الناجم عن تعطل اآلليات نسبة التعطل الناجم عن غياب العمال نسبة التعطل الناجم عن الظروف الجوية النتيجة عالية عالية عالية منخفضة ١ عالية عالية عالية متوسطة ٢ عالية عالية عالية عالية ٣ متوسطة عالية متوسطة منخفضة ٤ متوسطة متوسطة متوسطة منخفضة ٥ منخفضة منخفضة منخفضة منخفضة ٦ ً متوسطة متوسطة متوسطة متوسطة ٧ منخفضة منخفضة منخفضة متوسطة ٨ متوسطة عالية منخفضة متوسطة ٩ عالية عالية متوسطة متوسطة ١٠ عالية متوسطة عالية متوسطة ١١ متوسطة متوسطة منخفضة متوسطة ١٢ متوسطة منخفضة عالية متوسطة ١٣ منخفضة منخفضة متوسطة متوسطة ١٤ منخفضة متوسطة منخفضة منخفضة ١٥ ً متوسطة عالية منخفضة عالية ١٦ منخفضة منخفضة متوسطة منخفضة ١٧ منخفضة متوسطة منخفضة منخفضة ١٨ ً عالية متوسطة عالية عالية ١٩ متوسطة متوسطة عالية منخفضة ٢٠ متوسطة منخفضة عالية منخفضة ٢١ عالية منخفضة عالية عالية ٢٢ عالية عالية متوسطة عالية ٢٣ متوسطة عالية منخفضة عالية ٢٤ متوسطة منخفضة متوسطة عالية ٢٥ متوسطة متوسطة متوسطة عالية ٢٦ منخفضة منخفضة منخفضة عالية ٢٧ (a) 36 Mathkour et al.: The Development of a Bilingual Fuzzy Expert System Shell (b) Fig. 11. Result of observation 2. Observation 2: X is not low, Y is not high and Z is not high. It is observed (as illustrated in Figure 11) that all the implications has given the expected results according to the Table 9 except for Rgs and Rss. Examining the matrices that resulted from using Rgs and Rss we have found that they were very much close to the matrix of the expected result which is "not high". The result in words should have been “more or less not high", but because the shell is not designed to handle composite hedges the result matrix was translated to "more or less low" Observation 3: X is very low, Y is very high and Z is very high. It is observed (as illustrated in Figure 12) that the implications Rg, Rgs and Rgg have given the expected results according to Table 9. On the other hand, the implications Rs ,Rsg and Rss have given different results than expected. Upon examining the matrices that resulted from using Rs, Rsg and Rss we found that they were very much close to the matrix of the expected result which is "very high" . The result in words should have been " more or less very high" ,but because the shell is not designed to handle composite hedges the result matrix was translated to "more or less high". (a) (b) Fig. 12. Result of observation 3. Observation 4: X is more or less low, Y is more or less high and Z is more or less high. It is observed (as illustrated in Figure 13) that the implications Rs, Rsg, and Rss all have given the expected results according to Table 9. But for the rest of the implication methods the results are close to the expected. (a) (b) Fig. 13. Result of observation 4. 5. Comparison with Existing Tools We demonstrate that the proposed fuzzy shell performs in a natural manner that mirrors the human inferences of real world problems and yields the expected conclusions that conform to human possible conclusions as compared to other tools. The comparison is made with tools that use the inference methods in (Mamdani, 1977) which are commonly J. King Saud University, Vol. 21, Comp. & Info. Sci, Riyadh (2009/1430H.) 37 used in commercial fuzzy expert system shells such as FuzzyClips, and with tools such as FuzzyJ Toolkit that allows for different inference methods including those discussed in (Aly and Vrana, 1977; Mamdani, 1977; Mizumoto, et all, 1979). We examine the performance of both FuzzyClips and FuzzyJ Toolkit shells and compare their results with that of our shell using the four observations detailed in Section 4. For this purpose, the English-translation of the data in Table 12 is used. 5.1 A comparison with fuzzyCLIPS Mamdani's inference methods are the most commonly used in commercial fuzzy expert system shells. It has been observed that inference methods in (Mamdani, 1977) do not satisfy human intuition (Fukami et al. 1980; Mizumoto, 1982). We demonstrate this observation by examining the behavior of FuzzyCLIPS with the rules and fuzzy variables of discussed in Section 4. Table 12. The only criteria satisfied by mamdani's methods Criterion I Ant1: IF x is A Then y is B Ant2: x is A Cons: y is B Criterion II-2 Ant1: IF x is A Then y is B Ant2: x is very A Cons: y is B The definitions of the fuzzy variables and the fuzzy rule using FuzzyCLIPS are shown in Figures 14 to 18 below: (deftemplate x ;definition of fuzzy variable ‘Machine cessation’ 15 45 ;Universe of Discourse ((low (16 0.1) (17 0.3) (18 0.5) (19 1) (20 0.6) (21 0.5) ( 22 0.3) (23 0.1)) (medium (21 0.1) (22 0.2) (23 0.4) (24 0.5) (25 0.6) (26 0.7) ( 27 0.6) (28 0.5) (29 0.4)(30 0.2)(31 1)) (high (31 0.1) (32 0.2) (33 0.3) (34 0.4) (35 0.5) (36 0.6) ( 37 0.7) (38 0.1) (39 0.7)(40 0.6)(41 0.5) (42 0.4) (43 0.3) (44 0.2) (45 0.1)) ) ) Fig. 14. Variable x definition using fuzzyCLIPS (deftemplate y ;definition of fuzzy variable ‘Absence of professionals’ 4 16 ;Universe of Discourse ( (low (5 0.2) (6 1) (7 0.2) ) (medium (7 0.4) (8 1) (9 1) (10 1) (11 1) (12 1)) (high (11 0.2) (12 0.5) (13 1) (14 0.5) (15 0.2)) ) ) Fig. 15. Variable y definition using fuzzyCLIPS. (deftemplate z ;definition of fuzzy variable ‘Weather changes’ 10 90 ;Universe of Discourse ( (low (10 0.2) (20 0.6) (30 1) (40 0.2)) (medium (40 0.7) (50 1) (60 0.7) ) (high (70 0.7) (80 1) ) )) Fig. 16. Variable z definition using fuzzyCLIPS. (deftemplate cons ;definition of fuzzy variable ‘Conclusion’ 20 100 ;Universe of Discourse ( (low (25 0.2) (30 0.5) (35 0.7) (40 1) (45 0.7) (50 0.5) (55 0.2) ) (medium (55 0.5) (65 1) (70 0.6) (75 0.5) ) (high (80 0.2) (85 1) (90 0.5) (95 0.2) ) ) ) Fig. 17. Conclusion definition using fuzzyCLIPS. (defrule r1 ; a rule that matches and asserts fuzzy facts (x low) (y high) (z high) => (assert (cons high) ) ) Fig. 18. Fuzzy rule definition using fuzzyCLIPS. 38 Mathkour et al.: The Development of a Bilingual Fuzzy Expert System Shell FuzzyCLIPS was run several times with the same observations in Section 4 and the results are as follows (Figures 19 to 22): Observation 1 : X is Low , Y is High and Z is high. FuzzyCLIPS gives the expected result according to Criteria I in Table 9. This is natural and expected as all observations match all antecedents. Fig. 19. FuzzyCLIPS result for Observation 1. Observation 2: X is not low, Y is not high and Z is not high. When the antecedents contain the NOT hedge, FuzzyCLIPS yields a fuzzy set that cannot be mapped to a linguistic expression. This is expected as Mamdani's methods do not satisfy Criterion IV-1 and Criterion IV-2 of Table 9. Fig. 20. FuzzyCLIPS result for observation 2. Observation 3: X is very low, Y is very high and Z is very high. FuzzyCLIPS gives the expected result according to Criteria II-2 in Table 9. Fig. 21. FuzzyCLIPS result for observation 3. Observation 4: X is more or less low, Y is more or less high and Z is more or less high. The resulting fuzzy set cannot be mapped to a linguistic expression. From the result shown in figure 9 it is clear that Mamdani's methods do not satisfy criterion III. Fig. 22. FuzzyCLIPS result for observation 4. 5.2 Comparison with FuzzyJ Toolkit FuzzyJ Toolkit is a set of Java classes that provide the capability to handle fuzzy concepts and reasoning (Orchard,2001). It allows for different inference methods including those in (Aly and Vrana, 2006; Mamdani, 1977). We examine the behavior of FuzzyJ Toolkit using the rules and fuzzy variables of Section 4. Figures 23 to 27 show the fuzzy variables and fuzzy rule definitions. J. King Saud University, Vol. 21, Comp. & Info. Sci, Riyadh (2009/1430H.) 39 Fig. 23. Definition of fuzzy variable x using fuzzyJ toolkit. Fig. 24. Definition of fuzzy variable y using fuzzyJ foolkit. Fig. 25. Definition of fuzzy variable z using fuzzyJ toolkit. 40 Mathkour et al.: The Development of a Bilingual Fuzzy Expert System Shell FuzzyJ was run several times with the same observations in Section 4 and with the inference method set to Larsen's inference method. The results of the inference are as follows: Observation 1 : X is Low, Y is High and Z is High. The resulting fuzzy set is {0/70, 0.2/75, 0.5/80, 1/85, 0.5/90, 0.2/95, 0/100}. Here the result given by FuzzyJ is "high" which is natural as all the observations match the all the antecedents of the fuzzy rule. Observation 2: X is not low, Y is not high and Z is not high. The resulting fuzzy set is {0/70 ,0.1/75, 0.25/80 ,0.5/85, 0.25/90 ,0.1/95, 0/100}. This result could not be mapped to a linguistic expression although it is rather close to the fuzzy set "high". Observation 3 & Observation 4: The resulting fuzzy set is {0/70, 0.2/75, 0.5/80, 1/85, 0.5/90 ,0.2/95, 0/100}. Here the result given by FuzzyJ is "high". Notice that this is the same result when no hedges were used. It is obvious that the use of the hedge "very" and the "more or less" hedge had no effect on the result. In our shell the Fig. 26. Definition of fuzzy variable conclusion using fuzzyJ toolkit. Fig. 27. Definition of the fuzzy rule using fuzzyJ. J. King Saud University, Vol. 21, Comp. & Info. Sci, Riyadh (2009/1430H.) 41 hedges were recognized through the calculation of the implication criteria of Table 9. 6. Conclusion In this paper, we discussed the development of our own bilingual fuzzy expert system shell. In the process, we have examined, evaluated, and compared various fuzzy expert system shells that adopt different inference methods for the sake of identifying desirable features and examining their performance. Our shell was developed using NetBeans 4.1 IDE. It has an Arabic user interface as well as an English user interface. The inference engine is a backward chaining inference engine. It uses the implication methods Rs, Rg, Rss, Rgg, Rgs and Rsg. Several tests have been performed on this shell to ascertain its proper functionality. Some of the tests have given the expected results that reflect human intuitions. Few tests have given results which are very close to the expected outcome. We observe that when the membership function of fuzzy values covers a wide range from 0 to 1, the shell produces more accurate results. Experimental results for our shell have been reported and analyzed. A comparison of the performance of our shell with other shells such as FuzzyCLIPS and FuzzyJ has also been discussed. We are in the process of extending the shell to allow for the processing of fuzzy terms in other natural languages. References Aly, S. and Vrana, I.. " Toward efficient modeling of fuzzy expert systems: a survey" Agric. Econom. - Czech, (2006), 456-460. Bandler, W. and Kohout, L. “Fuzzy power sets and fuzzy implication operators”, Fuzzy Sets and Systems, 4, (1980), 13-30. Council, C. and Orchard, R. “Fuzzy reasoning in jess: The fuzzyj toolkit and fuzzyjess”, The Proceedings of 3rd International Conference on Enterprise Information Systems (ICEIS 2001), Setubal, Portugal, (2001). Forgy, C.L. “The OPSS User Manual”, Technical Report, CMU- CS-81-135, Computer Science Department, Carnegie-Mellon University, Pittsburgh, (1981). Friedmann-Hill, E.J.. “Jess, The Java Expert System Shell”, http://herzberg.ca.sandia.gov/jess/, (1999). Fukami, S.; Mizumoto, M. and Tanaka, K. “Some Considerations on Fuzzy Conditional Inference”, Fuzzy Sets and Systems, 4, (1980), pp.243-273. Ganoud, A.; Ali, H. and Ibrahem, S. "Studying the Influence of random Factors on the Planning of Building Work", Tishreen University Journal for Studies and Scientific research- Engineering Science series, 27(3), (2005). Giarratano, J.C. “The CLIPS User’s Guide”, http://www.twine.com/twine/11lslvj57-19d/clips-expert- system-shell, (1998). Kent, J.R. and Denholm, P. “ESB-96 – A Graphical KBS Development Tool”, in Europal'90, Proceedings of the First European Conference on the Practical Application of Lisp. Dorking, March (1990). Ishihara, S.; Ishihara K.; Nagamachi, M. and Matsubara, Y. “An Automatic Builder for a Kansei Engineering Expert System using self- Organizing Neural Networks”, Int. Journal of Industrial Ergomonics, 15(1), (1995), 13-24. Kelmet, E. and Slany, W. “Fuzzy Logic in Artificial Intelligence”, Proceedings of the 8th Austrian Artificial Intelligence Conference, FLAI '93, Linz, Austria, (1993). Leung, K. S.; and Lam, W. "Fuzzy Concepts in Expert Systems",IEEE Computer, 21(9), (1988), 43-56. López-Ortega, O. " Java Fuzzy Kit (JFK): A shell to build fuzzy inference systems according to the generalized principle of extension", Expert Systems with Applications, (2006). Mamdani, E.H. “Application of fuzzy logic to approximate reasoning using linguistic systems”, IEEE Trans. Comput. 26, (1977), 1182-1191. Mathkour, H.; Al-Turaiki, I. and Touir, A. “The Development of an Arabized Fuzzy Expert System Shell” in int. conf on Information Management and Engineering ICIME, (2009). Mathworks, Inc. Fuzzy Logic Toolbox Use's Guide version 2, (1999). Mizumoto, M.; Fukami, S. and Tanaka, K. “Fuzzy conditional inference and fuzzy inference with fuzzy quantifiers”. in: Proc. of 6th Int. Conf. on Artificial Intelligence, Tokyo, (1979), 589-591. Mizumoto, M.; Fukami, S. and Tanaka, K. “Some methods of fuzzy reasoning”, in: M.M. Gupta et al., ads., Advances in Fuzzy Set Theory and application (North-Holland, Amsterdam, (1979), 117-136. Mizumoto, M. Note on the arithmetic rule by Zadeh for fuzzy conditional inference, Cybernetics and Systems, 12, (1981), 247-306. Mizumoto, M. “Fuzzy inferences under max- composition in the compositional rule of inference”, in: M.M. Gupta et al., ads., Fuzzy Information and Decision Processes, North-Holland, Information Sciences, 27, (1982), 183-209. Mizumoto, M.; Fukami S. and Tanaka, K. “Several methods for fuzzy conditional inference”, in: Proc. of IEEE Conf. on Decision & Control, Florida, (1979), 777-782. Orchard, R. “NRC FuzzyJ Toolkit for the Java™, Platform User’s Guide” [online]. National Research Council of Canada. Available from:http://ai.iit.nrc.ca/IR_public/fuzzy/fuzzyJDocs/index.ht ml, 2001. Orchard, R. "Fuzzy Reasoning in Jess: The FuzzyJ Toolkit and FuzzyJess", Proceedings of the ICEIS 2001, Third International Conference on Enterprise Information Systems, Setubal, Portugal. (2001), 533-542. Orchard, R. FuzzyClips version 6.10d User's Guide,National Research Council of Canada, (2004). Pan, J. FuzzyShell User's Manual,(1996). Philip, G., "A case study in selecting an expert system shell", Journal of Systems Management, (1991). Shalfield, R. FLINT Reference, Logic Programming Associates Ltd.,(2005). Siler, W. and Buckley, J. Fuzzy Expert Systems and Fuzzy Reasoning, John Wiley & Sons, Inc., New Jersey, (2005). Vasey P.; Westwood D. and Johns N. “Flex Expert System Toolkit”, Logic Programmers Associates Ltd. (1996). 42 Mathkour et al.: The Development of a Bilingual Fuzzy Expert System Shell Willmott, R., “Two fuzzier implication operators” in the theory of fuzzy power sets, Fuzzy Sets and Systems, 4, (1980), 31-37. Zadeh, L.A. Calculus of fuzzy restriction, in: L.A. Zadeh et al., Eds. Fuzzy Sets and Their Applications to Cognitive and Decision Processes (Academic Press, New York, (1975) 1-39. Zadeh, L.A. "From Computing with Numbers to Computing with Words− From Manipulation of Measurements to Manipulation of Perceptions", iee transactions on circuits and systems, Fundamental theory and applications, 45(1), JANUARY (1999), 105-119. Zadeh L.A. “Fuzzy sets” information and control 8, (1965), 338- 353. Zadeh, L. A. "Why the Success of Fuzzy Logic is not Paradoxial", IEEE Expert, 9, (1994), 43-45. J. King Saud University, Vol. 21, Comp. & Info. Sci, Riyadh (2009/1430H.) 43 تطوير صدفية ثنائية اللغة لبناء النظم الخبيرة الضبابية عامر عبداهللا طويرو حسن إسماعيل مذكور، إسراء محمد الطريقي قسم علوم احلاسب، كلية علوم احلاسب و املعلومات، العربية السعوديةجامعة امللك سعود، الرياض،اململكة mathkour@ksu.edu.sa )م٢/٣/٢٠٠٩؛ وقبل للنشر يف م٢١/١٢/٢٠٠٨(قدم للنشر يف يستخدم املنطق الضبايب يف بعض النظم اخلبرية حلل مشكالت من احلياة العملية واليت تتسم بالغموض. إذ . ملخص البحث استخدام املنطق الضبايب لربجمة التطبيقات اليت تعتمد على احلدس البشري من خالل بناء صدفيات لتطوير النظم من املمكن اخلبرية الضبابية من أحل التعامل مع املشكالت الضبابية و الغري واضحة. و توفر هذه الصدفيات تشكيلة من الدوال ميكن االت و التطبيقات مثل التطبيقات الطبية، استخدامها لتسهيل تطوير نظم خبرية ضبابية للتعا مل مع مشكالت يف خمتلف ا واهلندسية، واملالية. ال تتوفر حسب علمنا صدفيات برجمة ضبابية، لبناء النظم اخلبرية الضبابية، مطورة يف األصل للتعامل مع لضبابية ثنائية اللغة. اهلدف من هذه الصدفية اللغة العربية. تصف ورقة العمل هذه تطوير وجتربة صدفية لبناء النظم اخلبرية ا الربجمية هو استخدامها كأداة حبثية ملطوري النظم اخلبرية الضبابية يف البيئات ثنائية اللغة كما هو احلال يف العامل العريب حيث مع املصطلحات تكون استخدامات املطورين متعددة اللغات بسبب نظام تعليمهم وبيئة عملهم. تسمح الصدفية بالتعامل الضبابية يف اللغتني العربية واإلجنليزية. وتعترب صدفية برجمة عامة األهداف توفر للمستخدمني القدرة على تطوير نظم خبرية ضبابية عربية وإجنليزية باستخدام واجهة مستخدم مبسطة. وتقوم بتطبيق طرق االقتضاء يف حماكاة للحدس البشري. مت عند ارنة عدة صدفيات لتطويرو برجمة النظم اخلبرية الضبابية لتحديد نقاط القوة والضعف للربجميات املتوفرة. التطوير دراسة و مق كما مت إعداد تقرير عن جتربة و اختبار الصدفية املطورة ومقارنة أدائها مع الصدفيات املتوفرة واليت تستخدم طرقًا اقتضائية خمتلفة. 44 Mathkour et al.: The Development of a Bilingual Fuzzy Expert System Shell 
work_ulpzm6akbfcupnofubkdvzqq5q	----	S@ iv / S ? v 4 -8343% M,,JC- q 6 o 6 a t j - - Design Issues of a Knowledge-based Welding Advisor Stephen D. Kleban Sandia National Laboratories Albuquerque, NM 87 185 sdkleba@sandia.gov Abstract Expert system implementation can take numerous forms ranging from traditional declarative rule-based systems with if-then syntax to imperative programming languages that capture expertise in procedural code. The artificial intelligence community generally thinks of expert systems as rules or rule-bases and an inference engine to process the knowledge. The welding advisor developed at Sandia National Labs and described in this paper, deviates from this by codifying expertise using object representation and methods. Objects allow computer scientists to model the world as humans perceive it giving us a very natural way to encode expert knowledge. The design of the welding advisor, which generates and evaluates solutions, will be compared and contrasted to a traditional rule-based system. Background Welding in the US is a $50B economy which generates upwards of $7B of waste per year due to rework and scrap. Many of the problems associated with welding arise due to the high degree of "trial and error" that typically accompa- nies the deyelopmFnt ,pE weld .sphedules, weld joint designs, and material selection. Smartweld (Mitchider et al. 1996), which includes the weldmg 'advisor, developed at Sandia National Laboratories (SNL), is an attempt to decrease the amount of waste due to failed welds by providing expertise to both novice and expert welders as a check or first order recommendation. Different weld processes and joint designs have different properties, such as depth of metal penetration, heat input, residual stress, distortion, sensitivity to loading and sensitivity to corro- sive environments. Choosing the weld process and joint geometry is the core problem. Given weld requirements, the advisor suggests a weld- ing process, a joint geometry, and a welding schedule. The advisor currently houses knowledge about ten welding processes and forty different joint geometries. It accesses a database of proven historical weld schedules, and selects a schedule based on a similarity measure, comprised of material type, joint thickness, joint geometry, and welding process. Once a weld schedule is selected, the ultimate goal of this system is to translate this schedule into NC format to run the actual welding machinery. Problem Solving Structure The intent of a welding advisor is to advise, not dictate. With this in mind, it is very important not to recommend a single solution but a set of solutions where the user can be actively involved in making the fmal selection. A generate and test approach fulfills these requirements by first intelligently constructing a set of plausible weld process and joint geometry combinations and then evaluating each scenario according to a set of criteria. The final step is to rank the scenarios and present the results to the user for final selection. This system can be implemented many different ways, including a rule-based solution that can be either goal or 7 data driven. Rules are a declarative expression of expertise that are especially useful for problems that are inherently search based like planning or design (Lager & Stubblefield 1993). Control is implicit via the inference engine and the knowledge engineer needs only to represent ' the xomain expertise in rules. Rules can be used for many other types of problems, but may not be the best approach . or even a poor choice when trying to force fit a problem into a rule-based forinat. For example, a problem that has a great deal of structure and sequence has to be modeled in rules with control rules, priority, or an agenda. Even the famous MYCIN expert system (Buchanan & Shortliff 1984), a diagnosis system for meningitis, was criticized by the doctors because the interactive dialog appeared random and seem to not follow any line of reasoning. Implementing the welding advisor in a purely rule-based manner is acceptable but awkward given the inherently sequential nature of the problem. There is significant structure and sequence in the problem that would be wasted if everything was coded in rules and given to an inference engine for search. Structure and sequence could be implemented in rules but that would result in an inefficient and unnatural representation. An acceptable hybrid solution would have been to capture the structure and sequence in an object-oriented mailto:sdkleba@sandia.gov DISCLAIMER Portions of this document mag be illegiile in electronic image products. fmnes are produced from the best a d a b l e original d O r ? u m e l l t I T I ftamework and use small rule sets to encode discrete chunks of knowledge. This hybrid approach exploits the strength of both the rule-based and object-oriented approaches. The rule-based approach offers a natural way to declare knowledge and perform search through the rule space but is clumsy for its data expressiveness and sharing. The object-oriented paradigm is almost just the opposite with excellent data modeling through encapsulation and data sharing through messaging. Welding knowledge could be represented in rules with implicit control with the object-oriented framework providing the global explicit control. The purely object-oriented approach taken by the welding advisor is a result of evolutionary prototyping. We started implementing the generate and test control structure using KappaTM, an object-oriented development environment from Intellicorp, and continued in this environment codifying the welding expertise in methods. Although a hybrid solution is perhaps preferable, the object-oriented approach provided an invaluable solution to the difficult and time-consuming process of knowledge engineering and domain understanding. Rule-Based Approach A rule-based approach could solve this problem by takiig weld requirements as inputs to start the rule chaining and resulting in a ordered set of solutions. A high level rule to start the process would be: if GeneratePlausibleScenarios((IN)WeldRequirements, (0UT)Scenarios) and EvaluateEachScenario((IN)Scenarios, (0UT)EvaluatedScenarios) and RankAndScoreScenarios((IN)EvaluatedScenarios, DisplayResults((IN)RankedAndOrderedScenarios) Done. (0UT)RankedAndOrderedScenarios) and then In this statement we see the sequence coded in a rule. Each clause in turn would start the chaining process in a corresponding rule set. The frst two clauses would have rules that encode actual welding expertise. The last two have nothing to do with search or knowledge representa- tion and would be part of an awkward representation as mentioned above. It also has a large amount of complex knowledge that is handled nicely in a rule-base with opportunistic search. The unwieldy part is in specifying the control structures needed to satisfy the clause, EvaluateEachScenarioO, because that is where the wealth of welding knowledge resides and requires orchestration, sequence, and control to a fair degree. For example, a geometry with a backing bar requires additional analysis that could be implicitly invoked by pattern matching, but in methods, the actual testing for a backing bar is per- formed and the appropriate messages are sent. Object-Oriented Approach Encapsulation, polymorphism, and message passing are the aspects of object-orientation used most in the welding advisor. Objects are a very natural and powerfUl way to model the world (Rossen, 1990) and reasoning processes of human experts. When initially interviewing welding experts, the domain understanding proceeded very quickly because of the natural expressiveness of objects. It very quickly became clear what the high level objects were (JointGeometry, WeldProcess, Material, etc.) and what, at least on a simple level, were the relationships and interactions between these objects. For example, welding processes have preferred materials and joint geometries have preferred welding processes. This knowledge engineering could proceed independently of the design of the generate and test engine because the knowledge is completely separate. The only unanswered questions were what model to use for the computer assistant (best answer, set of answers, ranked and scored set of answers with what-if capability, etc.) and how to generate the solution (purely generative, candidate creation and evaluation, etc.). The goal was to build a system that was highly interactive where designers and engineers could examine in great detail the results of the advisor and perform extensive what-if analysis. A generate and test approach was selected because it typically generates a number of plausible scenarios offering an exploratory environment where the user can examine a wealth of detailed information as to why a scenario either passed or failed. The generate and test approach is palatable for the new user in that it will not absolutely dictate solutions but rather offer advice. As AI researchers are well aware, it is important not to threaten the users but to give them a tool that assists in their work. That is clearly the case with the welding advisor because welding is so complex it is nai’ve to imagine the welding advisor replacing welding engineers. The advisor could help train entry-level engineers, but will probably be most useful in its completeness of analysis of weld designs by reminding experts of some details that may be otherwise overlooked. The control of flow is shown in Figure 1 . Up front criteria for welding are determined and implemented as tests in the advisor. A designer or engineer presents requirements for a given weld and the evaluation object take those requirements, the tests, and any other inputs needed ftom databases, user queries, or deduced knowledge to produce a ranked and scored ordered set of possible weld designs. The wealth of welding knowledge resides in the tests: they use the joint geometry knowledge and weld process knowledge to evaluate a scenario against the given requirements. For what-if analysis, the user is provided with the capability to change the importance or weighting of any of the criteria. Requirements (Generate Scenarios) (Criteria) i-i Data Evaluate (Run all tests against each scenario) n . I ' I I Score&Rank p-1 Modifycriteria (Importance & Weighting) Figure 1. The flow of control of the generate and test approach in the welding advisor. Another advantage of the object-oriented approach is reuse. We abstracted the generate and test methodology into a generic advisor. The welding advisor specialized certain objects and methods to capture the details of the welding domain, but the structure of the generate and test approach was completely inherited. Another project at SNL will use this basic approach and specialize the generic advisor for advice on near-net shape determination (Rivera, 1996) as shown in Figure 2. 0 Requirements Figure 2. Object hierarchy specializing the requirements object of the generate and test approach, where lower level objects add specific data and methods. The Functionality of the Advisor Once the requirements for the weld are specified the user typically spends most of their time manipulating the table of results shown in Figure 3. The rows in the table represent the plausible welding scenarios. They were generated using heuristics for combining welding process and joint geometry configurations that address the design requirements of the weld. The fourth column is the Score awarded each scenario, based on its performance. The performance is assessed by examining the criteria in the form of tests. Each ranking criterion has an "importance" and a "weight" associated with it. "Required" criteria are hard constraints, and "desired criteria" are soft constraints. If a user wishes to omit one of the normal criteria fiom the evaluation func- tion, the "don't care" importance for that criteria is chosen. By default, all criteria are set to "required" importance, as opposed to "desirable" or "don't care" and each has a weight of 50. Both the importance and weight of each cri- terion can be changed as shown in Figure 4 by the user (lower right corner button in the window, Figure 3) to de- velop an understanding of the sensitivity of the weld sce- nario rankings to the importance and weighting of each criterion. This allows the user to rapidly perform trade-off assessments. Other "what-if' analysis is available by changing performance requirements by either changing specification of design or answers to the advisor's queries. The first column of Figure 3, the scenario number, has a color coded background that indicates if the scenario passed all hard or soft constraints (criteria). Global constraints are part of the performance requirements. If a scenario satisfied all constraints (required and desired) then it is awarded a green color. If one or more desired constraints are violated then the scenario is awarded a yelloy color and if a required constraint is violated it is red. The scenarios are sorted and displayed in descending order. The columns to the right of the table depict the individual scores of the tests performed on each scenario. The detailed analysis of each scenario is available by pressing the "E~planation~~ button at the lower portion of the window in Figure 3. The explanation includes details about each of the tests performed explaining how a particular scenario performed against the weld requirements. For example, a scenario's explanation might read, "For a Low Voltage Electron Beam hocess, a smooth underbead requirement, square joint geometry, we recommend a joint thickness between 0.05mm and 3.0mm, not 8.Omm as you have specified." Also provided in the explanation is any anecdotal information not pertaining to tests that may be interesting to the user, such as, "Since we have two sided access, we can satisfy the smooth underbead requirement by removing the backing bar." or "Weld only fiom one side." Future Work The welding advisor today has the capability t train new welding engineers and to guide, analyze, check, and verify the weld designs of more experienced personnel. Refining the knowledge in the advisor is a never ending task. In fact, the "smarter" it gets, the more users, and with more users come more input. In addition to refinement, new tests and performance criteria are being added. Work has begun to add knowledge about material preparation and purchase requirements. This summer we will be exploring both generative and adaptive weld schedule generation based on historical data and knowledge. Acknowledgments The author acknowledges the patience and enthusiasm of Ken Hicken and Jerry Knorovsky, the welding experts at Sandia, initial prototype and design of the advisor by Dave Messink of Intellicorp, and objects verses rules discussions with Bill Stubblefield and Dave Messink. Also, Kim Mahin and John Mitchiner are acknowledged for the support and development of the Smartweld program. This work was performed at Sandia National Laboratories supported by the U.S. Department of Energy under contract numbers DE-AC04-76DP00789 and DE-ACO4- 94AL85000. References Luger, G., and Stubblefield, B., 1993. Artificial Intelligence: Structures and Strategies for Complex Problem Solving. Redwood City, CA: Benjam WCummings. Mitchiner, J., Kleban, S., Hess, B., Mahin, K., Messink, D., 1996. Smartweld A Knowledge-based Approach to Welding. From these proceedings. Buchanan, B. and Shortliff, E., eds. 1984. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Hueristic Programming Project. Reading, MA: Addison- Wesley. Rossen, M. B., and Sherman R. Alpert, 1990.The Cognitive Consequences of Object-Oriented Design. Human-Computer Interaction, Volume 5,345-379. Rivera, J., Stubblefield, W, and Ames, A, 1996. From these proceedings. This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any'warranty, express or implied, or assumes any legal liability or responsi- bility for the accuracy. completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Refer- ence herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recom- mendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. 
work_un2s64p45jdxpl32xpood2vfoa	----	Problems Solving of Cell Subscribers based on Expert Systems Neural Networks (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 12, 2019 226 | P a g e www.ijacsa.thesai.org Problems Solving of Cell Subscribers based on Expert Systems Neural Networks Ahmad AbdulQadir AlRababah Faculty of Computing and Information Technology, King Abdulaziz University, Rabigh 21911, Kingdom of Saudi Arabia Abstract—With the growing demand for telecommunications services, the number of calls to telecommunications companies related to issues of using services, setting up and maintaining equipment, as well as resolving possible problems arising in the process of using services is also growing. From the point of view of system analysis, the problem is the mismatch between the existing and the required (target) state of the system for a given state of the environment at the moment in time. Based on this definition, we consider the problem of the subscriber of the cellular network a mismatch between the existing and the required state of the cellular network in this state of the environment at the moment in time. The state of the cellular network is characterized by the functioning of all devices, the proposed range of services. A short time for analyzing problem situations and making decisions, a large amount of information characterizing the current situation, the difficulty of solving poorly formalized and poorly structured tasks in the absence of complete and reliable information about the state of the cellular communication network and the functioning of its elements lead to a mismatch of human capabilities to effectively solve these problems. In this regard, the development and implementation of a precedent based neural network expert system for solving the problems of subscribers of a cellular communication network is an urgent scientific and technical task. Keywords—Neural network expert system; telecommunications companies; system analysis; cellular network; structured tasks; reliable information; human capabilities; telecommunications services I. INTRODUCTION Development of precedent based neural network expert system on an integrated approach to the problem of effective relationship management with subscribers of a cellular communication network [4,17,31], including the use of expert systems technologies, neural networks, case-based reasoning, as well as the creation of models [10,24,37], algorithms and support programs for decision-makers interacting with subscribers of a cellular communication network. Research Objectives of this manuscript are considered analysis of the most popular systems for managing cellular networks [5,19,32,40] and customer relationships; research of existing methods and techniques for decision support for tasks of managing relationships with subscribers of cellular communication networks; analysis of various technologies of intelligent systems, methods of their interaction and combination [15,29]; building mathematical models of precedent based neural network expert system components: a production fuzzy knowledge base about subscriber problems[16,30,42], a fuzzy controller based on a neural network, a knowledge base of precedents for problems; development of a complex of algorithms: processing subscriber applications, finding solutions to subscribers' problems based on precedents and using a fuzzy neural network[3,18,34]; creation of a neural network expert system based on precedents for solving problems of subscribers of a cellular communication network; study of the effectiveness of the developed precedent based neural network expert system using the following groups of indicators: functional suitability, efficiency, reliability, cost-effectiveness of the system[2,8,39]; development and implementation in trial operation of the precedent based neural network expert system software[6,13] to solve the problems of cellular network subscribers. Scientific novelties of this research are mentioned in the existing approaches to solving the problems of managing the relationships of operator companies with subscribers of cellular networks[11,25]; theoretically substantiated a new approach to building an intelligent system to solve the problems of subscribers of a cellular communication network, based on the integrated use of technology expert systems, neural networks, fuzzy logic and reasoning based on precedents; software was developed for building a precedent based neural network expert system, including: a production fuzzy knowledge base about problems of subscribers, a fuzzy controller based on a neural network[7,20], a knowledge base of precedents for problems; algorithms have been created for finding solutions to subscribers' problems based on precedents and using a fuzzy neural network; a neural network expert system based on precedents was developed to solve the problems of subscribers of a cellular communication network[21,33,38]; experimental studies were conducted to verify the effectiveness of the developed system. The practical value of this research was explored as a first version of the software neural network expert system based on precedents for solving the problems of subscribers of a cellular network after testing and evaluating specialists in 2009, where it was used to apply as an integrated approach to the development of precedent based neural network expert system, methods for building knowledge bases of fuzzy products and use cases[1,14,41], algorithms for finding solutions based on use cases and a neural networks. II. RESEARCH FRAMEWORK The first stage of the research will focus on analyzes methods and software tools for solving problems associated with servicing subscribers of cellular communication (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 12, 2019 227 | P a g e www.ijacsa.thesai.org networks[9,22]. The study showed that making decisions in the field of telecommunication network management, and in particular, in solving problems of servicing subscribers of such networks is a complex and multi-criteria task [23,28]. One of the bottlenecks in its solution is the procedure for finding the cause of problems with the provision of cellular services to the subscriber. This task requires a minimum of time to solve it and a maximum of reliability of the found solution to the identified problem [12,35]. The complexity of the problem also lies in the fact[26,38,43] that the state and dynamics of the processes of functioning of a cellular communication network cannot be unambiguously described using clear mathematical models. Based on the analysis of operators serving cellular network subscribers, it is concluded that it is necessary to use intelligent systems (IS) [7,39] that combine previously accumulated experience in the field of operation of a cellular communication network. As a result of research on various IS technologies and methods of their interaction, the feasibility of hybridization of various intellectual components is substantiated [5,27,36]. This stage of the work concludes with the goals and objectives of the study. III. PROPOSED METHODOLOGY The next stage of this research discusses the theoretical foundations and resolves issues of developing mathematical support for a precedent based neural network expert system to solve the problems of subscribers of a cellular communication network. The mathematical description of a neural network expert system based on precedents has the form:  pnnp p p IIIRpAKBKBNES ,,,),(,, 2 In a precedent based neural network expert system, the knowledge base contains knowledge in the form of products KB and in the form of precedents p KB , R  - systemic relations of IS. The search for solutions is divided into a neural network 2n I and case law p I with )( pA - an algorithm for determining similar use cases. Neural network training np I is based on data from precedents. A model of a system for solving subscriber servicing problems is presented as a nonlinear object (Fig. 1) with many inputs   i x and output variables  k y :          qkxxxfy nix nyk i ,1),,...,,( ;,1, 21 (1) Fig. 1. Model of a System for Solving Problems of Customer Service. Input variables are characteristics of the problem that the subscriber has. The output variables are the causes of the problem. Input   nix i ,1,  and outputs   qky k ,1,  variables can take only qualitative values, and the set of all possible values of these variables is known   mjj uuuU ,...,, 1  (2) Where j u - score corresponding to the lowest value of the input i x or outputs k y variable; m u - score corresponding to the largest input value i x (or output value k y ) variable m is the power of the set U. We accept that the vector  ** 2 * 1 * ,,, n xxxX  - fixed values of input variables of the considered system model, where niUx i ,1, *  . The task of finding a solution is based on the vector * X define output vector  ** 2 * 1 * ,,, q yyyY  . Input and output variables will be considered as linguistic variables defined on universal sets U. To evaluate linguistic variables, qualitative terms from the following term set are used:   mjj aaaA ,,, 1    , (3) Where A is the term set of variables i x and k y , j a - j-th linguistic term of a variable i x or k y , ni ,1 , mj ,1 , qk ,1 . Linguistic terms mjj aaa ,,, 1   are calculated as follows:     l p p j p j a j uua j 1  , where  pj a uj - degree of membership of an element Uu j  of term Aa j  , lp ,1 , mj ,1 . The task of constructing membership functions of elements Uu j  term set   mjj aaaA ,,, 1    comes down to determining degrees of belonging  p j a uj for all lp ,1 , mj ,1 . In accordance with (1), the MIMO structure (Multiple Input - Multiple Output) of a fuzzy knowledge base of the form: If (x1=a1 l ) and (x2=a2 l ) and … and (xn=am l ), then (y1=b1 l ) and (y2=b2 l ) and … and (yn=bm l ) (4) Where l is the rule number, Ll ,1 , L is the number of rules, l j a and l j b - fuzzy terms to evaluate the input variable x1 x2 xn y1 y2 yq Customer Service Problem Solving System . . . . . . (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 12, 2019 228 | P a g e www.ijacsa.thesai.org i x and output variable k y , ( ni ,1 , mj ,1 , qk ,1 ) in the l-th rule, respectively. We transform the system of logical statements (4) using operations (or), (and):       tt k p q k k p jk k p n i k p ji byax 1 11 1                  (5) Where qj ,1 , qk ,1 , k tp ,1 IV. IMPLEMENTATIONS AND EXPEREMENTAL RESULTS DISCUSSIONS To implement the process of extracting knowledge from a fuzzy knowledge base, a neuro-fuzzy logical inference mechanism is used in the form of a fuzzy controller based on a neural network - NNFLC (Neurons Network Fuzzy Logic Controller) (Fig. 2). Structurally, NNFLC is a multilayer network for direct signal propagation, and different layers perform different functions. Layer 1: represents membership functions implemented as radial basis neurons:      221 2exp iiii cxxy  . Layer 2: models the and- conditions of the rules:       11 1 2 ,...,min ni yyy  . Layer 3: is an OR combination of rules with identical terms in consequents:       22 1 3 ,...,max ni yyy  . In the training mode, the layer adjusts the parameters of the membership functions of the output variables. In operating mode, forms an output. Layer 4: in operating mode, neurons perform defuzzification:      j ijii yz 34  . In training mode, this is an additional input that performs normalization, which allows you to configure the membership function of the output variable:                j j jijiiii yzy  344 . The structure of a fuzzy neural network NNFLC is initialized on the basis of the formation of a complete matrix of rules. The mathematical definition of the knowledge base of precedents has the form:  IKBP n ,,,...,, 21  , where n  ,...,, 21 - a lot of precedents, I - a set of index terms that determine whether the precedent belongs to class K. Next stage of this research discusses the architectural features of precedent based neural network expert system (Fig. 3) to solve the problems of subscribers of a cellular communication network. Fig. 3 shows the structural diagram of the system. The input data for the precedent based neural network expert system is information from the application submitted by the network subscriber. The application contains general data, technical parameters, a list of actions carried out on the application. Fig. 2. The Structure of the Neuro-Fuzzy Controller NNFLC. Fig. 3. Structure of Neural Network Expert System. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 12, 2019 229 | P a g e www.ijacsa.thesai.org An algorithm for extracting precedents from the knowledge base using the Euclidean metric is developed. The input to the algorithm is: 1) a description of the subscriber’s problem   n pppP ,...,, 21  , including n values of parameters characterizing the situation; 2) BP is a non-empty set of precedents; 3)   n WWWW ,...,, 21  - weight (importance factors) parameters; 4) M - the number of precedents under consideration from the knowledge base, 5) K - threshold value of the degree of similarity. Output: a lot of precedents SP, which have a degree of similarity greater than (or equal to) the threshold value K. Step 1. SP=, j=1 and go to the next step. Step 2. If j≤M we choose the use case Aj from the set BP (AjBP) and go to step 3, otherwise the use cases are considered and go to step 7. Step 3. We calculate the distance according to the Euclidean metric between the selected use case Aj and the current situation P, taking into account the importance factors of the parameters:     i n i iij wpaAPD   1 2 , Step 4. We calculate the distance according to the Euclidean metric for the boundary values of the parameters:     i n i wppPD   1 2 minmaxmax Step 5. In this step, we calculate the degree of similarity     max 1, DDPAS  or as a percentage     %1001, max  DDPAS , if the threshold value of K is set in percent, and go to step 6. Step 6. If   KPAS , , then this precedent Aj is added to the result set SP (AjSP), means we extract this precedent from the knowledge base. After checking, j = j + 1 and go to step 2. Step 7. If SP=, then no precedents for the current problem situation were found and go to step 9 with a message for the operator about the need to reduce the threshold value K, otherwise the precedents for the current situation were successfully extracted and go to the next step. Step 8. The found precedents are sorted in decreasing order of similarity with the current situation and presented to the operator. Step 9.The end (completion of the algorithm). Neural network training is as follows: In step 1. A training sample is set, consisting of many examples of the following form:                 k q kkkk n kkk yyyyxxxx ,...,,,,,...,,, 321321 , Kk ,1 , where  k i x - values of input variables i x ( ni ,1 ) and  k j y - values of output variables j y ( qj ,1 )in k-th example; K is the total number of examples in the training set. The values of the input and output variables are determined by the membership functions of the corresponding terms:             k qqb k b k nna k a yyxx  ,...,,,..., 1111 , Kk ,1 . We introduce the notation  k i x and  k j y as values of membership functions of the terms of input and output variables, respectively. In step 2. Configuring the parameters of membership functions includes determining the centers i c and widths i  for membership functions represented by shape functions:               2 2 1 2 exp i ii i cx xy  The training data self-organization algorithm is used, which automatically divides the space into clusters. The center i c of the cluster is identified with the center of the radial basis function. The preliminary selection of centers is carried out randomly on the basis of uniform distribution. The matrix of bond weights ji  is specified following the conditions:         ,,0 1 1 ,1 ji T ji ii   where T is the number of neurons in the input layer. After presentation of the k-th vector  k x the center is selected from the training set i c closest to  k x according to the Euclidean metric:       T t iti k cxcx 1 2 . This center is subject to refinement in accordance with the winner algorithm (WTA algorithm):           kcxtkckc i k ii  1 , where  t - monotonously decreasing level of training. Other centers do not change. All training vectors  k x presented several times in random order until the stabilization of the values of the centers. Width Adjustment i  carried out heuristically, on the principle of "first closest neighbor":     ii , where  is the overlap parameter. The outputs of each layer are calculated by the formulas:       11 1 2 ,...,min ni yyy  ,       22 1 3 ,...,max ni yyy  ,              j j jijiii yy 34 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 12, 2019 230 | P a g e www.ijacsa.thesai.org Winner Algorithm Looks For Weight Matrix ji  , which evaluates the quality of relations between the left and right parts of the rules:       nn jiji 1 , where      jiij yy   23 . In step 3. The combination of rules is carried out with the participation of an expert. In step 4. Final configuration of membership functions is performed using the error back propagation algorithm for the error function   24 kii dye  regarding vectors  k x . For K learning pairs, the objective error function is defined as:                 K k K k T t kktikk dxydyE 1 1 2 0 424 2 1 2 1  The magnitude of the error determines the gradient vector of the objective function relative to specific centers ij c and width ij  : )( )()1( nc E ncnc ij ijij     , )( )()1( n E nn ij ijij      . Repeated training cycles lead to complete and fast network learning. The search for a solution to the subscriber’s problem using the neural network mechanism is carried out according to the created algorithm (Fig. 4). Fig. 4. Algorithm for Neural Network Search for a Solution to a Problem. The next main stage of this research discusses the features of the software implementation of a precedent based neural network expert system. The choice of the object-oriented programming language Java as a part of the Microsoft Visual Studio with a program development environment as the main precedent based neural network expert system programming tool is substantiated. The modern database development tools are analyzed and the choice is made in favor of the Microsoft SQL Server DBMS. A production fuzzy knowledge base has been developed for a system for solving problems of cellular network subscribers, training data sets have been generated for a neural network search mechanism. A software implementation of a fuzzy neural network was made, an algorithm for training and finding a solution using a neural network was implemented. A database of customer applications has been developed. A precedent presentation form is defined, a precedent knowledge base is built (Fig. 5), and a software implementation of the use case search is implemented. Fig. 5. Knowledge base Structure of Precedent based Neural Network Expert System. Use case 1 Use case M KB Parameter Description Possible Values Precedent Similarity Threshold Maximum search time Weight value (importance) Affinity metrics Use case 1 Use case N Parameter 1 to describe the problem Knowledge base structure Parameter N to describe the problem Knowledge base settings Klass 1 Klass K . . . . . . . . . . . . (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 12, 2019 231 | P a g e www.ijacsa.thesai.org To increase the speed and efficiency of the search for a precedent in the knowledge base, the full space of precedents is classified by the number of incidents in the precedent. V. RESULTS ANALYSIS AND PRACTICAL STRENGHS Next main important stage in this research is devoted to an experimental study of the operability and effectiveness of precedent based neural network expert system to solve the problems of subscribers of a cellular communication network. Four groups of indicators were selected as performance indicators: functional, operational, economic indicators and reliability indicators (Table I). TABLE. I. A LIST OF INDICATORS FOR ASSESSING THE EFFECTIVENESS OF PRECEDENT BASED NEURAL NETWORK EXPERT SYSTEMS № Indicator Assessment Object 1) Functionality 1. The total number of applications of subscribers System as a whole 2. Problem Identification Factor Use Case Search Subsystem Neural Network Search Subsystem System as a whole 3. The ratio of unprocessed applications Use Case Search Subsystem Neural Network Search Subsystem System as a whole 4. Problem Identification Criteria Accuracy Use Case Search Subsystem Neural Network Search Subsystem System as a whole 5. Problem Level System as a whole 2) Efficiency 6. Application Processing Duration System as a whole 7. Duration of finding a solution to a problem Intellectual part of the system The non-intellectual part of the system 8. Average processing time Use Case Search Subsystem Neural Network Search Subsystem The non-intellectual part of the system 9. The number of operators 3) Reliability 10. System uptime System as a whole 4) Profitability 11. Development cost System as a whole 12. Operation cost System as a whole 13. The average cost of processing one subscriber application System as a whole 14. Annual cost savings System as a whole 15. Annual economic effect System as a whole 16. Coefficient of economic efficiency System as a whole 17. Payback period (years) System as a whole A technique has been developed for conducting an experiment to check the operability and effectiveness of the developed information system. When testing a precedent based neural network expert system in real conditions, the system was installed on the computers of two operators of the call center of cellular network subscribers. One operator used only the database of customer requests, and another used an additional intelligent subsystem for servicing applications. The experiment was carried out during the week around the clock to take into account the influence of the day of the week and time of day on the frequency of receipt and processing time of subscribers' applications (Table II). An economic assessment of the effectiveness of the implementation of a neural network expert system based on precedents for solving the problems of subscribers of a cellular network was carried out according to the following particular criteria: development cost, operating cost, annual cost savings, annual economic effect, cost recovery period. The calculations showed that the costs of creating and maintaining the system pay off for 3.5 years of operation, and the use of intellectual support of the precedent based neural network expert system in the work of the mobile operator gives well enough annual savings. TABLE. II. RESULTS OF TESTING PRECEDENT BASED NEURAL NETWORK EXPERT SYSTEM IN REAL CONDITIONS Indicator Intelligent Subsystem The non- intellectual part of the system System as a whole Use Case Search Subsystem Neural Network Search Subsystem Total number of customer requests 347 471 572 1390 818 Problem Identification Factor 93,37% 100,00% 91,96% 95,11% 96,69% The ratio of unprocessed applications 6,63% 0,00% 8,04% 4,96% 6,63% Problem Identification Criteria Accuracy 96,50% 99,72% 81,33% 92,52% 98,11% Application processing time (hours) 35,16 24,54 143,00 168 59,59 Duration of finding a solution to a problem 1,39 0,59 12,45 4,51 1,09 Average processing time 5,54 3,13 15 7,59 4,33 The number of operators 1 1 2 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 12, 2019 232 | P a g e www.ijacsa.thesai.org VI. CONCLUSION The main scientific results of this research is a theoretical justification, a study of construction methods and the development of a neural network expert system based on precedents for solving the problems of subscribers of a cellular communication network. It was also investigated modern systems for managing cellular networks and customer relationships. The methods for solving problems that arise during the operation of a cellular communication network are analyzed, and it is concluded that it is advisable to hybridize various intelligent technologies in order to create a single advisory system for solving subscribers' problems. Also a mathematical model of a neural network expert system based on precedents for solving the problems of subscribers of a cellular communication network is developed; the system components and the processes of interaction of its intellectual components are mathematically described. In additional, a mathematical model of a fuzzy knowledge base with a MIMO structure has been built, including knowledge about the problems that subscribers have during the operation of a cellular network. The composition is determined and the characteristic of input and output linguistic variables and their terms is given. Beside the previous mentioned outputs and concludes a mathematical model of a fuzzy neural network output system using a fuzzy controller based on the NNFLC neural network has been developed. The use of a neural network approach to the implementation of fuzzy inference is justified. A module for explaining the solution obtained by the neural network search engine has been implemented. A neural network expert system based on precedents for solving the problems of subscribers of a cellular network is implemented programmatically. A database of subscriber applications has been developed, which is used both for registering applications and for creating precedents and knowledge base rules on their basis. Also performance indicators of the developed precedent based neural network expert system are determined. An experiment to verify the health and effectiveness of the system. The calculated indicators suggest that the developed precedent based neural network expert system has reliable software, good ability to identify the causes of subscribers' problems, and can identify these causes with a high degree of reliability and high speed. ACKNOWLEDGEMENT This project was funded by the Deanship of Scientific Research (DSR), King Abdulaziz University, Jeddah, under grant No. (DF-749-830-1441). The authors, therefore, gratefully acknowledge DSR technical and financial support. REFERENCES [1] Guimarães, Augusto Junio, et al. "Using fuzzy neural networks to the prediction of improvement in expert systems for treatment of immunotherapy." Ibero-American Conference on Artificial Intelligence. Springer, Cham, 2018.‏ [2] AlRababah, Ahmad Abdul Qadir. "Watermarking implementation on digital images and electronic signatures." International Journal of Advanced and Applied Sciences 4.10 (2017): 160-164.‏ [3] Bui, Nhan X., et al. "Reliable data reading with data set screening by error injection." U.S. Patent Application No. 10/262,681.‏ [4] AlRababah, Ahmad AbdulQadir. "Data Flows Management and Control in Computer Networks." International Journal of Advanced Computer Science and Applications 9.11 (2018): 207-217.‏ [5] Li, Xiaodong, et al. "Rapid, robust, and reliable blind deconvolution via nonconvex optimization." Applied and computational harmonic analysis ‏.893-934 :(2019) 47.3 [6] Rondinelli, Dennis A. Applied methods of regional analysis: the spatial dimensions of development policy. Routledge, 2019.‏ [7] Tang, Jinjun, et al. "Lane-changes prediction based on adaptive fuzzy neural network." Expert Systems with Applications 91 (2018): 452-463.‏ [8] Zhang, Chiya, et al. "Spectrum Sharing of Drone Networks." Handbook of Cognitive Radio (2019): 1279-1304.‏ [9] AlRABABAH, A.A., Implementation of Software Systems Packages in Visual Internal Structures. Journal of Theoretical & Applied Information Technology, 2017. 95(19). [10] Shen, Kai-wen, Xiao-kang Wang, and Jian-qiang Wang. "Multi-criteria decision-making method based on Smallest Enclosing Circle in incompletely reliable information environment." Computers & Industrial Engineering 130 (2019): 1-13.‏ [11] Bui, Dac-Khuong, et al. "A modified firefly algorithm-artificial neural network expert system for predicting compressive and tensile strength of high-performance concrete." Construction and Building Materials 180 ‏.320-333 :(2018) [12] AlRababah, Ahmad AbdulQadir. "Lempel-Ziv Implementation for a Compression System Model with Sliding Window Buffer."‏ [13] Huberman, Gur, Jacob Leshno, and Ciamac C. Moallemi. "An economic analysis of the Bitcoin payment system." Columbia Business School Research Paper 17-92 (2019).‏ [14] Anand, S. Krishna, TG Sundara Raman, and S. Subramanian. "Implementing a neuro fuzzy expert system for optimising the performance of chemical recovery boiler." International Journal of Artificial Intelligence and Soft Computing 4.2-3 (2014): 249-263.‏ [15] Mukherjee, Sayandev. "Distribution of downlink SINR in heterogeneous cellular networks." IEEE Journal on Selected Areas in Communications 30.3 (2012): 575-585.‏ [16] Deliyannis, Theodore, Yichuang Sun, and John Kelvin Fidler. Continuous-time active filter design. CRC press, 2019.‏ [17] Pelusi, Danilo, et al. "Neural network and fuzzy system for the tuning of Gravitational Search Algorithm parameters." Expert Systems with Applications 102(2018):234-244. Pearlson, Keri E., and Carol S. Saunders. Managing and using ‏ [18] information systems: A strategic approach. John Wiley & Sons, 2019.‏ [19] Silva Araújo, Vinícius Jonathan, et al. "Using resistin, glucose, age and bmi and pruning fuzzy neural network for the construction of expert systems in the prediction of breast cancer." Machine Learning and Knowledge Extraction 1.1 (2019): 466-482.‏ [20] AlRababah, A.A., A. AlShahrani, and B. Al-Kasasbeh, Efficiency Model of Information Systems as an Implementation of Key Performance Indicators. International Journal of Computer Science and Network Security (IJCSNS), 2016. 16(12): p. 139. [21] Reichling, Markus, and Tim Otto. "The environmental impact of the new economy: Deutsche Telekom, telecommunications services and the sustainable future." The ecology of the new economy. Routledge, 2017. ‏.119-129 [22] Naderializadeh, Navid, Mohammad Ali Maddah-Ali, and A. Salman Avestimehr. "Cache-aided interference management in wireless cellular networks." IEEE Transactions on Communications 67.5 (2019): 3376- ‏.3387 [23] Bar-Yam, Yaneer. Dynamics of complex systems. CRC Press, 2019.‏ [24] Kristjanpoller, Werner, and Marcel C. Minutolo. "A hybrid volatility forecasting framework integrating GARCH, artificial neural network, technical analysis and principal components analysis." Expert Systems with Applications 109 (2018): 1-11.‏ [25] AlRababah, A.A., A new model of information systems efficiency based on key performance indicator (KPI). management, 2017. 4: p. 8. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 10, No. 12, 2019 233 | P a g e www.ijacsa.thesai.org [26] AL-Qutami, Tareq Aziz, et al. "Virtual multiphase flow metering using diverse neural network ensemble and adaptive simulated annealing." Expert Systems with Applications 93 (2018): 72-85.‏ [27] Reichling, Markus, and Tim Otto. "The environmental impact of the new economy: Deutsche Telekom, telecommunications services and the sustainable future." The ecology of the new economy. Routledge, 2017. 119-129. [28] Al-Rababah, Ahmad A., Taghreed AlTamimi, and Najat Shalash. "A New Model for Software Engineering Systems Quality Improvement." Research Journal of Applied Sciences, Engineering and Technology ‏.2724-2728 :(2014) 7.13 Moustafa, Akram, et al. "A New Dynamic Model for Software Testing ‏ [29] Quality." Research Journal of Applied Sciences, Engineering and Technology 7.1 (2014): 191-197.‏ [30] Rendón, Claudio Marco Cartagena, et al. "Proposed Model for Measuring Customer Satisfaction with Telecommunications Services." Mediterranean Journal of Social Sciences 8.2 (2017): 15-25.‏ [31] Hadi, Mohammed S., et al. "Patient-Centric Cellular Networks Optimization using Big Data Analytics." IEEE Access 7 (2019): 49279- 49296. [32] AlRababah, Ahmad. "Digital Image Encryption Implementations Based on AES Algorithm." VAWKUM Transactions on Computer Sciences ‏.1-9 :(2017) 13.1 [33] Smith, Trevor D., et al. "Methods and systems for distributing fiber optic telecommunications services to local area." U.S. Patent No. 8,805,152. 12 Aug. 2014.‏ [34] Al Ofeishat, H.A. and A.A. Al-Rababah, Real-time programming platforms in the mainstream environments. IJCSNS, 2009. 9(1): p. 197. [35] Wang, Yueying, Xixiang Yang, and Huaicheng Yan. "Reliable fuzzy tracking control of near-space hypersonic vehicle using aperiodic measurement information." IEEE Transactions on Industrial Electronics ‏.(2019) [36] Chougdali, Sallami, et al. "New air traffic management approach based on expert system and using real-time scheduling algorithms." International Journal of Intelligent Engineering Informatics 4.3-4 (2016): 305-321. ,Zahariadis, Nikolaos. "The multiple streams framework: Structure‏ [37] limitations, prospects." Theories of the Policy Process, Second Edition. Routledge, 2019. 65-92.‏ [38] AlRababah, Ahmad Abdul Qadir. "On the associative memory utilization in English-Arabic natural language processing." International Journal of Advanced and Applied Sciences 4.8 (2017): 14-18.‏ [39] Bolz, Ray E. CRC handbook of tables for applied engineering science. CRC press, 2019.‏ [40] Zurkirch, Manfred, and Inge Reichart. "Environmental Impacts Of Telecommunications Services: Two life-cycle analysis studies." The Ecology of the New Economy. Routledge, 2017. 130-149.‏ [41] Chen, Min, et al. "Mobility-aware caching and computation offloading in 5G ultra-dense cellular networks." Sensors 16.7 (2016): 974.‏ [42] Goddard, John B., and Andrew E. Gillespie. "Advanced telecommunications and regional economic development." Managing the city. Routledge, 2017. 84-109.‏ [43] Al Rababah, Ahmad Abdul Qadir. "Embedded Architecture for Object Tracking using Kalman Filter." JCS 12.5 (2016): 241-245.‏ 
work_unb545vt5vhmhnf5eo3lzltypm	----	Expert system based parallel multi-1D block matching algorithm with implementation for motion estimation Expert Systems with Applications 39 (2012) 3249–3256 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Expert system based parallel multi-1D block matching algorithm with implementation for motion estimation q Shin-Yeu Lin a,⇑, Chong-Wei Su b, Jung-Shou Huang c a Department of Electrical Engineering & Green Technology Research Center at Chang Gung University, Taoyuan, Taiwan, ROC b Institute of Electrical and Control Engineering at National Chiao Tung University, Hsinchu, Taiwan, ROC c Elan Electronics Corporation, Hsinchu, Taiwan, ROC a r t i c l e i n f o Keywords: Expert system Knowledge base Inference engine Block matching algorithm Motion estimation Mixed signal 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.09.012 q This research work was supported in part by Natio under Grant NSC98-2221-E-182-065-MY2. ⇑ Corresponding author. Address: Department of El Technology Research Center, Chang Gung Universit Kwei-Shan, Tao-Yuan 333, Taiwan, ROC. Tel.: +886 3 2118026. E-mail addresses: shinylin@mail.cgu.edu.tw (S.- edu.tw (C.-W. Su), rong@emc.com.tw (J.-S. Huang). a b s t r a c t In this paper, we propose an expert-system based parallel multi-1-dimensional block matching algorithm (ESPM-1D-BMA) for motion estimation (ME). Instead of the conventional 2D block matching, we employ the parallel multi-1D blocks matching to improve the computing speed. To improve the ME accuracy, we design a knowledge base and inference engine to determine the true motion vector (MV) from the results of parallel multi-1D blocks matching. To speed up the computing speed further, we present a hardware architecture for implementing the ESPM-1D-BMA. We have demonstrated that the MV estimation accu- racy achieved by the proposed ESPM-1D-BMA is much better than the comparing fast block matching algorithms and is close to the 2 dimensional full search block matching algorithm (2D-FSBMA). We also demonstrate that the computing speed of the proposed ESPM-1D-BMA is about two times as fast as the mixed-signal 2D-FSBMA (MS-2D-FSBMA). � 2011 Elsevier Ltd. All rights reserved. 1. Introduction matching to reduce the computational complexity in searching Two dimensional (2D)-block matching algorithm (BMA) is a commonly adopted method for searching the motion vector (MV) between two image frames namely the reference and the search frames, such that the MV is obtained when the best matched 2D- blocks, the reference and one search blocks are found. Among vari- ous BMAs, the 2D-full search BMA (2D-FSBMA) using the mean square error (MSE) criteria (Gharavi & Mills, 1990) is considered to be the most accurate algorithm for searching the MV. However, the 2D-FSBMA is computationally complex. Hence, some fast BMAs were proposed, such as the new three-step search (NTSS) (Li, Zeng, & Liou, 1994), diamond search (DS) (Zhu & Ma, 1997) and the hexago- nal based search (HEXBS) (Zhu, Lin, & Chau, 2002). These algorithms use few search points to reduce computational complexity, however at the price of poor accuracy. Therefore, proposing a method to re- duce the computational complexity of 2D-FSBMA while maintaining its accuracy in searching the MV is the purpose of this paper. Instead of 2D-block matching, we will slice a 2D block into mul- tiple, say K, 1D blocks and employ a parallel multi-1D blocks ll rights reserved. nal Science Council in Taiwan ectrical Engineering & Green y, 259 Wen-Hwa 1st Road, 2118800x3221; fax: +886 3 Y. Lin), cwsu.ece97g@nctu. the MV. Due to the noise appearing in the search and reference frames, the 1D-block matching should be less accurate than the 2D-block matching in searching the MV. Additionally, the MVs determined in each of the K 1D-blocks matching may be different due to various noise contaminations in various 1D blocks. There- fore, to remedy the possible inaccuracy in searching the MV using parallel multi-1D blocks matching, we propose an expert system based parallel multi-1D-BMA (ESPM-1D-BMA). For the purpose of real-time motion estimation (ME), we need to improve the computing speed further by implementing the pro- posed algorithm in hardware. Therefore, we will present the hard- ware implementation architecture of the proposed algorithm. We organize our paper in the following manner. In Section 2, we will present the proposed ESPM-1D-BMA. In Section 3, we will present the hardware implementation architecture of the proposed algorithm. In Section 4, we will test the performance of the pro- posed algorithm and compare with other existing methods in terms of ME accuracy and the computing speed using comprehen- sive simulations. Finally, we will draw a conclusion in Section 5. 2. Expert system based parallel multi-1D block matching algorithm (ESPM-1D-BMA) 2.1. Review of 2D-FSBMA We let In and In�1 in Fig. 1(a) denote the reference and search frames, respectively; we let block A inside In denote the reference http://dx.doi.org/10.1016/j.eswa.2011.09.012 mailto:shinylin@mail.cgu.edu.tw mailto:cwsu.ece97g@nctu. edu.tw mailto:cwsu.ece97g@nctu. edu.tw mailto:rong@emc.com.tw http://dx.doi.org/10.1016/j.eswa.2011.09.012 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 3250 S.-Y. Lin et al. / Expert Systems with Applications 39 (2012) 3249–3256 block (RB) and let block B, which can be any block in In�1, denote the search block (SB). The idea of 2D-FSBMA is to search all possi- ble SBs and find the one that is most similar to A. This searching task is performed by computing the MSE between blocks A and B as described below. We assume that the sizes of the RB A (or SB B) and frame In (or In�1) are X � Y and M � N pixels, respectively, as shown in Fig. 1(b). The MSE between two blocks, RB A and SB B, induced in 2D-FSBMA, denoted by MSE2D, is defined as MSE2D ¼ 1 X � Y XX i¼1 XY j¼1 ðrði; jÞ� sði; jÞÞ2 ð1Þ where r(i, j) and s(i, j) denote the (i, j)th pixel values of the RB A and SB B, respectively. Therefore, the 2D-FSBMA will search through all possible SBs to find the one with smallest MSE, say B⁄. Then the MV is defined as the difference of position indices between RB A and SB B⁄ as shown in Fig. 1(a). 2.2. Motivation To improve the computing speed of 2D-FSBMA, we will slice the 2D block into multi-1D blocks and apply a parallel multi-1D blocks matching algorithm. However, the MVs determined in each of the multi-1D blocks matching may be different due to various noise contaminations in various 1D blocks. Therefore, we will use the ex- pert system concept to help determine the true MV. In the follow- ing, we will describe the proposed algorithm step by step. 2.3. The 1D block matching We let B denote the number of pixels in 1D block, then a 1D block is formed by the B consecutive pixel values from a row of the frame. We let r(i), i = 1, . . . , B and s(i), i = 1, . . . , B denote the B Fig. 1. Motion vector determination using 2D-FSBMA Fig. 2. A diagram of the K independen pixel values of the 1D reference block and the 1D search block, respectively. Then, the MSE between r(i), i = 1, . . . , B and s(i), i = 1, . . . , B can be computed as follows: MSEðSBCÞ¼ 1 B XB i¼1 ðrðiÞ� sðiÞÞ2 ð2Þ where SBC represents the 1D search block coordinate (SBC), which is identified by the coordinate of the first pixel of the 1D search block, s(1). 2.4. Parallel multi-1D blocks matching To improve the ME accuracy of 1D block matching, while keep- ing its computational efficiency, we can use a parallel multi-1D blocks matching. The multi-1D blocks matching consists of K 1D reference blocks, and each 1D reference block represents one inde- pendent 1D reference block in the reference frame as shown in Fig. 2, in which we assume K = 8. We let rk(i), i = 1, . . . , B denote the kth 1D reference block in In and let the MSEk(SBC) denote the MSE between rk(i), i = 1, . . . , B in In and the search 1D block, s(i), i = 1, . . . , B in In�1, then MSEk(SBC) for k = 1, . . . , K can be computed by MSEkðSBCÞ¼ 1 B XB i¼1 ðrkðiÞ� sðiÞÞ 2 ð3Þ which can be performed independently and in parallel for each k. 2.5. Expert system based parallel multi-1D blocks matching algorithm (ESPM-1D-BMA) As described earlier that the searched MV based on each of the K 1D-blocks matching may be different due to various noise . (b) Example reference block and image frame. t 1D blocks in a reference frame. S.-Y. Lin et al. / Expert Systems with Applications 39 (2012) 3249–3256 3251 contamination in the search and reference 1D blocks. However, we can view the MSEs computed from the 1D block matching for a ref- erence block in In as a result determined by an expert. Therefore, the MSEs resulted from K 1D blocks matching can be viewed as a result determined by K experts. Subsequently, to determine the true MV, we can employ the concept of expert system (Liao, 2005; Li & Sun, 2009; Sun & Li, 2008) to construct the knowledge base (KB) and the inference engine (IE) for the parallel multi-1D blocks matching as follows. The input of the employed expert system is the overall resulted MSEs of the K 1D-blocks matching. For each of the K 1D-blocks matching, the true MV should be among the MVs with top smallest MSEs. Therefore, the employed KB of the proposed algorithm can be stated as follows. For each of the K 1D blocks matching, we se- lect the P search blocks that correspond to the top P smallest MSEs and assign them with the marked numbers (MNs) P, P � 1, . . . , 1, such that the selected search block with smaller MSE is marked by a larger MN. For example, the MN assigned to the search block with smallest MSE is P. Since each selected search block may con- clude an MV, there will be K � P MVs resulted from the parallel multi-1D blocks matching, and each MV is associated with the MN of the corresponding search block. However, some of the K � P MVs may be the same. In general, the frequently appearing MVs and the MV with smaller MSE have higher probability to be the true MV. Consequently, if an MV ap- pears q times in the K � P MVs, it will associate with q MNs. There- fore, the IE of our expert system can be stated as follows. For each distinct MV with q copies in the resulted K � P MVs, we will sum the q MNs and define the resulting value as the accumulated MN (AMN) of the MV. Consequently, the MV with largest AMN is deter- mined to be the true MV. For the sake of illustration, we use the following example to explain the proposed ESPM-1D-BMA. We assume K = 8 and P = 3. In Table 1, the first column shows the K 1D reference blocks, and the second, the third and the fourth columns show the P (=3) MVs with P smallest MSEs resulted from the 1D block match- ing for each reference block. Then the MVs in columns 2, 3, and 4 are assigned with MNs 3, 2, and 1, respectively. Calculating the Table 1 The top 3 MV for each 1D reference block. Index of reference block MV MV with smallest MSE 1 (3, 2) 2 (3, 2) 3 (5, 8) 4 (8, 8) 5 (8, 9) 6 (5, 8) 7 (3, 2) 8 (3, 2) Fig. 3. The hardware implementation architecture of th AMNs of the seven distinct MVs presented in Table 1, we find that (3, 2) has the largest AMN, 17, and is considered to be the true MV. 3. Implementation for motion estimation For the purpose of real-time ME, we can speed up the comput- ing speed of the proposed algorithm further by hardware imple- mentation. However, a pure digital circuit implementation (Hsieh & Lin, 1992; Yang, Wolf, & Vijaykrishan, 2005) may suffer from some implementation problems, such as high power consumption and large chip size. To overcome these implementation problems, a mixed-signal approach that uses simple current-summation circuit to circumvent computationally complex digital MSE computation should be a good choice (Panovic & Demosthenous, 2006). In the following, we will present the implementation of the proposed algorithm using mixed-signal approach step by step. 3.1. Implementing 1D-block matching using mixed-signal approach First of all, we transformed the sensored pixel values into volt- ages within the range [0 V, 2.5 V] by dividing the range of pixel val- ues between black and white into 255 grey levels, which are represented by 255 least significant bits (LSBs), such that black and white correspond to 0 and 255 LSB, respectively. We let Vx de- note the transformed voltage of a pixel value of x LSB, then V x ¼ 2:5V�0V255 � x. We let voltages Vr(i) and Vs(i) denote the trans- formed voltages of r(i) and s(i), respectively. Assuming B = 8, the mixed signal approach for the 1D block matching hardware imple- mentation architecture is presented in Fig. 3, in which we preload the transformed voltage Vr(i) of the pixel value of the 1D reference block r(i), i = 1, . . . , B in In into the reference block memory (RBM). Then the transformed voltage Vs(i) of the pixel value of the search frame In�1 is serially fed into the search block memory (SBM) from left to right, from top to bottom, then the transformed voltage of the pixel value of the 1D search block will be fetched from the SBM as shown in Fig. 3. The two clock signals tc and tr are used MV with 2nd smallest MSE MV with 3rd smallest MSE (5, 8) (1, 5) (5, 5) (8, 9) (3, 2) (8, 8) (3, 2) (5, 8) (6, 3) (5, 5) (6, 3) (3, 2) (5, 5) (1, 5) (6, 3) (1, 5) e 1D block matching using mixed-signal approach. Fig. 4. The computing architecture of parallel multi-1D block matching. 3252 S.-Y. Lin et al. / Expert Systems with Applications 39 (2012) 3249–3256 to control the timing of the propagation of the pixel value fed into SBM and RBM. The SE(i) denoted by a circle in Fig. 3 represents the ith square error computing circuit to result in an output current kt(Vr(i) � Vs(i))2, where kt is the transconductance parameter. The summation of the output currents of SE(i), denoted by I(i) in Fig. 3, i = 1, . . . , 8, will flow into a single resistive load R as shown in Fig. 3. Then, the resulted voltage across R denoted by VO is pro- portional to the MSE of 1D block matching and will be input to a P winner comparator (PWC), which will be presented later. Clearly, the computation of the MSE for all possible 1D search blocks can be completed when the last pixel value in the last row of the frame In�1 is read out. 3.2. Implementing parallel multi-1D blocks matching The parallel computing architecture for the K 1D blocks matching is presented in Fig. 4. The K 1D reference blocks are preloaded into RBMk, k = 1, . . . , K, and the search block in the frame In�1 are serially fed into SBM. The MSEk, k = 1, . . ., K are computed in parallel to obtain MSEk(SBC), k = 1, . . . , K. The detailed structure of SBM, RBMk, and MSEk in Fig. 4 are the same as that presented in Fig. 3. Fig. 5. The circuit of PW 3.3. Implementing ESPM-1D-BMA To implement the ESPM-1D-BMA, we need a PWC to select the P search blocks that correspond to the top P smallest MSEs resulted from the 1D block matching for each 1D reference block. The cir- cuit of PWC for the kth 1D reference block, denoted by PWCk, is presented in Fig. 5, which is designed based on a sorting logic; the solid lines and dotted lines in this figure represent the trans- mission of data and signals, respectively. At the very beginning, the P (=3) sample and hold circuits (SHs) are reset to a default value, which is the largest value that SH can take; similarly, the corresponding P digital memories (DMs) are reset to null values. Then, for each incoming MSEk(SBC), we will compare it with the MSE values stored in the P SHs as indicated by the P compar- ators, denoted by COMPi, i = 1, . . . , P, shown in Fig. 5. COMPi will gen- erate an enable signal eni, whose value depends on the comparison result such that if MSEkðSBCÞ < MSESHi then eni = 1; otherwise eni = 0. The combination of en1, . . . , enP will indicate which of the following ranges that MSEk(SBC) lies: ½0; MSESH1�; ðMSESH1 ; MSESH2�; . . . ;ðMSESHP�1 ; MSESHP �, orðMSESHP ;1�as presented in the illustrative table in Fig. 5, in which we set P = 3. From the value range of the incoming MSEk(SBC), we can easily update the P winners as follows. If eni ¼ 1; MSESHi , the content of SHi, should be replaced by either MSESHi�1 for the case eni�1 = 1 or MSEk(SBC) for the case that eni�1 = 0 or i = 1, and the content of the corresponding DMi will be replaced by the proper SBC accordingly. To implement the above P winners updating logic, we will design a selection signal, seli, to determine what to replace the contents of SHi and DMi when the enable signal eni = 1 in the following manner. If eni = 1 and seli = 1, the contents of SHi and DMi will be replaced by SHi�1 and DMi�1, respectively. If eni = 1 and seli = 0, the contents of SHi and DMi will be replaced by MSEk (SBC) and the corresponding SBC, respectively. If eni = 0, SHi and DMiremain unchanged. The values of seli, i = 2, . . . , P, which are determined based on the values of eni, i = 1, . . . , P, are presented in the illustrative table of Fig. 5, and they can be generated using AND Ck and illustrations. S.-Y. Lin et al. / Expert Systems with Applications 39 (2012) 3249–3256 3253 gates as shown in Fig. 5. Notably, sel1 is not needed because if en1 ¼ 1; MSESH1 and DM1 must be replaced by MSEk(SBC) and the corresponding SBC. For the kth 1D reference block with reference block coordinate RBCk, the above comparison and replacement pro- cess will continue until the last pixel value in the last row of the search frame is read out. The final contents in SHi and DMi, i = 1, . . . , P are the MSEs and the SBCs of the P search blocks with top P smallest MSEs, respectively. Consequently, the P SBCs stored in DMi, i = 1, . . . , P and the kth reference block coordinate (RBCk) will be input to the subtracter, SUB, as shown in Fig. 5 to calculate the top P MVks corresponding to the P smallest MSEs. This constitutes the operations of PWCk. By the aid of PWCk, k = 1, . . . , K, we can design the IE using the following components. Except for the indicators for indicating the index of and the number of distinct MVs in the K � P MVs, we employ a counter to cyclically generate the MN for each MV, an identifier to recognize the distinct MVs, an adder to calculate the AMN for each distinct MV and a comparator to identify the MV with largest AMN. Since the circuit to interconnect the above mentioned components for implementing the IE and generating the true MV is complicated and tedious, we will present it in Appendix A. 3.4. Hardware Implementation Architecture of ESPM-1D-BMA for ME Now, combining the parallel multi-1D blocks matching comput- ing architecture (Figs. 3 and 4), PWCk (Fig. 5), k = 1, . . . , K, and IE Fig. 6. The block diagram of ESPM-1D-BMA for ME. Table 2 The average ME accuracies of the 5000 estimated MVs. K P Football Greens Concord 2 1 91.70 98.84 89.98 2 95.11 98.96 92.12 3 96.04 99.00 93.18 4 96.51 99.36 94.18 5 97.36 99.04 94.12 4 1 98.27 99.60 94.32 2 98.83 99.36 95.18 3 99.08 99.44 95.40 4 99.39 99.42 96.20 5 99.56 99.48 96.10 6 1 99.30 99.68 96.12 2 99.55 99.52 96.80 3 99.58 99.32 97.10 4 99.50 99.60 97.52 5 99.57 99.68 97.96 8 1 99.63 99.82 96.98 2 99.65 99.58 97.12 3 99.58 99.54 97.20 4 99.60 99.60 97.92 5 99.69 99.46 98.28 2D-FSBMA 100 99.998 99.947 (Fig. 8 in Appendix A), the hardware implementation architecture of the proposed ESPM-1D-BMA for ME can be described in Fig. 6. The solid lines and dotted lines in Fig. 6 represent the transmission of data and signals, respectively. For the sake of simplicity in illustration, we assume K = 3 in Fig. 6, where SBM, RBMk and MSEk, k = 1, . . . , K, are the same as those in Fig. 4. The Image Sensor, IS, shown in Fig. 6 is used to ob- tain the image data of the 1D search block and the K reference 1D blocks. The purpose of Digital Controller, DC, is to control the acti- vation timing of each unit and the synchronization of the parallel processing architecture. Therefore, the operations of the hardware implementation architecture presented in Fig. 6 can be described as follows. First of all, the image data of the K 1D reference blocks will be preloaded into RBMk, k = 1, . . . , K. The DC will send a control signal to the IS to obtain the image data of In�1, which will be input to SBM. Since MSEk, k = 1, . . . , K are analog circuits, they will com- pute MSEk(SBC), k = 1, . . . , K, directly and in parallel once the data are ready at the output of both SBM and RBMk, k = 1, . . . , K. The computed MSEk(SBC), k = 1, . . . , K will be input to PWCk, k = 1, . . . , K controlled by the signal output from DC to determine whether it is among the P smallest MSEks for the kth 1D reference block. The above process will repeat until the last pixel value in the last row of the search frame is read out. Then DC will send a signal to PWCk to output the top P MVks for k = 1, . . . , K in parallel. These K � P MVs will be input to IE to generate the true MV. 3.5. Time complexity We let T(�) denote the computation time or propagation time of unit (�). In addition toT(�), there are other time delays need be con- sidered such as (i) the time delay incurred from DC to guarantee a safety margin of controlling the action of the K units of same type of components in parallel and (ii) the wire delay between units. However, comparing with T(�), these time delays are negligible. Therefore, the estimated time needed to generate an MV of the ESPM-1D-BMA can be stated in the following: ðM � NÞðT SH þ T MSE þ T PWCÞþ T IE ð4Þ where TSH denotes the time for loading a pixel value into RBM or SBM, TMSE denotes the time required in computing the MSE of 1D block matching, which is the propagation time of the circuit Fabric Hestain Pears Tissue Board 96.62 98.60 81.16 98.04 98.34 98.33 99.14 86.20 97.82 98.76 98.89 99.30 90.16 97.92 99.00 99.08 99.20 91.84 97.82 98.70 99.17 98.92 91.30 98.04 98.72 99.59 99.54 93.74 98.66 98.72 99.53 99.16 95.96 98.30 99.98 99.59 99.04 97.24 98.28 99.70 99.79 99.20 97.60 98.32 99.66 99.70 99.12 97.80 98.12 99.66 99.92 99.44 96.96 98.60 100 99.65 99.18 97.54 98.54 99.44 99.68 99.22 98.24 98.44 99.62 99.63 99.32 98.30 98.54 99.60 99.62 99.24 98.32 98.52 99.72 99.94 99.48 97.70 98.88 99.98 99.72 98.98 98.70 98.78 99.64 99.57 98.90 99.28 98.52 99.70 99.63 99.36 99.14 98.78 99.74 99.54 99.24 98.98 98.58 99.70 100 99.625 99.923 99.285 100 3254 S.-Y. Lin et al. / Expert Systems with Applications 39 (2012) 3249–3256 presented in Fig. 3, TPWC denotes the propagation time of the PWCk presented in Fig. 5 and TIE denotes the processing time of the IE pre- sented in Fig. 8. Fig. 7. The pseudo-code of IE. 4. Test results and comparisons In this section, we will demonstrate the ME accuracy achieved by the proposed ESPM-1D-BMA in comparison with the 2D-FSBMA and some fast block matching algorithms (Li et al., 1994; Zhu et al., 2002; Zhu & Ma, 1997) using extensive simulations. We will also compare the computational efficiency of the proposed ESPM-1D- BMA with the 2D-FSBMA. We use Matlab and their eight pictures, football, greens, concord, fabric, hestain, pears, tissue and board, as our simulation tool and test bed, respectively. The eight tested pic- tures were all formatted as grey-level image in eight-bit. We set M = 24 and N = 24 for all tests, X = 8 and Y = 8 for 2D-FSBMA, and B = 8 for ESPM-1D-BMA. However, we will use various combina- tions of K and P to test the performance of ESPM-1D-BMA. For each tested picture, we arbitrarily pick an image frame of M � N pixels to serve as the reference frame In. For each reference frame In, we prepare the search frame In�1 by the following procedures. We ran- domly generate an MV ranging from �M�X2 to M�X 2 and from � N�Y 2 to N�Y 2 in x and y directions, respectively, and apply it to In to form a noise free In�1. For each pixel in the noise free In�1, we randomly generate a noisy signal based on a normal distribution with mean 0 LSB and variance 3 LSBs and add it to the pixel. The resulted frame will serve as the tested search frame In�1. For each one of the eight tested pictures, we prepare 5000 (In�1, In)s based on the above process. 4.1. Comparisons of ME accuracy Now for each of the eight tested pictures, we use the prepared 5000 (In�1, In) s to estimate the corresponding randomly generated MVs using the proposed ESPM-1D-BMA and the 2D-FSBMA, the average accuracy of the 5000 estimated MVs for each picture and for various combinations of K and P are presented in Table 2. From Table 2, we can observe that the larger the values of K and P in the proposed ESPM-1D-BMA, the more accurate the MV estimation will be. We can also observe that when K P 8 and P P 4, the pro- posed algorithm is at most 1% less accurate than the 2D-FSBMA on the average. Putting the test results of ESPM-1D-BMA for K = 8 and P = 3 pre- sented in Table 2 in the second row of Table 3 for reference, we also use the same 5000 prepared (In�1, In) s to test the three fast block matching algorithms, the DS, the NTSS, the HEXBS. The average ME accuracy resulted by these three methods are presented in the last three rows of Table 3, which show that the proposed ESPM-1D-BMA is far better than the three fast block matching algorithms in ME accuracy. 4.2. Comparison of computational efficiency Now, to evaluate the computing time of ESPM-1D-BMA and the mixed-signal approach 2D-FSBMA (MS-2D-FSBMA), we need to re- view the computational complexity of the latter first. Similar to the Table 3 Comparisons of ESPM-1D-BMA with the three fast block matching algorithms. Football Greens Concord ESPM-1D-BMA (K = 8, P = 3) 99.58 99.54 97.20 DS 45.93 69.04 81.79 NTSS 48.69 81.13 86.87 HEXBS 38.15 63.02 73.89 1 � B current summation circuit employed in 1D block matching presented in Fig. 3, the MS-2D-FSBMA employed a X � Y current- summation circuit to obtain the 2D-MSE (Panovic & Demosthen- ous, 2006). Instead of PWCk, k = 1, . . . , K, the MS-2D-FSBMA need only one single winner comparator (SWC), which consists of a com- parator, COMP, a SH, and a DM to identify the coordinate of the SB with minimum MSE. In the case of a frame with size M � N, a block with size X � Y, the computing time of the MS-2D-FSBMA for com- puting an MV can be calculated by: 2 � X � Y � T SH þðM � XÞðN � Y þ 1ÞðY � T SH þ T MSE2D þ T COMPÞþðN � YÞðX � T SH þ T MSE2D þ T COMPÞ ð5Þ where TSH is the same as that in (4); T MSE2D denotes the time for per- forming a 2D-MSE computation; since the most time consuming Fabric Hestain Pears Tissue Board 99.57 98.90 99.28 98.52 99.70 82.88 89.04 75.54 79.24 38.73 86.22 87.89 75.01 87.69 43.99 67.10 74.53 62.13 74.69 28.19 Fig. 8. Hardware architecture for implementing the IE. S.-Y. Lin et al. / Expert Systems with Applications 39 (2012) 3249–3256 3255 component in SWC is the COMP, the time needed for comparison in SWC is TCOMP. Based on the existing circuit for SH (Chen, Gu, Shen, Wu, & Hsu, 1998), MSE2D (Panovic & Demosthenous, 2004) and COMP (Razavi & Wooley, 1992), we obtain the following computing time using HSPICE simulation: T SH ¼ 50 ns; T MSE2D ¼ 10 ns, and TCOMP = 100 ns. Although the analog circuit for computing MSE presented in Fig. 3 is simpler than computing MSE2D in MS-2D-FSBMA, the operations of all SEs in both circuits are carried out in parallel. Therefore, T MSE ffi T MSE2D ¼ 10 ns. Similarly, the operations of PWCk (Fig. 5), k = 1, . . . , K are also carried out in parallel, and the most time consuming component in PWCk is COMPk, therefore, TPWC ffi TSWC = TCOMP = 100 ns. According to the hardware implementation architecture of the IE presented in Appendix A, the processing time for IE is TIE = K � P � Tclock, where Tclock = 13.1 ns is the critical path delay of IE. Therefore, for K = 8, P = 3, we have TIE = 314.4 ns. Conse- quently, based on the parameters employed in our tests, the com- puting time needed for an MV estimation of ESPM-1D-BMA with K = 8, P = 3 and MS-2D-FSBMA are 92474.4 ns and 153,280 ns, respectively. This demonstrates that the proposed ESPM1D-BMA uses only 60% of computing time of the MS-2D-FSBMA and achieves almost the same ME accuracy. Notably, the critical point of the computing efficiency achieved by the ESPM-1D-BMA is the small amount of time spending on loading the pixel values of the frame In�1 into the SBM. The part of loading time in (4) and (5) for ESPM-1D-BMA and MS-2D-FSBMA are M � N � TSH and 2 � X � Y � TSH + (M � X)(N � Y + 1)(Y � TSH) + (N � Y)(X � TSH), respec- tively, which constitutes 31.14% and 79.33% of the corresponding total computing time, respectively. 5. Conclusion In this paper, we have proposed a hardware implementable ESPM-1D-BMA and demonstrated that its ME accuracy is close to the 2D-FSBMA and better than the three comparing fast block matching algorithms. Above all, the computing speed of ESPM- 1D-BMA is about two times as fast as the MS-2D-FSBMA. Appendix A To implement the IE, we will first describe the pseudo code for carrying out the IE then present the hardware architecture for implementing the pseudo code. We let mvk,p and MNk,p, p = 1, . . . , P denote the P MVs that correspond to the top P smallest MSEs for the kth 1D block matching and the corresponding MN, respectively; we let D denote the number of distinct MVs among the K � P mvk,p’s and let Dmvd, d = 1,. . .,D, denote the D distinct MVs; we let AMNddenote the AMN of Dmvd and let wMV and wAMN denote the resulting best-so-far MV and the corresponding AMN during the process, respectively. Based on the above nota- tions, the pseudo code for carrying out the IE is presented in Fig. 7. The operations of the pseudo code can be summarized in the following. In the initialization step, we reset all the variables. In step 1, we check whether the incoming MV, mvk,p, matches any one of the stored distinct MVs, Dmvd. If it matches, we identify the matched distinct MV and output a flag new = 0; otherwise, we denote the incoming MV as a new distinct MV and set new = 1. In step 2, we compute the AMN of the identified distinct MV. In step 3, we check whether AMN_new of the distinct MV iden- tified previously greater than the wAMN of wMV and output a flag win = 1 if the result is positive; otherwise we set win = 0. In step 4, we update the distinct MV and the associated AMN and update wMV and wAMN if win = 1; furthermore, if the flag new resulted in step 1 is 1, we set D = D + 1 that is to increase the number of dis- tinct MVs by 1. The hardware implementation architecture of the pseudo code is presented in Fig. 8. In the leftest part of this architecture, we use a counter, counter_kp marked by (ii) in Fig. 8, to generate the index p = 1, . . . , P for each k = 1, . . . , K sequentially and use a KP-to-1 Multi- plexer, denoted by MUX and marked by (i) in Fig. 8, to select mvk,p based on the generated index; this corresponds to the statement smv = mvk,p in step 1 of Fig. 7. The smv will be input to the MV iden- tifier marked by (iii) in Fig. 8, which will compare the smv with the distinct MVs stored in the registers denoted by Distinct MV and AMN. The counter, counter_D marked by (iv) in Fig. 8, will generate the index D, such that if smv does not match any existing distinct MVs, this smv will be the Dth distinct MV, and the MV identifier will set the enable signal enD = 1 to activate the Distinct MV and AMN registers, marked by (v) in Fig. 8, to store the current smv and the corresponding AMN. In the meantime, the MV identifier also sets new = 1 to increase counter_D by 1. We will then use an adder marked by (vi) in Fig. 8 to update the AMN register for the current smv, which may be one of the existing distinct MVs or a new distinct MV, as follows. Add the MNk,p corresponding to the current smv to the AMN_old then output AMN_new, which will be fed back to the 3256 S.-Y. Lin et al. / Expert Systems with Applications 39 (2012) 3249–3256 Distinct MV and AMN registers. Notably, the MNk,p is generated by the counter_MN marked by (vii) in Fig. 8. Furthermore, the AMN_- new will compare with wAMN through the comparator, >, marked by (viii) in Fig. 8. If AMN_new > wAMN, we set win = 1, which will up- date wAMN by AMN_new and wMV by smv as shown by the blocks marked by (x) and (ix), respectively. This completes the implemen- tation of the pseudo code presented in Fig. 7. We design the clock period to be long enough such that the sig- nal mvk,p can travel through the critical path, which includes the propagation of counter_kp, KP-to-1 MUX, and MV identifier, the ac- cess of the Distinct MV and AMN registers, the processing time of adder and comparator, and the update of wMV and wAMN. Based on Taiwan Semiconductor Manufacturing Corporation (TSMC) 0.18 lm CMOS technology, the critical path’s delay of the IE pre- sented in Fig. 8 can be within 13.1 ns. That means we can design the clock period for IE as Tclock = 13.1 ns. Then, the time required for processing the IE is TIE = K � P � Tclock. Therefore, for K = 8, P = 3, the IE can obtain the true MV within 314.4 ns. References Chen, M. J., Gu, Y. B., Shen, W. C., Wu, T., & Hsu, P. C. (1998). A compact high-speed Miller-capacitance based sample-and-hold circuit. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 45, 198–201. Gharavi, H., & Mills, M. (1990). Block matching motion estimation algorithms new results. IEEE Transactions on Circuits and Systems, 37, 649–651. Hsieh, C. H., & Lin, T. P. (1992). VLSI architecture for block-matching motion estimation algorithm. IEEE Transactions on Circuits Systems Video Technology, 2(2). Liao, S. H. (2005). Expert systems methodologies and applications – a decade review from 1995 to 2004. Expert Systems with Applications, 28(1), 93–103. Li, H., & Sun, J. (2009). Majority voting combination of multiple case-based reasoning for financial distress prediction. Expert Systems with Applications, 36(3), 4363–4373. Li, R., Zeng, B., & Liou, M. L. (1994). A new three-step search algorithm for block motion estimation. IEEE Transactions on Circuits Systems for Video Technology, 4, 438–442. Panovic, M., & Demosthenous, A. (2004). A compact block matching cell for analogue motion estimation processors. Proceedings of 2004 IEEE International Symposium on Circuits and Systems (ISCAS’04) (Vol. 2, pp. 229–232). Canada: Vancouver. Panovic, M., & Demosthenous, A. (2006). Motion estimation processor using mixed- signal approach. IEEE Transactions on Circuits and Systems II, 53(6), 492–496. Razavi, B., & Wooley, B. A. (1992). Design techniques for high-speed, high-resolution comparators. IEEE Journal of Solid-State Circuits, 27(12), 1916–1926. Sun, J., & Li, H. (2008). Listed companies’ financial distress prediction based on weighted majority voting combination of multiple classifiers. Expert Systems with Applications, 35(3), 818–827. Yang, S., Wolf, W., & Vijaykrishan, N. (2005). Power and performance analysis of motion estimation based on hardware and software realizations. IEEE Transactions on Computers, 54, 714–726. Zhu, S. & Ma, K.-K. (1997). A new diamond search algorithm for fast block matching motion estimation. In IEEE international conference on communications and signal processing (pp. 292–296). Zhu, C., Lin, X., & Chau, L. P. (2002). Hexagon-based search pattern for fast block motion estimation. IEEE Transactions on Circuits Systems for Video Technology, 12, 349–355. Expert system based parallel multi-1D block matching algorithm with implementation for motion estimation 1 Introduction 2 Expert system based parallel multi-1D block matching algorithm (ESPM-1D-BMA) 2.1 Review of 2D-FSBMA 2.2 Motivation 2.3 The 1D block matching 2.4 Parallel multi-1D blocks matching 2.5 Expert system based parallel multi-1D blocks matching algorithm (ESPM-1D-BMA) 3 Implementation for motion estimation 3.1 Implementing 1D-block matching using mixed-signal approach 3.2 Implementing parallel multi-1D blocks matching 3.3 Implementing ESPM-1D-BMA 3.4 Hardware Implementation Architecture of ESPM-1D-BMA for ME 3.5 Time complexity 4 Test results and comparisons 4.1 Comparisons of ME accuracy 4.2 Comparison of computational efficiency 5 Conclusion Appendix A References 
work_unexgu5tqzez7pvos2fwoudg54	----	SEMA4AFinal SEMA4A: an Ontology for Emergency Notification Systems Accessibility A. Malizia a,1, T. Onorati a, P. Diaz a and I. Aedo a a Computer Science Dept., University Carlos III of Madrid, Avda. de la Universidad, 30, 28911-Leganés, Madrid, Spain Abstract Providing alert communication in emergency situations is vital to reduce the number of victims. Reaching this goal is challenging due to users’ diversity: people with disabilities, elderly and children, and other vulnerable groups. Notifications are critical when an emergency scenario is going to happen (e.g. a typhoon approaching) so the ability to transmit notifications to different kind of users is a crucial feature for such systems. In this paper we describe how ontologies are used to generate accessible notifications. An ontology was developed by investigating different sources: accessibility guidelines, emergency response systems, communication devices and technologies, taking into account the different abilities of people to react to different alarms (e.g. mobile phone vibration as an alarm for deaf blind people). We think that the proposed ontology addresses the information needs for sharing and integrating emergency notification messages over distinct emergency response information systems providing accessibility under different conditions and for different kind of users. Keywords: HCI, accessibility, ontology management, universal design, Emergency Response Information Systems (ERIS). 1 Introduction Within an emergency scenario sharing information and common knowledge about types of disasters, kinds of affected entities (people, infrastructures, communications,…), measures and alerts (depending on the kind of emergency) is crucial in order to reduce the number of victims or damages. Information technology (IT) is a relevant support when an emergency occurs, or is going to occur; furthermore IT tools like Emergency Response Information Systems (ERIS) can manage communications and information processing, can help in decision making and 1 Corresponding Author: Alessio Malizia, Universidad Carlos III, Dpto Informatica, Avda de la Universidad 30, 28911-Leganes, Spain. E- mail addresses: alessio.malizia@uc3m.es (A. Malizia), tonorati@inf.uc3m.es (T. Onorati), pdp@inf.uc3m.es (P. Diaz), aedo@ia.uc3m.es (I. Aedo) improve situational awareness when used in an emergency scenario. Technologies involved in ERIS can vary from mobile devices, to client-server applications, reaching complex distributed services architectures (Van de Walle, B. and Turoff, M., 2007). Nevertheless, one of the most relevant features of ERIS is alert notification since it can affect the citizens and their safety. Our approach consists of automatically adapting (with the use of our ontology) the notification of alerts to different kind of users (elderly, disabled, …) depending on the type of technologies (devices) they can access and considering the impact of the disaster on alerts communication and infrastructures. So for instance, when a fire is occurring in a building we know that we should communicate with people by audio notifications instead of visual ones since smoke can reduce visibility, but also that we could use the same kind of alert for compensating disabilities, e.g. alerting blind users of a critical or dangerous event. To provide emergency notifications is important but not trivial since in order to get an efficient communication many systems should interoperate with each other sharing a common knowledge with different terminologies and types of crisis or emergencies. From the semantics point of view, a common language (at least a glossary) is needed in order to have coordination among users, systems and communications. Thus, it is important to codify semantics in an accessible way so that it is easy for users to interpret notifications and for expert users to communicate among each other on relevant topics. To help in augmenting the interoperability among systems, and also among people, involved in such scenarios we developed an ontology2 called SEMA4A (Simple Emergency Alerts 4 [for] All) including concepts taken from emergency systems and control rooms but also from accessibility guidelines and interactive devices. The scope of this ontology is designing a common knowledge within ERIS and accessibility. Tim Bernes-Lee stated (Berners-Lee, 2007): “Disaster response is much about preparedness. If much relevant data is available in RDF, when a disaster strikes, those on the ground and across the world will be able to use it to know what best to do to respond”. That’s why we not only designed and developed our ontology integrating the categories of information described above, but also used the Web Ontology Language (OWL) to adhere to a standard codification and to offer an interoperable knowledge framework for enabling collaboration among different ERIS. The novelty of this 2 An ontology is defined (Neches, R., Fikes, R., Finin, T., Gruber, T,. Patil, R., Senatir, T., and Swartout W. R, 1991) as a set of basic terms and relationships of a vocabulary within an area or subject, as well as the rules to combine terms and relations that extend the vocabulary. work is both in providing accessible emergency notifications and supporting systems interoperability at syntactic and semantics levels by sustaining standard ontology codification and concepts mapping among different domains of accessibility and emergencies. In section 2 we will describe related works on HCI, accessibility and emergency response information systems. Successively, in section 3, we will present the proposed approach describing our ontology-driven methodology for semantic adaptation of emergency notifications to different types of users. Section 4 will be about evaluating the proposed ontology, while in section 5 we comment on obtained results and describe future works. 2 Background 2.1 HCI and Emergency Managing risk is crucial for Governments as well as for engaging stakeholders, and providing accurate information to allow public departments to make decisions on how to deal with risk (UK Resilience3). Providing effective communication of alerts and emergency situations is really important in order to reduce the number of victims. Reaching this goal is challenging because there is no effective standard for design due to different people characteristics: we considered users having different abilities, backgrounds, experiences, ages and sizes. During an emergency situation, all people have in some sense a kind of disability, some of them might be caused by stress, the environment or even by lack of information. Thus, it is important to design systems that effectively provide information to people included in vulnerable groups. This can be achieved by adopting Universal Design principles (Dix et al., 2003). As said by Information & Communications Technologies Standards Board, there are two ways to reduce the gap between products and human abilities: the Design for all approach and Assistive Technology (ICTSB, 2007). Aspects related to the implementation of Design for All in ICT from a developer oriented perspective have been also presented in Burzagli et al., (2009). 2.2 Accessibility Standards The problem of the accessibility is only partially related to adopting measures that 3 http://www.ukresilience.info/ compensate the disadvantages or that surpass the functional limitations of people with disabilities. In fact, adapting the environment to be accessible, not only is good for people with disabilities but also for those without disabilities, this is remarkable if we think about urban adaptation in big cities. One of the most important communication channels is the Internet and thus it is important to provide accessibility over the web. In order to ensure web accessibility, web developers and designers can follow guidelines established by the World Wide Web Consortium (W3C), through a special working group called Web Accessibility Initiative; these guidelines are called Web Content Accessibility Guidelines (WCAG) (WAI, 2006). It is also important to examine the approach taken by US, the so-called Section 5084, which aims at adapting information to people with disabilities. To verify the compliance of a specific web content to guidelines, developers may use accessibility checker software, such as: Bobby, LiFT, A-Prompt (Bobby 2006, LiFT, 2006, A- Prompt, 2006); this is not sufficient since these tools mainly perform a syntactic assessment of web pages, but they are not able to verify semantic contents and to adapt them for special needs users (Yang et al., 2007). We should mention that another interesting approach has been proposed by Gabrielli in (Gabrielli et al., 2004, Gabrielli et al., 2005), where authors provided an accessibility approach to e-learning for people with special needs. In their work, they highlighted a set of abilities together with a set of visual features by which learners, with different abilities, could interact with an e-learning system. We studied these difficulties since they are similar to the ones we found when adapting (or designing how to adapt) information to different kinds of users. 2.3 Emergency Systems In this section, we analyse information systems used to inform people about emergencies. These systems are called Emergency Notification Systems (ENS) and are included within Emergency Response Information Systems (ERIS). In these days, many emergency notification systems exist and are used to notify people in places like home, office, school, outdoor, etc. Notifications are often delivered by phone, e- mail or websites. Some systems also permit to deliver messages to pagers, faxes, VoIP (Voice 4 http://www.section508.gov/ over IP), SMSes (Short Message Service), and instant messengers; nevertheless, these are features that today are already included in mobile phones, smart phones or PDA (Personal Digital Assistant). In order to understand if existing ENS consider Accessibility principles, we studied the systems included in Table I. After surveying the systems presented in table 1, we found out that ENS are used by private companies, schools, government offices, Red Cross, fire-fighters, police, as well as many other institutions. Service technologies provided by these systems are almost the same (aee ‘Notification’ column). We found out that Waves Alerter was the only system providing accessible notifications; in particular for people with auditory deficiencies using TDD/TTY5, but this technology is old and non-standard. Therefore, even if ENS are intended to inform people about an emergency, we found out that these systems do not provide notifications in a format considering people’s profiles and preferences. This can be done by providing ENS with a model or base of knowledge (an ontology in our case) that reflects this information (users’ profile and preferences), as well as information related to accessibility, media and emergencies in order to provide effective and customized emergency notifications. System Web Comm. Type Source Notification Accessibility 3n http://www.3nonline.co m/ Emergency, Situational Alarm, Alert, System Status Phone, Web Phone, E-mail, Pager, Fax, SMS, PDA No AlertFind http://www.messageon e.com/crisis- communications/ Emergency, Situational Alarm, Alert Phone , Web Phone, E-mail, Pager, Fax, SMS, PDA No Arce https://arce.dei.inf.uc3 m.es/arce_demo/ Emergency, Situational Alarm, Alert, System Status E-mail, Web Web pages, E-mail No Command Caller http://www.voicetech.c om/Command_Caller_ 40.htm Emergency, Situational Alarm, Alert Phone, E-mail, Fax Phone, E-mail, Pager, Fax, SMS, PDA No RapidReach http://www.rapidreach. com/ Emergency, Situational Alarm, Alert Phone, Web Phone, Pager, Fax, SMS y E-mail No Sahana http://www.sahana.lk/ Emergency Web Web pages No Sigame http://www.sigame.es/ Emergency Web Web pages No 5 (Telecommunications Device for the Deaf/TeleTYpewriter) A user terminal with keyboard input and printer or display output used by the hearing and speech impaired (source: PCMag.com, http://www.pcmag.com/). SWN http://www.sendwordn ow.com/smart_alert_se rvice.aspx Emergency, Situational Alarm, Alert, System Status Phone, E-mail, SMS, Blackberry, Palm, Web Phone, E-mail, Pagers, SMS, MMS, VoIP, Skype, Chat y PDAs No WAVES Alerter http://www.madah.com /products/subpage.asp? mer_notf_sys Emergency, Situational Alarm, Alert, System Status Phone, Web Phone, E-mail, Fax, PDA and TDD /TTY Yes Table 1. Comparative survey of accessibility features in Emergency Notification Systems. Moreover, we have also taken into account the CAP (Common Alerting Protocol) 1.0 specification approved by the OASIS consortium. OASIS6 is the Organization for the Advancement of Structured Information Standards, and CAP is an XML-based data format for interchanging warnings and emergencies between alerting technologies. The scope of the CAP is focused on defining and exchanging the different kinds of alerts and types of notifications but does not take into account the different abilities of the users and does not, explicitly, model the relationships among the technologies and the kinds of emergencies. So we decided to use, directly, a structured knowledge in form of ontology to link all the dimensions of the alert’s notifications information space (types of notifications, users’ abilities to react or understand the notifications, available technologies, types of emergencies and impact of the emergency on the available technologies). This is exactly about supporting software interoperability since CAP is a standard way of communicating about emergency notifications within systems. We highlight that by integrating the CAP data structures into our ontology we also obtained interoperability at semantics level because we linked concepts already present in our ontology to the hierarchical structure extracted from the CAP. 3 Approach In order to deploy an accessible emergency notification system, our proposal focuses on developing an ontology that adapts notifications by using accessibility and usability concepts. This can be done by means of a knowledge base that reflects users’ needs as well as ways for effectively present them information according to their needs, the kind of emergency, and available technologies (depending on the users’ abilities and the type of emergency). Having such information base, assures us to be aware of stakeholders’ characteristics and needs; making it possible to reach more people. Designing such a complex system requires an articulated knowledge base that consists of 6 http://www.oasis-open.org/home/index.php knowledge in three areas: accessibility, user profiles, and devices. For these reasons we chose to model the knowledge base by an ontology which can include complex relationships among concepts and can provide tools like first order logic to verify the validity and integrity of the knowledge codified within it. Ontologies provide a semantic resource to describe information related to a specific domain. Ontologies will reduce the necessary effort to specify accessibility requirements for emergency notification systems because there will be a set of shared concepts that could be used for specifying requirements. An ontology-driven system could be developed on top of this codified knowledge that could be used to infer and adapt notifications in emergency scenarios depending on users’ special needs (e.g. a fire alarm in a building could cause the system to send a short video to deaf people showing the emergency exit locations to solve the problem of not hearing the alert sound). For these reasons, we studied different existing ontologies for accessibility and used information from these ontologies to render such concepts for adapting emergency notifications to different types of users. 3.1 Ontology and accessibility To clarify the way in which ontologies can be used in the field of accessibility, in this section we describe three works that are relevant to us and related to our approach. The first is the KAICO7 system, which uses the OntoSaw ontology, the second is the Dante tool that uses the Web Authoring for Accessibility (WAfA) ontology, and finally, we analise the AccessOnto8 architecture and ontology. OntoQuercus group, sub-group of Quercus Software Engineering Group (Quercus, 2007) of University of Extremadura developed the KAICO system, which is composed by an ontology as knowledge base, software applications that serve to add semantic tags to web pages elements, applications to extract information and specialised hardware to communicate with blind people (Lozano-Tello et al., 2003, Lozano-Tello et al., 2004). The OntoSaw ontology contains the conceptual model of elements that form web pages; the ontology is oriented to represent the necessary information so that the pages are accessible. With the purpose of identifying elements and attributes of the ontology, opinions from people 7 http://quercusseg.unex.es 8 http://shapevle.cant.ac.uk/AccessOntoTool.htm with visual disability were collected, together with observation of difficulties faced by them while visiting web sites with poor accessibility. In order to complement this ontology, elements and characteristics identified in WCAG 1.0 have been taken into account. This ontology has been constructed following the METHONTOLOGY methodology (Pinto, H., S., 2004), using Protégé-20009, and represented by the DARPA Agent Markup Language and Ontology Inference Layer or Ontology Interchange Language, the DAML+OIL10 language. Another relevant work comes from the Information Management Group of computer science department at University of Manchester, that has designed Dante11 (Yesilada et al., 2004), which is a semiautomatic tool that aims at improving navigation between web pages for people with visual deficiencies. The main objectives of Dante are: to analyze web pages in order to identify objects which improve the navigation, to translate into page elements the concepts from WAfA ontology, and to transform (transcode)12 the pages using these annotations so that they can be easily accessed using a screen reader. The WAfA ontology (Web Authoring for Accessibility13), is also known as the Travel Ontology because it is based on the analogy of tourists’ trips with web navigation; it represents concepts and relations necessary to automatically model the structural organization and navigation of web pages (Yesilada Y., 2005). In Dante, this ontology is implemented as a controlled vocabulary describing annotations and transformations. A different approach has been taken at the Businesses school at the Canterbury Christ Church University. AccessOnto has been developed (Masuwa-Morgan and Burrell, 2004) as a tool for requirements engineering managed as a repository of semantic requirements for accessibility (Guarino, 1997). The goal of this ontology is to extract a specification of requirements by using a knowledge base built on user’s characteristics. The repositories used by this system included two kinds of knowledge: declarative knowledge focused on a limited set of classes defined by the guidelines, interface objects and user’s characteristics, and procedural 9 Ontology Editor. http://protege.stanford.edu/ 10 http://www.daml.org/language/ 11 http://dante.cs.manchester.ac.uk/ 12 www.w3.org/1999/07/NOTE-annot-19990710 13 http://augmented.man.ac.uk/ontologies/wafa.owl knowledge composed of production rules capturing the ideas of adaptive programming and multiple relations management (entities and dependencies). The guidelines included in AccessOnto are proposed by: WCAG, Sun Micro System, IBM, Microsoft and Apple. Nevertheless, the intention is that in the future the repository will point directly to the sources of guidelines. At the moment, the problem is that existing guidelines do not use a standard format, which makes it difficult to integrate guidelines from different sources. Our purpose was to explore if ontology can be used to specify accessibility requirements so that it was feasible to create an emergency notification system that used this knowledge to decide the optimal media and format to adopt for notification messages. The three described approaches are a useful foundation for our work but there are several constraints: 1. The Kaiko idea is to add semantic marks to Web pages elements in order to be able to present them to blind people. It is based on OntoSaw ontology that contains elements to model Web pages, attributes and their relations. Some limitations of this approach are: • Generated code can present problems in being interpreted by some screen readers and browsers. • The ontology is focused only on blind people. 2. Dante aims at improving the navigation among web pages for people with visual deficiencies. Dante is based on the WAfA ontology, which contains concepts and relationships needed to model organization, structures and navigation of web sites. WAfA defines concepts about the rendering of objects in a web page (structural properties) and how these objects are used, in other words, WAfA encapsulates extensive knowledge to make explicit structural and navigation information of a web page. This approach is more complex than the one made by Kaiko, nevertheless it presents some problems: • Existing pages are transformed into small fragments that could cause the loss of users’ context. • Screen readers could have difficulties reading the web pages annotations. For this reason it might occur that screen readers present wrong information to users or cannot be able to interpret the entire code. • The ontology is only focused on people with visual disabilities. 3. AccessOnto is a requirements engineering tool in form of an accessibility requirements repository. This ontology is made of three information repositories: user profile repository, guidelines repository and interface object-action repository. AccessOnto also presents some weak points: • Structural information given by previous ontologies is not included. • The ontology is at an early stage, reason why it is not formalized using a standard language like OWL. After reviewing the existing approaches, we can point out that ontologies will clearly help the development of emergency notification systems by fulfilling accessibility and universal design principles; nevertheless there is a strong need of integrating or mapping information among them in order to take advantage of the various ontologies characteristics. The current version of SEMA4A includes information related to content design, accessibility guidelines, emergencies, devices and communication technologies organized in three main categories: (1) WAfA; (2) AccessOnto; (3) EMEDIA. Below we explain what information is included in each section, as well as the process followed in order to model and integrate this information. 3.2 The Proposed Ontology Designing and developing ontologies is a complex task involving knowledge management and domain experts (Pinto, H., S., et al., 2002). Ontologies have proliferated in these last years mainly thanks to the semantic web development. They have been also widely adopt in other areas like e-learning for adaptation purposes (Cristea, A. I., 2004, Pattuelli, M., C., 2008). Having this in mind, we developed the SEMA4A ontology, which has been created using the Web Ontology Language (OWL) as well as an OWL reasoning tool from Mindswap laboratory at Maryland University called Pellet to verify the consistency of existing classes in the ontology. Our purpose was to model a set of sharable concepts and knowledge considering a successive step the rigorous formalization and the development of an expert system based on the proposed ontology as suggested by Geller in (Geller, J., et al., 2004). The first part of the proposed ontology is EMEDIA (Emergency and MEDIA technologies); which is the portion of the SEMA4A ontology that provides concepts and relations about emergency and media technologies. We developed it through a semiautomatic procedure with two phases: the first phase was performed to extract concepts and relations concerning emergency and media technologies; the second one consisted of integrating new information within the existing ontology (adding relations with the others portions). We applied this technique to develop and expand the parts of our ontology related to the emergencies; we have also modelled how technologies affect accessibility. The first phase of our procedure consisted of, automatically, extracting concepts and relations from WordNet (Miller, G., A., et al., 1990). The starting point was a simple set of words, related to emergency and media technologies, found in MyFlorida.com (The Official Portal of the State of Florida. MyFlorida.com - Taxonomy - Disasters & Emergency Information14) and A Simple Taxonomy for Mobile Emergency Announcement Systems (Addams-Moring, R., et al. 2005), as suggested by emergency field experts (Spanish civil protection). We proceeded, initially, retrieving these terms in WordNet. For each meaning of each term, WordNet gave a set of synonyms, holonyms, hypernyms, hyponyms and meronyms. Synonyms represented new concepts that could be possibly added to the ontology. Examples of concepts used as a seed for retrieving related knowledge are: Avalanche, Blizzard, Colds, Cyclone, Drought, Earthquake, Epidemic, Eruption, Fire, Flood, Forest fire, Hailstorm, Heat wave, Hurricane, Ice storm, Lahars, Landslides, Limnic eruption, Maelstrom, Mudslide, Seiche, Sinkholes, Storm, Thunderstorm, Tornado, Tsunami, Typhoon, Volcano, and Wildfire. We iterated this procedure with all synonyms found as in an n-ary tree: each concept was a node having a child for each related synonym until a maximum of three levels (this threshold has been experimentally set; fewer levels generated few terms, while more levels added terms which were not really related to our domain). For each concept we stored all meanings and relations. In this way, we obtained a new taxonomy related to emergency and media. The final phase consisted of integrating the new taxonomy in the existent ontology as a new class, EMEDIA. After creating EMEDIA we extracted, refined, upgraded and linked information obtained from: (1) an ontology that contains concepts and relations needed to model the organization, 14 http://www.myflorida.com/taxonomy/floridian/disasters%20&%20emergency%20information/ structure and navigation of information contents (WAfA); (2) an ontology that includes accessibility guidelines, user’s profiles and actions that can be performed by users with different abilities (AccessOnto). In particular the basic classes included in this combination of ontologies are: Impairment, Age, and Expertise. Impairments class includes: Motor (coordination difficulty, reach limitations, no tactile sensation), Visual (color blindness dichromatic and color tones, low vision, blindness, deaf blind), Cognitive (word and spatial dyslexia, learning difficulty), Hearing (deafness, deaf blind). Age is subdivided in children and elderly (with their combination of disabilities and characteristics). Expertise is subdivided into: novice, intermediate, and expert; it includes all kind of disabilities originated by the level of computer education users have. We combined these two existing ontologies by including and mapping their concepts and relations into our ontology together with the EMEDIA part. We used different transformations and semantics paths to link related concepts in these that before were separated ontologies, thus creating a common knowledge base on accessibility guidelines and users’ characteristics and abilities. Summarising, SEMA4A counts on three basic classes: WAfA, AccessOnto and EMEDIA; it includes information related to concepts and relations needed to model organization, structure and navigation of information contents; it also includes accessibility guidelines, user’s profiles and actions that users can perform, as well as, information related to emergencies, notifications and devices. These main classes are linked with relations existing within their subclasses. In the next section we provide a use case that depicts the more common relations that exist in our ontology. Figure 1 shows the high level classes that form the SEMA4A ontology. Fig.ure 1. SEMA4A ontology fragment. 3.3 Use Case In order to show how the concepts and relationships included in SEMA4A could be used by ENS for providing accessible alerts notifications, we provide a use case. Imagine a deafblind person walking alone on the streets in a city that he/she is visiting for the first time. Before arriving the city, this deafblind has been subscribed to an ENS. Weather forecast for the city, where the deafblind person is touring, expects that a tornado arrives in around 6 hours. If the ENS wants to alert the person about this event, it is important to provide significant information in a format that he/she can access. When this person subscribed to the service, he/she communicated that the media device he/she carried was a Personal Digital Assistant (PDA) with Internet connection. This person also pointed out that he/she had a special program installed in his/her device for reading the screen (screen reader) that transforms text and images into Braille writing system. The ENS could use SEMA4A for alerting the deafblind person in an accessible way of the upcoming event taking into account his/her preferences and profile information as follows: • From the ontology, we have that deafblind is a visual and hearing impairment where people cannot hear or see either partially or totally; we also have that people with this disability may have difficulties using a standard mouse (or a pointing device like the pen used with PDAs). It is common that deaf blind people use speech input and speech output, sometimes with difficulties; they can also use tactile input and output (e.g. Braille line, which is a Braille display). It is also obtained from our ontology that deaf blind people can use keyboards, and can notice vibrations. According to the actual version of SEMA4A, deaf blind people can access text data using a Braille line. Figure 2. Definition of deaf blind. • As the ENS needs to alert a deaf blind person using a PDA with Internet access, about an approaching tornado, it obtains from our ontology that PDA (Personal Digital Assistant) is a media device that can communicate information contained in figures, sounds, text, as well as vibration signals (see Figure 3). From our ontology it is also obtained that tornado is an emergency that can be communicated using the Internet, TV and radio (see Figure 4). Moreover, SEMA4A defines that when a tornado alert should be communicated, this must be done following the Common Alerting Protocol (CAP) (see Figure 5). Figure 3. PDA communication features. Figure 4. Tornado and media that can be used to alert about. Figure 5. Common Alerting Protocol description. • According to the user’s profile and preferences, it is desired that the ENS notifies via the Internet. Having this in mind, from the ontology we know that using the Internet we can communicate employing multiple languages, text, figure, video, sound or emails, as shown in Figure 6. Figure 6. Internet communication capabilities. • In order to assure that this person can access the information, SEMA4A infers ontology fragments, as shown in Figure 7, to follow some guidelines from: Web Accessibility Initiative (WAI), Accessibility Quick Reference Guide, Custom Guidelines, Neuman’s guidelines, as well as guidelines from IBM relative to adapt content for blindness. Figure 7. Web accessibility guidelines. • Finally, SEMA4A infers Web accessibility guidelines specific for deafblind, as well as specific Web elements (e.g. figures or images, sounds, input controls) that need to be formatted or transformed for assuring that this deafblind person can access the tornado alert notification over his/her PDA in order to save his/her life (Figure 8). For instance images descriptions (which can be read by a text-reader software) replacing graphics. Figure 8. Gidelines for deafblind. 4 Ontology Evaluation 4.1 Evaluation Criteria There exist many different methods and techniques to validate and evaluate ontologies as already described in the introduction. We used an approach inspired by Spyns et al. (Spyns, P., Meersman, R. and Jarrar, M., 2002) based on triples extracted from the ontology which were defined as lexons. Formally, a lexon is as 5-tuple: <(G,L): term1 role co-role term2>. Where G is the context and L is the language. Co-role is the inverse of role; we can omit the (G,L) pair and describe a lexon as a 4-tuple like <President, directs, is directed by, enterprise>. Usually only the role is explicitly represented (while the inverse is implicit), thus a lexon is described a triple, and in fact could be described as a combination of an OWL triple and its inverse, or as a conceptual graph style relation (Sowa, J., F., 1984). Informally we can say that a lexon expresses that the term1 (or head term) can have term2 (or tail term) occur in an associating role with it. Conceptualisations can be represented in terms of lexons. We can represent our ontology in form of a list of lexons. The main research hypothesis in this evaluation is that lexons, representing the basic binary facts (contained in a corpus of documents or in ontology) expressed in natural language, can be extracted from the available textual sources, i.e. a corpus, using a specific parser, as well as, from our ontology as OWL-triples. In fact in a specific domain noun phrases (NP) carry important information about the domain itself, while verbs impose restriction on the nouns semantics. The considered corpus is composed by articles about emergency, accessibility and devices as suggested by the domains experts. In particular, for the emergency topic there are about sixty- seven article from the proceedings of the conference ISCRAM2007 (Intelligent Systems for crisis management), plus a manual developed by the North Central Texas regional government (Know what to do. Think. Prepare. Act.15) and papers on community emergency management (for example: Schafer, W.A., et. al, 2007). The total number of analysed words was about three-hundred thousand for five-hundred pages. For accessibility and supported devices, we considered the Web Content Accessibility Guidelines 1.0 with W3R Recommendation of 5 May 1999 and twenty-four articles from www.webaim.org (Web Accessibility in Mind). We extracted all words from the corpus and applied a tagging procedure for the analysis of corpus texts. We analysed the input texts (the whole corpus in textual form) and tagged each word with its syntactic function (nouns, verbs, adjectives, etc.). We considered only nouns and verbs at this stage and reduced them to the root form (singular nouns and verbs at infinitive). This normalised corpus has been used to help in generating the lexons validation lists submitted to the evaluators (domain experts) and to act as a corpus to be quantitatively evaluated against lexons extracted from our ontology. In fact, we could compare lexons extracted from the corpus with lexons obtained from our ontology. Successively, we applied a second iteration where we filtered our ontology by using feedbacks of the evaluators and contrasted it against the corpus to verify the accuracy of the contained information. 4.2 Qualitative Method Domain experts were asked to evaluate the value and usefulness of the ontology extracted lexons in building a base knowledge for his/her domain. In fact, according to the HCOME Methodology (Kotis, K., et al., 2006) knowledge workers should participate actively in the ontology engineering processes. We selected two evaluators: one is an expert of accessibility who worked several years for R&D projects; she is particularly expert on Infometrics (information measurement) applied to web accessibility. The second evaluator is an expert professional working for Spanish Civil Protection and developing documents, policies and recommendations on the emergency domain. Questions have been asked in form of a short evaluation questionnaire associated to the 15 http://www.knowwhat2do.com/en/pdf/KnoWhat2Do_Guide.pdf lexons list extracted from the ontology and for the respective domains. We reduced the number of presented lexons to hundreds of terms by matching the lexons contained in the ontology with the ones also present in the normalised corpus. This is due to the fact that a human evaluator can be prompted with hundreds of terms and not with thousands (used in the quantitative evaluation). It is important to notice that, to our knowledge, there are very few documents on both the fields, so about emergency and accessibility together, while our ontology includes concepts from both domains. Due to these reasons we generated two different lists, one with emergency related lexons evaluated by the emergency expert and another one with accessibility lexons evaluated by the accessibility expert. The evaluator as a domain expert (professional in its field with years of experience), was prompted with three criteria to rate the lexons extracted from SEMA4A ontology. • Coverage (have all the lexons to be discovered actually been discovered?) • Precision (are the lexons making sense for the domain?) • Accuracy (are the lexons not too general but reflecting the important terms of the domain?) Which specifically were translated in the following questions asked for each presented lexon: • A. Is the lexon in the domain? • B. What is the level of precision of the lexon? • C. Does the lexon make sense for the specific domain? We assigned different discrete values to the possible answers: 0/1 for yes/no to the questions of type A; 0/1 for yes/no to the questions of type B; and 1/2/3 for specific/not too specific/general to the questions of type C. The expert on accessibility evaluated, totally, 155 lexons extracted from our ontology in the accessibility domain. Results were: • Coverage 91% (have all the lexons to be discovered actually been discovered?) • Precision 84% (are the lexons making sense for the domain?) • Accuracy 79% (are the lexons not too general but reflecting the important terms of the domain?). With Accuracy corresponding to “specific”; the expert also rated a 9% of “not too specific”, and a 12% of “general” lexons. The expert on emergency evaluated, totally, 265 lexons extracted from our ontology in the emergency domain. Results were: • Coverage 66% (have all the lexons to be discovered actually been discovered?) • Precision 65% (are the lexons making sense for the domain?) • Accuracy 45% (are the lexons not too general but reflecting the important terms of the domain?). With Accuracy corresponding to “specific”, while having a 1% of “not too specific”, and a 54% of “general” lexons. These results were mainly due to the fact that the emergency portion of our ontology was automatically built by extracting relevant information from corpus of documents suggested by experts; while the accessibility part was built by integrating ontologies that were already verified and cleaned. Figure 9 shows the average scores on experts’ evaluations for accessibility and emergency components of SEMA4A. Figure 9. Average scores on Coverage, Precision and Accuracy as evaluated by experts on accessibility and emergency over lexons contained in the SEMA4A ontology. In the next session, we will show how we used the qualitative results to improve the ontology and thus increase accuracy and coverage when measuring these characteristic with a quantitative method. 4.3 Quantitative Method In a previous paper (Malizia et. al., 2008), we defined a quantitative measure and semi- automated evaluation procedure for measuring coverage and accuracy of lexons over an entire corpus of documents. The procedure was inspired by Zipf’s law (Zipf, G. K., 1949). The idea is that the frequency of occurrence of a word in a corpus of documents is inversely proportional to its frequency class; it means that words with higher frequency are less meaningful for the corpus domain than words with less frequency (even if words with a low frequency can be too peculiar to be relevant for a domain within a corpus of documents). Coverage has been measured by counting for each frequency class the number of lexon terms contained in the ontology that are identical with terms extracted from the corpus and comparing this number to the overall frequency class term count. Accuracy has been estimated on the basis of the coverage percentage for an interval of frequency classes. As the SEMA4A lexons consist of three words (two terms and one role), it is possible to investigate how much the produced lexons cover the corpus vocabulary, and more importantly how accurate they are. Regarding the accuracy, determining exactly which frequency classes contain the terms most characteristic for a domain still depends mainly on intuition and subjective opinions. It should also be point out that no stopword list has been defined because lexons have been produced extracting nouns and verbs. Figure 10. Coverage and accuracy of frequency classes of the ontology over the corpus. Lemmas are the terms contained in each group of lexons (2 for each lexon). In a previous preliminary work we presented the graphic representation in Figure 9 where the lowest (<9) and highest (>250) frequency classes were omitted since did not contain relevant words (they contained too specific or too general terms for being relevant to display). Figure 10 shows that the coverage improved with the increasing rank of the frequency classes (until FC=250 to have a clear view of the graph). On average, the coverage ratio was 32.56%. The accuracy (i.e. the coverage percentage for the selected interval) ratio for the 9-250 intervals was 42.24%. This last phenomenon was probably due to the fact that our corpus was made of documents about the different accessibility and emergency domains but including few existing documents on both domains and that had possibly reduced the specific terms but supported more general terms used in the technical domain senses. In this work we introduced the qualitative evaluation of experts and so we used these evaluations to filter our ontology and clean it from less specific or ambiguous terms that dropped down the accuracy and coverage of our previous approach (shown in Figure 11). We filtered out from our ontology terms that where judged too general or not specific for our domains by experts. Terms like “Link_location_attribute” or “Link_AccessKey” were judged too peculiar by accessibility experts (even if included in guidelines imported in our ontology); while terms like “novice” were too general for the accessibility domain. From the emergency point of view experts judged too general terms like “attention” or “removal”, while terms like “finite_quantity” or “unfortunate_person” were too peculiar for the emergency domain. So, after this filtering phase, we ran again our evaluation procedure onto the same corpus but with reduced set of concepts (many terms and thus lexons were filtered out following the qualitative evaluation). Thus, we had exactly the same ontology as before but filtered with the experts evaluation of too specific or too peculiar terms. Figure 11 shows that the coverage improved with the filtered rank of the frequency classes (until FC=250 to have a clear view of the graph). On average, the coverage ratio was 70.04%. The accuracy (i.e. the coverage percentage for the selected interval) ratio for the 1-100 intervals was 74.42%. Figure 11: Coverage and accuracy of frequency classes of the ontology over the corpus after filtering the terms by using the experts’ evaluation. Lemmas are the terms contained in each group of lexons (2 for each lexon). The frequency classes are lower than in the previous experiment since this time, after filtering, we have words that were much more relevant for the domain and so for the Zipf law they were all compressed within lower frequency classes; while, again for the same law, the number of lemmas was higher. 5 Conclusion When critical events can occur people has to be informed with complete and understandable information to reduce the damages or to inform about what measures can be taken for peoples’ safety. In fact, as stated by Spanish Civil Protection: “Everywhere, at every time an unexpected critical event can occur damaging people or things” (DGPCE, 2007). We showed that emergency information systems generally do not include information and knowledge about different kind of users to be notified about an emergency; there is a lack in taking into account differences in accessing the information sources (e.g. Internet or radio) or available resources depending on the cognitive and physical abilities of people. Providing accessible information within emergency notification systems is crucial, since, it can reduce the number of victims and strongly help users in receiving emergency and critical news. To solve this problem of communicating emergencies and critical information to different categories of users (impaired, aged, …) within different kind of emergencies using different technologies, we have developed a knowledge base codified as an ontology. We have also explored accessibility in general and investigated on broad guidelines and users’ experts guidelines coded as ontologies. We selected two ontologies: WAfA, and AccessOnto considering them as the most relevant for codifying accessibility concepts in emergency scenarios. We, then, developed another portion of our ontology called EMEDIA with a semi- automated technique including information on emergencies, and how they could affect technologies. Starting from the three portions we have developed a new ontology called SEMA4A including concepts and relationships considering different users profiles and abilities in conjunction with different communication medias (EMEDIA) and accessibility guidelines. The novelty of our approach consists of providing accessible notifications within ERIS. Furthermore we also included interoperability features both at syntactic level (supporting OWL and standard languages) and at semantics level by mapping and linking concepts coming from different fields (accessibility, emergency, multimedia) into the same ontology. We used a qualitative evaluation to test whether our ontology included the most relevant concepts in the domain of ERIS and accessibility. We obtained results from experts in both domains (emergency and accessibility) verifying that we have enough coverage on both the fields; thus, our ontology could be employed for developing systems that could infer information to automatically adapt emergency notifications depending on the recipients’ characteristics. Furthermore we employed a quantitative evaluation technique measuring the coverage and accuracy of our ontology over a corpus of documents (on accessibility, devices and emergencies) suggested by domain experts. We may notice that even if we had around 70% of coverage and 74% of accuracy, the corpus was mainly separated into two topics and only few documents shared the topics of accessibility and emergencies (this is exactly the lack of information we are trying to fill with our work). Nevertheless with the help of experts’ evaluation we improved the coverage and accuracy of our ontology with respect to our previous work. Future works include the possibility of integrating the ontology within an emergency system to test whether effective notification can be generated by an event-driven process refined by the knowledge base contained in the ontology to inform users according to their devices and abilities. In fact, the knowledge codified in SEMA4A can be used at different levels: • At static level as a set of facts and reusable content made by metadata (owl files) and possibly generating shared knowledge in for of HTML or XML files that could be used by different systems to interoperate. • At dynamic level as a knowledge base for developing web services (or semantic web services) able to query the knowledge base and derive actions to perform depending on the information codified in SEMA4A. Moreover we will automate such process by developing a fine-grained level in the ontology to exactly match users’ abilities with accessibility guidelines and new interactive media features, such as: touch sensitive, tactile, force-feedback (force/resistance), texture, heat, vibration, etc. We are also considering the evolution of the ontology when used by knowledge workers in emergency domain. The evolution of the ontology can be related to the evolution of its schema, structure, and the introduction of updated knowledge (Noy, N., F., et al., 2004). Finally we aim to validate our ontology within a wider audience of domain experts testing it within international organizations. In fact, we will use our ontology integrating it within a system (SIGAME16) managing real emergencies as a test bed among a community of users interoperating for solving emergency scenarios. Acknowledgment THIS WORK HAS BEEN PARTLY FUNDED BY UIA4SIGE (MINISTERIO DE EDUCACIÓN Y CIENCIA TSI2007-60388) AND MDDSIGE (CAM-UC3M CCG06-UC3M/TIC-0787) PROJECTS. AUTHORS ALSO WANT TO ACKNOWLEDGE F. ASTORGA-PALIZA FOR HIS COMMENTS AND PREVIOUS COLLABORATION. REFERENCES A-Prompt 2006. University of Toronto, A-Prompt, retrieved March 14, 2006 from http://aprompt.snow.utoronto.ca/. Abián M. A. 2005. Ontologías: qué son y para qué sirven. Web Semántica Hoy. http://www.wshoy.sidar.org/index.php?2005/12/09/30-ontologias-que-son-y-para-que-sirven. Addams-Moring,, R. Kekkonen, M. and Zhao, S. 2005. A Simple Taxonomy for Mobile Emergency. Proceedings of the 2nd International ISCRAM Conference (B. Van de Walle and B. Carlé, eds.), Brussels, Belgium, April 2005; pp. 309-316. Berners-Lee,T., (2007), Disaster Management Ontologies?. [Online] http://lists.w3.org/Archives/ Public/semantic-web/2007Apr/0094.html (Retrieved June, 2007). Bobby 2006. Watchfire Corporation, Bobby, retrieved March 14, 2006 from http://bobby.watchfire.com/. 16 https://www.sigame.es/ Borst, W.N. 1997. Construction of Engineering Ontologies. PhD thesis, University of Twente, Enschede, NL Centre for Telematica and Information Technology. Burzagli, L., Emiliani, P., Gabbanini, F., 2009. Design for All in action: An example of analysis and implementation. Expert Systems with Applications, 36 (2) Part 1, March 2009, Pages 985-994. Calero, C., Ruiz, F., AND Piattini, M. (Eds.) 2006. Ontologies for Software Engineering and Software Technology. ISBN: 10 3-540-34517 Springer. Carver, L., and Turoff, M. 2007. Human-computer interaction: the human and computer as a team in emergency management information systems. Commun. ACM 50, 3, 33-38. Cristea, A. I. What can the Semantic Web do for Adaptive Educational Hypermedia? International Peer- Reviewed On-line Journal "Educational Technology and Society" 7(4), Special Issue on Ontologies and the Semantic Web for E-learning, IEEE, LTSC, pp 40--58, 2004. Dix, A., Finley, J., Abowd, G., and Beale, R. 2003 Human-Computer Interaction (3rd Ed.). Prentice-Hall, Inc. DGPCE - Dirección General de Protección Civil y Emergencias, spain 2007. Conocimientos generales sobre protección civil. (Retrieved April 2007). http://www.proteccioncivil.org/centrodoc/publicaciones/conogralpc.pdf Geller, J., Perl, Y., Lee, J., (2004), “Editorial: Ontology Challenges: A Thumbnail Historical Perspective”, Knowledge and Information Systems 6(4): 375-379 (2004). Gabrielli, S., Mirabella, V., Kimani, S., AND Catarci, T. 2004. Steering the Development of Accessible e- Learning Content. In Proc. of the 3rd ECEL Conf., Paris (F), 517-526. Gabrielli, S., Mirabella, M., Teso, M., AND Catarci T. 2005. The Design of an Authoring Interface to Make eLearning Content Accessible. INTERACT 2005: 1100-1103. Gómez-Pérez, A. 1999. Ontological Engineering: A State Of The Art. Expert Update. Vol 2 Nº 3. Guarino, N. 1997. "Understanding, building, and using ontologies: A commentary to Using Explicit Ontologies in KBS Development", by van Heijst, Schreiber, and Wielinga. International Journal of Human and Computer Studies 46: 293-310. Hartmann J., Spyns P., Giboin A., Maynard D., Cuel R., Suárez-Figueroa M.C., and Sure Y., (2004), “Methods for ontology evaluation”, Knowledge Web Deliverable D1.2.3 ICTSB 2000. The ICT Standards Board Project Team. Design for all, final report. http://www.ictsb.org/. Kotis, K., Vouros, A., (2006), “Human-centered ontology engineering: The HCOME methodology”, Knowledge and Information Systems 10 (2006) 109–131. LiFT 2006. UsableNet, LiFT, retrieved March 14, 2006 from http://www.usablenet.com/. Lozano-Tello, A., Macías, M., Sánchez, F., and Sosa, E., (2003). “Uso de Ontologías en Páginas Web para Mejorar su Accesibilidad a Invidentes”. VIII Jornadas de Ingeniería del Software y Base de Datos (JISBD’03), págs 625-634. Alicante, Spain. Lozano-Tello, A., Macías, M., Prieto, A., Sánchez, F., and Sosa, E., (2004), “Contenidos Web Accesibles a Invidentes Meditante Ontologías”. II Congreso Virtual "Derecho y Discapacidad en el Nuevo Milenio". Cáceres, 17-18 Dec. 2004. Luhn, H. P., (1958), The automatic creation of literature abstracts. IBM Journal of Research and Development, 2(2):159 – 195, 1958. Malizia, A., Astorga-Paliza, F., Onorati, T., Díaz, P., and Aedo, I. 2008. Emergency Alerts for all: an ontology based approach to improve accessibility in emergency alerting systems. To appear in Proceeding of the ISCRAM 2008. Masuwa-Morgan, K. R., and Burrell, P. 2004. Justification of the need for an ontology for accessibility requirements (Theoretic framework). Interacting with Computers. Volume 16, Issue 3, June 2004, Pages 523- 555. Miller, G., A., Beckwith, R., Felbaum, C., Gross, D., and Miller, K., (1990), "Introduction to WordNet : An On- line Lexical Database", International Journal of Lexicography, Vol. 3, No. 4, 1990, 235 - 244. Neches, R., Fikes, R., Finin, T., Gruber, T,. Patil, R., Senatir, T., and Swartout W. R. 1991. Enabling Technology for Knowledge Sharing. AI Magazine, 12(3), pp 36-56. Noy, N., F., Klein, M., C., A., (2004), “Ontology Evolution: Not the Same as Schema Evolution”, Knowledge and Information Systems 6(4): 428-440 (2004). Pattuelli, M., C., (2008), “Teachers’ perspectives and contextual dimensions to guide the design of N.C. history learning objects and ontology”, Information Processing & Management 44(2): 635-646 (2008). Pinto, H., S., and Martins, J., P., (2002), “Ontologies: How can They be Built?”, Knowledge and Information Systems, 6(4):441 – 464, 2002. Plessers, P., Casteleyn, S., Yesilada, Y., De Troyer, O., Stevens, R., Harper, S., and Goble, C. 2005. Accessibility: a Web engineering approach. In Proceedings of the 14th international Conference on World Wide Web (Chiba, Japan, May 10 - 14, 2005). WWW '05. ACM Press, New York, NY, 353-362. QUERCUS 2007. Software Engineering Group. http://quercusseg.unex.es. Schafer, W.A., Ganoe, C.H., and Carroll, J.M. (2007), Supporting Community Emergency Management Planning through a Geocollaboration Software Architecture. Journal of Computer Supported Cooperative Work (CSCW). Volume 16, Numbers 4-5.pp. 501-537. Springer, October, 2007. Sowa, J., F., (1984), Conceptual Structures: Information Processing in Mind and Machine. Addison- Wesley, 1984. Spyns, P., (2005), Evalexon: Assessing triples mined from texts. Technical report09, STAR Lab, Brussels, Belgium, 2005. Spyns, P., Meersman, R. and Jarrar, M., (2002), Data modelling versus ontology engineering. SIGMOD Record Special Issue, 31 (4): 12 - 17, 2002. Telefónica 2005. Comunicación para todos. Pautas para la comunicación accesible. http://www.telefonica.es/manualdecomunicacion/index.html Van de Walle, B., and Turoff, M. 2007. Emergency Response Information systems: Emerging Trends and Rechnologies. Commun. ACM 50, 3, 29-31. WAI 2006. Web Accessibility Initiative. http://www.w3.org/WAI/ Yang, S. J. H., Shao, N. W. Y., 2007. Enhancing pervasive Web accessibility with rule-based adaptation strategy. Expert Systems with Applications, 32 (4), May 2007, Pages 1154-1167. Yesilada, Y., Harper, S., Goble C., AND Stevens R. 2004. Screen Readers Cannot See (Ontology Based Semantic Annotation for Visually Impaired Web Travellers). In Nora Koch, Piero Fraternali, and Martin Wirsing, editors, Web Engineering - 4th International Conference, ICWE 2004 Proceedings (LNCS 3140), pages 445-458. Springer, 2004. Yesilada Y. 2005. Annotation and transformation of Web pages to improve mobility for visually impaired users. PhD Thesis. School of Computer Science at the University of Manchester, UK. Zipf, G. K., (1949), Human Behaviour and the Principle of Least-Effort. Addison-Wesley, Cambridge MA, 1949. Appendix A: SEMA4A Classes In this appendix we give a brief description of the main classes and concepts included in the ontology to give an overall idea of the specific knowledge codified in SEMA4A. The main classes, as presented in Figure 1, are: EMEDIA (including concepts and relations on emergency and communication devices and technologies), AccessOnto (including concepts on disabilities and accessibility guidelines), and WAFa (including structural definitions for formatting content on the web, to which usually guidelines are applied),. EMEDIA EMEDIA (Emergency and MEDIA technologies) is the portion of the SEMA4A ontology that provides concepts and relations about emergency and media technologies. We developed it trough a semiautomatic procedure with two phases: the first phase was performed to extract new concepts and relations from WordNet (concerning emergency and media technologies); the second to integrate new information within the existing ontology (adding relations with the others portions). We applied this technique to develop and expand part of our ontology related to the emergencies and how they can affect technologies accessible to the users. The considered corpus is composed by articles about emergency, accessibility and devices as suggested by the domains experts. In particular, for the emergency topic there are about sixty- seven article from the proceedings of the conference ISCRAM2007 (Intelligent Systems for crisis management), plus a manual developed by the North Central Texas regional government (Know what to do. Think. Prepare. Act.17) and papers on community emergency management (for example: Schafer, W.A., et. Al, 2007). For accessibility and supported devices, we considered the Web Content Accessibility Guidelines 1.0 with W3R Recommendation of 5 May 1999 and twenty-four articles from www.webaim.org (Web Accessibility in Mind). The total number of analysed words is about three-hundred thousand for five-hundred pages. 17 http://www.knowwhat2do.com/en/pdf/KnoWhat2Do_Guide.pdf WAfA Web Authoring for Accessibility (WAfA) is an existing ontology also known as Travel Ontology because it is based on the analogy of web navigation with tourists’ trips. This ontology represents concepts and relations necessary to automatically model the structural organization and navigation of web pages (Yesilada, Y, 2005) to users’ profiles. This ontology has been evaluated with real users, contains information on how to model content for being accessible, and it is codified using OWL; we extended our ontology including WAfA concepts, defining a class called WAfA that contains concepts and relations needed to model organization, structure and navigation of sites. AccessOnto AccessOnto in an ontology in form of an accessibility requirements repository from which it is possible to extract requirements using an accessibility knowledge base (AKB) built on user’s characteristics (Masuwa-Morgan, K. and Burrellb, P, 2004). It includes guidelines from Web Accessibility Initiative, Sun Micro Systems, IBM, Microsoft, and Apple guidelines. In our ontology we created a class called AccessOnto that contains information related to Web accessibility guidelines, users’ profiles and actions that users can perform. We created this class translating information from XML (AccessOnto is codified in XML) to OWL; after this phase, we established relations that linked concepts contained in WAfA and in AccessOnto sections, as we will show in the use case section. Summarizing, SEMA4A counts on three basic classes: WAfA, AccessOnto and EMEDIA; including information related to concepts and relations needed to model organization, structure and navigation of information contents; accessibility guidelines, user’s profiles and actions that users can perform; as well as information related to emergencies, notifications and devices. These main classes are linked with relations existing within their subclasses. The following section we provide a use case that depicts the more common relations that exist in our ontology. 
work_ungntmprafc7zgznwu6bag5amu	----	An Approximate Microaggregation Approach for Microdata Protection Xiaoxun Sun1, Hua Wang1, Jiuyong Li2 Yanchun Zhang3 1Department of Mathematics & Computing University of Southern Queensland, Australia 2School of Computer and Information Science University of South Australia, Australia 3School of Engineering and Science Victoria University, Australia Abstract Microdata protection is a hot topic in the field of Statistical Disclosure Control, which has gained special interest after the disclosure of 658000 queries by the America Online (AOL) search engine in August 2006. Many algorithms, methods and properties have been proposed to deal with microdata disclosure. One of the emerging concepts in mi- crodata protection is k-anonymity, introduced by Samarati and Sweeney. k-anonymity provides a simple and efficient approach to protect private individual information and is gaining increasing popularity. k-anonymity requires that every record in the microdata table released be indistinguishably related to no fewer than k respondents. In this paper, we apply the concept of entropy to propose a distance metric to eval- uate the amount of mutual information among records in microdata, and propose a method of constructing dependency tree to find the key attributes, which we then use to process approximate microaggregation. Further, we adopt this new microaggregation technique to study k-anonymity problem, and an efficient algorithm is developed. Ex- perimental results show that the proposed microaggregation technique is efficient and effective in the terms of running time and information loss. 1 1 Introduction British politicians gasped with astonishment when they were told on November 20th, 2007, that two computer disks full of personal data of 25m British individuals had gone missing [24]. The fate of the disks is unknown and the privacy of the individuals, whose personal data are lost, is in danger. Unfortunately, this is the latest in a series of similar incidences. In October, HM’s Revenue and Customs (HMRC) lost another disk containing pension records of 15,000 people, and it also lost a laptop containing personal data on 400 people in September [7]. Data on 26.5m people were stolen from the home of an employee of the Department of Veterans Affairs in America in 2006, and 658000 queries were disclosed by the AOL search engine in August of the same year [15]. These pitfalls are not new. Due to the great advances in the information and communication technologies, it is very easy to gather large amounts of personal data, and mistakes such as those described are magnified. There are many real-life situations in which personal data is stored: For example: (i) Electronic commerce results in the automated collection of large amounts of consumer data. These data, which are gathered by many companies, are shared with subsidiaries and partners. (ii) Health care is a very sensitive sector with strict regulations. In the U.S., the Privacy Rule of the Health Insurance Portability and Accountability Act (HIPAA [16]) requires the strict regulation of protected health information for use in medical research. In most western countries, the situation is similar, (see e.g. [2]). (iii) Cell phones have become ubiquitous and services related to the current position of the user are growing fast. If the queries that a user submits to a location-based server are not securely managed, it could be possible to infer the consumer habits of the user [33]. (iv) The massive deployment of the Radio Frequency IDentification (RFID) technology is a reality. On the one hand, this technology will increase the efficiency of supply chains and will eventually replace bar codes. On the other hand, the existence of RFID tags in almost every object could be seen as a privacy problem [34]. In addition to these real-life situations, most countries have legislation which compels national statistical agencies to guarantee statistical confidentiality when they release data collected from citizens or companies; see [23] for regulations in the European Union, [26] for regulations in Canada, [27] for regulations in the U.S, and [28] for regulations in Australia. Thus, protecting individual privacy is a key issue for many institutions, especially statistical 2 agencies, Internet companies, manufacturers, etc; and many efforts have been devoted to develop techniques guaranteeing some degree of personal privacy. In order to protect privacy, Samarati and Sweeney [31, 38, 29, 30] proposed the k- anonymity model, where some of the quasi-identifier fields are suppressed or generalized so that, for each record in the modified table, there are at least k − 1 other records in the mod- ified table that are identical to it with respect to the quasi-identifier attributes. The general approach adopted in the literatures to achieve k-anonymity is suppression/generalization, so that minimizing information loss translates to reducing the number and/or the magnitude of suppressions and generalizations [1, 29, 38, 35, 37, 39, 20, 19, 21]. Another method to achieve anonymity is through microaggregation [12, 11, 32]. Microag- gregation is a Statistical Disclosure Control (SDC) technique consisting in the aggregation of individual data. It can be considered as an SDC sub-discipline devoted to the protection of microdata. Microaggregation can be seen as a clustering problem with constraints on the size of the clusters. It is somehow related to other clustering problems (e.g., dimension reduction or minimum squares design of clusters). However, unlike clustering, microaggregation is not considered with the number of clusters or the number of dimensions, but only the minimum number of elements that are grouped in each cluster. 1.1 Motivation As stated in [9, 10, 11], the result and execution time of miroaggregation depends on the number of the variables used in the microaggregation process. Microaggregation using fewer variables sometimes offer the best solution. The question of interest is: Do we have to use all the dimension resources (attributes) in the microaggregation, or can we use only a small number of the attributes in the microaggregation process and obtain better solutions? This paper is highly motivated by this. To answer the question, we introduce the concept of entropy, an important concept in information theory, and propose a distance metric to evaluate the amount of the mutual information among records in the microdata, and propose the method of constructing dependency tree to find the key attributes, which we can use to process approximate microaggregation. Further, we apply this new microaggregation technique to solve k-anonymity problem, and an efficient algorithm is developed. Finally, experimental 3 results show that the proposed microaggregation technique is efficient and effective in terms of running time and information loss. Our Contributions: • We propose a novel metric to measure the mutual information between attributes in the microdata based on the concept of entropy, which captures the expected uncertainty in the attribute pairs and the mutual information between them. We also discuss the properties of this metric. • Based on this mutual information measure, we develop a simple, yet efficient algorithm to find the best dependency tree from the given microdata, and we also discuss how to select key attributes from the best dependency tree, and how to use it for the approximate microaggregation. • We apply our technique to k-anonymity problem, and develop an efficient algorithm for it. Experimental results show that the proposed microaggregation technique is effective and efficient compared with the previous microaggregation method. Running Example ID A1 A2 A3 A4 A5 A6 r1 0 0 0 1 1 1 r2 0 1 1 0 1 0 r3 1 1 0 1 0 0 r4 0 0 1 1 1 1 r5 0 1 1 1 0 0 r6 0 0 1 0 0 1 r7 1 1 1 0 0 1 r8 0 1 1 0 0 0 r9 1 1 1 0 1 1 r10 0 1 1 1 0 1 r11 0 1 1 1 0 0 r12 1 1 1 1 1 1 Table 1: Sample data 4 For the simplicity of illustration, we use the data shown in Table 1 as our running example. There are 12 records {r1, r2, · · · , r12} in the sample data and each record contains 6 attributes {A1, · · · , A6}. For each attribute Ai (1 ≤ i ≤ 6), we define the probability P(Ai = x) as the fraction of rows whose projection onto Ai is equal to x, where x ∈ {0, 1}. For instance, P(A1 = 1) = 1/3, P(A3 = 0) = 1/6 and P(A1 = 1, A3 = 0) = 1/12. 2 Background Many techniques have been proposed to deal with the anonymity problem. In this section, we introduce some basic concepts regarding this. First, we take a look at some fundamental concepts of microaggregation and k-anonymity. Then, we show how to achieve k-anonymity through microaggregation. 2.1 Microaggregation Statistical Disclosure Control (SDC) seeks to transform data in such a way that the data can be publicly released whilst preserving utility and privacy, where the latter means avoiding disclosure of information that can be linked to specific individual or corporate respondent entities. Microaggregation is an SDC technique consisting in the aggregation of individual data. It can be considered as an SDC sub-discipline devoted to the protection of the micro- data. Microaggregation can be seen as a clustering problem with constraints on the size of the clusters. It is somehow related to other clustering problems (e.g., dimension reduction or minimum squares design of clusters). However, the main difference of the microaggregation problem is that it does not consider the number of clusters to generate or the number of dimensions to reduce, but only the minimum number of elements that are grouped in the same cluster. Microaggregation has been used for several years in different countries. It started at Eurostat [8] in the early nineties, and has since then been used in Germany [25] and several other countries [13]. Microaggregation is relevant not only with SDC, but also in artificial intelligence [10]. In the latter field, the application is to increase the knowledge of a system for decision making and domain representation. Microaggregation techniques may also be 5 Gender Age Postcode Problem male middle 4350 stress male middle 4350 obesity male young 4351 stress female young 4352 obesity female old 4353 stress female old 4353 obesity Table 2: A raw microdata Gender Age Postcode Problem male middle 4350 stress male middle 4350 obesity ∗ young 435∗ stress ∗ young 435∗ obesity female old 4353 stress female old 4353 obesity Table 3: A 2-anonymous microdata 69 5 11104 5 55 1010 10 Average Origianl data Microaggregated data record Figure 1: Example of microaggregation used in data mining in order to scale down or even compress the data set while minimizing the information loss. When we microaggregate data we have to keep two goals in mind: (i) Preserving data utility. To do this, we should introduce as little noise as possible into the data; i.e., we should aggregate similar elements instead of different ones. In the example in Figure 1, groups of three elements are built and aggregated. Note that elements in the same aggregation group are similar. (ii) Protecting the privacy of the individuals. Data have to be sufficiently modified to make re-identification difficult; i.e., by increasing the number of aggregated elements, we increase data privacy. In the example in Figure 1, after aggregating the chosen elements, it is impossible to distinguish them, so that the probability of linking any individual is inversely proportional to the number of aggregated elements. In order to determine whether two elements are similar, a similarity function such as the Euclidean distance, Minkowski distance or Chebyshev distance can be used. A common 6 measure is the Sum of Squared Errors (SSE). The SSE is the sum of squared distances from the centroid of each group to every record in the group, and is defined as: SSE = s∑ i=1 ni∑ j=1 (xij − x̄i)′(xij − x̄i) (1) where s is the number of groups, ni is the number of records in the i th group, xij is the j th record in the ith group and x̄i is the average record of the i th group. Optimal multivariate microaggregation, that is, with minimum SSE, was shown to be NP-hard in [22]. The only practical microaggregation methods are heuristic. 2.2 K-Anonymity k-anonymity, suggested by Samarati and Sweeney [31, 38, 29, 30], is an interesting approach to reduce the conflict between information loss and privacy protection. To define of k-anonymity, we need to enumerate the various types of attributes that can appear in a microdata set T : • Identifier attributes that can be used to identify a record, such as Name and Medicare card. Since our objective is to prevent sensitive information from being linked to specific respondents, we will assume in what follows that identifier attributes in the microdata have been removed or encrypted in a pre-processing step. • Quasi-identifier (QI) attributes are those, such as Postcode and Age, that in combination, can be linked with external information to re-identify (some of) the respondents to whom (some of) the records in the microdata belong. Unlike identifier attributes, QI attributes can not be removed from the microdata, because any attribute is potentially a QI attribute. • Sensitive attributes that are assumed to be unknown to an intruder and need to be pro- tected, such as Disease or ICD-9 Code1. 1International Statistical Classification of Diseases and Related Health Problems: ICD-9, which provides multiple external links for looking up ICD codes. Available http://icd9cm.chrisendres.com/. 7 Definition 1 (k-anonymity). A protected microdata set is said to satisfy k-anonymity, if, for each combination of QI attributes, at least k records exist in the microdata sharing that combination Note that, if a protected microdata T ′ satisfies k-anonymity, an intruder trying to link T ′ with an external non-anonymous data source will find at least k records in T ′ that match any value of the QI attributes the intruder use for linkage. Thus re-identification, i.e., map- ping a record in T ′ to a non-anonymous record in the external data source, is not possible. For example, Table 3 is a 2 anonymous view of Table 2 if QI attributes are {Gender, Age, Postcode}. If for a given k, k-anonymity is assumed to be enough protection for respondents, one can concentrate on minimizing information loss with the only constraint that k-anonymity should be satisfied. This is a clean way of solving the tension between data protection and data utility. The general approach adopted in the literature to achieve k-anonymity is suppres- sion/generalization, so that minimizing information loss translates to reducing the number and/or the magnitude of suppressions and generalizations [29, 38, 35]. Generalization consists in substituting the values of a given attribute with more general values. We use ∗ to denote the more general value. For instance, in Table 3, Postcode 4351 and 4352 are generalized to 435∗. Suppression refers to removing the part or entire value of attributes from the microdata. Note that suppressing an attribute to reach k-anonymity can equivalently be modeled via a generalization of all the attribute values to ∗. The drawbacks of partially suppressed and coarsened data for analysis were highlighted in [12]: 1. Satisfying k-anonymity with minimum data modification using generalization (recod- ing) and local suppression was shown to be NP-hard by Meyerson and Williams [21], Aggarwal et al. [1] and Sun et al. [36]; 2. Using global recoding for generalization causes too much information loss, and using local recoding complicates data analysis by causing old and new categories to co-exist in the recoded data; 8 3. There is no standard way of using local suppression and analyzing partially suppressed data usually requires specific software; 4. Last but not least, when numerical attributes are generalized, they become non-numerical. Joint multivariate microaggregation of all QI attributes with minimum group size k was proposed in [12] as an alternative to achieve k-anonymity. Besides being simpler, this alterna- tive has the advantage of yielding complete data without any coarsening (nor categorization in the case of numerical data). Other proposals [18, 35, 36, 37] generalize ordinal numeri- cal data, replacing numerical data by intervals. In the case of the k-anonymity application, micro-aggregation is performed on the projection of records on QI attributes. The first algorithm, known as Maximum Distance to Average Vector (MDAV), to achieve microaggregation through k-anonymity was proposed in [11]. The MDAV algorithm works as follows: First, it computes the centroid (average record) of records in the data set, and find the most distant record r from the centroid and the most distant record s from r. Second, it forms two groups around r and s: the first group contains r and the k − 1 records closest to r; the other group contains s and the k − 1 records closest to s. Finally, the two group are microaggregated and removed from the original dataset. The steps are repeated until there are no records in the original dataset. Although MDAV generates groups of fixed size k, it lacks flexibility for adapting the group size to the distribution of the records in the data set, which may result in poor homogeneity in a group. Variable-size MDAV (V-MDAV) was proposed to overcome this limitation by computing a variable-size group, and a detailed analysis can be found in [32]. In the next section, we will propose our approximate microaggregation technique, and show how to apply it to solve k-anonymity in order to overcome most of the problems of generalization/suppression listed above. 3 Approximate Microaggregation The work presented in this paper is based on information theory, and is related to the applica- tion of dependency tree of information theory in data mining and databases. In this section, 9 we first introduce the concept of entropy, and the mutual information measure, which cap- tures the mutual dependency between attributes. Then we introduce our microaggreation technique by constructing the dependency tree, and finally, we apply this microaggregation technique to k-anonymity problem, and an efficient algorithm is proposed. 3.1 Mutual Information Measure We are more surprised when an unlikely outcome happens than a likely one occurs. A useful measure of the surprise of an event with probability p is −log2p. The main concept of infor- mation theory is that of entropy, which measures the expected uncertainty or the amount of information provided by a certain event. The entropy of X is defined by: H(X) = − ∑ x P(X = x)log2P(X = x) with 0log20 = 0 by convention. It can be shown that 0 ≤ H(X) ≤ log2|X|, with H(X) = log2|X| only for the uniform distribution, P(X = x) = 1/|x| for all x ∈ X. For instance, in the given running example, H(A1) = −(8/12)log2(8/12) − (4/12)log2(4/12) = 0.9183, H(A2) = 0.8113 and H(A1, A2) = 1.5546. The conditional entropy H(Y |X) of a random variable Y given X is then defined as: H(Y |X) = − ∑ x,y p(x, y)log2p(y|x) where p(x, y) is the joint distribution of variables X and Y . The conditional entropy has the following properties: Proposition 1: Let H(Y |X) be the conditional entropy for Y given X, then, (1) 0 ≤ H(Y |X) ≤ H(Y ); (2) H(X, Y ) = H(X) + H(Y |X) = H(Y ) + H(X|Y ); (3) H(X, Y ) ≤ H(X) + H(Y ) 10 The proof of Proposition 1 is given in [40]. According to the proposition, the condi- tional entropy H(Y |X) can be rewritten as: H(Y |X) = H(X, Y ) − H(X), which provides an alternative and easy way to compute the conditional entropy H(Y |X). For instance, in our running example, H(A1|A2) = H(A1, A2) − H(A1) = 1.5546 − 0.9183 = 0.6363 and H(A2|A1) = 0.7433. We adopt the conditional entropy to measure the mutual information, which is a distance metric. Definition 2 (Mutual Information Measure). The mutual information measure with re- gard to two random variables A and B is defined as: MI(A, B) = H(A|B) + H(B|A) (2) Mutual information measure is a measure of how independent are the two random variables when the value of each random variable is known. Two events A and B are independent if and only if their mutual information measure achieves the maximum H(A) + H(B). Therefore, the less the value of the mutual information measure is, the more dependent the two random variables are. According to this measure, A is said to be more dependent on B than C, if MI(A, B) ≤ MI(A, C). Theorem 1: The mutual information measure MI(A, B) satisfies the following properties: (1) MI(A, B) ≥ 0; (2) MI(A, B) = MI(B, A); (3) MI(A, B) + MI(B, C) ≥ MI(A, C) Proof: The first two are easy to be verified. Here, we give the detail for the third one. Note that, 11 H(A|C) ≤ H(A, B|C) (3) ≤ H(B|C) + H(A|B, C) − H(C) (4) ≤ H(B|C) + H(A|B) + H(C) − H(C) (5) = H(B|C) + H(A|B) (6) The inequalities (3) and (4) hold because of Proposition 1(1) and (2). (5) holds due to Proposition 1(3) and (6) holds because of Proposition 1(2). Then, MI(A, B) + MI(B, C) (7) = H(A|B) + H(B|A) + H(B|C) + H(C|B) (8) = (H(A|B) + H(B|C)) + (H(C|B) + H(B|A)) ≥ H(A|C) + H(C|A) (9) = MI(A, C) (10) The equality (8) holds because of the definition of mutual information measure and the inequality (9) holds because of (6). It is easy to verify that MI(A, B) = 0 if and only if there is a one-to-one function mapping between A and B. Since when H(B|A) = 0, B is a function of A, then when MI(A, B) = 0 if and only if H(B|A) = 0 and H(A|B) = 0; i.e, there is a one-to-one function mapping between A and B. In this sense, the mutual information measure MI(A, B) we defined is a distance metric. 3.2 Dependency Tree Dependency tree was introduced by Chow and Liu [4], in which they introduced an algorithm for fitting a multivariate distribution with a tree (i.e., a density model that assumes that there are only pairwise dependency between variables). In the maximum likelihood sense, the 12 dependency tree is the best tree to fit the dataset, and it uses mutual information measure to estimate the dependency of two random variables. The dependency tree has been used in finding dependency structure in the features which improve the classification accuracy of the Bayes network classifiers [14]. [5] uses the depen- dency tree to represent a set of frequent patterns, which can be used to summarize patterns into few profiles. [17] presents a large node dependency tree, in which the nodes are subsets of variables of dataset. The large node dependency tree is applied to density estimation and classification. Definition 3 (Dependency Matrix). Given microdata T with n records {r1, r2, · · · , rn}, where each record contains m attributes {A1, A2, · · · , Am}, the dependency matrix DT is de- fined as: DT = (MI(i, j))m×m where MI(i, j) is the mutual information measure, i, j ∈ {A1, A2, · · · , Am}. For instance, the dependency matrix in our running example is as follows:  0 1.3796 1.5339 1.8777 1.8777 1.8126 1.3796 0 1.3753 1.7772 1.6681 1.3180 1.5339 1.3753 0 1.3368 1.6217 1.6217 1.8777 1.7772 1.3368 0 1.9586 1.9586 1.8777 1.6681 1.6217 1.9586 0 1.7510 1.8126 1.3180 1.6217 1.9586 1.7510 0   With the dependency matrix, we could construct a fully connected weighted graph G = (V, E, ω), where V = {v1, v2, · · · , vm} is the set of vertices, which corresponds to the attributes in T , and for each pair of vertices (vi, vj) there is an edge eij connecting them, and ω(eij) refers to the weight of each eij between vi and vj, which can be obtained from the dependency matrix. An example of such a fully connected graph is shown in Figure 2(Left). We observe that ω(eij) represents to what extent vertex vi (or attribute Ai) is dependent on vj (or Aj). Although, in the worst case, any pair of attributes can be dependent, however, as stated in [4], we could simplify by using an approximation which ignores the conditions on 13 A1 A2 A3 A4 A5 A6 G A1 A2 A3 A4 A5 A6 T G 1.3796 1.3753 1.7772 1.6217 1.3180 Figure 2: Left: Fully connected graph G; Right: Its minimum spanning tree TG (Right) v1 v2 v m· · · · · ·v3 v1 v2 · · · · · · vm (a) (b) Figure 3: Proof of Theorem 2 multiple attributes, and retaining only dependency in at most a single attribute at a time, which results in a tree-like structure. It is easy to see that in the fully connected weighted graph G, there are a large number of trees, each of which represents a unique approximation dependency structure. Here, in order to reduce the uncertainty in the dataset and maximize the mutual information among the attributes simultaneously, we find the minimum spanning tree as our best dependency tree from the fully connected graph G based on our proposed mutual information measure. Here, we use the Kruskal algorithm [6], which is essentially a greedy algorithm. The candidate edges are sorted in increasing order of their weights (i.e. mutual information measure). Then, starting with an empty set E0, the algorithm examines one edge at a time (in the order resulting from the sort operation), checks if it forms a cycle with the edges already in E0 and, if not, adds it to E0. The algorithm ends when m−1 edges have been added to E0, where m refers to the number of vertices in G. 14 Algorithm 1: Finding best dependency tree 1. Compute the mutual information measure between each pair of attributes in T and construct the dependency matrix DT . There are m(m − 1)/2 weight need to be calculated, since T has m attributes. 2. Construct a fully connected graph, where the nodes correspond to the attributes in T . The weight of each edge refers to their mutual information measure. 3. Find the best dependency tree by the the minimum spanning tree algorithm. The algorithm of finding best dependency tree is briefly described in Algorithm 1 and an example of the found out best dependency tree is shown in Figure 2(Right). After finding out the best dependency tree, we need to set out rules to select the key attributes from the dependency tree to process approximate microaggregation. Definition 4 (Degree of The Vertex). Let G = (V, E) be a graph, where V = {v1, v2, · · · , vm}. Then, the degree of the node vi is the number of edges incident to the nodes, denoted by deg(vi). For example, in Figure 2(Right), deg(A2) = 4, and deg(A3) = 2. Let TG be the best dependency tree found in G. We then compute the degree of each vertex in TG and sort them in decreasing order. Without loss of generality, we assume that deg(v1) ≥ deg(v2) ≥ · · · ≥ deg(vm) after they are sorted in decreasing order. Then, the principle of choosing the key attributes is as follows: Definition 5 (Choosing Key Attributes). Suppose deg(v1) ≥ deg(v2) ≥ · · · ≥ deg(vm) after they are sorted. Then, the vertices v1, v2, · · · vk are chosen as the key attributes if the following two requirements are satisfied at the same time: 15 Algorithm 2: k-anonymity through approximate microaggregation Input: Microdata set T consisting of n records having m attributes each. Output: Microaggregated microdata T ′ satisfying k-anonymity property 1. Find out the best dependency tree by Algorithm 1 and select the key attributes 2. Project the records of T to the key attributes. 3. Computes the centroid (average record) x̄ of records in the projected data set, and find the most distant record r from the centroid and the most distant record s from r. 4. Form two groups around r and s: the first group contains r and the k − 1 records closest to r; The other group contains s and the k − 1 records closest to s. 5. If there are at least 2k records which do not belong to any of the groups formed in Step 4, go to Step 3, taking the previous set of records minus the groups formed in the latest instance of Step 4, as the new set of records. 6. If there are between k and k − 1 records which do not belong to any of the groups formed in Step 4, form a new group with those records and exit the algorithm. 7. If there are less than k remaining records which do not belong to any of the groups formed in Step 4, add them to the group formed in Step 4 whose centroid is closest to the centroid of the remaining records. 8. Return microaggregated data T ′ by replacing each record by the centroid of the group it belongs to. k−1∑ i=1 deg(vi) < m (11) k∑ i=1 deg(vi) ≥ m (12) For example, for the minimum spanning tree TG in Figure 2, we choose attributes A2 and A3 as the key attributes, since according to the principle described above, deg(A2) < 6 and deg(A2) + deg(A3) = 6. Theorem 2: Let TG be the best dependency tree of G, with V = {v1, v2, · · · , vm}, and N be the number of selected key attributes. Then, 2 ≤ N ≤ m/2. 16 Proof: Since in a tree-like structure, the maximum degree of a vertex is m−1 [6], and without loss of generality, we assume that deg(v1) = m − 1, and in this case, the best dependency tree found has the form as shown in Figure 3(a), and then according to Definition 5, only two vertices will be selected as key attributes, say v1 and v2. This is the situation when the number of the selected key attributes reaches the minimality. On the other hand, when the number of the selected key attributes reaches the maximality, the structure of the best dependency tree has the form as shown in Figure 3(b), and in this case, at most m/2 key attributes will be selected. So, 2 ≤ N ≤ m/2. Theorem 2 assures that at most half the amount of dimension resources are needed in the microaggregation process with our technique, which could significantly reduce the execution time. In the next section, we discuss in detail how to apply this technique to k-anonymity problem. 3.3 Application to K-Anonymity Our aim is to obtain k-anonymous microdata without coarsened nor partially suppressed data. This makes their analysis and exploitation easier, with the additional advantage that numerical continuous attributes are not categorized. In this section, we adopt the approximate microaggregation technique to solve k-anonymity problem. Our algorithm receives as input a microdata set T consisting of n records having m at- tributes each. The result of the algorithm is a k-partition used to microaggregate the original microdata set and to generate a microaggregated data set T ′ that fulfils the k-anonymity property. Instead of taking all the attributes into the microaggregation process, we only use the selected key attributes, which captures the dependency between attributes, to microag- gregate the data. The novelty and difference from the previous microaggregation methods exist here. Our proposed approach is effective and efficient in terms of running time and information loss. The first two steps of the algorithm builds the initial dataset for microaggregation. It selects the key attributes from the best dependency tree and returns a projected dataset, which has the same number of records as T, but each record only contains the value of key 17 attributes. Once the average record is computed, the algorithm looks for other records which are distant to it and adds records to it until it reaches a minimum cardinality k (Step 3- 4). After repeating this process several times, a set of groups satisfying the k-anonymity property is obtained. However, a number of records can remain unassigned, and they must be distributed amongst the previously created groups (Step 5-7). Finally, the algorithm further microaggregates the original microdata T by replacing each record in T by the centroid of the group to which it belongs (Step 8). The algorithm is outlined in Algorithm 2. In this section, we discuss in detail how to apply our microaggregation technique to solve k-anonymity in order to overcome most of the problems of generalization/suppression listed in Section 2 in the following aspects: • Approximate microaggregation is a unified approach, unlike the dual method combining generalization and suppression. • It does not complicate data analysis by adding new categories to the original scale, unlike generalization/suppression. • It does not result in suppressed data, which makes analysis of k-anonymous data easy. • It is suitable to protect continuous data without removing their numerical semantics. 4 Experimental Results 4.1 Data set We employ a real-life CENSUS data set downloadable at http://www.ipums.org in the ex- perimental study. The CENSUS data set contains the personal information of 500K American adults. The data set has 9 discrete attributes summarized in Table 4. From CENSUS, we create two sets of micro tables, in order to examine the influence of dimensionality and the impact of cardinality. The first set has 6 tables, denoted as CENSUS-20%, · · · , CENSUS- 100%, respectively. Specifically, CENSUS-t% (20 ≤ t ≤ 100) indicates the data set consisting of t% records randomly sampled from the whole CENSUS data set, and each record has 9 18 Attribute Number of distinct values Age 78 Gender 2 Education 17 Marital 6 Race 9 Work-class 8 Country 83 Occupation 50 Salary-class 50 Table 4: Summary of attributes in CENSUS attributes shown in Table 4. The second set contains 5 tables, denoted as 5-CENSUS, · · · , 9-CENSUS, respectively, where n-CENSUS (3 ≤ n ≤ 9) represents the data set with the first n attributes selected from Table 4, and each data set has the same number of records as the whole CENSUS data set. 4.2 Experiment setup Our aim is to test the efficiency and effectiveness of the proposed approximate microaggrega- tion algorithm for k-anonymity. We denote our proposed algorithm as MA, and we compare it with the previous MDAV-based algorithm [11], denoted as MA. We first evaluate the execution time of our approach by varying the cardinality of the data sets, the number of attributes and the value of k. In order to compare the effectiveness, for each data set, we adopt two measurements. One is to measure the information loss in terms of SSE/SST , where SSE is the sum of square errors as defined in equation (1), and SST refers to the sum of square errors applied over the whole dataset. The other metric is to compare the number of key attributes projected in the microaggregation. 4.3 Results Efficiency: Figures 4(a)-(c) show the comparison of execution time of two microaggregation methods. In this set of experiments, we fixed k = 20 and vary the data percentage. Figure 19 Census−20% Census−40% Census−60% Census−80% Census−100% 10 20 30 40 50 60 R u n n in g t im e ( S e c ) MA AMA (a) K=5 K=10 K=15 K=20 K=25 0 10 20 30 40 50 60 70 80 R u n n in g t im e ( S e c ) MA AMA (b) 5−Census 6−Census 7−Census 8−Census 9−Census 20 25 30 35 40 45 50 55 60 R u n n in g t im e ( S e c ) AMA MA (c) Figure 4: Running time comparison between different methods 4(a) plots the result by varying the data percentage of the whole Census data set from 20% to 100%. As we can see, the AMA incurs less computation time than MA method. This is expected since in the AMA process, less attributes are used in the microaggregation. We can see the difference of the computation cost is getting larger with the increased data cardinality. Figure 4(b) describes the running time comparison when varying the privacy parameter k. The computation cost of both MA and AMA algorithms is increasing with k, but AMA consistently outperforms MA method. Figure 4(c) shows the computation overhead differences by altering the number of attributes. The computation overhead of both methods is increasing when enlarging the number of attributes. The result is expected since the overhead is increased with the more dimensions. The AMA method performs better than MA algorithm since we use a part of the attributes instead of the whole dimensional resources, which is significantly reduce the amount of computation. Effectiveness: Having verified the efficiency of our technique, we proceed to test its effective- ness. We measure the utility in terms of SSE/SST , where SSE is the sum of square errors as defined in equation (1), and SST refers to the sum of square errors applied over the whole data set. Figure 5 shows the number of key attributes used in MA and AMA approaches. As we can see, the number remains the same for MA method, since it projects all the attributes into the microaggregation process. On the contrary, the number of key attributes used in 20 Census−20% Census−40% Census−60% Census−80% Census−100% 0 5 10 15 N o . o f K e y A tt ri b u te s MA AMA Figure 5: No. of key attributes comparison AMA is less than half of that used by MA approaches, which verifies the results in Theorem 2. Figures 6(a) and (b) show the information loss by applying MA and AMA algorithms. Figure 6(a) is plotted by changing the percentage of data set. Although the result indicates that AMA generates a little bit more information loss than MA, the difference is not enlarged when the data cardinality is increased. Similar tread is obtained in Figure 6(b) by varying the value k. The information loss is increased with k, since larger k demands more strict privacy requirement, which reduces the utility of the data. Summary: Overall, the AMA outperforms MA in terms of efficiency, and the difference is getting larger when the volume and dimension of data are increasing. Although AMA generates a little bit more information loss than MA, it is still practical since AMA only uses at most half of the attributes in the microaggregation process. 5 Related Work Privacy preservation is an important issue in the release of data for mining purpose. The k- anonymity model, which was introduced for protecting individual identification by Samarati and Sweeney [29, 31], has been extensively investigated for its simplicity and effectiveness [1, 29, 38, 35, 37, 39, 20, 19, 21]. k-anonymity requires that each record in the anonymous table 21 Census−20% Census−40% Census−60% Census−80% Census−100% 0 10 20 30 40 50 60 In fo rm a ti o n L o s s AMA MA (a) K=5 K=10 K=15 K=20 K=25 20 30 40 50 60 70 In fo rm a ti o n l o s s MA AMA (b) Figure 6: Information loss comparison between different methods be indistinguishable with at least k−1 other records within the dataset with respect to a set of quasi-identifier attributes. In this case, individuals cannot be uniquely identified by adversary, so the individuals’ privacy can be preserved. Started from [29, 31, 30], the general approach adopted in the literature to achieve k-anonymity is based on generalization/suppression, which has some defects on efficiency, information loss and implementation. Our work in this paper is related to the microaggregation technique, which has been introduced to implement k- anonymous data set recently and remedies most of defects of generalization and suppression [12, 10, 11, 22, 9]. In the previous research, all the dimensional resources (attributes) are required in the microaggregation process. However, as mentioned in [10, 11, 9], the result and execution time of the microaggregation highly depends on the number of the variables used in the mi- croaggregation process, since few variables sometimes offers the better solutions. Different from previous microaggregation methods, in this paper, we propose a new approach to select only a small number of dimensional resources that captures the maximal dependency rela- tionship among resources and as experiments show that the new technique achieves better microaggregation results. Specifically, our microaggregation method is effective and efficient in terms of information loss and running time. In the case of k-anonymity problem, the microaggregation approach presented in this paper could overcome most of the problems of 22 generalization/suppression. (1) Our method is a unified approach, unlike the dual method combining generalization and suppression. (2) It does not complicate data analysis by adding new categories to the original scale, unlike generalization/suppression. (3) It does not result in suppressed data, which makes analysis of k-anonymous data easy. (4) It is suitable to protect continuous data without removing their numerical semantics. From a different perspective, the microaggregation technique discussed in this paper produces better solutions compared with previous ones. Our work is also related to the application of dependency tree of information theory in data mining and databases. The dependency tree has been used in finding dependency structure in the features which improve the classification accuracy of the Bayes network classifiers [14]. [5] uses the dependency tree to represent a set of frequent patterns, which can be used to summarize patterns into few profiles. [17] presents large node dependency tree, in which the nodes are subsets of variables of data set. The large node dependency tree is applied to density estimation and classification. As far as its application to privacy preserving data mining, fewer results are obtained. In this paper, we introduce the concept of entropy and propose the mutual information measure to evaluate the mutual dependency between attributes, and the method to construct the dependency tree. We also discuss how to select key attributes from the constructed dependency tree, and how to use them in the approximate microaggregation. We prove theoretically that at most half the amount of resources are needed with our approach. 6 Conclusion and Future Work k-anonymity is a property that, when satisfied by the microdata, can help increase the privacy of the respondents whose data is being used. Previous approaches to obtain microdata sets fulfilling the k-anonymity property were mainly based on suppression and generalization. In this article, we have shown how to achieve the same property by means of approximate microaggregation, which, different from the previous microaggregation method, uses a part of the dimensional resources. It works by selecting key attributes from the best dependency tree, which is constructed based on a new mutual information measure based on information theory, 23 which captures the dependency between attributes in the microdata. The experimental results show that the proposed technique is efficient and effective in the terms of running time and information loss. A number of other sophistication of k-anonymity for protecting against attribute disclosure have recently been proposed, such as (p+, α)-sensitive k-anonymity [37], l-diversity [20], (α, k)- anonymity [39], t-closeness [19]. All of them rely on generalizations, so the microaggregation approach proposed in this paper would be a novelty in all of them. The technique proposed in this paper restricted its focus on numerical attributes, and it is interesting to investigate the extension to other types of attributes. References [1] G. Aggarwal, T. Feder, K. Kenthapadi, R. Motwani, R. Panigrahy, D. Thomas and A. Zhu. Anonymizing tables. In Proc. of the 10th International Conference on Database Theory (ICDT’05), pp. 246-258, Edinburgh, Scotland. [2] C. Boyens, R. Krishnan, and R. Padman. On privacy-preserving access to distributed het- erogeneous healthcare information. In I. C. Society, editor, Proceedings of the 37th Hawaii International Conference on System Sciences HICSS-37, Big Island, HI., 2004. [3] R. Brand, J. Domingo-Ferrer, and J. M. Mateo-Sanz, Reference data sets to test and compare sdc methods for protection of numerical microdata, 2002, European Project IST-2000-25069 CASC, http://neon.vb.cbs.nl/casc. [4] C. Chow and C. Liu. Approximating discrete probability distributions with dependence trees. IEEE Transactions on Information Theory, 14,3:462-467, 1968. [5] G. Cong, B. Cui, Y. Li, and Z. Zhang. Summarizing frequent patterns using profiles. In Database Systems for Advanced Applications, 11th International Conference, DASFAA, 2006. [6] T. Cormen, C. Leiserson, R. Rivest, C. Stein. Introduction to Algorithms, second edition, MIT Press and McGraw-Hill. ISBN 0-262-53196-8. [7] Data lost by Revenue and Customs. BBC News. http://news.bbc.co.uk/1/hi/uk/7103911. stm 24 [8] D. Defays and P. Nanopoulos, Panels of enterprises and confidentiality: the small aggregates method, in Proc. of 92 Symposium on Design and Analysis of Longitudinal Surveys. Ottawa: Statistics Canada, 1993, pp. 195-204. [9] J. Domingo-Ferrer, V. Torra. Aggregation Techniques for Statistical confidentiality. In: Aggre- gation operators: new trends and applications, pp. 260-271. Physica-Verlag GmbH, Heidelberg (2002) [10] J. Domingo-Ferrer and V. Torra, On the connections between statistical disclosure control for microdata and some artificial intelligence tools, Information Sciences, vol. 151, pp. 153-170, May 2003. [11] J. Domingo-Ferrer and J. M. Mateo-Sanz, Practical data-oriented microaggregation for statis- tical disclosure control, IEEE Transactions on Knowledge and Data Engineering, vol. 14, no. 1, pp. 189-201, 2002. [12] J. Domingo-Ferrer and V. Torra, Ordinal, continuous and heterogenerous k-anonymity through microaggregation, Data Mining and Knowledge Discovery, vol. 11, no. 2, pp. 195-212, 2005. [13] E. C. for Europe, Statistical data confidentiality in the transition countries: 2000/2001 winter survey, in Joint ECE/Eurostat Work Session on Statistical Data Confidentiality, 2001, invited paper n.43. [14] N. Friedman, D. Geiger, and M. Goldszmid. Bayesian network classifiers. Machine Learning, 29:131-163, 1997. [15] S. Hansell. AOL removes search data on vast group of web users. New York Times, Aug 8 2006. [16] HIPAA. Health insurance portability and accountability act, 2004. http://www.hhs.gov/ocr/ hipaa/. [17] K. Huang, I. King, and M. Lyu. Constructing a large node chow-liu tree based on frequent itemsets. In Proceedings of the International Conference on Neural Information Processing, 2002. [18] K. LeFevre, D. DeWitt, and R. Ramakrishnan. Incognito: Efficient Full-Domain k-Anonymity. In ACM SIGMOD International Conference on Management of Data, June 2005. 25 [19] N. Li, T. Li and S. Venkatasubramanian. t-Closeness: Privacy Beyond k-anonymity and l- diversity. ICDE 2007: 106-115 [20] A. Machanavajjhala, J. Gehrke, D. Kifer, and M. Venkitasubramaniam. l-Diversity: Privacy beyond k-anonymity. ICDE 2006. [21] A. Meyerson and R. Williams. On the complexity of optimal k-anonymity. In Proc. of the 23rd ACM-SIGMOD-SIGACT-SIGART Symposium on the Principles of Database Systems, pp. 223- 228, Paris, France, 2004. [22] A. Oganian and J. Domingo-Ferrer, On the complexity of optimal microaggregation for sta- tistical disclosure control, Statistical Journal of the United Nations Economic Comission for Europe, vol. 18, no. 4, pp. 345-354, 2001. [23] Euro. Parliament. DIRECTIVE 2002/58/EC of the European Parliament and Council of 12 july 2002 concerning the processing of personal data and the protection of privacy in the electronic communications sector (Directive on privacy and electronic communications), 2002. http:// europa.eu.int/eur-lex/pri/en/oj/dat/2002/l_201/l_20120020731en00370047.pdf. [24] E. Pfanner. Data Leak in Britain Affects 25 Million. The New York Times. http://www. nytimes.com/2007/11/22/world/europe/22data.html, November 22, 2007. [25] M. Rosemann, Erste Ergebnisse von vergleichenden Untersuchungen mit anonymisierten und nicht anonymisierten Einzeldaten am Beispiel der Kostenstrukturerhebung und der Umsatzs- teuerstatistik, in G. Ronning and R. Gnoss (editors), Anonymisierung wirtschaftsstatistischer Einzeldaten, Wiesbaden: Statistisches Bundesamt, 2003, pp. 154-183. [26] Ca. Privacy. Canadian privacy regulations, 2005. http://www.media-awareness.ca/english/ issues/privacy/canadian_legislation_privacy.cfm. [27] US. Privacy. U.S. privacy regulations, 2005. http://www.media-awareness.ca/english/ issues/privacy/us_legislation_privacy.cfm. [28] Aus. Privacy. Review of Australian Privacy Law, Discussion Paper 72, (DP 72), September 2007, http://www.austlii.edu.au/au/other/alrc/publications/dp/72/. 26 [29] P. Samarati and L. Sweeney. Protecting privacy when disclosing information: k-anonymity and its enforcement through generalization and suppression. Technical Report SRI-CSL-98-04, SRI Computer Science Laboratory, 1998. [30] P. Samarati. Protecting respondents’ identities in microdata release. IEEE Transactions on Knowledge and Data Engineering, 13(6): pp: 1010-1027. 2001. [31] P. Samarati. and L. Sweeney. Generalizing data to provide anonymity when disclosing infor- mation (Abstract). In Proc. of the 17th ACM-SIGMODSIGACT- SIGART Symposium on the Principles of Database Systems, p. 188, Seattle, WA, USA, 1998. [32] A. Solanas and A. Martinez-Balleste, A Multivariate Microaggregation With Variable Group Size. In 17th COMPSTAT Symposium of the IASC, Rome (2006). [33] A. Solanas and A. Martinez-Balleste. Privacy protection in location-based services through a public-key privacy homomorphism. In EuroPKI’07, LNCS 4582, pages 362-368. Springer, June 2007 [34] A. Solanas, J. Domingo-Ferrer, A. Martinez-Balleste and V. Daza. A Distributed Architecture for Scalable RFID Identification. Computer Networks, 51, 2007 [35] X. Sun, M. Li, H. Wang and A. Plank. An efficient hash-based algorithm for minimal k- anonymity problem. 31st Australasian Computer Science Conference (ACSC 2008), Wollon- gong, NSW, Australia. CRPIT 74, pp: 101-107. [36] X. Sun, H. Wang and J. Li. On the complexity of restricted k-anonymity problem. 10th Asia Pacific Web Conference (APWeb 2008), LNCS 4976, pp: 287-296, Shenyang, China. [37] X. Sun, H. Wang, J. Li, T. M. Traian and P. Li. (p+, α)-sensitive k-anonymity: a new enhanced privacy protection model. In 8th IEEE International Conference on Computer and Information Technology (IEEE-CIT 2008), 8-11 July 2008, Sydney, Australia. pp:59-64. [38] L. Sweeney. Achieving k-anonymity privacy protection using generalization and suppression. International Journal on Uncertainty, Fuzziness, and Knowledge-based Systems, 10(5):571-588, 2002. [39] R. Wong, J. Li, A. Fu, K. Wang. (α, k)-anonymity: an enhanced k-anonymity model for privacy preserving data publishing. KDD 2006: 754-759. 27 [40] J. C. A. van der Lubbe. Information Theory. Cambridge University Press. 1997. 28 
work_uo4gaxtmyvbm7f7z5jt2my4m2e	----	Microsoft Word - 14. Drage Petreski.doc 850 V O JN O T E H N IČ K I G LA S N IK / M IL IT A R Y T E C H N IC A L C O U R IE R , 2 01 6. , V ol 6 4, N o 3 EXPERT SYSTEM FOR MANAGING LOGISTIC PROCESSES Drage T. Petreskia, Andrej P. Ilievb, Lazar B. Gjurovc, Jugoslav Z. Ackoskid, Military Academy „General Mihailo Apostolski“, Skopje, Republic of Macedonia, Aleksandra D. Petreskae, Faculty of Electrical Engineering and Information Technologies, Skopje, Republic of Macedonia a e-mail: drage_petreski@yahoo.com, ORCID iD: http://orcid.org/0000-0002-5830-1389 b e-mail: andrej220578@gmail.com, ORCID iD: http://orcid.org/0000-0002-7917-1966 c e-mail: lazar.gjurov@gmail.com, ORCID iD: http://orcid.org/0000-0002-7608-2802 d e-mail: jugoslav.ackoski@ugd.edu.mk, ORCID iD: http://orcid.org/0000-0003-2782-3739 e e-mail: petreska_aleksandra@yahoo.com ORCID iD: http://orcid.org/0000-0003-3315-8966 DOI: 10.5937/vojtehg64-10258 FIELD: Defense Management - Logistics ARTICLE TYPE: Professional Paper ARTICLE LANGUAGE: English Abstract: The management of resources is one of the three most important areas of management in defense, besides the information management and the human resources management. By defining the concepts of Information Systems, decision support systems, expert systems, as well as their component parts, the possibilities they offer and the way of their functioning, we have made an attempt to present the Logistics Information System as an Expert System. The growing importance and the need for fast, timely and continuous logistic support of military units imposes the need for introducing a computer automated system which will increase the work efficiency and will save time to managers and decision makers at all levels of the Command and Control in the Ministry of Defense of the Republic of Macedonia. Having accurate information in appropriate time, especially regarding the availability of material resources, is a prerequisite for effectively managing logistic processes and for successfully performing tasks by military units. Key words: information systems, systems, resources, military, manager, management of resources, logistics, information use, expert systems. 851 P et re sk i, D ., et a l, E xp er t s ys te m fo r m an ag in g lo gi st ic p ro ce ss es , p p. 8 50 –8 65 Introduction The management of modern business processes must constantly find solutions to all turbulence in the working environment and pay attention to developing one’s own strategy in the area of: operations, automation, integration, information and utilization of resources. Knowledge, information and achievements are an imperative of our time and a major precondition for success in the work of any individual or organization. Today, business entities are trying to communicate quickly and accurately to send all the information to necessary users. This kind of needs exists in every sphere of the human activity: defense, science, technology, economics, arts, history, and more. The opportunity is available for instant access to information and its rapid exchange today, which is not a privilege of individuals or groups, but generally is a common good for everyone. The rapid development of technology of computer systems created an opportunity of their participation in the transmission of information. Applications for this purpose are being developed in various fields of life. Modern scientific and technological development imposes a substantial change in the manner of data collection, processing and presenting. The first rudiments of automated flow of information were made in the logistics system in the United States Army in terms of planning. The complexity of planning logistic support in multiple phase operations of modern armies overcomes the method of board and pen that was used before. Modern planning procedures depend on the use of automated programs to ensure their speed and to include all the details which have to match the pace of the operation (Edwards, 2004). For quality management decisions, we need quality knowledge which is inherent to human experts. This knowledge can be included in the software for supporting decision making. Software that integrates the knowledge from experts is called a support system in decision-making based on knowledge (Knowledge Based Decision Support System - KBDSS) or an intelligent system for supporting decision-making (Intelligent Decision Support System - IDSS). Support systems for decision making based on knowledge can increase the capabilities for decision-making by supplying a software tool that will directly support the decision. One of the main features of the decision-making process is based on information. Timely and accurate information is a key element and a prerequisite in the process of decision-making and the management of resources. Without the necessary information, data and parameters, it is impossible to make a decision. A system for decision support is one of the most modern tools used in decision-making because it allows the integration of knowledge and experience of the decision-maker with the database, which facilitates the decision-making process. 852 V O JN O T E H N IČ K I G LA S N IK / M IL IT A R Y T E C H N IC A L C O U R IE R , 2 01 6. , V ol 6 4, N o 3 The most commonly used software tools for supporting decision- making are expert systems. They are used in all areas of operation, in defense management as well, especially in providing logistic support. Logistics information management is a continuous process of: collecting, processing, preserving, analysing and presenting information on the course of the operation of the planned logistical support at all levels of command and control. Modern logistics systems and processes are almost unimaginable without adequate information support. A huge amount of information which has to be processed daily, the need for quality management of resources and quick decision-making cannot be realized without work environment based on information technology. It supports the decision- making bodies and optimizes the management of logistic support. The duty of all responsible authorities is to define precisely information objects (vehicles, persons, units, etc.), their characteristics that will be monitored and the form of reports which will be submitted to higher levels. The Logistics Information System as an expert system is a computer- supported system the goal of which is to support the management of logistic activities. It provides an opportunity for permanent optimization and improvement of working processes and the logistic support operation in general. Our aim is to present the logistic information system as an expert system, which is an important link in the decision-making process in the Ministry of Defense and the Army of the Republic of Macedonia. Categories of research Permanent social and technological development which attempt to promote and facilitate work imply the need to introduce information systems that contribute to speed, timely and accurate receiving, processing and distributing information to all levels of management. In the following part of our paper, we will examine several categories relevant to the subject of research: information system, decision support system, expert system, management, logistics management and logistics information system. The International Federation for Information Processing - IFIP defines the information system as follows: The information system is a system that collects, complements, stores, processes and delivers information relevant to the organization and society, in order to be available and usable by anyone who wants to use it, including the management, customers, employees and others. The Information System is an active social system that may or may not have to use information technology (International Federation for Information Processing: Technical Committees and Working Groups- Information Systems, 2013). 853 P et re sk i, D ., et a l, E xp er t s ys te m fo r m an ag in g lo gi st ic p ro ce ss es , p p. 8 50 –8 65 The information system is a set of methods, procedures and resources designed to facilitate the achievement of certain goals (Thierauf, 2001). In terms of the systematic approach, the information system is a sorted set of methods, processes and operations for: collection, storage, processing, transmission and distribution of data in one organization, including equipment used for these purposes and people engaged in these activities (Pojam informacionog sistema, 2013). The information system should be placed in a way that: it will be understood by all users, will be simple in presentation of information, will be secure and will allow presentation of the processed information in a very short time interval. There are different divisions of information systems. For example, some authors are dividing modern information systems in: operational information systems and systems for supporting decision-making in management. Systems for decision support Decision Support Systems-DSSs represent a powerful tool in making complex decisions and they are used in many fields, including defense. As an update of various disciplines, primarily management and information science, they have roots in the decision-making theory and are used in various areas of operation from natural sciences, technology, economics to education (Veljović, 2007, p.1). Systems for decision support are supporting all stages in the decision- making process, starting with the stage of formulating the problem through the design phase and the phase of selection, to implementation. Furthermore, they support various processes and styles of decision-making, they are adaptive to new operating needs, easy to use, they enable greater efficiency in the decision-making process, etc. They also need to provide timely information to managers, which would also be accurate, relevant and complete. The information should be shown in a proper form in order to be easily understood and managed (Veljović, 2007, pp.1-2). Many authors during their development had different interpretations of the term ‘system decision support’. Turban gives a very precise definition which, according to him, is: an interactive, flexible and adaptive computer-based information system designed primarily to support the solving of unstructured management problems for improving the decision-making process (Turban, 1995, p.5). Expert systems Expert systems in the artificial intelligence are computer programs designed to solve complex problems using reasoning based on expert knowledge which does not follow the procedures developed in the case of conventional programming (Turban, Aronson, 1988, p.15). 854 V O JN O T E H N IČ K I G LA S N IK / M IL IT A R Y T E C H N IC A L C O U R IE R , 2 01 6. , V ol 6 4, N o 3 The appearance of expert systems has greatly facilitated the work at places where complex and important decisions are made. These systems represent computer programs that simulate the behavior of persons or organizations with expert knowledge and experience in a particular area. Typically, these systems contain a knowledge base which contains accumulated experience and a set of rules to change the basis of knowledge of each particular situation described in the program. Sophisticated expert systems can be improved by adding knowledge to the base or set of rules. According to Edward Feigenbaum from Stanford University, one of the best experts in this area, who presents expert systems in the best way, uniting both of the above aspects, expert systems are intelligent computer programs that use the knowledge and procedures of locking to solve complex problems, which require human expertise and skill. The required level of knowledge along with locking mechanism for solving the problem can be considered as a model which is simulating the best expert in that area (Mišković, 2013, p.150). Besides the general division of expert systems which are divided into those that analyze a problem and those which perform a synthesis in the process of solving the problem, the expert systems are distributed based on the type of information they offer, such as: independent, consulting and systems "would be if ..." (Mišković, 2013, p.139). The key components of each expert system are: (Fig.1) knowledge base, deduction mechanism and user interface (King, 1989, p.21). Figure 1 – The basic structure of expert systems Рис. 1 – Базовая структура экспертной системы Slika 1 – Osnovna struktura ekspertskog sistema 855 P et re sk i, D ., et a l, E xp er t s ys te m fo r m an ag in g lo gi st ic p ro ce ss es , p p. 8 50 –8 65 The knowledge base is the basis of the expert system and contains the necessary knowledge for understanding, formulation and problem- solving, which contains information and regulations for the specific problem area (real system) that we want to make. Besides general knowledge, it contains special heuristics - rules and subjective assessments that guide the knowledge. A deduction mechanism has a task to find the right knowledge from the knowledge base and to develop new knowledge. It can be said that the mechanism of deduction is the most important activity which takes place in the expert system. Thanks to this activity, expert systems are intelligent systems. The process of locking is the one which animates the otherwise lifeless knowledge stored in the knowledge base and which knows how to apply it to the existing problems. Without the process of locking, expert systems would be only an Encyclopedia of information. The user interface is the third major component of any projected expert system. The main role of this component is to provide simple communication between the system and the user, i.e. to provide the following: easy communication, simple running of the program, control of input and output and quality interpretation of results (King, 1989, pp.22- 23). The process of expert systems functioning can be divided into the following five components: acquisition of knowledge, presentation / exhibition of knowledge, knowledge processing, interface components and explanation. Management of logistics information Management as a phenomenon is present in nearly every scientific discipline and covers all functions that successful functioning of society depends on. It represents a base of the modern approach to the management and training of organizations in a line with the long-term needs and requirements of the environment. According to Shermerhorn (Schermerhorn, 2002), management is a process of planning, organizing, managing and controlling the process and use of resources to achieve the objectives. He provides a schematic view of the management process, shown in Figure 2. 856 V O JN O T E H N IČ K I G LA S N IK / M IL IT A R Y T E C H N IC A L C O U R IE R , 2 01 6. , V ol 6 4, N o 3 Figure 2 – Schematic representation of the process of managing Рис. 2 – Схематическое изображение процесса управления Slika 2 – Šematski prikaz procesa upravljanja According to Woodruff and Thomas (Thomas, Woodruff, 1999), managers are responsible for four functions of the management process. Decision-making is the primary job of any manager, no matter the level or function. A decision is always a reaction to a situation, something that should be taken or needs to be changed. In order to make good decisions, managers must be familiar with modern methods and principles of decision- making. Modern managerial decision-making is based on scientific principles, quality information and effective assessment processes. As we already pointed out, management is an essential feature of modern activity of any organization, including the Ministry of Defense and the Army of the Republic of Macedonia. Logistic support management is a process that includes planning, decision-making, organization, cooperation and coordination, monitoring and reporting the implementation of logistic support. It is a segment in the overall system of defense management, implemented at all decision- making levels to optimize logistic support (Ministry of Defence of the Republic of Macedonia, 2011, pp.22-26). The management of logistics information connects the information technology with logistics processes and enables the authorities to facilitate decision-making, delivering actionable information in real time, as well considering the actual needs for logistic support (Andrejić, et al., 2009, p.59). The management of logistics information should cover all logistic functions and should enable the link between them and with other functional areas in the defense; hence the need to introduce an information system in logistics or the Logistics Information System. The logistics Information system is a computer-supported system which completely provides support in managing integrated logistic activities and management with the logistics system. Generating quality reports and timely information necessary for decision-making is the primary task of the logistics information system. 857 P et re sk i, D ., et a l, E xp er t s ys te m fo r m an ag in g lo gi st ic p ro ce ss es , p p. 8 50 –8 65 The mission of the logistics information system in the sphere of defense is seen in the planning, creation, development, monitoring, maintenance or continuous support of efficient and effective defense forces (Andrejić, et al., 2009, p.60). Logistics information system as an expert system for managing logistics processes Working with logistics information is a continuous process of collecting, processing, preservation, analysis and presentation of information on the course of operation of the planned logistic support at all levels. All this is aimed at supporting the decision-making bodies and at the optimization of logistics management support. The duty of all authorities is to define precisely the information objects (vehicles, persons, units) and their characteristics which will be monitored in the form of reports submitted to higher levels. The periodic and additional records, reports and collected information are in accordance with the needs and the available data. Modern working processes, modernization of institutions, influences from developed countries, emerging challenges and trends show the need for keeping up with these developments in all spheres of society, including the sphere of defense. The problems of today’s working practices (characterized by outdated methods, business techniques, technology and systems, data inconsistency, poor visibility of fixed assets and supplies, lack of accurate information and slow flow of information) have imposed the necessity of introducing an automated information system that will satisfy the following needs and requirements: compatibility with the NATO standards, the need for accurate and timely data, precision in planning, budget versus reality, better consumption control, sustainable and proactive planning, procurement and monitoring of supplies. In accordance with the Project for improvement of the Macedonian infrastructure and business services (MIESU - Macedonian Infrastructure and Enterprise Services Upgrade) two subprojects were initiated and conducted, related to: upgrading the infrastructure of Global Communications Information System (GCIS) and implementation of the application for the Logistics information system (The Government of the Republic of Macedonia, 2011, pp.4-5). The aim of the project concerning the global information communication system was building new IT infrastructure and improving the existing one in the country for the needs of the Ministry of Defence and the Army of the Republic of Macedonia. Now the entire network and communications equipment and services are based on open standards and they are compatible with the existing global information infrastructure. 858 V O JN O T E H N IČ K I G LA S N IK / M IL IT A R Y T E C H N IC A L C O U R IE R , 2 01 6. , V ol 6 4, N o 3 The purpose of the rest of the project was configuring, developing and constructing a logistics information system which will be implemented in the Ministry of Defense and the Army of the Republic of Macedonia. It is planned that this system consists of three subsystems: Information supply systems (ISS), Information maintenance system (IMS) and Information system for movement and transportation (ISMT). These three subsystems will be developed successively. The first phase is currently being implemented, which is the development and implementation of the Information Supply System. This system provides support for managing materials, resources and requirements for procurement and supply of background material items for the Ministry of Defense and the Army. The Logistics information system provides support for the automated information system for the following functional processes: forecasting and planning materials, requirements for the purchase of stocks, origin of supply, contracts, purchasing, receiving, inventory transfers between units, delivery / dispensing material items, requests for supply recharge and asset management. These eleven procedural models are closely interrelated and there is a need for exchanging information. The interconnection between the processes is shown in the diagram below (Figure 3) in order to present the flow of information between each model. Figure 3 – Procedural areas of the Logistics Information System Рис. 3 – Процессные области Логистической информационной системы Slika 3 – Procesne oblasti Logističkog informacionog sistema 859 P et re sk i, D ., et a l, E xp er t s ys te m fo r m an ag in g lo gi st ic p ro ce ss es , p p. 8 50 –8 65 In the framework of the logistics information system based on the needs for accessing data and their analysis in daily operations, the key logistics customers are defined. The logistics information system has led to changes in the performance of daily tasks of employees in the military, especially those involved in the logistics business processes. Documents and data which used to be processed manually and on paper now are stored electronically in the system for promoting greater efficiency and greater control. It was already stated that the Logistics information system was based on eleven contractual process areas classified in three modules: • From finance to control - module which is covering the business areas: integration with financial and material planning and forecasting; • From procurement to payment - module which is covering the business areas: procurement requirements, sources of supply, contracts, procurement, requests for recharging; • From admission to delivery - module which is covering the business areas: reception of materials, inventory transfers between units and issuing material delivery and asset management. Additionally provided requirements are fuel and planning projects (Government of the Republic of Macedonia, 2010). Let us remember that according to the definition, the expert system is a computer program that simulates the behavior of one person or organization that has expert knowledge and experience in a particular area. Indeed, this definition proves to be correct because the Logistics information system was created by experts whose expertise and experience it incorporated and transformed into a knowledge base which is the basis of this system. Two groups of experts were involved in the creation of the Logistic information system. First, many experts from different fields: finance, supply management, logistics, procurement, cost management, warehouse management, etc. This group of experts created the user software Dynamics AX. The other group was composed of experts from the Ministry of Defence from the respective areas who incorporated their expertise and long experience to modify the Dynamics AX and to make an adequate knowledge base which is the basis for the Logistic information system. In the time of establishing the system, the experts provided an opportunity for constant updating and modernization depending on the needs and requirements brought by the new era. Benefits from the use of the Logistics information system Logistics management is considered as a critical link in the process of defense management. For making better decisions and for proper management of resources, the necessary information is required to be 860 V O JN O T E H N IČ K I G LA S N IK / M IL IT A R Y T E C H N IC A L C O U R IE R , 2 01 6. , V ol 6 4, N o 3 timely and accurate. Due to the need for the codification of assets and compatibility of the domestic logistics system with the NATO systems, and the need for the accuracy of inventory planning, a project was launched for modernizing the way of working in the field of logistics which resulted in the decision to implement the Logistics information system. The new Logistics information system has these sets of capabilities: development, construction, implementation and support of eleven agreed processes / modules, fuel management, export data to other systems, prototype of functionality for maintenance, repair and overhaul, as well as determining the needs for the module of transport and distribution. The list of benefits from the implementation of the Logistics information system is remarkable, but, as some of the most important and greatest benefits, we will mention the following: • speed and accuracy of data; • efficiency in the functioning; • interoperability with NATO; • reduced operating costs; • sustainable Logistics information system and architecture that is easy to upgrade; • benefit on the national level, and others. In the following part of our paper, we will give a short description for each of the benefits. Speed and accuracy of data The proper planning and forecasting of purposes are essential for the proper resource allocation and management of courses to ensure their rapid and accurate determination. Also an important element is the corresponding warehouses for storing goods providing tactical and strategic flexibility. The Logistics information system has replaced the outdated system which consisted of a lot of data sources. Their inconsistency, poor visibility of the underlying assets and consumables were replaced with high precision in the planning process, improving the control over the consumption and, very importantly, the sustainable and proactive planning of procurement and monitoring of supplies. Month-long data updates, slow flow and lack of accurate information are replaced with accurate and timely data given in the Logistics information system which maintains the accuracy in daily operations. All this in reality means that daily transactions in terms of material resources or material lists are transferred to a responsible person with one click, instead of spending time in a time-consuming process. All this affects the 861 P et re sk i, D ., et a l, E xp er t s ys te m fo r m an ag in g lo gi st ic p ro ce ss es , p p. 8 50 –8 65 accuracy of data, because in real time we have precise data about what and where something can be found, going to the tiniest details, for example: Which shelf, room or building has what we are looking for. Efficient functioning Defense forces must be ready at any time to carry out their tasks and missions in a controlled manner, thereby avoiding unwanted situations of lack of exact quantity supplies, ammunition, food, fuel and spare parts. This system forms the basis for the efficient functioning of the armed forces in a way that enables decision makers at every level to have visibility of all necessary data. Interoperability The desire of Macedonia to become a NATO member and the change of the mission of the armed forces are leading to the needs for the Army to be flexible in terms of fulfillment of the standards. The system is designed to enable complete management of logistics defense processes and to allow a kind of transformation of the organization of the defense at the strategic level. The Logistics information system significantly improves the interoperability with allied Logistics systems and the codifing standards. Reduced operating costs The possibility of obtaining accurate data for the assets in each segment of the Army or unit, warehouse, etc, which contributes to more accurate creation of financial plans and, accordingly, the plans for procurement, which results in proper and economical use of the budget. With visibility of materials on inventories in warehouses, managers are allowed to perform restricted resources from the nearest storehouse, which leads to reduced costs in terms of movement and transport. Sustainable Logistics information system and architecture that is easy to upgrade The Logistics information system, as an expert system, is selected for use in the defense sector because of its scalable architecture that allows extensibility of processes, in addition to the latest trends of modernization and transformation of the Army. 862 V O JN O T E H N IČ K I G LA S N IK / M IL IT A R Y T E C H N IC A L C O U R IE R , 2 01 6. , V ol 6 4, N o 3 After the implementation of the first stage, we have already seen a possibility of extending the decision for closing the gap for financial management, transportation and distribution and resources management of selected tracking assets as the next stages of our engagement. The benefits at the national level The state will have a possibility to implement advanced technology available from global partners, such as Hewlett Packard to identify, design and implement technology for transforming and extending the knowledge of employees in the government and in the military sector using the best practices and leading technologies. Conclusion In the decision-making process, it is important to consider alternative solutions, recognize possible consequences of their use and compare options. The process is further complicated with the external environment (new technologies, information systems, advanced research and globalization) and the fact that everything is becoming more and more complex. Because of all the above reasons, managers are forced to use new techniques and tools for the decision-making process and they need to rely on information technology. Modern logistics systems and processes are unthinkable without adequate information support. A huge amount of information which needs to be processed daily, the need for quality management of resources and quick decision- making, and the reduction of personnel in order to use resources more economically, cannot be realized without adequate support of a quality information system. Today we spent less energy, time and personnel for processing simple information which would be adequately used in the process of decision-making from the lowest tactical to the strategic level. With the introduction of the Logistics information system, we will have lower economic costs in the logistics support system. The rapid transmission of information, timely locating the needs of units, available data for the status of resources, affects on reduction of storage material assets in stocks (Andrejić, et al., 2009, p.15). This application itself was a part of Microsoft software, specifically Microsoft Dynamics AX. This solution was chosen because of several key elements, such as: timing of implementation which entails reduced costs; the fact that it is an integrated ERP (Enterprise Resource 863 P et re sk i, D ., et a l, E xp er t s ys te m fo r m an ag in g lo gi st ic p ro ce ss es , p p. 8 50 –8 65 Planning)-solution concerning financial management, payment and purchasing, and supply chain of logistics; its flexible configuration (application which is global, but adapted to the Macedonian conditions), and in particular the important ladder architecture which allows the extensibility of processes and roles. The last element relates to the intentions of the Ministry of Defense to integrate all functions of the Logistics system though different stages. Currently, the Logistics information system is in its first phase, which supports the supply area with their supplying subsystems. References Andrejić, D.M., Milenkov A.M., Sokolović S.V., 2009, Logistički informacioni sistem, Beograd, Vojna akademija – Katedra logistike, pp.59-60. Edwards, J.E., 2004, Combat service support guide - 4 th Edition, Mechanicsburg, PA 17055, USA, Stackpole Books. Government of Republic of Macedonia, Project: MIESU - Macedonian Infrastructure and Enterprise Services Upgrade, 2011, Skopje, pp.4-5. Government of Republic of Macedonia, Strategy for development of e-Contents 2010 - 2015, 2010, Ministry of Information Society, Skopje. International Federation For Information Processing: Technical Committees and Working Groups–Information Systems, Retrieved from: http://www.ifip.org/index.php?option=com_content&task=view&id=185&Itemid=509, 10.09.2013. King, D., 1989, Modeling and reasoning: Inegrating Decision support with Experts Systems-Englewood Cliffs, NJ, Prentice-Hall, pp.21-23. Ministry of Defence on the Republic of Macedonia, 2011, Concept for logistic support of defense of the Republic of Macedonia, Skopje, GS ARM, pp.22-26. Mišković, V., 2013, Sistemi za podrsku odlucivanju, Beograd, Univerzitet Singidunum, pp.139-150. Pojam informacionog sistema, Retrieved from: http://www.znanje.org/knjige/computer/access/access_01/POJAM%20INFORMACIONO G%20SISTEMA.htm, 11.10.2013. Schermerhorn, R.J., 2002, Management Seventh Edition, Marshall Texas, USA, Wiley College. Thierauf, R., 2001, Effective business intelligence systems, London, Quorum Books. Thomas, R.R., Woodruff, M.I., 1999, Building a house for diversity, New York, Amacom. Turban, E., 1995, Decision support and expert systems: management support systems-Englewood Cliffs, N.J., Prentice Hall, p.5. Turban, E., Aronson, J., 1988, Decision Support Systems and Intelligent Systems- Upper Saddle River, NJ: Prentice Hall, Inc, p.15. Veljović, A., 2007, Sistem za podršku odlučivanja, Beograd, CET Čitalište 67, pp.1-2. 864 V O JN O T E H N IČ K I G LA S N IK / M IL IT A R Y T E C H N IC A L C O U R IE R , 2 01 6. , V ol 6 4, N o 3 ЭКСПЕРТНАЯ СИСТЕМА МЕНАДЖМЕНТА В ЛОГИСТИЧЕСКОМ ПРОЦЕССЕ Драге T. Петрескиa, Андрей П. Илиевa, Лазарь Б. Гюровa, Югослав З. Aцкоскиa, Александра Д. Петрескаб a Военная академия „Генерал Михаило Aпостолски“, Скопье, Республика Македония б Факультет электротехники и информационных технологий, Скопье, Республика Македония ОБЛАСТЬ: менеджмент в области обороны, логистика ТИП СТАТЬИ: профессиональная статья ЯЗЫК СТАТЬИ: английский Резюме: Менеджмент ресурсов наряду с информационным менед- жментом и менеджментом персонала является одной из важне- йших областей управления в системе обороны. В результате определения таких понятий как: информаци- онная система, система поддержки принятия решений, модуль- ная система экспериментов, и их составляющих, а также в ре- зультате исследований их возможностей и эффективности, предпринята попытка преобразования логистической информа- ционной системы в экпертную систему управления. Для быстрой, своевременной и бесперебойной логистиче- ской поддержки военных подразделений большое значение игра- ет автоматизированная информационная система, повышающ- ая эффективность работы и экономящая время менеджерам и командующим на принятия решений и мониторинг в Министер- стве обороны Республики Македония. Своевременное получение точной информации о наличии материальных ресурсов является главным условием эффектив- ного управления логистическим процессом и успешного выпол- нения поручений, выданных воинскими подразделениями. Ключевые слова: информационные системы, системы, ресурсы, военные, менеджер, менеджмент логистики, преобразование данных, экспертная система. EKSPERTSKI SISTEM ZA MENADŽIRANJE LOGISTIČKIH PROCESA Drage T. Petreskia, Andrej P. Ilieva, Lazar B. Gjurova, Jugoslav Z. Ackoskia, Aleksandra D. Petreskab a Vojna akademija „General Mihailo Apostolski”, Skoplje, Republika Makedonija b Fakultet elektrotehnike i informacionih tehnologija, Skoplje, Republika Makedonija OBLAST: menadžment u odbrani, logistika VRSTA ČLANKA: stručni članak JEZIK ČLANKA: engleski 865 P et re sk i, D ., et a l, E xp er t s ys te m fo r m an ag in g lo gi st ic p ro ce ss es , p p. 8 50 –8 65 Sažetak: Menadžment resursa predstavlja jednu od tri najvažnije oblasti za menadžment u odbrani, između informatičkog menadžmenta i me- nadžmenta ljudskim resursima. Preko definisanja pojmova informatički sistem, sistem za podršku odlučivanju, sistem za eksperimentisanje, kao i komponente koje ih sačinjavaju, mogućnosti koje nude i način nji- hovog funkcionisanja, učinjen je pokušaj da se prevede logistički infor- matički sistem u ekspertski sistem. Veliki značaj za brzu, pravovreme- nu i kontinuiranu logističku podršku vojnih jedinica nameće potrebu za uvođenjem informatički automatizovanog sistema koji će da poveća efikasnost u radu i uštedi vreme menadžerima i donosiocima odluka komandovanja i kontrole u Ministarstvu odbrane Republike Makedoni- je. Pravovremeno raspolaganje tačnim informacijama koje se odnose na raspoloživost materijalnih resursa predstavlja preduslov za efikasno upravljanje logističkim procesima i uspešno izvršavanje zadataka voj- nih jedinica. Ključne reči: informacioni sistemi, sistemi, resursi, vojni, menadžer, menadžment resursa, logistika, koričenje informacija, ekspertski si- stem. Paper received on / Дата получения работы / Datum prijema članka: 13. 02. 2016. Manuscript corrections submitted on / Дата получения исправленной версии работы / Datum dostavljanja ispravki rukopisa: 23. 03. 2016. Paper accepted for publishing on / Дата окончательного согласования работы / Datum konačnog prihvatanja članka za objavljivanje: 25. 03. 2016. © 2016 The Authors. Published by Vojnotehnički glasnik / Military Technical Courier (www.vtg.mod.gov.rs, втг.мо.упр.срб). This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/rs/). © 2016 Авторы. Опубликовано в "Военно-технический вестник / Vojnotehnički glasnik / Military Technical Courier" (www.vtg.mod.gov.rs, втг.мо.упр.срб). Данная статья в открытом доступе и распространяется в соответствии с лицензией "Creative Commons" (http://creativecommons.org/licenses/by/3.0/rs/). © 2016 Autori. Objavio Vojnotehnički glasnik / Military Technical Courier (www.vtg.mod.gov.rs, втг.мо.упр.срб). Ovo je članak otvorenog pristupa i distribuira se u skladu sa Creative Commons licencom (http://creativecommons.org/licenses/by/3.0/rs/). << /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Dot Gain 20%) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Warning /CompatibilityLevel 1.3 /CompressObjects /Tags /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /LeaveColorUnchanged /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams false /MaxSubsetPct 100 /Optimize true /OPM 1 /ParseDSCComments true /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo true /PreserveFlatness true /PreserveHalftoneInfo true /PreserveOPIComments true /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Preserve /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true /AgencyFB-Bold /AgencyFB-Reg /Albertus-ExtraBold /Albertus-Medium /AlbertusMT /AlbertusMT-Italic /AlbertusMT-Light /Algerian /AntiqueOlive /AntiqueOlive-Bold /AntiqueOlive-Compact /AntiqueOlive-Italic /AntiqueOlive-Roman /Apple-Chancery /Arial-Black /Arial-BoldItalicMT /Arial-BoldMT /Arial-ItalicMT /ArialMT /ArialNarrow /ArialNarrow-Bold /ArialNarrow-BoldItalic /ArialNarrow-Italic /ArialRoundedMTBold /ArialUnicodeMS /AvantGarde-Book /AvantGarde-BookOblique /AvantGarde-Demi /AvantGarde-DemiOblique /BaskOldFace /Bauhaus93 /BellMT /BellMTBold /BellMTItalic /BerlinSansFB-Bold /BerlinSansFBDemi-Bold /BerlinSansFB-Reg /BernardMT-Condensed /BlackadderITC-Regular /Bodoni /Bodoni-Bold /Bodoni-BoldItalic /Bodoni-Italic /BodoniMT /BodoniMTBlack /BodoniMTBlack-Italic /BodoniMT-Bold /BodoniMT-BoldItalic /BodoniMTCondensed /BodoniMTCondensed-Bold /BodoniMTCondensed-BoldItalic /BodoniMTCondensed-Italic /BodoniMT-Italic /BodoniMTPosterCompressed /Bodoni-Poster /Bodoni-PosterCompressed /BookAntiqua /BookAntiqua-Bold /BookAntiqua-BoldItalic /BookAntiqua-Italic /Bookman-Demi /Bookman-DemiItalic /BookmanITCbyBT-Demi /BookmanITCbyBT-DemiItalic /BookmanITCbyBT-Light /BookmanITCbyBT-LightItalic /Bookman-Light /Bookman-LightItalic /BookmanOldStyle /BookmanOldStyle-Bold /BookmanOldStyle-BoldItalic /BookmanOldStyle-Italic /BookshelfSymbolSeven /BradleyHandITC /BritannicBold /Broadway /BrushScriptMT /CalifornianFB-Bold /CalifornianFB-Italic /CalifornianFB-Reg /CalisMTBol /CalistoMT /CalistoMT-BoldItalic /CalistoMT-Italic /Candid /Castellar /Centaur /Century /CenturyGothic /CenturyGothic-Bold /CenturyGothic-BoldItalic /CenturyGothic-Italic /CenturySchoolbook /CenturySchoolbook-Bold /CenturySchoolbook-BoldItalic /CenturySchoolbook-Italic /CGOmega /CGOmega-Bold /CGOmega-BoldItalic /CGOmega-Italic /CGTimes /CGTimes-Bold /CGTimes-BoldItalic /CGTimes-Italic /CHelv /CHelvBold /CHelvBoldItalic /CHelv-Italic /Chicago /Chiller-Regular /CHVojska /CHVojska-Bold /CHVojska-BoldItalic /CHVojska-Italic /CirTimes /CirTimes_New_Roman /CirTimesBold /CirTimesBoldItalic /CirTimesItalic /Clarendon /Clarendon-Bold /Clarendon-Condensed-Bold /Clarendon-Light /ColonnaMT /ComicSansMS /ComicSansMS-Bold /CooperBlack /CooperBlack-Italic /CopperplateGothic-Bold /CopperplateGothic-Light /Copperplate-ThirtyThreeBC /Copperplate-ThirtyTwoBC /Coronet /Coronet-Regular /Courier /Courier-Bold /Courier-BoldOblique /CourierNewPS-BoldItalicMT /CourierNewPS-BoldMT /CourierNewPS-ItalicMT /CourierNewPSMT /Courier-Oblique /CTVojska /CTVojska-Bold /CTVojska-BoldItalic /CTVojska-Italic /CurlzMT /Decor /EdwardianScriptITC /Elephant-Italic /Elephant-Regular /English157BT-Regular /EngraversMT /ErasITC-Bold /ErasITC-Demi /ErasITC-Light /ErasITC-Medium /EstrangeloEdessa /Euclid /Euclid-Bold /Euclid-BoldItalic /EuclidExtra /EuclidExtra-Bold /EuclidFraktur /EuclidFraktur-Bold /Euclid-Italic /EuclidMathOne /EuclidMathOne-Bold /EuclidMathTwo /EuclidMathTwo-Bold /EuclidSymbol /EuclidSymbol-Bold /EuclidSymbol-BoldItalic /EuclidSymbol-Italic /Eurostile /Eurostile-Bold /Eurostile-BoldExtendedTwo /Eurostile-ExtendedTwo /FelixTitlingMT /FencesPlain /FootlightMTLight /FormalScript421BT-Regular /ForteMT /FranklinGothic-Book /FranklinGothic-BookItalic /FranklinGothic-Demi /FranklinGothic-DemiCond /FranklinGothic-DemiItalic /FranklinGothic-Heavy /FranklinGothic-HeavyItalic /FranklinGothic-Medium /FranklinGothic-MediumCond /FranklinGothic-MediumItalic /FreestyleScript-Regular /FrenchScriptMT /Garamond /Garamond-Antiqua /Garamond-Bold /Garamond-Halbfett /Garamond-Italic /Garamond-Kursiv /Garamond-KursivHalbfett /Gautami /Geneva /Georgia /Georgia-Bold /Georgia-BoldItalic /Georgia-Italic /Gigi-Regular /GillSans /GillSans-Bold /GillSans-BoldCondensed /GillSans-BoldItalic /GillSans-Condensed /GillSans-ExtraBold /GillSans-Italic /GillSans-Light /GillSans-LightItalic /GillSansMT /GillSansMT-Bold /GillSansMT-BoldItalic /GillSansMT-Condensed /GillSansMT-ExtraCondensedBold /GillSansMT-Italic /GillSans-UltraBold /GillSans-UltraBoldCondensed /GloucesterMT-ExtraCondensed /Goudy /Goudy-Bold /Goudy-BoldItalic /Goudy-ExtraBold /Goudy-Italic /GoudyOldStyleT-Bold /GoudyOldStyleT-Italic /GoudyOldStyleT-Regular /GoudyStout /Haettenschweiler /HarlowSolid /Harrington /Helvetica /Helvetica-Bold /Helvetica-BoldOblique /Helvetica-Condensed /Helvetica-Condensed-Bold /Helvetica-Condensed-BoldObl /Helvetica-Condensed-Oblique /HelveticaLat /HelveticaLatBold /Helvetica-Narrow /Helvetica-Narrow-Bold /Helvetica-Narrow-BoldOblique /Helvetica-Narrow-Oblique /Helvetica-Oblique /HighTowerText-Italic /HighTowerText-Reg /HoeflerText-Black /HoeflerText-BlackItalic /HoeflerText-Italic /HoeflerText-Ornaments /HoeflerText-Regular /Impact /ImprintMT-Shadow /InformalRoman-Regular /JoannaMT /JoannaMT-Bold /JoannaMT-BoldItalic /JoannaMT-Italic /Jokerman-Regular /JuiceITC-Regular /Kartika /KristenITC-Regular /KunstlerScript /Latha /LatinWide /LetterGothic /LetterGothic-Bold /LetterGothic-BoldSlanted /LetterGothic-Italic /LetterGothic-Slanted /LHVojska /LHVojska-Bold /LHVojska-BoldItalic /LHVojska-Italic /LTVojska /LTVojska-Bold /LTVojska-BoldItalic /LTVojska-Italic /LubalinGraph-Book /LubalinGraph-BookOblique /LubalinGraph-Demi /LubalinGraph-DemiOblique /LucidaBright /LucidaBright-Demi /LucidaBright-DemiItalic /LucidaBright-Italic /LucidaCalligraphy-Italic /LucidaConsole /LucidaFax /LucidaFax-Demi /LucidaFax-DemiItalic /LucidaFax-Italic /LucidaHandwriting-Italic /LucidaSans /LucidaSans-Demi /LucidaSans-DemiItalic /LucidaSans-Italic /LucidaSans-Typewriter /LucidaSans-TypewriterBold /LucidaSans-TypewriterBoldOblique /LucidaSans-TypewriterOblique /LucidaSansUnicode /Magneto-Bold /MaiandraGD-Regular /Mangal-Regular /Marigold /MaturaMTScriptCapitals /MicrosoftSansSerif /Mistral /Modern-Regular /Monaco /MonaLisa-Recut /MonotypeCorsiva /MonotypeSorts /MSOutlook /MSReferenceSansSerif /MSReferenceSpecialty /MT-Extra /MT-Symbol /MVBoli /MyriadWebPro /MyriadWebPro-Bold /MyriadWebPro-Condensed /MyriadWebPro-CondensedItalic /MyriadWebPro-Italic /NewCenturySchlbk-Bold /NewCenturySchlbk-BoldItalic /NewCenturySchlbk-Italic /NewCenturySchlbk-Roman /NewYork /NiagaraEngraved-Reg /NiagaraSolid-Reg /OCRAbyBT-Regular /OCRAExtended /OCRB10PitchBT-Regular /OldChurchSlavonicCyr /OldEnglishTextMT /Onyx /Optima /Optima-Bold /Optima-BoldItalic /Optima-Italic /Oxford /PalaceScriptMT /Palatino-Bold /Palatino-BoldItalic /Palatino-Italic /PalatinoLinotype-Bold /PalatinoLinotype-BoldItalic /PalatinoLinotype-Italic /PalatinoLinotype-Roman /Palatino-Roman /Papyrus-Regular /Parchment-Regular /Perpetua /Perpetua-Bold /Perpetua-BoldItalic /Perpetua-Italic /PerpetuaTitlingMT-Bold /PerpetuaTitlingMT-Light /Playbill /PoorRichard-Regular /Pristina-Regular /Raavi /RageItalic /Ravie /Rockwell /Rockwell-Bold /Rockwell-BoldItalic /Rockwell-Condensed /Rockwell-CondensedBold /Rockwell-ExtraBold /Rockwell-Italic /ScriptMTBold /Serbian-Elegant /Serbian-Elegant-Bold /Serbian-Elegant-Bold-Italic /Serbian-Elegant-Italic /ShowcardGothic-Reg /Shruti /SnapITC-Regular /StempelGaramond-Bold /StempelGaramond-BoldItalic /StempelGaramond-Italic /StempelGaramond-Roman /Stencil /Sylfaen /Symbol /SymbolMT /Taffy /Tahoma /Tahoma-Bold /TempusSansITC /Times-Bold /Times-BoldItalic /Times-Italic /TimesNewRoman /TimesNewRomanBold /TimesNewRomanBoldItalic /TimesNewRomanItalic /TimesNewRomanPS-BoldItalicMT /TimesNewRomanPS-BoldMT /TimesNewRomanPS-ItalicMT /TimesNewRomanPSMT /Times-Roman /Trebuchet-BoldItalic /TrebuchetMS /TrebuchetMS-Bold /TrebuchetMS-Italic /Tunga-Regular /TwCenMT-Bold /TwCenMT-BoldItalic /TwCenMT-Condensed /TwCenMT-CondensedBold /TwCenMT-CondensedExtraBold /TwCenMT-Italic /TwCenMT-Regular /Univers /Univers-Bold /Univers-BoldExt /Univers-BoldExtObl /Univers-BoldItalic /Univers-BoldOblique /Univers-Condensed /Univers-CondensedBold /Univers-Condensed-Bold /Univers-Condensed-BoldItalic /Univers-CondensedBoldOblique /Univers-Condensed-Medium /Univers-Condensed-MediumItalic /Univers-CondensedOblique /Univers-Extended /Univers-ExtendedObl /Univers-Light /Univers-LightOblique /Univers-Medium /Univers-MediumItalic /Univers-Oblique /UstavIzvorni-Medium /Verdana /Verdana-Bold /Verdana-BoldItalic /Verdana-Italic /VinerHandITC /Vivaldii /VladimirScript /Vrinda /Webdings /Wingdings2 /Wingdings3 /Wingdings-Regular /YUTimesNewRoman /YUTimesNewRomanBold /YUTimesNewRomanBoldItalic /YUTimesNewRomanItalic /ZapfChancery-MediumItalic /ZapfDingbats /ZWAdobeF ] /NeverEmbed [ true ] /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 300 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 600 /ColorImageDepth 8 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.00000 /EncodeColorImages true /ColorImageFilter /FlateEncode /AutoFilterColorImages false /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /ColorImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 300 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 600 /GrayImageDepth 8 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.00000 /EncodeGrayImages true /GrayImageFilter /FlateEncode /AutoFilterGrayImages false /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /GrayImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 1200 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 600 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 >> /AllowPSXObjects false /CheckCompliance [ /None ] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description << /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000500044004600206587686353ef901a8fc7684c976262535370673a548c002000700072006f006f00660065007200208fdb884c9ad88d2891cf62535370300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef653ef5728684c9762537088686a5f548c002000700072006f006f00660065007200204e0a73725f979ad854c18cea7684521753706548679c300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /DAN <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> /DEU <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> /ESP <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> /FRA <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> /ITA <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> /JPN <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> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020b370c2a4d06cd0d10020d504b9b0d1300020bc0f0020ad50c815ae30c5d0c11c0020ace0d488c9c8b85c0020c778c1c4d560002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken voor kwaliteitsafdrukken op desktopprinters en proofers. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <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> /PTB <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> /SUO <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> /SVE <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> /ENU (Use these settings to create Adobe PDF documents for quality printing on desktop printers and proofers. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.) >> /Namespace [ (Adobe) (Common) (1.0) ] /OtherNamespaces [ << /AsReaderSpreads false /CropImagesToFrames true /ErrorControl /WarnAndContinue /FlattenerIgnoreSpreadOverrides false /IncludeGuidesGrids false /IncludeNonPrinting false /IncludeSlug false /Namespace [ (Adobe) (InDesign) (4.0) ] /OmitPlacedBitmaps false /OmitPlacedEPS false /OmitPlacedPDF false /SimulateOverprint /Legacy >> << /AddBleedMarks false /AddColorBars false /AddCropMarks false /AddPageInfo false /AddRegMarks false /ConvertColors /NoConversion /DestinationProfileName () /DestinationProfileSelector /NA /Downsample16BitImages true /FlattenerPreset << /PresetSelector /MediumResolution >> /FormElements false /GenerateStructure true /IncludeBookmarks false /IncludeHyperlinks false /IncludeInteractive false /IncludeLayers false /IncludeProfiles true /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe) (CreativeSuite) (2.0) ] /PDFXOutputIntentProfileSelector /NA /PreserveEditing true /UntaggedCMYKHandling /LeaveUntagged /UntaggedRGBHandling /LeaveUntagged /UseDocumentBleed false >> ] >> setdistillerparams << /HWResolution [2400 2400] /PageSize [1734.803 2245.040] >> setpagedevice 
work_upk45tn4afenfhldwybnpdskpy	----	Evolutionary Coincidence–Based Ontology Mapping Extraction Vahed Qazvinian Hassan Abolhassani Seyed H. HAERI (Hossein) Babak Bagheri Hariri Web Intelligent Lab, Department of Computer Engineering, Sharif University of Technology and School of Computer Science, Institute for Studies in Theoretical Physics and Mathematics (IPM) Tehran, Iran qazvinian@ce.sharif.edu, abolhassani@sharif.edu, shhaeri@ce.sharif.edu, hariri@ce.sharif.edu Abstract Ontology Matching is a process for selection of a good alignment across entities of two (or more) Ontologies. This can be viewed as a two phase process of: 1) ap- plying a similarity measure to find the correspondence of each pair of entities from two ontologies, and 2) Ex- traction of an optimal or near optimal mapping. This paper is focused on the second phase and introduces our evolutionary approach for that. To be able to do so, we need a mechanism to score different possi- ble mappings. Our solution is a weighting mechanism named Coincidence-Based Weighting – as explained in the paper. On that basis, a Genetic Algorithm is then introduced to create better mappings in succes- sive iterations. We will explain how we code a map- ping as well as our Crossover and Mutation functions. Evaluations of the algorithm is shown and discussed in the paper too. 1 INTRODUCTION Semantic web is rather a new concept, and is defined so that the agents will be able to understand Web content and communicate through it, as like as hu- man beings doing so now. Traditional knowledge- based systems were centralized, but on the other hand, semantic web is distributed and heterogenous. As a matter of fact, and according to (Mitra, Noy & Jaiswal 2003): “Information sources, even those from the same domain, are heterogenous in nature”. This heterogeneity has resulted in designing ontolo- gies to lessen the difficulties of agents’ understandings and communications. However, still another problem exists: ontologies themselves may have heterogene- ity. This is when two ontologies are trying to express same knowledge or concepts but they use different languages or words (Euzenat & Valtchev 2004). Ontology Alignment(OA) is a proposed solution to this problem by introducing a (proper) mapping of en- tities in two ontologies from two (different) domains. (Bouquet, Euzenat, Franconi, Serafini, Stamou & Tessaris 2004a) defines OA as: ... given two ontologies which describe each a set of discrete entities (which can be classes, properties, rules, predicates, etc.), find the relationships (e.g., equivalence or subsumption) holding between these enti- ties. Figure 1 shows a simplified OA framework. As shown in the figure, to extract an alignment, it is customary to first apply some measures (simple or complex) to reach to some initial similarity values. To this respect, there are already a vast amount of research works in the literature discussing about lexical and structural measures suitable for OA. Having such similarity val- ues, the next problem is how to form an ideal map- ping. We refer to this problem as Mapping Extrac- tion. The goal is to find correspondence of entities among two ontologies such that the overall similarity value is maximal. Therefore it can be viewed as an op- timization problem in which evolutionary approaches could be a legal solution. Figure 1: A Simplified Alignment Framework In this paper we introduce a genetic based algo- rithm for the Mapping Extraction problem. First we have an explanation of the related works in section 2. Then in section 3 explanations about graph theo- retical bases we use through the paper is given. To have a measure for calculation of how good a mapping is, the paper discusses coincidence based weighting in section 4. Our evolutionary algorithm is explained in section 5 and in section 6 evaluations are discussed. We also provide conclusions and explanations about our future works in Section 7. 2 RELATED WORKS Unfortunately and as stated in (Bouquet, Euzenat, Franconi, Serafini, Stamou & Tessaris 2004b), works on ontology extractions are not so common. How- ever, current researches on ontology mapping and its applications entails a large number of fields ranging from machine learning, concept lattices, and formal theories to heuristics, and linguistics. There are some similar works to match graphs, and trees (Hopcroft & Karp 1973, Papadimitriou & Steiglitz 1998), data- base schemas (Rahm & Bernstein 2001) and even in clustering compound objects with a machine learning technique (Bisson 1992). (Kalfoglou & Schorlemmer 2003) have come with a comprehensive review and presentations on the methods and approaches and the state of the art in ontology aligning. Related works to our research are of the following two categories: • The Ontology Alignment weighting and similar- ity measures. These works focus on introduction of new similarity measures between concepts of two ontologies, and a weight function to evaluate an alignment between two ontologies. • The Ontology Mapping Extractions, in which the researches try to address the problem of align- ment extraction and propose methods to find a (proper) alignment among many different candi- dates. There are also some works which address both prob- lems simultaneously. We will have a quick review on each category in the following two subsections. 2.1 Similarity Measures and Ontology Align- ment There are considerable amount of previous works on similarity measures and ontology alignment. Some standards of metrics are acknowledged and defined as in the CommonKADS methodology (Schreiber, de Hoog, Akkermans, Anjewierden, Shadbolt & de Velde 2000), or OntoWeb EU thematic network (OntoWeb. A survey on ontology tools. 2002), which are partly endorsed by recognized bodies. Also there have been some works on finding similari- ties of entities in two ontologies based on their struc- tural standings. (Valtchev 1999) computes the dis- similarity of elements in a hierarchy based on their distances from closest common parent. The Up- ward Cotopic distance is introduced by (Maedche & Zacharias 2002) where they find dissimilarity of en- tities in hierarchies of ontologies. (Zhong, Zhu & Y. Li 1995) introduces a measure to calculate simi- larity of WordNet1 concepts, i.e. a single hierarchy. In it, the similarity is computed based on the closest common parent and distance of the two entities from the root. On the other hands, some methods tend to- ward a trade-off between different features such as ef- ficiency and quality, as in QOM (Ehrig & Staab 2003), and some have used approaches to integrate various similarity methods (Ehrig & Sure 2004). Also compound metrics get use of simple measures by combining them, and hoping to improve the result of the mapping between two ontologies. One approach has been to define each measure as a dimension to find the Minkowski distance of two objects (Euzenat, Bar- rasa, Bouquet & Bo 2004). As introduced in (Euzenat et al. 2004) another approach for this problem has been a weighted average of features in which weights can be calculated by a machine learning technique. For example, Glue (Doan, Domingos & Halevy 2003) builds the similarity matrix by a machine learning ap- proach. Also in APFEL (M. Ehrig 2005) weights for each feature is calculated using Decision Trees. The user only has to provide some ontologies with known correct alignments. The learned decision tree is then used for aggregation and interpretation of the sim- ilarities. Abolhassani et al. (Abolhassani, Haeri & Hariri 2006) introduces a new method for compound measure creation without any need to the mapping extraction phase. 2.2 Mapping Extraction Previous works, do not especially cover the prob- lem of alignment extraction. A method for ontol- ogy alignment extraction is proposed by (Dieng & Hug 1998) which examines linguistical features to compare two ontologies on the basis of a IS-A rela- tionship. In (Melnik, Garcia-Molina & Rahm 2002), to extract a reasonable extraction, Stable Marriage (Gibbons 1985) problem is discussed. There are some other approaches, e.g. a machine learning approach to the problem is discussed in (Doan, Madhavan, Domingos & Halevy 2003), and (Mitra et al. 2003) describe a probabilistic based model. Staab et al. (Staab & Mdche 2002) have focused on structural and taxonomic comparison of two trees 1wordnet.princeton.edu to extract an alignment, in which dissimilarity of each two concepts is calculated based on their super- classes and subclasses. Stumme et al. (Stumme & Mdche 2001) uses shared instances of two ontologies that are to be mapped, however this work ignores the properties of classes. Zhdanova et al. (Zhdanova & Shvaiko 2006) expand the notion of ontology matching to a community- driven approach to enable web communities to es- tablish and reuse ontology mappings to achieve an adequate and timely domain representation. (Johnson, Cohen, Baumgartner, Lu, Bada, Kester, Kim & Hunter 2006) models inter-ontology relation- ship detection as an information retrieval task, where relationship is defined as any direct or indirect asso- ciation between two ontological concepts. Wand et al. (Wang & Gasser 2002) presents a specific formalization and algorithm for local interpretation of shared representations to build global semantic coher- ence for the distributed actions of individual agents, known as ”Mutual Online Ontology Alignment”. LOM, as described in (Li 2004), is a semi-automatic lexicon-based ontology-mapping tool that supports a human mapping engineer with a first-cut compari- son of ontological terms between the ontologies to be mapped, based on their lexical similarity and simple heuristic methods. 3 PRELIMINARIES AND NOTATIONS In this section, we define some necessary mathemati- cal concepts which are used throughout this paper. 3.1 Basic Definitions A graph Gi, by definition, consists of two sets: V (Gi), E(Gi), where V (Gi) is the set of vertices, and E(Gi) is the set of edges. The size of a graph is |V (Gi)|, which is denoted by |G|. Let us assume that labels assigned to nodes are chosen from a fi- nite alphabet Σ. Let λ /∈ Σ be a null character, and Σλ = Σ ∪ λ. 3.2 Metric Space According to (Rudin 1976), a set of points X along with a function, is said to be a Metric Space if the function associates a real num- ber with any pair of points p, q, denoted by d(p, q), and called the distance p, q, such that: ∀x, y ∈ O, δ(x, y) ≥ 0 (positiveness) ∀x ∈ O, ∀y, z ∈ O, δ(x, x) ≥ δ(y, z) (maximality) ∀x, y ∈ O, δ(x, y) = δ(y, x) (symmetry) Any function δ, satisfying the above conditions is said to be a distance function or a metric. In fact, the distance of two concepts belonging to two different ontologies is described as the distance of their labels in a metric space, and usually this metric distance is described by a distance function described above. 3.3 Typed Graph An ontology Oi, in this paper, is considered as a typed graph Gi. A typed graph, as defined in (Haeri, Hariri & Abolhassani 2006), is denoted by Gi(V, E, T ), for which E is of type: E : V ×V → T . Members of T are all from Σλ. In such a graph an edge e of type t be- tween vertices vij and vik is denoted by: e(vij , vik ) : t. A homeomorphism from a typed graph G(V, E, T ) to another typed graph G′(V ′, E′, T ) is a one-to-one correspondence between V and V ′. In this paper, each ontology Oi is modeled using a typed graph Gi where concepts of Oi are nodes of Gi, and the rela- tions/properties of Oi are typed edges of the graph. Figure 2: A sample alignment of two graphs G, G′ 3.4 Edge Preservation We will call an edge e(v1j , v1k ) : t ∈ E(Gi1 ) preserved under the mapping M, if and only if there is an edge e(M(v1j ), M(v1k )) : t ∈ E(Gi2 ). In other words, an edge e is preserved under mapping M if and only if ∃e′ ∈ E(Gi2 ) : e′ = (M(v1j ), M(v1k )), M(e) = e′, and is not preserved otherwise. The preservation of edges between corresponding nodes is the key point to find an ideal mapping. In fact in an ideal alignment most of the edges of one ontology are preserved in the second one. 3.5 Ontology Alignment (OA) In this section we will discuss our own understanding of a one to one alignment of two ontologies. A one to one alignment of two ontologies Oi1 , Oi2 is denoted by M : Oi1 → Oi2 and is a one to one correspondence between nodes of the two graphs of Oi1 , Oi2 : (Gi1 , Gi2 ). (Ehrig & Sure 2004) defines the mapping function in the following way: • M : Oi1 → Oi2 • ∀v ∈ Gi1 : M(v) = v′ if v′ ∈ Gi2 and δ(v, v′) < t, for t being a threshold v′ is the corresponding node of v under the mapping M. Figure 2 illustrates a sample alignment for two exam- ple ontologies G, G′. 3.6 Our Formulation of OA We denote the correspondence from ontology Oi to Oj while described by the concept of typed graphs, i.e. Gi to Gj by M : Gi → Gj . It is defined as follows: 1. ∀vi ∈ Gi, vi corresponds to only one vertex vj in Gj (denoted by M(vi) = vj ), or does not corre- spond to any vertex in Gj (denoted by M(vi) = null). And if v1, v2 ∈ Vi, v1 6= v2, M(v1) 6= null, M(v2) 6= null then M(v1) 6= M(v2) 2. The correspondence of edges, is determined by the correspondence of nodes: ∀ei = (vi1 , vi2 ) : t ∈ E(Gi) : if M(vi1 ) = vj1 6= null ,M(vi2 ) = vj2 6= null, and ej = (vj1 , vj2 ) : t ∈ E(Gj ) then ei corresponds to ej , M(e(vi1 , vi2 )) = e(vj1 , vj2 ), else ei does not correspond to any edge in Gj (denoted by (M(e(vi1 , vi2 )) = null) 3. Let M be a correspondence from Gi to Gj . We call M a map from Gi to Gj if ∀vi ∈ Vi, M(vi) 6= null, i.e.M(vi) ∈ Vj 4. Each map, M from Gi to Gj has a weight and this weight is defined by the coincidence-based technique described in section 4. The problem which is addressed in this paper is for- mulated as follows: INPUT: Two ontologies together with a matrix, rows and columns of which, stands for concepts of ontologies, and each cell shows the distance of the two concepts as given by a distance measure. OUTPUT: A proper alignment. In what follows, we explain a technique to score pos- sible mappings of ontologies, so called coincidence– based mapping, in the next section and then use this weight function to extract a proper alignment with evolutionary approaches in section 5. 4 COINCIDENCE BASED WEIGHTING In this section we introduce and discuss a new weight- ing model for an alignment, with which we will later design our genetic algorithm. The coincidence based alignment weight function is sufficiently discussed in (Haeri et al. 2006), and here, we will have an overview of it. Before talking about the weight itself, lets take some time, and dis- cuss the matter. There is a set of properties that we believe any mapping should convince. Considering a mapping M, between two ontologies with graphs Gi1 , Gi2 , and two nodes v1j , v1k ∈ V (Gi1 ) and their matches M(v1j ), M(v1k ), the weighting system should result a high weight if v1j is close to M(v1j ) and also v1k is close to M(v1k ) and when e = (v1j , v1k ) : t ∈ E(Gi1 ) is preserved under M . This case is considered to be the most desired one and should be given the highest value. An alternative is when the edge is not preserved. Here, a negligible negative point should be given. The reason for negative point is the fact that, the edge is not preserved and the structural matching of the graphs is interrupted. In this case the nodes are very close but the edge is missing. The farther any of the nodes is, from its match, the lower should be the positive value of the mapping. If the edge is preserved, we give this mapping a low positive value. But when the edge is not preserved, in fact it is an undesired mapping. So we give it a negative point. In this case not only the nodes are far from their matches, but also the edge is not preserved. According to the above considerations there should be six different categories which are graphi- cally shown in the Figure 3 (In the following expla- nations we assume that G, G′ are graphs of two on- tologies O, O′ to be aligned. a, b are concepts from G, and a′, b′ from G′): • Category I. a and a′ are too close2, and b, b′ are close as well, and the edge between a, b is preserved under the matching process. This cat- egory is of much importance. This is because actually the two edges coincide too much. To clarify the point suppose the case when a and b are “means of communication” and “mail” re- spectively, and a′, b′ are “communication” and “email”. The fact that there is an edge, (i.e. 2in terms of a distance function described before Figure 3: Different Properties of mappings in a metric space. Dotted edges with type t′ show that there might be an edge of type t′ or there might be no edge present. Dashed arrowed lines shows the mapping elements. (I) shows the first category where vertices (concepts) are close and the edge is preserved. (II) is the dual of category I, and different in that the edge is not preserved. (III) shows the case when the edge is preserved but only one of the endpoints of the two edges are close to each other. (IV) is dual case of category III, and different in that the edge is not preserved (V) The edge is preserved, but none of the endpoints are close to what they have been matched (VI) is the dual case of category V, but different in that the edge is not preserved. rdf s : type) between both a, b and a′, b′ shows that the two edges coincide too much, and that the two ontolgoies are describing the same world. • Category II. In this category, the two peers of an edge are close to their matches, that means, a is close to a′ and b is close to b′ as well. The only difference between this category and the previous one is that here, the edge is not preserved. As de- scribed in (Haeri et al. 2006) consider, e.g. when O describes the Glazing Technology, whilst O′ is the ontology of simple glasses manufacturing stu- dio. Let a, b be “Glass” and “Frame”, and a′ , b′ the same respectively. Although δ(a, a′), δ(b, b′) are both small, the non-preservation of the edge (a, b) is a negative point. The fact that the ver- tices coincide, make us not to penalize this cat- egory much, because at least concepts are close to their matches and vertices coincide. • Category III. In this category the edge is pre- served but only one of the a or b is close to its match. This is good but not as much as the previous category. Consider two ontologies, de- scribing two different worlds. Suppose O is de- scribing a “High Tech manufacturing” while O′ is describing a “Supermarket”. Let a, b be “Laptop Computer” and “Product” respectively and a′, b′ be “Laptops” and “On Sale Item” respectively. The two ontologies are describing two totally dif- ferent domains, whilst a and a′ are close. So it seems as if such ontologies are getting close “ from the side of a”. We would like to give such category a large weight, yet smaller than that of the category I. • Category IV. As category II can be consid- ered to be the dual of category I, this category can be the dual of category III. The reason ori- gins in that only one peer of the edge in O is too close to what it is matched to, yet the edge is not preserved under the matching. As it is clear in Figure 3 IV the edge between a, b is not pre- served, and b is far from b′. The only positive point of such a matching is the fact that a and a′ are close. As an example, just to make things clearer, consider O to be describing a hotel’s ser- vices, where a is “egg” and b is “omelette” and O′ is describing a “Supermarket” and a′, b′ are “egg” and “shampoo” respectively. This match- ing, which maps a to a′ and b to b′ is not de- sired and is most probably getting into mistake. However, this mistake should not be penalized as much as the mistake in category VI. • Category V. The last two categories describe the situation where none of the peers of an edge is close to what it is matched to. Even though the edge might be preserved (as in category V) the two edges do not coincide at either end points. In other words, in these categories, both a, a′ and b, b′ are far from one another, and the difference is in the preservation of edges. The fact that the vertices are not close to their matches, is quite enough to make us indifferent about the edge preservation. Because even if the edge is preserved, the two edges are not that much co- incident. Both cases are not desired and should obtain low points. A clear example of category V, would be when a is “Plant”, b is “Water”, a′ is “Car”, and b′ is “Fuel”. No need to discuss that this mapping is not a desired one. • Category VI. This case is even worse in cat- egory VI than that of category V, where the edge is not even preserved. An example would be when a, b are “mammal” and “elephant” in O which is describing a “zoo” and a′, b′ are “glasses” and “frame” respectively in O′ which is incidently describing a “glasses manufactur- ing company”. There is neither a similarity be- tween endpoints of the two edges, nor is there any preservation. The vertices in this category are mapped to what have no similarities, seman- tically. According to the above cases, the following weigh function is suggested: w(M) = w0(M) − wl(M) − wr(M) w0(M) = ∑ (v1,v2):t∈E(G) , (M(v1),M(v2)):t∈E(G′) f (v1)+f (v2) wl(M) = ∑ (v1,v2):t∈E(G) , (M(v1),M(v2)):t /∈E(G′) g(v1)+g(v2) wr(M) = ∑ (v1,v2):t /∈E(G) , (M(v1),M(v2)):t∈E(G′) g(v1)+g(v2) The functions f and g, referred to as Normaliza- tion Functions (Haeri et al. 2006) , are in the form: f : R → R+ g : R → R+ f, g are related to the distance function. In fact, f should be a positive decreasing function, so that if δ(v, M(v)) grows, it decreases to reduce the positive point. And on the other hand g should be a posi- tive increasing function to grow with the growth of δ(v, M(v)) to increase the negative point for that match. In any other cases, in one of the above six categories w will misbehave. Normalization functions are defined by tuning the system. This will be de- scribed again later. According to (Haeri et al. 2006): “no [much] work is so far done on the problem of Ontology Alignment or Ontology Matching in which graph theoretic back- bone of problem is scrutinized.”. With the use of graph theory and such a modeling we believe that there is a vast area for new work on the problem of on- tology aligning. The coincidence measure explained in this section is a step forward in this direction. We believe it can be used in various ways for the Map- ping Extraction problem. The sole usage of it in this paper is introduced in the next section. 5 GENETIC ALGORITHM (GA) This section describes the developed genetic algo- rithm. Matching two general graphs in polynomial running time algorithms is impossible, because the problem in its general case is MAX SNP-Hard (Arora, Lun, Mot- wani, Sudan & Szegedy 1992). So a random search algorithm could be a good idea when designed care- fully. This leads us to the idea of using genetic algo- rithms. Genetic Algorithmic solutions are evolutionary algo- rithms, which will approach the final state by grad- ually improving the solution. Any problem which is solved by a GA, should first be coded in a way that it can be easily handled in different parts of the GA. Each coding of a solution will form an individual. Some individuals which are stored and processed in each iteration form the population. the population improves in each step with the use of crossovers and mutations and the best individual will finally be re- ported as the answer of the GA. In the following sub- sections we explain about different parts of our GA solution. 5.1 Coding a Mapping To code a mapping we use hashmaps (Cormen, Leiser- son, Rivest & Stein 2001) in which keys are concepts of one ontology and entries are concepts of another one. This data structure help us easily manipulate a one-to-one alignment, with a search of concepts in O(1). Entry for each key is actually the corresponding node of that key in the mapping. That is, if vi ∈ O is mapped to v′i ∈ O′ then in the hashmap h we’ll have the vi as the key, and h(vi) = v′i. 5.1.1 Pairs According to the coincidence-based weighting, we de- fine a Pair, as two concepts from one ontology, be- tween which there is a relation (So there is an edge in the graph of that ontology between them). Fig- ure 2 shows the alignment of two ontologies, in which (v1, v2), (v1, v3), (v3, v4), (v2, v4) are pairs of G. A pair also has a weight according to the alignment it involves in. Clearly speaking, a pair is a function of the form: P : V × V × T −→ R where V is the set of vertices in ontology graph G and T is a set of labels in Σλ. So an ontology in a matching has a limited number of pairs. The weight of a pair depends on the alignment in which the ontology is involved. Let Gi1 , Gi2 be two graphs of two aligned ontologies, and v1j , v1k ∈ V (Gi1 ). Also assume an edge between v1j , v1k to be of type t, e1jk = e(v1j , v1k ) : t. P (v1j , v1k , t) in the alignment of two ontologies is given by: P (v1j , v1k , t) = { f (v1j ) + f (v1k ) e1jk preserved −(g(v1j ) + g(v1k )) e1jk not preserved −∞ if e1jk /∈ E(Gi1 ) For a couple of concepts which do not form a pair the value of P function is set to be −∞. Definition of pairs is useful in crossover function which will try to improve the structural matching. In the alignment of two ontologies, Oi1 (Gi1 ) , Oi2 (Gi2 ), say M : Oi1 → Oi2 , we also define the weight of a single concept from one ontology, W (v1j ) where v1j ∈ V (Gi1 ), as follows: if v1j ∈ V (Gi1 ), M(v1j ) ∈ V (Gi2 ) W (v1j ) = ∑ ∀v∈Gi1 :e(v1j ,v):t∈E(Gi1 ) P (v1j , v, t) 5.1.2 An Example To make things clear about the definition of pair and its corresponding weights described above, we give an example on how to compute these weights. In the Figure 2 we have: P (v1, v2, t2) = f (v1) + f (v2) P (v1, v3, t1) = f (v1) + f (v3) P (v3, v4, t3) = −g(v3) − g(v4) P (v2, v4, t4) = −g(v2) − g(v4) P (v1, v4, ti) = P (v2, v3, ti) = −∞ W (v1) = (f (v1) + f (v2)) + (f (v1) + f (v3)) W (v2) = (f (v2) + f (v1)) − (g(v2) + g(v4)) W (v3) = (f (v3) + f (v1)) − (g(v3) + g(v4)) W (v4) = −(g(v4) + g(v3)) − (g(v4) + g(v2)) Now, with these definitions, it is the time, to de- scribe the steps of our genetic algorithm. 5.2 Initialization As of any other Genetic Algorithms, a primary pop- ulation is needed. A population is made up of some individuals each of which is a solution to the problem (a mapping in this problem). The start population, is initialized randomly, with an initial size of 1000 in- dividuals. The ideal mapping can be reached more quickly if the initial individuals, are made on the ba- sis of the labels of concepts, that is if v1j in Gi1 and v2j in Gi2 have same labels, then let v1j corresponds to v2j in the initial mapping. 5.3 Selection In each iteration, we sort the 1000 individuals accord- ing to their fitness described in section 4 (coincidence based weight function), and we select the 500 best in- dividuals as parents of the next step. From these 500 individuals, with the use of crossover and mutation functions (as we will see later), 1000 new individuals are created. These 1000 individuals are sent to the next iteration as parents. Two crossover functions are designed, one based on pairs, and the other based on solitary vertices. In the following two subsections we explain each one in detail. 5.4 Crossover I. In the first crossover, Crossover I, the pairs are in primary concern and best pairs from the parents are preserved in offsprings. For every single node in the first ontology graph, the pairs of that node are examined in the two mappings (i.e. two parents), and the best pair, which has the highest value of weight is copied in the offspring. If a matched node in ontology graph G′ is already as- signed to some other node of G, the assignment is done to a random unassigned node. In figure 4 two mappings as parents and the result of crossover I are shown. The pair (vi, vj , t) in parent 1, has greater weight than in parent 2. It is resulted from computing P (vi, vj , t) in both alignments: In parent 1 we have P (vi, vj , t) = f (vi) + f (vj ) because the edge e(vi, vj ) : t is preserved under M, but in parent 2 it is P (vi, vj , t) = −(g(vi) + g(vj )) because the edge e(vi, vj ) : t is not preserved. So the pair in parent 1 is chosen for offspring. Since f, g are positive functions, in the offspring we have M(vi) = v′i, M(vj ) = v′j . It is clear that in this crossover the pairs in off- springs are not worse than those of parents. 5.5 Crossover II. In this crossover function, single nodes are compared according to their weights. As we described before, the weight of a single node in a mapping is the sum of weights of pairs in which, that node is included. Consider two parents in two ontology graphs Gi1 , Gi2 . Figure 4: crossover I. (a) part of parent 1 mapping. (b) part of parent 2 mapping (c) part of offspring To make an offspring from two parents, for every node in Gi1 , say v1j , the mapping with larger W (v1j ) is copied to the offspring. if M (v1j ) in Gi2 is already assigned with some other node of Gi1 , then v1j is put in a reserved list. At the end of the complete iter- ation of nodes in Gi1 , the nodes in the reserved list are randomly mapped to the unassigned nodes of Gi2 . The random assignment is not done in the middle of an iteration to prevent nodes of Gi2 to be assigned to some random nodes that can be assigned to better nodes later in the iteration. So this random assign- ment is postponed until all nodes of Gi1 are examined to map to nodes in Gi2 . As an example suppose v1m ∈ V (Gi1 ) should be mapped to v2m ∈ V (Gi2 ) and v2m is already mapped by some node from Gi1 , so if at that time we assign v1m to some random node like v2n ∈ V (Gi2 ), it will prevent a possible good mapping of v1n to v2n later in the iteration. So this random assignment is delayed until no more assignment is possible. In Figure 5 two mappings between two ontologies O, O′ are shown, and we want to decide the match node for vi ∈ V (G) in the offspring. In parent 1 we have: W (vi) = P (vi, vj , t2)+P (vi, vk, t1) = (f (vi)+f (vj ))+ (f (vi) + f (vk)) and in parent 2 we have: W (vi) = P (vi, vj , t2) + P (vi, vk, t1) = −(g(vi) + g(vj )) + (f (vi) + f (vk)) Again it should be noted that f and g are positive functions so the value of W (vi) in parent 1 (Figure 5 (a)) is greater than that of in parent 2 (Figure 5 (b)). So as it is shown in part c of the Figure, the corresponding node of vi ∈ V (G) in offspring is chosen by the mapping from parent 1, and therefore is v′i ∈ V (G′). This kind of crossover seems reasonable because the mapping of a single node in the offspring is not worst than that of the two parents. So by this as- Figure 5: crossover . (a) part of parent 1 mapping. (b) part of parent 2 mapping (c) part of offspring . sumption, little by little, mappings of nodes will con- verge to ideal ones. 5.6 Mutation A proportion of the population are mutated with some probability, different in various iterations. In mutation of a mapping of two ontologies with graphs Gi1 , Gi2 , two random nodes from Gi1 are chosen, and their matches in Gi2 are substituted with each other. Let v1j , v1k ∈ V (Gi1 ) are chosen randomly, also let M(v1j ) = v2j ∈ V (Gi2 ), M(v1k ) = v2k ∈ V (Gi2 ). In the mutation process we just substitute the match nodes of the selected ones. So the new mapping will be M(v1j ) = v2k ∈ V (Gi2 ), M(v1k ) = v2j ∈ V (Gi2 ). 5.7 Continuation One important issue with any Genetic Algorithm, is how to get use of the crossover and mutation func- tions, and how to create the new population, based on the old one. In our solution, the two previously explained crossover functions are invoked on the ith and i + 1th parents to create two offsprings. The last parent is mixed with the first one to produce last two offsprings. Our population as described before, contains 500 in- dividuals. These individuals are sorted decreasingly, and the sorted array forms the parents of the current step. In each iteration, these 500 parents, with the help of a series of crossovers and mutations (as ex- plained before) produce 1000 new individuals. The resultant array of individuals is then sorted and the best 500 of them are selected as parents of the next step. Figure 6 shows this process. Figure 6: Population generation in each step of GA 5.8 End Condition Basically there is no end for the execution of any evo- lutionary algorithm and specifically a genetic algo- rithm, where a user must pause the run process if she thinks the results are reasonable. However, to end the iteration of our GA, we used a threshold for conver- gence. The sequential GA is continued until the best mapping among all individuals in the population does not improve for more than 15 steps. Such mapping is reported as the answer for the problem of a proper alignment. The alignment process of two ontologies is then finalized. 6 EVALUATION To evaluate our Genetic Algorithm, we designed three kinds of experiments. In the first experiment, we tested the Genetic Algorithm with diverse mutation probability. In the second experiment, we tried to align two identical ontologies (actually we aligned one ontology with itself). This experiment helped us ex- amine the efficiency and accuracy of the algorithm, when two ontologies are more similar to each other. To verify our contribution, we came with a third ex- periment, in which we used a naive local search align- ment method. We already discussed about similarity measures in section 2. It is actually an important issue to pay attention in the ontology alignment and extraction process. There are some similarity measures pro- posed by some researchers. For example, (Euzenat & Altchev 2004) developed a similarity metric between concepts in OWL ontologies, which is a weighted com- bination of similarities of various features in OWL concept definitions: labels, domains, ranges of prop- erties, restrictions on properties, types, IS–A rela- tionships. (Mitra, Wiederhold & Decker n.d.) and (Mitra, Wiederhold & Decker 2001) use the combi- nation of interactive specifications of mappings and heuristics to propose proper mappings. In this paper we used the Levenshtein (Levenshtein 1966) string– based similarity measure for the concepts, where the dissimilarity of any two concepts (from two ontolo- gies) is calculated by the Levenshtein distance. This measure is suitable for our experiments since most of the heterogeneity in the Ontologies in our test collec- tion comes from lexical difference. However, it should be noted that our coincidence-based weighting and hence our GA solution is independent of any similar- ity measure. To apply it to any other alignment, one can select another suitable measure for that domain and find similarity values to be used in our algorithm3 6.1 Limitations The coincidence-based weighting has an innovative idea behind, however there are essential practical lim- itations to apply this method. The most important limitation is the available ontologies and test collec- tions. Most of them do not have a large taxonomic structure and so the method does not have enough merit for them. However, in our search for a suit- able test collection we found “Tourism” ontologies (Tourism ontology FOAM n.d.) a good one with ap- proximately 340 classes and concepts. 6.2 Various Experiments Characteristics For the tourism ontologies, an ideal alignment is in- cluded in the test collection. We use such information to calculate the precision (Baeza-Yates & Ribeiro- Neto 1999) for each experiment. Consider M to be a mapping: MO → O′. To find the accuracy of the method and calculate the precision, we need a model alignment M′ which is formed and extracted by some expert. Let G, G′ be the graphs for O, O′ respectively. Also suppose, S is the set of vertices, vi, in V (G) where M(vi) = v′i = M′(vi). In other words: ∀vi ∈ V (G) : vi ∈ S ⇔ M(vi) = M′(vi) Now, the precision of the alignment M is given by: P recision(M) = |S||V (G)| • Experiment 1 As discussed previously, in this experiment we aligned “TourismA” with “TourismB”. This is the main experiment to check the efficiency of our coincidence-based genetic algorithm. In this experiment, from each two parents, we made two offsprings one with the use of crossover I function and the other with crossover II. From the four different individuals (parents and the offsprings) we chose two best of them, to introduce as children of this amalgamation. Normalization Functions are as follows: f (v) = 1 eδ(v,M (v)) g(v) = 1 emax(5,15−δ(v,M (v))) These functions actually satisfy the characteris- tics expected from f, g (explained in Sec. 4). f is a decreasing function and decreases with the growth of δ and g is increasing. Exponential functions were chosen for f, g so that f, g would have close and comparable values. In fact, these functions match the discussions on positive and negative points for different categories of a coin- cidence based weight. – Experiment 1.1 After the 1000 individuals are created, we mutate the lower half of them (with the mutation function described before) with a probability of 0.7. 3as cited in related works section, there are some good works for selection of an appropriate measure for a domain which can be applied here. Figure 7: Precision result of experiments – Experiment 1.2 After the 1000 individuals is created, we mutate the lower half of the them (with the mutation function described before) with a probability of 0.3. – Experiment 1.3 Mutation was done on each one of the 1000 individual in the all of the 1000 children with the probability of 0.5. • Experiment 2 In this experiment we are aligning “TourismA” with itself. This actually is a verification that shows how efficient the genetic algorithm will work, if two ontologies are more similar and actually more coincide. The generation summary is similar to the previous experiment, and mutation was done on the lower half of the individuals, with the probability of 0.5. Normalization Functions are similar to the previous experiment, f (v) = 1 eδ(v,M (v)) g(v) = 1 emax(5,15−δ(v,M (v))) • Experiment 3 This experiment actually provides a baseline comparison of the GA method with a naive lo- cal search method. In this part, we implemented a naive hill-climbing local search method. For the start point, we made an initial alignment. In this initial alignment, all concepts in “TourismA” is matched with concepts in “TourismB”. For a node vj in TourismA if there were a node v′j with label label(vj ) in TourismB, we matched vj with v′j . Otherwise we mapped vj to a random node of TourismB. After that, in each iteration, the best single change (mutation) was preformed to improve the weight value of alignment. We iterated the method until almost 1000 steps, where, the re- sults did not improve for more than 15 steps. 6.3 Results Figure 7 shows the result of the above experiments according to the precision measure. As it is shown, with identical graphs, Genetic algorithm finds the best mapping and precision is 1. With other exper- iments, however, the result is a little inaccurate in Figure 8: Convergence of results in GA comparison with the ideal alignment and precision is approximately 0.8. In our experiments, the distance threshold, which we talked about in section 3.5, is set to be 4. We chose this number by experience, however there could be other solutions to determine this number, like ma- chine learning techniques, etc. It is also possible to add one level of iteration to our genetic algorithm to test different values for distance threshold and select the threshold that results the best total weight for mapping. Since the labels for concepts in a typical Ontology normally does not exceed the length of an English (or other natural language) word the upper limit of the distance threshold can be said to be less than 20. So by execution of our genetic algorithm with different distance threshold vales (e.g. 1,2 to 20) we may find the best threshold value for an alignment task. We also did an investigation on iteration number and convergence of the result in this genetic algorithm for Experiment 1. The results are shown in figure 8. Figure 8 shows the weight of the best alignment, i.e. the best individual in the population, reached in each iteration of the running process in the Ge- netic Algorithm. The figure shows the best align- ment weight for all three experiments, 1.1 in red, 1.2 in blue, and 1.3 in green. As it is clear, in all experi- ments the Genetic Algorithm converges before almost 35 iterations. The precision average for convergence is 79%. This means that the GA will not find a better solution, if it runs more than this number of itera- tions. Usually at such circumstances all individuals in the population, converge to a single individual and equal to each other, so that the combination of them in the form of crossovers will not improve the solution and will result in the same individuals. Besides, the mutation in most cases will result in a worse individ- ual than a better one. The reason is quite clear. The optimal states (individuals) are much less than worse states, and the mutation is absolutely random. Muta- tion is mostly useful when the population is trapping in local maxima, and the mutation will move it to another point in the Genetic Algorithm. It should be noted that in this area the main cri- teria of the goodness of an approach is the precision of the extracted mapping and reaching to a solution with the optimal time complexity is beyond the state of the researches. 7 CONCLUSION AND FUTURE WORK In this paper, we discussed our method for Map- ping Extraction in an Ontology Alignment frame- work. First, we introduced our modeling of an on- tology with the concept of typed graphs. Then our coincidence-based measure for weighting of different candidate alignments were discussed. We would like to state that this experience gives us the impression that our solution which gets help from graph the- ory (and which was contrived for ontology alignment only) can be even further extended to a wider range of problems. Graph matching and similarity measures of graphs have many applications in machine vision, pattern recognition, bio–informatics, and etc. We defined two cross-over functions based on coincidence-based measure. They are crafted in a way that ensure the new generations are not worse than previous ones. In our experiments our GA solu- tion converges very rapidly, for example after approx- imately 40 iterations, which, in the order of magni- tudes, is considerably better than a naive ”candidate mapping generation and test” approach. This number can even be reduced more by choosing a biased initial population, where labels can be involved to choose better initial mappings. There are also some weaknesses with Genetic Algo- rithms. One of them is the dependency of results to initial population. The more significant weakness is when the two ontologies are sparse graphs, or even worse is when there are like forests. In these cases the domain for crossover is not a soft one, and small changes in an individual in crossover or mutation might take it to a very far point. The reason for this anomaly is that in sparse graphs or jungles, the number of disconnected vertices is more, and there- fore a small change in the alignment will map one node to another vertex which, more probably, will not form a coincident mapping. Roughly speaking, the more connected and taxonomic the two ontologies are, the better the results of the Genetic Algorithmic coincidence-based extraction will be. In continuation of our research, work is now be- ing done on tree-like ontologies (which seem being a most common form). Once we can align tree ontolo- gies, we can model ontologies as trees and align the resulting trees. We are also interested in extending our theory and mechanisms for matching ontologies based on their various graphical shapes, properties of subgraphs, etc. 8 Acknowledgments This research was in part supported by a grant from IPM (No. CS1385-4-01). References Abolhassani, H., Haeri, S. H. & Hariri, B. (2006), ‘On ontology alignment experiments’, Webology 3(3). Arora, A., Lun, C., Motwani, R., Sudan, M. & Szegedy, M. (1992), Proof verification and hard- ness of approximation problems, in ‘33rd IEEE symposium on the Foundations of Computer Sci- ence (FOCS)’, pp. 14–23. Baeza-Yates, R. & Ribeiro-Neto, B. (1999), Modern Information Retrieval, Addison Wesley. Bisson, G. (1992), Learning in fol with similarity mea- sure, in ‘Proceedings of the 10th American Asso- ciation for Artificial Intelligence conference, San- Jose (CA US)’, pp. 82–87. Bouquet, P., Euzenat, J., Franconi, E., Serafini, L., Stamou, G. & Tessaris, S. (2004a), Specification of a common framework for characterizing align- ment, deliverable 2.2.1, Knowledge web NoE. Bouquet, P., Euzenat, J., Franconi, E., Serafini, L., Stamou, G. & Tessaris, S. (2004b), ‘Specification of common network framework for characterizing alignments’, Knowledge web NoE 2.2.1. Cormen, T. H., Leiserson, C. E., Rivest, R. L. & Stein, C. (2001), Introduction to Algorithms, sec- ond edition, MIT Press. Dieng, R. & Hug, S. (1998), Comparison of p̈ersonal ontologiesr̈epresented through concep- tual graphs, in ‘13th ECAI, Brighton (UK)’, pp. 341–345. Doan, A., Domingos, P. & Halevy, A. (2003), ‘Learn- ing to match the schemas of data sources: A multistrategy approach’, Machine Learning 50(3), 279–301. Doan, A., Madhavan, J., Domingos, P. & Halevy, A. (2003), ‘Ontology matching: A machine learning approach’, Handbook on Ontologies in Informa- tion Systems. Springer-Velag, 2003. . Ehrig, M. & Staab, S. (2003), Qom quick ontology mapping, in ‘Proc. ISWC-2003.’. Ehrig, M. & Sure, Y. (2004), Ontology mapping — an integrated approach, in ‘1st European Semantic Web Symposium (ESWS)’, pp. 76–91. Euzenat, J. & Altchev, P. (2004), Similarity–based ontology alignment in owl–lite, in ‘The 16th Eu- ropean Conference on Artificial Intelligence’. Euzenat, J., Barrasa, J., Bouquet, P. & Bo, J. (2004), ‘State of the art on ontology alignment’, Knowl- edge Web, Statistical Research Division . Euzenat, J. & Valtchev, P. (2004), An integrative proximity measure for ontology alignment, in ‘Proc. 3rd International Semantic Web Confer- ence (ISWC2004) .’. Gibbons, A. (1985), Algorithmic Graph Theory., Cambridge University Press,. Haeri, H., Hariri, B. & Abolhassani, H. (2006), Coincidence-based refinement of ontology align- ment, in ‘3rd Intl. Conference on Soft Computing and Intelligent Systems’. Hopcroft, J. & Karp, R. (1973), ‘An n5/2 algorithm for maximum matchings in bipartite graphs’, SIAM Journal on Computing 2(4), 225–231. Johnson, H. L., Cohen, K. B., Baumgartner, W. A., Lu, Z., Bada, M., Kester, T., Kim, H. & Hunter, L. (2006), ‘Evaluation of lexical methods for de- tecting relationships between concepts from mul- tiple ontologies’, Pac Symp Biocomput pp. 28– 39. Kalfoglou, Y. & Schorlemmer, M. (2003), ‘Ontology mapping: the state of the art’, The Knowledge Engineering Review 18, 1–31. Levenshtein, V. I. (1966), ‘Binary codes capable of correcting deletions, insertions and reversals.’, Sov. Phys. Dokl. 6, 707–710. Li, J. (2004), Lom: A lexicon-based ontology mapping tool, in ‘Proceeding of the Perfor- mance Metrics for Intelligent Systems (Per- MIS04).Information Interpretation and Integra- tion Conference (I3CON), Gaithersburg, MD.’. M. Ehrig, S. Staab, Y. S. (2005), Bootstrapping on- tology alignment methods with apfel., in ‘Pro- ceedings of the 4th International Semantic Web Conference (ISWC-2005), ser. Lecture Notes in Computer Science’, pp. 186–200. Maedche, A. & Zacharias, V. (2002), Clustering on- tologybased metadata in the semantic web., in ‘Proceedings of the 13th ECML and 6th PKDD’. Melnik, S., Garcia-Molina, H. & Rahm, E. (2002), Similarity flooding: a versatile graph matching algorithm, in ‘Proc. 18th International Confer- ence on Data Engineering (ICDE), San Jose (CA US)’, pp. 117–128. Mitra, P., Noy, N. F. & Jaiswal, A. R. (2003), Omen: A probabilistic ontology mapping tool, in ‘4th international semantic web conference (ISWC 2005)’, Vol. 3729, pp. 537–547. Mitra, P., Wiederhold, G. & Decker, S. (2001), A scal- able framework for interoperation ofinformation sources, in ‘In The 1st International Semantic Web Working Symposium (SWWS01), Stanford University, Stanford, CA,’. Mitra, P., Wiederhold, G. & Decker, S. (n.d.), ‘The prompt suite: Interactive tools for ontology merging and mapping.’, International Journal of Human-Computer Studies, volume = . OntoWeb. A survey on ontology tools. (2002), avail- able online : www.ontoweb.org/deliverable.htm. Papadimitriou, C. H. & Steiglitz, K. (1998), Combi- natorial Optimization Algorithms and Complex- ity, Prentice-Hall. Rahm, E. & Bernstein, P. (2001), ‘A survey of ap- proaches to automatic schema matching’, VLDB Journal 10(4), 334–350. Rudin, W. (1976), Principles of Mathematical Analy- sis, third edn, McGraw-Hill, New York. Schreiber, G., de Hoog, R., Akkermans, H., Anjewier- den, A., Shadbolt, N. & de Velde, W. V. (2000), Knowledge Engineering and Management, MIT Press. Staab, S. & Mdche, A. (2002), ‘Measuring similar- ity between ontologies’, Lecture notes in artifi- cial intelligence (2473), 251–263. Stumme, G. & Mdche, A. (2001), Fca-merge: bottom-up merging of ontologies, in ‘In Proc. 17th IJCAI, Seattle (WA US)’, pp. 225–230. Tourism ontology FOAM (n.d.). avail- able online: http://www.aifb.uni- karlsruhe.de/WBS/meh/foam/ontologies/. Valtchev, P. (1999), ‘Construction automatique de taxonomies pour laide la reprsentation de con- naissances par objets.’, Ph.D. Dissertation, Uni- versite Grenoble. . Wang, J. & Gasser, L. (2002), Mutual online ontology alignment, in ‘Proc. of the AAMAS 2002 Work- shop’. Zhdanova, A. V. & Shvaiko, P. (2006), Community- driven ontology matching., in ‘Proc. of ESWC’, pp. 34–49. Zhong, J., Zhu, H. & Y. Li, a. Y. Y. (1995), Using information content to evaluate semantic simi- larity in a taxonomy, in ‘Proceedings of the 14th International Joint Conference on Artificial In- telligence (IJCAI-95)’. 
work_upn2iwecqvhozhd2y7ajcu64a4	----	An expert system for the diagnosis and management of oral ulcers Available online at www.sciencedirect.com ScienceDirect Tanta Dental Journal 11 (2014) 42e46 www.elsevier.com/locate/tdj An expert system for the diagnosis and management of oral ulcers Sh.A. Ali, H.I. Saudi* Periodontology, Oral Diagnosis and Radiology, Tanta University, Egypt Received 14 December 2013; revised 5 March 2014; accepted 8 March 2014 Available online 11 May 2014 Abstract The present research was conducted to introduce a Visual Basic expert system to help the postgraduate students in the diagnosis and treatment of the most common and rare oral ulcers. A total of sixty postgraduate students shared in the study. They received the system on a CD-ROM and asked to evaluate it using a printed questionnaire. The oral ulcerative conditions merged in the database included, aphthous ulcers, chancre, traumatic ulcers, histoplasmosis ulcers, acute herpetic ulcers, burns, herpangina, tuberculosis, syphilis, gonorrhea, acute necrotizing ulcerative gingivitis, herpes labialis, herpes zoster, erythema multiform, Steven Johnson's syndrome. The conditions also included, allergy, Behcet's syndrome, neutropenia, necrotizing sialometaplasia, pemphigus, ul- cerative lichen planus, vitamin deficiency, anemia, lupus erythematosis, epidermolysis bullosa, pemphigoid, squaemous cell carcinoma, desquamative gingivitis and others. Results showed that the mean overall scoring quality of the program was 3.0 ± 0.7 with 75% success rate. It can be concluded that the introduced expert system was helpful to the postgraduate students in the diagnosis and treatment of the most common oral ulcers. © 2014, Production and Hosting by Elsevier B.V. on behalf of the Faculty of Dentistry, Tanta University. Keywords: Expert system; Diagnosis and treatment of oral ulcers Open access under CC BY-NC-ND license. 1. Introduction Oral ulcerations is a common mucosal disorder. It may be caused by physical or chemical trauma, viral, fungal or bacterial infections, allergy, malignancy or a manifestation of systemic diseases [1]. The process of oral ulceration causes a breach in the oral epithelium, * Corresponding author. þ20 1005218804 (mobile). E-mail address: Hussein_saudi@hotmail.com (H.I. Saudi). Peer review under the responsibility of the Faculty of Dentistry, Tanta University Production and hosting by Elsevier 1687-8574 © 2014, Production and Hosting by Elsevier B.V. on behalf of http://dx.doi.org/10.1016/j.tdj.2014.03.0Open access under CC BY-NC-ND license. which typically exposes nerve endings in the under- lying lamina propria, resulting in pain or soreness, especially when eating spicy foods or citrus fruits [2]. As the majority of the ulcers require treatment of the underlying cause, proper diagnosis will lead to suc- cessful treatment and prevention of the lesions [1]. The arrival of the twenty-first century has suddenly forced on dentistry a new paradigm regarding expected standards for state-of-the art patient care. Traditional methods and procedures that have served the profes- sion well are being questioned within the context of evidence based rationales and emerging information technologies [3]. In the area of dental informatics, the application of computer and information sciences to improve dental research, education, and practice has been particularly the Faculty of Dentistry, Tanta University. 05 mailto:Hussein_saudi@hotmail.com http://crossmark.crossref.org/dialog/?doi=10.1016/j.tdj.2014.03.005&domain=pdf www.sciencedirect.com/science/journal/16878574 http://dx.doi.org/10.1016/j.tdj.2014.03.005 http://dx.doi.org/10.1016/j.tdj.2014.03.005 http://www.elsevier.com/locate/tdj http://creativecommons.org/licenses/by-nc-nd/4.0/ http://creativecommons.org/licenses/by-nc-nd/4.0/ 43Sh.A. Ali, H.I. Saudi / Tanta Dental Journal 11 (2014) 42e46 noteworthy and the use of electronic teaching tools and learning environments as CD-ROM or web-based has increased dramatically [4,5]. Expert system is an intelligent computer program that uses knowledge and inference procedures to solve problems at the level of a human expert. It emulates the decision making ability of a human expert in a particular field. The knowledge in expert system may be either expertise or knowledge that is available from books, journals, conferences, magazines, knowledge persons and internet [6]. So, the aim of the present research is to introduce an expert system for the diagnosis and treatment of the most common oral ulcers based on the algorithm of Morris (2004) [7]. 2. Materials and methods The knowledge database was collected from Bur- ket's Oral Medicine (2008) [8], strategies in dental diagnosis and treatment planning of Morris (2004) [7] and many health care websites as Mayo Clinic. The data was classified, stratified, organized then trans- ferred to the expert system database. The expert system was developed using Visual Basic Ver. 5 language. The system contained many forms that welcome the user and start step by step to introduce the major descriptive categories of the oral ulcers whether occurring for the first time, recurrent or chronic in nature. Then starts to branch according to the user selections to reach the most descriptive con- dition that meets his or her patient. When the user selects the final condition, the corresponding form is located by the search engine and loaded to the screen displaying the diagnosis, the picture of the ulcer and the suggested treatment options (Figs. 1e3). To facilitate the user in decision making, a back- ward button was added in all selection forms to revise or change the selection. Also to make the decisions without confusions the introduced selection options were programmed so that the user was not allowed to select more than one option. The oral ulcerative conditions merged in the data- base included, aphthous ulcers, chancre, traumatic ul- cers, histoplasmosis ulcers, acute herpetic ulcers, burns, herpangina, tuberculosis, syphilis, gonorrhea, acute necrotizing ulcerative gingivitis, herpes labialis, herpes zoster, erythema multiform, Steven Johnson's syndrome. In addition to allergy, Behcet's syndrome, neutropenia, necrotizing sialometaplasia, pemphigus, ulcerative lichen planus, vitamin deficiency, anemia, lupus erythematosis, epidermolysis bullosa, pemphigoid, squaemous cell carcinoma, desquamative gingivitis and others. The compiled software was copied to CD-ROMs and distributed to 60 postgraduate dental students and house officers of Tanta University Faculty of Dentistry. The users were trained on using the software, asked to use it in their clinics and to fill a questionnaire about the size of the program, the loading time, the clarity of the introduced selection options, the number of correctly diagnosed ulcers and to score the quality of the overall program from 1(inadequate) to 5 (excellent) [9]. 3. Results The questionnaire papers were collected from the postgraduate students and data were tabulated and analyzed. So as to facilitate data scoring for statistical purposes the selection options were converted into numerical scores, the item “Not clear” ¼ 0, “Confused” ¼ 1 and “Clear” ¼ 2. Also the overall scoring quality of the program were coded as “Inadequate” ¼ 0, “Adequate” ¼ 1, “Good” ¼ 2, “Very good” ¼ 3 and “excellent” ¼ 4. The mean scores for the included ulcers were rep- resented in Table 1. The overall scoring quality of the program was 3.0 ± 0.7 with 75% success rate. 4. Discussion Early detection of diseases enables their prevention and appropriate treatment. Computer based methods are increasingly used to improve that medical service [10]. The present research was conducted to introduce a newly developed concise expert system as an aid in the proper diagnosis and treatment of the most common oral ulcers for the postgraduate students. The oral ul- cers were selected as they represent a common complaint of the outpatient clinics [11]. The percentage of diagnosis for the recurrent, first time ulcers and ulcers with well circumscribed borders was 90%, 85% and 85% respectively. Such results may be attributed to the increased prevalence of such ulcers. Also the percentage of the ulcers with ill defined bor- ders, ulcers without systemic manifestations and mul- tiple ulcers was 80%, 80% and 75% respectively. Results may be attributed to prevalence of such type of ulcers, the introduced data from the expert system and the back knowledge of the postgraduate students. The percentage of diagnosis for the chronic ulcers, ulcers with purulent material, single ulcers and ulcers with systemic manifestations was 70%, 70%, 65% and 65% respectively. Such scores may be attributed to low Fig. 1. The software layout showing the starting page. 44 Sh.A. Ali, H.I. Saudi / Tanta Dental Journal 11 (2014) 42e46 prevalence of some of such ulcers such as histoplas- mosis ulcer, oral chancre, ulcers of syphilis and gonorrhea, tuberculosis ulcers and ulcers of neu- tropenia. In addition to some difficulties in diagnosis due to presence of other systemic signs and symptoms that may confuse the postgraduate student or the data Fig. 2. The expert system showing the options for the ulcers with introduced by the system was not enough to reach a diagnosis. The expert system used in the present research was useful and the results were supported by many researches. Expert systems were used in many fields, medically Smitha and Rohini (2010) [6] had used an expert well circumscribed borders. The first option was selected. Fig. 3. The system displays the result of the selection of the first option. 45Sh.A. Ali, H.I. Saudi / Tanta Dental Journal 11 (2014) 42e46 system for the diagnosis of diabetes. Their system was able to decide whether the person was diabetic or not and his level of illness, possible complications and the most appropriate treatment plan. Moreover Entacher et al. (2011) [12] had used the expert system in preoperative testing for patients before surgery. The system was used to collect the patient data as history, general information, and con- dition of lung, kidney, coagulation, drugs, heart and hematology. The collected data helped to prevent double examinations and unnecessary test to reach a proper diagnosis and treatment plan. In the dental field Abbey (1987) [13] had used an expert system for the oral diagnosis concerning the periapical radiolucency. The system takes the patient data and links them to a large fixed database then suggests the causes and the management. Table 1 Showing the mean scores for the ulcers with their detection percentage. Ulcer Mean scores Percentage First time 1.7 ± 0.3 85% Recurrent 1.8 ± 0.4 90% Chronic 1.4 ± 0.3 70% Well circumscribed borders 1.7 ± 0.2 85% Ill defined borders 1.6 ± 0.3 80% With purulent material 1.4 ± 0.3 70% Single 1.3 ± 0.5 65% Multiple 1.5 ± 0.2 75% Without systemic manifestations 1.6 ± 0.3 80% With systemic manifestations 1.3 ± 0.3 65% The results of the present research agree with those of Firriolo and Wang (1993) [14] who used an expert computer system for the diagnosis of selected painful pulpal conditions. The results were promising and successful in endodontic diagnosis. Moreover the results of the present research agree with those of Mc Cracken et al. (2000) [15] who used a computer assisted learning package in the management of traumatized incisors for the general practitioners. They used a questionnaire to evaluate the overall pro- gram and 57% found it easily to use and improved their knowledge in the management of fractured incisors. The results agree with those of Saudi (2002) [16] who used an expert system to help in decision- making and treatment of the most common peri- odontal conditions. The system idea was based on the collection of the patient's data such as age, oral hy- giene, attachment loss, bleeding, recession, tooth mobility, alveolar bone loss, width of the attached gingiva, occlusion, and fitness for periodontal surgery. When the user selects an item, the system displays the available options which mostly influence the decision making in periodontal therapy. An alert database con- necting module was developed to evaluate all entered data and the safest clinical solutions. Also the system could display and print a report for each case. The system was tried by 20 dentists and the mean agree- ment level was 82.3% in 30 periodontal conditions of 35 patients. It was concluded that the system was helpful for general practitioners. 46 Sh.A. Ali, H.I. Saudi / Tanta Dental Journal 11 (2014) 42e46 Also the results agree with those of Xue-jun et al. (2006) [17] who had realized the use of expert system in periodontal disease. They used a Visual Basic 6.0 program to collect the patient data and compare them with those stored in the system to produce the proper diagnosis, prognosis and treatment plan. They concluded that the system was effective in diagnosis and the suggested treatment plans. Moreover the results of the present research agree with those of Shankarapillal et al. (2010) [18] who used the neural artificial intelligence to test the accuracy of periodontal disease risk assessment. Data was collected from 230 patients concerning age, gender, family history of periodontitis, history of periodontal surgery, diet, smoking history, pan chewing habit, history of diabetes, history of hypertension, presence of sub gingival resto- rations, bleeding on probing, debris index average pocket probing depth, presence of root calculus, pres- ence of furcation involvement and vertical bone loss. Periodontitis risk assessment was done on a grade of 1e5. Results showed that neural networks can be used effectively for periodontitis risk prediction. Furthermore the results agree with those of Allah- verdi and Akcan (2011) [19] who used an expert sys- tem for the diagnosis of periodontal diseases. The system receives the patient data including clinical and radiographic findings and determines the diagnosis, disease severity and the treatment methods. It was concluded that the system helped the dentists by facilitating their job with the most correct diagnosis and treatment method. In addition, reducing time loss with some advantages compared with traditional diagnosis and treatment methods. Based on the previous results of the current research, it seems that the introduced expert system may be helpful to the postgraduate students in the diagnosis and management of the most common oral ulcers. Although the artificial intelligence and expert systems may pro- vide an additional help in many medical fields but the human expert factor could not be neglected as the human resources are more wise and merciful. References [1] Muhaidat ZH, Rodan RE. Prevalence of oral ulceration among Jordanian people. Pak Oral Dental J 2013;33(1):42e9. [2] Scully C, Felix D. Oral medicine-update for the dental practi- tioner. Aphthous and other common ulcers. Br Dent J 2005;199:259e64. [3] Lacopino AM. The influence of new science on dental educa- tion: current concepts, trends and models for the future. J Dent Educ 2007;71(4):450e62. [4] Jasinevicius TR, Landers M, Nelson S, Urbankova A. An evaluation of two dental simulation systems: virtual reality versus contemporary non-computer-assisted. J Dent Educ 2004;68:1151e62. [5] Hilenburg KI, Cederberg RA, Gray SA, Hurst CL, Johonson GK, Potter GK. E-learning and the future of dental education: options of administrators and information technol- ogy specialists. Eur J Dent Educ 2006;10:169e77. [6] Smitha V, Rohini V. An expert system for diabetes diagnosis. M.D. thesis. Bangalore: Christ University; 2010. pp. 11e21. [7] Morris RB. Strategies in dental diagnosis and treatment plan- ning. e Book. 2nd ed. Taylor and Francis e-Library; 2004. pp. 77e100. [8] Woo SB, Martin S, Greenberg M. Ulcerative, vesicular and bullous lesions. In: Greenberg M, Glick M, Ship J, editors. Burket's oral medicine. 11th ed. Hamilton: BC Decker Inc.; 2008. pp. 41e75. [9] Leao JC, porter SR. Telediagnosis of oral diseases. Braz Dent J 1999;10:1e6. [10] Parta PS, Sahu DP, Mandel I. An expert system for diagnosis of human diseases. Int J Comput Appl 2010;1:71e3. [11] Field EA, Allan RB. Review article: oral ulceration- aetiopathogenesis, clinical diagnosis and management in gastrointestinal clinic. Aliment Pharmacol Ther 2003;18:949e62. [12] Entacher K, Fritsch G, Huber V, Klausner S. In: Holzinger, Simonic, editors. PROPeA medical expert system for preop- erative testing. USAB, LNCS; 2010. pp. 417e28. [13] Abbey LM. An expert system for oral diagnosis. J Dent Educ 1987;51:475e80. [14] Firrilo F, Wang T. Diagnosis of selected pulpal pathosis using an expert computer system. Oral Surg Oral Med Oral Pathol 1993;76:390e6. [15] McCracken G, Nunn J, Hobson R, Stephenson J, Jepson N. Evaluation of a computer-assisted learning package on the management of traumatized incisors by general dental practi- tioners. Endod Dent Traumatol 2000;16:40e2. [16] Saudi H. A periodontal expert system as an aid in periodontal treatment planning. Egypt Dent J 2002;48:1087. [17] Xue-jun G, Bing L, Hui-zhen C, Shu-yue C, Ke-qin S, Xiu- yun R. Realization of the expert system for periodontal disease. Chin J Geriatr Dent 2006;1:2e11. [18] Shankarapillal R, Mathur L, Nair M, Rai N, Mathur A. Peri- odontitis risk assessment using two artificial neural networks-a pilot study. Int J Dental Clin 2010;2:36e40. [19] Allahverdi N, Akcan T. A fuzzy expert system design for diagnosis of periodontal disease. In: Application of information and communication technologies, 5th International Conference; 2011. pp. 1e5. An expert system for the diagnosis and management of oral ulcers 1 Introduction 2 Materials and methods 3 Results 4 Discussion References 
work_uq3vqye7gzbq7isiz7yigh7m7u	----	A personalized trustworthy seller recommendation in an open market A personalized trustworthy seller recommendation in an open market Seungsup Lee a, Keunho Choi b,⇑, Yongmoo Suh b a LS Global Inc., Sanbon 1-Dong, Gunpo-Si, Gyeonggi-Do, Republic of Korea b Business School, Korea University, Anam-Ro 145, Seongbuk-Gu, Seoul 136-701, Republic of Korea a r t i c l e i n f o Keywords: Recommendation system Content-based filtering Data mining Classification analysis a b s t r a c t Although more and more customers are buying products on online stores, they have a difficulty in select- ing a both trustworthy and suitable seller who sells a product they want to buy since there is a plenty number of sellers who sell the same product with different options. Therefore, the objective of this research is to propose a personalized trustworthy seller recommendation system for the customers of an open market in Korea. To that end, we first developed a module which classifies sellers into trustwor- thy one or not using a classification technique such as decision tree, and then developed another module which makes use of the content-based filtering method to find best-matching top k sellers among the selected trustworthy sellers. Experimental results show that our approach is worthwhile to take. This study makes a contribution at least in that to our knowledge it is the first attempt to recommend sellers, not products as done in most other studies, to customers. Crown Copyright � 2012 Published by Elsevier Ltd. All rights reserved. 1. Introduction The wide spread of Internet infrastructure facilitated e-transac- tions all around the world since early 1990s. Especially in Korea, this type of phenomenon has been observed since then. According to OECD reports, the percentage of the households which have ac- cess to the internet is 80%, the deployment ratio of DSL (Digital Subscriber Line) is about 99.5%, and that of broadband cable is approximately 57%, in Korea. Consequently, such high coverage of infrastructure stimulated e-commerce transactions more than in other countries. Behind the ever-expanded e-commerce industry, there have been new types of problem. That is, relatively low cost of starting a new business in e-commerce would cause more fierce competi- tion among sellers, while providing more choices to the customers (Choi, Stahl, & Whinston, 1997). Therefore, the participants of online transactions can commonly face with significant problems such as ‘‘trust’’ (Chen & Barnes, 2007; Hoffman, Novak, & Peralta, 1999; McKnight & Chervany, 2001) and ‘‘information overload’’ (Mooney & Roy, 2000; Yu, Schwaighofer, Tresp, Xu, & Kriegel, 2004). In order to mitigate the trust problem in online shopping, sev- eral approaches have been proposed. For example, the online shop- ping service providers such as eBay and Amazon suggested reputation systems to promote more honest transactions between sellers and customers (Dellarocas, 2003). The reputation systems allow customers to share their opinion and experiences on prod- ucts, services, and the online shopping mall through a large-scale of word-of-mouth (Dellarocas, 2006). However, the reputation systems are so simple that most of them fail to consider the collu- sive attempts by some fraudulent sellers who increase their own ratings (Wang & Chiu, 2008). In order to deal with information overload problem, on the other hand, fundamental techniques have been studied to develop rec- ommendation systems, including content-based filtering (CBF) (Lang, 1995; Mooney & Roy, 2000; Pazzani & Billsus, 1997) and collaborative filtering (CF) (Yu et al., 2004). Those systems were developed mainly to recommend products. However, recommend- ing sellers who are trustworthy and suitable to each customers has been addressed by few researches, although it is usual that custom- ers have difficulty in finding trustworthy sellers suitable to them in e-malls. Therefore, it would be very helpful to customers if a few trustworthy sellers suitable to each of them are recommended when they want to buy a product, especially in an open market. The objective of this research is two-fold. One is to evaluate the sellers registered on an open market by classifying them as either trustworthy or not using a classification technique such as decision tree, and to select only trustworthy sellers. The other is to propose a recommendation system which recommends best-matching top k sellers to a target customer among the selected trustworthy sell- ers based on the CBF method. The rest of this paper is organized as follows. Section 2 presents literature review regarding classification model, CBF, and seller recommendation. Section 3 describes the seller recommendation system proposed in this paper. Section 4 shows our experimental results from the recommendation system. The last section con- cludes this paper with a summary, implications and limitations of this paper. 0957-4174/$ - see front matter Crown Copyright � 2012 Published by Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2012.08.054 ⇑ Corresponding author. Tel.: +82 2 3290 1945; fax: +82 2 922 7220. E-mail address: keunho@korea.ac.kr (K. Choi). Expert Systems with Applications 40 (2013) 1352–1357 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2012.08.054 mailto:keunho@korea.ac.kr http://dx.doi.org/10.1016/j.eswa.2012.08.054 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 2. Previous works 2.1. Classification model A number of classification techniques have been developed and used for various purposes such as classification, estimation or pre- diction, including decision tree, artificial neural network, Bayesian network, support vector machine, etc. They have been applied to various domains including finance, marketing, health, and so on. For example, credit scoring, which has been widely used espe- cially in finance domain, was designed to classify a loan applicant’ credit status into either ‘‘good’’ or ‘‘bad’’ using such information as marital status, income, and age (Chen & Huang, 2003). It could be adopted to make a decision on loaning (West, 2000). During the early years in that domain, linear discriminant analysis was the first model used for credit scoring (Hsieh, 2004; Hsieh, 2005). Due to its shortcomings (Chen & Huang, 2003), many studies have been later conducted to develop more sophisticated ways of modeling such as logistic regression, kernel density estimation, k-nearest neighbor, decision tree, and artificial neural network (West, 2000). Among them, data mining techniques such as decision tree and artificial neural network have been widely used for the purpose of credit evaluation because of their accuracy of classifying (Witten & Frank, 2005). There are a number of researches in which data mining tech- niques were used for the purpose of credit evaluation in different domains such as loan, bank customer evaluation, and bankruptcy prediction (Hsieh, 2004; 2005; Hu, Patuwo, & Indro, 1999; Ong, Huang, & Tzeng, 2005; Zhang, Hu, Patuwo, & Indro, 1999). Zhang et al. (1999) built artificial neural network model for bankruptcy prediction, and compared its performance with that of logistic regression model. Ong, Huang, and Tzeng (2005) suggested a credit scoring model based on genetic algorithm. We built a classification model using a data mining technique, for the purpose of classifying sellers as either trustworthy or not, which is similar, in some sense, to the credit scoring method. Since decision tree algorithm has been widely used in many domains due to its simplicity and accuracy, we used it to build a classification model to identify trustworthy sellers. We then recommend some of the sellers to each customer using CBF technique. That is, we recommend them, if they deal with the product the customer wants to buy and if historical records of sellers and those of cus- tomers show high similarity in terms of some characteristics of products. 2.2. Content-based filtering With the wide spread of Internet technology in early 1990’s, web-based recommendation systems have been developed mostly taking either collaborative filtering (CF) or content-based filtering (CBF) approach. CF approach recommends items based on the rat- ings of like-minded people, while CBF recommends items based on the similarity between items to recommend and items already purchased. Thus, CF requires ratings of customers on items and CBF requires item profiles, showing their characteristics. Some recommendation systems take advantages of both approaches. Since our data does not satisfy the requirements for CF, we decided to take CBF approach to recommending sellers to customers. So far, there have been many recommendation systems devel- oped taking CBF approach. They were used to recommend books (Mooney & Roy, 2000), music (Kodama et al., 2005), news (Lang, 1995), web pages (Pazzani & Billsus, 1997) and supermarket prod- uct (Lawrence, Almasi, Kotlyar, Viveros, & Duri, 2001). Most of these systems recommend various kinds of products, but few sys- tems recommend sellers of open markets in spite of the growing importance of seller recommendation systems in the e-commerce environment, where customers usually do not have trust in some sellers. 2.3. Seller recommendation To our knowledge, there have been only a few studies on rec- ommending sellers. Xu and Zhang (2009) suggested seller credit evaluation method based on analytic hierarchy process (AHP) and set pair analysis (SPA). The advantage of their method is the fact that they suggested a way of using both qualitative and quan- titative indicators to define a unified score. This method, however, has a limitation in that it cannot control the malicious evaluations on sellers caused by several customers. And recently, Wang and Chiu (2008) described a way of recommending trustworthy sellers using social network analysis. By the structure of social relation- ship of sellers which was made up of historical transactional data between traders, they suggested a collaboration-based recommen- dation system and attempted to lessen the problem of collusive or fraudulent rating for sellers. As mentioned above, although a few studies have been con- ducted to recommend trustworthy sellers, few studies have been conducted to develop a system which recommends trustworthy sellers suitable to each customer. In this paper, we intend to develop such a recommendation system, using CBF technique, in an open market in Korea. 3. The proposed seller recommendation system This section presents the overview of the recommendation sys- tem proposed in this paper, followed by a detailed description of each step of the system. 3.1. System overview The overall framework of our proposed seller recommendation system consists of two main modules (i.e., classification and rec- ommendation modules), as shown in Fig. 1. First, in classification module, we build a classification model using the features selected from a table which integrated customer table, transaction table and seller table, and then identify trustwor- thy sellers using the classification model, with the assumption that the degree of customer’s purchase satisfaction determines whether the sellers are trustworthy or not. That is, sellers are believed to be trustworthy if they have provided customers with good purchase satisfaction, since high satisfaction forms a customer’s trust in sellers (Kim, Ferrin, & Rao, 2003). In recommendation module, profiles of both customers and the identified trustworthy sellers are built, and then, the similarity be- tween the profiles of a target customer and each seller who deals with items the target customer want to buy are calculated. Finally, the top k sellers whose profiles are most similar to that of target customer are recommended. Detailed explanation on each module is provided in following sections. 3.2. Classification module Our proposed system aims to recommend trustworthy sell- ers who provides product or services that each customer wants. To that end, we first built a classification model which classifies sellers into one of two classes, trustworthy and untrustworthy. In order to select features relevant to our classification analysis, we reviewed the previous researches which examine the factors affecting the trust of e-sellers. Through the literature review, we S. Lee et al. / Expert Systems with Applications 40 (2013) 1352–1357 1353 https://isiarticles.com/article/15196 
work_uri7vtryhnc2tjlxc6tsbfvpqe	----	Microsoft Word - NeetaSingh _115-120_.doc Asia Pac J Clin Nutr 2007;16 (Suppl 1):116-121 116 Original Article Expert system prototype of food aid distribution Neeta Singh PhD Department of Nutrition, University of the Incarnate Word, United States of America The aim of this study was to improve efficiency of the food aid distribution process of international food relief organizations. An overall objective of this study was to develop a prototype expert system for monitoring and evaluating food aid by international disaster relief organizations. The research identifies data related to monitor- ing and evaluation processes of various international food-aid organizations. It then applies an artificial intelli- gence-based expert system to develop a prototype for those processes. Existing data related to monitoring and evaluation program cycles were obtained. An expert system shell called CLIPS© (National Aeronautics Space Administration) was used to develop a prototype system named Food Aid Monitor, a rule-based expert system, which uses facts and heuristic rules to provide an adaptive feedback regarding monitoring and evaluating proc- esses at various stages of food aid operation. The Food Aid Monitor was evaluated and validated by three expert panels checking the prototype system for completeness, relevancy, consistency, correctness, precision, and us- ability. Finally, the panels indicated a belief that the system could have an overall positive impact on the stages of monitoring and evaluating food aid processes of the food relief organizations. Key Words: food aid, expert system, disaster relief, monitoring, evaluation Introduction Natural disasters, including floods, famine, fires and other calamities triggered by natural forces, fill the chronicles of recorded history. Along with wars, disasters were, for centuries, the principle events by which people marked the transition from one epoch to another. Man-made and natu- ral disasters create famines and, therefore, the scarcity of food that could affect many people.1 Disaster relief aid is a type of humanitarian assistance that has, primarily, short-term goals. It has usually been directed not just at relieving immediate distress, but also towards rebuilding the infrastructure. While donors were often motivated by a desire to help the unfortunate people affected by a disaster, an operation’s purpose was to bene- fit aid recipients by alleviating their immediate hardship. Thus, the objectives are transparently the same for both donor and recipients. Donors want to help and gain good- will; recipients want help, and in most cases, remember the goodwill.2 Distribution of food aid to beneficiaries has been the last link in the food aid delivery chain. Being the most visible part of the process, it naturally attracts the most attention. Since it generally determines the image of an operation, little attention is paid to the earlier stages of the process. The distribution stage is the one over which donors have least control, since all action and mobilization involved in this step takes place within the territory of the recipient country. Consequently, it is this stage of the operation that is of greatest concern to donors.3 This con- cern is frequently expressed through requests to undertake evaluation and participate in monitoring activities. Donors at all levels are concerned about this final stage of the distribution process, for it is at this point that they discover whether food is getting where it is intended to go. This is the period at which relief organizations are held account- able to provide the outcome that donors expect. Thus, this is when monitoring and evaluation processes play their most critical role. Food organizational inefficiency, politics, and corruption are among some of the problems that can impede the pro- gress of relief operations in terms of food aid distribution or medical services. They not only deprive the victims of life giving support, but also discourage potential sources of contributions. The role of politics in disaster areas is by no means limited to the government of the affected countries alone. Donor government organizations have also been known to delay operations or overlook irregularities in food aid distribution. It should be a major concern for any relief-providing organization to preclude disaster victims from the above-described suffering.4 Some by-products of modernization have led to an explo- sive population growth and have put drastic stress on the fragile balance of the ecosystem. Hence, natural disasters could easily increase the toll of fatalities. The improvement of long-distance communication, transportation, and com- puter technology permits international relief agencies to reach remote disaster areas without much delay.4 These technologies provide the means that alleviate much of the distress caused by a natural disaster. They should be readily available to all victims. Corresponding Author: Assistant Professor Neeta Singh, De- partment of Nutrition, University of the Incarnate Word, 4301 Broadway, San Antonio, Texas 78209 USA Tel: 1 210 829 3167; Fax: 1 210 829 3153 Email: singh@uiwtx.edu 117 N Singh Only rarely has it been unfeasible, with the available pre- sent day technology, to bring aid directly to the place where it was needed. Hence, the obstacles to effective relief in the 1990s were primarily organizational and po- litical. Measures should have been aimed at improving organizational deficiencies that allow disasters to become far more tragic than they need be.5 Although there has always been politics in food aid distribution, now it has emerged as a focused issue. Food has become a controversy and a contest in which the ar- gument revolves around how it should be distributed and how the efficiency of aid distribution may be improved. It has been suggested that monitoring and evaluation can provide a range of quantitative answers with respect to the efficiency of food aid operations, particularly the question of end use.6 Monitoring and evaluation reports have been shown to help administrators and the decision-makers of food aid distribution programs make improvements to increase their effectiveness. A method identified as being particularly effective in improving decision-making strategies is an expert system based method. An expert system is, typically, a computer program that simulates the performance of human experts in a specific field or domain.7 The presence of technology in all aspects of life has enabled solutions to real-life problems that were either difficult or unfeasible. The availability of computer-based database management, monitoring, and maintenance systems and its world wide web-based accessibility have revolutionized information processing for organizations that operate over a vast re- gion and are spread all over the world with active units in various countries. The basic fabric of food aid distribution, monitoring, and evaluation, requires efficient information disseminating systems similar to the ones which are the backbone of all large-scale inventory, marketing, and sales infrastructure for multinational organizations.8 Intelligent inference machines based upon available data are being implemented throughout the industry. Typically, expert system-based decision-making software is employed to review hundreds of gigabytes of data, from a database, and assist human managers to interpret and implement their decisions. It has been shown that, in association with their computer colleagues, human man- agers have been able to improve efficiencies of their op- eration significantly.9 Due to the multinational presence, the mammoth amount of data and the distributed nature of the operation of a typical food aid provider, such a database and expert system, could potentially revolutionize the food aid dis- tribution process the way it has impacted the commercial and industrial sectors. Thus, an expert system-based method of decision-making and data gathering was pro- posed for this research study. Accordingly, a prototype expert system for the monitoring and evaluation processes of food aid was developed. The prototype system has the obvious potential to be useful for food relief organizations seeking ways to improve their monitoring and evaluation processes. Furthermore, the ultimate beneficiaries of the research would be disaster victims. Such an expert system can help to get a higher percentage of aid to the intended beneficiaries (disaster victims). The outcome of this re- search was a prototype expert system, which is intended to facilitate monitoring and evaluation processes at vari- ous food relief organizational levels.2-9 Methods Design The goal of this research was to improve disaster food aid monitoring and evaluation processes by utilizing artificial intelligence technology to build and validate a prototype expert system. To accomplish the above-stated goal, the following objectives were laid down for this research: Gather data (organizational policy/procedure manu- als/program cycles) on monitoring and evaluation proc- esses of food aid from international relief organizations to develop a database. Classify and structure the database into domains, facts, and rules according to the CLIPS©10 expert system syntax. Convert the database into a knowl- edge base. Develop a prototype expert system using the CLIPS© expert system shell. Verify and validate the pro- totype for consistency, correctness, precision (with the goal of software operational accuracy of 95% or more), completeness, and usability. Expert system-based structure The main objective of this research was to develop a pro- totype system for monitoring and evaluating emergency food aid. The system uses artificial intelligence technol- ogy, an expert system-based structure ideally suited for complex knowledge transfers. The monitoring and evaluation processes among international food relief agencies are extremely complicated and intensive.3 Ide- ally the processes require abiding by extensive guidelines, continuous feedback, and decision-making. This can be further complicated when the operations are conducted far away from the donor organizations. Several independent and dependent variables such as number of distribution sites, available staff, food spoilage, and location of food storage need to be considered before decisions are made.3,11 Control problems The problem to be solved in this research was a control problem that required determination of control decisions based upon real-life feedback data. Control problems in- clude processes that require interpretation, monitoring, planning, and prognosis. The expert system requires de- termination of the inference engine based upon the type of problem. An inference engine of an expert system is a component that draws conclusions to execute the highest probable rule, based upon the facts in the available data- base. For example, diagnostic problems are better solved with backward chaining, while prognosis, monitoring, and control are better done with forward chaining.12 Control problems tend to be well-suited for forward chaining rule-based language because of their data driven nature. Forward chaining is reasoning from facts to a con- clusion. For example, if the fact database infers that there is a shortage of food supply in an inventory, one of the conclusions drawn by the inference engine may be to or- der more food supplies to continue the food distribution process. Typically, sets of input values are read during each program execution cycle. Inferencing occurs until all possible conclusions that can be derived from the input Food aid distribution prototype 118 data are reached. This is consistent with a data-driven approach in which reasoning occurs from the data and to the conclusions that can be derived from the data.13 Prototype development The design of this research was prototype development. Prototype development is an initial version of an expert system that is developed to test the effectiveness of the overall knowledge representation and inference strategies being employed to solve a particular problem.12 Prototype development includes defining a problem, identifying an authentic source of data or knowledge, verifying an ap- proach for building a knowledge base, and selecting hardware and software to construct a trial version of a system. The main goal of this research was to develop a prototype expert system for monitoring and evaluating food aid distribution processes, thereby validating the conceptual feasibility of such a system.12,13 Data Source The data sources for this research were the operational information and database maintained by international food relief organizations. Initially, three international food relief organizations were selected for obtaining in- formation regarding each organization’s monitoring and evaluation processes program cycle. The International Council of Red Cross (ICRC), United States Agency for International Development (USAID) and World Food Program (WFP) were the organizations that were con- tacted.14 Rationale The primary reason for selecting three organizations was to comparatively verify the completeness of the documen- tation of their monitoring and evaluation program cycles. Data from the ICRC lacked the required information and hence was not considered for further evaluation in this research. Data sets obtained from the USAID and WFP were sufficiently complete to allow further investigation. It was observed from the USAID and WFP relief opera- tion reports that they worked together for food aid relief operations in cases for which the operational data were obtained. Additionally, published reports indicated that USAID funds the WFP for various emergency operations.4 It was observed, however, that the WFP was equipped with more infrastructure than the USAID for international emergency operations. The USAID and WFP evaluation offices were contacted mainly through electronic mail regarding the requests for operational data. These requests were re-routed to their internal libraries for a data-search by the evaluation offices. The libraries conducted the re- quired data search and provided electronic files with the relevant information. The electronic documents obtained from these libraries consisted primarily of extensive key- word searches and available documents on the requested topics. The available data on the monitoring and evaluation program cycles were adequate to conduct further research though missing in parts. Hence, the missing parts were reconstructed using a partial “dummy” database that also emulated real-life operational conditions. This resulted in a more complete data set for the monitoring and evalua- tion program cycles solely for the research purposes. This information was then structured to analyze the appropri- ateness of the CLIPS© syntax. Finally, the structured information was converted into pseudocode, which is a notation resembling a programming language but not in- tended for actual compilation. System development The CLIPS© software program was used to develop a prototype expert system intended to serve as an aid to monitor and evaluate food aid distribution processes. The development steps included (1) identifying developmental tools and system requirements for building the prototype, (2) determining the method of data analysis, (3) incre- mental buildup of rules to solve the problem, and finally (4) validation and verification of the prototype. Database development In CLIPS©, knowledge can be encapsulated in rules and objects. Furthermore, rules can match patterns or objects as well as facts and objects can operate independent of rules. A rule-based system was developed during the course of this research. The reasons for using a rule-based approach were due to its ease of encapsulation of knowl- edge and future expandability. 12,13 Testing While developing the rules, they were formatted in vari- ous ways using template features of the system shell. Ini- tially, testing was done on a small section of data. The section used for testing was about one fifth of the data set. The data were structured in various formats to find out the best way to encapsulate knowledge and create a knowl- edge base for the system. A pseudocode was developed from the same portion of data. The code was used to de- termine the validity of the syntax and verify the structure of the data. Simultaneously, a tentative knowledge base was generated based on the template creation feature of CLIPS©. This method was inadequate to handle such a varied data type, and it lost much information. It was later realized that the template-based knowledge encapsulation was appropriate for shallow knowledge, whereas the available data characterized a knowledge type, catego- rized by CLIPS©, as deep knowledge. 12,13 Due to the complexity of the available data, the infor- mation was re-formatted into “if” and “then” rules. This rule syntax was able to accommodate more detailed in- formation. This turned out to be the desired method of creating the knowledge base for the system. Trial runs on the same section of data helped to determine any syntax error and programming problems. Finally, upon determi- nation of the validity of the syntax and program execution for the same portion of data, all of the information was converted into “if” and “then” rules. Before programming, the rules were further divided into three files according to various food aid operation stages: pre-operation, opera- tion, and post operation. The pre-operation batch file con- tained information regarding monitoring and evaluation processes before relief operations start. Operation batch files were further divided into three sections that con- tained data regarding the operation stage. The last batch 119 N Singh file contained data regarding the post-operation stage of monitoring and evaluation processes of food aid distribu- tion. 15 Prototyping “Expert system prototype” refers to a scaled down version of a larger system. It typically includes representation of knowledge captured in a manner that will enable quick inference and the major components of the expert system on a rudimentary basis. Prototype development includes definition of a problem, verification of an approach, iden- tification of an authentic source of data or knowledge, and selection of hardware and software. After the prototype is developed, the system developer and expert panels verify and validate the results, and make improvements to the basic system in order to include the required enhance- ments. The system typically requires addition of complete information about a process to the knowledge base and goes through several iterations with required refinements before the final expert system becomes available for gen- eral use. 16 System verification and validation The final stage in developing an expert system involves validation and verification of the system. The process of confirming accuracy and effectiveness of the methodol- ogy, as applied to the product, is known as verification. Verification means building the system right, that is, en- suring that the system correctly implements the specifica- tion.17 It determines if the knowledge base conforms to its design requirements and the software syntax from which it is built.18 In order to test the accuracy of the prototype system, it must be subjected to the food distribution proc- ess during a disaster. For this research the programmer performed verifica- tion of the system. The verification phase of the prototype development included using built-in CLIPS© verification features. While programming, the rules were checked for inconsistencies, syntax error, and duplication. The knowl- edge base and the inference engine system were subjected to verification by checking their rule syntax. Typical checks included rules with same name, rules with incor- rect syntax, redundant rules, isolated rules, subsumed rules, conflicting rules, and circular rules. These pro- gramming errors would have prevented rule production by the system.19 Validation is the process of ensuring that the method- ology, as applied, produces the desired results for the pro- totype expert system for monitoring and evaluation of food aid. In expert system terminology, determining that a chain of correct inferences leads to the correct answer is called “validation”.17-19 Real-life emulation for the valida- tion and verification of the methodology would require the development of a food aid distribution, monitoring, and evaluation plan (validation) and also subjecting the plan to a disaster scenario to observe the outcome (verifi- cation). Unfortunately, employing the prototype for a food aid distribution scenario is well beyond the scope of the present research. Validation, therefore, was limited to expert reviews. This process allowed experts and potential users of the methodology to determine the desirability and the appli- cability of inputs and outputs of the prototype system for practical purposes. While this method was primarily a subjective approach, it still had the capability to address practical situations and generate successful results by in- corporating specific requirements based upon the experi- ence of experts in real-life disaster situations. Based upon review of literature,20 questionnaires and checklists were developed to facilitate the validation process. The validation process consisted of checking the prototype system for completeness by a university-faculty expert panel; consistency, correctness and precision by a software-engineering expert panel; and usability by a field expert panel. The prototype validation for completeness included checking the knowledge base for satisfactory antecedent and consequent parts of the rules. It also required check- ing for relevancy of each rule to the subject matter. This was achieved by having the faculty expert panel check the rule combination matrix of the prototype knowledge base. The software validation process for checking consis- tency, correctness, and precision included running the prototype system and using a rule combination matrix to detect system errors and complete the checklist. A panel of software engineers checked the prototype for fact vali- dation, unused facts, unused rules, multiple methods, run- time errors, and unfired rules. The objective of this stage was to achieve an accuracy of 95 percent or more. Finally, a panel of field experts checked the usability of the proto- type by reviewing the knowledge base for the applicabil- ity of the facts and rules to real-life situations and com- pleted a questionnaire. Three expert panels were selected for the validation process. The first panel was responsible for evaluating relevancy (completeness) of the prototype knowledge base. Faculty members from Oregon State University (Corvallis, Oregon) were chosen as expert panelists for this purpose. Individuals with software engineering exper- tise were the panelists who reviewed the prototype source code for accuracy (consistency, correctness and preci- sion). Applicability (usability) of the prototype was evaluated by an expert panel consisting of individuals with knowledge and experience in the field of interna- tional development (diverse background in food aid relief operations such as working for international development projects and/or field experience with food aid). The expert panelists were selected based on informa- tion obtained from various international food relief or- ganizations, software associations, educational institu- tions, and personal knowledge. Ten panelists (three fac- ulty members, three software engineers, and four workers from food relief organizations) agreed to participate in the research and completed the validation activities. The validation activities consisted of checking the rule combination matrix of the knowledge base and/or testing the prototype system. Expert panelists were asked to run the prototype system on a PC and/or check the rule com- bination matrix and finally, complete question- naires/checklists. Cover letters along with the check- lists/questionnaires/Food Aid Monitor (FAM) prototype were mailed to all the expert panelists. The process of running the prototype system was facilitated by instruc- tions on how to start the program as well as a CLIPS© Food aid distribution prototype 120 programming guide. The last part of the validation activi- ties included completing checklists/questionnaire. Changes in the system were made based on the panel’s comments and are presented in the results and discussion section. Results and discussion An expert system prototype for monitoring and evaluation of food aid was developed and named FAM. The FAM is a rule-based prototype and was developed using both facts and heuristic rules. The FAM was developed for use as an aid for decision-making regarding food aid monitor- ing and evaluation processes at various stages of food relief operations (pre-operation, operation, and post- operation). This allowed for forward chaining or the data driven approach, which starts with available information and draws conclusions from that information. Rules within the knowledge base were written in “IF” (antece- dent), and “THEN” (consequent) statements. While the information was derived from international food aid or- ganizations, the literature review determined the facts and rules that were used as the knowledge base of the proto- type expert system. The FAM was designed for use by international food relief organizations for decision-making regarding moni- toring and evaluating processes. When installed on an organizations’ computing system, the general capability of the system was to provide advice and recommenda- tions regarding food aid monitoring and evaluation proce- dures at various stages of the relief operation, thereby improving the process. FAM creates advice and recom- mendations for the food aid organizations, which, when used internally (within the organization), provide critical need-to-know information to employees. The prototype could be considered a useful tool for decision-makers at the headquarters level as well as for the operation staff of food relief organizations. The design of the FAM essentially consisted of devel- oping the rule matrix, generating the pseudocode, and, finally, writing the source code in CLIPS© syntax. As mentioned before, the FAM is an expert system prototype designed to be used as an aid for decision making regard- ing monitoring and evaluation of food aid at various stages of food relief operations. Hence, it is a rule-based prototype and was developed using both facts and heuris- tic rules. This allowed for forward chaining or the data driven approach, which starts with available information and draws conclusions from that information. Rules within the knowledge base were written in “IF” (antece- dent), and “THEN” (consequent) statements. An overall objective of this research was to develop a prototype expert system for monitoring and evaluating food aid as well as validating the system source code for structural problems. The research was supported by a re- view of literature, which indicated lack of food aid moni- toring and evaluations among food aid relief organiza- tions. No previous research was directly related to this topic, which made the analysis and comparison of the prototype extremely difficult. The task of evaluation be- comes difficult if there is no previously established yard- stick to measure performance. Thus, the only justifiable way to evaluate the performance of the prototype was to use validation by various experts. The faculty experts panel was mainly responsible for identifying completeness of the source code, which is required for useful and meaningful rule production by the system. The experts expressed that the validation process could have been improved if the researcher had provided more materials such as organizational charts and a de- tailed explanation of monitoring and evaluation program cycles. Some activities, like demonstration of the source code in a panel meeting prior to completing the validation checklist, would have helped the experts to have a better understanding of the program and its limitations. The rea- sons for omitting the meeting and keeping the validation process short and concise was due to time constraints ex- pressed by the panel. The validation of the FAM prototype by the engineer- ing experts went very smoothly, due to the fact that they were only responsible for checking the program for struc- tural errors and not data deficiencies. Software develop- ment easily accommodates numerical data or shallow knowledge, as opposed to knowledge intensive data or deep knowledge. The reported problems were corrected after round one of the source code validation process. Finally, the field experts’ validation of the FAM proto- type indicated a need for such a system among food aid relief organizations. As expected, however, the panelists repeatedly commented upon the generic nature of the pro- totype and the need to incorporate detailed practical is- sues for real-life implementation. The field experts be- lieved that the system might have an overall positive im- pact on the stages of monitoring and evaluation. Nonethe- less, they pointed out a few stages and examples of the process where the system needed to be updated according to their organizational needs and particular disaster sce- nario. The panel found that the system could be extremely helpful in planning, feedback process, conducting inven- tory audits, data collection/archive storage, and in follow- ing operational guidelines. Overall, the FAM validation process by the three expert panels helped to detect struc- tural errors and source code errors, and to analyze the usability of the system. It should be noted here that the FAM was an attempt to develop a proof-of-concept prototype system to verify the applicability of an expert system for the food aid distribu- tion process. While the FAM demonstrated the efficacy of an expert system based monitoring and evaluation process for food aid distribution, the specific and detailed practi- cal issues were beyond the scope of this research. Hence, as stated before, implementation would require the addi- tion of various practical issue rules to the basic knowl- edge base developed in this work. References 1. Tisch JS, Wallace BM. Dilemmas of development assis- tance. San Francisco: Waterview Press. 1994. 2. Cassen R. Does aid work?: report to an intergovernmental task force. Oxford: Clarendon Press. 1986. 3. Benini AA. Uncertainty and information flow in humani- tarian agencies. Disasters 1997; 21: 335-353. 121 N Singh 4. Committee on Government Operations. World food pro- gram: funding and management improvements can strengthen delivery of food aid. Washington, DC. 1994. 5. Cathie J. The political economy of food aid. New York: St. Martin’s Press. 1984. 6. Singer H, Wood J, Jennings T. Food aid: the challenge and the opportunity. Oxford: Clarendon Press. 1987. 7. Bowen JT, Clinton DN. Expert systems: advisor on a disk. Cornell Hotel Restaurant Adm Q 1998; 29: 62-67. 8. Green S. International disaster relief. New York: McGraw- Hill. 1980. 9. Masuch M. Organizations, management, and expert sys- tems. New York: Walter De Gruyter. 1990. 10. CLIPS©. Available at: http://www.ghg.net/clips/CLIPS. html. Accessed in January 1998. 11. Marsi A, Moore J. Integrated information systems: disas- ter-planning analysis. J of Urban Plng and Devel 1995; 20; 19-21. 12. Giarratano J, Riley G. Expert systems: principles and pro- gramming. Boston: PWS Publishing Company. 1994. 13. Medsker L, Liebowitz J. Design and development of expert systems and neural network. New York: Macmillan Pub- lishing. 1994. 14. Brown BJ. Disaster preparedness and the United Nations. New York: Pergamon Press. 1991. 15. Frenzel LE. Crash course in artificial intelligence and ex- pert systems. Indianapolis: HW Saws. 1987. 16. Duan Y. The use of expert systems for decision making in organizations. Birmingham: Aston University. 1994. 17. Cragun BJ. A decision-table based processor for checking completeness and consistency in rule-based expert systems. Int J of Man-Machine Stud 1987; 26: 633-648. 18. Nguyen TA. Verifying consistency of production systems. Proceedings of the IEEE Third Conference on Artificial In- telligence Applications. Orlando, Florida.1987 Feb 23-27. 19. Bahill T. Verifying and validating personal computer-based expert systems. Upper Saddle River: Prentice Hall. 1991. 20. Fields DL. The application of computer-aided expert deci- sion support systems to developing countries: a case of ru- ral development in Kenya. Urbana-Champaign: University of Illinois at Urbana Champaign. 1992. 
work_urmbglu2djbjvmdzmvfengjs7y	----	Page 1/19 Determining the performance of website-based relationship marketing Kerstin Schäfer, Tyge-F. Kummer Keywords: Website performance evaluation, relationship marketing, clickstream data, online consumer behavior Page 2/19 Abstract Company websites are an important instrument for relationship marketing activities. We present a methodological framework that aligns website performance assessment and marketing intelligence for evaluating the performance of relationship marketing activities. In this context, we develop an extended web mining approach that integrates managerial perspectives in the analyst’s investigation of the customer-website interaction based on historical clickstream data. This approach enables quantification of the moderating effect of a website’s structure and content regarding website-based relationship marketing. The applicability of our approach is demonstrated by clickstream data of 477,471 visitor sessions on a software developer’s website. The results provide detailed insights into the usage behavior on a website and the mechanisms to enhance e-commerce efficiency via website optimization. Page 3/19 1. Introduction Today’s companies use their websites as windows through which much of the world will see it (Winter, Saunders, & Hart, 2003). In contrast to traditional e-commerce sites or portals, modern corporate websites provide a broad range of information about themselves, their products and related services to the public (Stuart & Jones, 2004). As a result, the online presence becomes the most important communication channel a company possesses; determining how effective it supports this communication function is crucial in the overall webpage evaluation (Weinberg, Parise, & Guinan, 2007). Several approaches have been introduced to improve web performance covering models based on web mining and click stream analysis (Cho, Kim, & Kim, 2002; Chou, Li, Chen, & Wua, 2010; Kalczynski, Senecal, & Nantel, 2006; Kim & Yum, 2011), web-based marketing (Tsai, Chou, & Leu, 2011; Wang, Wang, & Farn, 2009) as well as online effectiveness and website evaluation (Chou & Cheng, 2012; Kim, Kwon, & Chang, 2011). Due to its practical importance, this issue is for example also addressed by popular evaluation instruments such as the Web Effectiveness Index, which helps investigating how a corporate website reacts on global communication trends and how it performs against its peers by quantifying the serving potential for different public target groups (BowenCraggs & Co, 2011). Although such performance indicators are beneficial for companies, they often exclude direct business related measurements, because they cannot quantify website performance in their serving character for the customer value (Financial Times/Bowen Craggs & Co, 2011). But this is of utmost importance, as besides benchmarking among peers, companies need to assess customer value first for achieving a competitive advantage (Woodruff, 1997). Business relevant measures can be included by referring to the company’s applied relationship marketing strategies (RMS), which aim at establishing, developing, and maintaining successful relational exchanges with its customers (Morgan & Hunt, 1994). These strategies are directly reflected in online performance, as the Internet also became the most important communication channel for a company’s RM effort (Srinivasan & Moorman, 2005). The evaluation of such RM effectiveness is based on customer behavior and allows increasing a company’s return of RM investment by offering demand-actuated strategies and services (Reinartz & Kumar, 2003). In this line of arguing, customer value can be determined in two steps: firstly, by measuring how well the corporate website moderates the company RMS, which is expressed in how well it serves its customers online; and secondly, by inferring involved relationship economics (Palmatier, Dant, Grwal, & Evans, 2006). As a result, the evaluation of corporate online performance can be integrated with internal business measures and is substantial for marketing decisions (Weischedel, Matear, & Deans, 2005). Page 4/19 Similar to the spread of relationship marketing evaluation that mainly uses customer data, website performance assessment increasingly takes place on the customer/visitor-level, empirically using clickstream data (Ayanso & Yoogalingam, 2009; Hahn & Kauffman, 2004; Kalczynski, et al., 2006; Olbrich & Holsing, 2011/12). From the viewpoint of business practice, particularly the “easy-to-use” web analytic approaches that use internal web page-tagging, have been widely employed as they provide a fast method for surveying visitor behavior (Kohavi, Rothleder, & Simoudis, 2002). However, their scope of analysis is usually limited to a structural, aggregated level of internal page frequentation and related web metrics such as stickiness or prominent exit pages (Booth & Jansen, 2009; Olbrich & Holsing, Winter 2011-12), which are difficult to relate to customer value. Investigating whole visitor paths using clickstream data can set a much broader scope; further, this allows to infer business relevant website performance measures from visitor usage behavior, as specified in various web mining approaches (Lee, Podlaseck, Schonberg, & Hoch, 2001). Nevertheless, such elaborate approaches are highly underutilized in business practice due to their increased complexity in data storage, data preparation and analysis in comparison to typical web metric-based methodologies (Sen, Dacin, & Pattichis, 2006). However, only an investigation on website performance based on the individual visitor session level enables a valid demand-actuated evaluation of customer/visitor value by quantifying the effectiveness of a company’s RM effort in their website usage. Though, suitable methodological approaches that are able to enrich existing company data by providing internal measure of website performance in the context of relational marketing analysis are rare (Hahn & Kauffman, 2004; Hochsztain, Millán, Pardo, Pena, & Menasalvas, 2003). Furthermore, as far as we know, approaches that directly link the evaluation of a company’s RM effectiveness to website performance, empirically based on customer website usage, do not exist. In this study we attempt to overcome this shortage by presenting a methodological framework that aligns marketing analysis with web mining approaches on the theoretical level. Therein, we develop a methodological approach that integrates managerial perspectives into the modeling and analyzing of a company’s RM effectiveness in website performance based on customer website usage data. The knowledge gained supports customer value determination by quantifying RM effective website usage behavior and also enables a demand- actuated website optimization and customization for different relational marketing strategies. Furthermore, we extend existing web mining approaches by integrating managerial perspectives on website usage. Addressing Internet based marketing communication environments finally enriches traditional marketing analysis approaches. Page 5/19 The paper is structured as follows: in the next section (2) we outline the related research. In section 3, we describe the developed methodology in detail. The practical relevance of our approach will be demonstrated in section 4 by analyzing historical clickstream data of a corporate web site from the software development sector. Finally, we conclude our research contribution and the limitations of our approach in section 5. 2. Related Research in Performance Assessment Corporate websites are an important channel for Internet based marketing communication in a company’s multi- channel marketing efforts (Weinberg, et al., 2007). They support three relational marketing strategies (RMS) in particular: (1) the building of brand equity, which is the sum of the intangible assets of the corporate brand that are supported by factors such as name awareness, perceived quality and customer loyalty (Aaker, 1993); (2) the creation and maintenance of relationships at reduced costs (Sheth & Parvatiyar, 2000); and (3) the creation of customer satisfaction by delivering superior products and services (Gale, 1994). Success in any of these strategies leads to an increase in repeat purchases, insulation from price increases and improved responsiveness to marketing communication by customers. Thus, a corporate website can maximize the impact of a company’s RM efforts (Argyriou, Kitchen, & Melewar, 2006). The strength of this contribution depends on how well the website performs with respect to customer information needs. Therefore it is necessary to develop metrics that evaluate the effectiveness in relation to customer website usage of the RM effort in website performance. For business orientated performance measurements, marketing analysis provides a rich set of methodologies, which have been adapted to the Internet context subsequently. Initially, business performance assessment was conceptualized as assessment of product and service quality (i.e. SERVQUAL: (Carman, 1990; Parasuraman, Zeithaml, & Berry, 1988). In these studies, the general measurement model determines a gap (the deviation) between managerial expectations towards performance and the actual performance based on customer evaluation of products and services (Asubonteng, McCleary, & Swan, 1996; Teas, 1993). In the context of RM, the effectiveness evaluation uses data intensive methodologies based on customer purchase behavior (Vercellis, 2009). In related research this is commonly discussed in the notion of relationship economics that includes metrics such as customer lifetime value, acquisition, retention and cross-selling (Büchner & Mulvenna, 1998; Chi & Tavella, 2008). Subsequent, in IS an adaption of these marketing approaches took place. Website performance is either measured in terms of single marketing outcomes such as web-customer satisfaction and loyalty (McKinney, Yoon, & Zahedi, 2002), or by deploying so-called WEBQUAL studies that focus rather on quality assessment and benchmarking (Lee, Strong, Kahn, & Wang, 2002). Page 6/19 Additionally, other IS studies investigate website performance directly by inferring customer value of the customer-website interaction in methodological approaches similar to marketing intelligence (Senecal, Kalczynski, & Nantel, 2005; Spiliopoulou & Pohle, 2001). In these studies, different effectiveness metrics are deployed depending on the website type. For transaction-orientated e-commerce websites, performance is assessed in terms of revenue. Thus, metrics are used that relate the customer lifecycle to customer web usage such as reach, acquisition, conversion, click through and look-to-buy (Lee, Hoch, Podlasek, Schonberg, & Gomory, 1999; Teltzrow & Berendt, 2003). Investigating customer web site usage in terms of customer behavior on the other hand assesses the performance of information and service provision orientated websites. Here, the clickstream data is interpreted as implicit feedback for meeting customer information needs (Bucklin & Sismeiro, 2003) and effectiveness can be determined by evaluating the degree or amount of this serving (Stolz, Viermetz, Neuneier, & Skubacz, 2005). 3. Towards a Methodology for RMS Effectiveness Evaluation 3.1 Relationship Marketing Website Typology Integrating relationship marketing strategies (RMS) with related research from website performance assessment, an allocation of typical RMS and related metrics is possible: According to the functions a corporate website possesses, different RMS are supported, and different effectiveness metrics need to be distinguished. Adjusting the typology of website functions and metrics suggested by (Booth & Jansen, 2009) for corporate website functions, as displayed in table 1, we specify firstly that a commerce orientation focuses on getting customers who visit the site to purchase goods or services directly from the website. The building of brand equity is the most important RMS here, because it addresses customer loyalty in terms of repeated purchases and perceived product quality. Clickstream-based effectiveness metrics for commerce emphasize transactions such as purchase or downloads. Secondly, content and media provision focuses on drawing in visitors and immersing them within the site. The relationship between a company and its customers can be strengthened as a result. Clickstream- based RMS effectiveness metrics are in this case concerned with visitor engagement, for instance a high number of visit actions indicating browsing, and long session duration. Finally, the support and service function helps users to find specialized answers for specific problems. This increases customer satisfaction at reduced external support costs. RMS effectiveness depends largely upon the provision of the searched information with regard to the structure of the website; thus, low page depth is a useful indicator in clickstream data. Table 1 Website functions, RMS and metrics (Booth & Jansen, 2009) Page 7/19 Website function Related relationship marketing strategy Focus of web metric Commerce Brand equity Conversion Content/media Relationship maintenance Visit depth, engagement Support/self service Customer Satisfaction Page depth, fact-finding In the following, a general research framework is presented and operationalized that integrates RMS effectiveness determination in a web mining approach for modeling, evaluating and predicting RMS effectiveness in corporate website performance. 3.2 An Extended Web Mining Approach The allocation of relational marketing strategies (RMS) to website functions is necessary for developing an extended web mining approach that integrates managerial perspectives in the analyst’s investigation of the customer-website interaction based on historical clickstream data. This integration is reflected in our three-phase approach that contains six single steps: in phase 1, the business expert provides domain knowledge for identifying and characterizing typical corporate RMS in different website areas and related visitor activities (step 1). This domain knowledge is necessary for modeling and validating visitor activities in the clickstream data set by an analyst using statistical procedures (step 2). In Phase 2, the business expert further weights his previously specified visitor activities according to their relevance for achieving RMS effectiveness website area specific (step 3). Subsequently, the analyst calculates overall RMS effectiveness values in the clickstream data set among the modeled visitor activities using these relevance weights (step 4). In phase 3, the analyst determines RMS effectiveness factors by initially measuring the relative contribution of the modeled visitor activities to overall RMS effectiveness and by finally calculating the gap between the business expert’s expectations towards the relevance of an activity in reaching RMS effectiveness and its empirical contribution to overall RMS effectiveness (step 5). The gained insights are then used as managerial implications by the business experts and transferred into website optimization requirements (step 6). Figure 1 gives an overview of the evaluation procedure. Fig. 1 Overview of developed methodology for RMS effectiveness evaluation Page 8/19 Phase 1: Characterizing and modeling visitor activities Step 1. In an initial, exploratory approach a business expert provides the necessary corporate domain knowledge that guides the main analysis (Pyle, 1999). In it, he defines corporate RMS that are moderated by website performance and relates them to different website areas. For it, the expert manually explores the website structure and content and also uses an exploration data set (random sample of visitor sessions, n ≈ 50). This enables him to discriminate typical visitor activities which represent customer website interaction and to set thresholds of the used variables that are used in the clickstream analysis for each website area. Such typical activities are information gathering or browsing, fact-finding, and all possible transactions such as buying or downloading (Kellar, Watters, & Shepherd, 2006). In clickstream data, browsing is usually characterized by a high number of clicks (visit actions) and a long session duration in a visitor session. Fact-finding, can be described by a low page depth, which implies a low number of visit actions and a certain style of interacting with the structural elements of the website content (Bucklin & Sismeiro, 2003). Integrating a distinction of different page types, i.e. into action/navigational and target pages (Spiliopoulou & Pohle, 2001), fact-finding is expressed by visiting only few navigational pages in order to find specific content on target pages. Step 2. These insights are particularly relevant for the analyst, who uses this qualitative activity characterization for the specification of quantitative models and their validation in a training data set (random sample of visitor sessions from the historical clickstream, n ≈ 5000) With the help of these models, the analyst infers visitor Page 9/19 activities from the historical clickstream data using statistical classification methods (decision tree) (Fox, Karnawat, Mydlands, Dumais, & White, 2005). Therein, previously defined expert characterizations are used both for modeling of different visitor activities, as well as for identifying and selecting activity representing nodes in the decision tree. Comparing different models for each activity further ensures statistical validity. Phase 2: Relevance weighting of visitor activities for RMS effectiveness Step 3. Assuming that the corporate website supports different RMS, the initially characterized visitor activities are then weighted by the business experts according to their relevance for achieving RMS effectiveness in customer website usage. We express this formally: the strategy set S ={s1, s2, …, sn}represents the current relational marketing strategies of a company that are moderated by the performance of its corporate website. The customer website usage set U ={u1, u2, …, um} represents different activities a visitor can engage in when interacting with the corporate website. Further, we define wij as the weight to each relational marketing strategy, where i represents the i th visitor activity and j the j th relational marketing strategy. It is assumed that business experts assign weights differently for each visitor activity in every relational marketing strategy. The weight function ij w depends on UuSs ii  , and has the following properties: 0 £ w ij £ 1 with w ij = 1 j=1 m å i=1 n å (1) For each RMS the sum of all visitor activities is always 1; when an activity possesses no weight, then it is not relevant for this RMS from a managerial point of view. Table 2 displays an example allocation of weighting properties, with “-“ indicating no relevance of an activity for this website area. Table 2 RMS specific relevance weighting of visitor activities Visitor activity Brand equity (s1) Relationship maintenance (s2) Customer satisfaction (s3) Long stay/engagement (u1) - s2u1 - Information- gathering/Browsing (u2) - s2u2 - Fact-finding (u3) s1u3 s2u3 s3u3 Buy, download (u4) s1u4 - s3u4 Page 10/19 Step 4. This relevance weighting of visitor activities helps the analyst to determine overall RMS effectiveness in the clickstream data. For doing so, the analyst enriches the statistically identified and validated activities in the visitor sessions with the specified weights for each website area. Then, an effectiveness heuristic is applied for determining the amount of visitor sessions with RMS relevant customer website usage behavior in a website area, using a linear equation L that specifies a dichotomous values: if the visitor session contains any RMS relevant activities, then the website successfully moderates a company’s RM effort in this area; if the user session does not contain any RMS relevant activities, than the website fails to moderate the RM effort in this area. This is formally expressed by: L i = s i u j j=1 m å . (i.e. L1 = s1u1 +s1u2 +... +s1uj ) (2)  If then RMS effectiveness is determined in a visitor session,  If than RMS failure is determined in a visitor session. Phase 3: RMS effectiveness contributing factors Step 5. Such a dichotomous overall RMS effectiveness value is further necessary for the last step of our evaluation procedure, where the analyst estimates the relative contribution of modeled visitor activities to overall RMS effectiveness website area specific using discriminant analysis. Further, by relating the actual contribution of a visitor activity to the specified relevance for achieving RMS effectiveness, the moderating effect of corporate website performance is quantified using gap analysis. Conceptually similar to Brown and Swartz’s (Brown & Swartz, 1989) gap determination procedure, we calculate a “RMS effectiveness contribution- relevance gap” calculated visitor activity wise for each website area and its attached RMS. In it, the relative relevance values ij w with UuSs ii  , that are generated by business experts in step 3 (Eq. 1) are subtracted from the contribution activity values cij with UuSs ii  , (see Eq. 2) that are estimated by discriminant analysis in step 5. The determination of gap values Gij is expressed: G ij = c ij -w ij (3)  If Gij ≥ 0 than the website supports the website moderation of the company’s RM effort as reflected in customer website usage;  If Gij < 0 than the website underperforms in moderating the companies RM effort. Page 11/19 Step 6. These findings finally allow the analyst to formulate demand-actuated optimization recommendations that can increase RMS effectiveness of the corporate website and are used by the business expert for managerial decision-making. 4. Example Implementation and Empirical Results The developed approach was applied for evaluating the effectiveness of the relational marketing strategies (RMS) in the website performance of a German software developer. The website is dynamic and contains over 25 billion distinct URLs for content generation. Further, the website possesses different functions that are reflected in distinct website areas such as commerce (shop), content and media provision (product catalogue, media center, company, community) and service and support (support). As specified by business experts, relevant RMS that are moderated by website performance are: (1) building of brand equity (shop area); (2) relationship maintenance (all content providing areas); and (3) creation and increase of customer satisfaction (support area). 4.1 Clickstream Data Collection and Data Preparation For investigating RMS effectiveness in website performance, we collected server log files of the clickstream over a period of four weeks (6.5 GB). In the following data preparation and transformation process, according to (Cooley, Mobasher, & Srivastava, 1999; Markov & Larose, 2007), data was cleaned and sessionized using MySQL 5.5.9. In total, we identified and prepared 1,389,413 sessions with a session duration longer than 10 seconds and containing two or more visit actions for statistical analysis. For integrating structural information that allow distinguishing navigational and content page view objects, additional data was collected with a java- based agent: the crawler requested the HTML page title of each URL in the log files, which offers information about the website area and parsed them either as navigational or content objects. To this end, we deployed a decision heuristic: page view objects with more than 30 links are navigational; those with less than 30 links are content objects. This classification was successfully evaluated and validated by comparing the mean of session duration values (Spiliopoulou & Pohle, 2001). In total, we identified 143546 distinct navigational, and 11075070 content objects in our data set. For applying our developed web mining approach, we generated four data sets that allocate visitor sessions according to their clicks in a website area such as products, shop, community and support (multiple response possible). In further data quality monitoring, outliers were identified and removed using the interquartile range method (Larose, 2005) for the ratios of visit actions and session duration in each data set. Table 3 displays the Page 12/19 average session ratio values that were used for model specification after outlier removal in the net samples (sum of net samples = 477,471 sessions). These values also depict differences in customer website usage across website areas, which supports our website area specific evaluation approach. Table 3 Average session ratio values in website areas Mean per website area Products (n = 114,264) Shop (n = 158,529) Community (n = 20,052) Support (n = 184,626) Session duration per website area (SD) 20.11 156.36 277.14 146.80 Session duration per navigational object (SD_Nav) 14.86 18.82 21.40 33.52 Session duration per content object (SD_Con) 5.25 137.55 255.74 113.28 Visit actions in website area (VA) 2.85 3.85 3.98 3.42 Visit actions navigational objects (VA_Nav) 2.06 .59 .56 .80 Visit actions content objects (VA_Con) .79 3.26 3.41 2.62 4.2 Data Analysis and Results Statistical analysis was performed using PASW Statistics 18 and PASW Modeler 13. Data analysis proceeded according to the developed approach in three phases and six steps (figure 1). In step 1, the business expert executed a qualitative characterization and discrimination of RMS relevant visitor activities using the website and a provided exploration data set. As a result, thresholds for activity-modeling using the ratios visit actions (VA) and session duration (SD), and other values of the indicator variables that describe visitor activities were defined. In the subsequent classification procedure in step 2, a C&RT regression tree was used for predicting the values of the visitor activities browsing and fact-finding, and for classifying visitor activities, first within a training data set and after successful validation in the whole data sets of the net samples. Measurement validity was ensured using the Gini-index as measure of impurity and a test-sample cross validation for tree-selection. The identification and selection of visitor activity representing nodes in the regression trees deployed the expert characterization. Then, two data models for each activity were created and compared, using one dependent (DV) and one or more independent variables (IV). For valid activity inference, the model with the lower standard error of the estimate (s) was finally preferred. Then, an indicator variable was created in the visitor sessions for each inferred activity (Breiman, Friedman, Stone, & Olshen, 1984). The regression models and the results of the model evaluation that prefers model 1, as well as the predicted values of the ratios in model 1 are shown in the Page 13/19 Appendix (Table 8, 9). In the step 3, the business expert then weighted visitor activities according to their impact for achieving effective RMS in each investigated website area using Eq. 1, as presented in Table 4. Table 4 Relevance weighting of visitor activities Visitor activities Products (Relationship maintenance) Shop (Building of brand equity) Community (Relationship maintenance) Support (Creation of customer satisfaction) Weights of visitor activities Long stay .5 - .5 - Browsing .2 - .3 - Fact-finding .2 .1 .1 .9 Download .1 .1 .1 .1 Purchase - .8 - - Sum 1 1 1 1 Subsequently, in step 4 overall RMS effectiveness was determined deploying the decision heuristic the Eq. 2 (Table 5). Results demonstrate that the company’s RM effort is moderated differently in each website area: for relationship maintenance in the product area, only 45 % of all visitor sessions were effective; whereas in the community area, which provides an exchange forum between customers, a high effectiveness of 77% in the visitor sessions was achieved. For the building of brand equity in the shop area, the website performed moderately efficient, with 48% of visitors sessions supporting RMS effectiveness. The creation of customer satisfaction displays the highest RMS failure, including 75% of all visitor sessions. Table 5 Results of overall RMS effectiveness determination in visitor sessions Products (Relationship maintenance) Shop (Building of brand equity) Community (Relationship maintenance) Support (Creation of customer satisfaction) N % N % N % N % RMS effectiveness 51,308 44.9 74,638 47.1 15,579 77.7 46,375 25.1 RMS failure 62,956 55.1 83,891 52.9 4,473 22.3 138,251 74.9 Total 114,264 100 158,529 100 20,052 100 184,626 100 In step 5 of our evaluation, discriminant analysis was used for determining the relative contribution of the weighted activities to RMS effectiveness, which was also used as input for the following gap analysis procedure. For ensuring measurement validity of the discriminant function, the assumption of homogeneity of variance was Page 14/19 fulfilled; analysis could thus be conducted. In model evaluation the quality of discrimination was very high, as shown in high eigenvalues and low values of Wilk’s Lambda (eigenvalues between 5.04 and 8.84; Wilk’s Lambda between .10 and .17 with p < .001). Further, classification results for the predicted group membership (RMS failure/effectiveness) are satisfying with 95.3 % (products), 95.8 (community), 99% (support), and 100% (shop) for the correct prediction assumed a priori probability. Table 6 presents the average discriminant values and relative importance of the weighted visitor activities for RMS effectiveness. Discriminant analysis results indicate that the relative contribution of website usage to RMS effectiveness varies across the website area and supports different visitor segments in customer value creation. For effective relationship maintenance, long stay is the greatest contributing factor in the products and community area. Customer value can be gained by those visitors who browse in the product area and stay long in the community area because both activities indicate high visitor engagement with the website content and the company’s products. Thus, the active and loyal customers from these activity segments are highly responsive to marketing communication. An effective building of brand equity in the shop area and effective creation of customer satisfaction in the support area are both determined by fact-finding as prominent activity. Thus, customer value is created by visitors who actively search for specific information and possess a structured interaction with the website structure and its content and therefore require less website external support. Table 6 Discriminant analysis results Products (Relationship maintenance) Shop (Building of brand equity) Community (Relationship maintenance) Support (Creation of customer satisfaction) Average discriminant values RMS effectiveness -2.68 -2.24 -4.19 -1.31 RMS failure 3.29 2.52 1.20 3.91 Standardized discriminant coefficients of relevance weighted activities for RMS effectiveness Long stay .17 .88 Browsing .65 .46 Fact-finding .09 .36 .12 .71 Download .11 .08 .09 .15 Purchase .30 Integrating the companies understanding of visitor activity relevance and the empirical contribution of these activities to RMS effectiveness, we nevertheless identify a gap between both, as shown in Table 7. These gaps finally provide the basis of our website optimization recommendations, given to the business experts in step 6 of Page 15/19 our evaluation approach. For an effective relationship maintenance three optimization factors occur that drastically influence customer value in terms of growing responsiveness to marketing communication, conversion and decrease of external support cost. First, long stay as indicator of high visitor engagement underperforms in the product area, although it is defined as an important effectiveness RMS effectiveness contribution factor. Second, also fact-finding underperforms in the product area, indicating difficulties accessing relevant content. Third, download provision also slightly underperforms in the community area. Hence, for increasing relationship maintenance in these areas, website optimization should engage in providing content that enables a higher visitor engagement and matches the visitors information needs, as well as structurally improving access to requested information and download libraries. With regard to an effective building of brand equity the website seriously underperforms in terms of all commerce related transactions in the shop. Consequently, sales and downloads need to be increased by deploying transaction-enhancing mechanisms, usually referring to prices, discounts and other product characteristics. For an effective creation of customer satisfaction based on effective support and service provision, the website generally underperformances in terms of enabling the visitors’ fact-finding activities. Optimization should engage here in structurally improving access to the searched information. Table 7 Visitor activity contribution-relevance website performance gap Products (Relationship maintenance) Shop (Building of brand equity) Community (Relationship maintenance) Support (Creation of customer satisfaction) Performance gap in visitor activities Long stay -.33 .38 Browsing .45 .16 Fact-finding -.11 .26 .02 -.19 Download .1 -.02 -.01 .05 Purchase -.50 5. Conclusion and further research In this study we provide an effectiveness evaluation instrument for corporate websites that can be integrated with internal business measures and is substantial for marketing decisions. This approach overcomes the shortcomings of relationship marketing evaluations using clickstream data which are difficult to relate to customer value. Instead, our approach quantifies the effectiveness of a company’s relationship marketing (RM) effort in its online performance empirically based on customer/visitor website usage. This allowed us further to Page 16/19 determine customer value in two steps: firstly, by calculating the moderating degree of the corporate website for RM strategy effectiveness, which reflects the company’s ability of serving customer information needs online; and secondly, by discussing involved relationship economics. Methodically, our evaluation approach aligns marketing analysis with web mining approaches. Results of our evaluation procedure support customer value determination and also enable a demand-actuated website optimization and customization for different relational marketing strategies. The developed, marketing and IS research uniting methodology addresses both practitioners and researchers, as we developed a rigor methodological approach for website performance evaluation that is also highly applicable. Our approach comprises six steps conducted in three phases by domain experts and analysts. Its clear structure enables guidance regarding the complex evaluation procedure by allocating tasks, and supporting communication between both stakeholder groups. We illustrated our approach by evaluating the website performance of a software developer. Here, we gave a detailed description of all involved methodological procedures including web data preparation, statistical analysis and effectiveness assessment. Finally, our results are used to derive managerial implications in order to improve company’s relationship marketing. For researchers, our findings can be used as a starting point regarding the determination of customer value in company’s RM. Additional case studies should be conducted to investigate the benefits of our approach. Furthermore, future applications based on this methodology could be developed supporting online analytical processing by providing visitor model development, identification and classification. The main drawback of our developed method is the characterization and discrimination of visitor activities by a business expert, which impacts the quality and quantity of the activities modeled in the subsequent classification procedure. Further, the selection and weighting of RMS relevant visitor activities may not be adequate and exhaustive in a company’s multi-channel marketing context. The extent of this problem can be limited by increasing external reliability. Here, inter-rater reliability could be enhanced by including several business experts in the qualitative procedures (step 1, 3), which allows measuring the degree of agreement in their specifications, i.e. by using Cohen’s kappa. Then, the inferential approach of modeling visitor behavior itself is limited and may not be representative for all customer segments, which differ in information needs and thus, usage behavior. Modeling and inferring customer segments prior to analysis could improve general results. Furthermore, the developed methodology has been applied on historical data only and needs to become extended, i.e. for integrating seasonal changes in website content, structure and possible changes in visitor behavior for enabling longitudinal evaluations of website performance. Eventually, the mentioned example only Page 17/19 illustrates our approach. It does not enable an entire evaluation of our developed approach. This should be the aim of further publications in this domain. Acknowledgements Masked for review. Appendix Table 8 Data models for visitor activity “browsing” Products Community C&RT model 1 Dependent variable Visit actions > 3.5 > 3.5 Session duration navigational objects < .5 < .5 Independent variable Session duration content objects > 12.5 > 39.5 Model evaluation s .02 .10 C&RT model 2 Dependent variable Session duration > 15.5 > 32.5 Independent variable Visit actions navigational objects < 1 < 1 Visit actions content objects > 2.5 < 4 Model evaluation s 4.34 1836.77 Table 9 Data models for visitor activity “fact-finding” Products Shop Comm- unity Support C&RT model 1 Dependent variable Visit actions < 3.5 < 3.5 < 3.5 < 3.5 Session duration navigational objects < .5 < .5 < .5 < .5 Independent variable Session duration content objects > 2.5 > 39.5 > 31.5 > 14.5 Model evaluation s .01 .03 .10 .02 C&RT model 2 Dependent variable Independent variable Session duration < 10 < 8 < 15 < 4 Visit actions navigational objects < 1 < 1 < 3 < .5 Page 18/19 Visit actions content objects < 5 < 4 < 3 <5.5 Model evaluation s 4.33 167.44 1790.04 7196.89 References List Aaker, D. (1993). Are brand equity investments really worthwhile? In D. A. Aaker & A. Biel (Eds.), Brand equity and advertising: Advertising's role in building strong brands (pp. 333-341). Hillsdale, NJ: Erlbaum. Argyriou, E., Kitchen, P. J., & Melewar, T. C. (2006). The relationship between corporate websites and brand equity: A conceptual framework and research agenda. International Journal of Market Research, 48, 575-599. Asubonteng, K., McCleary, K. J., & Swan, J. E. (1996). SERVQUAL revisited: A critical review of service quality. The Journal of Services Marketing, 10, 62-81. Ayanso, A., & Yoogalingam, R. (2009). Profiling retail website functionalities and conversion rates: A cluster analysis. International Journal of Electronic Commerce, 14, 79-114. Booth, D., & Jansen, B. J. (2009). A review of methodologies for analyzing websites. In B. J. Jansen, A. Spink & I. Taksa (Eds.), Handbook of research on web log analysis (pp. 141-163). Hershey, NY: IGI Global. BowenCraggs & Co. (2011). Index of corporate web effectiveness. In. Breiman, L., Friedman, J., Stone, C. J., & Olshen, R. (1984). Classification and regression trees. Wadsworth, Belmont: Chapman & Hall. Brown, S. W., & Swartz, T. (1989). A gap analysis of professional service quality. Journal of Marketing, 53, 92- 98. Büchner, A. G., & Mulvenna, M. D. (1998). Discovering Internet marketing intelligence through online analytical web usage mining. SIGMOD Record, 27, 54-61. Bucklin, R. E., & Sismeiro, C. (2003). A model of web site browsing behavior estimated on clickstream data. Journal of Marketing Research, 40, 249-267. Carman, J. M. (1990). Consumer perceptions of service quality: An assessment of the SERVQUAL dimensions. Journal of Retailing, 66, 33-55. Chi, S., & Tavella, D. (2008). Data mining and market intelligence for optimal marketing returns. Oxford, UK: Elsevier. Cho, Y. H., Kim, J. K., & Kim, S. H. (2002). A personalized recommender system based on web usage mining and decision tree induction. Expert Systems with Applications, 23, 329–342. Chou, P.-H., Li, P.-H., Chen, K.-K., & Wua, M.-J. (2010). Integrating web mining and neural network for personalized e-commerce automatic service. Expert Systems with Applications, 37 2898–2910. Chou, W.-C., & Cheng, Y.-P. (2012). A hybrid fuzzy MCDM approach for evaluating website quality of professional accounting firms. Expert Systems with Applications, 39, 2783–2793. Cooley, R., Mobasher, B., & Srivastava, J. (1999). Data preparation for mining world wide web browsing patterns. Journal of Knowledge and Information Systems, 1, 5-32. Financial Times/Bowen Craggs & Co. (2011). Financial Times Bowen Craggs Index 2011. In. Fox, S., Karnawat, K., Mydlands, M., Dumais, S., & White, T. (2005). Evaluating implicit measures to improve web search. ACM Transactions on Information Systems, 23, 147-168. Gale, B. (1994). Managing Customer Value: Creating Quality and Service that Customers Can See. New York, NY: The Free Press. Hahn, J., & Kauffman, R. J. (2004). A methodology for business value-driven website evaluation: A data envelopment analysis approach. In Proceedings of the Third Annual Workshop on HCI Research in MIS (pp. 45-49). Washington, D.C. Hochsztain, E., Millán, S., Pardo, B., Pena, J. M., & Menasalvas, E. (2003). A framework to integrate business goals in web usage mining. In E. Menasalvas, J. Segovia & P. S. Szczepaniak (Eds.), Advances in Web Intelligence: First International Atlantic Web Intelligence Conference (AWIC 2003) (pp. 28-45). Berlin, New York: Springer. Kalczynski, P. J., Senecal, S., & Nantel, J. (2006). Predicting On-Line Task Completion with Clickstream Complexity Measures: A Graph-Based Approach. International Journal of Electronic Commerce, 10, 121-141. Kellar, M., Watters, C., & Shepherd, M. (2006). A goal-based classification of web information tasks. Proceedings of the American Society for Information Science and Technology, 43, 1-22. Kim, C., Kwon, K., & Chang, W. (2011). How to measure the effectiveness of online advertising in online marketplaces. Expert Systems with Applications, 38 4234–4243. Kim, Y. S., & Yum, B.-J. (2011). Recommender system based on click stream data using association rule mining. Expert Systems with Applications, 38 13320–13327. Page 19/19 Kohavi, R., Rothleder, N. J., & Simoudis, E. (2002). Emerging trends in business analytics. Communications of the ACM, 45, 45-48. Larose, D. T. (2005). Discovering knowledge in data: An introduction to data mining. New York, NY: Wiley. Lee, J., Hoch, R., Podlasek, M., Schonberg, E., & Gomory, S. (1999). Analysis and visualization of metrics for online merchandising, web usage analysis and user profiling. In Proceedings of the International WEBKDD’99 Workshop (pp. 126-141). San Diego, USA. Lee, J., Podlaseck, M., Schonberg, E., & Hoch, R. (2001). Visualization and Analysis of Clickstream Data of Online Stores for Understanding Web Merchandising. Data Mining and Knowledge Discovery, 5, 59- 84. Lee, Y. W., Strong, D. M., Kahn, B. V., & Wang, R. Y. (2002). AIMQ: a methodology for information quality assessment. Information and Management, 40, 133-146. Markov, Z., & Larose, D. T. (2007). Data mining the web: Uncovering patterns in web content, structure, and usage. New York, NY: Wiley. McKinney, V., Yoon, K., & Zahedi, F. M. (2002). The measurement of web-customer satisfaction: An expectation and disconfirmation approach. Information Systems Research, 13, 296-315. Morgan, R. M., & Hunt, S. D. (1994). The commitment-trust theory of relationship marketing. The Journal of Marketing, 58, 20-38. Olbrich, R., & Holsing, C. (Winter 2011-12). Modeling Consumer Purchasing Behavior in Social Shopping Communities with Clickstream Data. International Journal of Electronic Commerce, 16, 15-40. Palmatier, R. W., Dant, R. P., Grwal, D., & Evans, K. R. (2006). Factors influencing the effectiveness of relationship marketing: A meta-analysis. Journal of Marketing, 70, 136-153. Parasuraman, A., Zeithaml, V., & Berry, L. (1988). SERVQUAL: A Multiple-Item Scale for Measuring Consumer Perceptions of Service Quality. Journal of Retailing, 64, 12-40. Pyle, D. (1999). Data preparation for data mining (3rd ed.). San Francisco: Morgan Kaufmann. Reinartz, W. J., & Kumar, V. (2003). The impact of customer relationship characteristics on profitable lifetime duration. Journal of Marketing, 67, 77–99. Sen, A., Dacin, P. A., & Pattichis, C. (2006). Current trends in web data analytics. Communications of the ACM, 49, 85-91. Senecal, S., Kalczynski, P. W., & Nantel, J. (2005). Consumer’s decision-making process and their online shopping behavior: A clickstream analysis. Journal of Business Research, 58, 1599-1608. Sheth, J., & Parvatiyar, A. (2000). Relationship marketing in consumer markets: Antecedents and consequences. In J. Sheth & A. Parvatiyar (Eds.), Handbook of Relationship Marketing. Thousand Oaks, CA: Sage. Spiliopoulou, M., & Pohle, C. (2001). Data mining for measuring and improving the success of web sites. Data Mining and Knowledge Discovery, 5, 85-114. Srinivasan, R., & Moorman, C. (2005). Strategic firm commitments and rewards for customer relationship management in online retailing. Journal of Marketing, 69, 193-200. Stolz, C., Viermetz, M., Neuneier, R., & Skubacz, M. (2005). Web performance indicator by implicit user feedback – applications and formal approach. In Proceedings of the sixth International Conference of Web Information Systems Engineering (WISE 2005) (pp. 689-700). New York, NY. Stuart, H., & Jones, C. (2004). Corporate branding in marketspace. Corporate Reputation Review, 7, 84-93. Teas, R. K. (1993). Expectations, performance evaluation, and consumer's perception of quality. The Journal of Marketing, 57, 18-34. Teltzrow, M., & Berendt, B. (2003). Web-usage-based success metrics for multi-channel businesses. In Proceedings of the ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (WEBKDD Workshop on Web Mining 2003) (pp. 17-27). Washington, D.C. Tsai, W.-H., Chou, W.-C., & Leu, J.-D. (2011). An effectiveness evaluation model for the web-based marketing of the airline industry. Expert Systems with Applications, 38, 15499–15516. Vercellis, C. (2009). Business intelligence: Data mining and optimization for decision making. New York, NY: Wiley. Wang, K., Wang, E. T. G., & Farn, C.-K. (2009). Influence of Web Advertising Strategies, Consumer Goal- Directedness, and Consumer Involvement on Web Advertising Effectiveness. International Journal of Electronic Commerce, 13, 67-95. Weinberg, B. D., Parise, S., & Guinan, P. J. (2007). Multichannel marketing: Mindset and program development. Business Horizons, 50, 385-394. Weischedel, B., Matear, S., & Deans, K. R. (2005). The use of emetrics in strategic marketing decisions: A preliminary investigation. Journal of Internet Marketing and Advertising, 2, 109-125. Winter, S. J., Saunders, C., & Hart, P. (2003). Electronic window dressing: Impression management with websites. European Journal of Information Systems, 12, 309-322. Woodruff, R. B. (1997). Customer value: The next source for competitive advantage. Journal of the Academy of Marketing Science, 25, 139-153. 
work_uskxb65xc5ajbkgwcoetuvztf4	----	PDF hosted at the Radboud Repository of the Radboud University Nijmegen The following full text is a publisher's version. For additional information about this publication click this link. http://hdl.handle.net/2066/112337 Please be advised that this information was generated on 2021-04-06 and may be subject to change. http://hdl.handle.net/2066/112337 Mikrochim. Acta [Wien] 1991, II, 493-503 Mikrochimica Acta �9 by Springer-Verlag 1991 Printed in Austria Expert Systems for Method Development and Validation in HPLC M a r y M u l h o l l a n d 1 '*, N. W a l k e r 1, J. A. v a n Leeuwen, L u i t g a r d Buydens 2 , F. Maris 3 , H. H i n d r i k s 3, a n d Peter J. S c h o e n m a k e r s 4 1 Philips Scientific, Cambridge CB1 2PQ, UK 2 University of Nijmegen, Nijmegen, The Netherlands 30rganon International, Oss, The Netherlands 4 Philips Research Laboratories, NL-5600 JA Eindhoven, The Netherlands Abstract. ESCA, Expert Systems Applied to Chemical Analysis, started its research in M a r c h 1987, with the aim of building p r o t o t y p e expert systems for H P L C m e t h o d development. Results of this research have been published as the w o r k has progressed. T h e project is n o w c o m p l e t e d a n d this p a p e r summarises s o m e o f the overall project conclusions. Seven different expert systems have been built which tackle p r o b l e m s t h r o u g h o u t the process of m e t h o d development, four s t a n d - a l o n e systems a n d three integrated systems. T h e object of ESCA was to evaluate the applicability of expert system t e c h n o l o g y to analytical chemistry a n d n o t all the systems were built for c o m m e r c i a l uses. M a n y of the systems tackle p r o b l e m s specific to one o r m o r e of the p a r t n e r s a n d t h u s m a y n o t be useful outside this e n v i r o n m e n t . However, the results of the w o r k are still p e r t i n e n t to analysts wishing to build their o w n systems. These results are described, however, the emphasis of the p a p e r is o n those systems developed for m e t h o d validation. M e t h o d validation for H P L C is a complex task which requires m a n y charac- teristics o f the m e t h o d to be tested, e.g. accuracy, precision, etc. T h e expert systems built within E S C A c o n c e r n the validation of precision. T w o systems were d e v e l o p e d for repeatability testing a n d ruggedness testing. T h e m e t h o d validation process can be divided into several discrete stages, these include: (1) T h e selection o f the m e t h o d feature to test, for instance which factors can influence the ruggedness of a m e t h o d . (2) The definition o f a test procedure, for instance a n efficient statistical design. (3) T h e execution of experiments a n d the i n t e r p r e t a t i o n of results. (4) A diagnosis o f any observed problem. This p a p e r describes these two systems in some detail a n d summarises s o m e of the results o b t a i n e d f r o m their evaluation. It concludes t h a t expert systems can be useful in solving analytical p r o b l e m s a n d the i n t e g r a t i o n of several expert systems can p r o v i d e extremely powerful tools for the analyst. * To whom correspondence should be addressed: Department of Chemistry, The University of New South Wales, Sydney, NSW 2033, Australia 494 M. Mulholland et al. Key words: expert systems, m e t h o d development, m e t h o d validation, H P L C . ESCA started its research in March 1987, with the aim of building prototype expert systems for H P L C m e t h o d development and validation. The starting point of the project involved two separate tasks. Firstly, expert system development software was evaluated to enable the project to select the best available packages for our application [1]. Secondly, the chemical application area had to be defined [2, 3]. Initially, four separate expert system prototypes were built using various expert system development software. The first system tackled the problem of selecting a best starting column and mobile phase for the H P L C of central nervous system drugs [4]. The second system uses expertise to r e c o m m e n d the best criterion for the selectivity optimisation [5]. The third system optimises the physical parameters of an H P L C method, e.g. flow rate and column dimensions [6]. The final system tackled the problem of ruggedness testing H P L C methods [7]. As the project progressed it became apparent that although each of the four systems performed well in its application, analysts require a combination of these tasks. Therefore, the expert systems were integrated in such a way that communica- tion between the various stages of m e t h o d development was possible. Three possible integrations were proposed, an integration of the stages of m e t h o d development, an integration of the ruggedness test with system optimisa- tion and an integration of repeatability with system optimisation. The purpose of this paper is firstly to summarise the objectives of each inte- gration and to describe the results of testing these systems. Each team is preparing publication of the detailed results from the evaluation of the individual systems. However, some conclusions were c o m m o n t h r o u g h o u t the work. It is the aim of this paper to describe these as they reflect both the successes and limitations of current expert system technology. Integration of the Method Development Stages The first step in the definition of an integrated system for H P L C m e t h o d devel- opment was to identify areas of knowledge which were missing from the existing prototype expert systems. The existing prototype for column and mobile phase selection only concerned the specific application of central nervous system (CNS) drugs. In order to widen the scope of this system two other applications were added. Prior to the ESCA project the Free University of Brussels had developed an expert system which could define methods for Label Claim Analysis. This involved devel- oping methods for a very wide range of drug formulations to ascertain the correctly labelled dosage content [8]. This system was implemented in the same expert system shell as the CNS drugs prototype and thus they could be combined in an integrated system. A second system was also added which could refine methods taken from the literature. The addition of these two modules considerably expanded the application scope of the first stage in the m e t h o d development process. In order to combine the first guess systems with a selectivity optimisation stage it was necessary to build an extra system which optimised the retention range of the sample components. Expert Systems for Method Development ESCA Expert Systems Applied to Chemical Analysis 495 First Guess 1 Retention Optimisation Selectivity Optimisation CNS Actives Label Claim Literature System Optimisation I Flow Rate Column Flow Cell Fig. 1. The integrated system for method development Selection of Criteria Solvent Optimisation pH Optimisation The selectivity o p t i m i s a t i o n stage was e x p a n d e d by including m o d u l e s for the o p t i m i s a t i o n o f solvent c o m p o s i t i o n a n d pH. A simplex p r o c e d u r e was used to optimise solvents a n d the D o e h l e r t design was used to optimise pH. Finally, the system o p t i m i s a t i o n p r o t o t y p e c o u l d be integrated to define the physical p a r a m e t e r s of the m e t h o d . These systems were integrated such t h a t c o m m u n i c a t i o n between the stages was possible via a supervisor system with a c o m m o n d a t a base [9]. This system is illustrated in Fig. 1. Integration of the Method Validation Sta#es T h e aim of validating a m e t h o d is to identify any sources of error in the m e t h o d a n d to check w h e t h e r these errors are within acceptable ranges. However, sometimes a m e t h o d can s h o w errors which are unacceptable for its application. It is at this stage in the validation process w h e n a link to a re-optimisation m o d u l e is required. T o p r o v i d e this link t h e ruggedness test system was integrated to the system o p t i m i s a t i o n p r o g r a m . Earlier in the project a small expert system for repeatability testing was built [10]. This system was used as a starting p o i n t for a third integrated system which c o u l d c o m b i n e a repeatability test with the system o p t i m i s a t i o n p r o g r a m . F o r b o t h these systems the process of the validation was divided into four stages, which are illustrated in Fig. 2. T h e first stage required the definition o f the character- istic to test. F o r b o t h repeatability a n d ruggedness testing this was precision. N e x t the test was defined as either ruggedness or repeatability a n d a n experimental p r o c e d u r e was recommended. Finally the results are processed in a diagnosis m o d u l e . This included a pass/fail criterion, a m e t h o d for interpreting a n d curing any p r o b l e m s a n d a m e a n s to re-optimise a failed m e t h o d . T h e i n t e g r a t i o n o f the system 496 M. Mulholland et al. Method Validation Define Characteristic Accuracy Precision Sensitivity Define test Repeatability Ruggedness Experimental Diagnosis Pass/Fail Cure Problem Re-optimise Method Fig. 2. The four stages of method validation Repeatability Reduced test Re-optimise ~ [ Method feature 1 Test Design 1 Test Measure 1 Diagnosis ~" [ Sample Preparation 'I in oction I 'ISing'oPoiotl Concentration Range lOut.ers Ori t I Fig. 3. The repeatability text expert system o p t i m i s a t i o n p r o g r a m in the d i a g n o s i s m o d u l e a l l o w e d t h e r e s o l u t i o n o f a critical p a i r o f p e a k s t o be i m p r o v e d . Fig. 3 illustrates the r e p e a t a b i l i t y test i n t e g r a t e d system. T h e m e t h o d f e a t u r e t o b e t e s t e d for r e p e a t a b i l i t y c o u l d be t h e s a m p l e p r e p a r a t i o n o r t h e i n j e c t i o n p r o c e d u r e . T h e e x p e r i m e n t a l d e s i g n p r o p o s e d c o u l d test t h e r e p e a t a b i l i t y across a c o n c e n t r a t i o n r a n g e o r at a single c o n c e n t r a t i o n p o i n t . T h e m e a s u r e m e n t s m a d e w e r e for t h e relative s t a n d a r d d e v i a t i o n o f c o n c e n t r a t i o n , p e a k heights, r e t e n t i o n t i m e s a n d p e a k areas. T h e s e results c o u l d t h e n be c h e c k e d for a n y o u t l i e r s o r drift. T h e d i a g n o s i s m o d u l e t h e n c o u l d identify a n y p r o b l e m s a n d s u g g e s t a cure. W h e n t h e p r o b l e m was d u e t o i n a d e q u a t e r e s o l u t i o n b e t w e e n a p a i r o f p e a k s , t h e s y s t e m Expert Systems for Method Development Ruggedness Test 497 ) Method Feature I Flow rate ) Column Manufacturer Solvent Test Design I ) Factorial Design I Test Measure ] ) Standard Errors Main Effects Interaction Effects t ~176 I IOia I Report Fig. 4. The ruggedness test expert system Example Diagnosis If column temperature causes a large effect on retention times Then conclude recommend frequent recalibration If a factor causes loss of resolution Then conclude consult system optimisation to increase resolution Fig. 5. An example diagnosis rule r o u t e d t o t h e r e - o p t i m i s a t i o n m o d u l e . T h e n e w m e t h o d was t h e n r e - t e s t e d for r e p e a t a b i l i t y . T h e r u g g e d n e s s test is i l l u s t r a t e d in a similar w a y in Fig. 4. A r u g g e d n e s s test n o r m a l l y r e q u i r e d t h e t e s t i n g o f b e t w e e n t w o a n d t e n f e a t u r e s o f t h e m e t h o d . T h e s e w e r e c h o s e n b y c o n s i d e r i n g t h e i r p o t e n t i a l influence o n t h e r u g g e d n e s s o f t h e m e t h o d . F a c t o r i a l d e s i g n s w e r e e m p l o y e d t o test t h e effect o f c h a n g i n g t h e s e m e t h o d features, a n d s t a n d a r d errors, m a i n effects a n d i n t e r a c t i o n effects w e r e m e a s u r e d . T h e d i a g n o s i s m o d u l e t h e n c h e c k e d these m e a s u r e m e n t s a g a i n s t pass/fail criteria a n d identified p o t e n t i a l p r o b l e m s . W h e n r e q u i r e d it r o u t e d t o t h e s y s t e m o p t i m i s a - t i o n m o d u l e t o i m p r o v e t h e r e s o l u t i o n . A n e x a m p l e d i a g n o s i s rule is s h o w n in Fig. 5. Evaluation of the Validation Expert Systems T h e e x p e r t s y s t e m s r e q u i r e d t w o stages o f testing, v a l i d a t i o n a n d e v a l u a t i o n . V a l i d a - t i o n o f t h e s o f t w a r e w a s p e r f o r m e d b y t h e i m m e d i a t e t e a m m e m b e r s a n d t h e o b j e c t 498 M. Mulholland et al. was to test the correct operation of the software and to check that the decisions m a d e by the expert system agree with those of the expert. Evaluation was performed by other team members, who were not directly involved with the development of the software, and by third parties external to ESCA. The objective of the evaluation was to determine how well the expert system performed in practice. The validation of the software involved checking the system for logical reasoning and testing for bugs. Any inconsistencies or bugs could then be removed at this stage. The expert then selected a n u m b e r of problems which demonstrated a variety of situations. The expert predicted the answers, and then tested the problem on the software. If the expert and expert system disagreed the reason needed to be identified and a solution found. Knowledge could be incorrect in which case it should be corrected. Knowledge could be missing which then should be added. In this way, the validation process showed whether the systems provided g o o d answers to problems falling within the intended scope of expertise. In order to validate the ruggedness test expert system, the expert selected ten H P L C methods representing a variety of pharmaceutical applications. The expert selected the factors to test and a statistical design to perform the experiments. These methods were then applied to the expert system and the answers compared. Several modifications and additions were then made to the software until the agreement between consultations was acceptable. The expert and expert system always agreed on the factors to be tested within a difference of two factors, i.e. there were never m o r e than two factors over which the expert and expert system did not agree. W h e n the software was successfully validated and its performance was con- sidered sufficiently consistent with that of the expert, it could then be given to external evaluators. The evaluation then involved putting problems to the systems in a practical laboratory environment. Two types of evaluators could be dis- tinguished, those who were experts themselves in the application area and those who had H P L C experience but did not have the specific expertise contained in the software. It was i m p o r t a n t to involve both types of evaluator to assess both the quality and overall usefulness of the advice given by the expert systems. A list of evaluation criteria was proposed by the knowledge engineers and experts within ESCA. This list is shown in Table 1. The evaluators were then allowed to select relevant criteria from this list to test the software against. The list consisted of three types of criteria, these concerned the user interface, the consistency of advice and the limitations of the system. These criteria allowed the expert evaluators to contribute to identifying the system limits, whereas the non-experts could evaluate the ease of use and consistency of the software. The time scales of the evaluations varied from a couple of weeks to over a month, so some evaluations were more t h o r o u g h than others. D u e to the fact that the repeatability system was built at such a late stage in the project there was some overlap between the validation and evaluation of this package. Results o f the Evaluation The Repeatability Expert System This system was evaluated in two laboratories, at O r g a n o n International and Philips Research in Eindhoven. Expert Systems for Method Development Table 1. List of evaluation criteria Man machine interface (user interface) Choice of phrases Explanation Operation (mouse, keyboard, etc.) Ease of use Consistency testin9 Accuracy (correct answers, quality of advice) Robustness (does the system crash or lock-up) Reproducibility (same input, same output) System limits Conflict (do answers provide conflicting advice) Missing knowledge Strange or inexplicable advice Technical content 499 T h e e v a l u a t i o n at O r g a n o n w a s c a r r i e d o u t in t h r e e s e p a r a t e ways: 1. M e t h o d d e v e l o p m e n t a n d v a l i d a t i o n o f n e w H P L C m e t h o d s . 2. R e p e a t a b i l i t y testing o f v a l i d a t e d m e t h o d s . 3. T r o u b l e s h o o t i n g . F o r e a c h t y p e o f a p p l i c a t i o n , different e x p e r i m e n t s w e r e selected t o p e r f o r m the e v a l u a t i o n a n d to d e m o n s t r a t e the a p p l i c a b i l i t y o f this s y s t e m t o the p h a r m a c e u t i c a l industry. A full d e s c r i p t i o n o f this e v a l u a t i o n a n d its results is c u r r e n t l y in p r e p a r a - tion for p u b l i c a t i o n . H o w e v e r , a s u m m a r y o f t h e p r o b l e m s identified a n d the s u b s e q u e n t a c t i o n t a k e n is s h o w n in T a b l e 2. This s h o w s the t y p e o f p r o b l e m s w h i c h w e r e identified, it also d e m o n s t r a t e s h o w t h e v a l i d a t i o n o f the s o f t w a r e h a d t o b e c o m b i n e d w i t h the e v a l u a t i o n . Ideally, p r o b l e m s such as i n c o r r e c t c a l c u l a t i o n s s h o u l d h a v e b e e n identified a n d fixed d u r i n g a s e p a r a t e v a l i d a t i o n stage. T h e s e c o n d e v a l u a t i o n o f t h e r e p e a t a b i l i t y s y s t e m a p p l i e d t h e s o f t w a r e t o a n H P L C m e t h o d for t h e d e t e r m i n a t i o n o f caffeine in coffee. T h e m e t h o d w a s tested at t w o different injection volumes. I n the p r o c e s s o f this e v a l u a t i o n several limitations o f the s y s t e m were high- lighted. Table 2. Summary of some of the problems observed Problem description Action Usage requirement unclear Inaccurate retention times Inaccurate RSD calculation Poor file handling No outlier test Drifting baseline not observed Text altered Calculation corrected Calculation corrected Load options added Outlier test added In study 500 M. M u l h o l l a n d et al. (1) The diagnosis o f p r o b l e m s by the system was n o t completely correct. How- ever, the evaluators u n d e r s t o o d t h a t this was p r o b a b l y inevitable, because a real expert was also expected to c o m e u p with several possible causes for a single observed problem. Hence the criteria used by the evaluators were (a) w h e t h e r the suggestions were sensible a n d (b) w h e t h e r the real cause o f the p r o b l e m was a m o n g those suggested. P u t in this perspective, the diagnosis m o d u l e was considered to function correctly. (2) A design editor which w o u l d allow changes to the suggested experiments was requested. (3) F o r the system to be used in a practical e n v i r o n m e n t it w o u l d need to be linked to a c h r o m a t o g r a p h y d a t a station. (4) T h e system c a n n o t be used by c h r o m a t o g r a p h e r s w i t h o u t a certain a m o u n t of knowledge. It was n o t a tool to educate novices. D u r i n g the evaluation of this software m a n y of the identified p r o b l e m s were rectified a n d requested additions were made. This d e m o n s t r a t e d the c o n t r i b u t i o n m a d e to the overall value of the software due to the evaluation process. The Ruggedness Test Expert System The ruggedness test system was evaluated in several p h a r m a c e u t i c a l laboratories. The detailed results will be published in a separate paper. The evaluations were divided into three stages: 1. The i n p u t of a variety of m e t h o d s to the system. 2. A critical evaluation of the factors selected. 3. T h e collection o f d a t a a n d a n evaluation o f the diagnosis. The p u r p o s e of applying several different m e t h o d s to the expert system was to allow the evaluators to find any m e t h o d features which could n o t be defined by the system. Also any missing k n o w l e d g e could be identified. Table 3 summarises some of the results from this stage of testing. T h e first c o l u m n lists the m e t h o d features where k n o w l e d g e was missing. T h e second c o l u m n lists s o m e of the m e t h o d features which c o u l d n o t be defined in the i n p u t module. It was recognised that to some extent this was inevitable as H P L C is very complex a n d it was difficult to p r o g r a m all the available expertise. However, it was r e c o m m e n d e d t h a t m o d u l e s be a d d e d to the system at a later date with k n o w l e d g e o n o t h e r types of detection. The rugged- ness test system did allow the user to use his o w n k n o w l e d g e to select factors when necessary. This was invaluable in these cases where knowledge was missing. Table 3. Results on the input of methods Missing knowledge M e t h o d features not defined Fluorescence detection Ion c h r o m a t o g r a p h y G r a d i e n t methods Chiral c h r o m a t o g r a p h y D u a l detection Solid phase extraction Both sonicate and shake Sample solvent composition Expert Systems for Method Development 501 T h e selection o f factors m o d u l e was critically evaluated by c o m p a r i n g the factors selected with the o p i n i o n s of the evaluator. T h e following list shows examples o f the c o m m e n t s made: 1. T h e selection o f centrifuge time was n o t necessary. 2. I n s o m e instances the levels selected for solvent mix were t o o extreme. 3. C o l u m n b a t c h variation was often preferred to c o l u m n m a n u f a c t u r e r variation. T h e selection of factors m o d u l e was very flexible a n d any changes in factors o r levels required by the user could be made. The analysts used this system in a completely different m a n n e r to a c o n v e n t i o n a l software p r o g r a m . T h e y e x a m i n e d the advice a n d the explanations a n d accepted o r rejected suggestions as they required. This is exactly the same as a typical c o n s u l t a t i o n with an expert a n d it e m p h a s i z e d the i m p o r t a n c e of flexibility in expert systems. It was n o t always possible to o v e r c o m e conflicts o f o p i n i o n between evaluators a n d expert. F o r instance, some evaluators felt restricted by the allowed statistical designs. However, the expert felt t h a t t o o m u c h flexibility c o u l d lead to errors. T h e diagnosis m o d u l e of the system was also f o u n d to have missing k n o w l e d g e a n d some differences in o p i n i o n were noted. T h e following list summarises some c o m m e n t s : 1. T h e m a i n effects were n o t tested o n their statistical significance. 2. T h e pass/fail levels were sometimes t o o strict. 3. T h e h a n d l i n g o f internal s t a n d a r d m e t h o d s was inadequate. 4. The system suitability criteria did n o t include noise. T o s o m e extent the flexibility of the system c o u l d o v e r c o m e these problems. However, it was clear t h a t a d d i t i o n a l k n o w l e d g e is required, particularly for h a n d l i n g internal s t a n d a r d s a n d noise m e a s u r e m e n t s . Conclusions The Repeatability Test Expert System It was clear t h a t there was insufficient time to evaluate the c o m p l e t e system a n d especially the link f r o m the diagnosis m o d u l e to the system o p t i m i s a t i o n module. This p a r t o f the system w o u l d only be c o n s u l t e d if the d i a g n o s e d p r o b l e m was d u e to i n a d e q u a t e resolution. This d i d n o t occur for any o f the test cases. F o r this reason, it was difficult to c o n c l u d e o n the usefulness of the integrated expert system. A l t h o u g h the evaluators d i d n o t use the r e - o p t i m i s a t i o n m o d u l e , they agreed t h a t it could p r o v e invaluable for the relatively few times it w o u l d be required. Initially, the software h a d several p r o b l e m s with flexibility a n d missing k n o w l - edge, b u t these were solved in later u p d a t e s o f the software. This process o f evalua- tion, r e c o m m e n d a t i o n s a n d i m p r o v e m e n t s was invaluable to the quality o f the final software. The user interface o f this p a c k a g e was considered excellent, with g o o d edit a n d explain facilities. The quality of advice was g o o d a n d it c o u l d deal with relatively difficult problems. T h e evaluators c o n c l u d e d t h a t the system was of use to analysts with little software knowledge. However, s o m e experience o f H P L C was required a n d training in the use o f the p a c k a g e w o u l d be necessary. 502 M. Mulholland et al. The Ruggedness Test Expert System As for the repeatability system the integrated module for re-optimisation was not fully evaluated. N o n e of the test cases required an improvement in resolution. The evaluation of this expert system clearly illustrated the difference between an expert system and a conventional software package. Analysts were happy to accept advice or modify it when required. This is exactly how a typical consultation with an expert would be. Generally the expert system was found to be useful, particularly for analysts who had not been involved in the m e t h o d development process. There were areas of missing knowledge which should eventually be added to enhance its usefulness. General Conclusions As the ESCA research progressed the team built a total of seven different expert systems and it was concluded that these systems provided solutions to m a n y problems in H P L C m e t h o d development. They were successfully used at each stage of the m e t h o d development process and an integration of these stages provided useful communication links. W h e n the evaluation of these packages began m a n y problems were predicted. This was due to the unique nature of expert systems. They do not give right or wrong answers, but good or bad advice. The quality of this advice would be very difficult to measure. Users are expected to interact with expert systems in a completely different way to other software packages. They need to question the system and confirm or modify advice. In some cases they were required to supplement the system with their own knowledge. In fact, we found that analysts welcomed this kind of interaction and were perfectly capable of critically evaluating the advice given. The overall conclusion was that expert systems do not replace analysts as indeed experts cannot directly replace analysts. However, they provide easy access to expertise which can help them work more efficiently. References [1] J. A. van Leeuwen, B. G. M. Vandeginste, G. Postma, G. Kateman, Chemom. Intel. Lab. Syst. 1989, 6, 239. [2] D. Goulder, T. Blaffert, A. Blokland, L. Buydens, A. Chabra, A. Cleland, N. Dunand, H. Hindriks, G. Kateman, J. A. van Leeuwen, D. Massart, M. Mulholland, G. Musche, P. Naish, A. Peeters, G. Postma, P. Schoenmakers, M. de Smet, B. Vandeginste, J. Vink, Chromatographia 1988, 26, 37. [-3] P.J. Schoenmakers, M. Mulholland, Chromatographia 1988, 25, 737. [4] H. Hindricks, F. Maris, J. Vink, A. Peeters, M. de Srnet, D. L. Massart, L. Buydens, J. Chromatogr. 1989, 485, 255. [5] A. Peeters, L. Buydens, D. L. Massart, P. J. Schoenmakers, Chromatographia 1988, 26, 101. [-6] P.J. Schoenmakers, N. Dunand, A. Cleland, G. Musche, T. Blaffert, Chromatographia 1988, 26, 37. [-7] M. Mulholland, N. Dunand, A. Cleland, J. A. van Leeuwen, B. G. M. Vandeginste, J. Chromatogr. 1989, 485, 283. 1-8] M. de Smet, A. Peeters, L. Buydens, D. L. Massart, J. Chromatogr. 1989, 485, 283. Expert Systems for Method Development 503 [9] P. Conti, H. Piryns, N. Van den Driessche, M. de Smet, T. Hamoir, F. Marls, H. Hindriks, P. Schoenmakers, D. L. Massart, Chromatographia (accepted). [10] M. Mulholland, J. A. van Leeuwen, B. G. M. Vandeginste, Anal. Chim. Acta 1989, 223, 183. Received August 31, 1990. Revision February 12, 1991. 
work_ut27gja5vbcgrlyb27mnoaiofi	----	Formal neuron, Learning process, Recognition American Journal of Intelligent Systems 2015, 5(1): 34-41 DOI: 10.5923/j.ajis.20150501.04 Development of the Learning Process for the Neuron and Neural Network for Pattern Recognition Ramaz Khurodze Department of Artificial Intelligence, Georgian Technical University, Tbilisi, Georgia Abstract The goal of our work is development of a neuron and a neural network which recognize any object without mistakes. It worked out by learning method, with totality of simple (single-layer) neurons. The work deals with the formal neuron and neural network learning process through the realizations of a set of learning sets. At the same time, the features and the feature space used for the recognition process are evaluated. The feature space for all the types and realizations is of the same dimension and binary. The learning process is carried out by means of recognition procedures; this, in case of incorrect recognition, as much as possible, draws together the changes in neuron weighting coefficients as well as the threshold (structuring etalon descriptions) with the neural process of recognition. A set of realizations for the learning clusters of each pattern is used for the learning process. The learning algorithm comprises two stages. The first stage represents structuring etalon description of its own, the second – the correction of the received description in relation to other patterns of descriptions by using the same patterns of the learning set’s realizations. The correction of the results received in the recognition process is carried out by means of changing the weighting coefficients through using the award algorithm (procedure). In case of the incorrect recognition of some realization, it is presented to the neuron until we get the correct recognition through coefficients changing (error correction of mistakes) which may require neuron threshold changing. Keywords Formal neuron, Learning process, Recognition 1. Introduction Elaborating Neural Network has a long history of development. McCulloch and Pitts (1943) [1] developed models of neural networks based on their understanding of neurology. In the late 1940s psychologist Donald Hebb [2] created a hypothesis of learning based on the mechanism of neural plasticity that is now known as Hebbian learning. These ideas started being applied to computational models in 1948 with Turing's B-type machines Farley and Wesley A. Clark [3] (1954) first used computational machines, then called calculators, to simulate a Hebbian network at MIT. Other neural network computational machines were created by Rochester, Holland, Habit, and Duda [4] (1956). Rosenblatt (1958) [5] stirred considerable interest and activity in the field when he designed and developed the Perceptron. The Perceptron had three layers with the middle layer known as the association layer. (Three layers: I layer – output of feature space, II layer – conversion of the output; III layer decision-making). Another system was the ADALINE (ADAptive LInear Element) which was developed in 1960 by Widrow and Hoff (of Stanford * Corresponding author: mariam.chkhaidze@bk.ru (Ramaz Khurodze) Published online at http://journal.sapub.org/ajis Copyright © 2015 Scientific & Academic Publishing. All Rights Reserved University). In 1969 Minsky and Papert wrote a book in which they generalised the limitations of single layer Perceptrons to multilayered systems [7]. Grossberg's (Steve Grossberg and Gail Carpenter in 1988) influence founded a school of thought which explores resonating algorithms. They developed the ART (Adaptive Resonance Theory) networks based on biologically plausible models. Anderson and Kohonen developed associative techniques independent of each other. Klopf (A. Henry Klopf) in 1972, developed a basis for learning in artificial neurons based on a biological principle for neuronal learning called heterostasis. Werbos (Paul Werbos 1974) developed and used the back-propagation learning method, however several years passed before this approach was popularized. Amari (A. Shun-Ichi 1967) was involved with theoretical developments: he published a paper which established a mathematical theory for a learning basis (error-correction method) dealing with adaptive patern classification. While Fukushima (F. Kunihiko) developed a step wise trained multilayered neural network for interpretation of handwritten characters. The original network was published in 1975 and was called the Cognitron. The backpropagation algorithm was created by Paul Werbos [6] (1975). The parallel distributed processing of the mid-1980s became popular under the name connectionism. The text by David E. Rumelhart and James McClelland [8] (1986) provided a full exposition on the use of connectionism in American Journal of Intelligent Systems 2015, 5(1): 34-41 35 computers to simulate neural processes. In the 1990s, neural networks were overtaken in popularity in machine learning by support vector machines and other, much simpler methods such as linear classifiers. Renewed interest in neural nets was sparked in the 2000s by the advent of deep learning. Between 2009 and 2012, the recurrent neural networks and deep feedforward neural networks developed in the research group of Jürgen Schmidhuber at the Swiss AI Lab IDSIA have won eight international competitions in pattern recognition and machine learning. [9] For example, multi-dimensional long short term memory (LSTM) [10] [11] won three competitions in connected handwriting recognition at the 2009 International Conference on Document Analysis and Recognition (ICDAR), without any prior knowledge about the three different languages to be learned. Variants of the back-propagation algorithm as well as unsupervised methods by Geoff Hinton and colleagues at the University of Toronto [12] [13] can be used to train deep, highly nonlinear neural architectures similar to the 1980 Neocognitron by Kunihiko Fukushima, [14] and the "standard architecture of vision", [15] inspired by the simple and complex cells identified by David H. Hubel and Torsten Wiesel in the primary visual cortex. Deep, highly nonlinear neural architectures similar to the 1980 neocognitron by Kunihiko Fukushima [14] and the "standard architecture of vision" [15] can also be pre-trained by unsupervised methods [16] [17] of Geoff Hinton's lab at University of Toronto. A team from this lab won a 2012 contest sponsored by Merck to design software to help find molecules that might lead to new drugs. [18]. Nowadays, there are some scientists working in the field of Artificial neuroscience and making significant contribution to it: Yann LeCun, Pierre Sermanet and more. We are using the terms which are used already in the works: [19][20]. Used Symbols: Set of patterns to be recognized: { } { }1 2, ,.......... ,.......... , IA A A Ai A I Card A= = (1.1) Set of given realizations for each pattern { }X { } 1 2, ,..., ,...,i IX X X X X= ; (1.2) A set of realizations for each pattern consists of learning, testing and unknown realization sets; for example, for {Xi } set of Ai pattern we will have: 1 2, ,..., ,...,i i i mi MiX X X X X= , 1; ; 1;i im M i I= = ;(1.3) Any type of realization represents N dimensional vector with binary components. For example, for Ai type mi realization we will have: { } 1 2; , ,..., ,..., ; 1; ; 1; i i i i im m m m mmi i i i ni Nii iX X X x x x x i I n N ∈ = = = ; (1.4) Spatial dimension for N = Card(xn ) symbols; ∀ {0;1}.imnx = Due to the fact that for different patterns we may have different quantities of learning set realizations, we will have: { }immi niX x= , where 𝑚𝑚𝑖𝑖 = 1; 𝑀𝑀𝑖𝑖������ ; (1.5) We have each neuron weighing coefficient for each sign of the sign space, for example, we will have a set of weighting coefficients {xn } for a {wn } number of symbols: { } 1 2, ,..., ,...,n NW w w w w= ; (1.6) For each Aj member of 𝐴𝐴 set of patterns we will correspondingly have Wi set: { }i niW w= ; (1.7) For each pattern we have one neuron designated as Ne, for example we will have Nei neuron for Ai pattern, which will be represented by its own weighting coefficient and in the form of neuron threshold we will accordingly have sets of Nei neurons and Zi thresholds : {Ne} = Ne1 , Ne2 , … , Nei , … , NeI , (1.8) {Z} = Z1 , Z2 , … , Zi , … , ZI ; (1.9) For the purpose of compactly describing the recognition process through neurons in the given work original symbols have been used; these symbols reflect separate stages of recognition. 1. The introduction (arrival) of unknown realizations for recognition at the neuron entry is expressed by means of the predicate (presentation): “presentation”; The predicate “presentation” implies that an unknown realization with all its components is included into neuron 2. Predicate “is accepted” implies the result obtained through some mathematical operation: ⇒ “ is accepted”. By considering one or two points, the recognition process in neurons can be presented as follows: X Ne{W} ⇒ � XW = net ⇒ F{net} ⇒ Z ⇒ out = {0; 1} ; (1.10) F(net) function is termed as the activation function which is nonlinear as a rule, e.g. sigmoid or hyperbolic tangent, or some other one chosen heuristically. F(. ), ”net” and “out” symbols are widely spread in the corresponding literature, therefore, we don’t change them. However, let us present their mathematical description and function: n nn net x w= ∑ ; (1.11) which represents the coordinate product of the realization and neuron weighting coefficients by means of which the 36 Ramaz Khurodze: Development of the Learning Process for the Neuron and Neural Network for Pattern Recognition modeling of charge accumulation process in soma is performed; out = {0; 1}; (1.12) Through this pattern the description of the decision making process at Ne neuron exit, namely: If out = 1, then X ∈ Aj , if out = 0, then X ∉ Aj Aj ∈ {A}. (1.13) 2. Description of the Correct and Incorrect Recognition Processes As the goal of our work is development of a neural network which recognizes correctly (without mistakes), then at first, explain the meaning of correct or incorrect recognition processes: The result of the recognition process is correct, if the right decision is obtained due to it: the realization to be recognized belongs to “its own” type. The result of the recognition process is incorrect if an incorrect decision is obtained due to it: the realization to be recognized does not belong to “its own” type, or belongs to a different type. For example, if Aj type realization is presented for recognition from Xj learning set, then the recognition is correct if making a decision results in Xj ∈ Aj, (2.1) And the recognition is incorrect if we have: Xi ∉ Ai . Taking into consideration the symbols obtained in the previous chapter and 1,10 patterns, the correct and incorrect recognition processes can be described as follows: Correct recognition Xi Nei ⇒ ∑ Xi Wi = ∑ xni wnin = neti ⇒ F{neti } Zi ⇒ outi = 1 ⇒ Xi ∈ Ai ; (2.2) incorrect recognition: Xi Nei ⇒ ∑ Xi Wi = ∑ xni wnin = neti ⇒ F{neti } Zi ⇒ outi = 0 ⇒ Xi ∉ Ai . (2.3) The process of correct recognition is completely described by 2.2 expression, whereas the incorrect recognition process can be varied. It depends on the number of patterns to be recognized, on the similarity measure, on the decision making algorithm. In our case we deal with the recognition process through neurons or a fixed similarity measure as well as the decision making algorithm. Therefore, the list of enumerated incorrect recognitions is considerably shortened. In particular, if the realization to be recognized is presented to a neuron of a “different” type, for example, Xi ⇒ Nej , where i, j = 1; I���� , i≠j, Xi ∈ Aj . When the number of recognitions is more than two in the decision making algorithm the neuron threshold initial (heuristically chosen) value is frequently admissible; if the value is exceeded, out = 1 for several patterns. In such a case the decision of attributing it to the type for which the given condition is true, the decision is made according to the net function maximum value or creating a different algorithm. Later on we will discuss the process of correcting incorrect recognition results for 2.3 expression and for the cases when a realization of one type is presented for recognition to a “different” type of neuron. It should be taken into consideration that the above-mentioned situations fully comprise the list of incorrect (false) recognition processes. Let us imagine a process of incorrect recognition when Xj realization is to be recognized and is presented to Ai type Nei neuron: Xi Nei = Wi ⇒ ∑ xnj wnjn = netji ⇒ F�netji� Zi ⇒ outji = 1 ⇒ Xj ∈ Ai . (2.4) In 2.3 and 2.4 expressions weighting coefficients Wi = {Wni } of one, namely, Nei neuron are used; these coefficients must ensure, in one case, the correct recognition of ”their own” realizations, in the other case, - discarding realizations of a different type. Let us consider that F(net) function is generally nonlinear but it has a linear section as well. Later on we will consider that F(net) function parameters and weighting coefficients are chosen in such a way that the relation between net parameters and F(net) function is linear and monotonously growing. Therefore, in the further process of correcting recognition errors we can directly use net function values considering the fact that scaling of net function values is possible during the process of correction as the need arises. Let us consider F(net) function in the given range as linear which satisfies monotony conditions. Then we will have the following expression for describing the correct and incorrect recognition process: F(net) = k ∙ net = k ∑ xni wnin , where k > 0, 𝑘𝑘 = 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐. (2.5) In fact k, represents a scaling coefficient which changes in the process of adjusting the weighting coefficient. Later on, in the process of correcting mistakes we will have that the initial value of k coefficients equals one and in the process of correlation the scaling coefficient is automatically established at the necessary level. According to the above-mentioned, for describing the expressions 1.10 and 2.3 of the incorrect recognition process we will have: 1. While presenting its own realization we have (Xi Nei ) ni ni in k x w Z<∑ ; (2.6) 2. While presenting a different realization we have (Xi Nei ) Expressions 1,10 and 2,4 nj ni ink x w Z≥∑ . (2.7) Let us assume that k = 1, then 2.6 and 2.7 inequalities will be expressed as follows: ni ni in x w Z<∑ (2.8) American Journal of Intelligent Systems 2015, 5(1): 34-41 37 nj ni in x w Z≥∑ (2.9) It should be taken into account that for accomplishing the recognition process (algorithm) it is necessary to give initial values to the weighting coefficients and thresholds. In literature, there are numerous methods and algorithms for forming such values. The majority of authors select {wni } {Zi } parameters heuristically and randomly. In our case, as it was mentioned above, in the process of recognition and correction we only use award procedures, e.g. for {wni } coefficients we have: wni [ρ + 1] = wni [ρ] + α (2.10) where ρ = 0,1,2, … is the number of the step. From 2.4 expression it follows that we can receive that [0] 0niw∀ = , 1; , 1; ,n N i I= = (2.11) In a special case when the condition 1.10 and α = 1 is fulfilled then we get the so-called statistical etalons [1] where we have: 1 , 1; , 1; , 1;imni ni i in i w X m M i I n N M = = = =  (2.12) where Mi is the number of realizations in Ai type learning cluster. By means of 2.4 expression we get the realization of overlay procedure for a learning cluster of any type of set. The received set of weighting coefficients can be used for initial Wn [0] values. Finally, for the correct recognition the following inequality is true: ni ni in x w Z≥∑ (2.13) If the recognition is incorrect, the following inequality is true: ni ni in x w Z<∑ (2.14) 3. Correction of the Errors in the Process of Recognizing “Their Own” Realizations The incorrect recognition process is described by means of 2.3 expression. If we take into account 1.10 expression, then the incorrect recognition process can be described in more detail by means of the following expression (2.3 expression modification): Xi Nei ⇒ ∑ xnj wnin = netji ⇒ F�netji� Zi ⇒ outi = 0 ⇒ Xi ∉ Ai (3.1) The expression 2.7 is analogous to 3.1 expression received by means of assumptions (error making) in the previous chapter; this expression can be presented in the contracted form as follows: 1;ni ni in x w Z i I< =∑ (3.2) For correcting the recognition mistake it is necessary to change the inequality symbol in 3.2 expression, which is possible through changing the elements of vectors (matrices) xni or wni . It is obvious that realization elements cannot be changed, therefore, it is only admissible to change or, in our case to increase the weighting coefficients which is termed awarding in the theory of neural networks. Awarding can be carried out by applying iteration procedures, e.g. through 2.10 expression where the initial values of the weighting coefficients can be given according to 2.11 expression. Let us assume that according to step we received the Wni [ρ] set of weighing coefficients. In this case, if the error is not corrected, then 3.2 expression can be presented as follows: ∑ xnin wni [ρ] < Zi (3.3) If 3.3 expression is true, we will have to continue the iteration (awarding) process by 2.10 expression i.e. to increase wni = [ρ] weighing coefficients until the inequality symbol in 3.3 expression is changed: ni ni in x w Z≥∑ (3.4) where wni is a set of weighing coefficients for which 3.4 inequality is true. If we use 2.10 procedure for all Ai –type realizations, where an error occurred, then due to the fact that applying the awarding procedure does not change the correct recognition results, we will get a correct recognition of Ai -type learning cluster of { }imiX realizations. Let us carry out the above-mentioned error correction and the recognition process for the realization of all set patterns of learning sets, accordingly we will have the correct recognition of all the realizations for recognizing patterns of a learning set. There is an obstacle in the described procedure of error correction; this obstacle is termed as “repletion” in the theory of neural networks. Repletion implies such a growth of the weighing coefficients in the process of error correction, when the values of a separate and total coefficients exceed the possibilities of a neural network and the computer. In literature we have numerous methods for overcoming this obstacle; the simplest of them implies the proportional decreasing of neural network parameters which corresponds to the division of both sides of 3.4 expression (inequality) by nonnegative integral number. This problem is topical when we have a great number of patterns or in case of an increased sign space dimension. In this respect the activation function nonlinearity is more important; it suppresses the values of the activation function which correspond to a high value of the argument. In addition, the method of awarding infinitely small positive quantities to initial values of weighting coefficients is applied. 3.1. Correction of the Errors in the Process of Recognizing “Different” Realizations The error at recognizing a different type of realization is 38 Ramaz Khurodze: Development of the Learning Process for the Neuron and Neural Network for Pattern Recognition given in 2.3 expression and 2.9 expressions obtained through further assumptions and simplifications: nj ni in X W Z≥∑ Due to the fact that wni weighing coefficients have changed in the process of error correction given in the previous paragraph, instead of 2.9 expression we will have; * ni nj in x w Z≥∑ (4.1) If 4.1 inequality is true, then we will have an incorrect recognition result Xnj ∈ Ai (expression 2.4). To correct this error it is necessary to change wni ∗ and Zi parameters so that the results of correct recognitions obtained in the previous chapter are not replaced by an incorrect recognition. It means that wni ∗ and Zi parameters should be changed in such a way that the inequality sign in 4.1 expression is reversed. Let us take into account that we use an awarding procedure (expression 2.10) as a learning method; this is caused by the necessity of error correction in the process of recognizing realizations of their own pattern. Let us use the awarding procedure for the right side of 4.1 expression, that is for Zi threshold of Nei neuron. Similar to 2.10 procedure we will have: Zi [α] = Zi [α − 1] + ΔZ (4.2) Where α = 0,1,2, … is the step number. We increase the neuron threshold by 4.2 iterative expression until the inequality sign in 4.1 expression is reversed; as a result we will get the following expression: * * i[ ] Z ni ni in x w Z α< =∑ (4.3) where Zi ∗ is the required value of the threshold. If 4.3 expression is true, then the recognition process described by 4.1 expression is incorrect, that is the error made in the process of recognizing Xj realization is corrected. In order to avoid errors that Zi threshold increase may cause during the process of recognizing by Nei neuron its own Xi realizations it is necessary to take certain measures that will enable us to determine that error making in the process of recognizing its own realizations can be avoided. Let us consider that set pattern realizations are binary (condition 1.4), which enables us to make a list of possible situations for different patterns of realizations, for example, Xi and Xj realizations for Xni and Xnj components (table 4.1). It is evident, that realizations may belong to any two different patterns; the n index value of their properties (characteristic features) is the same. Table 4.1 Situation 1 2 3 4 Xni 0 1 0 1 Xnj 1 0 0 1 The recognition of Ai type Xj realizations is correct if 3.4 inequality is true. If we consider that the neuron threshold has increased according to 4.2 and 4.3 expressions, then 3.4 expression will be presented as follows: I) *ni ni in x w Z≥∑ recognition is correct or II) *ni ni in x w Z<∑ recognition is incorrect (expression 4.4). Situation 1: Xni = 0; Xnj = 1 In the first case which is described by 4.4 expression, the recognition is correct, so it is not necessary to change wni weighing coefficient; in the second case the increase in wni will give no result (no error connection) because wni = 0, as we see in case of situation 1 the increase in wni weighting coefficient is ineffective. Situation 2: Xni = 1; Xnj = 0 Proceeding from 4.4 expression the increase in wni coefficients helps to make the first inequality of the expression stricter or, which is the same, strengthens the process of correct recognition; in case of the second inequality the increase in Wni weighting coefficient becomes the necessary condition for correcting the recognition error. Consequently, it becomes evident that in case of situation 2 the increase in weighting coefficient is effective according to the increase in the reliability of the recognition Situation 3: Xni = 0; Xnj = 0 In case of situation three, on the basis of 4.4 expression the recognition result (correct or incorrect) will not change, as we have weighing coefficients multiplied by zero. Situation 4: Xni = 1; Xnj = 1 In case of situation four, the recognition of Xi realization will improve. But it is possible to make errors in the Xj realizations recognition process, which occurs when Zj < Zi , so we must consider other circumstances in case of the increase in wni coefficients, or we must avoid the increase procedure and search for another way out to correct the erroneous recognition. The analysis of the above-given situation shows that for the correct recognition of any type of realization it is necessary for the etalon description of this pattern to have at least one sign whose realization probability for one pattern is equal to one, whereas it is equal to zero for other patterns. If there is such a sign, then the correction of the recognition error is possible through using an awarding procedure. In other cases it is necessary to use the existing methods for evaluation or to work out a new one. 4. Signs Evaluation (Ranging) According to Learning Set Realizations Let us assume that we have chosen (heuristically or American Journal of Intelligent Systems 2015, 5(1): 34-41 39 according to other factors) signs-characteristic features that represent A set patterns in this sign space. The presentation of patterns in the sign space occurs by means of realizations whose totality for each pattern forms or must form a separate cluster (group). It is implied that the number of clusters must be equal to the number of patterns or exceed it. The position of clusters in the sign space determines their separateness degree and, accordingly, the reliability of recognition process. Namely, the more separated the type-reflecting clusters are from one another, the more reliable the recognition process is and vice versa; in case of closely positioned clusters the recognition process becomes more complicated and the recognition becomes unreliable. The process of selecting signs, as well as its result, greatly influences the position of clusters in the sign space, so the selection of signs should be based on the existing scientific information (if it is possible) in the given sphere. The next stage may be selected according to evaluating signs together or separately. In our case the process of evaluating-selecting signs is considered according to its reliability which makes it necessary to introduce some concepts – terms through which we will characterize certain properties (signs) from the point of view of recognition, content by means of formalizing these concepts later on. To evaluate a separate sign in the sign space let us use general and content criteria: To what extent a concrete pattern is characterized by a given sign. To what extent a given concrete pattern of a sign differs from the same sign of another pattern. Characteristic features represent the property-characteristic of the object which the object possesses always or very frequently; this is expressed by the fact that we have this property in the object realizations or if it is measurable, its values are steadily placed in a certain line of values; the sign possesses a property distinguishing it from other patterns, if it satisfies the condition of specificity and, at the same time, this condition is not fulfilled clearly in relation to other signs. According to the condition, 1.4 expression, the sign space and, correspondingly, pattern realizations are binary. At the same time, any realization of any pattern (vectors or matrices) has equal dimensions, which makes it possible to present the concepts “characteristic” and “distinguishing” in a formalized way. Let us place the realizations of a learning cluster according to types, after this with the help of superposition (overlay) let us make the so-called “statistical” description (etalon) for each type separately [1]. Let us designate the Ai type statistical etalon with Si (vector or matrix) the coordinates of which are: S = s1 , s2 , … , sni , … , sNi , … i = 1; I���� (5.1) Where for ∀sni we have: 1;ini i R S i I M = = (5.2) Where Ri is the number of those realizations in Ai pattern learning set in which Xni sign gets the value equal to one. Proceeding from 5.2 expression we can formalize “characteristic” and “distinguishing” signs: Sign xni is absolutely characteristic of Ai type if Sni = 1. Sign xni is absolutely distinguishing for Ai type if Sni = 1 and, Snj = 0, j = 1; J����, j ≠ i For the situations discussed in chapter 2 xni sign is absolutely characteristic for Ai pattern if xni = 1 (situation 2 and situation 4) for all the realizations of Ai pattern learning set. The same sign is absolutely distinguishing for Ai pattern if the following condition is fulfilled: xni = 1; xnj = 0, (situation 2). ∀i, j ∈ I; i ≠ 0. It is evident that the greater the number of the absolutely characteristic and distinguishing signs in the description of the given pattern, the higher the chance of correct recognition for this pattern of realization and vice versa. Unfortunately, the combined occurrence of characteristic and different signs for a real object is quite rare; due to this it becomes necessary to introduce intermediate concepts and their formalized definitions. In particular, let us assume that xni xni sign is partially characteristic of Ai pattern if the following condition is fulfilled: 0,5 ≤ Sni ≤ 1 (5.3) xni sign is partially distinguishing for Ai pattern if 5.3 condition and the following condition are fulfilled: 0 ≤ Sni ≤ 0,5 j = 1; J����, i ≠ j (5.4) It is obvious that the greater the number of signs satisfying the 5.3 and 5.4 conditions, the higher the chance of correct recognition and vice versa. It can be said that through carrying out the learning processes discussed in chapters 3 and 4 the error-free recognition of learning set realizations is achieved. If there are no absolutely characteristic and distinguishing signs for the patterns, then it becomes necessary to find satisfying signs for conditions 5.3 and 5.4; this will make it possible to correctly evaluate the recognition process for such cases. Let us designate the set of Ai pattern signs for Aj type which satisfy 5.3 and 5.4 conditions �Sij� and by Sji set - Aj pattern signs which satisfy the same conditions in relation to Ai pattern; the correlation of the whole set of signs {X} and Sij Sji , Sji sets is given in Picture 5.1: Let us assume that the number of signs in Sij set is equal to Rij , whereas in Sji set it is equal to Rji ; for the corresponding indexation of these signs in order to describe their total sequence let us use symbols 1; ; 1;ij i ji jr R r R= = ; let us consider expressions: ; rij ij rji jir rS Q S Q= =∑ ∑ (5.5) 40 Ramaz Khurodze: Development of the Learning Process for the Neuron and Neural Network for Pattern Recognition Picture 5.1 As in equalities Rij < 𝑁𝑁 and Rji < 𝑁𝑁 are fulfilled, the following inequalities are also true: ] ; ij ni i ji nj jn nQ S C Q S C≤ =∑ ∑ (5.6) We can characterize (range) separate elements or groups of elements of a sign set from the point of view of recognition reliability which depends on how big is Qij value which constitutes Ci value and on how big is Qji value which constitutes Cj value. The following inequalities are received considering 5.6 expression: 0; , 1, , i ijC Q i j I i j− ≥ = ≠ (5.7) 1. A sign is absolutely useful if 5.7 inequality is true for ∀i and ∀j. In this case there is at least one sign in the sign space which ensures the Ai pattern error-free recognition for learning set realizations of the given set of patterns. 2. A sign is useful if 5.7 expression is fulfilled for a part of patterns, e.g. for half of the patterns: 1 10; 1, , 0,5 i ijC Q j I I I I− ≥ = ≤ < (5.8) 3. A sign is less useful if 5.7 expression is true for less than a half of patterns; 2 20; 1, , 0,5 i ijC Q j I I I− ≥ = < (5.9) The I1, I2 coefficient values are taken heuristically, their improvement is possible and should be made for a concrete set of types and signs. It is obvious that absolutely useful signs for real objects and the sign space are very rare, whereas useful and less useful signs for the vast majority of patterns and signs occur quite often. It is evident that the greater the number of absolute or useful signs is, the greater is the possibility of correct recognition of learning processes described in the previous chapters and vice versa. 5. The Plan and Algorithm for Experimental Research The error correction methods and sign evaluation algorithms given in the previous chapters are based on the exact knowledge of the elements of pattern realizations set as well as the realization of their learning set. According to the condition the set of signs for all the patterns and accordingly for all the realizations should be equal and binary. To carry out the realizations of learning set patterns it must be determined which pattern each of them belongs to; this must be considered attentively as otherwise this will cause the incorrect description of patterns in the learning process and eventually we will get incorrect recognition results. The stages of experimental research are as follows: Building neural networks or selecting them from the existing ones; The setting up – calculating the threshold coefficient for each neuron in the network; Awarding (calculating) the weighting coefficient for each sign; Error correction in the process of recognizing their own realizations; Error correction during the recognition of different realizations; Carrying out feature’s evaluation (algorithm). In order to carry out the above mentioned stages it is necessary to create the so-called “statistical” descriptions, after which it becomes possible to calculate the threshold for each neuron by way of conducting the recognition process (process algorithm). If necessary, it is possible to get the error-free recognition of learning set patterns realizations by means of changing-improving the weighting coefficients of a neuron. This article has a theoretical character, so explained method may be used for practical tasks, which are executed by neural nets. REFERENCES [1] McCulloch, Warren; Walter Pitts. "A Logical Calculus of Ideas Immanent in Nervous Activity". Bulletin of Mathematical Biophysics 5 (4): 115–133. 1943, doi:10.1007/ BF02478259. [2] Hebb, Donald. The Organization of Behavior. New York: Wiley, 1949. [3] Farley, B.G.; W.A. Clark "Simulation of Self-Organizing Systems by Digital Computer". IRE Transactions on Information Theory 4 (4): 76–84, 1954. doi:10.1109/TIT.195 4.1057468. [4] Rochester, N.; J.H. Holland, L.H. Habit, and W.L. Duda. "Tests on a cell assembly theory of the action of the brain, using a large digital computer". IRE Transactions on Information Theory 2 (3): 80–93. 1956 doi:10.1109/TIT.195 6.1056810. [5] Rosenblatt, F. "The Perceptron: A Probalistic Model For Information Storage And Organization In The Brain". Psychological Review 65 (6): 386–408. 1958, doi:10.1037/h 0042519. PMID 13602029. [6] Werbos, P.J.. Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences, 1975. [7] Minsky, M.; S. Papert. An Introduction to Computational Geometry. MIT Press. ISBN 0-262-63022-2, 1969. [8] Rumelhart, D.E; James McClelland Parallel Distributed Processing: Explorations in the Microstructure of Cognition. American Journal of Intelligent Systems 2015, 5(1): 34-41 41 Cambridge: MIT Press, 1986. [9] http://www.kurzweilai.net/how-bio-inspired-deep-learning-k eeps-winning-competitions, 2012 Kurzweil AI Interview with Jürgen Schmidhuber on the eight competitions won by his Deep Learning team 2009–2012. [10] Graves, Alex; and Schmidhuber, Jürgen; Offline Handwriting Recognition with Multidimensional Recurrent Neural Networks, in Bengio, Yoshua; Schuurmans, Dale; Lafferty, John; Williams, Chris K. I.; and Culotta, Aron (eds.), Advances in Neural Information Processing Systems 22 (NIPS'22), December 7th–10th, 2009, Vancouver, BC, Neural Information Processing Systems (NIPS) Foundation, 2009, pp. 545–552. [11] A. Graves, M. Liwicki, S. Fernandez, R. Bertolami, H. Bunke, J. Schmidhuber. A Novel Connectionist System for Improved Unconstrained Handwriting Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 31, no. 5, 2009. [12] http://www.scholarpedia.org/article/Deep_belief_networks / [13] Hinton, G. E.; Osindero, S.; Teh, Y. "A fast learning algorithm for deep belief nets". Neural Computation 18 (7): 1527–1554. doi:10.1162/neco.2006.18.7.1527.PMID 16764513, 2006. [14] Fukushima, K. "Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position". Biological Cybernetics36 (4): 93–202. doi:10.1007/BF00344251. PMID 7370364, 1980. [15] M Riesenhuber, T Poggio. Hierarchical models of object recognition in cortex. Nature neuroscience, 1999. [16] Deep belief networks at Scholarpedia. [17] Hinton, G. E.; Osindero, S.; Teh, Y. W. "A Fast Learning Algorithm for Deep Belief Nets". Neural Computation 18 (7): 1527–1554.doi:10.1162/neco.2006.18.7.1527. PMID 167645 13. edit, 2006. [18] John Markoff. "Scientists See Promise in Deep-Learning Programs". New York Times. November 23, 2012. [19] Verulava O., Khurodze R. “Theory of Rank of Links: Modeling of Recognition Process”, Nova Science Publishers, Inc. New York, 2011. [20] Verulava O. Khurodze R., Clustering Analysis and Decision-making by “Rank of Links”, Mathematical Problems in Engineering, 2002, Vol. 8(4-5), pp. 475-492, Taylor & Francis group. 
work_ut5scaz4jncc5b67akup37xhpe	----	EKAW European Knowledge Acquisition Workshop Intellectual Capital: From Intangible Assets to Fitness Landscapes Brendan Kitts Vignette Corporation, 19 Newcrossing Road, Reading, MA. 01867. USA. Ph: (US) 781 942-3600x136, Fax: (US) 781 942-2163, Email: bkitts@vignette.com Leif Edvinsson Skandia Corporation, Skandia Future Center, Villa Askudden, PO Box 153, S-185 22 Vaxholm, Sweden. Email: ledvinsson@skandia.com Tord Beding Hagatornet AB, Haga Nygata 28, S-411 22 Gothenburg, Sweden, Email: hagatornet@goteborg.mail.telia.com Abstract Intellectual Capital (IC) has been proposed by Edvinsson and Malone (1997) as a technique for quantifying a company’s intangible assets. A careful analysis can result in hundreds of variables, and extracting knowledge from these measurements can be difficult. We introduce a knowledge management technique called IC mapping that attempts to synthesize this data into a fitness landscape. Using the map, managers can query the surrounding landscape, view past progress as a trajectory across the landscape, and calculate what parameters need to be changed to reach new locations. IC mapping provides a novel knowledge management tool for understanding, managing, and representing a company’s intangible knowledge assets. mailto:bj@vignette.com mailto:ledvinsson@skandia.com mailto:hagatornet@goteborg.mail.telia.com Kitts, Edvinsson and Beding _______________________________________ Page: 2 Introduction Managing corporations is a difficult and risky business. Doing it well requires the ability to understand factors including market, customers, employees, technology, culture, history, and opportunities. Management is made all the more difficult because variables interact. For instance, increasing expenditure on information technology may decrease next year’s profitability, but increase the number of projects completed on-time. Such variable interactions may be separated in time, non-causal, non-linear, and involve multiple variables. This is where a knowledge management tool and can help. This paper introduces a knowledge management technique we call “IC mapping”. IC mapping extracts knowledge from historical company data and converts them into an interactive three-dimensional landscape. Using this map, managers can interactively query the landscape, perform what-if analysis, identify problems in their business, and understand their company’s performance in relation to others. Multivariate measures of company performance With the growth of the services sector in major industrialized countries (USBLS, 1999), many authors have suggested that non-traditional or “intangible” assets of business operations - such as customer relationships, and skills of employees may be increasingly important, and worthy of reporting on profit-loss sheets in their own right. In effect, a wider measurement net needs to be cast in order to capture the fitness of a company (Karlgaard, 1993; Kaplan and Norton, 1996; Edvinsson and Malone, 1997; Madison Valuation Associates, 1999). Kitts, Edvinsson and Beding _______________________________________ Page: 3 Leif Edvinsson pioneered the development of diversified measures of company performance from 1990-1997 (Edvinsson and Malone, 1997). The system he devised, “Intellectual Capital” (IC), measures performance in five major areas: Financial, what appears on ordinary balance sheets; Human, the skills and experience of the employees; Customer, goodwill, relationships, and brandname; Process, which measures how efficient internal functions are; and Renewal, which measures growth and long-term research and development. By using variables from diverse areas, Edvinsson hoped to probe for problems that would remain hidden in ordinary profit and loss balance sheets. This information could then be used to inform strategic decisions as to expenditure in different organizational areas, new investments, and business reorganization. Top variables for predicting future earnings 1 year into the future Rank Variable Type Correlation Rank correlation F number of observations 1 Operating result F 0.7632 0.7544 17.4723 32 2 Number of contracts C 0.4182 0.4794 4.7225 29 3 Number of contracts / employee P 0.4642 0.4395 3.6638 19 4 Points of sale C 0.77 0.8076 3.557 8 5 Percent managers H -0.9422 -0.9918 3.5511 6 6 Operating income F 0.8034 0.8143 3.2272 7 7 Value added / employee F 0.7022 0.7594 2.9586 8 8 Assets under management F 0.4033 0.4888 2.6027 18 9 Change and development of existing holdings R 0.701 0.4113 1.9656 6 10 Percent of managers who are women H 0.457 0.4434 1.8793 11 11 Payroll costs / administrative expenses (%) P 0.5933 0.6609 1.7602 7 12 Savings / contract (000s) C 0.5388 0.6096 1.7421 8 13 Adm exp / gross premiums written (%) P -0.358 -0.3498 1.538 14 14 Return on capital employed (%) F 0.5238 0.5479 1.3718 7 15 Developmental expense as a percentage of Gross profit H 0.4203 0.4942 1.0599 8 Kitts, Edvinsson and Beding _______________________________________ Page: 4 2 years into the future Rank Variable Type Correlation Rank correlation F number of observations 1 Operating result F 0.6172 0.6609 6.0944 18 2 Number of contracts / employee P 0.6717 0.8518 4.5123 12 3 Number of contracts C 0.5135 0.5869 4.2187 18 4 Assets under management F 0.5133 0.5834 2.3714 11 5 Increase in number of contracts (%) R -0.3894 -0.3053 1.8196 14 6 Adm exp / gross permiums written (%) P -0.4827 -0.4598 1.6308 9 3 years into the future Rank Variable Type Correlation Rank correlation F number of observations 1 Number of contracts / employee P 0.6643 0.8015 3.0894 9 2 Adm exp / gross permiums written (%) P -0.7027 -0.8822 2.4691 7 3 Number of contracts C 0.4315 0.4077 1.6759 11 21* Operating result F -0.112 0.0972 0.1004 10 * included for comparison Figure 1: Top variables (in terms of strength of effect) for predicting Total Operating Result (MSEK) at Skandia companies 1, 2 and 3 years into the future. These results were generated by taking data from Skandia’s Annual reports, joining together variables which were common between these companies, and then collecting together training data and operating result 1, 2 and 3 years after training data sample is taken. Past Operating Result becomes unpredictive for predicting revenue three years into the future (F=0.1, R=-0.1), but Number of Contracts Per Employee remains reliable over all years (F>3, R>0.4). Our own detailed analysis of Skandia companies has confirmed the importance of diverse company measurements. We acquired data from Skandia over the period 1991 to 1997 and ran a correlation between variables and future earnings. The problem was to predict Total Operating Result in the future, based on observed variables for a company. Kitts, Edvinsson and Beding _______________________________________ Page: 5 Fifty-nine variables were available, and variables were transformed into zscores within their company timeseries, to prevent any effect from the absolute operating size of each company. Results are shown in figure 1. The best variable for predicting operating income 12 months into the future is Total Operating Result in previous year, which is a financial variable. However, as the prediction horizon extends to 2 and 3 years into the future, the predictive utility of financials decreases. At 2 and 3 years in the future, Contracts Processed Per Employee, and Value-Add Per Employee become more strongly correlated with future success than past operating income. Administrative Expenses and Percent Of Employees Who Are Managers have negative correlations with future Operating Result. Unfortunately, even with detailed statistical analysis, extracting knowledge can still be difficult. With potentially hundreds of interacting, non-linear and co-linear IC variables, it can be difficult for a human being to grasp the overall state of an organization. Edvinsson and Malone recognized this problem, and wrote of the need for a way to collect these variables together into a form which could be easily comprehended by a human being: “Somehow, then, all of these regions must be pulled together into an overall format …. what is needed in Intellectual Capital is a map that captures all of the value of an enterprise, color-coded so that one can quickly ascertain the quality of the topology – where there are swamps and lush forests, mountains and deserts.” (Edvinsson and Malone, 1997, pp. 67) Kitts, Edvinsson and Beding _______________________________________ Page: 6 The ideal method of presentation should make it possible to see such interactions, and allow human beings to quickly comprehend what was happening within the company. This article describes a new method that we believe fulfills these requirements. The technique extracts knowledge from historical IC measurements, and represents them using a knowledge map. The map shows the fitness of a company for different combinations of parameters, and supports interactive what-if analysis. The promise of mapping IC Mapping is designed to provide a visual tool that unifies the operating state of the company, and is easy to use. The method is based on a body of well-known statistical methods including Multi-dimensional scaling, which was pioneered by Torgerson (1958), Kruskal (1978), Shepard (1974), Young (1987; 1998), and others; and non-parametric estimation (applied widely in the field of Neural Networks), investigated by Girosi, Jones and Poggio (1993) among others. Kitts, Edvinsson and Beding _______________________________________ Page: 7 Figure 3 shows an example of an IC map, with a company traveling across a fitness landscape. Each point on the map represents a possible company state vector. At each location, the height represents the fitness of that state. This gives rise to mountains and valleys for different company states. The user can see their company in relation to other companies (and their own company’s past states) on the map. In this way, the user can use those companies like navigation stars, to chart a course across the landscape. The applications of this map are numerous. Policy change comes into effect Position of the company 24 months ago Extrapolated position of company +6months Region of poor fitness Current position of the company Figure 3: Viewing a company’s weight timeseries allows us to plot its speed and direction of movement, and allows us to anticipate problems it will encounter in the future. Kitts, Edvinsson and Beding _______________________________________ Page: 8 a) Viewing Company Trajectory across landscape: By connecting the company’s positions in chronological order, we can view the historical movement of the company as a trail of points across the map. b) Landscapes constructed using multiple companies Multiple companies can be displayed on the same fitness landscape, allowing a company to compare itself to competitors, and get some idea of the shape of the fitness terrain far from its present location. c) Labeling important terrain Important points or regions, such as a company that went bankrupt, can be explicitly labeled on the map. These will become landmarks, and companies can identify if they are heading towards or away from them. d) Predicting future location and avoiding poor regions of fitness By extrapolating the trajectory of the company, for instance, drawing a line through the last few points, we can also estimate where the company will be in the next period of time. This may enable decision makers to recognize that their company is moving towards a bankruptcy several years before reaching that condition. e) Time to reach destination The time to reach a new destination can be estimated by using historical parameter change speed as the basis for predicting time to reach new points. Kitts, Edvinsson and Beding _______________________________________ Page: 9 f) What-if analysis We can interactively test the outcome of making changes to one or more variables, for instance, doubling the administration budget. After a change is made, the updated position of the company can be displayed on the landscape. g) Means-ends analysis We can look for desirable regions, and find out the variable settings that underlie that point, and are required to reach that location. Means-ends analysis involves building an inverse model mapping 2D surface position back to high-dimensional IC vector. h) Path planning Optimal trajectories can be calculated by finding a path that maximizes fitness between source and destination, constrained by time allowed to reach destination; in other words, maximizing the path integral, bounded by a particular time allowed. The path planning model can be made more elaborate by estimating “currents” or natural interactions between variables which will tend to move parameters in particular directions. Exploitation of drift currents might enable positions to be reached faster at lower cost. Building a Map The mapping process consists of two stages. The first task is to project the high dimensional company data into two dimensions. The second is to add a fitness variable as a third dimension, Kitts, Edvinsson and Beding _______________________________________ Page: 10 and then interpolate between those points to predict the shape of the fitness surface connecting these regions. Step 1: Multidimensional scaling The objective of MDS is to represent high dimensional data in 1-3 dimensions so that a human being can visually understand the data. Because its not possible to directly visualize more than three spatial dimensions, the method focuses instead on preserving the inter-point distances of the high-dimensional data. MDS therefore finds a set of points in two dimensions which have distances as close as possible to the inter-point distances in the high dimensional space. If the method succeeds, then a human looking at the points will be able to correctly judge that their present position is “very different from” some other point on the map, and that judgement will also hold in the difference between the two high-dimensional vectors. (Young, 1998; Young and Hamer, 1987; Norusis, 1997) It should be noted that the axes of the new, low-dimensional space are not intuitively interpretable since each is constructed out of convenience to retain the distance relationships. The only features that make sense in the new landscape are the concepts of “more similar to-“ or “more different from-“ x, where x is a known case. Thus, this is somewhat analogous to a paper topographic map, where landmarks are shown (the other data points), and the user can see how distant these landmarks are to his or her position. Algorithms for Multi-dimensional Scaling Kitts, Edvinsson and Beding _______________________________________ Page: 11 To generate the new coordinates one has to minimize the error between the distance between points in the orignal data dij, and in the 2D representation ij. This “error” is known as Stress, and was introduced by Kruskal and Shepard (Kruskal, 1978; Shepard, 1974):     n i n j 2 ij 2 ijij )(d )S tress(d, Many different algorithms can be used to minimize stress (Kohonen, 1996; Young, 1987). In our own work we have found that the Torgerson Projection (Torgerson, 1958) followed by several iterations of a Nelder-Mead simplex optimization (Nelder and Mead, 1965; Betteridge, et. al, 1985) can generate very good results. Other empirical comparisons of MDS algorithms can be found in Li et. al. (1993, 1995) and Duch and Naud (1998a, 1998b). Torgerson’s Classic metric MDS algorithm Torgerson showed in the 1950s that if a distance matrix was double centered such that, given a raw distance matrix B, the double-centered matrix D is defined as dij = -0.5 (bij 2 – bi. 2 – b.j 2 + b.. 2 ) then the following relationship held: D = XX T (2) (3) (1) Kitts, Edvinsson and Beding _______________________________________ Page: 12 This meant that any distance matrix could be changed into a coordinate matrix by taking a matrix square root of D. To derive a set of coordinates with less than the original number of dimensions, singular value decomposition could be used to identify the two largest eigenvalues, and then reconstruct coordinates with only the two largest eigenvalues. This will result in coordinates which capture as much of the variance as possible in the selection of coordinates. UU T = D X = U 1/2 Self-organizing map Self-organizing maps are a new technique originally formulated as a model of human visual cortex (von der Malsburg, 1973). The self-organizing map consists of an array of neurons which each store a prototype of the input they most prefer (their tuning curve). As input comes in, the cell with the prototype vector most similar to the input “wins”, and adjusts its prototype to be more similar to the input. Each cell also has a “neighborhood” of other cells. This can be seen in the figures below (figure 4 and 9) and as a mesh connecting the cells. When all the cells start off, they are in a perfect grid. Over time this grid will deform to “map” the input. Whenever one neuron adjusts its vector, the other cells in the neighborhood also have their prototypes adjusted in exactly the same direction, but to a smaller degree. The result is that neighboring cells represent similar inputs. If we then “read off” what each cell represents, we will find that input vectors have migrated to different regions of the self-organizing map. These (4) (5) Kitts, Edvinsson and Beding _______________________________________ Page: 13 different regions are the low-dimensional manifestation of the high-dimensional data. Other details of the Self-organizing map can be found in Kohonen (1996). 70 75 80 85 90 95 100 105 1.25 1.3 1.35 1.4 1.45 1.5 1.55 1.6 Figure 4: Top-view of the Kohonen self-organizing map, as it attempts to map a higher dimensional object. Areas of “compression” in the map represent regions where high concentrations of datapoints exist. High concentrations of points encourage centroids in the Kohonen map to “crowd in” try to to map those denser regions of points. The equations for a Kohonen net are as follows: Given a learning constant 1>>0, a cv matrix of centroids which is initially random, the centroids are changed as follows, upon presentation of x to the network: Kitts, Edvinsson and Beding _______________________________________ Page: 14   )c(xccF dt c nminn N(min)n    α xcmincc iimin                       gsize i gsizei, gsize i grid(i) nsized, otherwise0, nsize d 1 F(d)        nsizegrid(j)grid(i)N(i)j  where nsize is neighborhood size, gsize is gridsize, The first equation states that a neighbor cn of the closest centroid cmin, is moved towards the input x at a rate equal to a function of its distance on the mesh to cmin. This speed function, F, is linear with decreasing activation from distance 0 from cmin to 0 at distance nsize. Thus, the closest centroid moves fastest towards the input. N(cmin) is a set of neighbors of cmin. In this implementation neighbors form a square grid around cmin, where their grid coordinates are given by grid(.). In addition to the learning algorithm above, the SOM also undergoes annealing of the neighborhood size nsize, and learning rate . These details can be found in Kohonen (1996). Typical parameters used in our implementation were =0.005, gsize=10, nsizeinit=gsize/2. (6) (7) (8) (9) (10) Kitts, Edvinsson and Beding _______________________________________ Page: 15 Comparison of map quality for MDS methods Duch and Naud (1998a, 1998b) developed a novel test set to evaluate the quality of Multi- dimensional scaling methods. They proposed applying MDS to n-dimensional equilateral simplexes. These simplexes are easy to build, since their distance matrix is simply all 1s. When these simplexes are reconstructed in two dimensions using MDS, intricate symmetrical geometric representations are created. The quality of the MDS method can be visually checked by merely looking at the symmetry of the reconstructed shape. MDS results using Duch and Naud’s test set are shown in figure 5 and 6. Figure 5: 7-dimensional simplex, 8-dimensional simplex, stress=3.610143. 12-dimensional simplex, stress 5.3911, generated using Nelder-Mead optimization with 10,000 iterations Torgerson’s method (figure 6 right) was the poorest on Duch and Naud’s test set. Torgerson’s method works by recovering coordinates from the distance matrix, and then finding the principal components of the coordinates, and projecting the coordinates onto these principal components. The resulting 2 dimensional coordinates capture as much of the variance as possible. However, all of the distances between each vertex are equal to 1. Therefore, in this particular application all Kitts, Edvinsson and Beding _______________________________________ Page: 16 of the dimensions are equally important, and taking the principal components just results in loosing d-2 important dimensions. Self-organizing maps (figure 6 middle) fair better, but are still twice the stress of Nelder-mead MDS (figure 6 left). Nelder-mead optimization generates the best results by far. Figure 6: 6D simplexes represented in two-dimensions using three different algorithms. (left) MDS stress minimization, Iterations = 10,000, Stress = 2.786404, (middle) SOM 10x10 grid, Stress = 5.078828, (right) Torgerson, stress = 20.9641 Step 2: Fitness as the Third dimension The most important feature of the landscape we want to plot is fitness. Fitness refers to the health of the company, and examples can include “total operating income”, or “net profit after tax”. Because knowing the fitness of the company at a given position is so important, we carry fitness into the 2D projection unaltered, and have it plotted as the third dimension. Kitts, Edvinsson and Beding _______________________________________ Page: 17 After the above step we have a set of 3D points. We can think of these 3D points as a scaffold for the landscape that we want to build (figure 7,8,10 shows an example for one company Skandiabanken). Unfortunately, this scaffold is still difficult to interpret. To infer the appearance of a lanscape, we need a model which predicts fitness at any point on the 2D map, and so infers the topographic surface around the known points. To generate this, we can use function estimation techniques such as neural networks, splines, and regression to fill in away from the observed datapoints. We have found that spline estimators (Karur and Ramachandran, 1995; Girosi, Jones and Poggio, 1993) give the best results in generating our surfaces. Unlike polynomials, splines fit surfaces around local knot points, allowing them to model local surface features. Splines Splines are a non-parametric regression technique which approximate a function by (a) finding a set of high-density “centroids” in the function (b) projecting all data onto a new coordinate system where each axis is a centroid, and the value on the axis is the distance from this data point to that centroid, (c) performing a least squares mapping from the new points to the target. The spline model is defined as follows: Let C be a c2 basis matrix of centroid vectors, sometimes called knot points. Let S be a rc basis-transformed input matrix, and let W be a c1 matrix of weights. Given an r2 data matrix X, a spline approximates a function by applying the following formula: Kitts, Edvinsson and Beding _______________________________________ Page: 18 where S is a representation of the input which has been transformed by the basis function given by G(.). G is usually one of the radial functions given below (Karur and Ramachandran, 1995). The Gaussian spline has come to be known as a “Radial Basis Function network” in the neural networks literature (Girosi, Jones, and Poggio, 1993; Orr, 1996), however all of the functions below are actually radial functions, so ‘RBF’ is a slight misnomer. Radial Basis net (Gaussian spline)           σ 2 XCD XC e)G(D Thin-plate spline 1)log(DD)G(D XCXCXC  2 Cubic 3 XCXC D1)G(D  Multi-quadratic WS  )G(DS XC  (13) (14) (15) (16) (17) (18) Kitts, Edvinsson and Beding _______________________________________ Page: 19  222 n XCXC D)G(D   Linear spline XCXC D1)G(D  2 , log and e operate over the individual elements of the matrix. The DXC term is the distance between each row of X and the centroid points, C. Therefore, S can be thought of as a dot product between the input and the bases, normalized by the size of both X and C, and then put through a non-linear transform. We also ensured that all activations were normalized such that Where V is the number of centroids. Normalization meant that extrapolation outside the range of datapoints did not result in overly large or small values. Instead, the surface far from the points converges to the average of the datapoint values. Given the small number of datapoints, we also set the rows of C to be equal to the existing datapoints, so each centroid was a datapoint. Examples of linear, cubic, and thin-plate spline surface approximation for Skandiabanken’s fitness landscape are shown in figures 11, 12, 13 and 14. XXCCX2CD TTT XC  1Sr V 1v rv    (19) (20) (21) Kitts, Edvinsson and Beding _______________________________________ Page: 20 Adding details to the map Once the final landscape is built, a variety of details can be added to the basic map. Some observations may coincide with an important event, for instance “competitor came onto the market”, “company went public”, and so on. We can carry these labels from our high dimensional data into our 2D map, and display them on the map. This can help the user navigate the map. Finally if the trajectory across the landscape is itself observed once every 12 months, the points in-between are not known. Since linearly connecting these points would be assuming a linear interpolation, we have connected points using smoothing splines to convey less certainty as to the nature of the points in-between the observations. Time to reach destination In the early days of navigation, ship captains used calipers to measure the distance between locations on their map. They would measure speed by dropping buoys and using a stopwatch to calculate speed across water, a process known as “heaving the log” (Bowditch, 1826). Using these tools, ships could estimate the time to reach their destination. On IC Maps the same calculation can be performed. A simple method is to compare the parameter difference needed to reach the new location, with the company’s historical parameter change speed per year: Kitts, Edvinsson and Beding _______________________________________ Page: 21 T(C,a,b) = Absolute parameter change / Average Parameter change per year          1T 1t i C 1t C i ii it,xi,x 1T 1 ba b)a,ΔT(C, where a and b are the points we want to travel between, C is the company which will be travelling, xti is a historical company value for variable i at time t, and T is the total number of historical observations. This assumes that a company is able to change each of its parameters equally well, and is limited by an inherent magnitude of change per year. Other methods for estimating time can also be used; for instance it is also possible to take into account historical interactions between the IC variables to estimate “currents” that might increase or decrease time to reach destination. Kitts, Edvinsson and Beding _______________________________________ Page: 22 Figure 7: Scaffold for Skandiabanken’s low-dimensional fitness landscape generated using Nelder-Mead stress optimization (Stress = 2.001988). Height represents total operating income for years 1994-1997, and 2D position represents 11-dimensional state. Figure 8: Scaffold for Skandiabanken data, Torgerson MDS (Stress = 42.18) -10 -5 0 5 -2 -1 0 1 2 200 250 300 350 400 -4 -3 -2 -1 0 1 2 3 4 -2 -1 0 1 2 200 250 300 350 400 Total operating income Kitts, Edvinsson and Beding _______________________________________ Page: 23 Figure 9: Side view of the Kohonen net, showing how it tries to “reach” out to the different points, constrained by its 2D mesh topology. Figure 10: Scaffold for Skandiabanken generated by Self-organizing map, (Stress = 3.59) 70 75 80 85 90 95 100 105 1.2 1.3 1.4 1.5 1.6 12 14 16 18 20 22 24 26 2 4 6 8 10 12 5 10 15 20 200 250 300 350 400 Figure 11 and 12: (top) Linear spline and (bottom) cubic spline approximation to SkandiaBanken’s surface March 20, 1999___________________________________________________ Page: 25 Figure 13 and 14: Different methods for showing the landscape. 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20 360 3 8 0 3 8 0 34 0 36 0 3 8 0 3 2 0 3 8 0 96 3 0 0 thin plate spline | stddev 56 | nd modelerror 172 | 2d stress 2.001988 | 2d modelerror 125 3 4 0 36 0 2 8 0 32 0 2 6 0 2 4 0 3 6 0 95 30 0 3 4 0 3 2 0 2 8 0 26 0 220 3 0 0 94 24 0 2 8 0 26 0 March 20, 1999___________________________________________________ Page: 26 Experiment 1: Skandia Fitness Landscape In order to test whether IC Mapping could be used for real companies, we selected four companies from the large multinational Skandia investment group to be projected together onto a fitness map. The companies were Intercaser, Dial, American, and UK Life. Intercaser, American, and UK Life are the Spanish, American, and UK branches of Skandia Investment firm. Dial is a telemarketing insurance company which operates throughout Nordic countries. The state of each company was fixed with a vector of five variables - one from each IC focus area. The five variables used were total operating result, number of customers, number of employees, contracts processed per employee, and % new contracts. Missing values were estimated using stepwise timeseries imputation. The variables were transformed into Z-scores. Ideally the fitness metric for the companies should be predictive of future earnings potential, and we could have developed a regression equation for this purpos. However, for the purposes of the test, we decided to let fitness be a linear standardized sum of the IC focus areas. The multi-dimensional scaling was performed using a Torgerson projection (Torgerson, 1958), and fine-tuned via a Nelder-Mead simplex optimization. Optimizations took a couple of hours on a 200 MHz Pentium computer, and resulted in a stress of approximately 17.32 with correlation between original and projected distances equal to 0.91 (figures 19, 20). Tests using other methods achieved similar stresses and 2D point March 20, 1999___________________________________________________ Page: 27 positions, lending confidence to the landscape produced (figures 15, 16 shows scaffolds generated using different methods). -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 0 2 4 Kruskal non-monotonic MDS scaffold of points -4 -3 -2 -1 0 1 2 3 4 -3 -2 -1 0 1 2 3 0 2 4 Torgerson/Nelder-Mead scaffold of points Figure 15 and 16: Scaffold of points for the Multiple company Skandia map March 20, 1999___________________________________________________ Page: 28 The final Skandia map is shown in figures 17 and 18. The map shows American Skandia rapidly zigzagging across the landscape, and increasing its Intellectual Capital. Intercaser is also doing well, and is headed out of the map area, while Link is maintaining steady fitness. UK Life, however, is losing Intellectual Capital and slipping down the slope. Figure 17: Skandia IC landscape. Height dimension is IC. Am = American Skandia, UK = UK Life, In = Intercaser, Di = Dial March 20, 1999___________________________________________________ Page: 29 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20 3 3.5 Uk92.5 Uk 92 Uk 93 Uk93.5 Uk 94Uk 94 Uk94.5 3 4 2 .5 Uk 91Uk 91Uk 91Uk 91Uk 91 Uk 95 Uk95.5 Uk 96 2. 5 Uk 97 Am 97 3 .5 Uk96.5 2 .5 Am96.5Am96.5Am96.5 2 2 3 1 .5 Am 96 Am95.5Am95.5 Am 95 linear spline | stddev 0.620000 | nd modelerror 0.110000 | 2d stress 17.318763 | 2d modelerror 0.240000 2 1 2. 5 Am94.5 Am 94 2 1 1.5 In 92 1.5 Am93.5 Am 93 Am92.5 In92.5 Am 92 In93.5 1 Lk 93 Ba91.5 In 94 In94.5 In 93 Lk93.5 Lk 94 Lk 92 1 .5 Lk92.5 Lk 95 In95.5 Lk94.5 In 95 2 Ba 91Ba 91 2 Lk95.5 In 96 2 Lk 91 Lk 96 Lk 97Lk 97 Lk96.5Lk96.5 2 .5 In96.5In96.5 In 97 Figure 18: Top-down view of Skandia IC Map Independent Mean absolute error R Rank correlation 5D company vector 0.18 0.99 0.98 2D map location 0.66 0.91 0.83 Figure 19: Prediction accuracy of Skandia Map calculated using 2-fold cross-validation. The accuracy of using a 5-dimensional vector to predict fitness was also tested, to gauge the degradation in accuracy due to the low-dimensional map. The test revealed that accuracy did degrade, although the map predictions of fitness were still strongly correlated to actual values. March 20, 1999___________________________________________________ Page: 30 Figure 20: Distances on the Skandia IC Map versus the high-dimensional points. Distances on the IC map are very close to their high-dimensional counterparts, with a correlation of 0.91. Thus map positions a good representation of company differences. Application American Skandia has experienced sharp growth in the 1990s, rising from a small company of 94 employees, to a successful investment company employing over 599. The older UK Life, has experienced slower growth during the 1990s. Should UK Life adopt operating practices more like those of American Skandia? March 20, 1999___________________________________________________ Page: 31 The main difference between UK Life in 97 and 93 are drops in Customer and Renewal capital (figure 24). UK Life would need to increase its number of customers and rate of new contracts in order to get back to the position it occupied in 1993. American Skandia 1997 meanwhile, has a lower level of overall fitness, and reaching Am97 would take 2.8 years, whilst UK93 is 2.2 years away. As a result, management should conclude that it is better for UK Life to return to operating practices similar to 1993 by increasing expenditure on customer satisfaction and running programs to acquire new customers. Financial Customer Human Process Renewal UK93 1.64 1.19 1.77 -0.65 -0.39 UK97 0.41 0.27 1.62 -0.65 -0.98 UK93-UK97 +1.23 +0.93 +0.15 0 +0.59 Figure 24: Difference between UK97 and UK93 Destination Mean Standard deviations difference Time to reach location Fitness at new location (estimate) Changes required Present location 0 0 0.67 No changes UK93 0.58 2.2 years 3.30 Increase Renewal and Customer Am97 0.74 2.8 years 0.06 Increase Renewal, decrease Customer Figure 25: Time for UK Life to reach various destination on the landscape (assuming straight-line movement and no interactions). UK Life has changed its parameters at an average rate of 0.266 standard deviations per year, so to reach UK93 would require 0.5794 / 0.266 = 2.2 years to reach. March 20, 1999___________________________________________________ Page: 32 Experiment 2: UN Map Each year the World Bank Organization publishes a document titled The World Development Indicators Report, in which statistics for 140 countries over 35 years are listed. From the electronic version of this report, we obtained 64 variables with low percentages of missing values, and chose 44 countries for analysis. For purposes of the experiment, we have chosen to use an extremely simple indicator of the well-being, the “life expectancy” for citizens in these countries. The resulting landscape is shown in figure 19. The map is stretched along one dimension, with developing countries at one end, and developed countries at the other. This stretching could be caused because many of the variables are co-linear, and correlated with the same underlying cause, such as gross domestic product. One of the nice surprises about this landscape, is that many developing countries have progressed up the landscape in the 35 year measurement period. Papua New Guinea has moved from a life expectancy of 46 to 58. India and Pakistan have also both improved their life expectancies from 49 to 63 and 49 to 64 respectively. India and Pakistan have similar positions on the map (figure 26). Most of these countries, including the United States and Japan, are traveling “up” the landscape in a direction of greater life expectancy. March 20, 1999___________________________________________________ Page: 33 The country with poorest life expectancy on earth in 1998 was Sierra Leone. Sierra Leone briefly appears to be getting better from 1980-1990, however by 1995, the direction of movement is at tangent to other countries who are marching up the life- expectancy landscape, and into territory no other nation has traversed. This is probably a bad sign! . Figure 26: Fitness landscape for 44 nations from the World Bank’s World Development Indicator data, from 1970 to 1995. Height axis is life-expectancy at birth, with lighter colors representing better life expectancy. Close-up of India, Pakistan, and Sierra Leone. US and Japan are shown in the top-left climbing towards a life-expectancy peak. March 20, 1999___________________________________________________ Page: 34 Conclusion We have developed a method for organizing company knowledge into a form which people can understand, and used it to provide advanced automated decision support functions such as predicting company fitness at untested parameters (seen as the landscape around known points), what-if analysis, estimation of time and cost required to reach new states. The techniques used to construct the map are mathematically transparent and robustly minimize error in the low-dimensional representation. The maps themselves are intuitive, compelling, and graphically depict the movement of companies across the landscape. There are many areas in which these maps can be improved. Error estimates can be provided by adding error-bars to the surface. Graphically these could be represented as an “atmosphere” or “mist” that covers the landscape above and below. The rate of change of the topographic fitness landscape itself is another issue - the surface may deform over time as the economy or market changes. We hope that this geological change will be small compared to the speed of company movement so that it won’t greatly impact planning, and future work will need to calculate fair planning horizons are on particular IC maps. March 20, 1999___________________________________________________ Page: 35 References Betteridge, D., Wade, A. and Howard, A. (1985), Reflections on the modified simplex II, Talanta, Vol. 32, 8B, pp. 723-734. Bowditch, N. (1826), The New American Practical Navigator, sixth edition, Edmund L. Blunt, New York. http://www.iws.net/wier/logline.html Duch, W. and Naud, A. (1998a), Simplexes, Multi-Dimensional Scaling and Self- Organized Mapping, Technical Report, Department of Computer Methods, Nicholas Copernicus University, Poland. http://www.phys.uni.torun.pl/kmk Duch, W. and Naud, A. (1998b), Multidimensional scaling and Kohonen’s Self- organizing maps, Technical Report, Department of Computer Methods, Nicholas Copernicus University, Poland. http://www.phys.uni.torun.pl/kmk Edvinsson, L. and Malone, M. (1997) Intellectual Capital, Harper. Girosi, F., Jones, M. and Poggio, T. (1993), Priors, Stabilizers and Basis Functions: from regularization to radial, tensor and additive splines, AI Memo 1430, CBCL Paper 75, http://www.ai.mit.edu/people/girosi/home-page/memos.html http://www.iws.net/wier/logline.html http://www.phys.uni.torun.pl/kmk http://www.phys.uni.torun.pl/kmk http://www.ai.mit.edu/people/girosi/home-page/memos.html March 20, 1999___________________________________________________ Page: 36 Kaplan, R. and Norton, D. (1996), The Balanced Scorecard: Translating Strategy into Action, Harvard Business School Press. MA. Karlgaard, R. (1993), Rest in Peace, Book Value, Forbes ASAP, October 25, pp. 9. Karur, S. and Ramachandran, P. (1995), Augmented Thin Plate Spline Approximation in DRM, Boudnary Elements Communications, Vol. 6, pp. 55-58. http://wuche.wustl.edu/~karur/papers.html Kohonen, T. (1996), Self-Organizing Maps, second edition, Springer Series in Information Sciences, Vol. 30, Springer, Berlin. Kruskal, J. (1978), Multidimensional Scaling, Sage University series, Beverly Hills, CA. Li, S. (1993), Dimensionality Reduction using The Self-Organizing map, Honours Thesis, James Cook University, North Queensland. http://www.cs.jcu.edu.au/ftp/pub/ Li, S., de Vel, O., Coomans, D. (1995), Comparative Performance Analysis of Non-linear Dimensionality Reduction Methods, Proceedings of the Fifth International Workshop on Artificial Intelligence and Statistics, Fort Lauderdale, Florida. http://www.cs.jcu.edu.au/ftp/pub/techreports/94-8.ps.gz http://wuche.wustl.edu/~karur/papers.html http://www.jcu.edu.au/ http://www.cs.jcu.edu.au/ftp/pub/techreports/94-8.ps.gz March 20, 1999___________________________________________________ Page: 37 von der Malsburg (1973), Self-organization of orientation sensitive cells in striate cortex, Kybernetik, Vol. 14, pp. 85-100. Nelder, J. and Mead, R. (1965), A simplex method for function minimization, Computer Journal, Vol. 7, pp. 308-313. Norusis, M. (1997), Multidimensional scaling Examples, SPSS Professional Statistics 7.5 User Manual. Orr, M. (1996), Introduction to Radial Basis Function Networks, Centre for Cognitive Science, University of Edinburgh, 2 Buccleuch Place, Edinburgh EH8 9LW, Scotland, http://www.cns.ed.ac.uk/people/mark/intro/intro.html Shepard, R. (1974), Representation of structure in similarity data, Psychometrika, Vol. 39, pp. 373-421. Torgerson, W. (1958), Multidimensional scaling, in P. Colgan (ed), Quantitative ethology, pp. 175-217., Wiley, NY. Young, F. and Hamer, R. (1987), Multidimensional scaling: History, theory and applications, Lawrence Erlbaum, NJ. http://www.cns.ed.ac.uk/people/mark/intro/intro.html March 20, 1999___________________________________________________ Page: 38 Young, F. (1998), Multidimensional scaling, Lecture notes, University of North Carolina, http://forrest.psych.unc.edu/teaching/p230/p230.html World Development Indicators 1998 CD-ROM, The World Bank, ISBN: 0-8213-4375-0, http://www.worldbank.org/html/extpb/wdi99.htm Comparative Civilian Labor Force Statistics, Ten Countries, 1959-1999, US Department of Labor, Bureau of Labor Statistics, Office of Productivity and Technology, May 27, 1999. http://stats.bls.gov/flsdata.htm An Introduction to Business Valuation, Madison Valuation Associates, 1999. http://www.madval.com/introduction.html http://forrest.psych.unc.edu/teaching/p230/p230.html http://www.worldbank.org/ http://stats.bls.gov/flsdata.htm March 20, 1999___________________________________________________ Page: 39 
work_utk4nc3tvjcdfdaejpaqxr2xna	----	Multi-instance genetic programming for web index recommendation Expert Systems with Applications 36 (2009) 11470–11479 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Multi-instance genetic programming for web index recommendation A. Zafra a, C. Romero a, S. Ventura a,*, E. Herrera-Viedma b a Dept. of Computer Sciences and Numerical Analysis, University of Córdoba, Campus de Rabanales, edificio Albert Einstein, 14071 Córdoba, Spain b Dept. of Computer Sciences and Artificial Intelligence, University of Granada, Periodista Daniel Saucedo Aranda s/n, 18071 Granada, Spain a r t i c l e i n f o a b s t r a c t Keywords: Grammar-guided genetic programming Multiple instance learning User modelling Web mining 0957-4174/$ - see front matter � 2009 Elsevier Ltd. A doi:10.1016/j.eswa.2009.03.059 * Corresponding author. Tel.: +34 957212218; fax: E-mail addresses: azafra@uco.es (A. Zafra), cr sventura@uco.es (S. Ventura), viedma@decsai.ugr.es ( This article introduces the use of a multi-instance genetic programming algorithm for modelling user preferences in web index recommendation systems. The developed algorithm learns user interest by means of rules which add comprehensibility and clarity to the discovered models and increase the quality of the recommendations. This new model, called G3P-MI algorithm, is evaluated and compared with other available algorithms. Computational experiments show that our methodology achieves competitive results and provide high-quality user models which improve the accuracy of recommendations. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction In the last few years, the quantity of information available on Internet has been growing so rapidly that it now exceeds human processing capabilities. Users feel overwhelmed by the amount of information available and are usually unable to locate really rele- vant information that suits their individual needs in a limited amount of time. In this situation, there is a pressing need for tools that anticipate the preferences of users and provide recommenda- tions about whether or not a particular item will be of interest to the user. Such systems, referred to in the literature as recommen- dation systems (Felfernig, Friedrich, & Schmidt-Thieme, 2007), have features similar to traditional information retrieval ap- proaches but differ from them, especially in the use of models that contain information about user tastes, preferences and needs. This information differs according to the type of processing performed by the system. So, in collaborative filtering recommender systems (Schafer, Herlocker, & Sen, 2007) this model reflects similar users’ preferences or needs, while in content-based recommender sys- tems (Pazzani & Billsus, 2007) this information maps the relation- ship between the items to be recommended and the preferences of a given user. In modelling user preferences, an interesting problem is the classifying of web index pages into two categories (according to whether or not they are pertinent for a user), because this allows us to build a user model for a content-based recommendation sys- tem. The main difficulty in this problem lies in training set repre- sentation; web index pages are those which contain references or brief summaries of other pages and where there is a different num- ll rights reserved. +34 957218630. omero@uco.es (C. Romero), E. Herrera-Viedma). ber of references on each page. Moreover, the information available about the user is imprecise. We know if the user is interested in an index page or not, instead of determining exactly which concrete links the user really considers to be of interest. Recently, Zhou, Jiang, and Li (2005) have solved the problem from a multi-instance learning perspective, adapting the well known k-Nearest Neighbor (k-NN) algorithm to this new learning framework. Experimental results show that this approach greatly improves supervised learn- ing algorithm approaches. In spite of the interesting results reported by Zhou et al. (2005), their proposal presents two major limitations. The first one is re- lated to sparsity and to scalability, as the k-NN algorithm requires computations that grow linearly with the number of items, which makes it hard to scale when the number of items is high and maintain reasonable prediction performance and accuracy. The second one is related to the interpretability of new-found knowl- edge. The K-NN algorithm is a black box algorithm, that is, it sim- ply classifies web index pages as being ‘‘of interest” or ‘‘not of interest”, without providing additional information about user preferences. This is not a desirable property in recommendation systems, where any information that allows us to learn more about the interest of the user is of outmost interest for facilitating new recommendations. To overcome the aforementioned drawbacks, we propose the use of G3P-MI, a grammar-guided genetic programming algorithm for multiple instance learning. This algorithm learns prediction rules which provide information on whether any of the links con- tained on a given web index page are of interest to a given user. Experimental results concerning several benchmarks show that this approach obtains competitive results in terms of accuracy, re- call and precision. Moreover, it adds comprehensibility and clarity to the knowledge discovery process which is such an important characteristic for obtaining high predictive accuracy since the mailto:azafra@uco.es mailto:cromero@uco.es mailto:sventura@uco.es mailto:viedma@decsai.ugr.es http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa A. Zafra et al. / Expert Systems with Applications 36 (2009) 11470–11479 11471 system’s results can be interpreted easily (understandable user models) and this data can be used to obtain further information about the user thus generating even more appropriate recommendations. The rest of this paper is organized as follows. Section 2 is de- voted to introducing the multi-instance learning paradigm, and Section 3 describes the proposed G3P-MI algorithm. Section 4 pre- sents Web Index Recommendation as a multi-instance learning problem. Sections 5 and 6 presents and analyses the experimental results of our system. Finally, Section 6 presents conclusions and future work. 2. Multiple instance learning The term Multiple Instance Learning was coined by Dietterich, Lathrop, and Lozano-Perez (1997) when investigating a qualitative structure–activity relationship problem. In this problem, the task consisted of determining if a given substance does or does not present pharmacological activity in information about its molecu- lar structure. The difficulty of this task is due to the fact that a sub- stance can present more than one spatial configuration, each of which showing different structural properties. Because conforma- tions can most naturally be represented as fixed-length vectors of attribute values (or instances), the most convenient form for representing these learning examples seems to be in collections of these instances, with one associated class label representing the concept that we can learn. This is called ‘‘multiple instance” and contrasts with the manner of representing examples in super- vised learning, where each example contains only one labeled instance. To solve this last problem, Dietterich et al. (1997) proposed as the learning hypothesis that an example should be considered po- sitive (that is, it represents the concept that we can learn) if it con- tains at least one instance that represents this concept. On the other hand, an example should be considered negative if it does not contain any instances representing the concept of learning. Using this learning hyphothesis, they developed three Axis-Parallel Rectangle (abbreviated as APR) algorithms, which attempt to search for appropriate axis-parallel rectangles constructed by the conjunction of their features. Their best performing algorithm (iterated-discrim) starts with a point in the feature space and ‘‘grows” a box with the goal of finding the smallest box that can cover at least one instance from each positive bag and no instances from any negative bags. The resulting box was then expanded (via a statistical technique) to get optimum results. Following Dietterich et al.’s study, Auer (1997) tries to avoid some potentially difficult computational problems that were re- quired by the heuristics used in the iterated-discrim algorithm, presenting a theoretical algorithm that does not require product distribution, MULTINST. With a new approach, Maron and Loz- ano-Pérez (1997) proposed one of the most famous multi-instance learning algorithms, Diverse Density (DD). The diverse density of a point, p, in the feature space is defined as a probabilistic measure taking into consideration how many different positive bags have an instance near p, and how far the negative instances are from p. This algorithm was combined with the Expectation Maximiza- tion (EM) algorithm, resulting in EM-DD (Zhang & Goldman, 2001), a general-purpose MI algorithm whose basic premise is to show which instance corresponds to the bag labeled as a missing attribute which can be estimated using the EM approach. Recently, Pao, Chuang, Xu, and Fu (2008) have proposed an EM based learn- ing algorithm to provide a comprehensive procedure to maximize the measurement of DD on given multiple instances. In 1998, Long and Tan (1998) described a polynomial-time the- oretical algorithm showing that if the instances in the bags drawn from product distribution are independent, then the APR is PAC- learnable. Continuing with PAC-learnable research, Kalai and Blum (1998) described a reduction in the problem of PAC-learning in the MIL framework as compared to PAC-learning with one- sided random classification noise, and presented a theoretical algorithm with less complexity than the algorithm described in Auer (1997). The first approaches using lazy learning, decision trees and rule learning were studied during the year 2000. In the lazy learning context, Wang and Zucker (2000) proposed two variants of the K-nearest neighbor algorithm (kNN) that they referred to as Cita- tion-kNN and Bayesian-kNN, these algorithms extending the K- nearest neighbor algorithm for MIL, adopting Hausdorff distance. With respect to decision trees and learning rules, Zucker and Che- valeyre (2000) implemented ID3-MI and RIPPER-MI, which are multi-instance versions of decision tree algorithm ID3 and rule learning algorithm RIPPER, respectively. At that time, Ruffo (2000) presented a multi-instance version of the C4.5 decision tree, which was called RELIC. There are also many other supervised learning algorithms which have been adapted to MIL. Thus, we can find the contribu- tion of Ramon and De Raedt (2000) which extends standard neural networks to MIL. After this work, further studies appeared improv- ing or extending it (Chai & Yang, 2007; Zhang, Jack, & Nandi, 2005; Zhang & Zhou, 2004, 2006). Another approach that has been adapted to the MIL framework is Support Vector Machines (SVM). There are numerous proposals in these approaches, Gärtner, Flach, Kowalczyk, and Smola (2002) adapted kernel methods to work with MIL data by modifying the kernel distance measures to handle sets. Using a related approach, Chen and Wang (2004) and Chen et al. (2006) adapted SVMs by modifying the form of the data rather than changing the underlying SVM algorithms while Andrews, Tsochantaridis, and Hofmann (2002) adapted the SVM kernels directly to produce one of the best MIL classification systems currently available. Recently, we can also find the propos- als by Mangasarian and Wild (2008) and Gu et al. (2008). Finally, there are works such as those by Zhang et al. (2005) and Zhou and Zhang (2007) that show the use of ensembles to enhance mul- ti-instance learners. 3. Grammar-guided genetic programming for multiple instance learning In this section we introduce G3P-MI, a grammar-guided genetic programming algorithm for multi-instance learning. In the next sections, we will introduce the following design aspects: individual representation, genetic operators, fitness function and evolution- ary process. 3.1. Individual representation In G3P-MI, an individual represents rules that determine if a gi- ven pattern should be considered positive (that is, is an example of the concept we want to represent) or negative (if it is not): if ðcondBðbÞÞ then pattern b is an instance of the concept: else pattern b is not an instance of the concept: end if ð1Þ where condB is a condition that is applied to the pattern. Consider- ing the Dietterich hypothesis, which states that a pattern is positive if any of its instances represent the concept that we want to learn and negative otherwise, we could represent condB as: 11472 A. Zafra et al. / Expert Systems with Applications 36 (2009) 11470–11479 condBðbÞ¼ true; 9i 2 ½1; sizeðbÞ�=condIðinstanceði; bÞÞ¼ true false; otherwise: � ð2Þ where sizeðbÞ returns the number of instances in pattern b, instanceði; bÞ is a function that returns the ith instance to bag b and condI is a condition that is applied to an instance contained in a given bag. Considering the properties of the disjunction opera- tor, Eq. (2) can be rewritten as condBðbÞ¼ _ i¼1;sizeðbÞ condIðinstanceði; bÞÞ ð3Þ where _ is the disjunction operator. As can be seen, condI is the only variable part in Eqs. (1)–(3) that can experiment an evolutionary process. So, in G3P-MI, an individual’s genotype is a syntax tree that contains the code of function condI , while the individual’s pheno- type is the whole rule that is applied to the bags (Eq. 1). Fig. 1 shows the context free grammar that represents these individuals in a gen- eral form. As can be seen, the code of the condition can contain one or several valid clauses, that check conditions related to instances contained in patterns. In the case of there being more than one clause, they can be combined by the conjunctions or disjunctions found in any order. The format of the clauses depends on the type of data contained in the instances analyzed (in Section 5 we will de- scribe the format of the clauses used in the web index recommen- dation problem). 3.2. Initialization To initialize the population in the algorithm, the procedure used is inspired by that defined by Geyer-Schulz (1995). This procedure builds a new syntax tree at random given the maximum number of derivations. To guarantee that the syntax tree generated is valid and uses a maximum number of derivations, the system calculates a selection probability for each symbol in the grammar for a spec- ified number of available derivations. This table of probabilities, although implying some computational input, only has to be calcu- lated once since it is saved with the rest of the structural informa- tion about individuals. To guarantee greater diversity in the number of individuals gen- erated, the initialization procedure generates individuals of differ- ent sizes, with two parameters as the minimum number of derivations and maximum number of individuals in the population. 3.3. Genetic operators G3P-MI uses two genetic operators, called respectively selective crossover and selective mutation, to generate new individuals in a given generation of the evolutionary algorithm. Both operators were proposed by Whigham (1996) in the definition of grammar- based genetic programming. In this section, we will briefly de- scribe their basic principles and functioning. 3.3.1. Selective crossover This operator creates new programs by mixing the contents of two parent programs. To do so, a non-terminal symbol is chosen at random and two sub-trees (one from each parent) are selected whose roots coincide with the symbol adopted. Fig. 2 shows how this operation is performed. Fig. 1. Grammar used for representing individuals’ genotypes in G3P-MI. Selective crossover presents several configuration parameters. On one hand, a list of eligible symbols can be defined in order to increase the probability of crossover for certain symbols and lessen that probability for certain others. On the other hand, in order to reduce bloating (Banzhaf, Francone, Keller, & Nordin, 1998), there is a parameter that defines the largest size possible for offspring generated in a crossover. Surpassing this size, the operations will reproduce the original parents which would also occur if one of the parents does not possess the symbol in question. 3.3.2. Selective mutation Mutation is associated with a small change in the structure of the representation it is applied to. As can be seen in Fig. 3, this is achieved by randomly selecting a sub-tree within the individual to be mutated, and replacing this sub-tree with a new randomly- generated one. The procedure used to generate this sub-tree is the same as that used to create new individuals. As in the case of selective crossover, this operator presents two configuration parameters: on one hand, the list of eligible non-ter- minal symbols as the root of the sub-tree to be mutated, and on the other, a maximum size for the offspring generated. 3.4. Evolutionary algorithm G3P-MI follows the structure of a classical generational and elit- ist evolutionary algorithm. First of all, there is the creation of an initial population, following the procedure described in Section 3.2. Once the individuals are evaluated, we encounter the main loop of the algorithm composed of the following operations: (1) Parents selection. The procedure followed to choose the indi- viduals to be reproduced by crossover and/or mutation is roulette selection. (2) Parents reproduction. Once the parents are obtained, the crossover operator is applied with a certain probability, and later on the mutation operator as well, also with a deter- mined probability. The offspring obtained through this pro- cedure are then evaluated. (3) Population update. The population is updated by direct replacement, that is, the resulting offspring replace the pres- ent population. To guarantee that the best individual in the population is not lost during the updating process, the algo- rithm employs elitism. Finally, there are two conditions for exiting from the main loop. On one hand, the algorithm ends if the maximum number of genera- tions defined by the user is surpassed and, on the other, it also ends if the best individual in the population achieves the quality objec- tives indicated by the user. 4. Web index recommendation: a multiple instance problem Web Index Pages are pages that provide titles or brief summa- ries of other pages. These pages contain a lot of information through references, leaving detailed presentations to their linked pages. An example of a web index page is http://health.yahoo.com as shown in Fig. 4. The web index recommendation problem consists of building a model to establish exactly which web page index it is that interests a given user from among the contents of a myriad of web index pages that have already been labeled as being ‘‘of interest” or ‘‘not of interest” for this particular user. This problem is more dif- ficult than other analogous ones related to the construction of user models for content-based recommendation systems (Mooney, Ray- mond, & Loriene, 2000; Pazzani & Billsus, 2007) since, in this case, http://health.yahoo.com Parent 1 Parent 2 <cond > OR <cmp> <cond > <op> <term-name> <term-freq> LT term1 value1 <cmp> <op> <term-name> <term-freq> <cond > AND <cmp> <cond > <op> <term-name> <term-freq> <op> <term-name> <term-freq> GE term4 value4 AND <cond > <cmp> <op> <term-name> <term-freq> <cmp> Offspring 1 Offspring 2 LE term2 value2 GT term3 value3 GE term5 value5 GE term5 value5 <cond > OR <cmp> <cond > <op> <term-name> <term-freq> LT term1 value1 <op> <term-name> <term-freq> GE term4 value4 <cmp> <cond > AND <cmp> <cond > <op> <term-name> <term-freq> AND <cond > <cmp> GT term3 value3 GE term5value5 <cmp> <op> <term-name> <term-freq> LE term2 value2 <op> <term-name> <term-freq> Fig. 2. Selective crossover. Parent Offspring <op> <term-name> <term-freq> LT term1 value1 <op> <term-name> <term-freq> <op> <term-name> <term-freq> <op> <term-name> <term-freq> <op> <term-name> <term-freq> LE term2 value2 AND <cond > <cmp> GT term3 value3 LT term1 value1 GE term4 value4 <cmp> <comp> OR <cmp> <cond > <cond > <cond > >OR <cmp> <cond Fig. 3. Selective mutation. A. Zafra et al. / Expert Systems with Applications 36 (2009) 11470–11479 11473 the information on the user’s preferences has to do with web index pages as a whole and not with the items contained on the page in question. For this reason, supervised learning algorithms could not produce results that met with expectations (Zhou et al., 2005). On observing this problem in detail, it is seen to adjust perfectly to multi-instance representation. In fact, the web index page could be represented by a bag, and each of the connected pages is an example of a bag. Moreover, Dietterich’s hypothesis also seems Fig. 4. Web index recommendation problem. Table 1 Experimental data sets. Data sets V1 V2 V3 V4 V5 V6 V7 V8 V9 Training Positive 17 18 14 56 62 60 39 35 37 Negative 58 57 61 19 13 15 36 40 38 Test Positive 4 3 7 33 27 29 16 20 18 Negative 34 35 31 5 11 9 22 18 20 hese datasets and a more detailed description of them can be downloaded from llowing Internet address: http://cs.nju.edu.cn/zhouzh/zhouzh.files/publication/ x/milweb-datafile.htm. 11474 A. Zafra et al. / Expert Systems with Applications 36 (2009) 11470–11479 appropriate for resolving this problem. A web index page is consid- ered to be of interest if at least one of the links on it leads to a page of interest for the user, while a page is not of interest if none of the pages indicated are useful for the user. Based on previously mentioned considerations, we have consid- ered the use of multi-instance representation for web index recom- mendation problems. Two different codification schemes normally used to categorize documents (Sebastiani, 2002) have been taken into consideration to build this type of instance. The first scheme is representing a page as a vector of Boolean values, each of which representing the presence (true) or absence (false) of a term in the document. The second scheme represents the page as a vector of integer values, each element represents the absolute frequency of the appearance of term on the page. As seen in Section 5, each of these coding schemes requires learning rules with different com- parison operators and thus the result is two different versions of the G3P-MI algorithm. 5. Experimental setup This section describes the data sets that have been used in the experimentation as well as several especially relevant methodo- logical and configuration aspects. 5.1. Data sets In order to evaluate our G3P-MI algorithm in the web index rec- ommendation task, and to compare with other results previously published, we have used the data sets prepared by Prof. Z.H. Zhou’s team. These datasets have been used in a previous paper about cat- egorization of web index pages (Zhou et al., 2005). There are nine data sets and in each a different volunteer labeled 113 web index pages according to his/her interests. For each data set, 75 web in- dex pages are randomly selected as training bags while the remain- ing 38 index pages are used as test bags. Table 1 summarizes the information about these data sets.1 To be able to use the original datasets, it has been necessary to preprocess information in order to adapt the original data to the format represented in our algorithms. In fact, although Zhou et al. represent web index pages as bags with one or various in- stances, they represent each instance by a list containing the N terms that most frequently appear on the page (where N takes val- ues between 5 and 15). In our case, each instance is represented in vector format, where each component of the vector is related to a term in the corpus of available documents. The concrete value of each component can be binary (if we opt for Boolean representa- tion) or whole if we opt for representation based on absolute frequencies. Both these data sets are characterized by a high degree of dimensionality and great sparsity. In order to reduce both effects and probably also thus improve the yield of the algorithm, the ori- ginal sets were preprocessed to eliminate the terms that appear only in a few documents and others that appear in most of them. In both cases, term discriminating capacity is so limited that their elimination does not reduce the information available. After a ser- ies of preliminary tests, it was decided to eliminate those terms that appeared in less than 3 documents or more than 40. In this way the total number of attributes is reduced from 5763 to 600. In consequence, there are 4 different types of data sets: Boolean all. These 9 data sets use Boolean representation (pres- ence or absence of a term in a document), and each document vector presents a size of 5763 components. It is considered all terms belonging to the original corpus of documents. 1 T the fo anne http://cs.nju.edu.cn/zhouzh/zhouzh.files/publication/annex/milweb-datafile.htm http://cs.nju.edu.cn/zhouzh/zhouzh.files/publication/annex/milweb-datafile.htm Fig. 6. Grammar used for learning from numerical (term frequency) datasets. Table 2 Configuration parameters for all G3P-MI versions. General Population size = 1000 individuals A. Zafra et al. / Expert Systems with Applications 36 (2009) 11470–11479 11475 Boolean filtered. These 9 data sets use Boolean representation (presence or absence of a term in the document), and each doc- ument vector is made up of 600 components. It is considered the terms from the corpus that appear in more than 3 and less than 40 documents. Frequency all. These 9 data sets use numeric representation (fre- quency of appearance of a term in a document), and each doc- ument vector is composed of 5763 elements. It is considered all terms from the original corpus of documents. Frequency filtered. These 9 data sets use numerical representa- tion (frequency of appearance of a term in a document), and each document vector has 600 components. It is considered the terms that appear in more than 3 and less than 40 docu- ments of the corpus. As will be seen now, one of the objectives of this study is to eval- uate which of the representation schemes (and associated versions of the G3P-Ml algorithm) produces the best results. Then, we will compare our best version with others algorithms which have solved the same problem. 5.2. Algorithm details This section presents details about the adaptation of the G3P-MI algorithm for the task of recommendation of web index pages. First of all there are algorithm variants that have been defined to resolve this problem using, respectively, boolean and integer representa- tions. The fitness function used in all the tests done is then ex- plained. Finally there is a brief presentation of other configuration parameters of interest. 5.2.1. G3P-MI variants As has already been mentioned, two variants of the G3P-MI algorithm have been specially developed for the two types of rep- resentations. Actually these variants only differ in the format of the prediction rules learned in the evolutionary process. Thus in the case of boolean representation, the rules generated refer to the presence or absence of a term on a given page, which respectively uses the functions ‘‘CONTAINS” and ‘‘DOES NOT CONTAIN”, whose operator is the name of any of the terms that appear in the corpus of training documents (see Fig. 5). When the representation of the pages is numerical (based on frequencies), the rules generated inform about whether a given term appears on a given page more or less frequently than a de- fined threshold. In this case the comparison operators ‘‘GT” (great- er than), ‘‘GE” (greater than and equal to), ‘‘LT” (less than) and ‘‘LE” (less than and equal to) are used to compare the frequency of appearance of a term in the document with the threshold (an inte- ger value). Fig. 6 shows the grammar that represents this type of rules. 5.2.2. Fitness function The problem of developing good metrics to measure the effec- tiveness of recommendations is not a trivial task (Herlocker, Kon- stan, Borchers, & Riedl, 1999, 2004; Yang & Padmanabhan, 2005). To begin with, as it deals with a classification task, it would be log- ical to think that the use of such measures as accuracy (the percent- Fig. 5. Grammar used for learning from boolean datasets. age of web index pages that are correctly classified by the classifier) would be appropriate. However, in recommender systems it is very common to work with unbalanced data sets, where the number of pages of interest is much lower than the number of pages that are not of interest. In this case, by solely using accuracy, the classifier obtains better re- sults if no examples of the minority class (interesting examples) are classified than if only some are classified and mistakes are made with respect to the bigger class (uninteresting examples). For that reason, other measures are considered, such as precision, the percentage of accepted documents that are in fact relevant to a given user, that is precision ¼ #interesting and recommended #recommended ð4Þ In other cases, the problem is that there are many more interesting examples than uninteresting ones. In this case, precision is not the best measure to take into account, and it is necessary to use another measure such as recall recall ¼ #interesting and recommended #interesting ð5Þ For the above-mentioned reasons, we have tried to build a fitness function combining accuracy, precision and recall. Of all the combi- nations tested, the product has been that which produced the best results since it weighs the 3 measurements equally, as well as penalizing those individuals with a value of 0 in any of the above measurements. In consequence, fitness ¼ accuracy � recall � precision ð6Þ has been the fitness function used in all the experiments. 5.2.3. Other algorithm parameters Table 2 shows the configuration parameters used in the exper- iments carried out with all the versions of the G3P-MI algorithm. As can be seen, the only difference is the depth of the minimum depth used, which is less in the Boolean version than in the numer- ical one (perfectly logical taking into account that the number of substitutions needed to create a valid syntax tree is lower than in the Boolean version). All of the algorithms used have been devel- oped in Java, using the Java Class Library for Evolutionary Compu- tation (Ventura, Romero, Zafra, Delgado, & Hervás, 2008). Maximum of generations = 100 Initialization Maximum tree depth = 15 Minimum tree depth = 5/6 Crossover Crossover probability = 0.95 Eligible symbols = all non-terminals Maximum depth for sons = 15 Mutation Mutation probability = 0.05 Eligible symbols = all non-terminals Maximum depth for sons = 15 Table 4 Average rankings of the algorithms. Algorithm Ranking Accuracy Recall Precision Boolean all 2.6111 2.7777 2.7777 Boolean filtered 2.3888 3.3333 1.8888 Frequency all 2.8333 2.2777 2.8888 Frequency filtered 2.1666 1.6111 2.4444 11476 A. Zafra et al. / Expert Systems with Applications 36 (2009) 11470–11479 6. Results and discussion We carry out two types of experiments. The first experiment compares the performance of our proposals with respect to the problem of Web Index Recommendation. The second experiment compares the performance of our best algorithm to other classifi- cation techniques to solve this problem. This section describes these experiments and the results obtained. Also, at the end of the section we will comment on the type of knowledge discovered with G3P-MI algorithms. 6.1. Comparing different versions of G3P-MI As mentioned previously, the objective of this first experiment was to analyze which variant of G3P-MI yields the best results in the web index recommendation task. To do so, we have executed every variant of the G3P-MI algorithm 10 times, with different ran- dom seeds, for the nine available data sets. Table 3 shows average values for accuracy, precision and recall. To determine if there are significant differences between the re- sults shown in Table 3, we have performed a Friedman test (Dem- sar, 2006). This is a non-parametric test that compares the average ranks of the algorithms, where the algorithm with the best results for a dataset is given a rank of 1 and the algorithm with the worst results is given a ranking of 4 (the ranking averages of all the avail- able datasets are shown in Table 4). According to the Friedman test results, we accept the null-hypothesis, that is, with the available information, there do not exist significative differences between the different variants of G3P-MI analyzed. In spite of the results yielded by the Friedman test, an inspec- tion of Table 4 reveals that the versions that use the filtering oper- ation (boolean filtered and frequency filtered) give slightly better results than their counterparts that do not have this preprocessing operation. This result is consistent with the fact that a reduction in search space facilitates the recommendation problem and achieves better results. Another interesting question that can be deduced from the analysis of Table 4 is that, in general, boolean versions of G3P-MI seem to yield more precise results, while numerical versions of the algorithm are better in recall results. This seems to indicate that the algorithm has not been able to reach a complete trade- off between the metrics that are in conflict. Table 3 Results obtained with four different versions of G3P-MI (see text for details). Algorithm Accuracy V1 V2 V3 V4 Boolean all 0.8158 0.8684 0.8421 0.8684 Boolean filtered 0.8421 0.7368 0.8421 0.8421 Frequency all 0.9211 0.7368 0.8684 0.8947 Frequency filtered 0.8684 0.9211 0.8684 0.8947 Precision V1 V2 V3 V4 Boolean all 0.2857 0.2500 0.5714 0.8684 Boolean filtered 0.3333 0.2308 0.5714 0.9355 Frequency all 1.0000 0.2308 0.7500 0.8919 Frequency filtered 0.4286 0.5000 0.6250 0.8919 Recall V1 V2 V3 V4 Boolean all 0.5000 0.3333 0.5714 1.0000 Boolean filtered 0.5000 1.0000 0.5714 0.8788 Frequency all 0.2500 1.0000 0.4286 1.0000 Frequency filtered 0.7500 1.0000 0.7143 1.0000 6.2. Comparing G3P-MI with other proposals In this section, we compare our algorithm with other proposals which have resolved this problem. These proposals are Frecit-kNN, Citation-kNN and Txt-kNN (Zhou et al., 2005). Txt-kNN is obtained through adapting the standard kNN algorithm to textual objects. Citation-kNN is similar to the latter algorithm but considers both the references and the citers of an unseen object in prediction. Fre- cit-kNN is similar to Citation-kNN, but is a multi-instance learning algorithm while the former is a single-instance learning algorithm. After analyzing our algorithm in the previous section, we choose the configuration which uses filtered data sets to compare these techniques. Table 5 shows a summary with the average values ob- tained by different algorithms using different data sets. In order to make an empirical comparison between the different methods, we use Friedman test (Demsar, 2006) on the results re- ported in Table 5. The results of applying the Friedman test are re- ported in Table 6, for precision, recall and accuracy measurements. The test accepts the null-hypothesis for accuracy values and so we can determine that there are no significant differences for this mea- surement. For precision and recall measurements, the test rejects the null-hypothesis, indicating the existence of significant differ- ences between the results of different algorithms. Due to these re- sults, a posteriori statistical analysis is needed. In Fig. 7a and b, we show the application of the Bonferroni–Dunn test. We mark the interval of one critical difference (CD) to the left and right in Fig. 7. Any algorithm ranking outside this area is significantly differ- ent from the control algorithm (the control algorithm is that with the lowest ranking). We can see that with recall value, our proposal obtains the best value and we can assert that five algorithms are worse than our proposal because they rank outside the (CD) value. In the case of precision value, Frecit-kNN obtains the best value, and V5 V6 V7 V8 V9 0.7632 0.7895 0.7632 0.6842 0.6316 0.8421 0.7895 0.8421 0.6053 0.6842 0.7895 0.7632 0.4211 0.5263 0.4737 0.7632 0.8158 0.4737 0.5526 0.5526 V5 V6 V7 V8 V9 0.7500 0.7838 0.6522 0.6429 0.5667 0.8889 0.8621 0.9167 0.5758 0.6875 0.7714 0.7632 0.4211 0.5263 0.4737 0.7500 0.8056 0.4444 0.5405 0.5143 V5 V6 V7 V8 V9 1.0000 1.0000 0.9375 0.9000 0.9444 0.8889 0.8621 0.6875 0.9500 0.6111 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 Table 5 Results summary. Algorithm Accuracy V1 V2 V3 V4 V5 V6 V7 V8 V9 Txt-kNN 0.8632 0.7948 0.7262 0.8894 0.7318 0.7946 0.5524 0.5896 0.5688 Citation-kNN 0.8788 0.7682 0.7632 0.8156 0.7578 0.8156 0.6842 0.6842 0.6528 Fretcit-kNN 0.9106 0.8682 0.8576 0.8474 0.8526 0.8632 0.7578 0.6898 0.6476 Boolean filtered 0.8421 0.7368 0.8421 0.8421 0.8421 0.7895 0.8421 0.6053 0.6842 Frequency Txt-kNN 0.8738 0.7422 0.7682 0.8948 0.7370 0.8054 0.6474 0.6002 0.5526 Frequency Citation-kNN 0.8526 0.8104 0.8368 0.8630 0.6686 0.8158 0.6792 0.7054 0.6370 Frequency Fretcit-kNN 0.9002 0.8366 0.8734 0.8264 0.8052 0.8000 0.7738 0.7160 0.7052 Frequency filtered 0.8684 0.9211 0.8684 0.8947 0.7632 0.8158 0.4737 0.5526 0.5526 Precision V1 V2 V3 V4 V5 V6 V7 V8 V9 Txt-kNN 0.3700 0.2710 0.3648 0.9046 0.7346 0.7886 0.4778 0.5974 0.5392 Citation-kNN 0.3998 0.1998 0.4088 0.9460 0.8302 0.8486 0.6132 0.7322 0.7582 Fretcit-kNN 0.5734 0.3476 0.6202 0.9538 0.8548 0.8792 0.7280 0.7388 0.6874 Boolean Filtered 0.3333 0.2308 0.5714 0.9355 0.8889 0.8621 0.9167 0.5758 0.6875 Frequency Txt-kNN 0.4238 0.1996 0.3542 0.9054 0.7612 0.7974 0.5564 0.6010 0.5290 Frequency Citation-kNN 0.3434 0.2762 0.6700 0.9214 0.8210 0.8316 0.6158 0.7316 0.7370 Frequency Fretcit-kNN 0.5000 0.6670 0.7140 0.9346 0.8096 0.8244 0.7224 0.7878 0.8946 Frequency filtered 0.4286 0.5000 0.6250 0.8919 0.7500 0.8056 0.4444 0.5405 0.5143 Recall V1 V2 V3 V4 V5 V6 V7 V8 V9 Txt-kNN 0.4500 0.9334 0.5714 0.9758 0.9778 1.0000 0.7128 0.7300 0.7222 Citation-kNN 0.3000 0.6002 0.4004 0.8362 0.8298 0.9244 0.6626 0.6300 0.3888 Fretcit-kNN 0.6000 0.6670 0.5996 0.8664 0.9556 0.9522 0.7002 0.6410 0.4666 Boolean Filtered 0.5000 1.0000 0.5714 0.8788 0.8889 0.8621 0.6875 0.9500 0.6111 Frequency Txt-kNN 0.5500 0.7332 0.3146 0.9820 0.9186 1.0000 0.8126 0.7100 0.7002 Frequency Citation-kNN 0.3500 0.7336 0.2860 0.9210 0.6818 0.9520 0.6504 0.7000 0.4332 Frequency Fretcit-kNN 0.5334 0.2826 0.6428 0.8604 0.9482 0.9382 0.7504 0.6310 0.4332 Frequency filtered 0.7500 1.0000 0.7143 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 Table 6 Test Friedman ðp < 0:01Þ. v2 Ranking Accuracy Recall Precision 18.475 15.7314 28.5185 37.111 A. Zafra et al. / Expert Systems with Applications 36 (2009) 11470–11479 11477 in this case, the Bonferroni–Dunn test does not consider there to be differences with our proposal, with p ¼ 0:05. 5678 Txt-kNN Frequency Txt-kNN Frequency Citation-kNN Frequency Filtered Citation-kNN Boolean Filtered Frequency Citation-kNN Frequency Fretcit-kNN 5678 Fig. 7. Comparison of one classifier against the oth Summing up, according to statistical tests, Fretcit-kNN is the control algorithm for recall measures, but our proposal is not sig- nificantly different according to the Bonferroni–Dunn test. On the other hand, our proposal for the precision measurement is consid- ered the control algorithm and, in this case, the Bonferroni–Dunn test results show that both Fretcit-kNN versions are significantly different with respect to it. Therefore, from a statistical point of view, we can indicate that G3P-MI results are slightly better than Fretcit-kNN results. However, the most significant contribution of 1234 Citation-kNN Fretcit-kNN Boolean Filtered Frequency Fretcit-kNN Txt-kNN Fretcit-kNN Frequency Txt-kNN Frequency Filtered 1234 ers with the Bonferroni–Dunn test, ðp < 0:05Þ. 11478 A. Zafra et al. / Expert Systems with Applications 36 (2009) 11470–11479 our proposal is that it adds comprehensibility and clarity to the knowledge acquired. This point is dealt in more depth in the fol- lowing section. 6.3. Knowledge acquired As we noted in the introduction, the greatest advantage of the G3P-MI algorithm over other black-box proposals is the compre- hensibility of the knowledge acquired, that is presented to the user in the form of prediction rules. In this section we introduce exam- ples of these rules, and we discuss the advantages of using Boolean representation instead of frequency representation for comprehen- sibility criterion. Firstly, we show a rule obtained for the first user/dataset using Boolean representation. IF ((contains planet AND contains forecast ) OR (contains atmospheric) OR (not_contains football)) THEN Recommend page to V1 user. ELSE No recommend page to V1 user. By means of this rule we can learn what topics can be recom- mended to the user. Thus, user 1 is interested in such topics as ecology and environment (with words like planet, forecast and atmospheric) and is not interested in football. Secondly, we show a rule obtained for the first user/dataset using numerical representation. IF ((frech > 16) _ (house > 11) _ (science > 2) _ (aol 6 20 ^ on-line > 4))) THEN Recommend page to V1 user. ELSE No recommend page to V1 user. We can see that this rule is more complex because the words are limited by their frequency and it is more difficult to identify user preferences. For this, although both representations obtain similar results, after this study we can conclude that numerical represen- tation is less effective because it obtains less comprehensive rules. 7. Conclusions and future work This study describes the use of the G3P-MI algorithm for recom- mending Web Index Pages. This algorithm applies grammar- guided genetic programming to learn rules about whether or not a page referred to on a Web Index Page is of interest to a given user. To represent the Web Index Page, this algorithm applies the con- cept of multi-instances, representing the web pages as a set of in- stances where each instance represent the different referenced pages and stores information related to reference page. Two ver- sions of the algorithms have been developed, one which is applied when pages are represented in Boolean form and the other which is applied when the representation format is based on the fre- quency of the appearance of certain terms on a page. There has also been analyzed the possibility of carrying out a previous filtering of information to eliminate less discriminating terminology. The experiments carried out show that although there is no significant differences in the application of a Friedman test, the techniques using a previous information filtering produce better results than those that do not while those that use numerical representation produce a less understandable knowledge. Moreover, our proposal have been compared to the results of various versions of the kNN algorithm published by the Zhou team. The statistical tests carried out do not show significant differences in accuracy, although they do for precision and recall. In fact, our proposal is the one that produces the best results with respect to precision, while the Fretcit-kNN algorithm is that with the best re- sults for recall. Finally, some examples of the rules discovered with the G3P-MI algorithm are presented. These rules show the high de- gree of comprehensibility of the knowledge acquired with the G3P- MI algorithm as compared to in black box methods like those based on kNN. Although the results obtained are of great interest, we feel that the yield of the G3P-MI algorithm in the task of Web Index Page rec- ommendation could be improved in some ways. On one hand the algorithm has not always been able to establish an equilibrium among conflicting objectives. In this sense the problem could be considered from a multi-objective perspective to see if the balance of different objectives could be achieved more appropriately. On the other hand it has been confirmed that a reduction in the space ded- icated to characteristics can improve the yield of the algorithm. For this reason we would consider it to be of special interest to study the application of selection techniques for characteristics and com- pare the effect these would produce on the yield of our system. Acknowledgments This work has been subsidised in part by the research project SAINFOWEB (P05-TIC-00602) and the TIN2005-08386-C05-02, TIN2007-61079 and TIN2008-06681-C06-03 projects of the Span- ish Inter-Ministerial Commission of Science and Technology (CI- CYT) and FEDER funds. References Andrews, S., Tsochantaridis, I., & Hofmann, T. (2002). Support vector machines for multiple-instance learning. In NIPS’02: Proceedings of neural information processing system (pp. 561–568). Auer, P. (1997). On learning from multi-instance examples: Empirical evaluation of a theoretical approach. In ICML’97: Proceedings of the 14th international conference on machine learning (pp. 21–29). San Francisco, CA, USA: Morgan Kaufmann Publishers. Banzhaf, W., Francone, F. D., Keller, R. E., & Nordin, P. (1998). Genetic programming: An introduction. On the automatic evolution of computer programs and its applications. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc. Chai, Y.-M., & Yang, Z.-W. (2007). A multi-instance learning algorithm based on normalized radial basis function network. LNCS (Vol. 4491). Chen, Y., Bi, J., & Wang, J. (2006). Miles: Multiple-instance learning via embedded instance selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(12), 1931–1947. Chen, Y., & Wang, J. Z. (2004). Image categorization by learning and reasoning with regions. Journal of Machine Learning Research, 5, 913939. Demsar, J. (2006). Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research, 7, 1–30. Dietterich, T. G., Lathrop, R. H., & Lozano-Perez, T. (1997). Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 89(1-2), 31–71. Felfernig, A., Friedrich, G., & Schmidt-Thieme, L. (2007). Guest editors’ introduction: Recommender systems. IEEE Intelligent Systems, 22(3), 18–21. Gärtner, T., Flach, P. A., Kowalczyk, A., & Smola, A. J. (2002). Multi-instance kernels. In ICML’02: Proceedings of the 19th international conference on machine learning (pp. 179–186). San Francisco, CA, USA: Morgan Kaufmann Publishers Inc. Geyer-Schulz, A. (1995). Fuzzy rule-based expert systems and genetic machine learning (1st ed.). Heidelberg, Germany: Physica Verlag. Gu, Z., Mei, T., Tang, J., Wu, X., & Hua, X. (2008). MILC2: A multi-layer multi-instance learning approach to video concept detection (Vol. 4903). Herlocker, J., Konstan, J., Borchers, A., & Riedl, J. (1999). An algorithmic framework for performing collaborative filtering. In SIGIR’99: Proceedings of the 22nd annual international ACM conference on research and development in information retrieval, Berkeley, California, United States (pp. 230–237). Herlocker, J., Konstan, J., Terveen, L., & Riedl, J. (2004). Evaluating collaborative filtering recommender systems. ACM Transaction Information Systems, 22(1), 5–53. Kalai, A., & Blum, A. (1998). A note on learning from multiple-instance examples. Machine Learning, 30(1), 23–30. Long, P. M., & Tan, L. (1998). PAC learning axis-aligned rectangles with respect to product distributions from multiple-instance examples. Machine Learning, 30(1), 7–21. Mangasarian, O. L., & Wild, E. W. (2008). Multiple instance classification via successive linear programming. Journal of Optimization Theory and Applications, 137(3), 555–568. A. Zafra et al. / Expert Systems with Applications 36 (2009) 11470–11479 11479 Maron, O., & Lozano-Pérez, T. (1997). A framework for multiple-instance learning. NIPS’97: Proceedings of neural information processing system (Vol. 10, pp. 570–576). Cambridge, MA, USA: MIT Press. Mooney, Raymond, J., & Loriene, R. (2000). Content-based book recommending using learning for text categorization. In Proceedings of the ACM international conference on digital libraries (pp. 195–204). Pao, H. T., Chuang, S. C., Xu, Y. Y., & Fu, H. . (2008). An em based multiple instance learning method for image classification. Expert Systems with Applications, 35(3), 1468–1472. Pazzani, M. J., & Billsus, D. (2007). Content-based recommendation systems. In P. Brusilovsky, A. Kobsa, & W. Nejdl (Eds.), The adaptive web: Methods and strategies of web Personalization. Lecture notes in computer science (Vol. 4321, pp. 325–341). Springer. Ramon, J., & De Raedt, L. (2000). Multi-instance neural networks. In Proceedings of the ICML-workshop on attribute-value and relational learning. Ruffo, G. (2000). Learning single and multiple instance decision tree for computer security applications. Ph.D. thesis, Department of Computer Science. University of Turin, Torino, Italy. Schafer, J. B., Herlocker, D. F. J., & Sen, S. (2007). Collaborative filtering recommender systems. In P. Brusilovsky, A. Kobsa, & W. Nejdl (Eds.), The adaptive web: Methods and strategies of web personalization. Lecture notes in computer science (Vol. 4321, pp. 291–324). Springer. Sebastiani, F. (2002). Machine learning in automated text categorization. ACM Computing Surveys, 34(1), 1–47. Ventura, S., Romero, C., Zafra, A., Delgado, J. A., & Hervás, C. (2008). JCLEC: A java framework for evolutionary computation soft computing. Soft Computing, 12(4), 381–392. Wang, J., & Zucker, J.-D. (2000). Solving the multiple-instance problem: A lazy learning approach. In ICML’00: Proceedings of the 17th international conference on machine learning (pp. 1119–1126). San Francisco, CA, USA: Morgan Kaufmann Publishers Inc. Whigham, P. A. (1996). Grammatical bias for evolutionary learning. Ph.D. thesis, School of Computer Science, University College, University of New South Wales, Australian Defence Force Academy, Canberra, Australia. Yang, Y., & Padmanabhan, B. (2005). Evaluation of online personalization systems: A survey of evaluation schemes and a knowledge-based approach. Journal of Electronic Commerce Research, 6(2), 112–122. Zhang, L., Jack, L. B., & Nandi, A. K. (2005). Fault detection using genetic programming. Mechanical Systems and Signal Processing, 19(2), 271–289. Zhang, M.-L., & Zhou, Z.-H. (2004). Improve multi-instance neural networks through feature selection. Neural Processing Letters, 19(1), 1–10. Zhang, M.-L., & Zhou, Z.-H. (2006). Adapting rbf neural networks to multi-instance learning. Neural Processing Letters, 23(1), 1–26. Zhang, Q., & Goldman, S. (2001). EM-DD: An improved multiple-instance learning technique. In NIPS’01: Proceedings of neural information processing system (Vol. 14). Zhou, Z.-H., Jiang, K., & Li, M. (2005). Multi-instance learning based web mining. Applied Intelligence, 22(2), 135–147. Zhou, Z.-H., & Zhang, M.-L. (2007). Solving multi-instance problems with classifier ensemble based on constructive clustering. Knowledge and Information Systems, 11(2), 155–170. Zucker, J.-D., & Chevaleyre, Y. (2000). Solving multiple-instance and multiple-part learning problems with decision trees and decision rules. Application to the mutagenesis problem. In Proceedings of the 14th Canadian conference on artificial intelligence. Lecture notes in artificial intelligence, Ottawa, Canada (pp. 204–214). Multi-instance genetic programming for web index recommendation Introduction Multiple instance learning Grammar-guided genetic programming for multiple instance learning Individual representation Initialization Genetic operators Selective crossover Selective mutation Evolutionary algorithm Web index recommendation: a multiple instance problem Experimental setup Data sets Algorithm details G3P-MI variants Fitness function Other algorithm parameters Results and discussion Comparing different versions of G3P-MI Comparing G3P-MI with other proposals Knowledge acquired Conclusions and future work Acknowledgments References 
work_uudn7xmzevfhlkatd6wdjrnypm	----	The MICROSTROKE expert system for stroke type diagnosis. 1353 The MICROSTROKE Expert System for Stroke Type Diagnosis Klaus Spitzer, PhD, MD, Andreas Thie, MD, Louis R. Caplan, MD, and Klaus Kunze, MD MICROSTROKE is a prototype expert system designed to categorize and diagnose stroke types based on clinical information. The knowledge base of MICROSTROKE includes information from large stroke registries. The system first queries the physician-user for details of the patient's history, information about the onset of stroke, accompanying symptoms, and pertinent neurologic findings and then sums the individual data items, factors in the a priori odds, and arrives at the probabilities of different stroke types for a given patient. Specific diagnosis of stroke type includes thrombosis, embolus, lacune, intracerebral hemorrhage, and subarachnoid hemorrhage. Stroke type diagnoses by MICROSTROKE were correct in 72.8% of 250 cases in the Hamburg Stroke Data Bank, MICROSTROKE runs on any MS-DOS microcomputer and is intended as a practical aid for physicians not fully familiar with the diagnosis of stroke types. (Stroke 1989;20:1353-1356) For the bedside assessment of stroke type,knowledge of the frequency distributions ofsigns, symptoms, and ecological data asso- ciated with the different stroke types can be of prime importance. Neurologists collect historical data, neurologic signs, and symptoms to arrive at a "best guess" as to stroke type, which then forms the basis for performing further diagnostic proce- dures such as computed tomography (CT scan) or cardiologic or cerebrovascular tests.1 During the last decade, much research effort has been devoted to the development of expert systems to cope with complex medical decision-making. An expert system is "an embodiment within a com- puter of a knowledge-based component, from an expert skill, in such a form that the system can offer intelligent advice or make an intelligent decision about a processing function."2 Such a system uses expert knowledge to attain high levels of perfor- mance in a narrow problem area.3 We present MICROSTROKE, the prototype of an expert system for computer-supported stroke type diagnosis based only on clinical and historical patient From the Neurologische Universitatsklinik Hamburg- Eppendorf, Hamburg, Federal Republic of Germany (K.S., A.T., K.K.) and the Department of Neurology, Tufts-New England Medical Center, Boston, Massachusetts (L.R.C.). Presented in part at the 13th International Joint Conference on Stroke and Cerebral Circulation, February 18-20, 1988, San Diego, California. Address for correspondence: Dr. Spitzer, Neurologische Uni- versitatsklinik Hamburg-Eppendorf, MartinistraBe 52, D-2000 Hamburg, Federal Republic of Germany. Received August 4, 1988; accepted May 16, 1989. data available at the bedside, MICROSTROKE serves as an aid in the diagnostic work-up of stroke as both a stroke patient data bank and as an educational tool in clinical teaching. Another expert system, TOPOSCOUT, has been independently developed for stroke localization.4 Materials and Methods MICROSTROKE includes three knowledge data bases. The first contains frequency distributions of clinical and ecological parameters for the stroke types throm- bosis, embolus, lacune, intracerebral hemorrhage (ICH), and subarachnoid hemorrhage (SAH). Data are derived from the Stroke Data Bank,5 the Michael Reese Stroke Registry,6 and the Harvard Coopera- tive Stroke Registry7 (Figure 1). The frequency distribution tables contain relative frequencies of single items (symptoms or historical data) or of a set of alternative symptoms for all stroke patients and for each stroke type, respectively. A second knowledge data base is implemented using rule-based information coding. Rule-based systems depend on the hypotheses that expert knowl- edge consists of many independent, situation- specific rules and that computers can simulate expert reasoning by stringing these rules together in chains of deduction.8 The "if" part of a rule (the premise) contains the pattern or attributes that must be matched for the rule to be used. The "then" part (the conclusion) contains an assertion to be made when the premise is satisfied. A typical rule is "If there is hypertension at onset and early course of deficit is gradual smooth progression of symptoms, D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1354 Stroke Vol 20, No 10, October 1989 a l l thr l i e u b ICH SiH OnMt on i r i i i a g . 31 40 BO 1 7 1 3 I E Physical S t r t u at o n u t . 4 1 1 S 10 16 Ordinary daily a c t i v i t y a t o m t t . B9 64 47 68 64 64 Othar/oakson. 7 6 2 10 13 6 Sourer Harvard Straki R t g i i t r y . ( p a r c u t a g t ) FIGURE 1. Representation of MICROSTROKE'S knowledge base of frequency distributions, all, all stroke types; thr, thrombosis; lac, lacune; emb, embolus; ICH, intra- cerebral hemorrhage; SAH, subarachnoid hemorrhage. then display warning for ICH.'' Certain rules include combinations of symptoms that are associated with a high probability of intracranial hemorrhage. At present, MICROSTROKE matches the premises of 11 rules with data from a current patient to prompt a warning for ICH, SAH, or both if they apply. The third knowledge data base consists of exclu- sively text information, used by the tutorial module of MICROSTROKE and stored as an American Stan- dard Code for Information Interchange (ASCII) file. This knowledge data base serves only educational purposes and has no influence on calculations of stroke type diagnostic probabilities. MICROSTROKE acquires knowledge by interac- tively asking the physician, user for details of the patient's history, information about the onset of stroke, and accompanying symptoms in a question- naire comprising 38 items for which frequency distribution tables in different types of stroke are available. The answers accepted are yes, no, unknown, or an option if a multiple-choice question is presented. The inference engine of any expert system is the computer program that provides its general problem- solving capabilities. The inference engine is sepa- rated from the collection of domain knowledge, the knowledge data base, MICROSTROKE'S inference engine calculates probabilities of different stroke types using modified Bayesian inference tech- niques. Each stroke type is attributed an account. Starting from initial accounts representing the a priori odds for the five stroke types, accounts are recalculated after each question depending on the physician-user's answer. The order of the questions presented depends on the MICROSTROKE mode selected. If there is no laboratory data available, intracranial hemorrhage cannot be excluded, and MICROSTROKE'S first goal is to detect signs of ICH or SAH. If intracranial hemorrhage has already been excluded, for example, by CT scan and lumbar puncture, data are acquired to differentiate isch- emic stroke types. The accounts are displayed as probabilities of stroke types by multiplying each with a constant, yielding an account sum of 100. Figure 2 shows the calculation process of MICRO- STROKE'S inference engine in some detail. It must be emphasized that the weight attributed to an item depends on the physician-user's answer regarding this item. Thus, a positive answer may influence Let type be an element of {thrombus,embolus,lacune, ICH,SAH}, a(iype)[n] the account of type after the n-th question, and p(iype)[n] the correspondent prob- ability of type after the n-th question. The five accounts a(thrombus)[0]=34, a(embolus)[0]=31, a(lacune)[0]=19, a(ICB)[0]=W, a(SAH)[0j=B represent the a priori odds before any patient infor- mation is available. The formula "(tVP')[" + 1] = <"(<W«) W x pr""' with pr = 1, pr = ptsymptom\type), pr = p{notsymptom\type), vitx = 1, wei = 0.5, wti = 0.25, if answer unknown; if answer yes; if answer no; if weight(question, answer) = 10; if weightiquestion, answer) = 5; '•£ weighi\queetion,answer) = 2; is used for iterative recalculation after each question is answered. The correspondent probability is given by a(type)[n] x 100 p(type)[n] = The dependent probabilities p(symptom\type) and p{notsymplom\type) are ex- tracted from the frequency distribution tables of the knowledge base. The weight(question, answer) represents the weight of a question dependent from the answer given by the user. Weights are stored in the knowledge base. FIGURE 2. MICROSTROKE'S rules for calculating stroke type probabilities. ICH, intracerebral hemorrhage; SAH, subarachnoid hemorrhage. calculation more than a negative one; for example, "coma at onset" raises the account of ICH and SAH more than "normal level of consciousness" lowers it. This strategy and the mode-dependent modification of the questionnaire if intracranial hem- orrhage has been excluded distinguishes MICRO- STROKE from classical Bayesian-based systems. MICROSTROKE'S knowledge bases are strictly sep- arate from its inference engine. Thus, frequency tables, weights and selection of questions and their order, combinations of symptoms that invoke the ICH/SAH warning module, and the tutorial text information can be easily modified using any word processor. MICROSTROKE is written in TURBO-PASCAL. It runs on any IBM-compatible microcomputer. Results MICROSTROKE consists of five modules that can be switched by entering simple keystrokes in screen menus. In the first, the physician-user must select an item from the knowledge base menu that includes information on the history, neurologic signs, and accompanying findings, and MICROSTROKE then dis- plays relative frequencies of the item selected for all patients and each stroke type from one of the three stroke registries. Frequency distribution tables (as in Figure 1) are presented. The second, tutorial, module is used for computer- aided instruction. We have created text screen images that are linked by a logical relation, provid- ing a stepwise explanation of stroke problems (e.g., pathophysiology, diagnosis, and management). MICROSTROKE'S text data base has the ability to store complete publications on the subject of stroke (e.g., this paper) to be read on the computer screen during a session. Furthermore, the tutorial module has the capacity for explanation, displaying path- D ow nloaded from http://ahajournals.org by on A pril 5, 2021 Spitzer et al MICROSTROKE Expert System 1355 probability probability probability probability probability taning: of of of of of throiboiii lacuna: ubolis: ICH: Slfl: : 18 a 21 E 64 ICH tttt$tttt t ttitttttm m umttttmttttmtttttttt o r SAH EEC 19: Otnder ula. Toing*r thin 56. So atb.«roscltrosis. Ho iyp*rt*n- •ion at onut. Haadach* at onitt. Hoi about tht tarly count of dificit ? ->1 1 Haxiial dtfieit at oistt. 2 Stuttering or itapwis* coirs*. 3 Gradial nooth C O U T H . i Flictiant coirs*. 6 Early conn* unknown. (1-5) FIGURE 3. Excerpt from typical MICROSTROKE session. ICH, intracerebral hemorrhage; SAH, subarachnoid hemorrhage. ways of reasoning, physician-user's instructions, and help screens. In the third, expert system, module, the physician- user is first asked whether intracranial bleeding has already been excluded. If so, MICROSTROKE omits questions especially designed to detect ICH and SAH. MICROSTROKE starts out with the a priori odds of different stroke types and modifies them accord- ing to the physician-user's consecutive answers. Figure 3 shows a typical MICROSTROKE screen. The upper part displays the calculated odds of different stroke types for the patient whose entered charac- teristics are listed in the middle, MICROSTROKE'S questions and the answers that will be accepted by the inference engine appear in the lower part. In this particular case, a warning message appears on the screen, drawing attention to the possibility of intra- cranial bleeding in this young patient presenting with headache. The patient's data profile is reviewed continuously during the session for patterns of symptoms commensurate with ICH or SAH. At the end of the session, MICROSTROKE has summed up all entered information to present its final odds. The physician-user is asked for a final diagnosis based on laboratory studies, if available. When all data of authentic cases are stored, including the physician-user's final diagnosis con- firmed by laboratory studies, MICROSTROKE updates its stroke registry (fourth module). Data profiles can be replayed for later reviews of particular cases, and information can also be transferred to other programs for statistical analyses or scientific reports. MICROSTROKE'S registry has the ability to store large quantities of data, providing easy access for patient care and research. MICROSTROKE'S fifth, control unit, module has access to data files of large stroke registries that are not included in its knowledge base. Cases stored in these stroke registries may be presented to MICRO- STROKE in an automatic fashion so that the data of a particular stroke registry, rather than the physician- user, will answer the system's questions. The final odds calculated by MICROSTROKE may then be com- pared with the diagnosed stroke types in the regis- try, thereby achieving quality control of the system. Using the control unit, MICROSTROKE'S guesses have been prospectively tested for conformity with the final diagnoses of 250 cases in the Hamburg Stroke Data Bank.9 MICROSTROKE was correct in 72.8% of all, in 11 of 12 SAH (91.7%), in 17 of 27 ICH (63.0%), in 96 of 128 embolic (75.0%), and in 58 of 82 thrombotic strokes (70.7%). Nine ICH, one SAH, and 12 embolic strokes were incorrectly diagnosed as thrombosis, and one lacune was taken for an embolus. The control unit also checks the quality of MICROSTROKE'S diagnostic assessment of authentic cases from previous sessions using the built-in stroke registry. Discussion MICROSTROKE is designed to serve as a computer- based diagnostic tool and a knowledge base for stroke type diagnosis, MICROSTROKE is intended as a practical aid for practitioners and neurologists involved in stroke management on wards, in stroke units, and in emergency rooms. Our expert system is hardware-independent, allowing it to be run on a wide variety of small computers, including laptops. These computers are now available in almost all medical institutions. Application of computer- assisted methods for faster and more accurate diag- nostic routines could save time, permitting physician- users to give more time to patient care or teaching.10 MICROSTROKE'S ability to detect intracranial hemor- rhage with high sensitivity could prompt physician- users to make appropriate diagnostic tests very early during the clinical course. Simulating a physician's decision-making pro- cess, MICROSTROKE incorporates modified Bayesian statistics based on frequency distributions of symp- toms combined with information stored in rule- based knowledge data banks.11-13 In contrast to statistically oriented expert systems, a class of expert systems based on inexact reasoning has been developed by investigators in artificial intelligence, taking into account the judgmental, subjective, and nonprobabilistic characteristics of medical knowl- edge.14-16 TOPOSCOUT, an expert system for stroke localization, belongs to this class of expert systems.4 Though there are major objections to the use of Bayesian techniques in medical diagnosis.17'18 Bay- esian classification systems may be more valuable than is commonly believed for some medical problems.19-20 The major advantage of computer programs based on Bayesian statistics is their abil- ity to incorporate data from large patient registries. For example, MOS is a module that deduces stroke type by using clinical and historical data from the Michael Reese Stroke Registry.21 Like MOS, many expert systems developed in the medical field are connected to clinical data banks.22 Furthermore, MICROSTROKE exhibits self-learning characteristics, making use of its stroke registry module, so that D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1356 Stroke Vol 20, No 10, October 1989 with each new patient, the amount of stored infor- mation increases. If these data are made available to the inference engine, MICROSTROKE simulates the growing clinical experience of a human expert. Since expertise is the cornerstone upon which an entire expert system is built,23 the true power of MICROSTROKE lies in its knowledge data base. Yet, the process of extracting expertise from an expert physician and transforming it into a formal structure that can be understood and used by an expert system is a long, difficult, and error-prone task.16-24 For this reason, we decided to implement only published information from large stroke registries. We are aware that important clinical items are missing from MICROSTROKE'S questionnaire, but we preferred to start out with a limited but well- documented knowledge base. This restriction of the knowledge base was mainly responsible for the 27.2% incorrect stroke type diagnoses, especially in ICH and thrombotic strokes. For example, due to the lack of discriminating questions, MICROSTROKE revealed difficulties in distinguishing patients with cerebral infarcts who presented with impaired lev- els of consciousness from patients with ICH. As work progresses, MICROSTROKE'S knowledge base will be enlarged by published cases that have been studied in greater detail to improve its performance. Nevertheless, in converting this prototype expert system into a high-performance one we may meet with unexpected difficulties.25 MICROSTROKE suggests that computer-assisted stroke type diagnosis is feasible at the bedside.26 A prospective study currently in progress at Tufts- New England Medical Center and Hamburg Uni- versity Hospital must be concluded to appraise MICROSTROKE'S true value. In the future, efforts to improve MICROSTROKE'S performance will include expansion of its knowledge bases, incorporation of its own experience contained in its stroke registry, and further problem-specific modification of the underlying inference techniques. Finally, we want to emphasize that physicians disagree among themselves as much as computer programs disagree with doctors,17 especially in the diagnosis of stroke type.27 References 1. Caplan LR, Stein RW: Stroke. A Clinical Approach. Boston, Butterworths, 1986 2. Forsyth R (ed): Expert Systems. London, Chapman & Hall, 1984 3. Waterman DA: A Guide to Expert Systems. Reading, Mass, Addison-Wesley, 1986 4. Spitzer K, Thie A, Caplan LR, Kunze K: The TOPOSCOUT expert system for stroke localization. Stroke 1989; 20:1195-1201 5. Foulkes MA, Wolf PA, Price TR, Mohr JP, Hier DB: The Stroke Data Bank: Design, methods, and baseline charac- teristics. Stroke 1988;19:547-554 6. Caplan LR, Hier DB, D'Cruz I: Cerebral embolism in the Michael Reese Stroke Registry. Stroke 1983;14:530-536 7. Mohr JP, Caplan LR, Melski JW, Goldstein RJ, Duncan GW, Kistler JP, Pessin MS, Bleich HL: The Harvard Cooperative Stroke Registry: A prospective registry. Neu- rology 1978;28:754-762 8. Schwartz B, Patil RS, Szolovits P: Artificial intelligence in medicine. Where do we stand? N Engl J Med 1987; 316:685-688 9. Spitzer K, Becker V, Thie A, Kunze K: The Hamburg Stroke Data Bank: Goals, design and preliminary results. / Neurol 1989;236:139-144 10. Richard RH: Evaluation of a medical data system. Comput Biomed Res 1970;3:415-425 11. Davis R, Buchanan B, Shortliffe E: Production rules as a representation for a knowledge based consultation program. Artiflntell 1977;7:15-45 12. Hudson DL, Estrin T: Derivation of rule-based knowledge from established medical outlines. Comput Biol Med 1984; 14:3-13 13. Clancey WJ: The epistemology of a rule-based expert sys- tem—A framework for explanation. Artif Intell 1983; 20:215-251 14. Shortliffe EH: Computer-Based Medical Consultations: MYCIN. New York, Elesevier, 1976 15. Szolovits P: Artificial Intelligence in Medicine. Boulder, Colo, Westview Press, 1982 16. Torasso P: Knowledge based expert systems for medical diagnosis. Stat Med 1985;4:317-325 17. Bouckaert A: Medical diagnosis: Are expert systems needed? IntJ Biomed Comput 1987;20:123-133 18. Szolovits P, Pauker SG: Categorical and probabilistic rea- soning in medical diagnosis. Artiflntell 1978;11:115-144 19. Chard T: Human versus machine: A comparison of a com- puter 'expert system' with human experts in the diagnosis of vaginal discharge. Int J Biomed Comput 1987;20:71-78 20. Zagoria RJ, Reggia JA: Transferability of medical decision support systems based on Bayesian classification. Med Decis Making 1983;3:503-509 21. Hier DB, Atkinson GD, Perline R, Hill H, Evans M, Desai B, McCormick WC, Caplan LR: Can a patient data base help build a stroke diagnostic expert system? Med Inf (Lond) 1986;11:75-81 22. Hier DB, Caplan LR, Hill H, Evens M, Sinha A: A microcomputer-based expert system to assist in localization of anatomic damage after stroke (abstract). Neurology 1984; 34(Suppl 1):83 23. Hayes-Roth F, Waterman DA, Lenat DB (eds): Building Expert Systems. Reading, Mass, Addison-Wesley, 1983 24. Mars NJ, Miller PL: Knowledge acquisition and verification tools for medical expert systems. Med Decis Making 1987; 7:6-11 25. Alvey PL, Myers CD, Greaves MF: High performance for expert systems: I. Escaping from the demonstrator class. Medrnf(Umd) 1987;12:85-95 26. Tuhrim S, Reggia JA: Feasibility of physician-developed expert systems. Med Decis Making 1986;6:23-26 27. Gross CR, Shinar D, Mohr JP, Hier DB, Caplan LR, Price TR, Wolf PA, Kase CS, Fishman IG, Calingo S: Interob- server agreement in the diagnosis of stroke type. Arch Neurol 1986;43:893-898 KEY WORDS • expert systems cerebrovascular disorders diagnosis D ow nloaded from http://ahajournals.org by on A pril 5, 2021 
work_uus5mifswfhjrdx4l3rjafh7lm	----	SOFTWARE—PRACTICE AND EXPERIENCE, VOL. 1(1), 1 (DECEMBER 1997) Object-Oriented Patterns: Lessons from Expert Systems TIM MENZIES Department of Artificial Intelligence, School of Computer Science and Engineering, The University of New South Wales, Sydney, Australia, 2052 timm@cse.unsw.edu.au SUMMARY Three benefits are typically claimed for object-oriented (OO) patterns: (i) reusing parts of the conceptual models of old implementations; (ii) guiding the current development based using successful previous devel- opments; and (iii) communicating existing systems to newcomers. We will argue that a similar idea can be found in the expert systems literature dating from the early-1980s. The goal of ��� or knowledge-level mod- elling (e.g. KADS) is to identify abstract patterns of inference that appear in many expert systems. Such abstract patterns of inference and program structure, it is argued, are productivity tools for the creation of software applications; i.e. ��� argues for a similar reuse benefit as OO patterns. Recently, however, an alternative view has emerged. While such abstract patterns are good for communications and guidance, the reuse benefits may never be realised. Patterns may be best viewed as tools for structuring an argument, rather than recording a conclusion. KEY WORDS OO patterns expert systems knowledge-level modelling INTRODUCTION An exciting idea in current OO thinking is the pattern, i.e. a fragment of a high-level conceptual model which may be useful in many applications. Using patterns may have three benefits: � The re-use benefit: A designer can bootstrap themselves into better systems using proven old systems. Example reuse patterns can be found in 1-4. Note that analysts may not use the patterns verbatim. Reuse patterns are like the logical design which may require some configuration/ al- teration for the physical implementation of any particular system. Nevertheless, the essence of the physical implementation will be the reuse pattern.� The guidance benefit: Not all patterns are reusable libraries of OO classes. Guidance patterns serve to direct the analyst’s focus onto a set of issues that previous analysts have found insightful. Example guidance patterns are CHECKS 5 and Caterpillar’s Fate 6.� The communication benefit: Patterns are a succinct tool for explaining existing systems. When we tutor OO, we find patterns to be a useful final initiation ritual for a novice OO developer. When they “get” patterns, we know that they are capable of comparing and contrasting a wide range of OO systems. Despite the current level of enthusiasm for OO reuse patterns, we find it necessary to sound a note of caution. A similar idea, called ��� or knowledge-level modelling can be found in the expert systems literature dating from the early-1980s 7, 8. This paper tries to bridge the gap between abstract conceptual models proposed for OO and abstract conceptual models proposed for expert systems. In all, we will say more about expert systems than OO. In particular, after over a decade of experience with ��� , we CCC 0038–0644/97/010001–01 Received 1 October 1996 c � 1997 by John Wiley & Sons, Ltd. Revised 10 May 1996 2 T.J. MENZIES Class Hierarchy Browser Number Fraction Complex Integer - * + - aNumber denominator) - ^((numerator * aNumber (denominator * aNumber (denominator * aNumber numerator))/ denominator) Figure 1. A class hierarchy browser can make some clear statements about its pitfalls. We will argue that we can extrapolate the lessons of ��� modelling to OO reuse patterns. That is, the problems already seen by ��� will have to be faced in the future by OO pattern researchers. In particular: we doubt the reuse benefit but not the communication benefit or the guidance benefit. This paper is structured as follows. Due to the hybrid nature of the content of this paper (OO and expert systems), we will first introduce OO design patterns to our expert systems audience. Since our concern is primarily with the reuse benefit and not the guidance benefit, this section will focus on the GOF/GOV/Fowler/Coad-style reuse design patterns . Next, we introduce ��� to our software engineering audience and then offer a mapping from OO patterns to ��� . Known pitfalls of ��� will be reviewed. Our discussion section offers two action plans based on this revised view of patterns. Roughly speaking, we say that patterns may be best viewed as tools for structuring an argument, rather than recording a conclusion. REUSE OO PATTERNS This section is a short introductory tutorial on OO patterns. Consider the class hierarchy browser of Figure 1. When a class name is selected in the upper-left list box, the methods of that class are displayed in the upper-right list box. If one of these methods is selected, then the source code for that method is displayed in the bottom text pane. Now compare this class hierarchy browser with the disk browser shown in Figure 2. When a directory name is selected in the upper-left list box, the files in that directory are displayed in the upper-right list box. If one of these files is selected, then the contents of that file are displayed in the bottom text pane. Clearly, there is some similarity in the two browsers. Containers (classes or directories) are shown top-left. The things in the containers that are not themselves containers (methods and files) are shown top-right. The contents of these non-container things are shown in the bottom pane. If we rename containers composites and the non-containers leaves then we can design one composite browser class that handles both class hierarchies and directory trees (see Figure 3). That is, our disk browser and class hierarchy browser are both presentations of nested composites. Figure 4 shows the inner structure of the composite browser. Composites contain either other com- posites or leaves. Leaves compile the contents of lower text-pane. Once this structure is in place, all that is required to convert a disk browser to a class hierarchy browser is to: � Change the title of the window for “Disk Browser” to “Class Hierarchy Browser”. GOF= the “gang of four” 1 ; GOV= the “gang of five” 2; Fowler 3 ; Coad 4 OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 3 c:\ bin docs wp docs words.dvi words.tex words.ps \begin{document} \author{Tim Menzies, Department \title{OO Patterns: Lessons} of Artificial Intelligence Figure 2. A disk browser � Define Class beneath Composite and Method beneath Leaf.� Implement the different compiles methods in Method. Compiling file contents implies transferring text to primary storage. Compiling method contents implies parsing the source code, etc. We have just isolated a “pattern”: a fragment of a high-level conceptual model which may be useful in many applications. The above design can be used to (i) browsing a disk; (ii) browse a class hierarchy; or, more generally, browse any 1-to-many nested aggregation (e.g. players in teams, persons in com- panies, stock on shelves). This composite pattern is one of the 23 OO reuse design patterns listed by GOF. The above example suggests the power of such OO reuse patterns. Seemingly different problems can be resolved to a single design. OO reuse patterns could become a repository for experience which can benefit new designers. OO reuse patterns could also serve to unify the terminology of OO design, allowing experience from one application to migrate into another area. Patterns have been documented in many formats. The GOV prefer the format: context, problem, solution 2. The context describes a design situation. The problem describes the set of forces that re- peatedly occur in that situation while the solution describes a configuration to balance those forces. This solution contains a description of the static components and the runtime behaviour. In OO pat- terns: (i) the static components are described using class hierarchies and their relationships; and (ii) the runtime behaviours are described using some variant on collaboration diagrams 9. The GOV argue that this format of a pattern is compatible with numerous other patterns researchers (page 11 of 2). Patterns can be at different layers of abstraction. The GOV describe three layers of pattern ab- Composite Broswer Composites hierarchy Leafs at selected composite Fraction Editor for the selected leaf. Figure 3. A composite browser 4 T.J. MENZIES Method X Y isa XY YX Association Aggregation YX (1 to 1) Class name methods attributes File Component Leaf Class Directory edits Composite children title CompositeBrowser browse compiles (many X to 1 Y) � ����� Figure 4. Object model for the composite browser straction: (i) low-level language-dependent idiom patterns; middle-layer language independent design patterns describing a programmer’s key mechanisms (e.g. the GOF patterns); (iii) and top-level ar- chitectural patterns that spread across the entire application. Examples of architectural patterns are layered architectures (e.g. the three tiered database-model-dialogue systems found commonly in stan- dard management information system-style applications); pipe-and-filter (e.g. the dominant paradigm in UNIX shell scripts); or blackboards (an expert systems technique). Patterns can be pitched at different audiences. For example, the GOV and GOF patterns are intended for programmers or implementation-aware analysts. Fowler describes analysis patterns’; i.e. high-level conceptual patterns which are used to communicate a design to the user community. Fowler was in- volved in the development of a large medical system. Analysis patterns were used to discuss the design of the system with doctors and nurses. Some patterns found in that medical system (chapter 3 of 3) were also useful in a corporate finance applications (chapter 4 of 3). ��� INFERENCE SKELETONS Our general claim will be that OO reuse patterns and ��� (e.g. KADS) are similar enough for lessons from ��� to be relevant to OO reuse patterns. This section describes ��� . For the moment, we will use the term “inference skeletons” to describe the conceptual models in ��� since we have yet to prove that ��� equals OO patterns (this will be done below). The goal of ��� modelling is to identify abstract reusable inference skeletons that appear in many ex- pert systems; e.g. diagnosis, classification, monitoring, etc. Such abstract reusable inference skeletons, it is argued, are productivity tools for the creation of expert systems. Examples of ��� are Problem Solving Types 7, Clancey’s model construction operators 10, the PROTEGE-II project 11, TINA 12, SPARK/ BURN/ FIREFIGHTER (SBF) 13 and KADS 14. In terms of mature ��� methodologies, OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 5 Data abstractions Solution abstractions HEURISTIC MATCH ABSTRACTION DATA Data Solutions REFINEMENT Figure 5. Heuristic classification KADS is the dominant ��� approach. Wielinga et. al. note that, as of 1992, KADS has been used in some 40-to 50 knowledge-based systems (KBS) projects, 17 of which are described in published papers 14. In his classic Heuristic Classification paper, Clancey reverse-engineered 10 expert systems written in a variety of tools and languages. He found that all these systems shared the same abstract inference skeletons, which he called heuristic classification (shown in Figure 5). For example, Figure 6 shows Clancey’s analysis of the MYCIN 15, 16 inference structure (abstract schema at top, followed by an example). MYCIN was a backward-chaining rule-based system that prescribed antibiotics. MYCIN worked by building an abstract model of the patient from the patient’s data. This is then matched across to a hierarchy of disease classes. The final diagnosis is produced by following the class diseases hierarchy downwards, looking for the most specific disease that is relevant to this patient. Note the similarity with Figure 5. Figure 7 shows Clancey’s analysis of another system, written in LISP which diagnoses an electronic circuit in terms of the component that is causing faulty behaviour. The abstract schema is shown on top, and an example is shown underneath. Note the similarity with Figure 6 and 5. After the Heuristic Classification paper, Clancey refined his inference skeletons. In Model Con- struction Operators 10, Clancey argued that rules like Figure 8 contain domain-specific terminology (see Figure 9) as well as reusable inference strategies (see Figure 10). If these are removed from the rule, then not only have we isolated the true business knowledge in the rule (see Figure 11), but we also have found inference knowledge we can reuse elsewhere. Clancey’s preferred architecture for expert systems is (i) a library of pre-defined problems solving strategies such as Figure 10; and (ii) a separate knowledge base containing the special domain heuristics like Figure 11. Inspired by Clancey’s work, subsequent researchers sought other abstract inference skeletons. Tans- ley & Hayball 17 list over two dozen reusable inference skeletons including systematic diagnosis (lo- calisation and causal tracing), mixed mode diagnosis, verification, correlation, assessment, monitoring, simple classification, heuristic classification, systematic refinement, prediction, prediction of behaviour and values, design (hierarchical and incremental), configuration(simple and incremental) planning, and scheduling. These skeletons are recorded using the KADS notation of Figure 12 in which rectangles are data structures and ovals are functions. Given a complaint, the KADS abstract pattern for diagnosis is that a system model is decomposed into hypothetical candidate faulty components. A norm value is collected from the system model. An observation for that candidate is requested from the observables (stored internally as a finding). The candidate hypothesis is declared to be the diagnosis based on the difference between the norm value and the finding. Note the shaded portions of Figures 12 & 13. We will return to these shaded portions below (see Figure 18). As another example, Figure 13 shows the KADS abstract inference skeleton for monitoring. A parameter is selected from a system model. The model’s expected normal value is generated 6 T.J. MENZIES HEURISTIC Compromised Host infection Gram-negative Immunnosuppressed GENERALISATION GENERALISATION Leukopenia DEFINITIONAL Low WBC QUALITATIVE WBC < 25 E. coli Infections SUBTYPE ������� ���� !�"�� $#&%$'�� Patient Data Disease REFINEMENT ABSTRACTION DATA Patient Abstractions Disease Classes HEURISTIC MATCH �)( � '�*�+,% Figure 6. MYCIN from the model and collected from the observable-s (stored as a finding). The current state of the monitoring system us reported as a discrepancy class after comparing the finding with the expected normal value. Note that Figures 5,12, & 13 do not imply a particular execution order of their functions. Concep- tually each function can be driven forwards or backwards to connect inputs to outputs or visa versa. The heuristic classification pattern of Figure 5 could be driven from data to solutions to perform diag- nosis; i.e. given the data, execute forwards data-abstraction then heuristic match, then refinement. Alternatively, it could be driven from solutions to data to perform intelligent data col- lection; i.e. given solutions, execute backwards refinement, then heuristic match, then data abstraction. In this backwards reasoning, the generated data items become requests back to the environment in order to rule out certain possibilities. KADS explicitly models this procedural ordering of the function calls in a separate task layer diagram. All known abstract inference patterns skeletons are really combinations of a small number of reusable inference subroutines. Some of the reusable inference subroutines are shown in Figures 12 & 13 (e.g. select, specify, compare). We will see more inference subroutines in the TINA system, below. OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 7 HEURISTIC DATA HEURISTIC MATCH of Ports Qualitative value ABSTRACTION Qualtitative Circuit Behaviour Circuit Behaviour Measurements Behaviour at Some Port of Some Module in Behavior Lattice REFINEMENT Component Fault ������� ���- .�"�� $#&%$'�� QUALITATIVE Variable Voltage(Voltage N11 N14) Reference is High or OK is High Q5 Collector Open CAUSE (Voltage N11 N14) > 31 V �)( � '�*-+/% Figure 7. SOPHIE III New abstract reusable inference skeletons can be quickly built out of these lower-level inference primi- tives. Libraries of abstract reusable inference skeletons become a productivity tool for building a wide- variety of expert systems. Marques et. al. report significantly reduced development times for expert systems using a library of 13 reusable inference subroutines (including eliminate, schedule, present, monitor, classify, select and dialog-mgr) in their SPARK/ BURN/ FIRE- FIGHTER (SBF) environment. In the nine studied applications, development times changed from one to 17 days (using SBF) compared to a range of 63 to 250 days (without using SBF) 13. To our knowl- edge, this is the largest documented evidence of productivity gains in any software approach (be it knowledge based, object-oriented, or otherwise). if the infection in meningitis and the type of infection in bacterial and the patient has undergone surgery and the patient has undergone neurosurgery and the neurosurgery-time was less than 2 months ago and the patient received a ventricular-urethral-shunt then infection = e.coli (.8) or klebsiella (.75) Figure 8. A domain rule with hidden reusable inference fragments. From 10. 8 T.J. MENZIES subtype( meningitis, bacteriaMenigitis ). subtype( bacteriaMenigitis, eColi ). subtype( bacteriaMenigitis, klebsiella ). subsumes( surgery, neurosurgery ). subsumes( neurosurgery, recentNeurosurgery ). subsumes( recentNeurosurgery, ventricularUrethralShunt). causalEvidence( bacteriaMenigitis, exposure ). circumstantialEvidence( bacteriaMenigitis, neurosurgery ). Figure 9. Domain-specific terms from Figure 8. The above description is only a partial description of ��� in general and KADS in particular. For an on-line versions of the KADS documentation, see 021!143�5�6!687.7!7�9;: 7=<>9)3=:@?�9)A!B2CD9)E�F!683!G2H.I!J-K@1L:M6N H@O!O=HME2P.Q!R�S26MT2JM3-HMG.1L:>9)0.18O=F and 0!1!1.3�5�6.6�: 7�<>9U3=:�?�9VA2B2CW9UE-F!6 32G�H.I.J�K�1L:86 N H@O.O=HME.P!Q.R2S268X�C�3-JMG&:Y9U021 O�F . See 17 for a detailed text on KADS. See the Related Work section of 14 for a discussion of the differ- ences in the various ��� approaches. For a tutorial introduction to KADS, see 18, 8. For an advanced use of KADS-type models, see the TINA system (below). ��� = OO PATTERNS This section argues that it is inappropriate to declare ��� inference skeletons to be different from OO patterns based on their observed notational differences. The intention of the ��� researchers is the same as the OO patterns. Inference patterns satisfy the GOV definition of a pattern; i.e. they have context, problem, and solution. Hence, we argue that ��� inference skeletons are essentially the same as OO patterns. Further, we will demonstrate that the ��� patters are, in some respects, better patterns than the OO patterns. Are the Notational Differences Significant? OO patterns are usually expressed in a different notation to ��� inference skeletons (compare Fig- ure 4 with Figure 12). However, just because ��� inference skeletons are not expressed in an OO format, that does not mean they aren’t patterns: � OO patterns researchers agree that patterns need not be expressed only as networks of classes (pages 23-24 of 2). For example, Coplien refers to the 150 patterns gathered at Bell Labs for telecommunication systems. He remarks that “none are really object-oriented” 19. Strategy Description exploreAndRefine Explore super-types before sub-types. findOut If an hypothesis is subsumed by other findings which are not present in this case then that hypothesis is wrong. testHypothesis Test causal connections before mere circumstantial evidence. Figure 10. Problem solving strategies from Figure 8. OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 9 if the patient received a ventricular-urethral-shunt then infection = e.coli (.8) or klebsiella (.75) Figure 11. The business knowledge of Figure 8. � Patterns can be found in the data-modelling world (chapter 4 of 3) where they are expressed in an entity-relationship format.� Patterns were first described by Alexander as a tool for architectural design 20; i.e. originally, patterns were expressed in a non-programming format.� Shaw & Garlan’s text discusses common software architectures 21 using a free text format. The GOV’s section on architectural patterns translates the Shaw & Garlan, into an OO notation under the headings “context-problem-solution”. We would argue that the GOV translation did not add significantly to the Shaw & Garlan material. That is, Shaw & Garlan were discussing “patterns”; they just structured their material in a different manner. Another reason to reject ��� inference skeletons as being true patterns is that they are not described in the way proposed by the GOV; i.e. context, problem, and solution. Such a rejection may not be justified: � We will argue below that tacit in the ��� inference skeletons is a very strong notion of context, complaint select decompose observable hypothesis specify select finding compare norm system model difference Figure 12. KADS: diagnosis 10 T.J. MENZIES difference discrepancy class classify historical data select observable finding system model select parameter specify normcompare Figure 13. KADS: monitoring problem, and solution.� Several prominent patterns texts do not use the context/ problem/ solution format explicitly (i.e. Fowler 3 and Coad 4), yet clearly represent patterns research. For example, several of the patterns found in Coad can also be found in the GOF text, with small changes. Also, the foreword of the Fowler book is an enthusiastic patterns-based endorsement by one of the GOF authors. EveningTime DayTime Table DayRate Table EveningRate Activity State Taxes Compute Taxes State Allocation Time BasicTime Figure 14. Fowler’s process modelling notation; from page 153 of 3. OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 11 Interestingly, sometimes an “OO pattern” can look a lot like a “ ��� inference skeleton”. For exam- ple, consider Fowler’s proposed high-level notation for dynamic events Figure 14. Fowler notes that this notation is experimental, but stresses the importance of a high-level functional view like Figure 14. In Fowler’s proposed notation, ovals are event-driven controllers or triggers on tables which, when required, will call methods inside the classes (polygons). For example, when day time writes to day rate table, the triggers in that table will call methods within the activity class. If we (i) divide Fowler’s class polygons into one rectangle for each public method of each class, and (ii) replace Fowler’s ovals with ellipses, then the Fowler diagram become very similar to the KADS inference skeletons shown in Figures 12 & 13. Goals & Uses Notational comparisons aside, the intent of the ��� modelers and the OO patterns researchers is clearly the same. Consider the quote Clancey used to open the classic ��� paper Heuristic Classifica- tion 22. We believe that this quote reflects something of the same intent as the OO patterns community: To understand something as a specific instance of a more general case- which is what understanding a more fundamental principle of structure means- is to have learned not only a specific thing but also a model for understanding other things like it that one may encounter. J.S. Bruner Both ��� and OO patterns are examples of the same software abstraction process. Experienced software engineers can describe different applications using a common, abstract, language. Both ��� and OO design languages seek some characterisation of a design that is implementation independent and re-usable. The ��� literature shows that they seek to use ��� inference skeletons in a similar manner to OO patterns: � The SHELLEY 14 workbench project seeks to gain from the reuse benefit. Using SHELLEY, knowledge engineering becomes a structured search for an appropriate inference pattern. Knowl- edge engineers search requirement documents for a match between the stated requirements and the library of known abstract reusable inference patterns. Once such a pattern is found or devel- oped, then systems development becomes a process of filling in the details required to implement that abstract inference pattern. productivity tools for new applications.� Numerous ��� researchers argue for the communication benefit. For example, practioners find this retrospective second-glance at their systems useful for developing more generalised archi- tectures for future work 18. Expert systems theoreticians have used ��� to assess and clarify the essential features and differences of applications 23. Lastly, knowledge engineering novices can use a ��� analysis of classic expert systems to quickly review successful techniques.� However, to our knowledge, ��� researchers do not argue for the guidance benefit. Structure To demonstrate conclusively that our ��� inference skeletons are the same as the patterns seen in the OO world, we must show that ��� inference skeletons match the GOV definition; i.e. context, problem/solution, and descriptions of static components with their runtime behaviour. This section will argue that not only do ��� inference skeletons contain “context-problem-solution”, but the under- standing of the mapping between problems and solutions is more advanced in ��� than in OO patterns. 12 T.J. MENZIES Context Clearly, all the ��� inference skeletons have a context. The contexts of the Tansley & Hayball work were listed above and included systematic diagnosis (localisation and causal tracing), mixed mode diagnosis, verification, correlation, assessment, monitoring, simple classification, heuristic classifica- tion, systematic refinement, prediction, prediction of behaviour and values, design (hierarchical and incremental), configuration(simple and incremental) planning, and scheduling. Problems & Solutions ��� can offer detailed descriptions of problems and their mapping to solutions. For example: � The TINA 12, 24 system offers a 45 word language for describing problem constraints divided up into “epistemological criteria” (e.g. fault behaviour not constrained), “environ- mental criteria” (e.g. imprecise values). and “assumption criteria” (e.g. single fault assumption).� SBF 13 allows users to graphically draw a network reflecting their business process. The problems/solutions component of the GOV definition is often expressed as an IF-THEN rule. In some cases, the rules describing the mapping between problems and solutions in ��� libraries is so well understood that they can be directly executed. The result can be an automatically configured system. For example: � SBF can automatically map its business function graphs into its library of inference sub-routines. Once this mapping has been made, a rule-base can be generate which solves the business prob- lem. SBF worked in the context of configuration.� A similar tool for automatically generating a runtime solution from a high-level problem descrip- tion can be found in Protege-II 11. Protege-II is intended to be a multi-context tool but most of its published applications have been in the area of skeletal plan refinement (instantiating a gen- eral plan to a particular circumstance) or propose-and-revise (proposing an initial design, then modifying inappropriate portions of that design).� Benjamin describes TINA, an automatic configuration device for configuring solutions to differ- ent problems in the context of diagnosis 12, 24. TINA was a “proof-of-concept” prototype only and is not as sophisticated as SBF or Protege-II. However, the TINA technique is quite succinct. We will use this system to demonstrate how ��� researchers automatically map problem statements into solutions. In TINA, a solution is a problem solving method (PSM) which must be configured for a particular sub-context using a set of primitive inference techniques. For example: � Sub-contexts of diagnosis are defined by constraints within the domain such as the availability or absence of simulation rules.� A primitive inference within prediction based filtering could be a set intersection sub-routine. A TINA problem is described via suitability criteria categorised into a small number of types. For example inference rules and simulation rules are suitability criteria with the same type of constraint suspension method. The TINA system can automatically reflect over a set of rules describing the transformation process from problems to solutions (or, in the language of TINA, types of suitability criteria into PSMs). A simplified version of some of TINA’s rules is given in Fig- ure 15. Types of suitability criteria are shown in the when sections. When executing a then section, OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 13 1 diagnosis 2 if prime_diagnostic_method 3 then symptom_detection and 4 hypothesis_generation and 5 hypothesis_discrimination. 6 7 symptom_detection 8 when ask_user_method 9 then apply_user_judgment. 10 11 symptom_detection 12 when compare_symptom or 13 detection_method 14 then generate_expectation and 15 compare. 16 17 hypothesis_generation 18 when empirical_hypothesis_ generation_method 19 then associate and 20 prediction_filter. 21 22 hypothesis_generation 23 when model_based_hypothesis_method 24 then find_contributors and 25 transform_to_hypothesis_set and 26 prediction_based_filtering. 27 28 hypothesis_generation 29 when hypothesis_generation_method 30 then select_hypothesis and 31 collect_data and 32 interpret_data. Figure 15. Portions of the TINA rules used for converting problems descriptions into solutions. Adapted from 12. . if the PSM component is named in another rule, the reflection can recurse. For example, executing line 3 symptom detection makes TINA test the suitability of the rules at lines 7 and 11. A sample fragment of TINA output is shown in Figure 16. The trace back method traces back the dependents of the broken component to find potential contributors to the fault. In the case of mul- tiple contributors, TINA is saying that in this sub-context, they can be simply intersected. The resulting contributors set is assessed using the corroboration method. Innocent contributors are deleted (innocence is computed via running a high-level simulation). The remaining contributors are potentially guilty of the faults and another sub-routine is called to discriminate between them. Note that this corroboration method was generated when TINA explored prediction based filtering on line 26 of Figure 15 (using rules not shown in this article). For full details of this example, see 12. Static & Runtime Behaviour One difference between ��� patterns and OO patterns is their different emphasis of descriptions of static components vs runtime behaviours. As we move from the GOF to the GOV to Fowler, we see a common approach to static structure descriptions (OO notations) and an increasing focus on describing 14 T.J. MENZIES model_based_hypothesis_generation_method { trace_back_method; intersection_method; corroboration } trace_back_method { find_upstream } intersection_method { intersection } corroboration_method { select_random; simulate_hypothesis; compare; delete } Figure 16. After exploring its problems/solution mappings, TINA can automatically generate a PSM for diagnosis. Adapted from 12 and converted into a procedural formalism. the runtime behaviour. ��� patterns offer a very rich description of the runtime behaviour but are less focused on the static descriptions. The ��� patterns we have seen to date are modelled in a functional decomposition style. Hence: � Unlike OO, ��� operations are not modelled with data structures. Rather, they are modelled separately as inference sub-routines like classify (Figure 12) and select random (Fig- ure 16).� KADS does not show the relationships between their entities. The only “relationships” modelled are those representing data flows between functions. If “patterns” was purely an OO concept, then this functional decomposition approach to the speci- fication of ��� inference skeletons would disqualify them as patterns. However, we argued above that the concept of a pattern transcends the OO paradigm. ��� Patterns Z OO Patterns? OO patterns dates back to the early 1990s 25 while ��� dates back to nearly a decade earlier 7. ��� patterns research is more advanced than OO patterns research: � ��� patterns have been refined to the point where libraries of pattern solution components can be automatically configured into executable solutions from a problem description.� Librarian software has been built to support the intelligent indexing of ��� patterns. For example, if TINA finds that its problem cannot be solved using its known solutions, it conducts a dialogue with the user about which modelling assumptions can be relaxed. Further, a programmer can go to TINA with only a partial description of their proposed solution, and TINA will fill in the rest after a search for patterns related to that description. OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 15 Model % disorders identified % knowledge fragments identified 1: epistemological 50 28 2: KADS 55 34 3: no model 75 41 Figure 17. Analysis via different models LIMITS TO ��� We have made our case that ��� research is essentially the same as OO patterns excepting that the former uses a functional decomposition notation while that latter uses an OO notation. This section discusses the significance of that equivalence. What can OO patterns research learn from the ��� work? On the positive side, OO patterns researchers can import a large, well-defined body of knowledge about patterns in inference. OO patterns are mostly specified via a description of their static compo- nents. ��� patterns are specified via a description of their runtime components. OO patterns researchers could learn some useful tricks about runtime behaviour from their ��� counterparts. On the negative side, even after a decade of research, the productivity benefits of patterns has yet to be conclusively demonstrated in the ��� field. The rest of this section discusses these productivity issues. Clancey’s Heuristic Classification paper 22 offered a unified retrospective view on numerous, seemingly different, expert systems. Similar (but smaller) studies (e.g. 26, 18) suggest that ��� can retro- spectively clarify historical expert systems design issue. However several important productivity issues remain outstanding: do ��� abstraction assist in the design of new systems?; are the ��� abstractions re-usable?; and does the extra level of abstractions used in ��� overly-complicate the design process?. This questions are discussed below. Do ��� Abstractions Assist in the Design of New Systems? Corbridge et. al. reports a study in which subjects had to extract knowledge from an expert dialogue using a variety of abstract pattern tools 27. In that study, subjects were supplied with transcripts of a doctor interviewing a patient. From the transcripts, it was possible to extract 20 respiratory disorders and a total of 304 “knowledge fragments” (e.g. identification of routine tests, non-routine tests, relevant parameters, or complaints). Subjects were also supplied with one of three abstract reusable patterns representing models of the diagnostic domain. Each model began with the line “To help you with the task of editing the tran- script, here is a model describing a way of classifying knowledge”. Model one was an “epistemological model” that divided knowledge into various control levels of the diagnosis process. Model one was the “straw man”; it was such a vague description of how to do analysis that it should have proved use- less. Model two was a more sophisticated version of Figure 12. Model three was “no model”; i.e. no guidance was given to subjects as to how to structure their model. The results are shown in Figure 17. The statistical analysis performed by Corbridge et. al. found a significant difference between the performance of groups 3 compared to groups 1 and 2. No significant difference could be found between the poor-abstract-model group (model 1) and the group that was using a very mature abstract model (model 2). These are very counter-intuitive results. Using a hastily-built abstraction was just as useful as using a mature abstraction. And using no abstractions worked best of all! Far from challenging this result, the ��� community is now exploring empirical methods for exploring its approach in the 16 T.J. MENZIES Sisyphus-3 project . Are ��� Abstractions Re-usable? It is not clear that the problem solving methods found by ��� are truly reusable. Often, researchers build their own preferred patterns rather than use existing ones. The Tansley & Hayball patterns 17 are very different to the patterns offered by other researchers; e.g. Bredeweg’s qualitative prediction method 29. In the literature already reviewed by this paper, we can find four different versions of diagnosis: � Benjamin’s TINA’s view,� KADS (Figure 12),� Heuristic classification (Figure 6, & 7),� Tansley & Hayball’s approach 17. To this list, we can add numerous other approaches to diagnosis: � An OO-based approach described in chapter 3 of Fowler 3.� Our own graph-theoretic approach 30.� DeKleer & William’s approach based around a distinction between a problem solver and a assumption-based truth maintenance system 31.� Poole’s approach based on Bayesian reasoning 32.� To name but a few (for more examples, see 33). While some of the these patterns share some common features, they reflect fundamentally divergent different views on how to perform diagnosis. For example, like DeKleer, we view assumption space management as the key inference strategy within diagnosis 30. In assumption space management, mu- tually exclusive assumptions are managed in separate logical worlds. An expert system can reflect over that assumption space to intelligently select the next test to perform (a “good” test costs very little and culls much of the assumptions space). The implementation of such a multiple-worlds approach is a non-trivial task. It adds significant amounts of extra architecture to the implementation. This approach is not explored by Fowler, KADS, or Tansley & Hayball. We therefore note that, at least in the case of diagnosis: � The pattern has not stabilised with time;� The pattern may not do so in the foreseeable future. More generally, between the various camps of ��� researchers, there is little agreement on the details of the inference patterns. The list of inference sub-routines from KADS 14 and SBF 13 are significantly different. Also, the number and nature of the problem solving methods is not fixed. Often when a domain is analysed using ��� a new method is induced 18. In summary, since inference patterns have not stabilised over time, then extensive reuse is unlikely. Does ��� Modelling Overly-Complicated Modelling? Certain ��� authors note certain similarities between the different abstract reusable inference pat- terns proposed by KADS: Sisyphus is an attempt by the international KA community to define reproducible KA experiments 28. OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 17 finding compare difference norm specifyobservable select parameter/ hypothesis Figure 18. Overlap between KADS diagnosis and monitoring � Scheduling, planning and configuration are actually the same problem, divided on two dimen- sions (“goal states known or not” and “temporal factors considered or not” (Figure 12.3 of 17).� There exists a common sub-graph between Figures 12 & 13 (see Figure 18). Having noted that such similarities exist, the ��� researchers do not take the next step and sim- plify their distinctions (e.g. by combining diagnosis and monitoring). In other work 30, 34, we have explored unifying knowledge-based processing by inferencing over and-or graphs. Such graphs could be computed from an OO design if we ignore all encapsulation boundaries and just map the depen- dencies between instance variables. In terms of abstraction, we would characterise such a modelling technique as a very-low level tool. Nevertheless, we have found that a single inference procedure (ab- duction) can be applied over such a representation to implement many of the ��� abstract reusable inference patterns (e.g. diagnosis, case-based reasoning, explanation, prediction, prediction, classifica- tion, planning, monitoring, qualitative reasoning, verification, multiple-expert knowledge acquisition, explanation, single-user decision support systems, multiple-user decision support systems, natural- language processing, design, visual pattern recognition, analogical reasoning, financial reasoning, ma- chine learning, and case-based reasoning). Our general point here is that, in the case of ��� , abstractions have confused rather than clarified the modelling process. In this regard, the analysis of Motta & Zdrahal of the Sisyphus-2 applications to be particularly interesting. Motta & Zdrahal discuss the various Sisyphus-2 ��� implementations using their special knowledge of constraint satisfaction algorithms 35. We find their lower level more insightful into the construction process than the less-detailed, high-level ��� approach. This low-level view of a problem can find errors that experienced ��� practioners cannot. For example, Motta & Zdrahal argue that one declarative translation of the procedures in the Sisyphus-2 specification blurred the distinction between hard constraints (which must not be violated) and soft constraints (which can be optionally violated) 36. DISCUSSION Based on our experience with an equivalent line of research in expert systems, we have argued that OO patterns are not a productive reuse tool. This is a counter-intuitive position. The overwhelming intuition is that the abstractions offered by OO reuse design patterns and ��� are true and useful for the construction of new systems. Ralph Johnson comments: 18 T.J. MENZIES The reason that I believe that OO design patterns are reusable is because I reuse them. The reason I believe they help people design is because people who learn them tell me that they help them design. The reason that I am sure that different groups of designers use different patterns is because I’ve seen it happen 37. This quote sounds a lot like something from the KADS community; i.e. abstract descriptions of old designs are a penetrating and useful insight into the design process. However, after a decade of ��� research, we now doubt this insight, at least for ��� : � Abstract patterns developed by different developers can be different.� The number of abstract patterns seems unbounded; practioners keep inventing new one.� Practioners don’t reuse each others’ supposedly reusable abstractions.� When we actually experiment with supposedly reusable patterns and productivity, (e.g. the Cor- bridge study) we see evidence to support the counter-intuitive conclusion that well-formed ma- ture supposedly reusable patterns are less productive than no pattern at all. So, what is the appropriate use of the reuse patterns offered by (e.g.) GOF, GOV, Fowler, KADS, etc? We make two suggestions. Firstly, we should monitor OO patterns for ��� -type problems. Secondly, we not use them as objective canonical versions of truth, but as an assistant in analysis and design. Monitoring Potential Problems with Patterns Reuse We hope we have, at the very least, motivated the need for experimentation to test if patterns are indeed reusable. This section describes a series of such tests, If, in the year 2007 we see the following, then we can conclude that the reuse benefits for OO patterns are illusionary: � Between different OO “reusable” patterns, common processing elements are identified suggest- ing that seemingly different “reusable” patterns have a significant overlap.� The extra layer of abstraction used in OO “reusable” patterns confuses rather than clarifies design issues.� When controlled experiments are performed measuring development productivity, then no dif- ference is detected between those applications that use and do not use OO “reusable” design patterns.� In the future, we find that low-level details (e.g. constraint-satisfaction algorithms or and-or graph processing) become the focus of the design process.� OO “reusable” patterns prove not to be reusable.� Different developers working on the same problem develop different “reusable” patterns.� For the above two reasons, libraries of “reusable” design patterns from different sources contain significant deviations. On the last point, we note that we can already see differences of opinion of the exact nature of the patterns: � The GOV’s design-level patterns are different to those of the GOF (page 380 of 2);� There exists a non-trivial overlap between 3 of the GOV patterns: MVC (page 125 of 2), View Handler (page 291 of 2), and Publisher/Subscribe (page 239 of 2). We would propose to combine them. OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 19 � We place a very different emphasis to Fowler on graph-theoretic designs. Fowler offers 1.5 pages on “plans and protocols as graphs”(pages 166-68 of 3). We have argued above that generalised and-or graphs are a engineering framework useful in numerous contexts. For example, recall TINA’s search for candidates that could contribute to the current diagnostic problem (Figure 16). Such a search is greatly simplified if the program can access the dependency graph between business concepts. So, in our graph-theoretic view, we would propose to replace much of the diagnostic architecture of Fowler chapters 3 & 4 with a more intricate version of “plans and protocols as graphs”. With respect to these last two points, it is a debatable point whether or not our design proposals are better than those given by the GOV or Fowler. We argue below that it is important that such debates take place, in public, and in front of students of OO design. Constructivist Patterns Having raised some doubts over patterns, we still believe patterns are useful. When training OO practioners, we have observed that patterns are a powerful tool for communicating the insights of experienced designers to less-experienced designers. However, it may be a mistake to assume that just because students study patterns, that they will actually use them in their own work. This section uses some theory from computer-based education to claim that while patterns may not be useful as objective records of canonical truth, they may be very useful in supporting analysts and designers as they construct their systems. That is, while we doubt the reuse benefit, we are encouraged by the guidance benefit of patterns. Patterns may be best viewed as tools for structuring an argument, rather than recording a conclusion. Objectivism vs Constructivism Education theorists are turning away from objectivist and towards constructivist approaches to edu- cation 38. In objectivism, the teacher presents canonical forms of truth to students. In constructivism, the teacher presents a range of different viewpoints on the one issue to students. The training session and the students’ role is different in both approaches. In objectivism, students must memorise facts or models which they will apply later. In constructivism, (i) teachers support students as they explore a debate over some issue; and (ii) students try to build their own views of best practice. An objectivist teacher may present some impressive theoretical mental framework summarising their own view of a field. This summary may have taken years to synthesis. An constructivist teacher knows that while such summaries marked a significant watershed in their own understanding of a field, they may not be the final goal for the students. Rather, such theoretical frameworks may be useful to students only as guides to reviewing a field while each student constructs their own personal mental frameworks. Our reading of the current ��� and OO patterns literature is that it is tacitly objectivist: i.e. the libraries are assumed to be canonical forms of best solutions and a source of power in their own right. The constructivist process of using a pattern to explore options has not been a major focus of the current OO or ��� patterns literature. A Constructivist use of OO Patterns A hypothetical constructivist patterns toolkit would support the following features: � Practioners could browse a library of ��� or OO solutions to the same problem. Instead of reading up on “one” solution, a constructivist patterns text would present multiple solutions to 20 T.J. MENZIES (e.g.) the diagnostic problem (as done above when we discussed are ��� abstractions reusable?) or the update problem between multiple views. Techniques from case-based reasoning may be relevant here 39.� Each alternative solution to a problem is annotated with its strengths and weaknesses. Within the OO world, the Fowler, Coad, and Riel 40 texts are a first step in this direction. Fowler presents his patterns in an evolutionary form. From some similar initial design, Fowler takes the reader through a discussion of when the current form would suffice and when it should be made more intricate. Coad takes a simpler “extend, discuss, refine” approach, but in greater detail. Riel presents small designs (less than ten classes) along with a dialogue describing the strengths and weaknesses of those design. He then makes some preliminary remarks about patterns in design modifications which he calls transformational design patterns (chapter 10 of 40).� Practioners can try parts of those models on their own problem and when they do, some tool encourages them to explore the strengths and weaknesses of that approach. Exactly how this is done is an open research issue. Riel’s design heuristics might be suitable for small-designs (though they are far from being universally accepted). However, heuristics for assessing larger- scale 00 architectures or alternative ��� inference skeletons are still poorly documented. Interim Advice on Using Patterns The open issues in the last two points of the previous section could be the basis for an research programme that could take many years to terminate. Practioners cannot wait on the outcomes of such a programme. In the interim, we offer the following advice on how to use the current generation of OO or ��� patterns in a constructivist manner: � Seek alternative patterns for the same context.� When studying the alternatives, form a group and argue each alternative with respect to some concrete design problem.� Record not only the resolution of the debate, but the steps in the argument that resulted in that resolution.� Heuristics suitable for your local institution will be found within those argument steps. Whenever options are being considered, try and discover how your group explores and evaluates options.� Seek techniques for optimising this exploration/evaluation process. We speculate that these locally-generated heuristics for reviewing a design may become a tool at least as powerful as reading patterns. CONCLUSION We have distinguished three potential benefits from OO patterns: reuse, guidance, and communica- tion. Based on our experience with similar patterns research in expert systems, we doubt the reuse benefit, but endorse the communication and guidance benefit. The real power of patterns is how they encourage practioners to succinctly describe and review the essence of complex architectures. Pat- terns should not be viewed as objective canonical knowledge, but tools for supporting a constructivist approach to analysis and design. ACKNOWLEDGEMENTS Ralph Johnson’s enthusiasm for exploring an alternative view made this paper possible. Norm Kerth pointed out the guidance benefit of patterns. The comments of the anonymous reviewers prompted the OBJECT-ORIENTED PATTERNS: LESSONS FROM EXPERT SYSTEMS 21 creation of half of the discussion section and a significant clarification of our mapping from ��� to OO patterns. REFERENCES 1. E. Gamma, R. Helm, R. Johnson, and J. Vlissides, Design Patterns: Elements of Reusable Object-Oriented Software, Addison-Wesley, 1995. 2. F. Buschmann, R. Meunier, H. Rohnert, P. Sommerlad, and M. Stal, A System of Patterns: Pattern-Oriented Software Architecture, John Wiley & Sons, 1996. 3. M. Fowler, Analysis Patterns: Reusable Object Models, Addison Wesley, 1997. 4. P. Coad, D. North, and M. Mayfield, Object Models: Strategies, Patterns, and Applications, Prentice Hall, 1997. 5. W. Cunningham, ‘The checks pattern language of information integrity’, J. Coplien and D. Schmidt (eds.), Pattern Languages of Program Design. Addison-Wesley, 1995. Also available at http://c2.com/ppr/checks.html. 6. N. Kerth, ‘Caterpillar’s fate: A pattern language for transformation from analysis to design’, J. Coplien and D. Schmidt (eds.), Pattern Languages of Program Design. Addison-Wesley, 1995. Also available from http://c2.com/ppr/catsfate.html. 7. B. Chandrasekaran, ‘Towards a Taxonomy of Problem Solving Types’, AI Magazine, 9–17 (1983). 8. T.J. Menzies, ‘Limits to Knowledge Level-B Modeling (and KADS)’, Proceedings of AI ’95, Australia. World-Scientific, 1995. 9. G. Booch, I. Jacobsen, and J. Rumbaugh, Version 1.0 of the Unified Modeling Language, Rational, 1997. http://www.rational.com/ot/uml/1.0/index.html. 10. W.J. Clancey, ‘Model Construction Operators’, Artificial Intelligence, 53, 1–115 (1992). 11. T.E. Rothenfluh, J.H. Gennari, H. Erikson, A.R. Puetra, W. Tu, and M.A. Musen, ‘Reusable ontologies, knowledge-acquisition tools and performance systems: Protege-II solutions to sisyphus-2’, B.R. Gaines and M. Musen (eds.), Proceedings of the 8th AAAI-Sponsored Banff Knowledge Acquisition for Knowledge-Based Systems Workshop, 1994, pp. 43.1–43.30. 12. R. Benjamins, ‘On a role of probem solving methods in knowledge acquisition- experiments with diagnostic strategies’, Proceedings of the European Knowledge Acquisition Workshop, 1994, 1994. 13. D. Marques, G. Dallemagne, G. Kliner, J. McDermott, and D. Tung, ‘Easy Programming: Empowering Peo- ple to Build Their own Applications’, IEEE Expert, 16–29 (1992). 14. B.J. Wielinga, A.T. Schreiber, and J.A. Breuker, ‘KADS: a Modeling Approach to Knowledge Engineering.’, Knowledge Acquisition, 4, 1–162 (1992). 15. B.G. Buchanan and E.H. Shortliffe, Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, Addison-Wesley, 1984. 16. V.L. Yu, L.M. Fagan, S.M. Wraith, W.J. Clancey, A.C. Scott, J.F. Hanigan, R.L. Blum, B.G. Buchanan, and S.N. Cohen, ‘Antimicrobial Selection by a Computer: a Blinded Evaluation by Infectious Disease Experts’, Journal of American Medical Association, 242, 1279–1282 (1979). 17. D.S.W. Tansley and C.C. Hayball, Knowledge-Based Systems Analysis and Design, Prentice-Hall, 1993. 18. M. Linster and M. Musen, ‘Use of KADS to Create a Conceptual Model of the ONCOCIN task’, Knowledge Acquisition, 4, 55–88 (1992). 19. J.O. Coplien, ‘Idioms and patterns as architectural literature’, IEEE Software, 36–42 (1997). 20. C. Alexander, S. Ishikawa, S. Silverstein, I. Jacobsen, I. Fiksdahl-King, and S. Angel, A Pattern Language, Oxford University Press, 1977. 21. M. Shaw and D. Garlan, Software Architecture: Perspectives on an Emerging Discipline, Prentice Hall, 1996. 22. W. Clancey, ‘Heuristic Classification’, Artificial Intelligence, 27, 289–350 (1985). 23. A.T. Schreiber, B.J. Wielinga, and J.M. Akkermans, ‘Using KADS to Analyse Problem Solving Methods’, in J. Balder and H. Akkermans (eds.), Formal Methods for Knowledge Modeling in the CommonKADS Method- ology: A Compilation (KADS-EE/T1.2/TR/ECN/014/1.0), Netherlands Energy Research Foundation, 1992, pp. 53–90. 24. V.R. Benjamins, ‘Problem-solving methods for diagnosis and their role in knowledge acquisition’, Interna- tional Journal of Expert Systems: Research & Applications, 8(2), 93–120 (1995). 25. R. Johnson, ‘Documenting frameworks using patterns’, Proceedings of OOPSLA’92, ACM SIGPLAN, 1992. 22 T.J. MENZIES 26. H. Akkermans, F. van Harmelen, G. Schreiber, and B. Wielinga, ‘A Formalisation of Knowledge-Level Mod- els for Knowledge Acquisition’, in J. Balder and H. Akkermans (eds.), Formal Methods for Knowledge Mod- eling in the CommonKADS Methodology: A Compilation (KADS-EE/T1.2/TR/ECN/014/1.0), Netherlands Energy Research Foundation, 1992, pp. 53–90. 27. C. Corbridge, N.P. Major, and N.R. Shadbolt, ‘Models Exposed: An Empirical Study’, Proceedings of the 9th AAAI-Sponsored Banff Knowledge Acquisition for Knowledge Based Systems, 1995. 28. M. Linster, ‘A review of sisyphus 91 and 92: Models of problem-solving knowledge’, in N. Aussenac, G. Boy, B. Gaines, M. Linser, J.-G. Ganascia, and Y. Kordratoff (eds.), Knowledge Acquisition for Knowledge-Based Systems, Springer-Verlag, 1992, pp. 159–182. 29. B. Bredeweg, ‘Expertise in qualitative prediction of behaviour’, Ph.D. Thesis, University of Amsterdam, 1992. 30. T.J. Menzies, ‘Applications of abduction: Knowledge level modeling’, International Journal of Human Com- puter Studies, 45, 305–355 (September, 1996). 31. J. DeKleer and B.C. Williams, ‘Diagnosing Multiple Faults’, Artificial Intelligence, 32, 97–130 (1987). 32. D. Poole, ‘Probabilistic Horn abduction and Bayesian networks’, Artificial Intelligence, 64(1), 81–129 (1993). 33. W. Hamscher, L. Console, and J. DeKleer, Readings in Model-Based Diagnosis, Morgan Kaufmann, 1992. 34. T.J. Menzies and P. Compton, ‘Applications of abduction: Hypothesis testing of neuroendocrinological qual- itative compartmental models’, Artificial Intelligence in Medicine (1997). To appear. 35. Z. Zdrahal and E. Motta, ‘An In-Depth Analysis of Propose & Revise Problem Solving Methods’, R. Mi- zoguchi, H. Motoda, J. Boose, B. Gaines, and P. Compton (eds.), Proceedings of the Third Japanese Knowl- edge Acquisition for Knowledge-Based Systems Workshop: JKAW ’94, 1994. 36. E. Motta and Z. Zdrahal, ‘The trouble with what: Issues in method-independent task specifications’, Pro- ceedings of the 9th AAAI-Sponsored Banff Knowledge Acquisition for Knowledge-Based Systems Workshop Banff, Canada, 1995. 37. R. Johnson. (personal communication). 38. T.C. Reeves, ‘Evaluating what really matters in computer-based education’, in M. Wild and D. Kirkpatrick (eds.), Computer education: New Perspectives, MASTEC, 1994, pp. 219–246. 39. J.L. Kolodner, ‘Improving Human Decision Making Through Case-Based Decision Aiding’, AI Magazine, 68 (1991). 40. A.J. Riel, Object-Oriented Design Heuristics, Addison-Wesley, 1996. Note that some of the Menzies references can be obtained from 021.1.3�5�6!6�7!7!7�9;K.:MJW9)A4E=: 7�9[J4\4A�9]CMA-6^ 1L<_O!O�683.A!`&64\2H�K!:46�3�CM3�J4G-:MH8E-FM?a9)0.18O�F . 
work_uwalrhrrubc2nc6hsopdwz6rte	----	Trust based recommender system using ant colony for trust computation Trust based recommender system using ant colony for trust computation Punam Bedi, Ravish Sharma ⇑ Department of Computer Science, Faculty of Mathematical Sciences, Opposite Daulat Ram College, University of Delhi, Delhi 110007, India a r t i c l e i n f o Keywords: Collaborative Filtering Trust Recommender system Ant colony Pheromone updating a b s t r a c t Collaborative Filtering (CF) technique has proven to be promising for implementing large scale recom- mender systems but its success depends mainly on locating similar neighbors. Due to data sparsity of the user–item rating matrix, the process of finding similar neighbors does not often succeed. In addition to this, it also suffers from the new user (cold start) problem as finding possible neighborhood and giving recommendations to user who has not rated any item or rated very few items is difficult. In this paper, our proposed Trust based Ant Recommender System (TARS) produces valuable recommendations by incorporating a notion of dynamic trust between users and selecting a small and best neighborhood based on biological metaphor of ant colonies. Along with the predicted ratings, displaying additional information for explanation of recommendations regarding the strength and level of connectedness in trust graph from where recommendations are generated, items and number of neighbors involved in pre- dicting ratings can help active user make better decisions. Also, new users can highly benefit from pher- omone updating strategy known from ant algorithms as positive feedback in the form of aggregated dynamic trust pheromone defines ‘‘popularity’’ of a user as recommender over a period of time. The per- formance of TARS is evaluated using two datasets of different sparsity levels viz. Jester dataset and MovieLens dataset (available online) and compared with traditional Collaborative Filtering based approach for generating recommendations. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction The complexity of the recommendation problem is due to its vast space of possibilities. Recommender systems are important tools that overcome the information overload by sifting through the large set of data and recommending information relevant to the user. Rec- ommendation techniques have a number of possible classifications including content based, collaborative, knowledge-based, demo- graphic, and utility based (Burke, 2002; Resnick & Varian, 1997; Schafer, Konstan, & Reidl, 1999; Terveen & Hill, 2001). Collaborative Filtering, Content based approach and Hybrid methods are the pre- valent three approaches to developing recommender systems. In Collaborative Filtering (CF) approach, opinions from users in the form of ratings on various items are collected. The recommendations produced are based only on the opinions of users similar to the active user. Active user refers to recommendation seeker for whom recom- mendations are generated. Content based approach suggests items that are similar to the ones the active user has shown a preference for in the past rather than on the preferences of other users. Mostly CF technique is combined with one or more recommendation techniques i.e. content based, knowledge-based, demographic, or utility based and is called Hybrid approach. CF takes its roots from something humans have been doing for centuries i.e. sharing opinions with others and brings together the opinions of large interconnected communities on the web (Schafer, Frankowski, Herlocker, & Sen, 2007). Collaborative Filtering based recommenders work best for a user who fits into a niche with neighbors of similar taste. This approach doesnot need a represen- tation of the items in terms of features but is based only on the judgment of the user community. Because of its simplicity in both theory and implementation, CF can be applied to virtually any kind of item viz. papers, news, web sites, movies, songs, books, jokes, locations of holidays, stocks etc. The three-step conceptual model of the operation of the Collaborative Filtering process is as follows: (i) Construct rating matrix using user–item data. (ii) Locate people (neighbors) with similar profiles (similar preferences). (iii) Unify neighbors’ ratings to form recommendations. There are several limitations and challenges to CF based recom- mender systems due to traditional emphasis on user similarity. One of the main limitations is that it provides recommendations based solely on the opinion of the users whose preferences best matches the taste of active user. Approaches incorporating trust 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.07.124 ⇑ Corresponding author. Address: B-408, Flex Apartments, Plot No. C-58/22, Sector 62, Noida 201301, U.P., India. Tel.: +91 09313907606; fax: +91 11 27662553. E-mail addresses: pbedi@cs.du.ac.in (P. Bedi), sharma_ravish123@rediffmail.com (R. Sharma). Expert Systems with Applications 39 (2012) 1183–1190 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.07.124 mailto:pbedi@cs.du.ac.in mailto:sharma_ravish123@rediffmail.com http://dx.doi.org/10.1016/j.eswa.2011.07.124 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa models into recommender systems are gaining momentum, syn- thesizing recommendations based upon opinions from trusted peers rather than most similar ones (Ziegler & Nilcolas, 2007). In online recommendation generation process, a number of users form a virtual social network where trust relationship is the key of users’ human relations and trust relationship of mutual interde- pendence forms a so-called Web of trust (Wei, 2007). The three properties assigned to trust as discussed in literature (Ries, Kanga- sharju, & Muhlhauser, 2006) are as follows: (i) Trust is subjective and therefore asymmetric. (ii) Trust is context dependent. (iii) Trust is dynamic which means that it can increase with posi- tive experience and decrease with negative experience or over time without any experience. The dynamic property of trust can be viewed as trust intensity among the users. For example, the trust intensity between recom- mender and active user may increase or decrease depending on recommendations generated by the recommender. Trust intensity being time-based information, can be analyzed using pheromone updating strategy known from ant algorithms. Ant algorithms are based on emulation of behavior of real ants and build better solu- tions by communicating through artificial pheromone. In many ant species, ants tend to lay a substance called pheromone while walk- ing from their nests to food source and vice versa. Ants do not com- municate directly with each other, but they follow pheromone trails laid by other ants. Ants are attracted by pheromones coming from fellow type (members of same species) ants and repulsed by pheromone of non-fellow type ants. As the time passes, paths that are marked by stronger amount of pheromone are chosen with higher probability than those that have weaker amount of phero- mone deposit (Blum, 2005; Bonabeau, Dorigo, & Theraulaz, 1999; Dorigo, Maniezzo, & Colorni, 1996; Dorigo & Stutzle, 2004). One of the basic ingredients of the ant algorithms is a randomized gree- dy construction heuristic in which each option is selected with a probability proportional to its perceived quality. As the algorithm progresses, the probabilities are altered to encourage the selection of elements that have featured good solutions. Ant algorithms repeatedly construct new solutions from scratch throughout the duration of the search. Moreover, this collaborative behavior be- tween fellow type ants has an analogy with the collaborative world as people mostly collect opinions from their like-minded friends, neighbors etc. Work by Bedi, Sharma, and Kaur (2009) and Sharma, Singh, Makkar, Kaur, and Bedi (2007) combine artificial pheromone information with similarity measure for choosing best matching clusters providing active user with good set of alternative recom- mendations. In addition to the taste of active user, clusters that are marked by stronger amount of pheromone have the higher probability of being chosen than those that have weaker amount of pheromone deposit. Also, one of the fundamental challenges for recommender systems is to improve the quality of the pre- dicted ratings. In such a scenario, pheromone updating strategy of ants guides the search to a better recommendation. The positive feedback in the form of pheromone deposition results in achieving an emergent, unified behavior for the recommender system as a whole, and produces a robust system capable of finding improved quality recommendations. Still, there is a need for producing valu- able recommendations due to data sparsity of input-rating matrix and solving new user problem. These problems are intrinsic in the process of finding similar neighbors. In our proposed approach TARS (Trust based Ant Recommender System), sparseness in user similarity due to data sparsity of input matrix is reduced while creating directed trust graph for each user. Weight on the edge rep- resents strength of connectedness i.e. trust intensity between the two recommendation partners (recommender and active user) at time t. Due to its dynamic property, it is called Dynamic Trust Pher- omone (DTP) and is updated using pheromone updating strategy known from ant algorithms. As the time passes, recommendations are produced by continuously updating dynamic trust between users and selecting a small and best neighborhood using ant colony metaphor. Also, new users can highly benefit from pheromone updating strategy known from ant algorithms. The rest of the paper is organized as follows: Section 2 reviews the related research work and presents the motivation for our work. Our proposed system TARS with a detailed description of each step is presented in Section 3. Section 4 explains the experi- mental methodology employed and the two datasets used in our experiments. Section 5 concludes the paper. 2. Related work In this section, review of literature related to recommender sys- tems and motivation for our work is presented. Many collaborative systems have been developed in the acade- mia and the industry. Using stereotypes, the Grundy system was developed to build individual user models which were used to rec- ommend relevant books to each user (Rich, 1979). Later on, the Tapestry system relied on each user to identify like-minded users manually (Goldberg, Nicholas, Oki, & Terry, 1992). GroupLens (Konstan et al., 1997; Resnick, Lakovou, Sushak, Bergstrom, & Riedl, 1994) in the domain of Usenet newsgroup articles, Bellcore’s Video Recommender (Hill, Stead, Rosenstein, & Furnas, 1995) in the domain of movies and Ringo (Shardanand & Maes, 1995) in the do- main of music and musical artists were the first systems to use Col- laborative Filtering algorithms to automate predictions. Other examples of collaborative recommender systems include the book recommendation system from Amazon.com, the PHOAKS system that helps people find relevant information on the WWW (Terveen, Hill, Amento, McDonald, & Creter, 1997), and the Jester system that recommends jokes (Goldberg, Roeder, Gupta, & Perkins, 2001). Dif- ferent item-based recommendation algorithms are analyzed by Sarwar, Karypis, Konstan, and Riedl (2001). Researchers have also applied Swarm Intelligence techniques to recommender systems. Ujjin and Bentley (2003) have described Particle Swarm Optimization (PSO) recommender system in which PSO algorithm has been employed to learn personal preferences of users and provide tailored suggestions. A system called CASIS has been developed combining case based reasoning approach with a metaphor from colonies of social insects, namely the honey bee dance. This combination has been used in the retrieval step of the recommendation cycle (Lorenzi, Santos, & Bazzan, 2005). Web-based system user interface hybrid recommendation method is presented in (Sobecki, 2007) where ant colony metaphor is used for selecting the most optimal path in the user interface graph. In addition to traditional emphasis on user similarity, trustwor- thiness of users has been an important consideration by research- ers. They have suggested that the advantage of combining users’ trust network and user–item rating matrix solves CF sparseness problem to some extent. Two computational models of trust namely profile-level trust and item level trust have been developed and incorporated into standard Collaborative Filtering frameworks (O’Donovan and Smith, 2005). Work by Massa and Bhattacharjee (2004) and Massa and Avesani (2009, 2007, 2004) focus on using the explicit trust as input along with user–item rating matrix to predict ratings. They have analyzed explicit trust from the popular Internet web site epinions.com and shown that by exploiting the web of trust, it is possible to propagate trust and infer an additional weight for other users. New users can also benefit from trust prop- agation as long as the users provide at least one trusted friend. Although the explicit trust based recommender system models have high rating prediction coverage and high rating prediction 1184 P. Bedi, R. Sharma / Expert Systems with Applications 39 (2012) 1183–1190 https://isiarticles.com/article/7769 
work_uysid7rerjf55i5cagmsmarm4e	----	A combined methodology for supplier selection and performance evaluation A combined methodology for supplier selection and performance evaluation Mithat Zeydan a,⇑, Cüneyt Çolpan b, Cemal Çobanoğlu c a Department of Industrial Engineering, Engineering Faculty, Erciyes University, 38039 Kayseri, Turkey b Department of Quality Assurance, 2nd Turkish Air Supply and Maintenance Center Command, Kayseri, Turkey c Department of Procurement, Hyundai Assan Automotive Factory, _Izmit/Kocaeli, Turkey a r t i c l e i n f o Keywords: Supplier evaluation Multi-criteria decision making techniques Performance improvement a b s t r a c t Today, organizations that wish to carry on the sustainable growing need a robust strategic performance measurement and evaluation system because of changing demands of consumers, reduced product life cycle, competitive and globalised markets. In this study, a new methodology is introduced and proposed for increasing the supplier selection and evaluation quality. The new approach considers both qualitative and quantitative variables in evaluating performance for selection of suppliers based on efficiency and effectiveness in one of the biggest car manufacturing factory in Turkey. This new methodology is realized in two steps. In the first stage, qualitative performance evaluation is performed by using fuzzy AHP (Ana- lytical Hierarchical Process) in finding criteria weights and then fuzzy TOPSIS (Technique for Order Pref- erence by Similarity to Ideal Solution) is utilized in finding the ranking of suppliers. So, qualitative variables are transformed into a quantitative variable for using in DEA (Data Envelopment Analysis) methodology as an output called quality management system audit. In the second stage, DEA is per- formed with one dummy input and four output variables, namely, quality management system audit, warranty cost ratio, defect ratio, quality management. As a result, comparing with the present system applied by the car factory, the new method seems to be some advantages and superiorities for making the decision in buying the quality car luggage side part (panel) by selecting the suitable supplier(s) in an automotive factory of Turkey. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction To choose the right supplier deals, with an important evalua- tion, and selection problems in the purchasing function of a busi- ness. A good supplier selection makes a significant difference to an organization’s future to reduce operational costs and improve the quality of its end products. There have been a lot of factors in today’s global market in which that influence companies to search for a competitive advantage by focusing on purchasing raw mate- rials and component parts represents the largest percentage of the total product cost. For instance, high technology products such as motor vehicles, railroad&transport equipment, machinery&equip- ment components, purchased materials and services account for up to 80% of the total product cost. Therefore, selecting the right suppliers is a key to the procurement process and represents a ma- jor opportunity for companies to reduce costs. On the other hand, selecting the wrong suppliers can cause operational and financial problems (Weber, Current, & Benton, 1991). The traditional ap- proach to supplier selection has been to select suppliers solely on the basis of price for many years. However, as companies have learned that price as a single criterion for supplier selection is insufficient, they have turned into more comprehensive multi-cri- teria decision making techniques. Recently, these criteria have be- come increasingly complex as environmental, social, political, and customer satisfaction concerns have been added to the traditional factors of quality, delivery, cost, and service. Apart from cost reduc- tion, companies continuously work with suppliers to remain com- petitive by reducing product development time, improving product quality, and reducing lead times. For instance, a qualified base of suppliers helps a company achieve greater innovation through im- proved product design and increased flexibility. Some authors have identified several criteria for supplier selection, such as the net price, quality, delivery, historical supplier performance, capacity, communication systems, service, and geographic location, among others (Dempsey, 1978). These evaluation criteria involve trade- offs and are a key issue in the supplier assessment process since it measures the performance of suppliers. For example, one vendor may offer inexpensive parts of slightly below average quality, while another vendor may offer higher quality parts, with uncer- tain delivery thus setting up trade-offs. In addition, the importance of each criterion, varies from one purchase to the next and is com- plicated further by the fact that some criteria are quantitative (price, quality, etc.), while others are qualitative (service, flexibil- ity, etc.). Thus, a technique is needed that can adjust for the deci- sion maker’s attitude toward the importance of each criterion 0957-4174/$ - see front matter � 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.08.064 ⇑ Corresponding author. Tel.: +90 352 4374901/32454; fax: +90 352 4375784. E-mail address: mzeydan@erciyes.edu.tr (M. Zeydan). Expert Systems with Applications 38 (2011) 2741–2751 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2010.08.064 mailto:mzeydan@erciyes.edu.tr http://dx.doi.org/10.1016/j.eswa.2010.08.064 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa and incorporates both qualitative and quantitative factors (Bhutta & Huq, 2002). The overall objective of the supplier evaluation pro- cess is to reduce risk and maximize overall value to the purchaser. An effective supplier survey should have certain characteristics such as comprehensiveness, objectiveness, reliability, flexibility and finally, it has to be mathematically straightforward. It can be concluded that important savings can be realized through effective purchasing strategies. This study helps decision makers reduce a base of potential suppliers to a manageable number and make the supplier selection by means of multi-criteria techniques. This new methodology was applied to a car manufacturing facility in Turkey. 2. Literature review Supplier evaluation is a multi-objective and criteria decision making problem containing many quantitative and qualitative fac- tors because there are typically more than one criterion (attitude) needed to be taken into consideration in evaluating a supply source. All of supply sources are focused on their performance such as delivery, quality, service and price as the main factors that all firms use for evaluating sources of supply (Ha & Krishnan, 2008). Many firms and researchers have been working on the supplier evaluation problem over the past decade to develop decision mak- ing models which can effectively deal with this problem. According to Ghodsypour and O’Brien (1998), optimization models for sup- plier evaluation can be classified into two groups: single objective models which are used to consider one criterion as the objective function and other criteria as constraints. The single objective models have two disadvantages: all criteria are equally weighted, which rarely happens in practice, and they have significant difficul- ties in considering qualitative factors. In contrast, the multiple objective models have been applied to a supplier evaluation prob- lem. Relying on a single criterion makes the supplier selection pro- cess risky. Therefore, a multi-criteria approach is recommended. A pioneering work in supplier selection criteria was that of Dickson in 1966. Despite the multiple criteria nature of the problem, very little work has been devoted to the study of the supplier selection problem by using multi-criteria techniques such as goal program- ming, multi-objective programming, or other similar approaches. Kahraman, Cebeci, and Ulukan (2003) used fuzzy AHP to select the best supplier for a manufacturer firm established in Turkey. Bevilacqua and Petroni (2002) developed a system for supplier selection using fuzzy logic. Some authors as used in this paper have combined decision models in the supplier selection process, for example, Weber, Current, and Desai (1998) combined DEA and mathematical programming models. This combination provided decision makers with a tool for negotiating with suppliers. Dickson (1966) developed a model combining mathematical programming model and TCO (total cost of ownership). They derived the inven- tory management policy using activity-based costing information. Ghodsypour and O’Brien (1998) used AHP and mathematical pro- gramming to determine the best order quantity allocation while considering qualitative criteria into the analysis. Xia and Wu (2007) presented an integrated approach of AHP improved by rough sets theory and multi-objective mixed integer programming. Dulmin and Mininno (2003) applied a model to a mid-sized Italian firm operating in the field of public road and rail transportation by applying a multi-criteria decision making technique (promethee/ gaia) to supplier selection problem. The supplier selection problem is complicated and risky, owing to a variety of qualitative and quantitative factors affecting the decision making process. There have been several supplier selection methods available in the liter- ature. Some authors propose linear weighting models in which suppliers are rated on several criteria and in which these ratings are combined into a single score. These models include the cate- gorical method which relies heavily on the experience and ability of the individual buyer, the weighted point (Timmerman, 1986) and the analytical hierarchical process (Nydick & Hill, 1992). Total cost approaches attempt to quantify all costs related to the selec- tion of a vendor in monetary units, this approach includes cost ra- tio (Timmerman, 1986) and total cost of ownership (Dulmin & Mininno, 2003). Mathematical programming models often con- sider only the more quantitative criteria; this approach includes the principal component analysis (Petroni & Braglia, 2000) and neural network (Lovell & Pastor, 1999). The neural network for supplier selection is another method that has been developed to help selecting the best supplier. Com- paring to conventional models for decision support system, neural networks save a lot of time and money of system development. The supplier-selecting system includes two functions: one is the func- tion measuring and evaluating performance of purchasing (quality, quantity, timing, price and costs) and storing the evaluation in a database to provide data sources to neural network. The other is the function using neural network to select suppliers. ANN was also applied to the supplier evaluation problem by imitating the decision process of a buyer for supplier selection (Lovell & Pastor, 1999). Nevertheless, these models are still lacking of the capability to deal with uncertainty which is usually present in the supplier selection problem. Carrera and Mayorga (2008) proposed a Fuzzy Inference System (FIS) approach in supplier selection for new prod- uct development. Experts agree that no best way exists to evaluate and select suppliers (Bello, 2003), and thus organizations use a variety of approaches and implements the one that suits best depending on the company’s particular requirements. Many previ- ous researches, in vendor evaluation, emphasizes conceptual and empirical decision support models that may suffer from certain shortcomings, such as being mathematically too complex or too subjective. Practical appreciation needs a methodology that is sim- ple to use and understand, but yet it shall produce reasonably accurate results. There have been a lot of hybrid methods em- ployed in the last 10 years at the literature in terms of supplier evaluation and selection methods (Morlacchi, 1999; Simpson, Si- guaw, & White, 2003; Weber, Current, & Desai, 2000; Wang, Huang, & Dismukes, 2004; Bello, 2003). Fuzzy AHP, fuzzy TOPSIS and DEA are commonly used in the literature separately or some- times their combinations can be used at the same time. There has not been any study in the literature about a hybrid fuzzy AHP/fuz- zy TOPSIS/DEA approach before. When the literature is widely looked through, MCDM (Multi-Criteria Decision Making) tech- niques generally used are focused on TOPSIS (or fuzzy TOPSIS), AHP (or fuzzy AHP) and DEA. There have been some advantages and disadvantages when compared with each other, in terms of AHP and TOPSIS (Zeydan & Çolpan, 2009). Also, fuzzy AHP and fuz- zy TOPSIS are combined in this study. But, it is the first time in the literature that weights are used by transforming qualitative vari- ables into only one quantitative variable in fuzzy TOPSIS and found as triangular fuzzy numbers with fuzzy AHP. The hybrid method in the first step uses fuzzy AHP to assign criteria weight and then ranks all suppliers for the qualitative selection by using fuzzy TOP- SIS. The result obtained from fuzzy TOPSIS is used in DEA as an out- put variable called quality management system audit (QMSA). In the second step, it uses DEA methodology in order to choose effi- cient vendors in the final selection process. 3. Proposed hybrid method for supplier selection and evaluation We used three multi-criteria decision making method to find efficient and inefficient suppliers sensitively. In the first step, these are fuzzy AHP for the determination of criteria weights and fuzzy 2742 M. Zeydan et al. / Expert Systems with Applications 38 (2011) 2741–2751 https://isiarticles.com/article/19255 
work_uzz76ezssrdgdloaq22rqagi4e	----	Dynamic Operator Instructions Based on Augmented Reality and Rule-based Expert Systems Available online at www.sciencedirect.com 2212-8271 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the scientific committee of 48th CIRP Conference on MANUFACTURING SYSTEMS - CIRP CMS 2015 doi: 10.1016/j.procir.2015.12.113 Procedia CIRP 41 ( 2016 ) 346 – 351 ScienceDirect 48th CIRP Conference on MANUFACTURING SYSTEMS - CIRP CMS 2015 Dynamic operator instructions based on augmented reality and rule-based expert systems Anna Syberfeldt*,a, Oscar Danielsson a, Magnus Holm a, Lihui Wang a b aUniversity of Skövde, PO-54148, Skövde, Sweden bRoyal Institute of Technology, PO-1148, Stockholm, Sweden. * Corresponding author. Tel.: +46(0)500448577. E-mail address: anna.syberfeldt@his.se Abstract Augmented reality is currently a hot research topic within manufacturing and a great potential of the technique is seen. This study aims to increase the knowledge of the adaptation and usability of augmented reality for the training of operators. A new approach of using dynamic information content is proposed that is automatically adjusted to the individual operator and his/her learning progress for increased efficiency and shorter learning times. The approach make use of the concept of expert systems from the field of artificial intelligence for determine the information content on-line. A framework called "Augmented Reality Expert System" (ARES) is developed that combines AR and expert systems. A proof-of-concept evaluation of the framework is presented in the paper and possible future extensions are discussed. © 2015 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the Scientific Committee of 48th CIRP Conference on MANUFACTURING SYSTEMS - CIRP CMS 2015. Keywords: Augmented reality; Expert systems; Industrial operator; Information content 1. Introduction Augmented reality (AR) is a hot topic in manufacturing research today. With AR, it is possible to give industrial operators access to information that their ordinary senses could not have gathered from reality, and to give this information in the context of where it is needed. The basic concept behind AR is to overlay digital information about the environment and its objects on the real world, and thereby enhance the perception of reality. Over the last few years, the technology enabling AR has advanced rapidly and a number of real-world applications of AR can be seen today; mainly within areas such as gaming, sports and tourism. Within the context of industrial applications, AR has been discussed for over 20 years. There exist plenty of studies on the topic (see for example [1-4]), but few practical applications. Tiefenbacher et al. [5] discuss that the limited success of AR in industrial applications is mainly due to the fact that the industrial setting is highly challenging and complex. For the big breakthrough of AR, additional research is needed on how to tackle these complexities and successfully develop useful applications. An industrial application area were the research on AR has been specifically limited is the adaptation and usability of AR for training of industrial operators within the manufacturing domain. Lee [6] discusses that there exist only a few general studies on the topic of industrial training, and none of these are in the manufacturing domain. This study aims to contribute to an increased knowledge of the adaptation and usability of AR for training of operators within manufacturing. The authors believe that increased knowledge is critical for utilizing the full potential of AR, and to achieve a widespread use of the technique within the manufacturing domain. The standpoint in the paper is that a new approach for information content handling is necessary for the effective use of AR for training purposes and the learning a new tasks. That is, a new approach for what information that is presented to an individual operator, at which point in time and in what form. Previous studies on AR in the industrial context have paid little attention to the information content, instead mainly focusing on evaluating the concept of AR as such. A common property of the majority of the existing AR solutions is that the information content displayed to the operator is © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the scientifi c committee of 48th CIRP Conference on MANUFACTURING SYSTEMS - CIRP CMS 2015 347 Anna Syberfeldt et al. / Procedia CIRP 41 ( 2016 ) 346 – 351 predetermined and static. This might work well when the purpose of the AR solution is simply to guide the operator through a number of steps, but from a learning perspective it can be argued that this approach is too inflexible. Since learning is a dynamic and individual process, the authors believe that the information content must be dynamic and individual as well. In this paper, an approach of using dynamic information content is proposed that is automatically adjusted to the individual operator and his/her learning progress. The approach make use of the concept of rule-based expert systems from the field of artificial intelligence for determine the information content on-line. An expert system is basically a computer program that emulates the knowledge and experience of a human expert based on rules. By using an expert system, an operator can learn how to perform a task based on expert knowledge – without the need to actually have human expert physically present. For realizing the new approach of combining AR and expert systems, a framework called "Augmented Reality Expert System" (ARES) is developed as part of the study. In the next section, the concepts of AR and experts systems are described in further detail for the unversed reader. In Chapter 3, the framework ARES is presented. Chapter 4 describes a proof-of-concept evaluation of ARES, while Chapter 5 presents conclusions from the study and outlines important topics to consider in the future. 2. Background In this chapter, AR and expert systems are presented in further detail. 2.1. Augmented reality The term “augmented reality” was first used by Caudel and Mizell in 1992 to refer to a heads-up, see-through display combined with head position sensing and workplace registration systems [11, p.660]: “This technology is used to “augment” the visual field of the user with information necessary in the performance of the current task, and therefore we refer to the technology as 'augmented reality' (AR).”. Later on, Krevelen and Poelman [12] formally defined an AR system as having the following features (based on [13]): (a) ability to combine real and virtual objects in a real environment, (b) ability to register (align) real and virtual objects with each other, and (c) ability to run interactively, in three dimensions, and in real time. It can be noticed that this definition does not limit AR to specific technologies, which is also supported by other definitions. For example, Azuma et.al. [14] and Kipper and Rampolla [15] state that AR can be applied to all senses, including smell, touch and hearing. In general, AR is implemented through some form of anchor in the real world for the AR system to be able navigate. The most common way to implement anchors is to use target images. By connecting virtual information and target images, the AR system can orientate the virtual objects correctly in relation to the real world objects. 2.2. Expert systems In the research field of artificial intelligence, an expert system is a computer program that emulates the decision- making ability of a human expert. Basically, an expert system is designed to solve complex problems by reasoning about knowledge that is represented as if–then rules. A recent definition of expert systems has been provided by Patel, Virparia and Patel [16, p. 11]: "An expert system can be defined as a set of programs that use the human expertise as knowledge which is stored in an encoded form and may manipulate it to solve problems in a specialized domain. An expert system’s knowledge must be coded and stored in the form which the system can use in its reasoning processes performed by the inference engine.”. The main motivation behind expert systems is the fact that human experts are a scarce resource and their time is highly valuable. If an expert system is able to take over some of the work of the expert it cannot only free time and save money, but also give unlimited access to expert knowledge 3. Framework This chapter describes the motivation behind the ARES framework, related work and the overall framework design. 3.1. Motivation The motivation behind ARES ("Augmented Reality Expert System") is the assumption that the guidance and instructions in an AR solution must depend on the previous experience and skills of the individual operator in order for him/her to complete and learn the task at hand in shortest time possible. The authors believe that with customized information content, the effectiveness of an AR solution can be significantly increased and thereby also the efficiency of the operator. Dynamically customizing instructions for individual operators is, however, no easy task. For each task and operator it must be decided in real-time what information that is important to show at the moment and in which form, as well as what information not to show. The latter is important since too much information is not only frustrating for the operator, but might even be detriment due to the high cognitive load imposed. For solving the problems of dynamic instructions and realizing customized information content, the authors propose to use the concept of expert systems. The idea is to use an expert system with the knowledge of human cognition to dynamically determine the information content in the AR solution. Through this approach, it is intended to provide the operator an enhanced view of reality optimized for his/her individual need at the moment. With such support, the hypothesis is that the operator can learn a new task more efficiently compared to using the common approach of predetermined, static information content. 348 Anna Syberfeldt et al. / Procedia CIRP 41 ( 2016 ) 346 – 351 3.2. Related work As far as the authors are aware, expert systems have not previously been used for enabling dynamic, individual information content in AR applications. However, there are previous studies on related thoughts and issues. For example, Hajovy and Christensen [7] discussed already back in 1991 the possibilities to use dynamic expert systems in teaching applications. The idea in [7] is to adjust the instructions presented to a user based on individual needs, instead of using static instructions. Ten years later, Pan [8] evolved the topic and explored the use of strategies that enable an expert system to adapt to, or learn from, interactions with users. The aim of the study was to develop a concept of an expert system that created and adjusted its rules based on the feedback received from the human user in real-time. Neither [7] nor [8] performed their studies within the context of AR, but in 2003 Wiedenmaier et al. [9] discussed expert knowledge in relation to AR applications. Wiedenmaier et al. did not use expert systems, but they compared AR with paper instructions and expert guidance for a typical industrial assembling task and discovered that the assembling was completed in shortest time when the user was guided by an expert, followed by the use of AR support and in last place paper instructions. This finding was important since it indicates the value of expert knowledge in AR applications. Radowski et al. [10] recently presented a study that investigated different visual features for AR-based assembly instructions. Although expert systems was not used in the study, it is relevant in the context since the authors found out through user tests that the visual features used to explain a particular assembly operation must be adjusted to its relative difficulty level. This is indicates the need of dynamic information content in AR applications. 3.3. Overall design ARES consist of two systems; a rule-based expert system and an AR system. From an overall perspective, the expert system is responsible for determine the information content (what information to show and when), while the AR system is responsible for displaying the information properly in the user interface. The overall design of ARES is shown in Figure 1 below. Fig. 1. ARES framework. The AR system in ARES is developed in C# in Unity 3D with the use of the Vuforia for realizing the AR functionality. Vuforia is a software development kit (SDK) for mobile devices that supports the creation of augmented reality application. The SDK implements vision technology to recognize and track image targets and simple 3D objects in real-time. With the image recognition feature it is possible to position and orient virtual objects in relation to real world objects when these are viewed through the camera of a mobile device. The virtual object tracks the position and orientation of the image in real-time so that the viewer’s perspective on the object corresponds with their perspective on the image target. In that way, it appears that the virtual object is a part of the real world. More information about Vuforia can be found at www.qualcomm.com. The rendering of AR objects is determined on-line by the expert system. The expert system is implemented based on rules in the OpenRules format (http://openrules.com/). OpenRules utilizes the concepts of workbooks, worksheets, and tables to build rule repositories. Basically, each OpenRules workbook is comprised of one or several worksheets that separate information into different categories. The logical and physical organization of OpenRules repositories is shown in Figure 3. In ARES, the rules are parsed dynamically and the AR content is updated in real-time according to the outcome of the rules. The rules are created by a human expert in an ordinary Excel file. Using an Excel file has the advantage that is makes it easy for the expert to create and change rules without the need to manipulate programming code. An example of three simple rules is given in Figure 2. These rules specify the rendering of a certain AR object (called “Object 1”) and is based on the time elapsed in combination with the status of the current subtask. Condition Condition Conclusion Seconds StepStatus Active Render Object1 < 5 is TRUE is FALSE >= 5 is TRUE is TRUE is FALSE is FALSE Fig. 2. Example of rules in expert system. Fig. 3. Architecture of OpenRules (from http://openrules.com). Expert system AR system User interface ARES framework Information content Grapichal AR objects User input Content request 349 Anna Syberfeldt et al. / Procedia CIRP 41 ( 2016 ) 346 – 351 The parsing of rules is done in an if-then manner; for instance, the second row in Table 1 is interpreted by the system in the following way: “If the operator has spent 5 seconds or more on the current subtask and if the currently checked subtask is active then the guidance provided to the operator should be in the form defined by Object1”. 4. Proof-of-concept evaluation For a proof-of-concept evaluation of ARES, a practical scenario in which ARES is used has been set up as part of the study. In this scenario, an operator has to perform battery maintenance of a solar- and wind powered street lamp. The street lamp, shown in Figure 4, has been developed as a student project at the university. The battery to be changed is placed inside the light grey box at the bottom on the green fixture shown in the pictures. The top cover of the box is locked with two screws that have to be removed before any work can be done on the battery. As can be seen in Figure 3, the operator uses a tablet which runs ARES. The tablet is a Nvidia SHIELD tablet with a 9” screen (an off-the-self product). For the scenario, six rules have been defined. These rules are based on three conditions: a) the number of seconds elapsed since starting ARES, b) the competence of the operator, and c) if the top cover of the battery box is removed or not. The rules are shown in Figure 5 below. The defined rules are relatively simple, but perfectly serve the purpose of the proof-of-concept evaluation. Fig. 4. Proof-of-concept evaluation. Condition Condition Condition Conclusion Seconds Competence Top cover Render < 15 is beginner is on Basic text instruction <= 30 is beginner is on Detailed text instruction > 30 is beginner is on Detailed text instruction + pointing arrows & tool to use > 0 is beginner is off Text instruction & graphical highlights > 60 is experienced is off Text instruction > 90 is experienced is off Graphical highlights Fig. 5. Implemented rules. In Figure 6-8 below, the outcome of the first three rules in Table 2 is shown as example (screenshots from user interface in the tablet). As can be seen in the pictures, the information content shown to the operator varies depending on which rule that is active. The image on the top of the cover showing the university’s logo is the target image for the AR functionality. Results from using ARES in the described scenario shows that the framework is fully functional and that AR and expert systems can be successfully integrated for enabling dynamic information content. In testing ARES in the scenario, four persons from the university’s staff were engaged. None of these had any prior knowledge of how to maintain the street lamp battery or any previous experience of using AR. The test persons were given the task of changing the battery without any additional instructions on how to do it or how to use the system. All four persons succeeded to perform the task immediately and without flaws, which indicates that implementation worked as intended. Fig. 6. Rule 1 - basic text instruction is shown. Fig. 7. Rule 2 - detailed text instructions is shown. 350 Anna Syberfeldt et al. / Procedia CIRP 41 ( 2016 ) 346 – 351 Fig. 8. Rule 3 - graphical objects in the form of pointing arrows and picture of the tool to use. In the next chapter, possible improvements of ARES are further discusses and conclusions from the study are outlined. 5. Conclusions and future work In this paper, it is argued that a new way of handling information content is necessary for the effective use of AR for training operators in learning new tasks. A new approach of using dynamic information content is proposed in which the information is automatically adjusted to the individual operator and his/her learning progress for increased efficiency and shorter learning times. To do this, a rule-based expert system with the knowledge of human cognition is implemented to dynamically determine the information content in the AR solution. Through this approach, it is intend to provide the operator an enhanced view of reality optimized for his/her individual need at the moment. With such support, the authors believe that the operator can learn a new task more efficiently compared to using the common approach of predetermined, static information content. For realizing the new approach, a framework called "Augmented Reality Expert System" (ARES) is developed that combines AR and expert systems. A proof-of-concept evaluation of the framework is performed using a scenario of battery maintenance. The evaluation shows that it is possible to successfully combine AR and expert systems, and that such approach has potential to be a value support for operators in the manufacturing industry. For proving the full potential of the approach, however, it must be evaluated on also on real, industrial tasks of high complexity. The authors intend to perform such evaluations in the near future in collaboration with the manufacturing company Volvo Cars in Sweden. Furthermore, the authors plan to perform extensive user studies involving industrial operators with different skills and experience in order to evaluate the actual efficiency gain as well as the operators’ experiences. Regarding future research, an important topic to investigate further is the possibility to automatically generate and modify the rules in the expert system. As it is now, the rules must be entered manually by a human expert. Although this is a one- time effort, it might be time consuming and the quality of the rules is dependent on the quality of the expert. A valuable improvement would be the possibility for the system automatically generate proper rules and to modify existing rules in case of changed circumstances. Significant benefits of such self-adaptive rules can be seen, but it is most probably non-trivial to implement. One possible way to realize it might be to use artificial neural networks (ANNs). In general terms, an ANN is a non-linear statistical data modelling method used to model complex relationships between inputs and outputs [17]. Originally, the inspiration for the technique was from the area of neuroscience and the study of neurons as information processing elements in the central nervous system. ANNs have universal approximation characteristics and the ability to adapt to changes through training. Instead of only following a set of rules, ANNs are able to learn underlying relationships between inputs and outputs from a collection of training examples, and to generalize these relationships to previously unseen data. Theoretically, these attributes make ANNs very suitable to be used for implementing self-adaptive rules in an expert system. Whether this is possible in practice will be studied in the future with the aim of implementing a solution in ARES. At bit of topic, but yet of high relevance, is finally a comment on the hardware technology enabling AR. Today, there exist mainly three approaches of implementing AR from a hardware perspective: hand-held devices (e.g. tablets), head- worn devices (e.g. “smart” glasses) and spatial devices (e.g. projectors projecting information on real object surfaces). The first approach, hand-held, has the obvious disadvantage that is requires the operator use his/her hands to hold a device, which prevents to perform the work task. The second approach, head-worn device, does not have this problem as it frees the operator’s hands. Currently, there are two implementations of head-worn solutions on the commercial market; video-based and optical. An optical see- through solution has the advantage of leaving the view of the real-world almost intact, attempting to merge a virtual image into the view of the real-world. A disadvantage of the solution, however, is that image updating at head moves often lags behind which affect the user experience negatively. The visual lag is caused by, for example, communication or rendering delays and means that the virtual objects do not stay in the correct real-world position when the user moves. Even normal head movements require an extremely fast and frequent image updating in order to avoid visual lag, which is not possible to ensure even with the newest hardware technology. In comparison, a video see-through solution does not have the problem of lag as the real view and the virtual view is merged into the same view, but on the other hand it has the disadvantage of blocking out the real world. From a security perspective this is severe as the operator’s sight relies completely on hardware, which is dangerous in an industrial setting. The third approach, spatial devices, has the advantages that it does not directly affect the operator’s sight, nor does it require the operator to hold a device. However, it has a major disadvantage in that it requires hardware equipment to be permanently installed in the working environment, which is both expensive and inflexible. In summary, it can establish that for industrial applications, the AR technology still has a way to go. Solving the various problems with AR enabling hardware and to adjust it for 351 Anna Syberfeldt et al. / Procedia CIRP 41 ( 2016 ) 346 – 351 industrial settings is certainly a key factor for the success of technology in the future. References [1] Zauner J, Haller M, Brandl A and Hartman W. Authoring of a Mixed Reality Assembly Instructor for Hierarchical Structures. In: Proc. of International Symposium on Mixed and Augmented Reality, 2003, pp. 237–246. [2] Nilsson S and Johansson B. Fun and Usable: Augmented Reality Instructions in a Hospital Setting. In: Proc. of 19th Australasian Conference on Computer-Human Interaction, 2007, pp. 123-130. [3] Sääski J, Salonen T, Hakkarainen M, Siltanen S, Woodward C and Lempiäinen J. Integration of design and assembly using augmented reality. Micro Assembly Technologies and Applications 2008; 260: 395- 404. [4] Henderson S, and Feiner S. Exploring the Benefits of Augmented Reality Documentation for Maintenance and Repair. IEEE Transactions on Visualization and Computer Graphics 2011; 17(10): 1355-1368. [5] Tiefenbacher P, Lehment N, Rigoll G. Augmented Reality Evaluation: A Concept Utilizing Virtual. In: Proc. of Virtual, Augmented and Mixed Reality, LNCS 8525, 2014; 226–236. [6] Lee K. Augmented Reality in Education and Training. TechTrends 2012;2:13-21.. [7] Hajovy H, Christensen D. Intelligent Computer-Assisted Instruction: The Next Generation. In: Expert systems and intelligent computer-aided instruction. NJ Educational Technology Publications Englewood Cliffs. 1991; 191-196. [8] Pan X. An Experimental Adaptive Expert System. In: Proc. of InterSymp-2000. Baden-Baden, Germany, July 31 – August 4, 2000. [9] Wiedenmaier S, Oehme O, Schmidt L and Luczak H. Augmented Reality (AR) for Assembly Processes Design and Experimental Evaluation, International Journal of Human-Computer Interaction 2003; 16(3): 497- 514. [10] Radkowski R, Herremaa J and Olivera J. Augmented reality based manual assembly support with visual features for different degrees-of- difficulty. International Journal of Human-Computer Interaction 2015. [11] Caudel T, Mizell D. Augmented Reality: An Application of Heads-Up Display Technology to Manual Manufacturing Processes. In: Proc. of the Twenty-Fifth Hawaii International Conference on System Sciences (HICSS). Volume II, 1992, pp. 659-669. [12] Krevelen D, Poelman R. A Survey of Augmented Reality Technologies, Applications and Limitations. International Journal of Virtual Reality 2010; 9(2): 1-20. [13] Azuma R. A Survey of Augmented Reality. Malibu, Hughes Research Laboratories, 1997. [14] Azuma R, Baillot Y, Behringer R, Feiner S, Julier S, MacIntyre B. Recent Advances in Augmented Reality. Computer Graphics and Applications 2001; 21(6): 34-47 [15] Kipper G, Rampolla J. Augmented Reality: An Emerging Technologies Guide to AR. 1st Edition, Syngress, 2012. [16] Patel M, Virparia P, Patel D. Web based Fuzzy Expert System and its Applications – a Survery. International Journal of Applied Information Systems 2012; 11-15. [17] Mehrota K, Mohan C, Ranka S. Elements of artificial neural networks. The MIT Press, 1996. 
work_v235goyagzaddouqdgpxiwrmsi	----	doi:10.1016/j.eswa.2004.05.016 Constructing detection knowledge for DDoS intrusion tolerance Shun-Chieh Lin, Shian-Shyong Tseng* Department of Computer and Information Science, National Chiao Tung University, 1001 Ta Hsueh Road, Hsinchu 300, Taiwan, ROC Abstract Intrusion tolerance is the ability of a system to continue providing (possibly degraded but) adequate services after a penetration. With the rapid development of network technology, distributed denial of service (DDoS) attacks become one of the most important issues today. In this paper, we propose a DDoS ontology to provide a common terminology for describing the DDoS models consisting of the Profile model (the representation of the behaviors of system and users) and the Defense model (the descriptions of Detection and Filter methodologies). Also, the Evaluation strategy based upon current statuses of users’ behaviors is used to evaluate the degree of the intrusion tolerance of the proposed models during DDoS attacks. Based upon the ontology, four KCs (Profile model, Evaluation strategy, Detection methodology, and Filter methodology Knowledge Classes) and their relationships are then proposed, where each KC may contain a set of sub-KCs or knowledge represented as a natural rule format. For an arbitrarily given network environment, the default knowledge in the Profile KC and the Evaluation KC, the appropriate detection features in the Detection KC, and the suitable access control list policies in the Filter KC can be easily extracted and adopted by our proposed integrated knowledge acquisition framework. We are now implementing a NORM-based DDoS intrusion tolerance system for DDoS attacks to evaluate the proposed models. q 2004 Elsevier Ltd. All rights reserved. Keywords: Distributed denial of service (DDoS); Intrusion tolerance; Ontology; Knowledge acquisition; NORM 1. Introduction Intrusion tolerance is the ability of a system to continue providing (possibly degraded but) adequate services after a penetration (Stavridou, 2001). With the rapid development of network technology, the network security becomes one of the most important issues today, especially for distributed denial of service (DDoS) attacks. A DDoS attack is a simple but ferocious attack since it could be launched by someone who has DDoS attacking tools. In recent years, many e-commerce enterprises have been attacked by DDoS attacks, such as Yahoo, eBay, Amazon, Key Internet Computers and so on, which cause great damage on these enterprises (Bridis, 2002; CERT/CC, 2003; Xu & Lee, 2003). Various approaches and monitoring policies have then been developed to detect and filter the malicious traffic of DDoS attacks. Unfortunately, it is very difficult to keep up with the rapidly growing expertise or knowledge on their domains due to lack of the common terminology. Therefore, an ontology, which could be acquired from domain experts, is needed for providing a sharing vocabulary to integrate and categorize the increasing growth of knowledge of DDoS intrusion tolerance. Due to the difficulty of handling the complicated DDoS characteristics, maintaining the huge amount of knowledge for detecting and preventing DDoS attacks becomes more and more important. Accordingly, we will propose a knowledge base to store the characteristics of DDoS attacks, which may be obtained by analyzing the traffic behaviors of the previously discovered DDoS attacking tools, to help mitigate the malicious traffic caused by the DDoS attacks. As we know, the knowledge of DDoS intrusion tolerance could be divided into several knowledge objects including various kinds of DDoS attacking tools/methods and defending heuristics. The relationships between attacking characteristics of DDoS attacking tools have not been clearly defined and the defending heuristics may be too complicated to maintain; moreover, the different character- istics and the relationships of DDoS attacks in DDoS intrusion tolerance may incrementally increase. A knowl- edge based system (KBS) with object-oriented model and rich relationships is required to achieve the modularity, sharability, and reusability of the dramatically increasing DDoS characteristics. Therefore, a DDoS ontology according to the experience of the observations for DDoS intrusion tolerance will be 0957-4174/$ - see front matter q 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2004.05.016 Expert Systems with Applications 27 (2004) 379–390 www.elsevier.com/locate/eswa * Corresponding author. Tel.: þ886-3-5712121x56658; fax: þ886-3- 5721490. E-mail address: sstseng@cis.nctu.edu.tw (S.-S. Tseng). http://www.elsevier.com/locate/eswa proposed to describe the DDoS models and the Evaluation strategy to evaluate the models. Each knowledge class (KC) could be represented as a node of the ontology to achieve the modularity of DDoS knowledge containing several sub-KCs and the relationships between KCs could be represented as edges. The DDoS models consist of Profile KC (including system state and role state diagrams to model the behaviors of system and users, respectively) and Defense models (including Detection and Filter KCs to collect the knowledge for defending DDoS). To evaluate the tolerance degree during DDoS attacks, the Evaluation KC consisting of the system capacity and network bandwidth utilization criteria could be acquired from domain experts. In order to obtain the more complicated accumulating expertise and speed up the collection of expertise for DDoS intrusion tolerance, an integrated knowledge acquisition (KA) frame- work including interviewing, training process, and learning process will be proposed based upon the ontology. The Profile KC could be easily acquired from domain experts by interviewing with domain experts. Traditionally, NORMAL and DEAD states are usually enough to describe the system status because the victim would be crashed simultaneously when the DDoS attacks. By applying the various filtering polices including white list and black list for DDoS attacks according to the role state diagram, the survival time of the system can then be extended and a SURVIVAL state is further added to represent such situation. Since the only way we can stop a DDoS attack once it starts is to identify the addresses of all agents (zombies) sending DDoS packets and shut off traffic from them (Garber, 2000), the system will tend to move SURVIVAL state quickly backed to NORMAL state with the heuristic of setting more restricted filtering policies in the Filter methodology KC. On the other hand, the behaviors of the DDoS intrusion tolerance system will be expected to be the NORMAL-SURVIVAL-NORMAL pattern instead of the traditional NORMAL-DEAD pattern. To determine the new useful filter policy for mitigating the damage causing by DDoS attacks, the role state diagram including CANDIDATE, TRUSTED, SUSPECTED, and UNTRUSTED states will also be proposed to represent the behaviors of each user. The Detection methodology KC, which is responsible for detecting and predicting the occurrence of the DDoS attacks, including attack predicting, attack detecting, and attack tracebacking rules could be obtained by analyzing the behaviors of DDoS attacking tools. Besides, a NORM-based intrusion tolerance system will be implemented to evaluate the proposed models. 2. Related work As mentioned above, many different DDoS attacking tools and defending methods to help mitigate the malicious traffic developed result in the rapid growth of complicated characteristics of DDoS intrusion tolerance in recent years. 2.1. Basic of DDoS The DDoS appeared in June of 1998 firstly means that many attackers launch malicious traffic to the same victims together and make the victims too busy to respond all the traffic including legitimate requests. Several different DDoS attacks which were developed (Barlow & Thrower, 2001; Cabreraa et al., 2001; Dittrich, 2000) from 1998 to 2002 can be divided into several categories including UDP flood, ICMP flood, TCP flood, and Smurf as shown in Table 1, so the common characteristics of each category of DDoS attacks could be easily observed and extracted. For example, the TFN2K, discovered in November 1999, is a kind of DDoS attacking tools. It can launch UDP flood, ICMP flood, Smurf, or TCP-SYN flood attacking method to attack victims in one time. All of the observed knowledge can be represented as a natural rule format and stored into a knowledge base. The general topology of DDoS attack shown in Fig. 1 could be divided into control stage and attack stage. In control stage, a scan is performed in large ranges of network to find the list of vulnerable hosts. Generally speaking, the vulnerable hosts consist of handlers and agents, where the handlers (the first level vulnerable hosts) are controlled Table 1 The DDoS attacks developed from 1998 to 2002 Date Attack Attack method UDP flood ICMP flood Smurf TCP SYN TCP ACK TCP RST TCP SYN/ACK 1998/5 FAPI W W W 1999/6 – 1999/7 Trin00 W 1999/8 – 1999/9 TFN W W W W W 1999/9 – 1999/10 Stacheldracht W W W W 1999/11 TFN 2K W W W W 2000/1 – 2000/2 Shaft W W W 2000/2 Trinity W W W W 2002/1 DRDoS W 2002/4 Mstream W S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390380 by attackers and agents (the second level vulnerable hosts) are also controlled by attackers through handlers. Most of the controlling traffic, the traffic of communication in control stage, is signal direction between attacker and handler but is bi-direction between handlers and agents. The two level topology results in the locations of attackers can be hidden. After the control stage, the list of vulnerable hosts is then used to launch the distributed attacking traffic in attack stage. The attacking traffic including UDP flood, ICMP flood, Smurf, TCP SYN, TCP ACK, TCP RST, and TCP SYN/ACK as shown in Table 1 could overwhelm the victim. As we know, there are two different types of attack technique in DDoS attacks: bandwidth consumption and resource consumption. The bandwidth consumption means that the attacking traffic launched by the compromised hosts, which are controlled by attackers, is aggregated to a single huge flood and overwhelms the victim. The resource consumption means that attackers make use of the leak of the network protocol or the system security such as the techniques of SYN flood, land and Teardrop, resulting in the starvation of system resources (CERT/CC, 2003). As the DDoS attack tools have become more compli- cated in recent years, the maintenance of the characteristics of DDoS attacks is becoming more difficult despite the previously known common characteristics of each category of discovered DDoS attacks. Therefore, we will propose a knowledge base to store the characteristics of DDoS attacks, which may be obtained by analyzing the traffic behaviors of the DDoS attacking tools, for DDoS intrusion tolerance. Besides, two criteria considering the difference between two types of DDoS attacks will be proposed to evaluate the degree of intrusion tolerance. 2.2. Intrusion tolerance of DDoS Intrusion tolerance is the ability of a system to continue providing (possibly degraded but) adequate services after a penetration (Stavridou, 2001) As mentioned above, it is very hard to detect and prevent the DDoS attacks. Therefore, the intrusion tolerance of DDoS attacks is an important issue to mitigate the damage during DDoS attacks for providing the critical services continuously on Internet. Park and Lee (2001) suggested a method to install packet filters at different autonomous system on the Internet to filter out attacking traffic traveling between them. The method is very effective but not practical to defend DDoS attacks because of requiring the cooperation of thousands of autonomous systems on Internet. Chang (2002) also introduced the concept of Internet firewall and four detecting and filtering approaches consist- ing of ubiquitous ingress packet filtering, router-based packet filtering, local attack detecting, and distributed attack detecting for defending flooding-based DDoS attacks. He also indicated that more effective detect-and-filter approaches, such as distributed attack detecting, should be developed for DDoS intrusion tolerance. However, all of them lack a systematic approach to integrate the knowledge of DDoS intrusion tolerance. Xu and Lee (2003) proposed a DDoS intrusion tolerance system to sustain the availability of web service under DDoS attacks. The main idea is to isolate and protect legitimate traffic from huge volumes of DDoS traffic when an attack occurs. Unfortunately, it only aims on protecting web service instead of protecting the network. Although a variety of methods have been proposed to mitigate the damage during DDoS attacks for providing the critical services continuously, it is still very difficult to keep up with the rapid growth of DDoS expertise in their studies. To solve this problem, a DDoS ontology is proposed to provide a common vocabulary among domain experts and an integrated KA framework is then proposed to assist in quickly accumulating their expertise. We will also use the behaviors of access control list (ACL) to evaluate the performance of the DDoS models. 2.3. Knowledge based system and NORM Although expert systems simulating the decision-making ability of a human expert have been widely applied in many domains, only little research focuses on the DDoS intrusion tolerance. As we know, the knowledge of DDoS intrusion tolerance could be divided into several knowledge objects including various kinds of DDoS attacking tools and defending heuristics. The relationships between attacking characteristics of DDoS attacking tools and the defending heuristics may be too complicated to maintain; moreover, the different characteristics and the relationships of DDoS attacks in DDoS intrusion tolerance may be incrementally increased. A KBS with object-oriented model and rich relationships is required to achieve the modularity, sharability, and reusability of the rapidly increasing DDoS characteristics. In Lin et al. (2003), a new object-oriented rule model (NORM) was proposed based on the concept of learning and thinking behaviors of human to provide high maintainability Fig. 1. The general topology of DDoS attacks. S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390 381 and reusability through object-oriented concept. There are four basic relations between knowledge concepts defined in NORM: Reference, Extension-of, Trigger and Acquire. The Reference relation represents the association of two different KCs if the KCs have common piece of knowledge, which is useful for using original knowledge to construct new knowledge. Extension-of relation is used to extend or modify the KC constructed by other people, which is useful for knowledge sharing and exchanging. The Trigger and Acquire relations are used to represent the interaction of different KCs. Drama, a NORM knowledge modeled rule base system platform implemented using pure Java language, includes Drama Server, Console, Knowledge Extractor, and Rule Editor. Also, it provides Application Programming Interface (API) to access Drama server in Drama integrated systems. In our prototype system, the relations of NORM are used to represent the interaction of different KCs of DDoS intrusion tolerance. 3. Ontology and model of DDoS As we know, the traditional methods for detecting and filtering DDoS attacks are monitoring the status of network and system, specifying the alert thresholds, defining detection rules, and setting filter policies by domain experts. Based upon interviewing with domain experts, the DDoS ontology proposed to models the behaviors of system and users, the methodologies of defense, and the strategies of evaluation are described as follows. 3.1. Ontology of DDoS Before understanding the more complicated DDoS knowledge, an ontology, which is needed for sharing knowledge with a common terminology among numerous experts of the DDoS domain, could be divided into three parts: Profile model, Defense model, and Evaluation strategy as shown in Fig. 2. The Profile model is proposed to describe the behaviors of system and users according to the state and user state diagrams. The Defense model consisting of Detection and Filter methodologies is then used to resist the DDoS attacking. Finally, the Evaluation strategy including system and network evaluations is proposed to evaluate the performance of our proposed DDoS model. Hence, the ontology includes Profile model, Detection methodology, Filter methodology and Evaluation strategy knowledge classes (KCs), each may include several sub-KCs and may be obtained by interviewing with domain experts, e.g. Attack predicting and Attack detecting are sub- KCs of the Detection KC. The Profile model KC (herein the Profile KC) and Evaluation strategy KC (herein the Evaluation KC) men- tioned above could be easily acquired from domain experts. The Detection methodology KC (herein the Detection KC) including attack preventing, attack detecting, and attack tracebacking technologies to defend DDoS attacks describes some useful features which could be selected by analyzing the characteristics/behaviors of attacking tools and the profiles defined in Profile KC. The predicting features including default communication behaviors, encryption behaviors, and encoding behaviors to predict the occurrence of DDoS attacks could be obtained by analyzing the behaviors of controlling traffic in DDoS attack tools. As for detecting DDoS, the behaviors of DDoS attacking traffic become most useful recently; e.g. three kinds of detecting features including TCP, UDP and ICMP are often used to detect the occurrence of DDoS attacks. Due to the massive of forge IP addresses and port numbers used by the spoofing techniques, both users and port numbers are divided into Trusted and Untrusted. Hence, the behaviors of Untrusted source IP could be used to detect the DDoS attacks. Similarly, the randomly generated port numbers are usually used in most DDoS attacking tools; Fig. 2. The ontology of DDoS. S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390382 the behaviors of Untrusted destination port numbers could also be used to detect the DDoS attacks. In order to systematically select the suitable features for DDoS intrusion tolerance, a Characteristics Trainer is then proposed to obtain the significant characteristics, e.g. the specific attacking behaviors and the communication beha- viors of DDoS attacks, which may be found by comparing the attacking traffic with the normal traffic according to the expertise. The attack tracebacking technology is still evolving and out of the scope of this paper. The Filter methodology KC (herein the Filter KC) including black list and white list policies provides an efficient way to mitigate the malicious traffic during the DDoS attack. Various filtering policies, which are respon- sible for filtering out the malicious traffic, could be chosen in Filter KC according to the current system or network environments. To help domain experts adaptively tune the policies for filtering out the malicious traffic according to the response from other KCs, a recommendation of black list and white list policies can be provided by a User’s Behavior Learner which is proposed using the role state diagram in Profile KC. 3.2. The relationships between knowledge classes Four basic relations between KCs have been defined in Drama/NORM (Lin et al., 2003): Acquire, Trigger, Reference, and Extension-of relation. These relations are helpful in describing the relationships among KCs. Trigger relation triggers another KC with current facts as knowledge transfer. In other words, the remnant knowledge in original KC should not be necessarily considered. Acquire rep- resents the acquirement relation. After Acquire process, the original inference process will continue and only facts predefined in the acquired KC will be carried back in Acquire relationship. Reference is used to represent the associations between different KCs. Through the Reference relation, the knowledge contained in referred KC is regarded as the base knowledge and will be taken into consideration together with the knowledge defined in the KC. On the other hand, Reference can be thought as an unconditional Acquire relation between KCs. Unlike the Reference relation, the Extension-of relation makes a new KC to include all the knowledge contents of an existent KC. The activities of Extension-of relation include extension and modification. Therefore, it must support the overriding mechanism, including the overriding of facts and rules. As shown in Fig. 3, the Filter KC referencing the Profile KC can be treated as a filter to filter out the malicious traffic and can be triggered by the outside traffic events. Also, it can set the new filter policy and acquire the Evaluation KC to evaluate the system performance. The Detection KC triggered by the Filter KC could be treated as a detector to detect the occurrence of DDoS attacks and could trigger the Filter KC according to the specific detection events for dynamically setting the new filtering policy to filter the malicious packets. Also, the Detection KC can acquire the Profile KC to set the suitable attack detecting or attack predicting sub-KCs, which are included by the Extension-of relation. 3.3. Profile model According to the expert’s experiences of defending DDoS attacks, the Profile model including system state and role state diagrams shown in Figs. 4 and 5, respectively, are proposed to represent the behaviors of system and users through interviewing with domain experts. 3.3.1. System state diagram In the traditional systems, NORMAL and DEAD states are usually enough to describe the system status, because the time is too short to respond the DDoS attacks for the administrator before the victim system moving to DEAD state. Therefore, the behaviors of the most traditional information systems will be NORMAL-DEAD pattern when they are attacked. By applying the various filtering polices for DDoS attacks according to the users’ behaviors, the survival time of the system can then be extended and the SURVIVAL state is further added to represent such situation. With the heuristic of setting more restricted filtering policies, the system will tend to move SURVIVAL Fig. 3. Relationships between of knowledge classes. Fig. 4. System state diagram. Fig. 5. Role state diagram. S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390 383 state quickly backed to NORMAL state. On the other hand, the behaviors of the DDoS intrusion tolerance system will be expected to be NORMAL-SURVIVAL-NORMAL pattern during DDoS attacks. To describe the relationships between states of the system conceptually, the system state diagram including NORMAL, SURVIVAL, and DEAD states are proposed as shown in Fig. 4. A system capability and the bandwidth utilization could be computed by the CPU load, the memory usage and the behaviors of ACL to determine the transition of the system state diagram according to the new policy setting event, the attacking traffic coming event, and the system overloading event. 3.3.2. Role state diagram We assume the behaviors of normal users may not be changed dramatically in a short period of time. To represent the behaviors of each user, a role state diagram based upon an efficient ACL including white list and black list policy is proposed. We assume the DDoS attackers may frequently request service during the abnormal network status instead of normal ones. To monitor users’ behaviors, the historical behaviors of the users based upon the current network status in a short time slice are proposed to distinguish the abnormal users from normal users efficiently, where each historical behavior is represented as a sequence of current network status. According to the white list and black list shown in Fig. 5, all users could be categorized into TRUSTED, UNTRUSTED and CANDIDATE states. The N and A indicate that the current network status is normal and abnormal, respectively, and the user in CANDIDATE or SUSPECTED state is moved to CANDIDATE state when current network status is N; otherwise, he/she is moved to SUSPECTED state. Then, the MðqÞ and MðqÞ0 indicate that the historical behavior of the user q; which is acquired from Users’ Log DB, is normal and abnormal, respectively. The user in CANDIDATE state will be moved into TRUSTED state if he/she has accumulated sufficient normal behaviors; it means that the historical behavior of the user q should be normal, MðqÞ: On the other hand, the user in SUSPECTED state will be switched to UNTRUSTED state if he/she has accumulated sufficient abnormal behaviors, MðqÞ0: The users in TRUSTED and UNTRUSTED states will be moved to CANDIDATE and SUSPECTED states, respect- ively, if their historical behaviors match the corresponding constraints. With the role state diagram, the filtering policies could be adaptively generated based upon dynamically changing network environment. 3.4. Evaluation strategy As mentioned above, there are two different types of DDoS attacking technique: bandwidth consumption and resource consumption. Since the only way we can stop a DDoS attack once it starts is to identify the addresses of all agents (zombies) sending DDoS packets and to shut off traffic from them, the behaviors of black list and white list are considered to monitor the potential latency of the network and the CPU usage and memory usage are used to monitor the degrading rate of system performance when suffering the DDoS attack and the Tolerance of the initial ACL is assumed to be maximal in this paper. During resource consumption DDoS attacks, the system capacity of the victim is always decreasing (i.e. the system resource usage r is increasing), causing Tolerance low. On the other hand, the bandwidth consumption DDoS attacks may increase the members of the black list and decrease the members of the white list due to filtering the malicious traffic; hence the Tolerance will become small if the filter policy could not be performed well. Thus, the following formula combining network tolerance (Tolerancenetwork) and system tolerance (Tolerancesystem) is given to represent the degree of DDoS intrusion tolerance, where a and b are used to indicate the weights of the tolerance. If we focus on protecting network performance, a is set to be larger than b: Otherwise, a large b is recommended to protect the system performance. Tolerance ¼ a £ Tolerancenetwork þ b £ Tolerancesystem Since the members in the white list (WL) may represent they could be served and the members in the black list (BL) may represent they could be blocked by the victim, the ratio of the WL and the BL is used to evaluate the network bandwidth utilization (Bw). A large BL implies the filtering policies are set to be more restricted and the tolerance may become small; otherwise, a large WL may represent the system is with more tolerance since more users can access the services of the victim. The number of users who have been moved from white list to black list (W2B) is further regarded as a penalty weight for network tolerance. The formula of the Tolerancesystem and the Tolerancenetwork are shown as follows: Tolerancesystem ¼ 1 2 r Tolerancenetwork ¼ WL BL þ W2B ð1 2 BwÞ The system state will be set as NORMAL when the Tolerance is larger than the predefined thresholds. Other- wise, the system state will be in SURVIVAL state and the new filter policy based upon Evaluation knowledge will be generated to move the state to NORMAL. As mentioned above, each KC may include several sub- KCs due to the hierarchy of the knowledge in DDoS intrusion tolerance. In order to obtain the knowledge of each KC, an integrated KA framework including interviewing with domain experts, training the predicting and detecting features, and learning the filter policies for adaptively filtering malicious traffic is proposed. All knowledge of DDoS intrusion tolerance can be represented as a natural rule format, IF Conditions Then Conclusions. The bodies of S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390384 Conditions are the facts generated from network traffic flow, detecting results, and filtering policy. The Conclusions may include the alarming events, system state changing events, and user state changing events. Furthermore, constructing the knowledge base can facilitate the maintenance of the knowledge for defending DDoS and can help the adminis- trators manage their networks. 4. Knowledge base construction Due to the difficulty of acquiring and collecting the various DDoS characteristics from domain experts, an integrated KA framework including three related KA methods is used for reducing the effort of accumulating the expertise and speeding up the knowledge collection of DDoS characteristics. 4.1. The integrated knowledge acquisition (KA) framework The problems that we are faced with during the KA process are usually very hard. In general, KA involves: (1) elicitation (gathering) of data from the expert, (2) interpretation of the data to infer the underlying knowledge or reasoning procedure and (3) guided by this interpretation, creation of a model of the expert’s domain knowledge and performance. Although quite many different kinds of KA approaches have been proposed in many research studies (Hirasawa & Chu, 2003; Nikiforou & Fink, 2002; Rafea & Hassen, 2003; Tsai & Liu, 1999; Webb & Wells, 1999; Yan & Jiang, 2002) including interviewing with experts, Repertory Grids, and machine learning, few studies have focused on integrating various kinds KA approaches for an application due to the lack of a common vocabulary. Based upon the DDoS ontology we mentioned above, an integrated KA framework is proposed in this paper. Through the KA framework shown in Fig. 6, four kinds of KCs including Profile KC, Evaluation KC, Detection KC, and Filter KC could be easily obtained by various KA processes. Other new discovered or defined KCs can also be easily added or modified in our knowledge base due to the nature of KBS. As shown in Fig. 6, all of these KCs of expertise can be obtained in the integrated KA framework, which includes modeling the DDoS environment through interviewing with domain experts, selecting useful features by analyzing the attacking tools in the Characteristic Trainer, and adaptively learning filter policies in the User’s Behavior Learner. The behaviors of users, communication signatures, and other useful features are the characteristics of the Detection KC, so called detecting rules, which can be obtained by the Characteristic Trainer. On the other hand, the User’s Behavior Learner is responsible for generating the various filtering policies in Filter KC. All other KCs including Profile KC and Evaluation KC can be directly obtained by interviewing with experts. 4.2. The knowledge obtained by interviewing with domain experts As interviewing is one of the traditional approaches to acquire the expertise from domain experts by knowledge engineers, many approaches have been proposed to acquire expertise from experts through interviewing. As mentioned above, the Profile KC and the Evaluation KC could be modeled by domain experts using interviewing approach. Besides, the default knowledge including default communication ports and the white list and black list policy in the Detection KC and the Filter KC could be also acquired by interviewing with domain experts. Because the characteristics of DDoS attacking tools and the filtering knowledge are dramatically increasing, the Characteristic Trainer and the User’s Behavior Learner are proposed to obtain the useful knowledge for DDoS intrusion tolerance. 4.3. Training the detection KC by Characteristic Trainer To obtain the previously undiscovered DDoS character- istics/behaviors, a training process, namely Characteristic Fig. 6. The framework of KA process. S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390 385 Trainer, is proposed to learn the useful, new features and store them into the Detection KC for predicting and detecting DDoS attacks. 4.3.1. Training process of Characteristic Trainer Fig. 7 shows the new features of DDoS attacks can be selected by comparing the normal behaviors during the NORMAL system state with the attacking behaviors launched by DDoS attacking tools in the systematic training process of Characteristic Trainer. Thus, to distinguish attacking behaviors from normal behaviors, each kind of characteristics represented from the DDoS behaviors could be easily identified using a Repertory Grids approach, which is a table with four attributes including the feature, the feature threshold d; the feature operation u; and the corresponding actions. The d is a parameter adaptively determined in a short period by different features. For example, if one feature value is larger than d; it needs to be considered as abnormal behaviors. Therefore, the small d will increase more false alarms. On the contrary, larger d will treat more attacking behaviors as normal. After a DDoS attacking feature is detected, the corresponding action, e.g. trigger the Filter KC, alarm the attacking traffic coming, or specify the attacking type, must be taken. In addition to the above previously known features, new characteristics/features may be observed by using the Repertory Grids approach after analyzing fingerprints of DDoS attacks such as spoofing, flooding-based and communication techniques and more attacks could then be detected and predicted. The Characteristic Trainer algorithm Input: Training cases TC; Actions A; and feature set F with n Features Output: The Detection KC Step 1: Select a skeleton feature set Fk ¼ {f1; f2; …; fk } # F by interviewing with domain experts. Step 2: Set di; for each feature fi in F in each TC: Step 3: Choose the proper actions Ai for the feature fi selected in Step 2. Step 4: Generate the detection rule as ‘IF fi uI di THEN Ai’. Step 5: Repeat Steps 2 – 4 until all detection rules have been generated. 4.3.2. Examples of detailed rules obtained by the Characteristic Trainer According to the Characteristic Trainer algorithm, the following six examples show the obtained partial character- istics/features and the corresponding detection rules † Example 1: the ratios of untrusted IP addresses and ports. Since the attacking traffic of DDoS is launched from the compromised hosts which disperse on Internet, most of the attacking traffic comes from the untrusted IP addresses. Also, some DDoS attacking tools may generate the destination port of attack packets randomly. There- fore, we select the ratios of traffic from untrusted IP addresses and ports as two important characteristics in Step 1 to detect the occurrence of DDoS attacking traffic, set dangerous ratio d ¼ 50 by comparing the list of ports and ACL defined by ISPs in Step 2, and the action ‘Trigger Filter KC’ for these two detecting features is chosen in Step 3. Finally, the rule ‘IF (the ratio of untrusted IP addresses . 50) OR (the ratio to untrusted ports . 50) Then (Spoofing DDoS attacking) AND (Trigger Filter KC)’ is generated to discover DDoS attacks. † Example 2: number of flows from the same source increase rapidly. Generally speaking, the number of flows from one source normally does not increase dramatically if the attacks without spoofing did not occur. Hence, we select the number of flows from one specific source as a characteristic of suspected user to detect the DDoS attacks in Step 1. Next, the rapidly increasing rate d ¼ 50 in Step 2. In Step 4, the rule ‘IF (Number of flows from the same source . 50) Then (Same source DDoS attacking) AND (Trigger Filter KC)’ is generated to discover DDoS attacks. † Example 3: number of flows in the state of ‘SYN_RE- CEIVED’. Resource consumption attacks often use the SYN flood technique (Criscuolo, 2000), such as TFN, TFN2K, stacheldracht and so on, to overwhelm the victim, where client sends a fake packet to Server, and Server accepts this SYN packet, responds the SYN/ACK packet to the fake address, and waits for the response of SYN/ACK packet. But the response will never arrive due to the faked source address. Therefore, the number of flows in the state of SYN_RECEIVED is first selected to detect the DDoS attacks in Step 1 and the d is set to 1000 in Step 2 in this example. Finally, the rule such as ‘IF (#SYN_RECIEIVED . 1000) Then (SYN-flooding DDoS attacking) AND (Trigger Filter KC)’ is generated to discover DDoS attacks. † Example 4: percentages of UDP and ICMP packets. The percentages of UDP and ICMP packets usually repre- senting the error control messages during communi- cations are always small in normal traffic and they will become large if the bandwidth consumption DDoS attack, the UDP or ICMP flood, could be launched to overwhelm victims. Since the percentages of UDP and ICMP packets are various for different autonomous Fig. 7. The training process of Characteristic Trainer. S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390386 systems, the environment-dependent characteristics on monitoring the huge traffic of UDP and ICMP packets are required and selected to detect the DDoS attacks in Step 1. The d is then set to be 30 in Step 2. Finally, the corresponding detection rules such as ‘IF (P(UDP) . 30) OR (P(ICMP) . 30) Then (Flooding-based DDoS attacking) AND (Trigger Filter KC)’ are generated to discover DDoS attacks. † Example 5: oversize of UDP packets and ICMP packets. As mentioned above, the UDP and ICMP packets are usually the error control messages during communication, the encryption of payload in packet for DDoS tools is usually used for communication between attackers, handlers, and agents in controlling traffic, since the controlling traffic stores the source addresses of attackers, handlers and victims and attacks do not want to be revealed. Besides, the accounts, the passwords, and the commands to control handlers and agents from attackers are also necessarily encrypted. Since encrypting the information may enlarge the size of UDP and ICMP packets, the ratio of oversize UDP and ICMP packets might be selected as good features in Step 1 and set threshold value ¼ 100. Finally, the corresponding predic- tion rules such as ‘IF (#Oversize . 100) Then (Control- ling traffic attacking) AND (Trigger DDoS Prediction KC)’ are generated to predict the DDoS attacks. † Example 6: percentage of BASE64 encoding packets. Once the communication packets are encrypted, the resulting special characters should be encoded to avoid the occurrence of errors in communication during DDoS attacks. BASE64 encoding is very popular in solving this problem, but the payload of packet contains only alphanumeric character and 0 £ 41(‘A’) always appears at the end of the BASE64 encoding packet in the TFN2K attack. Both may be selected as useful features to predict the occurrence of DDoS attacks in Step 1. Finally, the predicting rule ‘IF (AlphaBeta ¼ Yes) AND (Tailing ¼ ‘A’) Then (TFK2K Communication traffic attacking) AND (Trigger Filter KC)’ is generated. The following table shows the summarization of Examples 1 – 6 4.4. Learning the filter knowledge class by User’s Behavior Learner In traditional network management system such as firewall, intrusion detection system, network management system, etc. ACL is widely used not only to filter suspected connection from untrusted sources but also to admit the access from trusted sources black list and white list strategies used in ACL are included in the Filter KC. The former is used to interdict the access right, but the latter is used to permit the access right. Moreover, the various filtering policies can be set according to the configurations of current system and network environments. In order to dynamically construct suitable ACL to mitigate the damage of DDoS attack, a learning process is also proposed to generate appropriate filtering policies for various network environments. The black list is used to drop the malicious attacking traffic and the white list is used to allow the trusted IP to access the critical service of the victim server. The principles of the User’s Behavior Learner for ACL are twofold in this paper. One is to keep the legal users, whose behaviors are determined as normal, in the ACL. The other is to remove the possible suspected users or the users, who do not request critical service for a long time, from the ACL. Since the normal user may not change his/her behavior rapidly, a user behavior scoring function is designed to calculate the score of each user by his/her own behavior to determine his/her status of historical behavior. The basic idea of the scoring function is to incrementally adjust the weight. If a current network status is normal, the score becomes larger; otherwise, it becomes smaller. The initial score value of each user is given 0 and the behavior of user changes from A to N or from N to A; the score is no change. To evaluate the historical network status of the user q; G is defined by expertise to determine the user’s historical behavior network status. The historical behavior of the user ranging from [MAX, MIN ] is determined as normal (MðqÞ) if the score is larger than G: Otherwise, it is judged as abnormal (MðqÞ0) when the value is smaller than 2G: Feature name (f ) Operator (u) Threshold (d) Actions (A) Ratio of untrusted IP . 50 (Spoofing DdoS attacking) AND (Trigger Filter KC) Ratio of untrusted ports . 50 (Spoofing DDoS attacking) AND (Trigger Filter KC) Ratio of same IP . 50 (Same IP DDoS attacking) AND (Trigger Filter KC) #SYN_RECIEIVED . 1000 (SYN flooding attacking) AND (Trigger Filter KC) UDP percentage . 30 (Flooding-based attacking) AND (Trigger Filter KC) #Oversize packets . 100 (Controlling traffic) AND (Trigger Prediction KC) AlphaBeta ¼ Yes (Communication traffic) AND (Trigger Filter KC) Packet tailing ¼ ‘A’ (Communication traffic) AND (Trigger Filter KC) S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390 387 In Fig. 8, an example of P ¼ kN; N; N; N; A; A; A; N; N; A; N; Nl and Q ¼ kA; N; A; A; A; A; N; A; N; A; A; Nl is given to explain our proposed scoring function. The final score of P and Q are 4 and 25, respectively. If G ¼ 3; the behavior of P is M and the behavior of Q is M0 in this example. To more accurately categorize the user behavior, a penalty weight (PW) could be attached when the character- istic of DDoS attack is discovered from individual user. The range of PW could be also defined according to the degree of dangerous behavior. 5. The verification of selected features/characteristics of DDoS To evaluate the selected features/characteristics of DDoS attacks, the selected characteristics of DDoS which are similar to the normal behaviors will be eliminated due to the reduction of the false detection rate in the DDoS intrusion tolerance system. And then the Drama-based DDoS intrusion tolerance system will be implemented to evaluate the performance of detection power by the selected characteristics. 5.1. NORM-based DDoS intrusion tolerance system using KCs In order to evaluate the efficiency of the KCs, the DDoS intrusion tolerance system using KCs is implemented by an inference engine Drama for detecting and predicting the occurrence of DDoS attacks. As shown in Fig. 9, the four KCs could be easily used to infer the other system components through the inference engine, and each component can be easily replaced according to different configurations of the network environment. The network traffic can be character- ized as feature vectors (Lin et al., 2002) by Feature Vector Constructor. The Facts Collection is used to store all facts including the feature vectors, detecting results, system status, user states, and system evaluation results. When the anomaly network traffic is detected in the Detection KC, the event of attacking traffic coming is triggered and the Profile KC is acquired to change the state of system. And the alarm event is thus triggered to set the suitable ACL in Filter KC for dropping the huge traffic from the attackers and allow the legitimate traffic from trusted users. Since then, the event of updating filter rules would be triggered to generate a new filter policy for dropping the malicious traffic. It implies the Evaluation KC will be acquired and used to compute the tolerance of the system for updating the filtering policy of the Filter KC. Finally, the Profile KC would be triggered to indicate the proper system state. However, the complex attacking behaviors sometimes make the filtering policy fail. When it is failing, the event of generating new filter policy would be triggered again until system state in Profile KC is stable. Otherwise, experts are asked to solve the problem of DDoS attacks. 5.2. An example of knowledge classes The following shows a simple example of KC to defend the TFN2K attack, where the Profile KC includes system state transition rules and user state transition rules, the Detection KC includes attack detecting and attack predict- ing sub-KCs, the Detection KC includes rules focusing on detecting TFN2K attack, the Filter KC lists the heuristics of policy learning proposed in this paper, and the Evaluation KC illustrates the evaluating rules including system performance evaluating rules and the network performance evaluating rules. A partial Profile KC System state transition rules IF Ss ¼ NORMAL AND Traffic ¼ Normal Then Ss ¼ NORMAL //Ss is system state IF Ss ¼ NORMAL AND Traffic ¼ Attack Then Ss ¼ SURVIVAL IF Ss ¼ SURVIVAL AND Policy_Set ¼ Enable Then Ss ¼ NORMAL IF Ss ¼ SURVIVAL AND System ¼ Overload Then Ss ¼ DEAD User state transition rules IF Us ¼ TRUSTED AND Uhb ¼ M0 Then Us ¼ CANDIDATE //Us is user state IF Us ¼ CANDIDATE AND Uhb ¼ M Then Us ¼ TRUSTED //Uhb is historical user behavior IF Us ¼ CANDIDATE AND Ncb ¼ N Then Us ¼ CANDIDATE //Ncb is current user behavior Fig. 8. An example of users’ behaviors. Fig. 9. The NORM-based DDoS intrusion tolerance system. S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390388 IF Us ¼ CANDIDATE AND Ncb ¼ A Then Us ¼ SUSPECTED IF Us ¼ SUSPECTED AND Ncb ¼ N Then Us ¼ CANDIDATE IF Us ¼ SUSPECTED AND Ncb ¼ A Then Us ¼ SUSPECTED IF Us ¼ SUSPECTED AND Uhb ¼ M0 Then Us ¼ UNTRUSTED IF Us ¼ UNTRUSTED AND Uhb ¼ M Then Us ¼ SUSPECTED A partial Detection KC Attack detecting sub-KC IF P(Protocol) . dProtocol Then Flooding-based DDoS attacking AND Trigger Filter KC //P(Protocol) is the percentage of protocol IF IR(U) . 50% Then Same source DDoS attacking AND Trigger Filter KC //IR(U) is the increase rate of the user IF (Untrusted . 50%) AND (#Request . 200) Then Spoofing DDoS attacking AND Trigger Filter KC //UNTRUSTED is the rate of users in Untrusted IF #SYN_RECEIVED . dSYN Then SYN-flooding DDoS attacking AND Trigger Filter KC Attack predicting sub-KC IF AlphaBeta ¼ Yes AND Tailing ¼ ‘A’ Then TFK2K Communication traffic attacking AND Trigger Filter KC //BASE64 encoding IF #OverSize . 10 Then Controlling traffic attacking AND Trigger DDoS Prediction KC //Oversize packet A partial Filter KC Acquire Evaluation KC IF Ncb ¼ N AND Us ¼ CANDIDATE Then S(U) ¼ S(U) þ 1 //S(U) is the historical behavior score of user IF Ncb ¼ A AND Us ¼ SUSPECTED Then S(U) ¼ S(U) 2 1 IF Ncb ¼ N AND NclðUÞ ¼ N AND Us ¼ TRUSTED Then S(U) ¼ S(U) þ 1 //Ncl is the last network state IF Ncb ¼ A AND NclðUÞ ¼ A AND Us ¼ TRUSTED Then S(U) ¼ S(U) 2 1 IF Ncb ¼ N AND NclðUÞ ¼ N AND Us ¼ UNTRUSTED Then S(U) ¼ S(U) þ 1 IF Ncb ¼ A AND NclðUÞ ¼ A AND Us ¼ UNTRUSTED Then S(U) ¼ S(U) 2 1 IF SðUÞ $ G Then Uhb ¼ M IF Us ¼ TRUSTED Then Set U in WL IF SðUÞ # 2G Then Uhb ¼ M 0 IF Us ¼ UNTRUSTED Then Put U in BL IF U in BL Then (Block U) AND (Policy_Set ¼ Enable) IF Policy_Set ¼ Enable Then Acquire Evaluation KC IF U in WL Then Trigger Detection KC A partial Evaluation KC IF CPU ¼ X AND MEM ¼ Y Then r ¼ AVGðX; YÞ IF r $ 90 Then System ¼ Overload IF r # 45 Then Traffic ¼ Normal AND Policy_Set ¼ Disable IF (r . 45) AND (r , 90) Then Traffic ¼ Attack IF Bw $ 50 Then Ncb ¼ A IF Bw , 50 Then Ncb ¼ N When the TFN2K traffic is coming, the Filter KC will be firstly triggered to filter the users in black list. Next, it will acquire the Evaluation KC to obtain the current network situation and then report abnormal ðNcb ¼ AÞ: In the meanwhile, the Filter KC will also trigger the Detection KC for detecting the users who pass the Filter KC. If the attacker lunched ‘TCP-SYN flood’ attack, then the rule ‘(IF #SYN_RECEIVED . dSYN Then SYN-flooding DDoS attacking AND Trigger Filter KC)’ will be matched; hence the Filter KC is triggered to adapt the ACL. If one user’s behavior is abnormal comparing to his/her historical behavior, he/she is gradually moved to black list and the Policy_Set is enabled. After new policy is set, the Filter KC will trigger Evaluation KC to determine the performance of the policy. If the system capacity become high, then system state will be transformed into NORMAL; otherwise, the policy will be reset if necessary. 6. Concluding remarks In this paper, the ontology of DDoS is firstly given. Based upon the ontology, the models (including Profile model and Defense model), four KCs, and the relationships between these KCs are then proposed accordingly, where each KC may contain a set of sub-KCs or knowledge. To evaluate the degree of the Tolerance during DDoS attacks, two criteria including system capacity and network bandwidth utilization are proposed. For capturing the knowledge in the various KCs, a KA framework integrating interviewing, the Characteristics Trainer, and the User’s Behavior Learner has also been proposed. Now, we are implementing a NORM-based DDoS intrusion tolerance system for DDoS attacks to evaluate the proposed models. In the future, we are going to develop an XML-based language for modeling the complicated environment of DDoS intrusion tolerance. Acknowledgements This work was partially supported by MOE Program for Promoting Academic Excellence of Universities under grant number 89-E-FA04-1-4, High Confidence Information Systems, and National Science Council of the Republic of China under grant No. NSC93-2752-E-009-006-PAE. S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390 389 References Barlow, J., & Thrower, W. (2001). TFN2K—An analysis by Jason Barlow and Woody Thrower, http://www2.axent.com/swat/swat.htm. Bridis, T (2002). Powerful attack cripples majority of key Internet computers. Yahoo! News, Oct. 22, 2002. Cabreraa, J. B. D., et al. (2001). Proactive detection of distributed denial of service attacks using MIB traffic variables—A feasibility study. Proceedings of integrated network management, 2001 IEEE/IFIP international symposium. CERT Coordination Center, (2003). DDoS attacks, http://www.cert.org, 2002. Chang, K. C. (2002). Defending against flooding-based distributed denial of service attacks: A tutorial. IEEE Communications Magazine, 40(10). Criscuolo, P. J (2000). Distributed denial of service-Trin00, Tribe Flood Network, Tribe Flood Network 2000, and Stacheldraht. Technical Report CIAC-2319, Department of Energy—Computer incident advi- sory capability, February 2000. Dittrich, D. (2000). DDoS: Is there really a threat? Proceedings of USENIX Security Symposium, August 16. Garber, L. (2000). Denial-of-service attacks rip the Internet. IEEE Computer, 33(4), 12 – 17. Hirasawa, S., & Chu, W. W. (2003). Knowledge acquisition from documents with both fixed and free formats (Vol. 5). Proceedings of IEEE international conference on systems, man and cybernetics 2003, (pp. 4694 – 4699). Lin, S. C., et al. (2002). A new mechanism of mining network behavior. Proceedings of PAKDD’2002, Taipei, Taiwan, May 6 – 8, and Lecture Notes in Artificial Intelligence, 2336. Lin, Y. T., et al. (2003). Design and implement of new object-oriented rule base management system. Expert Systems with Applications, 25, 369 – 385. Nikiforou, S., & Fink, E. (2002) (Vol. 1). Proceedings of IEEE international conference on systems, man and cybernetics, 6 – 9 Oct, (pp. 66 – 71). Park, K., & Lee, H. (2001). On the effectiveness of route-based packet filtering for distributed DOS attack prevention in power-law Internets. Proceedings of ACM Sigcomm, August. Rafea, A., & Hassen, H. (2003). Automatic knowledge acquisition tool for irrigation and fertilization expert systems. Expert Systems with Applications, 24(1), 49 – 57. Stavridou, V., et al. (2001). Intrusion tolerant software architectures (Vol. 2). Proceedings of DARPA information survivability con- ference and exposition (DISCEX II’01), Anaheim, California, USA, June 12 – 14. Tsai, J. J. P., & Liu, A. (1999). Knowledge-based software architectures: Acquisition, specification, and verification. IEEE Transactions On Knowledge and Data Engineering, 11(1), 187 – 201. Webb, G., & Wells, J. (1999). An experimental evaluation of integrating machine learning with knowledge acquisition. Machine Learning, 35(1), 5 – 23. Xu, J., & Lee, W. (2003). Sustaining availability of web services under distributed denial of service attacks. IEEE Transactions on Computers, 52(2), 2003. Yan, H. M., & Jiang, Y. T. (2002). Internet-based knowledge acquisition and management method to build large-scale medical expert systems (Vol. 3). Proceedings of the second joint conference on EMBS/BMES, (pp. 1885 – 1886). S.-C. Lin, S.-S. Tseng / Expert Systems with Applications 27 (2004) 379–390390 http://www2.axent.com/swat/swat.htm http://www.cert.org Constructing detection knowledge for DDoS intrusion tolerance Introduction Related work Basic of DDoS Intrusion tolerance of DDoS Knowledge based system and NORM Ontology and model of DDoS Ontology of DDoS The relationships between knowledge classes Profile model Evaluation strategy Knowledge base construction The integrated knowledge acquisition (KA) framework The knowledge obtained by interviewing with domain experts Training the detection KC by Characteristic Trainer Learning the filter knowledge class by User’s Behavior Learner The verification of selected features/characteristics of DDoS NORM-based DDoS intrusion tolerance system using KCs An example of knowledge classes Concluding remarks Acknowledgements References 
work_v4suihtgkfdi7hyyeooer4htzq	----	A patent intelligence system for strategic technology planning A patent intelligence system for strategic technology planning Hyunseok Park a, Kwangsoo Kim b, Sungchul Choi c, Janghyeok Yoon d,⇑ a Department of Technology and Innovation Management, Pohang University of Science and Technology, 77 Cheongam-ro, Nam-gu, Pohang 790-784, Republic of Korea b Department of Industrial and Management Engineering, Pohang University of Science and Technology, 77 Cheongam-ro, Nam-gu, Pohang 790-784, Republic of Korea c Department of Technology Strategy, Samsung Advanced Institute of Technology, PO Box 111, Suwon 440-600, Republic of Korea d Department of Industrial Engineering, Konkuk University, 120 Neungdong-ro, Gwangjin-gu, Seoul 143-701, Republic of Korea a r t i c l e i n f o Keywords: Technology intelligence Technology planning Subject–Action–Object SAO Function Patent map Patent network Patent mining Intelligence system a b s t r a c t Patent intelligence—the transformation of content found in multiple patents into technical, business, and legal insight—is considered a key factor in gaining a competitive advantage in technologically competi- tive business environments. Although keyword-based patent intelligence tools are widely used due to their simplicity and ease of use, they are limited in that they cannot represent key technological concepts and inventive knowledge by relying only on the frequency of occurrence of defined keywords. As a rem- edy, this paper proposes a Subject–Action–Object (SAO)-based patent intelligence system. SAO structures that can be extracted from textual patent information are known as the expertise and inventive findings of the relevant patent. On the basis of semantic analysis of patent SAO structures, our proposed intelli- gence system constructs patent maps and patent networks. Building on the maps and networks, the sys- tem provides specific functionalities including identification of technology trends and significant patents, detection of novel technologies, and identification of potential infringement. This paper describes the architecture of our proposed patent intelligence system in detail, and illustrates the system’s functional- ities using case studies. We anticipate that our proposed system will be incorporated into the technology planning process to assist experts in the formulation of technology strategies. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction In today’s competitive business environments, patent intelli- gence—the transformation of content found in patents into techni- cal, business, and legal insight—is getting much attention as a tool to aid in efforts to secure competitive advantages. Analyzing pat- ents that are representative industrial property provides informa- tion about specific conditions in relation to technological or market-related development, which aids decision makers in the tracking of competitors’ activities and new innovation trends (Cantrell, 1996). Recently, companies have increasingly been mak- ing patent applications in order to protect their inventive knowl- edge. This increase in the number of patents makes patent intelligence a vital tool for formulating strategic technology plan- ning (Yoon & Kim, 2012b). Patent intelligence tools incorporate a variety of functionalities including technology monitoring, tech- nology assessment, and technology forecasting (Lichtenthaler, 2004). They provide several advantages over qualitative patent analysis done by human experts. These include (1) ability to ana- lyze large amounts of patent data that cannot be analyzed by hu- mans alone, (2) ability to generate much useful information, such as visual relationships between technology and companies (char- acteristic of statistical analysis technology), which humans cannot produce, (3) ability to support decision making processes by pro- viding relevant information including technology assessment and technology forecasting (Yoon, 2008; Yoon & Kim, 2012b). Many patent intelligence tools have been developed to support decision-making in technology planning. Overall, they can be clas- sified into two approaches—the bibliographic approach and the content-based approach. The bibliographic approach uses biblio- graphic patent information including citations, applicants, inven- tors, and international patent classification (IPC) codes. Although the bibliographic approach is widely used to identify meso- and macro-trends in technologies, companies, and inventors (No & Park, 2010), it cannot identify detailed technological features and significant insights because it mainly relies on bibliographic infor- mation, which is considered to be superficial data (Yoon & Park, 2004). In this regard, the content-based approach emphasizes technologically significant patterns, trends, and opportunities by extracting useful information such as abstracts, detailed descrip- tion of invention and claims from patent text (Tseng, Lin, & Lin, 2007; Yoon, Choi, & Kim, 2011). Content-based patent intelligence tools are increasingly being proposed by researchers. One representative content-based 0957-4174/$ - see front matter � 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2012.10.073 ⇑ Corresponding author. E-mail addresses: howgood@postech.ac.kr (H. Park), kskim@postech.ac.kr (K. Kim), blissray@gmail.com (S. Choi), janghyoon@konkuk.ac.kr (J. Yoon). Expert Systems with Applications 40 (2013) 2373–2390 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2012.10.073 mailto:howgood@postech.ac.kr mailto:kskim@postech.ac.kr mailto:blissray@gmail.com mailto:janghyoon@konkuk.ac.kr http://dx.doi.org/10.1016/j.eswa.2012.10.073 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa approach method is keyword-based analysis (KWA). Many researchers have developed keyword-based patent intelligence tools to identify trends in high-technology (Yoon & Park, 2004), to discover new technological opportunities from patent vacuums (Lee, Yoon, & Park, 2009), to forecast new technological concepts (Yoon & Park, 2005), and to develop technology roadmaps (Lee, Seol, & Park, 2008; Yoon, Phaal, & Probert, 2008). The keyword- based patent intelligence approach uses occurrence information including frequencies of defined keywords and co-occurrences among the keywords. In general, by exploiting the vector space model (Salton, Wong, & Yang, 1975) used in information retrieval, KWA transforms each patent into a keyword vector, identifies sim- ilarities among patents using similarity measures including Cosine similarity and Euclidean distance, and then constructs patent maps and patent networks. Despite its simplicity and ease of use, KWA is limited in that it cannot incorporate key technological concepts such as objectives, uses, and structures of the relevant patent (Cascini, Fantechi, & Spinicci, 2004; Cascini & Zini, 2008; Yoon et al., 2011). Furthermore, several studies have indicated that frequencies and co-occurrences of keywords in patents cannot represent the inventive knowledge and method of relevant patents (Park, Ree, & Kim, 2013; Wanner et al., 2008). As a remedy, this paper proposes an SAO-based patent intelli- gence system. SAO structures are the grammatically sequenced sentence of a subject, a verb, and an object that can be extracted by exploiting natural language processing (NLP) of textual patent information (Cascini et al., 2004). Specifically, SAO structures that are obtainable from a patent are considered as being able to pro- vide the expertise, know-how, and significant findings of the pat- ent (Bergmann et al., 2008; Moehrle, Walter, Geritz, & Müller, 2005; Sternitzke & Bergmann, 2009). Building on the semantic similarities of patents’ SAO structures, the system constructs pat- ent maps and patent networks that provide several functionalities for patent intelligence. Using the patent maps and patent net- works, the system creates significant information to support deci- sion-making for technology planning. In this paper, we describe the architecture of the proposed intelligence system and illustrate the system’s functionalities using several practical case studies. We anticipate that the proposed patent intelligence system will be incorporated into the technology planning process to assist experts in the formulation of technology strategies. The rest of this paper is organized as follows. Section 2 over- views related work while Section 3 describes the architecture and functionalities of the proposed system. In Section 4, the func- tionalities of the system are illustrated using case studies, and finally, Section 5 concludes the paper by outlining future research topics. 2. Related work Our proposed patent intelligence system is based on SAO-based patent analysis. Thus, this section lays the groundwork by explain- ing patent intelligence for technology planning and SAO-based pat- ent analysis. 2.1. Patent intelligence for technology planning Technology planning—a process that sets the goals of technol- ogy development and formulates specific strategies on how to ac- quire, manage, and exploit technologies—is recognized as the single most important ingredient for achieving the future vision or end-state of organizations (Bettis & Hitt, 1995; Kerr, Mortara, Phaal, & Probert, 2006; Phaal, Farrukh, & Probert, 2004). For suc- cessful technology planning, the knowledge and insight of human experts who are conversant with a specific domain play significant roles in various technology related decision-making processes. However, in today’s fast changing technology environment, depending solely on them (experts and managers) is insufficient, as they cannot capture every small, but important, signals of tech- nological change. Therefore, the decision-making processes of ex- perts should be supported by a systematic approach that provides valuable technological information gleaned from big technology data. Accordingly, technology intelligence—an activity that identifies the technological opportunities and threats that can affect an orga- nization’s business—has been receiving increasing attention and has come to be regarded as a key component in technology plan- ning (Cooper & Schendel, 1976; Lichtenthaler, 2003; Porter & Detampel, 1995; Utterback, 1996). In particular, Patent intelli- gence—the transformation of content found in patents into techni- cal, business, and legal insight—is receiving a great deal of attention because patents are the most abundant and well-orga- nized source of technological information. According to many sur- veys of authorities, patents cover more than 90% of the latest technical information in the world, and 80% of the patent informa- tion is not published in any other form (Zha & Chen, 2010). The use of patent intelligence tools, which facilitate the gathering and anal- ysis of patents and provides decisive information for technology planning, has several advantages. They can handle the analysis of massive numbers of patents that would otherwise require consid- erable labor or may be impossible using human experts only. In addition, they can generate the characteristics of technologies by statistical analysis and visualize technological relationships in the form of patent maps and patent networks (Yoon, 2008; Yoon & Kim, 2012). Patent maps—snapshots of the technology landscape of a technology industry—and patent networks, which show tech- nological relationships among technologies in a technology indus- try, can provide overall and specific understandings of a given technology domain, and thus they are recognized as crucial out- puts from patent intelligence tools that can increase user understandability. 2.2. SAO-based patent analysis Technological information in a patent is described in textual form in sections such as the title, abstract, summary, detailed description, and claims. In order to mine crucial information from patents, a content-based approach is adopted. Although KWA has been widely adopted due to its simplicity and easy application, it cannot incorporate the key concepts and inventive knowledge of patents. Recently, SAO-based patent analysis has been receiving attention as an alternative that overcomes KWA’s limitations. An SAO structure is composed of Subject (noun phrase), Action (verb phrase), and Object (noun phrase)—a canonical form that represents the main idea in a complex sentence. In the technology field, SAO structures have mainly been used to represent the func- tionalities of technologies. The Theory of Invention Problem Solv- ing (Russian acronym: TRIZ) adopted an SAO structure to express and analyze functional relationships among components of a tech- nical system. Because a ‘function’, which can be defined as an ac- tion or task that a system is able to perform, can be represented as Action-Object (AO), and a ‘tool’ or ‘method’ which makes a func- tion do can be represented as Subject (S) in a technological sen- tence, the functionalities of technical systems and their components can be clearly expressed using SAO structures. In par- ticular, S and O represent components of a technical system and A specifies a functional relationship between the components (Cascini et al., 2004). For example, a technological sentence such as ‘Shock-waves remove small particles’ is composed of subject (‘shock-waves’), action (‘remove’), and object (‘small particles’), and ‘remove’ clearly represents a structural relationship between 2374 H. Park et al. / Expert Systems with Applications 40 (2013) 2373–2390 https://isiarticles.com/article/26667 
work_v747xjw75jfu5drznwrk6szweu	----	Automatic recognition of quarantine citrus diseases Georgina Stegmayera,b, Diego H Milonea, Sergio Garranc, Lourdes Burdync aResearch Center for Signals, Systems and Computational Intelligence, FICH-UNL, CONICET, Ciudad Universitaria UNL, Santa Fe, (3000), Argentina, d.milone@ieee.org bCentro de Investigación en Ingenieŕıa en Sistemas de Información, CONICET, Lavaise 610, Santa Fe, (3000), Argentina, gstegmayer@santafe-conicet.gov.ar cInstituto Nacional de Tecnoloǵıa Agropecuaria (INTA), Estacion Experimental Concordia Abstract Citrus exports to foreign markets are severely limited today by fruit diseases. Some of them, like citrus canker, black spot and scab, are quarantine for the markets. For this reason, it is important to perform strict controls before fruits are exported to avoid the inclusion of citrus affected by them. Nowa- days, technical decisions are based on visual diagnosis of human experts, highly dependent on the degree of individual skills. This work presents a model capable of automatic recognize the quarantine diseases. It is based on the combination of a feature selection method and a classifier that has been trained on quarantine illness symptoms. Citrus samples with citrus canker, black spot, scab and other diseases were evaluated. Experimental work was performed on 212 samples of mandarins from a Nova cultivar. The proposed approach achieved a classification rate of quarantine/not-quarantine samples of over 83% for all classes, even when using a small subset (14) of all the available features (90). The results obtained show that the proposed method can be suitable for helping the task of citrus visual diagnosis, in particular, quarantine diseases recognition in fruits. Preprint submitted to Expert Systems with Applications November 13, 2012 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. Keywords Pattern recognition, multiclass classification, neural networks, citrus dis- eases 1. Introduction Three diseases have currently quarantine restrictions for the European Union (EU) and the United States of America citrus markets: citrus canker, black spot and scab. Despite this severe limitation, many regions continue to export to these markets by following the guidelines of the so-called Sys- tem Approach, individually agreed with the EU [1], which includes different methods to provide the quarantine security required by the citrus trade with the EU and to certify citrus quarantine with minimal risk. A key point to the success of these programs is the effectiveness of the audit work carried out in the field by inspectors, both in the packaging area where the items are processed as well as at the boarding ports. Since tolerance to the presence of symptoms of these diseases is zero, it is essential the early detection of items with such symptoms, especially when they can reach detectable levels in the ports and markets of arrival. At present, the diagnosis of these diseases, both in field and packing points, depends on the visual method based on the presence of symptoms. Due to the characteristics of the harvest and export operations in the case of having the dubious presence of symptoms, the diagnosis must be realized immediately. However, the visual diagnosis presents a number of disadvan- 2 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. tages including the fact that the accuracy and reliability of the diagnostic procedure is subject to the personal capacity of those who make it [2]. More- over, the decision-making process frequently involves subjective factors that provide some degree of variability to the result [3]. There is a clear need for computational tools to identify situations in which the value of an entire production is compromised. The literature currently available on the symptoms of diseases affecting citrus fruits is abundant [4, 5, 6]. However, symptoms of each quarantine infection are described based on a small number of very characteristic at- tributes [7, 8]. This helps in identifying and solving the diagnoses of the most typical symptoms of each disease, but are insufficient to diagnose those that are less frequent and/or share similar attributes and variants with symptoms of other non quarantine diseases. Alternative diagnostic techniques can be used, such as the incubation of fruits under temperature and light controlled conditions, or practices of isolation of the causal agent of dubious symptoms. However, they have proved to be very slow and subject to methodological and experimental errors [3]. In recent years, highly sensitive biochemical techniques have been devel- oped, which allow diagnosis within hours. However, they are highly expen- sive and require instrumental and specific training. Thus, they are not yet available today for the in situ use required in the field and packaging ar- eas [7, 9, 10]. For these reasons, efforts are needed in order to improve the current procedure of visual diagnosis for early detection of diseases [11, 12], such as technological strategies using machine learning to achieve intelligent farming [13]. Current proposals study the problem of post-harvest processing 3 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. of citrus [2], in particular, visual detection defects through image analysis to classify the fruit depending on appearance (texture and colour [14]) in an unsupervised way [15, 12]. This paper presents a classifier able to distinguish among the three quar- antine diseases mentioned, based on a binary description of the presence or absence of disease symptoms. Such classifier could be stored in a cen- tral server that could be accessed online through a simple portable device, without special equipment nor computational processing requirements. An inspector could check the symptoms that he can see on the suspected fruits, send the data to the server and receive a response from it, indicating whether the fruit may have the quarantine disease or not. The symptomatological descriptions available in the current literature are rather general in nature, aiming to allow the diagnosis of the most typical symptoms only and described using few attributes. Besides, some symptoms are common to symptoms of other non quarantine diseases. In this work, an accurate data set of symptoms has been created through careful observations and descriptions of different types and variations of symptoms caused by the quarantine diseases of interest. After that, a feature selection analysis on the attributes of diseases and their variants has been performed to select the most representatives ones. This allows to minimize the number of features needed to be loaded into a portable device. Several classifiers have been trained with the selected features that better represent each of the three quarantine infections of interest. Results for each classifier have been obtained using cross-validation. Then, the best classifier has been selected for further study, calculating specific performance metrics to deeply analyze its results, class 4 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. by class. This paper is organized as follows. Section 2 presents the materials used in this study. Section 3 explains in detail the proposed approach for quarantine diseases recognition, which includes a feature selection step and a classifier training. The performance measures used in this work are presented in Sec- tion 4. Section 5 shows the results obtained and their discussion. Finally, the conclusions and future work can be found in Section 6. 2. Materials Data set used in this sudy includes citrus canker, black spot and scab symptoms on a group of 212 Nova mandarins grown along the Uruguay River citrus growing area. The database of symptoms of each quarantine disease was manually built because, while a number of the studied diseases symptoms are described in [8], the variability observed in them is large in practice. Symptoms described in the data set were those ones that represent variants with respect to typical symptoms, for example the four typical symptoms of black spot [8]: freckle spot, hard spot, virulent or spreading spot and speckle blotch, and that are observed less frequently. The data recolection period lasted one week on May 2011, where the fruits were all mature. Although infections mainly occur during the early growth of the fruits, new symptoms can be observed several months later, when the fruits reach their final color and commercial maturity [16]. This long incubation period hampers the effectiveness of both field and packaging site monitoring inspections for the detection of the disease. The number of samples of each class (quarantine disease) was quite balanced, according to 5 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. the following detail: 54 citrus canker, 43 black spot, 45 scab and 70 other (non quarantine) diseases. The whole set was randomly divided into two subsets: data set 1 (DS1) having 25% (71 samples) of the total data for feature and model selection; data set 2 (DS2) having the remaining 75% (141 samples) for training and cross-validation testing. 3. Quarantine diseases recognition 3.1. Feature selection Feature or attribute selection is an active research area in pattern recog- nition, statistics, and data mining. Its main idea is to eliminate features with little or no predictive information and select only a subset of relevant fea- tures for building robust learning models. Feature selection can significantly improve the performance of learning models by removing most irrelevant and redundant features from the data, thus achieving better generalization to test points. Besides, it can help to improve model interpretability and comprehension [17]. The problem of feature subset selection is that of finding a subset of the original features of a dataset, such that an algorithm that is run on data containing only these features be able to generate a classifier with the highest possible accuracy. Given an algorithm I that will be used for classification and a dataset D with features F1,F2, · · · ,Fn, from a distribution over the labeled instance space, an optimal feature subset, Fopt, is a subset of the features such that the accuracy of the classifier C = I(D) is maximal [18]. Techniques for feature selection can be divided in two approaches: feature ranking, where features are ranked by some criteria and then features above 6 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. a defined threshold are selected; and subset selection, where one searches a space of feature subsets for the optimal subset. Such approach works by using a function (for example, classifier accuracy) that takes a subset and generates an evaluation value for that subset. A search is performed in the subsets space until the best solution can be found. For example, best-first search is a commonly used search algorithm which explores a state space by expanding the node with the best score first [19]. An evaluation function is used to assign a score to each candidate node. The algorithm maintains two lists, one containing a list of candidates yet to explore, and one containing a list of visited nodes. This algorithm always chooses the best of all unvisited nodes, rather than being restricted to only a small subset, such as immediate neighbours. Other search strategies, such as depth-first and breadth-first, have this restriction. In this work we have used the best-first search method for feature selection, which searches the attribute subset space by finding low-dimensional projections of the data that score highly. The features that have the largest projections in the lower dimensional space are then selected [20]. 3.2. Classification A classifier is a mapping from the space of feature values to the set of class values. Each technique uses a learning algorithm to identify a model that best fits the relationship between the attribute set and the class labels in the training data. The models should fit the training data well and also correctly predict the class label of points not seen during the training process [20]. After the feature selection step explained above, using only the selected attributes, we have trained three different classifiers (decision trees, neural 7 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. networks and naive Bayes. 3.2.1. Classification and regression tree (CART) CART is a classification method that uses data to construct a so-called decision-tree, which is then used to classify new data. The goal of CART is to create a model that predicts the value of a target variable based on several input variables. CART can handle numerical as well as categorical variables. Decision trees are formed by a set of rules, based on variables in the training set, selected to get the best split to differentiate observations based on the independent variables (the classes). Once a rule is selected and splits a node into two, the same process is applied to each child node of the resulting tree, recursively, until no further changes can be made. That is to say, until a node has the same value of the target variable, or when splitting no longer adds value to the predictions. Each branch of the tree ends in a terminal node. Each observation falls into one and exactly one terminal node, and each terminal node is uniquely defined by a set of rules [21]. In summary, 1. Take all of the data in the training set. 2. Consider all possible values of all variables. 3. Select the variable/value that produces the greatest separation in the target (x = ti is called a split). 4. If x < ti then send the data to the left part of the tree; otherwise, send data point to the right branch of the tree. 8 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. 5. Repeat same process from 3 on these two nodes of the tree and the data in each node, until no further changes can be made. 3.2.2. Naive Bayes (NB) The naive Bayes model for joint distributions has been studied exten- sively in the pattern recognition literature [22]. A naive Bayes classifier is a simple probabilistic classifier based on applying Bayes’ theorem with strong (naive) independence assumptions. The naive Bayes model assumes the con- ditional independence of all effect variables, given a single cause variable. In this model, the class variable (which is to be predicted) is the root and the attribute variables are the leaves. The model is naive because it assumes that the attributes are conditionally independent of each other, given the class. Once the model has been trained, it can be used to classify new examples for which the class variable is unobserved. A deterministic prediction can be obtained by choosing the most likely class [19]. The probability model for a classifier can be stated as a conditional model p(C|F1, ...,Fn) over a dependent class variable C with a small number of outcomes or classes, conditional on several feature variables F1 through Fn. Using Bayes’ theorem, we can write p(C|F1,F2, . . . ,Fn) = p(C)p(F1, ...,Fn|C) p(F1, ...,Fn) . (1) In practice we are only interested in the numerator of that fraction, since the denominator does not depend on C and the values of the features Fi are given, so that the denominator is effectively constant. The numera- tor is equivalent to the joint probability model p(C,F1, . . . ,Fn). Under the conditional independence assumption, we assume that each feature Fi is con- 9 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. ditionally independent of every other feature Fj for j 6= i. This means that p(Fi|C,Fj) = p(Fi|C) for j 6= i and so the joint model can be expressed as p(C,F1,F2, . . . ,Fn) ∝ p(C) n∏ i=1 p(Fi|C). (2) The training of a naive Bayes model is computed by simple frequen- cies (maximum likelihood estimate). The class distribution is estimated by p(C) = #(C)/|D|, where #(C) is the number of times the class C shows up in the training data D, with the denominator being the total number of training instances (each instance has a unique class). The naive Bayes classifier combines the naive Bayes probability model with a decision rule, such as selecting the hypothesis that is most probable; this is known as the maximum a posteriori rule. The corresponding classifier is a function that can be defined as follows: ĉ(f1,f2, . . . ,fn) = arg max c p(C = c) n∏ i=1 p(Fi = fi|C = c). (3) This means that for each possible class label, the conditional probability of each feature has to be multiplied together, given the class label. The label for which the largest product is obtained is the label returned by the classifier. 3.2.3. Multilayer perceptron (MLP) MLP is a type of artificial neural network model, which can be loosely defined as a large set of interconnected units (neurons) that are executed in parallel to perform a common global task. The units undergo a learn- ing or training process in response to input signals, adjusting the internal parameters of the neural model (weights between neurons). 10 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. The MLP model, which is the most widely used for classification problems [23], has distinct layers such as input, hidden and output, with no connec- tions among neurons belonging to the same layer. The number of layers and neurons in each layer is chosen a-priori, as well as the type of activation func- tions for the neurons. Each neuron j in each layer computes a weighted sum of its inputs and then applies an activation function to produce an output, as follows: yj = φj ( n∑ i=1 wjixi ) , (4) where yj is the neuron output, n is the number of inputs, wji is the synaptic weight connecting the input signal xi to the neuron j and φj is the activation function, which will provide a nonlinear mapping between input and target signals [24]. In the MLP model, learning is supervised and the basic learning algorithm used is backpropagation, which uses gradient descend to minimize a cost function. The cost function is generally defined as the mean square error E = 1 2 p∑ k=1 (tk −yk)2 (5) between the desired or target output (tk) and each actual network output (yk), for p output neurons. During learning, the error propagates backwards through the network and the model parameters are changed according to the so-called delta rule [24]: δwji = −η ∂E ∂wji . (6) 11 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. 4. Performance measures The objective in supervised learning is to approximate an unknown func- tion, using a set of data, searching for the model that better predicts the outputs of the unknown function. To fit the parameters of the different mod- els, their performance is compared on a dataset not used during training, to evaluate the generalization capability of each model. This is named cross- validation and it is an effective method for estimating the prediction error of a classifier to an independent data set. To reduce variability, multiple folds of cross-validation can be performed using different partitions, and the validation results are averaged over the folds. In k-fold cross-validation, the original data set is randomly partitioned into k subsets: k-1 subsets are used as training data and the remaining single subset is used for testing the model. The cross-validation process is then repeated k times (the folds), with each of the k subsets used exactly once. The k results from the folds can then be averaged to produce a single estimation [24]. In classification problems, the primary source of performance measure- ments in classification is the overall accuracy of a classifier estimated through the classification rate or accuracy (A), which is the proportion of correctly classified examples, calculated as A = tp + tn M , (7) where M is the total number of samples, tp (true positives) is number of correct predictions of a class sample; and tn (true negatives) is the number of correct predictions of a no-class sample. 12 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. Another usual performance measures commonly used in pattern recogni- tion are precision and recall [25], defined as: P = tp tp + fp , (8) R = tp tp + fn , (9) where fp (false positives) is the number incorrect predictions of a class exam- ple; fn (false negatives) corresponds to the number of incorrect prediction of a no-class example. Precision for a class can be defined as the ratio between samples correctly classified and the total number of samples assigned to a class. That is to say, it is a measure of the fraction of classified samples that are relevant. Recall for a class is the ratio between samples correctly detected over the total number of samples that actually belong to a class. Precision can be seen as a measure of exactness (fidelity), recall is a mea- sure of completeness, and the F-score is a measure that combines precision and recall through their harmonic mean as follows: F = 2 · P ·R P + R . (10) Relative operating characteristic (ROC) curves can be used to visualize the achieved trade-offs between correctly classifying positive and negative cases. A ROC curve is a graphical plot of the true positive rate also known as sensitivity, versus the false positive rate or one minus the specificity, for a classifier system as its discrimination threshold is varied. Each point on the ROC curve represents a classifier with a particular trade-off between sensi- tivity and specificity. Comparing the performance of multiple classification schemes with statistical tools requires the information represented by the 13 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. ROC curve to be collapsed into a single response variable [26]. To this end, the area under the entire ROC curve (AUC) was proposed as a suitable per- formance index [27] since it is a value between 0 and 1 and makes easier the comparison of classifiers among them. When AUC is close to 1, it means that most of the positive class samples have been assigned a score higher than any no-class sample, meaning that there is a threshold that perfectly separates the classes. 5. Results and discusion This section reports the results obtained on the data set described in Section 2. First of all, the feature selection step has been performed. Clas- sification rates for three classifiers are reported before and after the feature selection step. After that, using only the selected features, the three classi- fiers have been tuned and compared in order to select the most adequate for the recognition of the three quarantine diseases of interest. The results on global, as well as per class classification rates are reported and discussed. 5.1. Feature and model selection The feature selection, as well as the classification models tuning, have been performed through cross-validation on the smaller subset only, as sug- gested in the pattern recognition literature [28]. The data set 2 was used for model testing and discussion of results. First of all, all of the features (90) of the quarantine diseases of interest in this study were used to train and test three classifiers. Table 6 shows the global classification rate obtained with each of the studied classifiers (in 14 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. columns) after a 5-fold cross-validation procedure performed on each subset (in rows). Table 1: Accuracy (A %) for 5-fold cross-validation without feature selection (90 features) in both datasets. Data set CART NB MLP DS1 (25% data) 54.93 76.06 60.56 DS2 (75% data) 68.43 79.43 73.76 It can be seen that for each partition of the original dataset, the global classification rate is higher than 60% for both NB and MLP, being significa- tively lower in the case of CART for the data set 1. In the case of the data set 2, accuracies higher than 70% are achieved. Note that these rates are ob- tained if all the 90 features (symptoms) are used for the classification task. This implies that, at the moment of diagnosis, for example at a packaging site, an inspector having a mobile device would have to fullfill 90 boxes in a form in order to send all of the required information to a remote server and get a classifier response. However, if a feature selection procedure is performed over the original 90 features, only the most important features would be used. This would significatively speed the in situ diagnosis task, by reducing the amount of information to be provided to the classifier. This approach, however, implies that less information is given to the classifier, which could reduce its performance. We will show, however, that the per- formance of the classifier can be maintained, and even, increased, if only the relevant features to discriminate between classes is provided, and noisy and 15 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. redundant features are not considered for the classification. For the feature selection step, a 5-fold cross-validation procedure has been performed to decide which features to use. In this step, DS1 was used for feature and model structure selection, and DS2 for estimation of the final error [29]. Once the feature selection step was completed by performing a classical best-first search, 14 of a total of 90 features were selected (see Table 2). For CART, the minimal number of observations at the terminal nodes was 2 and minimal cost-complexity pruning has been used [21]. For the MLP model, the best results have been obtained after 100 training epochs with a topology of 5 hidden neurons and using a 10% of the training set for monitoring the generalization peak. 5.2. Classifiers training and testing on selected features After the feature and model selection steps, the classifiers were trained using the WEKA software1 by performing 5-fold cross-validation on DS2 [30]. With this approach, the data set was divided into 5 mutually exclusive folds with approximately the same class distribution as the original data set. Each fold was used once to test the performance of the classifier that was generated from the combined data of the remaining folds, leading to 5 independent performance estimates [26]. Table 3 reports the classifiers performance obtained on DS2 using the 14 main features previously selected. It can be seen that, while NB has maintained its classification rate, both CART and MLP have even increased their performance. In fact, the MLP model has reached a very high clas- 1www.cs.waikato.ac.nz/ml/weka/ 16 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. sification rate (almost 84%) on the 5-fold test partitions of the larger data subset. This result can be explained with the fact that, once only the most informative features haven been considered for classification, removing noisy and redundant information, the generalization capability of the models has been improved. Moreover, by using the best features for class discrimination, we are performing a better training of the models because there is a better relationship between data size and number of parameters to estimate. Once the better classifier was obtained, per class performance has been analyzed. Table 4 shows the detail of the performance measures used in this study (columns) for each of the classes (quarantine diseases) included (rows). It can be noticed that the best classifier found (the MLP model) has very high values, almost all of them are near the optimum (1.0). In particular, for the ROC area index (AUC), the values obtained assure a very high performance for the automatic recognition of any of the quarantine diseases under study, since the obtained values are, in all cases, higher than 0.90. Table 5 shows the performance of the MLP classifier for the recognition of each of the diseases, through a confusion matrix. Looking at the detail of the confusion matrix, it can be seen that the higher mis-classification occurs with the other class, not the classes of interest. That is to say, at the presence of a sample of a quarantine disease the classification will be correct. However, at the presence of a sample of a non-quarantine disease, there are 11 samples that would be indicated as quarantine, and 13 quarantine samples that would be indicated as other disease. It is important to highlight the fact that the model has no confusion among the three classes of interest. That is to say, for the quarantine diseases 17 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. of interest in this study, there are no misclassified elements: each sample of each quarantine disease has been correctly identified. For the classification problem stated as binary, that is to say, if only two classes are considered: quarantine disease and not quarantine disease, the corresponding confusion matrix is shown in Table 6. The classification rate corresponding to the binary problem and the MLP model is of 83%, which maintains the result obtained with this classifier in the multiclass problem. If the confusion matrix is analyzed in detail, it can be seen that, as stated before, the quarantine diseases samples of interest are correctly recognized in most cases: in 92 out of 96 examples of quarantine disease class are correctly classified. The other class has less examples and the confusion if higher. This fact should be tackled in future works, increasing the number of non quar- antine diseases samples or through a better classifier design. Furthermore, existing proposals for visual detection of citrus defects through images and machine vision [31], that is to say, depending on the texture and colour of the fruit [15, 12], could be used as a preliminary step to our classifier, obtaining the features automatically through image analysis. 6. Conclusions and future work It has been explained how citrus exports to foreign markets are limited today, mainly, by fruit diseases. Some of them are quarantine for the mar- kets and have zero-tolerance at the destination market. For this reason, it is important to perform good controls before fruits are exported. Nowadays, technical decisions are highly dependent on the degree of individual skills on human experts, with previous experience in visual diagnosis. This work has 18 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. presented a model capable of recognizing three quarantine diseases (citrus canker, black spot and scab) in an automatic way and achieving high classi- fication rates, by using barely a bit more than a dozen of characteristics or symptoms that can be seen in a fruit. The proposed approach is based on the combination of a feature selec- tion method and a classifier that has been trained on the illness symptoms. Experimental work was performed on 212 Nova mandarins. The proposed approach achieved a classification ratio of quarantine/not-quarantine sam- ples of 83% for all classes, even when using a small subset (14) of all the available features (90). When the problem was stated as multiclass, also high classification rates of 84% was achieved and AUC values higher than 95%, very close to the optimum. Since only the most informative features have been considered, removing noisy and redundant information, the gen- eralization capability of the models has been improved, with a direct impact on the model performance. The high classification rates that have been obtained on the task of au- tomatic recognition of quarantine citrus diseases show the usefulness of the proposed approach. Another advantages of the proposed method is the sig- nificant reduction of the number of features that have to be used in order to obtain a high classifier response, which could be very helpful for visual inspection on the field, if, for example, the classifier was implemented in a mobile device used at field and packaging site monitoring inspections for the detection of the diseases. The results obtained show that the proposed method can be suitable for helping the task of citrus visual diagnosis, in particular, quarantine diseases 19 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. recognition in fruits in the field. All of the quarantine diseases samples are correctly recognized into each corresponding class. As future work, it would be very interesting to design a hierarchical classifier, for the not quarantine class only, capable of better discerning with which quarantine disease the sample is being confused. References [1] SENASA, Programa de Certificación de Fruta Fresca Ćıtrica de la Región NEA para exportacion a la Unión Europea. Resolución 78/2001, SENASA, 2001. [2] J. Gmez-Sanchis, J. D. Martn-Guerrero, E. Soria-Olivas, M. Martnez- Sober, R. Magdalena-Benedito, J. Blasco, Detecting rottenness caused by penicillium genus fungi in citrus fruits using machine learning tech- niques, Expert Systems with Applications 39 (1) (2012) 780 – 785. [3] N. A. Peres, R. Harakava, G. Carroll, J. Adaskaveg, L. Timmer, Com- parison of molecular procedures for detection and identification of guig- nardia citricarpa and g. mangiferae, Plant Disease 91 (5) (2007) 525–531. [4] R. Bernal, Some aspects of the epidemiology of citrus scab (Elsinoe spp) in Satsuma mandarin cv Owari and Valencia orange ( Citrus sinensis (L) Osbeck, in: in Proceedings 9 th. International Society of Citriculture Congress. Orlando. Florida. EEUU, Vol. 1, 2000, pp. 990–993. [5] B. Canteros, Gúıa para la Identificación y Manejo de las Enfermedades Fúngicas y Bacterianas en Citrus, CFI-INTA, 2009. 20 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. [6] M. Dewdney, L. Timmer, Alternaria Brown Spot. UF IFAS Extension, Publication PP152, 2011. [7] T. Gottwald, J. Graham, T. Schubert, Citrus canker: The pathogen and its impact, Plant Management Network. [8] J. Kotze, Black spot. in: Compendium of citrus diseases, W.L. Timmer and S.M. Garnsey and J.H. Graham, editors. Second Edition. APS Press, 2000. [9] R. Baayen, et al., Nonpathogenic isolates of the citrus black spot fungus, guignardia citricarpa, identified as a cosmopolitan endophyte of woody plants, g. mangiferae (phyllosticta capitalensis), Phytopathology 92 (5) (2002) 464–477. [10] L. Meyer, G. Sanders, R. Jacobs, L. Korsten, A one-day sensitive method to detect and distinguish between the citrus black spot pathogen guig- nardia citricarpa and the endophyte guignardia mangiferae, Plant Dis- ease 90 (1) (2006) 97–101. [11] T. Brosnan, D.-W. Sun, Inspection and grading of agricultural and food products by computer vision systems - a review, Computers and Elec- tronics in Agriculture 36 (2) (2002) 193–213. [12] S. Sankaran, A. Mishra, R. Ehsani, C. Davis, A review of advanced techniques for detecting plant diseases, Computers and Electronics in Agriculture 72 (1) (2010) 1–13. [13] T. Pydipati, T. Burks, W. Lee, Identification of citrus disease using color 21 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. texture features and discriminant analysis, Computers and Electronics in Agriculture 52 (1) (2006) 49–59. [14] F. Bianconi, E. Gonzlez, A. Fernndez, S. A. Saetta, Automatic clas- sification of granite tiles through colour and texture features, Expert Systems with Applications 39 (12) (2012) 11212 – 11218. [15] F. Lopez-Garcia, G. Andreu-Garcia, J. Blasco, N. Aleixos, J.-M. Va- liente, Automatic detection of skin defects in citrus fruits using a mul- tivariate image analysis approach, Computers and Electronics in Agri- culture 71 (2) (2010) 189–197. [16] S. Garran, R. Garin, Evolucion pre- y postcosecha de los sintomas de mancha negra (guignardia citricarpa kiely) en naranja valencia late y mandarina nova en la region citrcola de concordia en entre rios, in: Programa y Resumenes V Congreso Argentino de Citricultura, 2005, p. 84. [17] I. Guyon, A. Elisseeff, An introduction to variable and feature selection, Journal of Machine Learning Research. Special Issue on Variable and Feature Selection 3 (3) (2003) 1157–1182. [18] R. Kohavi, G. John, Wrappers for feature subset selection, Artificial intelligence 97 (2) (1997) 273–324. [19] S. Russell, P. Norvig, Artificial Intelligence: A Modern Approach (3rd Edition), Ed. Prentice Hall, New York, 2009. 22 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. [20] I. H. Witten, E. Frank, M. Hall, Data Mining: Practical Machine Learn- ing Tools and Techniques. Third edition, Morgan Kaufmann Publishers Inc. San Francisco, CA, USA, 2011. [21] L. Breiman, J. Friedman, R. Olshen, C. Stone, Classification and regres- sion trees, Wadsworth Brooks/Cole Advanced Books Software, Mon- terey, CA., 1984. [22] R. Duda, P. Hart, Pattern classification and scene analysis, Wiley, New York, 1973. [23] H. Zheng, S. Fang, H. Lou, Y. Chen, L. Jiang, H. Lu, Neural network prediction of ascorbic acid degradation in green asparagus during ther- mal treatments, Expert Systems with Applications 38 (5) (2011) 5591 – 5602. [24] S. Haykin, Neural Networks and learning machines (3rd Edition), Ed. Prentice Hall, New York, 2008. [25] D. L. Olson, D. Delen, Advanced Data Mining Techniques, Springer, 2008. [26] D. Pietersma, R. Lacroix, D. Lefebvre, K. M. Wade, Performance anal- ysis for machine-learning experiments using small data sets, Computers and Electronics in Agriculture 38 (1) (2003) 1–17. [27] A. P. Bradley, The use of the area under the ROC curve in the evaluation of machine learning algorithms, Pattern Recognition 30 (7) (1997) 1145– 1159. 23 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. [28] J. Demsar, Statistical comparisons of classifiers over multiple data sets, The Journal of Machine Learning Research 7 (1) (2006) 1–30. [29] S. Arlot, A. Celisse, A survey of cross-validation procedures for model selection, Statist. Surv. 4 (1) (2010) 40–79. [30] R. Remco, et al., Weka-experiences with a java open-source project, Journal of Machine Learning Research 11 (1) (2010) 2533–2541. [31] Z. Wen, Y. Tao, Building a rule-based machine-vision system for de- fect inspection on apple sorting and packing lines, Expert Systems with Applications 16 (3) (1999) 307 – 313. 24 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. Table 2: Selected features. All features are binary. Feature Description Shape Circular spot Shape of the symptom in the infected area Topography Depressed Topography of the symptom of the surface Prominent on the surface of the infected area Deepness Shallow Deepness reached by the necrotic tissue within the infected area Transition Constellation Kind of transition zone between zone Aureole healthy and necrotic tissue Color of the Oily-water-soaked Color-aspect of the transition transition zone zone between healthy - ill tissue Central color White Predominant color of the symptom central zone Ruggedness of the Eruptive Type of the surface of the central surface Edge perimeter central zone of the symptom Pattern of the Flat and smooth External aspect of the central zone central zone Central Corky & granular Texture of the tissue in the texture Scabby central zone of the symptom Presence of Pycnidia present Presence of pycnidia on the fruting bodies central zone of the symptom 25 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. Table 3: Accuracy (A %) for 5-fold cross validation on DS2 after the feature selection step performed on DS1 (14 features selected). Class CART NB MLP citrus canker 83.80 91.90 94.60 black spot 70.00 83.30 80.00 scab 72.40 82.80 86.20 global 70.92 78.72 83.69 Table 4: Detailed performance by class of the MLP (best) model. Class P R F AUC citrus canker 0.92 0.95 0.93 0.99 black spot 0.80 0.80 0.80 0.94 scab 0.93 0.86 0.89 0.95 Table 5: Confusion matrix for the best classification model (MLP) - multiclass problem. classified as → a b c d a = citrus canker 35 0 0 2 b = black spot 0 24 0 6 c = scab 0 0 25 4 d = other 3 6 2 34 26 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. Table 6: Confusion matrix for the best classification model (MLP) - two class problem. classified as → a b a = quarantine disease 92 4 b = not quarantine disease 20 25 27 si nc (i ) R es ea rc h C en te r fo r S ig na ls , S ys te m s an d C om pu ta ti on al I nt el li ge nc e (f ic h. un l. ed u. ar /s in c) G . S te gm ay er , D . H . M il on e, S . G ar ra n & L . B ur dy n; " A ut om at ic r ec og ni ti on o f qu ar an ti ne c it ru s di se as es " E xp er t S ys te m s W it h A pp li ca ti on s, 2 01 2. 
work_v7ptiurqnvegdi23ii6ufbnu2i	----	Riss et al, Expert systems 283 J. Perinat. Med. 16 (1988) 283 Development and application of simple expert systems in obstetrics and gynecology Paul A. Riss, Heinz Koelbl, Alexander Reinthaller, and Josef Deutinger 2nd Department of Obstetrics and Gynecology, University Hospital, A-1090 Vienna, Austria 1 Introduction In recent years the accepted roles of computers have expanded from the traditional ones of data retrieval, storage and processing into the area of what is often called "artificial intelligence"; that is into areas which would have been called "intelli- gent" if they had been performed by man. Any field of medicine, in which the number of answers to a particular question is limited — e.g. the number of possible diagnoses — and in which it is possible to contruct a line of reasoning is well suited for computers using intelligent programs [4]. We have developed simple rule-based expert sys- tems for 2 clearly defined applications in obstetrics and gynecology: cycle stimulation and assessment of urinary incontinence. In the present paper we describe the development and clinical application of these 2 expert systems, and discuss the advan- tages and limitations of computer, in assisting medical decision making. 2 Hardware and software We used an IBM-PC with 2 disk drives and 256k RAM. The expert system was developed using the expert system shell EXSYS (Exsys Inc, Albuquer- que, New Mexico, U. S. A.), a rule based system consisting of an editor to develop and edit the knowledge base (the rules), and a RUNTIME program to run the expert system. The expert has to define and write the rules. The structure of the rules is in the form of "IF...THEN..." (table I), in which "THEN" is Curriculum vitae PAUL A. RISS was born in Vienna, Austria, in 1948. He studied medicine in Vienna and Lausanne, and began his residency in ob- stretrics and gynecology at the University Hospital of Vienna in 1977. In addition to clinical work, he did re- search on questions related to beta-endorphin, and be- came assistant professor in 1985. Since 1982 he has been involved with hospital in- formation systems, artificial intelligence and decision ana- lysis. His other area of interest is that of gynecological urology. either one of the solutions, or a statement which is used in a subsequent rule. All possible solutions — called "choices" — have to be defined at the outset and must be found in the "THEN" part of at least on rule. The expert can assign probabilities to the different rules. Notes and references can be added to every rule to provide additional infor- mation concerning each rule and to explain how the expert arrived at the solution (table I). When in use, the system asks specific questions of the user and they are answered either by giving one of a group of several responses (multiple choice form) or by entering numeric data. After having prompted the user for the required infor- mation the system uses backward chaining to cal- culate the most likely solutions, which are then displayed in decreasing order of probability. 1988 by Walter de Gruyter & Co. Berlin · New York 284 Riss et al, Expert systems Table I. Example of rule for cycle day 9 to 12 suggesting continued cycle stimulation Rule number: 12 IF: and and and and THEN: [CD] > 8 [CD] < 13 [E2] < 1000 [DIFFE2] > = 0 [DIFFP] < = 50 CONTINUE STIMULATION - Probability = 10/10 Reference: Rule for continuation of cycle stimulation from cycle day 9-12 if E2 is still below 1000 and in the absence of luteinization 3 Development of a rule-base expert system In building an expert system — and in particular a rule based expert system — one has to follow certain steps [1]. After defining the application the first step consists in identifying the possible out- comes (choices, solutions, diagnoses, suggested ac- tions, etc). The next step is the selection of the variables which will be used in the expert system: the variables should be easily and consistently available, reliable, and actually used in clinical work. Then the rules have to be written, before the expert system can be tested and modified. A very important point, which will not be addressed in the present paper, is the maintenance and reg- ular updating of the expert system (table II). Before starting with an expert system, however, one has to decide which strategy to adopt. In the present paper we describe two possible approaches — statistical and heuristic — in the construction of rule based expert systems for limited applica- tions in clinical medicine. 4 Expert system for cycle stimulation In the first application we developed an expert system for stimulation of the menstrual cycle in our in vitro fertilization program. The menstrual cycle was stimulated with clomiphene and human menopausal gonadotropin (HMG), and ovulation was induced with human chorionic gonadotropin (HCG). Beginning on cycle day 8 the patient is Table Π. Steps in the development of an expert system 1. Definition of the problem and of possible solutions 2. Selection of the variables to be used 3. Development of the rules 4. Preliminary testing and modification of the rules 5. Evaluation of the expert system (Prospective testing) [6. Maintenance and updating] monitored by daily serum hormone level deter- minations and ultrasound examinations [5], Seven variables were used for the expert system, the end point was one of 3 possible lines of action: 1) continue stimulation, 2) induce ovulation, and 3) cancel the treatment cycle. These solutions are mutually exclusive, we did not add probabilities (or certainty factors) to the solutions. We now wrote a total of 27 rules, trying to cover every possibility applying to the cycle days 8 to 14. The rules as well as the order in which the rules were placed in the knowledge base, i. e. the expert system, were decided by the developers and based entirely on experience and intuition. After preliminiary testing and modification of the rules we tested the expert system on 62 consecutive patients of our in vitro fertilization program (244 cycle days). The misclassification rate was 5% for continuation of stimulation and 10% for induc- tion of ovulation. 5 Expert system for diagnosis of female urinary incontinence Urodynamic examination is done preoperatively to make a differential diagnosis of urinary incon- tinence in incontinent women. All urodynamic investigations performed in our department in the' years 1984 and 1985 were used for the develop- ment of the expert system. The only conditions needing to be satisfied were that the patient com- plained of involuntary loss of urine and that the urodynamic examination was complete (N = 465). The prevalence of the 5 possible di- agnoses was as follows: stress incontinence 50.5%, motor urge incontinence 1.7%, sensory urge 2.6%, mixed incontinence 18.3%, and no incontinence demonstrable 26.9%. The variables obtained at urodynamic examination were used for the con- struction of the expert system. J. Perinat. Med. 16 (1988) Riss et al, Expert systems 285 In contrast to the first application we used a twofold approach for the development of the ex- pert system for the diagnosis of incontinence [6]: firstly, we calculated the predictive value of each vailable variable for each of the five urodynamic diagnoses. For every diagnosis we selected the 4 variables with the highest predictive values and designed a tree diagram, using the variables in the order of their predictive values [7, 9]. This process of building an empirical tree diagram by repeti- tively splitting the patient population into smaller and smaller categories is called "recursive parti- tioning analysis" [3]. For each diagnosis we thus obtained several sub- sets of patients, and for each subset we calculated the predictive values of the combination of uro- dynamic variables characterizing the subset. We used the combinations of variables for each subset to develop of rule (IF... AND... AND... THEN DIAGNOSIS X PROBABILITY Y) and assigned to the diagnosis the probability calculated as the predictive value for the given subset. We then developed a number of rules to either exclude or establish a diagnosis on the basis of our experience. Finally, after preliminary testing of the expert system we either modified some rules or the assigned probability. The expert system consists of 44 rules. Table III shows the result of prospective testing of the expert system on 54 consecutive patients complaining of incontinence. 6 Comment Expert systems have become very popular, ranging from simple programs consisting of a few rules to large systems requiring years of development. We present 2 simple applicatins which we were able to develop in the course of a few weeks, using commercial software and a standard personal computer. However, we feel that the development of an ex- pert system is a major intellectual task: the prob- lems have to be clearly defined, variables must be available and reliable, and the right approach to the development of the knowledge base has to be chosen. While in theory this seems simple, in prac- tice it can be very time-consuming and difficult to dissect medical knowledge and experience into the small pieces required for the knowledge base. The statistical approach described here offers the advantage of yielding probabilities for the differ- ent solutions, and these probabilities can be com- bined. On the other hand, the major drawback of the purely statistical approach is that every com- bination of variables not only has to be found in the study group but has to exist with sufficient prevalence in order to permit the calculation of meaningful predictive values [8]. The purely heuristic approach — as used in the expert system for cycle stimulation — does not permit the assignment of calculated probabilities. In addition, it is almost impossible for an expert to think of all possibilities and to cover every aspect of a solution. Both expert systems presented here are in actual use in our department. In summary, we can draw several conclusions with regard to the develop- ment and the use of expert systems: It is very difficult to cover every aspect of a prob- lem with an expert system. While the statistical approach chosen in the second application assigns certainty factors to every combination of variables seen in the last 2 years, the heuristic approach to cycle stimulation can only be refined through con- tinued testing; even then the developer can never be sure that he has thought of all possibilities. We doubt that an expert system for diagnosis or treatment can be better than a human expert [2]. An expert system might give a more complete list of solutions, or might suggest solutions the phy- sician may not have thought of, or might do Table ΙΠ. Results of prospective testing of the expert system for assessment of urinary incontinence on 54 patients (PV = predictive value) Sensitivity Specificity PVpos PVneg Genuine stress incontinence Motor urge incontinence Mixed incontinence No incontinence demonstrable 96.7 50.0 50.0 100.0 91.7 100.0 95.5 97.6 93.6 100.0 71.4 92.9 95.7 98.1 89.4 100.0 J. Perinat. Med. 16 (1988) 286 Riss et al, Expert systems calculations which are difficult or time consuming to do by hand. However, the essential aspect of an expert is his (or her) ability to learn from experience and to evaluate a situation on the basis of a year long and constantly updated experience. In contrast, most expert systems — and certainly simple rule-based systems — are not able to learn. On the other hand, expert system are very valuable in teaching situations: the different steps in the development of a diagnosis can be clearly shown and explained. Last not least, expert systems can be very helpful when a "true" expert is not readily available. Abstract Expert systems have become increasingly popular in medicine for the support of medical decisions, e. g. di- agnosis or treatment. We describe the development and application of 2 simple rule-based expert systems, one used for cycle stimulation in our in vitro fertilization program and the other for preoperative assessment of urinary incontinence. The programs were written using a commercially available expert system shell, are run on a standard personal computer, and are in actual use in our department. Though it is doubtful that simple expert systems can be superior to a human expert we feel that expert systems are very useful in the standardization of protocols and are a valuable teaching instrument. Keywords: Artificial intelligence, cycle stimulation, expert system, urinary incontinence. Zusammenfassung Entwicklung und Anwendung yon einfachen Expertensy- stemen in der Gynäkologie und Geburtshilfe Wir beschreiben die einzelnen Schritte in der Konstruk- tion von 2 einfachen Expertensystemen: Definition des Problems, Identifizierung der möglichen Endpunkte (Diagnosen, Therapien, Handlungsstrategien, etc), Aus- wahl der zu verwendenden Variablen, Schreiben der Re- geln, Testen und Modifizieren des Expertensystems. Die Programme wurden mittels des Expertensystem-Umge- bung EXSYS geschrieben und laufen auf Personal Com- putern. Als erste Anwendung entwickelten wir ein Expertensy- stem für die Stimulierung des Menstruationszyklus mit- tels clompiphen und humanem menopausalen Goando- tropinen (HMG) für unser in vitro Fertilisierungs (IVF) Programm. 7 Variable wurden für das Expertensystem verwendet, die Endpunkte waren 3 mögliche Handlungs- stragegien: 1) Stimulation weiterführen, 2) Ovulation induzieren, und 3) Behandlungszyklus abbrechen. Wir schrieben insgesamt 27 Regeln, basierend auf Erfahrung und Intuition. Wir testeten das Expertensystem an 62 aufeinanderfolgenden Patientinnen unseres IVF-Pro- gramms (244 Behandlungstage). Die Rate an falschen Klassifizierungen durch das Expertensystem betrug 5% bei der Diagnose: Stimulation weiterführen und 10% in Bezug auf Auslösung der Ovulation. Im Gegensatz zur ersten Anwendung benutzten wir bei der Entwicklung des zweiten Expertensystems zur Un- terstützung der Diagnose der weiblichen Harninkonti- nenz durch die urodynamische Untersuchung eine zwei- fache Strategie: zunächst berechneten wir den Vorher- sagewert in Bezug auf 5 mögliche urodynamische Dia- gnosen für verschiedene Untergruppen von Patientin- nen. Wir verwendeten diese Vorhersagewerte bei der Entwicklung der Regeln für das Expertensystem, indem wir die Vorhersagewerte den einzelnen Diagnosen als Wahrscheinlichkeiten zuordneten. Dann schrieben wir aufgrund unserer Erfahrung zusätzlich einige Regeln zur definitiven Bestätigung einzelner Diagnosen. Das Ex- pertensystem besteht aus 44 Regeln und wurde an 54 Patientinnen getestet. Die Vorhersagewerte des Exper- tensystems lagen zwischen 70 und 100%. Zusammenfassend konnten wir die Möglichkeit der Konstruktion von einfachen Expertensystemen mit einer Kombination von statistischen und heuristischen Ver- fahren zeigen. Obwohl wir bezweifeln, daß einfache Ex- pertensysteme besser als menschliche Experten sein kön- nen sehen wir den Nutzen von Expertensystemen unter anderem in der Standardisierung von Protokollen und im Einsatz im Unterricht. Schlüsselwörter: Expertensystem, Harninkontinenz, künstliche Intelligenz (KI), Zyklusstimulation. J. Perinat. Med. 16 (1988) Riss et al, Expert systems 287 Resume Developpement et application de systemes experts simples en obstetrique et gynecologic Nous decrivons les etapes du developpement de systemes experts a regies simples: definition de Fapplication, iden- tification des points terminaux possibles (diagnostics, evolutions, choix), selection de variables, ecritures des regies, test et modification du Systeme expert. Les pro- grammes ont etc rediges ä l'aide d'un Systeme expert du commerce (EXSYS) et utilises sur un ordinateur person- nel standard. Comme premiere application nous avons mis au point un Systeme expert pour l'induction du cycle menstruel par clomiphene et gonadodrophines humaines en pro- venance de femmes menopausees (HMG) dans notre programme de fecondation in vitro (FIV). On a utilise 7 variables pour le Systeme expert, le point terminal etait Tune des 3 actions possibles: 1) poursuite de la stimu- lation, 2) induction de Fovulation, et 3) abandon du cycle de traitement. Au total, nous avons redige 27 regies fondees sur Fexperience et l'intuition du realisateur. Nous avons teste le Systeme expert chez 62 patientes consecutives de FIV (244 jours de cycle). Le pourcentage d'erreur de classification a etc de 5% pour la poursuite de la stimulation et de 10% pour Tinduction de Fovu- lation. Nous avons utilise une approche double pour le deve- loppement d'un Systeme expert pour le diagnostic d'in- continence urinaire feminine par exploration urodyna- mique, ce qui contraste avec la premiere application: dans un premier temps nous avons calcule les valeurs predictives de 5 diagnostics urodynamiques possibles pour differents groupes de patientes. Ensuite nous avons utilise les valeurs predictives pour mettre en ceuvre les regies et nous avons attribue au dignostic la probabilite calculee comme valeur predictive pour le groupe donne. Nous avons ensuite developpe un certain nombre de regies soit pour exclure soit pour etablir un diagnostic sur la base de notre experience. Le Systeme expert comporte 44 regies et a ete teste chez 54 patientes conse- cutives. Les valeurs predictives positives vont de 70 a 100%. En resume, nous avons demontre la faisabilite du de- veloppement de systemes experts simples par la combi- naison de principes heuristiques et statistiques. Bien que puisse douter que des systemes experts simples puissent etre superieurs ä un expert humain, nous res- sentons que les systemes experts sont tres utiles pour la standardisation de protocoles et qu'ils representent un instrument d'enseignement de valeur. Mots-cles: Incontinence urinaire, intelligence artificielle, stimulation de cycle, Systeme expert. References [1] ADLASSNIG KP, G KOLARZ, W SCHEITHAUER, H EFFENBERGER, G GRABNER: CADIAG: Approaches to computer-assisted medical diagnosis. Comp Biol Med 15 (1985) 315 [2] CHARD T: Human versus machine: a comparison of a computer "expert system" with human experts in the diagnosis of vaginal discharge. Int J Biomed Comp 20 (1987) 71 [3] FRIEDMAN JH: A recursive partitioning decision rule for nonparametric classification. IEEE Trans Corn- put 16 (1977) 404 [4] REGGIA JA, S THURIM: An introduction to com- puter-assisted medical decision making II. MD- Computing 2 (1985) 40 [5] REINTHALLER A, P RISS, J DEUTINGER, E MUELLER- TYL: Factors influencing successful in vitro fertil- ization and embryo transfer: a matched pair study. Fertil Steril 46 (1986) 511 [6] RISS P, H KOELBL: Strategien zur Entwicklung eines Expertensystems am Beispiel der urodynamischen Untersuchung. In: RAPPELSBERGER P, P PFUNDNER, G GELL: Medizin-Technik Medizinische Informatik '86, pp 351. R. Oldenbourg, Wien-München 1986 [7] RISS P, H KOELBL: The predictive values of single and combined urodynamic parameters. Acta obstet gynecol scand (in print) [8] STEMPEL LE: Eenie, meenie, minie, mo ... What do the data really show? Am J Obstet Gynecol 144 (1982) 745 [9] WASSON JH, HC Sox, RK NEFF, L GOLDMAN: Clin- ical prediction rules. Applications and methodolog- ical standards. N Engl J Med 313 (1985) 793 Dr. Paul A. Riss 2. Univ. Frauenklinik Spitalgasse 23 A-1090 Vienna, Austria J. Perinat. Med. 16 (1988) 
work_vckd3ttwbjgldfrgzguiw7srf4	----	doi:10.1016/j.eswa.2007.08.049 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 35 (2008) 1177–1185 Expert Systems with Applications Incremental clustering of mixed data based on distance hierarchy Chung-Chian Hsu a, Yan-Ping Huang a,b,* a Department of Information Management, National Yunlin University of Science and Technology, Taiwan b Department of Information Management, Chin Min Institute of Technology, Taiwan Abstract Clustering is an important function in data mining. Its typical application includes the analysis of consumer’s materials. Adaptive resonance theory network (ART) is very popular in the unsupervised neural network. Type I adaptive resonance theory network (ART1) deals with the binary numerical data, whereas type II adaptive resonance theory network (ART2) deals with the general numer- ical data. Several information systems collect the mixing type attitudes, which included numeric attributes and categorical attributes. However, ART1 and ART2 do not deal with mixed data. If the categorical data attributes are transferred to the binary data format, the binary data do not reflect the similar degree. It influences the clustering quality. Therefore, this paper proposes a modified adaptive resonance theory network (M-ART) and the conceptual hierarchy tree to solve similar degrees of mixed data. This paper utilizes artificial simulation materials and collects a piece of actual data about the family income to do experiments. The results show that the M-ART algorithm can process the mixed data and has a great effect on clustering. � 2007 Elsevier Ltd. All rights reserved. Keywords: Adaptive resonance theory network; Conceptual hierarchy; Clustering algorithm; Unsupervised neural network; Data mining 1. Introduction Clustering is the unsupervised classification of patterns into groups. It is an important data analyzing technique, which organizes a collection of patterns into clusters based on similarity (Hsu, 2006; Hsu & Wang, 2005; Jain & Dubes, 1988). Clustering is useful in several exploratory pattern-analysis, grouping, decision-making, and machine- learning situations. This includes data mining, document retrieval, image segmentation, and pattern classification. Clustering methods have been successfully applied in many fields including pattern recognition (Anderberg, 1973), biology, psychiatry, psychology, archaeology, geology, geography, marketing, image processing (Jain & Dubes, 1988) and information retrieval (Rasmussen, 1992; Salton & Buckley, 1991). Intuitively, patterns with a valid cluster 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.08.049 * Corresponding author. Address: Department of Information Man- agement, National Yunlin University of Science and Technology, Taiwan. Tel.: +886 37627153; fax: +886 37605684. E-mail addresses: hsucc@yuntech.edu.tw (C.-C. Hsu), sunny@ms. chinmin.edu.tw (Y.-P. Huang). are more similar to each other than they are to a pattern belonging to a different cluster. Data clustering has been considered as a primary data mining method for knowledge discovery. There have been many clustering algorithms in the literature. In general, major clustering methods can be classified into the hierar- chical or the partition category. A hierarchical method cre- ates a hierarchical decomposition of the given set of data patterns. A partition approach produces k partitions of the patterns, where each partition represents a cluster. Fur- ther classification in each of the categories is possible (Jain & Dubes, 1988). In addition, Jian (1999) discussed some cross-cutting issues that might affect all of the different approaches regardless of their placement in the categories (Jain, Murty, & Flynn, 1999). Being non-incremental or incremental is one of the issues (Hsu, 2006; Hsu & Wang, 2005). Non-incremental clustering methods process all the data patterns at a time. These algorithms usually require the entire datasets being loaded into memory and therefore have high requirement in memory space. The major advantage with the incremental clustering algorithms is that it is not necessary to store the entire mailto:hsucc@yuntech.edu.tw mailto:sunny@ms. chinmin.edu.tw mailto:sunny@ms. chinmin.edu.tw 1178 C.-C. Hsu, Y.-P. Huang / Expert Systems with Applications 35 (2008) 1177–1185 pattern matrix in the memory. So, the space requirements of incremental algorithms are very small. Incremental clus- tering considers input patterns one at a time and assigns them to the existing clusters (Jain & Dubes, 1988). Here, a new input pattern is assigned to a cluster without affect- ing the existing clusters significantly. Moreover, a major advantage of the incremental clustering algorithms is their limited space requirement since the entire dataset is not necessary to store in the memory. Therefore, these algo- rithms are well suited for a dynamic environment and for very large datasets. They have already been applied along these directions (Can, 1993; Ester, Kriegel, Sander, Wim- mer, & Xu, 1998; Somlo & Adele, 2001). Most of clustering algorithms consider either categorical data or numeric data. However, many mixed datasets including categorical and numeric values existed nowadays. A common practice to clustering mixed dataset is to trans- form categorical values into numeric values and then pro- ceed to use a numeric clustering algorithm. Another approach is to compare the categorical values directly, in which two distinct values result in distance 1 while identical values result in distance 0. Nevertheless, these two methods do not take into account the similarity information embed- ded between categorical values. Consequently, the cluster- ing results do not faithfully reveal the similarity structure of the dataset (Hsu, 2006; Hsu & Wang, 2005). This article is based on distance hierarchy (Hsu, 2006; Hsu & Wang, 2005) to propose a new incremental cluster- ing algorithm for mixed datasets, in which the similarity information embedded between categorical attribute is considered during clustering. In our setting, each attribute of the data is associated with a distance hierarchy, which is an extension of the concept hierarchy (Somlo & Adele, 2001) with link weights representing the distance between concepts. The distance between two mixed data patterns is then calculated according to distance hierarchies. It is worth mentioning that the representation scheme of distance hierarchy can generalize some conventional dis- tance computation schemes including the simple matching and the binary encoding for categorical values, and the subtraction method for continuous values and ordinal values. The rest of this article is organized as follows. Section 2 reviews clustering algorithms and discusses the shortcom- ings of the conventional approaches to clustering mixed data. Section 3 presents distance hierarchy for categorical data and proposes the incremental clustering algorithm based on distance hierarchies. In Section 4, experimental results on synthetic and real datasets are presented. Con- clusions are given in Section 5. 2. Literature review Adaptive resonance theory neural networks model real- time prediction, search, learning, and recognition. ART networks function as models of human cognitive informa- tion processing (Carpenter, 1997; Carpenter & Grossberg, 1993; Grossberg, 1980, 1999, 2003). A central feature of all ART systems is a pattern matching process that com- pares an external input with the internal memory of an active code. ART1 deals with the binary numerical data and ART2 deals with the general numerical data (Gross- berg, 1999). However, these two methods do not deal with mixed data attributes. About clustering mixed data attributes, there are two approaches for mixed data. One is resorted to a pre-process, which transferred the data to the same type, either all numeric or all categorical. For transferring continuous data to categorical data, some metric function is employed. The function is based on simple matching in which two distinct values result in distance 1, with identical values of distance 0 (Guha, Rastogi, & Shim, 1999). The other is to use a metric function, which can handle mixed data (Wilson & Martinez, 1997). Overlap metric is for nominal attributes and normal- ized Euclidean distance is for continuous attributes. Among problems with simple matching and binary encoding, a common approach for handling categorical data is simple matching, in which comparing two identical categorical values result in distance 0, while two distinct values result in distance 1 (Ester et al., 1998; Wilson & Martinez, 1997). In this case, the distance between patterns of Gary and John in the previous example becomes d(Gary, John) = 1, which is the same as d(John, Tom) = d(- Gary, Tom) = 1. Obviously, the simple matching approach disregards the similarity information embedded in categor- ical values. Another typical approach to handle categorical attri- butes is to employ binary encoding that transforms each categorical attribute to a set of binary attributes and a cat- egorical value is then encoded to a set of binary values. As a result, the new relation contains all numeric data, and the clustering is therefore conducted on the new dataset. For example, as the domain of the categorical attribute: Favor- ite_Drink. The set of it is {Coke, Pepsi, Mocca}. Favor- ite_Drink is transformed to three binary attributes: Coke, Pepsi and Mocca in the new relation. The value Coke of Favorite_Drink in a pattern is transformed to a set of three binary values in the new relation, i.e. {Coke = 1, Pepsi = 0, Mocca = 0}. The Manhattan distance of patterns Gary and John is dM(Gary, John) = 2, which is the same as dM(Gary, Tom) and dM(John, Tom), according to the new relation. Traditional clustering algorithm transfers Favor- ite_Drink categorical attributes into a binary numerical attribute type as shown in Fig. 1. After transformation, each original categorical attribute handled by the binary encoding approach contributes as twice as that by the simple matching approach, as shown in the above example of distance (Gary, John). Conse- quently, when the binary encoding approach is adopted by a clustering algorithm, categorical attributes have larger influence on clustering data than those adopting the simple matching approach. The ART network is a popular incremental clustering algorithm (Jain & Dubes, 1988). It has several variants ID Favorite Drink Amt. Gary Coke 70 John Pesi Tom Coffee ID Coke Pepsi Coffee Amt. Gary 0 0 70 John 0 1 0 70 Tom 0 0 1 70 70 70 1 Fig. 1. Traditional clustering algorithm transfers Favorite_Drink cate- gorical attributes into binary numerical attribute type. Any Refrig. Pepsi TV Y Z X Coke Drink Appliance 1 1 1 1 1 1 Any Refrig. Pepsi TV Coke .5 .5 .5 .5 MAX MIN w Fig. 2. (a) A distance hierarchy with weight 1, (b) two-level distance hierarchy for simple matching approach, and (c) degenerated distance hierarchy with w = (max � min) for a numeric attribute. C.-C. Hsu, Y.-P. Huang / Expert Systems with Applications 35 (2008) 1177–1185 1179 (Carpenter & Grossberg, 1987; Carpenter, Grossberg, & Rosen, 1991), in which ART1 handles only the binary data and ART2 can handle only the arbitrary continuous data. K-prototype (Huang, 1998) is a recent clustering algorithm for mixed data. It transfers categorical data attributes to the binary data format, however, the binary data do not reflect the similar degree. It influences the clustering qual- ity. Therefore, this paper proposes a modified adaptive res- onance theory network algorithm and the conceptual hierarchy tree to solve the similar degree of mixed data. 3. Clustering hybrid data based on distance hierarchy This paper proposes the distance hierarchy tree structure to overcome the expression for similar degree. This dis- tance hierarchy tree algorithm combines the adaptive reso- nance theory network algorithm and it can be effective with mixed data in data clustering. This section presents dis- tance hierarchy for categorical data and it proposes the incremental clustering algorithm based on distance hierarchies. 3.1. Distance hierarchy The distance hierarchy tree is a concept hierarchy struc- ture. It is also a better mechanism to facilitate the represen- tation and computation of the distance between categorical values. A concept hierarchy consists of two parts: a node set and a link set (Dash, Choi, Scheuermann, & Liu, 2002; Hsu, 2006; Hsu & Wang, 2005; Maulik & Bandyo- padhyay, 2002). According to binary encoding approach, it does not reflect the similar degree. However, it influences the clustering quality. Maintenance was difficult when the domain of a categorical attribute changes, because the transformed relation schema also needs to be changed. The transformed binary attributes cannot preserve the semantics of the original attribute. Because of the draw- backs resulting from the binary-encoding approach, this paper uses distance hierarchy to solve the similar degree of mixed data. A concept hierarchy extends with distance weights as shown in Fig. 2. This paper extends the distance hierarchy structure with link weights. Each link has a weight representing a dis- tance. Link weights are assigned by domain experts. There are several assignment alternatives. The simplest way is to assign all links as a uniform constant weight. Another alternative is to assign heavier weights to the links closer to the root and lighter weights to the links away from the root. For simplicity, unless stated explicitly, each link weight is set to 1 in this article. The distance of two con- cepts at the leaf nodes is the total weight between those two nodes. A point X in a distance hierarchy consists of two parts, an anchor and a positive real-value offset, denoted as X(N, d), that is, anchor(X) = N and offset(X) = d. The anchor is a leaf node and the offset represents the distance from the root of the hierarchy to the point. A point X is an ancestor of Y if X is in the path from Y to the root of the hierarchy. If neither one of the two points is an ancestor of the other point, then the least common ancestor, denoted as LCA(X, Y), is the deepest node that is an ancestor of X as well as Y. Let X(NX, dX) and Y(NY, dY) be two points, the distance between X and Y can be defined as jX � Y j ¼ d X þ d Y � 2d LCPðX ;YÞ ð1Þ where LCP(X, Y) is the least common point of X and Y in the distance hierarchy. dLCP(X,Y) is the distance between the least common point and the root. The distance between two equivalent points is 0, i.e. |X � Y| = 0. For example, W = (Coke, 1.4), X = (Coke, 0.3), Y = (Pepsi, 0.3) and Z = (Pepsi, 1.4). Both X and Y are ancestors of W and Z. X is equivalent to Y, and their distance is 0. Both X and Y are the least common points of these two points. They are the least common points of Y and W, as well as those of Y and Z. The distance between nodes Any and Drink is 1. The distance between Y and Z is |1.4 + 0.3 � 2 * 0.3| = 1.1. The least common point, LCP(W, Z), of W and Z is the node Drink, which is also the least common ancestor of these two points. Therefore, dLCP(W,Z) = 1. Furthermore, the distance between W and Z is |1.4 + 1.4 � 2 * 1| = 0.8. A special distance hierarchy calls numeric distance hier- archy for a numeric attribute, say xi, is a degenerate one, which consists of only two nodes, a root MIN and a leaf MAX (e.g. Fig. 2c), and has the link weight w being the domain range of xi, i.e. w = (maxi � mini). A point p in such a distance hierarchy has the value (MAX, dp) where the anchor is always the MAX and the offset dp is the dis- tance from the point to the root MIN. About measuring distance, the distance between two data points can be measured as follows: Let x = [x1 x2 . . . xn] and y = [y1 y2 . . . yn]. The distance between a Fig. 3. Modified adaptive neural network architecture. 1180 C.-C. Hsu, Y.-P. Huang / Expert Systems with Applications 35 (2008) 1177–1185 training pattern x and an M-ART neuron y is measured as the square root of the sum of the square differences between each-paired components of x and y. Specifically, x andy represent a training data and a map neuron, respec- tively, with n-dimension, and C is a set of n distance hier- archies, then the distance between x and y can be expressed as dðx; yÞ¼ kx � yk¼ X i¼1;n wiðxi � yiÞ 2 !1=2 ¼ X i¼1;n wiðhðxiÞ� hðyiÞÞ 2 !1=2 ð2Þ where h(xi) and h(yi) are the mapping of xi and yi to their associated distance hierarchy hi and wi, the attribute weight, is a user specified parameter allowing the domain expert to give different weights. For a numeric attribute I, h(xi) � h(yi) is equal to xi � yi, since h(xi) � h(yi) = (MIN, dh(xi)) � (MIN, dh(yi)) = (MIN, xi � mini) � (MIN, yi � mini) = (xi � yi). The attribute weight wi can be used to remedy the mixed depth effect of distance hierarchies, especially when all the attribute domains are normalized to a small range, such as [0 1]. For example, the difference between any two dis- tinct values of an attribute with a two-level distance hierar- chy, like Fig. 2b, is always 1 while the difference between two distinct values of a three-level distance hierarchy can be the 0.5 (minimum) or 1 (maximum). Therefore, in a mul- tidimensional data set with attributes associated with vari- ous depths of distance hierarchies, attributes with shallow distance hierarchies tend to dominate the distance compu- tation. Decreasing the weight of attributes with a shallow hierarchy or in contrast, increasing the weight of an attri- bute with a deep hierarchy can alleviate the mixed depth effect. 3.2. Incremental clustering of hybrid data The unsupervised clustering algorithm has two layers, the input layer and output layer. There is one distance tree. As shown in Fig. 3, each layer is explained as follows. The input layer is used for receiving the data or the vec- tor. In the input layer, each neuron corresponds to the input vector. An input vector can be a numerical attribute or a categorical attribute. The output layer presents the clustering result. Each neuron represents a clustering result. Each related prototype vector is the representative of a clustering vector. Particularly, supposes D = {1, 2, . . ., n} is the training dataset. Input neurons hx1, x2, . . ., xni receive the input data vector. A neuron xk represents an input vector, an attribute or a variable. The sets of distance tree are {dt1, dt2, . . ., dtn}, each dti expresses the training dataset’s attribute xi. It rep- resents the concept distance in the special and general relation. A neuron yi of the output layer is a vector. yi = [y1i, y2i, . . ., yni]. The neuron yki = (N, d) is composed of two parts: N is a symbol and d is a real number. When yki corresponds to a categorical attribute, N corresponds to a categorical attribute value. When yki corresponds to a numerical attribute, N corresponds to a numerical symbol. As the traditional self-organizing maps and adaptive neural network, the vector has two functions. One is the vector in the output layer and the other is mapped each other in the training dataset. Each dti in the distance tree relates to the attribute xi in the training data and the neu- ron attribute yik in the output layer. The algorithm has the following steps: Input: N records in the training datasets, which include distance hierarchy sets, warning value and stopping parameter value. Output: Z neurons (the prototype of clustering). Step 1. Read the first record and map it into the first neuron’s vector. Step 2. Read the next record until the record is empty. Find out the most similar output neuron. If similar degrees exceed the warning value, then join this record to the group and adjust the vector else set up a new neu- ron. And set the vector as a new neuron’s vector. Step 3. Meet the stopping conditions and then stop else repeat Step 2 until the record is empty. 3.3. Evaluating clustering results About time complexity, the main decisive factors of the complexity of time in the training algorithm are the train- ing data record N, data dimension P, output neuron num- ber Z and train round C. For each round in the training data, the vectors compare all the attributes with the output neurons. So the time complexity of the whole process is O(C * N * Z * P). The other method is the significance test on external variables. This technique compares the clusters on vari- ables, but it does not generate. One way of doing this is to compute the expected entropy of the clusters using a var- iable, say a class attribute C that does not participate in the C.-C. Hsu, Y.-P. Huang / Expert Systems with Applications 35 (2008) 1177–1185 1181 clustering. The expected entropy of an attribute C in a set of clusters can be computed as follows. First, it calculates the entropy of an attribute C in each cluster. Then it sum- mates all the entropies weighted by its cluster size. The rela- tionship is: E �ðCÞ¼� X k jCkj jDj X j PðC ¼ V jÞ log PðC ¼ V jÞ ! ð3Þ where Vj denotes one of the possible values that the attri- bute C can take, |Ck| is the size of the cluster k, and |D| is the size of the data set. The categorical utility function (Gluck & Corter, 1985) attempts to maximize the probability that the two objects in the same cluster have attribute values in common and the probability that the objects from different clusters have different attributes. The categorical utility of a set of clus- ters can be calculated as CU ¼ X k jCkj jDj X i X j ½PðAi ¼ V ijjCkÞ 2 � PðAi ¼ V ijÞ 2� ! ð4Þ Here, P(Ai = Vij|Ck) is the conditional probability that the attribute i has the values Vij given the cluster Ck, and P(Ai = Vij) is the overall probability of the attribute i hav- ing the values Vij in the entire set. The function aims to measure if the clustering improves the likelihood of similar values falling in the same cluster. Obviously, the higher the CU values, the better the clustering result (Barbara, Couto, & Li, 2002). For numeric attributes, the standard deviation repre- sents the dispersion of values. Variance (r2) can be used for evaluating the quality of clustering numeric data. Here, it can sum up the respective variance of every numeric attri- bute in all the clusters to evaluate the quality of clustering. The method of calculating the variance is shown in Eq. (5), where V ki;avg and V k i;j are the average and the jth record value of attribute i in cluster k, respectively. Incidentally, the attribute values have been normalized before clustering. Apparently, the lower the variance values, the better the clustering results: r2 ¼ X k 1 jCkj X i X j ðV ki;j � V k i;avgÞ 2 ð5Þ Several cluster validity indices, such as Davies–Bouldin (DB) Index and Calinski Harabasz (CH) Index (Halkidi, Table 1 Synthetic dataset Sex Age Amt Depart F(50%) M(50%) N(20, 2) 70–90 EE ME ID VC MBA MIS Batistakis, & Vazirgiannis, 2001; Maulik & Bandyopadhy- ay, 2002), have been published; however, they are only suit- able for the numeric data. Hence, in order to evaluate the effectiveness of clustering mixed data, this paper uses CV index (Hsu & Chen, 2007), which combined the category utility (CU) function with variance. The CV is defined as in Eq. (6), where the CU and variance are the validity index for categorical and numeric data, respectively. The higher the CV values, the better the clustering result: CV ¼ CU 1 þ Variance ð6Þ 4. Experiments and discussion This paper develops a prototype system with Borland C++ Builder 6. A series of experiments have been per- formed in order to verify the method. A mixed synthetic dataset and a UCI dataset have also been designed to show the capability of the M-ART in reasonably expressing and faithfully preserving the distance between the categorical data. It also reports the experimental results of artificial and actual data. 4.1. Synthetic data sets The synthetic dataset consisted of three categorical attri- butes: Sex, Department and Product. The two numeric attributes are Age and Amt. One class level attribute is Col- lege. The total records are 600. Table 1 shows all distribu- tions of synthetic categorical dataset. Fig. 4 shows the concept hierarchies for the synthetic categorical dataset. The M-ART parameters are established as follows: The initial warning value is 0.5 and it increases progressively by 0.05 until 0.7. The initial learning rate is 0.4–0.6. The stop condition is occurs when the momentum of the output layer is lower than 0.000015. The input sequence influences the performance of M-ART and ART2 algorithms. There- fore M-ART uses five different introductory orders. The experimental results are shown in Table 2. In the K-proto- type method, the initial central points are influenced by clustering. There are five experiments by randomly choos- ing one central point. Each parameter in ART2 algorithm is established a = 0.1, b = 0.1, c = 0.1, d = 0.9, theta = 0.1– 0.2, rho = 0.755–0.76, bottom to top weight = 2. The experimental result is shown in Table 2. The entropy values can be used for evaluating the quality of ment Product College Count Coke Engineering 100 Pepsi 100 Bread Design 100 Rice 100 Apple Management 100 Orange 100 Any Grain Drink Fruit Rice Bread Coke Pepsi Orange Apple Engineering Management ID VC EE ME MIS MBA Any Desing M F Any Fig. 4. The concept hierarchies for the synthetic datasets. Table 2 Entropy values for Student dataset with clusters by ART2, K-prototype and M-ART Order ART2 K-prototype M-ART 1 0.47 0.44 0 2 0.42 0.38 0 3 0.43 0.38 0 4 0.42 0.37 0 5 0.42 0.39 0 Mean 0.43 0.39 0 1182 C.-C. Hsu, Y.-P. Huang / Expert Systems with Applications 35 (2008) 1177–1185 clustering. The value is small and the quality of clustering is good. In these five experiments, entropy values are zero in M-ART. That means that the classification value in each group is the same. That is to say, the datasets are divided into three groups such as Engineering, Design, and Man- agement. The experimental results show that, in the same input order, the quality of clustering M-ART method is the highest, K-prototype method is second, and traditional ART2 is the lowest. ART2 and K-prototype are unable to divide the datasets into three groups. 4.2. UCI adult data These experiments use a real Adult dataset from the UCI repository [Murphy 1992] with 48,842 records of 15 attributes, including eight categorical attributes, six numer- ical attributes, and one class attribute. This experiment uses seven attributes, which include three categorical attributes, such as Relationship, Mari- tal_status, and Education; and four numeric attributes, Capital_gain, Capital_loss, Age, and Hours_per_week. The concept hierarchies are constructed in Fig. 5. The M arried_A F _spouse W idow ed S eparated N ever_m arried M arried_spouse_absen t M arried_civ_spouse D ivorced Single Couple Any P reschool 10th 12th 11th 7th_8th 9th 5th_6th 1st_4th HighSchoolLittle Junior Any Fig. 5. Concept hierarchies for (a) Marital-status, (b) Educa M-ART parameters are established as follows: the initial warning value is 0.55 and it increases progressively 0.05 until 0.75. The initial learning rate is 0.9. The stop condi- tion t occurs when the momentum of the output layer is lower than 0.000015. 4.3. Experiments and discussion This paper collects the dataset with different methods. These methods divide adult datasets into 5, 6, 7 and 8 groups. Concerning the CU values for categorical attri- butes, the higher the CU values, the better the clustering result. The CU value of clustering M-ART method is the highest, K-prototype method is second, and traditional ART2 is the lowest. The symbol ‘‘***’’ means that it does not find the suitable parameter to divide into group with the datasets. The parameter of ART2 reaches 7, it is unable to divide seven groups all the time. The problem occurs because there are too many parameters in ART2. Table 3 shows the CU values of the clustering results by M-ART, ART2 and K-prototypes of each categorical attribute on level 1 and the leaf level in individual concept hierarchies with cluster numbers 5, 6, 7 and 8. The parameter of M- ART is established as follows: the initial warning value is 0.53–0.58, and the initial learning rate is 0.7. This paper normalizes the variance between 0 and 1 for numeric results. The normalized variance is useful in CV index. Table 4 shows the CV values of the clustering results by M-ART, ART2 and k-prototypes on level 1 and the leaf level in individual concept hierarchies with cluster numbers 5, 6, 7 and 8. The higher the value of CV values, the better the clustering result. The CV value in M-ART method is M aster P rof_school H S _grad A ssoc_acdm B achclors S om e_college D octorate A ssoc_voc College Advanced H usband W ife U nm arried O w n_child O ther_relative N ot_in_fam ily Any tion and (c) Relationship attributes of the Adult dataset. Table 3 The CU values for Adult dataset with 5, 6, 7 and 8 clusters by M-ART, ART2 and K-prototypes Clusters Leaf_Level Level 1 Increased (%) M-ART 5 1.069 1.16 8.5 6 1.113 1.20 7.8 7 1.115 1.21 8.5 8 1.177 1.31 11.3 K-prototype 5 0.859 0.834 �2.9 6 1.002 0.977 �2.5 7 1.039 0.919 �11.6 8 1.088 1.087 �0.3 ART2 5 0.001 0.00081 �19 6 0.0043 0.0057 3.3 7 *** *** *** 8 0.0073 0.00757 3.7 C.-C. Hsu, Y.-P. Huang / Expert Systems with Applications 35 (2008) 1177–1185 1183 the highest, K-prototype method is second, and traditional ART2 is the lowest. After M-ART clustering, each clustering has a proto- type vector to represent the characteristic of the group. Table 5 shows the result. There is a binary set in a proto- type vector. The binary set includes the anchor and Offset. The anchor shows the mode of this field in this group. The Offset is the distance vector to the root. The offset means that a rough proportion of response is accounted for all number values. The higher the value of Offset values, the bigger the mode proportion of the anchor result. For exam- ple the fourth group of anchors is HS-grad. The Offset value is 0.84 (Offset value regularized between 0 and 1). The proportion of HS-grad is 73%. Further, the anchor of 2, 3, 6 and 8 groups is all HS-grad. The Offset value approaches zero. It expresses the HS-grad node near the root. Table 6 shows the result. As an example in fifth group with Marital-status and Relationship, the Offset value is quite big. The populations of Married-civ-spouse and Hus- Table 4 The CV values for Adult dataset with 5, 6, 7, 8 clusters by M-ART, ART2 an Clustering number CU Variance Leaf_Level Level 1 Age Capital gain Cap M-ART 5 1.069 1.16 0.127 0.022 0.0 6 1.113 1.2 0.163 0.023 0.0 7 1.115 1.21 0.182 0.011 0.0 8 1.177 1.31 0.218 0.014 0.0 K-prototype 5 0.859 0.834 0.114 0.033 0.0 6 1.002 0.977 0.424 0.72 1.4 7 1.039 0.919 0.515 0.893 1.8 8 1.088 1.087 0.61 1.068 2.1 ART2 5 0.001 0.00081 0.176 0.027 0.0 6 0.0043 0.0057 0.211 0.033 0.0 8 0.0073 0.00757 0.281 0.044 0.0 band of this group are 99% and 89% separately. The result shows each prototype vector can reflect characteristics of each group. For example, the fourth group of anchors is HS-grad. The Offset value is 0.84 (Offset value is regularized between 0 and 1). The proportion of HS-grad is 73%. Further, the anchor of 2, 3, 6 and 8 groups is all HS-grad. The Offset value approaches zero. It expresses the HS-grad node near the root. Table 6 shows the result. As an example in the fifth group with Marital-status and Relationship, the Offset value is quite big. The populations of Married-civ-spouse and Husband of this group are 99% and 89%, respectively. The result shows that each prototype vector can reflect the characteristics of each group. The prototype vector reflects the characteristic in this group. A comparison with these prototype vectors and Sal- ary attribute of each group can obtain some characteristics about Salary attribute. Table 5 shows that the salary was over 50k in an order. In the fifth group, all values in the Salary attribute is over 50k. In the first group, the popula- tion is 1%. From these prototype vectors, the characteris- tics in the high-income group can be understood. The salary is over 50k; the proportion is big in 2, 3 and 5 groups. The characteristics are steady marriage states and family relationships, the age is generally relatively longer, and the working hours are relatively longer. The value of Gain and Loss are relatively bigger. Contrarily, the propor- tion is small in the 1, 4, 6 and 8 groups. The characteristic is not married or had divorced, the age is younger, and work- ing hours is relatively lower. The value of Gain and Loss are relatively small. It can find out the obvious differences in groups 1 and 5. The comparison with K-prototype and M-ART, K-pro- totype can present the prototype vector of each group. In the categorical attribute, K-prototype is a representation of the mode as the prototype vector but unable to express the weight of this group. The M-ART algorithm uses the Offset value to understand rough population. d K-prototypes CV Increased (%) ital_loss Hours_per_week Leaf_Level Level 1 38 0.071 0.85 0.59 31.08 44 0.085 0.85 0.60 29.05 44 0.1 0.83 0.61 27.05 52 0.115 0.84 0.65 23.00 46 0.073 0.68 0.46 32.85 54 2.937 0.15 0.16 1.72 01 3.635 0.13 0.12 7.01 54 4.342 0.12 0.13 5.56 42 0.079 0.001 0.001 6.64 51 0.095 0.003 0.005 55.87 68 0.128 0.005 0.006 26.71 Table 5 The categorical attribute prototype for Adult dataset with 1–8 clusters by M-ART C_no >50k Education Marital-status Relationship Age Hours Gain Loss 5 100 (Prof-school, 0.34) (Married-civ-spouse, 0.99) (Husband, 0.89) 48 51 99,999 0 2 46 (HS-grad, 0.03) (Married-civ-spouse, 0.99) (Wife, 0.98) 40 37 1000 131 3 44 (HS-grad, 0) (Married-civ-spouse, 1) (Husband, 1) 44 44 1000 131 7 13 (Some-college, 0.5) (Never-married, 0.61) (Not-in-family, 1) 38 41 1000 87 4 7 (HS-grad, 0.84) (Never-married, 0.5) (Not-in-family, 1) 40 40 0 87 6 6 (HS-grad, 0) (Divorced, 0.5) (Unmarried, 1) 40 39 0 44 8 3 (HS-grad, 0.05) (Never-married, 0.5) (Other-relative, 1) 33 37 0 44 1 1 (Some-college, 0) (Never-married, 0.38) (Own-child, 1) 25 33 0 44 Table 6 The distribution of Education attribute in each cluster Education C5 C2 C3 C7 C4 C6 C8 C1 Preschool 0 0 0 0 0 0 0 0 1st–4th 0 0 1 0 1 1 2 0 5th–6th 1 1 1 1 1 1 4 0 7th–8th 0 1 3 1 2 2 3 1 9th 0 1 2 1 2 2 3 1 10th 1 2 2 0 6 4 4 4 11th 0 2 2 0 6 4 6 9 12th 0 1 1 0 2 1 3 3 HS-grad 12 31 33 0 73 38 40 30 Some-college 6 20 19 37 0 22 20 34 Assoc-voc 2 5 5 7 0 5 3 3 Assoc-acdm 0 5 3 7 0 4 2 3 Bachelors 24 20 18 34 0 10 9 10 Masters 13 8 7 2 5 4 0 1 Prof-school 33 2 3 8 1 1 1 0 Doctorate 8 1 2 1 1 1 0 0 1184 C.-C. Hsu, Y.-P. Huang / Expert Systems with Applications 35 (2008) 1177–1185 In the C7 prototype vector, the value of Education field is (HS-grad, 0.84): Anchor is HS-grad, and the Offset is 0.84. It expresses that the leaf node in the hierarchy tree is near the root node and the distance is 0.84. The distance between leaf node and HS-grad node is 0.16. It shows that HS-grad has taken heavy proportion in this group. In Table 6, the heavy proportion is 73%. Otherwise, the Offset value of Relationship field is about 1. It means that the mode proportions in each group are between 96% and 100%. 5. Conclusions Most traditional clustering algorithms can only handle either categorical or numeric value. Although some research results have been published for handling mixed data, they still cannot reasonably express the similarities among categorical data. The paper presents a MART algo- rithm, which can handle mixed dataset directly. The exper- imental results on synthetic data sets show that the proposed approach can better reveal the similarity struc- ture among data, particularly when categorical attributes are involved and have different degrees of similarity, in which the traditional clustering approaches do not perform well. The experimental results on the real dataset have bet- ter performances than other algorithms. The future work will try to use this method in finding out the pattern rules from a large time series database. References Anderberg, M. (1973). Cluster analysis for applications. New York: Academic Press. Barbara, D., Couto, J., & Li, Y. (2002). COOLCAT: An entropy-based algorithm for categorical clustering. In Proceedings of the 11th international conference on information and knowledge management (pp. 582–589). Can, F. (1993). Incremental clustering for dynamic information process- ing. ACM Transaction for Information Systems, 11, 143–164. Carpenter, G.-A. (1997). Distributed learning, recognition, and prediction by ART and ARTMAP neural networks. Neural Networks, 10(8), 1473–1494. Carpenter, G.-A., & Grossberg, S. (1987). ART2: Self-organization of stable category recognition codes for analog input patterns. Applied Optics: Special Issue on Neural Networks, 26, 4919–4930. Carpenter, G.-A., & Grossberg, S. (1993). Normal and amnesic learning, recognition, and memory by a neural model of cortico-hippocampal interactions. Trends in Neuroscience, 16(4), 131–137. Carpenter, G., Grossberg, A.-S., & Rosen, D.-B. (1991). Fuzzy ART: Fast stable learning and categorization of analog patterns by an adaptive resonance system. Neural Networks, 4, 759–771. Dash, M., and Choi, K., Scheuermann, P., and Liu, H., (2002). Feature selection for clustering – a filter solution. In IEEE International Conference on Data Mining, pp. 115–122. Ester, M., Kriegel, H.-P., Sander, J., Wimmer, M., & Xu, X. (1998). Incremental clustering for mining in a data warehousing environment. In Proceedings of the 24th international conference on very large data bases (VLDB) (pp. 323–333). Gluck, M.-A., & Corter, J.-E. (1985). Information, uncertainty, and the utility of categories. In Proceedings of the seventh annual conference of the cognitive science society. Grossberg, S. (1980). How does a brain build a cognitive code? Psychological Review, 87, 1–51. Grossberg, S. (1999). The link between brain, learning, attention, and consciousness. Consciousness and Cognition, 8, 1–44. Grossberg, S. (2003). How does the cerebral cortex work? Development, learning, attention, and 3D vision by laminar circuits of visual cortex. Behavioral and Cognitive Neuroscience Reviews, 2(1), 47–76. Guha, S., Rastogi, R., & Shim, K. (1999). ROCK: A robust clustering algorithm for categorical attributes. In Proceedings of the IEEE conference on data engineering (pp. 512–521). Halkidi, M., Batistakis, Y., & Vazirgiannis, M. (2001). On clustering validation techniques. Journal of Intelligent Information Systems, 17, 107–145. Hsu, C.-C., & Chen, Y.-C. (2007). Mining of mixed data with application to catalog marketing. Expert Systems with Applications, 32(1), 12–23. Hsu, C.-C. (2006). Generalizing self-organizing map for categorical data. IEEE Transactions on Neural Networks, 17(2), 294–304. C.-C. Hsu, Y.-P. Huang / Expert Systems with Applications 35 (2008) 1177–1185 1185 Hsu, C.-C., & Wang, S.-H. (2005). An integrated framework for visualized and exploratory pattern discovery in mixed data. IEEE Transactions on Knowledge and Data Engineering, 18(2), 161–173. Huang, Z. (1998). Extensions to the k-means algorithm for clustering large data sets with categorical values. Data Mining and Knowledge Discovery, 2(3), 283–304. Jain, A., & Dubes, R. (1988). Algorithms for clustering Data. Englewood Cliffs, NJ: Prentice-Hall. Jain, A.-K., Murty, M.-N., & Flynn, P.-J. (1999). Data clustering: A review. ACM Computing Surveys, 31(3), 264–323. Maulik, U., & Bandyopadhyay, S. (2002). Performance evaluation of some clustering algorithms and validity indices. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(12), 1650–1654. Rasmussen, E. (1992). Clustering algorithms. In William B. Frakes & Ricardo Baeza-Yates (Eds.), Information retrieval: Data structures & algorithms. Prentice Hall. Salton, G. & Buckley, C. (1991). Automatic text structuring and retrieval- experiments in automatic encyclopedia searching. In Proceedings of the 14th international ACM SIGIR conference on research and development in information retrieval (pp. 21–30). Somlo, G.-L. & Adele, E.-H. (2001). Incremental clustering for profile maintenance in information gathering web agents. In Proceedings of the fifth international conference on autonomous agents (pp. 262–269). Wilson, D.-R., & Martinez, T.-R. (1997). Improved heterogeneous distance functions. Journal of Artificial Intelligence Research, 6, 1–34. Incremental clustering of mixed data based on distance hierarchy Introduction Literature review Clustering hybrid data based on distance hierarchy Distance hierarchy Incremental clustering of hybrid data Evaluating clustering results Experiments and discussion Synthetic data sets UCI adult data Experiments and discussion Conclusions References 
work_vesmay3hzrdnlde2lq3iqj4odu	----	Introduction An Expert System for the Composition of Formal Spanish Poetry Pablo Gervás Universidad Europea - CEES, Madrid pg2@dinar.esi.uem.es www.uem.esi.es/~pg2 Abstract: The present paper presents an application that composes formal poetry in Spanish in a semiautomatic interactive fashion. ASPERA is a forward reasoning rule- based system that obtains from the user basic style parameters and an intended message; applies a knowledge-based preprocessor to select the most appropriate metric structure for the user's wishes; and, by intelligent adaptation of selected examples from a corpus of verses, carries out a prose-to- poetry translation of the given message. In the composition process, ASPERA combines natural language generation and CBR techniques to apply a set of construction heuristics obtained from formal literature on Spanish poetry. If the user validates the poem draft presented by the system, the resulting verses are analysed and incorporated into the system data files. 1. Introduction The automatic generation of text is a well established and promising problem in AI, with numerous practical applications waiting in the sidelines for efficient and acceptable solutions. Existing systems have shown reasonable results in restricted domains [3,7,9], opening the way to considering how more elaborate texts - from the point of view of aesthetics and reader satisfaction - can be obtained [2,4,8]. The composition of poetry ranks among the most challenging problems of language generation, and is therefore a good testbed for techniques designed to improve the quality of generated texts. ASPERA (Automatic Spanish Poetry Expert and Rewriting Application) is a prose-to-poetry semiautomatic translator. By ingenious use of well accepted AI techniques (Natural Language Processing, Case Based Reasoning, Knowledge Based Systems), the application obtains from the user a prose description of the intended message and a rough specification of the type of poem required (length, mood, topic); selects appropriate metre and stanza (by resorting to a knowledge base on literary style); generates a draft of the poem (by applying CBR techniques to a database of previous poems); requests modification or validation of the draft by the user; and updates its own database of information (using NLP techniques to extract all the required linguistic information from the validated poem). ASPERA is designed to be used as a teaching aid for specific information technology subjects of the degree on Translation and Interpretation at the Universidad Europea - CEES. 2. Problem Description On first acquaintance, the generation of poetry involves advanced linguistics skills and common sense, two of the major challenges that face AI in general. On the positive side, poetry has the advantage of not requiring exaggerate precision. To a certain extent, imposing restrictions over the form of a poem implies a slight relaxation on the specification of the content. Under this assumption, the general problem becomes tractable. There are three main challenges to be faced: 1. a specification of the formal requirements that define a correct poem under classical literary rules (in a format that can be used to drive the construction process) 2. appropriate management of an extensive vocabulary (the choice of words plays an important role in determining the quality of a poem) 3. correct combination of words in the poem to match both the intended message and the chosen metric structure 2.1 The Formal Rules of Spanish Poetry The fact that the application was to be developed in Spanish presents important advantages over other languages. The phonetics of Spanish are quite straightforward to obtain from the written word. Most letters in Spanish sound the same wherever they appear in a piece of text, so the metrics, or the syllabic division, of a verse can be worked out algorithmically [11]. Spanish scholars have a love for rules, and there is a good set of formal rules [10] describing the conditions that a poem must fulfil in order to be acceptable. The challenge becomes a simple problem of transforming the given evaluation rules (designed to be applied to an existing poem in order to ascertain its acceptability) into the corresponding construction rules. Given that words are divided into syllables and each word has a unique syllable that carries the prosodic stress, the constraints that the rules have to account for are the following: 1. Specific strophic forms require different number of syllables to a line 2. Syllables may be run together under specific conditions - this is a form of poetic license called synaloepha - thereby shortening the syllable count of the line involved 3. The position of the stressed syllable of the last word of a line affects the metric count for that line (adding or subtracting one syllable to the count) 4. Not all possible stress patterns for a line are valid (depending on the length) 5. The rhyme of a word is made up of the last half of its stressed syllable and all following syllables to the end of the word, and each strophic form requires a different rhyming pattern A poem may be an unstructured sequence of lines, but for the purpose of this application only poems of regular strophic forms are considered. For the purposes of the present application, the following strophic forms are relevant: 1. romances, a stanza of several lines of eight syllables where all even numbered lines rhyme together 2. cuartetos, a stanza of four lines of eleven syllables where the two outer lines rhyme together and the two inner lines rhyme together 3. tercetos encadenados, a longer poem made up of stanzas of three lines of eleven syllables linked together by their rhyme in a simple chain pattern ABA BCB CDC... These three types have been chosen because each shows a different structural characteristic. Type 1 presents a recurring rhyme that flows all along the poem. It uses only one rhyme, so many words that rhyme together are required for an acceptable result (our starting data have proven to be poor choices in this respect). Type 2 presents a very simple but rigid structure. It stands for the simplest possible stanza with enough complexity to be distinguishable from prose (the simplification employed with respect to syntax/semantics makes it difficult for shorter poems to sound acceptable). Type 3 presents a simple structure that recurs throughout the poem, but with the rhyme changing slowly as it moves down. A constructive implementation of these rules was developed for the WASP system [4]. WASP was a forward reasoning rule-based system that used a set of construction heuristics obtained from these constraints to produce a poem from a set of words and a set of line patterns provided by the user. The system followed a generate and test method by randomly producing word sequences that met the formal requirements. Output results were impeccable from the point of view of formal metrics, but they were clumsy from a linguistic point of view and made little sense. An improved version of the construction strategies developed in WASP is the starting point of the generating module of ASPERA. 2.2 Guiding Word Choice The set of words available as building blocks for the composition process play a crucial role in determining the quality of the results. Too narrow a choice of vocabulary can lead to formal requirements not being met. Too wide a choice can seriously reduce the efficiency of the system, leading to lengthy searches over vast amounts of words. A trade off must be found: enough words to compose the poem must be available, but they must all be words with a reasonable chance of finding their way into the final result. This is a part of the problem where intelligent solutions have a good chance of outperforming purely computational techniques. The ASPID system [5] provided specific algorithms for the selection of a working set of words from an initial vocabulary using methods based on similarity calculations between the message proposed by the user for his poem and a corpus of already validated verses. Based on the similarity calculations, the system established a set of priorities over the complete available vocabulary. The next word to be added to the poem draft was initially looked for only among words marked with the highest priority, with the search extending in subsequent steps to words of lower priority only if none had been found in the previous step. This procedure improved search times considerably and it made possible computations with wider vocabulary coverage and narrower constraints on strophic forms. However, above a certain threshold (of vocabulary size and/or number of constraints imposed on the poem) even the method of establishing a priority ordering on the available words failed to ensure successful termination. The ASPID method of priority assignment is retained in ASPERA, but a prior step of knowledge-based pre-selection of the vocabulary based on user requirements has been added. 2.3 Fitting Words to Message and Metric Structure There are several restrictions on the actual process of composition that must be observed. The length of a poem is given by the number of lines (and the length of the lines) in the chosen strophic form. The intended message must be shortened or extended to adjust it to the length of the poem. The basic unit for poem composition is not the sentence but the line. A poem may contain one or several sentences, but it is in its subdivision into lines that the constraints (position of stressed syllables, and rhyme) are imposed. A step of planning is required to distribute the contents of the intended message over the required number of lines. Words at the end of lines may have additional constraints imposed on them (rhyme) by the chosen strophic form. These restrictions must be taken into account when planning the poem. The words in a poem must be combined according to the syntax of the language in question, and must make sense according to their semantics. There are two alternative ways of achieving this: to provide the system with adequately rich lexicon, syntax and semantics for the language involved (as done in [8]) for English poetry), or to develop engineering solutions that achieve equivalent results without attempting to model the imposing complexity of the human language faculty. The former solution requires powerful tools capable of representing and manipulating the syntax and semantics of language. For the present application, the latter approach is preferred. ASPERA resorts to a radical simplification of the linguistic skills underlying poem composition. The exhaustive knowledge approach is abandoned in favour of a heuristic engineering solution. Only the barest outline of a grammatical outline is provided (in the form of a line pattern) to ensure syntactic correctness. Semantic correctness is not enforced strictly, rather an approximate result is obtained by intelligent pre-selection of the choice of words together with a memory based approach to word combination. The system is provided with a corpus of already validated verses. A Case Based Reasoning approach [1] is applied to this corpus in order to generate new verses. The words in these verses are marked with their corresponding part-of-speech (POS) tag. A line pattern is generated from these POS tags. Each line pattern is a set of tags, and each tag acts as place keeper for a possible word of the line. The tag is actually a string that represents information about part of speech, number, and gender of the word that would stand in that particular place in the pattern. Patterns act as seed for lines, therefore a pattern determines the number of words in a line, the particular fragment of sentence that makes up the line, and the set of words that can be considered as candidates for the line. By following this heuristic shortcut, the system is able to generate verses with no knowledge about grammar or meaning. 3. ASPERA: Automatic Spanish Poetry Expert - a Rewriting Application ASPERA is a forward reasoning rule-based system that performs the following sequence of operations: 1. interacts with the user to obtain a specification of the desired poem (intended message, mood, setting); 2. searches its knowledge base to find the most appropriate strophic form to match that specification; 3. plans a poem draft by distributing the intended message over the chosen strophic form; 4. pre-selects and loads a task-specific vocabulary and corpus of verse examples for each fragment of the poem by adequately combining its data files (CBR Retrieve step I); 5. generates each of the lines of the poem draft by mirroring the POS structure of one (CBR Retrieve step II) of the pre-selected line examples combined with the words in the pre-selected vocabulary (CBR Reuse step); 6. presents the draft to be validated or corrected by the user (CBR Revise step); and 7. carries out a linguistic analysis of any validated poems in order to add the corresponding information to its data files to be used in subsequent computations (CBR Retain step). ASPERA is written in CLIPS, a rule-based system shell developed by NASA. Earlier implementations were attempted using Prolog, but the nature of the line generation part of the problem, being a constructive combination of a set of elementary ingredients towards the achievement of a single whole which is the final poem, lends itself more easily to a forward-reasoning mode of operation. The system's poetry expert knowledge base had originally been coded in Prolog, but translation onto the new paradigm was presented no special problems. 3.1 Collection of Poem Specifications The system interacts with the user to obtain specifications of the desired poem. The user is first asked a set of questions designed to ascertain a few elementary facts about his intentions. These will be used by the knowledge-based module of the system to select the right vocabulary and an adequate set of verse examples from the available corpus. At present, the system is designed to operate with the following parameters: • approximate length of the poem: The user has to provide a numerical value for the number of lines he would like to obtain. The system uses this number in working out an adequate strophic form (or combination of strophic forms) for the poem. The user is also asked whether the stated length is a rigid or flexible constraint. • rhyme structure: Some strophic forms have rigid rhyme structures (there are constraints on the rhyme of every line) and others are more flexible (there are constraints on only some of the lines). • degree of formality: Certain strophic forms are more formal than others. This distinction plays an important role in selecting the appropriate form. • setting: The system provides a choice between an urban setting or a rural setting. This is a very restricted choice, but it may be extended at later stages. • mood: As a first approximation, mood is interpreted either as positive or negative. Length of poem and degree of formality are used specially to determine the strophic form. Setting and mood are used to select the correct vocabulary. Once the basic parameters have been set, the user is asked to provide a prose paraphrase of this intended message. This paraphrase need not be grammatically sound, and may range from a full explicit text to a set of keywords that the user would like to show up in the poem. The intended message is stored as a list of words in a format amenable for later use by the system: (message Peter loves Mary they go together to the beach) Every word in the message is considered a good candidate to contribute to the final poem. Statistical methods of natural language processing applied in the field of information retrieval favour an initial trimming of user queries with a view to retaining for processing only terms that have relevance for the retrieval task. Stop lists are applied in order to eliminate empty words (pronouns, articles, prepositions...). For the present purposes, however, the presence of words of these types allows the user to control the style and appearance of the final result even more (or at least as much as) nouns, verbs and adverbs. 3.2 The Poetry Expert Knowledge Base The Poetry Expert knowledge base contains two distinct sets of rules that encode specific knowledge concerning the relationship between strophic forms, the user's wishes, and the available data file. Their job is to select the best available combination of strophic form and data files to suit the user's request. The first set of rules concerns the choice of strophic form and the selection of data files with the required sets of verse examples/patterns. They have been obtained by a knowledge acquisition process performed systematically over the author's personal experience, careful analysis of classical Spanish poetry, and available references on poetic analysis [10]. This gave rise to a complex set of rules that relate possible combinations of input parameters with possible outputs (strophic forms or combinations of strophic forms). The relationship is mostly associative, but requires a step of arithmetic calculation when processing the required length. The decisions embodied in the rules follow roughly the following guidelines: • formal poem → strophic forms of 11 syllables. • informal → strophic forms of 8 syllables. • long and flexible → romance or terceto encadenado • short and concise → terceto • fully rhymed → cuarteto or terceto encadenado • loosely rhymed → romance (long) or terceto (short) • many similar rhymes → dos cuartetos (not too long) or romance (long) The second set of rules associates the information obtained from the user regarding setting and mood desired for the poem with the various data files over which the available vocabulary is distributed. The application of the Poetry Expert knowledge base results in a message to the user proposing a specific strophic form. If the user accepts the proposal, the corresponding data files are loaded and the system moves onto the next stage. If the user rejects the proposal, the system allows the users choice of strophic form to override its own suggestion. The knowledge base is then addressed for the right combination of data files and these are loaded. Mismatching specifications of poem length or degree of formality and choice of strophic form may lead to non termination. 3.3 The Elementary Planning Step Given the user's choice of basic parameters the system uploads an appropriate combination of data files containing vocabulary extracts. A vocabulary entry for a word is a CLIPS fact of the form: (word (cual desde_n) (numsil 2) (accent 2) (emp 0) (term 0) (cat NCMS) (rima en) (level nil)) An entry must provide all the necessary information for imposing the metric restrictions: length of the word in syllables (numsil), position of stresses syllable (accent), whether the word starts (emp) or ends (term) with a vowel, the POS tag for the word (cat), and the rhyme (rima). An additional level field is provided to enable priority marking of the words during later processing. For the resulting choice of strophic form specific line patterns /examples are loaded. A line pattern/example is a CLIPS fact of the form: ((sample (n 72) (patt NCMS PDEL NCMS NCMS DET PPO3FS NCFS) (words desde_n del cielo error de la ventura) (beg 0) (end 1) (level nil)) where n is a unique identifier for the specific sample, patt is the set of POS tag corresponding to the sample - which acts as line pattern -, words is the actual line example, beg encodes information about whether such line starts a sentence, end encodes information about whether the line ends a sentence, and the level field is used for priority settings. The words in the intended message provided by the user are distributed into as many fragments as there are lines in the chosen strophic form, retaining within each fragment the original order of appearance. For instance, supposing the message presented above were to be rewritten as a terceto, it would be split by the system into the following facts: (fragment 1 Peter loves Mary) (fragment 2 they go together) (fragment 3 to the beach) This is not considered a draft of the poem, but only as an approximate distribution of information over lines of the poem. Figure 1 Basic structure of the ASPERA system A module of the system works out the similarities between the line fragments and the available line examples/patterns. Each fragment/pattern pair is indexed with a numerical value indicating their similarity. In the initial approximation, this similarity value is worked out as the number of POS tags they have in common. The results of these similarity calculations are later used during the CBR Retrieve step. Another module counts the number of rhymes available (if any) for each of the pre- selected words. These numbers are later used to assign specific words and patterns to each fragment during the CBR Retrieve step: a pattern is only good for a given line in the poem if there is a word of the necessary rhyme whose POS matches the last POS tag in the pattern. The CBR Retrieve step can be considered to start during the planning stage, since line patterns/examples are assigned to the different fragments of the poem at this stage. This is because the retrieval of resources for one line is not independent from the assignation of resources to its neighbours. It is important, for instance, to ensure that consecutive lines are assigned sets of patterns that allow a correct linguistic fit between the resulting lines (the sentence fragment in the first line can be continued with the sentence in the following one, or there are patterns for starting a sentence on the following line if patterns in the first line include an end of sentence at the end of the line). 3.4 The CBR Approach to Line Generation A different CBR process is set in motion for each line in the poem. This is done sequentially starting from the first line, but none of the different CBR processes progresses onto its next stage until all the other processes are ready to do so as well: all processes retrieve their vocabularies and line examples, all processes adapt their selection to generate a new line (the revision and analysis phases can be done in any order). This ensures that no two lines in the same poem are built following the same pattern and that words are not repeated in a poem. Figure 2 Verse generation using CBR 3.4.1 Retrieve step For each fragment of the poem, adequate words and adequate line patterns/examples must be selected. Line examples/patterns have already been assigned during the planning stage. A module assigns priorities to all the available words, so that only the most adequate words are considered for each line. The criteria used in this assignation are, in decreasing order of priority: words actually appearing in the corresponding fragment of the intended message, words with the required rhyme for the line. These priorities can be modified once the initial lines have been generated, to include any words in the fragment of the intended message corresponding to previous lines that have not been used in the generation of the actual verse. 3.4.2 Reuse step For each set of fragment of the intended message, set of line examples /patterns, and set of selected words, the basic generation algorithm following the constructive rules for the appropriate line length is applied. The elementary generation units are the line pattern being followed (which acts as guiding form) and the draft of the current line. At each step of the generation the words are chosen to match the next POS tag in the line pattern (always according to their established priority ordering). If the chosen word meets the constructive requirements, it is appended to the draft of the current line and the process is iterated. The constructive requirements are implemented in the form of imperative Boolean functions. If no valid extension is found, the system allows limited backtracking: the last word to be added can be removed and an alternative attempted. An additional process of annotation takes place to avoid a repetition of alternatives that have already been tried. Additionally, the assignation of line examples/patterns to each fragment of the poem can be revised during the Reuse step if no solution is found with the existing assignation once all possible backtracking alternatives over the assigned words have been exhausted. If the assignation of line example/pattern is modified, the assignation of words may have to be revised. This new word assignation is now not only constrained by the words assigned to previous fragments, but also by the words already assigned to following fragments. However, if the generation process for any previous line has already finished at that point, any words of its fragment of the intended message that have not been used can be considered. 3.4.3 Revise step As soon as all the generation processes have concluded, the whole poem draft is presented to the user. The user is required to validate each of the lines. If the user comes up with possible modifications, the system accepts them instead of the generated verse. The modified versions suggested by the user ought to be tested for metric correctness. This feature is not currently contemplated, under the assumption that any corrections the user makes will either be well founded or they will be in repair of a serious syntactic or semantic error introduced by the heuristic approximation. 3.4.4 Retain step All verses validated by the user are processed to create a personal data file containing the corresponding line examples/patterns in the correct format. Any new words introduced by the user either as part of the intended message or modifications of the proposed draft are also added to the data. This will ensure that regular use of the system will improve the quality of results only if the degree of satisfaction with the initial poems is reasonable high. This implies that although the system is capable of a certain amount of learning once a reasonable standard (of vocabulary and stored examples) has been achieved, an initial threshold of vocabulary and corpus adequacy must be met for the system to function adequately. 3.5 ASPERA Results: Examples The following examples illustrate system input and resulting poem for specific instances of use of the ASPERA system. Suppose the user expressed a desire for a short, fully rhymed, formal poem, with a rural setting and a positive mood. The following specification was given to guide the search in example 1: tan hermoso viento fue corazón mudo The system suggested a terceto should be attempted (lines 11 syllables long rhyming ABA), and provided a set of model poems from its case base. Model poems were of the form given in example 0: Sabed que en mi perfecta edad y armado con mis ojos abiertos me he rendido al niño que sabéis ciego y desnudo. Example 0. Model poem Since rhyme restrictions are imposed by the generation mechanisms independently of the model poems, these need not follow the rhyming pattern required. The model poems provide guidance as to the number and specific POS elements to be used in constructing each line. Vocabulary files appropriate to the setting and mood where loaded and the system produced the following poem: Ladrará la verdad el viento airado en tal corazón por una planta dulce al arbusto que volais mudo o helado. Example 1. Three -viento, corazón and mudo - of the six words in the specification have made it into the twenty words of the final verse. Six words of the poem originate in the selected vocabulary -ladrará, planta, and arbusto through the rural requirement and dulce, verdad, and volais from the positive mood. Two words - airado and helado - come from the model poems. These words are forced to appear in the poem by the additional restrictions posed by the rhyme on the final words of the first and last line. As inexperienced human poets do, ASPERA simply chooses two words that rhyme and happens to find more of those among the model poems1. The rest of the words form part of ASPERA's elementary vocabulary and are used to hold the poem together with an approximation of sense. In this particular case, the last line of the resulting poem closely follows the structure of the last line of the sample model poem presented. The structure of the other lines is based on different model poems, chosen by the planning stage of the system to ensure minimal overall coherence of the poem. A similar analysis can be carried out for example 2: Andando con arbusto fui pesado vuestras hermosas nubes por mirarme quien antes en la liebre fue templado. Example 2. resulting from similar general choices and the specification: 1 This behaviour may improve if enough rhyming words are provided either in the selected vocabulary or the specification. andando por el monte hermosas nubes una liebre The poem as it stands is quite obscure, however, it can be seen very clearly in this case how the specification and the motives - rural and positive - drive the system. 4. Benefits to be Expected from ASPERA ASPERA may prove beneficial to a certain extent in two different contexts. Within the restricted domain of tuition in the field of literature, and more widely where it may shed light on general language problems. 4.1.1 As a Pedagogical Tool In the long run, it is hoped that ASPERA will provide guidelines to develop better pedagogical tools for the field of Spanish poetry. For the average person, a poem is either pleasing or not, and the introduction of - sometimes very complex - rules to be computed as part of its enjoyment goes against the grain on first time acquaintance. ASPERA provides the means for students to see how the imposition of the formal rules may affect the form of a message of their choosing, without requiring them to have learnt all possible combinations of the rules. It is expected that interacting with the system may help the students to develop - without learning the formal rules - the instinctive feel for 'correctness' of a verse that a poet achieves naturally. 4.1.2 As a Potential Source for Slim Language Processing Heuristics The solutions applied in ASPERA to language problems are all highly heuristic in nature, and resort to no complex linguistics techniques. The examples presented show the advantages and the shortcomings of the approach: message content gets scrambled for the sake of form, yet applications relying very heaviliy on specific formats and not requiring originality may find the simplicity of the techniques compensates for the loss of precision. From this point of view the ASPERA system, as it is evolved to more refined heuristics solutions to language generation, constitutes a benchmark for simplified approaches. 5. Conclusion ASPERA is the computing heart of an application with a great potential for user satisfaction. In its current version, it is handicapped by clumsy interfaces that are unable to compete with state of the art GUI technology. However, the underlying methodologies and techniques have shown their adequacy to the task even in their budding state of development. The heuristics techniques applied to resolve linguistic problems at all levels - syntactic and semantic - open interesting avenues of research in language processing. The system is open to enhancement and adaptation once user feedback has been gathered during evaluation. References [1] Aamodt, A. & Plaza, E. (1994). Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches. AI Communications, 7(i), pp.39-59. [2] Bailey, P., 'A reader-based model of story generation', Symposium on Artificial Intelligence and Creative Language: Stories and Humour, AISB'99 Convention, 6th-9th April 1999, Edinburgh College of Art & Division of Informatics, University of Edinburgh. [3] Chevreau, K, Coch, J., García Moya, J.A., and Alonso, M., 'Generación multilingüe de boletines meteorológicos', Procesamiento del Lenguaje Natural, SEPLN, N. 25, September 1999, pp 51-58. [4] Gervás, P., 'WASP: Evaluation of Different Strategies for the Automatic Generation of Spanish Verse', in: Time for AI and Society, Proceedings of the AISB-00 Symposium on Creative & Cultural Aspects and Applications of AI & Cognitive Science, 17th-18th April 2000, University of Birmingham, England, pp 93-100. [5] Gervás, P., 'Un modelo computacional para la generación automática de poesía formal en castellano', in: Actas del XVI Congreso de la SEPLN (Sociedad Española para el Procesamiento del Lenguaje Natural), Vigo, España, September 2000. [6] Gervás, P., 'A Logic Programming Application for the Analysis of Spanish Verse', in: Lloyd, J., et al, Computational Logic - CL 2000, First International Conference, London, UK, July 2000, Proceedings, Springer Verlag. [7] Horacek, H. and Busemann, S., 'Towards a Methodology for Developing Application-Oriented Report Generation', in: Günter, A. and Herzog, O. (eds.), 22nd German Conference on Artificial Intelligence (KI-98), Proceedings, Bremen, Germany, 1998. [8] Hisar Maruli Manurung, Graeme Ritchie and Henry Thompson, 'Towards a computational model of poetry generation', in: Time for AI and Society, Proceedings of the AISB-00 Symposium on Creative & Cultural Aspects and Applications of AI & Cognitive Science, 17th-18th April 2000, University of Birmingham, England. [9] Nederhof, M.-J., 'Efficient generation of random sentences', Encyclopaedia of Computer Science and Technology, Vol.41, Marcel Dekker, 1999, pp 45- 65. [10] Quilis, A., (1985) 'Métrica española', Ariel, Barcelona. [11] Real Academia Española (Comisión de Gramática), (1986) 'Esbozo de una nueva gramática de la lengua española', Espasa-Calpe, Madrid. [12] Sánchez León, F., Corpus Resources And Terminology ExtRaction project (MLAP-93/20), 'Spanish tagset of the CRATER project', ftp://ftp.lllf.uam.es/pub/CRATER/esT-tagger-1-0.tar.gz Introduction Problem Description The Formal Rules of Spanish Poetry Guiding Word Choice Fitting Words to Message and Metric Structure ASPERA: Automatic Spanish Poetry Expert - a Rewriting Application Collection of Poem Specifications The Poetry Expert Knowledge Base The Elementary Planning Step The CBR Approach to Line Generation Retrieve step Reuse step Revise step Retain step ASPERA Results: Examples Benefits to be Expected from ASPERA As a Pedagogical Tool As a Potential Source for Slim Language Processing Heuristics Conclusion References 
work_vfmson6knzf7xcg7eigifbw2ue	----	Environmental management for Agriculture: An Expert System Approach P. Tucker, K.A. Lewis & J.A. Skinner Division of Environmental Sciences, University of Hertfordshire Abstract The development of quantitative measures for self-assessment of the environmental performance of farming enterprises is described. A rule-based expert system approach is adopted based on well established sources of guidance. Value judgements are introduced as configurable coefficients. The system derives a compound performance indicator which can provide the basis for inter- or intra-farm comparisons and also provides an objective criterion on which the concept of continuing environmental performance can be judged. This method can also be used to provide an empirical check on environmental sustainability. The evaluations within the system are based on whether actions have, on balance, the potential to result in net environmental harm or benefit. The role of such evaluations within an integrated decision support framework is discussed. Introduction Farming as an activity operates through modifying the natural environment. Although in the past the consequences of this were often accepted without question, there is now a mounting concern over both the short- and long-term environmental impacts that might occur as the result of farming activities. Increasingly stringent regulation is forcing the farming community to substitute many traditional practices with more environmentally acceptable alternatives, e.g. for stubble· disposal, and is prompting a reconsideration of the potential impacts from many routine practices such as agrochemical application. Public and supply chain pressures are also beginning to become significant driving forces towards more sustainable practices. To move towards a more environmentally conscious farming system can present the farmer with many options to consider and often with difficult decisions to make. The solutions are not straightforward. Many diverse factors need to be interrelated, often depending on value judgements rather than hard facts. Options can conflict. For example, some mitigating measures to control nitrate leaching could result in increased emissions of greenhouse gases. How does a farmer rank one against the other? How can the totality of the impact be rationalized? Past activities should not be neglected as poor historical practices might have already degraded environmental quality - for example, through erosion, mineral excesses or toxic residues in soils. Natural attenuation of excesses and residues can take considerable time. Inappropriate action now can similarly leave its own legacy for future generations. The problem of how to sustain environmental improvement while maintaining economic growth presents challenges to the farmer and requires careful appreciation of all the risks and benefits that can ensue. These are all complex problems and, as yet, no universal solution can be offered, although much guidance is available. The Codes of Good Agricultural Practice (Anon., 1991; 1992; 1993) published by the UK Ministry of Agriculture, Fisheries and Food (MAFF)provide sound environmental advice for the management of soil, water and air and are complemented by other good reference sources (e.g. Canter, 1986; Park, 1988; Powlson, 1993; Association of Applied Biologists, 1992;Davis et al., 1993). There is not a paucity of information. The problem facing the farmer is how to distil this information into a coherent action plan. The commitment to good environmental management still relies heavily on voluntary action. Formal environmental management schemes such as BS7750 (British Standards Institution, 1994) can be applied to agricultural enterprises and again professional and general guidance is becoming available for the implementation of such schemes (Anon., 1995a; 1995b). If, however, a less formal approach is preferred, then it should be pursued. One underlying concept which needs to be considered in any approach is a commitment to continuing environmental improvement. This implies periodic evaluation against a baseline, the establishment of such a baseline being an early priority. One approach for this is through a self assessment questionnaire or checklist. The LEAF (Linking Environment and Farming) organization adopted this type of approach using a broadly based, though largely qualitative' audit' questionnaire (Blake, 1994). Although providing a useful catalyst for self- analysis, the scope of the LEAF audit does not, however, readily lend itself to the establishment of qualitative performance indicators nor to any holistic appraisal of the overall farm performance. This paper presents the next step in farm appraisal methodologies. It describes how the environmental performance baseline might be established on a more quantitative basis and how continuing environmental improvement could be evaluated. This evaluation, however, must be viewed as just one facet of the decision support that should be available to the farming industry. Access to sources of guidance and awareness of legislation, for example, are equally important to the farmer in formulating any sustainable development plan. The paper concludes by summarizing our initial decision support system, based on these concepts, for sustainable development in the agricultural industry. A Farm Performance Evaluation Environmental performance evaluation needs to set objective criteria against which actual performance can be judged. Such criteria can be set empirically as specific targets of a farm environmental policy or can be expressed more generically in terms of comparable or absolute measures. Comparative measures can include both time-based assessments (continuing improvement) or assessments against the performance of a 'model' farm where good practice is observed. Absolute measures could include a check for sustainability. The work presented here concentrates on the development of some possible generic measures. Clearly many individual and diverse factors will contribute in combination to the overall performance. An important requirement of any evaluation system is a mechanism to enable these factors to be combined in a rational, reproducible and acceptable way. The problems in establishing such a mechanism are well known. One commonly used approach is a simple scoring system, each impact or indicator being given a numerical score based on its magnitude and/or significance, with the total score representing the total impact. Examples can be cited for environmental impact assessment (e.g. Leopold et al., 1971;Solomon et al., 1977;Lee, 1987), life cycle analysis and site assessment studies (Tucker and Perguson, in press). Although often criticized for over-simplicity (e.g. Thompson, 1988), this method can provide an essential starting point, and perhaps the only practical means, for systems whose underlying complexities are not fully understood or agreed. The scoring approach was adopted here. The mapping of individual practices or actions into a numerical score was accomplished through a set of 'expert system' rules. The score range was set at +10 (maximum positive impact) to -10 (maximum negative impact) for individual impacts. The value 0 represents an environmentally neutral practice. The kernel of the rule base was established from UK Government publications, including MAFF Codes of Good Agricultural Practice, Codes of Recommendations for the Welfare of Domestic Fowls (Anon., 1990a), Codes of Recommendations for the Welfare of Pigs (Anon., 1990b), Codes of Practice for the Safe Use of Pesticides on Farms and Holdings (Anon., 1990)and Fertiliser Recommendations for Agriculture and Horticultural Crops, RB 209 (Anon., 1994). Additional heuristics were elicited from leading practitioners in the industry. Examples of the derived rules are given in the following. These examples make specific reference to fertilizer application. Rule 1: nitrogen application rates (nitrate leaching) This rule compares actual amounts of soil mineral N with that recommended by RB 209 (Anon., 1994). Recommendations are crop-specific and consider soil type, the amount of mineral already in the soil, either by analysis or based on the previous crop and any organic manure, sewage sludge or other soil conditioner applications. The scoring system is based on the relative error on the recommended application. Score (S1) = integer [10 x (recommended - actual/recommended)] (1) where if [S1] > 10 the [S1] = 10. Clearly, if less fertilizer than recommended is applied, the potential for leaching is low and the possibility of consequential environmental impact is small. Although -this- results in a positive environmental impact from this rule, this does not imply that it is in fact good practice. Crop yields may suffer and soil degradation may occur. Resultant soil degradation is accounted for in a separate rule. Rule 2: N, P, K levels (soil degradation) If phosphate (P) and potash (K) fall below maintenance levels, then soil degradation may occur (Davis et al., 1993). Equivalent rules can be assigned (heuristically) for N. If soil is not being sustained, S2 = -10, else S2 = 0. Rule 3: rainfall (nitrate leaching) If rainfall is above average (a default value of 900 mm per year is taken as the average), then if the actual N applications exceed the recommended levels, the risk of leaching is increased. The scale of this risk depends on the soil type, the depth of soil and the underlying rock. These are considered as far as possible in the soil type classification. Score (S3) = integer [10 x T x (average – actual rainfall/average)] (2) If the actual rainfall is greater than average, else S3 = O. T is a scaling parameter for soil type; T = 1.0 (sandy or shallow soils), 0.8 (organic or peat), 0.6 (silt) or 0.4 (clay). Rule 4: fertilizer application timing (nitrate leaching) Correct timing is important to ensure that crops make the maximum use of the N available. Too early an application increases potential leaching. Too late an application may result in impaired crop growth. Recommendations are given in RB 209, with additional advice provided by the Code of Good Agricultural Practice for the protection of Water. The guidance on timing in RB 209 is often specific (e.g. early May, six weeks after planting, etc.), which in practice may not always be acceptable. For example, if it rains when the application should have been made, then the farmer cannot be penalized. More environmental damage may be done by rigidly meeting the deadline. The application rules have therefore been designed to provide a wider tolerance on timing (e.g. spring, seedbed, split seedbed/ spring, etc.). These rules, by necessity, are crop-specific and consider existing soil N, recommended additions of N, soil type and in some instances individual crop variety. For example, consider main-crop potatoes on a sandy soil with average N levels. If the crop is irrigated, then fertilizer addition should be split, half in the seedbed and half at tuber initiation: S4 = +10 if the application timing is correct; S4 = +5 if the application is all in the seedbed or all in spring; S4 = -10 if the application is late autumn or winter and the crop does not require it; and S4 = 0 otherwise. Summary In all, well over 70 rules have been identified for fertilizer application alone, with special rules considering fertilizer application in environmentally sensitive areas. Similar rule bases are under development for pesticides, farm waste management, soil erosion and general performance assessment. These general performance assessment rules cover diverse concerns, not considered elsewhere, such as hedgerow management, woodland management and animal welfare. Approximately 300 rules have so far been identified for these concerns. An overall performance assessment is achieved by aggregating the individual scores into a single performance parameter S. S = [(Si • Wi)i] (3) Where wi is a weighting factor describing the relative importance of each rule against all the others. Currently values of wi have been set according to the authors' best opinion. These default values are not considered to be prescriptive and will inevitably evolve as wider expert consensus is reached and more practical understanding is gained. If a system user has good reasons to override the default value, then they should do so. The overall performance indicator, S, provides the comparative measure of environmental performance. The methodology will generate S indices most precisely when the evaluation is performed on a field by field basis, although aggregated whole farm values can equally be estimated. The scale of S values given by Equation (3) is fairly arbitrary, though consistent for both inter-farm and intra-farm comparisons. A positive S value indicates a probable overall beneficial affect on the environment, whereas a negative S value implies a probable detrimental effect. The value S = a (environmentally neutral) might be taken as the threshold of sustainability. It is often more helpful to the practitioner if the evaluation scale is normalized and bounded, say +100 to –100, so that it can then be related in absolute terms to performance against the best and worst case scenarios. This normalization, however, limits the scope for inter-farm comparisons unless local factors happen to be similar between the two units. User Input Requirements The demands of the 'expert system' rule base establish precise input requirements for the system. No redundant information needs to be requested. The first stage of the assessment involves the specification of local farm data. This includes, for example, the farm's geographical location, size, local climate, soil types, soil nutrient analyses, soil pH and outline details of farming operations. Other relevant information may also be required, such as the presence of nearby surface waters, archaeological features and whether or not the land is classified under any special environmental status. These data should be readily available to the farmer. Once the local information has been established, the user must then respond to an 'audit' questionnaire. The audit is divided into sub-audits relating to soil fertilizer application, pesticides, energy management, waste management, habitat conservation and animal welfare. The audit and its evaluation, if preferred, can be restricted to individual issues, such as energy management. The requested response can range from a simple categorical reply (yes/no or true/false), through multiple choice answers to the input of specific quantitative information, e.g. 'how much fertilizer N is applied (kg/ha)?' The multiple choice approach helps to avoid leading questions, which can prove particularly important when elements of good and bad practice are found together with a single activity. Discussion The performance measure, S, can provide a farm with a local baseline on which an assessment of environmental improvement can be based. However, although these methods will perform 'better' or 'worse' comparisons adequately, they cannot give absolute measures other than on an empirical local scale. This scale depends on value judgements made on the relative importance of diverse effects and practices. These value judgements are not prescriptive and can be configured to meet local objectives and criteria or made to reflect specific expert opinion. There is no right answer. This lack of absoluteness should, however, not be seen as a cause for concern. The system is, after all, only a support tool to help farmers make more informed decisions and to help reduce the risk involved in those decisions. It is not intended to exercise judgement on the farmer. The index S is based primarily on the evaluation of activities and, as such, provides a measure of whether individual activities have the potential to cause effects that are not either beneficial or detrimental to the environment. The value of S does not establish whether or not the activities will cause such impacts. For example, although a potential for leaching might be identified (say, for pesticides), it does not follow that the pesticides will be leached into a water course - they may be sorbed onto soil particles or degraded in the soil- nor does it follow that, if on reaching a water course, concentration levels will be sufficient to give concern (the EU legal limit of 0.1 ug/l might be taken as the threshold for concern). To evaluate whether an activity will actually result in an impact would require a second stage of analysis. This stage will be subject to additional uncertainties and require new assumptions based largely on an understanding of the externalities involved. In some instances it might be possible to use quantitative models for this analysis and to estimate the resulting environmental burdens, although value judgements would be necessary to interpret these burdens as environmental impacts. An Integrated Decision Support System A means for a structured farm evaluation provides one tool for realizing the commitment to good environmental management. Any recommendations arising from the evaluation need to be supported by sound evidence. This is provided through cross-referencing the evaluation system with the source publications that provide the basis for the rules. In the software implementation of the system, context-sensitive mapping provides hypertext jumps from the evaluation system directly to the relevant source-material. These sources of guidance can also be accessed independently of the evaluation system, with cross- referencing within and between individual texts being provided by further context-sensitive jumps. A third component of the system provides the technical support, including models and databases (for example, a database on average rainfall for England and Wales). The technical support module is also free-standing and is automatically consulted by the evaluation system when required. The system is shown structurally in Figure 1. The integrated system is written primarily in Visual Basic and is PC-based running under the Windows operating system. Conclusions The drive towards increasing environmental accountability is progressively affecting the agricultural industry. Farmers are faced with complex decisions on how best to improve environmental performance without jeopardizing profitability. Decision support to farmers is already provided in much reference material, although some of this advice may appear to conflict according to the environmental emphasis that has been adopted. Farmers also need to develop a means of evaluating environmental performance against a defined benchmark. A mechanism for this evaluation has been developed based on an audit' questionnaire and 'expert system' analysis using established rules of guidance extracted from quality reference sources. It is recognized that such a system can never produce an absolute measure of environmental performance, nor can it ever be free of value judgement. To be flexible and acceptable to all, these value judgements must be both transparent and configurable. In the system developed here the judgements are represented as a series of adjustable numerical coefficients. Their values are not prescriptive and will evolve with time as understanding develops. The evaluation must be constrained (for the time being) to estimates of whether actions, policy or procedures, singly or in combination, produce the potential to cause environmental harm or benefit. The performance criteria derived can form the basis for an environmental performance evaluation: (i) against expectations for a 'model' farm; (ii) as a test of continuing environmental improvement within the context of a farm environmental management system; and (iii) as an indicative check for sustainable development. These evaluations are not absolute measures; they simply provide reasoned support to help the farmer make some of the complex decisions involved in running a business. The system prototype is currently under evaluation by leading practitioners from the farming industry. The information and rule bases within the prototype have been focused on England and Wales; however, it is believed that the concepts should be generally applicable. Acknowledgement The work was funded by the UK Ministry of Agriculture, Fisheries and Food and was carried out in collaboration with IACR Rothamsted and ADAS. The advice of John Catt (IACR- Rothamsted) and Nick Nicholson and Brian Chambers (ADAS) is gratefully acknowledged. The views expressed in this paper are those of the authors and are not necessarily those held by the Ministry. References Anon. (1990) Code of Practice for the Safe Use of Pesticides on Farms and Holdings, HMSO, London. Anon. (1990a) Codes of Recommendations for the Welfare of Domestic Fowls, MAFF Publications, London. Anon. (1990b) Codes of Recommendations for the Welfare of Pigs, MAFF Publications, London. Anon. (1991) Code of Good Agricultural Practice for the Protection of Water, MAFF Publications, London. Anon. (1992) Code of Good Agricultural Practice for the Protection of Air, MAFF Publications, London. Anon. (1993) Code of Good Agricultural Practice for the Protection of Soil, MAFF Publications, London. Anon. (1994) Fertiliser Recommendations for Agriculture and Horticultural Crops (RB 209), HMSO, London. Anon. (1995a) An Introductory Guide to BS 7750 for Agriculture and Commercial Horticulture, ATB Landbase, Stoneleigh Warwickshire. Anon. (1995b) BS 7750: Sector Application Guidelines for Agriculture and Commercial Horticulture, ATB Landbase, Stoneleigh Warwickshire. Association of Applied Biologists (1992) Aspects of Applied Biology 30, Nitrate and Farming Systems, Horticulture Research International, Wellesbourne. Canter, L.W. (1986) Environmental Impacts of Agricultural Production Activities, Lewis, Chelsea. Blake, A. (1994) Leaf audit: step in right direction. Farmers Weekly, 2 Sep. British Standards Institution (BSI) (1994) British Standard BS 7750, BSI, London. Davies B., Eagle D. and Finney B. (1993) Soil Management, Farming Press, Ipswich. Lee, N. (1987) Environmental Impact Assessment: a Training Guide, Occasional Paper 18, Department of Town and Country Planning, University of Manchester. Leopold, L.B., Clarke, F.E., Hanshaw, B.B. and Basley J.R. (1971) A procedure for evaluating environmental impact, Circular 645, US Geological Survey, Washington. Park, J.R. (Ed.) (1988) Environmental Management in Agriculture: European Perspectives, Belhaven Press, London. Powlson, D.S. (1993)Agriculture, Scientific Basis for Codes of Good Agricultural Practice. Farming, Fertilisers and the Nitrate and Phosphate Problems, Commission of the European Communities, Luxembourg. Solomon, R., Colbert, B.K, Hansen, W.J., Richardson, S.E., Canter, L. and Vlachos, E.C. (1977) Water Resources Assessment Methodology (WRAM) Impact Assessment and Alternative Evaluation, Technical Report Y-77-l, US Army Corps of Engineers, Vicksburg. Thompson, M.A. (1990) Determining impact significance in EIA: a review of 24 methodologies, Journal of Environmental Management, 30. Tucker, P. and Ferguson C. A methodology for optimising sampling designs for contaminated sites, Mathematical Geology, in press. Biography The authors may be contacted at: Division of Environmental Science, University of Hertfordshire, Hatfield Campus, College Lane, Hatfield, Hertfordshire ALIO 9AB, UK. Tel.: 01707 284528. Fax.: 01707284514. E-mail: KA.Lewis@herts.ac.uk 
work_vfqugwm6pbe3pmbkovbeewlzta	----	Requirements for an expert system explanation facility Edinburgh Research Explorer Requirements for an expert system explanation facility Citation for published version: Moore, JD & Paris, CL 1991, 'Requirements for an expert system explanation facility', Computational Intelligence, vol. 7, no. 4, pp. 367-370. https://doi.org/10.1111/j.1467-8640.1991.tb00409.x Digital Object Identifier (DOI): 10.1111/j.1467-8640.1991.tb00409.x Link: Link to publication record in Edinburgh Research Explorer Document Version: Publisher's PDF, also known as Version of record Published In: Computational Intelligence General rights Copyright for the publications made accessible via the Edinburgh Research Explorer is retained by the author(s) and / or other copyright owners and it is a condition of accessing these publications that users recognise and abide by the legal requirements associated with these rights. Take down policy The University of Edinburgh has made every reasonable effort to ensure that Edinburgh Research Explorer content complies with UK legislation. If you believe that the public display of this file breaches copyright please contact openaccess@ed.ac.uk providing details, and we will remove access to the work immediately and investigate your claim. Download date: 06. Apr. 2021 https://doi.org/10.1111/j.1467-8640.1991.tb00409.x https://doi.org/10.1111/j.1467-8640.1991.tb00409.x https://www.research.ed.ac.uk/portal/en/publications/requirements-for-an-expert-system-explanation-facility(ef7ff645-9d9f-49ff-8d47-34b60b987091).html 361 Requirements for an expert system explanation facility JOHANNA D. MOORE Department of Computer Science and Learning Research and Development Center, University of Pittsburgh, Pittsburgh, PA 15260, U.S.A. AND CI~CILE L. PARIS In formation Sciences Institute, University of Southern California, 4676 AdmiraIty Way, Marina del Rey, Received August 9, 1991 Revision accepted September 24, 1991 CA 90292-669.5, U.S.A. Comput. Inttll. 7, 367-370 (1991) For the past several years, we have worked on building an explanation component for an expert system building framework (or “shell”), the Explainable Expert System (EESQ Framework. In this short paper, we describe the char- acteristics that we believe to be essential for an explanation component of an expert system. We then identify important features of the EES architecture that support the desired capabilities. Finally, we discuss some areas where fruitful work remains to be done. 1. Characterisitcs of a good explanation facility Every type of natural language application has its own requirements. Based on studies of human explanation giving and on our system-building experience, we outline here the requirements we have identified for an expert system expla- nation facility. Fidelity. In order to be trusted, an explanation must accurately reflect the system’s knowledge and reasoning (Swartout 1983, 1990). To ensure fidelity, an explanation facility must synthesize text directly from the same knowl- edge sources used for problem solving. It cannot rely on templates or canned text written a priori by a programmer because, as the system evolves, it is impossible to ensure that the corresponding text is an accurate reflection of the pro- gram’s behavior. Knowledge from multiplesources. To support the range of questions users wish to ask, an expert system must provide several different knowledge sources, including terminological knowledge, factual domain knowledge, problem-solving knowledge, and an execution history (see Swartout et al. 1991). An explanation facility must have strategies that enabie it to decide on the type(s) of information needed and extract it from these knowledge sources. Naturalness. Expert systems are often required to pro- duce multisentential explanations, for example, t o provide justifications of the system’s actions or recommendations, descriptions of its problem-solving strategies, or definitions of the terms it uses. The explainer must thus be able to organize information into a coherent presentation. To gen- erate coherent natural language explanations, the system’s explanations should follow standard patterns of discourse employed by humans. To do so, the explanation facility must have knowledge about discourse structure and strate- gies for employing that knowledge. The naturalness criterion applies not only to individual explanations, but to entire explanation dialogues as well. Printed in Canada / ImDrimC au Canada Taken as a whole, the explanations produced by the system during a dialogue must form a coherent set. Responsiveness. Analyses of naturally occurring advisory dialogues show that advice-seekers often do not understand the advisor’s explanations and frequently ask follow-up questions (Moore 1989b). The system must thus be able to answer follow-up questions or offer an alternative explana- tion if a user is not satisfied with a given explanation. Flexibility. An explanation facility must have a variety of strategies for answering a given question. This is impor- tant for at least two reasons. First, if a given explanation is not understood, the system must be able to offer an alter- native explanation. Second, the system must be able to pre- sent the same information from different perspectives, depending on contextual factors such as the user’s knowl- edge or goals. Sensitivity. Explanations should be influenced by infor- mation about the user’s knowledge and goals, the problem- solving situation, and the previous dialogue. Question and answer pairs cannot be treated independently; later explana- tions must take prior explanations into account. Extensibility. An explanation facility must be able to handle many different types of questions, some of which may not be foreseen at design time. It must be possible to easily extend the explanation facility to respond to additional question types or to add new strategies for answering existing question types. Portability. Because our generation facility is a com- ponent of an expert system shell, it can include only domain- independent strategies. The architecture must thus aid users in customizing the explainer by adding domain-dependent strategies to suit their requirements. Adaptive capabilities. An explanation facility should have the ability to automatically modify its existing strategies and learn new strategies based on experience and interactions with users (Paris 1990). 2. Brief system overview We have devised an explanation architecture that explicitly plans explanations to achieve discourse goals denoting what information should be communicated to the user (e.g., make the user know a certain concept, persuade the user to per- form an action). When a goal is posted, the planner searches its library of explanation strategies looking for candidates that could achieve it. In general, there will be several can- didate strategies for achieving a goal. The planner employs 368 COMPUT. INTELL. VOL. 7, 1991 (GOAL USER (DO USER REPIACE-1)) (RECOMMEND SYSTEM USER REPLACE-1) (PERSUADED "You rhouMloplw (SETO X 1 ) (GOAL USER (DO USER REPLACE-I))) with ( S m X 1)" (MOTIVATION REPLACE-1 ENHANCE-MAINTAINABILIW) I (BEL USER (SOMEREF (DIFFERENCES SETF SETO ENHANCE-MAINTAINABILITY))) I (BEL USER (REF (DIFFERENCES SETF SETQ ) USE)) / \ (CONTRAST (USE SETF ASSIGN-TOGV)) N (INFORM SYSTEM USER "SETO can on& be used lo assign a value (USE SETO ASSIGN-TO-SV)) lo (INFORM a simple SYSTEM v a d l e . " USER (USE SETF / \ ASSIGN-TOGV)) " S E F can be wed to assign a value (BEL USER (CONCEPT GENERALIZED-VARIABLE)) lo any generaked variable. " " I (ELABORATION GENERALIZED-VARIABLE) (INFORM SYSTEM USER \ (CLASS-ASCRIPTION GENERALIZED-VARIABLE SToRAGE-LoCATIoN))) (ELABORATION -OBJECT-ATTRIBUTE GENERALIZED-VARIABLE NAMED-BY) .* A generalized w h b l e is a storage bcarion" I (INFORM SYSTEM USER (NAMED-BY GENERALIZED-VARIABLE ACCESS-FUNCTION)) "rhaf can be named by any access function" Sample explanation generated by this text plan: You should replace (SHTQ X 1) with (SETF X 1 ) . SETQ can only be used to assign a value to a simple-variable. SETF can be used to assign a value t o any generalized- variable. A generalized-variable is a storage location that can be named by any access function. FIG. 1. A completed text plan and the explanation produced. a set of selection heuristics to determine which strategy is most appropriate in the current situation. These selection heuristics take into account information about the user's knowledge and goals (as recorded in the user model), the conversation that has occurred so far (as recorded in the dia- logue history), and information about whether or not a strat- egy requires assumptions to be made. Once a strategy is selected, it may in turn post subgoals for the planner to refine. Planning continues in a top-down fashion until all goals are refined into speech acts, such as INFORM and RECOMMEND. As the system plans explanations, it records the goal struc- ture of the response being produced. In addition, it keeps track of any assumptions it makes about what the user knows, as well as alternative strategies that could have been chosen at any point in the planning process. The result is a text plan for achieving the original discourse goal. Text plans are recorded in the dialogue history before being passed to the Penman text generation system (Mann and Matthiessen 1983), which performs the process of realiza- tion into English text. In our system, a completed text plan is more than simply a specification for the realization process. It is an explicit representation of the planning or "design" process that pro- duces an explanation. To briefly illustrate this point, we show a sample text plan in Fig. 1 accompanied by the text it produces. This text plan comes from an interaction with the program enhancement advisor (PEA) (Neches et al. 1985), an advice-giving system constructed using the EES framework that is intended to aid users in improving their Common Lisp programs.' The system produces the text plan shown in Fig. 1 to satisfy the discourse goal of achieving the state in which the user has adopted the domain goal of replacing (SETQ X 1 ) 'PEA recommends transformations that improve the "style" of the user's code. It does not attempt to understand the content of the user's program. MOORE AND PARIS 369 with (SETFX l ) , denoted by (GOAL USER (DO USER REPLACE-1)) in the figure. Basically this text plan does the following. It recommends that the user performs the replace- ment and then attempts to persuade the user to do this act. To persuade the user to replace SETQ with SETF, it motivates this act in terms of the shared domain goal of enhancing the maintainability of the program. To motivate this act, the system then chooses a strategy that contrasts the object being replaced (SETQ) with the object replacing it (SETF) in terms of the shared domain goal ENHANCE- MAINTAINABILITY. Although there are many differences between SETQ and SETF in the expert system’s knowledge base, the difference relevant t o the current domain goal of enhancing maintainability is in the generality of their usage. Thus, the system informs the user that SETQ can be used to assign a value to a simple variable and contrasts this with the use of SETF, namely to assign a value to any generalized variable. Finally, as the planner constructs the speech act to inform the user that SETF may be used to assign a value to any generalized variable, it determines from the user model that it must introduce a concept, GENERALIZED- VARIABLE, that is not known t o the current user. This causes the planner to post the additional subgoal (BEL USER (CONCEPT GENERALIZED-VARIABLE)) to introduce this new term. This goal is achieved by providing the elaborating information defining the concept in terms of its class membership and defining attributes. As shown in Fig. 1 and described more fully in Moore and Paris (1 989), a text plan represents the roles individual clauses in the text play in achieving discourse goals, as well as how the clauses relate to one another rhetorically. For example, contrasting the usage of SETQ with the usage of SETF is intended to achieve the discourse goal of persuading the user to do the replacement act. Moreover, the portion of text that contains the contrast is rhetorically related to the recommendation by a MOTIVATION relation. Similarly, defining the term GENERALIZED-VARIABLE is intended to make the user know this concept, and this text serves as ELABORATION for the INFORM act which first introduces the term. In addition, information about what entities are salient at each point in the explanation (attentional informa- tion) can be derived from a text plan. 3. Important architectural features In our system, knowledge about explanation is represented explicitly in a set of plan operators that were derived by studying naturally occurring explanations. These operators integrate multiple sources of knowledge. First, they encode stantlard ways that discourse (i.e., intentional) goals are achieved by rhetorical means, thus achieving our goal of naturalness. For example, as shown in the figure above, one strategy persuades the user to perform a replacement act using the rhetorical strategy of motivating the act by con- trasting the replacee with the replacer in terms of achieving shared goal@). . _ _ _ _ _ Second, operators contain applicability con- straints that specify the knowledge that must be available if the operator is to be used. These criteria can refer to the expert system’s knowledge bases, the user model, or the dia- logue: history. For example, the operator for explaining a concept by analogy requires that there be an analogous con- cept that is familiar to the user or has been mentioned pre- viously in the dialogue. Planning using these operators allows the system to produce coherent explanations directly _- from the expert system’s domain knowledge (fidelity). Because operator constraints also reference the user model, the system can tailor the content and organization of its explanations t o the individual user (sensitivity) (Paris 1990; Moore and Paris 1992).* A second important feature of our architecture is that explanations themselves are structured objects that the sys- tem can reason about. We have demonstrated that the text plans recorded by our system provide the context necessary for handling a range of dialogue phenomena. More specifi- cally, by reasoning about the prior explanations it has pro- duced, our system is able t o interpret users’ follow-up ques- tions and answer them in context of the ongoing dialogue (Moore and Swartout 1989; Moore 1989b), select a perspec- tive when describing or comparing objects (Moore 1989a), and avoid repeating information that has already been com- municated or that the user already knows (Moore and Paris 1989). Thus our system exhibits responsiveness and some types of Sensitivity t o dialogue context. 4. Future directions While we have made considerable progress towards achieving our goals, there are several areas in which much work remains. First, there is considerable discussion about the type of planning architecture that is best suited for explanation generation. Some argue for an approach in which an explanation is completely planned and revised by critics before it is generated (Suthers 1991; Lester and Porter 1991). Others advocate a simple goal refinement strategy which facilitates the interleaving of planning with realiza- tion (Cawsey 1989; Moore 19890). There is a clear tradeoff here. The former approach allows explanations to be opti- mized along various dimensions, but must expend consid- erable effort planning and revising without feedback from the user. The latter approach allows incremental explana- tion generation and thus the developing explanation can adapt t o the changing context quickly as user feedback is received. This tradeoff should be further explored through empirical tests. Yet others argue that planning text requires several inherently different types of reasoning, and therefore that a single top-down planning mechanism is insufficient (Hovy 1988; Mooney et al. 1991; Suthers 1991). We believe that it is crucial t o clearly identify the range of explanation tasks and the control strategies best suited to achieving each of these tasks. Suthers (1991) has begun this process. Another issue that concerns us is the feasibility of pro- viding domain-independent explanation strategies, which we must do if we are to provide a shell for constructing explain- able expert systems. EES has already been used to construct expert systems in several different domains. In addition to the program enhancement advisor, EES has been used t o construct expert systems for diagnosing faults in a space sta- tion, diagnosing faults in local area networks, and aiding physicians in prescribing treatment for patients in the cardiac intensive care unit. We have been able to reuse a considerable number of explanation ‘operators across these various domains. These include operators for justifying the system’s actions, defining terms, describing objects, and describing the system’s general problem-solving strategies. While it is ’In addition, Bateman and Paris (1989) have begun to examine how to tailor the phrasing of the generated texts to the users. 370 COMPUT. INTELL. VOL. 7, 1991 likely that additional domain-specific strategies will be needed in some domains (Kittredge et al. 1991), our initial experience is promising. Finally, we believe that more work must be done in under- standing t h e nature of explanation dialogues. Explanation dialogues are largely user-controlled, a n d thus the structure of the dialogue emerges only as t h e interaction proceeds. W e have only begun to understand this structure and the processes t h a t access it. Many interesting questions remain. H o w should the dialogue history be managed and used? Can things be forgotten (i.e., removed f r o m the dialogue his- tory)? Which things and when? How should later explana- tions be affected by previous utterances? What information must be recorded in order to guarantee that, t a k e n as a whole, t h e explanations produced by t h e system f o r m a coherent set? What is t h e relationship between t h e dialogue history a n d user model? A r e these separate entities? If so, should information migrate from t h e dialogue history to the user model? What information? H o w can we handle other dialogue phenomena such as interruptions and subdialogues? D o we require meta-level strategies for managing t h e dia- logue, such as the discourse plans o f Litman a n d Allen (1987) or those of Cawsey (1989)? We hope to be able to study some of these questions in t h e future. Acknowledgment The research described in this paper was supported by the Defense Advanced Research Projects Agency (DARPA) under a NASA Ames cooperative agreement number NCC 2-520. BATEMAN, J.A., and PARIS, C.L. 1989. Phrasing a text in terms the user can understand. Proceedings of the Eleventh Interna- tional Joint Conference on Artificial Intelligence, Detroit, MI, CAWSEY, A. 1989. Generating explanatory discourse: a plan-based interactive approach. Ph.D. thesis, University of Edinburgh, Edinburgh, United Kingdom. HOVY, E.H. 1988. Two types of planning in language generation. Proceedings of the Twenty-Sixth Annual Meeting of the Asso- ciation for Computational Linguistics, State University of New York, Buffalo, NY, pp. 179-186. KITTEDGE, R., KORELSKY, T., and RAMBOW, 0. 1991. On the need for domain communication knowledge. computational Intelligence, 7(4): this issue. LESTER, J., and PORTER, B. 1991. An architecture for planning multi-paragraph pedagogic explanations. Proceedings of the pp. 1511-1517. AAAI-91 Workshop on Comparative Analysis of Explanation Planning Architectures, Anaheim, CA, pp. 27-41. LITMAN, D.J., and ALLEN, J.F. 1987. A plan recognition model for subdialogues in conversations. Cognitive Science. 11: 163-200. MANN, W.C., and MATTHIESSEN, c. 1983. Nigel: a systemic grammar for text generation. Technical Report RR-83-105. Infor- mation Sciences Institute, University of Southern California, Marina del Rey, CA. MOONEY, D.J.. CARBERRY, M., and McCoy. K.F. 1991. Captur- ing high-level structure of naturally occurring extended explana- tions using bottom-up strategies. Computational Intelligence, 7(4): this issue. MOORE, I.D. 1989~. A reactive approach to explanation in expert and advice-giving systems. Ph.D. thesis, University of California, Los Angeles, CA. 19896. Responding to “huh?”: answering vaguely articulated follow-up questions. Proceedings of the Conference on Human Factors in Computing Systems, Austin, TX, pp. 91-%. MOORE, J.D., and PARIS, C.L. 1989. Planning text for advisory dialogues. Proceedings of the Twenty-Seventh Annual Meeting of the Association for Computational Linguistics, Vancouver, B.C., ___ 1992. User models and dialogue: an integrated approach to producing effective explanations. In User modeling and user- adapted interaction. Kluwer Academic Publishers, Dordrecht, The Netherlands. In press. MOORE, J.D., and SWARTOUT, W.R. 1989. A reactive approach to explanation. Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, Detroit, MI, pp. 1504-1510. NECHES, R., SWARTOUT, W.R. and MOORE, J.D. 1985. Enhanced maintenance and explanation of expert systems through explicit models of their development. lEEE Transactions on Software Engineering, SE-ll(11): 1337-1351. PARIS, C.L. 1990. Generation and explanation: building an explanation facility for the explainable expert systems frame- work. In Natural language generation in artificial intelligence and computational linguistics. Kluwer Academic Publishers, Boston/Dordrecht/London. pp. 49-82. SLITHERS, D.D. 1991. Task-appropriate hybrid architectures for explanation. Computational Intelligence, 7(4): this issue. SWARTOUT, W.R. 1983. XPLAIN: a system for creating and explaining expert consulting systems. Artificial Intelligence, 21(3): 285-325. Also available as Report ISI/RS-83-4, Informa- tion Sciences Institute, University of Southern California, Marina del Rey, CA. 1990. Evaluation criteria for expert system explanation. Proceedings of the AAAI-90 Workshop on Evaluation of Natural Language Generation Systems, Boston, MA. SWARTOUT, W.R. PARIS, C.L., and MOORE, J.D. 1991. Design for explainable expert systems. IEEE Expert, 6(3): 58-64. pp. 203-211. 
work_vgbbzggjsncmvk2od4ozpjsvme	----	Simultaneous people tracking and motion pattern learning Expert Systems with Applications 41 (2014) 7272–7280 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Simultaneous people tracking and motion pattern learning http://dx.doi.org/10.1016/j.eswa.2014.05.019 0957-4174/Crown Copyright � 2014 Published by Elsevier Ltd. All rights reserved. ⇑ Corresponding author. Tel.: +61 295147629; fax: + 61 295142655. E-mail addresses: Sarath.Kodagoda@uts.edu.au (S. Kodagoda), Stephan.Sehestedt@ uts.edu.au (S. Sehestedt). Sarath Kodagoda ⇑, Stephan Sehestedt Centre for Autonomous Systems, Faculty of Engineering and Information Technology, University of Technology, Sydney, PO Box 123, Broadway, NSW 2007, Australia a r t i c l e i n f o a b s t r a c t Article history: Available online 29 May 2014 Keywords: People tracking Motion pattern learning Human Robot Interaction The field of Human Robot Interaction (HRI) encompasses many difficult challenges as robots need a better understanding of human actions. Human detection and tracking play a major role in such scenarios. One of the main challenges is to track them with long term occlusions due to agile nature of human naviga- tion. However, in general humans do not make random movements. They tend to follow common motion patterns depending on their intentions and environmental/physical constraints. Therefore, knowledge of such common motion patterns could allow a robotic device to robustly track people even with long term occlusions. On the other hand, once a robust tracking is achieved, they can be used to enhance common motion pattern models allowing robots to adapt to new motion patterns that could appear in the environment. Therefore, this paper proposes to learn human motion patterns based on Sampled Hidden Markov Model (SHMM) and simultaneously track people using a particle filter tracker. The proposed simultaneous people tracking and human motion pattern learning has not only improved the tracking robustness compared to more conservative approaches, it has also proven robustness to prolonged occlusions and maintaining identity. Furthermore, the integration of people tracking and on-line SHMM learning have led to improved learning performance. These claims are supported by real world experi- ments carried out on a robot with suite of sensors including a laser range finder. Crown Copyright � 2014 Published by Elsevier Ltd. All rights reserved. 1. Introduction Successful Human Robot Interaction (HRI) requires a robot to have advanced abilities to carry out complex tasks. One such abil- ity is robust people tracking. It has been identified as an important tool in HRI not only for safe operation (Schulz, Burgard, Fox, & Cremers, 2003) but also for collision avoidance (Bennewitz, Burgard, Cielniak, & Thrun, 2005) or to implement following behaviors (Bolić & Fernández-Caballero, 2011; Gockley, Forlizzi, & Simmons, 2007; Prassler, Bank, & Kluge, 2002). Conventional way of people tracking is to use known motion models (Kluge, Kohler, & Prassler, 2001; Montemerlo, Thrun, & Whittaker, 1999) including probabilistic motion models (Tadokoro, Hayashi, Manabe, Nakami, & Takamori, 1995; Zhu, 1991). However, those methods have limited ability to model agile human movements. Therefore, some researchers opt not to use motion models (Francesc Serratosa & Amézquitaa, 2012; Kluge et al., 2001). Another type of techniques can track people in the vicinity of sensors (Montemerlo et al., 1999; Schulz, Burgard, Fox, &Cremers,Crem). However, they do not provide solutions for tracking with long term occlusions. Rosencrantz, Gordon, and Thrun (2003) have overcome the problem of tracking with temporary occlusions, however the tech- niques is not reliable when it is tracking outside the sensory ranges. It has been observed that human motion often follows place dependent patterns as it is influenced by a combination of social, psychological and physiological constraints (Altman, Rapoport, & Wohlwill, 1980; Arechavaleta, Laumond, Hicheur, & Berthoz, 2006; Dean, 1996; Hall, 1969). Therefore, if such a model could be learned by a robot, it could subsequently be used to improve people tracking. Approaches to the learning of place dependent motion pattern models have been proposed in the past. Bennewitz et al. (2005) used a network of laser scanners to learn motion patterns of indi- vidual occupants of an office environment. After collecting a data set, Expectation Maximization (EM) was used to cluster trajecto- ries for building a Hidden Markov Model (HMM). The HMM was used to implement collision avoidance behaviors. However, this technique relies on environment mounted sensors, which needs infrastructure modifications and hence it leads to difficulty in deployment. In Kanda, Glas, Shiomi, and Hagita (2009), multiple laser scanners were used to learn activity patterns of people in a shopping mall. The idea was to automatically identify potential customers in order to communicate with them. The approach http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.05.019&domain=pdf http://dx.doi.org/10.1016/j.eswa.2014.05.019 mailto:Sarath.Kodagoda@uts.edu.au mailto:Stephan.Sehestedt@uts.edu.au mailto:Stephan.Sehestedt@uts.edu.au http://dx.doi.org/10.1016/j.eswa.2014.05.019 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa S. Kodagoda, S. Sehestedt / Expert Systems with Applications 41 (2014) 7272–7280 7273 seems appealing, however it uses six infrastructure mounted laser scanners and off-line learning framework which may limit the feasibility to be used in many other robotic applications. Using stationary laser scanners, Luber, Tipaldi, and Arras (2011) proposed to learn a spatial affordance map based on a Poisson process. The approach was shown to be a viable extension to Multi Hypothesis Tracking (MHT) with improved tracking capabilities and robust- ness. However, it uses stationary sensors. Vasquez, Fraichard, and Laugier (2009) proposed Growing Hidden Markov Models to learn motion patterns. Although it implements an on-line learning algo- rithm, the technique requires the observer to be stationary. Lookingbill, Lieb, Stavens, and Thrun (2005) used a small helicopter with a camera as a semi-stationary observer to monitor a single roundabout to build an activity histogram. This histogram was shown to improve target tracking notably, however it may not be feasible nor suitable to be used in the indoor scenarios like the one proposed in our paper. Bruce and Gordon (2004) used training data to learn goal locations in a small environment. This informa- tion was then used to provide a particle filter based tracker with an improved prediction model, which was shown to perform better than Brownian motion for prediction. This work relies on previ- ously learned goal locations in a given environment. In general, the methods proposed in the literature have many limitations as described in the previous paragraphs such as the requirement of infrastructure based sensors, difficulty of operation under partial observability or occlusions, limited on-line adaptabil- ity, and on-line operation. Further, they are not capable of simulta- neously improving both the model learning and tracking. Therefore, here we propose simultaneous people tracking and motion pattern learning based on our previous work on SHMMs (Sehestedt, Kodagoda, & Dissanayake, 2010), which uses robot mounted sensors for on-line learning and can effectively handle partial observations with limited FOV of the sensors. This frame- work allows the robot to improve its tracking abilities with the availability of a learned model, while improving the model learn- ing with the feedback from the improved tracker. The SHMM based framework was tested in an office environment using the robot LISA shown in Fig. 1 with appealing results. More specifically, the contributions of this paper are, 1. Formulation of the Sampled Hidden Markov Model for effective handling of on-line learning and model adaptation to capture the changes in the human motion patterns. 2. Synthesis of the theoretically sound probabilistic algo- rithm for the simultaneous people tracking and motion pattern learning for improving both aspects. 3. Implementing and testing the algorithm and obtaining superior results for long term occlu- sions when comparing with conventional (model based) methods. This paper is organized as follows. Section 2 gives a brief intro- duction to SHMMs for on-line learning of common human motion patterns. Section 3 introduces and formulates the simultaneous Fig. 1. The LISA robot. people tracking and motion pattern model learning. Section 4 pre- sents experimental results showing the viability and effectiveness of the proposed approach. Finally, Section 5 summarizes the con- tributions and briefly discusses current and future work. 2. Sampled Hidden Markov Models In this section, a brief introduction to SHMM learning is given. SHMMs provide a sparse representation of common motion pat- terns and can be learned on-line using a mobile robot’s on-board sensors. The importance of this can be seen in Fig. 2, which illustrates a robot as a red circle and its current laser scan as a red outline. It can be observed that the field of view (FOV) of the robot is a small fraction of the size of the operating environment. Thus, any learning algorithm that can be used in such environ- ments needs the ability to incrementally learn with partial observations. An SHMM is defined by its states S and state transition matrix A where each state is represented by a set of weighted samples. Although, these sample sets can represent arbitrary probability distributions, for notational simplicity, the states are defined by their means l and covariances R as, S ¼ sðiÞ ¼ lðiÞ RðiÞ " # 1 6 i 6 N ð1Þ where N is the number of states in the model. Note that the model is time dependent as learning is done incrementally, however it is omitted for convenience of notation. The state transition matrix contains the probabilities of transitions from state i to state j as A ¼ aðijÞ ¼ K ðijÞ PðsðjÞjsðiÞÞ " # 1 6 i 6 N ð2Þ where KðijÞ is the number of times a transition was observed, from which the probability of the transition PðsðjÞjsðiÞÞ can be calculated. A particle filter based people tracker is used for learning the SHMM. Consider the situation in Fig. 3(a) where a person walked along the trajectory indicated by the arrow. The tracking algorithm produced a series of sample clusters, each of which could be inter- preted as one state of an SHMM. To obtain a more sparse represen- tation of the observed trajectory, a subset of those sample clusters was used as shown in Fig. 3(b), where means, covariances and transition of states are shown as red squares, red ellipses and red lines respectively. It is to be noted that the state transitions could be directly derived from the sequence of sample clusters as the temporal order of the clusters are known. Suppose another person is observed walking along the trajec- tory shown by the arrow in Fig. 3(c), where the information needs Fig. 2. A small robot observing its environment. Fig. 3. A basic example of SHMM learning. 7274 S. Kodagoda, S. Sehestedt / Expert Systems with Applications 41 (2014) 7272–7280 to be immediately added to the available motion pattern model. As part of the new trajectory is overlapped with the existing trajec- tory, a fusion mechanism is essential. Further, the non-overlapping states need to be appropriately added to the model. The symmetric Kullback–Leibler divergence (KLD) (Kullback & Leibler, 1951) is used for this purpose, which finds associations between states in the SHMM and sample clusters obtained from people tracking as, KLDðsðiÞksðjÞ�Þ¼ KLDsðsðiÞksðjÞ�Þþ KLDs�ðsðjÞ�ksðiÞÞ ð3Þ with 1 6 i 6 N and 1 6 j 6 K, where K is the number of sample clusters (candidate states, sðjÞ�) reported by the tracker. Whenever an association is found, the new information is added to the corre- sponding state and state transitions can be updated by counting. The non-associated sections are added as new states. Now the probabilistic formulation of the SHMM learning approach is given by the belief of motion patterns Dt at time t conditioned on the robot’s location estimate and people tracking results. That is the belief, BelðDtÞ¼ PðDtjnt; ft; zt; . . . ; n0; f0; z0Þ ð4Þ is to be computed, where ft is the robot localization hypothesis, nt is the people tracking result and zt the sensor reading at time t. From this an incremental update rule can be derived using the Bayes theorem as, BelðDtÞ¼ PðntjDt; ft; zt; nt�1; . . . ; n0; f0; z0Þ � PðDtjft; zt; nt�1; . . . ; n0; f0; z0Þ Pðntjft; zt; nt�1; . . . ; n0; f0; z0Þ ð5Þ Since the denominator is independent of Dt , it can be written as, BelðDtÞ¼ gPðntjDt; ft; zt; nt�1; . . . ; n0; f0; z0Þ � PðDtjft; zt; nt�1; . . . ; n0; f0; z0Þ ð6Þ where g ¼ Pðntjft; zt; nt�1; . . . ; n0; f0; z0Þ is a constant. This is the belief of D at time t given all past observations, sensor readings and observer poses. Obviously, this is not an efficient solution for updating the belief since all observations of moving people, sensor data and observer poses up until time t would have to be remembered. Therefore, it is assumed that obser- vations and poses are conditionally independent of past observa- tions and poses given ft and Dt , i.e. the system is Markov. Therefore, BelðDtÞ¼ gPðntjDt; ft; ztÞPðDtjft; zt; nt�1; . . . ; n0; f0; z0Þ zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{prior belief ð7Þ In fact the last term of this equation is the belief at time t � 1 and thus the final update rule is written as BelðDtÞ¼ g PðntjDt; ft; ztÞ zfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflffl{people tracking BelðDt�1Þ ð8Þ As given in the above equation, the belief BelðDtÞ can now be updated using the most recent observations of moving people. 3. Simultaneous people tracking and motion pattern learning In this section, the formulation is extended to simultaneous people tracking and motion model learning. The idea is to tightly integrate learning and tracking in order to improve both perfor- mances. Therefore, the goal is to estimate PðDt; nt; ftjztÞ ð9Þ which is the joint probability of the motion pattern model Dt , the position estimates of people in the robot’s FOV nt and the robot’s loca- tion estimate ft at time t given the latest sensor reading zt . Assuming independence between these variables, it can be shown that PðDt; nt; ftjztÞ¼ PðDtjnt; ztÞPðftjztÞ YK k¼1 PðnðkÞt jztÞ ð10Þ where K denotes the number of tracked people. It is to be noted that the first term on the right hand side is conditioned on nt . Further- more, it can be observed that the estimate of Dt is conditionally independent of zt given nt which is estimated from the latest sensor reading, leading to PðDt; nt; ftjztÞ¼ PðDtjntÞPðftjztÞ YK k¼1 PðnðkÞt jztÞ ð11Þ The second term on the right hand side defines robot localiza- tion and the estimate is commonly conditioned on control input given as ut resulting in PðDt; nt; ftjzt; utÞ¼ PðDtjntÞPðftjzt; utÞ YK k¼1 PðnðkÞt jztÞ ð12Þ Intuitively, this equation could be solved from the right to the left, i.e. after states of people have been estimated, the result could be used to improve the robot localization. The localization data in turn would be used to determine the global positions of detected people. Moreover, the result could be used in the first term to update the model of motion patterns. However, to fully exploit the idea of an incrementally learned model of motion patterns Dt , the simultaneous utilization and learning of the model is desir- able. To accomplish this, the last term of above equation could be made dependent on Dt . S. Kodagoda, S. Sehestedt / Expert Systems with Applications 41 (2014) 7272–7280 7275 PðDt; nt; ftjzt; utÞ¼ PðDtjntÞPðftjzt; utÞ YK k¼1 PðnðkÞt jzt; D_tÞ ð13Þ Note that when people tracking is performed, although time has already incremented to the current discrete time step t, belief at this point is still Dt ¼ Dt�1 , which is indicated by the notation D_t . We now have a formulation of SHMM learning, which explicitly accounts for a mobile observer and takes advantage of improved tracking while learning. 3.1. Motion prediction There is no general model for human motion prediction as such a very conservative prediction models like Brownian motion or constant velocity model is commonly used (Luber et al., 2011). In a b c d i dx dy f g h j y x Person e y Fig. 4. Motion predictio D D D HC Fig. 5. The map used for experiments. Desk areas, corridors and common are marked as laser reading is shown as a red outline. Note the limited FOV of LISA. (For interpretatio version of this article.) Fig. 6. A panoramic view of the la rare cases, less conservative learned models have been presented (Bennewitz et al., 2005; Bruce & Gordon, 2004; Luber et al., 2011), however, they often use stationary observers or off-line learning, which limits the potential mobile robotic applications. Here, we propose to utilize the learned motion models at the prediction stage of a particle filter (PF) based people tracker (Arulampalam, Maskell, & Gordon, 2002) as Pðntjnt�1; D_tÞ ð14Þ Consider the scenario given in Fig. 4(a) with a person (blue ellipse) associated with state ‘‘a’’ in the model, state ellipses (red) and state transition lines (thicknesses are proportional to the magnitude of transition probabilities). Then in the next state, the location of the person can be predicted as indicated by the purple arrow based on dx and dy. a b c d i f g h j x Person e n using an SHMM. D D D D H H ‘‘D’’, ‘‘H’’ and ‘‘C’’ respectively. LISA’s pose is shown by a red circle and the current n of the references to color in this figure caption, the reader is referred to the web rge open office environment. pe rs on tra je ct or y ro bo t Fig. 7. Model learning with real world data. 7276 S. Kodagoda, S. Sehestedt / Expert Systems with Applications 41 (2014) 7272–7280 In our current implementation, half of the samples are dedicated to SHMM based prediction and the other half is used for the prediction based on the constant velocity model to cater for some previously unlearned motion patterns. In more complex situations, there can be more state transitions based on the learned model (see Fig. 4(b)). There the person is associated with state d has two possible transitions, e and f. As the transition probabilities are available in the model, they are used to determine the number Fig. 8. Experimental set up for the evaluation of the tracking performance. Two trajectories (blue and green) were considered. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.) Table 1 Number of track losses in 22 trials. Experiment Sample size 50 100 200 500 1000 Stage 1 13 8 5 2 0 Stage 2 3 0 0 0 0 Stage 3 22 12 9 4 1 Stage 4 2 0 0 0 0 S. Kodagoda, S. Sehestedt / Expert Systems with Applications 41 (2014) 7272–7280 7277 of samples that are allocated to each prediction. That is, the predic- tions with higher transition probabilities are allocated more sam- ples than that of the prediction with lower transition probabilities. Hence the number of samples are estimated based on NðjÞD ¼ Pðs ðjÞjsðiÞÞ � ND; 1 6 i 6 N 1 6 j 6 N ð15Þ where N is the number of states of the SHMM, NðjÞD denotes the num- ber of samples dedicated to the transition from the current state i to state j. 3.2. Long term prediction Long term prediction of people locations without observations is a complex and a non-trivial problem due to the agile and compli- cated nature of human movements and the presence of occlusions. Without having prior knowledge about movements nor recent track new Fig. 9. Learning while observations, the conventional model based predictions tend to deviate from the actual people movements leading to track losses. However, this may be solved to a greater extent, if the knowledge of human motion patterns are available. In our approach, we utilize the learned human models based on SHMM in the prediction stage of the particle filter as described in the previous section. However, it is important to note that there is an inherent uncertainty in the learned model due to the gross nature of motion pattern learning. It gives rise to slight growing uncertainty of the estimator over time. However, as the tracker is still managed to follow the general trend of people motion, it still has significantly lower track losses. 4. Experimental results Experiments were conducted using the Lightweight Integrated Social Autobot (LISA) shown in Fig. 1. LISA is a low cost robotic platform designed for fast and easy deployment. The base is an iRobot create equipped with a small Intel Atom X86-64 computer together with a Hokuyo UTM-30LX laser range scanner (http:// www.hokuyo-aut.jp) for perception. The software development environment is Player/Stage (http://playerstage.sourceforge.net/) and all the algorithms were implemented in C++ within the Orca software framework (Brugali et al., 2007). Fig. 5 shows a Simultaneous Localization and Mapping (SLAM) generated map of the environment where desk areas and corridors are marked appropriately. The LISA robot is shown as a red circle with the red outline illustrating the observed laser reading. Being a small robot, it has a significantly limited field of view due to the presence of furniture. The map spans approximately 32 m � 20 m. Fig. 6 shows a panoramic view of the office environ- ment. It is important to note the complexity of the environment with large amount of clutter and semi-static objects like trash cans. Furthermore, some of the walls are made out of glass contributing to perception issues. 4.1. Learning motion patterns In this section SHMM learning is presented with the robot LISA in the aforementioned office environment. Ten different subjects were included in this experiment and no modifications were done track losses occur. http://www.hokuyo-aut.jp http://www.hokuyo-aut.jp http://playerstage.sourceforge.net/ 7278 S. Kodagoda, S. Sehestedt / Expert Systems with Applications 41 (2014) 7272–7280 to the environment. The limited observability at most times means that the robot had to explore the environment to build a model of motion patterns. Furthermore, in order to observe longer trajecto- ries the robot had to follow people, hence it was a truly mobile observer. The series of figures in Fig. 7 show the evolution of an SHMM. Fig. 7(a) shows the robot following a person, where the person is represented by a yellow cylinder and the trajectory is shown as an orange line. The robot is shown as a red circle, where the red outline indicates the observed reading of the forward looking laser sensor. The observed trajectory exhibits a typical human motion and accordingly it is represented in the initial model in Fig. 7(b). Fig. 7(c) shows the model after more than 70 observed trajecto- ries while the robot was on the move. The trajectories were successfully joined and compactly represented. The final represen- person1 person2 prediction1 prediction2 Fig. 10. Motion prediction during sensor failure in a complex situation. (a) Two people LISA’s sensor fails. Consequently, the position estimates of people were updated based maintained using the information from the SHMM. (d) After the sensor resumed opera maintained. tation including more than 80 trajectories are shown in the (Fig. 7(d)) as a unimodal Gaussians distribution. It could be noted that trajectories are positioned correctly on free spaces rather than through obstacles on the map. Further, compared to grid based representations of motion patterns, a greater efficiency is achieved as the belief has to be maintained only in the relevant areas (with human motion) of interest rather than over the entire space. 4.2. Tracking robustness In this experiment, the robot was positioned at an intersection while observing people trajectories as shown in the Fig. 8. The test data consists of 22 observed trajectories for the right turn and the same for the left turn (44 in total). The experiment was divided into four stages. In the first stage (stage 1) people followed the prediction1 prediction2 person1 person2 were tracked and associated with an SHMM with two disconnected trajectories. (b) on the SHMM. (c) Over an extended period of time, the tracks and identities were tion, both tracks could be recovered successfully and the identities were correctly robot person robot person robot person robot person Fig. 11. Long term motion prediction. S. Kodagoda, S. Sehestedt / Expert Systems with Applications 41 (2014) 7272–7280 7279 trajectory indicated by the blue arrow (right turn) to evaluate the tracking performance. In the second stage (stage 2,) a model of the blue trajectory was learned and used in the tracker to predict the motion. In the third stage (stage 3), the model from stage two was used, however, people followed the green trajectory (left turn), i.e. they exhibited a drastically different behavior compared to the previous observations. In the fourth stage (stage 4) of the experiment, the model was extended to include both types of trajectories and tested on people following either trajectory. Based on these tests, the tracking performance was evaluated in terms of the number of track losses as shown in Table 1. In stage 1, the subjects took a sharp right turn, as indicated by the blue arrow in Fig. 8, while being observed by the robot. As there was no earlier learned SHMM, tracking was done using the near constant velocity model for motion prediction. Sample sizes of 50, 100, 200, 500 and 1000 were considered and the observed track loss rates are summarized in the first row of Table 1. Not sur- prisingly, it can be seen that the tracking with fewer number of samples provided large number of track losses. This is due to the sharp turn, where the constant velocity model could not adequately provide a reasonable prediction. However, with the increasing number of samples, the tracker could achieve lower track losses. In stage 2, the SHMM model was learned first with sharp right turn data before being used in the tracker. The algorithm was tested by replaying the data in stage 1 of the exper- iment. As shown in Table 1, the tracker performance was signifi- cantly improved with no track losses for more than 100 samples used in the PF. In stage 3, data of the people following the right turn was used to learn the model, however tested on people following a left turn, i.e. their motion pattern would diverge from the learned model. As expected the number of tracks losses has increased significantly, even more than that of using a constant velocity model (stage 1). It can also be seen that the tracking performance has an adverse effect with the small number of samples used. Finally, in stage 4, after adaptively learning the left turn with the previously learned right turn, the tracker performed significantly better with only loss of 2 tracks even with 50 samples used in the PF. 4.3. Learning with track losses The findings in stage 1 and 3 of above experiment can be further analysed to understand the process of learning in the presence of track losses. As discussed before, in stage 1, LISA observed people following the blue trajectory in Fig. 8 without having a priory motion pattern model. The People were tracked with a sample size of 100 in this experiment. Fig. 9(a) shows the tracking result of the first observed person, where the track was lost during the sharp turn. Fig. 9(b) illustrates the learning result based on that trajec- tory. As could be seen in Table 1, most of the tracks were lost during the process without contributing much to the learned model. However, once an extended track was observed as shown in Fig. 9(c), the result was immediately used to update the model as shown in Fig. 9(d). This model was further tuned with other few more extended tracks to achieve the gross motion model. This alleviates the need to keep all the trajectories by just keeping the combined trajectory. 4.4. Occlusion handling and long term prediction Consider a part in an office environment with two strictly sep- arated motion trajectories as shown in Fig. 10(a). The figure further shows two people who were tracked by LISA while predicting both motions based on the motion pattern model. Then, in Fig. 10(b) a sensor failure has occurred, which is indicated by the absence of the red outline of the laser scan. From this point, the two people tracks were maintained purely based on the predictions made by the motion pattern model and the predicted estimates are shown as transparent purple cylinders in the figure. The sensor failure persisted for a while and the two people turned to their respective rights. While the constant velocity model based tracker failed to track both targets, our SHMM based tracker could successfully track in such agile situations as shown in Fig. 10(c). Furthermore, it could successfully maintain the identities of people. Finally, as in Fig. 10(d), after the sensor resumed it’s normal operation, the 7280 S. Kodagoda, S. Sehestedt / Expert Systems with Applications 41 (2014) 7272–7280 two people were re-detected correcting associated uncertainties of the filter. Consider a hallway which is partly divided by a wall as shown in Fig. 11(a), where the dividing wall is approximately eight meters long. A motion pattern model has been learned for this environ- ment. At one time, there was a person appeared while the robot was observing the hallway from the left side. The person moved upwards in the figure and took the path right of the dividing wall leaving the field of view of the robot. Once the person was occluded by the wall, as there were no observations, all samples were predicted according to the SHMM (Fig. 11(b)) for continuous tracking. Although, unavailability of observations lead to slight increase of uncertainty (which is indicated by the green covariance ellipse in Fig. 11(c)), it did not contribute to track losses. The con- stant velocity model based implementation lost track within a cou- ple of iterations of the particle filter. Once the observations were available towards the end of the track, the filter converged success- fully as shown in Fig. 11(d). 5. Conclusions and future work 5.1. Conclusions Model based conventional trackers have significant difficulties in tracking with long term occlusions. Therefore, this paper pro- posed to use human motion pattern models in people tracking within a probabilistic framework. We presented simultaneous peo- ple tracking and human motion pattern learning based on Sampled Hidden Markov Models and particle filter tracker. Learning is achieved on a mobile observer where only the robot’s on-board sensors were used without relying on cumbersome infrastructural sensors. Since there was no dedicated learning phase, all new infor- mation was incorporated in the model immediately improving the adaptability. This property is of central importance which particu- larly distinguish our work with many of the other previously pre- sented approaches. The representation of motion as an SHMM is more memory efficient than grid based approaches as the model only represents motion in areas of interest rather than the whole space. Furthermore, as the model is not fully connected, the transi- tion matrix can be replaced by more compact data structures. The knowledge of motion patterns has been shown to be useful in not only handling long term occlusions but also maintaining identities. While a reduction in the tracking errors could be observed with SHMMs integrated tracker, the most significant improvement was observed in the tracking robustness especially in complex situations like a sudden change of direction. The num- ber of lost tracks were greatly reduced even when using small sam- ple sizes for PF tracking. These effects in turn led to a better convergence in motion pattern learning, which again highlights the benefits of on-line learning of human motion patterns. The integration of tracking and learning was shown to have mutual benefits. 5.2. Future work Once the robot is placed in a new environment, it takes some time for learning the motion patterns. Until a reasonable motion pattern model is learned, the tracking performance of the current algorithm suffers in accuracy. Therefore, learning the human behaviors directly from the environmental features rather than relying on observing humans in the environment (through object affordance and hallucinated humans) is a research direction that is of future interest. Further, in the context of learning of human spatial behavior, this paper has presented some of the fundamental abilities needed for lifelong learning such as on-line learning, adaptability and the ability to include incomplete knowledge in incremental learning. One another key competency for lifelong learning under uncertainty in a changing environment is the for- getting of obsolete information, which is not addressed in this research and is considered as a great future direction. Acknowledgments This work was supported by the ARC Centre of Excellence pro- gramme, funded by the Australian Research Council (ARC) and the New South Wales State Government. References Altman, I., Rapoport, A., & Wohlwill, J. (1980). Environment and culture. Springer. Arechavaleta, G., Laumond, J. P., Hicheur, H., & Berthoz, A. (2006). The nonholonomic nature of human locomotion: A modeling study. In IEEE/ RAS- EMBS international conference on biomedical robotics and biomechatronics. Arulampalam, M. S., Maskell, S., & Gordon, N. (2002). A tutorial on particle filters for online nonlinear/non-Gaussian Bayesian tracking. IEEE Transactions on Signal Processing, 50, 174–188. Bennewitz, M., Burgard, W., Cielniak, G., & Thrun, S. (2005). Learning motion patterns of people for compliant robot motion. International Journal of Robotics Research, 24. Bolić, J. M. G., & Fernández-Caballero, A. (2011). Agent-oriented modeling and development of a person-following mobile robot. Expert Systems with Applications, 38, 4280–4290. Bruce, A., & Gordon, G. (2004). Better motion prediction for people-tracking. In IEEE international conference on robotics and automation (ICRA), Citeseer. Brugali, D., Brooks, A., Cowley, A., Côté, C., Domínguez-Brito, A., Létourneau, D., Michaud, F., & Schlegel, C. (2007). Trends in component-based robotics. Springer tracts in advanced robotics (Vol. 30). Berlin/Heidelberg: Springer. Dean, D. (1996). Museum exhibition: theory and practice. Routledge. Francesc Serratosa, R. A., & Amézquitaa, N. (2012). A probabilistic integrated object recognition and tracking framework. Expert Systems with Applications, 39, 7302–7318. Gockley, R., Forlizzi, J., & Simmons, R. (2007). Natural person-following behavior for social robots. In ACM/IEEE international conference on human-robot interaction (HRI) (pp. 17–24). New York, NY, USA: ACMhttp://doi.acm.org/10.1145/ 1228716.1228720. Hall, E. T. (1969). The hidden dimension – man’s use of space in public and private. London: The Bodley Head Ltd. Kanda, T., Glas, D., Shiomi, M., & Hagita, N. (2009). Abstracting people’s trajectories for social robots to proactively approach customers. IEEE Transactions on Robotics, 25, 1382–1396. http://dx.doi.org/10.1109/TRO.2009.2032969. Kluge, B., Kohler, C., & Prassler, E., 2001. Fast and robust tracking of multiple moving objects with a laser range finder. In IEEE international conference on robotics and automation (ICRA) (pp. 1683–1688). Kullback, S., & Leibler, R. (1951). On information and sufficiency. The Annals of Mathematical Statistics, 22, 79–86. Lookingbill, A., Lieb, D., Stavens, D., & Thrun, S. (2005). Learning activity-based ground models from a moving helicopter platform. In IEEE international conference on robotics and automation (ICRA) (pp. 3948–3953). Luber, M., Tipaldi, G. D., & Arras, K. O. (2011). Place-dependent people tracking. The International Journal of Robotics Research http://dx.doi.org/10.1177/ 0278364910393538. Montemerlo, M., Thrun, S., & Whittaker, W. (1999). Conditional particle filters for simultaneous mobile robot localization and people-tracking. In IEEE international conference on robotics and automation (ICRA) (pp. 695–701). Prassler, E., Bank, D., & Kluge, B. (2002). Motion coordination between a human and a mobile robot. In IEEE/RSJ international conference on intelligent robots and systems (IROS) (pp. 1228–1233). Rosencrantz, M., Gordon, G., & Thrun, S. (2003). Locating moving entities in dynamic indoor environments with teams of mobile robots. In Second joint international conference on autonomous agents and multi agent systems (pp. 233–240). Schulz, D., Burgard, W., Fox, D., & Cremers, A. B. (2003). People tracking with mobile robots using sample-based joint probabilistic data association filters. The International Journal of Robotics Research, 22, 99–116. Sehestedt, S., Kodagoda, S., & Dissanayake, G. (2010). Models of motion patterns for mobile robotic systems. In IEEE/RSJ international conference on intelligent robots and systems (IROS) (pp. 4127–4132). Tadokoro, S., Hayashi, M., Manabe, Y., Nakami, Y., & Takamori, T. (1995). On motion planning of mobile robots which coexist and cooperate with human. In IEEE/RSJ international conference on intelligent robots and systems (IROS) (pp. 518–523). Vasquez, D., Fraichard, T., & Laugier, C. (2009). Growing hidden markov models: An incremental tool for learning and predicting human and vehicle motion. The International Journal of Robotics Research, 28, 1486–1506. Zhu, Q. (1991). Hidden markov model for dynamic obstacle avoidance of mobile robot navigation. IEEE Transactions on Robotics and Automation, 7, 390–397. http://refhub.elsevier.com/S0957-4174(14)00299-1/h0060 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0065 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0065 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0065 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0070 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0070 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0070 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0075 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0075 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0075 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0080 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0080 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0080 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0085 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0090 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0090 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0090 http://doi.acm.org/10.1145/1228716.1228720 http://doi.acm.org/10.1145/1228716.1228720 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0105 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0105 http://dx.doi.org/10.1109/TRO.2009.2032969 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0115 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0115 http://dx.doi.org/10.1177/0278364910393538 http://dx.doi.org/10.1177/0278364910393538 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0120 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0120 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0120 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0125 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0125 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0125 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0130 http://refhub.elsevier.com/S0957-4174(14)00299-1/h0130 Simultaneous people tracking and motion pattern learning 1 Introduction 2 Sampled Hidden Markov Models 3 Simultaneous people tracking and motion pattern learning 3.1 Motion prediction 3.2 Long term prediction 4 Experimental results 4.1 Learning motion patterns 4.2 Tracking robustness 4.3 Learning with track losses 4.4 Occlusion handling and long term prediction 5 Conclusions and future work 5.1 Conclusions 5.2 Future work Acknowledgments References 
work_vh63ojb7zrewbdz7l3biyqd4zq	----	A model for aircraft evaluation to support strategic decisions This is a repository copy of A model for aircraft evaluation to support strategic decisions. White Rose Research Online URL for this paper: http://eprints.whiterose.ac.uk/96373/ Version: Accepted Version Article: Bruno, G., Esposito, E. and Genovese, A. (2015) A model for aircraft evaluation to support strategic decisions. Expert Systems with Applications, 42 (13). pp. 5580-5590. ISSN 0957-4174 https://doi.org/10.1016/j.eswa.2015.02.054 Article available under the terms of the CC-BY-NC-ND licence (https://creativecommons.org/licenses/by-nc-nd/4.0/) eprints@whiterose.ac.uk https://eprints.whiterose.ac.uk/ Reuse This article is distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs (CC BY-NC-ND) licence. This licence only allows you to download this work and share it with others as long as you credit the authors, but you can’t change the article in any way or use it commercially. More information and the full terms of the licence here: https://creativecommons.org/licenses/ Takedown If you consider content in White Rose Research Online to be in breach of UK law, please notify us by emailing eprints@whiterose.ac.uk including the URL of the record and the reason for the withdrawal request. mailto:eprints@whiterose.ac.uk https://eprints.whiterose.ac.uk/ 1 Introduction In the contemporary global market, the civil aviation industry is characterized by an ever increasing attention towards passengers’ needs, service quality, comfort and environmental issues (Hope, 2005; Newson & Cairns, 2006; Brindisi & Concilio Concilio, 2008; Hope, 2005; ICAO, 2013; Miyoshi & Mason, 2009; ICAO, 2013Newson & Cairns, 2006). In this context, industry stakeholders (including airlines and aircraft manufacturing companies) can no longer consider the provision of air transport exclusively as a cost-oriented problem; consequently, airlines should select aircraft in such a way that they assure the best combination among costs, technical characteristics, passenger comfort, and environmental impact. On the other hand, large manufacturing companies, prior to the launch of capital intensive production programmes, have to clearly identify the set of characteristics that better satisfy airlines’ requirements (Esposito & Passaro, 1997; Esposito & Raffa Raffa, 1994). The misalignment with respect to airlines requirements could jeopardize the success of new programmes and generate relevant financial losses (Esposito & Raffa, 2007; Ferreri, 2003). Unsurprisingly, in recent years, the academic literature has been focusing on many aspects in terms of aircraft performances, besides the traditional ones. Several papers analyse analyze environmental impact from different points of views, by highlighting: - the environmental impact of emissions and noise due to air traffic (Abeyratne, 2003; Sen, 1997Armstrong, Allen, & Denning, 1997; Borken-Kleefeld, Berntsen, & Fuglestvedt, 2010; Price & Probert, 1995; Sen, 1997; Tsilingiridis, 2009); - the social costs of aircraft noise and emissions (Brueckner & Zhang, 2010; Lu & Morrell, 2006; Schipper, 2004); - strategies and goals of national and international organizations in reducing noise and emissions from air transport (Abeyratne, 2002; Girvin, 2010; Koblen, Szabo, & Krnáčová, 2013; Ott, 2007) and the economic impact and ecological effects of such strategies (Brueckner & Girvin, 2008; Vespermann & Wald, 2011); - the proposal of new products, new technologies and new materials (such as new engines, alternative fuel, etc.) in order to reduce the negative environmental impact of the aviation industry (Dray, 2013; Frota, 2010; Haddad & Fawaz, 2013; Dray,2013). A Model model for Aircraft Evaluation aircraft evaluation to Support Strategic Decisionssupport strategic decisions Giuseppe Brunoa giuseppe.bruno@unina.it Emilio Espositoa emilio.esposito@unina.it Andrea Genoveseb, ⁎ a.genovese@shef.ac.uk aDepartment of Industrial Engineering (DII), University of Naples Federico II, Piazzale Tecchio, 80 – 80125 Naples, Italy bManagement School, University of Sheffield, Conduit Road, S1 4DT, Sheffield, UK ⁎Corresponding author. Abstract In the contemporary air transport industry, many factors, such as environmental impact, service quality and comfort are becoming increasingly crucial. In this context, airlines and manufacturing companies can no longer consider air transport exclusively as a cost-oriented problem; thus, for airlines, the choice concerning the purchase of the aircraft is no longer a simple matter of minimizing operative costs, but rather a multi-attribute problem, characterized by a high complexity level in which a variety of factors play a pivotal role. Given this scenario, the aim of this paper is to propose a novel model for aircraft evaluation, based on the investigation of airlines’ needs. The model is based on the two main approaches proposed in literature to address generic evaluation problems, Analytic Hierarchy Process and Fuzzy Set Theory, proposing a hybrid approach which combines some the strengths of the two methodologies. The usability of the hybrid model for the stakeholders of the air transport industry is investigated through an empirical study. Keywords: AHP; Fuzzy Set Theory; Aircraft evaluation; Decision-making elsevier_ESWA_9904 Also, the amount of papers dealing with comfort and service quality issues is increasing, focusing on: - the impact of airline service quality and comfort on passenger choices (Balcombe, Fraser, & Harris, 2009; Jiang, 2013; Martin, Roman, & Espino, 2008; Park, Robertson, & Wu, 2004; Park, Robertson, & Wu, 2004, 2006; Wojahn, 2002; Jiang, 2012;Pennig, Quehl, & Rolny, 2012; 2012; Wojahn, 2002; Yang, Hsieh, Li, & Yang, 2012; Zhang Y., Zhang, 2012); - the attributes of the airline service quality (Babbar & Koufteros, 2008; Curry & Gao, 2012; De Jager, Van Zyl, & Toriola, 2012; Kim & Lee, 2009; Martin, Roman, & Espino, 2011; Wen & Yeh, 2010) and the customer-value drivers (Boetsch, Bieger, & Wittmer, 2011; Park, Robertson, & Wu, 2009); - the evaluation of airline service quality (Chen & Chang, 2005; Cheng & Chang, 2006; Chou, Liu, Huang, Yih, & Han, 2011; Higgins, Lawphongpanich, Mahoney, & Yin, 2008; Liou & Tzeng, 2007; Pakdil & Aydin, 2007; Tsaur, Chang, & Yen, 2002); - the effect of service quality on airlines’ performance (Sim, Koh, & Shetty, 2006), the proposal of methods and strategies for improving airline service quality (Liou, Tsai, Lin, & Tzeng, 2011; Liou, Yen, & Tzeng, 2010; Maji, 2012); - the important role of comfort in the interior design of airplanes (Brindisi & Concilio Concilio, 2008; Lee & Luengo-Prado, 2004; Vink, Bazley, Kamp, & Blok, 2012). The extant literature suggests that, in the air transport industry, aircraft selection and evaluation issues have become a multi-attribute problem, characterized by a high complexity level in which a variety of quantitative and qualitative factors play a crucial role. In this context, the aim of this paper is to propose a novel model for aircraft evaluation, based on the investigation of airlines’ needs. This model considers not only traditional characteristics (operating costs and technical performance, such as cruise speed) but also a variety of features whose importance is strikingly increasing, such as environmental impact and aircraft interior quality. The proposed model can be useful both for airlines, in their process of selecting the most suitable aircraft for their fleet, and for manufacturing companies in their process of designing future aircraft. The model is based on two popular approaches proposed in literature to address evaluation problems, the Analytic Hierarchy Process (AHP) (Saaty, 1980; Saaty, 1980, 2001) and the Fuzzy Set Theory (FST) (Zadeh, 1965), proposing a hybrid approach combining the main strengths of the two methodologies. The usability of the model for the stakeholders of the air transport industry and its adaptability to contexts characterized by complexity, high technological level, and increasing requirements in terms of sustainability and environmental regulations, are investigated through an empirical study focused on regional transport aircraft. The paper is organized in the following sections. Section 2 provides further information about the extant literature on evaluation models in the aviation industry. Then, in Section 3, the hybrid model is introduced and described. In Section 4, the proposed model is implemented and a case study related to the evaluation of three regional aircraft is analyzed and discussed. Finally, some conclusions are drawn. 2 Background Since the 1960s, international air transport organizations and large aircraft manufacturing companies have developed accurate methodologies to calculate aircraft direct operating costs (AEA, 1985; Airbus, 1988; ATA, 1964; ATA, 1964, 1967; Boeing, 1972). For many years direct operating costs (DOCs) and Net Present Value (NPV), that in turn includes DOCs, were the main two indicators being utilized in the aircraft evaluation process and in purchasing decisions. Under this approach, Gibson and Morrell (2004) proposed the use of a variant of the well-documented adjusted present value (APV) concept and suggested a methodology for aircraft financial evaluation using Monte Carlo simulation and Real Option Analysis. Although this contribution has the merit of proposing a robust methodology, which overcomes the simple analysis based on DOCs or NPV, it addresses the evaluation exclusively from the financial point of view, thus neglecting issues related to quality and environmental impact. In the early 2000s, a number of airlines introduced empirical multi-criteria procedures that included not only DOCs or NPVs, but also a number of indicators, such as speed, range and passengers’ capacity (Ferreri, 2003). The weight of each criterion was identified on the basis of past experience or through heuristic decision-making procedures. Even today, many airlines use simple empirical multi-criteria procedures. In fact, despite the fact that the literature is rich of contributes suggesting a variety of tools to deal with multi-criteria problems, only few papers focus their attention on aircraft evaluation. Among these, See et al. (2004) presented a multi-attribute methodology for selecting the best aircraft among a set of alternatives. Authors used the method of the hypothetical equivalents and inequivalents (Wu, 1996), basing their choice on three criteria: speed, range and number of passengers. This paper has the undoubted advantage of dealing with the problem of aircraft evaluation through a multi-criteria approach; nevertheless it neglects important criteria such as costs, aircraft interior quality and environmental issues. In addition, when the number of alternatives or the number of criteria increase, the proposed method becomes extremely farraginous and provided results are not easy to be interpreted. Yeh and Chang (2009) proposed a fuzzy multi-criteria decision making algorithm. The evaluation process is based on three criteria (technological advances, social responsibility and economical efficiency), further articulated into eleven elsevier_ESWA_9904 sub-criteria. For each criterion, the performance of each aircraft is evaluated through a fuzzy rating. The authors use a pair-wise comparison process to assess the weights among the three criteria, and between sub-criteria within each criterion. The crisp weights are translated into fuzzy numbers and then aggregated with the fuzzy rating of performances. The result is an overall fuzzy preference value for each aircraft type. The overall fuzzy preference value is then defuzzified to obtain a crisp preference value for each aircraft type. The proposed algorithm appears not easy to use. Nevertheless, it could be partially simplified avoiding the fuzzification of the crisp weights. In fact, the aggregation of the crisp weights with fuzzy performances provides overall fuzzy preference values that are highly correlated with the results of the proposed algorithm, revealing that fuzzification of weights could be an unnecessary complexification. Gomes, Mattos Fernandes, and Mello (2012) proposed a fuzzy stochastic approach to the multi-criteria selection of aircraft based on the NAIADE method (Novel Approach to Imprecise Assessment and Decision Environments). The process of evaluation is based on three criteria (Financial, Logistics, Quality) further articulated in twelve sub-criteria (Acquisition Cost, Liquidity, Operating Costs, Range, Flexibility, Cruising Speed, Replacement Parts Availability, Landing and take-off distance, Comfort, Avionics, Availability, Safety). The output is the ranking of alternatives by means of an outranking procedure adapted from the PROMÉTHÉE method. The proposed method is conceptually simple. Nevertheless, as also authors underline, in the practical applications the mathematical calculations become complex and the analysis of alternative hard to evaluate and interpret. In this context this paper proposes a novel model for aircraft evaluation, which aims to overcome the weaknesses of the previous models; specifically, coherently to the above-mentioned nature of the aircraft evaluation problem, the model should be based on a multi-criteria framework, including the most relevant categories for stakeholders in the civil aviation industry. Also, the usability of the model and its applicability to real-world decision-making should be ensured. 3 A Hybrid Model hybrid model for aircraft evaluation The proposed model is based on the two main approaches suggested in literature to address evaluation problems, the Analytic Hierarchy Process (AHP) and the Fuzzy Set Theory (FST). In particular, our approach starts from the analysis of respective weaknesses and strengths of AHP and FST and proposes a hybrid model combines some of the positive features of the two approaches. From the assessment of AHP and FST approaches, some positive and negative sides related to their usage emerge. Indeed, there are some distortions introduced by AHP and FST techniques in the perception, evaluation and computation respectively of performances and weights associated with criteria adopted in the evaluation process. AHP models can significantly bias the evaluation of performances associated with the criteria due to the mechanisms that rule the conversion of qualitative judgements into numerical values (as illustrated in Bruno et al., 2012); this generally produces final rankings that tend to flatten the differences among alternatives themselves, causing then troubles to the decision-maker who cannot clearly identify the best solution. For these reasons, when AHP is adopted for performance evaluation, differences are not properly tracked and the ending outcomes of the model may appear significantly altered. When FST approaches, instead of AHP, are applied for dealing with criteria performances, the distortions illustrated above are almost eliminated. Indeed, in FST models, the qualitative scale defined for each criterion is not treated as a flat scale, as membership functions are defined for each one of the levels. In this way, a fuzzy variable is associated with each crisp numerical value of the indicators related to the evaluation criteria. This fuzzy variable keeps track of the degree of membership of the measured value to each defined qualitative range; hence, biases introduced by AHP approaches are almost overcome. On the other side, from an accurate investigation on FST models, and in particular on their application in real-world practice, some other shortcomings can be highlighted. When decision-makers are inquired to state judgments about the weights associated with different criteria, a flattening/overestimating effect of weights assessment is triggered. Firstly differences between levels of importance of criteria are lost and then lots of criteria are overestimated. This happens for the uncontrolled decision-makers propensity to judge criteria importance equally high or very high for all the factors considered when absolute qualitative judgments are inquired. When AHP is applied for weights determination, the drawbacks illustrated above are overcome thanks to pair-wise evaluations between criteria that allow detecting even little differences perceived by decision makers about importance assigned to different criteria. Therefore, AHP-based models appear relevantly suitable for weights determination, meanwhile for performance evaluation they lead to some biased results; FST-based models, instead, seem to be very suitable for performance estimations, but, on the other side, they introduce some distortions in the weights assessment process as well. For this reason, we propose a hybrid model which combines the methodology to determine criteria weights typical of the AHP approach with performance estimation drawn from a FST based model. In particular, the use of FST (through the use of linguistic scales) in the performance evaluation process could be very beneficial in this case, as experts conducting the evaluation process may not be able to estimate, in a precise way, different nuances in performance levels; therefore, it may be important to use methods that capture the imprecision, rather than to use crisp scales. In the literature there are several proposals of Fuzzy-AHP models for dealing with problems in many different fields (Buyukozkan, 2012; Calabrese, Costa, & Menichini, 2013; Chen & Chao, 2012; Ertay, Kahveci, & Tabanli 2011; Tabanli, 2011; Lee, Nha Le, Genovese, & Koh, 2011; Calabrese, Costa, & Menichini, 2013;Li, Liu, & Chen Chen, 2012; Shaw, Shankar, Yadav, & Thakur, 2012; Zeydan, Colpan, & Cobanoglu Cobanoglu, 2011). All these papers propose the “fuzzification” of the AHP through the adoption of a fuzzy pair-wise comparison matrix (see also, for instance, Abdullah & Najib, 2014; Abdullah & Zulkifli, 2015; Cho & Lee, 2013; Ishizaka & Nguyen, 2013; Rezaei, Fahim, & Tavasszy, 2014). In this process, they practically insert the Fuzzy Logic within the core of the AHP methodology. As stated by Saaty and Tran (2007), the result is a contamination of the AHP with the Fuzzy Logic that implies a higher computational complexity (involving the calculation of eigenvalues, eigenvectors and consistency ratios) and a more complex framework, which increases the difficulties in the phase of implementation and reduces them usability in the practical applications, as also results interpretations may become more difficult. Moreover, the fuzzification of pair-wise comparison matrix is unnecessary since the matrix in itself is able to catch the nuances of the judgements (Barbarosoglu & Yazgac, 1997; 1997; Cheng, Chen, elsevier_ESWA_9904 Chang, & Chou, 2007; Eagan, 1999; Narasimhan, 1983; Nydick & Hill, 1992; Saaty, 1992; Saaty & Tran, 2007; Soukoup, Soukup, 1987). It is also worth to notice that several adaptations of decision support models based on a combination of Fuzzy Logic and AHP have been applied to deal with evaluation problems in the aviation industry (Aydogan, 2010) (Aydogan, 2011) and aeroengine health assessment (Wang, Fan, & Wang, 2010); however, the combined application of the two methodologies for the aircraft evaluation problem is a novel contribution. Moreover, differently from what is illustrated in the literature, the hybrid model proposed and assessed in this paper combines the procedure for deriving criteria weights typical of the AHP approach with performances drawn from a FST approach. In other words, the model we propose does not contaminate the AHP with FST (and vice-versa), vice versa), while keeping the two approaches separate and using each of the two for its respective strengths. In this way the computational complexity of the Fuzzy-AHP application is reduced, and the practical application is facilitated. While Fuzzy Set Theory and AHP are well established methodologies, their application in the way proposed in this paper represents another contribution, as, to the best of our knowledge, has not been proposed in similar studies. The implementation of the model is characterized by the following steps (Fig. 1). Step 1. The starting point of the hybrid approach is represented (as in the case of the AHP and FST-based approaches) by the identification of airlines’ needs. Step 2. After the identification of airlines’ needs, the characteristics of the formal model are then defined in terms of evaluation criteria, sub-criteria, alternatives and decision makers. Step 3. Once criteria and sub-criteria have been identified, these ones, along with alternatives and decision makers, are arranged according to a hierarchical scheme, as prescribed by the AHP approach. Step 4. The determination of the weights is performed following an AHP approach; in particular, after the definition of the hierarchical scheme, the relative importance of evaluation criteria and sub-criteria at each level of the hierarchy is evaluated through the pair-wise comparison method according to Saaty (1980) scale reported below: 1: Equal importance of two elements 3: Moderate importance of one over another 5: Strong or essential importance 7: Very strong or demonstrated importance 9: Extreme importance Within this framework, 2, 4, 6, 8 are utilized to keep track of intermediate values, while reciprocals are utilized for inverse comparisons (Saaty, 1980). elsevier_ESWA_9904 Step 5 (From here onwards, the indentation should be back to normal.). Consistency of the obtained pair-wise comparison matrices is then verified according to Saaty (1980) procedure. For each pair-wise comparison matrix of dimension n, this is done, by extracting the maximum eigenvalue λmax and computing the Consistency Index (CI) as: Then, the computed CI is compared to the Random Index (RI), representing an average CI for a huge number of randomly generated matrices of the same order, to obtain the Consistency Ratio (CR) as: Typical values for RI depending on the dimension of the matrix n are reported in Table 1. Values of CR lower than 0.1 indicate an acceptable level of inconsistency in the pair-wise comparison matrices, that can be therefore utilized for the extraction of the weights. Step 6. Once consistency has been proved, the final pair-wise comparison matrices are then utilised utilized to derive a coherent vector of priorities corresponding to criteria weights, by extracting from each of them the eigenvector associated with the maximum eigenvalue and normalising normalizing its components (Saaty, 1980). Step 7. For the sake of performance evaluation, a FST based approach is proposed; in particular, criteria performances are defined as linguistic variables on five qualitative levels: Very Poor (VP), Poor (P), Medium (M), Good (G), Very Good (VG). Thereafter, the five qualitative levels are translated into fuzzy numbers. In particular, the membership functions of the fuzzy numbers representing the terms VP, P, M, G, VG associated with the linguistic variable performance are evaluated by a direct estimation method (Watanabe, 1979). The range Fig. 1 Hybrid approach (step by step). elsevier_ESWA_9904 where the trapezoidal membership function assumes maximum value (equal to 1), is defined as the range corresponding to the intersection of the judgments collected from multiple decision makers; the border values (where the trapezoidal membership function assumes values equal to 0), instead, are labelled labeled on the extreme values of the range given by the union of the judgments collected from them. Step 8. Performance evaluation is then conducted. This step consists of a measurement process of indicators associated with criteria and sub-criteria, through an appropriate data collection process. Step 9. Performances are then fuzzified, by translating numerical values given by indicators into fuzzy numbers. In this case the numerical values measured for each criterion are compared to the corresponding term set of linguistic variables. The output values for membership functions are combined according to the inferred weights of the members through a fuzzy weighted operator. The result of this procedure is a fuzzy number translating the crisp value measured for a specific couple criterion-alternative. This procedure has to be applied for all the criteria (and related sub-criteria) of the hierarchy. Step 10. As depicted in Fig. 1, fuzzy performances and crisp weights need to be combined to get the final vendor rating. The issue here is represented by the intent to pick the fuzzy aggregation operator in a way to avoid to spread the entropy related to fuzzy numbers when they are combined. Fuzzy weighted mean operator is judged as the most suitable one to adopt for weights, performances and combined aggregation since it is a convex composition of several fuzzy sets with coefficients which indicate the ‘percentage’ of a given set in the aggregation. It allows combining fuzzy variables in a simple way, not requiring time expensive or complex calculations, and resembling human decision making. Step 11. The ranking of trapezoidal fuzzy numbers representing the score associated with the different alternatives represents the last step associated with hybrid model implementation. It appears fundamental to utilise utilized the final fuzzy scores of alternatives for profiling a final rank of the alternatives, in order to identify the best one, as this is an important component of the decision process. According to the review performed by Abbasbandy and Hajjari (2009), more than 30 fuzzy ranking indices have been proposed, although a heated debate has been developing about the counter-intuitiveness and absence of discrimination capability of many of these methods. According to the seminal work of Bortolan and Degani (1985), each ranking method involves some losing of information; still, nowadays, there is a lack of a golden choice that allows the identification of a universally accepted ranking methodology (Abbasbandy & Hajjari, 2009; Kaufmann & Gupta, 1988); Brunelli and Mezei (2013) have proven that rankings may differ significantly depending on the adopted methodology. Table 1 Random Index values. n 1 2 3 4 5 6 7 8 R.I. 0 0 0.52 0.89 1.11 1.25 1.35 1.40 Within ranking methods, defuzzification techniques provide a way to associate a crisp real number to fuzzy sets, in such a way that a ranking can be developed by utilizing a simple ordering relation. These specific techniques can be classified in three main categories (Saletic, Velasevic, & Mastorakis, 2002): distribution techniques, maxima techniques, area techniques. Distribution and area techniques are suggested for use in fuzzy controllers; the maxima techniques are suggested for use in general fuzzy expert systems and fuzzy decision-making systems (Saletic et al., 2002), mainly for the low computational complexity which characterizes them. Therefore, since the main purpose of the step is to provide a simple and straightforward ranking to industrial decision-makers, maxima technique (and, in particular, the Middle of Maxima defuzzification method) are judged as the most suitable to profile the final ranking of the alternatives. 4 Empirical Studystudy To test and to verify the usability of hybrid model in the air transport industry, in this section an empirical study involving three regional aircraft is illustrated. Step 1. The identification of airlines needs is performed considering the case of Air Italy, an Italian airline, which was interested to select aircraft to be purchased within a set of three candidates: Bombardier CRJ1000 (first flight 2008), Sukhoi SSJ100 (first flight 2008), and Embraer ERJ190 (first flight 2004). Step 2. The formal model setting was focused on the definition of the characteristics of the model, identifying the evaluation criteria capable of better representing the airline’s requirements. In this case the evaluation criteria were derived merging some insights coming from the literature with Air Italy requirements. The overall set of provided criteria (including economic performance, technical performance, aircraft interior quality, and environmental impact) was then discussed and articulated in specific sub-criteria (including operating cost, aircraft price, cruise speed, autonomy, seat comfort, cabin luggage compartment size, environmental pollution, and noise); for each sub-criterion, specific indicators were defined and normalized in a range from 0 to 1 (as reported in Table 2). Step 3. Collected criteria and related sub-criteria were organized in homogeneous groups by interviewees involved in the aircraft evaluation process. The output of such gathering is the hierarchy of criteria indicated in Fig. 2. Step 4. Table 3 reports the pair-wise matrix that is the result of the focus group meeting with the Airline decision makers. It represents the priority establishment between the criteria at the first level of the hierarchy according to the Saaty’s (1980) scale. A similar procedure was followed also for the other elements in the hierarchy. Step 5. The principal eigenvalue (λ(λmax) of the matrix reported in Table 4 was computed to evaluate the consistency index (CI). Then, the latter was matched with random index (RI) to derive the consistency ratio (CR). As reported in Table 3 the final consistency ratio was equal to 0.023, which is less than the threshold (0.1) needed to assure consistency. A similar procedure was followed also for the other elements in the hierarchy. elsevier_ESWA_9904 Step 6. Once the consistency of the matrix had been verified, the final priority vector was obtained (Table 4). The eigenvector associated with the principal eigenvalue (λ(λmax), was calculated and normalized to obtain the final priority vector. Its components, as indicated in Table 5, correspond to the weights associated with the variables at the first level of the hierarchy. Following the same procedure, the weights associated with the other variables of the hierarchy were extracted obtaining the final results depicted in Fig. 3. Step 7. An example of definition of performances as linguistic variables is reported in Figs. 4a and 4b. Through the direct estimation method, qualitative ranges associated with five levels (very poor, poor, medium, good and very good) were defined. Fig. 4a reports the direct estimation process (based on judgments collected from four experts) for the qualitative performance level Good under the criterion Speed. It can be seen as the intersection and the union of the four judgements represent, respectively, the core and the support of the fuzzy number. Results of the same process applied to the other performance levels for the same criterion are reported in Fig. 4b. Step 8 and Step 9. An example of performance fuzzification is reported with reference to the criterion Cruise Speed. Fig. 5 shows that, starting from the crisp numerical performances measured trough specific indicators associated with the sub-criteria reported in the hierarchy (which, in this case is equal to 0.40, for the CRJ1000), this performance is then fuzzified by comparing its crisp numeric value to the term set of the linguistic variables defined for that criterion. The fuzzy numbers corresponding to crossed membership functions, combined with the inferred weights through a fuzzy weighted operator allow obtaining the final fuzzy number representing the performance of the alternative CRJ1000 for the sub-criterion Cruise Speed. This same “fuzzification” method was applied to get fuzzy numbers representative of the performances of the criteria belonging to the last level of the hierarchy and also of all the respective sub-criteria, which were directly measured to evaluate the score for each aircraft evaluated. Step 10. Crisp weights and fuzzy performances were aggregated across the hierarchy, according to the aggregation of crisp weights and fuzzy performances to obtain the fuzzy preference index through the weighted fuzzy operator (see Figs. 6–8). Step 11. Finally, in adherence to the ranking of the fuzzy preference index for each alternative step, the fuzzy numbers representative of the aircraft rating were defuzzified adopting the middle of maxima (MoM) defuzzification method. Results are reported in Table 6. Table 2 Criteria, Sub-criteria and indicators. Criteria Sub-criteria Indicator Pre-normalization range Economic performance Unit operational costs Cost per passenger per NM on a typical 300 NM route [0.098; 0.158] Aircraft price Purchasing price in M$ [30;60] Technical performance Cruise speed Maximum speed in KTS [400; 500] Autonomy Maximum range in NM [1000; 3500] Aircraft interior quality Seat comfort Square meters per passenger [0.51; 0.75] Cabin luggage compartment size (kg) Kilograms of overhead baggage per passenger [5;10] Environmental impact Environmental pollution kg CO2-eq per passenger emitted on a 300 KTS route [40; 79] Noise EPNdB measurement in critical stages [250; 280] elsevier_ESWA_9904 Table 3 Pair-wise comparison matrix. Vendor rating (This should read "Aircraft Score") Economic performance Technical performance Aircraft interior quality Environmental impact Economic performance 1.00 5.00 5.00 5.00 Technical performance 0.20 1.00 1.00 2.00 Comfort 0.20 1.00 1.00 2.00 Environmental impact 0.20 0.50 0.50 1.00 Table 4 Principal eigenvalue, consistency index and consistency ratio. Vendor rating (This should read "Aircraft Score") λmax 4.061 CI 0.020 RI 0.890 CR 0.023 Table 5 Priority vector. Priority vector 0.62 0.15 0.15 0.09 Fig. 2 Hierarchy of criteria. elsevier_ESWA_9904 Fig. 3 Weights associated with evaluation criteria. Fig. 4a (It would be better to have a better balance of text and figures, by including figures exactly where specified in the submitted version of the manuscript.)Direct estimation for the qualitative performance level “Good” for the criterion “Speed”. elsevier_ESWA_9904 Fig. 4b Definition of performance as linguistic variable for the criterion “Speed”. Fig. 5 Performance evaluation and fuzzification for the criterion “Speed”. Fig. 6 Application of the model to the alternative Sukhoi SSJ100. elsevier_ESWA_9904 Table 6 Aircraft final score. Aircraft Fuzzy rating Crisp score Sukhoi SSJ100 (0.464; 0.631; 0.677; 0.775) 0.637 Bombardier CRJ1000 (0.385; 0.512; 0.588; 0.714) 0.550 Embraer ERJ190 (0.310; 0.493; 0.546; 0.675) 0.506 Table 6 shows that the Sukhoi SSJ 1000 has a higher score than Bombardier CRJ1000 and Embraer ERJ190. The practical relevance of the hybrid approach does not consist only in its usefulness as an evaluation system but rather in Fig. 7 Application of the model to the alternative Bombardier CRJ1000. Fig. 8 Application of the model to the alternative Embraer ERJ190. elsevier_ESWA_9904 the opportunity to adopt it as a strategic tool by both airlines and manufacturing companies. Airlines may use the model to clearly identify their own requirements and ask to manufacturer to modify specific characteristics to better satisfy their own needs. Vice versa, manufacturers may use the model in the phase of design for better meet the airlines requirements. For example, the Embraer ERJ190 appears to be the aircraft with the lowest score. However the analysis of the performances (Fig. 8) provides useful suggestions to the manufacturer about how it could be possible to gain positions in the ranking, mainly by increasing the value of the criterion Operating costs, which is definitely lower than the one reported by other aircraft and whose associated weight is considerably high (0.75). If Embraer engineers are able to re-design a new version of the ERJ190 aligning operating costs towards an average value of the performances estimated for the other competitors (Table 7), ERJ190 could increase its overall rating from 0.506 to 0.596 in such a way to become second instead of last in the new ranking. Table 7 Hybrid model application as strategic tool to better improve aircraft rating. Aircraft Operating costs Updated operating costs Sukhoi SSJ100 (0.505; 0.677; 0.737; 0.788) Bombardier CRJ1000 (0.611; 0.736; 0.833; 0.861) Embraer ERJ190 (0.356; 0.560; 0.606; 0.684) (0.558; 0.707; 0.785; 0.820) 5 Conclusions In this paper a model for aircraft evaluation has been proposed. The model includes four criteria (economic performance, technical performance, aircraft interior quality, and environmental impact) and eight sub-criteria (aircraft price, operative cost, cruise speed, autonomy, seat comfort, cabin luggage compartment size, noise, and environmental pollution). The model framework is articulated in eleven steps and it is based on a hybrid approach that uses both the Analytic Hierarchy Process and the Fuzzy Set Theory. The AHP approach is utilised utilized for criteria weights determination because it ensures to keep track of the differences between importance associated with diverse criteria thanks to pair-wise evaluations between them; the FST approach, instead, is adopted to deal with aircraft performances because it allows representing the vagueness associated with criteria evaluation and indicating the “nuances” of airlines perceptions about aircraft performances without losing information. In this way the proposed model overtakes some AHP and FST weaknesses and combines some of their strengths. The proposed model also overtakes weaknesses of previous models dealing with aircraft evaluation. Differently by Gibson and Morrell (2004) and See et al al. (2004), the hybrid model includes a variety of criteria in line with the new requirements of airlines, such as seat comfort, cabin luggage compartment size, noise, and environmental pollution. Compared with models from Yeh and Chang (2009) and Gomes et al. (2012), the hybrid model has a lower level of computational complexity, which facilitates its practical application. The model has been tested starting from the requirements of an Italian airline (Air Italy) and considering three potential candidates (Sukhoi SSJ 1000, Bombardier CRJ1000, and Embraer ERJ190) to be selected for enlarging the airline’s fleet. The results of the model implementation highlight how the practical relevance of the hybrid approach does not consist only in its usefulness as an evaluation system but rather in the opportunity to adopt it as a strategic tool. In particular, airlines may use the model both ex-ante, in order to clearly identify their own requirements, and ex-post, in order to choose the aircraft that better fits to their requirements; manufacturers operating in global markets may use the model both ex-ante, for meeting, in a better way, airlines’ requirements, and ex-post, for comparing their own aircraft with those of competitors to propose new models or modified versions of old models. Further researches could be addressed at extending the model towards a multi-stakeholder perspective, in which not only the airline point of view is considered but also the one of other players in the civil aviation industry. In this way, the model could become a strategic tool for detecting preferences misalignment and gaps in the industry, leading to interesting considerations. Future studies will also address the issue of the actual usability of such model, by deploying more real-world case studies reflecting on further issues arising from its implementation. Also, the methodological framework could be further developed, by introducing a specific module for performance aggregation and ranking determination based on the use of well-established techniques (such as, for instance, outranking methods, TOPSIS); the integration of the model in a dynamic decision support system could be also proposed.(Aydogan, 2011; Cheng et al., 2007; Jiang, 2013; Lee et al., 2011; Yang et al., 2012). References Abeyratne R., Some recent developments in aviation and environmental protection regulation, Environmental Policy and Law 32 (1), 2002, 32–40. Abeyratne R., Air transport and sustainable development, Environmental Policy and Law 33 (3–4), 2003, 138–143. elsevier_ESWA_9904 Abbasbandy S. and Hajjari T., A new approach for ranking of trapezoidal fuzzy numbers, Computers and Mathematics with Applications 57 (3), 2009, 413–419. Abdullah L. and Najib L., A new type-2 fuzzy set of linguistic variables for the fuzzy analytic hierarchy process, Expert Systems with Applications 41 (7), 2014, 3297–3305. Abdullah, L., & Zulkifli, N. (in press). Integration of fuzzy AHP and interval type-2 fuzzy DEMATEL: An application to human resource management. Expert Systems with Applications. AEA (1985). Requirements for future advanced short/medium range aircraft. Association of European Airlines Report, Bruxelles. Airbus (1988). Airbus project D.O.C. method. Airbus Industries Report, Toulouse. Armstrong F.W., Allen J.E. and Denning R.M., Fuel-related issues concerning the future of aviation, Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering 211 (1), 1997, 1–11. ATA (1964). Introduction to airline operating economics. Air Transport Association of US, Washington D.C. ATA (1967). Standard method of estimating comparative direct operating costs of turbine powered transport airplanes. Air Transport Association of US, Washington D.C. Aydogan E.L., Performance measurement model for Turkish aviation firms using the rough-AHP and TOPSIS methods under fuzzy environment, Expert Systems with Applications 38 (4), 2011, 3992–3998. Babbar S. and Koufteros X., The human element in airline service quality: Contact personnel and the customer, International Journal of Operations & Production Management 28 (9–10), 2008, 804–830. Balcombe K., Fraser I. and Harris L., Consumer willingness to pay for in-flight service and comfort levels: A choice experiment, Journal of Air Transport Management 15 (5), 2009, 221–226. Barbarosoglu G. and Yazgac T., An application of the analytic hierarchy process to the supplier selection problem, Production and Inventory Management Journal (1st quarter), 1997, 14–21. Boeing (1972). Update of ATA formula for direct operating cost. Boeing Report, Washington D.C. Boetsch T., Bieger T. and Wittmer A., A customer-value framework for analyzing airline services, Transportation Journal 50 (3), 2011, 251–270. Borken-Kleefeld J., Berntsen T. and Fuglestvedt J., Specific climate impact of passenger and freight transport, Environmental Science and Technology 44 (15), 2010, 5700–5706. Bortolan G. and Degani R., A review of some methods for ranking fuzzy numbers, Fuzzy Sets and Systems 15, 1985, 1–19. Brindisi A. and Concilio A., Passengers’ comfort modeling inside aircraft, Journal of Aircraft 45 (6), 2008, 2001–2008. Brueckner J.K. and Girvin R., Airport noise regulation, airline service quality, and social welfare, Transportation Research Part B: Methodological 42 (1), 2008, 19–37. Brueckner J.K. and Zhang A., Airline emission charges: Effects on airfares, service quality, and aircraft design, Transportation Research Part B: Methodological 44 (8–9), 2010, 960–971. Brunelli M. and Mezei J., How different are ranking methods for fuzzy numbers? A numerical study, International Journal of Approximate Reasoning 54 (5), 2013, 627–639. Buyukozkan G., An integrated fuzzy multi-criteria group decision-making approach for green supplier evaluation, International Journal of Production Research 50 (11), 2012, 2892–2909. Calabrese A., Costa R. and Menichini T., Using fuzzy AHP to manage intellectual capital assets: An application to the ICT service industry, Expert Systems with Applications 40 (9), 2013, 3747–3755. Chen F.-Y. and Chang Y.-H., Examining airline service quality from a process perspective, Journal of Air Transport Management 11 (2), 2005, 79–87. Chen Y.-H. and Chao R.-J., Supplier selection using consistent fuzzy preference relations, Expert Systems with Applications 39 (3), 2012, 3233–3240. Cheng C.-H. and Chang J.-R., MCDM aggregation model using situational ME-OWA and ME-OWGA operators, International Journal of Uncertainty Fuzziness and Knowledge-Based Systems 14 (4), 2006, 421–443. Cheng S.C., Chen M.Y., Chang H.Y. and Chou T.C., Semantic-based facial expression recognition using analytical hierarchy process, Expert Systems with Applications 33 (1), 2007, 86–95. Cho J. and Lee J., Development of a new technology product evaluation model for assessing commercialization opportunities using Delphi method and fuzzy AHP approach, Expert Systems with Applications 40 (13), 2013, 5314–5330. Chou C.-C., Liu L.-J., Huang S.-F., Yih J.-M. and Han T.-C., An evaluation of airline service quality using the fuzzy weighted SERVQUAL method, Applied Soft Computing 11 (2), 2011, 2117–2128. elsevier_ESWA_9904 Curry N. and Gao Y., Low-cost airlines – A new customer relationship? An analysis of service quality, service satisfaction, and customer loyalty in a low-cost setting, Services Marketing Quarterly 33 (2), 2012, 104–118. De Jager J.W., Van Zyl D. and Toriola A.L., Airline service quality in South Africa and Italy, Journal of Air Transport Management 25, 2012, 19–21. Dray L., An analysis of the impact of aircraft lifecycles on aviation emissions mitigation policies, Journal of Air Transport Management 28, 2013, 62–69. Eagan P., Application of analytic hierarchy process techniques to streamlined life-cycle analysis of two anodizing processes, Environmental Science and Technology 33 (9), 1999, 1495–1500. Ertay T., Kahveci A. and Tabanli R.M., An integrated multi-criteria group decision-making approach to efficient supplier selection and clustering using fuzzy preference relations, International Journal of Computer Integrated Manufacturing 24 (12), 2011, 1152–1167. Esposito E. and Raffa M., The evolution of Italian subcontracting firms: empirical evidence, European Journal of Purchasing and Supply Management 1 (2), 1994, 67–76. Esposito E. and Passaro R., Material requirement planning and the supply chain at Alenia Aircraft, European Journal of Purchasing and Supply Management 3 (1), 1997, 43–51. Esposito E. and Raffa L., Global reorganisation in a high-technology industry: The aircraft industry, International Journal of Globalisation and Small Business 2 (2), 2007, 166–184. Ferreri D., Marketing and management, 2003, Praeger; Westport, Connecticut, USA. Frota J., NACRE novel aircraft concepts, Aeronautical Journal 114 (1156), 2010, 399–404. Girvin R., Aircraft noise regulation, airline service quality, and social welfare: The monopoly case, Journal of Transport Economics and Policy 44 (1), 2010, 17–35. Gomes L.F.A.M., Mattos Fernandes J.E. and Mello J.C.C., A fuzzy stochastic approach to the multicriteria selection of an aircraft for regional chartering, Journal of Advanced Transportation 48 (3), 2012, 223–237. Haddad M. and Fawaz Z., Evaluation of microalgal alternative jet fuel using the AHP method with an emphasis on the environmental and economic criteria, Environmental Progress and Sustainable Energy 32 (3), 2013, 721–733. Higgins K.M., Lawphongpanich S., Mahoney J.F. and Yin Y., Evaluating airline service quality by data envelopment analysis, Journal of the Transportation Research Board 2052 (1), 2008, 1–8. Hope, J. (2005). Emission from international air-transport and related emission policy. ICAO Report. ICAO (2013). Environmental report 2013 – Aviation and climate change. ICAO Report. Ishizaka A. and Nguyen N.H., Calibrated fuzzy AHP for current bank account selection, Expert Systems with Applications 40 (9), 2013, 3775–3783. Jiang H., Service quality of low-cost long-haul airlines – The case of Jetstar Airways and AirAsia X, Journal of Air Transport Management 26, 2013, 20–24. Kaufmann A. and Gupta M.M., Fuzzy mathematical models in engineering and management science, 1988, Elsevier Science. Kim Y.-K. and Lee H.-R., Passenger complaints under irregular airline conditions – Cross-cultural study, Journal of Air Transport Management 15 (6), 2009, 350–353. Koblen I., Szabo S. and Krnáčová K., Selected information on European Union research and development programmes and projects focused on reducing emissions from air transport, Naše more, Znanstveno-stručni časopis za more i pomorstvo 60 (5–6), 2013, 113–122. Lee D. and Luengo-Prado M.J., Are passengers willing to pay more for additional legroom?, Journal of Air Transport Management 10 (6), 2004, 377–383. Lee T.R., Nha Le T.P., Genovese A. and Koh S.C.L., Using FAHP to determine the criteria for partner’s selection within a green supply chain: The case of hand tool industry in Taiwan, Journal of Manufacturing Technology Management 23 (1), 2011, 25–55. Li Y., Liu X. and Chen Y., Supplier selection using axiomatic fuzzy set and TOPSIS methodology in supply chain management, Fuzzy Optimization and Decision Making 11 (2), 2012, 147–176. Liou J.J.H., Tsai C.-Y., Lin R.-H. and Tzeng G.-H., A modified VIKOR multiple-criteria decision method for improving domestic airlines service quality, Journal of Air Transport Management 17, 2011, 57–61. Liou J.J.H. and Tzeng G.-H., A non-additive model for evaluating airline service quality, Journal of Air Transport Management 13 (3), 2007, 131–138. elsevier_ESWA_9904 Liou J.J.H., Yen L. and Tzeng G.-H., Using decision rules to achieve mass customization of airline services, European Journal of Operational Research 205 (5), 2010, 287–288. Lu C. and Morrell P., Determination and applications of environmental costs at different sized airports – Aircraft noise and engine emissions, Transportation 33 (1), 2006, 45–61. Maji L., A note on “A modified VIKOR multiple-criteria decision method for improving domestic airlines service quality”, Journal of Air Transport Management 20, 2012, 364–368. Martin J.C., Roman C. and Espino R., Willingness to pay for airline service quality, Transport Reviews 28 (2), 2008, 199–217. Martin J.C., Roman C. and Espino R., Evaluating frequent flyer programs from the air passengers’ perspective, Journal of Air Transport Management 17 (6), 2011, 7–8. Miyoshi C. and Mason K.J., The carbon emissions of selected airlines and aircraft types in three geographic markets, Journal of Air Transport Management 15 (3), 2009, 138–147. Narasimhan R., An analytic approach to supplier selection, Journal of Purchasing and Materials Management (Winter), 1983, 27–32. Newson C. and Cairns S., Predict and decide: Aviation, climate change and UK policy, 2006, Report of UK Environmental Change Institute. Nydick R.L. and Hill R.P., Using the analytic hierarchy process to structure the supplier selection procedure, International Journal of Purchasing and Materials Management 28 (2), 1992, 31–36. Ott J., Clearing the air, Aviation Week and Space Technology 167 (8), 2007, 54–58. Pakdil F. and Aydin O., Expectations and perceptions in airline services: An analysis using weighted SERVQUAL scores, Journal of Air Transport Management 13 (4), 2007, 229–237. Park J.-W., Robertson R. and Wu C.-L., The effect of airline service quality on passengers’ behavioural intentions: A Korean case study, Journal of Air Transport Management 10 (6), 2004, 435–439. Park J.-W., Robertson R. and Wu C.-L., Modelling the impact of airline service quality and marketing variables on passengers’ future behavioural intentions, Transportation Planning and Technology 29 (5), 2006, 359–381. Park J.-W., Robertson R. and Wu C.-L., Differences in air passengers’ buying behaviour: Findings from Korean and Australian international passengers, Transportation Planning and Technology 32 (5), 2009, 441–460. Pennig S., Quehl J. and Rolny V., Effects of aircraft cabin noise on passenger comfort, Ergonomics 55 (10), 2012, 1252–1265. Price T. and Probert D., Environmental impacts of air traffic, Applied Energy 50 (2), 1995, 133–162. Rezaei J., Fahim P.B. and Tavasszy L., Supplier selection in the airline retail industry using a funnel methodology: Conjunctive screening method and fuzzy AHP, Expert Systems with Applications 41 (18), 2014, 8165–8179. Saaty T.L., The analytic hierarchy process, 1980, McGraw Hill International; New York. Saaty T.L., Decision making with Dependence and Feedback- The Analytic Network Process, 2nd ed., 2001, RWS Publications; Pittsburgh. Saaty T.L. and Tran L.T., On the invalidity of fuzzifying numerical judgments in the analytic hierarchy process, Mathematical and Computer Modelling 46 (7), 2007, 962–975. Saletic, D., Velasevic, D., Mastorakis, N. (2002). Analysis of basic defuzzification techniques. In: Proceedings of the 6th WSES international multiconference on circuits, systems, communications and computers. Schipper J., Environmental costs in European aviation, Transport Policy 11 (2), 2004, 141–154. Sen O., The effect of aircraft engine exhaust gases on the environment, International Journal of Environment and Pollution 8 (1–2), 1997, 148–157. Shaw K., Shankar R., Yadav S.S. and Thakur L.S., Supplier selection using fuzzy AHP and fuzzy multi-objective linear programming for developing low carbon supply chain, Expert Systems with Applications 39 (9), 2012, 8182–8192. Sim K.L., Koh H.C. and Shetty S., Some potential issues of service quality reporting for airlines, Journal of Air Transport Management 12 (6), 2006, 293–299. Soukup W.R., Supplier selection strategies, Journal of Purchasing and Materials Management 23 (3), 1987, 7–12. Tsaur S.-H., Chang T.-Y. and Yen C.-H., The evaluation of airline service quality by fuzzy MCDM, Tourism Management 23 (2), 2002, 107–115. Tsilingiridis J., Aircraft air pollutant emissions in Greek airports, Global Nest Journal 11 (4), 2009, 528–534. elsevier_ESWA_9904 Queries and Answers Query: Your article is registered as a regular item and is being processed for inclusion in a regular issue of the journal. If this is NOT correct and your article belongs to a Special Issue/Collection please contact c.samiullah@elsevier.com immediately prior to returning your corrections. Answer: This is correct. The article belongs to a normal issue. Query: Please confirm that given name(s) and surname(s) have been identified correctly. Answer: Names are correctly stated. Query: Please check whether the designated corresponding author is correct, and amend if necessary. Answer: Yes, Dr. Andrea Genovese is the corresponding author. Vespermann J. and Wald A., Much ado about nothing? – An analysis of economic impacts and ecologic effects of the EU-emission trading scheme in the aviation industry, Transportation Research Part A: Policy and Practice 45 (10), 2011, 1066–1076. Vink P., Bazley C., Kamp I. and Blok M., Possibilities to improve the aircraft interior comfort experience, Applied Ergonomics 43 (2), 2012, 354–359. Wang J., Fan K. and Wang W., Integration of fuzzy AHP and FPP with TOPSIS methodology for aeroengine health assessment, Expert Systems with Applications 37 (12), 2010, 8516–8526. Watanabe N., Statistical methods for estimating membership functions, Japanese Journal of Fuzzy Theory and Systems 5 (4), 1979. Wen C.-H. and Yeh W.-Y., Positioning of international air passenger carriers using multidimensional scaling and correspondence analysis, Transportation Journal 49 (1), 2010, 7–23. Wojahn O.W., The impact of passengers’ preferences regarding time and service quality on airline network structure, Journal of Transport Economics and Policy 36 (1), 2002, 139–162. Wu G., Exercises on tradeoffs and conflicting objectives, Harvard Business School Case Studies 9, 1996, 307–396. Yang K.-C., Hsieh T.-C., Li H. and Yang C., Assessing how service quality, airline image and customer value affect the intentions of passengers regarding low cost carriers, Journal of Air Transport Management 20, 2012, 52–53. Yeh C.-H. and Chang Y.-H., Modeling subjective evaluation for fuzzy group multicriteria decision making, European Journal of Operational Research 194, 2009, 464–473. Zadeh L.A., Fuzzy sets, Information and Control 8, 1965, 338–353. Zeydan M., Colpan C. and Cobanoglu C., A combined methodology for supplier selection and performance evaluation, Expert Systems with Applications 38 (3), 2011, 2741–2751. Zhang Y., Are Chinese passengers willing to pay more for better air services?, Journal of Air Transport Management 25, 2012, 5–7. Highlights • This paper presents a novel multi-criteria decision-making model for aircraft evaluation. • The model is based on a combination of Analytic Hierarchy Process and Fuzzy Set Theory. • A variety of criteria aligned to contemporary airlines’ requirements is included. • The model overcomes limitations of existing approaches proposed to deal with the problem. • A real-world empirical study testifies the usability and usefulness of the model. elsevier_ESWA_9904 Query: References “Gibson and Morrell (2004), See et al. (2004), Bruno et al. (2012)” are cited in the text but not provided in the reference list. Please provide them in the reference list or delete these citations from the text. Answer: Required references can be found below: Bruno, G., Esposito, E., Genovese, A., & Passaro, R. (2012). AHP-based approaches for supplier evaluation: Problems and perspectives. Gibson, W., & Morrell, P. (2004). Theory and practice in aircraft financial evaluation. See, T. K., Gurnani, A., & Lewis, K. (2004). Multi-attribute decision making using hypothetical equivalents and inequivalents. Query: Please provide an update for reference “Abdullah and Zulkifli (in press)”. Answer: The updated reference can be found below: Abdullah, L., & Zulkifli, N. (2015). Integration of fuzzy AHP and interval type-2 fuzzy DEMATEL: An application to human resource management. Expert Systems with Applications. Volume 42, Issue 9, pp. 4397–4409 Query: The decimal comma has been changed to a decimal point in Tables 1, 6, 7. Please check, and correct if necessary. Answer: All changes are correct. elsevier_ESWA_9904 http://www.sciencedirect.com/science/journal/09574174/42/9 
work_u2ztrkxxu5cspbri4ek4xp45fm	----	SHYSTER: A Pragmatic Legal Expert System SHYSTER: A Pragmatic Legal Expert System JAMES DAVID POPPLE A thesis submitted for the degree of Doctor of Philosophy of The Australian National University April 1993 c� James Popple 1993 Doonesbury (page 131) copyright G. B. Trudeau Reprinted with permission of Universal Press Syndicate All rights reserved Earlier versions of some parts of this thesis have appeared in: Proceedings of the Thirteenth Australian Computer Science Conference c� Australian Computer Science Association 1990 Advances in Computing and Information: Proceedings of the International Conference on Computing and Information c� International Conference on Computing and Information 1990 The Australian Computer Journal c� Australian Computer Society Inc. 1991 National Library of Australia Cataloguing-in-Publication entry Popple, James David, 1964– . SHYSTER: a pragmatic legal expert system. Bibliography. Includes indexes. ISBN 0 7315 1827 6. 1. SHYSTER (Computer file). 2. Law — Methodology — Data processing. 3. Expert systems (Computer science). I. Australian National University. Faculty of Engineering and Information Technology. Dept of Computer Science. II. Title. 340.11 Except where otherwise indicated, this thesis is my own original work. James Popple 29 April 1993 shy.ster \�sh̄ıst�(r)\ n -s [prob. after Scheuster fl 1840 Am. attorney frequently rebuked in a New York court for pettifoggery] : one who is professionally unscrupulous esp. in the practice of law or politics . . . Webster’s Third New International Dictionary (1961)1 shyster (��a�st�(r)) . . . [Of obscure origin. It might be f. shy a. (sense 7, disreputable) + -ster ; but this sense of the adj. is app. not current in the U.S.] . . . ‘A lawyer who practises in an unprofessional or tricky manner; especially, one who haunts the prisons and lower courts to prey on petty criminals; hence, any one who conducts his business in a tricky manner’ (Funk’s Stand. Dict. 1895). Also attrib. or adj. Orig. and chiefly U.S. slang . . . The Oxford English Dictionary (1989)2 shyster. An unscrupulous lawyer (note that the definition presumes the existence of scrupulous ones) . . . The term does not come from—as suggested in various dictionaries—the surname Scheuster, supposedly a lawyer noted for shyster-like practices; from the name of the Shakespearean char­ acter, Shylock; . . . or from any of the various meanings of shy (e.g., to be shy of money). Rather . . . shyster evolved from the underworld use of shiser, a worthless fellow, which derived in turn from the German scheisse, excrement, via scheisser, an incompetent person (specifically, one who cannot control his bodily functions) . . . A Dictionary of Invective (1991)3 v Abstract Most legal expert systems attempt to implement complex models of legal reas­ oning. But the utility of a legal expert system lies not in the extent to which it simulates a lawyer’s approach to a legal problem, but in the quality of its predic­ tions and of its arguments. A complex model of legal reasoning is not necessary: a successful legal expert system can be based upon a simplified model of legal reasoning. Some researchers have based their systems upon a jurisprudential approach to the law, yet lawyers are patently able to operate without any jurisprudential insight. A useful legal expert system should be capable of producing advice similar to that which one might get from a lawyer, so it should operate at the same pragmatic level of abstraction as does a lawyer—not at the more philosophical level of jurisprudence. A legal expert system called SHYSTER has been developed to demonstrate that a useful legal expert system can be based upon a pragmatic approach to the law. SHYSTER has a simple representation structure which simplifies the problem of knowledge acquisition. Yet this structure is complex enough for SHYSTER to produce useful advice. SHYSTER is a case-based legal expert system (although it has been designed so that it can be linked with a rule-based system to form a hybrid legal expert system). Its advice is based upon an examination of, and an argument about, the similarities and differences between cases. SHYSTER attempts to model the way in which lawyers argue with cases, but it does not attempt to model the way in which lawyers decide which cases to use in those arguments. Instead, it employs statistical techniques to quantify the similarity between cases. It decides which cases to use in argument, and what prediction it will make, on the basis of that similarity measure. vii viii Abstract SHYSTER is of a general design: it can provide advice in areas of case law that have been specified by a legal expert using a specification language. Hence, it can operate in different legal domains. Four different, and disparate, areas of law have been specified for SHYSTER, and its operation has been tested in each of those domains. Testing of SHYSTER in these four domains indicates that it is exception­ ally good at predicting results, and fairly good at choosing cases with which to construct its arguments. SHYSTER demonstrates the viability of a pragmatic approach to legal expert system design. Acknowledgements Thanks . . . to the members of my supervisory panel—Robin Stanton, Roger Clarke, Peter Drahos and Malcolm Newey: for their counsel and encouragement. to Robin Creyke and Phillipa Weeks: for giving so freely of their time and legal expertise. to Peter Bailey, David Cullen, Kate Lazenby-Cohen and Kevin Popple: for their comments on various parts of various drafts. to Graham Jefferson, Chris Johnson, Brendan McKay, Gavin Michael, Bob Moles, Neil Tennant and Trevor Vickers: for their suggestions and advice. to Sally Begg, Jeannie Haxell, Bev Johnstone and Jacquie Wilson: for adminis­ trative assistance. to David Hawking and the programming staff: for technical assistance. to Richard Walker and Markus Zellner: for TEXnical assistance. to Zdzis�law �Loboz, Rachel Oñate, Dharmendra Sharma, Wang Xiaoji and Yang Jian: for sharing, with me, different offices at different times. to Diane Hutchens: for all her help during my time with PARSA. to my parents—Marli and Kevin: for their unfailing love and support. and to Paula Fearn: for whom, but for whom . . . This research was supported by an Australian National University PhD Scholarship, funded by the Centre for Information Science Research. Additional financial support was provided by Sigma Data Corporation. ix Contents Abstract vii Acknowledgements ix Figures xxi Chapter 1: Introduction 3 1.1 The aims of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.2 SHYSTER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.3 The structure of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Chapter 2: Legal analysis systems 7 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.1.1 Sources of law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.1.2 The doctrine of precedent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.1.3 The structure of this chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.1.4 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.2 Jurisprudence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.2.1 The importance of jurisprudence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.2.2 Scientific and mechanical jurisprudence . . . . . . . . . . . . . . . . . . . . . . . 13 2.2.3 Judgment machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.2.4 Petrifaction of the law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.2.5 Clear rules and clear cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2.6 Legal realism and rule scepticism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.2.7 A jurisprudential consensus? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.3 Jurimetrics and the behaviourists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.3.1 Kort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.3.2 Lawlor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 xi xii Contents 2.3.3 Nagel and Schubert . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.3.4 Haar, Sawyer and Cummings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.3.5 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.4 Rule-based systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.4.1 McCarty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 2.4.2 Bench-Capon, Kowalski and Sergot . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.4.3 Gardner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.4.4 Susskind . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.4.5 Knowledge acquisition and representation . . . . . . . . . . . . . . . . . . . . . 36 2.4.6 Fact representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 2.4.7 The inadequacy of rules for case law . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.5 Case-based systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 2.5.1 Nearest neighbour analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 2.5.2 FINDER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 2.5.3 HYPO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 2.5.4 The inadequacy of semantic networks . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.6 Hybrid systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.7 Conceptual models: deep and shallow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 Chapter 3: A pragmatic approach to case law 51 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.2 Design criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.2.1 Users and output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.2.2 A pragmatic model of legal reasoning . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.2.3 Knowledge representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.2.4 Generality of application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.2.5 Prediction and argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.2.6 Evaluation of advice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.2.7 A hybrid structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.3 A model of legal reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.3.1 Private and public law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.3.2 The functions of legal reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.3.3 Adopting a model of legal reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.4 Experts, users, and expert users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 3.5 Knowledge representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 3.5.1 Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 3.5.2 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 3.5.3 Leading cases and attribute values . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 3.5.4 Ideal points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 3.6 A specification language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 Contents xiii 3.7 Weighting attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 3.8 Detecting attribute dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 3.8.1 Functional dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 3.8.2 Stochastic dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 3.9 Calculating distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 3.9.1 Similarity measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 Distance measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 Association coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 Correlation coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 3.9.2 Weighted similarity measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 3.9.3 Choosing a similarity measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 3.9.4 Infinite distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 3.10 Nearest cases and nearest results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 3.10.1 Nearest cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 3.10.2 Nearest results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 3.10.3 Equidistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 3.11 Writing a report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 3.11.1 Arguing with the instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Choosing cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Using cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 3.11.2 Arguing with instantiations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 3.11.3 Arguing with hypotheticals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 3.12 Safeguards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 3.12.1 Extra similarity measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 3.12.2 Ideal points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 3.12.3 Centroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 3.12.4 Attribute direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 3.12.5 Equidistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 3.12.6 Instantiations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 3.13 Testing and evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 3.13.1 Choosing a test domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 3.13.2 Evaluating SHYSTER’s opinion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 3.13.3 The paucity of test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 3.13.4 Generated tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 3.13.5 Reflexive tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 3.14 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 3.14.1 Comparisons with other approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 98 3.14.2 A Sisyphean journey? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 Chapter 4: Implementing SHYSTER 103 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 4.1.1 Internal representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 4.1.2 Output files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 4.2 The Shyster module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 4.3 The Statutes module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 xiv Contents 4.4 The Cases module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 4.5 The Tokenizer module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 4.5.1 Tokens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 4.5.2 Comments and whitespace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 4.6 The Parser module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 4.6.1 Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 4.6.2 Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 4.6.3 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 Local attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 External attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 4.6.4 Leading cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 4.6.5 Ideal points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 4.7 The Dumper module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 4.7.1 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 Local attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 External attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 4.7.2 Leading cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 4.7.3 Ideal points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 4.8 The Checker module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 4.8.1 Detecting dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 4.8.2 Writing matrices of probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 4.9 The Scales module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.9.1 Calculating weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.9.2 Writing tables of weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.10 The Adjuster module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.11 The Consultant module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.12 The Odometer module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 4.12.1 Calculating distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.12.2 Writing tables of distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.13 The Reporter module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 4.13.1 Arguing with the instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 Arguing for the nearest result . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 Arguing for other results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.13.2 Arguing with instantiations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.13.3 Arguing with hypotheticals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.13.4 Safeguards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 4.14 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 Chapter 5: Case studies 131 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 5.2 A simulation of FINDER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 5.2.1 The law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 Contents xv 5.2.2 The Finder specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 Leading cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 Attribute dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 5.2.3 Test case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 Parker v. British Airways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 5.2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 5.3 The authorization of copyright infringement . . . . . . . . . . . . . . . . . . . . . . . . . . 140 5.3.1 The law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 5.3.2 The Authorization specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 Leading cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 Attribute dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 5.3.3 Test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 CBS Inc. v. Ames Records and Tapes Ltd . . . . . . . . . . . . . . . . . . 148 WEA International Inc. v. Hanimex Corporation Ltd . . . . . . . 149 CBS Songs Ltd v. Amstrad Consumer Electronics plc . . . . . . . 151 Australasian Performing Right Association Ltd v. Jain . . . . . . 152 Hypothetical case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 Hypothetical case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 5.3.4 Generated tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 Generated test 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 Generated test 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 Generated test 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 5.3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 5.4 Employees and independent contractors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 5.4.1 The law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 5.4.2 The Employee specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 Leading cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 Attribute dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 5.4.3 Test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 Narich Pty Ltd v. Commissioner of Pay-roll Tax . . . . . . . . . . . 160 Stevens v. Brodribb Sawmilling Company Pty Ltd . . . . . . . . . . . 163 Re Porter; Re Transport Workers Union of Australia . . . . . . . 163 Building Workers’ Industrial Union of Australia v. Odco Pty Ltd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 Hypothetical case 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 Hypothetical case 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 xvi Contents 5.4.4 Generated tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 Generated test 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 Generated test 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 Generated test 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 5.4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 5.5 The implication of natural justice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 5.5.1 The law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 5.5.2 The Natural specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 Natural area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 Affected area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 Expectation area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 Leading cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Attribute dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 5.5.3 Test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 Twist v. Council of Municipality of Randwick . . . . . . . . . . . . . . 179 Salemi v. MacKellar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 Heatley v. Tasmanian Racing and Gaming Commission . . . . . 183 Cole v. Cunningham . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 Ackroyd v. Whitehouse (Director of National Parks & Wildlife Service) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 Hodgens v. Gunn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 Western Australia v. Bropho . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 Ainsworth v. Criminal Justice Commission . . . . . . . . . . . . . . . . 187 5.5.4 Generated tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 Generated test 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 Generated test 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 5.5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 5.6.1 Test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 5.6.2 Generated tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 5.6.3 Reflexive tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 Chapter 6: Conclusion 195 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 6.2 Evaluating SHYSTER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 6.2.1 A useful, working system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 6.2.2 A general approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 6.2.3 Good advice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 6.2.4 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 6.2.5 Possible enhancements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 6.3 Future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 6.4 The contribution of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 Contents xvii Appendix A: Case law specifications 205 A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 A.2 The Finder specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 Finder area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 Cases in which the finder wins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 Cases in which the finder loses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 A.3 The Authorization specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 Authorization area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 Opening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 Cases in which the accused authorized the infringement . . . . . . . . . 219 Cases in which the accused did not authorize the infringement . . . 222 Cases in which the accused is liable (directly or vicariously) for the infringement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 A.4 The Employee specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 Employee area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 Opening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 Cases in which the worker is an employee . . . . . . . . . . . . . . . . . . . . . 231 Cases in which the worker is an independent contractor . . . . . . . . . 240 A.5 The Natural specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 Cases in which the decision affected the property, right, interest, Cases in which the decision did not affect the property, right, Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 Natural area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 Opening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250 Cases in which a duty to observe natural justice is implied . . . . . . 252 Cases in which a duty to observe natural justice is not implied . . . 259 Affected area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266 status, or legitimate expectation of the applicant . . . . . . . . 267 interest, status, or legitimate expectation of the applicant 273 xviii Contents Expectation area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274 Cases in which the applicant had a legitimate expectation Cases in which the applicant did not have a legitimate Opening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275 which was affected by the decision . . . . . . . . . . . . . . . . . . . . 276 expectation which was affected by the decision . . . . . . . . . . 280 Appendix B: Example reports 283 B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 B.2 Report file for Parker v. British Airways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 Finder area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 Instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 Hypothetical 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286 Hypothetical 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288 B.3 Report file for APRA v. Jain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 Authorization area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 Instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 Hypothetical 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292 Hypothetical 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 Hypothetical 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294 B.4 Report file for Re Porter; Re TWU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296 Employee area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296 Instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296 Instantiation 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298 Hypothetical 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299 Hypothetical 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301 B.5 Report files for Ainsworth v. CJC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302 Expectation area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302 Instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302 Hypothetical 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 Affected area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 Instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 Hypothetical 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 Natural area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307 Instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307 Hypothetical 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 Appendix C: A complete example 311 C.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312 C.2 Log file for BWIU v. Odco . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312 C.3 Employee specification file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 Contents xix C.4 Dump file for Employee specification (LaTEX input) . . . . . . . . . . . . . . . . . . . 328 C.5 Dump file for Employee specification (LaTEX output) [see §A.4] . . . . . . . . . 331 C.6 Probabilities file for Employee specification (LaTEX input) . . . . . . . . . . . . . . 332 C.7 Probabilities file for Employee specification (LaTEX output) . . . . . . . . . . . . . 333 Employee area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333 C.8 Weights file for Employee specification (LaTEX input) . . . . . . . . . . . . . . . . . . 334 C.9 Weights file for Employee specification (LaTEX output) . . . . . . . . . . . . . . . . . 335 Employee area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335 C.10 Distances file for BWIU v. Odco (LaTEX input) . . . . . . . . . . . . . . . . . . . . . . . . 335 C.11 Distances file for BWIU v. Odco (LaTEX output) . . . . . . . . . . . . . . . . . . . . . . . 336 Employee area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337 Instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337 Instantiation 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338 Instantiation 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339 Instantiation 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340 Instantiation 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341 Hypothetical 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342 Hypothetical 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343 C.12 Report file for BWIU v. Odco (LaTEX input) . . . . . . . . . . . . . . . . . . . . . . . . . . 344 C.13 Report file for BWIU v. Odco (LaTEX output) . . . . . . . . . . . . . . . . . . . . . . . . . 346 Employee area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346 Instant case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346 Hypothetical 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348 Hypothetical 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350 Appendix D: Reflexive tests 353 D.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354 D.2 The Finder specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354 Hannah v. Peel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354 Bridges v. Hawkesworth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354 Armory v. Delamirie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355 Moffatt v. Kazana . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355 City of London Corporation v. Appleyard (1) . . . . . . . . . . . . . . . . . . . . . . 355 City of London Corporation v. Appleyard (2) . . . . . . . . . . . . . . . . . . . . . . 356 South Staffordshire Water Co. v. Sharman . . . . . . . . . . . . . . . . . . . . . . . . 356 Elwes v. Brigg Gas Co. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356 D.3 The Authorization specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357 University of New South Wales v. Moorhouse . . . . . . . . . . . . . . . . . . . . . 357 Australasian Performing Right Association Ltd v. Canterbury- Bankstown League Club Ltd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357 Winstone v. Wurlitzer Automatic Phonograph Co. of Australia Pty Ltd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358 Mellor v. Australian Broadcasting Commission . . . . . . . . . . . . . . . . . . . . 358 xx Contents Falcon v. Famous Players Film Co. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358 RCA Corporation v. John Fairfax and Sons Ltd . . . . . . . . . . . . . . . . . . . 359 Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd . . . . . . 359 A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd . . . . . . . . . . . . . . 359 Australasian Performing Right Association Ltd v. Miles . . . . . . . . . . . . 360 D.4 The Employee specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361 Federal Commissioner of Taxation v. J. Walter Thompson Performing Right Society Ltd v. Mitchell & Booker (Palais de Ready Mixed Concrete (South East) Ltd v. Minister of Pensions Zuijs v. Wirth Brothers Pty Ltd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361 Cam and Sons Pty Ltd v. Sargent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361 (Australia) Pty Ltd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362 Australian Timber Workers Union v. Monaro Sawmills Pty Ltd . . . . . 362 Ferguson v. John Dawson & Partners (Contractors) Ltd . . . . . . . . . . . 362 Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (2) . . . . 363 Danse) Ltd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363 Humberstone v. Northern Timber Mills . . . . . . . . . . . . . . . . . . . . . . . . . . . 363 Queensland Stations Pty Ltd v. Federal Commissioner of Taxation . . . 364 Price v. Grant Industries Pty Ltd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364 Australian Mutual Provident Society v. Chaplin . . . . . . . . . . . . . . . . . . . 364 Massey v. Crown Life Insurance Co. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364 Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (1) . . . . 365 and National Insurance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366 D.5 The Natural specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366 FAI Insurances Ltd v. Winneke . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366 Haoucher v. Minister of State for Immigration and Ethnic Affairs . . . 366 Annetts v. McCann . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 Kioa v. West . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 Commissioner of Police v. Tanos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 Marine Hull & Liability Insurance Co. Ltd v. Hurford . . . . . . . . . . . . . . 367 Macrae v. Attorney-General for New South Wales . . . . . . . . . . . . . . . . . 368 Attorney-General of Hong Kong v. Ng Yuen Shiu . . . . . . . . . . . . . . . . . . 368 Durayappah v. Fernando . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 South Australia v. O’Shea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369 Bread Manufacturers of New South Wales v. Evans . . . . . . . . . . . . . . . . 369 Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd . . 369 Nashua Australia Pty Ltd v. Channon . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370 Council of Civil Service Unions v. Minister for the Civil Service . . . . . 370 McInnes v. Onslow Fane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370 Notes 375 Cases 403 Statutes 409 Bibliography 411 Figures Figures marked with a star are extracted from SHYSTER output. 3.1 Forms of functional dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 3.2 Example attribute values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 3.3 Probabilities for example attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 3.4 Two-way association table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 3.5 Example distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 3.6 SHYSTER’s algorithm for choosing cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 3.7 SHYSTER’s algorithm for using cases: part A . . . . . . . . . . . . . . . . . . . . . . . . . 82 3.8 SHYSTER’s algorithm for using cases: part B . . . . . . . . . . . . . . . . . . . . . . . . . 83 3.9 SHYSTER’s algorithm for using cases: part C . . . . . . . . . . . . . . . . . . . . . . . . . 84 4.1 SHYSTER’s command line switches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 4.2 Keywords in SHYSTER’s case law specification language . . . . . . . . . . . . . . . . 108 4.3 EBNF description of SHYSTER’s case law specification language . . . . . . . . . 110 � 4.4 Attribute values for Employee area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 � 4.5 Extract from probabilities for Employee area . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 4.6 Similarity measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 � 4.7 Distances for BWIU v. Odco . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 � 4.8 Extract from log file for BWIU v. Odco . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 � 5.1 Probabilities for Finder area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 � 5.2 Weights for Finder area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 � 5.3 Distances for Parker v. British Airways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 � 5.4 Probabilities for Authorization area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 � 5.5 Weights for Authorization area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 � 5.6 Distances for CBS v. Ames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 � 5.7 Distances for Narich v. CPT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 5.8 Relationship between areas in Natural specification . . . . . . . . . . . . . . . . . . . 173 � 5.9 Probabilities for Natural specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 xxi xxii Figures � 5.10 Weights for Natural specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 � 5.11 Distances for Twist v. Randwick , Natural area . . . . . . . . . . . . . . . . . . . . . . . . . 181 5.12 Summary of test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 5.13 Summary of generated testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 D.1 Summary of reflexive testing of Finder specification . . . . . . . . . . . . . . . . . . . 356 D.2 Summary of reflexive testing of Authorization specification . . . . . . . . . . . . 360 D.3 Summary of reflexive testing of Employee specification . . . . . . . . . . . . . . . . . 365 D.4 Summary of reflexive testing of Natural specification . . . . . . . . . . . . . . . . . . 371 Chapters 1 1 Introduction Dicke the Butcher: “The first thing we do, let’s kill all the Lawyers.” William Shakespeare (1591) Henry VI, Part 2 4 Attending a Cabinet when there was a tendency to Mutiny in the Fleet, Sir Thomas Troubridge, . . . who was a most excellent Officer, was asked his opinion what was best to be done. He said let me hang a hundred Lawyers, and we shall hear no more of the business. I asked what he could mean—what were these People that he called Lawyers. He replied, Fellows that can read and write. They are the Fellows, that I call Lawyers, and make the whole of the Mischief. Lord Eldon (1827) Lord Eldon’s Anecdote Book 5 To his colleagues, [he] was a lonely and solitary nut. To an objective outsider he might have seemed a bold crusader, waging war against the forces of darkness. The truth was somewhere in between. He was in the early stages of a PhD thesis. Alan Plater (1985) The Beiderbecke Affair 6 3 4 Chapter 1: Introduction 1.1 The aims of this thesis The history of the development of legal expert systems has, for the most part, been characterized by the development and implementation of complex models of legal reasoning. This thesis aims to show that a legal expert system need not be based upon a complex model of legal reasoning in order to produce useful advice. It advocates a pragmatic approach to legal expert system design based on the way in which lawyers deal with the law on a day-to-day basis. It argues that a system based upon a simple model of legal reasoning can still produce good advice, where that advice is evaluated by reference to the accuracy of its predictions and to the quality of its arguments. Furthermore, such a sys­ tem, with its simpler knowledge representation structure, makes commensurately simpler the process of knowledge acquisition. These arguments are made theoretically, and then by example. A legal expert system which is based on a pragmatic approach to the law has been developed by the author. The development and testing of that system, called SHYSTER, is described in this thesis. 1.2 SHYSTER SHYSTER was developed to demonstrate that a useful, working legal expert sys­ tem could be based upon a pragmatic approach to the law. SHYSTER is of a general design, so that it can operate in different legal domains. It was designed to provide advice in areas of case law that have been specified by a legal expert using a specially developed specification language. SHYSTER is a case-based legal expert system. Its knowledge of the law is acquired, and represented, as information about cases. It produces its advice by examining, and arguing about, the similarities and differences between cases. By contrast, a rule-based expert system represents the law using rules. A hybrid system uses both rule-based and case-based techniques. SHYSTER has been de­ signed so that it can be linked with a rule-based system to form a hybrid legal expert system. Although SHYSTER attempts to model the way in which lawyers argue with cases, it does not attempt to model the way in which lawyers decide which cases to use in those arguments. It uses statistical techniques to quantify the similarity between cases, and chooses cases on the basis of that similarity measure. SHYSTER’s representation structure was designed so as to be as simple as possible while complex enough to allow SHYSTER to produce good advice. This simple structure greatly simplifies the process of knowledge acquisition. � 1.3 The structure of this thesis 5 1.3 The structure of this thesis The body of this thesis is divided into six chapters: • Chapter 2 discusses previous work of relevance to the development of legal analysis systems, especially legal expert systems. The value of jurisprudence to legal expert system development is also discussed, and the adoption of a pragmatic approach (as opposed to a jurisprudentially pure approach) is recommended. This approach holds that legal expert systems should operate at the same level of abstraction at which lawyers operate on a day-to-day basis. • In chapter 3, a pragmatic approach is proposed for developing legal expert systems. It incorporates a simple model of legal reasoning, and uses a simple knowledge representation structure. Comparisons are made between this approach, and approaches adopted by other legal expert system developers. This approach is adopted for the development of SHYSTER. Specific design criteria for SHYSTER are detailed, and methods of testing and evaluating the system are discussed. • The implementation of SHYSTER is explained in chapter 4. The twelve modules that comprise SHYSTER are described, and demonstrated using examples. • Four different specifications have been written for SHYSTER, and these are used as the basis of case studies in chapter 5. Each specification represents a different area of case law. Several different methods are employed to test SHYSTER and these specifications. • In chapter 6, conclusions are drawn about SHYSTER and its approach to case law. Some enhancements to SHYSTER are suggested, and avenues of future research are identified. Finally, the nature of the contribution made by this thesis is discussed. There are four appendices to this thesis. Appendix A contains each of the four specifications used to test SHYSTER in chapter 5. Six example reports, demonstrating the use of each of these four specifications, are given in appendix B. Each report is SHYSTER’s opinion on one of the test cases used in chapter 5. A complete example of SHYSTER’s input and output files for another of those test cases is given in appendix C. Appendix D gives details of one of the methods of testing, as applied to each of the four specifications. The thesis concludes with the endnotes to the chapters and appendices, a table of cases, a table of statutes, and a list of bibliographical references. 2 Legal analysis systems In the yahoo world of computer public relations, it is often the loudest mouths which lead the unsuspecting computer user into that lonely canyon of empty pockets and broken promises. It now seems . . . that parts of the academic world are fast approaching that decibel level so far achieved only by computer salesmen. Thus . . . it seems that space can only be booked on the band-wagon if one is prepared to make more outrageous claims than the next man. And such claims are being made for the application of AI to the law. Philip Leith (1986)7 The computer scientists, encouraged by the modern positivists, fail to recognize . . . that law, positive morality and ethics are inseparably connected parts of a vast organic whole. Judgments are involved at every stage of the legal process and machines cannot make judgments. In stating that legal rules can be applied without further judgment; that they apply in an all or nothing fashion; that legal decision making follows the form of the syllogism or that it is a pattern- matching routine, the modern positivists, joined now by the computer scientists take us along a dangerous road. Robert N. Moles (1987) Definition and Rule in Legal Theory: A Reassessment of H. L. A. Hart and the Positivist Tradition8 Aussi, lorsqu’un homme se rend plus Thus when a man takes on absolute power, he first absolu, songe-t-il d’abord ` thinks of simplifying the law. a simplifier In such a state one ´ les lois. On commence, dans cet Etat, begins to be more affected by technicalities than by à être plus frappé des inconvénients the freedom of the people, about which one no longer particuliers, que de la liberté des sujets cares at all. dont on ne se soucie point du tout. Montesquieu (1748) De l’Esprit des lois9 7 � � � 8 Chapter 2: Legal analysis systems 2.1 Introduction The range of computer applications in the law is wide. It extends from general applications, of use to lawyers, to applications designed specifically for the law. This thesis is concerned only with a subset of those systems that make use of artificial intelligence (AI) techniques to solve legal problems. Legal AI systems can usefully be divided into two categories: legal retrieval systems and legal analysis systems.10 Legal retrieval systems allow lawyers to search through databases containing details of statutes and decided cases. AI techniques may be employed to simplify this task: e.g. by searching for keywords which have not been input by the user but are deduced to be equivalent to, or sufficiently related to, the input keywords.11 Legal analysis systems take inform­ ation about a set of facts and determine the ramifications of those facts in a given area of law. Mehl claims that there is no fundamental difference between these two cat­ egories—that the difference is one of degree only.12 Shannon and Golshani suggest that the difference between systems based on a “conceptual model of legal ana­ lysis” and text retrieval systems is that the latter do not “understand” any area of the law.13 Similarly, Susskind says that “knowledge-based” systems, as opposed to “database” systems are capable of “applying their knowledge of the law to the problem data presented to them.”14 A better distinction—one which avoids the vexed question of whether any AI system can be said to really know or understand anything—can be made by reference to the output of the system. The output from a legal analysis system is such that, if it had been produced by a human, that human would be said to have legal expertise. By contrast, the output from a legal retrieval system could be produced by a human possessed of no legal expertise; such output is used in, and is not the product of, legal analysis.15 This thesis is concerned only with legal analysis systems. Legal analysis systems can be divided into two categories: judgment machines and legal expert systems. A judgment machine is a machine designed to replace a human judge. Such machines were first proposed over forty years ago, though no such proposals have been made in the last decade. Writings on judgment machines are discussed in §2.2.3 because they are of historical interest, and because the idea of a judgment machine raises some issues which are also relevant to the second category of legal analysis systems: legal expert systems. A legal expert system, as the term is used in this thesis, is a system cap­ able of performing at a level expected of a lawyer. AI systems which merely assist a lawyer in coming to legal conclusions or preparing legal argument are not here considered to be legal expert systems; a legal expert system must exhibit 9 � 2.1 Introduction some legal expertise itself. This definition does not exclude systems that can only be used by legal experts; several systems—including SHYSTER—have been developed for exclusive use by lawyers.16 2.1.1 Sources of law The law in so-called “common law countries”—e.g. Australia, Britain, Canada, USA—is derived from legislation and from case law. Legislation, also referred to as statute law, consists of statutes and delegated legislation. In Australia, statutes are made by Federal Parliament, or a State Parliament, and enacted when given royal assent by the Governor-General, or a State Governor. Delegated legislation (rules, regulations, ordinances, by-laws, etc.) is made by a person or body to whom legislative power has been delegated by statute. Case law, or the common law, is judge-made law: judicial resolutions of spe­ cific disputes. The sources of case law are the published reports of the cases as heard before various courts. Legislation takes precedence over case law;17 parliaments can override judge- made law by legislative enactment. However, judges have the task of determining the meaning of legislation. A legal analysis system that seeks to deal with the law in a common law country must account for both statute law and case law. 2.1.2 The doctrine of precedent The development of case law is based upon the principle of stare decisis which holds that courts should apply the doctrine of precedent. Morris et al. summarize the general rules of the doctrine of precedent as follows: • each court is bound by decisions of courts higher in its hierarchy; • a decision of a court in a different hierarchy may be of considerable weight, but will not be binding; • only the ratio decidendi of a case is binding; • any relevant decisions, although not binding, may be considered and fol­ lowed; and precedents are not necessarily abrogated by lapse of time.18 • (The ratio decidendi of a case is, literally, the reason for deciding; rationes are “pronouncements of legal principle necessary for the judge’s decision on the es­ tablished facts of the case”. 19 They are different from pronouncements of legal principle which may be illustrative or clarifying, but are not strictly relevant to the case in issue. These are called obiter dicta, and are not binding on other courts.) 10 Chapter 2: Legal analysis systems Courts have a tendency to follow similar cases, even when not strictly bound to do so. According to Cross: It is a basic principle of the administration of justice that like cases should be decided alike. This is enough to account for the fact that, in almost every jurisdiction, a judge tends to decide a case in the same way as that in which a similar case has been decided by another judge.20 There is considerable argument about the extent to which judges actually apply the doctrine of precedent.21 Some legal theorists contend that the doctrine is simply part of the public discourse that judges use to justify their decisions. Stone claims that “the degree of certainty and stability in the law secured by the doctrine of stare decisis is far less than it appears to be,” and that one of the important social functions of the doctrine is “to maintain at a maximum the feeling and appearance of certainty and stability.”22 Yet, judges of the highest court in Australia have been known to follow previous cases which they believed to be wrongly decided.23 What is clear is that judges use previously decided cases to justify their decisions—if not to reach them—and that lawyers use previously decided cases in legal argument. Any legal analysis system which deals with case law assumes the application of the doctrine of precedent, at least to this extent. 2.1.3 The structure of this chapter This chapter reviews previous work of relevance to the design of legal expert systems. Jurisprudence and its value to the development of such systems is discussed (§2.2) and the field of “jurimetrics” and some behaviouristic research is examined (§2.3). The bulk of this chapter concerns the development of legal expert systems. This is discussed under three headings: rule-based, case-based, and hybrid systems. Rule-based systems are examined first (§2.4), and special attention is paid to the work of four projects (§2.4.1–§2.4.4). The problems of knowledge acquisition and representation, and fact representation are also discussed (§2.4.5 and §2.4.6). Rule-based systems have been used to represent statutes and cases. However, as explained in §2.4.7, rule-based systems are fundamentally inadequate for repres­ enting case law. Case-based systems are examined (§2.5) and the work of three projects is focused upon (§2.5.1–§2.5.3). Systems that use rules in order to represent case law are considered to be rule-based systems; only systems which adopt case-based reasoning methods are considered to be case-based. Applying this distinction is usually straightforward, though not always.24 Attempts have been made to use semantic networks to represent case law but, as explained in §2.5.4, they are not suited to the task. Hybrid systems (§2.6) employ both rule-based and case-based methods. � 2.2 Jurisprudence 11 The search for conceptual models of legal reasoning is discussed (§2.7). Finally, conclusions are drawn from the literature as to the best approach to the design of legal expert systems (§2.8). (The development of legal expert systems raises legal questions itself. For example, who—if anyone—is liable for “bad” advice provided by such a sys­ tem? These questions are beyond the scope of this thesis, but are examined elsewhere.25) 2.1.4 Terminology For consistency, the following three terms are used wherever possible in this thesis for a variety of terms that are used throughout the literature. An attribute is a legally important fact; some use the words “descriptor,” “dimension,” “factor,” “feature” or “variable.” The instant case is the situation about which the expert system is interrogated: this is sometimes called the “actual case,” the “current fact situation,” the “hypothetical case” or the “new case.” Leading cases make up the case base of a legal expert system; some call these the “reference set” or the “training set.” 2.2 Jurisprudence For centuries, lawyers and philosophers have written on, and argued about, the nature of human laws. This field is called jurisprudence. It would seem sensible for the developer of a legal expert system to have regard to jurisprudential the­ ory before designing her/his system. However, as Susskind complained in 1987, most of the published research on various legal expert systems makes no use of jurisprudential resources.26 2.2.1 The importance of jurisprudence Some expert system researchers have been doubtful as to the value of jurispru­ dence; Niblett claims that “a successful expert system is likely to contribute more to jurisprudence than the other way round”. 27 Susskind disagrees: It is beyond argument . . . that all expert systems must conform to some jurisprudential theory because all expert systems in law necessarily make assumptions about the nature of law and legal reasoning. To be more spe­ cific, all expert systems must embody theories of legal knowledge, legal science, the structure of rules, the individuation of laws, legal systems and sub-systems, legal reasoning, and of logic and the law (as well perhaps as elements of a semantic theory, a sociology, and a psychology of law), theor­ ies that must all themselves rest on more basic philosophical foundations. If this is so, it would seem prudent that the general theory of law implicit in expert systems should be explicitly articulated using (where appropriate) 12 Chapter 2: Legal analysis systems the relevant works of seasoned theoreticians of law. Perhaps the reason that there is, as yet, no overwhelmingly successful system is that the vast corpus of apposite jurisprudential material has not yet been tapped in the construction process.28 Although it is true that all legal expert systems necessarily make assump­ tions about the nature of law and legal reasoning, it does not follow that they must conform to some jurisprudential theory. A lawyer must have a model of the law (maybe unarticulated) which includes assumptions about the nature of law and legal reasoning, but that model need not rest on basic philosophical found­ ations. It may be a pragmatic model, developed through experience within the legal system. Many lawyers perform their work with little or no jurisprudential knowledge,29 and there is no evidence to suggest that they are worse, or better, at their jobs than lawyers well-versed in jurisprudence. Susskind concedes that it is possible to build a legal expert system “without jurisprudential insight”, but suggests that such a system would be of very poor quality: Because successful legal knowledge engineering presupposes so profound a familiarity with the nature of law and legal reasoning, it is scarcely ima­ ginable that such a mastery could be gained other than through immersion in jurisprudence.30 Harris, himself a jurisprudent, takes a different view: People acquire those technical skills of legal reasoning and legal argument­ ation which make up the concept of ‘good lawyer’ by immersing themselves in substantive legal subjects. Jurisprudence has to do, not with the law­ yer’s role as a technician, but with any need he may feel to give a good account of his life’s work—either to fellow citizens, or to himself, or to any gods there be.31 If Harris is right—and the existence of good (though jurisprudentially illiterate) lawyers suggests that he is—then the importance of jurisprudence to legal expert system design is questionable. A legal expert system need only operate at the same level of abstraction as does a lawyer, rather than at the philosophical level of a jurisprudent. The fact that many lawyers have mastered the process of legal reasoning, without having been immersed in jurisprudence, suggests that it may indeed be possible to develop legal expert systems of good quality without jurisprudential insight. This does not mean that legal expert systems designers should completely ignore jurisprudential literature. However, as a legal expert system need not conform to any jurisprudential theory, a pragmatic approach to expert system design may be preferable to a jurisprudentially pure one. Susskind, as Niblett complains, “gives the role of jurisprudence in the design of legal expert systems a greater significance than it deserves.”32 13 � 2.2 Jurisprudence It is well beyond the scope of this thesis to cover “the vast corpus of apposite jurisprudential material”,33 however several areas of jurisprudence of relevance to expert system design are discussed below. 2.2.2 Scientific and mechanical jurisprudence The dictionary defines jurisprudence as: The science which treats of human laws (written and unwritten) in general; the philosophy of law.34 Pound refers to the “scientific character” of the law: Sir Frederick Pollock gives us the clew when he defines the reasons that compel law to take on this scientific character as three: the demand for full justice, that is for solutions that go to the root of controversies; the demand for equal justice, that is a like adjustment of like relations under like conditions; and the demand for exact justice, that is for a justice whose operations, within reasonable limits, may be predicted in advance of action. In other words, the marks of a scientific law are, conformity to reason, uniformity, and certainty.35 This approach to the law has been much criticized. Frank contends that uncer­ tainty is inherent in the legal process, and that seeking certainty in general legal principles is simply “an expression of infantile emotional attitudes which have persisted into adulthood.”36 In his 1908 paper, Mechanical Jurisprudence, Pound distinguishes scientific jurisprudence from mechanical jurisprudence: Roman law in its decadence furnishes a striking example [of mechanical jurisprudence]. The Valentinian “law of citations” made a selection of jur­ isconsults of the past and allowed their writings only to be cited. It declared them, with the exception of Papinian, equal in authority. It confined the judge, when questions of law were in issue, to the purely mechanical task of counting and of determining the numerical preponderance of authority. Principles were no longer resorted to in order to make rules fit cases. The rules were at hand in a fixed and final form, and cases were to be fitted to the rules.37 By contrast, Pound says, scientific law is “a reasoned body of principles for the administration of justice”. 38 Loevinger also argues a role for science in law, but sees little science in juris­ prudence. In 1949 he proposed an alternative to jurisprudence which he termed jurimetrics: The next step forward in the long path of man’s progress must be from jurisprudence (which is mere speculation about law) to jurimetrics—which 14 Chapter 2: Legal analysis systems is the scientific investigation of legal problems . . . The inescapable fact is that jurisprudence bears the same relation to a modern science of jurimet­ rics as astrology does to astronomy, alchemy to chemistry, or phrenology to psychology. It is based upon speculation, supposition and superstition; it is concerned with meaningless questions; and, after more than two thousand years, jurisprudence has not yet offered a useful answer to any question or a workable technique for attacking any problem.39 This proposal spawned a new field of study, some work from which is discussed in §2.3 below. Loevinger calls those who fear the dangers of mechanized jurisprudence “quix­ otic and uncomprehending”40—yet he was not a mechanical jurisprudent himself (as explained in §2.3). Others, however, saw the development of computers as an opportunity to develop a judgment machine: a machine that could replace a judge. 2.2.3 Judgment machines In 1955, Lasswell predicted that: When machines are more perfect [sic] a bench of judicial robots . . . can be constructed. The machine would apply a system of “weights” to allegations of fact made by parties to a controversy, and also to the justifications advanced in support of the claims put forward by participants. Litigation can proceed by counsel for the plaintiff and the defendant pressing buttons that translate their cases into the physical signs built into the machine. Many results would be “no decision.” However, the machine could be designed to settle a controversy of this kind with a “random” operation (by lot).41 He was not altogether serious: It is a challenging task for legal historians to assist in constructing a robot whose weights would give substantially the same result as those produced by the [US Supreme] Court at various periods. The task would not be too difficult for some justices on some issues. But a robot facsimile of the less repetitive members of the Court would provide a genuine challenge to the engineers.42 However, some have taken the idea of a judgment machine very seriously indeed. In 1949, Frank wrote that a “logic machine” might “disclose all possible available alternative legal rules,” although judges would still have to exercise the “the sovereign prerogative of choice” between the rules on the basis of the judges’ “conscious or unconscious notions of policy.”43 Similarly Mehl wrote, in 1959, that a machine could perform some of the functions of a judge, but that a role for humans would remain because “the solution to a legal problem may depend upon extra-rational factors, involving the whole of human experience”. 44 15 � 2.2 Jurisprudence As recently as 1977, it has been seriously suggested that a judgment machine could replace a human judge. D’Amato proposed that a machine could take the relevant facts of a case as input and produce a number in the range −1 to 1 (where a positive number indicates a victory for the plaintiff). Given the multiplicity of factors, he claims, a result of zero would be extremely unlikely,45 although it is not clear why this would be so. These facts would be determined by a jury, but the law would be decided by the machine. Somewhat grudgingly, he allows for some vestige of human control: an appeal court could review all of the machine’s determinations in a certain numerical range (e.g. −0.05 to 0.05) within which the cases would be so close that a re-examination might be required. The review court’s subsequent decision would then be incorporated into the system.46 The idea of human judges being replaced by machines has been trenchantly criticized. According to Weizenbaum: The very asking of the question, “What does a judge . . . know that we cannot tell a computer?” is a monstrous obscenity. That it has to be put into print at all, even for the purpose of exposing its morbidity, is a sign of the madness of our times. Computers can make judicial decisions . . . They can flip coins in much more sophisticated ways than can the most patient human being. The point is that they ought not be given such tasks. They may even be able to arrive at “correct” decisions in some cases—but always and necessarily on bases no human being should be willing to accept.47 But D’Amato sees advantages in replacing human judges by machines: Would we lose a judge’s “judgment,” and how important would such a loss be to our legal system? Surely computers do not make “judgments” the way humans do, and so we would lose the “human” aspect of legal judgments. But what specifically do we lose when we lose the humanness of judgments? Is human judgment just a euphemism for arbitrariness, discretion, or bias?48 By contrast, Stone stresses the importance of legal change springing from “devi­ ance, tentativeness, and even indecision in judgment.” It is these phenomena above all which promote a judge’s sensitivity to new ideas or to newly perceived social situations and stir Hamlet-like introspec­ tion. On these there rest some of the main foundations of the social good we call “justice.”49 Proponents of the idea of automated judges claim that such systems would reduce the cost of the legal system, find inconsistencies in the law, and provide a level of certainty in the law which does not exist at present.50 D’Amato claims that a judgment machine would allow people to live under the rule of law and not under the “rule of persons.”51 16 Chapter 2: Legal analysis systems D’Amato’s argument is based upon “two bold initial premises.” He assumes that, from a jurisprudential point of view, the law has been made completely determinable. He also assumes that, if a computer can be programmed to make judicial decisions, human “discretion” will have been completely removed.52 His views on judgment machines are extreme, and have not been widely ac­ cepted. For the most part, expert system designers have focused on building systems which provide advice to lawyers or to laypeople, rather than usurping the role of judges. However (as is discussed in §2.4 below) many expert systems designers have made D’Amato’s first assumption—that the law is completely determinable—often tacitly. 2.2.4 Petrifaction of the law The idea of a judgment machine removing uncertainty and ambiguity in the law raises the possibility of the petrifaction of the law. This possibility is also relevant to legal expert system design. D’Amato claims that, once programmed, the law would become settled. The computer would stop “progress,” although the legislature could always step in if anomalous results were being produced.53 Similarly, Kayton believes that the use of propositional logic could identify legal ambiguity and provide guidelines for its resolution. He asks: Would not each concrete demonstration of ambiguity and its resolution tend to rigidify the law and eventually destroy the very flexibility which has made the common law viable? . . . Despite the dangers of general­ ization, it is submitted that . . . [this question] should be answered in the negative. Information brought to bear in a rational pursuit is always better than ignorance or confusion. The elimination of ambiguities, rather than ossify the law, would produce an optimum condition for the purposeful, intelligent, and efficient development of the law.54 Schubert, another early researcher in this field, is not convinced of the possibility, or the importance, of certainty in law. He writes: . . . the ideal of certainty in law is tolerable only in the context of an em­ pirical world in which forces inducing change are so manifold that the attainment of the goal is never possible.55 Tyree argues that a deterministic computer judge could “overrule himself” if cases are added as they are decided.56 But as Pound pointed out more than ninety years ago, the problem is more than mere ossification of the law: The effect of all system is apt to be petrifaction of the subject systemat­ ized. Perfection of scientific system and exposition tends to cut off indi­ vidual initiative in the future, to stifle independent consideration of new problems and of new phases of old problems, and to impose the ideas of one generation upon another.57 17 � 2.2 Jurisprudence This problem is inherent in expert system design—in the law and in other domains. There are dangers, for example, in the use of a legal expert system by judges. Stone considers the use of predictive techniques in the law and warns that: . . . reliance at the judgment seat on new predictive techniques would ser­ iously threaten the judge’s central concern with justice. For if the results thus predicted for him do affirmatively guide him in present decisions, each judge will tend to vote somewhat more consistently with his past record. The margin of deviation would accordingly disappear in deference to pre­ dicted patterns, aided by the human tendency to follow the less agonizing because already trodden path . . . 58 Judges, Stone says, have a duty to strive wholeheartedly for justice “at the mo­ ment of judgment.”59 Although an expert system may be of some use to judges of lower courts, there is a critical border between a machine serving the legal order and the dangers of subversion of that order. For whenever an appellate judge faces the duty to do justice in this sense, he must address himself to it with his own present experience and insights. It would be a corruption of justice for him to shortcut this duty by resting on predictions of his future decisions based on his own past behavior—still more so if the basis is the average past behavior of a group of judges.60 So, as Stone points out, the “pioneers and experts of the new techniques” must face “the limits of the contributions they can make.”61 2.2.5 Clear rules and clear cases H. L. A. Hart was the most famous of the positivist legal theorists. His major work, The Concept of Law, was published in 1961. Hart realizes that because the law is expressed in natural language, it is subject to considerable semantic indeterminacy. This he terms the open texture of law. However, he claims that it is possible to use rules deductively to solve “clear cases”: that is “those in which there is general agreement that they fall within the scope of a rule.”62 He also contends that: . . . the result of the English system of precedent has been to produce, by its use, a body of rules of which a vast number, of both major and minor importance, are as determinate as any statutory rule. They can now only be altered by statute, as the courts themselves often declare in cases where the ‘merits’ seem to run counter to the requirements of the established precedents.63 18 Chapter 2: Legal analysis systems As Susskind writes, one ramification of Hart’s analysis is that: . . . all expert systems in law whose inference procedures are solely deductive will function exclusively in the clear case domain, and will be of no aid in the solving of “problems of the penumbra.”64 But what constitutes a “clear case” is itself far from clear. As Hart concedes: . . . it is a matter of some difficulty to give an exhaustive account of what makes a ‘clear case’ clear or makes a general rule obvious and uniquely applicable to a particular case.65 Moles disputes the very existence of clear rules and clear cases. He performs a detailed analysis of the application of an example of an ostensibly clear rule in a British statute: a provision which prescribes the circumstances in which an injunction should be issued in domestic violence cases.66 Moles summarizes the effect of three cases as follows: (1) Before B v. B 67 we are concerned with a statutory provision which on its face appears to be clear and comprehensible (at least to non-lawyers) and it would appear that an injunction should issue in cases such as those we have looked at. (2) After Cantliff, 68 the provision has been considered twice by the [English] Court of Appeal within two weeks. All of the six judges who considered the matter are in agreement that the injunction should not issue and that the rule is clear. (3) The matter is further considered by a specially constituted Court of Appeal of five,69 who decide that the rule is clear, but different from that in (2), and that the injunction should issue.70 Moles also analyses the judicial application of the rule of precedent—which, with this statutory rule, “must be regarded as amongst the clearest available to us”71— to demonstrate that “our experience of the legal system bears little relationship to Hart’s account of it”. 72 Leith developed a legal expert system which operated on “clear rules” in the law.73 He later recanted, saying that: . . . the very idea of a clear rule is inherently confusing and is not observable in the real operation of the judicial process . . . judicial creativity is not an aberation [sic] of the legal process, but (as Moles . . . suggests), the very heart of the law.74 As is explained in §2.4 below, many expert system designers have adopted a view of the law that allows for clear rules and clear cases—without considering the possibility that the law may not be like that at all. � 19 � 2.2 Jurisprudence 2.2.6 Legal realism and rule scepticism In the 1920s and 1930s, a movement called legal realism developed in America.75 Realists rejected the importance that mechanical jurisprudence placed on rules. However, the rule scepticism of the American Realists is meek by comparison with the strong indeterminacy thesis. Drahos and Parker explain that: According to Karl Llewellyn a characteristic feature of Realism was the rejection of simple (by which he meant general) rules and the substitution of more detailed classificatory schemes which better captured the specific nature of judicial rule-making . . . On Llewellyn’s account, rule scepticism emerges as a set of doubts about the veracity of legal actors’ claims to be following the legal rules they say they are. This is not rule scepticism in the strong sense of denying the existence of rules, however. Legal actors may simply be following some other rules. The rule scepticism of American Realism could perhaps be more accurately described as rule cynicism.76 Kripke, by contrast, is a true rule sceptic. His argument is an example of the Wittgensteinian Paradox.77 Consider the two functions “plus” (+) and “quus” (�). The + function is the mathematical function, addition. The � func­ tion is defined as follows: x + y, if x, y < 57; x � y = 5, otherwise. Suppose that Kripke has used + in the past, but always with values of x and y smaller than 57. He performs the computation 68 + 57 and gets a result of 125. Yet it could be said that when he thought he was using the “plus” function in the past, he was in fact using “quus.” As Kripke explains: . . . in this new instance, I should apply the very same function or rule that I applied so many times in the past. But who is to say what function this was? In the past I gave myself only a finite number of examples instantiating this function. So perhaps in the past I used ‘plus’ and ‘+’ to denote . . . ‘quus’ . . . Who is to say that [‘�’] is not the function I previously meant by ‘+’ ?78 There is no justification for Kripke answering 125 rather than 5. There is no way of determining (from his past behaviour—even his past thoughts) whether by “plus” Kripke meant + or �. Rules are derived from a finite number of examples; for any given rule, there is always an alternative rule which also explains those examples. Mathematicians could counter that the meaning of + is well defined. But Kripke argues that: . . . scepticism about arithmetic should not be taken to be in question: we may assume, if we wish, that 68 + 57 is 125 . . . I cannot doubt coherently that ‘plus’, as I now use it, denotes plus! Perhaps I cannot . . . doubt this 20 Chapter 2: Legal analysis systems about my present usage. But I can doubt that my past usage of ‘plus’ denoted plus . . . . . . There is no objective fact—that we all mean addition by ‘+’, or even that a given individual does—that explains our agreement in particular cases. Rather our license to say of each other that we mean addition by ‘+’ is part of a ‘language game’ that sustains itself only because of the brute fact that we generally agree.79 Drahos and Parker write that: The consequences of this argument, if valid, are shattering. Rules turn out to be no more than leaps in the dark and the whole notion of rule following seems illusory.80 The implications of Kripke’s argument have been realized by legal theorists.81 As Drahos and Parker point out: Saying that there are no rules to follow, only social practices, means that propositions about the law are potentially open to wild fluctuations. The argument also puts paid to any possibility of a correspondence theory of truth in law.82 They propose a solution to this problem: The Kripkean argument does not prevent people from saying that they are following rule X. Rather it stops them from being able to justify the existence of that rule by reference to some objective meaning. This still leaves the practice, as opposed to the justification, of successful rule following to be accounted for.83 Their solution is to view the law as a set of rules and conventions: “a type of rule used to fix or interpret the meaning of other rules.”84 They claim there is a distinction between rule knowledge and rule understanding, the latter being “a matter of absorbing conventions relating to rule use.”85 A lawyer requires knowledge of the rules, and understanding as to how to apply those rules prop­ erly. Drahos and Parker argue that the problems facing designers of legal expert systems “flow from the difficulties of representing rule understanding rather than rule knowledge.” Rules can be used to represent rule knowledge; the problem is “how to represent with tolerable accuracy . . . conventions which confer rule understanding.”86 But they concede that, as with other rules, conventions have to “run the gauntlet of scepticism”. 87 2.2.7 A jurisprudential consensus? The only major examination of the role of jurisprudence in the development of legal expert systems is Susskind’s 1987 book Expert Systems in Law: A Juris­ prudential Inquiry. 88 As mentioned in §2.2.1 above, he argues that all expert systems must conform to some jurisprudential theory. 21 � 2.2 Jurisprudence He also sets out to find consensus in jurisprudential theory. As there is far too much literature in the field for one person to cover it all, he limits his choice of material: “the vast majority” of sources that he uses are British analytical juris­ prudential writings since the mid-fifties and early sixties, “the impetus for which”, he concedes, “was derived very largely from the publications of H. L. A. Hart.”89 Susskind concludes that “there are no theoretical obstacles, from the point of view of jurisprudence, to the development of rule-based expert systems in law of limited scope.”90 He also claims that the divergence of views within jurisprudence has been overstated because legal theorists tend to focus on the differences. He claims that there is a jurisprudential consensus, “albeit of mundane and limited application”. 91 Susskind’s work provided the theoretical justification for not only his own development work,92 but for much of the subsequent work of other developers of legal expert systems. It could also be said to have provided, retrospectively, legal theoretical justification for most of the work on rule-based expert systems that preceded his book. However, Susskind’s approach to jurisprudence is fundamentally flawed. As Moles points out: [Susskind] said that he would carry out a survey of the jurisprudential literature. He acknowledged, of course, that it would not be possible to survey the whole of the jurisprudential literature. In fact, he determined that law was a system of rules by “surveying” only those whose avowed position was based on the fact that the law was a system of rules . . . I would venture to suggest that this is in fact a misuse of the survey technique . . . Susskind was perfectly familiar with the work of Hart and his followers, and was well able, therefore, to find any number of books and articles which supported the “law as rules” view. He then developed his position . . . on the basis of what this purported consensus within jur­ isprudence had to say. . . . Of course, Susskind was telling certain sections of the AI and Law community what they wanted to hear, and hence their enthusiasm for it.93 Moles also notes Susskind’s admission that “the most rigorous of these writings constituted the source materials with greatest potential given the overall purpose of the project.”94 Because the first objective of Susskind’s work was to design, develop and implement an expert system in law,95 Moles cites this admission as evidence that Susskind prejudged his survey. In the light of this criticism, it is ironic to note Susskind’s own caution that: . . . a little jurisprudential knowledge can be a dangerous thing! It is tempt­ ing for the jurisprudential neophyte to become an ardent devotee of a par­ ticular school of thought within legal theory and to go on from there to implement all and only the teachings of that school. This course of action 22 Chapter 2: Legal analysis systems should be avoided at all costs. Familiarity with a wide range of works should be achieved prior to commitment to any particular jurisprudential posture.96 Clark, another critic of Susskind, says that: The shortcomings of the book coincide with the limits of positivist legal theory, and even then some doubt must remain as to whether expert sys­ tems in law mark a revival of the kind of “mechanical jurisprudence” which Hart opposed so vigorously from within the confines of positivist legal thought.97 But the strongest criticism of Susskind comes from Leith: . . . [Susskind] believes that he can come to some sort of compromise with the various theoretical positions taken by the renowned thinkers of the field and produce “a general theory” (an indication, I might suggest, of [his] poor theoretical conceptions) . . . . . . Susskind sees jurisprudence as providing a variety of theoretical models which can be modelled mathematically and translated into com­ puter programs. . . . of necessity formal specification requires a formalisation of law; there can be no “informal models” which are mysteriously formalised into a computer model . . . . . . Therefore, in order to use formal specifications, Susskind must provide a formal specification of law which can then be incorporated into the rule-format of a computer program . . . And, of course, formal specific­ ations of law are renowned for their theoretical and practical inadequacy. Legal formalism can, surely, hardly be the compromise he wishes might arise from the conflicting positions of Kelsen, Hart, Dworkin et al.98 . . . if he really does believe that his informal theoretical models can be transformed into formal theoretical models without loss of their informal attributes, then I must suggest that he has really little understanding of the discipline of computer science.99 Leith’s comments are consistent with the point made in §2.2.1 above: lawyers and legal expert systems operate at a lower level of abstraction than the philosophical level of jurisprudence. Susskind’s claim to have found a consensus in jurisprudence—even one of mundane and limited application—is absurd. As shown in §2.2.5 above, the views of just two jurisprudents (Hart and Moles) are completely irreconcilable. Similarly, Hart’s views are totally at odds with those of Kripke (as explained in §2.2.6). Susskind states his theory, and develops an expert system which conforms to that theory. His mistake is to claim that his theory is definitive, in that it represents a jurisprudential consensus. � 2.3 Jurimetrics and the behaviourists 23 2.3 Jurimetrics and the behaviourists As mentioned in §2.2.2, the term “jurimetrics” was coined by Loevinger. He defines “jurimetrics” as “the study of law and legal problems by scientific methods and concepts, the employment of science in law to the extent that it is applicable or adaptable.”1 Loevinger is neither a mechanical jurisprudent nor a positivist: To begin with we must be clear that science offers us neither ultimate nor certain answers to legal problems. The dream that science might someday tell us which of several competing interests is the more important is a vain one. Science essays no such answers in any field. Science does not assign social or ethical values. Science may, indeed, provide data from which social or ethical judgments may be made; but the judgments will remain with man. . . . There is no prospect of any process that will preclude consideration of social desirability or wisdom. The opportunity will always be available to argue that precedent should not be followed, and that considerations of policy, or expediency, require a different rule or a special result . . . . . . Science does not and will not offer us any law machines that give automatic answers to specific questions put to them, whether as to partic­ ular cases or as to ultimate legal issues such as the relative importance of interests that may be in conflict. By the same token, science will provide us with no formulae or calculus that will give us certainty either of predic­ tion, analysis or answers to ultimate questions such as which interest is to be preferred or which desire has greater social value.2 Although these comments are eminently reasonable, some of Loevinger’s turns of phrase are exasperating. In 1949 he claimed that putting the law on a “rational basis” was the “indispensable condition” of the survival of the human race.3 In the light of such a ridiculous statement, and after his trenchant criticism of the value of jurisprudence (quoted in §2.2.2 above), it is not surprising that, as Gardner says: The attitudes Loevinger and his colleagues expressed were never adopted by the legal profession generally . . . As a movement within the legal pro­ fession, jurimetrics has not been much heard from since the early 1970s.4 Nevertheless, some of the work of his colleagues in predicting judicial decisions is worthy of comment here. A number of researchers in the early 1960s focused on the statistical analysis of the behaviour of judges. On the basis of this analysis, they claimed that they could predict the future behaviour of individual judges, and of courts.5 Such predictions were justified on the basis of the application of the doctrine of precedent. As Lawlor writes: Even if they are man-made, the principles of stare decisis are akin to the all embracing assumption of uniformity of natural science. Without such a principle to guide us, prediction of legal decisions is impossible.6 � 24 Chapter 2: Legal analysis systems 2.3.1 Kort Kort uses mathematical expressions to represent judicial decisions.7 He identifies attributes which are of importance and represents decided cases in the following form: n Aij wi = Vj i=1 where n is the number of attributes, Aij is the value of the ith attribute for the jth case (attribute values being 0 or 1), wi is the weight of the ith attribute, and Vj is the number of votes of judges favourable to the party seeking redress in the jth case. This approach is justified on the basis that the decisions of split courts do not necessarily constitute two opposite extremes, but represent certain degrees of support for one party.8 By solving these simultaneous equations, weights are obtained for the attrib­ utes. The sum of the weights of those attributes present in the instant case is, according to Kort, the number of likely judicial votes in favour of the party seeking redress. 2.3.2 Lawlor Lawlor uses logical expressions to represent previously decided cases.9 He ana­ lyzes right-to-counsel cases heard before the US Supreme Court over thirty years, and finds that in all of these cases each judge behaved consistently with his own “personal stare decisis.” He identifies legally significant attributes and builds a logical expression which represents the behaviour of each judge. A program uses these logical expressions to predict the likely outcome, given a composition of the Court specified by the user. Lawlor’s system successfully predicted the US Supreme Court’s overruling of Betts v. Brady 10 in Gideon v. Wainright. 11 But it predicted a 5:4 majority; the Supreme Court’s decision was unanimous. Applying “traditional” stare decisis (i.e. considering the court as a whole, rather than the decisions of individual judges) Lawlor’s system did not predict the Supreme Court’s change. Given the same cases, Kort’s approach did not predict the change either. This failure draws criticism from those who disagree with such a “logarithmic approach to justice.”12 Wiener complains that “advocates of the computer” rest their arguments on an assumption that courts will adhere to the doctrine of stare decisis, which does not always hold.13 In defence of the behaviourists, Kayton refers to their failure to predict the decision in Gideon v. Wainright and writes: Should we have expected otherwise? Of course not! A reversal is by definition a logical inconsistency. That which by stare decisis had been called black is now called white.14 25 � 2.3 Jurimetrics and the behaviourists The developers of these prediction systems accept that they rely on the doctrine of precedent. Hence, a reversal of previous authority will always be beyond the predictive capacity of these systems—as it is beyond the predictive capacity of most lawyers. 2.3.3 Nagel and Schubert Nagel and Schubert15 examine the personal attitudes of US Supreme Court judges towards various political and economic relations, and claim that these “off-the­ bench” attitudes (as Nagel calls them) affect the judges’ decision-making. Correlations were . . . made between responses to specific items and various decisional propensities. For example, there was a high and statistically significant correlation between disagreeing with [one questionnaire item] (“Our treatment of criminals is too harsh; we should try to cure, not to punish them”) and being above the average of one’s court with regard to the proportion of times one voted for the prosecution in criminal cases. Off-the-bench judicial attitudes thus do seem to correlate in a meaningful way with on-the-bench judicial decisions.16 2.3.4 Haar, Sawyer and Cummings In the mid-1970s Haar, Sawyer and Cummings made use of regression analysis to build a predictive model for zoning amendment cases in Connecticut.17 Re­ gression analysis is a statistical technique for analyzing the relationship of a set of independent variables to a dependent variable. For Haar et al. the dependent variable is the outcome of the cases, the independent variables are attributes. They identify 167 attributes which “appeared to be important” to the courts, 40 of which were deemed significant using a σ2 (“chi-square”) test for association. This is “too many to use in a regression analysis”,18 so Haar et al. employ two different methods to reduce this number further: grouping attributes on the basis of “experience, knowledge, and intuition”,19 and factor analysis. 2.3.5 Prediction The behaviourists’ predictive research has been criticized because they do not attempt to model legal reasoning.20 But, in not so doing, the behaviourists just reflect the influence upon them of the American Realists. They adopt Holmes’s analysis that for any individual “the law is simply a prediction of the way in which the public force possessed by the government will act upon him.”21 Loev­ inger claims that some method of legal prediction is “indispensable”;22 Lawlor says that the ultimate goal of all scientific methods is reliable prediction of future events and “[r]eliable prediction is also one of the ultimate goals of law.”23 Suss- kind (certainly no behaviourist) also stresses the importance of the prediction of judicial decisions.24 26 Chapter 2: Legal analysis systems Tapper writes that: Although the statistical techniques employed by some of these workers have been criticized, there can be no real doubt that this work provides a successful approach to the analysis of the decisions which have been reached in the past, and at least as satisfactory a method of predicting future decisions as can be arrived at by native wit and unaided intuition.25 However, he criticizes the conclusions that have been drawn using these tech­ niques: Either the behaviourist is to content himself with observing the objective phenomena, in which case he can conclude nothing as to motivation, or he is to ascribe motivation to the phenomena, in which case he ceases to be objective. It is precisely at the point at which decisions of the [US] Supreme Court, for example, are characterized as pro-civil liberties or anti-labour that doubts arise as to the real objectivity of the studies. The same general line of argument may be advanced against the fact- oriented approach . . . Here the point at which the study loses its objectivity is in the characterization of the facts present in the case.26 (Although this may be a valid criticism of a behaviouristic study, it is not an argument against the use of predictive techniques where the characterization of facts has been performed—subjectively, admittedly—by a legal expert, as is done for SHYSTER.) Gardner is also critical of the jurimetric programs: . . . programs, done from a political science viewpoint rather than a legal one, in which the data concern legally irrelevant matters such as the ideol­ ogy and social background of individual judges.27 By “legally irrelevant matters” Gardner means those matters to which judges do not—and/or should not—explicitly have regard when coming to their decisions. But if a system is designed to predict decisions then any attribute which assists in prediction should be included—regardless of whether regard ought to be had to that attribute. These “legally irrelevant matters” are not necessarily irrelevant to the prediction of judicial decisions. Stone dismisses concern about “jurimetric” prediction: “the behavioralists [sic] may have no conscious designs on the integrity of the decisional process. The judgment of justice may be of no concern to them.”28 He defines the “judgment of justice” by reference to a situation where a judge makes a decision which: . . . does not merely declare the existing law but decides what justice re­ quires that the law should be. (. . . [W]e include tacitly here the reversal of earlier decisions, that is, creative decisions which unsettle and resettle law: this is a fortiori creative.) It is this kind of creative judgment which we have here termed “a judgment of justice.”29 � 2.4 Rule-based systems 27 The behaviourists are, Stone says: . . . only observers looking at what has already been done in judgment. And when in the course of prediction they turn their attention to future judgments, it is to ask not what the judge should do to further justice but only what kind of decision he will give if he acts consistently with values attributed to him on the basis of his past decisions.30 In other words, their work attempted to be predictive not normative. When a lawyer gives advice, she/he is expected to make some prediction as to the likely result. This prediction need not be emphatic: it may be no more than a tentative statement about the strength of a person’s legal position. Hence, legal expert systems should have some degree of predictive capacity. Like the behaviourists’ research, a legal expert system’s predictive power is a projection based on past cases. There is—and can be—no allowance for changing social mores. This is not a serious limitation upon such a system’s utility: predicting a judge-made change in the law is beyond all but the very best lawyers. However, as discussed in §2.5 (especially §2.5.2) below, the justification that a legal expert system provides for its prediction is also important. 2.4 Rule-based systems There are many examples of rule-based legal expert systems, the most important of which are discussed in detail below: the work of McCarty (§2.4.1), Bench- Capon, Kowalski and Sergot (§2.4.2), Gardner (§2.4.3) and Susskind (§2.4.4). The first proposal for a rule-based legal expert system was made by Buchanan and Headrick in 1970. They complained that interdisciplinary work between lawyers and computer scientists had “floundered on the misconceptions that each has of the other’s discipline”,31 and suggested that the computer modelling of legal reasoning would be a fruitful area for research.32 Their proposal asserts that “[i]n the absence of any reason to speculate on how they carry on their work, [lawyers] now apply complex sets of rules without being aware of the rules themselves.”33 They make no reference to jurisprudential writ­ ing, but they do make it clear that their approach is based upon two assumptions about human problem-solving in general: “(1) problems can be broken down into a set of subproblems, and (2) the solution to any subproblem requires a series of decisions that are governed by decision rules.”34 The literature is replete with examples of projects in which the assumption that lawyers work with rules is unstated or unsupported. Maggs and deBessonet, for example, tacitly make this assumption.35 The law, they claim, can be ex­ pressed as rules using propositional calculus. A program which implements these rules could answer questions from a user, and allow the checking of statute law for redundancy and contradiction. (It is not clear what role case law plays in their system, or indeed in their model of the law.) 28 Chapter 2: Legal analysis systems Popp and Schlink’s JUDITH system36 uses rules to represent parts of the Ger­ man Civil Code. The principles behind JUDITH are “strikingly similar” to those of the MYCIN system (an expert system dealing with bacterial infections37)—so similar that, its developers claim, it would be possible to create a legal knowledge base for MYCIN and a medical knowledge base for JUDITH. 38 Michaelsen and Michie’s TAXADVISOR system39 is implemented using MYCIN. 40 TAXADVISOR was designed to assist lawyers to advise clients on taxation and estate planning.41 Meldman’s system uses two different kinds of rules: general rules which define the elements of the claim, and specific rules extracted from cases.42 Things and relations are used to represent the “everyday world of human affairs”,43 and are classified hierarchically into categories. A fact comprises two things and a relation between them; facts are assembled into situations. These situations are compared with the situation of the instant case, and the system determines the extent to which the instant case falls within or near the law of intentional torts (e.g. assault and battery). Waterman and Peterson’s LDS system44 is a rule-based system for the field of product liability. It does not determine whether liability exists, but is designed to assist legal experts in settling product liability cases. LDS is a typical example of a system which developed without any legal theoretical justification. After admitting that there is no “deep model of the legal process”,45 Waterman and his colleagues proceed on the unstated assumption that any such model must involve rules: One might expect that the large body of legal rules and regulations that have been accumulated and formalized in the legal domain would make expert system development easier. Unfortunately, this is not the case. Instead, this characteristic of the domain, having rules that already exist, has led to trouble . . . First, the formal rules that define and regulate legal activity are often ambiguous, contradictory and incomplete. And second, there exists a body of informal rules or procedures about how to access, interpret and use the ‘formal’ rules. Without these informal rules the formal rules can not be used in any efficient or cost-effective way.46 This body of rules, they write, “needs to be mapped into code”47 before a legal expert system can be built. Bing’s SARA system48 is designed to analyze discretionary decisions. It uses rules to represent legal norms: some strict and some discretionary. These discre­ tionary norms are weighted using correlation techniques. SARA allows a lawyer “to back up his qualitative legal reasoning by quantitative indications.”49 Stamper’s LEGOL language50 also uses rules to represent legal norms. Pattison and Ciesielski use a rule-based system to review contracts.51 SoftLaw’s STATUTE, a commercially successful system, uses rules to represent statutes, regulations and departmental guidelines in taxation, social security and veterans’ affairs law.52 29 � 2.4 Rule-based systems 2.4.1 McCarty McCarty has been described as “the father of AI and law”. 53 His TAXMAN project54 was concerned with the development of a computational theory of legal reasoning, using corporate tax law as an experimental problem domain. He claims that a computer-based legal consultation system must be able to represent the “facts,” at some comfortable level of abstraction, and the “law,” which would consist of a system of “concepts” and “rules.” These concepts and rules are relatively abstract (i.e. they subsume large classes of lower-level factual descriptions), and they have normative implications (i.e. they specify which actions are permitted and which are obligatory). Legal analysis, in its simplest form, would then be a process of applying the “law” to the “facts”. Put this way, the paradigm seems to be an ideal candidate for an artificial intelligence approach: the “facts” would be represented in a lower-level semantic network, perhaps; the “law” would be represented in a higher-level semantic description; and the process of legal analysis would be represented by a pattern-matching routine.55 However, McCarty concedes, the representation of facts in such a system is more difficult than in other expert systems, because “the facts of a legal case typically involve all the complexities of daily life: human actions, beliefs, intentions, mo­ tivations, etc.”56 Even if the facts can be represented, he writes, the rules will often be problematic: Some rules, usually those embodied in statutes, have a precise logical struc­ ture, and this makes them amenable to the existing artificial intelligence techniques. But it is a commonplace among lawyers that the most im­ portant legal rules do not have this form at all: instead they are said to have an “open texture”; their boundaries are not fixed, but are “construc­ ted” and “modified” as they are applied to particular factual situations. A sophisticated legal consultation system would not be able to ignore these complexities, but would have to address them directly.57 McCarty also makes the startling claim that the “simplest” problems for first- year law students are the hardest for an AI system because the student draws upon ordinary human experience: Paradoxically, the cases that are most tractable for an artificial intelligence system are those cases, usually involving commercial and corporate mat­ ters, which a lawyer finds most complex. There is a simple reason why this is so. A mature legal system in an industrialized democracy is composed of many levels of legal abstractions . . . Because of their technical complexity, the legal rules at the top levels of this conceptual hierarchy are difficult for most lawyers to comprehend, but this would be no obstacle for an artificial intelligence system.58 � � � 30 Chapter 2: Legal analysis systems McCarty chose the area tax law for his system because “commercial abstrac­ tions, in fact, are artificial and formal systems themselves, drained of much of the content of the ordinary world” and, by legal standards, well structured.59 The field of corporate tax law, he says, is “very near the apex of the hierarchy of commercial abstractions”60—“whatever that means”, comments Moles.61 In TAXMAN I, McCarty’s first prototype system, the basic “facts” of a corporate case are captured in a relatively straightforward representation (e.g. a corporation issues securities). Below this level is an expanded representation of the meaning of various entities (e.g. a security interest) in terms of their component rights and obligations. Above this level—presumably above both levels, although this is not made clear—is the “law” (statutory rules which classify transactions as taxable or non-taxable etc.).62 McCarty found that, although the rules are complex, the underlying represent­ ations are manageable. He concludes from his early work that “the construction of an expert consultation system in this area of the law is a feasible proposition.”63 McCarty sees the development of legal expert systems as an opportunity to contribute to jurisprudence. Although the jurisprudential literature includes “many illuminating examples and many valuable insights about the structure and dynamics of legal concepts”, he complains that “taken as a whole” it is “no­ toriously imprecise”. 64 The TAXMAN system adds a strong dose of precision and rigor to these discussions of linguistic and conceptual problems. Its critical task is to clarify the concepts of corporate reorganization law in such a way that they can be represented in computer programs.65 Moles is extremely critical of rule-based expert systems designers, and Mc- Carty in particular. McCarty takes a positivist approach to the law, and Moles complains that: The computer scientists have taken the de-humanizing aspects of modern positivism to their extreme. . . . The computer scientists, encouraged by the modern positivists, fail to recognize the point which Austin66 correctly emphasized throughout his work—that law, positive morality and ethics are inseparably connected parts of a vast organic whole.67 Most damning is Moles’s statement that: The sad thing is that [McCarty] has not shown the slightest awareness of the nature of the legal enterprise. Far from having emphasized any difficulties, he shows that he simply does not understand what they are.68 � � � 31 � 2.4 Rule-based systems McCarty does identify two major limitations to the approach taken in TAX­ MAN I. Firstly, he concedes that the factual descriptions, although manageable in the corporate domain, would be too complex in (for example) the average con­ tract or tort problem. Secondly, the higher-level conceptual representations are not adequate for all domains because judicially created concepts are incurably open-textured, have a dynamic structure with the capacity to evolve and adapt to new situations, and the evolution of these concepts is governed by a sense of purpose.69 Having recognized these problems, McCarty set out to solve them with TAXMAN II. TAXMAN II70 uses “prototypes and deformations” to represent a legal concept by specifying the prerequisite conditions for that concept, a set of cases (real and hypothetical) in which that concept does or does not apply, and a set of transformations for getting from one case to another. If a given case (representing a legal concept) can be transformed into the instant case, while still satisfying the prerequisite conditions for the concept, then it can be used in argument as an example of that concept. Ashley sees McCarty’s work as a significant advance, but points to several shortcomings of TAXMAN II: The work is largely an exercise in knowledge representation. McCarty does not set forth a control or process model that clarifies how a program would actually generate a legal argument . . . The reported research involves a hand simulation of the arguments in one US Supreme Court case71 and work done on hand simulations of several subsequent cases. . . . [TAXMAN II] has no mechanisms for comparing cases in terms of how on point they are, for distinguishing cases, or selecting the best pre­ cedents . . . McCarty’s model assumes a much neater domain than exists in law. He assumes that in reality, legal cases are consistently allocated as positive and negative exemplars of concepts. They are not. He assumes that there is a near match between concepts and the features of a case that are relevant to the concept. There is not.72 McCarty’s definition of legal primitives has also been criticised. Moles writes: McCarty appears not to appreciate that ‘corporations’, ‘securities’, ‘prop­ erty’, ‘dividends’ and so on are not subsumed ‘beneath the law’, but are each the products of complex legal analysis. The question of whether certain transactions are taxable or not is intimately tied into that legal analysis.73 Ashley makes a similar point: The problem with such primitives, if they are taken seriously as a means for defining concepts, is that they assume what is to be shown. Far from being a primitive, that someone has a right or a duty in a given fact situation is an arguable legal conclusion that must be justified by citing authorities.74 32 Chapter 2: Legal analysis systems 2.4.2 Bench-Capon, Kowalski and Sergot Bench-Capon, Kowalski, Sergot and their colleagues use PROLOG to model stat­ utes.75 Their approach to legislation is the most extreme of all expert system developers, due to their attitude towards knowledge acquisition. They write: The formalisation of legislation by means of rules has almost all the charac­ teristics of an expert system. It differs, however, in one important respect. In a classical expert system, before knowledge can be formalized, it has to be elicited from the subconscious of an expert. Eliciting this knowledge is generally regarded as the main bottle-neck in the construction of expert systems. It is entirely absent, however, in the case of legislation which is already formulated and written down. Thus the use of expert system tech­ niques for representing legislation has virtually all the advantages of expert systems without the attendant disadvantages of eliciting the knowledge.76 This statement is nothing less than astounding. Even if it is accepted that statute law can be represented using PROLOG clauses, it is a bold claim indeed to assert that constructing those clauses requires no expertise. As Moles says, “[c]learly these researchers do not distinguish between the writing (which is the legislation) and the meaning of that writing.”77 Bench-Capon et al. have worked with several statutes. Most famous is their work with the British Nationality Act 1981 (UK).78 Consistent with their ap­ proach to knowledge acquisition, their PROLOG representation of the British Na­ tionality Act was implemented in two months by a student, “without any expert legal assistance.”79 They also represented the Supplementary Benefits Act 1976 (UK) and its reg­ ulations using a similar method, and made a similar claim about the importance of legal expertise: For our project, the accuracy of the representation was not a critical con­ sideration at this [early] stage. Our formalisation could therefore be un­ dertaken with no expert legal assistance . . . In general, accuracy of the formalisation is, of course, critical, particularly if one were constructing a representation to be used in practice.80 Predictably—and correctly—this approach has been strongly criticized. Moles points out that: This is to assume . . . that the problem with regard to “accuracy” is merely a matter of changing the detail of content. It fails to appreciate that an expert may have a great many useful things to say about how one goes about the process of interpretation. The expert advice will therefore have implications for the method being employed and the way in which the knowledge is structured.81 33 � 2.4 Rule-based systems Kowalski and Sergot make this assumption explicit: Access to an expert adviser might well have changed the exact form of the rules in our program, but it would not have changed the method we used to formulate and compute with the rules.82 But how can they possibly be sure? On the need for legal expertise in the development of legal expert systems, Susskind quotes Hayes-Roth, Waterman and Lenat: It is very easy to be deluded into thinking one knows a great deal about the domain. Remember: the expert became one only after years of training and experience.83 The work of Bench-Capon et al. has been most vehemently criticised by Leith. He quotes them: The formalisation of the British Nationality Act is an axiomatic theory similar, for example, to an axiomatisation of Euclidean Geometry. In prin­ ciple, any logical consequence of the axiomatisation can be generated and tested mechanically . . . 84 then comments: Does this sound like a wild claim? I suspect so. No lawyer, with whom I have so far discussed Kowalski’s “legal” work, has been impressed in the least: the common reply is, “Surely he doesn’t believe that?” And this is the rub. In AI research, funders are prepared to provide research funding for the most inane of ideas without appealing to those with some expertise in the area of “intelligence” being researched.85 2.4.3 Gardner The aim of Gardner’s research86 is not to develop a program that “solves” legal problems: Instead, the objective is to enable the program to recognize the issues a problem raises and to distinguish between those it has enough information to resolve and those on which competent human judgments might differ. Toward this end a heuristic distinction between hard and easy questions is proposed.87 Her chosen domain is an aspect of the law of contract that is dominated by case law: offer and acceptance. She claims that, although an area based on statute might seem easier for an AI program to handle, the reverse is true. Case law must be taken into account in statutory areas too, and statutory interpretation raises its own problems. Hence, she concludes, “[b]eginning from statutes therefore seems likelier to add a layer of complication than to remove one.”88 Susskind 34 Chapter 2: Legal analysis systems agrees: “Gardner . . . is right in recognising that it is unrealistic to focus on statute at the expense of precedent. There is a lesson here for all workers in the field.”89 Gardner’s system has four different levels, the last of which is not imple­ mented. The network level is an augmented transition network which represents legal states and events. The rule level is a set of rules which operate on objects at the first level. These rules are “definitions of the major concepts”90 in the law of offer and acceptance. The third level is a set of examples which explain those predicates in the rule level which are undefined, but whose resolution is “clear”; this includes (non-legal) common sense knowledge. Gardner, citing the legal positivists, assumes that open texture does not render all cases hard: The first three levels of the program are devoted to identifying the hard questions. From an opposite viewpoint, the first three levels try to identify the easy questions and resolve them.91 The fourth (unimplemented) level would deal with the hard questions. Susskind comments that: Ultimately, the human expertise that Gardner tries to encapsulate in heur­ istic form is the ability to assess, in advance (in a sense), whether a case does indeed raise easy or hard questions. This will indicate if a conclu­ sion may be inferred without further ado or whether human judgement is required. She makes a brave attempt, but . . . I was left wondering about the generalisability of her ideas and their applicability in other, far more extensive, branches of law.92 Ashley, too, is critical of Gardner’s approach: . . . an attorney distinguishes hard from easy questions in terms of com­ paring the strengths of the best argument he or she can make with the best arguments an opponent can make. Gardner’s program provides no measure for evaluating the strengths of competing arguments . . . Even if Gardner’s program evaluated case-based arguments, it could not do so very realistically given the way cases are represented. Distinguishing, for example, is not possible because there is nothing to distinguish.93 Unlike McCarty and Bench-Capon et al., Gardner does confront the argu­ ments of the (American) Realists: If legal realism is right, it appears to make the AI paradigm of rule-based expert systems inappropriate, at least with any simple mapping from legal rules to knowledge-base rules. There are several directions one might go instead. One direction would emphasize the idea that it is individual de­ cisions, not general rules, that have authoritative status as law. With this emphasis, one might look for a method of reasoning from the decisions in � 2.4 Rule-based systems 35 past cases to a conclusion in a present case . . . 94 But this cannot be the whole story. If it were, why would present-day law professors, thoroughly aware of the insights of realism, continue to expect their students to know rules?95 This is a strange argument; it says more about the way in which the law is taught than it does about whether a rule-based approach to the law is appropriate. (And it must be remembered that the American Realists were not true rule sceptics, as discussed in §2.2.6 above.) Gardner identifies two other directions to take “if legal realism is right”: em­ phasizing the behaviouristic side of legal realism (e.g. the jurimetrics research discussed in §2.3), and a third direction which Gardner herself adopts—“to re­ tain an important place for legal rules but to reinterpret their significance.”96 She claims that: Rulelike sentences can be understood as useful cognitive constructs, needed to find order in (or impose order upon) an unwieldy mass of individual decisions. Once articulated, they can provide guidance as to how future decisions can be kept in some rough conformance with this order; or, if the articulated rule seems to be a bad rule, it can suggest a way of saying how the course of decisions ought to be changed.97 This does not really represent a reinterpretation of the significance of rules. Only the most ardent rule-follower would suggest that rules were anything else. Gardner develops a rule-based model for offer and acceptance cases, but makes the significant concession that: . . . a rule-based model of case law must be understood, like any academic legal writing, as a secondary source. The official sources are the decisions. There has never been agreement on what it would mean for rulelike gener­ alizations from decisions to be both accurate and appropriate. Thus, the basis for even uncontroversial rules remains undefined in legal theory.98 This is an interesting, and quite pragmatic concession, coming as it does from someone whose work has been characterized as “fairly clearly purist in nature”. 99 2.4.4 Susskind Susskind’s work on the importance of jurisprudence to the development of legal expert systems is discussed in detail in §2.2.1 and §2.2.7 above. He chose the Scottish law of divorce as his first experimental legal domain.1 He has also been involved in the development of the Latent Damage System: a legal expert system concerned with the time periods within which claimants may start negligence proceedings where they have suffered latent damage or loss.2 Capper and Suss- kind claimed in 1988 that this was the first legal expert system built in the UK by lawyers for lawyers.3 36 Chapter 2: Legal analysis systems Susskind believes that statutes and some cases—the clear cases—can and should be represented using rules. With clear cases, he claims, it is possible to “draw legal conclusions on the basis of literal interpretations of the formal legal sources.”4 However: . . . it should not be taken for granted that the entire common law system can be reduced to a collection of rules . . . Simpson has forcefully conten­ ded that such a reductionist model misrepresents the common law and is inconsistent with its development, content, and scope . . . 5 Other methods, he says, would be required in order to represent cases which are not clear: methods which reason with uncertainty and draw “probabilistically phrased conclusions.”6 Yet, he contends: . . . although the common law may not be sufficiently represented in terms of rules, it cannot be doubted that it is invariably possible, desirable, and necessary to interpret individual cases in the form of individuated rules.7 The author doubts that it is possible, disputes that it is desirable, and demon­ strates—with SHYSTER—that it is not necessary to represent cases using rules. 2.4.5 Knowledge acquisition and representation Feigenbaum wrote in 1981 that “[t]here are many important problems of know­ ledge representation, utilisation, and acquisition that must be solved, but the acquisition problem is the most critical ‘bottleneck’ problem.”8 As discussed in §2.4.2 above, Bench-Capon et al. claim that there is no such bottleneck for knowledge acquisition from statute law because legislation is “already formulated and written down.”9 However, this demonstrates a serious misunderstanding of the law. Knowledge acquisition is as much a problem in the legal domain (stat­ ute and case law) as it is in any other. The author proposes a solution to this problem, at least in regard to case law, in chapter 3. In order to avoid the knowledge acquisition bottleneck, some researchers have examined the possibility of the machine-processing of statutes. If the only input is the words of the statute itself then this problem is one of understanding natural language—a research field in which results have not fulfilled original hopes. A more feasible approach is to convert the words of the statute into a machine- readable form. Allen has developed one such form and claims that if statutes were drafted using this form, not only would the automatic logical analysis of their contents be possible, but humans would be able to read and work with them more easily.10 Unless and until legislation is expressed in machine-readable form—an extremely unlikely event, the desirability of which is far from clear—builders of legal expert systems which deal with statute law must use human expertise to extract knowledge from statutes. 37 � 2.4 Rule-based systems The process of acquiring knowledge from statutes is usually seen as a process of writing rules. However, Shannon and Golshani warn that doing this in an ad hoc fashion is unsatisfactory because the rules cannot be checked for correctness, and such an approach may lead to dissimilar rule formulations which do not work well together.11 Instead they recommend following precise and consistent methods (like those of Allen) to formulate rules. This, they say, would reduce the scope for dissimilar rule formulations.12 Shannon and Golshani’s belief that there is a “correct” interpretation of a statute is shared by many researchers in this field. In believing this, they go even further than the legal positivists on whose theories much of their research is based.13 2.4.6 Fact representation The representation of legal facts in a rule-based system—indeed any expert system—is difficult. As with the representation of statutes, researchers have developed normalized forms for fact representation. An example is deBessonet and Cross’s atomically normalized form (ANF).14 Shannon and Golshani use a modified example, taken from some of deBessonet and Cross’s work modelling the Louisiana law of causality, to demonstrate how a statement of fact can be represented using ANF. The statement: A lessor believes that the lessee caused a defect in the leased premises which requires that the lessee fix the defect is represented in ANF as: (Lessor Believes ((Lessee Caused (Property Has Defect)) Causes (Lessee Must-Fix Defect))). This decomposes into clauses as follows: “Lessor Believes A,” where A is: “B1 Causes B2 ,” and B1 is: “Lessee Caused C,” and B2 is: “Lessee Must-Fix Defect,” and C is: “Property Has Defect.”15 In this ANF representation, and the decomposed clauses, the lessor believes that the lessee caused a defect in the leased premises and that (as a result) the lessee is required to fix the defect. 38 Chapter 2: Legal analysis systems There is at least one other interpretation of the original statement: “Lessor Believes B1 ” and “A” where A and B1 have the same meanings as above. This interpretation may not be “better” than Shannon and Golshani’s, but it is certainly plausible. In attempting to demonstrate the uses of one method of fact representation they (unwittingly) provide an excellent example of the enormous difficulties that rep­ resenting facts can present.16 It makes no more sense to say that any given representation of the facts is “correct” than it does to make the same claim of a representation of the law. 2.4.7 The inadequacy of rules for case law Although several developers have used rule-based systems to model statute law and case law, rules are fundamentally inadequate for representing cases. Tyree, Greenleaf and Mowbray claim that it is inappropriate to model case law using a rule-based system: not because it is theoretically impossible, but that “it is not the natural way in which lawyers reason with cases.”17 As dis­ cussed in §2.2.6 above, it may be theoretically impossible to write such rules: the sceptical view of rules applies to representing statutes and cases. But there is a fundamental difference between these two sources of law which means that, even if rules are appropriate for representing statutes, they are not appropriate for representing cases. Lawyers apply statutes in a rule-like fashion. This is understandable given the rule-like form in which they are written. However, lawyers reason with cases by arguing about their similarities and differences. As Tyree et al. explain, it used to be thought that each decided case stood for a rule of law. “It is now clear that any interpretation of the legal significance of a case must be in the larger context of the legal material in which it is embedded”. 18 In support of this assertion, they cite Stone who argues that: . . . however much we try to conceal the truth by using singular terms like “case”, “precedent”, “decision” or “holding”, the truth is that the ratio decidendi of a case has always to be sought in a body of judicial discourse, that is, of communications by judges which enter the legal materials as a more or less complex collocation of words in a written report.19 Perhaps, Tyree et al. suggest, this is why reasoning with case law using the usual production rule formulation has had little success.20 (The way in which lawyers argue with statutes and cases is examined further in §3.3.) The fact that case law is embodied in cases makes knowledge acquisition for a rule-based system extremely problematic. Extracting rules from case law is difficult for a legal expert, because that is not the way in which legal experts � � � � 2.5 Case-based systems 39 view case law. Kowalski identifies this “profiling” of cases as the bottleneck in the construction of case-based reasoning legal expert systems.21 It is not possible to automate this knowledge acquisition process using in­ ductive methods because the number of decided cases in any given area of law is usually so small that inductive inference algorithms cannot be used.22 Even if an area of law has a sufficiently large number of cases, those cases are, in a sense, unrepresentative of that area: so-called “problems of the penumbra”23 are more likely to require determination by the courts than are straightforward cases. Once such a problem has been resolved, other cases in which the same problem arises are less likely to be taken to court—if they are taken to court they are less likely to be reported. (Reasons for the paucity of reported cases are discussed in detail in §3.13.3.) This further complicates the induction of sensible rules. And, of course, it is not at all clear how any such rules could be used in legal reasoning with cases even if they could be induced. Developers of case-based systems like FINDER (§2.5.2) and HYPO (§2.5.3), and, to some extent, the hybrid CABARET system (§2.6) have recognized the in­ adequacy of rules for representing case law. However this view has not been uni­ versally accepted. Notable amongst its critics is Berman. His argument against case-based reasoning is examined, and refuted, in §3.14.2—after the pragmatic approach to case law adopted in chapter 3 has been fully explained. 2.5 Case-based systems Bench-Capon and his colleagues chose the British Nationality Act as an object of research because, as a fairly recently enacted statute, they claim it was “free of the complicating influence of case law.”24 In Australian courts, even new Acts may be interpreted in the light of pre­ viously decided cases. This applies to decisions interpreting an expression in a similar statute, in the same or in a different jurisdiction, and (to a lesser extent) to the re-enactment of a statutory provision after a judicial decision as to its meaning.25 If Bench-Capon et al. had sought expert advice they would have learnt that, similarly, in the United Kingdom the prior legal history of the language and concepts used in a statute are relevant to its interpretation.26 A statute—even a newly enacted statute—must always be interpreted in the light of case law. In 1988, according to Pearce and Geddes, approximately 50% of recent re­ ported Australian cases required the court to rule upon the meaning of some legislative instrument. In a further 25% of cases courts were required to apply legislation, “its meaning this time not being in dispute.”27 Clearly, any useful legal expert system must be able to take account of the legal effect of previously decided cases. � � � 40 Chapter 2: Legal analysis systems Ashley claims that the following operations are general to all case-based reason­ ing: . . . (1) ordering relevant cases and potentially relevant cases in terms of how analogous they are to the problem situation, (2) selecting the most analogous cases, (3) identifying configurations of counterexamples, (4) hy­ pothetically modifying the problem situation to explore contingencies, and (5) comparing case-based analyses of different problem situations to ex­ plain differences.28 In order to have some basis upon which to determine whether two cases are similar, information about certain attributes of those cases must be gathered. As Lambert and Grunewald explain: The first task of [case-based] reasoning is to pick, from the infinite number of respects in which cases can be similar and dissimilar, a manageable set of [attributes] that could support a conclusion that one case is so similar to another that it will likely have the same outcome. Without such con­ straint, one would be faced with the commonsense impracticability, if not the jurisprudential impossibility, of defining the entire set of [attributes] that any case in the domain can have, together with the full ranges of possible values that those [attributes] could take. Taking this arbitrarily restricted, but practically necessary, case struc­ ture, one could in principle generate the complete set of cases belonging to the domain and produce thereby a case base containing one case that would be exactly the same as any possible test case in the domain. But the size of such casebases would be intractable for case structures possess­ ing more than an extremely small number of [attributes]. Therefore the next task of the reasoner is to install in the case base a set of cases con­ sidered typical of those one is likely to encounter in the domain. These can be either real or hypothetical cases that an expert concludes collectively capture the essence of the domain.29 As with statute law, there is no “correct” answer to a question of case law. So, as Ashley writes, a case-based legal expert system: . . . does not “decide” a case; it makes arguments on behalf of the respective parties but does not necessarily determine a winner. The program would be useful as an attorney’s assistant, spotting issues, strengths, weaknesses, and precedents that an attorney representing a client in the [instant case] would want to take into account, or as part of a legal tutoring system.30 Nevertheless, as discussed in §2.3.5 above, an ability to predict the likely outcome of a case is a component of legal advice, both human and computerized. Three examples of case-based systems are discussed below: the nearest neigh­ bour analysis of Mackaay and Robillard (§2.5.1), and the FINDER (§2.5.2) and HYPO (§2.5.3) systems. These examples are compared with SHYSTER in §3.14.1. 41 � 2.5 Case-based systems 2.5.1 Nearest neighbour analysis Mackaay and Robillard were the first to examine the use of nearest neighbour analysis in predicting judicial decisions.31 Nearest neighbour analysis is a stat­ istical technique first developed in the early 1950s.32 As Cover and Hart describe it: The nearest neighbor decision rule assigns to an unclassified sample point the classification of the nearest of a set of previously classified points . . . If it is assumed that the classified samples . . . are independently identic­ ally distributed . . . it is reasonable to assume that observations which are close together (in some appropriate metric) will have the same classifica­ tion, or at least will have almost the same posterior probability distribu­ tions on their respective classifications.33 Mackaay and Robillard chose as their domain Canadian capital gain cases decided in the ten years before 1968. These same cases were the subject of earlier research by Lawlor.34 They screened the set of “initial or standard” cases and removed those in which the decision did not coincide with the majority decision amongst its nearest neighbours. This, they claim, “has the advantage of purifying the standard or reference cases by eliminating those that appear to be erroneous in relation to the other ones.”35 Mackaay and Robillard used, as their similarity metric, the number of different attributes. They weighted each attribute equally, on the basis that such a method is more reliable than differential weighting via multiple regression.36 Their results compare favourably with those of Lawlor, using the same cases.37 2.5.2 FINDER Tyree, Greenleaf and Mowbray use nearest neighbour analysis in their FINDER system.38 FINDER is a case-based system which gives advice in the law of trover— the law concerning the rights of the finders of lost chattels. This area of law is unusual in that it is based entirely on cases. FINDER has a database of leading trover cases, and a set of attributes which were of legal significance in those cases: e.g. “Was the chattel attached to the land or premises where it was found?” For each of the leading cases, FINDER has a vector of attribute values; each attribute value (yes or no) answers the corresponding attribute’s question for that case. Hence, each vector of attribute values represents the facts of that case. The user provides FINDER with the facts of the instant case by giving a yes or no answer to each of the attribute questions. FINDER assigns a weight to each attribute, equal to the inverse of the variance of the values of that attribute across all the cases.39 FINDER uses these weights to find the weighted Euclidean distance between the instant case and each of the leading cases. It uses the nearest case, and the nearest case with the opposite res­ ult, to build an argument about the likely result in the instant case: i.e. whether 42 Chapter 2: Legal analysis systems or not the finder should be allowed to keep the found chattel. Several statistical techniques are employed to reduce the possibility of giving bad advice. FINDER was designed to provide a comprehensive report for the user as to its opinion. As Tyree explains: Usually explanation and justification are provided in the [non-legal expert] system as a means of establishing the user’s confidence in the advice which is given by the system . . . By contrast, the legal expert system provides the justification, that is, the reasoning process, as one of its major products. The user, particularly if a professional, may care little for the prediction of the system, but the reasons provided for the prediction could be useful even if the predictions were always wrong. If the expert system can provide good arguments, then these are useful as a product in themselves.40 Tyree understates the role of prediction in a legal expert system. Certainly, a legal expert system’s argument is important; so, too, is its predictive capacity (as discussed in §2.3.5 above). SHYSTER adopts and expands upon FINDER’s approach to case law. SHY­ STER’s approach is detailed in chapter 3. Further details of FINDER are explained in §3.14.1 by comparison with SHYSTER. FINDER has been simulated using SHYSTER as described in §5.2.41 2.5.3 HYPO Ashley and Rissland’s HYPO system is a case-based legal expert system which makes use of hypotheticals in building its arguments.42 Their aim was to build a working model of “making reasonable arguments in law,” a “messy domain . . . that lacks a strong theoretical model that would support deductive reasoning techniques.”43 HYPO has four knowledge sources. Its case representation language—the first knowledge source—has two tiers: “legal case frames” are used to store basic in­ formation about cases (including the instant case) and hypotheticals, and the factual objects and relations that are important in those cases; “factual predic­ ates” are used to summarize the facts of a case represented by the legal case frames. “They are generalized factual statements that confirm whether certain legally significant relationships are true in the case”. 44 HYPO’s case base—its second knowledge source—contains thirty legal cases (including hypotheticals) concerning trade secrets law. Its third knowledge source are its dimensions. These are constructs for rep­ resenting factors: “stereotypical facts of legal cases important for the strength of a plaintiff’s position on a particular kind of claim.”45 A claim can be thought of as a result that one of the parties desires.46 � � � 43 � 2.5 Case-based systems Each dimension has a list of prerequisite factual predicates: information about those facts which make a case more or less extreme along the dimension, and how a change in those facts makes the case better or worse for the plaintiff along the dimension. A dimension is applicable if all of its prerequisites are satisfied; a dimension is a near miss if all of its prerequisites are satisfied except those associated with facts which locate the instant case somewhere along the dimension. These dimensions: . . . are not definitional elements of a claim; they do not purport to spe­ cify necessary and sufficient conditions for determining the existence of a claim. Instead, they represent collections of facts that tend to strengthen or weaken an assertion that the claim applies to a fact situation.47 HYPO’s fourth knowledge source is a set of criteria for evaluating the strength of an argument: e.g. if one case is more “on point” than another case, the former case is better.48 Given the facts of an instant case, HYPO selects the relevant cases (the cases which share at least one dimension with the instant case) and generates an argu­ ment based on those cases. HYPO’s arguments are “three-ply”; it makes a point for one side (“drawing the analogy between the problem and the precedent”49), responds with a point for the other side (attempting to distinguish the cited case, and citing other cases as counterexamples), then makes a final rebuttal (attempting to distinguish the counterexamples). The best cases to cite in the first instance are those in which the result favours the chosen side, and which share (with the instant case) at least one applicable dimension favouring that side. HYPO also uses “most-on-point near-miss cases” in its arguments. These are cases that would be analogous to the fact situation if all of the dimensions that were near misses applied. Of course, as Ashley points out, there may be more than one analogous case. Because the evaluation criteria are not well defined, the choice of most analogous case may depend on which evaluation technique is used. “Thus it may be useful for decision-making or explanation to define ‘most analogous’ less restrictively to yield a larger set of alternatives.”50 The most analogous cases may lead to conflicting results. However: . . . comparing and contrasting the conflicting most analogous cases in a symbolic way can help educate the decision maker. She or he sees the alternative ways of answering a question and is better prepared to make a wise decision. She or he also sees how small changes in the problem could lead to different results. The law’s adversarial system institutionalized this approach to decision making.51 HYPO also generates hypothetical variants of the instant case; variants that would strengthen/weaken the case for each side.52 44 Chapter 2: Legal analysis systems HYPO is a sophisticated case-based system. But, its complicated structure for knowledge representation means that knowledge acquisition is difficult. Although HYPO is based on a general theory of law, it was developed and tested using only one domain: American trade secrets law. The applicability of HYPO’s approach in other domains has not been demonstrated. 2.5.4 The inadequacy of semantic networks Semantic networks represent another potential approach to representing case law. Hafner’s LRS system, for example, uses a semantic network to represent statutes and cases of relevance to negotiable instruments law.53 Branting cites two reasons for using a semantic network, rather than a “feature-vector” representation of cases, in his GREBE system:54 First, any particular set of case features can represent only a small portion of the nearly limitless variety of event sequences that can give rise to legal claims. A second and more fundamental reason is that determining the legally relevant aspects of a new case is frequently the most difficult step in legal reasoning. Systems limited to case descriptions consisting of sets of legally relevant features ignore this step or force it onto the user. The capacity to represent and create multiple, competing arguments about the legal consequences of a set of facts depends on a representation that is free of bias towards any particular analysis. Any such unbiased representation must be of a finer granularity than legally relevant features.55 The author contends that no representation can be unbiased. Furthermore, rep­ resentations with exceptionally fine granularity face the same criticism as do deep conceptual models of legal reasoning (discussed in §2.7 below): viz. they operate at too high a level of abstraction for use in an expert system. Using semantic networks raises other problems. As Branting concedes: Unfortunately, representing complex cases in a semantic network formal­ ism is extremely laborious and difficult. Knowledge representation is a sufficiently immature field that each new case may raise representational issues that are more difficult than the legal issues posed by the case.56 Semantic networks are of dubious use for representing case law. At the very least, their use is not consistent with a pragmatic approach to case law which, it is argued, is appropriate for expert system design. 2.6 Hybrid systems The law in Australia and other common-law countries is based on both statutes and cases. For a legal expert system to be of use in most legal domains, it must be able to take account of statute law and case law. � � � � � � � � � 45 � 2.6 Hybrid systems The first legal expert systems used rule-based methods to represent both statutes and cases. It was not until the late 1980s that a few researchers examined the possibility of combining rule-based and case-based methods to produce a hybrid legal expert system. Rissland and Skalak wrote in 1989 that “[t]his sort of hybrid architecture . . . has not been much researched to date”,57 and only a few projects have arisen since. The most important of these hybrid systems is CABARET. 58 CABARET deals with a small area of US taxation law: home office deductions. It treats its rule-based and case-based systems as co-reasoners, each capable of operating on its own. Some thirty heuristics control how the two systems work together. For example, some of its heuristics concern how to “broaden” a near miss rule (i.e. one in which all but one conjunct can be established): • Use CBR [case-based reasoning] to find cases where the rule did not fire, but the consequent of the rule still held. (That is, show that the missing conjunct is not necessary to fire the rule.) • Use CBR to find cases where the rule did fire, and point out the similarities between those cases and the present case. (Show that effectively you have the missing conjunct.) Use CBR to find similar cases where the rule did not fire, but the ultimate • disposition of the case was consistent with the user’s point of view. (Show that the rule firing is not necessary for the ultimate result the user wants.)59 As can be seen from these heuristics, CABARET treats cases as examples of rules firing or not firing. CABARET’s hybrid structure is a mixed one: it mixes rules and cases. Branting’s GREBE system60 deals with Texas worker’s compensation law. GREBE is a hybrid system. Its rule base contains statutory rules and common-law rules. It uses a semantic network to represent case facts,61 and utilizes precedent con­ stituents: “Each precedent constituent acts as a warrant connecting some subset of the facts of a precedent to one of eight distinct legal predicates.”62 These precedent constituents are (effectively) rules which allow GREBE to use portions of precedents. Branting claims that this improves the system’s case matching capacity; it can match portions of cases where the entire cases would not match. Of course, all case-based systems should be capable of matching cases which are not completely identical. GREBE’s precedent constituents simply constitute a different approach to the notion of similarity between cases than that adopted by (for example) FINDER and SHYSTER. � � � 46 Chapter 2: Legal analysis systems PROLEXS (in its earlier versions63) was a hybrid system. If the sum of the weights of a list of weighted factors in the instant case exceeded a certain threshold, the instant case was said to match the “case-prototype” and “other types of (mostly rule-based) reasoning took over”. 64 The weights and the threshold were determined by the knowledge engineer. However, the developers have since adopted a different approach. PROLEXS now uses a neural network to model the Dutch law of “suitable employment.”65 Its developers claim that neural networks have advantages over rule-based sys­ tems because a neural network generalizes from the examples it is provided with, though they concede that a neural network cannot explain its decision by referring to explicit rules since its knowledge is not symbolic.66 PROLEXS’s developers comment that “the generalizing capacity of the network is already evident: in a sense the network has discovered a new rule.”67 Clearly PROLEXS’s approach to the law is, now, rule-based. All of these, and other,68 hybrid systems use rules and cases to represent case law. A cleaner division between the rule base and the case base—one which does not mix rules and cases—is proposed in chapter 3 (see especially §3.2.7). 2.7 Conceptual models: deep and shallow As Susskind says, “purely rule-based systems can cope only with problems for which they have explicitly represented and applicable rules”. 69 In response to this problem—“to fill the gaps in rule-based legal knowledge bases”70—several researchers have sought to develop conceptual models of the legal domain.71 Fore­ most amongst these is McCarty72 who identifies “the construction of a conceptual model of the relevant legal domain” as “the most critical task in the development of an intelligent legal information system”. 73 McCarty advocates the development of what he calls “deep conceptual mod­ els,” as opposed to “shallow” ones. As Greenleaf, Mowbray and Tyree comment, “[t]he meaning to be given to ‘deep’ and ‘shallow’ is not always clear.”74 It seems that by “deep,” McCarty means a conceptual model that is detailed enough to express the important facts about a particular legal world, yet abstract enough to suppress the irrelevant detail.75 According to Shannon and Golshani, truly deep conceptual models closely approach human reasoning, because they model the meaning behind words, not just the words themselves.76 Greenleaf et al. claim that there are at least three levels at which a legal expert system might be expected to model legal knowledge: (i) the system could include only the heuristics of legal experts as to the outcomes which are likely in particular situations, but without provision of any justification based on primary legal sources; 47 � 2.7 Conceptual models: deep and shallow (ii) the representation could include justification based on the primary legal sources, but without any explicit model of those sources; certain heuristics concerning the relationship between these sources, e.g., principles of interpretation, may be implied in the representation or the inference system; (iii) the system could include an explicit causal model which serves to define the relationships among the concepts employed in the primary sources. Justification would then, presumably, be based on the model.77 A “deep” model, then, would be one of type (iii). Susskind claims that the results obtained by researchers who are trying to develop conceptual models “are not universalizable in so far as they do not seem to be offering any coherent guidance regarding the development of conceptual models in other legal domains.”78 But the problem is more fundamental than this. Deep conceptual models of legal reasoning, like jurisprudence, operate at too high a level of abstraction to be of use in legal expert system development. Furthermore, developing a conceptual model of legal reasoning amounts to the writing of meta-rules;79 meta-rules have all the limitations inherent in rule-based systems, with the single exception that they can cope with problems for which a rule-based system does not have explicit rules. A conceptual model does not meet the arguments put by rule sceptics (see §2.2.6 above). The author agrees with Greenleaf et al.: We believe that the absence of anything even resembling a “deep” model of more than the smallest subset of the legal domain means that expert systems of type (iii) are far in the future and will require very substantial resources to build. Whether the expenditure of these resources is neces­ sary or even justified, at least for the purpose of building expert systems, will depend upon the performance of level (ii) type systems. This is an empirical problem which may only be resolved by building systems and evaluating their performance.80 SHYSTER is a system of type (ii). As to the value of developing conceptual models of the law, it is worth noting the words of Stone: When excessive pretensions are avoided, it may be a worthwhile intellectual activity to construct general concepts of law, or of particular notions found within legal orders, and to draw from these logical implications concerning the conceivably possible arrangements and contents of legal orders . . . Such jurisprudential activity . . . does not, despite many misconceptions, help to discover or create any actual law. Its raison d’être is basically to extend knowledge, and to order complex legal materials for mnemonic purposes of legal study and legal reform. Of course, like all efforts to extend knowledge, analytical jurisprudence also serves to sharpen the mind. Also, 48 Chapter 2: Legal analysis systems since it exposes premises from which existing legal rules may claim to have been inferred, logical analysis may provide a basis for substantive criticism of law which includes the rationalising and the testing of rationalisations offered for such rules. These are all legitimate outcomes of logical analysis; but they must always be carefully distinguished from erroneous uses of these outcomes. Among these erroneous uses is their tactical use in legal tasks, for instance to persuade courts to a particular view of the law.81 But it is just these legal tasks with which a legal expert system should be designed to deal. McCarty describes his “language for legal discourse” as taking “a first concrete step” towards the realization of a deep conceptual model.82 Moles is disparaging: “McCarty is only taking his first steps after some 15 years (although apparently he is going to continue to use the same tools as previously).”83 2.8 Conclusion Susskind argues that “jurisprudence can and ought to supply the models of law and legal reasoning that are required for computerized implementation in the process of building all expert systems in law.”84 In fact, jurisprudence is of limited value to developers of legal expert systems. For a legal expert system to be capable of producing advice similar to that which one might get from a lawyer, it needs to operate at the same pragmatic level of abstraction as does a lawyer— not at the philosophical level of jurisprudence. Many lawyers operate without jurisprudential insight; why not, then, develop a legal expert system based upon a similar pragmatic approach? Jurisprudence is of greater value to developers of judgment machines; ma­ chines concerned more with the nature of law and justice than with the nature of lawyers’ arguments. But the social desirability of judgment machines is question­ able, and whether such machines are possible is debatable. It is doubtful that anyone would seriously advocate their development today. The expectations of legal analysis systems have changed since the proposals of the 1940s and 1950s. Interest now focuses not upon judgment machines but upon legal expert systems: systems which are designed not to pass judgment but to provide legal advice. Developers have, as Stone urged them to in 1964, faced the limits of the contributions they can make.85 Providing legal advice, in an adversarial legal system like that in Australia, re­ quires the construction and analysis of arguments and counter-arguments. It also requires prediction of the likely outcome, or at least some comment on the relative strengths of the arguments. The jurimetrics researchers focused on prediction; case-based systems like FINDER and HYPO focus on argument construction, al­ though they can identify one side’s argument as being stronger than another’s. � � � 49 � 2.8 Conclusion Although a legal expert system should have a degree of predictive power, it should not be normative. Its predictions must be based on projections from the past. It need make no allowance for new issues of policy, changing social mores, or other factors to which a judge may, openly or otherwise, have regard when coming to a decision. Without regard to these factors, a legal expert system can still provide useful legal advice. Taking all of these factors into account requires a level of predictive skill which is beyond all expert systems—and beyond all but the most prescient of lawyers. Susskind blames the lack of successful working expert systems on the developers’ failure to use jurisprudence.86 Moles attributes it to the developers’ “uncritical acceptance of law as a system of rules.”87 Certainly, most legal expert system designers have embraced the work of Hart and the legal positivists, without con­ sidering the work of the legal realists or the rule sceptics. Although a rule-based approach may be appropriate for representing statute law, it is not appropriate for representing case law, and a case-based approach is clearly inappropriate for representing statute law. A legal expert system should account for both statute law and case law. A hybrid approach, utilizing rule-based and case-based techniques, addresses these representational problems. Despite what Bench-Capon et al. say,88 legal expertise is essential in the de­ velopment of a legal expert system. Knowledge acquisition is as much a problem for the expert system designer in the legal domain as it is in any other; legal expert systems are subject to the knowledge acquisition bottleneck. Deep conceptual models of legal reasoning are inappropriate for legal expert systems. They contribute to the difficulty of knowledge acquisition. They also operate at the same level of abstraction as jurisprudence, so they have little relevance to the pragmatic level of abstraction at which lawyers operate. A conceptual model of legal reasoning attempts to precisely model a process which is not fully understood. Hence, developers of these deep models confuse precision with accuracy. A legal expert system should operate at the same pragmatic level of abstrac­ tion as does a lawyer. The approach taken in the development of SHYSTER, detailed in the next chapter, is based upon this conclusion. 3 A pragmatic approach to case law Hector Frome: “Justice is a machine that, when someone has once given it the starting push, rolls on of itself.” John Galsworthy (1910) Justice89 Then there is the doctrine of precedent, one of my favourite doctrines. I have managed to apply it at least once a year since I’ve been on the Bench. The doctrine is that whenever you are faced with a decision, you always follow what the last person who was faced with the same decision did. It is a doctrine eminently suitable for a nation overwhelmingly populated by sheep. As the distinguished chemist, Cornford, said: “The doctrine is based on the theory that nothing should ever be done for the first time.” Lionel Murphy (1979) The Responsibility of Judges90 51 52 Chapter 3: A pragmatic approach to case law 3.1 Introduction The adoption of a pragmatic approach to legal expert system design is advocated in the previous chapter. This chapter describes a pragmatic approach to case law—an approach which was adopted for the development of SHYSTER. SHYSTER’s design criteria are set out in §3.2. SHYSTER is based upon a model of legal reasoning which is described in §3.3. The need for users of a system like SHYSTER to have some legal expertise is explained (§3.4). SHYSTER’s knowledge representation structure is described in §3.5. This structure was designed to facilitate specification of different areas of case law using a specification language which is described in §3.6. Areas of case law are specified in terms of the cases and attributes of importance in those areas. SHYSTER weights its attributes (as described in §3.7) and checks for dependence between them (§3.8). In order to choose cases upon which to construct its opinions, SHYSTER calcu­ lates distances between cases (§3.9) and uses these distances to determine which are the nearest leading cases to the instant case (§3.10). SHYSTER uses informa­ tion about these cases to construct a report (§3.11). Several safeguards are em­ ployed so as to warn the user when SHYSTER’s opinion may be suspect (§3.12). Methods of testing and evaluating SHYSTER’s performance are discussed in §3.13. Conclusions are drawn in §3.14. SHYSTER’s approach to case law is compared with those of other systems (§3.14.1), and an argument that an approach like SHYSTER’s is inappropriate for case law is examined and refuted (§3.14.2). 3.2 Design criteria The following design criteria were used in the development of SHYSTER. They are based upon the conclusions drawn in chapter 2 (see §2.8), and are presented here in the future tense. 3.2.1 Users and output SHYSTER will be designed to be used by lawyers. Hence, SHYSTER will attempt to imitate the manner in which lawyers write advice for their clients, and for each other; its output will be constructed so as to resemble legal advice produced by and for a lawyer. The user will be assumed to have legal expertise, though no specific expertise in the area of law about which SHYSTER is interrogated. This approach will not significantly restrict the utility of the system, and has been adopted by other expert system developers. Susskind, for example, suggests that: . . . the users of expert systems in law should be lawyers, or at least those with considerable familiarity with the workings of the legal and court sys­ tems . . . for a system to be used responsibly, the user must be aware of 53 � 3.2 Design criteria the possible role in legal reasoning of ‘principles’ . . . and of ‘purpose’ . . . Moreover, he must be sensitive to the drawbacks and implications of ‘com­ partmentalizing’ the law . . . and capable too of recognizing those occasions when some legal aid cannot help him with any problem at hand . . . 91 A system designed to be used exclusively by lawyers is still a legal expert system as defined in §2.1. (The role of legal expertise in the development and use of a legal expert system is further discussed in §3.4 below.) 3.2.2 A pragmatic model of legal reasoning As SHYSTER’s advice will be in a form that might be produced by a lawyer, SHYSTER will be designed to operate at the same level of abstraction as does a lawyer. Lawyers operate on a day-to-day basis at a pragmatic level of abstraction which is different to the philosophical level of a jurisprudent. A legal expert system should be built upon a model of legal reasoning, but this model need not conform to any jurisprudential theory about the nature of law. The model of legal reasoning adopted for SHYSTER will reflect the way in which lawyers reason with statutes and cases in areas of private law.92 (The model of legal reasoning adopted for SHYSTER is explained in §3.3 be­ low.) 3.2.3 Knowledge representation The representational structure used for SHYSTER will be as simple as possible while complex enough to allow SHYSTER to produce good advice. Simple knowledge representation is consistent with the choice of a pragmatic model of legal reasoning over a more complex, or deeper, model. Furthermore, complex knowledge representation requires commensurately complex knowledge acquisition. A simpler representation—one which makes no attempt to model accurately the way lawyers represent legal problems—will obviate, to some extent, the knowledge acquisition problem. The knowledge representation structure will be similar to that used for the FINDER system (§2.5.2): cases will be represented as points in space, the dimen­ sionality of which is the number of relevant attributes. Like FINDER, SHYSTER will allow attributes to have yes or no values. In addition, SHYSTER will allow an attribute’s value to be unknown. The smaller the distance between two points in this space, the more similar they will be considered to be. Using such a structure, SHYSTER will choose cases to use in argument on the basis of similarity. This structure will be less sophisticated than, for example, HYPO (§2.5.3). However, this simplicity can be justified on the grounds that it will greatly simplify the knowledge acquisition process—avoiding, to some degree, 54 Chapter 3: A pragmatic approach to case law the knowledge acquisition bottleneck. Despite this simpler structure, SHYSTER will still be capable of arguing with hypotheticals like HYPO, though not neces­ sarily with the same sophistication. (SHYSTER’s knowledge representation structure is detailed in §3.5 below. Dif­ ferences between SHYSTER and FINDER are explained in §3.14.1.) 3.2.4 Generality of application SHYSTER will be of general design, so that it can operate in more than one legal domain. The only restriction upon applicable domains will be that they conform to the model of legal reasoning adopted for the system. Generalizability has been claimed by several legal expert system developers,93 but to date none has been demonstrably generalizable (or even demonstrably widely applicable). SHYSTER will be designed so that once information about different areas of case law has been specified for it, it can be interrogated as to the ramifications of any or all of those areas of law for a given situation. (SHYSTER’s case law specification mechanism—a specification language—is introduced in §3.6, and its use is illustrated by example in §4.5 and §4.6.) 3.2.5 Prediction and argument SHYSTER will be designed to make a prediction about the likely result in a case. This prediction will be based upon previously decided cases, assuming (as must any legal case-based expert system) the application of the doctrine of precedent. SHYSTER will also produce legal argument supporting, and opposing, the pre­ dicted outcome. The calculations which SHYSTER uses to reach its conclusions and to construct its legal arguments will not be part of those arguments, although they will be accessible. A legal expert system’s predictive ability and its ability to construct legal argument are both important: prediction is a valuable component of legal advice, but the nature of the adversarial system requires that a lawyer be able to argue a case, and be prepared to respond to counter-arguments. Of course, SHYSTER will not be normative. A prediction will merely be a statement about the likely outcome—a statement about the relative strengths of the arguments that are constructed. (SHYSTER’s method of constructing arguments is explained in §3.11.) 3.2.6 Evaluation of advice SHYSTER’s advice will be evaluated by reference to the accuracy of its predictions and the quality of its arguments. These are two of the three criteria by which a lawyer’s advice is evaluated. (The third criterion is the cost of the advice which, for the purposes of this thesis project, is ignored.) � 3.3 A model of legal reasoning 55 As a principal method of evaluating its advice, SHYSTER will be given details on real cases and its output compared with the results and legal arguments in those cases. (This and other methods of testing and evaluation are discussed in detail in §3.13.) 3.2.7 A hybrid structure SHYSTER will be designed so that it can be linked with a rule-based system to form a hybrid system. A case-based system is inappropriate for representing statute law (see §2.4.7). Yet some researchers have recommended the use of both rules and cases to rep­ resent case law. As discussed in §2.4.4, Susskind advocates the use of rules to represent statutes and clear cases, and hints that other methods of representation could be employed for the hard cases. In effect he proposes dividing a rule-based system from a case-based system at the boundary between clear and hard cases. The major drawback with this approach is that, as explained in §2.2.5, there is considerable doubt as to whether there is such a thing as a clear case—and even if there is, no-one has devised a method of identifying one. CABARET (see §2.6) uses rules and cases to represent case law, but its struc­ ture is a mixed one. Rules are used to guide case-based arguments. Although it is not clear exactly where the two systems meet, the division falls somewhere in case law. The problem with a mixed approach is that it complicates the model of legal reasoning and, consequently, the system’s knowledge representation. Deciding exactly how to divide the law into rule-based and case-based sources is an arbitrary process. For SHYSTER, a clean and intuitive approach will be adopted. Rule-based techniques will be used only for the representation of statute law; case-based techniques will be employed for case law. This division can be clearly defined because the sources of law (statutes and cases) are clearly defined. It also simplifies the knowledge acquisition process. 3.3 A model of legal reasoning A legal expert system must be based upon a (possibly implicit) model of legal reasoning. For the development of SHYSTER, a pragmatic model of legal reasoning was adopted. That model is explained and set out here. 3.3.1 Private and public law Traditionally, a distinction has been drawn between private and public law.94 Private law concerns relationships between citizens; public law concerns relation­ ships between citizens and the state. Private law is characterized by commercial law; public law is characterized by constitutional law or international law. 56 Chapter 3: A pragmatic approach to case law This distinction has been rejected by some jurisprudents.95 However, some areas of law are more overtly concerned with matters of policy than are others. In these areas—public law areas—precedent is given less weight than it is in private law areas. In areas of private law, where predictability is crucial, the doctrine of precedent is given greater weight (at least ostensibly96). This distinction is not sharp: precedent still applies in public law, and matters of policy underlie much of private law. The difference is a matter of emphasis. However, this difference in emphasis affects the way in which lawyers reason in different areas of law. SHYSTER’s model of legal reasoning assumes the applica­ tion of the doctrine of precedent, and has no regard to matters of policy. This model reflects the way in which lawyers reason with statutes and cases in areas of private law. 3.3.2 The functions of legal reasoning Susskind identifies three functions of legal reasoning: justification, prediction and persuasion. Judges, he says, need to provide “at least ostensible reasons”97 to justify their decisions. Lawyers try to predict judicial or official behaviour, and try to persuade the courts.98 These three functions, Susskind points out, are “not fundamentally incompatible with one another”. 99 Indeed a legal expert system designer does not have to choose one of these three functions: For there is an underlying, more restricted, and yet fundamental, model of legal reasoning, common to all three accounts of function just noted, that should be at the core of all current systems.1 Susskind, it must be remembered, claims to have found a consensus in jurispru­ dential theory.2 This explains his reference to a model. There is no jurisprudential consensus, and there is no single model of legal reasoning. However, he is correct in that if a model of legal reasoning is to be used for the development of an expert system, it should be “common to all three accounts of function” noted above. 3.3.3 Adopting a model of legal reasoning For the purposes of the development of SHYSTER, the following model of the process of reasoning with statutes and case law was adopted.3 A lawyer examines the facts of the case in question—the instant case—and determines which area of law, and which statutes (if any) apply. These statutes are applied to the facts of the instant case. The meaning of a concept in a statute may be open-textured, and may determine the result of the application of that statute to the instant case. A lawyer argues about the meaning of an open-textured concept by reference to the facts of the instant case and those of previously decided cases. The results of some cases are desirable in that they ascribe a meaning to an open-textured � � � 57 � 3.3 A model of legal reasoning concept which (when the statute is applied) leads to a desired result in the in­ stant case. No two cases can be completely identical, given the plethora of facts associated with any given case. Some of these differences may be insignificant, and much of a lawyer’s reasoning by analogy concerns the legal significance of these differences. A lawyer argues with cases in the following fashion: • If the result of a previously decided case is desirable, she/he argues that there are no legally significant differences between the previous case and the instant case, so the previous case should be followed. • If the result of a previously decided case is undesirable, she/he argues that there is some legally significant difference between the previous case and the instant case upon which the previous case should be distinguished. However, as Hart points out, “the class of such differences can never be exhaust­ ively determined.”4 This model allows that some concepts in a statute may be open-textured, but assumes that these concepts are amenable to full definition—at some level of abstraction—by reference to case law. That is, open-textured statutory con­ cepts can be defined by arguing with cases, a process which (in turn) may involve arguing with more open-textured concepts. These open-textured case-based con­ cepts are also assumed to be amenable to definition by further reference to case law. The model does not allow that a case-based open-textured concept may be defined by reference to statutes. This assumption simplifies the model and is not a significant restriction upon its application. The appropriate level of abstraction below which a concept is considered to be fully defined depends upon the legal expertise of the user (as explained in §3.4 below). That part of this model that deals with cases is consistent with what Ashley describes as the standard model of analogical legal reasoning.5 This model has three steps: identifying a proper precedent; analyzing the facts and comparing and contrasting the precedent with the instant case; and determining whether the factual similarities or the differences are more important under the circumstances (i.e. deciding whether to follow or distinguish the precedent). As Ashley points out, this model provides no guidance as to which similarities and differences are more important, or for deciding between competing analogies. He argues that “HYPO’s model of analogical legal reasoning meets both of these criticisms”,6 and he is right in the sense that HYPO has a well-defined algorithm for choosing important similarities and differences and for choosing between the most analogous precedents. So too with SHYSTER, as explained in §3.7 and §3.9 below. 58 Chapter 3: A pragmatic approach to case law 3.4 Experts, users, and expert users As already discussed,7 legal expertise is essential to the development of a legal expert system. The author, who developed SHYSTER, is a computer scientist and a lawyer.8 He tested SHYSTER with the assistance of three lawyers expert in different areas of law.9 In the model of legal reasoning described in §3.3.3 above, the level of abstrac­ tion at which open-textured concepts are considered to be fully defined depends on the level of legal expertise possessed by the user of the system. Take, for example, Shannon and Golshani’s discussion of the development of a rule-based expert system. They identify one of the major problems of expert system design as “the overwhelming problem of open texture, illustrated by the swelling volume of case law.”10 In fact, they don’t see the problem as being lit­ erally overwhelming, for they propose a solution. If a predicate is open-textured, they say, the designer has two options: to rely on the user’s assessment of whether the predicate is true, or to produce new rules which define that open-textured concept. As explained in §2.4.7, the second option is inappropriate. The first option represents a very high level of abstraction at which open-textured statutory con­ cepts are deemed to be fully defined. It is, of course, not a very practical level of abstraction to choose: the only people qualified to use such a system would have no need of it. The approach adopted in SHYSTER is to assume that the user has legal ex­ pertise, though no specific expertise in the area of law that is represented.11 Legal expertise is required at the first step of the model of legal reasoning described above: i.e. determining which statute applies. Unless an expert system can cover the whole field of the law—an unlikely prospect, discussed in §3.5.1— then legal expertise is required at the top level to choose which expert system applies. SHYSTER’s ability to argue with instantiations of the facts of the instant case means that a user who is unable to answer any of SHYSTER’s questions can force SHYSTER to consider all the possibilities (see §3.11.2 below). For the remainder of this thesis, the term the legal expert is used to refer to the person who specifies areas of law for SHYSTER. The user of the system—also a legal expert, though with different expertise—is called the user. 3.5 Knowledge representation SHYSTER adopts and expands upon the approach to case law adopted by FINDER (§2.5.2). However, where FINDER gives advice only in the law of trover, SHYSTER gives advice in an area of case law specified by a legal expert. SHYSTER’s method 59 � 3.5 Knowledge representation of representing knowledge about case law was designed so as to be complex enough to allow the production of good advice, yet simple enough to facilitate knowledge acquisition and avoid the knowledge acquisition bottleneck. A program written in SHYSTER’s case law specification language is called a specification. A specification may contain any number of areas. Each area is specified in terms of the attributes in that area, leading cases, hypothetical cases (ideal points),12 and relationships between these entities. The legal expert may also specify a hierarchy of courts which applies in all areas of the specification. This hierarchy allows SHYSTER to take account of the relative ranking of leading cases when constructing its opinion. 3.5.1 Areas For SHYSTER, an area of law represents an open-textured concept. Each area has at least two results; a result is a possible resolution of the open-textured concept that the area represents.13 If the attributes within an area are also open-textured, they may be defined by reference to other areas in the same specification. For example, an area may be specified in order to define a statutory open-textured concept. That area may have several open-textured attributes which are linked to other areas, which in turn may include open-textured attributes. This linkage is achieved by binding the result from one area to a value for the open-textured attribute. There is no theoretical limit to the number of levels of areas that may be used, but every open-textured attribute must be defined, at some level, in terms of areas which have no open-textured attributes; circular definitions are not allowed. The same area may be linked to more than one attribute. This structure can be thought of as a directed acyclical graph. Each area is an internal node, with a child node for each of its attributes. The root node is the top level—usually statutory—open-textured concept. The user need only provide values for the leaf nodes; SHYSTER determines a value for each internal node until a value is obtained for the root node.14 In 1959, Mehl suggested that “a machine covering the whole field of law would be simpler and less cumbersome than a series of machines handling separate legal sectors.”15 Susskind disagrees: . . . the practical problems faced in engineering a system to function even in a limited legal domain of application are so numerous that it is likely that the only way ‘a vast field of law’ [Mehl’s words16] could be catered for is through the networking of smaller systems.17 The approach to case law adopted for SHYSTER lies somewhere between these two. Theoretically a single specification could represent a vast field of law using a large number of areas. 60 Chapter 3: A pragmatic approach to case law 3.5.2 Attributes For SHYSTER, there are two types of attribute. An open-textured attribute is called an external attribute because it is defined by reference to another area of case law: an external area. Its value is determined by the result in that external area. The value of a local attribute must be input by the user. Each external attribute has an association between each of the possible results in its external area and an attribute value. When the result of the external area is determined, the appropriate value is given to the attribute. Each local attribute has an associated question. The user gives the attribute its value by answering that question. An area’s attributes should represent all of (what the legal expert deems to be) the relevant similarities and differences between cases in that area. These may be questions of fact or—because an attribute may represent an open-textured concept—questions of law. A question of fact is represented by a local attribute; a question of law can be represented by a local or an external attribute, depending on the extent of its open-texture and the assumed legal expertise of the user. A point of law which is below the appropriate level of abstraction is considered to be fully defined (i.e. answerable by the user) and represented by a local attribute despite its open texture. In allowing attributes to be questions of fact and of law, this approach differs from those discussed in chapter 2 where an attribute is considered to be a “legally important fact.” However, this approach is a natural consequence of the model of legal reasoning adopted in §3.3.3 above.18 An attribute may be a political attribute: i.e. one which Gardner would char­ acterize as “legally irrelevant” (see §2.3.5). The fact that an attribute ought not to be relevant is not a reason to exclude it; SHYSTER is predictive, not normat­ ive. An attribute need not have been judicially enunciated for the legal expert to include it in the specification.19 So, for example, if the legal expert advises that the skin colour of the person seeking relief in a certain area of law has proved to be relevant in the leading cases, an appropriate attribute should be included despite the fact that it ought not to be relevant. Attribute direction allows any value for any attribute (local or external) to be “directed” towards a result (or results). Attribute direction indicates that the occurrence of that value for that attribute suggests that result (or results). Such an occurrence is not conclusive, merely suggestive. SHYSTER uses attribute direction as a safeguard, as explained in §3.12.4 below. 3.5.3 Leading cases and attribute values Each of SHYSTER’s areas includes details of the important cases decided in that area of the law. These are the leading cases. Not all cases in the area should be specified, only the important ones: the best cases for SHYSTER to use in legal argument. � 3.6 A specification language 61 By letting the legal expert decide which are the important cases, SHYSTER avoids the need to “screen” or “purify” the leading cases, as Mackaay and Robil­ lard do. The legal expert’s choice of cases is assumed to be a good one, even if some cases “appear to be erroneous in relation to the other ones.”20 For each leading case, the legal expert specifies various items of information including a fact vector. The fact vector is a vector of attribute values: one value for each attribute specified in the area. An attribute may have one of three values: yes, no, or unknown. The fact vector represents the relevant facts of the case—although some of these “facts” may actually be questions of law.21 Every attribute has a value in the fact vector (wherever possible, a known value: yes or no) even if that attribute was not the subject of legal argument and/or was not judicially considered in the case. This is because lawyers in their arguments, and judges in their judgments, may not address all of the matters of legal significance in a case.22 Furthermore, if a case is decided before the case in which an attribute is first judicially enunciated, that attribute is assumed to have always been important, even though not mentioned until the later case. This assumption is consistent with the declaratory theory of law. 23 That the­ ory has been strongly criticised,24 but this assumption is justified on the basis that the previously decided cases used in a specification are assumed to be con­ sistent with each other and (taken together) to represent the law as it is, not as it used to be.25 This is not to deny that judges make law, and that case law evolves over time. It is simply assumed that the body of cases which the lawyer will use in argument in a given area of law form a consistent whole. A result must be associated with each case. This corresponds to the decision reached by the court. 3.5.4 Ideal points The legal expert may also specify ideal points. An ideal point represents the best case for a given result.26 A fact vector is specified, representing the ideal combination of attribute values for that result. For an ideal point, an attribute value of unknown indicates that the value of that attribute does not matter. Only one ideal point may be specified for each result.27 It is useful to have ideal points in the case base because the leading cases may represent extreme combinations of attribute values. SHYSTER uses ideal points to provide a safe­ guard against giving erroneous advice (see §3.12.2 below). 3.6 A specification language A language was designed in which areas of case law can be specified for use by SHYSTER. Programs written in this language—specifications—reflect the struc­ ture of the SHYSTER’s knowledge representation (described in §3.5 above). 62 Chapter 3: A pragmatic approach to case law Each specification can contain any number of areas. Each area contains in­ formation on results, attributes, cases and ideal points. Results are represented as strings of characters: statements in English as to the effect of that outcome. Attributes are represented by a collection of strings explaining the effect of each possible attribute value. Local attributes have another string: a question to ask the user when determining the value of this attribute for the instant case. Leading cases are represented by citation information (names, dates, etc.) and a vector of attribute values: one value for each of the attributes specified in this area. A vector is used to represent each ideal point. Identifiers are used to link these concepts. So, for example, each result has an identifier and each leading case is linked to a result by use of one of those identifiers. Areas also have identifiers, so they can be accessed by name by a rule-based system and/or linked to external attributes in other areas. The result from an area is indicated to the rule-based system, or the external attribute in the case-based system, that invoked it using a result identifier. Identifiers are attached to the attribute values for an external attribute so that the result from one area can be bound to a value for that external attribute. Identifiers are also used to indicate attribute direction (linking a value for a particular attribute to a particular result). Each specification can include a hierarchy of courts with a list of strings: each describing a court and each with an identifier. Any case in any area in the specification can be linked to a court in that hierarchy using its identifier. In this fashion, identifiers are used to link entities within, and between, areas in the same specification. The use of this language to specify areas of case law for SHYSTER is illustrated by example in §4.5 and §4.6. A formal definition of the syntax of SHYSTER’s specification language is given in figure 4.3. 3.7 Weighting attributes In order to construct its opinion, SHYSTER quantifies the “distance” between the instant and the leading cases. These calculations, explained in §3.9 below, use weighted attributes. The question of how—indeed, whether—to weight attributes to account for their relative importance is controversial. One approach is to ask the legal expert to quantify the weight to be given to each attribute. For example, the developers of LESTER28 assign weights that, they say, “reflect the associations and relative significance an expert would attach to particular case features, with some adjustments we have arrived at through trial and error.”29 But lawyers are not used to thinking in this fashion. Ashley and Rissland write: In the legal domain, attorneys do know what [attributes] are important in a particular legal claim. Although they may be willing to say in the abstract � � � 63 � 3.7 Weighting attributes that a certain [attribute] is more important than other [attributes], they almost never will venture numerical weights to distinguish the [attributes’] importance.30 Lawyers are unwilling, or unable, to give numerical weights to attributes. If attributes are to be weighted then their weights should be determined without assistance from the legal expert. Aldenderfer and Blashfield warn: While the concept of weighting is simple, its practice is difficult, and very few guidelines exist. Williams31 describes five types of weighting, the most common being the a priori manipulation of [attributes]. Sneath and Sokal32 . . . argue strongly against a priori weighting, and suggest that the appropriate way to measure similarity is to give all [attributes] equal weight.33 However, Everitt points out that: . . . the consequences of choosing one or other of the plethora of similarity indices are in many cases equivalent to the adoption of different schemes of [attribute] weighting, and so the concept of “equal weighting” is not as simple as it seems at first sight. It should also be remembered that [at­ tributes] not included in the analysis are effectively given a zero weighting compared with those included.34 SHYSTER adopts the approach to attribute weighting used by FINDER. Each attribute value of yes is assigned a value of 1; each no is assigned a value of 0. (These values are completely arbitrary—although they must be different—as they form the basis of calculations to determine relative weights of attributes.) Each attribute is assigned a weight equal to the inverse of the variance of the numerical values of that attribute across all the leading cases. unknown values are ignored for this purpose. It would be inappropriate to give unknown a numerical value of (say) 0.5 on the basis that it is neither yes nor no. unknown is not a halfway point between yes and no; it simply indicates that a value is not known. Using the inverse of the variance is diametrically opposite to the standard approach adopted in statistical classification problems which deems high-variance variables to be the most important.35 As Tyree explains: It is not that low-variance facts are of themselves important, but that low-variance relevant facts are more important than high-variance relevant facts. They are the facts which have been included by the expert in spite of the fact that they do not appear to assist greatly in the separation of the cases into two classes.36 � � � � � � � 64 Chapter 3: A pragmatic approach to case law The aim is to quantify the importance that the law attaches to an attribute, not that attribute’s efficiency in discriminating between the leading cases.37 The variance �i 2 of the numerical values of an attribute Ai across n cases is defined as follows:38 �i 2 = 1 n (Aij − Āi)2 n j=1 ¯where Aij is the value of the ith attribute for the jth case, and Ai is the mean of all attribute values for the ith attribute. Because Aij is either 0 or 1, �i 2 ranges from 0 to 0.25. Consequently an attribute’s weight ranges from 4 to infinity. An attribute with all known values the same has zero variance and is assigned infinite weight.39 This seems paradoxical, but is actually appropriate. Consider an area of law in which a value of yes for an attribute AX is enormously suggestive of a result R. 40 In an instant case in which AX = yes it may be clear that, on the strength of AX alone, the likely result is R. Such a case may never reach a court; a result of R may be so likely that no-one seeks judicial determination of the instant case. So, it is possible that AX = no in all of the leading cases in this area. If SHYSTER is asked for its opinion on an instant case where AX = yes, it is appropriate that SHYSTER treats that attribute value as being of considerable importance. HYPO does not assign weights to attributes because, according to Ashley, the concept of an attribute’s weight “though intuitively attractive, is, on closer view, highly problematic.”41 He gives five reasons for not giving attributes numerical weights. Three of these can be easily countered in SHYSTER’s case. • Domain experts may not reason in terms of weighting schemes, especially numerical ones. Attorneys generally concede that [attributes] are useful in analyzing legal problems, but they rarely are willing to apply weights or probabilities to those [attributes]. For this reason, as discussed above, SHYSTER does not require the legal expert to weight attributes. • Weighting [attributes] is not justified by any authoritative means. Even if attorneys did assign weights to [attributes,] they would disagree on what those weights should be. In addition, attorneys would not actually be able to cite [attribute] weights in their arguments to a court because weighting is not an accepted kind of argument. Bing is unconvinced by Ashley’s reference to “authoritative means”, which he sees as criticizing a weighting scheme “for not being an objective method, or a method which is related to the legal argument.”42 SHYSTER’s method is object­ ive, though, Bing claims, “its relation to legal argument is less obvious.”43 65 � 3.7 Weighting attributes And, as explained in §3.13.2 below, SHYSTER’s advice is evaluated by refer­ ence to the accuracy of its predictions and the quality of its arguments. It does not use attribute weights in its arguments, only in choosing the cases to use in those arguments, so the extent to which its weighting method relates to legal argument is irrelevant. Furthermore, the fact that lawyers would disagree on weights is no reason not to weight attributes. It means that no weighting method can be definitive; it does not mean that no weighting method can be effective. • Reduction to numerical weights obliterates information needed for sym­ bolic comparison of cases. The business of attorneys is arguing about the competing [attributes] in the light of the precedents. If the [attributes] are collapsed into a number, there is nothing left to argue about. Although this may be true given HYPO’s approach to argument (discussed below, and in §2.5.3) it is not true of SHYSTER. SHYSTER does not collapse attribute information into a number, but uses numbers to decide which cases to use in argument. The choice of cases determines the manner in which attributes are used in argument. Two of Ashley’s reasons for not weighting attributes (both closely related) do apply to SHYSTER: • [An attribute’s] weight is highly contextual and depends on individual problem situations. Although an attorney may consider one factor gen­ erally to be more important than another, she or he is always mindful of peculiar cases where the opposite is true . . . • Premature commitment to a weighting scheme may cut off fruitful lines of inquiry. A rigid scheme may cause an attorney to overlook a factor that, although not generally important, is crucial in a particular situation.44 To address these concerns, Ashley and Rissland advocate a “symbolic least com­ mitment approach” to attribute weighting, in which weighting is postponed for as long as feasible.45 This approach is adopted in HYPO. As Ashley explains: HYPO clusters the applicable [attributes] according to how they appear in the most-on-point cases, interprets the effect of the clustered [attributes] in the light of the most-on-point cases, and criticizes and tests the inter­ pretations in the light of the salient differences among the most-on-point cases by distinguishing precedents, citing counterexamples, and posing hy­ potheticals that change magnitudes and combinations of [attributes] in the problem . . . Although by the end of [this process] HYPO has not actually assigned weights to the competing [attributes], it has dealt symbolically with the problem of weighting. It has generated precedent-citing arguments in favor of alternative interpretations of the weights of the [attributes] within the context of the problem.46 66 Chapter 3: A pragmatic approach to case law SHYSTER’s simple knowledge representation means that HYPO’s approach to weighting is inappropriate for SHYSTER. Nevertheless, there can be no doubt that, in the legal domain, some attributes are of greater importance than others. For this reason, and despite the problems identified by Ashley, SHYSTER weights its attributes. However, it is important not to lose sight of what Lambert and Grunewald call “the necessary arbitrariness of any weighting function.”47 SHYSTER’s approach to weighting is discussed in §6.3 in the light of the case studies performed in chapter 5. 3.8 Detecting attribute dependence A functional dependence exists between two attributes when there is a function which maps the values of one of the attributes directly to the values of the other. A stochastic dependence exists when the occurrence of one event affects the prob­ ability of the occurrence of another event. Functional dependence or stochastic dependence between attributes may indicate shortcomings in the legal expert’s specification. Consider this (extreme) example. The legal expert chooses two attributes which, although worded differently, are identical: i.e. they ask precisely the same question, in different ways. The values of these two attributes across the leading cases are identical, and each attribute is assigned the same weight. By effectively choosing the same attribute twice, the legal expert has given that attribute twice its appropriate weight. In this example, there is both a functional dependence and a stochastic de­ pendence between the two attributes.48 Alternatively, if two attributes differ only very slightly in their values across the cases then there is a stochastic depend­ ence (though no functional dependence) between them: i.e. the occurrence of a given value for one attribute affects the probability of the occurrence of a given value for the other. The legal expert may have chosen two very similar attrib­ utes. Dependency also exists where two attributes are completely, or very nearly completely, different. For every pair of attributes there is a pair of attribute values corresponding to each of the leading cases. SHYSTER detects attribute dependence by examining the known pairs (pairs where both the attribute values are known) for each pair of attributes in the area. If either, or both, of a pair of attribute values is unknown then both values are ignored. (An unknown value is not a value that could form part of a dependency; it is an absence of known values.) Attribute dependence does not necessarily mean that the specification must be re-written. Even if two attributes have exactly the same values across the leading cases, the legal expert may decide that they do not ask the same question: i.e. that a case can be imagined where the values of those two attributes are different, despite the fact that they do not differ in any of the leading cases.49 SHYSTER 67 � 3.8 Detecting attribute dependence Equivalence function: yes ≡⇐ yes, no ≡⇐ no Inverse function: yes ≡⇐ no, no ≡⇐ yes yes function: yes ≡⇐ yes, no ≡⇐ yes no function: yes ≡⇐ no, no ≡⇐ no Figure 3.1: The four forms of functional dependence. Only the equivalence and inverse functions are detected by SHYSTER; the yes and no functions do not indicate a relationship between two attributes. warns the legal expert of attribute dependence so that she/he can reconsider her/his choice of attributes. It is not possible to prove that two attributes are stochastically independent, hence it is not possible to detect stochastic dependence with complete certainty. However, SHYSTER can detect evidence of stochastic dependence, and warn the legal expert that such evidence exists. 3.8.1 Functional dependence Detecting functional dependence between attributes is straightforward. There are only four functions which map an attribute AX to an attribute AY . These functions—the four forms of functional dependence—are described in figure 3.1. The equivalence function produces the same value in AY as in AX ; the inverse function produces the opposite value; the YES function sets AY to yes regardless of the value of AX ; and the NO function sets AY to no regardless of the value of AX . SHYSTER checks all pairs of attributes, and warns of any equivalence func­ tion, or inverse function. yes functions and no functions are ignored. Each is a constant function: i.e. each produces the same result regardless of its argu­ ment. Hence, a yes or no function does not indicate a relationship between two attributes.50 3.8.2 Stochastic dependence Two events E1 and E2 are said to be stochastically independent if the probability of their both occurring is equal to the product of the probabilities of each of them occurring: P (E1 ⇒ E2) = P (E1) P (E2). There are four types of known pair: yes/yes, yes/no, no/yes, and no/no. Let N be the number of known pairs in the attributes AX and AY . Let x be the number of yess in AX , and let y be the number of yess in AY . Let n be the = � �� � � �� �� � � � � 68 Chapter 3: A pragmatic approach to case law AX : no no no no no yes yes yes yes yes no no yes AY : no yes no no no no yes yes yes yes yes no yes Figure 3.2: Example attribute values for two attributes AX and AY . number of yes/yes pairs for the attributes AX and AY . Because known attribute values are binary, the number of any type of known pair is linked to the numbers of all other types, and can be expressed in terms of N, x, y and n: number of yes/yes pairs = n (by definition), number of yes/no pairs = x − n, number of no/yes pairs = y − n, number of no/no pairs = N − x − y + n. So, in order to detect evidence of stochastic dependence it is only necessary to examine the expectation and actual occurrence of one of the four possible pairs. SHYSTER examines only the yes/yes pairs. Let P (n) be the probability of there being exactly n yes/yes pairs given N, x and y, and assuming random data: number of configurations with exactly n yes/yes pairs P (n) = total number of configurations N N−n N−x n x−n y−n N N � �� x �y y N−y = n � x�−n . N x Because the variables N, x, y, and n specify a combination of attribute values, P (n) is the probability of that combination occurring. The manner in which SHYSTER uses this formula for P (n) to check for evidence of stochastic dependence between attributes is best illustrated by example. Consider two attributes AX and AY , with known attribute values as shown in figure 3.2.51 There are thirteen known pairs: N = 13. There are six yess in AX : x = 6. There are seven yess in AY : y = 7. The formula for P (n) can be used to calculate the probability that n = i for i = 0 . . . N . Figure 3.3 lists these probabilities in the column labelled “P (n = i)”. There are five yes/yes pairs in AX and AY : n = 5. Figure 3.3 shows that the probability of there being exactly five yes/yes pairs (given the distribution of yess and nos in AX and AY , and assuming random data) is 0.0734. 69 � 3.8 Detecting attribute dependence i 0 1 2 3 4 5 6 7 . . . 13 P (n = i) 0.0006 0.0245 0.1836 0.4079 0.3059 0.0734 0.0041 0.0000 . . . 0.0000 P (n ∗ i) 0.0006 0.0251 0.2087 0.6166 0.9225 0.9959 1.0000 1.0000 . . . 1.0000 P (n ≤ i) 1.0000 0.9994 0.9749 0.7913 0.3834 0.0775 0.0041 0.0000 . . . 0.0000 Figure 3.3: The probabilities for the example attributes AX and AY defined in figure 3.2. N = 13, x = 6, y = 7. But, the probability of there being exactly five yes/yes pairs is not of much use on its own. What is important is whether the number of yes/yes pairs is unusually high, or unusually low. The “P (n ∗ i)” column gives the probability of there being i yes/yes pairs or fewer: i.e. the cumulative total of the “P (n = i)” column. The “P (n ≤ i)” column gives the probability of there being i yes/yes pairs or more: i.e. the cumulative total of the “P (n = i)” column, summing backwards from the bottom. The probability of there being five yes/yes pairs or fewer is 0.9959. The probability of there being five yes/yes pairs or more is 0.0775. It is necessary to choose a threshold of likelihood, below which a given number of yes/yes pairs will be considered unusually high or unusually low. The method that SHYSTER uses to detect stochastic dependence is analogous to using the σ2 (“chi-square”) test for independence,52 except that that test assumes a normal distribution and, hence, a large sample while the test used by SHYSTER calculates probabilities exactly and does not require a large sample. As such SHYSTER’s method is an example of what is called Fisher’s exact test.53 When using the σ2 test, it is common practice to use a threshold of 0.05. Us­ ing this threshold, the occurrence of five yes/yes pairs in AX and AY is neither unusually high nor unusually low. This means that the combination of attrib­ ute values in those attributes is not unusual: there is no evidence of stochastic dependence between AX and AY . Had there been six yes/yes pairs, that would have been unusually high be­ cause P (n ≤ 6) = 0.0041. One or no yes/yes pairs would have been unusually low, because P (n ∗ 1) = 0.0251. � � � 70 Chapter 3: A pragmatic approach to case law 3.9 Calculating distances SHYSTER’s choice of the cases with which to construct its arguments is based upon a notion of similarity between cases. SHYSTER quantifies this similarity using distance measures: the smaller the distance between two cases, the more similar they are. SHYSTER calculates two different types of distance: the known distance is defined as the sum of the weights of every attribute for which those two cases have different known values; the unknown distance is defined as the sum of the weights of every attribute for which either of the two cases has an unknown value. A known distance of zero indicates that the two cases are identical—at least as far as the known attributes are concerned. A large unknown distance indic­ ates that the values of some important (i.e. heavily weighted) attributes were unknown, casting some doubt on the reliability of the known distance calculation for those two cases. The unknown distance can be thought of as a measurement of possible error: it is the maximum distance that could be added to the known distance if all of the unknown attribute values were known. Known and unknown distances are calculated between the instant case and each of the leading cases. These distance measurements are treated as character­ istics of each leading case. So, for example, a statement that case j has a smaller known distance than case k means that the known distance between case j and the instant case is less than the known distance between case k and the instant case. Statisticians make use of many different similarity measures. However, two as­ pects of SHYSTER’s approach to case law mean that each of the commonly used similarity measures reduces to one of three measures—six, allowing for attribute weighting. This is because SHYSTER’s known values are binary, and because its choice of cases is based on the relative distance/similarity between cases: i.e. it has regard to the fact that a case j is further from a case k than a case �, but not how much further. The decision to use known and unknown distance as SHYSTER’s similarity measure is explained in §3.9.3 below, after a brief survey of the different tech­ niques of measuring similarity. 3.9.1 Similarity measures Cluster analysis is the separation of data into groups on the basis of similarity. Entities are grouped so that two entities in the same group are more similar than two entities in different groups. Similarity measures are used to quantify the similarity between every entity and every other entity. SHYSTER does not • � 71 � 3.9 Calculating distances perform cluster analysis, but it does quantify the similarity between the instant case and each of the leading cases. This process takes O(n) time, where n is the number of cases, whereas cluster analysis takes O(n2) time. Many different types of similarity measure have been developed. The most widely used measures are called metrics. Aldenderfer and Blashfield explain: The quantitative estimation of similarity has been dominated by the con­ cept of metrics; this approach to similarity represents cases as points in a coordinate space such that the observed similarities and dissimilarities of the points correspond to metric distances between them . . . The dimen­ sionality of the space is determined by the number of [attributes] used to describe the cases.54 For a similarity measure to be a metric it must satisfy the following four criteria:55 • Given two cases j and k, djk = dkj ≤ 0, where djk is the distance between case j and case k. • Given three cases j, k and �, djk ∗ dj� + dk� (this is called the triangle inequality or the metric inequality). Given two cases j and k, if djk = 0, then j is not identical to k. • Given two identical cases j and j�, djj� = 0. Many researchers have argued against the use of similarity measures which do not meet these criteria. Aldenderfer and Blashfield point out that: [Measures] that are not metrics may not be jointly monotonic; that is, the values of different [measures] used with the same data will not necessarily vary conjointly, raising the disturbing issue that these [measures] could suggest quite different relationships among the entities.56 However, it is not essential that a similarity measure be a metric. For ex­ ample, the Pearson product-moment correlation coefficient (“a popular similarity measure”,57 discussed below) is not a metric. There are four different kinds of similarity measures: distance measures, as­ sociation coefficients, correlation coefficients, and probabilistic similarity coeffi­ cients. Strictly speaking, probabilistic similarity coefficients do not actually calculate the similarity between two cases. They take into account the distribution of the frequencies of the attribute values over all the cases, and are calculated during the formation of clusters.58 Probabilistic similarity coefficients are inappropriate for use with SHYSTER because it does not perform cluster analysis. The three kinds of similarity measure appropriate for SHYSTER are distance measures, association coefficients, and correlation coefficients. ���� � � � � � 72 Chapter 3: A pragmatic approach to case law Distance measures Two popular distance measures are Euclidean distance and Manhattan distance. The Euclidean distance between two cases j and k is defined as n Aij − Aik � i=1 �2 where Aij is the value of the ith attribute for the jth case,59 and n is the number of attributes.60 The Manhattan distance is defined as n ⎬⎬⎬ ⎬⎬⎬Aij − Aik . i=1 Both Euclidean distance and Manhattan distance are specific examples of the class of metric distance functions known as Minkowski metrics. The Minkowski metric is defined as n �1/K⎬⎬⎬ ⎬⎬⎬ K Aij − Aik i=1 where K is some constant: for Euclidean distance, K = 2; for Manhattan dis­ tance, K = 1. In the binary case (e.g. SHYSTER), ⎬⎬⎬ ⎬⎬⎬ 0, if Aij = Aik;K Aij − Aik = 1, if Aij =� Aik. Hence, for binary attribute values, the Minkowski metric reduces to �jk 1/K where �jk is the number of differences in the corresponding attribute values of case j and case k. For increasing values of �jk, this metric always increases regardless of the value of K. As SHYSTER is concerned only with the relative distance between cases, it does not matter which value of K is used. Choosing K = 1 gives a distance measure djk: djk = �jk. This distance measure is a metric as it satisfies all four metric criteria. It ranges from 0 to n: the smaller the value of djk, the nearer case j is to case k. Association coefficients Association coefficients are used to describe similarity between cases with binary attributes. These coefficients are usually expressed in terms of a, b, c and d; the two-way association table in figure 3.4 defines these variables. 73 � 3.9 Calculating distances case k yes no a is the number of yes/yes pairs; a b case j yes b is the number of yes/no pairs; etc. no c d Figure 3.4: A two-way association table defining a, b, c and d. Many different association coefficients have been proposed—Aldenderfer and Blashfield state that there are more than thirty.61 According to Everitt, the reason for this proliferation of coefficients is: . . . uncertainty over how to incorporate negative matches [d] into the coef­ ficients, and also whether or not matched pairs of [attributes] are equally weighted, or carry twice the weight of unmatched pairs, or unmatched pairs carry twice the weight of matched pairs.62 Everitt gives several examples of association coefficients. Jaccard’s coefficient, a , a + b + c is one measure of the association between two cases j and k. It ignores d—the number of no/no pairs. This coefficient was developed for use in applications where it would be inappropriate to treat two cases as being similar for lacking the same features as well as sharing the same features.63 Similarly, 2a a and 2a + b + c a + 2(b + c) ignore negative matches—and give double weight to matched pairs and un­ matched pairs, respectively. And a a + b + c + d gives no positive significance to negative matches. Coefficients such as these are inappropriate for use by SHYSTER because the occurrence of no/no pairs is as important as the occurrence of yes/yes pairs. An attribute value of no does not indicate the lack of a feature; it means that the answer to the attribute’s question is “no.” Simply rephrasing attribute questions could turn all yes/yes pairs into no/no pairs. 74 Chapter 3: A pragmatic approach to case law The other two example coefficients that Everitt gives are a + d 2(a + d) and . a + b + c + d 2(a + d) + b + c The first, called the simple matching coefficient, is the number of matching pairs as a fraction of the total number of pairs. The second gives matching pairs twice the weight of non-matching pairs. Coefficients such as these are appropriate for use by SHYSTER because they do not distinguish between yes/yes pairs and no/no pairs. As �jk = b+c and n = a+b+c+d, coefficients of this kind can be generalized to K(n − �jk) K(n − �jk) + �jk where K is some constant. For increasing values of �jk, this coefficient always decreases regardless of the value of K. Because SHYSTER is concerned only with relative measures of similarity, it does not matter which value of K is used. The simplest is K = 1, giving n − �jk n which ranges from 0 to 1. Subtracting it from 1 yields an inversely proportional association coefficient Sjk which satisfies all four metric criteria: �jk Sjk = . n This association coefficient also ranges from 0 to 1: the smaller the value of Sjk, the nearer case j is to case k. (Note that, in SHYSTER, n may vary within the same area of law because n is the number of pairs of attributes with known values. Hence, Sjk is not proportional to djk as derived above.) Correlation coefficients The most popular correlation coefficient is Pearson’s product-moment correlation coefficient.64 This coefficient is defined as ⎫ � �� �n ¯ ¯Aij − Aj Aik − Ak i=1 rjk = ⎪ n � �2 n � �2⎫ ⎫¯ ¯Aij − Aj Aik − Ak i=1 i=1 where n is the number of attributes, Aij is the value of the ith attribute for the ¯jth case, and Aj is the mean of all attribute values for the jth case. � � 75 � 3.9 Calculating distances In general, this correlation coefficient has several limitations.65 It is not a true metric.66 Furthermore, as Aldenderfer and Blashfield explain: . . . the use of [this] method to calculate the correlation of cases does not make statistical sense, because one must obtain the mean value across different [attribute] types rather than across cases, as in the standard use of the method. The meaning of the “mean” across these [attributes] is far from clear.67 Although this is true in general, in SHYSTER all attributes are of the same type. Hence it does make sense to calculate the mean across the attributes. Despite its drawbacks this coefficient has been widely used. It ranges from −1 to 1:68 the larger the value of rjk, the nearer case j is to case k. 3.9.2 Weighted similarity measures Each of the three similarity measures discussed in §3.9.1 above can be weighted to take account of the importance of each attribute. The weighted distance measure d� isjk n ⎬ ⎬ d�jk = ⎬⎬Aij − Aik⎬⎬ × wi i=1 ⎫ where wi is the weight of the ith attribute. It ranges from 0 to i n =1 wi: the smaller the value of djk � , the nearer case j is to case k. The weighted association coefficient Sjk � is n⎫⎬⎬ ⎬⎬⎬Aij − Aik⎬ × wi S� = i=1 ⎫ .jk n wi i=1 It ranges from 0 to 1: the smaller the value of Sjk � , the nearer case j is to case k. The weighted correlation coefficient r� isjk ⎫ � �� �n Aij × wi − Ā�j Aik × wi − Ā�k = ⎪ i=1 rjk� n � �2 n � �2⎫ ⎫ Aij × wi − Āj� Aik × wi − Āk� i=1 i=1 where Ā�j is the weighted mean of all attribute values for the jth case. It ranges from −1 to 1: the larger the value of rjk, the nearer case j is to case k. � � � 76 Chapter 3: A pragmatic approach to case law 3.9.3 Choosing a similarity measure For SHYSTER—and allowing for the weighting of attributes—each of the com­ monly used similarity measures reduces to one of six measures: djk, d� �jk, Sjk, Sjk, rjk and r�jk. Mackaay and Robillard (§2.5.1) use djk as their similarity measure; FINDER (§2.5.2) uses d�jk. According to Ashley, HYPO (§2.5.3) uses four comparison metrics:69 The basic measure is on pointness: the degree of overlap of dimensions shared by the instant case and a given case, relative to that of other cases. HYPO also has regard to the outcome of a leading case (because “depending on the procedural context of the case, some outcomes are more determinative than others”70), the magnitude of shared dimensions, and a case’s potential relevance as a near miss. A near miss case is examined to see whether a small hypothetical change would make that case more on point. Strictly speaking, HYPO uses only one similarity measure: on-pointness. The outcome of a case and the magnitude of shared dimensions are weighting con­ siderations;71 making hypothetical changes to a case is an argument technique. None of these three is a similarity measure (in the sense that that term is used in this thesis). When choosing an appropriate similarity measure for a particular set of data, it is important to have regard to the sort of data, and to what is to be measured. Because SHYSTER treats cases as points in n-dimensional space, where n is the number of attributes, a metric should be an appropriate similarity measure. As discussed in §3.7 above, attribute weighting is also appropriate. SHYSTER’s known and unknown distance measures are variations on the weighted distance measure d� which is a metric. jk However, it is important to note Sneath and Sokal’s warning that “when all is said and done, the validation of a similarity measure . . . in a given field has so far been primarily empirical”. 72 The testing of SHYSTER is described in chapter 5. In order to compare em­ pirically the various similarity measures, SHYSTER also calculates values of djk, Sjk, Sjk � , rjk and r� These extra measures are used as safeguards (see §3.12.1jk. below). 3.9.4 Infinite distance As explained in §3.7 above, each attribute’s weight ranges from 4 to infinity. In determining a distance—known or unknown—SHYSTER may have to deal with one or more attributes with infinite weight. To handle such possibilities, SHYSTER’s distances have an infinite and a finite component. The infinite component is the number of infinitely weighted attrib­ utes that differ between the two cases. A distance of “2� + x” is considered to be 77 � 3.10 Nearest cases and nearest results greater than “� + y”, regardless of the values of x and y (the finite components). The infinite component of this first distance should not be thought of as being the sum of two infinities. It represents the effect, upon the distance between two cases, of an infinite weight in each of two dimensions in n-dimensional space. 3.10 Nearest cases and nearest results Having determined the distance between the instant case and each of the leading cases, SHYSTER can decide which of the leading cases is nearest to the instant case. Having chosen a nearest case, SHYSTER can decide upon the most likely result: the result in the nearest case. In fact, SHYSTER deals with several nearest cases and can deal with several cases which are equidistant from the instant case. 3.10.1 Nearest cases The nearest known neighbour is the case with the smallest known distance, and with no unknown distance (i.e. with an unknown distance of zero). The nearest unknown neighbour is the case which has the smallest sum of the known and unknown distances, and non-zero unknown distance. The nearest of these, the case with the smallest sum of known and unknown distances, is called the nearest neighbour. Consider the example distances in figure 3.5, taken from one of the tests performed in chapter 5.73 There are fourteen cases: C1 . . . C14. Each case has a known and an unknown distance. Zero distances are indicated by “–”. The nearest known neighbour (and the nearest neighbour) is C5; the nearest unknown neighbour is C3. 3.10.2 Nearest results The result of the nearest neighbour is termed the nearest result. Applying the doctrine of precedent, SHYSTER assumes that the decision in the instant case will be the same as that in the nearest neighbour. SHYSTER also finds the nearest known and nearest unknown neighbours for every other result—i.e. from amongst the cases in which another result was reached. These are termed the nearest known other and the nearest unknown other. The nearest of these cases is called the nearest other ; there is a nearest other for each other result. The example cases whose distances are given in figure 3.5 are grouped by result: the first seven cases have one result, the second seven have another.74 The nearest result is the result of the nearest neighbour: C5. The nearest known other (and the nearest other) is C12; the nearest unknown other is C13. 78 Chapter 3: A pragmatic approach to case law Case C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 Known distance 43.57 36.45 17.76 42.85 17.09 14.92 59.32 54.48 71.79 55.74 55.21 46.13 28.54 49.58 Unknown distance – – 5.63 4.02 – 62.65 – – – – – – 54.97 – Figure 3.5: Distances for fourteen example cases. The cases are grouped by res­ ult: the first seven have one result, the second seven have another. The nearest known neighbour (and the nearest neighbour) is C5; the nearest unknown neigh­ bour is C3. The nearest known other (and the nearest other) is C12; the nearest unknown other is C13. 3.10.3 Equidistance In each comparison made during these classifications of cases and results it is possible that the distances being compared are equal. If two distances are the same then their cases (or results, as applicable) are said to be equidistant, because they are equidistant from the instant case. SHYSTER allows any number of equidistant cases in all of the categorizations explained above. Equidistant cases are used in argument (one after another, in order of importance) wherever a single case would be used. SHYSTER also allows equidistant results—results whose nearest neighbours are equidistant—but only if those results are “other” results. There can only be one nearest result so that the open-textured concept can be defined for use by other areas in the case base or by a rule-based system. Mackaay and Robillard consider the problem of equidistance and propose two solutions.75 The nearest result can be that which has the greatest number of equidistant cases. If there is no majority, then no prediction can be made. Alternatively, the set of nearest neighbours can be repeatedly extended to include 79 � 3.11 Writing a report the next nearest neighbour(s) until a majority is achieved. This approach is not adopted for use with SHYSTER. If there is a great distance between the nearest neighbours and the next nearest neighbours, it is inappropriate to use those next nearest neighbours to resolve the problem. SHYSTER chooses between equidistant results by reference to the relative im­ portance of the equidistant cases and (if further resolution is required) the year in which those cases were decided (more recently decided are considered to be more important). If the equidistance remains unresolved, SHYSTER refuses to venture an opinion, writes an error message and stops. It was intended to add further comparisons, resolving equidistance by reference to the nearest ideal points, the nearest centroids (centroids are described in §3.12.3 below), and attribute direc­ tion. However, testing of SHYSTER indicates that the two stage approach adopted is adequate. The utility of going further is doubtful; to make further distinctions would be to split a very fine hair. SHYSTER uses equidistance as a safeguard against giving bad advice (see §3.12.5 below). 3.11 Writing a report SHYSTER constructs a report—a legal opinion—about the instant case. This opinion includes a statement as to the likely result in the instant case, and jus­ tification of that statement. This justification is crucially important. As Tyree, Greenleaf and Mowbray point out: . . . the justification in a legal system is the main product, for the justi­ fication is no more and no less than the legal arguments which support the suggested outcome. It is in the nature of legal reasoning that these arguments must also address the support for the opposite outcome. It is these arguments which, if the matter goes to court, must be presented for adjudication.76 SHYSTER’s approach to report generation is based on that of FINDER, but SHYSTER takes account of unknown distance and equidistance, and argues about the effect upon its argument of hypothetical changes to the attribute values of the instant case. This last feature allows SHYSTER to adopt, to some extent, aspects of HYPO’s approach to arguing with hypotheticals. HYPO: Summarizes the cases that can be cited in favor of a position, characterizes how strongly they support the position, focuses the attorney’s attention on the most significant cases and hypotheticals, . . . and facilitates comparing arguments between cases and hypotheticals . . . 77 So too does SHYSTER, although the sophistication of its reporting is limited by its main advantage: the simplicity of its knowledge representation. 80 Chapter 3: A pragmatic approach to case law 3.11.1 Arguing with the instant case SHYSTER opens its report with some introductory comments about the area of law, then boldly declares its opinion78 that the result in the instant case will be the nearest result. It then uses the nearest neighbours and nearest results to justify that statement, before closing with some concluding remarks about the area. The opening and closing remarks are general comments provided by the legal expert. How SHYSTER chooses the cases to use in argument, and how those cases are used, is described below. Choosing cases Figure 3.6 gives a pseudo-code description of SHYSTER’s algorithm for choosing leading cases to use in argument, and the order in which to use those cases. SHYSTER will always use the nearest known neighbour in argument. It will also use the nearest unknown neighbour in two circumstances: if it is the nearest neighbour, or if it would be the nearest neighbour but for its unknown distance. As defined in §3.9 above, the unknown distance can be thought of as a meas­ urement of error: it is the maximum distance that could be added to the known distance if all of the unknown attribute values were known. Whichever of the nearest known neighbour and the nearest unknown neighbour is the nearer is used first. A slightly different approach is used with the nearest others. For every other result SHYSTER uses the nearest known other in argument. The nearest unknown other is used in two circumstances: if it is the nearest other, or if it would be nearer the instant case than the nearest neighbour—not just the nearest other— were it not for unknown distance. The nearest unknown other is used in this second circumstance because, if all of the unknown attribute values were known, the nearest unknown other could be the nearest neighbour and SHYSTER’s opin­ ion as to the likely result would be different. The nearest known other is used in argument before the nearest unknown other, except in either of these two circumstances. This description is simplified in one respect: it assumes that there is only one nearest known case and one nearest unknown case for each result. If there are two or more equidistant cases, SHYSTER uses each case—one after the other, in order of importance. Using cases How each leading case is used in argument varies depending on several factors. Figures 3.7, 3.8 and 3.9 give pseudo-code descriptions of SHYSTER’s algorithm for using cases. Each description is a refinement of some steps in the algorithm � � ������� � ������� � � � ���������� � ���������� 81 � 3.11 Writing a report FOR the nearest result DO IF the nearest known neighbour is the nearest neighbour THEN use the nearest known neighbour; IF (were it not for unknown distance) the nearest unknown neighbour would be nearer the instant case than the nearest neighbour THEN use the nearest unknown neighbour; END A ELSE use the nearest unknown neighbour; use the nearest known neighbour; END END FOR every other result DO B IF the nearest unknown other is the nearest other OR (were it not for unknown distance) the nearest unknown other would be nearer the instant case than the nearest neighbour THEN C use the nearest unknown other; END use the nearest known other; END Figure 3.6: SHYSTER’s algorithm for choosing the cases upon which to base its opinion. The nearest known neighbour is the case with the smallest known distance, and with no unknown distance; the nearest unknown neighbour is the case which has the smallest sum of known and unknown distances, and non-zero unknown distance. The nearest neighbour is the case with the smallest sum of known and unknown distances. The nearest result is the result of the nearest neighbour. Similarly, for every other result there is a nearest known other and a nearest unknown other ; the nearest of these is the nearest other. The steps marked A are refined in figure 3.7; B in figure 3.8; C in figure 3.9. For simplicity, this description and the descriptions in these other figures assume no equidistance. for choosing cases described in figure 3.6. Words within quotation marks are paraphrasings of words that SHYSTER uses in its opinion. Each case is summarized, then the similarities and differences between the case and the instant case are explained. If the case is an unknown case, the attributes for which values are unknown are also detailed. 82 Chapter 3: A pragmatic approach to case law IF (were it not for unknown distance) the nearest unknown neighbour would be nearer the instant case than the nearest neighbour THEN predict that the result will be the nearest result, citing the nearest known neighbour and the nearest unknown neighbour; ELSE predict that the result will be the nearest result, citing just the nearest known neighbour; END FOR the nearest known case DO summarize the case; IF there is no distance between the case and the instant case THEN ‘‘the two cases are identical”; ELSE list the similarities between the two cases; list the differences between the two cases; ‘‘nevertheless, the case should still be followed”; END END IF (were it not for unknown distance) the nearest unknown neighbour would be nearer the instant case than the nearest neighbour THEN cite the nearest unknown neighbour as another case in which the nearest result was reached; FOR the nearest unknown neighbour DO summarize the case; IF there is some known distance between the case and the instant case THEN list the similarities between the two cases; list the differences between the two cases; ELSE ‘‘the two cases may be identical, and . . .”; END ‘‘I would have suggested, that this case be followed instead of the nearest neighbours except that . . .”; list the unknown attributes; END END Figure 3.7: Part of SHYSTER’s algorithm for using cases in argument: a refinement of the steps marked A in the algorithm described in figure 3.6. � � � 83 � 3.11 Writing a report predict that the result will be the nearest result, citing the nearest unknown neighbour and the nearest known neighbour; FOR the nearest unknown neighbour DO summarize the case; list the similarities between the case and the instant case; list the differences between the two cases; list the unknown attributes; ‘‘nevertheless, the case should still be followed”; END FOR every nearest known neighbour DO summarize the case; list the similarities between the case and the instant case; list the differences between the two cases; ‘‘nevertheless, the case should still be followed”; END Figure 3.8: Part of SHYSTER’s algorithm for using cases in argument: a refinement of the steps marked B in the algorithm described in figure 3.6. For the nearest result, it is argued that (because of the similarities, and despite the differences) the result in the instant case should be the same. For every other result, it is argued that (because of the differences, and despite the similarities) the result in the instant case should be different. If SHYSTER’s opinion is a desirable result for the user, she/he can use SHY­ STER’s discussion of the nearest case, and the differences between the nearest case and the instant case, as the basis for a legal argument. Alternatively, if SHYSTER’s opinion is not a desirable result for the user, she/he can base a legal argument upon SHYSTER’s discussion about the nearest others. This description also assumes that there is only one nearest known case and one nearest unknown case for each result. Two or more equidistant cases are used one after the other. The report makes no reference to “nearest cases,” “nearest others,” etc. No weights or distance measures are included or discussed. The calculations which SHYSTER uses to reach its conclusions are not part of its report, although they are written to various intermediate files. The report refers only to the similarities and differences between the instant case and the leading cases. 84 Chapter 3: A pragmatic approach to case law IF the nearest unknown other is the nearest other OR (were it not for unknown distance) the nearest unknown other would be nearer the instant case than the nearest neighbour THEN ‘‘if the nearest unknown other and the nearest known other are followed then the result will be this (other) result”; IF (were it not for unknown distance) the nearest unknown other would be nearer the instant case than the nearest neighbour THEN FOR the nearest unknown other DO summarize the case; IF there is some known distance between the case and the instant case THEN list the similarities between the two cases; list the differences between the two cases; ELSE ‘‘the two cases may be identical, and . . .”; END ‘‘I would have suggested, that this case be followed instead of the nearest neighbours except that . . .”; list the unknown attributes; ‘‘nothing in the case warrants changing the prediction”; END ELSE FOR the nearest unknown other DO summarize the case; list the similarities between the case and the instant case; list the differences between the two cases; list the unknown attributes; ‘‘nothing in the case warrants changing the prediction”; END END ELSE ‘‘if the nearest known other is followed then the result will be this (other) result”; END FOR the nearest known other DO summarize the case; list the similarities between the case and the instant case; list the differences between the two cases; ‘‘nothing in the case warrants changing the prediction”; END Figure 3.9: Part of SHYSTER’s algorithm for using cases in argument: a refinement of the steps marked C in the algorithm described in figure 3.6. 85 � 3.11 Writing a report Both FINDER and HYPO report on the similarities and differences in this fashion. However, as Ashley concedes: We expect more from law students’ explanations . . . (and, presumably, from lawyers’ explanations) . . . of why a precedent should or should not be followed than a discussion of the superficial, factual similarities and differences associated with factors. We expect their explanations to invoke some principled analysis of why the similarities and differences matter and to structure their explanations to reflect the relevant statutes and court-made rules. Thus, at first glance, HYPO’s approach to identifying important similarities and differences may not seem philosophically satisfying . . . 79 The same criticism applies to SHYSTER, but is countered on the pragmatic grounds that the aim is to produce a working expert system, and to obviate the problem of knowledge acquisition. Similarly, Ashley makes a pragmatic ar­ gument in defence of HYPO: Undoubtedly fundamental legal principles play a role in legal analogical reasoning . . . But one cannot hope to model that kind of adversarial reas­ oning until one understands the simpler, more factual analogical compar­ isons among precedents . . . 80 3.11.2 Arguing with instantiations Apart from using unknown distance, SHYSTER has a second method of taking unknown attribute values into account. SHYSTER instantiates the unknown attribute values in the instant case to create instantiations of the instant case in which all the attribute values are known. SHYSTER treats each instantiation as if it were a new instant case, and determines the nearest cases and nearest results. Because there are only two known values there are 2n different instantiations, where n is the number of unknowns in the instant case. To avoid writing un­ necessarily long reports, SHYSTER only reports on an instantiation if its nearest result is different to that of the instant case. Instantiation is a useful feature. If the user is unable to answer an attribute question, she/he simply answers “unknown.” If having known values for the unknown attribute values could make a difference to the result, the relevant instantiations are reported on in full.81 This feature can also be used where the user knows the answer to an attribute question, but wants to test the effect of providing a different answer.82 Instantiation is used as a safeguard against giving erroneous advice (see §3.12.6 below). It is also used in chapter 5 to perform generated tests (which are described in §3.13.4). 86 Chapter 3: A pragmatic approach to case law 3.11.3 Arguing with hypotheticals HYPO generates hypotheticals: hypothetical variations of the instant case that are stronger or weaker for a particular side. There are five heuristics for modifying the instant case: make a near-miss dimension apply; strengthen or weaken a case along an applicable dimension; move a case along a related dimension; make a case extreme along a dimension; and make a case into a near-win given a target.83 Ashley identifies several uses for hypotheticals in legal argument: to factor a complex situation into component parts (e.g. by exaggerating strengths, weaknesses, or by hypothetically eliminating features); to create a test case that puts an issue or pits competing attributes against each other; to present, support, and attack positions in an argument (e.g. by testing consequences of a tentative conclusion); etc.84 Ashley describes hypothetical reasoning as the key to “exploring the dialectics between cases and principles or between cases and rules.”85 Whether or not it is appropriate to think of cases as rule exemplars (as Ashley seems to), SHYSTER’s simplified representation of case law does not allow it to reason with hypotheticals to the same extent as does HYPO. However, SHYSTER does examine hypothetical variations in order to alert the user to the effect of such variations. Showing how a case can be strengthened and weakened can be particularly useful where there is some uncertainty surrounding one or more “known” attribute values. For SHYSTER, the number of possible hypothetical variations upon instant case is 2n, where n is the number of known attribute values.86 Generating all the possible hypothetical variations is of little use, as well as being computation- ally intensive. Instead, SHYSTER examines all possible variations which can be achieved by making no more than a certain number of changes to the known attribute values in the instant case; that number is specified by the user.87 As with instantiations, SHYSTER treats each hypothetical as if it were a new instant case, and determines the nearest cases and nearest results. A hypothetical is considered eligible to be reported on if its nearest result is different to that of the instant case, or if it has the same nearest result but its nearest neighbour is nearer the instant case than that of the instant case: i.e. it is a better case for the nearest result. Of these eligible hypotheticals, only the nearest are chosen to be reported on—up to a certain (user-specified) number for each result. The hypothetical reports give the user information about how the argument about the instant case can be strengthened and weakened by changing only a spe­ cified number of attributes. (If the user wishes to examine the effect of varying a specific attribute or attributes, she/he can use SHYSTER’s instantiation feature.) � � 87 � 3.12 Safeguards 3.12 Safeguards SHYSTER employs several safeguards so that it can warn the user in situations where its advice may be suspect. In certain circumstances, the results of these safeguards are logged, for the user, in a log file. Only in some of these circum­ stances is a warning issued. SHYSTER’s safeguards are based on the extra similarity measures, ideal points specified by the legal expert, centroids (described in §3.12.3 below), attribute direction, equidistance and instantiations. 3.12.1 Extra similarity measures As well as calculating known and unknown distance, SHYSTER also determines the similarity between the instant case and each of the leading cases using the extra similarity measures discussed in §3.9.3 above: viz. djk, Sjk, Sjk, rjk and rjk. If any of these measures suggests different nearest neighbours or nearest others then those differences are logged. In this way, cases which might be useful to the user’s argument, but are not mentioned in SHYSTER’s report, are brought to the user’s attention. A warning is issued if either of the extra weighted measures (the weighted association coefficient or the weighted correlation coefficient) suggests that a case with a different result to that of the nearest neighbour ought to be the nearest neighbour. This may be cause for concern about SHYSTER’s opinion as to the likely result. A different result suggested by one of the unweighted measures is not con­ sidered important enough to warn the user about, on the basis that attribute weighting is essential in the legal domain (see §3.7 above). 3.12.2 Ideal points As discussed in §3.5.4 above, the legal expert may specify ideal points. SHYSTER treats each ideal point as if it were one of the leading cases, and determines the distance between it and the instant case. It is reasonable to expect that the result in the nearest ideal point will be the same as that of the nearest neighbour; after all, the nearest ideal point represents the ideal combination of attributes for its result. If the nearest ideal point has a different result to that of the nearest neighbour, this fact is logged. 88 Chapter 3: A pragmatic approach to case law More serious would be a situation where an ideal point with a different result is at least as near to the instant case as is the nearest neighbour. If this occurs, a warning is issued. 3.12.3 Centroids For each result, SHYSTER calculates the mean of each attribute for the cases with that result and creates an average attribute value vector for that result.88 These vectors of attribute value are called centroids. SHYSTER treats each centroid as if it were one of the leading cases, and determines the distance between it and the instant case. As with ideal points, it is reasonable to expect that the result in the nearest centroid will be the same as that of the nearest neighbour. The nearest centroid represents the average combination of attributes for its result. If the nearest centroid has a different result to that of the nearest neighbour then this fact is logged because, as Tyree et al. explain, the instant case “is, in some sense, near the common boundary of the [result] groups and so must be considered as a ‘difficult’ case.”89 However, a warning is not issued unless a centroid with a different result is at least as near to the instant case as is the nearest neighbour. 3.12.4 Attribute direction As mentioned in §3.5.2 above, SHYSTER allows the legal expert to specify attrib­ ute directions. These indicate that the occurrence of a certain value for a certain attribute suggests a certain result or results. For each result, SHYSTER sums the weights of each attribute for which the value of that attribute in the instant case is directed towards that result.90 This sum is termed a direction. The larger—the stronger—a direction, the greater the extent to which the attributes in the instant case “direct” SHYSTER towards that result. There are three types of attribute direction for each result. The specified direction is calculated using the legal expert’s attribute direction in the case law specification. The ideal point direction is calculated using ideal points; if an ideal point is the only ideal point in the area with a given value for an attribute then that value for that attribute is considered to be directed towards the ideal point’s result. By an analogous method, the centroid direction is calculated using each result’s centroid. For each type of direction, the result with the strongest direction is said to be suggested by that direction. If any of these three directions suggest a result different to the nearest result, that fact is logged. Only if the specified direction 89 � 3.13 Testing and evaluation suggests a different result is a warning issued. This distinction between directions is made because the specified direction, based as it is on information provided by the legal expert, is less likely than the other directions to suggest a different result anomalously. 3.12.5 Equidistance As explained in §3.10.3 above, SHYSTER chooses between equidistant results by reference to the relative importance of the equidistant cases and the year in which those cases were decided. If equidistance has to be resolved in such a fashion, SHYSTER issues a warning; equidistant results cast doubt on SHYSTER’s choice of nearest result. 3.12.6 Instantiations As explained in §3.11.2 above, SHYSTER creates instantiations of the instant case in which all attribute values are known, and treats each as if it were a new instant case. SHYSTER reports on any of these instantiations of the instant case which have a different result to that of the uninstantiated instant case. In such an event, doubt is cast on SHYSTER’s conclusion; there is a combination of known attribute values which is consistent with the instant case but which leads SHYSTER to a different result, so a warning is issued. 3.13 Testing and evaluation SHYSTER’s general structure allows the testing of SHYSTER’s approach to case law in different areas of law. Testing, using four different case law specifications, is described in chapter 5. The choice of test domains—deciding which areas of law to specify for testing purposes—is discussed in §3.13.1 below. The principal testing method involves specifying an area of law for SHYSTER, then giving it information about a case which was actually decided in that area but which is not one of the leading cases in the specification. Unfortunately there is a paucity of such test cases, as explained in §3.13.3. Even if the specification is written so as to represent the law as it was a few years ago, the number of cases decided since then (all potential test cases) is small. So other testing methods are also employed. Generated testing (§3.13.4) uses SHYSTER’s ability to generate instantiations to create many different imaginary cases. Reflexive testing (§3.13.5) involves removing a leading case and testing it using the case base from which it was removed. (Cases used with the principal testing method are referred to as “test cases”, to distinguish them from those used in generated tests or reflexive tests.) � � � 90 Chapter 3: A pragmatic approach to case law Because all three methods of testing make use of a specification, the quality of SHYSTER’s advice is a function of both SHYSTER’s approach to case law and of the quality of the specification; even assuming that SHYSTER’s approach is a good one, its advice is only as good as its specification. Methods of evaluating the quality of SHYSTER’s advice are examined in §3.13.2. For simplicity, the tests described in chapter 5 are often referred to as tests of their specification: of course, those tests are actually testing SHYSTER— i.e. SHYSTER’s approach to case law—using that specification. (This is true for test cases and generated tests. For reasons discussed in §3.13.5 below, re­ flexive testing does not test SHYSTER’s approach although it does provide some information about the specification.) 3.13.1 Choosing a test domain Susskind suggests that (insofar as any area of the law is self-contained) an area of law chosen as a test domain should be relatively autonomous: its sources should be limited in number and reasonably well defined. It should be small enough to allow extensive coverage and its problems should not require the use of “a great deal of ‘common-sense’ knowledge”:91 not because reasoning in the law doesn’t require common sense—on the contrary—but because artificial intelligence tech­ niques cannot cope satisfactorily with common sense. Given Susskind’s consen­ sual approach to jurisprudence (discussed in §2.2.7), it is not surprising that he also suggests that there should be agreement amongst experts as to the scope and content of the test domain.92 Furthermore, he says, the domain must be one in which problem solving requires expertise. He is critical of Leith’s ELI system,93 which deals with British welfare rights—a domain which was: . . . chosen because of its simplicity which allowed Leith (not himself trained in law) to “become ‘expert’ in it.” The system was not constructed, there­ fore, with the assistance of a legal expert, there could not have been any inclusion of experts’ heuristics, and this factor might incline us to doubt whether the designation “expert system” is appropriate. Moreover, if a non-lawyer could, in a fairly short period, develop expertise in an area of law, then we might justifiably query whether that chosen area is indeed a suitable domain of application. For the chosen legal domain ought to be one whose problems do indeed require expertise (normally acquired over many years), and not relatively brief research, for their resolution.94 As Clark comments, Susskind’s criteria of suitability “would seem to narrow the scope of expert systems in law quite considerably.”95 91 � 3.13 Testing and evaluation Each of the four specifications used to test SHYSTER is quite different from the others. The Finder specification (§5.2) is a simulation of the FINDER system, and deals with a completely case-based area of law. The Authorization spe­ cification (§5.3) deals with the definition of a specific open-textured statutory concept, and is an example of an area of law in which there are more than two possible results. The Employee specification (§5.4) defines an open-textured concept which is common to several different statutes and areas of case law. Gardner justifies her choice of an area of the law on the grounds that it is “relatively well developed and stable.”96 Although that is true of the first three specifications discussed in this chapter, it is not true of the fourth. The Natural specification (§5.5) deals with a recently developed (and still developing) area of Australian administrative law. These four specifications satisfy most of Susskind’s criteria. They are all specified in terms of a limited number of well-defined cases. Each domain is fairly small, yet sufficiently complex that solving problems requires legal expertise. However, no claim is made that any of these specifications embodies a legal consensus as to what the law is. Apart from the Finder specification, each was developed by the author—a lawyer—with the assistance of an expert in the relevant field. As such, each specification represents one interpretation of its field. 3.13.2 Evaluating SHYSTER’s opinion There are a number of levels at which SHYSTER’s opinion can be evaluated. It states its prediction as to the likely outcome, on the assumption that the result will be the same as it was in the case which it deems most similar to the instant case. SHYSTER’s prediction of the likely result is considered “good” if it is the same as the result in the actual case, and “bad” otherwise.97 In their written judgments, judges will often explicitly follow, or refer favour­ ably, to certain cases before coming to their conclusion. Sometimes they will explicitly refuse to follow a case on the basis that it is distinguishable on its facts. Some reported judgments also include precis of the arguments put to the judges by counsel for the parties involved. SHYSTER’s opinion is only as good as the cases it chooses upon which to base its arguments. If a case which SHYSTER chooses is referred to by a judge, or cited in argument before the court, then it is considered a “good” case for SHYSTER to have chosen. This is true for cases upon which SHYSTER bases its arguments and its counterarguments.98 SHYSTER is proposing these cases as the best arguments for each of the possible results. For example, a judge may explicitly follow a case in coming to her/his con­ clusion. If SHYSTER chooses that same case as the basis of its counterargument then it will come to a bad conclusion, by following another case. However, it has identified a good case upon which to base an argument for one party; its conclusion is bad, but at least one of its chosen cases is good. � � � � � � 92 Chapter 3: A pragmatic approach to case law Generally, SHYSTER’s choice of a case is considered “bad” if that case was neither cited in forensic argument nor in judgment. In some circumstances, however, it is argued that SHYSTER’s choice of a case is good even though it was not used by counsel or by the court. All such divergences from the general rule are discussed in detail in chapter 5. Sometimes the case which is used as a test was decided before one of the leading cases that SHYSTER chooses. If this happens, SHYSTER’s choice is only characterized as “good” or “bad” on the advice of the legal expert. As discussed in §3.12 below, SHYSTER employs several safeguards to protect against giving bad advice. If SHYSTER issues a warning and SHYSTER’s predic­ tion as to the likely result is bad, that warning is characterized as “good”; if the predicted result is good, the absence of a warning is also considered to be “good.” If SHYSTER issues a warning even though SHYSTER’s prediction is good, the warning is “bad”; similarly if the prediction is bad and SHYSTER issues no warning, the absence of a warning is deemed “bad.” When lawyers argue with a case, it may be that only one aspect of that case is considered. This is true of the test cases, and of the leading cases that make up the specifications in appendix A. For example, Salemi v. MacKellar 99 is used as a test of the Natural specification in §5.5.3 because the High Court had to determine whether Salemi had a right to be heard before the Minister ordered his deportation. In that case, the Court also considered the issue of whether news releases were “instruments under the hand of the Minister.” But that issue is ignored in the discussion of Salemi v. MacKellar in chapter 5. It is not relevant to the specification that the case is used to test, and is not taken into account when evaluating SHYSTER’s opinion. 3.13.3 The paucity of test cases One of the major obstacles to developing1 and testing a case-based legal expert system like SHYSTER is the paucity of reported cases. This can be attributed to the filtering effect of various stages of the legal system. Consider a client who seeks legal advice as to her/his legal options in some matter. Good legal advice will filter out a hopeless case.2 If, for example, a lawyer recognizes the matter as being identical to a previously decided case which suggests that the client will lose, her/his advice will probably be not to proceed. So, for any given leading case in a SHYSTER specification, there may be many substantially identical cases which proceeded no further than a lawyer’s office. Even if the client’s legal case is a strong one, there are many reasons why it might never reach court. Most people find the cost of legal redress prohibitively expensive. Further, the time delay involved may well dissuade a person from � � � � � � 93 � 3.13 Testing and evaluation taking legal action, or defending an action. For any number of non-legal reasons, prospective parties to a case may choose not to proceed, and/or to “settle out of court.” Finally, not all cases that are decided in court are reported. Only those cases which the court reporters deem significant are included in the law reports. The legal domain is quite different from other areas of case-based expert system development where the developer may have access to a large number of cases which are relatively unfiltered. In the face of this lack of test cases, Gardner used examination problems to test her system on the basis that, although they may not be as complex as real cases, they are reasonably difficult and the sorts of questions that lawyers are expected to be able to handle.3 This is certainly true—and SHYSTER is tested with some hypothetical cases in this way—but the author contends that actual cases make better tests. Although the legal system greatly restricts the number of cases that are reported, it ensures that those cases that are reported make good tests. If a case had been straightforward, it is highly unlikely that it would have proceeded through the various filters that the legal system provides to emerge as a reported judgment. The four specifications described in chapter 5 are tested with a total of seven­ teen previously decided cases which are not amongst the specified leading cases. Although a small number, it compares well with the testing of other case-based systems. Tyree, Greenleaf and Mowbray tested FINDER with one case.4 Ashley evaluated HYPO’s performance using only four cases—all of which were taken from the case base.5 In addition to the real cases used to test SHYSTER, there are four hypothetical test cases. These were proposed by the legal experts as interesting tests that have not yet been before the courts. 3.13.4 Generated tests Generated tests are performed using SHYSTER’s capacity to argue with instanti­ ations (see §3.11.2 above). SHYSTER can be used to generate, and give advice on, all possible cases in an area. However, there is no point in generating the entire search space in this fashion. In an area of reasonable size, the search space would be so large as to make evaluation of SHYSTER’s advice in each case impractical.6 Instead, the legal expert is asked to specify a number of attribute values whose presence in a case means that a certain result, or results, follows (i.e. that result, or results, would be reached if that case were to be heard by a court). A fact vector is constructed containing these known attribute values, with the remaining attribute values unknown, and SHYSTER is made to generate all instantiations of that fact vector. � � � 94 Chapter 3: A pragmatic approach to case law Although it is not possible to evaluate SHYSTER’s choice of cases in all of these instantiations—these generated cases—it is to be hoped that SHYSTER will choose a good result (i.e. one of the results specified by the legal expert) in most, if not all, of them. The number of good choices of result as a fraction of the number of generated cases is SHYSTER’s success rate for a generated test. Some of these generated cases may represent paradoxes: combinations of attribute values which are impossible. Paradoxical generated cases are ignored when determining SHYSTER’s success rate.7 SHYSTER’s ideal point warnings8 are of particular importance to the generated tests. Many of the generated cases may be extremely unusual: i.e. combinations of attribute values that are very unlikely to occur, though not impossible. Assuming that its leading cases are well chosen, SHYSTER is more likely to choose a bad nearest case when given an extremely unlikely and unusual instant case than if it were given a realistic combination of attribute values as its instant case. Ideal point warnings are designed to detect these extremes, and are taken into account when evaluating the results of generated tests. 3.13.5 Reflexive tests A reflexive test is performed in the same fashion as a test performed using the principal testing method, except that one of the leading cases in the current specification is used as the test case. Of course, if SHYSTER is presented with a fact vector which is identical to one of the leading cases in its specification it will simply follow that leading case: the two cases are identical. Such a test would demonstrate nothing. So, before a reflexive test is performed, the leading case that is to be used as a test case is removed from the specification. SHYSTER is effectively asked “if this case were decided now, in the light of all the leading cases except itself, how would it have been decided?” (The term “reflexive” is coined to describe such a test because the case is being applied to the specification whence it came.) Mackaay and Robillard criticize Lawlor for adopting a reflexive approach to testing. They point out that: . . . when all cases are used to determine optimal [attribute] weights . . . there can be no surprise to find that on the basis of those weights each case is correctly classified; the results would look unduly promising.9 This criticism does not apply to reflexive testing of SHYSTER because SHYSTER recalculates attribute weights (without the removed leading case) for each of the reflexive tests. � � � 95 � 3.13 Testing and evaluation The specification that is used for a reflexive test is diminished: it excludes the test case. It must be assumed that the case that has been removed belongs in the specification, otherwise the legal expert would not have included it. Hence the diminished specification no longer represents the area of law that it was written to represent. This means that the result of a reflexive test does not assist in evaluating SHYSTER’s approach to case law: SHYSTER’s opinion and the cases it chooses cannot sensibly be compared with the judgment in the actual case. However, the results of reflexive tests do provide information about the specification itself. It is very unlikely that a specification of reasonable size would have a leading case for every possible combination of attribute values. Yet it is highly desirable that a specification be capable of handling new cases which are not identical to any of the leading cases. Consider a specification in which conducting a reflexive test for every leading case in that specification yields a good result. All of the leading cases contribute to the extent that each case’s result can be determined from the other cases. Such a set of results would indicate that the specification is successful at handling new combinations of attributes. Conversely, consider a specification in which every reflexive test yields a bad result. That specification is clearly very poor at handling new fact situations. In reality, a given specification is unlikely to reach good conclusions in all of its reflexive tests, or in none of them. The number of good conclusions will probably lie somewhere in between. If a large proportion of a specification’s reflexive tests yield a good result, that is indicative of that specification’s suitability to handle new cases. It is important to note the following points about reflexive testing. Even if SHYSTER’s opinion in a reflexive test is very good—i.e. it comes to the same conclusion as did the court in that case, and for the same reasons—that is no reason to remove that case from the specification permanently. It does not follow that, because the diminished specification is all SHYSTER needs to reach a good conclusion, the excluded case adds nothing to the specification. A new case could occur in which the best opinion would be to follow the excluded case, and not the cases that SHYSTER chooses in the reflexive test. This will almost certainly be true for a new case with the same attributes as the excluded case, but the attribute values need not be identical. The diminished specification may reach a good conclusion, but not necessarily for good reasons.10 Removing a case from a specification can introduce attribute dependencies which were not there before. It is also quite likely that one of the cases which SHYSTER chooses upon which to base its opinion was decided after the instant case. If that happens, it is impossible to evaluate SHYSTER’s choice of cases. � � � 96 Chapter 3: A pragmatic approach to case law The weight given to the decisions of a given court may vary between jurisdic­ tions. For example, decisions of the English House of Lords are binding on lower courts in England, but only strongly persuasive in Australian courts. Three of the four specifications discussed in chapter 5 use Australian and English cases,11 and each specification aims to represent the Australian law. If an English case is used for a reflexive test, SHYSTER may choose a case which was not referred to in the test case but which might have been a good case to choose had the test case been heard in an Australian court. Reflexive tests are not tests of SHYSTER’s approach to case law, though these tests do provide information about a particular specification. A reflexive test is performed for every leading case in each of the four specifications discussed in chapter 5. The results of these tests are set out and discussed in appendix D. 3.14 Conclusion SHYSTER treats cases as points in space, the dimensionality of which is the num­ ber of attributes. The instant case is placed at its appropriate point in this space, and the nearest leading cases are determined. These nearest cases are used to produce an argument (based on similarities and differences between the cases) about the likely outcome in the instant case. That argument relies on the doc­ trine of precedent; it assumes that the instant case will be decided the same way as was the nearest case. SHYSTER can also instantiate unknown attributes, thus testing all possible configurations of the instant case. A limited number (specified by the user) of hypothetical variations of the instant case can also be tested, to see whether the case can be strengthened toward some result. SHYSTER also applies several safeguards and the user is warned if SHYSTER has some doubt about the veracity of its advice. The report that SHYSTER generates makes a prediction and justifies that prediction by reference only to cases and their similarities and differences: the numbers which SHYSTER uses in coming to its opinion do not appear in that opinion. SHYSTER models legal knowledge at the second of the three levels that Green- leaf, Mowbray and Tyree identify (as quoted in §2.7): its representation includes justification based on the primary legal sources, but without any explicit model of those sources; principles of interpretation are implied in the representation.12 � � � � � � 97 � 3.14 Conclusion By taking a pragmatic approach to representing case law, SHYSTER follows one of the approaches that Gardner recommends taking “if legal realism is right”: emphasizing the behaviouristic side of legal realism.13 Gardner herself takes a different approach,14 and says of the behaviourists’ programs: These are not AI programs . . . The programs are concerned with predict­ ing judicial decisions, or more generally with analyzing judicial behavior, from a data base in which legal rules have no role. Traditional modes of reasoning are replaced by mathematical methods . . . 15 Gardner does not explain why she thinks that these “prediction programs” are not AI programs. If a program can successfully predict judicial decisions—even without adopting a “traditional mode of reasoning”—it is at least arguable that it is artificially intelligent. Areas are used in SHYSTER to represent open-textured concepts: statutory or case-based. This structure facilities the linking of a rule-based system with SHYSTER’s case-based system to form a hybrid system, capable of dealing with statutes and case law.16 A hybrid approach has a number of benefits. It has the advantage of approx­ imating the approach which a lawyer would take when given a legal problem. The rules, derived from a statute, are applied until the meaning of some open- textured concept is required. Faced with this problem, a lawyer would turn to the common law in order to further clarify the meaning of the statute. So, too, does SHYSTER: the lawyer’s two-stage approach is clearly modelled. A consequence of SHYSTER’s approach is that it returns a result from each area of case law, for use in another area, or by a statutory rule base. It is certainly true, as Ashley points out, that “the goal of a theory of analogical legal argument should not be to explain what the right answer is.”17 But in returning a result from an area, SHYSTER does not pretend to be giving the “right answer”; it is merely attempting to predict the “likely answer,” applying the principle of stare decisis. SHYSTER’s opinion is predictive, not normative, and (as explained in §2.3.5) a good legal expert system should have predictive power. Gardner cites various writers who identify the following as difficulties associated with this approach to case law: separating findings of fact from legal conclusions; determining what the judges in the leading cases believed the facts to be; cat­ egorizing facts appropriately; and deciding what aspects of the facts should be included.18 But these difficulties are not peculiar to legal expert systems which 98 Chapter 3: A pragmatic approach to case law take a statistical approach to case law: all of these difficulties are inherent in the problem of dealing with case law. SHYSTER expects the legal expert to have regard to these difficulties when specifying an area of case law. This is not an unreasonable expectation: knowing how to overcome these difficulties is one of the characteristics of legal expertise. 3.14.1 Comparisons with other approaches SHYSTER adopts and expands on the nearest neighbour approach to case law taken by FINDER. Mackaay and Robillard apply nearest neighbour techniques to cases, too, although they did not actually develop an expert system. SHYSTER also adopts aspects of HYPO’s approach to reasoning with hypotheticals. One difference between SHYSTER and both FINDER and HYPO is SHYSTER’s generality; while FINDER deals only with the law of trover and HYPO deals with trade secrets law, SHYSTER allows a legal expert to define an arbitrary number of areas of case law in each specification. SHYSTER’s knowledge representation is more complex than FINDER’s, but simpler than HYPO’s.19 Unlike SHYSTER, FINDER does not allow the specification of unknown attribute values: unknown values are entered as nos. Tyree suggests that an alternative approach, “possibly better,” would be to select a value at random;20 SHYSTER’s use of unknowns is better still. SHYSTER allows any number of possible results in an area. Both HYPO and FINDER treat all cases (including the instant case) as having one of two possible results. Ashley claims that: . . . subject to some qualification, there are only two possibilities: either the plaintiff won the case or did not. The qualification is that depending on the procedural context of the case, some outcomes are more determinative than others.21 This qualification relates not to the number of possibilities but to the importance of the leading cases (which is captured by SHYSTER using its hierarchy of courts). Although there may be only two possible results in the areas of law covered by FINDER and HYPO, this is not true in other areas.22 Because of its approach to weighting attributes, SHYSTER assumes attrib­ ute independence, and checks for attribute dependence in all areas. SHYSTER uses the same method to weight its attributes as does FINDER; hence, FINDER also assumes attribute dependence, although it does not address the problem. Fortunately, for FINDER, there is neither functional dependence, nor evidence of stochastic dependence, between its attributes.23 SHYSTER, like FINDER but un­ like HYPO, assumes that each attribute has the same weight in all cases (see §3.5.3 above). � � � 99 � 3.14 Conclusion SHYSTER’s distance metric is similar to FINDER’s, except that it takes account of unknown distance and infinitely weighted attributes.24 FINDER does not use extra similarity measures, specified directions, equidistance or instantiations as safeguards, as does SHYSTER. FINDER does use centroids as a safeguard. The use of ideal points as a safeguard is suggested by Tyree et al. but is not implemented in FINDER. 25 Apart from a few cosmetic additions, the structure of SHYSTER’s reports is similar to those of FINDER. However, SHYSTER’s algorithm for choosing which cases to use in argument is more complicated because of the need to account for unknown distance. Unlike FINDER, SHYSTER handles equidistant cases and argues with instantiations and hypotheticals. Two of SHYSTER’s features—specified direction and hypotheticals—are de­ signed to incorporate some of HYPO’s functionality, without the need for HYPO’s more complex knowledge representation. HYPO represents the attributes that favour each side, and treats a problem as a collection of possibly competing attributes. All attributes are required to favour one side or the other.26 Similarly, SHYSTER’s specified direction allows the legal expert to specify a result favoured by a certain value for a certain attribute. However, SHYSTER does not require that all attributes (or attribute values) be directed, for the simple reason that not all attribute values can be directed towards a result or results.27 SHYSTER’s reasoning with hypotheticals is not as sophisticated as HYPO’s, due to SHYSTER’s simpler representation of case law. However, SHYSTER is able to examine limited hypothetical variations, and inform the user if the effect of these variations is to strengthen or weaken its argument about the instant case. 3.14.2 A Sisyphean journey? Now that SHYSTER’s approach to case law has been explained, it is possible to mount a detailed refutation (foreshadowed in §2.4.7) of Berman’s argument that case-based reasoning techniques are inappropriate for modelling case law. Berman is a proponent of deep rule-based models. He believes that “lawyers make their decisions on the basis of their judgment as to whether a particular rule will be applied to the facts of a specific case.”28 It is argued elsewhere that developers of conceptual models of legal reasoning confuse precision with accuracy.29 The refutation that follows deals only with Berman’s arguments against case-based reasoning, with particular reference to SHYSTER. Berman claims that the choice for the developer in the legal domain is a “Sisyph­ ean journey” with case-based reasoning or “down hill with rules”. 30 He refers to work in case-based legal reasoning employing “frame-based structures, transition 100 Chapter 3: A pragmatic approach to case law nets, semantic networks, discrimination trees, connectionist models, etc.,31 and claims that: To legal scholars well versed in the subtleties of legal reasoning these par­ ticular representations of legal cases, though seminal works of considerable scientific importance, constitute a mere simulacrum of legal thought. First, the models do not contain choice of law rules to account for those legal cases that implicate the law of more than one jurisdiction . . . This could be achieved using a hybrid system. “Choice of law” rules could be implemented in the rule base (or a meta-rule base) to ensure that the appropriate areas of case law were applied. Alternatively, assuming that the user is an expert, the user could make this choice her/himself (as discussed in §3.4 above, such an assumption is reasonable). Second, the models do not account for the fact that some precedents are weakened by divided courts . . . SHYSTER allows the legal expert to specify a hierarchy of courts. Several courts in this hierarchy could be different compositions of the same court (e.g. one, three, five or seven judges of the High Court of Australia). The legal expert could even include a minority judgment as a separate case in the specification if it was deemed sufficiently important. SHYSTER only uses its hierarchy for resolving equidistance, but there is no reason why a case-based system could not make more use of such a hierarchy in its operation. Third, the models do not take into consideration that judicial opinions carry varying precedential values . . . SHYSTER takes this into consideration in two ways: by use of a hierarchy, and by allowing a legal expert to choose the important cases and, by implication, to reject the less important cases. Fourth, the models do not account for the fact that the precedential weight may turn on when the case was decided . . . As with the relative importance of a case, SHYSTER takes the year in which a case was decided into account when resolving equidistance—there is no reason why a case-based reasoning system could not give the time since a case was decided more importance than does SHYSTER. Further, the legal expert should be expected not to include cases which are so old as to be of little precedential weight. Fifth, the models do not account for sub silentio overruling—the disregard of precedents which have been so often distinguished or ignored that they lack precedential value. SHYSTER assumes that all of the cases chosen by the legal expert are “good law.” Disregarded precedents should be disregarded. � � � 101 � 3.14 Conclusion Sixth, except for the works of McCarty . . . the models lack mechanisms for resolving tensions between conflicting lines of authority. But conflicting lines of authority are unresolvable. The appropriate response to a conflicting line of authority is to present all the arguments to assist the user in constructing her/his argument. Seventh, the model does not account for judicial decisions motivated by political considerations unarticulated in opinions . . . Political considerations, if the legal expert deems them important, can be included as attributes in one of SHYSTER’s case law specifications.32 Eighth, these models have not provided for the computational representa­ tion of legal fictions where the concepts of contracts, easements and notice become spurious easements, quasi contracts, and constructive notice. Why these examples could not be represented in a case-based system is not clear. Ninth, these models do not consider that the precedential value of a case may turn on the prestige of the judge who wrote the opinion. As discussed above, under Berman’s second point, this presents no difficulty for SHYSTER. Tenth, and most importantly, the model does not represent accurately the procedural posture of a case so that the resulting arguments fail to distinguish cases in which courts have ruled on matters of law from cases where appellate courts have merely affirmed findings of fact . . . 33 If a case involves merely an affirmation of a finding of fact, and no ruling on a matter of law, then the legal expert is unlikely to choose it as a leading case. Berman confuses the role of a case-based system with the role of a legal expert in the knowledge acquisition process. With the possible exceptions of Bench- Capon et al. (§2.4.2), no-one would deny that legal expertise is required in the development of a legal expert system. Hence, it is reasonable to use the fact of legal expert specification in SHYSTER to refute several of Berman’s points. Berman concedes that rules “fall far short of fully representing legal know­ ledge”, but: For developers, as contrasted to researchers, the issue is not whether the resulting base is “complete” or even “accurate” or “self-modifying”—but whether the resulting rule base is sufficiently complete and accurate to be “useful”.34 The author agrees that usefulness should be the principal criterion in legal expert system design, but argues that, in a case law domain at least, a case-based system can better satisfy that criterion. 4 Implementing SHYSTER I know you Lawyers can, with ease, Twist words and meanings as you please; That language, by your skill made pliant, Will bend to favour ev’ry client; That ’tis the fee directs the sense To make out either side’s pretense. When you peruse the clearest case, You see it with a double face; For scepticism’s your profession; You hold there’s doubt in all expression. Hence is the bar with fees supply’d, Hence eloquence takes either side . . . John Gay (1732) The Dog and the Fox 35 He is no lawyer who cannot take two sides. Charles Lamb (1833)36 Thelma Todd: “I didn’t know you were a lawyer. You’re awfully shy for a lawyer.” Groucho Marx: “You bet I’m shy. I’m a shyster lawyer.” Monkey Business (1931)37 103 104 Chapter 4: Implementing SHYSTER 4.1 Introduction SHYSTER’s approach to case law is described in the previous chapter. In this chapter, the implementation of that approach is described, and illustrated using examples. SHYSTER is implemented using a dozen modules, written in ISO C. 38 (The code for each module is listed in full elsewhere.39) The structure of this chapter mirrors that of SHYSTER, with the description of the implementation divided into descriptions of each module. The Shyster module (§4.2) is the top-level module for the whole system. The Statutes module (§4.3) is the top-level module for a rule-based system, presently unimplemented. The Cases module (§4.4) is the top-level module for the case-based system. The Tokenizer and Parser modules (§4.5 and §4.6) tokenize and parse a program written in SHYSTER’s case law specification lan­ guage. The Dumper module (§4.7) displays the information that has been parsed. The Checker module (§4.8) checks for evidence of dependence between the at­ tributes. The Scales module (§4.9) determines the weight of each attribute. The Adjuster module (§4.10) allows the legal expert to adjust the weights of the attributes. The Consultant module (§4.11) interrogates the user as to the attribute values in the instant case. The Odometer module (§4.12) determines the distances between the leading cases and the instant case, and the Reporter module (§4.13) writes SHYSTER’s legal opinion. Except where otherwise indicated, the examples used in this chapter are taken from the Employee specification, which is explained in §5.4 and appears in §A.4. The Employee specification is used as the basis of the complete example in appendix C. 4.1.1 Internal representation SHYSTER uses records to represent entities (courts, areas, results, attributes, cases, etc.) and makes multiply linked lists of these entities to reflect the rela­ tionships between them. For example, each area record has a pointer to the head of a linked list of result records. Each result record has a pointer to the head of a linked list of case records. Each case record has a pointer to the head of a linked list of attribute value records. Each attribute value record is an element in two lists, linked by case and by attribute, forming part of a matrix of attribute values. Using linked lists allows the manipulation of these entities and relationships without requiring the imposition of any limits upon their numbers. The only limits that affect the user of SHYSTER or the writer of specifications are on the maximum length of a filename, and the maximum length of an identifier.40 The information contained in each record, and the links between those records, are built up by each of SHYSTER’s modules until, by the time that the Reporter module is invoked, the structure is complete. 105 � 4.1 Introduction 4.1.2 Output files Each time SHYSTER’s case-based system is invoked it writes the following files: • a log file (introduced in §3.12) which summarizes SHYSTER’s operation and includes any warnings that have been issued; • a dump file which is a dump of SHYSTER’s internal representation of the case law specification; • a probabilities file which gives probability figures for each attribute pair in each area; • a weights file which gives details of the weights that SHYSTER gives to each attribute in each area; a distances file for each invoked area which includes distances and other • similarity measures for the instant case, instantiations and hypotheticals; and • a report file for each invoked area which is SHYSTER’s legal opinion—its argument about the likely result in that area for the instant case. Only the first and the last of these are intended for the user. SHYSTER’s report files are completely self-contained; the user need have no knowledge of SHYSTER and its operation in order to understand its reports. The log file should be com­ prehensible to anyone with only passing acquaintance with SHYSTER’s operation. A user who finds the information contained in a log file arcane may safely ignore the file: of its contents, only the warnings are crucial, and these are also written to the standard error stream.41 All other files are intermediate files that SHYSTER produces on its way to its report. They provide details about SHYSTER’s internal workings for the inform­ ation of the knowledge engineer and the legal expert. All of these output files are plain text files. With the exception of the log file, they are all in LaTEX format: i.e. they are suitable for processing by the LaTEX document processor.42 This contributes to SHYSTER’s portability, as LaTEX is widely available on many platforms. Using LaTEX simplifies the footnoting of text, allows some data to be displayed in a clear and economical tabular format, and ensures the aesthetic quality of the output. All of the examples of SHYSTER output in this thesis—the dump files in appendix A, the example reports in appendix B, the “LaTEX output” files in the complete example in appendix C, and extracts in fourteen of the figures in this chapter and the next43—appear exactly as produced by SHYSTER after processing by LaTEX. 44 106 Chapter 4: Implementing SHYSTER -a Enable weight adjustment (see §4.10) -c specification Read the case law specification from “specification.cls” -d distances Write each distances file to “distances-area.tex” -D dump Write the dump file to “dump.tex” -e Enable echo mode (see §4.11) -h r c Hypothesize, reporting on r hypotheticals per result with a limit of c changes -i Write LaTEX code that can be included in another LaTEX document (i.e. not stand-alone code) -l log Write the log file to “log.log” -p probabilities Write the probabilities file to “probabilities.tex” -q Enable quiet mode (don’t summarize cases, etc.) -r report Write each report file to “report-area.tex” -w weights Write the weights file to “weights.tex” Figure 4.1: The UNIX command line switches and arguments recognized by SHYSTER. Apart from the -c switch, all switches are optional. 4.2 The SHYSTER module The Shyster module is the top-level module for SHYSTER. It extracts the options and arguments from the UNIX command line, initializes the rule-based system and the case-based system, then invokes the rule-based system. Figure 4.1 lists the command line switches and arguments that SHYSTER recognizes. Only the -c switch must be used; all other switches are optional. So, for example, if the -d switch is not used, no dump file is written. However, if the -l switch is not used then the information that would have been written to the log file is written to the standard output stream.45 4.3 The STATUTES module The Statutes module is the top-level module for a rule-based system. This module provides a skeletal structure within which a rule-based system could be developed, and linked with the case-based system to form a hybrid system. The module has two functions. The Initialize Statutes function presently returns a pointer to a dummy struc­ ture. If implemented, it would initialize the rule-based system, by reading a stat­ ute law specification, and return a pointer to SHYSTER’s internal representation of that specification. � 4.4 The CASES module 107 The Statute Law function would invoke the rule-based system proper. Each statutory open-textured concept in the rule base would be associated with an identifier corresponding to an area in the case base. The rule-based system would invoke the case-based system using that identifier. The case-based system returns an identifier corresponding to the result from the appropriate area. That result would be bound in the rule-base to a value for the statutory open-textured con­ cept. At present, the Statute Law function prompts the user for an identifier, then invokes the case-based system seeking advice in the area corresponding to that identifier. The result returned by the case-based system is written to the log file. 4.4 The CASES module The Cases module is the top-level module for the case-based system. Its two major functions correspond to the two functions in the Statutes module. The Initialize Cases function calls the Tokenizer and Parser to read the case law specification and build an internal representation of that specification. The Dumper is invoked to dump that internal representation to the dump file. The Checker is used to check for attribute dependence. Finally, the Scales module is called to assign weights to all the attributes. The function returns a pointer to SHYSTER’s internal representation of the specification with all attrib­ utes weighted. The Case Law function takes a pointer to that internal representation, and an identifier corresponding to one of the areas in that specification. If weight adjustment is enabled, it calls the Adjuster. It then invokes the Consultant to interrogate the user as to the attribute values in the instant case. (It is the Consultant which recursively invokes the Case Law function, if required, to resolve open textured—external—attributes.) Having ascertained the attribute values, the Case Law function calls the Odometer to calculate the distances between the instant case and the lead­ ing cases, and to determine the nearest cases and results. Then the Reporter is invoked to write a report about the instant case. The Cases module instantiates any unknown variables in the instant case and invokes the Odometer and the Reporter to recalculate the distances and argue using the instantiation. If the user has requested it, the Cases module also makes hypothetical variations to the instant case and invokes the Odometer and the Reporter yet again to recalculate distances and argue with the hypothetical. Output from these invocations of the Odometer and the Reporter is written to the distances file and report file for the current area. 108 Chapter 4: Implementing SHYSTER AREA CLOSING HELP OPENING SUMMARY ATTRIBUTE COURT HIERARCHY QUESTION UNKNOWN CASE EXTERNAL IDEAL RESULT YEAR CITATION FACTS NO RESULTS YES Figure 4.2: The keywords in SHYSTER’s case law specification language. 4.5 The TOKENIZER module The Tokenizer module reads the case law specification file, and breaks it into tokens. The Tokenizer is invoked repeatedly by the Parser; each time it is invoked, it returns the next token in the specification. 4.5.1 Tokens There are seven different valid types of token. An identifier is any sequence of alphabetic characters, numeric characters, and the - character (starting with an alphabetic character). If an identifier is more than 16 characters long it is truncated to its first 16 characters and a warning is issued. An identifier’s case is significant. A keyword is an identifier that has a special meaning to SHYSTER. There are twenty keywords, and they are listed in figure 4.2. Keywords are reserved: i.e. they cannot be used as identifiers. A string is a sequence of characters, enclosed between a pair of " characters.46 There must be at least one character between the two quotation marks, but SHYSTER imposes no upper limit on the length of a string. SHYSTER converts a pair of consecutive " characters within a string into a single " character.47 A year is a positive integer of up to four digits. An attribute vector is a sequence of Y, N and U characters within parentheses.48 The remaining two tokens are the = character and a token which indicates that the end of the specification file has been reached. 4.5.2 Comments and whitespace When the Tokenizer reads a % character in the specification (except in a string), it skips over the rest of that line: any characters between the % and the next carriage return are ignored. This allows the legal expert to put comments in the specification file. Whitespace49 is required between adjacent identifiers/keywords, and between adjacent strings. The Tokenizer treats each occurrence of whitespace in the specification as a single space. Hence, extra whitespace can be freely added � 4.6 The PARSER module 109 between tokens making the specification easier to read without changing the way that it is tokenized. This also applies inside strings; if a string is too long to fit on a single line it can be split over several lines. 4.6 The PARSER module The Parser module parses50 the case law specification using the tokens provided by the Tokenizer. A formal definition of the syntax of the specification language is given, in Extended Backus-Naur Form (EBNF),51 in figure 4.3. The specification of case law using this language is best illustrated by example. 4.6.1 Hierarchy A specification starts with an (optional) hierarchy. This binds a court identifier to a string which describes that court. For example: HIERARCHY HC-5 "five judges of the High Court of Australia" HC-4 "four judges of the High Court of Australia" HC-3 "three judges of the High Court of Australia" HC "a single judge of the High Court of Australia" FCA-3 "three judges of the Federal Court of Australia" PC "the Judicial Committee of the Privy Council" CA "the English Court of Appeal" KB "the King’s Bench Division of the English High Court" = QB "the Queen’s Bench Division of the English High Court" The courts are listed in descending order of seniority; the earlier in the list that a court appears, the better the authority of its cases. An = character separates courts of equivalent rank. It is not necessary to specify every court in the juris­ diction, because courts are ranked relatively: the above example tells SHYSTER that FCA-3 is more important than PC—but not how much more important. Each court identifier in the hierarchy must be unique. Note that there are three different sorts of identifier: court identifiers, area identifiers and result identifiers. SHYSTER knows which sort of identifier to expect in various places in the specification. Hence, the same identifier can be used (for example) to refer to a result and to a court in the same area. Court identifiers apply in all areas in the specification. The scope of a result identifier is the area in which it is specified: i.e. the same identifier can be used for two different results in two different areas.52 110 Chapter 4: Implementing SHYSTER specification = [ hierarchy ] area { area }. hierarchy = hierarchy-header hierarchy-block. hierarchy-header = "HIERARCHY". hierarchy-block = court-identifier string { [ "=" ] court-identifier string }. court-identifier = identifier. area = area-header area-block. area-header = "AREA" area-identifier. area-block = [ opening ] [ closing ] results attribute { attribute } case { case } { ideal-point }. area-identifier = identifier. opening = "OPENING" string. closing = "CLOSING" string. results = results-header results-block. results-header = "RESULTS". results-block = result-identifier string result-identifier string { result-identifier string }. result-identifier = identifier. attribute = attribute-header attribute-block. attribute-header = "ATTRIBUTE". attribute-block = local-attribute external-attribute. | local-attribute = "QUESTION" string [ "YES" string { result-identifier } ] [ "NO" string { result-identifier } ] [ "UNKNOWN" string { result-identifier } ] [ "HELP" string ]. external-attribute = "AREA" area-identifier [ "YES" string { result-identifier } [ "EXTERNAL" result-identifier { result-identifier } ] ] [ "NO" string { result-identifier } [ "EXTERNAL" result-identifier { result-identifier } ] ] [ "UNKNOWN" string { result-identifier } [ "EXTERNAL" result-identifier { result-identifier } ] ] Figure 4.3: A formal definition, in Extended Backus-Naur Form (EBNF), of the syntax of SHYSTER’s case law specification language. continued next page 111 � 4.6 The PARSER module case = case-header case-block. case-header = "CASE" string [ string ]. case-block = "CITATION" string "YEAR" year [ "COURT" court-identifier ] "FACTS" attribute-vector "RESULT" result-identifier [ "SUMMARY" string ]. ideal-point = ideal-point-header ideal-point-block. ideal-point-header = "IDEAL". ideal-point-block = "FACTS" attribute-vector "RESULT" result-identifier. attribute-vector = "(" attribute-value { attribute-value } ")". attribute-value = "Y" "N" "U".| | string = """" character { character } """". identifier = letter { letter | digit | "-" }. year = digit [ digit ] [ digit ] [ digit ]. Figure 4.3 (continued). 4.6.2 Areas A specification can contain any number of areas. An area commences with an area header: AREA Employee which binds an identifier to the area that is about to be specified: the identifier with which the case-based system is invoked in order to use this area to define an open-textured concept. An opening string may be specified. This should be a brief introduction appropriate to any opinion given in this area. It is written at the beginning of SHYSTER’s report for this area. Similarly a closing string may be specified, containing concluding remarks. At least two results must be specified; this involves binding identifiers to strings. For example: RESULTS Employee "the worker is an employee" Contractor "the worker is an independent contractor" Each string is a statement which holds when the relevant result occurs. It must be cast so that is makes sense when prefixed with the following: “If case is followed then . . .” 112 Chapter 4: Implementing SHYSTER 4.6.3 Attributes Any number of attributes may be specified. Each attribute must be either local or external. Local attributes For a local attribute, the first keyword after ATTRIBUTE is QUESTION. For ex­ ample:53 ATTRIBUTE QUESTION "Did the employer pay the worker by time" YES "the employer paid the worker by time" Employee NO "the employer did not pay the worker by time" Contractor UNKNOWN "it is not known whether the employer paid the worker by time" HELP "The employer could pay the worker by time (e.g. by the hour, or by the week) or by results." The QUESTION string is the question that the user will be asked if the value of the attribute needs to be determined. The YES, NO and UNKNOWN strings are optional, but the user will only be allowed to answer the question with a value for which there is a string: e.g. if no UNKNOWN string is defined, the user will not be allowed to answer unknown. The HELP string is displayed at the user’s request, and should provide further information to assist her/him in answering the question. The result identifiers after the strings specify attribute direction: A value of yes for this attribute is directed towards the Employee result; a value of no is directed towards Contractor. External attributes For an external attribute, the first keyword after ATTRIBUTE is AREA. This ex­ ample is taken from another specification:54 ATTRIBUTE AREA Expectation YES "the applicant had a legitimate expectation which was affected by the decision" Affected EXTERNAL Expectation NO "the applicant did not have a legitimate expectation which was affected by the decision" EXTERNAL No-Expectation UNKNOWN "it is not known whether the applicant had a legitimate expectation which was affected by the decision" � 4.6 The PARSER module 113 The value of this attribute is to be resolved by reference to the Expectation area.55 The identifiers after the EXTERNAL keywords are external result identifiers: result identifiers in the external area. These results are associated with values for this attribute. If SHYSTER returns a result of Expectation for the Expectation area, the value of this attribute is set to yes; if it returns a result of No-Expectation, the value is set to no. 56 If it returns any other identifier, SHYSTER exits with an 57error. A value of yes for this attribute is directed towards the Affected result in this area. 4.6.4 Leading cases Any number of leading cases may be specified as follows:58 CASE "Queensland Stations Pty Ltd Commissioner of Taxation" v. Federal "Queensland Stations v. FCT" CITATION YEAR "(1945) 70 CLR 539" 1945 COURT HC-3 FACTS RESULT (NYNYNNYYYNNNNNNNYN) Contractor SUMMARY "agreements were entered into between Queensland Stations and some drovers. The agreements stated that the drovers would . . ." The first CASE string is the full name of the case; the second is a shorter version of the name, and is optional. The CITATION string is the case citation, and the YEAR is the year of the decision—not necessarily the same as the year in the citation.59 The (optional) COURT identifier links the case to the court in which it was decided. The FACTS attribute vector has one attribute value for each attribute defined in the area. The order of these values corresponds to the order in which the attributes have been defined. The Y, N and U characters represent the values yes, no and unknown, respectively. The SUMMARY string is the legal expert’s summary of the case. The string may contain many sentences, but must be cast so as to make sense when prefixed with the following: “In case, a decision of court, . . .” This string may (but need not) include LaTEX commands which will be processed when the case is listed in the dump file, and if the case is used in a report file: e.g. \footnote. A warning is issued if a result has been specified with no cases, or (worse still) with neither cases nor an ideal point.60 And if two cases in an area have the same attribute values, or the same attribute values except for unknown values, a warning is issued. If these two cases have different results, this is also mentioned 114 Chapter 4: Implementing SHYSTER in the warning. If two or more leading cases have identical attribute values but different results then one of those cases was wrongly decided or the area needs another attribute to distinguish between those two cases.61 4.6.5 Ideal points Ideal points are specified as follows: IDEAL FACTS (YNYNYYYNNYYYNYYYYU) RESULT Employee This represents the ideal combination of attribute values for the specified result. 4.7 The DUMPER module The Dumper module writes the dump file: a formatted version of the case law specification that has just been read. This file is easier to read (once processed by LaTEX) than its corresponding specification, simplifying the development and amendment of a specification. It also reflects SHYSTER’s internal representation of the specification. Each of the four specifications given in appendix A is a complete dump file. The Dumper begins the dump file by displaying the hierarchy of cases in a tabular format. The courts are numbered; courts of the same rank share the same number. For each area in the specification, the Dumper starts by displaying the cases and their attribute values in a matrix format. An example is given in figure 4.4. The attribute values are represented by a symbol for yes, a × symbol for no,• and a blank space for unknown. 62 The attributes are named A1 . . . A18 in the order of their appearance in the specification. The cases have been grouped by their result, and named C1 . . . C14 in the order of their importance within their group. (The more important of two cases is considered to be the one decided by the more important court in the hierarchy or, if their courts are equally important, the case that is more recent.) The ideal points are also represented: IEmployee and IContractor. The rank of the court in which each case was decided is indicated in the column labelled “c”; these numbers correspond to those in the hierarchy display. This matrix allows the user easily to compare the attribute values in all the specified cases and ideal points. The Dumper writes out the opening and closing strings (if they were spe­ cified). It then writes the result identifiers and their strings. 115 � 4.7 The DUMPER module A tt ri bu te s C as e c R es ul t A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 C 1 × • • × • • • × × × • • × × • × × × 1 C 2 • • • × • × × × • × × × × × × × × × 2 C 3 • × • × • • • × × • × • × × × × × 4 C 4 • × • • × • • × × × × × × × × × × 5 E m p lo ye e C 5 • × • × • × • × × × • • × × × × × • 7 C 6 × • • × • × × × × • × 7 C 7 • • • • × • • × × • • • × • × × × × 8 I E m p lo y e e • × • × • • • × × • • • × • • • • C 8 × • × • × × × × • × × × × × × × × × 3 C 9 × • × • × × • • • × × × × × × × • × 3 C 1 0 × • • • × × × × • × × × × × • × × × 5 C 1 1 C o n tr a ct o r × • • • × × × • • • × × × × × × × • 6 C 1 2 • • • × • • × • • × • • × × × • × • 7 C 1 3 × • × × • × × × × • • × 7 C 1 4 × • • • × × × × • • × × × × × × × • 8 I C o n tr a c to r × • × • × × • • × × × • × × × • F ig u re 4 .4 : T h e at tr ib u te v al u e m at ri x fo r th e E m p lo ye e ar ea ( S H Y S T E R o u tp u t) . 116 Chapter 4: Implementing SHYSTER 4.7.1 Attributes The Dumper writes details on all attributes: local and external. Local attributes Local attributes are displayed in the following format:63 A12: Did the employer pay the worker by time? yes: the employer paid the worker by time. ⊃ Employee no: the employer did not pay the worker by time. ⊃ Contractor unknown: it is not known whether the employer paid the worker by time. The employer could pay the worker by time (e.g. by the hour, or by the week) or by results. Specified direction is indicated using a ⊃ symbol: e.g. a value of yes is directed towards the Employee result.64 The last paragraph is the help string. External attributes External attributes are displayed in the following format:65 A4: ∀ Expectation area yes: the applicant had a legitimate expectation which was affected by the decision. � Expectation ⊃ Affected no: the applicant did not have a legitimate expectation which was affected by the decision. � No-Expectation unknown: it is not known whether the applicant had a legitimate expectation which was affected by the decision. The link between this attribute and the Expectation area is indicated using a ∀ symbol. The association of results from the external area with values of this attribute is indicated using a � symbol: e.g. if the Expectation area returns a result of Expectation the value of this attribute is set to yes. As with local attributes, specified direction is indicated using a ⊃ symbol. � 4.8 The CHECKER module 117 4.7.2 Leading cases Leading cases appear grouped under the appropriate result heading: e.g. “Cases in which the worker is an independent contractor.” They are displayed in the following format: C9: Queensland Stations Pty Ltd v. Federal Commissioner of Taxation (1945) 70 CLR 539 (“Queensland Stations v. FCT ”) A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. . . . A18: the employer and the worker did not express any intention that the relationship would be one of principal and independent con­ tractor. In Queensland Stations Pty Ltd v. Federal Commissioner of Taxa­ tion, 66 a 1945 decision of three judges of the High Court of Australia, agreements were entered into between Queensland Stations and some drovers. The agreements stated that the drovers would . . . The short case name appears in parentheses after the full citation. The attribute values are described in full. The summary is formatted as it would be if the case were used in a report, complete with footnotes—in the above excerpt, an endnote. 4.7.3 Ideal points Ideal points are displayed using a similar format to that used for leading cases (§4.7.2). Details of the ideal point’s attribute values are given under an appro­ priate result heading. For example: IEmployee (the ideal case in which the worker is an employee): 4.8 The CHECKER module The Checker module examines every pair of attributes in each area for functional dependence and evidence of stochastic dependence. � �� � 118 Chapter 4: Implementing SHYSTER 4.8.1 Detecting dependence Detecting functional dependence is straightforward. Detecting evidence of stoch­ astic dependence (as explained in §3.8.2) involves calculating the probability P (n) of there being exactly n yes/yes pairs, assuming random data: y N−y P (n) = n � x�−n N x where N is the number of known pairs in two attributes AX and AY , x is the number of yess in AX , and y is the number of yess in AY . For every pair of attributes, the Checker counts the actual number of yes/yes pairs and calculates the probability of there being that number of yes/yes pairs or fewer, and the probability of there being that number of yes/yes pairs or more. The Checker does not need to build a complete probability table, like that in figure 3.3, in order to calculate these two probabilities for each attribute pair. ⎫iThe probability of i yes/yes pairs or fewer is P (n ∗ i) = n=0 P (n) and the probability of i yes/yes pairs or greater is P (n ≤ i) = 1 − ⎫i−1 P (n). So, P (n)n=0 needs only to be calculated for n = 0 . . . i. The Checker does not use the formula for P (n) directly when calculating probabilities. Because n ≤ max(0, x + y − N ), the first non-zero probability is either P (0) or P (x + y − N ). By substitution into the formula for P (n), P (0) = (N − x)! (N − y)! , x + y ∗ N, N ! (N − x − y)! and x! y! P (x + y − N ) = N ! (x + y − N )! , x + y ≤ N. The Checker starts with the first non-zero probability and successively multiplies it by P (i + 1)/P (i): by substitution, P (i + 1) = (i − x)(i − y) . P (i) (i + 1)(N − x − y + i + 1) This is a more efficient method of calculating each probability than applying the formula for P (n) directly. P (i + 1)/P (i) becomes zero when i = x or i = y, so all values of P (i) for i > min(x, y) are zero. This limit on i is never exceeded because n ∗ min(x, y) and, as explained above, the Checker only calculates P (n) for n = 0 . . . i. 119 � 4.8 The CHECKER module A2 A3 0.05 1.00 1.00 0.15 0.45 1.00 A4 0.30 0.95 0.90 0.50 0.50 0.90 A5 0.95 0.30 0.50 0.90 0.90 0.50 0.00 1.00 A6 1.00 0.06 0.27 0.97 1.00 0.23 0.50 0.87 0.87 0.50 A7 0.96 0.28 0.09 1.00 0.77 0.77 0.50 0.88 0.88 0.50 0.99 0.12 A8 0.69 0.80 1.00 0.58 0.58 0.89 0.88 0.56 0.56 0.88 0.80 0.71 0.58 0.88 A9 0.41 0.91 1.00 0.19 0.56 0.88 1.00 0.08 0.08 1.00 0.22 0.98 0.02• 1.00 1.00 0.16 A10 0.62 0.85 1.00 0.45 1.00 0.45 1.00 0.10 0.10 1.00 0.73 0.77 0.50 0.91 0.89 0.58 0.88 0.56 · · · A1· · · A2· · · A3· · · A4· · · A5· · · A6· · · A7· · · A8· · · A9· · · . . . .. . Figure 4.5: An extract from the probabilities matrix for the Employee area (SHYSTER output). The Employee area has 18 attributes; only 45 of the 153 cells in the complete matrix are shown here. 4.8.2 Writing matrices of probabilities Probability figures are written, in matrix form, to SHYSTER’s probabilities file. Each cell in this matrix has two probabilities:67 1.00 probability of the actual number of yes/yes pairs or fewer; 0.08 probability of the actual number of yes/yes pairs or more. If a probability figure is not above the threshold of 0.05 (chosen in §3.8.2) it is marked with a symbol and a message is written to the log file to warn • the legal expert that there is evidence of stochastic dependence. If there is an equivalence function or an inverse function mapping one attribute to the other, the first probability is marked with a symbol and a warning message is written. An extract from the probabilities matrix for this example is given in fig­ ure 4.5. (The complete probabilities file is in §C.7.) There is functional depend­ ence between attributes A4 and A5. There is evidence of stochastic dependence between attributes A7 and A9. 120 Chapter 4: Implementing SHYSTER 4.9 The SCALES module The Scales module assigns weights to attributes and writes, to the weights file, a table of weights for each area in the specification. 4.9.1 Calculating weights For each attribute, the Scales module calculates a weight and some result weights (one for each result). The weight (as defined in §3.7) is the inverse of the variance68 and it ranges from 4 to infinity.69 A warning is issued if an attribute has infinite weight. A result weight is the inverse of the variance of the attribute values for that result. Result weights are used to calculate the strength of attribute directions which represent the extent to which the attribute values of the instant case suggest a given result (see §3.12.4). 4.9.2 Writing tables of weights The Scales module writes to the weights file a table of weights for each area in the specification. The weights file for the Employee specification is given in §C.9. Each table has columns for the mean µ, variance �2 and weight w for each result and for the attribute as a whole. The mean column for each result is that result’s centroid (see §3.12.3). 4.10 The ADJUSTER module The Adjuster module allows any of the attribute weights (including result weights) in the current area to be set to any desired value. This feature is intended for use by the legal expert—not the user, who need not be aware of the manner in which SHYSTER comes to its conclusions. Using the Adjuster, the legal expert can test the effect of changing weights during the development of a specification. The Adjuster is not invoked unless weight adjustment is enabled using the -a switch on the command line. If a weight is changed, an adjusted weights file is written for the current area. 4.11 The CONSULTANT module The Consultant module determines the attribute values for the instant case in the current area. If an attribute is external, the Consultant invokes the Cases module again and assigns the attribute a value on the basis of the result identifier that is returned. If an attribute is local, the user is interrogated as to the attribute values. 121 � 4.12 The ODOMETER module distance djk = �jk measure: association �jk Sjk = coefficient: n ⎫ � �� �n ¯ ¯Aij − Aj Aik − Ak correlation rjk = ⎪ i=1 n � �2 n � �2 coefficient: ⎫ ⎫¯ ¯Aij − Aj Aik − Ak i=1 i=1 weighted n�⎬ ⎬ distance d�jk = ⎬⎬Aij − Aik⎬⎬ × wi measure: i=1 n⎫⎬⎬ ⎬⎬ weighted ⎬Aij − Aik⎬ × wi association Sjk � = i=1 n ⎫ coefficient: wi i=1 ⎫ � �� �n weighted Aij × wi − Āj� Aik × wi − Āk� � = i=1correlation rjk ⎪ n � �2 n � �2 coefficient: ⎫ Aij × wi − Āj� ⎫ Aik × wi − Ā�k i=1 i=1 Figure 4.6: Measures of similarity between a case j and a case k (discussed in §3.9.1 and §3.9.2). �jk is the number of differences in the corresponding attribute values of the two cases; n is the number of attributes; Aij is the value of the ith attribute for the jth case; Ā j is the mean of all attribute values for the jth case; wi is the weight of the ith attribute; and Ā�j is the weighted mean of all attribute values for the jth case. At present, the Consultant uses a simple scrolling prompted dialogue. If the user has enabled echo mode (using the -e switch on the command line) then the Consultant echoes the user’s input by writing the appropriate attribute string to the terminal. This ensures that the user understands the meaning of the attribute value that she/he has entered. SHYSTER’s modular design is such that any desired interface could be employed by changing just this module. 4.12 The ODOMETER module The Odometer module performs all of SHYSTER’s distance calculations, and writes tables of distances to a distances file. 122 Chapter 4: Implementing SHYSTER 4.12.1 Calculating distances As explained in §3.9, for SHYSTER each of the commonly used similarity measures reduces to one of six measures, and a variant of the weighted distance metric d�jk was chosen. All six similarity measures are summarized in figure 4.6 (see previous page). The Odometer calculates the distances (known and unknown) between the instant case and the leading cases, the ideal points, and the centroids. It also cal­ culates values for the extra similarity measures, and the strength of the attribute directions. Using the known and unknown distances it determines which are the nearest cases and the nearest result. If necessary, it resolves equidistant results by reference to the rank of the courts involved in the nearest cases, and the recentness of those cases. A warning is issued if equidistance has to be resolved in this fashion. SHYSTER uses a precision threshold of two decimal places in all of its compar­ isons. This allows for the possibility of rounding errors having been introduced in SHYSTER’s arithmetic, and recognizes the danger of relying too much on pre­ cise quantification of abstract notions. In addition there is a second precision threshold for use in distance comparisons.70 This is the threshold within which two cases will be considered equidistant. The Odometer is invoked for the instant case, and for each instantiation and each hypothetical.71 It treats each instantiation or hypothetical as if it were the instant case for the purposes of performing its calculations. 4.12.2 Writing tables of distances The Odometer writes, to the distances file, a table of distances for the instant case, each instantiation, and each of the chosen hypotheticals.72 A distances file for the Employee area is given in full in §C.11; the instant case is Building Workers’ Industrial Union of Australia v. Odco Pty Ltd (discussed in detail in §5.4.3). An extract from that file—the table of distances for the uninstantiated and unhypothesized instant case—is given in figure 4.7. Tables of distances extracted from other distances files can be found in figures 5.3, 5.6, 5.7 and 5.11. The table in figure 4.7 includes an attribute value matrix, similar to the one in figure 4.4, with the addition of the instant case CInstant and a centroid for each result: µEmployee and µContractor. As before, the rank of the court that decided each case is in the column labelled “c”. The known and unknown distances are in columns labelled “dK” and “dU”. Unweighted distance measures are in the column labelled “�”—not “d”, to avoid confusion with the known and unknown distances. Values for the other similarity measures (S, S�, r and r�) are labelled appropriately.73 � • • • � � � 123 C a se C In s t a n t C 1 C 2 C 3 C 4 A tt r ib u te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 × • × • × × × × × • × × × × × • × • • × • • • × × × • • × × • × × × 1 • • • × • × × × • × × × × × × × × × 2 • × • × • • • × × • × • × × × × × 4 • × • • × • • × × × × × × × × × × 5 d K d U 3 4 .3 9 1 0 .1 3 3 0 .9 5 1 0 .1 3 5 0 .3 1 1 0 .1 3 3 3 .2 2 1 4 .1 6 7 7 9 7 S S � 0 .4 4 0 .3 3 0 .4 4 0 .2 9 0 .5 6 0 .4 8 0 .4 7 0 .3 3 r r � R e su lt I 0 .0 7 0 .1 1 − 0. 0 8 0 .0 9 − 0. 2 2 − 0. 2 6 E m p lo y e e − 0. 1 1 − 0. 1 8 � 1 9 .3 0 • × • × • × • × × × × × × × × C 5 7 2 7 .9 6 1 0 .1 3 6 0 .3 8 0 .2 7 0 .1 5 0 .1 1 1 9 .3 0 µ 1 2 .1 0 C 6 × • • × • × × × × • × 7 1 8 .8 4 6 8 .2 9 4 0 .4 0 0 .4 0 0 .1 7 0 .2 3 C 7 I Em p lo ye e µ Em p lo ye e C 8 C 9 C 1 0 C 1 1 • • • • × • • × × • • • × • × × × × 8 • × • × • • • × × • • • × • • • • • • • × • • • × × × • • × × × × × × × • × • × × × × • × × × × × × × × × 3 × • × • × × • • • × × × × × × × • × 3 × • • • × × × × • × × × × × • × × × 5 × • • • × × × • • • × × × × × × × • 6 4 2 .3 1 1 0 .1 3 7 9 .3 2 1 5 .0 3 3 1 .2 8 1 0 .1 3 1 2 .9 2 1 0 .1 3 2 4 .6 1 1 0 .1 3 2 6 .0 6 1 0 .1 3 1 9 .9 0 1 0 .1 3 7 1 2 7 3 5 5 4 0 .4 4 0 .4 0 0 .8 0 0 .7 9 0 .4 4 0 .3 0 0 .1 9 0 .1 2 0 .3 1 0 .2 3 0 .3 1 0 .2 5 0 .2 5 0 .1 9 I 0 .2 2 0 .0 6 − 0. 4 5 − 0. 4 5 0 .0 6 0 .0 2 0 .4 6 0 .5 6 0 .2 3 0 .2 5 0 .2 3 0 .2 2 C o n tr a c to r 0 .4 5 0 .4 6 + 5 5 .9 0 � 2 � � 4.12 The ODOMETER module C 1 2 7 3 9 .4 9 1 0 .1 3 7 0 .4 4 0 .3 7 0 .2 2 0 .1 4 • • • × • • × • • × • • × × × • × • + 5 2 .6 3 µ � 2 � C 1 4 8 1 9 .9 0 1 0 .1 3 4 0 .2 5 0 .1 9 0 .4 5 0 .4 6 × • • • × × × × • • × × × × × × × • I Co n tr a ct o r × • × • × × • • × × × • × × × • 2 1 .1 1 2 2 .1 7 3 0 .2 1 0 .2 3 0 .5 2 0 .2 9 µ Co n tr a ct o r 1 8 .8 6 1 0 .1 3 4 0 .2 5 0 .1 8 0 .5 3 0 .5 5 × • • • × × × × • × × × × × × × × × F ig u re 4 .7 : T h e ta b le o f d is ta n ce s fo r B ui ld in g W or ke rs ’ In du st ri al U n io n o f A us tr al ia v . O dc o P ty L td ( S H Y S T E R ou tp u t) . T h e n ea re st n ei gh b ou r is C 8 , so t h e n ea re st r es u lt i s C o n tr a ct o r; t h e n ea re st o th er i s C 5 . S H Y S T E R ’s o p in io n o n B W IU v . O dc o is d is cu ss ed i n § 5. 4. 3. C 1 3 × • × × • × × × × • • × 7 2 0 .5 8 6 0 .6 1 4 0 .3 6 0 .3 8 0 .2 1 0 .2 0 3 3 .3 3 � 124 Chapter 4: Implementing SHYSTER The strength of each non-zero attribute direction is displayed in the “Result” A ⊃ symbol indicates specified direction; a ⊃ symbol indicates ideal point direction; a ⊃ symbol indicates centroid direction. (If all three directions I µ column. are of the same strength, a ⊃ symbol and a single number appear in their place.) Each number represents the strength of the direction towards that result in the instant case. 4.13 The REPORTER module The principal output from SHYSTER is contained in its report file, written by the Reporter module. This module also implements SHYSTER’s safeguards and writes information about them, together with a general summary of its report, to the log file. In each area, the Reporter is invoked once for the instant case, and once for each instantiation and for each chosen hypothetical.74 Only if an instantiation has a different result to that of the instant case does the Reporter write a report on that instantiation. A report file for the Employee area is given in full in §C.13. Again, the instant case is BWIU v. Odco. Except where otherwise indicated, the examples in this section are taken from that report file. Other example report files are given in appendix B. 4.13.1 Arguing with the instant case The Reporter argues that the result in the instant case will be the nearest result, then builds a counterargument for each of the nearest other results. This same process is also used to argue with instantiations and with hypotheticals (see §4.13.2 and §4.13.3 below). Arguing for the nearest result For the instant case, the Reporter starts by writing the opening string to the report file. It then states the facts of the instant case and declares its opinion as to the likely result—the nearest result: In my opinion—following Humberstone v. Northern Timber Mills—the worker is an independent contractor. Humberstone v. NTM (C8 in figure 4.7) is the nearest (unknown) neighbour. There are no nearest known cases; all of the leading cases have some unknown distance because of the two unknown attribute values in the instant case. 125 � 4.13 The REPORTER module The Reporter then summarizes the nearest neighbour:75 In Humberstone v. Northern Timber Mills, 76 a 1949 decision of three judges of the High Court of Australia, Humberstone carried goods for NTM. He had originally held himself out as a carrier, . . . If the case is identical to the instant case—if it has the same attribute values— then the Reporter announces that the two cases are “on all fours.”77 If there is no known distance between the cases but there is some unknown distance, it declares that the two cases “may be on all fours” and explains its reservation on the basis of the unknown attributes. In this example there are some known differences. First, the Reporter lists all of the similarities: There are several significant similarities between the instant case and Hum­ berstone v. NTM : the employer did not direct the manner in which the work was to be done; the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; . . . Three or more similarities are characterized as “significant”; two similarities are “very significant”; a solitary similarity is called “extremely significant,” as it must be for such a dissimilar case to be the nearest neighbour. Then the Reporter lists all of the differences: However, the instant case is not on all fours with Humberstone v. NTM. In that case the worker was not in business on her/his own account; the worker was allowed to employ others to assist with her/his work; the worker was not required to work at specified times; the employer did not pay the worker by time; and the employer and the worker did not express any intention that the relationship would be one of principal and independent contractor. And if there are any unknowns in the leading case—there are none in this example—then they are explained too. Despite the differences, SHYSTER stands by its original statement: Nevertheless, I believe that Humberstone v. NTM should be followed. If there are equidistant nearest neighbours they are all handled in this fashion, with a linking paragraph written between each pair of cases. This paragraph explains which case is the more important and why, as in this example taken from the report on the second hypothetical for BWIU v. Odco: In 1967, Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance78 was decided by the Queen’s Bench Division of the English High Court. (A case decided by the Queen’s Bench Division of the English High Court is not as good authority as a case decided by the Judicial Committee of the Privy Council—like AMP v. Chaplin; further­ more Ready Mixed v. Minister is 11 years older than AMP v. Chaplin.) 126 Chapter 4: Implementing SHYSTER Arguing for other results For every other result, the Reporter discusses the nearest other. It explains the effect of following the nearest other instead of the nearest neighbour, then summarizes the nearest other: If Ferguson v. John Dawson & Partners (Contractors) Ltd is followed then the worker is an employee. In Ferguson v. John Dawson & Partners (Contractors) Ltd, 79 a 1976 decision of the English Court of Appeal, Ferguson fell off a roof while re­ moving some scaffolding boards. He claimed damages against John Dawson (the building contractors) for breach of statutory duty relying on . . . Once again all of the similarities are listed, but this time the differences are decisive: However, there are several significant differences between the instant case and Ferguson v. Dawson. In that case the employer directed the manner in which the work was to be done; the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was an integral part of the employer’s business; . . . (Like the similarities in the nearest neighbours, the differences in the nearest others are characterized as “significant,” “very significant” or “extremely signi­ ficant,” depending on their number.) The Reporter also compares the importance of this case with that of the nearest neighbour,80 before reiterating its opinion: Note also that Ferguson v. Dawson is only a decision of the English Court of Appeal and not as good authority as a case decided by three judges of the High Court of Australia—like Humberstone v. NTM. Consequently, there is nothing in Ferguson v. Dawson to warrant any change in my conclusion. The Reporter is not swayed from its conclusion on the basis of the importance of the nearest other. For example, in the report on the first hypothetical for BWIU v. Odco it writes: Despite the fact that Cam v. Sargent is a decision of four judges of the High Court of Australia (and better authority than a case decided by three judges of the High Court of Australia—like Humberstone v. NTM ), there is nothing in Cam v. Sargent to warrant any change in my conclusion. Equidistant nearest others are handled in a similar fashion to that used for equidistant nearest neighbours. The report on the instant case concludes with the closing string. 127 � 4.13 The REPORTER module Variations Although the Reporter relies heavily on the strings supplied by the legal expert in the specification, each report is more than a mere regurgitation of these strings. The structure of each report varies depending on the circumstances. For example, where a case would have been the nearest case if not for its unknown distance, the Reporter makes this clear—without any mention of unknown distance. This example is taken from another report file:81 I would have suggested that Stevenson v. Macdonald (1) be followed (in­ stead of Massey v. Crown Life) except that it is not known whether the employer supervised or inspected the work; it is not known whether the employer paid the worker by time; it is not known whether the money that the employer paid to the worker was stated to be a “fee”; . . . Different situations are handled in different ways at several stages in the process of preparing the report. For this reason, there are many different possible reports, even ignoring the difference in case names, summaries, and other strings. This means that SHYSTER’s reports read quite well. Although their style is a little stilted, it is just possible that they could be mistaken for the work of a lawyer— although that is not one of the aims of this thesis project. 4.13.2 Arguing with instantiations For the purposes of writing reports on instantiations, the Reporter treats each instantiation as if it were the instant case and follows the steps outlined in §4.13.1 above.82 The only differences are in the introductory comments, and in the fact that the instant case is referred to as the “instantiated case.” For example, a report on an instantiation from another test case commences as follows:83 It may be that the following is true of the instant case: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; and the employer neither supervised nor inspected the work. If that is so then in my opinion—following Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance—the worker is an independent contractor. 4.13.3 Arguing with hypotheticals As with instantiations, to write a report on a hypothetical the Reporter treats that hypothetical as if it were the instant case and follows the steps outlined in §4.13.1 above.84 Again, the introductory comments are different, and the instant case is referred to as the “hypothetical case.” 128 Chapter 4: Implementing SHYSTER The report on the first hypothetical in the report file for BWIU v. Odco starts with these words: Consider the instant case changed so that the following is true: the worker was allowed to employ others to assist with her/his work; and the employer and the worker did not express any intention that the relationship would be one of principal and independent contractor. If that were so then I would be more strongly of the opinion that— following Humberstone v. Northern Timber Mills—the worker is an inde­ pendent contractor. If the hypothetical variations lead SHYSTER to a different conclusion, as in the second hypothetical in the example report file, the words are different again: Consider the instant case changed so that the following is true: the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; and the worker was an integral part of the employer’s business. If that were so then my opinion would be that—following Ferguson v. John Dawson & Partners (Contractors) Ltd —the worker is an employee. 4.13.4 Safeguards As well as constructing its reports, the Reporter implements the safeguards explained in §3.12. It determines which would be the nearest neighbours applying, in turn, each of the extra similarity measures: djk, Sjk, Sjk � , rjk and rjk � . If these measures suggest different nearest neighbours from those chosen by SHYSTER using its known and unknown distance, that fact is noted in the log file. Figure 4.8 is extracted from the log file for this example (the complete log file is in §C.2). The fact vector for the instant case is given, then the short case names of the nearest neighbour and the nearest other. All of the extra similarity measures suggest the same nearest neighbour; there are no safeguard warnings. The nearest result is declared to be Contractor (this is the identifier that is returned by this invocation of the case-based system). The fact vector of the first instantiation is given, and the differences between it and the instant case are marked with carets. In this instantiation, the nearest neighbour and nearest other are the same as in the uninstantiated instant case. However, some of the extra similarity measures now disagree with the choice of nearest neighbour. These disagreements are set out under the “Safeguards” heading.85 Extra cases which the extra measures suggest as the nearest neigh­ bour are marked with a +; if an extra case has a different result to that of the instantiation, it is marked with a * and its result appears in parentheses after the case name. If an extra measure does not suggest one of the nearest neighbours it is marked with a -. � � 129 � 4.13 The REPORTER module Fact vector is (NYNYNNNUNNYUNNNNNY). Nearest neighbours: Contractor: C8 Humberstone v. NTM Nearest others: Employee: C5 Ferguson v. Dawson Nearest result for the instant case is Contractor. Instantiation 1 is (NYNYNNNYNNYYNNNNNY). Nearest neighbours: Contractor: C8 Humberstone v. NTM Nearest others: Employee: C5 Ferguson v. Dawson Safeguards: Distance measures: * C6 Stevenson v. Macdonald (2) (Employee) + C11 AMP v. Chaplin + C13 Stevenson v. Macdonald (1) Association coefficients: + C11 AMP v. Chaplin Correlation coefficients: - C8 Humberstone v. NTM + C11 AMP v. Chaplin Weighted correlation coefficients: - C8 Humberstone v. NTM + C11 AMP v. Chaplin Nearest result for instantiation 1 is Contractor. Figure 4.8: An extract from the log file for Building Workers’ Industrial Union of Australia v. Odco Pty Ltd (SHYSTER output). All of these differences are logged, but only some of them will cause a warning to be issued. If one or both of the weighted safeguard metrics suggest that a case with a different result should be the nearest neighbour then a warning is issued. In this example, only the (unweighted) distance measures suggest a different result, so no warning is issued. 130 Chapter 4: Implementing SHYSTER The Reporter also writes to the log file if the ideal points, centroids or any of the three attribute directions (specified, ideal point and centroid directions) suggest a different result. However, a warning is not issued unless an ideal point or a centroid with a different result is at least as near to the instant case (or instantiation or hypothetical) as is the nearest neighbour, or if the specified dir­ ections suggest a different result. Warnings are written to the log file and to the standard error stream. In summary, SHYSTER will issue a warning in each of the following circum­ stances: • the weighted association coefficients suggest that a case, with a different result than that of the nearest neighbour, ought to be the nearest neighbour; • the weighted correlation coefficients suggest that a case, with a different result than that of the nearest neighbour, ought to be the nearest neighbour; • an ideal point suggesting a different result is at least as near to the instant case as is the nearest neighbour; • a centroid suggesting a different result is at least as near to the instant case as is the nearest neighbour; or • the specified directions suggest a different result or results. As explained in §3.12.5, SHYSTER will also issue a warning if: • there are two or more equidistant results. And, as explained in §3.12.6, a warning is issued if: an instantiation of the instant case has a different result to that of the • uninstantiated instant case. 4.14 Conclusion SHYSTER is a working implementation of the pragmatic approach to case law adopted, and justified, in the previous chapter. It has been designed to allow a legal expert to specify areas of case law using a simple specification language. SHYSTER’s case-based system has been constructed so as to facilitate its linking with a rule-based system to form a hybrid legal expert system. Within the limitations imposed by its internal representation of case law, SHYSTER is capable of producing sophisticated reports on the likely outcome in user-specified instant cases. In the process of constructing its report, SHYSTER produces several intermediate files containing information of use to the knowledge engineer and the legal expert. The next chapter discusses the specification of areas of law and the use of those specifications, and various testing methods, to evaluate SHYSTER’s output—to evaluate SHYSTER’s approach to case law. 5 Case studies . . . the notion that a computer can predict the course of judicial decision rests on assumptions that are demonstrably untenable, does violence to the very nature of law, and is moreover certain to blunt the professional techniques of any lawyer who relies on machines rather than on his own powers of reasoning and advocacy. Frederick Bernays Wiener (1962) Decision Prediction by Computers: Nonsense Cubed—and Worse86 G. B. Trudeau (1975) The Doonesbury Chronicles87 131 132 Chapter 5: Case studies 5.1 Introduction SHYSTER is unusual amongst legal expert systems in that it has not been de­ veloped specifically for a given area of the law. Instead, its case-based system has been designed to allow the specification of different areas of law. In order to test SHYSTER, and its approach to case law, four specifications were written. Each specification aims to represent an area of Australian law. The complete specifications are given in appendix A, and discussed in de­ tail in §5.2–§5.5 below. For each specification, the law is briefly stated, and the specification and its testing are described. The methods used to test each specification, and SHYSTER, are explained in §3.13. The facts of cases—cases which are not part of the specification—were given to SHYSTER, and SHYSTER’s opinions were compared with the judgments in those cases. If a specification includes ideal points, those ideal points were also used as test cases. Each of these tests is discussed, individually, below. The results of these tests are summarized in figure 5.12 near the end of this chapter, and conclusions are drawn from them in §5.6.1. Generated tests (explained in §3.13.4) were performed for each specification for which legal expertise was available: i.e. all but the FINDER specification. Generated testing is also discussed below, and summarized in figure 5.13 at the end of this chapter. Conclusions are drawn from the results of these tests in §5.6.2. The results of reflexive testing (explained in §3.13.5) of every leading case in each of the four specifications are set out and discussed in appendix D, and summarized in the four figures in that appendix. Conclusions are drawn from these results in §5.6.3. Conclusions are drawn from all of this testing in the next chapter. 5.2 A simulation of FINDER FINDER (§2.5.2) is a case-based legal expert system. Because SHYSTER’s ap­ proach to case law expands upon that of FINDER, a specification can be written which causes SHYSTER to simulate FINDER. The simulation does not make use of all of SHYSTER’s features (as explained in §5.2.2 below). However, SHYSTER’s safeguard mechanisms, and the opportun­ ity to perform reflexive tests, mean that FINDER’s approach to case law can be tested more extensively using SHYSTER than was possible using FINDER itself. 5.2.1 The law The law of trover—the law concerning the rights of the finders of lost chattels— is unusual in that it is based entirely on cases. The 1772 case of Armory v. Delamirie88 established that the finder of any article which has been lost has a general right to that article as against all the world except the true owner.89 Only � 5.2 A simulation of FINDER 133 a few reported trover cases have been decided in the 270 years since. Although they are few, these cases have qualified considerably the general rule stated in Armory v. Delamirie, and the law of trover has become complicated. In 1948, Lord Goddard CJ of the King’s Bench Division of the English High Court referred to several trover cases (Bridges v. Hawkesworth, 90 Elwes v. Brigg Gas Co., 91 South Staffordshire Water Co. v. Sharman 92 and Hannah v. Peel 93) and said: These cases, or rather the first three of them, have long been the delight of professors and text writers, whose task it often is to attempt to reconcile the irreconcilable. It is, however, right to say that in recent years both the Corpus Professor of Jurisprudence at Oxford and the Professor Emeritus of English Law at Cambridge have expressed the opinion that Bridges v. Hawkesworth was wrongly decided. If it was, the difficulty largely dis­ appears. But that much-battered case has lately been re-invigorated by Birkett J’s decision in Hannah v. Peel, and I am glad to think that . . . it is still for wiser heads than mine to end a controversy which will no doubt continue to form an appropriate subject for moots till the House of Lords lays it to rest for all time.94 Although wiser heads than Lord Goddard’s have not yet laid it to rest for all time, Donaldson LJ made what has been called “an unusually clear statement of the principles to be applied”95 in the 1982 case of Parker v. British Airways. 96 He listed the following rights and obligations of the finder: 1. The finder of a chattel acquires no rights over it unless (a) it has been abandoned or lost and (b) he takes it into his care and control. 2. The finder of a chattel acquires very limited rights over it if he takes it into his care and control with dishonest intent or in the course of trespassing. 3. Subject to the foregoing and to point 4 below, a finder of a chattel, whilst not acquiring any absolute property or ownership in the chattel, acquires a right to keep it against all but the true owner or those in a position to claim through the true owner or one who can assert a prior right to keep the chattel which was subsisting at the time when the finder took the chattel into his care and control. 4. Unless otherwise agreed, any servant or agent who finds a chattel in the course of his employment or agency and not wholly incidentally or collaterally thereto and who takes it into his care and control does so on behalf of his employer or principal who acquires a finder’s rights to the exclusion of those of the actual finder. 5. A person having a finder’s rights has an obligation to take such measures as in all the circumstances are reasonable to acquaint the true owner of the finding and the present whereabouts of the chattel and to care for it meanwhile.97 134 Chapter 5: Case studies His lordship then listed the following rights and obligations of the occupier of the premises where the chattel was found: 1. An occupier of land has rights superior to those of a finder over chattels in or attached to that land and an occupier of a building has similar rights in respect of chattels attached to that building, whether in either case the occupier is aware of the presence of the chattel. 2. An occupier of a building has rights superior to those of a finder over chattels on or in, but not attached to, that building if, but only if, before the chattel is found, he has manifested an intention to exercise control over the building and the things which may be on or in it. 3. An occupier who manifests an intention to exercise control over a building and the things which may be on or in it so as to acquire rights superior to those of a finder is under an obligation to take such measures as in all the circumstances are reasonable to ensure that lost chattels are found and, on their being found, whether by him or by a third party, to acquaint the true owner of the finding and to care for the chattels mean­ while. The manifestation of intention may be express or implied from the circumstances including, in particular, the circumstance that the occupier manifestly accepts or is obliged by law to accept liability for chattels lost on his ‘premises’ e.g. an innkeeper’s or carrier’s liability. 4. An ‘occupier’ of a chattel, e.g. a ship, motor car, caravan or aircraft, is to be treated as if he were the occupier of a building for the purposes of the foregoing rules.98 It is clear that the question of who owns a found chattel remains, in Lord Goddard’s words, “a really difficult area of law”. 99 5.2.2 The FINDER specification The dump file for the Finder specification is given in full in §A.2. The specifica­ tion contains a single area: the Finder area. This area does not make use of all of SHYSTER’s features—it has no opening string, no closing string, no help strings, no unknown attribute values, no ideal points and no attribute direction—because FINDER does not have these features. Only two features of this specification are not part of FINDER. First there is the inclusion of a hierarchy of courts. This is not strictly necessary; SHYSTER does not require that a hierarchy be specified. The effect of this addition is merely cosmetic: it allows SHYSTER to mention the court in which a case was heard when it discusses that case. There are only three courts in this hierarchy representing two of the three divisions of the English High Court: the Chancery division, and the King’s/Queen’s Bench division. All three courts are of equal importance; the inclusion of the hierarchy does not affect which cases SHYSTER chooses to justify its opinions. 135 � 5.2 A simulation of FINDER The second feature of this specification that is not part of FINDER is the inclusion of unknown strings in each attribute. This allows the user to answer “unknown” to any of the attribute questions, and force SHYSTER to examine instantiations of the instant case. Results The Finder area has two results: Win and Lose. These correspond to FINDER’s two possible outcomes: “the finder wins” and “the finder loses.” The area—like FINDER—actually represents only part of the law of trover: the resolution of conflict between the finder of a chattel and another person who is not the true owner of the chattel. Attributes There are ten attributes, whose questions are taken (slightly reworded) from the questions asked by FINDER: A1: Was the finder the occupier of the premises where the chattel was found? A2: Was the chattel attached to the land or premises where it was found? A3: Was the other claimant (the non-finder) the owner of the premises where the chattel was found? A4: Was the other claimant the true owner of the chattel or did she/he claim through the rights of the true owner? A5: Did the finder hand over the chattel to the other claimant after the finding? A6: Did one of the parties rely on the terms of an agreement made with the other which purported to give her/him the right to the chattel? A7: Was the finder a servant of the other claimant? A8: Was the chattel hidden or in a position so as to be difficult to find? A9: Was an attempt made to find the true owner of the chattel or, altern­ atively, was the chattel clearly abandoned? A10: Did either of the parties know of the existence of the chattel prior to the finding? Answering these questions should be fairly straightforward. Only the use of the word “servant” in A7 could cause difficulty. Lawyers use the word to refer to an employee as opposed to an independent contractor—a distinction which is the basis of the Employee specification described in §5.4.2 below. Tyree suggests that 136 Chapter 5: Case studies “servant” can be read as “employee”. 1 However, in one (and perhaps both) of the FINDER cases in which the value of A7 is yes, the finder was an independent contractor of the other claimant.2 Clearly “servant” has a broad meaning in FINDER; it would probably be better to ask “Was the finder an employee of, or independent contractor to, the other claimant?” However, in the interests of an accurate simulation, the question for A7 has not been rephrased. Leading cases The same cases used by FINDER are used as leading cases in the specifica­ tion. As well as the 270 year old case of Armory v. Delamirie, 3 there are three nineteenth century cases—Bridges v. Hawkesworth, 4 Elwes v. Brigg Gas Co.5 and South Staffordshire Water Co. v. Sharman 6—and three twentieth century cases—Hannah v. Peel, 7 City of London Corporation v. Appleyard 8 and Mof­ fatt v. Kazana. 9 All of these cases are English. Nevertheless, they represent the law as it applies in Australia. London v. Appleyard is used twice. In that case some workers (including Ap­ pleyard) were employed by a firm which was engaged by another firm (Yorkwin Investments Ltd) to perform some construction work on premises leased by York- win from the owner (the City of London).10 The workers found some banknotes during construction. These banknotes were claimed by Appleyard (and his col­ leagues), by Yorkwin and by the City of London. Before McNair J of the Queen’s Bench Division of the English High Court the case was considered as two sep­ arate conflicts: Appleyard (as finder) against Yorkwin, and Yorkwin (as finder) against the City of London. Hence the case appears in the specification as City of London Corporation v. Appleyard (1) and as City of London Corporation v. Appleyard (2). 11 SHYSTER treats each as a separate case; the facts are different, although the result happens to be the same in both cases. Unlike SHYSTER, FINDER does not allow the specification of unknown attrib­ ute values: unknown values are coded as nos in FINDER. 12 Two of the no values in FINDER are actually unknown,13 and two other attribute values in FINDER are incorrect.14 Nevertheless, the fact vectors in the specification are those used in FINDER; in the interests of accurate simulation, no unknowns are included and the errors are not rectified. As it happens, including unknown values and recti­ fying errors changes neither SHYSTER’s opinion nor the cases SHYSTER chooses to justify its opinion in §5.2.3 below. The summary strings for each case are taken (slightly reworded) from FINDER. Attribute dependence SHYSTER detects no functional dependence, and no evidence of stochastic de­ pendence, between the attributes in this specification. The probabilities matrix for the Finder area is given in figure 5.1. 137 � 5.2 A simulation of FINDER A2 A3 A4 A5 A6 A7 A8 A9 A10 0.93 0.82 1.00 0.71 1.00 0.36 1.00 1.00 0.89 A10.50 0.71 0.38 0.82 0.11 1.00 0.36 0.63 0.64 0.93 0.50 0.50 1.00 1.00 1.00 1.00 0.21 A20.50 1.00 0.93 0.21 0.21 0.21 0.50 1.00 0.38 0.82 1.00 0.64 0.89 1.00 0.11 A31.00 0.71 0.36 0.89 0.64 0.38 1.00 0.63 0.75 0.75 1.00 1.00 1.00 A41.00 1.00 1.00 0.75 0.88 0.25 0.89 0.36 0.11 0.38 0.89 A50.64 1.00 1.00 1.00 0.64 0.54 1.00 1.00 0.54 A61.00 0.54 0.75 1.00 1.00 1.00 0.54 A70.54 0.75 1.00 1.00 0.46 A80.25 0.96 0.25 A91.00 Figure 5.1: The probabilities matrix for the Finder area (SHYSTER output). Weights The table of weights for the Finder area, as extracted from the weights file, is given in figure 5.2. The rightmost column indicates that two attributes are of equal greatest importance:15 A4: Was the other claimant the true owner of the chattel or did she/he claim through the rights of the true owner? A9: Was an attempt made to find the true owner of the chattel or, altern­ atively, was the chattel clearly abandoned? That SHYSTER deems A4 to be important is pleasing. The finder has a general right to the found article as against all the world except the true owner. If the other claimant is the true owner then the finder should lose (as happened in Moffatt v. Kazana). Although A9 is important, it probably does not deserve to be considered as important as A4. � � � � � � � 138 Chapter 5: Case studies Win Lose Attr. µ �2 w µ �2 w µ �2 w A1 0.00 0.00 0.60 0.24 4.17 0.38 0.23 4.27 0.00 0.00 0.80 0.16 6.25 0.50 0.25 4.00 A2 0.67 0.22 4.50 0.60 0.24 4.17 0.63 0.23 4.27 A3 0.00 0.00 0.20 0.16 6.25 0.13 0.11 9.14 A4 0.67 0.22 4.50 0.20 0.16 6.25 0.38 0.23 4.27 A5 0.00 0.00 0.40 0.24 4.17 0.25 0.19 5.33 A6 0.00 0.00 0.40 0.24 4.17 0.25 0.19 5.33 A7 0.33 0.22 4.50 1.00 0.00 0.75 0.19 5.33 A8 0.67 0.22 4.50 1.00 0.00 0.88 0.11 9.14 A9 0.33 0.22 4.50 0.20 0.16 6.25 0.25 0.19 5.33 A10 Figure 5.2: The table of weights for the Finder area (SHYSTER output). 5.2.3 Test case A major drawback of the Finder area as a case study for SHYSTER is the dearth of finder cases. Only one such case has been decided since the most recent of FINDER’s leading cases. Parker v. British Airways In Parker v. British Airways, 16 Parker (an airline passenger) found a gold brace­ let in the British Airways international executive lounge at Heathrow Airport, London. He handed the bracelet to a British Airways employee, and left his name and address, asking that the bracelet be returned to him if the true owner could not be found.17 The owner never claimed the bracelet, and British Airways sold it and kept the proceeds: £850. When Parker discovered this, he sued British Airways in the Brentford County Court and was awarded £850 damages plus £50 interest. Brit­ ish Airways appealed to the English Court of Appeal claiming that, as occupiers of the premises, they had rights superior to Parker’s. Counsel for Parker relied heavily on Bridges v. Hawkesworth. 18 Counsel for British Airways, on the other hand, submitted that Bridges could be distinguished in favour of South Staffordshire Water Co. v. Sharman. The three judges who heard Parker v. British Airways in the Court of Appeal unanimously found for Parker. British Airways was held not to have sufficiently manifested an intention to exercise control over lost property before it was found in the executive lounge. 1982 � � � � � � 139 � 5.2 A simulation of FINDER The English Court of Appeal is the highest court to have heard a case in this area. Although not strictly bound by the previously decided cases, the court took those cases into account. Donaldson LJ, who delivered the major judgment, exhaustively discussed the authorities before stating the rights and obligations of the finder and of the occupier in such cases (as quoted in §5.2.1 above). He gave considerable emphasis to Bridges v. Hawkesworth. 19 Eveleigh LJ and Sir David Cairns agreed, Sir David declaring that Bridges v. Hawkesworth “is the closest case on its facts to the present case.”20 The complete report file for Parker v. British Airways is given as an example in §B.2. SHYSTER’s fact vector is (NNNNYNNNYN); to quote the report file: . . . the finder was not the occupier of the premises where the chattel was found; the chattel was not attached; the other claimant was not the owner of the premises where the chattel was found; the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner; the finder handed over the chattel to the other claimant after the finding; neither party relied on the terms of an agreement regarding the right to the chattel; the finder was not a servant of the other claimant; the chattel was not hidden and was not in a position so as to be difficult to find; an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned; and neither party knew of the existence of the chattel prior to the finding. Of these attribute values, only the value of A9 (an attempt was made to find the true owner of the chattel) is contentious. This contention could be important; A9 is one of the most important attributes, according to SHYSTER. The British Airways employee to whom Parker gave the bracelet handed it to the company’s lost property department. The true owner never claimed it. It is not clear from the report of the case whether Parker or British Airways made any attempt to find the true owner beyond leaving the bracelet in the lost property department for some time (e.g. they may have advertised the fact that the bracelet had been found, in the executive lounge). It is at least arguable that the correct value of A9 is unknown, although all three judges who decided the case believed that all reasonable steps to find the true owner had been taken.21 In any event, changing the value of A9 from yes to unknown makes no difference despite that attribute’s weight; SHYSTER comes to the same conclusion on the basis of the same cases, and both instantiations have the same nearest result as does the uninstantiated instant case. 140 Chapter 5: Case studies SHYSTER concludes that Parker should win. SHYSTER’s table of distances, as extracted from the distances file, is given in figure 5.3. The nearest neighbour is Bridges v. Hawkesworth (C2); the nearest other is City of London Corporation v. Appleyard (1) (C5). SHYSTER issues no warnings. SHYSTER comes to a good conclusion, and its choice of nearest neighbour is a good one. Its choice of nearest other is not unjustifiable: the court in London v. Appleyard (1) explicitly followed South Staffordshire v. Sharman—the case upon which British Airways relied.22 As can be seen in figure 5.3, South Staffordshire v. Sharman (C7) is SHYSTER’s (equal) second-nearest other. SHYSTER reports that Parker v. British Airways is not on all fours with Bridges v. Hawkesworth because, in that case, the other claimant was the owner of the premises where the chattel was found. Nevertheless, SHYSTER states its opinion that Bridges v. Hawkesworth should be followed. SHYSTER’s report on this case is substantially identical to that of FINDER given the same facts,23 although this is hardly surprising. 5.2.4 Conclusion SHYSTER is able to simulate FINDER. However due to an almost total lack of test cases, the Finder specification is not of much use in the testing of SHYSTER. The reflexive tests performed in §D.2, and summarized in figure D.1, indicate that the Finder specification is capable of handling new fact situations. 5.3 The authorization of copyright infringement The Authorization specification deals with the meaning of “authorization” in the Australian Copyright Act. This specification is designed to be used with a rule-based system representing the Copyright Act. 5.3.1 The law The Copyright Act 1968 (Cth) is the statutory component of the law of copyright as it applies in Australia. Sections 13(2), 36(1) and 101(1) forbid the authoriza­ tion of copyright infringement: 13. (2) For the purposes of this Act, the exclusive right to do an act in relation to a work, an adaptation of a work or any other subject-matter includes the exclusive right to authorize a person to do that act in relation to that work, adaptation or other subject-matter. � 141 C as e C In st a n t C 1 C 2 C 3 µ W in C 4 C 5 C 6 C 7 C 8 µ Lo se A tt ri bu te s A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 × × × × • × × × • × × × • × × × × • • × × × • × • × × × • × × × × × • × × × × • × × • × • × × × • × • × × • × × × • • • × • × × × × • • • × • • • × • • × • • × × • • × × × • • • × • • • × × • × • • × • • • × × × × • • × c 1 1 1 1 1 1 1 1 d K d U 13 .8 7 – 4. 27 – 14 .4 8 – 4. 27 – 28 .3 4 – 18 .9 3 – 23 .2 0 – 23 .2 0 – 27 .4 7 – 22 .1 3 – 3 1 2 1 5 4 5 5 6 5 S S � 0. 30 0 .2 5 0. 10 0 .0 8 0. 20 0 .2 6 0. 10 0 .0 8 0. 50 0 .5 0 0. 40 0 .3 4 0. 50 0 .4 1 0. 50 0 .4 1 0. 60 0 .4 9 0. 50 0 .3 9 r r � R es ul t 0. 22 0 .6 5 0. 76 0 .9 1 W in 0. 38 0 .1 0 µ + 9. 00 ⊃ 2 � 0. 69 0 .8 6 0. 00 0 .3 7 0. 10 0 .5 6 0. 33 0 .6 6 L o se 0. 00 0 .5 1 0. 10 0 .4 7 − 0. 10 0 .6 8 F ig u re 5 .3 : T h e ta b le o f d is ta n ce s fo r P ar ke r v. B ri ti sh A ir w ay s (S H Y S T E R o u tp u t) . T h e n ea re st n ei gh b ou r is B ri dg es v . H aw ke sw or th ( C 2 ), s o th e n ea re st r es u lt is W in ; t h e n ea re st o th er is C it y of L on do n C or po ra ti on v . A pp le ya rd ( 1) ( C 5 ). � 5.3 The authorization of copyright infringement � � � 142 Chapter 5: Case studies 36. (1) Subject to this Act, the copyright in a literary, dramatic, mu­ sical or artistic work is infringed by a person who, not being the owner of the copyright, and without the licence of the owner of the copyright, does in Australia, or authorizes the doing in Australia of, any act comprised in the copyright. 101. (1) Subject to this Act, a copyright subsisting by virtue of this Part [copyright in subject-matter other than works] is infringed by a person who, not being the owner of the copyright, and without the licence of the owner of the copyright, does in Australia, or authorizes the doing in Australia of, any act comprised in the copyright. Nowhere in the Act is there any other reference to authorization; the meaning of the word is left to the courts to interpret. From 1912, until the 1968 Act came into force (in 1969), the Copyright Act 1911 (UK) was in force in Australia by virtue of the Copyright Act 1912 (Cth). Section 1(2) of the UK Act also forbids authorization of copyright infringement: 1. (2) For the purposes of this Act, “copyright” means the sole right to produce or reproduce the work or any substantial part thereof in any material form whatsoever, to perform, or in the case of a lecture to de­ liver, the work or any substantial part thereof in public; if the work is unpublished, to publish the work or any substantial part thereof; and shall include the sole right,— (a) to produce, reproduce, perform, or publish any translation of the work; (b) in the case of a dramatic work, to convert it into a novel or other non-dramatic work; (c) in the case of a novel or other non-dramatic work, or of an artistic work, to convert it into a dramatic work, by way of performance in public or otherwise; (d) in the case of a literary, dramatic, or musical work, to make any record, perforated roll, cinematograph film, or other contrivance by means of which the work may be mechanically performed or delivered, and to authorize any such acts as aforesaid. As with the 1968 Act, “authorize” is not defined in the UK Act. Since early this century there have been several cases in which courts have had to decide whether copyright infringement had been authorized under one of these Acts. Because of the similarities between the UK Act and the 1968 Act, Australian courts have taken into account cases concerning the former when interpreting the latter.24 143 � 5.3 The authorization of copyright infringement In one of the earliest of these authorization cases, Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd, Scrutton LJ expressed his view that the words “to authorize any such acts as aforesaid” at the end of s. 1(2) of the UK Act “are superfluous and add nothing to the definition.”25 He considered them to be a reference to acting through an agent. However, only two years later Bankes LJ in Falcon v. Famous Players Film Co. decided that “authorize” should be given its ordinary dictionary meaning of “sanction, approve, countenance”. 26 Bankes LJ’s interpretation was adopted by Isaacs J of the High Court of Australia in 1928,27 and has been used ever since. As Ricketson points out, this means that: . . . the notion of “authorisation” extends beyond the authority given to an agent. Thus, if an infringing act is committed by an agent acting within his real or ostensible authority, his principal will be directly liable on the ordinary rules governing the relationship of principal and agent. The same, of course, applies where the infringing act is committed by an employee acting within the course of his employment.28 Gibbs J discussed the meaning of authorization in some detail in University of New South Wales v. Moorhouse. 29 He pointed out that “authorize” can also mean “permit,” though this is tempered by the principle that a person cannot be said to authorize an infringement of copyright unless she/he has the power to prevent it.30 Furthermore: Express or formal permission or sanction, or active conduct indicating approval, is not essential to constitute an authorization . . . However, the word “authorize” connotes a mental element and it could not be inferred that a person had, by mere inactivity, authorized something to be done if he neither knew nor had reason to suspect that the act might be done.31 Having cited several authorities, Gibbs J summarized the law as follows: It seems to me to follow from these statements of principle that a person who has under his control the means by which an infringement of copyright may be committed . . . and who makes it available to other persons, know­ ing, or having reason to suspect, that it is likely to be used for the purpose of committing an infringement, and omitting to take reasonable steps to limit its use to legitimate purposes, would authorize any infringement that resulted from its use.32 However, he warned that: . . . the question of whether one person authorizes another to commit an infringement depends upon all the facts of the case so that a decision on a particular set of circumstances may be of no assistance in other cases.33 144 Chapter 5: Case studies 5.3.2 The AUTHORIZATION specification The author—having consulted a legal textbook,34 the published judgments in the leading cases to which that textbook refers, and a legal expert35—wrote a specification of the meaning of “authorization” in the Copyright Act. This specification was written so as to represent the law as it was in 1983—the textbook states the law as it was then36—so that important cases which have been decided since that time can be used as test cases. The dump file for the Authorization specification is given in full in §A.3. It contains a single area: the Authorization area. This area makes use of several of the features not used in the Finder specification. There is a hierarchy which ranks seven courts at six distinct levels, there is an opening string (although there is no closing string), attributes have help strings, some of the attribute values of the leading cases are unknown, and ideal points and attribute direction are included. Results The Authorization area has three results: Auth, Not-Auth and Liable. These cor­ respond to “the accused authorized the infringement,” “the accused did not au­ thorize the infringement” and “the accused is liable (directly or vicariously) for the infringement.” (An employer is vicariously liable for an employee’s infringing act if the employee is acting within the course of her/his employment.) These are three quite distinct results; the Authorization area demonstrates the need for a case-based system to be able to handle more than two possible outcomes, as discussed in §3.14.1. Attributes There are seven attributes: A1: Was the infringer an employee of the accused? A2: Was the infringer an independent contractor to the accused? A3: Did the accused sell or hire the infringer the means of infringing? A4: Did the accused have the power to prevent the infringement? A5: Did the accused take reasonable steps to avoid the infringement? A6: Did the accused know, or have reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done? A7: Was the specific infringement causally related to an incitement to infringe on the part of the accused? 145 � 5.3 The authorization of copyright infringement Although there will be times when the answers to A1 and A2 are obvious, these questions are not always easy to answer. The Employee specification de­ scribed in §5.4.2 below is concerned solely with deciding whether a worker is an employee or an independent contractor. If the Employee area from that specifica­ tion were included in the Authorization specification then either or both of A1 and A2 could be made external attributes; i.e. rather than asking the user a single question, the attribute value would be determined by reference to the Employee area, which has 18 attributes.37 The Natural specification described in §5.5.2 illustrates this process: it has three areas, two of which have external attributes. However, a simpler approach was adopted for the Authorization specific­ ation. Both A1 and A2 have help strings which give assistance to the user in answering the attribute question. For example, the help string for A1 is: If the accused had control over the infringer’s manner of doing her/his work then the infringer was an employee of the accused. If the infringer undertook to do something for the accused and had discretion as to the manner in which it was to be done then the infringer was an independent contractor to the accused, not an employee. (This is a greatly simplified statement of the law, as can be seen by comparison with the Employee specification.) Of the other attribute questions, only A5 should present any difficulty. This attribute is very important. If the accused did take reasonable steps to avoid the infringement then she/he will not be said to have authorized the infringement,38 although she/he could still be directly or vicariously liable. In all of the cases that make up the Authorization area the value of A5 is no: i.e. in none of the cases did the accused take reasonable steps. This is not surprising; if reasonable steps had been taken, the cases would probably never have been taken to court. It also means that SHYSTER assigns A5 infinite weight. The attribute has a fairly unhelpful help string: “Whether particular steps were reasonable depends on the facts of the case.” But the nebulous concept of “reasonableness” is one with which lawyers very often have to deal. A lawyer would probably not have much difficulty in answering this question, and she/he always has the option of answering unknown and forcing SHYSTER to examine instantiations. If A5 is changed to unknown in all of the test cases in §5.3.3, SHYSTER comes to the same conclusion on the basis of the same cases. This does not mean that A5 is irrelevant; its inclusion (and its infinite weighting) means that if the accused did not take reasonable steps to avoid the infringement, the specified directions and ideal point directions will strongly suggest Not-Auth. 39 In theory, it should be possible to specify another area of case law which defines this very open-textured concept of reasonableness, however that might prove practically impossible.40 146 Chapter 5: Case studies A2 A3 A4 A5 A6 A7 0.75 0.50 1.00 1.00 1.00 1.00 A11.00 1.00 0.63 1.00 0.88 0.50 0.21 1.00 1.00 0.25 0.79 A21.00 0.36 1.00 1.00 0.79 0.40 1.00 1.00 0.64 A30.95 1.00 0.56 0.83 1.00 0.67 0.95 A41.00 1.00 0.40 1.00 1.00 A51.00 1.00 1.00 A60.44 Figure 5.4: The probabilities matrix for the Authorization area (SHYSTER out­ put). Leading cases There are nine leading cases in the Authorization area. Three of them arose under the Copyright Act 1911 (UK): Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd, 41 Falcon v. Famous Players Film Co.42 and A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd. 43 Four arose under the UK Act, in force in Aus­ tralia by virtue of the Copyright Act 1912 (Cth): Mellor v. Australian Broad­ casting Commission, 44 Winstone v. Wurlitzer Automatic Phonograph Co. of Aus­ tralia Pty Ltd, 45 Australasian Performing Right Association Ltd v. Miles46 and Australasian Performing Right Association Ltd v. Canterbury-Bankstown League Club Ltd. 47 And two arose under the Copyright Act 1968 (Cth): University of New South Wales v. Moorhouse48 and RCA Corporation v. John Fairfax and Sons Ltd. 49 SHYSTER warns that APRA v. Canterbury-Bankstown and Mellor v. ABC have identical facts (except for unknown values), and that APRA v. Canterbury- Bankstown and APRA v. Miles have identical facts (except for unknown values) and different results. The first draft of the specification included an early King’s Bench Division case: Performing Right Society Ltd v. Bradford Corporation. 50 That case has an identical fact vector to that of Mellor v. ABC. On reflection, the former case was omitted in favour of the latter, later, and more important, Privy Council decision. (See §5.5.2 below for a discussion of identical cases.) � � � � � � � � � � � � � � 147 � 5.3 The authorization of copyright infringement Auth Not-Auth Liable Attr. µ �2 w µ �2 w µ �2 w µ �2 w A1 0.00 0.00 0.00 0.00 1.00 0.00 0.13 0.11 9.14 0.25 0.19 5.33 0.33 0.22 4.50 0.00 0.00 0.25 0.19 5.33A2 0.60 0.24 4.17 0.33 0.22 4.50 0.00 0.00 0.44 0.25 4.05A3 0.80 0.16 6.25 0.33 0.22 4.50 1.00 0.00 0.67 0.22 4.50A4 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00A5 1.00 0.00 0.67 0.22 4.50 1.00 0.00 0.89 0.10 10.13 A6 0.80 0.16 6.25 0.00 0.00 1.00 0.00 0.56 0.25 4.05A7 Figure 5.5: The table of weights for the Authorization area (SHYSTER output). Attribute dependence SHYSTER detects no functional dependence, and no evidence of stochastic de­ pendence, between the attributes in this specification. The probabilities matrix for the Authorization area is given in figure 5.4. Weights The table of weights for the Authorization area, as extracted from the weights file, is given in figure 5.5. The rightmost column of the table indicates that, as discussed above, A5 is given infinite weight.51 The next most important attributes are: A6: Did the accused know, or have reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done? A1: Was the infringer an employee of the accused? These attributes are indeed important, although A4 (“Did the accused have the power to prevent the infringement?”) should probably be more heavily weighted. 5.3.3 Test cases The Authorization specification was written so as to represent the law as it was in 1983. Three authorization cases have been decided since then, and they are used to test SHYSTER. Another case was decided before 1983 but was not considered important enough to include in the specification. That case, CBS Inc. v. Ames Records and Tapes Ltd, 52 is used as a test case too. There are also two hypothetical test cases, chosen by the legal expert. The results of all these tests are summarized in figure 5.12. � � � � � � 148 Chapter 5: Case studies The Authorization area has three results, and SHYSTER will always look for nearest cases for each result. So, SHYSTER argues about the Liable result even in those test cases where the courts did not explicitly consider the possibility that the accused was directly liable. It is not inappropriate to argue about the Liable result in such cases; the judges must have (at least tacitly) rejected that conclusion themselves. It would only be cause for concern if SHYSTER’s opinion were that the accused was directly liable in such a case—but it does not make that mistake in any of these tests. CBS Inc. v. Ames Records and Tapes Ltd 1982 In CBS Inc. v. Ames Records and Tapes Ltd, 53 Ames operated a record lending library. As well as hiring records, Ames sold blank cassettes. CBS sold records to Ames, and held the copyright in those sound recordings. CBS claimed that Ames knew that it was likely that their customers would copy the records onto blank cassettes at home, and that hence Ames was “authorizing” infringement within the meaning of 1(2) of the Copyright Act 1956 (UK). Before Whitford J of the Chancery Division of the English High Court, CBS relied on Falcon v. Famous Players Film Co., Vigneux v. Canadian Performing Right Society Ltd 54 (not one of the leading cases in the Authorization specific­ ation), Winstone v. Wurlitzer Automatic Phonograph Co. of Australia Pty Ltd and University of New South Wales v. Moorhouse. They urged that A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd be distinguished on the basis that “here the intention and desire was to take advantage of home taping.”55 Ames argued that, applying the dictionary meaning of authorization as did Bankes LJ in Falcon v. Famous Players, they had not authorized any infringe­ ment. They also claimed that UNSW v. Moorhouse was wrongly decided.56 Ames relied on A&M v. Audio Magnetics, pointing to the absence of evidence of a par­ ticular case of infringement. Whitford J agreed with Ames, although he made no mention of A&M v. Audio Magnetics in his judgment. SHYSTER’s fact vector is (NNYNNYN); to quote its report file: . . . the infringer was not an employee of the accused; the infringer was not an independent contractor to the accused; the accused sold or hired the infringer the means of infringing; the accused did not have the power to prevent the infringement; the accused did not take reasonable steps to avoid the infringement; the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done; and the specific infringement was not causally related to an incitement to infringe on the part of the accused. � � � � 5.3 The authorization of copyright infringement 149 SHYSTER agrees with Whitford J that Ames did not authorize the infringement. SHYSTER’s table of distances, as extracted from the distances file, is given in fig­ ure 5.6. The nearest neighbour is A&M v. Audio Magnetics (C8): one of the cases upon which Ames relied; the nearest others are Falcon v. Famous Players (C5: Auth)—upon which CBS relied—and Australasian Performing Right Association Ltd v. Miles (C9: Liable). SHYSTER warns that the specified directions suggest the Auth result, as can be seen in figure 5.6: the strength of the specified direction for Auth (3�+4.16) is greater than that for the other two results. WEA International Inc. v. Hanimex Corporation Ltd 1987 In WEA International Inc. v. Hanimex Corporation Ltd, 57 Hanimex advertised one of their cassette products on radio. These humorous advertisements implied that, while the cassette recordings of some artists’ songs would melt in a hot car, Hanimex’s cassettes would not: “If you don’t want your favourite recordings ruined, use Fuji GTI Car Tapes.” WEA owned the copyright in sound recordings made by the singer Madonna— one of the artists mentioned by name in the advertisements. WEA claimed that Hanimex had authorized members of the public to infringe their copyright, con­ trary to the Copyright Act 1968 (Cth). The case was heard in the Federal Court of Australia by Gummow J. His judg­ ment includes a long history of authorization provisions in copyright legislation since the turn of the century. Of the recent cases, he referred to University of New South Wales v. Moorhouse, CBS Inc. v. Ames Records and Tapes Ltd and RCA Corporation v. John Fairfax and Sons Ltd. 58 He pointed out that RCA v. Fairfax was authority for the proposition that authorization of copyright infringement requires an actual act of infringement, and observed that: . . . it has not even been shown that there has been any unauthorized repro­ duction by any particular person of any of the sound recordings in which [WEA] hold copyright.59 In addition, Gummow J held that the advertisements were not “an invitation or incitement to or approval of” copyright infringement.60 SHYSTER’s fact vector is (NNYNNYN); as it is in CBS v. Ames. Hence, SHYSTER’s distances and report files are identical to those for that case. SHYSTER agrees that Hanimex did not authorize the infringement, although it warns that the specified directions suggest Auth. The nearest neighbour is A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd. That this case was not mentioned by Gummow J in his judgment is most surprising. The facts of A&M v. Audio Magnetics are very similar to those of × × × × • × × � 150 A tt ri bu te s C as e c d K d U � S r r � R es ul t A 1 A 2 A 3 A 4 A 5 A 6 A 7 C In st a n t × × • × × • × C 1 × × • • × • × 1 4. 50 – 1 0. 14 0. 73 0. 90 C 2 × • × • • 2 12 .6 0 14 .4 8 3 0. 60 − 0. 17 0. 67 A u th C 3 × × • • × • • 3 8. 55 – 2 0. 29 0. 55 0. 85 ⊃ 3 � + 4. 17 4. 17 C 4 C 5 × • × • × • • 4 4. 05 – 1 0. 14 × × • × × • • 5 17 .9 3 – 4 0. 57 0. 09 0. 61 − 0. 73 0. 92 13 .8 8 – 3 0. 43 0. 40 0. 77 I ⊃ µ ⊃ 4 .1 7 I A u th × • • • × • • µ Au th × × • • × • • 8. 55 – 2 0. 29 0. 51 0. 87 C 6 C 7 × × × × × • × 3 24 .0 1 – 4 0. 57 × • × • × × × 5 4. 05 – 1 0. 14 0. 65 0. 92 0. 40 − 0. 35 1. 00 1. 00 N o t- A u th C 8 – – – – × × • × × • × 6 I N o t- A u th I µ − ⊃ 2 � + 9. 00 ⊃ � + 4. 50 + 14 .1 8 – 3 0. 43 0. 26 − 0. 23 − 0. 71 ⊃ � Li a b le + 4. 50 µ No t- A u th × × × × × • × 4. 05 – 1 0. 14 0. 91 • × × • × • • 3 I L ia b le 13 .6 4 8. 10 2 0. 40 21 .7 4 – 4 0. 57 0. 09 0. 46 C 9 − • × • × • 21 .7 4 – 4 0. 57 0. 09 0. 46 − 0. 41 0. 62 ⊃ 3 � µ Li a b le • × × • × • • F ig u re 5 .6 : T h e ta b le o f d is ta n ce s fo r C B S I n c. v . A m es R ec or ds a n d T ap es L td ( S H Y S T E R o u tp u t) . T h e n ea re st n ei gh b ou r is A & M R ec or ds I n c. v . A ud io M ag n et ic s In c. ( U K ) L td (C 8 ), s o th e n ea re st re su lt i s N o t- A u th ; th e n ea re st o th er s ar e F al co n v . F am ou s P la ye rs F il m C o. ( C 5 ) an d A us tr al as ia n P er fo rm in g R ig ht A ss oc ia ti on L td v . M il es ( C 9 ). N ot e th at t h e sp ec ifi ed d ir ec ti on s (⊃ ) su gg es t A u th . Chapter 5: Case studies � � � � 5.3 The authorization of copyright infringement 151 WEA v. Hanimex : they both concern advertisements claiming high durability of blank cassettes and referring to specific recording artists. A&M v. Audio Magnet­ ics is a decision of the Chancery Division of the English High Court, and not as good authority in Australia as a decision of the NSW Supreme Court, like RCA v. Fairfax. Nevertheless, it is amazing, given the length of his discussion of the development of the law in Australia, England and the US, that Gummow J did not refer to A&M v. Audio Magnetics. Both A&M v. Audio Magnetics and RCA v. Fairfax were decided on the same point: that there must be an actual infringing act before someone can be said to have authorized an infringement. The judge in RCA v. Fairfax explicitly followed A&M v. Audio Magnetics, amongst other cases, in coming to his decision.61 Hence, the author claims that SHYSTER’s choice of nearest neighbour is a good one, despite the fact that it was not referred to in WEA v. Hanimex. Gummow J made no reference to SHYSTER’s nearest others: Falcon v. Famous Players Film Co. (Auth) and Australasian Performing Right Association Ltd v. Miles (Liable). CBS Songs Ltd v. Amstrad Consumer Electronics plc 1988 In CBS Songs Ltd v. Amstrad Consumer Electronics plc, 62 Amstrad manufactured twin-deck tape-recording machines. They were designed so as to facilitate tape- to-tape copying, and were advertised in a manner that was likely to encourage home taping and copying. However, the advertisements warned that some copy­ ing could not be done without permission, and made it clear that Amstrad had no authority to grant such permission. CBS and other record companies sued Amstrad claiming that they had au­ thorized users of their twin-deck recorders to infringe copyright. Before the House of Lords, CBS cited many cases in argument, including Falcon v. Famous Players Film Co. Amstrad relied on A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd, amongst others. As in CBS Inc. v. Ames Records and Tapes Ltd, A&M v. Audio Magnetics— although used in argument—was not referred to in the judgments. Instead, Lord Templeman (with whom the other four law lords agreed) found for Amstrad on the basis of Bankes LJ’s definition of authorization in Falcon v. Famous Players, and following CBS Inc. v. Ames Records and Tapes Ltd. SHYSTER’s fact vector is (NNYNNYN); as it is in CBS v. Ames and WEA Interna­ tional Inc. v. Hanimex Corporation Ltd. � � � 152 Chapter 5: Case studies SHYSTER agrees that Amstrad did not authorize the infringement. Its choice of cases is good: the nearest neighbour is A&M v. Audio Magnetics; the nearest others are Falcon v. Famous Players (Auth) and Australasian Performing Right Association Ltd v. Miles (Liable). However, SHYSTER warns that the specified directions suggest Auth. Australasian Performing Right Association Ltd v. Jain 1990 In Australasian Performing Right Association Ltd v. Jain, 63 Jain was the director and principal executive officer of a company (Valamo Pty Ltd) which was the proprietor of a tavern. Live bands, and a video music system, performed works in which the APRA held copyright. It was conceded that Valemo had infringed copyright, but Jain claimed that he was not personally liable. The day-to-day operations of the tavern were controlled by another person (the licensee). Jain claimed that the licensee was responsible for all the music at the tavern, and that he (Jain) only took an interest if takings were down, in which case he would advise the licensee to engage a different band. The Full Court of the Federal Court of Australia held that Jain had authorized the infringement. Sheppard, Foster and Hill JJ held that Jain had allowed: . . . a situation to develop and to continue in which he must have known that it was likely that the [APRA’s] music would be played without any licence from it. It was within his power to control what was occurring but he did nothing at all.64 In coming to their conclusion, their honours mentioned many cases. To the extent that they followed any particular case, they followed University of New South Wales v. Moorhouse. 65 The complete report file for APRA v. Jain is given as an example in §B.3. SHYSTER’s fact vector is (NYNYNYN). SHYSTER agrees with the Federal Court that Jain authorized the infringement. However, none of the cases chosen by SHYSTER was mentioned by the Court. The nearest neighbour is Mellor v. Aus­ tralian Broadcasting Commission; the nearest others are RCA Corporation v. John Fairfax and Sons Ltd (Not-Auth) and Australasian Performing Right Association Ltd v. Miles (Liable). There are no warnings. Hypothetical case 1 The legal expert suggested a hypothetical case for testing the Authorization specification: a fact situation about which he had been asked for advice, but on which no court has been forced to rule. � � � � � � 153 � 5.3 The authorization of copyright infringement A residential college on a university campus provides personal computer fa­ cilities for the use of its residents. It also makes available to the residents several items of software. Neither the college nor the residents hold the copyright in this software, although the software could easily be copied. The college is concerned as to whether, if a resident were to copy an item of software, the college could be said to have authorized that infringement. The legal expert advises that the college could be said to have authorized the infringement. SHYSTER’s fact vector is (NNNYNYN). SHYSTER’s conclusion is that the college would have authorized the infringement. The nearest neighbour is University of New South Wales v. Moorhouse; the nearest others are RCA Corporation v. John Fairfax and Sons Ltd (Not-Auth) and Australasian Performing Right Association Ltd v. Miles (Liable). However, it warns that the specified directions suggest Liable. The legal expert advises that SHYSTER’s choice of cases, nearest neighbour and nearest others, is good. Hypothetical case 2 The legal expert suggested a second hypothetical authorization case. A language school lends audio tapes to its students and allows them to take those tapes home. One of the language school’s employees gives a student one of these tapes and tells him that “lots of people copy them.” Could the school be said to have authorized an infringement if the student makes a copy of the tape? The legal expert advises that the language school would probably not be held to have authorized the infringement. SHYSTER’s fact vector is (NNNNNYY). Its conclusion is that the language school would not have authorized the student’s infringement. The nearest neighbour is RCA Corporation v. John Fairfax and Sons Ltd ; the nearest others are Falcon v. Famous Players Film Co. (Auth) and Australasian Performing Right Association Ltd v. Miles (Liable). However, SHYSTER warns that two of these cases are equidistant from the instant case; it chooses RCA v. Fairfax because it is a decision of the New South Wales Supreme Court whereas Falcon v. Famous Players is merely a decision of the English Court of Appeal. The weighted correlation coefficients suggest that Falcon v. Famous Players should be the nearest neighbour, and the specified directions suggest Liable. Again, the legal expert advises that SHYSTER’s choice of cases is good. 154 Chapter 5: Case studies 5.3.4 Generated tests The legal expert provided three examples for generating tests using the Author­ ization specification: generated tests 1, 2 and 3. The results of these tests are summarized in figure 5.13. Generated test 1 The legal expert advises that, if the following is true, either the accused au­ thorized the infringement, or the accused is directly or vicariously liable for the infringement: . . . the infringer was an employee of the accused; the infringer was not an independent contractor to the accused; and the accused did not take reasonable steps to avoid the infringement. Given the fact vector (YNUUNUU), SHYSTER generates 16 instantiations: one eighth of the total search space. In all but two of these instantiations, SHYSTER’s chosen results are good (i.e. Auth or Liable). In the two instantiations in which the result is Not-Auth, the nearest ideal point is that of a good result, but those ideal points are not nearer the instant case than the nearest neighbour so no warnings are issued. In both of those instantiations, the accused did not have the power to prevent the infringement, and the specific infringement was not causally related to an incitement to infringe on the part of the accused. The legal expert agrees that these attribute values justify SHYSTER’s choice of result in both instantiations. Generated test 2 The legal expert also advises that, if the following is true, the accused did not authorize the infringement: . . . the infringer was not an employee of the accused; the accused did not have the power to prevent the infringement; and the accused took reasonable steps to avoid the infringement. The fact vector is (NUUNYUU). SHYSTER generates 16 instantiations, in all but four of which SHYSTER’s chosen result is good (i.e. Not-Auth). In each of those four instantiations, SHYSTER warns that an ideal point sug­ gesting the good result is at least as near to the instant case as is the nearest neighbour. The only attribute for which these four instantiations have common values (apart from the three attributes with known values in the above fact vector) is the last: in each, the specific infringement was causally related to an incitement to infringe on the part of the accused. Although this attribute value does suggest Auth or Liable, 66 the three known attribute values in the above fact vector are more persuasive: SHYSTER’s choice of results in these four instantiations is bad. � � � � 5.4 Employees and independent contractors 155 Generated test 3 The legal expert advises that, if the accused took reasonable steps to avoid the infringement, she/he cannot have authorized the infringement (although she/he may be directly or vicariously liable for the infringement). SHYSTER’s fact vector is (UUUUYUU). It generates 64 instantiations: half the search space. In 13 of these 64 instantiations, SHYSTER’s chosen result is bad (i.e. Auth). In all 13, SHYSTER warns that an ideal point suggesting a good result is at least as near to the instant case as was the nearest neighbour. The fact that, in each of these 13 instantiations, the infringer was not an employee of the accused, does not justify SHYSTER’s choice of result in these instantiations. 5.3.5 Conclusion In the test cases described in §5.3.3, SHYSTER chooses a good result every time, and 75% of its chosen cases are good. However, the small number of attributes specified for the Authorization area means that SHYSTER has difficulty distinguish­ ing between cases: e.g. the first three test cases share the same fact vector. This may indicate a deficiency in the specification, or it may be indicative of a feature of the area. As shown in figure 5.13, SHYSTER’s choice of result is good in 80.21% of the instantiations in all three generated tests. If a bad choice of result is con­ sidered good where an ideal point warning is issued, SHYSTER’s success rate rises to 87.50%, 100.00% and 100.00% for each test: 97.92% overall. (In none of the instantiations in these three tests does SHYSTER issue a warning about a nearer ideal point when the choice of results is good; i.e. none of SHYSTER’s ideal point warnings is bad.) The Authorization area is a good example of an area which could be used to define an open-textured statutory concept in a rule base: the meaning of “authorization” is undefined in sections 13(2), 36(1) and 101(1) of the Copyright Act. It is also a good example of an area of law in which there are more than two possible results. SHYSTER (when using the Authorization specification) produces good ad­ vice in the test cases, and chooses good results in almost all of the generated tests. The reflexive tests performed in §D.3, and summarized in figure D.2, indicate that the Authorization specification is capable of handling new fact situations. 5.4 Employees and independent contractors The Employee specification deals with the legal categorization of a worker as an employee or as an independent contractor. 156 Chapter 5: Case studies 5.4.1 The law If a person works for another, one of two relationships exists: either they are employer and employee—and theirs is a contract of service—or they are principal and independent contractor—and theirs is a contract for services. (The parties to a contract of service used to be called “master” and “ser­ vant,” but this terminology is no longer used.67 When it is not yet clear which sort of relationship exists, it is convenient to use generic terms: “employer” and “worker” are used to mean “employer” and “employee”—in the case of a con­ tract of service—or “principal” and “independent contractor”—in the case of a contract for services.) This distinction is important; the law of employment is concerned almost totally with contracts of service.68 As mentioned in §5.3.2 above, an employer is vicariously liable for the actions of an employee if the employee is acting within the course of her/his employment. A principal would not normally be vicariously liable for the actions of an independent contractor. The distinction also affects the terms that will be implied into the employ­ ment contract in the absence of an express agreement between the parties, the applicability of industrial awards, the applicability of statutes which may affect workers’ compensation, occupational health and safety, long-service leave, annual holidays, unfair dismissal, payroll tax, fringe benefits tax, etc.69 As Creighton, Ford and Mitchell point out: It is most important therefore to be able to determine whether any given worker is engaged under a contract of employment. Unfortunately neither courts nor parliaments have been consistent in the attitude they have ad­ opted to this process of categorization. In particular they have contrived to make it virtually impossible to draw a clear and consistent distinction between ‘employees’ and ‘independent contractors.’ . . . Parliaments tend to adopt one of two approaches to the process of categorization. The first, and simplest, is to ignore the issue entirely, and to leave the matter to be determined by reference to common law criteria . . . The other approach is to include some form of definition, usually in the interpretation section. In some instances this definition is purely circular in form . . . whereas in others it consists of either a broadening or a narrowing of the common law concept (sometimes going as far as to include ‘independent contractors’). Whatever approach is adopted, the final answer is to be found in the common law.70 The most commonly applied criterion for distinguishing between a contract of service and a contract for services is the “control test.” If an employer has a right of control over a worker’s manner of doing her/his work then the control test suggests that the contract is of service and that the worker is an employee. Conversely, if a worker has discretion as to the manner in which the work is to be done then the test suggests that the contract is for services.71 157 � 5.4 Employees and independent contractors However, courts have explicitly had regard to many different factors in decid­ ing whether a worker is an employee or an independent contractor. In Price v. Grant Industries Pty Ltd, for example, the Full Court of the Federal Court ex­ amined in detail fifteen points which Price claimed were indicative of a contract of service; Grant Industries relied on eleven different points which it claimed were indicative of a contract for services.72 This is an area of the law in which many factors have to be balanced before a conclusion can be reached. 5.4.2 The EMPLOYEE specification The author—having consulted a legal textbook,73 Halsbury’s, 74 the published judgments in the leading cases to which they refer, and a legal expert75—wrote a specification of the distinction between an employee and an independent con­ tractor. This specification was written so as to represent the law as it was in 1982 so that important cases which have been decided since then can be used as test 76cases. The Employee specification is used as the basis of the complete example of SHYSTER’s output files in appendix C. The specification file, written in SHYSTER’s case law specification language and input by SHYSTER, is in §C.3. The dump file for the Employee specification is given in full in §A.4. The specification contains a single area: the Employee area. There is a hier­ archy which ranks nine courts appropriately, and there is an opening string (al­ though there is no closing string). Results The Employee area has two results: Employee and Contractor. These correspond to “the worker is an employee” and “the worker is an independent contractor.” Attributes In keeping with the large number of important factors in this area, there are eighteen attributes: A1: Did the employer direct not only what work was to be done, but also the manner in which it was to be done? A2: Was the worker allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand? A3: Was the worker an integral part of the employer’s business? A4: Did the worker own the tools or provide the transport with which she/he performed the work? A5: Would the employer make a profit/loss if the work performed by the worker cost less/more than expected? 158 Chapter 5: Case studies A6: Was the work performed on the employer’s premises? A7: Did the employer supervise or inspect the work? A8: Was the worker in business on her/his own account? A9: Was the worker allowed to employ others to assist with her/his work? A10: Was the worker obliged to work only for the employer? A11: Was the worker required to work at specified times? A12: Did the employer pay the worker by time? A13: Was the money that the employer paid to the worker stated to be a “fee”? A14: Was the money that the employer paid to the worker stated to be “wages” or “salary”? A15: Did the employer deduct PAYE77 tax instalments from the worker’s pay? A16: Did the employer pay the worker sick pay or holiday pay? A17: Did the employer and the worker express an intention that the rela­ tionship would be one of employer and employee? A18: Did the employer and the worker express an intention that the rela­ tionship would be one of principal and independent contractor? Answering these questions should be fairly straightforward. Only A3 could cause difficulty. The help string for that attribute suggests that: If the worker was “part and parcel” of the employer’s business then she/he was an integral part of the business, not merely accessory to it. And the user always has the option of answering unknown if unsure. The first draft of the specification included three extra attributes: • Did the employer have to pay the costs of performing the work? • Did the employer stand to make a profit if the work performed by the worker cost less than expected? • Did the employer bear the risk of loss if the work performed by the worker cost more than expected? These attributes were chosen because courts have had regard to these questions in employee/independent contractor cases. However, SHYSTER detected functional dependence between each of these attributes and the other two: viz. every case in 159 � 5.4 Employees and independent contractors the specification had the same attribute values for each of these three attributes. It was decided that these three attributes were, in fact, three different ways of asking the same question, so they were removed and replaced with a single attribute: A5: Would the employer make a profit/loss if the work performed by the worker cost less/more than expected? Leading cases There are thirteen leading cases in the Employee area. Three of them concerned the applicability of an industrial award: Cam and Sons Pty Ltd v. Sargent, 78 Aus­ tralian Timber Workers Union v. Monaro Sawmills Pty Ltd 79 and Price v. Grant Industries Pty Ltd. 80 Three others concerned the payment of tax or payments to a government contribution scheme: Federal Commissioner of Taxation v. J. Walter Thompson (Australia) Pty Ltd, 81 Queensland Stations Pty Ltd v. Federal Commissioner of Taxation82 and Ready Mixed Concrete (South East) Ltd v. Min­ ister of Pensions and National Insurance. 83 Two concerned workers’ compens­ ation: Zuijs v. Wirth Brothers Pty Ltd 84 and Humberstone v. Northern Timber Mills. 85 Two arose out of copyright claims: Stevenson Jordon and Harrison Ltd v. Macdonald and Evans 86 and Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd. 87 One was a claim for damages for breach of statutory duty: Ferguson v. John Dawson & Partners (Contractors) Ltd. 88 Another con­ cerned a worker’s entitlement to long service leave: Australian Mutual Provident Society v. Chaplin. 89 And one was a claim for compensation for unfair dismissal: Massey v. Crown Life Insurance Co.90 Stevenson v. Macdonald is used twice. In that case Macdonald and Evans claimed copyright in a book because its author had written it while he was work­ ing for them. The book was divided into sections, which were written under different circumstances. The English Court of Appeal looked at each section separately, to determine whether that section had been written by its author as an employee of Macdonald and Evans. The case appears as Stevenson v. Macdonald (1)—which deals with the first section—and as Stevenson v. Macdon­ ald (2)—which deals with the second. SHYSTER treats each as a separate case; the facts and the result are different in both. Attribute dependence The probabilities file for the Employee specification is given in full in §C.7. SHYSTER detects functional dependence between A4 and A5. Although, in the leading cases, the inverse function maps the attribute values of one to those of the other, the two attributes represent different questions: there is no reason to believe that there is any relationship between the worker not owning the tools 160 Chapter 5: Case studies or providing the transport with which she/he performs the work (A4), and the employer standing to make a profit/loss if the work performed by the worker costs less/more than expected (A5). Hence, both attributes remain in the specification. SHYSTER also detects evidence of stochastic dependence between A3 and A17, A4 and A11, A5 and A11, A7 and A9, A9 and A11, and A11 and A12. Only in two of these pairs—the first and the last—is there any danger that the attributes are actually asking significantly similar (or dissimilar) questions. After careful consideration it was decided that all of these attributes, even those in the first and last pairs, are different questions and important in distinguishing between employees and independent contractors. So these attributes, too, remain in the specification. Weights The weights file for the Employee specification is given in full in §C.9. The rightmost column of the table of weights indicates that three attributes are of equal greatest importance in the Employee area:91 A13: Was the money that the employer paid to the worker stated to be a “fee”? A14: Was the money that the employer paid to the worker stated to be “wages” or “salary”? A16: Did the employer pay the worker sick pay or holiday pay? The fact that the courts have had regard to so many different factors when deciding employee cases, makes it difficult to judge SHYSTER’s choice of most important attributes. However, A1 and A2 should probably be more heavily weighted than they are. 5.4.3 Test cases The Employee specification was written so as to represent the law as it was in 1982. Four actual cases are used to test the Employee specification—cases which were decided after that time. There are also two hypothetical cases, chosen by the legal expert. The results of all these tests are summarized in figure 5.12. Narich Pty Ltd v. Commissioner of Pay-roll Tax 1983 In Narich Pty Ltd v. Commissioner of Pay-roll Tax, 92 Narich held the Australian franchise of Weight Watchers. Lecturers taught classes, following a program detailed in the Weight Watchers handbook, and deducted their fees from the � � � 161 � 5.4 Employees and independent contractors money that they collected from class members. Clause 3 of the agreement between the lecturers and Narich stated that the lecturers were not employees of Narich, but independent contractors. The Commissioner claimed that the lecturers were actually employees of Narich, and that Narich was liable, under the Pay-roll Tax Act 1971 (NSW), to pay payroll tax on the money paid to the lecturers. The Judicial Committee of the Privy Council cited Australian Mutual Provid­ ent Society v. Chaplin as authority for the principle that a statement in an agreement that a worker is an “independent contractor” or an “employee” is not decisive. Their lordships held that, despite clause 3, the effect of the agree­ ment between the lecturers and Narich was that the lecturers were employees of Narich. SHYSTER’s fact vector is (YNYNYNNNNNYYYNNNNY); to quote its report file: . . . the employer directed the manner in which the work was to be done; the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was an integral part of the employer’s business; the worker neither owned the tools nor provided the transport with which she/he performed the work; the employer would make a profit/loss if the work performed by the worker cost less/more than expected; the work was not performed on the employer’s premises; the employer neither supervised nor inspected the work; the worker was not in business on her/his own account; the worker was not allowed to employ others to assist with her/his work; the worker was not obliged to work only for the employer; the worker was required to work at specified times; the employer paid the worker by time; the money that the employer paid to the worker was stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. SHYSTER agrees with the Privy Council that the lecturers are employees, al­ though it warns that the specified directions suggest Contractor. Its table of distances, as extracted from the distances file, is given in figure 5.7. None of SHYSTER’s chosen cases was cited in the judgment: the nearest neighbour is Fer­ guson v. John Dawson & Partners (Contractors) Ltd (C5); the nearest other is Massey v. Crown Life Insurance Co. (C12). � � � 162 Chapter 5: Case studies A tt r ib u te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 • × • × • × × × × × • • • × × × × • × • • × • • • × × × • • × × • × × × 1 • • • × • × × × • × × × × × × × × × 2 • × • × • • • × × • × • × × × × × 4 • × • • × • • × × × × × × × × × × 5 S S � 0 .3 9 0 .3 8 0 .3 3 0 .3 2 0 .2 4 0 .1 6 0 .4 7 0 .3 8 C a se C In s t a n t C 1 C 2 C 3 C 4 r r � R e su lt 0 .2 0 0 .0 1 − 0 .2 7 0 .1 0 0 .5 1 0 .8 0 E m p lo y e e I − 0. 0 2 − 0. 0 1 � 2 � + 2 9 .9 8 d K d U 4 3 .5 7 – 3 6 .4 5 – 1 7 .7 6 5 .6 3 4 2 .8 5 4 .0 2 7 6 4 8 C 5 7 1 7 .0 9 – 2 0 .1 1 0 .1 5 0 .7 7 0 .4 2 • × • × • × • × × × • • × × × × × • 9 8 5 1 0 1 3 1 0 µ I� 2 � 0 .0 3 0 .1 8 − 0 .2 1 0 .2 5 − 0 .4 7 0 .3 7 − 0. 3 6 − 0. 2 9 − 0. 5 6 − 0. 4 5 − 0. 2 4 − 0. 1 7 C o n tr a c to r − 0. 1 7 − 0. 1 4 � 2 � + 3 8 .7 5 + 3 4 .8 8 C 6 × • • × • × × × × • × 7 1 4 .9 2 6 2 .6 5 3 0 .2 7 0 .2 8 0 .4 5 0 .4 1 3 0 .8 0 C 7 I Em p lo ye e µ Em p lo ye e C 8 C 9 C 1 0 C 1 1 • • • • × • • × × • • • × • × × × × 8 • × • × • • • × × • • • × • • • • • • • × • • • × × × • • × × × × × × × • × • × × × × • × × × × × × × × × 3 × • × • × × • • • × × × × × × × • × 3 × • • • × × × × • × × × × × • × × × 5 × • • • × × × • • • × × × × × × × • 6 5 9 .3 2 – 6 8 .4 5 4 .9 0 3 2 .2 9 – 5 4 .4 8 – 7 1 .7 9 – 5 5 .7 4 – 0 .5 0 0 .5 1 0 .4 7 0 .6 2 0 .2 8 0 .2 8 0 .5 6 0 .4 7 0 .7 2 0 .6 2 0 .5 6 0 .4 8 0 .5 6 0 .4 8 5 5 .2 1 – 1 0 C 1 2 7 4 6 .1 3 – 6 0 .3 3 0 .4 0 0 .4 0 0 .0 4 • • • × • • × • • × • • × × × • × • − + 3 0 .5 8 µ � 2 � C 1 4 8 4 9 .5 8 – 9 0 .5 0 0 .4 3 0 .0 8 0 .0 7 × • • • × × × × • • × × × × × × × • − − I Co n tr a ct o r 4 2 .1 2 1 2 .0 4 9 0 .5 6 0 .4 1 0 .1 6 0 .5 0 × • × • × × • • × × × • × × × • − µ Co n tr a ct o r 4 8 .5 4 – 9 0 .5 0 0 .4 2 0 .2 0 0 .3 1 × • • • × × × × • × × × × × × × × × − − F ig u re 5 .7 : T h e ta b le o f d is ta n ce s fo r N ar ic h P ty L td v . C om m is si on er o f P ay -r ol l T ax ( S H Y S T E R o u tp u t) . T h e n ea re st n ei gh b ou r is F er gu so n v . Jo hn D aw so n & P ar tn er s (C on tr ac to rs ) L td ( C 5 ), s o th e n ea re st r es u lt i s E m p lo ye e; t h e n ea re st ot h er i s M as se y v. C ro w n L if e In su r a n ce C o. ( C 1 2 ). N ot e th at t h e sp ec ifi ed d ir ec ti on s (⊃ ) su gg es t C o n tr a ct o r. 1 5 .3 7 C 1 3 × • × × • × × × × • • × 7 2 8 .5 4 5 4 .9 7 5 0 .4 2 0 .4 7 0 .1 2 − 0. 1 2 � � � � � � � 5.4 Employees and independent contractors 163 Stevens v. Brodribb Sawmilling Company Pty Ltd 1986 In Stevens v. Brodribb Sawmilling Company Pty Ltd, 93 Brodribb employed snig­ gers to fell trees, and truck drivers to carry the trees to the sawmill. They used their own vehicles, set their own hours of work, and were paid according to the volume of timber that they delivered to the sawmill. Stevens, a truck driver, was injured due to the negligence of a snigger. He claimed that the snigger was an em­ ployee of Brodribb and, hence, Brodribb was vicariously liable for the snigger’s negligence. He also claimed that he was an employee of Brodribb and, hence, Brodribb was personally liable to him for breaching the duty of care owed by an employee to another employee. Five judges of the High Court of Australia applied Humberstone v. Northern Timber Mills and Australian Mutual Provident Society v. Chaplin, and decided that neither the truck driver nor the snigger were Brodribb’s employees.94 SHYSTER’s fact vector is (NYYYNUNYYNNNNNNNNN). SHYSTER agrees with the High Court that Stevens was an independent contractor. The nearest neighbour is Australian Mutual Provident Society v. Chaplin: one of the cases that the High Court applied. But none of their honours referred to SHYSTER’s nearest other: Cam and Sons Pty Ltd v. Sargent. SHYSTER issues no warnings. Re Porter; Re Transport Workers Union of Australia 1989 In Re Porter; Re Transport Workers Union of Australia, 95 five truck drivers, who owned their own trucks, had nominated for various offices in the Victorian branch of the TWU. The returning officer rejected their nominations on the grounds that only members who were employees were eligible for election, and they were independent contractors. The drivers challenged the returning officer’s decision. In the Federal Court, the TWU relied heavily on four Australian cases (only one of which is part of the Employee specification) in which owner-drivers of trucks had been held not to be employees of the company for whom they worked: Humberstone v. Northern Timber Mills, Wright v. Attorney-General for Tasmania, 96 Barro Group Pty Ltd v. Fraser 97 and Stevens v. Brodribb Sawmilling Company Pty Ltd. The union also relied on the English case of Ready Mixed Con­ crete (South East) Ltd v. Minister of Pensions and National Insurance. Gray J said that, despite the consistency of these cases, “a balancing exercise is always involved in the determination whether an employment relationship exists.”98 On balance, he held, the five drivers were employees. � � � 164 Chapter 5: Case studies The complete report file for Re Porter; Re TWU is given as an example in §B.4. SHYSTER’s fact vector is (YNYYUNUNNYNYNNYNNY). 99 SHYSTER agrees with Gray J that the drivers were employees. The nearest neighbour is Ferguson v. John Dawson & Partners (Contractors) Ltd, which was not mentioned in the case. The nearest other is Ready Mixed v. Minister : one of the cases cited by the TWU. SHYSTER warns that one of the instantiations has a different result to that of the instant case. If the employer would not make a profit/loss when the work performed by the drivers cost less/more than expected (A5), and the em­ ployer neither supervised nor inspected the drivers’ work (A7)—attributes with unknown values in the instant case—then SHYSTER’s opinion would be that the drivers were independent contractors (following Ready Mixed v. Minister ). The legal expert confirms that this is a sensible distinction to make. Building Workers’ Industrial Union of Australia v. Odco Pty Ltd 1991 In Building Workers’ Industrial Union of Australia v. Odco Pty Ltd, 1 Odco provided labour to builders in Victoria.2 In the agreement between Odco and the workers it was stated that the workers were contractors, not employees. They undertook to work for an agreed hourly rate, regardless of any industrial awards which might apply. Odco encountered “hostility from within the trade union movement, especially amongst officials of building unions”3 and “incidents” oc­ curred which led Odco to bring an action against the union for breach of s. 45d of the Trade Practices Act 1974 (Cth). The question of whether the workers were employees or independent contractors of Odco was relevant to establishing a defence for the union under the Act. The case was heard by the Federal Court. Wilcox, Burchett and Ryan JJ ap­ plied Stevens v. Brodribb Sawmilling Company Pty Ltd and held that the workers were not employees of Odco.4 The fact that parties to an agreement label workers as employees or as independent contractors is not decisive—their honours cited Ferguson v. John Dawson & Partners (Contractors) Ltd 5—however they held that, in this case, the “contractor” label was correct. BWIU v. Odco is used as the test case in the complete example in appendix C. SHYSTER’s fact vector is (NYNYNNNUNNYUNNNNNY). 6 The distances file for BWIU v. Odco is given in full in §C.11. The report file is in §C.13. SHYSTER’s conclusion is that the workers were independent contractors. Its choice of cases is good. The nearest neighbour is Humberstone v. Northern Timber Mills (C8), which was applied by the High Court in Stevens v. Brodribb which was � � � 165 � 5.4 Employees and independent contractors applied by the Federal Court in BWIU v. Odco (Stevens v. Brodribb is not one of the leading cases in the Employee specification). The nearest other is Ferguson v. John Dawson & Partners (Contractors) Ltd (C5), which their honours cited in their judgment. No warnings are issued. Hypothetical case 3 Both hypothetical case 3 and case 4 are taken from a question set in an examin­ ation for a university employment law course.7 A secretary, Amber, is employed in the Department of Legal Studies in an Australian university. The written con­ tract between the university and Amber provides that salary, hours, allowances, and leave shall be in accordance with the terms of a specified award. Both the university and Amber contribute to a superannuation scheme for Amber’s bene­ fit. Amber’s appointment is to continue until she is 65 years of age, after an initial probationary period. There is a duty statement for Amber’s position, and she performs duties as required by the Head of the Department. It is clear, and the legal expert confirms, that Amber is an employee of the university. SHYSTER’s fact vector is (YNYNYYYNNUYNNYYYYN); SHYSTER agrees that Amber is an employee. The nearest neighbour is Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd ; the nearest other is Massey v. Crown Life Insurance Co. Neither is a good choice. SHYSTER warns that the weighted correlation coefficients suggest that Steven­ son Jordon and Harrison Ltd v. Macdonald and Evans (1), in which the worker was an independent contractor, should be the nearest neighbour. Hypothetical case 4 From the same examination question as hypothetical case 3 comes another hy­ pothetical case. Bonny works as a secretary in the same department as Amber. Bonny is registered with an employment agency. The Department and the employment agency entered into a contract in which the agency agreed to provide a competent secretary for a three month period, and retained the right to provide a substitute at any time. The university indemnified the employment agency for any liability arising out of Bonny’s work. When Bonny registered with the employment agency she signed a contract which stated that any work which she procured through the agency she would perform as an independent contractor. Bonny is paid a specified daily rate, is not eligible for any leave, and is responsible for injury insurance and taxation liabilities. Bonny has had her work period extended several times, and has worked � � � 166 Chapter 5: Case studies in the Department of Legal Studies for a year. Bonny and Amber have similar duties and patterns of work. They both follow guidelines issued by the Head of the Department. The legal expert advises that, unlike Amber and on the basis of BWIU v. Odco, Bonny is an independent contractor. SHYSTER’s fact vector is (YNYNYYYNNNYYNNNNNY); SHYSTER disagrees with the expert, and concludes that Bonny is an employee of the university. The nearest neighbour is Ferguson v. John Dawson & Partners (Contractors) Ltd. As in this hypothetical case, both the employer and the worker in Ferguson v. Dawson agreed that the worker was an independent contractor.8 In fact the only differ­ ence that SHYSTER sees between Ferguson v. Dawson and hypothetical case 4 is that in this case the work was performed on the employer’s premises. Neverthe­ less, SHYSTER’s choice of nearest neighbour is a bad one. The nearest other is Massey v. Crown Life Insurance Co., a good choice. In Massey v. Crown Life, Massey was an employee for two years, then performed the same duties for an­ other two years as an independent contractor (for tax purposes). The relationship between Massey as employee and Massey as independent contractor, is similar to that between Amber and Bonny. SHYSTER issues no warnings. The legal expert advises that one of the aims of this examination question was to demonstrate the absurdity of the law in this area: Amber and Bonny perform very similar jobs, yet the law sees them as having quite different relationships with the university. Hypothetical case 4 is an example of a tripartite employment relationship: a relationship where one party pays an employment agency which pays the worker. BWIU v. Odco was the first case in which a court had to determine the employment status of workers in such relationships. It was decided in 1991 and so is not part of the Employee specification. SHYSTER’s failure in this hypothetical case illustrates the unsuitability of the Employee specification for providing advice on such tripartite relationships. 5.4.4 Generated tests The legal expert provided three examples for generating tests using the Employee specification: generated tests 4, 5 and 6. The results of these tests are summarized in figure 5.13. Generated test 4 The legal expert advises that, if the following is true, the worker is an employee: . . . the employer directed the manner in which the work was to be done; the worker was an integral part of the employer’s business; the employer would make a profit/loss if the work performed by the worker cost less/more than 167 � 5.4 Employees and independent contractors expected; the employer supervised or inspected the work; the worker was not allowed to employ others to assist with her/his work; the money that the employer paid to the worker was stated to be “wages” or “salary”; the employer deducted PAYE tax instalments from the worker’s pay; and the employer paid the worker sick pay or holiday pay. SHYSTER’s fact vector is (YUYUYUYUNUUUUYYYUU). It generates 1024 instanti­ ations: a mere 1/256th of the search space. However, it is not possible that the money that the employer paid to the worker was stated to be both a “fee” and “wages.” This paradox occurs whenever the values of both A13 and A14 are yes. Because A14 is always yes, setting A13 to no avoids these impossible cases and halves the size of the search space. In 149 of the 512 instantiations in this reduced search space, SHYSTER’s choice of result is bad (i.e. Contractor): a success rate of 70.90%. In all 149 the nearest ideal point is that of the expected result; and in all but 4 of those 149, SHYSTER warns that that ideal point is at least as near to the instant case as is the nearest neighbour. There is no (originally unknown) attribute for which these 149 instantiations have common values. However, in 131 (or 87.92%) of them the worker was in business on her/his own account. The legal expert advises that although this (in combination with the originally specified attribute values) does not present a paradox, it is an almost inconceivable set of attributes. Generated test 5 The legal expert also advises that, if the following is true, the worker is an independent contractor: . . . the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the worker was in business on her/his own account; the worker was allowed to employ others to assist with her/his work; the employer did not deduct PAYE tax instalments from the worker’s pay; and the employer paid the worker neither sick pay nor holiday pay. Given the fact vector (UUUUNUUYYUUUUUNNUU), SHYSTER generates 8192 instan­ tiations: 1/32nd of the search space. Once again, a paradox occurs whenever the values of both A13 and A14 are yes. Because both attributes are unknown, removing these impossible cases reduces the size of the search space by a quarter. In 1946 of the 6144 instantiations in this reduced search space, SHYSTER’s choice of result is bad (i.e. Employee): a success rate of 68.33%. 168 Chapter 5: Case studies Despite the fact that the nearest ideal point is that of the expected result in 1353 of those 1946, only 492 of them cause SHYSTER to issue a warning. Furthermore, in 111 instantiations SHYSTER issues an ideal point warning, even though the choice of result is good. Generated test 6 The legal expert further advises that, if the following is true, the worker is an independent contractor: . . . the worker was allowed to use her/his own discretion in doing an as­ pect of the work that was not specified beforehand; the worker owned the tools or provided the transport with which she/he performed the work; the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the worker was in business on her/his own account; the money that the employer paid to the worker was stated to be a “fee”; the employer did not deduct PAYE tax instalments from the worker’s pay; and the employer paid the worker neither sick pay nor holiday pay. SHYSTER’s fact vector is (UYUYNUUYUUUUYUNNUU). It generates 2048 instanti­ ations. As with generated test 4, removing the paradoxical cases halves the size of the search space. In 149 of the 1024 instantiations in this reduced search space, SHYSTER’s choice of result is bad (i.e. Employee): a success rate of 85.45%. In all 149 the nearest ideal point is that of the expected result; and in all but 37 of these 149, SHYSTER warns that that ideal point is at least as near to the instant case as is the nearest neighbour. The fact that, in 120 (or 80.54%) of these 149, the work was performed on the employer’s premises, is not enough to justify SHYSTER’s choice of result in these instantiations. 5.4.5 Conclusion In all but one of the tests described in §5.4.3 above (i.e. 83.34%) SHYSTER reached a good conclusion, but in only 41.67% of those tests did it choose good cases to use in argument. As shown in figure 5.13, SHYSTER’s choice of result is good in only 70.78% of the instantiations in all three generated tests. (This takes into account the reduction of the size of the search space due to the removal of instantiations where the values of both A13 and A14 are yes. Such instantiations do not represent plausible cases.9) � � � � � � 169 � 5.4 Employees and independent contractors This performance is dominated by generated test 5:10 the test which produced, from SHYSTER, the worst performance of all the generated tests described in this chapter. Alone amongst all the generated tests, this test causes SHYSTER to issue ideal point warnings for instantiations in which its choice of result is good. If a bad choice of result is considered good where an ideal point warning is issued and a good choice of result is considered bad where an ideal point warning is issued, SHYSTER’s success rate for the three generated tests of the Employee specification rises to 99.22%, 74.53% and 96.39%, respectively: 79.09% overall. There are two tripartite employment cases amongst the test cases: BWIU v. Odco and hypothetical case 4. In both of these tests, it is important to be clear as to who is the “employer” for the purposes of answering SHYSTER’s questions: in BWIU v. Odco it is Odco (the employment agency); in hypothetical case 4 it is the university. Although SHYSTER’s choices of result and cases for BWIU v. Odco are good, the ability of the Employee specification to handle tripartite employment relationships is stretched by hypothetical case 4. The Employee area also demonstrates one of SHYSTER’s minor shortcomings. In its report, SHYSTER’s summary of the facts in a case can make difficult read­ ing when, as here, there is a large number of attributes. The wording of the strings for the last two attributes means, for example, that the statement of facts in SHYSTER’s report on Stevens v. Brodribb Sawmilling Company Pty Ltd is a 253-word sentence that concludes: . . . the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker did not express any intention that the relationship would be one of principal and independent contractor. It is to be hoped that no human legal expert would express the facts so clum­ sily. However, this problem is inherent in SHYSTER’s method of writing reports; possible solutions are discussed in §6.2.5. The Employee area could be linked to several statutory open-textured concepts, as represented in a rule-based system. It could also be linked to areas (like the Authorization area) in which the distinction between employees and independent contractors is important. 170 Chapter 5: Case studies SHYSTER (when using the Employee specification) is capable of providing good advice in cases involving bipartite employment relationships (as it was de­ signed to do). It is not so well suited to cases involving tripartite relationships. The reflexive tests performed in §D.4, and summarized in figure D.3, indicate that the Employee specification is capable of handling new fact situations. The specification chooses good results in most of the generated tests. It per­ forms fairly well in an area of law in which, according to Clark and Wedderburn, the juridical concepts “approach anarchy.”11 5.5 The implication of natural justice The last specification discussed in this chapter concerns the implication of natural justice principles in administrative decision-making. 5.5.1 The law The law in Australia has developed the notion of affording natural justice to a person who is affected by an administrative decision. Natural justice principles might require, for example, that the person affected be given the right to a hearing—to make representations—before the decision is made. Or the decision- maker might be obliged to give reasons for the decision, after it has been made. These requirements are sometimes called requirements of “procedural fairness.” The development of this area of law has been convoluted: The evolution of the doctrine of natural justice has been a journey in fluctuation, typified by the regular pitch of basic principles from one leaning to another.12 In Kioa v. West, Mason J said: The law has now developed to a point where it may be accepted that there is a common law duty to act fairly, in the sense of according procedural fairness, in the making of administrative decisions which affect rights, in­ terests and legitimate expectations, subject only to the clear manifestation of a contrary statutory intention.13 But McMillan asserts that: The main residual difficulty concerning the implication of natural justice has been the tendency to apply uncritically [this] superficial test . . . the affect [sic] of a decision on a person’s interest is only one of many factors that is relevant to deciding whether a natural justice obligation attaches to the decision.14 � � � 171 � 5.5 The implication of natural justice The notion of a “legitimate expectation”—or a “reasonable expectation”15— expands the area considerably: it has been judicially stated that legitimate ex­ pectations “are capable of including expectations which go beyond enforceable legal rights, provided they have some reasonable basis”. 16 McMillan discusses different ways of determining whether natural justice prin­ ciples apply to a given administrative procedure: An option at one end of the spectrum has always been to treat the oblig­ ation of procedural fairness as attaching universally to all decisions, and to concentrate instead on the practical content of that obligation in any particular instance. This option . . . was described by Deane J in Haoucher as “conceptually more satisfying”.17 However, that approach is too broad. As Wilcox J points out in Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd: . . . the law has not yet reached the stage of applying the obligation of natural justice to every decision which disadvantages individuals.18 McMillan proposes an alternative: . . . a preferable approach would be to determine whether natural justice applied by examining in each case a number of different factors. While the weight attached to some factors is clearly greater, it is the overall balance that is important. This approach was propounded by the Privy Council in Durayappah v. Fernando, 19 was initially applied in Australia in Salemi v. MacKellar (No. 2)20 and FAI Insurances Ltd v. Winneke, 21 but has not been developed further.22 His factors are: • the nature of the property, right, interest, status, or reasonable expectation; • the effect or impact of the decision; • the nature of the power being exercised; • the statutory and factual criteria according to which the decision was made; • the nature of the officer making the decision; • the statutory procedural framework under which the decision was made; and the circumstances in which the decision was made.23 • � � � 172 Chapter 5: Case studies McMillan discusses these factors in detail. It is worth discussing here the law relating to one of these factors as it significantly affects one of the test cases in §5.5.3 below.24 The statutory procedural framework under which the decision was made may grant a person a right to appeal against a decision. The existence of a statutory right to appeal used to suggest that natural justice would not be implied. For example, in Twist v. Council of Municipality of Randwick, Mason J stated: Having regard to the subject matter of the section, . . . and more particu­ larly the comprehensive nature of the appeal to a District Court judge, I am of opinion that s. 317b(5) [granting the right of appeal] should be read as providing the exclusive remedy available to an owner who wished to challenge the validity or correctness of an order made under s. 317b(1).25 And, in the first Marine Hull & Liability Insurance Co. Ltd v. Hurford, Wilcox J held that the Treasurer should have applied the principles of natural justice before making the contested direction—however, because the relevant Act provided for a review of any such directions by the Administrative Appeals Tribunal: . . . the legislature must be taken to have evinced an intention that, in the event of the Treasurer failing to so act, the directions are not to be regarded as being invalid in law. They are merely susceptible of challenge before the Tribunal.26 Recently, there has been a tendency to interpret the existence of a statutory right to appeal as indicating that natural justice should be implied. In the second Marine Hull & Liability Insurance Co. Ltd v. Hurford, for example, Davies J stated that: The existence of a right to have a matter reconsidered and of a right to have a matter reviewed by the Administrative Appeals Tribunal may well affect the nature of the procedures which ought to be adopted in complying with the rules of natural justice but, ordinarily, it does not exclude them.27 And in Bropho v. Western Australia, Rowland J held that the existence of a right of appeal to the Supreme Court indicated that the decision-maker should be influenced by the same considerations as would a judge of that Court; it indicated that questions of natural justice should not be excluded.28 Rowland J’s approach is diametrically opposed to that of the High Court in Twist v. Randwick, decided fourteen years earlier. His decision was overturned on appeal by the Full Court of the Supreme Court of Western Australia in Western Australia v. Bropho29 (another of the test cases in §5.5.3 below). So, although the law may be evolving towards Rowland J’s interpretation, it has not yet reached that stage. 173 � 5.5 The implication of natural justice Natural area: A1 A2 A3 A4 A5 A6 A7 ! ! ! !! � � � a a a � � � �� Affected area: A1 A2 A3 A4 �� �� �� �! ! ! !! � � � a a a � � � �� ������� Expectation area: A1 A2 A3 A4 A5 A6 Figure 5.8: The relationship between the attributes in the three areas which make up the Natural specification. 5.5.2 The NATURAL specification The author—having consulted a legal expert,30 and the published judgments in the leading cases—wrote a specification of the implication of the duty to observe natural justice. The specification is based on the factors identified by McMillan.31 The dump file for the Natural specification is given in full in §A.5. It contains three areas: the Natural area, the Affected area and the Expectation area. The relationship between the attributes in these areas is indicated in figure 5.8. There is a hierarchy which ranks nine courts, and the Affected and Expectation area both have opening strings. Natural area The Natural area has two results: Implied and Not-Implied. These correspond to “a duty to observe natural justice is implied” and “a duty to observe natural justice is not implied.” There are seven attributes in the Natural area, corresponding to the seven factors identified by McMillan. The term applicant is used to refer to the person who seeks the implication of natural justice: A1: Did the decision affect the property, right, interest, status, or legitim­ ate expectation of the applicant? A2: Is the decision apt to have a discrete impact on the interests of the applicant? A3: Is the power of a nature that would suggest that procedural fairness would be applied? A4: Did the statutory or factual criteria focus on matters which were dis­ crete to the interests of the applicant? A5: Was the decision-maker a high-level policy-maker? 174 Chapter 5: Case studies A6: Is there a statutory right to appeal against the decision?32 A7: Were there circumstances which make an obligation to observe natural justice inappropriate? Only A6 is straightforward. But (with the exception of A1) answering the other questions—with the assistance of the help strings which are included in the specification—should present no major difficulties. For example, the help string for A4 is: The decisional criteria are of two kinds: the spectrum of considerations to which the decision maker was authorized to have regard (the statutory criteria), and the specific considerations to which regard was had in fact (the factual criteria). Either set of criteria can focus on matters which are discrete to the interests of the applicant, or on matters of policy or public interest. And the help string for A5 explains that: Ministers, members of Cabinet, Governors and the Governor-General are high-level policy-makers. However, A1 is such a difficult question to answer that it is defined as an ex­ ternal attribute, linked to the Affected area. The user is never asked the A1 ques­ tion, but is asked questions from the Affected area instead. Affected area The Affected area has two results: Affected and Unaffected. These correspond to “the decision affected the property, right, interest, status, or legitimate expecta­ tion of the applicant” and “the decision did not affect [those interests].” A result of Affected produces a value of yes for A1 in the Natural area; a result of Unaffected produces a value of no. There are four attributes in the Affected area: A1: Did the decision affect a financial, property or occupational interest of the applicant? A2: Did the decision affect the applicant’s personal liberty? A3: Did the decision affect the applicant’s reputation? A4: Did the applicant have a legitimate expectation which was affected by the decision? Only A4 is not a straightforward question. It, too, is defined as an external attribute, and linked to the Expectation area. 175 � 5.5 The implication of natural justice Expectation area The Expectation area also has two results: Expectation and No-Expectation. These correspond to “the applicant had a legitimate expectation which was affected by the decision” and “the applicant did not have a legitimate expectation which was affected by the decision.” A result of Expectation produces a value of yes for A4 in the Affected area; a result of No-Expectation produces a value of no. There are six attributes in the Expectation area: A1: Did the decision-maker break a promise or undertaking? A2: Did the decision-maker go against an established course of practice? A3: Did the decision involve a refusal to renew an existing interest? A4: Did the decision-maker or a statutory provision suggest that an initial interest would be granted? A5: Did the decision affect an established liberty or interest? A6: Was there a standard administrative procedure which the decision- maker did not follow? There is a subtle distinction between A1, A2 and A6 which is best illustrated by two examples. In Haoucher v. Minister of State for Immigration and Ethnic Affairs, 33 the Minister had told Parliament that recommendations of the Administrative Ap­ peals Tribunal, on deportation matters, would be overturned only in “exceptional circumstances” and when justified by “strong evidence.” The AAT recommended that the Minister’s order to deport Haoucher be revoked, but the Minister re­ jected the recommendation, without giving Haoucher details of any “exceptional circumstances” or “strong evidence.” The value of A1 is no: there may indeed have been “exceptional circum­ stances”—the Minister did not promise to give details of those circumstances. A2 is yes: the Minister had an established course of practice of following AAT recommendations. However, A6 is no: the Minister went against practice, not procedure. In Attorney-General of Hong Kong v. Ng Yuen Shiu, 34 an immigration official had announced that certain illegal immigrants would be interviewed, and each case “treated on its merits.” Shiu came forward the next day and was questioned, but given no opportunity to put the merits of his case. Although the decision-maker broke a promise (A1 = yes), there had not been time to establish a course of practice (A2 = no), and there was no standard procedure to be followed (A6 = no). � � � 176 Chapter 5: Case studies Leading cases There are fifteen leading cases in the Natural area. Seven of them were decided by the High Court: FAI Insurances Ltd v. Winneke, 35 Haoucher v. Minister of State for Immigration and Ethnic Affairs, 36 Annetts v. McCann, 37 Kioa v. West, 38 Commissioner of Police v. Tanos, 39 South Australia v. O’Shea40 and Bread Man­ ufacturers of New South Wales v. Evans. 41 Two were decided by the Federal Court of Australia: Marine Hull & Liability Insurance Co. Ltd v. Hurford 42 and Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd. 43 Two were decided by the Supreme Court of New South Wales: Macrae v. Attorney-General for New South Wales44 and Nashua Australia Pty Ltd v. Channon. 45 Two are decisions of the Judicial Committee of the Privy Council: Attorney-General of Hong Kong v. Ng Yuen Shiu46 and Durayappah v. Fernando. 47 One was decided by the House of Lords: Council of Civil Service Unions v. Minister for the Civil Service. 48 And one was a decision of the Chancery Division of the English Court of Appeal: McInnes v. Onslow Fane. 49 Nine of these are also leading cases in the Affected area. And six of them, plus two further cases—Heatley v. Tasmanian Racing and Gaming Commission50 and Cole v. Cunningham51—are used in the Expectation area. For reasons discussed in §5.5.3 below, the prominent High Court cases of Twist v. Council of Municip­ ality of Randwick 52 and Salemi v. MacKellar 53 were not used in the Natural specification. A peculiarity of the Natural specification is that, as SHYSTER warns, there are three pairs of cases in the Natural area which have identical facts and different results.54 The existence of two or more leading cases with identical facts but different results indicates either that one of the cases was wrongly decided, or that there is a need for more attributes in the area (to distinguish between the cases). All of the leading cases in the Natural area were correctly decided, and the attributes in that area are based on the factors identified as being important by McMillan. Hence, no change is made to the specification. If two or more of these cases with identical facts but different results are the nearest cases, SHYSTER warns the user and chooses the most important of them as the case upon which to base its conclusion. Attribute dependence The probabilities matrices for the three areas in the Natural specification, as extracted from the probabilities file, are given in figure 5.9. SHYSTER detects no functional dependence, and no evidence of stochastic dependence, between any of the attributes in any of the three areas. 177 � 5.5 The implication of natural justice A2 A3 A4 A5 A6 A7 A1 A2 A3 A4 A5 A6 1.00 1.00 1.00 0.73 1.00 1.00 0.13 0.47 0.57 1.00 0.87 0.93 1.00 1.00 0.94 1.00 1.00 0.20 0.31 0.48 0.74 0.87 0.95 0.34 1.00 0.47 0.30 0.95 0.27 1.00 0.17 0.84 0.57 0.99 0.69 1.00 0.48 1.00 0.94 0.73 0.87 1.00 A2 A3 A4 A1 A2 A3 0.12 0.95 0.83 1.00 0.40 0.64 0.24 0.95 1.00 0.40 0.60 0.88 A2 A3 A4 A5 A6 A1 A2 A3 A4 A5 0.36 0.75 1.00 0.46 0.75 1.00 1.00 1.00 0.96 1.00 1.00 1.00 1.00 0.63 0.38 1.00 0.36 1.00 1.00 1.00 0.88 1.00 0.75 1.00 1.00 1.00 1.00 1.00 1.00 0.75 Figure 5.9: The probabilities matrices for the three areas in the Natural spe­ cification: the Natural area, the Affected area and the Expectation area (SHYSTER output). � � � � � � � � � � � � � � � 178 Chapter 5: Case studies Implied Not-Implied Attr. µ �2 w µ �2 w µ �2 w A1 1.00 0.00 0.83 0.14 7.20 0.93 0.06 16.07 1.00 0.00 0.67 0.22 4.50 0.87 0.12 8.65A2 0.67 0.22 4.50 0.33 0.22 4.50 0.53 0.25 4.02A3 0.63 0.23 4.27 0.17 0.14 7.20 0.43 0.24 4.08A4 0.78 0.17 5.79 0.67 0.22 4.50 0.73 0.20 5.11A5 0.22 0.17 5.79 0.00 0.00 0.13 0.12 8.65A6 0.00 0.00 0.17 0.14 7.20 0.07 0.06 16.07 A7 Affected Unaffected Attr. µ �2 w µ �2 w µ �2 w A1 0.50 0.25 4.00 0.00 0.00 0.44 0.25 4.05 0.38 0.23 4.27 0.00 0.00 0.33 0.22 4.50 A2 0.38 0.23 4.27 0.00 0.00 0.33 0.22 4.50 A3 0.50 0.25 4.00 0.00 0.00 0.44 0.25 4.05 A4 Expectation No-Expectation Attr. µ �2 w µ �2 w µ �2 w A1 0.33 0.22 4.50 0.00 0.00 0.25 0.19 5.33 0.50 0.25 4.00 0.00 0.00 0.38 0.23 4.27 A2 0.17 0.14 7.20 0.00 0.00 0.13 0.11 9.14 A3 0.00 0.00 0.00 0.00 0.00 0.00A4 0.83 0.14 7.20 0.50 0.25 4.00 0.75 0.19 5.33 A5 0.17 0.14 7.20 0.00 0.00 0.13 0.11 9.14 A6 Figure 5.10: The tables of weights for the three areas in the Natural spe­ cification: the Natural area, the Affected area and the Expectation area (SHYSTER output). Weights The tables of weights for the three areas in the Natural specification, as extrac­ ted from the weights file, are given in figure 5.10. In the Natural area, the rightmost column indicates that two attributes are of equal greatest importance:55 A1: Did the decision affect the property, right, interest, status, or legitim­ ate expectation of the applicant? A7: Were there circumstances which make an obligation to observe natural justice inappropriate? � 5.5 The implication of natural justice 179 SHYSTER’s weighting of attributes in this area is fairly good. A1 and A7 are very important attributes. Indeed if the value of A1 is no, then natural justice should never be implied.56 However, the legal expert advises that A2 should be weighted more heavily than it is. All four attributes in the Affected area are weighted (almost) the same.57 This is appropriate—each of these attributes is equally indicative of the applicant having been affected. One of the attributes in the Expectation area is infinitely weighted: A4: Did the decision-maker or a statutory provision suggest that an initial interest would be granted? This is an important attribute, but its infinite weighting probably weights it too highly in relation to the other attributes in this area.58 5.5.3 Test cases Eight cases are used to test the Natural specification—seven of which were considered for the specification, but deemed not to be leading cases. The last case, Ainsworth v. Criminal Justice Commission, 59 is a very recently reported decision of the High Court of Australia. The results of these tests are summarized in figure 5.12. Twist v. Council of Municipality of Randwick 1976 In Twist v. Council of Municipality of Randwick, 60 Twist owned land in Rand- wick. The Council resolved, under s. 317b(1) of the Local Government Act 1919 (NSW), to demolish a building on his land. Twist was given sixty days to demolish the building, and then an extension of another three months. Section 317b(5) gave Twist a right of appeal to the District Court against the demolition order, but the Court’s rules required that such appeals be made within sixty days. Twist did not appeal in time, and was not allowed an extension of time. When the Council informed Twist that it would demolish the building, he appealed to the New South Wales Supreme Court, and then to the High Court. Twist claimed that the Council was bound by the rules of natural justice to have given him an opportunity to be heard before the demolition order was made. He relied on Ridge v. Baldwin, 61 a decision of the House of Lords. The Council relied on the nineteenth century English case of Vestry of St James and St John, Clerkenwell v. Feary. 62 (Neither decision is part of the Natural specification.) � � � 180 Chapter 5: Case studies Three judges of the High Court unanimously found against Twist. Barwick CJ and Mason J held that the existence of a right of appeal showed that Twist had no right to be heard before the Council made its demolition order.63 Jacobs J disagreed, but held that the existence of a right of appeal did show that the Council’s failure to allow a hearing did not render its decision void.64 SHYSTER’s fact vectors are (NNNNYN), (YNNN) and (YYYYNYN)—for the Expecta­ tion, Affected and Natural areas, respectively. To quote its report files: . . . the decision-maker did not break a promise or undertaking; the decision- maker did not go against an established course of practice; the decision did not involve a refusal to renew an existing interest; neither the decision- maker nor a statutory provision suggested that an initial interest would be granted; the decision affected an established liberty or interest; and there was no standard administrative procedure which the decision-maker should have followed. . . . the decision affected a financial, property or occupational interest of the applicant; the decision did not affect the applicant’s personal liberty; the decision did not affect the applicant’s reputation; and the applicant did not have a legitimate expectation which was affected by the decision. . . . the decision affected the property, right, interest, status, or legitimate expectation of the applicant; the decision is apt to have a discrete impact on the interests of the applicant; the power is of a nature that would suggest that procedural fairness would be applied; the statutory or factual criteria focused on matters which were discrete to the interests of the applicant; the decision-maker was not a high-level policy-maker; there is a statutory right to appeal against the decision; and there were no circumstances which would have made an obligation to observe natural justice inappropriate. In the Expectation area, SHYSTER concludes (as did the High Court) that Twist did not have a legitimate expectation which was affected by the decision. The nearest neighbour is South Australia v. O’Shea; the nearest others are Haoucher v. Minister of State for Immigration and Ethnic Affairs and Heatley v. Tasmanian Racing and Gaming Commission. SHYSTER warns that the specified directions suggest Expectation. In the Affected area, SHYSTER concludes that the decision affected Twist’s property, right, interest, status, or legitimate expectation. The nearest neigh­ bour is Bread Manufacturers of New South Wales v. Evans ; the nearest other is Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd. However, SHYSTER’s conclusion in the Natural area is that a duty to observe natural justice is implied. Its table of distances, as extracted from the appropriate distances file, is given in figure 5.11. The nearest neighbour is Commissioner of Police v. Tanos (C5), one of the cases to which Mason J referred in coming to his decision; the nearest other is McInnes v. Onslow Fane (C15). • • • � 181 C as e C In st a n t C 1 C 2 C 3 C 4 C 5 C 6 A tt ri bu te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 • • • • × • × • • × • • × × 1 • • × • • × × 2 • • • • × × × 2 • • • • • × × 2 • • • • × • × 3 d K d U 17 .7 9 17 .7 9 8. 65 13 .7 7 – S S � – 3 0. 43 0 .2 8 – 3 0. 43 0 .2 8 – 1 0. 14 0 .1 4 – 2 0. 29 0 .2 2 – – – – r r � 0. 09 0. 73 0. 09 0. 73 0. 73 0. 84 0. 30 0. 74 1. 00 1. 00 R es ul t Im p li ed 2 � + 14 .5 5 3 � + 14 .5 5 � 5.5 The implication of natural justice C 7 • • • × • • × 4 17 .8 5 – 3 0. 43 0 .2 8 0. 09 0. 73 • • • × • × × 5 • × × 7 – 4 0. 57 0 .3 5 – 1 0. 14 0 .1 4 – 2 0. 29 0 .2 2 – 3 0. 43 0 .2 8 – 4 0. 57 0 .4 1 – 6 0. 86 0 .7 4 9. 20 – 2 0. 29 0 .1 5 0. 30 0. 89 ⊃ 8 .7 7 C 8 13 .7 7 4. 08 2 0. 33 0 .2 4 0. 25 0. 74 C 9 I I m p li e d µ Im p li e d C 1 0 C 1 1 C 1 2 C 1 3 C 1 4 • • × × • × × 7 • • • • × × × • • • • • × × • • • × • × × 2 • × × × × × × 2 × × × × • × × 4 • • × × • × × 6 8 • • × × • × • I 0. 55 0. 25 µ I ⊃⊃ − 0. 09 0. 73 0. 73 0. 84 0. 40 0. 84 0. 09 0. 73 0. 26 0. 78 − 0. 65 − 0. 46 N o t- Im p li ed − 0. 09 0. 73 ⊃ � − ⊃ � C 1 5 21 .8 7 8. 65 13 .7 7 17 .8 5 25 .4 1 46 .5 9 21 .8 7 – 4 0. 57 0 .3 5 37 .9 4 – 5 0. 71 0 .6 1 8. 65 – 1 0. 14 0 .1 4 • • • • × × × 9 I N o t- Im p li e d 54 .0 1 – 6 0. 86 0 .8 6 0. 73 − 0. 48 × × × × • • • 21 .8 7 – 4 0. 57 0 .3 5 − 0. 03 0. 71 µ No t- Im p li e d • • × × • × × − F ig u re 5 .1 1 : T h e ta b le o f d is ta n ce s fo r T w is t v. C ou n ci l of M un ic ip al it y of R an dw ic k in t h e N a tu ra l ar ea ( S H Y S T E R o u tp u t) . T h e n ea re st n ei gh b ou r is C om m is si on er o f P ol ic e v. T an os ( C 5 ), s o th e n ea re st re su lt i s Im p li ed ; th e n ea re st o th er i s M cI n n es v . O n sl ow F an e (C 1 5 ). 0. 73 0. 84 � � � 182 Chapter 5: Case studies This conclusion is at odds with the decision of the High Court. However, as discussed in §5.5.1 above, the significance of the existence of a statutory right to appeal has changed over the years since Twist v. Randwick. The author respect­ fully submits (and the legal expert agrees) that Twist v. Randwick was wrongly decided, at least insofar as it concerns the significance of a statutory right to appeal. The case was not included in the Natural specification for this reason. Apart from Commissioner of Police v. Tanos, all of SHYSTER’s chosen cases were decided after Twist v. Randwick. Salemi v. MacKellar 1977 In Salemi v. MacKellar, 65 Salemi was an Italian citizen who entered Australia under a temporary entry permit. He stayed in the country after the permit ex­ pired, and became a “prohibited immigrant” within the meaning of the Migration Act 1958 (Cth). The Minister of Immigration announced, in a press release, an amnesty under which prohibited immigrants who met certain conditions would be granted resident status if they came forward before a certain date. Salemi appeared to meet those conditions, but the Minister refused his application for the benefit of the amnesty and proposed to deport him under s. 18 of the Act. Salemi cited Schmidt v. Secretary of State for Home Affairs66 and claimed that the news release had given him a “legitimate expectation” that he would be entitled to remain in Australia. Hence the Minister should have given him an opportunity to be heard before deciding to issue the deportation order. A majority of the High Court held that Salemi had no “legitimate expect­ ation,”67 and half of the Court held that the legislative scheme indicated that Parliament did not intend that people affected by deportation orders issued under s. 18 should have a right to be heard.68 SHYSTER’s fact vectors are (YUNYYN), (UYNY) and (YYYYYNN). In the Expectation area, SHYSTER concludes that the applicant did have a legitimate expectation which was affected by the decision. The nearest neighbour is Attorney-General of Hong Kong v. Ng Yuen Shiu; the nearest other is South Australia v. O’Shea. SHYSTER warns that the weighted correlation coefficients suggest that SA v. O’Shea should be nearest neighbour. In the Affected area, SHYSTER concludes that the decision affected the prop­ erty, right, interest, status, or legitimate expectation of the applicant. Although not discussed in the case (apart from the issue of legitimate expectation) this is clearly true. The nearest neighbours are Haoucher v. Minister of State for Immigration and Ethnic Affairs and Kioa v. West ; the nearest other is Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd. � � � � � � � 5.5 The implication of natural justice 183 In the Natural area, SHYSTER disagrees with the High Court and concludes that a duty to observe natural justice is implied. The nearest neighbour is Kioa v. West ; the nearest other is SA v. O’Shea. However, according to the legal expert, it is now generally believed that Salemi v. MacKellar was wrongly decided. Salemi did have a legitimate expectation, and natural justice should have been implied. The case was not included in the Natural specification for this reason. All of SHYSTER’s chosen cases were decided after Salemi v. MacKellar. Heatley v. Tasmanian Racing and Gaming Commission 1977 Heatley v. Tasmanian Racing and Gaming Commission69 is one of the leading cases in the Expectation area, and its facts are summarized in the dump file for the Natural specification in §A.5. It does not appear in either of the Affected or Natural areas, so it is used here to test those two areas. SHYSTER’s fact vectors are (UNYY) and (YYYUNNN). In the Affected area, SHYSTER agrees with the High Court that the decision affected the property, right, interest, status, or legitimate expectation of the applicant. The nearest neighbour is FAI Insurances Ltd v. Winneke; the nearest other is Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd. In the Natural area, SHYSTER concludes, as did the High Court, that a duty to observe natural justice is implied. The nearest neighbour is Annetts v. McCann; the nearest other is McInnes v. Onslow Fane. However, SHYSTER warns that these two cases are equidistant from Heatley v. TRC ; it chooses Annetts v. McCann because it is a decision of five judges of the High Court whereas McInnes v. Onslow Fane is merely a decision of the Chan­ cery Division of the English High Court. It also warns that both the weighted association and weighted correlation coefficients prefer McInnes v. Onslow Fane as nearest neighbour. All of SHYSTER’s chosen cases were decided after Heatley v. TRC, however the legal expert points out that the cases it has chosen in the Affected area both con­ cerned licence interests; SHYSTER’s choice would have been better if the chosen cases had been about a mere interest. Cole v. Cunningham 1983 Like Heatley v. Tasmanian Racing and Gaming Commission, Cole v. Cunning­ ham70 is one of the leading cases in the Expectation area, and its facts are summar­ ized in the dump file for the Natural specification in §A.5. It does not appear in either of the Affected or Natural areas, so it is used here to test those two areas. � � � 184 Chapter 5: Case studies SHYSTER’s fact vectors are (YNNY) and (YYNYNNN). In the Affected area, SHYSTER concludes that the decision affected the prop­ erty, right, interest, status, or legitimate expectation of the applicant—as did the Full Court of the Federal Court. The nearest neighbour is Council of Civil Service Unions v. Minister for the Civil Service; the nearest other is Minister for Environment v. Peko-Wallsend. SHYSTER agrees with the Court that a duty to observe natural justice is implied. The nearest neighbour is Annetts v. McCann; the nearest other is McInnes v. Onslow Fane to which no reference was made. As with Heatley v. TRC, SHYSTER warns that these two cases are equidistant from Cole v. Cunningham and chooses Annetts v. McCann over McInnes v. Onslow Fane because it was decided in a higher court. It also warns that both the weighted coefficients prefer McInnes v. Onslow Fane. Apart from McInnes v. Onslow Fane, all of SHYSTER’s chosen cases were decided after Cole v. Cunningham. The legal expert approves of SHYSTER’s choice of nearest neighbour in the Affected area, but not its choice of nearest other: Minister for Environment v. Peko-Wallsend concerned a property interest, where Cole v. Cunningham concerned a promissory interest. Ackroyd v. Whitehouse (Director of National Parks & Wildlife Service) 1985 In Ackroyd v. Whitehouse (Director of National Parks & Wildlife Service), 71 Ack­ royd was a bird trapper. His licence was cancelled without notice, hearing or reasons, by the Director under s. 134(1) of the National Parks and Wildlife Act 1974 (Cth). Section 135 of the Act granted a person whose licence had been cancelled the right to appeal to the Minister. Ackroyd claimed that the Minister ought to have applied natural justice principles before making his decision. The Director cited Twist v. Council of Municipality of Randwick claiming that the right of appeal to the Minister meant that natural justice ought not to be implied. The New South Wales Court of Appeal disagreed, and found for Ackroyd. Kirby P held that the fact that Ackroyd had had his licence renewed annually for several years “might reasonably give rise to a legitimate expectation” that the licence would continue in operation unless some misconduct was proved.72 SHYSTER’s fact vectors are (NUYNYN), (YNNY) and (YYYYNYN). In the Expectation area, SHYSTER agrees with Kirby P that Ackroyd had a legitimate expectation which was affected by the decision. The nearest neighbour is FAI Insurances Ltd v. Winneke; the nearest other is South Australia v. O’Shea, which was decided after Ackroyd v. Whitehouse. � � � � 5.5 The implication of natural justice 185 In the Affected area, SHYSTER concludes (as did the Court) that the decision affected the property, right, interest, status, or legitimate expectation of the ap­ plicant. The nearest neighbour is Council of Civil Service Unions v. Minister for the Civil Service; the nearest other is Minister for Arts Heritage and Environ­ ment v. Peko-Wallsend Ltd, which was decided afterwards. Finally, in the Natural area, SHYSTER agrees with the Supreme Court that a duty to observe natural justice is implied. The nearest neighbour is Commissioner of Police v. Tanos; the nearest other is McInnes v. Onslow Fane. Of those of SHYSTER’s chosen cases which were decided before Ackroyd v. Whitehouse, only McInnes v. Onslow Fane was cited in argument—none was cited in the Court’s judgment. However, the legal expert approves of SHYSTER’s choice of cases in the Expectation area. SHYSTER issues no warnings. Hodgens v. Gunn 1989 In Hodgens v. Gunn, 73 Hodgens had been convicted of ill-treating dogs. The Magistrates Court had made an order under s. 19(2) of the Animals Protection Act 1925 (Qld) permanently prohibiting him from having a dog. He continued to breed dogs, which were seized by the police. The Minister made an order under s. 11(4) of the Act forfeiting the dogs to the Crown. Hodgens appealed to the Full Court of the Supreme Court of Queensland, claiming that the principles of natural justice required that he should have been given a hearing before the Minister made his order. The Court agreed: since the Minister’s order involved the expropriation of property without compensation, and the Minister had absolute discretion in the exercise of this power, natural justice was implied. (However, Hodgens was denied the remedy he sought because he had waited six months before seeking a Court order.) SHYSTER’s fact vectors are (NNNNNN), (YNNN) and (YYNYYNY). In the Expectation area, SHYSTER concludes that Hodgens did not have a le­ gitimate expectation which was affected by the decision. This is appropriate; legitimate expectation was not an issue in Hodgens v. Gunn. The nearest neigh­ bour is Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd; the nearest other is Cole v. Cunningham. In the Affected area, SHYSTER agrees with the Court that the decision affected Hodgens’s property, right, interest, status, or legitimate expectation. The nearest neighbour is Bread Manufacturers of New South Wales v. Evans ; the nearest other is Minister for Environment v. Peko-Wallsend. However, in the Natural area SHYSTER concludes (unlike the Supreme Court) that a duty to observe natural justice is not implied. The nearest neighbour is Council of Civil Service Unions v. Minister for the Civil Service; the nearest others are FAI Insurances Ltd v. Winneke and Haoucher v. Minister of State for Immigration and Ethnic Affairs. � � � 186 Chapter 5: Case studies Interestingly, all of SHYSTER’s unweighted measures suggest that FAI v. Win­ neke and Haoucher v. Minister for Immigration should be the nearest neighbours. Furthermore the ideal points, centroids, and specified and ideal point directions all suggest that the result should be Implied, but only the specified directions give rise to a warning (see §3.12.4). Of SHYSTER’s chosen cases, only FAI v. Winneke was judicially cited.74 The legal expert agrees with SHYSTER’s choice of cases in the Affected area, and with its choice of nearest neighbour in the Expectation area. However, she disagrees with its choice of Cole v. Cunningham as nearest other in the Expectation area because that case did not concern property rights. Western Australia v. Bropho 1991 In Western Australia v. Bropho, 75 land on the Swan River was converted into a Crown reserve and a proposal put forward to renovate an old brewery on the site. However, the Aboriginal Heritage Act 1972 (WA) required that a Committee be asked to recommend to the Minister as to whether there was any Aboriginal site on the land. The Committee invited public submissions, and recommended that the site was a significant Aboriginal site and that the development should not be allowed. Nevertheless, the Minister gave her consent to the development. Bropho was a custodian of the site, in accordance with the customs of his people. He claimed that the Minister was obliged to give him a hearing before making her decision. The Full Court of the Supreme Court of Western Australia held that the fact that Bropho had not made a submission to the Committee, before it made its recommendation to the Minister, meant that he had no legitimate expectation of a hearing before the Minister. The Minister was obliged to follow natural justice principles in coming to her decision, but the opportunity to be heard before the Committee meant that this obligation had been met. Malcolm CJ stated that Bropho had a “special interest in the subject matter of the action” and was “more particularly affected by the proposed development than ordinary members of the public”. 76 But Franklyn and Anderson JJ were not so sure. They said that Bropho may only have had a “spiritual concern rather than a special interest” in the matter.77 SHYSTER’s fact vectors are (NNNNNN), (NNNN) and (NYYNYYN). In the Expectation area, SHYSTER agrees with the Supreme Court that Bro­ pho did not have a legitimate expectation which was affected by the decision. The nearest neighbour is Minister for Arts Heritage and Environment v. Peko- Wallsend Ltd ; the nearest other is Cole v. Cunningham. In the Affected area, SHYSTER concludes that the decision did not affect Bro­ pho’s property, right, interest, status, or legitimate expectation. The nearest neighbour is Minister for Environment v. Peko-Wallsend; the nearest other is � � � � 5.5 The implication of natural justice 187 Bread Manufacturers of New South Wales v. Evans. This conclusion is in line with the reservations expressed by Franklyn and Anderson JJ, although their honours appear to have continued on the assumption that Bropho did have an interest which was affected. In the Natural area, SHYSTER concludes, as did the Court, that a duty to ob­ serve natural justice is implied. The nearest neighbour is Marine Hull & Liability Insurance Co. Ltd v. Hurford ; the nearest other is Minister for Environment v. Peko-Wallsend. Of course, if a decision does not affect the interests of an applicant (as in this case) then natural justice should never be implied. This fact is captured, in a limited sense, by the specified direction towards Not-Implied of a value of no for A1—and SHYSTER warns that the specified directions suggest Not-Implied in WA v. Bropho. 78 None of SHYSTER’s chosen cases was cited in the published judgment. Ainsworth v. Criminal Justice Commission 1992 In Ainsworth v. Criminal Justice Commission, 79 the Commission’s advice had been sought on the implementation of the introduction of poker machines into Queensland. Its report dealt with matters of general concern and with particular poker machine suppliers and manufacturers (including Ainsworth’s companies). It was very critical of their conduct and recommended that “the Ainsworth group of companies not be permitted to participate in the gaming machine industry in Queensland.” The Commission had formed its opinion on the basis of reports of other bodies. Ainsworth had not been informed of the Commission’s interest in his group of companies, and was not given any opportunity to make representations on the matter before the Commission made its report. Ainsworth claimed that the business reputation of the companies was an interest which was adversely affected by the Commission’s report. The Full Court of the High Court unanimously agreed. It followed Annetts v. McCann and held that the Commission ought to have applied natural justice principles.80 The report files for Ainsworth v. CJC (one for each area) are given as examples in §B.5. SHYSTER’s fact vectors are (NNNNNN), (NNYN) and (YYYYNNN). In the Expectation area, SHYSTER concludes that the applicant did not have a legitimate expectation which was affected by the decision. The nearest neighbour is Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd; the nearest other is Cole v. Cunningham. This is appropriate; legitimate expectation was not an issue in Ainsworth v. CJC. 188 Chapter 5: Case studies In the Affected area, SHYSTER concludes, as did the High Court, that the decision affected the property, right, interest, status, or legitimate expectation of the applicant. The nearest neighbour is Annetts v. McCann: the same case as the High Court followed; the nearest other is Minister for Environment v. Peko-Wallsend. In the Natural area, SHYSTER agrees with the Court that a duty to observe natural justice is implied. The nearest neighbour is also Annetts v. McCann; the nearest other is McInnes v. Onslow Fane. However, SHYSTER warns that these two cases are equidistant from Ainsworth v. CJC ; it chooses Annetts v. McCann because it is a decision of five judges of the High Court whereas McInnes v. Onslow Fane is merely a decision of the Chan­ cery Division of the English High Court. It also warns that both the weighted association and weighted correlation coefficients prefer McInnes v. Onslow Fane as nearest neighbour. 5.5.4 Generated tests The legal expert provided two examples for generating tests using the Natural area of the Natural specification: generated tests 7 and 8. The results of these tests are summarized in figure 5.13. Generated test 7 The legal expert advises that, if the decision did not affect the property, right, interest, status, or legitimate expectation of the applicant, then natural justice will not be implied. The fact vector is (NUUUUUU). 81 SHYSTER generates 64 in­ stantiations, half the search space, and in 56 of them the choice of result is good (i.e. Not-Implied). In one of the 8 instantiations in which the choice of result is bad, SHYSTER warns that an ideal point suggesting the good result is at least as near to the instant case as is the nearest neighbour. In all 8, the decision is apt to have a discrete impact on the interests of the applicant. This does not justify SHYSTER’s choice of result in those instantiations. Generated test 8 The legal expert also advises that, if the decision is not apt to have a discrete impact on the interests of the applicant, natural justice will not be implied. The required fact vector is (UNUUUUU). But this cannot be given to SHYSTER directly because the value of A1 cannot be unknown: it is an external attribute defined by reference to the Affected area. So, this test is divided into two. In the first test (“8a” in figure 5.13), the value of A1 is yes; in the second (“8b”) it is no: i.e. the fact vectors are (YNUUUUU) and (NNUUUUU). 82 Each of these two tests generates half of the required search space. � � � � 5.6 Conclusion 189 In 55 of the total of 64 instantiations generated for test 8, the choice of result is good (i.e. Not-Implied). In one of the 9 instantiations in which the choice of result is bad, SHYSTER issues an ideal point warning. In each of these 9 instan­ tiations, the decision affected the property, right, interest, status, or legitimate expectation of the applicant. This does not justify SHYSTER’s choice of result in those instantiations. 5.5.5 Conclusion The tests described in §5.5.3 indicate that, using the Natural specification, SHYSTER is exceptionally good at coming to the right conclusion (95.46% good results), though not very good at choosing good cases to use in argument (40%). As shown in figure 5.13, SHYSTER’s choice of result is good in 86.72% of the instantiations in both generated tests. If a bad choice of result is considered good where an ideal point warning is issued, SHYSTER’s success rate rises only slightly to 89.06% and 85.94% for each test: 87.50% overall. (In none of the instantiations in these tests does SHYSTER issue a warning about a nearer ideal point when the choice of results is good; i.e. none of SHYSTER’s ideal point warnings is bad.) As discussed in §5.5.2, three pairs of cases in the Natural area have identical facts yet different results. Because all of the leading cases in the Natural area were correctly decided, this inconsistency indicates a need for more attributes to distinguish between cases in the area. Extra attributes were not added, because the seven attributes were based on the seven factors identified by McMillan as being important.83 The Natural area is a good example of an area with attributes which re­ quire further definition by reference to other areas. The Natural specification represents—and represents effectively—an area of law that is quite fluid and expanding rapidly: as discussed in §5.5.3 above, both Twist v. Council of Mu­ nicipality of Randwick and Salemi v. MacKellar would be decided differently today—seventeen years after they were heard by the High Court. SHYSTER (when using the Natural specification) produces good advice in the test cases, and chooses good results in the vast majority of the generated tests. However, the reflexive tests performed in §D.5, and summarized in figure D.4, indicate that the Natural specification is not as capable of handling new fact situations as are the other three specifications. 5.6 Conclusion The results of all the testing described in this chapter—test cases, generated tests and reflexive tests—are discussed below. General conclusions are drawn from these results in the next chapter. � � � � � � 190 Chapter 5: Case studies 5.6.1 Test cases The results of the testing using test cases are summarized in figure 5.12. Several special symbols are used in that figure. A tick is used to indicate a “good” result or choice of cases; a cross indicates a “bad” result or choice of cases. (These terms are defined in §3.13.2.) A dash indicates that a chosen case was not decided before the test case and that it is not possible to evaluate that choice. A blank space indicates that no symbol is applicable. The “Warnings” column summarizes the safeguard warnings that were issued during each test. A warning that one of the weighted similarity measures sugges­ ted a different result is indicated by the appropriate symbol: S� for the weighted association coefficients; r� for the correlation coefficients. An I or a µ indicates that an ideal point or a centroid suggesting a different result was at least as near to the instant case as was the nearest neighbour.84 A ⊃ indicates that the specified directions suggested a different result. An indicates that the test → produced equidistant results. A ∨ in the warnings column indicates that one or more unknown values was instantiated. The fraction after each ∨ is the number of instantiations that pro­ duced a different result to that of the nearest neighbour, over the total number of instantiations; SHYSTER only issues a warning if an instantiation has a different result. At the right of the warnings column is a tick or a cross which indic­ ates whether the warning (or absence thereof) was “good” or “bad”—as defined in §3.13.2. These symbols are also used in the figures in appendix D which summarize the reflexive testing described in that appendix. As figure 5.12 illustrates, SHYSTER has proven itself exceptionally able to choose good results: a 95.24% success rate (including ideal point tests). However, its ability to choose good cases with which to argue varies dramatically between the four different specifications. For the Finder specification, SHYSTER’s choice of cases is excellent, but there is only one test case for that specification. For the Authorization specification, 75% of its chosen cases are good. However, only 41.67% and 40% of the cases chosen from the Employee and Natural specifications were good. One reason for this discrepancy may be the greater number of leading cases in the Employee and Natural specifications (15 each85) than in the Finder and Authorization specifications (8 and 9, respectively). Judges and lawyers will often refer only to a handful of cases; only rarely does a judge discuss exhaustively all of the important cases in a field.86 Given an area of law in which there were comparatively few leading cases, SHYSTER would be more likely to chose a good case (all other things being equal). � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � × × � � � � � � � – – � – – – ×� � � � – – – – – – ×� � � – – � � � – �×× � � � � � � � � � � � � � 191 � 5.6 Conclusion Nearest Nearest Result Case n’bours others Warnings Parker v. British Airways � � � � N a t u r a l E m p l o y e e A u t h o r i z a t i o n F i n d e r CBS v. Ames WEA v. Hanimex CBS v. Amstrad APRA v. Jain Hypothetical case 1 Hypothetical case 2 IAuth INot-Auth ILiable × × × ⊃ ⊃ ⊃ ⊃ r� ⊃ → ⊃ r� ∨ 4 0 × × × × × × × Narich v. CPT Stevens v. Brodribb Re Porter; Re TWU BWIU v. Odco Hypothetical case 3 Hypothetical case 4 IEmployee IContractor × × × × × × ⊃ 0 2 ∨ 1 4 ∨ 0 4 ∨ r� ∨ 2 0 0 2 ∨ r� ∨ 4 0 × × × × Twist v. Randwick Salemi v. MacKellar Heatley v. TRC Cole v. Cunningham Ackroyd v. Whitehouse Hodgens v. Gunn WA v. Bropho Ainsworth v. CJC IImplied INot-Implied × × ×× � � � � – � � �× �× � � ××× ××× ×× ⊃ 0 r� ∨ 2 S� r� → ∨ 2 0 S� r� → 0 2 ∨ ⊃ ⊃ S� r� → S� r� × × × × × Figure 5.12: A summary of the testing using test cases performed in this chapter. These tests are described in §5.2.3, §5.3.3, §5.4.3 and §5.5.3. The meaning of the symbols used here is explained in §5.6.1. Symbols for tests in the Natural specification refer to the Expectation, Affected and Natural areas, respectively. Safeguard warnings in the Natural specification are from the Natural area—except those marked with an asterisk, which are from the Expectation area. � � � 192 Chapter 5: Case studies The appropriateness of SHYSTER’s warnings also varies between specifications. In the single test case for the Finder specification, the (absence of a) warning is good. But warnings in the Authorization, Employee and Natural specifica­ tions are generally bad: only 22.22%, 50% and 30%, respectively, good warnings. For the Authorization specification, all but one of the seven bad warnings are due to specified directions. Across all four specifications, the rate of good warnings would rise from 35.71% to 60.71% if the specified direction warnings were ignored. Yet specified directions represent the expert’s opinion as to which values for attributes suggest certain results; warnings based upon the specified directions should be sound. It is possible that the preponderance of bad specified direction warnings in tests of the Authorization specification may be due to the fact that the Author­ ization area has more than two results. However, that is not the case in the other three specifications. In each of the nine tests in which a warning was issued due to an extra similarity measure (i.e. S� or r�) the warning was bad. Without those warnings due exclusively to either or both extra similarity measures, 53.57% of the warnings are good. That these warnings are so bad indicates that the choice of similarity measure made in §3.9.3 was a good one: using known and unknown distance measures (variations on the weighted distance measure d� ) SHYSTER produces jk better results than it would using either of the other two weighted measures. In only one of the twenty-one tests did an instantiation of the instant case have a different result to that of the uninstantiated instant case: Re Porter; Re Transport Workers Union of Australia, 87 a case used to test the Employee specification in §5.4.3. As explained there, the legal expert agrees with SHYSTER that that instantiation would indeed lead to a different result. 5.6.2 Generated tests The results of the generated testing are summarized in figure 5.13. The “� Results” column gives the number of generated tests in which SHYSTER’s choice of result is good, as a fraction of the total number of generated tests. That fraction is also expressed as a percentage. For tests in which SHYSTER’s choice of result is bad (i.e. “× Results”) the number of tests in which ideal point warnings were issued is given in the “Wrn” column. The “IP” column gives the number of tests in which the nearest ideal point was that of a different result, but in which no warning was issued because that ideal point was not nearer the instant case than was the nearest neighbour. Corresponding figures for those tests in which SHYSTER’s choice of result is good are given in the next column. 193 � 5.6 Conclusion × Results � Results � Results No. � Results % %IP Wrn IP Wrn and warnings 1 14/16 87.50 2 0 0 0 14/16 87.50 2 12/16 75.00 0 4 0 0 16/16 100.00 3 51/64 79.69 0 13 0 0 64/64 100.00 77/96 80.21 2 17 0 0 94/96 97.92 4 363/512 70.90 4 145 0 0 508/512 99.22 5 4198/6144 68.33 861 492 144 111 4579/6144 74.53 6 875/1024 85.45 37 112 0 0 987/1024 96.39 5436/7680 70.78 902 749 144 111 6074/7680 79.09 7 56/64 87.50 1 1 10 0 57/64 89.06 8 55/64 85.94 1 0 16 0 55/64 85.94 8a 23/32 71.87 1 0 12 0 23/32 71.87 8b 32/32 100.00 0 0 4 0 32/32 100.00 111/128 86.72 2 1 26 0 112/128 87.50 Figure 5.13: A summary of the generated testing performed in this chapter. These tests are described in §5.3.4, §5.4.4 and §5.5.4. This table is explained in §5.6.2. Note that results for generated tests of the Employee specifica­ tion take into account the removal of paradoxical cases from the search space (see §5.4.4). In the “� Results and warnings” column, a bad choice is considered good if an ideal point warning is issued. The number of good choices is also expressed as a fraction of the total number of generated tests, and as a percentage. Except in one generated test, SHYSTER never issues an ideal point warning when a good result has been chosen. The bad warnings issued in that test—generated test 5— are accounted for in this column because a good choice of result is considered bad if an ideal point warning is issued. A total for each column is given after each group of tests. Ignoring warnings, SHYSTER has a success rate of 80.21%, 70.78% and 86.72%, respectively, for each of the three specifications. If warnings are taken into account the success rate rises to 97.92%, 79.09% and 87.50%. 5.6.3 Reflexive tests The results of the reflexive tests are described in §D and summarized in fig­ ures D.1, D.2, D.3 and D.4. As explained in §3.13.5, reflexive tests do not test SHYSTER’s approach to case law; they do, however, provide information about a specification. The reflexive tests indicate that the Finder, Authorization and Employee specifications are capable of handling new fact situations, but that the Natural specification is less capable. N a t u r a l E m p l o y e e A u t h . 6 Conclusion Among the vital processes occurring at [the] intellectually intractable points of precedent growth are the modification, supplementation, or even abandonment of reigning legal precepts by reference to contemporary social ideas and ideals. Because the doctrine of precedent, as many lawyers and social scientists still misunderstand it, does not allow for the operation of these processes, their occurrence is an evercontinuing source of trauma. Julius Stone (1964) Man and Machine in the Search For Justice88 We do not use to judge of Cases by fractions. John Finch (1637) R v. Hampden89 Donne, I suppose, was such another Who found no substitute for sense, To seize and clutch and penetrate; Expert beyond experience . . . T. S. Eliot (1920) Whispers of Immortality90 195 196 Chapter 6: Conclusion 6.1 Introduction SHYSTER was developed to support the main argument of this thesis: that a use­ ful, working legal expert system can be based upon a pragmatic approach to the law. In this chapter, SHYSTER and its approach to case law are evaluated (§6.2), and avenues of future research are identified (§6.3). The contribution made by this thesis to the field of legal expert systems development is discussed in §6.4. 6.2 Evaluating SHYSTER This evaluation of SHYSTER is carried out under five headings: its usefulness, its generality, the quality of its advice, its limitations, and possible enhancements that could be made to it. Although there is some overlap between these five topics—and in the identification of areas of future research in §6.3 below—it is convenient to evaluate SHYSTER under these headings. 6.2.1 A useful, working system Susskind claims that in 1986 there was “an embarrassing lack of demonstrable expert systems in law”. 91 Although several systems have been developed since then, the embarrassment has continued. In 1990, Kowalski and Sergot announced that they had decided not to finish their system because their experience sugges­ ted that the final stage of the development process “can involve a considerable amount of work and extra programming effort.”92 That decision made feeble their claim that “there are no outstanding technical obstacles which need to be overcome”,93 and left Kowalski and Sergot particularly vulnerable to Moles’s sug­ gestion that the real reason for the abandonment of the project was a realization of the inadequacy of their approach.94 The development of (at least) a working prototype of a system is an important part of any expert system project. In order to demonstrate the efficacy of a pragmatic approach to case law, SHYSTER’s case-based system has been fully implemented, as explained in chapter 4. SHYSTER has been tested, as explained in chapter 5. That testing was not as comprehensive as is desirable but was as comprehensive as possible within the restrictions of this thesis project, and compares favourably with the testing of other comparable systems.95 Ashley, as quoted in §2.5, lists the following tasks as being general to all case-based reasoning: . . . (1) ordering relevant cases and potentially relevant cases in terms of how analogous they are to the problem situation, (2) selecting the most analogous cases, (3) identifying configurations of counterexamples, (4) hy­ 197 � 6.2 Evaluating SHYSTER pothetically modifying the problem situation to explore contingencies, and (5) comparing case-based analyses of different problem situations to ex­ plain differences.96 SHYSTER performs all five of these tasks, within the restrictions placed upon it by its simple knowledge representation. SHYSTER’s usefulness for the user—a lawyer—is due to the fact that it is based upon a pragmatic model of legal reasoning; it operates at the same prag­ matic level of abstraction as do lawyers. SHYSTER’s Reporter module goes to some considerable trouble to ensure that its reports are in a form that might be produced by a lawyer. Details of SHYSTER’s distance calculations are not part of its reports. During the development and testing of SHYSTER, an unexpected aspect of the system’s usefulness became apparent. Three of the four specifications described in chapter 5 were developed by the author in consultation with lawyers with expertise in the relevant fields. Two of those legal experts found that the process of specifying an area of law for SHYSTER, and examining and evaluating its reports, changed the way that they looked at that area. They felt that the process had contributed to their knowledge and understanding of the field. For example, the legal expert who was consulted about the Employee specification found that the process showed that that area of law was more systematic than she had first thought. Ashley says that a case-based legal expert system could be useful as part of a legal tutoring system.97 The experiences of the legal experts during the development of the specifications suggest that SHYSTER has the potential to assist users to gain similar insights. 6.2.2 A general approach The general applicability of SHYSTER’s approach to case law is an open question. SHYSTER has been designed so that it can provide advice in different areas of case law, specified by legal experts. However, as explained in §3.3.1, the model of legal reasoning adopted for SHYSTER in §3.3.3 reflects the way in which lawyers reason with statutes and cases in areas of private law. SHYSTER has not been designed to deal with areas of public law in which the doctrine of precedent is given less weight. Nevertheless, the four specifications described and tested in chapter 5 suggest a wide application for SHYSTER’s approach to case law. The specifications are quite different from each other: they are of differing sizes (in terms of the number of attributes and the number of cases), and they represent disparate areas of the law. The Finder specification represents a completely case-based area of law. The Authorization specification represents an area of case law which has developed 198 Chapter 6: Conclusion from specific statutory provisions. The Employee specification represents an area of case law which has developed alongside several different statutes. In the cases that make up the Natural specification, judges were overtly concerned with matters of policy. That specification is more like a public law area than any of the other three specifications.98 Mendelson applied dimensional analysis (as performed by HYPO) to the law governing government appeals in US criminal cases.99 His goal was to determ­ ine how “domain-sensitive” was HYPO’s approach to case law. His dimensional analysis was unsuccessful, because, he concludes, the area of law “is unsettled, fact-sensitive and conflicting.”1 As explained in §5.5, the Natural specification shares all three of these characteristics. This makes SHYSTER’s success with that specification (although not as consistent as its success with other specifications) particularly interesting. Although results vary between the four specifications (as discussed in §5.6.1, and in §6.2.3 below) they are generally good in all four specifications. This thesis does not demonstrate the generality of SHYSTER’s approach to private case law, but its does suggest its broad applicability. 6.2.3 Good advice As discussed in §3.13.2, the quality of SHYSTER’s advice is evaluated on the basis of its prediction as to the likely result, and its choice of cases for use in argument. Within the limitations imposed by its internal representation of case law, SHYSTER is capable of producing quite sophisticated reports—but these reports can only be as good as the cases it chooses. (What is meant by a “good” result or a “good” choice of cases is defined in §3.13.2.) As is shown in chapter 5, SHYSTER has proved remarkably good at choosing a result: in test cases and in generated tests. It is, however, quite inconsistent (across the four specifications) in the quality of its choice of cases. It should be remembered, however, that the test cases used in chapter 5 are difficult tests. As explained in §3.13.3, the same filters in the legal system which restrict the number of cases that are reported also ensure that those cases that are reported make difficult tests.2 Furthermore, it is a little harsh to rule SHYSTER’s choice of a case “bad” on the basis that that case was not cited in judgment or argument by a judge or by counsel. Some judges will not look beyond the cases cited in argument by counsel. For all sorts of non-legal reasons (lack of time, incompetence, etc.) counsel may not cite all the good cases in argument. And, of course, opinion as to what is a good case to use in argument differs. However, harsh as it may be, some method of evaluating the quality of SHYSTER’s choice of cases had to be adopted.3 � � � � � � 199 � 6.2 Evaluating SHYSTER 6.2.4 Limitations Although SHYSTER has proved quite successful, it has several limitations. As well as being restricted (for the most part4) to areas of private law, there are some areas of private law to which SHYSTER is not suited. These restric­ tions upon domains—restrictions which apply to all legal expert systems—are examined in §3.13.1. “Reasonableness,” for example, is a difficult concept. As discussed in §5.3.2, it plays an important part in the Authorization specification, yet it is so neb­ ulous a concept that it would probably be impossible to specify for SHYSTER. Furthermore, the meaning of “reasonableness” may well be very different within different areas of the law: even if it could be specified for one area, that specific­ ation might not be appropriate for another area within which the concept is also important. SHYSTER’s solution to this problem in the Authorization specification is to leave the user to answer the question of whether the accused took reasonable steps.5 This is not a major limitation: it is quite possible that the user (a lawyer) knows the answer to that question, but not to the larger question of whether the accused authorized the infringement. In any event, the user has the option of answering unknown to difficult questions, forcing SHYSTER to examine instantiations. As discussed in §5.6.1, if many leading cases are specified then SHYSTER’s ability to choose good cases is reduced. This problem can be avoided by choosing a dif­ ferent threshold within which SHYSTER will consider two cases to be equidistant.6 If such a change is made, SHYSTER will tend to choose more cases to use in ar­ gument. This increases the likelihood of SHYSTER choosing good cases, but also increases the number of bad cases chosen. The fact that SHYSTER will construct arguments for all possible results means that sometimes it will produce extraneous arguments. For example, as explained in §5.3.3, given an authorization case, SHYSTER will argue about the Liable result even if the question of direct or vicarious liability does not arise. In an area where there are several results, this problem would be exacerbated. One solution to this problem would be to restrict SHYSTER to arguing only about the nearest n results (where n is 2 or more): i.e. the result of the nearest neighbour, and the results of the n − 1 nearest others. This solution reduces the likelihood of SHYSTER building an irrelevant argument, but increases the likelihood of it missing a relevant one. � � � � � � � � � 200 Chapter 6: Conclusion 6.2.5 Possible enhancements The model of legal reasoning adopted in §3.3.3 allows for open-textured concepts to be defined by reference to case law. However, there may be circumstances in which a case-based open-textured concept is defined by reference to statute law. SHYSTER’s model of legal reasoning does not allow for such circumstances, though it could be augmented to allow resolution of open texture using legislation. Implementing such an augmented model should not pose any major difficulties. SHYSTER’s case law specification language would have to be extended to allow external attributes to be linked to rules in the rule base in the same fashion that external attributes are linked to areas in the case base at present. SHYSTER allows the specification of no more than one ideal point for each result in each area. It is possible that there might be more than one best combination of attribute values for a given result. There is no reason why SHYSTER could not be modified to allow multiple ideal points for the same result. As shown in §5.4.5, SHYSTER can produce clumsily worded sentences. The qual­ ity of SHYSTER’s reports could be improved by changing the way in which it uses attribute strings. The number of attribute strings that are written could be restricted on the basis of their importance (defined by their weights), but sometimes the user may be interested in a less important attribute. Attribute strings could be combined in more sophisticated ways: rather than writing “x is not true; and y is not true,”SHYSTER could write “neither x nor y is true.” This would require changes to SHYSTER’s case law specification language, as well as to the Reporter module. 6.3 Future research The most obvious avenue of future research using SHYSTER is the development of a rule-based system, and the linking together of that rule-based system with the existing case-based system to form a hybrid system. As argued in chapter 2, a hybrid approach is desirable where the system seeks to represent statute law and case law. SHYSTER has been designed to facilitate such an articulation. The mechanism by which this could be achieved is explained in §4.3. There are also several areas of further research which could be carried out using the existing SHYSTER system. New specifications could be written, and tested in the same manner as are the four specifications described in chapter 5. As well as writing different specifications, different legal experts could be asked to � � � � � � � � � 201 � 6.3 Future research specify the same area of law. It would be interesting to examine the differences (if any) between the specifications and the ways in which those specifications make SHYSTER behave. Research could also be done into the process of updating specifications to take account of changes in the law. Most changes could be captured by the addition of new cases to, and the removal of out-of-date cases from, the specification—both operations are easily performed using SHYSTER’s case law specification language. More significant changes in the law might require the addition of new attributes, or the complete rewriting of a specification. SHYSTER’s method of assigning weights to attributes deserves further examin­ ation. As explained in §3.7, SHYSTER adopts FINDER’s method of attribute weighting. This method contributes significantly to SHYSTER’s distance calcu­ lations, and so has an important effect on SHYSTER’s opinions. As has been shown, SHYSTER’s opinions are generally good which suggests that its method of attribute weighting is also a good one. Yet, as demonstrated in chapter 5, SHYSTER’s weighting of attributes is not always appropriate.7 Some experimentation by the author indicates that weighting attributes ac­ cording to the inverse of the variance of each attribute across the leading cases (as SHYSTER does) produces better results than weighting according to the vari­ ance (the standard approach adopted in statistical classification problems8) or weighting each attribute equally. Further testing could be performed. SHYSTER’s Adjuster provides an ideal tool for experimenting with the effect of different weighting schemes. The choice of SHYSTER’s similarity measure, made in §3.9.3, is vindicated by the number of bad warnings generated using the other similarity measures (see §5.6.1). Nevertheless, further research could also focus on the refinement of SHYSTER’s similarity measure: its known and unknown distance. SHYSTER makes only lim­ ited use of the hierarchy of courts that can be part of any specification.9 It would be possible to take account of the rank of a leading case when quantifying the dis­ tance between it and the instant case. Another possibility would be to determine the court in which the instant case is likely to be heard, and to apply rigorously the doctrine of precedent as it relates to a court being bound by decisions of courts higher in its hierarchy.10 As demonstrated in several of the examples in chapter 5, there are occasions when specifying an area of law seems to call for a more disjunctive structure than SHYSTER permits: i.e. occasionally there is a need for a conditional attribute. 202 Chapter 6: Conclusion For example, A5 in the Authorization area of the Authorization specification concerns whether reasonable steps were taken to avoid an infringement of copy­ right. If the accused does take reasonable steps to avoid an infringement then she/he will not be said to have authorized the infringement,11 although she/he may still be directly or vicariously liable. Other examples can be found in the generated tests discussed in chapter 5: each of these relies on some attribute value or combination of attribute values implying a given result or results. The idea of conditional attributes can be captured, in a limited way, by at­ tribute direction.12 But the idea of allowing a given result to be conditional upon certain attribute values, in a strict sense, could be investigated. This would rep­ resent a major change to SHYSTER’s operation, and raises questions of attribute dependence. 6.4 The contribution of this thesis This thesis has argued that a legal expert system need not be based upon a complex or deep model of legal reasoning in order to be successful. It has recom­ mended a pragmatic approach to legal expert system design. SHYSTER is a working example—just one possible example—of such an expert system. Despite its simple knowledge representation structure, it has shown itself capable of producing good advice. And its simple structure has facilitated the specification of several different areas of law. This thesis expands upon the work of previous researchers in this field. It has adopted the approach taken by the developers of FINDER, and expanded it so that it incorporates some of the capabilities of more sophisticated systems (e.g. HYPO) without greatly increasing the complexity of the underlying model. SHYSTER incorporates a simple model of the way in which lawyers argue with cases. However, no attempt has been made to model the way in which lawyers decide which cases to use in those arguments. It is not claimed that lawyers choose the cases that they use in argument by reference to distance calculations in n-dimensional space. SHYSTER’s is not a deep model of legal reasoning. As Ashley points out: Even an only partially successful computational theory of jurisprudence may still prove useful. For example LEXIS and WESTLAW [legal retrieval systems13] represent a theory about legal reasoning: that relevance can be assessed by the appearance of keywords in past cases. No one would accept that theory as adequate, and yet it yields a practical tool.14 SHYSTER, too, is a practical tool. It demonstrates that a pragmatic approach to legal expert system development can be a successful one. Appendices 203 A Case law specifications The Law is the true embodiment Of everything that’s excellent. It has no kind of fault or flaw, And I, my Lords, embody the Law. William Gilbert (1882) Iolanthe; or The Peer and the Peri15 It is one of the maxims of the civil law, that definitions are hazardous. Samuel Johnson (1751) The Rambler 16 205 206 c Appendix A: Case law specifications A.1 Introduction The complete dump files for the four specifications used to test SHYSTER in chapter 5 are given in this appendix. The Finder specification (described in §5.2.2) is in §A.2. The Authorization specification (§5.3.2) is in §A.3. The Employee specification (§5.4.2) is in §A.4. (The complete specification file for the Employee specification, written in SHY­ STER’s case law specification language and input by SHYSTER, is in §C.3.) The Natural specification (§5.5.2) is in §A.5: external attributes in the Natural area are defined by reference to the Affected and Expectation areas. Each dump file is shown exactly as output by LaTEX, except that the attribute matrix in the Employee specification has been reduced in size so as to fit within the margins of this thesis. A.2 The FINDER specification Hierarchy Court 1 the Chancery Division of the English High Court the King’s Bench Division of the English High Court the Queen’s Bench Division of the English High Court Finder area Attributes Case c Result A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 C1 × × • × × × × • • × C2 × × • × • × × × • × C3 × × × × • × × × × • C4 • × × • × × × • • • C5 × • × × × × • • • × C6 • • • × • • × • • × C7 × • • × × × • • • × C8 • • • × × • × • • × Results Win: the finder wins. Lose: the finder loses. 1 1 Win 1 1 1 1 Lose 1 1 207 � A.2 The FINDER specification Attributes A1: Was the finder the occupier of the premises where the chattel was found? yes: the finder was the occupier of the premises where the chattel was found. no: the finder was not the occupier of the premises where the chattel was found. unknown: it is not known whether the finder was the occupier of the premises where the chattel was found. A2: Was the chattel attached to the land or premises where it was found? yes: the chattel was attached. no: the chattel was not attached. unknown: it is not known whether the chattel was attached. A3: Was the other claimant (the non-finder) the owner of the premises where the chattel was found? yes: the other claimant was the owner of the premises where the chattel was found. no: the other claimant was not the owner of the premises where the chattel was found. unknown: it is not known whether the other claimant was the owner of the premises where the chattel was found. A4: Was the other claimant the true owner of the chattel or did she/he claim through the rights of the true owner? yes: the other claimant was the true owner of the chattel or was claiming through the rights of the true owner. no: the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner. unknown: it is not known whether the other claimant was the true owner of the chattel or was claiming through the rights of the true owner. A5: Did the finder hand over the chattel to the other claimant after the finding? yes: the finder handed over the chattel to the other claimant after the finding. no: the finder did not hand over the chattel to the other claimant after the finding. unknown: it is not known whether the finder handed over the chattel to the other claimant after the finding. 208 Appendix A: Case law specifications A6: Did one of the parties rely on the terms of an agreement made with the other which purported to give her/him the right to the chattel? yes: one of the parties relied on the terms of an agreement made with the other which purported to give her/him the right to the chattel. no: neither party relied on the terms of an agreement regarding the right to the chattel. unknown: it is not known whether one of the parties relied on the terms of an agreement regarding the right to the chattel. A7: Was the finder a servant of the other claimant? yes: the finder was a servant of the other claimant. no: the finder was not a servant of the other claimant. unknown: it is not known whether the finder was a servant of the other claimant. A8: Was the chattel hidden or in a position so as to be difficult to find? yes: the chattel was hidden or was in a position so as to be difficult to find. no: the chattel was not hidden and was not in a position so as to be difficult to find. unknown: it is not known whether the chattel was hidden or was in a position so as to be difficult to find. A9: Was an attempt made to find the true owner of the chattel or, alternatively, was the chattel clearly abandoned? yes: an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned. no: no attempt was made to find the true owner of the chattel and the chattel was not clearly abandoned. unknown: it is not known whether an attempt was made to find the true owner of the chattel or whether the chattel was clearly abandoned. A10: Did either of the parties know of the existence of the chattel prior to the finding? yes: one of the parties knew of the existence of the chattel prior to the finding. no: neither party knew of the existence of the chattel prior to the finding. unknown: it is not known whether either of the parties knew of the existence of the chattel prior to the finding. 209 � A.2 The FINDER specification Cases in which the finder wins C1: Hannah v. Peel [1945] 1 KB 509 A1: the finder was not the occupier of the premises where the chattel was found. A2: the chattel was not attached. A3: the other claimant was the owner of the premises where the chattel was found. A4: the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner. A5: the finder did not hand over the chattel to the other claimant after the finding. A6: neither party relied on the terms of an agreement regarding the right to the chattel. A7: the finder was not a servant of the other claimant. A8: the chattel was hidden or was in a position so as to be difficult to find. A9: an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned. A10: neither party knew of the existence of the chattel prior to the finding. In Hannah v. Peel, 1 a 1945 decision of the King’s Bench Division of the English High Court, a brooch was found by the plaintiff who was a lance-corporal sta­ tioned in a house owned by the defendant. The house had been requisitioned by the army during the war and had never been occupied by the defendant. The plaintiff was adjusting the black-out curtains when he touched something on the top of the window-frame. He thought the object to be a piece of dirt or plaster and he dropped it on the outside window ledge. On the following morning, he saw that it was a brooch and, on the advice of his commanding officer, turned it over to the police for the purpose of finding the owner. In the following year, the police returned the brooch to the defendant who sold it to a jeweller. The plaintiff at all times maintained his rights to the brooch against all persons other than the true owner. The Court found for the plaintiff on the basis of Bridges v. Hawkesworth2 after a thorough review of the authorities. The Court further noted that the defendant was never in possession of the premises, that the brooch was never his, and that he had no knowledge of it until it was brought to his notice by the finder. 1[1945] 1 KB 509. 2(1851) 21 LJQB 75. 210 Appendix A: Case law specifications C2: Bridges v. Hawkesworth (1851) 21 LJQB 75 A1: the finder was not the occupier of the premises where the chattel was found. A2: the chattel was not attached. A3: the other claimant was the owner of the premises where the chattel was found. A4: the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner. A5: the finder handed over the chattel to the other claimant after the finding. A6: neither party relied on the terms of an agreement regarding the right to the chattel. A7: the finder was not a servant of the other claimant. A8: the chattel was not hidden and was not in a position so as to be difficult to find. A9: an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned. A10: neither party knew of the existence of the chattel prior to the finding. In Bridges v. Hawkesworth, 3 an 1851 decision of the Queen’s Bench Division of the English High Court, the plaintiff found a bundle of banknotes on the floor of the public area of a shop. He handed the notes to the shopkeeper in order that the true owner of the notes might be found. Although the owner was never found, the shopkeeper refused to return the notes to the finder. The Court found for the finder, holding that there is a “general right of [a] finder to any article which has been lost as against all the world except the true owner”.4 It was further noted that the notes had never been in the custody of the shopkeeper nor within the protection of his house as might be the case had they intentionally been deposited there. C3: Armory v. Delamirie (1722) 1 Str 505 A1: the finder was not the occupier of the premises where the chattel was found. A2: the chattel was not attached. A3: the other claimant was not the owner of the premises where the chattel was found. A4: the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner. A5: the finder handed over the chattel to the other claimant after the finding. A6: neither party relied on the terms of an agreement regarding the right to the chattel. A7: the finder was not a servant of the other claimant. 3(1851) 21 LJQB 75. 4ibid. at 77, per Patteson J. 211 � A.2 The FINDER specification A8: the chattel was not hidden and was not in a position so as to be difficult to find. A9: no attempt was made to find the true owner of the chattel and the chattel was not clearly abandoned. A10: one of the parties knew of the existence of the chattel prior to the finding. In Armory v. Delamirie, 5 a 1722 decision of the King’s Bench Division of the English High Court, a chimney sweep found a jewel. Although the report does not expressly say so, most commentators assume that the jewel was found in a chimney in the course of the sweep’s occupation. The jewel was handed to a goldsmith for appraisal who, under the pretence of weighing it, extracted the stone from its setting and offered the sweep three halfpence for it. The sweep refused the offer, but the goldsmith refused to return the stone. The goldsmith was held liable in trover and ordered to return the stone to the sweep or, failing that, to pay him a sum equivalent to what a stone of that size “of the finest water” would be worth. Cases in which the finder loses C4: Moffatt v. Kazana [1969] 2 QB 152 A1: the finder was the occupier of the premises where the chattel was found. A2: the chattel was not attached. A3: the other claimant was not the owner of the premises where the chattel was found. A4: the other claimant was the true owner of the chattel or was claiming through the rights of the true owner. A5: the finder did not hand over the chattel to the other claimant after the finding. A6: neither party relied on the terms of an agreement regarding the right to the chattel. A7: the finder was not a servant of the other claimant. A8: the chattel was hidden or was in a position so as to be difficult to find. A9: an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned. A10: one of the parties knew of the existence of the chattel prior to the finding. In Moffatt v. Kazana, 6 a 1967 decision of the Queen’s Bench Division of the English High Court, the occupant of a house found a biscuit tin which contained a large number of banknotes. The tin was discovered during the course of some work in the kitchen of the house. To do that work it was necessary to dislodge some bricks from a point at which the main kitchen chimney joined a false flue. The tin fell out when the bricks were dislodged. 5(1722) 1 Str 505. 6[1969] 2 QB 152. 212 Appendix A: Case law specifications It emerged from the evidence that the defendant occupant had purchased the house from the previous plaintiff owner who had secreted the tin of notes during the time before the current owner had occupied the house. In selling the bunga­ low, the seller had wholly forgotten the existence of the tin and the buyer was, of course, unaware of its existence. The Court held that the conveyance of the house and land did not suffice to convey the chattels in the house, a consequence of s. 62 of the governing Law of Property Act 1965 (UK). Consequently, the plaintiff remained the true owner of the tin of notes and was entitled to prevail. Although the decision depends upon the provisions of a specific statute, it is clear that the true owners of chattels may displace the prima facie title of the owner of the land and the possessory title of any finder. C5: City of London Corporation v. Appleyard (1) [1963] 1 WLR 982 (“London v. Ap­ pleyard (1)”) A1: the finder was not the occupier of the premises where the chattel was found. A2: the chattel was attached. A3: the other claimant was not the owner of the premises where the chattel was found. A4: the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner. A5: the finder did not hand over the chattel to the other claimant after the finding. A6: neither party relied on the terms of an agreement regarding the right to the chattel. A7: the finder was a servant of the other claimant. A8: the chattel was hidden or was in a position so as to be difficult to find. A9: an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned. A10: neither party knew of the existence of the chattel prior to the finding. In City of London Corporation v. Appleyard (1), 7 a 1963 decision of the Queen’s Bench Division of the English High Court, workmen employed by Wates Ltd were engaged in cutting a key-way into a cellar wall for the purposes of securing a foundation when they found an old wall-safe built into a recess of the old wall. Inside was a wooden box which contained a large number of Bank of England notes. The notes were handed over to the City of London police who sought interpleader proceedings to determine who was entitled to the possession of the notes. 7[1963] 1 WLR 982. 213 � A.2 The FINDER specification Wates Ltd was an independent contractor engaged by Yorkwin Investments Ltd for a construction project. Yorkwin was lessee in possession of the property which was owned in fee simple by the City of London. The Court followed the decision in South Staffordshire Water Co. v. Sharman8 in holding that the occupier is, in the absence of a better title elsewhere, entitled to the possession of objects which are attached to or under the land. Consequently, since the notes were in a wooden box within a safe built into the wall of the old building, the safe formed part of the demised premises. Yorkwin, being in lawful possession of the premises, were in de facto possession of the safe, even though ignorant of its existence. Although Yorkwin was entitled to possession as against the finders, they in turn were displaced by the City of London which relied successfully on a term in the lease which granted them the right to certain objects found on the premises. C6: City of London Corporation v. Appleyard (2) [1963] 1 WLR 982 (“London v. Ap­ pleyard (2)”) A1: the finder was the occupier of the premises where the chattel was found. A2: the chattel was attached. A3: the other claimant was the owner of the premises where the chattel was found. A4: the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner. A5: the finder handed over the chattel to the other claimant after the finding. A6: one of the parties relied on the terms of an agreement made with the other which purported to give her/him the right to the chattel. A7: the finder was not a servant of the other claimant. A8: the chattel was hidden or was in a position so as to be difficult to find. A9: an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned. A10: neither party knew of the existence of the chattel prior to the finding. In City of London Corporation v. Appleyard (2), 9 a 1963 decision of the Queen’s Bench Division of the English High Court, workmen employed by Wates Ltd were engaged in cutting a key-way into a cellar wall for the purposes of securing a foundation when they found an old wall-safe built into a recess of the old wall. Inside was a wooden box which contained a large number of Bank of England notes. The notes were handed over to the City of London police who sought interpleader proceedings to determine who was entitled to the possession of the notes. 8[1896] 2 QB 44. 9[1963] 1 WLR 982. 214 Appendix A: Case law specifications Wates Ltd was an independent contractor engaged by Yorkwin Investments Ltd for a construction project. Yorkwin was lessee in possession of the property which was owned in fee simple by the City of London. The Court found that the safe formed part of the demised premises and that, consequently, Yorkwin was entitled to the notes as against the workmen. The lease contained a clause which purported to grant the rights to “every relic or article of antiquity rarity or value” to the City of London. The sole issue was to determine if the notes fell into that description. The Court could find no reason for limiting the generality of the words and so found for the City of London. C7: South Staffordshire Water Co. v. Sharman [1896] 2 QB 44 (“South Staffordshire v. Sharman”) A1: the finder was not the occupier of the premises where the chattel was found. A2: the chattel was attached. A3: the other claimant was the owner of the premises where the chattel was found. A4: the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner. A5: the finder did not hand over the chattel to the other claimant after the finding. A6: neither party relied on the terms of an agreement regarding the right to the chattel. A7: the finder was a servant of the other claimant. A8: the chattel was hidden or was in a position so as to be difficult to find. A9: an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned. A10: neither party knew of the existence of the chattel prior to the finding. In South Staffordshire Water Co. v. Sharman, 10 an 1896 decision of the Queen’s Bench Division of the English High Court, the defendant was a workman em­ ployed by the plaintiff to clean out a pool located on land owned by the plaintiff. During the operation the defendant found two gold rings embedded in the mud at the bottom of the pool. Although the plaintiff demanded the rings, the defendant refused to give them up. He placed them in the hands of police authorities who unsuccessfully endeavoured to find the owners of the rings. The police returned the rings to the defendant who was then sued in detinue for the recovery of the rings. It was proved at the trial that there was no special contract between the parties which called upon the defendant to give up any articles which might be found. Although the county court held in favour of the defendant on the basis of Bridges v. Hawkesworth, 11 the appeal found for the plaintiff on the basis that 10[1896] 2 QB 44. 11(1851) 21 LJQB 75. 215 � A.2 The FINDER specification they had, as owners of the land and pool, the right to exercise control over the same. Bridges v. Hawkesworth was distinguished on the grounds that the notes in that case were in a public part of the shop and the shopkeeper did not in any sense control them. The Court stated a general principle: where a person has possession of a house or land with a manifest intention to exercise control over it and the things which may be upon or in it, then there is a presumption that things found there are in the possession of the owner. C8: Elwes v. Brigg Gas Co. (1886) 33 ChD 562 (“Elwes v. Brigg Gas”) A1: the finder was the occupier of the premises where the chattel was found. A2: the chattel was attached. A3: the other claimant was the owner of the premises where the chattel was found. A4: the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner. A5: the finder did not hand over the chattel to the other claimant after the finding. A6: one of the parties relied on the terms of an agreement made with the other which purported to give her/him the right to the chattel. A7: the finder was not a servant of the other claimant. A8: the chattel was hidden or was in a position so as to be difficult to find. A9: an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned. A10: neither party knew of the existence of the chattel prior to the finding. In Elwes v. Brigg Gas Co., 12 an 1886 decision of the Chancery Division of the English High Court, a lessee of land found, during the course of some excavations, an ancient boat buried in the soil. The Court held that the boat belonged to the owner of the land at the time when the lease was granted since: (a) if the boat was regarded as a part of the soil or as a mineral in the soil, it was clearly a part of the land to which the plaintiff was entitled; alternatively, (b) if the boat was considered as a chattel, then the plaintiff was in lawful possession of everything that lay beneath the surface. Since a trespasser could not have taken possession of the boat, it followed that only the original owner could have a better title, but obviously such rights could no longer be established. It followed from the above that the defendant’s claim must rest on the terms of the lease. Although the lease contemplated the excavations which were done, it was silent as to what was to be done with the soil excavated. It was impossible to imply a term into the contract which would give the defendant lessee the rights to the boat. 12(1886) 33 ChD 562. 216 c Appendix A: Case law specifications A.3 The AUTHORIZATION specification Hierarchy Court 1 three judges of the High Court of Australia 2 the Full Court of the Supreme Court of New South Wales 3 the Supreme Court of New South Wales the Supreme Court of Victoria 4 the Judicial Committee of the Privy Council 5 the English Court of Appeal 6 the Chancery Division of the English High Court Authorization area Attributes Case c Result A1 A2 A3 A4 A5 A6 A7 C1 × × • • × • × 1 C2 × • × • • 2 AuthC3 × × • • × • • 3 C4 × • × • × • • 4 C5 × × • × × • • 5 IAuth × • • • × • • C6 × × × × × • × 3 Not-AuthC7 × • × • × × × 5 C8 × × • × × • × 6 INot-Auth × × × × • × × LiableC9 • × × • × • • 3 ILiable • × • × • Opening The notion of authorization extends beyond the authority given to an agent. The word “authorize” should be “understood in its ordinary dictionary sense of ‘sanction, approve, and countenance.’ ”1 “[A] person who has under his control the means by which an infringement of copy­ right may be committed . . . and who makes it available to other persons, knowing, or having reason to suspect, that it is likely to be used for the purpose of committing an infringement, and omitting to take reasonable steps to limit its use to legitimate purposes, would authorize any infringement that resulted from its use.”2 1Falcon v. Famous Players Film Co. [1926] 2 KB 474 at 491, per Bankes LJ. 2University of New South Wales v. Moorhouse (1975) 133 CLR 1 at 13, per Gibbs J. 217 � A.3 The AUTHORIZATION specification Results Auth: the accused authorized the infringement. Not-Auth: the accused did not authorize the infringement. Liable: the accused is liable (directly or vicariously) for the infringement. Attributes A1: Was the infringer an employee of the accused? yes: the infringer was an employee of the accused. ⊃ Liable no: the infringer was not an employee of the accused. ⊃ Auth ∧ Not-Auth unknown: it is not known whether the infringer was an employee of the accused. If the accused had control over the infringer’s manner of doing her/his work then the infringer was an employee of the accused. If the infringer undertook to do something for the accused and had discretion as to the manner in which it was to be done then the infringer was an independent contractor to the accused, not an employee. A2: Was the infringer an independent contractor to the accused? yes: the infringer was an independent contractor to the accused. ⊃ Auth no: the infringer was not an independent contractor to the accused. ⊃ Not-Auth ∧ Liable unknown: it is not known whether the infringer was an independent contractor to the accused. If the infringer undertook to do something for the accused and had discretion as to the manner in which it was to be done then the infringer was an independent contractor to the accused. If the accused had control over the infringer’s manner of doing her/his work then the infringer was an employee of the accused, not an independent contractor. A3: Did the accused sell or hire the infringer the means of infringing? yes: the accused sold or hired the infringer the means of infringing. ⊃ Auth no: the accused did not sell or hire the infringer the means of infringing. unknown: it is not known whether the accused sold or hired the infringer the means of infringing. 218 Appendix A: Case law specifications A4: Did the accused have the power to prevent the infringement? yes: the accused had the power to prevent the infringement. ⊃ Auth ∧ Liable no: the accused did not have the power to prevent the infringement. ⊃ Not-Auth unknown: it is not known whether the accused had the power to prevent the infringement. If the accused had some control over the infringer, or the means by which the infringement was committed, then the accused had the power to prevent the infringement. A5: Did the accused take reasonable steps to avoid the infringement? yes: the accused took reasonable steps to avoid the infringement. ⊃ Not-Auth no: the accused did not take reasonable steps to avoid the infringement. ⊃ Auth ∧ Liable unknown: it is not known whether the accused took reasonable steps to avoid the infringement. Whether particular steps were reasonable depends on the facts of the case. A6: Did the accused know, or have reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done? yes: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. ⊃ Auth ∧ Liable no: the accused did not know, and had no reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. ⊃ Not-Auth unknown: it is not known whether the accused knew, or had reason to anti­ cipate or suspect, that the infringing act was to be, or was likely to be, done. Authorization requires a mental element. However, it is not necessary for the accused to have known that a particular infringing act was to occur; merely that, from her or his awareness of the circumstances, she or he recognized that it was likely that such an act might occur. A7: Was the specific infringement causally related to an incitement to infringe on the part of the accused? yes: the specific infringement was causally related to an incitement to infringe on the part of the accused. ⊃ Auth ∧ Liable 219 � A.3 The AUTHORIZATION specification no: the specific infringement was not causally related to an incitement to infringe on the part of the accused. ⊃ Not-Auth unknown: it is not known whether the specific infringement was causally related to an incitement to infringe on the part of the accused. General exhortations to infringe will not amount to authorization unless specific acts of infringement can be established. There must be some relationship creating a link or connection, however tenuous, between the authorizer and the infringer. Cases in which the accused authorized the infringement C1: University of New South Wales v. Moorhouse (1975) 133 CLR 1 (“UNSW v. Moor- house”) A1: the infringer was not an employee of the accused. A2: the infringer was not an independent contractor to the accused. A3: the accused sold or hired the infringer the means of infringing. A4: the accused had the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was not causally related to an incitement to infringe on the part of the accused. In University of New South Wales v. Moorhouse, 3 a 1975 decision of three judges of the High Court of Australia, a graduate of the University used a photocopy machine in the University library to make two copies of a story from a library copy of a book of short stories. McTiernan ACJ, Gibbs and Jacobs JJ held that the University had authorized the infringement within the meaning of s. 36(1) of the Copyright Act 1968 (Cth); it had the power to prevent infringements, but had not taken reasonable steps to prevent them.4 Gibbs J’s statement about what constitutes authorization of an infringement is quoted above. C2: Australasian Performing Right Association Ltd v. Canterbury-Bankstown League Club Ltd [1964–65] NSWR 138 (“APRA v. Canterbury-Bankstown”) A1: it is not known whether the infringer was an employee of the accused. A2: it is not known whether the infringer was an independent contractor to the accused. A3: the accused did not sell or hire the infringer the means of infringing. A4: the accused had the power to prevent the infringement. 3(1975) 133 CLR 1. 4The addition, in 1980, of s. 39a to the Copyright Act ameliorated the effect of UNSW v. Moorhouse as far as photocopying in libraries is concerned. 220 Appendix A: Case law specifications A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was causally related to an incitement to infringe on the part of the accused. In Australasian Performing Right Association Ltd v. Canterbury-Bankstown Lea­ gue Club Ltd, 5 a 1964 decision of the Full Court of the Supreme Court of New South Wales, the club engaged a dance band to play music for dances that it held at its premises. The choice of the music to be played was left to the band leader. Herron CJ, Ferguson and Asprey JJ held that whether the bandleader was an employee or an independent contractor was immaterial. “He was authorized to play and was allowed a discretion to select whatever music he liked. He was thus given a general authority to play whatever music he liked irrespective of copyright.”6 So (if he was an employee) the club was vicariously liable for—or (if he was an independent contractor) the club was liable for the authorization of—the bandleader’s breach of the Copyright Act 1912 (Cth). C3: Winstone v. Wurlitzer Automatic Phonograph Co. of Australia Pty Ltd [1946] VLR 338 (“Winstone v. Wurlitzer ”) A1: the infringer was not an employee of the accused. A2: the infringer was not an independent contractor to the accused. A3: the accused sold or hired the infringer the means of infringing. A4: the accused had the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was causally related to an incitement to infringe on the part of the accused. In Winstone v. Wurlitzer Automatic Phonograph Co. of Australia Pty Ltd, 7 a 1946 decision of the Supreme Court of Victoria, Wurlitzer installed a juke-box in a shop and had an agreement with the shop’s proprietor by which Wurlitzer maintained and repaired the machine, and supplied it with records which Wurl­ itzer selected. The juke-box played a musical work, the copyright in which was owned by Winstone. Herring CJ held that the proprietor of the shop had publicly performed the musical work and—because of nature of the agreement between Wurlitzer and the shop’s proprietor, and because Wurlitzer selected the records—Wurlitzer had 5[1964–65] NSWR 138. 6ibid. at 140, per Ferguson J, with whom Herron CJ agreed. 7[1946] VLR 338. 221 � A.3 The AUTHORIZATION specification authorized that infringing performance within the meaning of s. 1(2) of the Copy­ right Act 1911 (UK) which was in force in Australia by virtue of the Copyright Act 1912 (Cth). C4: Mellor v. Australian Broadcasting Commission [1940] AC 491 (“Mellor v. ABC ”) A1: the infringer was not an employee of the accused. A2: the infringer was an independent contractor to the accused. A3: the accused did not sell or hire the infringer the means of infringing. A4: the accused had the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was causally related to an incitement to infringe on the part of the accused. In Mellor v. Australian Broadcasting Commission, 8 a 1940 decision of the Ju­ dicial Committee of the Privy Council, Mellor and others held the sole right to perform in public in Australia some musical works arranged for performance by brass and military bands. They published and distributed advertising pamphlets which included a statement that all of their sheet music was “ ‘Free for Public Performance’ anywhere . . . We have paid for the performing rights of every piece we issue.”9 The ABC engaged bands to play some of this music, and broadcast the bands’ performances on radio. The Privy Council held that the ABC had authorized the bands to perform the musical works within the meaning of s. 1(2) of the Copyright Act 1911 (UK) which was in force in Australia by virtue of the Copyright Act 1912 (Cth). However, the ABC had not infringed the plaintiffs’ sole right to authorize public performance because the statements made in the pamphlets amounted to consent. C5: Falcon v. Famous Players Film Co. [1926] 2 KB 474 (“Falcon v. Famous Players”) A1: the infringer was not an employee of the accused. A2: the infringer was not an independent contractor to the accused. A3: the accused sold or hired the infringer the means of infringing. A4: the accused did not have the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was causally related to an incitement to infringe on the part of the accused. 8[1940] AC 491. 9ibid. at 498–9. 222 Appendix A: Case law specifications In Falcon v. Famous Players Film Co., 10 a 1926 decision of the English Court of Appeal, the author of a play assigned to Falcon the sole right to perform the play in the United Kingdom and, twenty-one years later, sold to Famous Players the film rights to the play throughout the world. Famous Players made a film of the play in America, imported it to England, and purported to let the right to exhibit it to the proprietor of a cinema. Falcon brought an action to restrain Famous Players from infringing his performing right. Famous Players denied that Falcon had such an exclusive right and claimed that, even if he had, they had not infringed it. The Court of Appeal held that Falcon did have an exclusive right to perform the play in the UK, and that Famous Players had infringed it. Scrutton LJ referred to the hiring agreement that impliedly stipulated that the cinema proprietor should exhibit. “They have imposed an obligation upon him that he shall perform, and in my view persons who do that perform themselves.”11 Hence Scrutton LJ found no need to consider whether Famous Players had au­ thorized the cinema proprietor to infringe. Bankes and Atkin LJJ held that Famous Players had authorized the infringement within the meaning of s. 1(2) of the Copyright Act 1911 (UK). Bankes LJ’s view (quoted above) that the word “authorize” should be understood in its ordinary dictionary sense of “sanction, approve, and countenance” has been adopted by most subsequent courts in the UK and in Australia. IAuth (the ideal case in which the accused authorized the infringement): A1: the infringer was not an employee of the accused. A2: the infringer was an independent contractor to the accused. A3: the accused sold or hired the infringer the means of infringing. A4: the accused had the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was causally related to an incitement to infringe on the part of the accused. Cases in which the accused did not authorize the infringement C6: RCA Corporation v. John Fairfax and Sons Ltd [1981] 1 NSWLR 251 (“RCA v. Fairfax ”) A1: the infringer was not an employee of the accused. A2: the infringer was not an independent contractor to the accused. A3: the accused did not sell or hire the infringer the means of infringing. 10[1926] 2 KB 474. 11ibid. at 495. 223 � A.3 The AUTHORIZATION specification A4: the accused did not have the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was not causally related to an incitement to infringe on the part of the accused. In RCA Corporation v. John Fairfax and Sons Ltd, 12 a 1981 decision of the Su­ preme Court of New South Wales, the Fairfax newspaper the Sun-Herald carried an article which pointed out that, using cassette tapes and good quality taping equipment, the same album can be taped by many people. It also discussed how the advent of FM radio had made it easy for people to tape new album and single releases without buying the discs: “Why spend nearly $10 on the new David Bowie album when you can tape it from 2JJJ?”13 Kearney J held that “authorization involves some element of causation—and hence the necessity for some relationship creating a link or connection however tenuous between the authorizer and the infringer.”14 There was no such link, so Fairfax had not authorized any infringement within the meaning of s. 13(2) of the Copyright Act 1968 (Cth). C7: Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd [1924] 1 KB 1 (“PRS v. Ciryl”) A1: the infringer was not an employee of the accused. A2: the infringer was an independent contractor to the accused. A3: the accused did not sell or hire the infringer the means of infringing. A4: the accused had the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused did not know, and had no reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was not causally related to an incitement to infringe on the part of the accused. In Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd, 15 a 1923 de­ cision of the English Court of Appeal, the syndicate was the lessee of a theatre. The managing-director of the syndicate produced a play at that theatre, and engaged a band to perform at the theatre under the direction of a bandmaster. In the absence of the managing-director, and without his knowledge, the band performed works the copyright in which was owned by the Performing Right Society. 12[1981] 1 NSWLR 251. 13ibid. at 252. 14ibid. at 259. 15[1924] 1 KB 1. 224 Appendix A: Case law specifications Bankes, Scrutton and Atkin LJJ held that the managing-director had not author­ ized the infringing performances, within the meaning of s. 1(2) of the Copyright Act 1911 (UK), because the infringement occurred without his knowledge and he had no reason to anticipate or suspect that the band was likely to give per­ formances which would breach copyright. C8: A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd [1979] FSR 1 (“A&M v. Audio Magnetics”) A1: the infringer was not an employee of the accused. A2: the infringer was not an independent contractor to the accused. A3: the accused sold or hired the infringer the means of infringing. A4: the accused did not have the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was not causally related to an incitement to infringe on the part of the accused. In A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd, 16 a 1978 decision of the Chancery Division of the English High Court, A&M Records and twenty-three others alleged that Audio Magnetics was inciting the public to infringe their copyright in sound recording by advertising blank cassette tapes. Foster J held that there was no “particular specific authorisation”;17 there was not sufficient causal relationship between the alleged authorization and the actual breach. “It was not sufficient to allege authorisation at large. Authorisation meant sanctioning, express approval or countenancing of an actual breach of copyright by some act directly related to that breach.”18 INot-Auth (the ideal case in which the accused did not authorize the infringement): A1: the infringer was not an employee of the accused. A2: the infringer was not an independent contractor to the accused. A3: the accused did not sell or hire the infringer the means of infringing. A4: the accused did not have the power to prevent the infringement. A5: the accused took reasonable steps to avoid the infringement. A6: the accused did not know, and had no reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was not causally related to an incitement to infringe on the part of the accused. 16[1979] FSR 1. 17ibid. at 10. 18ibid. at 2. 225 � A.3 The AUTHORIZATION specification Cases in which the accused is liable (directly or vicariously) for the infringement C9: Australasian Performing Right Association Ltd v. Miles [1962] NSWR 405 (“APRA v. Miles”) A1: the infringer was an employee of the accused. A2: the infringer was not an independent contractor to the accused. A3: the accused did not sell or hire the infringer the means of infringing. A4: the accused had the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: the specific infringement was causally related to an incitement to infringe on the part of the accused. In Australasian Performing Right Association Ltd v. Miles, 19 a 1961 decision of the Supreme Court of New South Wales, the Dee Why RSL Club engaged a band to play at a dance held at the club. During the dance the band played I’ve Got a Lovely Bunch of Coconuts, the copyright in which was owned by the Australasian Performing Right Association. Jacobs J held that the members of the band were servants of the club, because “the club through its officers was exercising a control over the work performed in such a way as to show that there was an authority to command the orchestra in its performance.”20 So the members of the club, through the band, performed the musical work and infringed the copyright under s. 2(1) of the Copyright Act 1911 (UK) which was in force in Australia by virtue of the Copyright Act 1912 (Cth). ILiable (the ideal case in which the accused is liable (directly or vicariously) for the infringement): A1: the infringer was an employee of the accused. A2: the infringer was not an independent contractor to the accused. A3: it is not known whether the accused sold or hired the infringer the means of infringing. A4: the accused had the power to prevent the infringement. A5: the accused did not take reasonable steps to avoid the infringement. A6: the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. A7: it is not known whether the specific infringement was causally related to an incitement to infringe on the part of the accused. 19[1962] NSWR 405. 20ibid. at 407. 226 c Appendix A: Case law specifications A.4 The EMPLOYEE specification Hierarchy Court 1 five judges of the High Court of Australia 2 four judges of the High Court of Australia 3 three judges of the High Court of Australia 4 a single judge of the High Court of Australia 5 three judges of the Federal Court of Australia 6 the Judicial Committee of the Privy Council 7 the English Court of Appeal 8 the King’s Bench Division of the English High Court the Queen’s Bench Division of the English High Court Employee area Attributes Case c Result A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A12 A13 A14 A15 A16 A17 A18 C1 × • • × • • • × × × • • × × • × × × 1 C2 • • • × • × × × • × × × × × × × × × 2 C3 • × • × • • • × × • × • × × × × × 4 C4 Employee • × • • × • • × × × × × × × × × × 5 C5 7• × • × • × • × × × • • × × × × × • C6 × • • × • × × × × • × 7 C7 • • • • × • • × × • • • × • × × × × 8 IEmployee • × • × • • • × × • • • × • • • • C8 × • × • × × × × • × × × × × × × × × 3 C9 × • × • × × • • • × × × × × × × • × 3 C10 × • • • × × × × • × × × × × • × × × 5 C11 6 Contractor × • • • × × × • • • × × × × × × × • C12 7• • • × • • × • • × • • × × × • × • C13 × • × × • × × × × • • × 7 C14 8 × • • • × × × × • • × × × × × × × • IContractor × • × • × × • • × × × • × × × • Opening The law distinguishes between a contract of service (between employer and employee) and a contract for services (between principal and independent contractor). This dis­ tinction affects the terms that will be implied in the absence of an express agreement, the liability of the employer to third parties, the applicability of industrial awards, the applicability of statutes which may affect workers’ compensation, occupational health and safety, long-service leave, fringe benefits tax, etc. The terms “employer” and “worker” are used here to mean “employer” and “employee” (in the case of a contract of service) or “principal” and “independent contractor” (in the case of a contract for services). 227 � A.4 The EMPLOYEE specification Results Employee: the worker is an employee. Contractor: the worker is an independent contractor. Attributes A1: Did the employer direct not only what work was to be done, but also the manner in which it was to be done? yes: the employer directed the manner in which the work was to be done. ⊃ Employee no: the employer did not direct the manner in which the work was to be done. ⊃ Contractor unknown: it is not known whether the employer directed the manner in which the work was to be done. If the employer had a right of control over how the worker did the work then the employer had the power to direct not only what work was to be done, but also the manner in which it was to be done. A2: Was the worker allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand? yes: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. ⊃ Contractor no: the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. ⊃ Employee unknown: it is not known whether the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: Was the worker an integral part of the employer’s business? yes: the worker was an integral part of the employer’s business. ⊃ Employee no: the worker was not an integral part of the employer’s business, but was accessory to it. ⊃ Contractor unknown: it is not known whether the worker was an integral part of the em­ ployer’s business or was merely accessory to it. If the worker was “part and parcel” of the employer’s business then she/he was an integral part of the business, not merely accessory to it. 228 Appendix A: Case law specifications A4: Did the worker own the tools or provide the transport with which she/he performed the work? yes: the worker owned the tools or provided the transport with which she/he performed the work. ⊃ Contractor no: the worker neither owned the tools nor provided the transport with which she/he performed the work. ⊃ Employee unknown: it is not known whether the worker owned the tools or provided the transport with which she/he performed the work. A5: Would the employer make a profit/loss if the work performed by the worker cost less/more than expected? yes: the employer would make a profit/loss if the work performed by the worker cost less/more than expected. no: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected. unknown: it is not known whether the employer would make a profit/loss if the work performed by the worker cost less/more than expected. A6: Was the work performed on the employer’s premises? yes: the work was performed on the employer’s premises. no: the work was not performed on the employer’s premises. ⊃ Contractor unknown: it is not known whether the work was performed on the employer’s premises. A7: Did the employer supervise or inspect the work? yes: the employer supervised or inspected the work. ⊃ Employee no: the employer neither supervised nor inspected the work. ⊃ Contractor unknown: it is not known whether the employer supervised or inspected the work. A8: Was the worker in business on her/his own account? yes: the worker was in business on her/his own account. ⊃ Contractor no: the worker was not in business on her/his own account. ⊃ Employee unknown: it is not known whether the worker was in business on her/his own account. 229 � A.4 The EMPLOYEE specification A9: Was the worker allowed to employ others to assist with her/his work? yes: the worker was allowed to employ others to assist with her/his work. ⊃ Contractor no: the worker was not allowed to employ others to assist with her/his work. ⊃ Employee unknown: it is not known whether the worker was allowed to employ others to assist with her/his work. A10: Was the worker obliged to work only for the employer? yes: the worker was obliged to work only for the employer. ⊃ Employee no: the worker was not obliged to work only for the employer. ⊃ Contractor unknown: it is not known whether the worker was obliged to work only for the employer. A11: Was the worker required to work at specified times? yes: the worker was required to work at specified times. ⊃ Employee no: the worker was not required to work at specified times. ⊃ Contractor unknown: it is not known whether the worker was required to work at specified times. A12: Did the employer pay the worker by time? yes: the employer paid the worker by time. ⊃ Employee no: the employer did not pay the worker by time. ⊃ Contractor unknown: it is not known whether the employer paid the worker by time. The employer could pay the worker by time (e.g. by the hour, or by the week) or by results. A13: Was the money that the employer paid to the worker stated to be a “fee”? yes: the money that the employer paid to the worker was stated to be a “fee”. ⊃ Contractor no: the money that the employer paid to the worker was not stated to be a “fee”. ⊃ Employee unknown: it is not known whether the money that the employer paid to the worker was stated to be a “fee”. 230 Appendix A: Case law specifications A14: Was the money that the employer paid to the worker stated to be “wages” or “salary”? yes: the money that the employer paid to the worker was stated to be “wages” or “salary”. ⊃ Employee no: the money that the employer paid to the worker was not stated to be “wages” or “salary”. ⊃ Contractor unknown: it is not known whether the money that the employer paid to the worker was stated to be “wages” or “salary”. A15: Did the employer deduct PAYE tax instalments from the worker’s pay? yes: the employer deducted PAYE tax instalments from the worker’s pay. ⊃ Employee no: the employer did not deduct PAYE tax instalments from the worker’s pay. ⊃ Contractor unknown: it is not known whether the employer deducted PAYE tax instal­ ments from the worker’s pay. A16: Did the employer pay the worker sick pay or holiday pay? yes: the employer paid the worker sick pay or holiday pay. ⊃ Employee no: the employer paid the worker neither sick pay nor holiday pay. ⊃ Contractor unknown: it is not known whether the employer paid the worker sick pay or holiday pay. A17: Did the employer and the worker express an intention that the relationship would be one of employer and employee? yes: the employer and the worker expressed an intention that the relationship would be one of employer and employee. ⊃ Employee no: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. unknown: it is not known whether the employer and the worker expressed an intention that the relationship would be one of employer and employee. For example, if the employer and the worker characterized their agreement as being a “contract of service,” that would be an expression of an intention that the relationship would be one of employer and employee. 231 � A.4 The EMPLOYEE specification A18: Did the employer and the worker express an intention that the relationship would be one of principal and independent contractor? yes: the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. ⊃ Contractor no: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. unknown: it is not known whether the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. For example, if the employer and the worker characterized their agreement as being a “contract for services,” that would be an expression of an intention that the relationship would be one of principal and independent contractor. Cases in which the worker is an employee C1: Zuijs v. Wirth Brothers Pty Ltd (1955) 93 CLR 561 (“Zuijs v. Wirth Brothers”) A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker neither owned the tools nor provided the transport with which she/he performed the work. A5: the employer would make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was performed on the employer’s premises. A7: the employer supervised or inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was not allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was required to work at specified times. A12: the employer paid the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer deducted PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. 232 Appendix A: Case law specifications In Zuijs v. Wirth Brothers Pty Ltd, 1 a 1955 decision of five judges of the High Court of Australia, Zuijs was an acrobat who fell during a trapeze act at one of Wirth Brothers’ circuses. He sought compensation under the Worker’s Compens­ ation Act 1926 (NSW), claiming to be an employee of Wirth Brothers. Wirth Brothers claimed that, because of the high degree of skill and personal judgment that he had to exercise in his work, Zuijs was an independent contractor and therefore not entitled to compensation. The High Court unanimously agreed with Zuijs. “Even if [Wirth Brothers] could not interfere in the actual technique of the acrobats and in the character of the act, no reason appears why the appellant should not be subject to his directions in all other respects . . . There are countless examples of highly specialized functions in modern life that must as a matter of practical necessity and sometimes even as a matter of law be performed on the responsibility of persons who possess particular knowledge and skill and who are accordingly qualified. But those engaged to perform the functions may nevertheless work under a contract of service.”2 C2: Cam and Sons Pty Ltd v. Sargent (1940) 14 ALJ 162 (“Cam v. Sargent”) A1: the employer directed the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker neither owned the tools nor provided the transport with which she/he performed the work. A5: the employer would make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was not performed on the employer’s premises. A7: the employer neither supervised nor inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was not required to work at specified times. A12: the employer did not pay the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. 1(1955) 93 CLR 561. 2ibid. at 571–2, per Dixon CJ, Williams, Webb and Taylor JJ. 233 � A.4 The EMPLOYEE specification A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. In Cam and Sons Pty Ltd v. Sargent, 3 a 1940 decision of four judges of the High Court of Australia, Sargent was the master of a ship. He entered into an agreement with Cam and Sons that claimed that the ship was hired by Cam and Sons to Sargent and his fellow contractors (called “the partnership”). However, it was doubtful whether that agreement actually deprived Cam and Sons of any control over the ship. The partnership was to use the ship only to carry coal from Swansea to Sydney. Cam and Sons were sole agents of the partnership for securing cargoes for the ship, and for collecting money due to the partnership. The partnership paid nothing for the “hire” of the ship, but received a specified sum for each return trip of a certain tonnage plus (in certain circumstances) 5% of the earnings, the balance of which was retained by Cam and Sons. Cam and Sons had to approve people employed by the partnership. Sargent claimed that he (and others in the partnership) were employed by Cam and Sons, and therefore came within the terms of an industrial award. Cam and Sons claimed that members of the partnership were independent contractors. The High Court unanimously agreed with Sargent. Rich J came to the conclusion that the agreement was an attempt to evade the terms of the industrial award.4 C3: Federal Commissioner of Taxation v. J. Walter Thompson (Australia) Pty Ltd (1944) 69 CLR 227 (“FCT v. Thompson”) A1: the employer directed the manner in which the work was to be done. A2: the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker neither owned the tools nor provided the transport with which she/he performed the work. A5: the employer would make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was performed on the employer’s premises. A7: the employer supervised or inspected the work. A8: it is not known whether the worker was in business on her/his own account. 3(1940) 14 ALJ 162. 4ibid. at 163. 234 Appendix A: Case law specifications A9: the worker was not allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was required to work at specified times. A12: the employer did not pay the worker by time. A13: the money that the employer paid to the worker was stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. In Federal Commissioner of Taxation v. J. Walter Thompson (Australia) Pty Ltd, 5 a 1944 decision of a single judge of the High Court of Australia, the FCT claimed that payments made to radio artists by Thompson were “wages” within the meaning of the Pay-roll Tax Assessment Act 1941 (Cth) and therefore tax­ able. The artists were selected by a producer and paid to appear in radio plays. They were paid a “fee” for each performance, but were paid nothing for attend­ ing (compulsory) rehearsals. Thompson claimed that the artists were presumed to know their work and “to render services in the same manner as a professional man, such as a surgeon or an architect, not being subject . . . to detailed control as to the manner in which those services are to be performed.”6 Hence, Thompson claimed, they were independent contractors. Latham CJ held that the radio actors were employed “to co-operate with others in a team under the control of the producer to bring about a result, the details of which must in great measure be determined by the producer.”7 Hence the artists were employed by Thompson; the fee they were paid was subject to payroll tax. C4: Australian Timber Workers Union v. Monaro Sawmills Pty Ltd (1980) 29 ALR 322 (“ATWU v. Monaro Sawmills”) A1: the employer directed the manner in which the work was to be done. A2: the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker owned the tools or provided the transport with which she/he performed the work. 5(1944) 69 CLR 227. 6ibid. at 231. 7ibid. at 232. 235 � A.4 The EMPLOYEE specification A5: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was performed on the employer’s premises. A7: the employer supervised or inspected the work. A8: the worker was not in business on her/his own account. A9: it is not known whether the worker was allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was not required to work at specified times. A12: the employer did not pay the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. In Australian Timber Workers Union v. Monaro Sawmills Pty Ltd, 8 a 1980 de­ cision of three judges of the Federal Court of Australia, Wales was a tree feller who cut timber exclusively for Monaro Sawmills. He performed his work in an area allotted to him by Monaro Sawmills. He, and other fellers, were paid by the amount of millable wood they cut. Wales provided his own tools and transport, but was (with the other fellers) covered by Monaro Sawmill’s workers’ compens­ ation policy. The union sought an order that a penalty be imposed on Monaro Sawmills for breaching the Timber Industries Consolidated Award 1974 by failing to pay Wales money in lieu of annual leave. Monaro Sawmills claimed that Wales was an independent contractor, and so was not subject to the award. Sweeney and Evatt JJ examined the circumstances of Wales’s employment and held that those circumstances clearly pointed to the existence of a relationship of employer and employee. They could not see “any sense in which it could be said that Wales was conducting some sort of business of his own.”9 8(1980) 29 ALR 322. 9ibid. at 329. 236 Appendix A: Case law specifications C5: Ferguson v. John Dawson & Partners (Contractors) Ltd [1976] 1 WLR 1213 (“Fer­ guson v. Dawson”) A1: the employer directed the manner in which the work was to be done. A2: the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker neither owned the tools nor provided the transport with which she/he performed the work. A5: the employer would make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was not performed on the employer’s premises. A7: the employer supervised or inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was not allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was required to work at specified times. A12: the employer paid the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. In Ferguson v. John Dawson & Partners (Contractors) Ltd, 10 a 1976 decision of the English Court of Appeal, Ferguson fell off a roof while removing some scaffolding boards. He claimed damages against John Dawson (the building contractors) for breach of statutory duty relying on the Construction (Working Places) Regulations 1966 (UK). This duty would only be owed if Ferguson was an employee of John Dawson. Megaw and Browne LJJ held that, despite the fact that both parties labelled Ferguson a “self-employed labour only subcontractor”,11 the reality of the rela­ tionship between them was that of employer/employee. 10[1976] 1 WLR 1213. 11ibid. at 1219, per Megaw LJ; at 1225, per Lawton LJ; at 1228, per Browne LJ. 237 � A.4 The EMPLOYEE specification C6: Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (2) [1952] 1 TLR 101 (“Stevenson v. Macdonald (2)”) A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker neither owned the tools nor provided the transport with which she/he performed the work. A5: the employer would make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was not performed on the employer’s premises. A7: it is not known whether the employer supervised or inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was not allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was required to work at specified times. A12: it is not known whether the employer paid the worker by time. A13: it is not known whether the money that the employer paid to the worker was stated to be a “fee”. A14: it is not known whether the money that the employer paid to the worker was stated to be “wages” or “salary”. A15: it is not known whether the employer deducted PAYE tax instalments from the worker’s pay. A16: it is not known whether the employer paid the worker sick pay or holiday pay. A17: it is not known whether the employer and the worker expressed an intention that the relationship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. In Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (2), 12 a 1952 de­ cision of the English Court of Appeal, Evans-Hemming was an accountant who had been employed (first as a servant, then as an executive officer) by Macdonald and Evans. Shortly after he left them, he wrote a textbook on business manage­ ment and submitted the manuscript to Stevenson Jordon and Harrison (a firm of publishers). He died before the book was published. Macdonald and Evans claimed that the book was written while Evans-Hemming was their employee, and so they owned the copyright in the work under s. 5(1)(b) of the Copyright Act 1911 (UK). 12[1952] 1 TLR 101. 238 Appendix A: Case law specifications The book was divided into five sections. The second section was written in its final form while Evans-Hemming was employed by Macdonald and Evans. The Court of Appeal held that he wrote the second section as an employee, and hence the copyright in the second section was in Macdonald and Evans. C7: Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd [1924] 1 KB 762 (“PRS v. Palais de Danse”) A1: the employer directed the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker owned the tools or provided the transport with which she/he performed the work. A5: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was performed on the employer’s premises. A7: the employer supervised or inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was not allowed to employ others to assist with her/his work. A10: the worker was obliged to work only for the employer. A11: the worker was required to work at specified times. A12: the employer paid the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. In Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd, 13 a 1924 decision of the King’s Bench Division of the English High Court, the defendants were the occupiers of a dance hall. They engaged a band to provide music in the hall. The agreement provided that the band should not infringe 13[1924] 1 KB 762. 239 � A.4 The EMPLOYEE specification copyright, and that the band would be liable for damages and costs caused by any such infringement. There was also a notice displayed in the hall stating that “[o]nly such music as may be played without fee or licence is allowed to be played in this Hall.”14 The band performed several pieces of music, the copyright in which was held by the Performing Right Society, without its permission. The defendants did not know, and had no reasonable grounds for suspecting, that the infringement was to take place. The PRS abandoned its earlier claim that the defendants had “permitted” the infringement under s. 2(3) of the Copyright Act 1911 (UK). However, it claimed that the band were the defendants’ employees, and so the defendants were vi­ cariously liable for the infringement. McCardie J examined the agreement and found that it gave “to the defendants the right of continuous, dominant, and detailed control on every point, including the nature of the music to be played”.15 Hence the band members were employees of the defendant company, which was liable for the infringement. IEmployee (the ideal case in which the worker is an employee): A1: the employer directed the manner in which the work was to be done. A2: the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker neither owned the tools nor provided the transport with which she/he performed the work. A5: the employer would make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was performed on the employer’s premises. A7: the employer supervised or inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was not allowed to employ others to assist with her/his work. A10: the worker was obliged to work only for the employer. A11: the worker was required to work at specified times. A12: the employer paid the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was stated to be “wages” or “salary”. A15: the employer deducted PAYE tax instalments from the worker’s pay. A16: the employer paid the worker sick pay or holiday pay. 14ibid. at 764. 15ibid. at 771. 240 Appendix A: Case law specifications A17: the employer and the worker expressed an intention that the relationship would be one of employer and employee. A18: it is not known whether the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. Cases in which the worker is an independent contractor C8: Humberstone v. Northern Timber Mills (1949) 79 CLR 389 (“Humberstone v. NTM ”) A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was not an integral part of the employer’s business, but was accessory to it. A4: the worker owned the tools or provided the transport with which she/he performed the work. A5: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was not performed on the employer’s premises. A7: the employer neither supervised nor inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was not required to work at specified times. A12: the employer did not pay the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. In Humberstone v. Northern Timber Mills, 16 a 1949 decision of three judges of the High Court of Australia, Humberstone carried goods for NTM. He had originally held himself out as a carrier, prepared to carry for anyone, but for over twenty years he had carried goods solely for NTM (although he would, infrequently, 16(1949) 79 CLR 389. 241 � A.4 The EMPLOYEE specification carry back-loads for NTM’s customers). Humberstone owned the truck, and paid for petrol and repairs. He was paid weekly on a weight-mileage basis. He was a licenced carrier, and had his name printed on the side of his truck with the description “carrier.” On the way back from a job, he had a puncture. He went home to change the wheel, but exerted himself so strenuously in trying to remove the tyre from the wheel that he became ill and later lapsed into a coma, from which he did not recover. Section 3 of the Worker’s Compensation Act 1928 (Vic) had been amended about a year before Humberstone’s death so as to include independent contractors in its definition of a “worker” covered by the Act. However, the High Court held that the amendment applied only to contracts entered into after it came into operation. Further, the Court decided that Humberstone was not an employee of NTM. Hence he was not a “worker” under the Act and his widow was not entitled to compensation under the Act. C9: Queensland Stations Pty Ltd v. Federal Commissioner of Taxation (1945) 70 CLR 539 (“Queensland Stations v. FCT ”) A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was not an integral part of the employer’s business, but was accessory to it. A4: the worker owned the tools or provided the transport with which she/he performed the work. A5: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was not performed on the employer’s premises. A7: the employer supervised or inspected the work. A8: the worker was in business on her/his own account. A9: the worker was allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was not required to work at specified times. A12: the employer did not pay the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker expressed an intention that the relationship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. 242 Appendix A: Case law specifications In Queensland Stations Pty Ltd v. Federal Commissioner of Taxation, 17 a 1945 decision of three judges of the High Court of Australia, agreements were entered into between Queensland Stations and some drovers. The agreements stated that the drovers would “serve” Queensland Stations and take charge of a spe­ cified number of cattle, and deliver them to a specified place. The drovers were paid a specified rate per head of cattle successfully delivered. Each drover was responsible for hiring help, and paying for feed for the cattle. The drovers were to “obey and carry out all lawful instructions and to use the whole of [their] time, energy and ability in the careful droving of the stock.”18 The FCT claimed that payments made to drovers were “wages” within the meaning of the Pay-roll Tax Assessment Act 1941 (Cth), and that Queensland Stations was liable to payroll tax. The High Court held that the drovers were independent contractors, so the pay­ ments were not “wages”. Rich J pointed out that drovers were traditionally free from the control of owners of cattle. “The obligation imposed on the drover to obey and carry out all lawful instructions is not a reservation of detailed control and possession having regard to the terms of the agreement as a whole.”19 C10: Price v. Grant Industries Pty Ltd (1978) 21 ALR 388 (“Price v. Grant Industries”) A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker owned the tools or provided the transport with which she/he performed the work. A5: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was not performed on the employer’s premises. A7: the employer neither supervised nor inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was not required to work at specified times. A12: the employer did not pay the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. 17(1945) 70 CLR 539. 18ibid. at 540 19ibid. at 549. 243 � A.4 The EMPLOYEE specification A15: the employer deducted PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. In Price v. Grant Industries Pty Ltd, 20 a 1978 decision of three judges of the Federal Court of Australia, Grant Industries manufactured and sold wardrobes, which Price (and others) delivered and installed. Price and each of the other “contractors” (as Grant Industries called them) had to provide and maintain a suitable truck to deliver the wardrobes, and provide the tools required to install them. Price sought an order that a penalty be imposed on Grant Industries for breaching the Furnishing Trades (Consolidated) Award 1975 by not paying Price the appropriate rate of wages, and not giving him annual leave. The award only applied to “employees” of specified employers. The Federal Court examined the facts, and the provisions of the agreement, and held that Price was an independent contractor and, therefore, not subject to the award. C11: Australian Mutual Provident Society v. Chaplin (1978) 18 ALR 385 (“AMP v. Chaplin”) A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker owned the tools or provided the transport with which she/he performed the work. A5: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was not performed on the employer’s premises. A7: the employer neither supervised nor inspected the work. A8: the worker was in business on her/his own account. A9: the worker was allowed to employ others to assist with her/his work. A10: the worker was obliged to work only for the employer. A11: the worker was not required to work at specified times. A12: the employer did not pay the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. 20(1978) 21 ALR 388. 244 Appendix A: Case law specifications A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. In Australian Mutual Provident Society v. Chaplin, 21 a 1978 decision of the Ju­ dicial Committee of the Privy Council, Chaplin was a representative of AMP. A clause of the agreement between them stated that the relationship was one of “principal and agent” and not one of “master and servant”. Chaplin claimed that he was employed under a contract of service, and therefore a “worker” under the Long Service Leave Act, 1967 (SA) and entitled to certain benefits. The Privy Council found that there was no reason to think that the clause was not a genuine statement of the parties’ intentions. Examining the agreement, their lordships concluded that it provided for a contract of agency. The fact that Chaplin was given the power of unlimited delegation of the whole perform­ ance of his work was “almost conclusive against the contract being a contract of service.”22 C12: Massey v. Crown Life Insurance Co. [1978] ICR 590 (“Massey v. Crown Life”) A1: the employer directed the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker neither owned the tools nor provided the transport with which she/he performed the work. A5: the employer would make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was performed on the employer’s premises. A7: the employer neither supervised nor inspected the work. A8: the worker was in business on her/his own account. A9: the worker was allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was required to work at specified times. 21(1978) 18 ALR 385. 22ibid. at 391. 245 � A.4 The EMPLOYEE specification A12: the employer paid the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker sick pay or holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. In Massey v. Crown Life Insurance Co., 23 a 1978 decision of the English Court of Appeal, Massey was the manager of a branch of Crown Life. He had been an employee for two years, then he and Crown Life entered into a new agreement whereby Massey continued to perform the same duties as before, but was self- employed. This arrangement had tax advantages for Massey. After a further two years, Crown Life terminated the agreement and Massey sought compensation for unfair dismissal under the Trade Union and Labour Relations Act 1974 (UK). Compensation was only payable if Massey was employed under a contract of service. Lord Denning MR stated that “if the true relationship of the parties is that of master and servant under a contract of service, the parties cannot alter the truth of that relationship by putting a different label upon it.”24 However, he (and the rest of the Court of Appeal) held that the agreement was genuinely intended to establish Massey as being self-employed: an independent contractor. C13: Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (1) [1952] 1 TLR 101 (“Stevenson v. Macdonald (1)”) A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was not an integral part of the employer’s business, but was accessory to it. A4: the worker neither owned the tools nor provided the transport with which she/he performed the work. A5: the employer would make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was not performed on the employer’s premises. 23[1978] ICR 590. 24ibid. at 594. 246 Appendix A: Case law specifications A7: it is not known whether the employer supervised or inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was not allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was required to work at specified times. A12: it is not known whether the employer paid the worker by time. A13: it is not known whether the money that the employer paid to the worker was stated to be a “fee”. A14: it is not known whether the money that the employer paid to the worker was stated to be “wages” or “salary”. A15: it is not known whether the employer deducted PAYE tax instalments from the worker’s pay. A16: it is not known whether the employer paid the worker sick pay or holiday pay. A17: the employer and the worker expressed an intention that the relationship would be one of employer and employee. A18: the employer and the worker did not express any intention that the rela­ tionship would be one of principal and independent contractor. In Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (1), 25 a 1952 de­ cision of the English Court of Appeal, Evans-Hemming was an accountant who had been employed (first as a servant, then as an executive officer) by Macdonald and Evans. Shortly after he left them, he wrote a textbook on business manage­ ment and submitted the manuscript to Stevenson Jordon and Harrison (a firm of publishers). He died before the book was published. Macdonald and Evans claimed that the book was written while Evans-Hemming was their employee, and so they owned the copyright in the work under s. 5(1)(b) of the Copyright Act 1911 (UK). The book was divided into five sections. The first section consisted of the text of three public lectures that Evans-Hemming had given while employed by Mac­ donald and Evans. The Court of Appeal held that he had given these lectures as an independent contractor. As Denning LJ said, “under a contract of service, a man is employed as part of the business, and his work is done as an integral part of the business; whereas, under a contract for services, his work, although done for the business, is not integrated into it but is only accessory to it . . . The lectures were, in a sense, part of the services rendered by Mr Evans-Hemming for the benefit of the company. But they were in no sense part of his service. It follows that the copyright in the lectures was in Mr Evans-Hemming.”26 25[1952] 1 TLR 101. 26ibid. at 111 247 � A.4 The EMPLOYEE specification C14: Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance [1968] 2 QB 497 (“Ready Mixed v. Minister ”) A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was an integral part of the employer’s business. A4: the worker owned the tools or provided the transport with which she/he performed the work. A5: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected. A6: the work was not performed on the employer’s premises. A7: the employer neither supervised nor inspected the work. A8: the worker was not in business on her/his own account. A9: the worker was allowed to employ others to assist with her/his work. A10: the worker was obliged to work only for the employer. A11: the worker was not required to work at specified times. A12: the employer did not pay the worker by time. A13: the money that the employer paid to the worker was not stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: the employer and the worker did not express any intention that the rela­ tionship would be one of employer and employee. A18: the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. In Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance, 27 a 1967 decision of the Queen’s Bench Division of the English High Court, Latimer worked for Ready Mixed as an “owner-driver.” He was paid at mileage rates, and was obliged to buy the truck through a financial organization associated with Ready Mixed. The truck was painted in the company’s colours, and he had to wear a Ready Mixed uniform. Latimer was obliged to meet the 27[1968] 2 QB 497. 248 Appendix A: Case law specifications costs of maintenance, repair and insurance of the truck (and the attached mixing unit, which belonged to Ready Mixed). The Minister determined that Latimer was employed under a contract of service and therefore an “employed person” under s. 1(2) of the National Insurance Act 1965 (UK), making Ready Mixed liable to make weekly contributions. MacKenna J examined the contract and held that the rights it conferred, and the duties it imposed, between Latimer and Ready Mixed were not such as to make it a contract of service. IContractor (the ideal case in which the worker is an independent contractor): A1: the employer did not direct the manner in which the work was to be done. A2: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand. A3: the worker was not an integral part of the employer’s business, but was accessory to it. A4: the worker owned the tools or provided the transport with which she/he performed the work. A5: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected. A6: it is not known whether the work was performed on the employer’s premises. A7: the employer neither supervised nor inspected the work. A8: the worker was in business on her/his own account. A9: the worker was allowed to employ others to assist with her/his work. A10: the worker was not obliged to work only for the employer. A11: the worker was not required to work at specified times. A12: the employer did not pay the worker by time. A13: the money that the employer paid to the worker was stated to be a “fee”. A14: the money that the employer paid to the worker was not stated to be “wages” or “salary”. A15: the employer did not deduct PAYE tax instalments from the worker’s pay. A16: the employer paid the worker neither sick pay nor holiday pay. A17: it is not known whether the employer and the worker expressed an intention that the relationship would be one of employer and employee. A18: the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. 249 c � A.5 The NATURAL specification A.5 The NATURAL specification Hierarchy Court 1 seven judges of the High Court of Australia 2 five judges of the High Court of Australia 3 three judges of the High Court of Australia 4 the Full Court of the Federal Court of Australia 5 the New South Wales Court of Appeal 6 the Supreme Court of New South Wales 7 the Judicial Committee of the Privy Council 8 the House of Lords 9 the Chancery Division of the English High Court Natural area Attributes Case c Result A1 A2 A3 A4 A5 A6 A7 C1 • • × • • × × 1 C2 • • × • • × × 2 C3 • • • • × × × 2 C4 • • • • • × × 2 C5 • • • • × • × 3 Implied C6 • • • × • • × 4 C7 • • • × • × × 5 C8 • • • • × × 7 C9 • • × × • × × 7 IImplied • • • • × × × C10 • • • × • × × 2 C11 • × × × × × × 2 C12 × × × × • × × 4 C13 • • × × • × × 6 Not-Implied C14 • • × × • × • 8 C15 • • • • × × × 9 INot-Implied × × × × • • • Opening In recent years courts have tended to imply a duty to observe the principles of natural justice. It has been said that “[t]he law has now developed to a point where it may be accepted that there is a common law duty to act fairly, in the sense of according pro­ cedural fairness, in the making of administrative decisions which affect rights, interests 250 Appendix A: Case law specifications and legitimate expectations, subject only to the clear manifestation of a contrary stat­ utory intention.”1 However, there are some circumstances in which a duty to observe natural justice will not be implied: “the law has not yet reached the stage of applying the obligation of natural justice to every decision which disadvantages individuals.”2 Results Implied: a duty to observe natural justice is implied. Not-Implied: a duty to observe natural justice is not implied. Attributes A1: ∀ Affected area yes: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. � Affected ⊃ Implied no: the decision did not affect the property, right, interest, status, or legitimate expectation of the applicant. � Unaffected ⊃ Not-Implied unknown: it is not known whether the decision affected the property, right, interest, status, or legitimate expectation of the applicant. A2: Is the decision apt to have a discrete impact on the interests of the applicant? yes: the decision is apt to have a discrete impact on the interests of the applicant. ⊃ Implied no: the decision is not apt to have a discrete impact on the interests of the applicant. ⊃ Not-Implied unknown: it is not known whether the decision is apt to have a discrete impact on the interests of the applicant. If the decision affects the applicant differently to the way in which it affects others then the decision has a discrete impact on the interest of the applicant. Note that the applicant must suffer the detriment as a direct and immediate effect of the decision, not as a contingent result. 1Kioa v. West (1985) 159 CLR 550 at 584, per Mason J. 2Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd (1987) 75 ALR 218 at 251, per Wilcox J. 251 � A.5 The NATURAL specification A3: Is the power of a nature that would suggest that procedural fairness would be applied? yes: the power is of a nature that would suggest that procedural fairness would be applied. ⊃ Implied no: the power is of a nature that would suggest that procedural fairness would not be applied. ⊃ Not-Implied unknown: it is not known whether the power is of a nature that would suggest that procedural fairness would be applied. Some prerogative powers by their nature suggest that procedural fairness would not apply in their exercise (e.g. international relations, security, defence, emer­ gency). The exercise of a high-level policy-making power, a broad unfettered discretion, or (in some circumstances) a power which is recommendatory only, suggests that procedural fairness would not apply. However, the exercise of an administrative power or discretion that has been curtailed by statute suggests that procedural fairness would apply. A4: Did the statutory or factual criteria focus on matters which were discrete to the interests of the applicant? yes: the statutory or factual criteria focused on matters which were discrete to the interests of the applicant. ⊃ Implied no: the statutory or factual criteria focused on matters of policy or public in­ terest. ⊃ Not-Implied unknown: it is not known whether the statutory or factual criteria focused on matters which were discrete to the interests of the applicant, or on matters of policy or public interest. The decisional criteria are of two kinds: the spectrum of considerations to which the decision maker was authorized to have regard (the statutory criteria), and the specific considerations to which regard was had in fact (the factual criteria). Either set of criteria can focus on matters which are discrete to the interests of the applicant, or on matters of policy or public interest. A5: Was the decision-maker a high-level policy-maker? yes: the decision-maker was a high-level policy-maker. ⊃ Not-Implied no: the decision-maker was not a high-level policy-maker. ⊃ Implied unknown: it is not known whether the decision-maker was a high-level policy- maker. Ministers, members of Cabinet, Governors and the Governor-General are high- level policy-makers. 252 Appendix A: Case law specifications A6: Is there a statutory right to appeal against the decision? yes: there is a statutory right to appeal against the decision. ⊃ Not-Implied no: there is no statutory right to appeal against the decision. ⊃ Implied unknown: it is not known whether there is a statutory right to appeal against the decision. A7: Were there circumstances which make an obligation to observe natural justice inappropriate? yes: there were circumstances which made an obligation to observe natural justice inappropriate. ⊃ Not-Implied no: there were no circumstances which would have made an obligation to observe natural justice inappropriate. unknown: it is not known whether there were circumstances which made an obligation to observe natural justice inappropriate. For example, a prompt or urgent decision may have been necessary, or they may have been national security considerations. Cases in which a duty to observe natural justice is implied C1: FAI Insurances Ltd v. Winneke (1982) 151 CLR 342 (“FAI v. Winneke”) A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would not be applied. A4: the statutory or factual criteria focused on matters which were discrete to the interests of the applicant. A5: the decision-maker was a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In FAI Insurances Ltd v. Winneke, 3 a 1982 decision of seven judges of the High Court of Australia, Winneke was the Governor of Victoria. The Workers Com­ pensation Act 1958 (Vic) made accident insurance compulsory for all employers, and required them to obtain that insurance from the Insurance Commissioner or from an insurer approved by the Governor in Council. Regulations made un­ der the Act allowed the Governor in Council to grant an approval for a period 3(1982) 151 CLR 342. 253 � A.5 The NATURAL specification not exceeding twelve months, and (on application) to renew an approval for any period not exceeding twelve months “if the Governor in Council thinks fit.” FAI had had its approval to provide worker’s compensation renewed every year for twenty years, until 1981. Gibbs CJ, Stephen, Mason, Aickin, Wilson and Brennan JJ (Murphy J dissent­ ing) held that in deciding whether to renew an approval previously given, the Governor in Council is subject to the requirements of natural justice. “In these circumstances, a company which becomes an approved insurer has a legitimate expectation that its approval will be renewed unless some good reason exists for refusing to renew it. It would not be fair to deprive a company of the ability to carry on its business without revealing the reason for doing so”.4 C2: Haoucher v. Minister of State for Immigration and Ethnic Affairs (1990) 169 CLR 648 (“Haoucher v. Minister for Immigration”) A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would not be applied. A4: the statutory or factual criteria focused on matters which were discrete to the interests of the applicant. A5: the decision-maker was a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In Haoucher v. Minister of State for Immigration and Ethnic Affairs, 5 a 1990 de­ cision of five judges of the High Court of Australia, Haoucher had been convicted of an offence for which he had been sentenced to imprisonment for a period not less than one year. Hence—as Haoucher was not an Australian citizen—the Min­ ister had the power to order his deportation under s. 12 of the Migration Act 1958 (Cth), and did so. The Minister had previously told Parliament that the Govern­ ment’s policy was that a deportee had the right to appeal to the Administrative Appeals Tribunal, and that “recommendations of the Administrative Appeals Tribunal should be overturned by the Minister only in exceptional circumstances and only when strong evidence can be produced to justify his decision.” Haoucher appealed to the AAT which recommended that the deportation order be revoked. The Minister decided not to accept that recommendation. Deane, Toohey and McHugh (Dawson and Gaudron dissenting) held that Haoucher was entitled to know the matters which constituted “exceptional circumstances” and 4ibid. at 348, per Gibbs CJ. 5(1990) 169 CLR 648. 254 Appendix A: Case law specifications “strong evidence” sufficient for the Minister to depart from the general policy of following the AAT’s recommendations. The fact that the Minister had not given Haoucher such details, and had given him no opportunity to make representa­ tions, was a denial of natural justice. C3: Annetts v. McCann (1990) 170 CLR 596 A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would be applied. A4: the statutory or factual criteria focused on matters which were discrete to the interests of the applicant. A5: the decision-maker was not a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In Annetts v. McCann, 6 a 1990 decision of five judges of the High Court of Australia, a coroner had been conducting an inquest into the death of a 16-year old boy. The boy’s parents (Mr and Mrs Annetts) sought to make a submission before the coroner made a finding. The coroner decided that the Coroner’s Act 1920 (WA) gave him the discretion (which he chose to exercise) to disallow their submission. The Annettses appealed. The High Court (Mason CJ, Brennan, Deane, Toohey and McHugh JJ) held that their son’s reputation gave the Annettses an interest in the Coroner’s inquiry. “A finding in an inquest into a death is naturally likely to deal with the conduct of the deceased leading to death. An unfavourable reflection on the deceased is usually a matter of concern to her or his parents, spouse or children and, if they choose to appear at the inquest in order to safeguard the reputation of the deceased, the familial relationship suffices, in my view, to establish the deceased’s reputation as a relevant interest which should not be adversely affected without according natural justice to those who are seeking to safeguard that reputation.”7 The Court held that the fact that the coroner’s decision was merely recommend­ atory (whether or not to prosecute) was not sufficient to avoid the implication of natural justice; the coroner was bound to hear the Annettses before making any finding adverse to them or their son.8 6(1990) 170 CLR 596. 7ibid. at 612, per Brennan J. 8ibid. at 603, per Mason CJ, Deane and McHugh JJ; at 612 per Brennan J; at 621 per Toohey J. Note, however, that Brennan and Toohey JJ dismissed the appeal because they believed that the decision of the Full Court of the Supreme Court of Western Australia (from which the Annettses appealed) was right on the material before it. 255 � A.5 The NATURAL specification C4: Kioa v. West (1985) 159 CLR 550 A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would be applied. A4: the statutory or factual criteria focused on matters which were discrete to the interests of the applicant. A5: the decision-maker was a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In Kioa v. West, 9 a 1985 decision of five judges of the High Court of Australia, Kioa and his wife were Tongan citizens. They were each granted a temporary entry permit. When their entry permits expired, they stayed in Australia and their daughter was born (becoming an Australian citizen). The Minister’s del­ egate made an order for their deportation under the Migration Act 1958 (Cth). They were given no opportunity to answer prejudicial statements against them made to the Minister’s delegate. Mason, Wilson, Brennan and Deane JJ (Gibbs CJ dissenting) held that, in the absence of statutory provisions to the contrary, the requirements of natural justice should have been observed in relation to the making of the deportation order. The deportation order was set aside. C5: Commissioner of Police v. Tanos (1958) 98 CLR 383 A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would be applied. A4: the statutory or factual criteria focused on matters which were discrete to the interests of the applicant. A5: the decision-maker was not a high-level policy-maker. A6: there is a statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In Commissioner of Police v. Tanos, 10 a 1958 decision of three judges of the High Court of Australia, a judge had made a declaration that a restaurant run by Tanos was a “disorderly house” pursuant to s. 3(1)(b) of the Disorderly Houses 9(1985) 159 CLR 550. 10(1958) 98 CLR 383. 256 Appendix A: Case law specifications Act 1943 (NSW). A police inspector had sworn an affidavit as to his suspicion and belief that liquor had been unlawfully sold or supplied on the premises and was likely to unlawfully sold or supplied on the premises again. The judge made the declaration ex parte. The stated grounds for the police inspector’s suspicion concerned things that Tanos’s husband had done when he had been running the restaurant. Tanos claimed that things had changed since she had taken over: “the patronage of a much more desirable class of customer was obtained, a class which would not demand wine with their food.”11 Tanos appealed successfully to the Supreme Court. Dixon CJ, Webb and Taylor JJ held that Tanos’s appeal should have been dis­ missed by the Supreme Court; once the declaration was made, the Act placed the burden on Tanos to prove that liquor had never been sold or supplied on the premises, and she had not proved that. However, the High Court also held that, except in exceptional and special circumstances, an owner/occupier should have an opportunity to be heard before her/his premises are declared “disorderly”. Tanos had been denied that opportunity, so the declaration was set aside. C6: Marine Hull & Liability Insurance Co. Ltd v. Hurford (1986) 67 ALR 77 (“Marine Hull v. Hurford”) A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would be applied. A4: the statutory or factual criteria focused on matters of policy or public in­ terest. A5: the decision-maker was a high-level policy-maker. A6: there is a statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In Marine Hull & Liability Insurance Co. Ltd v. Hurford, 12 a 1986 decision of the Full Court of the Federal Court of Australia, Hurford was Acting Treasurer. He forbade Marine Hull from issuing or renewing insurance policies, using his powers under s. 62 of the Insurance Act 1973 (Cth). Marine Hull complained that they had not been granted a hearing before the Minister made his direction. In the first case, Wilcox J held that the Treasurer should have applied the prin­ ciples of natural justice before making his direction under s. 62. However, because s. 63 of the Act provided for a review of any such directions by the Administrative Appeals Tribunal, “the legislature must be taken to have evinced an intention 11ibid. at 389, per Dixon CJ and Webb J. 12(1986) 67 ALR 77. 257 � A.5 The NATURAL specification that, in the event of the Treasurer failing to so act, the directions are not to be regarded as being invalid in law. They are merely susceptible of challenge before the Tribunal.”13 Marine Hull appealed to the Full Court, which agreed with Wilcox J that natural justice had not been denied in this case; the Acting Treasurer’s direction was not invalidated by his failing to allow Marine Hull a prior hearing. However, the Full Court also held that the “[t]he existence of a right to have a matter reconsidered . . . may well affect the nature of the procedures which ought to be adopted in complying with the rules of natural justice but, ordinarily, it does not exclude them.”14 C7: Macrae v. Attorney-General for New South Wales (1987) 9 NSWLR 268 (“Macrae v. AG for NSW ”) A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would be applied. A4: the statutory or factual criteria focused on matters of policy or public in­ terest. A5: the decision-maker was a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In Macrae v. Attorney-General for New South Wales, 15 a 1987 decision of the New South Wales Court of Appeal, five magistrates who had been appointed under the Justices Act 1902 (NSW) were not appointed under the Local Courts Act 1982 (NSW). The new Act had reorganized the magistracy in NSW, and magistrates appointed under the old Act were entitled to apply for appointment as magistrates under the new Act. The five had applied and were interviewed. Allegations were made privately to the Attorney-General claiming that they were unfit to be appointed, but these allegations were not brought to their notice at the time of the interviews. The Court of Appeal held that the Attorney-General’s decision not to recom­ mend the appointment of the magistrates was void because they were denied their legitimate expectation of procedural fairness. “They have not been treated fairly.”16 13(1985) 62 ALR 253 at 266. 14(1986) 67 ALR 77 at 81, per Davies J. 15(1987) 9 NSWLR 268. 16ibid. at 283, per Kirby P. 258 Appendix A: Case law specifications C8: Attorney-General of Hong Kong v. Ng Yuen Shiu [1983] 2 AC 629 (“AG of Hong Kong v. Shiu”) A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would be applied. A4: it is not known whether the statutory or factual criteria focused on matters which were discrete to the interests of the applicant, or on matters of policy or public interest. A5: the decision-maker was a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In Attorney-General of Hong Kong v. Ng Yuen Shiu, 17 a 1983 decision of the Judicial Committee of the Privy Council, Shiu was an illegal immigrant in Hong Kong. The Hong Kong Government’s immigrant policy allowed illegal entrants to stay if they had reached the urban areas without being arrested. This policy was changed, and it was announced that illegal immigrants would be repatriated. Many illegal immigrants, who (like Shiu) had come from China via Macau, were fearful of being repatriated to China. A senior immigration official announced that each illegal immigrant from Macau would be interviewed, and each case “treated on its merits”. Shiu was questioned by an immigration officer, but was given no opportunity to make representations, and the Director of Immigration ordered his removal. The Privy Council held that, because the government had promised to follow a certain procedure before reaching its decision, it should honour that promise. The removal order was set aside because Shiu had not been asked whether he wished to make representations as to why he should not be removed. C9: Durayappah v. Fernando [1967] 2 AC 337 A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would not be applied. A4: the statutory or factual criteria focused on matters of policy or public in­ terest. A5: the decision-maker was a high-level policy-maker. 17[1983] 2 AC 629. 259 � A.5 The NATURAL specification A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In Durayappah v. Fernando, 18 a 1967 decision of the Judicial Committee of the Privy Council, following complaints as to the conduct of the Jaffna Municipal Council, the Ceylonese Minister of Local Government sent a Commissioner to Jaffna to inquire into the matter. The Commissioner examined the Council’s records, but did not ask any questions of members of the Council, or give them any opportunity to put their views to him. The Commissioner reported to the Minister who, pursuant to s. 277(1) of the Municipal Councils Ordinance 1947 (Ceylon), made an order stating that the Council was not competent to perform its duties and dissolved it. The Mayor of Jaffna (Durayappah) sought writs to quash the Minister’s order and to annul the appointments of the special commissioners who had taken over the running of Jaffna, and a declaration that he was the duly elected mayor. The Privy Council held that the Minister had no right to dissolve the Council without allowing it the right to be heard. However, the Minister’s order was voidable only after a complaint by the Council. Durayappah (as Mayor) had no right to complain independently of the Council: he could only complain if he was representing the Council—and he was not. IImplied (the ideal case in which a duty to observe natural justice is implied): A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would be applied. A4: the statutory or factual criteria focused on matters which were discrete to the interests of the applicant. A5: the decision-maker was not a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. Cases in which a duty to observe natural justice is not implied C10: South Australia v. O’Shea (1987) 163 CLR 378 (“SA v. O’Shea”) A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. 18[1967] 2 AC 337. 260 Appendix A: Case law specifications A3: the power is of a nature that would suggest that procedural fairness would be applied. A4: the statutory or factual criteria focused on matters of policy or public in­ terest. A5: the decision-maker was a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In South Australia v. O’Shea, 19 a 1987 decision of five judges of the High Court of Australia, O’Shea had been convicted of two offences of indecent assault of young children. He was released on licence and remained at liberty after the licence expired. Over a year later, after allegations had been made against him, O’Shea was apprehended and detained. The parole board recommended his release on licence on various conditions, but the Governor in Council resolved to take no action. O’Shea had been given a hearing by the Parole Board, but he claimed he was entitled to a further hearing before the Governor in Council could exercise his discretionary powers under s. 77a(7a) of the Criminal Law Consolidation Act, 1935 (SA). Mason CJ, Wilson, Brennan and Toohey JJ (Deane J dissenting) held that O’Shea was not entitled to a further hearing. “Given the nature of this decision, it cannot be said that Mr O’Shea could have more than a hope that the Governor would be prepared to act on the recommendation of the Board. Hope, of itself, is not sufficient to ground an expectation that will attract legal consequences. So far as the concept of legitimate expectation is concerned, Mr O’Shea must be taken to know that the Act committed to the Governor, with the advice and consent of the Executive Council, the responsibility for determining where the public interest lay . . . The nature of the decision that they were required to make was such that participation by Mr O’Shea was inappropriate.”20 C11: Bread Manufacturers of New South Wales v. Evans (1981) 38 ALR 93 (“Bread Manufacturers v. Evans”) A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is not apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would not be applied. A4: the statutory or factual criteria focused on matters of policy or public in­ terest. A5: the decision-maker was not a high-level policy-maker. 19(1987) 163 CLR 378. 20ibid. at 402, per Wilson and Toohey JJ. 261 � A.5 The NATURAL specification A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In Bread Manufacturers of New South Wales v. Evans, 21 a 1981 decision of five judges of the High Court of Australia, the Bread Manufacturers claimed that an order made by the Prices Commission was void. The order affected the classific­ ation of bread products and had an incidental effect on the price of hamburger buns. The Bread Manufacturers complained that they should have been given the right to put their case to the Commission. The Prices Regulation Act 1948 (NSW) provided that a public inquiry had to be held before an order could be made setting prices, except where the Minister consented to dispensing with the inquiry. The Minister had dispensed with an inquiry before this order was made. Hence, “[t]he argument that the Commission was bound to disclose to the Association the fact that it proposed to make an order which would have the incidental effect of reducing the price of hamburger buns can only succeed if the Commission, although not bound to hold an inquiry, was bound to observe the rules of natural justice”.22 The High Court held that there was no denial of natural justice in relation to the order, because “the reduction of the maximum price in respect of one item was simply a minor incident in a major revision of the price framework covering the whole range of bread products. The effect of that major revision was generally to increase prices. There was, in our opinion, no obligation on the Commission to give advance notice of this development or of the possibility of its occurrence.”23 C12: Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd (1987) 75 ALR 218 (“Minister for Environment v. Peko-Wallsend”) A1: the decision did not affect the property, right, interest, status, or legitimate expectation of the applicant. A2: the decision is not apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would not be applied. A4: the statutory or factual criteria focused on matters of policy or public in­ terest. A5: the decision-maker was a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. 21(1981) 38 ALR 93. 22ibid. at 101, per Gibbs CJ. 23ibid. at 119, per Mason and Wilson JJ, with whom Murphy and Aickin JJ agreed on this point. 262 Appendix A: Case law specifications In Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd, 24 a 1987 decision of the Full Court of the Federal Court of Australia, Peko-Wallsend held various mining interests in Stage 2 of Kakadu National Park. Federal Cabinet decided to nominate Stage 2 for inclusion in the World Heritage List, so it be­ came “identified property” within the meaning of s. 3(2) of the World Heritage Properties Conservation Act 1983 (Cth). This meant that the Governor-General could, by proclamation, make mining operations unlawful in the area. The de­ cision did not affect Peko-Wallsend’s mining rights which were preserved under s. 8b of the National Parks and Wildlife Conservation Act 1975 (Cth). Before Cabinet’s decision, Peko-Wallsend had lobbied Ministers and other offi­ cials extensively, seeking to preserve their mining interests. After the decision they commenced proceedings to prevent the Government from taking any fur­ ther steps to have Stage 2 nominated on the World Heritage List, claiming that Cabinet was bound by the rules of natural justice and had failed to give Peko- Wallsend an opportunity to be heard. Beaumont J (a Federal Court judge) agreed, and held the Cabinet decision void.25 The Full Court of the Federal Court disagreed. Bowen CJ decided that “it would . . . be inappropriate for this court to interfere to set aside a Cabinet decision involving such complex policy considerations”.26 Both Sheppard and Wilcox JJ held that Peko-Wallsend had had adequate opportunity to put their case to relevant Ministers and officials before the Cabinet decision, and were not denied natural justice.27 However, Wilcox J (with whose reasons the other two judges generally agreed) held that the Cabinet’s decision in this case did not attract the obligations of natural justice.28 C13: Nashua Australia Pty Ltd v. Channon (1981) 36 ALR 215 (“Nashua v. Channon”) A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would not be applied. A4: the statutory or factual criteria focused on matters of policy or public in­ terest. A5: the decision-maker was a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. 24(1987) 75 ALR 218. 25Peko-Wallsend Ltd v. Minister for Arts Heritage and Environment (1986) 70 ALR 523. 26(1987) 75 ALR 218 at 225. 27ibid. at 228 per Sheppard J; at 253 per Wilcox J. 28ibid. at 253. 263 � A.5 The NATURAL specification In Nashua Australia Pty Ltd v. Channon, 29 a 1981 decision of the Supreme Court of New South Wales, Nashua applied for a by-law to be made in respect of specially coated paper that they wanted to import from America. They claimed that suitable paper was not available in Australia. Channon, as delegate of the Minister, made a determination under s. 273 of the Customs Act 1966 (Cth) that enabled the paper to be imported at a duty of 2% (instead of 25%) for a period of over two years. Less than four months into that period, the determination was revoked because (according to Channon) suitably equivalent paper was available from an Australian manufacturer. Nashua claimed that they had been denied natural justice as they had been given no notice of the intended revocation and no opportunity to make representations that the revocation should not be made. Lee J held that, given the purposes and operation of the Act, the rules of natural justice did not apply to the revocation of a determination under s. 273. However, he held the revocation invalid because Channon himself had not exercised the discretion entrusted to him. C14: Council of Civil Service Unions v. Minister for the Civil Service [1985] AC 374 (“CCSU v. Minister for the Civil Service”) A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would not be applied. A4: the statutory or factual criteria focused on matters of policy or public in­ terest. A5: the decision-maker was a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were circumstances which made an obligation to observe natural justice inappropriate. In Council of Civil Service Unions v. Minister for the Civil Service, 30 a 1984 de­ cision of the House of Lords, the Minister gave an instruction varying the terms and conditions for staff working at the Government Communications Headquar­ ters. The effect of the variation was to prohibit staff from belonging to national trade unions. Staff had been allowed to belong to these unions for almost forty years. There was a well-established practice of consultation between the gov­ ernment and the trade unions about important alterations to the staff’s terms and conditions, but there was no consultation with staff or the unions before the Minister issued her instruction. The unions claimed that the Minister had been under a duty to act fairly by consulting those concerned before issuing the instruction. The Minister produced an affidavit from the Secretary to the Cabinet stating that recent industrial 29(1981) 36 ALR 215. 30[1985] AC 374. 264 Appendix A: Case law specifications action had led the Cabinet to believe that prior consultation about the Minister’s instruction could have precipitated further disruption and would have indicated vulnerable areas of the Government Communications Headquarters’ operations. The House of Lords held that the unions had a legitimate expectation that the Minister would consult with them on this matter. However, their Lordships held that the requirements of national security outweighed those of fairness: the Minister “had shown that her decision was one which not only could reasonably have been based, but was in fact based, on considerations of national security, which outweighed what would otherwise have been the reasonable expectation on the part of the [unions] for prior consultation.”31 The Minister’s instruction was valid. C15: McInnes v. Onslow Fane [1978] 3 All ER 211 A1: the decision affected the property, right, interest, status, or legitimate ex­ pectation of the applicant. A2: the decision is apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would be applied. A4: the statutory or factual criteria focused on matters which were discrete to the interests of the applicant. A5: the decision-maker was not a high-level policy-maker. A6: there is no statutory right to appeal against the decision. A7: there were no circumstances which would have made an obligation to observe natural justice inappropriate. In McInnes v. Onslow Fane, 32 a 1978 decision of the Chancery Division of the English High Court, McInnes had held, at various times, licences to promote, train and act as master of ceremonies in professional boxing. All his licences were revoked by the British Boxing Board of Control. He made five unsuccessful applications for a manager’s licence. With his sixth application he requested an oral hearing and prior notification of anything that might prevent the area council (to which he applied) making a favourable recommendation to the board. The board refused his applications without giving him an oral hearing or informing him of the case against him. Megarry V-C held that the board was under no duty to provide reasons to McInnes or to allow him a hearing: “This is not a case in which there has been any suggestion of the board considering any alleged dishonesty or morally culpable conduct of the plaintiff. A man free from any moral blemish may nevertheless be wholly unsuitable for a particular type of work . . . In such circumstances, in the absence of anything to suggest that the board have been affected by dishonesty or bias or caprice, or that there is any other impropriety, I think that the board are fully entitled to give no reasons for their decision, and to decide the application 31ibid. at 403, per Lord Fraser of Tullybelton. 32[1978] 3 All ER 211. • • • • 265 � A.5 The NATURAL specification without any preliminary indication to the plaintiff of those reasons. The board are the best judges of the desirability of granting a licence, and in the absence of any impropriety the court ought not to interfere.”33 INot-Implied (the ideal case in which a duty to observe natural justice is not implied): A1: the decision did not affect the property, right, interest, status, or legitimate expectation of the applicant. A2: the decision is not apt to have a discrete impact on the interests of the applicant. A3: the power is of a nature that would suggest that procedural fairness would not be applied. A4: the statutory or factual criteria focused on matters of policy or public in­ terest. A5: the decision-maker was a high-level policy-maker. A6: there is a statutory right to appeal against the decision. A7: there were circumstances which made an obligation to observe natural justice inappropriate. Affected area Case C1 C2 C3 C4 C5 C6 C7 C8 IAffected C9 IUnaffected Results Attributes c Result A1 A2 A3 A4 • × • • 1 × • × • 2 × × • × 2 × • × × 2 Affected× • × • 2 • × × × 2 • × • × 3 • × × • 8 Unaffected× × × × 4 × × × × Affected: the decision affected the property, right, interest, status, or legitimate expect­ ation of the applicant. Unaffected: the decision did not affect the property, right, interest, status, or legitimate expectation of the applicant. 33ibid. at 223. 266 Appendix A: Case law specifications Attributes A1: Did the decision affect a financial, property or occupational interest of the applic­ ant? yes: the decision affected a financial, property or occupational interest of the applicant. ⊃ Affected no: the decision did not affect a financial, property or occupational interest of the applicant. unknown: it is not known whether or not the decision affected a financial, property or occupational interest of the applicant. The interest must be an existing interest: one which existed when the decision was made. A2: Did the decision affect the applicant’s personal liberty? yes: the decision affected the applicant’s personal liberty. ⊃ Affected no: the decision did not affect the applicant’s personal liberty. unknown: it is not known whether the decision affected the applicant’s personal liberty. A3: Did the decision affect the applicant’s reputation? yes: the decision affected the applicant’s reputation. ⊃ Affected no: the decision did not affect the applicant’s reputation. unknown: it is not known whether the decision affected the applicant’s repu­ tation. A4: ∀ Expectation area yes: the applicant had a legitimate expectation which was affected by the de­ cision. � Expectation ⊃ Affected no: the applicant did not have a legitimate expectation which was affected by the decision. � No-Expectation unknown: it is not known whether the applicant had a legitimate expectation which was affected by the decision. 267 � A.5 The NATURAL specification Cases in which the decision affected the property, right, interest, status, or legitimate expectation of the applicant C1: FAI Insurances Ltd v. Winneke (1982) 151 CLR 342 (“FAI v. Winneke”) A1: the decision affected a financial, property or occupational interest of the applicant. A2: the decision did not affect the applicant’s personal liberty. A3: the decision affected the applicant’s reputation. A4: the applicant had a legitimate expectation which was affected by the de­ cision. In FAI Insurances Ltd v. Winneke, 34 a 1982 decision of seven judges of the High Court of Australia, Winneke was the Governor of Victoria. The Workers Com­ pensation Act 1958 (Vic) made accident insurance compulsory for all employers, and required them to obtain that insurance from the Insurance Commissioner or from an insurer approved by the Governor in Council. Regulations made un­ der the Act allowed the Governor in Council to grant an approval for a period not exceeding twelve months, and (on application) to renew an approval for any period not exceeding twelve months “if the Governor in Council thinks fit.” FAI had had its approval to provide worker’s compensation renewed every year for twenty years, until 1981. Gibbs CJ, Stephen, Mason, Aickin, Wilson and Brennan JJ (Murphy J dissent­ ing) held that in deciding whether to renew an approval previously given, the Governor in Council is subject to the requirements of natural justice. “In these circumstances, a company which becomes an approved insurer has a legitimate expectation that its approval will be renewed unless some good reason exists for refusing to renew it. It would not be fair to deprive a company of the ability to carry on its business without revealing the reason for doing so”.35 C2: Haoucher v. Minister of State for Immigration and Ethnic Affairs (1990) 169 CLR 648 (“Haoucher v. Minister for Immigration”) A1: the decision did not affect a financial, property or occupational interest of the applicant. A2: the decision affected the applicant’s personal liberty. A3: the decision did not affect the applicant’s reputation. A4: the applicant had a legitimate expectation which was affected by the de­ cision. 34(1982) 151 CLR 342. 35ibid. at 348, per Gibbs CJ. 268 Appendix A: Case law specifications In Haoucher v. Minister of State for Immigration and Ethnic Affairs, 36 a 1990 de­ cision of five judges of the High Court of Australia, Haoucher had been convicted of an offence for which he had been sentenced to imprisonment for a period not less than one year. Hence—as Haoucher was not an Australian citizen—the Min­ ister had the power to order his deportation under s. 12 of the Migration Act 1958 (Cth), and did so. The Minister had previously told Parliament that the Govern­ ment’s policy was that a deportee had the right to appeal to the Administrative Appeals Tribunal, and that “recommendations of the Administrative Appeals Tribunal should be overturned by the Minister only in exceptional circumstances and only when strong evidence can be produced to justify his decision.” Haoucher appealed to the AAT which recommended that the deportation order be revoked. The Minister decided not to accept that recommendation. Deane, Toohey and McHugh (Dawson and Gaudron dissenting) held that Haoucher was entitled to know the matters which constituted “exceptional circumstances” and “strong evidence” sufficient for the Minister to depart from the general policy of following the AAT’s recommendations. The fact that the Minister had not given Haoucher such details, and had given him no opportunity to make representa­ tions, was a denial of natural justice. C3: Annetts v. McCann (1990) 170 CLR 596 A1: the decision did not affect a financial, property or occupational interest of the applicant. A2: the decision did not affect the applicant’s personal liberty. A3: the decision affected the applicant’s reputation. A4: the applicant did not have a legitimate expectation which was affected by the decision. In Annetts v. McCann, 37 a 1990 decision of five judges of the High Court of Australia, a coroner had been conducting an inquest into the death of a 16-year old boy. The boy’s parents (Mr and Mrs Annetts) sought to make a submission before the coroner made a finding. The coroner decided that the Coroner’s Act 1920 (WA) gave him the discretion (which he chose to exercise) to disallow their submission. The Annettses appealed. The High Court (Mason CJ, Brennan, Deane, Toohey and McHugh JJ) held that their son’s reputation gave the Annettses an interest in the Coroner’s inquiry. “A finding in an inquest into a death is naturally likely to deal with the conduct of the deceased leading to death. An unfavourable reflection on the deceased is usually a matter of concern to her or his parents, spouse or children and, if they choose to appear at the inquest in order to safeguard the reputation of the deceased, the 36(1990) 169 CLR 648. 37(1990) 170 CLR 596. 269 � A.5 The NATURAL specification familial relationship suffices, in my view, to establish the deceased’s reputation as a relevant interest which should not be adversely affected without according natural justice to those who are seeking to safeguard that reputation.”38 The Court held that the fact that the coroner’s decision was merely recommend­ atory (whether or not to prosecute) was not sufficient to avoid the implication of natural justice; the coroner was bound to hear the Annettses before making any finding adverse to them or their son.39 C4: South Australia v. O’Shea (1987) 163 CLR 378 (“SA v. O’Shea”) A1: the decision did not affect a financial, property or occupational interest of the applicant. A2: the decision affected the applicant’s personal liberty. A3: the decision did not affect the applicant’s reputation. A4: the applicant did not have a legitimate expectation which was affected by the decision. In South Australia v. O’Shea, 40 a 1987 decision of five judges of the High Court of Australia, O’Shea had been convicted of two offences of indecent assault of young children. He was released on licence and remained at liberty after the licence expired. Over a year later, after allegations had been made against him, O’Shea was apprehended and detained. The parole board recommended his release on licence on various conditions, but the Governor in Council resolved to take no action. O’Shea had been given a hearing by the Parole Board, but he claimed he was entitled to a further hearing before the Governor in Council could exercise his discretionary powers under s. 77a(7a) of the Criminal Law Consolidation Act, 1935 (SA). Mason CJ, Wilson, Brennan and Toohey JJ (Deane J dissenting) held that O’Shea was not entitled to a further hearing. “Given the nature of this decision, it cannot be said that Mr O’Shea could have more than a hope that the Governor would be prepared to act on the recommendation of the Board. Hope, of itself, is not sufficient to ground an expectation that will attract legal consequences. So far as the concept of legitimate expectation is concerned, Mr O’Shea must be taken to know that the Act committed to the Governor, with the advice and consent of the Executive Council, the responsibility for determining where the public interest lay . . . The nature of the decision that they were required to make was such that participation by Mr O’Shea was inappropriate.”41 38ibid. at 612, per Brennan J. 39ibid. at 603, per Mason CJ, Deane and McHugh JJ; at 612 per Brennan J; at 621 per Toohey J. Note, however, that Brennan and Toohey JJ dismissed the appeal because they believed that the decision of the Full Court of the Supreme Court of Western Australia (from which the Annettses appealed) was right on the material before it. 40(1987) 163 CLR 378. 41ibid. at 402, per Wilson and Toohey JJ. 270 Appendix A: Case law specifications C5: Kioa v. West (1985) 159 CLR 550 A1: the decision did not affect a financial, property or occupational interest of the applicant. A2: the decision affected the applicant’s personal liberty. A3: the decision did not affect the applicant’s reputation. A4: the applicant had a legitimate expectation which was affected by the de­ cision. In Kioa v. West, 42 a 1985 decision of five judges of the High Court of Australia, Kioa and his wife were Tongan citizens. They were each granted a temporary entry permit. When their entry permits expired, they stayed in Australia and their daughter was born (becoming an Australian citizen). The Minister’s del­ egate made an order for their deportation under the Migration Act 1958 (Cth). They were given no opportunity to answer prejudicial statements against them made to the Minister’s delegate. Mason, Wilson, Brennan and Deane JJ (Gibbs CJ dissenting) held that, in the absence of statutory provisions to the contrary, the requirements of natural justice should have been observed in relation to the making of the deportation order. The deportation order was set aside. C6: Bread Manufacturers of New South Wales v. Evans (1981) 38 ALR 93 (“Bread Manufacturers v. Evans”) A1: the decision affected a financial, property or occupational interest of the applicant. A2: the decision did not affect the applicant’s personal liberty. A3: the decision did not affect the applicant’s reputation. A4: the applicant did not have a legitimate expectation which was affected by the decision. In Bread Manufacturers of New South Wales v. Evans, 43 a 1981 decision of five judges of the High Court of Australia, the Bread Manufacturers claimed that an order made by the Prices Commission was void. The order affected the classific­ ation of bread products and had an incidental effect on the price of hamburger buns. The Bread Manufacturers complained that they should have been given the right to put their case to the Commission. The Prices Regulation Act 1948 (NSW) provided that a public inquiry had to be held before an order could be made setting prices, except where the Minister consented to dispensing with the inquiry. The Minister had dispensed with an inquiry before this order was made. Hence, “[t]he argument that the Commission was bound to disclose to the Association the fact that it proposed to make an 42(1985) 159 CLR 550. 43(1981) 38 ALR 93. 271 � A.5 The NATURAL specification order which would have the incidental effect of reducing the price of hamburger buns can only succeed if the Commission, although not bound to hold an inquiry, was bound to observe the rules of natural justice”.44 The High Court held that there was no denial of natural justice in relation to the order, because “the reduction of the maximum price in respect of one item was simply a minor incident in a major revision of the price framework covering the whole range of bread products. The effect of that major revision was generally to increase prices. There was, in our opinion, no obligation on the Commission to give advance notice of this development or of the possibility of its occurrence.”45 C7: Commissioner of Police v. Tanos (1958) 98 CLR 383 A1: the decision affected a financial, property or occupational interest of the applicant. A2: the decision did not affect the applicant’s personal liberty. A3: the decision affected the applicant’s reputation. A4: the applicant did not have a legitimate expectation which was affected by the decision. In Commissioner of Police v. Tanos, 46 a 1958 decision of three judges of the High Court of Australia, a judge had made a declaration that a restaurant run by Tanos was a “disorderly house” pursuant to s. 3(1)(b) of the Disorderly Houses Act 1943 (NSW). A police inspector had sworn an affidavit as to his suspicion and belief that liquor had been unlawfully sold or supplied on the premises and was likely to unlawfully sold or supplied on the premises again. The judge made the declaration ex parte. The stated grounds for the police inspector’s suspicion concerned things that Tanos’s husband had done when he had been running the restaurant. Tanos claimed that things had changed since she had taken over: “the patronage of a much more desirable class of customer was obtained, a class which would not demand wine with their food.”47 Tanos appealed successfully to the Supreme Court. Dixon CJ, Webb and Taylor JJ held that Tanos’s appeal should have been dis­ missed by the Supreme Court; once the declaration was made, the Act placed the burden on Tanos to prove that liquor had never been sold or supplied on the premises, and she had not proved that. However, the High Court also held that, except in exceptional and special circumstances, an owner/occupier should have an opportunity to be heard before her/his premises are declared “disorderly”. Tanos had been denied that opportunity, so the declaration was set aside. 44ibid. at 101, per Gibbs CJ. 45ibid. at 119, per Mason and Wilson JJ, with whom Murphy and Aickin JJ agreed on this point. 46(1958) 98 CLR 383. 47ibid. at 389, per Dixon CJ and Webb J. 272 Appendix A: Case law specifications C8: Council of Civil Service Unions v. Minister for the Civil Service [1985] AC 374 (“CCSU v. Minister for the Civil Service”) A1: the decision affected a financial, property or occupational interest of the applicant. A2: the decision did not affect the applicant’s personal liberty. A3: the decision did not affect the applicant’s reputation. A4: the applicant had a legitimate expectation which was affected by the de­ cision. In Council of Civil Service Unions v. Minister for the Civil Service, 48 a 1984 de­ cision of the House of Lords, the Minister gave an instruction varying the terms and conditions for staff working at the Government Communications Headquar­ ters. The effect of the variation was to prohibit staff from belonging to national trade unions. Staff had been allowed to belong to these unions for almost forty years. There was a well-established practice of consultation between the gov­ ernment and the trade unions about important alterations to the staff’s terms and conditions, but there was no consultation with staff or the unions before the Minister issued her instruction. The unions claimed that the Minister had been under a duty to act fairly by consulting those concerned before issuing the instruction. The Minister produced an affidavit from the Secretary to the Cabinet stating that recent industrial action had led the Cabinet to believe that prior consultation about the Minister’s instruction could have precipitated further disruption and would have indicated vulnerable areas of the Government Communications Headquarters’ operations. The House of Lords held that the unions had a legitimate expectation that the Minister would consult with them on this matter. However, their Lordships held that the requirements of national security outweighed those of fairness: the Minister “had shown that her decision was one which not only could reasonably have been based, but was in fact based, on considerations of national security, which outweighed what would otherwise have been the reasonable expectation on the part of the [unions] for prior consultation.”49 The Minister’s instruction was valid. IAffected (the ideal case in which the decision affected the property, right, interest, status, or legitimate expectation of the applicant): A1: the decision affected a financial, property or occupational interest of the applicant. A2: the decision affected the applicant’s personal liberty. A3: the decision affected the applicant’s reputation. A4: the applicant had a legitimate expectation which was affected by the de­ cision. 48[1985] AC 374. 49ibid. at 403, per Lord Fraser of Tullybelton. 273 � A.5 The NATURAL specification Cases in which the decision did not affect the property, right, interest, status, or legitimate expectation of the applicant C9: Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd (1987) 75 ALR 218 (“Minister for Environment v. Peko-Wallsend”) A1: the decision did not affect a financial, property or occupational interest of the applicant. A2: the decision did not affect the applicant’s personal liberty. A3: the decision did not affect the applicant’s reputation. A4: the applicant did not have a legitimate expectation which was affected by the decision. In Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd, 50 a 1987 decision of the Full Court of the Federal Court of Australia, Peko-Wallsend held various mining interests in Stage 2 of Kakadu National Park. Federal Cabinet decided to nominate Stage 2 for inclusion in the World Heritage List, so it be­ came “identified property” within the meaning of s. 3(2) of the World Heritage Properties Conservation Act 1983 (Cth). This meant that the Governor-General could, by proclamation, make mining operations unlawful in the area. The de­ cision did not affect Peko-Wallsend’s mining rights which were preserved under s. 8b of the National Parks and Wildlife Conservation Act 1975 (Cth). Before Cabinet’s decision, Peko-Wallsend had lobbied Ministers and other offi­ cials extensively, seeking to preserve their mining interests. After the decision they commenced proceedings to prevent the Government from taking any fur­ ther steps to have Stage 2 nominated on the World Heritage List, claiming that Cabinet was bound by the rules of natural justice and had failed to give Peko- Wallsend an opportunity to be heard. Beaumont J (a Federal Court judge) agreed, and held the Cabinet decision void.51 The Full Court of the Federal Court disagreed. Bowen CJ decided that “it would . . . be inappropriate for this court to interfere to set aside a Cabinet decision involving such complex policy considerations”.52 Both Sheppard and Wilcox JJ held that Peko-Wallsend had had adequate opportunity to put their case to relevant Ministers and officials before the Cabinet decision, and were not denied natural justice.53 However, Wilcox J (with whose reasons the other two judges generally agreed) held that the Cabinet’s decision in this case did not attract the obligations of natural justice.54 50(1987) 75 ALR 218. 51Peko-Wallsend Ltd v. Minister for Arts Heritage and Environment (1986) 70 ALR 523. 52(1987) 75 ALR 218 at 225. 53ibid. at 228 per Sheppard J; at 253 per Wilcox J. 54ibid. at 253. 274 Appendix A: Case law specifications IUnaffected (the ideal case in which the decision did not affect the property, right, in­ terest, status, or legitimate expectation of the applicant): A1: the decision did not affect a financial, property or occupational interest of the applicant. A2: the decision did not affect the applicant’s personal liberty. A3: the decision did not affect the applicant’s reputation. A4: the applicant did not have a legitimate expectation which was affected by the decision. Expectation area Case A1 Attributes A2 A3 A4 A5 A6 c Result C1 C2 C3 C4 C5 C6 IExpectation × • • × • × 1 × • × × • × 2 × × × × • • 2 × • × × • × 2 • × × × × × 4 • × × × • × 7 • • • • • • Expectation C7 C8 INo-Expectation × × × × • × 2 × × × × × × 4 × × × × × × No-Expectation Opening If the applicant had a legitimate expectation—or a reasonable expectation—which was affected by the decision, natural justice may be implied. “ ‘[L]egitimate expectations’ . . . are capable of including expectations which go beyond enforceable legal rights, provided they have some reasonable basis”. 55 Results Expectation: the applicant had a legitimate expectation which was affected by the de­ cision. No-Expectation: the applicant did not have a legitimate expectation which was affected by the decision. 55Cole v. Cunningham (1983) 49 ALR 123 at 131, per Bowen CJ, Sheppard and Morling JJ. 275 � A.5 The NATURAL specification Attributes A1: Did the decision-maker break a promise or undertaking? yes: the decision-maker broke a promise or undertaking. ⊃ Expectation no: the decision-maker did not break a promise or undertaking. unknown: it is not known whether the decision-maker broke a promise or un­ dertaking. For example, if the decision went against a published policy then it was in breach of a promise or undertaking. A2: Did the decision-maker go against an established course of practice? yes: the decision-maker went against an established course of practice. ⊃ Expectation no: the decision-maker did not go against an established course of practice. unknown: it is not known whether the decision-maker went against an estab­ lished course of practice. There may have been an expectation that an established course of practice would be adopted, or that notice (and an opportunity to comment) would be given before the practice was abandoned. A3: Did the decision involve a refusal to renew an existing interest? yes: the decision involved the refusal to renew an existing interest. ⊃ Expectation no: the decision did not involve a refusal to renew an existing interest. unknown: it is not known whether the decision involved a refusal to renew an existing interest. For example, if the decision was to refuse to renew a licence, then it was a refusal to renew an existing interest. A4: Did the decision-maker or a statutory provision suggest that an initial interest would be granted? yes: the decision-maker or a statutory provision suggested that an initial interest would be granted. ⊃ Expectation no: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. unknown: it is not known whether the decision-maker or a statutory provision suggested that an initial interest would be granted. 276 Appendix A: Case law specifications A5: Did the decision affect an established liberty or interest? yes: the decision affected an established liberty or interest. ⊃ Expectation no: the decision did not affect an established liberty or interest. unknown: it is not known whether the decision affected an established liberty or interest. For example, a decision to warn the applicant off a racetrack would affect the applicant’s established liberty to go to the races. And a deportation order would affect an applicant’s established liberty to remain in the country. A6: Was there a standard administrative procedure which the decision-maker did not follow? yes: there was a standard administrative procedure which the decision-maker did not follow. ⊃ Expectation no: there was no standard administrative procedure which the decision-maker should have followed. unknown: it is not known whether there was a standard administrative pro­ cedure which the decision-maker should have followed. Cases in which the applicant had a legitimate expectation which was affected by the decision C1: FAI Insurances Ltd v. Winneke (1982) 151 CLR 342 (“FAI v. Winneke”) A1: the decision-maker did not break a promise or undertaking. A2: the decision-maker went against an established course of practice. A3: the decision involved the refusal to renew an existing interest. A4: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. A5: the decision affected an established liberty or interest. A6: there was no standard administrative procedure which the decision-maker should have followed. In FAI Insurances Ltd v. Winneke, 56 a 1982 decision of seven judges of the High Court of Australia, Winneke was the Governor of Victoria. The Workers Com­ pensation Act 1958 (Vic) made accident insurance compulsory for all employers, and required them to obtain that insurance from the Insurance Commissioner or from an insurer approved by the Governor in Council. Regulations made un­ der the Act allowed the Governor in Council to grant an approval for a period 56(1982) 151 CLR 342. 277 � A.5 The NATURAL specification not exceeding twelve months, and (on application) to renew an approval for any period not exceeding twelve months “if the Governor in Council thinks fit.” FAI had had its approval to provide worker’s compensation renewed every year for twenty years, until 1981. Gibbs CJ, Stephen, Mason, Aickin, Wilson and Brennan JJ (Murphy J dissent­ ing) held that in deciding whether to renew an approval previously given, the Governor in Council is subject to the requirements of natural justice. “In these circumstances, a company which becomes an approved insurer has a legitimate expectation that its approval will be renewed unless some good reason exists for refusing to renew it. It would not be fair to deprive a company of the ability to carry on its business without revealing the reason for doing so”.57 C2: Haoucher v. Minister of State for Immigration and Ethnic Affairs (1990) 169 CLR 648 (“Haoucher v. Minister for Immigration”) A1: the decision-maker did not break a promise or undertaking. A2: the decision-maker went against an established course of practice. A3: the decision did not involve a refusal to renew an existing interest. A4: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. A5: the decision affected an established liberty or interest. A6: there was no standard administrative procedure which the decision-maker should have followed. In Haoucher v. Minister of State for Immigration and Ethnic Affairs, 58 a 1990 de­ cision of five judges of the High Court of Australia, Haoucher had been convicted of an offence for which he had been sentenced to imprisonment for a period not less than one year. Hence—as Haoucher was not an Australian citizen—the Min­ ister had the power to order his deportation under s. 12 of the Migration Act 1958 (Cth), and did so. The Minister had previously told Parliament that the Govern­ ment’s policy was that a deportee had the right to appeal to the Administrative Appeals Tribunal, and that “recommendations of the Administrative Appeals Tribunal should be overturned by the Minister only in exceptional circumstances and only when strong evidence can be produced to justify his decision.” Haoucher appealed to the AAT which recommended that the deportation order be revoked. The Minister decided not to accept that recommendation. Deane, Toohey and McHugh (Dawson and Gaudron dissenting) held that Haoucher was entitled to know the matters which constituted “exceptional circumstances” and “strong evidence” sufficient for the Minister to depart from the general policy of following the AAT’s recommendations. The fact that the Minister had not given Haoucher such details, and had given him no opportunity to make representa­ tions, was a denial of natural justice. 57ibid. at 348, per Gibbs CJ. 58(1990) 169 CLR 648. 278 Appendix A: Case law specifications C3: Kioa v. West (1985) 159 CLR 550 A1: the decision-maker did not break a promise or undertaking. A2: the decision-maker did not go against an established course of practice. A3: the decision did not involve a refusal to renew an existing interest. A4: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. A5: the decision affected an established liberty or interest. A6: there was a standard administrative procedure which the decision-maker did not follow. In Kioa v. West, 59 a 1985 decision of five judges of the High Court of Australia, Kioa and his wife were Tongan citizens. They were each granted a temporary entry permit. When their entry permits expired, they stayed in Australia and their daughter was born (becoming an Australian citizen). The Minister’s del­ egate made an order for their deportation under the Migration Act 1958 (Cth). They were given no opportunity to answer prejudicial statements against them made to the Minister’s delegate. Mason, Wilson, Brennan and Deane JJ (Gibbs CJ dissenting) held that, in the absence of statutory provisions to the contrary, the requirements of natural justice should have been observed in relation to the making of the deportation order. The deportation order was set aside. C4: Heatley v. Tasmanian Racing and Gaming Commission (1977) 137 CLR 487 (“Heat­ ley v. TRC ”) A1: the decision-maker did not break a promise or undertaking. A2: the decision-maker went against an established course of practice. A3: the decision did not involve a refusal to renew an existing interest. A4: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. A5: the decision affected an established liberty or interest. A6: there was no standard administrative procedure which the decision-maker should have followed. In Heatley v. Tasmanian Racing and Gaming Commission, 60 a 1977 decision of five judges of the High Court of Australia, Heatley was warned off racecourses in Tasmania by the Commission, using its powers under s. 39(3) of the Racing and Gaming Act 1952 (Tas). Heatley had been given no notice that the Com­ mission intended warning him off (and hence was given no opportunity to make representations to the Commission) and was given no reasons. Stephen, Mason, Murphy and Aickin JJ (Barwick CJ dissenting) held that the Commission was bound by the rules of natural justice. In the absence of an emer­ gency, the Commission should have given Heatley notice of its intention to warn 59(1985) 159 CLR 550. 60(1977) 137 CLR 487. 279 � A.5 The NATURAL specification him off, and the grounds for taking such action. Further, Heatley should have been given an opportunity to make representations to the Commission before it made its decision. Aickin J (with whom Stephen and Mason JJ agreed) held that all members of the public—including Heatley—have a legitimate expectation that they will be allowed onto racecourses.61 C5: Cole v. Cunningham (1983) 49 ALR 123 A1: the decision-maker broke a promise or undertaking. A2: the decision-maker did not go against an established course of practice. A3: the decision did not involve a refusal to renew an existing interest. A4: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. A5: the decision did not affect an established liberty or interest. A6: there was no standard administrative procedure which the decision-maker should have followed. In Cole v. Cunningham, 62 a 1983 decision of the Full Court of the Federal Court of Australia, Cunningham had been encouraged to resign from the Public Service because his superiors believed he had been guilty of misconduct in the perform­ ance of his duties. He had formed an attachment and begun to live with a Fijian women whose permit extension application he had processed. He was threatened with criminal prosecution and told that “[i]f you resign now it will be a normal resignation and you’ll leave with a clean record.”63 About eighteen months later, Cunningham sought reappointment to the Pub­ lic Service and was told that he would be offered a position subject to police and ASIO clearances. The next day he was told that he have been given an unsatisfactory report based on the earlier events. Bowen CJ, Sheppard and Morling JJ held that, in general, applicants for appoint­ ment or reappointment to the public service are not entitled to natural justice because they have no legitimate expectation which can be affected by a refusal to appoint. However, Cunningham did have a legitimate expectation that any decision to reappoint him would not be made on the basis of his past record. C6: Attorney-General of Hong Kong v. Ng Yuen Shiu [1983] 2 AC 629 (“AG of Hong Kong v. Shiu”) A1: the decision-maker broke a promise or undertaking. A2: the decision-maker did not go against an established course of practice. A3: the decision did not involve a refusal to renew an existing interest. A4: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. 61ibid. at 509. 62(1983) 49 ALR 123. 63ibid. at 125. 280 Appendix A: Case law specifications A5: the decision affected an established liberty or interest. A6: there was no standard administrative procedure which the decision-maker should have followed. In Attorney-General of Hong Kong v. Ng Yuen Shiu, 64 a 1983 decision of the Judicial Committee of the Privy Council, Shiu was an illegal immigrant in Hong Kong. The Hong Kong Government’s immigrant policy allowed illegal entrants to stay if they had reached the urban areas without being arrested. This policy was changed, and it was announced that illegal immigrants would be repatriated. Many illegal immigrants, who (like Shiu) had come from China via Macau, were fearful of being repatriated to China. A senior immigration official announced that each illegal immigrant from Macau would be interviewed, and each case “treated on its merits”. Shiu was questioned by an immigration officer, but was given no opportunity to make representations, and the Director of Immigration ordered his removal. The Privy Council held that, because the government had promised to follow a certain procedure before reaching its decision, it should honour that promise. The removal order was set aside because Shiu had not been asked whether he wished to make representations as to why he should not be removed. IExpectation (the ideal case in which the applicant had a legitimate expectation which was affected by the decision): A1: the decision-maker broke a promise or undertaking. A2: the decision-maker went against an established course of practice. A3: the decision involved the refusal to renew an existing interest. A4: the decision-maker or a statutory provision suggested that an initial interest would be granted. A5: the decision affected an established liberty or interest. A6: there was a standard administrative procedure which the decision-maker did not follow. Cases in which the applicant did not have a legitimate expectation which was affected by the decision C7: South Australia v. O’Shea (1987) 163 CLR 378 (“SA v. O’Shea”) A1: the decision-maker did not break a promise or undertaking. A2: the decision-maker did not go against an established course of practice. A3: the decision did not involve a refusal to renew an existing interest. A4: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. 64[1983] 2 AC 629. 281 � A.5 The NATURAL specification A5: the decision affected an established liberty or interest. A6: there was no standard administrative procedure which the decision-maker should have followed. In South Australia v. O’Shea, 65 a 1987 decision of five judges of the High Court of Australia, O’Shea had been convicted of two offences of indecent assault of young children. He was released on licence and remained at liberty after the licence expired. Over a year later, after allegations had been made against him, O’Shea was apprehended and detained. The parole board recommended his release on licence on various conditions, but the Governor in Council resolved to take no action. O’Shea had been given a hearing by the Parole Board, but he claimed he was entitled to a further hearing before the Governor in Council could exercise his discretionary powers under s. 77a(7a) of the Criminal Law Consolidation Act, 1935 (SA). Mason CJ, Wilson, Brennan and Toohey JJ (Deane J dissenting) held that O’Shea was not entitled to a further hearing. “Given the nature of this decision, it cannot be said that Mr O’Shea could have more than a hope that the Governor would be prepared to act on the recommendation of the Board. Hope, of itself, is not sufficient to ground an expectation that will attract legal consequences. So far as the concept of legitimate expectation is concerned, Mr O’Shea must be taken to know that the Act committed to the Governor, with the advice and consent of the Executive Council, the responsibility for determining where the public interest lay . . . The nature of the decision that they were required to make was such that participation by Mr O’Shea was inappropriate.”66 C8: Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd (1987) 75 ALR 218 (“Minister for Environment v. Peko-Wallsend”) A1: the decision-maker did not break a promise or undertaking. A2: the decision-maker did not go against an established course of practice. A3: the decision did not involve a refusal to renew an existing interest. A4: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. A5: the decision did not affect an established liberty or interest. A6: there was no standard administrative procedure which the decision-maker should have followed. In Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd, 67 a 1987 decision of the Full Court of the Federal Court of Australia, Peko-Wallsend held various mining interests in Stage 2 of Kakadu National Park. Federal Cabinet decided to nominate Stage 2 for inclusion in the World Heritage List, so it be­ came “identified property” within the meaning of s. 3(2) of the World Heritage Properties Conservation Act 1983 (Cth). This meant that the Governor-General 65(1987) 163 CLR 378. 66ibid. at 402, per Wilson and Toohey JJ. 67(1987) 75 ALR 218. 282 Appendix A: Case law specifications could, by proclamation, make mining operations unlawful in the area. The de­ cision did not affect Peko-Wallsend’s mining rights which were preserved under s. 8b of the National Parks and Wildlife Conservation Act 1975 (Cth). Before Cabinet’s decision, Peko-Wallsend had lobbied Ministers and other offi­ cials extensively, seeking to preserve their mining interests. After the decision they commenced proceedings to prevent the Government from taking any fur­ ther steps to have Stage 2 nominated on the World Heritage List, claiming that Cabinet was bound by the rules of natural justice and had failed to give Peko- Wallsend an opportunity to be heard. Beaumont J (a Federal Court judge) agreed, and held the Cabinet decision void.68 The Full Court of the Federal Court disagreed. Bowen CJ decided that “it would . . . be inappropriate for this court to interfere to set aside a Cabinet decision involving such complex policy considerations”.69 Both Sheppard and Wilcox JJ held that Peko-Wallsend had had adequate opportunity to put their case to relevant Ministers and officials before the Cabinet decision, and were not denied natural justice.70 However, Wilcox J (with whose reasons the other two judges generally agreed) held that the Cabinet’s decision in this case did not attract the obligations of natural justice.71 INo-Expectation (the ideal case in which the applicant did not have a legitimate expecta­ tion which was affected by the decision): A1: the decision-maker did not break a promise or undertaking. A2: the decision-maker did not go against an established course of practice. A3: the decision did not involve a refusal to renew an existing interest. A4: neither the decision-maker nor a statutory provision suggested that an initial interest would be granted. A5: the decision did not affect an established liberty or interest. A6: there was no standard administrative procedure which the decision-maker should have followed. 68Peko-Wallsend Ltd v. Minister for Arts Heritage and Environment (1986) 70 ALR 523. 69(1987) 75 ALR 218 at 225. 70ibid. at 228 per Sheppard J; at 253 per Wilcox J. 71ibid. at 253. B Example reports The law is not a series of calculating machines where definitions and answers come tumbling out when the right levers are pushed. William O. Douglas (1948)17 An argument which does not convince yourself, may convince the Judge to whom you urge it; and if it does convince him, why, then, Sir, you are wrong, and he is right. Samuel Johnson (1768)18 283 284 Appendix B: Example reports B.1 Introduction This appendix presents six example reports from four test cases, demonstrating the use of each of the four specifications described in chapter 5 and given in appendix A. For each example, SHYSTER was requested to hypothesize with a limit of two changed attribute values, and to report on up to one hypothetical per result. Parker v. British Airways 19 is used as a test case for the Finder specification. SHYSTER’s opinion on that case is evaluated in §5.2.3, and its report file is in §B.2. Australasian Performing Right Association Ltd v. Jain20 is used in §5.3.3 as a test case for the Authorization specification, and SHYSTER’s advice is evaluated there. SHYSTER’s report file for APRA v. Jain is in §B.3. Re Porter; Re Transport Workers Union of Australia21 is used as a test case for the Employee specification. SHYSTER’s opinion on Re Porter; Re TWU is evaluated in §5.4.3 and its report file is in §B.4. Ainsworth v. Criminal Justice Commission22 is used in §5.5.3 as a test case for the Natural specification, and SHYSTER’s advice is evaluated there. SHYSTER prepares a report file for each of the three areas in that specification: the Expect­ ation, Affected and Natural areas. All three reports, combined, are in §B.5.23 (The report file for Building Workers’ Industrial Union of Australia v. Odco Pty Ltd 24 is given in §C.13 as part of the complete example, using the Employee specification, in appendix C.) B.2 Report file for Parker v. British Airways Finder area Instant case In the instant case, the finder was not the occupier of the premises where the chattel was found; the chattel was not attached; the other claimant was not the owner of the premises where the chattel was found; the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner; the finder handed over the chattel to the other claimant after the finding; neither party relied on the terms of an agreement regarding the right to the chattel; the finder was not a servant of the other claimant; the chattel was not hidden and was not in a position so as to be difficult to find; an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned; and neither party knew of the existence of the chattel prior to the finding. In my opinion—following Bridges v. Hawkesworth—the finder wins. In Bridges v. Hawkesworth, 1 an 1851 decision of the Queen’s Bench Division of the English High Court, the plaintiff found a bundle of banknotes on the floor of the public 1(1851) 21 LJQB 75. 285 � B.2 Report file for Parker v. British Airways area of a shop. He handed the notes to the shopkeeper in order that the true owner of the notes might be found. Although the owner was never found, the shopkeeper refused to return the notes to the finder. The Court found for the finder, holding that there is a “general right of [a] finder to any article which has been lost as against all the world except the true owner”.2 It was further noted that the notes had never been in the custody of the shopkeeper nor within the protection of his house as might be the case had they intentionally been deposited there. There are several significant similarities between the instant case and Bridges v. Hawkesworth: the finder was not the occupier of the premises where the chattel was found; the chattel was not attached; the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner; the finder handed over the chattel to the other claimant after the finding; neither party relied on the terms of an agreement regarding the right to the chattel; the finder was not a servant of the other claimant; the chattel was not hidden and was not in a position so as to be difficult to find; an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned; and neither party knew of the existence of the chattel prior to the finding. However, the instant case is not on all fours with Bridges v. Hawkesworth. In that case the other claimant was the owner of the premises where the chattel was found. Nevertheless, I believe that Bridges v. Hawkesworth should be followed. If City of London Corporation v. Appleyard (1) is followed then the finder loses. In City of London Corporation v. Appleyard (1), 3 a 1963 decision of the Queen’s Bench Division of the English High Court, workmen employed by Wates Ltd were engaged in cutting a key-way into a cellar wall for the purposes of securing a foundation when they found an old wall-safe built into a recess of the old wall. Inside was a wooden box which contained a large number of Bank of England notes. The notes were handed over to the City of London police who sought interpleader proceedings to determine who was entitled to the possession of the notes. Wates Ltd was an independent contractor engaged by Yorkwin Investments Ltd for a construction project. Yorkwin was lessee in possession of the property which was owned in fee simple by the City of London. The Court followed the decision in South Staffordshire Water Co. v. Sharman4 in holding that the occupier is, in the absence of a better title elsewhere, entitled to the possession of objects which are attached to or under the land. Consequently, since the notes were in a wooden box within a safe built into the wall of the old building, the safe formed part of the demised premises. Yorkwin, being in lawful possession of the premises, were in de facto possession of the safe, even though ignorant of its existence. Although Yorkwin was entitled to possession as against the finders, they in turn were displaced by the City of London which relied successfully on a term in the lease which granted them the right to certain objects found on the premises. 2ibid. at 77, per Patteson J. 3[1963] 1 WLR 982. 4[1896] 2 QB 44. 286 Appendix B: Example reports There are several similarities between the instant case and London v. Appleyard (1): the finder was not the occupier of the premises where the chattel was found; the other claimant was not the owner of the premises where the chattel was found; the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner; neither party relied on the terms of an agreement regarding the right to the chattel; an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned; and neither party knew of the existence of the chattel prior to the finding. However, there are several significant differences between the instant case and Lon­ don v. Appleyard (1). In that case the chattel was attached; the finder did not hand over the chattel to the other claimant after the finding; the finder was a servant of the other claimant; and the chattel was hidden or was in a position so as to be difficult to find. Despite the fact that London v. Appleyard (1) and Bridges v. Hawkesworth are both decisions of the Queen’s Bench Division of the English High Court, there is nothing in London v. Appleyard (1) to warrant any change in my conclusion. Hypothetical 1 Consider the instant case changed so that the following is true: the other claimant was the owner of the premises where the chattel was found. If that were so then I would be more strongly of the opinion that—following Bridges v. Hawkesworth—the finder wins. Details of Bridges v. Hawkesworth are summarized above. The hypothetical case is on all fours with Bridges v. Hawkesworth. If City of London Corporation v. Appleyard (2) or South Staffordshire Water Co. v. Sharman are followed then the finder loses. In City of London Corporation v. Appleyard (2), 5 a 1963 decision of the Queen’s Bench Division of the English High Court, workmen employed by Wates Ltd were engaged in cutting a key-way into a cellar wall for the purposes of securing a foundation when they found an old wall-safe built into a recess of the old wall. Inside was a wooden box which contained a large number of Bank of England notes. The notes were handed over to the City of London police who sought interpleader proceedings to determine who was entitled to the possession of the notes. Wates Ltd was an independent contractor engaged by Yorkwin Investments Ltd for a construction project. Yorkwin was lessee in possession of the property which was owned in fee simple by the City of London. The Court found that the safe formed part of the demised premises and that, consequently, Yorkwin was entitled to the notes as against the workmen. 5[1963] 1 WLR 982. 287 � B.2 Report file for Parker v. British Airways The lease contained a clause which purported to grant the rights to “every relic or article of antiquity rarity or value” to the City of London. The sole issue was to determine if the notes fell into that description. The Court could find no reason for limiting the generality of the words and so found for the City of London. There are several similarities between the hypothetical case and London v. Apple- yard (2): the other claimant was the owner of the premises where the chattel was found; the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner; the finder handed over the chattel to the other claimant after the finding; the finder was not a servant of the other claimant; an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned; and neither party knew of the existence of the chattel prior to the finding. However, there are several significant differences between the hypothetical case and London v. Appleyard (2). In that case the finder was the occupier of the premises where the chattel was found; the chattel was attached; one of the parties relied on the terms of an agreement made with the other which purported to give her/him the right to the chattel; and the chattel was hidden or was in a position so as to be difficult to find. Despite the fact that London v. Appleyard (2) and Bridges v. Hawkesworth are both decisions of the Queen’s Bench Division of the English High Court, there is nothing in London v. Appleyard (2) to warrant any change in my conclusion. In 1896, the Queen’s Bench Division of the English High Court also decided South Staffordshire Water Co. v. Sharman. 6 (Note, however, that London v. Appleyard (2) is 67 years more recent than South Staffordshire v. Sharman.) In South Staffordshire v. Sharman, the defendant was a workman employed by the plaintiff to clean out a pool located on land owned by the plaintiff. During the operation the defendant found two gold rings embedded in the mud at the bottom of the pool. Although the plaintiff demanded the rings, the defendant refused to give them up. He placed them in the hands of police authorities who unsuccessfully endeavoured to find the owners of the rings. The police returned the rings to the defendant who was then sued in detinue for the recovery of the rings. It was proved at the trial that there was no special contract between the parties which called upon the defendant to give up any articles which might be found. Although the county court held in favour of the defendant on the basis of Bridges v. Hawkesworth, 7 the appeal found for the plaintiff on the basis that they had, as owners of the land and pool, the right to exercise control over the same. Bridges v. Hawkesworth was distinguished on the grounds that the notes in that case were in a public part of the shop and the shopkeeper did not in any sense control them. The Court stated a general principle: where a person has possession of a house or land with a manifest intention to exercise control over it and the things which may be upon or in it, then there is a presumption that things found there are in the possession of the owner. 6[1896] 2 QB 44. 7(1851) 21 LJQB 75. 288 Appendix B: Example reports There are several similarities between the hypothetical case and South Stafford­ shire v. Sharman: the finder was not the occupier of the premises where the chattel was found; the other claimant was the owner of the premises where the chattel was found; the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner; neither party relied on the terms of an agreement regarding the right to the chattel; an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned; and neither party knew of the existence of the chattel prior to the finding. However, there are several significant differences between the hypothetical case and South Staffordshire v. Sharman. In that case the chattel was attached; the finder did not hand over the chattel to the other claimant after the finding; the finder was a servant of the other claimant; and the chattel was hidden or was in a position so as to be difficult to find. Despite the fact that South Staffordshire v. Sharman and Bridges v. Hawkesworth are both decisions of the Queen’s Bench Division of the English High Court, there is nothing in South Staffordshire v. Sharman to warrant any change in my conclusion. Hypothetical 2 Consider the instant case changed so that the following is true: the finder was a servant of the other claimant; and the chattel was hidden or was in a position so as to be difficult to find. If that were so then my opinion would be that—following City of London Corporation v. Appleyard (1)—the finder loses. Details of London v. Appleyard (1) are summarized above. There are several significant similarities between the hypothetical case and London v. Appleyard (1): the finder was not the occupier of the premises where the chattel was found; the other claimant was not the owner of the premises where the chattel was found; the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner; neither party relied on the terms of an agreement regarding the right to the chattel; the finder was a servant of the other claimant; the chattel was hidden or was in a position so as to be difficult to find; an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned; and neither party knew of the existence of the chattel prior to the finding. However, the hypothetical case is not on all fours with London v. Appleyard (1). In that case the chattel was attached; and the finder did not hand over the chattel to the other claimant after the finding. Nevertheless, I believe that London v. Appleyard (1) should be followed. If Hannah v. Peel is followed then the finder wins. In Hannah v. Peel, 8 a 1945 decision of the King’s Bench Division of the English High Court, a brooch was found by the plaintiff who was a lance-corporal stationed in a 8[1945] 1 KB 509. 289 � B.3 Report file for APRA v. Jain house owned by the defendant. The house had been requisitioned by the army during the war and had never been occupied by the defendant. The plaintiff was adjusting the black-out curtains when he touched something on the top of the window-frame. He thought the object to be a piece of dirt or plaster and he dropped it on the outside window ledge. On the following morning, he saw that it was a brooch and, on the advice of his commanding officer, turned it over to the police for the purpose of finding the owner. In the following year, the police returned the brooch to the defendant who sold it to a jeweller. The plaintiff at all times maintained his rights to the brooch against all persons other than the true owner. The Court found for the plaintiff on the basis of Bridges v. Hawkesworth9 after a thorough review of the authorities. The Court further noted that the defendant was never in possession of the premises, that the brooch was never his, and that he had no knowledge of it until it was brought to his notice by the finder. There are several similarities between the hypothetical case and Hannah v. Peel : the finder was not the occupier of the premises where the chattel was found; the chattel was not attached; the other claimant was not the true owner of the chattel and was not claiming through the rights of the true owner; neither party relied on the terms of an agreement regarding the right to the chattel; the chattel was hidden or was in a position so as to be difficult to find; an attempt was made to find the true owner of the chattel or, alternatively, the chattel was clearly abandoned; and neither party knew of the existence of the chattel prior to the finding. However, there are several significant differences between the hypothetical case and Hannah v. Peel. In that case the other claimant was the owner of the premises where the chattel was found; the finder did not hand over the chattel to the other claimant after the finding; and the finder was not a servant of the other claimant. Despite the fact that Hannah v. Peel is a decision of the King’s Bench Division of the English High Court (and as good authority as a case decided by the Queen’s Bench Division of the English High Court—like London v. Appleyard (1)), there is nothing in Hannah v. Peel to warrant any change in my conclusion. B.3 Report file for Australasian Performing Right Association Ltd v. Jain Authorization area Instant case The notion of authorization extends beyond the authority given to an agent. The word “authorize” should be “understood in its ordinary dictionary sense of ‘sanction, approve, and countenance.’ ”1 9(1851) 21 LJQB 75. 1Falcon v. Famous Players Film Co. [1926] 2 KB 474 at 491, per Bankes LJ. 290 Appendix B: Example reports “[A] person who has under his control the means by which an infringement of copyright may be committed . . . and who makes it available to other persons, knowing, or having reason to suspect, that it is likely to be used for the purpose of committing an infringement, and omitting to take reasonable steps to limit its use to legitimate purposes, would authorize any infringement that resulted from its use.”2 In the instant case, the infringer was not an employee of the accused; the infringer was an independent contractor to the accused; the accused did not sell or hire the infringer the means of infringing; the accused had the power to prevent the infringement; the accused did not take reasonable steps to avoid the infringement; the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done; and the specific infringement was not causally related to an incitement to infringe on the part of the accused. In my opinion—following Mellor v. Australian Broadcasting Commission—the accused authorized the infringement. In Mellor v. Australian Broadcasting Commission, 3 a 1940 decision of the Judicial Committee of the Privy Council, Mellor and others held the sole right to perform in public in Australia some musical works arranged for performance by brass and military bands. They published and distributed advertising pamphlets which included a statement that all of their sheet music was “ ‘Free for Public Performance’ anywhere . . . We have paid for the performing rights of every piece we issue.”4 The ABC engaged bands to play some of this music, and broadcast the bands’ performances on radio. The Privy Council held that the ABC had authorized the bands to perform the musical works within the meaning of s. 1(2) of the Copyright Act 1911 (UK) which was in force in Australia by virtue of the Copyright Act 1912 (Cth). However, the ABC had not infringed the plaintiffs’ sole right to authorize public performance because the statements made in the pamphlets amounted to consent. There are several significant similarities between the instant case and Mellor v. ABC : the infringer was not an employee of the accused; the infringer was an independ­ ent contractor to the accused; the accused did not sell or hire the infringer the means of infringing; the accused had the power to prevent the infringement; the accused did not take reasonable steps to avoid the infringement; and the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. However, the instant case is not on all fours with Mellor v. ABC. In that case the specific infringement was causally related to an incitement to infringe on the part of the accused. Nevertheless, I believe that Mellor v. ABC should be followed. If RCA Corporation v. John Fairfax and Sons Ltd is followed then the accused did not authorize the infringement. 2University of New South Wales v. Moorhouse (1975) 133 CLR 1 at 13, per Gibbs J. 3[1940] AC 491. 4ibid. at 498–9. 291 � B.3 Report file for APRA v. Jain In RCA Corporation v. John Fairfax and Sons Ltd, 5 a 1981 decision of the Supreme Court of New South Wales, the Fairfax newspaper the Sun-Herald carried an article which pointed out that, using cassette tapes and good quality taping equipment, the same album can be taped by many people. It also discussed how the advent of FM radio had made it easy for people to tape new album and single releases without buying the discs: “Why spend nearly $10 on the new David Bowie album when you can tape it from 2JJJ?”6 Kearney J held that “authorization involves some element of causation—and hence the necessity for some relationship creating a link or connection however tenuous between the authorizer and the infringer.”7 There was no such link, so Fairfax had not authorized any infringement within the meaning of s. 13(2) of the Copyright Act 1968 (Cth). There are several similarities between the instant case and RCA v. Fairfax : the infringer was not an employee of the accused; the accused did not sell or hire the infringer the means of infringing; the accused did not take reasonable steps to avoid the infringement; the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done; and the specific infringement was not causally related to an incitement to infringe on the part of the accused. However, there are two very significant differences between the instant case and RCA v. Fairfax. In that case the infringer was not an independent contractor to the accused; and the accused did not have the power to prevent the infringement. Despite the fact that RCA v. Fairfax is a decision of the Supreme Court of New South Wales (and better authority than a case decided by the Judicial Committee of the Privy Council—like Mellor v. ABC ), there is nothing in RCA v. Fairfax to warrant any change in my conclusion. If Australasian Performing Right Association Ltd v. Miles is followed then the accused is liable (directly or vicariously) for the infringement. In Australasian Performing Right Association Ltd v. Miles, 8 a 1961 decision of the Supreme Court of New South Wales, the Dee Why RSL Club engaged a band to play at a dance held at the club. During the dance the band played I’ve Got a Lovely Bunch of Coconuts, the copyright in which was owned by the Australasian Performing Right Association. Jacobs J held that the members of the band were servants of the club, because “the club through its officers was exercising a control over the work performed in such a way as to show that there was an authority to command the orchestra in its performance.”9 So the members of the club, through the band, performed the musical work and infringed the copyright under s. 2(1) of the Copyright Act 1911 (UK) which was in force in Australia by virtue of the Copyright Act 1912 (Cth). 5[1981] 1 NSWLR 251. 6ibid. at 252. 7ibid. at 259. 8[1962] NSWR 405. 9ibid. at 407. 292 Appendix B: Example reports There are several similarities between the instant case and APRA v. Miles: the accused did not sell or hire the infringer the means of infringing; the accused had the power to prevent the infringement; the accused did not take reasonable steps to avoid the infringement; and the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. However, there are several significant differences between the instant case and APRA v. Miles. In that case the infringer was an employee of the accused; the infringer was not an independent contractor to the accused; and the specific infringement was causally related to an incitement to infringe on the part of the accused. Despite the fact that APRA v. Miles is a decision of the Supreme Court of New South Wales (and better authority than a case decided by the Judicial Committee of the Privy Council—like Mellor v. ABC ), there is nothing in APRA v. Miles to warrant any change in my conclusion. Hypothetical 1 Consider the instant case changed so that the following is true: the specific infringement was causally related to an incitement to infringe on the part of the accused. If that were so then I would be more strongly of the opinion that—following Mellor v. Australian Broadcasting Commission—the accused authorized the infringement. Details of Mellor v. ABC are summarized above. The hypothetical case is on all fours with Mellor v. ABC. If RCA Corporation v. John Fairfax and Sons Ltd is followed then the accused did not authorize the infringement. Details of RCA v. Fairfax are summarized above. There are several similarities between the hypothetical case and RCA v. Fairfax : the infringer was not an employee of the accused; the accused did not sell or hire the infringer the means of infringing; the accused did not take reasonable steps to avoid the infringement; and the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. However, there are several significant differences between the hypothetical case and RCA v. Fairfax. In that case the infringer was not an independent contractor to the accused; the accused did not have the power to prevent the infringement; and the specific infringement was not causally related to an incitement to infringe on the part of the accused. Despite the fact that RCA v. Fairfax is a decision of the Supreme Court of New South Wales (and better authority than a case decided by the Judicial Committee of the Privy Council—like Mellor v. ABC ), there is nothing in RCA v. Fairfax to warrant any change in my conclusion. If Australasian Performing Right Association Ltd v. Miles is followed then the accused is liable (directly or vicariously) for the infringement. 293 � B.3 Report file for APRA v. Jain Details of APRA v. Miles are summarized above. There are several similarities between the hypothetical case and APRA v. Miles: the accused did not sell or hire the infringer the means of infringing; the accused had the power to prevent the infringement; the accused did not take reasonable steps to avoid the infringement; the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done; and the specific infringement was causally related to an incitement to infringe on the part of the accused. However, there are two very significant differences between the hypothetical case and APRA v. Miles. In that case the infringer was an employee of the accused; and the infringer was not an independent contractor to the accused. Despite the fact that APRA v. Miles is a decision of the Supreme Court of New South Wales (and better authority than a case decided by the Judicial Committee of the Privy Council—like Mellor v. ABC ), there is nothing in APRA v. Miles to warrant any change in my conclusion. Hypothetical 2 Consider the instant case changed so that the following is true: the accused did not know, and had no reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. If that were so then my opinion would be that—following Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd —the accused did not authorize the infringement. In Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd, 10 a 1923 decision of the English Court of Appeal, the syndicate was the lessee of a theatre. The managing- director of the syndicate produced a play at that theatre, and engaged a band to perform at the theatre under the direction of a bandmaster. In the absence of the managing-director, and without his knowledge, the band performed works the copyright in which was owned by the Performing Right Society. Bankes, Scrutton and Atkin LJJ held that the managing-director had not authorized the infringing performances, within the meaning of s. 1(2) of the Copyright Act 1911 (UK), because the infringement occurred without his knowledge and he had no reason to anticipate or suspect that the band was likely to give performances which would breach copyright. The hypothetical case is on all fours with PRS v. Ciryl. If Mellor v. Australian Broadcasting Commission is followed then the accused author­ ized the infringement. Details of Mellor v. ABC are summarized above. There are several similarities between the hypothetical case and Mellor v. ABC : the infringer was not an employee of the accused; the infringer was an independent contractor to the accused; the accused did not sell or hire the infringer the means of infringing; the accused had the power to prevent the infringement; and the accused did not take reasonable steps to avoid the infringement. 10[1924] 1 KB 1. 294 Appendix B: Example reports However, there are two very significant differences between the hypothetical case and Mellor v. ABC. In that case the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done; and the specific infringement was causally related to an incitement to infringe on the part of the accused. Despite the fact that Mellor v. ABC is a decision of the Judicial Committee of the Privy Council (and better authority than a case decided by the English Court of Appeal—like PRS v. Ciryl), there is nothing in Mellor v. ABC to warrant any change in my conclusion. If Australasian Performing Right Association Ltd v. Miles is followed then the accused is liable (directly or vicariously) for the infringement. Details of APRA v. Miles are summarized above. There are several similarities between the hypothetical case and APRA v. Miles: the accused did not sell or hire the infringer the means of infringing; the accused had the power to prevent the infringement; and the accused did not take reasonable steps to avoid the infringement. However, there are several significant differences between the hypothetical case and APRA v. Miles. In that case the infringer was an employee of the accused; the infringer was not an independent contractor to the accused; the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done; and the specific infringement was causally related to an incitement to infringe on the part of the accused. Despite the fact that APRA v. Miles is a decision of the Supreme Court of New South Wales (and better authority than a case decided by the English Court of Appeal— like PRS v. Ciryl), there is nothing in APRA v. Miles to warrant any change in my conclusion. Hypothetical 3 Consider the instant case changed so that the following is true: the infringer was an employee of the accused; and the infringer was not an independent contractor to the accused. If that were so then my opinion would be that—following Australasian Performing Right Association Ltd v. Miles—the accused is liable (directly or vicariously) for the infringement. Details of APRA v. Miles are summarized above. There are several significant similar­ ities between the hypothetical case and APRA v. Miles: the infringer was an employee of the accused; the infringer was not an independent contractor to the accused; the accused did not sell or hire the infringer the means of infringing; the accused had the power to prevent the infringement; the accused did not take reasonable steps to avoid the infringement; and the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done. However, the hypothetical case is not on all fours with APRA v. Miles. In that case the specific infringement was causally related to an incitement to infringe on the part of the accused. Nevertheless, I believe that APRA v. Miles should be followed. 295 � B.3 Report file for APRA v. Jain If University of New South Wales v. Moorhouse is followed then the accused authorized the infringement. In University of New South Wales v. Moorhouse, 11 a 1975 decision of three judges of the High Court of Australia, a graduate of the University used a photocopy machine in the University library to make two copies of a story from a library copy of a book of short stories. McTiernan ACJ, Gibbs and Jacobs JJ held that the University had authorized the infringement within the meaning of s. 36(1) of the Copyright Act 1968 (Cth); it had the power to prevent infringements, but had not taken reasonable steps to prevent them.12 Gibbs J’s statement about what constitutes authorization of an infringement is quoted above. There are several similarities between the hypothetical case and UNSW v. Moor- house: the infringer was not an independent contractor to the accused; the accused had the power to prevent the infringement; the accused did not take reasonable steps to avoid the infringement; the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done; and the specific infringement was not causally related to an incitement to infringe on the part of the accused. However, there are two very significant differences between the hypothetical case and UNSW v. Moorhouse. In that case the infringer was not an employee of the accused; and the accused sold or hired the infringer the means of infringing. Despite the fact that UNSW v. Moorhouse is a decision of three judges of the High Court of Australia (and better authority than a case decided by the Supreme Court of New South Wales—like APRA v. Miles), there is nothing in UNSW v. Moorhouse to warrant any change in my conclusion. If RCA Corporation v. John Fairfax and Sons Ltd is followed then the accused did not authorize the infringement. Details of RCA v. Fairfax are summarized above. There are several similarities between the hypothetical case and RCA v. Fairfax : the infringer was not an independent con­ tractor to the accused; the accused did not sell or hire the infringer the means of infringing; the accused did not take reasonable steps to avoid the infringement; the accused knew, or had reason to anticipate or suspect, that the infringing act was to be, or was likely to be, done; and the specific infringement was not causally related to an incitement to infringe on the part of the accused. However, there are two very significant differences between the hypothetical case and RCA v. Fairfax. In that case the infringer was not an employee of the accused; and the accused did not have the power to prevent the infringement. Despite the fact that RCA v. Fairfax and APRA v. Miles are both decisions of the Supreme Court of New South Wales, there is nothing in RCA v. Fairfax to warrant any change in my conclusion. 11(1975) 133 CLR 1. 12The addition, in 1980, of s. 39a to the Copyright Act ameliorated the effect of UNSW v. Moorhouse as far as photocopying in libraries is concerned. 296 Appendix B: Example reports B.4 Report file for Re Porter; Re Transport Workers Union of Australia Employee area Instant case The law distinguishes between a contract of service (between employer and employee) and a contract for services (between principal and independent contractor). This dis­ tinction affects the terms that will be implied in the absence of an express agreement, the liability of the employer to third parties, the applicability of industrial awards, the applicability of statutes which may affect workers’ compensation, occupational health and safety, long-service leave, fringe benefits tax, etc. The terms “employer” and “worker” are used here to mean “employer” and “em­ ployee” (in the case of a contract of service) or “principal” and “independent con­ tractor” (in the case of a contract for services). In the instant case, the employer directed the manner in which the work was to be done; the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was an integral part of the employer’s business; the worker owned the tools or provided the transport with which she/he performed the work; it is not known whether the employer would make a profit/loss if the work performed by the worker cost less/more than expected; the work was not performed on the employer’s premises; it is not known whether the employer supervised or inspected the work; the worker was not in business on her/his own account; the worker was not allowed to employ others to assist with her/his work; the worker was obliged to work only for the employer; the worker was not required to work at specified times; the employer paid the worker by time; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer deducted PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. In my opinion—following Ferguson v. John Dawson & Partners (Contractors) Ltd —the worker is an employee. In Ferguson v. John Dawson & Partners (Contractors) Ltd, 1 a 1976 decision of the Eng­ lish Court of Appeal, Ferguson fell off a roof while removing some scaffolding boards. He claimed damages against John Dawson (the building contractors) for breach of stat­ utory duty relying on the Construction (Working Places) Regulations 1966 (UK). This duty would only be owed if Ferguson was an employee of John Dawson. 1[1976] 1 WLR 1213. 297 � B.4 Report file for Re Porter; Re TWU Megaw and Browne LJJ held that, despite the fact that both parties labelled Fer­ guson a “self-employed labour only subcontractor”,2 the reality of the relationship between them was that of employer/employee. There are several significant similarities between the instant case and Ferguson v. Dawson: the employer directed the manner in which the work was to be done; the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was an integral part of the employer’s business; the work was not performed on the employer’s premises; the worker was not in business on her/his own account; the worker was not allowed to employ others to assist with her/his work; the employer paid the worker by time; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, the instant case is not on all fours with Ferguson v. Dawson. In that case the worker neither owned the tools nor provided the transport with which she/he performed the work; the employer would make a profit/loss if the work performed by the worker cost less/more than expected; the employer supervised or inspected the work; the worker was not obliged to work only for the employer; the worker was required to work at specified times; and the employer did not deduct PAYE tax instalments from the worker’s pay. Nevertheless, I believe that Ferguson v. Dawson should be followed. If Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National In­ surance is followed then the worker is an independent contractor. In Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance, 3 a 1967 decision of the Queen’s Bench Division of the English High Court, Latimer worked for Ready Mixed as an “owner-driver.” He was paid at mileage rates, and was obliged to buy the truck through a financial organization associated with Ready Mixed. The truck was painted in the company’s colours, and he had to wear a Ready Mixed uniform. Latimer was obliged to meet the costs of maintenance, repair and insurance of the truck (and the attached mixing unit, which belonged to Ready Mixed). The Minister determined that Latimer was employed under a contract of ser­ vice and therefore an “employed person” under s. 1(2) of the National Insurance Act 1965 (UK), making Ready Mixed liable to make weekly contributions. MacKenna J examined the contract and held that the rights it conferred, and the duties it imposed, between Latimer and Ready Mixed were not such as to make it a contract of service. There are several similarities between the instant case and Ready Mixed v. Minister : the worker was an integral part of the employer’s business; the worker owned the tools or provided the transport with which she/he performed the work; the work was not 2ibid. at 1219, per Megaw LJ; at 1225, per Lawton LJ; at 1228, per Browne LJ. 3[1968] 2 QB 497. 298 Appendix B: Example reports performed on the employer’s premises; the worker was not in business on her/his own account; the worker was obliged to work only for the employer; the worker was not required to work at specified times; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, there are several significant differences between the instant case and Ready Mixed v. Minister. In that case the employer did not direct the manner in which the work was to be done; the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the employer neither supervised nor inspected the work; the worker was allowed to employ others to assist with her/his work; the employer did not pay the worker by time; and the employer did not deduct PAYE tax instalments from the worker’s pay. Note also that Ready Mixed v. Minister is only a decision of the Queen’s Bench Division of the English High Court and not as good authority as a case decided by the English Court of Appeal—like Ferguson v. Dawson. Consequently, there is nothing in Ready Mixed v. Minister to warrant any change in my conclusion. Instantiation 4 It may be that the following is true of the instant case: the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; and the employer neither supervised nor inspected the work. If that is so then in my opinion—following Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance—the worker is an independent contractor. Details of Ready Mixed v. Minister are summarized above. There are several significant similarities between the instantiated case and Ready Mixed v. Minister : the worker was an integral part of the employer’s business; the worker owned the tools or provided the transport with which she/he performed the work; the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the work was not performed on the employer’s premises; the employer neither supervised nor inspected the work; the worker was not in business on her/his own account; the worker was obliged to work only for the employer; the worker was not required to work at specified times; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. 299 � B.4 Report file for Re Porter; Re TWU However, the instantiated case is not on all fours with Ready Mixed v. Minister. In that case the employer did not direct the manner in which the work was to be done; the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was allowed to employ others to assist with her/his work; the employer did not pay the worker by time; and the employer did not deduct PAYE tax instalments from the worker’s pay. Nevertheless, I believe that Ready Mixed v. Minister should be followed. If Ferguson v. John Dawson & Partners (Contractors) Ltd is followed then the worker is an employee. Details of Ferguson v. Dawson are summarized above. There are several similarities between the instantiated case and Ferguson v. Dawson: the employer directed the manner in which the work was to be done; the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was an integral part of the employer’s business; the work was not performed on the employer’s premises; the worker was not in business on her/his own account; the worker was not allowed to employ others to assist with her/his work; the employer paid the worker by time; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, there are several significant differences between the instantiated case and Ferguson v. Dawson. In that case the worker neither owned the tools nor provided the transport with which she/he performed the work; the employer would make a profit/loss if the work performed by the worker cost less/more than expected; the employer supervised or inspected the work; the worker was not obliged to work only for the employer; the worker was required to work at specified times; and the employer did not deduct PAYE tax instalments from the worker’s pay. Despite the fact that Ferguson v. Dawson is a decision of the English Court of Appeal (and better authority than a case decided by the Queen’s Bench Division of the English High Court—like Ready Mixed v. Minister ), there is nothing in Ferguson v. Dawson to warrant any change in my conclusion. Hypothetical 1 Consider the instant case changed so that the following is true: the worker was not obliged to work only for the employer; and the employer did not deduct PAYE tax instalments from the worker’s pay. If that were so then I would be more strongly of the opinion that—following Ferguson v. John Dawson & Partners (Contractors) Ltd —the worker is an employee. Details of Ferguson v. Dawson are summarized above. There are several significant similarities between the hypothetical case and Ferguson v. Dawson: the employer dir­ ected the manner in which the work was to be done; the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; 300 Appendix B: Example reports the worker was an integral part of the employer’s business; the work was not performed on the employer’s premises; the worker was not in business on her/his own account; the worker was not allowed to employ others to assist with her/his work; the worker was not obliged to work only for the employer; the employer paid the worker by time; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, the hypothetical case is not on all fours with Ferguson v. Dawson. In that case the worker neither owned the tools nor provided the transport with which she/he performed the work; the employer would make a profit/loss if the work performed by the worker cost less/more than expected; the employer supervised or inspected the work; and the worker was required to work at specified times. Nevertheless, I believe that Ferguson v. Dawson should be followed. If Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National In­ surance is followed then the worker is an independent contractor. Details of Ready Mixed v. Minister are summarized above. There are several similarities between the hypothetical case and Ready Mixed v. Minister : the worker was an integral part of the employer’s business; the worker owned the tools or provided the transport with which she/he performed the work; the work was not performed on the employer’s premises; the worker was not in business on her/his own account; the worker was not required to work at specified times; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, there are several significant differences between the hypothetical case and Ready Mixed v. Minister. In that case the employer did not direct the manner in which the work was to be done; the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the employer neither supervised nor inspected the work; the worker was allowed to employ others to assist with her/his work; the worker was obliged to work only for the employer; and the employer did not pay the worker by time. Note also that Ready Mixed v. Minister is only a decision of the Queen’s Bench Division of the English High Court and not as good authority as a case decided by the English Court of Appeal—like Ferguson v. Dawson. Consequently, there is nothing in Ready Mixed v. Minister to warrant any change in my conclusion. 301 � B.4 Report file for Re Porter; Re TWU Hypothetical 2 Consider the instant case changed so that the following is true: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; and the employer did not deduct PAYE tax instalments from the worker’s pay. If that were so then my opinion would be that—following Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance—the worker is an independ­ ent contractor. Details of Ready Mixed v. Minister are summarized above. There are several significant similarities between the hypothetical case and Ready Mixed v. Minister : the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was an integral part of the employer’s business; the worker owned the tools or provided the transport with which she/he performed the work; the work was not performed on the employer’s premises; the worker was not in business on her/his own account; the worker was obliged to work only for the employer; the worker was not required to work at specified times; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, the hypothetical case is not on all fours with Ready Mixed v. Minister. In that case the employer did not direct the manner in which the work was to be done; the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the employer neither supervised nor inspected the work; the worker was allowed to employ others to assist with her/his work; and the employer did not pay the worker by time. Nevertheless, I believe that Ready Mixed v. Minister should be followed. If Ferguson v. John Dawson & Partners (Contractors) Ltd is followed then the worker is an employee. Details of Ferguson v. Dawson are summarized above. There are several similarities between the hypothetical case and Ferguson v. Dawson: the employer directed the manner in which the work was to be done; the worker was an integral part of the employer’s business; the work was not performed on the employer’s premises; the worker was not in business on her/his own account; the worker was not allowed to employ others to assist with her/his work; the employer paid the worker by time; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express 302 Appendix B: Example reports any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, there are several significant differences between the hypothetical case and Ferguson v. Dawson. In that case the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker neither owned the tools nor provided the transport with which she/he performed the work; the employer would make a profit/loss if the work performed by the worker cost less/more than expected; the employer supervised or inspected the work; the worker was not obliged to work only for the employer; and the worker was required to work at specified times. Despite the fact that Ferguson v. Dawson is a decision of the English Court of Appeal (and better authority than a case decided by the Queen’s Bench Division of the English High Court—like Ready Mixed v. Minister ), there is nothing in Ferguson v. Dawson to warrant any change in my conclusion. B.5 Report files for Ainsworth v. Criminal Justice Commission Expectation area Instant case If the applicant had a legitimate expectation—or a reasonable expectation—which was affected by the decision, natural justice may be implied. “ ‘[L]egitimate expectations’ . . . are capable of including expectations which go beyond enforceable legal rights, provided they have some reasonable basis”. 1 In the instant case, the decision-maker did not break a promise or undertaking; the decision-maker did not go against an established course of practice; the decision did not involve a refusal to renew an existing interest; neither the decision-maker nor a statutory provision suggested that an initial interest would be granted; the decision did not affect an established liberty or interest; and there was no standard administrative procedure which the decision-maker should have followed. In my opinion—following Minister for Arts Heritage and Environment v. Peko-Wall­ send Ltd —the applicant did not have a legitimate expectation which was affected by the decision. In Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd, 2 a 1987 decision of the Full Court of the Federal Court of Australia, Peko-Wallsend held various mining interests in Stage 2 of Kakadu National Park. Federal Cabinet decided to nominate Stage 2 for inclusion in the World Heritage List, so it became “identified property” 1Cole v. Cunningham (1983) 49 ALR 123 at 131, per Bowen CJ, Sheppard and Morling JJ. 2(1987) 75 ALR 218. 303 � B.5 Report files for Ainsworth v. CJC within the meaning of s. 3(2) of the World Heritage Properties Conservation Act 1983 (Cth). This meant that the Governor-General could, by proclamation, make mining operations unlawful in the area. The decision did not affect Peko-Wallsend’s mining rights which were preserved under s. 8b of the National Parks and Wildlife Conservation Act 1975 (Cth). Before Cabinet’s decision, Peko-Wallsend had lobbied Ministers and other officials extensively, seeking to preserve their mining interests. After the decision they com­ menced proceedings to prevent the Government from taking any further steps to have Stage 2 nominated on the World Heritage List, claiming that Cabinet was bound by the rules of natural justice and had failed to give Peko-Wallsend an opportunity to be heard. Beaumont J (a Federal Court judge) agreed, and held the Cabinet decision void.3 The Full Court of the Federal Court disagreed. Bowen CJ decided that “it would . . . be inappropriate for this court to interfere to set aside a Cabinet decision involving such complex policy considerations”.4 Both Sheppard and Wilcox JJ held that Peko- Wallsend had had adequate opportunity to put their case to relevant Ministers and officials before the Cabinet decision, and were not denied natural justice.5 However, Wilcox J (with whose reasons the other two judges generally agreed) held that the Cabinet’s decision in this case did not attract the obligations of natural justice.6 The instant case is on all fours with Minister for Environment v. Peko-Wallsend. If Cole v. Cunningham is followed then the applicant had a legitimate expectation which was affected by the decision. In Cole v. Cunningham, 7 a 1983 decision of the Full Court of the Federal Court of Australia, Cunningham had been encouraged to resign from the Public Service because his superiors believed he had been guilty of misconduct in the performance of his duties. He had formed an attachment and begun to live with a Fijian women whose permit extension application he had processed. He was threatened with criminal prosecution and told that “[i]f you resign now it will be a normal resignation and you’ll leave with a clean record.”8 About eighteen months later, Cunningham sought reappointment to the Public Service and was told that he would be offered a position subject to police and ASIO clearances. The next day he was told that he have been given an unsatisfactory report based on the earlier events. Bowen CJ, Sheppard and Morling JJ held that, in general, applicants for appoint­ ment or reappointment to the public service are not entitled to natural justice because they have no legitimate expectation which can be affected by a refusal to appoint. However, Cunningham did have a legitimate expectation that any decision to reappoint him would not be made on the basis of his past record. 3Peko-Wallsend Ltd v. Minister for Arts Heritage and Environment (1986) 70 ALR 523. 4(1987) 75 ALR 218 at 225. 5ibid. at 228 per Sheppard J; at 253 per Wilcox J. 6ibid. at 253. 7(1983) 49 ALR 123. 8ibid. at 125. 304 Appendix B: Example reports There are several similarities between the instant case and Cole v. Cunningham: the decision-maker did not go against an established course of practice; the decision did not involve a refusal to renew an existing interest; neither the decision-maker nor a statutory provision suggested that an initial interest would be granted; the decision did not affect an established liberty or interest; and there was no standard administrative procedure which the decision-maker should have followed. However, there is one extremely significant difference between the instant case and Cole v. Cunningham. In that case the decision-maker broke a promise or undertaking. Despite the fact that Cole v. Cunningham and Minister for Environment v. Peko- Wallsend are both decisions of the Full Court of the Federal Court of Australia, there is nothing in Cole v. Cunningham to warrant any change in my conclusion. Hypothetical 1 Consider the instant case changed so that the following is true: the decision-maker broke a promise or undertaking. If that were so then my opinion would be that—following Cole v. Cunningham—the applicant had a legitimate expectation which was affected by the decision. Details of Cole v. Cunningham are summarized above. The hypothetical case is on all fours with Cole v. Cunningham. If Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd is followed then the applicant did not have a legitimate expectation which was affected by the decision. Details of Minister for Environment v. Peko-Wallsend are summarized above. There are several similarities between the hypothetical case and Minister for Environment v. Peko-Wallsend: the decision-maker did not go against an established course of practice; the decision did not involve a refusal to renew an existing interest; neither the decision- maker nor a statutory provision suggested that an initial interest would be granted; the decision did not affect an established liberty or interest; and there was no standard administrative procedure which the decision-maker should have followed. However, there is one extremely significant difference between the hypothetical case and Minister for Environment v. Peko-Wallsend. In that case the decision-maker did not break a promise or undertaking. Despite the fact that Minister for Environment v. Peko-Wallsend and Cole v. Cun­ ningham are both decisions of the Full Court of the Federal Court of Australia, there is nothing in Minister for Environment v. Peko-Wallsend to warrant any change in my conclusion. Affected area Instant case In the instant case, the decision did not affect a financial, property or occupational interest of the applicant; the decision did not affect the applicant’s personal liberty; the decision affected the applicant’s reputation; and the applicant did not have a legitimate expectation which was affected by the decision. 305 � B.5 Report files for Ainsworth v. CJC In my opinion—following Annetts v. McCann—the decision affected the property, right, interest, status, or legitimate expectation of the applicant. In Annetts v. McCann, 9 a 1990 decision of five judges of the High Court of Australia, a coroner had been conducting an inquest into the death of a 16-year old boy. The boy’s parents (Mr and Mrs Annetts) sought to make a submission before the coroner made a finding. The coroner decided that the Coroner’s Act 1920 (WA) gave him the discretion (which he chose to exercise) to disallow their submission. The Annettses appealed. The High Court (Mason CJ, Brennan, Deane, Toohey and McHugh JJ) held that their son’s reputation gave the Annettses an interest in the Coroner’s inquiry. “A finding in an inquest into a death is naturally likely to deal with the conduct of the deceased leading to death. An unfavourable reflection on the deceased is usually a matter of concern to her or his parents, spouse or children and, if they choose to appear at the inquest in order to safeguard the reputation of the deceased, the familial relationship suffices, in my view, to establish the deceased’s reputation as a relevant interest which should not be adversely affected without according natural justice to those who are seeking to safeguard that reputation.”10 The Court held that the fact that the coroner’s decision was merely recommendat­ ory (whether or not to prosecute) was not sufficient to avoid the implication of natural justice; the coroner was bound to hear the Annettses before making any finding adverse to them or their son.11 The instant case is on all fours with Annetts v. McCann. If Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd is followed then the decision did not affect the property, right, interest, status, or legitimate expectation of the applicant. In Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd, 12 a 1987 decision of the Full Court of the Federal Court of Australia, Peko-Wallsend held various mining interests in Stage 2 of Kakadu National Park. Federal Cabinet decided to nominate Stage 2 for inclusion in the World Heritage List, so it became “identified property” within the meaning of s. 3(2) of the World Heritage Properties Conservation Act 1983 (Cth). This meant that the Governor-General could, by proclamation, make mining operations unlawful in the area. The decision did not affect Peko-Wallsend’s mining rights which were preserved under s. 8b of the National Parks and Wildlife Conservation Act 1975 (Cth). 9(1990) 170 CLR 596. 10ibid. at 612, per Brennan J. 11ibid. at 603, per Mason CJ, Deane and McHugh JJ; at 612 per Brennan J; at 621 per Toohey J. Note, however, that Brennan and Toohey JJ dismissed the appeal because they believed that the decision of the Full Court of the Supreme Court of Western Australia (from which the Annettses appealed) was right on the material before it. 12(1987) 75 ALR 218. 306 Appendix B: Example reports Before Cabinet’s decision, Peko-Wallsend had lobbied Ministers and other officials extensively, seeking to preserve their mining interests. After the decision they com­ menced proceedings to prevent the Government from taking any further steps to have Stage 2 nominated on the World Heritage List, claiming that Cabinet was bound by the rules of natural justice and had failed to give Peko-Wallsend an opportunity to be heard. Beaumont J (a Federal Court judge) agreed, and held the Cabinet decision void.13 The Full Court of the Federal Court disagreed. Bowen CJ decided that “it would . . . be inappropriate for this court to interfere to set aside a Cabinet decision involving such complex policy considerations”.14 Both Sheppard and Wilcox JJ held that Peko- Wallsend had had adequate opportunity to put their case to relevant Ministers and officials before the Cabinet decision, and were not denied natural justice.15 However, Wilcox J (with whose reasons the other two judges generally agreed) held that the Cabinet’s decision in this case did not attract the obligations of natural justice.16 There are several similarities between the instant case and Minister for Environ­ ment v. Peko-Wallsend : the decision did not affect a financial, property or occupational interest of the applicant; the decision did not affect the applicant’s personal liberty; and the applicant did not have a legitimate expectation which was affected by the decision. However, there is one extremely significant difference between the instant case and Minister for Environment v. Peko-Wallsend. In that case the decision did not affect the applicant’s reputation. Note also that Minister for Environment v. Peko-Wallsend is only a decision of the Full Court of the Federal Court of Australia and not as good authority as a case decided by five judges of the High Court of Australia—like Annetts v. McCann. Consequently, there is nothing in Minister for Environment v. Peko-Wallsend to warrant any change in my conclusion. Hypothetical 1 Consider the instant case changed so that the following is true: the decision did not affect the applicant’s reputation. If that were so then my opinion would be that—following Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd—the decision did not affect the property, right, interest, status, or legitimate expectation of the applicant. Details of Minister for Environment v. Peko-Wallsend are summarized above. The hypothetical case is on all fours with Minister for Environment v. Peko-Wallsend. If Bread Manufacturers of New South Wales v. Evans is followed then the decision affected the property, right, interest, status, or legitimate expectation of the applicant. 13Peko-Wallsend Ltd v. Minister for Arts Heritage and Environment (1986) 70 ALR 523. 14(1987) 75 ALR 218 at 225. 15ibid. at 228 per Sheppard J; at 253 per Wilcox J. 16ibid. at 253. 307 � B.5 Report files for Ainsworth v. CJC In Bread Manufacturers of New South Wales v. Evans, 17 a 1981 decision of five judges of the High Court of Australia, the Bread Manufacturers claimed that an order made by the Prices Commission was void. The order affected the classification of bread products and had an incidental effect on the price of hamburger buns. The Bread Manufacturers complained that they should have been given the right to put their case to the Commission. The Prices Regulation Act 1948 (NSW) provided that a public inquiry had to be held before an order could be made setting prices, except where the Minister consented to dispensing with the inquiry. The Minister had dispensed with an inquiry before this order was made. Hence, “[t]he argument that the Commission was bound to disclose to the Association the fact that it proposed to make an order which would have the incidental effect of reducing the price of hamburger buns can only succeed if the Commission, although not bound to hold an inquiry, was bound to observe the rules of natural justice”.18 The High Court held that there was no denial of natural justice in relation to the order, because “the reduction of the maximum price in respect of one item was simply a minor incident in a major revision of the price framework covering the whole range of bread products. The effect of that major revision was generally to increase prices. There was, in our opinion, no obligation on the Commission to give advance notice of this development or of the possibility of its occurrence.”19 There are several similarities between the hypothetical case and Bread Manufactur­ ers v. Evans: the decision did not affect the applicant’s personal liberty; the decision did not affect the applicant’s reputation; and the applicant did not have a legitimate expectation which was affected by the decision. However, there is one extremely significant difference between the hypothetical case and Bread Manufacturers v. Evans. In that case the decision affected a financial, property or occupational interest of the applicant. Despite the fact that Bread Manufacturers v. Evans is a decision of five judges of the High Court of Australia (and better authority than a case decided by the Full Court of the Federal Court of Australia—like Minister for Environment v. Peko-Wallsend), there is nothing in Bread Manufacturers v. Evans to warrant any change in my con­ clusion. Natural area Instant case In recent years courts have tended to imply a duty to observe the principles of natural justice. It has been said that “[t]he law has now developed to a point where it may be accepted that there is a common law duty to act fairly, in the sense of according pro­ cedural fairness, in the making of administrative decisions which affect rights, interests 17(1981) 38 ALR 93. 18ibid. at 101, per Gibbs CJ. 19ibid. at 119, per Mason and Wilson JJ, with whom Murphy and Aickin JJ agreed on this point. 308 Appendix B: Example reports and legitimate expectations, subject only to the clear manifestation of a contrary stat­ utory intention.”20 However, there are some circumstances in which a duty to observe natural justice will not be implied: “the law has not yet reached the stage of applying the obligation of natural justice to every decision which disadvantages individuals.”21 In the instant case, the decision affected the property, right, interest, status, or legit­ imate expectation of the applicant; the decision is apt to have a discrete impact on the interests of the applicant; the power is of a nature that would suggest that procedural fairness would be applied; the statutory or factual criteria focused on matters which were discrete to the interests of the applicant; the decision-maker was not a high-level policy-maker; there is no statutory right to appeal against the decision; and there were no circumstances which would have made an obligation to observe natural justice inappropriate. In my opinion—following Annetts v. McCann—a duty to observe natural justice is implied. In Annetts v. McCann, 22 a 1990 decision of five judges of the High Court of Australia, a coroner had been conducting an inquest into the death of a 16-year old boy. The boy’s parents (Mr and Mrs Annetts) sought to make a submission before the coroner made a finding. The coroner decided that the Coroner’s Act 1920 (WA) gave him the discretion (which he chose to exercise) to disallow their submission. The Annettses appealed. The High Court (Mason CJ, Brennan, Deane, Toohey and McHugh JJ) held that their son’s reputation gave the Annettses an interest in the Coroner’s inquiry. “A finding in an inquest into a death is naturally likely to deal with the conduct of the deceased leading to death. An unfavourable reflection on the deceased is usually a matter of concern to her or his parents, spouse or children and, if they choose to appear at the inquest in order to safeguard the reputation of the deceased, the familial relationship suffices, in my view, to establish the deceased’s reputation as a relevant interest which should not be adversely affected without according natural justice to those who are seeking to safeguard that reputation.”23 The Court held that the fact that the coroner’s decision was merely recommendat­ ory (whether or not to prosecute) was not sufficient to avoid the implication of natural justice; the coroner was bound to hear the Annettses before making any finding adverse to them or their son.24 The instant case is on all fours with Annetts v. McCann. 20Kioa v. West (1985) 159 CLR 550 at 584, per Mason J. 21Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd (1987) 75 ALR 218 at 251, per Wilcox J. 22(1990) 170 CLR 596. 23ibid. at 612, per Brennan J. 24ibid. at 603, per Mason CJ, Deane and McHugh JJ; at 612 per Brennan J; at 621 per Toohey J. Note, however, that Brennan and Toohey JJ dismissed the appeal because they believed that the decision of the Full Court of the Supreme Court of Western Australia (from which the Annettses appealed) was right on the material before it. 309 � B.5 Report files for Ainsworth v. CJC If McInnes v. Onslow Fane is followed then a duty to observe natural justice is not implied. In McInnes v. Onslow Fane, 25 a 1978 decision of the Chancery Division of the English High Court, McInnes had held, at various times, licences to promote, train and act as master of ceremonies in professional boxing. All his licences were revoked by the British Boxing Board of Control. He made five unsuccessful applications for a manager’s licence. With his sixth application he requested an oral hearing and prior notification of anything that might prevent the area council (to which he applied) making a favourable recommendation to the board. The board refused his applications without giving him an oral hearing or informing him of the case against him. Megarry V-C held that the board was under no duty to provide reasons to McInnes or to allow him a hearing: “This is not a case in which there has been any suggestion of the board considering any alleged dishonesty or morally culpable conduct of the plaintiff. A man free from any moral blemish may nevertheless be wholly unsuitable for a particular type of work . . . In such circumstances, in the absence of anything to suggest that the board have been affected by dishonesty or bias or caprice, or that there is any other impropriety, I think that the board are fully entitled to give no reasons for their decision, and to decide the application without any preliminary indication to the plaintiff of those reasons. The board are the best judges of the desirability of granting a licence, and in the absence of any impropriety the court ought not to interfere.”26 The instant case is on all fours with McInnes v. Onslow Fane. Note, however, that McInnes v. Onslow Fane is only a decision of the Chancery Division of the English High Court and not as good authority as a case decided by five judges of the High Court of Australia—like Annetts v. McCann. Consequently, there is nothing in McInnes v. Onslow Fane to warrant any change in my conclusion. Hypothetical 1 Consider the instant case changed so that the following is true: the statutory or factual criteria focused on matters of policy or public interest; and the decision-maker was a high-level policy-maker. If that were so then my opinion would be that—following South Australia v. O’Shea—a duty to observe natural justice is not implied. In South Australia v. O’Shea, 27 a 1987 decision of five judges of the High Court of Aus­ tralia, O’Shea had been convicted of two offences of indecent assault of young children. He was released on licence and remained at liberty after the licence expired. Over a year later, after allegations had been made against him, O’Shea was apprehended and detained. The parole board recommended his release on licence on various conditions, 25[1978] 3 All ER 211. 26ibid. at 223. 27(1987) 163 CLR 378. 310 Appendix B: Example reports but the Governor in Council resolved to take no action. O’Shea had been given a hearing by the Parole Board, but he claimed he was entitled to a further hearing before the Governor in Council could exercise his discretionary powers under s. 77a(7a) of the Criminal Law Consolidation Act, 1935 (SA). Mason CJ, Wilson, Brennan and Toohey JJ (Deane J dissenting) held that O’Shea was not entitled to a further hearing. “Given the nature of this decision, it cannot be said that Mr O’Shea could have more than a hope that the Governor would be prepared to act on the recommendation of the Board. Hope, of itself, is not sufficient to ground an expectation that will attract legal consequences. So far as the concept of legitimate expectation is concerned, Mr O’Shea must be taken to know that the Act committed to the Governor, with the advice and consent of the Executive Council, the responsibility for determining where the public interest lay . . . The nature of the decision that they were required to make was such that participation by Mr O’Shea was inappropriate.”28 The hypothetical case is on all fours with SA v. O’Shea. If Macrae v. Attorney-General for New South Wales is followed then a duty to observe natural justice is implied. In Macrae v. Attorney-General for New South Wales, 29 a 1987 decision of the New South Wales Court of Appeal, five magistrates who had been appointed under the Justices Act 1902 (NSW) were not appointed under the Local Courts Act 1982 (NSW). The new Act had reorganized the magistracy in NSW, and magistrates appointed under the old Act were entitled to apply for appointment as magistrates under the new Act. The five had applied and were interviewed. Allegations were made privately to the Attorney-General claiming that they were unfit to be appointed, but these allegations were not brought to their notice at the time of the interviews. The Court of Appeal held that the Attorney-General’s decision not to recommend the appointment of the magistrates was void because they were denied their legitimate expectation of procedural fairness. “They have not been treated fairly.”30 The hypothetical case is on all fours with Macrae v. AG for NSW. Note, however, that Macrae v. AG for NSW is only a decision of the New South Wales Court of Appeal and not as good authority as a case decided by five judges of the High Court of Australia—like SA v. O’Shea. Consequently, there is nothing in Macrae v. AG for NSW to warrant any change in my conclusion. 28ibid. at 402, per Wilson and Toohey JJ. 29(1987) 9 NSWLR 268. 30ibid. at 283, per Kirby P. C A complete example . . . in order to enter the magic circle of employees, the worker has to solve not only the riddle of “contract” but also that of “employment.” And no sooner has the hapless worker tried to answer the riddles, but the judicial wizards have changed the question. Bob Hepple (1986) Restructuring Employment Rights25 Fury said to a mouse, That he met in the house, “Let us both go to law: I will prosecute you. Come, I’ll take no denial: We must have the trial; For really this morning I’ve nothing to do.” Said the mouse to the cur, “Such a trial, dear sir, With no jury or judge, would be wasting our breath.” “I’ll be judge, I’ll be jury,” said cunning old Fury: “I’ll try the whole cause, and condemn you to death.” Lewis Carroll (1865) Alice’s Adventures in Wonderland 26 311 312 Appendix C: A complete example C.1 Introduction This appendix gives a complete example of SHYSTER’s input and output files for one of the cases used to test the Employee specification in §5.4.3: Building Workers’ Industrial Union of Australia v. Odco Pty Ltd. 27 The log file (§C.2) summarizes SHYSTER’s operation for this test case. The complete case law specification file for the Employee specification (Employee.cls) is given in §C.3. The only other input to SHYSTER for this example is the test case’s fact vector as entered by the user, one attribute value at a time, in response to the attribute questions asked by SHYSTER. The fact vector for BWIU v. Odco is given in the log file. SHYSTER was requested to hypothesize with a limit of two changed attribute values, and to report on up to one hypothetical per result. The dump file Dump.tex is shown in raw form as output by SHYSTER (i.e. ready to be input by LaTEX) in §C.4; the dump file as output by LaTEX appears with those of the other specifications in appendix A (§A.4). The remaining files are shown here as both LaTEX input and LaTEX output: viz. the probabilities file Probabilities.tex (§C.6 and §C.7), the weights file Weights.tex (§C.8 and §C.9), the distances file Distances-Employee.tex (§C.10 and §C.11) and the report file Report-Employee.tex (§C.12 and §C.13). Most of the LaTEX input shown here is in expurgated form; 28 vertical ellipses are used to indicate elision. A � symbol indicates that the line above is too wide ⇐ to fit on the page, and continues after the symbol. The level of indentation of a long line is maintained by indenting the continuation appropriately. LaTEX output is complete, and unaltered—except that two formatting changes were made so that all of the output would fit within the margins of this thesis: the probabilities matrix in §C.7 has been reduced in size, and all seven distance tables in §C.11 have been rotated anti-clockwise by 90 degrees. C.2 Log file for Building Workers’ Industrial Union of Australia v. Odco Pty Ltd SHYSTER version 1.0 Copyright James Popple 1993 Reading case law specification from "Employee.cls" ... 9 courts in the hierarchy. Employee area: 2 results 18 attributes 14 cases 2 ideal points Case law specification is valid. � � 313 � C.2 Log file for BWIU v. Odco Writing dump to "Dump.tex". Writing probabilities to "Probabilities.tex". WARNING (Checker): evidence of stochastic dependence between A3 and A17 in Employee area. WARNING (Checker): functional dependence (inverse) between A4 and A5 in Employee area. WARNING (Checker): evidence of stochastic dependence between A4 and A11 in Employee area. WARNING (Checker): evidence of stochastic dependence between A5 and A11 in Employee area. WARNING (Checker): evidence of stochastic dependence between A7 and A9 in Employee area. WARNING (Checker): evidence of stochastic dependence between A9 and A11 in Employee area. WARNING (Checker): evidence of stochastic dependence between A11 and A12 in Employee area. Writing weights to "Weights.tex". Case-based system called with area identifier "Employee". Area is Employee. Writing distances to "Distances-Employee.tex". Writing report to "Report-Employee.tex". Fact vector is (NYNYNNNUNNYUNNNNNY). Nearest neighbours: Contractor: C8 Humberstone v. NTM Nearest others: Employee: C5 Ferguson v. Dawson Nearest result for the instant case is Contractor. Instantiation 1 is (NYNYNNNYNNYYNNNNNY). Nearest neighbours: Contractor: C8 Humberstone v. NTM Nearest others: Employee: C5 Ferguson v. Dawson � � � � 314 Appendix C: A complete example Safeguards: Distance measures: * C6 Stevenson v. Macdonald (2) (Employee) + C11 AMP v. Chaplin + C13 Stevenson v. Macdonald (1) Association coefficients: + C11 AMP v. Chaplin Correlation coefficients: - C8 Humberstone v. NTM + C11 AMP v. Chaplin Weighted correlation coefficients: - C8 Humberstone v. NTM + C11 AMP v. Chaplin Nearest result for instantiation 1 is Contractor. Instantiation 2 is (NYNYNNNYNNYNNNNNNY). Nearest neighbours: Contractor: C8 Humberstone v. NTM Nearest others: Employee: C2 Cam v. Sargent Safeguards: Distance measures: + C11 AMP v. Chaplin Association coefficients: + C11 AMP v. Chaplin Correlation coefficients: - C8 Humberstone v. NTM + C11 AMP v. Chaplin Weighted correlation coefficients: - C8 Humberstone v. NTM + C11 AMP v. Chaplin Nearest result for instantiation 2 is Contractor. Instantiation 3 is (NYNYNNNNNNYYNNNNNY). Nearest neighbours: Contractor: C8 Humberstone v. NTM Nearest others: Employee: C5 Ferguson v. Dawson � � � � �� 315 � C.2 Log file for BWIU v. Odco Safeguards: Distance measures: * C6 Stevenson v. Macdonald (2) (Employee) + C13 Stevenson v. Macdonald (1) Nearest result for instantiation 3 is Contractor. Instantiation 4 is (NYNYNNNNNNYNNNNNNY). Nearest neighbours: Contractor: C8 Humberstone v. NTM Nearest others: Employee: C2 Cam v. Sargent Nearest result for instantiation 4 is Contractor. Hypothetical 1 is (NYNYNNNUYNYUNNNNNN). Nearest neighbours: Contractor: C8 Humberstone v. NTM Nearest others: Employee: C2 Cam v. Sargent Nearest result for hypothetical 1 is Contractor. Hypothetical 2 is (NNYYNNNUNNYUNNNNNY). Nearest neighbours: Employee: C5 Ferguson v. Dawson Nearest others: Contractor: C11 AMP v. Chaplin C14 Ready Mixed v. Minister Safeguards: Distance measures: + C6 Stevenson v. Macdonald (2) * C11 AMP v. Chaplin (Contractor) * C14 Ready Mixed v. Minister (Contractor) Association coefficients: * C11 AMP v. Chaplin (Contractor) * C14 Ready Mixed v. Minister (Contractor) Correlation coefficients: * C11 AMP v. Chaplin (Contractor) * C14 Ready Mixed v. Minister (Contractor) 316 Appendix C: A complete example Ideal points: Contractor Centroids: Contractor Specified directions: Contractor Ideal point directions: Contractor Centroid directions: Contractor WARNING (Reporter): the specified directions suggest a different result or results. Nearest result for hypothetical 2 is Employee. All 4 instantiations have the same nearest result as does the instant case. Reported on 2 hypotheticals of 136 (limit of 2 changes). Case-based system returned result identifier "Contractor". Finished. C.3 EMPLOYEE specification file HIERARCHY = HC-5 HC-4 HC-3 HC FCA-3 PC CA KB QB "five judges of the High Court of Australia" "four judges of the High Court of Australia" "three judges of the High Court of Australia" "a single judge of the High Court of Australia" "three judges of the Federal Court of Australia" "the Judicial Committee of the Privy Council" "the English Court of Appeal" "the King’s Bench Division of the English High Court" "the Queen’s Bench Division of the English High Court" AREA Employee OPENING "The law distinguishes between a contract of service (between employer and employee) and a contract for services (between principal and independent contractor). This distinction affects the terms that will be implied in the absence of an express agreement, the liability of the employer to third parties, the applicability of industrial awards, the applicability of statutes which may affect workers’ compensation, occupational health and safety, long-service leave, fringe benefits tax, etc.\par The terms ‘‘employer’’ and ‘‘worker’’ are used here to mean ‘‘employer’’ and ‘‘employee’’ (in the case of a contract of service) or ‘‘principal’’ and ‘‘independent contractor’’ (in the case of a contract for services)." � C.3 EMPLOYEE specification file 317 RESULTS Employee Contractor "the worker is "the worker is an an employee" independent contractor" ATTRIBUTE % Employer had control over manner in which work was done QUESTION "Did the employer direct not only what work was to but also the manner in which it was to be done" be done, YES NO "the employer directed the manner in which the work was to be done" Employee "the employer did not direct the manner in which the work was to be done" Contractor UNKNOWN "it is not known whether the employer directed the which the work was to be done" manner in HELP "If the employer had a right of control over how the worker did the work then the employer had the power to direct not only what work was to be done, but also the manner in which it was to be done." ATTRIBUTE % Worker had discretion as to how to do work QUESTION YES "Was the worker allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand" "the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand" Contractor NO UNKNOWN "the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand" Employee "it is not known whether the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand" ATTRIBUTE % Worker was an integral part of employer’s business QUESTION YES NO UNKNOWN HELP "Was the worker an integral part of the employer’s business" "the worker was an integral part of the employer’s business" Employee "the worker was not an integral part of the employer’s business, but was accessory to it" Contractor "it is not known whether the worker was an integral part of the employer’s business or was merely accessory to it" "If the worker was ‘‘part and parcel’’ of the employer’s business then she/he was an integral part of the business, not merely accessory to it." ATTRIBUTE % Worker used own tools or provided transport QUESTION YES NO UNKNOWN "Did the worker own the tools or provide the transport with which she/he performed the work" "the worker owned the tools or provided the transport with which she/he performed the work" Contractor "the worker neither owned the tools nor provided the transport with which she/he performed the work" Employee "it is not known whether the worker owned the tools or provided the transport with which she/he performed the work" 318 Appendix C: A complete example ATTRIBUTE % Employer would make profit/loss QUESTION YES NO UNKNOWN "Would the employer make a profit/loss if the work performed by the worker cost less/more than expected" "the employer would make a profit/loss if the work performed by the worker cost less/more than expected" "the employer would not make a profit/loss if the work performed by the worker cost less/more than expected" "it is not known whether the employer would make a profit/loss if the work performed by the worker cost less/more than expected" ATTRIBUTE % Work performed on employer’s premises QUESTION YES NO UNKNOWN "Was the work performed on the employer’s premises" "the work was performed on the employer’s premises" "the work was not performed on the employer’s premises" Contractor "it is not known whether the work was performed on the employer’s premises" ATTRIBUTE % Employer supervised/inspected work QUESTION YES NO UNKNOWN "Did the employer supervise or inspect the work" "the employer supervised or inspected the work" Employee "the employer neither supervised nor inspected the work" Contractor "it is not known whether the employer supervised or inspected the work" ATTRIBUTE % Worker in business on her/his own account QUESTION YES NO UNKNOWN "Was the worker in business on her/his own account" "the worker was in business on her/his own account" Contractor "the worker was not in business on her/his own account" Employee "it is not known whether the worker was in business on her/his own account" ATTRIBUTE % Worker could sub-contract QUESTION YES NO UNKNOWN "Was the worker allowed to employ others to assist with her/his work" "the worker was allowed to employ others to assist with her/his work" Contractor "the worker was not allowed to employ others to assist with her/his work" Employee "it is not known whether the worker was allowed to employ others to assist with her/his work" ATTRIBUTE % Worker obliged to work only for employer QUESTION YES "Was the worker obliged "the worker was obliged Employee to to work only for the employer" work only for the employer" � C.3 EMPLOYEE specification file 319 NO UNKNOWN "the worker was not obliged to work only for the employer" Contractor "it is not known whether the worker was obliged to work only for the employer" ATTRIBUTE % Worker required to work at specified times QUESTION YES NO UNKNOWN "Was the worker required to work at specified times" "the worker was required to work at specified times" Employee "the worker was not required to work at specified times" Contractor "it is not known whether the worker was required to work at specified times" ATTRIBUTE % Worker paid by time QUESTION YES NO UNKNOWN HELP "Did the employer pay the worker by time" "the employer paid the worker by time" Employee "the employer did not pay the worker by time" Contractor "it is not known whether the employer paid the worker by time" "The employer could pay the worker by time (e.g. by the hour, or by the week) or by results." ATTRIBUTE % Payment was called a "fee" QUESTION YES NO UNKNOWN "Was the money that the employer paid to the worker stated to be a ‘‘fee’’" "the money that the employer paid to the worker was stated to be a ‘‘fee’’" Contractor "the money that the employer paid to the worker was not stated to be a ‘‘fee’’" Employee "it is not known whether the money that the employer paid to the worker was stated to be a ‘‘fee’’" ATTRIBUTE % Payment was called "wages" or "salary" QUESTION YES NO UNKNOWN "Was the money that the employer paid to the worker stated to be ‘‘wages’’ or ‘‘salary’’" "the money that the employer paid to the worker was stated to be ‘‘wages’’ or ‘‘salary’’" Employee "the money that the employer paid to the worker was not stated to be ‘‘wages’’ or ‘‘salary’’" Contractor "it is not known whether the money that the employer paid to the worker was stated to be ‘‘wages’’ or ‘‘salary’’" ATTRIBUTE % Employer deducted PAYE tax instalments from worker’s pay QUESTION YES NO UNKNOWN "Did the employer deduct PAYE tax instalments from the worker’s pay" "the employer deducted PAYE tax instalments from the worker’s pay" Employee "the employer did not deduct PAYE tax instalments from the worker’s pay" Contractor "it is not known whether the employer deducted PAYE tax instalments from the worker’s pay" 320 Appendix C: A complete example ATTRIBUTE % Employer paid worker sick/holiday pay QUESTION YES NO UNKNOWN "Did the employer pay the worker sick pay or holiday pay" "the employer paid the worker sick pay or holiday pay" Employee "the employer paid the worker neither sick pay nor holiday pay" Contractor "it is not known whether the employer paid the worker sick pay or holiday pay" ATTRIBUTE % Expressed intention: employee/employer QUESTION YES NO UNKNOWN HELP "Did the employer and the worker express an intention that the relationship would be one of employer and employee" "the employer and the worker expressed an intention that the relationship would be one of employer and employee" Employee "the employer and the worker did not express any intention that the relationship would be one of employer and employee" "it is not known whether the employer and the worker expressed an intention that the relationship would be one of employer and employee" "For example, if the employer and the worker characterized their agreement as being a ‘‘contract of service,’’ that would be an expression of an intention that the relationship would be one of employer and employee." ATTRIBUTE % Expressed intention: principal/independent contractor QUESTION "Did the employer and the worker express an intention that the relationship would be one of principal and independent contractor" YES "the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor" Contractor NO UNKNOWN HELP "the employer and the worker did not express any intention that the relationship would be one of principal and independent contractor" "it is not known whether the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor" "For example, if the employer and the worker characterized their agreement as being a ‘‘contract for services,’’ that would be an expression of an intention that the relationship would be one of principal and independent contractor." CASE "Performing Right Society Ltd Danse) Ltd" "PRS v. Palais de Danse" v. Mitchell & Booker (Palais de CITATION YEAR "[1924] 1 KB 762" 1924 COURT KB FACTS RESULT (YYYYNYYNNYYYNYNNNN) Employee � C.3 EMPLOYEE specification file 321 SUMMARY "the defendants were the occupiers of a dance hall. They engaged a band to provide music in the hall. The agreement provided that the band should not infringe copyright, and that the band would be liable for damages and costs caused by any such infringement. There was also a notice displayed in the hall stating that ‘‘[o]nly such music as may be played without fee or licence is allowed to be played in this Hall.’’\footnote{ibid. at 764.}\par The band performed several pieces of music, the copyright in which was held by the Performing Right Society, without its permission. The defendants did not know, and had no reasonable grounds for suspecting, that the infringement was to take place.\par The PRS abandoned its earlier claim that the defendants had ‘‘permitted’’ the infringement under s. 2(3) of the Copyright Act 1911 (UK). However, it claimed that the band were the defendants’ employees, and so the defendants were vicariously liable for the infringement.\par McCardie J examined the agreement and found that it gave ‘‘to the defendants the right of continuous, dominant, and detailed control on every point, including the nature of the music to be played’’.\footnote{ibid. at 771.} Hence the band members were employees of the defendant company, which was liable for the infringement." CASE "Cam and Sons Pty Ltd "Cam v. Sargent" v. Sargent" CITATION YEAR "(1940) 14 ALJ 162" 1940 COURT HC-4 FACTS RESULT (YYYNYNNNYNNNNNNNNN) Employee SUMMARY "Sargent was the master of a ship. He entered into an agreement with Cam and Sons that claimed that the ship was hired by Cam and Sons to Sargent and his fellow contractors (called ‘‘the partnership’’). However, it was doubtful whether that agreement actually deprived Cam and Sons of any control over the ship. The partnership was to use the ship only to carry coal from Swansea to Sydney. Cam and Sons were sole agents of the partnership for securing cargoes for the ship, and for collecting money due to the partnership. The partnership paid nothing for the ‘‘hire’’ of the ship, but received a specified sum for each return trip of a certain tonnage plus (in certain circumstances) 5% of the earnings, the balance of which was retained by Cam and Sons. Cam and Sons had to approve people employed by the partnership.\par Sargent claimed that he (and others in the partnership) were employed by Cam and Sons, and therefore came within the terms of an industrial award. Cam and Sons claimed that members of the partnership were independent contractors.\par 322 Appendix C: A complete example The High Court unanimously agreed with Sargent. Rich J came to the conclusion that the agreement was an attempt to evade the terms of the industrial award.\footnote{ibid. at 163.}" CASE "Federal Commissioner of Taxation (Australia) Pty Ltd" "FCT v. Thompson" v. J. Walter Thompson CITATION YEAR "(1944) 69 CLR 227" 1944 COURT HC FACTS RESULT (YNYNYYYUNNYNYNNNNN) Employee SUMMARY "the FCT claimed that payments made to radio artists by Thompson were ‘‘wages’’ within the meaning of the {\it Pay-roll Tax Assessment Act 1941\/} (Cth) and therefore taxable. The artists were selected by a producer and paid to appear in radio plays. They were paid a ‘‘fee’’ for each performance, but were paid nothing for attending (compulsory) rehearsals. Thompson claimed that the artists were presumed to know their work and ‘‘to render services in the same manner as a professional man, such as a surgeon or an architect, not being subject \dots\ to detailed control the manner in which those services are to be as to performed.’’\footnote{ibid. at 231.} Hence, Thompson claimed, they were independent contractors.\par Latham CJ held that the radio actors were employed ‘‘to co-operate with others in a team under the control of the producer to bring about a result, the details of which must in great measure be determined by the producer.’’\footnote{ibid. at 232.} Hence the artists were employed by Thompson; the fee they were paid was subject to payroll tax." CASE "Queensland Stations Pty Ltd v. Taxation" "Queensland Stations Federal Commissioner of v. FCT" CITATION YEAR "(1945) 70 CLR 539" 1945 COURT HC-3 FACTS RESULT (NYNYNNYYYNNNNNNNYN) Contractor SUMMARY "agreements were entered into between Queensland Stations and some drovers. The agreements stated that the drovers would ‘‘serve’’ Queensland Stations and take charge of a specified number of cattle, and deliver them to a specified place. The drovers were paid a specified rate per head of cattle successfully delivered. Each drover was responsible for hiring help, and paying for feed for the cattle. The drovers were to ‘‘obey and carry out all lawful instructions and to use the whole of [their] time, energy and ability in the � C.3 EMPLOYEE specification file 323 careful droving of the stock.’’\footnote{ibid. at 540} The FCT claimed that payments made to drovers were ‘‘wages’’ within the meaning of the {\it Pay-roll Tax Assessment Act 1941\/} (Cth), and that Queensland Stations was liable to payroll tax.\par The High Court held that the drovers were independent contractors, so the payments were not ‘‘wages’’. Rich J pointed out that drovers were traditionally free from the control of owners of cattle. ‘‘The obligation imposed on the drover to obey and carry out all lawful instructions is not a reservation of detailed control and possession having regard to the terms of the agreement as a whole.’’\footnote{ibid. at 549.}" CASE "Humberstone v. Northern Timber Mills" "Humberstone v. NTM" CITATION YEAR "(1949) 79 CLR 389" 1949 COURT HC-3 FACTS RESULT (NYNYNNNNYNNNNNNNNN) Contractor SUMMARY "Humberstone carried goods for NTM. He had originally held himself out as a carrier, prepared to carry for anyone, but for over twenty years he had carried goods solely for NTM (although he would, infrequently, carry back-loads for NTM’s customers). Humberstone owned the truck, and paid for petrol and repairs. He was paid weekly on a weight-mileage basis. He was a licenced carrier, and had his name printed on the side of his truck with the description ‘‘carrier.’’\par On the way back from a job, he had a puncture. He went home to change the wheel, but exerted himself so strenuously in trying to remove the tyre from the wheel that he became ill and later lapsed into a coma, from which he did not recover. Section 3 of the {\it Worker’s Compensation Act\/} 1928 (Vic) had been amended about a year before Humberstone’s death so as to include independent contractors in its definition of a ‘‘worker’’ covered by the Act. However, the High Court held that the amendment applied only to contracts entered into after it came into operation. Further, the Court decided that Humberstone was not an employee of NTM. Hence he was not a ‘‘worker’’ under the Act and his widow was not entitled to compensation under the Act." CASE "Stevenson Jordon and Harrison Ltd "Stevenson v. Macdonald (1)" v. Macdonald and Evans (1)" CITATION YEAR "[1952] 1 TLR 101" 1952 COURT CA FACTS RESULT (NYNNYNUNNNYUUUUUYN) Contractor 324 Appendix C: A complete example SUMMARY "Evans-Hemming was an accountant who had been employed (first as a servant, then as an executive officer) by Macdonald and Evans. Shortly after he left them, he wrote a textbook on business management and submitted the manuscript to Stevenson Jordon and Harrison (a firm of publishers). He died before the book was published. Macdonald and Evans claimed that the book was written while Evans-Hemming was their employee, and so they owned the copyright in the work under s. 5(1)(b) of the Copyright Act 1911 (UK).\par The book was divided into five sections. The first section consisted of the text of three public lectures that Evans-Hemming had given while employed by Macdonald and Evans. The Court of Appeal held that he had given these lectures as an independent contractor. As Denning LJ said, ‘‘under a contract of service, a man is employed as part of the business, and his work is done as an integral part of the business; whereas, under a contract for services, his work, although done for the business, is not integrated into it but is only accessory to it \dots\ The lectures were, in a sense, part of the services rendered by Mr Evans-Hemming for the benefit of the company. But they were in no sense part of his service. It follows that the copyright in the lectures was in Mr Evans-Hemming.’’\footnote{ibid. at 111}" CASE "Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (2)" "Stevenson v. Macdonald (2)" CITATION "[1952] 1 TLR 101" YEAR 1952 COURT CA FACTS (NYYNYNUNNNYUUUUUUN) RESULT Employee SUMMARY "Evans-Hemming was an accountant who had been employed (first as a servant, then as an executive officer) by Macdonald and Evans. Shortly after he left them, he wrote a textbook on business management and submitted the manuscript to Stevenson Jordon and Harrison (a firm of publishers). He died before the book was published. Macdonald and Evans claimed that the book was written while Evans-Hemming was their employee, and so they owned the copyright in the work under s. 5(1)(b) of the Copyright Act 1911 (UK).\par The book was divided into five sections. The second section was written in its final form while Evans-Hemming was employed by Macdonald and Evans. The Court of Appeal held that he wrote the second section as an employee, and hence the copyright in the second section was in Macdonald and Evans." � C.3 EMPLOYEE specification file 325 CASE "Zuijs "Zuijs v. v. Wirth Brothers Pty Ltd" Wirth Brothers" CITATION YEAR "(1955) 93 CLR 561" 1955 COURT HC-5 FACTS RESULT (NYYNYYYNNNYYNNYNNN) Employee SUMMARY "Zuijs was an acrobat who fell during a trapeze act at one of Wirth Brothers’ circuses. He sought compensation under the Worker’s Compensation Act 1926 (NSW), claiming to be an employee of Wirth Brothers. Wirth Brothers claimed that, because of the high degree of skill and personal judgment that he had to exercise in his work, Zuijs was an independent contractor and therefore not entitled to compensation.\par The High Court unanimously agreed with Zuijs. ‘‘Even if [Wirth Brothers] could not interfere in the actual technique of the acrobats and in the character of the act, no reason appears why the appellant should not be subject to his directions in all other respects \dots\ There are countless examples of highly specialized functions in modern life that must as a matter of practical necessity and sometimes even as a matter of law be performed on the responsibility of persons who possess particular knowledge and skill and who are accordingly qualified. But those engaged to perform the functions may nevertheless work under a contract of service.’’\footnote{ibid. at 571--2, per Dixon CJ, Williams, Webb and Taylor JJ.}" CASE "Ready Mixed Concrete (South East) Ltd and National Insurance" v. Minister of Pensions "Ready Mixed v. Minister" CITATION YEAR "[1968] 2 QB 497" 1967 COURT FACTS RESULT QB (NYYYNNNNYYNNNNNNNY) Contractor SUMMARY "Latimer worked for Ready Mixed as an ‘‘owner-driver’’. He was paid at mileage rates, and was obliged to buy the truck through a financial organization associated with Ready Mixed. The truck was painted in the company’s colours, and he had to wear a Ready Mixed uniform. Latimer was obliged to meet the costs of maintenance, repair and insurance of the truck (and the attached mixing unit, which belonged to Ready Mixed). The Minister determined that Latimer was employed under a contract of service and therefore an ‘‘employed person’’ under s. 1(2) of the National Insurance Act 1965 (UK), making Ready Mixed liable to make weekly contributions.\par 326 Appendix C: A complete example MacKenna J examined the contract and held that the rights it conferred, and the duties it imposed, between Latimer and Ready Mixed were not such as to make it a contract of service." CASE "Ferguson "Ferguson v. v. John Dawson & Partners (Contractors) Ltd" Dawson" CITATION YEAR COURT FACTS RESULT "[1976] 1 WLR 1213" 1976 CA (YNYNYNYNNNYYNNNNNY) Employee SUMMARY "Ferguson fell off a roof while removing some scaffolding boards. He claimed damages against John Dawson (the building contractors) for breach of statutory duty relying on the {\it Construction (Working Places) Regulations 1966\/} (UK). This duty would only be owed if Ferguson was an employee of John Dawson.\par Megaw and Browne LJJ held that, despite the fact that both parties labelled Ferguson a ‘‘self-employed labour only subcontractor’’,\footnote{ibid. at 1219, per Megaw LJ; at 1225, per Lawton LJ; at 1228, per Browne LJ.} the reality of the relationship between them was that of employer/employee." CASE "Massey "Massey v. v. Crown Life Insurance Co." Crown Life" CITATION YEAR COURT FACTS RESULT "[1978] ICR 590" 1978 CA (YYYNYYNYYNYYNNNYNY) Contractor SUMMARY "Massey was the manager of a branch of Crown Life. He had been an employee for two years, then he and Crown Life entered into a new agreement whereby Massey continued to perform the same duties as before, but was self-employed. This arrangement had tax advantages for Massey. After a further two years, Crown Life terminated the agreement and Massey sought compensation for unfair dismissal under the Trade Union and Labour Relations Act 1974 (UK). Compensation was only payable if Massey was employed under a contract of service.\par Lord Denning MR stated that ‘‘if the true relationship of the parties is that of master and servant under a contract of service, the parties cannot alter the truth of that relationship by putting a different label upon � C.3 EMPLOYEE specification file 327 it.’’\footnote{ibid. at 594.} However, he (and the rest of the Court of Appeal) held that the agreement was genuinely intended to establish Massey as being self-employed: an independent contractor." CASE "Australian Mutual Provident Society "AMP v. Chaplin" v. Chaplin" CITATION YEAR "(1978) 18 ALR 385" 1978 COURT PC FACTS RESULT (NYYYNNNYYYNNNNNNNY) Contractor SUMMARY "Chaplin was a representative of AMP. A clause of the agreement between them stated that the relationship was one of ‘‘principal and agent’’ and not one of ‘‘master and servant’’. Chaplin claimed that he was employed under a contract of service, and therefore a ‘‘worker’’ under the {\it Long Service Leave Act, 1967\/} (SA) and entitled to certain benefits.\par The Privy Council found that there was no reason to think that the clause was not a genuine statement of the parties’ intentions. Examining the agreement, their lordships concluded that it provided for a contract of agency. The fact that Chaplin was given the power of unlimited delegation of the whole performance of his work was ‘‘almost conclusive against the contract being a contract of service.’’\footnote{ibid. at 391.}" CASE "Price "Price v. v. Grant Industries Pty Ltd" Grant Industries" CITATION YEAR "(1978) 21 ALR 388" 1978 COURT FCA-3 FACTS RESULT (NYYYNNNNYNNNNNYNNN) Contractor SUMMARY "Grant Industries manufactured and sold wardrobes, which Price (and others) delivered and installed. Price and each of the other ‘‘contractors’’ (as Grant Industries called them) had to provide and maintain a suitable truck to deliver the wardrobes, and provide the tools required to install them. Price sought an order that a penalty be imposed on Grant Industries for breaching the Furnishing Trades (Consolidated) Award 1975 by not paying Price the appropriate rate of wages, and not giving him annual leave. The award only applied to ‘‘employees’’ of specified employers.\par The Federal Court examined the facts, and the provisions of the agreement, and held that Price was an independent contractor and, therefore, not subject to the award." 328 Appendix C: A complete example CASE "Australian Timber Workers Union v. Monaro Sawmills Pty Ltd" "ATWU v. Monaro Sawmills" CITATION "(1980) 29 ALR 322" YEAR 1980 COURT FCA-3 FACTS (YNYYNYYNUNNNNNNNNN) RESULT Employee SUMMARY "Wales was a tree feller who cut timber exclusively for Monaro Sawmills. He performed his work in an area allotted to him by Monaro Sawmills. He, and other fellers, were paid by the amount of millable wood they cut. Wales provided his own tools and transport, but was (with the other fellers) covered by Monaro Sawmill’s workers’ compensation policy.\par The union sought an order that a penalty be imposed on Monaro Sawmills for breaching the Timber Industries Consolidated Award 1974 by failing to pay Wales money in lieu of annual leave. Monaro Sawmills claimed that Wales was an independent contractor, and so was not subject to the award.\par Sweeney and Evatt JJ examined the circumstances of Wales’s employment and held that those circumstances clearly pointed to the existence of a relationship of employer and employee. They could not see ‘‘any sense in which it could be said that Wales was conducting some sort of business of his own.’’\footnote{ibid. at 329.}" IDEAL FACTS (YNYNYYYNNYYYNYYYYU) RESULT Employee IDEAL FACTS (NYNYNUNYYNNNYNNNUY) RESULT Contractor C.4 Dump file for EMPLOYEE specification (LATEX input) % Dump file % Produced by SHYSTER version 1.0 % Copyright James Popple 1993 % This is not a stand-alone LaTeX file. % Include it in a LaTeX document using the \input command. % Use LaTeX version 2.09 <25 March 1992> and TeX version 3.141. � C.4 Dump file for EMPLOYEE specification (LATEX input) 329 \subsection*{Hierarchy} \begin{small} \begin{trivlist}\item[] \begin{tabular}{|r|l|}\hline \multicolumn{1}{|c|}{$c$}&\multicolumn{1}{c|}{\it Court\/}\\\h � line\hline⇐ 1&five judges of the High Court of Australia\\ 2&four judges of the High Court of Australia\\ 3&three judges of the High Court of Australia\\ 4&a single judge of the High Court of Australia\\ 5&three judges of the Federal Court of Australia\\ 6&the Judicial Committee of the Privy Council\\ 7&the English Court of Appeal\\ 8&the King’s Bench Division of the English High Court\\ &the Queen’s Bench Division of the English High Court\\\hline \end{tabular} \end{trivlist} \end{small} \subsection*{Employee area} \begin{small} \begin{tabular}{*{2}{|c}*{17}{@{\hspace{0.4em}}c}|r|c|}\hline &\multicolumn{18}{|c|}{\it Attributes\/}&&\\ \smash{\raisebox{0.6\ht\strutbox}{\it Case\/}}&$A_{1}$&$A_{2}$&$A_ � {3}$&$A_{4}$&$A_{5}$&$A_{6}$&$A_{7}$&$A_{8}$&$A_{9}$&$A_{10}$&⇐ � $A_{11}$&$A_{12}$&$A_{13}$&$A_{14}$&$A_{15}$&$A_{16}$&$A_{17}$⇐ � &$A_{18}$&\multicolumn{1}{c|}{\smash{\raisebox{0.6\ht\strutbox⇐ � }{$c$}}}&\smash{\raisebox{0.6\ht\strutbox}{\it Result\/}}\\\hl⇐ � ine\hline⇐ $C_{1}$&$\times$&$\bullet$&$\bullet$&$\times$&$\bullet$&$\bullet$& � $\bullet$&$\times$&$\times$&$\times$&$\bullet$&$\bullet$&$\tim⇐ � es$&$\times$&$\bullet$&$\times$&$\times$&$\times$&1&\\⇐ $C_{14}$&$\times$&$\bullet$&$\bullet$&$\bullet$&$\times$&$\times$& � $\times$&$\times$&$\bullet$&$\bullet$&$\times$&$\times$&$\time⇐ � s$&$\times$&$\times$&$\times$&$\times$&$\bullet$&8&\\\cline{2­⇐ � 20}⇐ $I_{\mbox{\scriptsize\sf Contractor}}$&$\times$&$\bullet$&$\times$ � &$\bullet$&$\times$&&$\times$&$\bullet$&$\bullet$&$\times$&$\t⇐ � imes$&$\times$&$\bullet$&$\times$&$\times$&$\times$&&$\bullet$⇐ � &&\\\hline⇐ \end{tabular} \end{small} \subsubsection*{Opening} \begin{list}{}{\leftmargin=0mm}\item[] The law distinguishes between a contract of service (between employer and employee) and a contract for services (between principal and independent contractor). This distinction affects the terms that ... mailto:\begin{tabular}{*{2}{|c}*{17}{@{\hspace{0.4em}}c}|r|c|}\hline 330 Appendix C: A complete example will be implied in the absence of an express agreement, the liability of the employer to third parties, the applicability of industrial awards, the applicability of statutes which may affect workers’ compensation, occupational health and safety, long-service leave, fringe benefits tax, etc.\par The terms ‘‘employer’’ and ‘‘worker’’ are used here to mean ‘‘employer’’ and ‘‘employee’’ (in the case of a contract of service) or ‘‘principal’’ and ‘‘independent contractor’’ (in the case of a contract for services). \end{list} \subsubsection*{Results} \begin{description} \item[\rm{\sf Employee}:] the worker is an employee. \item[\rm{\sf Contractor}:] the worker is an independent contractor. \end{description} \subsubsection*{Attributes} \begin{description} \item[\rm$A_{1}$:] Did the employer direct not only what work was to be done, but also the manner in which it was to be done? \begin{description} \item[\sc yes:] the employer directed the manner in which the work was to be done. $\Rightarrow$ {\sf Employee} \item[\sc no:] ... \end{description} If the employer had a right of control over how the worker did the work then the employer had the power to direct not only what work was to be done, but also the manner in which it was to be done. \item[\rm$A_{2}$:] Was the worker allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand? ... \end{description} � C.5 Dump file for EMPLOYEE specification (LATEX output) 331 \subsubsection*{Cases in which the worker is an employee} \begin{description} \item[\rm$C_{1}$:]\frenchspacing {\it Zuijs v. Wirth Brothers Pty Ltd\/} \nonfrenchspacing (1955) 93 CLR 561 (‘‘\frenchspacing {\it Zuijs v. Wirth Brothers\/}\nonfrenchspacing’’) \begin{description} \item[\rm$A_{1}$:] the employer did not direct the manner in which the work was to be done. \item[\rm$A_{2}$:] ... \end{description} In \frenchspacing {\it Zuijs v. Wirth Brothers Pty Ltd}\nonfrenchspacing,% \footnote{(1955) 93 CLR 561.} a 1955 decision of five judges of the High Court of Australia, Zuijs was an acrobat who fell during a trapeze act at one of Wirth Brothers’ circuses. He sought compensation under the Worker’s Compensation Act 1926 (NSW), claiming to be an employee of Wirth ... \item[$I_{\mbox{\scriptsize\sf Employee}}$] (the ideal case in which the worker is an employee): ... \end{description} \subsubsection*{Cases in which the worker is an independent contractor} ... C.5 Dump file for EMPLOYEE specification (LATEX output) The dump file for the Employee specification, as output by LaTEX, appears with those of the other specifications in appendix A (§A.4). � \ 332 Appendix C: A complete example C.6 Probabilities file for EMPLOYEE specification (LATEX input) % Probabilities file % Produced by SHYSTER version 1.0 % Copyright James Popple 1993 % This is not a stand-alone LaTeX file. % Include it in a LaTeX document using the \input command. % Use LaTeX version 2.09 <25 March 1992> and TeX version 3.141. \subsection*{Employee area} \begin{small} \def\arraystretch{0} \begin{tabular}{*{17}{|c}|@{}p{\doublerulesep}@{}|c|}\cline{1-17} \smash{\raisebox{0.6\ht\strutbox}{$A_{2}$}}&\smash{\raisebox{0.6\h � t\strutbox}{$A_{3}$}}&\smash{\raisebox{0.6\ht\strutbox}{$A_{4}⇐ � $}}&\smash{\raisebox{0.6\ht\strutbox}{$A_{5}$}}&\smash{\raiseb⇐ � ox{0.6\ht\strutbox}{$A_{6}$}}&\smash{\raisebox{0.6\ht\strutbox⇐ � }{$A_{7}$}}&\smash{\raisebox{0.6\ht\strutbox}{$A_{8}$}}&\smash⇐ � {\raisebox{0.6\ht\strutbox}{$A_{9}$}}&\smash{\raisebox{0.6\ht\⇐ � strutbox}{$A_{10}$}}&\smash{\raisebox{0.6\ht\strutbox}{$A_{11}⇐ � $}}&\smash{\raisebox{0.6\ht\strutbox}{$A_{12}$}}&\smash{\raise⇐ � box{0.6\ht\strutbox}{$A_{13}$}}&\smash{\raisebox{0.6\ht\strutb⇐ � ox}{$A_{14}$}}&\smash{\raisebox{0.6\ht\strutbox}{$A_{15}$}}&\s⇐ � mash{\raisebox{0.6\ht\strutbox}{$A_{16}$}}&\smash{\raisebox{0.⇐ � 6\ht\strutbox}{$A_{17}$}}&\smash{\raisebox{0.6\ht\strutbox}{$A⇐ � _{18}$}}&\multicolumn{2}{c}{\raisebox{\ht\strutbox}{\strut}}\\⇐ � \cline{1-17}⇐ \multicolumn{19}{c}{\rule{0mm}{\doublerulesep}}\\\cline{1-17}\clin � e{19-19}⇐ 0.05&1.00&0.30&0.95&1.00&0.96&0.69&0.41&0.62&0.95&0.97&1.00&1.00&0 � .23&1.00&0.27&0.83&&\\⇐ 1.00&0.15&0.95&0.30&0.06&0.28&0.80&0.91&0.85&0.30&0.27&0.50&0.50&1 � .00&0.50&1.00&0.59&&\smash{\raisebox{0.6\ht\strutbox}{$A_{1}$}⇐ � }\\\cline{1-17}\cline{19-19}⇐ \multicolumn{1}{c|}{}&0.45&0.90&0.50&0.27&0.09&1.00&1.00&1.00&0.50 � &0.75&0.25&1.00&1.00&1.00&1.00&0.67&&\\⇐ \multicolumn{1}{c|}{}&1.00&0.50&0.90&0.97&1.00&0.58&0.19&0.45&0.90 � &0.76&1.00&0.75&0.55&0.75&0.58&0.82&&\smash{\raisebox{0.6\ht\s⇐ � trutbox}{$A_{2}$}}\\\cline{2-17}\cline{19-19}⇐ \multicolumn{2}{c|}{}&0.50&0.90&1.00&0.77&0.58&0.56&1.00&0.90&1.00 � &1.00&1.00&1.00&1.00&0.04\rlap{\makebox[\tabcolsep]{$\bullet$}⇐ � }&1.00&&\\⇐ \multicolumn{2}{c|}{}&0.90&0.50&0.23&0.77&0.89&0.88&0.45&0.50&0.42 � &0.83&0.83&0.68&0.83&1.00&0.33&&\smash{\raisebox{0.6\ht\strutb⇐ � ox}{$A_{3}$}}\\\cline{3-17}\cline{19-19}⇐ \multicolumn{3}{c|}{}&0.00\rlap{\makebox[\tabcolsep]{\rule[0.25ex] � {0.35em}{0.35em}}}&0.50&0.50&0.88&1.00&1.00&0.01\rlap{\makebox⇐ � [\tabcolsep]{$\bullet$}}&0.15&0.42&1.00&0.68&0.42&0.73&0.72&&\⇐ ⇐ � C.7 Probabilities file for EMPLOYEE specification (LATEX output) 333 \multicolumn{3}{c|}{}&1.00&0.87&0.88&0.56&0.08&0.10&1.00&0.99&1.00 � &0.58&0.85&1.00&0.81&0.72&&\smash{\raisebox{0.6\ht\strutbox}{$⇐ � A_{4}$}}\\\cline{4-17}\cline{19-19}⇐ ... \multicolumn{15}{c|}{}&0.92&1.00&&\\ \multicolumn{15}{c|}{}&1.00&0.33&&\smash{\raisebox{0.6\ht\strutbox � }{$A_{16}$}}\\\cline{16-17}\cline{19-19}⇐ \multicolumn{16}{c|}{}&0.46&&\\ \multicolumn{16}{c|}{}&1.00&&\smash{\raisebox{0.6\ht\strutbox}{$A_ � {17}$}}\\\cline{17-17}\cline{19-19}⇐ \end{tabular} \end{small} C.7 Probabilities file for EMPLOYEE specification (LATEX output) Employee area A2 A3 A4 A5 0.05 1.00 0.30 0.95 1.00 0.15 0.95 0.30 0.45 0.90 0.50 1.00 0.50 0.90 0.50 0.90 0.90 0.50 0.00 1.00 A6 1.00 0.06 0.27 0.97 1.00 0.23 0.50 0.87 0.87 0.50 A7 0.96 0.28 0.09 1.00 0.77 0.77 0.50 0.88 0.88 0.50 0.99 0.12 A8 0.69 0.80 1.00 0.58 0.58 0.89 0.88 0.56 0.56 0.88 0.80 0.71 0.58 0.88 A9 0.41 0.91 1.00 0.19 0.56 0.88 1.00 0.08 0.08 1.00 0.22 0.98 0.02• 1.00 1.00 0.16 A10 0.62 0.85 1.00 0.45 1.00 0.45 1.00 0.10 0.10 1.00 0.73 0.77 0.50 0.91 0.89 0.58 0.88 0.56 A11 0.95 0.30 0.50 0.90 0.90 0.50 0.01• 1.00 1.00 0.01• 0.99 0.13 0.99 0.12 0.56 0.88 0.00• 1.00 0.50 0.90 A12 0.97 0.27 0.75 0.76 1.00 0.42 0.15 0.99 0.99 0.15 0.99 0.15 0.97 0.27 0.72 0.79 0.09 1.00 0.76 0.75 1.00 0.01• A13 1.00 0.50 0.25 1.00 1.00 0.83 0.42 1.00 1.00 0.42 1.00 0.42 1.00 0.50 1.00 1.00 0.36 1.00 0.75 1.00 1.00 0.42 0.67 1.00 A14 1.00 0.50 1.00 0.75 1.00 0.83 1.00 0.58 0.58 1.00 1.00 0.42 1.00 0.50 0.73 1.00 0.36 1.00 1.00 0.25 1.00 0.42 1.00 0.33 0.92 1.00 A15 0.23 1.00 1.00 0.55 1.00 0.68 0.68 0.85 0.85 0.68 0.85 0.68 0.77 0.77 0.51 1.00 0.62 0.89 0.55 1.00 0.85 0.68 0.91 0.58 0.83 1.00 0.83 1.00 A16 A17 A18 1.00 0.27 0.83 A10.50 1.00 0.59 1.00 1.00 0.67 A20.75 0.58 0.82 1.00 0.04 1.00• A30.83 1.00 0.33 0.42 0.73 0.72 A41.00 0.81 0.72 1.00 0.81 0.72 A50.42 0.73 0.72 1.00 0.36 0.55 A60.42 1.00 0.87 0.50 1.00 0.27 A71.00 0.50 0.97 1.00 0.95 0.99 A80.27 0.45 0.20 1.00 0.68 0.95 A90.64 0.85 0.34 0.75 0.58 0.99 A101.00 1.00 0.18 1.00 0.81 0.72 A110.42 0.73 0.72 1.00 0.67 0.93 A120.33 1.00 0.41 0.92 0.92 0.67 A131.00 1.00 1.00 0.92 0.92 0.67 A141.00 1.00 1.00 0.83 0.83 0.42 A151.00 1.00 1.00 0.92 1.00 A161.00 0.33 0.46 A171.00 334 Appendix C: A complete example C.8 Weights file for EMPLOYEE specification (LATEX input) % Weights file % Produced by SHYSTER version 1.0 % Copyright James Popple 1993 % This is not a stand-alone LaTeX file. % Include it in a LaTeX document using the \input command. % Use LaTeX version 2.09 <25 March 1992> and TeX version 3.141. \subsection*{Employee area} \begin{small} \begin{tabular}{|c|*{3}{c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}} � r|}}\hline⇐ &\multicolumn{3}{c|}{\sf Employee}&\multicolumn{3}{c|}{\sf Contrac � tor}&&&\\⇐ \smash{\raisebox{0.6\ht\strutbox}{\it Attr.}}&$\mu$&$\sigma�2$&\mu � lticolumn{1}{c|}{$w$}&$\mu$&$\sigma�2$&\multicolumn{1}{c|}{$w$⇐ � }&\smash{\raisebox{0.6\ht\strutbox}{$\mu$}}&\smash{\raisebox{0 � ⇐ .6\ht\strutbox}{$\sigma�2$}}&\multicolumn{1}{c|}{\smash{\raise⇐ � box{0.6\ht\strutbox}{$w$}}}\\\hline\hline⇐ $A_{1}$&0.71&0.20&4.90&0.14&0.12&8.17&0.43&0.24&4.08\\ $A_{2}$&0.57&0.24&4.08&1.00&0.00&\multicolumn{1}{c|}{$\infty$}&0.7 � 9&0.17&5.94\\⇐ $A_{3}$&1.00&0.00&\multicolumn{1}{c|}{$\infty$}&0.57&0.24&4.08&0.7 � 9&0.17&5.94\\⇐ $A_{4}$&0.29&0.20&4.90&0.71&0.20&4.90&0.50&0.25&4.00\\ $A_{5}$&0.71&0.20&4.90&0.29&0.20&4.90&0.50&0.25&4.00\\ $A_{6}$&0.57&0.24&4.08&0.14&0.12&8.17&0.36&0.23&4.36\\ $A_{7}$&0.83&0.14&7.20&0.17&0.14&7.20&0.50&0.25&4.00\\ $A_{8}$&0.00&0.00&\multicolumn{1}{c|}{$\infty$}&0.43&0.24&4.08&0.2 � 3&0.18&5.63\\⇐ $A_{9}$&0.17&0.14&7.20&0.86&0.12&8.17&0.54&0.25&4.02\\ $A_{10}$&0.14&0.12&8.17&0.29&0.20&4.90&0.21&0.17&5.94\\ $A_{11}$&0.71&0.20&4.90&0.29&0.20&4.90&0.50&0.25&4.00\\ $A_{12}$&0.50&0.25&4.00&0.17&0.14&7.20&0.33&0.22&4.50\\ $A_{13}$&0.17&0.14&7.20&0.00&0.00&\multicolumn{1}{c|}{$\infty$}&0. � 08&0.08&13.09\\⇐ $A_{14}$&0.17&0.14&7.20&0.00&0.00&\multicolumn{1}{c|}{$\infty$}&0. � 08&0.08&13.09\\⇐ $A_{15}$&0.17&0.14&7.20&0.17&0.14&7.20&0.17&0.14&7.20\\ $A_{16}$&0.00&0.00&\multicolumn{1}{c|}{$\infty$}&0.17&0.14&7.20&0. � 08&0.08&13.09\\⇐ $A_{17}$&0.00&0.00&\multicolumn{1}{c|}{$\infty$}&0.29&0.20&4.90&0. � 15&0.13&7.68\\⇐ $A_{18}$&0.14&0.12&8.17&0.43&0.24&4.08&0.29&0.20&4.90\\\hline \end{tabular} \end{small} � � � � � � � � C.10 Distances file for BWIU v. Odco (LATEX input) 335 C.9 Weights file for EMPLOYEE specification (LATEX output) Employee area Employee Contractor Attr. µ �2 w µ �2 w µ �2 w 0.71 0.20 4.90 0.14 0.12 8.17 0.43 0.24 4.08A1 0.57 0.24 4.08 1.00 0.00 0.79 0.17 5.94A2 1.00 0.00 0.57 0.24 4.08 0.79 0.17 5.94A3 0.29 0.20 4.90 0.71 0.20 4.90 0.50 0.25 4.00A4 0.71 0.20 4.90 0.29 0.20 4.90 0.50 0.25 4.00A5 0.57 0.24 4.08 0.14 0.12 8.17 0.36 0.23 4.36A6 0.83 0.14 7.20 0.17 0.14 7.20 0.50 0.25 4.00A7 0.00 0.00 0.43 0.24 4.08 0.23 0.18 5.63A8 0.17 0.14 7.20 0.86 0.12 8.17 0.54 0.25 4.02A9 0.14 0.12 8.17 0.29 0.20 4.90 0.21 0.17 5.94A10 0.71 0.20 4.90 0.29 0.20 4.90 0.50 0.25 4.00A11 0.50 0.25 4.00 0.17 0.14 7.20 0.33 0.22 4.50A12 0.17 0.14 7.20 0.00 0.00 0.08 0.08 13.09 A13 0.17 0.14 7.20 0.00 0.00 0.08 0.08 13.09 A14 0.17 0.14 7.20 0.17 0.14 7.20 0.17 0.14 7.20A15 0.00 0.00 0.17 0.14 7.20 0.08 0.08 13.09 A16 0.00 0.00 0.29 0.20 4.90 0.15 0.13 7.68A17 0.14 0.12 8.17 0.43 0.24 4.08 0.29 0.20 4.90A18 C.10 Distances file for Building Workers’ Industrial Union of Australia v. Odco Pty Ltd (LATEX input) % Distances file % Produced by SHYSTER version 1.0 % Copyright James Popple 1993 % This is not a stand-alone LaTeX file. % Include it in a LaTeX document using the \input command. % Use LaTeX version 2.09 <25 March 1992> and TeX version 3.141. \subsection*{Employee area} \subsubsection*{Instant case} \begin{small} \begin{tabular}{*{2}{|c}*{17}{@{\hspace{0.4em}}c}|r|r@{\hspace{\tabcol � sep}}r|r|c@{\hspace{\tabcolsep}}c|r@{\hspace{\tabcolsep}}r|c|}\hli⇐ � ne⇐ mailto:\begin{tabular}{*{2}{|c}*{17}{@{\hspace{0.4em}}c}|r|r@{\hspace{\tabcol 336 Appendix C: A complete example &\multicolumn{18}{|c|}{\it Attributes\/}&&&&&&&&&\\ \smash{\raisebox{0.6\ht\strutbox}{\it Case\/}}&$A_{1}$&$A_{2}$&$A_ � {3}$&$A_{4}$&$A_{5}$&$A_{6}$&$A_{7}$&$A_{8}$&$A_{9}$&$A_{10}$&⇐ � $A_{11}$&$A_{12}$&$A_{13}$&$A_{14}$&$A_{15}$&$A_{16}$&$A_{17}$⇐ � &$A_{18}$&\multicolumn{1}{c|}{\smash{\raisebox{0.6\ht\strutbox⇐ � }{$c$}}}&\multicolumn{1}{c}{\smash{\raisebox{0.6\ht\strutbox}{⇐ � $d_{\rm K}$}}}&\multicolumn{1}{c|}{\smash{\raisebox{0.6\ht\str⇐ � utbox}{$d_{\rm U}$}}}&\multicolumn{1}{c|}{\smash{\raisebox{0.6⇐ � \ht\strutbox}{$\Delta$}}}&\smash{\raisebox{0.6\ht\strutbox}{$S⇐ � $}}&\smash{\raisebox{0.6\ht\strutbox}{$S’$}}&\multicolumn{1}{c⇐ � }{\smash{\raisebox{0.6\ht\strutbox}{$r$}}}&\multicolumn{1}{c|}⇐ � {\smash{\raisebox{0.6\ht\strutbox}{$r’$}}}&\smash{\raisebox{0.⇐ � 6\ht\strutbox}{\it Result\/}}\\\hline\hline⇐ $C_{\rm Instant}$&$\times$&$\bullet$&$\times$&$\bullet$&$\times$&$ � \times$&$\times$&&$\times$&$\times$&$\bullet$&&$\times$&$\time⇐ � s$&$\times$&$\times$&$\times$&$\bullet$&\multicolumn{9}{c|}{}\⇐ � \\hline\hline⇐ $C_{1}$&$\times$&$\bullet$&$\bullet$&$\times$&$\bullet$&$\bullet$& � $\bullet$&$\times$&$\times$&$\times$&$\bullet$&$\bullet$&$\tim⇐ � es$&$\times$&$\bullet$&$\times$&$\times$&$\times$&1&34.39&10.1⇐ � 3&7&0.44&0.33&0.07&0.11&\\⇐ ... $C_{12}$&$\bullet$&$\bullet$&$\bullet$&$\times$&$\bullet$&$\bullet � $&$\times$&$\bullet$&$\bullet$&$\times$&$\bullet$&$\bullet$&$\⇐ � times$&$\times$&$\times$&$\bullet$&$\times$&$\bullet$&7&39.49&⇐ � 10.13&7&0.44&0.37&0.22&0.14&$\stackrel{\scriptscriptstyle I~}{⇐ � \Rightarrow}$\,$2\infty$+52.63\\⇐ ... $I_{\mbox{\scriptsize\sf Contractor}}$&$\times$&$\bullet$&$\times$ � &$\bullet$&$\times$&&$\times$&$\bullet$&$\bullet$&$\times$&$\t⇐ � imes$&$\times$&$\bullet$&$\times$&$\times$&$\times$&&$\bullet$⇐ � &&21.11&22.17&3&0.21&0.23&0.52&0.29&\\⇐ $\mu_{\mbox{\scriptsize\sf Contractor}}$&$\times$&$\bullet$&$\bull � et$&$\bullet$&$\times$&$\times$&$\times$&$\times$&$\bullet$&$\⇐ � times$&$\times$&$\times$&$\times$&$\times$&$\times$&$\times$&$⇐ � \times$&$\times$&&18.86&10.13&4&0.25&0.18&0.53&0.55&\\\hline⇐ \end{tabular} \end{small} \subsubsection*{Instantiation 1} ... C.11 Distances file for Building Workers’ Industrial Union of Australia v. Odco Pty Ltd (LATEX output) � • • • � � � C a se C In s t a n t C 1 C 2 C 3 C 4 A tt r ib u te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 × • × • × × × × × • × × × × × • × • • × • • • × × × • • × × • × × × 1 • • • × • × × × • × × × × × × × × × 2 • × • × • • • × × • × • × × × × × 4 • × • • × • • × × × × × × × × × × 5 d K d U 3 4 .3 9 1 0 .1 3 3 0 .9 5 1 0 .1 3 5 0 .3 1 1 0 .1 3 3 3 .2 2 1 4 .1 6 7 7 9 7 S S � 0 .4 4 0 .3 3 0 .4 4 0 .2 9 0 .5 6 0 .4 8 0 .4 7 0 .3 3 I r r � R e su lt 0 .0 7 0 .1 1 − 0. 0 8 0 .0 9 − 0. 2 2 − 0. 2 6 E m p lo y e e − 0. 1 1 − 0. 1 8 � 1 9 .3 0 • × • × • × • × × × × × × × × µ C 5 7 2 7 .9 6 1 0 .1 3 6 0 .3 8 0 .2 7 0 .1 5 0 .1 1 1 9 .3 0 C 6 × • • × • × × × × • × 7 1 8 .8 4 6 8 .2 9 4 0 .4 0 0 .4 0 0 .1 7 0 .2 3 1 2 .1 0 C 7 I Em p lo ye e µ Em p lo ye e C 8 C 9 C 1 0 C 1 1 • • • • × • • × × • • • × • × × × × 8 • × • × • • • × × • • • × • • • • • • • × • • • × × × • • × × × × × × × • × • × × × × • × × × × × × × × × 3 × • × • × × • • • × × × × × × × • × 3 × • • • × × × × • × × × × × • × × × 5 × • • • × × × • • • × × × × × × × • 6 4 2 .3 1 1 0 .1 3 7 9 .3 2 1 5 .0 3 3 1 .2 8 1 0 .1 3 1 2 .9 2 1 0 .1 3 2 4 .6 1 1 0 .1 3 2 6 .0 6 1 0 .1 3 1 9 .9 0 1 0 .1 3 7 1 2 7 3 5 5 4 0 .4 4 0 .4 0 0 .8 0 0 .7 9 0 .4 4 0 .3 0 0 .1 9 0 .1 2 0 .3 1 0 .2 3 0 .3 1 0 .2 5 0 .2 5 0 .1 9 µ I 0 .2 2 0 .0 6 − 0. 4 5 − 0. 4 5 0 .0 6 0 .0 2 0 .4 6 0 .5 6 0 .2 3 0 .2 5 0 .2 3 0 .2 2 C o n tr a c to r 0 .4 5 0 .4 6 + 5 5 .9 0 � 2 � � 2 � � C.11 Distances file for BWIU v. Odco (LATEX output) 337 Employee area Instant case C 1 2 7 3 9 .4 9 1 0 .1 3 7 0 .4 4 0 .3 7 0 .2 2 0 .1 4 • • • × • • × • • × • • × × × • × • + 5 2 .6 3 C 1 3 × • × × • × × × × • • × 7 2 0 .5 8 6 0 .6 1 4 0 .3 6 0 .3 8 0 .2 1 0 .2 0 3 3 .3 3 C 1 4 8 1 9 .9 0 1 0 .1 3 4 0 .2 5 0 .1 9 0 .4 5 0 .4 6 × • • • × × × × • • × × × × × × × • I Co n tr a ct o r × • × • × × • • × × × • × × × • 2 1 .1 1 2 2 .1 7 3 0 .2 1 0 .2 3 0 .5 2 0 .2 9 µ Co n tr a ct o r 1 8 .8 6 1 0 .1 3 4 0 .2 5 0 .1 8 0 .5 3 0 .5 5 × • • • × × × × • × × × × × × × × × � • • • � � � 338 C a se C In s t -1 C 1 C 2 C 3 C 4 A tt r ib u te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 × • × • × × × • × × • • × × × × × • × • • × • • • × × × • • × × • × × × 1 • • • × • × × × • × × × × × × × × × 2 • × • × • • • × × • × • × × × × × 4 • × • • × • • × × × × × × × × × × 5 d K d U 4 0 .0 3 – 4 1 .0 8 – 5 4 .8 1 5 .6 3 4 3 .3 5 4 .0 2 8 9 1 0 9 S S � 0 .4 4 0 .3 5 0 .5 0 0 .3 6 0 .5 9 0 .5 0 0 .5 3 0 .3 9 r r � R e su lt I 0 .0 8 0 .0 7 − 0. 1 8 − 0. 0 4 − 0. 2 8 − 0. 3 0 E m p lo y e e − 0. 2 1 − 0. 2 7 � 2 3 .3 0 • × • × • × • × × × × × × × × µ C 5 7 3 3 .6 0 – 7 0 .3 9 0 .2 9 0 .1 6 0 .1 0 2 3 .3 0 C 6 × • • × • × × × × • × 7 2 4 .4 7 6 2 .6 5 5 0 .4 5 0 .4 6 0 .0 7 0 .1 0 1 6 .1 0 C 7 I Em p lo ye e µ Em p lo ye e C 8 C 9 C 1 0 C 1 1 • • • • × • • × × • • • × • × × × × 8 • × • × • • • × × • • • × • • • • • • • × • • • × × × • • × × × × × × × • × • × × × × • × × × × × × × × × 3 × • × • × × • • • × × × × × × × • × 3 × • • • × × × × • × × × × × • × × × 5 × • • • × × × • • • × × × × × × × • 6 4 7 .9 4 – 8 4 .9 5 4 .9 0 3 6 .9 1 – 2 3 .0 6 – 2 9 .1 1 – 3 6 .2 0 – 2 4 .4 0 – 8 1 3 8 5 6 7 5 0 .4 4 0 .4 2 0 .7 6 0 .7 7 0 .4 4 0 .3 2 0 .2 8 0 .2 0 0 .3 3 0 .2 5 0 .3 9 0 .3 1 0 .2 8 0 .2 1 µ I 0 .1 6 0 .0 2 − − 0. 4 3 − 0. 4 9 − 0. 0 3 − 0. 1 1 0 .3 2 0 .3 9 0 .2 5 0 .3 0 0 .0 9 0 .0 7 C o n tr a c to r 0 .4 0 0 .4 5 + 5 9 .9 8 � 2 � � 2 � Appendix C: A complete example Instantiation 1 C 1 2 7 3 9 .4 9 – 7 0 .3 9 0 .3 4 0 .3 2 0 .2 2 • • • × • • × • • × • • × × × • × • + 5 6 .7 2 C 1 3 × • × × • × × × × • • × 7 2 6 .2 2 5 4 .9 7 5 0 .4 2 0 .4 3 0 .1 2 0 .0 9 3 3 .3 3 C 1 4 8 3 0 .0 4 – 6 0 .3 3 0 .2 6 0 .2 5 0 .2 5 × • • • × × × × • • × × × × × × × • I Co n tr a ct o r × • × • × × • • × × × • × × × • 2 5 .6 1 1 2 .0 4 4 0 .2 5 0 .2 5 0 .4 7 0 .2 8 µ Co n tr a ct o r 2 9 .0 0 – 6 0 .3 3 0 .2 5 0 .4 3 0 .4 4 × • • • × × × × • × × × × × × × × × � � � C a se C In s t -2 C 1 C 2 C 3 C 4 A tt r ib u te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 × • × • × × × • × × • × × × × × × • × • • × • • • × × × • • × × • × × × 1 • • • × • × × × • × × × × × × × × × 2 • × • × • • • × × • × • × × × × × 4 • × • • × • • × × × × × × × × × × 5 d K d U 4 4 .5 3 – 3 6 .5 8 – 5 0 .3 1 5 .6 3 3 8 .8 5 4 .0 2 9 8 9 8 S S � 0 .5 0 0 .3 9 0 .4 4 0 .3 2 0 .5 3 0 .4 6 0 .4 7 0 .3 5 I r r � R e su lt − 0. 0 6 − 0. 0 3 − 0. 1 1 0 .0 3 − 0. 1 8 − 0. 2 3 E m p lo y e e − 0. 1 3 − 0. 2 0 � 1 9 .3 0 • × • × • × • × × × × × × × × µ C 5 7 3 8 .1 0 – 8 0 .4 4 0 .3 3 0 .0 1 0 .0 3 1 9 .3 0 • • • − � C 6 × • • × • × × × × • × 7 2 4 .4 7 6 2 .6 5 5 0 .4 5 0 .4 6 0 .0 7 0 .1 0 1 2 .1 0 C 7 I Em p lo ye e µ Em p lo ye e C 8 C 9 C 1 0 C 1 1 • • • • × • • × × • • • × • × × × × 8 • × • × • • • × × • • • × • • • • • • • × • • • × × × • • × × × × × × × • × • × × × × • × × × × × × × × × 3 × • × • × × • • • × × × × × × × • × 3 × • • • × × × × • × × × × × • × × × 5 × • • • × × × • • • × × × × × × × • 6 5 2 .4 4 – 8 9 .4 5 4 .9 0 4 1 .4 1 – 1 8 .5 6 – 2 4 .6 1 – 3 1 .7 0 – 1 9 .9 0 – 9 1 4 9 4 5 6 4 0 .5 0 0 .4 5 0 .8 2 0 .8 1 0 .5 0 0 .3 6 0 .2 2 0 .1 6 0 .2 8 0 .2 1 0 .3 3 0 .2 7 0 .2 2 0 .1 7 µ I 0 .0 6 0 .0 6 − − 0. 5 5 − 0. 5 1 − 0. 0 8 − 0. 1 3 0 .3 9 0 .4 5 0 .3 5 0 .3 9 0 .1 7 0 .1 4 C o n tr a c to r 0 .5 2 0 .5 5 + 6 7 .1 8 � 2 � � 2 � � C.11 Distances file for BWIU v. Odco (LATEX output) 339 Instantiation 2 C 1 2 7 4 3 .9 9 – 8 0 .4 4 0 .3 8 0 .2 4 0 .1 9 • • • × • • × • • × • • × × × • × • + 6 3 .9 2 C 1 3 × • × × • × × × × • • × 7 2 6 .2 2 5 4 .9 7 5 0 .4 2 0 .4 3 0 .1 2 0 .0 9 4 0 .5 3 C 1 4 8 2 5 .5 4 – 5 0 .2 8 0 .2 2 0 .3 5 0 .3 4 × • • • × × × × • • × × × × × × × • I Co n tr a ct o r × • × • × × • • × × × • × × × • 2 1 .1 1 1 2 .0 4 3 0 .1 9 0 .2 0 0 .5 9 0 .3 7 µ Co n tr a ct o r 2 4 .5 0 – 5 0 .2 8 0 .2 1 0 .5 3 0 .5 4 × • • • × × × × • × × × × × × × × × � 340 C a se C In s t -3 C 1 C 2 C 3 C 4 C 5 C 6 C 7 I Em p lo ye e µ Em p lo ye e C 8 C 9 C 1 0 C 1 1 C 1 2 C 1 3 C 1 4 I Co n tr a ct o r µ Co n tr a ct o r A tt r ib u te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 × • × • × × × × × × • • × × × × × • × • • × • • • × × × • • × × • × × × 1 • • • × • × × × • × × × × × × × × × 2 • × • × • • • × × • × • × × × × × 4 • × • • × • • × × × × × × × × × × 5 • × • × • × • × × × • • × × × × × • 7 × • • × • × × × × • × 7 • • • • × • • × × • • • × • × × × × 8 • × • × • • • × × • • • × • • • • • • • × • • • × × × • • × × × × × × × • × • × × × × • × × × × × × × × × 3 × • × • × × • • • × × × × × × × • × 3 × • • • × × × × • × × × × × • × × × 5 × • • • × × × • • • × × × × × × × • 6 • • • × • • × • • × • • × × × • × • 7 × • × × • × × × × • • × 7 × • • • × × × × • • × × × × × × × • 8 × • × • × × • • × × × • × × × • × • • • × × × × • × × × × × × × × × d K d U 3 4 .3 9 – 3 5 .4 5 – 5 4 .8 1 5 .6 3 3 7 .7 2 4 .0 2 2 7 .9 6 – 1 8 .8 4 6 2 .6 5 4 2 .3 1 – 7 9 .3 2 4 .9 0 3 1 .2 8 – 1 7 .4 2 – 3 4 .7 4 – 3 0 .5 6 – 3 0 .0 4 – 4 5 .1 3 – 2 0 .5 8 5 4 .9 7 2 4 .4 0 – 3 1 .2 5 1 2 .0 4 2 3 .3 6 – 7 8 1 0 8 6 4 7 1 2 7 4 7 6 6 8 4 5 5 5 S S � 0 .3 9 0 .3 0 0 .4 4 0 .3 1 0 .5 9 0 .5 0 0 .4 7 0 .3 4 0 .3 3 0 .2 4 0 .3 6 0 .3 6 0 .3 9 0 .3 7 0 .7 1 0 .7 2 0 .3 9 0 .2 7 0 .2 2 0 .1 5 0 .3 9 0 .3 0 0 .3 3 0 .2 6 0 .3 3 0 .2 6 0 .4 4 0 .3 9 0 .3 3 0 .3 4 0 .2 8 0 .2 1 0 .3 1 0 .3 0 0 .2 8 0 .2 0 r r � 0 .1 9 0 .2 0 − 0. 1 1 0 .0 5 − 0. 2 8 − 0. 3 0 − 0. 1 3 − 0. 1 9 0 .2 7 0 .2 3 0 .2 1 0 .2 7 0 .3 1 0 .1 2 − 0. 2 5 − 0. 3 6 0 .1 2 0 .0 8 0 .3 9 0 .4 9 0 .0 9 0 .1 1 0 .1 7 0 .1 7 0 .2 7 0 .2 8 0 .2 4 0 .1 3 0 .2 5 0 .2 4 0 .3 5 0 .3 8 0 .3 1 0 .1 6 0 .4 1 0 .4 2 R e su lt E m p lo y e e � � + 2 3 .3 0 I � � + 2 3 .3 0 µ � 1 6 .1 0 C o n tr a c to r � 2 � + 5 5 .9 0 I � 2 µ � + 5 2 .6 3 � 3 3 .3 3 Appendix C: A complete example Instantiation 3 � C a se C In s t -4 C 1 C 2 C 3 C 4 C 5 C 6 C 7 I Em p lo ye e µ Em p lo ye e C 8 C 9 C 1 0 C 1 1 C 1 2 C 1 3 C 1 4 I Co n tr a ct o r µ Co n tr a ct o r A tt r ib u te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 × • × • × × × × × × • × × × × × × • × • • × • • • × × × • • × × • × × × 1 • • • × • × × × • × × × × × × × × × 2 • × • × • • • × × • × • × × × × × 4 • × • • × • • × × × × × × × × × × 5 • × • × • × • × × × • • × × × × × • 7 × • • × • × × × × • × 7 • • • • × • • × × • • • × • × × × × 8 • × • × • • • × × • • • × • • • • • • • × • • • × × × • • × × × × × × × • × • × × × × • × × × × × × × × × 3 × • × • × × • • • × × × × × × × • × 3 × • • • × × × × • × × × × × • × × × 5 × • • • × × × • • • × × × × × × × • 6 • • • × • • × • • × • • × × × • × • 7 × • × × • × × × × • • × 7 × • • • × × × × • • × × × × × × × • 8 × • × • × × • • × × × • × × × • × • • • × × × × • × × × × × × × × × d K d U 3 8 .8 9 – 3 0 .9 5 – 5 0 .3 1 5 .6 3 3 3 .2 2 4 .0 2 3 2 .4 6 – 1 8 .8 4 6 2 .6 5 4 6 .8 1 – 8 3 .8 2 4 .9 0 3 5 .7 8 – 1 2 .9 2 – 3 0 .2 4 – 2 6 .0 6 – 2 5 .5 4 – 4 9 .6 3 – 2 0 .5 8 5 4 .9 7 1 9 .9 0 – 2 6 .7 5 1 2 .0 4 1 8 .8 6 – 8 7 9 7 7 4 8 1 3 8 3 6 5 5 9 4 4 4 4 S S � 0 .4 4 0 .3 4 0 .3 9 0 .2 7 0 .5 3 0 .4 6 0 .4 1 0 .3 0 0 .3 9 0 .2 8 0 .3 6 0 .3 6 0 .4 4 0 .4 1 0 .7 6 0 .7 6 0 .4 4 0 .3 1 0 .1 7 0 .1 1 0 .3 3 0 .2 6 0 .2 8 0 .2 3 0 .2 8 0 .2 2 0 .5 0 0 .4 3 0 .3 3 0 .3 4 0 .2 2 0 .1 7 0 .2 5 0 .2 6 0 .2 2 0 .1 6 r r � 0 .0 6 0 .1 0 − 0. 0 3 0 .1 3 − 0. 1 8 − 0. 2 3 − 0. 0 5 − 0. 1 2 0 .1 2 0 .0 9 0 .2 1 0 .2 7 0 .2 1 0 .0 7 − 0. 3 8 − 0. 3 9 0 .0 8 0 .0 6 0 .4 8 0 .5 8 0 .1 9 0 .2 0 0 .2 7 0 .2 5 0 .4 0 0 .4 0 0 .1 5 0 .1 0 0 .2 5 0 .2 4 0 .4 7 0 .4 9 0 .4 5 0 .2 6 0 .5 2 0 .5 4 R e su lt E m p lo y e e � � + 1 9 .3 0 I � � + 1 9 .3 0 µ � 1 2 .1 0 C o n tr a c to r � 2 � + 6 3 .1 0 I � 2 µ � + 5 9 .8 3 � 4 0 .5 3 � C.11 Distances file for BWIU v. Odco (LATEX output) 341 Instantiation 4 � � � 342 C a se C H y p o -1 C 1 C 2 C 3 C 4 A tt r ib u te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 × • × • × × × • × • × × × × × × × • • × • • • × × × • • × × • × × × 1 • • • × • × × × • × × × × × × × × × 2 • × • × • • • × × • × • × × × × × 4 • × • • × • • × × × × × × × × × × 5 d K d U 3 3 .5 2 1 0 .1 3 2 2 .0 2 1 0 .1 3 4 9 .4 3 1 0 .1 3 2 8 .3 2 1 4 .1 6 7 5 9 6 S S � 0 .4 4 0 .3 2 0 .3 1 0 .2 1 0 .5 6 0 .4 7 0 .4 0 0 .2 8 r r � R e su lt I 0 .0 7 0 .1 4 0 .2 3 0 .3 4 − 0. 2 2 − 0. 2 6 E m p lo y e e 0 .0 0 0 .0 8 1 2 .1 0 − � • × • × • × • × × × × × × × × µ C 5 7 3 6 .8 9 1 0 .1 3 8 0 .5 0 0 .3 5 0 .1 5 0 .2 0 1 2 .1 0 • • • − − � C 6 × • • × • × × × × • × 7 1 7 .9 6 6 8 .2 9 4 0 .4 0 0 .3 8 0 .1 7 0 .2 7 4 .9 0 C 7 I Em p lo ye e µ Em p lo ye e C 8 C 9 C 1 0 C 1 1 • • • • × • • × × • • • × • × × × × 8 • × • × • • • × × • • • × • • • • • • • × • • • × × × • • × × × × × × × • × • × × × × • × × × × × × × × × 3 × • × • × × • • • × × × × × × × • × 3 × • • • × × × × • × × × × × • × × × 5 × • • • × × × • • • × × × × × × × • 6 4 1 .4 3 1 0 .1 3 8 3 .3 4 1 5 .0 3 3 0 .4 0 1 0 .1 3 4 .0 0 1 0 .1 3 1 5 .6 8 1 0 .1 3 1 7 .1 4 1 0 .1 3 2 0 .7 8 1 0 .1 3 7 1 3 7 1 3 3 4 0 .4 4 0 .3 9 0 .8 7 0 .8 3 0 .4 4 0 .2 9 0 .0 6 0 .0 4 0 .1 9 0 .1 5 0 .1 9 0 .1 6 0 .2 5 0 .2 0 µ I 0 .2 2 0 .0 9 − 0. 6 6 − 0. 5 8 0 .0 7 0 .0 4 0 .8 3 0 .8 8 0 .5 4 0 .4 8 0 .5 4 0 .4 5 C o n tr a c to r 0 .4 5 0 .4 1 + 5 9 .9 8 � 2 � � 2 � Appendix C: A complete example Hypothetical 1 C 1 2 7 4 0 .3 7 1 0 .1 3 7 0 .4 4 0 .3 8 0 .2 2 0 .1 0 • • • × • • × • • × • • × × × • × • + 5 6 .7 2 C 1 3 × • × × • × × × × • • × 7 1 9 .7 1 6 0 .6 1 4 0 .3 6 0 .3 6 0 .2 1 0 .2 4 4 1 .5 0 C 1 4 8 2 0 .7 8 1 0 .1 3 4 0 .2 5 0 .2 0 0 .4 5 0 .4 1 × • • • × × × × • • × × × × × × × • I Co n tr a ct o r × • × • × × • • × × × • × × × • 2 1 .9 9 2 2 .1 7 3 0 .2 1 0 .2 4 0 .5 2 0 .2 5 µ Co n tr a ct o r 9 .9 4 1 0 .1 3 2 0 .1 3 0 .0 9 0 .7 5 0 .6 8 × • • • × × × × • × × × × × × × × × � C a se C H y p o -2 C 1 C 2 C 3 C 4 C 5 C 6 C 7 I Em p lo ye e µ Em p lo ye e C 8 C 9 C 1 0 C 1 1 C 1 2 C 1 3 C 1 4 I Co n tr a ct o r µ Co n tr a ct o r A tt r ib u te s c A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 0 A 1 1 A 1 2 A 1 3 A 1 4 A 1 5 A 1 6 A 1 7 A 1 8 × × • • × × × × × • × × × × × • × • • × • • • × × × • • × × • × × × 1 • • • × • × × × • × × × × × × × × × 2 • × • × • • • × × • × • × × × × × 4 • × • • × • • × × × × × × × × × × 5 • × • × • × • × × × • • × × × × × • 7 × • • × • × × × × • × 7 • • • • × • • × × • • • × • × × × × 8 • × • × • • • × × • • • × • • • • • • • × • • • × × × • • × × × × × × × • × • × × × × • × × × × × × × × × 3 × • × • × × • • • × × × × × × × • × 3 × • • • × × × × • × × × × × • × × × 5 × • • • × × × • • • × × × × × × × • 6 • • • × • • × • • × • • × × × • × • 7 × • × × • × × × × • • × 7 × • • • × × × × • • × × × × × × × • 8 × • × • × × • • × × × • × × × • × • • • × × × × • × × × × × × × × × d K d U 3 4 .3 9 1 0 .1 3 3 0 .9 5 1 0 .1 3 3 8 .4 3 1 0 .1 3 2 1 .3 4 1 4 .1 6 1 6 .0 8 1 0 .1 3 1 8 .8 4 6 8 .2 9 4 2 .3 1 1 0 .1 3 6 7 .4 4 1 5 .0 3 3 1 .2 8 1 0 .1 3 2 4 .8 0 1 0 .1 3 3 6 .4 8 1 0 .1 3 2 6 .0 6 1 0 .1 3 1 9 .9 0 1 0 .1 3 3 9 .4 9 1 0 .1 3 3 2 .4 6 6 0 .6 1 1 9 .9 0 1 0 .1 3 3 2 .9 9 2 2 .1 7 1 8 .8 6 1 0 .1 3 7 7 7 5 4 4 7 1 0 7 5 7 5 4 7 6 4 5 4 S S � 0 .4 4 0 .3 3 0 .4 4 0 .2 9 0 .4 4 0 .3 6 0 .3 3 0 .2 1 0 .2 5 0 .1 5 0 .4 0 0 .4 0 0 .4 4 0 .4 0 0 .6 7 0 .6 7 0 .4 4 0 .3 0 0 .3 1 0 .2 4 0 .4 4 0 .3 5 0 .3 1 0 .2 5 0 .2 5 0 .1 9 0 .4 4 0 .3 7 0 .5 5 0 .5 9 0 .2 5 0 .1 9 0 .3 6 0 .3 5 0 .2 5 0 .1 8 r r � 0 .0 7 0 .1 1 − 0. 0 8 0 .0 9 0 .0 7 0 .0 4 0 .2 1 0 .3 4 0 .4 5 0 .5 9 0 .1 7 0 .2 3 0 .2 2 0 .0 6 − 0. 0 8 − 0. 1 5 0 .2 5 0 .3 2 0 .0 9 0 .0 1 − − 0. 0 8 − 0. 1 7 0 .2 3 0 .2 2 0 .4 5 0 .4 6 0 .2 2 0 .1 4 − 0. 1 8 − 0. 3 0 0 .4 5 0 .4 6 0 .1 9 0 .0 3 − 0 .3 1 0 .2 4 R e su lt E m p lo y e e � � + 2 3 .3 8 I � � + 2 3 .3 8 µ � 1 2 .1 0 C o n tr a c to r � � + 5 1 .8 2 I � � + 4 8 .5 5 µ � 3 3 .3 3 � C.11 Distances file for BWIU v. Odco (LATEX output) 343 Hypothetical 2 344 Appendix C: A complete example C.12 Report file for Building Workers’ Industrial Union of Australia v. Odco Pty Ltd (LATEX input) % Report file % Produced by SHYSTER version 1.0 % Copyright James Popple 1993 % This is not a stand-alone LaTeX file. % Include it in a LaTeX document using the \input command. % Use LaTeX version 2.09 <25 March 1992> and TeX version 3.141. \subsection*{Employee area} \subsubsection*{Instant case} The law distinguishes between a contract of service (between employer and employee) and a contract for services (between principal and independent ... \medskip\noindent In the instant case, the employer did not direct the manner in which the work was to be done; the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; ... \medskip\noindent In my opinion---following \frenchspacing {\it Humberstone v. Northern Timber Mills\/}\nonfrenchspacing---% the worker is an independent contractor. \medskip\noindent In \frenchspacing {\it Humberstone v. Northern Timber Mills}\nonfrenchspacing,% \footnote{(1949) 79 CLR 389.} a 1949 decision of three judges of the High Court of Australia, Humberstone carried goods for NTM. He had originally held himself out as a carrier, prepared to carry for anyone, but for over twenty years he had ... There are several significant similarities between the instant case and \frenchspacing {\it Humberstone v. NTM\/}\null\nonfrenchspacing: the employer did not direct the manner in which the work was to be done; the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; ... However, the instant case is not on all fours with \frenchspacing {\it Humberstone v. NTM}\null\nonfrenchspacing. � C.12 Report file for BWIU v. Odco (LATEX input) 345 In that case the worker was not in business on her/his own account; the worker was allowed to employ others to assist with her/his work; ... Nevertheless, I believe that \frenchspacing {\it Humberstone v. NTM\/} \nonfrenchspacing should be followed. \medskip\noindent If \frenchspacing {\it Ferguson v. John Dawson \& Partners (Contractors) Ltd\/} \nonfrenchspacing is followed then the worker is an employee. \medskip\noindent In \frenchspacing {\it Ferguson v. John Dawson \& Partners (Contractors) Ltd}\nonfrenchspacing,% \footnote{[1976] 1 WLR 1213.} a 1976 decision of the English Court of Appeal, Ferguson fell off a roof while removing some scaffolding boards. He claimed damages against John Dawson (the building contractors) for ... There are several similarities between the instant case and \frenchspacing {\it Ferguson v. Dawson\/}\null\nonfrenchspacing: the work was not performed on the employer’s premises; the worker was not allowed to employ others to assist with her/his work; ... However, there are several significant differences between the instant case and \frenchspacing {\it Ferguson v. Dawson}\null\nonfrenchspacing. In that case the employer directed the manner in which the work was to be done; the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; ... Note also that \frenchspacing {\it Ferguson v. Dawson\/} \nonfrenchspacing is only a decision of the English Court of Appeal and not as good authority as a case decided by three judges of the High Court of Australia% ---like \frenchspacing {\it Humberstone v. NTM}\null\nonfrenchspacing. Consequently, there is nothing in \frenchspacing {\it Ferguson v. Dawson\/} \nonfrenchspacing to warrant any change in my conclusion. \subsubsection*{Hypothetical 1} ... 346 Appendix C: A complete example C.13 Report file for Building Workers’ Industrial Union of Australia v. Odco Pty Ltd (LATEX output) Employee area Instant case The law distinguishes between a contract of service (between employer and employee) and a contract for services (between principal and independent contractor). This dis­ tinction affects the terms that will be implied in the absence of an express agreement, the liability of the employer to third parties, the applicability of industrial awards, the applicability of statutes which may affect workers’ compensation, occupational health and safety, long-service leave, fringe benefits tax, etc. The terms “employer” and “worker” are used here to mean “employer” and “em­ ployee” (in the case of a contract of service) or “principal” and “independent con­ tractor” (in the case of a contract for services). In the instant case, the employer did not direct the manner in which the work was to be done; the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was not an integral part of the employer’s business, but was accessory to it; the worker owned the tools or provided the transport with which she/he performed the work; the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the work was not performed on the employer’s premises; the employer neither supervised nor inspected the work; it is not known whether the worker was in business on her/his own account; the worker was not allowed to employ others to assist with her/his work; the worker was not obliged to work only for the employer; the worker was required to work at specified times; it is not known whether the employer paid the worker by time; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. In my opinion—following Humberstone v. Northern Timber Mills—the worker is an independent contractor. In Humberstone v. Northern Timber Mills, 1 a 1949 decision of three judges of the High Court of Australia, Humberstone carried goods for NTM. He had originally held himself out as a carrier, prepared to carry for anyone, but for over twenty years he had carried 1(1949) 79 CLR 389. � C.13 Report file for BWIU v. Odco (LATEX output) 347 goods solely for NTM (although he would, infrequently, carry back-loads for NTM’s customers). Humberstone owned the truck, and paid for petrol and repairs. He was paid weekly on a weight-mileage basis. He was a licenced carrier, and had his name printed on the side of his truck with the description “carrier.” On the way back from a job, he had a puncture. He went home to change the wheel, but exerted himself so strenuously in trying to remove the tyre from the wheel that he became ill and later lapsed into a coma, from which he did not recover. Section 3 of the Worker’s Compensation Act 1928 (Vic) had been amended about a year before Humberstone’s death so as to include independent contractors in its definition of a “worker” covered by the Act. However, the High Court held that the amendment applied only to contracts entered into after it came into operation. Further, the Court decided that Humberstone was not an employee of NTM. Hence he was not a “worker” under the Act and his widow was not entitled to compensation under the Act. There are several significant similarities between the instant case and Humber- stone v. NTM : the employer did not direct the manner in which the work was to be done; the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was not an integral part of the employer’s business, but was accessory to it; the worker owned the tools or provided the transport with which she/he performed the work; the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the work was not performed on the employer’s premises; the employer neither supervised nor inspected the work; the worker was not obliged to work only for the employer; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; and the employer and the worker did not express any intention that the relationship would be one of employer and employee. However, the instant case is not on all fours with Humberstone v. NTM. In that case the worker was not in business on her/his own account; the worker was allowed to employ others to assist with her/his work; the worker was not required to work at specified times; the employer did not pay the worker by time; and the employer and the worker did not express any intention that the relationship would be one of principal and independent contractor. Nevertheless, I believe that Humberstone v. NTM should be followed. If Ferguson v. John Dawson & Partners (Contractors) Ltd is followed then the worker is an employee. In Ferguson v. John Dawson & Partners (Contractors) Ltd, 2 a 1976 decision of the Eng­ lish Court of Appeal, Ferguson fell off a roof while removing some scaffolding boards. He claimed damages against John Dawson (the building contractors) for breach of stat­ utory duty relying on the Construction (Working Places) Regulations 1966 (UK). This duty would only be owed if Ferguson was an employee of John Dawson. 2[1976] 1 WLR 1213. 348 Appendix C: A complete example Megaw and Browne LJJ held that, despite the fact that both parties labelled Fer­ guson a “self-employed labour only subcontractor”,3 the reality of the relationship between them was that of employer/employee. There are several similarities between the instant case and Ferguson v. Dawson: the work was not performed on the employer’s premises; the worker was not allowed to employ others to assist with her/his work; the worker was not obliged to work only for the employer; the worker was required to work at specified times; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, there are several significant differences between the instant case and Fer­ guson v. Dawson. In that case the employer directed the manner in which the work was to be done; the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was an integral part of the employer’s business; the worker neither owned the tools nor provided the transport with which she/he performed the work; the employer would make a profit/loss if the work performed by the worker cost less/more than expected; the employer supervised or inspected the work; the worker was not in business on her/his own account; and the employer paid the worker by time. Note also that Ferguson v. Dawson is only a decision of the English Court of Appeal and not as good authority as a case decided by three judges of the High Court of Australia—like Humberstone v. NTM. Consequently, there is nothing in Ferguson v. Dawson to warrant any change in my conclusion. Hypothetical 1 Consider the instant case changed so that the following is true: the worker was allowed to employ others to assist with her/his work; and the employer and the worker did not express any intention that the relationship would be one of principal and independent contractor. If that were so then I would be more strongly of the opinion that—following Humber- stone v. Northern Timber Mills—the worker is an independent contractor. Details of Humberstone v. NTM are summarized above. There are several significant similarities between the hypothetical case and Humberstone v. NTM : the employer did not direct the manner in which the work was to be done; the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was not an integral part of the employer’s business, but was accessory to it; the worker owned the tools or provided the transport with which she/he performed the work; the employer would not make a profit/loss if the work performed by the worker 3ibid. at 1219, per Megaw LJ; at 1225, per Lawton LJ; at 1228, per Browne LJ. � C.13 Report file for BWIU v. Odco (LATEX output) 349 cost less/more than expected; the work was not performed on the employer’s premises; the employer neither supervised nor inspected the work; the worker was allowed to employ others to assist with her/his work; the worker was not obliged to work only for the employer; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker did not express any intention that the relationship would be one of principal and independent contractor. However, the hypothetical case is not on all fours with Humberstone v. NTM. In that case the worker was not in business on her/his own account; the worker was not required to work at specified times; and the employer did not pay the worker by time. Nevertheless, I believe that Humberstone v. NTM should be followed. If Cam and Sons Pty Ltd v. Sargent is followed then the worker is an employee. In Cam and Sons Pty Ltd v. Sargent, 4 a 1940 decision of four judges of the High Court of Australia, Sargent was the master of a ship. He entered into an agreement with Cam and Sons that claimed that the ship was hired by Cam and Sons to Sargent and his fellow contractors (called “the partnership”). However, it was doubtful whether that agreement actually deprived Cam and Sons of any control over the ship. The partnership was to use the ship only to carry coal from Swansea to Sydney. Cam and Sons were sole agents of the partnership for securing cargoes for the ship, and for collecting money due to the partnership. The partnership paid nothing for the “hire” of the ship, but received a specified sum for each return trip of a certain tonnage plus (in certain circumstances) 5% of the earnings, the balance of which was retained by Cam and Sons. Cam and Sons had to approve people employed by the partnership. Sargent claimed that he (and others in the partnership) were employed by Cam and Sons, and therefore came within the terms of an industrial award. Cam and Sons claimed that members of the partnership were independent contractors. The High Court unanimously agreed with Sargent. Rich J came to the conclusion that the agreement was an attempt to evade the terms of the industrial award.5 There are several similarities between the hypothetical case and Cam v. Sargent: the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the work was not performed on the employer’s premises; the employer neither supervised nor inspected the work; the worker was allowed to employ others to assist with her/his work; the worker was not obliged to work only for the employer; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the 4(1940) 14 ALJ 162. 5ibid. at 163. 350 Appendix C: A complete example worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker did not express any intention that the relationship would be one of principal and independent contractor. However, there are several significant differences between the hypothetical case and Cam v. Sargent. In that case the employer directed the manner in which the work was to be done; the worker was an integral part of the employer’s business; the worker neither owned the tools nor provided the transport with which she/he performed the work; the employer would make a profit/loss if the work performed by the worker cost less/more than expected; the worker was not in business on her/his own account; the worker was not required to work at specified times; and the employer did not pay the worker by time. Despite the fact that Cam v. Sargent is a decision of four judges of the High Court of Australia (and better authority than a case decided by three judges of the High Court of Australia—like Humberstone v. NTM ), there is nothing in Cam v. Sargent to warrant any change in my conclusion. Hypothetical 2 Consider the instant case changed so that the following is true: the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; and the worker was an integral part of the employer’s business. If that were so then my opinion would be that—following Ferguson v. John Dawson & Partners (Contractors) Ltd —the worker is an employee. Details of Ferguson v. Dawson are summarized above. There are several significant similarities between the hypothetical case and Ferguson v. Dawson: the worker was not allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was an integral part of the employer’s business; the work was not performed on the employer’s premises; the worker was not allowed to employ others to assist with her/his work; the worker was not obliged to work only for the employer; the worker was required to work at specified times; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, the hypothetical case is not on all fours with Ferguson v. Dawson. In that case the employer directed the manner in which the work was to be done; the worker neither owned the tools nor provided the transport with which she/he performed the work; the employer would make a profit/loss if the work performed by the worker cost less/more than expected; the employer supervised or inspected the work; the worker was not in business on her/his own account; and the employer paid the worker by time. Nevertheless, I believe that Ferguson v. Dawson should be followed. � C.13 Report file for BWIU v. Odco (LATEX output) 351 If Australian Mutual Provident Society v. Chaplin or Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance are followed then the worker is an independent contractor. In Australian Mutual Provident Society v. Chaplin, 6 a 1978 decision of the Judicial Committee of the Privy Council, Chaplin was a representative of AMP. A clause of the agreement between them stated that the relationship was one of “principal and agent” and not one of “master and servant”. Chaplin claimed that he was employed under a contract of service, and therefore a “worker” under the Long Service Leave Act, 1967 (SA) and entitled to certain benefits. The Privy Council found that there was no reason to think that the clause was not a genuine statement of the parties’ intentions. Examining the agreement, their lordships concluded that it provided for a contract of agency. The fact that Chaplin was given the power of unlimited delegation of the whole performance of his work was “almost conclusive against the contract being a contract of service.”7 There are several similarities between the hypothetical case and AMP v. Chaplin: the employer did not direct the manner in which the work was to be done; the worker was an integral part of the employer’s business; the worker owned the tools or provided the transport with which she/he performed the work; the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the work was not performed on the employer’s premises; the employer neither supervised nor inspected the work; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, there are several significant differences between the hypothetical case and AMP v. Chaplin. In that case the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was in business on her/his own account; the worker was allowed to employ others to assist with her/his work; the worker was obliged to work only for the employer; the worker was not required to work at specified times; and the employer did not pay the worker by time. Despite the fact that AMP v. Chaplin is a decision of the Judicial Committee of the Privy Council (and better authority than a case decided by the English Court of Appeal—like Ferguson v. Dawson), there is nothing in AMP v. Chaplin to warrant any change in my conclusion. In 1967, Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance8 was decided by the Queen’s Bench Division of the English High Court. (A case decided by the Queen’s Bench Division of the English High Court is not as 6(1978) 18 ALR 385. 7ibid. at 391. 8[1968] 2 QB 497. 352 Appendix C: A complete example good authority as a case decided by the Judicial Committee of the Privy Council—like AMP v. Chaplin; furthermore Ready Mixed v. Minister is 11 years older than AMP v. Chaplin.) In Ready Mixed v. Minister, Latimer worked for Ready Mixed as an “owner-driver.” He was paid at mileage rates, and was obliged to buy the truck through a financial organization associated with Ready Mixed. The truck was painted in the company’s colours, and he had to wear a Ready Mixed uniform. Latimer was obliged to meet the costs of maintenance, repair and insurance of the truck (and the attached mixing unit, which belonged to Ready Mixed). The Minister determined that Latimer was employed under a contract of service and therefore an “employed person” under s. 1(2) of the National Insurance Act 1965 (UK), making Ready Mixed liable to make weekly contributions. MacKenna J examined the contract and held that the rights it conferred, and the duties it imposed, between Latimer and Ready Mixed were not such as to make it a contract of service. There are several similarities between the hypothetical case and Ready Mixed v. Minister : the employer did not direct the manner in which the work was to be done; the worker was an integral part of the employer’s business; the worker owned the tools or provided the transport with which she/he performed the work; the employer would not make a profit/loss if the work performed by the worker cost less/more than expected; the work was not performed on the employer’s premises; the employer neither supervised nor inspected the work; the money that the employer paid to the worker was not stated to be a “fee”; the money that the employer paid to the worker was not stated to be “wages” or “salary”; the employer did not deduct PAYE tax instalments from the worker’s pay; the employer paid the worker neither sick pay nor holiday pay; the employer and the worker did not express any intention that the relationship would be one of employer and employee; and the employer and the worker expressed an intention that the relationship would be one of principal and independent contractor. However, there are several significant differences between the hypothetical case and Ready Mixed v. Minister. In that case the worker was allowed to use her/his own discretion in doing an aspect of the work that was not specified beforehand; the worker was not in business on her/his own account; the worker was allowed to employ others to assist with her/his work; the worker was obliged to work only for the employer; the worker was not required to work at specified times; and the employer did not pay the worker by time. Note also that Ready Mixed v. Minister is only a decision of the Queen’s Bench Division of the English High Court and not as good authority as a case decided by the English Court of Appeal—like Ferguson v. Dawson. Consequently, there is nothing in Ready Mixed v. Minister to warrant any change in my conclusion. D Reflexive tests Great cases, like hard cases, make bad law. Oliver Wendell Holmes (1904) Northern Securities Co. v. United States29 This is the tale of the 1401, The law clerk that was nobody’s son. It spends its days in a furious hunt For authorities, dictum, and argumunt. But after it found them, it burned with shame; The Supreme Court reversed it just the same. T. H. Lassagne (1963)30 353 354 Appendix D: Reflexive tests D.1 Introduction A reflexive test is performed by removing one of the leading cases from a spe­ cification, and using it as a test case with the diminished specification. Reflexive tests, and their limitations, are discussed in §3.13.5. A reflexive test was performed for every leading case in each of the four specifications described in chapter 5. Details of each of the cases are summarized in the dump files of the specifications given in appendix A. The results of these reflexive tests are described below (§D.2–§D.5), and sum­ marized in four figures in this appendix. The symbols used in these figures are the same as those used in figure 5.12 and are explained in §5.6. Conclusions are drawn from the results of this reflexive testing in §5.6.3. D.2 The FINDER specification The results of the reflexive testing of all cases in the Finder specification are described here, and summarized in figure D.1. Hannah v. Peel 1945 In Hannah v. Peel 31 (C1), Birkett J of the King’s Bench Division of the English High Court followed Bridges v. Hawkesworth, and distinguished South Stafford­ shire Water Co. v. Sharman and Elwes v. Brigg Gas Co.32 Hannah, the finder, won. SHYSTER concludes that Hannah should have lost. The nearest neighbour is South Staffordshire v. Sharman; the nearest other is Bridges v. Hawkesworth. Its choice of cases is good, although SHYSTER chooses to follow the wrong one. However, SHYSTER warns that these two cases are equidistant from Hannah v. Peel ; it only chooses South Staffordshire v. Sharman because it is the more recent of the two—both were decided by the Queen’s Bench Division. It also warns that the weighted association coefficients suggest that Bridges v. Hawkesworth should be the nearest neighbour. Bridges v. Hawkesworth 1851 In Bridges v. Hawkesworth33 (C2), Patteson J of the Queen’s Bench Division could find no authority “directly in point” and found no reason not to apply the general rule, as stated in Armory v. Delamirie, that the finder is entitled to the found chattel as against all parties except the real owner.34 SHYSTER agrees that Bridges should have won. The nearest neighbour is Hannah v. Peel ; the nearest other is City of London Corporation v. Appleyard (2). Both of these cases were decided after Bridges v. Hawkesworth. The results of this test are consistent with the results of a similar reflexive test performed by Tyree, Greenleaf and Mowbray using FINDER. 35 � D.2 The FINDER specification 355 (With Bridges v. Hawkesworth removed from the specification, there is an equivalent functional dependence between A8 and A9.) Armory v. Delamirie 1722 Armory v. Delamirie36 (C3) is the oldest of all the finder cases. Pratt CJ of the King’s Bench Division had no previous authorities upon which to base his decision, which was in favour of the finder, Armory. SHYSTER agrees that the finder should have won. The nearest neighbour is Bridges v. Hawkesworth; the nearest other is Moffatt v. Kazana. Of course, both of these cases were decided after Armory v. Delamirie. However, SHYSTER warns that the weighted correlation coefficients suggest that six other cases should be the nearest neighbour—in five of which, the finder lost. (Without Armory v. Delamirie, there is an equivalent functional dependence between A4 and A10. Further, A9 has infinite weight.) Moffatt v. Kazana 1967 In Moffatt v. Kazana37 (C4), Wrangham J of the Queen’s Bench Division fol­ lowed City of London Corporation v. Appleyard and found against Kazana (the finder).38 SHYSTER concludes, unlike the Court, that Kazana should have won. The nearest neighbour is Hannah v. Peel ; the nearest others are City of London Cor­ poration v. Appleyard (1) and Elwes v. Brigg Gas Co. (With Moffatt v. Kazana removed from the specification, there are two func­ tional dependencies: equivalence between A1 and A6, and inverse between A9 and A10. Further, A4 has infinite weight.) City of London Corporation v. Appleyard (1) 1963 As discussed in §5.2.2, City of London Corporation v. Appleyard appears in the specification twice. In London v. Appleyard (1)39 (C5), McNair J resolved the conflict between Ap­ pleyard (as finder) and Yorkwin by finding against Appleyard. In coming to this decision, he followed South Staffordshire Water Co. v. Sharman and Hannah v. Peel. 40 SHYSTER agrees. The nearest neighbour is South Staffordshire v. Sharman; the nearest other is Hannah v. Peel. (With London v. Appleyard (1) removed from the specification, there is an inverse functional dependence between A3 and A10.) � � � � � � � � � � � � � � � � 356 Appendix D: Reflexive tests Case Hannah v. Peel Bridges v. Hawkesworth Armory v. Delamirie Moffatt v. Kazana London v. Appleyard (1) London v. Appleyard (2) South Staffordshire v. Sharman Elwes v. Brigg Gas Result × × Nearest n’bours – – × × – – Nearest others – – × – – Warnings S� �→ r� × × Figure D.1: A summary of the reflexive testing of the Finder specification, described in §D.2. The meaning of the symbols used here is explained in §5.6.1. City of London Corporation v. Appleyard (2) 1963 In City of London Corporation v. Appleyard (2)41 (C6), McNair J resolved the conflict between Yorkwin (as finder) and the City of London by finding against Yorkwin. His decision was based on the construction of a clause of the lease agreement between the two parties. SHYSTER agrees that the finder should lose. The nearest neighbour is Elwes v. Brigg Gas Co.; the nearest others are Hannah v. Peel and Bridges v. Hawkes­ worth. (With London v. Appleyard (2) removed from the specification, there is an inverse functional dependence between A5 and A8.) South Staffordshire Water Co. v. Sharman 1896 In South Staffordshire Water Co. v. Sharman 42 (C7), Lord Russell of Killowen CJ and Wills J distinguished Bridges v. Hawkesworth and found against Sharman (the finder).43 SHYSTER agrees that Sharman should have lost. The nearest neighbour is City of London Corporation v. Appleyard (1); the nearest other is Hannah v. Peel. Both of these cases were decided after South Staffordshire v. Sharman. Elwes v. Brigg Gas Co. 1886 In Elwes v. Brigg Gas Co.44 (C8), Chitty J found against the finder. SHYSTER agrees that Brigg Gas should have lost. The nearest neighbour is City of London Corporation v. Appleyard (2); the nearest other is Hannah v. Peel. Both of these cases were decided after Elwes v. Brigg Gas Co. � D.3 The AUTHORIZATION specification 357 D.3 The AUTHORIZATION specification The results of the reflexive testing of all cases in the Authorization specification are described here, and summarized in figure D.2. University of New South Wales v. Moorhouse 1975 In University of New South Wales v. Moorhouse45 (C1), the High Court ap­ plied Falcon v. Famous Players Film Co. and held that the infringement was authorized.46 SHYSTER agrees that the University authorized the infringement. The nearest neighbour is Winstone v. Wurlitzer : a case of which Gibbs J approved;47 the nearest others are A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd (Not- Auth), which was decided after UNSW v. Moorhouse, and Australasian Performing Right Association Ltd v. Miles (Liable). However, SHYSTER warns that two of these cases are equidistant from UNSW v. Moorhouse; it chooses Winstone v. Wurlitzer because it is a decision of the Su­ preme Court of Victoria whereas A&M v. Audio Magnetics is merely a decision of the Chancery Division of the English High Court. It also warns that the weighted correlation coefficients suggest that A&M v. Audio Magnetics should be the nearest neighbour, and that the specified directions suggest Liable. Australasian Performing Right Association Ltd v. Canterbury-Bankstown League Club Ltd 1964 Australasian Performing Right Association Ltd v. Canterbury-Bankstown League Club Ltd 48 (C2), appears in the specification as a case in which the accused authorized an infringement. However, in that case Asprey J cited Australasian Performing Right Association Ltd v. Miles and held that the club had performed the work, through the orchestra, and was directly liable for the infringement or, if that were not so, that the club had authorized the infringement.49 Ferguson J (with whom Herron CJ agreed) came to a similar conclusion.50 SHYSTER concludes that the Club is liable (directly or vicariously) for the in­ fringement. The nearest neighbour is APRA v. Miles; the nearest others are Mel­ lor v. Australian Broadcasting Commission (Auth) and RCA Corporation v. John Fairfax and Sons Ltd (Not-Auth), which was decided after APRA v. Canterbury- Bankstown. However, SHYSTER warns that two of these cases are equidistant from APRA v. Canterbury-Bankstown; it chooses APRA v. Miles because it is a decision of the Supreme Court of New South Wales whereas Mellor v. ABC is merely a decision of the Judicial Committee of the Privy Council. It also warns that the weighted correlation coefficients suggest that Mellor v. ABC should be the nearest neigh­ bour. Furthermore, two of the four instantiations have a nearest result of Auth. 358 Appendix D: Reflexive tests Winstone v. Wurlitzer Automatic Phonograph Co. of Australia Pty Ltd 1946 In Winstone v. Wurlitzer Automatic Phonograph Co. of Australia Pty Ltd 51 (C3), Herring CJ applied Bankes LJ’s test, from Falcon v. Famous Players Film Co., and decided that Wurlitzer had authorized the infringement.52 SHYSTER agrees, but all of its chosen cases were decided after Winstone v. Wurlitzer. The nearest neighbour is University of New South Wales v. Moor- house; the nearest others are A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd (Not-Auth) and Australasian Performing Right Association Ltd v. Miles (Li­ able). SHYSTER warns that the specified directions suggest Liable. Mellor v. Australian Broadcasting Commission 1940 In Mellor v. Australian Broadcasting Commission53 (C4), the Judicial Committee of the Privy Council made reference to no authorization cases in deciding that the ABC had authorized the infringement (although Mellor was held to have consented to that infringement). SHYSTER also concludes that the ABC authorized the infringement. The nearest neighbours are Winstone v. Wurlitzer Automatic Phonograph Co. of Aus­ tralia Pty Ltd and Australasian Performing Right Association Ltd v. Canterbury- Bankstown League Club Ltd ; the nearest others are Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd (Not-Auth) and Australasian Performing Right Association Ltd v. Miles (Liable). With the exception of PRS v. Ciryl, all of these cases were decided after Mellor v. ABC. (With Mellor v. ABC removed from the specification, there is an inverse func­ tional dependence between A2 and A6.) Falcon v. Famous Players Film Co. 1926 In Falcon v. Famous Players Film Co.54 (C5), a majority of the English Court of Appeal held that Famous Players had authorized the infringement; Bankes LJ made his oft-quoted statement that “authorize” should be “understood in its ordinary dictionary sense of ‘sanction, approve, and countenance.’ ”55 SHYSTER disagrees with the Court of Appeal and concludes that Famous Players did not authorize the infringement. It warns that the specified directions suggest Liable. The nearest neighbour is A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd ; the nearest others are Winstone v. Wurlitzer Automatic Phono­ graph Co. of Australia Pty Ltd (Auth) and Australasian Performing Right As­ sociation Ltd v. Miles (Liable). All of these cases were decided after Falcon v. Famous Players. � D.3 The AUTHORIZATION specification 359 RCA Corporation v. John Fairfax and Sons Ltd 1981 In RCA Corporation v. John Fairfax and Sons Ltd 56 (C6), Kearney J referred to several cases, especially University of New South Wales v. Moorhouse and A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd, before deciding that Fairfax had not authorized any infringement. SHYSTER also concludes that Fairfax did not authorize the infringement, al­ though it warns that the specified directions suggest Auth or Liable. The nearest neighbour is A&M v. Audio Magnetics: one of the cases to which his honour referred. The nearest others are Falcon v. Famous Players Film Co. (Auth) and Australasian Performing Right Association Ltd v. Miles (Liable), neither of which was mentioned in the judgment. Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd 1923 Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd 57 (C7) is the oldest of the leading cases. The only authorization case cited in argument was Performing Right Society Ltd v. Bradford Corporation. As discussed in §5.3.2, PRS v. Bradford had an identical fact vector to Mellor v. Australian Broadcasting Commission, and was omitted from the Authorization specification. SHYSTER comes to a different conclusion to that of the English Court of Ap­ peal; it concludes that Ciryl authorized the infringement. The nearest neighbour is Mellor v. ABC which (being identical to the omitted PRS v. Bradford ) is a good choice. The nearest others are RCA Corporation v. John Fairfax and Sons Ltd (Not-Auth) and Australasian Performing Right Association Ltd v. Miles (Li­ able). Of course, all of these cases were decided after PRS v. Ciryl. However, SHYSTER warns that the weighted correlation coefficients suggest that six other cases should be the nearest neighbour—including RCA v. Fairfax and another case in which the accused did not authorize the infringement, and APRA v. Miles. It also warns that the specified directions suggest Not-Auth. (With PRS v. Ciryl removed from the specification, there is an inverse func­ tional dependence between A5 and A6. Further, both of those attributes have infinite weight—instead of just A5, as before.) A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd 1978 In A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd 58 (C8), Foster J applied Bankes LJ’s definition of “authorization” from Falcon v. Famous Players Film Co. and held that Audio Magnetics had not authorized any specific infringement.59 � � � � � � � � � � � � 360 Appendix D: Reflexive tests Case UNSW v. Moorhouse APRA v. Canterbury-Bankstown Winstone v. Wurlitzer Mellor v. ABC Falcon v. Famous Players RCA v. Fairfax PRS v. Ciryl A&M v. Audio Magnetics APRA v. Miles Result × × Nearest n’bours – – – – Nearest others – – – – – × – × – Warnings r� ⊃ → × r� → ∨ 4 2 × ⊃ × × ⊃ × r� �⊃ r� ⊃ → × Figure D.2: A summary of the reflexive testing of the Authorization specific­ ation, described in §D.3. The meaning of the symbols used here is explained in §5.6.1. SHYSTER agrees with Foster J that Audio Magnetics did not authorize the in­ fringement. The nearest neighbour is RCA Corporation v. John Fairfax and Sons Ltd , which was decided after A&M v. Audio Magnetics; the nearest others are Falcon v. Famous Players (Auth) and Australasian Performing Right Association Ltd v. Miles (Liable). However, SHYSTER warns that two of these cases are equidistant from A&M v. Audio Magnetics; it chooses RCA v. Fairfax because it is a decision of the Su­ preme Court of New South Wales whereas Falcon v. Famous Players is merely a decision of the English Court of Appeal. It also warns that the weighted cor­ relation coefficients suggest that Falcon v. Famous Players should be the nearest neighbour, and that the specified directions suggest Auth. Australasian Performing Right Association Ltd v. Miles 1961 In Australasian Performing Right Association Ltd v. Miles60 (C9), Jacobs J fol­ lowed Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd— not a leading case in this specification, though it does appear in the Employee specification—and held the members of the Dee Why RSL Club directly liable for the infringing acts of the orchestra. SHYSTER concludes that Miles authorized the infringement. This is not sur­ prising. With APRA v. Miles removed from the specification, there is no case in which the result was Liable—and SHYSTER warns that this is so. Further, SHYSTER warns that the ideal point for Liable is at least as near to the instant case as is the nearest neighbour. � D.4 The EMPLOYEE specification 361 The nearest neighbours are Winstone v. Wurlitzer Automatic Phonograph Co. of Australia Pty Ltd and Australasian Performing Right Association Ltd v. Canterbury-Bankstown League Club Ltd ; the nearest other is RCA Corporation v. John Fairfax and Sons Ltd (Not-Auth). Apart from Winstone v. Wurlitzer, these were decided after APRA v. Miles. The choice of APRA v. Canterbury-Bankstown is a good one, despite the fact that its result is Auth. As discussed above, the judges in that case held that the club was either directly liable for, or had authorized, the infringement. (Without APRA v. Miles there is an equivalent functional dependence between A1 and A5, and each of those attributes—not just A5—has infinite weight.) D.4 The EMPLOYEE specification The results of the reflexive testing of all cases in the Employee specification are described here, and summarized in figure D.3. Zuijs v. Wirth Brothers Pty Ltd 1955 In Zuijs v. Wirth Brothers Pty Ltd 61 (C1), Zuijs cited several cases, notably Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd and Stevenson Jordon and Harrison Ltd v. Macdonald and Evans. In response, Wirth Brothers relied on Humberstone v. Northern Timber Mills. The Full Court of the High Court found for Zuijs: he was held to be an employee of Wirth Brothers. SHYSTER disagrees. The nearest neighbours are Price v. Grant Industries Pty Ltd and Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (1); the nearest other is Ferguson v. John Dawson & Partners (Contractors) Ltd. Apart from Stevenson v. Macdonald, these cases were decided after Zuijs v. Wirth Brothers. However, SHYSTER issues four safeguard warnings. It warns that the weighted association coefficients and the weighted correlation coefficients suggest that Stevenson v. Macdonald (2)—in which the result was Employee—should be the nearest neighbour. It also warns that the ideal point and the centroid for Employee are both at least as near to the instant case as is the nearest neighbour, and that the specified directions suggest Employee. Cam and Sons Pty Ltd v. Sargent 1940 In Cam and Sons Pty Ltd v. Sargent 62 (C2), four judges of the High Court agreed with Sargent who claimed that he was an employee of Cam. SHYSTER concludes, instead, that Cam was an independent contractor. The nearest neighbour is Humberstone v. Northern Timber Mills; the nearest others are Australian Timber Workers Union v. Monaro Sawmills Pty Ltd and Fer­ guson v. John Dawson & Partners (Contractors) Ltd. All three cases were de­ cided after Cam v. Sargent. 362 Appendix D: Reflexive tests SHYSTER warns that the weighted correlation coefficients suggest that Steven­ son Jordon and Harrison Ltd v. Macdonald and Evans (2)—in which the result was Employee—should be the nearest neighbour. Federal Commissioner of Taxation v. J. Walter Thompson (Australia) Pty Ltd 1944 In Federal Commissioner of Taxation v. J. Walter Thompson (Australia) Pty Ltd 63 (C3), Latham CJ of the High Court held that the radio artists were em­ ployees of J. Walter Thompson, and hence their fees were subject to payroll tax. SHYSTER agrees. The nearest neighbour is Ferguson v. John Dawson & Part­ ners (Contractors) Ltd ; the nearest other is Massey v. Crown Life Insurance Co., both of which were decided after FCT v. Thompson. In both instantiations, the result is the same as in the instant case. SHYSTER warns that the ideal point for Contractor is at least as near to the instant case as is the nearest neighbour, and that the specified directions suggest Contractor. (With FCT v. Thompson removed from the specification, there is an equivalent functional dependence between A11 and A12. Further, A13 has infinite weight.) Australian Timber Workers Union v. Monaro Sawmills Pty Ltd 1980 In Australian Timber Workers Union v. Monaro Sawmills Pty Ltd 64 (C4), Sween­ ey and Evatt JJ of the Federal Court applied the “organization test” as proposed by Denning LJ in Stevenson Jordon and Harrison Ltd v. Macdonald and Evans and Bank Voor Handel en Scheepvaart NV v. Slatford. 65 SHYSTER agrees with their honours and concludes that Wales (the worker) was an employee. However its chosen cases were not cited in the judgment: the nearest neighbour is Cam and Sons Pty Ltd v. Sargent; the nearest other is Humberstone v. Northern Timber Mills. SHYSTER reaches the same conclusion in both instantiations. SHYSTER warns that the centroid for Contractor is at least as near to the instant case as is the nearest neighbour. Ferguson v. John Dawson & Partners (Contractors) Ltd 1976 Megaw and Brown JJ in Ferguson v. John Dawson & Partners (Contractors) Ltd 66 (C5) held (Lawton LJ dissenting) that despite the label that the parties put on their relationship, it was in fact one of employer and employee. They based their decision on Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance and Market Investigations Ltd v. Minister of Social Security. 67 � D.4 The EMPLOYEE specification 363 SHYSTER agrees that Ferguson was an employee. The nearest neighbours are Zuijs v. Wirth Brothers Pty Ltd and Australian Timber Workers Union v. Monaro Sawmills Pty Ltd ; the nearest other is Massey v. Crown Life Insurance Co. Apart from Zuijs v. Wirth Brothers, these cases were decided after Ferguson v. Dawson. Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (2) 1952 As discussed in §5.4.2, Stevenson Jordon and Harrison Ltd v. Macdonald and Evans appears in the specification twice. In Stevenson v. Macdonald (2)68 (C6), the Court of Appeal held that Evans- Hemming was an employee of Macdonald and Evans when he wrote the second section of his book. Scant reference was made to previously decided cases. SHYSTER agrees with the Court that Evans-Hemming was an employee. The nearest neighbour is Zuijs v. Wirth Brothers Pty Ltd, which was decided after Stevenson v. Macdonald . The nearest other is Stevenson v. Macdonald (1). In 120 of the 128 instantiations, SHYSTER reaches the same conclusion. Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd 1924 Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd 69 (C7) is the oldest of the cases in the Employee specification. McCardie J applied the control test and held that the band members were employees of Palais de Danse. SHYSTER agrees. The nearest neighbours are Australian Timber Workers Union v. Monaro Sawmills Pty Ltd are Zuijs v. Wirth Brothers Pty Ltd ; the nearest others are Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (1) and Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance. All, of course, were decided after PRS v. Palais de Danse. (Without PRS v. Palais de Danse, A14 has infinite weight.) Humberstone v. Northern Timber Mills 1949 In Humberstone v. Northern Timber Mills70 (C8), three High Court judges fol­ lowed Queensland Stations Pty Ltd v. Federal Commissioner of Taxation and applied the control test.71 They concluded that Humberstone was not an em­ ployee of NTM. SHYSTER agrees that Humberstone was an independent contractor. The nearest neighbour is Price v. Grant Industries Pty Ltd, which was decided after Humberstone v. NTM ; the nearest other is Cam and Sons Pty Ltd v. Sargent. (With Humberstone v. NTM removed from the specification, there is an inverse functional dependence between A3 and A17.) 364 Appendix D: Reflexive tests Queensland Stations Pty Ltd v. Federal Commissioner of Taxation 1945 In Queensland Stations Pty Ltd v. Federal Commissioner of Taxation72 (C9), three judges of the High Court based their arguments on Logan v. Gilchrist, Watt and Cunningham, 73 another droving case. Latham CJ and Dixon J followed that case, and held that the drovers were independent contractors. SHYSTER agrees with their honours’ decision. The nearest neighbours are Humberstone v. Northern Timber Mills and Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (1), both of which were decided after Queensland Stations v. FCT ; the nearest other is Cam and Sons Pty Ltd v. Sargent. (With Queensland Stations v. FCT removed from the specification, there is an inverse functional dependence between A7 and A9.) Price v. Grant Industries Pty Ltd 1978 In Price v. Grant Industries Pty Ltd 74 (C10), Smithers, Evatt and Keely JJ of the Federal Court followed Humberstone v. Northern Timber Mills and held that Price was an independent contractor. SHYSTER agrees. The nearest neighbour is Humberstone v. NTM ; the nearest other is Cam and Sons Pty Ltd v. Sargent. However, SHYSTER warns that the weighted correlation coefficients suggest that Zuijs v. Wirth Brothers Pty Ltd —in which the result was Employee—should be the nearest neighbour. Australian Mutual Provident Society v. Chaplin 1978 In Australian Mutual Provident Society v. Chaplin75 (C11), the Judicial Commit­ tee of the Privy Council held that Chaplin’s power of unlimited delegation was almost conclusive against Chaplin being an employee.76 SHYSTER agrees that Chaplin was an independent contractor. The nearest neighbour is Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance; the nearest other is Cam and Sons Pty Ltd v. Sargent. The Privy Council mentioned neither of these cases. Massey v. Crown Life Insurance Co. 1978 In Massey v. Crown Life Insurance Co.77 (C12), the English Court of Appeal unanimously distinguished Ferguson v. John Dawson & Partners (Contractors) Ltd, 78 and Lord Denning MR referred to Stevenson Jordon and Harrison Ltd v. Macdonald and Evans, before holding that Massey was an independent con­ tractor. SHYSTER disagrees, however its choice of cases is good. It chooses both Ferguson v. Dawson and Stevenson v. Macdonald (1): the former is the nearest neighbour, which the latter should be. Australian Mutual Provident Society v. 365 � D.4 The EMPLOYEE specification Nearest Nearest Result Case n’bours others Warnings Zuijs v. Wirth Brothers Cam v. Sargent FCT v. Thompson ATWU v. Monaro Sawmills Ferguson v. Dawson Stevenson v. Macdonald (2) PRS v. Palais de Danse Humberstone v. NTM Queensland Stations v. FCT Price v. Grant Industries AMP v. Chaplin Massey v. Crown Life Stevenson v. Macdonald (1) Ready Mixed v. Minister × × � � � – – – × – – – – × – S� r� I µ r� I ⊃ ∨0 2 µ ∨0 2 � � × × � � � – – – ∨ 8 128 × � � � � � × × � – – � × � – × × × × � × × r� ⊃ r� ∨48 64 � � × � � � � Figure D.3: A summary of the reflexive testing of the Employee specification, described in §D.4. The meaning of the symbols used here is explained in §5.6.1. Chaplin, which was decided in the same year as was Massey v. Crown Life, is also chosen as a nearest other. SHYSTER warns that the specified directions suggest Contractor. (With Massey v. Crown Life removed from the specification, there is an in­ verse functional dependence between A9 and A11, and A16 has infinite weight.) Stevenson Jordon and Harrison Ltd v. Macdonald and Evans (1) As discussed in §5.4.2, Stevenson Jordon and Harrison Ltd v. Macdonald and Evans appears in the specification twice. In Stevenson v. Macdonald (1)79 (C13), the Court of Appeal held that Evans- Hemming was not an employee of Macdonald and Evans when he wrote the first section of his book. The court made very few references to previously decided cases in making its decision. SHYSTER disagrees with the Court. The nearest neighbour is Stevenson Jor­ don and Harrison Ltd v. Macdonald and Evans (2)—an unsurprising choice. The nearest other is Queensland Stations Pty Ltd v. Federal Commissioner of Taxa­ tion. 1952 366 Appendix D: Reflexive tests However, SHYSTER warns that the weighted correlation coefficients suggest that Queensland Stations v. FCT should be the nearest neighbour. And in 48 of the 64 instantiations, the nearest result is Contractor. Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance 1967 In Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance80 (C14), held that the contract between Latimer and Ready Mixed was a contract of carriage and not a contract of service.81 SHYSTER agrees that Latimer was an independent contractor. The nearest neighbour is Australian Mutual Provident Society v. Chaplin, which was decided after Ready Mixed v. Minister ; the nearest other is Cam and Sons Pty Ltd v. Sargent, which was not referred to in argument or in MacKenna J’s judgment. D.5 The NATURAL specification Reflexive testing of the Natural specification was limited to those cases in the Natural area—the principal area in that specification. The results of those reflexive tests are described here, and summarized in figure D.4. FAI Insurances Ltd v. Winneke 1982 In FAI Insurances Ltd v. Winneke82 (C1), Gibbs CJ, Stephen, Mason, Aickin, Wilson and Brennan JJ (Murphy J dissenting) held that FAI had a legitimate expectation that its approval would be renewed. Hence the Governor in Council was required to apply principles of natural justice before deciding not to renew that approval. SHYSTER agrees that a duty to observe natural justice is implied. The nearest neighbour is Haoucher v. Minister of State for Immigration and Ethnic Affairs, which was decided after FAI v. Winneke; the nearest other is Nashua Australia Pty Ltd v. Channon, which was not referred to by any of the judges. Haoucher v. Minister of State for Immigration and Ethnic Affairs 1990 In Haoucher v. Minister of State for Immigration and Ethnic Affairs83 (C2), the majority of the High Court84 applied several cases, including FAI Insurances Ltd v. Winneke, 85 in concluding that Haoucher had been entitled to make rep­ resentations to the Minister before the Minister ordered his deportation. SHYSTER agrees that a duty to observe natural justice is implied. The nearest neighbour is FAI v. Winneke; the nearest other is Nashua Australia Pty Ltd v. Channon, to which no reference was made. � D.5 The NATURAL specification 367 Annetts v. McCann 1990 In Annetts v. McCann86 (C3), five judges of the High Court unanimously held that the coroner ought to have accorded natural justice to the parents of the boy who was the subject of the inquest.87 SHYSTER disagrees, and concludes that a duty to observe natural justice is not implied. The nearest neighbour is McInnes v. Onslow Fane to which no reference was made; the nearest other is Kioa v. West, to which all but one of their honours referred in their judgments.88 SHYSTER warns that the ideal point for Implied is at least as near to the instant case as is the nearest neighbour, and that the specified directions suggest Implied. Kioa v. West 1985 In Kioa v. West 89 (C4), Mason, Wilson, Brennan and Deane JJ (Gibbs CJ dis­ senting) held that natural justice ought to have been observed in the making of the deportation order. SHYSTER agrees that a duty to observe natural justice is implied. The nearest neighbours are FAI Insurances Ltd v. Winneke, Haoucher v. Minister of State for Immigration and Ethnic Affairs and Attorney-General of Hong Kong v. Ng Yuen Shiu; the nearest other is South Australia v. O’Shea. Gibbs CJ, Mason and Brennan JJ made reference to FAI v. Winneke and AG of Hong Kong v. Shiu. 90 The other two of SHYSTER’s chosen cases were decided after Kioa v. West. Commissioner of Police v. Tanos 1958 Commissioner of Police v. Tanos91 (C5) is the oldest of all the natural justice cases. Dixon CJ, Webb and Taylor JJ held that Tanos ought to have been accorded natural justice before her restaurant was declared a “disorderly house.” SHYSTER agrees that a duty to observe natural justice is implied. The nearest neighbour is Marine Hull & Liability Insurance Co. Ltd v. Hurford ; the nearest other is McInnes v. Onslow Fane. Of course, both of these cases were decided after Commissioner of Police v. Tanos. Marine Hull & Liability Insurance Co. Ltd v. Hurford 1986 In Marine Hull & Liability Insurance Co. Ltd v. Hurford 92 (C6), Fox, Davies and Morling JJ held that natural justice was implied, but that Marine Hull had not been denied it in this case. SHYSTER also concludes that a duty to observe natural justice is implied. The nearest neighbour is Commissioner of Police v. Tanos, to which Morling J referred;93 the nearest other is South Australia v. O’Shea, which was decided after Marine Hull v. Hurford. 368 Appendix D: Reflexive tests Macrae v. Attorney-General for New South Wales 1987 In Macrae v. Attorney-General for New South Wales 94 (C7), Kirby P, Mahoney and Priestly JJA held that the magistrates were denied their legitimate expect­ ation of procedural fairness. The legal expert advises that this was a somewhat surprising outcome. SHYSTER disagrees with their honours. It concludes that a duty to observe natural justice is not implied. The nearest neighbour is South Australia v. O’Shea which was decided in the same year as was Macrae v. AG for NSW ; the nearest other is Durayappah v. Fernando to which no reference was made. SHYSTER warns that both the weighted coefficients prefer Attorney-General of Hong Kong v. Ng Yuen Shiu (in which natural justice was implied) as the nearest neighbour. It also warns that the specified directions suggest Implied. Attorney-General of Hong Kong v. Ng Yuen Shiu 1983 In Attorney-General of Hong Kong v. Ng Yuen Shiu95 (C8), the Judicial Commit­ tee of the Privy Council held that Shiu ought to have been given an opportunity to make a representation to the Minister. SHYSTER disagrees. The nearest neighbour is South Australia v. O’Shea; the nearest others are Kioa v. West and Macrae v. Attorney-General for New South Wales. However, SHYSTER warns that all three chosen cases are equidistant from AG of Hong Kong v. Shiu; SA v. O’Shea and Kioa v. West are both decisions of five judges of the High Court, but SHYSTER prefers the former because it is two years more recent. (Macrae v. AG for NSW is merely a decision of the New South Wales Court of Appeal.) SHYSTER also warns that both the weighted coefficients prefer SA v. O’Shea and Kioa v. West as the nearest neighbours, that the centroid for Implied is at least as near to the instant case as is the nearest neighbour, and that the specified directions suggest Implied. In one of the two instantiations, the result was Implied. Kioa v. West was not referred to in any of the judgments, or in argument, and the other two chosen cases were decided after AG of Hong Kong v. Shiu. Durayappah v. Fernando 1967 In Durayappah v. Fernando96 (C9), the Judicial Committee of the Privy Council held that the Minister ought to have applied natural justice principles before dissolving the Jaffna Municipal Council. Unlike the Privy Council, SHYSTER concludes that a duty to observe natural justice is not implied. The nearest neighbour is Nashua Australia Pty Ltd v. Channon; the nearest others are FAI Insurances Ltd v. Winneke and Haoucher v. Minister of State for Immigration and Ethnic Affairs. All of these cases were decided after Durayappah v. Fernando. SHYSTER warns that the specified directions suggest Implied. � D.5 The NATURAL specification 369 South Australia v. O’Shea 1987 In South Australia v. O’Shea97 (C10), Mason CJ, Wilson, Brennan and Toohey JJ (Deane J dissenting) held that O’Shea was not entitled to a further hearing. SHYSTER disagrees: it concludes that a duty to observe natural justice is implied. The nearest neighbour is Macrae v. Attorney-General for New South Wales, which was decided in the same year as was SA v. O’Shea; the nearest other is Nashua Australia Pty Ltd v. Channon, to which no reference was made. Bread Manufacturers of New South Wales v. Evans 1981 In Bread Manufacturers of New South Wales v. Evans 98 (C11), Gibbs CJ, Mason, Murphy, Aickin and Wilson JJ held that the Prices Commission was not obliged to apply natural justice principles. SHYSTER agrees. The nearest neighbours are Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd and Nashua Australia Pty Ltd v. Channon, which were decided after Bread Manufacturers v. Evans; the nearest other is Durayappah v. Fernando to which none of their honours referred. However, SHYSTER warns that all three chosen cases are equidistant from Bread Manufacturers v. Evans; SHYSTER prefers both Minister for Environ­ ment v. Peko-Wallsend and Nashua v. Channon (decisions of the Full Court of the Federal Court and the New South Wales Supreme Court, respectively) to Durayappah v. Fernando (a Privy Council decision). SHYSTER warns that the weighted association coefficients prefer Durayap­ pah v. Fernando as the nearest neighbour, and that the weighted correlation coefficients prefer Annetts v. McCann (in which natural justice was implied). It also warns that the specified directions suggest Implied. Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd 1987 In Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd 99 (C12), Wilcox J (with whom Bowen CJ and Sheppard J generally agreed) held that Federal Cabinet was not obliged to apply natural justice principles. SHYSTER agrees that a duty to observe natural justice is not implied. The nearest neighbour is Bread Manufacturers of New South Wales v. Evans ; the nearest other is Durayappah v. Fernando. Neither case was mentioned in the Federal Court’s judgment. However, SHYSTER warns that the weighted correlation coefficients suggest that twelve other cases should be the nearest neighbour—in eight of which, nat­ ural justice was implied. 370 Appendix D: Reflexive tests Nashua Australia Pty Ltd v. Channon 1981 In Nashua Australia Pty Ltd v. Channon1 (C13), Lee J held that the Minister was not required to afford Nashua natural justice before making his determination. SHYSTER disagrees. The nearest neighbour is Durayappah v. Fernando, which Lee J applied;2 the nearest other is South Australia v. O’Shea, which was decided after Nashua v. Channon. However, it warns that the centroid for Not-Implied is at least as near to the instant case as is the nearest neighbour. Council of Civil Service Unions v. Minister for the Civil Service 1984 In Council of Civil Service Unions v. Minister for the Civil Service3 (C14), the House of Lords held that the Minister was not obliged to accord the unions natural justice. SHYSTER agrees. The nearest neighbour is Nashua Australia Pty Ltd v. Chan­ non; the nearest other is Durayappah v. Fernando. Neither case was mentioned in the law lords’ judgment. However, SHYSTER warns that these two cases are equidistant from CCSU v. Minister for the Civil Service; it chooses Nashua v. Channon because it is a de­ cision of the Supreme Court of New South Wales, while Durayappah v. Fernando is a Privy Council decision. It also warns that the specified directions suggest Implied. McInnes v. Onslow Fane 1978 In McInnes v. Onslow Fane4 (C15), Megarry V-C held that the British Boxing Board of Control had no duty to afford McInnes natural justice. SHYSTER comes to the opposite conclusion. The nearest neighbour is An- netts v. McCann; the nearest other is South Australia v. O’Shea. Both cases were decided after McInnes v. Onslow Fane. � � � � � � � � � � � � � � � � � � µ � � 371 C a se F A I v. W in n ek e H a o u ch er v . M in is te r fo r Im m ig ra ti o n A n n et ts v . M cC a n n K io a v . W es t C o m m is si o n er o f P o li ce v . T a n o s M a ri n e H u ll v . H u rf o rd M a cr a e v. A G f o r N S W A G o f H o n g K o n g v. S h iu D u ra y a p pa h v . F er n a n d o S A v . O ’S h ea B re a d M a n u fa ct u re rs v . E va n s M in is te r fo r E n vi ro n m en t v. P ek o -W a ll se n d N a sh u a v . C h a n n o n C C S U v . M in is te r fo r th e C iv il S er vi ce M cI n n es v . O n sl o w F a n e R e su lt × × × × × × × N ea re st n ’b o u r s – × – – – – – – × × – N ea re st o th e r s × × – – – × – – × × × – × – W a rn in gs I ⊃ � S � r � ⊃ � S � r � µ ⊃ → ∨ 1 2 ⊃ � × S � r � ⊃ → × r � × ⊃ → × × � D.5 The NATURAL specification F ig u re D .4 : A s u m m ar y of t h e re fl ex iv e te st in g of t h e N a t u r a l s p ec ifi ca ti on ( N a tu ra l ar ea o n ly ), d es cr ib ed i n § D .5 . T h e m ea n in g of t h e sy m b ol s u se d h er e is e xp la in ed i n § 5. 6. 1. Notes Cases Statutes Bibliography 373 Notes 1. at 2110. 2. Second edition, vol. XV at 403. 3. Rawson 1991 at 355–6. Chapter 1: Introduction 4. Wells and Taylor 1986, 4.2.78/2238. Original title: The First Part of the Con­ tention of the Two Famous Houses of Yorke and Lancaster. Dated by Wells and Taylor 1987 at 111. 5. Quoted by Muir 1976 at 81. Muir does not date this quotation but, according to Twiss 1844 at 16, Lord Eldon finished his anecdote book in 1827. 6. Plater 1985 at 57. Chapter 2: Legal analysis systems 7. Leith 1986b at 97. 8. Moles 1987 at 271. 9. Montesquieu 1973, liv. VI, ch. II at 84, footnote omitted. Translation by Peter Brown, Lecturer, Department of Modern European Languages, Faculty of Arts, The Australian National University. 10. McCarty 1980a identifies a third category of legal AI systems: “integrated legal systems.” He cites, as an example, computerized title registration systems which make decisions about people’s rights and obligations (at 2). It is hard to see why such a system would not be better classified as a legal analysis system, albeit with some of the features of a legal retrieval system. 11. On legal retrieval systems see Tapper 1963, 1984; Bing and Selmer 1980, chs 8–15; Bing 1984a, 1984b, 1987, 1989; Croydon 1980; Dick 1991. For an example of early work in the field see Allen, Brooks and James 1962. Sprowl 1979 describes a doc­ 375 376 [pages 8–13] Notes to chapter 2 ument preparation system (see also Cook, Hafner, McCarty, Meldman, Peterson, Sprowl, Sridharan and Waterman 1981). For a detailed summary of Australian legal retrieval systems (CLIRS, SCALE, etc.) see Greenleaf, Mowbray and Lewis 1988. 12. Mehl 1959 at 759. 13. Shannon and Golshani 1988 at 306. 14. Susskind 1987a at 3. 15. This is, of course, a variation on the famous test for machine intelligence proposed by Turing 1950. 16. For example, HYPO (§2.5.3) and the Latent Damage System (see §2.4.4). Apropos SHYSTER and legal expertise, see §3.4. 17. The doctrine of parliamentary supremacy over the courts developed in England in the seventeenth century: see Morris, Cook, Creyke and Geddes 1992 at 8. 18. Morris et al. 1992 at 31. 19. ibid. at 40. 20. Cross 1961 at 4. 21. See Stone 1964a, 1985; Harris 1980; Krygier 1986. See also the discussion in §2.2.5. 22. Stone 1964a at 325. 23. In Queensland v. Commonwealth of Australia (1977) 139 CLR 585, Gibbs J (at 601) and Stephen J (at 604) of the High Court of Australia felt bound by the principle of stare decisis to follow Western Australia v. Commonwealth of Australia (1975) 134 CLR 201—a decision in which they had been in the minority—despite still believing that that case had been wrongly decided. 24. Bain’s JUDGE system, although strictly speaking a rule-based system, adopts an approach which is similar to that of a case-based system. It analyzes the similarities and differences between cases to give advice about sentencing for certain crimes including murder, assault and manslaughter. JUDGE starts with a rule about a crime (provided by the user) and an appropriate sentence taken from a case. When a new case is added, JUDGE finds similar crimes and generates a rule about the new case on the basis of the similarities and differences between the new case and those similar crimes. 25. Thomas 1979; Campbell 1984; Boden 1985; Willick 1985; Clarke 1988; Capper and Susskind 1988 at 101–38. 26. Susskind 1987a at 19. 27. Niblett 1981 at 3. 28. Susskind 1987a at 20, emphasis in the original. 29. In 1987, only four of the twelve law schools in Australia had compulsory courses in jurisprudence: Pearce, Campbell and Harding 1987 at 106–7. 30. Susskind 1987a at 26. 31. Harris 1980 at 2. 32. Niblett 1988 at 33. 33. Susskind 1987a at 20, as quoted above. 34. The Oxford English Dictionary 1989, second edition, vol. VIII at 321. 35. Pound 1908 at 605. 36. Loevinger 1949 at 472, citing Frank 1930. 37. Pound 1908 at 607. Notes to chapter 2 [pages 13–18] 377 38. ibid. at 605. 39. Loevinger 1949 at 483, footnote omitted. See also Loevinger 1963. 40. Loevinger 1963 at 35. 41. Lasswell 1955 at 398. As Gardner 1987 comments, “[w]hy one robot was not enough was not discussed; the idea may have been that hard cases would be decided by a vote of the random number generators the robots were to incorporate” (at 81–2). 42. Lasswell 1955 at 398. 43. Frank 1949 at 207. 44. Mehl 1959 at 758. 45. D’Amato 1977 at 1280. 46. ibid. at 1290–1. 47. Weizenbaum 1976 at 226–7. 48. D’Amato 1977 at 1281. 49. Stone 1964b at 555. 50. This increased level of certainty would be a result of freely available (automated) judicial advisory opinions. 51. D’Amato 1977 at 1278. 52. ibid. at 1279. 53. ibid. at 1290. 54. Kayton 1964 at 302. 55. Schubert 1968 at 63, emphasis omitted. 56. Tyree 1980. 57. Pound 1908 at 606. 58. Stone 1964b at 554. See also Spengler 1963: an earlier cautionary article about judgment machines. 59. Stone 1964b at 558, emphasis omitted. 60. ibid. at 559. 61. ibid. at 560. 62. Hart 1983 at 106. 63. Hart 1961 at 132. As discussed in §2.2.6, a true rule sceptic would agree with Hart that the rules embodied in cases are “as determinate as any statutory rule”—that is, not at all. 64. Susskind 1986 at 189–90. 65. Hart 1983 at 106. 66. Section 13(2) of the Domestic Violence and Matrimonial Proceedings Act 1976 (UK). See Moles 1987 at 142–57. 67. [1978] 1 All ER 821. 68. Cantliff v. Jenkins [1978] 1 All ER 836. 69. Davis v. Johnson [1978] 1 All ER 841. 70. Moles 1987 at 155, footnotes added. Compare Moles’s analysis with that of MacCormick 1978. In order to demon­ strate the possibility of purely deductive legal justification he provides a detailed analysis of the judgment of Lewis J in Daniels and Daniels v. R. White and Sons and Tarbard [1938] 4 All ER 258. He argues that that judgment is an example of deductive reasoning—a series of applications of modus ponens: p ⊕ q ; p ; � q. 71. Moles 1987 at 172. 378 [pages 18–24] Notes to chapter 2 72. id. 73. Leith’s system was called ELI (see also §3.13.1). 74. Leith 1985b at 9. See also Leith 1985a, 1986a. 75. See Llewellyn 1931, 1951, 1962. See also Holmes 1897. 76. Drahos and Parker 1992 at 110, referring to Llewellyn 1931 at 1237. 77. Kripke 1982 bases his argument on Wittgenstein’s Philosophical Investigations (1974), but Drahos and Parker 1992 cite devotees of Wittgenstein who deny that he was the source of Kripke’s argument or conclusion (at 111). 78. Kripke 1982 at 8–9. 79. ibid. at 13, 97. 80. Drahos and Parker 1992 at 112. 81. For example Yablon 1987 who agrees with Kripke, and Bjarup 1988 who does not. 82. Drahos and Parker 1992 at 113. It also puts paid to Susskind’s claim of a consensus in legal theory (see §2.2.7): Hart’s and Kripke’s viewpoints are totally irreconcil­ able. 83. Drahos and Parker 1992 at 112. 84. ibid. at 114. 85. id. 86. ibid. at 115. 87. ibid. at 114. 88. Susskind 1987a. 89. ibid. at 27–8. See also Susskind 1987b at 2. 90. Susskind 1987a at vii. 91. ibid. at 27. 92. Susskind’s development work, as opposed to his jurisprudential work, is discussed in §2.4.4. 93. Moles 1991 at 153–4. 94. Susskind 1987b at 2. 95. ibid. at 1. 96. ibid. at 2. Unaccountably, Moles 1991 does not make this point in his criticism of Susskind. 97. Clark 1988 at 429. 98. Kelsen 1945, 1967; Hart 1961, 1983; Dworkin 1977, 1978. 99. Leith 1987 at 131–2, emphases omitted, footnote added. 1. Loevinger 1961 at 259. 2. ibid. at 266, 269–70, 274. 3. Loevinger 1949 at 493. 4. Gardner 1987 at 69. Note, however, that the Jurimetrics Journal (ISSN 0022-6793 and ISSN 0897-1277) is published quarterly, and has been published since 1966. 5. As well as the references made in this section, see Tanenhaus, Schick, Muraskin and Rosen 1963. 6. Lawlor 1963 at 341, footnote omitted. 7. Kort 1963a, 1963b. 8. Kort 1963b at 145. 9. Lawlor 1963, 1968, 1981. 10. (1942) 316 US 455. Notes to chapter 2 [pages 24–28] 379 11. (1963) 372 US 335. 12. Wiener 1962 at 1023. 13. ibid. at 1024. 14. Kayton 1964 at 312, some emphasis omitted, some emphasis added. 15. Nagel 1960, 1963, 1986, 1989; Schubert 1963a, 1963b. See also Spaeth 1963 and Aubert 1963. 16. Nagel 1963 at 40, emphases and footnotes omitted. 17. Haar, Sawyer and Cummings 1977. 18. ibid. at 758. 19. id. 20. See Gardner’s comments quoted in §3.14. Note, however, that Gardner is less critical of Haar et al. because “[b]y reflecting on their statistical results in the light of a traditional legal analysis of the same cases, the authors are able to provide recommendations for attorneys and litigants in zoning cases and to raise questions about zoning law for the attention of both court and legislature” (Gardner 1987 at 76, referring especially to Haar et al. 1977 at 742–50). 21. Loevinger 1949 at 461 citing Holmes 1897. 22. Loevinger 1961 at 269. 23. Lawlor 1963 at 339. 24. Susskind 1987a at 95–6. 25. Tapper 1973 at 249, footnote omitted. 26. ibid. at 250. 27. Gardner 1987 at 196 (a note to p. 75). Gardner cites as examples Schubert 1975 and Goldman and Sarat 1978, though her criticism clearly applies to Nagel’s work too. 28. Stone 1964b at 518. See also Stone 1966 at 688. Stone refers to Schubert 1963b and Ulmer 1963 (see also Ulmer 1964). 29. Stone 1964b at 550, emphases in the original. 30. ibid. at 553, emphasis in the original. 31. Buchanan and Headrick 1970 at 40. 32. ibid. at 41. 33. ibid. at 45. 34. id. 35. Maggs and deBessonet 1972. 36. Popp and Schlink 1975. 37. Buchanan and Shortliffe 1984. 38. Popp and Schlink 1975 at 309. 39. Michaelsen and Michie 1983. 40. More precisely, EMYCIN: MYCIN with its domain knowledge removed. 41. The domain of taxation law had been widely used in rule-based expert systems research. Apart from McCarty’s TAXMAN project (§2.4.1), see Torsun 1987 and Sherman 1987, 1989. See also the hybrid CABARET system (§2.6). 42. Meldman 1977. See also Cook et al. 1981. 43. Cook et al. 1981 at 692. 44. Waterman and Peterson 1980. See also Cook et al. 1981 and Waterman, Paul and Peterson 1987. 380 [pages 28–32] Notes to chapter 2 45. Waterman et al. 1987 at 26. 46. ibid. at 27–8. These “informal rules” are similar to the “conventions” proposed by Drahos and Parker 1992, as discussed in §2.2.6. 47. Waterman et al. 1987 at 28. 48. Bing 1980; Borchgrevink and Hansen 1980. 49. Bing 1980 at 132. 50. Stamper 1980, 1982. See also Cook et al. 1981. Stamper later decided that LEGOL was inadequate, and developed NORMA (1986). 51. Pattison and Ciesielski 1990. See also Ciesielski 1990. 52. Johnson and Mead 1991. 53. Susskind 1989a at 29. 54. McCarty 1977, 1980a, 1980b, 1980c. 55. McCarty 1980a at 3. 56. id. 57. id. 58. ibid. at 4. 59. id. 60. id. 61. Moles 1987 at 270. 62. McCarty 1980a at 4–5. 63. ibid. at 5. 64. McCarty 1977 at 839. 65. ibid. at 839–40. 66. Austin 1885. 67. Moles 1987 at 270–1, footnote added. 68. ibid. at 269. 69. McCarty 1980c at 29–30. 70. McCarty and Sridharan 1981. 71. Eisner v. Macomber (1920) 252 US 189 (see McCarty 1982). 72. Ashley 1990 at 224–5, footnote added, extraneous punctuation removed. Ashley’s (i.e. HYPO’s) mechanism for determining how “on point” are two cases is discussed in §3.9.3. 73. Moles 1987 at 270, referring to the structure of TAXMAN I. 74. Ashley 1990 at 226–7. Ashley criticizes Goldman, Dyer and Flowers 1987 for a similar use of primitives in their STARE system. 75. Kowalski and Sergot 1985, 1990; Sergot, Cory, Hammond, Kowalski, Kriwaczek and Sadri 1986a, 1986b; Bench-Capon, Robinson, Routen and Sergot 1987; Bench- Capon and Sergot 1988; Kowalski 1989; Bench-Capon 1991. 76. Sergot et al. 1986a at 49. Note that this statement was qualified in another paper by the same authors published in the same year: “The knowledge elicitation problem is almost entirely absent in the formalization of legislation” (Sergot et al. 1986b at 383, emphasis added). 77. Moles 1991 at 144. 78. Sergot et al. 1986a, 1986b; Kowalski and Sergot 1985; Bench-Capon and Sergot 1988; Kowalski 1989. 79. Sergot et al. 1986a at 41. Notes to chapter 2 [pages 32–37] 381 80. Bench-Capon et al. 1987 at 192. 81. Moles 1991 at 159. He also refers to Bench-Capon 1991 (a collection of articles by Bench-Capon, Sergot, and others) and notes that, “[a]lthough the book is called Knowledge-Based Systems and Legal Applications, it would appear from [the bio­ graphical notes] that there was very little emphasis on involving people with legal expertise” (at 146). Moles’s article sparked a debate: see Tyree 1992 and Moles and Dayal 1992. 82. Kowalski and Sergot 1990 at 207. 83. Hayes-Roth, Waterman and Lenat 1983 at 165, quoted by Susskind 1986 at 176. 84. Sergot et al. 1986a at 46. 85. Leith 1986b at 97, some extraneous punctuation removed. Leith’s remarks are particularly interesting given Susskind’s criticism of Leith for having built a legal expert system without legal expertise (see §3.13.1). 86. Gardner 1983, 1984, 1985, 1987. 87. Gardner 1987 at 4. 88. ibid. at 6. 89. Susskind 1990 at 222. 90. Gardner 1987 at 120. 91. ibid. at 119. 92. Susskind 1990 at 225. Grunbaum 1988 is similarly critical of Gardner’s “cour­ ageous” attempt (at 639). Jones 1988 also doubts “[t]he utility of the system outside this domain” (at 32). 93. Ashley 1990 at 226. 94. See Levi 1949. 95. Gardner 1987 at 22, footnote added. 96. id. 97. id. 98. ibid. at 33. 99. Susskind 1989a at 30. 1. Gold and Susskind 1986; Susskind 1986, 1987a, 1987b. 2. Capper and Susskind 1988; Susskind 1989b. 3. Capper and Susskind 1988 at vii. See also Susskind 1989b at 24. 4. Susskind 1987a at 255. 5. ibid. at 87, citing Simpson 1986. 6. Susskind 1987a at 256. 7. ibid. at 87. 8. Feigenbaum 1981 at 226. See also Feigenbaum and McCorduck 1983 at 75. 9. Sergot et al. 1986a at 49. 10. Allen 1963, 1965, 1968, 1981, 1982; Allen and Caldwell 1963; Allen and Saxon 1985, 1988. See also Bench-Capon 1987. 11. Shannon and Golshani 1988 at 308. They also warn that ad hoc rule formulation means that subsequent designers/users cannot trace the evolution of a set of rules from the words of the statute. However, the author suggests that this could be obviated by the sensible use of comments. 12. ibid. at 308–9. 382 [pages 37–41] Notes to chapter 2 13. However, Dworkin 1978, for example, contends that legal problems do have a cor­ rect answer. 14. deBessonet and Cross 1985, 1986, 1988. 15. Shannon and Golshani 1988 at 310, footnotes omitted, one footnote added. An extraneous right parenthesis has been removed from the ANF representation. 16. This criticism of Shannon and Golshani’s modified example does not apply to the original example given by deBessonet and Cross 1985. They refer, in much less ambiguous terms, to the lessee’s belief about damage caused by her/himself (at 208). 17. Tyree, Greenleaf and Mowbray 1988 at 232. 18. id. 19. Stone 1985 at 56. See also Tyree 1989 at 136. 20. Tyree et al. 1988 at 232. 21. Kowalski 1991 at 30. 22. Tyree et al. 1988 at 232. Quinlan, Compton, Horn and Lazarus 1987 give two examples of the use of an inductive technique. The first involves 2647 cases, from which a training set of 2000 and a test set of 647 are extracted (at 165–6). The second involves 3066 cases: a training set of 2300 and a test set of 766 (at 168–70). These examples use an inductive inference tool called C4: a descendent of ID3. See also Quinlan 1986, 1988. 23. Susskind 1986 at 190. 24. Sergot et al. 1986b at 370. 25. Pearce and Geddes 1988 at 56–62. 26. Bankowski and MacCormick 1991 at 377. 27. Pearce and Geddes 1988 at 1. 28. Ashley 1990 at 127. 29. Lambert and Grunewald 1991 at 194. 30. Ashley 1985 at 116. See Ashley and Aleven 1991 for a description of their research into developing an intelligent tutoring system for teaching law students to argue with cases. 31. Mackaay and Robillard 1974. They did not actually develop an expert system, but their work is relevant to the development of legal expert systems and, hence, is discussed here. 32. The first formulation of the nearest neighbour rule appears to have been made by Fix and Hodges in 1951. 33. Cover and Hart 1967 at 21. 34. Lawlor reported his results in a paper presented to the 1971 Annual Meeting of the American Political Science Association. Excerpts from that paper, published in 1972, omit details of his work with these cases. Lawlor’s approach was to find a weight for each of the attributes such that the sum of those weights, where the attribute is present, is less than 1 if the case was decided against a given party, and greater than 1 otherwise. 35. Mackaay and Robillard 1974 at 308. 36. See Schmidt 1971 and Claudy 1972. Notes to chapter 2 [pages 41–46] 383 37. Mackaay and Robillard 1974 at 322. They also examine the visual representation of cases, scaling the cases onto one- and two-dimensional maps using the same similarity metric as they use for their nearest neighbour analysis. These maps allow easy visual identification of the borderline cases. 38. Tyree 1985, 1986, 1989 and Tyree, Greenleaf and Mowbray 1988, 1989. FINDER’s approach to case law is based upon a mathematical concept of the similarity of cases described by Tyree 1977. 39. Treating each yes as a 1, and each no as a 0. 40. Tyree 1989 at 12–13. 41. The complete Finder specification is given in §A.2. 42. Ashley 1985, 1986, 1989a, 1989b, 1990; Rissland 1983, 1985, 1990; Rissland, Val­ carce and Ashley 1984; Ashley and Rissland 1987a, 1987b, 1988; Rissland and Ashley 1987, 1989. 43. Ashley 1990 at 2–3. 44. ibid. at 37. 45. id. 46. For example, a finding that there has been trade secrets misappropriation. 47. Ashley 1990 at 112–3. 48. ibid. at 279. “On pointness” in HYPO is discussed in §3.9.3. 49. ibid. at 20. 50. ibid. at 142. 51. id. 52. See Rissland 1989 for a discussion of argument with hypotheticals. 53. Hafner 1981. See also Cook et al. 1981. Negotiable instruments are cheques and promissory notes. 54. GREBE is described in §2.6. 55. Branting 1989 at 109. 56. id. 57. Rissland and Skalak 1989a at 48. 58. Rissland and Skalak 1989a, 1989b; Skalak 1989; Skalak and Rissland 1991, 1992. See also Rissland 1990. 59. Rissland and Skalak 1989b at 527. 60. Branting 1989, 1991. 61. See §2.5.4 for a discussion of the inadequacy of semantic networks for representing case law. 62. Branting 1991 at 150. 63. Oskamp, Walker, Schrickx and van den Berg 1989; Walker, Zeinstra and van den Berg 1989. 64. van Opdorp, Walker, Schrickx, Groendijk and van den Berg 1991 at 280. 65. Oskamp 1989; Oskamp et al. 1989; Walker et al. 1989; Wolstenholme 1989; van Opdorp et al. 1991. 66. van Opdorp et al. 1991 at 280. 67. ibid. at 285. 68. Berman and Hafner 1991; Vossos, Zeleznikow, Dillon and Vossos 1991. See also Sanders 1991 (a hybrid planner). 69. Susskind 1987a at 151. 384 [pages 46–57] Notes to chapter 3 70. ibid. at 153. 71. Hafner 1981, 1987; deBessonet and Cross 1985, 1986, 1988. 72. McCarty 1984, 1986, 1989, 1990, 1991. 73. McCarty 1984 at 126. 74. Greenleaf, Mowbray and Tyree 1987 at 11. 75. McCarty 1983 at 267–8. See also Bench-Capon 1989. 76. Shannon and Golshani 1988 at 311. They propose a “happy medium” between shallow and deep conceptual models. 77. Greenleaf et al. 1987 at 11. 78. Susskind 1987a at 153. 79. Meta-rules are similar to the “conventions” proposed by Drahos and Parker 1992, as discussed in §2.2.6. 80. Greenleaf et al. 1987 at 11. 81. Stone 1985 at 46, footnotes omitted, emphases in the original. 82. McCarty 1989 at 180. 83. Moles 1991 at 163. 84. Susskind 1987a at 20. 85. Stone 1964b at 560, as quoted in §2.2.3. 86. Susskind 1987a at 20. 87. Moles 1987 at 7. 88. See §2.4.2. Chapter 3: A pragmatic approach to case law 89. Galsworthy 1941, act II at 36. 90. Murphy 1980 at 5. Opening address delivered at the First National Conference of Labor Lawyers, Adelaide, 29 June 1979. 91. Susskind 1987a at 12. 92. The difference between private and public law is explained in §3.3.1. 93. For example, Susskind 1987a at 255. 94. See Stone 1964a at 120; Kelman 1987; Bottomley, Gunningham and Parker 1991. 95. For example, Kelsen 1945, 1967, and the authors referred to in the previous note. 96. See §2.1.2. 97. Susskind 1987a at 41. 98. id. Susskind also claims that law reformers use legal reasoning to try to persuade the legislature to change the law. This is debatable. In a sense, law reformers operate outside the realm of legal reasoning; their concern is that the application of legal reasoning to the existing law produces undesirable results. 99. id. 1. ibid. at 42. 2. See §2.2.7. 3. See Morris, Cook, Creyke and Geddes 1992, especially chs 3–4, 8. 4. Hart 1961 at 131. 5. Ashley 1990 at 229, citing Burton 1985 at 40. Notes to chapter 3 [pages 57–61] 385 6. Ashley 1990 at 230, capitalization added. 7. See §2.4.2, and the beginning of §2.5. 8. BA (Hons); LLB; Barrister of the High Court of Australia and of the Supreme Court of New South Wales; Barrister and Solicitor of the Supreme Court of the Australian Capital Territory. 9. See chapter 5. 10. Shannon and Golshani 1988 at 315. 11. See the discussion of A5 in the Authorization area of the Authorization specifica­ tion (§5.3.2). 12. The word “hypothetical” is used here in its general sense; ideal points are different to the hypotheticals which SHYSTER uses in argument (see §3.11.3). 13. For example, the Employee area of the Employee specification (§5.4.2) represents the open-textured concept of a worker’s employment status. There are two results: the worker is an employee or the worker is an independent contractor. 14. The Natural specification (§5.5.2) contains three areas. The root node (whether a duty to observe natural justice is implied in a particular decision-making process) is represented by the Natural area. One of its seven children is an internal node (whether the decision affected an interest of the applicant) represented by the Affected area. One of the four children of that node is also internal (whether the applicant had a legitimate expectation which was affected by the decision) and is represented by the Expectation area, which has six children—all leaf nodes. The relationship between the attributes in these areas is indicated in figure 5.8. 15. Mehl 1959 at 768. 16. id. 17. Susskind 1987a at 54, emphasis omitted. He continues: “[b]ut that is not a job for the near future.” 18. Berman and Hafner 1991 discuss a formal model for distinguishing between ques­ tions of fact, questions of law, and mixed questions. 19. See, for example, the development of the Natural specification (§5.5.2) from the work of McMillan 1991 (discussed in §5.5.1). 20. Mackaay and Robillard 1974 at 308. See §2.5.1. 21. As discussed in §3.5.2. 22. An example of this can be found in the Employee specification described in §5.4.2. The legal expert identified the question of whether the worker’s employer deducted “pay as you earn” (PAYE) tax instalments from the worker’s pay as being an attribute in the Employee area (A15). In Zuijs v. Wirth Brothers Pty Ltd (1955) 93 CLR 561, Wirth Brothers deducted PAYE tax instalments from Zuijs’s pay. Dixon CJ, Williams, Webb and Taylor JJ of the High Court mention this fact (at 567), but do not use it in coming to their decision. 23. Blackstone 1773 (especially vol. 1 at 69–70); Krygier 1986. 24. For example, Stone 1964a, 1985. 25. The Natural area of the Natural specification (§5.5.2) affords an example of this. A3 concerns the question whether the administrative power in question is of a nature that would suggest that procedural fairness would be applied in its exercise. When 386 [pages 61–68] Notes to chapter 3 Bread Manufacturers of New South Wales v. Evans (1981) 38 ALR 93 was decided, the answer in that case would have been yes. The law has developed to the extent that, if that case were to be heard now, the answer would be no. 26. Ideal points are so-named because, as explained in §3.9.1, SHYSTER treats cases as points in space, the dimensionality of which is the number of attributes. 27. Apropos specifying more than one ideal point, see §6.2.5. 28. Lambert and Grunewald 1989. LESTER deals with unjust discharge from employ­ ment under collective bargaining agreements. 29. ibid. at 91. 30. Ashley and Rissland 1988 at 239. 31. Williams 1971. 32. Sneath and Sokal 1973. 33. Aldenderfer and Blashfield 1984 at 21, emphases added, footnotes added. 34. Everitt 1974 at 50. Note that Haar, Sawyer and Cummings 1977 use statistical methods to determine which are the important attributes, and ignore the unim­ portant ones. 35. Tyree 1989 at 149, n. 16. 36. ibid. at 141, emphases added. Tyree refers to separating the cases into two classes because FINDER allows only two results for each leading case. SHYSTER allows an arbitrary number of results in each specified area of law. 37. Tyree, Greenleaf and Mowbray 1988 at 245, n. 8. 38. Neter, Wasserman and Whitmore 1982 at 185. This formula defines the variance of a finite number of values—a finite population—and is different from the formula that defines the variance of a sample of values. 39. FINDER does not deal with infinite weight as none of its attributes has zero vari­ ance. 40. Real examples are A5 in the Authorization area of the Authorization specifica­ tion (§5.3.2), and A4 in the Expectation area of the Natural specification (§5.5.2). 41. Ashley 1990 at 175. 42. Bing 1992 at 104. 43. id. Bing’s reference is to FINDER’s method, but his comment applies to SHY­ STER’s method too. 44. Ashley 1990 at 175–6. 45. Ashley and Rissland 1988. 46. Ashley 1990 at 176–8, capitalization added. 47. Lambert and Grunewald 1991 at 194. 48. There is only a stochastic dependence in this example if there is some variation of values within the attributes. In the extreme circumstance where all of the attribute values are the same, the two attributes are (strictly) stochastically independent. 49. An example of this can be found in the discussion of attribute dependence in the Employee specification in §5.4.2. 50. If an attribute’s known values are all yess, or all nos, the legal expert is warned that the attribute has been given infinite weight (as explained in §3.7). 51. These attributes are A4 and A9, respectively, from the Employee area of the Employ­ ee specification explained in §5.4.2 (see also the attribute value matrix in fig­ ure 4.4). Notes to chapter 3 [pages 69–85] 387 52. Conover 1971 at 154–8. 53. Fisher 1970. 54. Aldenderfer and Blashfield 1984 at 18, emphasis omitted. 55. id. Everitt 1974 at 56 lists all but the third criterion. 56. Aldenderfer and Blashfield 1984 at 18–19. 57. ibid. at 19. 58. Sneath and Sokal 1973 at 140–5; Aldenderfer and Blashfield 1984 at 33. 59. Attribute values are often standardized before Euclidean distance is calculated. Standardization is not necessary in SHYSTER as all of the attribute values are of the same type. 60. In SHYSTER, n is the number of known pairs of attribute values. Attribute pairs with one or two unknowns are ignored for these purposes; they are taken into account in the calculation of unknown distance. 61. Aldenderfer and Blashfield 1984 at 28. 62. Everitt 1974 at 51. 63. Gower’s coefficient, another measure of similarity, is identical to Jaccard’s coeffi­ cient in the binary case: Aldenderfer and Blashfield 1984 at 31. 64. Aldenderfer and Blashfield 1984 at 22. 65. See ibid. at 23–4, and Everitt 1974 at 53–4. 66. Identical cases do not have a coefficient of zero, and the coefficient often fails to satisfy the triangle inequality. 67. Aldenderfer and Blashfield 1984 at 23. 68. A correlation coefficient of −1 indicates a negative linear association; a value of 1 indicates a positive linear association. A value of zero indicates no linear associ­ ation. 69. Ashley 1990 at 127. 70. ibid. at 128. 71. As explained in §3.7, HYPO adopts a “symbolic least commitment approach” to attribute weighting. 72. Sneath and Sokal 1973 at 146. 73. These distances are for Narich Pty Ltd v. Commissioner of Pay-roll Tax : one of the cases used to test the Employee specification described in §5.4.2. Narich v. CPT is described in §5.4.3. Its complete table of distances, as extracted from SHYSTER’s distances file, is given in figure 5.7. 74. The Employee and Contractor results, respectively (see §5.4.2). 75. Mackaay and Robillard 1974 at 309–10. They term this the problem of “tied distances amongst neighbours.” 76. Tyree, Greenleaf and Mowbray 1989 at 47. 77. Ashley 1990 at 39. 78. Insofar as a program can have an opinion. 79. Ashley 1990 at 230–1, capitalization added. 80. ibid. at 232. 81. Note that the use of unknown distance as a measurement of error (discussed in §3.9) does not account for the possibility of an instantiation producing a different result. An unknown attribute value in the instant case has the effect of adding the weight � � 388 [pages 85–92] Notes to chapter 3 of the attribute to the unknown distances of all the leading cases: i.e. the leading cases are all pushed away from the instant case by the same (unknown) distance. 82. Used in this fashion, instantiation is similar to SHYSTER’s use of hypotheticals (§3.11.3), only with a higher degree of user control. 83. Ashley 1990 at 85. 84. ibid. at 233. See also Rissland 1989. 85. Ashley 1990 at 232. 86. unknown values are dealt with by instantiation: see §3.11.2. 87. The number of hypotheticals with exactly k differences from the instant case, given n attributes, is n = n(n−1)···(n−k+1) . So, the number of hypotheticals h with no k � �k! � � � � more than k differences is n + n + + n . 1 2 k· · · The Employee area of the Employee specification (§5.4.2), for example, has 18 attributes. Allowing no more than 2 differences (k = 2) restricts SHYSTER to 171 hypotheticals. If k = 3, h = 987; and if k = 4, h = 4047—still only a small fraction of the maximum: 218 = 262 144 hypotheticals. 88. Each yes is assigned a value of 1, and each no is assigned a value of 0. For the purposes of constructing the centroid, each mean (a number n in the range 0 to 1) is rounded to the nearest attribute value (0 ∗ n < 0.5: no; 0.5 ∗ n ∗ 1: yes). However, the unrounded means are used when calculating the correlation coefficients r and r� for those centroids. 89. Tyree et al. 1988 at 241. 90. SHYSTER uses the attribute’s result weight (explained in §4.9.1) when calculating the strength of directions. 91. Susskind 1987a at 55. 92. ibid. at 53–5. 93. Leith 1985a, 1986a. 94. Susskind 1986 at 175–6. It is probably unnecessary to point out that Susskind wrote this article for a legal audience. He continues: “This is not to belittle Leith’s achievements, for he was clearly working with limited resources” (at 176). Leith bitterly disputes Susskind’s com­ ments, including his definition of what constitutes a legal expert system. “If there is no agreement [over suitable terminology in the field], why should computer scient­ ists accept the definition proposed by a lawyer? Will Susskind accept the definition of ‘divorce’ proposed by a computer scientist?” (1987 at 130). Susskind’s early ex­ perimental work was with the Scottish law of divorce. 95. Clark 1988 at 428. 96. Gardner 1987 at 33. Gardner’s chosen domain was one aspect of the law of contract (see §2.4.3). 97. There are two exceptions to this rule—two tests where SHYSTER’s choice of nearest result is deemed “good” despite disagreeing with the actual cases: Twist v. Council of Municipality of Randwick (1976) 136 CLR 106 and Salemi v. MacKellar (1977) 137 CLR 396. These exceptions are explained and justified in §5.5.3. 98. The case upon which SHYSTER bases its argument is the nearest neighbour ; the case upon which the counterargument is based is the nearest other (see §3.10.1). 99. (1977) 137 CLR 396. Notes to chapter 3 [pages 92–99] 389 1. As discussed in §2.4.7, the number of decided cases in any given area of law is usually so small that inductive inference algorithms cannot be used. 2. And bad legal advice may filter out a hopeful case. 3. Gardner 1987 at 6. 4. Tyree 1986, 1989 and Tyree et al. 1988, 1989. Tyree et al. 1989 also describe (at 49–50) a reflexive test using Bridges v. Hawkesworth (1851) 21 LJQB 75. 5. Ashley 1990 at 183–93: Crown Industries, Inc. v. Kawneer Co. (1971) 335 F Supp 749; Structural Dynamics Research Corporation v. Engineering Mechanics Research Corporation (1975) 401 F Supp 1102; USM Corporation v. Marson Fastener Cor­ poration (1979) 379 Mass 90; Amoco Production Co. v. Lindley (1980) 609 P 2d 733. 6. For example, the Employee area of the Employee specification (§5.4.2) has 18 at­ tributes and, hence, a search space of 218 = 262 144 different cases. 7. Some of the cases generated for the Employee specification represent paradoxes. These impossible cases are discussed in §5.4.4. 8. SHYSTER issues a warning if the instant case is nearer to the ideal point for another result than it is to that of the nearest result (see §3.12.2). 9. Mackaay and Robillard 1974 at 312. Lawlor 1968 terms this reflexive approach “cyclical sampling” (at 110–11). 10. Compare this with the approach taken by Mackaay and Robillard 1974. They say that cases which are “incorrectly” predicted in a reflexive test are “suspect” because “[s]uch cases are difficult to explain on the basis of the remaining ones” (at 312). Mackaay and Robillard were dealing with a large set of cases, not a smaller set of expertly chosen cases as in SHYSTER. 11. The Finder specification (§5.2.2) uses only English cases. 12. Greenleaf, Mowbray and Tyree 1987 at 11. 13. Gardner 1987 at 22. 14. See §2.4.3. 15. Gardner 1987 at 74–5. Gardner refers to several of the researchers discussed in chapter 2, including Kort 1963a, Lawlor 1963, 1972, Mackaay and Robillard 1974, Haar et al. 1977, Borchgrevink and Hansen 1980, and Tyree 1981. 16. Greenleaf, Mowbray and Tyree 1991 propose a similar mechanism for their DATA­ LEX project (at 223), though the extent to which it has been implemented is not clear. See also Tyree et al. 1989 at 49. 17. Ashley 1990 at 254. 18. Gardner 1987 at 75. Questions of fact and questions of law are discussed in §3.5.2. 19. HYPO’s knowledge representation is explained in §2.5.3. 20. Tyree 1977 at 414, n. 40. There are two unknown values coded as no in FINDER (see §5.2.2). 21. Ashley 1990 at 128. 22. For example, the Authorization area of the Authorization specification (§5.3.2) has three results. 23. See §5.2.2. 24. Because FINDER does not allow unknown values, it has no unknown distance. None of its attributes is infinitely weighted. 390 [pages 99–108] Notes to chapter 4 25. Tyree et al. 1988 at 241. 26. Ashley 1990 at 249. 27. An example is the Authorization area of the Authorization specification (§5.3.2). An answer of yes to A3 (“Did the accused sell or hire the infringer the means of infringing?”) can be directed towards the result Auth, but an answer of no or unknown for that attribute cannot be directed towards either of the other results: Not-Auth or Liable. The fact that the accused did not sell or hire the infringer the means of infringing (A3 = no) suggests neither that the accused did not authorize the infringement (Not-Auth) nor that the accused is directly or vicariously liable for the infringement (Liable); similarly for A3 = unknown. Another example is A6 in the Employee area of the Employee specification de­ scribed in §5.4.2. 28. Berman 1991 at 308. 29. See §2.7 and §2.8. 30. Berman 1991. 31. Including Gardner 1987, Goldman, Dyer and Flowers 1987, Hafner 1987, Rissland and Ashley 1987, Branting 1989, and Ashley and Aleven 1991. 32. For example, several attributes in areas in the Natural specification (§5.5.2) em­ body political considerations. 33. Berman 1991 at 307–8. 34. ibid. at 308. Chapter 4: Implementing SHYSTER 35. Gay 1967, vol. II at 1–2 (fable I). 36. Lucas 1935 at 394. Letter to Samuel Rogers, probably 21 December 1833. 37. Paramount. Screenplay by Arthur Sheekman. 38. International standard ISO/IEC 9899: 1990; Australian standard AS 3955–1991. Kernighan and Ritchie 1988 describe ANSI C which is the same as ISO C. 39. Popple 1993. 40. 256 characters and 16 characters, respectively. 41. UNIX’s stderr stream. 42. Lamport 1986 describes LaTEX which is a set of macros for Knuth’s TEX system (1984). SHYSTER’s LaTEX output is suitable for processing by LaTEX version 2.09 �25 March 1992� and TEX version 3.141; it should also be suitable for processing by later versions. 43. Figures extracted from SHYSTER output are marked with a star in the list of figures that immediately precedes chapter 1. 44. The SHYSTER output that appears in appendices A, B and C was produced using the -i switch which causes SHYSTER to write LaTEX code suitable for inclusion in another LaTEX document. 45. UNIX’s stdout stream. 46. A string, as defined in figure 4.3, includes the enclosing quotation marks. The Tokenizer discards these quotation marks when the specification is read. Hence, the word string is used in two different ways in this description of SHYSTER: before Notes to chapter 4 [pages 108–114] 391 being tokenized, a string includes the enclosing quotation marks; once internally represented by SHYSTER, a string is only those characters which appeared between the enclosing quotation marks in the specification file. 47. A string may contain commands which will be processed by LaTEX if the string is in­ cluded in one of SHYSTER’s files. LaTEX commands are prefixed with a \ character. The $, & and % characters have special meanings in LaTEX; if one of these characters appears in a string, SHYSTER converts it into the command that produces the appropriate character: \$, \& or \%. 48. As with strings (see the note before last), attribute vectors are defined in figure 4.3 to include the enclosing parentheses, but the Tokenizer discards these parentheses when the specification is read. 49. Whitespace is any sequence of spaces, carriage returns, tabs, vertical tabs, or form feeds. 50. SHYSTER’s is an LL(1) parser. 51. EBNF is described by Wirth 1977. 52. Note that a result identifier from another (explicitly named) area can be referred to within the specification of an external attribute, as explained in §4.6.3. 53. This is A12 from the Employee area of the Employee specification (see §A.4 and §C.3). 54. This is A4 from the Affected area of the Natural specification (see §A.5). 55. The external area does not need to have been specified before it is used in an external attribute. Forward references are allowed, and are not checked until in­ vocation. Recursive references (circular definitions) are not allowed but SHYSTER does not detect them, unless the reference is to the same area in which the attribute is specified. 56. The specification of an UNKNOWN string is not strictly necessary; no external result identifier can produce a value of unknown in this attribute. However, if it is not specified and one of the leading cases or ideal points has a value of unknown for this attribute, the Dumper module will issue a warning. 57. Result identifiers are not checked for validity: i.e. the legal expert should ensure that these identifiers match result identifiers in the external area. If the same result identifier is specified for two or all of the possible attribute values it is linked to the first specified (the order being yes, no then unknown). 58. The order in which cases are specified is unimportant. SHYSTER groups cases by their result, and orders the cases in each group according to the importance of the court in which each case was decided. The cases in the Employee specification appear in the specification file (§C.3) in the order in which they were decided. 59. See, for example, Moffatt v. Kazana [1969] 2 QB 152 in the Finder area of the Finder specification (§5.2.2), and Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance [1968] 2 QB 497 in the Employee area of the Employee specification (§5.4.2). Both cases were decided in 1967. 60. This warning is actually issued by the Cases module, not the Parser, but it is convenient to mention it here. 61. This happens in the Natural specification (§5.5.2), but note the discussion there as to why that specification was not changed. 392 [pages 114–125] Notes to chapter 4 62. Some experimentation by the author has shown that this sort of graphical repres­ entation conveys its information more clearly than does one using Ys, Ns and Us, or ones and zeros. 63. As with the example specification of a local attribute in §4.6.3, this is A12 from the Employee area of the Employee specification (see §A.4). 64. If an attribute value is directed towards more than one result, result identifiers are displayed separated by a ∧ symbol. 65. As with the example specification of an external attribute in §4.6.3, this is A4 from the Affected area of the Natural specification (see §A.5). 66. (1945) 70 CLR 539. 67. This cell contains the probabilities for the attribute values in figure 3.2, used as an example in §3.8.2. These example attributes are A4 and A9 from the Employee area of the Employee specification (§5.4.2). Figure 4.5 is an extract from the probabilities matrix for that area. 68. Because all of SHYSTER’s known attribute values are assigned numerical values 2of 0 or 1, the variance �i 2 (defined in §3.7) simplifies to Ā i − Ā i . SHYSTER uses this simplified formula to calculate the variance. 69. If an attribute has no known attribute values, i.e. all its values are unknown, it is given no weight and a warning message is issued. A further warning is issued (by the Odometer) if the instant case or an ideal point has a known attribute value for the weightless attribute. 70. Distance comparisons, like other comparisons, are precise to two decimal places. They need not be; by changing one constant, SHYSTER can be rebuilt with a different distance comparison threshold. 71. The Odometer is invoked twice for some hypotheticals. The first time it performs its calculations without writing anything to the distances file. Only some hypo­ theticals are chosen to be reported upon; for these hypotheticals the Odometer is called a second time to perform the calculations again and to write details to the distances file. Only a fraction of the hypotheticals are chosen; performing the calculations twice for the chosen hypotheticals avoids having to store the results of calculations from previous hypotheticals. 72. How hypotheticals are chosen is explained in §3.11.3. 73. If one of the attributes is infinitely weighted, the values obtained for S� become meaningless because both the numerator and the denominator are infinite. In this event, values for S� are not written to the dump file, and S� is not used as a safeguard. Calculation of the weighted correlation coefficient r� is complicated when one of the attributes is infinitely weighted. If this happens, a very large weight is used in place of an infinite one. If either of two cases has all attribute values the same, the correlation coeffi­ cients r and r� are meaningless and are ignored for safeguard purposes. 74. How hypotheticals are chosen to be reported on is explained in §3.11.3. 75. Cases are not summarized more than once in the same report—and not summarized at all if the user has enabled quiet mode (using the -q switch on the command line). In quiet mode, opening and closing strings are not written either. 76. (1949) 79 CLR 389. Notes to chapter 5 [pages 125–136] 393 77. The rather quaint expression “on all fours” is used by FINDER (see, for example, Tyree 1989 at 160). Although this expression may seem twee, it is preferable to “the same,” “identical,” etc., because two cases are never completely identical. Its meaning is clear from the context, and its oddness draws the reader’s attention to the fact that something less than sameness is meant. 78. [1968] 2 QB 497. 79. [1976] 1 WLR 1213. 80. If there is more than one nearest neighbour, the comparison is made with the most important nearest neighbour. 81. This quotation is taken from the report file for Narich Pty Ltd v. Commissioner of Pay-roll Tax (1983) 50 ALR 417 which is used as a test case in §5.4.3. 82. The opening and closing strings are not written again when arguing with instanti­ ations. 83. This quotation is taken from the report on the fourth instantiation in the report file for Re Porter; Re Transport Workers Union of Australia (1989) 34 IR 179 which is given in full in §B.4. Re Porter; Re TWU is used as a test case in §5.4.3. 84. As with instantiations, the opening and closing strings are not written again when arguing with hypotheticals. 85. “Weighted association coefficients” does not appear as a subheading in this example because those coefficients agree with the choice of nearest neighbour. Chapter 5: Case studies 86. Wiener 1962 at 1023–4. 87. Doonesbury copyright G. B. Trudeau. Reprinted with permission of Universal Press Syndicate. All rights reserved. 88. (1722) 1 Str 505. 89. Bridges v. Hawkesworth (1851) 21 LJQB 75 at 77, per Patteson J. 90. (1851) 21 LJQB 75. 91. (1886) 33 ChD 562. 92. [1896] 2 QB 44. 93. [1945] 1 KB 509. 94. Hibbert v. McKiernan [1948] 2 KB 142 at 149 (footnotes omitted). Lord God­ dard CJ, Humphreys and Pritchard JJ decided that the law of trover did not apply in that case. 95. Tyree 1989 at 119. 96. [1982] 1 All ER 834. 97. ibid. at 843. 98. id. Tyree 1989 says that Donaldson LJ’s statement of the principles to be applied is “a good candidate for a direct translation into a rule-based system” (at 119). He develops some rules to reflect this statement and concludes that “building even small rule-based systems is not a task which may be completed easily” (at 126); the rule base is considerably more complicated than both FINDER and SHYSTER’s specification of FINDER. 99. [1948] 2 KB 142 at 149. 1. Tyree 1989 at 131, n. 12. 394 [pages 136–143] Notes to chapter 5 2. In City of London Corporation v. Appleyard (1), “the finders . . . were not the servants of Yorkwin but the servants of Wates Ltd, who were engaged by Yorkwin as independent contractors”: [1963] 1 WLR 982 at 988, per McNair J. In South Staffordshire Water Co. v. Sharman, the finder was “employed” by the company. It is not clear whether the finder was an employee or an independent contractor, although some of what Lord Russell of Killowen CJ says in the case suggests the latter: [1896] 2 QB 44 at 46. 3. (1722) 1 Str 505. 4. (1851) 21 LJQB 75. 5. (1886) 33 ChD 562. 6. [1896] 2 QB 44. 7. [1945] 1 KB 509. 8. [1963] 1 WLR 982. 9. [1969] 2 QB 152. 10. In fact, the premises were leased by Venture Property and Development Co. Ltd which held the premises on trust and to the order of Yorkwin. Venture and York- win were “associated companies.” This added complication is not relevant for FINDER’s (or SHYSTER’s) purposes, and is ignored. 11. In FINDER these “two cases” are referred to as Yorkwin v. Appleyard and Corpor­ ation of London v. Yorkwin, respectively. 12. See §3.14.1. 13. A2 and A8 in Armory v. Delamirie. 14. A10 in Armory v. Delamirie, and A5 in City of London Corporation v. Apple- yard (2); both are yess and both should be nos. 15. In the Finder area, w4 = w9 > w6 = w7 = w8 = w10 > w1 = w3 = w5 > w2 (where wi is the weight of Ai). 16. [1982] 1 All ER 834. 17. Donaldson LJ put it somewhat melodramatically: “On 15 November 1978 Mr Alan George Parker had a date with fate, and perhaps with legal immortality” (ibid. at 835). 18. [1982] 1 All ER 834 at 838, per Donaldson LJ. 19. ibid. at 838–40. His lordship was surprised to find the law so complex. “It is astonishing that there should be any doubt as to who is right. But there is. Indeed, it seems that academics have been debating this problem for years” (at 836). 20. ibid. at 845. 21. ibid. at 843, per Donaldson LJ (with whom Eveleigh LJ and Sir David Cairns agreed). 22. [1963] 1 WLR 982 at 986, per McNair J. 23. Tyree 1986 at 18; 1989 at 160; Tyree, Greenleaf and Mowbray 1988 at 246–7; 1989 at 50–1. 24. e.g. University of New South Wales v. Moorhouse (1975) 133 CLR 1; RCA Cor­ poration v. John Fairfax and Sons Ltd [1981] 1 NSWLR 251; WEA International Inc. v. Hanimex Corporation Ltd (1987) AIPC ¶90-428; Australasian Performing Right Association Ltd v. Jain (1990) AIPC ¶90-718. 25. [1924] 1 KB 1 at 12. 26. [1926] 2 KB 474 at 491. Notes to chapter 5 [pages 143–149] 395 27. Adelaide Corporation v. Australasian Performing Right Association Ltd (1928) 40 CLR 489. 28. Ricketson 1984 at 229. 29. (1975) 133 CLR 1 at 12–13. 30. ibid. at 12, citing Adelaide Corporation v. Australasian Performing Right Associ­ ation Ltd (1928) 40 CLR 481 at 497–8, 503. Although Adelaide v. APRA is cited as authority for several propositions in this area, it is not one of the leading cases in the Authorization specification because it was not actually a case about authoriz­ ation. The question in that case was whether the City of Adelaide had “permitted” a song to be sung contrary to s. 2(3) of the Copyright Act 1911 (UK), in force in Australia by virtue of the Copyright Act 1912 (Cth). 31. (1975) 133 CLR 1 at 12–13. 32. ibid. at 13. 33. ibid. at 12. 34. Ricketson 1984 at 229–34. 35. Peter Drahos, Lecturer, Faculty of Law, The Australian National University. 36. Ricketson 1984 states the law as it was on 1 May 1983. 37. The Employee specification would have to be modified before it could be linked to the Authorization specification in this fashion. At present, the Employee area assumes that there is a worker who is either an employee or an independent con­ tractor (see §5.4.2). In the Authorization area, the infringer may not work for the accused at all. 38. (1975) 133 CLR 1 at 13, per Gibbs J. 39. See §6.3 for a discussion of conditional attributes. 40. See §6.2.4. 41. [1924] 1 KB 1. 42. [1926] 2 KB and 474. 43. [1979] FSR 1. 44. [1940] AC 491. 45. [1946] VLR 338. 46. [1962] NSWR 405. 47. [1964–65] NSWR 138. 48. (1975) 133 CLR 1. 49. [1981] 1 NSWLR 251. 50. [1917–23] MacGillivray’s Copyright Cases 309. 51. In the Authorization area, w5 > w6 > w1 > w2 > w4 > w3 = w7 (where wi is the weight of Ai). 52. [1982] Ch 91. 53. id. 54. [1945] AC 108, a decision of the Judicial Committee of the Privy Council on appeal from the Supreme Court of Canada. 55. [1982] Ch 91 at 95. 56. ibid. at 95–6. 57. (1987) AIPC ¶90-428. 58. ibid. at 32 764–5. 396 [pages 149–163] Notes to chapter 5 59. ibid. at 37 764. 60. ibid. at 37 766. 61. [1981] 1 NSWLR 251 at 257, per Kearney J. 62. [1988] AC 1013. 63. (1990) AIPC ¶90-718. 64. ibid. at 36 648. 65. id. 66. In the Authorization specification, a value of yes for A7 in the Authorization area is directed towards Auth or Liable. 67. Creighton, Ford and Mitchell 1983 at 25. 68. Brooks 1990 at 1. 69. id. 70. Creighton et al. 1983 at 25. 71. This simplified statement of the law is incorporated in the help strings for A1 and A2 in the Authorization specification (see §5.3.2). 72. (1978) 21 ALR 388 at 399–402, per Smithers, Evatt and Keely JJ. 73. Creighton et al. 1983 at 25–44. 74. Brooks 1990; Smith 1992. 75. Phillipa Weeks, Senior Lecturer, Faculty of Law, The Australian National Uni­ versity. 76. Creighton et al. 1983 state the law as it was on 31 March 1982. 77. “Pay as you earn.” 78. (1940) 14 ALJ 162. 79. (1980) 29 ALR 322. 80. (1978) 21 ALR 388. 81. (1944) 69 CLR 227. 82. (1945) 70 CLR 539. 83. [1968] 2 QB 497. 84. (1955) 93 CLR 561. 85. (1949) 79 CLR 389. 86. [1952] 1 TLR 101. 87. [1924] 1 KB 762. 88. [1976] 1 WLR 1213. 89. (1978) 18 ALR 385. 90. [1978] ICR 590. 91. In the Employee area, w13 = w14 = w16 > w17 > w15 > w2 = w3 = w10 > w8 > w18 > w12 > w6 > w1 > w9 > w4 > w5 > w7 > w11 (where wi is the weight of Ai). 92. (1983) 50 ALR 417. 93. (1986) 160 CLR 16. 94. ibid. at 24, 26, per Mason J (with whom Brennan and Deane JJ generally agreed); at 38, per Wilson and Dawson JJ. 95. (1989) 34 IR 179. 96. (1954) 94 CLR 409. 97. [1985] VR 577. 98. (1989) 34 IR 179 at 181. Notes to chapter 5 [pages 164–174] 397 99. There were five different owner-drivers in Re Porter; Re TWU, and the circum­ stances for all five were very similar. These attribute values are for the first driver mentioned in the judgment—Newman: (1989) 34 IR 179 at 188–92. The facts of Re Porter; Re TWU are summarized in the third paragraph of SHYSTER’s report file in §B.4: “In the instant case . . .” 1. (1991) 99 ALR 735. 2. In this matter, Odco was trading as “Trouble Shooters Available.” 3. (1991) 99 ALR 735 at 736. 4. ibid. at 752–6. 5. ibid. at 756. 6. The facts of Building Workers’ Industrial Union of Australia v. Odco Pty Ltd are summarized in the third paragraph of SHYSTER’s report file in §C.13: “In the instant case . . .” 7. Individual Employment Law, 1991, Faculty of Law, The Australian National Uni­ versity, question 4. 8. [1976] 1 WLR 1213 at 1217–19. 9. These paradoxes could not be removed from the search space by giving A13 and A14 known values in the initial fact vector. It is only a yes/yes pair which is impossible; the other three combinations are possible. 10. Generated test 5 contributes 80% of the instantiations in these three tests. 11. Clark and Wedderburn 1983 at 153. 12. McMillan 1991 at 63. 13. (1985) 159 CLR 550 at 584. 14. McMillan 1991 at 68. 15. Either term can be used: e.g. in Haoucher v. Minister of State for Immigration and Ethnic Affairs (1990) 169 CLR 648, McHugh J used “legitimate” and Deane J used “reasonable.” 16. Cole v. Cunningham (1983) 49 ALR 123 at 131, per Bowen CJ, Sheppard and Morling JJ. 17. McMillan 1991 at 70, citing (1990) 169 CLR 648 at 653. 18. (1987) 75 ALR 218 at 251. 19. [1967] 2 AC 337. 20. (1977) 137 CLR 396. 21. (1982) 151 CLR 342. 22. McMillan 1991 at 71. 23. ibid. at 71–4. 24. Twist v. Council of Municipality of Randwick (1976) 136 CLR 106. 25. ibid. at 117. The section references are to the Local Government Act 1919 (NSW). 26. (1985) 62 ALR 253 at 266. 27. (1986) 67 ALR 77 at 81. 28. (1990) 21 ALD 730 at 735. 29. (1991) 24 ALD 768. 30. Robin Creyke, Senior Lecturer, Faculty of Law, The Australian National University. 31. McMillan 1991. 32. The significance of this attribute is discussed in §5.5.1. For reasons discussed there, the specified directions direct a value of yes towards Not-Implied, and a value of no 398 [pages 175–183] Notes to chapter 5 towards Implied. Note that there are only two leading cases in the Natural area with a value of yes for A6 —Commissioner of Police v. Tanos and Marine Hull & Liability Insurance Co. Ltd v. Hurford—and natural justice was implied in both, but for other reasons. 33. (1990) 169 CLR 648. 34. [1983] 2 AC 629. 35. (1982) 151 CLR 342. 36. (1990) 169 CLR 648. 37. (1990) 170 CLR 596. 38. (1985) 159 CLR 550. 39. (1958) 98 CLR 383. 40. (1987) 163 CLR 378. 41. (1981) 38 ALR 93. 42. (1986) 67 ALR 77. 43. (1987) 75 ALR 218. 44. (1987) 9 NSWLR 268. 45. (1981) 36 ALR 215. 46. [1983] 2 AC 629. 47. [1967] 2 AC 337. 48. [1985] AC 374. 49. [1978] 3 All ER 211. 50. (1990) 169 CLR 648. 51. (1983) 49 ALR 123. 52. (1976) 136 CLR 106. 53. (1977) 137 CLR 396. 54. Annetts v. McCann (C3) and McInnes v. Onslow Fane (C15), Macrae v. Attorney- General for New South Wales (C7) and South Australia v. O’Shea (C10), and Durayappah v. Fernando (C9) and Nashua Australia Pty Ltd v. Channon (C13). 55. In the Natural area, w1 = w7 > w2 = w6 > w5 > w4 > w3 (where wi is the weight of Ai). 56. This is the basis of generated test 7 (see §5.5.4). 57. In the Affected area, w2 = w3 > w1 = w4. 58. In the Expectation area, w4 > w3 = w6 > w1 = w5 > w2. 59. (1992) 175 CLR 564. 60. (1976) 136 CLR 106. 61. [1964] AC 40. 62. (1890) 24 QBD 703. 63. (1976) 136 CLR 106 at 111-12, per Barwick CJ; at 117, per Mason J. 64. ibid. at 118–19. 65. (1977) 137 CLR 396. 66. [1969] 2 Ch 149. 67. (1977) 137 CLR 396, per Barwick CJ, Gibbs, Stephen, Jacobs and Aickin JJ (Murphy J dissenting). 68. ibid. per Barwick CJ, Gibbs and Aickin JJ. 69. (1977) 137 CLR 487. 70. (1983) 49 ALR 123. Notes to chapter 6 [pages 184–198] 399 71. [1985] 2 NSWLR 239. 72. ibid. at 246. 73. (1989) 18 ALD 536. 74. ibid. at 537, per Thomas J (with whom Shepherdson and Williams JJ generally agreed). 75. (1991) 24 ALD 768. 76. ibid. at 770. 77. ibid. at 768. 78. See §6.3 for a discussion of conditional attributes. 79. (1992) 175 CLR 564. 80. ibid. at 578, per Mason CJ, Dawson, Toohey and Gaudron JJ; at 585, per Bren­ nan J. 81. A value of no for A1, an external attribute in the Natural area, is achieved as a result of a fact vector of (NNNNNN) in the Affected area. Similarly, a value of no for A4, an external attribute in the Affected area, is achieved as a result of a fact vector of (NNNN) in the Expectation area. 82. The way in which A1 in the Natural area is given a value of no is explained in the previous note. Similarly, fact vectors of (YYYYYY) and (YYYY) for the Affected and Expectation areas result in a value of yes for A1 in the Natural area. 83. McMillan 1991 at 71–4. See §5.5.1. 84. Neither I nor µ appear in figure 5.12. They appear in the figures in appendix D which summarize the reflexive testing, so they are explained here with the other symbols. 85. The Natural specification has 15 cases in the Implied area, 9 in the Affected area and 8 in the Expectation area. There are 17 different cases in the specification. 86. For example, Donaldson LJ’s discussion of finder cases in Parker v. British Airways (see §5.2.1 and §5.2.3), and Gummow J’s discussion of authorization cases in WEA International Inc. v. Hanimex Corporation Ltd (see §5.3.3). 87. (1989) 34 IR 179. Chapter 6: Conclusion 88. Stone 1964b at 556. 89. (1637) 3 How St Tr 825 at 969. 90. Eliot 1963 at 55. 91. Susskind 1989b at 24. 92. Kowalski and Sergot 1990 at 208. 93. id. 94. Moles 1991 at 162–3. 95. See §3.13.3. 96. Ashley 1990 at 127. 97. Ashley 1985 at 116, as quoted in §2.5. See also Ashley and Aleven 1991. 98. The distinction between private and public law is discussed in §3.3.1. 99. Mendelson 1989. 1. ibid. at 137. 400 [pages 198–312] Notes to appendix A 2. The hypothetical test cases are also difficult, although the ideal points (which are also used as test cases) are not. 3. The harshness of this method is mitigated somewhat by taking the legal expert’s opinion into account (see, for example, the discussions on Cole v. Cunningham and Ackroyd v. Whitehouse (Director of National Parks & Wildlife Service) in §5.5.3). 4. Note the comments on the Natural specification in §6.2.2. 5. See the discussion on A5 in the Authorization area of the Authorization specification (§5.3.2). 6. SHYSTER’s use of precision thresholds is discussed in §4.12.1. 7. See §5.2.2, §5.3.2, §5.4.2 and §5.5.2. 8. See §3.7. 9. The rank of courts is used as the first step in the resolution of equidistance (§3.10.3), and when referring to the relative importance of different leading cases (see §4.13). 10. See §2.1.2. 11. (1975) 133 CLR 1 at 13, per Gibbs J. This is the basis of generated test 3 (see §5.3.4). 12. See, for example, the discussion on Western Australia v. Bropho in §5.5.3. 13. LEXIS: Stanley 1980; WESTLAW: Herman 1980. 14. Ashley 1990 at 254, capitalization added. Appendix A: Case law specifications 15. Act I; Gilbert 1977 at 212. 16. No. 125, 28 May 1751, emphasis omitted; Johnson 1823 at 344. Appendix B: Example reports 17. Douglas 1948 at 105. 18. Quoted by Boswell 1960 at 342. 19. [1982] 1 All ER 834. 20. (1990) AIPC ¶90-718. 21. (1989) 34 IR 179. 22. (1992) 175 CLR 564. 23. The Reporter module will not summarize the same case twice in the same report file. However, there is some repetition in §B.5, because there are three separate report files for Ainsworth v. CJC—one for each area. (Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd (1987) 75 ALR 218 and Annetts v. McCann (1990) 170 CLR 596 are both summarized twice.) 24. (1991) 99 ALR 735. Appendix C: A complete example 25. Hepple 1986 at 71. 26. Carroll 1982 at 19. 27. (1991) 99 ALR 735. 28. The complete LaTEX input is shown for the weights file in §C.8. Notes to appendix D [pages 353–363] 401 Appendix D: Reflexive tests 29. 193 US 197 at 400. 30. Quoted by Lawlor 1963 at 344. 31. [1945] 1 KB 509. 32. ibid. at 520–1. 33. (1851) 21 LJQB 75. 34. ibid. at 77–8. 35. Tyree, Greenleaf and Mowbray 1989 at 49. 36. (1722) 1 Str 505. 37. [1969] 2 QB 152. 38. ibid. at 157. 39. [1963] 1 WLR 982. 40. ibid. at 986–7. 41. ibid. 42. [1896] 2 QB 44. 43. ibid. at 46, 48. 44. (1886) 33 ChD 562. 45. (1975) 133 CLR 1. 46. ibid. at 12, per Gibbs J; at 20–1, per Jacobs J (with whom McTiernan ACJ agreed). 47. ibid. at 13. 48. [1964–65] NSWR 138. 49. ibid. at 149–50. 50. ibid. at 140. 51. [1946] VLR 338. 52. ibid. at 355. 53. [1940] AC 491. 54. [1926] 2 KB 474. 55. ibid. at 491. 56. [1981] 1 NSWLR 251. 57. [1924] 1 KB 1. 58. [1979] FSR 1. 59. ibid. at 9–10. 60. [1962] NSWR 405. 61. (1955) 93 CLR 561. 62. (1940) 14 ALJ 162. 63. (1944) 69 CLR 227. 64. (1980) 29 ALR 322. 65. [1952] 2 All ER 956. 66. [1976] 1 WLR 1213. 67. [1969] 2 QB 173. 68. [1952] 1 TLR 101. 69. [1924] 1 KB 762. 70. (1949) 79 CLR 389. 402 [pages 363–370] Notes to appendix D 71. ibid. at 396–7, per Latham CJ; at 399, per Rich J; at 405, per Dixon J. Latham CJ also followed Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd. 72. (1945) 70 CLR 539. 73. (1927) 33 Argus LR 321. 74. (1978) 21 ALR 388. 75. (1978) 18 ALR 385. 76. ibid. at 391. 77. [1978] ICR 590. 78. ibid. at 596, per Lord Denning MR; at 597, per Lawton LJ; Eveleigh LJ agreed with both. 79. [1952] 1 TLR 101. 80. [1968] 2 QB 497. 81. ibid. at 525. 82. (1982) 151 CLR 342. 83. (1990) 169 CLR 648. 84. Deane, Toohey and McHugh JJ (Dawson and Gaudron JJ dissenting). 85. (1990) 169 CLR 648 at 651, per Deane J; at 679, per McHugh J. 86. (1990) 170 CLR 596. 87. Note, however, that Brennan and Toohey JJ dismissed the appeal on the basis that the decision of the Full Court of the Supreme Court of Western Australia (from which the Annettses had appealed) was right on the material before it. 88. (1990) 170 CLR 596 at 598, per Mason CJ, Deane and McHugh JJ; at 606, per Brennan J. 89. (1985) 159 CLR 550. 90. ibid. at 563, 567, per Gibbs CJ; at 582–3, 587 per Mason J; at 610, 616–7, per Brennan J. 91. (1958) 98 CLR 383. 92. (1986) 67 ALR 77. 93. ibid. at 88. 94. (1987) 9 NSWLR 268. 95. [1983] 2 AC 629. 96. [1967] 2 AC 337. 97. (1987) 163 CLR 378. 98. (1981) 38 ALR 93. 99. (1987) 75 ALR 218. 1. (1981) 36 ALR 215. 2. ibid. at 229. 3. [1985] AC 374. 4. [1978] 3 All ER 211. Cases Page numbers set in boldface type refer to summaries of case details. Page num­ bers set in italics refer to SHYSTER output. “4003” refers to note 3 on page 400. “1505.6” refers to figure 5.6 on page 150. Leading cases in SHYSTER specifications are marked with a star. Test cases are marked with a diamond. � A&M Records Inc. v. Audio Magnetics Inc. (UK) Ltd [1979] FSR 1: 146, 148, 149, 1505.6, 151, 152, 224, 357–359, 360, 360D.2 ∞ Ackroyd v. Whitehouse (Director of National Parks & Wildlife Service) [1985] 2 NSWLR 239: 184, 185, 1915.12, 4003 Adelaide Corporation v. Australasian Performing Right Association Ltd (1928) 40 CLR 481: 39527, 39530 AG of Hong Kong v. Shiu: see Attorney-General of Hong Kong v. Ng Yuen Shiu ∞ Ainsworth v. Criminal Justice Commission (1992) 175 CLR 564: 179, 187, 188, 1915.12, 284, 302–310, 40023 Amoco Production Co. v. Lindley (1980) 609 P 2d 733: 3895 AMP v. Chaplin: see Australian Mutual Provident Society v. Chaplin � Annetts v. McCann (1990) 170 CLR 596: 176, 183, 184, 187, 188, 254, 268, 305, 306, 308, 309, 367, 369, 370, 371D.4, 39854, 40023 APRA v. Canterbury-Bankstown: see Australasian Performing Right Association Ltd v. Canterbury-Bankstown League Club Ltd APRA v. Jain: see Australasian Performing Right Association Ltd v. Jain APRA v. Miles: see Australasian Performing Right Association Ltd v. Miles � Armory v. Delamirie (1722) 1 Str 505: 132, 133, 136, 210, 211, 354, 355, 356D.1, 39413, 39414 � Attorney-General of Hong Kong v. Ng Yuen Shiu [1983] 2 AC 629: 175, 176, 182, 258, 279, 280, 367, 368, 371D.4 403 http:1915.12 http:1915.12 404 [ATW–DAV] Cases ATWU v. Monaro Sawmills: see Australian Timber Workers Union v. Monaro Sawmills Pty Ltd � Australasian Performing Right Association Ltd v. Canterbury-Bankstown League Club Ltd [1964–65] NSWR 138: 146, 219, 220, 357, 358, 360D.2, 361 ∞ Australasian Performing Right Association Ltd v. Jain (1990) AIPC ¶90-718: 152, 1915.12, 284, 289–295, 39424 � Australasian Performing Right Association Ltd v. Miles [1962] NSWR 405: 146, 149, 1505.6, 151–153, 225, 291, 292–295, 357–360, 360D.2, 361 � Australian Mutual Provident Society v. Chaplin (1978) 18 ALR 385: 125, 1294.8, 159, 161, 163, 243, 244, 314, 315, 327, 351, 352, 364, 365D.3, 366 � Australian Timber Workers Union v. Monaro Sawmills Pty Ltd (1980) 29 ALR 322: 159, 234, 235, 328, 361, 362, 363, 365D.3 B v. B [1978] 1 All ER 821: 18 Bank Voor Handel en Scheepvaart NV v. Slatford [1952] 2 All ER 956: 362 Barro Group Pty Ltd v. Fraser [1985] VR 577: 163 Betts v. Brady (1942) 316 US 455: 24 � Bread Manufacturers of New South Wales v. Evans (1981) 38 ALR 93: 176, 180, 185, 187, 260, 261, 270, 306, 307, 369, 371D.4, 38625 � Bridges v. Hawkesworth (1851) 21 LJQB 75: 133, 136, 138–140, 1415.3, 209, 210, 214, 215, 284, 285–289, 354, 355, 356, 356D.1, 3894, 39389 Bropho v. Western Australia (1990) 21 ALD 730: 172 ∞ Building Workers’ Industrial Union of Australia v. Odco Pty Ltd (1991) 99 ALR 735: 122, 1234.7, 124–126, 128, 1294.8, 164, 165, 166, 169, 1915.12, 284, 312, 312–316, 335–352, 3976 BWIU v. Odco: see Building Workers’ Industrial Union of Australia v. Odco Pty Ltd � Cam and Sons Pty Ltd v. Sargent (1940) 14 ALJ 162: 126, 159, 163, 232, 233, 314, 315, 321, 349, 350, 361, 362–364, 365D.3, 366 Cantliff v. Jenkins [1978] 1 All ER 836: 18, 37768 ∞ CBS Inc. v. Ames Records and Tapes Ltd [1982] Ch 91: 147, 148, 149, 1505.6, 151, 1915.12 ∞ CBS Songs Ltd v. Amstrad Consumer Electronics plc [1988] AC 1013: 151, 1915.12 CCSU v. Minister for the Civil Service: see Council of Civil Service Unions v. Minister for the Civil Service � City of London Corporation v. Appleyard [1963] 1 WLR 982: 136, 140, 1415.3, 212, 213, 285, 286, 287–289, 354, 355, 356, 356D.1, 39411, 39414, 3942 � ∞ Cole v. Cunningham (1983) 49 ALR 123: 176, 183–187, 1915.12, 27455, 279, 3021, 303, 304, 39716, 4003 � Commissioner of Police v. Tanos (1958) 98 CLR 383: 176, 180, 1815.11, 182, 185, 255, 271, 367, 371D.4, 39832 � Council of Civil Service Unions v. Minister for the Civil Service [1985] AC 374: 176, 184, 185, 263, 272, 370, 371D.4 Crown Industries, Inc. v. Kawneer Co. (1971) 335 F Supp 749: 3895 Daniels and Daniels v. R. White and Sons and Tarbard [1938] 4 All ER 258: 37770 Davis v. Johnson [1978] 1 All ER 841: 37769 http:1915.12 http:1915.12 http:1915.12 http:1815.11 Cases [DUR–MEL] 405 � Durayappah v. Fernando [1967] 2 AC 337: 171, 176, 258, 259, 368, 369, 370, 371D.4, 39854 Eisner v. Macomber (1920) 252 US 189: 38071, 422 � Elwes v. Brigg Gas Co. (1886) 33 ChD 562: 133, 136, 215, 354–356, 356D.1 � FAI Insurances Ltd v. Winneke (1982) 151 CLR 342: 171, 176, 183–186, 252, 267, 276, 366, 367, 368, 371D.4 � Falcon v. Famous Players Film Co. [1926] 2 KB 474: 143, 146, 148, 149, 1505.6, 151–153, 2161, 221, 222, 2891, 357, 358, 359, 360, 360D.2 FCT v. Thompson: see Federal Commissioner of Taxation v. J. Walter Thompson (Australia) Pty Ltd � Federal Commissioner of Taxation v. J. Walter Thompson (Australia) Pty Ltd (1944) 69 CLR 227: 159, 233, 234, 322, 362, 365D.3 � Ferguson v. John Dawson & Partners (Contractors) Ltd [1976] 1 WLR 1213: 126, 128, 1294.8, 159, 161, 1625.7, 164–166, 236, 296, 297–302, 313–315, 326, 345, 347, 348, 350–352, 361, 362, 363, 364, 365D.3 Gideon v. Wainright (1963) 372 US 335: 24 � Hannah v. Peel [1945] 1 KB 509: 133, 136, 209, 288, 289, 354, 355, 356, 356D.1 � Haoucher v. Minister of State for Immigration and Ethnic Affairs (1990) 169 CLR 648: 171, 175, 176, 180, 182, 185, 186, 253, 267, 268, 277, 366, 367, 368, 371D.4, 39715 � ∞ Heatley v. Tasmanian Racing and Gaming Commission (1977) 137 CLR 487: 176, 180, 183, 184, 1915.12, 278 Hibbert v. McKiernan [1948] 2 KB 142: 39394 ∞ Hodgens v. Gunn (1989) 18 ALD 536: 185, 1915.12 � Humberstone v. Northern Timber Mills (1949) 79 CLR 389: 124, 124–126, 128, 1294.8, 159, 163, 164, 240, 313–315, 323, 344–346, 347–350, 361–363, 364, 365D.3 � Kioa v. West (1985) 159 CLR 550: 170, 176, 182, 183, 2501, 255, 270, 278, 30820, 367, 368, 371D.4 Logan v. Gilchrist, Watt and Cunningham (1927) 33 Argus LR 321: 364 London v. Appleyard : see City of London Corporation v. Appleyard � McInnes v. Onslow Fane [1978] 3 All ER 211: 176, 180, 1815.11, 183–185, 188, 264, 309, 367, 370, 371D.4, 39854 � Macrae v. Attorney-General for New South Wales (1987) 9 NSWLR 268: 176, 257, 310, 368, 369, 371D.4, 39854 Marine Hull & Liability Insurance Co. Ltd v. Hurford (1985) 62 ALR 253: 172 � Marine Hull & Liability Insurance Co. Ltd v. Hurford (1986) 67 ALR 77: 172, 176, 187, 256, 367, 371D.4, 39832 Market Investigations Ltd v. Minister of Social Security [1969] 2 QB 173: 362 � Massey v. Crown Life Insurance Co. [1978] ICR 590: 127, 159, 161, 1625.7, 165, 166, 244, 245, 326, 362, 363, 364, 365, 365D.3 � Mellor v. Australian Broadcasting Commission [1940] AC 491: 146, 152, 221, 290, 291–294, 357, 358, 359, 360D.2 http:1915.12 http:1815.11 406 [MIN–SOU] Cases � Minister for Arts Heritage and Environment v. Peko-Wallsend Ltd (1987) 75 ALR 218: 171, 176, 180, 182–188, 2502, 261, 262, 273, 281, 302, 303–305, 306, 307, 30821, 369, 371D.4, 40023 � Moffatt v. Kazana [1969] 2 QB 152: 136, 137, 211, 355, 356D.1, 39159 ∞ Narich Pty Ltd v. Commissioner of Pay-roll Tax (1983) 50 ALR 417: 160, 1625.7, 1915.12, 38773, 39381 � Nashua Australia Pty Ltd v. Channon (1981) 36 ALR 215: 176, 262, 263, 366, 368–370, 371D.4, 39854 Northern Securities Co. v. United States (1904) 193 US 197: 353 ∞ Parker v. British Airways [1982] 1 All ER 834: 133, 138, 139, 140, 1415.3, 1915.12, 284, 284–289, 39986 Peko-Wallsend Ltd v. Minister for Arts Heritage and Environment (1986) 70 ALR 523: 26225, 27351, 28268, 3033, 30613 Performing Right Society Ltd v. Bradford Corporation [1917–23] MacGillivray’s Copyright Cases 309: 146, 359 � Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd [1924] 1 KB 1: 143, 146, 223, 293, 294, 358, 359, 360D.2 � Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd [1924] 1 KB 762: 159, 165, 238, 320, 321, 360, 361, 363, 365D.3, 40271 ∞ Porter, Re; Re Transport Workers Union of Australia (1989) 34 IR 179: 163, 164, 1915.12, 192, 284, 296–302, 39383, 39799 � Price v. Grant Industries Pty Ltd (1978) 21 ALR 388: 157, 159, 242, 243, 327, 361, 363, 364, 365D.3 PRS v. Bradford : see Performing Right Society Ltd v. Bradford Corporation PRS v. Ciryl: see Performing Right Society Ltd v. Ciryl Theatrical Syndicate Ltd PRS v. Palais de Danse: see Performing Right Society Ltd v. Mitchell & Booker (Palais de Danse) Ltd Queensland v. Commonwealth of Australia (1977) 139 CLR 585: 37623 � Queensland Stations Pty Ltd v. Federal Commissioner of Taxation (1945) 70 CLR 539: 113, 117, 159, 241, 242, 322, 363, 364, 365, 365D.3, 366 R v. Hampden (1637) 3 How St Tr 825: 195 � RCA Corporation v. John Fairfax and Sons Ltd [1981] 1 NSWLR 251: 146, 149, 151–153, 222, 223, 290, 291, 292, 295, 357, 359, 360, 360D.2, 361, 39424 � Ready Mixed Concrete (South East) Ltd v. Minister of Pensions and National Insurance [1968] 2 QB 497: 125, 127, 159, 163, 164, 247, 297, 298–302, 315, 325, 351, 352, 362–364, 365D.3, 366, 39159 Ridge v. Baldwin [1964] AC 40: 179 SA v. O’Shea: see South Australia v. O’Shea ∞ Salemi v. MacKellar (No. 2) (1977) 137 CLR 396: 92, 171, 176, 182, 183, 189, 1915.12, 38897 Schmidt v. Secretary of State for Home Affairs [1969] 2 Ch 149: 182 � South Australia v. O’Shea (1987) 163 CLR 378: 176, 180, 182–184, 259, 260, 269, 280, 281, 309, 310, 367–369, 370, 371D.4, 39854 http:1915.12 http:1915.12 http:1915.12 http:1915.12 Cases [SOU–ZUI] 407 � South Staffordshire Water Co. v. Sharman [1896] 2 QB 44: 133, 136, 138, 140, 213, 214, 285–287, 288, 354–356, 356D.1, 3942 ∞ Stevens v. Brodribb Sawmilling Company Pty Ltd (1986) 160 CLR 16: 163, 164, 165, 169, 1915.12 � Stevenson Jordon and Harrison Ltd v. Macdonald and Evans [1952] 1 TLR 101: 127, 1294.8, 159, 165, 237, 245, 246, 314, 315, 323, 324, 361–363, 364, 365, 365D.3 Structural Dynamics Research Corporation v. Engineering Mechanics Research Corporation (1975) 401 F Supp 1102: 3895 Timmons v. McCall (1939): 131 ∞ Twist v. Council of Municipality of Randwick (1976) 136 CLR 106: 172, 176, 179, 1815.11, 182, 184, 189, 1915.12, 38897, 39724 � University of New South Wales v. Moorhouse (1975) 133 CLR 1: 143, 146, 148, 149, 152, 153, 2162, 219, 2194, 2902, 295, 29512, 357, 358, 359, 360D.2, 39424 UNSW v. Moorhouse: see University of New South Wales v. Moorhouse USM Corporation v. Marson Fastener Corporation (1979) 379 Mass 90: 3895 Vestry of St James and St John, Clerkenwell v. Feary (1890) 24 QBD 703: 179 Vigneux v. Canadian Performing Right Society Ltd [1945] AC 108: 148 WA v. Bropho: see Western Australia v. Bropho ∞ WEA International Inc. v. Hanimex Corporation Ltd (1987) AIPC ¶90-428: 149, 151, 1915.12, 39424, 39986 ∞ Western Australia v. Bropho (1991) 24 ALD 768: 172, 186, 187, 1915.12, 40012 Western Australia v. Commonwealth of Australia (1975) 134 CLR 201: 37623 � Winstone v. Wurlitzer Automatic Phonograph Co. of Australia Pty Ltd [1946] VLR 338: 146, 148, 220, 357, 358, 360D.2, 361 Wright v. Attorney-General for Tasmania (1954) 94 CLR 409: 163 � Zuijs v. Wirth Brothers Pty Ltd (1955) 93 CLR 561: 159, 231, 232, 325, 331, 361, 363, 364, 365D.3, 38522 http:1815.11 http:1915.12 http:1915.12 http:1915.12 Statutes Page numbers set in boldface type refer to statute extracts. Page numbers set in italics refer to SHYSTER output. “39530” refers to note 30 on page 395. Aboriginal Heritage Act 1972 (WA): 186 Coroner’s Act 1920 (WA): 254, 268, Administrative Decisions (Judicial 305, 308 Review) Act 1977 (Cth): 423 Criminal Law Consolidation Act, 1935 Animals Protection Act 1925 (Qld) (SA) s. 11(4): 185 s. 77a(7a): 260, 269, 281, 310 s. 19(2): 185 Customs Act 1966 (Cth) s. 273: 263 British Nationality Act 1981 (UK): 32, 33, 39, 427 Disorderly Houses Act 1943 (NSW) Copyright Act 1911 (UK): 142, 146 s. 3(1)(b): 255, 271 s. 1(2): 142, 143, 221, 222, 224, 290, Domestic Violence and Matrimonial 293 Proceedings Act 1976 (UK) s. 2(1): 225, 291 s. 13(2): 37766 s. 2(3): 239, 321, 39530 Insurance Act 1973 (Cth) s. 5(1)(b): 237, 246, 324 s. 62: 256 Copyright Act 1912 (Cth): 142, 146, s. 63: 256 220, 221, 225, 290, 291, 39530 Copyright Act 1956 (UK) Justices Act 1902 (NSW): 257, 310 s. 1(2): 148 Copyright Act 1968 (Cth): 140, 142, Law of Property Act 1965 (UK) 144, 146, 149, 155 s. 62: 212 s. 13(2): 140, 155, 223, 291 Local Courts Act 1982 (NSW): 257, 310 s. 36(1): 140, 142, 155, 219, 295 Local Government Act 1919 (NSW): s. 39a: 2194, 29512 39725 s. 101(1): 140, 142, 155 s. 317b: 172, 179 409 410 [LON–WOR] Statutes Long Service Leave Act, 1967 (SA): 244, 327, 351 Migration Act 1958 (Cth): 182, 255, 270, 278 s. 12: 253, 268, 277 s. 18: 182 Municipal Councils Ordinance 1947 (Ceylon) s. 277(1): 259 National Insurance Act 1965 (UK) s. 1(2): 248, 297, 325, 352 National Parks and Wildlife Act 1974 (Cth) s. 134(1): 184 s. 135: 184 National Parks and Wildlife Conservation Act 1975 (Cth) s. 8b: 262, 273, 282, 303, 305 Pay-roll Tax Act 1971 (NSW): 161 Pay-roll Tax Assessment Act 1941 (Cth): 234, 242, 322, 323 Prices Regulation Act 1948 (NSW): 261, 270, 307 Racing and Gaming Act 1952 (Tas) s. 39(3): 278 Supplementary Benefits Act 1976 (UK): 32 Trade Practices Act 1974 (Cth) s. 45d: 164 Trade Union and Labour Relations Act 1974 (UK): 245, 326 Worker’s Compensation Act 1926 (NSW): 232, 325, 331 Worker’s Compensation Act 1928 (Vic) s. 3: 241, 323, 347 Worker’s Compensation Act 1958 (Vic): 252, 267, 276 World Heritage Properties Conservation Act 1983 (Cth) s. 3(2): 262, 273, 281, 303, 305 Bibliography Aldenderfer, Mark S. and Blashfield, Roger K. 1984, Cluster Analysis, Sage Uni­ versity Paper Series on Quantitative Applications in the Social Sciences, no. 07-044, Sage, Beverly Hills. ISBN 0 8039 2376 7. Allen, Layman E. 1963, “Beyond document retrieval toward information retrieval”, Minnesota Law Review, vol. 47, pp. 713–67. ISSN 0026-5535. Allen, Layman E. 1965, “Usefulness of modern logic to the readers and writers of legal documents”, in Communication Sciences and Law: Reflections from the Jur­ imetrics Conference, eds Layman E. Allen and Mary E. Caldwell, Bobbs-Merrill, Indianapolis, ch. 13, pp. 83–112. From the Conference on the Implications of Developments in the Communication Sciences for Legal Education in the Next Decade of the Jurimetrics Committee of the Association of American Law Schools, Yale Law School, New Haven, Connecticut, 5–7 September 1963. Allen, Layman E. 1968, “A language-normalization approach to information retrieval in law”, Jurimetrics Journal, vol. 9, no. 1, September, pp. 41–56. ISSN 0022-6793. Allen, Layman E. 1981, “Analysis of law by symbolic logic”, in Computers and the Law: An Introductory Handbook (third edition), ed. Robert P. Bigelow, American Bar Association, pp. 84–90. Allen, Layman Edward 1982, “Towards a normalized language to clarify the struc­ ture of legal discourse”, in Deontic Logic, Computational Linguistics and Legal In­ formation Systems, ed. Antonio A. Martino, vol. II, North-Holland, Amsterdam, pp. 349–407. ISBN 0 444 86415 6, 0 444 86413 X (set). From the International Conference on “Logic, Informatics, Law”, Florence, Italy, April 1981. Allen, Layman E., Brooks, Robin B. S. and James, Patricia A. 1962, Automatic Retrieval of Legal Literature: Why and How, Walter E. Meyer Research Institute of Law, New Haven, Connecticut. 411 412 [ALL–ASH] Bibliography Allen, Layman E. and Caldwell, Mary Ellen 1963, “Modern logic and judicial decision making: A sketch of one view”, Law and Contemporary Problems, vol. 28, no. 1, Winter, pp. 213–70. ISSN 0023-9186. Republished in Baade 1963. Allen, Layman E. and Saxon, Charles S. 1985, “Computer aided normalizing and unpacking: Some interesting machine-processable transformations of legal rules”, in Computing Power and Legal Reasoning, ed. Charles Walter, West Publishing Company, St Paul, ch. 20, pp. 495–572. ISBN 0 314 95570 4. Allen, Layman and Saxon, Charles 1988, “Exploring computer-aided generation of questions for normalizing legal rules”, in Computer Power and Legal Language: The Use of Computational Linguistics, Artificial Intelligence, and Expert Sys­ tems in the Law, ed. Charles Walter, Quorum, New York, ch. 19, pp. 243–316. ISBN 0 89930 306 4. From the Second Annual Conference on Law and Technology, University of Houston, 24–28 June 1985. Ashley, Kevin D. 1985, “Reasoning by analogy: A survey of selected AI research with implications for legal expert systems”, in Computing Power and Legal Reas­ oning, ed. Charles Walter, West Publishing Company, St Paul, ch. 6, pp. 105–27. ISBN 0 314 95570 4. Ashley, Kevin D. 1989a, “Toward a computational theory of arguing with precedents: Accomodating [sic] multiple interpretations of cases”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 93–102. ISBN 0 89791 322 1. Ashley, Kevin D. 1989b, “Defining salience in case-based arguments”, in Proceedings of the Eleventh International Joint Conference on Artificial Intelligence (IJCAI-89), Detroit, Michigan, 20–25 August, vol. 1, pp. 537–42. ISBN 1 55860 094 9. ISSN 1045-0823. Ashley, Kevin D. 1990, Modeling Legal Argument: Reasoning with Cases and Hypo­ theticals, Artificial Intelligence and Legal Reasoning Series, MIT Press (Bradford), Cambridge, Massachusetts. ISBN 0 262 01114 X. Ashley, Kevin D. and Aleven, Vincent 1991, “Toward an intelligent tutoring system for teaching law students to argue with cases”, in Proceedings of the Third Inter­ national Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 42–52. ISBN 0 89791 399 X. Ashley, Kevin D. and Rissland, Edwina L. 1986, “Toward modelling legal argument”, in Automated Analysis of Legal Texts: Logic, Informatics, Law, eds Antonio A. Martino and Fiorenza Socci Natali, Elsevier Science (North-Holland), Amsterdam, pp. 19–30. ISBN 0 444 70111 7. From the Second International Conference on ‘Logic, Informatics, Law’, Florence, Italy, September 1985. Bibliography [ASH–BEN] 413 Ashley, Kevin D. and Rissland, Edwina L. 1987a, “But, see, accord: Generating ‘blue book’ citations in HYPO”, in Proceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 67–74. ISBN 0 89791 230 6. Ashley, Kevin D. and Rissland, Edwina L. 1987b, “Compare and contrast, a test of expertise”, in Proceedings of the Sixth National Conference on Artificial Intelligence (AAAI-87), Seattle, Washington, 13–17 July, vol. 1, pp. 273–8. ISBN 0 934613 42 7. Ashley, Kevin D. and Rissland, Edwina L. 1988, “Waiting on weighting: A symbolic least commitment approach”, in Proceedings of the Seventh National Conference on Artificial Intelligence (AAAI-88), St Paul, Minnesota, 21–26 August, vol. 1, pp. 239– 44. ISBN 0 929280 00 8. Aubert, Vilhelm 1963, “Conscientious objectors before Norwegian military courts”, in Judicial Decision-Making, ed. Glendon Schubert, International Yearbook of Polit­ ical Behavior Research, vol. 4, Free Press of Glencoe, New York, ch. vii, pp. 201–19. Austin, John 1885, Lectures on Jurisprudence or The Philosophy of Positive Law (fifth edition), John Murray, London. Two volumes, revised and edited by Robert Campbell. Baade, Hans W. (ed.) 1963, Jurimetrics, Basic Books, New York. Originally published as Law and Contemporary Problems, vol. 28, no. 1, Winter 1963. Bain, William M. 1986, “A case-based reasoning system for subjective assessment”, in Proceedings of the Fifth National Conference on Artificial Intelligence (AAAI-86), Philadelphia, Pennsylvania, 11–15 August, vol. 1, pp. 523–7. ISBN 0 934613 13 3. Bankowski, Zenon and MacCormick, D. Neil 1991, “Statutory interpretation in the United Kingdom”, in Interpreting Statutes: A Comparative Study, eds D. Neil Mac- Cormick and Robert S. Summers, Dartmouth Series in Applied Legal Philosophy, Dartmouth, Aldershot, ch. 10, pp. 359–406. ISBN 1 85521 183 1. Bench-Capon, T. J. M. 1987, “Support for policy makers: Formulating legislation with the aid of logical models”, in Proceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 181–9. ISBN 0 89791 230 6. Bench-Capon, T. J. M. 1989, “Deep models, normative reasoning and legal expert systems”, in Proceedings of the Second International Conference on Artificial In­ telligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 37–45. ISBN 0 89791 322 1. Bench-Capon, T. J. M. (ed.) 1991, Knowledge-Based Systems and Legal Applications, APIC Series, no. 36, Academic Press, London. ISBN 0 12 086441 X. Bench-Capon, T. J. M., Robinson, G. O., Routen, T. W. and Sergot, M. J. 1987, “Logic programming for large scale applications in law: A formalisation of supple­ mentary benefit legislation”, in Proceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 190–8. ISBN 0 89791 230 6. 414 [BEN–BLA] Bibliography Bench-Capon, Trevor and Sergot, Marek 1988, “Toward a rule-based representa­ tion of open texture in law”, in Computer Power and Legal Language: The Use of Computational Linguistics, Artificial Intelligence, and Expert Systems in the Law, ed. Charles Walter, Quorum, New York, ch. 6, pp. 39–60. ISBN 0 89930 306 4. From the Second Annual Conference on Law and Technology, University of Houston, 24–28 June 1985. Berman, Donald H. 1991, “Developer’s choice in the legal domain: The Sisyphean journey with CBR or down hill with rules”, in Proceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 307–9. ISBN 0 89791 399 X. A working paper for the Case–Rules Panel. Berman, Donald H. and Hafner, Carole D. 1991, “Incorporating procedural context into a model of case-based legal reasoning”, in Proceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 12–20. ISBN 0 89791 399 X. Bing, Jon 1980, “Legal norms, discretionary rules and computer programs”, in Com­ puter Science and Law, ed. Bryan Niblett, Cambridge University Press, Cambridge, ch. 7, pp. 119–36. ISBN 0 521 23451 4. Bing, Jon (ed.) 1984a, Handbook of Legal Information Retrieval, Elsevier Science (North-Holland), Amsterdam. ISBN 0 444 87576 X. Bing, Jon 1984b, “Legal information services: Some trends and characteristics”, in Data Processing and the Law, ed. Colin Campbell, Sweet and Maxwell, London, pp. 29–45. ISBN 0 421 32000 1. Bing, Jon 1987, “Designing text retrieval systems for ‘conceptual searching’ ”, in Pro­ ceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 43–51. ISBN 0 89791 230 6. Bing, Jon 1989, “The law of the books and the law of the files: Possibilities and problems of legal information systems”, in Advanced Topics of Law and Information Technology, ed. G. P. V. Vandenberghe, Computer/Law Series, no. 3, Kluwer Law and Taxation Publishers, Deventer, pp. 151–82. ISBN 90 6544 391 6. Bing, Jon 1992, [Review of Ashley 1990], Artificial Intelligence and Law, vol. 1, no. 1, pp. 103–7. ISSN 0924-8463. Bing, Jon and Selmer, Knut S. (eds) 1980, A Decade of Computers and Law, Uni­ versitetsforlaget, Oslo. ISBN 82 00 05376 8. Bjarup, Jes 1988, “Kripke’s case: Some remarks on rules, their interpretation and application”, Rechtstheorie, vol. 19, pp. 39–49. ISSN 0034-1398. Blackstone, William 1773, Commentaries on the Laws of England (fifth edition), Oxford (Clarendon). Four volumes. Bibliography [BOD–CIE] 415 Boden, Margaret A. 1985, “Panel: Artificial intelligence and legal responsibility”, in Proceedings of the Ninth International Joint Conference on Artificial Intel­ ligence (IJCAI-85), Los Angeles, California, 18–23 August, vol. 2, pp. 1267–8. ISBN 0 934613 02 8. ISSN 1045-0823. Borchgrevink, Mette and Hansen, Johannes 1980, “SARA: A system for the analysis of legal decisions”, in A Decade of Computers and Law, eds Jon Bing and Knut S. Selmer, Universitetsforlaget, Oslo, ch. 18, pp. 342–75. ISBN 82 00 05376 8. Boswell, James 1960?, The Life of Samuel Johnson, vol. I, Heron Books, London. First published in 1791. Bottomley, Stephen, Gunningham, Neil and Parker, Stephen 1991, Law in Con­ text, Federation Press, Leichhardt. ISBN 1 86287 061 6. Branting, L. Karl 1989, “Representing and reusing explanations of legal precedents”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 103– 10. ISBN 0 89791 322 1. Branting, L. Karl 1991, “Reasoning with portions of precedents”, in Proceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 145–54. ISBN 0 89791 399 X. Brooks, Adrian 1990, “Employment”, in Australian Commentary on Halsbury’s Laws of England (fourth edition), ed. Garfield Barwick, Butterworths, Sydney, ch. 56. ISBN 0 409 30930 3. Buchanan, Bruce G. and Headrick, Thomas E. 1970, “Some speculation about artifi­ cial intelligence and legal reasoning”, Stanford Law Review, vol. 23, no. 1, November, pp. 40–62. ISSN 0038-9765. Buchanan, Bruce G. and Shortliffe, Edward H. (eds) 1984, Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, Addison-Wesley Series in Artificial Intelligence, Addison-Wesley, Reading, Mas­ sachusetts. ISBN 0 201 10172 6. Burton, Steven J. 1985, An Introduction to Law and Legal Reasoning, Little, Brown and Company, Boston. ISBN 0 316 11786 2. Campbell, J. A. 1984, “The expert computer and professional negligence: who is liable?”, in Artificial Intelligence: Human Effects, eds Masoud Yazdani and Ajit Narayanan, Ellis Horwood Series in Artificial Intelligence, Ellis Horwood, Chichester, pp. 37–51. ISBN 0 85312 577 5, 0 470 20092 8. Capper, Phillip and Susskind, Richard E. 1988, Latent Damage Law—The Expert System, Butterworths, London. ISBN 0 406 02362 X. Carroll, Lewis 1982, The Complete Illustrated Works of Lewis Carroll, Avenel Books, New York. ISBN 0 517 38566 X. Ciesielski, Victor B. 1990, Some Experiments in the Use of Expert Systems for Le­ gislation and Regulations, Technical Report 21, Key Centre for Knowledge Based Systems, Departments of Computer Science, Royal Melbourne Institute of Techno­ logy and The University of Melbourne, 27 April. 416 [CLA–DEB] Bibliography Clark, Andrew 1988, [Review of Susskind 1987a], Journal of Law and Society, vol. 15, no. 4, Winter, pp. 426–30. ISSN 0263-323X. Clark, Jon and Wedderburn, Kenneth 1983, “Modern labour law: Problems, func­ tions, and policies”, in Labour Law and Industrial Relations: Building on Kahn- Freund, eds Kenneth Wedderburn, Roy Lewis and Jon Clark, Oxford University Press (Clarendon), Oxford, ch. 6, pp. 127–242. ISBN 0 19 825393 1. Clarke, Roger 1988, “Legal aspects of knowledge-based technology”, Journal of In­ formation Technology, vol. 3, no. 1, March, pp. 9–16. ISSN 0268-3962. Claudy, John G. 1972, “A comparison of five variable weighting procedures”, Edu­ cational and Psychological Measurement, vol. 32, no. 2, Summer, pp. 311–22. ISSN 0013-1644. Conover, W. J. 1971, Practical Nonparametric Statistics, John Wiley, New York. ISBN 0 471 16851 3. Cook, Sandra, Hafner, Carole D., McCarty, L. Thorne, Meldman, Jeffrey A., Peterson, Mark, Sprowl, James A., Sridharan, N. S. and Waterman, D. A. 1981, “The applications of artificial intelligence to law: A survey of six current projects”, AFIPS Conference Proceedings, vol. 50, pp. 689–96. ISSN 0449-1173. From the National Computer Conference, Chicago, Illinois, 4–7 May 1981. Cover, T. M. and Hart, P. E. 1967, “Nearest neighbor pattern classification”, IEEE Transactions on Information Theory, vol. IT-13, no. 1, January, pp. 21–7. ISSN 0018-9448. Creighton, W. B., Ford, W. J. and Mitchell, R. J. 1983, Labour Law: Materials and Commentary, Law Book Company, Sydney. ISBN 0 455 20413 6. Cross, Rupert 1961, Precedent in English Law, Clarendon Law Series, Oxford Uni­ versity Press (Clarendon), London. Croydon, Stanley O., Jr 1980, “JURIS: A tool for legal research”, in Legal and Legis­ lative Information Processing, ed. Beth Krevitt Eres, Greenwood Press, Westport, Connecticut, ch. 11, pp. 163–72. ISBN 0 313 21343 7. D’Amato, Anthony 1977, “Can/should computers replace judges?”, Georgia Law Re­ view, vol. 11, no. 5, September, pp. 1277–1301. ISSN 0016-8300. deBessonet, Cary G. and Cross, George R. 1985, “Representation of some aspects of legal causality”, in Computing Power and Legal Reasoning, ed. Charles Walter, West Publishing Company, St Paul, ch. 11, pp. 205–14. ISBN 0 314 95570 4. deBessonet, Cary G. and Cross, George R. 1986, “Conceptual retrieval and legal decision-making”, in Automated Analysis of Legal Texts: Logic, Informatics, Law, eds Antonio A. Martino and Fiorenza Socci Natali, Elsevier Science (North- Holland), Amsterdam, pp. 219–27. ISBN 0 444 70111 7. From the Second International Conference on ‘Logic, Informatics, Law’, Florence, Italy, September 1985. Bibliography [DEB–GAL] 417 deBessonet, Cary and Cross, George 1988, “Distinguishing legal language-types for conceptual retrieval”, in Computer Power and Legal Language: The Use of Com­ putational Linguistics, Artificial Intelligence, and Expert Systems in the Law, ed. Charles Walter, Quorum, New York, ch. 14, pp. 155–66. ISBN 0 89930 306 4. From the Second Annual Conference on Law and Technology, University of Houston, 24–28 June 1985. Dick, Judith P. 1991, “Representation of legal text for conceptual retrieval”, in Pro­ ceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 244–53. ISBN 0 89791 399 X. Douglas, William O. 1948, “The dissent: A safeguard of democracy”, Journal of the American Judicature Society, vol. 32, December, pp. 104–7. ISSN 0022-5800. Drahos, Peter and Parker, Stephen 1992, “Rule following, rule scepticism and in­ determinacy in law: A conventional account”, Ratio Juris, vol. 5, no. 1, March, pp. 109–19. ISSN 0952-1917. Dworkin, Ronald 1977, Taking Rights Seriously, Duckworth, London. ISBN 0 7156 0715 4. Dworkin, Ronald 1978, “No right answer?”, New York University Law Review, vol. 53, no. 1, April, pp. 1–32. ISSN 0028-7881. Eliot, T. S. 1963, Collected Poems 1909–1962, Faber and Faber, London. ISBN 0 571 05549 4. Everitt, Brian 1974, Cluster Analysis, Social Science Research Council Reviews of Current Research, no. 11, Heinemann Educational Books, London. ISBN 0 435 82297 7. Feigenbaum, E. A. 1981, “Expert systems in the 1980s”, in Machine Intelligence, ed. A. H. Bond, Infotech State of the Art Report Series 9, no. 3, Pergamon Infotech, Maidenhead, Berkshire, pp. 219–29. ISBN 008 028556 2. ISSN 0734-8487. Feigenbaum, Edward A. and McCorduck, Pamela 1983, The Fifth Generation: Ar­ tificial Intelligence and Japan’s Computer Challenge to the World, Addison-Wesley, Reading, Massachusetts. ISBN 0 201 11519 0. Fisher, Ronald A. 1970, Statistical Methods for Research Workers (fourteenth edi­ tion), Hafner, New York. Fix, E. and Hodges, J. L., Jr 1951, Discriminatory Analysis, Non-parametric Discrim­ ination, Project 21-49-004, Report 4, United States Air Force School of Aviation Medicine, Randolph Field, Texas, February. Frank, Jerome 1930, Law and the Modern Mind, Brentano’s, New York. Frank, Jerome 1949, Courts on Trial: Myth and Reality in American Justice, Prince- ton University Press, Princeton, New Jersey. Galsworthy, John 1941, Ten Famous Plays, Duckworth, London. 418 [GAR–GRE] Bibliography Gardner, Anne von der Lieth 1983, “The design of a legal analysis program”, in Proceedings of the National Conference on Artificial Intelligence (AAAI-83), Wash­ ington, DC, 22–26 August, pp. 114–18. ISBN 0 86576 065 9. Gardner, Anne von der Lieth 1984, An Artificial Intelligence Approach to Legal Reas­ oning, PhD Thesis STAN-CS-85-1045, Department of Computer Science, Stanford University, Stanford, California, June. Gardner, Anne v. d. L. 1985, “Overview of an artificial intelligence approach to legal reasoning”, in Computing Power and Legal Reasoning, ed. Charles Walter, West Publishing Company, St Paul, ch. 13, pp. 247–74. ISBN 0 314 95570 4. Gardner, Anne von der Lieth 1987, An Artificial Intelligence Approach to Legal Reas­ oning, Artificial Intelligence and Legal Reasoning Series, MIT Press (Bradford), Cambridge, Massachusetts. ISBN 0 262 07104 5. Gay, John 1967, Fables, William Andrews Clark Memorial Library, University of Cali­ fornia, Los Angeles. Reprint of the 1727 (vol. I) and 1738 (vol. II) editions. Gilbert, W. S. 1977, The Savoy Operas: Being the Complete Text of the Gilbert and Sullivan Operas as originally produced in the years 1875–1896, Macmillan, London. ISBN 0 333 06901 3. Gold, David I. and Susskind, Richard E. 1986, “Expert systems in law: A jurispru­ dential and formal specification approach”, in Automated Analysis of Legal Texts: Logic, Informatics, Law, eds Antonio A. Martino and Fiorenza Socci Natali, Elsevier Science (North-Holland), Amsterdam, pp. 625–42. ISBN 0 444 70111 7. From the Second International Conference on ‘Logic, Informatics, Law’, Florence, Italy, September 1985. Goldman, Seth R., Dyer, Michael G. and Flowers, Margot 1987, “Precedent-based legal reasoning and knowledge acquisition in contract law: A process model”, in Proceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 210–21. ISBN 0 89791 230 6. Goldman, Sheldon and Sarat, Austin (eds) 1978, American Court Systems: Readings in Judicial Process and Behavior, W. H. Freeman, San Francisco. ISBN 0 7167 0061 1. Greenleaf, G. W., Mowbray, A. S. and Lewis, D. P. 1988, Australasian Computer­ ised Legal Information Handbook, Butterworths, Sydney. ISBN 0 409 49445 3. Greenleaf, Graham, Mowbray, Andrew and Tyree, Alan L. 1987, “Expert systems in law: The DataLex project”, in Proceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 9–17. ISBN 0 89791 230 6. Greenleaf, Graham, Mowbray, Andrew and Tyree, Alan 1991, “The DataLex legal workstation—integrating tools for lawyers”, in Proceedings of the Third Inter­ national Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 215–24. ISBN 0 89791 399 X. Bibliography [GRU–KAY] 419 Grunbaum, W. F. 1988, [Review of Gardner 1987], Computing Reviews, vol. 29, no. 12, December, p. 639. ISSN 0010-4884. Haar, Charles M., Sawyer, John P., Jr and Cummings, Stephen J. 1977, “Computer power and legal reasoning: A case study of judicial decision prediction in zoning amendment cases”, American Bar Foundation Research Journal, no. 3, Summer, pp. 651–768. ISSN 0361-9486. Hafner, Carole D. 1981, “Representation of knowledge in a legal information re­ trieval system”, in Information Retrieval Research, eds R. N. Oddy, S. E. Robertson, C. J. van Rijsbergen and P. W. Williams, Butterworths, London, ch. 9, pp. 139–53. ISBN 0 408 10775 8. From a symposium “Research and Development in Information Retrieval”, St John’s Col­ lege, Cambridge, June 1980. Hafner, Carole D. 1987, “Conceptual organization of case law knowledge bases”, in Proceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 35–42. ISBN 0 89791 230 6. Harris, J. W. 1980, Legal Philosophies, Butterworths, London. ISBN 0 406 59361 2. Hart, H. L. A. 1961, The Concept of Law, Clarendon Law Series, Oxford University Press (Clarendon), London. Hart, H. L. A. 1983, Essays in Jurisprudence and Philosophy, Oxford University Press (Clarendon), London. ISBN 0 19 825387 7. Hayes-Roth, Frederick, Waterman, Donald A. and Lenat, Douglas B. (eds) 1983, Building Expert Systems, Teknowledge Series in Knowledge Engineering, Addison- Wesley, Reading, Massachusetts. ISBN 0 201 10686 8. Hepple, Bob 1986, “Restructuring employment rights”, The Industrial Law Journal, vol. 15, pp. 69–83. ISBN 0 421 37340 7. Herman, Theodor 1980, “WESTLAW: Computerized legal research program of West Publishing Company”, in Legal and Legislative Information Processing, ed. Beth Krevitt Eres, Greenwood Press, Westport, Connecticut, ch. 10, pp. 157–61. ISBN 0 313 21343 7. Holmes, O. W. 1897, “The path of the law”, Harvard Law Review, vol. 10, no. 8, 25 March, pp. 457–78. ISSN 0017-811X. Johnson, Peter and Mead, David 1991, “Legislative knowledge base systems for public administration—Some practical issues”, in Proceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 108–17. ISBN 0 89791 399 X. Johnson, Samuel 1823, The Works of Samuel Johnson, vol. III, Rivington et al., Lon­ don. Jones, R. 1988, [Review of Gardner 1987], Computers and Law, no. 56, June, p. 32. ISSN 0140-3249. Kayton, Irving 1964, “Can jurimetrics be of value to jurisprudence?”, The George Washington Law Review, vol. 33, no. 1, October, pp. 287–317. ISSN 0016-8076. 420 [KEL–LAM] Bibliography Kelman, Mark 1987, A Guide to Critical Legal Studies, Harvard University Press, Cambridge, Massachusetts. ISBN 0 674 36755 3. Kelsen, Hans 1945, General Theory of Law and State, 20th Century Legal Philosophy Series, vol. I, Russell and Russell, New York. ISBN 0 8462 0215 8. Kelsen, Hans 1967, Pure Theory of Law, University of California Press, Berkeley. Translation of Reine Rechtslehre (1960, second revised edition). Translated by Max Knight. Kernighan, Brian W. and Ritchie, Dennis M. 1988, The C Programming Language (second edition), Prentice Hall, Englewood Cliffs, New Jersey. ISBN 0 13 110362 8. Knuth, Donald E. 1984, The TEXbook, Addison-Wesley, Reading, Massachusetts. ISBN 0 201 13448 9. Kort, Fred 1963a, “Content analysis of judicial opinions and rules of law”, in Judi­ cial Decision-Making, ed. Glendon Schubert, International Yearbook of Political Behavior Research, vol. 4, Free Press of Glencoe, New York, ch. vi, pp. 133–97. Kort, Fred 1963b, “Simultaneous equations and boolean algebra in the analysis of judicial decisions”, Law and Contemporary Problems, vol. 28, no. 1, Winter, pp. 143– 63. ISSN 0023-9186. Republished in Baade 1963. Kowalski, Andrzej 1991, “Case-based reasoning and the deep structure approach to knowledge representation”, in Proceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 21–30. ISBN 0 89791 399 X. Kowalski, Robert 1989, “The treatment of negation in logic programs for represent­ ing legislation”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 11–15. ISBN 0 89791 322 1. Kowalski, Robert and Sergot, Marek 1985, “Computer representation of the law”, in Proceedings of the Ninth International Joint Conference on Artificial Intel­ ligence (IJCAI-85), Los Angeles, California, 18–23 August, vol. 2, pp. 1269–70. ISBN 0 934613 02 8. ISSN 1045-0823. Kowalski, Robert and Sergot, Marek 1990, “The use of logical models in legal problem solving”, Ratio Juris, vol. 3, no. 2, July, pp. 201–18. ISSN 0952-1917. Kripke, Saul A. 1982, Wittgenstein on Rules and Private Language: An Elementary Exposition, Basil Blackwell, Oxford. ISBN 0 631 13077 2. Krygier, Martin 1986, “Julius Stone: Leeways of choice, legal tradition and the de­ claratory theory of law”, University of New South Wales Law Journal, vol. 9, no. 2, pp. 26–38. ISSN 0313-0096. Lambert, Kenneth A. and Grunewald, Mark H. 1989, “LESTER: Using paradigm cases in a quasi-precedential legal domain”, in Proceedings of the Second Inter­ national Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 87–92. ISBN 0 89791 322 1. Bibliography [LAM–LLE] 421 Lambert, Kenneth A. and Grunewald, Mark H. 1991, “Legal theory and case-based reasoners: The importance of context and the process of focusing”, in Proceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 191–5. ISBN 0 89791 399 X. Lamport, Leslie 1986, LATEX: A Document Preparation System, Addison-Wesley, Read­ ing, Massachusetts. ISBN 0 201 15790 X. Lasswell, Harold D. 1955, “Current studies of the decision process: Automation versus creativity”, The Western Political Quarterly, vol. 8, no. 3, September, pp. 381–99. ISSN 0043-4078. Lawlor, Reed C. 1963, “What computers can do: Analysis and prediction of judicial decisions”, American Bar Association Journal, vol. 49, no. 4, April, pp. 337–44. ISSN 0002-7596. Lawlor, Reed C. 1968, “Personal stare decisis”, Southern California Law Review, vol. 41, pp. 73–118. ISSN 0038-3910. Lawlor, Reed C. 1972, “Fact content of cases and precedent—A modern theory of precedent”, Jurimetrics Journal, vol. 12, no. 4, June, pp. 245–70. ISSN 0022-6793. Excerpts from a paper presented at the September 1971 Annual Meeting of the American Political Science Association. Lawlor, Reed C. 1981, “Analysis and prediction of judicial decisions”, in Computers and the Law: An Introductory Handbook (third edition), ed. Robert P. Bigelow, American Bar Association, pp. 81–4. Leith, Philip 1985a, “An IKBS implementation”, Software—Practice and Experience, vol. 15, no. 1, January, pp. 65–86. ISSN 0038-0644. Leith, Philip 1985b, Law and Computer Program: The Limits to Logic, Paper, De­ partment of Jurisprudence, Queen’s University, Belfast, March. Leith, Philip 1986a, “Clear rules and legal expert systems”, in Automated Analysis of Legal Texts: Logic, Informatics, Law, eds Antonio A. Martino and Fiorenza Socci Natali, Elsevier Science (North-Holland), Amsterdam, pp. 661–79. ISBN 0 444 70111 7. From the Second International Conference on ‘Logic, Informatics, Law’, Florence, Italy, September 1985. Leith, Philip 1986b, [Review of Michie and Johnston 1984, and Yazdani and Naray­ anan 1984], Northern Ireland Legal Quarterly, vol. 37, no. 1, Spring, pp. 97–100. ISSN 0029-3105. Leith, Philip 1987, “The Emperor’s new expert system”, The Modern Law Review, vol. 50, no. 1, January, pp. 128–32. ISSN 0026-7961. Levi, Edward H. 1949, An Introduction to Legal Reasoning, University of Chicago Press (Phoenix), Chicago. Llewellyn, Karl N. 1931, “Some realism about realism—Responding to Dean Pound”, Harvard Law Review, vol. 44, no. 8, June, pp. 1222–64. ISSN 0017-811X. 422 [LLE–MCC] Bibliography Llewellyn, K. N. 1951, The Bramble Bush: On Our Law and its Study, Oceana Pub­ lications, New York. Llewellyn, Karl N. 1962, Jurisprudence: Realism in Theory and Practice, University of Chicago Press, Chicago. Loevinger, Lee 1949, “Jurimetrics: The next step forward”, Minnesota Law Review, vol. 33, no. 5, April, pp. 455–93. ISSN 0026-5535. Loevinger, Lee 1961, “Jurimetrics: Science and prediction in the field of law”, Min­ nesota Law Review, vol. 46, no. 2, December, pp. 255–75. ISSN 0026-5535. Loevinger, Lee 1963, “Jurimetrics: The methodology of legal inquiry”, Law and Con­ temporary Problems, vol. 28, no. 1, Winter, pp. 5–35. ISSN 0023-9186. Republished in Baade 1963. Loevinger, Lee 1964, “Jurimetrics”, in Judicial Behavior: A Reader in Theory and Re­ search, ed. Glendon Schubert, Rand McNally Political Science Series, Rand McNally and Company, Chicago, ch. I.C.7, pp. 72–6. Extracts from Loevinger 1949. Lucas, E. V. (ed.) 1935, The Letters of Charles Lamb to which are added those of his sister Mary Lamb, vol. III, Dent and Methuen, London. McCarty, L. Thorne 1977, “Reflections on TAXMAN: An experiment in artificial intelligence and legal reasoning”, Harvard Law Review, vol. 90, no. 5, March, pp. 837– 93. ISSN 0017-811X. McCarty, L. Thorne 1980a, Some Requirements for a Computer-Based Legal Con­ sultant, Technical Report LRP-TR-8, Laboratory for Computer Science Research, Rutgers University, New Brunswick, New Jersey, 1 July. Published, with minor abbreviations, as McCarty 1980b. McCarty, L. Thorne 1980b, “Some requirements for a computer-based legal consult­ ant”, in Proceedings of the First Annual National Conference on Artificial Intelli­ gence (AAAI-80), Stanford University, 18–21 August, pp. 298–300. McCarty, L. Thorne 1980c, “The TAXMAN project: Towards a cognitive theory of legal argument”, in Computer Science and Law, ed. Bryan Niblett, Cambridge University Press, Cambridge, ch. 3, pp. 23–43. ISBN 0 521 23451 4. McCarty, L. Thorne 1982, “A computational theory of Eisner v. Macomber”, in Ar­ tificial Intelligence and Legal Information Systems, ed. Constantino Ciampi, vol. I, North-Holland, Amsterdam, pp. 329–55. ISBN 0 444 86414 8, 0 444 86413 X (set). From the International Conference on “Logic, Informatics, Law”, Florence, Italy, April 1981. McCarty, L. Thorne 1983, “Intelligent legal information systems: Problems and prospects”, Rutgers Computer and Technology Law Journal, vol. 9, no. 2, pp. 265–94. ISSN 0735-8938. Republished as McCarty 1984. McCarty, L. Thorne 1984, “Intelligent legal information systems: Problems and prospects”, in Data Processing and the Law, ed. Colin Campbell, Sweet and Max­ well, London, pp. 125–51. ISBN 0 421 32000 1. Bibliography [MCC–MEL] 423 McCarty, L. Thorne 1986, “Permissions and obligations: An informal introduction”, in Automated Analysis of Legal Texts: Logic, Informatics, Law, eds Antonio A. Martino and Fiorenza Socci Natali, Elsevier Science (North-Holland), Amsterdam, pp. 307–37. ISBN 0 444 70111 7. From the Second International Conference on ‘Logic, Informatics, Law’, Florence, Italy, September 1985. McCarty, L. Thorne 1989, “A language for legal discourse: I. Basic features”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 180– 9. ISBN 0 89791 322 1. McCarty, L. Thorne 1990, “Artificial intelligence and law: How to get there from here”, Ratio Juris, vol. 3, no. 2, July, pp. 189–200. ISSN 0952-1917. McCarty, L. Thorne 1991, “On the role of prototypes in appellate legal argument (abstract)”, in Proceedings of the Third International Conference on Artificial Intel­ ligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 185– 90. ISBN 0 89791 399 X. McCarty, L. Thorne and Sridharan, N. S. 1981, “The representation of an evolving system of legal concepts: II. Prototypes and deformations”, in Proceedings of the Seventh International Joint Conference on Artificial Intelligence (IJCAI-81), Uni­ versity of British Columbia, Vancouver, British Columbia, Canada, 24–28 August, vol. I, pp. 246–53. ISBN 0 86576 059 4. ISSN 1045-0823. MacCormick, Neil 1978, Legal Reasoning and Legal Theory, Clarendon Law Series, Oxford University Press (Clarendon), Oxford. ISBN 0 19 876080 9. Mackaay, Ejan and Robillard, Pierre 1974, “Predicting judicial decisions: The nearest neighbour rule and visual representation of case patterns”, Datenverarbei­ tung im Recht, vol. 3, no. 3/4, November, pp. 302–31. ISSN 0301-2980. McMillan, John 1991, “Developments under the ADJR Act: The grounds of review”, Federal Law Review, vol. 20, no. 1, pp. 50–82. ISSN 0067-205X. Revised version of a paper presented at the Conference “Ten Years of the Federal Admin­ istrative Decisions (Judicial Review) Act”, Canberra, September 1990. Maggs, Peter B. and deBessonet, Cary G. 1972, “Automated logical analysis of systems of legal rules”, Jurimetrics Journal, vol. 12, no. 3, March, pp. 158–69. ISSN 0022-6793. Mehl, Lucien 1959, “Automation in the legal world: From the machine processing of legal information to the ‘law machine’ ”, in Mechanisation of Thought Processes, vol. II, Her Majesty’s Stationery Office, London, pp. 755–87. From a symposium held at the National Physical Laboratory, 24–27 November 1958. Meldman, Jeffrey A. 1977, “A structural model for computer-aided legal analysis”, Rutgers Journal of Computers and the Law, vol. 6, no. 1, pp. 27–71. ISSN 0735-8938. 424 [MEN–NAG] Bibliography Mendelson, Simon 1989, “An attempted dimensional analysis of the law governing government appeals in criminal cases”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 128–37. ISBN 0 89791 322 1. Michaelsen, Robert and Michie, Donald 1983, “Expert systems in business”, Datamation, vol. 29, no. 11, November, pp. 240–6. ISSN 0011-6963. Michie, Donald and Johnston, Rory 1984, The Creative Computer: Machine Intel­ ligence and Human Knowledge, Penguin (Viking), Harmondsworth. ISBN 0 670 80060 0. Moles, Robert N. 1987, Definition and Rule in Legal Theory: A Reassessment of H. L. A. Hart and the Positivist Tradition, Basil Blackwell, Oxford. ISBN 0 631 15342 X. Moles, Robert N. 1991, “Logic programming—An assessment of its potential for ar­ tificial intelligence applications in law”, Journal of Law and Information Science, vol. 2, no. 2, pp. 137–64. ISSN 0729-1485. Moles, Robert N. and Dayal, Surendra 1992, “There is more to life than logic”, Journal of Law and Information Science, vol. 3, no. 2, pp. 188–218. ISSN 0729-1485. Montesquieu 1973, De l’Esprit des lois, vol. I, Garnier Frères, Paris. First published in 1748. Morris, Gwen, Cook, Catriona, Creyke, Robin and Geddes, Robert 1992, Laying Down the Law: The Foundations of Legal Reasoning, Research and Writing in Aus­ tralia and New Zealand (third edition), Butterworths, Sydney. ISBN 0 409 30541 3. Muir, Frank 1976, The Frank Muir Book: An Irreverent Companion to Social History, Heinemann, London. ISBN 0 434 48150 5. Murphy, Lionel 1980, “The responsibility of judges”, in Law, Politics and the Labor Movement, ed. Gareth Evans, Legal Service Bulletin, Law Faculty, Monash Uni­ versity, Clayton, ch. 1, pp. 2–10. ISBN 0 9594727 0 3. Nagel, Stuart 1960, “Using simple calculations to predict judicial decisions”, American Behavioral Scientist, vol. 4, no. 4, December, pp. 24–8. ISSN 0002-7642. Nagel, Stuart S. 1963, “Off-the-bench judicial attitudes”, in Judicial Decision-Making, ed. Glendon Schubert, International Yearbook of Political Behavior Research, vol. 4, Free Press of Glencoe, New York, ch. ii, pp. 29–53. Nagel, Stuart S. 1986, “Microcomputers and judicial prediction”, in Automated Analysis of Legal Texts: Logic, Informatics, Law, eds Antonio A. Martino and Fiorenza Socci Natali, Elsevier Science (North-Holland), Amsterdam, pp. 681–702. ISBN 0 444 70111 7. From the Second International Conference on ‘Logic, Informatics, Law’, Florence, Italy, September 1985. Nagel, Stuart S. 1989, Decision-Aiding Software and Legal Decision-Making: A Guide to Skills and Applications Throughout the Law, Quorum, New York. ISBN 0 89930 382 X. Bibliography [NET–POP] 425 Neter, John, Wasserman, William and Whitmore, G. A. 1982, Applied Statistics (second edition), Allyn and Bacon, Boston. ISBN 0 205 07613 0. Niblett, Bryan 1981, “Expert systems for lawyers”, Computers and Law, no. 29, Au­ gust, pp. 2–4. ISSN 0140-3249. Niblett, Bryan 1988, [Review of Susskind 1987a], Computers and Law, no. 56, June, pp. 32–3. ISSN 0140-3249. Oskamp, A. 1989, “Knowledge representation and legal expert systems”, in Ad­ vanced Topics of Law and Information Technology, ed. G. P. V. Vandenberghe, Com­ puter/Law Series, no. 3, Kluwer Law and Taxation Publishers, Deventer, pp. 195– 211. ISBN 90 6544 391 6. Oskamp, A., Walker, R. F., Schrickx, J. A. and van den Berg, P. H. 1989, “PROLEXS, divide and rule: a legal application”, in Proceedings of the Second In­ ternational Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 54–62. ISBN 0 89791 322 1. Pattison, Michael and Ciesielski, Victor B. 1990, Using Expert Systems and Con­ ceptual Graphs to Review Legal Contracts, Technical Report 22, Key Centre for Knowledge Based Systems, Departments of Computer Science, Royal Melbourne Institute of Technology and The University of Melbourne, 10 May. Pearce, Dennis, Campbell, Enid and Harding, Don 1987, Australian Law Schools: A Discipline Assessment for the Commonwealth Tertiary Education Commission, vol. 1, Australian Government Publishing Service, Canberra. ISBN 0 644 06182 0, 0 644 06186 3 (set). Pearce, D. C. and Geddes, R. S. 1988, Statutory Interpretation in Australia (third edition), Butterworths, Sydney. ISBN 0 409 49525 5. Plater, Alan 1985, The Beiderbecke Affair, Methuen, London. ISBN 0 413 57750 3. Popp, Walter G. and Schlink, Bernhard 1975, “JUDITH, a computer program to advise lawyers in reasoning a case”, Jurimetrics Journal, vol. 15, no. 4, Summer, pp. 303–14. ISSN 0022-6793. Popple, James 1990a, “Legal expert systems: The inadequacy of a rule-based ap­ proach”, in Proceedings of the Thirteenth Australian Computer Science Con­ ference (ACSC-13), Monash University, Melbourne, 7–9 February, pp. 303–13. ISSN 0157-3055. Australian Computer Science Communications, vol. 12, no. 1. Popple, James 1990b, “Legal expert systems: The inadequacy of a rule-based ap­ proach”, in Advances in Computing and Information: Proceedings of the Interna­ tional Conference on Computing and Information (ICCI-90), eds Selim G. Akl, Frank Fiala and Waldemar W. Koczkodaj, Niagara Falls, Canada, 23–26 May, Ca­ nadian Scholars’ Press, Toronto, pp. 348–51. ISBN 0 921627 70 X. Popple, James 1991, “Legal expert systems: The inadequacy of a rule-based ap­ proach”, The Australian Computer Journal, vol. 23, no. 1, February, pp. 11–16. ISSN 0004-8917. 426 [POP–RIS] Bibliography Popple, James 1993, SHYSTER: The Program, Technical Report TR-CS-93-13, De­ partment of Computer Science, Faculty of Engineering and Information Technology, The Australian National University, Canberra, December. Pound, Roscoe 1908, “Mechanical jurisprudence”, Columbia Law Review, vol. 8, no. 8, December, pp. 605–23. ISSN 0010-1958. Quinlan, J. R. 1986, “Induction of decision trees”, Machine Learning, vol. 1, no. 1, pp. 81–106. ISSN 0885-6125. Quinlan, J. R. 1988, “Induction, knowledge and expert systems”, in Artificial Intelli­ gence Developments and Applications, eds John S. Gero and Robin Stanton, Elsevier Science (North-Holland), Amsterdam, ch. 17, pp. 253–71. ISBN 0 444 70465 5. From the Australian Joint Artificial Intelligence Conference, Sydney, Australia, 2–4 Novem­ ber 1987. Quinlan, J. R., Compton, P. J., Horn, K. A. and Lazarus, L. 1987, “Inductive knowledge acquisition: a case study”, in Applications of Expert Systems, ed. J. Ross Quinlan, Addison-Wesley, Sydney, ch. 9, pp. 157–73. ISBN 0 201 17449 9. Based on the proceedings of the Second Australian Conference on Applications of Expert Systems, May 1986. Rawson, Hugh 1991, A Dictionary of Invective: A Treasury of Curses, Insults, Put- Downs, and Other Formerly Unprintable Terms from Anglo-Saxon Times to the Present, Robert Hale, London. ISBN 0 7090 4399 6. Ricketson, Staniforth 1984, The Law of Intellectual Property, Law Book Company, Sydney. ISBN 0 455 20553 1. Rissland, Edwina L. 1983, “Examples in legal reasoning: Legal hypotheticals”, in Proceedings of the Eighth International Joint Conference on Artificial Intel­ ligence (IJCAI-83), Karlsruhe, West Germany, 8–12 August, vol. 1, pp. 90–3. ISBN 0 86576 064 0. ISSN 1045-0823. Rissland, Edwina L. 1985, “Argument moves and hypotheticals”, in Computing Power and Legal Reasoning, ed. Charles Walter, West Publishing Company, St Paul, ch. 7, pp. 129–43. ISBN 0 314 95570 4. Rissland, Edwina L. 1989, “Dimension-based analysis of hypotheticals from Supreme Court oral argument”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Van­ couver, 13–16 June, pp. 111–20. ISBN 0 89791 322 1. Rissland, Edwina L. 1990, “Artificial intelligence and law: Stepping stones to a model of legal reasoning”, The Yale Law Journal, vol. 99, no. 8, June, pp. 1957– 81. ISSN 0044-0094. Rissland, Edwina L. and Ashley, Kevin D. 1987, “A case-based system for trade secrets law”, in Proceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 60–6. ISBN 0 89791 230 6. Bibliography [RIS–SER] 427 Rissland, Edwina L. and Ashley, Kevin D. 1989, “HYPO: A precedent-based legal reasoner”, in Advanced Topics of Law and Information Technology, ed. G. P. V. Vandenberghe, Computer/Law Series, no. 3, Kluwer Law and Taxation Publishers, Deventer, pp. 213–34. ISBN 90 6544 391 6. Rissland, Edwina L. and Skalak, David B. 1989a, “Interpreting statutory predic­ ates”, in Proceedings of the Second International Conference on Artificial Intelli­ gence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 46–53. ISBN 0 89791 322 1. Rissland, Edwina L. and Skalak, David B. 1989b, “Combining case-based and rule- based reasoning: A heuristic approach”, in Proceedings of the Eleventh International Joint Conference on Artificial Intelligence (IJCAI-89), Detroit, Michigan, 20–25 August, vol. 1, pp. 524–30. ISBN 1 55860 094 9. ISSN 1045-0823. Rissland, Edwina L., Valcarce, Eduardo M. and Ashley, Kevin D. 1984, “Ex­ plaining and arguing with examples”, in Proceedings of the National Conference on Artificial Intelligence (AAAI-84), University of Texas at Austin, 6–10 August, pp. 288–94. ISBN 0 86576 080 2. Sanders, Kathryn E. 1991, “Representing and reasoning about open-textured pre­ dicates”, in Proceedings of the Third International Conference on Artificial Intelli­ gence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 137–44. ISBN 0 89791 399 X. Schmidt, Frank L. 1971, “The relative efficiency of regression and simple unit pre­ dictor weights in applied differential psychology”, Educational and Psychological Measurement, vol. 31, no. 3, Autumn, pp. 699–714. ISSN 0013-1644. Schubert, Glendon 1963a, “Civilian control and stare decisis in the Warren Court”, in Judicial Decision-Making, ed. Glendon Schubert, International Yearbook of Polit­ ical Behavior Research, vol. 4, Free Press of Glencoe, New York, ch. iii, pp. 55–77. Schubert, Glendon 1963b, “Judicial attitudes and voting behavior: The 1961 term of the United States Supreme Court”, Law and Contemporary Problems, vol. 28, no. 1, Winter, pp. 100–42. ISSN 0023-9186. Republished in Baade 1963. Schubert, Glendon 1968, “The importance of computer technology to political science research in judicial behavior”, Jurimetrics Journal, vol. 8, no. 3, March, pp. 56–63. ISSN 0022-6793. Schubert, Glendon 1975, Human Jurisprudence: Public Law as Political Science, Uni­ versity Press of Hawaii, Honolulu. ISBN 0 8248 0294 2. Sergot, Marek, Cory, Therese, Hammond, Peter, Kowalski, Robert, Kriwaczek, Frank and Sadri, Fariba 1986a, “Formalisation of the British Nationality Act”, in Yearbook of Law Computers and Technology, ed. Christopher Arnold, vol. 2, Butterworths, London, pp. 40–52. ISBN 0 406 18701 0. Sergot, M. J., Sadri, F., Kowalski, R. A., Kriwaczek, F., Hammond, P. and Cory, H. T. 1986b, “The British Nationality Act as a logic program”, Commu­ nications of the ACM, vol. 29, no. 5, May, pp. 370–86. ISSN 0001-0782. 428 [SHA–STA] Bibliography Shannon, David T. and Golshani, Forouzan 1988, “On the automation of legal reasoning”, Jurimetrics Journal, vol. 28, no. 3, Spring, pp. 305–15. ISSN 0897-1277. Sherman, David M. 1987, “A PROLOG model of the Income Tax Act of Canada”, in Proceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 127–36. ISBN 0 89791 230 6. Sherman, David M. 1989, “Expert systems and ICAI in tax law: Killing two birds with one AI stone”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 74–80. ISBN 0 89791 322 1. Simpson, Brian 1986, “The common law and legal theory”, in Legal Theory and Common Law, ed. William Twining, Basil Blackwell, Oxford, ch. 2, pp. 8–25. ISBN 0 631 14477 3. Skalak, David B. 1989, “Taking advantage of models for legal classification”, in Pro­ ceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 234–41. ISBN 0 89791 322 1. Skalak, David B. and Rissland, Edwina L. 1991, “Argument moves in a rule-guided domain”, in Proceedings of the Third International Conference on Artificial Intelli­ gence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 1–11. ISBN 0 89791 399 X. Skalak, David B. and Rissland, Edwina L. 1992, “Arguments and cases: An inevitable intertwining”, Artificial Intelligence and Law, vol. 1, no. 1, pp. 3–44. ISSN 0924-8463. Smith, Ian 1992, “Employment”, in Halsbury’s Laws of England (fourth edition reissue), ed. Lord Hailsham of St Marylebone, vol. 16, Butterworths, London. ISBN 0 406 03465 5, 0 406 03400 1 (set). Sneath, Peter H. A. and Sokal, Robert R. 1973, Numerical Taxonomy: The Prin­ ciples and Practice of Numerical Classification, W. H. Freeman, San Francisco. ISBN 0 7167 0697 0. Spaeth, Harold J. 1963, “Warren Court attitudes toward business: The ‘B’ scale”, in Judicial Decision-Making, ed. Glendon Schubert, International Yearbook of Polit­ ical Behavior Research, vol. 4, Free Press of Glencoe, New York, ch. iv, pp. 79–108. Spengler, Joseph J. 1963, “Machine-made justice: Some implications”, Law and Con­ temporary Problems, vol. 28, no. 1, Winter, pp. 36–52. ISSN 0023-9186. Republished in Baade 1963. Sprowl, James A. 1979, “Automating the legal reasoning process: A computer that uses regulations and statutes to draft legal documents”, American Bar Foundation Research Journal, no. 1, Winter, pp. 1–81. ISSN 0361-9486. Stamper, Ronald 1980, “LEGOL: Modelling legal rules by computer”, in Computer Science and Law, ed. Bryan Niblett, Cambridge University Press, Cambridge, ch. 4, pp. 45–71. ISBN 0 521 23451 4. Bibliography [STA–TAN] 429 Stamper, Ronald 1986, “A non-classical logic for based on the structures of behaviour [sic]”, in Automated Analysis of Legal Texts: Logic, Informatics, Law, eds Antonio A. Martino and Fiorenza Socci Natali, Elsevier Science (North-Holland), Amsterdam, pp. 115–39. ISBN 0 444 70111 7. From the Second International Conference on ‘Logic, Informatics, Law’, Florence, Italy, September 1985. Stamper, Ronald, Tagg, Clare, Mason, Peter, Cook, Sandra and Marks, Jo 1982, “Developing the LEGOL semantic grammar”, in Artificial Intelligence and Legal Information Systems, ed. Constantino Ciampi, vol. I, North-Holland, Amsterdam, pp. 357–79. ISBN 0 444 86414 8, 0 444 86413 X (set). From the International Conference on “Logic, Informatics, Law”, Florence, Italy, April 1981. Stanley, Karlyn D. 1980, “LEXIS: Legal research and litigation support”, in Legal and Legislative Information Processing, ed. Beth Krevitt Eres, Greenwood Press, Westport, Connecticut, ch. 9, pp. 149–56. ISBN 0 313 21343 7. Stone, Julius 1964a, Legal System and Lawyers’ Reasonings, Maitland, Sydney. Stone, Julius 1964b, “Man and machine in the search for justice”, Stanford Law Re­ view, vol. 16, May, pp. 515–60. ISSN 0038-9765. Stone, Julius 1966, Social Dimensions of Law and Justice, Maitland, Sydney. Stone, Julius 1985, Precedent and Law: Dynamics of Common Law Growth, Butter- worths, Sydney. ISBN 0 409 49304 X. Susskind, Richard E. 1986, “Expert systems in law: A jurisprudential approach to artificial intelligence and legal reasoning”, The Modern Law Review, vol. 49, no. 2, March, pp. 168–94. ISSN 0026-7961. Susskind, Richard E. 1987a, Expert Systems in Law: A Jurisprudential Inquiry, Oxford University Press (Clarendon), Oxford. ISBN 0 19 825582 9. Susskind, Richard E. 1987b, “Expert systems in law: Out of the research laboratory and into the marketplace”, in Proceedings of the First International Conference on Artificial Intelligence and Law (ICAIL-87), Boston, Massachusetts, 27–29 May, pp. 1–8. ISBN 0 89791 230 6. Susskind, Richard 1989a, “Pragmatism and purism in artificial intelligence and legal reasoning”, AI and Society, vol. 3, no. 1, January–March, pp. 28–38. ISSN 0951-5666. Susskind, Richard E. 1989b, “The Latent Damage System: A jurisprudential analysis”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 23–32. ISBN 0 89791 322 1. Susskind, Richard E. 1990, [Review of Gardner 1987], in Yearbook of Law Computers and Technology, ed. Kenneth Russell, vol. 4, Butterworths, London, pp. 221–8. ISBN 0 406 18703 7. Tanenhaus, Joseph, Schick, Marvin, Muraskin, Matthew and Rosen, Daniel 1963, “The Supreme Court’s certiorari jurisdiction: Cue theory”, in Judicial Decision- Making, ed. Glendon Schubert, International Yearbook of Political Behavior Re­ search, vol. 4, Free Press of Glencoe, New York, ch. v, pp. 111–32. 430 [TAP–TYR] Bibliography Tapper, Colin 1963, “Lawyers and machines”, The Modern Law Review, vol. 26, no. 2, March, pp. 121–37. ISSN 0026-7961. Tapper, Colin 1973, Computers and the Law, Law in Context, Weidenfeld and Nicolson, London. ISBN 0 297 76571 X. Tapper, Colin 1984, “An experiment with citation vectors”, in Data Processing and the Law, ed. Colin Campbell, Sweet and Maxwell, London, pp. 90–109. ISBN 0 421 32000 1. Thomas, Bart T. 1979, “Unauthorized practice and computer aided legal analysis systems”, Jurimetrics Journal, vol. 20, no. 1, Fall, pp. 41–51. ISSN 0897-1277. Torsun, I. S. 1987, “PAYE: A tax expert system”, in Research and Development in Expert Systems III, ed. M. A. Bramer, British Computer Society Workshop Series, Cambridge University Press, Cambridge, pp. 69–80. ISBN 0 521 34145 X. From Expert Systems ’86: the Sixth Annual Technical Conference of the British Computer Society Specialist Group on Expert Systems, Brighton, 15–18 December 1986. Trudeau, G. B. 1975, The Doonesbury Chronicles, Holt, Rinehart and Winston, New York. ISBN 0 03 015256 9. Turing, A. M. 1950, “Computing machinery and intelligence”, Mind, vol. 59, no. 236, October, pp. 433–60. ISSN 0026-4423. Twiss, Horace 1844, The Public and Private Life of Lord Chancellor Eldon, with Se­ lections from his Correspondence (second edition), vol. III, John Murray, London. Tyree, Alan L. 1977, “The geometry of case law”, Victoria University of Wellington Law Review, vol. 8, pp. 403–20. ISSN 0042-5117. Tyree, Alan L. 1980, “Can a ‘deterministic’ computer judge overrule himself?”, Rutgers Journal of Computers, Technology and the Law, vol. 7, no. 2, pp. 381–4. ISSN 0735-8938. Tyree, Alan L. 1981, “Fact content analysis of case law: Methods and limitations”, Jurimetrics Journal, vol. 22, no. 1, Fall, pp. 1–33. ISSN 0897-1277. Tyree, A. 1985, “FINDER: An expert system”, in Proceedings of the Fortieth Annual Conference of the Australasian Universities Law Schools Association, University of Adelaide, 26–29 August. Tyree, Alan 1986, “Expert systems and the law: Will justice fall to bits?”, Current Affairs Bulletin, vol. 62, no. 10, March, pp. 13–18. ISSN 0011-3182. Tyree, Alan 1989, Expert Systems in Law, Prentice Hall, New York. ISBN 0 13 295650 0. Tyree, Alan L. 1992, “The logic programming debate”, Journal of Law and Informa­ tion Science, vol. 3, no. 1, pp. 111–15. ISSN 0729-1485. Bibliography [TYR–WEI] 431 Tyree, Alan L., Greenleaf, Graham and Mowbray, Andrew 1988, “Legal reas­ oning: the problem of precedent”, in Artificial Intelligence Developments and Ap­ plications, eds John S. Gero and Robin Stanton, Elsevier Science (North-Holland), Amsterdam, ch. 16, pp. 231–47. ISBN 0 444 70465 5. From the Australian Joint Artificial Intelligence Conference, Sydney, Australia, 2–4 Novem­ ber 1987. Tyree, Alan L., Greenleaf, Graham and Mo[w]bray, Andrew 1989, “Generat­ ing legal arguments”, Knowledge-Based Systems, vol. 2, no. 1, March, pp. 46–51. ISSN 0950-7051. Ulmer, S. Sidney 1963, “Quantitative analysis of judicial processes: Some practical and theoretical applications”, Law and Contemporary Problems, vol. 28, no. 1, Winter, pp. 164–84. ISSN 0023-9186. Republished in Baade 1963. Ulmer, S. Sidney 1964, “Homeostasis in the Supreme Court”, in Judicial Behavior: A Reader in Theory and Research, ed. Glendon Schubert, Rand McNally Political Science Series, Rand McNally and Company, Chicago, ch. II.B.6, pp. 162–80. van Opdorp, G. J., Walker, R. F., Schrickx, J. A., Groendijk, C. and van den Berg, P. H. 1991, “Networks at work: A connectionist approach to non-deductive legal reasoning”, in Proceedings of the Third International Conference on Artifi­ cial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25–28 June, pp. 278–87. ISBN 0 89791 399 X. Vossos, George, Zeleznikow, John, Dillon, Tharam and Vossos, Vivian 1991, “An example of integrating legal case based reasoning with object-oriented rule- based systems: IKBALS II”, in Proceedings of the Third International Conference on Artificial Intelligence and Law (ICAIL-91), St Catherine’s College, Oxford, 25– 28 June, pp. 31–41. ISBN 0 89791 399 X. Walker, R. F., Zeinstra, P. G. M. and van den Berg, P. H. 1989, “A model to model knowledge about knowledge or implementing meta-knowledge in PROLEXS”, in Ad­ vanced Topics of Law and Information Technology, ed. G. P. V. Vandenberghe, Com­ puter/Law Series, no. 3, Kluwer Law and Taxation Publishers, Deventer, pp. 235–60. ISBN 90 6544 391 6. Waterman, Donald A., Paul, Jody and Peterson, Mark 1987, “Expert systems for legal decision making”, in Applications of Expert Systems, ed. J. Ross Quinlan, Addison-Wesley, Sydney, ch. 2, pp. 23–47. ISBN 0 201 17449 9. Based on the proceedings of the Second Australian Conference on Applications of Expert Systems, May 1986. Waterman, D. A. and Peterson, Mark 1980, “Rule-based models of legal expertise”, in Proceedings of the First Annual National Conference on Artificial Intelligence (AAAI-80), Stanford University, 18–21 August, pp. 272–5. Weizenbaum, Joseph 1976, Computer Power and Human Reason: From Judgment to Calculation, W. H. Freeman, San Francisco. ISBN 0 7167 0464 1. 432 [WEL–YAZ] Bibliography Wells, Stanley and Taylor, Gary (eds) 1986, William Shakespeare: The Complete Works (original-spelling edition), Oxford University Press (Clarendon), Oxford. ISBN 0 19 812919 X. Wells, Stanley and Taylor, Gary 1987, William Shakespeare: A Textual Companion, Oxford University Press (Clarendon), Oxford. ISBN 0 19 812914 9. Wiener, Frederick Bernays 1962, “Decision prediction by computers: Nonsense cubed—and worse”, American Bar Association Journal, vol. 48, no. 11, November, pp. 1023–8. ISSN 0002-7596. Williams, W. T. 1971, “Principles of clustering”, Annual Review of Ecology and Sys­ tematics, vol. 2, pp. 303–26. ISSN 0066-4162. Willick, Marshal S. 1985, “Professional malpractice and the unauthorized practice of professions: Some legal and ethical aspects of the use of computers as decision-aids”, in Computing Power and Legal Reasoning, ed. Charles Walter, West Publishing Company, St Paul, ch. 28, pp. 817–63. ISBN 0 314 95570 4. Wirth, Niklaus 1977, “What can we do about the unnecessary diversity of notation for syntactic definitions?”, Communications of the ACM, vol. 20, no. 11, November, pp. 822–3. ISSN 0001-0782. Wittgenstein, Ludwig 1974, Philosophical Investigations (third edition), Basil Black- well, Oxford. ISBN 0 631 14670 9. Translation of Philosophische Untersuchungen (1953). Translated by G. E. M. Anscombe. Wolstenholme, David E. 1989, “Amalgamating regulation- and case-based advice systems through suggested answers”, in Proceedings of the Second International Conference on Artificial Intelligence and Law (ICAIL-89), University of British Columbia, Vancouver, 13–16 June, pp. 63–7. ISBN 0 89791 322 1. Yablon, Charles M. 1987, “Law and metaphysics”, The Yale Law Journal, vol. 96, no. 3, January, pp. 613–36. ISSN 0044-0094. Review of Kripke 1982. Yazdani, Masoud and Narayanan, Ajit (eds) 1984, Artificial Intelligence: Human Effects, Ellis Horwood Series in Artificial Intelligence, Ellis Horwood, Chichester. ISBN 0 85312 577 5, 0 470 20092 8. [Title page] [Verso title page] [Statement] [Etymology] Abstract Acknowledgments Contents Figures 3.1 Forms of functional dependence 3.2 Example attribute values 3.3 Probabilities for example attributes 3.4 Two-way association table 3.5 Example distances 3.6 SHYSTER's algorithm for choosing cases 3.7 SHYSTER's algorithm for using cases: part A 3.8 SHYSTER's algorithm for using cases: part B 3.9 SHYSTER's algorithm for using cases: part C 4.1 SHYSTER's command line switches 4.2 Keywords in SHYSTER's case law specification language 4.3 EBNF description of SHYSTER's case law specification language 4.4 Attribute values for Employee area (SHYSTER output) 4.5 Extract from probabilities for Employee area (SHYSTER output) 4.6 Similarity measures 4.7 Distances for BWIU v. Odco (SHYSTER output) 4.8 Extract from log file for BWIU v. Odco (SHYSTER output) 5.1 Probabilities for Finder area (SHYSTER output) 5.2 Weights for Finder area (SHYSTER output) 5.3 Distances for Parker v. BA (SHYSTER output) 5.4 Probabilities for Authorization area (SHYSTER output) 5.5 Weights for Authorization area (SHYSTER output) 5.6 Distances for CBS v. Ames (SHYSTER output) 5.7 Distances for Narich v. CPT (SHYSTER output) 5.8 Relationship between areas in Natural specification 5.9 Probabilities for Natural specification (SHYSTER output) 5.10 Weights for Natural specification (SHYSTER output) 5.11 Distances for Twist v. Randwick, Natural area (SHYSTER output) 5.12 Summary of test cases 5.13 Summary of generated testing D.1 Summary of reflexive testing of Finder specification D.2 Summary of reflexive testing of Authorization specification D.3 Summary of reflexive testing of Employee specification D.4 Summary of reflexive testing of Natural specification Chapters Chapter 1: Introduction 1.1 The aims of this thesis 1.2 SHYSTER 1.3 The structure of this thesis Chapter 2: Legal analysis systems 2.1 Introduction 2.1.1 Sources of law 2.1.2 The doctrine of precedent 2.1.3 The structure of this chapter 2.1.4 Terminology 2.2 Jurisprudence 2.2.1 The importance of jurisprudence 2.2.2 Scientific and mechanical jurisprudence 2.2.3 Judgment machines 2.2.4 Petrifaction of the law 2.2.5 Clear rules and clear cases 2.2.6 Legal realism and rule scepticism 2.2.7 A jurisprudential consensus? 2.3 Jurimetrics and the behaviourists 2.3.1 Kort 2.3.2 Lawlor 2.3.3 Nagel and Schubert 2.3.4 Haar, Sawyer and Cummings 2.3.5 Prediction 2.4 Rule-based systems 2.4.1 McCarty 2.4.2 Bench-Capon, Kowalski and Sergot 2.4.3 Gardner 2.4.4 Susskind 2.4.5 Knowledge acquisition and representation 2.4.6 Fact representation 2.4.7 The inadequacy of rules for case law 2.5 Case-based systems 2.5.1 Nearest neighbour analysis 2.5.2 FINDER 2.5.3 HYPO 2.5.4 The inadequacy of semantic networks 2.6 Hybrid systems 2.7 Conceptual models: deep and shallow 2.8 Conclusion Chapter 3: A pragmatic approach to case law 3.1 Introduction 3.2 Design criteria 3.2.1 Users and output 3.2.2 A pragmatic model of legal reasoning 3.2.3 Knowledge representation 3.2.4 Generality of application 3.2.5 Prediction and argument 3.2.6 Evaluation of advice 3.2.7 A hybrid structure 3.3 A model of legal reasoning 3.3.1 Private and public law 3.3.2 The functions of legal reasoning 3.3.3 Adopting a model of legal reasoning 3.4 Experts, users, and expert users 3.5 Knowledge representation 3.5.1 Areas 3.5.2 Attributes 3.5.3 Leading cases and attribute values 3.5.4 Ideal points 3.6 A specification language 3.7 Weighting attributes 3.8 Detecting attribute dependence 3.8.1 Functional dependence 3.8.2 Stochastic dependence 3.9 Calculating distances 3.9.1 Similarity measures Distance measures Association coefficients Correlation coefficients 3.9.2 Weighted similarity measures 3.9.3 Choosing a similarity measure 3.9.4 Infinite distance 3.10 Nearest cases and nearest results 3.10.1 Nearest cases 3.10.2 Nearest results 3.10.3 Equidistance 3.11 Writing a report 3.11.1 Arguing with the instant case Choosing cases Using cases 3.11.2 Arguing with instantiations 3.11.3 Arguing with hypotheticals 3.12 Safeguards 3.12.1 Extra similarity measures 3.12.2 Ideal points 3.12.3 Centroids 3.12.4 Attribute direction 3.12.5 Equidistance 3.12.6 Instantiations 3.13 Testing and evaluation 3.13.1 Choosing a test domain 3.13.2 Evaluating SHYSTER's opinion 3.13.3 The paucity of test cases 3.13.4 Generated tests 3.13.5 Reflexive tests 3.14 Conclusion 3.14.1 Comparisons with other approaches 3.14.2 A Sisyphean journey? Chapter 4: Implementing SHYSTER 4.1 Introduction 4.1.1 Internal representation 4.1.2 Output files 4.2 The Shyster module 4.3 The Statutes module 4.4 The Cases module 4.5 The Tokenizer module 4.5.1 Tokens 4.5.2 Comments and whitespace 4.6 The Parser module 4.6.1 Hierarchy 4.6.2 Areas 4.6.3 Attributes Local attributes External attributes 4.6.4 Leading cases 4.6.5 Ideal points 4.7 The Dumper module 4.7.1 Attributes Local attributes External attributes 4.7.2 Leading cases 4.7.3 Ideal points 4.8 The Checker module 4.8.1 Detecting dependence 4.8.2 Writing matrices of probabilities 4.9 The Scales module 4.9.1 Calculating weights 4.9.2 Writing tables of weights 4.10 The Adjuster module 4.11 The Consultant module 4.12 The Odometer module 4.12.1 Calculating distances 4.12.2 Writing tables of distances 4.13 The Reporter module 4.13.1 Arguing with the instant case Arguing for the nearest result Arguing for other results Variations 4.13.2 Arguing with instantiations 4.13.3 Arguing with hypotheticals 4.13.4 Safeguards 4.14 Conclusion Chapter 5: Case studies 5.1 Introduction 5.2 A simulation of FINDER 5.2.1 The law 5.2.2 The Finder specification Results Attributes Leading cases Attribute dependence Weights 5.2.3 Test case Parker v. BA 5.2.4 Conclusion 5.3 The authorization of copyright infringement 5.3.1 The law 5.3.2 The Authorization specification Results Attributes Leading cases Attribute dependence Weights 5.3.3 Test cases CBS v. Ames WEA v. Hanimex CBS v. Amstrad APRA v. Jain Hypothetical case 1 Hypothetical case 2 5.3.4 Generated tests Generated test 1 Generated test 2 Generated test 3 5.3.5 Conclusion 5.4 Employees and independent contractors 5.4.1 The law 5.4.2 The Employee specification Results Attributes Leading cases Attribute dependence Weights 5.4.3 Test cases Narich v. CPT Stevens v. Brodribb Re Porter BWIU v. Odco Hypothetical case 3 Hypothetical case 4 5.4.4 Generated tests Generated test 4 Generated test 5 Generated test 6 5.4.5 Conclusion 5.5 The implication of natural justice 5.5.1 The law 5.5.2 The Natural specification Natural area Affected area Expectation area Leading cases Attribute dependence Weights 5.5.3 Test cases Twist v. Randwick Salemi v. MacKellar Heatley v. TRGC Cole v. Cunningham Ackroyd v. Whitehouse Hodgens v. Gunn WA v. Bropho Ainsworth v. CJC 5.5.4 Generated tests Generated test 7 Generated test 8 5.5.5 Conclusion 5.6 Conclusion 5.6.1 Test cases 5.6.2 Generated tests 5.6.3 Reflexive tests Chapter 6: Conclusion 6.1 Introduction 6.2 Evaluating SHYSTER 6.2.1 A useful, working system 6.2.2 A general approach 6.2.3 Good advice 6.2.4 Limitations 6.2.5 Possible enhancements 6.3 Future research 6.4 The contribution of this thesis Appendices Appendix A: Case law specifications A.1 Introduction A.2 The Finder specification Hierarchy Finder area Results Attributes Cases in which the finder wins Cases in which the finder loses A.3 The Authorization specification Hierarchy Authorization area Opening Results Attributes Cases in which the accused authorized the infringement Cases in which the accused did not authorize the infringement Cases in which the accused is liable (directly or vicariously) for the infringement A.4 The Employee specification Hierarchy Employee area Opening Results Attributes Cases in which the worker is an employee Cases in which the worker is an independent contractor A.5 The Natural specification Hierarchy Natural area Opening Results Attributes Cases in which a duty to observe natural justice is implied Cases in which a duty to observe natural justice is not implied Affected area Results Attributes Cases in which the decision affected the property, right, interest, status, or legitimate expectation of the applicant Cases in which the decision did not affect the property, right, interest, status, or legitimate expectation of the applicant Expectation area Opening Results Attributes Cases in which the applicant had a legitimate expectation which was affected by the decision Cases in which the applicant did not have a legitimate expectation which was affected by the decision Appendix B: Example reports B.1 Introduction B.2 Report file for Parker v. BA Finder area Instant case Hypothetical 1 Hypothetical 2 B.3 Report file for APRA v. Jain Authorization area Instant case Hypothetical 1 Hypothetical 2 Hypothetical 3 B.4 Report file for Re Porter Employee area Instant case Instantiation 4 Hypothetical 1 Hypothetical 2 B.5 Report files for Ainsworth v. CJC Expectation area Instant case Hypothetical 1 Affected area Instant case Hypothetical 1 Natural area Instant case Hypothetical 1 Appendix C: A complete example C.1 Introduction C.2 Log file for BWIU v. Odco C.3 Employee specification file C.4 Dump file for Employee specification (LaTeX input) C.5 Dump file for Employee specification (LaTeX output) C.6 Probabilities file for Employee specification (LaTeX input) C.7 Probabilities file for Employee specification (LaTeX output) Employee area C.8 Weights file for Employee specification (LaTeX input) C.9 Weights file for Employee specification (LaTeX output) Employee area C.10 Distances file for BWIU v. Odco (LaTeX input) C.11 Distances file for BWIU v. Odco (LaTeX output) Employee area Instant case Instantiation 1 Instantiation 2 Instantiation 3 Instantiation 4 Hypothetical 1 Hypothetical 2 C.12 Report file for BWIU v. Odco (LaTeX input) C.13 Report file for BWIU v. Odco (LaTeX output) Employee area Instant case Hypothetical 1 Hypothetical 2 Appendix D: Reflexive tests D.1 Introduction D.2 The Finder specification Hannah v. Peel Bridges v. Hawkesworth Armory v. Delamirie Moffatt v. Kazana London v. Appleyard (1) London v. Appleyard (2) South Staffordshire v. Sharman Elwes v. Brigg D.3 The Authorization specification UNSW v. Moorhouse APRA v. Canterbury-Bankstown Winstone v. Wurlitzer Mellor v. ABC Falcon v. Famous Players RCA v. Fairfax PRS v. Ciryl A&M v. Audio Magnetics APRA v. Miles D.4 The Employee specification Zuijs v. Wirth Cam v. Sargent FCT v. Thompson ATWU v. Monaro Ferguson v. Dawson Stevenson v. Macdonald (2) PRS v. Palais de Danse Humberstone v. NTM Queensland Stations v. FCT Price v. Grant AMP v. Chaplin Massey v. Crown Life Stevenson v. Macdonald (1) Ready Mixed v. Minister D.5 The Natural specification FAI v. Winneke Haoucher v. Minister Annetts v. McCann Kioa v. West Commissioner v. Tanos Marine Hull v. Hurford Macrae v. AG AG v. Ng Durayappah v. Fernando SA v. O'Shea Bread Manufacturers v. Evans Minister v. Peko-Wallsend Nashua v. Channon CCSU v. Minister McInnes v. Onslow-Fane Notes/Cases/Statutes/Bibliography Notes Chapter 1: Introduction Chapter 2: Legal analysis systems Chapter 3: A pragmatic approach to case law Chapter 4: Implementing SHYSTER Chapter 5: Case studies Chapter 6: Conclusion Appendix A: Case law specifications Appendix B: Example reports Appendix C: A complete example Appendix D: Reflexive tests Cases Statutes Bibliography 
work_vi52ffbwhzf2zmdwj2o35h7h5m	----	OCAMS Methods AIEDAM v2 wjc Draft: May 1, 2008 1 Workflow Agents vs. Expert Systems: Problem Solving Methods in Work Systems Design William J. Clancey NASA Ames Research Center & Florida Institute for Human & Machine Cognition Maarten Sierhuis RIACS, NASA Ames Research Center Chin Seah QSS, NASA Ames Research Center Prepared for special issue of Artificial Intelligence for Engineering Design, Analysis, and Manufacturing (AIEDAM) — “Problem Solving Methods: Past, Present, and Future,” Spring 2008 Corresponding Author: William.J.Clancey@NASA.gov Intelligent Systems Division, M/S 269-3 NASA Ames Research Center Moffett Field, CA 94035 650-604-2526, fax 4-4036 Short Title: “The Role of Method Abstraction in Work Systems Design” Number of pages (excluding title, abstract, and biographies): 55 Number of figures: 2 Draft: May 1, 2008 2 Workflow Agents vs. Expert Systems: Problem Solving Methods in Work Systems Design Abstract During the 1980s, a community of artificial intelligence researchers became interested in formalizing problem solving methods as part of an effort called “second generation expert systems” (2nd GES). How do the motivations and results of this research relate to building tools for the workplace today? We provide an historical review of how the theory of expertise has developed, a progress report on a tool for designing and implementing model-based automation (Brahms), and a concrete example how we apply 2nd GES concepts today in an agent-based system for space flight operations (OCAMS). Brahms’ incorporates an ontology for modeling work practices, what people are doing in the course of a day, characterized as “activities.” OCAMS was developed using a simulation-to-implementation methodology, in which a prototype tool was embedded in a simulation of future work practices. OCAMS uses model-based methods to interactively plan its actions and keep track of the work to be done. The problem solving methods of practice are interactive, employing reasoning for and through action in the real world. Analogously, it is as if a medical expert system were charged not just with interpreting culture results, but actually interacting with a patient. Our perspective shifts from building a “problem solving” (expert) system to building an actor in the world. The reusable components in work system designs include entire “problem solvers” (e.g., a planning subsystem), interoperability frameworks, and workflow agents that use and revise models dynamically in a network of people and tools. Consequently, the research focus shifts so “problem solving methods” include ways of knowing that models Draft: May 1, 2008 3 do not fit the world, and ways of interacting with other agents and people to gain or verify information and (ultimately) adapt rules and procedures to resolve problematic situations. Keywords: Work systems design, work practice simulation, model-based automation, problem solving agent, situated cognition Draft: May 1, 2008 4 Introduction During the 1980s, a community of computer scientists working with the area of artificial intelligence became interested in formalizing problem-solving methods (PSMs) in an effort called “second generation expert systems” (2nd GES; David et al., 1993). What motivated this abstraction process and what did it accomplish? How do problem solving methods relate to building tools for the workplace today? Indeed, what have we learned about problem solving since the 1980s? To answer these questions, we provide an historical review of how the theory of expertise has developed, a progress report on tools for designing and implementing model-based automation, and a concrete example how we apply 2nd GES concepts today. The 2nd GES effort often involved analyzing the expert systems built in the 1970s to abstract their design and operation (e.g., Clancey & Letsinger, 1981). The motivations included: producing higher-level frameworks that would make building expert systems more efficient (called “knowledge acquisition”), making the programs more robust and powerful (by incorporating general principles rather than many isolated facts and heuristics), facilitating explanation of reasoning to users (especially in instructional applications), and facilitating reuse of the constructs in developing other expert systems (through PSM libraries). The 2nd GES analytic effort was part of the study of human problem solving (Newell & Simon, 1972), which showed that there were patterns, called “methods,” by which people applied and configured finer-grained “operators” in a process called “search in a problem space.” Analyzing dozens of programs in different domains, ranging from Draft: May 1, 2008 5 medicine to physics, and spanning a variety of kinds of problems (e.g., troubleshooting, design), 2nd GES researchers identified these additional patterns: o All expert systems are “model-based”—the representation methods of AI programming, specifically in expert systems, introduces a new modeling method to science and engineering, namely a means of modeling processes qualitatively (Clancey, 1986; 1989) in contrast with purely numeric programming. o Domain ontologies should be represented separately from the procedures that manipulate models, facilitating explanation and reuse (Clancey & Letsinger, 1981). o Solving particular problems (e.g., diagnosing a patient) involves creating situation-specific models; this is formalized most clearly in the blackboard architecture (Nii, 1986), which makes explicit the posting and comparison of model components (“hypotheses”) (Clancey, 1992). o Processes for manipulating models can be abstracted on different levels ranging from graph operators to entire frameworks for doing diagnosis, design, etc. (Clancey, 1985; Clancey & Barbanson, 1993). People summarized the conclusions of 2nd GES research in different ways, because models and procedures for manipulating models vary a great deal and have been represented in very different formalisms. Frameworks for organizing modeling languages and methods are provided by Chandrasekaran & Johnson (1993) and Clancey (1992). Depending on theoretical and practical objectives, different analytic perspectives will be preferred and useful. However the motivations of 2nd GES were broadly shared and not Draft: May 1, 2008 6 much in dispute, and the representation of domain entities and processes in classification and causal models seemed tractable. Instead, the debate and ambiguities concerned how to abstract and describe the procedures that manipulated models in the expert system (i.e., the PSMs). Two decades later, software engineering has not been transformed into a process of assembling PSMs from a library, as 2nd GES researchers imagined. Has there been a failure to appreciate the benefits of a PSM library or was the vision incomplete? We argue that there is a parallel between the limitations of expert systems and limitations of the value of PSMs for software engineering. Evaluating the success of “the PSM movement” requires also evaluating the success of “the expert system movement.” These limitations are founded in the narrow view of expertise that led 2nd GES researchers to believe that all software tools would be (or at least could contain) expert systems. The limitations in the “cognitivist” (Wallace, et al., 2007) view of knowledge, expertise, and problem solving—epitomized by the expert systems movement—explain why PSMs as conceived in David, et al. (1993) play only a small part in practical tools in the workplace. On the other hand, the motivations for 2nd GES are just as important and useful today, and the effort to abstract, formalize, and reuse software components in program libraries (e.g., for C++, JAVA) has in some respects accomplished what we hoped in the 1980s. Yet writing complex systems remains a craft; it seems we are always striving for the imagined libraries of modeling components. Is a kind of 3rd GES effort required or is invention and tedious adaptation inherent in the problem of developing software tools? Draft: May 1, 2008 7 To answer this question, we provide a broad review of problem-solving abstraction; the development of knowledge engineering tools; the shift in perspective about models, knowledge, and work from goals and reasoning to activities and interactive behavior; and the development of a tool for building work systems by modeling practice, called Brahms (Clancey, et al., 1998; 2005b; Seah, et al., 2005; Sierhuis, 2001; Sierhuis, et al., 2003). We provide an example of how Brahms has been used to design and implement a workflow tool for communications between NASA’s Mission Control ground support and the astronaut crew of the International Space Station (ISS; Clancey, et al., in press). We analyze the use of abstraction in this workflow tool and relate its architecture and methods to 2nd GES terminology. We conclude that the motivations of 2nd GES and PSM abstraction in particular has been transformed from configuring a problem solver to a higher-level problem of designing agents in a work system. The reusable components in work systems design include entire “problem solvers” (e.g., a planning subsystem), interoperability frameworks (relating hardware and software on different platforms), and interactive systems (“workflow agents”) that use and revise models dynamically in a network of people and tools. Historical Review of Problem Solving Methods Problem Solving is Reasoning The idea of problem solving methods is continuous with the long-term interest in mechanizing human reasoning (see for example, the review by Agre, 1997). The premises are simple and powerful: Good judgment is principled – knowledge is true belief — reasoning is mental. In the cognitivist paradigm, represented especially well by Draft: May 1, 2008 8 Newell & Simon (1972), logical reasoning, formulated most notably by Whitehead & Russell (1910), is the essence of cognition, connecting perception and action. In early modeling frameworks, called the “Logic Theorist” and “General Problem Solver,” theorem proving was the problem being studied. In subsequent frameworks (e.g., Green,1969), theorem proving then becomes a method for solving real-world problems: States in the world and goals are expressed in predicate calculus, and developing a plan of action to achieve a goal is reformulated as finding the states and actions (functions) that satisfies a theorem. Thus, the “logicist” formulation became the foundation of cognitivism (Wallace, et al., 2007), the analytic framework in which human “knowledge” consists of facts and rules (axioms and theorems in predicate calculus), and “reasoning” is logical manipulation (e.g., resolution theorem proving). These researchers viewed problem solving as logical, mental manipulation of beliefs about representations.1 The Power of Generality: Heuristic Methods Abstracting and generalizing is an essential aspect of human learning, in formulating a scientific model, a policy or law, or even an everyday principle for managing one’s day. People naturally express models of the world and behavior as generalizations that in logical terms involve predicates and variables (e.g., “For every weekday, if 4 PM > Time 1 For example, when Simon and Lea (1974) emphasized the role of what they called “information gathering” in problem solving, they didn’t mean observing the world, but rather “the degree of similarity or difference between the expressions contained in a given knowledge state and the goal expression” (p. 333)—information about mental constructs, as a measure of progress in a theorem proving process. Draft: May 1, 2008 9 > 7 PM, then avoid the freeways.”). Thus, the vision of Newell and Simon in their magnum opus (1972) was that a single program that represented and manipulated generalizations (theorems) could be directly applied to solve many problems in different domains, hence the name “General Problem Solver.” They demonstrated generality by applying the GPS method in cryptarithmetic and chess. Most notably, Newell and Simon (1972) formulated the problem solving process as a variety of methods such as “generate and test” and “recognition.” They define a method as “a collection of information processes that combine a series of means to attain an end” (p. 91). Their focus was not on the generality of the domain knowledge per se (that the theorems had variables and hence were general was essential for playing different games of chess, for example). Rather they focused on the generality of the “heuristic search” problem solving method (Newell & Simon, 1972, p. 101). This method was expressed as a program with steps such as “select-operator” and “decide-next-step.” They emphasized that the formulation of heuristic search that they presented is an “encompassing scheme from which more methods can be derived,” depending on the application. They also list some known alternative realizations of the undefined processes; for example, “decide- next-step” could be carried out by a fixed strategy of “always continue (one-level breadth-first search).” In summary, it was clear from early on, at least a decade before the term “expert system” was coined, that problem solving programs could be constructed from a general problem solving method, consisting of mental operations, variously called “heuristics,” “operators,” and “methods” (cf. Newell & Simon, 1972, pps, 101-103, 416-417) that are general processes, instantiated and combined in different ways, depending on the task Draft: May 1, 2008 10 environment. The overall program is referred to as a “heuristic;” thus for example, they refer to “the heuristic of means-ends analysis” and also “the basic system of heuristic of GPS.” This wording, where “heuristic” means “discovery” or “exploratory searching”— hence in effect, “system of discovery”—sounds awkward today because a different interpretation emerged in applying the framework to real-world problems in the 1970s. Then “heuristic” shifted from meaning the model manipulation operators to meaning the domain relations that the operators manipulated, aka “domain knowledge.” Building Expert Systems: Heuristic Knowledge Edward Feigenbaum, a student of Simon’s, moved the mechanization of problem solving from logic and chess, which were than characterized as “game playing”, to the broader realm of scientific human expertise, starting with chemistry (Feigenbaum, 1977). In 1965 Feigenbaum, et al. (1971) at Stanford started developing DENDRAL, a program for inferring molecular composition from spectral analysis, using the Plan-Generate-and- Test method, a variation of GPS. They concluded that real world problem solving, that is, problems that involved modeling empirical phenomena, was greatly aided by including more domain relationships, in particular uncertain syllogisms called “production rules.” These rules were called “heuristics,” emphasizing that they made the search process tractable. After DENDRAL, a broader effort in domains of medicine and biology was initiated as the Heuristic Programming Project in 1970. The body of domain heuristics became known as a “knowledge base” in the early 1970s, and the HPP was renamed the Knowledge Systems Laboratory in 1982. Because computer scientists worked with experts to formulate these rules, the common view was that an expert’s knowledge Draft: May 1, 2008 11 consists of production rules stored in long-term memory, and thus problem solving involves instantiating rules into chains of inference relating facts to actions.2 At the same time, an alternative formulation based on the notion of “schemas” or “frames” developed most notably at MIT (Minsky, 1985) and Yale (Schank, 1982). In this framework, expert knowledge consists of concepts that describe patterns in objects and events through relations (or attributes). Problem solving instantiates schemas and relates them to interpret situations, construct designs, formulate plans, etc. At the same time, the GPS formulation of operators and methods for applying operators was formalized in another framework called the “blackboard architecture” (Nii, 1986), a shared “working memory” in which operators post, relate, and refine alterative models of the world and actions. In related work, Pople (1977) formalized diagnosis as operators for constructing disease models that provided multiple and alternative explanations of or causal processes. Researchers also recognized that generalizing heuristics made knowledge bases more concise and the generalizations might be useful across domains. For example, Davis (1980) formalized metarules that could heuristically control how more specific domain rules were applied. Also in the 1970s, researchers applying educational psychology to computer-aided instruction emphasized “strategic knowledge,” “problem solving strategy,” and “meta-cognition” as powerful ways of thinking that could be taught (Greeno, 1980). 2 Representing concepts and their relations (attributes) in “semantic networks,” which dominated in the 1960s (e.g., see Minsky, 1969), was useful for tasks such as question answering; problem solving required conditional and higher-order associations. Draft: May 1, 2008 12 In summary, the idea of “problem solving methods” had different levels of emphasis—from heuristic processes that manipulated operators, to simply logical inference, to domain-general methods of analysis (e.g., Papert’s (1972) “thinking like a mathematician”). Researchers agreed on the overall methodology: formalizing problems in some representational language, abstracting domain knowledge and the problem solving procedure. But the different domain representations, inference methods, and analytic perspectives led to a Babel of languages and terms. A workshop was held to relate the perspectives (Hayes-Roth, et al., 1983); in some respects its effect was to promote the 2nd GES analytic effort. The Modeling Perspective Neomycin (Clancey & Letsinger,1981), a 2nd GES, was one of the first efforts to bridge between different problem solving formulations, proving that a domain ontology and diagnostic procedure were implicitly represented in Mycin’s production rules. Clancey (1984) formulated this procedure as a Hypothesize-Refine-Test method inspired by the terminology of GPS, implemented as a hierarchical set of “tasks” consisting of metarules for constructing a situation-specific model (e.g., a patient diagnosis) from the domain model. Clancey (1985; 1986; 1989; 1992) also claimed that all expert systems necessarily contained models and that all heuristic programs were necessarily constructing situation-specific models by applying operators that manipulated a general model of facts and associations. The “model construction operators” framework claims that whether one views knowledge-based programs as modeling human knowledge or just as automation tools— and whether one views “knowledge acquisition” as extracting what was already stored in Draft: May 1, 2008 13 the expert’s brain or collaboratively developing new theories of a domain and expert behavior (Clancey, 1993a)—expert systems are computer programs that contain domain models and procedures for manipulating models to apply them in specific situations. The kinds of problems (the modeling purpose) and the kinds of modeling methods fall into general categories, leading to the System-Task-Operator formulation of problem solving: 1) Problem solving involves modeling one or more systems in the world.3 2) The task is the purpose, what one wants to do with the system, that is, “the problem”: diagnosis, repair, configuration (design, modification, and/or plan), and control.4 3) Knowledge representation languages (e.g., production rules, schemas) provide a means of modeling processes occurring in the system (Clancey, 1989) using qualitative relations (e.g., type, cause, part-of, adjacency, co-occurrence). 4) Problem solving procedures (methods) consist of operators for manipulating models (e.g., chaining situation-specific models of different systems together in the Heuristic Classification Method).5 3 A system could be a naturally occurring physical system (e.g., the human body), designed artifact (e.g., a computer system), formally defined (e.g., a chess game), or some combination (e.g., a work process involving human behavior and computer tools). 4 Predicting the future state of a system is useful for many tasks, but not in itself a “problem.” 5 “The typology of problem tasks refers to why the system is being modeled; the typology of inference methods refers to how the model is developed” (cite AIJ retrospective). Generally speaking, when 2nd GES researchers referred to PSMs, they described what Draft: May 1, 2008 14 As detailed later, the enduring value of these methods for developing sophisticated automation systems seems assured. However, by the late 1980s a different perspective on knowledge and expertise suggested that problem solving consisted of much more than manipulating models, and thus experts could not easily be replaced by expert systems, and fitting such tools into the work place required much more than reliable machines and good interfaces. Theoretical Rebuttal: Reasoning is an Interactive Behavior From the very start, the logicist view was controversial. For example, the philosopher- educator, Dewey (1896), argued that inquiry consisted of reasoning-in-action in an early critique of stimulus-response theory, and criticized Russell’s equating reasoning with logic (Dewey, 1939). Wittgenstein (1953) rejected his own early support for Russell and like Dewey argued against a reductionist view of concepts and rule following (see also Wallace & Ross, 2006, p. 142). Damasio (1994) argued that emotion was essential to judgment, undermining the emotion vs. logic dichotomy that dominated the study of human intelligence. Many more trends of thought throughout the 20th century in ethology, cybernetics, anthropology, and systems theory built on a modeling framework often called “systems thinking.” These fields, operating unknown to mainstream cognitive psychologists and AI researchers, developed a theory of cognition that placed were variously called procedures, methods, or operators that combined how processes are modeled (the representation of the system being reasoned about) with how the models were manipulated. This is not necessarily wrong or unexpected, for operators for manipulating models would necessarily be stated in terms of the relations in the model (e.g., causality, subtype, temporality, adjacency). Draft: May 1, 2008 15 human knowledge, memory, and reasoning within a complex system of biological, psychological, and social processes (Clancey, in press). This perspective eventually became known in the fields of AI and cognitive science as “situated cognition.” From the perspective of expert systems research, the essential claim of situated cognition is that human knowledge cannot be equated with models, or put another way conceptualizing does not consist of simply retrieving and instantiating stored relational networks and procedures. In effect, the nature of human conceptualization has never been properly understood or replicated because the nature of memory as a storage place is incorrect (Clancey, 1997a; 1999). The implications for advancing theories of knowledge and action are immense. For example, the nature and role of consciousness becomes clearer: In attentively controlling behavior, a person is always conceiving “what I am doing now” (WIDN; Clancey, 1999), the present activity. This ongoing conception is effectively the construction of identity, a social-psychological construct. In particular, this conception must relate constraints of multiple, blended identities, involving different and perhaps conflicting obligations (accountability) and methods (what I might do now). This understanding of WIDN is dynamically and mutually developing within the conception of “the situation” (Clancey, 2004). Knowledge systems developers had a great deal of difficulty at first understanding the situated cognition perspective because it violated the very assumptions of the information processing paradigm. “Situated” does not just mean located in the world (which is of course true) or “extracting information from the surrounding physical world” (Chandrasekaran & Johnson, 1993). Rather the person’s perceiving, interpreting, and Draft: May 1, 2008 16 acting is conceptually organized with respect to WIDN. Behavior is situated because the person is acting while (and by) dynamically conceiving what constitutes “the situation.” Reflective feedback (Schön’s [1987] “knowledge-in-action”) is potentially fast (e.g., in dance or conversation) and often involves different conceptual organizers acting sequentially and/or simultaneously. Conceiving itself occurs in experience as a (necessarily conscious) behavior. You only know what you are going to say when you say it (even when you say it to yourself first). Action changes perception, and hence the conception of activity changes what constitutes information. (See Clancey [1997a, 1999] for discussion and references.) Furthermore, not every human activity (the conception of WIDN) involves a problem to be solved (Clancey, 2002). Activity theory, another parallel school of thought dating from 50 years before Newell and Simon, provides a much broader view of motives, goals, and operations, including non-problematic goals (e.g., reading a magazine to relax), how conceptualization of the social setting affects the choice of methods (i.e., norms), and how a structured environment can provide an interactive scaffolding for guiding information gathering and reasoning (Hutching & Palen, 1997). Situated cognition is not a behaviorist movement, as the cognitivists feared (Vera & Simon, 1993), but rather one that more radically turns from behaviorism than information processing was able—by emphasizing that information is not given, or simply “extracted,” rather, the environment is both dynamically perceived in action and modified in action (dynamic interaction).6 In contrast, cognition in the conventional information processing paradigm is more reactive, assuming a kind of given stimulus and a packaged 6 See discussion of ecological psychology in Clancey (1997a). Draft: May 1, 2008 17 response (a plan), such that human problem solving can be replicated and improved upon (not just modeled) by situation-action (stimulus-response) rules and stored schemas. Cognitivism thus narrowed the study of problem solving to the internal (mental) manipulation of models of the world and behavior, viewing the getting of information and subsequent action as the inputs and outputs of reasoning. As many have noted, this dichotomy between “mental processes” and “behavior” is just a continuation of Descartes’ separation of mind and body (e.g., Damasio, 1994; Wallace, et al., 2007). If problem solving (reasoning) is actually a behavior, then the notion of “problem-solving method” can be greatly broadened and hence very different kinds of abstractions formulated for software engineering. In particular, as explained below, analyzing work in terms of activities—what people do and how they conceive of what they are doing— provides a context for how tasks are discovered, defined, and handled. As has been shown in two decades of Applications of AI conferences, we can develop useful model-based tools. But these tools operate within a complex system involving other tools and human interactions. For example, Mycin’s design assumed a culture has already been taken, presuming a certain medical setting with sophisticated caretakers, even if they are not antimicrobial experts. Developing tools that fit into a workplace involves a more sophisticated theory of problems and problem solving than was assumed in first developing expert systems. Most importantly, the analytic perspective “technical rationality” (Schön, 1987) is reductionist because it presumes that gathering and interpreting information and decision making always has a routine character. Ethnomethodologists have shown how everyday work requires deciding how to categorize “situations” and deal with conflicts and Draft: May 1, 2008 18 shortcomings in procedures (Clancey, 2006). Choosing among courses of action requires interpreting how actions will be evaluated in the current social-organizational context. For example, medical practitioners need to relate the choice of tests to the policies of the patient’s insurance company. This involves judgmental reasoning to be sure, but moves beyond scientific models of the human body to a realm of negotiation and compromise (consider the work of patients themselves in attempting to overturn insurance decisions). Model-based automation can be very powerful for handling routine work, but there must be means for non-programmers to revise ontologies and rules, monitor the program, and modify operations on a case-by-case basis. This in turn requires new roles, work processes, and organizational policies, leading to an analytic framework called “work systems design”—a far broader problem than the view of automation expressed in the effort we called “building expert systems.” In summary, the limitations of expert systems stem from the limits of process models, procedures, and policies for detailing in advance how work must be (or could be) done. As conceived by people, the meanings of these formalizations (whether conceptual networks or texts) are not reducible to more models. People necessarily and opportunistically reconceive what categories and rules mean, and they often do this with other people. Even if one rejects the strong claim—that in principle human conceptualizations cannot be reduced to concept-relation networks7—and even allowing that meanings, justifications, sources etc. can be modeled so automation is more adaptive to circumstances—the automation cannot itself be left alone. For the seeable future at 7 Despite the promise of neural net mechanisms (e.g., Elman, 2004), we have not yet figured out how to replicate human conceptualization (Clancey, 1999). Draft: May 1, 2008 19 least, interactions with people are required so they can ensure that models, procedures, and policies embedded in automation tools are properly interpreted and adapted in practice. This interaction itself must be flexible and sensitive to contexts. The expert system must become an actor in a “web of practices” (Wallace & Ross, 2006), an agent that can interact with people and other tools in a dynamic physical-organizational work system. As Pollack (1991) said, “We want to build intelligent actors, not just intelligent thinkers. Indeed, it is not even clear how one could assess intelligence in a system that never acted – or, put otherwise, how a system could exhibit intelligence in the absence of action.” Put another way, we need to formalize automation behaviors with respect to human activities. Brahms: Modeling and Facilitating Human Activities Motivation for Simulating Work Practice In the early 1990s AI researchers at NYNEX Science & Technology Research Center in New York City were unsuccessful in fielding an expert system. An anthropologist was hired to study the knowledge and work relationships of line craftsmen in Manhattan. She brought to the AI group three areas of new expertise: 1) a work practice analysis approach that related tools to how the work was actually done, 2) an ethnographic method for gathering information about the workplace, and 3) a participatory design approach for bringing workers into the tool development process. At the same time, the NYNEX Expert Systems group was competing with other systems analysts who espoused the “business process re-engineering” approach of modeling and optimizing workflows, using a business process modeling tool called Sparks (Clancey, et al., 1998; Sierhuis & Clancey, 1997). To be competitive, the team of Draft: May 1, 2008 20 AI and social scientists needed to advocate their work system redesigns using a similar simulation that predicted timelines and costs. From the social scientists’ perspective, the main requirement was to make social processes visible—put people on the screen so workers could participate in the modeling process and visualize how the graphics related to their own roles and situations. The new team of AI expert systems developers and social scientists hired by Sachs used Sparks to model work practices as best they could, but were hampered by the lack of explicit modeling constructs for representing people, tools, documents, workplace layouts, communications, and movement of people and objects in geographic space. They were attempting to model not the idealized and abstracted “work flow” of business processes, but how artifacts and information were actually modified and conveyed by interactions among people and automated tools. For example, a work practice model would represent not just that a job order moved from one business functional unit to another (e.g., sales to provisioning to installation), but how the order was represented in a document and transmitted by fax, and how the manager of a particular business office would handle the incoming faxes. Using Sparks, it was particularly difficult to explicitly represent how three or more people coordinated their actions (e.g., in testing a circuit across Manhattan) without jury- rigging the constructs in Sparks’ manufacturing-inspired paradigm that centered on functional changes to the product as opposed to behaviors of the people. Multitasking, informal assistance (working on a task to which you were not assigned), dealing with breakdowns (e.g., inconsistent orders), interruption and resumption of activities (e.g., when answering a phone call) were all very difficult to express in Sparks’ assembly-line Draft: May 1, 2008 21 framework. In effect, business process modeling tools enabled representing how work flows through an organization, but not the work people were actually doing so that jobs and information actually moved along from one person or tool to the next (Wynn, 1991). A work practice simulation has advantages over a model that only represents formal organizational roles and procedures: o Reveals informal practices–what is not in the procedures but affects the quality of the work, including informal assistance, learning, sharing of information, workarounds, variations for efficiency, ways of satisfying the customer when the rules cannot be strictly followed, etc. o Reveals tacit assumptions about work that a functional abstraction into tasks and methods ignores, e.g., who notices that an order fax arrives and how are questions about the order resolved? When informal and tacit aspects are revealed (i.e., logistic issues and hidden side- benefits are articulated) then we can be more confident that workers’ methods are not obstructed and are appropriately supported when new roles, procedures, schedules, tools, documents, etc. are introduced. Modeling work practice requires representing details of workflow coordination that business process models usually omit (e.g., fax machines) and moving beyond individual reasoning to simulate interactions among groups (e.g., office workers). The essence is always to understand and model how work actually gets done, not just what is supposed to happen. The key constructs in a work practice model are: o Activities (chronological behaviors of people), not just tasks (functional transformations of work products). Activities (conceptualizations of WIDN) Draft: May 1, 2008 22 are effectively subsumed and simultaneous on different organizational and temporal levels: Living and working in New York, working for NYNEX, Installing a circuit for a customer on-site, Testing the circuit. Personal activities (e.g., Being a Parent) are dynamically blended with these work activities during the day (e.g., how a call from home is handled may depend on the ongoing work activity or may override work concerns). o Tools, Documents, Communications, Areas, Objects with behaviors, Movements. Using these constructs, one models facilities, object layouts in space, vehicles, communication devices, etc. In summary, a work practice simulation simulates behaviors of people, which includes simulating their reasoning about objects in a simulated environment. The emphasis is on chronological behaviors of people (how they organize their time, e.g., “reading email first thing in the morning”) instead of only functional behaviors (transformations of work products, e.g., “filing out a purchase order”) as in business process models, or just reasoning as in expert systems. Of course, some activities (e.g., constructing a plan, troubleshooting a device) are like expert system tasks. In effect, the nature of the domain broadens: A work practice simulation models the structure and behavior of human organizations, which includes modeling the structure and behavior of objects (e.g., an electronic circuit) that reasoning operates upon. Because a work system includes objects, models, procedures, and policies, and a simulation of work must show how tasks are performed, a work practice model contains models of problem solving within it. Draft: May 1, 2008 23 Brahms: A Work Practice Modeling Approach In late 1992 NYNEX and the Institute for Research on Learning formed a partnership, with a primary objective of developing a work systems design simulation tool that would facilitate work practice analysis, ethnography, and participatory design. In developing the tool, which became known as Brahms, it was apparent from early on that the modeling language must enable representing interactions between people doing activities, objects having structure and behaviors, and geographic areas in which people and objects were located and moved. Although the AI members of the group could not fully explain at the time how the social scientists’ concept of “activities” related to “tasks” of expert systems (Clancey, 1997b), it was possible to develop an architecture that incorporated the desired constructs for modeling chronological behaviors. Three existing computational ideas were merged in the Brahms architecture: o Neomycin’s “metacognitive” architecture was adapted for its flexibility for organizing and controlling high-level processes. Neomycin’s strategic methods called “tasks”8 became Activities in Brahms; Metarules became Workframes; “end-conditions” became Detectables). o Activities are activated and “running” in a subsumption architecture (Brooks, 1991) instead of being invoked like functions (e.g., like Neomycin’s “tasks”). o Following the “Distributed AI” approach (Bond & Gasser, 1988), agents and objects interact in a modeled environment—blending ideas from Cohen, et al.’s (1989) simulation of fire-fighting, SimLife’s simulation of animals (a 8 In Clancey’s (1992) reformulation, Neomycin’s “tasks” were renamed “methods.” Draft: May 1, 2008 24 game by Maxis), and the then nascent work on “simulating societies” (Gilbert & Doran, 1993)9. Key concepts in modeling human behavior are made explicit in the Brahms language to constitute a particular type of multiagent system: o Groups of Agents with individual Beliefs interact while doing personal and inherited group Activities. o Behaviors in Activities represented as conditional actions (Workframes), which are sequences or alternative ways of doing something; Activities can be aborted, interrupted and resumed. o Inference occurs within the context of Activities (Thoughtframes) o Perceiving is an experience while acting (Detectables within Workframes) o World Facts (the modeler’s God’s eye view of the environment) are distinguished from agent Beliefs about the world (e.g., the simulation may represent that an object is in a location with a state, but an agent may have arbitrary beliefs about the object) 9 Brahms was first presented at the Second International Conference on Multiagent Systems in 1996. The ideas of “multiagent systems” and “agent-based modeling” were in the air when the architecture was invented in early 1993. For example, Carley (1990) presents a “socio-cognitive model of the interface between self and society,” combining social and cognitive model constructs. However, her formalism does not have the construct of an “agent” with simulated behaviors in a simulated the environment. Individuals only interact in an abstract sense, which causes “exchange of information.” Draft: May 1, 2008 25 o Conceptual Objects represent mental constructs about Agents, Groups, and Activities (e.g., jobs, phases in an activity: preparation for, during, and after journey of STS to ISS); objects may be an instance of a class. o Agents and Objects are contained within Areas; an area may be an instance of an area class, Part Of another area or connected by a Path. Brahms is a natural extension of the knowledge-based systems concept, applied to modeling people at work. In original inspiration, each agent in Brahms is like one knowledge-based system, but not all agents are people: Some devices with sensors and complex behaviors are modeled as agents (e.g., robots); simpler objects (or systems modeled as simple objects) can have behaviors, too (e.g., an email program). In 1998 Brahms development shifted from IRL/NYNEX to NASA Ames Research Center. Sierhuis (2001) simulated aspects of Apollo lunar operations, followed by simulations of mission operations on the ISS (Acquisti, et al., 2002), a Mars analog habitat (Clancey, et al., 2005b), and planning operations for controlling the Mars Exploration Rover (Seah, et al., 2005). Amazingly, the notion of geography shifted from Manhattan to the moon and Mars. Modeled objects shifted from telephones to robots. Simulations of operations showed lack of connectivity and how breakdowns in flows (e.g., missing steps in procedures) were detected and handled in practice. In summary, the Brahms modeling framework constitutes a schema for simulating work practice, very much in the spirit of the 2nd GES effort to develop domain-general abstractions, but shifting from a focus on modeling problem solving processes to modeling work systems. In effect, a model of work practice involves modeling how problems arise and are recognized, formulated, and resolved; the roles of different people Draft: May 1, 2008 26 and the tools they use (e.g., the instruments that provide data input to expert systems), and the environment in which all this occurs. This notion of problem solving is much broader than is formalized in the model manipulation processes of PSMs. Using Agents to Implement Automation Tools Modeling problem solving as it occurs in the world, within activities, provides a context for defining automation, specifically model-based tools. In an elegant formulation, a prototype tool can be embedded in the Brahms work practice simulation, prompting the modeler to investigate and determine, for example, how information is gathered to use a tool and how tools are interactively related to the work of other people and tools, which might involve documents, communicating, networks, and so on. Thus a tool can be designed to fit the work practice, and then extracted from the simulation and deployed as an agent-based workflow system. We began the simulation-to-implementation approach in the Mobile Agents Project (2001-2006), where we used the Brahms architecture as a runtime system to develop a series of distributed workflow tools (Clancey, et al., 2005b). In the runtime configuration, one or more agents are located on a given computer platform and communicate in real time with each other and to get data from and control external devices and software systems. Thus, runtime agents are interacting processes that interpret data, communicate, and take action in the world and may cause their platform to move (e.g., a robot) or be moved about in the world (e.g., a computer on a backpack). Generally, each person using Mobile Agents has a “personal agent” with which he or she communicates by voice and/or a GUI. The “world facts” of the Brahms simulation are replaced by the world Draft: May 1, 2008 27 itself, which must be inspected, instrumented, and manipulated by the agents in order to get information. In the Brahms runtime configuration, an “agent” is a subsystem within a larger environment of agents. Comparing to Schreiber, et al. (1994, p. 29), “A KBS [knowledge-based system] is only one agent among many—human and nonhuman—and carries out only a fraction of the organization’s tasks,” we would say that a workflow tool consists of many agents, each of which is a KBS. Correspondingly, a workflow tool constructed from Brahms doesn’t consist of a set of modules such as “Inference,” “Communication,” and “Domain Knowledge Base”—the common components of an expert system—but has a higher-level physical and functional architecture (e.g. the rover’s agents include a “navigation agent,” “panoramic camera agent,” and “a speech agent”). Each agent has inferential, communication, and belief maintenance capabilities provided by the Brahms Virtual Machine (engine; Sierhuis, et al., 2007). In particular, depending on its roles and location in the real world, including other systems to which it is coupled, each agent attends to different data, forms its own beliefs, and carries out its own activities. Diagrams of a Brahms multiagent system (see Figure 1 for the OCAMS tool described subsequently) show how the agents are distributed on platforms and their functional interactions; we also represent resulting behaviors in timelines (the AgentViewer; Sierhuis, et al., 2007). In summary, just as an expert system was not in general conceived as being embedded in a work system, a library of PSMs is not sufficient for building workflow tools. The expert system notion of an “executive” interpreting data and applying schemas Draft: May 1, 2008 28 or rules applies most directly to the functions of the Brahms Virtual Machine.10 The Agent-Thoughtframe-Belief framework is based on the architecture of expert systems, but is contained within a broader work system schema (group, agent, activity, detectable, communication act, area, movement) that enables simulating parallel, dynamic interactions among people and objects in their environment. << INSERT FIGURE 1 HERE >> Practical Perspectives: Insights from Using Brahms in Practice Distinguishing Brahms from other NASA tools (Freed, et al., 1998) and cognitive task analysis (Vicente, 1999) led us to better articulate the practical nature of activity models and use of simulation for work systems design. That is, we came to understand how to disentangle a number of theoretical and technical issues by recognizing how modeling, tools, and simulations are successfully configured in practice. Perspective on Models (circa 2001) o Activities and tasks are analytic abstractions (Clancey, 2002). There is no single, correct way to model and simulate work. The choice of analytic perspective(s) depends on the purpose of the model. Also, we can derive workflow diagrams from Brahms simulation runs; because these sequences are not necessarily built into the model, their emergence in particular cases can provide new information about the work system design. o A Brahms model is not the actual knowledge, conceptualizations, situations, people, communications, etc. of practice—it is just a model. Because 10 The Brahms Virtual Machine achieves a parallel, discrete simulation by managing communications, belief revision, agent movements, activity/workframe activations, scoping of detectables, and application of thoughtframes (Sierhuis, et al., 2007). Draft: May 1, 2008 29 cognitivism equated human knowledge with models11, the model was viewed not just as a process theory to be evaluated contextually, but as the very stuff of cognition itself—so it necessarily was either correct, incomplete, or wrong. Perspective on Tools o People in their everyday lives create models as tools, including models of other people’s behavior and knowledge (Schön, 1987). Models are guides or, broadly speaking, maps that people interpret through conceptualizations. o In predictable, patterned work settings with well-defined operations, models can be used to automate the work—to replicate routine human behavior. But when a model-based program is put into different value-laden and non-routine contexts, its operation may be interpreted as wrong and/or requiring a workaround. This means that at a minimum we must build into the tools and work practices means for adapting and/or circumventing the automation. o Consequently, building workplace tools requires working with the people who will use them, not just the people who are being replaced (if any). Contrast 11 Vera and Simon (1993) wrote: “Patterns of neurons and neuronal relations… bear a one to one relationship to the Category 4 [stored programs and data] symbol structures in the corresponding program” (p. 120). They provided no neuropsychological evidence of this isomorphism. But when Clancey (1993b) said that “Every act…is a new neurological coordination” citing neuropsychological models by Edelman and Freeman, as well as the psychological analyses of Dewey, Bartlett, Sacks, and Vygotsky, Vera and Simon replied, “He provides no evidence for these flat assertions” (p. 124). Draft: May 1, 2008 30 building a medical expert system by working with nurses who will use the tool, not just interviewing physicians (e.g., Greenbaum & Kyng, 1991). Perspective on Simulation o In shifting from a model of mental processes to a model of work systems, the issue of building knowledge bases is replaced by designing work systems (e.g., including roles, facilities, operational procedures). We move from a tool for building tools to a tool for simulating tools and the context in which they will be used. We develop Brahms simulations to get insights about the workplace, particularly how problems are solved in practice, and thus guidance for knowing what tool to build. o A practical simulation is not just descriptive, but makes predictions about timing, flows and bottlenecks, and costs. This resolves a scoping issue: What aspects of human life should we simulate if we are not (just) developing a model of the technical domain and reasoning? A good approach is to design the simulation to provide metrics that answer questions having a bearing on proposed changes to the work practices (e.g., schedules, automation, roles, product flows). o A work practice model reveals unanticipated or missing interactions. By not directly modeling (building in) the workflows that form the backbone of a product-centered simulation, a work practice simulation enables evaluating more basic aspects of how the work comes together (e.g., whether work schedules of different roles interact to cause delays). Draft: May 1, 2008 31 EXAMPLE: The OCAMS Workflow Automation Tool In this section we illustrate and develop some of the points about problem solving and tools further by analyzing a mission operations workflow tool developed at NASA using the Brahms modeling and simulation tool. We describe the context and design constraints, the methodology, the use of abstraction in the solution, and the practical implications for the “library of methods” approach. Objectives The project is to automate some (and eventually perhaps all) of the file management operations between support groups and the astronauts onboard the International Space Station, as performed by a job position called the OCA Officer.12 The broader organizational objective is to improve efficiency of mission operations by reducing personnel costs by 30% by 2012. A secondary objective is to bring NASA’s research results into practical application by establishing partnerships between research and operations organizations. Demonstrating practical applications of agent-based systems integration in ground flight operations will promote the use of such tools in lunar surface operations (prototyped in the Mobile Agents field experiments). Design Constraints The project has the following design constraints: o Automate routine operations of the OCA officer: mirroring,13 archiving, up/downlink to the ISS, notification.14 12 OCA = Orbital communications adapter, a card used that effectively enables a personal computer to FTP files on a satellite network. OCAMS = OCA Mirroring System. 13 Mirroring involves replicating on a local network, called the Mirror LAN, the file operations performed on the ISS file system. Draft: May 1, 2008 32 o Enable the OCA officer to retain responsibility and authority by allowing for manual overrides of all system operations. o Enable OCA officers to modify how the system operates without programming (sustainability and adaptability in practice). o Be sensitive to the current practices for workflow (e.g., receipt of new jobs and transmission of results), timing, communications, and authority for variances in routines. o Respect the work practices of shift handovers that involve restarting software tools, recording file management statistics in a handover log, retaining records of incomplete work or unresolved problems, etc. Simulation to Implementation Methodology We partnered with operations personnel to create two simulations: current operations (in which mirroring is done manually) and future operations (in which mirroring is done with a distributed multiagent workflow tool). The future operations simulation effectively includes the OCAMS tool used by a simulated OCA officer; it includes a prototype GUI by which someone can control the automation to understand what is happening. Both simulations model work shifts, handovers between OCA officers, and maintaining handover logs. The current and future simulations ran on one month of previously recorded data, allowing comparisons of the OCA Officer’s work with and without the 14 Notification includes speaking on the “voice loop” (a programmable network of intercoms), modifying a Flight Note (a message posted in a workflow tool), sending email, telephoning someone, and broadcasting a remark outloud in the room). Draft: May 1, 2008 33 tool, based on actual data about the work products of an ISS mission (Clancey, et al., in press). Subsequently the agents comprising the tool were extracted from the future simulation and reconfigured on multiple platforms. This overall approach of transforming a current simulation into a future simulation and then a tool is called “simulation to implementation,” and represents an important example of how models can be reused for design within a project (contrasted with the PSMs library idea of reuse across projects). Analysis of File Management Process Table 1 details how the work of ISS file management can be abstracted into an ontology of file types and handling methods. The principles for doing the three file management operations (mirroring, archiving, and notifying) are not based on the encoding of the file (e.g., a text document vs. a JPG image), but the functional relation of the file to the mission (operational plans/procedures and software, private data, and exceptions to these).15 The 31 file types are acronyms assigned by OCA officers (e.g., JEDI for certain procedures; NAV for antivirus software; BME is medical; NFH for “news from home”). Files may be transferred up to the ISS (uplink), down to earth, or both. 15 Not shown are file name templates used to recognize the file type. The input to OCAMS is a log of operations carried out by the OCA officer in transferring files between the ground and ISS. OCAMS infers the file type from the file path and name (e.g., a file name of the form “DOUG/flights/…pkg” is file type DOUG and should be mirrored). Draft: May 1, 2008 34 By examining the file type ontology, we can better understand the role of PSMs (either actual or potential) in the construction of OCAMS. The ontology can be summarized by the following four principles: 1) Operational data is mirrored and archived. Medical or personal data is neither mirrored nor archived. 2) Most down-linked items are not mirrored. Exceptions: a) Files that stay onboard after downlink; b) Changes the crew has made to the onboard software that must also be implemented on the Mirror LAN. 3) Exception to archiving: Keep a rolling archive of imagery because of the volume. 4) Items deleted onboard are also deleted on the Mirror LAN. << INSERT TABLE 1 HERE >> Without considering exceptions, the principles given above suggest these categories: 1. <Operational: UP: Mirror: Archive> 2. <Private: BOTH: Don’t Mirror: Don’t Archive> 3. <Operational: DOWN: Don’t mirror: Archive) But of the 60 possible combinations of {Transfer x Mirror x Archive x Notify}, 11 categories are actually required to cover the 31 file types. The exceptions and variations in notification cause file handling to be highly dependent on the type of file and customer. For example, notification must take into account how the file was delivered and whether the customer is on the voice loop system. We find some principles (e.g., modify the flight note if any). But when we add further exceptions (e.g., is this a shuttle flight or a “stage” in the ISS expedition?), we end up with seven categories that are Draft: May 1, 2008 35 exceptions to the rules. Not surprisingly, the OCA officers’ manual provides one procedure per file type, rather than a set of principles (e.g., “what to do with a down- linked image file” or “how to handle operational software data”). Clancey (1992, pp 15-16) found that the relations required in an ontology depended on the modeling purposes (e.g., diagnosis, teaching, knowledge acquisition). We find in OCAMS this same process-specific, conditional character—for each function added (mirroring, archiving, notifying) the ontology becomes more branched and specific, such that 3 categories cover only 12/31 = 38% of the file types, and 8 more categories are required to cover the remaining 62%. Six categories include only one or two file types, and most are exceptions to the general rule of “mirror uplinks, don’t mirror downlinks.”16 Following the four principles listed above, we could probably write rules to infer whether a new file type provided by a customer is mirrored, archived, and how the customer is notified. But this much is obvious to the customer, too. Automation to infer file handling is not required, a customer could add new file types to the system by filling out a simple form like Table 1. Further modeling and automation could eliminate the need for a flight controller to approve modifications to the OCAMS rules operation, but this would clearly fall into the realm of a more advanced tool, after OCAMS has been deployed and its methods and 16 Every combination of {Direction x Mirror} occurs except BOTH/DOWN, DOWN/YES, and UP/NO, fitting the principle to mirror uplinks (hence the rule is BOTH/BOTH or BOTH/DOWN) and not to mirror downlinks (hence the rule is BOTH/UP and DOWN/NO). But there are exceptions, namely BOTH/BOTH for software configuration files (which violates both rules) and BOTH/YES (for email). Draft: May 1, 2008 36 capability accepted and understood in practice. That is to say, the nature of the file management work system, which already has a practice for changing procedures, suggests a manual method for updating at first, rather than attempting to eliminate human checks and balances. Towards a Work Systems Design Library: Components Reused in OCAMS Can we relate the design of OCAMS to the generic tasks and problem solving methods of 2nd GES? A first step is to ask what components are reused in Brahms models and workflow tools, besides of course the work practice schema in the Brahms language. Here are examples from OCAMS that relate to systems integration: o Agents that communicate with external systems (Comm Agents) inherit behaviors from a group (AbstractCommunicationAgent) that handles memory management and supports both simulation and real-time modes. o An FTP client library has been reused in multiple Mobile Agents configurations.17 o The Brahms “base library” includes: o Basic file operations (copy, delete, checksum verification, etc) represented as a Brahms Input/Output group with file manipulation Java activities that other Brahms agents can inherit. 17 One of the first examples of an “intelligent agent” was a program that managed files using FTP (Anderson & Gillogly, 1976). A favorite joke was that if the agent were told to move a directory in the most efficient manner possible, it might first delete all the files. Draft: May 1, 2008 37 o A Brahms Communicator group with activities to create and read Communicative Acts (inspired by Searle’s [1969] speech act theory and based on the FIPA standard agent communication language for multi-agent systems). o A Brahms JavaUtility group with activities to manipulate Java objects, read Java object values, and manage properties. We have used a table-driven method for different purposes in Brahms models. For example, in OCAMS and two previous simulations of operations, a spreadsheet representing a work schedule for one or more groups is interpreted to initialize agent beliefs about what activities are done when (i.e., the schedule timeline) (Seah, et al., 2005). Some of the specific workframe and thoughtframes that relate to schedules are reused. We expect that this method for modeling scheduled operations will play a role in most mission simulations, and more generally applicable for scheduling an agent in any domain. In OCAMS the table-driven method is also adapted to generate thoughtframes and workframes about file types and handling rules (beliefs about attribute/values of file paths, file names, file extensions, etc.). That is, the spreadsheet serves as a knowledge- acquisition method by organizing information required from domain specialists, a means of presenting the model to others, and a means for changing the model (an agent’s initial beliefs and activities). After building a variety of mission operations simulations, it seems clear that the simulating and automating workflow operations requires agents to maintain beliefs about the work in process, represented as sets of objects (e.g., the files being uplinked and Draft: May 1, 2008 38 downlinked), constituting a central part of the agent’s “situation-specific model” of the work system. Other aspects of the SSM include beliefs about the shift schedule (who is coming in to take over the role). We can expect these constructs to be adapted in the future. In summary, the Brahms language enables formalizing and reusing model constructs, realizing the 2nd GES concept of representational tools with components more specific than “inference rule” and “schema.” However, 2nd GES research especially focused on abstraction of methods for manipulating models. Of what value is abstracting tasks and methods in developing a system like OCAMS? Applying the System-Task-Operator Framework to OCAMS: From Expert Systems to Workflow Systems Representing an ontology of file types and different file handling operations as agents (mirroring, monitoring, archiving) derives directly from 2nd GES approach of developing a domain ontology and functional (task-specific) operators. Here is how OCAMS fits the System-Task-Operator framework: o The system being modeled is the ISS file system, including workstations, directory structure, and types of files. These file types are related to types of customers (e.g., physicians, mission planners) and two broad functions in which the files play a part (operational and medical/personal). o With respect to the ISS file system, the task is configuration— assembling/maintaining another file system with certain properties (mainly mirroring the ISS file system minus medical/personal files and images). Put another way, the configuration task here is to replicate a given structure (the ISS file structure) on a “mirror” server, in which the structure of the secondary Draft: May 1, 2008 39 system (the Mirror LAN) is subject to certain general constraints (namely, what file types are mirrored), which constitute “configuration rules.” The problem-solving method is simple classification: The file name defines the file type and this defines how the file is to be configured in the Mirror LAN (the options are: Copy, Unzip and monitor for errors, Delete, and Do nothing). An obvious reaction to presenting OCAMS as an example for appraising the value of 2nd GES analysis is that OCAMS is not a heuristic program (yet), so the most trivial method—simple classification—suffices. One could argue that ISS file management is not the kind of systems modeling task addressed by the 2nd GES analysts.18 However, the file management problem actually being solved by OCAMS is more complex than it might first appear: o OCAMS is actually building a physical system (the Mirror LAN), not just a representation of a design for the file system. o OCAMS is coordinating the general model (ontology and file handling rules) with a situation-specific model of the desired configuration (SSM-CONFIG— what files need to be mirrored, archived, and monitored at this time) with a 18 Errors do occur and put the file system (and agent system) into an uncertain state. But rather than modeling and reasoning in an expert system “diagnosis and repair” approach, failure handling is automated in OCAMS by an “administration agent” that simply restarts the agent processes and redoes the operations in the current “batch” of files. When more is required, a person usually needs to do something that is out of the scope of automation (e.g., deciding how to diplomatically handle an ISS crew member’s overgrown mail file). Draft: May 1, 2008 40 situation-specific model of the current state of the world (SSM-MIRROR— the state of the Mirror LAN and SSM-ISS—the state of the ISS file system). o Besides using simple classification for creating the SSM-CONFIG from the SSM-ISS, OCAMS uses the methods of queuing, handshake protocol, retry- iteration, and synchronization to maintain the Mirror LAN system (according to SSM-MIRROR). In other words, OCAMS is not just a reasoning system, a model manipulator. OCAMS is an interactive system, an actor in the world, which uses model-based methods to plan its actions (SSM-CONFIG) and keep track of the work to be done (SSM- MIRROR). In some respects, it is as if Mycin were charged not just with interpreting culture results, but actually treating a patient. More prosaically, this is the difference between an expert system and a workflow system. Consequently, the software engineering problems in designing OCAMS are complex and go well beyond the expert systems framework. In particular, coordinating the operations of OCAMS agents that run as distributed processes on six or more workstations is not trivial. File manipulation is the easy part. The dominant effort in requirements definition and system building involved: 1) Enabling communications between computers on different networks owned by different organizations, 2) Security of mission systems and private data using secure communication (SSL), 3) Customization of file management for special requests, 4) Verification and notification that customer requests are complete, and 5) Recordkeeping for handover logs and ongoing mission documentation. These are recurrent software engineering considerations for building office workflow tools. Draft: May 1, 2008 41 Because of the advent of distributed computing, the internet, security concerns, multimodal interfaces, multiple vendor platforms, etc., having a library of PSMs is just one part of what is required in a practical toolkit for building model-based systems today. To further understand the requirements, we will consider how extending OCAMS’ functionality involves much more than assembling additional PSMs or configuring them differently. A Broader View: The Mission Operations Work System As we broaden OCAMS’ original mirroring function to cover other work done by the OCA officer, our perspective of “the system” being reasoned about and manipulated changes, and the focus on maintaining proper interactions with other players (people and tools) becomes more central. When we include the customers who are delivering files for uplink and receiving down-linked files, we see that the work system involves people performing other roles in the Mission Operations Directorate,19 astronaut family members, and flight controllers in other countries in support of three to thirteen astronauts (assuming a full Shuttle flight of seven astronauts occurs with a full contingent of six onboard the ISS). This is a work system, a distributed collaboration among people using diverse representations and tools—physicians, aeronautics engineers, planners, robotics engineers, power and propulsion flight controllers, family members, etc. 19 MOD is the organization within the NASA Johnson Space Center that operates the Mission Control Center, usually associated with a room with three large monitors called the Flight Control Center (FCR). The OCA officers work within MCC (a secure building), but in another room, of one several “backrooms” where people support the flight controllers in the FCR, whom they can hear and speak to via the voice loop. Draft: May 1, 2008 42 From this perspective, an extended OCAMS that automates all of the work of the OCA officer must be designed as an actor (agent) in a work system. This agent would be responsible for retrieving files from different locations (file servers, hard drives), interpreting documents, controlling different subsystems (e.g., software programs such as FTP), creating structured documents (logs), and communicating with people (via a GUI, email, and perhaps someday by speaking on the voice loop). The comprehensive process would require multiple ontologies: file types, roles in operations, mission phases. File management correspondingly becomes a higher-order system configuration problem—such as prioritizing file transfers for different customers, given limited bandwidth and fragmented communication windows between the ground and ISS. If OCAMS were extended so it required such planning, we would probably couple it to a constraint-based tool, such as SPIFe (McCurdy, et al., 2006), rather than represent a planning capability in Brahms. Indeed, some of the files managed by OCAMS are maintained by a model-based scheduler (OSTPV; Frank, et al., in press). This example suggests, just like the functional-location decomposition of OCAMS into agents, that the preferred software engineering approach today is not construction of single programs from components, but integration of modules—often running on different platforms— that can flexibly communicate in real-time settings.20 20 For example, in problem solving research the notion of “memory” focuses on efficient matching; for an agent in an operational environment, the practical issue broadens to storing facts to engage in discourse about events that occurred days or months ago. In MOD, a flight controller might ask his/her personal agent, “Have we uplinked files like this to crew members on previous flights?” Draft: May 1, 2008 43 So again, we see that reusability of components is important, but the systems engineering problem now includes integration and interoperability of tools or services, not just copying a formalism used in one program into another. Furthermore, just as the shift from a product-centered model to a behavioral-practice model constituted a shift from of detail from functions to transactions (client-server-product relations), this shift to an agent in a work system must focus on maintenance of transactions, which involves substantial negotiation of requirements and resources, whose status must be tracked, confirmed, communicated, and sometimes renegotiated. Conclusions In applying 2nd GES methods, we have developed Brahms, a tool for modeling work practice. Brahms is not a program that automates the design of work systems, which was the original expert systems vision. Rather, Brahms is a language and tool for expert designers.21 By enabling designers to model and simulate alternative work system designs in different scenarios, Brahms serves like any simulation model in science and engineering. It enables better understanding causal processes (e.g., relations between roles, schedules, procedures, and workplace automation), measuring work systems flows (e.g., productivity), identifying bottlenecks, predicting how the work system might fail, and evaluating hypothesized improvements. In using the simulation-to-implementation methodology we have found that a work practice simulation has a range of purposes over time: formalizing a particular aspect of practice to produce metrics useful for improving how the work is done; modeling and 21 Brahms is currently developed and maintained with the NASA Ames group called “Work Systems Design and Evaluation.” Draft: May 1, 2008 44 simulating agents that automate aspects of the work; simulating how a workflow tool would be used in practice (again with metrics); deploying an agent-based workflow tool on distributed platforms; and then by comparison with the tool in use, potentially improving the formalization of work systems and human behavior in the Brahms language and engine. In developing OCAMS, we have shifted our perspective from building a “problem solving” (expert) system to building a workflow system, which inherently must become an actor in the world. Consequently, the library of reusable components is at a higher level than the model manipulation processes formalized in PSMs, involving integration with particular types of subsystems (e.g., handling high-volume telemetry), assisting people over time (e.g., managing a work plan), relating different representations (e.g., maps and databases), communicating in different modes (e.g., email, voice mail, conversations), and even methods for moving and behaving in location-dependent ways (e.g., robotic systems that avoid obstacles and follow people). Nevertheless, the main idea behind 2nd GES research, that abstraction of systems, tasks, and methods could make building future systems easier, has certainly been central in our methodology. The abstractions that have guided the development of OCAMS combine concepts and methods from software engineering and from our own multiagent systems: 1) a layered architecture, 2) functional decomposition of services into agents, 3) providing a “personal agent” for interacting with the person using the tool, 4) distributed implementation capability, 5) abstraction of domain relations into a separate domain model, enabling, for example, a table-driven process, 6) general methods for systems integration (namely, using JAVA to write a “comm agent” that mediates between an API Draft: May 1, 2008 45 and other Brahms agents), 7) handshake communication protocols for tracking the status of subsystems, 8) categorizing agent messages (e.g., request, information, subscription, proposal, in the Brahms Communication Library). Reflecting on the appropriate “grain size” for PSMs (or software reuse more generally), the trend appears to be towards programs that use specialized representations (e.g., a science database), carry out high-level modeling tasks (e.g., scheduling), or mediate between such programs (e.g., Brahms Comm Agents). This confirms the perspective of Newell and Simon (1972), restated by McDermott (1988), that the reusable component or method is the overall computational process, a self-contained package (such as a Brahms agent). Other reports of progress in developing software libraries affirm our experience that a significant opportunity for abstraction and reuse lies in higher-level components (Choo & Skura, 2004, p. 4): SciBox uses the data analysis components from [the System Independent] layer to build data analysis packages specific to space operation simulations but not specific to any particular space mission. Examples of SciBox software components are common mathematical algorithms used in celestial mechanics and astronomy, map projection, coordinate transformation, and scheduling and commanding. From yet another perspective, the Brahms work systems ontology (i.e., groups, agents, activities, workframes, etc.) for simulating practices remains fixed across domains and applications, providing an interesting twist on the idea of formalizing PSMs (Clancey, et al., 1998; Sierhuis, 2001). Brahms is analogous to the knowledge engineering tools used in the 1980s with their built-in abstractions for modeling and Draft: May 1, 2008 46 reasoning about causal processes. However, Brahms’ framework is in effect a language for modeling problem solving methods, but in a much broader sense intended in the 1980s, namely how models are created, manipulated, and used when detecting and solving problems in the real world—work practice. We conclude with a claim and a hypothesis. Our claim is that work on 2nd GES was a reasonable, well-grounded engineering phase of research that aimed to analyze expert systems, abstract methods, and potentially make system building more efficient through tools with libraries of PSMs. Our hypothesis is that such libraries never became widely used in software engineering for multiple reasons: o The dominant challenge in developing a new workplace tool is proper integration into a complex, distributed work system involving people and other tools; o The most obvious automation opportunities can be handled algorithmically because people need to be responsible for value-based judgments requiring diplomacy and sensitivity (e.g., how to handle a new astronaut’s request to introduce a new procedure) and people usually need to be involved where physical reconfigurations are required (e.g., substituting a new computer); o When components are “reused” they are usually large programs (e.g., a planner) or hardware (e.g., camera) integrated into a larger workflow system, and such integration is specialized because it involves representational mapping between ontologies (e.g., integrating a camera with email and a database; Clancey, et al., 2005a). Draft: May 1, 2008 47 More broadly, by outlining the historical development of problem solving research, expert systems, and situated cognition, we argued that PSMs are not adequate for modeling how people discover, articulate and solve problems, that is, the practice of human problem solving. The issue is not so much that people might model the world differently (an issue central to developing instructional programs), but that simulating what people are doing in the course of a day is better characterized in terms of “activities,” rather than only “solving problems.” The methods of practice are interactive, employing reasoning for and through action in the real world. Interactive methods relate internal system models to actions in the world in a manner that carries out the agent’s responsibilities in a sustained way over time, including direct observation, communicating with people and other tools, coping with failure (retrying, reconciling models and reality), and detecting when assistance is required. The twist is that a problem solver must not just model the world to reason about it, but actually uses such models to keep the world in order. In conclusion, the analytic thrust of 2nd GES research was appropriate and still makes sense, however the belief that many practical tools would be constructed from PSM primitives alone was wrong. Automation tools are not standalone problem solvers— whether diagnosticians, therapists, or designers—but like human physicians and engineers, such tools are properly conceived as agents, which are frequently interacting with people and other systems, to categorize, negotiate, and communicate their requests and contributions in the work environment. This cooperative endeavor can be viewed as a higher-order “modeling problem” of configuration, diagnosis, planning, and so on. But the work itself is emergent, out of the control of any particular person or tool. Thus for an Draft: May 1, 2008 48 “expert system” the most practical, reusable methods are ways of interacting with people and other systems to handle discrepancies between models (beliefs, plans, procedures, theories) and the world. In the expert systems formulation, the paradigm is applying a model to solve a problem. In practice, expertise includes knowing how to deal with models that don’t apply, such as seeking supervisory assistance for dealing with a procedural category that doesn’t fit the current situation, negotiating with peers across disciplines or shifts about who will take responsibility for certain problems, and understanding economic and political perspectives by which actions will be evaluated. These concepts—assistance or permission, responsibility, and non-technical perspectives—move work practice into the realm characterized by Simon as “ill- structured problems” (1973). Complex events can call into question the validity of models and policies. “Tear” in models (Burton & Brown, 1979, p. 95) is sometimes handled by creating new categories, giving new interpretive twists to rules and procedures by blending otherwise conflicting values, and other methods of deferring, reassigning, or even defining away the problematic situation.22 Such adaptation can be difficult when different analytic perspectives (e.g., scientific, ethical, economic, political world views) are at cross purposes (Schön, 1987). Here, in saying that cognition is situated we mean that expertise is inherently distributed, not all technical, and dynamically constructed in an ongoing social process. Ill-structured problems transcend 22 Here we are reminded of the Columbia disaster, in which a simulation model was interpreted to argue that foam could not damage the Space Shuttle, and hence photographs of possible damage (taken from Earth) would not be necessary (Columbia Accident Investigation Board, 2003). Draft: May 1, 2008 49 the ontology and library of methods. The problem solver asks: Are my models adequate? Have I interpreted them appropriately? Do I need to work harder to prove my proposed actions are valid? As software engineers move into this realm, with programs becoming actors in the workplace, the challenge of designing problem solving agents is to facilitate human responsibility, not just by automating routine tasks, but by deferring to people when necessary, and revealing how the models might be wrong. Acknowledgements OCAMS has been developed in partnership with the OCA officers in Mission Operations Directorate of NASA Johnson Space Center, particularly Chris Buckley, Deborah Hood, Skip Moore, Fisher Reynolds, and Karen Wells. We are grateful for the vision and support of Tim Hall and Brian Anderson at NASA JSC and Mike Shafto at NASA Ames. Mike Scott and Ron van Hoof (QSS Group, NASA Ames) have played key roles in implementing Brahms and OCAMS. This work has been supported in part by funding from NASA’s Constellation Program. Draft: May 1, 2008 50 References Acquisti, A., Sierhuis, M., Clancey, W. J., and Bradshaw, J. M. (2002). Agent-based modeling of collaboration and work practices onboard the International Space Station. Proc. Eleventh Computer-Generated Forces and Behavior Representation Conf, pp. 181-188. Bond, A. H. & Gasser, L. (1988). Readings in Distributed Artificial Intelligence. San Mateo, CA: Morgan Kaufmann. Anderson, R. H. (1977). The Use of Production Systems in RITA to Construct Personal Computer “Agents.” SIGART Newsletter, 63 (June), 23-28. Anderson, R. H. & Gillogly, J. J. (1976). Rand intelligent terminal agent (RITA): Design philosophy. RAND Report R-1809-ARPA, The Rand Corporation. Brooks, R.A. (1991). How to build complete creatures rather than isolated cognitive simulators. In Architectures for Intelligence (VanLehn, K., Ed.), pp. 225-239. Hillsdale, NJ: Lawrence Erlbaum Associates. Burton, R. R., & Brown, J. S. (1979). An Investigation of Computer Coaching for Informal Learning Activities. Int. J. of Man-Machine Studies, 11 (1), 5-24. Carley, K. (1990). Group Stability: A Socio-Cognitive Approach. In Advances in Group Processes, Advances in Group Processes: Theory and Research (Lawler, E. J., Markovsky, B., Ridgeway, C., Walker, H. A., Eds.), Vol. 7, pp. 1-44. Greenwich, CT: JAI Press. Chandrasekaran, B. & Johnson, T. R. (1993). Generic Tasks And Task Structures: History, Critique and New Directions. Second Generation Expert Systems (David, J. M., Krivine, J. P., & Simmons, R., Eds.), pp. 239-280. New York: Springer Verlag. Draft: May 1, 2008 51 Choo, T. H. & Skura, J. P. (2004). SciBox: A software library for rapid development of science operation simulation, planning, and command tools. Johns Hopkins APL Tech. Dig. 25(2), 154–162. Clancey, W. J. (1984). Methodology for building an intelligent tutoring system. Method and Tactics in Cognitive Science, (Kintsch, W., Miller, J. R., and Polson, P. G., Eds.), pp. 51-83. Hillsdale, NJ: Lawrence Erlbaum Associates. Clancey, W. J. (1985). Heuristic classification. Artificial Intelligence, 27, 289-350. Clancey, W. J. (1986). Qualitative student models. Annual Review of Computer Science, pp. 381-450. Palo Alto: Annual Reviews Inc. Clancey, W. J. (1989). Viewing knowledge bases as qualitative models. IEEE Expert: Intelligent Systems and Their Applications, 4 (2), 9-15, 18-23. Clancey, W. J. (1992). Model construction operators. Artificial Intelligence, 53(1), 1-124. Clancey, W. J. (1993a). The knowledge-level reinterpreted: Modeling socio-technical systems. Int. J. of Intelligent Systems, 8(1), 33-49. Clancey, W. J. (1993b). Situated action: A neuropsychological interpretation (Response to Vera and Simon). Cognitive Science, 17(1), 8-116. Clancey, W. J. (1997a). Situated Cognition: On Human Knowledge and Computer Representations. New York: Cambridge University Press. Clancey, W. J. (1997b). The conceptual nature of knowledge, situations, and activity. In Human and Machine Expertise in Context (Feltovich, P., Ford, K., & Hoffman, R., Eds.), pp. 247–291. Menlo Park, CA: AAAI Press. Clancey, W. J. (1999). Conceptual Coordination: How the mind orders experience in time. Hillsdale, NJ: Lawrence Erlbaum. Draft: May 1, 2008 52 Clancey, W. J. (2002). Simulating activities: Relating motives, deliberation, and attentive coordination. Cognitive Systems Research, 3(3), 471-499. Clancey, W. J. (2004). Roles for agent assistants in field science: Personal projects and collaboration. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews (Special Issue on Human-Robotic Interaction). 34(2), 125- 137. Clancey, W. J. (2005). Towards on-line services based on a holistic analysis of human activities In Towards the learning GRID: Advances in human learning services. Series Frontiers in Artificial Intelligence and Applications (Ritrovato, P., Allison, C., Cerri, S. A., Dimitrakos, T., Gaeta, M., Salerno, S., Eds.), pp. 3-11. Amsterdam: IOS Press. Clancey, W. J. (2006). Observation of work practices in natural settings. In Cambridge Handbook on Expertise and Expert Performance (Ericsson, A., Charness, N., Feltovich, P., & Hoffman, R., Eds.), pp. 127-145. New York: Cambridge University Press. Clancey, W. J. (in press). Scientific antecedents of situated cognition. To appear in Cambridge Handbook of Situated Cognition (Robbins, P., & Aydede, M., (Eds.). New York: Cambridge University Press. Clancey, W. J. & Barbanson, M. (1991). TOPO: Implications of the system-model- operator metaphor for knowledge acquisition. IEEE Expert, 6(5), 61-65. Clancey, W. J. & Letsinger, R. (1981). NEOMYCIN: Reconfiguring a rule-based expert system for application to teaching. Proc. of Seventh IJCAI, pp. 829-826. Clancey, W. J., Sachs, P., Sierhuis, M., & van Hoof, R. (1998). Brahms: Simulating practice for work systems design. Int. J. Human-Computer Studies, 49, 831-865. Clancey, W. J., Sierhuis, M., Alena, R., Berrios, D., Dowding, J., Graham, J. S., Tyree, K. S., Hirsh, R. L., Garry, W.B., Semple, A., Buckingham Shum, S.J., Shadbolt, N. Draft: May 1, 2008 53 and Rupert, S. (2005a). Automating CapCom using Mobile Agents and robotic assistants. American Institute of Aeronautics and Astronautics 1st Space Exploration Conf. NASA TP 2007-214554, Available: http://ntrs.nasa.gov. Clancey, W. J., Sierhuis, M., Damer, B., and Brodsky, B. (2005b). The cognitive modeling of social behavior. In Cognitive modeling and multi-agent interaction (Sun, R., Ed.), pp. 151-184. New York: Cambridge University Press. Clancey, W. J., Sierhuis, M., Seah, C., Buckley, C., Reynolds, F., Hall, T., and Scott, M. (in press). “Multi-Agent Simulation to Implementation: A Practical Engineering Methodology for Designing Space Flight Operations.” Engineering Societies of Agents Workshop, Athens, Greece. Cohen, P. R., Greenberg, M. L., Hart, D. M., & Howe, A. E. (1989). Trial by fire: Understanding the design requirements for agents in complex environments. AI Magazine 10(3), 34-48. Columbia Accident Investigation Board. (2003). CAIB Report, Volume 1. NASA. (Available: http://www.caib.us/news/report/volume1/default.html). Damasio, A. (1994). Descartes’ error: Emotion, reason, and the human brain. New York: G. P. Putnam’s Sons. David, J. M., Krivine, J. P., & Simmons, R., Eds. (1993). Second Generation Expert Systems. New York: Springer Verlag. Davis, R. (1980). Metarules: Reasoning about Control. Artificial Intelligence, 15, 179- 222. Dewey, J. (1896). The reflex arc concept in psychology. Psychological Review, 3, 357- 370. Draft: May 1, 2008 54 Dewey, J. (1939). Experience, Knowledge and Value: A Rejoinder. In Philosophy of John Dewey (Schilpp, P. A., & Hahn, L. E., Eds.), 3rd ed. La Salle: Open Court. Feigenbaum, E. A. (1977). The Art of Artificial Intelligence: Themes and Case Studies of Knowledge Engineering. Proc. IJCAI, pp. 1014-1029 Elman, J. L. (2004). An alternative view of the mental lexicon. Trends in Cognitive Science, 7, 301-306. Feigenbaum, E. A., Buchanan, B. G., & Lederberg, J. (1971). On generality and problem solving: A case study using the DENDRAL program. In Machine Intelligence (Meltzer, B., & D. Michie, D., Eds), Vol. 6, pp. 165-190. New York: American Elsevier. Frank, J., Morris, P. H., Green, J., & Hall, T. (in press). The Challenge of Evolving Mission Operations Tools for Manned Spaceflight. Proc. 9th iSAIRAS 2008. Freed, M., Shafto, M. G., Remington, R. W. (1998). Employing Simulation to Evaluate Designs: The APEX Approach. In Engineering for Human-Computer Interaction (Chatty, S., & Dewan, P., Eds.), pp. 207-223. Dordrecht: Kluwer Academic. Gilbert, N. & Doran, J. (1993). Simulating Societies: The computer simulation of social phenomena. London: UCL Press. Green, C. (1969). Applications of theorem proving to problem solving. In Proc. of the Int. Joint Conf. on Artificial Intelligence (Walker, D. E., & Norton, L. M., Eds.), pp. 219-239. Greenbaum, J., & Kyng, M. (Eds.). (1991). Design at work: Cooperative design of computer systems. Hillsdale, NJ: Lawrence Erlbaum. Draft: May 1, 2008 55 Greeno, J. G. (1980). Some examples of cognitive task analysis with instructional implications. In Aptitude, learning, and instruction: Vol. 2. Cognitive process analyses of learning and problem solving (Snow, R. E., Federico, P. A., & Montague W. E., Eds.), pp. 1-21. Hillsdale, NJ: Lawrence Erlbaum. Hayes-Roth, F., Waterman, D. A., & Lenat, D., Eds. (1983). Building Expert Systems. Reading, MA: Addison-Wesley. Hutchins, E., & Palen, L. (1997). Constructing meaning from space, gesture, and speech. In Discourse, Tools and Reasoning: Essays on Situated Cognition (Resnick, L., Säljö, R., Pontecorvo, C., & Burge, B., Eds.), pp. 23-40. Berlin: Springer-Verlag. McCurdy, M., Pyrzak, G., Ratterman, C., & Vera, A. (2006). The Design of Efficient Ground Software Tools. In Proc. of the 2nd IEEE Int. Conf. on Space Mission Challenges For Information Technology, p. 257. McDermott, J. (1988). Preliminary steps toward a taxonomy of problem-solving methods, In Automating Knowledge Acquisition for Expert Systems (Marcus, S., Ed.), pp. 225- 256. Boston: Kluwer Academic Publishers. Minsky, M. (1969). Semantic Information Processing. Cambridge, MA: MIT Press. Minsky, M. (1985). The Society of Mind. New York: Simon and Schuster. Newell, A., & Simon, H. A. (1972). Human problem solving. Englewood Cliffs, NJ: Prentice-Hall. Nii, H. P. (1986). Blackboard Systems. AI Magazine, 7(2), 38-53 and 7(3), 82-106. Papert, S. (1972). Teaching Children to Be Mathematicians vs. Teaching About Mathematics. Int. J. of Mathematics Education and Science Technology, 3, 249-262. Pollack, M. (1992). The Uses of Plans. Artificial Intelligence, 57, 43-68. Draft: May 1, 2008 56 Pople, H. E., (1977). The Formation of Composite Hypotheses in Diagnostic Problem Solving: An Exercise in Synthetic Reasoning. Proc. IJCAI, pp. 1030-1037. Cambridge, MA. Schank, R. C. (1982). Dynamic memory: A theory of reminding and learning in computers and people. Hillsdale, NJ: Lawrence Erlbaum. Schön, D. A. (1987). Educating the reflective practitioner: Toward a new design for teaching and learning in professions. San Francisco: Jossey-Bass. Schreiber, A. T., Wielinga, B. J., de Hoog, R., Akkermans, J. M., & Van de Velde , W. (1994). CommonKADS: A comprehensive methodology for KBS development. IEEE Expert, 9(6), 28-37. Seah, C., Sierhuis, M., & Clancey W. J. (2005). Multi-agent modeling and simulation approach for design and analysis of MER mission operations. Proc. Int. Conf. on Human-Computer Interface Advances for Modeling and Simulation, pp. 73-78. Searle R. (1969). Speech Acts: An Essay in Philosophy of Language. Cambridge University Press. Sierhuis, M. (2001). Modeling and simulating work practice. Ph.D. thesis, Social Science and Informatics (SWI), University of Amsterdam, The Netherlands. Sierhuis, M., & Clancey, W. J. (1997). Knowledge, Practice, Activities, and People. In Proc. AAAI Spring Symposium on Artificial Intelligence in Knowledge Management (Gaines, B., Ed.), pp. 142-148. Sierhuis, M., W. J. Clancey, Seah, C., Trimble, J., & Sims, M. H. (2003). Modeling and simulation for mission operations work systems design. J. of Management Information Systems, 19(4), 85-128. Draft: May 1, 2008 57 Sierhuis, M., Clancey, W. J., & van Hoof, R. (2007). Brahms: A multiagent modeling environment for simulating work practice in organizations. Int. J. for Simulation and Process Modeling, 3(3), 134-152. Simon, H. A. (1973). The structure of ill-structured problems. Artificial Intelligence, 4(3), 181-202. Simon, H. & Lea, G. (1974). Problem solving and rule induction. Models of Thought (Simon, H. A., Ed.), pp. 329-346. New Haven: Yale University Press. Vera, A., and Simon, H. (1993). Situated Action: Reply to William Clancey. Cognitive Science, 17(1), 117-135. Vicente, K. J. (1999). Cognitive work analysis: Toward safe, productive, and healthy computer-based work. Mahwah, NJ: Erlbaum. Wallace, B., & Ross, A. (2006). Beyond Human Error: Taxonomies and Safety Science. Boca Raton, FL: CRC Press. Wallace, B., Ross, A., Davies, J. B., & Anderson T., (Eds.) (2007). The Mind, the Body and the World: Psychology after Cognitivism. London: Imprint Academic. Whitehead, A. N. & Russell, B. (1910/1913). Principia Mathematica. Cambridge, Cambridge University Press. Wittgenstein, L. [1953] (1958). Philosophical Investigations. New York: Macmillan. Wynn, E. (1991). Taking Practice Seriously. In Design at Work: Cooperative design of computer systems (Greenbaum, J., & Kyng, M., Eds.), pp. 45-64. Hillsdale, NJ: Lawrence Erlbaum. Draft: May 1, 2008 58 Author Biographies William J. Clancey works at the NASA Ames Research Center and Florida Institute of Human and Machine Cognition. He received a BA degree in Mathematical Sciences from Rice University (1974) and a PhD in Computer Science from Stanford University (1979). Chief Scientist for Human-Centered Computing, in the Intelligent Systems Division at Ames, he has extensive experience in medical, educational, and financial software and was a founding member of the Institute for Research on Learning. He is especially interested in relating social science and neuropsychology to descriptive (symbolic) models of cognition to understand the nature of consciousness. Maarten Sierhuis is a Senior Research Scientist and Lead of the Autonomy and Decision Support group at RIACS/USRA, located at NASA Ames Research Center. He is a Co-Principal Investigator for the Brahms project, working in the Work Systems Design & Evaluation group in the Collaborative and Assistant Systems area within the Intelligent Sciences Division at NASA Ames Research Center. Previously, he worked at NYNEX Science & Technology. He received a PhD in Social Science Informatics from the University of Amsterdam and holds an engineering degree in Informatics from the Polytechnic University in The Hague, The Netherlands. Chin Seah is a computer scientist at Science Applications International Corporation (SAIC), working at NASA Ames Research Center on the Brahms project. He has applied the Brahms work system design and modeling approach to the Mars Exploration Rover and International Space Station mission operations. Before joining the Brahms team, he worked as a business process management consultant at Andersen Consulting and as a knowledge engineer at Mindbox, Inc. implementing rule-based and case-based expert Draft: May 1, 2008 59 systems. He has a B.S. in computer engineering from Santa Clara University and an M.S. in computer information science from the University of Pennsylvania. Draft: May 1, 2008 1 Figure 1. OCA Mirroring System (OCAMS) using model-based systems integration to automate some of the file management between ground support and the International Space Station (ISS). Draft: May 1, 2008 1 Table 1. OCAMS Ontology of File Types and Handling Procedures. Thirty-one file types are categorized by the function of the data and/or customer providing or using the data. Transfer directions refer to “UP” to the International Space Station and “DOWN” from ISS to the Earth. A mirrored file is copied to a ground-based duplicate of the ISS file system. An archived file is saved in a dated folder indicating its source. Ground support customers are notified in different ways, including a workflow “flight note” system, a speaker-headset intercom (“voice loop”), email, telephone, or by speaking outloud to someone across the room at another workstation. File Types Type of Data Transfer Direction Mirror? Archive? Notify? <Symbolic Name> { Operational Plans & Procedures | Operational Software & Schedule Changes | Personal or Medical | Exceptions} {UP | DOWN | BOTH} {YES | NO} {YES | NO} {FlightNote | VoiceLoop | Email | Phone | Outloud} 
work_viehff56qvdjrhuqk3pach75ge	----	A Robust Profit Measure for Binary Classification Model Evaluation∗ Franco Garrido1, Wouter Verbeke2, and Cristián Bravo3 1Programa de Maǵıster en Gestión de Operaciones, Universidad de Talca, fgarridoc@alumnos.utalca.cl 2Faculty of Economic and Social Sciences and Solvay Business School, Vrije Universiteit Brussel, Belgium, wouter.verbeke@vub.be 3Department of Decision Analytics and Risk, Southampton Business School, University of Southampton, c.bravo@soton.ac.uk Abstract Using profit-based evaluation measures is a necessity in business- oriented contexts, as they aid companies in making cost-optimal de- cisions. Among the measures that effectively include the true na- ture of costs and benefits in binary classification, the expected max- imum profit (EMP) has been used successfully for churn prediction and credit scoring, and defined in general for binary classification problems. However, despite its competitive results against the most frequently used measures, the EMP relies on a fixed probability dis- tribution of costs and benefits, the range of which in real applications is not entirely known. In this paper, we propose to extend this mea- sure by adding random shocks to these distributions. We call this new measure the R-EMP, following the convention of the analogous EMP measure. Our metric adds a stochastic component to each point of ∗ NOTICE: this is the author’s version of a work that was accepted for publication in Expert Systems with Applications in September 19, 2017, published online as a self-archive copy after the 24 month embargo period. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. Please cite this paper as follows: Franco Garrido, Wouter Verbeke, Cristin Bravo, A Robust Profit Measure for Binary Classification Model Evaluation, In Expert Systems with Applications, 2017, Accepted: Available Online https://doi.org/10.1016/j.eswa.2017.09.045. 1 https://doi.org/10.1016/j.eswa.2017.09.045. the cost-benefit distributions, assuming that costs and benefits have a fixed probability, but its distribution range is subject to an external shock, which can be different for each cost or benefit. The experimen- tal set-up is focused on a credit scoring application using a dataset of a Chilean financial institution, with the attribute selection for a logistic regression being accomplished using the AUC, EMP, H-measure, and R-EMP as the selection criteria. The results indicate that the R-EMP measure is the most robust metric for achieving the greatest profit for the company under uncertain external conditions. Keywords: Supervised Binary Classification; Business analytics; Per- formance measures; Profit-driven analytics 1 Introduction The development of performance measures for classification methods has become an important task in data analytics, given their critical role in oper- ations management (Baesens et al., 2009). In many industries, information analysis has become the only method of differentiation (Davenport, 2006). McAfee and Brynjolfsson (2012) stated the following: ”You can’t manage what you don’t measure”. The most common predictive analytics problem, binary classification, has the goal of classifying elements into one of two classes. Most models that are used to solve this problem, such as logistic regression or neural networks, return a continuous score that indicates how likely each case is to belong to one of the two classes, and it is up to the practitioners to determine the threshold that defines the frontier between the two classes. There is a wide variety of measures available for evaluating the performances of algorithms, among which the receiver operating characteristic (ROC) curve and, espe- cially, the area under the ROC curve (AUC) are the most frequently used (Bradley, 1997). Researchers have demonstrated that measures like the AUC are not suit- able for environments in which misclassification costs are different (Hand, 2009). There are measures that consider the true nature of costs effectively; among them, the H-measure (Hand, 2009), the maximum profit (MP) mea- sure (Verbeke et al., 2012), and the expected maximum profit (EMP) measure (Verbraken et al., 2013) are some of the best known. The last two measures are designed as total profit measures; the former (MP) assumes certainty in 2 cost parameters, obtaining the maximum benefit and the optimal threshold, while the latter (EMP) is a stochastic version of the MP measure, in which cost and benefit parameters are described by a probability distribution, lead- ing to the estimation of the expected maximum profit. In real applications, the MP and EMP measures have contributed to the selection of models aligned with the nature of costs and benefits. Some ap- plications of these measures include determining the optimal fraction of the consumers to be targeted for a churn prevention campaign at a telecommuni- cations company (Verbeke et al., 2012) and estimating the maximum profit of credit scoring models (Verbraken et al., 2014a). This paper presents a more robust version of the measure for the case in which the uncertainty comes not only from the profit parameters but also from external random shocks, which we call the Robust Expected Maximum Profit (R-EMP) measure. Our method is based on the rationale that random shocks can modify an originally rigid profit estimation; thus, if the distribu- tion of these potential shocks can be known, then the R-EMP will fit the profit estimation, taking into account this information. This paper is structured as follows: in Section 2, we describe the state-of- the-art profit-based performance measures for evaluating classification mod- els. Section 3 shows the R-EMP formulation, specifying its structure and all the considerations regarding its implementation. Section 4 presents the experimental design of this work, which consists of a synthetic case (Section 5) and an empirical case using loan data (Section 6). This section includes both the benchmark of the R-EMP against other measures and a case in which we show the use of the measure as a decision-making tool. Finally, conclusions are presented in Section 7. 2 Evaluation Measures for Classification Mod- els Within the field of predictive analytics, a classification problem refers to the task of determining a class label for an element from a set of known labels. When the number of possible labels is only two, this task is known as binary classification. Data mining/analytics models for supervised classification al- low determining the labels for new cases with unobserved labels, and binary classifiers usually return a probability of belonging to one of two classes, lead- 3 ing to the necessity of defining a threshold that separates the two classes, i.e., a cut-off value. According to Ali and Smith (2006), there is no unique measure that can be used to find the best classification model. Baldi et al. (2000) showed that the most frequently used measures are percentages, different kinds of distances, correlation, entropy, mutual information, and ROC curves. Various authors have applied ROC curves in many applications and used the area under the ROC curve, i.e., the AUC, as the performance measure, mainly because the AUC does not depend on a cut-off value and is insensitive to class distri- bution (Bradley, 1997). The AUC is also easily implemented (Brown and Davis, 2006) and interpreted (Fawcett, 2006a). Fawcett (2006b) showed that ROC graphs are not a suitable reference when there are instance-varying, i.e., case-dependent, classification costs. To fix this problem, he developed a variant called the ROCIV. Hand (2009) detected an additional AUC weak- ness and proposed an alternative measure known as the H-measure, which is coherent and therefore should yield a more reliable indication of performance than the area under the ROC curve. Correa Bahnsen et al. (2014) indicated a significant need that exists for measures that are sensitive to classification costs. They proposed an algorithm for credit scoring that allows construct- ing a classifier while simultaneously taking into account the variable nature of costs. Several publications presented measures or techniques for evaluat- ing classification models. Most of the proposed measures were compared to the AUC measure. Among these works, we find McDonald (2006) introduc- ing a measure that has the characteristic of allowing an unbiased (with or without cost sensitivity) comparison between different classifiers; De Bock and Van den Poel (2011) presenting a methodology that considers a rota- tive evaluation of performance measures; and Aman et al. (2015) proposing a set of measures that allows comparing models in terms of independence, reliability, volatility, and cost. Later, Clemente-Ćıscar et al. (2014) proposed two measures, one based on benefits and another based on returns, with the objective of evaluating the performance of a customer retention campaign. In this paper, we elaborate upon the MP framework as introduced by Verbeke et al. (2012) for supervised binary classification problems. The MP measure Verbeke et al. (2012), which is the first measure in the MP frame- work, considers the different costs of classification and at the same time facil- itates the obtaining of the optimal cut-off value to be applied when operating the obtained classification model, which is a practical advantage when com- pared to alternative measures. Verbraken et al. (2013) developed a stochastic 4 version of the MP measure, called the EMP, which models each cost using a probability distribution, allowing the estimation of the expected value of the maximum profit. The MP and EMP measures have been implemented and adopted successfully in churn prediction (Verbraken et al., 2014b) and credit scoring (Verbraken et al., 2014a). Both the MP and EMP measures are discussed in more detail in the next section. 2.1 Profit-based Evaluation Measures The MP measure is designed as a profit-based function, in which the param- eters b0 and c0 (b1 and c1) are, respectively, the benefit and cost associated with correctly and incorrectly classifying a good (bad) case, F0(t) and F1(t) denote the cumulative fraction of, respectively, goods and bads, with a score assigned by the classifier below the variable cut-off t. The average classi- fication profit per case resulting from adopting a threshold t is defined as follows: P(t; b0,c0,b1,c1) = b0π0F0(t)+b1π1(1−F1(t))−c0π0(1−F0(t))−c1π1F1(t) (1) Since all parameters in this function, i.e., b0, b1, c0, and c1, are assumed to be positive, then it follows that the theoretical overall maximum profit can be attained when F0(t) = 1 and F1(t) = 0. This, however, only occurs when a classifier perfectly discriminates between goods and bads. The max- imum profit that can be obtained for a non-perfect classifier is defined in Equation (2), where T is the optimal cut-off value that defines the threshold score separating the two classes. MP = max ∀t P(t; b0,c0,b1,c1) = P(T ; b0,c0,b1,c1) (2) The value of T can be obtained under the maximization of the profit function and satisfies the first-order condition for the maximization of the average profit, P : f0(t) f1(t) = π1(b1 + c1) π0(b0 + c0) = π1θ π0 (3) with π0 and π1 being the prior class probabilities and θ = b1+c1 b0+c0 being the cost-benefit ratio. Hence, the optimal threshold T depends on the cost- benefit ratio θ. The MP measure has the merit of being oriented toward the 5 central business objective, i.e., profit maximization, and also the practical benefit of providing the optimal cut-off value. More recently, the EMP measure has been proposed as an extension of the MP (Verbraken et al., 2013). This measure was designed considering that in real application settings, it is often difficult to estimate accurate values for benefit and cost parameters or costs and benefits may be case dependent; therefore, these parameters were modeled using a probability distribution. The EMP measure is presented in Equation (4) and accounts for the involved uncertainty, with ω(b0,c0,b1,c1) being the conjoint probability distribution of the cost and benefit parameters. For each possible combination of the cost and benefit parameters (b0,c0,b1,c1), the optimal threshold T is determined using Equation (3) as a function of the cost-benefit ratio θ. EMP = ∫ b0 ∫ c0 ∫ b1 ∫ c1 P(T(θ); b0,c0,b1,c1) ·ω(b0,c0,b1,c1)db0dc0db1dc1 (4) 3 The R-EMP measure In this article, we extend the EMP measure by acknowledging that in addition to the uncertainty regarding the cost and benefit parameters, as captured by ω(b0,c0,b1,c1) in the EMP measure, these parameters can change because of a random shock. Such a random shock can either be an external or internal event, or, despite the name, a steady evolution of the operational setting in which the classification model functions. For instance, changes in economic conditions or technological evolutions can have an impact on the operational setting and are examples of external shocks that affect profitability. On the other hand, changes in customer behavior, customer relationship manage- ment, business strategies and business processes are examples of internal shocks affecting profitability. Therefore, the presented R-EMP measure ex- tends the EMP approach, which models the benefit and cost parameters using a probability distribution to capture either uncertainty in estimating the exact values or to account for inherent variability across cases, by su- perimposing a perturbation or uncertainty on top of these distributions to capture the effect of such random shocks on profitability. As such, we aim to achieve a more robust measure and, through the use of this measure, as will be explained in a later section, to obtain more robust classification models for improved decision-making under variable conditions. Thus, we include the 6 impact of external information, given by the random shock, and the potential correlation between the components of profit, as the random shock can affect each measure separately or the same shock can affect multiple parameters at once. For example, both the benefits and the costs of an application can be affected by inflation, which is both external to any intrinsic uncertainty and the same for all measures, thus creating correlation between the costs and benefits. If Equation (4) is considered to give an estimation of the maximum profit but there is an external event η (that is out of our control) affecting this value, then we can extend the definition in Equation (4) to incorporate such a random shock. This leads to the definition of the extended, more robust R-EMP measure. For this purpose, the benefit and cost parameters are defined by a probability function and, in addition, by a random shock. As explained, these random shocks correspond to perturbations of benefits and costs. The extended expressions for the benefit and cost parameters are given in Equations (5), (6), (7) and (8). b′0 = f (b0,ηb0 ) (5) c′0 = f (c0,ηc0 ) (6) b′1 = f (b1,ηb1 ) (7) c′1 = f (c1,ηc1 ) (8) Then, b′0 represents the stochastic benefit of correctly classifying a case of class 0, which is a function of the original stochastic variable b0, as adopted in the EMP measure, and additionally of a random variable (ηb0 ), representing an external random shock. b′1 is the benefit of correctly classifying a case of class 1 and is defined as a function of the stochastic variables b1 and ηb1 . The cost variables c′0 and c ′ 1 define the cost of classifying a case as class 0 that belongs to class 1 or vice versa, respectively; these variables have been defined in a manner similar to that of the benefit parameters that include a stochastic random shock. If ω represents the joint distribution function of these parameters, then the R-EMP measure is defined by the following equation: 7 R−EMP = ∫ b′0 ∫ c′0 ∫ b′1 ∫ c′1 P(T(θ′); b′0,c ′ 0,b ′ 1,c ′ 1) ·ω(b ′ 0,c ′ 0,b ′ 1,c ′ 1)db ′ 0dc ′ 0db ′ 1dc ′ 1 (9) Note that the definitions of the random shocks, ηj, allow the specification and inclusion of highly complex probability distributions for the cost and benefit parameters, incorporating both stochastic effects that are intrinsic to the operation and random shocks that are external to the user. For example, by defining the distribution d of each component ηj as d(ηj) = d(νj,ε), with νj being an internal random shock affecting only one parameter and ε being an external random shock affecting every parameter j ∈ {b′0,b′1,c′0,c′1}, then each parameter will depend on internal and external stochastic effects. 3.1 R-EMP for Credit Scoring The R-EMP measure proposed in the previous section is a generic profit measure that can be adapted towards application in any business setting that requires accounting for stochastic costs and benefits, which may be subject to shocks. In this section, we define the functional form of the R-EMP measure for credit scoring. In credit risk management, one critical decision involves whether or not to grant a loan to a consumer. Credit scorecards, especially application scorecards, are classification models that are typically developed for making this decision in a data-driven manner (Thomas et al., 2002; Siddiqi, 2016). By defining the cost and benefit parameters and establishing the involved profit formula, Verbraken et al. (2014a) adapted the EMP measure as defined in Equation (4) to evaluate credit scorecards: EMPCS = ∫ 1 0 P(T(θ); λ,ROI) ·h(λ)dλ (10) with P(t,λ,ROI) = λ ·π0F0(t) −ROI ·π1F1(t) (11) The EMP measure for credit scoring defined in Equation (10) involves a single uncertain parameter, i.e., the loss fraction of a loan, λ, which is the benefit of correctly classifying a bad applicant (i.e., b1). This fraction is defined in Equation (12) below as the amount owed in the case of default, 8 i.e., the exposure at default (EAD), multiplied by the loss after all collection measures have been exhausted, i.e., the loss given default (LGD), and divided by the original amount of the loan (A). b1 = λ = LGD · EAD A (12) Additionally, when an application scorecard wrongly classifies a good ap- plicant as a bad applicant, an opportunity cost is incurred equal to the total return over the investment (ROI). This cost is considered relative to the amount of the requested loan (A) and will be assumed to be constant across cases (Verbraken et al., 2014a). In defining the R-EMP measure for credit scoring, we adopt the same approach as for the EMP measure, except that the λ variable is replaced in the same way that b1 is replaced in Equation (7), i.e., instead of λ, we now consider λ′, which is expressed by λ′ = f (λ,ηλ). This function adds a random shock ηλ to the stochastic loss fraction λ defined in Equation (12), which impacts the potential losses. The distribution of the loss fractions can be impacted, for instance, by changing the economic conditions and collateral prices. The R-EMP measure for credit scoring is defined as follows: R−EMPCS = ∫ 1 0 (f (λ,ηλ) ·π0F0(t) −ROI ·π1F1(t)) ·h(λ′)dλ′ (13) In the following sections, we will extensively illustrate the use of this measure, using both a synthetic and a real credit scoring dataset. 4 Experimental Settings An evaluation measure has practical use when it assists in decision-making during model development and operation. Common decisions that have to be made are, for example, choosing model parameters to maximize the classifi- cation performance, choosing the best attributes to include in a classification model, or choosing the cut-off point that is to be used for making a binary decision (e.g., to accept or reject a loan application) based upon the con- tinuous score that is produced by the classification model. In the following sections, we will illustrate how the R-EMP measure supports and improves such decision-making in a business context. 9 For this purpose, we conducted experiments on a synthetic and on a real credit scoring dataset. The experiments on the synthetic dataset focus on pinpointing differences between the R-EMP and other commonly used eval- uation measures, such as the AUC, H-measure and EMP, while the empirical case study focuses on the use of the R-EMP measure for practical decision- making. In the empirical case, given that we have data spanning 12 years, we will show both a comparison of measures overall with an out-of-time benchmark and a year-by-year comparison. This allows observing the behavior of these measures both over short and long terms and to assess their robustness, which was the main objective in developing the R-EMP measure. 5 Synthetic Case For the experiments reported in this section, a synthetic dataset has been cre- ated to compare the characteristics of a selection of measures often adopted to evaluate credit scorecards. The goal of this experiment is to compare how the applied measures behave when we subject the profit to a higher level of uncertainty. For this, we built a synthetic dataset, with a binary target variable (default and non-default) following a binomial distribution with size n = 1000 and probability of success p = 0.5. We created 12 attributes originating from six distributions (binomial, exponential, normal, Poisson, uniform and Weibull). Each attribute had different parameters per class; thus, there was a slight overlap (15%) between the distributions for each class. This process resulted in attributes that do not allow for linearly separable classes but do allow the achievement of a very high predictive accuracy. To construct the costs and benefits, we first created the loan amount A for each of the n cases, and then the EAD and the LGD were set based on this value. Benefits were defined as b = LGD ·EAD and costs as c = ROI · A, meaning that there are benefits when defaulters are identified correctly because we are avoiding the loss of b and that there are losses when rejecting good applicants because we are not earning c. Following this, we calculated the value of λ using Equation (12). The dataset was then randomly divided into a training and a test set. To introduce a shock, we perturbed the value of λ following Equation (14). 10 λ′ = λ + N(µ = 0,σ2 = 0.2 ·λ2) (14) We want to study the performance of different measures when more noise is added to the dataset to study the behavior of different performance mea- sures in this situation. Because each variable is simulated with an overlap, as more variables become available, the uncertainty (noise) in the model will increase. We selected random forest as our underlying classifier to extract as much information as possible from the variables, to filter random noise that can be easily eliminated by a model, and to focus on the impact of the uncertainty that cannot be filtered and the impact of the profit subject to a random shock. The simulation starts with no variables, and in each iteration, we include in the model the attribute that maximizes each evaluation measure. Once the procedure converges, i.e., when no further improvements can be achieved by adding more attributes, the resulting profit is calculated for the test sample. This process is repeated 100 times, generating new attributes without varying the underlying distributions. After all profits are calculated, the maximum profit achieved across all 100 iterations is calculated and used to normalize the results. Hence, the experiment simulates different samples of the same population with the same statistical structure but also with different external shocks to the profit structure. Figure 1 shows the profit in the test sample as a proportion of the max- imum profit for that sample. Here, the power of the profit-driven measures is demonstrated. The AUC shows a high variability, with profit proportions ranging from 0% to 70% and an average of just 42%. The H-measure per- forms slightly better, with less variability and a somewhat higher average profit proportion being achieved (54%), but yields results much lower than those achieved by the EMP and R-EMP. Both of these measures yield an average profit of 80% (EMP: 79.8%, R-EMP: 79.6%). It can also be seen that the R-EMP achieves a smaller overall standard deviation for most ex- periments but that there are outliers, which occur when the maximum profit is achieved. This effect is consistent with the design of the measure: the model will generally select the most robust measure (small deviation), but that measure will be a maximum profit measure. The robustness of the measure is shown in Table 1. In 57 out of 100 times, the R-EMP reaches the maximum value. The small increased standard deviation of the R-EMP occurs only because of the outlier value (without this 11 Figure 1: Synthetic Dataset - Profit Out-of-time by Measure ● ● ● ● 0 25 50 75 100 AUC EMP H−measure R−EMP Measures P ro fit % Measures AUC EMP H−measure R−EMP value, the R-EMP falls to 5.9%), and acknowledging some small difference, the means are basically the same. We can conclude that the R-EMP is equivalent to the EMP in that it presents better behavior when the noise in the sample is bigger; explicit information regarding this variability can be captured by adding a new exogenous variable. Table 1: Results of synthetic experiments Measure Average s.d. Times best measure R-EMP 79.6% 6.3% 57 EMP 79.8% 6.1% 43 AUC 42.2% 13.0% 0 H-Measure 53.8% 13.5% 0 6 Empirical Case The dataset, consisting of loans for small businesses, that was used to eval- uate the presented empirical case was provided by a Chilean financial in- stitution. The dataset contains 9 attributes, which after preprocessing and 12 one-hot encoding result in 16 predictive variables, as shown in Table 2. The attributes can be grouped into three types: • Loan variables: which describe the characteristics of the operation. Besides the amount and the term - in two forms, to account for different types of loans that might be either very short or very long term -, whether the borrower has collateral, or if the loan has a guarantor. • Sociodemographic variables: These variables describe the owner of the business. It is composed of the age of the borrower, a grouping of the zip code where the borrower operates in terms of default rates, and the ownership status of their main plot of land. For the last variable, the borrower can either rent, own, have free use of the land, part-own, or have other arrangements, resulting in five categories. • Business segment: These variables describe the business segment in which the borrower operates. The number of plots was segmented considering the default rates at each group, resulting in three categories: one, two or more than two plots. The second variable, the business segment, was divided following the same logic into four categories, each describing a macro-economic sector in the economy. Table 2: Variables used in empirical experiment. Variable Description Type Guarantor If the borrower had a guarantor for the loan Loan Collateral If the borrower had collateral on the loan Loan Term (months) Term of the loan (in months) Loan Term (year) Term of the loan (in years) Loan Age Age of the borrower Sociodemographic Housing Ownership status Sociodemographic ZipGroup Region of the country Sociodemographic Properties Number of plots of land Business EconSector Main economic sector Business The number of cases in the dataset is approximately 40, 000, with a de- fault rate of 24%. The dataset includes loan applications from 1996 to 2008, but the cases are not distributed equally across years: 35% of the cases stem from the first two years and the last four years, with 13% of the total cases corresponding to the most recent year. The last four years are selected as 13 the out-of-time sample, and the cases occurring between 1996 and 2004 are randomly divided into a training and a test sample in a 70% versus 30% proportion, respectively. For more details regarding this dataset, one may refer to Bravo et al. (2013). 6.1 Using R-EMP as a Profit Measure We first repeat the experimental setup applied in Section 5 for the synthetic dataset, allowing us to study how the R-EMP measure behaves when shocks affect a dataset with more complex data structures. As before, we want to choose the best model as more information becomes available, following a forward selection-like procedure. We stop adding information once there is no improvement in the measure and record the profit over the test set. As in the synthetic case, we assume a normal distribution to model the random shock (ηλ), and the number of replications (ne) is again set to 100. From Figure 2, we can see that the AUC and the H-measure generally lead to a similar number of attributes being selected; the EMP selects a smaller number of attributes; and the R-EMP selects the highest number. The R- EMP, AUC and H-measure select very similar average numbers of attributes, with the EMP being slightly off this mark. Again, as in the previous section, the R-EMP selects the models more robustly, with only small deviations with respect to the mean number of attributes, instead of either less or more attributes than average typically being selected, as with the other measures. Using the test sample, which covers the same time period as the training sample, we obtain the results shown in Table 3. The main insight from this table is that the R-EMP outperforms the other measures in all years except for 1999. This result again exemplifies the goal of the R-EMP, i.e., to provide a measure that is resistant to random shocks and variability. The selected model exhibits a stable, high performance over the selected period of time, as opposed to exhibiting such performance only in average years, as observed for the EMP and the H-measure. Additionally, note that in 2002, which involved a severe economic downturn, the model that was developed using the R-EMP measure is the only model that yielded a positive profit. From 2002 onward, the R-EMP model clearly outperforms the other models. Upon calculating the total average profit, the R-EMP is found to achieve the highest profit, followed at some distance by the EMP, H-measure, and AUC. Finally, the total average standard deviation is calculated over the full set of experimental results. As expected, the R-EMP has the lowest standard 14 Figure 2: Empirical Dataset - Number of Selected Attributes by Measure ● ● ● ●● ● ● ● ● ●●● ● ● ● ●● ● ●● ● ● ●●●●● ●● ● ●● ● ● ●●●● ● ●●●●●●●●●●● ● ●●● ● ●●● ● ●● ●● ● ● ●●●● ● ●● ● ●●●●●● ● ● ●●●● ●● ● ●●● ● ●● ● ●●● ● ● ●● ● ● ● ●● ●●●●●● ●● ● ● ● ● ● ●●●● ●● ● ● ● ● ●●● ● ●●●●● ● ● ●● ●● ●● ● ●● ● ● ● ●●● ● ● ● ● ● ●●● ●● ● ●●● ●●●●●●●● ● ● ● ● ● ●●●● ●●● ●● ● ● ●●●●● ● ●● ●● ● ● ● ● ● ●●● ●●● ●●●● ● ●● ● ●● ● ● ●●● ●● ● ●● ● ● ●● ●● ● ●● ● ● ● ● ●●● ●●●●● ● ●●●●●●●● ● ● ● ● ● ● ●●● ● ●●●● ●● ● ●●●●●● ● ●●● ● ●●● ● ● ●●● ● ● ● ● ● ● ●●●● ● ●●● ●●●●●● ●● ● ●●●●● ● ●● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ●● ● ●●● ●● ● ●●●● ● ●●●●●●● ● ●● ● ● ●● ●●●● ● ● ●●● ● ● ● ● ●●● ●● ● ● ●● ● ● ●●●● ● ● ● ● ●●●●●● ●●● ● ● ● ● ● ●●●● ●●● ● ●●● ● ● ● ● ●●●● ● ● ●● ● ● ●●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ●●●● ●●● ● ● ●● ●● ● ●●●●● ●● ●●●●●●● ●● ●● ●●●●● ● ● ●● ● ●● ●● ● ● ●● ● ● ●● ●● ●● ●● ● ● ● ● ●● ●● ●●● ●●●● ●● ●● ●●● ●● ● ● ●●● ●● ●●● ● ● ● ●●● ●● ●●● ● ●● ●● ● ● ● ●● ● ● ●●●●● ● ● ● ● ●●●● ●●● ● ● ● ●● ● ●●● ●● ● ●● ●●● ● ●● ● ● ●● ●●●● ● ●● ● ●● ● ● ● ●● ●● ● ●● ●● ●● ● ● ● ● ● ● ●● ●●●●● ● ● ● ●●●● ● ● ●● ●● ● ●● ●● ● ● ●● ● ●●●● ●●●● ● ● ●● ● ● ●● ● ●●●● ● ● ● ● ● ●● ●● ●●● ●●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ●●●● ●●● ● ●● ● ●●● ●●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●●●●● ●● ● ● ● ●● ●● ●●●●●●●●●●●●●●●●●● 7.5 10.0 12.5 15.0 AUC EMP H−measure R−EMP Measures N u m b e r o f se le ct e d a tt ri b u te s Measures AUC EMP H−measure R−EMP deviation with respect to profit. A final comparison between the models developed using the different mea- sures considers their predictive ability for an out-of-time test sample. The results are given in Figure 3. According to this figure, it is possible to ob- serve that the density of profit obtained using the EMP and R-EMP is more concentrated than that obtained using the AUC and H-measure; also, the EMP and R-EMP yield less dispersion, again hinting at the robustness of the proposed measure. The average profit for the out-of-time test sample is only slightly different across the models. According to Table 4, the R-EMP yields the highest profit (51,930 EUR), with the EMP in second place, closely followed by the AUC and the H-measure. Both Table 4 and Figure 3 indicate that even though the difference in average profit is small, the use of the R-EMP does consistently yield a higher profit model, which, importantly, is either less disperse or more robust. This outcome is exactly what the measure was designed to accomplish. 6.2 R-EMP as a decision-making tool This section is devoted to showing how the R-EMP can be used as a decision- making tool. Therefore, two additional experiments are conducted: the first 15 Table 3: Empirical Dataset - Average Profit Year-by-year in EUR by Measure Year AUC EMP R-EMP H-measure 1996 17,273 18,010 18,373 18,277 1997 6,391 5,043 7,877 6,511 1998 41,981 47,271 52,202 42,487 1999 121,923 121,625 109,097 123,311 2000 143,919 143,667 145,059 144,901 2001 103,714 103,375 104,923 103,987 2002 -3,683 -2,686 1,776 -2,856 2003 59,515 68,550 78,670 59,893 2004 41,235 50,974 59,539 40,548 Total average 532,269 555,827 577,515 537,059 Total standard deviation 217,210 216,181 209,535 218,179 Table 4: Empirical Dataset - Profit Out-of-time ± Standard Deviation in EUR by Measure AUC EMP R-EMP H-measure 51,270 ± 13,138 51,732 ± 10,162 51,930 ± 9,057 50,665 ± 13,936 experiment concerns parameter tuning, while the second experiment concerns determining a cut-off value. For this purpose, and for illustrating the use of the developed profit driven evaluation measure for decision making, an artificial neural network is trained. More specifically, since often adopted in a business analytics con- text (Verbeke et al., 2012), a multilayer perceptron (MLP) with one hidden layer is trained, also given the importance of tuning the characteristics of an MLP, for which various evaluation measures can be adopted. As a result, we obtain an indication of the potential gain from a profit perspective that may be achieved when consistently adopting the proposed measure during development of a classification model, in this case, a credit risk model. Note that the evaluation measure can be adopted in combination with any supervised learning approach, including alternative neural networks and deep learning approaches (Schmidhuber, 2015). A broad benchmarking study to compare various supervised learning techniques and performance evaluation measures, is beyond the scope of this paper, but is considered an important topic for further research. 16 Figure 3: Empirical Dataset - Profit Out-of-time in EUR (thousands) by Measure ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 20 40 60 80 AUC EMP H−measure R−EMP Measures P ro fit Measures AUC EMP H−measure R−EMP An MLP requires a large number of hyperparameters to be tuned to func- tion optimally. One typical method for setting parameter values is using a grid-search over various combinations of parameter values, thus limiting the infinite search space. The parameters tuned in this experiment are the num- ber of units in the hidden layer of the network and the maximum number of iterations of the algorithm. By using a training sample that includes data from the real dataset (described in the previous section) for the years from 1996 to 2004 and an out-of-time test sample that includes data for the years from 2005 to 2008 to evaluate the performance, we select the best combina- tion of parameters using the Accuracy, AUC and R-EMP measures. Given the strong correlation between the AUC and H-measure that was observed in the previous sections, we have omitted the H-measure from this experiment and instead included the Accuracy. The same reasoning was used for the EMP and R-EMP, focusing, of course, on the latter. Based on the number of attributes, the size of the hidden layer is tested from 8 neurons to 32 neurons in steps of 1, and the number of iterations is set from 50 iterations to 1000 iterations in steps of 50. In Table 5, the optimal values of the parameters using the different mea- sures are shown. The table shows the values for the iterations and hidden 17 layer size, the value of the performance measure (PM) at which that param- eter combination occurred, and, for the R-EMP, the optimal fraction (cut-off value) suggested for the model. Table 5: Parameter selection decision driven by different measures Performance Measure Iterations Hidden layer size Value of PM Optimal fraction Accuracy 800 17 0.6945 N/A AUC 50 20 0.7364 N/A R-EMP 50 30 0.0127 6.37% To operate a credit scorecard in practice to decide whether to accept or reject loan applications, a cut-off value needs to be adopted. Setting a cut-off value also allows a straightforward comparison of the performances of the various models in terms of profitability. The R-EMP measure implies a cut-off to be used, reported as the optimal fraction in Table 5. However, for the Accuracy and AUC, we need to choose the cut-off, which for Accuracy is selected based on the score of the test sample for which the accuracy was maximal, while for the AUC, we select the score for which the tangent of the ROC curve is equal to the proportion between average costs for acceptance and rejection, following Hand (2009). In Table 6, the behavior of the models based on the selected cut-off value and parameters is reported. The R-EMP, AUC and Accuracy measures are considered as alternatives, and a baseline scenario, in which no model is used to make a decision, i.e., loans are always granted, is reported as a reference. Under the baseline scenario, 4,566 loans are granted, leading to a total negative profit of -382,197 EUR. When using the Accuracy-based model and selecting the optimal cut-off for the validation sample, there is an improvement in terms of profit, resulting in a positive number of 11,567 EUR. AUC-based decision-making leads to an improvement in the total profit of up to 22,609 EUR. Note that the test accuracy is considerably lower when using the AUC and that the number of rejected loans is relatively larger. Finally, when adopting the R-EMP-based model, we further improve upon the AUC-based model profit, yielding 45,028 EUR and significantly reducing the number of rejected loans. As reported in Table 6, the R-EMP achieves both the highest accuracy and profit for the test sample. The accuracy of the Accuracy-based model is very similar to the accuracy of the model using the R-EMP, with the marginal gain in accuracy when using the R-EMP probably due to the robustness of 18 Table 6: Cut-off value decision driven by different measures Model Cut-off Test accuracy Total profit (EUR) Profit/loan (EUR) Number of granted loans No model N/A 80.20% -382,197 -83.70 4,566 Accuracy-based 0.78 57.60% 11,567 2.74 4,220 AUC-based 0.65 47.82% 22,609 6.89 3,283 R-EMP-based 0.75 58.18% 45,028 10.53 4,275 the measure, as it considers distributions over the sample as opposed to only the averages. The result in terms of the achieved profit is expected, as the R-EMP measure is designed to maximize profit over populations using distributions over samples, whereas the other measures are not. These results show that the R-EMP can be used with confidence to make decisions during model development and for model selection, supporting the method as a decision-making tool. 7 Conclusions In business environments, it is imperative to strive for optimal and robust decision-making, taking into account risks and evolving conditions. This ar- ticle contributes to achieving optimal and robust decision-making by propos- ing a novel performance metric for improving the decision-making process in developing classification models, i.e., by designing a variation of the EMP measure for evaluating classification performance when random shocks may affect the distribution of the profit parameters. The novel measure, dubbed the R-EMP, is experimentally evaluated using both a synthetic and real credit scoring dataset to assess its appropriateness and to compare its characteris- tics with those of the EMP measure as well as the AUC and H-measure. The results of the experiments indicate that the R-EMP effectively out- performs the EMP, as well as the AUC and H-measure, when there are ex- ternal factors affecting the profit variables, thus demonstrating that taking into consideration the impact of random shocks improves the quality of deci- sions in terms of profit. Additionally, the experiments provide evidence that the use of the EMP and R-EMP effectively leads to the selection of different models and yields better performance in terms of achieved profits. More- over, using the R-EMP for decision-making results in reduced variability and therefore improved robustness. Moreover, the presented results show that the novel measure can be reliably used to select the best model and to define the cut-off point for future use. 19 The experiments on a real dataset confirm that the use of the R-EMP results in a more robust model than the use of the EMP, which leads to the conclusion that by adding perturbations to the profit variables, the EMP measure can effectively be improved. For credit scoring applications, the R- EMP measure leads to better decisions than the main measures reported in the literature. Hence, the R-EMP appears to be a robust measure for building predictive analytics models within highly variable business environments. Acknowledgments We acknowledge the support of Conicyt Fondecyt Initiation Into Research 11140264. References References Ali, S., Smith, K. A., 2006. On learning algorithm selection for classification. Applied Soft Computing 6 (2), 119–138. Aman, S., Simmhan, Y., Prasanna, V. K., 2015. Holistic measures for eval- uating prediction models in smart grids. Transactions on Knowledge and Data Engineering, IEEE 27 (2), 475–488. Baesens, B., Mues, C., Martens, D., Vanthienen, J., 2009. 50 years of data mining and OR: upcoming trends and challenges. Journal of the Opera- tional Research Society 60 (1), S16–S23. Baldi, P., Brunak, S., Chauvin, Y., Andersen, C. A., Nielsen, H., 2000. Assessing the accuracy of prediction algorithms for classification: an overview. Bioinformatics 16 (5), 412–424. Bradley, A. P., 1997. The use of the area under the ROC curve in the eval- uation of machine learning algorithms. Pattern Recognition 30 (7), 1145– 1159. 20 Bravo, C., Maldonado, S., Weber, R., 2013. Granting and managing loans for micro-entrepreneurs: New developments and practical experiences. Eu- ropean Journal of Operational Research 227 (2), 358 – 366. Brown, C. D., Davis, H. T., 2006. Receiver operating characteristics curves and related decision measures: A tutorial. Chemometrics and Intelligent Laboratory Systems 80 (1), 24–38. Clemente-Ćıscar, M., San Mat́ıas, S., Giner-Bosch, V., 2014. A methodology based on profitability criteria for defining the partial defection of customers in non-contractual settings. European Journal of Operational Research 239 (1), 276–285. Correa Bahnsen, A., Aouada, D., Ottersten, B., 2014. Example-dependent cost-sensitive logistic regression for credit scoring. In: International Con- ference on Machine Learning and Applications. p. 7. Davenport, T. H., 2006. Competing on analytics. Harvard Business Re- view (84), 98–107. De Bock, K. W., Van den Poel, D., 2011. An empirical evaluation of rotation- based ensemble classifiers for customer churn prediction. Expert Systems with Applications 38 (10), 12293–12301. Fawcett, T., 2006a. An introduction to ROC analysis. Pattern Recognition Letters 27 (8), 861–874. Fawcett, T., 2006b. ROC graphs with instance-varying costs. Pattern Recog- nition Letters 27 (8), 882–891. Hand, D. J., 2009. Measuring classifier performance: a coherent alternative to the area under the roc curve. Machine Learning 77 (1), 103–123. McAfee, A., Brynjolfsson, E., 2012. Big data: the management revolution. Harvard Business Review (90), 60–6. McDonald, R. A., 2006. The mean subjective utility score, a novel metric for cost-sensitive classifier evaluation. Pattern Recognition Letters 27 (13), 1472–1477. Schmidhuber, J., 2015. Deep learning in neural networks: An overview. Neu- ral networks 61, 85–117. 21 Siddiqi, N., 2016. Intelligent Credit Scoring: Building and Implementing Better Credit Risk Scorecards. John Wiley & Sons. Thomas, L. C., Edelman, D. B., Crook, J. N., 2002. Credit Scoring and its Applications. SIAM. Verbeke, W., Dejaeger, K., Martens, D., Hur, J., Baesens, B., 2012. New insights into churn prediction in the telecommunication sector: A profit driven data mining approach. European Journal of Operational Research 218 (1), 211–229. Verbraken, T., Bravo, C., Weber, R., Baesens, B., 2014a. Development and application of consumer credit scoring models using profit-based classifica- tion measures. European Journal of Operational Research 238 (2), 505–513. Verbraken, T., Verbeke, W., Baesens, B., 2013. A novel profit maximizing metric for measuring classification performance of customer churn pre- diction models. Transactions on Knowledge and Data Engineering, IEEE 25 (5), 961–973. Verbraken, T., Verbeke, W., Baesens, B., 2014b. Profit optimizing customer churn prediction with bayesian network classifiers. Intelligent Data Anal- ysis 18 (1), 3–24. 22 Introduction Evaluation Measures for Classification Models Profit-based Evaluation Measures The R-EMP measure R-EMP for Credit Scoring Experimental Settings Synthetic Case Empirical Case Using R-EMP as a Profit Measure R-EMP as a decision-making tool Conclusions 
work_vrnc3jbhjvblxi64li5uowbbzi	----	1 Aided Diagnose of Structural Pathologies with an Expert System Ernest Bernat1, Lluís Gil Polytechnic University of Catalonia – BarcelonaTECH. Department of Structures and Material Resistance. ETSEIAT – Terrassa (Spain) Abstract Sustainability and safety are social demands for long-life buildings. Suitable inspection and maintenance tasks on structural elements are needed for keeping buildings safely in service. Any malfunction that causes structural damage could be called pathology by analogy between structural engineering and medicine. Even the easiest evaluation tasks require expensive training periods that may be shortened with a suitable tool. This work presents an expert system (called Doctor House or DH) for diagnosing pathologies of structural elements in buildings. DH differs from other expert systems when it deals with uncertainty in a far easier but still useful way and it is capable of aiding during the initial survey ‘in situ’, when damage should be detected at a glance. DH is a powerful tool that represents complex knowledge gathered from bibliography and experts. Knowledge codification and uncertainty treatment are the main achievements presented. Finally, DH was tested and validated during real surveys. Keywords: Expert System; Buildings; Maintenance; Inspection; Diagnosis; Assessment. 1. Introduction During the past decade construction has been a driving force of the economies of many countries. This surge has created an increasing volume of buildings that will have to be 1Corresponding author: ernest.bernat@upc.edu. Colom 11 TR45 D.137, Terrassa 08222 Spain. Telephone: +034937398728. Fax:+034937398994 mailto:ernest.bernat@upc.edu maintained in the near future. Moreover, nowadays the industry is slowly changing from classical construction towards sustainable building. Therefore, maintenance and renovation are processes that enlarge the life-cycle of buildings. In short, most of the recently constructed building will have to be surveyed and/or repaired during the following years. Taken together, these factors point to an explosion in the demand for qualified personnel for diagnosing pathologies of buildings, infrastructure and other forms of construction—of which there will be a greater number in use and to be conserved. This backdrop is the motivation for the development of a program to provide support to the new inspectors who will have to diagnose the pathologies of this immense volume of buildings. This type of program is not intended to substitute technicians; rather, its main objective is to assist in the training of new experts and provide them with greater confidence in the most complex diagnoses. Furthermore, it helps free these highly qualified workers from mundane tasks, enabling them to focus on the most complex cases. Described here are the steps followed in the development of a first operating prototype of the software. It was focused on buildings for several reasons, including their great economic weight, user demands and the large volume of buildings already constructed. Another reason was to obtain a more useful tool: buildings generally have a lower inspection level than large public works, therefore they were considered to suffer more damage. Knowledge codification, which includes obtaining the information, ordering it and codifying it in form of objects and logical rules, was the most challenging part of the investigation together with the necessary uncertainty treatment. How these tasks have been carried out is fully detailed in their corresponding sections, but in general lines the first one was overcome with the support of a team of experts form private companies and university whose experience, which covered all structural elements of a building, was classified. The second one was faced by simplifying the uncertainty levels for not losing the qualitative approach and still offering a reliable and transparent tool. Classifying symptoms (observable information) into required ones and recommended ones was an original improvement. This paper is divided as follows. First, past and present artificial intelligence (AI) programs developed for the construction field are described. The objectives of developing DH are then outlined. Next, expert systems are briefly described and are evaluated in the context of other methods. The process employed for developing DH, and the results of initial trials, are then explained in detail. A brief qualitative comparison with similar tools is also included and, lastly, conclusions are presented. 2. Review on Artificial Intelligence in construction and Aim of the work The term artificial intelligence (AI) was first coined by John McCarthy at a conference in Dartmouth in 1956. Since then, AI has been used for a myriad of applications, including modelling and decision making in the construction field. One of the most important milestones in the practical use of intelligent reasoning was the development of expert systems, which are explained in more detail below. Briefly, an expert system can be defined as a computer program that is designed to emulate a human expert for solving problems in a specific area of knowledge as it is explained in Harmon and King (1985) or Giarratano and Riley (1994). The recent advent of artificial neural networks, particle swarm optimization method or genetic algorithms have provided a new practical stimulus for AI in the field of decision-making around engineering activities. Nowadays, AI includes a lot of different fields, like robotics, representation of natural language, human reasoning and others, but the most commonly used in construction are artificial neural networks, Case-Based Reasoning (CBR) and expert systems. However, recent investigations on genetic algorithms like Cheng et al. (2002) and particle swarm method presented in Zhang and Xing (2010) or Jiang et al. (2011) among many others, suggest them to be widely used in the development of the future most complex decision-making tools. Although construction was one of the last fields to use this technology, see Kaetzel and Clifton (1995), over time several tools and lines of research have since been developed. These include: • Support for decision making in project design and planning: e.g. CONBPS (Construction Best Practice System) for modelling of the construction process by Poon (2004); neural networks to calculate the shear bearing capacity of FRP – fibre reinforced polymer – strengthened reinforced concrete beams by Tanarslan et al. (2012); or for the assessment of construction environmental impact in Zhao et al. (2006). • Support for decision making in the building process: e.g. OLSC (On-Line System for Construction) for multi-criteria analysis presented by Kaklauskas (2007); CoSPES (Construction Site Preparation Expert System) for construction tasks written by Albert and Wu (1997); CONBPS by Poon (2004); for construction project supplies there are many example like Kumaraswamy and Dissanayaka (2001); for determining the rippability of rocks in Tiryaki (2007); for adapting tunnel construction in base of its displacements using Support Vector Machine (SVM) together with a particle swarm as shown in Jiang et al. (2011); or a fuzzy multi-objective particle swarm to optimise time- cost-quality in construction presented in Zhang and Xing (2010). • Support for the diagnosis and repair of construction pathologies: e.g. MDDS (Masonry Damage Diagnostic System) for diagnosing pollution caused pathologies in masonry projects proposed by Van Balen (1996 and 2001); RCDES (Reinforced Concrete Diagnosis Expert System) for diagnosing reinforced concrete structures in Peter (1996); REPCON (Repair of Concrete) for repairing concrete structures written by Moodi and Knapton (2003); concrete bridge management by Brito et al. (1997); DIASYN (Diagnosis Synthesis) for concrete bridges used in Zhao and Chen (2002); maintenance of cement bridge decks as in Yehia et al. (2008); for evaluating bridge damage in Lee et al. (1999); for diagnosing cracks in reinforced concrete in Chao and Cheng (1998) and Yeong et al. (2007); and for diagnosing corrosion of pipes included in Najjaran et al. (2006); These have been applied to buildings: e.g. for evaluating damage to buildings from cyclones in Murlidharan et al. (1997); for diagnosing framed reinforced concrete structures in Lu and Simmonds (1997); for diagnosing building defects in Koppány (2005); SCAMS (Subsidence Case Management System) for evaluation of pathologies due to soil settlement presented in Anumba and Scout (2001); and for diagnosing seismic damage in metallic structures like in González and Zapico (2007). Many of the expert systems for construction that were developed up to the early 1990’s can be found in Kaetzel and Clifton (1995). A list of expert systems for construction project planning can be found in Yamazaki et al. (1990). Obviously, when the problem to be solved is very specific it can be handled in greater depth and the results will be more accurate. However, it was observed that there was no software developed for the global inspection of buildings, as all available programs focus on a specific material or structural type. So, having observed that there was no generic building- inspection software on the market, we sought to develop an operative prototype in order to prove the technical feasibility of such program. Our initial objectives comprised: a) Researching, summarising and normalising the symptoms associated with each building pathology; b) Developing a prototype of a knowledge-based expert system for diagnosing building pathologies, including ordering, encoding and programming of symptoms data according to the programming tool chosen; and c) Evaluating the resulting software, and then generating conclusions. 3. Used tool: Expert Systems Expert Systems (ES), which fall under AI, are computer programs that are designed to imitate human experts in solving problems from a very specific area. Modern expert systems bear little resemblance to their earliest predecessors, which arose in the 1960’s and which were originally designed to solve general problems (Harmon and King 1985). They have three fundamental components: an inference motor, a knowledge base and an interface. The inference motor is in charge of obtaining new data based on the existing data in the knowledge base and that provided by the user. It can control this data collection process according to various criteria, including forward reasoning and backward reasoning. The forward reasoning is based on performing a basic search to determine which data is needed for continuing with the reasoning. Albeit this leads to a disordered and chaotic appearance from the user’s perspective, it allows execution of planning tasks for which solutions are not known a priori. In contrast, the backward reasoning performs a deep search which is focused on an objective (a hypothesis which is to be verified) and determines the information that must be confirmed in order to validate the chosen line of reasoning. This method of inference is much more natural, logical and ordered from the user’s perspective, and is especially suited for cases in which there are a limited number of known solutions, as it is diagnosis of building pathologies. The knowledge base stores the information from the human expert that the program may use in searching for a solution. The knowledge base is encoded differently in each expert system, but the most common and powerful way is through objects, which are related to each other by production rules (i.e. through if/then logic rules, which are associated with the inference via modus ponens, see below). Finally, the interface is a crucial part of an expert system, as it must be adapted to allow easy entry and modification of information from the knowledge base and enable justification of the diagnosis. Moreover, it has to be user-friendly, since it is used for data entry by the user and for the delivery of results. For greater detail on the structure and function of expert systems, the reader is referred to Harmon and King (1985), Giarratano and Riley (1994) and Kaetzel and Clifton (1995). 3.1. Why an expert system? The decision to employ software in the form of an Expert System was based on the ease of programming of this method as compared to conventional programming. This selection was also done because ES show a more natural and predefined solving behaviour than artificial neural networks (ANN) whose drawbacks are shown below. Moreover, ES are far simpler and more organised than genetic algorithms (GA) or particle swarm optimisation methods (PSO) which could work together with support vector machines (SVM) as it was done in Jiang at al. (2011). The main advantages of an expert system over conventional programming comprise: • The possibility of completely separating the solution method for the problem from the data, which allows ready modification, correction or expansion of the knowledge base, as well as creation of shell-type tools, which already contain the inference motor and only require the knowledge base; • The use of symbolic programming, which allows incorporation of symbolic variables (i.e. words from natural language, such as “crack”, “oxidation”, etc.), between which logic-type relationships can be established (e.g. if/then production rules, or qualifiers such as and, or, equal to, different than, etc.). Just as other programs based on knowledge and experience, the work described here adjusts better to the programming patterns of expert systems than to those associated with classic programming, which is focused on the use of numbers and the procedural solution of problems. The main advantages of an expert system over ANN in diagnosis of building pathology comprise: • Expert systems receive information from a human expert (once that information has been properly ordered and encoded), whereas neural networks learn by examples. Therefore, for a field such as building pathology, in which there are so many possible combinations and symptoms, an extremely large number of cases would be required in order to train a neural network. • Neural networks automatically establish the relationships between variables from the examples provided. This means that solutions that are initially unknown to the expert can be found. In the field of diagnosis, this possibility is counter-productive, as the aim is to provide confidence in the identification of the cause of damage, rather than to “discover” pathologies based on atypical symptoms. Furthermore, the automatic encoding of knowledge complicates its subsequent modification since the information introduced does not remain structured, but rather is stored diffusely throughout the network. In fact, several authors have referred to neural networks as a “black box” whereby the relation between each input and its corresponding output remains obscure to the user as González and Zapico (2007) pointed out. For these reasons, neural networks can be considered more amenable to problems that are poorly structured or that feature non-explicit knowledge—the very opposite case to building damage diagnosis, in which the causes and effects are well known and can be easily correlated. Other actual tools, like SVM could overcome the extra-learning problem of ANN. They base their learning on statistics in order to solve multidimensional functions whose domain is previously delimited as explained in Jiang et al. (2011). However, its working procedure is “hidden” from user in the same way it was in ANN, so it is not suitable in developing a tool aimed to aid new professionals and to help them in their formation. Like ANN, GA and PSO are mainly oriented to solve problems from incomplete and bad-structured information. Cheng et al. (2002) used GA to deal with the problem of flood prevision which is a clear example of incomplete and bad-structured data. Modelling algal blooms in Muttil et al. (2006) or trying to foresee sunspots as shown in Xie et al. (2006) could be another two examples. As structural damage diagnosis at first visual inspection has a relatively little field of possible solutions and related data can be organised, GA and PSO are not as suitable as ES are to deal with this problem. 3.2. Expert Systems drawbacks Today, Expert Systems are not widely used. Lately, this programming tool has been progressively substituted by CBR (Case-Based Reasoning) firstly and ANN (Artificial Neural Network) recently, due to its difficulty to structure and validate the general expertise of most of the nowadays common problems. Nevertheless, ES are still useful in some fields, like diagnosis of building pathologies (or detection of the damage’s cause), in which knowledge is very well organised, and the development of an expert system is easier than in other fields. The use of an expert system can be seen as step back, but in our opinion, it can break the current trend, which is turning to numerical treatment of any problem. It loses touch with the real problem, and all is reduced to adjusting many factors in order to reach the expected solution whereas the logical reasoning process and the justification of the obtained solution are absolutely forgotten. As there is the aim to develop a tool that could be also useful at training new human inspectors for diagnosing structural damage, it seems suitable for this tool to be transparent, to use qualitative information rather than quantitative coefficients decided by a group of experts and the most important, to have the possibility of giving an explanation about how the conclusion was reached in basis of the information and the logical inferences. 4. Developing the prototype. Methodology Outlined below is the process followed to develop the prototype of the new tool for construction damage diagnosis, from the scope delimitation to the translation possibilities, including the knowledge codification and the uncertainty handling. 4.1. Focusing on a specific area Once the problem was centred on building pathology, it was decided that the level of casuistry should be minimised, since the objective was merely to create a prototype to prove the technical feasibility of this type of program. It was ultimately decided that the problem would focus on pathologies that affect structural elements or that directly derive from these elements. This choice was due to the fact that the principal repair costs in building pathology are those related to structural damage as pointed out in Calavera (2005) or by The Masonry Society Construction Practices Committee (2004). Different pathologies were not treated with the same level of depth; the south-European most common type of structure was emphasised: a framed concrete building with dividing masonry walls. Special attention was paid to the principal structural elements: pillars, beams and floor slabs. Altercations to masonry elements were also given special attention, as the number of these pathologies is much higher than other types, and they are generally caused by the malfunction of the structure of a building. In terms of pathology types, cracks are the most common form of damage among structures, as presented in Rodriguez (2000). As such, they were studied in more detail than other symptoms. Because of the aim of DH, the range of pathologies to be diagnosed by the expert system was limited by including only data that could be obtained through visual inspection of a building with the naked eye; complex assays were therefore omitted. Thus, the resulting software was honed for providing a preliminary qualitative inspection in a transparent way. 4.2. Choosing the programming tool A shell-type program—which already contains the inference motor and the interface—was chosen as the development platform for the prototype. This choice was based on various factors, including the knowledge (data) itself to be introduced as well as the scope of the work at hand (for a first prototype). The information collected on symptoms was generally well ordered. For this reason, it was preferable to use encoding based on objects and values. This way, the tool had to use inference through modus ponens and relate objects with production rules. Modus ponens is the most widely used inference method. It establishes if/then type of relationships. Hence, once a rule has been introduced (A ⇾ B), and the left-hand side factor (A) has been confirmed, the result provided by the inference is affirmation of the right-hand side factor (B). An object is the symbolic representation of a characteristic, fact or real element (e.g. “direction of the crack”). Each object can adopt different values (e.g. [vertical, horizontal or diagonal] for “direction of the crack”). Considering that the developed expert system is a diagnostics program aimed to train novel inspectors in on-field preliminary inspection, the inference process had to be ordered according to user needs, and consequently, the shell had to allow backward reasoning. The shell also had to be monotonic: the inference process could not be allowed to change the information introduced by the user, since the program would then be diagnosing a different damage from that observed. Furthermore, given that the developed software was designed to simplify the diagnostic process as well as the data required, the use of either numerical uncertainty or diffuse logic was ruled out as it is outlined at subsection 4.4. Commercial software Acquire 2.1 was selected from among the various software packages that were available at the time of development, as it responded to all of the programming needs for the work at hand and has features which are highly amenable to the creation of a prototype of a small expert system. 4.3. Encoding and introducing the information using Acquire 2.1 The first step in encoding the information comprised classifying the pathologies according to the structural element affected and the constituent material of that element. Many small and focused expert systems (modules) were created in this way, each one for dealing with the diagnosis of pathologies of one structural element or one material. Most of the information related to the symptoms of the building pathologies was extracted from documentation like Calavera (2005) and The Masonry Society Construction Practices Committee (2004), and from some experts interviews. In total, seven specialists were interview. Two engineers form Construction Engineering Department, two more from Structures and Material Resistance Department, all them from Polytechnic University of Catalonia – BarcelonaTech, plus two architects and another engineer from the company Asistencia Técnica Industrial S.A.E, ATISAE, who are expert inspectors. The aim of the interviews and meetings was to get a deeper description of selected pathologies. They were also oriented to improve the knowledge acquired from bibliography and to obtain a more confident response. As all pathologies included in DH are common and can be diagnosed at glance by a professional, the team of experts did not show any discrepancy about the diagnosis nor the causes of damages. This is one of the reasons, as it is widely shown later, why fuzzy logic and numeric evaluation of uncertainty were not considered in the development of this software whose purpose is to help inexperienced inspectors at diagnosing common structural pathologies of buildings. The way the knowledge was acquired is following described. First of all, the major part of it was extracted from bibliography. After this, an expert from Structures and Material Resistance Department was selected as the conductor of the interviews and the meetings. The first meeting was set to discuss the scope of the selected damages and the main variables that had to be included, distinguishing between the defining symptoms and the ones that do not need to be observed to come up with the diagnosis. Then, a preliminary version of the prototype was implemented and individual interviews were carried out with each expert in order to improve the result. As no contradictions between experts were observed, every opinion was taken into account. The information given by the experts in the individual interviews was complementary and never confronts other’s advice. After improving DH, a full meeting was arranged to present the result to the experts and discuss about the last possible improvements that mostly affect superficial things like the interface language or the distribution packaging. The knowledge base was consolidated after this meeting. Because of the scope and the aim of DH it was not considered to deal with uncertainty in an analytical way. Moreover, no discrepancies between experts’ recommendations were observed, so quantifying uncertainty in common pathologies would not worth the effort as the diagnosis were clear. For these two reasons it was decided that the selected team of experts was enough to provide the information needed. For further research including more complex pathologies, uncertainty would have to be taken into account in a numeric way because different opinions are expected in less common problems, so a bigger group of experts (statistically significant) would be required in order to value this uncertainty in the conventional quantitative approach. Once the pathologies had been classified, the next step was to encode the symptoms of each one by transforming the collected information into objects and rules, which are schematically represented in Figure 1. The objects created for encoding information in DH can be classified according to its function. First of all, there are starting objects (continuous striped rectangles in Figure 1) which are symbolic representations of the symptoms. When the expert system asks the user about a particular symptom, its answer will be reflected in the value taken by the corresponding starting object. Secondly, there are auxiliary objects (discontinuous horizontal striped rectangles in Figure 1) whose main function is to give internal structure to the program, thus providing a more logical and ordered presentation of the inference for the user. The use of these objects makes heredity possible, so at the moment at which the inference process assigns a value to an auxiliary object, this value implicitly represents all of the information provided up to that moment by the user over the course of a DH query. Finally, there are conclusion objects (spotted rectangle in Figure 1) which symbolically represent pathologies that can be diagnosed by the expert system. At the moment these objects adopt a favourable value (either “Yes” or “Yes, but”, as explained below) a diagnostic is reached. In terms of the relationships among objects, Acquire 2.1 can use production rules (Figure 2) and action tables. These two methods only allow a single value to be assigned to a unique object as a conclusion of the relationship. Production rules are best suited for cases in which many objects are considered. These objects should adopt very specific values in order to reach a conclusion. This is the case of the last step indicated by the dashed lines in Figure 1, where it can be observed that, for a given combination of values for some different objects, one value is obtained for the conclusion object. In contrast, action tables are indicated for cases in which very few objects are considered and in which nearly all of the combinations of values of these objects are valid for generating a conclusion in the inference. Roughly speaking, an action table can be considered as a set of production rules grouped together as a matrix. Action tables are used in practically all of the relationships that assign a value to an auxiliary object. Each step in Figure 1 that is marked with a solid line corresponds to an action table. A complete diagnostic tree, following the encoding process described, is shown in Figure 3. 4.4. How to handle uncertainty DH considers uncertainty, which is intrinsic to any type of heuristics-based reasoning. In general, uncertainty might be encountered in the inspection as well as in the logic rules employed for diagnosing the pathology. However, it has been shown previously that there was no uncertainty in the logic rules for DH as experts and bibliography did not show discrepancies in their diagnosis. It is because DH is aimed to help at diagnosing well-known cases to train new inspectors or free highly qualified workers from ordinary activities. Despite the fact that most of the current expert systems handle uncertainty through inexact reasoning and fuzzy logic –examples of this are Zhao and Chen (2002), Murlidharan (1997), Lu and Simmonds (1997), Chao and Cheng (1998), Yeong et al. (2007), Najjaran et al. (2006) or Anumba and Scout (2001) – the prototype presented here was designed to treat uncertainty qualitatively, which is simpler and more transparently for the user than the conventional quantitative approach. It allows DH to be used as a learning tool. Actually, this approach goes back to the empirical treatment used in the original expert systems. The expert system was thus conferred with a tool that enabled it to deal with uncertain information, although with no complex mathematics required, as these tend to overcomplicate the justification of the obtained solutions to the user and make them less transparent—when these explanations are in fact one of the main advantages of expert systems. This is because current numerical treatment is based on the use of coefficients that are chosen by experts according to their personal experience (in order to represent the scenario that certain symptoms correspond to a specific pathology) and these coefficients are usually hidden to users, who do not receive all the inference information. Instead, showing all the mathematically represented uncertainty information to the user is not the best solution either as the amount of data would make it difficult to understand the reasoning procedure anyway. This, together with the nature of the problem faced in this research lead us to develop a prototype with a different system that would be more general, simpler and more intuitive. As a drawback, it would not give any numerical value to describe the confidence on the solution. Therefore, the empirical treatment of uncertainty used here is not based on a scientific approach like probability. We assume the limitation of this approach and admit this technique could not be suitably extended to the development on an expert system that had to deal with more complex cases in which conflicting views between experts were likely to rise. In this other situation, fuzzy-logic would be likely to face the problems with more success than the original method proposed here. In conclusion, it could be said that the way the uncertainty is treated in DH is the best option for the specific faced problem and the particular aim of the program but it could not be the best for developing tools oriented at helping the highly-qualified professionals at taking the most difficult decisions. The goal for representing uncertainty in DH was to establish two levels of confidence in the solution. This was based on the fundamental hypothesis that there are two levels of symptoms in any pathology: Required ones and Recommended ones. Required symptoms are those which are most characteristic of a particular type of damage and which must be observed in order to diagnose it, whereas Recommended symptoms are those whose observation—though not required for diagnosis—allow to reach a greater confidence in the diagnosis. This treatment of uncertainty was encoded by assigning one of either two values (“Yes”, or “Yes, but”) to the conclusion object which confirms diagnosis of a pathology with different levels of confidence. If all of the observed symptoms correspond with the expected symptoms, then the conclusion object takes the value “Yes”, and consequently, the diagnosis is confirmed with maximum confidence. If any of the Recommended symptoms are missing, then the conclusion object takes the value “Yes, but”, and therefore, the diagnosis is made with less certainty. In this last case, the results show which of the Recommended symptoms are missing. In contrast, if one of the Required symptoms is not observed, then the pathology is not diagnosed. These three scenarios are summarised in Table 1. Scenario Result Required symptoms = expected values Recommended symptoms = expected values Diagnosis with maximum confidence (Value “Yes” in pathology object) Required symptoms = expected values Recommended symptoms ≠ expected values Diagnosis with less confidence (Value “Yes, but” in pathology object) Required symptoms ≠ expected values Undiagnosed pathology (Value “No” in pathology object) Table 1. Combinations of symptoms shown with their corresponding levels of diagnosis confidence It should be mentioned that all of the auxiliary objects are associated with Required symptoms, and that the Recommended symptoms are the last ones that the user is asked about; hence, they are incorporated in the last step of the inference, the production rule. Along the development process, it was stipulated that any symptom which exhibits a high degree of variability in the real world, or which is difficult to identify, can never constitute a Required symptom for the diagnosis. A clear example of this would be the width of cracks, which, although indicative of the magnitude of the damage, is rarely meaningful for determining the cause of the pathology. 4.5. User interface and translation possibilities In the developed prototype, the user interface was limited to two environments: data entry by the user (as requested by the program through a series of window pop-ups) and the presentation of results in the form of a report which includes either the diagnosed pathology or a course of action in the event that no diagnosis was obtained. Furthermore, the entire expert system’s interface can be easily translated into any language by incorporating additional objects and relationships that establish one-to-one relationships between the new inputs and the original ones, which consequently become auxiliary objects. 5. Running DH: How does it diagnose? Since the development of DH has been detailed, now here it is presented how it works and which its main features are through the analysis of a real case. However, the basic steps the program always takes should be explained first. Observing the flowchart in Figure 4, it is noticed that at the beginning of the diagnosis the user have to provide, at least, the information related with the structure typology and material and the kind of affectation. After this, the program chooses a possible diagnostic and tries to confirm its main symptoms (those called “Required” in the section 4). If some of these symptoms are not confirmed by the user, DH changes its assumption and tries to confirm another pathology that actually matches with the information already provided. If there are no pathologies in the knowledge base that agree with the observed main symptoms, DH cannot diagnose the affectation. Else if one possible diagnostic is confirmed by the observation of its main symptoms, the program follows to the next level and tries to be fully certain of the diagnostic that has been reached. At this point, more specific symptoms (previously called “Recommended”) are asked to be observed, and if the inspector-user confirms all them, DH achieves a fully certain diagnostic. If not, the diagnostic is anyway presented but with no fully certainty. In this case, an explanation with which the secondary expected symptoms were is also showed. For better understanding, a real diagnostic case is outlined below. Each step in decision- making progress is explained, and both the questions of the expert system and the answers from the user are indicated. There are also pictures and sketches of the analysed pathology obtained in the field inspection in cooperation with the company ATISAE. Step Question Answer 1st Structure typology and material Concrete Beam (*) 2nd Kind of pathology Crack (*) 3rd When does the problem become visible? In use (**) 4th Where are the cracks located in the beam? End (*) 5th How is the crack orientated? Inclined (*) 6th Is the beam supporting a stairs or a cantilever? Or is the beam placed in the edge of a floor? No 7th How many cracks do you see? A few (2 to 4) (*) 8th Has the beam been constructed in two times? If you select Yes, means that it's a beam constructed in two times or that it's a continuous beam constituted by two simple beams that have been united in situ. No (*) 9th Is the end of the beam connected with a rigid node? If you choose Yes, it is assumed that the analysed element is a span of a longer continuous beam. No 10th Where in the section are the cracks located? Around the upper head (*) 11th How wide is the crack? Variable (*) 12th How deep is the crack? Cross the beam (º) (*) See Figures 5 and 6 (**) It was reported by the building users (º) Assumption from the fact that the same cracks are observed in both sides of the beam (see Figures 5 and 6) Table 2. Summary of the running for a case diagnosed by DH The analysed building is a framed, reinforced concrete structure that comprises four floors plus a basement level. The damage was observed in a secondary girder on the 2nd floor. Figure 5 and Figure 6 are pictures of the crack to diagnose from either side. Once defined the problem, now the different steps DH took are summarized in Table 2. First of all, the user selected the structure typology and material (concrete beam) and the kind of pathology (a crack). From this point, DH automatically selected a suitable pathology (any among all possibilities that could explain a concrete beam cracking) and tried to confirm it. At each of the eight following questions, DH could have been changing its initial assumption to a pathology that definitely agreed with the defining symptoms introduced by the user. For example, the question in 6th step (Table 2) is oriented to confirm or discard on of the possible causes for inclined cracks at the end of a concrete beam: torsion. In the same way, questions in 8th and 9th steps tried to find out if the pathology was associated with the construction procedure or with an excessive negative momentum at a rigid node respectively. The next steps (11th to 12th) and questions DH made were only to add certainty to its decision of a shear failure. So, finally, DH reached a diagnostic and presented a report with the pathology and also added a warning due to the high danger level of this kind of pathology. A screenshot of the program running is presented in Figure 7. In conclusion, Doctor House diagnoses the case as shearing cracks due to excessive stress in the concrete. This result is totally consistent with analysis of samples extracted from the building, which revealed that the resistance of the concrete used was much lower than that recommended for this type of structure. The program also warns of sudden collapse, which is associated with said pathology. Lastly, the program's results were further supported by the conclusions of a team of experts in the field of building pathology, who concluded that the cracks were indeed due to excessive shearing force. 6. Results, Discussion and Comparison with similar Expert Systems Once the prototype of DH was finalised, its functions were evaluated in a series of tests. The initial round of test cases was taken from the literature. Especially, from Muñoz (1988) which summarises many cases on real pathologies diagnosed by human experts. In order to ensure objectivity, a different source of information was used for this evaluation than that used for the programming of the expert system. The results obtained (Table 3) are satisfactory and the correct pathology was diagnosed in most of the test cases. Only a few discrepancies were observed and they were the result of missing information in the test case example, which was limited to a photograph and an expert diagnosis description in the corresponding book. In three of the cases summarised in Table 3 the result was no fully satisfactory because it exceeded the knowledge limit of DH. It is that the pathologies to be diagnosed were out of the scope of the expert system. It is not such a big problem as DH is a prototype which was developed just to prove the technical feasibility of this kind of software. On the other hand, the cases for which the statement “Insufficient information of the test example” is displayed in Table 3 need further justification. For all these cases the problem was that there were validation cases from bibliography so access to the real structure was impossible and the available information was limited to a few pictures and a brief description of the problem that was not enough to provide the expert system required information to achieve a correct and reliable diagnostic. For example, the torsion of concrete beams was not diagnosed with full confidence because the beam position related to the stairs or cantilever elements was unknown. In the same way there was no information about the cracks in the walls of the lower floors to give plenty confidence on the diagnostic of differential settlement of a pillar. The lack of an accurate description of the cracks at the down side of the slab made it impossible to diagnose the punching cause without any doubt, and finally, there was no information about the construction procedure that might be the basis for the identification of an initial thermal reaction as the problem that caused the 10th pathology summarized in Table 3. All these problems arisen from the impossibility to have direct contact with the structure would not be found in real inspection like the one fully explained at the previous section because the access to the building visible part of the structure should be granted. DH Module Pathology Result Problem Reinforced concrete pillars Hydraulic retraction - Knowledge limit exceeded Excessive tensile stress + Concrete beams Torsion ~+ Insufficient information of the test example Bending + Masonry elements Differential settlement of a pillar ~+ Insufficient information of the test example Absence of joints + Slabs Punching ~+ Insufficient information of the test example Excessive bending of a beam ~- Short reinforced concrete cantilevers and nodes Horizontal tensile + Reinforced concrete windscreens and walls Initial thermal retraction ~ Insufficient information of the test example and Knowledge limit exceeded Facades Map cracking + Metallic elements Generalised corrosion ~+ Knowledge limit exceeded Wooden elements Wood worm + Table 3. Results of the first round of evaluation tests A second round was employed to validate DH with real cases (together with ATISAE Company). Results from all 5 tests on damages affecting concrete beams and masonry walls were successful and the diagnosis coincided with the one made by another unassisted inspector. One of these tests has been explained in section 5 as an example. The final result is a decision-making program that comprises ten modules that can diagnose pathologies associated with reinforced concrete pillars; concrete beams; frameworks; nodes in framed concrete structures; short reinforced concrete cantilevers; masonry elements; reinforced concrete windscreens and walls; facades; foundations; metallic elements; and wooden elements. However, it should be noted that in the initial prototype, the last three modules have a rather limited scope. In total, more than 900 objects and over 1,200 rules were required to develop the software, which can diagnose over 200 pathologies of building structures. The results obtained in the various phases of evaluation are totally satisfactory, indicating that the prototype of DH can diagnose dozens of pathologies with a level of certainty greater than 80% (in agreement with the tests performed). A great part of the errors detected were due to the knowledge limit of this prototype or the lack of in situ information. Furthermore, the level of certainty grows up to 90% if all of the information required for answering the questions is available. Finally, if the main features of DH are compared with the capacities of other Expert Systems in the field of maintenance of buildings some conclusions arise. First of all, DH is an expert system that deals with the diagnosis of structural pathologies in different structural elements and materials, whereas most of the similar expert systems focus in one specific structural element, eg. concrete bridge decks in Yehia et al. (2008), or one specific pathology, eg. cracks in reinforced concrete in Chao and Cheng (1998). However, DH is not yet capable of analysing together all the pathologies of a building for a full and auto-complementary diagnosis. Secondly, the aim of DH is quite different from the aim of most of the other similar expert systems because it is not oriented to achieve a full evaluation of the structural security of a building or to obtain a damage level, but to help a little-experienced human inspector in the first on field inspections at the detection of the most important and common pathologies in order to reach a most efficient resources distribution, assigning the high-experienced inspectors to the unusual or most complex pathologies. However, in order to have some simple damage level indicator, DH displays a unique warning message if it diagnoses one of the most dangerous pathologies. Comparing DH with other significant expert systems in the field of building inspection it is possible to obtain a better perspective of which are the achievements of the work herein presented. For example, MDDS (Masonry Damage Diagnostic System) from Van Balen (1996 and 2001) is an expert system aimed to carry out in situ inspections of masonry damage due to air pollution. Most of the information included in MDDS came from literature what also happens with DH but the main difference is how this knowledge is codified. As MDDS only uses decision tables and no production rules, the last decision step done in DH (usually a production rule) is not so easy in MDDS. This makes difficult to deal with uncertainty even in the simpler way used in DH. Comparing MDDS with DH other similarities are observed: both used an expert system shell to be developed, are goal oriented so only the necessary questions are asked and divide the variables between those that are fundamental to come up with a diagnostic and those complementary which are not even used in MDDS while in DH are considered in the last questions to give more confidence in the system output. After doing the tests with the developed expert system, two main advantages of MDDS are noted comparing with DH: the possibility of on-line actualisation and the recommendation of further tests to improve the diagnosis. Other possible comparison that could be made is between DH and REPCON, which was developed by Moodi and Knapton (2003). REPCON is aimed for offering a repairing solution for damages in concrete elements. Indeed, its knowledge is structured in 13 decision tables and just one of them is dedicated to diagnosis while it is the full purpose of DH which also has a greater scope. The main difference between REPCON and other expert systems in the field of damage evaluation of buildings that have been studied is that this one is able to evolve with the knowledge of the users. To do that it first assign a confidence level to the user based on the answers the expert system asked they. Then, if it is proved the user is an expert it is allowed to modify the data base. The confidence of the information is also evaluated taking into account only the source of it. On the contrary, DH has a static knowledge which offers more confidence in the obtained results but less upgrading velocity. However, in this branch of knowledge where almost every single pathology has already been studied it does not seem a big deal. Finally, whilst DH is clearly oriented to the formation of novel inspectors, REPCON has also this possibility but it is not directly aimed for it, resulting in a more complex system that covers a deeper analysis than DH in concrete elements. Finally, a comparison between DH and the KBES (Knowledge Based Expert System) presented by Lu and Simmonds (1997) has been carried out to highlight the possibility of incorporating calculus results as an input data to the expert system and, more important, to remark the possibility of reaching a global evaluation of the whole building from the evaluation of each element. This is a clear advantage in front of DH that shows not enough efficiency at gather all the diagnostic to offer a global assessment. On the other hand, the way the uncertainty is handled in the KBES presented by Lu and Simmonds (1997) is not the most useful for teaching purposes as it requires the user to assign a confidence level to each observed symptom. Moreover, the aim of DH differs from the aim of KBES by Lu and Simmonds (1997) because the latter makes an assessment of the building but does not identify the causes of the damages. To conclude with all these comparisons, one more time (see section 4.4) it is pointed out that the way DH deals with uncertainty is simpler than the common way other expert systems analysed by the authors (REPCON or KBES by Lu and Simmonds) do because, from the point of view of the developers of DH, in a first on field inspection it is not necessary to assign a probability percentage to a detected pathology and it is even less necessary to request a confidence level on the observation made when the user is being trained. It is also worth highlighting the way DH has its knowledge encoded using both production rules and decision tables together, which results in a more flexible system, with the variables clearly classified by their importance and allows the possibility of handling uncertainty in the way it has been done. A sample of DH (Doctor House) can be freely distributed under request. For greater detail on the procedure followed and the results obtained, the reader is referred to Bernat (2007). 7. Conclusions Nowadays the industry is slowly changing from classical construction towards sustainable building. Therefore, most of the recently constructed building will have to be surveyed and/or repaired during the following years. The field of construction is clearly marked by an ever increasing amount of “illnesses”, but “doctors” are not being trained quickly enough to be able to address these cases with a level of quality required by the market. Hence, there is a pressing need for a commercially available tool for this area, which would not only increase productivity in technical inspections but would also allow for on-going training of the professionals who use it, as well as it could easily become a research and data compilation tool on the “health” of buildings. After DH was developed and evaluated in the first phase, the following conclusions were reached: • Expert systems were determined to be the best platform on which to develop software for decision-making on diagnosis of building pathologies, despite the difficulty in establishing internal order that is inherent to these types of programs. • The type of damage that is easiest to classify is that affecting concrete structures, as they share many similarities, which limits casuistry and the data required for diagnosis. In contrast, the most difficult ones are those that affect masonry elements, which are more complex and are difficult to encode. Diagnosis of these pathologies requires diverse data. This means that the module rules for diagnosing masonry elements are more basic, less compact and have more entry objects. • Treating uncertainty qualitatively is a feasible approach; it provides good results. Indeed, for building pathology diagnosis, this treatment is actually recommended, as it provides greater transparency to the user and greater value in terms of critical analysis of the presented results. • The interface could be easily improved and the knowledge base could be expanded in all areas of building pathology. The last and most important conclusion is that the development of an expert system for diagnosing building pathology is feasible; indeed, creating a new, commercially available tool for practical applications is merely a question of providing the needed resources. 7.1. Future work In 6th section some weaknesses of DH have arisen from analysing its results and comparing it with other expert systems which means some improvements should be done in the future. These future developing directions would include the following activities: • to widely promote the use of DH among professional inspectors for detecting the main weaknesses and come up with a statistically more significant range of results, • to spread the knowledge data base which should grow to cover all possible common pathologies in every structural element, • to join the actual independent modulus in order to face a global assessment of the building instead of the evaluation of each structural element by their own, • to include graphical information to assure a better understanding of the damages, • to increase the functionality including suggestions about further tests or reparation techniques, • to open a new development branch to create a professional version aimed at diagnosing the most complex pathologies that should include a quantitative approach in handling uncertainty, • to include a function aimed for recollecting the data of the real uses in order to obtain new information about real damages and its frequency, • to interconnect DH’s users to create a network that would help further improvements and faster upgrading, • to add an optional calculus module in the hypothetical professional version in order to obtain analytical evidences that could back the obtained solution, • to move the expert system on other informatics support allowing it to be executed in nowadays smartphones and making it universally available. It is thought all this future work is necessary and would be worthy once the technical feasibility of this kind of software has been proved. Acknowledgements The authors gratefully acknowledge the support provided by company ATISAE, Project under contract MATE and Project under contract VALTEC of Generalitat de Catalunya. The first author also enjoys a Ph. D. grant offered by the Spanish Institution of Civil Engineers. References Albert, K. and Wu, W. (1997). “A knowledge-based application in engineering”, Advances in Engineering Software, Vol. 28, No.8, pp. 469-486. Anumba, C. J. and Scout, D. (2001). “Intelligent pathological assessment of housing subsidence”, Structures & Buildings, Vol. 146, No. 2, pp. 183-193. Bernat, E. (2007). Sistema Expert per a la diagnosi de patologies en elements estructurals, Degree Thesis, ETSECCPB, Universitat Politècnica de Catalunya, Barcelona, Spain. Brito, J. and Branco, F.A. (1997). “An expert system for concrete bridge management”, Engineering Structures, Vol. 19, No. 7, pp. 519-526. Calavera, J. (2005). Patología de estructuras de hormigón armado y pretensado, INTEMAC S.A., Madrid, Spain. Chao, C.J. and Cheng, F.P. (1998). “Fuzzy Pattern Recognition Model for Diagnosing cracks in RC Structures”, Journal of computing in civil engineering, Vol.12, No.2, pp. 111-119. Cheng, C.T., Chunping, O. and Chau, K.W. (2002). “Combining a fuzzy optimal model with a genetic algorithm to solve multiobjective rainfall-runoff model calibration”, Journal of Hydrology, Vol. 268, No. 1-4, pp. 72-86 Giarratano, J. and Riley, G. (1994). Expert Systems. Principles and programming, PWS Publishing Co., Boston, MA, USA. González, M.P., and Zapico, J.L. (2007). “Seismic damage identification in buildings using neural networks”, Computers & Structures, Vol.86, No. 3-5, pp. 416-426 Harmon, P. and King, D. (1985). Expert Systems: Artificial Intelligence in Business, John Wiley & Sons Inc., New York, USA. Jiang, A.N., Wang, S.Y. and Tang, S.L. (2011). “Feedback analysis of tunnel construction using a hybrid arithmetic based on Support Vector Machine and Particle Swarm Optimisation”, Automation in Construction, Vol.20, No. 4, pp. 482-489. Kaetzel, L.J. and Clifton, J.R. (1995). “Expert/Knowledge based systems for materials in the construction industry: State-of-the-art report”, RILEM Journal of Materials and Structures, Vol.28, No. 177, pp. 160-174. Kaklauskas, A., Zavadskas, E. and Trinkunas, V. (2007). “A multiple criteria decision support on-line system for construction”, Engineering Applications of Artificial Intelligence, Vol. 20, No. 2, pp. 163-175. Kim, Y.M. and Kim, C.K. (2007). “Fuzzy set based crack diagnosis system for reinforced concrete structures”. Computers & Structures, Vol. 85, No. 23-24, pp. 1828-1844. Koppány, A. (2005). “Diagnostic tools and a new building defect registration system”, 10th DBMC International Conference On Durability of Building Materials and Components, International Council for Research and Innovation in Building and Construction, ASTM and CSTB eds., Lyon, France, 17-20 April, pp. 1067-1074. Kumaraswamy, M.M. and Dissanayaka, S.M. (2001). “Developing a decision support system for building project procurement”, Building and Environment, Vol. 36, No. 3, pp. 337-349. Lee, J., Liu, K.F.R. and Chiang, W. (1999). “A Fuzzy Petri Net-Based Expert System and Its Application to Damage Assessment of Bridges”, IEEE Transactions on systems, man, and cybernetics – Part B: Cybernetics, Vol. 29, No. 3, pp. 350-369. Lu, X. and Simmonds, S.H. (1997). “KBES for evaluating R.C. framed buildings using fuzzy sets”, Automation in Construction, Vol. 6, No. 2, pp. 121-137. Moodi, F. and Knapton, J. (2003). “Research into a Management System for Diagnosis, Maintenance, and Repair of Concrete Structures”, Journal of Construction engineering and management, Vol. 129, No. 5, pp. 555-561. Muñoz, M. (1988). Conceptos y patología en la edificación, M. Muñoz ed., Sevilla, Spain. Murlidharan, T.L., Durgaprasad, J. and Appa Rao, T.V.S.R. (1997) “Knowledge-based expert system for damage assessment and vulnerability analysis of structures subjected to cyclones”, Journal of Wind Engineering and Industrial Aerodynamics, Vol. 72, No. November-December, pp. 479-491. Muttil, N. and Chau, K.W. (2006). “Neural network and genetic programming for modelling coastal algal blooms”, International Journal of Environment and Pollution, Vol. 28, No. 3-4, pp. 223-238. Najjaran, H., Sadiq, R. and Rajani, B. (2006). “Fuzzy Expert System to Assess Corrosion of Cast/Ductile Iron Pipes from Backfill Properties”, Computer-Aided Civil and Infrastructure Engineering, Vol. 21, No. 1, pp. 67–77. Peter, P.F.C. (1996). An expert system for diagnosis of problems in reinforced concrete structures, Minor Thesis, Royal Melbourne Institute of Technology, City Campus, Melbourne, Australia. Poon, J. (2004). “Development of an expert system modelling the construction process”, Journal of Construction Research, Vol. 5, No. 1, pp. 145-158. Rodríguez, A. (2000). Siniestros más frecuentes en la construcción de edificios, MUSAAT, Madrid, Spain. The Masonry Society Construction Practices Committee, eds., (2004). Masonry Inspection Checklist, Boulder, CO, USA. Tiryaki, B. (2007). “Application of artificial neural networks for predicting”, Tunnelling and Underground Space Technology, Vol. 23, No. 3, pp. 273-280. Tanarslan, H.M., Secer, M. and Kumanlioglu, A. (2012). “An approach for estimating the capacity of RC beams strengthened in shear with FRP reinforcements using artificial neural networks”, Construction and Building Materials, Vol. 30, No. May, pp. 556-568. Van Balen, K. (1996). “Expert system for evaluation of deterioration of ancient brick masonry structures”, Science of the Total Environment, Vol. 28, No. October, pp. 247-254. Van Balen, K. (2001). “Learning from damage of masonry structures, expert systems can help!”, III International Seminar on historical constructions, Lourenço, P.B. and Roca, P., eds., University of Minho, Guimaraes, Portugal, November, pp. 15-27. Xie, J.X., Cheng, C.T., Chau, K.W. and Pei, Y.Z. (2006). “A hybrid adaptative time-delay neural network model for multi-step-ahead prediction of sunspot activity”, International Journal of Environment and Pollution, Vol. 28, No. 3-4, pp. 364-381. Yamazaki, Y., Uchiyama, Y., Ito, K., Akimoto, M. and Terada, N. (1990). “Expert Systems and integrated construction planning”, Habitat International, Vol. 14, No. 2-3, pp. 95-106. Yehia, S., Abudayyeh, O., Fazal, I and Randolph, D. (2008). “A decision support system for concrete bridge deck maintenance”, Advances in Engineering Software, Vol. 39, No. 3, pp. 202- 210. Zhang, H. and Xing, F. (2010). “Fuzzy-multi-objective particle swarm optimization for time- cost-quality tradeoff in construction”, Automation in Construction, Vol. 19, No. 8, pp. 1067- 1075. Zhao, Z. and Chen, C. (2002). “A fuzzy system for concrete bridge damage diagnosis”, Computers & Structures, Vol.80, No. 7-8, pp. 629-641. Zhao, M.Y., Cheng, C.T., Chau, K.W. and Li, G. (2006). “Multiple criteria data envelopment analysis for full ranking units associated to environmental impact assessment”, International Journal of Environment and Pollution, Vol. 28, No. 3-4, pp. 448-464. Figure 1. Encoding knowledge Figure 2. A production rule of the developed prototype Figure 3. A pathology diagnostic tree Figure 4. Flowchart of DH Figure 5. Crack to be analysed from side 1 Figure 6. Crack to be analysed from side 2 Figure 7. A screenshot of DH 1. Introduction 2. Review on Artificial Intelligence in construction and Aim of the work 3. Used tool: Expert Systems 3.1. Why an expert system? 3.2. Expert Systems drawbacks 4. Developing the prototype. Methodology 4.1. Focusing on a specific area 4.2. Choosing the programming tool 4.3. Encoding and introducing the information using Acquire 2.1 4.4. How to handle uncertainty 4.5. User interface and translation possibilities 5. Running DH: How does it diagnose? 6. Results, Discussion and Comparison with similar Expert Systems 7. Conclusions 7.1. Future work 
work_vs53ce2hjzfbha3l3pbs4j24be	----	1 Combining Domain Knowledge and Machine Learning for Robust Fall Detection Mirchevska Violeta 1 , Luštrek Mitja 2 , Gams Matjaž 2 (1) Result d.o.o., Bravničarjeva 11, Ljubljana, Slovenia (2) Jožef Stefan Institute, Jamova 39, Ljubljana, Slovenia E-mail: violeta.mircevska@result.si Abstract: This paper presents a method for combining domain knowledge and machine learning (CDKML) for classifier generation and online adaptation. The method exploits advantages in domain knowledge and machine learning as complementary information sources. While machine learning may discover patterns in interest domains that are too subtle for humans to detect, domain knowledge may contain information on a domain not present in the available domain dataset. CDKML has three steps. First, prior domain knowledge is enriched with relevant patterns obtained by machine learning to create an initial classifier. Second, genetic algorithms refine the classifier. Third, the classifier is adapted online based on user feedback using the Markov decision process. CDKML was applied in fall detection. Tests showed that the classifiers developed by CDKML have better performance than ML classifiers generated on a one-sided training dataset. The accuracy of the initial classifier was 10 percentage points higher than the best machine learning classifier and the refinement added 3 percentage points. The online adaptation improved the accuracy of the refined classifier by additional 15 percentage points. Keywords: prior domain knowledge, inductive machine learning, ambient intelligence, fall detection 1. INTRODUCTION The training dataset to which machine learning is applied is often one-sided, not representing all real-life cases (Li et al., 2007; Yang & Kecman, 2009). A typical example is medical studies based on laboratory samples upon which machine learning methods are applied. There is an important difference between laboratory samples that consider a limited number of clear-case scenarios and real life. For a classifier induced by machine learning to work in the general case, it must be induced using a sufficiently large and representative training dataset. Because such data is not always available, this can be partially countered by expert domain knowledge. Expert domain knowledge may be related to examples not present in the available domain dataset and thus may improve the generality and robustness of classifiers induced on such datasets. This paper addresses the problem of combining domain knowledge (DK) and machine learning (ML). It contributes a method for combining DK and ML (CDKML) for classifier generation and online adaptation. We demonstrate our method on a fall detection task. This task is relevant for the elderly and the European Union, whose population is rapidly ageing. Predictions made by the Statistical Office of the European Communities state that the over-65 population in EU27 expressed as a percentage of the working-age population (aged between 15 and 64) will rise from 26% in 2010 to 53% in 2060 (Eurostat, 2011). This demographic change will make medical and care services scarce, increasing the need to motivate and assist the elderly to stay independent as long as possible. Innovative ICT systems can help 2 the elderly live independently for longer and counteract reduced capabilities caused by age. One such system, developed as part of a European FP7 project, is the Confidence – Ubiquitous care system to support independent living (Confidence, 2011). Confidence aims to develop a system to monitor the health conditions of its elderly users in real-time. It detects falls and behaviour changes, including limping and physical inactivity. Confidence is based on wearable tags attached to a user reporting x, y and z tag coordinates with around 15 cm accuracy. Polls show that, with respect to privacy-violation issues, such hardware is more acceptable to users than, for example, video cameras. We used CDKML to develop the fall detection classifier in Confidence. Confidence has three fall detection properties that make it challenging from an ML perspective. First, a representative dataset for falls is difficult to obtain because of the variety of fall types (in consultation with medical experts, we compiled a list of 18 fall types from over 40 listed in the literature), variations depending on the user, as well as ethical issues and injury dangers that prevent collecting large amounts of data from healthy persons simulating falls or, even worse, the elderly. Second, developing a classifier to suit each user in each possible circumstance from the start is difficult. Confidence detects falls as situations in which the user is lying/sitting motionless on the ground for a prolonged period of time. However, it is difficult to set a period of time to suit each user. For example, one user might never voluntarily lie or sit on the ground because of a physical disability that prevents him/her from getting up again, whereas another might exercise regularly on the living room carpet. Therefore, an online classifier adaptation is needed. Third, because of noise in the sensor data, misclassifications between similar postures occur. For example, sitting on a low chair may be misclassified as sitting on the ground. Such misclassifications of the posture directly influence the output of the fall detection model. Motivation for developing the CDKML method lies in addressing ML shortcomings through DK. Using only initial clear-cases of the domain of interest, our method can create classifiers with improved general performance than ML classifiers induced on a one-sided dataset. The method also encompasses online classifier adaptation using information obtained from user feedback. The feedback is obtained occasionally and contains information about false negatives (i.e., the system did not detect the class of interest when there was one) or false positives (i.e., the system detected the class of interest when there was not one). The paper is organized as follows. In Section 2, we present related work about combining DK and ML for classifier generation. In Section 3, we present the CDKML method for combining DK and ML for classifier generation and online adaptation. In Section 4, we present the experiments used to test the proposed method and obtained results. Section 5 concludes the paper. 2. RELATED WORK Cognitive psychology research shows that human concept-learning considers both prior DK and interest concept examples (Wisniewski & Medin, 1994; Feldman, 1993; Heit, 2000). In principle, one information source offsets information missing from another source. DK influences interpreting examples. Before 3 obtaining a considerable amount of concept examples, humans base their judgements mainly on prior DK. Conversely, examples affect DK. As the number of observed items of the interest concept increases, judgment relies increasingly on the actual observations and less on prior DK. ML literature includes examples of concept learning using both prior DK and interest concept examples. A comprehensive overview of methods for incorporating prior DK into inductive ML is presented in (Yu, 2007). Yu categorizes these methods into four groups: (1) methods that use prior DK to prepare training examples, (2) methods that use prior DK to initiate the hypothesis or hypothesis space, (3) methods that use prior DK to alter the search objective and (4) methods that use prior DK to augment the search. The first group of methods incorporates prior DK into the training dataset used for induction by inserting virtual examples into the training dataset (Kambar, 2005; Niyogi et al., 1998; Poggio & Vetter, 1992). Niyogi et al. (1998) showed that adding virtual examples is mathematically equivalent to incorporating the prior DK as a regulariser in function learning in certain restricted domains. In the second group, prior DK determines the part of the hypothesis space searched during induction (Zhu & Liu, 2010; Burns & Danyluk, 2000; Thrun, 1996). This is achieved by determining which part of the hypothesis space satisfies prior DK and using ML to search for a hypothesis in it, or by creating an initial hypothesis from the prior DK and using ML to refine it. The third group incorporates the DK into the inductive bias that guides the search through the hypothesis space. This is achieved by modifying the goal criterion to satisfy both DK and training examples, as in learning with constraints (Sabzekar et al., 2011; Chen et al., 2011; Davidson & Ravi, 2005), or by weighting the examples' influence in the training dataset (Brown et al., 2000; Wu & Srihari, 2004; Wang et al., 2004). The fourth group produces hypothesis candidates and adjusts the hypothesis space using DK during the on-going search (Decoste & Scholkopf, 2002; Pazzani et al., 1991). In all cases, incorporating the DK aims to improve the generality of the induced ML model and/or the efficiency of the learning process. The interactive ML field also explores methods for concept learning using both prior DK and interest concept examples. Compared to the previously described methods that incorporate DK into the ML algorithm, interactive ML is basically an iterative process of classifier generation through human- computer interaction (Benyon, 2001). Two strategies for model generation using interactive ML can be distinguished: (1) iterative improvement of a single model by refining the input information used during ML induction (Sun & Hardoon, 2010; Stumpf et al., 2009; Bramer, 2005) and (2) generating multiple models to select one or several that are the most relevant from the user's perspective (Vidulin & Gams, 2011; Osei-Bryson, 2004). The CDKML method belongs to the group of methods that use prior DK to initiate the hypothesis or hypothesis space. The domain expert determines the initial classifier and hypothesis space using DK only or using interactive ML. Genetic algorithms then refine the initial classifier using the available training dataset. Here, DK is included as a set of constraints on the classifier form (e.g., for a classifier with the form of a rule set, relations between rule parameters). Finally, the Markov decision process enables online classifier adaptation based on user feedback. DK specifies the mapping from obtained user feedback to changes of the state rewards. We are unaware of any work similar to combining the three steps. The final step is also novel. 4 3. CDKML METHOD Figure 1 presents a general CDKML schema – a method for combining DK and ML for classifier generation and online adaptation. The schema contains three phases: (1) initialization, (2) refinement and (3) online adaptation. In the first phase, the domain expert specifies the hypothesis space and initial classifier. The domain expert may apply ML to the available training dataset, generating human-understandable classifiers to obtain additional insight in the interest domain (Vidulin & Gams, 2011) and include parts of these ML classifiers in the initial classifier. After determining the initial classifier, genetic algorithms refine it in the second phase, under expert supervision. The third phase adapts the classifier online using feedback information obtained from the user, who may indicate that the output class was incorrect. The adaptation is defined as a Markov decision process where user feedback is considered a reward signal from the environment. CDKML is not bound to a specific classifier form, but requires a human- understandable form. In the following subsections, we present each CDKML phase in detail. Initial model Refined model Adapted model 3. ONLINE ADAPTATION 2. REFINEMENT1. INITIALIZATION Learning human- understandable classifiers Domain knowledge User feedback Dataset OUTPUT PHASE ML INPUT Genetic algorithms Markov decision processes Figure 1 CDKML – method for combining DK and ML for classifier generation and online adaptation The CDKML presentation is accompanied by examples from its application in the fall detection domain. In this specific domain, we selected the classifiers in the form of a set of disjunctive rules: IF condition1 AND condition2 AND … AND conditionN THEN class. We chose this classifier form because it can be constructed manually or with the help of supervised learning and modified by genetic algorithms or Markov decision processes. 5 3.1. Initialization In the first CDKML phase (initialization), a domain expert defines an initial classifier. For example, in the fall detection domain, an expert may specify that if an elderly person is lying or sitting on the ground for a long period of time, then there is high probability of a fall, as elderly people are unlikely to lie or sit on the ground. Figure 2 presents an outline of this phase. Function Initialization Input: training examples Ex Output: initial classifier CLinit begin CLinit := empty_set_of_rules add rules in CLinit from domain knowledge // Explore human understandable ML models (e.g. decision tree, rule set) ML_ModelType := {decision tree, set of rules, etc.} for each ML_ModelType do create ML model on Ex using different initial parameters and attribute vectors explore patterns from the induced ML model add relevant rules in CLinit end do end Figure 2 Phase 1 – initialization As an aid for designing the initial classifier, the expert may also examine human-understandable classifiers induced by supervised learning on the available training data. An example is presented in (Mirchevska et al., 2009), where several decision trees are iteratively created to explore the space of possible classifiers. From these classifiers, the expert may obtain additional insight in the domain, modify DK, or add extracted patterns in the initial classifier. In the fall detection example, the starting rule-based classifier contained the following rule types: 1. IF falling activity within T1fall seconds AND the user was lying/sitting on the ground P1activity% of T1activity seconds AND the user was not moving P1moving% of T1moving seconds THEN fall 2. IF falling activity within T2fall seconds AND the user was lying/sitting on the ground area afterwards P2activity% of T2activity seconds THEN fall 6 3. IF the user was lying/sitting on the ground for P3activity% of T3activity seconds AND the user was not moving P3moving% of T3moving seconds THEN fall 4. IF the user was lying/sitting on the ground for P4activity% of T4activity THEN fall The expert specified these types of important fall patterns. However, specifying exact values for the parameters in the rules, for example specifying exact values for the parameters P4activity and T4activity in the last rule of the fall detection classifier, presented an issue for the expert, as the values may be influenced by system features, such as the noise in the sensor data or the ability of the system to correctly detect the lying/sitting posture. While the expert specified some initial values, and some values were obtained from the generated classifiers, the expert was not confident in them. 3.2. Refinement The second CDKML phase (refinement) refines the initial classifier set from the domain expert to conform to system-related and general-user characteristics evident from the training dataset. As the rule structure in the classifier is fixed, standard rule induction methods are unsuitable for the desired training. Genetic algorithms (Eiben & Smith, 2003) thus tune the initial classifier parameters to maximize its performance on the training dataset. Figure 3 presents an outline of this phase. We apply genetic algorithms thus: We use the Pittsburgh approach, where each individual in the population represents one possible solution. The individual is a vector containing parameters of all rules in the rule-based classifier. For example, if the rule-based classifier contains 8 rules with 4 parameters each, the individual is 32 elements long. The elements are real values inside an interval defined by the expert. The fitness function for evaluating the quality of each individual is accuracy on the training dataset. Fitness values fall within the interval [0, 1]. We thus want to tune the parameters of the rule- based classifier to the training dataset, hopefully avoiding overfitting, as the domain expert defines the structure of the rule-based classifier. We used elitism, meaning that the best individual is always transferred to the new population. Using genetic algorithms allows constraining relations between rules and parameters within a rule. In the presented fall detection classifier, rule strictness decreases from rule type 1 to rule type 4. The first rule type requires detecting falling activity and the user to be immovable and lie/sit on the ground to detect a fall, whereas the fourth rule type requires only the user to lie/sit on the ground. The duration of lying/sitting on the ground needed for the first rule type to detect a fall should be the shortest (the combination with other evidence more quickly assures that a fall happened) and should increase toward rule type 4. The relation between the required periods of lying/sitting on the ground in the rules is expressed through constraints. Additionally, if the rule requires detecting falling activity to detect a fall, the falling activity should be detected before the person lied/sat on the ground. This relation is also represented as a constraint. The fitness of individuals representing rule-based classifiers that violate the constraints is set to minimal fitness. 7 Function Refinement Input: initial classifier CLinit; constraints between the rule parameters of the initial classifier Constraints; training examples Ex; parameters of the genetic algorithms POPULATION_SIZE, CROSSOVER_RATE, MUTATION_RATE, STOP_CRITERION, MAX_ITERATIONS Output: refined classifier CLref begin //Create individual Ibase representing the initial classifier CLinit using the Pittsburgh //approach Ibase := empty_vector for each rule r in CLinit do put the parameters of rule r in a single vector Vecr append Vecr to Ibase end do //Create initial population Pinit := empty_set add Ibase to Pinit for i:=1 to POPULATION_SIZE do Ii := create an individual by random changes of Ibase add Ii to Pinit end do //Evolve population Ibest:= Find_fittest_individual(Pinit, Constraints, Ex) //function defined below iter := 0, Pnew:= Pinit while ((accuracy(Ibest) < STOP_CRITERION) AND (iter < MAX_ITERATIONS)) do Pold := Pnew, Pnew := empty_set, iter:= iter+1 add Ibest to Pnew //Use elitism for i:=1 to (POPULATION_SIZE/2) do select two parents from Pold by tournament selection crossover parents with probability CROSSOVER_RATE mutate individuals obtained by crossover with probability MUTATION_RATE add new individuals to Pnew end do Icur_best := Find_fittest_individual(Pnew, Constraints, Ex) 8 if(accuracy(Icur_best) > accuracy(Ibest) then Ibest:= Icur_best end if end do CLref := update the values of the parameters in CLinit with the values in Ibest end Function Find_fittest_individual Input: population P; constraints between the parameters of the initial classifier Constraints; training examples Ex Output: individual Iresult begin for each I in P do if I violates Constraints fitness(I) := 0 else fitness(I) := accuracy(I) on Ex end if end for each Iresult := individual in P with highest fitness value end Figure 3 Phase 2 – refinement The genetic algorithm outputs the final general rule-based classifier. The expert should observe various classifiers generated with different input parameters and choose the best one in his/her opinion, not solely based on accuracy. 9 3.3. Online adaptation The third CDKML phase (online adaptation) adapts the general rule-based classifier online using feedback obtained from a particular user. We explain the adaptation process using the example rule: “IF the user was lying on the ground for Pactivity% in Tactivity THEN fall”. Figure 4 presents an outline of this phase. Function Initialize_MDPs Input: refined classifier CLref Output: set of MDPs begin for each rule r in CLref do create a MDPR(S, A, P, R) for r initialize the set of states S to a n-dimensional state space, where n is the number of parameters in r initialize the set of actions A to all possible parameter value changes initialize the transition probability P(s, a, s’) to be 0 or 1 initialize the elements of the reward matrix R to zero initialize current state MDPR.current to the parameter values of r end do end Function Update_rules Input: current classifier CLcurrent; user feedback UF {false positive, false negative} accompanied with the triggering example Ex; penalty amount for false positive PaFp; penalty amount for false negative PaFn Output: adapted rule-based classifier CLadpt begin if UF= false positive then Rfp := set of all rules of CLcurrent that caused a false positive for each rule r in Rfp do in MDPR reduce the utility of the current state and all states that it dominates by PaFp Cstates := set of neighbouring states of MDPR.current with highest utility 10 MDPR.current := state from Cstates with maximum distance from Ex end do else // UF = false negative find rule r in CLcurrent with minimum distance from Ex in MDPR reduce the utility of the current state and all states that dominate it by PaFn Cstates := set of neighbouring states of MDPR.current with highest utility MDPR.current := state from Cstates with minimum distance from Ex end if end Figure 4 Phase 3 – online adaptation The problem of adapting a rule in the rule-based classifier is defined as a Markov decision process (Russell & Norvig, 2003), MDPR(S, A, P, R). The state space S of each rule is N-dimensional, where N is the number of adjustable parameters in the rule. The example rule space is two-dimensional, with one dimension representing the set of possible percentage values and the other representing possible time interval values (Figure 5). We use discrete parameters. In each step, we can increase or decrease the value of each parameter by one unit. The set of actions A are combinations of such parameter value changes. Parameter value changes are deterministic; the values in the transition probability matrix P(s, a, s’), denoting the probability of transitioning from state s to s’ when executing action a, are 0 or 1. The elements of the reward matrix R reflect the obtained user feedback (a reward signal from the environment) and may change. Translating user feedback to the appropriate state reward requires information of how each parameter influences the output of the rule, as specified by the domain expert. The MDP goal state is the combination of rule parameter values that best separates fall events from non- fall events and depends on the needs of a particular user and may change through time. Figure 5 presents the process of adapting the example rule. Current rule parameter values (MDPR.current) are highlighted with a black rectangle. First of all, the reward matrix R of MDPR is initialized to zero for all states and MDPR.current is set to the rule's values in the refined classifier (Figure 5a). We assume that, after a certain period of time, a false positive feedback is obtained. In this concrete rule, a false positive feedback reduces the current state reward and all states dominated by it (states with less strict parameter values than the current state’s parameter values) by a penalty amount paFp, which in our example is -1, because a false positive indicates that the parameters of the rule must be made stricter (Figure 5b). After obtaining such feedback, the set of neighbouring MDPR.current states with the highest utility is determined, and the new MDPR.current value is the state with the maximum parameter distance from the example that caused the false positive. In the example rule, the distance from a state s to an example Ex that triggered user feedback is calculated thus: 11 where T is the set of possible time interval values in the rule, stimeInterval and spercentage represent the rule parameter values represented by state s, and Expercentage(t) represents the amount of lying on the ground in time interval t in Ex. The new MDPR.current value in Figure 5b has stricter values for both the time and percentage parameters. We again assume that, after a certain period of time, a false negative feedback is obtained. A false negative feedback reduces the reward of the current state and all states that dominate it (states with stricter parameter values than the current state’s parameter values) by a penalty amount paFn, which in our example is -1, because a false negative indicates that the parameters of the rule are too strict and must be relaxed (Figure 5c). Again, the set of neighbouring MDPR.current states with the highest utility is determined, and the new value of MDPR.current is the state with the minimum parameter distance from the example that caused the false negative. Figure 5c presents a case where the feedback result reduced the strictness of the percentage parameter of MDPR.current, while the time parameter remained unchanged. The initial state was avoided because of the negative reward received during the first false positive. Rule parameters values are adapted in this way after each obtained user feedback. Figure 5 Online classifier adaptation using user feedback 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 IF person lying on the ground 70% in 9 seconds THEN alarm p e rc e n ta g e 8 9 10 11 12 FALSE POSITIVE time interval (s) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 IF person lying on the ground 80% in 10 seconds THEN alarm p e rc e n ta g e 5 0 6 0 7 0 8 0 9 0 1 0 0 time interval (s) 8 9 10 11 12 F A L S E N E G A T IV E 0 0 -1 -1 -1 0 0 -1 -1 -1 0 0 -1 -1 -1 -1 -1 0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 IF person lying on the ground 80% in 9 seconds THEN alarm p e rc e n ta g e 5 0 6 0 7 0 8 0 9 0 1 0 0 time interval (s) 8 9 10 11 12 FALSE POSITIVE 0 0 -1 -1 -1 0 0 -1 -1 -1 -1 -1 -1 -1 -1 -2 -2 0 0 0 -2 -2 0 0 0 -2 -2 0 0 0 IF person lying on the ground 70% in 10 seconds THEN alarmp e rc e n ta g e 5 0 6 0 7 0 8 0 9 0 1 0 0 time interval (s) 8 9 10 11 12 5 0 6 0 7 0 8 0 9 0 1 0 0 a) b) d) c) 12 User feedback does not affect all rules in the rule-based classifier. If false positive feedback is obtained, only the rules that incorrectly classified the concrete example as positive are adapted. If false negative is obtained, only the rule that needs the least change to cover the concrete example is adapted. 4. EVALUATION This section evaluates CDKML on the fall detection task. We used CDKML to build a rule-based classifier for fall detection as part of the fall detection module in the Confidence system. We first describe the fall detection module in the Confidence system. We then present the data on which the rule-based fall detection classifier was evaluated. Finally, we comment on the obtained results. 4.1 Fall detection in Confidence This section presents the fall detection module of the Confidence system by which CDKML was evaluated. Figure 6 presents a simplified version of the part of the Confidence system related to fall detection. Detailed system descriptions can be found in literature (Lustrek et al., 2011; Kaluza et al., 2010) and a detailed description of the fall detection module in (Mirchevska et al., 2010). Preprocessing and filtering Attribute computation Activity recognition Fall detection Preprocessed RTLS data Attributes: distances between tags, tag velocity and similar Current user activity; User level of movement User has fallen/has not fallen Raw RTLS data U s e r fe e d b a c k Figure 6 Fall detection in the Confidence system 13 In the Confidence system, the user is equipped with wearable tags whose coordinates are detected by radio sensors. The experiments presented in this paper used the real-time localization system (RTLS) Ubisense (Ubisense, 2011) for this purpose. In a typical open-environment, the localization accuracy of Ubisense is on average about 15 cm but in practice may occasionally drop to 200 cm or more. The raw RTLS data is first preprocessed to reduce noise. The preprocessed RTLS data is then given as input to the attribute computation module. This module computes characteristics of the user's body, including tag velocity and amount of movement, and relations between body parts, including the distance between tags. The activity recognition module uses these characteristics to classify the user’s activity into one of seven classes: standing, sitting, lying, standing up, going down, falling, or on all fours. Additionally, if the system detects lying or sitting, it determines whether these activities are done at appropriate places, including a bed for lying or chair for sitting, or at inappropriate places, such as on the ground. The activity recognition module's output is given as input to the fall detection module. The fall detection module uses data concerning user activity history and user movement levels to detect falls, using the four rule types shown in Section 3.1, which mostly depend on whether an elderly person is lying or sitting at an inappropriate place (e.g., on the ground) for a long period of time, resulting in a high probability of a fall. Fall detection does not rely only on detecting the falling activity (high acceleration toward the ground), as it always lasts a very short time and is thus difficult to recognize. Compared to detecting falling activity, lying and sitting on the ground are easier to detect, which makes them convenient for fall detection. However, this approach has certain issues. Activity on all fours may be misclassified as lying on the ground. Because lying on the ground indicates a fall, such misclassifications may lead to false positives. However, activity on all fours that occurs when a person is searching for something on the ground is shorter than the period of lying/sitting on the ground that follows a fall and includes more movement. Another common misclassification occurs when a person is sitting on a low chair. Sitting on a low chair may be misclassified as sitting on the ground because of the noise in the localization system measurements and may cause false positives. However, the amount of sitting on the ground recognized when the person is sitting on a low chair should be lower than the amount of this activity recognized when the person is sitting on the ground. Therefore, the main challenge faced when developing the fall detection classifier is providing reliable and robust fall detection even in various complex real life circumstances. 4.2 Data We designed a test scenario to investigate the generality and robustness of the developed rule-based classifiers, as well as their adaptation capabilities. The scenario (Table 1) contains two types of events: straightforward and complex events. Straightforward events represent typical fall and non-fall events. Both fall events (1 and 2) involve high acceleration toward the ground during the falling activity. High acceleration during the falling activity is a characteristic feature of falls, and setting thresholds for it is a common way of detecting falls. The user lands lying (1) or sitting (2) on the ground after the fall. Non-fall events contain activities commonly done 14 at home, including walking, sitting on a chair, or lying in bed (3). Additionally, searching for something on the ground on all fours or lying (4) is added as a non-fall event. Complex events represent atypical falls and non-fall events that may be particularly easily misclassified. One type of non-fall event is lying down quickly on a bed or sitting down quickly on a chair (7). This event includes high acceleration during the lying/sitting down activity, which is a characteristic feature of falls. However, the lying/sitting that follows is on the bed/chair, enabling the rule-based classifier to differentiate falls from non-falls. The other non-fall event is sitting on a low chair (8). Five non-fall events of sitting on a low chair are present in the scenario. They differ in the position of the user’s body on the chair: the user sits straight or leans forward, backward, to the left, or to the right. In complex fall events (5 and 6), the user slowly descends to the ground, trying to hold onto nearby furniture. However, after the falling activity, the user lands lying/sitting on the ground. Table 1 Test scenario STRAIGHTFORWARD EVENTS COMPLEX EVENTS Description Fall Description Fall 1 Tripping, landing flat on the ground Yes 5 Falling slowly (trying to hold onto furniture), landing flat on the ground Yes 2 Falling when trying to stand up, landing sitting of the ground Yes 6 Falling slowly when trying to stand up (trying to hold onto furniture), landing sitting on the ground Yes 3 Normal everyday behaviour, such as walking, sitting on a chair, lying in bed No 7 Lying down quickly on the bed / Sitting down quickly on the chair No 4 Searching for something on the ground on all fours and lying No 8 Sitting on a low chair No We selected the falls in the test scenario from a list of 18 fall types, compiled in consultation with medical personnel. The falls were demonstrated by a physician, who also provided guidance during initial recordings. All events present in the test scenario were recorded in single recordings interspersed with short periods of walking. Each recording lasted around 20 minutes. The recordings were made by 5 healthy volunteers (3 male and 2 female), 5 times by each. Figure 7 presents the total number of fall and non-fall examples in the recorded data. The large number of non-fall events among the complex events is due to the examples of sitting on a low chair. We recorded many such examples because the adaptation (in the third phase) primarily occurred on them. 15 Fall events Non-fall events Complex eventsStraightforward events No. of examples 150 100 50 Figure 7 Total number of fall and non-fall examples in the recorded data 4.3 Results We evaluated the first and second CDKML phases as follows: The domain expert first specified the initial classifier. Genetic algorithms then refined the initial classifier based only on examples of straightforward events (to demonstrate laboratory testing). We used leave-one-person-out evaluation, where the refined classifier was generated from examples of four people and tested on examples of the fifth, which was excluded from the training dataset. This was repeated five times, using a different person for testing each time. The accuracy of the refined classifier was tested on both straightforward and complex events of the person excluded from the training dataset, thus illustrating real-life performance, which includes both clear and complex cases. The test on the straightforward events shows how well the classifier performs on events present in the training dataset. The test on the complex events, conversely, tests the generality and robustness of the generated classifier, as the complex events are not present in the learning process. We evaluated the online classifier adaptation part, i.e., phase 3 of CDKML, thus: The refined classifier was adapted to a concrete user using examples of both straightforward and complex concrete user events, because we wanted to test the ability of the method to learn new cases while preserving its performance on the cases present in the training dataset in phase 2 of CDKML. Four of the five concrete user scenario recordings were randomly presented one by one to the fall detection classifier. The fall detection classifier classified each event as fall or non-fall, then feedback was provided and the fall detection classifier was adapted, as necessary, before the next event. The final adapted classifier evaluation was done on the recording, which was not used in the adaptation phase. For comparison, fall detection classifiers were induced using ML only. The attributes were the time since detecting the last falling activity, the amount of each type of activity in time intervals from 5 to 15 seconds, and the amount of user movement in this interval range. The attributes are equivalent to the parameters of the rules in the rule-based fall detection classifier. We used the following ML algorithms: decision trees (J48), rules (JRip), support vector machines (SMO), random forest (RandomForest), and Naïve Bayes (NaiveBayes). In brackets, we give the Weka implementation (Hall et al., 2009) for these algorithms. 16 We evaluated fall detection classifier performance using two measures: accuracy on a subset of events ACCevents and F-measure on a subset of events FMevents. The accuracy on a subset of events ACCevents is The F-measure on a subset of events FMevents is where Pevents is the precision and Revents is the recall of the classifier of events E belonging to the set events. Table 2 presents the performance of the induced fall detection classifiers on straightforward events only, on complex events only, and on the whole sequence with respect to the accuracy on fall examples ACCf, accuracy on non-fall examples ACCnf, overall accuracy ACCall and overall F-measure FMall. Table 3 presents the accuracy of the induced classifiers on each event in the test scenario separately ACCe. The measures were computed for each person separately, and the values in Tables 2 and 3 represent the averages. Additionally, the refinement CDKML phase was performed five times in each test run, because of the stochastic nature of the genetic algorithm and the average value was considered. Table 2 shows that the best overall accuracy among ML classifiers was obtained by support vector machines with an ACCall of 53 percentage points. The ML classifiers tended to be biased towards fall recognition. They had maximal ACCf; however, they raised many false positives, as indicated by the low ACCnf values. The overall accuracy of the initial classifier was 10 percentage points higher than support vector machines. It slightly decreased on the ACCf, from 100 to 98 percentage points, but increased greatly on the ACCnf from 30 to 46 percentage points. The refinement of the initial classifier based on straightforward-event examples contributed to a 3 percentage point increase in accuracy. The ACCnf increased to 53 percentage points at the cost of a slight decrease in ACCf, which was 93 percentage points. The adapted classifier outperformed the refined classifier in accuracy by 15 percentage points; however, as mentioned above, it had an advantage over the previous classifiers, because it obtained examples of both straightforward and complex events during learning, and the examples came from the concrete user on which the tests were made. The adapted classifier had the highest ACCnf, 85 percentage points, whereas its ACCf had 71 percentage points. The tests concerning F-measure, which compensates for uneven class distribution, also confirmed the performance improvement. 17 Table 2 Classifier comparison using accuracy on the fall examples (Accf), accuracy on the non-fall examples (ACCnf), overall accuracy (ACCall) and overall F-measure (FMall). 1) CLASSIFIER M L C D K M L J48 JRip SMO Random Forest Naïve Bayes Initial classifier Refined classifier Adapted classifier Training dataset SF events SF events SF events SF events SF events SF events SF events All events T e st in g d a ta se t S tr a ig h tf o rw a rd e ve n ts o n ly ACCf 1.00 1.00 1.00 1.00 1.00 0.98 0.91 0.71 ACCnf 0.68 0.68 0.68 0.70 0.30 0.82 0.94 0.99 ACCall 0.84 0.84 0.84 0.85 0.65 0.90 0.92 0.85 FMall 0.86 0.86 0.86 0.87 0.74 0.91 0.92 0.83 C o m p le x e ve n ts o n ly ACCf 1.00 1.00 1.00 1.00 1.00 0.98 0.96 0.72 ACCnf 0.15 0.12 0.17 0.14 0.04 0.34 0.39 0.81 ACCall 0.37 0.34 0.38 0.36 0.28 0.50 0.53 0.79 FMall 0.44 0.43 0.45 0.44 0.41 0.49 0.51 0.63 A ll e ve n ts ACCf 1.00 1.00 1.00 1.00 1.00 0.98 0.93 0.71 ACCnf 0.28 0.26 0.30 0.28 0.10 0.46 0.53 0.85 ACCall 0.52 0.51 0.53 0.52 0.40 0.63 0.66 0.81 FMall 0.58 0.57 0.59 0.58 0.53 0.64 0.65 0.71 Table 3 compares the performance of the induced classifiers on each event separately. As mentioned above, the ML classifiers detected all fall events; however, they performed poorly on all non-fall events. Introducing domain knowledge to the initial classifier significantly improved the ACCe on the normal behaviour non-fall event. The refinement improved ACCe on the non-fall event searching on the ground. This event was included in the training data for the refinement phase, so increased performance was expected; it was achieved at the cost of neglecting certain fall events. The adapted classifier correctly recognized almost all falls after which the user lay on the ground, but it had difficulties with falls after which the user sat on the ground. Sitting on the ground is a rare event in real life. Sitting on a low chair, an event for which ACCe significantly increased, is a much more common real life event. The classifier frequently confused these two activities for one another. Not only is the user’s posture similar, but they can both last a long time, during which the user is immovable. Some examples of sitting on a low chair are in fact undistinguishable from falls because of the noise in the measurements of the sensors used. Adapting the fall detection classifier establishes a tradeoff between these events. As sitting on a low chair is far more frequent then falls after which a user sits on the ground in a normal sitting position, 1) The abbreviation SF stands for straightforward events. 18 misclassifications of this event are more costly. The adapted classifier is thus inclined to reduce misclassifications during sitting on a low chair at the cost of not detecting certain falls after which the user lands sitting on the ground. In any case, user immovability after falls for additional or prolonged time should enable detecting these false negatives; however, because of practical reasons, we could not capture this in the recordings. Table 3 Classifier comparison using the accuracy on each event (ACCe) in the test scenario CLASSIFIER/ ACCe S T R A I G H T F O R W A R D T E S T S C O M P L E X T E S T S FALLS NON-FALLS FALLS NON-FALLS Tripping (1) Falling landing sitting (2) Normal behavi- our (3) Searching on the ground (4) Falling slowly (5) Falling slowly landing sitting (6) Lying/ Sitting down quickly (7) Sitting on low chair(8) -avg. value M a ch in e L e a rn in g J48 1.00 1.00 0.68 0.68 1.00 1.00 0.64 0.06 JRip 1.00 1.00 0.76 0.60 1.00 1.00 0.60 0.02 SMO 1.00 1.00 0.76 0.60 1.00 1.00 0.88 0.03 Random Forest 1.00 1.00 0.76 0.64 1.00 1.00 0.76 0.02 Naïve Bayes 1.00 1.00 0.12 0.44 1.00 1.00 0.20 0.01 Initial classifier 1.00 0.96 0.96 0.68 0.96 1.00 0.96 0.22 Refined classifier 0.96 0.86 0.96 0.92 0.96 0.96 1.00 0.27 Adapted classifier 0.96 0.46 1.00 0.98 0.84 0.60 1.00 0.77 5. CONCLUSION We presented the CDKML – method for combining DK and ML for classifier generation and online adaptation. The method has three phases: initialization, refinement, and online adaptation. The domain expert specifies the initial CDKML classifier in the first phase. The expert may also use hypotheses induced by ML as help. In the second phase, genetic algorithms adjust the initial classifier to suit system- and general-user characteristics. Tests show that the classifiers developed after the first two phases are already more reliable and robust than ML classifiers built from limited examples from the domain of interest. In addition to general classifier generation, in the third phase, we presented a method for 19 online classifier adaptation based on specific user feedback indicating incorrect system output. Tests show that, during online adaptation, the classifier is adjusted to correctly recognize events not present in the training dataset, making tradeoffs between contradictory examples based on the cost of each misclassification. The method is suitable for domains that have limited obtainable training data and available domain knowledge. Online adaptation is essential if specific characteristics of the objects of interest must be accommodated; the user must be able to give feedback about misclassifications to allow the online adaptation. One such domain is fall detection, on which we evaluated the method. The method is suitable for modelling behaviour in general and for studies in the medical or biological fields, for which large, representative training datasets are typically difficult to obtain. ACKOWLEDGEMENTS This research is partly financed by the European Union, European Social Fund and partly by the European Community's Framework Programme FP7/2007–2013 under grant agreement No. 214986. References BENYON, D. (2001) The new HCI? navigation of information space, Knowledge-Based Systems, 14(8), 425- 430. BRAMER, M. (2005) Inducer: a public domain workbench for data mining, International Journal of Systems Science, 36(14), 909-919. BROWN, M., GRUNDY, W., LIN, D., CRISTIANINI, N., SUGNET, C., FUREY, T., et al. (2000) Knowledge-based analysis of microarray gene expression data by using support vector machines, Proceedings of the National Academy of Sciences, 97, 262-267. BURNS, B. D., and DANYLUK, A. P. (2000) Feature Selection vs Theory Reformulation: A Study of Genetic Refinement of Knowledge-based Neural Networks, Machine Learning - Special issue on multistrategy learning, 89-107. CHEN, M.-C., CHAO, C.-M., and WU, K.-T. (2011) Pattern filtering and classification for market basket analysis with profit-based measures. Expert Systems. CONFIDENCE. (2011) Confidence, Retrieved August 23, 2011, from http://www.confidence-eu.org/ DAVIDSON, I., and RAVI, S. S. (2005) Hierarchical clustering with constraints: Theory and practice, Proceedings of the Nineth European Principles and Practice of KDD (PKDD), 59-70. DECOSTE, D., and SCHOLKOPF, B. (2002) Training invariant support vector machines machine learning, Machine Learning, 46, 161-190. EIBEN, A. E., & SMITH, J. E. (2003) Introduction to Evolutionary Computing, Springer-Verlag. 20 EUROSTAT. (2011) Eurostat, Retrieved August 23, 2011, from http://epp.eurostat.ec.europa.eu/tgm/table.do?tab=table&init=1&language=en&pcode=tsdde511&plugi n=1 FELDMAN, R. S. (1993) Understanding Psychology, New York: Mc Graw-Hill. HALL, M., FRANK, E., HOLMES, G., PFAHRINGER, B., REUTEMANN, P., and WITTEN, I. H. (2009) The WEKA Data Mining Software: An Update, SIGKDD Explorations, 11(1) , 10-18. HEIT, E. (2000) Background Knowledge and Models of Categorization, In U. Hahn, & M. Ramscar, Similarity and Categorization, 155-178. KALUŽA, B., MIRCHEVSKA, V., DOVGAN, E., LUŠTREK, M., and GAMS, M. (2010) An agent-based approach to care in independent living, Proceedings of Ambient Intelligence, 177-186. KAMBAR, S. (2005) Generating synthetic data by morphing transformation for handwritten numeral recognition (with ν-svm) (Master thesis). LI, D.-C., YEH, C.-W., TSAI, T.-i., FANG, Y.-H., and HU, S. C. (2007) Acquiring knowledge with limited experience, Expert Systems, 24(3), 162-169. LUŠTREK, M., GJORESKI, H., KOZINA, S., CVETKOVIĆ, B., MIRCHEVSKA, V., and GAMS, M. (2011) Detecting Falls with Location Sensors and Accelerometers, Proceedings of Innovative Applications of Artificial intelligence. MIRCHEVSKA, V., KALUŽA, B., LUŠTREK, M., and GAMS, M. (2010) Real-Time Alarm Model Adaptation Based on User Feedback, Proceedings of Workshop on Ubiquitous Data Mining in ECAI 2010 , 39-43. MIRCHEVSKA, V., LUŠTREK, M., VELEZ, I., VEGA, N. G., and GAMS, M. (2009) Classifying Posture Based on Location of Radio Tags. Ambient Intelligence Perspectives II - Selected papers from the Second International Ambient Intelligence Forum 2009, 85-92. NIYOGI, P., GIROSI, F., and POGGIO, T. (1998) Incorporating prior information in machine learning by creating virtual examples, Proceedings of the IEEE, 2196-2209. OSEI-BRYSON, K.-M. (2004) Evaluation of decision trees: a multi-criteria approach, Computers and Operations Research, 1933-1945. PAZZANI, M., BRUNK, C., and SILVERSTEIN, G. (1991) A knowledge-intensive approach to learning relational concepts, The Eighth International Workshop on Machine Learning, 432-436. POGGIO, T., and VETTER, T. (1992) Recognition and structure from one 2D model view: Observations on prototypes, object classes and symmetries, Tech. Rep. AIM-1347. Cambridge, MA, USA: Massachusetts Institute of Technology. RUSSELL, S., & NORVIG, P. (2003) Artificial intelligence A Modern Approach, Prentice Hall. 21 SABZEKAR, M., YAZDI, H. S., & NAGHIBZADEH, M. (2011) Relaxed Constraints Support Vector Machine, Expert Systems. STUMPF, S., RAJARAM, V., LI, L., WONG, W.-K., BURNETT, M., DIETTERICH, T., et al. (2009) Interacting meaningfully with machine systems: Three experiments, International journal of human-computer studies, 67, 639-662. SUN, S., and HARDOON, D. R. (2010) Active learning with extremely sparse labeled examples, Neurocomputing, 73, 2980-2988. THRUN, S. (1996) Explanation-Based Neural Network Learning: A Lifelong Learning Approach. Boston, MA: Kluwer Academic Publishers. UBISENSE (2011) Ubisense, Retrieved August 23, 2011, from http://www.ubisense.net/ VIDULIN, V., and GAMS, M. (2011) Impact of higher-level knowledge on economic welfare through interactive data mining, Aplied artificial intelligence, 25(4), 267-291. WANG, L., XUE, P., and CHAN, K. L. (2004) Incorporating prior knowledge into SVM for image retrieval, Proceedings of the 17th International Conference on Pattern Recognition, 981-984. WISNIEWSKI, E. J., and MEDIN, D. L. (1994) On the Interaction of Theory and Data in Concept Learning, Cognitive science, 18(2), 221-281. WU, X., and SRIHARI, R. (2004) Incorporating prior knowledge with weighted margin support vector machines, Proceedings of the tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 326-333. YANG, T., and KECMAN, V. (2009) Adaptive local hyperplane algorithm for learning small medical data sets, Expert systems, 26(4), 355-359. YU, T. (2007) Incorporating Prior Domain Knowledge into Inductive Machine Learning Its implementation in contemporary capital markets (PhD thesis). ZHU, Z., and LIU, P. (2010) Feasibility research of text information filtering based on genetic algorithm, Scientific Research and Essays, 5(22) , 3405-3410. The authors Violeta Mirchevska Violeta Mirchevska is a researcher at Result d.o.o., cooperating closely with the Department of Intelligent Systems at the Jožef Stefan Institute. She completed her Bachelor's in computer science and automation at the Faculty of Electrical Engineering and Information Technologies, Ss. Cyril and 22 Methodius University, Macedonia in 2007 and is currently a PhD candidate at Jožef Stefan International Postgraduate School, Slovenia. Her research focuses on modelling agent behaviour based on observing low-level action sequences by leveraging both domain knowledge and machine learning. The main application areas of her research are user profiling, remote health monitoring and security. Mitja Luštrek Dr. Mitja Luštrek is a researcher at the Jožef Stefan Institute in Slovenia. He is the head of the ambient intelligence group at the Department of Intelligent Systems. His main research area is ambient intelligence with a focus on human behavior analysis. He has experience with wearable inertial sensors and real-time locating systems, and has used artificial intelligence techniques for activity recognition and detection of anomalous behaviors. He has also worked in several other areas of artificial intelligence, including game playing, heuristic search, and machine learning in bioinformatics. He is an editor of the Informatica journal and a member of the executive board of the Slovenian Artificial Intelligence Society. Matjaž Gams Prof. dr. Matjaž Gams (http://dis.ijs.si/Mezi/) is a senior researcher at the Jožef Stefan Institute, Ljubljana, Slovenia. His research interests include artificial intelligence, intelligent systems, intelligent agents, machine learning, cognitive sciences, and information society. His publication list includes 500 items, 70 in scientific journals. Matjaž Gams is heading the Department of Intelligent Systems at the Jožef Stefan Institute. He is currently president of ACM Slovenia and the cofounder of the Engineering Academy, Artificial Intelligence Society, and Cognitive Sciences Society in Slovenia. He is an executive contact editor of the journal Informatica and member of the editorial board of several international journals. He headed several major applications in Slovenia, including a virtual agent for the Slovenian employment agency, an expert system controlling the quality of nearly all national steel production and a text-to-speech system in Slovenian, donated to several thousand users. Matjaž Gams also teaches several courses in computer sciences at the graduate and postgraduate levels. http://dis.ijs.si/Mezi/ 
work_vvejlao4xbbttpydvf3tbow73m	----	A semantic framework for textual data enrichment Accepted Manuscript A semantic framework for textual data enrichment Yoan Gutiérrez , Sonia Vázquez , Andrés Montoyo PII: S0957-4174(16)30143-9 DOI: 10.1016/j.eswa.2016.03.048 Reference: ESWA 10617 To appear in: Expert Systems With Applications Received date: 23 December 2015 Revised date: 29 February 2016 Accepted date: 25 March 2016 Please cite this article as: Yoan Gutiérrez , Sonia Vázquez , Andrés Montoyo , A semantic framework for textual data enrichment, Expert Systems With Applications (2016), doi: 10.1016/j.eswa.2016.03.048 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. http://dx.doi.org/10.1016/j.eswa.2016.03.048 http://dx.doi.org/10.1016/j.eswa.2016.03.048 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Highlights  A semantic framework for recommender systems is presented  An in-depth analysis of different Natural Language Processing resources is showed  A description of different Natural Language Processing approaches is addressed  Related research works are described  A case of study to evaluate our proposal with real data is presented ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT A semantic framework for textual data enrichment Yoan Gutiérrez, Sonia Vázquez * and Andrés Montoyo Department of Software and Computing Systems. University of Alicante, Spain. ygutierrez, svazquez, montoyo@dlsi.ua.es Abstract: In this work we present a semantic framework suitable of being used as support tool for recommender systems. Our purpose is to use the semantic information provided by a set of integrated resources to enrich texts by conducting different NLP tasks: WSD, domain classification, semantic similarities and sentiment analysis. After obtaining the textual semantic enrichment we would be able to recommend similar content or even to rate texts according to different dimensions. First of all, we describe the main characteristics of the semantic integrated resources with an exhaustive evaluation. Next, we demonstrate the usefulness of our resource in different NLP tasks and campaigns. Moreover, we present a combination of different NLP approaches that provide enough knowledge for being used as support tool for recommender systems. Finally, we illustrate a case of study with information related to movies and TV series to demonstrate that our framework works properly. Keywords: Recommender Systems, Framework, Integrated Semantic Resources, Sentiment Analysis, Word Sense Disambiguation, Content Categorization 1. Introduction Recent advances in modern technologies have motivated the development of different techniques to improve human-machine communication. Internet and new communication tendencies such as: short messages, forum participations, social networks, etc, have led to a revolution in the way in which people work, communicate and manage their free time. As a consequence of this technological revolution, a huge quantity of information is generated in different social contexts via diverse sources such as: forums, blogs, microblogs, social networks, etc. As a result, people are able to share their knowledge, expectations and emotions through Internet and they may also influence political, economic or social behaviour. At this point, governments, enterprises or even celebrities need to manage this information in order to extract relevant knowledge, social tendencies, etc. Because of this new context, research community in Natural Language Processing (NLP) have developed different tools with which to analyse news and opinions in order to discover what people think or how they perceive past, present and future. At present, personalization and recommender systems have gained popularity. In fact, recommender systems began to appear in the market in 1996 (Udi et al., 2000). Since then, several approaches have been developed (Gediminas and Alexander, 2005):  Content-based: these systems try to find products, services or contents that are similar to those already evaluated by the user. In this kind of systems, user‟s feedback (that can be collected in many ways) are essential to support and accomplish recommendations (Marco de et al., 2008).  Knowledge-based: these systems model the user profile in order to, through inference algorithms, identify the correlation between their preferences and existing products, services or content. (Walter et al., 2012)  Collaborative filtering: these systems create/classify groups of users that share similar profiles/behaviours in order to recommend products, services or content that has been well evaluated by the group to which a user belongs (Perner et al., 2007).  Hybrid: these systems combine two or more techniques previously mentioned to improve the „„quality‟‟ of recommendations (Shinde and Kulkarni, 2012). Dealing with textual information and obtaining valuable knowledge require advanced natural language techniques to solve different kinds of problems: document correction, automatic translation, ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT summary elaboration, opinion extraction, word sense disambiguation, etc. Solving all of these problems requires a considerable linguistic knowledge and, even more importantly, a high computational cost. In the vast majority of tasks in NLP it is necessary to use external resources such as: Machine- readable dictionaries 1 , Thesaurus 2 , Ontologies 3 and others. These resources have different internal structures, interfaces, concept relations and other characteristics. One of the most frequently used resources in its different versions is WordNet 4 (WN) (Miller et al., 1990). Various semantic resources related to WordNet have consequently been developed in different domains or by using semantic integration. But it is still difficult to find resources that provide semantic integration in different domains and which are useful for specific NLP tasks. In this work we present a new semantic resource (ISR-WN) and a set of different methods to take advantage of it with the aim of enriching texts with semantic information. As a result, we provide a semantic framework suitable of being used as support tool for content-based recommender systems by annotating texts with different features such as: sentiments, polarities or domain labels. In order to analyse the results of the semantic enrichment process, we have carried out a comprehensive case of study using texts from movies and TV series reviews obtained from IMDb 5 . Finally, we have evaluated how our proposed framework works comparing our results with real ratings. To summarize, we point out the main contributions of this work:  Taking advantage of a semantic resource with different dimensions previously developed (ISR-WN)  The use of a set of NLP methods based on ISR-WN to take advantage of each one of its semantic dimensions  Providing a new semantic framework that is able to enrich texts in several dimensions with the aim of obtaining a support tool for content-based recommender systems  An exhaustive evaluation with real datasets to demonstrate how it works The document is structured as follows. After this introduction, each semantic resource used in ISR- WN is described, and an in-depth analysis of the different approaches for semantic integration resources in NLP is also presented. Having evaluated previous proposals, in Section 3 we go on to show how ISR- WN was developed. An evaluation according to its integration effectiveness is then provided in Section 4. In Section 5 we provide a brief description of the different NLP tasks selected to enrich texts. Section 6 describes the characteristics of a case of study to illustrate how our framework works with real data obtained from IMDb. In Section 7 we show some examples of how the semantic enrichment approaches are used to annotate texts. Section 8 provides the experiment results of the case of study and Section 9 presents a discussion about the results obtained. Finally, the conclusions and further works are presented in Section 10. 2. Related Work This section presents the different semantic resources that are integrated into ISR-WN and a comparison with other semantic integration resources. 2.1 WordNet As mentioned in the previous section, WN is one of the most frequently used semantic resources in computational linguistics (Navigli, 2009). WN is a lexical database for the English language also considered as ontology. It was created at the University of Princeton 6 and it represents a semantic conceptual and structured network of nouns, verbs, adjectives and adverbs. The basic unit of knowledge is the synset (synonym sets), which represents a lexical concept (Ševčenko, 2003). A synset is associated with a unique eight-digit number called an offset (this number is the position in the data file). Each synset * Corresponding author. Tel. +34 965 90 37 72; Fax: +34 965 90 93 26. E-mail address: svazquez@dlsi.ua.es 1 Dictionaries of words available in electronic format. 2 Provides relationships among words (i.e., synonyms, antonyms and others). 3 Conceptualization of a domain in order to share information among different agents. 4 http://wordnet.princeton.edu/ 5 http://www.imdb.com 6 http://wordnet.princeton.edu/wordnet/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT represents different senses which are related through the use of semantic, conceptual or lexical connections. The result of this set of connections is a wide navigable network with a high number of interrelations among different word senses. The semantic relations among synsets are:  Synonymy  Antonymy  Hyponymy / Hyperonymy  Meronymy / Holonymy  Entailment and Cause,  and others…(more details in 7 ) WN establishes the frequency of usage of each word sense (synset)in its internal relationships 8 . For example, the word image has eight senses in WN 2.0 (see Table 1). As we can observe, one word has different senses, there is a sentence (gloss) which describes each one and each sense has a set of synonyms that are ordered by their frequency of usage. POS Offset Lemmas Gloss Noun 00039630 effigy, image, simulacrum a representation of a person (especially in the form of sculpture); "the coin bears an effigy of Lincoln"; "the emperor's tomb had his image carved in stone" Noun 00043454 Picture ,image, icon, ikon a visual representation (of an object or scene or person or abstraction) produced on a surface; "they showed us the pictures of their wedding"; "a movie is a series of images projected so rapidly that the eye integrates them" Noun 00047709 persona, image (Jungian psychology) a personal facade that one presents to the world; "a public image is as fragile as Humpty Dumpty" Noun 00053779 image, mental_image an iconic mental representation; "her imagination forced images upon her too awful to contemplate" Noun 00053832 prototype, paradigm, epitome, image a standard or typical example; "he is the prototype of good breeding"; "he provided America with an image of the good father" Noun 00059547 trope, figure_of_speech, figure, image language used in a figurative or non-literal sense Noun 00074537 double, image, look-alike someone who closely resembles a famous person (especially an actor); "he could be Gingrich's double"; "she's the very image of her mother Verb 00109926 visualize, visualize, envision, project, fancy, see, figure, picture, image imagine; conceive of; see in one's mind; "I can't see him on horseback!"; "I can see what will happen"; "I can see a risk in this strategy" Table 1. Word senses of “image” in WN 2.0 It is important to emphasize that WN has been adapted to different languages: English, Spanish, Dutch, Italian, German, French, Czech, Estonian, Swedish, Norwegian, Danish, Greek, Portuguese, Basque, Catalan, Romanian, Lithuanian, Russian, Bulgarian, Slovenian and others that are under development. These versions have been developed under the supervision of the University of Princeton and later under that of the Global WordNet Association 9 . This research work is based on two versions of WN: WN 1.6 with 99,643 synsets, of which 66,025 are nouns, 17,915 are adjectives, 3,575 are adverbs and 12,127 are verbs, and WN 2.0 with 115,424 synsets, of which 79,689 are nouns, 18,563 are adjectives, 3,664 are adverbs and 13,508 are verbs. 2.2 Semantic resources aligned to WordNet 7 http://wordnet.princeton.edu/man/wninput.5WN.html 8 https://wordnet.princeton.edu/man/cntlist.5WN.html 9 http://www.globalwordnet.org/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Owing to the fact that WN has been used in many NLP research works, a set of different semantic resources aligned to WN synsets has been developed with the aim of obtaining more knowledge. Some of these resources were created from WN, such as: WordNet Domains 10 (Magnini and Cavaglia, 2000), WordNet Affect 11 (Magnini and Cavaglia, 2000, Sara and Daniele, 2009) and Semantic Classes 12 (Izquierdo et al., 2007). Others emerged from the association of pre-produced tags, i.e., SUMO 13 . The resources used in our proposed semantic integration resource (ISR-WN) are described in detail below. 2.2.1 WordNet Domains This is a resource for the English language. WordNet Domains (WND) includes a set of Subject Field Codes (SFC) (Magnini and Cavaglia, 2000) with which to enrich WN synsets. Each SFC groups a set of words related to the same domain. On the one hand, these domains identify the context of the definition and on the other, they allow a quick search of concepts to take place. For example, if we are searching for the meaning of disc in the Computer Science context, we need only check the domain label preceding each definition (in this case, Computer Science) until we find the correct definition ([magnetic_disc, magnetic_disc, disck, disk]). This resource therefore provides the integration of domain labels in WN with the aim of reducing WN sense granularity by grouping different senses in the same domain or semantic category. In WND, WN has been annotated by using a semi-automatic process that assigns one or more domain labels to each synset. Domain labels are selected from a set of 200 hierarchically organized labels. In this research we use 172 of these labels. The main purposes of annotating WN with SFC are:  To create new relations among words. Domain labels allow us to relate words that pertain to different grammatical categories.  Semantic annotation. Domain labels are associated with synsets, signifying that the annotation takes place at the semantic level rather than at the word level.  Synsets pertaining to different syntactic categories can be included in the same domain label.  Word senses pertaining to different sub-hierarchies of WN can be included in the same domain label.  To reduce word sense granularity. Grouping different senses of the same word in the same domain label reduces the polysemy of words. Sense Domain Gloss man#1 person an adult male person (as opposed to a woman); “there were two women and six men on the bus” man#2 military someone who serves in the armed forces; “two men stood sentry duty” man#3 person the generic use of the word to refer to any human being; “it was every man for himself” man#4 factotum all of the inhabitants of the earth; “all the world loves a lover” man#5 biology, person any living or extinct member of the family Hominidae man#6 person a male subordinate; “the chief stationed two men outside the building”; “he awaited word from his man in Havana” man#7 person an adult male person who has a manly character(virile and courageous competent); “the army will make a man of you” man#8 person (informal) a male person who plays a significant role (husband or lover or boyfriend) in the life of a particular woman; “she takes good care of her man” man#9 person a manservant who acts as a personal attendant to his employer; “Jeeves was Bertie Wooster’s man” man#10 play a small object used in playing certain board games; “he taught me to set up the men on the chess board”; “he sacrificed a piece to get a strategic advantage” Table 2. Domain labels and senses for man 10 http://wndomains.fbk.eu/ 11 http://wndomains.fbk.eu/wnaffect.html 12 http://rua.ua.es/dspace/bitstream/10045/2522/1/ranlp07BLC2.pdf 13 http://www.ontologyportal.org/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT For example, the word man has ten senses in WN (see Table 2). However, in WND we can group different senses according to their domain labels, and the number of senses is thus reduced from ten to four (Magnini et al., July 2002). Moreover, there are different specification levels in the domain hierarchy. The deeper the level, the greater the specialisation is. Fig 1 shows a brief excerpt of the WND hierarchy from (Luisa Bentivogli, 2005). Fig 1. WordNet Domains hierarchy Among all the SFC domain labels there is a special one called Factotum. This domain label was created in order to group two types of synsets (Magnini and Cavaglia, 2000):  Generic Synsets. Those that it is difficult to classify into a particular domain label.  Stop senses. Those that frequently appear in different contexts, such as: numbers, days of the week, colours, etc. 2.2.2 WordNet Affect WordNet Affect (WNA) is an extension of WND (Magnini and Cavaglia, 2000, Sara and Daniele, 2009). It contains different subsets of affective concepts that group together synsets which denote emotional states. This resource was labelled by following a similar process to that of WND. Some of the represented concepts are: moods, situations eliciting emotions or emotional responses. This resource was extended with a set of additional labels called emotional categories. It has a hierarchical structure in which hyperonymy is used to relate the affective concepts of WN (Valitutti et al., 2004). In a second revision, some modifications were made in order to differentiate those senses that are closer to emotional labels, and new labels were also included: positive, negative, ambiguous and neutral:  The first pertains to positive emotions. For example, it includes synsets such as: joy#1 or enthusiasm#1.  Negative defines negative states such as: anger#1 o sadness#1.  Ambiguous represents synsets whose semantics depends on the contexts in which they appear: surprise#1  Neutral represents synsets that refer to mental states but which are not characterised by valence. One important property of WNA labels is that they associate the nouns and adjectives involved in emotional states. In this case, the adjective modifies the state of the noun, and may in some situations determine the modified noun state, i.e.: cheerful / happy boy (Strapparava and Valitutti, 2004). In other words, if the adjective pertains to an emotional state it could indicate how the noun is related. Table 3 shows a list of affective labels associated with synsets. 300 labels are integrated into our research work. Doctrines Archaeology Astrology History Linguistic Psychology Art Religion Heraldry Grammar Psychoanalysis Music Dance Drawing Photography Plastic_arts Theatre Mythology Occultism Roman_catholic Theology Painting Philately Jewelry Numismatic Sculpture ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Affective labels Examples emotion noun anger#1, verb fear#1 mood noun animosisy#1, adjective amiable#1 trait noun aggressiveness#1, adjective competitive#1 cognitive state noun confusion#2, adjective dazed#2 physical state noun illness#1, adjective all in#1 hedonic signal noun hurt#3, noun suffering#4 emotion-eliciting situation noun awkwardness#3, adjective out of danger#1 emotional response noun cold sweat#1, verb tremble#2 behaviour noun offense#1, adjective inhibited#1 attitude noun intolerance#1, noun defensive#1 sensation noun coldness#1, verb feel#3 Table 3. WordNet Affects labels and their associated synsets. 2.2.3 SUMO SUMO 14 (Suggested Upper Merged Ontology) is considered to be an upper level ontology. It provides definitions for general terms and can be used as a basis for domain specific ontologies. It was created from the combination of different ontological contents in one single cohesive structure. It currently contains around 1,000 terms and 4,000 assertions (Niles and Pease, 2003). Our research work only uses 568 of the concepts that are aligned with WN. SUMO was obtained from the information of: Ontolingua 15 and the developed ontologies of ITBM- CNR 16 (Unrestricted-Time, Representation, Anatomy, Biologic-Functions, and Biologic-Substances). It uses a standard language representation called SUO-KIF (Pease, 2007) obtained from KIF (Knowledge Interchange Format) (Genesereth and Fikes, 1992). SUMO was built by dividing the concepts into two groups: high level concepts and low level concepts. In the first group, the ontology of John Sowa (Sowa, 1999), and Russell and Norvig (Russell and Norvig, 1994) were considered, while the remaining concepts were included in the second group. Finally, a unique conceptual structure combining the two high level ontologies was created. The remaining low level class contents were included after the combination. Fig 2 shows the high level concepts. Fig 2. SUMO high level concepts (Ševčenko, 2003). The higher level concept is the entity category, as occurs in most hierarchies. The entity concept groups the rest of the concepts, and the physical and abstract concepts are closest to it. An example of the word sense bank#1 is shown in Fig 3, along with the SUMO hierarchy. 14 http://suo.ieee.org/SUO/SUMO/index.html 15 http://www.ksl.stanford.edu/software/ontolingua/ 16 http://www.ontologyportal.org/SUMOhistory/ Entity Abstract Physical Quantity Attribute SetOrClass Proposition Relation ProcessObject ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Fig 3. SUMO hierarchy for bank#1. 2.2.4 Semantic classes A semantic class is a sense conceptualisation that can be manually or semi automatically created from different abstraction levels and in different domains. WN is composed of a set of related and connected synsets with different semantic relations. Each of these synsets represents a concept and contains a set of words referring to the concept it describes (synonyms). Synsets are classified into forty-five groups which include lexical categories (nouns, verbs, adjectives and adverbs) and semantic groupings (person, phenomenon, sentiment, place, etc.) (Fellbaum, 1998). There are twenty six categories for nouns, fifteen for verbs, three for adjectives and one for adverbs in WN (Izquierdo, 2010). The organisational design of WN helps lexicographers to obtain a structure with which to create and edit a set of words and senses in the same semantic class. These semantic categories are considered to be Semantic Classes that are more general than senses, in which different WN senses are grouped in a semantic class. It is also possible to use this sense conceptualisation in different languages because all the WN senses (in English) are linked to EuroWordNet which contains WNs in different languages. It is important to note that despite the fact that WN was generated by following a top-down process, it is difficult to distinguish between those synsets that were originally semantic classes and those that were not. If we wish to obtain the sense categorisation it is therefore necessary to apply a method with which to generate the Semantic Classes. The main goal of Semantic Class (SC) generation (Izquierdo et al., 2007) is to reduce the polysemy, and various techniques with which to group senses have therefore been developed. In all cases, senses of the same word have been grouped together, thus reducing polysemy and improving WSD system results. The SC resource consists of a set of Base Level Concepts (BLC) obtained from WN using a bottom- up process with hyperonymy relations. For each synset of WN, its BLC is obtained from the first local maxim according to its relative number of relations. As a result, semantic classes have a set of BLCs that are linked to different synsets. The process follows an ascendant itinerary by using the hyperonymy relations in WN. In the case of one synset having several hyperonyms, the path with the maximum number of relations is selected. The process ends when a set of initial concepts with the synsets selected as BLCs for other synsets is selected. In some cases, there are BLCs that do not represent an adequate number of concepts or synsets. This situation is avoided by developing a final filtering process for these false BLCs, in which those BLCs that do not represent a minimum number of concepts are eliminated. Each BLC therefore has a minimum threshold associated with it, as is described in (Izquierdo et al., 2010). A set of different BLCs is obtained by combining different thresholds and types of relations (all or only hypo/hyperonyms). Those synsets that do not have a BLC associated with them (because their numbers of relations do not have the minimum threshold required) are processed again to select another BLC from their hyperonymy relations. This eventually results in a new number of labels with which to categorise a set of senses. SCs are applied with the use of different repositories. Those currently being used are WN1.6 and WN2.0. depository fina ncia l institution, ba nk, ba nking concern, ba nking compa ny Object Physical Entity Collection Agent Corporation Organization Group ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 2.2.5 SentiWordnet SentiWordNet 17 (SWN) (Esuli and Sebastiani, 2006, Baccianella et al., 2010) is a lexical resource in which each synset is associated with three different sentiment categories: objectivity, positivity and negativity. Each category has a score that moves from [0..1] and the addition of the three scores is always 1. This means that one synset could have scores that are different from 0 for each of the categories. For example, atrocious#3 In our proposal we use SentiWordNet 3.0. This version was obtained in two steps: first a semi- supervised learning phase and then a random walk. The first step consists of four sub-steps, as in the first version of SWN:  Semi-supervised learning o First sub-step. Two small sets of seeds are used, one in which all the synsets contain seven paradigmatically positive terms and the other in which all the synsets contain seven paradigmatically negative terms (Turney and Littman, 2003). Both are expanded by using WN binary relations in order to connect synsets with the same polarity. The expansion is carried out with a specific k radius. o Second sub-step. A previous set of synsets is used with another set of synsets from the Objectivity category in order to build a set of training synsets with which to create a ternary classifier (one synset is classified as Pos, Neg or Obj). The classifier uses synset glosses to conduct the process. In SWN 1.0, a bag of words model is used in which the bag of words is obtained from gloss words (those frequent words that are most important). In SWN 3.0, rather than using words from glosses (with the ambiguity problem) a bag of synsets is used. This sub-step is improved by varying the displacement radius. o Third sub-step. All WN synsets are classified as being either Pos, Neg or Obj via the classifier generated in the second sub-step. o Fourth sub-step. The second sub-step can be performed by using different values of the k radius and different supervised learning technologies.  Step 2 (random walk). WN 3.0 is considered to be a graph and an iterative process is conducted. In this random walk Pos(s) and Neg(s) (and consequently Obj(s)) are determined using the previous step. The process ends when the iterations have converged. There are 117,659 WN synset descriptions in our proposal. 2.2.6 eXtendedWordNet eXtendedWordNet 18 (XWN) (Sanda M. Harabagiu, 1999) is a lexical resource created at the University of Texas. This resource was developed to improve semantic information in WN in its different versions. The goal is to add semantic information to glosses and establish new relations among words (now labelled with their senses) from glosses and synsets. The new annotated version of WN only uses information from gloss definitions, and gloss examples and other information are discarded. eXtendedWordNet was created by applying three processes:  Syntactic analysis. A voting process with two syntactic analysers was applied using the outputs of (Brill, 1995) as inputs. The content of the glosses was extended with: o Adverbs. Glosses of adverbs were extended by adding an adverb + is at the beginning of the gloss and a full stop at the end of the definition. For example, the gloss for automatically would be: automatically is in a reflex manner. A direct semantic annotation is thus obtained between word and gloss. o Adjectives. Glosses of adjectives were extended by adding an adjective + is something at the beginning of the gloss and a full stop at the end of the definition. For example, the gloss for pure would be: pure is something not mixed. o Verbs. Glosses of verbs were extended by adding to + verb + is to at the beginning of the gloss and a full stop at the end of the definition. For example, the gloss for shed would be: to shed is to cast off hair, skin, horn, or feathers. 17 http://gandalf.aksis.uib.no/lrec2006/pdf/384_pdf.pdf 18 http://xwn.hlt.utdallas.edu/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT o Nouns. Glosses of nouns were extended by adding noun + is at the beginning of the gloss and a full stop at the end of the definition. For example, the gloss for play would be: play is the act using a sword (or other weapon) vigorously and skilfully.  Logical analysis. After applying a syntactic analysis, a logical form transformation is then applied. Syntactic relations, syntactic objects, prepositional links, complex nominalizations, etc.  Semantic analysis. Two versions were created: automatic and manual. Automatic annotation was applied by using one specific system to disambiguate WN glosses (XWN WSD) and another system to disambiguate free text. The decision process was a voting system that selected those senses with a coincidence between both systems with a precision of 90%. There were three different categories for semantic annotation (the verbs to be and to have were managed in a special way): o GOLD: manual annotation. o SILVER: both WSD methods returned the same value. o NORMAL: only uses XWN WSD value. This annotation obtained 100% coverage and 70% precision. Words labelled with the same sense of both systems obtained 90% precision. In our approach 551,551 relations were integrated using XWN1.7 and 419,387 using XWN3.0. 2.3 Semantic integration resources Additional information with which to solve different NLP problems was obtained by using different semantic integration resources (ISR). However, one of the main problems is decentralisation. WN can solve this problem because it has been used as a basic resource to build others. Different resources currently use WN as basis to build their structures. For example:  MultiWordNet 19 (MWN) (Pianta et al., 2002) is a project that is being carried out by the ITC-IRST group in Trento, Italy in order to build an Italian WN that is aligned with the English WN (Princeton). Its first version contains around 37,000 Italian words that are organised in 28,000 synsets and their connections to English synsets. MultiWordNet is different from EuroWordNet because at least two methods can be used to build a multilingual WN. The first method was used in the EuroWordNet project and consists of building the WNs of specific languages independently, while correspondences are found in a second phase (Vossen, 1998). The second method was used in MultiWordNet and consists of building WNs of specific languages from semantic relations in the English WN. New synsets are obtained from English synsets. In MWN, the domain information has been automatically transferred from English to Italian, thus obtaining WND (Bentivogli et al., 2004). EuroWordNet (EWN) (Dorr and Castellón, 1997, Vossen, 1998) was developed to align English, Spanish, Dutch, Italian, German, French, Czech and Estonian. New versions included Swedish, Norwegian, Danish, Greek, Portuguese, Basque, Catalan, Romanian, Lithuanian, Russian, Bulgarian and Slovenian. The ILI (Inter-Lingual-Index) is used to connect each language (Vossen et al., 1999). The use of ILI allows closer senses to be obtained among each language, thus reducing polysemy and obtaining a large connection among different languages.  Meaning: the Multilingual Central Repository (Meaning project MCR) (Atserias et al., 2004) is integrated into the EWN framework with five local WNs, including the English WN, and uses an improved version of Superior Concept ontology of EWN, MWN Domains, SUMO (Suggested Upper Merged Ontology) (Zouaq et al., 2009) and new semantic relations from corpus. The first version of MCR includes only conceptual Knowledge with semantic relations among synsets from local WNs. The last version of MCR integrates: o ILI based on WN1.6, including base concepts from EWN, Superior Concept ontology of EWN, MultiWordNet Domains and SUMO; o Local WNs (Basque, Catalan, Italian and Spanish) related to ILI, including WN English versions 1.5, 1.6, 1.7 and 1.7.1. o Semantic preferences collections, from SemCor and BNC, and nominal entities.  UBY: this is a large-scale lexical-semantic resource for NLP based on the ISO standard Lexical Markup Framework (LMF). UBY combines a wide range of information from expert and collaboratively constructed resources for English and German. It presents a web browser and an API 19 http://multiwordnet.itc.it ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT that provides facilities that allow this tool to be used. UBY currently holds structurally and semantically interoperable versions of ten resources in two languages: o EnglishWordNet, Wiktionary, Wikipedia, FrameNet and VerbNet, o German Wikipedia, Wiktionary, GermaNet and IMSLex-Subcat, and multilingual Omega Wiki.  BabelNet: this is a very large multilingual semantic network with millions of concepts obtained from: o An integration of WordNet and Wikipedia based on an automatic mapping algorithm, and o Translations of the concepts (i.e., English Wikipedia pages and WordNet synsets) based on Wikipedia cross-language links and the output of a machine translation system. As will be noted, these examples attempt to build semantic networks with a common interface. In most cases, ISRs apply lexical integration with a few conceptual resources but in our proposal ISR-WN provides semantic information and other types of relations. ISRs are resources that help to improve the results of tasks such as: document classification, entity discrimination, author detection, etc. The improvement is owing to the fact that ISRs provide the contexts analysed with additional information (for example, subjectivity detection, contextual domain, etc). In this work we present a resource that provides navigability from any word sense in WN, domain labels (and sentiments from WNA), SUMO categories or Semantic Classes through the use of semantic relations, as shown in Fig 4. As a result, we can extract the multidimensionality of each sentence through the use of all the concepts and words related in a semantic network. We can also detect sentiment polarities in each sentence from SWN and relate them with WND, WNA, SUMO and SC. Fig 4. Sentence semantic characteristics extraction (only a few labels). After using the WN interface to study different lexical and conceptual resources, our goal is to develop a tool with which to align different WN-based resources and exploit all their relations: hyperonymy, meronymy, synonymy, etc. The result will be a graph with a set of nodes that represent WN synsets, WND concepts, SUMO categories, WNA concepts, SC labels or SWN sentiment polarities. Trait Social Science Quality PedagogyFactotum Administration Root Domain Free Time WNA WND SUMO Subjective Assessment Attribute Removing Occupational Role Educational Process Organization Normative Attribute Relational Attribute Attribute Abstract Entity Group Collection Object Physical Social Role Organizational Process International Process Process Transfer Motion use.v.04 be.v.08 exist.v.01 be.v.01 be.v.05 be.v.02 constitute.v.01 be.v.03 equal.v.01 comunicate.v.02 metalic_element.n.01 strike.v.01 travel.v.01 move.v.02 get_rid_of.v.01 sell.v.01 geographical_area.n.01 natural_process.n.01 content.n.01 teacher.n.01 change_of_state.n.01 structure.n.01 cognition.n.01 organization.n.01 gathering.n.01 SC Positive SWN Negative SWN Objetive SWN “But it is unfair to dump on teachers as distinct from the educationalestablishment” ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 3. New integrated resource: ISR-WN Our semantic integration proposal is denominated as ISR-WN (Integrated Semantic Resource aligned with WordNet). It consists of the integration of different resources that are isolated but which can be aligned with WN. Considering that WordNet provides synsets (S) that contain a set of synonym words and WND, WNA, SUMO, SC and SWN provide a set of concepts (C) we can relate each concept with different synsets from WordNet. Therefore, ISR-WN is a Lexical Knowledge Base (LKB) that is represented as a complete non directed graph 20 . Vertexes are represented with concepts or senses , that is, . Relations between two vertexes i, j are represented with an edge . According to the different resources taken into account, we have developed the integration knowledge base of ISR-WN based on the following resources: WN, WND, SUMO, WNA, Semantic Classes (SC), SWN and eXtended WordNet (XWN) 1.7 and 3.0. Note that XWN provides additional semantic relations among WordNet synsets. Bellow, we describe each version and the integration process in each case. 3.1 Integration process The integration process takes into account the above mentioned resources. The aim is to obtain a new environment from which to retrieve semantic information using a unified set of resources. It is necessary to mention that WN is one of the most frequently used resources during NLP, since it provides a sense inventory. Its possibilities have motivated several researchers to develop taxonomies with new semantic information (Magnini and Cavaglia, 2000, Valitutti et al., 2004, Niles and Pease, 2003, Niles, 2001, Sara and Daniele, 2009, Forner, 2005, Strapparava and Valitutti, 2004). With regard to the resulting resources, we can mention SUMO (Niles and Pease, 2001) , WND (Sara and Daniele, 2009), WNA (Strapparava and Valitutti, 2004), SC (Izquierdo et al., 2007), SWN (Esuli and Sebastiani, 2006), eXtended WordNet (Sanda M. Harabagiu, 1999) and others. Several authors such as (Gliozzo et al., 2004, Magnini et al., 2002, Magnini et al., July 2002, Vázquez, 2009, Vázquez et al., 2004) have developed methods and systems based on these resources, and have also demonstrated improvements in several tasks: Information Extraction, Automatic Summarization, Document Indexing and Lexical Disambiguation. However, these authors have developed their approaches using one or two semantic resources, since there is no tool or resource with which to integrate all the semantic resources that are mapped onto WN. Bearing in mind both this fact and researchers‟ current need to use a single tool or resource to obtain semantic information from different resources, this work is focused on integrating the greatest possible quantity of semantic resources mapped onto WN into a single tool. ISR-WN includes WN as its lexical nucleus because its internal structure and relations provide relevant information for many NLP tasks. Fig 5 represents the conceptual model, in which each dimension (resource) is aligned with WN through the use of semantic interconnections (Gutiérrez et al., 2010a) (Gutiérrez et al., 2011a). The main challenge as regards the integration consists of dealing with different versions of WN and different versions of the rest of the resources involved in this proposal. It is therefore necessary to match the mappings to the WN versions used by each resource. See Fig 6. This integration resource involves a set of semantic resources considering the following element distribution: 99,643 synsets for WN1.6 21 and 115,424 synsets for WN2.0 22 ; WND (with 172 labels), WNA (with 300 labels), SUMO (with 568 labels), SWN (with 117,659 labels), SC (with 1,231 labels), and XWN (551,551 relations for XWN1.7 and 19,387 new synset relations for XWN3.0). This is known as “Integration of Semantic Resources based on WordNet” (ISR-WN) (Gutiérrez et al., 2010a) (Gutiérrez et al., 2011a). It is important to note that all the labels and relations included in both versions have been reused from the resources cited. The elements of which ISR-WN is composed are shown in Section 2. 20 This is a non-directed graph because each relation connects two vertexes, and and connects and in both senses and is denoted with another semantic relation. The internal relations of WN are described in http://wordnet.princeton.edu/man/wninput.5WN.html. 21 66,025 nouns, 17,915 adjectives, 3,575 adverbs and 12,127 verbs. 22 79,689 nouns, 18,563 adjectives, 3,664 adverbs and 13,508 verbs. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Fig 6. Logic Model of the Integration of Semantic Resources (ISR-WN). 3.2 Integration architecture This section describes some of the process features used to integrate all the resources by using WN as an interlinking nucleus. This process takes into account each of their particularities and integrated resource annotations based on different WN versions, in this case by considering version 1.6 and 2.0. Despite the fact that this version only uses the English language, it can be extensible to other languages if the Interlingua Index (ILI) of EWN (Dorr and Castellón, 1997, Vossen, 1998) or a similar technology is used to align it to them. Fig 7 shows WN synsets linked to several taxonomies (SUMO, WND and WNA), and descriptions of SC and SWN respectively. In many cases the links are established through the use of mapping files, thus allowing the resources tagged in the WN versions to be interlinked. This results in the creation of an enriched semantic graph that is highly suitable for NLP applications. Fig 7 shows how all the semantic resources (the taxonomies of SUMO (in green), WND (in purple) and WNA (in red), the SC labels (in orange) and the SWN descriptions (in grey) are linked to the WN synsets. Particular aspects from each of the resources involved have therefore been taken into consideration. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Fig 7. Architecture. With regard to integrating all the resources mentioned, mapping files are suitable for interlinking different WN versions, and are used by the semantic resources to map their labels with WN definitions. Owing to the word sense granularity of the WN versions (word senses that exist in one version cannot exist in another) many word senses cannot be taken into account when a mapping occurs. A suitable solution to this problem is to navigate all the necessary mapping files in order to take into account the greatest quantity of possible interlinks. For instance, WNA has been used in this proposal, and has been tagged with WN 2.0, while SWN has been tagged with 1.6 and 2.0 and SC with 2.0. There is, therefore, no need to use mapping files to integrate them. It is worth noting that those semantic resources that are annotated with both WN 1.6 and 2.0 reduce the lost interlinks to 0%. Fig 7 shows a model that allows one of the two nuclei, WN 1.6 or 2.0, to be chosen depending on the user‟s needs. This decision is not limited to the inclusion of other WN versions for further ISR-WN approaches. As can be seen when comparing the second version to the first of WN, the former has new semantic labels and the option of selecting the WN nucleus in order to build the semantic graph on demand. This integration model therefore reuses all the semantic relations included in the Princeton WN. Furthermore, in ISR-WN the SUMO categories also keep their relations with their respective WN mappings; moreover, the semantic relations of WND and WNA are now hypernym (owing to the is_a relation) and hyponym (owing to the is_child relation) in order for their hierarchy taxonomy to make sense. Pertainym is, correspondingly, the semantic relation used to link the WN synset to each label of both taxonomies (e.g. a synset can pertain to one or many WND and WNA labels). It is important to stress that the WNA for version 1.1 has been populated with new affective relations, which do not exist in WNA1.0 (e.g. entailment and cause). New relations in WNA1.1 can therefore link verbs, adjectives and adverbs with those nouns from which these are derived. All these considerations have been taken into account when developing ISR-WN resource. The resulting semantic interconnections of this knowledge base permit navigation among all the resources involved. For instance, if the English word atrocious is explored using WN 1.6 as a nucleus, a list of WN synsets can be obtained as regards this word. Once a synset (e.g. offset 00193347) has been selected from its representative list, the following information is retrieved.  00193347 atrocious [Adjective] o Similar-To (00192906 alarming [Adjective]) o Pertainym (Psychological_Feature [Domain]) o Hypernym (Subjetive_Assessment_Attribute [SUMO]) ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT o Pertainym (Emotion [Affect] ) o Pertainym (Horror [Affect] o SentiWN-Description ( atrocious#3 Pos: 0|Neg: 0.625 |Obj: 0.375 [SentiWordNet]) o Cause (05591212 horror [Noun]) As will be noted, the offset, part-of-speech, synonym list and gloss are retrieved for each WN synset. The description of each semantic label from each particular resource linked to the synset being explored is also retrieved. As a search tool, ISR-WN also allows labels to be discovered in all the resources included, thus providing theirs associated information and establishing semantic navigation in order to find interesting semantic paths. An interesting case may be discussed as regards the new features introduced in this integration resource for WNA. For instance, in the case of exploring the label Horror from WNA, we find that this label is not linked directly with the third synset of atrocious (e.g atrocious#3 with offset 00193347). Moreover, it has been assumed that if a synset (e.g. atrocious#3) is linked with a noun (e.g. horror#1) by an affective relation (e.g. entailment, cause) which has been obtained from WNA1.1, the synset will be linked to all the WNA labels linked to its noun. Upon applying this procedure the synset horror#1, which is linked with the WNA label Horror provokes atrocious#3 to be additionally linked to Horror. Fig 8 shows the aforementioned example after applying the previously mentioned procedure, focusing on the synset atrociuos#3 when linked to two WNA labels (e.g. Emotion and Horror). Fig 8. ISR-WN including new affective relation from WNA1.1. As part of the semantic improvement strategy for ISR-WN, the semantic links suggested by XWN1.7 and XWN3.0 have also been considered. Their semantic relations to the synsets for ISR-WN have been tagged as follows:  XWN17_Relation_as_Synset,  XWN17_Relation_as_Gloss,  XWN30_Relation_as_Synset,  XWN30_Relation_as_Gloss. Different features are involved in the resulting semantic graph depending on the nucleus selected (WN 1.6 or 2.0) in order to load the semantic knowledge base on demand. It is important to highlight that Relation_as_Synset and Relation_as_Gloss represent a bidirectional relation between synset and gloss. In the case of using WN1.6 as a nucleus, 505,755 and 199,123 relations will be involved, and are from XWN1.7 and XWN3.0, respectively. However, when WN2.0 is used to load ISR-WN 551,065 relations from XWN1.7 will be involved, in addition to 358,843 relations from XWN3.0. Table 4 shows all the semantic relations included in this integration resource. Note that XWN1.7 is annotated with WN1.7, and it is therefore necessary to apply a conversion of WN versions in order to map their offset of version 1.7 onto the WN versions used as nucleus (1.6 or 2.0). The integration of XWN3.0, which is annotated with WN3.0, is performed in the same way. These mappings result in some relations being missed owing to the conversion issues mentioned above. An in-depth analysis of this can be found in the evaluation section. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT An exploration of ISR-WN is presented as follows, in which WN1.6 is used to load the knowledge base. This exploration reuses the example of the word love, in which an increase in the semantic information is evident as regards the target synset:  01211759 love [Verb] o Antonym (01211167 hate, detest [Verb]) o Hyponym (01212004 love [Verb]) o Hyponym (012125039 care_for, cherish, hold_dear, teasure [Verb]) o Hyponym (01213640 dote [Verb]) o Hyponym (01213998 adore [Verb]) o Pertainym (Factotum [Domain]) o Hypernym (Intentional_Psychological_Process [SUMO]) o Pertainym (Emotion [Affect]) o Pertainym (Love [Affect]) o Stative (05607724 love [Noun]) o Hypernym (love.v.01 [SemanticClass]) o SentiWN-Description ( love#1 Pos: 0.5|Neg: 0 |Obj: 0.5 [SentiWordNet]) o XWN17_Relation_as_Synset (05608483 affection, affectionateness, fondness, tenderness, heart, warmheartedness [Noun]) o XWN17_Relation_as_Synset (05573285 liking [Noun]) o XWN17_Relation_as_Synset (01508689 have, have_got, hold [Verb]) o XWN17_Relation_as_Synset (01332909 great [Adjective]) o XWN17_Relation_as_Gloss (07626109 sweetheart, sweetie, steady, truelove [Noun]) o XWN17_Relation_as_Gloss (07462325 patriot, nationalist [Noun]) o XWN17_Relation_as_Gloss (07111212 bibliophile, booklover, book_lover [Noun]) o XWN17_Relation_as_Gloss (07073765 amorist [Noun] o XWN17_Relation_as_Gloss (06950629 lover [Noun]) o XWN17_Relation_as_Gloss (06930637 Brunnhilde, Brynhild [Noun]) o XWN17_Relation_as_Gloss (06921123 Psyche [Noun]) o XWN17_Relation_as_Gloss (06897084 Adonis [Noun]) o XWN17_Relation_as_Gloss (01402650 unloved ,not loved [Adjective]) o XWN30_Relation_as_Gloss (05315655 hyperbaton [Noun]) o XWN30_Relation_as_Gloss (06950629 lover [Noun]) o XWN30_Relation_as_Gloss (07111212 bibliophile, booklover, book_lover [Noun]) o XWN30_Relation_as_Gloss (08287863 strawflower, golden_everlasting, yellow_paper_daisy, Helichrysum_bracteatum [Noun]) o XWN30_Relation_as_Synset (05608483 affection, affectionateness, fondness, tenderness, heart, warmheartedness [Noun]) o XWN30_Relation_as_Synset (05573285 liking [Noun]) o XWN30_Relation_as_Synset (01213998 adore [Verb]) o XWN30_Relation_as_Synset (01214144 idolize, worship, hero-worship, revere [Verb]) As can be appreciated in this example, the exploration for the same word love (offset 01211759 for the verb love) has been considerably enriched. The semantic information obtained combines many labels from different resources and establishes new relations among WN synsets. This integration of resources allows to discovery new semantic interconnections never before shown in a single tool. Interesting features such as WND or SUMO concepts with positive or negative tendencies could also be mentioned. To return to the example of the word love, an exploration of one level in depth has been applied. However, if we were to navigate in greater depth, many more semantic elements would be discovered. Table 4 shows all the semantic relations. This table is used as a matrix to relate each resource by rows. Note that in Table 5 each relation in ISR-WN has one inverse relation. Most of the relation pairs have been taken from WN 23 and the other resources involved, while the rest (those marked with *) have been created in this research work. The aim of creating bilateral relations has been to allow forwards and backwards navigation through the semantic links. 23 http://wordnet.princeton.edu/man/wninput.5WN.html ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT WN WND WNA SUMO SC SWN WN Also see Pertainym Pertainym Hypernym Hypernym SentiWN_Description* Antonym Hyponym Hyponym Attribute Cause Derivationally related form Derived from adjective Domain of synset - REGION Domain of synset - TOPIC Domain of synset - USAGE Entailment Hypernym Hyponym Instance Hypernym Instance Hyponym Member holonym Member meronym Member of this domain - REGION Member of this domain - TOPIC Member of this domain - USAGE Part holonym Part meronym Participle of verb Pertainym (pertains to noun) Similar to Substance holonym Substance meronym Verb Group XWN17_Relation_as_Gloss* XWN17_Relation_as_Synset* XWN30_Relation_as_Gloss* XWN30_Relation_as_Synset* WND Pertainym Hypernym Hyponym WNA Pertainym Hypernym Hyponym SUMO Hypernym Hypernym Hyponym Hyponym SC Hypernym Hypernym Hyponym Hyponym SWN SentiWN_Description* Table 4. Relations between semantic labels on ISR-WN. Bidirectional relations on WordNet Pointer Reflect Antonym Antonym Hyponym Hypernym Hypernym Hyponym Instance Hyponym Instance Hypernym Instance Hypernym Instance Hyponym Holonym Meronym Meronym Holonym Similar to Similar to Attribute Attribute Verb Group Verb Group Derivationally Related Derivationally Related Domain of synset Member of Doman Bidirectional relations added by ISR-WN ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Pointer Reflect Entailment Cause Participle Participle Pertainym Pertainym XWN17_Relation_as_Gloss* XWN17_Relation_as_Synset* XWN30_Relation_as_Gloss* XWN30_Relation_as_Synset* SentiWN_Description* SentiWN_Description* Stative Cause Synonymy Synonymy Table 5. Bidirectional relations on ISR-WN. 4. Integration process evaluation In this section the results obtained are analysed by taking into consideration the semantic elements that are available for the integration and those eventually integrated. Once developed ISR-WN following the aforementioned process, an evaluation took place. This was carried out in order to obtain a statistic assessment of the quantity of WN synsets that should be linked as regards those established. Two evaluations were applied based on the semantic knowledge interlinked depending on the nucleus (WN 1.6 or WN2.0) selected for loading the semantic graph. 4.1 Evaluation based on WN1.6 as a nucleus Table 6 shows a distribution of the semantic labels involved in ISR-WN with WN1.6 as a nucleus. In this analysis the greatest possible quantity of labels has been extracted for inclusion, and the labels which have not been included are justified below. An important factor that supports the 100% of alignments of WND, SUMO and SC has been the reduction in the usage of mapping files. This was possible since these three resources were built based on WN 1.6 and the current evaluated nucleus is WN1.6. WND 2.0 SUMO WNA 1.0-1.1 SC SWN 3.0 XWN 1.7 XWN 3.0 Mean # Labels 172 568 300 1,231 117,659 - - Synsets to link n 86,901 67,923 1,256 66,025 82,114 - - a 19,322 18,531 2,418 - 18,157 - - v 12,843 12,469 801 12,127 13,767 - - r 3,735 3,627 614 - 3,621 - - Total of synsets to link 12,2801 102,550 5,089 78,152 117,659 551,551 419,387 Linked Synsets n 86,901 67,923 1,096 66,025 56,563 - - a 19,322 18,531 2,125 8,757 - - v 12,901 12,469 474 12,127 9,223 - - r 3,735 3,627 549 2,101 - - Total of linked synsets 122,801 102,550 4,244 78,152 76,644 505,755 119,123 Difference 0 0 845 0 41,015 45,796 300,264 Linked % 100.00 100.00 83.40 100.00 65.14 91.70 28.40 81.23 Table 6. Synsets linked to each resource by using WN1.6 as a nucleus. In this integration resource we have considered both WNA versions (1.0 and 1.1) to be a single resource and have mixed them. This mixture has been developed by involving the taxonomy of WNA1.1 and the WN mappings from both versions (1.0 and 1.1). However, a difference in labels (as regards all linked synsets with regard to those synsets that should be linked) between WNA 1.0 and WNA 1.1 has caused 845 missing links (see Table 6). This special case is owing to the fact that most of the WNA1.0 labels are included in WNA1.1, but several WNA1.1 labels are not included in WNA1.0. Some links have therefore been removed for the integration. All the affective labels not considered by ISR-WN because they are not represented in the WNA1.1 taxonomy are shown as follows: attitude, emotional response, psy, man, sympathy, sta, softheartedness, joy-pride, identification, levity-gaiety, general-gaiety, empathy, positive-concern, compatibility, kindheartedness and buck-fever. Note that the taxonomies used are the most recent for WND and WNA (i.e. WND 3.2 and WNA1.1). ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT WN mapping files have also been considered for the alignment of WNA to WN. WNA1.1 has a peculiar feature consisting of the fact that only nouns are linked to affective labels. Adjectives, verbs and adverbs are therefore linked to these nouns at sense level by means of derived relations (i.e. entailment and cause). The description of this issue can be found in the section “3.1 Integration process”. Furthermore, when evaluating SWN integration based on WN 1.6 as a nucleus, it was found that the main difference between the mismatching of SWN3.0 alignments had occurred because this resource had been developed on the basis of a very different WN version, WN3.0. It is therefore possible to find that several SWN features that are annotated with WN3.0 do not exist for WN1.6. This justifies all the lost links shown in Table 6. Table 6 also shows how the integration of XWN1.7 reached 91.70%. Moreover, with XWN3.0 it reached a lower integration of 28.40%. A similar phenomenon occurred with SWN3.0 in which several mapping WN versions were involved. This is responsible for the low enrichment of ISR-WN as regards XWN3.0 when WN1.6 is used as a nucleus. Graph-based approaches, such as (Gutiérrez, 2012) which uses ISR-WN with XWN3.0, cannot therefore obtain the same relevant results as those obtained using XWN1.7. Once ISR-WN is loaded with WN1.6, it contains a total of 178,558 vertexes, of which 99,643 represent WN synsets, 172 WND domains, 558 SUMO categories, 300 WNA affective labels, 1,231 SC semantic classes and, finally, 76,644 SWN descriptions. ISR-WN therefore includes a total of 232,6211semantic relations with WN1.6 as a nucleus. 4.2 Evaluation based on WN2.0 as a nucleus Table 7 shows a similar comparative study to Table 6, but here WN2.0 is considered as a nucleus. WND 3.2 SUMO WNA 1.0-1.1 SC SWN 3.0 XWN 1.7 XWN 3.0 Mean # Labels 172 568 300 1,231 117,659 - - Synsets to link n 103,504 79,688 1,256 66,025 82,114 - - a 19,398 18,564 2,418 - 18,157 - - v 19,398 13,507 801 12,127 13,767 - - r 3,835 3,663 614 - 3,621 - - Total of synsets to link 146,135 115,422 5,089 78,152 117,659 551,551 419,387 Linked Synsets n 103,504 79,688 1,089 65,904 78,061 - - a 19,398 18,564 2,118 11,052 - - v 19,398 13,507 473 12,064 13,207 - - r 3,835 3,663 580 3,428 - - Total of linked synsets 146,135 115,422 4,260 77,968 105,748 505,755 119,123 Difference 0 0 829 184 11,911 486 60,544 Linked % 100.00 100.00 83.71 99.76 89.88 99.91 85.56 94.11 Table 7. Synsets linked to each resource by using WN2.0 as a nucleus. As can be seen in Table 7, in ISR-WN SC has some missing links, since it has been developed using WN1.6 and has to connect with WN2.0 through the use of mapping files. When this occurs, some synsets are not considered since some of them do not exist for both versions (1.6 and 2.0). It will, however, be noted that SWN links increase in comparison to the previous distribution table because the mapping versions are reduced. WNA 1.0 and WNA 1.1 have been distributed together in both tables, since both have been seen as a single resource. All the WNA1.1 labels and the synset-label relations from WNA 1.0 and 1.1 have therefore been considered. However, only all those WNA1.0 labels that do not exist in the WNA1.1 taxonomy have not been considered. Both versions of WNA are thus considered to be a single resource. As Table 7 shows, when using WN2.0 as a nucleus the incorporation of XWN increases significantly, particularly in the case of XWN3.0. It is most appropriate to use ISR-WN with WN2.0 since the mean of the integration percentage was 94.11%, while for WN1.6 it was 81.23%. It is therefore evident that the best integration of resources based on WN occurs when WN2.0 is taken into account. In this case, 223,448 vertices have been computed, of which 115,424 are WN synsets, 172 are WND domains, 568 are SUMO categories, 300 are WNA affective domains, 1,231 are SC labels and 105,748 are SWN descriptions. All these vertices are related to 2,893,838 semantic relations. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 5. NLP tasks description As we have mentioned before, our goal in this work is to provide a semantic framework able to enrich texts with semantic information and use this new information to classify, obtain opinions or recommend similar texts in open domains. To do this, we use ISR-WN as knowledge source for different NLP tasks. Basically, we have conducted four different tasks to add new information from different point of views: 1- Word Sense Disambiguation (WSD). We need to obtain the correct sense of each word in order to obtain accurate results. Due to the fact that one word has different meanings depending on the context, it is necessary to decide which word sense is the correct because it will determine a proper understanding. 2- Semantic similarity. The goal of this task is to discover similar content. According to the words, senses, content structure, etc., we will be able to obtain similar opinions that could be expressed with different words but the meaning is still similar. 3- Domain classification. This task will provide the most relevant domains of a text a set of texts. This information will be useful to annotate texts in order to recommend similar content. 4- Sentiment analysis. The goal of this task is to decide whether an opinion is positive or negative automatically. Moreover, many scientific researchers have used the semantic features that ISR-WN provides to deal with different NLP tasks. Some of these tasks and their contributions are shown below. Notice that most of semantic NLP challenges are based on these WN versions, therefore this integration resource is very useful for dealing with these tasks (see http://senseval.org/). Word Sense Disambiguation approaches (WSD): In order to create effective NLP systems it is necessary to transform the information extracted from the words in plain text into a conceptual level to detect meaningful word senses. In WSD the goal is to determine the words‟ senses in a text in which a word may have different senses depending on the context that it appears in. So that, the main purpose of the above listed approaches is to automatically choose the intended sense (meaning) of a word in a particular context. To be able to do that, it can be used two types of approaches: knowledge based approaches (Zhibiao and Martha, 1994), (Leacock and Chodorow, 1998) and corpus-based approaches (Peter, 2001). The former requires lexical resources such as WordNet, Roget's Thesaurus 24 , BabelNet 25 , DBpedia 26 , etc. to obtain semantic similarities, while the latter uses co-occurrences to measure the similarity between words. In our case, by conducting knowledge based approaches through ISR-WN, we find all the possible synsets for the word in WordNet and also their links to different concepts of the aforementioned resources (i.e. WND, WNA, SUMO, SC) being able to group word synsets under semantic concepts facilitating their distinctions depending on the context. For example for the concept “Sport” of WND we can find three word senses with respect to the word “exercise”, two for “School”, one for “Sociology” and one for “Pedagogy”. So that, if we determine the context of the sentence where “exercise” is being used, we can reduce the number of possible meanings to assign. The advantage to use ISR-WN is we can do this procedure not just using WND, we can also utilise the rest of the conceptual resources at the same platform. WSD can be also considered a task for doing text‟s semantic enrichment, since by applying this technique and depending on the knowledge base used; the processed text can be tagged or linked with different sources. This phenomenon can be associated to the task Entity Linking in which the matching of a textual entity mention to a knowledge base is performed. This entity mention can be a Wikipedia page, DBpedia entry, or a specific URI that identify an entity, that is a canonical entry for that entity (Rao et al., 2013). The reason whereby we are introducing the Entity Linking task is because we also can integrate BabelNet with ISR-WN for dealing with it, even for different languages. Such is the case of (Gutiérrez et al., 2013) in Semeval-2013 for Task 12 “Multilingual Word Sense Disambiguation” 27 . A list of WSD approaches that have been supported the conceptual integration of ISR-WN are described below. Among them, the second approach not just describes a word sense disambiguation approach, it also describes how by using ISR-WN relevant conceptual trees of every conceptual resource 24 http://www.thesaurus.com/roget 25 http://babelnet.org/ 26 http://dbpedia.org 27 http://www.cs.york.ac.uk/semeval-2013/task12/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT integrated in ISR-WN, can be obtained. These trees can be interpreted as a manner of obtain those concepts that classify a text. For example, in the sentence taken from an IMDb‟s 28 user review “Grey's Anatomy stars Ellen Pompeo (who has starred in a few movies but was never really noticeable) as the narrator - Meredith Grey”, using this approach based on WND this text can be classified with the following concepts [“Cinema”, “Art”, “Humanities”]. The following researches, which are based on ISR- WN, support the tasks mentioned:  Participation of the UMCC-DLSI research team in Semeval-2010 for Task 17: All-words Word Sense Disambiguation on Specific Domain (Agirre et al., 2010). A description of this approach can be found in (Gutiérrez et al., 2010b).  Research paper (Gutiérrez et al., 2011b) which was discussed in the RANLP 2011 conference 29 and which combined ISR-WN (considering , WN, WND, WNA and SUMO) and word sense frequencies to solve Word Sense Disambiguation. This proposal attained higher results in comparison to top proposals at a world level.  Research paper (Gutiérrez et al., 2011d) which was discussed in the NLDB2011 conference 30 and which used ISR-WN (considering WN, WND, WNA and SUMO) as a graph-based approach to deal with Word Sense Disambiguation. This proposal attained important results in comparison with the reported submissions of Senseval-2 (Cotton et al., 2001). Participation of the UMCC-DLSI research team in Semeval-2013 for Task 12 “Multilingual Word Sense Disambiguation” 31 . This approach attained the first position of the rank. A description of this approach can be found in (Gutiérrez et al., 2013) In order to show the results obtained by WSD approaches supported by ISR-WN we describe the two types used in this work: graph-based (Gutiérrez et al., 2013) and tree-based (Gutiérrez et al., 2011b). The graph-based approach takes into account the whole ISR-WN graph as configuration, thus it includes all ISR-WN resources and their links for applying WSD tasks. As part of the evaluation of this approach we participated in the Semeval-2013 for Task 12 “Multilingual Word Sense Disambiguation” Campaign. For example, for evaluating WSD in English language, 1,931 instances of Babelnet, 1,242 of Wikipedia and 1,644 of WordNet were considered. Note that in this case ISR-WN and BabelNet, were aligned through WordNet, since WordNet is common for both. Based on the corpus 32 provided by Task 12 “Multilingual Word Sense Disambiguation”, our WSD approach was able to be placed at the top of the campaign ranking, reaching a F1 around 68.5%, 54.6% and 64.7% for BabelNet, Wikipedia and WN respectively. More details about the approach and its experiments on Semeval-2013 campaign see (Gutiérrez et al., 2013). Resources Files Experiments WN WNA SUMO WND MFS Precision Recall F1 F1 Difference with the Best system Corpus d00.txt [648 instances] Exp 1 (MFS) X 0.565 0.564 0.564 Exp 2 X 0.572 0.572 0.572 Exp 3 X 0.561 0.56 0.560 Exp 4 X 0.555 0.554 0.554 Exp 5 X 0.572 0.572 0.572 Exp 6 (Voting) all ISR-WN as configuration X X X X X 0.575 0.575 0.575 Whole corpus (d00.txt[648 instances], d01.txt[1032 instances], d02.txt[757 instances]) Exp 7 (MFS) X 0.601 0.599 0.600 0.090 Exp 8 (Voting) all ISR-WN as configuration X X X X X 0.610 0.609 0.609 0.081 Best system of Senseval-2 0.690 0.690 0.690 Average of the Senseval-2 Results 0.499 0.360 0.391 Worst system of Senseval-2 0.370 0.345 0.357 Table 8. Evaluation results of the tree based WSD approach with Senseval-2 corpora 28 http://www.imdb.com/ 29 http://lml.bas.bg/ranlp2011/ 30 http://gplsi.dlsi.ua.es/congresos/nldb11/ 31 http://www.cs.york.ac.uk/semeval-2013/ 32 https://www.cs.york.ac.uk/semeval-2013/task12/index.php%3Fid=data.html ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT In order to show how to use each resource individually (i.e. WN, WNA, SUMO and WND) we describe another WSD approach: tree-based (Gutiérrez et al., 2011b). In next table we show the evaluation, using the corpora provided by the public Rada Mihalcea‟s repository. This repository includes 2,447 instances from the Senseval-2 competition “The English All-Words Task” (Cotton et al., 2001). Table 8 shows the results by using different configurations: each resource individually or all together. Note that the most frequent word sense (MFS) is also considered as a dimension for the voting. More details in (Gutiérrez et al., 2011b). Opinion Mining approaches: Nowadays, Opinion Mining (OM), also known as Sentiment Analysis (SA), as part of an NLP task has become a popular discipline due to its wide-relatedness to social media behaviour studies. OM is commonly used to analyse the comments that people post on social networks. Also, it allows identifying the preferences and criteria of users about situations, events, products, brands, etc. In order to exploit the potential that provides ISR-WN with the integration of semantic, conceptual, affective and sentiment scoring resources some approaches have been produced for the detection of opinions, relevance, and polarity classification and also measuring the impact of the sentiment polarity detection on Textual Entailment tasks. Those approaches are the following:  The multidimensional point of view of ISR-WN has been useful in the Opinion Mining area. In this case it was presented in the WASSA‟11 Workshop 33 , dealing with three opinion mining tasks in the MOAT (NTCIR Multilingual Opinion Analysis Task 34 ) competition system. These tasks consist of identifying whether or not a sentence represents an opinion. Another task involves identifying the polarity of the opinion sentence, while the last one consists of aligning questions with topics. This proposal used ISR-WN to extract relevant concepts which represent the sentences analysed. The Opinion Mining tasks were carried out by taking these concepts and linking them with SWN (in Enriched ISR-WN) sentiment polarities. This research attained relevant results which could be comparable with the first places attained in the MOAT campaign (Gutiérrez et al., 2011c)  Paper research “Approaching Textual Entailment with Sentiment Polarity” (Fernández et al., 2012b), discussed in the ICAI‟12 conference: The 2012 International Conference on Artificial Intelligence. This takes ISR-WN as a knowledge base and uses the (Gutiérrez et al., 2011c) approach to determine Textual Entailment with Opinion Mining techniques. Semantic Textual Similarity (STS) approaches: STS is related to Textual Entailment35 (TE) and Paraphrase36 tasks. The main difference is that STS assumes bidirectional graded equivalence between the pair of textual snippets. In case of TE, the equivalence is directional (e.g. a student is a person, but a person is not necessarily a student). In addition, STS differs from TE and Paraphrase in that, rather than being a binary yes/no decision, STS is a similarity-graded notion (e.g. a student is more similar to a person than a dog to a person). This graded bidirectional is useful for NLP tasks such as Machine Translation (MT), Information Extraction (IE), Question Answering (QA), and Summarization. Several semantic tasks could be added as modules in the STS framework, such as WSD and Induction, Lexical Substitution, Semantic Role Labelling, Multiword Expression detection and handling, Anaphora and Co- reference resolution, Time and Date resolution and Named Entity, among others”. Below can be found different participations on international competitions where STS systems made use of ISR-WN a semantic nucleus allowing judging textual similarities from multidimensional perspectives such as from domain and affective points of view and others. The following researches support STS tasks based on ISR-WN:  Participation of the UMCC-DLSI research team in Semeval-2012 in Task 6 “Semantic Textual Similarity”(Agirre et al., 2012) . A description of this approach can be found in (Fernández et al., 2012a).  Participation of the UMCC-DLSI research team in Semeval-2013 Task 5 37 . “UMCC_DLSI- (EPS): Paraphrases Detection Based on Semantic Distance” (Dávila et al., 2013) discussed in Second Joint Conference on Lexical and Computational Semantics (*SEM). It took ISR-WN as a knowledge base to determine the semantic similarity of different short texts. 33 http://gplsi.dlsi.ua.es/congresos/wassa2011/ 34 http://research.nii.ac.jp/ntcir/ntcir-ws8/ws-en.html 35 http://aclweb.org/aclwiki/index.php?title=Recognizing_Textual_Entailment 36 The adaptation or alteration of a text or quotation to serve a different purpose from that of the original. 37 http://www.cs.york.ac.uk/semeval-2013/task5/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT  Participation of the UMCC-DLSI research team in Semeval-2012 in Task 6 “Semantic Textual Similarity” 38 . A description of this approach can be found in (Chávez et al., 2013). Ongoing approach: In order to capture the semantics of a textual question, we pretend to use the aforementioned approaches (WSD and STS) supported by ISR-WN, to make interpretations of both questions written in natural language and ontology node‟s descriptions (and also considering the name of the adjacent nodes) for finding further alignments with existing ontologies. Reusing the work below, which has been the first approach by using ISR-WN, we would be able to recover those ontology nodes that involve textual descriptions relevant to textual question. The following research is part of this ongoing approach.  Paper research “Semantic Information extraction method on ontologies”(Dávila et al., 2012) discussed in SEPLN‟12: XXVIII Congreso de la Sociedad Española para el Procesamiento del Lenguaje Natural, in which semantic ontology searches are applied on the basis of ISR-WN and natural language adaptation. Next, we describe the annotation process through a case of study taken from IMDb. 6. Case study In order to enrich texts using ISR-WN, we have developed different approaches to automatically process textual information and obtain the semantic knowledge associated to each word or to the general context. As a result, we provide a semantic framework suitable of being used as support of content-based recommender systems that is able to find contents that are similar to those already evaluated by the users by considering semantic information. In this section we illustrate the annotation process with an example obtained from IMDb 39 (Internet Movie Database) and explain the different approaches used to conduct the annotation process. Our experiments are carried out using textual information from IMDb. IMDb is the world‟s most popular and authoritative source for movie, TV and celebrity content. This website has more than 250 million unique monthly visitors and offers a searchable database of more than 185 million data items including more than 3 million movies, TV and entertainment programs and more than 6 million cast and crew members. As we have mentioned above, IMDb provides a lot of information related to movies, TV shows or TV series. However, due to the fact that we are interested in people‟s interest and opinions we only focus our attention over the reviews & ratings section. In this case, for each movie or TV series the reviews & ratings section has a list of reviews from different users that give their opinions about general concepts, feelings, likes and dislikes, etc. Our purpose is to collect textual information provided by the reviews of different users about TV series or movies in order to enrich it with semantic information and therefore being able to rate or recommend similar content according to people feelings, interests or likes. With this case study, we aim at demonstrating the practical use of our proposed resource. 7. Semantic enrichment approaches In this section, we briefly describe how to enrich textual information by using ISR-WN. As we have mentioned in Section 3, we have integrated a set of different resources in order to obtain new relations and therefore being able to build new connections. With ISR-WN we can enrich texts in different dimensions using: affection labels, domain labels, semantic classes, etc. To conduct our experiments and annotate texts we have developed a framework that uses different techniques in order to take into account each dimension of ISR-WN. 7.1 WSD First of all, it is important to extract the correct senses of words. To do this, our framework uses an unsupervised multilingual approach of WSD. It works using a words window and applies a modification of the Personalizing Page Rank (Ppr) algorithm (Agirre and Soroa, 2009) considering as knowledge base the ISR-WN semantic network. More in detail, it uses senses frequencies to rank the synsets of each 38 http://ixa2.si.ehu.es/sts/ 39 http://www.imdb.com ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT lemma (in a descending order) according to a calculated factor of relevance (Ppr+Freq algorithm, more details see in Section 7 or (Gutiérrez et al., 2013)). This approach was evaluated in Semeval-2013 obtaining the best results of this campaign. An example of the results after processing a text fragment, taken form IMDb‟s user reviews, is shown in next table: Original fragment Format: [word][(word position in the sentence] WSD results Format: [position of the word in the sentence][word sense key suggested as appropriated in this context][gloss of the word sense suggested] Her mother(1) is(2) the famous(3) surgeon(4) and she is trying(5) to follow(6) in her mother's(7) footsteps(8) (1) mother%1:18:00:: "a woman who has given birth to a child (also used as a term of address to your mother); \"the mother of three children\" (2) be%2:42:03:: "have the quality of being; (copula, used with an adjective or a predicate noun); \"John is rich\"; \"This is not a good answer\" (3) famous%5:00:00:known:00 " widely known and esteemed; \"a famous actor\"; \"a celebrated musician\"; \"a famed scientist\"; \"an illustrious judge\"; \"a notable historian\"; \"a renowned painter\" surgeon%1:18:00:: "a physician who specializes in surgery " (4) be%2:42:03:: " have the quality of being; (copula, used with an adjective or a predicate noun); \"John is rich\"; \"This is not a good answer\" (5) try%2:41:00:: " make an effort or attempt; \"He tried to shake off his fears\"; \"The infant had essayed a few wobbly steps\"; \"The police attempted to stop the thief\"; \"He sought to improve himself\"; \"She always seeks to do good in the world\" " (6) follow%2:38:00::" to travel behind, go after, come after; \"The ducklings followed their mother around the pond\"; \"Please follow the guide through the museum\" (7) mother%1:18:00:: " a woman who has given birth to a child (also used as a term of address to your mother); \"the mother of three children\" (8) footstep%1:23:00:: " the distance covered by a step; \"he stepped off ten paces from the old tree and began to dig\" " Table 9: Disambiguation example As Table 9 shows, each word (nouns, verbs, adjectives and adverbs) is disambiguated. The annotation process after disambiguating each word obtains: lemma and word sense (mother%1:18:00::) and a gloss or definition with examples ("a woman who has given birth to a child (also used as a term of address to your mother); \"the mother of three children\"). With this information we are able to discriminate among a set of senses and understand the meaning of words if they appear in different contexts. 7.2 Concepts and Polarities In order to obtain relevant concepts and extract sentiment information we propose an unsupervised knowledge-based approach, described in section before, that uses the RST (Relevant Semantic Trees) technique (Gutiérrez et al., 2011c). The result is a set of relevant semantic trees associated to each sentence that provides sentiment polarity values. An example of the relevant semantic tree technique based on WordNet Domains resource is shown in Fig 9, where the number between parentheses of each concept indicates the order of relevance from 1 to 7 according to the next fragment. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Fig 9. Relevant Semantic Tree (RST) based on WordNet Domains (WND) This tree represents the most relevant domains related to the next text fragment extracted from a review of Grey‟s Anatomy from IMDb: “…The story revolves around her time as an intern and the people she meets and sort of is portrayed as "survival camp for medical students." The minute she arrives at work, she meets Christina Yang (Sandra Oh - flawless in her bitchy supporting role), George O'Malley (T.R. Knight - one word: breakthrough performer), Izzie Stevens (Katherine Heighl - very, very believable as a model who is more like the girl-next-door), and Alex Karev (Justin Chambers - plays sort of a not-so-likable person). Most of all, there is Dr. Derek Shepherd (Patrick Dempsey - very attractive), the man that Meredith had a one- night stand with - he just happens to be her boss. This is a show that wants to be liked….”. As we can see in Fig 9, there are seven domains that are closely related to the meaning of the context, starting from Person to Telecommunications. Moreover, WND has a hierarchy where all the domain labels (172 labels) are connected and Fig 9 also shows how the relevant domains are arranged into the hierarchy. This information is useful in order to classify texts into different categories and to obtain related information through a previous categorization. 7.3 Semantic Textual Similarity Last step to enrich textual information is to extract semantic textual similarities. Our approach is a Machine Learning System (MLS) that uses several algorithms to extract all features: similarity measures, lexical-semantic alignment, semantic alignment (Fernández et al., 2012a). In order to extract the semantic features it uses the multidimensional resource ISR-WN. Thus our STS approach provides a value scale to decide whether a pair of contexts can be considered semantically similar or not. The scale has 5 values that go from 1 to 5. Where 1 indicates that there is no semantic relation between two pair of contexts and 5 indicates that a pair of contexts is semantically equivalent. See the evaluation described in next section. 8. Experiment setup and results Due to the fact that we need real texts to evaluate our framework, our experiments have been conducted by using texts from IMDb user reviews. These texts contain information related to user feelings, expectations, complains, acceptance, etc. Moreover, each review has associated an overall rating. In order to achieve comprehensive results we have extracted information related to medical TV series: House M.D., Grey Anatomy, etc. Accordingly to our data set, we randomly have selected 3 reviews to show how texts are annotated. Table 10 shows a brief part of the original reviews that have been annotated. TV series/Movies Original Texts Grey‟s Anatomy If you think ABC can't get any better - you're wrong. With the great success over smash-hits, "Desperate Housewives" and "Lost" they also picked up a few shows over mid-season hoping for more success. They got "Jake in Progress," "Eyes" and "Grey's Anatomy" - but "Grey's Top Level (1) Person Doctrines History (5) Heraldry (4) Art (2) Theatre Social Science (7) Telecommunications (3) Cinema Applied Science (6) Medicine ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Anatomy" is definitely considered the best out of those three. Grey's Anatomy stars Ellen Pompeo (who has starred in a few movies but was never really noticeable) as the narrator - Meredith Grey. Her mother is the famous surgeon and she is trying to follow in her mother's footsteps.… House M.D. Let me put it simply. I am a physician, and as an inviolable rule, I HATE medical shows. Granted, TV series tend to be one dimensional, due to inherent difficulties in the genre, but "doctor shows" are something I avoid like the proverbial plague. And then one evening I caught "House, MD" and was completely drawn into the show. In House I find the anti-hero that I've been waiting for in a medical show. The guy who knows everything, but is wrong often enough to keep us all guessing. I enjoy the contrast of House and his cadre of young fresh faced colleagues… Tomorrowland When the director of "The Incredibles" signed for this film, I was looking forward to the same amount of humour and exhilaration present in that animated masterpiece, something similar to what the little boy expresses when he realizes he can run on water. Nothing remotely close occurs in "Tomorrowland", a film that suffers from having too big a budget and hardly any original or exciting thoughts. It is also hindered by the fact that almost all of the actors appear clueless and not quite matching their characters. There's something about George Clooney… Table 10: Original reviews obtained from IMDb IMDb provides the overall rating for each TV show and movie in their database. For Grey‟s Anatomy the overall rating is 7.7/10 from 140,600 users, for House M.D. is 8.9/10 from 261,929 users and for Tomorrowland 6.7/10 from 39,939 users. After processing each text we obtain a set of different features that we use to enrich the original information. Table 11 shows the results of annotation for Affect labels and domain labels including a rating of positiveness, negativeness and objectiveness, respectively: Title Affects P N O Domains P N O Grey‟s Anatomy love annoyance liking anger anxiety 0.82 0.18 0.64 Person Theatre Cinema Art Heraldry Medicine Telecommunication 0.72 0.28 0.93 House M.D. sensation love wonder 0.73 0.27 0.68 Psychology Telecommunication Humanities Literature Theatre Medicine Person 0.72 0.28 0.94 Tomorrowland closeness belonging joy behaviour exhilaration sensation 0.49 0.51 0.67 Racing Telecommunication Sociology Theatre Cinema Sport Radio+Tv 0.50 0.50 0.96 Table 11: Results of annotation of three IMDb reviews Once got over the annotation process we are able to infer whether the review is positive or negative and also which are the general domains of each context. This information is useful if we want to recommend similar shows to other people because we can analyse the general content and extract common domains. Moreover, we are able to rate each show according to people‟s reviews gathering if they are positive, negative or objective. Fig 10 shows the annotation of sentiment positiveness of each IMDb review in comparison with the real IMDb Rating: ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Fig 10. Annotation of sentiment positiveness of three IMDb reviews As we can observe, our predictions for Grey‟s Anatomy and Tomorrowland are very similar to real IMDb rating. In these cases the textual enrichment process obtains very good results and the semantic classes and SUMO dimensions are nearby to real ranking. In case of House M.D. there are two points of difference between our predictions and real rating. This is due to the review selected mixture a set of negative and positive reactions to the show. For example, it begins saying “I HATE medical shows” but next says “I enjoy the contrast of House and his cadre of young fresh faced colleagues, complete with starched white lab coats, who struggle as much with their professionally imposed constraints, and sense of decorum, as they do with his personality”. So, in this case it is important to remark that despite the fact that there are some negative aspects our framework is able to establish that there is some positive things that must be taken into account. Another dimension to enrich texts is provided by the STS (Semantic Textual Similarity) approach. In STS, three features are used: syntactic (similar words, syntactic distances, etc), sentiment and semantics. As a result, we are able to identify similar texts by measuring them in a scale from 1 to 5 (where 1 indicates that there is no semantic relation between a pair of contexts and 5 indicates that both are semantically equivalents). In Table 12 we can observe the results obtained after applying STS. The results indicate that there is a slight relation between Grey‟s Anatomy and House M.D. (1.53) and there is no relation between Grey‟s Anatomy and Tomorrowland (0). Semantic Textual Similarity Grey's Anatomy House M.D. Tomorrowland Grey's Anatomy - 1.53 0.00 Machine Learning technique Support Vector Machine 40 Approach STS approach described in Section 7 Table 12: STS results to measure contexts similarities In order to demonstrate how our framework would help to predict the rating with a set of multiple user reviews, we present an exhaustive evaluation over the 10 first IMDb reviews for each mentioned TV shows. Table 13 lists the title of each TV show considered on these evaluations and their ratings, which were manually set by the users. Title No. Review title IMDb Rating (0-1) Grey's Anatomy 1 So far, so good 1 2 Addictive show 0.8 3 Excellent Series 1 40 http://www.support-vector-machines.org/ 0.570 0.551 0.427 0.77 0.89 0.67 0.000 0.100 0.200 0.300 0.400 0.500 0.600 0.700 0.800 0.900 1.000 Grey's Anatomy House Tomorrowland P Affect P Domain P SUMO P Semantic Class P Synset P Mean Real IMDB Rating ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 4 Waste of Time 0.4 5 Boring! 0.1 6 This show is absolutely wonderful!!! 1 7 Wonderful! 1 8 Can someone explain me WHY this is a Top 10 Show? 0.4 9 Started well, now its just more of the same... 0.6 10 I'm Done--and SHAME on me for taking so long 0.1 House 11 Abrasive medical doctor saves lives - no matter what the cost. 1 12 Extremely formulaic fantasy medical detective show 0.1 13 Ingenious and Compelling!!! 1 14 The Greatest Medical Show Ever 1 15 Best Show Ever! 1 16 Watch it for the characters, not the medicine! 0.7 17 One of the best things on the box 0.9 18 Very Good and Getting Better 0.7 19 Everybody lies... 1 20 Pllllllease watch a full episode. 1 Tomorrowland 21 Do not listen to the critics on this one! A vastly under-appreciated tale of promise and hope. 1 22 A fantastic sci-fi experience 1 23 A spectacular Disney sci-fi thrill ride 0.9 24 Misunderstood 0.9 25 Why do people hate this movie? 1 26 Maybe a bit childish but enjoyable sci-fi tale none the less. 0.8 27 Upbeat positive story for the whole family. 0.8 28 A great movie! 0.9 29 Fun kid safe sci fi movie that looks great 0.6 30 Offensively shallow 0.1 Table 13. List of reviews used for automatically rating these three movies Once processed the 30 reviews, Positive, Negative and Objective scores were obtained by a tree based OM approach (Gutiérrez et al., 2011c) using as configuration: WNA (Affect), WND (Domain), SUMO, SC (Semantic Class) and WN (Synsets). Moreover, an additional experiment was conducted to take into account the Harmonic Mean 41 among all positive outputs per review. As Table 14 shows different types of evaluations were performed. One evaluation measure is the Pearson correlation that compares the real rating with the positiveness automatically obtained by our framework. Another one is the combination of all the outputs with the Harmonic Mean where we reach a 69% of correlation. Related to each resource, we can notice that WNA (Affect) obtains the best results of correlation. It means that this resource provides better knowledge to deal with OM challenges due to the fact that WNA conceptualize and link only affective words. On the other hand, if we pay attention to the evaluation shown in Table 14, it shows the results of the mean difference between the real rating and the positiveness obtained by using each resource. In Table 14, WNA reaches the smallest mean difference with 0.22 perceptual points of margin. It is very difficult to compare this kind of reviews with OM system outputs because the users usually set dramatic scores (e.g. 1 or 2) however; their reviews are not as cruel as their scores are. For example, we invite to take a look at the following IMDb review: number 12 of Table 13 titled “Extremely formulaic fantasy medical detective show” with a score of 0.1. In Table 14 it has been highlighted those cases in which the ratings are negatives. In those cases our OM functionality is able to identify that the scoring should be lower than the others but it is not able to set how much. Thus, as future work we propose to carry out new evaluations for introducing OM outputs in a machine learning system. At this way the framework could be able to learn from the different patterns provided by theses outputs and suggest more balanced results. 41 It is one of several kinds of average, and in particular one of the Pythagorean means. Typically, it is appropriate for situations when the average of rates is desired. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Tit. No. Rating (0-1) Affect Domain SUMO Semantic Class Synset Harmonic Mean P N O P N O P N O P N O P N O P N O G r e y 's A n a to m y 1 1 0.81 0.19 0.66 0.78 0.22 0.96 0.59 0.41 0.88 0.69 0.31 0.92 0.58 0.42 0.72 0.68 0.28 0.81 2 0.8 0.92 0.08 0.68 0.84 0.16 0.92 0.81 0.19 0.83 0.71 0.29 0.94 0.6 0.4 0.68 0.76 0.17 0.79 3 1 0.9 0.1 0.66 0.88 0.12 0.95 0.65 0.35 0.91 0.62 0.38 0.93 0.75 0.25 0.8 0.74 0.18 0.84 4 0.4 0.55 0.45 0.76 0.52 0.48 0.93 0.44 0.56 0.89 0.55 0.45 0.93 0.56 0.44 0.7 0.52 0.47 0.83 5 0.1 0.25 0.75 0.76 0.64 0.36 0.96 0.52 0.48 0.92 0.58 0.42 0.94 0.66 0.34 0.7 0.47 0.43 0.84 6 1 0.91 0.09 0.62 0.83 0.17 0.94 0.77 0.23 0.87 0.74 0.26 0.93 0.64 0.36 0.71 0.77 0.18 0.79 7 1 0.93 0.07 0.62 0.56 0.44 0.93 0.68 0.32 0.89 0.63 0.38 0.93 0.66 0.34 0.73 0.67 0.2 0.8 8 0.4 0.62 0.38 0.65 0.41 0.59 0.94 0.44 0.56 0.88 0.69 0.31 0.94 0.67 0.33 0.77 0.54 0.4 0.82 9 0.6 0.79 0.21 0.66 0.59 0.41 0.89 0.55 0.45 0.87 0.61 0.39 0.92 0.62 0.38 0.73 0.62 0.34 0.8 10 0.1 0.54 0.46 0.58 0.58 0.42 0.95 0.57 0.43 0.89 0.62 0.38 0.92 0.6 0.4 0.74 0.58 0.42 0.79 H o u se 11 1 0.92 0.08 0.65 0.72 0.28 0.94 0.64 0.36 0.9 0.61 0.39 0.92 0.66 0.34 0.73 0.69 0.21 0.81 12 0.1 0.58 0.42 0.64 0.49 0.51 0.92 0.59 0.41 0.83 0.53 0.47 0.92 0.67 0.33 0.76 0.57 0.42 0.8 13 1 0.5 0.5 0.68 0.52 0.48 0.97 0.6 0.4 0.93 0.58 0.42 0.93 0.73 0.27 0.77 0.58 0.39 0.84 14 1 0.72 0.28 0.73 0.73 0.27 0.96 0.6 0.4 0.89 0.65 0.35 0.93 0.65 0.35 0.76 0.67 0.32 0.84 15 1 0.76 0.24 0.55 0.7 0.3 0.94 0.56 0.44 0.91 0.61 0.39 0.94 0.66 0.34 0.73 0.65 0.33 0.78 16 0.7 0.81 0.19 0.72 0.76 0.24 0.95 0.6 0.4 0.91 0.65 0.35 0.93 0.58 0.42 0.69 0.67 0.29 0.83 17 0.9 0.69 0.31 0.68 0.64 0.36 0.95 0.63 0.37 0.89 0.63 0.37 0.93 0.67 0.33 0.74 0.65 0.35 0.82 18 0.7 0.61 0.39 0.68 0.5 0.5 0.95 0.46 0.54 0.91 0.59 0.41 0.92 0.42 0.58 0.78 0.5 0.47 0.83 19 1 0.84 0.16 0.7 0.79 0.21 0.95 0.65 0.35 0.92 0.67 0.33 0.94 0.69 0.31 0.79 0.72 0.25 0.85 20 1 0.78 0.22 0.6 0.8 0.2 0.94 0.69 0.31 0.89 0.68 0.32 0.93 0.62 0.38 0.76 0.71 0.27 0.8 T o m o r r o w la n d 21 1 0.74 0.26 0.6 0.83 0.17 0.95 0.77 0.23 0.88 0.72 0.28 0.93 0.65 0.35 0.75 0.74 0.25 0.8 22 1 0.8 0.2 0.65 0.64 0.36 0.95 0.68 0.32 0.88 0.66 0.34 0.93 0.61 0.39 0.71 0.67 0.31 0.81 23 0.9 0.58 0.42 0.65 0.56 0.44 0.93 0.51 0.49 0.87 0.57 0.43 0.92 0.65 0.35 0.71 0.57 0.42 0.8 24 0.9 0.59 0.41 0.62 0.51 0.49 0.51 0.58 0.42 0.89 0.59 0.41 0.93 0.65 0.35 0.76 0.58 0.41 0.7 25 1 0.69 0.31 0.59 0.63 0.37 0.94 0.57 0.43 0.91 0.59 0.41 0.93 0.6 0.4 0.75 0.61 0.38 0.8 26 0.8 0.55 0.45 0.69 0.46 0.54 0.95 0.49 0.51 0.89 0.57 0.43 0.92 0.64 0.36 0.73 0.53 0.45 0.82 27 0.8 0.76 0.24 0.6 0.7 0.3 0.92 0.62 0.38 0.87 0.7 0.3 0.92 0.62 0.38 0.77 0.68 0.31 0.8 28 0.9 0.68 0.32 0.59 0.64 0.36 0.93 0.65 0.35 0.85 0.63 0.37 0.91 0.62 0.38 0.69 0.64 0.35 0.77 29 0.6 0.76 0.24 0.69 0.73 0.27 0.93 0.66 0.34 0.85 0.49 0.51 0.93 0.56 0.44 0.71 0.62 0.33 0.81 30 0.1 0.57 0.43 0.74 0.35 0.65 0.93 0.33 0.67 0.89 0.42 0.58 0.92 0.7 0.3 0.76 0.44 0.48 0.84 Pearson correlation 0.62 0.54 0.56 0.52 0.09 0.69 Difference (Rating-P) Mean 0.22 0.25 0.28 0.29 0.31 0.26 Table 14. Opinion Mining evaluation over 30 reviews Another evaluation was carried out in order to compare the rating obtained by each movie (by considering 10 reviews per movie) with the rating suggested by our framework. Fig 11 shows how our suggestions were very consistent regarding to Grey‟s Anatomy. However, there are some differences with the results for Tomorrowland and House M.D, but not so distant. This is due to our results provide positive sentiment polarities and not specifically user ratings. Even though there are some differences, we can still obtain accurate results according to WNA or Domains dimensions. Fig 11. Review rating (rating mean) and OM rating (positiveness mean) comparison Grey's Anatomy House Tomorrowland IMDB Rating (Mean) 0.640 0.840 0.800 Affect Positiveness (Mean) 0.722 0.721 0.672 Domain Positiveness (Mean) 0.664 0.663 0.605 SUMO Positiveness (Mean) 0.602 0.601 0.585 Semantic Class Positiveness (Mean) 0.643 0.619 0.593 Synset Positiveness (Mean) 0.634 0.636 0.632 Mean Positiveness (Mean) 0.635 0.640 0.609 0.000 0.100 0.200 0.300 0.400 0.500 0.600 0.700 0.800 0.900 1.000 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT We want to remark that our framework is able to provide the OM outputs as inputs to a recommender system by different ways. One way is to provide the individual analytics considering each semantic resource (in this case SUMO, WND, WNA, SC, and the taxonomy of WN) separately. Another way could be considering as input all resources to get a single sentiment analysis result. And finally, we can provide all these outputs with a mean calculation. 9. Discussion According to the experiments conducted and the results obtained, we can conclude that our resource ISR- WN and the different approaches that take advantage of it are suitable of being used to acquire enough knowledge to make recommendations for TV-shows or films. Moreover, we are also able to classify texts into different labels with the purpose of ranking a set of TV-shows according to their similarities. Moreover, due to the fact that user reviews have associated a set of feelings to indicate likes or dislikes we are also able to detect emotions and classify texts into three different sections: positive, negative or objective. In some cases, there are contexts that mixed a set of emotions and can confuse the framework. This is the case of the review for “House M.D.”. However, we can establish the appropriate feelings by taking into account all the context information. We would like to mention another similar work such as (Briguez et al., 2014) that presents a novel framework for the specific domain of movie recommendation. This work proposes a complete set of postulates accounting for both quantitative and qualitative aspects of the movie domain implemented by means of DelP (Defeasible Logic Programming). Related to our work, there are some differences: first of all, we want to remark that we present a semantic framework as a tool to help recommender systems so, taking into account our proposal, we can mention that the way we provide information to recommend similar movies is established by measuring different dimensions (i.e. feelings, sentiment polarities, word senses, textual similarities), on the contrary in (Briguez et al., 2014) their system uses specific rules to determine which characteristics have to be shared. Despite the fact that the process of building specific rules helps to obtain better user recommendations, it is designed to one specific domain. In case of applying this system to another domain it would be necessary to create additional rules to obtain accurate results. On the other hand, our approach can be easily adapted to other domains or languages (since WN is a nucleus for other WN languages 42 ) with minimal changes. Another difference with our work is that in (Briguez et al., 2014) the system provides a combination of quantitative and qualitative aspects however, we only have focused on quantitative aspects. The results obtained after mixing qualitative and quantitative aspects demonstrate that the incorporation of qualitative aspects introduce significant improvements. In fact, it would be a reasonable feature to take into consideration for future works. Related to evaluation results, we cannot provide an in depth comparison because our datasets are different. Moreover, as far as we are concerned we have not found any example in the literature using exactly the same dataset employed in this work. Even though our experiments have been carried out with user reviews for TV-shows we can deal with other domains. Moreover, we can deal with different languages by using a specific version of WN for each language. This fact has been demonstrated with our participation in different competitions where our framework was evaluated over different languages. One interesting usage of our proposal is shown in Fig 12. Through our semantic labels and features provided by our framework, one user could navigate among different TV-shows or movies that have semantic similarities and thus, provide a great experience of knowing automatically which are the movies or shows more appropriated to each one. 42 http://globalwordnet.org/wordnets-in-the-world/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Fig 12. Conceptual similarities among different IMDB reviews 10. Conclusions and future works In this work our goal has been to enrich texts by measuring similarities among different comments/reviews, grading the positiveness or negativeness by using textual information and using also this information to create a recommendation rate according to people interests. To do this, we have developed a framework that allows the integration of several semantic resources in order to provide to recommender systems enough knowledge for being able to: extract similarities among texts, measure sentiment polarities and obtain a semantic analysis to understand contexts meanings. In order to build the knowledge database of the framework we have selected a set of resources by conducting an in-depth analysis of the different semantic resources used in NLP. Just like we have discussed in previous sections, several authors have created integrated resources but most of them have been based on linguistic features rather than on semantic or conceptual features. In this research work we have therefore considered previous studies in order to develop a multidimensional knowledge network that integrates different semantic resources. Therefore, an integrated resource (ISR-WN) has been proposed. This resource integrates semantic resources (WN, WND, WNA, SUMO, SC and new semantic relations provided by XWN and a sentiment analysis resource (SWN), using WN, from versions 1.6 and 2.0). The knowledge database obtained in the integration process was evaluated in order to detect any problems in the alignment by using different intermediate versions of WN. So, depending on the nucleus of WN, WND and SUMO used, an accuracy of 100% was obtained, while accuracies of 99.76% and 100% were obtained for SC related to WN1.6 and 2.0, respectively. In addition to this research we have included a set of research works that contains interesting knowledge bases used to deal with NLP tasks. One of the most interesting resources is SWN. It supplies the positivity of a sense, domain, category, emotion or semantic class. It is important to stress that this resource has been used in several research works to take advantage of its semantic multidimensionality. Based on the semantic multidimensionality that ISR-WN provides we are able to extract new information to classify, obtain opinions or recommend similar texts in open domains. Basically, we have applied four different tasks to add new information from different point of views: extract similarities among texts, measure sentiment polarities and obtain a semantic analysis for understanding contexts meanings. Moreover, we have illustrated a case of study with information related to movies and TV series reviews to demonstrate that our framework works properly. As future work we plan to enrich the ISR- WN resource with collocation sense relations 43 . This type of information provides better results in tasks such as WSD (Gutiérrez, 2012). In (Gutiérrez et al., 2011c) ISR-WN was used to combine conceptualizations with polarities. We propose adding resources such as Micro-WNOp 44 (a corpus labelled with sentiments with around 1,105 WN synsets). Moreover, we are working to align ISR-WN with WNs in other languages in order to create a multilingual resource. 43 Synset pairs that commonly appear together in corpus. 44 http://www.unipv.it/micrownop Love Love Sensation n Sensation Medicine Medicine ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT In order to add more knowledge to ISR-WN, we propose the integration of different ontologies by using the RDF model to align WN synsets with ontological concepts. This kind of information will help us to apply different techniques (i.e. Domain Classification) with specific domains such as: medicine, technology, tourism, pharmaceutical, etc. Moreover, we want to integrate BabelNet (Navigli and Ponzetto, 2010) (a resource that connects the multilingual web encyclopaedia Wikipedia with WN) with ISR-WN. BabelNet was considered as a sense inventory in Semeval2013 45 in Task 12: Multilingual Word Sense Disambiguation. Finally, we plan to use the ISR-WN functionalities in order to help summarization systems to customize text summaries using different domains, categories, emotions, sentiment polarities, etc. Acknowledgments This research work has been partially funded by the University of Alicante, Generalitat Valenciana, Spanish Government and the European Commission through the projects, TIN2015-65136-C2-2-R, TIN2015-65100-R, SAM (FP7-611312), and PROMETEOII/2014/001. References Agirre, E., Cer, D., Diab, M. & Gonzalez-Agirre, A. (2012) SemEval-2012 Task 6: A Pilot on Semantic Textual Similarity. {*SEM 2012}: The First Joint Conference on Lexical and Computational Semantics -- Volume 1: Proceedings of the main conference and the shared task, and Volume 2: Proceedings of the Sixth International Workshop on Semantic Evaluation {(SemEval 2012)}. Montreal, Canada, Association for Computational Linguistics. Agirre, E., Lacalle, O. L. D., Fellbaum, C., Hsieh, S.-K., Tesconi, M., Monachini, M., Vossen, P. & Segers, R. (2010) SemEval-2010 task 17: All-words word sense disambiguation on a specific domain. Proceedings of the 5th International Workshop on Semantic Evaluation. Los Angeles, California, Association for Computational Linguistics. Agirre, E. & Soroa, A. (2009) Personalizing PageRank for Word Sense Disambiguation. Proceedings of the 12th conference of the European chapter of the Association for Computational Linguistics (EACL-2009). Athens, Greece. Atserias, J., Villarejo, L., Rigau, G., Agirre, E., Carroll, J., Magnini, B. & Vossen, P. (2004) The MEANING Multilingual Central Repository. Proceedings of the Second International Global WordNet Conference (GWC’04). . Brno, Czech Republic. Baccianella, S., Esuli, A. & Sebastiani, F. (2010) SENTIWORDNET 3.0: An Enhanced Lexical Resource for Sentiment Analysis and Opinion Mining. IN 2010, L. (Ed.) 7th Language Resources and Evaluation Conference. Valletta, MALTA. Bentivogli, L., Forner, P., Magnini, B. & Pianta, E. (2004) Revising the WORDNET DOMAINS Hierarchy: semantics, coverage and balancing. Proceedings of COLING 2004 Workshop on "Multilingual Linguistic Resources". Geneva, Switzerland. . Briguez, C. E., Budan, M. C. D., Deagustini, C. A. D., Maguitman, A. G., Capobianco, M. & Simari, G. R. (2014) Argument-based mixed recommenders and their application to movie suggestion. Expert Systems with Applications, 41, 6467-6482. Brill, E. (1995) Transformation-based error-driven learning and natural language processing: a case study in part-of- speech tagging. MIT Press. Cotton, S., Edmonds, P., Kilgarriff, A. & Palmer, M. (2001) English All word. IN LINGUISTICS, A. F. C. (Ed.) SENSEVAL-2: Second International Workshop on Evaluating Word Sense Disambiguation Systems. Toulouse, France, Association for Computational Linguistics. Chávez, A., Dávila, H., Gutiérrez, Y., Collazo, A., Abreu, J. I., Fernández Orquín, A., Montoyo, A. & Muñoz, R. (2013) UMCC_DLSI: Textual Similarity based on Lexical-Semantic features. Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task: Semantic Textual Similarity. Atlanta, Georgia, USA, Association for Computational Linguistics. Dávila, H., Fernández, A., Gutiérrez, Y., Muñoz, R., Montoyo, A. & Vázquez, S. (2012) Semantic Information Extraction method on ontologies. SEPLN 2012: XXVIII CONGRESO DE LA SOCIEDAD ESPAÑOLA PARA EL PROCESAMIENTO DEL LENGUAJE NATURAL. Castellón, Spain. Dávila, H., Fernández Orquín, A., Chávez, A., Gutiérrez, Y., Collazo, A., Abreu, J. I., Montoyo, A. & Muñoz, R. (2013) UMCC_DLSI-(EPS): Paraphrases Detection Based on Semantic Distance. Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Proceedings of the Seventh International 45 http://www.cs.york.ac.uk/semeval-2013/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Workshop on Semantic Evaluation (SemEval 2013). Atlanta, Georgia, USA, Association for Computational Linguistics. Dorr, B. J. & Castellón, M. A. M. A. I. (1997) Spanish EuroWordNet and LCS-Based Interlingual MT. AMTA/SIG-IL First Workshop on Interlinguas. San Diego, CA. Esuli, A. & Sebastiani, F. (2006) SentiWordNet: A Publicly Available Lexical Resource for Opinion Mining. IN 2006, L. (Ed.) Fifth international conference on Languaje Resources and Evaluation Genoa - ITaly. Fellbaum, C. (1998) WordNet. An Electronic Lexical Database, University of Cambridge. Fernández, A., Gutiérrez, Y., Dávila, H., Chávez, A., González, A., Estrada, R., Castañeda, Y., Vázquez, S., Montoyo, A. & Muñoz, R. (2012a) UMCC_DLSI: Multidimensional Lexical-Semantic Textual Similarity. {*SEM 2012}: The First Joint Conference on Lexical and Computational Semantics -- Volume 1: Proceedings of the main conference and the shared task, and Volume 2: Proceedings of the Sixth International Workshop on Semantic Evaluation {(SemEval 2012)}. Montreal, Canada, Association for Computational Linguistics. Fernández, A., Gutiérrez, Y., Muñoz, R. & Montoyo, A. (2012b) Approaching Textual Entailment with Sentiment Polarity. ICAI'12 - The 2012 International Conference on Artificial Intelligence. Las Vegas, Nevada, USA. Forner, P. (2005) WordNet Domains 2.0. ITC-irst, Povo-Trento, Italy. Gediminas, A. & Alexander, T. (2005) Toward the Next Generation of Recommender Systems: A Survey of the State-of-the-Art and Possible Extensions. IEEE Educational Activities Department. Genesereth, M. R. & Fikes, R. E. (1992) Knowledge Interchange Format. IN UNIVERSITY, C. S. D. S. (Ed.) version 3.0 Reference Manual ed. Stanford, Computer Science Department. Gliozzo, A., Strapparava, C. & Dagan, I. (2004) Unsupervised and Supervised Exploitation of Semantic Domains in Lexical Disambiguation. Computer Speech and Language. Gutiérrez, Y. (2012) Análisis Semántico Multidimensional aplicado a la Desambiguación del Lenguaje Natural. Departamento de Lenguajes y Sistemas Informáticos. Alicante, Alicante. Gutiérrez, Y., Castañeda, Y., González, A., Estrada, R., Piug, D. D., Abreu, J. I., Pérez, R., Fernández Orquín, A., Montoyo, A., Muñoz, R. & Camara, F. (2013) UMCC_DLSI: Reinforcing a Ranking Algorithm with Sense Frequencies and Multidimensional Semantic Resources to solve Multilingual Word Sense Disambiguation. Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Proceedings of the Seventh International Workshop on Semantic Evaluation (SemEval 2013). Atlanta, Georgia, USA, Association for Computational Linguistics. Gutiérrez, Y., Fernández, A., Montoyo, A. & Vázquez, S. (2010a) Integration of semantic resources based on WordNet. IN 2010, S. (Ed.) XXVI Congreso de la Sociedad Española para el Procesamiento del Lenguaje Natural. Universidad Politécnica de Valencia, Valencia, SEPLN 2010. Gutiérrez, Y., Fernández, A., Montoyo, A. & Vázquez, S. (2010b) UMCC-DLSI: Integrative resource for disambiguation task. Proceedings of the 5th International Workshop on Semantic Evaluation. Uppsala, Sweden, Association for Computational Linguistics. Gutiérrez, Y., Fernández, A., Montoyo, A. & Vázquez, S. (2011a) Enriching the Integration of Semantic Resources based on WordNet. Procesamiento del Lenguaje Natural, 47, 249-257. Gutiérrez, Y., Vázquez, S. & Montoyo, A. (2011b) Improving WSD using ISR-WN with Relevant Semantic Trees and SemCor Senses Frequency. Proceedings of the International Conference Recent Advances in Natural Language Processing 2011. Hissar, Bulgaria, RANLP 2011 Organising Committee. Gutiérrez, Y., Vázquez, S. & Montoyo, A. (2011c) Sentiment Classification Using Semantic Features Extracted from WordNet-based Resources. Proceedings of the 2nd Workshop on Computational Approaches to Subjectivity and Sentiment Analysis (WASSA 2.011). Portland, Oregon., Association for Computational Linguistics. Gutiérrez, Y., Vázquez, S. & Montoyo, A. (2011d) Word Sense Disambiguation: A Graph-Based Approach Using N- Cliques Partitioning Technique. IN MUÑOZ, R., MONTOYO, A. & MÉTAIS, E. (Eds.) Natural Language Processing and Information Systems. Springer Berlin / Heidelberg. Izquierdo, R. (2010) Una Aproximación a la Desambiguación del Sentido de las Palabras Basada en Clases Semánticas y Aprendizaje Automático. Departamento de Lenguajes y Sistemas Informáticos. Alicante, Universidad de Alicante. Izquierdo, R., Suárez, A. & Rigau, G. (2007) A Proposal of Automatic Selection of Coarse-grained Semantic Classes for WSD. Procesamiento del Lenguaje Natural, 39, 189-196. Izquierdo, R., Suárez, A. & Rigau, G. (2010) GPLSI-IXA: Using Semantic Classes to Acquire Monosemous Training Examples from Domain Texts Proceedings of the 5th International Workshop on Semantic Evaluation. Uppsala, Sweden, Association for Computational Linguistics. Leacock, C. & Chodorow, M. (1998) Using Corpus Statistics and WordNet Relations for Sense Identification. Computational Linguistics. Luisa Bentivogli, P. F., Bernardo Magnini, Emanuele Pianta (2005) Revising the WORDNET DOMAINS Hierarchy: semantics, coverage and balancing. ITC-irst – Istituto per la Ricerca Scientifica e Tecnologica Via Sommarive 18, Povo – Trento, Italy, 38050. Magnini, B. & Cavaglia, G. (2000) Integrating Subject Field Codes into WordNet. Proceedings of Third International Conference on Language Resources and Evaluation (LREC-2000). Magnini, B., Satrapparava, C., Pezzulo, G. & Gliozzo, A. (July 2002) The Role of Domains Informations in Word Sense Disambiguatios. Treto, Cambridge University Press. Magnini, B., Strapparava, C., Pezzulo, G. & Gliozzo, A. (2002) Comparing Ontology-Based and Corpus-Based Domain Annotations in WordNet. Proceedings of the First International WordNet Conference. Mysore, India. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Marco De, G., Pasquale, L., Giovanni, S. & Pierpaolo, B. (2008) Integrating tags in a semantic content-based recommender. Proceedings of the 2008 ACM conference on Recommender systems. Lausanne, Switzerland, ACM. Miller, G. A., Beckwith, R., Fellbaum, C., Gross, D. & Miller, K. (1990) Five papers on WordNet. Princenton University, Cognositive Science Laboratory. Navigli, R. (2009) Word sense disambiguation: A survey. ACM Comput. Surv., 41, 10:1--10:69. Navigli, R. & Ponzetto, S. P. (2010) BabelNet: Building a Very Large Multilingual Semantic Network. Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics. Uppsala, Sweden, Association for Computational Linguistics. Niles, I. (2001) Mapping WordNet to the SUMO Ontology. Teknowledge Corporation. Niles, I. & Pease, A. (2001) Origins of the IEEE Standard Upper Ontology. Working Notes of the IJCAI-2001 Workshop on the IEEE Standard Upper Ontology. Seattle, Washington, USA. Niles, I. & Pease, A. (2003) Linking Lexicons and Ontologies: Mapping WordNet to the Suggested Upper Merged Ontology. Pease, A. (2007) Standard Upper Ontology Knowledge Interchange Format. Perner, P., Candillier, L., Meyer, F. & Boulle, M. (2007) Comparing State-of-the-Art Collaborative Filtering Systems. Machine Learning and Data Mining in Pattern Recognition. Springer Berlin Heidelberg. Peter, D. T. (2001) Mining the Web for Synonyms: PMI-IR versus LSA on TOEFL. Proceedings of the 12th European Conference on Machine Learning. Springer-Verlag. Pianta, E., Bentivogli, L. & Girardi, C. (2002) MultiWordNet. Developing an aligned multilingual database. Proceedings of the 1st International WordNet Conference. Mysore, India. Rao, D., Mcnamee, P. & Dredze, M. (2013) Entity linking: Finding extracted entities in a knowledge base. Multi- source, multilingual information extraction and summarization. Springer. Russell, S. & Norvig, P. (1994) A Modern, Agent-Oriented Approach to Introductory Artificial Intelligence. Sanda M. Harabagiu, G. A. M., Dan I. Moldovan (1999) Wordnet 2-A morphologically and semantically enhanced resource. SIGLEX99: Standardizing Lexical Resources. Sara, T. & Daniele, P. (2009) New features for FrameNet: WordNet mapping. Proceedings of the Thirteenth Conference on Computational Natural Language Learning. Boulder, Colorado, Association for Computational Linguistics. Ševčenko, M. (2003) Online Presentation of an Upper Ontology. CTU Prague, Dept of Computer Science. Shinde, S. K. & Kulkarni, U. (2012) Hybrid personalized recommender system using centering-bunching based clustering algorithm. Expert Systems with Applications, 39, 1381-1387. Sowa, J. F. (1999) Knowledge representation: logical, philosophical, and computational foundations. Course Technology. Strapparava, C. & Valitutti, A. (2004) WordNet-Affect: an affective extension of WordNet. Proceedings of the 4th International Conference on Language Resources and Evaluation (LREC 2004). Lisbon. Turney, P. D. & Littman, M. L. (2003) Measuring Praise and Criticism: Inference of Semantic Orientation from Association. ACM Transactions on Information Systems (TOIS), 21, 315–346. Udi, M., Ash, P. & John, R. (2000) Experience with personalization of Yahoo! , ACM. Valitutti, A., Strapparava, C. & Stock, O. (Eds.) (2004) Developing Affective Lexical Resources, ITC-irst, Trento, Italy, PsychNology Journal. Vázquez, S. (2009) Resolución de la ambigüedad semántica mediante métodos basados en conocimiento y su aportación a tareas de PLN. Depto. de Lenguajes y Sistemas Informáticos. Alicante, Spain., Universidad de Alicante. Vázquez, S., Montoyo, A. & Rigau, G. (2004) Using Relevant Domains Resource for Word Sense Disambiguation. IC-AI’04. Proceedings of the International Conference on Artificial Intelligence. Ed: CSREA Press. Las Vegas, E.E.U.U. Vossen, P. (1998) EuroWordNet: A Multilingual Database with Lexical Semantic Networks, Dordrecht, Kluwer Academic Publishers. Vossen, P., Peters, W. & Gonzalo, J. (1999) Towards a Universal Index of Meaning. proceedings of the ACL-99 Siglex workshop. University of Maryland. Walter, C.-N., Maria Luisa, H.-A., Rafael, V.-G. & Francisco, G.-S. (2012) Social knowledge-based recommender system. Application to the movies domain. Pergamon Press, Inc. Zhibiao, W. & Martha, P. (1994) Verbs semantics and lexical selection. Proceedings of the 32nd annual meeting on Association for Computational Linguistics. Las Cruces, New Mexico, Association for Computational Linguistics. Zouaq, A., Gagnon, M. & Ozell, B. (2009) A SUMO-based Semantic Analysis for Knowledge Extraction. Proceedings of the 4th Language & Technology Conference. Poznań, Poland. 
work_vw3k7kxkz5dxpm3xbsboiqxzza	----	Fuzzy data envelopment analysis: A fuzzy expected value approach Fuzzy data envelopment analysis: A fuzzy expected value approach q Ying-Ming Wang a,⇑, Kwai-Sang Chin b a School of Public Administration, Fuzhou University, Fuzhou 350002, PR China b Department of Manufacturing Engineering and Engineering Management, City University of Hong Kong, 83 Tat Chee Avenue, Kowloon Tong, Hong Kong a r t i c l e i n f o Keywords: Fuzzy data envelopment analysis Fuzzy expected values Double frontier analysis (DFA) Flexible manufacturing system a b s t r a c t Performance assessment often has to be conducted under uncertainty. This paper proposes a ‘‘fuzzy expected value approach’’ for data envelopment analysis (DEA) in which fuzzy inputs and fuzzy outputs are first weighted, respectively, and their expected values then used to measure the optimistic and pes- simistic efficiencies of decision making units (DMUs) in fuzzy environments. The two efficiencies are finally geometrically averaged for the purposes of ranking and identifying the best performing DMU. The proposed fuzzy expected value approach and its resultant models are illustrated with three numer- ical examples, including the selection of a flexible manufacturing system (FMS). � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Traditional data envelopment analysis (DEA) (Charnes, Cooper, & Rhodes, 1978) requires crisp input and output data, which may not always be available in real word applications. Significant ef- forts have been made to handle fuzzy input and fuzzy output data in DEA. For example, Sengupta (1992) incorporated fuzziness into DEA by defining tolerance levels for both the objective function and violations of constraints and proposed a fuzzy mathematical programming approach. Triantis and Girod (1998) transformed fuzzy input and fuzzy output data into crisp data using member- ship function values and suggested a mathematical programming approach in which efficiency scores were computed for different values of membership functions and then averaged. Guo and Tanaka (2001) converted fuzzy constraints such as fuzzy equalities and fuzzy inequalities into crisp constraints by predefining a pos- sibility level and using the comparison rule for fuzzy numbers and presented a fuzzy CCR model. León, Liern, Ruiz, and Sirvent (2003) suggested a fuzzy BCC model based on the same idea. Lertworasirikul, Fang, Joines, and Nuttle (2003) proposed a pos- sibility DEA model for fuzzy DEA. In the special case that fuzzy data are trapezoidal fuzzy numbers, the possibility DEA model became a linear programming (LP) model. They (Lertworasirikul, Fang, Joines, & Nuttle, 2003) also presented a credibility approach as an alternative way for solving fuzzy DEA problems. The possibility and credibility approaches were further extended to fuzzy BCC model in Lertworasirikul, Fang, Nuttle, and Joines (2003) by the same authors. Wu, Yang, and Liang (2006) applied the possibility DEA model for efficiency analysis of cross-region bank branches in Canada. Garcia, Schirru, and Melo (2005) utilized the possibility DEA model for failure mode and effects analysis (FMEA) and presented a fuzzy DEA approach to determining ranking indices among failure modes. Wen and Li (2009) employed credibility measure to represent fuzzy CCR model as an uncertain program- ming and solved it with a hybrid intelligent algorithm which inte- grates fuzzy simulations and genetic algorithms. Kao and Liu (2000a, 2000b, 2003, 2005) transformed fuzzy in- put and fuzzy output data into intervals by using a-level sets and Zadeh’s extension principle, and built a family of crisp DEA models for the intervals. Based on their crisp DEA models for a-level sets, Liu (2008) and Liu and Chuang (2009) took further into consider- ation the concept of assurance region (AR) and developed a fuzzy DEA/AR model for the selection of flexible manufacturing systems (FMSs) and the assessment of university libraries, respectively. Saati, Menariani, and Jahanshahloo (2002) defined fuzzy CCR mod- el as a possibilistic-programming problem and transformed it into an interval programming by specifying a a-level set. Their ap- proach was further extended in Saati and Memariani (2005) so that all decision making units (DMUs) could be evaluated with a com- mon set of weights under a given a-level set. Entani, Maeda, and Tanaka (2002) and Wang, Greatbanks, and Yang (2005) also chan- ged fuzzy input and fuzzy output data into intervals by using a- level sets, but suggested two different interval DEA models. Dia (2004) proposed a fuzzy DEA model based upon fuzzy arith- metic operations and fuzzy comparisons between fuzzy numbers. The model requires the decision maker (DM) to specify a fuzzy aspiration level and a safety a-level so that the fuzzy DEA model could be transformed into a crisp DEA model for solution. Wang, Luo, and Liang (2009) constructed two fuzzy DEA models from the perspective of fuzzy arithmetic to deal with fuzziness in input 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.03.049 q The work described in this paper was supported by the National Natural Science Foundation of China (NSFC) under the Grant Nos. 70771027 and 70925004 and also partially supported by a grant from CityU (Project No. 7002571). ⇑ Corresponding author. Tel.: +86 591 22866681; fax: +86 591 22866677. E-mail addresses: msymwang@hotmail.com, ymwang@fzu.edu.cn (Y.-M. Wang). Expert Systems with Applications 38 (2011) 11678–11685 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.03.049 mailto:msymwang@hotmail.com mailto:ymwang@fzu.edu.cn http://dx.doi.org/10.1016/j.eswa.2011.03.049 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa and output data in DEA. The two fuzzy DEA models were both for- mulated as linear programs and could be solved to determine fuzzy efficiencies of DMUs. Triantis (2003) introduced a fuzzy DEA approach to calculate fuzzy non-radial technical efficiencies and implemented the ap- proach in a newspaper preprint insertion manufacturing process. Soleimani-damaneh, Jahanshahloo, and Abbasbandy (2006) addressed some computational and theoretical pitfalls of the fuzzy DEA models developed in Kao and Liu (2000a), León et al. (2003) and Lertworasirikul et al. (2003) and provided a fuzzy DEA model to yield crisp efficiencies for the DMUs with fuzzy input and fuzzy output data. Jahanshahloo, Soleimani-damaneh, and Nasrabadi (2004) extended a slack-based measure (SBM) of efficiency in DEA to fuzzy settings and developed a two-objective nonlinear DEA model for fuzzy DEA. Existing fuzzy DEA models exhibit some drawbacks. For in- stance, fuzzy DEA models derived from the direct fuzzification of crisp DEA models ignore the fact that a fuzzy fractional program cannot be transformed into an LP model in the traditional way that we do for a crisp fractional program. Fuzzy DEA models built on the basis of a-level sets require the solution of a series of LP models and thus considerable computational efforts. Fuzzy DEA models constructed from the perspective of fuzzy arithmetic demand a ra- tional yet easy-to-use ranking approach for fuzzy efficiencies. To overcome these drawbacks, we propose in this paper a ‘‘fuzzy ex- pected value approach’’ for fuzzy DEA, which first weights fuzzy in- puts and fuzzy outputs, respectively, and then utilizes their expected values for measuring the performances of DMUs in fuzzy environments. The paper is organized as follows. Section 2 introduces the mea- sures of fuzzy expected values and develops fuzzy expected value models for fuzzy DEA. Section 3 illustrates the developed fuzzy ex- pected value models with three numerical examples, including the selection of a FMS. Section 4 concludes the paper. 2. Fuzzy expected values and fuzzy DEA models Fuzzy numbers are convex fuzzy sets, characterized by given intervals of real numbers, each interval with a grade of member- ship between 0 and 1. The most commonly used fuzzy numbers are triangular and trapezoidal fuzzy numbers defined by the following membership functions, respectively: leA 1ðxÞ¼ ðx � aÞ=ðb � aÞ; a 6 x 6 b; ðd � xÞ=ðd � bÞ; b 6 x 6 d; 0; otherwise; 8>< >: ð1Þ leA 2ðxÞ¼ ðx � aÞ=ðb � aÞ; a 6 x 6 b; 1; b 6 x 6 c; ðd � xÞ=ðd � cÞ; c 6 x 6 d; 0; otherwise: 8>>>< >>>: ð2Þ For brevity, triangular and trapezoidal fuzzy numbers are often denoted as (a, b, d) and (a, b, c, d). It is evident that triangular fuzzy numbers are special cases of trapezoidal fuzzy numbers with b = c. For any two positive trapezoidal fuzzy numberseA ¼ðaL; aM; aN; aUÞ and eB ¼ðbL; bM; bN; bUÞ, fuzzy addition and fuzzy multiplication on eA and eB are respectively defined aseA þ eB ¼ðaL þ bL; aM þ bM; aN þ bN; aU þ bUÞ and eA � eB �ðaL bL; aM bM; aN bN; aU bUÞ. Let n be a fuzzy variable with a membership function l: R !½0; 1� and r be a real number. The possibility and the necessity of {n P r} are respectively defined by Posfn P rg¼ sup xPr lðxÞ; ð3Þ Necfn P rg¼ 1 � Posfn < rg¼ 1 � sup x<r lðxÞ; ð4Þ which show respectively the possibility and the necessity degrees to which n is not smaller than r. Pos and Nec are a pair of dual fuzzy measures in the sense that Pos{A} = 1 � Nec{AC} with AC is the com- plement of A. Based upon the possibility and the necessity mea- sures, credibility measure is defined as Crfn P rg¼ 1 2 Posfn P rgþ Necfn P rgð Þ: ð5Þ The fuzzy expected value of n can thus be defined as (Liu & Liu, 2002) E½n� ¼ Z 1 0 Crfn P rgdr � Z 0 �1 Crfn 6 rgdr: ð6Þ It has been shown (Liu & Liu, 2002) that if fuzzy variable n is re- placed with a random variable whose probability density function is / and Cr is replaced with the probability measure Pr, then there exists R1 0 Prfn P rgdr � R0 �1 Prfn 6 rgdr ¼ Rþ1 �1 x/ðxÞdx, which is exactly the expected value of the random variable n. It is also shown (Liu & Liu, 2002) that if n is a trapezoidal fuzzy variable (r1, r2, r3, r4), then the expected value of n is (1/4)(r1 + r2 + r3 + r4). In particular, if n is a triangular fuzzy variable (r1, r2, r3), then the expected value of n is (1/4)(r1 + 2r2 + r3). Suppose we have n DMUs to be evaluated in terms of m inputs and s outputs. Let xij (i = 1, . . . , m) and yrj (r = 1, . . . , s) be the input and out- put data of DMUj (j = 1, . . . , n). Without loss of generality, all input and output data xij and yrj are assumed to be uncertain and character- ized by trapezoidal fuzzy numbers ~xij ¼ xLij; x M ij ; x N ij ; x U ij � � and ~yrj ¼ yLrj; y M rj ; y N rj; y U rj � � with xLij P 0 and y L rj P 0 for i = 1 to m, r = 1 to s, and j = 1 to n. Crisp data and triangular fuzzy data are treated as special cases of trapezoidal fuzzy data ~xij and ~yrj with xLij ¼ x M ij ¼ x N ij ¼ x U ij ; y L rj ¼ y M rj ¼ y N rj ¼ y U rj , and x M ij ¼ x N ij ; y M rj ¼ y N rj , respectively. The total fuzzy weighted output (FWO) and the total fuzzy weighted input (FWI) of DMUj are given by FWOj ¼ Xs r¼1 ~ur ~yrj ¼ Xs r¼1 uLr; u M r ; u N r ; u U r � � � yLrj; y M rj ; y N rj; y U rj � � ; ð7Þ FWIj ¼ Xm i¼1 ~v i~xij ¼ Xm i¼1 v Li ; v M i ; v N i ; v U i � � � xLij; x M ij ; x N ij ; x U ij � � ; ð8Þ where ~ur ¼ uLr; uMr ; uNr ; uUr � � and ~v i ¼ v Li ; v M i ; v N i ; v U i � � are fuzzy weights for fuzzy input ~xij and fuzzy output ~yrj, respectively. Accord- ing to fuzzy addition and fuzzy multiplication operations on two positive fuzzy numbers, (7) and (8) can be approximately expressed as FWOj � Xs r¼1 uLr y L rj; Xs r¼1 uMr y M rj ; Xs r¼1 uNr y N rj; Xs r¼1 uUr y U rj ! ; ð9Þ FWIj � Xm i¼1 v Li x L ij; Xm i¼1 v Mi x M ij ; Xm i¼1 v Ni x N ij ; Xm i¼1 v Ui x U ij ! ; ð10Þ which can be viewed as two trapezoidal fuzzy variables, whose expected values can therefore be determined as EðFWOjÞ¼ 1 4 Xs r¼1 uLr y L rj þ Xs r¼1 uMr y M rj þ Xs r¼1 uNr y N rj þ Xs r¼1 uUr y U rj ! ¼ 1 4 Xs r¼1 uLr y L rj þ u M r y M rj þ u N r y N rj þ u U r y U rj � � ; ð11Þ Y.-M. Wang, K.-S. Chin / Expert Systems with Applications 38 (2011) 11678–11685 11679 https://isiarticles.com/article/16224 
work_vwu6oibbn5c3xcnly3yuyubxqi	----	* * "AN EVALUATION AND SELECTION METHODOLOGY FOR EXPERT SYSTEM SHELLS" by Louis LE BLANC and Tawfik JELASSI** N° 90/39/TM Associate Professor of Policy and Administration, School of Public and Environmental Affairs, Indiana University, Bloomington, Indiana 47405 U.S.A. Associate Professor of Information Systems, INSEAD, Boulevard de Constance, Fontainebleau 77305 Cedex, France Printed at INSEAD, Fontainebleau, France AN EVALUATION AND SELECTION METHODOLOGY FOR EXPERT SYSTEM SHELLS * Louis A. Le Blanc Policy and Administration Faculty School of Public and Environmental Affairs Indiana University Bloomington, Indiana 47405 U.S.A. and Tawfik Jelassi Technology Management Area INSEAD Boulevard de Constance 77305 Fontainebleau France May, 1990 * Forthcoming in Expert Systems with Applications: The International Journal, 1991. AN EVALUATION AND SELECTION METHODOLOGY FOR EXPERT SYSTEM SHELLS ABSTRACT This paper illustrates an evaluation and selection methodology for expert system (ES) shells. The methodology incorporates three stages: 1) ES shell screening; 2) shell evaluation; and, 3) assurance of final ES shell selection. Initially, developing a short list through screening of commercial shell products determines whether appropriate software exists and narrows the field of available expert system software products for detailed consideration. The second stage determines which of the remaining ES shells (the finalists) best meets the needs of the organization, from both functional and technical perspectives. The final stage compares user requirements with the features of the selected ES software by defining how these requirements will be satisfied by building expert system applications with the selected product. The methodology also considers the possibility that, at any stage of the process, no expert system shell is suitable and that a system must be developed with programming languages such as LISP, PROLOG, or some conventional programming language. AN EVALUATION AND SELECTION METHODOLOGY FOR EXPERT SYSTEM SHELLS TABLE OF CONTENTS 1.0 Introduction 1.1 Overview of Expert System Shells 1.2 ES Shell Definition 1.3 The ES Software Selection Problem 1.4 Structure of the Paper 2.0 A Decision Methodology for ES Software Selection 2.1 ES Shell Screening 2.1.1 Identify Candidate Software 2.1.2 Screening Criteria 2.1.3 Pick Finalists 2.2 ES Shell Evaluation 2.2.1 Expand Evaluation Criteria 2.2.2 Obtain Package Information 2.2.3 Evaluate ES Shells 2.2.4 Additional Selection Requirements 2.3 Specific ES Design 2.3.1 Alter Functional Requirements 2.3.2 ES Software Modifications and Supporting Programs 2.3.3 Finalize ES Shell Selection 3.0 Summary and Conclusion Page 1 AN EVALUATION AND SELECTION METHODOLOGY FOR EXPERT SYSTEM SHELLS 1.0 INTRODUCTION 1.1 Overview of Expert System Shells An expert system (ES) is a computer program for capturing human knowledge of a specific problem domain and delivering for decision making purposes. This expertise is often represented in the form of if-then rules, though it may take other forms of knowledge representation, such as "semantic networks" or "frames" (see, e.g., Harmon and King, 1985). Many organizations are developing ES applications with only minimal assistance from their information systems department by using the relatively inexpensive and widely available "off-the-shelf" ES shells. Historically, rather than develop a custom-built ES application, it was possible at times to borrow extensively from a previously constructed ES. This practice has resulted in the development of software that is known as ES shells or skeletal systems. The shells are ES with their knowledge component (which contains the facts and rules of a specific problem domain) removed and leaving only the explanation, inference, and user interface components. Currently, these commercial software packages include a ready- built inference engine, user interface, and framework for the knowledge base. All the system developer has to do is fill in the rules and facts of the knowledge base, state the goals of the system, and make the right connections with the user interface module. This module can provide a variety of techniques to interact with an end user. Page 2 Increasingly, it also includes the ability to integrate with existing databases and transaction processing systems of an organization. The knowledge base and inference engine are the heart of the ES shell. The inference engine, which is the reasoning component of the ES, examines the knowledge base and determines what to do in order to achieve the goal which it has been instructed to perform. 1.2 SS Shell Definition An ra Shell is "the dialog structure and inference engine which, when linked to a knowledge base, functions as a fully operational expert system" (Liebowitz, 1988). It offers several advantages over building ES from scratch. In addition to saving significant resources, using an ES shell allows the knowledge engineer to concentrate on the development of the knowledge base, from which the expert system derives its power. An ES shell constitutes one of the three technological levels of software that comprise the ES development environment (Henderson, 1987; Sprague, 1980; and Turban, 1990). Shells are essentially integrated packages of software that provide a set of capabilities to build specific ES quickly, easily, and at low cost. The other two levels are: Specific Ea, which are operational systems that actually support and/or advise managers/end-users on a specific issue; and Ea tools, which are software elements that provide developers with the facility to construct both specific ES and ES shells. Examples of such ES tools include procedural programming languages, such as C and PASCAL, and artificial intelligence languages like LISP, PROLOG, and FORTH. (Insert Figure 1 About Here) Page 3 Figure 1 (adapted from Sprague, 1980) illustrates the three ES technological levels defined by Turban (1990) and the relationships between them. Specific ES applications can be developed either directly from ES tools or by adapting the ES shell to satisfy the application requirements. In the latter situation, the ES builder could employ the iterative design approach to rapidly select capabilities from the ones available in the ES shell (or delete unnecessary features) as required by the specific ES (represented by the iterative cycling between the ES shell and the specific ES in Figure 1). A "prototyping" methodology could be employed between the "low" technological level of ES tools and a specific ES application. Given the ES development capabilities of a shell, ES applications of all dimensions can be rapidly prototyped. Expert system shells are available for a number of hardware platforms, including mainframes, mini and microcomputers, as well as special-purpose machines such as parallel processors. According to Newquist (1988), a new generation of ES vendors is emerging and placing ES software (i.e., ES shells) into the information systems mainstream. As companies like Apple Computer, IBM, Cullinet Software, and McCormack & Dodge become involved with ES technology, the choices for organi- zations will multiply. The options will range from the choice of hardware platform to the choice of computer language and software interfaces. Page 4 1.3 The ES Software Selection Problem Publications which have addressed the issue of evaluating and selecting ES software (Waterman and Hayes-Roth, 1982; Hayes-Roth, Waterman and Lenat, 1983; Liebowitz, 1985; Harmon and King, 1985; Forman and Nagy, 1985 and 1987; Harmon, et al., 1989; and Vetter, 1989) identified criteria, especially user-related ones, which are critical in selecting a suitable ES shell. However, these authors did not suggest how to incorporate multiple user criteria, as well as technical attributes, into a complete and thorough evaluation and selection process. Furthermore, Lynch (1984 and 1985) and Meador and Mezger (1984) suggest that inadequate examination of prospective software packages leads to serious difficulties if not failures when implementing information systems. Although a number of approaches to selecting application software for transaction processing systems (TPS) and management information systems (MIS) have been proposed (Breslin, 1986; Curry and Bonner, 1983; Gray, 1987; Martin and McClure, 1983), some very critical factors were omitted. These factors include assuring that the selected software package is superior to a custom alternative, or that a screening process is provided to reduce the number of packages subjected to detailed evaluation. To advance the importance of evaluation and selection of ES software (as compared to that of other information systems), it should be noted that ES shells are used to develop multiple management support applications. From the same ES shell, an organization may develop specific ES applications ranging from consultative and training to task execution in functional areas as diverse as marketing, manufacturing and finance. However, other types of software is usually employed only Page 5 for a single application, for example a general ledger or a material requirements planning system which respectively only support a single functional area. To efficiently develop specific ES applications using the iterative design approach (i.e., prototyping), an ES shell need be available (Naumann and Jenkins, 1982; Leary, 1987). The basic objectives of the ES shell are: 1) to permit quick and easy development of a wide variety of specific ES; and, 2) facilitate the iterative development process by which specific ES can respond quickly to changes. In a top-down fashion in order to satisfy these overall objectives, an ES shell must satisfy a number of general criteria or requirements. These criteria can be categorized according to ES components, such as user interface, inference engine and knowledge base. Once the ES shell is established, an enterprise can implement specific ES rapidly. Therefore, the very critical software evaluation and selection process for ES shells should take place prior to any systems analysis and design (iterative development) efforts for specific ES applications (Turban, 1990). 1.4 Structure of the Paper This paper illustrates a method to select the most appropriate ES shell where multiple criteria exist not only from functional requirements but also from technical and vendor-support perspectives. As an essential part of the methodology, an initial stage determines whether an ES software product is even suitable for a particular application, or should an ES be developed from programming languages or other ES tools. Page 6 [Insert Figure 2 About Here] The proposed selection process in this paper also ensures that, at each successive stage of the methodology, an ES shell is superior to a custom-built ES application (see Figure 2). It continually reduces the number of ES software products under consideration until a final selection of a shell is made or constructing an ES application from tools is chosen as the best alternative. This paper is primarily addressed to academics interested in software selection methodologies as well as practitioners faced with ES-related problems. Section Two suggests a multiple criteria method- ology for ES shell selection. The three stages of this process - ES software screening, ES shell evaluation, and ES software assurance - are described. Section Three concludes the paper with some final remarks. 2.0 A DECISION METHODOLOGY FOR ES SOFTWARE SELECTION There are three principal stages in the proposed ES shell evaluation and selection methodology: 1) screening of prospective candidates and development of a short list of ES software packages; 2) selecting an ES shell, if any, which best suits the application requirements; and, 3) matching user requirements to the features of the selected shell and describing how these requirements will be satisfied through the building of prototypes for specific ES. The detailed procedures involved in each stage of the selection process are described in the following sections. Page 7 The complete methodology provides a project outline for the evaluation and selection of ES software. At the end of each stage, the methodology produces a deliverable - a well-defined output or result (i.e., the short list of software packages at the conclusion of the initial stage). Given the project nature of this software evaluation process, scheduling (i.e., determination of start and finish dates) for each stage, as well as the entire project, is possible. (A Gantt chart or other similar device can be used for this purpose.) Staffing requirements can also be determined, including both the number of staff members and their required skills. To evaluate and select ES software, a project team should be assembled with personnel from user departments as well as information systems professionals. 2.1 ES Software Screening During this first stage of the evaluation and selection methodology, three key issues must be addressed: 1) Is there an ES shell that can be used or should a specific ES be developed from tools?; 2) What ES shells are available?; and, 3) Which ES software packages should be seriously considered and evaluated in detail? Examples of commercially available mainframe and microcomputer-based ES shells are given in Table 1. For more examples, see Computerworld (1988). [Insert Table 1 About Here] Page 8 The purpose of developing a short list of shell products is to narrow the field of available ES software for consideration during Shell Evaluation. A short list of candidate ES software (two or three) eliminates any unnecessary effort or confusion which might result because too many alternative ES shells are evaluated. 2.1.1 Identify Candidate ES Software The project team must first identify available ES shells that operate within the enterprise's specific computer hardware and are compatible with its operating system. An enterprise may have dedicated hardware for running ES applications, such as a LISP machine. This would certainly be a technical constraint for ES shells under consideration. To initially identify appropriate ES shells for a particular application, there are several publishers (e.g., catalogs published from Datapro and ICP, as well as Expert Systems Strategies by Harmon and Associates) who provide profiles of ES software vendors and the products they offer. 2.1.2 Screening Criteria The list of screening criteria will contain relatively few items and should concentrate on functional requirements not commonly provided by ES shells and which are very specific to the organization evaluating ES software. Most commercial software packages share many standard functions and capabilities. At the same time, organizations have many identical requirements for software. However, software packages have unique capabilities and features which can readily distinguish one package from another. Similarly, organizations have unique requirements for any software package to satisfy. Page 9 [Insert Figure 3 About Here] The screening process matches the unique capabilities of commercial software packages with the unique requirements of the organization. Figure 3 illustrates the overlap between unique capabilities of software with the unique requirements of an organi- zation. The commercial software packages (e.g., ES shell) with the greatest overlap of unique capabilities with unique organizational requirements would be retained for more detailed scrutiny in the next stage of the evaluation and selection methodology (i.e., ES Shell Evaluation). At this point in the process, the concern is not how well the respective software packages meet the screening criteria, but only if the screening criteria (i.e. unique capabilities) are offered by the software. In the next stage of the methodology, all criteria (including both standard and unique capabilities) are evaluated as to how well they are performed by each respective software package on the short list. The screening criteria can be categorized into four major types: 1) technical requirements; 2) functional requirements; 3) documentation and training; and 4) vendor information. Technical Requirements - An organization's hardware and software strategy will likely dictate the screening criteria in the technical area. To be considered, a shell must fit the framework of the proposed system; it must be compatible with the hardware and software direction already identified. For example, the Social Security Administration required ES shell software to operate on their current hardware and operating system platform (Leary, 1989). The operating system is Page 10 clearly a strict technical requirement. Others could include programming languages, peripherals, memory needs, or data communication capabilities. Functional Requirements - The functional requirements of an ES shell can be classified according to the following system components: 1) User Interface; 2) Inference Engine; and 3) Knowledge Base. The functional requirements associated with each of these three components readily distinguish ES shell evaluation and selection from other software appraisal efforts (Harmon and King, 1985; Harmon, et al., 1989). The user interface component of an ES provides the dialog structure through which the user can access the ES. Normally, the user interacts with the ES via a consultative mode; and many ES interface components include an explanation module. The inference engine component allows hypotheses to be generated from information contained in the knowledge base; while the latter is comprised of facts and rules of thumb based upon experience. [Insert Table 2 About Here] Table 2 lists several examples of functional criteria to conduct the first-cut screening according to the ES components. (Forman and Nagy (1987) offer a more exhaustive list of ES shell criteria.) Possible screening criteria for the user interface component would be that the ES shell offers certain necessary features for the knowledge engineer (e.g., knowledge base editor and rule trace) as well as needed features for the user (e.g., prompted-menu display and explanations and justifications). The inference engine component could necessitate generating new facts by modus ponens and decision tress, while control Page 11 strategies should include forward chaining and procedural control. The knowledge base criteria might require representing facts by object- attribute-value (O-A-V) triplets and frames, handle uncertainty, and model relationships with variable rules. Documentation and Training - ES software packages normally include the documentation required to install and support the ES shell. It should be detailed, complete, and easy to understand. The availability of vendor-developed training sessions and materials may be very important, especially when the organization's personnel are inexperienced in implementing software. Vendor Information - A vendor's ability to support its package through training, consultation, installation, and maintenance assistance is an important consideration in evaluating ES software packages. The vendor should also be able to refer an evaluation team to a user who is willing to talk to them about the ES package and the accompanying support. Some organizations may wish to acquire a run-time version of an ES shell. This shell is modified to incorporate a specific knowledge base and to deactivate certain programming features. Not every vendor may offer such a product. The financial stability of a vendor can also be an important consideration. Financially successful vendors that have been in existence for more than a few years are more likely to adequately support their packages initially and in the future because such vendors attract and retain competent personnel. Page 12 But financial success alone does not ensure adequate and continued support. Vendor image, package reputation, the unit price, and the number of installations are also important considerations. Either the vendors themselves or the users to whom they directed the prospective buyer should be able to provide the needed information in these areas. 2.1.3 Pick Finalists The matching of the screening criteria against the list of ES software and their capabilities will cause the elimination of many (but hopefully not all) shells. The following are typical reasons to eliminate potential ES software candidates: 1) a vendor has only a few employees and has been in business less than a year; 2) operating systems software and hardware is not supported by a vendor; and, 3) system documentation is inadequate. By reducing the number of ES software packages under consider- ation from as many as twenty to two or three, a project team can more effectively devote its attention to the critical details that can make the difference between selecting an adequate ES shell and selecting a superior one. Moreover, by determining which ES software packages are available for the application, the screening process also determines whether an ES shell can be used or if a specific expert system should be constructed from ES tools (see Figure 2). Page 13 2.2 ES Shell Evaluation This second stage focuses on the two or three ES shells that were identified in the screening of ES software. The objective is to evaluate in detail the ES shell finalists and select the one software product that best meets the needs of the organization. The primary tasks of ES software evaluation are: 1) to further define the detailed evaluation criteria; 2) obtain shell product information; and, 3) evaluate the ES software finalists and pick one as the best alternative. 2.2.1 Expand Evaluation Criteria The screening criteria are expanded in more detail and fall into the same four categories: 1) technical requirements; 2) functional requirements; 3) documentation and training; and 4) vendor information. Although all categories are expanded during shell evaluation, the functional requirements receive the main attention and are related to the interface, inference and knowledge components of the ES shell. The purpose of this task is to develop a rather comprehensive functional view of the proposed system and to summarize the requirements that must be satisfied by the ES. As the project team defines the functional criteria, they should also document the levels of importance and need to the user. The following functional requirements for ES software, identified in Harmon and King (1985) are those for a hypothetical enterprise: 1) user interface features, including those for knowledge engineering and data extraction from external sources; 2) inference engine features, including generating new facts and control strategies; and 3) knowledge base features for handling facts, relationships, and Page 14 uncertainty. Table 3 exhibits an expansion of the above summary list of functional shell requirements. [Insert Table 3 About Here] 2.2.2 Obtain Package Information Once the system requirements have been established and the criteria reviewed, the capability of each ES shell to satisfy the requirements must be measured. Several techniques may be used to gather enough information to determine how well each package meets the requirements. In many cases, the project team can meet directly with the vendor sales and support personnel and discuss each requirement. But if requirements are so comprehensive and detailed that a more formal procedure should be followed, a request for proposal (RFP) can be submitted to vendors. It should be noted that preparing an RFP can be time-consuming and costly. In situations where the requirements are not so detailed and complex, it can be replaced by a less formal and more direct procedure such as a basic letter of request. 2.2.3 Evaluate ES Shells Once the vendors' responses to requirements have been received, the actual evaluation process can begin. The review is very detailed at this point, since the project team is looking for specific strengths and weaknesses of each package. Page 15 The project team is searching for deciding factors - not only what ES software packages have and how well they provide it, but also what they don't have. Detailed information is desired on the functions of the ES software and its related processing, including if and how functions that are not included in the ES shell could be implemented. Shovel and Lugasi (1987) compared three models for evaluating and selecting a computer system from four alternatives in a manufacturing case study. The three scoring models were: 1) the additive weight model; 2) the eigen-vector model; and, 3) the multi-attribute utility model. The authors reported that these models provided identical rankings of combinations of computer hardware, software availability and system support. Given a short list of software alternatives, any scoring model will likely yield identical rankings. In evaluating ES shells, Forman and Nagy (1987) applied the Analytic Hierarchy Process (Saaty, 1980) to select a shell from among just three alternative software packages. However, this study did not indicate how the short list was developed nor were any comparisons made with different scoring or ranking techniques. Generating the short list is more critical than the method of ranking the software packages. Criticism levied against weighting schemes for software selection decisions (Klein and Beck, 1987) can be minimized through the screening process and development of a short list. Naumann and Pelvis (1982) successfully applied weighting and scoring measures to select a systems development tool from a relatively short list (four) of candidate techniques. By weighting criteria for only two or three packages rather than for a dozen (in which case the aforementioned criticism is probably valid), the proposed evaluation Page 16 process allows for a very detailed and focused inspection of just the few best alternative ES software products. 2.2.4 Additional Selection Requirements The ranking scores are not necessarily the determining factor for selecting a particular ES shell. They should be used as a decision tool - a means for organizing and summarizing the significant quantity of information that the project team has collected. The highest score may not always indicate the best ES shell. The scores may not accurately reflect certain intangible factors such as the cosmetic appearance of reports and screens, how easy it will be to use the ES software, etc. Tailoring - The scores may not indicate how much time or the level of technical expertise needed to "tailor" the ES shell. Tailoring can be either costly if it is relatively extensive, or difficult if the internal structure of the software is complex. The importance of the technical processing architecture will depend on how much tailoring is anticipated. Furthermore, the architecture of the ES shell also determines how much modification is even possible. Documentation - A decision to use a particular ES shell should not be made on the basis of functional requirements alone. The ES software's documentation is a very important non-functional factor. Its accuracy and level of detail can affect the time it will take to evaluate and modify the package. Comparing documentation is sometimes very difficult at this stage of the evaluation process due to differences in format, style, etc. Still, it is important to review the vendors' documentation and to Page 17 reconfirm that the information collected on maintenance and support, for instance, is accurate and correct. The availability of and easy access to vendor-developed training sessions is very important. Training materials, such as user tutorials or video-based learning aids, may be especially critical when company personnel are inexperienced in developing ES applications. At this stage, the vendor should be able to refer the project team to current users of his software. The comments of these customers should prove invaluable. Site visits and demonstrations of the ES software in operation may be helpful. 2.3 Specific ES Design Assuming that ES software which is anticipated to provide satisfactory performance has been selected (see Figure 2), the project team is ready to confirm the selection by developing some specific ES prototypes based on the chosen shell. The primary reasons for this stage are to ensure that the ES package can be used effectively and to provide one last chance to reconsider the ES software decision. It is often difficult to determine the degree of user satisfaction until the design process has begun for specific applications utilizing the selected ES software. Therefore, this stage involves the design of demonstration prototypes of specific ES built from the ES shell. Such prototypes can provide significant benefits before finalizing the selection decision (Alavi, 1984; Meador and Mezger, 1984; and Janson, 1986). Page 18 2.3.1 Alter Functional Requirements Based on the capabilities of the selected shell as experienced in the prototyping exercise of specific ES, the definition of user requirements might be altered to include package features not previously considered, or to change or eliminate others. The modified requirements should be reviewed with the users. The effect of ES software deficiencies perhaps can be minimized by altering user procedures or postponing the implementation of some requirements until shell enhancements could be made. 2.3.2 ES Software Modifications and Supporting Programs Typically, the specific ES being developed requires certain functions and/or interfaces not provided by the software. If an ES shell does not meet all the functional requirements of a system, the following alternatives should be considered: 1) persuade the vendor to include additional features; 2) develop supplemental software; and, 3) modify the vendor's software. The chosen alternative will depend on the extent of the ES shell's deficiencies, the potential costs and benefits of altering the software, and the size and technical skills of the programming staff. Vendor-Supplied Enhancements - If possible, the vendor should be persuaded to do the modification for the purchaser. This is often the best alternative, since the vendor will usually update and maintain the software on a routine basis. Page 19 Supporting Programs - Developing software to supplement the vendor's ES package is often the most practical alternative. The vendor will normally continue to service the ES shell; but if this alternative is selected, the supplemental software should conform to the standards used by the vendor in developing the ES shell. Alter Code - Modifying an ES shell is usually not recommended. If the software is modified, the vendor may be reluctant or may even refuse to service the package. Updates to the software may not be compatible with the modifications effected. In some cases, this may not even be an option, since the purchaser of the ES shell does not have (or cannot get at any price) a copy of the source code. In this instance, all that the purchaser can do is to build a front-end or back-end to the software package. 2.3.3 Finalize ES Shell Selection It is not unusual for an organization to complete the last stage of the evaluation and selection process for ES software, only to realize that the selected shell is not the best choice. Perhaps too many compromises have been made and users are no longer satisfied. Possibly, the tailoring effort has become so extensive that a custom ES (i.e., specific ES application built from tools) would be a better choice (see Figure 2). Therefore, a final commitment to using a particular shell should be avoided until the design of specific ES using the potential software package has progressed to the point where user satisfaction is ensured. Page 20 3.0 SUMMARY AND CONCLUSION Using ES shells for the development of specific ES will reduce personnel requirements and development costs. Conducting the evaluation of ES software increases the effort necessary for developing specific ES, but this undertaking is offset by the aforementioned advantages of using a shell package. Despite the promises offered by ES software, the performance of some ES shells is much less than expected. Weak or non-existent selection procedures may explain much of this poor implementation record. The methodology proposed in this paper will hopefully reduce the risks associated with ES software and facilitate success in developing specific ES from shells (Alavi, 1984; Janson, 1986). The most critical phases of the methodology are determining the criteria for an ES shell as incorporated in the first phase (the development of a short list) and the confirmation element of the third phase (design of specific ES with the selected shell). Initially, the screening process determines whether a shell is feasible and reduces the number of ES software packages to be evaluated in detail. Finally, the development of specific ES with the selected shell ensures that the ES software can be used effectively and provides a last chance to consider building specific ES from tools. Page 21 REFERENCES Alavi, M., "An Assessment of the Prototyping Approach to Information Systems Development," Communications gf the gO, 1984, 27 (6), 556-563. Breslin, J. Selecting and Installing Software packages. Westport, CN: Quorom Books, 1986. Domputerworld, "Expert System Shells," October 17, 1988, 90-96. Curry, J. W., and Bonner, D. M. Mow tg Find and Bu y good Software: A Guide for Business Anal Professional people. Englewood Cliffs, NJ: Prentice-Hall, Inc., 1983. Forman, E. H., and Nagy, T. J. "A Multicriteria Model to Select an Expert System Generator." In proceedings g the Expert Systems in Government Conference, Washington, DC: IEEE/MITRE, 1985. Forman, E. H., and Nagy, T. J. "EXSYS vs. TOPSI/OPS5 vs. MICRO-PS: A Multicriteria Model to Select an Expert System Generator." Telematics And Informatics, 1987, 4 (1), 37-54. Gray, C. D. The bight Choice: A Complete Guide IQ Evaluating, Selecting, and Installing MEEII Software. Essex Junction, VT: Oliver Wight Limited Publications, Inc., 1987. Harmon, P., and King, D. Expert Systems: Artificial Intelligence in business. New York, NY: John Wiley & Sons, Inc., 1985. Harmon, P., Maus, R., and Morrissey, W. Expert Systems: Tools & Applications. New York, NY: John Wiley & Sons, Inc., 1985. Hayes-Roth, F., Waterman, D. A., and Lenat, D. B. Building Expert Systems. Reading, MA: Addison-Wesley, 1983. Henderson, J. C., "Finding Synergy Between Decision Support Systems and Expert Systems Research," Decision Sciences, 1987, 18 (3), 333-349. Janson, M., "Applying a Pilot System and Prototyping Approach to Systems Development and Implementation," Information And Management, 1986, 10 (4), 209-216. Klein, G., and Beck, P. O., "A Decision Aid for Selecting among Information System Alternatives," MIS Quarterl y , 1987, 11 (2) 177-185. Leary, E., "Collecting Shells for Rapid Prototyping," Computerworla, November 23, 1987, p. S2. Leary, E., "Expert Systems at the Social Security Administration," Journal QL policy Analysis and Management, 1989, 8 (2), 200-203. Page 22 Liebowitz, J., "Determining Functional Requirements for NASA Goddard's Command Management System Software Design Using Expert Systems," D. Sc. dissertation, George Washington University, Washington, DC, 1985. Liebowitz, J. introduction t2 Expert Systems. Watsonville, CA: Mitchell Publishing, 1988. Lynch, R. K., "Implementing Packaged Application Software: Hidden Costs and New Challenges," Systems, Objectives, Solutions, 1984, 4 (4), 227-234. Lynch, R. K., "Nine Pitfalls in Implementing Packaged Applications Software," Journal gf Information Systems Management, 1985, 2 (2), 88-92. Martin, J., and McClure, C., "Buying Software off the Rack," Harvard Business Review, 1983, 61 (6), 32-47. Meador, G. L., and Mezger, R. A., "Selecting An End User Programming Language For DSS Development," ma Quarterly, 1984, 8 (4), 267-281. Naumann, J. D., and Jenkins, A. M., "Prototyping: The New Paradigm For Systems Development," MIS Ouarterly, 1982, 6 (3), 29-44. Naumann, J. D., and Palvia, S., "A Selection Model for Systems Development Tools," MI$ Ouarterly, 1982, 6 (1), 39-48. Newquist, H., "AI Adapts to Use of the Vernacular," Computerwor14, October 17, 1988, pp. 82-83. Saaty, T. L., The Analytic Hierarchy Process, York, NY; McGraw-Hill, 1980. Shoval, P. and Y. Lugasi, "Models for Computer System Evaluation and Selection," Information k Management, 1987, 12 (3), 117-129. Sprague, R. H., Jr., "A Framework for the Development of Decision Support Systems," MIS Quarterly, 1980, 4 (4), 1-26. Turban, E. Decision Support And Expert Systems (second Edition). New York, NY: Macmillan Company, 1990. Vedder, R. G., "PC-Based Expert System Shells: Some Desirable and Less Desirable Characteristics," Expert Systems, 1989, 6 (1), 28-42. Waterman, D. A. A Guide tga £pert Systems. Reading, MA: Addison-Wesley, 1986. Waterman, D. A., and Hayes-Roth, F. An investigation gf Tools for Building Expert Systems. Santa Monica, CA: Rand Corporation, 1982. FIGURE 1. LEVELS OF ES TECHNOLOGY Sisecific ES Applicanons MoralIv• Deveicio►ent IS Soo* No guild Swine ES horn Tools Evaluate U Shells Yes Confirm U Shell Selection FIGURE 2. A MULTIPLE CRITERIA DECISION METHODOLOGY FOR ES SHELL SELECTION Develop Mon list IS Software Purchase ES Software Standard R•quinments Standard N CapablIttios I U U' N U/ / FIGURE 3. MATCHING OF SOFTWARE SCREENING CRITERIA Software Organization * TABLE 1. REPRESENTATIVE ES SHELL PRODUCTS Large Machine Shells** Product KEE ART ESE EMYCIN S.1 Level 5 KES Knowledge Craft IBM Knowledgetool Rulemaster 2 KBMS ADS Microcomputer Shells Product 1st-Class EXSYS M.1 KES-PC GURU Personal Consultant VP-Expert Knowledgepro Nexpert Vendor Intellicorp Inference Corp. IBM Stanford University Teknowledge, Inc. Information Builders Software A & E Carnegie Group, Inc. IBM Radian Corp. AI Corporation AION Corporation Vendor lst-Class Expert Systems EXSYS, Inc. Teknowledge, Inc. Software A & E Micro Database Systems Texas Instruments Paperback Software Knowledge Garden, Inc. Neuron Data * Numerous shells are available in both large machine and microcomputer versions. ** Includes minicomputers, mainframes and specialized LISP machines. TABLE 2. POTENTIAL SCREENING CRITERIA OF ES COMPONENTS User Inference Knowledge Interface Engine Base Knowledge Base Modus Ponens Editor Decision Tree Prompted-Menu Algorithm Display Forward Explanation Chaining Justification Backward Chaining Graphics Mixed Chaining Procedural Control Database/ Spreadsheet Links Hypertext O-A-V Triplets Frames Variable Rules Certainty Factors TABLE 3. DETAILED FUNCTIONAL CRITERIA FOR ES SOFTWARE Interface Component Knowledge Engineering Facilities Knowledge Base Editor Traces and Probes Graphic Display (Windows) End-User Elements Explanations and Justifications Line-Oriented Display Prompted-Menu Display Sources of Data Instruments Data Bases Other Languages and Procedures Inference Component Generating New Facts Modus Ponens Resolution Decision Tree Algorithm Control Strategies Backward Chaining Forward Chaining Depth-First Search Breadth-First Search Procedural Control Knowledge Component Representing Facts A-V Pairs O-A-V Triplets Frames Relationships If-Then Rules Variable Rules Examples Uncertainty Inheritance Certainty Factors Probabilities 1986 86/01 Arnoud DE MEYER 86/02 Philippe A. NAERT Marcel VEVERBERGN and Guido VERSVIJVEL 86/03 Michael BRIMN 86/04 Spyros MAKRIDAKIS and Michele HIBON 86/05 Charles A. VYPLOSZ 86/06 Francesco GIAVAllI, Jeff R. SHEEN and Charles A. VYPLOSZ 86/07 Douglas L. MacLACHLAN and Spyros MAKRIDAKIS 86/08 Jose de la TORRE and David H. NECKAR 86/09 Philippe C. HASPESLACH 86/10 R. MOENART, Arnoud DE MEYER, J. BARBE and D. DESCHOOLMEESTER. 86/11 Philippe A. NAERT and Alain BULTEZ 86/12 Roger BETANCOURT and David GAUTSCHI 86/13 S.P. ANDERSON and Damien J. NEVEN "The R i D/Production interface". "Subjective estimation in integrating communication budget and allocation decisions: a case study", January 1986. "Sponsorship and the diffusion of organizational innovation: a preliminary view". "Confidence intervals: an empirical investigation for the series in the M- Competition" . "A note on the reduction of the workweek", July 1985. "The real exchange rate and the fiscal aspects of a natural resource discovery", Revised version: February 1986. "Judgmental biases in sales forecasting", February 1986. "Forecasting political risks for international operations", Second Draft: March 3, 1986. "Conceptualizing the strategic process in diversified firms: the role and nature of the corporate influence process", February 1986. "Analysing the issues concerning technological de-maturity". "From "Lydiametry" to "Pinkhamization": misspecifying advertising dynamics rarely affects profitability". "The economics of retail firms", Revised April 1986. "Spatial competition a la Cournot". 86/16 B. Espen ECKBO and Hervig M. LANGOHR 86/17 David B. JEMISON 86/18 James TEBOUL and V. MALLERET 86/19 Rob R. VEITZ 86/20 Albert CORHAY, Gabriel HAVAVINI and Pierre A. MICHEL 86/21 Albert CORHAY, Gabriel A. HAVAVINI and Pierre A. MICHEL 86/22 Albert CORNAY, Gabriel A. HAVAVINI and Pierre A. MICHEL 86/23 Arnoud DE MEYER 86/24 David GAUTSCNI and Vithala R. RAO 86/25 H. Peter GRAY and Ingo WALTER 86/26 Barry EICHENGREEN and Charles VYPLOSZ 86/27 Karel COOL and Ingemar DIERICKX 86/28 Manfred KEYS DE VRIES and Danny MILLER 86/29 Manfred KETS DE VRIES 86/30 Manfred KETS DE VRIES 86/31 Arnoud DE MEYER "Les primes des offres publiques, la note d'information et le marche des transferts de contr8le des societie. "Strategic capability transfer in acquisition integration", May 1986. "Towards an operational definition of services", 1986. "Nostradamus: a knowledge-based forecasting advisor". "The pricing of equity on the London stock exchange: seasonality and size premium", June 1986. "Risk-premia seasonality in U.S. and European equity markets", February 1986. "Seasonality in the risk-return relationships some international evidence", July 1986. "An exploratory study on the integration of information systems in manufacturing", July 1986. "A methodology for specification and aggregation in product concept testing", July 1986. "Protection", August 1986. "The economic consequences of the Franc Poincare", September 1986. "Negative risk-return relationships in business strategy: paradox or truism?", October 1986. "Interpreting organizational texts. "Why follow the leader?". "The succession game: the real story. "Flexibility: the next competitive battle", October 1986. INSEAD WORKING PAPERS SERIES "Comparaison Internationale des marges brutes du commerce", June 1985. "Nov the managerial attitudes of firms vith FMS differ from other manufacturing firms: survey results", June 1986. 86/31 Arnoud DE MEYER, Jinichiro NAKANE, Jeffrey G. MILLER and Kasra FERDOVS 86/32 Karel COOL and Dan SCHENDEL "Flexibility: the next competitive battle", Revised Version: March 1987 Performance differences among strategic group members", October 1986. 86/14 Charles WALDMAN 86/15 Mihkel TOMBAK and Arhoud DE MEYER 87/06 Arun K. JAIN, "Customer loyalty as a construct in the Christian PINSON and marketing of banking services", July 1986. Naresh K. MALHOTRA 86/33 Ernst BALTENSPERGER and Jean DERMINE 86/34 Philippe HASPESLAGH and David JEMISON 86/35 Jean DERMINE 86/36 Albert CORHAY and Gabriel HAVAVINI 86/37 David GAUTSCHI and Roger BETANCOURT 86/38 Gabriel HAVAVINI 86/39 Gabriel HAVAVINI Pierre MICHEL and Albert CORHAY 86/40 Charles VYPLOSZ 86/41 Kasra FERDOWS and Wickham SKINNER 86/42 Kasra FERDOWS and Per LINDBERG 86/43 Damien NEVEN 86/44 Ingemar DIERICKX Carmen MATUTES and Damien NEVEN 1987 87/01 Manfred KETS DE VRIES 87/02 Claude VIALLET 87/03 David GAUTSCHI and Vithala RAO 87/04 Sumantra CHOSHAL and Christopher BARTLETT 87/05 Arnoud DE MEYER and Kasra FERDOWS "The role of public policy in insuring financial stability: a cross-country, comparative perspective", August 1986, Revised November 1986. "Acquisitions: myths and reality", July 1986. "Measuring the market value of a bank, a primer", November 1986. "Seasonality in the risk-return relationship: some international evidence", July 1986. "The evolution of retailing: a suggested economic interpretation". "Financial innovation and recent developments in the French capital markets", Updated: September 1986. "The pricing of common stocks on the Brussels stock exchange: a re-examination of the evidence", November 1986. "Capital flovs liberalization and the EMS, a French perspective", December 1986. "Manufacturing in a new perspective", July 1986. "FMS as indicator of manufacturing strategy", December 1986. "On the existence of equilibrium in hotelling's model", November 1986. "Value added tax and competition", December 1986. "Prisoners of leadership". "An empirical investigation of international asset pricing", November 1986. "A methodology for specification and aggregation in product concept testing", Revised Version: January 1987. "Organizing for innovations: case of the multinational corporation", February 1987. "Managerial focal points in manufacturing strategy", February 1987. "Equity pricing and stock market anomalies", February 1987. "Leaders who can't manage", February 1987. "Entrepreneurial activities of European MBAs", March 1987. "A cultural view of organizational change", March 1987 "Forecasting and loss functions", March 1987. "The Janus Bead: learning frost the superior and subordinate faces of the manager's job", April 1987. "Multinational corporations as differentiated networks", April 1987. "Product Standards and Competitive Strategy: An Analysis of the Principles", May 1987. "KETAFORECASTING: Vmys of improving Forecasting. Accuracy and Usefulness", May 1987. "Takeover attempts: what does the language tell us?, June 1987. "Managers' cognitive maps for upward and dovnvard relationships", June 1987. "Patents and the European biotechnology lag: a study of large European pharmaceutical firms", June 1987. "Why the EMS? Dynamic games and the equilibrium policy regime, May 1987. "A new approach to statistical forecasting", June 1987. "Strategy formulation: the impact of national culture", Revised: July 1987. "Conflicting ideologies: structural and motivational consequences", August 1987. "The demand for retail products and the household production model: nev views on complementarity and substitutability". 87/07 Rolf BANZ and Gabriel HAVAVINI 87/08 Manfred KETS DE VRIES 87/09 Lister VICKERY, Mark PILKINGTON and Paul READ 87/10 Andre LAURENT 87/11 Robert FILDES and Spyros MAKRIDAKIS 87/12 Fernando BARTOLOME and Andre LAURENT 87/13 Sumantra GHOSHAL and Nitin NOHRIA 87/14 Landis GABEL 87/15 Spyros MAKRIDAKIS 87/16 Susan SCHNEIDER and Roger DUNBAR 87/17 Andre LAURENT and Fernando BARTOLOME 87/18 Reinhard ANGELMAR and Christoph LIEBSCHER 87/19 David BEGG and Charles VYPLOSZ 87/20 Spyros MAKRIDAKIS 87/21 Susan SCHNEIDER 87/22 Susan SCHNEIDER 87/23 Roger BETANCOURT David GAUTSCHI 87/29 Susan SCHNEIDER and Paul SHRIVASTAVA "The internal and external careers: a theoretical and cross-cultural perspective", Spring 1987. "The robustness of MDS configurations in the face of incomplete data", March 1987, Revised: July 1987. "Demand complementarities, household production and retail assortments", July 1987. "Is there a capital shortage in Europe?", August 1987. "Controlling the interest-rate risk of bonds: an introduction to duration analysis and immunization strategies", September 1987. "Interpreting strategic behavior: basic assumptions themes in organizations", September 1987 87/41 Cavriel HAVAVINI and Claude VIALLET 87/42 Damien NEVEN and Jacques-F. THISSE 87/43 Jean CABSZEVICZ and Jacques-F. THISSE 87/44 Jonathan HAMILTON, Jacques-F. THISSE and Anita WESKAMP 87/45 Karel COOL, David JEMISON and Ingemar DIERICKX 87/46 Ingemar DIERICKX and Karel COOL "Seasonality, size premium and the relationship between the risk and the return of French common stocks", November 1987 'Combining horizontal and vertical differentiation: the principle of max-min differentiation", December 1987 "Location", December 1987 "Spatial discrimination: Bertrand vs. Cournot in a model of location choice", December 1987 "Business strategy, market structure and risk- return relationships: a causal interpretation", December 1987. "Asset stock accumulation and sustainability of competitive advantage", December 1987. 87/24 C.B. DERR and Andre LAURENT 87/25 A. K. JAIN, N. K. MALHOTRA and Christian PINSON 87/26 Roger BETANCOURT and David GAUTSCHI 87/27 Michael BURDA 87/28 Gabriel HAVAVINI 87/30 Jonathan HAMILTON "Spatial competition and the Core", August W. Bentley MACLEOD 1987. 1988 and J. F. THISSE 87/31 Martine OUINZII and J. F. THISSE 87/32 Arnoud DE MEYER 87/33 Yves DOZ and Amy SHUEN 87/34 Kasra FERDOWS and Arnoud DE MEYER 87/35 P. J. LEDERER and J. F. THISSE 87/36 Manfred KETS DE VRIES 87/37 Landis LABEL 87/38 Susan SCHNEIDER 87/39 Manfred KETS DE VRIES 1987 87/40 Carmen MATUTES and Pierre REGIBEAU "On the optimality of central places", September 1987. "German, French and British manufacturing strategies less different than one thinks", September 1987. "A process framework for analyzing cooperation between firms", September 1987. "European manufacturers: the dangers of complacency. Insights from the 1987 European manufacturing futures survey, October 1987. "Competitive location on networks under discriminatory pricing", September 1987. "Prisoners of leadership", Revised version October 1987. "Privatization: its motives and likely consequences", October 1987. "Strategy formulation: the impact of national culture", October 1987. 'The dark side of CEO succession", November "Product compatibility and the scope of entry", November 1987 88/01 Michael LAWRENCE and Spyros MAKRIDAKIS 88/02 Spyros MAKRIDAKIS 88/03 James TEBOUL 88/04 Susan SCHNEIDER 88/05 Charles WYPLOSZ 88/06 Reinhard ANGELMAR 88/07 Ingemar DIERICKX and Karel COOL 88/08 Reinhard ANGELMAR and Susan SCHNEIDER 88/09 Bernard SINCLAIR- DESGAGNe 88/10 Bernard SINCLAIR- DESGAGN4 88/11 Bernard SINCLAIR- DESGAGNe "Factors affecting judgemental forecasts and confidence intervals", January 1988. "Predicting recessions and other turning points", January 1988. "De-industrialize service for quality", January 1988. "National vs. corporate culture: implications for human resource management", January 1988. "The svinging dollar: is Europe out of step?", January 1988. "Les conflits dans les canaux de distribution", January 1988. "Competitive advantage: a resource based perspective", January 1988. "Issues in the study of organizational cognition", February 1988. "Price formation and product design through bidding", February 1988. "The robustness of some standard auction game forms", February 1988. "When stationary strategies are equilibrium bidding strategy: The single-crossing property", February 1988. 88/12 Spyros MAKRIDAKIS "Business firms and managers in the 21st century", February 1988 88/13 Manfred KETS DE VRIES "Alezithymia in organizational life: the organization man revisited", February 1988. 88/14 Alain NOEL "The interpretation of strategies: a study of the impact of CEOs on the corporation", March 1988. 88/15 Anil DEOLALIKAR and Lars-Hendrik ROLLER 88/16 Gabriel HAWAWINI 88/17 Michael BURDA 88/18 Michael BURDA 88/19 M.J. LAWRENCE and Spyros MAKRIDAKIS 88/20 Jean DERMINE, Damien NEVEN and J.F. THISSE 88/21 James TEBOUL 88/22 Lars-Hendrik ROLLER 88/23 Sjur Didrik FLAM and Georges ZACCOUR 88/24 B. Espen ECKBO and Hervig LANGOHR 88/25 Everette S. GARDNER and Spyros MAKRIDAKIS "The production of and returns from industrial innovation: an econometric analysis for a developing country", December 1987. "Market efficiency and equity pricing: international evidence and implications for global investing", March 1988. "Monopolistic competition, costs of adjustment and the behavior of European employment", September 1987. "Reflections on "Wait Unemployment" in Europe", November 1987, revised February 1988. "Individual bias in judgements of confidence", March 1988. "Portfolio selection by mutual funds, an equilibrium model", March 1988. "De-industrialize service for quality", March 1988 (88/03 Revised). "Proper Quadratic Functions with an Application to AT&T", May 1987 (Revised March 1988). "Equilibres de Nash-Cournot dans le marche europeen du gaz: un cas oil les solutions en boucle ouverte et en feedback coincident", Mars 1988 "Information disclosure, means of payment, and takeover premia. Public and Private tender offers in France", July 1985, Sixth revision, April 1988. "The future of forecasting", April 1988. 88/26 Sjur Didrik FLAM and Georges ZACCOUR 88/27 Murugappa KRISHNAN Lars-Hendrik ROLLER 88/28 Sumantra GHOSHAL and C. A. BARTLETT "Semi-competitive Cournot equilibrium in multistage oligopolies", April 1988. "Entry game with resalable capacity", April 1988. "The multinational corporation as a network: perspectives from interorganizational theory" May 1988. 88/29 Naresh K. MALHOTRA, Christian PINSON and Arun K. JAIN 88/30 Catherine C. ECKEL and Theo VERMAELEN 88/31 Sumantra CHOSHAL and Christopher BARTLETT 88/32 Kasra FERDOVS and David SACKRIDER 88/33 Mihkel M. TOMBAK 88/34 Mihkel M. TOMBAK 88/35 Mihkel M. TOMBAK 88/36 Vikas TIBREVALA and Bruce BUCHANAN 88/37 Murugappa KRISHNAN Lars-Hendrik ROLLER 88/38 Manfred KETS DE VRIES 88/39 Manfred KETS DE VRIES 88/40 Josef LAKONISHOK and Theo VERMAELEN 88/41 Charles WYPLOSZ 88/42 Paul EVANS 88/43 B. SINCLAIR-DESGAGNE 88/44 Essam MAHMOUD and Spyros MAKRIDAKIS 88/45 Robert KORAJCZYK and Claude VIALLET 88/46 Yves DOZ and Amy SHUEN "Consumer cognitive complexity and the dimensionality of multidimensional scaling configurations", May 1988. "The financial fallout from Chernobyl: risk perceptions and regulatory response", May 1988. "Creation, adoption, and diffusion of innovations by subsidiaries of multinational corporations", June 1988. "International manufacturing: positioning plants for success", June'1988. "The importance of flexibility in manufacturing", June 1988. "Flexibility: an important dimension in manufacturing", June 1988. "A strategic analysis of investment in flexible manufacturing systems", July 1988. "A Predictive Test of the NBD Model that Controls for Non-stationarity", June 1988. "Regulating Price-Liability Competition To Improve Welfare", July 1988. "The Motivating Netale of Envy : A Forgotten Factor in Management, April 88. "The Leader as Mirror : Clinical Reflections", July 1988. "Anomalous price behavior around repurchase tender offers", August 1988. "Assymetry in the EMS: intentional or systemic?", August 1988. "Organizational development in the transnational enterprise", June 1988. "Croup decision support systems implement Bayesian rationality", September 1988. "The state of the art and future directions in combining forecasts", September 1988. "An empirical investigation of international asset pricing", November 1986, revised August 1988. "From intent to outcome: a process framework for partnerships", August 1988. 88/47 Alain BULTEZ, Els GIJSBRECHTS, Philippe NAERT and Piet VANDEN ABEELE 88/48 Michael BURDA 88/49 Nathalie DIERKENS 88/50 Rob WEITZ and Arnoud DE MEYER 88/51 Rob VEITZ 88/52 Susan SCHNEIDER and Reinhard ANCELMAR 88/53 Manfred KETS DE VRIES 88/54 Lars-Hendrik ROLLER and Mihkel M. TOMBAK 88/55 Peter BOSSAERTS and Pierre HILLION 88/56 Pierre HILLION 88/57 Wilfricd VANHONACKER and Lydia PRICE 88/58 B. SINCLAIR-DESGAGNE and Mihkel M. TOMBAK "Asymmetric cannibalism between substitute items listed by retailers", September 1988. "Reflections on 'Wait unemployment' in Europe, II", April 1988 revised September 1988. "Information asymmetry and equity issues", September 1988. "Managing expert systems: from inception through updating", October 1987. "Technology, work, and the organization: the impact of expert systems", July 1988. "Cognition and organizational analysis: vho's minding the store?", September 1988. "Whatever happened to the philosopher king: the leader's addiction to power, September 1988. "Strategic choice of flexible production technologies and welfare implications", October 1988 "Method of moments tests of contingent claims asset pricing models", October 1988. "Size-sorted portfolios and the violation of the random walk hypothesis: Additional empirical evidence and implication for tests of asset pricing models", June 1988. "Data transferability: estimating the response effect of future events based on historical analogy", October 1988. "Assessing economic inequality", November 1988. 88/63 Fernando NASCIHENTO and Vilfried R. VANHONACKER 88/64 Kasra FERDOWS 88/65 Arnoud DE MEYER and Kasra FERDOVS 88/66 Nathalie DIERKENS 88/67 Paul S. ADLER and Kasra FERDOWS 1989 89/01 Joyce K. BYRER and Tavfik JELASSI 89/02 Louis A. LE BLANC and Tavfik JELASSI 89/03 Beth H. JONES and Tavfik JELASSI 89/04 Kasra FERDOWS and Arnoud DE MEYER 89/05 Martin KILDUFF and Reinhard ANCELHAR 89/06 Mihkel M. TOMBAK and B. SINCLAIR-DESGACNE "Strategic pricing of differentiated consumer durables in a dynamic duopoly: a numerical analysis", October 1988. "Charting strategic roles for international factories", December 1988. "Quality up, technology down", October 1988. "A discussion of exact measures of information assymetry: the example of Myers and Majluf model or the importance of the asset structure of the firm", December 1988. "The chief technology officer", December 1988. "The impact of language theories on DSS dialog", January 1989. "DSS software selection: a multiple criteria decision methodology", January 1989. "Negotiation support: the effects of computer intervention and conflict level on bargaining outcome", January 1989. "Lasting improvement in manufacturing performance: In search of a new theory", January 1989. "Shared history or shared culture? The effects of time, culture, and performance on institutionalization in simulated organizations", January 1989. "Coordinating manufacturing and business strategies: I", February 1989. 88/59 Martin KILDUFF "The interpersonal structure of decision 89/07 Damien J. NEVEN "Structural adjustment in European retail making: a social comparison approach to banking. Some view from industrial organizational choice", November 1988. organisation", January 1989. 88/60 Michael BURDA 88/61 Lars-Hendrik ROLLER 88/62 Cynthia VAN HULLE, Theo VERMAELEN and Paul DE WOUTERS "Is mismatch really the problem? Some estimates of the Chelvood Cate I/ model vith US data", September 1988. "Modelling cost structure: the Bell System revisited", November 1988. "Regulation, taxes and the market for corporate control in Belgium", September 1988. 89/08 Arnoud DE MEYER and Hellmut SCHUTTE 89/09 Damien NEVEN, Carmen MATUTES and Marcel CORSTJENS 89/10 Nathalie DIERKENS, Bruno GERARD and Pierre MILLION "Trends in the development of technology and their effects on the production structure in the European Community", January 1989. "Brand proliferation and entry deterrence", February 1989. "A market based approach to the valuation of the assets in place and the growth opportunities of the firm", December 1988. 89/11 Manfred KETS DE VRIES and Alain NOEL 89/12 Vilfried VANHONACKER 89/13 Manfred KETS DE VRIES 89/14 Reinhard ANGELMAR 89/15 Reinhard ANGELMAR 89/16 Vilfried VANHONACKER, Donald LEHMANN and Fareena SULTAN 89/17 Gilles AMADO, Claude FAUCHEUX and Andre LAURENT 89/18 Srinivasan BALAK- RISHNAN and Mitchell KOZA 89/19 Vilfried VANHONACKER, Donald LEHMANN and Fareena SULTAN 89/20 Vilfried VANHONACKER and Russell VINER 89/21 Arnoud de MEYER and Kasra FERDOWS 89/22 Manfred KETS DE VRIES and Sydney PERZOW 89/23 Robert KORAJCZYK and Claude VIALLET 89/24 Martin KILDUFF and Mitchel ABOLAFIA 89/25 Roger BETANCOURT and David GAUTSCHI 89/26 Charles BEAN, Edmond MALINVAUD, Peter BERNHOLZ, Francesco GIAVAllI and Charles VYPLOSZ "Understanding the leader-strategy interface: application of the strategic relationship interview method", February 1989. "Estimating dynamic response models when the data are subject to different temporal aggregation", January 1989. "The impostor syndrome: a disquieting phenomenon in organizational life", February 1989. "Product innovation: a tool for competitive advantage", March 1989. "Evaluating a firm's product innovation performance", March 1989. "Combining related and sparse data in linear regression models", February 1989. "Changement organisationnel et rialitis culturelles: contrastes franco-americains", March 1989. "Information asymmetry, market failure and joint-ventures: theory and evidence", March 1989 "Combining related and sparse data in linear regression models", Revised March 1989 "A rational random behavior model of choice", Revised March 1989 "Influence of manufacturing improvement programmes on performance", April 1989 "What is the role of character in psychoanalysis? April 1989 "Equity risk premia and the pricing of foreign exchange risk" April 1989 "The social destruction of reality: Organisational conflict as social drama" April 1989 "Two essential characteristics of retail markets and their economic consequences" March 1989 "Macroeconomic policies for 1992: the transition and after", April 1989 89/27 David KRACKHARDT and Martin KILDUFF 89/28 Martin KILDUFF 89/29 Robert GOGEL and Jean-Claude LARRECHE 89/30 Lars-Hendrik ROLLER and Mihkel M. TOMBAK 89/31 Michael C. BURDA and Stefan GERLACH 89/32 Peter HAUG and Tavfik JELASSI 89/33 Bernard SINCLAIR- DESGAGNE 89/34 Sumantra CHOSHAL and Nittin NOHRIA 89/35 Jean DERMINE and Pierre BILLION 89/36 Martin KILDUFF 89/37 Manfred KETS DE VRIES 89/38 Manfrd KETS DE VRIES 89/39 Robert KORAJCZYK and Claude VIALLET 89/40 Balaji CHAKRAVARTHY 89/41 B. SINCLAIR-DESGAGNE and Nathalie DIERKENS 89/42 Robert ANSON and Tavfik JELASSI 89/43 Michael BURDA 89/44 Balaji CHAKRAVARTHY and Peter LORANGF. 89/45 Rob WEITZ and Arnoud DE MEYER "Friendship patterns and cultural attributions: the control of organizational diversity", April 1989 "The interpersonal structure of decision making: a social comparison approach to organizational choice", Revised April 1989 "The battlefield for 1992: product strength and geographic coverage", May 1989 "Competition and Investment in Flexible Technologies", May 1989 "Intertemporal prices and the US trade balance in durable goods", July 1989 "Application and evaluation of a multi-criteria decision support system for the dynamic selection of U.S. manufacturing locations", May 1989 "Design flexibility in monopsonistic industries", May 1989 "Requisite variety versus shared values: managing corporate-division relationships in the it-Form organisation", May 1989 "Deposit rate ceiliess and the market value of banks: The case of France 1971-1981", May 1989 "A dispositional approach to social networks: the case of organizational choice", May 1989 "The organisational fool: balancing a leader's hubris", May 1989 "The CEO blues", June 1989 "An. empirical investigation of international asset pricing", (Revised June 1989) "Management systems for innovation and productivity", June 1989 "The strategic supply of precisions", June 1989 "A development fraaevork for computer supported conflict resolution", July 1989 "A note on firing costs and severance benefits in equilibrium unemployment", June 1989 "Strategic adaptation in multi-business firms", June 1989 "Managing expert systems: a framevork and case study", June 1989 89/46 Marcel CORSTJENS, Carmen MATUTES and Damien NEVEN 89/47 Manfred KETS DE VRIES and Christine MEAD "Entry Encouragement", July 1989 "The global dimension in leadership and organization: issues and controversies", April 1989 89/64 Enver YUCESAN and "Complexity of simulation models: A graph (TM) Lee SCHRUBEN theoretic approach", November 1989 89/65 Soumitra DUTTA and "MARS: A mergers and acquisitions reasoning (TM, Piero BONISSONE system", November 1989 AC, PIN) 89/48 Damien NEVEN and Lars-Hendrik ROLLER 89/49 Jean DERMINE 89/50 Jean DERMINE 89/51 Spyros MAKRIDAKIS 89/52 Arnoud DE MEYER 89/53 Spyros MAKRIDAKIS 89/54 S. BALAKRISHNAN and Mitchell KOZA 89/55 H. SCHUTTE 89/56 Vilfried VANHONACKER and Lydia PRICE 89/57 Taekvon KIM, Lars-Hendrik ROLLER and Mihkel TOMBAK 89/58 Lars-Hendrik ROLLER (EP,TM) and Mihkel TOMBAK 89/59 Manfred KETS DE VRIES, (08) Daphna ZEVADI, Alain NOEL and Mihkel TOMBAK 89/60 Enver YUCESAN and (TM) Lee SCHRUBEN 89/61 Susan SCHNEIDER and (All) Arnoud DE MEYER "European integration and trade flows", August 1989 "Home country control and mutual recognition", July 1989 "The specialization of financial institutions, the EEC model", August 1989 "Sliding simulation: a new approach to time series forecasting", July 1989 "Shortening development cycle times: a manufacturer's perspective", August 1989 "Why combining works?", July 1989 "Organisation costs and a theory of joint ventures", September 1989 "Buro-Japanese cooperation in information technology", September 1989 "On the practical usefulness of meta-analysis results", September 1989 "Market growth and the diffusion of multiproduct technologies", September 1989 "Strategic aspects of flexible production technologies", October 1989 "Locus of control and entrepreneurship: a three-country comparative study", October 1989 "Simulation graphs for design and analysis of discrete event simulation models", October 1989 "Interpreting and responding to strategic issues: The impact of national culture", October 1989 "On the regulation of procurement bids", November 1989 "Market microstructure effects of government intervention in the foreign exchange market", December 1989 89/66 B. SINCLAIR-DESGAGNE (TM,RP) 89/67 Peter BOSSAERTS and (PIN) Pierre HILLION 89/62 Arnoud DE MEYER "Technology strategy and international B 6 D (TM) operations", October 1989 89/63 Enver YUCESAN and "Equivalence of simulations: A graph theoretic (TM) Lee SCHRUBEN approach", November 1989 90/16 Richard LEVICH and FIN Ingo WALTER 90/17 Nathalie DIERKENS PIN 90/18 Wilfried VANHONACKER MKT 90/19 Beth JONES and TM Tavfik JELASSI 90/20 Tavfik JELASSI, TM Gregory KERSTEN and Stanley ZIONTS 90/21 Roy SMITH and FIN Ingo WALTER 90/22 Ingo WALTER FIN 90/23 Damien NEVEN 90/24 Lars Tyge NIELSEN 90/25 Lars Tyge NIELSEN 90/26 Charles KADUSHIN and OB/BP Michael BRIBE 90/27 Abbas FOROUGHI and TM Tavfik JELASSI 90/28 Arnoud DE MEYER TM 90/29 Nathalie DIERKENS FIN/AC 90/30 Lars Tyge NIELSEN FIN/EP 90/31 David GAUTSCHI and MKT/EP Roger BETANCOURT 90/32 Srinivasan BALAK- SM RISHNAN and Mitchell KOZA "Tax-Driven Regulatory Drags European Financial Centers in the 1990's", January 1990 "Information Asymmetry and Equity Issues", Revised January 1990 "Managerial Decision Rules and the Estimation of Dynamic Sales Response Models", Revised January 1990 "The Effect of Computer Intervention and Task Structure on Bargaining Outcome", February 1990 "An Introduction to Group Decision and Negotiation Support", February 1990 "Reconfiguration of the Global Securities Industry in the 1990's", February 1990 "European Financial Integration and Its Implications for the United States", February 1990 "EEC Integration towards 1992: Some Distributional Aspects", Revised December 1989 "Positive Prices in CAPE", January 1990 "Existence of Equilibrium in CAPM", January 1990 "Why netvorking Fails: Double Binds and the Limitations of Shadow Networks", February 1990 "NSS Solutions to Major Negotiation Stumbling Blocks", February 1990 • "The Manufacturing Contribution to Innovation", February 1990 "A Discussion of Correct Measures of Information Asymmetry", January 1990 "The Expected Utility of Portfolios of Assets", March 1990 "What Determines U.S. Retail Margins?", February 1990 "Information Asymmetry, Adverse Selection and Joint-Ventures: Theory and Evidence", Revised, January 1990 1990 90/01 B. SINCLAIR-DESGAGNE "Unavoidable Mechanisms", January 1990 TM/EP/AC 90/02 Michael BURDA "Monopolistic Competition, Costs of EP Adjustment, and the Behaviour of European Manufacturing Employment", January 1990 90/03 Arnoud DE MEYER "Management of Communication in International TM Research and Development", January 1990 90/04 Gabriel HAVAVINI and "The Transformation of the European Financial FIN/EP Eric RAJENDRA Services Industry: From Fragmentation to Integration", January 1990 90/05 Gabriel HAVAVINI and, "European Equity Markets: Toward 1992 and FIN/EP Bertrand JACOUILLAT Beyond", January 1990 90/06 Gabriel HAVAVINI and "Integration of European Equity Markets: FIN/EP Eric RAJENDRA Implications of Structural Change for Key Market Participants to and Beyond 1992", January 1990 90/07 Gabriel HAVAVINI "Stock Market Anomalies and the Pricing of FIN/EP Equity on the Tokyo Stock Exchange", January 1990 90/08 Tavfik JELASSI and "Modelling vith MCDSS: What about Ethics'', TM/EP B. SINCLAIR-DESGACNE January 1990 90/09 Alberto CIOVANNINI "Capital Controls and International Trade EP/FIN and Jae VON PARK Finance", January 1990 90/10 Joyce BRYER and "The Impact of Language Theories on 055 TM Tavfik JELASSI Dialog", January 1990 90/11 Enver YUCESAN "An Overview of Frequency Domain Methodology TM for Simulation Sensitivity Analysis', January 1990 90/12 Michael BURDA "Structural Change, Unemployment Benefits and El nigh Unemployment: A U.S.-European Conoarison", January 1990 90/13 Soumitra DUTTA and "Approximate Reasoning about Temporal TM Shashi SHEKHAR Constraints in Real Time Planning and Search", January 1990 90/14 Albert ANGEHRN and "Visual Interactive Modelling and Intelligent TM Bans-Jakob LOTHI DSS: Putting Theory Into Practice", January 1990 90/15 Arnoud DE MEYER, "The Internal Technological Renewal of a TM Dirk DESCHOOLMEESTER, Business Unit with a Mature Technology", Rudy MOENAERT and January 1990 Jan BARBE EP/SM FIN/EP FIN/EP 90/33 Caren SIEHL, 08 David BOWEN and Christine PEARSON "The Role of Rites of Integration in Service Delivery", March 1990 90/34 Jean DERMINE "The Gains fro. European Banking Integration, PIN/EP a Call for a Pro-Active Competition Policy", April 1990 90/35 Jae Von PARK "Changing Uncertainty and the Time-Varying EP Risk Premia in the Term Structure of Nominal Interest Rates", December 1988, Revised March 1990 90/36 Arnoud DE MEYER "An Empirical Investigation of Manufacturing TM Strategies in European Industry", April 1990 90/37 William CATS-BARIL "Executive Information Systems: Developing an TM/08/SM Approach to Open the Possibles", April 1990 90/38 Wilfried VANHONACKER "Managerial Decision Behaviour and the MKT Estimation of Dynamic Sales Response Models", (Revised February 1990) Page 1 Page 2 Page 3 Page 4 Page 5 Page 6 Page 7 Page 8 Page 9 Page 10 Page 11 Page 12 Page 13 Page 14 Page 15 Page 16 Page 17 Page 18 Page 19 Page 20 Page 21 Page 22 Page 23 Page 24 Page 25 Page 26 Page 27 Page 28 Page 29 Page 30 Page 31 Page 32 Page 33 Page 34 Page 35 Page 36 Page 37 Page 38 Page 39 Page 40 Page 41 
work_vx6n4xg2sbb3pbzlcba5y5vyo4	----	Fuzzified Expert System for Employability Assessment Procedia Computer Science 62 ( 2015 ) 99 – 106 Available online at www.sciencedirect.com 1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of organizing committee of The 2015 International Conference on Soft Computing and Software Engineering (SCSE 2015) doi: 10.1016/j.procs.2015.08.420 ScienceDirect The 2015 International Conference on Soft Computing and Software Engineering (SCSE 2015) Fuzzified Expert System for Employability Assessment Rajani Kumaria*, Sandeep Kumara, Vivek Kumar Sharmaa aJagannath University, Jaipur, India Abstract Employability is somebody’s prospective for gaining and maintaining employment. Basically employability depends on basic three parameters and these parameters are as education, understanding power and personal development. It is the capability to achieve the preliminary employment, to continue it and to obtain different one, if it is required. This paper introduced an innovative knowledgeable system for valuation of employability through some fuzzy rules. The purpose and scope of this concern research is to observe the optimal valuation for employability. This concern research considers three employability skills as an input namely Education, Understanding power and Personal development and find out a novel crisp value for employability which is basically characterize the ability of employee. This paper uses twenty seven fuzzy rules, by using Mamdani type fuzzy inference system in Mat-lab for catches solitary value of output named as employability. © 2015 The Authors. Published by Elsevier B.V. Peer-review under responsibility of organizing committee of The 2015 International Conference on Soft Computing and Software Engineering (SCSE 2015). Keywords: Employability Skills; Fuzzy Logic; Personal Developmen; Understanding Power; Education 1. Introduction The two highest apprehensions of employers are finding good workers and prepare them. The skills gap is a gap between the skills desired on the job and those influenced by applicants. Companies would prefer to appoint people who are trained and prepared to go to work; they are generally willing to deliver the job specific training which essential for those lacking such skills. Most negotiation concerning workforce ultimately turn to employability skills. Employability skills are the elementary skills essential for receiving, observing, and doing well on a job. These are * Corresponding author. Tel.: +91-946-210-3099 E-mail address: rajanikpoonia@gmail.com. © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of organizing committee of The 2015 International Conference on Soft Computing and Software Engineering (SCSE 2015) http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.08.420&domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.08.420&domain=pdf 100 Rajani Kumari et al. / Procedia Computer Science 62 ( 2015 ) 99 – 106 the skills, attitudes and activities that allow workers to acquire along with their associated workers and administrators to make reverberation and precarious decisions. Employability skills are in general divided into three skill sets first is basic academic skills (Education), second is personal qualities (Personal Development) and third is higher-order thinking skills (Understanding Power). The employability skill education includes reading, writing, oral communication and listening. The second one includes learning, reasoning, thinking, creatively decisions, making problem solving. The employability skill personal development includes self-confidence, self-control, social skills, honest, integrity, adaptable, self-motivated, self-management, self-directed, good work, attitude, well groomed and cooperative. The devotion within advanced education to increasing student’s employability potentially helps a number of significant purposes. First, it reacts to students’ motivations for inflowing higher education. A survey of school students originate that the most significant personal reasons mentioned for going to university were beside to study a subject that really suits me and three vocationally oriented causes to increase job prospects, to have a professional career and to achieve entrance to a well-paid career. Each of these four causes was evaluated by around four- fths as extremely or very important [1]. Where tuition fees are payable, such vocational inspirations are likely to be supported. Second, it reacts to policy concerns in two salutations; an important part of the motivation for the huge sums which the Government finances in higher education is the involvement which it creates to the development of the country’s human capital [2]. The further employable students are the superior the economic revenue is likely to be from this investment. Increasing higher education is also considered to attend social-equity goals by growing access for disadvantaged groups. To achieve such goals, devotion needs to be rewarded not only to confirming the participation of these groups in higher education but also to enhancing their consequent success in the employment market [3].Third, faraway from depression wider academic principles it can be deduced as reinforcing such values in three salutations by underlining generic aptitudes rather than straight subject relevance, it can help to fight stealing occupational in terms of course content, and to legitimize the remaining value of traditional academic corrections. Over two-thirds of graduate opportunities in the UK are for graduates in any subject [4]; companies tend to be much more concerned with standard graduate attributes than with subject knowledge [5]. A detailed study of employability in Information Communication Technology carried out by M. de Hoyos [18]. Employability in Europe: enhancing post graduate complementary skills training projected by GP Wall and CP Welsch [19]. Factors Influencing the achievement of Employability Aids by Students of Selected Technical Secondary School in Malaysia proposed by Jovinia Danialand Shamsiah Mohamed [20]. Research for Presentation of the “Internet of Things” in University's Training Management proposed by CJ Zong, BX Jia, and Y Zhang [21]. 2. Employability Employability is a set of three achievements that are skills, understandings and delicate attributes which creates gradates more probable to increase employment and be successful in their preferred occupations, which profits themselves, the community and the budget, the labor force. It moreover boosts student’s capability to protected rewarding and satisfying outcomes in their social, economic and the public lives. In the present scenario almost each job includes the need for staff that communicates professionally with each other, understand the needs and give good service to their clients, work in a team, adaptability, willingness and flexibility. Employability has three abilities: Achieve preliminary employment Preserving employment Gaining new employment if required The most employability skills which look in potential employees areas are as follows: 2.1. Teamwork Teamwork incorporates the working with others to accomplish results and identifying the value of others contributions and concepts. The effective teamwork skills are established by taking different roles in a team, functioning independently or as a part of a team, giving beneficial criticism, being able to recognize strengths and faults of team participants and working with people of different religions, genders, races or political influences. 101 Rajani Kumari et al. / Procedia Computer Science 62 ( 2015 ) 99 – 106 Fig. 1. Employability Skills 2.2. Planning and Organizing Planning and Organizing involves the ability to recognize that what is essential in a specified situation and to manage persons and resources efficiently to succeed results. It also includes being able to manage time powerfully and primacies what tasks essential to be completed to achieve an inclusive goal. The effective Planning and Organizing skills are established by managing time and primacies, allocating people and further resources to tasks, establishing clear project goals and deliverables, collecting, analyzing and organizing information and time management. 2.3. Self-Management Self-management skills rise to the ability to revenue duty for your o wn schedules and to set objectives and effectively complete them. It includes setting practicable goals and using your time and resources successfully to achieve them. The effective self-management skills are established by taking responsibility, expressing one's ideas and visualization, planning fast and consuming a personal vision and goals and estimating and observing one's own performance. 2.4. Communication Communication is maybe the most required after skills by most employers and includes fundamentals such as being a worthy listener, explaining things to persons from different circumstances and presenting a perfect case. EMPLOYABILITY SKILLS 102 Rajani Kumari et al. / Procedia Computer Science 62 ( 2015 ) 99 – 106 The effective self-management skills are established by negotiating, writing and speaking in languages other than English and listening and understanding evidence. 2.5. Problem Solving and Creativity Problem Solving and Creativity involves being able to propose an explanation to a problem by examining a situation and employed out how to reach at a satisfactory outcome. It often includes making optimal use of accessible resources and procuring others to achieve an outcome. The effective self-management skills are established by solving problems in groups, applying a variety of approaches to problem solving and determining customer complaints satisfactorily. 2.6. Learning Learning skills refers to your ability to accomplish your own knowledge and contribute to ongoing enhancement and growth in your own skill set and knowledge. The effective self-management skills are established by contributing to the knowledge community at the workstation, open to innovative thoughts and techniques and prepared to spend time and effort into learning innovative skills. 2.7. Use of Information Technology Information Technology involves being able to preserve abreast of present technology and apply it to difficulties, as well as the capability to embrace life-long knowledge in the field of technology. The effective self- management skills are established by being prepared to acquire new IT skills, selecting the suitable technology for a specified task and having a variety of basic IT skills. 2.8. Leadership The act of leading a group of people, an association, or the talent to do this is leadership. The skill to persuade others to act in order to accomplish a common goal and to use the skills and knowledge of team members to work productively together is leadership. Leadership in association has a specific focus on decision-making leadership in large. Organizations and is an effort at bridge the inlet between academics and management practitioners. 3. Fuzzy Logic Knowledge occurs in two different ways, one is the objective knowledge that occurs in mathematical system is used in engineering difficulties and another one is subjective knowledge that occurs in linguistic system, which is generally impossible to calculate. Fuzzy Logic can synchronize these two systems of knowledge in a logical technique. It deals with reasoning that is approximate rather than accurately gathered from classical predicate logic. The theory of Fuzzy Logic was founded by Lotfi Zadeh [22], a professor of computer science at the University of California. Fuzzy Logic is a problem-solving controller system approach that provides itself to implementation in systems ranging from small, simple, multi-channel PC, networked, or workstation-based data acquirement and control systems. It provides a modest way to reach at a definite assumption based upon ambiguous, vague, noisy, imprecise, or missing input material. It can be applied in software, hardware or a combination of both. It includes a modest, rule-based IF X AND Y THEN Z methodology to a solving control problem rather than trying to model a system mathematically. This model is empirically-based, trusting on an operator's knowledge rather than their technical appreciative of the system. It involves some numerical constraints in order to activate such as what is measured substantial error and substantial rate-of-change-of-error. It measured as an enhanced technique for organizing and handling data but has verified to be an outstanding choice for various control system applications meanwhile it simulators human control logic. A fuzzy set is a set that allows its members to have different degree of membership, called as membership function. The interval of membership function is [0 1]. 103 Rajani Kumari et al. / Procedia Computer Science 62 ( 2015 ) 99 – 106 Innumerable applications of fuzzy logic have pointed a way for an effective exploitation of fuzzy logic in the framework of difficult processes. H. Ramazi and A. Amini [6] applied fuzzy logic control for compiling multi geohazards macro-zone maps. It applied for find flexibility of protein motifs by L.A. Zadeh[7]. Fuzzy logic used to outline differences between an assortment of poly nucleotides by Y. Huang and Y. Li [8]. Adaptive connotation of fuzzy theory [7] analyzed data for experimental appearance [9] by Z. Xiu-fen, P. Zi-shu, K. Le-shan and Z. Chu-yu. A multi-variable fuzzy logic control system for a class of distributed parameter systems proposed by YQ Ren, XG Duan, HX Li, CLP Chen [10]. PID controller enhanced by I. Pan and S.Das [11]. Active bus suspension system designed with help of fuzzy logic control by M. Turkkan and N. Yagiz [12]. A fuzzy matchmaking based system- oriented grid scheduler proposed by A.I. Saleh [16]. Fuzzy logic control in air conditioning system [13], ducting system [14], CPU scheduling [15] and Job shop scheduling [17] employed by R. Kumari, V. K. Sharma and S. Kumar. C. W. Chen uses neural-network-based fuzzy logic control in a nonlinear time-delay chaotic system [23]. C. Cecati et al. developed a multilevel inverter for photovoltaic systems using fuzzy logic control [24]. B. Y. Shih et al. developed a self-governing navigation system for transistor frequency documentation portable robot e-book reader [25]. A detailed list of applications of fuzzy logic control listed in [26]-[29]. 4. Proposed Work This paper initiates a ground-breaking expert system for appraisal of employability by using certain fuzzy rules. These fuzzy rules are mainly used for examine the optimum valuation for employability. The employability deals with certain fuzzy logic rules and these fuzzy logic rules are based on employability skills. The concern research is used to compute the Employability Skills for any employee with the help of Mamdani type inference. Fig. 2. Architecture of the Proposed System This paper use appropriate linguistic variables as input and output for compute a crisp value for employability skills. The linguistic input variables named as Education (E), Understanding Power (UP) and Personal Development (PD) measured as Low, Medium and High. And the linguistic output variable named as Employability skills (ES) measured as Very Low, Low, Medium, High and Very High. The proposed skills is an assortment of linguistic fuzzy rules which define the relationship between defined input variables (E, UP and PD) and output (ES). 4.1. Fuzzy Rules Table 1 covers the range of input variables named as education, understanding power and personal development. Table 2 outlines the range of output variable named as employability. Table 3 outlines the experimental results. This concern research use the twenty seven rules which are founded based on the IF THEN statement such as IF E is low and UP is low and PD is low THEN ES is low FUZZIFICATION DEFUZZIFICATION INFERENCE ENGINE FUZZY LOGIC RULES (27) Education Personal Development Understanding Power Employability 104 Rajani Kumari et al. / Procedia Computer Science 62 ( 2015 ) 99 – 106 These rules compute the crisp value using centroid defuzzification method of Mamdani inference in Matlab that represents the employability skill of each and every employee. Figure 3 outlines rules of employability. Figure 4 shows the surface viewer of employability. Table 1. Range of input variables “Education”, “Understanding Power” and “Personal Development” Education Understanding Power Personal Development Range Low Low Low 0-4 Medium Medium Medium 2-8 High High High 6-10 Fig. 3. Rules of employability Table 2. Range of output variable “Employability” Employability Range Very Low 0-2 Low 1-4 Medium 3-6 High 5-8 Very High 7-10 105 Rajani Kumari et al. / Procedia Computer Science 62 ( 2015 ) 99 – 106 Fig 4. Surface viewer of employability Table 3 Experimental Results Education Understanding Power Personal Development Employability 1 2 1 0.75 4 2 3 1.98 1 2 9 2.5 6 5 8 6.5 7 8 8 8.76 10 9 9 8.98 5. Conclusion This paper anticipated a fuzzified expert system for employability assessment. The apprehension research finds the capability or level of several employees through these employability skills. The proposed expert system is useful for organization to calculate employability smoothness for different persons in a very simple manner. Through the proposed expert system employer can easily filter best suitable applicants based on their education, understanding power and personal development. This concern system manipulates above three inputs based on fuzzy rules and calculates employability. In future this system may be improved using more appropriate rule base as per requirements. Neuro fuzzy system may be used instead of fuzzy control system. 106 Rajani Kumari et al. / Procedia Computer Science 62 ( 2015 ) 99 – 106 References [1] Connor H., R. Burton, R. Pearson, E. Pollard, and J. Regan. Making the Right Choice: How Students Choose Universities and Colleges, Appendix 2, the year 11/S4 survey. London: CVCP, 1999. [2] Dearing R., Higher education in the learning society [Dearing report]. 1997. [3] Morey A., L. Harvey, J. Williams, A. Saldana, P. Mena. HE Careers Services and Diversity. Manchester: Careers Services Unit. 2003. [4] Prospects G. Prospects Directory salary and vacancy survey. Graduate Market Trends, Winter, 2005:11-17. [5] Harvey L., S. Moon, V. Geall, & R. Bower. Graduates’ Work: Organisational Change and Students’ Attributes. Birmingham: Centre for Research into Quality, University of Central England in Birmingham & Association of Graduate Recruiters. 1997. [6] Ramazi H., A, Amini. Fuzzy logic application in compiling multi geohazards macro-zone maps; case study: Rahdar, 1: 25,000 Quadrangle, Khuzestan, Iran. Arabian Journal of Geosciences 7, no. 8 :2014: p. 3243-3249. [7] Zadeh L. A. Fuzzy logic, neural networks, and soft computing. Communications of the ACM 37(3);1994: p. 77-84. [8] Huang Y., Y. Li, Prediction of protein subcellular locations using fuzzy k-NN method, Bioinformatics, Vol. 20(1); 2004: p. 21-28. [9] Xiu-fen Z., P. Zi-shu, K. Le-shan, Z. Chu-yu, Evolutionary computation techniques for Protein structure prediction: A Survey, Wuhan University Journal of Natural Sciences, Vol. 8(1B);2003: p. 297-302. [10] Ren Y.Q., X.G. Duan, HX Li, CLP Chen, Multi-variable fuzzy logic control for a class of distributed parameter systems, Journal of Process Control 23(3); 2013: p. 351-358. [11] Pan I., S. Das, Enhancement of Fuzzy PID Controller with Fractional Calculus, In Intelligent Fractional Order Systems and Control. Springer Berlin Heidelberg. 2013: p. 159-193. [12] Turkkan M. and Yagiz N., (2013) Fuzzy logic control for active bus suspension system, Journal of Physics: Conference Series. Vol. 410(1); 2013: p. 012006. IOP Publishing. [13] Kumari R., S. Kumar, V. K. Sharma. Air Conditioning System with Fuzzy Logic and Neuro-Fuzzy Algorithm. In Proceedings of the Second International Conference on Soft Computing for Problem Solving (SocProS 2012), December 28-30, 2012, pp. 233-242. Springer India, 2014. [14] Kumari R., V. K. Sharma, S. Kumar. Two Way Ducting System Using Fuzzy Logic Control System. International journal of Electronics and Communication Engineering & Technology (IJECET), 4(7); 2013: p. 309-317. [15] Kumari R., V. K. Sharma, S. Kumar. Design and Implementation of Modified Fuzzy based CPU Scheduling Algorithm. International Journal of Computer Applications 77(17); 2013:1-6. [16] Saleh A.I. An efficient system-oriented grid scheduler based on a fuzzy matchmaking approach. Engineering with Computers, 29(2); 2013: p. 1-22. [17] Kumari R., V. K. Sharma, S. Kumar. Fuzzified Job Shop Scheduling Algorithm. International Journal of Technology Innovations and Research (IJTIR), vol 7; 2014: p. 1-20. [18] de Hoyos M., A. Green, S. A. Barnes, H. Behle, B. Baldauf, D. Owen. Literature Review on Employability, Inclusion and ICT, Part 2: ICT and Employability. No. JRC78601. Institute for Prospective and Technological Studies, Joint Research Centre, 2013. [19] Wall G. P., C. P. Welsch. Employability in Europe: enhancing post graduate complementary skills training. Proceedings of the HEA STEM Learning and Teaching Conference. 2013: p. 151-157 [20] Danial J., S. Mohamed. Factors Influencing the Acquisition of Employability Skills by Students of Selected Technical Secondary School in Malaysia. International Education Studies 7(2);2014: p.117. [21] Zong C. J., B. X. Jia, Y. Zhang. Research on Application of the Internet of Things in University's Teaching Management. Advanced Materials Research 860; 2014: p. 3017-3020. [22] Zadeh L. A. Fuzzy sets. Information and control 8(3);1965: p. 338-353. [23] Chen C. W., RETRACTED: Applications of neural-network-based fuzzy logic control to a nonlinear time-delay chaotic system. Journal of Vibration and Control 20(4);(2014: p. 589-605. [24] Cecati C., F. Ciancetta, P. Siano. A multilevel inverter for photovoltaic systems with fuzzy logic control. Industrial Electronics, IEEE Transactions on 57(12); 2010: p. 4115-4125. [25] Shih, B. Y., M. C. Lin, C. Y. Chen. Autonomous navigation system for radiofrequency identification mobile robot e-book reader. Journal of Vibration and Control.;2012. [26] Hayward G., V. Davidson. Fuzzy logic applications." Analyst 128(11); 2003: p. 1304-1306. [27] Yager R. R., L. A. Zadeh. An introduction to fuzzy logic applications in intelligent systems. Kluwer Academic Publishers.1992 [28] Yen J. R. Langari, L. A. Zadeh. Industrial applications of fuzzy logic and intelligent systems. IEEE Press. 1995. [29] Ross T. J., Fuzzy logic with engineering applications. John Wiley & Sons. 2009. 
work_vxyqed2ltzfyflzv6tdionhim4	----	VALIDATION OF EXPERT SYSTEMS­ WITH APPLICATIONS 10 AUDITING AND ACCOUNTING EXPERT SYSTEMS* Daniel E. O'Leary Graduate School of Business, University of Southern California, Los Angeles, CA 90089 ABSTRACT This paper proposes a set of definitions for the concepts "validation" and "assessment" applied to expert systems (ESs). It develops a framework for this validation and demonstrates the framework on existing accounting and auditing ESs to elicit some of the research issues involved in ES validation. Validation is critical to the design and implementation of decision-making ESs. In a setting where objectivity is sought and variance is avoided, validation ascertains what a system knows, knows incorrectly. or does not know. Validation ascertains the system's level of exper­ tise and investigates the theoretical basis on which the system is based. It evaluates the reliabili­ ty of decisions made by the system. The validation framework developed in this paper is research methods. It is designed to reflect the unique aspects of ESs (in contrast to other types of computer programs) and can be used by ES developers as a basis from which to perform validation and by researchers as a framework to elicit research issues in validation. Subject Areas: Accounting Theory and Auditing. INTRODUCTION In addition to the processes of designing, developing, and implementing ex­ pert systems (ESs), validation is important to the decision-making success of a sys­ tem and to the continued use of an ES. An ES that has not been validated suffi­ ciently may make poor decisions. This can lead to a loss of confidence in the par­ ticular ES or in other systems, resulting in discontinued use and financial loss. A number of different approaches to validating particular auditing and ac­ counting ESs [16] [17] [18] [38] [39] and medical ESs [52] [8] [7] have been reported. Validation of general ESs [21] [33] and potential bases for the validation of ESs [4] also have been discussed. This paper presents a theory-based framework that is useful not only for guiding the validation of an ES, but also for eliciting other validation research issues. Validation of ESs Developing, designing, and implementing systems for making expert decisions requires analyzing the knowledge base and the decision-making capabilities of the system. That process, referred to as validation, requires 1. ascertaining what the system knows, does not know, or knows incorrectly *The author would like to thank Doug Andrews, Nils Kandelin, Chen-en Ko, and Paul Watkins from the University of Southern California 1986 Conference on Expert Systems; participants at the American Accounting Association 1986 national meeting; Paul Watkins and the anonymous referees 468 1987J O'Leary 469 2. ascertaining the level of expertise of the system 3. determining if the system is based on a theory for decision making in the particular domain 4. determining the reliability of the system Once these concerns have been satisfactorily addressed, the system is updated to reflect the findings and may be revalidated. The process should be performed in an environment designed to provide an objective and cost-effective validation. Validation ascertains what the system knows, does not know, or knows incor­ rectly. For example, when the expert system Rl (a system for configuring computers used by Digital Equipment [36]) was validated, errors and omissions in the knowl­ edge base were identified. These errors and omissions were corrected before the system was placed in service [21]. Validation ascertains the level of decision-making expertise of the system. The types of problems a system can solve and the quality of its solutions define the level of expertise. For example, in education it is common to define an individual's level of accomplishment based on the types of problems that the individual can solve and the quality of the solutions produced. Validation determines if the ES is theory based. Davis [12] [13] argued that basing an expert system on a theory is an efficient approach to developing such a system. Lack of a theory base has resulted in the failure of at least one system [48]. Validation analyzes the reliability ofthe ES. Given similar inputs, the ES should produce similar outputs. In addition, before and after revalidation a system generally should give similar responses to similar sets of test problems. Validation Process Previous analyses of ES validation have stressed the importance of periodic informal validation rather than a single, formal validation at the end of a project [44]. This validation will not be the same in each situation but will differ in formal­ ity and in the extent to which the validation process is implemented. As suggested by software engineering [44] and ES validation [2l], formal acceptance testing may be appropriate. (A formal acceptance test might consist of a test where the user formally signs off on the quality of the decisions made by the system.) ES Assessment Although the focus of this paper is on ES validation, other issues addressed in developing, designing, and implementing ESs do not relate to decisions that the system makes; nevertheless, these other issues do contribute to the overall success of a system. The assessment process involves analyzing the user interface and sup­ porting the quality of the development effort. For example, it involves 1. ascertaining the quality of the user interface [31] .. of an earlier version of this paper; and, in particular, John Henderson for their comments. Earlier ver­ sions of this paper were presented at the University of Southern California Symposium on Expert Systems, February 1986, and the American Accounting Association National Meeting, August 1986. ,' 470 Decision Sciences [Vol. 18 2. evaluating the documentation of the system (a typical weakness of most systems, with good reason; since ESs are evolutionary, the documenta­ tion does not evolve at the same rate as the rest of the system) 3. determining what language or shell should be used to develop the system 4. analyzing the quality of the system programming Paper Objectives This paper has four primary objectives: to propose a set of definitions for the concepts "validation" and "assessment" applied to ESs; to develop a framework for the validation of ESs; to demonstrate that framework on some previously pub­ lished auditing and accounting ESs; and to use that demonstration to elicit some of the research issues involved in ES validation. The paper proceeds as follows. The objective of developing a framework is to take advantage of the unique aspects of ESs. Therefore the next section discusses some unique aspects of ESs in general and of auditing and accounting ESs in par­ ticular. Using a research methods approach, these unique characteristics are incor­ porated into a framework for ES validation. The framework is demonstrated on accounting and auditing ESs and that demonstration is used to elicit some of the research issues involved in the validation of ESs. UNIQUE CHARACTERISTICS OF AUDITING AND ACCOUNTING ESS If ESs are like other computer systems, then validating an ES should be the same as validating any other computer system. However if ESs are different, then their unique characteristics can be used to develop a specific framework for their validation. A number of technical, environmental, design, and domain charac­ teristics distinguish ESs from other computer-based systems. The technical aspects which distinguish ESs include the following. First, ESs process symbolic information (e.g., "If... then" rules) rather than just numeric infor­ mation [1] [43]. This ability to process symbolic information allows ESs to solve nonnumeric-like problems, which generally are less precise than numeric problems. Second, experience with representing knowledge shows that a fraction (less than 10 percent) of this knowledge escapes standard representation schemes and requires special "fixes" [19] to make it accessible. Since other systems do not use knowledge representation, they do not face this problem. Third, ESs often are developed using either artificial intelligence (AI) languages or ES shells [25]. An AI language is a computer language aimed at processing symbolic information (e.g., Prolog [11] or Lisp [5l]). An ES shell is software designed to aid the development of an ES by prespecifying the inference engine and making it easier to input knowledge into the system. Some of the first ES shells were EMYCIN [7] and AL/X [15]. More recently developed shells include Texas Instrument's Pef§Onal Consultant Plus, Teknowledge's M.I, and Inference Corporation's ART. The characteristics of ES shells and AI languages (e.g., their ease of examination by nonprogrammers) can be made a part of an ES validation framework. .­ 1987] O'Leary 471 Environmental characteristics which distinguish ESs include the following. First. ESs directly influence or make decisions [1] [25] [49]. Other systems (e.g., decision support systems (DSSs» simply support decision making or have an in­ direct impact on decisions. Second, the expertise being modeled by an ES generally is in short supply, is an expensive resource, or is not readily available at a particular geographic location [20]. This is in contrast to other computer systems (e.g.• account­ ing systems) where generally a number of personnel understand what the system is designed to accomplish. In addition to these characteristics that differentiate ESs from virtually all other types of computer systems, the dominant design methodology for both ESs and DSSs differentiates them from traditional computer systems. First, ESs (and DSSs) often are developed using a "middle-out" design rather than the traditional data processing approach of top-down or bottom-up design. A middle-out design phi­ losophy starts with a prototype and gradually expands the system to meet the needs of the decision [27]. Second, like DSSs, ESs evolve over time-the system changes as the decision-making process gradually is understood and modeled [28]. As a consequence, traditional validation models based on other design philosophies are not likely to meet the needs of an ES. Finally, some domain characteristics of auditing and accounting decisions (and other business-based decisions) often distinguish these decisions from decisions made in other domains. First, in cpntrast to some scientific decisions that have a unique solution (such as those represented by the ES DENDRAL [1]), auditing and accounting decisions generally do not have a single solution. Second, the deci­ sion reached often can be evaluated only by how similar it is to decisions other decision makers develop (i.e., by consensus); there may be no way to rank the deci­ sions a priori. Third, different schools of thought may represent alternative knowl­ edge bases. This means experts may not agree on a recommended solution. (This can apply to other disciplines as well.) Fourth, a "good" decision does not necessarily result in "good" consequences. Decisions are based on information available at the time they are made; this does not guarantee desirable consequences at a later time. Fifth, in contrast to some disciplines where a decision has no direct dollar value, the decision modeled by an ES can have substantial monetary value. VALIDATION FRAMEWORK Software engineering encompasses the general set of tools and techniques that aid programmers in software development. Since ESs are computer programs, soft­ ware engineering might appear to be a likely candidate to supply a framework for ES validation. However, the unique characteristics of ESs in general, and of account­ ing and auditing ESs in particular, indicate that such a framework will not be appropriate. An examination of one such appro8,fh [44] supports this view. First, in soft­ ware engineering the focus of validation is on finding errors in the program. The validation of ESs is more than just a process of finding errors. Here the validation process meets the development needs of derming the level of expertise of the system, ," 472 Decision Sciences [Vol. 18 identifying what the system can and cannot do, understanding the quality of the decisions produced by the system, and describing the theory on which the system is based. Because software engineering generally is not concerned directly with human expertise, it is not directly concerned with these issues. Second, the unique characteristics of ESs indicate they are different from other computer programs. Since ESs process symbolic information, use ES shells and AI languages, and require special fixes in their knowledge bases, they also require different validation approaches than other computer programs. Further, ESs directly affect decisions and they model expertise that is in short supply; the environment in which they operate is different from that of other computer programs. Third, the domain-based decisions of auditing and accounting ESs often are not well enough understood to use traditional software engineering approaches. While software engineering generally uses structured top-down or bottom-up ap­ proach in the design and evaluation of software, ESs are evolutionary and often are developed using a middle-out approach. One alternative to the software engineering approach is a research methods approach. This approach views the development of ESs as experimental representa­ tions of human expertise, that is, as research designs. Kerlinger defined research design as "the plan, structure and strategy of investigation conceived so as to ob­ tain answers to research questions and to control variance" [30, p. 3001. In a similar sense, validation can be defined as the p~an, structure, and strategy of investiga­ tion conceived so as to obtain answers to questions about the knowledge and deci­ sion processes used in ESs and to control variance in that process. Kerlinger also noted, "research designs are invented to enable the researcher to answer research questions as validly, accurately, objectively, and economically as possible" [30, p. 301]. He also noted that accuracy consists of four concepts: reliability, systematic variance, extraneous systematic variance, and error variance. We use these characteristics of research design (validity, objectivity, economics, and accuracy) to formulate a framework (Thble 1) for the validation of ESs. Since Kerlinger noted the existence of three types of validity in research methods (con­ tent validity, criterion-related validity, and construct validity), each will be treated separately. Content Validity "Content validity is the representativeness or sampling adequacy of the content­ the substance, the matter, the topics" [30, p. 4581. In validating ESs, content validity refers to ascertaining what the system knows, does not know, or knows incorrectly. This can be operationalized in at least two ways: by direct examination of the system components or by testing the system. Direct Examination of the System Components. An ES is based on the knowledge of experts. Accordingly, it is importfnt that these experts know what is contained in the system. The expert can examine the knowledge base directly in one of two ways. First, the ES could develop, for example, a list of the rules in its knowledge base for periodic review or a summary of the process it uses for " 1987] O'Leary 473 Table 1: Summary of validation framework. 1. Content validity a. Direct examination of the system by the expert b. System test against human experts (Thring test) (1) Intraexpert test (2) Interexpert test c. System test against other models 2. Criterion validity a. Definition of the level of expertise of the system (1) Human evaluation criteria (2) lest problems to define the level of expertise (3) Quality of responses defined b. Knowledge-base criteria c. Clarification of evaluation criteria 3. Construct validity 4. Objectivity a. Programmer validation b. Independent administration of validation Co Sponsor/end-user validation d. Biasing and blinding e. Different development and test ~ata 5. Economics (cost-benefit) 6. Reliability a. System test against itself (sensitivity analysis) b. lest problems for revalidation 7. Systematic variance (experimental variance) a. Problems reflecting range of problems encountered b. Variation in the test problems c. Number of test problems d. 'TYpe I and 'TYpe II errors 8. Extraneous variance a. Complexity of the system b. ES's location in the system life cycle c. Recognition, examination, and testing of special fixes d. Location of judges during testing e. Learning on part of judges 9. Error variance the inference engine. Second, the expert could examine the storage of the informa­ tion by the ES directly. The first solution might produce reports that could be read easily but, being an intermediate step, it also would require validation. As a result, it would be preferable if the expert could examine the ,nowledge base directly. In the second case, the primary concern is the format of the knowledge to be reviewed. If the system were built using an ES shell or an AI language such as Prolog, direct review likely is feasible. ---................... ..... -----­~-- 474 Decision Sciences [Vol. 18 A knowledge expert might not be able to investigate the inference engine to see whether it is correct because of the complexity of the computer code. However, if an ES shell is used, the inference engine normally would be prespecified. The direct examination of the components has some limitations. First, fear of computers [21] may generate hostility toward the system. Second, human infor­ mation processing is limited. Direct examination may not correctly process the links between parts of the knowledge base; also, the breadth of information contained in the knowledge base could limit the success of a direct investigation. Third, the direct approach may require substantial resources. If the expertise is scarce, purchas­ ing expertise may be costly. Fourth, direct examination may be a tedious job that could lead to errors in the validation. Fifth, if the knowledge base is large or com­ plex, examination could prove very time consuming and complex and lead to infor­ mation overload. Sixth, current technology is not designed to facilitate direct examination. Further, any direct analysis of the knowledge base is limited to looking at the pieces of the base and not at how they interact. In addition, translation of the com­ puter code makes any direct analysis of the heuristics in the inference engine dif­ ficult. Accordingly, the best solution is to test the system as a whole. System Test against Human Expert. If the system is designed to perform as an expert, it should be tested against an expert(s). To ensure that the ES has cap­ tured the expertise of the expert it Wij.S designed around, its decisions should be compared to that expert's. However, such an intraexpert test procedure has the potential to introduce bias into the validation process through the acquisition of information from memory or the manner in which information is processed [26]. In terms of ESs, this means that the knowledge base or the relationships between sets of rules or the heuristics in the inference engine all may contain bias. This suggests that the ES also must be tested against other experts from the same "school" as the original expert but who are not biased or invested in the particular ES. Since such interexpert tests when conducted using experts from alternative schools may yield contradictory decisions, this type of comparison is not recom­ mended. However, alternative views of the world might produce additional knowledge for the system. System Test against Other Models. System tests against human experts may be preferred to tests against other models. However, tests against human experts can be expensive. Experts also face time constraints [7]. Since the system is a model, one important characteristic should be its relationship to other models. An ES should 1. perform in a similar or better fashion than other models for the same problem 2. be able to solve problems that are not amenable to other solution methodologies ,. One type of model used to analyze decisions is regression analysis [6] [32] where independent variables represent the variables used by a decision maker. Other types of models may be used, but the model used largely is a function of the problem " 1987] O'Leary 475 and previous recommended solutions to the problem. In particular, the preferred model would be the one that has provided the best solution to the problem so far. For example, an ES for production scheduling would be tested against an existing operations research model. Unfortunately, in some cases comparison of the ES to regression or to other approaches may not be feasible because of the structure of the ES knowledge base, (e.g., if "If. .. then" statements are present). It may be very difficult to translate such rules into numeric variables. However, simulation generally is a feasible alter­ native [9] [10]. Criterion Validity "Criterion-related validity is studied by comparing test or scale scores with one or more external variables, or criteria, known or believed to measure the attri­ bute under study" [30, p. 459]. In ES validation, this refers to the criteria used to validate the system, for example, to ascertain the level of the system's expertise. The primary criterion for system validation is the relationship between the deci­ sions developed by the system and decisions developed by human experts. In AI this is referred to broadly as a Thring test. However, this sort of relationship is not evaluated easily. Difficulties arise for a number of reasons: differing definitions of the level of expertise, differing kn<?wledge-base criteria, and lack of clarity about what is to be evaluated. Definition of the Level ofExpertise. One way to define and measure the level of expertise the ES has attained is to use the same criteria identified for defining expertise in human experts [7]. Where feasible, this is a good option. However, in many auditing- and accounting-based decision-making situations specific criteria are not available for specific decisions. Instead, a portfolio of decisions is subject to a set of criteria. For example, managers may be evaluated on their monthly prof­ it and loss statements. These statements reflect a portfolio of decisions. An alternative approach is to evaluate the system's performance on a set of test problems. There are at least two possible ways to do this. First, in analyzing a human's knowledge, the problems that the person can solve determine his/her level of accomplishment. Similarly, the difficulty of test problems that a system can solve dermes the system's level of expertise. Second, in analyzing a human's knowledge, the qUality of the solutions also helps determine the level of accomplish­ ment. Similarly, the quality of the responses to test problems defines the level of expertise of an ES. Quality, of course, is a difficult concept to measure; its defini­ tion often is situation-based. However, with humans the number of correct responses obtained often is used to measure quality. This criterion also has been used to validate ESs. Knowledge-Base Criteria. There are three generic knowledge-base criteria that suggest what to test: consistency, accuracy, and completeness. Consistency refers to the relationship between the informati<,n in the knowledge base and the ability of the inference engine to process the knowledge base in a consistent manner. Ac­ curacy refers to the correctness of the knowledge in the knowledge base. Complete­ ness refers to the amount of knowledge built into the knowledge base. " 476 Decision Sciences [Vol. 18 Clarification of Evaluation Criteria. Although general approaches may be developed to test the above generic criteria, a particular application also may re­ quire specific evaluation criteria [21]. For example, when evaluating an automobile a number of criterion (such as luxury, economy, and sportiness) might be developed. Particular measures then must be developed to characterize these concepts. For ex­ ample, ''miles per gallon" can be used to measure economy. Because many measures are possible, researchers suggest that the specific criteria and characterizations to be used be established and agreed on before validation begins. Construct Validity Construct validity refers to those constructs or factors that the test is designed to discover. "The significant point about construct validity, that sets it apart from other types of validity is its preoccupation with theory [and] theoretical constructs" [30, p. 461]. In terms of ESs, this type of validity indicates the importance of the existence of a theory on which the system is based. Construct V'dlidity suggests that a purely empirical approach may be inappropriate. Davis [12] [13] suggested that an empirical approach is not as efficient as an approach based on a theory (or at least on an understanding of the problem at hand) when developing an ES. He suggested instead the use of systems that reason from first principles, that is, from an understanding of causality in the system being examined. McDermott (see [48]) indicated that a primary reason ES development may fail is the lack of a theory on which the knowledge encoded in the ES is based. One problem with using construct validity as a criterion is that conflicting or alternative theories or first principles may be available. This can lead to difficul­ ties in establishing interexpert validation tests. Objectivity In variance terms, objectivity refers to minimizing observer variance [30, p. 491]. Accordingly, in validating an ES any judgmental variance should be minimized. This includes eliminating expert bias by administering the test independently, and using blinding techniques [7]. Programmer Validation. Because of the programmer's knowledge of the system and the nature of the development process, the system programmer generally will perform periodic informal validation on the system. If the programmer does not have a vested interest in the ES, this can be a way of developing an independent validation. However, programmers typically do not perform all the validation pro­ cesses required. They may have a vested interest in the system, or they may not understand the problem being addressed by the system. This suggests using a com­ bination of programmer and other validation. Independent Administration of Validation. The importance of independence is recognized generally in research design. I' also is recognized in accounting when a CPA performs an external audit or when internal auditors perform audits for internal purposes. If the model builder also validates the model, conflicts of in­ terest can arise. " 1987] O'Leary 477 Sponsor/End User Validation. Software engineering uses an acceptance test by the sponsor or the end user as the final step in validating the computer pro­ gram. Gaschnig, Klahr, Pople, Shortliffe, and Thrry [21] suggested a similar type of formal test for an ES. That is, the user must "sign-off' on the system. If the end user has expertise in the area, then this acceptance test can be a critical test of user acceptance. However, if the sponsor or end user does not have sufficient expertise to judge the quality of the system, he or she may not be an appropriate judge. Blinding 7echniques to Eliminate Bias. Buchanan and Shortliffe [7] reported that some human expert validators are biased against computers making certain kinds of decisions. In order to minimize this limitation, the validation of MYCIN used a "blinded" study design to remove that source of bias. Using Different Development and Test Data. Test problems often are used when developing systems to ensure consistent and reliable performance. However, it is important that the problems used to test the system not be limited to those prob­ lems used in developing the system; otherwise, these are not true test problems. Economics (Cost-Benefit) Cost-benefit decisions permeate research methods. For example, Simon noted "you will want to invest your time and energy in the work that will be the most valuable" [46, p. 100]. '. A key aspect of any economic activity is cost-benefit analysis. Since auditing­ and accounting-based ESs are developed and validated using resources and may make decisions that have economic consequences, cost-benefit analysis is a major concern in their validation. One important factor used to determine system benefit is how the system ulti­ mately will be used. The system may be used for commercial purposes or it may be a prototype designed to investigate the feasibility of developing a system for a particular purpose. In either case, benefit may be difficult to measure. 1\vo of the main factors that determine the cost of validating a system are the formality and extent of validation required. Formality indicates the breadth of application of the validation-for example, is there an independent validation of the system? Extent indicates the depth of validation of the system-for exam­ ple, how many test problems are used? Thus a prototype ES designed to determine whether a particular process can be represented as an ES will not receive a valida­ tion that has the same formality or extent ,as a commercially based system. Cost-benefit affects not only the scope of validation but also the development of the system because the amount of validation defines the extent to which, for example, it can be determined if the knowledge in the knowledge base is correct or complete. ReliabUity One synonym for accuracy is reliability. "Reliability is the accuracy of preci­ sion of a measuring instrument" [30, p. 491]. In ES validation, the stability of the .. 478 Decision Sciences [Vol. 18 system and the ability of the system to generate identical solutions given identical inputs measures the reliability of the system. System Test against Itself (Sensitivity Analysis). One possible validation pro­ cedure is the analysis of a program's sensitivity to slight changes in the knowledge base or in the weights [7]. That is, the model can be tested against itself for stability. If the system produces several solutions over minor parametric shifts, the system may be unstable. Alternatively, in certain environments a highly sensitive ES may be required. In this case it can be difficult to differentiate instability from an appro­ priate model response. Standard Test Problems. Standard test problems may be used in the revalida­ tion process. These problems should be designed to test standard system responses to ensure, for example, that additions to the knowledge base do not result in con­ tradictions. The limitations of standard problems are discussed in [21]. Systematic Variance (Experimental Variance) "If the independent variable does not vary substantially, there is little chance of separating its effect from the total variance of the dependent variable, so much of which is often due to change" [30, p. 308]. In validating an ES's performance, the test problems used need to allow the validator to distinguish between systematic variance and chance. This can be accomplished in a number of different ways. Test Problems Reflect the Range -oj Problems Encountered. Test problems should be representative of problems the expert encounters in practice. Problems should reflect not only common "middle of the road" situations, but also some of the more unusual occurrences [8]. Variation in Test Problems. Validation success is based to a large extent on the test problems used in the validation process. Variation in the problems must exist if the entire set of parameters in the model and the range of the parameters is to be tested. If the problems are not varied enough in terms of the set and range of the parameters, the validation process and the test of any variations in the system's behavior will be too limited. Number oj Test Problems. A sufficient number of test problems must be used to ensure the statistical significance of the model. A test of only a few problems will not adequately test the knowledge base or range and set of parameters nor will it provide statistical significance [8]. 1)lpe I and 1)lpe II Errors. A 'TYpe I error occurs when we incorrectly reject a null hypothesis. A lYPe II error occurs when we accept a null hypothesis that is false. Both types of errors need to be considered when validating an ES. In some decision problems (e.g., bankruptcy-no bankruptcy) a large majority of the test problems will result in null hypotheses being accepted (and only a few rejected) because of the nature of the process itself. In..these cases, there is a large possibility of 'TYpe II errors. In such a decision setting, there may be few test problems that actually test the abilities of the system. Accordingly, the test problems must be chosen carefully in light of this concern. " 479 1987] O'Leary Extraneous Variance "The control of extraneous variance means that the influences of indepen­ dent variables extraneous to the purposes of the study are minimized, nullified or isolated" [30, p. 309]. At least five extraneous variables can influence the quality of system output: complexity of the system, location in the system life cycle, special fixes, location of the judges, and learning by the judges. Complexity of the System. Complexity of the system is one of the most impor­ tant variables in determining how difficult the validation process will be in software engineering. Software engineering uses a classification schema to measure the com­ plexity of a computer program [44]: a simple program is one with fewer than 1,000 statements, written by one programmer, with no interactions with any other systems; an intermediate program is one with fewer than 10,000 statements, written by one to five programmers, with few interactions with other systems. These criteria are not reasonable standards to use for evaluating ESs. Tradi­ tional software engineering computer programs do not use a knowledge base or an inference engine. In addition, in software engineering problems the process being modeled generally is well defined. Therefore, these software engineering classifica­ tion schema cannot be used to evaluate ESs. The criteria, however, can form the basis for other measures of complexity. Unfortunately, a ready replacement for general evaluation is not available. Some suggest the number of rules in the knowledge base as a standard, but this ignores connections between the rules and the inference engine. In spite of a lack of clear criteria, complexity remains an important variable in determining how difficult the validation process will be. ES's Location in the System Life Cycle. Gaschnig et al. [21] suggested that the ES's current position in the development cycle is critical in determining the ex­ tent of validation desired. Validation is less critical in the early stages of develop­ ment (but also less expensive and less difficult) than it is in later stages. In the ear­ ly stages, a system is not very knowledgeable. Accordingly, simple validation tests will determine whether or not the system has an adequate set of working knowledge. In the latter stages, however, the ES may be facing adoption by external users and may require extensive validation. As a result, the ES's location in the life cycle is an important variable in determining the extent of the validation process. Recognition, Examination, and Testing ofSpecial Fixes. As noted in Fox, "ex­ perience with representing large varieties of knowledge show that a small fraction « 10010) escape standard representation schemes, requiring specialized 'fixes'" [19, p. 282]. That is, in order to ensure that the knowledge base is complete, special features (fixes) are added to the system. These special fixes are of concern in ES validation. They need to be summarized so that special validation procedures can be developed to address them. Since these are "special" fixes, it is difficult to generalize any further. 'l Location of Judges during Testing. A critical distinction is made in research methodology between laboratory and field test settings. In the laboratory, many variables that may be faced by the judges who will be compared to the ES can 480 Decision Sciences [Vol. 18 be controlled. On the other hand, a field setting offers a richer decision-making environment and a more realistic test of the ES. Learning during the Validation Process. Although most business-based ESs do not learn from processing decision problems, this is not necessarily true of the judges to whom the system is compared. For human judges, the ability to learn from the decision problems may be an extraneous variable. Learning could occur from the order of the test problems or from the type of test problems used. The amount of the learning due to order can be assessed by changing the order of the test prob­ lems. Learning that occurs from the process of making decisions on test problems is a more difficult problem and often is ignored in 'TIlring tests. This problem can be addressed by determining whether an expert who has experience with the test prob­ lems makes decisions similar to those made by experts with less system experience. Error Variance In a discussion of error variancf; Kerlinger noted There are a number of determinants of error variance, for instance, factors associated with individual differences among subjects. Ordinarily we call this var­ iance due to individual differences "systematic variance" But when such variance cannot be, or is not identified and controlled, we lump it with the error variance. [30, p. 3121 This kind of variance can be caused by variation in the experts' responses from trial to trial, guessing, momentary inattention, slight temporary fatiguf; lapses of memory, transient emotional states, etc. [30]. The validation framework for the ES's attempts to minimize error variance by controlling many of the variables in controlling the extraneous variance is the same as in all experimental designs. However, much of this minimization of error variance occurs by preventing errors and ambiguities, that is, by using proven measurement scales and unambiguous questions. RELATIONSHIP OF PREVIOUS AUDITING AND ACCOUNTING ESS 10 THE PROPOSED VALIDATION FRAMEWORK A limited number of auditing and accounting-based ESs currently exist. The few that are being used (or are planned for use) on a commercial level (e.g., [50] and [45]) are proprietary. Accordingly, the present analysis is limited to prototype auditing- and accounting-based ESs. The review and analysis of ESs in this paper is limited to those accounting and auditing systems generally available in the literature at the time the paper was writ­ ten (February 1986) and to information published about the validation processes used in developing and designing those systems.! Accordingly, [3], [14], [22], and .. !ESs egmined in this paper were found by egmining three sources: Peat. Marwick, Mitchell & COo's Research Opportunities in Auditing (1985) [42]. Miller's 1984 Inventory of Expert Systems [40], and Ph.D. dissertations through calendar year 1984 (listed in University Microfilms). In addition, some systems mentioned in presentations and papers the author was aware of also are included. 1987] O'Leary 481 [24] are not discussed because they are unpublished papers; [3], [14], [20], [23], [29], and [34] are excluded because they contain little or no information about how the systems were validated. The validation procedures used in those prototype ESs discussed in previous sec~ tions offer us a chance to examine the ability of our framework to capture the valida­ tion process. These findings are summarized in Thble 2. Some framework items (eco­ nomics and error variance) from Thble 1 are not included in Table 2 because of the prototype nature of the systems. This analysis is not to be considered as critical of these systems-each system was an innovative prototype effort. The framework also can be used to analyze some of the potential research issues involved in ES validation. Content Validity The validators used examined the system's overall behavior rather than individ­ ual components. Future research needs to develop cost-benefit tools that allow indi~ vidual components (such as the knowledge base) to be validated rather than simply comparing the system's overall performance to that of another expert [41]. The system tests analyzed each used human experts (in both interexpert and intraexpert tests) as the basis for validation. However, probably due to the prototype nature of the systems, only a small number of human experts generally were used. In addition, the use of interexpert tests may have introduced variance because the different experts used may have held contradictory views about the subject area. In some cases, the small number of human experts available was supplemented by the use of alternative models or standard tests of validation. Unfortunately, developing an alternative model may cost more than having a human expert serve as a standard by responding to a set of test problems. Perhaps as a result, the com­ parison models analyzed were relatively unsophisticated. Generally, little research compares the performance of ESs to other sophisticated models. Criterion Validity The similarity of solutions proposed by the ES and the human expert(s) was used as the single criterion for success of the model. This standard ignored the possibility that an ES might derive better solutions than those of a human expert. It is not unusual for other types of models to outperform their human counter­ parts [26]. However, no such comparison has been made with ESs. If ESs can in fact consistently outperform humans, then other criteria will have to be developed to ensure appropriate validation. Because the developers of these prototypes were involved in their validation, there apparently was little difficulty clarifying the evaluation criteria. However, little research has been done on developing effective evaluation criteria when outside validators are used. In some cases, identifying the level of difficulty of the test problems established the expertise of the system. For exampl~ AUDITOR [16] [17] [18] used test prob­ lems drawn from the work papers of audits. However, currently there is no general way of determing the system's level of expertise. .. 00 .". Table 2: Validation characteristics of selected accounting ESs. N Dungan and Clarkson [9] (10) Bouwman (5) Chandler [l6] [l7) [18) Michaelsen [38] Steinbart (47) Meservy (37) I. Content validity Direct examination No No No No No No Human expert comparison Yes Yes Yes Yes Yes Yes lntraexpert test Yes Yes No (multiple experts) No (author was No Yes expert) lnterexpert test No Yes (18) \b (2) Yes (2) Yes (6) Yes (3) Other model test Yes No Yes No No No (random & naive) 2, Criterion validity Level of system Expert Expert Expert Expert Expert Expert 3, Construct validity Largely empirical Largely Weights generated Books Audit manuals Prototypic ~ 4. Objectivity empirical empirically & textbooks reasoning {;i' 5' ::s Independent No No \b Yes Yes Yes 5, Reliability ~ ~' Sensitivity No No No No No No ::s 6, Systemdtic variance Actual problems Apparently NA Yes Yes Yes Experiment ~ ... (experiment) similar to actual Variation Small range Not at extremes small range; Chosen by subjects Difficult to Sought prob­ available; funds actual $ amounts ascertain lems differ- of 22k-37,Sk Oikely) ent than development problems Sample Size 1\vo sets of four Four sets of one 11 cases (one company) 1\vo cases (one Thirteen actual Three sets of cases case company) problems one case 'tYpe I and type II errors No No No No No No 7, Extraneous variance Location of expert Few controls Laboratory Office of expert Office of expert Office of expert Unclear ~ -00 " 1987] O'Leary 483 The primary knowledge-base criterion used in the studies examined was accu­ racy of the system. However, little analysis was performed by the validators to en­ sure completeness or consistency of the system. In part, this may be because few methods for analyzing the completeness or consistency of a system are available. Future research could focus on developing such tools. Construct Validity Some of the prototype ESs were tied to a theoretical construct. However, some apparently were designed to discover how decisions are made, without the benefit of a normative decision-making model. The danger in this approach is that the model may be too "expert specific" with no ability to generalize or be compared to other expert systems or humans. On the other hand, of course, one of the primary research benefits of developing an ES is to understand better the nature of the par­ ticular decision process being modeled. Objectivity Generally, the validation processes used in these studies did not have an inde­ pendent valida tor. Since the developers were involved in administering the valida­ tion, bias may have been introduced. Commercially developed packages, however, likely will have independent validation',or at the least a user acceptance test. Generally, no blinding techniques were used in administering the validation tests, and this may have affected the results. Cost-Benefit Analysis Each of the ESs examined was a prototype system. As a result, validation was minimally cost beneficial. As system development moves away from research­ oriented prototypes, however, the cost-benefit relationship favors more validation. Generally, the costs of validation can be specified, but the benefits of validation are not as easily measured. Research could be directed toward measuring the benefits of validation. Reliability Since the ESs were prototype systems, there was no need to develop test prob­ lems for revalidation. However, for commercial systems, such test problems can be used after revision to ensure that no inconsistent or incorrect information has entered in the knowledge base. Test problems, of course, test only facets of the knowl­ edge base. Research could be directed toward developing other methods to test for reliability. Further, if revising the ES adds new knowledge to the knowledge base, can tools such as test problems be used effectively? With new knowledge in the system, a test problem may have different correct answers before and after revision. " 484 Decision Sciences [Vol. 18 Systematic Variance A major recognized problem of auditing- and accounting-based ESs is the small number of test problems used in the sample. In addition, the studies looked at ex­ amined test problems from a "middle of the road" viewpoint. While many real­ world situations are of this type, such a test does not examine system behavior in extreme situations and may not provide enough variation to test the system's overall behavior. This potential limitation can be overcome by broadening the range of test problem parameters. Further, validation of the prototypes studied did not take into account the frequency of 1YPe I and 1YPe II errors in developing the test problems, nor did it consider the effect the number of the test problems used would have. Both omis­ sions could produce a false sense of security in the test results for a system. Extraneous Variance The researchers in the prototype studies did not account for all extraneous system variables (e.g., complexity) in their validation efforts. Little research elicits or characterizes extraneous variables. Information on the special flxes and the valida­ tion of those special fIXes also was lacking. This may be because a relatively small number of fixes were required or because the fixes were application specific. However, research into the extent of these special fixes could provide a guid'e to the amount of effort that should be expended in validating these special fIXes. The effect of learning on the part of the judges apparently was not a concern in validating these prototype ESs. The location of the judges did vary from study to study. The extent to which learning and location affect a judge's decisions would make an interesting research question. CONCWSION This paper has developed a framework for validating ESs. The framework is based on research methods and can be used by ES developers to guide validation efforts. It also can be used to elicit further research topics in validation. The framework addresses validation in terms of validity, objectivity, cost­ beneflts, and accuracy. Using previously developed ESs as examples, the framework elicited some directions for future research in validation. Those directions include the development of tools to aid in cost-benefit validation, analysis of intervening variables and their characteristics, and comparison of ESs to other models. In ad­ dition, the framework suggests some important validation questions including, "What validation criteria can be used if the ES is expected to outperform a human expert?" and "How can we best identify the system's level of expertise?" [Received: April I, 1986. Accepted: January 28, 1987.] REFERENCts [I] Barr, A., & Feigenbaum, E. A. The handbook oj artifiCial intelligence. Stanford, CA: Heuristech Press and Los Altos, CA: William Kaufmann, 1981. I 1987] O'Leary 485 [2] Bennett, J. (Ed.). Building decision support systems. Reading, MA: Addison-Wesley, 1983. [3] Biggs, S. F., & Selfridge, M. GC-X: A prototype expert system for the auditor's going concern judgment. Presented at the University of Southern California Symposium on Expert Systems, Los Angeles, CA, February 1986. [4] Blanning, R. W. Knowledge acquisition and system validation in expert systems for management. Human Systems Management, 1984, 4,280-285. [5] Bouwman, M. J. Human diagnostic reasoning by computer. Management Science, 1983,29,653-672. (6) Bowman, E. H. Consistency and optimality in management decision making. Management Science, 1963, 9, 310-321. [7) Buchanan, B. G., & Shortliffe, E. H. Rule-based expert systems. Reading, MA: Addison-Wesley, 1984. [8] Chandrasekaran, B. On evaluating AI systems for medical diagnosis. AIMagazine, 1983,4(2),34-38. [9) Clarkson, G. P. E. Portfolio selection: A simulation of trust investment. Englewood Cliffs, NJ: Prentice-Hall, 1962. [10] Clarkson, G. P. E. A model of the trust investment process. In E. A. Feigenbaum & J. Feldman (Eds.), Computers and thought. New York: McGra~-Hill, 1963. [II] Clocksin, w. F., & Mellish, C. S. Programming in Prolog. New York: Springer-Verlag, 1984. [12] Davis, R. Reasoning from first principles in electronic troubleshooting. International Journal of Man-Machine Studies, 1983, 19, 403-423. [13] Davis, R. Diagnostic reasoning based on structure and behavior. Artificial Intelligence, 1984,24, 347-410. [14] Dillard, J. F., & Mutchler, J. F. Knowledge based expert systems for audit opinion decisions. Pre­ sented at the University of Southern California Symposium on Expert Systems, Los Angeles, CA, February 1986. [l5) Duda, R. 0., & Gaschnig, J. G. Knowledge-based systems come of age. Byte, September 1981, [16) pp. 238-281. Dungan, C. A model of audit judgment- in the form ofan expert system. UnpUblished Ph.D. dis­ p' sertation, University of Illinois, 1983. [17] Dungan, C., & Chandler, J. S. Analysis of audit judgment through an expert system (Faculty working paper no. 982). University of Illinois, College of Commerce and Business Administra­ tion, 1983. [18] Dungan, C., & Chandler, J. S. Auditor: a microcomputer-based expert system to support auditors in the field. Expert Systems, 1985, 2(4), 210-221. [19] Fox, M. S. On inheritance in knowledge representation. In Proceedings ofthe Sixth International Joint Conference on Artificial Intelligence (Vol. I). Menlo Park, CA: American Association for Artificial Intelligence, 1979. [20] Fox, M. Artificial intelligence in manufacturing. Presented at the CPMS Seminar on Expert Systems, Pittsburgh, PA, December 1984. [21] Gaschnig, J., Klahr, P., Pople, H., Shortliffe, E., & lefry, A. Evaluation of expert systems. In F. Hayes-Roth, D. A. Waterman, & D. B. Lenat (Eds.), Building expert systems. Reading, MA: Addison-Wesley, 1983. [22) Grudnitski, G. A prototype of an internal control expert system for the sales/accounts receivable application. Presented at the University of Southern California Symposium on Expert Systems, Los Angeles, CA, February 1986. [23) Hansen, J. V., & Messier, W. F. A knowledge-based expert system for auditing advanced compu­ ter systems (ARC working paper 83-5). University of Florida, Department of Accounting, 1985. [24] Hansen, J. V., & Messier, W. F. A preliminary test of EDP-XPERT. Presented at the University of Southern California Symposium on Expert Systems, Los Angeles, CA, February 1986. (25) Hayes-Roth, F., Waterman, D. A., & Lenat, D. B. (Eds.). Building expert systems. Reading, MA: Addison-Wesley, 1983. [26] Hogarth, R. Judgement and choice. New Yor,k: Wiley, 1980. [27] Hurst, E., Ness, D., Gambina, T., & Johnson, T. Growing DSS: A flexible, evolutionary approach. In J. Bennett (Ed.), Building decision support systems. Reading, MA: Addison-Wesley, 1983. [281 Keen, P., & Scott Morton, M. S. Decision support systems. Reading, MA: Addison-Wesley, 1978. 486 Decision Sciences [Vol. 18 [29] Kelly, K. P. Expert problem solving for the audit planning process. Unpublished Ph.D. disserta­ tion, University of Pittsburgh, 1984. [30] Kerlinger, F. Foundations of behavioral research. New York: Holt, Rinehart & Winston, 1973. [31] Kidd, A., & Cooper, M. Man-machine interface issues in the construction and use of an expert system. International Journal of Man-Machine Studies, 1985, 22, 91-102. [32] Libby, R. Accounting and human iriformation processing: Theory and applications. Englewood Cliffs, NJ: Prentice-Hall, 1981. [33] Liebowitz, 1. Useful approach for evaluating expert systems. Expert Systems, 1986,3(2), 86-99. [34] McCarty, L. T. Reflections on TAXMAN: An experiment in artificial intelligence and legal reason­ ing. Harvard Law Review, 1977,90,827-893. [35] McDermott, J. Background. theory and implementation of expert systems, II. Presented at the CPMS Seminar on Expert Systems, Pittsburgh, PA, December 1984. [36] McDermott, J. Rl revisted: Four years in the trenches. AI Magazine, 1984, 5(3), 21-35. [37] Meservy, R. D. Auditing internal controls: A computational model of the review process. Unpub­ lished Ph.D. dissertation, University of Minnesota, 1985. [38] Michaelsen, R. H. A knowledge-based system for individual income and transfer tax planning. Unpublished Ph.D. dissertation, University of Illinois, 1982. [39] Michaelsen, R. H. An expert system for federal tax planning. Expert Systems, 1984,1(2), 149-167. [40] Miller, R. K. The inventory of expert systems. Madison, GA: SEAl Institute, 1984. [41] O'Leary, D. Validation techniques for expert systems. Presented at the ORSA/TIMS Meeting, Miami, FL, October 1986. [42] Peat, Marwick. Mitchell & Co. Research opportunities in auditing. New York: Peat, Marwick, Mitchell Foundation, 1985. [43] Rich, E. Artificial intelligence. New York: McGraw-Hili, 1983. [44] Shooman, M. L. Software engineering. New York: McGraw-Hili, 1983. [45] Shpilberg, D., & Graham, L. E. Developing Expefl'AP: An expert system for corporate tax ac­ crual and planning. Presented at the University of Southern California Symposium on Expert Systems, Los Angeles, CA, February 1986. [46] Simon, 1. Basic research methods in social science. New York.: Random House, 1978. [47] Steinbart, P. J. The construction of an expert system to make materiality judgments. Unpub­ lished Ph.D. dissertation, Michigan State University, 1984. [48] Thxas Instruments. The Second Artificial Intelligence Satellite Symposium. Dallas, TX: 1&s instru­ ments, 1986. [49] Thrban, E., & Watkins, P. Integrating expert systems and decision support systems. MIS Quarterly, 1986, 10(2), 121-136. [50] Willingham, J., & Wright, W. Development qfa knowledge-based system for auditing the collect­ ability of a commercial loan. Presented at the ORSA/TIMS Meeting, Boston, MA, 1985. [51] Winston, P., & Horn, B. LISP. Reading, MA: Addison-Wesley, 1981. [52] Yu, V., Fagan, L., Wraith, s., Clancey, W., Scott, A., Hannigan, 1., Blum, R., Buchanan, B., & Cohen, S. Antimicrobial selection by computer: A blinded evaluation by infectious disease experts. Journal of the American Medical Association, 1979, 242, 1279-1282. (See also Buchanan and Shortliffe [7, chapter 31].) Daniel E. O'Leary is Assistant Professor in the Graduate School of Business, University of Southern California. Dr. O'Leary received his B.S. from Bowling Green State University, his M.B.A. from the University of Michigan, and his Ph.D. from Case Western Reserve University. His research interests include the theory of and methods for validating and assessing expert systems, and natural language interfaces. Dr. O'Leary has published or had accepted for publication papers on artificial intelligence and expert systems in various journals and book.s. He also has presented papers on artificial intelligence and expert systems at national and international symposia. Dr. O'Leary is the chair of the Measure­ ment in Management Section of TIMS and the editor of Exftert Systems in Business and Accounting, a newsletter published by the University of Southern California. " 
work_vy66sd74xbe2lbu3fpbuqxixai	----	TQM measurement model for the biotechnology industry in Taiwan Expert Systems with Applications 36 (2009) 8789–8798 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a TQM measurement model for the biotechnology industry in Taiwan Jui-Kuei Chen a,1, I-Shuo Chen b,* a Tamkang University, 4F, No. 20, Lane 22, WenZhou Street, Taipei City 10616, Taiwan b National Chiao Tung University, 4F, No. 20, Lane 22, WenZhou Street, Taipei City 10616, Taiwan a r t i c l e i n f o a b s t r a c t Keywords: Total quality management (TQM) TQM measurement model Biotechnology industry Fuzzy analytic network process (FANP) 0957-4174/$ - see front matter � 2008 Elsevier Ltd. A doi:10.1016/j.eswa.2008.11.013 * Corresponding author. Tel.: +886 911393602. E-mail addresses: chen3362@ms15.hinet.net (J.- com.tw (I.-S. Chen). 1 Tel.: +886 912272961. Due to the importance of the growth of biotechnology in Taiwan, the question of how the operation per- formance of the industry can be upgraded to sustain its competitive advantages has become an important question for Taiwanese biotechnology insiders. Recently, a growing number of organizations have imple- mented TQM in order to give them a competitive advantage [Chan, T. H., & Quazi, H. A. (2002). Overview of quality management practices in selected Asian countries. Quality Management Journal, 9(1), 172–180; Nilsson, L., Johnson, M. D., & Gustafsson, A. (2001). The impact of quality practices on customer satisfac- tion and business results: Product versus service organizations. Journal of Quality Management, 6, 5–27]. Meanwhile, quality related awards are also appearing for examining and identifying whether or not the overall quality of a firm is high. Each award, however, focuses only on examining certain items. It is impossible for a firm to make each examined item perfect. In addition to this, since the overall quality of the technology and products of some Taiwanese biotechnology corporations has been decreasing, the prestige of the Taiwanese biotechnology industry as a whole has decreased in the global market. Thus, in order to solve the above difficulties, this study attempts to find the most suitable measurement model for the biotechnology industry to enact quality improvement. The study first reviews a substantial body of literature on total quality management and categorizes measurement criteria. The study then proceeds with in-depth interviews with relevant background experts in order to extract and verify the most suitable measurement criteria. In addition to this, a FANP is utilized in order to analyze the weights of different measurement dimensions and criteria. The value of this study is its construction of a TQM measurement model for the Taiwanese biotechnology industry that can be used to improve its quality as well as to regain a higher market share of the global market. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction Industries have come to understand that, in order to stay com- petitive globally, a continuous improvement in organizational qual- ity performance is a necessity (Mele & Colucio, 2006; Sitalakshmi, 2007). Thus, a body of organizations started to implement total quality management (TQM) in order to generate a competitive advantage (Chan & Quazi, 2002; Nilsson , Johnson, & Gustafsson, 2001). In Taiwan, the importance and growth of the biotechnology industry has, in recent years, increased dramatically; it is now Tai- wan’s second most profitable industry, the high-tech industry being the highest. Moreover, with the large amount of capital investment by the Taiwanese government, biotechnology related corporations are flourishing; nevertheless, due to the poor quality of products and technology from some of them, both the prestige and market share of Taiwanese biotechnology corporations are decreasing in ll rights reserved. K. Chen), ch655244@yahoo. the global market. In this regard, a precise and effective way to mea- sure and operate quality improvement is necessary today. Measuring overall quality is complex. A great number of quality awards exist, such as the European Quality Award, the Malcolm Baldrige National Quality Award, the Asia-Pacific Business Excel- lence Standard, the Vietnam Quality Award, QS 9000, and IS 9000 (Dinh, Barbara, & Tritos, 2006), all of which are trying to play a role in standardizing the overall quality of an organization. Each award, however, focuses on certain examined items. It is impossible for a firm to make each item perfect. Although the national quality award (NQA) is widely used in different industries in Taiwan, there is little evidence that it can improve a specific industry’s overall quality, due to each industry’s different features. In light of this, a literature summarizing method is adapted in order to integrate related measurement criteria. A fuzzy analytic network process (FANP) was used to overcome the problem of dependence and feedback among dimensions or criteria. This is a general form of the FAHP that relieves hierarchical structural restrictions (Liou, Tzeng, & Chang, 2007). In this study, a literature summarizing method that is based on the NQA is utilized by com- bining it with FANP to create a new TQM measurement model. mailto:chen3362@ms15.hinet.net mailto:ch655244@yahoo.com.tw mailto:ch655244@yahoo.com.tw http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 8790 J.-K. Chen, I.-S. Chen / Expert Systems with Applications 36 (2009) 8789–8798 2. Total quality management The definition of quality, in the past, was the degree of confor- mance to a standard (Sitalakshmi, 2007). Sallis (1993) defined quality as that which best satisfies and exceeds customer needs and wants. Typically, quality has a variety of meanings (Sita- lakshmi, 2007) and its range of meanings causes confusion as each individual’s perception of quality differs (Shields, 1999). Since quality is proven to contribute to greater market share and return on investment (Cole, 1992; Philips, Chang, & Buzzell, 1983), lower manufacturing costs, improved productivity (Garvin, 1983) and improved strategic performance (Zhang, 2000), there are more and more organizations emphasizing the importance of increasing the quality of their service and products. Within the field of quality improvement, total quality manage- ment (TQM) is the most referred as well as most used criterion for enhancing organizational quality (Chan & Quazi, 2002; Nilsson et al., 2001). Recent studies have defined TQM as an holistic man- agement philosophy that strives for continuous organizational improvement (Kaynak, 2003). Furthermore, TQM can be seen as a management style based on producing quality service as defined by the customer, or based on achieving an organizational strategic imperative through continuous process improvement (Tseng, Lin, Chiu, & Liao, 2007). Additionally, TQM is also an integrated man- agement philosophy aimed at continuously improving the perfor- mance of products, processes, and services in order to achieve and surpass customer expectations (Ozden & Birsen, 2006). To summarize the above literature, TQM can be defined as a manage- rial method both for improving an organization’s core competitive ability and for gaining the maximization of market share within the industry in which it belongs. There is a stream of recent literature showing that if a firm implements TQM, it will gain many advantages, such as helping companies to improve their performance (Chase, Jacobs, & Aqui- lano, 2006; Han, Chen, & Ebrahimpour, 2007; Knod & Schonberger, 2001; Wadsworth, Stephens, & Godfrey, 2002), reducing rework, a Table 1 Measurement criteria based on TQM. Authors TQM factors Brah, Wong, and Rao (2000) Top management support, customer focus, em management, process improvement, service organization, customer satisfaction Antony et al. (2002) Management commitment, role of the qualit improvement, supplier partnership, product/ improve quality, and customer satisfaction o Sila and Ebrahimpour (2002) Top management commitment, employee in focus, process management, information and especially process management Shieh and Wu (2002) Leadership, human resource management, p Sureshchandar, Rajendran, and Anantharaman, (2002) Top management commitment and visionary analysis systems, benchmarking, continuous responsibility, services capes, and service cu Besterfield (2003) Quality culture, the quality chain, quality ass management Prajogo and Sohal (2003) Product innovation impacts the performance Jacqueline, Coyle, and Paula (2003) Statistical process control, the commitment Wanger and Schaltegger (2004) Leadership Escrig-Tena (2004) and Kenneth and Cynthia (2004) Financial performance Ozden and Birsen (2006) Customer focus, continuous improvement, an Nusrah, Ramayah, and Norizan (2006) Employee empowerment, information and co Ismail (2006) Leadership, strategic planning, customer focu supplier management, human resource resul Dinh et al., (2006) Strategic planning Keng, Nooh, Veeri, Lorraine, and Loke (2007) Teamwork, reward and recognition, custome management commitment at all levels, emp Han et al. (2007) Supplier relationship, customer involvement reduction in the costs related to poor quality (e.g., scrap, rework, late deliveries, warranty, replacement, etc.) (Antony, Leung, Know- les, & Gosh, 2002), and generating more unique competitive advan- tages (Reed, Lemak, & Mero, 2000). Hence, many quality related awards have arisen, such as the European Quality Award, the Mal- colm Baldrige National Quality Award, the Asia-Pacific Business Excellent Standard, the Vietnam Quality Award, QS 9000, and IS 9000 (Dinh et al., 2006; Uzumeri, 1997). The purpose of each of these awards is to examine the performance of each firm’s operat- ing TQM. Each award, however, has its own examination items. It is difficult for a firm to focus on every item while trying to improve. Consequently, it is important for firms to be able to discover which criteria are most critical for them to engage. 3. Measurement criteria of TQM The criteria for measuring TQM are various from one author to another (Ozden & Birsen, 2006). One of the early research works defining what constitutes TQM practice was conducted by Saraph et al. in 1989 (Joo & Yong, 2006). Since then, numerous related studies have been conducted by authors including Flynn, Schroe- der, and Sakakibara (1994), Black and Porter (1996), Choi and Eboch (1998), Samson and Terziovski (1999) and Kaynak (2003). This study summarizes the different criteria proposed by related researchers in Table 1. Based on the above literature, TQM criteria can be categorized into the following five dimensions: leaders, employees, customers, IT, and operating process. In Taiwan, the National Quality Award (NQA) is the most frequently utilized in some industries. It con- tains seven measurement dimensions: leadership and operation ideals, strategy management, the development of customers and a market, human resources and knowledge management, the applications and management of information strategy, process management, and operation performance. Because the measure- ment dimensions of NQA are similar to the above dimensions, the research structure in this study is based on NQA, with fixed ployee involvement, employee training, employee empowerment, supplier quality design, quality improvement rewards, benchmarking, and cleanliness and y department, training and education, employee involvement, continuous service design, quality policies, quality data and reporting, communication to rientation volvement, employee empowerment, education and training, teamwork, customer analysis systems, strategic planning, open organization, a service culture, and rocess management, supply chain management and information management leadership, human resource management, technical systems, information and improvement, customer focus, employee satisfaction, union intervention, social lture urance, commitment to continuous improvement, and the support of top of total quality management of top management, empowerment, and appropriate culture d teamwork mmunication, customer focus, and continuous improvement s, information and analysis, human resource management, process management, ts, customer results, and organizational effectiveness r focus, organizational trust, extensive training, high level of communication, loyee involvement, empowerment and organizational culture , training, top management commitment, and product design Table 2 Current status of biotechnology development in Taiwan (2005–2006). Industry Biotechnology (narrow) Pharmaceutical Medial devices Total Year 2005 2006 2005 2006 2005 2006 2005 2006 Revenue 386 434 624 660 590 697 1,600 1,791 Number of companies 253 268 419 368 484 500 1,156 1,136 Size of work force (number) 8090 8570 14,995 12,224 15,000 16,350 38,085 37,144 Export valuea 153 176 115 137 270 293 538 606 Import valuea 161 187 577 698 395 447 1133 1332 Domestic sales vs. export 60:40 60:40 82:18 79:21 54:46 58:42 66:34 66:34 Domestic market demanda 394 445 1086 1221 715 851 2195 2517 Source: Development Center for Biotechnology, Taiwan Institute of Economic Research, Biotechnology and Pharmaceutical Industries Program Office, MOEA, 2007. a Units: NT$100 million. Equally Moderatel Strongly Very Strong Extremel J.-K. Chen, I.-S. Chen / Expert Systems with Applications 36 (2009) 8789–8798 8791 criteria chosen using in-depth interviews and referring to the above summarized literature. Furthermore, in accordance with re- lated expert opinions, the sample industry of this study focuses mainly on research and development; therefore, the study finally concludes a measurement dimension, R&D and innovation, to be the base for developing measurement criteria. 4. The biotechnology industry in Taiwan 4.1. Definition and the technology range of the biotechnology industry According to the Taiwanese Biotechnology and Pharmaceutical Industries Program Office’s report (2007), Biotechnology is an inte- grated science of a multi-research field. It can be seen as a basic tool in researching life science, medical science, and agronomy. Biotechnology consists of the science of using biological processes and technology to both solve problems and produce useful prod- ucts, where examples include utilizing the characteristics or ingre- dients of microorganisms, vegetation, or animals to produce products; applying the design of products that enter molecules in order to understand life phenomena; and developing the techno- logical platform to solve the above problems. The overall goal of these efforts is to improve the quality of life of human beings. Therefore, biotechnology includes traditional biotechnology (e.g., microorganism fermentation), modern biotechnology (e.g., bioen- gineering), and new biotechnology (e.g., the integration of hybrid high technology and life science). Moreover, in Taiwan, in order to promote the domestic biotechnology industry, biotechnology related medical industries such as the pharmaceutical industry, the medical device industry, the agriculture and food industry, the biologic resource industry, and the service industry that con- tains the medical and R&D service industries are included in the field of biotechnology industry. Therefore, in Taiwan, the technol- ogy range of biotechnology industry can be formally categorized into three dimensions: first, the biotechnology industry (i.e., un- ique biochemistry, biomedical, agriculture, and environmental manufacturing and service), also called the narrow biotechnology industry, the pharmaceutical industry, and the medical devices industry. ( )A xµ L M 1 0 U ~ Fig. 1. A triangular fuzzy number. 4.2. Recent status on biotechnology industry In accordance with the latest survey of the Development Center for Biotechnology (2007), the biotechnology industry in Taiwan formally includes biotechnology, pharmaceuticals and medical de- vices. In 2006, the total annual revenue for these industries in Tai- wan was nearly NT$179.1 billion, of which NT$43.4 billion came from biotechnology, with 434 companies. The scope of businesses covered included genomics, drugs, diagnostics, agricultural bio- technology, environmental biotechnology, protein drugs, contract research organizations, biochips and bioinformatics. The island’s pharmaceutical industry returned NT$66.0 billion with 368 companies active in this sector. The medical devices industry returned NT$69.7 billion coming from 397 companies. The biotechnology workforce size is 37,144, of which 8570 are in the biotechnology industry, 12,224 in the pharmaceutical industry and 16,350 in the medical devices industry. To present the growth of the biotechnology industry in Taiwan, the current status is pro- vided in Table 2. Due to both the importance of the growth of the biotechnology and the large capital investment from the Taiwanese government, biotechnology related corporations are growing quickly; nonethe- less, since some of these corporations produce products of inferior quality, both the prestige and global market share of Taiwanese companies are shrinking. In this regard, knowing the most precise and effective way to measure and operate quality improvement is extremely important. This study, by utilizing both the literature and the opinions of experts, aims to arrive at an effective way for these corporations to conduct quality improvement. 1 3 5 7 9 Fig. 2. Fuzzy memberships function for linguistic values for attributes. Table 3 Definition and membership function of fuzzy numbers. Fuzzy number Linguistic variable Triangular fuzzy number ~9 Extremely important/preferred (7, 9, 9) ~7 Very strongly important/preferred (5, 7, 9) ~5 Strongly important/preferred (3, 5, 7) ~3 Moderately important/preferred (1, 3, 5) ~1 Equally important/preferred (1, 1, 3) 8792 J.-K. Chen, I.-S. Chen / Expert Systems with Applications 36 (2009) 8789–8798 5. Fuzzy analytic network process 5.1. Analytic hierarchy process Analytic hierarchy process (AHP) was developed by Thomas L. Saaty in 1971. The AHP method is known as an eigenvector meth- od. It indicates that the eigenvector corresponding to the largest eigenvalue of the pairwise comparisons matrix gives the relative priorities of the factors, and it preserves ordinal preferences among the alternatives. This means that if an alternative is preferred to another, its eigenvector component is larger than that of the other. A vector of weights obtained from the pairwise comparisons ma- trix reflects the relative performance of the various factors. A growing literature now argues, however, that AHP has its drawbacks. Studies have concluded that AHP can be applied to spe- cific, but not fuzzy decision making, that AHP evaluates questions using different criteria for different parts of the test set, that AHP cannot include uncertainty factors of people toward objects, and that the priority of AHP is unspecific. Furthermore, AHP is based on the concept of independence among each factor; however, this does not fit real environments. Thus, the study employed a modi- fied form of AHP, called fuzzy ANP (FANP), in order to arrive at more precise results. 5.2. Fuzzy set theory Fuzzy set theory was first developed in 1965 when Professor L.A. Zadeh was trying to solve fuzzy phenomenon problems that exist in the real world, such as uncertain, incomplete, nonspecific, and fuzzy situations. Fuzzy set theory has an advantage over tradi- tional set theory in describing set concepts in human language. It demonstrates specific and fuzzy characteristics in language on the evaluation, and it uses a membership function concept to rep- resent the field in which a fuzzy set can permit a situation such as ‘‘incompletely belongs to” and ‘‘incompletely does not belong to”. 5.3. Fuzzy number We order the Universe of Discourse such that U is a whole tar- get, and each target in the Universe of Discourse is called an ele- ment. Fuzzy ~A states for U that random x ? U, appointing a real number l~AðxÞ! ½0; 1�. We call anything above that level of x under A. The set of real numbers R is a triangular fuzzy number (TFN): ~A, which means that x 2 R, appointing l~AðxÞ 2 ½0; 1�, and l~AðxÞ¼ ðx � LÞ=ðM � LÞ; L � x � M; ðU � xÞ=ðU � MÞ; M � x � U; 0; otherwise; 8>< >: The triangular fuzzy number above can be written as ~A ¼ðL; M; UÞ, where L and U represent the fuzzy probability be- tween the lower and upper boundaries of evaluation information, as shown in Fig. 1. Assume two fuzzy numbers ~A1 ¼ðL1; M1; U1Þ and ~A2 ¼ðL2; M2; U2Þ : (1) ~A1 � ~A2 ¼ðL1;M1;U1Þ�ðL2;M2;U2Þ¼ðL1 þL2;M1 þM2;U1 þU2Þ. (2) ~A1 � ~A2 ¼ ðL1; M1; U1Þ�ðL2; M2; U2Þ¼ ðL1 L2; M1M2; U1U2Þ; Li > 0; Mi > 0; Ui > 0. (3) ~A1 � ~A2 ¼ðL1;M1;U1Þ�ðL2;M2;U2Þ¼ðL1 �L2;M1 �M2;U1 �U2Þ. (4) ~A1 � ~A2 ¼ ðL1; M1; U1Þ�ðL2; M2; U2Þ¼ ðL1=U2; M1=M2; U1=L2Þ: Li > 0; Mi > 0; Ui > 0. ~�1 �1 Fig. 3. Case 1 structure. A1 ¼ðL1; M1; U1Þ ¼ ð1=U1; 1=M1; 1=L1Þ; Li > 0; Mi > 0; Ui > 0: 5.4. The fuzzy linguistic variable The fuzzy linguistic variable is a variable that reflects the different levels of human language. Its value represents the range from natural to artificial language. When precisely reflecting the value or meaning of a linguistic variable, there must be an appropriate way for the var- iable to change. Variables representing a human word or sentence can be divided into numerous linguistic criteria, such as equally impor- tant, moderately important, strongly important, very strongly impor- tant, and extremely important, as shown in Fig. 2; the definitions and descriptions of these terms are given in Table 3. For the purpose of the present study, the five criteria above (i.e., equally important, moder- ately important, strongly important, very strongly important and ex- tremely important) are used. 5.5. Analytic network process (ANP) The purpose of the ANP approach is to solve the problem of interdependence and feedback between criteria and alternatives. The ANP is the general form of the analytic hierarchy process (AHP), which has been used in multicriteria decision making (MCDM) to release the restriction of hierarchical structure. Fur- thermore, it has been applied to project selection, product plan- ning, strategic decision making, optimal scheduling, etc. (Huang, Tzeng, & Ong, 2005). The first phase of the ANP is comparing the measurement criteria of the overall system to form the super matrix. This step can be accom- plished through pairwise comparisons. The relative importance value of pairwise comparisons can be assigned on a scale of 1–9, represent- ing equal importance to extreme importance (Saaty, 1980). The fol- lowing shows the general form of the super matrix: where Cm denotes the mth cluster, emn denotes the nth element in the mth cluster, and Wij is the principal eigenvector influencing the elements compared in the jth cluster to the nth cluster. In addi- tion, if the jth cluster has no influence on the nth cluster, then Wij = 0. Fig. 4. Case 2 structure. J.-K. Chen, I.-S. Chen / Expert Systems with Based on the above, the form of the super matrix relies on the variety of the structure. Several structures were proposed by Saaty, including hierarchy, holarchy, suparchy, intarchy, etc. (Huang et al., 2005). In order to demonstrate how the structure is affected by the super matrix, Huang et al. (2005) offer two simple cases that both have three clusters to show how to form the super matrix in accordance with the structures (Fig. 3). Table 4 Research structure of this study. Goal Evaluation dimensions Biotechnology industry KSF exploration based on TQM Leadership and operation ideals (D Strategy management (D2) R&D and innovation (D3) The development of customers an Human resource and knowledge m The application and management (D6) Process management (D7) Operation performance (D8) Table 5 Pairwise comparison matrix and weights of measurement dimension. Measurement dimension Local weight Inner dependence weight D1 0.063 0.079 0.042 0 D2 0.036 0.064 0.080 0 D3 0.243 0.228 0.238 0 D4 0.183 0.213 0.192 0.271 D5 0.119 0.063 0.076 0.099 0 D6 0.156 0.107 0.090 0.090 0 D7 0.053 0.037 0.031 0.041 0 D8 0.147 0.288 0.294 0.377 0 Based on Fig. 3, we can form the super matrix as follows: . In Fig. 4, the second case is more complex than the first one. Applications 36 (2009) 8789–8798 8793 From Fig. 4, we can form the super matrix as follows: When the super matrix has been formed, the weighted super matrix can then be derived by transforming all columns so that they sum exactly to unity (Huang et al., 2005). This step is similar Evaluation criteria 1) Operational values and ideals restructure (C1) Organizational quality improving mission (C2) Top managers’ leading style (C3) TQM culture construction (C4) The increasing of social contribution (C5) Organization of operation strategy planning (C6) Operation structure adjustment (C7) The quality improvement of strategy (C8) Process redefinition of R&D and innovation (C9) Input of R&D and innovation (C10) Evaluation of R&D and innovation results (C11) d a market (D4) Market operation strategy development (C12) Business relation management (C13) Customer relationship management (C14) anagement (D5) Human resource planning (C15) Human resource development (C16) Human resources utilization (C17) Employee relationship management (C18) Knowledge management (C19) of information strategy Biotechnology information receiving channel (C20) Internet applications (C21) Biotechnology information utilization (C22) Manufacturing process management (C23) Supportive activity planning (C24) Cross-unit (Department) Management (C25) Customer satisfaction (C26) Market enlargement performance (C27) Financial performance (C28) Human resource development performance (C29) Information management performance (C30) Process management performance (C31) Unique competitive ability gaining performance (C32) Prestige measurement (C33) Interdependence weight .046 0.045 0.050 0.036 0.065 0.0548 .089 0.053 0.062 0.043 0.089 0.0535 .312 0.315 0.290 0.306 0.363 0.2378 0.099 0.140 0.150 0.234 0.1725 .057 0.081 0.060 0.075 0.0936 .083 0.135 0.063 0.139 0.1214 .041 0.055 0.025 0.036 0.0446 .371 0.299 0.352 0.342 0.2219 Table 6 Pairwise comparison matrix and weights of D1 criteria. Measurement criteria 1 C1 C2 C3 C4 C5 Local weight Global weight C1 1.000 1.000 3.000 0.554 0.629 0.905 0.212 0.252 0.378 0.201 0.231 0.315 4.395 6.795 7.825 0.126 0.0069 C2 1.104 1.591 1.806 1.000 1.000 3.000 0.135 0.160 0.273 0.353 0.481 0.713 4.395 5.658 7.187 0.155 0.0085 C3 2.646 3.974 4.711 3.659 6.240 7.399 1.000 1.000 3.000 1.403 2.080 2.836 1.685 2.265 3.000 0.408 0.0224 C4 3.177 4.327 4.983 1.403 2.080 2.836 0.353 0.481 0.713 1.000 1.000 3.000 1.383 2.107 2.720 0.251 0.0138 C5 0.128 0.147 0.228 0.139 0.177 0.228 0.333 0.442 0.594 0.368 0.475 0.723 1.000 1.000 3.000 0.060 0.0033 Table 7 Pairwise comparison matrix and weights of D2 criteria. Measurement criteria 2 C6 C7 C8 Local weight Global weight C6 1.000 1.000 3.000 1.834 3.000 3.709 0.244 0.289 0.417 0.2789 0.0149 C7 0.270 0.333 0.545 1.000 1.000 3.000 0.353 0.481 0.713 0.1701 0.0091 C8 2.399 3.455 4.091 1.403 2.080 2.836 1.000 1.000 3.000 0.5510 0.0295 Table 8 Pairwise comparison matrix and weights of D3 criteria. Measurement criteria 3 C9 C10 C11 Local weight Global weight C9 1.000 1.000 3.000 1.326 1.910 2.362 0.255 0.306 0.454 0.2279 0.0542 C10 0.423 0.523 0.754 1.000 1.000 3.000 0.177 0.231 0.357 0.1406 0.0335 C11 2.203 3.267 3.923 2.798 4.327 5.658 1.000 1.000 3.000 0.6314 0.1502 Table 9 Pairwise comparison matrix and weights of D4 criteria. Measurement criteria 4 C12 C13 C14 Local weight Global weight C12 1.000 1.000 3.000 0.353 0.481 0.713 0.177 0.231 0.357 0.1348 0.0232 C13 1.403 2.080 2.836 1.000 1.000 3.000 0.231 0.333 0.467 0.2373 0.0409 C14 2.798 4.327 5.658 2.140 3.000 4.327 1.000 1.000 3.000 0.6280 0.1083 Table 10 Pairwise comparison matrix and weights of D5 criteria. Measurement criteria 5 C15 C16 C17 C18 C19 Local weight Global weight C15 1.000 1.000 3.000 2.140 3.000 4.327 0.121 0.135 0.189 0.201 0.231 0.315 0.167 0.212 0.298 0.0685 0.0064 C16 0.231 0.333 0.467 1.000 1.000 3.000 0.167 0.212 0.298 0.128 0.147 0.228 0.167 0.212 0.298 0.0444 0.0042 C17 5.278 7.399 8.277 3.361 4.711 5.984 1.000 1.000 3.000 0.400 0.481 0.628 0.160 0.201 0.273 0.1807 0.0169 C18 3.177 4.327 4.983 4.395 6.795 7.825 1.593 2.080 2.498 1.000 1.000 3.000 0.255 0.306 0.454 0.2520 0.0236 C19 3.361 4.711 5.984 3.361 4.711 5.984 3.659 4.983 6.240 2.203 3.267 3.923 1.000 1.000 3.000 0.4544 0.0425 Table 11 Pairwise comparison matrix and weights of D6 criteria. Measurement criteria 6 C20 C21 C22 Local weight Global weight C20 1.000 1.000 3.000 2.646 3.309 4.327 0.160 0.167 0.251 0.2197 0.0267 C21 0.231 0.302 0.378 1.000 1.000 3.000 0.231 0.278 0.429 0.1158 0.0141 C22 3.984 5.984 6.240 2.330 3.603 4.327 1.000 1.000 3.000 0.6645 0.0807 Table 12 Pairwise comparison matrix and weights of D7 criteria. Measurement criteria 7 C23 C24 C25 Local weight Global weight C23 1.000 1.000 3.000 4.395 5.658 7.187 0.192 0.219 0.289 0.2573 0.0115 C24 0.139 0.177 0.228 1.000 1.000 3.000 0.160 0.167 0.251 0.0765 0.0034 C25 3.460 4.576 5.196 3.984 5.984 6.240 1.000 1.000 3.000 0.6662 0.0297 8794 J.-K. Chen, I.-S. Chen / Expert Systems with Applications 36 (2009) 8789–8798 T ab le 1 3 P ai rw is e co m p ar is o n m at ri x an d w ei gh ts o f D 8 cr it er ia . M ea su re m en t cr it er ia 8 C 2 6 C 2 7 C 2 8 C 2 9 C 3 0 C 3 1 C 3 2 C 3 3 Lo ca l w ei g h t G lo b al w ei g h t C 2 6 1 .0 0 0 1 .0 0 0 3 .0 0 0 0 .2 3 1 0 .2 7 8 0 .4 2 9 0 .1 2 8 0 .1 4 7 0 .2 2 8 1 .2 7 2 1 .8 0 6 2 .1 6 9 2 .6 4 6 3 .9 7 4 4 .7 1 1 2 .5 3 7 3 .7 5 7 4 .3 2 7 0 .1 6 7 0 .2 1 2 0 .2 9 8 2 .1 4 0 3 .0 0 0 4 .3 2 7 0 .0 8 1 9 0 .0 1 8 2 C 2 7 2 .3 3 0 3 .6 0 3 4 .3 2 7 1 .0 0 0 1 .0 0 0 3 .0 0 0 0 .1 6 7 0 .1 7 7 0 .2 7 3 3 .6 5 9 6 .2 4 0 7 .3 9 9 2 .6 4 6 4 .1 4 9 5 .6 5 8 4 .3 9 5 6 .7 9 5 7 .8 2 5 0 .3 0 2 0 .4 0 0 0 .5 4 5 3 .6 5 9 5 .1 9 6 6 .7 9 5 0 .1 6 4 6 0 .0 3 6 5 C 2 8 4 .3 9 5 6 .7 9 5 7 .8 2 5 3 .6 5 9 5 .6 5 8 5 .9 8 4 1 .0 0 0 1 .0 0 0 3 .0 0 0 3 .6 5 9 6 .2 4 0 7 .3 9 9 3 .4 6 0 5 .9 8 4 7 .3 9 9 2 .7 9 8 4 .3 2 7 5 .6 5 8 0 .3 0 2 0 .4 0 0 0 .5 4 5 4 .3 9 5 5 .6 5 8 7 .1 8 7 0 .2 6 5 0 0 .0 5 8 8 C 2 9 0 .4 6 1 0 .5 5 4 0 .7 8 6 0 .1 3 5 0 .1 6 0 0 .2 7 3 0 .1 3 5 0 .1 6 0 0 .2 7 3 1 .0 0 0 1 .0 0 0 3 .0 0 0 0 .3 3 3 0 .3 6 8 0 .5 4 5 2 .6 4 6 3 .3 0 9 4 .3 2 7 0 .1 2 3 0 .1 3 9 0 .2 0 9 2 .6 4 6 3 .9 7 4 4 .7 1 1 0 .0 5 0 6 0 .0 1 1 2 C 3 0 0 .2 1 2 0 .2 5 2 0 .3 7 8 0 .1 7 7 0 .2 4 1 0 .3 7 8 0 .1 3 5 0 .1 6 7 0 .2 8 9 1 .8 3 4 2 .7 2 0 3 .0 0 0 1 .0 0 0 1 .0 0 0 3 .0 0 0 1 .4 0 3 2 .0 8 0 2 .8 3 6 0 .1 2 8 0 .1 4 7 0 .2 2 8 3 .1 7 7 4 .3 2 7 4 .9 8 3 0 .0 5 8 0 0 .0 1 2 9 C 3 1 0 .2 3 1 0 .2 6 6 0 .3 9 4 0 .1 2 8 0 .1 4 7 0 .2 2 8 0 .1 7 7 0 .2 3 1 0 .3 5 7 0 .2 3 1 0 .3 0 2 0 .3 7 8 0 .3 5 3 0 .4 8 1 0 .7 1 3 1 .0 0 0 1 .0 0 0 3 .0 0 0 0 .1 2 3 0 .1 3 9 0 .2 0 9 3 .1 7 7 4 .3 2 7 4 .9 8 3 0 .0 3 6 3 0 .0 0 8 1 C 3 2 3 .3 6 1 4 .7 1 1 5 .9 8 4 1 .8 3 4 2 .4 9 8 3 .3 0 9 1 .8 3 4 2 .4 9 8 3 .3 0 9 4 .7 8 5 7 .1 8 7 8 .1 6 0 4 .3 9 5 6 .7 9 5 7 .8 2 5 4 .7 8 5 7 .1 8 7 8 .1 6 0 1 .0 0 0 1 .0 0 0 3 .0 0 0 4 .3 9 5 5 .6 5 8 7 .1 8 7 0 .3 2 0 2 0 .0 7 1 0 C 3 3 0 .2 3 1 0 .3 3 3 0 .4 6 7 0 .1 4 7 0 .1 9 2 0 .2 7 3 0 .1 3 9 0 .1 7 7 0 .2 2 8 0 .2 1 2 0 .2 5 2 0 .3 7 8 0 .2 0 1 0 .2 3 1 0 .3 1 5 0 .2 0 1 0 .2 3 1 0 .3 1 5 0 .1 3 9 0 .1 7 7 0 .2 2 8 1 .0 0 0 1 .0 0 0 3 .0 0 0 0 .0 2 3 4 0 .0 0 5 2 J.-K. Chen, I.-S. Chen / Expert Systems with Applications 36 (2009) 8789–8798 8795 to the Markov chain concept for ensuring that the sum of the prob- abilities of all states equals one. Then, we limit the power of the weighted super matrix by using Eq. (1) to get the global weights: lim k!1 Wk: ð1Þ In this step, if the super matrix is cyclic, then more than one limiting super matrix exists. More precisely, there are two or more limiting super matrices in this case, and the Cesaro sum would be calculated to determine which matrix has priority. The Cesaro sum is formulated as Eq. (2) lim k!1 1 N � �XN k¼1 Wk ð2Þ to calculate the average effect of the limiting super matrix, or the super matrix could be raised to large powers to get the priority weights. From the above description of the ANP approach, the most crit- ical advantage of this analysis is that, first, that it is appropriate for both quantitative and qualitative data types; and second, that it can solve the problem of interdependence and feedback between whole factors. Due to these two reasons, the ANP approach is adopted by this study. 5.6. Steps of fuzzy analytic network process calculation Summarizing the above literature, the steps of fuzzy ANP calcu- lation in this study are provided as follows: Step 1: Confirm both the dimensions and criteria of the model. Step 2: Hierarchically build up an ANP model containing goals, dimensions, and criteria. Step 3: Determine the local weights of both dimensions and cri- teria by utilizing pairwise comparison matrices (assuming that there is no interdependence among the criteria). The relative importance value of pairwise comparisons is provided in Table 2. Step 4: Determine the inner dependence matrix of each dimen- sion with respect to the other dimensions. In step 3, the depen- dence of the local weights of the dimensions of the inner matrix has already been found. This next step is to calculate the inter- dependent weights of the dimensions. Step 5: Calculate the global weights for the criteria. These can be acquired by multiplying the local weight of the criteria with the interdependent dimension weights above each. 6. Empirical study of TQM measurement model The TQM measurement model for the biotechnology industry is a hybrid composition; furthermore, it is obvious that the criteria explored must be context dependent and in accord with real situ- ations. Before sending out questionnaires, five senior biotechnol- ogy industry background experts were consulted, and recent research studies and national quality awards (NQA) were referred to in order to construct eight dimensions (Leadership and Opera- tion Ideals, Strategy Management, R&D and Innovation, The Devel- opment of Customers and a Market, Human Resource and Knowledge Management, The Application and Management of Information Strategy, Process Management, and Operation Perfor- mance) with each dimension having three to eight criteria. A ques- tionnaire was used to find out from seven groups comprising 31 experts – 10 from the narrow biotechnology industry, 11 from pharmaceutical industry, and 10 from medial devices industry – their ranking of each dimension and criterion with respect to the Table 14 The new TQM measurement model. Goal Measurement dimension Measurement criteria TQM achievement for Taiwanese biotechnology industry Market focus Customer relationship management sustention and reinforce Organization focus Unique competitive ability development Process focus Information utilization maximization Result focus Financial evaluation and improvement R&D and innovation productivity evaluation and developing Table 15 The fuzzy ANP results of the study. Goal Evaluation dimensions Global weight Ranking Evaluation criteria Local weight Global weight Ranking Biotechnology industry KSF exploration based on TQM Leadership and operation ideals (D1) 0.0548 7 Operational values and ideals restructure (C1) 0.126 0.0069 28 Organizational quality improving mission (C2) 0.155 0.0085 26 Top managers’ leading style (C3) 0.408 0.0224 16 TQM culture construction (C4) 0.251 0.0138 21 The increasing of social contribution (C5) 0.060 0.0033 33 Strategy management (D2) 0.0535 8 Organization of operation strategy planning (C6) 0.2789 0.0149 19 Operation structure adjustment (C7) 0.1701 0.0091 25 The quality improvement of strategy (C8) 0.5510 0.0295 12 R&D and innovation (D3) 0.2378 1 Process redefinition of R&D and innovation (C9) 0.2279 0.0542 6 Input of R&D and innovation (C10) 0.1406 0.0335 10 Evaluation of R&D and innovation results (C11) 0.6314 0.1502 1 The development of customers and a market (D4) 0.1725 3 Market operation strategy development (C12) 0.1348 0.0232 15 Business relation management (C13) 0.2373 0.0409 8 Customer relationship management (C14) 0.6280 0.1083 2 Human resource and knowledge management (D5) 0.0936 5 Human resource planning (C15) 0.0685 0.0064 29 Human resource development (C16) 0.0444 0.0042 31 Human resources utilization (C17) 0.1807 0.0169 18 Employee relationship management (C18) 0.2520 0.0236 14 Knowledge management (C19) 0.4544 0.0425 7 The application and management of information strategy (D6) 0.1214 4 Biotechnology information receiving channel (C20) 0.2197 0.0267 13 Internet applications (C21) 0.1158 0.0141 20 Biotechnology information utilization (C22) 0.6645 0.0807 3 Process management (D7) 0.0446 6 Manufacturing process management (C23) 0.2573 0.0115 23 Supportive activity planning (C24) 0.0765 0.0034 32 Cross-unit (Department) management (C25) 0.6662 0.0297 11 Operation performance (D8) 0.2219 2 Customer satisfaction (C26) 0.0819 0.0182 17 Market enlargement performance (C27) 0.1646 0.0365 9 Financial performance (C28) 0.2650 0.0588 5 Human resource development performance (C29) 0.0506 0.0112 24 Information management performance (C30) 0.0580 0.0129 22 Process management performance (C31) 0.0363 0.0081 27 Unique competitive ability gaining performance (C32) 0.3202 0.0710 4 Prestige measurement (C33) 0.0234 0.0052 30 8796 J.-K. Chen, I.-S. Chen / Expert Systems with Applications 36 (2009) 8789–8798 TQM measurement, utilizing a 5-point scale ranging from 9 (extre- mely important) to 1 (no effect) as Table 3 shows. The top five cri- teria from each dimension were chosen to construct the final model for measuring TQM of the biotechnology industry (as Table 4). After forming the TQM measurement model, local weights of the dimensions and criteria which were in the second and third levels of the TQM measurement model were calculated first. All of the fuzzy measurement matrices were developed in the same way. Additionally, pairwise comparison matrices and local weights J.-K. Chen, I.-S. Chen / Expert Systems with Applications 36 (2009) 8789–8798 8797 were analyzed. The local weights for the dimensions were calcu- lated in a similar way to the fuzzy measurement matrices (given in the second column of Table 5). Then, inner dependence weights of the dimensions were calculated next and the dependencies among the dimensions were considered. The resulting inner dependence weights are given in the third to tenth columns of Ta- ble 5. Next, using the computed inner dependence weights above, interdependent weights of the dimensions were computed by tim- ing the inner dependence weights matrix of the dimensions with the local weights of dimensions. The result is given in the last col- umn of Table 5. Using the interdependent weights of each dimension and local weights of criteria, global weights for the criteria were subse- quently calculated. Global criteria weights were computed by tim- ing the local weight of the criteria (given in the last second column of Tables 6–13), with the interdependent weights of the dimen- sions to which it belongs. All global weights are provided in the last column of Tables 6–13. Based on the global weights calculated above, the top five critical criteria that can significantly improve the overall quality performance of the biotechnology industry are ‘‘Evaluation of R&D and Innovation Results”, ‘‘Customer Relation- ship Management”, ‘‘Biotechnology Information Utilization”, ‘‘Un- ique Competitive Ability Gaining Performance”, and ‘‘Financial Performance”. After in-depth interviews with experts from three different dimensions of the biotechnology industry in Taiwan, the study found that the most effective measurement (helpful) criteria in improving quality performance can be re-categorized into four fo- cus dimensions: market focus (customer relationship management sustention and reinforce), organization focus (unique competitive ability development), process focus (information utilization), and result focus (R&d and innovation productivity evaluation and developing and financial evaluation and improvement). Hence, biotechnology corporations that are devoted to improving their quality more efficiently and precisely should utilize these four ‘‘focuses”. The study proposes the final TQM measurement model as given in Table 14 and summarizes the result with rankings as gi- ven in Table 15. 7. Conclusion Because of the increasing emphasis on the biotechnology indus- try in Taiwan, the government has invested a great amount of cap- ital in order to turn it into one of the most profitable Taiwanese industries of the future. There is, consequently, a growing body of studies on such corporations over the past few years. Recently, however, the poor quality of some of those corporations has re- sulted in dangerous products and decreased the prestige of these corporations around the world. Upgrading the quality of the bio- technology industry is necessary for Taiwan if it is to reclaim pres- tige and market share in the global market. In the past, the national quality award has been the most useful tool for tackling different kinds of industries. Due to different characteristics between each industry, however, it is necessary for Taiwan’s newest profitable industry to construct a suitable quality measuring model for assessments. Based on several quality measurement criteria pro- posed by numerous researchers, we utilized in-depth interviews and FANP to develop a TQM measurement model that considers interdependence between a range of dimensions and criteria and their weighting. The study, reflecting on the results, proposes a new TQM measurement model for the Taiwanese biotechnology industry. References Antony, J., Leung, K., Knowles, G., & Gosh, S. (2002). Critical success factors of TQM implementation in Hong Kong industries. International Journal of Quality and Reliability Management, 19(5), 551–566. Besterfield, D. (2003). Quality control (7th ed). NY: Prentice Hall. Biotechnology and Pharmaceutical Industries Program Office (2007). Development center for biotechnology, Taiwan institute of economic research, biotechnology and pharmaceutical industries program office. <http://www.biopharm.org.tw/ 2t2s/situation.html>, MOEA Retrieved 09.05.08. Black, S. A., & Porter, L. J. (1996). Identification of critical factors of TQM. Decision Sciences, 27(1), 1–21. Brah, S. A., Wong, J. L., & Rao, B. M. (2000). TQM and business performance in the service sector: A Singapore study. International Journal of Operations and Production Management, 20(11), 1293–1312. Chan, T. H., & Quazi, H. A. (2002). Overview of quality management practices in selected Asian countries. Quality Management Journal, 9(1), 172–180. Chase, R. B., Jacobs, F. R., & Aquilano, N. J. (2006). Operations management for competitive advantage (11th ed). NY: McGraw-Hill. Choi, T., & Eboch, K. (1998). The TQM paradox: Relations among TQM practices, plant performance, and customer satisfaction. Journal of Operations Management, 17(1), 59–75. Cole, R. E. (1992). The quality revolution. Production and Operations Management, 1(1), 118–220. Dinh, T. H., Barbara, I., & Tritos, L. (2006). The impact of total quality management on innovation findings from a developing country. International Journal of Quality and Reliability, 23(9), 1092–1117. Escrig-Tena, A. B. (2004). TQM as a competitive factor: A theoretical and empirical analysis. International Journal of Quality and Reliability Management, 21(6), 612–637. Flynn, B. B., Schroeder, R. G., & Sakakibara, S. (1994). A framework for quality management research and an associated measurement instrument. Journal of Operations Management, 11, 339–366. Garvin, D. A. (1983). Quality on the line. Harvard Business Review, 61, 64–75. Han, S. B., Chen, S. K., & Ebrahimpour, B. (2007). The impact of ISO 9000 on TQM and business performance. Journal of Business and Economic Studies, 13(2), 1–21. Ismail, S. (2006). Examining the effects of contextual factors on TQM and performance through the lens of organizational theories: An empirical study. Journal of Operations Management, 25, 83–109. Jacqueline, A. M., Coyle, S., & Paula, C. M. (2003). The role of individual differences in employee adoption of TQM orientation. Journal of Vocational Behavior, 62, 320–340. Joo, Y. J., & Yong, J. W. (2006). Relationship between total quality management (TQM) and continuous improvement of international project management (CIIPM). Technovation, 26, 716–722. Kaynak, H. (2003). The relationship between total quality management practices and their effects on firm performance. Journal of Operations Management, 21, 405–435. Keng, B. O., Nooh, A. B., Veeri, A., Lorraine, V., & Loke, A. K. Y. (2007). Does TQM influence employees’ job satisfaction: An empirical case analysis? International Journal of Quality and Reliability Management, 24(1), 62–77. Kenneth, M. Y., & Cynthia, E. M. (2004). Causation or covariation: An empirical re- examination of the link between TQM and financial performance. Journal of Operations Management, 22, 291–311. Knod, E. M., Jr., & Schonberger, R. J. (2001). Operations management: Meeting customer’s demands (7th ed). NY: McGraw-Hill. Liou, J. H., Tzeng, G. H., & Chang, H. C. (2007). Airline safety measurement using a novel hybrid model. Journal of Air Transport Management, 13(4), 243–249. Mele, C., & Colucio, M. (2006). The evolving path of TQM: Towards business excellence and stakeholder value. International Journal of Quality and Reliability Management, 23(5), 464–489. Nilsson, L., Johnson, M. D., & Gustafsson, A. (2001). The impact of quality practices on customer satisfaction and business results: Product versus service organizations. Journal of Quality Management, 6, 5–27. Nusrah, S., Ramayah, T., & Norizan, M. S. (2006). TQM practices, service quality, and market orientation: Some empirical evidence from a developing country. Management Research News, 29(11), 713–728. Huang, J. J., Tzeng, G. H., & Ong, C. S. (2005). Multidimensional data in multidimensional scaling using the analytic network process. Pattern Recognition Letters, 26(6), 755–767. Ozden, B., & Birsen, K. (2006). An analytical network process-based framework for successful total quality management (TQM): An assessment of Turkish manufacturing industry readiness. International Journal of Production Economics, 105, 79–96. Philips, L. W., Chang, D. R., & Buzzell, R. D. (1983). Product quality, cost position business performance. A test of some key hypotheses. Journal of Marketing, 46, 26–43. Prajogo, D. I., & Sohal, A. S. (2003). The relationship between TQM practices, quality performance, and innovation performance. An empirical examination. International Journal of Quality and Reliability Management, 20(8), 901–918. Reed, R., Lemak, D. J., & Mero, N. P. (2000). Total quality management and sustainable competitive advantage. Journal of Quality Management, 5, 5–26. Saaty, J. T. (1980). The analytic hierarchy process. Columbus: McGraw-Hill. Sallis, E. (1993). Total quality management in education. London: Kogan Page. http://www.biopharm.org.tw/2t2s/situation.html http://www.biopharm.org.tw/2t2s/situation.html 8798 J.-K. Chen, I.-S. Chen / Expert Systems with Applications 36 (2009) 8789–8798 Samson, D., & Terziovski, M. (1999). The relationship between total quality management practices and operational performance. Journal of Operations Management, 17, 393–409. Shieh, H. M., & Wu, K. Y. (2002). The relationship between total quality management and project performance in building planning phase: An empirical study of real estate industries in Taiwan. Total Quality Management, 13(1), 133–151. Shields, P. M. (1999). Zen and the art of higher education maintenance: Bridging classic and romantic notions of quality. Journal of Higher Education Policy and Management, 21(2), 165–172. Sila, I., & Ebrahimpour, M. (2002). An investigation of the total quality management survey based research published between 1989 and 2000: A literature review. International Journal of Quality and Reliability Management, 19(7), 902–970. Sitalakshmi, V. (2007). A framework for implementing TQM in higher education programs. Quality Assurance in Education, 15(1), 92–112. Sureshchandar, G. S., Rajendran, C., & Anantharaman, R. N. (2002). The relationship between management’s perception of total quality service and customer perceptions of service quality. Total Quality Management, 13(1), 69–88. Tseng, M. L., Lin, Y. H., Chiu, S. F., & Liao, C. H. (2007). A structural equation model of total quality management and cleaner production implementation. The Journal of American Academy of Business, 11(1), 65–71. Uzumeri, M. V. (1997). ISO 9000 and other metastandards: Principles for management practice? Academy of Management Executive, 11(1), 21–36. Wadsworth, H. M., Stephens, K. S., & Godfrey, A. B. (2002). Modern methods for quality control and improvement (2nd ed). NY: Wiley. Wanger, M., & Schaltegger, S. (2004). The effect of corporate environment strategy choice and environmental performance on competitiveness and economic performance: An empirical study of EU manufacturing. European Management Journal, 22(5), 557–572. Zhang, Z. H. (2000). Implementation of total quality management: An empirical study of Chinese manufacturing firms. PhD unpublished thesis. Groningen: University of Groningen. TQM measurement model for the biotechnology industry in Taiwan Introduction Total quality management Measurement criteria of TQM The Biotechnology biotechnology industry in Taiwan Definition and the technology range of the biotechnology industry Recent status on biotechnology industry Fuzzy analytic network process Analytic hierarchy process Fuzzy set theory Fuzzy number The fuzzy linguistic variable Analytic network process (ANP) Steps of fuzzy analytic network process calculation Empirical study of TQM measurement model Conclusion References 
work_vybda6uyqnfrtaswahcnifxzzq	----	Breast cancer classification using deep belief networks Expert Systems With Applications 46 (2016) 139–144 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Breast cancer classification using deep belief networks Ahmed M. Abdel-Zaher∗, Ayman M. Eldeib Department of Systems and Biomedical Engineering, Cairo University, Giza, Egypt a r t i c l e i n f o Keywords: Breast cancer diagnosis CAD Classification Deep learning based classifier Pattern recognition a b s t r a c t Over the last decade, the ever increasing world-wide demand for early detection of breast cancer at many screening sites and hospitals has resulted in the need of new research avenues. According to the World Health Organization (WHO), an early detection of cancer greatly increases the chances of taking the right decision on a successful treatment plan. The Computer-Aided Diagnosis (CAD) systems are applied widely in the detection and differential diagnosis of many different kinds of abnormalities. Therefore, improving the accuracy of a CAD system has become one of the major research areas. In this paper, a CAD scheme for detection of breast cancer has been developed using deep belief network unsupervised path followed by back propagation supervised path. The construction is back-propagation neural network with Liebenberg Marquardt learning function while weights are initialized from the deep belief network path (DBN-NN). Our technique was tested on the Wisconsin Breast Cancer Dataset (WBCD). The classifier complex gives an accuracy of 99.68% indicating promising results over previously-published studies. The proposed system provides an effective classification model for breast cancer. In addition, we examined the architecture at several train-test partitions. © 2015 Elsevier Ltd. All rights reserved. 1 n c c c f & w t C d p a e i w s c t i 2 ( 2 c i e t q L w c v a o ( t ( 9 t a m o h 0 . Introduction Breast cancer is the most common cancers among women with early 1.7 million new cases diagnosed in 2012 (Centers for disease ontrol and prevention, cancer prevention control, 2014) (World can- er research fund, 2014). Breast cancer represents 18.3% of the total ancer cases in Egypt. A percentage of 37.3% of breast cancer could be ully healed especially in case of early detection (Salama, Abdelhalim, Zeid, 2012). In Egypt and Arab countries, the breast cancer targets omen in the age of 30 and represents 42 cases per 100 thousand of he population (Salama et al., 2012). An accurate classifier is the most important component of any AD scheme that is developed to assist medical professionals in early etecting mammographic lesions. CAD systems are designed to sup- ort radiologists in the process of visually screening mammograms to void miss-diagnosis because of fatigue, eyestrain, or lack of experi- nce. The use of an accurate CAD system for early detection could def- nitely save precious lives. In this study, back propagation neural net- ork initialized by weights from a trained deep belief network with imilar architecture (DBN-NN) was used to diagnose the breast can- er. Our data source is the Wisconsin Breast Cancer Dataset (WBCD) aken from the University of California at Irvine (UCI) machine learn- ng repository (Wisconsin breast cancer dataset (WBCD) (original), 014). ∗ Corresponding author. Tel.: +20 1114488419. E-mail addresses: ahmedallah.m.s@gmail.com (A.M. Abdel-Zaher), eldeib@ieee.org A.M. Eldeib). n N w ttp://dx.doi.org/10.1016/j.eswa.2015.10.015 957-4174/© 2015 Elsevier Ltd. All rights reserved. . Background A variety of classification techniques were developed for breast ancer CAD systems. The accuracy of many of them was evaluated us- ng the dataset taken from the UCI machine-learning repository. For xample, Goodman, Boggess, and Watkins, tried different methods hat produced the following accuracies: optimized learning vector uantization (optimized-LVQ) method’s performance was 96.7%, big- VQ method reached 96.8%, and the last method, they proposed AIRS, hich depending on the artificial immune system, obtained 97.2% of lassification accuracy (Goodman, Boggess, & Watkins, 2002). Quinlan reached 94.74% classification accuracy using 10-fold cross alidation with C4.5 decision tree method (Quinlan, 1996). Abonyi nd Szeifert used Supervised Fuzzy Clustering (SFC) technique and btained 95.57% accuracy (Abonyi & Szeifert, 2003). Salama et al. 2012) performed an experiment on WBC dataset and results showed hat the fusion between MLP and J48 classifiers with feature selection PCA) is superior to the other classifiers. Hamilton, Shan, and Cercone (1996) with RIAC method obtained 6% accuracy. Polat and Günes (2007) examined the robustness of he least square Support Vector Machine (SVM) by using classification ccuracy, analysis of sensitivity and specificity, k-fold cross-validation ethod, and confusion matrix. They obtained classification accuracy f 98.53%. Nauck and Kruse (1999) obtained 95.06% with neuro-fuzzy tech- iques. Pauline and Santhakumaran used Feed Forward Artificial eural Networks and back propagation algorithm to train the net- ork (Pauline, 2011).The performance of the network is evaluated http://dx.doi.org/10.1016/j.eswa.2015.10.015 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.10.015&domain=pdf mailto:ahmedallah.m.s@gmail.com mailto:eldeib@ieee.org http://dx.doi.org/10.1016/j.eswa.2015.10.015 140 A.M. Abdel-Zaher, A.M. Eldeib / Expert Systems With Applications 46 (2016) 139–144 Fig. 1. Confusion matrix of DBN-NN. i S t v s k a r v l m c & ( l h a l t t R d h b a b ( w s t a using Wisconsin breast cancer dataset for various training algo- rithms. The highest accuracy of 99.28% is achieved when using Lev- enberg Marquardt algorithm. The accuracy obtained by Pena-Reyes & Sipper (1999) was 97.36% using fuzzy-GA method. Akay (2009) combined SVM with fea- ture selection obtaining highest classification accuracy (99.51%) for SVM model that contains five features. Moreover, Setiono (2000) was reached 98.1% using the Neuro-rule method. Übeyli (2007) used SVM and obtain 99.54% accuracy at 37% train and 63% test partition. Mert, Kılıç, Bilgili, and Akan (2015)., explored features reduction properties of independent component analysis (ICA) on breast cancer decision support system. They proofed that a one-dimensional fea- tures vector obtained from (ICA) causes Radial Bases Function Neural Network (RBFNN) classifier to be more distinguishing with the in- creased accuracy from 87.17% to 90.49%. Nahato, Nehemiah, and Kannan (2015), used a rough set indis- cernibility relation method with back propagation neural network (RS-BPNN). This work has two stages. The first stage handles miss- ing values to obtain a smooth data set and to select appropriate at- tributes from the clinical dataset by indiscernibility relation method. The second stage is classification using back propagation neural net- work. The accuracy obtained from the proposed method was 98.6% on breast cancer dataset. Dheeba, Singh, and Selvi (2014), investigated a new classifica- tion approach for detection of breast abnormalities in digital mam- mograms using Particle Swarm Optimized Wavelet Neural Network (PSOWNN). The proposed abnormality detection algorithm is based on extracting Laws texture energy measures from mammograms and classifying the suspicious regions by applying a pattern classifier. They achieved 93.671%, 92.105% and 94.167% for accuracy, specificity, and sensitivity, respectively. In our study, we applied deep belief network (DBN) in an unsu- pervised phase to learn input features statistics of the original WBCD dataset. Then, we transferred the obtained network weight matrix of DBN to back propagation neural network with similar architecture to start the supervised phase. In supervised phase, we tested both conjugate gradient and Levenberg-Marquardt algorithm for learning back propagation neural network. 3. From back propagation (BP) to deep belief network (DBN) In 1985, the second-generation neural networks with back prop- agation algorithm have emerged. However, the learning algorithm struggle to adjust network weights so that output neurons state y rep- resent the learning example t. A common method for measuring the discrepancy between the expected output t and the actual output y is using the squared error measure: E = (t − y)2 (1) The change in weight, which is added to the old weight, is equal to the product of the learning rate and the gradient of the error function, multiplied by −1: �wi j = − ∂E ∂wi j (2) where almost all data is unlabeled. However, back propagation neural network requires a labeled training data. Therefore, the biggest issue with back propagation NN appears as its possibility to get stuck in poor local optima and the learning time is huge with multiple hidden layers. In 1963, Vapnik et al. invented the original support vector machine (SVM) algorithm. Boser, Guyon, and Vapnik (1992) suggested a way to create nonlinear classifiers by applying the kernel trick to maximum- margin hyperplanes. In classification task, the weight of each feature s computed by optimization technique. In non-linear classification, VMs can efficiently perform the task using what is called the kernel rick by mapping their inputs. The non-linear classification task con- erted to linear classification problem in high-dimensional feature paces. The biggest limitation of SVM approach lies in choice of the ernel. In practice, the most serious problem with SVMs is the high lgorithmic complexity and extensive memory requirements of the equired quadratic programming in large-scale tasks (Suykens, Hor- ath, Basu, Micchelli, & Vandewalle, 2003). In recent years, the attention has shifted to deep learning. Deep earning is a set of algorithms in machine learning that attempts to odel high-level abstractions in data by using model architectures omposed of multiple non-linear transformations (Bengio, Courville, Vincent, 2013; Schmidhuber, 2014). Restricted Boltzmann Machine RBM) is a generative stochastic artificial neural network that can earn a probability distribution over its set of inputs. On the other and, Deep Belief Network (DBN) is a generative graphical model, or lternatively a type of deep neural network, composed of multiple ayers of latent variables (“hidden units”), with connections between he layers but not between units within each layer (Hinton, 2009b). From Hinton’s perspective, the DBN can be viewed as a composi- ion of simple learning modules each of which is a restricted type of BM that contains a layer of visible units. This layer represents the ata. Another layer of hidden units represents features that capture igher-order correlations in the data. The two layers are connected y a matrix of symmetrically weighted connections (W) and there re no connections within a layer (Hinton, 2009b). The key idea behind DBN is its weight (w), learned by a RBM define oth p(v|h, w)and the prior distribution over hidden vectors p(h|w) Hinton, 2009b). The probability of generating a visible vector, can be ritten as p(v) = ∑ h (p(h|w) p(v|h, w)) (3) As the learning of DBN is a computational intensive task, Hinton howed that RBMs could be stacked and trained in a greedy manner o form the DBN (Hinton, Osindero, & Teh, 2006). He introduced a fast lgorithm for learning DBN. The weight update between visible v and A.M. Abdel-Zaher, A.M. Eldeib / Expert Systems With Applications 46 (2016) 139–144 141 Fig. 2. Confusion matrix of RIW-BPNN. h � w a r e r i e t i d r P b & D t b a s t b t p 4 2 i w o p c m w d A o ( t m c N n d ( F R t idden h unites simply as: wi j = ε(〈vi.h j〉0 − 〈vi.h j〉1) (4) here 0 and 1 in the above equation designate to the network data nd reconstruction state, respectively. DBN is competitive for five reasons: DBN can be fine-tuned as neu- al networks, DBN has many non-linear hidden layers, DBN is gen- ratively pre-trained beside it can act as non-linear dimensionality eduction for input features vector, and finally the network teacher s another sensory input. In addition, the real performance of DBN ncourages using DBN. For example, in pattern recognition applica- ion, Hinton reported that the generalization performance of the DBN ig. 3. Dual vertical bar represents misclassified sample percentage. Therefore, shorter ba IW-BPNN (conjugate gradient case). The horizontal axis concentrate on percentage from (6 he references to color in this figure legend, the reader is referred to the web version of this a s 1.25% errors on the 10000 digits of the MNIST handwritten digits atabase (Hinton, 2009a). The DBN’s performance beats the 1.5% er- or achieved by the best back propagation nets (Hinton et al., 2006). latt obtained 1.6% using back propagation while K-Nearest Neigh- or produced 3.3% classification error (Lecun, LeCun, Bottou, Bengio, Haffner, 2001). It is still better than the 1.4% errors reported by ecoste and Schölkopf (2002) for SVM on the same task. In addition, Hinton discussed the possibility of unsupervised raining DBN following by back propagation pass. This will give a etter accuracy if we had good data priors (Hinton, 2009a). In 2010, n experiment performed by Erhan et al. (2010) suggested that un- upervised pre-training in prior to supervised learning tasks guides he learning towards basins of attraction of minima that support etter generalization from the training dataset. The evidence from hese results supported a regularization explanation for the effect of re-training. . Experiment conditions and methodology We used Matlab 2014a and Palm DBN implementation (Palm, 012). After the deep belief network fully trained, we transferred ts weights matrix to native Matlab back propagation neural net- ork with similar architecture, i.e. same number of input-hidden- utput neurons, and then we performed several supervised back- ropagation paths. We applied this approach on Wisconsin breast ancer original database with nine features and two classes (benign, alignant). The accepted samples are 690 from 699. Nine samples ere rejected for incomplete features. We reduced the used samples own to 683 entries to compare our results easier with others, e.g. kay (2009) used 683 samples. The Sampling is repeated randomly Sub-sampling validation f (train+validate) quanta relative to test quanta at different train+validate) to test partitions, which varied from (0.5–99.5%) o (80–20%) while train–validate partition is fixed at 70–30%. Our ethodology is to calculate misclassified sample percentage from the onfusion matrix of Randomly Initialized Weight Back-Propagation eural Network (RIW-BPNN) side by side with back propagation eural network initialized by weights obtained from a trained eep belief network with similar architecture (DBN-NN) at different train+validate) to test partitions. r is the better. The first bar (left/blue) for DPN-NN and the second (right/red) is for 0.5–39.5%) to (64.5–35.5%) of (train+validate) to test partitions. (For interpretation of rticle.) 142 A.M. Abdel-Zaher, A.M. Eldeib / Expert Systems With Applications 46 (2016) 139–144 Fig. 4. Confusion matrix of DBN-NN. Fig. 5. Confusion matrix of RIW-BPNN. T c s ( The RIW-BPNN and DBN-NN architecture is nine inputs – four hidden – two hidden - one output. To increase classifier perfor- mance for both architectures, we test conjugate gradient back prop- agation and Levenberg–Marquardt in neural network learning phase. he experiment conducted by Pauline and Santhakumaran indi- ating Levenberg–Marquardt learning algorithm gives better clas- ifier accuracy when used with back-propagation neural network Paulin, 2011). A.M. Abdel-Zaher, A.M. Eldeib / Expert Systems With Applications 46 (2016) 139–144 143 Fig. 6. Dual vertical bar represents misclassified sample percentage. Therefore, shorter bar is the better. The first bar (left/blue) for DPN-NN and the second (right/red) is for RIW-BPNN (Levenberg-Marquardt case). The horizontal axis concentrate on percentage from (53.5–46.5%) to (57.2–42.8%) of (train+validate) to test partitions. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) Table 1 Summary table of classifiers performance. Classifiers Classifier performance Accuracy (%) Sensitivity (%) Specificity (%) Train+validate to test partition (%) Ls-SVM, Akay (2009) paper Wisconsin 683 entries 99.51 100 97.91 80–20% Übeyli (2007) - SVM 99.54 – – 37–63% Dheeba et al. (2014)) - PSOWNN 93.67 94.17 92.11 – Mert et al. (2015)) - ICA-RBFNN 90.49 – – – Nahato et al. (2015)) - RS-BPNN 98.6 – – – Our tested - RIW-BPNN Wisconsin - 683 entries conjugate gradient back propagation 98.86 100 98.22 61.35–38.65%. Our tested - DBN-NN Wisconsin 683 entries conjugate gradient back propagation 99.59 100 99.39 63.84–36.16% Our tested - RIW-BPNN Wisconsin - 683 entries Levenberg–Marquardt 99.03 99.13 98.97 54.76–45.24% Our tested - DBN-NN Wisconsin 683 entries Levenberg–Marquardt 99.68 100 99.47 54.9–45.1% 5 5 m ( s c I a ( 6 a ( s t 5 a 9 . Results .1. Scaled conjugate gradient back propagation Figs. 1 and 2 show the confusion matrix obtained from experi- ent as (train+validate) to test partitions varied from (0.5–99.5%) to 80–20%), while train–validate partition is fixed at 70–30%. The re- ults show that the best classifier accuracy was 99.59% for DBN-NN omplex at (train+validate) to test partition equals to 63.84–36.16%. n comparison, the best accuracy of RIW-BPNN was 98.86% reached t ((train+validate) to test partition equals to 61.35–38.65%. At the best accuracy of DBN-NN, the total test samples (1−0.63836) ∗ 683)) =247 samples and True positive (TP) = 82, True negative (TN) = 164, False Positive (FP) = 1, False negative (FN) = 0 Sensitivity = 100 ∗ TP/(TP+FN) = 100% Specificity = 100 ∗ TN/(TN+FP) = 100 ∗ 164/165 = 99.39% At the best accuracy of RIW-BPNN, the total test samples ((1−0. 134) ∗ 683)) = 264 samples and True positive (TP) = 95, True negative (TN) = 166, False Positive (FP) = 3, False negative (FN) = 0 Sensitivity = 100 ∗ TP/(TP+FN) = 100% Specificity = 100 ∗ TN/(TN+FP) = 100 ∗ 166/169 = 98.22% Fig. 3 demonstrates part of the sample partition domain nd shows the accuracy reach 99.6% for DBN-NN complex at train+validate) to test partition equals to 63.84–36.16%. In compari- on with RIW-BPNN, best accuracy 98.86% reached at (train+validate) o test partition equals to 61.35–38.65%. .2. Levenberg–Marquardt Figs. 4 and 5 show the confusion matrix obtained from experiment t several (train+validate) to test partitions, which varied from (0.5– 9.5%) to (80–20%), while train - validate partition is fixed at 70–30%. 144 A.M. Abdel-Zaher, A.M. Eldeib / Expert Systems With Applications 46 (2016) 139–144 c t p R A A B B C D D E G H H H H L M N N P P P P Q S S S S Ü W W The best classifier accuracy of DBN-NN complex was 99.68% obtained at (train+validate) to test partition equals to 54.9–45.1% with misclas- sified sample percentage reached 0.32%. In comparison, RIW-BPNN best accuracy reached 99.03% with misclassified sample percentage 0.97% obtained at (train+validate) to test partition equals to 54.76– 45.24%. At the best accuracy of DBN-NN, the total test samples ((1−0.549) ∗ 683)) = 308 samples and True positive (TP) = 118, True negative (TN) = 189, False Positive (FP) = 1, False negative (FN) = 0 Sensitivity = 100 ∗ TP/(TP+FN) = 100% Specificity = 100 ∗ TN/(TN+FP) = 100 ∗ 189/190 = 99.47% At the best accuracy of RIW-BPNN, the total test samples were ((1−0. 548) ∗ 683)) = 309 samples and TP = 114, TN = 192, FP = 2, FN = 1 Sensitivity = 100 ∗ TP/(TP+FN) = 99.130% Specificity = 100 ∗ TN/(TN+FP) = 100 ∗ 192/194 = 98.969% From Fig. 6, we can observe that the accuracy of DPN-NN complex reached 99.68% obtained at (train+validate) to test partition equals 54.9–45.1%. Table 1 summarizes different classifier performance including our technique. 6. Conclusion In this research, we presented an automatic diagnosis system for detecting breast cancer based on DBN unsupervised pre-training phase followed by a supervised back propagation neural network phase (DBN-NN). The pre-trained back propagation neural network with unsupervised phase DBN achieves higher classification accuracy in comparison to a classifier with just one supervised phase. The ra- tionale behind this enhancement could be that the learning of input statistics from input feature space by DBN phase initializes back prop- agation neural network to search objective function near a good local optima in supervised learning phase. From our experiment at the specified network architecture, DBN-NN complex accuracy outperforms RIW-BPNN when back propagation neural network uses conjugate gradient algorithm for learning. DBN-NN still outperforms RIW-BPNN when we use Levenberg-Marquardt for training in back propagation neural net- work phase. The enhancement of overall neural network accuracy is reaching 99.68% with 100% sensitivity and 99.47% specificity in breast cancer case. Results show classifier performance improvements over previous studies. Although Hinton developed fast algorithm for training DBN, DBN’s learning process still require substantial computational effort on legacy hardware. Therefore, the main limitation/challenge of our ap- proach is to build a CAD scheme based on DBN using commercial hardware to assist medical professionals in the early detection pro- cess of breast abnormality. Future research effort should be allocated for evaluating such classifier complex for auto diagnosis of other abnormalities such as epilepsy based on EEG dataset, cardiac arrhythmia, and dia- betic retinopathy (DR). Further, the presence of general-purpose omputing on graphics processing units (GPGPU) and the distribu- ion nature of DBN may also encourage the developments of efficient arallel algorithm for learning such classifier. eferences bonyi, J., & Szeifert, F. (2003). Supervised fuzzy clustering for the identification of fuzzy classifiers. Pattern Recognition Letters, 24, 2195–2207. kay, M. F. (2009). Support vector machines combined with feature selection for breast cancer diagnosis. Expert Systems With Applications, 36, 3240–3247. engio, Y., Courville, A., & Vincent, P. (2013). Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35, 1798–1828. oser, B. E., Guyon, I. M., & Vapnik, V. N. (1992). A training algorithm for optimal margin classifiers. In Proceedings of the 5th annual ACM workshop on computational learning theory (pp. 144–152). ACM Press. enters for disease control and prevention (2014) <http://www.cdc.gov/cancer/dcpc/ data/women.htm>. Cancer prevention control. Accessed 03.09.14. ecoste, D., & Schölkopf, B. (2002). Training Invariant Support Vector Machines. Ma- chine Learning, 46, 161–190. heeba, J., Singh, N. A., & Selvi, S. T. (2014). Computer-aided detection of breast cancer on mammograms: A swarm intelligence optimized wavelet neural network ap- proach. Journal of Biomedical Informatics, 49, 45–52. rhan, D., Bengio, Y., Courville, A., Manzagol, P.-A., Vincent, P., & Bengio, S. (2010). Why does unsupervised pre-training help deep learning? Journal of Machine Learning Research, 11, 625–660. oodman, D. E., Boggess, L. C., & Watkins, a. B. (2002). Artificial immune system clas- sification of multiple-class problems. In Proceedings of the intelligent engineering systems (pp. 179–184). ASME. amilton, H.J., Shan, N., & Cercone, N. (1996). RIAC: A Rule Induction Algorithm Based on Approximate Classification. inton, G.E.(2009a).Deep belief nets. <http://www.cs.toronto.edu/∼hinton/ nipstutorial/nipstut3.pdf> Accessed 03.09.14. inton, G.E.(2009b). Deep belief networks. <http://www.scholarpedia.org/article/ Deep_belief_networks> Accessed 03.09. 14. inton, G. E., Osindero, S., & Teh, Y.-W. (2006). A fast learning algorithm for deep belief nets. Neural Computation, 18, 1527–1554. ecun, Y., LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (2001). Gradient-based learning applied to document recognition. Intelligent Signal Processing (pp. 306–351). IEEE Press. ert, A., Kılıç, N. Z., Bilgili, E., & Akan, A. (2015). Breast cancer detection with reduced feature set. Computational and Mathematical Methods in Medicine, 1–11. ahato, K. B., Nehemiah, H. K., & Kannan, A. (2015). Knowledge mining from clini- cal datasets using rough sets and backpropagation neural network. Computational Mathematical Methods in Medicine, 1–13. auck, D., & Kruse, R. (1999). Obtaining interpretable fuzzy classification rules from medical data. Artificial Intelligence in Medicine, 16, 149–169. alm, R.B.(2012). Deep learning toolbox. <https://github.com/rasmusbergpalm/ DeepLearnToolbox> Accessed 03.09.14. aulin, F. (2011). Classification of breast cancer by comparing backpropagation training algorithm. Intenational Journal on Computer Science and Engineering, 3, 327–332. ena-Reyes, C. A., & Sipper, M. (1999). A fuzzy-genetic approach to breast cancer diag- nosis. Artificial Intelligence in Medicine, 17, 131–155. olat, K., & Günes, S. (2007). Breast cancer diagnosis using least square support vector machine. Digital Signal Processing, 17, 694–701. uinlan, J. R. (1996). Improved use of continuous attributes in C4.5. Journal of Artificial Intelligence Research, 4, 77–90. alama, G. I., Abdelhalim, M. B., & Zeid, M. A. (2012). Breast Cancer diagnosis on three different datasets using multi-classifiers. International Journal of Computer and In- formation Technology, 1, 36–43. chmidhuber, J. (2014). Deep learning in neural networks: An overview. Neural Net- works, 61C, 85–117. etiono, R. (2000). Generating concise and accurate classification rules for breast can- cer diagnosis. Artificial Intelligence in Medicine, 18, 205–219. uykens, J. A. K., Horvath, G., Basu, S., Micchelli, C., & Vandewalle, J. (2003), Advances in Learning Theory: Vol. 190 p. 392. IOS Press. beyli, E. D. (2007). Implementing automated diagnostic systems for breast cancer de- tection. Expert Systems With Applications, 33, 1054–1062. isconsin breast cancer dataset (WBCD) (2014). (original). Accessed 03.09.14 <https: //archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+%28Original%29>. orld cancer research fund. (2014). <http://www.wcrf.org/int/cancer- facts- figures/ data- specific- cancers/breast- cancer- statistics> Accessed 03.09.14. http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0004 http://www.cdc.gov/cancer/dcpc/data/women.htm http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0009 http://www.cs.toronto.edu/~hinton/nipstutorial/nipstut3.pdf http://www.scholarpedia.org/article/Deep_belief_networks http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0014 https://github.com/rasmusbergpalm/DeepLearnToolbox http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref011 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00710-1/sbref0006 https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+%28Original%29 http://www.wcrf.org/int/cancer-facts-figures/data-specific-cancers/breast-cancer-statistics Breast cancer classification using deep belief networks 1 Introduction 2 Background 3 From back propagation (BP) to deep belief network (DBN) 4 Experiment conditions and methodology 5 Results 5.1 Scaled conjugate gradient back propagation 5.2 Levenberg-Marquardt 6 Conclusion References 
work_vyixrh453vfihees7ne3g4af7i	----	Contrast set mining in temporal databases Article DOI: 10.1111/exsy.12080 Contrast set mining in temporal databases André Magalhães and Paulo J. Azevedo* HASLab/ INESC TEC, Departamento de Informática, Universidade do Minho, Braga, Portugal E-mail: afgmagalhaes@gmail.com; pja@di.uminho.pt Abstract: Understanding the underlying differences between groups or classes in certain contexts can be of the utmost importance. Contrast set mining relies on discovering significant patterns by contrasting two or more groups. A contrast set is a conjunction of attribute–value pairs that differ meaningfully in its distribution across groups. A previously proposed technique is rules for contrast sets, which seeks to express each contrast set found in terms of rules. This work extends rules for contrast sets to a temporal data mining task. We define a set of temporal patterns in order to capture the significant changes in the contrasts discovered along the considered time line. To evaluate the proposal accuracy and ability to discover relevant information, two different real-life data sets were studied using this approach. Keywords: software engineering, artificial intelligence, knowledge acquisition, knowledge representation, knowledge base < system 1. Introduction Understanding the difference between contrasting groups is a fundamental data mining task (Bay & Pazzani, 1999; Webb et al., 2003). This task can be used in many different domains. For example, census data collected this year can be compared with the data collected in a previous census activity, contrasting the data collected this year against the one collected 30 years ago. This comparison involves two groups (2011 vs 1981), and in this scenario, it is fairly easy to infer some differences among these two groups in contrast: the mean number of children per couple should be lower nowadays, but the income and education likely follow the reverse trend. This notion can easily be extrapolated to other domains. Although association rule mining captures the relations between items present in the data, it does not discriminate in regard to difference towards those same items. Even so, one proposal has shown that a commercial association rule learner (Magnum_OPUS) with some tweaks could achieve this task fairly well (Webb et al., 2003). Because of this latent inability, some techniques derived from association rule mining have been proposed to tackle this problem. Contrast set mining (CSM) (Bay & Pazzani, 1999; Hilderman & Peckham, 2005; Azevedo, 2010) has emerged as a data mining task whose goal is to effectively collect contrast sets, a formalism used to represent group differences. Rules for contrast sets (RCS) is a proposal that redesigns an association rule engine to derive rules that describe contrast sets (Azevedo, 2010). By contrasting two or more groups, the aim is to obtain the attributes that distinguish them. Some proposals have been made in order to perform this task. However, none did consider further in order to contrast and differentiate groups along a time line. Such setting enables an analysis of how contrasts evolve along time. Two certain groups being contrasted in a certain point in time could have just a few distinguishable features. Nothing guarantees that the relation between them has suffered a meaningful modification somewhere in another period either in the past or in the future for a certain attribute or set of attributes. This proposal pretends, essentially, to look up on this matter. It tries to understand and identify the contrasts evolution along the defined periods, bridging together the areas of CSM and temporal data mining (TDM). The main contribution of this work is a proposal to represent group difference in a temporal database. In order to accomplish the objective of contrasting in a timely manner, we propose a set of temporal patterns. This set of patterns will allow to detect and represent situations of interest that mark a significant change in the contrasting behaviour. Potentially, this can be considered highly valuable information by the end-user. The paper is organized as follows: Section 2 briefly surveys CSM. The proposal with the patterns developed as well as the whole strategy to obtain and analyse them is described in Section 3. Section 4 presents the evaluation and application of the technique in two distinct case studies. Finally, conclusions are drawn regarding the work developed. 2. Contrast set mining Contrast set mining was first referred by Bay and Pazzani (1999), as the problem of finding all contrast sets whose support differ meaningfully across groups. Association rule mining usually deals with market basket data; however for this specific problem, the data are represented as relational © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 data. The itemset concept present in association rules can be extended to contrast sets as defined by (Bay & Pazzani, 1999). Definition 1 Let A1, A2, …, Ak be a set of k variables called attributes. Each Ai can take on values from the set {Vi1, Vi2, …, Vim}. Then a contrast set is a conjunction of attribute–value pairs defined on groups G1, G2, …, Gn. Example (sex = male) ∧ (occupation = manager) In this context, the support is considered in regard to the group and not to the whole data set, meaning that the support of a contrast set cs is the percentage of examples in group G where the contrast set is true. Formally, the objective is to find all the contrast sets (cs) that meet the following criteria: ∃ij P cs GiÞ ≠ P cs GjÞ � � �� � � (1) maxi; j sup cs; Gið Þ � sup cs; Gj � �� � � � ≥ δ (2) where δ is a user-defined threshold named minimum support difference. These two equations albeit different represent the same goal, finding contrast sets whose support differ mea- ningfully across groups. Equation (1) guarantees that the contrast set represents a true difference between at least a pair of groups (i.e. the basis of a statistical test of meaningful), and equation (2) ensures that only contrast sets whose difference is big enough to be considered relevant are obtained. The contrast sets that equation (1) is statistically valid are called significant, and those that met equation (2) are referred as large. If both criteria are met, they are considered as deviations. 2.1. Search and Testing for Understandable Consistent Contrasts Presented in the first paper that introduced CSM (Bay & Pazzani, 1999), Search and Testing for Understandable Consistent Contrasts (STUCCO) is still widely used for mining contrast sets. It is based on Max-Miner (Bayardo, 1998) rule discovery algorithm and uses a breadth-first search framework. In order to check for significant contrast sets (equation (1)), a statistical test is required. The null hypothesis to be tested is contrast set support is equal across all groups. The support counts needed for this are organized in a 2 × G contingency table where the row variable represents the truth of the contrast set and the columns represent each group considered. STUCCO uses a standard test for testing independence of variables in a contingency table, the chi-square test. A test α level has to be selected in order to check if the differences are significant. This sets the maximum probability of falsely rejecting the null hypothesis for each test. In a case where multiple tests have to be applied, this probability quickly rises. Incorrectly rejecting the null hypothesis (concluding that a difference exists when it does not) is known as a type I error or false positive. To reduce the chance of obtaining a false positive, STUCCO uses a specific Bonferroni adjustment that reduces the probability of false discoveries at lower levels, but it also decreases the number of contrast sets at these levels (those with a significant number of items). Regarding pruning, STUCCO prunes away all nodes that are not deviations. Nodes of the search tree are pruned on the basis of some criteria (Bay & Pazzani, 1999, 2001) when there is a guarantee that a node and its own subtree will not contribute for finding deviations; for this reason, they do not need to be visited further. After determining which contrast sets are interesting, STUCCO presents the results to the user in the following form: hours_per_week ¼ �20:6 : 40:2� 2880 857 161 j 0:537815 0:497388 0:389831 ¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼ d:f: chî 2 pvalue 2 38:37 4:65e-09 ¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼¼ This is a contrast set with just one item (hours per week) in a domain with three groups. In the second line, there are the absolute and relative values of support within each group and below the statistical values such as degrees of freedom, χ2 statistic value and its p-value. This representation has an intrinsic flaw because it does not show in which combination of groups there is a significant difference in support. 2.2. Rules for contrast sets Rules for Contrast Sets (RCS) (Azevedo, 2010) is a proposal that makes uses of an existing association rule engine (Azevedo, 2007) redesigned to mine contrast sets that are expressed in form of rules. Rules are known by their ease of interpretation and expressive power making this repre- sentation easier to read than the one STUCCO adopted. Like a frequent itemset algorithm, search space traversal is performed in a depth-first manner contrasting to other proposals such as STUCCO and CIGAR (Hilderman & Peckham, 2005) that do it in a breadth-first manner. This type of traversal does not fully exploit the downward closure property of support (Agrawal et al., 1993) but still leads to an efficient rule-based algorithm (Azevedo, 2010). Contrast sets mined by this algorithm have to meet equations (1) and (2). Although not specifically introduced as a equation, a minimum support criteria is also used here. This implementation, like CIGAR, uses 2 × 2 contingency tables. This allows to perceive between which exactly groups the differences are significant, but unlike STUCCO, a χ2 test is replaced by a Fisher-exact test that is directional (one-sided) to determine if the frequencies observed are significant. The p-value is computed as follows: p ¼ ∑ min b;cð Þ i¼0 a þ bð Þ! c þ dð Þ! a þ cð Þ! b þ dð Þ! a þ b þ c þ dð Þ! a þ ið Þ! b � ið Þ! c � ið Þ! d þ ið Þ! (3) where min(b, c) is the minimal value between values b and c. Fisher is an exact test and suitable for small samples. Being a directional test instead of a two-sided test, that is the case © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 of χ2, a smaller p-value is generally obtained for data sets with two groups allowing more patterns to be found within the same cutoff level (Azevedo, 2010). The directionality of the test implies a slightly different null hypothesis: H0 : P cs GiÞ ≤ P cs GjÞ � � �� � � (4) Because of this, instead of equation (1) it is used: ∃ i, j P (cs|Gi) > P(cs|Gj). Even being slightly different, the same principle applies as both derivations still capture the same significant patterns. False discoveries are controlled differently than STUCCO. The layered critical values (Webb, 2008) proposal is adapted for this context. The adjustment used in STUCCO (Bay & Pazzani, 2001) considers the number of hypothesis evaluated rather than the size of the search space, that is, only accounts for candidate patterns that met a set of constraints. However, candidates are the patterns most likely to pass the statistical test. Thus, the critical value should be adjusted by the number of patterns from which those to be tested are selected instead of the number of times the statistical test is applied. Work carried out by Webb (2007) introduces a form to calculate the search space size before deriving any rule either with market basket data or attribute–value data that are easily adapted to this scenario. This is shown in more detail in these works (Webb, 2008; Azevedo, 2010). Rules that describe the contrast sets whose support differs across groups are formally organized as follows: G1 > > G2, …, Gi > > Gj ← cs, where cs represents the contrast set and Gi each group. The pairs Gi > > Gj indicate the direction to where the support differs (Gi has bigger support than Gj). Consider the following example: The rule is to be read as follows: the occurrence of the contrast set ‘working for the state government and making an income of more than 50K’ is significantly larger within people holding a PhD than a MSc. The same occurs between PhD and BSc holders. Gsup refers to the support of the contrast set in the group (for example, 17.91% of the PhD holders are state governers and have a salary superior to 50000), and Sup(CS) is the support of the contrast set in the entire database. The p- value of the Fisher-exact test is also shown. This approach is much more readable than STUCCO output because of the rule format and the way evolved groups and direction in which difference between those groups occur are described. 3. Temporal data mining Temporal data mining is concerned with data mining of large sequential data sets (Laxman & Sastry, 2006). Temporal data may be categorized in two main types: sequential data, a sequence composed by a series of nominal symbols from a particular alphabet (Antunes & Oliveira, 2001), and time series data, also a sequence but composed of continuous and real-valued element values where each event has uniform distance in the time window (Shahnawaz et al., 2011). Time series analysis dates back longer than TDM. Stock market, medical care and weather forecasting are examples of the most common problems studied in this area (Antunes & Oliveira, 2001; Laxman & Sastry, 2006). TDM, however, has a different approach as the goals are somewhat distinct especially in the type of information expected to be retrieved (patterns versus predictions). 3.1. Sequence mining To put it simple, sequence mining seeks to unearth all patterns of interest (Shahnawaz et al., 2011) from sequential data. To discover such patterns in a sequence of events, three steps are usually associated with this approach (Antunes & Oliveira, 2001): representation and modelling of the data into a suitable form, definition of a similarity measure to compare and distinguish sequences and mining operation suitable to solve the task at hand; the general problem of sequence mining was stated by Pujari (2001) as follows: Definition 2 Let ∑ = {i1, i2, …, im} be a set of distinct items comprising the alphabet. An event is a non-empty, disordered collection of items denoted as (i1, i2, …, ik) where ij is an item in ∑. A sequence s = {t1, t2, …, tn} is an ordered set of events. 3.2. Frequent episodes Another approach for unearthing temporal patterns is the frequent episode discovery framework (Mannila et al., 1997). In this framework, the objective is to find temporal patterns (designated here as episodes) that appear a sufficient number of times from the event sequences given. Mannila and coauthors (Mannila et al., 1997) applied the framework in a telecommunication alarm setting. The main objective was to find relationships between alarms from the discovered episodes in order to better explain the problems that cause alarms to fire and to predict severe faults. The sequence of events composes the input provided. Each event has an associated time of occurrence. Given a set E of event types, an event is a pair (A, t) where A ∈ E and t is the time of the event. An event sequence s on E is three-tuple (s,Ts,Te), where s ={(A1,t1),(A2, t2), …,(An,tn)} is an ordered sequence of events such that Ai ∈ E for all i = 1,…, n, and ti ≤ ti + i for all i = 1,…, n � 1. Ts represents the starting time and Te the ending time with TS ≤ ti < Te for all i = 1,…, n. The concept of interest for a frequent episode is given by how close in time the events of an episode must arise. The user is able to define the width of the time window within which the episode must occur. Formally stated, a window Gsup = 0.17191 | 0.04121 p = 1.1110878451E-017 education=Doctorate >> education=Masters Gsup = 0.17191 | 0.01681 p = 3.0718399575E-040 education=Doctorate >> education=Bachelors Sup(CS) = 0.03097 < ��- workclass=State-gov& class > 50 K. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 on an event sequence s = (s, Ts, Te) is an event sequence w = (w, ts, te), where ts < Te and te > Ts, and w consists of those pairs (A, t) from s where ts ≤ t < te. The time elapsed te � ts is designated as the width of the window w. An episode is a partially ordered collection of events occurring together (Mannila et al., 1997).The notion of frequency of an episode assume a similar meaning as support. It is defined as the fraction of all fixed-width sliding windows over the data in which the episodes occur at least once (Mannila et al., 1997). 3.3. Contrast and temporal mining Some recent proposals exist in the literature that applies CSM in a temporal framework. For instance, (Wang et al., 2013) uses contrast mining to address action recognition in videos by modelling spatial–temporal structure of human poses. They use a method to estimate human join locations from videos. Data are obtained from these estimations. The authors apply emerging patterns (Dong & Li, 1999) to these data to obtain spatial and temporal parts sets. These patterns assist in deriving body actions. No patterns are derived describing how contrast evolve along time. (Yang et al., 2008) proposed a new methodology in the context of classification learning that carries out contrast mining by measuring the degree of conceptual equivalence between groups. The authors did not consider a temporal dimension along the used data. In (Langohr et al., 2012), the authors proposed to extend subgroup discovery, where interesting subsets of objects of a given class are found, by a second subgroup discovery step to find interesting subgroups of objects specific for a class in one or more contrasting classes. They applied this method to gene potate database with three time points of observation. This method does not derive patterns representing contrasts evolving along time. One of the main sources of failure in concurrent systems is unforeseen encompassed interleavings. (Leue & Befrouei, 2013) proposed to apply sequential mining methods for re- vealing unforeseen interleavings in the form of sequences of actions derived from counterexamples. The used data are sequential data representing sequences of actions, that is, traces. The proposed method is based on contrasting the pat- terns of a set of counter examples with the patterns of a set of correct traces that do not violate a desired property. Although sequential data carry a temporal dimension, this work does not consider how contrasting patterns evolve along time. 4. Proposal The method can be summed up in a three-step process that occurs in a serialized manner. Figure 1 represents the whole process and how each individual step is related, showing the output that is produced at each stage that serves as input for the next step. The output that is produced at each stage serves as input for the next step. The RCS is the first operator in the chain. It is the algorithm included in CAREN (Azevedo, 2007) for discovering contrast sets in any given data set. This will be used in order to obtain the contrasts at each observation (single period of time considered). Having individual data sets for each period, it will be executed as many times as the number of periods. For each execution, comma-separated values output file that contains all the contrasts found and related information such as group support values and p-values, among others, will be produced. Post-processing contrast set (PPCS) was developed in order to process the set of output files produced by RCS at each period and to yield the temporal patterns that occur in them. An additional file is produced in order to be used by the PPCS Viewer. The viewer emerges as an optional but recommended manner to interpret the output given in the last step. Because there is some inability to interpret the results given in a textual format (at least not in an easy and intuitive manner), a graphical tool (Viewer) was developed in order to surpass those difficulties. It makes use of graphical representations such as histograms, filtering and searching features that significantly improve readability and increase the user interactivity. 4.1. Comparing contrast sets from different periods Despite being a self-sufficient form of rule, contrast sets require that a measure of interest is defined. This will enable the contrast comparison from different periods and will be able to contribute for patterns finding. That measure would ideally capture the contrast’s own strength at each period. Its evolution along the time line would reveal if that specific contrast got ‘stronger’ or instead got ‘weaker’. Supdif (difference in group support) remains as the only and obvious choice, and it indeed answers successfully the questions posed before. It can act as a measure of interest to gauge the strength of a contrast (bigger the difference, stronger the contrast). Observations regarding how the contrast evolved along the time can be done with ease, and some patterns involving this notion will be presented next. 4.2. Patterns of interest From the contrasts present at each single point in time, the goal is to find patterns that can somehow express how some contrast has evolved along time. Those results are presented in the form of temporal patterns. 4.2.1. Increase and shrink The first patterns relate to the widening and narrowing of the support difference of the involved groups in consecutive periods. The issue here revolves around the quantification of how much does the contrast needs Figure 1: Process overview. © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 to grow or shrink to consider it a significant change. It is clearly dependent on the context involved that definitely imposes the requirement of some input from the user who should be able to suggest the adequate value due to its domain dependency. This threshold is called the sigdif and operates much like support and confidence. If the difference from one period to the next is greater (in module) than the sigdif value, then the variation is deemed as significant and should be reported. With this threshold, the first two patterns appear, and they are called increase and shrink. They are the dual of each other with the first referring to the situation where some contrast has its supdif grow bigger than the sigdif value from point N to point N+ 1. The second is the exact opposite. These two patterns assume an important role, because they alert the end-user to a relevant spike in a contrast by moving to the next period. This change highlights that the groups being contrasted suffered some kind of modification for that specific antecedent and that change might be potentially informative for the end-user. This might help to locate some specific contextual phenomenon that occurred at that time and thus enable him or her to establish some possible relation of cause–effect. 4.2.2. Spring up and fade out Two other patterns came up one opposite to another, much like the two listed earlier. This time, the goal here solely involves the appearance and disappearance of a contrast in the periods considered. Consider an example where a contrast is found for period N and N + 2 but not for period N + 1. This ‘hole’ should trigger the analyst to query what happened at that moment. Knowing exactly why there is no contrast might entail strategical and valuable information. From point N to point N + 1, there is the disappearance of the contrast. This kind of pattern is referred as fade out. That same contrast arises again from period N + 1 to period N + 2, an example of an occurrence of a spring up. 4.2.3. Flip The last pattern is the flip. The name selected is well representative of its nature because of the ‘180 turn’ notion that this pattern entails. Let’s consider that for some antecedent, there are two groups being contrasted, A and B. At some point in time, the contrast A > > B exists, but a few periods later, this contrast disappears and gives place to B > > A. Hence, the name flip because of the contrast directionality was turned around. Because of its specific nature, the flip is the less frequent temporal pattern. 4.3. Stability measure Apart from the patterns developed and described before, the lack of a global mean to evaluate a contrast motivated the development of a measure. The patterns introduced with exception of flip operate in consecutive periods (i.e. locally) and do not allow to categorize or obtain the general behaviour of a specific contrast in its whole lifetime. The existence of a numerical value that could gauge the variability of a contrast would provide an easy and intuitive manner to understand how the contrast evolved. For instance, it would enable to verify whether it suffers frequent abrupt changes or instead it has remained relatively stable in all considered periods. To achieve its purpose, this stability measure will be based on the following two premises: • The maximum score or value will be given to a contrast that appeared in all the periods considered and did not suffer any significant variations (no increase or shrink patterns). • Any pattern found will contribute to lower the score because they translate significant variations that affect what we consider contrast stability. The proposed formula for this measure that abides by the remarks stated earlier is as follows: Stability ¼ T � P 2 N � 1 (5) N represents the number of periods considered. T stands for the number of consecutive periods with contrasts found. P is the number of increase and shrink patterns found in the whole time line. These are the patterns that affect stability, but only a factor of 0.5 is considered to diminish the impact in the computed stability value. In the denominator, N � 1 simply represents the number of transitions present in the periods considered. Best case scenario is T = N � 1, which means that contrasts have been found for every single period. Thus, it becomes evident that stability varies from 0 to 1. If there are no two contrasts found in consecutive periods, then T = 0. Consequently, stability = 0, which seems adequate because there is absolutely no consistency as contrasts that appear in one period immediately disappear in the next one. 4.4. Post-processing contrast set implementation The application developed can be summarized in a high- level, simplified pseudo-code listed in algorithm 1. Algorithm 1. PPCS pseudo-code input : les F sigdif S output : results R 1 R := ∅; 2 D := ∅; 3 validateUserInput( ); 4 foreach file ∈ F do 5 D + = insertIntoDataStructure( file); 6 end 7 foreach antecedent ∈ D do 8 R + = findFlip(antecedent); 9 foreach contrast ∈ antecedent do 10 R + = _ndPatterns(antecedent, contrast, S; 11 R + = calculateStability(antecedent; contrast); 12 end 13 end 14 Return R; © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 The set of files (F) produced by CAREN will be read and its contents inserted into the data structure in a multi-threaded fashion having each thread processing one file. Then, the list of antecedents (A) and contrasts (C) will be traversed in order to find patterns and calculate stability. This algo- rithm has a time complexity of Θ(|F| + |A| × |C|). 4.5. Post-processing contrast set results viewer This application intends to be an alternative to the results expressed in a textual form. It makes use of visual representations and some other features to perform different tasks. By using graphical features, it will enrich the analysis, increasing the readability and understanding of the contrast sets previously found. After loading the output file from PPCS, the main frame containing all the features available is constructed, and it is represented in Figure 2. Every contrast set found is contained in a tree-like representation as seen in Figure 2, area 2. It follows a two- level approach allowing each antecedent to be expanded. It will show which contrasts were found for that same antecedent. This allows for a better organization of the contrast set and makes navigation more intuitive. By clicking in an antecedent or contrast, the graphical pane (area 3) is updated. What is represented is dependent on what is selected in the tree model. This entails the close relation between both components and how they depend on each other. A chart like an histogram expresses the contrasts evolution along the time line. It permits a quick identification of periods with and without contrasts as well as the patterns found. Tree model usually has a considerable number of items present, and locating some specific antecedent might involve some scrolling effort that is not desirable. The features present in area 4 have been implemented in order to improve that situation. The first one relates to a typical filter as indicated by the label and a text box in which the user can type. This works as a filter over antecedents if the introduced text matches. The other feature discards antecedents which number of items are inferior to the number present in the spinner. This is useful for finding the complex antecedents which can reveal interesting relations. Flips usually appear just a few times, and a mechanism to quickly spot them was developed in the form of a toggle button (area 5). When pressed on, the Tree model shows only the contrast sets that contain at least one flip pattern. In area 6, the antecedent relaxation feature is present. It attempts to help the user in finding a possible explanation to why some specific contrast was not discovered in a specific period. If an antecedent with minus one item than the antecedent being analysed has a contrast in that specific period, one might conclude that the item removed may be the main reason (or at least a contributor) for that event. For a given antecedent, all its one-step-above general- izations are considered. For the selected period and contrast, a list of antecedent generalizations is constructed. The option all periods widens the test not only to one period but also to all periods without contrasts. If the antecedent being tested has at least one contrast in a period where the more specific one did not, the more general antecedent is added to the list. Still, there is the possibility of another outcome. This happens when no antecedent with one less item has a contrast for the selected period (or periods). In that case, a dialogue pops up querying the user whether he or she intents to find a generalization of that antecedent (of any size) that has a contrast for that specific period. 5. Case studies and experimentation In order to ascertain the accuracy of the proposal, two distinct data sets were studied. First scenario involved the study of data collected from the Portuguese Ministry of Labor and Social Security for all employed individuals in the private sector ranging from 1986 to 2009 (except years 1990 and 2001 where data were not available). The main goal was to check how the gender of an individual affects attributes such as salary and education. Each year was considered as an observation, comprising a total of 22 time points. The results obtained were highly discriminative in regard to gender, and some early suspicions were confirmed. In Figure 2: Viewer main frame decomposed by areas. © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 Figure 3, for the higher tiers of income, the contrast of sex = male > > sex = female was always present regardless of the period considered. This confirms common sense believing that, in average, men earn more than women. One attribute that displayed effective modification in the years considered was the workers’ education. The current trend is that women pursue higher education more than male counterpart with the gap between them increasing at a steady pace. Figure 4 corroborates this situation. The other case study was related to sports, more specifically basketball. The analysis aims to understand how each position on the field affected the typical statistical contribution and how it evolves over the years. The data obtained ranged from 1946 to 2010 with every player totals from each regular season in the NBA. Each period was defined as a decade. Three groups were considered accor- ding to a broader set of positions: guards (G), forwards (F) and centres (C). The results pointed towards an increasingly positional discrepancy, where players tend to have a more specific skillset regarding their position on the field. In the early days, the difference between players playing different roles was not as significant as it is nowadays in NBA seasons. Another attribute that marked a clear change in the sport was the height of the players along each position. Labelled as a big men sport, this tendency is observed along time with growing emphasis. Figure 5 reveals the contrasts found for players that are 6 ft 6 in. tall (198 cm). In early days, the contrast pos = C > > pos = G is present, which states that players with less than 2 m were tall enough to be centre players (usually the biggest player in the team). Nowadays, players with that height play the guard position (smaller players in the field), justified by the contrast pos = G > > pos = C in later periods. This emerged a flip pattern. The obtained patterns enable to categorize each positional contribution in terms of the considered attributes. For guards, it was evident that the better three point, free throw percentage, the more steals and assists than players from other positions and with smaller height. Centre players exhibit better shooting percentage, the ability to get more Figure 3: Contrast found for employees whose income is over 2.1 (in natural log scale). Figure 4: Contrasts found for individuals with higher education. Figure 5: Contrasts found for height = 6 ft 6 in. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 rebounds and blocks, than other players and those with bigger height. As for forwards, they tend to stay in the middle ground between guards and centres. This seems to sustain their known versatility and mixed characteristics from the other positions. 6. Conclusion This paper aimed to bring the concept of discrimination pattern to a temporal setting in order to check how group differences evolved along time. The post-processing scheme employed has its merits because it was effortless to import the contrasts found by RCS usage into the application developed (PPCS). However, another approach was also possible that would integrate the role of RCS and PPCS into a single application. This could probably relax the process from the user standpoint. It would imply the reduction of the number of performed steps. There was also the possibility to obtain a faster execution time by removing certain steps like comma-separated values files creation and import. A frequent setback involved the presence of continuous, numerical attributes. This always called for a discretization treatment on this type of attributes. Research on CSM generally overlooks this situation and assumes that the data are composed exclusively of categorical attributes with a finite set of values, which is not always the case. The two data sets used required numerical discretization for a considerable number of attributes. There has been a proposal for the incorporation of the discretization process in the contrast set mining algorithm (Simeon & Hilderman, 2007) but has a significant drawback of increasing the size of the search space. This matter could be looked upon in a future iteration, as the results produced are directly affected by the specific chosen discretization method. The patterns developed for this effect had a significant impact in revealing intriguing situations and on others that contain the so-called common knowledge. Still, they tend to rely solely on the stepping from one period to the next, not focusing in a more global form of behaviour that considers a set of periods. Stability arose as a way to tackle this. Despite being able to characterize the contrast evolution in terms of its general behaviour, there are some situations that could benefit from a special emphasis given by a new pattern (or set of patterns). The example present in Figure 4 could be one of those cases. From 1998 onwards, the supdif is steadily increasing, but because it never increases more than the 1% sigdif threshold defined in each passing period, there are no increase patterns. Despite this, one of these sequences of continuous expansion could be meaningful, and future work could look upon this matter, obtaining new patterns to stress this potentially intriguing situations. A time window concept like the one used in the work by Mannila et al. (1997) could serve this purpose by defining a number of periods that could reveal a persistent trend. The graphical user interface application developed to inspect the obtained results, the Viewer, although being something not considered in an initial phase assumed a prominent role by significantly improving the analysis compared with the previous form (i.e. the results presented in a simple fashion, by just displaying them in a plain text file). Its features also allow to reduce the set the contrasts presented, detect the patterns found easily and provide visual elements and aspects that aid the comparison effort. References AGRAWAL, R., T. IMIELIŃSKI and A. SWAMI (1993) Mining association rules between sets of items in large databases, SIGMOD Record, 22, 207–216. ANTUNES, C. M. and A. L. OLIVEIRA (2001) Temporal data mining : an overview. Lecture Notes in Computer Science, 1–15. AZEVEDO, P. J. (2007) CAREN - class project association rule engine, http://www.di.uminho.pt/~pja/class/caren.html [28 July 2014]. AZEVEDO, P. J. (2010) Rules for contrast sets, Intelligent Data Analysis, 14, 623–640. BAY, S. D. and M. J. PAZZANI (1999) Detecting change in categorical data: mining contrast sets, in Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’99, New York, NY, USA: ACM, 302–306. BAY, S. D. and M. J. PAZZANI (2001) Detecting group differences: mining contrast sets, Data Mining and Knowledge Discovery, 5, 213–246. BAYARDO, R. J. JR. (1998) Efficiently mining long patterns from databases, in Proceedings of the 1998 ACM SIGMOD International Conference on Management of Data, SIGMOD ’98, New York, NY, USA: ACM, 85–93. DONG, G. and J. LI (1999). Efficient mining of emerging patterns: discovering trends and differences, in Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’99, New York, NY, USA: ACM, 43–52. HILDERMAN, R. J. and T. PECKHAM (2005) A statistically sound alternative approach to mining contrast sets, in Proceedings of the 4th Australasian Data Mining Conference (AusDM), Gold Coast, Queensland, Australia: AusDM 2007, 157–172. LANGOHR, L., V. PODPECAN, M. PETEK, I. MOZETIC and K. GRUDEN (2012) Contrast mining from interesting subgroups, in Bisociative Knowledge Discovery, Volume 7250 of Lecture Notes in Computer Science, Berthold, M. (ed.), Berlin, Heidelberg: Springer, 390–406. LAXMAN, S. and P. SASTRY (2006) A survey of temporal data mining, Sadhana, 31, 173–198. LEUE, S. and M. BEFROUEI (2013) Mining sequential patterns to explain concurrent counterexamples, in Model Checking Software, Volume 7976 of Lecture Notes in Computer Science, Bartocci, E. and C. Ramakrishnan (eds), Berlin Heidelberg: Springer, 264–281. MANNILA, H., H. TOIVONEN and A. INKERI VERKAMO (1997) Discovery of frequent episodes in event sequences, Data Mining and Knowledge Discovery, 1, 259–289. PUJARI, A. (2001) Data Mining Techniques. India, Hyderguda, Hyderabad: Universities Press. SHAHNAWAZ, M., A. RANJAN and M. DANISH (2011) Temporal data mining: an overview, International Journal of Engineering and Advanced Technology, 1, 20–24. SIMEON, M. and R. J. HILDERMAN (2007) Exploratory quantitative contrast set mining: a discretization approach, in ICTAI (2)’07, Patras, Greece: ICTAI 2007, 124–131. © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 http://www.di.uminho.pt/~pja/class/caren.html WANG, C., Y. WANG and A. YUILLE (2013) An approach to pose- based action recognition, in Computer Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on, Portland, Oregan, 915–922. WEBB, G. I. (2007) Discovering significant patterns, Machine Learning, 68, 1–33. WEBB, G. I. (2008) Layered critical values: a powerful direct- adjustment approach to discovering significant patterns, Machine Learning, 71, 307–323. WEBB, G. I., S. BUTLER and D. NEWLANDS (2003) On detecting differences between groups, in Proceedings of the Ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’03, New York, NY, USA: ACM, 256–265. YANG, Y., X. WU and X. ZHU (2008) Conceptual equivalence for contrast mining in classification learning, Data & Knowledge Engineering, 67, 413–429. The authors André Magalhães André Magalhães did his MSc (2012) at Departamento de Informática of Universidade do Minho in Data Mining. His research interests are pattern mining and contrast sets mining. He is currently at Deloitte (Portugal) enrolled in data warehousing projects. Paulo J. Azevedo Paulo J. Azevedo obtained his PhD from Imperial College at University of London (1995) and is now an assistant professor (with habilitation) at the Department of Informatics University of Minho. He is a member of the High-Assurance Software Lab (HASLab) – INESC Tec L. A. His research interests are knowledge discovery in databases (data mining) and machine learning, in particular association rules, subgroup discovery, motif discovery in time series, graph mining, bioinformatics and recommendation systems. His most recent research interests include distribution learning and social networks. He was a PC member of several data mining conferences such as DS, PKDD, ECML, ECAI and DaWak. He was also a reviewer for several data mining journals. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 
work_vz6x6d5pa5ccricho676aoithm	----	doi:10.1016/j.eswa.2007.04.001 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 2417–2423 Expert Systems with Applications Design of BOM configuration for reducing spare parts logistic costs Muh-Cherng Wu *, Yang-Kang Hsu Department of Industrial Engineering and Management, National Chiao Tung University, Hsin-Chu, Taiwan, ROC Abstract This paper proposes an approach to reduce the total operational cost of a spare part logistic system by appropriately designing the BOM (bill of material) configuration. A spare part may have several vendors. Parts supplied by different vendors may vary in failure rates and prices – the higher the failure rate, the lower the price. Selecting vendors for spare parts is therefore a trade-off decision. Consider a machine where the BOM is composed of s critical parts and each part has k vendors. The number of possible BOM configurations for the machine is then ks. For each BOM configuration, we can use OPUS10 (proprietary software) to calculate an optimum inventory policy and its associated total logistic cost. Exhaustively searching the solution space by OPUS10 can yield an optimal BOM configuration; however, it may be formidably time-consuming. To remedy the time-consuming problem, this research proposes a GA-neural network approach to solve the BOM configuration design problem. A neural network is developed to efficiently emulate the function of OPUS10 and a GA (genetic algorithm) is developed to quickly find a near-optimal BOM configuration. Experiment results indicate that the approach can obtain an effective BOM configuration efficiently. � 2007 Elsevier Ltd. All rights reserved. Keywords: Bill of material; Spare parts; Stocking policy; Genetic algorithm; Neural network 1. Introduction Machine availability is very important in capitally inten- sive industries. The higher the machine availability, the higher is the capacity. The level of machine availability partly depends on the inventory level of its spare parts. At a lower inventory level, the time required to repair a machine would be longer due to having higher possibility of lacking spare parts. A higher inventory level by contrast would increase machine availability at the expense of pay- ing more inventory cost. Since spare parts in capitally intensive industries are quite expensive, much research investigated the stocking policies for spare parts to resolve the trade-off decision (Sherbrooke, 2004). Spare parts are typically replenished through a multi- echelon supply chain system, which is a hierarchical structure comprising multiple layers of facilities (Fig. 1). A facility has two main functions: storing and repairing 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.04.001 * Corresponding author. Tel.: +886 35 731 913; fax: +886 35 720 610. E-mail address: mcwu@cc.nctu.edu.tw (M.-C. Wu). spare parts. Facilities in the lowest layer directly supply parts to machines, whereas those at a high layer supply parts to its succeeding lower layer facilities. Facilities at a higher layer are generally equipped with higher repairing capability. That is, a part that cannot be repaired by a par- ticular facility would be sent upward to its parent facility. In this paper, the information for characterizing such a supply chain system is called BOS (bill of stations). A machine is typically composed of several modules; each module comprises several assemblies that are assem- bled by subassemblies/parts (Fig. 2). The hierarchical rep- resentation for modeling the materials of a machine is called BOM (bill of materials), where a layer in the BOM structure is usually called an indenture in literature. In the BOM, each part is defined with several BOS-independent attributes such as cost and failure rate, and some BOS- dependent attributes such as eligible stations for repairing the part. At a globalization era, a part tends to have multiple ven- dors who may provide parts that are functionally identical but with various failure rates and costs. The lower the mailto:mcwu@cc.nctu.edu.tw Module Sub-assembly Assembly (LRU) Part (SRU) Part (DP) Part (SRU) Assembly (DU) Sub-assembly Assembly (LRU) Part (DP) Part (SRU) Part (SRU) Part (SRU) Part (SRU) Part (SRU) Machine A B C D E F G H I J Fig. 2. BOM of a multi-indenture machine. Central Distribution Center Local Stocking Base Regional Stocking Facility Regional Stocking Facility Local Stocking Base Local Stocking Base Vendors Local Stocking Base Local Stocking Base Local Stocking Base Fig. 1. BOS of a multi-echelon logistic system. 2418 M.-C. Wu, Y.-K. Hsu / Expert Systems with Applications 34 (2008) 2417–2423 failure rate, the higher is the price of a part. Purchasing more reliable parts (i.e. with lower failure rates) would reduce the inventory level of spare parts required to main- tain the target machine availability. As a result, this would reduce the holding cost of part inventory at the expense of increasing its purchasing cost. Choosing an appropriate configuration of BOM is therefore a very important way to reduce the total operational cost of a spare part supply chain system. Yet, this idea has been rarely noticed in literature. This paper proposes the idea of choosing appropriate BOM configurations, formulates the decision problem, and develops an efficient approach to solve the problem. Suppose a machine has s critical parts, each of which has k vendors. The possible number of BOM configurations may be quite huge (ks). It might take a formidable computation time if we exhaustively evaluate the performance of each BOM configuration by using the evaluation methods developed in literature. To efficiently solve the problem, we propose a GA-NN approach. The NN (neural network) technique is used to efficiently emulate the function of an existing method for evaluating BOM configurations, whereas the GA tech- nique is used to efficiently identify a near-optimal BOM con- figuration from the huge solution space. The remainder of this paper is organized as follows. Sec- tion 2 reviews the literature on stocking spare parts for a supply chain system. Section 3 introduces OPUS10 (pro- prietary software), which can be used to evaluate the perfor- mance of a BOM configuration. That is, OPUS10 can yield the optimal stocking policy and its associated total opera- tional cost for a BOM configuration. Section 4 describes the procedure for establishing the neural network that could emulate the function of OPUS10. Section 5 presents the genetic algorithm. Section 6 illustrates the experimental results and concluding remarks are placed in Section 7. 2. Related literature The inventory policy for a multi-echelon repairable item (spare part) system has been a research issue for several dec- ades. Much literature has been published and some of them have included a comprehensive survey (Diaz & Fu, 2005; Guide & Srivastava, 1997; Kennedy, Patterson, & Freden- dall, 2002; Rustenburg, 2000; Sleptchenko, 2002). These previous studies can be categorized into two main streams. The first stream addressed a scenario equipped with an infinite repair capacity. An early and representative study is the METRIC (Multi-echelon Technique for Recoverable Item Control) model developed by Sherbrooke (1968). Many studies that extend the METRIC model were subse- quently developed (Graves, 1985; Muckstadt, 1973; Sher- brooke, 1986, 2004; Simon, 1971; Slay, 1984). With the ample-server assumption, the queue-time for repair is negli- gible; therefore, these METRIC-variant models have been able to deal with large and complex systems. However, a real-world problem is typically equipped with limited repair capacity. The METRIC-variant models thus tend to under- estimate the stocking levels in some real applications. The various versions of the METRIC-variant models can be characterized from three perspectives: the demand pattern of spare part, BOM, and the complexity of supply chain. In the original METRIC model (Sherbrooke, 1968), the BOM is a single-indenture system involving multi- repairable-items, the demand is a compound Poisson process, and a replenishment (S, S � 1) policy is used throughout a two-echelon supply chain. Simon (1971) developed a METRIC-variant model that uses (s, S) policy in the second-echelon facilities. Muckstadt (1973) enhanced the METRIC model by including two-indenture BOM systems. Slay (1984), Graves (1985), and Sherbrooke (1986) further developed methods in order to estimate the variances of service levels. Hausman and Erkip (1994) broadened the METRIC model by including a scenario where an emergency-ordering policy is allowed. The second stream addressed a scenario with finite repair capacity, which leads to the need of modeling the machine-repair queueing behavior. The models in this stream, more realistic than the METRIC-variant models, are certainly more difficult to solve. Due to the inherent complexity, by the possible inclusion of enumeration tech- niques, most of these studies are computationally extensive (Gross, Miller, & Soland, 1983). Approximations to reduce the complexity have therefore been proposed in order to M.-C. Wu, Y.-K. Hsu / Expert Systems with Applications 34 (2008) 2417–2423 2419 develop efficient algorithms for determining the capacities of repair facilities as well as spares level (Albright, 1989; Albright & Soni, 1988a, 1988b; Diaz & Fu, 1997; Gross & Miller, 1984; Gross, Kioussis, & Miller, 1987; Kim, Shin, & Park, 2000; Rustenburg, van Houtum, & Zum, 2001; Sleptchenko, van der Heijden, & van Harten, 2002). In summary, most previous research focused on how to appropriately determine stocking level and repair capaci- ties to reduce the total logistic costs of spare parts. Our problem of interest—how to choose a BOM configuration to reduce the total logistic cost of spare parts has been scar- cely addressed. 3. OPUS10 OPUS10, proprietary software of a Swedish company, was developed based on the techniques of the METRIC- variant models. The input to OPUS10 involves the BOM/ BOS of a spare part supply chain system. The output of OPUS10 yields optimal stocking levels for achieving target machine availability, and computes the associated total operational cost of the supply chain. The BOS, BOM, and the input/output relationships of OPUS10 are described below. 3.1. BOS As shown in Fig. 1, the BOS of a spare part supply chain in OPUS10 is a hierarchical structure (also called multi- echelon in literature) consisting of multiple facilities. The attributes for describing a BOS includes the number of lay- ers in the hierarchy, the number of facilities at each layer, the replenishment/repair logistic relationships between any two facilities, the lead time between any two facilities, and the number of machines logistically supported by each facility at the lowest layer. The facility at the lowest layer, which directly supports machines, hereafter is called a ter- minal facility. A replenishment logistic relationship defines the eligibil- ity of a particular facility to supply quality parts to other facilities. A repair logistic relationship defines the eligibility of a particular facility to send a failed part to other facili- ties for repair. Two facilities, if not eligible for shipping parts between each other, do not have a logistic relation- ship. The more number of logistic relationships denotes the more flexibility is embedded in the hierarchy of BOS. Lateral transshipment (logistic relationships among facili- ties of the same layer are eligible), which has been investi- gated by some literature (Axsäter, 2003; Lee, 1987; Wong, Cattrysse, & van Oudheusden, 2005), can also be modeled in the BOS of OPUS10. 3.2. BOM As shown in Fig. 2, the BOM is a hierarchical structure that involves two main types of information: the hierarchi- cal relationship and the attributes of each part/module. The typical attributes of a part/module involve the failure rate, the unit cost, and the facilities that are eligible for repairing the part/module. The modules/parts of a BOM are classified into two types: replaceable and discardable. A replaceable part/mod- ule denotes that it is repairable, which involves two types: LRU and SRU. An LRU (line replacement unit) is a part/module that can be directly replaced by a terminal facility. Any LRU if failed can be individually taken out from the machine, replaced by a quality unit, and sent upward for repairing. Referring to Fig. 2, an SRU (shop replacement unit), a member of a particular LRU assem- bly, cannot be individually taken out from the machine. When an SRU fails, we cannot replace the SRU on line. Rather, we have to replace its parent LRU on line and sent the LRU upward to a shop-facility, where the LRU can be disassembled so that the SRU can be taken out from the LRU for repairing, and a quality SRU can be filled in the LRU. A discardable part/module denotes that it is non-repair- able, which also involves two types: DU and DP. A DU (discardable unit) can be individually taken out from the machine. When a DU fails, we can take it out on line, dis- card it, and replace it by a new quality DU. A DP (discard- able part), like SRU, cannot be individually taken out from the machine and always has a parent LRU. When a DP fails, we cannot discard the DP on line. Rather, we have to replace its parent LRU on line and sent the LRU upward to a shop-facility, where the LRU can be disassem- bled so that the failed DP can be taken out and discarded for filling in a new quality DP in the LRU. In summary, each part/module in the BOM has a replacement attribute to clarify where it is of LRU, SRU, DP, or DU. The distribution of the replacement attributes would significantly affect the total operational cost of a spare part logistic system and have to be modeled in the BOM. 3.3. Input/output relationship of OPUS10 The input/output relationships of OPUS10 in dealing with a multi-vendor spare part supply chain system that adopts (S � 1, S) stocking policy to achieve a specified machine availability can be formulated as follows: ðSij; TSCÞ¼ fðA; rj; Qj; T pr j ; P jk; F jk; T tr io; C tr io; T re ij ; C re ijÞ ð1Þ where i index of facility, 1 6 i 6 I j index of part, 1 6 j 6 J k index of vendor, 1 6 k 6 K Sij stocking level of part j at facility i TSC total operation costs of the supply chain A target average machine availability specified by users Qj quantity of part j in the BOM rj replacement attribute of part j, which denotes that part j is LRU, SRU, DP or DU Input Layers Hidden Layers Output Layers 1x 2x nx 11w 12w nw1 1mw 2mw mnw 1y my 2y Fig. 3. Architecture of a back propagation neural network. 2420 M.-C. Wu, Y.-K. Hsu / Expert Systems with Applications 34 (2008) 2417–2423 T prj production/purchasing lead time of part j Pjk unit price of part j provided by vendor k Fjk failure rate of part j provided by vendor k T trio transportation time between facilities i and o Ctrio transportation cost between facilities i and o T reij repair time of part j at facility i Creij repair cost of part j at facility i In Eq. (1), f denotes the function of OPUS10. The right- hand side represents the input parameters of OPUS10, which characterize the BOM/BOS of the supply chain and the user-specified machine availability (A). The left- hand side represents the output of OPUS10 – an optimal setting for stocking levels (Sij) and the total operational cost at this setting (TSC). In dealing with the decision of selecting BOM configura- tions, the function of OPUS10 can be further formulated as follows: ðSij; TSCÞ¼ fðV jA; rj; Qj; T pr j ; P jk; F jk; T tr io; C tr io; T re ij ; C re ijÞ ð2Þ where V = [vj]1·J, 1 6 vj 6 K, vj denotes the vendor that supplies part j. That is, V is used to model a BOM configuration, which represents the decision variables of this problem. Sij (1 6 i 6 I and 1 6 j 6 J) are intermediate output variables denoting optimal stocking policy, which can be used to compute TSC – the final output variable. The combination of these decision variables would yield K J possible BOM configurations, where KJ = 59,049 if K = 3, J = 10. Using a particular personal computer, our experiment indicates that it takes OPUS10 about 5 s to evaluate the performance (TSC) of a BOM configuration. For a decision problem with KJ = 59,049, the required computation time is about 3.4 days if we search these configurations exhaustively. In this paper, we propose a GA-NN technique to reduce the computation time. The NN (neural network) technique is used to emulate the func- tion of OPUS10 in a more efficient manner. The (GA) genetic algorithm technique is used to reduce the number of BOM configurations that are searched during the pro- cess of finding a near-optimal one. 4. Neural networks (NN) The procedure for establishing an NN for emulating the function of OPUS10 involves two major steps. First, we randomly sample n number of BOM configurations from the solution space and evaluate their performance at the specified machine availability (A) by using OPUS10. Refer- ring to Eq. (2), this sampling would yield n pairs of input/ output vectors, where V (BOM configuration) is the input and TSC (total operation cost) is the output. Second, from the n pairs of input/output vectors, n1 pairs are randomly sampled and used to train or develop a back-propagation neural network (Fausett, 1994). The remaining n � n1 pairs are used to test the effectiveness of the network. The trained network if effective can be used to emulate the function of OPUS10. The architecture of the back-propagation neural net- work involves three layers of neurons (Fig. 3). The first layer represents the input, the third layer represents the output, and the second layer (called hidden layer) tends to model the transformation mechanism from input to out- put. Each neuron in a layer and that in its subsequent layer is connected by a link on which a weight is to be found. Training or developing the back-propagation neural net- work is to determine appropriate weights of each link. The algorithm for training the NN iteratively changes the network weights by the following formula: wijkðt þ 1Þ¼ wijkðtÞþ g � wijkðtÞþ a � wijkðt � 1Þ ð3Þ where t denotes the index of change, i denotes a neuron in layer k, j denotes a neuron in the preceding layer (k � 1), and wijk represents the weight between the two neurons. Parameters g (learning constant) and a (momentum con- stant) are intended to adjust the speed of convergence. The detailed procedures for training the NN can be referred to Fausett (1994). The validity of the trained NN is evaluated by measur- ing the deviation between the predicted output of the NN and the actual output of OPUS10, in terms of the root- mean-square error (RMSE). During the network develop- ment process, the network architecture (number of neurons at each layer) and the training parameters (g and a) are iter- atively selected in order that the RMSE is minimized. 5. Using GA to find a near-optimal BOM configuration The solution space of BOM configuration may be quite huge. Though the proposed NN could efficiently evaluate the performance of a particular BOM configuration. However, exhaustively searching the space may still be computationally extensive. In order to reduce the computa- tion time, we proposed a GA (genetic algorithm) for effi- ciently finding a near-optimal BOM configuration from the solution space. GAs have been widely used in various M.-C. Wu, Y.-K. Hsu / Expert Systems with Applications 34 (2008) 2417–2423 2421 applications (Bäck, Hammel, & Schwefel, 1997; Stockton, Quinn, & Khalil, 2004a, Stockton, Quinn, & Khalil, 2004b) and found to be efficient and effective in solving a complex space-searching problem. Referring to Eq. (2), a BOM configuration that involves J parts and each part has K vendors on can be represented by V = (v1, v2, . . ., vJ), where 1 6 vi 6 K represents the ven- dor that supplies part i. In the GA, a BOM configuration V is called a chromosome and vi is a called a gene. The perfor- mance of a BOM configuration refers to the total operation cost (TSC) subject that the stocking policy is optimally determined by using the developed NN that emulates the function of OPUS10. The TSC of a chromosome V, com- puted by the NN rather than OPUS10, is represented by F(V), which is called the fitness of chromosome V. The procedure of the GA involves the following major steps. Step 1: Set t = 0. Randomly sample N chromosomes to form a population P(0). Step 2: Reproduction operation • S = / • Reproduce N chromosomes in P(t) and place them in set S. Step 3: Crossover operation • Create N · Pcr new chromosomes and place them in set S. Step 4: Mutation operation • Create N · Pmu new chromosomes and place them in set S. Step 5: Termination check • If a terminating condition is met, output the best chromosome from S and stop. • Otherwise, t = t + 1, P(t) S, go to Step 2. In the aforementioned GA procedure, the reproduction, crossover, and mutation operators as well as the terminat- ing conditions are further explained below. The reproduc- tion operator in Step 2 is intended to select ‘‘good’’ chromosomes from P(t) to form a new population; that is, a chromosome with higher fitness value has higher probability of being reproduced. We use the tournament Table 1 Failure rates and prices of critical parts provided by different vendors Critical part Quantity of each part per BOM Failure rates provided by each vendor FRT1 FRT2 A 3 54.75 33.46 B 3 59.59 31.21 C 3 102.61 98.43 D 20 116.81 55.86 E 3 114.68 97.69 F 3 95.31 88.48 G 3 52.75 38.36 H 24 37.82 33.37 I 6 127.58 77.31 J 6 104.25 91.08 selection method (Blickle, 1997, chap. C2.3: C2.3:1– C2.3:4; Goldberg, Korb, & Deb, 1989) in the reproduction process. This method randomly samples two chromo- somes, from which the one with higher fitness is selected and placed in set S. This selection procedure is repeated until N chromosomes have been chosen. The crossover operator in Step 3 is intended to create ‘‘new’’ chromosomes by using the single-point crossover technique (Booker, Fogel, Whitley, & Angeline, 1997, chap. C3.3:C3.3:1–C3.3:27; Goldberg, 1989; Spears, 1997, chap. E1.3:E1.3:1–E1.3:11). This technique firstly samples a pair of chromosomes randomly. Secondly, a cut-point is randomly chosen so that each chromosome is interpreted as having two segments. Thirdly, the left-hand segments of the two chromosomes are swapped to form a new chro- mosome, so does the right-hand segments. This as a result yields a new pair of chromosomes. The crossover operation is repeated until N · Pcr new chromosomes have been cre- ated, where Pcr is termed crossover rate whose value is manually given. The mutation operator in Step 4 is intended to create ‘‘new’’ chromosomes by changing the value of a particular gene. This operator firstly samples a chromosome ran- domly. Secondly, from the chromosome, one of its genes is randomly selected. Thirdly, the value of the gene (denot- ing a particular vendor) is replaced by randomly selecting a new vendor. The mutation operator is repeatedly per- formed until N · Pmu new chromosomes have been created, where Pmu is termed mutation rate. The GA procedure terminates either when the popula- tion has been updated Tf times (i.e., t = Tf) or when the best solution in P(t) keeps unchanged for Bf generations. 6. Numerical experiments The proposed GA-NN method for solving the decision problem of BOM configuration has been tested by numer- ical experiments. The tested scenario assumes the BOS/BOM structures as shown in Figs. 1 and 2 respectively. In the BOM, there are 10 critical parts (A–J), each of which has three vendors to choose. Table 1 shows the quantity of each part per BOM, (number of failures per 106 h) Unit price provided by each vendor FRT3 C1 C2 C3 30.42 $19,086 $31,231 $34,355 28.38 $20,385 $32,277 $38,733 60.36 $48,776 $75,040 $97,552 50.79 $34,672 $39,625 $55,476 67.46 $26,006 $29,474 $44,211 56.06 $35,026 $40,030 $56,042 23.98 $38,951 $50,936 $66,218 22.25 $38,465 $57,698 $69,238 60.28 $40,866 $70,588 $97,647 37.92 $14,261 $19,015 $28,523 Table 2 Comparing performance between BOM configurations A B C D E F G H I J TSC % of cost reduction GA 1 2 1 2 1 3 3 1 1 3 $9,304,599 Vendor 1 1 1 1 1 1 1 1 1 1 1 $11,243,894 16.13% Vendor 2 2 2 2 2 2 2 2 2 2 2 $10,906,230 14.58% Vendor 3 3 3 3 3 3 3 3 3 3 3 $10,553,935 11.82% 2422 M.-C. Wu, Y.-K. Hsu / Expert Systems with Applications 34 (2008) 2417–2423 the failure rates and the prices of each part offered by each vendor. As indicated in the table, the higher the price, the lower the failure rate. Assume that the desired machine availability (A) is 85%. Out of these 310 (59,049) BOM configurations, 1500 con- figurations are randomly sampled to train the neural net- work. Of these samples, 1200 samples are used to train the NN and the remaining 300 are used for testing. The NN is a 10-10-1 back-propagation neural network architec- ture trained by setting g = 0.10 and a = 0.80, where 10-10-1 denotes that 10 neurons in the input layer, 10 in the hidden layer, and 1 in the output layer. Experiment results indicate that the accuracy of the trained NN is RMSE = 0.00795. A personal computer equipped with 3.0 G CPU and 512 MB is used in the experiments. It takes about 5 s for OPUS10 to evaluate the performance of a BOM configura- tion. It takes about 1330 s to train the neural network. That is, the trained NN needs only about 10�4 s to evaluate the performance of a BOM configuration, about 50,000 times faster than OPUS10. Parameters of the GA are so defined: N = 100, Pcr = 0.80, Pmu = 0.05, Tf = 99,999, and Bf = 1000. The GA was executed with 50 replicates. Experiment results showed that the solutions of the 50 replicates are all the same. The computation time for executing the GA with one replicate takes about 1 s. Table 2 compares the performance of four BOM config- urations. The first row in the table is the solution proposed by the GA. Each of the remaining three rows denotes a BOM configuration whose parts are exclusively supplied by a single vendor, rather than multiple ones. The table shows that the BOM configuration proposed by the GA outperforms the other three alternatives, reducing the cost up to 16.13%. 7. Concluding remarks This paper proposes a new perspective to reduce the total operational cost of a spare part logistic system through appropriately selecting the BOM configuration. Most pre- vious studies on spare part logistics assumed that the BOM configuration is fixed and aimed to determine the optimal inventory levels. However, each part in a BOM may have several options in terms of failure rates. A part option with lower failure rate is generally more expen- sive. Therefore, a BOM configuration with lower failure rates tends to require less quantity of inventory, at the price of incurring higher unit inventory cost. This research enhances previous works by relaxing the assumption of fixed BOM configuration in order to further reduce the total operational cost of a spare part logistic system. The proposed solution method involves the use of OPUS10 as well as the development of a back-propagation neural network and a genetic algorithm. Given a BOM configuration, OPUS10, (proprietary software) can be used determine the optimal inventory levels for a spare part logistic system. However, using OPUS10 to exhaustively search a huge solution space of BOM configurations may be quite time-consuming. The time-consuming issue is resolved in twofold. First, through the development of an artificial neural network, we can reduce the computation time for estimating the performance of a BOM configura- tion. Second, through the development of a GA, we can greatly reduce the number of BOM configurations to be evaluated. Numerical experiments indicate that selecting an appropriate BOM configuration could significantly reduce the total operational cost. Other than redesigning BOM configuration, we may fur- ther reduce the total operational cost of a spare part logistic system by redesigning the BOS (bill of stations) and the asso- ciated transportation configurations. Possible extensions of this research involve how to develop an integrated method to comprehensively design a spare part logistic system; that is, such a design should consider simultaneously the effects of BOM, BOS, and transportation configurations. Acknowledgement This research was financially supported by National Sci- ence Council, Taiwan, under a research contract NSC94- 2623-7009-004. References Albright, S. C. (1989). An approximation to the stationary distribution of a multiechelon repairable-item inventory system with finite sources and repair channels. Naval Research Logistics Quarterly, 36, 179– 195. Albright, S. C., & Soni, A. (1988a). Markovian multiechelon repairable inventory system. Naval Research Logistics Quarterly, 35, 49–61. Albright, S. C., & Soni, A. (1988b). An approximation to the stationary distribution of a multidimensional Markov process. IIE Transactions, 20, 111–118. Axsäter, S. (2003). Evaluation unidirectional lateral transshipments and substitutions in inventory systems. European Journal of Operational Research, 149, 438–447. Bäck, T., Hammel, U., & Schwefel, H. P. (1997). Evolutionary compu- tation: Comments on the history and current state. IEEE Transactions on Evolutionary Computation, 1, 3–17. M.-C. Wu, Y.-K. Hsu / Expert Systems with Applications 34 (2008) 2417–2423 2423 Blickle, T. (1997). Tournament selection. In T. Bäck, D. B. Fogel, & Z. Michalewicz (Eds.), The handbook of evolutionary computation. Philadelphia, PA: IOP Publishing Ltd. and Oxford University Press. Booker, L. B., Fogel, D. B., Whitley, D., & Angeline, P. J. (1997). Recombination. In T. Bäck, D. B. Fogel, & Z. Michalewicz (Eds.), The handbook of evolutionary computation. Philadelphia, PA: IOP Publish- ing Ltd. and Oxford University Press. Diaz, A., & Fu, M. C. (1997). Models for multi-echelon repairable item inventory systems with limited repair capacity. European Journal of Operational Research, 97, 480–492. Diaz, A., & Fu, M. C. (2005). Multi-echelon models for repairable items: A review. Decision & Information Technologies Research Works Collection, Digital Repository at the University of Maryland. Fausett, L. (1994). Fundamental of neural networks: Architectures, algo- rithms, and applications. New Jersey: Prentice-Hall. Goldberg, D. E. (1989). Genetic algorithms in search optimization and machine learning. Reading, MA: Addison-Wesley. Goldberg, D. E., Korb, G., & Deb, K. (1989). Messy genetic algorithms: Motivation, analysis, and first results. Complex Systems, 3, 493–530. Graves, S. C. (1985). A multi-echelon inventory model for a repairable item with one-for-one replenishment. Management Science, 31(10), 1247–1256. Gross, D., Kioussis, L. C., & Miller, D. R. (1987). A network decomposition approach for approximating the steady-state behavior of Markovian multi-echelon repairable item inventory systems. Man- agement Science, 33, 1453–1468. Gross, D., & Miller, D. R. (1984). Multiechelon repairable-item provi- sioning in a time-varying environment using the randomization technique. Naval Research Logistics Quarterly, 31, 347–361. Gross, D., Miller, D. R., & Soland, R. M. (1983). A closed queueing network model for multi-echelon repairable item provisioning. IIE Transaction, 15, 344–352. Guide, V. D. R., Jr., & Srivastava, R. (1997). Reparable inventory theory: Models and applications. European Journal of Operational Research, 102, 1–20. Hausman, W. H., & Erkip, N. K. (1994). Multi-echelon vs. single-echelon inventory control policies for low-demand items. Management Science, 40(5), 597–602. Kennedy, W. J., Patterson, J. W., & Fredendall, L. D. (2002). An overview of recent literature on spare parts inventories. International Journal of Production Economics, 76, 201–215. Kim, J. S., Shin, K. C., & Park, S. K. (2000). An optimal algorithm for repairable-item inventory system with depot spares. Journal of the Operational Research Society, 51, 350–357. Lee, H. L. (1987). A multi-echelon inventory model for repairable items with emergency lateral transshipments. Management Science, 33(10), 1302–1316. Muckstadt, J. A. (1973). A model for a multi-item, multi-echelon, multi- indenture inventory system. Management Science, 20(4), 472–481. Rustenburg, W. D. (2000). A system approach to budget-constrained spare parts management. Ph.D. thesis. Enschede, Netherlands: Eind- hoven University of Technology. Rustenburg, W. D., van Houtum, G. J., & Zum, W. H. M. (2001). Spare parts management at complex technology-based organizations: An agenda for research. International Journal of Production Economics, 71, 177–193. Sherbrooke, C. C. (1968). METRIC: A multi-echelon technique for recoverable item control. Operational Research, 16, 122–141. Sherbrooke, C. C. (1986). VARI-METRIC: Improved approximation for multi-indenture, multi-echelon availability models. Operations Research, 34(2), 311–319. Sherbrooke, C. C. (2004). Optimal inventory modeling of systems: Multi- echelon techniques. MA: Kluwer Academic Publishers. Simon, R. M. (1971). Stationary properties of a two-echelon inventory model for low demand items. Operations Research, 19(3), 761–773. Slay, F. M. (1984). VARI-METRIC: An Approach to modeling multi- echelon resupply when the demand process is poisson with gamma prior, Technical Report AF301-3, Logistic Management Institute, Washington, DC. Sleptchenko, A. (2002). Integral inventory control in spare parts networks with capacity restrictions. Ph.D. thesis. Enschede, Netherlands: University of Twente. Sleptchenko, A., van der Heijden, M. C., & van Harten, A. (2002). Effects of finite repair capacity in multi-echelon, multi-indenture service part supply systems. International Journal of Production Economics, 79, 209–230. Spears, W. (1997). Recombination parameters. In T. Bäck, D. B. Fogel, & Z. Michalewicz (Eds.), The handbook of evolutionary computation. Philadelphia, PA: IOP Publishing Ltd. and Oxford University Press. Stockton, D. J., Quinn, L., & Khalil, R. A. (2004a). Use of genetic algorithms in operations management. Part 1: Applications. ProQuest Science Journals, 3, 315–327. Stockton, D. J., Quinn, L., & Khalil, R. A. (2004b). Use of genetic algorithms in operations management. Part 2: Results. ProQuest Science Journals, 3, 329–343. Wong, H., Cattrysse, D., & van Oudheusden, D. (2005). Inventory pooling of repairable spare parts with non-zero lateral transshipment time and delayed lateral transshipments. European Journal of Opera- tional Research, 165, 207–218. Design of BOM configuration for reducing spare parts logistic costs Introduction Related literature OPUS10 BOS BOM Input/output relationship of OPUS10 Neural networks (NN) Using GA to find a near-optimal BOM configuration Numerical experiments Concluding remarks Acknowledgement References 
work_vzc6tioriremjmtbvgl42jii34	----	Variance enhanced K-medoid clustering Expert Systems with Applications 38 (2011) 764–775 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Variance enhanced K-medoid clustering q Por-Shen Lai *, Hsin-Chia Fu Department of Computer Science, National Chiao-Tung University, Hsin-Chu 300, Taiwan, ROC a r t i c l e i n f o Keywords: Clustering K-medoid Data variance Polygon model Image similarity 0957-4174/$ - see front matter � 2010 Elsevier Ltd. A doi:10.1016/j.eswa.2010.07.030 q This research was supported in part by the Nat Grant NSC 95-2221-E-009-218. * Corresponding author. E-mail addresses: porshenlai@yahoo.com (P.-S. (H.-C. Fu). a b s t r a c t This paper proposes new variance enhanced clustering methods to improve the popular K-medoid algo- rithm by adapting variance information in data clustering. Since measuring similarity between data objects is simpler than mapping data objects to data points in feature space, these pairwise similarity based clustering algorithms can greatly reduce the difficulty in developing clustering based pattern rec- ognition applications. A web-based image clustering system has been developed to demonstrate and show the clustering power and significance of the proposed methods. Synthetic numerical data and real-world image collection are applied to evaluate the performance of the proposed methods on the pro- totype system. As shown as the web-demonstration, the proposed method, variance enhanced K-medoid model, groups similar images in clusters with various variances according to the distribution of image similarity values. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction Clustering techniques have been widely used in many applica- tions, such as pattern recognition, data mining (Pao, Chen, Lai, Xu, & Fu, 2008), user interface design, and so on. Using clustering methods (Fu et al., 2000; Haykin, 1994) to approximate the data distribution of target objects (pattern) in feature space, the ob- tained data clusters can be used to recognize target objects (pat- tern) by checking whether a new object is in a cluster or not. Besides, clustering methods (Guyon & Elisseeff, 2003; Schapire & Singer, 1999) are also useful in finding patterns of labeled data groups for data mining application. Using labeled training objects, supervised clustering methods (Haykin, 1994) learn rules to parti- tion data objects into clusters (Pao, Chuang, Xu, & Fu, 2008). On the other hand, unsupervised clustering methods (Comaniciu & Meer, 2002; Hastie, Tibshirani, & Friedman, 2001b, chap. 14.3) partition feature space to distribute data objects into disconnected groups. Generally, most clustering methods apply the feature extraction techniques to extract feature vectors, which usually represent the original data objects in numerical form to represent data objects as data points in feature space. Then, data points are partitioned into clusters according to the distribution in feature space. There- fore, the performance of data clustering is highly depending on the quality of extracted features. ll rights reserved. ional Science Council under Lai), hcfu@csie.nctu.edu.tw K-means (Hastie, Tibshirani, & Friedman, 2001c, chap. 14.3.6), a popular clustering algorithm, groups data objects into clusters according to the norm between extracted feature vectors. In other words, using K-means clustering method, data objects in a cluster are represented by the mean of feature vectors of data points, the cluster with the closest mean. Since K-means algorithm does not use variance information to cluster data points, thus K-means model is inappropriate to cluster data points or objects having different variances in data distribution. Mixture Gaussian model (Bilmes, 1997; Hastie, Tibshirani, & Friedman, 2001d, chap. 14.3.7) represents a data cluster by a mean vector and a covariance matrix, which, respectively, illustrate the center location and the variance of data distribution in each orientation. Therefore, the mixture Gaussian model is capable to cluster data points of non- uniform variances. Representing data objects by their feature vectors make the de- sign of clustering method simpler. However, extracting representa- tive feature vectors for data objects is still difficulty. For instance, although the similarity between two images can be measured based on some visual features, such as color, texture, or spatial dis- tribution of pixels, creating feature vectors to represent the pair- wise similarities between images by the norm between two feature vectors is still difficult. Therefore, several methods (Hastie, Tibshirani, & Friedman, 2001a, chap. 14.3.10) are proposed to clus- ter data objects using pairwise similarities between data objects instead of measuring the distance between extracted feature vec- tors of data objects. K-medoid algorithm (Hastie et al., 2001a) has been widely used to cluster data points according to their pairwise similarity values. Given similarity values between every pair of data objects, an http://dx.doi.org/10.1016/j.eswa.2010.07.030 mailto:porshenlai@yahoo.com mailto:hcfu@csie.nctu.edu.tw http://dx.doi.org/10.1016/j.eswa.2010.07.030 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 765 iterative learning process can group data objects into clusters to maximize the sum of similarity values for each cluster. That is, when the sum of similarity values is maximized, K-medoid algo- rithm groups data objects of equal to or lesser than similarity value in one cluster. The decision boundaries that are generated by a given K-medoid model are the perpendicular bisector hyperplane of the line segment from the medoid of one cluster to another. That is, the variance of data distribution in each cluster is assumed to be uniform. However, the variances of different orientations of the data distribution in a cluster may be different. In this paper, the concept of different vari- ances is suggested to be included in the original K-medoid model. Therefore, a new variance enhanced K-medoid model is proposed to group data objects in clusters with variant variances. In addition, in order to use the new K-medoid model to cluster similar images, a similarity measurement between two images is also proposed and briefed as follows. First, color information is used to segment image into regions of similar color. Then, the similarity between two images is measured by computing the bitmap difference in coverage region. Based on the proposed methods, a web-based prototype system (The demonstration of variance enhanced K-medoid model, http://www.csie.nctu.edu.tw/�pslai/ ASKMDDemo), which displays variable-sized thumbnails of images, is developed and tested. The prototype system groups similar images in a cluster, and the image at the cluster center (medoid) is shown in larger size to depict each group. And the other similar images in one group are displayed together and shown in smaller size to save display space in screen. This paper is organized as follows. The K-medoid algorithm is introduced in Section 2. Then, the variance enhanced K-medoid clustering is proposed and presented in Section 3. Section 4 intro- duces the proposed methods for the similarity measurement of a image collection. The web-based prototype system is then demon- strated and evaluated in Section 5. Finally, concluding remarks are drawn in Section 6. 2. K-medoid Given a data point set P, the medoid of P is the point pm with the smallest sum of distance from pm to all the other points pi in P. Medoid pm of P is mathematically defined as follows: pm ¼ arg min pi2P X pj2P Nðpi; pjÞ; where Nðpi; pjÞ is the norm between two points pi and pj. For one-dimensional data, medoid of P is actually the median. For a data point set P, let Q(x) be the sum of distance from point x to all the other points pi in P, Q(x) can be formulated as follows: QðxÞ¼ X pi2P Nðpi; xÞ: Let md be the medoid of P, and pi be another data point in P. Accord- ing to the definition of medoid, Q(md) � Q(pi) must be smaller or equal to zero. Fig. 1. The relation between medoid and median for one-dimensional data points. Let m be median and md be medoid. Since medoid md is the minimal of Q(md), medoid md is equal to median m. Fig. 1 illustrates the relation between median and medoid. Let m be the median, and md be the medoid of P. For each point pi in P, we may observe the followings. � When pi is located between median m and medoid md, then Nðpi; mdÞ¼Nðm; mdÞ�Nðm; piÞ. � When median m is located between pi and medoid md; Nðpi; mdÞ¼Nðpi; mÞþNðm; mdÞ. � When medoid md is located between median m and pi; Nðmd; piÞ¼Nðm; piÞ�Nðm; mdÞ. Q(md) � Q(m) can be rewritten as follows: QðmdÞ� QðmÞ¼ X p2ð�1;m� Nðm; mdÞ� X p2½md;1Þ Nðm; mdÞ þ X p2ðm;mdÞ Nðp; mdÞ�Nðm; pÞ ¼ X p2ð�1;m� Nðm; mdÞ� X p2½md;1Þ Nðm; mdÞ þ X p2ðm;mdÞ Nðp; mdÞ�Nðm; mdÞþNðp; mdÞð Þ ¼ X p2ð�1;m� Nðm; mdÞ� X p2ðm;1Þ Nðm; mdÞ þ 2 X p2ðm;mdÞ Nðp; mdÞ ¼Nðm; mdÞþ 2 X p2ðm;mdÞ Nðp; mdÞ Since Q(md) � Q(m) P 0, and Q(md) is the minima (the definition of medoid), the medoid md must be equal to median m. The same re- sults hold when md is smaller or equal to m. Therefore, for one- dimensional data, medoid is equal to median. Intuitively, the medoid has higher chance to be located in the high density region in a data set than the mean can be. Therefore, medoid may be a better representation of a data set. Similar to the popular K-means algorithm, K-medoid (Hastie et al., 2001a) can be used to cluster data points into K groups. Un- like K-means algorithm, K-medoid does not depend on the coordi- nates or values of each data points in feature space, the data required for medoid estimation is the norm between every pair of data elements. Suppose data points in P = {p1, . . ., pi, . . ., pN} are partitioned into K clusters. By assigning each data point pi to the cluster with min- imal norm between pi and the medoid of each cluster, a data set P can be iteratively partitioned into K clusters. 2.1. The learning algorithm of K-medoid clustering The K-medoid cluster learning algorithm (Hastie et al., 2001a) is briefly stated as follows. Let Nðpi; pjÞ be the norm between pi and pj in data set P, and K is the number of medoid. Step 1. Given a current set of cluster centers Mt ¼ mt1; . . . ; mtk � � , the current cluster assignments Ct(pi) of each data points pi 2 P are evaluated by computing the closest cluster cen- ter for each data point pi as follows: CtðpiÞ¼ arg min mt j 2Mt Nðmtj ; piÞ; where Nðmj; piÞ is the norm between mj and pi, and C t(pi) is the current cluster assignment for pi. Step 2. Given a set of current cluster assignments Ct, the new clus- ter centers Mtþ1 ¼ mtþ11 ; . . . ; m tþ1 j ; . . . ; m tþ1 k n o are estimated by computing the medoid among the data points in the same assignment as follows: Fig. 2. A K-medoid represent cluster ce cluster ce weight v location c (the dis-s points are 766 P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 mtþ1j ¼ arg min CtðpxÞ¼mtj X CtðpyÞ¼mtj Nðpx; pyÞ; where Ct(pi) is the cluster assignment for pi. Step 3. Iterate steps 1 and 2 until the assignments do not unchange. 2.2. Self-growing K-medoid Generally speaking, asking user to assign the number of re- quired clusters is not practical since the number of clusters is usu- ally the natural of data distribution and is often unknown to a user. In order to perform data clustering without knowing cluster number, the self-growing K-medoid clustering algorithm is pro- posed. By giving the maximal acceptable dis-similarity between a data point and its cluster center, the proposed method iteratively increases the number of clusters until the dis-similarity between any data point, and the associated cluster center is lower than a gi- ven threshold. The basic concepts of the proposed method can be illustrated by the example shown in Fig. 2. The y-axis shows the value of weight for each points, and the x-axis is the location of data points. Based on K-medoid clustering algorithm, the point p3 is first selected to be the first cluster center c1. The weights of each data point are up- dated according to their norms with respect to the selected cluster center. Specifically, the weights of data points that are close to the selected cluster center will be decreased more than the weights of data points that are far away from the selected cluster center. Then, another data point locates in the middle of data points with larger weights is selected as the second cluster center c2. After clustering all data points with the two selected cluster centers, the weights of data points are updated. The clustering processes are repeated to decrease the weights of data points until all weights are lower than to a given threshold. To formulate this idea, a weighted medoid estimation is defined as follows: 1-D data distribution example is used to illustrate the idea of self-growing Clustering algorithm. In each diagram, the heights (w) of a sample points s its weight values. (1) At first, the middle point p3 is selected as the first nter c1. (2) By decreasing the weights of data points close to the first nter c1, point p5, which is the middle point of data points with large alues, is selected to be the new cluster center c2. (3) After refining the luster centers by standard K-medoid clustering algorithm, the weights imilarity between data point and the closest cluster center) of all data less than a given threshold. pn ¼ arg min pj2P X pi2P wj wiNðpj; piÞ ! : The weighted medoid estimation organizes a new cluster center for data points with high weight values. Combining the weighted medoid estimation and the standard K-medoid algorithm, the self-growing K-medoid method algorithm is introduced as follows: Step 1. Initialize weight values wi of each data point pi to be 1. Step 2. Use the weighted medoid estimation to create a new clus- ter center. Step 3. Apply the standard K-medoid clustering algorithm to adjust the location of clusters centers pm. Step 4. Recompute the weights wi of each data points pi as follows: a Fig. 3. Th model. Sh boundary centers. T single var between from clus decision b between wi ¼ 1 � 1 þ min pm2M Nðpi; pmÞ � ��1 : Step 5. Iteratively repeat steps 2–4 until all weight values wi are smaller than a given threshold. Using the proposed self-growing K-medoid algorithm, the num- ber of required model (medoid) can be determined to let the dis- similarity between data point and closest cluster center be less than a given threshold. 2.3. Discussion Based on the mathematical representation of two clusters, a decision boundary can be determined to separate data points into two parts. As shown in Fig. 3(a), the decision boundary between two K-medoid clusters is the perpendicular bisectors of a line seg- ment from one cluster center to another. Since the data variance is not considered, these decision boundaries may not separate data clusters properly. To achieve a better separation of the data points, variance values should be considered in determining the decision b c e comparison among three kinds of decision boundary for clustering aded area are the distributed regions of data points. (a) The decision is the perpendicular bisector of the line segment between two cluster he variance of data distribution is not used in creating data clusters. (b) A iance for every direction is used for each cluster. For two cluster, the ratio the two variances of clusters is equal to the ratio between the distances ter centers to decision boundary. (c) For each two clusters, the location of oundary is decided according to the data variances along the line segment two cluster centers. P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 767 boundaries. An variance enhanced K-medoid is illustrated in Fig. 3(b). By aggregating the norm between data points and its cluster center (medoid), the variance of each cluster is estimated Again, since the variance orientation of data distribution is still not considered, and the decision boundaries may not partition the feature space well enough. For instance, without considering variance orientation, a cluster with widely data distributed in one direction and with vary little data distributed in a perpendic- ular direction may result a large overall variance along the perpen- dicular direction. Fig. 3(c) shows the decision boundaries in corresponding to data cluster with multiple variances being con- sidered along several orientations. Based on the location of data points in feature space, a K-means model can be improved by including the covariance matrix for each cluster, e.g., Gaussian mixture model. Since the location information of data points is not used in K-medoid model, improving the K-medoid model using the covariance matrix method seems impracticable. How to in- clude the oriented variance in K-medoid method and how to esti- mate the variances in various orientations become essential issues for variance enhanced K-medoid clustering algorithm. The detailed description of this method will be illustrated in Section 3. 3. Variance enhanced K-medoid clustering According to the discussion in Section 2.3, multiple variances along several orientations are required for K-medoid based cluster- ing method. However, since the locations of data point in feature space are not available, estimating variance orientation using covariance matrix seems impractical. Therefore, multiple variances along line segments between cluster centers are proposed instead. Let us see that the variance along a line segment between cluster centers should be measured according to the data points located along the line segment. As the example shown in Fig. 4, four cluster centers M1, M2, M3, and M4 are included, and bi’s are the decision boundaries between the line segments from M1 to Mi for i = 2 to 4. For instance, the variance along M1M3 associated with cluster center M1 is estimated by the data points located in the region of gray-scaled color. In general, decision boundaries segment the fea- ture space into polygon-shaped regions for each cluster in a data set. And, the polygon-shaped region for a cluster can be further di- vided into several smaller pyramid-shape segments, whose tops Fig. 4. Giving four clusters with centers at M1, M2, M3, and M4. For the center M1, there are three associated decision boundaries b2, b3, and b4. The variance between M1 and M3 is computed using the data points inside the pyramid-shape segment whose top is located at the cluster center M1, and the bottom edge is along with decision boundary b3. are the cluster centers and bottom hyperplane are along the deci- sion boundaries. The variance along the line segment M1 M3 is mea- sured using data points located in the pyramid-shape segment, whose bottom edge is along the decision boundary b3 which parti- tions two data clusters with centers at M1 and M3, respectively. To select appropriate data points for the computation of vari- ance along line segments between cluster centers, a polygon- shaped data model, called Polygon descriptor (Lai & Fu, 2007, 2010), is suggested and briefly described in Section 3.1. In the rest of this section, Polygon descriptor is introduced in Section 3.1. Then, the method that segments each cluster into sub-clusters, called side-clusters, along each line segment between cluster centers is depicted in Section 3.2. Using the projection method described in Section 3.3, the projection distribution of data points in each side cluster can be used to adjust decision bound- aries between every pair clusters in the data points by the pro- posed method in Section 3.4. Finally, the variance enhanced K- medoid clustering method is presented in Section 3.5. 3.1. Polygon-shaped data model As shown as Fig. 3, decision boundaries segment feature space into several polygon-shaped regions (data clusters). Polygon descriptor (Lai & Fu, 2007, 2010) was originally proposed for mod- eling a polygon-shaped data distribution. A Polygon descriptor uses a reference center and N normal vectors to describe a polygon-shaped data distribution. A normal vector represents: (1) the normal direction of a hyperplane that encloses the convex unit and (2) the distance from the reference center to each hyperplane. Fig. 5(a) shows an exemplar Polygon descriptor for the representation of a 2D data distribution in a convex unit. The reference center is located at (20, 20), and five normal vectors are 10 0 � � ; 5 5 � � ; �5 5 � � ; �10 0 � � ; and 0 �10 � � . By applying a 1-D probability function to each normal vectors, a polygon-shaped probability model is created as shown in Fig. 5(b). The polygon- shaped probability model (distribution) shown in Fig. 5 is centered at (20, 20), and the variance corresponds to each normal vector is the length of each normal vectors, respectively. According to the Polygon descriptor, one center point and M normal vectors are used to describe a M-side polygon. The center point is used as the reference origin of the coordinate system of a Polygon descriptor. And, M normal vectors are used to span the polygon-shaped coverage area of data points. Since normal vectors are orthogonal to the boundaries of a data distribution, the orien- tation of the boundaries of a data distribution can be represented by their normal vectors. Besides, the length of a normal vector indi- cates the data distribution variance along the normal vector. Fig. 6 shows the components in a polygon descriptor. Data points are partitioned into clusters, called side-cluster, according to the length and orientation of each normal vectors. As shown in Fig. 6, for a data distribution containing M normal vectors, M side-clusters are determined according to the M boundary edges. The side-clusters corresponding to each boundary edges can be used to measure the variance of data points along the correspond- ing normal vectors. 3.2. Side-cluster segmentation M normal vectors of a Polygon descriptor can be used to further segment a data cluster into M side-clusters. Let ~a1; . . . ; ~aj; . . . ; ~aM � � be the M normal vectors of a Polygon descriptor. Given a set P of data points pi, a sub-cluster assignment C(pi) for a data points pi can be formulated as follows: a b Fig. 5. (a) An example of a proposed polygon descriptor, where its center is at (20,20) and normal vectors (drawn as solid arrow) to each surrounding hyperplane are (10, 0)T, (5, 5)T, (�5, 5)T, (�10, 0)T, and (0,�10)T. (b) The pentagon-shaped probability distribution for the polygon descriptor of (a). Fig. 6. A diagram shows the components of an exemplar polygon descriptor. One center point C is used as the reference origin of the coordinate system of a Polygon descriptor. M(= 3) normal vectors are emitted from the origin to M directions. According to the length and orientation of M normal vectors, M side-clusters are determined corresponding to each boundary edges. That is, the clustering of a side- cluster is determined by the length and orientation of normal vectors. The length and orientation of normal vectors can be estimated according to the statistical properties of data points in each clusters. 768 P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 CðpiÞ :¼ arg max M j¼1 ~pi � ~aj ~aj � ~aj ; Fig. 7. Let ~a1 and ~a2 are normal vectors of side-clusters C1 and C2, respectively. The dash line shows the decision boundary between C1 and C2. For data vector ~p1 which is from cluster center O to data point p1 in C1, its projection ratio on ~a1 is larger than the projection ration on ~a2 , i.e., ~p1 � ~a1ð Þ � ~a1 �� ��2 > ~p1 � ~a2ð Þ� ~a2�� ��2 . Similarly, the projection ratio of the data vector ~p2 in C2 on ~a2 is larger than the projection ratio on ~a1 , i.e., ~p2 � ~a1ð Þ � ~a1 �� ��2 < ~p2 � ~a2ð Þ� ~a2�� ��2 . where C(pi) is cluster assignment for pi, and ~pi is the vector from ref- erence center to pi. According to the cluster assignments, there are M side-clusters, and how data points are partitioned into side clus- ters will be described in the following paragraph. Fig. 7 illustrates the basic concept of how a side-cluster is formed. Let ~a1 and ~a2 be normal vectors of boundary edges b1 adn b2, respectively. The decision boundary between two side- clusters C1 and C2 is drawn in dot line, which is a hyperplane pass- ing through the origin point O and the intersection point I of the two boundary edges b1 and b2. For a vector ~p1 from reference cen- ter to a point p1 in C1, its projection ratio on ~a1 is larger than the projection ratio on ~a2 . That is, P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 769 ~p1 � ~a1 ~a1 �� ��2 > ~p1 � ~a2 ~a2 �� ��2 : By segmenting the data points of a cluster into side clusters corre- sponding to the line segments between cluster centers and by pro- jecting the data points in a side cluster onto the corresponding line segment between cluster centers, an 1-D data distribution can be created. The 1-D data distribution can be used to measure the var- iance along the line segment between cluster centers or adjust the location of decision boundary between clusters. However, since the coordinates of data points for K-medoid model is absent, a new method is proposed in Section 3.3 to estimate the projection length without knowing the coordinate of data points. 3.3. Projections on the center-to-center link Computing variance along normal vector using projection ratio requires the coordinate values of data points in feature space. In or- der to computing the projection ratio on normal vectors without knowing the coordinates of data points in feature space, a new pro- jection method is proposed. As shown in Fig. 8, given a point p, and two clusters with center points at m1 and m2. m1pi !��� ���; m2pi! ��� ���; and m1 m2!��� ��� are the norms from p to m1, p to m2, and m1 to m2, respectively. Let d be the pro- jection of p on m1; m2 ! . m1 d ! ���� ���� is the projection length of m1 pi! on m1m2 ! . As shown in Fig. 8, the following relation holds: m1pi !��� ���2 ¼ m1d! ���� ���� 2 þ pi d ! ���� ���� 2 ; m2pi !��� ���2 ¼ m1 m2!��� ���� m1 d! ���� ���� � �2 þ pid ! ���� ���� 2 : Then, the length of m1 d ! ���� ���� can be derived as follows: m1m2 !��� ���� m1 d! ���� ���� � �2 ¼ m2pi !��� ���2 � m1pi! ��� ���2 þ m1d! ���� ���� 2 ; m1m2 !��� ���2 � 2 m1 m2!��� ��� m1d! ���� ���� ¼ m2pi! ��� ���2 � m1 pi! ��� ���2; m1d ! ���� ���� ¼ m1m2 !��� ���2 � m2 pi! ��� ���2 þ m1 pi! ��� ���2 2 m1m2 !��� ��� : By projecting data points onto the line segment between two clus- ter centers, the 1-D data distribution along the line segment be- comes available. Based on the 1-D data distribution, the decision boundary can therefore be adjusted to maximize the decision mar- gin between two estimated clusters. 3.4. Decision boundaries adjustment Using the projection method proposed in Section 3.3, data points in a cluster can be segmented into several side-clusters cor- Fig. 8. m1 and m2 are two cluster centers. p is a data point. d is the projection point on m1 m2 . responding to the line segments between each pair of cluster cen- ters. The projection of data points in a side-cluster on the corresponding line segment between cluster centers forms an 1-D data distribution. Based on the 1-D data distribution, the var- iance of data points along the line segment between cluster centers can be measured. Since the purpose of measuring variance is to ad- just the location of decision boundary, the estimation or the mea- suring of variance can be avoid by directly adjusting the location of decision boundary according to projected 1-D data distribution. As shown in Fig. 9, given two cluster centers m1 and m2. By pro- jecting data points pi (shown in black square dots) onto the line segment m1 m2 , a series of 1-D location values (shown in black tri- angle dots) are measured. The following iterative procedure is pro- posed to find a decision boundary, such that the distance between the means of two partitioned clusters can be maximized. These 1-D location values are partitioned into two groups namely g1 and g2 according to the decision boundary b12. The decision boundary b12 is determined to maximize the distance between l1 and l2, where l1 and l2 are the means of data groups g1 and g2, respectively. Let a be the target location of decision boundary, and the pro- jection ratio h(pi) for point pi is defined as hðpiÞ¼ m1pi ! �m1 m2 ! m1m2 ! �m1m2 ! : Actually, h(pi) is the projection length of vector m1pi ! on m1m2 divid- ing by the distance between two cluster centers. Then, a can be evaluated according to the following formula: a ¼ arg max a P hðpiÞ>a hðpiÞP hðpiÞ>a 1 � P hðpiÞ<a hðpiÞP hðpiÞ<a 1 ! ; where P hðpiÞ>a hðpiÞP hðpiÞ>a 1 is the average projection ratio that is larger than a, and P hðpiÞ<a hðpiÞP hðpiÞ<a 1 is the average projection ratio that is smaller than a. The detail processes of estimating the decision boundary be- tween two clusters are described as follows. Data points are clus- tered using the K-medoid method proposed in Section 2.2. For every pair of clusters Ci and Cj, for j > i, data points in Ci and Cj are projected onto mimj ! , where mi and mj are medoids of Ci and Cj, respectively. Then, the line segment between Ci and Cj is parti- tioned into 10 intervals. By counting the number of points locating in one interval, a histogram is drawn. Based on the histogram, the decision boundary can be properly located so that the distance be- tween two means of data sets, which are separated by the decision boundary, is maximized. All data points are then redistributed using the adjusted decision boundaries. The process repeats until the location of decision boundaries converged, such that the dis- tance between l1 and l2 is maximized. Fig. 9. Data points pi are projected to the line segment between cluster centers m1 and m2. The decision boundary is located at the place which partition the projected data points (triangle points) into two groups with farthest mean location l1 and l2. Fig. 10. The flow chart of variance enhanced K-medoid clustering. At first, general K-medoid algorithm is applied. According to the resulted clusters of general K-medoid algorithm, the cluster-to-cluster variance is estimated. Then, the weights for each data points are updated and new cluster is estimated based on the weighted data points. These steps are repeated until all weights are smaller than a given threshold q. 770 P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 3.5. Variance enhanced K-medoid clustering method The flow chart of variance enhanced K-medoid clustering is shown in Fig. 10. At first, one medoid is estimated. Then, the num- ber of medoid is gradually increased until the distance from each data point to the closest cluster center is smaller than a given threshold. For each medoid, its positions are estimated using gen- eral K-medoid algorithm first. Then, the 1-D data distributions can be established by projecting data points in side-clusters to the cor- responding line segment between cluster centers. According to these 1-D data distributions, the location of decision boundaries is adjusted to redistributed data points to each clusters. After the position of medoid converged, weights wi of each points pi is calcu- lated as the following formula: wi ¼ 1 � 1 þ min pm2M simðpi; pmÞ � ��1 ; where M is a set of estimated medoids, and sim (pi, pm) is the pro- jected similarity defined as follows: simðpi; pmÞ¼ arg max pj2M;pj – pm ~pi � ~pj � ~pm � a ~pj � ~pm � � a ~pj � ~pm � ; where a ~pj � ~pm � is the vector from pm to the decision boundary be- tween the clusters associated to pm and pj. Based on these weights, a new medoid is created to include high weighted data points. By repeating these processes, the number of medoid gradually increas- ing until the maximal weight is lower than a given threshold. Using the proposed variance enhanced K-medoid, the location of decision boundaries between each pair of cluster centers is re- lated to the data distribution along a line segment between cluster centers now. A real-world application, photo album preview with variable-sized thumbnails, is proposed and implemented in Sec- tion 4 to illustrate the performance of the proposed variance en- hanced K-medoid clustering method. 4. Real-world application In this section, variance enhanced K-medoid is used to intelli- gently create image thumbnails, where ‘‘intelligently” means that the size and the display order of image thumbnail are decided without human intervention. Generally, image gallery (Pao et al., 2008) uses thumbnail to provide a quick preview for each photo (Pao et al., 2008). However, even though the thumbnail is down- sized from the original photos, the display area is often still too small to contain all thumbnails. To contain all thumbnails of same size can be very difficult in selecting a proper size for all the thumbnails. Sometimes, it may be too crowded to clearly represent the original image, or it may be too large to put all thumbnails in a page. Therefore, variable-sized thumbnail seems a better alterna- tive to fit all thumbnails into a page without losing too much details. Since most of images in an image gallery may be similar to each other, these images can be partitioned into groups according to their similarity. For each group, a most representative image can be displayed by a normal sized thumbnail, and the rest of similar images can be shown by smaller thumbnails. The detail description of the proposed system is depicted in the following sections. Color information of an image is used to create four bitmaps to show the distribution of dominant color compo- nents. Based on these bitmaps, the similarity between two images is measured. Then, the proposed variance enhanced K-medoid is applied to cluster all images into groups. 4.1. Color clustering Giving an image I, K’s dominant color components are selected to establish K’s bitmaps which show the coverage of data distribu- tion of dominant color components. K-means algorithm is used to cluster the color value of each pixel px,y 2 I. Let K = 4, then four dominant colors mi, for i = 1, . . ., 4, are estimated by K-means algo- rithm. Then, four bitmaps I(x, y)i are created as follows: Iðx; yÞi ¼ 1 ; if arg min mi normðmi; ciÞ¼ mi 0 ; if arg min mi normðmi; ciÞ – mi 8>< >: for i ¼ 1; . . . ; 4; where I(x, y)i is the bitmap created using dominant color mi, and ci is the ith color component at (x, y). An example is shown in Fig. 11, the size of each image is nor- malized to 80 � 80 at first. Then, K-means algorithm is used to cluster pixels of original image to four color component groups according to the color values of pixels. And, four bitmaps are estab- lished to show the distribution of data points in each groups. These four bitmaps will be used to measure the similarity between images (in Sections 4.2 and 4.3). 4.2. Similarity measurement of two bitmaps Giving two normalized bitmaps that show the coverage area for data distribution of specified color components, the similarity S(A, B) between two bitmaps A and B is defined as follows: SðA; BÞ¼ A\Bk k A[Bk k ; where A and B are the sets containing the pixels with value 1 in bitmap A and B, respectively. As shown as Fig. 12, the similarity is defined according to the ratio between overlapped (d) and overall (a [ d [ b) area of A ða [ dÞ and B ðb [ dÞ, where a, b, and d repre- sent the shaded areas in Fig. 12. Since four bitmaps are extracted to represent the color charac- teristics of an color image, the similarity between two images can be estimated based on the group similarity Sg between two group of bitmaps. The computation of group similarity can be performed on a best match between the elements in bitmap groups. The best match can be searched by the method proposed in Section 4.3. According to the best match, the group similarity measurement Sg is defined as follows: Sg ¼ X i kAikþkBikP j kAjkþkBjk � � SðAi; BiÞ; Fig. 11. Using K-means algorithm, pixels of original image, for example in (a), are clustered into four dominant groups according its color value. Then, four bitmaps, for example (b), are established to show the data distribution of pixels in each dominant groups. Fig. 12. Giving two coverage area of data distribution a and b, the similarity between these two coverage area is defined according to how much portion of these two coverage area are overlapped (d). P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 771 whereAi and Bi are the sets containing the pixels with value 1 in bit- map Ai and Bi, respectively. That is, the group similarity measurement Sg is the weighted sum of similarities between each pair of bitmaps, among the total number of pixels in each pair of bitmaps. 4.3. Similarity matching between bitmap groups For each image, four bitmaps are extracted to represent the col- or characteristics of an image. To measure the similarity between two images, a one-to-one and onto match (Liu, 1985) between ele- ments of two bitmap groups can be performed by the following proposed method. Giving two bitmap groupsA¼ A1; . . . ; ANf g and B¼ B1; . . . ; BNf g, where Ai and Bi are bitmaps. An N � N matrix M is formed in the fol- lowing manner: M ¼ m11 � � � m1N .. . . . . .. . mN1 � � � mNN 2 664 3 775; where mij ¼ SðAi; BjÞ: 772 P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 Since an element mij represents the similarity value when Ai and Bi are considered to be matched with each other. Finding a one-to-one and onto match between A and B is equivalent to selecting N ele- ments from each column of M, and no two elements are selected from the same row. To have the largest overall similarity SgðA;BÞ of a one-to-one and onto match between two bitmap groups A and B, the N elements should be selected such that the sum of se- lected elements is maximized. In order to regularize the computing procedure, row exchanges are performed to have the N selected ele- ments aligned on the diagonal position. Thus, the similarity value between bitmap groups can be computed as the sum of diagonal elements. Fig. 13 shows the flow chart of the proposed method that max- imizes the sum of diagonal elements by properly exchanging rows. Giving a similarity matrix M. First, find an element mij with the largest value in a similarity matrix M. Then, swap the ith row and jth row such that the element mij is at (j, j). Then, mark all ele- ments in jth row and jth column. Then, find elements with maxi- mal value from un-marked elements and swap the corresponding raw and column, such that the maximal element is moved to diag- onal position. Repeat the same process, until all elements are marked. Although the procedures stated above are efficient, its results may not be good enough. Therefore, a checking step is applied on each rows and columns to see whether any possible row exchange Fig. 13. The flow chart of the proposed matching method. At the beginning, a similarity matrix is constructed. The order of rows is then initialized by greedily exchanging the largest element to diagonal position. Then, further checking steps are applied on the diagonal elements to maximize the sum of diagonal elements. can further maximize the sum of diagonal elements in the similar- ity matrix M. The checking step is described as follows: Fig co se sim th Th ro res IF mii + mjj > mij + mji THEN exchange the ith and jth rows. This checking step is performed repeatedly until no further row exchange can be performed to increase the sum of diagonal ele- ments. In other words, the near maximum similarity value be- tween bitmap group A and B is achieved. An example of the one-to-one onto match method is shown in Fig. 14. At first, since m11 = 178 is the largest value in the 5 � 4 ma- trix, the elements in first row and first column are marked. Since the value m23 = 169, the largest value among the unmarked ele- ments is located in the third row, and the second row and the third row are exchanged. Then, the elements in second row and second column are marked too. Similar process is repeated until all col- umns are marked. According to the checking step, the second and the fourth rows are exchanged, since 80 + 143 > 60 + 162. In order to evaluate the performance of the proposed one-to- one and onto match algorithm, synthesis testing data sets are gen- erated by a random number generator. To evaluate different size of similarity matrices, 10 groups of test data sets with row dimension from 2 to 11 are generated. For each group, random seeds from 0 to 5999 are used to generate test data. The testing results are listed in Table 1. Label OST (optimizing sorting test) is performed by the proposed method, and label TAT (Testing All Test) indicates an exhaustive search method that vali- dates all the possible solutions. The TAT results are considered as a b c . 14. An example of the proposed matching method. The similarity matrix that is nstructed from the pair-wise similarity values is shown in (a). At first, by lecting an element with the largest value among unmarked elements in a ilarity matrix, swapping the selected element to diagonal position, and marking e elements in the same row and column iteratively until all elements are marked. e result is shown as (b). Then, the checking steps are applied to see whether any w exchanges can be applied to maximize the sum of diagonal elements. The final ults are shown in (c). Table 1 The performance comparison of the proposed one-to-one match algorithm (OST) and the exhaustive search method (TAT). To evaluate different size of similarity matrices, 10 groups of test data sets with row dimension from 2 to 11 are generated. The data listed in table are the averaged computing time in milliseconds spent by the matching process. TAT/OST is the ratio of the spending time by these two methods. Matrix dimension Computing time (ms) TAT/OST TAT OST 2 206 175 1.2 3 295 214 1.4 4 855 216 4.0 5 5440 228 23.9 6 40,853 232 176.1 7 349,117 255 1369.1 8 3,582,596 301 11902.3 9 39,197,125 327 119868.9 10 469,318,748 362 1296460.6 11 6,219,635,476 408 15244204.6 P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 773 the ground truth. The performance is measured in milliseconds of the total computing time to process all the test data. Apparently, the proposed method is quite efficient. Besides, based on the 60,000 (10 group of 6000 random seeds) testing data, the proposed method generates the same results as the ground truth. 4.4. Images clustering by enhanced K-medoid method Giving an image set I ¼ I1; . . . ; INf g, a matrix M is constructed by measuring the similarity between two images using the meth- ods proposed in Sections 4.1, 4.2, and 4.3. The image similarity ma- trix M for image comparison is defined as follows: M¼ m11 � � � m1N .. . . . . .. . mN1 � � � mNN 2 664 3 775; where mij ¼ SgðIi; IjÞ: Fig. 15. According to the web-based image clustering prototype system, the proposed similarity threshold is set as 0.35. That is, the similarity between a data object and the rep Then, the proposed variance enhanced K-medoid model clusters a set of images using the pair-wise image similarities Sg represented in the image similarity matrix M. The image corresponding to the medoid of a data cluster is selected as the representative image for each image cluster. 5. System demonstration A web-based prototype system (http://www.csie.nctu.edu.tw/ pslai/ASKMDDemo) is implemented to demonstrate the proposed method for public trial and testing. Fig. 15 shows a web-based user interface for image clustering. As shown in Fig. 15, a user may select a test media (http://www.youtube.com/watch?v = MgpzUo_ kbFY) set from the selector at the top-right corner of the web interface. The threshold, which represents the maximal allowed difference between data objects and its cluster center, is selected by clicking at a proper position on the ruler at the top-left corner. Then, the images in the test media set are clustered. And, the re- sulted clusters are presented in the order of cluster size. For each cluster, the represented image of each cluster is displayed in larger size, and the rest of images in a cluster are shown in smaller-sized thumbnail. Figs. 15–17 show the testing results using three different thresholds on the same media data set. In general, using higher threshold may reduce the number of clusters. More specifically, enlarging similarity threshold may reduce the similarity among data elements in certain clusters, so as to receive more not so sim- ilar data elements in a cluster. That is, the maximum allowed var- iance is increased. However, for data clusters containing quiet similar data elements, enlarge the similarity threshold will only slightly increase the number of less similar data elements, thus the maximal allowed variance has little inference on these data clusters. In addition, noisy data elements are clustered into certain less-similar clusters instead of spreading to all clusters. method successfully clusters similar images together. In this demonstration, the resentative object (cluster center) of a cluster should be larger than 0.65(= 1 � 0.35). Fig. 16. By increasing the similarity threshold to 0.4, certain data clusters receive several not so similar data objects (red frame A). However, the data clusters, which are previously contained quiet similar data elements, still have almost the same similarity among each other (clusters marked with B). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.) Fig. 17. By further increasing the similarity threshold to 0.45, data clusters previously contained some not so similar objects will grow to accept more dis-similar objects (red frame A). However, these data clusters, which previously contained quiet similar objects, will still have very similar objects (clusters marked with B). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.) 774 P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 6. Conclusion Generally speaking, grouping similar data together is useful to help user browse data efficiently. However, the similarity of data are difficult to define. Often, the definition of similarity between data objects depends on their applications. On the other hand, most clustering algorithms are developed for data sets which are represented as data points in feature space. That is, before P.-S. Lai, H.-C. Fu / Expert Systems with Applications 38 (2011) 764–775 775 clustering similar data, a feature extraction process is applied to each data objects to represent data objects as points in feature space. However, since the definition of similarity is different for various applications, and mapping the data objects into a feature space reasonably can be difficult and also complicated. Most of the time, proper feature extraction becomes a bottleneck for appli- cation system development. Instead of mapping data objects into points in feature space, an- other way to show the similarity among data objects is to measure the pairwise similarity among data objects. Describing the similar- ity between two data objects is more instinctively than mapping data objects into a feature space since computing similarity be- tween pair of data objects only need the special relationship be- tween data objects. A popular clustering algorithm, called K-medoid, is often used to cluster data objects using the pairwise similarities between data objects. K-medoid clustering algorithm groups data objects into clusters by maximizing the sum of similarity between each data objects and their cluster center. That is, the variance of each cluster still has the same value. In this paper, variance enhanced K-medoid clustering is pro- posed to introduce the idea of data variance into the original K- medoid algorithm. Besides, multiple variances are used to describe data variances along line segments between cluster centers. That is, the variance of cluster in different orientation is considered. An intelligent image thumbnail system is prototyped to demon- strate the proposed variance enhanced K-medoid clustering meth- od. For each image, the covered regions of four dominant colors are segmented at first. Between two images, the differences in four dominant color regions are measured as the similarity between images. Then, the variance enhanced K-medoid model is applied on these pairwise image similarities to group images into clusters. And, the image at cluster center (medoid) is used as the key image for this group. The web-based thumbnail demonstration show that the system successfully groups similar images together, and the key image is a good representative of the images in the cluster. By increasing the similarity thresholds of the proposed clustering method, the number of clusters will decrease. And, the clusters that contained quiet similar images before the changes will only changes slightly, i.e., receive a few more images, although the clus- ter that contained some not so similar images will grow bigger due to more outlier images are received. Such properties of the pro- posed model are useful such as selecting representative frame-sets from video for various video summary applications. References Bilmes, J. (1997). A gentle tutorial on the em algorithm and its application to parameter estimation for Gaussian mixture and hidden markov models, Tech. rep., ICSI-TR- 97-021, University of Berkeley. Comaniciu, D., & Meer, P. (2002). Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5), 603–619. Fu, H.-C., Lai, P.S., Lou, R.S., & Pao, H.T. (2000). Face detection and eye localization by neural network based color segmentation. In: Neural networks for signal processing X, 2000. Proceedings of the 2000 IEEE signal processing society workshop (Vol. 2, pp. 507–516). Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature selection. The Journal of Machine Learning Research, 3, 1157–1182. Hastie, T., Tibshirani, R., & Friedman, J. (2001a). The Elements of Statistical Learning:Data Mining, Inference, and Prediction. Springer. Hastie, T., Tibshirani, R., & Friedman, J. (2001b). The Elements of Statistical Learning:Data Mining, Inference, and Prediction. Springer. Hastie, T., Tibshirani, R., & Friedman, J. (2001c). The Elements of Statistical Learning:Data Mining, Inference, and Prediction. Springer. Hastie, T., Tibshirani, R., & Friedman, J. (2001d). The Elements of Statistical Learning:Data Mining, Inference, and Prediction. Springer. Haykin, S. (1994). Neural Networks, A Comprehensive Foundation. Macmillan College Publishing Company, Inc.. Lai, P.-S. & Fu, H.-C. (2007). A polygon description based similarity measurement of stock market behavior. In: IEEE congress on evolutionary computation, 2007 (CEC 2007) Singapore (pp. 806–812). Lai, P.-S., & Fu, H.-C. (2010). A polygon description based similarity measurement of stock market behavior. Expert System with Applications, 37(2), 1113–1123. Liu, C. L. (1985). Elements of discrete mathematics. New York: McGraw-Hill. Ch. 4, p. 126. Pao, H., Chen, Y., Lai, P., Xu, Y., & Fu, H.-C. (2008). Constructing and application of multimedia tv-news archives. Expert Systems with Applications, 35, 1444–1450. Pao, H., Chuang, S., Xu, Y., & Fu, H.-C. (2008). An em based multiple instance learning method for image classification. Expert Systems with Applications, 35, 1468–1472. Schapire, R. E., & Singer, Y. (1999). Improved boosting algorithms using confidence- rated predictions. Machine Learning, 80–91. The demonstration of variance enhanced K-medoid model. Available from http:// www.csie.nctu.edu.tw�pslai/ASKMDDemo/. The test video. Available from http://www.youtube.com/watch?v=MgpzUo_kbFY. http://www.csie.nctu.edu.tw~pslai/ASKMDDemo/ http://www.csie.nctu.edu.tw~pslai/ASKMDDemo/ http://www.csie.nctu.edu.tw~pslai/ASKMDDemo/ http://www.youtube.com/watch?v=MgpzUo_kbFY Variance enhanced K-medoid clustering Introduction K-medoid The learning algorithm of K-medoid clustering Self-growing K-medoid Discussion Variance enhanced K-medoid clustering Polygon-shaped data model Side-cluster segmentation Projections on the center-to-center link Decision boundaries adjustment Variance enhanced K-medoid clustering method Real-world application Color clustering Similarity measurement of two bitmaps Similarity matching between bitmap groups Images clustering by enhanced K-medoid method System demonstration Conclusion References 
work_w5iiiaai3jgclg3ek6obibep2m	----	AIC2 1 Automation in Construction, Vol. 12, No. 5, 2003, pp. 589-602 A COUPLED KNOWLEDGE-BASED EXPERT SYSTEM FOR DESIGN OF LIQUID RETAINING STRUCTURES K.W. Chau1 & F. Albermani2 1Department of Civil & Structural Engineering, Hong Kong Polytechnic University, Hunghom, Kowloon, Hong Kong 2Department of Civil Engineering, University of Queensland, QLD 4072, Australia Abstract This paper describes a coupled knowledge-based system for the design of liquid retaining structures, which can handle both the symbolic knowledge processing based on engineering heuristics in the preliminary synthesis stage and the extensive numerical crunching involved in the detailed analysis stage. The prototype system is developed by employing blackboard architecture and a commercial shell VISUAL RULE STUDIO. Its present scope covers design of three types of liquid retaining structures, namely, a rectangular shape with one compartment, a rectangular shape with two compartments and a circular shape. Through custom-built interactive graphical user interfaces, the user is directed throughout the design process, which includes preliminary design, load specification, model generation, finite element analysis, code compliance checking and member sizing optimization. It is also integrated with various relational databases that provide the system with sectional properties, moment and shear coefficients and final member details. This system can act as a consultant to assist novice designers in the design of liquid retaining structures with increase in efficiency and optimization of design output and automated record keeping. The design of a typical example of the liquid retaining structure is also illustrated. Introduction Crack width control is imperative in liquid retaining structures, which are exposed to a corrosive environment. The design of these structures is quite specialized and requires assimilation of knowledge from heuristics, research findings and standard engineering methodology. It requires extensive studies of loading requirements, underground geotechnical conditions, wind force characteristics, load combinations, practical dimensions and distribution of reinforcement and concrete, material properties and the exposure environment. Moreover, deviations often exist between the assumed properties of components at the preliminary design stage and that at the detailed design stage. In such cases, re-analysis will be entailed and iterative steps such as numerical modeling, structural analysis and code conformance checking are usually involved. Existing computer models may be available to perform a particular task in the whole design process. It is difficult to code empirical rules or expert knowledge in a conventional algorithmic framework. The application of these individual programs requires the intensive knowledge of the structural designer and they are prone to human errors during the data transferring processes. Recent advances in artificial intelligence techniques have rendered it possible to incorporate the heuristic knowledge into the conventional algorithmic structural analysis models. A knowledge-based system (KBS) is an interactive decision-making computer tool that mimics the reasoning processes of experts in a problem domain. Basic components of a KBS are system context, knowledge base, inference engine, knowledge acquisition, a user interface and explanation facilities. During the last decade, KBSs have been widely adopted to solve problems in many different disciplines (Chau, 1992; Chau and Anson, 2002; Chau This is the Pre-Published Version. 2 and Chen, 2001; Chau and Ng, 1996; Chau and Yang, 1992; Chau and Zhang, 1995; Maher, 1987; Shwe and Adeli, 1991). For application in engineering design, the KBS framework is in addition required to combine symbolic processing and extensive numerical processing. Examples of such systems by employing various representation schemes are INDEX (Kumar, 1995) and LADOME (Lin and Albermani, 2001). The objective of this study is to develop a microcomputer KBS that can bring all the design stages together into a user-friendly environment. One of the primary motivations for implementing a KBS on microcomputers is the low cost and availability of sophisticated expert system shells for personal computers, which are widely available in the design office. This prototype system for design of liquid retaining structures, LIQSTR, can offer assistance and advice to the engineer in making decisions during design. Conventionally, design is considered to be a constraint-satisfying problem and it is cumbersome to effect the most structurally optimized solution. However, by adopting a coupled approach, the system solves this problem and provides a more rational basis for the whole process of structural design. Increases in efficiency, standardization and optimization of design output and automated record keeping are among the benefits of such a KBS. The system can be used effectively by engineers of varied technical backgrounds. Novice designers can use it as a self-explanatory design system, supported by a combined factual/heuristic knowledge base. Experienced engineers can also use the system as a working tool to conduct the design in a more efficient and consistent way. It can also help engineering students to understand the expert approach and to acquire synthetic experience of the solution process. Besides, it can be a valuable research tool for evaluation and validation of the empirical rules and procedures embedded in design of liquid retaining structures. System Development Environment The KBS development environment for LIQSTR is VISUAL RULE STUDIO, which acts as an ActiveX Designer under the Microsoft Visual Basic programming environment (Rule Machine Corporation, 1998). VISUAL RULE STUDIO is an application development environment that combines expert system technology with object-oriented programming, a relational database, graphics capabilities and debugging tools. It furnishes a variety of knowledge representation schemes, inference mechanisms and capabilities to interface with external programs in a windows environment. Various types of displays, such as form, checkbox group, list box, command button, textbox, option button, picture box, etc., can be defined as different classes inheriting common characteristics and possessing their own special properties. Architecture of System A blackboard architecture (Engelmore and Morgan, 1988) has been adopted due to the nature of structural design, the expert system shell and its capability for coupling all design stages for liquid retaining structures. Under this model, the knowledge base comprises both declarative and procedural knowledge. This well-organized architecture furnishes communication between diversified knowledge modules involved in structural design process as well as facilitates future possible extension. Figure 1 shows the architecture of LIQSTR. The knowledge base is mainly composed of knowledge modules and the blackboard. Knowledge modules correspond to procedural expertise knowledge in solving design problem and are divided into two groups, namely, Design Process and Process Control. Message communications between Design Process knowledge modules and Process Control knowledge modules are effected via the blackboard. 3 The blackboard contains only declarative knowledge and is divided into two groups, namely, Design Entities and Design Stage. The Design Entities represent the breakdown of design concepts of liquid retaining structures. Each object in the blackboard incorporates some attributes, which are used to express the physical structural components, the facts used in the design, or the pertinent design entities entailed during the design solution process. The attribute can be defined as one of the following types, namely, compound, multi-compound, instance reference, numeric, simple, string, interval, and time. Fact about an attribute can be represented by facets, which design the inference strategy for processing an attribute. A semantic network, as shown in Figure 2, illustrates the relationships among various objects in Design Entities and represents abstraction of the knowledge. The unique class in Design Stage includes several attributes that represent indicators tracking the current stage of every design context. The status or indicator of the attributes has to be one of the pre-defined values. The knowledge represented in this level will handle the order of execution via the Process Control knowledge modules, which are responsible for controlling the sequence of undertaking these tasks. Figure 3 shows attributes of the class Design Stage. An example of these attributes is “alternativeEvaluation”, which is of the compound type. During the design session, this compound type attribute must have one of the following values: “not completed” or “finished”. Both the initial and default values of this attribute are invariably defined as “not completed”. The search order is in the order of session context and then default value. Knowledge acquisition Knowledge acquisition is the process of gathering of rules and facts from literature, including handbooks, standards and codes, and interviews with experienced designers of liquid retaining structures. Domain knowledge such as the minimum percentage of reinforcement, maximum percentage of reinforcement, minimum grade of concrete, etc. are then translated into rules or methods using an object-oriented representation. For instance, the following rule group, which is expressed using the Production Rule Language, represents knowledge on the determination of concrete tensile strength for different concrete grades. !RULE GROUP: concreteTensileStrength OF BBSectionalProperties RULE to find concreteTensileStrength : 1 of 3 IF concreteGrade OF BBSectionalProperties = 40 THEN concreteTensileStrength OF BBSectionalProperties := 1.75 RULE to find concreteTensileStrength : 2 of 3 IF concreteGrade OF BBSectionalProperties = 35 THEN concreteTensileStrength OF BBSectionalProperties := 1.6 RULE to find concreteTensileStrength : 3 of 3 IF concreteGrade OF BBSectionalProperties = 30 THEN concreteTensileStrength OF BBSectionalProperties := 1.45 Integration Methodology Microsoft Visual Basic offers an interfacing facility in a fully integrated fashion. After the execution of the external algorithmic program, the KBS resumes its design session environment. 4 Finite element analysis program ABAQUS The finite element package ABAQUS has both input and output files in ASCII format (Hibbitt et al, 1998). When the requisite input files have been prepared following model generation, the knowledge base calls the finite element program to perform the analysis. Through the interfacing module, the output data files are saved in the knowledge base for retrieval and manipulation. Accurate conversion of data format and correct mapping of character have to be attained during the data file preparation. The KBS is not only aimed to act as a front-end to this finite element analysis package, but also to encapsulate knowledge on the entire design process. During the member sizing process, in order to achieve the optimum design, structural re-analysis is very often inevitable. In these design cycles, the ensuing action to be taken by the system depends largely on message transfer amongst different components.. Numerical model generator and code conformance checking programs Whilst existing algorithmic models generally deal with numerical data input only, the novice user usually finds it more convenient to express in a natural language. The numerical model generator converts these linguistic variables entered by the user in the knowledge base to numerical format conforming to stipulations of the structural analysis package. The code conformance checking program then checks the code requirements of BS 8007 (British Standards Institution, 1987). Apart from the finite element analysis, all number crunching and data manipulation procedures are written in Visual Basic. The communication between the programs and the knowledge base is performed mainly through the objects and attributes. The data can be used by all other components through the interfacing module. Databases External Access database files are accessed and updated whenever a consultation is made with the system. The database is composed of structural properties of reinforced concrete sections, moment and shear coefficients for various configurations in preliminary design, structural properties of proposed alternatives and final member details in detailed design. They provide useful information for decision-making and design of liquid retaining structures. The Data Control, which provides an efficient connection of the Visual Basic application to the databases, can open a specified database table or a set of records based on an SQL query on the database. Two key attributes of the Data Control are the DataBaseName property that specifies the name of the database and the RecordSource property that specifies the name of a table or a SQL query in the database. A Data-aware Control such as TextBox, DBGrid, etc., connecting fields of data in the database to the control, is then used to bind to the Data Control. Two key attributes of the Data-aware Control are the DataSource property that specifies the name of the Data Control and the DataField property that specifies the name of the field it is binding to. Explanation system Explanations of the reasoning process, liquid retaining structure types, procedures to take various design loading, various code provisions and expert comments regarding the design of liquid retaining structures are included in the explanation system of LIQSTR via the Help button. Explanation of the solution derived by the system is currently displayed by applying user interface facilities such as message boxes. The explanations consist of built-in specific texts together with associated values of design parameters generated by the knowledge base during system run time. Moreover, the user can get an explanation of the command button at any level of execution by viewing the ToolTipText with the mouse pointing over it. 5 User interface The user interface allows the user to specify all design requirements and acquire the output results from the design consultation. The user can monitor the performance of the system during the design process through the interface. In fact, in most KBSs, it has been reported that a large part of the codes normally deal with the system-user interface (Kumar, 1995). In this prototype system, graphical user interfaces, consisting of layers of display screens and pop-up windows, are used for message transfer so that data handling has been greatly simplified. Contrary to traditional algorithmic models, the user has control over the sequence of actions during the design process subject to conformance with the Process Control knowledge modules. System description of LIQSTR The opening menu of LIQSTR is shown in Figure 4. When a command button is clicked, the value of the attribute attached to the command button is altered and the method associated with the attribute is executed, which in turn leads to changes in the display. Designing a liquid retaining structure involves the determination of a suitable configuration and selection of structural components according to both the design specifications and engineering practices. In this prototype system, three basic shapes, i.e. rectangular with one compartment, rectangular with two compartments and circular, and two layouts, i.e. underground and above the ground, are considered. The members are then sized for the imposed load, and wind load if any, on the structure. Structural optimization is considered in terms of the total costs of concrete together with reinforcement. Major design tasks performed by the KBS are preliminary synthesis, detailed specification, numerical model generation, structural analysis, optimized member sizing, and code conformance checking. Preliminary synthesis The preliminary synthesis stage of the design is begun with the geometrical and structural specifications. This includes the volume and height required, and also the width to breadth ratio for a rectangular tank. The system applies engineering heuristics such as span depth ratio, crack width computation and moment and shear coefficients for different length and width ratios to evaluate each alternative. Table 1 shows an example of such a heuristic design table showing the moment and shear coefficients for the vertical walls of rectangular shaped liquid retaining structures with different height and width ratios. The KBS selects proposed alternatives from a database containing characteristics of available reinforced concrete sections, which are defined by the concrete cross-sectional area composing of wall or slab thickness times unit metre length, reinforcement diameter as well as reinforcement spacing. In order to narrow the search space, it lists 15 most feasible reinforced concrete sections in order of priority based on minimum total material cost. The alternative with the minimum cost will be recommended by the system as the selected proposed alternative, but the user can still opt for overruling this recommendation. In fact, LIQSTR plays the role of a knowledgeable assistant in the decision making process by supplying details of the evaluated parameters. Heuristic rules reduce the amount of computations needed for preliminary design of members. In calculating the dead weight, the approximate weight of the member yet to be designed is initially estimated on the basis of the total gravity load supported by the structure and the span length of the structure. A more accurate estimate of the member weight is included after the first cycle of iteration. Detailed specifications The selected alternative is specified in much greater details. In order to determine the structural response for the chosen configuration accurately, primary loading conditions, the 6 load combinations and the support conditions must be specified in detail. The system generates default loading and support conditions, which can be accepted or modified by the user. The user can also define specific load combination and wind speed. The system automatically calculates the structural dead weight and imposes it at the nodal points. The user is required to specify the imposed loads, which include the surcharge load and the level of liquid inside the tank. If the liquid retaining structure is chosen to be underground, the user is required to enter level of ground surface, level of water table, specific weight of soil and active soil pressure coefficient. If the liquid retaining structure is chosen to be above the ground, wind forces acting on the external surfaces of the structure are calculated according to the Code of Practice for Wind Effects in Hong Kong (Hong Kong Building Development Department, 1983). The design wind pressure profiles for various types of terrain have been calculated at different height based on the gust velocity in Hong Kong conditions. Various load combinations according to BS 8110 (British Standards Institution, 1985) are considered. For serviceability limit state, 3 load combinations, i.e., DL + WL, DL + LL, and DL + LL + WL, are considered, where DL represents dead load factor, LL stands for live load factor while WL means wind load factor. For ultimate limit state, 3 load combinations, i.e., 1.4DL + 1.4WL, 1.4DL + 1.6LL, and 1.2DL + 1.2LL + 1.2WL, are considered. The user can adopt another set of load factors by entering the required values on the screen as shown in Figure 5. Structural analysis, crack width checking and optimization In order to arrive at the optimized design solution, an iterative process is involved here. A detailed design analysis, which is involved mainly with computationally intensive numerical processing, is conducted to generate the structure with the minimum cost. Upon receiving messages from the knowledge base, a structural model is prepared by the model generator. The messages include the descriptions of configuration, structural dimension, support, wind loads, dead loads, member properties and load factors. Prior to the finite element analysis, LIQSTR generates the joint and member information of the liquid retaining structure automatically, thus relieving the user of the cumbersome task of manipulating a large amount of data manually. Heuristic knowledge is employed to generate joints and members information, which include joint numbering, joint coordinates, member numbering and member-joint connectivity data. The created numerical model is especially tailor-made for the ensuing analysis employing the finite element package ABAQUS. The required input file format is quite complicated. It may require a structural engineer novice to the package a significant effort to manually generating input file successfully. Now, by using this KBS, the knowledge on the preparation of the input data file of a liquid retaining structure for ABAQUS analysis is represented symbolically. This results in both timesaving and avoidance of errors during data preparation or data transfer. The system evaluates the structural stability from the analysis results and furnishes post-processing for these results. Then it checks the structural members according to the relevant code of practice, BS 8007. The computed crack width has to be less than the prescribed crack width, which depends on the exposure environment. The crack width, Cw, is determined using the following formula: )(21 3 min xh Ca a C cr mcr w − − + = ε where acr = distance from the point considered to the surface of the nearest longitudinal bar (Figure 6); Cmin = minimum cover to the longitudinal bar; εm = average stain at the level considered; h = overall depth of the member; and x = depth of the neutral axis. (1) 7 The average strain at the level at which cracking is being considered, is given by εm = ε1 - ε2 where ε1 = the strain at the level being considered = s s Exd fxh )( )( − − ; ε2 = the strain due to stiffening effect of concrete; fs = service stress in tension reinforcement obtained from finite element analysis; Es = modulus of elasticity of steel; and d = effective depth. Besides, the strain in tension reinforcement is not allowed to exceed 0.8fy/Es.. For crack width Cw of 0.2 mm, For crack width Cw of 0.1 mm, where b = width of the section; and As = area of tension steel. The optimized reinforced concrete section obtained is then compared with the initial section used in the structural analysis. If they are not the same, the next design cycle, which involves model generation, structural analysis and code conformance checking, is invoked again using the new section. Typically this phase of design involves several iterations. Once final design result is achieved, the system will present a summary of the final design and generates a printed design report. Figure 7 is the flowchart showing the structural optimization procedure. The optimum values obtained from interactive optimization can then be added to the knowledge base of the system, which is effectively extended by the knowledge acquired through machine experimentation. Machine learning can be effected since the final optimum structural section from the finite element structural analysis is added to the database containing the heuristics during the preliminary design. The system can also be used as a means for testing the available empirical knowledge since the heuristic moment and shear coefficients during the preliminary design stage can be validated with the model run. Case study A case study on a typical liquid retaining structure is used to demonstrate the application of LIQSTR. The spatial requirements of the structure are as follows: Shape = rectangular with two compartments Volume = 100 m3 Depth = 5 m Breadth/width ratio = 1.2 Location = above ground Exposure condition = severe Based on these geometric constraints and crack width requirement, the KBS performed a preliminary synthesis using heuristics. According to the span lengths of the structure, the system searches the databases on moment and shear coefficients and on sectional properties and proposes 15 feasible proposed configurations in order of priority of minimum costs. (2) )(3 )( 2 2 xdAE xhb ss − − =ε )(3 )(5.1 2 2 xdAE xhb ss − − =ε (3) (4) 8 During this preliminary synthesis, the system suggests an initial member thickness of 225 mm with reinforcement diameter 10 mm at 100 mm spacing as the most suitable alternative. As shown in Figure 8, the proposed alternative include the alternative number, slab thickness, bar size, bar spacing, bar area, span depth ratio and the total material cost. The total costs of the suggested feasible alternatives are listed in ascending order. The user can choose between system’s selection and user’s selection through an option button with a corresponding evaluation message displayed. In this case, the system’s selection is opted. The default design parameters are adopted and shown as follows: Concrete grade = 40 Reinforcement grade = high yield steel Concrete cover = 40 mm Aggregate type = granite Temperature variation = 15 °C Unit cost of concrete = $3000 per m3 of concrete Unit cost of reinforcement = $500 per tonne of reinforcement Detailed design specification, as shown in the following, is input for the selected alternative prior to detailed design and analysis. Support specification = fixed Surcharge load = 10 kN/m2 Level of liquid = 5 m Wind load determination = found by system Terrain type = general terrain Top level = 5 m Load combination = default load combinations (1.4DL + 1.4WL/ 1.4DL + 1.6LL/ 1.2DL + 1.2LL + 1.2WL) Number of elements per surface = 100 (10x10) Figure 9 shows the screen displaying the execution of the finite element package, in which the total number of nodes and elements are computed automatically to be 366 and 350, respectively. The iterative process of numerical model generation, structural analysis, code conformance checking and optimized member sizing is undertaken. Finally, the system finds out that a member thickness of 300 mm with reinforcement diameter 25 mm at spacing 225 mm is the optimum section. Figure 10 shows the member sizing record of the recommended section, which includes slab thickness, reinforcement diameter and spacing. The designed bending moments at serviceability and ultimate limit states and the designed shear force at ultimate limit state are also shown. Table 2 tabulates the comparison between the moments and shears in preliminary synthesis and in the final design. Although the values are basically of the same order of magnitude, there still exist some rooms for improvement on accuracy of the values from heuristics in preliminary design, given the lack of details at that stage. As such, the values acquired from the final design, together with detailed specifications, can be added to the knowledge base to improve the existing heuristic knowledge. Conclusions A coupled microcomputer KBS on design of liquid retaining structure (LIQSTR) was implemented to combine expert knowledge with object-oriented programming, graphics capabilities, KBS technologies, conventional algorithmic models and relational databases under a windowing environment. It can act as a repository of empirical knowledge provided 9 by experienced specialists. It not only reduces time and effort in design of various structural elements of liquid retaining structures, but also offers assistance and explanations to the user. The prototype system undertakes all major design stages including preliminary synthesis, detailed specification, numerical model generation, nonlinear finite element analysis, code conformance checking, and optimized member sizing. Moreover, it has been shown that machine intelligence can be introduced into the KBS in addition to the knowledge acquired from the human experts and the literatures. From this point of view, the system will be an ideal research tool to validate and enhance our empirical knowledge, which in turn may lead to more efficient and optimized structural design. References British Standards Institution, 1985, BS 8110: Structural Use of Concrete. British Standards Institution, 1987, BS 8007: Design of Concrete Structures for Retaining Aqueous Liquids. Chau, K.W., 1992, An expert system for the design of gravity-type vertical seawalls, Engineering Applications of Artificial Intelligence, 5(4), 363-367. Chau, K.W. and Anson, M., 2002, A knowledge-based system for construction site level facilities layout, Lecture Notes in Artificial Intelligence, 2358, 393-402. Chau, K.W. and Chen, W., 2001, An example of expert system on numerical modelling system in coastal processes, Advances in Engineering Software, 32(9), 695-703. Chau, K.W. and Ng, Vitus, 1996, A knowledge-based expert system for design of thrust blocks for water pipelines in Hong Kong, Water Supply Research and Technology - Aqua, 45(2), 96-99. Chau, K.W. and Yang, Wen-wu, 1992, A Knowledge-based expert system for unsteady open channel flow, Engineering Applications of Artificial Intelligence, 5(5), 425-430. Chau, K.W. and Zhang, X.N., 1995, An expert system for flow routing in a river network, Advances in Engineering Software, 22(3), 139-146. Engelmore, R. and Morgan, T., 1988, Blackboard Systems, Addison_Wesley, Wokingham. Hibbitt, Karlsson & Sorensen, Inc., 1998, ABAQUS User’s Manual, Version 5.8. Hong Kong Building Development Department, 1983, Code of Practice for Wind Effects in Hong Kong. Kumar, B., 1995, Knowledge Processing for Structural Design, Topics in Engineering Vol. 25, Computational Mechanics, Southampton. Lin, S. and Albermani, F., 2001, Lattice-dome design using a knowledge-based system approach, Computer-aided Civil and Infrastructure Engineering, 16(4), 268-286. Maher, M.L., 1987, Expert System in Structural Engineering, in M.L. Maher (Editor), Expert Systems for Civil Engineers: Technology and Application, American Society of Civil Engineers, New York. Rule Machine Corporation, 1998, Visual Rule Studio Developer’s Guide. Shwe, T.T. and Adeli, H., 1991, An expert system for earthquake design code processing, Structural Engineering Review, 3, 41-48. 10 b/a Moment Coefficients Shear Coefficients Mh Mv Vh Vv 0.5 9 15 16 20 1.0 29 35 23 32 1.5 42 60 26 41 2.0 62 86 25 46 2.5 76 108 27 48 3.0 89 126 31 50 4.0 101 148 34 50 5.0 103 158 35 50 Notes: a = vertical height (in m); b = horizontal width (in m); w = specific weight of liquid (in N/m3) Mh = horizontal moment coefficient; Mv = vertical moment coefficient Vh = horizontal shear coefficient; Vv = vertical shear coefficient Moment = moment coefficient x wa3/1000 (in kNm/m); Shear force = shear coefficient x wa2/1000 (in kN/m) Table 1. Moment and shear coefficients for a vertical wall of rectangular shaped liquid retaining structure with topside free and the other three sides fixed Preliminary design Detailed design SLS moment 34 50 SLS shear 69 164 ULS moment 47 60 ULS shear 95 197 Table 2. Comparison between moments and shears in preliminary synthesis and in detailed design for the case study 11 Figure 1. Architecture of the coupled knowledge-based system Microsoft Windows Environment Visual Basic Programming Environment Knowledge Acquisition Module Knowledge Base Shell (Visual Rule Studio) User Interfaces Explanation Module Knowledge Engineer Interview with Expert Knowledge from Literature User Input/ Output ASCII Data Files Finite Element Structural Analysis Package (ABAQUS) Microsoft Access Databases (Moment Coefficients, Sectional Properties, Final Member Details) System Interfaces Code Conformance Checking Module Knowledge Base Inference Engine Backward Chaining Forward Chaining Blackboard (Context) Knowledge Module Design Process Process Control Design Entities Design Stage Procedural Methods Model Generation Module Requirements of Code of Practice 12 Figure 2. Semantic network representing objects and their relationship in the blackboard Liquid Retaining Structure Configuration requirement Proposed Alternatives Feasible Proposed Alternatives Preliminary Proposed Alternative Final Member Details Database Final Member Details Output Data Files Finite Element Package Model Parameters Input Data Files Support Condition Wind Load Preliminary Parameters Dead Load Moment & Shear Coefficient Database Selected Member Properties Imposed Load Sectional Properties Database Load Combination is part of is part of produces is member of is member of depends on produces executes is prepared for depends on is specified for is used to establish is specified for is specified for is specified for is used to establish belongs to is specified for is member of is member of 13 CLASS DESIGN_STAGE WITH structuralSpecification COMPOUND not specified, specified WITH imposedLoad COMPOUND specified, not specified WITH windLoad COMPOUND not specified, specified WITH loadCombination COMPOUND not specified, specified WITH modelSpecification COMPOUND specified, not specified WITH supportSpecification COMPOUND specified, not specified WITH crackWidthCheck COMPOUND not executed, complete WITH analysisAndMemberSizing COMPOUND not ready, finished WITH restartApplication SIMPLE WITH alternativeEvaluation COMPOUND not completed, finished INIT not completed DEFAULT not completed SEARCH ORDER CONTEXT DEFAULT WITH memberSizingRecord COMPOUND not produced, produced WITH finalMemberDetail COMPOUND not produced, produced WITH designSpecification COMPOUND not completed, completed WITH designReport COMPOUND not generated, generated Figure 3. Characteristics of class design stage 14 Figure 4. Screen showing the main menu 15 Figure 5. Screen showing input of specified load combination overriding the default values 16 Figure 6. Design parameters in crack width calculation s acr Cmin d h s/2 Section Strain x εs=fs/Es ε1 εm 17 Figure 7. Flowchart showing the structural optimization procedure start Geometrical specification Estimate maximum moment and shear from heuristic moment & shear coefficient database For each alternative section Determine ULS/SLS moment & shear capacities Are moment & shear capacities > estimated moment & shear? Sort each feasible alternative based on minimum total material cost of reinforcement and concrete No Yes Select alternative Determine maximum moment and shear from finite element analysis For each alternative section Are moment & shear capacities > computed moment & shear? Get best alternative with minimum total material cost Is best alternative same as selected alternative? Revise selected member size Yes No No Stop yes Optimized section obtained 18 Figure 8. Screen showing the selection of proposed alternative 19 Figure 9. Screen showing structural analysis by finite element package 20 Figure 10. Screen showing the best member sizing record 
work_w67qgpvf3zae5pqnx5ewzoapb4	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220969515 Params is empty 220969515 exception Params is empty 2021/04/06-03:05:08 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220969515 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:08 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_w6ft4swj3fa5fjkapir2k2svdy	----	An Expert System for Tractor Selection Iowa State University From the SelectedWorks of Steven A. Freeman 1989 An Expert System for Tractor Selection Steven A. Freeman, Colorado State University P. D. Ayers, Colorado State University Available at: https://works.bepress.com/steven_freeman/38/ http://www.iastate.edu https://works.bepress.com/steven_freeman/ https://works.bepress.com/steven_freeman/38/ An Expert System for Tractor Selection S. A. F r e e m a n , P . D . Ayers STUDENT MEMBER ASSOC. MEMBER ASAE ASAE ABSTRACT A user friendly expert system to assist the farmer in tractor selection was developed based on the evaluation procedure developed by Rider et al. (1979). This evaluation procedure utilizes Nebraska Tractor Test data from 1980 to 1986 along with information provided by the farmer to make selection decisions. The expert system provides the farmer with a list of tractors that are best suited for a specific operation. However, it is only designed to aid the farmer in the tractor selection process, not to provide a definite solution. INTRODUCTION The present economic situation in which the American farmer operates demands that, in order to survive, the best possible use must be made of all the resources available. The key to an economically healthy agricultural operation is continued increases in the productivity of farm workers and the quality of farm products with a resulting reduction in the unit cost of farm products. In the past, this increase in production has largely been tied to the increased application of power. Continued increases in productivity and product quality will most likely follow from the intelligent management of power. In particular, expert systems offer an opportunity to more intelligently manage the resources of the agricultural science and education system. By increasing the management potential of system administrators, expert systems can help members of the agricultural science and education system continue to serve agriculture and the consumer, especially in a time of declining resources (Barrett et al., 1985). One of the most important machinery management decisions made by a farmer is the selection of a tractor. The cost of buying a farm tractor is such a significant expense in most agricultural operations that the farmer must weigh all possibilities very carefully before making a decision which could dictate financial stability and thus economic survival. In order to overcome such a complex problem, a logical procedure needs to be used. Numerical procedures for determining the desired tractor power require rigorous computations and a vast array of background knowledge (Hunt, 1983). Computer Article was submitted for publication in February, 1988; reviewed and approved for publication by the Power and Machinery Div. of ASAE in September, 1988. The authors are: S. A. FREEMAN, Student, and P. D. AYERS, Assistant Professor, Agricultural and Chemical Engineering Dept., Colorado State University, Ft. Collins. Contribution of Colorado Agricultural Experiment Station with funding support from AES 1-56031. programs have been developed to simplify these procedures. There are now several microcomputer models available to aid the farmer in determining the required power (Freesmeyer and Hunt, 1985). The most recent is one developed by Chen (1987) which includes improvements over earlier models by decreasing the run time and by making the model user-interactive. An expert system to aid in farm management decisions is also available. This system integrates a farm management linear program and a companion simulation model with a knowledge-based expert system (Kline et al., 1987). The problem, however, is not solved once the desired power has been calculated. The farmer must now decide which specific tractor model is best suited for the operation. This process involves quantitative and qualitative factors which are difficult to define mathematically. A procedure to aid the farmer in this process was developed by Rider et al. (1979). The farmer can benefit from this procedure; however, presently it does not enable the farmer to easily and efficiently review all the tractors available. A method which allows the farmer to benefit the most from this procedure could be accomplished with an expert system. An expert system can handle the quantitative and qualitative factors involved in this type of comparison more easily than a conventional computer program. An expert system would allow the farmer to take advantage of this procedure without having to know the variety of tractors available or the background data from the Nebraska tractor tests which are required for each specific model. OBJECTIVE The objectives of this project are: 1. To develop a user friendly expert system to assist the farmer in tractor selection. 2. To incorporate tractors ranging from 75 kW to 300 kW (100 hp to 400 hp) into the expert system. 3. To evaluate the system performance for ease of operation, validity of selection decisions and comparative worth. LITERATURE REVIEW Suitability of Specific Models The tractor selection process is not complete once the necessary horsepower has been determined. The decision of which specific manufacturer and model is best suited for a particular operation is a difficult one, especially if there are several models available within the acceptable power range. The farmer has many resources to help in this evaluation, including literature from the Vol. 5(2):June 1989 © 1989 American Society of Agricultural Engineers 0883-8542/89/0502-0123$03.00 123 manufacturer, Nebraska Tractor Test Reports, dealers and other farmers. Each source of information is useful, but screening the information without bias and determining which factors are most important is difficult. The relative importance of the determining factors may vary for different farmers; however, the farmer is allowed to define the relative importance of each main factor. This procedure is based on rating primary selection factors for each tractor in question and then comparing this total to the total of the other tractors being considered. The primary selection factors included in the quantitative evaluation procedure are fuel efficiency, lugging ability, transmission characteristics, sound level, reliability and dealer services. The values of these selection factors are obtained from Nebraska Tractor Test Reports and from information supplied by the farmer. A worksheet (Appendix A) describes the selection process. Expert System Applications in Agriculture Expert systems allow someone who is not an expert in a particular field to make decisions based on the knowledge of the experts in that field. Expert systems have many possible applications in agriculture — from diagnosing plant and animal diseases and soil acidity, to aiding in management decisions (Barrett et al., 1985). Wolfgram et al. (1987) listed several expert systems which are already available including: control of disease in winter wheat crops, control of plant life in ponds, crop rotation, management of apple orchards, rice disease diagnosis and a material handling equipment selector. Some additional expert system applications are Farm- Level Machinery Management (Kline et al., 1987), Management of a Crop Research Facility (Jones et al., 1986), Tomato Greenhouse Environment Controller (Jacobson et al., 1987), and Diagnosing Problems in Hydraulic Systems (Gaultney et al., 1987). The application possibilities of expert systems in agriculture are already numerous and will no doubt continue to expand in the future. EXPERT SYSTEM DEVELOPMENT The expert system was developed using the EXSYS Expert System Development Package Version 3.1 (EXSYS, 1985). EXSYS runs on any IBM compatible personal computer. It is a rule based expert system using an if-then-else format. As discussed earlier, the procedure developed by Rider et al. (1979) forms the basis for the expert system. A knowledge base of stored information necessary to solve the problem was required. The knowledge base was obtained from the Nebraska Tractor Tests. The data required for each specific tractor was determined using the worksheet and input to the knowledge base. The 75% pull was used for all comparisons in the calculation of the fuel efficiency index. The system is based on the assumption that the farmer knows the approximate tractor horsepower desired. Separate files were developed for each 7.5 kW (10 hp) increment from 75 kW (100 hp) to the largest tractor tested from 1980 to 1986 which was approximately 300 kW (400 hp). Each file covers 10% on both sides of the increment power value. For example, a tractor power of 75 kW (100 hp) evaluates tractors from 67.5 to 82.5 kW (90 to 110 hp). The files were written with the help of the EXSYS editor. The arrangement of the system demands an overlap between the files; thus, a specific tractor will appear on numerous files. A menu program accompanies the expert system. The menu program does three things for the user. First, it gives the user information about the range of the program and the option to obtain additional information. Secondly, it gives the user a short explanation of how to respond to questions asked by the computer. Most importantly, it provides the farmer with easy access to the expert system. The selection decisions utilized by the expert system are a direct implementation of the procedure developed by Rider et al. (1979). It was not the direct purpose of this project to evaluate their comparison method. Thus they are considered the "experts." The focus of this study was on the development of the tractor data files and the implementation of the overall expert system rather than on the quantitative analysis procedure. Data for 113 tractors from the 1980 to 1986 Nebraska Tractor Tests were incorporated into the expert systems. Thirty- one (31) data files were developed with the number of tractors in each file shown in Table 1. The system should be maintained to retain its usefulness. The data files can be updated each year when new data becomes available. EXPERT SYSTEM OPERATION The overall system is controlled by the menu program (Fig. 1). The user has a choice to review the introduction, review the instructions or run the tractor selection routine. When running the selection routine, the computer interacts with the user to obtain the desired horsepower for tractor comparison. It then uses this information to call the EXSYS runtime program and the appropriate tractor data file. On computers with two floppy disk drives, a request is issued to insert the appropriate disk (containing the knowledge base) into the second floppy disk drive. The user is asked a series of questions requesting information required to complete the worksheet shown in Appendix A. This information TABLE 1. Number of Tractors Evaluated for Each Power Group Selected Tractor power, kW (hp) 300.0 (400) 292.5 (390) 285.0 (380) 277.5 (370) 270.0 (360) 262.5 (350) 255.0 (340) 247.5 (330) 240.0 (320) 232.5 (310) 225.0 (300) 217.5 (290) 210.0 (280) 202.5 (270) 195.0 (260) (continues) Number of tractors evaluated 2 2 3 4 6 7 7 10 10 9 11 8 11 9 12 Tractor power, kW (hp) 187.5 (250) 180.0 (240) 172.5 (230) 165.0 (220) 157.5 (210) 150.0 (200) 142.5 (190) 135.0 (180) 127.5 (170) 120.0 (160) 112.5 (150) 105.0 (140) 97.5 (130) 90.0 (120) 82.5 (110) 75.0 (100) Number of tractors evaluated 14 14 14 14 14 19 16 27 22 18 19 13 19 20 20 21 124 APPLIED ENGINEERING in AGRICULTURE USER EXPERT SYSTEM EVALUATIONS Fig. 1—Schematic diagram of the tractor selection program. includes: (a) farmer weighting factors, (b) ability to sustain workload, (c) dealer location, (d) dealer parts inventory, and (e) dealer repair services. At this point the user becomes an "expert" at the local level and the selection is partially based on information from the user. In this manner, the tractor selection procedure conforms to the individual's farming operation meeting the individual's desires and perceived needs. The expert system calculates the comparative tractor index (CTI) for each tractor in accordance with the worksheet in Appendix A. Since this method is to be used as an aid in tractor selection, not to determine a final decision, the results of the CTI calculations are first formatted to a general 0 to 10 rating (with 10 being the highest). The tractor models and their ratings for tractors receiving a rating greater than " 5 " are presented to the user. If a tractor receives a rating of " 5 " or lower, it is considered not to be a viable choice and is thus not included in the list of results, although the user has the option of listing all possibilities. The comparative tractor indices for all tractors are also given for further comparison if two or more tractors should tie with the same rating. At this point, the EXSYS runtime program gives the user several options which are again listed at the bottom of the screen. One important option allows the user to change the answer to any of the questions the computer asked and then see what effect these changes have on the results. All of the options are self-explanatory, but if a question should arise, there is an extensive internal help file to assist the user. When the user is finished making changes, a hard copy may be obtained before continuing. As with many expert systems, it is apparent that this tractor selection procedure could have been implemented with a "numerical" language as opposed to an expert system shell. The use of an expert system shell to program the selection procedure was based on the following factors: (a) the shell provides easier user interaction, (b) the shell provides an internal help feature, (c) the shell provides the flexibility to change input quickly, and (d) the shell provides the ability for online description of the selection process. The expert system shell also allows the ability to update the tractor data files quickly. The tractor data files and menu program are available from the authors. However the EXSYS Runtime Program needed to run the expert system is a licensed product and may not be distributed by the authors. The evaluation procedure consisted of having six farmers and one farm equipment dealer run the system and then fill out a short evaluation form. This evaluation form asked the following questions concerning the ease of operation, performance of the selection procedure and improvements needed. The response to the questions asked during the evaluation are discussed below. Was this program easy to operate and understand? . If not please give examples of where the difficulties were encountered: All seven reviewers answered 'yes' to this question. Even the individuals with no previous computer experience were able to run the program without additional instructions. Do you agree with the selection decisions made by this program? Why? All reviewers, except one farmer, answered 'yes' to this question. The most common reasons for agreement were because the system was based on the Nebraska Tractor Test information, which is an independent study, or that the system chose the same tractor that they themselves would have chosen. The one dissenting opinion was given because the best tractor was a four-wheel drive tractor and the user's operation was all row-crop. He did, however, believe that the selection process was correct based on tractor performance. (Note: the second choice was a row-crop model, thus it would be the best choice for his operation and the system still accomplished its goal.) Do you think this program was helpful? If you were about to purchase a tractor and you had access to a program like this, would you use it? Why? All the test subjects thought the program was helpful. All but one said they would like to use a similar program if they were in the market for a tractor. The one farmer who was not sure if he would run the program said that his mind was already made up as to what kind of tractor he wanted and thus this program was not that useful to him. The others gave reasons for wanting to run the program that ranged from getting the information at a fraction of the time it would take without the program to the fact that the information provided was unbiased. Additional reasons presented were that the program will help to narrow the choices that should be considered and possibly provide a means for deciding among these choices, or even suggest a tractor that had not previously been considered. How could this program be improved? What changes would you like to see or what additions to the program are needed? The most common response was the need to include economic factors in the selection process. Another important suggestion was to tie into a program that calculated the required horsepower to provide the user with a complete evaluation package. SUMMARY AND CONCLUSIONS The conclusions of this project are: 1. A user friendly system was developed, based on Vol. 5(2):June 1989 125 a procedure from Rider et al. (1979) to aid the farmer in tractor selection. 2. The expert system evaluated tractors ranging from 75 kW to 300 kW (100 hp to 400 hp) utilizing 113 Nebraska Tractor Tests from 1980 to 1986. 3. The evaluations of the system by six farmers and a farm equipment dealer were favorable. According to the reviewers: a. The system was easy to use. b . The systems selection decisions were valid. c. The system was helpful and worth the time required for its use. References 1. Barrett, J. R., J. B. Morrison and L. F. Huggins. 1985. Artificial intelligence and expert systems in agricultural research and education. ASAE Paper No. 85-5516. St. Joseph, MI: ASAE. 2. Chen, L. H. 1987. A microcomputer model for selection of machines and tractors. ASAE Paper No. 87-1050. St. Joseph, MI: ASAE. 3. EXSYS, Inc. 1985. EXSYS: Expert system development package. Albuquerque, NM 87194. 4. Gaultney, L., S. Harlow and W. Ooms. 1987. An expert system for diagnosing problems in hydraulic systems. ASAE Paper No. 87-1025. St. Joseph, MI: ASAE. 5. Freesmeyer, S. R. and D. R. Hunt. 1985. A farm machinery selection program for personal computers. ASAE Paper No. 85-1022. St. Joseph, MI: ASAE. 6. Hunt, D. 1983. Farm power and machinery management. Ames: Iowa State University Press. 7. Jacobson, B. K., J. W. Jones and P. Jones. 1987. Tomato greenhouse environment controller: Real-time expert system supervisor. ASAE Paper No. 87-5022. St. Joseph, MI: ASAE. 8. Jones, P. and J. Halderman. 1986. Management of a crop research facility with a microcomputer-based expert system. Transactions of the ASAE 29(l):235-242. 9. Kline, D. E., D. A. Bender and B. A. McCarl. 1987. Farm- level machinery management using intelligent decision support systems. ASAE Paper No. 87-1046. St. Joseph, MI: ASAE. 10. Rider, A. R., K. Von Bargen and P. Jasa. 1979. Selection factors to quantitatively evaluate tractors. ASAE Paper No. 79-1526. St. Joseph, MI: ASAE. 11. Wolfgram, D. D., T. J. Dear and C. S. Galbraith. 1987. Expert systems for the technical professional. New York: John Wiley & Sons. APPENDIX A TRACTOR SELECTION WORKSHEET (from Rider et al., 1979) TRACTOR SELECTION WORKSHEET Tractor Manufacturer Nebraska Tractor Test Report Number FUEL EFFICIENCY HP-HR/GAl Drawbar Performance 75% Pull or 50% Pull at Maximum Power 50% Pull at Reduced Engine Speed TOTAL FUEL EFFICIENCY INOEX = T 0^A L = (line h) LUGGING ABILITY IN RATED GEAR Increase in % Pull at 90% Crankshaft Speed Increase in % Pull at 80% Crankshaft Speed Increase in % Pull at 70% Crankshaft Speed 2 = 1 = (line i) TRANSMISSION Speed Range (MPH) 4.0-4.A Number of Gears If 0 gears, enter 0 If 1 or more gears, enter 2 If 0 gears, enter 0 If 1 or more gears, enter 2 If 0 gears, ente If 1 or more gea If 0 gears, enter I If 1 gear, enter 1 If 2 or more gears ter 2 TOTAL TRANSMISSION INDEX ; TRACTOR SOUND LEVEL Maximum Sound Level Measured at operator ear a) If measured dB(A) less than 85, enter 10 or b) If measured dB(A) more than 95, enter 0 or c) If measured dB(A) between 85 and 95, enter 95 - Measured dB(A) = 95 - OIL CONSUMPTION To Motor Drained from Motor DIFFERENCE Oil Consumption = Ma SOUND LEVEL INDEX = GAL DIFFERENCE x 4000 ximum Drawbar Total Hours Horsepower * Engine Operated Oil Consumption Index a) If oil consumption is 1.0 or greater, enter 0 or b) If oil consumption is less than 1.0, enter 1 REPAIRS AND ADJUSTMENTS Repairs and Adjustment Index a) If significant repairs or adjustments, enter 0 or b) If no significant repairs or adjustments, enter 1 ABILITY TO SUSTAIN WORKLOAD Workload Index a) If good quality manufacturer and good model, enter 2 or b) If either the manufacturer or model has not performed to expectations, enter 1 or c) If either the manufacturer or model has a poor performance record, enter 0 DEALER LOCATION Dealer Location Index a) If dealer is in primary trade center and less than 50 miles from farm, enter 2 or b) If dealer is not in primary trade center and less than 50 miles from farm, enter 1 or c) If no dealer is located within 50 miles from farm, enter 0 DEALER PARTS INVENTORY Dealer Parts Index If two dealers within 50 miles from farm adequate parts inventory, enter 2 If only one dealer within 50 miles from farm maintains adequate parts inventory, enter 1 If no dealer with 50 miles from farm maintains adequate parts inventory, enter 0 a) b) c) maintain DEALER REPAIR SERVICES Dealer Service Index If service shop with qualified service personnel equipment is available within the primary trade enter 2 If service shop with qualified service personnel equipment is not available within the primary tr center, but within 50 miles from farm, enter 1 If no service shop with qualified service person equipment is available within 50 miles from the enter 0 RELAIABILITY AND DEALER SERVICES Oil Consumption Index Repairs and Adjustment Workload Index Dealer Location Index Dealer Parts Index Dealer Service Index COMPARATIVE TRACTOR INDEX Fuel Efficiency Index (from Lug Index (from Transmission Index (from Sound Level Index (from Service Index (from Index SERVICE line h line i line j line k line s ( from ( from (from (from ( from (from TOTAL INDEX Line Line ine Line Line Line = T m) n) °) p) q) r) DTAL Farmer Weighting Factor we ; : 0 to 5 = = = = = TOTAL lghted Index (line kj (line m) (line oj (line p) (line q) (line O (line s) C0MPARATI0N TRACTOR INDEX = TOTAL 126 APPLIED ENGINEERING in AGRICULTURE Iowa State University From the SelectedWorks of Steven A. Freeman 1989 An Expert System for Tractor Selection tmp7YT7DY.pdf 
work_w7624tydubbvncyg3hfuhqhfta	----	PII: S0098-1354(00)00590-1 Computers and Chemical Engineering 24 (2000) 2339 – 2350 Use of expert systems for the synthesis of downstream protein processes M. Elena Lienqueo *, Juan A. Asenjo Department of Chemical Engineering, Centre for Biochemical Engineering and Biotechnology, Uni6ersity of Chile, Beaucheff 861, Santiago, Chile Abstract This paper describes recent developments in rational process design and their application in biotechnology for large-scale downstream processes. For implementing an expert system, it is necessary to divide the separation process in two parts: recovery and purification, because the information and available heuristic knowledge are different in each part. In the case of the recovery process, the implementation of an expert system, called Prot – Ex, only use heuristic rules from literature and human experts. The sequences suggested were tested with a real example and it gave a satisfactory result. For the purification process two criteria have been defined for selecting the optimal sequences, the SSC criterion and the purity criterion. Both criteria were implemented in an expert system, Prot – Ex – Purification. This expert system was validated with experimental examples, in two cases and, the sequences suggested were adequate and valid. However, the sequences suggested by purity criterion have lesser steps than the sequences suggested by SSC criterion and the purity algorithm spends lesser time and computational resources than SSC algorithm, then the use of purity criterion is more recommendable for selection of purification sequences. Finally, we consider the proposed sequences found by the expert system a very good starting point for industrial process selection, a clear case of ‘expert amplification’. © 2000 Elsevier Science Ltd. All rights reserved. Keywords: Protein recovery; Protein purification; Downstream processing; Bioseparation; Expert system www.elsevier.com/locate/compchemeng 1. Introduction In the modern biotechnology industry, after success- ful fermentation or enzyme reactions the desired product must be separated and purified. The necessary steps to obtain these are commonly known as down- stream processing (DSP), which can account for up to 60% of the total cost, excluding the cost of the pur- chased raw materials (Lee, 1992). DSP is usually com- posed of a sequence of recovery and purification operations, whose final aim is to obtain the required protein at a prespecified level of purity. The recovery process is characterised by the objective of obtaining the product in solution from the production system. Purification takes the multiprotein solution and purifies the individual protein product to a high purity level, which is generally more than 90% pure. In addition, on a large scale, it is necessary to obtain the highest possible yield while minimizing the resources utilised and hence the cost (Asenjo, 1990). The main steps in a large-scale protein separation process are usually not more than eight, they are shown in Fig. 1. The principal unit operations used in DSP are shown in Table 1. At present, the bio-product markets are becoming very competitive. Therefore, it is important to choose the optimal flowsheet as early as possible, since once the process has been approved by regulatory agencies, its characteristics are fixed and are expensive to change. Thus, it becomes necessary to use rational design tools for this purpose. In order to achieve this we can recognise two different tasks: 1. Selection of separation sequence or flowsheet generations. 2. Design of individual unit operations. In this paper we will center our attention in the first task. Selection of separation sequences is a complex procedure in which a design evolves from a preliminary to a final stage in a trial and error fashion by repeatedly revising and refining the initial assumptions and restric- � II Pan American Workshop in Catalysis and Process Systems Engineering, September 2 – 3, 1999, Santa Fe, Argentina * Tel.: + 1-405-3255811; fax: + 1-405-3255813. E -mail address: bagajewicz@mailhost.ecn.uoknor.edu (M. Baga- jewicz). 0098-1354/00/$ - see front matter © 2000 Elsevier Science Ltd. All rights reserved. PII: S0098-1354(00)00590-1 M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 23502340 tions. For protein recovery processes, the flowsheet generations always have the same type of unit opera- tions (see Table 1). The problem that has to be solved is, e.g. choosing between alternative operations. For example, for cell disruption it is necessary to select between a homogenizer and a bed mill or for cell separation between cross-flow microfiltration and cen- trifugation. This selection is more or less done using heuristics, i.e. using rules of thumb to arrive at a rapid and reliable specification of equipment type (Leser & Asenjo, 1992). Another way to design is using appropri- ate mathematical correlations and models for optimising the task. On the other hand, for the protein purification pro- cesses the flowsheet consists of a chromatographic se- quence (see Table 1). This sequence has to be satisfied with maximum yield and a minimum number of steps (1, 2 or 3). This part of the process is undertaken in classical chemical process engineering, finding a rigorous solution using numerical methods like mathematical optimising techniques (Jennings, Teo, Wang & Yu, 1995) or using an expert systems approach (Eriksson, Sandahl, Brewer & Osterlund, 1991; Forslund, 1995). For the selection of an optimal sequence, purely mathematical techniques have limited use in biotechnology because of a lack of useful design equations and databases. The expert sys- tems approach is more attractive because it allows the use of empirical knowledge that is not rigorous in nature and is typical of that used by experts (Asenjo & Maugeri, 1992). In this paper, we show the implementa- tion and validation of an expert system for recovery operations and another one for purification processes. Fig. 1. General flowsheet for downstream processing of proteins. Table 1 Principal steps and unit operations used in downstream processing StepStages Unit operations Protein reco6ery Cell separation Centrifugation, membrane process (filtration, microfiltration), two phase aqueous partitioning Homogenisation, bead milling, chemical and enzymatic lysis and permeablizationCell disruption (only for intracellular proteins) Debris separation (only for Centrifugation, membrane processes (microfiltration, ultrafiltration), aqueous two-phase partitioningintracellular proteins) Concentration Ultrafiltration, precipitation Protein purification Pre-treatment or primary isolation Batch adsorption, ion exchange chromatography, affinity adsorption, hydrophobic interaction chromatography, aqueous two-phase partitioning Hydrophobic interaction chromatography (HIC), high resolution ion-exchangeHigh resolution purification chromatography, affinity chromatography, metal chelate chromatography (IMAC) dye-ligand chromatography HPLC Polishing of final product Gel filtration (GF) HPLC, ion-exchange chromatography M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 2350 2341 Table 2 Potential constraints ObservationsConstraints The product is only stable in a range of pHsRange of pH (e.g. 4–9) The product is stable only up to certainRange of temperature temperature Elimination of proteases, because they couldProteases degrade the protein product Inclusion bodies Inclusion bodies have to be solubilized and refolded for the production bulk industrial enzymes this is not the case. 1. How big is the flow rate? Bigger than 3 m3/h. Smaller than 3 m3/h. 3.2. Characterisation of starting material The second important point is the characterisation of the starting material. The process design will mainly depend on the physical, chemical and biochemical properties of the contaminating materials in the original broth and those of the protein that will constitute the final product. The properties of the starting material will be partially determined by: 1. Its fermentation source. Bacterial. Yeast. Mammalian cell. 2. How big is the cell concentration? 3. The type of cultivation medium used. Presence of albumin. Calf serum. Presence of proteases. Solid bodies like whole cells or cell debris. 4. Localization of the product. Intracellular. Extracellular. 5. Physicochemical properties of the product and the contaminant proteins in the final protein mixture in solution. Surface charge at different pHs and pI. Surface hydrophobicity. Molecular weight. Biospecificity toward certain ligands. 6. Stability of the final product is also of extreme importance as this will affect the types of operations that can be used as well as the conditioning and processing times that can be afforded. 3.3. Define possible separation steps and constraints It is necessary to have a database with all possible separation steps (Table 1) and to consider all potential constraints (Table 2). 3.4. E6aluation of potential process integration Finally, the production process should try to inte- grate as much as possible, upstream, fermentation and downstream process. Asenjo and Leser (1996) have described possible areas and levels of process integra- tion. Process integration can involve either the whole process, specific parts of it or particular unit operations as shown in Fig. 2. Examples of process integration between fermentation and downstream and between downstream operations are shown in Table 3. 2. General aspects for designing a protein separation process The design of a protein recovery and purification process shares many characteristics with other engineer- ing design activities. To design a process or an opera- tion requires: 1. To satisfy of a number of constraints, such as purity, quality, process temperature and pH, desired yield, etc. 2. To know properties of the materials, for instance chemical and biochemical properties, thermody- namic and fluid properties of the process material. Using the information and knowledge of (1) and (2) to propose a sequence of equipment interconnected in a particular order. 3. Basic information for designing a separation process The design of a process to economically purify a protein, maintaining a high yield, yet obtaining a virtu- ally pure product while minimizing the cost, requires four main considerations: 1. Defining final product. 2. Characterisation of starting material. 3. Define possible separation steps and constraints. 4. Evaluation of possible process integration. 3.1. Defining final product It is necessary to define the final product and to have information on its used. For instance: 1. How the product is going to be used? Industrial. Diagnostic. Laboratory reagent. Therapeutic. Questions regarding the use and purity are vital (e.g. 80, 99 or 99.98% pure) as well as allowable ranges of specific impurity concentrations. With therapeutic proteins any impurities have to be minimized whereas M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 23502342 Fig. 2. The possible areas of process integration. Fig. 3. The basic architecture of an expert system. The knowledge showed in the previous sections can be used for defining expert system for the rational selection of downstream protein processes. 4. Expert systems Expert systems are programs that attempt to solve problems in a way similar to how a human expert would solve them. They incorporate ‘rules of thumb’ that experts in the field have developed through years of experience. The problems tackled are not necessarily solved in a procedural fashion, they are often vague, complex, and can contain incomplete or inexact infor- mation (Nebendahl, 1988). Expert systems contain three basic elements: a knowledge base, an inference engine, and a user inter- face. The architecture of an expert system is shown in Fig. 3. The knowledge base contains the information, which the program uses to reach decisions. The infer- ence engine is the program that manipulates the knowl- edge base to reach these decisions. Finally, the user interface allows the program and the user to communi- cate with each other in an effective manner (Harmon & King, 1985). There are software systems, called shells, used for the manipulation of heuristics, databases, and algebraic equations (like design equations). For in- stance, Nexpert Object™ and Clips. There are many examples of the use of expert systems in chemical and biochemical engineering (Siletti, 1989; Mulholland, Walker, Manis, Hinriks, Buydens, Blaffer & Schoen- maker, 1991; Jakus, 1992; Forslund, 1995). 5. Implementation of expert system The implementation of an expert system for separa- tion processes for proteins was divided in two different parts: 1. A first expert system, called Prot – Ex, for recovery process selection. 2. A second one, Prot – Ex – Purification, for purifica- tion process selection. This division was carried out because both systems need different kind of information and the available heuristic knowledge is different, too (Asenjo, Herrera & Byrne, 1989). Prot – Ex uses only heuristic knowledge whereas Prot – Ex – Purification needs an important amount of quantitative data to make a good selection. In a consultation both expert systems are integrated and controlled by the main expert system. The main expert system: � Receives the information from the user, database or other sources. � Organizes and gives the information that each sub- expert system needs for suggesting a sequence. � Receives the sequences suggested by each sub-expert system � Gives the global sequences suggested to the user or other programme. Table 3 Examples of possible process integration Level of integration Action Cell immobilisation, cell recycling using membranes, use ofFermentation-downstream (the adequate design of the fermentation step can provide conditions to facilitate the downstream) expanded bed chromatography after fermentation Use of liquid-liquid separation, use of expanded bed for proteinInside downstream (recovery and purification) recovery, use of hydrophobic interaction chromatography after precipitation with ammonium sulphate M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 2350 2343 Fig. 4. Diagram of main expert system and relation with sub-expert systems Prot Ex and Pro Ex purification. The general diagram of the main expert system and sub-expert system with its relation is shown in Fig. 4. 5.1. Prot – Ex for reco6ery process The expert system for the recovery process was im- plemented using only heuristic knowledge. The heuristic rules were obtained from the literature and renowned industrial experts were consulted to clarify specific points for which knowledge is missing or there is am- biguous information. To do this, it was necessary to provide the experts with a questionnaire (Table 4). For recovery processes, it is only necessary to choose between two or three options. For example, for separa- tion of cells If is necessary to select between centrifuga- tion and cross-flow microfiltration. This selection is based on the following variables: � Thermal sensitivity of the product. � Shear liability of cells. � Size of the cells. � Process throughput. � Density difference between solid and liquid phase. � Cell concentration in the broth. Depending of the value of each variable it is possible to choose one or the other option. Considering the previous points, the expert system Prot – Ex was implemented. It has more than six hun- dred logical rules to emulate human reasoning. Its implementation was carried out in a commercial pro- gramme, Nexpert Object™ from Neuron Data. Consul- tations were carried out to validate the expert system Prot – Ex. The basic information used for the expert system is shown in Table 5. Table 6 displays a compari- son between a sequence suggested by Prot – Ex and a good process for recovery of somatotropin produced in Escherichia coli (Wheelwright, 1991). For this example, the recovery sequences suggested by Prot – Ex and the industrial process were very simi- lar. The difference with the published process is in step 1, where centrifugation was used instead of microfiltra- tion. For small cells such as E. coli microfiltration usually has clear economic advantage (Asenjo & Table 4 Typical questions presented to biotechnology experts for implementa- tion of the expert system for recovery processes (Leser, 1996) QuestionAspect Therapeutic What would your recommendation be if the product utilisation is therapeutic?product Are the common ‘broth conditioningCell separation processes’, like heating, freezing, coagulation, flocculation, enzymatic or other used to facilitate cell separation in recombinant DSP or not? Cell disruption Would you agree that the best suggestion of a mechanical method for breaking cells is ‘high pressure homogenizer’ for most real situations? Precipitation of What are the preferred methods for precipitation of nucleic acids?nucleic acid M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 23502344 Table 5 Basic information for designing a purification process of Somatotropin (Lienqueo et al., 1996) Defining final product Protein’s name Somatotropin 7.86 (Estimated)pI 22.0 KDa (Estimated)Molecular weight Surface hydrophobicity 0.93 M (Estimated) 25.0 mg/mlConcentration Titration curve pH 5.04.0 6.0 7.0 8.0 2.42 1.034.77 0.12Charge (coulomb/molecule)10−25 −0.03 Localization of the product Intracelullar, inclusion bodies IndustrialUtilization 4–9pH stability range 94%Final purity level Not knownSpecific affinity Characterisation of starting material E. coliFermentation source Patrick, 1990). Therefore, we considered that the se- quences suggested by Prot – Ex are adequate and it can be used as a starting point for the selection of a good industrial process. 5.2. Prot – Ex – Purification for selection of purification processes The selection of an efficient purification process is a critical step in downstream processes. These steps are not usually chosen in a rational manner, the common method being ‘trial and error’. Therefore, the use of some basic heuristic rules for separation processes has been proposed. These rules are: (Asenjo & Patrick, 1990; Prokopakis & Asenjo, 1990): 1. Choose the separation process based on the differ- ent physicochemical properties, such as: 1.1. Surface charge as a function of pH (titration curve) and pIs. 1.2. Surface hydrophobicity. 1.3. Molecular weight. 1.4. Biochemical properties such as biospecificity with different ligands. 1.5. Stability at different temperatures and pHs. 2. Eliminate those contaminant proteins and com- pounds that are found in greater percentage first. 3. Use a high-resolution step, as soon as possible. The chromatographic techniques ranked according to their efficiency are: 3.1. Affinity. 3.2. Ion exchange. 3.3. Hydrophobic interaction. 3.4. Gel filtration. 4. Do the most arduous purification step at the end of the process (final polishing). Considering those rules two selection criteria have been defined: Seperation Selection Coefficient (SSC) criterion and Purity criterion. The first one, SSC criterion, was developed by As- enjo and collaborators (Asenjo, 1990; Asenjo & Maugeri, 1992; Leser & Asenjo, 1992). The SSC crite- rion considers that the key parameter is SSC, which caracterizes the ability of the purification operation to separate two proteins. SSC value is defined as: SSC = DFhUi (1) Deviation factor (DF), this variable relates the differ- ence between a property of the target protein (for example, the dimensionless retention time in a specific chromatographic technique) and the same property of the contaminant. DF = �KDproduct − KDcontaminant� (2) Dimensionless retention time (KD), this variable rep- resents the behaviour of the proteins in a separation carried out by gel filtration, ion exchange or hydropho- bic interaction chromatography. Mathematical relation- ships for predicting KD have been derived using the physicochemical properties of proteins (Lienqueo, Leser & Asenjo, 1996). Table 6 Comparison between sequences suggested by Prot – Ex and industrial process for recovery of somatotropin (Lienqueo et al., 1996) Sequence suggested bySteps Industrial process Prot – Ex Centrifugation1 Crossflow microfiltration 2 High-pressureHigh-pressure homogenizationhomogenization 3 Centrifugation-pellet wash Disk centrifugation 4 SolubilizationSolubilization Renaturation5 Renaturation 6 Microfiltration Ultrafiltration 7 Concentration and diafiltration M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 2350 2345 Fig. 5. Algorithm of SSC criterion. Efficiency factor (h ) This parameter gives account of the unequal capability of the different separation pro- cesses to separate different proteins. Its value is con- stant for each type of separation and the chromatographic material used and have been mea- sured experimentally (Lienqueo et al, 1996). Concentration factor (Ui ) Its value represents the relative concentration of each contaminant. It will af- fect the selection criteria. In this way contaminants which are present in a high concentration have to be eliminated first. Ui = concentration of contaminanti concentration of all contaminants (3) The amount of contaminant eliminated after a chro- matographic step has been mathematically calculated M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 23502346 for different situations (Lienqueo, Salgado & Asenjo, 1999) Then the SSC criterion chooses as best chromato- graphic operation the one that has the highest SSC value (Leser, Lienqueo & Asenjo, 1996). After deter- mining which chromatographic technique gives the maximum SSC value, it is necessary to calculate the new concentration of all contaminant after the chro- matographic technique selected has been applied and construct a new database of contaminants concentra- tion. With this new database the system calculates the purity level and this value is compared with the level of purity required. The optimal sequence of steps is chosen until the required level of purity is reached. The al- gorithm used is shown in Fig. 5. Considering that the most important value is the final purity level and that now we had developed and al- gorithm to calculate the purity after a purification step, Fig. 6. Algorithm of purity criterion. M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 2350 2347 Table 7 Main variables used for SSC and purity criteria (Lienqueo et al., 1999) Variable Symbol Definition SSC criterion SSC = DFhUiSeparation SSC selection coefficient DF = �KDproduct−KDcontaminant�Deviation factor DF KD Relationships with physicochemicalDimensionless retention time properties of the proteins. (Lienqueo et al., 1999) Efficiency factor Values for each chromatographich techniques shown in Table 8 Peak width Values for each chromatographicS techniques shown in Table 8 UiConcentration Ui factor = concentration of contaminant i concentration of all contaminantsPurity criterion Pj PjPurity = concentration of the target protein Sconcentration of the proteins present one that has the least number of total steps. Consider- ing these criteria an expert system was implemented, Prot – Ex – Purification, and its implementation was car- ried out in the same shell than Prot – Ex. For testing these criteria, consultations were carried out to validate the expert system Prot – Ex – Purification using both criteria. Table 9 and Table 10 show the basic information used for purification of bovine serum albu- min (BSA) and b-glucanase, respectively and Table 11 and Table 12 show a comparison between sequences suggested by Prot – Ex – Purification and experimental sequence for both examples. 5.3. Comparison between SSC and Purity criteria In both examples the sequences suggested are valid and adequate. Nevertheless, in the first case, the se- quence suggested by the purity criterion has one fewer step than that suggested by the SSC criterion, that mean less capital cost. This situation takes place be- cause the SSC criterion considers the contaminant that gives the highest SSC value for one protein only. The purity criterion chooses the optimum chromatographic step considering all the contaminant proteins present. Hence, this situation can occur when there are several contaminants present in similar quantities as was the case in this example. On other hand, the purity criterion spends less time and computational resources than the SSC criterion, since the purity algorithm gives the optimal sequences in less time than the SSC algorithm. 5.4. Comparison between sequences suggested and experimental sequences In both examples the sequences suggested result in experimentally valid ones and are a good starting point for an industrial process. There are differences between the estimated and experimental final purity level, which are less than 20%. These differences could originate in different causes: � Hydrophobic interaction between the protein, this effect was not considered in the development of correlations for estimated dimensionless retention time for hydrophobic interaction chromatography (HIC) (Table 8). � Asymmetric chromatographic peaks, the peak shape could be different to a triangle shape. The differences between estimated and experimental purity level could be minimised: � Improving the prediction of dimensionless retention time for HIC. � Using a Gaussian approximation in the representa- tion of the chromatographic peak instead of triangle simplifications. Table 8 Expressions and parameters used for SSC criteria and purity criteria (Lienqueo et al., 1999) Efficiency factor h Peak width SChromatographic techniques Anion exchange 1.00 0.15 Cation exchange 0.151.00 0.86Hydrophobic interaction 0.22 Gel filtration 0.66 0.46 we implemented the purity criterion as a possible selec- tion criterion. This criterion compares the final purity level obtained after a particular chromatographic tech- niques has been applied. The purity concept is defined as: Purity = concentration of the target protein Sconcentration of the proteins present (4) After determining which chromatographic technique gives the highest purity level, the system chooses this as the technique to use at this step. A sequence of steps is chosen until the required level of purity is reached. Finally, the system creates a list with the defined se- quence of operations. The algorithm is shown in Fig. 6 The main parameters, variables and correlations used for SSC and purity criteria are summarized in Tables 7 and 8. Both criteria consider that the cost of each chromatographic option is equal, except for affinity chromatography. Therefore, the optimal system is the M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 23502348 6. Conclusions It is possible to use artificial intelligence tools, as expert systems, for selection of optimal protein separa- tion processes. For implementing an expert system, it is necessary to divide the separation process in two parts, because the information and available heuristic knowl- edge are different in each case. The first part for the recovery of proteins uses only heuristic rules from human experts and literature, the expert system was called Prot – Ex. The second part of a purification pro- cess is an hybrid expert system, which uses heuristic rules and mathematical correlations and models for selecting the optimal protein purification sequences, expert system called Prot – Ex – Purification. Both expert systems, implemented in the shell Nexpert Object™, were integrated by a main expert system. Considering the cases studied, like the recovery of somatotropin, the purification of BSA and the purifica- tion of b-glucanase, the sequences proposed by the expert system are a good starting point for a practical industrial process. Nevertheless, there are differences between estimated and experimental purity level, these could be minimized by improving correlations for pre- dicting retention time and approximation of the peak shape. In the specific case of Prot – Ex – Purification, the sequences suggested by the purity criterion have less Table 9 Basic information for desining a purification process of BSA from artificial mixture (Lienqueo et al., 1999) Defining final product Protein’s name Bovine serum albumin (BSA) 4.9p I 67.0 KDaMolecular weight Surface hydrophobicity 0.86 M Concentration 2.00 mg/ml Titration curve 5.0 6.0pH 7.04.0 8.0 −0.14 −1.16Charge −1.681.03 −2.5 (Coulomb/molecule) 10−25 Localization of the Extracellular product IndustrialUtilization pH stability range 4–9 Final purity level 94% Not knownSpecific affinity Characterisation of starting material Artificial mixture (BSA, ovalbumin, thaumatin, soybean trypsin inhibitor)Fermentation source Hydrophobicitymw (Kda)Contaminant Titration curve (coulomb/molecule)10−25Initial concentration (M)(mg/ml) pH 4.0 pH 5.0 pH 6.0 pH 7.0 pH 8.0 −2.3643.8 0.54 1.40 −0.76 −1.65Ovalbumin −2.202.0 SBTI 2.0 1.22 −0.76 −1.54 −2.17 −2.1324.5 0.90 0.89 1.94 1.90 1.98 1.87Thaumatin 2.0 0.9122.2 Table 10 Comparison between sequences suggested by Prot – Ex – Purification and experimental process for purification of BSA from a mixture of four proteins (Lienqueo et al., 1999) Sequence suggested by Prot – Ex – PurificationStep Experimental validation process Purity criterionSSC criterion Chromatography step Purity Chromatography step PurityPurity Chromatography step 64%Cation exchange pH 6.0 Anion exchange pH 7.0 64%33% Anion exchange pH 7.01 Hydrophobic interaction Hydrophobic interaction 80%49% Hydrophobic interaction2 95% 97%3 Anion exchange pH 7.0 M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 2350 2349 Table 11 Basic information for designing a purification process of b-1,3-glucanase (Lienqueo et al., 1999) Defining final product Enzyme name b-1,3-glucanase (bgl IIa) 4.1pI 31 KdaMolecular weight 0.00 MSurface hydrophobicity 0.62 mg/mlConcentration Titration curve 5.0 6.0 7.0 8.04.0pH 1.46 −0.62Charge −1.02 −2.33 −2.52 (Coulomb/molecule) 10−25 ExtracellularLocalization of the product Utilization Industrial pH stability range 4–9 70%Final purity level Not knownSpecific affinity Characterisation of starting material B.subtilus ToC46Fermentation source Contaminant mw (Kda) HydrophobicityInitial concentration pH 4.0 Titration curve (coulomb/molecule)10−25 (mg/ml) (M) pH 5.0 pH 6.0 pH 7.0 pH 8.0 41.0 1.5 1.461 0.262.74 −0.87 −1.65 −2.04 2 2.74 32.9 1.5 0.00 −2.70 −3.51 −3.51 3 35.50.25 0.2 −0.55 −0.22 −0.73 −1.82 62.5 0.0 −1.060.42 −1.174 −2.79 −3.32 40.6 0.0 −0.555 −0.220.09 −0.73 −1.82 69.6 0.0 −0.550.09 −0.226 −0.73 −1.82 0.257 40.6 0.0 1.46 −0.47 −1.06 −1.04 69.6 0.0 1.460.25 −0.478 −1.06 −104 Table 12 Comparison between sequences suggested by Prot – Ex – Purification and experimental process for purification of b-glucanase (Lienqueo et al., 1999) Sequence suggested by Prot – Ex – PurificationStep Experimental validation process Purity criterionSSC criterion Purity Chromatography step Purity Chromatography stepChromatography step Purity 32% Hydrophobic interaction 32%Hydrophobic interaction Hydrophobic interaction1 35% 70% Anion exchange pH 6.5 70% Anion exchange pH 6.52 60–70%Anion exchange pH 6.5 steps than the sequences suggested by SSC criterion, then the Purity criterion is more recommendable. On other hand, the use of expert systems for the synthesis of downstream protein processes is a clear case of ‘expert amplification’. Heuristic knowledge from experts was complemented with databases and design equations. To implement such a solution more globally there is a lack of databases of protein properties. With ad- vances in the area of proteomics and protein engineer- ing in a few years there could be sufficient information for such an approach. An effort should be made to generate such data. Acknowledgements Financial assistance from Fundación Andes and Vicerrectoria Académica, Universidad de Chile (Beca PG/080/97). This project was also supported by Proyecto Catedra Presidencial en Ciencias. M.E. Lienqueo, J.A. Asenjo / Computers and Chemical Engineering 24 (2000) 2339 – 23502350 References Asenjo, J. A., Herrera, L., & Byrne, B. (1989). Development of an expert system for selection and synthesis of protein purification processes. Journal of Biotechnology, 11, 275 – 298. Asenjo, J. A. (1990). Selection of operations in separation processes. In J. A. Asenjo, Separation processes in biotechnology (pp. 3 – 16). New York: Marcel Dekker. Asenjo, J. A., & Patrick, I. (1990). Large-scale protein purification. In E. L. V. Harris, & S. Angal, Protein purification applications: a practical approach (pp. 1 – 28). Oxford: IRL Press. Asenjo, J. A., & Maugeri, F. (1992). An expert system for selection and synthesis of protein purification processes. In P. Todd, S. K. Sikdar, & M. Bier, Frontiers in bioprocessing II (pp. 358 – 379). Washington: American Chemical Society. Asenjo, J. A., & Leser, E. W. (1996). Process integration in biotech- nology. In M. Verrall, Downstream processing of natural products (pp. 123 – 137). Chichester: Wiley. Eriksson, H., Sandahl, K., Brewer, J., & Osterlund, B. (1991). Reac- tive planning for chromatography: chemometrics and intelligent laboratory systems. Laboratory Information Management, 13, 185 – 194. Forslund, G. (1995). Designing for flexibility: a case study. Expert Systems, 12 (1), 27 – 37. Harmon, P., & King, D. (1985). Expert systems: artificial intelligence in business (p. 283pp). New York: Wiley. Jakus, W. (1992). Artificial intelligence in chemistry. Collect Czech Chemistry Communication, 57, 2414 – 2442. Jennings, L. S., Teo, K. L., Wang, F. Y., & Yu, Q. (1995). Optimal protein separation. Computers & Chemical Engineering, 19 (5), 567 – 573. Lee, J. M. (1992). Biochemical engineering (p. 320). Englewood Cliffs, NJ: Prentice-Hall. Leser, E. W., & Asenjo, J. A. (1992). Rational design of purification processes for recombinant protein. Journal of Chromatography, 584, 43 – 57. Leser, E. W. (1996) Prot – Ex: an expert system for selecting the sequences of processes for downstream purification of proteins, PhD thesis, University of Reading, England. Leser, E. W., Lienqueo, M. E., & Asenjo, J. A. (1996). Implementa- tion in an expert system of selection rationale for purification processes for recombinant proteins. Annals of the New York Acadamy of Sciences, 782, 441 – 455. Lienqueo, M. E., Leser, E. W., & Asenjo, J. A. (1996). An expert system for selection and synthesis of multistep protein separation processes. Computers & Chemical Engineering, 20, 189 – 194. Lienqueo, M. E., Salgado, J. C., & Asenjo, J. A. (1999). An expert system for selection of protein purification processes: experimental validation. Journal of Chemical Technology & Biotechnology, 74, 293 – 299. Mulholland, M., Walker, N., Manis, F., Hinriks, H., Buydens, L., Blaffer, T., & Schoenmakers, P. (1991). Expert system for re- peatability testing of high-performance liquid chromatographic methods. Journal of Chromatography, 550, 257 – 266. Nebendahl, D. (1988). Expert systems: introduction to the technology and applications (p. 211). New York: Wiley. Prokopakis, G. J., & Asenjo, J. A. (1990). Synthesis of downstream processes. In J. A. Asenjo, Separation processes in biotechnology (pp. 571 – 601). New York: Marcel Dekker. Siletti, C. A. (1989). Computer aided design of protein reco6ery pro - cesses, PhD thesis, (pp. 379). Massachussets Institute of Technol- ogy, Cambridge, USA. Wheelwright, S. M. (1991). Protein purification: design and scale up of downstream processing (p. 228). München: Carl Hanser. . 
work_w7ccj62ltvfrjcam7jnefiuiri	----	A bipolar consensus approach for group decision making problems Open Archive TOULOUSE Archive Ouverte (OATAO) OATAO is an open access repository that collects the work of Toulouse researchers and makes it freely available over the web where possible. This is an author-deposited version published in : http://oatao.univ-toulouse.fr/ Eprints ID : 15535 To link to this article : DOI : 10.1016/j.eswa.2014.09.061 URL : http://dx.doi.org/10.1016/j.eswa.2014.09.061 To cite this version : Bouzarour-Amokrane, Yasmina and Tchangani, Ayeley and Pérès, François A bipolar consensus approach for group decision making problems. (2015) Expert Systems with Applications, vol. 42 (n° 3). pp. 1759-1772. ISSN 0957-4174 Any correspondence concerning this service should be sent to the repository administrator: staff-oatao@listes-diff.inp-toulouse.fr http://oatao.univ-toulouse.fr/ http://dx.doi.org/10.1016/j.eswa.2014.09.061 mailto:staff-oatao@listes-diff.inp-toulouse.fr A bipolar consensus approach for group decision making problems Yasmina Bouzarour-Amokrane a,⇑, Ayeley Tchangani a,b, François Peres a a Université De Toulouse/LGP 47, Avenue d’Azereix, BP 1629, 65016 Tarbes Cedex, France b Université De Toulouse/IUT de Tarbes, 1, rue Lauréamont, 65016 Tarbes Cedex, France Keywords: Group decision making (GDM) Consensus process Bipolarity Concordance Discordance Influence a b s t r a c t This paper addresses the collaborative group decision making problems considering a consensus pro- cesses to achieve a common legitimate solution. The proposed resolution model is based on individual bipolar assessment. Each decision maker evaluates alternatives through selectability and rejectability measures which respectively represent the positive and negative aspects of alternatives considering objectives achievement. The impact of human behavior (influence, individualism, fear, caution, etc.) on decisional capacity has been taken into account. The influence degrees exerted mutually by decision makers are modeled through concordance and discordance measures. The individualistic nature of deci- sion makers has been taken into account from the individualism degree. In order to achieve a common solution(s), models of consensus building are proposed based on the satisficing game theory formalism for collective decision problems. An application example is given to illustrate the proposed concepts. 1. Introduction Nowadays, the increasing complexity of the socio-economic, engineering and environmental management make less possible decision by a single decision maker considering all aspects of a problem (Yue, 2011a). Therefore, the majority of decision problems are considered currently in the group decision process. This pro- cess is generally characterized by the existence of two or more per- sons (i) who have different perceptions, attitudes, motivations and personality, (ii) who recognize the existence of a common problem, and (iii) attempt to reach a collective decision (Herrera, Herrera- Viedma, & Verdegay, 1996b). Solving a group decision making (GDM) problem often goes through the following phases: elicitation phase where different characteristics of the problem are defined (objectives, alternatives, attributes, etc.), evaluation phase and a selection and recommen- dation phase. In the evaluation phase, the way information is man- aged can leads to two families of aggregation approaches, we speak of input and output aggregation (Leyva Lopez, 2010) or common value tree for all decision makers and a value tree for each decision maker (De Brucker & Macharis, 2010). In the first case, aggregation is performed at the input when the decision group is invited to agree on a common set of attributes, weights and other parameters, which amounts to solving a problem as a single decision making problem. In the second case, the individual evalu- ations are represented by individual value trees solved using stan- dard process of decision support. The output aggregation is performed at the end. The present paper focuses on this second type of problem dealing with group decision making problem based on individual assessments. Decision makers’ evaluations can be represented by preference order (where the alternatives are ranked from best to worst), a util- ity function (where the alternatives are represented by real value – physical or monetary value–), or a frequently used preference rela- tion (where alternatives are evaluated by pairwise comparison) (Herrera, Herrera-Viedma, & Chiclana, 2001). Depending on the nature of the data, the certainty of decision-makers, these prefer- ences can be modeled by absolute evaluations when information is known or fuzzy evaluation based on the theory of fuzzy set, introduced by Zadeh (1965) in case of uncertainty in order to man- age human subjectivity, imprecision and vagueness. The fuzzy evaluation is used in many areas due to the pressure, lack of knowledge and/or time. Decision-makers’ evaluations are then integrated into decision resolution procedures to reach an agreement on the selection of the best solutions. Traditionally, GDM problems have been solved by applying an alternative selection process in which the prefer- ences of each decision makers over the alternatives are gathered and the best alternative or subset of alternatives is chosen (Roubens, 1997). However, as a group decision members usually come from different horizons with different specialty areas and different levels of knowledge, each group member has distinct ⇑ Corresponding author. E-mail addresses: yasmina.bouzarour@gmail.com (Y. Bouzarour-Amokrane), ayeley.tchangani@iut-tarbes.fr (A. Tchangani), francois.peres@enit.fr (F. Peres). information and sharing in general a part of the objectives with other decision members (Xu & Wu, 2011). This implies that indi- vidual assessments rarely meet (Roselló, Prats, Agell, & Sánchez, 2010; Ben-Arieh, Easton, & Evans, 2009) and the divergence of opinions can generates conflict (disagreement) and/or agreement within the group decision making. The recommendation phase in this case usually requires the establishment of a ‘‘consensus’’ building process in order to lead actors to a common decision (Khorshid, 2010). To achieve a common accord, a variety of consensus reaching processes have been proposed in recent years (Eklund, Rusinowska, & de. Swart, 2008; Gong, Forrest, & Yang, 2013; Herrera-Viedma, Cabrerizo, Kacprzyk, & Pedrycz, 2014). These approaches going from mechanist models of operational research to more sophisticated and soft computing oriented models that attempt to integrate human attitude (emotion, affect, fear, egoism, altruism, selfishness, etc.). The soft computing oriented models are used increasingly due to their ability to tolerate imprecision, uncertainty and partial truth in order to simulate human behavior with low cost (Pal & Ghosh, 2004), they allow to take into account the ambiguity in human thinking and uncertainty of the real world (Ko, Tiwari, & Mehnen, 2010). 2. Consensus building processes Basically, group decision making aims at obtaining the consent, not necessarily the agreement of the participants by accommodat- ing views of all parties involved to attain a decision that will yield what will be beneficial to the entire group (Herrera-Viedma et al., 2014). This is why the group consensus is usually considered as a total and final agreement between the decision members (Leyva Lopez, 2010). To reach a consensus, the researchers first proposed consensus approaches with the objective of reaching a full degree of agreement in the group, i.e. unanimity (Kline, 1972). The earliest approaches proposed to use group consensus functions that aggre- gate decision maker evaluations in a unique value representing the common opinion. Several aggregation methods have been pro- posed in the literature, simple average (Wheeler, Hora, Cramond, & Unwin, 1989), geometric mean (Cook & Kress, 1985), Bayesian aggregation (Bonano & Apostolakis, 1991), aggregation using the analytic hierarchy process (AHP) (see eg (Bard & Sousk, 1990; Korpela & Tuominen, 1997; Lai, Wong, & Cheung, 2002; Tavana, Kennedy, & Joglekar, 1996)), fuzzy set theory (Hsu & Chen, 1996; Kacprzyk, Fedrizzi, & Nurmi, 1992; Moon & Kang, 1999; Yue & Jia, 2012), multi-criteria decision analysis methods (Hatami- Marbini & Tavana, 2011), etc. In the best possible way, consensus should refer to unanimity of individuals because the selected solution(s) will be best represen- tative for the entire group. Traditionally way to reach a consensus propose to model process by using matrix calculus or Markov chains to model the time evolution of changes of opinions toward consensus (Coch & French, 1948; French, 1956; Harary, 1959). However, unanimity may be difficult to attain, in particular in large and diversified groups of individuals as is the case in real world settings (Herrera-Viedma et al., 2014). This resolution scheme does not take into account the agree- ment level between decision makers and some actors may not accept the final decision because their individual preferences are not taken into account sufficiently (Butler & Rothstein, 1987; Kacprzyk & Fedrizzi, 1988). For this reason, consensus reaching processes based on agreement levels were introduced as an addi- tional phase in the resolution of GDM problems (Saint, 1994), as cited in (Palomares, Estrella, Martínez, & Herrera, 2014). In this context, the concept of a soft consensus was introduced by Kacprzyk and Fedrizzi (1988) where some researchers assume that unanimity is not required in the real decision problem and employed milder definitions of consensus (Butler & Rothstein, 1987; Verma, 2009) which consider for example a unanimity minus number of persons whose don’t support the decision, per- centage of actors (%) accepting decision, etc. Generally, the soft consensus reached process is based on mul- tistage setting where the individuals assumed collaborative, change their opinions until some consensus is reached. The individual evaluation settings can be realized in discussion phases where the analyst or moderator –who is responsible for running the consensus reaching session– intervenes to guide stakeholders towards a common output solution, recommending them, based on rational arguments, to settle their preferences, and keeping the process within a period of time considered (Butler & Rothstein, 1987). In some case, the individual settings are modeled by integrating the dynamic discussion phase in the reach consensus process and thus substituting the role of the analyst. (see for example (Choudhury, Shankar, & Tiwari, 2006)). Although the latter method is becoming increasingly popular in recent years, the method where moderator running a consensus reaching process is usually more effective and efficient (Herrera-Viedma et al., 2014). In the present paper, a new adaptive consensus reached process based on semi-automated feedback mechanism –where analyst can intervenes– is developed. Considering a bipolar framework, initial decision makers’ preferences are represented by bipolar measures that express the degree of supportability and rejectabil- ity of alternatives avoiding compensations. To more realistic model, human behavior aspects (positive affect, negative affect, selfish, prudence, etc.) are integrated in the evaluation and recommendation phases of proposed approach. By considering social ties, the mutual influence of positive and neg- ative interactions of the group members are integrated through individualism degree of each one. The weight of the decision makers which was the subject of several studies (Yue, 2011b; Yue, 2012a; Yue, 2012b; Yue & Jia, 2012) is also treated in this paper. To our knowledge and as underlined in (Herrera-Viedma et al., 2014), although some authors introduce the decision maker impor- tance degrees in the aggregation phase of actors’ opinions (Herrera, Herrera-Viedma, & Verdegay, 1996a; Herrera, Herrera-Viedma, & Verdegay, 1997a; Lee, 2002), no one considers them in the recom- mendation phase when advising to the decision actors how to change their preferences to increase the consensus level. To rem- edy this, the present contribution integrates importance degrees of actors in proximity and bipolar consensus measures to adjust the actors’ preference depending on his/her own knowledge level about the problem. Considering that local preferences can be represented by a set of satisficing and non-dominated alternatives, developed consensus processes are defined considering two cases: when the local pref- erences of actors cannot be modified and in the case when modifi- cations are possible. In the first case, consensus achievement is based on setting caution index and consensus measures are used in the second. The remainder of the paper is organized as follow. Section 3 traces the evolution of consensus approaches in group decision making (GDM) and give a general classification of these process. Section 4 describe global framework of bipolar modeling of deci- sion group problems when considering members interaction. The satisficing game theory used as aggregation tool is briefly described in this section. Section 4 presents proposed consensus and selection processes, an application example is given in Section 5. Eventually, Section 6 provides a conclusion and some perspectives. 3. Consensus approaches in GDM As mentioned in Herrera-Viedma et al. (2014), the first mathe- matical approaches of consensus reaching processes started with the pioneering works by French and his collaborators in the late 1940s and early 1950s (Coch & French, 1948; French, 1956). Authors employed matrix calculus to model the time evolution and reaching of the consensus process. They also describe the impact of involving people in changes that affect them through the introduction of a simple model of how a network of interper- sonal influence enters into the process of opinion formation. Draw- ing on the algebra of a Markov chain process, the consensus theory was developed in a more general form by Degroot (1974); French (1981) and Harary (1959). These initial researches describes the formation of group consensus, but do not provide an adequate account of settled patterns of disagreement. Later, many models of consensus reaching (formation) have been proposed, particularly in the realm of so called rational con- sensus where authors Lehrer and Wagner (1981) considered that a consensus reaching procedure is not just a pooling or aggregation but changes of individual preferences occur and are rationally motivated. In fuzzy environment, the first work which addresses the prob- lem of consensus reaching modeling in a fuzzy environment (Ragade, 1976) examined some applications of fuzzy set theory in the area of communications and information systems, and was followed by an increasing number of researches (see for example, (Yager, 1988; Kacprzyk, 1986; Szmidt & Kacprzyk, 2003; Fu & Yang, 2012)). Considering that the unanimous agreement is not necessary to lead a consensus, Loewer and Laddaga (1985) introduced the first approach for a soft consensus was introduced where a consensus degree, or proximity to the ‘‘ideal’’ consensus were defined. According to Loewer and Laddaga, other authors (Kacprzyk & Fedrizzi, 1988; Kacprzyk & Fedrizzi, 1989; Kacprzyk & Fedrizz, 1986) introduced the concept of a fuzzy majority using Zadeh’s fuzzy linguistic quantifier to define soft consensus measures. A new consensus model for GDM problems based on fuzzy lin- guistic preference relations was then defined in an ordinal fuzzy linguistic approach (Herrera, Alonso, Chiclana, & Herrera-Viedma, 2009; Herrera & Herrera-Viedma, 2000). As main novelty authors define two types of soft consensus measures, consensus degrees and proximity measures. Both measures are computed in the three representation levels of a preference relation: level of preference, level of alternative, and level of preference relation. The consensus degrees indicate how far the set of actors is from the maximum consensus, and the proximity measures indicate how far each indi- vidual is from current consensus labels over the preferences. In such a way, this proposal provides moderator a complete consen- sus instrument to control the consensus reaching process (Herrera-Viedma et al., 2014). The consistency measures were introduced by the same authors in (Herrera et al., 1997a) to allow analyst to guide the decision pro- cess towards more consistent solution. In order to consider decision situations in which the authors can use different representation formats to express their preferences, Herrera-Viedma, Herrera, and Chiclana (2002) proposed a new consensus model based on consensus measures computed by com- parison between actor’s solutions and not between actors’ prefer- ences. A feedback mechanism based on rules was then proposed to make settings to actors’ preferences. This proposition can lead to an automated consensus process which does not require the intervention of an analyst who may be considered subjective. Another automatic control system to guide the consensus pro- cess was proposed by Herrera-Viedma, Martinez, Mata, and Chiclana (2005) if decision makers could use different linguistic domains to express their opinions. This consensus model uses the consensus degrees to decide when the consensus process should finish and the proximity measures to define a recommenda- tion system that recommends actors about the preferences that they should change in the next consensus rounds. In presence of incomplete information or missing values in GDM problems, Herrera-Viedma, Alonso, Chiclana, and Herrera (2007) propose other seminal automatic consensus model contri- bution based on three kind of measures: consensus measures, con- sistency measures and incompleteness measures. The automated consensus models act generally in similar way during all consensus stages although the conditions of GDM prob- lem change (Cabrerizo, Pérez, & Herrera-Viedma, 2010; Cabrerizo et al., 2010; Chiclana, Mata, Martinez, Herrera-Viedma, & Alonso, 2008), however adapted consensus model considering agreement degree may be interesting and more effective. In this context, an adaptive consensus model has been proposed by Mata et al. in (Chiclana et al., 2008) in which authors propose to adapt the number of changes required to the actors in each round of con- sensus with regard to the agreement degree. In the present paper, a consensus model is adaptive and instructions follow the agree- ment degrees in bipolar context. 3.1. Consensus approaches classification The soft consensus process intervening in recommendation phase, can be divided on the following mains steps: 1. Evaluation of consensus measures: from individual preferences obtained in the evaluation phase. 2. Consensus control: where consensus measures are compared to fixed threshold level of agreement. If the consensus level is achieved, the group moves onto the final recommendation and selects solution(s), otherwise, the preference setting is nec- essary and another iteration is to do. 3. Preference setting or consensus progress: when the threshold level of agreement is not respected by actors, oriented instruc- tions for preferences setting are proposed to divergent actors in order to obtain sufficient agreement level. This actions can be realized by moderator or automated feedback mechanism. 4. Final recommendation: once a sufficient level of agreement is reached, the selection of the final solution can be done. According to defined soft consensus processes, consensus approaches defined in the literature can be, generally, distin- guished according to the following criteria (Herrera-Viedma et al., 2014; Palomares, Martinez, & Herrera, 2014): - Type of preference structures: the individual results of each alter- native can be obtained using tools as preference relations, pref- erence orderings, utility vectors (Herrera-Viedma et al., 2002; Yu & Lai, 2011), etc., or, aggregation operator based on distances to the positive and negative ideal solutions by considering weight of attributes and/or decision-makers (Li, 2007; Park, Park, Kwun, & Tan, 2011; Yue, 2011b; Yue & Jia, 2012), or infor- mation domains (e.g. numerical or linguistic information (Bordogna, Fedrizzi, & Pasi, 1997; Parreiras, Ekel, & Bernardes, 2012)) used by actors to express their preferences over alterna- tives, amongst others. Additionally, some models are focused on multiple criteria GDM problems (MCGDM) (Fu & Yang, 2010; Parreiras et al., 2012), in which information fusion approaches are often uti- lized to combine preferences evaluated according to several cri- teria, whilst other models have been defined to deal with a particular type of real-life decision problems (Choudhury et al., 2006; Eklund et al., 2008). - Reference domain: considering reference domain, the proposed consensus approaches can be divided into two categories: - Approaches based on the actors set (Carlsson et al., 1992; Fedrizzi, Kacprzyk, & Zadro_zny, 1988; Kacprzyk & Fedrizzi, 1988): in this case, consensus degrees of decision actors are obtained from an aggregation of an agreement degree of pair of individuals as to their opinions between all the pairs of options. - Approaches based on the alternative set (Cabrerizo, alonso, & herrera-viedma, 2009; Herrera et al., 1996a; Herrera- Viedma et al., 2007): in this case, the proposed approach allows analyst to guide decision actors to change their pref- erences during the discussion process considering three lev- els of preference relations: level of preference (which indicates the consensus degree existing among all the prefer- ence values attributed by the actors to a specific preference), level of alternative (which allows to measure the consensus existing over all the alternative pairs where a given alterna- tive is present), level of preference relation (which evaluates the social consensus, that is, the current consensus existing among all the actors about all the preferences). - Concept of coincidence: which means observing the existing coincidence among actors’ opinions, is used to compute soft consensus measures that can be valued in [0, 1] (where 0 indi- cates no consensus and 1 the unanimity) (Bryson, 1996; Herrera-Viedma et al., 2007; Kacprzyk et al., 1992; Wu & Xu, 2012; Xu, Li, & Wang, 2013) or using linguistic labels (Herrera et al., 1997a; Herrera, Herrera-Viedma, & Verdegay, 1997b; Pérez, Cabrerizo, & Herrera-Viedma, 2011). Three different approaches can be distinguished: - Strict coincidence among preferences: The coincidence is obtained by computing distances between actors prefer- ences, accepting only the total coincidence or null coinci- dence cases (Herrera et al., 1997a; Kacprzyk, 1987). - Soft coincidence among preferences: very extended in GDM and specially applied in contexts of GDM under prefer- ence relations, this approach considers also different partial coincidence degrees to obtain consensus. (Bordogna et al., 1997; Fedrizzi et al., 1988; Kacprzyk & Fedrizzi, 1988). - Coincidence among solutions: In this case, the coincidence is also a gradual concept assessed in [0,1] (Ben-Arieh & Chen, 2006; Chiclana, Herrera-viedma, Herrera, & Chiclana, 2002). Coincidence approach provides here more realistic consen- sus measure among actors thanks to the consideration of the position of individual and the collective solutions. This approach is used in decision situations under different for- mats of preference representation. - Generation method of recommendation: in the recommendation phase, some instructions are given to actors in order to improve the consensus statue and lead group decision to common legit- imate solution. Two recommendation modes are possible. - based on analyst guidance (feedback mechanism): in these consensus approaches (Bordogna et al., 1997; Kacprzyk, 1987; Kacprzyk & Fedrizzi, 1988; Kacprzyk et al., 1992; Mata, Martinez, & Herrera-Viedma, 2009), the analyst’s goal in each stage is to address the consensus reaching process towards success by achieving the maximum possible agree- ment degree and reducing the number of actors outside of the consensus. - based on automatically process (Cabrerizo et al., 2010; Chiclana et al., 2002; Herrera-Viedma et al., 2007; Tapia GarcíA, Del Moral, MartíNez, & Herrera-Viedma, 2012): to avoid possible subjectivity of analyst, some authors propose to make the automated consensus process by introducing feedback mechanisms. In these approaches, feedback mech- anism is based on the evaluation of the distance between individual preference measures and the collective one using proximity measures. One actors with less contribution to reach high consensus identified, some instructions are trans- mitted for them to obtain a solution with better consensus degree. The automated feedback mechanisms know actually a growing success specially when consensus processes are developed in crowded social environments, such as Web 2.0 (Alonso, Herrera-Viedma, Chiclana, & Herrera, 2010; Alonso & Pérez, 2013). - Guiding measures: used to guide decision makers toward a common solution through the analyst and/or feedback mecha- nisms. It is represented by: - consensus measures: as mentioned bellow, in the classical consensus process the consensus measures are used to eval- uate the current consensus stage. - others kind of measures: avoids misleading solutions, in some consensus approaches, others measures are added to guide actors in the process, as a consistency measures (Herrera-Viedma et al., 2007; Cabrerizo et al., 2010), or index of comparability used to identify actors that have faced dif- ficulties in expressing their preferences (Parreiras et al., 2012), or compatibility measure between actors’ assessment (Fu & Yang, 2010; Jiang, Xu, & Yu, 2013), etc. Considering group decision making problem, this paper present a new adaptive bipolar soft consensus building process based on flexible feedback mechanism where instructions follow the agreement degrees. According to previous definitions, the pro- posed model can be considered as an approaches based on the alternative set where coincidence is given among solutions. 4. Bipolar group decision making framework modeling under interactions This section discusses the structuring and evaluation procedure based on bipolar analysis approach. Considering Multi-Attributes Multi-Objectives Group Decision Making problem (MAMOGDM probem), the general framework of the bipolar approach is pre- sented and the satisficing game theory used as aggregation tool in individual evaluations is briefly described. The objective of pro- posed model is to provide a realistic framework to achieve a con- sensus, taking into account the potential impact of (positive/ negative) influence among decision makers and their individual- ism degrees in different phases of problem resolution. 4.1. Proposed bipolar structuration of MAMOGDM problem Let us consider a MAMOGDM problems characterized by the set of decision makers D = {d1,d2, . . . .dm} where each actor evaluate individually a common, finite and static set of alternatives noted A = {a1,a2, . . . an} to achieve fixed objectives O = {o1, o2, . . . op}. The alternative evaluation is done using a set of attributes (or criteria) noted Col ¼ fc1; c2; . . . ; cqg for each objective ol. In the bipolar framework, alternatives are evaluated through bipolar ‘a priori’ measures. These measures noted ljs0 ðaiÞ=l j r0 ðaiÞ, represent respec- tively the ‘supportability’ and the ‘rejectability’ degree accorded by the decision maker dj for an alternative ai without considering potential influence of other decision makers. These bipolar measures can be obtained in structured frame- work using BOCR analysis (Saaty, 2001; Saaty & Özdemir, 2005) that brings structured evaluation considering in comprehensive way benefit, opportunity, cost and risk aspects. The opportunities in the BOCR analysis include usually positive expectations, future profits and income of the positive developments, while benefits are current revenues or profits of the positive developments rela- tively certain. Similarly, risks allow identifying the expected conse- quences of future negative developments, while costs are losses (in progress) where the consequences of the downturn are relatively certain. Each decision maker can evaluate alternatives through b, o, c, r factors. The BOCR analysis can be associated with analytic hierar- chy process to obtain individual preferences. In this case, criteria can be hierarchically distributed in clusters going from general to operational levels and to evaluate alternatives by pairwise compar- isons (Bouzarour-Amokrane, Tchangani, & Pérès, 2012; Tchangani, Bouzarour-Amokrane, & Pérés, 2012). We denoted by Bjol ðaiÞ; O j ol ðaiÞ; C j ol ðaiÞ; R j ol ðaiÞ respectively the evaluations of (b), (o), (c), (r) factors given by the decision maker dj for the alternative ai with regard to achievement of the objective ol. These evaluations are aggregated for each decision maker over all objectives to represent each alternative ai with (b), (o), (c), (r) values noted BjðaiÞ; O jðaiÞ; C jðaiÞ; R jðaiÞ. Synthesis methods of BOCR factors are usually based on multi- plicative or additive expressions (Saaty, 2001; Saaty & Özdemir, 2005). However, due to the incommensurability of priority synthe- sis of the four factors, except in special cases, one cannot be certain that multiplicative or additive synthesis expression produces a cor- rect order of alternatives (Wijnmalen, 2007). Considering bipolar context, we propose to aggregate the bene- fit (b) and opportunity (o) factors in the ‘selectability’ measure and cost (c) and risk (r) factors in the ‘rejectability’ measure using the satisficing game theory introduced below. 4.2. Satisficing game theory The satisficing game theory is based on the fact that decision makers in solving real problems do not necessarily seek the opti- mum solution –often costly in terms of time and money– but a sat- isfactory solution whose capabilities are estimated fairly good regarding to objectives achievement (Tchangani, 2009). The satis- ficing game theory is based on this observation and provides ade- quate tools for the selection of satisficing alternatives and reaching consensus. The concept of being good enough is suitable for our approach where an alternative can be considered good enough when its supporting contribution exceeds the rejecting one. The satisficing game theory is characterized by the triplet hA, ls, lri where: A: is a set of alternatives (discrete set). ls(ai)/lr(ai): are mass functions defined in A on the interval where ls(ai),lr(ai) measure the degree to which ai works towards success in achieving the decision maker’s goal and costs associated with this alternative respectively (Tchangani, Bouzarour- Amokrane, & Pérés, 2011). Thus, in the considered BOCR analysis framework, the bipolar measures using satisficing game theory are obtained with follow- ing equation. ljsðaiÞ ¼ rjBjðaiÞ þ ð1 ÿ r jÞOjðaiÞ P ir jBjðaiÞ þ ð1 ÿ rjÞO jðaiÞ ð1Þ ljrðaiÞ ¼ ð1 ÿ rjÞCjðaiÞ þ r jRjðaiÞ P ið1 ÿ r jÞCjðaiÞ þ rjR jðaiÞ ð2Þ where 0 6 rj 6 1: is risk aversion index of decision maker dj. More rj is close to 1, greater is risk aversion of decision maker who, pes- simistic, will tend to give more importance to risk than cost in rejec- tability measure (Eq. (2)) and penalize opportunity in favor of benefit in selectability measure (Eq. (1)). Inversely, when the risk aversion index tends to 0, the decision maker is optimist. He will focus on opportunity to benefit in the selectability measure, and will overlook risk against cost in the rejectability measure. The satisficing game theory is characterized by the several sets that can be used on the recommendation phase. The satisficing set Sqj # A (at a boldness or caution index qj of decision maker dj) is the set of alternatives defined as following. Sqj ¼ i 2 A : l j sðaiÞ P qj l j rðaiÞ � ð3Þ where qj is the caution index of decision maker dj that can be used to adjust the aspiration level: increase qj allows reducing a satisfic- ing set Sqj (if too many alternatives are declared satisficing), on the contrary, decrease qj allows increasing satisficing set (if Sqj is empty for instance), see Fig. 1. To identify non-dominated alternatives which presenting a higher selectability measure and a lower rejec- tability measure an equilibrium set ej is defined by decision maker dj as follow. ej ¼ fai 2 A : DjðaiÞ ¼ ug ð4Þ where DðaiÞ is the set of alternatives that are strictly better than ai (see Fig. 2). The set DjðaiÞ is defined with Eq. (5) DjðaiÞ ¼ DsðaiÞ [ DrðaiÞ ð5Þ where DS(ai) and DR(ai) are defined by Eqs. (6) and (7) below DsðaiÞ ¼ ai0 2 A : lr ai0ð Þ < lrðaiÞ and lsðai0 Þ P lsðaiÞ � ð6Þ DrðaiÞ ¼ ai0 2 A : lr ai0ð Þ 6 lrðaiÞ and lsðai0 Þ > lsðaiÞ � ð7Þ Thus, the satisficing equilibrium set eS;jq is given as follow eS;jq ¼ ej \ Sqj ð8Þ Notice that the set eS;jq constitutes a Pareto-equilibria set with an incomparability between a pair of alternatives and a trade-off pro- cess can be necessary for final choice in case of indecision. These alternatives are those laying on the portion of broken line green curve (above which there is no alternatives meaning the Pareto equilibrium) that is above the straight red line (separating satisfic- ing and non-satisficing alternatives) of the following Fig. 3. To provide a framework for integrating human factors evalua- tion, the following section proposes to model the influence between decision makers through concordance and discordance Fig. 1. Graphical representation of satisficing set. Fig. 2. Graphic representation of non-dominated alternatives. measures. The degree of individualism and the weight of decision makers will also be defined and integrated into the resolution model. 4.3. Influence modeling procedure A group decision is usually composed of persons whose percep- tions, attitudes, motivations and personality are different. In the GDM problems with an important number of jugged suitable and unsuitable actors to solve decision process, it is difficult to achieve acceptable consensus. To consider the importance of each actor in the decisional pro- cess, some researches introduce the decision maker importance degrees in the aggregation phase of actors’ opinions (Herrera et al., 1996a; Herrera et al., 1997a; Lee, 2002), but without consider them in the recommendation phase. In its recent research Alonso and Pérez (2013) propose to integrate the actors importance in consensus process but considering only subgroup of actors consid- ered more important and final opinions given by this subgroup is used to obtain the final solution. In present contribution, a confidence degree is introduced in the context of collaborative group decision where all decision makers are brought to cooperate in the evaluation phase. Each actor has a confidence degree obtained considering the positive or negative opinions that other members on the group have of him. Depending on the personality of the actors, positive and/or neg- ative influences can be exercised; some persons in the decision group are exclusively individualistic and take into account only their point of view, while other persons say ‘holistic’ prefer to take into account the opinion of other actors according to the impor- tance degree assigned to them. To model the influence related to each decision maker’s opin- ion, this part proposes to define concordance and discordance degrees awarded by each decision maker to actors he considers influential. These measures allow decision makers to express their level of agreement or disagreement overlooked these actors. Indi- ces of concordance and discordance are defined following (Bouzarour-Amokrane, Tchangani, & Pérès, 2013; Tchangani, 2013). Definition 1. let V(j) be the vicinity of decision maker dj, which represents a set of mj decision makers whose opinion matters (influence) the decision maker dj in a positive or negative way. For each decision maker dj0 belonging to a vicinity of de dj, we define respectively xc jj 0 ; xd jj 0 as relative concordance and discordance degrees accorded by decision maker dj to opinion of decision maker dj0 compared to other members of the vicinity, where P d 0 j2VðjÞ xc jj 0 ¼ 1, P d 0 j2VðjÞ xd jj 0 ¼ 1. The selection of vicinity V(j) can be obtained for each decision maker dj in an instinctive way in the case of a small group of decision where a pairwise comparison can be done for example to select a favorite actors considering a weight threshold. Considering a large decision group, one can adapt recent models proposed in the literature to identify the trust network and deduce the subgroup of decision makers (Alonso & Pérez, 2013), or, detect non cooperative behaviors and dealing with them used fuzzy clustering as proposed in Palomares et al. (2014). Once vicinity sets are defined, several methods can be used to obtain concordance and discordance degrees, the analytic hierar- chy process (AHP) can be proposed for a relatively quick estimate. The AHP method is based on pairwise comparison by answering question of the form ‘‘what is the concordance (resp. discordance) degree accorded by the decision maker dj to the opinion of dj0 compared to the opinion of dj00 ’’ (where dj0 ; dj00 2 VðjÞ). Based on this definition, we can consider that a decision maker is even more important in the community if other decision makers give him a good confidence. Thus, the degree of importance of deci- sion maker dj can be defined as follows. Definition 2. let xc j 0 j ; xd j 0 j be respectively the relative concordance and discordance degrees accorded by the decision maker dj0 to an opinion of dj. The importance degree of dj can be defined by Eq. (9) Hj ¼ P j 0 max 0; xc j0j ÿ xd j0j � � Pm p¼1 P l maxð0; x c lp ÿ xd lp Þ n o ð9Þ Or, Eq. (10) which allows decision makers to have non-zero degrees of importance, unlike formulation given by Eq. (9) in which zero importance is possible. Expression (10) may be useful in the case of a holistic and collaborative group consists of decision makers sensitive to opinion of their neighborhood. Hj ¼ P j0 1 1þ exp ÿ/ xc j0j ÿxd j0j � �� �� � 0 @ 1 A Pm p¼1 P k 1 1þ exp ÿ/ xc lp ÿxd lp � �� �� � 0 @ 1 A 0 @ 1 A ð10Þ where a is a parameter setting. Considering the relative importance degree, the relative bipolar measures are then defined to include neighborhood effects in the final bipolar measures. In the BOCR analysis where decision makers evaluations are expressed by BjðaiÞ; O jðaiÞ; C jðaiÞ; R jðaiÞ. The relative bipolar evalu- ations in this case can be expressed using Eq. (11) and (12). lj=VðjÞs ðaiÞ ¼ P j 0 2VðjÞHj 0 ðrj 0 Bj 0 ðaiÞ þ ð1 ÿ r j 0 ÞOj 0 ðaiÞÞ P i P j 00 P j 00 2VðjÞHj 00 ðrj 00 Bj 00 ðaiÞ þ ð1 ÿ rj 00 ÞOj 00 ðaiÞÞ � � ð11Þ lj=VðjÞr ðaiÞ ¼ P j 0 2VðjÞHj 0 ðð1 ÿ rj 0 ÞCj 0 ðaiÞ þ r j 0 Rj 0 ðaiÞÞ P i P j00 ð P j002VðjÞHj 00 ðð1 ÿ rj 00 ÞCj 00 ðaiÞ þ rj 00 Rj 00 ðaiÞÞÞ ð12Þ where rj 0 : risk aversion degree of decision maker dk0 2 VðdkÞ, Hj 0 : relative importance degree of decision maker dk0 2 VðdkÞ. These equations mean that relative bipolar measures of deci- sion maker dj considers only the opinion or evaluation of his vicinity according to the importance of each one. The final bipolar measures (selectability and rejectability measures) are then obtained by Eqs. (13) and (14). ljsðaiÞ ¼ d jljs0 ðaiÞ þ ð1 ÿ d jÞlj=VðjÞs ðaiÞ ð13Þ ljrðaiÞ ¼ d jljr0 ðaiÞ þ ð1 ÿ d jÞlj=VðjÞr ðaiÞ ð14Þ where ljs0 ðaiÞ=l j r0 ðaiÞ represent a priori measures of alternative ai obtained by aggregating the corresponding factors. Fig. 3. Positions of alternatives in the plan ðlR; lSÞ. 0 6 dj 6 1, is the individualism degree of decision maker dj. While dj tends to 0, the decision maker is considered as ‘holistic’ (altruist) and gives a more importance to the global opinion represented by his vicinity. Inversely, if dj tends to 1, the decision maker is ‘individualist’ and considers his opinion above his vicinity. When decision makers present their evaluations directly with ‘a priori’ bipolar measures ljs0 ðaiÞ=l j r0 ðaiÞ, the relative bipolar mea- sures can be obtained using the following equations. l j=VðdjÞ s ðaiÞ ¼ P j 0 2VðdjÞ xc jj 0l j 0 s0 ðaiÞ þ x d jj 0l j 0 r0 ðaiÞ � � P i P j 0 2VðdjÞ xc jj0 lj 0 s0 ðaiÞ þ x d jj0 lk 0 r0 ðaiÞ � �� � ð15Þ l j=VðdjÞ r ðaiÞ ¼ P j 0 2VðjÞ x c jj 0l j0 r0 ðaiÞ þ x d jj 0l j0 s0 ðaiÞ � � P i P j02VðdjÞ xc jj 0l j0 r0 ðaiÞ þ x d jj 0l j0 s0 ðaiÞ � �� � ð16Þ For decision maker dj, these relative bipolar measures take into account the ‘a priori’ bipolar measures of his vicinity. In the select- ability measure, the decision maker considers also the ‘a priori’ rejectability measure of his vicinity V(j) according to their discor- dance degree. Inversely in relative rejectability measure, decision maker integers ‘a priori’ selectability measures of his vicinity V(j). The finale bipolar measures can then obtained by the Eqs. (13) and (14). From the individual evaluation results, the selection of a solution or set of solutions approved by the group decision generally requires a consensus building process. The following section provides a process achieved consensus from local prefer- ences represented by different sets obtained from the satisficing game theory. The scheme of proposed consensus model can be given as follow, Fig. 4. 5. Consensus and selection processes The individual bipolar measures are used by decision makers to select the best alternative or the set of the best alternatives for each one. According to the knowledge and perceptions of actors, the decision-makers choices can be contradictory. To find a com- mon satisficing alternative(s), a consensus processes is necessary. The reach consensus can be done in two ways: by aggregating final bipolar measures of decision makers into collective bipolar measures using Eqs. (17) and (18). Then selected alternative(s) can be obtained from satisficing equilibrium set according to pro- posed selection criteria. The second way to reach a consensus can be done using local preferences. The local preferences are defined as the selected alternatives by each decision maker according to his bipolar measures. The local preferences can be expressed by selection of unique alternative considered as ‘the best’ one, or by the equilibrium satisficing set defined in the satisficing game the- ory (Section 4.2). 4If selected alternative(s) is(are) the same for all decision makers, the common alternative can be chose easily, however, if selected alternatives are different a consensus process is necessary. In the following, we propose a consensus model for each case of unique selected alternative or selected alternatives set. lcsðaiÞ ¼ P jHjl j sðaiÞ P jHj ð17Þ lcrðaiÞ ¼ P jHjl j rðaiÞ P jHj ð18Þ where lcsðaiÞ and l c rðaiÞ represent respectively, the collective select- ability and rejectability measures. 5.1. Case of unique local preference (unique selected alternative) In this case, each decision maker dj selects his ‘optimal’ or ‘more satisficing’ alternative noted a�j (among a set of equilibrium satis- ficing for example). The selected alternative can be obtained by the following. a�j ¼ arg maxa2eS;jq ðpðaiÞÞ ð19Þ where pðaiÞ ¼ f l j sðaiÞ;�ı j rðaiÞ � � is value function that can be take par- ticular form depending on the decision goal of decision maker dj, for example : pðaiÞ ¼ l j sðaiÞ ÿ ql j rðaiÞ ð20Þ that gives the priority to alternatives with large difference between the selectability measure and the rejectability measure given the index of caution, or pðaiÞ ¼ l j sðaiÞ=l j rðaiÞ ð21Þ that considers alternatives with the largest index of caution, or pðaiÞ ¼ l j sðaiÞ respct: pðaiÞ ¼ 1 ljrðaiÞ ! ð22Þ that gives priority to alternatives with the largest selectability (respect. lowest rejectability); this later case is suitable when one of the measure is uniformly distributed over alternatives. If the selected alternative a�j is the same for all decision makers, the final solution can be deduced easily. Conversely, if selected alternative a�j varies from one person to other a reach consensus processes based on selection criteria is proposed to find a common acceptable alternative. According to the group nature, context and society where the decision problem is considered, the decisions rules can be changed. In his paper, Urfalino (2007) addresses a question of decision rules choice according to periods and societies. These parameters will favor unanimity, consensus or majority voting in order to achieve legitimate collective decision. The group nature and social ties are also determinant parameters in the decision rules choice. These Fig. 4. General scheme of proposed consensus models. parameters can be introduced with the ‘individualism’ and ‘holism’ notions that play important role in the rules selection. Considering these notions, we propose some possible selection criteria in the following. 1. Select alternative chosen by decision maker who presents the more important importance degree according by the group. 2. Select alternative that presents the more important selectability a� ¼ arg max ai2\e S;j q P jHjl j sðaiÞ � �� � . 3. Select alternative that presents the less important rejectability a� ¼ arg min ai2\e S;j q P jHjl j rðaiÞ � �� � . 4. Select the alternative that has the maximum benefit for all deci- sion makers max ai2\e S;j q P jHj l j sðaiÞ ÿ l j sðaiÞ � �� � . 5. Selecting a qualified majority: in some situations, the use of qualified majority decision is recommended in decision making. In this case, only the responses from a subset of makers are con- sidered. We define ‘Qualified Majority’ as follows. Definition 3. a ‘Qualified Majority’ (QM) is a set of decision makers who present a% of the group decision members and b% of total importance. QM ¼ fd1; d2; . . . dtg where t ¼ am and X t j¼1 Hj P b; X P j¼1 Hj ¼ 1 ð23Þ where a,b, are respectively possible rates that define ‘majority members’ and ‘global importance of the majority’, these rate can be defined by group decision or analyst. Consider the ‘qualified majority’ (QM), the selection of the final decision can be done using the following steps. Step1. If decision makers have a common alternative (obtained from the previous selection criteria) then this alternative is selected. Step2. Otherwise, make adjustments to the caution index qj of decision makers who show great caution successively in descending order until a common alternative. Note that the latter definition ensures to reach a final solution on a partial consensus scheme; the notion of consensus in this case should be considered with regards to the agreement about the parameters a and b by the members of the group. 5.2. Case of the set of local preferences (expressed by satisficing equilibrium set) In this part, local preferences are expressed according to the equilibrium satisficing set defined on satisficing game theory (see Section 4.2). Considering that each decision maker dj identified his equilibrium satisficing set eSqj , if Tm j¼1e S qj – ;, then the selection of common alternative can be done for example with the following. a�j ¼ arg maxai2 Tm j¼1 e S;j q ðpðaiÞÞ ð24Þ where pðaiÞ ¼ Qm j¼1 ljsðaiÞ ljrðaiÞ � � for example. Conversely, if Tm j¼1e S;j q ¼ ;, a reached consensus process is neces- sary. To deal with this, we propose two possible reaching consen- sus process; fist process is based on varying caution index of decision makers when second proposed process is based on the distance measures between individual preferences. These mea- sures are commonly called ‘proximity measures’ when it comes to comparing individual decision maker opinions to global or col- lective opinion on the ‘soft’ consensus process. 5.2.1. Consensus model based on caution index One possible way to reach a consensus is to vary caution index of some decision makers to enlarge their satisficing equilibrium set and tend towards a non-empty intersection of these sets. To select decision makers who must modify their caution index, some selec- tion criteria can be proposed as, importance of decision maker, sat- isficing equilibrium set dimensions ljsðaiÞ=l j rðaiÞ rapport, etc. We propose to decrease caution index of decision makers according to - The importance degree: decrease caution index of decision makers who presents the less important degree one by one until reach consensus Tm j¼1e S;j q – ; � � : - The qualified majority (QM): if the qualified majority of deci- sion maker present one or more common alternatives, accord- ing to the group nature relations and decisional context (unanimity or not, confidence or not, complementarity or not), select a more satisficing alternative among QM common alternatives or ask to the rest of group to enlarge their equilib- rium satisficing set to achieve a consensus using importance degree as regulation factor for example. - Satisficing equilibrium set dimension: decrease caution index of decision makers who present the smallest set of satisficing equilibrium, one by one until reach a consensus Tm j¼1e S;j q – ; � � . - Proximity limits: decision makers who present a large number of alternatives with ljsðaiÞ=l j rðaiÞ tend to q max j ¼ maxai2A ljsðaiÞ ljrðaiÞ � � are brought to reduce their index caution. In other words, deci- sion makers consider alternative solutions with selectability measures much higher rejectability measures are brought to expand their range of solutions by reducing their caution indi- ces one by one until reaching a consensus. This allows the deci- sion makers to include in their satisficing equilibrium sets alternatives with better spreads. 5.2.2. Consensus model based on proximity measures Supporting that a final agreement between decision makers is not necessary to resolve group decision problems, this section pro- poses to reach a common solution through a soft consensus pro- cess (Herrera et al., 1996a; Herrera-Viedma et al., 2002) based on proximity and bipolar consensus measures defined from final bipo- lar measures. The proximity measures are used to assess the gap between decision makers’ evaluation on each alternative while bipolar con- sensus measures allow evaluating the gap between decision maker dj and the rest of the group considering bipolar measures of alter- natives. These measures are part of a feedback mechanism pro- posed to guide decision makers in adjusting their evaluations to tend towards a single solution and reach a consensus. Unlike soft consensus process proposed in the literature (Herrera-Viedma et al., 2007; Khorshid, 2010), the model pre- sented here does not aim to converge all decision makers assess- ments but focuses on alternatives presenting initially the same evaluation trend (convergent). The targeted recommendations are given to decision-makers to correct their differences and to reach the common solution from selected alternatives (Bouzarour-Amokrane et al., 2013). In other hands, as mentioned bellow, some literature proposes to take into account the importance of decision makers when aggregating the actor’s opinions (Herrera et al., 1996a; Herrera et al., 1997a; Lee, 2002) but not when advising to the actors how to change their preferences to increase the consensus level. Pro- posed feedback mechanism consider this point and propose to use weighted consensus measures to integrate decision maker’s importance in the recommendation phase. Considering that the intersection of the equilibrium satisficing sets of decision makers is empty Tm j¼1e S;j q ¼ ;, reaching a consensus amounts to establishing an iterative process allowing to have sev- eral phases of consultation until an agreement. The objective is to identify alternatives on which decision-makers have convergent opinions and support them to adjust their evaluation in order to reach a final common solution based on initial convergent opinion. In each iteration, considering importance degree of decision mak- ers, the proximity measures are calculated to determine conver- gent alternatives (that present the smallest distance compared to others), when bipolar consensus measures are measured to iden- tify decision makers who present a large gap of supportability and/or rejectability evaluations for the preselected alternatives compared to the rest of the group. The main problem in this situation is the difficulty to determin- ing the way to convergences individual preferences (Tapia GarcíA et al., 2012). To achieve this convergence, boundary conditions (threshold) are fixed for each level (for proximity and bipolar con- sensus measures). This process is carried out in a discussion ses- sion led by a feedback mechanism that can guide decision makers in changing their opinions. The alternatives with confused evaluations will be readjusted by the decision makers who present a strong divergence. A discussion session is led by a feedback mechanism to lead decision makers towards a common legitimate solution. The parameters of the proposed reached consensus pro- cess are developed below. 5.2.2.1. Proximity and bipolar consensus measures. The proximity and bipolar consensus measures are defined as follow (Bouzarour-Amokrane et al., 2013). Definition 4. proximity measure di is used to calculate the average distance between decision maker evaluations considering the alternative ai. It is obtained as follows, Eq. (25). di ¼ P j P j 0 ;j 0 –j d jj 0 si � �2 þ d jj 0 ri � �2 � �1 2 m 2 � � ð25Þ where d jj 0 si ¼ Hjl j sðaiÞ ÿ Hj0l j 0 s ðaiÞ, distance between the supportability measure of decision maker dj and dj0 for the alternative ai, with regard to the importance of each decision maker. d jj 0 ri ¼ Hjl j rðaiÞ ÿ Hj0l j 0 r ðaiÞ, distance between the rejectability measure of decision maker dj and dj0 for the alternative ai, with regard to the importance of each decision maker. m 2 � � ¼ m!2!ðmÿ2Þ!, binomial coefficient takes into account combi- nations with avoiding redundancy distances (for example: d 12 si ¼ d 21 si ). Definition 5. bipolar consensus measures are used to calculate the distance between a decision maker dj and the rest of the group con- sidering supportability and rejectability evaluation for the alterna- tive ai. The bipolar consensus measures are given by Eq. (26) and (27) respectively. d j si ¼ P j0;j0–j d jj 0 si � � m ÿ 1 ð26Þ d j ri ¼ P j 0 ;j 0 –j d jj0 ri � � m ÿ 1 ð27Þ where d j si ; d j ri represent respectively selectability and rejectability consensus measures. 5.2.2.2. feedback mechanism. The feedback mechanism is able to assist the moderator in the process of reaching consensus or even replace it allowing actors to change their preferences in order to achieve a tolerated proximity degree. The main issues of this mechanism are: how to ensure that individual positions converge? And how to support a decision maker in obtaining and accepting a given solution? (Tapia GarcíA et al., 2012; Chiclana et al., 2002). Literature generally represents the feedback process by identifi- cation and a recommendation phases. According to this structure, the following section presents a feedback process proposed in con- nection with the bipolar approach. Identification phase. The identification phase evaluates the closeness of decision maker assessments for each alternative and compares them to tolerances threshold fixed by the decision group or by the moderator. The alternatives which represent a conver- gent opinion among group decision members (manifested by a smallest proximity measure) considered more interesting, are selected. Then, bipolar consensus measures are used to identify decision makers who present a significant deviation compared the rest of the group according selected alternatives. From these measurements, evaluations of alternatives can be modified using the following steps. (a) Identification of alternatives whose proximity measure di respects the condition (1) di 6 x, where x is the tolerance threshold representing the permissible difference between alternatives (average distances on the set of alternatives can be considered as tolerance, for example). This allows excluding alternatives that may create conflicts and focus on alternatives already having a certain convergence. (b) For each alternative ai that respect condition (1), identifica- tion of decision makers that don’t respect one or both following conditions: ð2Þd k si 6 xs; ð3Þd k ri 6 xr, where xs/xr are a tolerance thresholds of selectability/rejectability measure respectively. These conditions allow identifying decision makers who pres- ent a divergent opinion regarding a rest of group. The average of all alternatives can be proposed as threshold. Recommendation phase. In this phase, recommendations are given to decision makers do not fulfill the conditions (2) and (3) in order to change their assess- ments concerning alternatives that meet the condition (1). The rec- ommendations are based on the following rules: - for d j si > xs, the decision maker dj has a large gap related to its selectability measure which means that the selectability mea- sure given by dj concerning alternative ai, moves away from the evaluations given by other actors. To know the divergence direction and thus know if the considered alternative ai, has more important selectability measure (positive divergence) or a low measure (negative divergence) compared to other actors, the following equation is proposed. div j si ¼ P j 0 ;j 0 –j d jj0 si � � m ÿ 1 ð28Þ If div j si > 0, the alternative ai has a good selectability measure, no change is required. Otherwise div j si < 0 � � , the selectability measure is smaller than the measures average, an increase selectability mea- sure is recommended for considered alternative ai. The same recommendations are implemented for rejectability measure. - for d j ri P xr, the decision maker dj has a strong divergence from the rest of the group. The divergence direction indicates whether the alternative has a low rejectability (positive diver- gence) or a significant rejectability (negative divergence). The divergence direction of rejectability is given by Eq. (29). div j ri ¼ P j0;j0–j d jj 0 ri � � m ÿ 1 ð29Þ If div j ri > 0, the alternative ai has a low rejectability measure, no change is required. Otherwise div j ri < 0 � � , the rejectability mea- sure is greater than the measures average, reduction of rejectabil- ity measure is recommended for considered alternative ai. Once modifications realized, the equilibrium satisficing sets for each decision maker are reconstructed. The iterative process is stopped when Tm j¼1e S;j q – ;. If obtained solution satisfies group deci- sion maker, the process can be stopped. Else, other iteration can be proposed. The next section presents an application example to illustrate proposed approach. 6. Application example In this section, we use proposed group decision making bipolar analysis approach to resolve real size wind farm implantation problem adapted from Lee, Chen, and Kang (2009). Install a wind farm involves various actors in society such as; wind specialist, local administration and public authority. In France for example, the selection of potential implantation sites (selection of alternatives) returns generally to the local administration, which taking into account previous studies, will retain potential installa- tion areas approved by the prefect. A more sophisticated study is then performed to select the site chosen for the implementation, bipolar collective approach presented in this chapter can be used for analyze study. To illustrate the developed bipolar model, this section adapts original problem data of the of installing a wind farm (Lee et al., 2009), considering a decision committee composed of three entities: wind specialist, local administration and public authority denoted respectively d1,d2,d3. The goal is to select a loca- tion for a wind farm installation to achieve a set of objectives related to socio-economic, performance and operability notions. Five potential sites are under consideration. The importance degrees of socio-economic, performance and operability objectives are given by decision makers d1,d2,d3 respectively as follows: x1,O = [0.1 0.8 0.1], x2,O = [0.8 0.1 0.1], x3,O = [0.1 0.1 0.8]. It is assumed that specialists wind (entity d1) give more importance to the performance objective, their first goal is to establish a production site with a good yield. To manage the budget and preserve the municipal property, local administration (d2) promotes socio-economics aspects while public authority (d3) are supposed to pay more attention to the operability of the future site. The bipolar evaluation of alternatives is carried out through BOCR analysis that consists in evaluating alternatives over benefit, opportunity, cost and risk factors through a distribution of attri- butes on these factors (Bouzarour-Amokrane et al., 2012). The results are summarized in Tables 1 and 2, representing respectively the evaluation results of the BOCR analysis and ’a priori’ bipolar measures obtained for a risk aversion given by the vector rj = [0.3 0.5 0.7]. It is assumed in this case that the wind specialists (d1) have a low risk aversion against a medium risk aversion for local administration (d2) and an important caution about the pub- lic authority (d3). To represent the social link and the potential influences between decision makers, the proposed approach suggests model- ing the interactions between decision makers through concordance and discordance measures notes respectively xc jj0 =xd jj0 and repre- sented by the following matrices. xc jj 0 ¼ ÿ 0:1 0:9 0:7 ÿ 0:3 0:2 0:8 ÿ 2 6 4 3 7 5 xd jj 0 ¼ ÿ 0:7 0:3 0:2 ÿ 0:8 0:6 0:4 ÿ 2 6 4 3 7 5 These matrices show for example that decision maker d2 gives good confidence in decision maker d1 through a good degree of concor- dance and a low degree of discordance. Considering the concor- dance and discordance given above, the importance degree Hj of each decision maker dj is deduced from Eq. (10) is given by the fol- lowing vector. Table 1 Evaluation results of BOCR analysis. Decision makers Factors a1 a2 a3 a4 a5 d1 B 0.1892 0.2284 0.1498 0.2250 0.2076 O 0.1965 0.1990 0.1980 0.2020 0.2045 C 0.2072 0.1814 0.2132 0.2298 0.1684 R 0.1976 0.2064 0.1986 0.2058 0.1915 d2 B 0.1924 0.1900 0.2347 0.1581 0.2249 O 0.1964 0.1980 0.1898 0.2103 0.2055 C 0.1618 0.2172 0.1507 0.2216 0.2487 R 0.2000 0.2043 0.2045 0.1894 0.2018 d3 B 0.2137 0.1519 0.1957 0.2313 0.2073 O 0.1981 0.1945 0.2026 0.2005 0.2043 C 0.1902 0.1829 0.1891 0.1792 0.2587 R 0.2078 0.1955 0.2115 0.1994 0.1857 Table 2 ‘A priori’ bipolar measures. Decision makers Bipolar measures a1 a2 a3 a4 a5 d1 Selectability measure l1s0 � � 0.1943 0.2078 0.1835 0.2089 0.2055 Rejectability measure l1r0 � � 0.2043 0.1889 0.2088 0.2226 0.1753 d2 Selectability measure l2s0 � � 0.1944 0.1940 0.2122 0.1842 0.2152 Rejectability measure l2r0 � � 0.1809 0.2108 0.1776 0.2055 0.2252 d3 Selectability measure l3s0 � � 0.2090 0.1647 0.1978 0.2221 0.2064 Rejectability measure l3r0 � � 0.2025 0.1917 0.2048 0.1933 0.2076 Hj ¼ ½0:3334 0:3285 0:3381� ð30Þ The relatives measures taking into account the vicinity can be obtained using Eqs. (11) and (12) and bipolar final measures from Eqs. (13) and (14). Considering that the individualism degree is medium (dk = [0.5 0.5 0.5]), the results are respectively presented in Tables 3 and 4. The graphical representation of the evaluation results (Table 4) in the plane (lr,ls) is given by the following Fig. 5. The index of caution qj is assumed to be 1 for each decision maker dj. The satis- ficing equilibrium sets eS;jq � � of actors are as follows: eS;11 ¼ fa5; a2; a4g, e S;2 1 ¼ fa3g and e S;3 1 ¼ fa1; a4g. In this case there is no common solution ð T3 j¼1e S;j 1 ¼ ;Þ, we propose then to seek a consensus using the proposed models. By varying the caution index qj first, satisficing equilibrium sets of decision makers will be resized to make their non-empty intersection T3 j¼1e S;j q ¼ ; � � . The process reached consensus based on proximity and consensus measures, is proposed in a second time. 7. Consensus building process based on caution index variation The proposed model achieved with this consensus approach allows decision makers to avoid changing their evaluations. Only a concession on the caution index is required based on a selection criteria proposed in section 3.2.1 In this example, we note that the equilibrium satisficing sets of qualified majority of 67% composed of decision makers d1, d3 which consider the alternative a4 as a final solution ðeS;11 \ e S;3 1 ¼ fa4gÞ. Depending on the relationships nature and group decision-making context, the decision maker d2 can accept selecting the alternative a4 unsatisfactory for him with- out requiring modification. Otherwise, decision maker d2 will have to change its caution index to expand its satisficing equilibrium set until include alternative a4. The caution index variation may also be according to the importance degree of decision makers. In this case, the decision makers classification deduced from Eq. (26), (d3d1d2) indicates that the decision maker d2 must reduce its cau- tion index first. If the intersection is still empty after changing, the decision maker d1 changes its index of caution, and so on until a common solution is obtained. Considering that the caution index of decision maker d2 is given by q2 = 0.95 this modification allows to enlarge its satisficing equi- librium set to eS;20:95 ¼ fa3; a5g, however the intersection T3 j¼1e S;j 1 remains empty. The decision maker d1 must change its caution index, considering that caution index becomes q1 = 0.95, the satis- ficing equilibrium set of d1 remains identical e S;1 0:95 ¼ fa5; a2; a4g, the alternative a1 is satisficing but dominated and then excluded from satisficing equilibrium set eS;10:95. The intersection at this level is still empty. It then asks the decision maker to modify its caution index. Considering that q3 = 0.95, the satisficing equilibrium set of d3 is given by eS;30:95 ¼ fa1; a4g, the alternative a5 is satisficing but domi- nated by a4. In this case, a new iteration is realized, decision maker d2 must reduce its caution index more. Considering that q 2 = 0,91, the satisficing equilibrium set is equal to eS;20:91 ¼ fa5; a3g, where the alternative a4 is satisficing but dominated. It is noted that all the remaining alternatives can be satisfactory, but are still dominated. In this case, since the adjustments of assessments are not accepted, a possible solution amounts to consider the intersection of the sat- isficing equilibrium set of the most important decision maker with the satisficing sets of the rest of the group regardless of dominance for the rest of the group. This involves the selection of a qualified ‘more satisfactory’ solution by the largest maker. The solution in this case is the result of the following intersection eS;31 \ S S;1 0:95 \ S S;2 0:91 ¼ fa4g. Table 3 Relative bipolar measures. Decision makers Relative bipolar measures a1 a2 a3 a4 a5 d1 l 1=Vðd1Þ s 0.1976 0.1861 0.1926 0.2102 0.2135 l 1=Vðd1Þ r 0.1996 0.1893 0.2050 0.1951 0.2109 d2 l 2=Vðd2Þ s 0.2008 0.1929 0.1968 0.2060 0.2035 l 2=Vðd2Þ r 0.2049 0.1815 0.2013 0.2166 0.1957 d3 l 3=Vðd3Þ s 0.1947 0.1972 0.2014 0.2026 0.2041 l 3=Vðd3Þ r 0.1900 0.2043 0.1894 0.2040 0.2122 Table 4 Final bipolar measures. Decision makers Final bipolar measures a1 a2 a3 a4 a5 d1 Selectability measure l1s ÿ � 0.1960 0.1969 0.1881 0.2095 0.2095 Rejectability measure l1r ÿ � 0.2020 0.1891 0.2069 0.2089 0.1931 d2 Selectability measure l2s ÿ � 0.1976 0.1935 0.2045 0.1951 0.2093 Rejectability measure l2r ÿ � 0.1929 0.1961 0.1894 0.2111 0.2105 d3 Selectability measure l3s ÿ � 0.2019 0.1809 0.1996 0.2123 0.2053 Rejectability measure l3r ÿ � 0.1963 0.1980 0.1971 0.1987 0.2099 0.18 0.185 0.19 0.195 0.2 0.205 0.21 0.215 0.22 0.18 0.19 0.2 0.21 0.22 a 1 a 1 a 1 a 2 a 2 a 2 a 3 a 3 a 3 a 4 a 4 a 4 a 5 a 5 a 5 Rejectability measure S e le c ta b il it y m e a s u re d 1 d 2 d 3 q 1 q 2 q 3 Fig. 5. Graphic representation of final bipolar measures for each decision maker. 8. Consensus building process based on distance measures The second proposed process is based on distance measures that allow the analyst to identify alternatives and decision makers with strong differences. Using a feedback mechanism, decision makers have the ability to change their assessments based on rec- ommendations made by the analyst during the discussion sessions, with the aim to converge to a common solution. The first phase of the feedback mechanism uses proximity mea- sures and bipolar consensus to identify respectively, divergent alternatives and decision makers with assessments that deviate from those of other decision makers. The second phase allows the analyst to make targeted recommendations to divergent deci- sion makers. Identification phase. The identification of the alternatives with strong differences consists in calculating proximity measures defined by Eq. (25) from the final bipolar measures (Table 4). The obtained results are given by the following vector di = [0.0077 0.0079 0.0119 0.0113 0.0090]. Assuming that the average distances on the set of alternatives is the tolerance threshold, the proximity distance must not exceed 0.0096 (di 6 0.0096). We can deduce that the alternative a3 and a4 have widely divergent and should there- fore be discarded. Assuming that the thresholds ùs, ùr were obtained from averages of bipolar distances on the set of alterna- tives, the following Table 5 shows the gaps observed at the actor level for each alternative. Recommandation phase. Table 5 shows that decision maker d1 presents a deviation from the average concerning selectability measure of alternative a2. The meaning of the divergence is posi- tive div 1 S2 ¼ 0:003 � � , the selectability measure is important and cannot be modified. - Decision maker d2 presents a divergence regarding the rejec- tability measures of alternatives a1 and a5. The divergence direction of the rejectability measures is given by div 2 r1 ¼ 0:0035 and div 2 r5 ¼ ÿ0:0015. The negative divergence of alternative a5 leads to a recommendation to reduce its rejec- tability measure. Alternative a1 which has a low rejectability is spared. - Decision maker d3 presents a strong rejectability measure on alternative a5 compared to the rest of the group. The negative divergence direction ðdiv 3 r5 ¼ ÿ0:0042Þ implies a recommenda- tion of reducing this measure. The reduction of rejectability measures of alternatives a5 by decision makers d2 and d3 to �ı 2 r ða5Þ ¼ 0:2005 and �ı 3 r ða5Þ ¼ 0:1980 respectively allows for the following graphical representation (see Fig. 6). The satisficing equilibrium sets ðeS;jq Þ of each decision makers dk are deduced as follows; e S;1 1 ¼ fa5; a2; a4g, e S;2 1 ¼ fa3; a5g and eS;31 ¼ fa1; a4; a5g. The solution obtained by the intersection of the sets is the alternative a5 T3 j¼1e S;j 1 ¼ a5 � � . In the example discussed here, the integration of positive and negative influences of decision makers in the model and the relatively small number of decision makers allowed reaching a consensus quickly after a single recommendation step. The individ- ualism average rate considered for all decision makers allows also to nuance the individual assessments and reduce differences that a high degree of individualism could make appear as shown in Fig. 7 for individualism degrees dj equal to 0.9 for each decision maker. A sensitivity analysis can be performed by varying the caution index and/or individualism degree to test different possible scenarios and stability of recommended solutions. 9. Conclusion and perspectives Considering collaborative decision problem based on individual evaluation, this paper proposes a new resolution approach to deal with group decision problems in multicriteria framework. To more realistic model, human behavior aspects (positive and negative influences, selfish, prudence, etc.) are integrated in the evaluation and recommendation phases. Based on bipolar context, local preferences which considers the vicinity influence are expressed by selectability and rejectability measures, to avoid compensations. A consensus building processes are proposed from final bipolar measures in order to guide decision makers toward a common solution. Based on the formalism of the satisficing game theory, common solutions were represented by the result of the intersec- tion of satisficing equilibrium sets of decision makers. The first pro- posed process for achieving consensus is based on the caution index variation, considering that the adjustments and changes in individual assessments were not admitted. In this case, decision actors had to reduce their caution index degree based on a set of selection criteria. This readjustment has expanded the satisficing equilibrium sets and increase the chances of obtaining a non- empty sets intersection. When individual assessments adjustments are possible, an iterative process of reaching consensus based on proximity and bipolar consensus measures has also been proposed to find common solutions. Table 5 Bipolar consensus measures. D d k si d k ri d1 d2 d3 d1 d2 d3 a1 0,0017 0,0019 0,0031 0,0025 0,0035 0,002 a2 0,0033 0,0022 0,0034 0,0026 0,002 0,0032 a3 0,0046 0,0024 0,0026 0,0045 0,0056 0,0034 a4 0,0039 0,0067 0,0048 0,0014 0,0012 0,0023 a5 0,0007 0,0009 0,0005 0,0057 0,0033 0,0042 0.18 0.185 0.19 0.195 0.2 0.205 0.21 0.215 0.22 0.18 0.185 0.19 0.195 0.2 0.205 0.21 0.215 0.22 a 1 a 1 a 1 a 2 a 2 a 2 a 3 a 3 a 3 a 4 a 4 a 4 a 5 a 5 a 5 Rejectability measure S e le c ta b il it y m e a s u re d 1 d 2 d 3 q 1 q 2 q 3 Fig. 6. Graphic representation of final bipolar measures (iteration1). 0.16 0.17 0.18 0.19 0.2 0.21 0.22 0.23 0.16 0.17 0.18 0.19 0.2 0.21 0.22 0.23 a 1 a 1 a 1 a 2 a 2 a 2 a 3 a 3 a 3 a 4 a 4 a 4 a 5 a 5 a 5 Rejectability measure S e le c ta b il it y m e a s u re d 1 d 2 d 3 q 1 q 2 q 3 Fig. 7. Graphical representation of final bipolar measures for dj ¼ ½0:90:90:9�. In the principle, our proposition can be considered similar to other consensus models proposed in the literature (Herrera et al., 1996a; Herrera-Viedma et al., 2002; Yeh, Kreng, & Lin, 2001) in the sense that the consensus seeking consists in adjusting initial assess- ment of some parameters obtained by actors but in practice the implementation of approaches may completely differ. For example regarding proposed model in where the principle is similar. The approach developed in this paper goes from bipolar mea- sures then obtaining a satisficing short list of alternatives by each decision makers; the adjustment is considered only with regards to these satisficing alternatives using divergence measures contrary to what is proposed in Yeh et al., 2001 for example, where a math- ematical programming problem that seeks new assessments when minimizing a deviation measure is formulated and solved by a genetic algorithm. Moreover, in contrast to soft consensus processes proposed in the literature ((Herrera-Viedma et al., 2007; Khorshid, 2010) for example), the model presented here does not aim to converge indi- vidual assessments on the set of all alternatives but focuses mainly on alternative with the same trend assessments initially (conver- gent). Targeted recommendations are given to divergent actors (with incoherent opinion compared to other members) in adaptive process. Another feature of the proposed model lies in the fact that the importance degree of actors is seen in evaluation and recom- mendation phases by integrating it in the consensus reached pro- cess unlike what is done to present literature, to our known. Although the proposed model has some new features such as those just mentioned and flexibility and ease of immersive excur- sion, some weakness can be felt when applying the approach developed in this paper particularly due to the amount of parame- ters that have to be elicited and the interpretation issues by deci- sion makers mainly when they are trained in different backgrounds. But if the process is conducted by an analyst step by step, these difficulties may be reduced. To apply the approach developed in this paper, different mate- rials must be gathered; among these materials, the notion of vicin- ity and related influence parameters elicitation may raise difficult issues such how different partners do understand this notion. Therefore, future researches must consider addressing scientifi- cally this issue by formalizing this notion building on social net- work concept for instance. Human attitude modeling is another subject that we think need to be seriously considered in the future works as well as uncertainty consideration during selection pro- cess (fuzzy selectability and rejectability measures for instance). On the other hand, consensus process assumed generally that decision makers accept to give modifications to their preferences to achieve a common solution. However, some actors can decide not to accept the recommendations (Herrera-Viedma et al., 2014). In this case, two alternatives can be studied, the first amounts to consider the refusal of these decision makers in the consensus processes though settings made according to the impor- tance of refractory actors. Or, conversely, develop a support for the consensus process based on psychology concepts (social proof, authority, linking, scarcity, consistency, etc) to convince recalci- trant decision makers (Cialdini, 2001). References Alonso, S., Herrera-Viedma, E., Chiclana, F., & Herrera, F. (2010). A web based consensus support system for group decision making problems and incomplete preferences. Information Sciences, 180(23), 4477–4495. Alonso, S., & Pérez, I. J. (2013). A linguistic consensus model for Web 2.0 communities. Applied Soft Computing, 13(1), 149–157. Bard, J. F., & Sousk, S. F. (1990). A tradeoff analysis for rough terrain cargo handlers using the AHP: An example of group decision making. IEEE Transactions on Engineering Management, 37(3), 222. Ben-Arieh, D., & Chen, Z. (2006). Linguistic-labels aggregation and consensus measure for autocratic decision making using group recommendations. IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, 36(3), 558–568. Ben-Arieh, D., Easton, T., & Evans, B. (2009). Minimum cost consensus with quadratic cost functions. IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, 39(1), 210–217. Bonano, E. J., & Apostolakis, G. E. (1991). Theoretical foundations and practical issues for using expert judgements in uncertainty analysis of high-level radioactive waste disposal. Radioactive Waste Management and the Nuclear Fuel Cycle, 16(2), 137–159. Bordogna, G., Fedrizzi, M., & Pasi, G. (1997). A linguistic modeling of consensus in group decision making based on OWA operators. IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, 27(1), 126–133. Bouzarour-Amokrane, Y., Tchangani, A., & Pérès, F. (2012). Defining and measuring risk and opportunity in BOCR framework for decision analysis. In Proceedings of 9th international conference on modeling, optimization and simulation, MOSIM 2012, Bordeaux, France. Bouzarour-Amokrane, Y., Tchangani, A., & Pérès, F. (2013). Résolution des problémes de décision de groupe par analyse bipolaire. In 5émes Journées Doctorales (JDJN) GDRMACS, Strasbourg, France. Bryson, N. (1996). Group decision-making and the analytic hierarchy process: Exploring the consensus-relevant information content. Computers & Operations Research, 23(1), 27–35. Butler, C. L., & Rothstein, A. (1987). On conflict and consensus: A handbook on formal consensus decision making. Food Not Bombs Publishing. Cabrerizo, f. J., alonso, s., & herrera-viedma, E. (2009). A consensus model for group decision making problems with unbalanced fuzzy linguistic information. International Journal of Information Technology & Decision Making, 08(01), 109–131. Cabrerizo, F. J., Pérez, I. J., & Herrera-Viedma, E. (2010). Managing the consensus in group decision making in an unbalanced fuzzy linguistic context with incomplete information. Knowledge-Based Systems, 23(2), 169–181. Carlsson, C., Ehrenberg, D., Eklund, P., Fedrizzi, M., Gustafsson, P., Lindholm, P., et al. (1992). Consensus in distributed soft environments. European Journal of Operational Research, 61(1–2), 165–185. Chiclana, H.-V. H., Herrera-viedma, E., Herrera, F., & Chiclana, F. (2002). A consensus model for multiperson decision making with different preference structures. IEEE Transactions on Systems, Man and Cybernetics-Part A: Systems and Humans, 32, 394–402. Chiclana, F., Mata, F., Martinez, L., Herrera-Viedma, E., & Alonso, S. (2008). Integration of a consistency control module within a consensus model. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 16, 35. Choudhury, A. K., Shankar, R., & Tiwari, M. K. (2006). Consensus-based intelligent group decision-making model for the selection of advanced technology. Decision Support Systems, 42(3), 1776–1799. Cialdini, R. B. (2001). The science of persuasion. Scientific American, pp. 76–81.. Coch, L., & French, J. R. P. (1948). Overcoming resistance to change. Human Relations, 1(4), 512–532. Cook, W. D., & Kress, M. (1985). Ordinal ranking with intensity of preference. Management Science, 31(1), 26–32. De Brucker, K. & C. Macharis. (2010). Multi-actor multi-criteria analysis (MAMCA) as a means to cope with societal complexity. In EURO XXIV, Lisbon, Portugal. Degroot, M. H. (1974). Reaching a consensus. Journal of the American Statistical Association, 69(345), 118–121. Eklund, P., Rusinowska, A., & de. Swart, H. (2008). A consensus model of political decision-making. Annals of Operations Research, 158(1), 5–20. Fedrizzi, M., Kacprzyk, J., & Zadro_zny, S. (1988). An interactive multi-user decision support system for consensus reaching processes using fuzzy logic with linguistic quantifiers. Decision Support Systems, 4(3), 313–327. French, J. R. P. (1956). A formal theory of social power. Psychological Review, 63(3), 181–194. French, S. (1981). Consensus of opinion. European Journal of Operational Research, 7(4), 332–340. Fu, C., & Yang, S.-L. (2010). The group consensus based evidential reasoning approach for multiple attributive group decision analysis. European Journal of Operational Research, 206(3), 601–608. Fu, C., & Yang, S. (2012). An evidential reasoning based consensus model for multiple attribute group decision analysis problems with interval-valued group consensus requirements. European Journal of Operational Research, 223(1), 167–176. Gong, Z.-W., Forrest, J., & Yang, Y.-J. (2013). The optimal group consensus models for 2-tuple linguistic preference relations. Knowledge-Based Systems, 37, 427–437. Harary, F. (1959). On the measurement of structural balance. Behavioral Science, 4(4), 316–323. Hatami-Marbini, A., & Tavana, M. (2011). An extension of the Electre I method for group decision-making under a fuzzy environment. Omega, 39(4), 373–386. Herrera, F., Alonso, S., Chiclana, F., & Herrera-Viedma, E. (2009). Computing with words in decision making: Foundations, trends and prospects. Fuzzy Optimization and Decision Making, 8(4), 337–364. Herrera, F., & Herrera-Viedma, E. (2000). Linguistic decision analysis: Steps for solving decision problems under linguistic information. Fuzzy Sets and Systems, 115(1), 67–82. Herrera, F., Herrera-Viedma, E., & Chiclana, F. (2001). Multiperson decision-making based on multiplicative preference relations. European Journal of Operational Research, 129(2), 372–385. Herrera, F., Herrera-Viedma, E., & Verdegay, J. L. (1996a). A model of consensus in group decision making under linguistic assessments. Fuzzy Sets and Systemss, 78(1), 73–87. Herrera, F., Herrera-Viedma, E., & Verdegay, J. L. (1996b). Direct approach processes in group decision making using linguistic OWA operators. Fuzzy Sets and Systems, 79(2), 175–190. Herrera, F., Herrera-Viedma, E., & Verdegay, J. L. (1997a). A rational consensus model in group decision making using linguistic assessments. Fuzzy Sets and Systems, 88(1), 31–49. Herrera, F., Herrera-Viedma, E., & Verdegay, J. L. (1997b). Linguistic measures based on fuzzy coincidence for reaching consensus in group decision making. International Journal of Approximate Reasoning, 16(3–4), 309–334. Herrera-Viedma, E., Alonso, S., Chiclana, F., & Herrera, F. (2007). A consensus model for group decision making with incomplete fuzzy preference relations. IEEE Transactions on Fuzzy Systems, 15, 863–877. Herrera-Viedma, E., Cabrerizo, F. J., Kacprzyk, J., & Pedrycz, W. (2014). A review of soft consensus models in a fuzzy environment. Information Fusion, 17, 4–13. Herrera-Viedma, E., Herrera, F., & Chiclana, F. (2002). A consensus model for multiperson decision making with different preference structures. IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, 32(3), 394–402. Herrera-Viedma, E., Martinez, L., Mata, F., & Chiclana, F. (2005). A consensus support system model for group decision-making problems with multigranular linguistic preference relations. IEEE Transactions on Fuzzy Systems, 13(5), 644–658. Hsu, Hsi-Mei, & Chen, Chen-Tung (1996). Aggregation of fuzzy opinions under group decision making. Fuzzy Sets and Systems, 79(3), 279–285. Jiang, Y., Xu, Z., & Yu, X. (2013). Compatibility measures and consensus models for group decision making with intuitionistic multiplicative preference relations. Applied Soft Computing (0). Kacprzyk, J. (1986). Group decision making with a fuzzy linguistic majority. Fuzzy Sets and Systems, 18(2), 105–118. Kacprzyk, J. (1987). On some fuzzy cores and ‘‘soft’’ consensus measures in group decision making, 119–130. Kacprzyk, J., & Fedrizz, M. (1986). Soft consensus measure for monitoring real consensus reaching processes under fuzzy preferences. Control and Cybernetics, 309–323. Kacprzyk, J., & Fedrizzi, M. (1988). A ‘soft’ measure of consensus in the setting of partial (fuzzy) preferences. European Journal of Operational Research, 34(3), 316–325. Kacprzyk, J., & Fedrizzi, M. (1989). A ‘human-consistent’ degree of consensus based on fuzzy login with linguistic quantifiers. Mathematical Social Sciences, 18(3), 275–290. Kacprzyk, J., Fedrizzi, M., & Nurmi, H. (1992). Group decision making and consensus under fuzzy preferences and fuzzy majority. Fuzzy Sets and Systems, 49(1), 21–31. Khorshid, S. (2010). Soft consensus model based on coincidence between positive and negative ideal degrees of agreement under a group decision-making fuzzy environment. Expert Systems with Applications, 37(5), 3977–3985. Kline, J. A. (1972). Orientation and group consensus. Central States Speech Journal, 23(1), 44–47. Korpela, J., & Tuominen, M. (1997). Group decision support for analysing logistics development projects. In Proceedings of the thirtieth Hawaii international conference on system sciences (Vol. 2, pp. 493–502). Ko, M., Tiwari, A., & Mehnen, J. (2010). A review of soft computing applications in supply chain management. Applied Soft Computing, 10(3), 661–674. Lai, V. S., Wong, B. K., & Cheung, W. (2002). Group decision making in a multiple criteria environment: A case using the AHP in software selection. European Journal of Operational Research, 137(1), 134–144. Lee, H.-S. (2002). Optimal consensus of fuzzy opinions under group decision making environment. Fuzzy Sets and Systems, 132(3), 303–315. Lee, A. H. I., Chen, H., & Kang, H. (2009). Multi-criteria decision making on strategic selection of wind farms. Renewable Energy, 34(1), 120–126. Lehrer, K., & Wagner, C. (1981). Rational consensus in science and society: A philosophical and mathematical study. Springer Science & Business Media. Leyva Lopez, J. C. (2010). A consensus model for group decision support based on valued outranking relations. In EURO XXIV, Lisbon, Portugal. Li, D.-F. (2007). Compromise ratio method for fuzzy multi-attribute group decision making. Applied Soft Computing, 7(3), 807–817. Loewer, B., & Laddaga, R. (1985). Destroying the consensus. Synthese, 62(1), 79–95. Mata, F., Martinez, L., & Herrera-Viedma, E. (2009). An adaptive consensus support model for group decision-making problems in a multigranular fuzzy linguistic context. IEEE Transactions on Fuzzy Systems, 17(2), 279–290. Moon, J. H., & Kang, C. S. (1999). Use of fuzzy set theory in the aggregation of expert judgments. Annals of Nuclear Energy, 26(6), 461–469. Pal, S. K., & Ghosh, A. (2004). Soft computing data mining. Information Sciences, 163(1–3), 444. Palomares, I., Estrella, F. J., Martínez, L., & Herrera, F. (2014). Consensus under a fuzzy context: Taxonomy, analysis framework AFRYCA and experimental case of study. Information Fusion, 20, 252–271. Palomares, I., Martinez, L., & Herrera, F. (2014). A consensus model to detect and manage noncooperative behaviors in large-scale group decision making. IEEE Transactions on Fuzzy Systems, 22(3), 516–530. Park, J. H., Park, I. Y., Kwun, Y. C., & Tan, X. (2011). Extension of the TOPSIS method for decision making problems under interval-valued intuitionistic fuzzy environment. Applied Mathematical Modelling, 35(5), 2544–2556. Parreiras, R., Ekel, P., & Bernardes, F. Jr., (2012). A dynamic consensus scheme based on a nonreciprocal fuzzy preference relation modeling. Information Sciences, 211, 1–17. Pérez, I. J., Cabrerizo, F. J., & Herrera-Viedma, E. (2011). Group decision making problems in a linguistic and dynamic context. Expert Systems with Applications, 38(3), 1675–1688. Ragade, R. K. (1976). Fuzzy sets in communication systems and in consensus formation systems. Cybernetics and Systems, 6, 21–38. Roselló, L., Prats, F., Agell, N., & Sánchez, M. (2010). Measuring consensus in group decisions by means of qualitative reasoning. International Journal of Approximate Reasoning, 51(4), 441–452. Roubens, M. (1997). Fuzzy sets and decision analysis. Fuzzy Sets and Systems, 90(2), 199–206. Saaty, T. (2001). The analytic network process. Fundamentals of decision making and priority theory. Saaty, T., & Özdemir, M. (2005). The encyclicon; a dictionary of applications of decision making with dependence and feedback based on the analytic network process. RWS Publications. Saint, S. (1994). Rules for reaching consensus: A modern approach to decision making (1st ed.). Amsterdam, San Diego: Pfeiffer. Szmidt, E., & Kacprzyk, J. (2003). A consensus-reaching process under intuitionistic fuzzy preference relations. International Journal of Intelligent Systems, 18(7), 837–852. Tapia GarcíA, J. M., Del Moral, M. J., MartíNez, M. A., & Herrera-Viedma, E. (2012). A consensus model for group decision-making problems with interval fuzzy preference relations. International Journal of Information Technology & Decision Making, 11(04), 709–725. Tavana, M., Kennedy, D. T., & Joglekar, P. (1996). A group decision support framework for consensus ranking of technical manager candidates. Omega, 24(5), 523–538. Tchangani, A. (2009). Evaluation model for multiattributes-multiagents decision making: Satisficing game approach. International Journal of Information Technology and Decision Making, 8(1), 73–91. Tchangani, A. (2013). Bipolarity in decision analysis: A way to cope with human judgment, exploring innovative and successful applications of soft computing. IGI Global, 216–244. Tchangani, A., Bouzarour-Amokrane, Y., & Pérés, F. (2012). Evaluation model in decision analysis: Bipolar approach. Informatica, 23(3), 461–485. Tchangani, A., Bouzarour-Amokrane, Y., & Pérés, F. (2011). Bipolar evaluation model in decision analysis. In 12éme Congrés annuel de la société française de recherché opérationnelle et d’aide á la décision, ROADEF 2011, Saint Etienne, France. Urfalino, P. (2007). L’esprit des régles de décision. Vincent Descombes. Questions disputées, 199235. Verma, D. (2009). Decision making style: Social and creative dimensions. Global India Publications. Wheeler, T. A., Hora, S. C., Cramond, W. R., & Unwin, S. D. (1989). Analysis of core damage frequency from internal events: Expert judgment elicitation: Part 1, Expert panel results, Part 2, Project staff results. Wijnmalen, D. J. D. (2007). Analysis of benefits, opportunities, costs, and risks (BOCR) with the AHP-ANP: A critical validation. Mathematical and Computer Modelling, 46(7-8), 892–905. Wu, Z., & Xu, J. (2012). A concise consensus support model for group decision making with reciprocal preference relations based on deviation measures. Fuzzy Sets and Systems, 206, 58–73. Xu, Y., Li, K. W., & Wang, H. (2013). Distance-based consensus models for fuzzy and multiplicative preference relations. Information Sciences, 253, 56–73. Xu, J., & Wu, Z. (2011). A discrete consensus support model for multiple attribute group decision making. Knowledge-Based Systems, 24(8), 1196–1202. Yager, R. R. (1988). On ordered weighted averaging aggregation operators in multicriteria decisionmaking. IEEE Transactions on Systems, Man and Cybernetics, 18(1), 183–190. Yeh, J.-M., Kreng, B., & Lin, C. (2001). A consensus approach for synthesizing the elements of comparison matrix in the Analytic Hierarchy Process. International Journal of Systems Science, 32(11), 1353–1363. Yue, Z. (2011a). An extended TOPSIS for determining weights of decision makers with interval numbers. Knowledge-Based Systems, 24(1), 146–153. Yue, Z. (2011b). Deriving decision maker’s weights based on distance measure for interval-valued intuitionistic fuzzy group decision making. Expert Systems with Applications, 38(9), 11665–11670. Yue, Z. (2012a). Approach to group decision making based on determining the weights of experts by using projection method. Applied Mathematical Modelling, 36(7), 2900–2910. Yue, Z. (2012b). Extension of TOPSIS to determine weight of decision maker for group decision making problems with uncertain information. Expert Systems with Applications, 39(7), 6343–6350. Yue, Z., & Jia, Y. (2012). An application of soft computing technique in group decision making under interval-valued intuitionistic fuzzy environment. Applied Soft Computing (0). Yu, L., & Lai, K. K. (2011). A distance-based group decision-making methodology for multi-person multi-criteria emergency decision support. Decision Support Systems, 51(2), 307–315. Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338–353. 
work_w7p2v35cpfenzaihw4nyphoxeq	----	doi:10.1016/j.eswa.2007.05.013 ARTICLE IN PRESS www.elsevier.com/locate/eswa Expert Systems with Applications xxx (2007) xxx–xxx Expert Systems with Applications Collaborative recommender systems: Combining effectiveness and efficiency Panagiotis Symeonidis *, Alexandros Nanopoulos, Apostolos N. Papadopoulos, Yannis Manolopoulos Department of Informatics, Aristotle University, 54124 Thessaloniki, Greece Abstract Recommender systems base their operation on past user ratings over a collection of items, for instance, books, CDs, etc. Collabora- tive filtering (CF) is a successful recommendation technique that confronts the ‘‘information overload’’ problem. Memory-based algo- rithms recommend according to the preferences of nearest neighbors, and model-based algorithms recommend by first developing a model of user ratings. In this paper, we bring to surface factors that affect CF process in order to identify existing false beliefs. In terms of accuracy, by being able to view the ‘‘big picture’’, we propose new approaches that substantially improve the performance of CF algo- rithms. For instance, we obtain more than 40% increase in precision in comparison to widely-used CF algorithms. In terms of efficiency, we propose a model-based approach based on latent semantic indexing (LSI), that reduces execution times at least 50% than the classic CF algorithms. � 2007 Elsevier Ltd. All rights reserved. Keywords: Recommender system; Collaborative filtering; Nearest neighbors 1. Introduction The ‘‘information overload’’ problem affects our every- day experience while searching for knowledge on a topic. To overcome this problem, we often rely on suggestions from others who have more experience on the topic. How- ever, in Web case where there are numerous suggestions, it is not easy to detect the trustworthy ones. Shifting from individual to collective suggestions, the process of recom- mendation becomes controllable. This is attained with the introduction of CF, which provides recommendations based on the suggestions of users who have similar prefer- ences. Since CF is able to capture the particular preferences of a user, it has become one of the most popular methods in recommender systems. 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.05.013 * Corresponding author. E-mail addresses: symeon@csd.auth.gr (P. Symeonidis), ananopou@ csd.auth.gr (A. Nanopoulos), papadopo@csd.auth.gr (A.N. Papadopou- los), manolopo@csd.auth.gr (Y. Manolopoulos). Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 Two types of CF algorithms have been proposed in the literature: memory-based algorithms, which recommend according to the preferences of nearest neighbors, and model-based algorithms, which recommend by first devel- oping a model of user ratings. Both practical experience and related research have reported that memory-based algorithms (a.k.a. nearest-neighbor algorithms) present excellent performance, in terms of accuracy, for multi- value rating data. On the other hand, model-based algo- rithms are efficiently handle scalability to large data sets. 1.1. Motivation Nearest-neighbor CF is influenced by several factors. Related research on CF, during the past decade, approached some of these factors. However, existing approaches may not be considered complete, because they examine the various factors only partially. More specifi- cally, existing CF algorithms and their experimental evalu- ation focus only on parts of the CF process and do not e recommender systems: Combining effectiveness ..., Expert Sys- mailto:symeon@csd.auth.gr mailto:ananopou@ csd.auth.gr mailto:ananopou@ csd.auth.gr mailto:papadopo@csd.auth.gr mailto:manolopo@csd.auth.gr 2 P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS handle it as a whole. For the aspects that these partial con- siderations do not examine, they usually make choices, which our study demonstrates that can be misleading. Through our study we are also able to confirm that there exist dependencies between the factors affecting CF. There- fore, we have to perform an evaluation of the entire CF process in order to produce reliable conclusions. Moreover, to handle scalability, we have to extend our findings for nearest-neighbor CF algorithms through a model-based approach. This approach will combine the effectiveness of the nearest-neighbor CF algorithms in terms of accuracy, with the efficiency in terms of execution time. Towards this direction, latent semantic indexing (LSI) is a technique that has been extensively used in infor- mational retrieval. LSI detects latent relationships between documents and terms. In CF, LSI can be used to form users’ trends from individual preferences, by detecting latent relationships between users and items. Therefore, with LSI, a higher level representation of the original user-item matrix is produced, which presents a three-fold advantage: (i) it contains the main trends of users’ prefer- ences, (ii) noise is removed, (iii) it is much more condensed than the original matrix, thus it favors scalability. 1.2. Contributions In this work, first, we provide a thorough analysis of the factors involved in CF. Notably, we examine several simi- larity measures, various criteria for generating the recom- mendation list, the appropriateness of evaluation metrics, and the impact of CF in real-world applications, which is considered through user’s satisfaction (measured with the popularity of recommended items) and the division of the ratings of the test user in past and future sets. During the analysis we identify choices that have been incorrectly adopted and new issues that have not been considered so far. As a result, we propose several extensions and new approaches, which substantially improve the entire CF process. Moreover, we propose a new model-based CF approach, which is based on LSI to produce a condensed model for the user-item matrix that handles the factor of scalability. Our contributions are summarized as follows: • The proposed approach examines the factors involved through the entire CF process. This helps to: (a) reveal fallacies in existing beliefs for several of them, (b) better analyze their impact and provide insights, and (c) syn- thesize a novel method, which substantially improve the effectiveness of CF in terms of accuracy. • The previous findings are extended with an approach based on LSI. This way, execution times for CF are sig- nificantly reduced. Moreover, improvement is possible in effectiveness too, because the proposed model identi- fies the main trends and removes noise. Notice that dif- ferently from similar approaches in related work, we introduce the notion of pseudo-user in order to fold-in Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 the vectors of target users in the produced model (related work incorrectly assumes the knowledge of tar- get users during the building of the model). • We carried out extensive experimental evaluation, which, to our knowledge, considers all the involved fac- tors for the first time. Our experimental results demon- strate the superiority of the proposed methods (more than 40% improvements in terms of precision over widely-used CF algorithms and 50% in terms of execu- tion times). The rest of this paper is organized as follows. Section 2 summarizes the related work, whereas Section 3 contains the analysis of the CF factors. The proposed approaches are described in Sections 4 and 5. Experimental results are given in Section 6. Finally, Section 7 concludes this paper. 2. Related work In 1992, the Tapestry system (Goldberg, Nichols, Brian, & Terry, 1992) introduced collaborative filtering (CF). In 1994, the GroupLens system (Resnick, Iacovou, Suchak, Bergstrom, & Riedl, 1994) implemented a CF algorithm based on common users preferences. Nowadays, it is known as user-based CF algorithm, because it employs users’ similarities for the formation of the neighborhood of nearest users. Since then, many improvements of user- based algorithm have been suggested, e.g., (Breese, Hecker- man, & Kadie, 1998; Herlocker, Konstan, Borchers, & Riedl, 1999). In 2001, another CF algorithm was proposed. It is based on the items’ similarities for a neighborhood generation of nearest items (Karypis, 2001; Sarwar, Karypis, Konstan, & Riedl, 2001) and is denoted as item-based CF algorithm. Most recent work followed the two aforementioned directions (i.e., user-based and item-based). Herlocker, Konstan, and Riedl (2002) weight similarities by the num- ber of common ratings between users/items. Deshpande and Karypis (2004) apply item-based CF algorithm combined with conditional-based probability similarity and cosine-similarity. Xue, Lin, and Yang (2005) suggest a hybrid integration of aforementioned algorithms (mem- ory-based) with model-based algorithms. All aforementioned algorithms are memory-based. Their efficiency is affected from scalability of data. This means that they face performance problems, when the vol- ume of data is extremely big. To deal with this problem, many model-based algorithms have been developed (Breese et al., 1998). However, there are two conflicting challenges. If an algorithm spends less execution time, this should not worse its quality. The best result would be to improve qual- ity with the minimum calculation effort. Furnas, Deerwester, and Dumais (1988) proposed latent semantic indexing (LSI) in information retrieval area to deal with the aforementioned challenges. More specifically, LSI uses SVD to capture latent associations between the terms e recommender systems: Combining effectiveness ..., Expert Sys- Table 1 Factors affect CF algorithms Factor name Short description Stage Sparsity Limited percentage of rated products 1 Scalability Computation increase by the number of users and items 1 Train/test data size Data are divided to training and evaluation or test 1,3 Neighborhood size Number of neighbors used for the neighborhood formation 1 Similarity measure Measures that calculate proximity of two objects 1 Recommendation list size Number of top-N recommended items 2 Generation of recommendation list Algorithms for the top-N list generation 2 Positive rating- threshold Positive and negative ratings segregation 2,3 Evaluation metrics Metrics that evaluate the quality of top- N list 3 Setting a baseline method A simple method against which performance is compared 3 Past/future items The segregation between apriori known and unknown items 3 Table 2 Symbols and definitions Symbol Definition Symbol Definition P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx 3 ARTICLE IN PRESS and the documents. SVD is a well-known factorization tech- nique that factors a matrix into three matrices. Berry, Dum- ais, and O’Brien (1994) carried out a survey of the computational requirements for managing (e.g., folding- in1) LSI-encoded databases. He claimed that the reduced- dimensions model is less noisy than the original data. Sarwar, Karypis, and Konstan (2000, 2002) applied dimensionality reduction for the user based CF approach. He also used SVD for generating predictions. In contrast to our work, Sarwar et al. (2000, 2002) do not consider two significant issues: (i) Predictions should be based on the users’ neighbors and not on the test (target) user, as the ratings of the latter are not apriori known. For this rea- son we rely only on the neighborhood of the test user. (ii) The test users should not be included in the calculation of the model, because they are not known during the fac- torization phase. For this reason, we introduce the notion of pseudo-user in order to include a new user in the model (folding in), from which recommendations are derived. Other related work also includes Goldberg, Roeder, Gupta, and Perkins (2001), who applied Principal Compo- nents Analysis (PCA) to facilitate off-line dimensionality reduction for clustering the users, and therefore, manages to have rapid on-line computation of recommendations. Hofmann (2004) proposed a model-based algorithm which relies on latent semantic and statistical models. k Number of nearest neighbors �ri Mean rating value for item i N Size of recommendation list pu,i Predicted rate for user u on item i NN(u) Nearest neighbors of user u jTj Size of the test set NN(i) Nearest neighbors of item i c Number of singular values Ps Threshold for positive ratings A Original matrix I Domain of all items U Left singular vectors of A U Domain of all users S Singular values of A u, v Some users V0 Right singular vectors of A i,j Some items A* Approximation matrix of A Iu Set of items rated by user u u User vector Ui Set of users rated item i unew Inserted user vector ru,i The rating of user u on item i n Number of training users �ru Mean rating value for user u m Number of items 3. Factors affecting the CF process In this section, we identify the major factors that criti- cally affect all CF algorithms. Our analysis focuses on the basic operations of the CF process, which consists of three stages. • Stage 1: formation of user or item neighborhood, where objects inside the neighborhood have similar ratings and behavior. • Stage 2: top-N list generation with algorithms that con- struct a list of best items recommendations for a user. • Stage 3: quality assessment of the top-N list. In the rest of this section we elaborate on the aforemen- tioned factors, which are organized with respect to the stage that each one is involved. The examined fac- tors,which are detailed in the following, are described in Table 1. Table 2 summarizes the symbols that are used in the sequel. To ease the discussion, we will use the run- ning example illustrated in Fig. 1 where U1�10 are users I1�6 are items. As shown, the example data set is divided into a training and test set. The null cells (no rating) are represented as zeros. Note that related work has identified some few more factors, like the impact of subjectivity during the rating 1 Folding in terms or documents is a simple technique that uses existing SVD to represent new information. Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 or issues concerning the preprocessing of data (Breese et al., 1998; Mobasher, Dai, Luo, & Nakagawa, 2001; Sar- war et al., 2000). Nevertheless, we do not examine these factors, because their effect is less easily determinable. 3.1. First stage factors Sparsity: In most real-world cases, users rate only a very small percentage of items. This causes data sets to become sparse. The problem of sparsity is extensively studied. In e recommender systems: Combining effectiveness ..., Expert Sys- Fig. 1. (a) Training set (n · m) and (b) test set. 2 Means �ru, �rv are the mean ratings of u and v over their co-rated items. 3 Means �ru, �rv are taken over all ratings of u and v. 4 P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS such cases, the recommendation engine cannot provide pre- cise proposals, due to lack of sufficient information. A simi- lar problem of CF algorithms is the one of cold-start (O’Mahony, Hurley, Kushmerick, & Silvestre, 2004). In few existing works (Mobasher et al., 2001; Sarwar et al., 2000), there is a preprocessing step that fills missing values. Our experimental examination indicated that this approach incurs the loss of valuable information. For the same reasons, the aforementioned direction has not been followed by forthcoming works. Several recent works (Herlocker et al., 2002; Karypis, 2001; Sarwar et al., 2001) focus only on very sparse data. Also, related work provides benchmark data sets with dif- ferent sparsity, e.g., the Jester data set (Goldberg et al., 2001) is dense; in contrast the Movielens data sets (Her- locker et al., 1999) are relatively sparse. The degree of spar- sity, however, depends on the application type. To provide complete conclusions, someone has to experiment with the amount of sparsity as well. Scalability: Scalability is important, because in real- world applications the number of users/items is very large. Related work (Sarwar et al., 2000) has proposed the use of dimensionality reduction techniques, which introduce a trade-off between the accuracy and the execution time of CF algorithms. In Section 5, we will propose a model to confront with the scalability problem. Train/test data size: There is a clear dependence between the training set’s size and the accuracy of CF algorithms (Sarwar et al., 2001). Through our experimental study we verified this conclusion. Additionally, we saw that after an upper threshold of the training set size, the increase in accuracy is less steep. However, the effect of overfitting is less significant compared to general classification problems. In contrast, low training set sizes negatively impact accu- racy. Therefore, the fair evaluation of CF algorithms should be based on adequately large training sets. Though most related research uses a size around 80%, there exist works that use significantly smaller sizes (McLauglin & Herlocher, 2004). From our experimental results we con- cluded that an 75% training set size corresponds to an ade- quate choice. But we have to notice that training/test size should not be data set independent. (In the running exam- ple, we set training size at 60%). Neighborhood size: The number, k, of nearest neighbors used for the neighborhood formation produces a trade-off: a very small k results to low accuracy, because there are not enough neighbors to base prediction. In contrast, a very Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 large k impacts precision too, as the particularities of user’s preferences can be blunted due to the large neighborhood size. In most related works (Herlocker et al., 1999; Sarwar, Karypis, Konstan, & Riedl, 2000), k has been examined in the range of values between 10 and 100. The optimum k depends on the data characteristics (e.g., sparsity). There- fore, CF algorithms should be evaluated against varying k, in order to tune it(In the running example, we set k = 3). Similarity measure: Related work (Herlocker et al., 2002; McLauglin & Herlocher, 2004; Mobasher et al., 2001; Sarwar et al., 2001) has mainly used Pearson correla- tion and cosine-similarity. In particular, user-based (UB) CF algorithms use the Pearson correlation (Eq. (1)),2 which measures the similarity between two users, u and v. Item- based (IB) CF algorithms use a variation of adjusted cosine-similarity (Eq. (2)),3 which measures the similarity between two items, i and j, and has been proved more accu- rate (McLauglin & Herlocher, 2004; Sarwar et al., 2001), as it normalizes bias from subjective ratings. simðu; vÞ¼ P 8i2Sðru;i ��ruÞðrv;i ��rvÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP 8i2Sðru;i ��ruÞ 2 q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP 8i2Sðrv;i ��rvÞ 2 q ; S ¼ I u \ I v; ð1Þ simði; jÞ¼ P 8u2Tðru;i ��ruÞðru;j ��ruÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP 8u2U iðru;i ��ruÞ 2 q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP 8u2U jðru;j ��ruÞ 2 q ; T ¼ U i \ U j: ð2Þ Herlocker et al. (2002) proposed a variation of the pre- vious measures, which henceforth is denoted as Weighted Similarity (WS). If sim is a similarity measure (e.g., Pearson or cosine), then WS is equal to maxðc;cÞ c � sim, where c is the number of co-rated items. Eq. (1) takes into account only the set of items, S, that are co-rated by both users. This, however, ignores the items rated by only one of the two users. The number of the latter items denotes how much their preferences differ. Especially for the case of sparse data, by ignoring these items we dis- card significant information. Analogous reasoning applies for Eq. (2), which considers (in the numerator) only the set of users, T, that co-rated both the examined pair of items, and for WS, which is based on Eqs. (1) or (2). To e recommender systems: Combining effectiveness ..., Expert Sys- P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx 5 ARTICLE IN PRESS address the problem, in the following, we will examine alternative definitions for S and T. Another issue that has not been precisely clarified in related work, is whether we include in the neighborhood a user or item with negative similarity. In order to improve accuracy, we suggest keeping only the positive similarities for the neighborhood formation, even if less than the spec- ified number k of neighbors remain. This approach is also followed in several works (Mobasher et al., 2001). Our experiments have shown that, by including users or items with negative similarities (just to maintain a constant neighborhood size), can impact the accuracy. Moreover, we have to mention that, for the generation process of item similarities matrix, this can be made off- line. In contrast, the user similarities matrix should be cre- ated on-line. More specifically, for the user similarities matrix we use the test data, while for the item based simi- larities table we use only training data. The application of Pearson Correlation (Eq. (1)) to the running example is depicted in Fig. 2a and the resulting k-nearest neighbors (k-NN) are given in Fig. 2b. Respec- tively, the similarities between items, calculated with the adjusted cosine measure (Eq. (2)), are given in Fig. 2c and d depicts the nearest neighbors. As only positive values of similarities are considered during the neighborhood for- mation, the items have different neighborhood size. 3.2. Second stage factors Recommendation list’s size: The size, N, of the recom- mendation list corresponds to a trade-off: With increasing N, the absolute number of relevant items (i.e., recall) is expected to increase, but their ratio to the total size of the recommendation list (i.e., precision) is expected to decrease. (Recall and precision metrics are detailed in the following.) In related work (Karypis, 2001; Sarwar et al., 2001), N usually takes values between 10 and 50. (In the running example, we set N = 2). Positive rating threshold: It is evident that recommenda- tions should be ‘‘positive’’. It is not success to recommend an item that will be rated with 1 in scale 1–5. Nevertheless, Fig. 2. (a) Users’ similarities matrix, (b) users’ nearest neighbors in descend neighbors in descending similarity matrix. Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 this issue is not clearly defined in several existing works. We argue that ‘‘negatively’’ rated items should not contrib- ute to the increase of accuracy, and we use a rating-thresh- old, Ps, to recommended items whose rating is no less than this value. If not a Ps value is used, then the results can become misleading, since negative ratings can contribute to the measurement of accuracy. Generation of recommendation list: The most often used technique for the generation of the top-N list, is the one that counts the frequency of each item inside the found neighborhood, and recommends the N most frequent ones (Sarwar et al., 2000). Henceforth, this technique is denoted as Most-Frequent item recommendation (MF). MF can be applied to both user-based and item-based CF algorithms. For example, assume that we follow the aforementioned approach for the test user U7, for k = 3 and N = 2. For the case of user-based recommendation, the top-2 list includes items I3, I6. In contrast, for the case of item-based recommendation, the top-2 list includes two items of I6 or I2 or I5 because all three have equal presence. Fig. 3 describes the corresponding two algorithms (for the user and item-based CF, respectively). We have to mention that these two algorithms, in con- trast to the past work, include, in addition, the concept of positive rating-threshold (Ps). Thus, ‘‘negatively’’ rated items do not participate to the top-N list formation. More- over, it is obvious that the generation of top-N list for the user-based approach is more complex and time consuming. The reason is that the former algorithm finds, firstly, user neighbors in the neighborhood matrix and then counts presences of items in the user-item matrix. In contrast, with the item-based approach the whole work is completed in the item neighborhood matrix. Karypis (2001) reports another technique, which addi- tionally considers the degree of similarity between items. This takes into account that the similarities of the k neigh- bors may vary significantly. Thus, for each item in the neighborhood, this technique admeasures not just their number of appearances, but the similarity of neighbors as well. The N items with the highest sum of similarities are finally recommended. Henceforth, this technique is denoted ing similarity matrix, (c) items’ similarities matrix and (d) items’ nearest e recommender systems: Combining effectiveness ..., Expert Sys- Fig. 3. Generation of top-N list based on most frequent algorithm for (a) user-based algorithm and (b) item-based algorithm. 6 P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS as Highest-Sum-of-Similarities item recommendation (HSS). HSS is applicable only to item-based CF. For our example, the top-2 list based on this algorithm includes the items I6, I2 because they have the greater sum of similar- ities. Note that for both techniques, we have to remove from the recommendation items, these that are already rated from the test user. Based on the aforementioned rationale, we wanted to perform an examination of other additional criteria against MF (used in the majority of existing works), in order to examine if this direction is promising for future work, because besides (Karypis, 2001), very few works elaborate on this issue. 3.3. Third stage factors Evaluation metrics: For the evaluation of CF algorithms several metrics have been used in related work (Herlocker et al., 2002; Herlocker, Konstan, Terveen, & Riedl, 2004), for instance the Mean Absolute Error (MAE) or the Receiving Operating Characteristic (ROC) curve. Although MAE has been used in most of related works, it has received criticism as well (McLauglin & Herlocher, 2004). From our experimental study we understood that MAE is able to characterize the accuracy of prediction, but is not indicative for the accuracy of recommendation, as algorithms with worse MAE many times produce more accurate recommendations than others with better MAE. Since in real-world recommender systems the experience of users mainly depends on the accuracy of recommenda- tion, MAE may not be the preferred measure. Other exten- sively used metrics are precision and recall. These metrics are simple, well known, and effective to measure the accu- racy of recommendation procedure. For a test user that receives a top-N recommendation list, let R denote the number of relevant recommended items (the items of the top-N list that are rated higher than Ps by the test user). We define the following: Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 • Precision is the ratio of R to N. • Recall is the ratio of R to the total number of relevant items for the test user (all items rated higher than Ps by him). Notice that with the previous definitions, when an item in the top-N list is not rated at all by the test user, we con- sider it as irrelevant/and it counts negatively to precision (as we divide by N). In the following we also use F1 = 2 Æ recall Æ precision/(recall + precision). F1 is used because it combines both the previous metrics. Recently, (McLauglin & Herlocher, 2004) has proposed a modified precision metric for evaluating user experience to address obscure recommendations that cannot been identified with the MAE measure. To keep comparisons with prior work clear, we focused on the traditional preci- sion/recall metrics. However, we also consider the issue of items’ popularity in order to test if the examined algo- rithms recommend obscure items or not. Setting a baseline method: Existing experimental evalua- tions lack the comparison against a baseline algorithm. The baseline algorithm has to be simple and to indicate what can be attained with as little effort as possible. Through a baseline, we can see the actual improvement due to existing CF algorithms. Past/future data: In real-world applications, recommen- dations are derived only from the currently available rat- ings of the test user. However, in most of related works, all the ratings of each test user is considered apriori known. For a more realistic evaluation, recommendations should consider the division of items of the test user into two sets (Huang, Chen, & Zeng, 2004): (i) the past items of the test user and (ii) the future items of the test user. As most exist- ing works ignore this division, their reported performance corresponds to the best case, because they exploit apriori known information. To make it more clear, let’s see an example. Assume that a test user has rated 10 items. According to most previous works, all the 10 ratings are e recommender systems: Combining effectiveness ..., Expert Sys- P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx 7 ARTICLE IN PRESS used to calculate his similarity with each training user. In contrast, we want to measure how algorithms behave when we use a smaller number of items than those he will finally rate. This simulates the real-world applications, where the test user gradually rates items and receives recommenda- tions before he provides all his ratings. In the previous example, if we use two past-items, we have to compute sim- ilarities based only on the provided two ratings and we have to predict the remaining eight items (those in the future set). If only a fraction of items is included in the past set, then the accuracy is expected to decrease compared to the best case. For this reason, we study the effect of this issue. 4. Proposed extensions for memory-based algorithms In the sequel, we describe in more detail our proposed extensions. These extensions are based on the ‘‘big picture’’ of the CF process, which was obtained from the investiga- tion of factors described in Section 3. 4.1. First stage extension : the UNION similarity measures We first examined the two most important factors that are involved in the first stage: sparsity and the similarity measure. As mentioned, for the formation of sets S and T (see Eqs. (1) and (2)), past work (Herlocker et al., 2004; McLauglin & Herlocher, 2004; Sarwar et al., 2001) takes into account only the items that are co-rated by both users. Example a: In Fig. 4, Example a depicts the ratings of four users, U1–U4, over six items. Comparing the similarity of U1 and U2 users, when only co-rated items are consid- ered, then the similarity measure will be computed based only on the ratings for I1 and I3. In case of sparse data, we have a very small amount of provided ratings to compute the similarity measure. By additionally constraining S and T with co-rated items only, we reduce further the effective amount of used information. To avoid this, we consider alternative definitions for S and T, given in Eq. (3): S ¼ I u [ I v; T ¼ U i [ U j ð3Þ According to Eq. (3), S includes items rated by at least one of the users. In the Example a of Fig. 4, except the rat- ings for I1 and I3, the ratings for I2 and I4 will be considered Fig. 4. (a) Example of the ratings of four users over six items (dash denotes an empty value) and (b) example of items rated positively by a test user Utest and two other users U1 and U2. Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 too (the issue of how to treat items rated by only one user, will be discussed in the following). Similar reasoning is fol- lowed for the set T, in the case of IB CF. By combining the definitions of S and T given in Eq. (3) with the Pearson cor- relation and adjusted cosine-similarity measures, we get two reformed measures: UNION Pearson correlation (for UB) and UNION adjusted cosine (for IB), respectively.4 Notice that in case of UNION Pearson correlation, user mean ratings correspond to the average user ratings over all rated items. To further understand why should not base similarity only on co-rated items, consider the following example. Example b: In Fig. 4, Example b depicts the items rated positively (i.e., higher than Ps) by a test user Utest and two users, U1 and U2, belonging in the training set. Utest and U1 have co-rated items I1 and I2. Assume that Utest and U1 rated I1 with 5 in the 1–5 scale, whereas I2 have been rated by both of them with 4. Nevertheless, items I3 – I9 are rated only by Utest or U1, and not by both. In this case, the Pear- son measure of Eq. (1), which is based on co-rated items only, results to the maximum possible similarity value (i.e., equal to 1) between Utest and U1. However, this is based only on the two co-rated items and ignores the seven items that are rated only by one of them. On the other hand, assume that U2 rated I1 and I2 with 5 and 4, respec- tively. As previously, the Pearson measure of Eq. (1) results to the maximum possible similarity between Utest and U2, whereas Utest and U2 differ in three items rated by only one of them. This example reflects the impotence of Eq. (1) to capture the actual notion of similarity: despite the fact that Utest and U1 differ at seven items and Utest and U1 differ at three, it assigns the same similarity value in both cases. In the previous example, if we designate U1 as neighbor of Utest, we ignore two issues: (i) Items I3 and I4, will not be recommended to Utest by U1, as U1 has not rated them; this fact harms recall. (ii) Items I5–I9 will be recommended by U1, but as they have not been rated by Utest, this will harm precision. It follows that a desirable property from a simi- larity measure is to maximize the number, x, of items that are co-rated by the test user and each of his neighbors, rel- atively to the number, y, of items that are rated only by one of them (in the example of Fig. 4b, for Utest and U1, x = 2 and y = 7). In the best case, the ratio x/(x + y) has value equal to 1 and in the worst 0. To evaluate the previously described argument, we com- pared Pearson correlation against UNION by performing the following measurement. We used the MovieLens 100K data set and for each test user we computed its k nearest neighbors (k was set to 10) from the training set. Next, we measured x and y between each test user and each of his k neighbors. Fig. 5a illustrates for each x, the result- ing ratio x/(x + y). Fig. 5a clearly presents that Pearson 4 Henceforth, when it is clearly understood from the context whether we discuss about UB or IB, we use only the name UNION. e recommender systems: Combining effectiveness ..., Expert Sys- Fig. 5. (a) Measuring the ratio x/(x + y) and (b) impact of assigned value for unrated items. 8 P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS measure results to significantly lower ratios than UNION. This explains why UNION compares favorably to Pearson correlation in terms of precision and recall, as will be pre- sented experimentally in Section 6. (Due to lack of space we do not present the analogous comparison between adjusted cosine and UNION.) 5 If less than N items have positive ratings (i.e., not less than Ps), then less than N items remain in the list. 4.1.1. Assigning a value to unrated items To calculate the UNION Pearson correlation between two users U1 and U2, we have to assign a rating by U1 to an item that is rated only by U2 (e.g., I2 in the example of Fig. 4a) and vice-versa. The same requirement holds for the UNION adjusted cosine measure in the IB case. Notice that this problem cannot be effectively solved with existing techniques for filling missing values, because the sparsity of user-item matrices severely hinders this task (more than 99.9% missing values). For this reason we assign the same rating value to all the required cases. There could be several options for this value. For instance, in a 1– 5 scale, we can assign the 0 value, to reflect that the absence of a rating for an item denotes indifference. Assuming that user u did not rate item i, another option is to assign the average value of the provided ratings on other items by u, or the average of the provided ratings on i by all users. To examine the impact of the selected value, we mea- sured F1 versus the assigned value, which is depicted in Fig. 5b for the MovieLens 100K data set (it uses 1–5 scale). The dashed line in Fig. 5b corresponds to F1 of the Pearson correlation (it is independent from the assigned value). As shown, values between the positive threshold (in this case Ps was set to 3) and the maximum rating of the scale, result to reduced F1 (notice that this range also includes the user average rating value). The reason is that these values impinge the ability to distinguish the assigned values from the actual positive ratings. However, when we assign values smaller than Ps or outside the rating scale, F1 is not affected. The reason is that with such assigned values we do not miss the ability to distinguish the assigned values from the actual positive ratings (as the latter are always within the provided scale). Thus, we conclude that UNION is not significantly affected by the selection for the assigned ratings, as all values outside the rating scale or below Ps result to about the same F1. Even more, the values between Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 Ps and the upper limit of the scale result to significantly higher F1 than Pearson measure. Henceforth, we assume that the assigned value is equal to 0. 4.2. Second stage extensions: the HPR and HSR algorithms In Section 3.2, we described the two criteria, MF and HSS, that have been proposed so far for the generation of the top-N list. The ranking criteria are important, as they can significantly affect the performance of CF. For this reason we examined additional criteria, to see if this direction of research worths investigation. As a first extension of the existing ranking criteria, someone could use the predicted values for each item to rank them. Predicted values (McLauglin & Herlocher, 2004) are computed by Eqs. (4) and (6), for the cases of user-based and item-based CF, respectively. These equa- tions have been used in related work for the purpose of MAE calculation, whereas we use them for generation of top-N list. pu;i ¼ �ru þ P v2U simðu; vÞðrv;i ��rvÞP v2Ujsimðu; vÞj ð4Þ pu;i ¼ �ri þ P j2I simði; jÞðru;j ��rjÞP j2Ijsimði; jÞj ð5Þ MAE ¼ PT u¼1jru;i � pu;ij jT j ð6Þ Therefore, we sort (in descending order) the items according to predicted rating value, and recommend the first N of them.5 This ranking criterion, denoted as highest predicted rated item recommendation (HPR) is influenced by the good accuracy of prediction that existing related work reports through the MAE. HPR opts for recommending the items that are more probable to receive a higher rating. Nevertheless, as already mentioned, MAE is not indicative for the accuracy of recommendation. As our experimental results will demonstrate, HPR presents poor performance. This fact is another indication that MAE alone cannot e recommender systems: Combining effectiveness ..., Expert Sys- P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx 9 ARTICLE IN PRESS characterize the performance of CF algorithms. In Fig. 6a, we represent the aforementioned criterion for the user- based approach. As another criterion, which resulted from observations during our experimental investigation, we sum the positive rates of the items in the neighborhood, instead of just counting their frequency. This criterion is denoted as high- est sum of rates item recommendation (HSR). The top-N list consists of the N items with the highest sum. The intu- ition behind HSR is that it takes into account both the fre- quency (as MF) and the actual ratings, because it wants to favor items that appear most frequently in the neighbor- hood and have the best ratings. Assume, for example, an item i that has just a smaller frequency from an item j. If i is rated much higher than j, then HSR will prefer it from i, whereas MF will favor j. In the running example, for test user U8, for the user-based approach with k = 3 and N = 1, using this criterion for the export of the top-N list, we rec- ommend finally item I2 because its sum of rates in the neighborhood is 9(4 + 5) instead of item I3 where its sum of rates is 8(4 + 4). If we had used the most frequent top-N list generation these two items would have the same Fig. 7. The 20 most positively rated mo Fig. 6. Generation of top-N list based on (a) highest predicted Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 presences (2) in the neighborhood. So, both of them could be proposed. In Fig. 6b, we represent the aforementioned criterion for the user-based approach. 4.3. Third stage extension : a baseline algorithm For the third stage, we considered all the issues that are described in Section 3.3. One issue that needs further expla- nation is the definition of a baseline algorithm. We propose the one that recommends the N items that are most fre- quently rated positively in the entire training set. This algo- rithm is denoted as BL. BL is very simple and, as will be shown in our experimental results, it is quite effective. For instance, our experiments with Movielens-100K data set have shown that, with BL, when we simply propose the N = 20 most positively rated movies (20 most popular movies), precision reaches 40%. We examined these movies and verified that they are not obscure at all, as one can rec- ognize in Fig. 7. Therefore, the most frequently rated items are very probable to be selected by the majority of the users. For the aforementioned reasons, BL is a tool to clearly evaluate the actual improvement of existing CF vies of the Movielens 100K dataset. rated (HPR) algorithm and (b) highest sum of rates (HSR). e recommender systems: Combining effectiveness ..., Expert Sys- 10 P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS algorithms. We will show experimentally that, as we increase the k nearest neighbors, the performances of Pear- son and adjusted cosine become equivalent to BL. 5. Proposed method for an LSI-based algorithm Our approach, initially, applies Singular Value Decom- position (SVD) over the user-item matrix A. We tune the number, c, of principal components (i.e., dimensions) with the objective to reveal the major trends. The tuning of c is determined by the information percentage that is preserved compared to the original matrix. Therefore, a c-dimen- sional space is created and each of the c dimensions corre- sponds to a distinctive rating trend. Next, given the current ratings of the target user u, we enter pseudo-user vector in the c-dimensional space. Finally, we find the k nearest neighbors of pseudo-user vector in the c-dimensional space and apply either user- or item-based similarity to compute the top-N recommended items. Conclusively, the provided recommendations consider the existence of user rating trends, as the similarities are computed in the reduced c-dimensional space, where dimensions correspond to trends. To ease the discussion, we will use the running example illustrated in Fig. 8 where I1�4 are items and U1�4 are users. As shown, the example data set is divided into training and test set. The null cells(no rating) are presented as zeros. 5.1. Applying SVD on training data Initially, we apply SVD on training data n · m matrix A that produces three matrices. These matrices obtained by SVD can give by performing multiplication the initial matrix as the following Eq. (7) and Fig. 9 show: Fig. 8. (a) Training set (n · m) and (b) test set. Fig. 9. Example of: An·m (initial matrix A), Un·m (left singular vectors o Fig. 10. Example of: A�n�m (approximation matrix of A), Un·c (left singular vec Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 An�m ¼ U n�n � Sn�m � V 0m�m ð7Þ 5.2. Preserving the principal components It is possible to reduce the n · m matrix S to have only c largest singular values. Then, the reconstructed matrix is the closest rank-c approximation of the initial matrix A as it is shown in Eq. (8) and Fig. 10: A�n�m ¼ U n�c � Sc�c � V 0 c�m ð8Þ We tune the number, c, of principal components (i.e., dimensions) with the objective to reveal the major trends. The tuning of c is determined by the information percent- age that is preserved compared to the original matrix. Therefore, a c-dimensional space is created and each of the c dimensions corresponds to a distinctive rating trend. We have to notice that in the running example we create a 2-dimensional space using 83% of the total information of the matrix (12,88/15,39). In our experiments we have seen that only a 10% is adequate to provide accurate results. 5.3. Inserting a test user in the c-dimensional space Related work (Sarwar et al., 2000) has studied SVD on CF considering the test data as apriori known. It is evident that, for user-based approach, the test data should be con- sidered as unknown in the c-dimensional space. Thus a spe- cialized insertion process should be used. Given the current ratings of the test user u, we enter pseudo-user vector in the c-dimensional space using the following Eq. (9) (Furnas et al., 1988). In the current example,we insert U4 into the 2-dimensional space, as it is shown in Fig. 11: unew ¼ u � V m�c � S�1c�c ð9Þ In Eq. (9), unew denotes the mapped ratings of the test user u, whereas Vm·c and S �1 c�c are matrices derived from SVD. This unew vector should be added in the end of the Un ·c matrix which is shown in Fig. 10. Notice that the inserted vector values of test user U4 are very similar to these of U2 after the insertion. This is reasonable, because these two users have similar ratings as it is shown in Fig. 8. f A), Sn·m (singular values of A), V 0 m�m (right singular vectors of A). tors of A*), Sc·c (singular values of A*), V 0 c�m (right singular vectors of A *). e recommender systems: Combining effectiveness ..., Expert Sys- Fig. 11. Example of: unew (inserted new user vector), u (user vector), Vm·c (two left singular vectors of V), S�1c�c (two singular values of inverse S). P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx 11 ARTICLE IN PRESS Note that this process is omitted in the item-based approach, which means that this process is more applicable in real applications in terms of complexity. In our experi- ments, we will present this drawback of the user-based approach in terms of execution time. 5.4. Generating the neighborhood of users/items Having a reduced dimensional representation of the ori- ginal space, we form the neighborhoods of users/items in that space. For the user based approach, we find the k nearest neighbors of pseudo-user vector in the c-dimensional space. The similarities between training and test users can be based on cosine similarity. First, we compute the matrix Un·c Æ Sc·c and then we perform vector similarity. This n · c matrix is the c-dimensional representation for the n users. Note that in our experiments, we generate – as a sec- ond approach – the similarity matrix taking into account only Un·c matrix. As we will see, results are quite impres- sive, but have to do only with the user- based approach. For the item based approach, we find the k nearest neighbors of item vector in the c-dimensional space. First, we compute the matrix Sc·c Æ Vc·m and then we perform vector similarity. This c · m matrix is the c-dimensional representation for the m items. 5.5. Generating the recommendation list As it is mentioned in Section 3.2, existing ranking crite- ria, such as MF, are used for the generation of the top-N list in classic CF algorithms. We propose a ranking crite- rion that uses the predicted values of a user for each item. Predicted values are computed by Eqs. (4) and (6), for the cases of user-based and item-based CF, respectively. These equations have been used in related work for the purpose of MAE calculation, whereas we use them for generation of top-N list. Therefore, we sort (in descending order) the items according to predicted rating value, and recommend the first N of them.6 This ranking criterion, denoted as highest predicted rated item recommendation (HPR), is influenced by the good accuracy of prediction that existing related work reports through the MAE. HPR opts for recom- mending the items that are more probable to receive a 6 If less than N items have positive ratings (i.e., not less than Ps), then less than N items remain in the list. Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 higher rating. As our experimental results will demonstrate, HPR presents poor performance for the classic CF algo- rithms, but dramatically spectacular results when it is used in combination with SVD. The reason is that in the latter it is based only on the major trends of users. 5.6. Evaluation of the CF process Related work (Sarwar et al., 2000) proposed Eq. (10) for the generation of predictions, where � means the addition of a scalar with a matrix. pu;i ¼ �ru � U n�c � ffiffiffiffiffiffiffiffiffi Sc�c p ffiffiffiffiffiffiffiffiffi Sc�c p � V 0c�m ð10Þ We test this equation and find out that the predicted val- ues were calculated not from other users but from the user himself. So, the information of an inserted in c-dimensional space test user was used to predict his own real rates. These predicted values were so close to the real ones. Therefore, although the produced MAE may be good, this procedure is not prediction, because the test user is considered apriori known. In contrast, we use Eqs. (4) and (6) for prediction, to exploit information from other users or items. Thus, we use these predicted values for the calculation of MAE. 6. Performance study In the sequel, we study the performance of the described extensions against existing CF algorithms, by means of a thorough experimental evaluation. Both user-based and item-based algorithms are tested. Moreover, we study the performance of the described SVD dimensionality reduc- tion techniques against existing CF algorithms. Several factors are considered, like the sparsity, the similarity mea- sures, and criteria for generating the top-N list. The additional factors, that are treated as parameters, are the following: the neighborhood size (k, default value 10), the size of the recommendation list (N, default value 20), the size of training set (default value 75%), and the division between past/future data. Regarding WS, the c value was set to 5. Evaluation is performed with the preci- sion and recall metrics (given as percentages). We also use F1 metric. We perform experiments with four real data sets that have been used as benchmarks in prior work. In particular, we examined two MovieLens data sets: (i) the first one with 100,000 ratings assigned by 943 users on 1682 movies (This is the default data set. Any use of other data sets is categor- ically defined), and (ii) the second one with about 1 million ratings for 3592 movies by 6040 users. The range of ratings is between 1(bad)-5(excellent) of the numerical scale. More- over, we ran our experiments on the EachMovie data set (McJones & DeTreville, 1997), which contains 2,811,983 ratings entered by 72,916 for 1628 different movies, and the Jester data set, which contains 4.1 million ratings of 100 jokes from 73,496 users. Finally, in all data sets, we normalized the rating scale in the range 1–5, whereas Ps e recommender systems: Combining effectiveness ..., Expert Sys- 12 P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS is set to 3 and the value of an unrated item is considered equal to zero. 6.1. Results for user-based CF algorithms First, we examine user-based CF algorithms and com- pare the existing Pearson similarity and WS measures against the proposed UNION method. We also include the baseline (BL) algorithm. The results for precision and recall vs. k are displayed in Fig. 12a and b, respectively. As shown, the existing Pearson measure, which is based on co-rated items, performs worst than BL. This result is surprising, as BL is very simple. WS improves Pearson, because the disadvantage of Pearson, due to co-rated items, is downsized by the weighting with the number of common items. UNION clearly outperforms all other measures. The reason is that the MovieLens data set is sparse and these measures better exploit the available information. UNION reaches an optimum performance for a specific k. In con- trast, in the examined range of k values, the performance of Pearson increases with increasing k. Outside the exam- ined k range (not displayed), it stabilizes and never exceeds BL. We have to notice here, that as we increase the number of k nearest neighbors Pearson measure is literally trans- formed to BL. We now examine the MAE metric. Results are illus- trated in Fig. 13a (BL is only for recommendation, not pre- diction, thus omitted). As expected, Pearson yields the lowest MAE values, whereas WS is second best. This fact supports our explanation that MAE is indicative only for Fig. 12. Performance of user-based CF Fig. 13. Performance of user-based CF vs. Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 the evaluation of prediction and not of recommendation, as these measures did not attain the best performance in terms of precision and recall. To consider the impact of density, we also examine the Jester data set. The results for the F1 metric are depicted in Fig. 13b. In this case, the relative differences are smaller than for the case of sparse data. The reason is that dense data have a sufficient amount of information, thus there is less need to exploit information in the way UNION does. Finally, we test the described criteria for the top-N list generation: MF, HPR, and HSR. We used the UNION similarity measure, because it presented the best perfor- mance in the previous measurements. The results for preci- sion and recall are given in Fig. 14a and b, respectively. In Fig. 14a it is shown that HPR is clearly outperformed by the other two criteria, a fact that furthermore illustrates the unsuitability of the MAE metric to characterize the task of recommendation. On the other hand, MF and HSR present similar precision. Fig. 14b presents the results on recall (to make comparison between MF and HSR more clear, we omit HPR). HSR is constantly better than MF, though slightly. 6.2. Results of LSI application on user-based CF algorithms Regarding the performance of SVD dimensionality reduction in user-based approach, we preserve, each time, a different fraction of principal components of SVD model. More specifically, we preserve 90%, 70%, 50%, 30% and vs. k: (a) precision and (b) recall. k: (a) MAE and (b) F1 for dense data. e recommender systems: Combining effectiveness ..., Expert Sys- Fig. 14. Comparison of criteria for the generation of top-N list for user-based CF vs. k: (a) precision and (b) recall. Fig. 15. Performance of SVD dimensionality reduction of user-based CF vs. k: (a) precision and (b) recall. P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx 13 ARTICLE IN PRESS 10% of the total information of initial user-item matrix. The results for precision and recall vs. k are displayed in Fig. 15a and b, respectively. As we can see, the performance of SVD50, SVD70, SVD90 are similar in terms of precision and recall. The rea- son is that SVD50 is adequate for producing a good approximation of the original matrix. Thus, we will con- tinue our study with two representative SVD models, SVD50 and SVD10. These choices are indicative of the behavior of the SVD model. We now move on to comparison of existing user-based CF algorithm that uses Pearson similarity against the two representative SVD reductions. The results for precision and recall vs. k are displayed in Fig. 16a and b, respectively. As shown, the existing Pearson measure, which is based on co-rated items, performs worst than SDV reductions. The two proposed reductions clearly outperform Pearson Fig. 16. Performance of user-based CF Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 measure. We have to notice that SVD50 perfroms equally with the UNION algorithm. The reason is that the Movie- Lens data set is sparse and relatively large (high n value). The SVD reductions reveal the most essential dimensions and filter out the outliers and misleading information. So, with less information, we have almost the same result. SVD50 performs a little better than SVD10, as it uses more information. In contrast, the latter SVD reduction uses only 11 dimensions instead of 157 of the former. This means that the performance of the latter is faster than the former one. Now, we compare the SVD50 and UNION algorithms with the variation we described in Section 5.4 denoted as SVDU50. As we can see, SVDU50 performs even better than the UNION algorithm. We propose it, as a second way for producing the users’ similarity matrix. The posi- tion of the points between Un·c Æ Sc·c matrix and only the vs. k: (a) precision and (b) recall. e recommender systems: Combining effectiveness ..., Expert Sys- Fig. 18. Comparison HPR criteria for the generation of top-N (a) precision and (b) recall list for user-based CF vs. k Fig. 17. Performance of user-based CF vs. k: (a) F1 for 100k Movielens data set and (b) F1 for 1M Movielens data set. 14 P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS Un·c matrix are the same except that the former is stretched or shrunk in proportion to the corresponding diagonal ele- ments of Sc·c matrix. To consider the impact of scalability, we also examine the 1M data set. The results for the F1 metric are depicted in Fig. 17b. In this case, the relative difference broadens as k increases, than the case of 100K data set. The reason is that 1M data set is more sparse (the percentage of non rated items exceeds 95%) than 100K data set (the corre- sponding percentage is 93%). Moreover, the rank (i.e., number of independent dimensions) of the 1M data set is 3010 instead of 943 for the 100K. This means, that there are many dependent dimensions in the former that can be compressed. Finally, we test the described criteria for the HSR top-N list generation algorithm. The results for precision and recall are given in Fig. 18. As shown, the combination of the SVD similarity measure with HPR as list generation algorithm, clearly outperforms the Pearson with HPR. This is due to the fact that in the former the remaining dimen- sions are the determinative ones and outliers users have been rejected. Note that in the SVD50 we preserve only 157 basic dimensions instead of 943 for the latter. 7 We have examined a variation of adjusted cosine that uses only the co- rated items in the denominator. As expected, it resulted to worse precision and recall, but to better MAE. 6.3. Results for item-based CF algorithms We perform similar measurements for the case of item-based CF. Thus, we first examine the precision Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 and recall for the existing adjusted cosine measure (con- siders co-rated items) against UNION for the item-based case (that consider not only co-rated items through the extended definitions of T set). The results are depicted in Fig. 19 and are analogous to those of the user-based case. UNION clearly outperforms adjusted cosine and WS. Again, it is surprising to find the item-based CF algorithm with the adjusted cosine measure looses out by BL. Next, we compare adjusted cosine, UNION and WS against MAE. The results are illustrated in Fig. 20a. Dif- ferently from the case of user-based CF, all measures have similar MAE, and for larger k values (where all of them reach the optimum MAE value) UNION and the adjusted cosine are equal. The reason that adjusted cosine does not present better MAE, is that in its denominator it considers all items and not just the co- rated ones (see Eq. (2)). This improves its performance for the task of recommendation and worsens the perfor- mance of prediction.7 Regarding the examination of the dense data set (Jester), the results for the F1 metric are illustrated in Fig. 20b. Since item-based CF has been designed to suit the needs of sparse data, we find out that for dense data all item-based algo- e recommender systems: Combining effectiveness ..., Expert Sys- Fig. 19. Performance of item-based CF vs. k: (a) precision and (b) recall. Fig. 20. Performance of item-based CF vs. k: (a) MAE and (b) F1 for dense data. Fig. 21. Comparison of criteria for the generation of top-N list for item-based CF vs. k: (a) precision and (b) recall. P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx 15 ARTICLE IN PRESS rithms are outperformed by BL. This is the case even for UNION, although it performs better than adjusted cosine. This result clarifies the need to examine CF algorithms for all the involved factors, in this case density, in order to draw more complete conclusions. Finally, we test the three described criteria, MF, HPR, and HSS, for the top-N list generation by item-based CF algorithms. The results for precision are illustrated in Fig. 21a. Similarly to the user-based case, HPR performs very poorly. Fig. 21b shows the resulting recall (HPR is omitted to make comparisons more clear). MF performs a little better than HSS. However, no clear winner can be identified, because for the other data sets HSS performed a little better. Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 6.4. Results of LSI application on item-based CF algorithms Regarding the performance of SVD dimensionality reduction in item-based approach, we first examine the pre- cision and recall for the existing adjusted cosine Measure (considers co-rated items) against SVD50 and SVD10 for the item-based case. The results are depicted in Fig. 22 and are analogous to those of the user-based case. SVD50 and SVD10 clearly outperform adjusted cosine. Notice that unlike the user-based case, the difference between SVD50 and SVD10 is greater. This mean that item based algorithm cannot preserve accuracy in a satisfactory way when we decrease the percentage of dimension’s information. e recommender systems: Combining effectiveness ..., Expert Sys- Fig. 22. Performance of item-based CF vs. k: (a) precision and (b) recall. Fig. 23. Comparison between item-based and user-based CF vs. k. 16 P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS 6.5. Comparative results in terms of effectiveness In this section, we compare user-based and item-based CF using the best options as they have resulted from the previous measurements. Therefore, for user-based and item-based we use the UNION similarity measure. With respect to the criterion for the generation of the top-N list, for user-based we use HSR, whereas for the item-based we use MF. The result for F1 metric are displayed in Fig. 23a. These results demonstrate that user-based CF, when using the best options, compares favorably against item- based CF, even when the latter uses the best options. The difference in F1 is larger than 0.3. The difference with Fig. 24. Comparison between item-based and user-based CF in terms of e information. Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 respect to precision is larger of 10%, whereas with respect to recall, it exceeds 5% (we refer to the optimum values resulting from the tuning of k). This conclusion contrasts the existing one, that item-based CF is preferable than user-based for sparse data. The reason is that user-based CF is more focused towards the preferences of the target user. In contrast, with item-based CF, the recommended items may have been found similar by transactions of users with much different preferences than the ones of the target user. Thus, they may not directly reflect the preferences of the latter. The differences that have been reported in prior work can only be explained by the fact that the algorithms did not use the best possible options. In Fig. 23b, we pres- xecution time: (a) time vs. k and (b) time vs. percentage of preserved e recommender systems: Combining effectiveness ..., Expert Sys- P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx 17 ARTICLE IN PRESS ent again -for reasons of clearness- the comparison between UNION algorithm with SVDU50. As we can see, SVDU50 performs favorably against UNION algorithm. The reason is that the SVD reduction reveals the most essential dimen- sions and filters out the outliers and misleading informa- tion. So, by applying LSI in user-based algorithm we attain the best possible result in terms of accuracy. 6.6. Comparative results in terms of efficiency Regarding the execution time, we measured the wall- clock time for the on-line parts of the user-based and item-based algorithms. The results regarding the time needed for a user recommendation vs. k are presented in Fig. 24a, whereas the results for varying percentage of pre- served information, are depicted in Fig. 24b (in the latter measurement we set k = 10). As already mentioned, item based CF needs less time to provide recommendations than user-based CF. This holds for both the aforementioned measurements. The reason is that a user-rate vector in user-based approach has to be inserted in the c-dimensional space. Moreover, we have to mention that the generation of top-N list for the user- based approach further burdens the CF process. The rea- son is that the algorithm finds, firstly, user neighbors in the neighborhood matrix and then counts presences of items in the user-item matrix. In contrast, with the item- based approach the whole work is completed in the item Fig. 25. Comparison vs. N: (a Fig. 26. (a) Comparison vs. training set s Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 neighborhood matrix. So, in terms of execution item based approach is superior over user based. 6.7. Examination of the additional factors In this section we examine the impact of the additional factors. In our measurements we consider the existing two cases, that is, user-based CF with Pearson similarity and item-based CF with adjusted cosine (both the two existing cases consider co-rated items). First, we examine the impact of N (recommendation list’s size). The results are depicted in Fig. 25. As expected, with increasing N, recall increases and preci- sion decreases. The relative differences between the algo- rithms are coherent with those in our previous measurements. Next, we test the impact of the size of the training set, which is expressed as percentage of the total data set size. The results for the F1 metric are given in Fig. 26a (to make the graph more clear, we show results only for the two best user-based and item-based algorithms). As expected, when the training set is small, performance downgrades for both algorithms. Therefore, we should be careful enough when we evaluate CF algorithms and use adequately large train- ing sets. Similar to the previous measurements, in all cases user-based UNION is better than item-based UNION. The performance of both reaches a peak around 75%, after which it reduces. The reason for this is the overfitting that results from very large training sets. ) precision and (b) recall. ize and (b) comparison vs. past’s size. e recommender systems: Combining effectiveness ..., Expert Sys- Fig. 27. Measuring the percentage of items that are correctly predicted and top rated. 18 P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx ARTICLE IN PRESS We also examine the performance when considering the division between past and future data. For each user’s transaction in the test set we keep the 70% as future data and use a varying number of ratings from the rest 30% as past data. This way, we examine the impact of past’s size. We compare user-based UNION with item-based UNION using the F1 metric. The results are illustrated in Fig. 26b. As the size of the considered past reduces, the performance of both algorithms reduces. This result demonstrates the need to evaluate CF algorithms using the division between past and future data, because this division is more indica- tive about the actual performance in real-world CF appli- cations. Nevertheless, user-based UNION manages to keep a higher F1 value for small past size, whereas both algorithms converge to the same point for larger past. The merit of user-based UNION is evident, as in real- world applications the past size usually takes very small values (users do not have to wait long enough before start receiving recommendations). Finally, we test the examined algorithms with respect to the quality of the provided recommendations. The results are presented in Fig. 27. In particular, we are interested in determining whether these algorithms provide obscure recommendations or not. For this reason, we sort the items according to the number of positive ratings (i.e., higher than Ps) they received. Each time we keep a varying num- ber of these top rated items (e.g., the top 20, 40, etc.). These top rated items are considered as more popular, thus they do not comprise obscure recommendations. We measure the percentage of the top rated items that are correctly pre- dicted by each CF algorithm.8 This measurement indicates how much of the correct predictions belong to the most rated items. As the number of considered top rated items increases, the Pearson similarity measure (which is based on co-rated items) does not increase the percentage of the correctly pre- dicted items that belong to these top rated items. This means that the recommendations of this algorithm tend to be outside the range of the top rated items. Thus, Pear- 8 The alternative of just counting the predictions, regardless if they are correct, could provide misleading results and, thus, it is avoided. Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 son recommends more obscure items. The adjusted cosine similarity measure presents a small increase, which means that it presents the latter problem to a smaller degree. In contrast, both user-based UNION and item-based UNION clearly present the best percentage and their resulting slope is high. This means that these algorithms tend to avoid obscure items during their recommendations and concen- trate on the first positions of the top rated items. 7. Conclusions We have performed a thorough study of neighborhood- based CF, which brought out several factors that have not been examined carefully in the past. Based on our observa- tions, we proposed extensions of similarity measures, new criteria for generating the recommendation list, and new ways to evaluate the quality of the recommendation result. Additionally to these extensions, we proposed an LSI- based approach combining simultaneously effectiveness and efficiency of CF algorithms. We performed experimen- tal comparison of our proposed method against well known CF algorithms, like user-based or item-based meth- ods, with real data sets. Our experimentation reforms sev- eral existing beliefs and provides new insights. In particular, we highlight the following conclusions from our examination: • In contrast to what is reported in majority of related work, MAE is not indicative for the accuracy of the rec- ommendation process. It is, though, useful to character- ize the quality of the similarity measure (as reflected in the process of prediction). • Constraining similarity measures with co-rated items, weaknesses the measure. Though it is somewhat useful to consider the number of co-rated items (as WS does), the strict constraining inside the formulae for similarity measures is not suitable. • The proposed extensions that do not use co-rated items only, substantially improve the performance of CF, especially for sparse data, because they exploit more information. Actually, the existing approaches based on co-rated items are outperformed by a simple baseline algorithm we developed. • Our results showed that, following the best options for user-based and item-based CF, the former compares favorably to the latter. This contrasts with existing results in related work, because until now, comparison did not follow the best options we describe. • Our results have shown that item-based CF is not appropriate for dense data. • Our LSI-based significantly outperforms existing CF algorithms in terms of accuracy (measured through recall/precision). Moreover, in some cases, it can slightly improve UNION. The reason is that it is able to identify more clearly the correct recommended items by focusing on trends and isolating noisy users (e.g., outliers). In terms of execution times, due to the use e recommender systems: Combining effectiveness ..., Expert Sys- P. Symeonidis et al. / Expert Systems with Applications xxx (2007) xxx–xxx 19 ARTICLE IN PRESS of smaller matrices, execution times are substantially reduced. • In our experiments we have seen that only a 10% of the original matrix is adequate to provide accurate results. • The proposed HPR criterion in combination with SVD can compete the existing MF, which has been used by the majority of related work. • The process of folding-in is omitted in the item-based approach. Thus, item-based algorithms are faster than user-based ones. For this reason they may be more appropriate for real-world applications, despite their worse accuracy, since execution time is sine-qua-non in realistic deployments. Summarizing the aforementioned conclusions, we see that, on one hand, item-based algorithms are more appro- priate for off-line computations and, thus, better in on-line response. On the other hand, user-based algorithms can produce more effective recommendations with a small frac- tion of initial information and efficient responding time. With the proposed approaches we combine effectiveness and efficiency in contrast to the classic CF item and user- based algorithms for each approach individually. In our future work we will consider the issue of an approach that would unify the best characteristics of these two cases in an integrated approach. References Berry, M., Dumais, S., & O’Brien, G. (1994). Using linear algebra for intelligent information retrieval. SIAM Review, 37(4), 573–595. Breese, J., Heckerman, D. & Kadie, C. (1998). Empirical analysis of predictive algorithms for collaborative filtering. In Proceedings of the conference on uncertainty in artificial intelligence (pp. 43–52). Deshpande, M., & Karypis, G. (2004). Item-based top-n recommendation algorithms. ACM Transactions on Information Systems, 22(1), 143–177. Furnas, G., Deerwester, S. & Dumais, S., et al. (1988). Information retrieval using a singular value decomposition model of latent semantic structure. In Proceedings of the ACM SIGIR conference (pp. 465–480). Goldberg, D., Nichols, D., Brian, M., & Terry, D. (1992). Using collaborative filtering to weave an information tapestry. ACM Communications, 35(12), 61–70. Goldberg, K., Roeder, T., Gupta, T., & Perkins, C. (2001). Eigentaste: a constant time collaborative filtering algorithm. Information Retrieval, 4(2), 133–151. Please cite this article in press as: Symeonidis, P. et al., Collaborativ tems with Applications (2007), doi:10.1016/j.eswa.2007.05.013 Herlocker, J., Konstan, J., Borchers, A. & Riedl, J. (1999). An algorithmic framework for performing collaborative filtering. In: Proceedings of the ACM SIGIR conference (pp. 230–237). Herlocker, J., Konstan, J., & Riedl, J. (2002). An empirical analysis of design choices in neighborhood-based collaborative filtering algo- rithms. Information Retrieval, 5(4), 287–310. Herlocker, J., Konstan, J., Terveen, L., & Riedl, J. (2004). Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems, 22(1), 5–53. Hofmann, T. (2004). Latent semantic models for collaborative filtering. ACM Transactions on Information Systems, 22(1), 89–115. Huang, Z., Chen, H., & Zeng, D. (2004). Applying associative retrieval techniques to alleviate the sparsity problem in collaborative filtering. ACM Transactions on Information Systems, 22(1), 116–142. Karypis, G. (2001). Evaluation of item-based top-n recommendation algorithms. In Proceedings of the ACM CIKM conference (pp. 247–254). McJones, P. & DeTreville, J. (1997). Each to each programmer’s reference manual. Tech. Rep. Systems Research Center, 1997-023. McLauglin, R. & Herlocher, J. (2004). A collaborative filtering algorithm and evaluation metric that accurately model the user experience. In Proceedings of the ACM SIGIR conference (pp. 329–336). Mobasher, B., Dai, H., Luo, T. & Nakagawa, M. (2001). Improving the effectiveness of collaborative filtering on anonymous web usage data. In Proceedings of the workshop intelligent techniques for web person- alization (pp. 53–60). O’Mahony, M., Hurley, N., Kushmerick, N., & Silvestre, G. (2004). Collaborative recommendation: a robustness analysis. ACM Transac- tions on Internet Technology, 4(4), 344–377. Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P. & Riedl, J. (1994). Grouplens: an open architecture for collaborative filtering on netnews. In Proceedings of the conference computer supported collaborative work (pp. 175–186). Sarwar, B., Karypis, G., Konstan, J. & Riedl, J. (2000). Analysis of recommendation algorithms for e-commerce. In Proceeedings of the ACM electronic commerce conference (pp. 158–167). Sarwar, B., Karypis, G., Konstan, J. & Riedl J. (2000). Application of dimensionality reduction in recommender system – a case study. In ACM WebKDD workshop. Sarwar, B., Karypis, G., Konstan, J. & Riedl, J. (2001). Item-based collaborative filtering recommendation algorithms. In Proceedings of the WWW conf. (pp. 285–295). Sarwar, B., Karypis, G., Konstan, J. & Riedl, J. (2002). Incremental singular value decomposition algorithms for higly scalable recommender sys- tems. In Proceedings of the computer and information technology. Xue, G., Lin, C. & Yang, Q., et al. (2005). Scalable collaborative filtering using cluster-based smoothing. In: Proceedings of the ACM SIGIR conference (pp. 114–121). e recommender systems: Combining effectiveness ..., Expert Sys- Collaborative recommender systems: Combining effectiveness and efficiency Introduction Motivation Contributions Related work Factors affecting the CF process First stage factors Second stage factors Third stage factors Proposed extensions for memory-based algorithms First stage extension : the UNION similarity measures Assigning a value to unrated items Second stage extensions: the HPR and HSR algorithms Third stage extension : a baseline algorithm Proposed method for an LSI-based algorithm Applying SVD on training data Preserving the principal components Inserting a test user in the c-dimensional space Generating the neighborhood of users/items Generating the recommendation list Evaluation of the CF process Performance study Results for user-based CF algorithms Results of LSI application on user-based CFalgorithms Results for item-based CF algorithms Results of LSI application on item-based CF algorithms Comparative results in terms of effectiveness Comparative results in terms of efficiency Examination of the additional factors Conclusions References 
work_wadj5fvobbbelpdhjpdsjvhz4q	----	www.rspsciencehub.com Volume 02 Issue 09S September 2020 International Research Journal on Advanced Science Hub (IRJASH) 48 Special Issue of First International Conference on Innovations in Engineering Sciences (ICIES 2020) Analysis of Student Academic Performance through Expert systems Kandula Neha 1 , Dr. S Jahangeer Sidiq 2 1 Assistant Professor, Artificial Intelligence, Lovely Professional University (Punjab) 2 Assistant Professor, Computer Science and Engineering, Lovely Professional University (Punjab) 1 neha09kandula@gmail.com, 2 jahangir.25453@lpu.co.in Abstract Predicting the performance of students is one of the most important topics required for learning contexts such as colleges and universities, as it helps to design successful mechanisms that boost tutorial outcomes and prevent dropouts among various items. These are benefited by automating the many processes involved in the activities of usual students which handle huge volumes of information collected from package tools for technology-enhanced learning. Thus, the careful analysis and interpretation of these information would provide us with valuable data regarding the data of the students and therefore the relationship between them and hence the tutorial tasks. This data is the supply which feeds promising algorithms and methods able to estimate the success of the students. During this analysis, virtually many papers were analysed to show radically different trendy techniques widely applied to predict the success of students, along with the goals they need to achieve in this area. These computing-related techniques and approaches are mainly machine learning techniques, deep learning techniques, Artificial Neural Networks & Neural Networks Convolution, etc. This paper demonstrates the analysis and their comparisons of various methods used to forecast Student Academic success. Keywords: Deep learning, Machine Learning, Prediction, Academic Performance, Analysis 1. Introduction Educational facilities area unit running in Associate in Nursing more and more competitive and complicated environment and facing a burden in dealing with world and national economic, these facilities try to establish modifications such as the rising one to improve the magnitude relationship of scholars in specific disciplines and to ensure that the quality of learning programmes area units each Area of students unit the first stakeholders within the academic establishments and their success play a notable role in the social and financial growth of an extremely nation by delivering innovative alumni. There's a fundamental interest in pedantic institutions staying up and providing massive student datasets for useful simple leadership. Furthermore, the use of net innovation has become an integral part of this educational era in many universities, improving the essential exchange of information about students, teachers and their contact with learning and educational frameworks. Advanced education plays an significant part in a society's progress and growth. It is an environment that offers lots of membership information, such as students, teachers, offices and training programmes. Scholars' success can be a major concern for different stakeholders as well as researchers, superiors and organisations. Therefore students should strive flat out for glorious marks, so they can rise to educational partners' standards. Student's success can be an basic concern of academic institutions, for the glorious educational accomplishment of a particular university will cause that university's level to increase. The mailto:neha09kandula@gmail.com mailto:jahangir.25453@lpu.co.in www.rspsciencehub.com Volume 02 Issue 09S September 2020 International Research Journal on Advanced Science Hub (IRJASH) 49 success of the student may not be inheritable by operation of the co-curriculum and measurement of earnings. Most of the researchers, however, stated that graduation is the student's live performance. Data science stands out as a discipline to replace. It is seen as an associate in the incorporation of traditional disciplines such as data processing, analytics, distributed systems and databases into nursing. It was important to merge current methods to view abundantly usable information in useful data for society, individuals and organisations. Information science tools are widely used to identify procedures, evaluate bottlenecks, verify enforcement and even suggest changes. Data science depends on the knowledge from account data, observations, and prophetic models. This attempt involves curetting and cleaning up at one end, and disseminating information at the other. After the successful creation and copy of systems and software packages the unit area capable of efficiently store, retrieve and efficiently store information on methods, focus has now shifted to predictive and correlative analytics . The fields of machine learning and information science have been robustly illustrated in Hell, this demand for a new approach, one that has hitherto seen some example in the syllabus of education and in the literature of scientific discipline. To employ information science in education, students would obtain data on their success in terms of their colleagues' importance or development. On the contrary, directors and decision-makers will take advantage of the information science to try to improve the environment of tutoring. There is typically a positive thing that should be able to anticipate the actions of potential students in order to improve instructional style and set up programmes on the programme provided to the scholars for educational support and steering. This can be anywhere data processing is involved (DM comes in). DM techniques analyze datasets and extract info to rework it into graspable structures for later use. Machine Learning (ML), cooperative Filtering (CF), Recommender Systems (RS) and Artificial Neural Networks (ANN), Convolution Neural Networks (CNN) are the most machine techniques that method this info to predict students’ performance, their grades. Nowadays, there is a significant amount of analyses and studies that follow on the lines of predicting the actions of the students, among various related topics of interest within the educational space. In reality, several papers are printed in journals and conference on this subject. Hence the main objective of this analysis is to provide an exhaustive overview of the various theoretical techniques and algorithms applied to the present Fig 1 : Framework Methodologies Used : There are few Methodology techniques used for the prediction of students performance and few of them are listed below. Education Data Mining: Data mining, additionally popularly referred to as data Discovery in information, refers to extracting or “mining" data from massive amounts of information. data processing techniques area unit accustomed care for massive volumes of information to get hidden patterns and relationships useful in deciding. whereas data processing and data discovery in information area Demographic Data Academic Data Behavioural data Expert Systems Input Process Output Prediction www.rspsciencehub.com Volume 02 Issue 09S September 2020 International Research Journal on Advanced Science Hub (IRJASH) 50 unit of times treated as synonyms, data {processing} is truly a part of the data discovery process. Various techniques and few kind of algorithms like Classification, , Regression, Neural Networks, Rules of association, Trees, Genetic, Clustering , Techniques in Nearest Neighbour etc., area unit used for data discovery from databases. These techniques and have different ways for processing of data and temporary mention in understanding in own ways. A. Classification : Classification is that it uses the most efficiently applied data processing method, a series of pre-classified examples for the creation of a model that will classify the database population to huge. This time approach which is called tree or classification algorithms based on neural networks. The method of classification of knowledge applies to learning and classification. In Learning the unit of coaching information field defined by the law of classification. Consider the unit of information field used to estimate the consistency of classification rules in classification. If the precision is sufficient the rules would apply to the latest Tuples of information. The classifier-training rule uses these pre-classified examples to see the set of parameters needed for correct discrimination. The rule these parameters uses and then encodes into a model known as a classifier. Classification approach can be used for successful suggests that nevertheless it becomes costly to classify teams or groups of objects so bunch would be used as a pre-processing method for collection and classification of attributes set. [19] C. Regression technique For postulation, will be customised. Multivariate analysis will be used to model the relation between one or more variables depending on freelance and a few variables. We already appear to toll-known attributes in the processing of freelance variables, and response variables square measure what we would like to expect. Unfortunately, many real-world issues do not seem to be just a guess. Hence, several complicated techniques ( e.g., supply regression, call trees, or neural nets) may also be required to forecast future values. Typically similar varieties of the model can be used for each regression and classification. . for instance, the CART (Classification and Regression Trees) decision tree rule are accustomed build every classification tree (to classify categorical response variables) and regression trees (to forecast continuous response variables). Neural networks may also turn out every classification and regression models. [19] D. Association rule: Association and correlation is typically to look out frequent item set findings among huge data sets. this kind of finding helps businesses to create sure selections, like catalog vogue, cross-selling, and shopper wanting behaviour analysis. Association Rule algorithms have to be compelled to be able to generate rules with confidence values however one. however the quantity of gettable Association Rules for a given dataset is generally really giant and a high proportion of the principles unit usually of little or no (if any) worth. [19] E. Neural networks : Neural network may be a collection of connected input/output units and each affiliation options a weight gift with it. The network learns in the coaching half by changing weights so that the correct class Tuple labels can be predicted. Neural networks have the exceptional capacity to derive that indicates difficult or inaccurate data, and may well be familiar with extract patterns and spot trends that unit of measurement is too complex to be identified by either humans or different computer techniques. This measurement unit is like minded for inputs and outputs that are continuously measured. Neural networks unit of measurement best at distinctive patterns or data trends and as intended for analysis or declaration desires [19] Machine learning: Machine learning techniques such as fully connected feed forward Artificial Neural Network, Naïve Bayes, Decision Tree, have been used to build the machine learning model. Artificial Neural Networks Artificial Neural Network represents a group of input unites and output unites. They are linked by weighted connexions to every other. The ANN learns by excessively adjusting the weights of the links, so it can predict the right target mark for a few instances of the input file. Backpropagation algorithmic method is one of the noted learning algorithms used to train the ANN. ANN has some advantages, such as its high resistance to yelling data sets and its strong success in classifying patterns which have not been trained so it is used in things until there is some knowledge about the relationship between the category mark and the www.rspsciencehub.com Volume 02 Issue 09S September 2020 International Research Journal on Advanced Science Hub (IRJASH) 51 options inside the dataset. The ANNs have many worldwide applications such as image and written recognition , voice recognition, laboratory medicine, and pathology. There are several kinds of ANNs that would help their design and style to be listed. One example may be a fully connected multilayer feed forward ANN during which the network has an input layer Associate in Nursing, one or more hidden layers, and also the output layer. [17] In this analysis a three-layer fully linked feed forward ANN was used. The network is made up of input layer Associate in Nursing, 2 hidden layers, and also the output layer. The main layer consists of twenty input units, neurons, while the primary hidden layer includes six hidden units. The second layer concealed has 3 hidden units. The fourth layer is the output layer which only has 1 unit capacity. The linear measure Rectifier was used so the activation of the hidden units is done. Naive Bayes : Naive Bayes classification model is called because the theorem network's simplest variant is. This model assumes that given the target attribute status each function attribute is freelance from the opposite attributes. . instance x within the dataset contains values for the attributes 𝑎1,𝑎2,…,𝑎𝑖. The target perform f(x) equals any price from predefined finite set V=(𝑣1,𝑣2,…,𝑣𝑗). Naïve mathematician model uses the subsequent equation.[17] Decision Tree : A model for a decision tree represents a tree structure which is just like a flow diagram. -- internal node during this structure represents a check on the attribute of the dataset, while each limb represents the check result. Moreover, each leaf node represents a target function code, and the higher initial node within the tree also represents the base node. Decision trees can be either binary or non-binary. Call trees are common strategies for classifying victims that they do not want prior knowledge on the subject domain or sophisticated classification criteria structure. Moreover, they can actually be resurrected to the laws of classification and that they can be interpreted clearly. This / tree classification technique has been used in many real word applications such as monetary research , medicine, biology, development and uranology. Data Mining Techniques: As several machine learning algorithms square measure won’t to train the scholar prophetical models therefore after analyzing several of them for comparison functions we have a tendency to have elite supplying regression and support vector machine using sequent stripped improvement 2 wide famous classifiers for prediction. a) Logistics Regression: Regression of logistics is widely used anywhere categorical goal variable is used and the most famous constant quantity approach is taken into account. Generally, providing square measure linear models of regression models used for predicting some categorical dependent worth primarily based on some freelance predictor meaning. It can also solve the problems of binary (two classes) classification as multinomial (two or additional classes) values. Logistic regression model operates on logic perform design to find the linear combination or prevalence possibilities of any targeted group worth supporting predictor values (continuous or discrete). b) Support Vector Machine(SVM): Additionally named SVM support vector machine is one of the most common machine learning algorithms used for classification , prediction, and regression. Since these algorithms work best with analysing large datasets with many predictor attributes, SVM algorithms have a nice use in text recognition or categorization, and bioinformatics. The SVM model is regarded as a discriminative maximum- margin model and operates on the principle of classifying information point into 2 or additional categories by locating the associate degree optimistic decision boundary which is assumed to be N-dimensional hyper plane and should be at a massive distance from the datum of each category. Deep Learning : Deep learning classifiers are given to the data set collected from educational environments. Data is pre-processed and check for missing values. Classifiers are applied on the data set to build the models. Models are tested with test data to predict the students’ performance and the best models yielding high accuracy are considered. The Methodologies Proposed in this research work has the important phases such as Artificial Neural Network, Supervised and Unsupervised Learning, Deep Neural Networks and Convolution Neural Networks. Artificial Neural Network: Artificial neural network approaches used to predict academic www.rspsciencehub.com Volume 02 Issue 09S September 2020 International Research Journal on Advanced Science Hub (IRJASH) 52 success. This is also intended to identify aspirants to enter university profession at multiple levels in line with the probability of reaching a success standard. This study identifies shortcomings for students during the academic career from various viewpoints in courses, and offers ways to prevent them. Many educational writers concentrated on learning varieties and their recommendations. They depend on the fact that students have various temperament types that appear to have different learning varieties that influence students ' performance in each subject. Through this study we can build an artificial neural network to predict academic success that has supported the type of student training. Neural networks are efficient and widespread implementation in enormous variety of information mining applications that exceed classifiers according to the study. This main objective of this analysis is to examine whether square neural networks measure a correct classifier to predict LMS (learning management system) scholars performing in academic data processing context. Table.1 Comparative Research work of Data Mining and outcomes : Sl No Title Techniques used Outcome 1 Edifice an Educational Framework using Educational Data Mining and Visual Analytics[11] Classification and Association Rule mining Analysis of the students‟ performance. 2 Analysis of Girls Vocational High School Students Academic Failure Causes with Data Mining Techniques[12] Clustering Technique Analysis of failure causes of high school students 3 Educational Data Mining Classifier For Semester One Performance to Improve Engineering Students Achievement [13] Classification Technique and Neuro-Fuzzy classifier Assessment of students at early in the beginning of the course 4 Students‟ Employability Prediction Model through Data Mining [14] Classification Bayesian methods, Multilayer Perceptions and Sequential Minimal Optimization (SMO), Ensemble Methods and Decision Trees Predict the employability of the PG students 5 Mining Educational Data to Analyze Students' Performance [15] Classification decision tree method Predict the enrolment of students in a particular course for higher education system in the university and to identify the dropouts. 6 Students Academic Failure Prediction using Data Mining [16] Classification and Feature selection Algorithm Designed a model to predict student failure at the end of the semester. Convolution Neural Network : CNN is commonly used for prediction functions, as CNN provides several settings that need to be outlined and manipulated to achieve maximum precision in performance. These variables have a strong effect on the CNN model 's efficiency (i.e., efficiency of activation, variety of layers, variety of vegetative cells in each layer, and hyper-parameter). If the data set is obtained, the model will be coached with a series of tests and various models will be used to test the predicted model. Authors follow the variety of epoch and number of layers as key variables for manipulating until maximum precision is achieved. As a result of the www.rspsciencehub.com Volume 02 Issue 09S September 2020 International Research Journal on Advanced Science Hub (IRJASH) 53 experiments, the accuracy measures that have been obtained will be contrasted with few works in the literature that have discussed the success prediction of the behaviour of the students. End Results of Few Methodologies and their accuracy Table2 - Methods Accuracy Method Accuracy Multi Class Classifier 99.81 SVM 93.9 Naïve Bayes 97.45 Random Forest 99.3 Decision Tree 98.9 CNN 100 Conclusion: This survey paper gives the Prediction of students performance and how it has become an huge trend these days in research work. In order to predict their performances many algorithms and techniques have been designed and have been used to get the accurate values. These predictions are been done based on many parameters of the individual students performances like behaviour of students in their home and classroom , their extra circular activities, classroom activities, behaviours in classroom, participation in quizzes and group discussions etc. All these parameters makes an huge impact on their academic performances. A Survey has been done on many Deep Learning based algorithms such as Artificial Neural Networks, Machine Learning, Support Vector Machines , Convolution Neural Network etc to draw conclusions based on their accuracy and to select the best deep learning algorithm to predict the students performance. Reference : [1]. Sri Lakshmi, A V Krishna Prasad "Research Methodologies for Student Performance Evaluation Using Educational Analytics Tools and Approaches", International Journal of Recent Technology and Engineering (IJRTE) ISSN: 2277- 3878, Volume-8, Issue-1S4, June 2019 [2]. J. Han, M. Kamber, and J. Pei, Data Mining: Concepts and Techniques, 3rd ed. Morgan Kaufmann publications, 2012. [3]. Daud, Ali, et al., "Predicting student performance using advanced learning analytics," Proceedings of the 26th international conference on world wide web companion. International World Wide Web Conferences Steering Committee, 2017. [4]. Mansur, Andi Besse Firdausiah, Norazah Yusof, and Ahmad Hoirul Basori, "Personalized Learning Model based on Deep Learning Algorithm for Student Behaviour Analytic," Procedia Computer Science vol. 163, pp. 125-133, 2019. [5]. Patil, Akhilesh P., Karthik Ganesan, and Anita Kanavalli, "Effective deep learning model to predict student grade point averages," IEEE International Conference on Computational Intelligence and Computing Research (ICCIC), pp. 1-6, 2017. [6]. Waheed, Hajra, et al., "Predicting academic performance of students from VLE big data using deep learning models," Computers in Human Behavior, vol. 104, pp. 106189, 2020. [7]. Students' Academic Performance Dataset, [Online], available https://www.kaggle.com/aljarah/xAPI-Edu- Data, Access on August 2019. [8]. Predicting Students' Academic Performance Through Supervised Machine Learning, 978-1-7281-6899-9/20//$31.00 ©2020 IEEE [9]. R.Arora, "Comparative analysis of classification algorithms on different datasets using WEKA," International Journal of Computer Applications, vol. 54(13),2012. [10]. Arsad, PM, Buniyamin, N and Manan, J- lA. 2013. A neural network students’ performance prediction model (NNSPPM). 2013 IEEE International Conference on Smart Instrumentation, Measurement and Applications (ICSIMA), (November): 1–5. [11].Dr. S. Anupam Kumar (2016) “Edifice an Educational Framework using Educational Data Mining and Visual Analytics” I.J. Education and Management Engineering, 2, 24-30 . [12].Deperlioglu, Omer & Sibel Birtil, Fatma. (2016). Analysis of Girls Vocational High www.rspsciencehub.com Volume 02 Issue 09S September 2020 International Research Journal on Advanced Science Hub (IRJASH) 54 School Students' Academic Failure Causes with Data Mining Techniques. The Anthropologist. [13]. Usamah Bin Mat, Norlida Buniyamin and Pauziah Mohd Arshad (2016) Engineering & Technical Education Research Group (Enter), Faculty of Electrical Engineering, Universiti Teknologi MARA, Middle-East Journal of Scientific Research 24 (2): 338- 346, 2016. [14]. Mishra, Tripti & Kumar, Dharminder & Gupta, Sangeeta. (2016). Students‟ employability prediction model through data mining. 11. 2275-2282. [15].Brijesh Kumar Baradwaj, Saurabh Pal (2011) International Journal of Advanced Computer Science and Applications, (IJACSA) Vol. 2, No. 6 [16].Jantawan, Bangsuk & Tsai, Cheng-Fa. (2013). The Application of Data Mining to Build Classification Model for Predicting Graduate Employment. [17].Hussein Altabrawee ,Osama Abdul Jaleel Ali , Samir Qaisar Ajmi , " Predicting Students’ Performance Using Machine Learning Techniques", Journal of University of Babylon, Pure and Applied Sciences, Vol.(27), No.(1): 2019 [18].Student Academic Performance Prediction using Supervised Learning Techniques https://doi.org/10.3991/ijet.v14i14.10310 [19]. Mining Educational Data to Analyze Students‟ Performance, (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 2, No. 6, 2011. [20].A Comparative Analysis of Techniques for Predicting Student Performance, Proceedings of the 9th International Conference on Educational Data Mining 
work_wcurz4s52va3rcxqzojkxfwgx4	----	c371.dvi Journal of Computing and Information Technology - CIT 11, 2004, 4, 285–291 285 Developing a Spell Checker as an Expert System Šandor Dembitz�, Petar Knežević� and Mladen Sokele�� �Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia ��Croatian Telecom, Zagreb, Croatia In this paper we discuss some basic concepts of know- ledge engineering, expressing our feeling that there is something missing. In our opinion, the missing link, able to clarify all the basic concepts, should be the concept of energy, not in use in knowledge engineering practice and theory at all. Since energy is an extremely abstract concept, which led to misunderstandings even in physics, we humbly introduce it by describing an existing expert system, called Hascheck. Hascheck is a spell checker for Croatian language, implemented as learning �semi�automaton, accessible via E-mail and Web, and being in public use nearly 10 years. We also present here a mathematical model of Hascheck’s learning. The model led to the so-called cognoelectrical analogy, which allowed some energy considerations about learning, and knowledge. Keywords: spell checker, knowledge acquisition, mo- delling of learning, energy. 1. Introduction Knowledge management is currently attracting a great deal of interest in scientific and busi- ness communities. However, is knowledge en- gineering a scientifically well-founded disci- pline? There are some reasons to doubt this. “Expert systems have to be an applied disci- pline, not a theoretical one. In application, real- word needs such as knowledge-based mainte- nance and learning need to be accommodated. We need to see expert systems as primarily a discipline to do with supporting knowledge ap- plications; primarily an applied discipline, pri- marily concerned with the logic of the know- ledge and only secondarily a technological dis- cipline. This needs a modification of viewpoint on the part of many expert systems practition- ers.” �Taylor, 1999; p. 8� The quoted thoughts of R.M. Taylor bring know- ledge engineers in the position of ancient Ro- man civil engineers constructing bridges. It is not such a bad position. Not knowing Newton axioms, they constructed many bridges, some of which are still in operation, from Spain to the Near East, from England to Africa. But, for modern engineers, this position is very weird; they are taught to think axiomatically. It is not so easy when knowledge is concerned. On the other hand, there are some attempts to formalise fundamentals of knowledge engineer- ing. How it works will be illustrated by quoting only one of these attempts, �Uschold, 1998�. In his introduction M. Uschold says: “Very impor- tantly, this is a descriptive exercise, not a nor- mative one” �p. 5�. A few pages later he says: “In particular, readers are expected to know how the term KNOWLEDGE is being used, and what INFERENCE is. They should be familiar with fundamental notations of LANGUAGE in gen- eral and KNOWLEDGE REPRESENTATION LANGUAGES in particular, as well as what is meant by a KNOWLEDGE BASE” �p. 8�. Do we all mean the same when using terms like knowledge, reasoning or language? Surely not! Speaking of language, the authors of these lines primarily think of their own, Croatian language. Is there any knowledge and reasoning out of language? Of course there is. Artists produc- ing paintings or music have such knowledge and such reasoning. But, is there any sharable, common knowledge and reasoning out of natu- ral languages? We doubt it. 286 Developing a Spell Checker as an Expert System Regarding Uschold’s and others’ papers about theoretical foundations of knowledge engineer- ing we got a strong feeling that there is some- thing missing, something that could connect, physically, all basic, weakly defined concepts. Being engineers, we believe that we have found the missing link in the concept of energy. “Knowledge is power”, people say. Every- one will agree that learning is difficult. Thus, in everyday life we describe knowledge and knowledge acquisition, or learning, with energy terms. Nowadays many researchers and engi- neers are trying to build up learning automata, or knowledge-based systems. Do they consider their attempts through energy analysis? Usually not, because of the lack of scientific concepts which could support such considerations. Even in physics, energy is an extremely abstract concept, which led to many misunderstandings: only perpetuum mobile is to be mentioned as an example. Therefore, our paper has to be re- garded as a humble – but practical – attempt, pointing where to research further in order to enable energy considerations about knowledge and learning. At the very beginning we had a simple – for many years now commonly regarded as unattrac- tive – knowledge engineering task: to collect word types and build up an acceptable spell checker for Croatian language. Our approach was unconventional: we decided to implement our checker as a telematic service embedded in E-mail, so the users had to send their texts to a given address and to wait for an automatic re- ply in the form of the checker’s report. Texts sent for checking demonstrated to be a valuable source of new knowledge for the checker. This provoked our interest for learning in technical systems. The simplicity of our original task had two faces: positive and negative. Firstly the neg- ative: we dealt with common knowledge, cor- rectly and incorrectly written Croatian words. Each naı̈ve user, and most Hascheck users are “naı̈ve” in the sense of the spell checker – expert system – development problem, having some competence reparding Croatian language, could judge whether our product functioned well or not. Fortunately, they accepted Hascheck. The positive side of our job was that we did not have to cope much with structures, relations and sim- ilar aspects complicating most knowledge engi- neering tasks. These circumstances allowed us to go fairly deep in modelling learning. The results will be presented in this paper. Section 2 describes the checker. Section 3 pro- vides a mathematical model of the learning pro- cess based on data collected during the first 5 years of the checker’s life; data collected later brought nothing new to the model. Finally, Section 4 introduces a cognoelectrical analogy that allows some energy considerations about knowledge and learning. 2. Hascheck – A Learning (Semi)automaton “Spelling is one of the best examples I’ve seen of the need for prototyping: build something small, try it, see how useful it is in practice, then modify and extend” �Bentley, 1985; p. 460�. This idea led to Croatian Academic Spelling Checker, called Hascheck �an acronym derived from its full Croatian name: Hrvatski akademski spelling checker�. Hascheck is the first Croatian spell checker; some other Croatian checkers were developed later �Sokele, 1997�. Hascheck functions as a telematic service em- bedded in E-mail �the address: hacheck�fer� hr�. In the last few years it is also accessible via World Wide Web on the address http���www� hr�hacheck�. The service has been in operation since March 21, 1994. Hascheck’s initial dictionary counted less than 100,000 common word types. Here we must explain the term word and how we use it. A lemmatised word is what one finds on the left side of a conventional dictionary. By apply- ing its morphology, a natural language produces word forms for each lemmatised word. Not all produced word forms, regarded as alphabetic strings, are necessarily distinct. In English, for example, the noun work and the verb work are two distinct word forms, because they are two distinct lemmatised words, but one word type. Further, it is necessary to distinguish common words �common nouns, verbs, adjectives etc.� from names �proper nouns, acronyms, foreign words in original orthography etc.�, because of their different behaviour in texts. Finally, words in text are tokens. But, not all tokens are words. Some tokens are nonwords �misspellings or ty- Developing a Spell Checker as an Expert System 287 pos�, and the spell checker’s job is to find �de- tect� them. With a small initial dictionary we encountered several problems. Croatian is a Slavic lan- guage, like Russian, Polish, Bulgarian and oth- ers. They all belong to the group of highly inflected languages with a great variety of dif- ferent word forms – and word types – for each lemmatised word. A Russian estimation sets the lowest limit of dictionary size for an acceptable spell checker at 1,000,000 word types; the up- per limit is 100,000,000 word types �Dolgopov, 1986�. Having a dictionary 10 times smaller than the estimated lowest limit, we had to find, in order to make the Hascheck report useful for the users, the way to divide unrecognised tokens from a text into legal word types and nonwords, respectively. An idea describing how to treat this problem was found in a paper dating from the very beginning of practical spell checking �Morris&Cherry, 1975�. The AT&T Bell Labs’ spell checker typo – not in use anywhere for many years now – used the concept of string peculiarity. This concept – despite its original technical implementation – meant that strings like approachment and ap- prchment have to be treated differently. For any person having some knowledge of English the first string resembles an English word much more than the second one. Is it possible to build up a program able to make a similar distinction? An acceptable solution was found by apply- ing binary n-grams �n�3,4,5,6� derived from a common word dictionary. It is a classical AI technique used mostly for error corrections �Riseman&Hanson, 1974; Ullmann, 1977�. Fi- nally, a very human-like fuzzy evaluation of strings not contained in the dictionary appeared, classifying them in several classes of pecu- liarity. The solution proved to be language- dependant. We experimented both in Croat- ian and English �Dembitz, 1993�, but we never managed to put more than 50% of unknown En- glish word types in the class of the lowest pe- culiarity; the best result in Croatian was around 80%. By all these attempts the rate of typos and misspellings in the same class varied between 10–15% for both languages. Using our classifying algorithm we were able to bring majority of nonword types at the begin- ning of Hascheck report; legal Croatian word types not included in the dictionary, were pushed to the end of the report. Such selectivity, taking into account a small dictionary volume, i.e. a poor coverage of incoming texts at the begin- ning of service life, was of great importance for accepting Hascheck as a useful service. It is the variety of word endings that makes a language highly inflected. Text sent for check- ing became a valuable source of new word types unknown to Hascheck. Inspired by �Weischedel et al., 1993� we developed a tagging algorithm for spell checking purposes. The tagging al- gorithm is applied to collections of Hascheck reports recorded for learning purposes. The tagging for common words is applied on less peculiar strings. Strings beginning with cap- itals, or consisting of capital letters only are assumed to be potential names and are treated separately. The tagging for names is applied to all classes. Our word-guessing method succeeds in tagging correctly more than 70% of unknown Croatian word types. These data refer to common words �common nouns, verbs etc.�, while efficiency of the algorithm with name types �proper names, acronyms etc.� is around 50%. In both cases the noise – percentage of typos and misspellings in a set of tagged strings is under 10%. Final result of our research and development was a spell checker functioning as a learning �semi�automaton. The classifier classifies un- recognised tokens from a text as more or less peculiar. The tagger processes the less peculiar. According to capitalisation, the tagger treats po- tential common words and potential names sep- arately. It produces collections of new word types that should be learned. After minor hu- man supervision and correction �therefrom pre- fix semi- when describing Hascheck� new word types are added to Hascheck name and common word dictionary, respectively. If necessary, only accepted common word types are used for up- dating the n-gram base. Hascheck is open to all word processors. The service is fairly quick in responding: For a book of 100,000 tokens, it takes Hascheck less than 1 minute to reply. During the first 5 years of pub- lic life Hascheck received over 2,000 texts for processing, or a text corpus amounting to more than 12,000,000 tokens. Actual number of texts processed by Hascheck is over 3,000, which makes the text corpus larger than 35,000,000 288 Developing a Spell Checker as an Expert System tokens. The difference between the first and the second part of Hascheck’s life gives an idea about the quality of its service. As mentioned earlier, Hascheck started to ope- rate with a dictionary of less than 100,000 com- mon word types. In the first 5 years its know- ledge grew to more than 300,000 Croatian com- mon word types, and more than 80,000 name types. The increase in the next 5 years �ap- proximately� – this period was not taken into account by modelling that follows – was not so prominent. Only 100,000 new common word types and 50,000 new name types were learned from the corpus of over 20,000,000 tokens. 3. Learning Process Hascheck learns words to improve its cover- ing of Croatian texts. In order to keep track of learning, an elementary statistical record fol- lows each processing. On the basis of these records, two variables – text coverage �T C� and learning index �LI�, are calculated: TC � � 1 � NumberO f UnrecognizedTokens TextVolume � � 100 �1� LI � � NumberO f WordTy pesLearned TextVolume � � 100 �2� To describe Hascheck learning process over the time it was necessary to find the connection be- tween T C and LI, respectively, and the volume of previously processed corpus �PPC�. PPC, the total amount of text processed by Hascheck before the checking of a particular text file, is the only acceptable measure of Hascheck’s ma- turity. PPC is the natural substitution for time in analytical modelling of Hascheck’s learning. We took 20 LI values obtained by proofreading the Croatian Lexicon �Dembitz&Sokele, 1998�. The Croatian Lexicon was split into 20 text files. These files came in for checking sepa- rately, with long time spans, during the period between November 1994 and May 1997. The Lexicon proofreading was an extremely well controlled process. Therefore, we are sure that statistical data obtained from this process are very reliable. Using PPC as the time variable, we looked for functions that best fit the data. The number of free parameters in all tested functions was lim- ited to 3 or less. The best fit was obtained by using: LI � a � �100 � a� � e� PPC�∆t τ �3� Construction of the function �3� needs an ex- planation. Learning index LI, as defined in �2�, cannot exceed the value of 100. The extreme value of LI can be reached only at the very be- ginning of learning, when the “time”, PPC�∆t, equals 0. The “time” shift ∆t is also easy to ex- plain. To put Hascheck in operation, capable to receive the first user’s text and to give a satisfac- tory response, some amount of research, devel- opment and programming was needed, as well as processing of a certain amount of text in or- der to acquire basic Croatian word knowledge. All these are expressed through ∆t parameter. The result of fitting is presented in Fig. 1, to- gether with optimal values of free function pa- rameters a, ∆t and τ , and with obtained corre- lation coefficient r. It is worth mentioning that even physicists experimenting with some fun- damental physical law, would be satisfied with the correlation of 0.975. Fig. 1. Fitting of learning index LI �X � axis � PPC �tokens�; Y � axis � LI�. Independent of learning indexes we took 20 T C values representing average text coverage cal- culated on the basis of a 3-month period of pro- cessing. These 20 points have a total span of 5 years or 12,000,000 tokens, when counted in PPC. By fitting these data we fixed the parame- ters ∆t and τ on values obtained in the previous fitting. The function we chose was a natural choice: T C � b � � 1 � e� PPC�∆t τ � �4� Developing a Spell Checker as an Expert System 289 The result of this fitting is presented in Fig. 2. The fitting was again very good, since the cor- relation coefficient �parameter r in Fig. 2� was near 0.92. Fig. 2. Fitting of text coverage TC �X � axis � PPC �tokens�; Y � axis � TC �%��. Functions �3� and �4� correspond to human experience. Learning words is a never-ending process, since every living language constantly produces new words; this is expressed by the small constant a in Fig. 1. “3 to 5% of word to- kens are usually missing in the lexicon �diction- ary� when tagging a real-world text” �Mikheev, 1996�. The value obtained for b in Fig. 2 con- firms these borders. This correspondence with experience is giving us the faith that our mathematical model of learning makes sense. 4. Cognoelectrical Analogy Analogy is not the best way of scientific rea- soning. However, if there is no better way of thinking, one must take this one. Functions very similar to �3� and �4� describe the charging of a real capacitor connected to a real battery �Fig. 3� where τ � C � RC � Rs��RC � Rs�. According to the standard nota- tion in circuit theory, the battery is represented by an electromotive force voltage source with constant voltage E and source resistance Rs. The capacitor is represented by a capacitance C and capacitor resistance RC� When they are connected, the current �i� and the voltage �uC� follow the equation �5� and �6�, respectively. i � E RC � Rs � E Rs � RC RC � Rs � e� t τ �5� uC � E � RC RC � Rs � �1 � e� t τ � �6� The power function, p�t� � i�uC, has an extreme if RC � Rs. When the extreme of power exists, it is maximum in t � τ � ln�2RC��RC � Rs��. Fig. 3. Electrical equivalence of Hascheck learning process. Words are the charge for Hascheck. The flow of words, expressed by learning index LI, can be regarded as an equivalent to current i. Ana- logously, the text coverage can be regarded as voltage uC. Users, with their natural language competence, are the battery; Hascheck is the ca- pacitor. We call this a cognoelectrical analogy. Having equivalences between LI and i, and be- tween T C and uC, it is normal to construct and analyse the power of learning function �PoL�: PoL � h a � �100 � a� � e� PPC�∆t τ i � b � � � 1 � e� PPC�∆t τ � �7� 290 Developing a Spell Checker as an Expert System The function �7� has an extreme when the “time” has a value of �parameters are taken from Fig. 1�: PPC � ∆t � τ � ln 200 � 2a 100 � 2a � 1� 465� 315 �tokens� �8� Now we must go back to the consideration of the value of “time” shift ∆t��4� 774� 792 �to- kens�, obtained by fitting. As we have already said, this value expresses the fact that Hascheck needed acquisition of basic word knowledge be- fore it could become a useful service. However the “time” shift amounting of nearly 5 Mto- kens cannot be verified by the text corpus used for it; only 800,000 real tokens, processed by Hascheck before it was offered to public use, can be quoted �Dembitz, 1993�. The rest has to be considered as another type of time in know- ledge acquisition, the time needed for research, development and programming of Hascheck. Therefore we split the entire life of Hascheck in two periods: its virtual time, approximately 2τ long, when Hascheck was only an idea, and its real time, approximately 7τ long, when Hascheck was coping with real texts. Fig. 4. The Hascheck power of learning function: Region A = energy spent in virtual time of Hascheck’s life; Region B = energy spent in real time of Hascheck’s life; �X � axis � PPC �tokens�; Y � axis � PoL�. The Hascheck power of learning function PoL is presented in Fig. 4. It is obvious that the most energy-demanding period was the virtual time of Hascheck’s life. The doubts, when one is faced with many problems to be solved, con- sume much energy. During the real time of Hascheck’s life, when we had these problems mainly solved, the demand for energy dropped significantly. Here the analytic model proves our own experience. Generally speaking, these energy considera- tions do not contradict common sense. Au- tomata are built up to save human energy; users sent their text to Hascheck to save their own time needed for proofreading. In case of learn- ing automata – or semiautomata, like Hascheck – it should be compensated by the authors’ en- ergy spent in the course of making such an au- tomaton operable. To put it simply, learning automata operate due to the power = knowledge initially supplied by their authors. Whether they will function well or not, it depends upon the quality and quantity of knowledge initially sup- plied, and in each concrete case an estimation of these is still pure art. Since expert systems are still an applied dis- cipline, and not a theoretical one, let us con- clude with the last Hascheck practical achieve- ment. In the first week of September 2003 Hascheck checked the entire Comprehensive Dictionary of Croatian Language. The 4th edi- tion of Anić’s well known dictionary �Anić, 1991, 1994, 1998� is to appear soon under a new title, because of an increase of vocabu- lary volume covered by the Dictionary. Before its appearance in the book-shops the publisher wanted to check the final lay-out of the Dic- tionary, so he decided to use Hascheck. The results were astonishing �for the publisher�. In the text of 1,300,000 tokens Hascheck detected more than 600 typos and misspellings. Glob- ally, for the publisher, it was a very good rate because 0.05% error-rate is acceptable even in better lexicographies than Croatian, but the peo- ple who spent a decade or more working on the Dictionary project couldn’t believe they had never noticed these errors. From the Dictionary proofreading Hascheck learned only 2,000 new word types. 5. Conclusion The results we have presented here are conse- quences of a well-chosen and very unconven- tional approach. Being engineers, we always regarded natural language as an economic and energy-balanced system. This point of view is not new. It was very modern in technical Developing a Spell Checker as an Expert System 291 literature 50 years ago �Shannon, 1951; Zipf, 1949�, but was later replaced by structural and algebra-based approaches to language �Chom- sky, 1957�. Inspired by Martinet’s philosophy of language �Martinet, 1960�, we decided to explore the hidden, natural language energy, or geometry of language, in order to solve, with minimal effort, the problem of developing a good spell checker for our highly inflected language. We believe that our achievement, presented here, might produce some impact on the renewal of energy considerations about lan- guage and knowledge, since these two concepts are inseparable. References �1� ANIĆ, V., Rječnik hrvatskoga jezika (Dictionary of Croatian Language, 1stedition 1991, 2ndedition 1994, 3rdedition 1998), Novi Liber, Zagreb, Croatia, 1991, 1994, 1998. �2� BENTLEY, J., A Spelling Checker, Commun. of the ACM, 1985, 28�5�:456–462. �3� CHOMSKY, N., Syntactic Structures, Mouton&Co., The Hague, 1957. �4� DEMBITZ, S., Automatic Misspelling Detection and Telecommunication Services �in Croatian�, PhD Thesis, Faculty of Electrical Engineering, Univer- sity of Zagreb, 1993. �5� DEMBITZ, S. & SOKELE, M., Computational Proof- reading of the Croatian Lexicon, Proc. of the 9th Mediterranean Electrotechnical Conference, 1998, pp. 1370–1374, Tel-Aviv, Israel, May 18–20. �6� DOLGOPOV, A.S., Automatic Spelling Correction, Cybernetics, 1986, 22�3�:332–339. �7� MARTINET, A., Eléments de linguistique générale, Armand Colin, Paris, 1960. �8� MIKHEEV, A., Unsupervised Learning of Word- Category Guessing Rules, Proceedings of the 34th Annual Meeting of the Association for Computa- tional Linguistics, University of California, Santa Cruz, 1996, pp. 62–70. �9� MORRIS, R. & CHERRY, L.L., Computer Detection of Typographical Errors, IEEE Trans. on Professional Communications, 1975, PC-18�1�:54–64. �10� RISEMAN, E.M. & HANSON, A.R., A Contextual Postprocessing System for Error Correction Us- ing Binary n-Grams, IEEE Trans. Comput., 1974, C-23�5�:480–493. �11� SHANNON, C.E., Prediction and Entropy of Printed English, Bell System Technical Journal, 1951, 30�1�:50–64. �12� SOKELE, M., Croatian Spelling Checkers �in Croa- tian�, WIN.INI, 1997, 6�3�:38–49. �13� TAYLOR, R.M., Expert Systems in the Knowledge Management Era, Applications and Innovations in Expert Systems VI �R.Milne et al., Eds�, 1999, pp. 3–8, Springer-Verlag. �14� ULLMANN, J.R., A Binary n-Gram Technique for Automatic Correction of Substitution, Deletion, In- sertion and Reversal Errors in Words, Computer Journal, 1977, 20�2�:141–147. �15� USCHOLD, M., Knowledge Level Modelling: Con- cepts and Terminology, The Knowledge Engineering Review, 1998, 13�1�:5–29. �16� WEISCHEDEL, R., METEER, M., SCHWARTZ, R., RAMSHOW, L. & PALMUCCI, J., Coping with Am- biguity and Unknown Words through Probabilis- tic Models, Computational Linguistics, 1993, 19�2�:359–383. �17� ZIPF, G.K., Human Behavior and the Principle of Least Effort, Cambridge: Addison-Wesley, 1949. Received: April, 2002 Revised: November, 2003 Accepted: December, 2003 Contact address: Šandor Dembitz, Petar Knežević Faculty of Electrical Engineering and Computing University of Zagreb Unska 3 10000 Zagreb Croatia Phone: �385 1 6129760 Fax: �385 1 6129616 e-mail: sandor�dembitz�fer�hr petar�knezevic�fer�hr Mladen Sokele HT – Croatian Telecom Inc. Corporate Strategy and Development Department Hebrangova 32–34 10000 Zagreb Croatia e-mail: mladen�sokele�ht�hr ŠANDOR DEMBITZ received the BSc �1973�, MSc �1981� and PhD �1993� degrees in electrical engineering, all from the University of Za- greb, Croatia. He is presently an assistant professor at the Faculty of Electrical Engineering and Computing, University of Zagreb. His cur- rent research interests include natural language processing and expert systems. PETAR KNEŽEVIĆ received the BSc �1974�, MSc �1977� and PhD �1982� degrees in electrical engineering, all from the University of Zagreb, Croatia. He is presently an associate professor at the Faculty of Elec- trical Engineering and Computing, University of Zagreb. His current research interests include formal description techniques for protocol specification, performance analysis and system modelling. MLADEN SOKELE received the BSc �1983� and MSc �1987� degrees in electrical engineering, both from the University of Zagreb, Croatia. He is presently at the Corporate Strategy and Development Department of HT �Croatian Telecom�, where he runs the Research and Informa- tion Centre. His current research interests include economic indicators modelling and forecasting in telecommunications. 
work_weka6p4pybcplgk5msixjmvjui	----	doi:10.1016/j.aquaeng.2004.12.003 SEDPA, an expert system for disease diagnosis in eel rearing systems J.C. Gutiérrez-Estrada a,*, E. De Pedro Sanz b , R. López-Luque c , I. Pulido-Calvo a a Dep. Ciencias Agroforestales, Univ. Huelva, EPS, Campus Universitario de La Rábida, 21819 Palos de la Frontera (Huelva), Spain b Dep. Producción Animal, Univ. Córdoba, ETSIAM, Avda. Menéndez Pidal s/n, 14080 Córdoba, Spain c Dep. Fı́sica Aplicada, Univ. Córdoba, ETSIAM, Avda. Menéndez Pidal s/n, 14080 Córdoba, Spain Received 22 October 2003; accepted 1 December 2004 Abstract The design, development and testing of a prototype interactive expert system (SEDPA) capable of diagnosing eel pathologies is described. The system incorporates a multiple subprogram modular design, although only the inference engine is largely described. Its user interface incorporates a natural language module. Starting from this kind of information, the system obtains its conclusions treating this information with a fuzzy controller and transmitting the uncertainty using the Dempster– Shafer theory (DST). The system’s performance was evaluated in a series of tests. The results of a Fisher’s exact test of the system’s diagnoses versus those of the three fish pathologists for 29 eel pathologies indicated statistical differences in diagnostic performance with the two human experts. On the other hand, the use of different fuzzy associative memory (FAM) or representation of the different human experts’s knowledge level indicated the system adaptability to the different work scenarios. The results described in this study have demonstrated the validity of this software for eel pathological diagnosis. # 2005 Elsevier B.V. All rights reserved. Keywords: Expert system; Decision support system; Aquaculture; Fuzzy control; Dempster–Shafer theory www.elsevier.com/locate/aqua-online Aquacultural Engineering 33 (2005) 110–125 * Corresponding author. E-mail addresses: juanc@uhu.es (J.C. Gutiérrez-Estrada), emiliano.depedro@uco.es (E.D.P. Sanz), fa1lolur@uco.es (R. López-Luque), ipulido@uhu.es (I. Pulido-Calvo). 0144-8609/$ – see front matter # 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.aquaeng.2004.12.003 1. Introduction The development of aquaculture and sport fishing have shown the importance of diseases in fish populations. Often these activities are the propagation vector of some diseases (Höglund and Andersson, 1993). The disease is presented in the fish in the form of lesions and symptoms. This causes a decrease in the physical quality of fish and final yield and often the death of the affected fish. This morbid is caused by changes in the physical and chemical parameters of water and the activity of biological aggressors (virus, bacteria, fungus and animal parasites) (Kinkelin et al., 1991). On the other hand, the management confined fish populations requires the introduction of other factors (chemical substances, grading, feeding, transport, etc.) that may further increase the harmful effects of physical, chemical and biological factors present in water (Wickins, 1981). Therefore, a fish health policy has a great economic importance by requiring the control and elimination of pathological problems. Eel intensive rearing systems are clear examples of the induction of pathological agents. In this kind of system, disease is a consequence of a poor water quality, lack of hygiene, inadequate management (grading and transfer) and nutritional deficiencies, or by the combination of two or more of these factors (Munro and Fijan, 1981). The illnesses may be eliminated by applying an adequate pharmacological treatment which implies a previous diagnosis process. However, the diagnosis process is hindered by the shortage of ictiopathologists or by the abundance, dispersion and lack of the available information. This has a great influence on the time taken for observing the first symptoms until the application of an appropriate treatment. This characteristic justifies the utilisation of the expert systems. Expert systems arose in the 1970s in the field of artificial intelligence as computer software that employs knowledge captured in a computer program to solve problems that usually require human expertise. Well-designed expert systems imitate the reasoning process of human experts to solve specific problems and can be used by nonexperts to improve their problem-solving capabilities and by experts as knowledgeable assistants. Most famous expert systems were built as diagnosis assistants and therapy advisors in different medical areas (MYCIN, PIP, CENTAUR, INTERNIST, ONCOCIN, etc.) (Dı́ez et al., 1997) and almost all of them based their reasoning totally or partially on if/then rules which, combined with frames or objects, has constituted until the present the standard method for building expert systems. However, the use of rules raises serious problems with regard to the knowledge representation and uncertainty management. They apply a formalism developed for classical logic in which every proposition is either true or false and is not suitable to the management of uncertainty through numerical factors (Heckerman and Horvitz, 1988; Dı́ez et al., 1997). Our goal was to develop an expert system, not based on if/then rules, that incorporates most of the disease which affect eel fishfarms. This system could then be used by the experts and nonexpert users. We tried to lower the risk of obtaining a wrong decision and maximise the possibility of making a right decision by means of a fuzzy controller and the Dempster–Shafer theory. Fuzzy logic technology (Zadeh, 1992) has been recognised recently by the Institute of Electrical and Electronics Engineers (New York) as one of the J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125 111 three key information processing technologies. This fuzzy logic attribute allows the capture of human thought processes in an optimal manner for automation (Lee et al., 2000). Recent results have also shown that fuzzy logic systems are universal approximators for general nonlinear functional relationships to any desired degree of accuracy (Kosko, 1993a,b). This makes fuzzy logic modelling a powerful tool for exploring complex, nonlinear biological problems (Mackinson et al., 1999; Chen, 2001). On the other hand, Dempster–Shafer theory (mathematical theory of evidence) is a numerical method for evidential reasoning. The mathematical theory of evidence begins with the assumptions that a real number between 0 and 1 indicates the degree of support the evidence provides for a proposition and the beliefs are not necessarily additive (Shafer, 1976). Thus, Dempster–Shafer’s belief theory can be viewed as an extension of the Bayesian probability theory (Wong and Yao, 1992). This expert system, called SEDPA, can suggest several diseases with different belief levels, allowing the ultimate decision to be made by the manager of the fishfarm. 2. Materials and methods The methods discussed here were applied to Hidrorecursos S.A., an intensive eel fishfarm located in the province of Córdoba (southern Spain). The fishfarm has three biological filtration units for water reuse. Globally, the average flow through the system was 3185 � 1634.2 m3/day and the exchange rate in the biological filters were 10% per day. 2.1. General structure of the expert system The expert system consists of several interrelated modules. These relationships together with the data provided by the user allows obtaining a conclusion. This way, the expert system has one or several databases or knowledge bases, an inference engine or kernel of the system, an explicative subsystem (provides information about the inference engine’s logic), a proposal engine (help to user in the diagnostic process), a learning subsystem (incorporate knowledge in the system) and an user interface. In this paper, only SEDPA’s inference engine is described. 2.2. Domain databases The domain databases are the permanent memory of the system. They store necessary information that the system uses to obtain its conclusions. These data were obtained from two principle sources: (a) literature and aquacultural journals (Roberts, 1981; Clifton- Hadley et al., 1984; MacConnell et al., 1989; Mellergaard and Dalsgaard, 1989; Kinkelin et al., 1991; Shepherd and Bromage, 1999); and (b) the historic data of physical, chemical, biological and yield parameters obtained from 1997 in the farm. The first kind of data (from literature and aquacultural journals) is stored in the ‘Principal Domain Database’ (PDD). In PDD, the relationships between syntactic labels and the eel pathologies are established with ‘1’ and ‘0’ (Van Diest et al., 1994). For the organisation of the dictionary used in the J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125112 syntactic analysis, an adaptation of the lexical approximation is carried out applying a hierarchically structured dictionary (Steffens, 1994). Therefore, only the canonical form of the words are stored in the PDD (Winiwarter, 2000). On the other hand, the historic data of physical, chemical, biological and yield were stored in the ‘Secondary Domain Database’ (SDD). 2.3. Inference engine The inference engine is the brain of the expert system. This component provides the methodology for reasoning by using the information from the domain databases to formulate conclusions. In SEDPA, the inference engine has three principle parts: (a) an augmented transition network (ATN); (b) a fuzzy logic controller; and (c) a system of uncertainty transmission based on the Dempster–Shafer theory. 2.3.1. The augmented transition network (ATN) The ATN is a method of syntactic analysis that allows the inference engine to access the data stored in the domain databases, starting from the information in natural language entered by the user (Woods, 1973; De Carolis et al., 1996). The use of the ATN assures the information introduced to the expert system is coherent. This allows calculating the relative frequencies established between the pathologies associated with the symptoms/lesions observed and the synctactic labels stored in the PDD (Fig. 1). Initially, the system only processed information in Spanish but at the moment an English version of the ATN is being debugged. 2.3.2. The fuzzy logic controller Fuzzy logic utilises a many valued form of logic. Unlike the ‘crisp’ logic of a yes–no controller, a fuzzy logic controller has in-between values. That is to say, a fuzzy set is divided by the geometric partitions. This allows us to describe a point as a function of its membership in different sets (Zeldis and Prescott, 2000). In SEDPA, the fuzzy logic controller has two input fuzzy sets (1: the minimum relative frequency (MRF) for each symptom/lesion and 2: the incidence (I) of each one of these pathologies in eel rearing systems) and one output fuzzy set (the categorical belief level that one or more of these pathologies are responsible for the illness in the fishfarm). The explicit relationship between the partitions of the input fuzzy sets and the output fuzzy set is stored in fuzzy associative memory (FAM). A human expert creates a FAM and it is a reflex of his experience. In this paper, SEDPA obtains its conclusions using three different FAMs, each one developed by a different human expert (FAM A [expert A], B [expert B] and C [expert C]). The possible relationships between input and output fuzzy sets for FAM B are shown in Fig. 2. After the inputs (SL1 in Fig. 3) to the controller (MRF and I) have been processed by the control algorithm, the result is a fuzzy output mout( y) or a categorical belief level (CBEL) for the n pathologies that satisfy all the conditions imposed by the ATN (in Fig. 3, n = 3 [Pt1, Pt3 and Pt9] and MRF = 45.3%). For each pathology, selecting a crisp number y * , representative of mout( y), is a process known as defuzzification. In Fig. 3, defuzzification process for Pt1 is shown. In this case, the MRF for SL1 implies two rules in the MRF fuzzy input set: rule 1 = normal shared (NS, 0.75) and rule 2 = low shared (LS, 0.30). Over the years, several defuzzification techniques have been suggested (Yager and Filev, 1994; J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125 113 Tsoukalas and Uhrig, 1997). In SEDPA, the defuzzification technique is the center of area (COA) or the center of gravity defuzzification: y � ¼ PN i¼1 y CenterðiÞðmoutðyiÞÞPN i¼1ðmoutðyiÞÞ (1) where y Center(i) is the mass center of the ith fuzzy partition. In COA defuzzification, the crisp value y * is taken to be the geometrical center ( y Center(i) ) of the output fuzzy value mout( y i ), where mout( y) is formed by taking the union of all the contributions of rules whose degree of fulfillment is >0. In the example, the defuzzyficate value for Pt1 or Pt1 belief is 38.07 (Fig. 3). 2.3.3. Management of uncertainty: Dempster–Shafer evidence theory Management of uncertainty is crucial in the design of expert systems. Broadly speaking, three fundamental issues must be considered: the representation, propagation and the combination of uncertain information. A common method for managing uncertainty is the J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125114 Fig. 1. Augmented transition network (ATN) schematic representation of the symptom/lesion: ‘the eels of the tank A4 have white stains in the head’. bayesian network (Liu and Bundy, 1994). Bayesian networks provide a formalism for reasoning about the degrees of belief under the conditions of uncertainty. In this formalism, propositions are given the numerical probability values, signifying the degree of belief accorded them, and the values are combined and manipulated according to the rules of probability theory (Haddawy et al., 1994). However, when the information volume is very low and the use of probability functions is not possible, it is advisable to use other alternative techniques. In this case, a candidate for managing uncertainty in the expert systems is the Dempster–Shafer evidence theory (DST) (Hégarat-Mascle et al., 1997). DST represents the relevant characteristics of the world as a finite set of mutually exclusive propositions and assumptions called the frame of discernment or space of hypotheses (Q). In SEDPA, Q is a set of pathologies that the system may diagnose. DST allows considering any subset of Q (we denote 2 Q , the set of the subset of Q). In this case, a J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125 115 Fig. 2. A fuzzy associative memory (FAM) and the meaning of labels. The meaning of ‘shared’ in this context is: with how many pathologies a certain symptom is related to the Principal Domain Database (PDD)? J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125116 Fig. 3. The example represents the system running (steps 0–3) when the user introduces a first symptom/lesion (SL1). Step 0 is the introduction of the symptom/lesion in natural language form in SEDPA. Step 1 is the ATN creation and calculus of the minimum relative frequecy (MRF). Step 2 is the assignment of a particular belief level for each pathology that satisfies all the conditions imposed by the ATN (step 1). Step 3 is the calculus of average belief level for the symptomatic group associated to SL1. subset of Q (or symptomatic group) called Ai (Ai 2 2Q) is formed by n pathologies associated with an event (symptom/lesion) observed by the user. In Ai, each pathology has a belief level (previously assigned through the fuzzy logic controller). This way, the belief level of Ai is the average belief level of its n pathologies (Fig. 3, step 3). The DST provides a representation of both the imprecision and uncertainty through the definition of two functions, plausibility (PLS) and belief (QBEL) which are both derived from a mass function (m). This m is defined for every subset Ai of 2 Q such that the mass value m(Ai) belongs to the [0, 1] interval and m : mð?Þ ¼ 0X 8 i mðAiÞ ¼ 1 8 < : (2) where 1 is the empty set. DST provides a method to combine the data from different sources, that is to say, to combine different symptomatic groups from different symptoms/lesions observed by the user. (In Fig. 4, upper square, the results of SL1 are combined with results of SL2. The final result is three new symptomatic groups [A3, A4 and A5]. In this case, 93% of the belief is accumulated in the A3, A4 and A5 symptomatic groups.) If mi is the basic probability assignment provided by a source, the combination: m = m1 � . . . � m p, also called orthogonal sum, is defined according to the Dempster’s combination rule (Shafer, 1976) by mð?Þ ¼ 0 if K < 1; mðCÞ ¼ P B1 \ ... \ Bp¼C Y 1 i q; 1 j p miðBjÞ 1 � K where K ¼ X B1 \ ... \ Bp¼? Y 1 i q; 1 j p miðBjÞ 8 >>>>>>>< >>>>>>>: (3) where C and B j are sets of symptomatic sets. K represents the mass which would be assigned to the empty set, after combination, in the absence of normalisation (division by (1 – K)). Thus, K is often interpreted as a measure of conflict between the different sources and it is introduced in (3) as a normalisation factor (Hégarat-Mascle et al., 1997). In the example (Fig. 4, lower square), the combination of A3, A4 and A5 symptomatic groups with the new symptomatic groups from SL3 (A6) generates two empty symptomatic groups (A7 and A9). The final belief of no empty groups (A8, A10, A11, A12 and A13) is scaled for K = 0.50 (A7 belief + A9 belief). Having computed the mass, plausibility and belief values for each simple and compound hypothesis, we need a criterion, which is called decision rule, to decide which hypothesis is the most realistic (Shafer, 1976). In SEDPA, the decision rule is the maximum of belief. In the example (Fig. 4, lower aquare), A8 belief is 0.40. 2.3.4. Evaluation of SEDPA A basic requirement for the acceptability of expert systems for the clinical diagnosis is that, instead of providing unique solutions, they select and submit to the user the whole set of hypotheses which are compatible with the available data. Such a strategy was applied in the present study, assuming that the condition for reliability of the system conclusion was J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125 117 J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125118 Fig. 4. The example represents the system running when the user introduces a second (SL2) and third (SL3) symptom/lesion. Upper square represents the combination of SL1 and SL2 by Dempster–Shafer’ combination rule. The result of this combination is three new symptomatic groups (A3, A4 and A5). Lower square represents the combination of SL3 and SL1 \ SL2. In this case, the Dempster–Shafer’ combination rule generates two empty sets (A7 and A9). Consequently, the belief level of rest of the symptomatic groups (A8, A10, A11, A12, A13 and Q) is rescaled for K = 0.50 (A7 belief = 0.40 + A9 belief = 0.10). the inclusion of the reference diagnosis among those selected as compatible (Molino et al., 1996). According to such a strategy, system conclusions were regarded as ‘incorrect’ only when the reference diagnosis was missed. While all the hypotheses selected as compatible were accepted as reliable and classified either as ‘correct’ (when one pathology got the highest belief level) or ‘approximate’ (whenever the right pathology was selected together with other options). The reference set was compiled using 29 cases detected in Hidrorecursos S.A. The specific data from these 29 cases were not stored into the database domain of the system. The diagnoses of each case (1–29) was judged by three experts in fish pathology (experts RL1, RL2 and RL3) in a reference laboratory. The diagnosis of the reference laboratory experts was taken as the ‘gold standard’ (Van Diest et al., 1994). Later on, these 29 cases were judged by SEDPA and a panel of other three different pathologists (experts D1, D2 and D3) who had varying experience in fish pathology. These pathologists were asked to classify the cases in their usual way, allowing for the consultation of books and/or journals in order to simulate a real situation. Finally, the diagnoses from the SEDPA and D1, D2 and D3 experts were compared with the gold standard. 3. Results 3.1. Human experts versus SEDPA A comparison of SEDPA diagnoses versus those obtained from experts D1, D2 and D3 are shown in Table 1a and b. The exact relationship between the human experts and SEDPA diagnostics and the reference diagnostic or ‘gold standard’ is presented in the principle diagonal. For example, case 4 (column 4) was correctly judged by the experts D2 and D3 and SEDPA. However, in this case, expert D2 included in his answer possible pathologies of cases 5 and 6. In the same way, expert D3 included in his answer a possible diagnosis of case 5 and SEDPA included in its answers cases 6 and 8. Therefore, in case 4, the answers of experts D2 and D3 and SEDPA were classified as ‘approximate’. On the other hand, the expert D1 included in his answer only one pathology. This pathology was different from ‘gold standard’, therefore; it was classified as ‘incorrect’. In some cases (6, 15, 17–23, 25 and 27), the human experts included in their answer different pathologies neither of which was the ‘gold standard’. These diagnostics were included in the ‘others’ (O) category. Globally, in the validation phase, the best result was obtained by SEDPA (20 answers classified as ‘correct’, 8 answers classified as ‘approximate’ and 1 answer classified as ‘incorrect’). Among the human experts, expert D1 obtained the best result (14 answers classified as ‘correct’ and 15 answers classified as ‘incorrect’). The diagnostics of the experts D2 and D3 were similar (expert D2: 6 answers classified as ‘correct’, 7 answers classified as ‘approximate’ and 16 answers classified as ‘incorrect’; expert D3: 7 answers classified as ‘correct’, 3 answers classified as ‘approximate’ and 19 answers classified as ‘incorrect’). Therefore, the success rate of the experts D1, D2 and D3 were 48.3%, 20.7% and 24.1%, respectively. If the answers classified as ‘approximate’ are reclassified as J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125 119 J .C . G u tié rre z-E stra d a e t a l./ A q u a c u ltu ra l E n g in e e rin g 3 3 (2 0 0 5 ) 1 1 0 – 1 2 5 1 2 0 Table 1 (a and b) Diagnostic of human experts (D1 = 1, D2 = 2, and D3 = 3) and SEDPA (S) vs. the gold standard (GE) for each case GE Human experts (D1 = 1, D2 = 2, D3 = 3) and SEDPA (S) diagnostic Case 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 (a) 1 3, S 2 1, S 3 2, 3 3 1, 2, 3, S 2 2, 3 2 2, 3, S 4 2, 3, S 5 2 2 1, 2, 3 1, 2, 3, S 2 2, 3 1, 2 2 S 2 3 6 1 2, S S 2 7 S S S 8 S S 3 9 1, S S 10 1, 2, S 11 1 1, 2, 3, S 2 12 1, 2, 3, S 13 1, 2, S 14 2, 3, S 2 1, 2 3 2, 3 J .C . G u tié rre z-E stra d a e t a l./ A q u a c u ltu ra l E n g in e e rin g 3 3 (2 0 0 5 ) 1 1 0 – 1 2 5 1 2 1 Table 1 (Continued ) GE Human experts (D1 = 1, D2 = 2, D3 = 3) and SEDPA (S) diagnostic Case 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 (b) 15 S 1, 2 2 1 16 S 17 2, S 18 S 19 S 20 S S 21 1, 2, S 22 S 23 S S 24 1, 2, 3, S S 25 S 26 2 2 2 2 2 2 2 1 2 27 1, 3 2, S 28 1, 2 2, 3 1, 2, S 29 1, 2, 3, S O 1 1 2, 3 1 1, 2 2 2 1 1, 3 2 1 The principle diagonal (bold) indicates the correct diagnosis. Each case results are detailed by the columns. For one case (column), when a number (1, 2 or 3) or S is in more of one file means that human expert or SEDPA generated more of one answer for this case. For example, see case 1 (column 1). The category ‘O’ indicates a diagnosis whose result disagrees with the gold standards judged in the reference laboratory. ‘correct’, then the success rates were incremented to 48.3%, 44.8% and 34.5%, respectively. With regard to SEDPA, the success rate with answers strictly correct was 69.9%. When the ‘approximate’ answers were considered, the success rate was increased to 96.6%. The results of a Fisher’s exact test demonstrated no statistically significant difference in the number of cases agreeing with the ‘gold standard’ by SEDPA and expert D1 ( p = 0.0910). On the other hand, the differences were statistically significant when the results obtained by experts D2 and D3 were compared in the same way to the SEDPA results (expert D2 versus SEDPA: p < 0.001; expert D3 versus SEDPA: p < 0.001) (Table 2). When the approximate answers of the human experts and SEDPA were reclassified to ‘correct’, the statistical differences were significant in all the cases (experts D1, D2 and D3 versus SEDPA: p < 0.001). 4. Discussion Previous works in the field of expert systems applied to aquaculture have shown that this kind of computer program or software may be useful for a great variety of purposes: financial analysis, yield management, diagnosis and pathology treatment (Schulstad, 1997). These activities are intimately related to the correct management of any aquaculture system. This way, Zeldis and Prescott (2000) developed the FISH-VET system which may process information for different species and may diagnose pathologies associated with fish rearing. The general-purpose systems may provide a correct solution in a multitude of scenarios but their conclusions may be incomplete whenever the available information is scarce. Generally, this is because incorporating of this kind of information complicates the construction of the knowledge base used by the system. The special information is easier to integrate in a system with a limited workspace which makes it easier to obtain useful conclusions to the technician responsible for the health policy in the fish farm. However, the construction of expert systems under special conditions (data absence, fuzzy or redundant information) may imply a high economical and temporal investment that is directly related to the number and kind of rules included by the program designer. J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125122 Table 2 Fisher’s exact test of SEDPA’s diagnosis vs. human experts’s diagnosis (experts D1, D2 and D3) SEDPA’s diagnosis Human expert’s diagnosis Agree 20 14 (Expert D1) Disagree 8 15 (Expert D1) N = 29; p = 0.0910 Agree 20 6 (Expert D2) Disagree 8 23 (Expert D2) N = 29; p < 0.001 Agree 20 7 (Expert D3) Disagree 8 22 (Expert D3) N = 29; p < 0.001 We successfully developed and tested an expert system that can be used as a tool in the diagnosis of eel pathologies. The design of SEDPA permits the entry of data in natural language form, helping the technician to reach a diagnosis. Consequently, SEDPA could potentially improve the technician’s efficiency, shorten the time-consuming process of diagnosis, increase the economic efficiency of diagnosis and decrease the cost of mortality. Computationally, SEDPA combines some advantages of the two kinds of systems previously described. On the one hand, its workspace is reduced to only one species which resulted in success rates close to 70%. On the other hand, the knowledge acquired by SEDPA is not stored in the rule form, so that the system is very versatile and easily adapted to the different work scenarios. This program’s nature facilitates the implementation and adaptation of the system to eel fishfarms with different characteristics and permits the inclusion of the knowledge of human experts with different experience levels. Hopefully, the system could serve as the prototype for other species with similar knowledge levels because the binary relations in the PDD are easily replaceable. Thus, foregoing rules confers both important advantages and high obstacles too. Modelling the knowledge base with if/then rules makes it possible to integrate the heuristics that guide a human expert in the diagnostic decision process. Expert systems based on rules as GOLDFINDER (Hawkes, 1992), MUNIN (Andreassen et al., 1995) and XPHEMO (Nguyen et al., 1996) may provide information to the user about the reasoning process of the inference engine. This retrospective process is not possible in SEDPA. The results of the evaluation process show that the SEDPA’s behaviour was correct because SEDPA’s success rate was significantly higher than the human experts success rates. These differences could have several explanations. For example, the human expert process the information in natural language form assuming that this information is compounded by a labels set whose combinations make coherent sentences (Chomsky, 1959). Usually, in the eel health-policy context, the information obtained from the observation of symptoms and lesions has a ‘categorical character’. That is to say, the technician observations are used to define the ‘way’ or ‘route’ for the definitive pathology determination (chemical analysis of water, bacterial crop, historical data analysis and antibiogram). This ‘categorical character’ is enhanced when the action of the different pathological agents produces similar symptoms or lesions. However, SEDPA does not use this information in the same way. In this case, initially the information is statistically treated, calculating for each sentence’s label, the appearance frequency and analysing the relationships between the different label sets. Starting from these initial results, the system presents a high level of pathologies discrimination using a fuzzy controller. Lee et al. (2000) designed an expert system to study a denitrification system in a fishfarm using a fuzzy controller. The results showed that the expert system controlled the generated actions in conservative form and the discrimination capacity was better than those obtained when classical techniques were used. Mamdani (1977), Takagi and Sugeno (1985) and Whitsell and Lee (1994) obtained similar results using fuzzy controllers. The discrimination capacity of SEDPA is increased using a method classified as probabilistic technique. Although its use in practical cases has been very rare, Hájek (1994), Hégarat-Mascle et al. (1997) and Murphy (1998) report that the Dempster–Shafer theory is an effective method of uncertainty transmission. J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125 123 On the other hand, the absence of visual confirmation has a great influence on the human expert conclusions. Although the description of symptoms and lesions in each case was made by the three human experts the more ones systematically possible, one can argue that the information described may be misunderstood by other experts if the pathological effects cannot be personally observed. Acknowledgements The authors wishes to thank two anonymous reviewers for their constructive criticism of the system, suggestions and comments on an earlier version of the manuscript. References Andreassen, S., Rosenfalck, A., Falck, B., Olesen, K.G., Andersen, S.K., 1995. Evaluation of the diagnostic performance of the expert EMG assistant MUNIN. Electroencephalogr. Clin. Neurophysiol. 101, 129–144. Chen, D.G., 2001. Detecting environmental regimes in fish stock-recruitment relationships by fuzzy logic. Can. J. Fisheries Aquat. Sci. 58, 2139–2148. Chomsky, N., 1959. On certain formal properties of grammars. Inf. Control 2 (2), 137–167. Clifton-Hadley, R.S., Bucke, D., Richards, R.H., 1984. Proliferative kidney disease of salmonid fish: a review. J. Fish Dis. 7, 363–377. De Carolis, B., Derosis, F., Grasso, F., Rossiello, A., Berry, D.C., Gillie, T., 1996. Generating recipient-centered explanations about drug prescription. Artif. Intell. Med. 8 (2), 123–145. Dı́ez, F.J., Mira, J., Iturralde, E., Zubillaga, S., 1997. DIAVAL, a bayesian expert system for echocardiography. Artif. Intell. Med. 10, 59–73. Haddawy, P., Kahn, C.E., Butarbutar, M., 1994. A bayesian network model for radiological diagnosis and procedure selection: work-up of suspected gallbladder diseases. Med. Phys. 21 (7), 1185–1192. Hájek, P., 1994. Systems of conditional beliefs in Dempster–Shafer theory and expert systems. Int. J. Gen. Syst. 22, 113–124. Hawkes, D.D., 1992. GOLDFINDER: a knowledge-based system for mineral prospecting. J. Geol. Soc. 149, 465– 471. Heckerman, D.E., Horvitz, E.J., 1988. The myth of modularity in rule-based systems for reasoning with uncertainty. In: Lemmer, J.F., Kanal, L.N. (Eds.), Uncertainty in Artificial Intelligence, vol. 2. Elsevier, Amsterdam, pp. 23–34. Hégarat-Mascle, S., Bloch, I., Vidal-Madjar, D., 1997. Application of Dempster–Shafer evidence theory to unsupervised classification in multisourse remote sensing. IEEE Trans. Geosci. Remote Sensing 35, 1018–1031. Höglund, J., Andersson, J., 1993. Prevalence and abundance of Anguillicola crassus in the european eel (Anguilla anguilla) at a thermal discharge site on the Swedish coast. J. Appl. Ichtyol. 9, 112–115. Kinkelin, P., Michel, Ch., Ghittino, P., 1991. Tratado de las enfermedades de los peces. ACRIBIA, S.A., Zaragoza. Kosko, B., 1993a. Fuzzy Thinking: The New Science of Fuzzy Logic. Hyperion, New York. Kosko, B., 1993b. Fuzzy system as universal approximators. In: Proceedings of the 1992 IEEE Conference on Fuzzy System. IEEE Transactions on Computers. Lee, P.G., Lea, R.N., Dohmann, E., Prebilsky, W., Turk, P.E., Ying, H., Whitson, J.L., 2000. Denitrification in aquaculture systems: an example of a fuzzy logic control problem. Aquacult. Eng. 23, 37–59. Liu, W., Bundy, A., 1994. A comprehensive comparison between generalized incidence calculus and the Dempster–Shafer theory of evidence. Int. J. Hum.-Comput. Stud. 40, 1009–1032. MacConnell, E., Smith, C.E., Hedrick, R.P., Speer, C.A., 1989. Cellular inflammatory response of rainbow trout to the protozoan parasite that causes proliferative kidney diseases. J. Aquat. Anim. Health 1, 108–118. Mackinson, S., Vasconcellos, M., Newlands, N., 1999. A new approach to the analysis of stock–recruitment relationships: ‘‘model-free estimation’’ using fuzzy logic. Can. J. Fisheries Aquat. Sci. 56, 686–699. J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125124 Mamdani, E., 1977. Application of fuzzy logic to approximate reasoning using linguistic synthesis. IEEE Trans. Comput. 26 (12), 1182–1191. Mellergaard, S., Dalsgaard, I., 1989. Handbook of eel diseases. Damn Fisk og Havunders Rapport 293, 1–47. Molino, G., Molino, F., Furia, D., Bar, F., Battista, S., Cappello, N., 1996. Computer-aided diagnosis in jaundice: comparison of knowledge-based and probabilistic approaches. Methods Inf. Med. 35, 41–51. Munro, A.L.S., Fijan, N., 1981. Disease prevention and control. In: Proceedings of the World Symposium on Aquaculture ‘Heated Effluents and Recirculation System’, Berlin. Murphy, R.R., 1998. Dempster–Shafer theory for sensor fusion in autonomous mobile robots. IEEE Trans. Rob. Autom. 14 (2), 197–206. Nguyen, A.N.D., Hartwell, E.A., Milam, J.D., 1996. A rule-based expert system for laboratory diagnosis of hemoglobin disorders. Arch. Pathol. Lab. Med. 120, 817–827. Roberts, R.J., 1981. Patologı́a de los peces. Mundi-Prensa, Madrid. Schulstad, G., 1997. Design of a computerized decision support system for hatchery production management. Aquacult. Eng. 16, 7–25. Shafer, G., 1976. A Mathematical Theory of Evidence. Princeton University Press, Princeton. Shepherd, J., Bromage, N., 1999. Piscicultura Intensiva. ACRIBIA, S.A, Zaragoza. Steffens, P., 1994. Machine translation and the lexicon. In: Proceedings of the International EAMT Workshop. Takagi, T., Sugeno, M., 1985. Fuzzy identification of system and its applications to modeling and control. IEEE Trans. Syst. Man Cybern. 15, 116–132. Tsoukalas, L.H., Uhrig, R.E., 1997. Fuzzy and Neural Approaches in Engineering. Wiley-Interscience, New York. Van Diest, P.J., Beliën, J.A.M., Zanstra, P.E., Wilhelm, W.W., Baak, J.P.A., 1994. Integrated decision support system/image archive for histological typing of breast cancer using a relation oriented inference system. Histopathology 25, 253–259. Whitsell, A., Lee, P.G., 1994. A plug-and-play machine vision application for aquaculture. Sci. Comput. Autom. 10 (8), 29–32. Wickins, J.F., 1981. Water quality requirements for intensive aquaculture: a review. In: Proceedings of the World Symposium on Aquaculture ‘Heated Effluents and Recirculation System’, Berlin. Winiwarter, W., 2000. Adaptive natural language interfaces to FAQ knowledge bases. Data Knowl. Eng. 35, 181– 199. Wong, S.K.M., Yao, Y.Y., 1992. Characterization of comparative belief structures. Int. J. Man-Mach. Stud. 37, 123–133. Woods, W.A., 1973. Progress in natural language understanding: an application to Lunar geology. In: Proceedings of the AFIPS Conference 42. Yager, R.R., Filev, D.P., 1994. Essentials of Fuzzy Modeling and Control. John Wiley and Sons, New York. Zadeh, L.A., 1992. The calculus of fuzzy if/then rules. AI Expert 7 (3), 23–27. Zeldis, D., Prescott, S., 2000. Fish disease diagnosis program—problems and some solutions. Aquacult. Eng. 23, 3–11. J.C. Gutiérrez-Estrada et al. / Aquacultural Engineering 33 (2005) 110–125 125 SEDPA, an expert system for disease �diagnosis in eel rearing systems Introduction Materials and methods General structure of the expert system Domain databases Inference engine The augmented transition network (ATN) The fuzzy logic controller Management of uncertainty: Dempster-Shafer evidence theory Evaluation of SEDPA Results Human experts versus SEDPA Discussion Acknowledgements References 
work_wfkrssjq2beuzmkovhxwnymdae	----	Open Data for Anomaly Detection in Maritime Surveillance i Master’s Thesis Computer Science September 2012 School of Computing Blekinge Institute of Technology SE – 371 79 Karlskrona Sweden Open Data for Anomaly Detection in Maritime Surveillance Shahrooz Abghari Samira Kazemi This thesis is submitted to the School of Computing at Blekinge Institute of Technology in partial fulfillment of the requirements for the degree of Master of Science in Computer Science. The thesis is equivalent to 20 weeks of full time studies. Contact Information: Authors: Shahrooz Abghari E-mail: shahroozabghari@gmail.com Samira Kazemi E-mail: kazemi.samira@gmail.com University advisors: Dr. Henric Johnson, School of Computing Blekinge Institute of Technology Dr. Niklas Lavesson, School of Computing Blekinge Institute of Technology School of Computing Blekinge Institute of Technology SE – 371 79 Karlskrona Sweden Internet : www.bth.se/com Phone : +46 455 38 50 00 Fax : +46 455 38 50 57 mailto:shahroozabghari@gmail.com� mailto:kazemi.samira@gmail.com� i Abstract Context Maritime Surveillance (MS) has received increased attention from a civilian perspective in recent years. Anomaly detection (AD) is one of the many techniques available for improving the safety and security in the MS domain. Maritime authorities utilize various confidential data sources for monitoring the maritime activities; however, a paradigm shift on the Internet has created new sources of data for MS. These newly identified data sources, which provide publicly accessible data, are the open data sources. Taking advantage of the open data sources in addition to the traditional sources of data in the AD process will increase the accuracy of the MS systems. Objectives The goal is to investigate the potential open data as a complementary resource for AD in the MS domain. To achieve this goal, the first step is to identify the applicable open data sources for AD. Then, a framework for AD based on the integration of open and closed data sources is proposed. Finally, according to the proposed framework, an AD system with the ability of using open data sources is developed and the accuracy of the system and the validity of its results are evaluated. Methods In order to measure the system accuracy, an experiment is performed by means of a two stage random sampling on the vessel traffic data and the number of true/false positive and negative alarms in the system is verified. To evaluate the validity of the system results, the system is used for a period of time by the subject matter experts from the Swedish Coastguard. The experts check the detected anomalies against the available data at the Coastguard in order to obtain the number of true and false alarms. Results The experimental outcomes indicate that the accuracy of the system is 99%. In addition, the Coastguard validation results show that among the evaluated anomalies, 64.47% are true alarms, 26.32% are false and 9.21% belong to the vessels that remain unchecked due to the lack of corresponding data in the Coastguard data sources. Conclusions This thesis concludes that using open data as a complementary resource for detecting anomalous behavior in the MS domain is not only feasible but also will improve the efficiency of the surveillance systems by increasing the accuracy and covering some unseen aspects of maritime activities. Keywords: open data, anomaly detection, maritime security, maritime domain awareness ii Acknowledgment First and foremost, we would like to thank our supervisors Dr. Niklas Lavesson and Dr. Henric Johnson for their patient guidance, encouragement and advice throughout this thesis. We would also like to acknowledge the invaluable support of the Swedish Coastguard in this thesis. In particular, we are grateful to Peter Ryman, the law enforcement officer at the Coastguard, without whose knowledge and kind support this study would not have been successful. Special thanks are due to Martin Boldt for all his help and support during the tough time of system installation. Last but not least, we would like to express our heartfelt thanks to our beloved families, for their understanding and endless love through the duration of our studies. iii TABLE OF CONTENTS 1. INTRODUCTION ....................................................................................................................... 1 1.1 PROBLEM STATEMENT ........................................................................................................... 2 1.2 RESEARCH QUESTIONS .......................................................................................................... 2 1.3 AIMS AND OBJECTIVES .......................................................................................................... 2 1.4 CONTRIBUTION ...................................................................................................................... 3 1.5 OUTLINE ................................................................................................................................ 3 2. BACKGROUND .......................................................................................................................... 3 2.1 TERMINOLOGY ....................................................................................................................... 4 2.2 JDL DATA FUSION MODEL .................................................................................................... 4 2.3 RELATED WORK .................................................................................................................... 5 3. RESEARCH METHODOLOGY ............................................................................................... 7 3.1 OPEN DATA IN THE MS DOMAIN ........................................................................................... 7 3.2 IDENTIFICATION OF MARITIME ANOMALIES .......................................................................... 8 3.3 FRAMEWORK DESIGN .......................................................................................................... 10 3.4 IMPLEMENTATION ................................................................................................................ 11 3.4.1 Data Description ............................................................................................................ 11 3.4.2 Data Collector Module ................................................................................................... 12 3.4.3 Database ......................................................................................................................... 13 3.4.4 Anomaly Detector Module .............................................................................................. 13 3.4.5 Display Client ................................................................................................................. 16 3.5 EXPERIMENTAL DESIGN ....................................................................................................... 17 3.6 VALIDATION DESIGN ........................................................................................................... 17 4. VALIDITY THREATS ............................................................................................................. 18 5. VERIFICATION ....................................................................................................................... 19 6. EXPERIMENTAL RESULTS .................................................................................................. 19 7. VALIDATION RESULTS ........................................................................................................ 22 8. DISCUSSION ............................................................................................................................. 25 9. CONCLUSION AND FUTURE WORK ................................................................................. 26 APPENDIX A: OPEN AND CLOSED DATA SOURCES ............................................................. 28 APPENDIX B: PORTS REGIONS ................................................................................................... 35 APPENDIX C: ALGORITHMS ....................................................................................................... 39 APPENDIX D: USER INTERFACE ................................................................................................ 48 APPENDIX E: ACRONYMS ........................................................................................................... 49 REFERENCES ................................................................................................................................... 50 1 1. INTRODUCTION MS is the effective understanding of all maritime activities that could impact the security, safety, economy or environment1 In addition to having a complete RMP, the way that the MS systems are used by human operators plays an important role in the efficiency of the surveillance operations. Monitoring vast sea areas and trying to establish Maritime Domain Awareness (MDA) for human operators is a difficult and time-consuming task (Riveiro, Falkman, & Ziemke, 2008a). This is due in part to the large amounts of heterogeneous data from multiple sources but also to the difficulties in detecting anomalous behavior from normal maritime activities. Therefore, having an automatic detector of unusual activities would help decision makers to efficiently monitor the ongoing activities in the surveillance area. . In recent years, the MS domain has received increased attention because of terrorism, smuggling activities and illegal immigration. An efficient MS system requires a complete Recognized Maritime Picture (RMP), which can be defined as a composite picture of maritime activities over an area of interest (Lefebvre & Helleur, 2001). For national maritime sovereignty, the RMP should include all activities within the 200 nautical miles Exclusive Economic Zone (EEZ). However, for some purposes such as the detection of illegal vessel transits, the RMP could extend beyond this region (Ponsford, D’Souza, & Kirubarajan, 2009). Using today’s technology, continuous tracking of all maritime activities by a single sensor data is not sufficient since it cannot monitor everything that happens in the surveillance area. On the other hand, there are large amounts of data in the MS domain that are gathered from a variety of sensors, databases and information systems. Therefore, by taking advantage of all the available data sources it would be possible to obtain a complete RMP. According to the Department of Homeland Security2 As well as the AD techniques, the use of different data sources will highly influence the detection of suspicious activities. Usually, only the data received from sensors are used for AD but there are a number of additional data sources regarding maritime activities that can be useful for this purpose. These sources of data consist of open and closed data about vessels, cargos, crew, etc. The closed data are only accessible to the maritime authorities, such as the Coastguard, but, by contrast, the open data are available online and freely accessible and reusable to the public. For instance, there are different organizations such as ports that publish their vessel traffic data or their facilities information online. In addition to the organizations, there are different online communities such as blogs, forums and social , AD is one of the enabling techniques for MDA. However, there are various AD techniques available and it is essential to choose the appropriate techniques that accomplish the MS goals. Data-driven AD approaches find the anomalous behavior by constructing a model from normal data and calculating the deviation from that model. However, relying only on data-driven approaches for surveillance systems is not sufficient due to the lack of user involvement in the detection process (Riveiro & Falkman, 2010). Furthermore, because of the diverse and complicated nature of the activities in the surveillance area, some of the suspicious behaviors are not directly observable. Therefore, finding all types of anomalies by using data-driven approaches seems to be impossible. On the other hand, maritime domain experts have the required knowledge and experience for finding maritime anomalies. Including the expert’s knowledge about suspicious activities in the detection process can result in improved AD. 1 Integrating Maritime Surveillance, common information sharing environment (cise). Retrieved from http://ec.europa.eu/maritimeaffairs/policy/integrated_maritime_surveillance/documents/integrating_maritime_su rveillance_en.pdf 2 National plan to achieve maritime domain awareness for the national strategy for maritime security. Retrieved from www.dhs.gov/xlibrary/assets/HSPD_MDAPlan.pdf http://ec.europa.eu/maritimeaffairs/policy/integrated_maritime_surveillance/documents/integrating_maritime_surveillance_en.pdf� http://ec.europa.eu/maritimeaffairs/policy/integrated_maritime_surveillance/documents/integrating_maritime_surveillance_en.pdf� http://www.dhs.gov/xlibrary/assets/HSPD_MDAPlan.pdf� 2 networks which provide the possibility of sharing information about maritime events. Some of the advantages of open data in addition to the availability for the public and free accessibility are: first, open data can reveal some facts that are not reported to the maritime authorities or available in their databases and second, open data can be used in the global context and are not suffered from legitimate limitations of exchanging data between different countries. By applying the open data to the detection process, the AD can be done more wisely and the results can have more facts of interests for the maritime experts. 1.1 Problem Statement In the maritime domain, there are different kinds of data sources that provide heterogeneous data regarding maritime activities. The majority of data, which are used by the MS systems, belong to the surveillance area of each country and are obtained from a variety of sensors and databases that are only accessible by the countries authorities. For detecting some of the anomalous activities such as smuggling, the maritime data beyond the surveillance area of each country are required. In order to assure security, maritime organizations in different countries need to exchange their privileged data and for this purpose they should deal with the diverse regulation of the data protection in each land. Exchanging data among countries is difficult, time-consuming and in some cases impossible because of the legislative issues. Moreover, there are activities that are neither reported to the maritime organizations, nor recorded in their data sources but they can be useful for the surveillance purpose. The publicly accessible and reusable data that are free from the legislative issues and revealing the unseen aspects of maritime activities are referred as open data. Consequently, employing the open data along with other confidential data sources would be beneficial for the MS systems to achieve their goals. 1.2 Research Questions Given the context of available data sources and MS systems at the Swedish Coastguard and the types of anomalies that the subject matter experts at the Coastguard are interested in, the first research question is: 1. How accurate and valid are the results of an AD system that exploits open data as a complement to the available closed data? In addition, by considering the accuracy as the degree to which the aforementioned AD system is able to distinguish between the normal and anomalous activities, and the validity as the degree to which the system results are true in real life, the next question is: 2. What is the performance difference between the system accuracy and the validity of results? 1.3 Aims and Objectives The aim is to investigate the potential open data as a complementary resource for AD in the MS domain. Objectives: • Identify existing open data sources in the maritime domain. • Identify those open data sources that are suitable for being used for AD in the MS domain. • Propose a framework for AD in the MS domain based on using the open data. • Develop an AD system based on the proposed framework and evaluate the accuracy of the system. • Validate the implemented AD system in real life. 3 1.4 Contribution This thesis contributes with a deeper understanding of open data as a complementary resource for effectively establishing the MS operations. It provides a framework for using the open data sources together with other sources of data for AD in the MS domain. According to the framework, an AD system is developed which employs a number of algorithms to implement the expert rules for detecting anomalies. The final contribution is the evaluation of the implemented AD system via the Coastguard validation in real life and also an experiment. 1.5 Outline The remainder of this work is organized as follows: Section 2 reviews the background and related work regarding the open data, AD and data fusion in the MS domain. Section 3 presents the research methodology. Validity threats and verification are described in sections 4 and 5, respectively. Section 6 presents the experiment results and the validation results are shown in section 7. Section 8 features a detailed discussion about the obtained results. Finally, section 9 concludes the research with a discussion on the possible directions for future work. 2. BACKGROUND The idea behind open data has been established for a long time. Open data can be used in a variety of domains and can be obtained from any resource. The two major sources of open data are the open data in science and the open data in government1 According to the estimation by Dedijer and Jéquier (1987), 90% of all information is open source, 9% is grey information (preprints of scientific articles, rumors in business circles, project proposals submitted to a research-funding agency, discussions with well- informed specialists, etc.), 0.9% is secret and 0.1% is non-existent information (i.e. the information you have, but you are not aware of it). Considering the large ratio of the open data sources, there should be a great value in using them in different domains. In the MS systems, the majority of the exploited data are obtained from the confidential sources. However, in recent years the new concept of the Web, which takes the network as a platform for information sharing, interoperability and collaboration, has created new sources of data for MS. There are organizations and communities that provide their maritime related data online and make them accessible for the public. Therefore, it would be beneficial for the MS . The longstanding concept of open data in science tries to overcome the difficulties in the current system of scientific publishing such as the inability to access data or usage limitation that is applied by the publishers or data providers (Molloy, 2011). Different groups, individuals and organizations are gathered to participate in a movement toward reforming the process of scientific publication (Molloy, 2011). One of the outcomes of the open data movement in science is the online availability of large number of scientific datasets for the public by different organizations. As well as the open data movement in science, governments for over a decade attempt to publish government data online and make it publicly accessible, readily available, understandable and usable (Alonso et al., 2009). Sharing the government data with the public provides openness and transparency with citizens. It can also improve the degree of participation in the society activities and the efficiency and effectiveness of the government services and the operations within and between the governments (Dietrich et al., 2009). 1 Open definition. Retrieved from opendefinition.org/okd/ http://opendefinition.org/okd/� 4 systems if they can take advantage of the open data to increase the safety and security in their surveillance area. 2.1 Terminology According to the Department of Homeland Security1 Data fusion involves the process of combining data from multiple sources or sensors and making inferences that may not be possible from a single source or sensor (Hall & McMullen, 2004). , MDA would be achieved by monitoring the maritime activities, fusing and analyzing the data in a way that normal activities can be identified and anomalies differentiated. Therefore, AD and data fusion techniques are important technologies for MDA. AD is widely used in the areas such as video surveillance, network security and military surveillance. Chandola, Banerjee and Kumar (2009) define AD as: The problem of finding patterns in data that do not conform to expected behavior. Depending on the domain of study, the non-conforming patterns are called by different names such as anomalies, outliers, exceptions, etc. In the MS domain, these non-conforming patterns are referred as anomalies. Defense R&D Canada (Roy, 2008) provides the following definition for the term anomaly in the context of the MS domain: Something peculiar (odd, curious, weird, bizarre, atypical) because it is inconsistent with or deviating from what is usual, normal, or expected, or because it is not conforming to rules, laws or customs. The term Open Data refers to the idea of making data freely available to use, reuse or redistribute without any restriction. The open data movement follows the other open movements such as Open Access and Open Source. According to the Open Knowledge Foundation2 2.2 JDL Data Fusion Model , a community based organization that promote open knowledge (whether it is content, data or information-based), an open work should be available as a whole, with a reasonable reproduction cost, preferably downloading via the Internet without any charge and in a convenient and modifiable form. Furthermore, it should be possible to modify and distribute the work without any discrimination against persons, groups, fields or endeavor. In the scope of this thesis, the open data term refers to the publicly available data that may or may not require free registration. One of the most widely used data fusion models in the literature is the JDL data fusion model. It was developed by the US Joint Directors of Laboratories (JDL) data fusion sub- panel in 1985. A current version of the model consists of six different levels. Table 1 shows the description of these six levels (Hall & McMullen, 2004). In this work the main focus is on AD which is done in the Level 2 (situation assessment) of the JDL model. The input of this level is the identified entities and their related information regarding each other or the environment and the output would be the assessment of a situation as normal or anomalous (Brax, Niklasson, & Smedberg, 2008). 1 National plan to achieve maritime domain awareness for the national strategy for maritime security. Retrieved from www.dhs.gov/xlibrary/assets/HSPD_MDAPlan.pdf 2 Open definition. Retrieved from opendefinition.org/okd/ http://www.dhs.gov/xlibrary/assets/HSPD_MDAPlan.pdf� http://opendefinition.org/okd/� 5 Table 1 Summary of the JDL Data Fusion Model Components JDL model component Description Level 0 processing (Source preprocessing) At this level, preprocessing of data from sensors and databases would be done by means of image processing, signal processing and conditioning, unit conversions, bias corrections or feature extractions, etc. Level 1 processing (Object refinement) This level focuses on combining data from sensors and databases in order to obtain the most accurate and reliable estimates of an entity’s position, movement, attributes, characteristics and identity. Level 2 processing (Situation assessment) According to the obtained result of previous level, a description of current relationships among entities and their relationship to the environment would be developed in order to determine the interpretation of the situation. Level 3 processing (Impact assessment) This level focuses on the estimation and prediction of alternative futures and hypotheses concerning the current situation to determine the potential impacts or threats. Level 4 processing (Process refinement) This level is a meta-process that monitors the whole data fusion process to optimize the utilization of data sources and algorithms and improve the performance of the ongoing data fusion. Level 5 processing (Cognitive refinement) Level 5 focuses on transforming the result of data fusion in to the displays and understandable information for the user and improvement of human/computer effectiveness. 2.3 Related Work Data fusion techniques have been used for a long time in the MS domain. The majority of studies focused on target tracking, tactical situation awareness and threat assessment (Akselrod, Tharmarasa, Kirubarajan, Zhen Ding, & Ponsford, 2009; Bick & Barock, 2005; Danu, Sinha, Kirubarajan, Farooq, & Brookes, 2007; Di Lallo et al., 2006; Gad, 2009; Giompapa, Farina, Gini, Graziano, & Di Stefano, 2007; Giompapa et al., 2008; Hatch, Kaina, Mahler, & Myre, 1998; Henrich, Kausch, & Opitz, 2004; Jouan, Valin, Gagnon, & Bosse, 1999; Lefebvre & Helleur, 2001; Maresca et al., 2010; Vespe, Sciotti, & Battistello, 2008). Typically, in these works the combination of two or more sensors such as: Automatic Identification System (AIS), Infrared (IR), video camera, Synthetic Aperture Radar (SAR), Vessel Traffic Service (VTS) radar, Over The Horizon (OTH) radar, High Frequency Surface Wave Radar (HFSWR) and microwave radar is used and the surveillance area is limited to the coastal regions. In recent years, the number of studies that address the use of AD in the MS domain is increasingly growing. AD techniques are divided into two groups, namely data-driven and knowledge-driven approaches. There are a couple of works that proposed knowledge-based systems with different representation techniques and reasoning paradigms such as rule- based, description logic and case-based reasoning (Guyard, Roy, & Defence R&D Canada- Valcartier, 2009; Nilsson, van Laere, Ziemke, & Edlund, 2008; Roy & Davenport, 2010). A prototype for a rule-based expert system based on the maritime domain ontologies was developed (Edlund, Gronkvist, Lingvall, & Sviestins, 2006) that could detect some of the anomalies regarding the spatial and kinematic relation between objects such as simple scenarios for hijacking, piloting and smuggling. Another rule-based prototype was developed 6 by Defense R&D Canada (Roy, 2008, 2010). The aforementioned prototype employed various maritime situational facts about both the kinematic and static data in the domain to make a rule-based automated reasoning engine for finding anomalies. One of the popular data-driven AD approaches is the Bayesian network (Fooladvandi, Brax, Gustavsson, & Fredin, 2009; Johansson & Falkman, 2007; Lane, Nevell, Hayward, & Beaney, 2010). Johansson and Falkman (2007) used the kinematic data for creating the network; however, in the work that was done by Fooladvandi et al. (2009) expert’s knowledge as well as the kinematic data was utilized in the detection process. Moreover, Lane et al. (2010) presented five unusual vessel behaviors and the way of formulating them in an AD system that the estimation of the overall threat was performed by using a Bayesian network. Unsupervised learning techniques have been widely used for data-driven AD such as Trajectory Clustering (Dahlbom & Niklasson, 2007), Self Organizing Map (Riveiro, Johansson, Falkman, & Ziemke, 2008) and fuzzy ARTMAP neural network (Rhodes, Bomberger, Seibert, & Waxmamn, 2005). Some statistical approaches, such as Gaussian mixture model (Laxhammar, 2008), hidden Markov model (Andersson & Johansson, 2010), adaptive kernel density estimator (Ristic, La Scala, Morelande, & Gordon, 2008) and precise/imprecise state- based anomaly detection (Dahlbom & Niklasson, 2007) have been used in this context. The majority of the works that have been done in the context of AD only used the AIS data. There are a number of studies that employed data fusion techniques to fuse data from different sensors in AD systems (Carthel et al., 2007; Guerriero, Willett, Coraluppi, & Carthel, 2008; Rhodes, Bomberger, Seibert, & Waxman, 2006; Vespe, Sciotti, Burro, Battistello, & Sorge, 2008). In these studies, the surveillance area was restricted to the coastal regions and the combination of data from AIS, SAR, IR, video and radar was used in the fusion process to obtain the vessel tracks. Furthermore, there are some other works that focused on the fusion of both sensor and non-sensor data (Andler et al., 2009; Ding, Kannappan, Benameur, Kirubarajan, & Farooq, 2003; Fooladvandi et al., 2009; Lefebvre & Helleur, 2004; Mano, Georgé, & Gleizes, 2010; Riveiro & Falkman, 2009). For example, Lefebvre and Helleur (2004) and Riveiro and Falkman (2009) treated the expert’s knowledge as the non-sensor data. Riveiro and Falkman (2009) introduced a normal model of vessel behavior based on AIS data by using self organizing map and a Gaussian mixture model. According to the model, the expert’s knowledge about the common characteristic of the maritime traffic was captured as IF-THEN rules and the AD procedure was supposed to find the deviation from the expected value in the data. Lefebvre and Helleur (2004) fused radar data with user’s knowledge about the vessels of interests. The sensor data were modeled as track and the non-sensor data were modeled as templates. The track-template association was done by defining mathematical models for tracks and using fuzzy membership functions for association possibilities. Mano et al. (2010) proposed a prototype for the MS system that could collect data from different types of sensors and databases and regroup them for each vessel. Sensors like AIS, HFSWR and classical radars and databases such as environmental database, Lloyd’s Insurance and TF2000 Vessel DB were included in this prototype. By using multi-agent technology an agent was assigned to each vessel and anomalies could be detected by employing a rule-based inference engine. When the combination of anomalies exceeded a threshold, vessel status was informed to the user as an anomaly. The work presented by Ding, Kannappan, Benameur, Kirubarajan and Farooq (2003), proposed the architecture of a centralized integrated maritime surveillance system for the Canadian coasts. Sensors and databases included in this architecture were: HFSWR, ADS (Automatic Dependant Surveillance) reports, visual reports, information sources, microwave radar and radar sat. A common data structure was defined for storing data that were collected from different sensors. Andler et al. (2009) also described a conceptual MS system that integrated all available information such as databases and sensor systems (AIS, LRIT, intelligence reports, registers/databases of vessels, harbors, and crews) to help user to detect and visualize anomalies in the vessel traffic data in a worldwide scale. Furthermore, the authors suggested using open data in addition to other resources in the fusion process. 7 In conclusion, the main focus of the projects that have been done in the context of AD in the MS domain was related to using sensors data and mainly the AIS data to find the anomalies in the coastal region. Detection of some suspicious activities such as smuggling requires vessel traffic data beyond the coastal region. Maritime authorities in each country have overall information of maritime activities in their surveillance area. But exchanging information among different countries is a complicated procedure because of the diverse regulation of data protection in each land. Therefore, using data sources that are free from legislative procedures can be a good solution for providing information that belongs to the regions outside the land territory. Furthermore, all the information about maritime activities is not recorded in the authorities’ databases or reported to them. On the other hand, there are numerous open data sources consists of different websites, blogs and social networks that can be useful for observing the hidden aspects of maritime activities. Hence, this thesis will investigate the potential open data sources for maritime activities and exploit them to build an AD system. 3. RESEARCH METHODOLOGY This thesis starts by investigating the applicable open data sources for AD in the MS domain. Then, potential maritime anomalies that can be detected by the obtained open data sources and the way that the open data can be applied to the AD process are identified. The next step is to design and implement an anomaly detector system that can use open data in the detection process. For this purpose, a general framework for AD based on using both open and closed data sources in the MS domain is proposed. Then, an anomaly detector system is developed according to the proposed framework. In order to understand to what extent the system results are valid, the system is evaluated in real life by the subject matter experts from the Swedish Coastguard. However, the system accuracy should be evaluated before using the system in real life. Therefore, an experiment is conducted to measure the system accuracy. The following sections describe these steps in detail. 3.1 Open Data in the MS Domain Applicable data sources for AD in the MS domain can be divided into three categories. The first and main category consists of sensors. Sensors provide kinematic data for each object in their coverage area and can be categorized as passive and active. Active sensors do not require cooperation from objects and collect the data by active probing the environment such as radar and sonar (A. M. Ponsford et al., 2009). However, passive sensors rely on the data that are broadcasted by objects intentionally such as AIS or unintentionally such as Electronic Intelligence (ELINT) and SIGnal INTelligence (SIGINT) systems (A. M. Ponsford et al., 2009). More information about the main maritime sensors can be found in (İnce, Topuz, & Panayirci, 1999; Vespe, Sciotti, & Battistello, 2008). The second category of data sources includes the authorized databases which contain information about vessels, cargos, crew, etc1 1 Integrated maritime policy for the EU : Working document III on maritime surveillance systems. Retrieved from . The first and second categories are only accessible to the maritime authorities such as the Coastguard and can be referred as closed data sources. The third category belongs to the open data sources which are publicly available via the Internet and are free to access or reuse. These data sources consists of vessels traffic data and reports or news that are related to the maritime domain and can be found in different blogs, websites or social networks. epub.sub.uni-hamburg.de/epub/volltexte/2009/1750/pdf/maritime_surveillance_en.pdf http://epub.sub.uni-hamburg.de/epub/volltexte/2009/1750/pdf/maritime_surveillance_en.pdf� 8 To obtain the applicable open data for AD, the first step is started by looking through the information resources document1 3.2 Identification of Maritime Anomalies provided by the International Maritime Organization (IMO). IMO is the United Nations specialized agency with responsibility for the safety and security of shipping and the prevention of marine pollution by vessels. The document introduces 29 governmental and intergovernmental organizations that work in different fields related to the MS domain such as maritime safety, prevention of pollution from vessels, liability and insurance issues, shipping information, etc. All these 29 organizations’ websites and the links provided by each of them are investigated and a list of online data sources is prepared. The prepared list contains data sources that have information that can be qualified for the MS purpose such as information related to AIS, vessel characteristics, ports, maritime companies, suppliers’ information, weather, etc. Table A.1 of Appendix A, presents the obtained data sources. These sources of data are available online but having access to some of them needs non-free registration. Moreover, in the process of finding open data sources, it is attempted to obtain sources of data that are related to the Baltic region and mostly Sweden by use of the previously observed data sources and also Google search engine. An important aspect of the literature review is to find out more about the potential maritime anomalies. The main sources of information about maritime anomalies are reports of the two workshops that were held in Canada (Roy, 2008) and Sweden (Andler et al., 2009; van Laere & Nilsson, 2009). In these two workshops attendees were experts in the maritime domain and a variety of maritime anomalies were identified. The outcome of the Swedish workshop was the identification of the 31 most desired anomalous behaviors for the Swedish stakeholders. The identified anomalies belong to the following categories: tampering, owner/crew, history, rendezvous (object, location), movement and cargo. The outcome of the Canadian workshop was a taxonomy of maritime anomalies that categorized anomalies as static and dynamic. Static anomalies were related to the vessel characteristics such as name, IMO number, etc. Dynamic anomalies were divided into two groups, kinematic and non- kinematic anomalies. The kinematic anomalies included anomalies related to location, course, speed, reporting and maneuver of vessels. The non-kinematic anomalies were related to passengers, crew list, cargo list, last port of call and next port of call. According to the identified anomalies by the two workshops, a list of some potential maritime anomalies that can be detected by use of the available open data sources is prepared. Then, in a meeting with representatives of the Swedish Coastguard the types of anomalies that are interesting for them and the possibility of using open data for AD are discussed. During the meeting the prepared list of anomalies is presented and they are asked about the possibility of occurrence and their degree of interest for each anomaly. As an outcome of the meeting a number of scenarios are created and based on them 11 rules are defined. The first scenario refers to the anomalies related to the vessel static information such as name, owner, IMO number, dimensions, type and the status (in service or laid up). For example, sailing a vessel with a draught of 22 meters over an area with a 9 meters depth or observing a vessel that should be laid up or changing the name or the owner of a vessel during its voyage indicate the existence of suspicious activities. The second scenario is related to the prior arrival notification for vessels. Vessels should inform their arrival time to the ports at least 24 hours in advance. Each port also provides an online timetable for the incoming vessels. Therefore, any mismatch between the reported AIS data regarding the destination or the arrival time of a vessel and the destination port timetable needs to be checked by coastguards. The third scenario is related to ordering pilots. Usually, large vessels because of their size and weight need to be guided by pilots through dangerous and congested waters. Therefore, vessels need to submit their request for a pilot and also inform 1 Information resources on maritime security and ISPS code. Retrieved from www.imo.org/knowledgecentre/informationresourcesoncurrenttopics/maritimesecurityandispscode/documents/inf ormation%20resources%20on%20maritime%20security%20and%20isps%20code.pdf http://www.imo.org/knowledgecentre/informationresourcesoncurrenttopics/maritimesecurityandispscode/documents/information%20resources%20on%20maritime%20security%20and%20isps%20code.pdf� http://www.imo.org/knowledgecentre/informationresourcesoncurrenttopics/maritimesecurityandispscode/documents/information%20resources%20on%20maritime%20security%20and%20isps%20code.pdf� 9 the destination port. However, in some cases vessels order a pilot without informing the port. Such situations should be investigated. In the next meeting, the scenarios and the rules are presented to the representatives of the Swedish Coastguard and they are asked to comment or suggest new scenarios or rules. By getting the final approval from the Coastguard experts, one new rule (rule number 5) is added to the list. Table 2 shows the admitted rules by the experts. These identified maritime anomalies can be detected by use of AIS data, vessel traffic timetables in ports and pilots websites and the vessel characteristic data that are available in data sources such as Lloyd’s. A name is given to each anomaly that can be detected by the rules, and for the rest of this thesis anomalies will be referred by their names. Table 2 The identified anomalies that can be detected by open data (confirmed by the Swedish Coastguard) No. Expert rules Anomaly 1 If a vessel destination does not exist in the port schedule then anomaly. VESSEL_NOT_INFORMED_PORT (A1) 2 If a vessel ETA does not match with the port ETA for the vessel then anomaly. ARRIVAL_TIME_MISMATCHED (A2) 3 If a vessel entered a port without informing the port then anomaly. VESSEL_ENTERED_PORT_WITHOUT_ NOTICE (A3) 4 If a vessel has requested a pilot but has not used the service then anomaly. VESSEL_NOT_USED_PILOT (A4) 5 If vessel A which normally travels between ports X and Y, suddenly goes to port Z then anomaly. UNUSUAL_TRIP_PATTERN (A5) 6 If a vessel has not left a port according to the port schedule then anomaly. VESSEL_NOT_LEFT_PORT (A6) 7 If a vessel exists in a port schedule but it has not entered the port then anomaly. VESSEL_NOT_ENTERED_PORT (A7) 8 If a vessel does not exist in the port schedule and the vessel has requested a pilot then anomaly. VESSEL_ORDERED_PILOT_AND_ NOT_INFORMED_PORT (A8) 9 If a vessel has moored in a port and has been observed somewhere else then anomaly. VESSEL_MOORED_IN_PORT (A9) 10 If vessel A has not entered a port according to the port schedule instead vessel B enters the port at the same time slot then anomaly. WRONG_VESSEL_ENTERED (A10) 11 If a vessel with the laid up status has been observed somewhere else then anomaly. VESSEL_LAID_UP (A11) Note. ETA = estimated time of arrival 10 3.3 Framework Design The Open Data Anomaly Detection System (ODADS) is designed for traffic monitoring and detecting anomalies in the MS domain by using open and closed data sources. Figure 1 depicts the ODADS architecture. ODADS consists of three modules: 1) Data Collector, 2) Anomaly Detector and 3) Display Client. The Data Collector module is responsible for collecting open data from the Internet, preprocessing and storing the data in a database. The data can be related to vessel traffic (such as AIS reports, ports and pilots timetables), vessel characteristics, ports equipments and facilities, companies that are involved in maritime activities, news or reports about maritime events and activities available in different social media platforms (such as blogs and social networks), etc. The Data Store comprises a set of databases that contain data belong to different types of sensors, authorized databases and open data sources. The data in the Data Store can be fused or integrated before being used in the detection process. When the Data Collector completes its task, Anomaly Detector starts to work. The Anomaly Detector module analyzes the available data (open and closed data) and detects possible anomalies by utilizing both knowledge-driven and data-driven techniques. Different AD techniques are employed due to the distinct nature of anomalies and the complexity of the environment in the MS domain. Previously known anomalies can be detected by knowledge based techniques such as rule-based, but in real life an AD system must be able to detect the unseen anomalies, too. This is one of the benefits of using data- driven methods such as machine learning techniques. Therefore, detecting different types of anomalies seems to be possible by exploiting different techniques. The Display Client module is the user interface of the system. This module represents the cognitive refinement level (Level 5) of the JDL model. M. J. Hall, Hall and Tate (2000), argued that the effectiveness of a system can be affected by the way that the system produced information is comprehended by the human user. The cognitive refinement process involves traditional Human-Computer Interaction (HCI) utilities such as geographical display or advanced methods that support functionalities such as cognitive aids, negative reasoning enhancement, focus/defocus of attention and representing uncertainty. Section 3.4 describes the implementation details for each module. Figure 1. The Open Data Anomaly Detection System (ODADS) architecture. The Data Collector module collects data from the Internet and stores them in the database. The Anomaly Detector module detects anomalies by taking advantage of different techniques. The Display Client module displays the detected anomalies to the user and enables system- user interaction. Display Client Data Collector Anomaly Detector Knowledge-Driven Approaches Data-Driven Approaches Internet .... Weather Information Vessel Traffic Vessels, ports and companies information Social Networks Data Store Open data Closed data 11 3.4 Implementation ODADS is implemented by taking advantage of the identified maritime anomalies and the obtained open data sources that were discussed in the previous sections. To limit the scope, first of all only four types of vessels: passenger, ferry, cargo and tanker are considered and other types of vessels such as fishing and sailing vessels, which ports usually do not provide any information about them, are omitted. Secondly, the rule related to the vessel static information is ignored. Furthermore, the WRONG_VESSEL_ENTERED anomaly is excluded due to its complexity. As well as the A1-A9 anomalies, in further collaboration with the Coastguard representatives during the implementation phase, a new type of anomaly is proposed. This anomaly is called UNDER_SURVEILLANCE_VESSEL and occurs when a vessel of interest has any of the A1-A9 anomalies and the vessel exists in the vessels blacklist. 3.4.1 Data Description The required vessel traffic data can be obtained from AIS reports and ports and pilots timetables. The surveillance area is restricted to the north of the Baltic Sea and a part of the Gulf of Finland, the regional area between three European countries Sweden, Finland and Estonia. Figure 2 shows the surveillance area where the geographic coordinates lie between latitudes 58.49º - 60.24º N and longitudes16.19º – 25.00º E. This region is one of the high- traffic regions in the Baltic Sea and is surrounded by the four highly used ports. More information regarding each port can be found in Appendix B. Figure 2. The area of interest is restricted to the north of the Baltic Sea and a part of the Gulf of Finland. Ports from left to right are Norrköping, Stockholm group (Nynäshamn, Stockholm and Kapellskär), Helsinki and Tallinn (The image is adapted from Google Earth). 3.4.1.1 AIS data Due to inaccessibility to the raw AIS data in this thesis, the AIS reports that are provided by the MarineTraffic.com website1 1 are exploited. These reports consist of both static and dynamic types of data for each vessel during its voyage. Vessel static data include name, type, built year, size, draught, flag, call sign, Maritime Mobile Service Identity (MMSI), IMO identification number, origin (last known port), destination, Estimated Time of Arrival (ETA) and dynamic data are related to speed (max and average), position (longitude and latitude), Course Over Ground (COG), heading. www.marinetraffic2.aegean.gr/ais/getkml.aspx http://www.marinetraffic2.aegean.gr/ais/getkml.aspx� 12 Data Collector AIS Logger Web Scraper Database AIS Reports Table Port and Pilot Schedule Table 3.4.1.2 Ports and Pilots data Ports timetables are obtained from the websites of the high-traffic ports in the area of interest: Stockholm group (Stockholm, Kapellskär and Nynäshamn) and Norrkoping ports in Sweden, Helsinki port in Finland and Tallinn port in Estonia. The pilots timetables belong to the vessel traffic in Stockholm pilotage area in Sweden. Table A.2 of Appendix A, presents each data source in detail. 3.4.2 Data Collector Module The Data Collector module is responsible for extracting data from different open data sources via the Internet, data preprocessing and data storage. This module consists of two parts. The first part is the AIS Logger, a shell script program, which downloads and extracts the processed vessel AIS reports and tracks1 Figure 3. The Data collector module consists of two parts, AIS Logger that provides a list of vessels in the area of interest and the Web Scraper, a program which gathers vessels details information and ports and pilots timetables from the Internet. data from MarineTraffic.com website. Since the MarineTraffic.com website updates the data at 10-minute intervals, data extraction is done every 10 minutes and the collected data are stored in a file. When the file is prepared, the second part of the module, the Web Scraper, starts to work. It is a Java program that harvests data from different websites and stores them into the database. These data include ports and pilots timetables from their websites and the vessels details information from the Marinetraffic.com. Figure 3 illustrates the different parts of the Data Collector module. 3.4.2.1 Preprocessing Data preprocessing can be performed by a number of techniques such as data cleaning to remove noise, data integration to merge multiple data sources into an understandable data, data transformation such as normalization to improve the accuracy and data reduction to eliminate the redundant features (Han et al., 2011). In the first step of preprocessing, text values such as vessel name, origin, destination and company name are transformed to the same format (lowercase characters) and all special characters and whitespaces are removed. Since the data are collected from different sources in different countries, time values are converted to Central European Time (CET). Every data source has its own format of data representation. For instance, vessel arrival time in ports can be stored as two separated parts of date and time, a combination of date and time values or just as date. Moreover, some of the ports provide additional vessel information such as weight and length as well as the timetables. This diversity makes some parts of the provided data remain unused. Therefore, a common data representation format for ports and pilots data is defined that contains vessel name, vessel type, origin, destination, company name, vessel status and arrival/departure time. 1 www.marinetraffic.com/ais/gettrackxml.aspx?mmsi=xxxxxxxx http://www.marinetraffic.com/ais/gettrackxml.aspx?mmsi=xxxxxxxx� 13 Database AIS Reports Table Port and Pilot Schedule Table History Anomaly Table Anomaly Detector Data-Driven Approaches Statistical Techniques Knowledge-Driven Approaches Expert Rules 3.4.3 Database The Database contains four tables to store the extracted open data: 1) AIS Reports table, 2) Ports and Pilots Schedule table, 3) Vessel Trip History table and 4) Detected Anomaly table. There is also a table for storing vessels that are kept under surveillance because of their previous involvement in the criminal activities. The under surveillance vessels data are provided by the Coastguard officer via the Display Client module and stored in the Blacklist Table. The AIS Reports table contains the processed AIS data such as vessel name, type, flag, origin, destination, etc. The Ports and Pilots Schedule table is used for storing the vessel data including arrival and departure time gathered from ports and pilots data sources. In the Vessel Trip History table, data related to the frequency of vessels trips between different ports are stored and the Detected Anomaly table contains the history of all types of detected anomalies. 3.4.4 Anomaly Detector Module The Anomaly Detector module is the main part of the system which employs different techniques to detect anomalies. After investigating the nature of the anomalies and the potential techniques, it is determined that except for the UNUSUAL_TRIP_PATTERN anomaly, the other types can be detected by exploiting search techniques. Detection of the UNUSUAL_TRIP_PATTERN anomaly requires data-driven approaches such as machine learning or statistical techniques and also a history of the vessel traffic data for training the system. In case of using machine learning techniques, the system will detect the anomaly by learning the normal vessel trip pattern and highlighting the unusual trip and by using statistical approaches the system will make decision based on the frequency or the possibility of a vessel trip between two different ports. The next section describes the detection algorithm for each anomaly in detail. This module is depicted in Figure 4. Figure 4. The Anomaly Detector module 3.4.4.1 Algorithms To define the appropriate detection algorithms, the first step is to specify what types of data are required for detecting each anomaly (Table A.3 of Appendix A). Detection of each individual anomaly (except the UNUSUAL_TRIP_PATTERN anomaly) can be done by performing a search in the specified data for finding the desired match. If the match is not found then the vessel would be marked as an anomaly. Appendix C gives a detailed description about each algorithm. Algorithm 1 shows the main procedure of the Anomaly Detector module. Algorithms 2-4 and 6-8 present the described search process in detail. For the UNUSUAL_TRIP_PATTERN anomaly, data related to six months (September 15, 2011- March 15, 2012) of vessel traffic in the surveillance area are gathered. By monitoring the 14 activities during this period, the system will be able to find the normal pattern of vessels trips in the area of interest. For detecting this anomaly a simple statistical approach is used. A look up table is created and for each vessel the number of times that the vessel travels between two different ports is stored. For each vessel, if the frequency of travelling between its origin and destination is less than a predefined threshold then the vessel will be reported as an anomaly. Algorithms 5 and 16 describe the process in detail. After executing all the algorithms, the final type of anomaly for each vessel should be determined. There are some situations that multiple anomalies can occur in the same time for a specific vessel. These situations are presented in Table 5. For the combination of anomalies that do not have any feature in common new types of anomalies are defined. Table 5 The combined anomalies Anomaly Condition Anomaly Type A1,A5 Vessel did not inform its arrival to the destination port and its trip to the port is not common. UNUSUAL_TRIP_PATTERN_AND_ VESSEL_NOT_INFORMED_PORT A1,A8 Vessel did not inform its arrival to the destination port and it ordered a pilot. VESSEL_ORDERED_PILOT_AND_ NOT_INFORMED_PORT A1,A5,A8 Vessel did not inform its arrival to the destination port, it ordered a pilot and its trip to the port is not common. UNUSUAL_TRIP_PATTERN_AND_ VESSEL_ORDERED_PILOT_AND_ NOT_INFORMED_PORT_ INFORMED_PORT A2,A4 Vessel has delay and it ordered a pilot but did not use it. VESSEL_ARRIVAL_TIME_ MISMATCHED_AND_VESSEL_NOT_ USED_PILOT A2,A5 Vessel has delay and its trip to the port is not common. UNUSUAL_TRIP_PATTERN_AND_ VESSEL_ARRIVAL_TIME_ MISMATCHED A2,A4,A5 Vessel has delay and it ordered a pilot but did not use it and its trip to the port is not common. UNUSUAL_TRIP_PATTERN_AND_ VESSEL_NOT_USED_PILOT_AND_ VESSEL_ARRIVAL_TIME_ MISMATCHED A3,A6 Vessel entered a port without prior notification and has not left the port according to the plan. VESSEL_ENTERED_PORT_ WITHOUT_NOTICE_AND_NOT_ LEFT_PORT_ON_TIME A4,A5 Vessel ordered a pilot and informed to the port however it did not used the pilot and its trip to the port is not common. UNUSUAL_TRIP_PATTERN_AND_ VESSEL_NOT_USED_PILOT A4,A7 Vessel ordered a pilot but did not use it and it has not entered the port. VESSEL_NOT_ENTERED_PORT_ AND_NOT_USED_PILOT A4,A8 Vessel ordered a pilot but did not inform the port and it did not use the pilot. VESSEL_ORDERED_PILOT_AND_ NOT_INFORMED_PORT_AND_ VESSEL_NOT_USED_PILOT (Continued) 15 Anomaly Condition Anomaly Type A5,A7 Vessel trip to the destination port is not common and it has not entered the port. UNUSUAL_TRIP_PATTERN_AND_ VESSEL_NOT_ENTERED_PORT A7,A4,A5 Vessel ordered a pilot but did not use it and has an unusual trip and it has not entered the port. UNUSUAL_TRIP_PATTERN_AND_ VESSEL_NOT_ENTERED_PORT_ AND_VESSEL_NOT_USED_PILOT A8,A5 Vessel ordered a pilot but did not inform the port and the trip to the destination port is not common. UNUSUAL_TRIP_PATTERN_AND_ VESSEL_ORDERED_PILOT_AND_ NOT_INFORMED_PORT A9,[A1-A8] Vessel uses the information that belongs to a moored vessel in a port; since the vessel provides unreal information, it is normal that other anomalies are set for this vessel. VESSEL_MOORED_IN_PORT A8,A4,A5 Vessel ordered a pilot but did not inform the port and the trip to the destination port is not common and the vessel also did not use the pilot. UNUSUAL_TRIP_PATTERN_AND_ VESSEL_ORDERED_PILOT_AND_ NOT_INFORMED_PORT_AND_ VESSEL_NOT_USED_PILOT 3.4.4.2 String matching techniques Using exact string matching techniques for comparing vessel information from different data sources is inapplicable due to the potential errors that might occur because of different notations or even human operator mistakes during data entry. Therefore, a metric should be used for measuring the degree of similarity between two vessels from different sources. For this purpose, some string comparison methods are evaluated such as Damerau-Levenshtein (D-L) also known as edit distance (Damerau, 1964; Levenshtein, 1966), the n-gram technique (P. A. V. Hall & Dowling, 1980), Jaro (Jaro, 1972, 1989) and JaroWinkler (Winkler, 1990). D-L or edit distance is equal to the minimum number of edits (substitution, delete, insert and transposition) required to change one string to the other. The use of n-gram for string matching is performed by first constructing the n-grams (contiguous sequences of n items) for each string and then comparing the n-grams of both strings in order to find the number of common n-grams. Jaro and JaroWinkler (a variant of Jaro) measure the number and order of common characters in two strings and also the number of transposition that is needed to change one of the strings to the other. Jaro metrics are more similar to the human decision making compared to the D-L distance (Denk & Hackl, 2003). To evaluate these metrics, the D-L distance is implemented and for the other metrics SimMetrics1 The collected data related to one day of vessel traffic in the area of interest are taken with distinct vessels names, origins and destinations and stored in separate groups. Then, the string matching methods are applied to each group by considering all pair wise combination of the elements in each group. To evaluate the results, the obtained similarities are sorted in ascending order and it is verified whether or not the values in each pair can be considered similar by a human operator. Two values are considered equal if they are exactly the same or a bit different because of misspellings or shortening. Finally, based on the results it is , an open source library that contains a set of similarity metrics, is used. Regarding the n-gram technique, choosing the length of n-grams is an important factor since it can affect the processing time and accuracy of the method. According to Salton and McGill (1986) and Zamora, Pollock, and Zamora (1981), trigram is one of the n-gram methods (when n=3) that can achieve the best results in retrieving similar words. 1 www.aktors.org/technologies/simmetrics/index.html http://www.aktors.org/technologies/simmetrics/index.html� 16 concluded that for the available data, JaroWinkler has the best performance among all the other metrics and its results are closer to reality. Jaro and D-L are the second and the third best respectively and trigram is the worst. According to the degrees of similarity that are obtained from the result of JaroWinkler and the data characteristics, two similarity thresholds, one for the vessel name and the other for the origin/destination are defined. The vessel name similarity threshold is considered as 0.03 and value for the trip threshold is 0.06 (Given p as the similarity value, 0 ≤ p ≤ 1, p= 0 shows the exact match and p=1 indicates that the two strings are dissimilar). These thresholds are referred as VESSEL_NAME_SIMILARITY_THRESHOLD and TRIP_ SIMILARITY_THRESHOLD in Appendix C. 3.4.5 Display Client The Display Client module is the graphical user interface of ODADS. It is a web-based application and supports the functionalities of the JDL Level 5 processing such as HCI utilities, cognitive aids and focus/defocus attention. While designing the user interface, the six principles of the user interface design that are based on the usage-centered design approach are considered. According to Constantine and Lockwood (1999), these principles are: structure, simplicity, visibility, feedback, tolerance, and reuse. Figure D.1 of Appendix D presents a snapshot of the user interface. When a user logs in to the system, the main view of the ODADS user interface is displayed. It consists of a geographical display and four tabs, namely Anomaly, Dataset, Reports and Setting. The geographical display utilizes the Google maps API features to enable the vessel traffic monitoring in the area of interest. Vessels are represented on the map by different colors according to their types. Ferries and passenger vessels are green, cargo vessels are cyan and tankers are yellow. In order to give a better perspective of the surveillance area, tug and pilot vessels are also displayed on the map. The tug and pilot vessels are presented in gray and dark blue, respectively. By clicking on each vessel, an information window appears which contains the vessel information such as name, MMSI number, IMO number, flag, heading, origin, destination, EAT and anomaly type. The AD process is performed every 10 minutes and consequently, the map is updated and if any new anomaly is detected, the Anomaly tab becomes red in order to inform the user about the incoming suspicious activities. The newly detected anomalous vessels are displayed in red and the previously detected anomalies are shown in light red. The Anomaly tab presents the overall information about the number of detected anomalies, total number of vessels and the last data collection time. Moreover, anomalous vessels are shown in a table which has the search, pagination and zoom to map utilities. The Dataset tab shows the ports and pilots timetables within a time interval and in a common representation format. This common data representation format contains vessel information such as name, type, company name, origin, destination, time, status and the name of the data source. The Reports tab consists of two sections. The first section provides reports regarding the total number of each anomaly and detailed information about anomalous vessels in a time interval in the HTML or EXCEL formats. The second section presents graphical reports about daily statistics of the detected anomalies in a stacked chart and overall statistics of all detected anomalies in a pie chart. The Setting tab is designed for getting any kind of information or system settings from the user. In the current implementation this tab is responsible for presenting the information about vessels that should be kept under surveillance and providing the features to add, remove, edit or disable/enable a vessel. Any changes in this tab would be considered in the next period of the AD process. The technologies that are used for implementing this module are Java, Spring Framework, JSP, JavaScript, jQuery and Google maps API. 17 3.5 Experimental Design Before using ODADS in real life, it is important to figure out to what extent the results of the system are accurate. Accuracy is the degree to which the estimates or measurements of a quantity correctly describe the exact value of that quantity. In other words, accuracy is the proportion of true results in the population. To evaluate the system accuracy, the number of True Positives (TP), False Positives (FP), True Negatives (TN) and False Negatives (FN) are needed. Accuracy is calculated by the following formula (Han, Kamber, & Pei, 2011): 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃 + 𝑇𝑁 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 The first step in designing the experiment is to identify the population. The population consists of the vessel traffic data in the surveillance area. Since the population is too large and it is impossible to look into all members manually to count the number of TP, FP, TN and FN, a sample should be taken from the population. The second step is to identify the sampling procedure. The important factors in sampling are the sampling frame, method and the sample size. The sampling frame consists of all members of the population that have the chance to be included in a sample and must be representative of the population. In this thesis the sampling frame is the vessel traffic data related to AIS, ports and pilots in the surveillance area, which are provided by ODADS. Due to the high volume of traffic through the surveillance area, it is expected that the majority of anomalies can be observed in one week execution of the system. Therefore, one week of vessel traffic data in April 2012 is used as the sample frame. For choosing a sampling method, it is important to decide whether or not to use probability sampling. In probability sampling, all members of the population have the chance of being selected in the sample. If the members have an equal probability of selection, the sample can be unbiased. In this thesis it is possible to perform a random sampling with equal probabilities. However, because of the ODADS periodic data collection (every 10 minutes), the data regarding to a specific vessel can be selected multiple times. To omit the effects of repeated data in a sample, a multistage sampling is performed. In the first stage, a simple random sampling without replacement is done for selecting the time slots that ODADS attempts to collect and analyze the data. After selecting the time slots, the corresponding data for each time slot will be selected by a stratified sampling. Three strata are defined according to the type of vessels: ferry and passenger, cargo and tanker vessels. Selection of vessels is also limited to the vessels that are originated from or targeted to the four particular ports. The total number of time slots in the sample frame is 835. This means that on average, ODADS collects data 139 times a day. In the first stage of sampling a random timeslot is selected for each day, which results in 7 time slots for one week. Then, by considering the described limitation in the selection process, the average number of entire data in a selected time slot is about 100 records. Among these records, 30 records are selected by stratification. Almost 73 % of the vessels in each time slot are moored. Since the majority of the anomalies are related to the vessels trips, a limitation on the number of moored vessels in the samples is defined. In this way, it can be possible to check more anomalies in the evaluation process. The second stage of sampling is repeated by taking into consideration that the number of moored vessel in the sample cannot exceed from the half of the sample size (in this case 15). 3.6 Validation Design After completing the implementation of ODADS, a meeting is held with the Coastguard representatives. During the meeting the system is presented and it is determined that ODADS will be used by the Coastguard officers for a period of time to verify the validity of the detected anomalies and evaluate the usefulness of such system for the Coastguard. Therefore, ODADS is installed on an Internet server and another meeting is arranged with 18 the Coastguard to define the validation process. As the outcome of the meeting, it is decided the validation will take place at the Coastguard headquarters office in Karlskrona, Sweden for four weeks (April 23, 2012 to May 18, 2012), at any time during working hours (08:00 to 17:00). The officers are supposed to evaluate the detected anomalies by checking them against the available data in the systems and data sources that are used during the normal operational activities at the Coastguard. They are asked to provide weekly report about their evaluation results in order to decrease the possible malfunctioning of the system and the validation process. An Excel form is provided to the officers in order to receive the reports in a common format. While investigating the validity of an alarm for a specific vessel, the following information should be reported: detection time, vessel information (such as name, MMSI and IMO), trip information, type of the alarm (true/false), systems or data sources that are used to investigate the detected anomaly. The Coastguard officers are also asked for providing any information about vessels that are marked as anomaly according to the available systems at the Coastguard, but ODADS is unable to recognize them. Although having such information is useful for analyzing the validation results, it is impossible for the Coastguard to provide this information. A detected anomaly for a vessel is true if it can be confirmed by the available data sources at the Coastguard and consequently it is false if the authorized data sources provide any information that declines the detected anomaly. No further assessment is done regarding the classification of the detected anomalies to true and false alarms. 4. VALIDITY THREATS In this thesis there are some issues that can threat the validity of the results and they should be considered before developing the system and performing the evaluation and validation. Construct validity refers to the extent to which the results of a study reflect the theory or the concept behind (Shadish, Cook, & Campbell, 2002). The main issue that may threat the construct validity is the design and reliability of implementation. Results can be affected by the potential faults that may happen in the implementation either because of programming faults or lack of tuning. In addition, the inaccurate nature of the open data that are used may have some effect on the results. The open data can have errors due to the human operator mistakes. They do not follow a similar format and can be unavailable for a while or are not updated immediately. To diminish the undesirable effect of the data, creating a common data representation format, filtering of the data and using approximate string matching techniques would be helpful. For decreasing the effect of programming faults, the validity of implemented application should be tested different times with both real and manipulated data that contain anomalies. There are also some parameters in the application that are needed to be selected correctly. For instance, in order to match the vessel information, string matching techniques are required and to determine that two strings are matched with each other, a similarity threshold should be used. Choosing an inappropriate value for such parameters will lead to incorrect results. The other threat of validity, which can occur while performing the evaluation and validation, targets both the internal and external validities. Internal validity ensures that the observed relationship between the treatment and outcome is due to a casual relationship and it is not because of an uncontrolled factor. External validity refers to the ability of generalizing the result of the study to other domains, times or places (Shadish et al., 2002). The threat to internal and consequently the external validity may occur if the data that are used in the evaluation process are biased and not representative of the population. In such situation generalizing the results of the treatments to the whole system is unrealistic. To prevent this issue, the samples, which are taken from the population in the experiment, should be real representative of the population. As well as the experiment, while performing the validation by the subject matter experts the same procedure for selecting and checking the detected anomalies should be followed. 19 5. VERIFICATION To ensure that ODADS works properly, the system is tested manually with both real and manipulated data. The tests are performed during the implementation and also after completing the system. At first, a number of vessels with different types of anomalies are inserted to the real collected data to check whether all types of anomalies can be detected by ODADS. Then, the system is run for a period of time and the detected anomalies are checked manually against the available data to make sure about their correctness. During the test phase the Anomaly Detector module is updated and some of the detection conditions are narrowed down. The process is repeated until the system can detect all the anomalies correctly. Before starting the validation process, the system is used by the subject matter experts from the Coastguard and according to their comments some of the algorithms in the Anomaly Detector module are updated. For example, ferries usually provide their schedule monthly or once in a couple of months; therefore, their arrival is not always available in daily schedule of the ports and the VESSEL_NOT_INFORMED_PORT and UNUSUAL_ TRIP_PATTERN anomalies are often set for them. For this reason, these two anomalies will not be checked for ferries. 6. EXPERIMENTAL RESULTS The results of the system execution during the specified week are provided. To provide an overview of the detected anomalies by ODADS, Table 6 illustrates the total number of vessels in the surveillance area. Table 7 shows the number of detected types of anomaly for each day and finally Table 8 provides the total percentage of each anomaly type during that week. Table 6 Total number of vessels in the surveillance area Days Avg 1 2 3 4 5 6 7 Total number of vessels 614 623 663 665 669 695 784 673.29 Cargo, Tanker, Passenger and Ferry vessels 345 349 365 369 370 372 394 366.29 Cargo, Tanker, Passenger and Ferry vessels that are originated from or targeted to the specified ports 136 142 141 137 145 149 142 141.71 Table 7 Total and the average number of detected anomalies during one week of execution Anomaly Days Avg 1 2 3 4 5 6 7 ARRIVAL_TIME_MISMATCHED 23 24 27 23 23 26 21 23.86 VESSEL_NOT_INFORMED_PORT 8 22 13 12 16 14 18 14.71 (Continued) 20 Anomaly Days Avg 1 2 3 4 5 6 7 VESSEL_NOT_LEFT_PORT 5 11 10 11 8 9 4 8.29 UNUSUAL_TRIP_PATTERN 1 6 3 5 4 4 3 3.71 VESSEL_NOT_USED_PILOT 1 3 2 2 2 2 2 2.00 UNUSUAL_TRIP_PATTERN_AND_VESSEL_ ARRIVAL_TIME_MISMATCHED 3 1 1 1 4 1 1 1.71 VESSEL_MOORED_IN_PORT 2 1 1 1 1 1 2 1.29 UNUSUAL_TRIP_PATTERN_AND_ VESSEL_NOT_INFORMED_PORT 0 0 0 1 1 1 1 0.57 VESSEL_ENTERED_PORT_WITHOUT_NOTIE 0 0 1 0 0 1 0 0.29 VESSEL_NOT_ENTERED_PORT 0 2 0 0 0 0 0 0.29 UNDER_SURVEILLANCE_VESSEL 1 1 0 0 0 0 0 0.29 VESSEL_ARRIVAL_TIME_MISMATCHED_ AND_VESSEL_NOT_USED_PILOT 0 0 1 0 0 1 0 0.29 VESSEL_ORDERED_PILOT_AND_NOT_ INFORMED_PORT 0 1 0 0 0 0 0 0.14 UNUSUAL_TRIP_PATTERN_AND_VESSEL_ NOT_USED_PILOT 0 0 0 0 0 0 1 0.14 VESSEL_ORDERED_PILOT_AND_NOT_ INFORMED_PORT_AND_VESSEL_NOT_ USED_PILOT 0 1 0 0 0 0 0 0.14 UNUSUAL_TRIP_PATTERN_AND_ VESSEL_ORDERED_PILOT_AND_NOT_ INFORMED_PORT 0 0 0 0 0 0 0 0.00 UNUSUAL_TRIP_PATTERN_AND_VESSEL_ NOT_USED_PILOT_AND_VESSEL_ARRIVAL_ TIME_MISMATCHED 0 0 0 0 0 0 0 0.00 VESSEL_ENTERED_PORT_WITHOUT_ NOTICE_AND_NOT_LEFT_PORT_ON_TIME 0 0 0 0 0 0 0 0.00 VESSEL_NOT_ENTERED_PORT_AND_NOT_ USED_PILOT 0 0 0 0 0 0 0 0.00 UNUSUAL_TRIP_PATTERN_AND_VESSEL_ NOT_ENTERED_PORT 0 0 0 0 0 0 0 0.00 UNUSUAL_TRIP_PATTERN_AND_ VESSEL_NOT_ENTERED_PORT_AND_ VESSEL_NOT_USED_PILOT 0 0 0 0 0 0 0 0.00 UNUSUAL_TRIP_PATTERN_AND_VESSEL_ ORDERED_PILOT_AND_NOT_INFORMED_ PORT_AND_VESSEL_NOT_USED_PILOT 0 0 0 0 0 0 0 0.00 21 Table 8 The percentage of each anomaly during one week of execution Anomaly Count (%) ARRIVAL_TIME_MISMATCHED 41.34 VESSEL_NOT_INFORMED_PORT 25.49 VESSEL_NOT_LEFT_PORT 14.36 UNUSUAL_TRIP_PATTERN 6.43 VESSEL_NOT_USED_PILOT 3.47 UNUSUAL_TRIP_PATTERN_AND_VESSEL_ARRIVAL_TIME_MISMATCHED 2.96 VESSEL_MOORED_PORT 2.23 UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_INFORMED_PORT 0.99 VESSEL_ENTERED_PORT_WITHOUT_NOTICE 0.50 VESSEL_NOT_ENTERED_PORT 0.50 UNDER_SURVEILLANCE_VESSEL 0.50 VESSEL_ARRIVAL_TIME_MISMATCHED_AND_VESSEL_NOT_USED_PILOT 0.50 VESSEL_ORDERED_PILOT_AND_NOT_INFORMED_PORT 0.24 UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_USED_PILOT 0.24 VESSEL_ORDERED_PILOT_AND_NOT_INFORMED_PORT_AND_ VESSEL_NOT_USED_PILOT 0.24 UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_ENTERED_PORT 0.00 UNUSUAL_TRIP_PATTERN_AND_VESSEL_ORDERED_PILOT_AND_NOT_ INFORMED_PORT 0.00 VESSEL_ENTERED_PORT_WITHOUT_NOTICE_AND_NOT_LEFT_PORT_ON_ TIME 0.00 VESSEL_NOT_ENTERED_PORT_AND_NOT_USED_PILOT 0.00 UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_USED_PILOT_AND_ VESSEL_ARRIVAL_TIME_MISMATCHED 0.00 UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_ENTERED_ PORT_AND_VESSEL_NOT_USED_PILOT 0.00 UNUSUAL_TRIP_PATTERN_AND_VESSEL_ORDERED_PILOT_AND_NOT_ INFORMED_PORT_AND_VESSEL_NOT_USED_PILOT 0.00 After carrying out the sampling, all the samples are checked against the primary identified anomalies (A1-A9). To compute the number of TP, FP, TN and FN, a confusion 22 matrix is created based on the nine classes of anomalies and the normal class (Table 9). The accuracy of the system is calculated as follow: 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃 + 𝑇𝑁 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 = 17 + 192 17 + 192 + 0 + 1 = 0.99 The existing FN for the ARRIVAL_TIME_MISMATCHED anomaly is due to the wrong provided AIS data by the vessel or possibly the MarineTraffic.com website and also the limitation of the system for considering all conditions. In this case, the vessel arrival time belongs to a couple of days before the current date and for this reason it is ignored by the system. However, it is quite possible to handle such situations if additional sources of AIS data are available. Table 9 Confusion matrix for the nine classes of anomalies and the normal class Predicted class A1 A2 A3 A4 A5 A6 A7 A8 A9 Normal Total Actual class A1 6 - - - - - - - - - 6 A2 - 7 - - - - - - - 1 8 A3 - - - - - - - - - - 0 A4 - - - - - - - - - - 0 A5 - - - - 3 - - - - - 3 A6 - - - - - 1 - - - - 1 A7 - - - - - - - - - - 0 A8 - - - - - - - - - - 0 A9 - - - - - - - - - - 0 Normal - - - - - - - - - 192 192 Total 6 7 0 0 3 1 0 0 0 193 17 7. VALIDATION RESULTS The sea monitoring system that is used by the Coastguard officer is called SJÖBASIS1 • SafeSeaNet . SJÖBASIS aggregates the maritime data from different systems and agencies with the aim of improving the efficiency of MS. In SJÖBASIS, the required data that contain vessel position, speed, heading, arrival/departure time and trip, are obtained from the following sources: 2 1 : A vessel traffic monitoring and information system, which is the centralized European platform for maritime data exchange and linking maritime authorities across Europe. It enables the member countries to provide and receive information about vessels, vessels movements and hazardous cargoes. Main sources of information include AIS-based position reports and notification messages that are sent by designated authorities in the participating countries. This system is available via the Swedish Maritime Administration. www.kustbevakningen.se/sv/granslos-samverkan/sjoovervakningsuppdraget/samverkan-sjoinformation/ 2 www.emsa.europa.eu/operations/maritime-surveillance/safeseanet.html http://www.kustbevakningen.se/sv/granslos-samverkan/sjoovervakningsuppdraget/samverkan-sjoinformation/� http://www.emsa.europa.eu/operations/maritime-surveillance/safeseanet.html� 23 • SjöC: The Swedish naval forces surveillance center, which uses military sensors for correlating positions with AIS. • AIS: The international system for presenting the vessels position and identity. • HELCOM AIS1 The officer checks the validity of anomalies according to the priority that each anomaly has for him. Table 10 provides the priority of the anomalies for an officer with an interest in law enforcement. : The national AIS data that is consolidated in cooperation of countries around the Baltic Sea. Table 10 Prioritization of the identified anomalies from a law enforcement officer’s point of view Anomaly Priority ARRIVAL_TIME_MISMATCHED 5 VESSEL_ENTERED_PORT_WITHOUT_NOTICE 5 VESSEL_NOT_ENTERED_PORT 5 UNDER_SURVEILLANCE_VESSEL 5 VESSEL_NOT_LEFT_PORT 4 VESSEL_MOORED_IN_PORT 4 UNUSUAL_TRIP_PATTERN 3 VESSEL_NOT_INFORMED_PORT 2 VESSEL_NOT_USED_PILOT 1 VESSEL_ORDERED_PILOT_AND_NOT_INFORMED_PORT 1 Note. 5 shows the highest priority and 1shows the lowest. During the four-week validation period, ODADS is used at the Coastguard for 12 working days and in total 76 of the detected anomalies are evaluated. Table 11 presents the validation results. Among the evaluated anomalies, there are a number of anomalous vessels that remain unchecked due to a lack of corresponding data in the Coastguard systems. A large number of detected anomalies are related to the ARRIVAL_TIME_MISMATCHED anomaly that in many cases can be due to the inconsistent time formats (UTC, CET and EET either with or without daylight saving) in different data sources and various settings in the AIS transmitters. In these cases, the detected anomalies by ODADS are logically true, however, in actual fact they are not. For example, a vessel that is going from Helsinki to Stockholm is reporting its arrival time according to the Finland local time (EET format) instead of using the UTC format. Therefore, this time difference causes that the arrival time reported by the vessel does not match with the expected arrival time of the vessel at the destination port, which leads to set the ARRIVAL_TIME_MISMATCHED anomaly for that vessel. 1 www.helcom.fi/BSAP/ActionPlan/en_GB/SegmentSummary/ http://www.helcom.fi/BSAP/ActionPlan/en_GB/SegmentSummary/� 24 Table 11 Validation result of the Coastguard Alarms Anomaly True False Not Checked VESSEL_NOT_INFORMED_PORT 7 3 4 ARRIVAL_TIME_MISMATCHED 19 5 3 VESSEL_NOT_USED_PILOT 2 - - UNUSUAL_TRIP_PATTERN 7 6 - VESSEL_NOT_LEFT_PORT 1 1 - VESSEL_ORDERED_PILOT_AND_NOT_INFORMED_PORT 2 - - VESSEL_MOORED_IN_PORT - 3 - UNDER_SURVEILLANCE_VESSEL 1 - - UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_ INFORMED_PORT 5 1 - UNUSUAL_TRIP_PATTERN_AND_VESSEL_ ARRIVAL_ TIME_ MISMATCHED 4 - - UNUSUAL_TRIP_PATTERN_AND_VESSEL_ORDERED_ PILOT_AND_NOT_INFORMED_PORT - 1 - VESSEL_ORDERED_PILOT_AND_NOT_INFORMED_ PORT_AND_VESSEL_NOT_USED_PILOT 1 - - Total count 49 20 7 Total count (%) 64.47 26.32 9.21 In addition to the comparative analysis in the validation process, modus operandi of using an anomaly detector that takes advantage of open data is investigated. The Coastguard officer uses an analysis tool1 It is also possible to look into the relation between the types of vessels and the detected anomalies. Such assessment can be used for strategic and risk analysis. For tanker vessels the to analyze the ODADS excel reports and draw conclusions regarding the modus operandi of the system in emergency situations that can have impact on the MS operations. One of the possible analyses of the system reports can be the investigation of vessels with multiple anomalies. Here are some examples of the vessels with multiple anomalies. A cargo vessel has recurring anomalies related to the arrival/departure time, trip and notification to the port. From the types of anomalies that are detected for this vessel, it can be concluded that the vessel behavior points to a higher threat concerning customs and border. For a passenger vessel the following anomalies are often detected: VESSEL_NOT_INFORMED_PORT, ARRIVAL_TIME_MISMATCHED, VESSEL_NOT_ ENTERED_PORT and VESSEL_NOT_LEFT_PORT. These anomalies may happen because of inaccurate or wrong provided data by the vessel. The conclusion and modus operandi for this vessel is that the duty officer will contact the vessel to highlight the importance of submitting accurate information. In real life, occurring anomalies such as VESSEL_NOT_ENTERED_PORT or VESSEL_NOT_LEFT_PORT for a passenger vessel can be related to serious issues such as an accident and it will result in increased scrutiny for that vessel. 1IBM i2 Analyst's Notebook available at www.i2group.com/us/products/analysis-product-line/ibm-i2-analysts- notebook http://www.i2group.com/us/products/analysis-product-line/ibm-i2-analysts-notebook� http://www.i2group.com/us/products/analysis-product-line/ibm-i2-analysts-notebook� 25 most popular anomaly is VESSEL_NOT_INFORMED_PORT. This anomaly has a high priority for emergency preparedness for accidents involving tankers. Tankers with UNUSUAL_TRIP_PATTERN anomaly are a potential risk to other vessels and can cause accidents. From a risk assessment point of view the combination of this anomaly with the VESSEL_NOT_ENTERED_PORT or VESSEL_NOT_USED_PILOT anomalies can lead to high-risk situations. The most occurring anomaly for ferries and passenger vessels is ARRIVAL_TIME_MISMATCHED. In several cases this anomaly is detected incorrectly because of the wrong reported arrival time; however, this anomaly has great importance for the authorities to plan their operations regarding ferries and passenger vessels arrivals effectively. The most serious anomaly for ferries and passenger vessels is VESSEL_NOT_ LEFT_PORT and the authorities should suspect that some form of accident or difficulty is arising regarding to the departure of the vessels. For cargo vessels, the most recurring anomalies are VESSEL_NOT_INFORMED_PORT and ARRIVAL_TIME_MISMATCHED. However, some of the ARRIVAL_TIME_MISMATCHED anomalies are false alarms because of the incorrect data. The VESSEL_NOT_INFORMED_PORT anomaly is important for ports security and safety. The prior notification to the ports is obligatory for vessels, but the fine for breaking this rule is negligible which lets the vessels that are involved in illegal activities such as smuggling disobey this rule. The most serious anomalies for the cargo vessels are the VESSEL_NOT_ENTERED_PORT and UNDER_SURVEILLANCE_ VESSEL anomalies. Furthermore, looking into the most frequent anomalies for different ports will assists the maritime authorities to make their decisions more efficiently. According to the report analysis, the most frequent anomaly in Stockholm and Nynäshamn ports is ARRIVAL_ TIME_MISMATCHED, in port of Kapellskär is VESSEL_NOT_INFORMED_PORT and in Norrköping port is UNUSUAL_TRIP_PATTERN. One possible conclusion from the most popular anomalies for the ports is that the ports authorities should be informed of the divergence in the traffic flow and the operational management functions in order to plan and allocate resources efficiently. On the other hand, in some cases the anomalies are too common for a port because of the inaccurate provided data and they can be disregarded by the officer. The received feedback from the Coastguard representatives during the validation process shows that ODADS is implemented in a way that it can assist the operator to have a better understanding of the ongoing maritime activities. They believe that the ODADS results are reliable and the quality of the open data that are used is good and can be used in real life. The functionality, usability and visualization of ODADS are satisfying. In addition to illustrating the vessel traffic data in a simple, clear and informative way, ODADS can provide clear statement about the anomalies and its statistical reports are beneficial when the authorities and freight companies conduct strategic analysis of maritime traffic and risk assessment. Finally, the capability of automated detection of anomalies by utilizing open data and the form of data representation could prove to be a valuable asset to the Coastguard. 8. DISCUSSION Taking advantage of AD systems will assist authorities to tighten security in the MS domain. There are a number of studies which focused on developing AD systems by using knowledge-driven and data-driven approaches. For instance, Defense R&D Canada (Roy, 2008, 2010) developed a rule-based prototype for AD by exploiting maritime situational facts about both kinematic and static data of the domain. Edlund et al (2006) developed another prototype for a rule-based expert system to detect the anomalies regarding spatial and kinematic relation between objects. Riveiro and Falkman (2009) proposed using a combination of data-driven and knowledge-driven approaches to detect anomalies by use of a normal model of vessel behavior based on AIS data and experts rules. In the majority of 26 studies that addressed AD, the exploited data for the AD process were obtained from closed data sources and there is a lack of investigation on using open data sources for AD in the MS domain. Therefore, in this thesis ODADS is implemented by employing expert rules to investigate the potential open data as a complement to the closed data for AD in the MS domain. The accuracy of ODADS is evaluated via an experiment and the experimental results indicate an accuracy of 99% for the system. In addition, validity of the system results is evaluated in real life by the experts from the Swedish Coastguard. Despite the inaccurate nature of open data and by considering the fact that only open data sources are used in the system, the high number of true alarms (64.47%) in the validation process admits the validity of the system outcomes. Furthermore, there are no corresponding data in the authorized databases for 9.21% of the evaluated anomalies by the Coastguard. This fact refers to a potential information gap in the closed data sources. However, the considerable number of false alarms (26.32%) for a surveillance system is unsatisfying. The number of false alarms indicates the difference between the accuracy of the system and the validity of the results. Even though the data that are used in this thesis are obtained from relatively trusted data sources such as ports, the false alarms occur mostly because of the inaccurate data. The open data that are exploited by ODADS suffer from the errors due to human operator mistakes, irregular data update, data update latency and incompatible data format. In ODADS, there are situations that a detected anomaly is disappeared in the next periods of the system execution because of new arrival of the correct data. Frequent occurrence of false alarms distracts operator’s attention from the real anomalies in the surveillance area. To decrease the false alarms in ODADS, the main solution is to integrate the open data with the closed data which can cover the lack of information or inaccuracy in the open data. In addition, defining a probability for detected anomalies can decrease the number of false alarms. This would be possible by analyzing the history of vessels behavior as well as the current situation and defining a probability threshold to omit the anomalies that have a lower probability than the threshold. Furthermore, having extra information regarding vessels such as crew and cargo information, can affect the probability of being a real anomaly for a specific vessel. For example, if a vessel has the ARRIVAL_TIME_MISMATCHED anomaly and it has a crew member with a criminal record or a special cargo, then there is a possibility that the vessel is stopped somewhere to exchange something. Therefore, in such situation, the probability of being a true anomaly is high. According to the validation results, the UNUSUAL_TRIP_PATTERN anomaly creates the majority of false alarms. This is due in part to the statistical approach that is used for detecting this anomaly and also the wrong origin and destination information that the vessels provide. The lookup table that is created for storing the frequency of the trips between different places is not updated periodically. While populating the table, the ports timetables are used which can be incomplete. An alternative detection approach can be using of machine learning techniques which attempt to detect the anomaly according to the pattern of movements for individual vessels instead of the reported trip data by the vessels or ports. In local areas such as the surveillance area in this thesis, mainly because of large amount of quality assured data and the limited size of the surveillance area, it is easier for the maritime authorities to track and control the vessels activities. Therefore, use of open AIS data in this region is not required and it should be prohibited to decrease the negative impacts of open data on the system results. On the other hand, when the vessel information beyond the EEZ is required, the value of open data becomes more obvious. 9. CONCLUSION AND FUTURE WORK This thesis investigated the potential open data as a complementary resource for Anomaly Detection (AD) in the Maritime Surveillance (MS) domain. A framework for AD was proposed based on the usage of open data sources along with other traditional sources of 27 data. According to the proposed AD framework and the algorithms for implementing the expert rules, the Open Data Anomaly Detection System (ODADS) was developed. To evaluate the accuracy of the system, an experiment on the vessel traffic data was conducted and an accuracy of 99% was obtained for the system. There was a false negative case in the system results that decreased the accuracy. It was due to incorrect AIS data in a special situation that was not possible to be handled by the detection rules in the scope of this thesis. The validity of the results was investigated by the subject matter experts from the Swedish Coastguard. The validation results showed that the majority of the ODADS evaluated anomalies were true alarms. Moreover, a potential information gap in the closed data sources was observed during the validation process. Despite the high number of true alarms, the number of false alarms was also considerable that was mainly because of the inaccurate open data. This thesis provided insights into the open data as a complement to the common data sources in the MS domain and is concluded that using open data will improve the efficiency of the surveillance systems by increasing the accuracy and covering some unseen aspects of maritime activities. In the future, it is important to investigate how the open data sources in the maritime domain can be used in a global perspective. In this thesis, the surveillance area was limited to a local area which is fully covered by the authorities’ data sources. When the data beyond the exclusive economic zone are needed, it is more valuable to use open data sources. By taking advantage of the subject matter experts’ knowledge about MS, it would be possible to figure out how the global open data should be exploited for the surveillance purpose. Integration of the open data with maritime confidential data can improve the efficiency of MS and should be considered as a further improvement of the system. Another improvement can be considering a probability for each detected anomaly according to the history of the vessels behavior and the current situation. Moreover, further investigation on the other sources of open data such as social data, which is created and shared through social media platforms, and online videos from the ports activities in the high risk regions, will be useful. The data that are used in ODADS are relatively trusted, but in case of using other open data sources in the MS domain for AD, the quality assurance of the data should be investigated. As well as using knowledge based systems, taking advantage of data-driven approaches such as machine learning techniques can increase the efficiency of the MS systems. Finally, the next step for improving the MS systems after being equipped with the AD functionality is to predict the future threats or incoming anomalies based on the analysis of the current situation. 28 APPENDIX A: OPEN AND CLOSED DATA SOURCES Table A.1 Maritime open and closed data sources that are available via the Internet AR Organization name Categories NAR FR NFR Provider Baltic and International Maritime Council www.bimco.org/ Shipping information Consulting services ● Denmark Association of Ship Brokers and Agents, Inc. www.asba.org/ Ship brokers information ● USA Bureau International des Containers www.bic-code.org/ Freight containers-coding Identification Marking ● France European Maritime Safety Agency www.emsa.europa.eu/oil-recovery-vessels/vessel-technical- specifications.html Maritime safety Prevention of pollution from ships ● Portugal ICC Commercial Crime Services www.icc-ccs.org/ Fraud in international trade ● ● UK International Association of Classification Societies www.iacs.org.uk/shipdata/default.aspx Maritime safety Regulation ● ● UK International Group of P&I Clubs www.igpandi.org/Home liability and insurance issues ● ● UK International Association of Independent Tanker Owners www.intertanko.com/ Transportation safety Prevention of pollution from ships Free competition ● UK (Continued) http://www.bimco.org/� http://www.asba.org/� http://www.bic-code.org/� http://www.emsa.europa.eu/oil-recovery-vessels/vessel-technical-specifications.html� http://www.emsa.europa.eu/oil-recovery-vessels/vessel-technical-specifications.html� http://www.iacs.org.uk/shipdata/default.aspx� http://www.igpandi.org/Home� http://www.intertanko.com/� 29 AR Organization name Categories NAR FR NFR Provider Oil Companies International Marine Forum www.ocimf.com/Home Transportation safety Prevention of pollution from ships ● UK Internet Ships Register www.ships-register.com/ Ships information ● UK World Shipping Register www.world-register.org/ Ships information Ports information Companies information ● — Lloyd’s Register Ships In Class www.lrshipsinclass.lrfairplay.com/default.aspx Ships information ● UK International Telecommunication Union www.itu.int/ITU- R/index.asp?category=terrestrial&rlink=mars&lang=en Ships information Coasts information Addresses of accounting authorities, administrations which notify information MMSI assigned to search and rescue aircraft MMSI assigned to AIS Aids to Navigation ● ● Switzerland International Maritime Consultancy Specialists in VTS www.maritime-vts.co.uk/ Vessel traffic services and maritime organization ● UK Port Directory www.port-directory.com/ Ports information ● UK Equasis www.equasis.org/EquasisWeb/public/HomePage?fs=HomePage Ships information Companies information ● Portugal Q88.COM www.q88.com/Home.aspx?c=1 Questionnaire generator Ships information ● ● USA InforMare www.informare.it/indexuk.htm Shipping information ● Italy (Continued) http://www.ocimf.com/Home� http://www.ships-register.com/� http://www.world-register.org/� http://www.lrshipsinclass.lrfairplay.com/default.aspx� http://www.itu.int/ITU-R/index.asp?category=terrestrial&rlink=mars&lang=en� http://www.itu.int/ITU-R/index.asp?category=terrestrial&rlink=mars&lang=en� http://www.maritime-vts.co.uk/� http://www.port-directory.com/� http://www.equasis.org/EquasisWeb/public/HomePage?fs=HomePage� http://www.q88.com/Home.aspx?c=1� http://www.informare.it/indexuk.htm� 30 AR Organization name Categories NAR FR NFR Provider Paris Mou www.parismou.org/ Port State control ● Netherlands Tokyo Mou www.tokyo-mou.org/ Port State control ● Japan Australian Government Bureau of Meteorology www.bom.gov.au/ Weather, climate and water ● ● Australia Finnish Meteorological Institute en.ilmatieteenlaitos.fi/home Weather, climate and water ● Finland Meteo France france.meteofrance.com/ Weather, climate and water ● ● France Earth Science Office NASA weather.msfc.nasa.gov/GOES/ Weather, climate and water ● — The Weather Channel www.weather.com/ Weather, climate and water ● — Ocean Color WEB NASA oceancolor.gsfc.nasa.gov/ Weather, climate and water ● ● — Earth European Space Agency earth.esa.int/ers/eeo4.10075/atsr_med.html Weather, climate and water ● ● Italy Weather BBC www.bbc.co.uk/weather/ Weather, climate and water ● UK Sailwx.info www.sailwx.info/ Live marin information ● USA Ship.gr www.ship.gr/ Ship brokers information Ship suppliers Companies information ● — (Continued) http://www.parismou.org/� http://www.tokyo-mou.org/� http://www.bom.gov.au/� http://france.meteofrance.com/� http://weather.msfc.nasa.gov/GOES/� http://www.weather.com/� http://oceancolor.gsfc.nasa.gov/� http://earth.esa.int/ers/eeo4.10075/atsr_med.html� http://www.bbc.co.uk/weather/� http://www.sailwx.info/� http://www.ship.gr/� 31 AR Organization name Categories NAR FR NFR Provider American Bureau of Shipping www.eagle.org/eagleExternalPortalWEB/appmanager/absEagle/absEa gleDesktop?_nfpb=true&_pageLabel=abs_eagle_portal_home_page Classification Societies ● USA Det Norske Veritas www.dnv.com/ Classification Societies ● ● Norway Bureau Veritas Groups www.bureauveritas.com/wps/wcm/connect/bv_com/Group/Footer/Ho me/ Classification Societies ● ● — China Classifiaction Societies www.ccs.org.cn/en/index.htm Classification Societies ● China HELLENIC REGISTER OF SHIPPING www.hrs.gr/index.htm Classification Societies ● Greece Nippon Kaiji Kyokai www.classnk.or.jp/hp/en/index.aspx Classification Societies ● Japan Vesseltracker.com www.vesseltracker.com/en/VesselArchive.html AIS Data Ships information ● ● Germany Digital Seas www.digital-seas.com/start.html AIS Data Ships information ● ● Germany MarineTraffic.com www.marinetraffic.com/ais/ AIS Data Ships information ● Greece Shipspotting.com www.shipspotting.com/ AIS Data Ships information ● ● — International Maritime Organization www5.imo.org/SharePoint/mainframe.asp?topic_id=334&offset Piracy reports ● UK Copenhagen Malmö Port www.cmport.com/ Port Authorities ● Denmark (Continued) http://www.eagle.org/eagleExternalPortalWEB/appmanager/absEagle/absEagleDesktop?_nfpb=true&_pageLabel=abs_eagle_portal_home_page� http://www.eagle.org/eagleExternalPortalWEB/appmanager/absEagle/absEagleDesktop?_nfpb=true&_pageLabel=abs_eagle_portal_home_page� http://www.dnv.com/� http://www.bureauveritas.com/wps/wcm/connect/bv_com/Group/Footer/Home/� http://www.bureauveritas.com/wps/wcm/connect/bv_com/Group/Footer/Home/� http://www.ccs.org.cn/en/index.htm� http://www.hrs.gr/index.htm� http://www.classnk.or.jp/hp/en/index.aspx� http://www.vesseltracker.com/en/VesselArchive.html� http://www.digital-seas.com/start.html� http://www.marinetraffic.com/ais/� http://www.shipspotting.com/� http://www5.imo.org/SharePoint/mainframe.asp?topic_id=334&offset� http://www.cmport.com/� 32 AR Organization name Categories NAR FR NFR Provider PORT OF GOTHENBURG www.portgot.se/prod/hamnen/ghab/dalis2b.nsf Port Authorities ● Sweden Genoa Port Authority www.porto.genova.it/index.php/en Port Authorities ● Italy Port of Klaipeda www.portofklaipeda.lt/en.php Port Authorities ● Lithuania Philippine Ports Authority www.ppa.com.ph/ Port Authorities ● Philippine Panama Canal Authority www.pancanal.com/eng/index.html Port Authorities ● USA UK P&I CLUB www.ukpandi.com/ liability and insurance issues ● ● UK The American Club www.american-club.com/ liability and insurance issues ● ● USA Steamship Mutual www.simsl.com/ liability and insurance issues ● ● Bermuda SKULD www.skuld.com/ liability and insurance issues ● Norway North of England P&I Association www.nepia.com/home/ liability and insurance issues ● UK The standard club www.standard-club.com/ liability and insurance issues ● ● UK Baltic Ports Organization www.bpoports.com/ Port coordinator ● Denmark Note. NAR = no authorization required; AR= authorization required; FR = free registration; NFR = non-free registration; Dashes indicate undisclosed information. http://www.portgot.se/prod/hamnen/ghab/dalis2b.nsf� http://www.porto.genova.it/index.php/en� http://www.portofklaipeda.lt/en.php� http://www.ppa.com.ph/� http://www.pancanal.com/eng/index.html� http://www.ukpandi.com/� http://www.american-club.com/� http://www.simsl.com/� http://www.skuld.com/� http://www.nepia.com/home/� http://www.standard-club.com/� http://www.bpoports.com/� 33 Table A.2 Data sources that are used in the implementation Website name Data type Website URL Marinetraffic.com Real time information based on AIS systems www.marinetraffic.com/ais/ Swedish Maritime Administration (Sjofartsverket) Stockholm pilotage area www.sjofartsverket.se/sv/Infrastruktur- amp-Sjotrafik/Lotsning/Lotsinfo/ Ports of Stockholm, Kapellskär and Nynäshamn Vessels in port and expected arrival www.stockholmshamnar.se/en/Karta/Vessel -calls/ Port of Norrköping Vessels in port www.norrkoping- port.se/anlop.php?page=snabb_fih&link=11 0|111 Expected vessels arrival www.norrkoping- port.se/anlop.php?page=snabb_fih_ank&lin k=110|111 Port of Helsinki Cargo vessels in port www.portofhelsinki.fi/cargo_traffic/vessels _in_ports Expected cargo vessels arrival www.portofhelsinki.fi/cargo_traffic/arrival_ ships Expected passenger vessels departure www.portofhelsinki.fi/passengers/departure _times_and_terminals Expected passenger vessels arrival www.portofhelsinki.fi/passengers/arrival_ti mes_and_terminals Passenger vessels have visited the port before www.portofhelsinki.fi/passengers/cruise_shi ps_that_have_visited_the_port Port of Tallinn Vessels in port www.ts.ee/?op=ships_in_port&lang=eng Expected cargo vessels arrival www.ts.ee/?op=cargo_ships_arrivals&lang= eng Expected passenger vessels arrival www.ts.ee/?op=passenger_ship_arrivals&la ng=eng Expected passenger vessels departure www.ts.ee/?op=passenger_ship_departures &lang=eng http://www.marinetraffic.com/ais/� http://www.sjofartsverket.se/sv/Infrastruktur-amp-Sjotrafik/Lotsning/Lotsinfo/� http://www.sjofartsverket.se/sv/Infrastruktur-amp-Sjotrafik/Lotsning/Lotsinfo/� http://www.stockholmshamnar.se/en/Karta/Vessel-calls/� http://www.stockholmshamnar.se/en/Karta/Vessel-calls/� http://www.norrkoping-port.se/anlop.php?page=snabb_fih&link=110|111� http://www.norrkoping-port.se/anlop.php?page=snabb_fih&link=110|111� http://www.norrkoping-port.se/anlop.php?page=snabb_fih&link=110|111� http://www.norrkoping-port.se/anlop.php?page=snabb_fih_ank&link=110|111� http://www.norrkoping-port.se/anlop.php?page=snabb_fih_ank&link=110|111� http://www.norrkoping-port.se/anlop.php?page=snabb_fih_ank&link=110|111� http://www.portofhelsinki.fi/cargo_traffic/vessels_in_ports� http://www.portofhelsinki.fi/cargo_traffic/vessels_in_ports� http://www.portofhelsinki.fi/cargo_traffic/arrival_ships� http://www.portofhelsinki.fi/cargo_traffic/arrival_ships� http://www.portofhelsinki.fi/passengers/departure_times_and_terminals� http://www.portofhelsinki.fi/passengers/departure_times_and_terminals� http://www.portofhelsinki.fi/passengers/arrival_times_and_terminals� http://www.portofhelsinki.fi/passengers/arrival_times_and_terminals� http://www.portofhelsinki.fi/passengers/cruise_ships_that_have_visited_the_port� http://www.portofhelsinki.fi/passengers/cruise_ships_that_have_visited_the_port� http://www.ts.ee/?op=ships_in_port&lang=eng� http://www.ts.ee/?op=cargo_ships_arrivals&lang=eng� http://www.ts.ee/?op=cargo_ships_arrivals&lang=eng� http://www.ts.ee/?op=passenger_ship_arrivals&lang=eng� http://www.ts.ee/?op=passenger_ship_arrivals&lang=eng� http://www.ts.ee/?op=passenger_ship_departures&lang=eng� http://www.ts.ee/?op=passenger_ship_departures&lang=eng� 34 Table A.3 Required data for detection of each anomaly Anomaly Data VESSEL_NOT_INFORMED_PORT AIS - Port vessel arrival data ARRIVAL_TIME_MISMATCHED AIS - Port vessel arrival data VESSEL_ENTERED_PORT_ WITHOUT_NOTICE AIS - Port vessel arrival data - Port vessel departure/ in-port data VESSEL_NOT_USED_PILOT AIS - Pilot vessels service data UNUSUAL_TRIP_PATTERN AIS - Port vessel arrival data – Port vessel departure/ in-port data VESSEL_NOT_LEFT_PORT AIS - Port vessel departure data VESSEL_NOT_ENTERED_PORT AIS - Port vessel arrival data – Port vessel departure/ in-port data VESSEL_ORDERED_PILOT_AND_ NOT_INFORMED_PORT AIS – Port vessel arrival data – Pilot vessels service data VESSEL_MOORED_IN_PORT AIS – Port vessel departure/ in-port data 35 APPENDIX B: PORTS REGIONS This section provides the geographical information about each port and its harbors. Table B.1 Ports areas and their geographical positions Port Number of areaa Minimum longitude Maximum longitude Minimum latitude Maximum latitude Helsinki 1 24.899414 24.985057 60.140472 60.18381 Kapellskär 1 19.064606 19.079539 59.715444 59.728009 Norrköping 3 16.180591 16.228976 58.59246 58.614909 16.231232 16.254241 58.608807 58.625638 16.260810 16.269291 58.635730 58.640399 Nynäshamn 1 17.949054 17.982726 58.898194 58.932706 Stockholm 6 18.060464 18.108335 59.29992 59.332768 18.101335 18.147693 59.33865 59.3577276 18.023163 18.061236 59.318440 59.328193 18.017862 18.034127 59.312907 59.317289 17.821146 17.825557 59.360349 59.362994 18.163756 18.174964 59.318795 59.322375 Tallinn 6 24.624438 25.010307 59.425881 59.524365 24.021952 24.098137 59.327022 59.362901 22.221453 22.249035 58.526984 58.541255 24.793068 24.877004 59.553744 59.602800 24.452785 24.585510 59.464992 59.497719 23.845721 23.957913 59.345430 59.389721 Note. a In order to have a more precise AD, harbors and areas around each port that are used by vessels are identified. 36 Figure B.1. Helsinki port area (The image is adapted from Google Earth). Figure B.2. Kapellskär port area, Stockholm group (The image is adapted from Google Earth). 37 Figure B.3. Norrköping port area (The image is adapted from Google Earth). Figure B.4. Nynäshamn port area, Stockholm group (The image is adapted from Google Earth). 38 Figure B.5.Stockholm port areas, Stockholm group (The image is adapted from Google Earth). Figure B.6. Tallinn port areas (The image is adapted from Google Earth). 39 APPENDIX C: ALGORITHMS This section provides a list of all algorithms that are used for implementing Anomaly Detection module. Algorithm 1 Main procedure for the Anomaly Detector module while true 1. Extract AIS, Ports and Pilot data from database regarding to the latest time of data collection by the Data Collector module 2. Call methods regarding to the following algorithms: a. Algorithm 2 b. Algorithm 3 c. Algorithm 4 d. Algorithm 5 e. Algorithm 6 f. Algorithm 7 g. Algorithm 8 3. Determine the final type of anomaly for vessels (Algorithm 19) 4. Store detected anomalies in database 5. Wait until the new data are collected by the Data Collector module end while Algorithm 2 Detect VESSEL_NOT_INFORMED_PORT and ARRIVAL_TIME_MIS- MATCHED anomalies ARRIVAL_TIME_DELAY_THRESHOLD ← 30 minutes for each vessel ∈ AIS data do destination ← vessel destination if destination ≠ null and isRelatedToPorts(destination) and not isInPortsArea(vessel) then vessel_found ← false port_schedule ← vessels that are expected to arrive at destination port // For vessels that have not updated their trip information at the beginning of their trip if destination contains vessel origin and vesselLeftPortRecently(destination port, vessel) and not isTargetedPort(destination port, vessel) then continue end if for each port_vessel ∈ port_schedule do if (vessel arrival date = port_vessel arrival date or current date = port_vessel arrival date or // To consider an interval for changing dates |vessel arrival time - port_vessel arrival time| < FIVE_HOURS ) and isVesselSimilar(vessel, port_vessel) then vessel_found ← true if vessel arrival time ≠ port_vessel arrival time and |vessel arrival time - port_vessel arrival time| > ARRIVAL_TIME_DELAY_THRESHOLD then add(Vessel anomaly, ARRIVAL_TIME_MISMATCHED) 40 Algorithm 2 Detect VESSEL_NOT_INFORMED_PORT and ARRIVAL_TIME_MIS- MATCHED anomalies end if break end if end for if vessel_found ≠ true then add(Vessel anomaly, VESSEL_NOT_INFORMED_PORT) end if end if end for Algorithm 3 Detect VESSEL_ENTERED_PORT_WITHOUT_NOTICE anomaly for each port ∈ available ports port_vessels ← vessels that are currently available in port and arrived one hour ago for each port_vessel ∈ port_vessels do expected_vessels ← vessels that were expected to enter the port at vessel arrival time vessel_found ← false for each exp_vessel ∈ expected_vessels do if isVesselSimilar(port_vessel, exp_vessel) then vessel_found ← true break end if end for if vessel_found ≠ true then add(port_vessel anomaly, VESSEL_ENTERED_PORT_WITHOUT_NOTICE) (port_vessel) end if end for end for Algorithm 4 Detect VESSEL_NOT_USED_PILOT and VESSEL_ORDERED_PILOT_AND_NOT_INFORMED_PORT anomaly PILOT_START_TIME_DELAY_THRESHOLD ← 30 minutes for each vessel ∈ AIS data do for each pilot_vessel ∈ pilot schedule do if current date = pilot_vessel date and isVesselSimilar(vessel, pilot_vessel) then if vessel anomaly contains VESSEL_NOT_INFORMED_PORT then add(vessel anomaly,VESSEL_ORDERED_PILOT_AND_NOT_INFORMED_PORT) end if if pilot_vessel has not received service and pilot_vessel start_operation_time + PILOT_START_TIME_DELAY_THRESHOLD < current time then add(vessel anomaly , VESSEL_NOT_USED_PILOT) end if // To solve the issue for duplicate records with different status in pilot data if pilot_vessel has received service and vessel anomaly contains VESSEL_NOT_USED_PILOT then remove(vessel anomaly, VESSEL_NOT_USED_PILOT) break 41 Algorithm 4 Detect VESSEL_NOT_USED_PILOT and VESSEL_ORDERED_PILOT_AND_NOT_INFORMED_PORT anomaly end if end if end for end for Algorithm 5 Detect UNUSUAL_TRIP_PATTERN anomaly TRIP_COUNT_THRESHOLD ← 2 for each vessel ∈ AIS data do if vessel origin ≠ null and vessel destination ≠ null and vessel type ≠ FERRY and (isRelatedToPorts(vessel origin) or isRelatedToPorts(vessel destination)) and not isInPortsArea(vessel) then count ← getTripHistoryForVessel(vessel) if count ≠ -1 and count < TRIP_COUNT_THRESHOLD then add(vessel anomaly , UNUSUAL_TRIP_PATTERN) end if end if end for Algorithm 6 Detect VESSEL_NOT_ENTERED_PORT anomaly for each port ∈ available ports expected_vessels ← vessels that were expected to enter the port since one hour ago port_vessels ← vessels that are currently available in port for each exp_vessel ∈ expected_vessels do vessel_found ← false for each port_vessel ∈ port_vessels do if port_vessel time ≥ exp_vessel time and isVesselSimilar(exp_vessel, port_vessel) then vessel_found ← true break end if end for if vessel_found ≠ true then add(port_vessel anomaly, VESSEL_NOT_ENTERED_PORT) findVesselMatchInAISData(port_vessel) end if end for end for Algorithm 7 Detect VESSEL_NOT_LEFT_PORT anomaly for each port ∈ available ports port_vessels ← vessels that were expected to leave the port since one hour ago for each port_vessel ∈ port_vessels do for each vessel ∈ AIS data do if isVesselSimilar(vessel, port_vessel) and isMooredInPortsArea (vessel) then add(vessel anomaly , VESSEL_NOT_LEFT_PORT) 42 Algorithm 7 Detect VESSEL_NOT_LEFT_PORT anomaly break end if end for end for end for Algorithm 8 Detect VESSEL_MOORED_IN_PORT anomaly for each vessel ∈ AIS data do if not isInPortsArea(vessel) then vessel_found ← false for each port ∈ available ports do if vessel_found ≠ true then port_vessels ← vessels that are available in port for at least 30 minutes later for each port_vessel ∈ port_vessels do if isVesselSimilar(vessel, port_vessel) and vessel type = port_vessel type and isOtherSimilarVesselInPort(vessel) then add(vessel anomaly , VESSEL_MOORED_IN_PORT) vessel_found ← true break end if end for end if end for end if end for Algorithm 9 isInPortsArea(vessel) in_area ← false for each port ∈ available ports do rectangle ← get the pre-defined port area if vessel position is within the rectangle then in_area ← true break; end if end for return in_area Algorithm 10 isMooredInPortsArea(vessel) in_area ← false for each port ∈ available ports do create a rectangle from pre-defined coordinates for the port area if vessel is not moving and vessel position is within the rectangle then in_area ← true break; 43 Algorithm 10 isMooredInPortsArea(vessel) end if end for return in_area Algorithm 11 isVesselSimilar(vessel_1, vessel_2) VESSEL_NAME_SIMILARITY_THRESHOLD ← 0.03 similarity ← calculate the similarity between the vessel_1 name and vessel_2 name using a string matching technique is_similar ← false if similarity < VESSEL_NAME_SIMILARITY_THRESHOLD then is_similar ← true end if return is_similar Algorithm 12 isRelatedToPorts(name) is_related ← false for each port ∈ available ports do if name = port name or name contains port name or isTripSimilar(name, port name) then is_ related ← true break end if end for return is_related Algorithm 13 isTripSimilar (trip1, trip2) TRIP_SIMILARITY_THRESHOLD ← 0.06 similarity ← calculate the similarity between the two trips using a string matching technique is_similar ← false if similarity < TRIP_SIMILARITY_THRESHOLD then is_similar ← true end if return is_similar Algorithm 14 vesselLeftPortRecently(port, vessel) left_port ← false port_vessels ← vessels that were expected to leave the port since twelve hours ago for each port ∈ available ports do if isVesselSimilar(vessel, port_vessel) then left_port ← true break end if end for 44 Algorithm 14 vesselLeftPortRecently(port, vessel) return left_port Algorithm 15 isTargetedPort(port, vessel) is_targeted ← false horizontal_line ← get the pre-defined horizontal line for the port vertical_line ← get the pre-defined vertical line for the port if vessel position is on the left side of vertical_line and vessel position is below the horizontal_line and vessel heading is to the north east then is_targeted ← true else if vessel position is on the right side of vertical_line and vessel position is below the horizontal_line and vessel heading is to the north west then is_targeted ← true else if vessel position is on the right side of vertical_line and vessel position is above the horizontal_line and vessel heading is to the south west then is_targeted ← true else if vessel position is on the left side of vertical_line and vessel position is above the horizontal_line and vessel heading is to the south east then is_targeted ← true end if return is_targeted Algorithm 16 getTripHistoryForVessel(vessel) vessel_found ← false count ← 0 for each history_vessel ∈ vessel trip history do if isVesselSimilar(vessel, history_vessel) then vessel_found ← true trip ← vessel origin and destination if isTripSimilar(trip, history_vessel trip) then count ← history_vessel trip_count break end if end if end for if vessel_found ≠ true then count ← -1 return count Algorithm 17 isOtherSimilarVesselInPort(ais_vessel) is_similar ← false for each vessel ∈ AIS data do if (vessel name = ais_vessel name and vessel type = ais_vessel type and 45 Algorithm 17 isOtherSimilarVesselInPort(ais_vessel) vessel mmsi ≠ ais_vessel mmsi)) and isInPortArea(vessel) then is_similar ← true end if end for return is_similar Algorithm 18 findVesselMatchInAISData(port_vessel) for each vessel ∈ AIS data do if isVesselSimilar(vessel, port_vessel) and vessel type = port_vessel type and vessel type ≠ FERRY then if port_vessel anomaly = VESSEL_ENTERED_PORT_WITHOUT_NOTICE and not isInPortsArea(vessel) then continue else if port_vessel anomaly = VESSEL_NOT_ENTERED_PORT and (vessel has ARRIVAL_TIME_MISMATCHED or isInPortsArea(vessel)) then continue else add(vessel anomaly , port_vessel anomaly) end if end if end for Algorithm 19 Determine the final type of anomaly for vessels for each vessel ∈ AIS data do //A10 if vessel ∈ smuggling list and vessel anomaly count ≠ 0 then vessel anomaly ← UNDER_SURVEILLANCE_VESSEL continue end if if vessel anomaly count > 1 then //A9 if vessel anomaly contains VESSEL_MOORED_IN_PORT then anomaly ← VESSEL_MOORED_IN_PORT //A1,A5 else if vessel anomaly contains VESSEL_NOT_INFORMED_PORT and vessel anomaly contains UNUSUAL_TRIP_PATTERN then anomaly ← UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_INFORMED _PORT //A2,A5 else if vessel anomaly contains ARRIVAL_TIME_MISMATCHED and vessel anomaly contains UNUSUAL_TRIP_PATTERN then anomaly ← UNUSUAL_TRIP_VESSEL_ARRIVAL_TIME_MISMATCHED //A7,A5 else if vessel anomaly contains VESSEL_NOT_ENTERED_PORT and 46 Algorithm 19 Determine the final type of anomaly for vessels vessel anomaly contains UNUSUAL_TRIP_PATTERN then anomaly ← UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_ENTERED _PORT //A8,A5 else if vessel anomaly contains VESSEL_ORDERED_PILOT_AND_NOT_ INFORMED_PORT and vessel anomaly contains UNUSUAL_TRIP _PATTERN then anomaly ←UNUSUAL_TRIP_PATTERN_AND_VESSEL_ORDERED_PILOT _AND_NOT_INFORMED_PORT //A3,A6 else if vessel anomaly contains VESSEL_ENTERED_PORT_WITHOUT_NOTICE and vessel anomaly contains VESSEL_NOT_LEFT_PORT then anomaly ← VESSEL_ENTERED_WITHOUT_INFORMATION_NOT_LEFT _PORT_ ON_TIME //A2,A4 else if vessel anomaly contains ARRIVAL_TIME_MISMATCHED and vessel anomaly contains VESSEL_NOT_USED_PILOT then anomaly ← VESSEL_ARRIVAL_TIME_MISMATCHED_AND _VESSEL_NOT_USED_PILOT //A2,A4,A5 else if vessel anomaly contains ARRIVAL_TIME_MISMATCHED and vessel anomaly contains VESSEL_NOT_USED_PILOT and vessel anomaly contains UNUSUAL_TRIP_PATTERN then anomaly ← UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_USED_PILOT _AND_VESSEL_ARRIVAL_TIME_MISMATCHED //A4,A5 else if vessel anomaly contains VESSEL_NOT_USED_PILOT and vessel anomaly contains UNUSUAL_TRIP_PATTERN then anomaly ← UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_USED_PILOT //A7,A4 else if vessel anomaly contains VESSEL_NOT_ENTERED_PORT and vessel anomaly contains VESSEL_NOT_USED_PILOT then anomaly ← VESSEL_NOT_ENTERED_PORT_AND_NOT_USED_PILOT //A7,A4,A5 else if vessel anomaly contains VESSEL_NOT_ENTERED_PORT and vessel anomaly contains VESSEL_NOT_USED_PILOT and vessel anomaly contains UNUSUAL_TRIP_PATTERN then anomaly ← UNUSUAL_TRIP_PATTERN_AND_VESSEL_NOT_ENTERED _ PORT_ AND_VESSEL_NOT_USED_PILOT //A8,A4 else if vessel anomaly contains VESSEL_ORDERED_PILOT_AND_NOT_ INFORMED_PORT and vessel anomaly contains VESSEL_NOT_USED _PILOT then anomaly ← VESSEL_ORDERED_PILOT_AND_NOT_INFORMED_PORT _AND_VESSEL_NOT_USED_PILOT 47 Algorithm 19 Determine the final type of anomaly for vessels //A8,A4,A5 else if vessel anomaly contains VESSEL_ORDERED_PILOT_AND_NOT_ INFORMED_PORT and vessel anomaly contains VESSEL_NOT_USED _PILOT and vessel anomaly contains UNUSUAL_TRIP_PATTERN then anomaly ← UNUSUAL_TRIP_PATTERN_AND_VESSEL_ORDERED_PILOT _AND_NOT_INFORMED_PORT_AND_VESSEL_NOT_USED _PILOT else // this part is a default state and executes when the combination of anomalies is not // logical and it is due to the wrong and inconsistent data anomaly ← select one of the anomalies randomly end if vessel anomaly ← anomaly end if end for 48 APPENDIX D: USER INTERFACE Figure D.1. The main view of the ODADS user interface, the map displays the surveillance area, green vessels are Ferry and Passenger, Cargo vessels are cyan and Tankers are yellow. Grey and dark blue colors belong to Tug and pilot vessels, respectively. The four tabs on the left side contain the data related to detected anomalies, ports and pilot timetables, system reports and settings. In order to inform operators about newly detected anomalies, the Anomaly tab on the left side and the anomalous vessels become red. Moreover, previously detected vessels are shown in light red. By clicking on each vessel it is possible to view more information about the vessel and its track. 49 APPENDIX E: ACRONYMS AD Anomaly Detection ADS Automatic Dependant Surveillance AIS Automatic Identification System CET Central European Time COG Course Over Ground D-L Damerau-Levenshtein EET Eastern European Time EEZ Exclusive Economic Zone ELINT Electronic Intelligence ETA Estimated Time of Arrival FN False Negatives FP False Positives HCI Human-Computer Interaction HFSWR High Frequency Surface Wave Radar IMO International Maritime Organization IR Infrared JDL Joint Directors of Laboratories MDA Maritime Domain Awareness MMSI Maritime Mobile Service Identity MS Maritime Surveillance ODADS Open Data Anomaly Detection System OTH Over The Horizon RADAR RAdio Detection And Ranging RMP Recognized Maritime Picture SAR Synthetic Aperture Radar SIGINT SIGnal INTelligence TN True Negatives TP True Positives UTC Coordinated Universal Time VTS Vessel Traffic Service 50 REFERENCES Akselrod, D., Tharmarasa, R., Kirubarajan, T., Zhen Ding, & Ponsford, T. (2009). Multisensor-multitarget tracking testbed. Proceedings of the IEEE Symposium: Computational Intelligence for Security and Defence Applications, Canada, 1-6. doi:10.1109/CISDA.2009 .5356526 Alonso, J. M., Ambur, O., Amutio, M. A., Azañón, O., Bennett, D., Flagg, R., McAllister, D., et al. (2009, May 12). Improving access to government through better use of the web. Retrieved from World Wide Web Consortium website: www.w3.org/TR/ egov-improving/ Andersson, M., & Johansson, R. (2010, November). Multiple sensor fusion for effective abnormal behaviour detection in counter-piracy operations. Paper presented at the International Waterside Security Conference, Carrara, Italy. doi:10.1109/WSSC.2010 .5730221 Andler, S. F., Fredin, M., Gustavsson, P. M., van Laere, J., Nilsson, M., & Svenson, P. (2009). SMARTracIn: A concept for spoof resistant tracking of vessels and detection of adverse intentions. Proceedings of the SPIE - The International Society for Optical Engineering, USA, 7305, 73050G-1-73050G-9. doi:10.1117/12.818567 Avdic, A., Hedström, K., Rose, J., & Grönlund, Å. (Eds.). (2007). Understanding eparticipation : Contemporary PhD eParticipation Research in Europe. Örebro: Örebro universitet. Bick, E. T., & Barock, R. T. (2005). CENTURION harbor surveillance test bed. Proceedings of MTS/IEEE OCEANS, USA, 2, 1358-1363. doi:10.1109/OCEANS.2005. 1639943 Brax, C., Niklasson, L., & Smedberg, M. (2008). Finding behavioural anomalies in public areas using video surveillance data. Proceedings of the Eleventh International Conference on Information Fusion, Germany. doi:10.1109/ICIF.2008.4632410 Carthel, C., Coraluppi, S., & Grignan, P. (2007). Multisensor tracking and fusion for maritime surveillance. Proceedings of the Tenth International Conference on Information Fusion, Canada. doi:10.1109/ICIF.2007.4408025 Chandola, V., Banerjee, A., & Kumar, V. (2009). Anomaly detection: A survey. ACM Computing Surveys, 41(3), 15:1–15:58. doi:10.1145/1541880.1541882 Constantine, L. L., & Lockwood, L. A. D. (1999). Software for use: a practical guide to the models and methods of usage-centered design. Addison Wesley. Dahlbom, A., & Niklasson, L. (2007). Trajectory clustering for coastal surveillance. Proceedings of the Tenth International Conference on Information Fusion, Canada. doi:10.1109/ICIF .2007.4408114 51 Damerau, F. J. (1964). A technique for computer detection and correction of spelling errors. Communications of the ACM, 7(3), 171–176. doi:10.1145/363958.363994 Danu, D., Sinha, A., Kirubarajan, T., Farooq, M., & Brookes, D. (2007). Fusion of over-the- horizon radar and automatic identification systems for overall maritime picture. Proceedings of the Tenth International Conference on Information Fusion, Canada. doi:10.1109/ICIF.2007. 4408147 Dedijer, S., & Jéquier, N. (1987). Intelligence for economic development : an inquiry into the role of the knowledge industry. Oxford: Berg. Di Lallo, A., Farina, A., Fulconi, R., Stile, A., Timmoneri, L., & Vigilante, D. (2006, October). A real time test bed for 2D and 3D multi-radar tracking and data fusion with application to border control. Paper presented at the CIE International Conference on Radar, Shanghai, China. doi:10.1109/ICR.2006.343162 Dietrich, D., Gray, J., McNamara, T., Poikola, A., Pollock, P., Tait, J., & Zijlstra, T. (2009). Open data handbook. Retrieved from opendatahandbook.org/en/ Ding, Z., Kannappan, G., Benameur, K., Kirubarajan, T., & Farooq, M. (2003). Wide area integrated maritime surveillance: An updated architecture with data fusion. Proceedings of the Sixth International Conference of Information Fusion, Australia, 2, 1324-1333. doi:10.1109/ICIF.2003.177391 Edlund, J., Gronkvist, M., Lingvall, A., & Sviestins, E. (2006). Rule-based situation assessment for sea surveillance. Proceedings of the SPIE - The International Society for Optical Engineering, USA, 6242, 624203-1-624203-11. doi:10.1117/12.664410 Fooladvandi, F., Brax, C., Gustavsson, P., & Fredin, M. (2009). Signature-based activity detection based on Bayesian networks acquired from expert knowledge. Proceedings of the Twelfth International Conference on Information Fusion, USA, 436-443. Gad, A. S. (2009). A fuzzy logic-based multisensor data fusion for maritime surveillance - real data testing. Proceedings of the Twenty-Sixth National Radio Science Conference, Egypt. Giompapa, S., Farina, A., Gini, F., Graziano, A., & Di Stefano, R. (2007, April). Computer simulation of an integrated multi-sensor system for maritime border control. Paper presented at the IEEE Radar Conference, Boston, MA, USA, 308-313. doi:10.1109/ RADAR.2007.374233 Giompapa, S., Farina, A., Gini, F., Graziano, A., Croci, R., & Di Stefano, R. (2008, May). Study of the classification task into an integrated multisensor system for maritime border control. Paper presented at the IEEE Radar Conference, Rome, Italy. doi:10.1109/RADAR.2008 .4720807 52 Guerriero, M., Willett, P., Coraluppi, S., & Carthel, C. (2008). Radar/AIS data fusion and SAR tasking for maritime surveillance. Proceedings of the Eleventh International Conference on Information Fusion, Germany. doi.10.1109/ICIF.2008.4632409 Guyard, A. ., Roy, J., & Defence R&D Canada-Valcartier, V. Q. (2009). Towards case-based reasoning for maritime anomaly detection: A positioning paper. Proceedings of the Twelfth IASTED International Conference on Intelligent Systems and Control. USA. Hall, D. L., & Llinas, J. (1997). An introduction to multisensor data fusion. Proceedings of the IEEE, 85(1), 6-23. doi:10.1109/5.554205 Hall, D. L., & McMullen, S. A. H. (2004). Mathematical techniques in multisensor data fusion. Boston, London: Artech House. Hall, M. J., Hall, S. A., & Tate, T. (2000). Removing the HCI bottleneck: how the human computer interface (HCI) affect the performance of data fusion systems. MSS National Symposium on Sensor and Data Fusion (pp. 89-104). USA. Han, J., Kamber, M., & Pei, J. (2011). Data mining: Concepts and techniques. USA: Elsevier. Hatch, M. D., Kaina, J. L., Mahler, R. P., & Myre, R. S. (1998, November). Data fusion methodologies to support theater level and deployable surveillance systems. Paper presented at the Conference Record of Thirty-Second Asilomar Conference on Signals, Systems and Computers, Pacific Grove, CA, USA, 1, 563-567. doi:10.1109/ACSSC.1998 .750926 Henrich, W., Kausch, T., & Opitz, F. (2004). Data fusion for the finnish fast attack craft squadron 2000: Concept and architecture. Proceedings of the Seventh International Conference on Information Fusion, Sweden, 2, 842-847. İnce, A. N., Topuz, E., & Panayirci, E. (1999). Principles of integrated maritime surveillance systems. Boston, Dordrecht, London: Springer. Jaro, M. A. (1972). UNIMATCH: A computer system for generalized record linkage under conditions of uncertainty. Proceedings of the spring joint computer conference, USA, 523–530. doi:10.1145/1478873.1478943 Jaro, M. A. (1989). Advances in record-linkage methodology as applied to matching the 1985 Census of Tampa, Florida. Journal of the American Statistical Association, 84(406), 414-420. doi:10.2307/2289924 Johansson, F., & Falkman, G. (2007, December). Detection of vessel anomalies - a Bayesian network approach. Paper presented at the Third International Conference on Intelligent Sensors, Sensor Networks and Information, Melbourne, Qld., Australia, 395-400. doi:10.1109/ISSNIP.2007.4496876 53 Jouan, A., Valin, P., Gagnon, L., & Bosse, E. (1999). Airborne fusion of imaging and non- imaging sensor information for maritime surveillance. Proceedings of the Conference on Quality Control by Artificial Vision, Canada, 281-286. Lane, R. O., Nevell, D. A., Hayward, S. D., & Beaney, T. W. (2010). Maritime anomaly detection and threat assessment. Proceedings of the Thirteenth International Conference on Information Fusion, UK. Laxhammar, R. (2008). Anomaly detection for sea surveillance. Proceedings of the Eleventh International Conference on Information Fusion, Germany. Lefebvre, E., & Helleur, C. (2001). Multisource information adaptive fuzzy logic correlator for recognized maritime picture. Proceedings of the Fourth International Conference on Information Fusion, Canada, 2, ThB2-19-24. Retrieved from ftp.isif.org/fusion /proceedings/fusion01CD/fusion/searchengine/pdf/ThB23.pdf Lefebvre, E., & Helleur, C. (2004). Automated association of track information from sensor sources with non-sensor information in the context of maritime surveillance. Proceedings of the Seventh International Conference on Information Fusion, Sweden, 2, 1251-1256. Levenshtein, V. I. (1966). Binary codes capable of correcting deletions, insertions, and reversals. Soviet physics-Doklady, 10(8), 707–710. Retrieved from www.freearchive.org/o/964e741877a3ffa8f2195ee253597fbaeed69a207e77df15043980 2fd370b2f8 Mano, J. P., Georgé, J. P., & Gleizes, M. P. (2010). Adaptive multi-agent system for multi- sensor maritime surveillance. In Y. Demazeau, F. Dignum, J. M. Corchado, & J. B. Pérez (Eds.), Advances in Intelligent and Soft Computing Series: Vol. 70. Advances in Practical Applications of Agents and Multiagent Systems (pp. 285-290). Berlin, Heidelberg, Germany: Springer. doi: 10.1007/978-3-642-12384-9_34 Maresca, S., Greco, M., Gini, F., Grasso, R., Coraluppi, S., & Horstmann, J. (2010, June). Vessel detection and classification: An integrated maritime surveillance system in the Tyrrhenian Sea. Paper presented at the Second International Workshop on Cognitive Information Processing, Elba, Italy, 40-45. doi:10.1109/CIP.2010.5604209 Molloy, J. C. (2011). The Open Knowledge Foundation: Open Data Means Better Science. Public Library of Science Biology, 9(12). doi:10.1371/journal.pbio.1001195 Nilsson, M., van Laere, J., Ziemke, T., & Edlund, J. (2008). Extracting rules from expert operators to support situation awareness in maritime surveillance. Proceedings of the Eleventh International Conference on Information Fusion, Germany. doi:10.1109 /ICIF.2008.4632308 Ponsford, A. M., D’Souza, I. A., & Kirubarajan, T. (2009, May). Surveillance of the 200 nautical mile EEZ using HFSWR in association with a spaced-based AIS interceptor. 54 Paper presented at the IEEE Conference on Technologies for Homeland Security, Boston, MA, USA, 87-92. doi:10.1109/THS.2009.5168019 Rhodes, B. J., Bomberger, N. A., Seibert, M., & Waxman, A. M. (2005, October). Maritime situation monitoring and awareness using learning mechanisms. Paper presented at the IEEE Military Communications Conference, Atlantic City, NJ, USA, 1, 646-652. doi:10.1109/MILCOM.2005.1605756 Rhodes, B. J., Bomberger, N. A., Seibert, M., & Waxman, A. M. (2006, October). SeeCoast: Automated port scene understanding facilitated by normalcy learning. Presented at the IEEE Military Communications Conference, Washington, DC, USA. doi:10.1109 /MILCOM.2006.302306 Ristic, B., La Scala, B., Morelande, M., & Gordon, N. (2008). Statistical analysis of motion patterns in AIS Data: Anomaly detection and motion prediction. Proceedings of the Eleventh International Conference on Information Fusion, Germany. doi:10.1109/ICIF .2008.4632190 Riveiro, M., & Falkman, G. (2009). Interactive visualization of normal behavioral models and expert rules for maritime anomaly detection. Proceedings of the Sixth International Conference on Computer Graphics, Imaging and Visualization, China. 459-466. doi:10.1109/CGIV.2009.54 Riveiro, M., & Falkman, G. (2010). Supporting the analytical reasoning process in maritime anomaly detection: Evaluation and experimental design. Proceedings of the Fourteenth International Conference Information Visualisation, UK, 170-178. doi:10.1109/IV.2010.34 Riveiro, M., Falkman, G., & Ziemke, T. (2008a). Visual analytics for the detection of anomalous maritime behavior. Proceedings of the Twelfth International Conference Information Visualisation, UK, 273-279. doi:10.1109/IV.2008.25 Riveiro, M., Falkman, G., & Ziemke, T. (2008b). Improving maritime anomaly detection and situation awareness through interactive visualization. Proceedings of the Eleventh International Conference on Information Fusion, Germany. Riveiro, M., Johansson, F., Falkman, G., & Ziemke, T. (2008). Supporting maritime situation awareness using self organizing maps and gaussian mixture models. Proceedings of the Tenth Scandinavian Conference on Artificial Intelligence, 84–91. Retrieved from www.his.se/PageFiles/34259/SCAI2008.pdf Roy, J. (2008). Anomaly detection in the maritime domain. Proceedings of SPIE - The International Society for Optical Engineering, USA, 6945, 69450W-1-69450W-14. doi:10.1117/12.776230 55 Roy, J. (2010). Rule-based expert system for maritime anomaly detection. Proceedings of SPIE - The International Society for Optical Engineering, USA, 7666, 76662N-1- 76662N-12. doi:10.1117/12.849131 Roy, J., & Davenport, M. (2010, November). Exploitation of maritime domain ontologies for anomaly detection and threat analysis. Paper presented at the Waterside Security Conference, Carrara, Italy. doi:10.1109/WSSC.2010.5730278 Salton, G., & McGill, M. J. (1986). Introduction to Modern Information Retrieval. New York, NY, USA: McGraw-Hill, Inc. Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002). Experimental and quasi- experimental designs for generalized causal inference. Houghton Mifflin. van Laere, J., & Nilsson, M. (2009). Evaluation of a workshop to capture knowledge from subject matter experts in maritime surveillance. Proceedings of the Twelfth International Conference on Information Fusion,USA, 171-178. Vespe, M., Sciotti, M., & Battistello, G. (2008). Multi-sensor autonomous tracking for maritime surveillance. Paper presented at the International Conference on Radar, Adelaide, SA, Australia, 525-530. doi:10.1109/RADAR.2008.4653980 Vespe, M., Sciotti, M., Burro, F., Battistello, G., & Sorge, S. (2008). Maritime multi-sensor data association based on geographic and navigational knowledge. Paper pesented at the IEEE Radar Conference, Rome, Italy. doi:10.1109/RADAR.2008.4720782 Winkler, W. E. (1990). String comparator metrics and enhanced decision rules in the fellegi- sunter model of record linkage. Retrieved from www.eric.ed.gov/PDFS /ED325505.pdf Yin, R. K. (2009). Case study research: Design and methods. Sage Publications. Zamora, E. M., Po llock, J. J., & Zamora, A. (1981). The use of Trigram Analysis for Spelling Error Detection. Information Processing and Management 305–316. 1. Introduction 1.1 Problem Statement 1.2 Research Questions 1.3 Aims and Objectives 1.4 Contribution 1.5 Outline 2. Background 2.1 Terminology 2.2 JDL Data Fusion Model 2.3 Related Work 3. Research methodology 3.1 Open Data in the MS Domain 3.2 Identification of Maritime Anomalies Framework Design 3.4 Implementation 3.4.1 Data Description 3.4.1.1 AIS data 3.4.1.2 Ports and Pilots data 3.4.2 Data Collector Module 3.4.3 Database 3.4.4.1 Algorithms 3.4.4.2 String matching techniques 3.4.5 Display Client 3.5 Experimental Design 3.6 Validation Design 4. Validity threats 5. Verification 6. Experimental Results 8. Discussion 9. Conclusion and Future Work Appendix A: Open and Closed Data Sources Appendix B: Ports Regions Appendix C: Algorithms Appendix D: User Interface Appendix E: Acronyms References 
work_wfsjha4f5nb7vih3n7vddj25zu	----	ESWA2DOI.dvi Prediction of Survival Probabilities with Bayesian Decision Trees V. Schetinin, L. Jakaite Computer Science Department and Technology, University of Bedfordshire, UK W. Krzanowski College of Engineering, Mathematics and Physical Sciences, University of Exeter, UK Abstract Practitioners use Trauma and Injury Severity Score (TRISS) models for pre- dicting the survival probability of an injured patient. The accuracy of TRISS predictions is acceptable for patients with up to three typical injuries, but unacceptable for patients with a larger number of injuries or with atypi- cal injuries. Based on a regression model, the TRISS methodology does not provide the predictive density required for accurate assessment of risk. Moreover, the regression model is difficult to interpret. We therefore consider Bayesian inference for estimating the predictive distribution of survival. The inference is based on decision tree models which recursively split data along explanatory variables, and so practitioners can understand these models. We propose the Bayesian method for estimating the predictive density and show that it outperforms the TRISS method in terms of both goodness-of-fit and classification accuracy. The developed method has been made available for evaluation purposes as a stand-alone application. Keywords: Bayesian prediction, survival probability, Markov chain Monte Carlo, classification tree, trauma care. DOI: 10.1016/j.eswa.2013.04.009 1. Introduction For patients alive on arrival at a hospital, the probability of survival is calculated by using a logistic regression model that employs the Trauma and Injury Severity Score (TRISS) system [5, 4, 27, 12, 28]. TRISS-based models Preprint submitted to Expert Systems With Applications April 22, 2013 consider up to three most severe injuries which a patient can obtain in six regions of the body: head, face, chest, abdomen, extremities, and external (skin, subcutaneous tissue and burns). The model includes both continuous and categorical screening tests: the first type includes age, systolic blood pressure, and respiratory rate, while the second type includes the severity scores of injuries a patient can obtain in the six body regions, as well as Glasgow Coma Scale (GCS) and type of injury. The screening tests are performed on the patient’s arrival by a trained scorer. The calculation of survival probabilities has been made available online as a TRISS Calculator [7]. The screening tests are used to form two aggregated predictors, Injury Severity Score (ISS), and Revised Trauma Score (RTS). The use of such aggregated predictors has revealed unexplained fluctuations of ISS over ob- served (or actual) probabilities of survival, and some researchers have raised a concern about its predictive ability [38, 4, 28, 1, 39]. The TRISS determines the probability of survival, Ps, using the following logistic regression model: Ps = 1/(1 + e−b), (1) where b = b0 + b1 × RTS + b2 × ISS + b3 × A. Here b0, b1, b2, and b3 are the regression parameters, and A is is the dichotomised age: A = 0, if age < 55, and A = 1, otherwise. The regression parameters were defined for blunt and penetrating injuries in [5]. The analysis of actual survival against predicted probabilities is defined as calibration, and a visualized relationship between values of these probabilities is a calibration curve, see e.g. [22, 53]. An ideal calibration curve is a 45 degree line with zero intercept. In this light the calibration curve for TRISS models has been found to deviate significantly from the ideal [28, 44]. It has been shown that the accuracy of TRISS models can be improved by updating the model parameters on new data, see e.g. [8, 36]. This ap- proach can be implemented efficiently when an appropriate model structure is known. However in practice such a model is difficult to identify. When goodness or fitness of a model can be measured in terms of its likelihood, the model can be fitted to data with the likelihood maximization method. In practice, this approach requires much effort to overcome the optimization problems while the desired improvement cannot be guaranteed [16, 23, 2, 37]. Despite the problems, the accuracy of TRISS predictions has been found acceptable when the types and severities of patients injuries are typical. How- 2 ever, for cases with four or more injuries as well as with some atypical com- binations of injuries, the accuracy could be improved [8, 27, 28]. It is highly desirable for practitioners to estimate the uncertainty in pre- dictions of survival. In general, estimates of uncertainty are required to min- imize risks of mistaken decisions and, in particular, to estimate confidence intervals. These intervals can be accurately estimated when a predictive probability density is fully known, but the desired estimates cannot be pro- vided within a concept such as TRISS which employs a maximum likelihood method to fit a logistic regression to data [43, 1]. The above problems motivated us to develop a Bayesian method for pre- diction of survival. In our research we used the US National Trauma Data Bank (NTDB) which is the major data source of records of injured patients admitted to hospitals and emergency units; these data are available for re- search from [12]. The NTDB data include information about a patients age, gender, type and regions of injuries along with some clinical and background information about a patients state. The NTDB also includes information about TRISS prediction and outcome of care, alive or died, for each patient. In our research we use well-known Decision Tree (DT) models which are induced from given data in such a way that features which make a distin- guishable contribution to the model outcome are selected. These features make axis-parallel partitions, and as a result users find the DT models inter- pretable [6, 18]. We tested the proposed Bayesian method on a set of patients registered in the NTDB with multiple injuries. We expected that in such cases the proposed method would significantly outperform the TRISS method. How- ever, in this research we did not generalize our method to the entire NTDB population including about 2 millions of patients because such a general- ization would require much more intensive computational experiments. For evaluation of our method, we developed a Bayesian calculator of survival as a stand-alone application available from [51]. 2. Related Work In the related literature we found a number of machine learning and sim- ulation methods which are competitive to conventional statistical methods. Simulation was principally via Markov chain Monte Carlo (MCMC) meth- ods, while ML methods included artificial neural networks and support vector 3 machines which are well known for providing non-linear fitting of models to data [3, 56]. The goodness-of-fit or calibration of predictive models produced by these methods was measured in terms of least square error. As for conventional statistical methods, the performance was measured in term of the receiver operating characteristic curve (ROC), or more specifically, by the area under this curve (AUC), see e.g. [18, 31]. The performance was also assessed in terms of accuracy of classification or discrimination measured as the ability of a model to discriminate survived patients from ones that died [27]. 2.1. Machine Learning Methods The study by Sujin et al[55] compared the mortality prediction models built with different Machine Learning methods on a data set collected by the University of Kentucky Hospital. The authors of this study compared the ar- tificial neural networks, support vector machines, DT, and conventional logis- tic regression models. The data set used for the experiments included 38,474 patients, information about which was represented by 41 variables. It was reported that only 15 out of these variables made a significant contribution to the prediction, including blood pressure, respiration rate, GCS, comor- bidity, and blood serum composition. The best performance was achieved with the DT model, AUC = 0.892, whilst the AUC for the artificial neu- ral networks, support vector machines, and logistic regression models were 0.874, 0.876, and 0.871, respectively. All the methods were implemented in the SPSS Clementine software [24]. In [52], an artificial neural network-based approach was described. This approach was compared with a logistic regression model on 13,164 patients whose physiological information was represented by 17 variables. Both the neural networks and the regression models were built to provide non-linear fits to the data, and their goodness-of-fit (or calibration) was evaluated in terms of least square error. It was reported that the resultant neural network model slightly outperformed the logistic regression in terms of AUC. The authors of study [25] employed a probabilistic artificial neural net- work for predicting mortality in emergency room. The network with 10 input variables was trained with a genetic algorithm. The calibration was evalu- ated using Hosmer-Lemeshow statistic. The study conducted on a data set of 533 patients revealed that the performances of the neural network and logistic regression models were comparable. 4 In [40], such machine learning methods as Naive Bayesian classifier, DT, support vector machine and artificial neural network were used for predicting survival of burn patients. The authors reported that the prediction accuracies of these methods were comparable. All the methods were implemented in the Weka software [21]. The data were represented by 10 features, namely age, gender, and percentages of burn in the eight body regions. The DT was induced with the C4.5 algorithm [41]. The study by Clermont et al [11] compared an artificial neural network method with a logistic regression model for predicting mortality in the in- tensive care units. The study was carried out on 1,647 patient cases repre- sented by 24 input features. The features included patient’s age, the values of 16 physiologic variables including temperature, heart rate, blood pres- sure, respiratory rate, oxygenation, composition of the blood serum, GCS, as well as binary indicators for the absence or presence of chronic conditions. The goodness-of-fit was evaluated using Hosmer-Lemeshow statistics, and the comparisons were made in terms of classification accuracy as well as in terms of AUC value. It was shown that the both methods had similar performances. In [2], an artificial neural network was compared against a TRISS model on the UK Trauma data represented by 16 anatomical and physiological pre- dictor variables. Both models provided the similar accuracy of classification, but the TRISS model showed better performance in terms of AUC, 0.941 ver- sus 0.921. The neural network model provided a better calibration in terms of Hosmer-Lemeshow statistics, 58.3 versus 105.4. It was also found that the head injury, age, and chest injury made the most important contribution to the outcome, while respiration rate, heart rate, and systolic blood pressure were underestimated (their contribution was less important). The authors concluded that the TRISS model was adequate but not optimal. A Bayesian belief network described in [54] was built to predict morbidity and outcomes in wounded patients. Bayesian belief networks are known as graphical models capable of explaining relationships between predictor vari- ables; such models cannot be developed with conventional logistic regression methods. The study conducted on a data set of 22 patients revealed that the logistic regression outperformed the Bayesian belief networks in terms of AUC. The described Bayesian belief network was developed to estimate one of three types of outcomes: likelihood of infection, requirement for intensive care, and impaired wound healing. The relationships between clinical vari- ables and outcomes were identified using DecisionQ FasterAnalytics Bayesian 5 modelling software[13]. It was found that the likelihood of infection could be estimated using serum albumin, injury severity score, and initial requirement for blood transfusion. The likelihood of intensive care admission was esti- mated using the blood transfusion requirement, physiological variables and serum biomarkers. Impaired wound healing was estimated using the indi- cator of intensive care admission, serum biomarkers and estimated hospital length of stay. The network consisted of 12 nodes, with two-to-three states per node. 2.2. Simulation The authors of review [30] found that conventional modelling techniques used in critical care could be improved by simulation methods such as MCMC. They analysed the usefulness and limitations of applications of these methods presented in the related literature and concluded that simulation provided practitioners with additional information on risks. In [33], conventional and Bayesian logistic random effects regression mod- els were compared for predicting outcomes on a data set of 8,509 patients with Traumatic Brain Injury. The Bayesian method has been implemented in statistical packages such as WinBUGS [34], MLwiN [42], MCMCglmm [20], and SAS [46]. It was reported that both methods provided similar prediction accuracy. The results of the Bayesian method were critically dependant on the chosen priors as well as on the sampler’s settings which can affect the convergence of the Markov Chain if given inappropriately. 3. Bayesian Predictions When Bayesian inference is employed for predictions, it is typically as- sumed that there exist a number of models which can appropriately approx- imate the relationship between predictor variables and output variable (or outcome) observed in given data. Given models with parameters, we can fit them to the data. It is most often the case that none of the models describes the true relationship between input and output variables. However, we as- sume that averaging over the models could result in more accurate approxi- mation to the true relationship. The most efficient averaging over models is achieved within the Bayesian methodology. However, when Bayesian meth- ods are applied to real data, a number of problems are raised. One specific problem we address in the paper is associated with Bayesian averaging over hierarchical models, such as DTs. 6 When DT models are collected in an ensemble, for interpretation purpose a single DT model with Maximum a Posterior likelihood can be selected as described in [10]. However our technique proposed in [47] for interpretation of an ensemble of DT models has been capable of finding a single DT model providing a better accuracy of predicting survival probabilities. The methodology of Bayesian averaging over DT models has been made computationally feasible with the MCMC method [9, 14]. This method aims to explore a posterior density of model parameters by making random walk proposals. The desired density is approximated by drawing samples from areas with high posterior density of the model parameters (so-called areas of interest). In the case of DT models, a model parameter space is often of variable size, and the MCMC method is extended to Reversible Jump (RJ) [19]. The desired approximation is then achieved when the RJ MCMC algorithm can explore all areas of the posterior density in a model parameter space of a variable size. However, a posterior density function is often multimodal and the detailed exploration of the areas of interest cannot be achieved in a reasonable time, see e.g. [43]. This affects the accuracy of approximation as the Bayesian model averaging tends to act more as model selection [17, 14]. In practice, when DT models (and hence a model parameter space) are large, results of Bayesian averaging are critically dependent on prior informa- tion as shown in [14, 43]. When prior information is available, the averaging is mostly done over areas of high posterior, and the estimates of the desired predictive density are likely to be accurate. However, when prior information is absent, the areas of possible interest cannot be specified and hence detailed exploration may not be possible in a reasonable time [14, 45]. As a result, the MCMC sampler cannot explore all possible areas of interest proportionally to the posterior density of the parameters. One possible reason for the above disproportion is that the RJ MCMC algorithm tends to simulate samples from an oversized model parameter space [9, 14]. In our previous work we attempted to reduce factors that cause the RJ MCMC to sample from overgrown DT models and proposed a sweeping strategy [48]. In the experiments on the benchmark problems, we described in [49], this strategy has been shown more efficient than the restarting [9] and restricting [14] MCMC strategies. In these experiments the proposed strategy also outperformed the conventional random forest [15]. In our previous research, we also observed that when prior information on predictors was absent, the posterior density cannot be explored in detail, 7 and some DT models were disproportionally sampled [26]. When in the post-analysis phase we evaluated the frequencies of using predictor variables in the DT models, we found that some predictors were employed rarely. We assumed that these predictors made a weak contribution to the outcome. When we removed DT models which explored such weak predictors from the ensemble, we observed a decrease in entropy of the model mix. The fact of decreasing entropy has been proven as an indicator of improving estimates of the predictive density, see e.g. [32]. The above analysis motivated us to extend the methodology of Bayesian averaging over DT models for predicting survival probabilities. We attempt to improve the RJ MCMC method used for implementation of the Bayesian methodology. We are also interested in exploring the importance of the pre- dictive variables within the proposed method, and expect that the posterior information on predictors can be useful to optimize existing procedures of scoring injury severities. The methodology of Bayesian averaging over DT models is well developed and described in the literature (see e.g. [9, 14]). The details of the Bayesian method and the proposed MCMC strategy are described in the Appendix. In the following sections we explore the proposed strategy on a data set of patients from the NTDB. We attempt to improve the accuracy of estimates of predictive density. 4. Trauma Data We used data on patients registered in the NTDB who were alive on ar- rival at hospitals, and whose survival probabilities have been calculated with the TRISS method. Table 1 shows the screening tests denoted as variables x1, . . . , x17 which were used for the TRISS predictions. Variables x1 (age), x4 (blood pressure), and x5 (respiration rate) are continuous, and the others are categorical. The predicted output is the discharge status, y = {0, 1}. Figure 1 shows that the TRISS predictions and the observed survival probabilities progressively decrease with the number of injuries ranging from 1 to 20. The actual survival of patients with one injury was 0.98 while for patients with 20 injuries it was 0.71. The proportions of these patients were 0.17 and 0.0002, respectively. We see that the predicted values are below the actual frequencies and that the difference between their values, or prediction error, tends to increase with the number of injuries obtained by a patient. 8 Table 1: Screening Tests Test Name Range x1 Age 0-99 x2 Gender female, male x3 Injury type penetrating, blunt, burn x4 Blood pressure 0-299 x5 Respiration rate 0-59 x6 GCS Eye 1-5 x7 GCS Verbal 1-5 x8 GCS Motor 1-6 x9 Head severity 0-6 x10 Face severity 0-6 x11 Neck severity 0-6 x12 Thorax severity 0-6 x13 Spine severity 0-6 x14 Abdomen severity 0-6 x15 Upper extremity severity 0-6 x16 Lower extremity severity 0-6 x17 External severity 0-6 y Discharge status alive, dead The majority of patients were registered with 1 to 3 injuries and the TRISS predictions for this largest group of patients are close to the observed survival. The proportions of patients with a larger number of injuries are smaller and so the TRISS model does not fit these data well. In our research we attempt to improve the accuracy of predictions for such groups of pa- tients. In particular, we target a group which includes all the 14,840 patients registered in the NTDB with 11 to 15 injuries. This set does not include about 2% of patient’s records we found with one or more missing values as we did not attempt to fill the absent values in this research. Figure 2 shows the calibration curve of the TRISS model for this group of patients. We see that the observed probabilities are significantly higher than the predicted values. The difference is largest for patients with predicted survival between 0.1 to 0.5. We can evaluate quantitatively the goodness-of-fit (or calibration) of the 9 0 2 4 6 8 10 12 14 16 18 20 0.4 0.2 0.6 0.8 1 1.2 Number of injuries A ct u a l a n d T R IS S s u rv iv a l TRISS Prediction Actual Survival Figure 1: Observed and predicted survival probabilities for patients with differ- ent numbers of injuries. The TRISS predictions and the observed survival probabilities progressively decrease with the number of injuries ranging from 1 to 20. TRISS model for this group of patients by using the Hosmer-Lemeshow (HL) statistic used in the related literature [27, 4, 53]. When we split the data into 30 equally sized subsets, its value was 3680.5. 5. Experiments We used the above data to test the proposed MCMC simulation strat- egy. The experiments were run within 3-fold cross-validation. We compared the TRISS and Bayesian predictions in terms of classification accuracy, and Hosmer-Lemeshow statistic. The results were obtained using settings which allowed us to achieve a stationary distribution of the Markov chain and an efficient acceptance rate, while DT models were of a reasonable size. These settings are described in the Appendix. Figure 3 shows the calibration curve for the proposed Bayesian model. We can see that the Bayesian predictions are much closer to the observed survival rate than the TRISS predictions shown in Figure 2. The value of the Hosmer-Lemeshow statistic was significantly reduced from 3680.5 to 93.4. We compared the classification accuracies of the Bayesian and TRISS 10 Figure 2: Calibration curve for TRISS model for patients with 11-15 injuries. The observed probabilities of survival are significantly higher than the predicted values. The difference is largest for patients with predicted survival between 0.1 to 0.5. methods. Table 2 shows the classification accuracy (AC), true positive (TP), false negative (FN), true negative (TN), false positive (FP), sensitivity (SE), and specificity (SP) which were calculated by assigning the outcome alive if a patient’s survival prediction is higher than 0.5. We can observe that the accuracy of the Bayesian method is higher by 6%. Having a significantly higher FN rate (0.129 versus 0.024), the TRISS model provides the better TN rate (0.121 versus 0.078). Using the standard bootstrap method, we found that the p-values were less than 0.001 for all the statistical tests. Table 2 shows that the accuracy of the proposed Bayesian method is higher than that of the TRISS method by approx 6%. This part of classi- fication error is reducible, and this error appears when a decision threshold t used for classification differs from the optimal and thus becomes biased. Setting threshold t below the optimal value decreases the FP and increases the FN rates and, vice versa, setting the t above the optimal value increases the false positive and decreases the false negative rates. Therefore a value of t can be set so as to find a compromise between the SE and SP of a diag- 11 Figure 3: Calibration curve for the proposed Bayesian model. The Bayesian predictions are much closer to the observed survival rate than the TRISS predictions. nostic method, calculated as SE=TP/(TP+FN) and SP = TN/(TN+FP). The sensitivity, SE, shows how the diagnostic method is sensitive to posi- tive results, and the specificity, SP, shows how the method specifies negative results. In our case, positive results are associated with the conditions of a ”died” patient, and negative with those of an ”alive” patient. The sensitivity of the Bayesian method shown in Table 2 for the threshold t = 0.5 is 0.4868, that is less than the 0.7588 provided by the TRISS method. To achieve a similar sensitivity rate, we can increase the threshold t. For example, we can set t = 0.74 to increase the sensitivity to 0.7496, accepting a small decrease in the accuracy from 0.8936 to 0.8671, as shown in the third line of Table 2. In this case the differences between the values of confusion matrices for the Bayesian and TRISS methods remain significant with p-value < 0.001. We also tested the restarting MCMC strategy of sampling DT models [9, 14] on the same set of patients. The Hosmer-Lemeshow statistic was 17% higher on average than that for the proposed method. The classification accuracy was not significantly lower, however the uncertainty of estimates 12 Table 2: Performances of the TRISS and Bayesian Model (BM) with thresholds t Model AC TP FN TN FP SE SP TRISS 0.832 0.121 0.038 0.708 0.129 0.759 0.846 BM, t = 0.50 0.894 0.078 0.082 0.816 0.024 0.487 0.971 BM, t = 0.74 0.867 0.120 0.040 0.747 0.093 0.750 0.890 of the predictive density in terms of entropy of model mixing was signifi- cantly higher than that for the proposed method. This allows us to conclude that the proposed MCMC simulation strategy is capable of providing better conditions for Bayesian averaging over DT models. 6. Importance of Screening Tests The ensemble of DT models collected during MCMC simulation allows us to estimate the contribution of the predictive variables (screening tests) to the outcome. The importance of the variables can be estimated in terms of frequencies of using them in the ensemble. These frequencies (or posterior probabilities) are shown in Table 3. We can see that the most important contribution is made by the variables x4 (Blood pressure), x9 (Head severity), and x15 (Upper extremity severity). By contrast, the variables x5 (Respiration rate), x11 (Neck severity), and x17 (External severity) are least important, and therefore their contribution can be insignificant for predicting the survival of patients in the target group. 7. Bayesian Calculator of Survival Probabilities For evaluating the proposed Bayesian method, we developed a Calculator for predicting survival and tested it on the target data set described in Sec- tion 4. The Calculator allows the user to compare the Bayesian and TRISS predictions for a random patient drawn from the data set. The comparison is made in terms of prediction accuracy as described in Section 5. The user can also input new values of the screening tests to make predictions. It is important that the Calculator allows the user to estimate the pre- dictive probability density in order to assess the confidence intervals, which are associated with risk of making mistaken decisions. These estimates are 13 Table 3: Importance of Screening Tests Test Name Importance x4 Blood pressure 0.157 x9 Head severity 0.122 x15 Upper extremity severity 0.115 x1 Age 0.107 x13 Spine severity 0.093 x12 Thorax severity 0.081 x16 Lower extremity severity 0.064 x2 Gender 0.053 x10 Face severity 0.052 x8 GCS Motor 0.050 x14 Abdomen severity 0.045 x6 GCS Eye 0.018 x7 GCS Verbal 0.014 x3 Injury type 0.014 x5 Respiration rate 0.006 x11 Neck severity 0.004 x17 External severity 0.003 made individually for each patient, whilst the TRISS method is unable to provide such estimates. Figure 4 presents a screenshot of the calculator interface. The first column in the table Screening Tests shows the 17 screening tests that are described in Table 1. The second column shows the ranges of these tests. The third column displays values which the user can input or edit within the specified ranges. The graph Predicted Probabilities of Survival displays the probabilities of survival for a patient with the given screening tests. Each of the predicted probabilities can be interpreted as a hypothesis which is tested on the data set in the context of Bayesian inference. The bars on the graph show the observed probabilities of these hypotheses. The estimates of the predictive density shown in the graph provide all the information required to calculate the confidence intervals. Consider the example shown in Figure 4 for patient No 11311 with TRISS 14 Figure 4: Bayesian Calculator Screenshot. The table Screening Tests shows the 17 screening tests and the values of these tests which the user can input or edit within the specified ranges. The graph Predicted Probabilities of Survival shows the probabilities of survival for a patient. The estimates of the predictive density shown in the graph provide the information required to calculate the confidence intervals. survival probability 0.680 and outcome alive. For this patient, the Calcula- tor predicts a survival probability of 0.605. As this value exceeds 0.5, the predicted outcome is alive. The locations and heights of the bars shown in the graph present the estimates of predictive density. All the bars on the left from the 0.5 mark on the x-axis represent low probabilities of survival associated with the outcome died whereas all the bars on the right represent high probabilities of survival. Observing these probabilities, the user can analyse the risk for this patient. In this example, the sum over the first bars (on the left from 0.5) is smaller than the sum over the bars on the right. The substantial proportion of the former bars warns the user about a high death risk attached to this prediction. These bars make the predicted probability distribution wider and the uncertainty interval larger. In addition to a predicted probability of survival, the user can analyse the confidence interval calculated for a given patient. The lengths of the intervals can be associated with the difficulty of treatment for patients – the larger the interval, the more difficult is the case. We also observed some bars with unex- 15 pectedly low values of the observed probabilities, see the bar located around 0.51 in the graph, which is a case of a multimodal distribution. The user can conclude that patients with such predictions require special attention, and an investigation of similar cases could provide additional information about possible causes of the uncertainty. The Calculator can be downloaded from a web page [51] to be installed and run on a Windows XP 32-bit or Linux 64-bit machine The Bayesian risk assessments are computationally expensive, and so a high performance machine with a 64-bit processor and 4 GB memory is recommended. The Calculator can be used for evaluation purposes and allows users to observe both the predictive distribution and the calculated confidence intervals. 8. Discussion In the related publications reviewed in Section 2, the main Machine Learn- ing methods (Artificial Neural Networks, Support Vector Machines, and De- cision Trees) have not been found to outperform the TRISS method in terms of accuracy of predicting survival on the trauma data. In the Introduction we noted that the Machine Learning methods cannot provide an assessment of the predictive posterior distribution as proposed in our work. None of these methods is currently used in emergency care. It is important to note also that the use of Machine Learning methods (in particular Support Vec- tor Machines) requires careful and accurate setting of parameters (such as kernels) as well as of a proper model structure. The use of Artificial Neural Networks requires using a proper back-propagation algorithm and its learn- ing parameters. Finding proper settings typically requires the analysis of results of experiments run with different parameters of a machine Learning method and a model structure. However when such experiments are done, one might question whether the settings were explored within only a limited range of possible values. This is what we called the problem of likelihood maximization. In particular, we found that a single classification tree model [6], con- structed with the Matlab Statistics Toolbox [35], can provide only the fre- quentist estimation of survival, although its discrimination performance was comparable with that of the proposed method. As we discussed in the Introduction, when a model structure is defined properly we can use the likelihood maximization method to fit the model to a given data set. Optimal results will be guaranteed if the model likelihood 16 function is unimodal. However, often we do not know whether our model structure is given properly or whether its likelihood function is unimodal. A more practical scenario is when we have a few models, each of which provides a suitable goodness-of-fit in a particular area of the model parameter space. The use of Bayesian methodology in such cases allows us to average outcomes of the models according to their likelihood values and given prior information. The Bayesian average has been shown improving results even if there is no prior information on models and the so-called uniform (or flat) prior is used, see e.g. [14]. The improvement is, however, achievable if the models do not suffer from the overfitting problem as discussed in [17]. Therefore the average provides a practical way to combine prior information (if available) with a set of models which the user could consider competitive to a single model believed to be of a suitable structure. In this context, the strength of the Bayesian method is in averaging over possible models regardless of how much we know about a proper model structure and its parameters. When a single DT model is built, its likelihood is estimated for a new split made by an assigned feature. If the likelihood is increased, the new split is included in the model, and the assigned feature is assumed making a distinguishable contribution to the classification. Otherwise, the new split is unlikely included in the model as the splitting feature is unable to make a significant contribution. Therefore, a resultant DT model includes only fea- tures that make a significant contribution. The importance of these features cannot be evaluated within a single DT. However, the MCMC method allows us to generate DT models of different configurations, and so we can calculate the frequencies of using the features in these models; these frequencies allow us to estimate the desired feature importance. 9. Conclusion We analysed conventional TRISS-based regression models for predicting survival probabilities of injured patients. In the related literature we found firstly that the prediction accuracy of the TRISS method can be improved and secondly that such attempts have been made by employing the ML and simulation methods. However, we have not found evidences that any of these attempts significantly improved the TRISS method. Based on a regression model, the TRISS method cannot provide the esti- mates of predictive probability density that is required to evaluate confidence intervals. The ISS, which is used as an aggregated predictor for TRISS, is ob- 17 served with unexplainable fluctuations and so may be misleading. Moreover practitioners find that the TRISS models are difficult to interpret. This moti- vated us to explore Bayesian model averaging over DT models for predicting survival. In this paper, we analysed the implementation of the Bayesian method- ology with RJ MCMC simulation and found that in the burn-in simulation phase DT models tend to grow excessively. The existing MCMC strategies were unable to manage the excessive growth efficiently in terms of diversity of model mixing. Moreover these strategies require additional settings which have to be adjusted experimentally. To provide better conditions for detailed exploration of the posterior den- sity during the simulation, we proposed a sweeping strategy. This strategy was tested on a large set of patients registered in the NTDB with multi- ple injuries. The results showed that Bayesian model averaging significantly outperforms the TRISS-based model in terms of goodness-of-fit and classifi- cation accuracy. Moreover, the proposed strategy has slightly outperformed the conventional MCMC strategy. The use of DT models for the Bayesian averaging allowed us to estimate the contribution of each screening test to the outcome. This is a reasonable argument for practitioners to use the DT models discussed in Section 3 in the context of the transparency of models and their ability to select important explanatory variables. The above results allow us to conclude that the proposed method is ca- pable of improving the accuracy of predictions for survival of a patient with multiple injuries. The desired confidence intervals can be accurately esti- mated for each patient. Information about the importance of screening tests could be useful for cost analysis and for further improvement of the prediction accuracy. References [1] T. Bailey, R. Everson, J. Fieldsend, W. Krzanowski, D. Partridge, V. Schetinin, Representing classifier confidence in the safety critical do- main an illustration from mortality prediction in trauma cases, Neural Computing and Applications 16 (2007) 1–10. [2] D. Becalick, T. Coats, Comparison of artificial intelligence techniques with UK TRISS for estimating probability of survival after trauma, Journal of Trauma 51 (2001) 123–133. 18 [3] C.M. Bishop, Pattern Recognition and Machine Learning, Springer, 2007. [4] O. Bouamra, A. Wrotchford, S. Hollis, A. Vail, M. Woodford, F. Lecky, A new approach to outcome prediction in trauma: A comparison with the TRISS model, Journal of Trauma 61 (2006) 701–710. [5] C.R. Boyd, M.A. Tolson, W.S. Copes, Evaluating trauma care: The TRISS method, Journal of Trauma 27 (1984) 370–378. [6] L. Breiman, J. Friedman, R. Olshen, C. Stone, Classification and Re- gression Trees, Chapman and Hall, 1984. [7] K. Brohi, TRISS - Overview and desktop calculator, http://www.trauma.org/index.php/main/article/387/, 2012. [8] M. Chawda, F. Hildebrand, H. Pape, P. Giannoudis, Predicting outcome after multiple trauma: which scoring system?, Injury 35 (2004) 347–358. [9] H. Chipman, E. George, R. McCullock, Bayesian CART model search, Journal of American Statistics 93 (1998) 935–960. [10] H.A. Chipman, E.I. George, R.E. McCulloch, Bayesian treed models, Mach. Learn. 48 (2002) 299–320. [11] G. Clermont, D.C. Angus, S.M. DiRusso, M. Griffin, W.T. Linde- Zwirble, Predicting hospital mortality for patients in the intensive care unit: a comparison of artificial neural networks with logistic regression models, Critical Care Medicine 29 (2001) 291–296. [12] Committee on Trauma. American College of Surgeons, NTDB Version 7.2, http://www.facs.org/trauma/ntdb/ntdbapp.html, 2007. [13] DecisionQ Corporation, Turning data into decisions, http://www.decisionq.com, 2010. [14] D. Denison, C. Holmes, B. Mallick, A. Smith, Bayesian Methods for Nonlinear Classification and Regression, Wiley, 2002. [15] T.G. Dietterich, An experimental comparison of three methods for con- structing ensembles of decision trees: Bagging, boosting, and random- ization, Mach. Learn. 40 (2000) 139–157. 19 [16] S. DiRusso, T. Sullivan, C. Holly, S. Cuff, J. Savino, An artificial neural network as a model for prediction of survival in trauma patients: Vali- dation for a regional trauma area, Journal of Trauma 49 (2000) 220–223. [17] P. Domingos, Bayesian averaging of classifiers and the overfitting prob- lem, in: The 17th International Conference on Machine Learning, Mor- gan Kaufmann Publishers, 2000, pp. 223–230. [18] R.O. Duda, P.E. Hart, D.G. Stork, Pattern Classification, Wiley & Sons, Inc., New York, NY, USA, 2nd edition, 2001. [19] P.J. Green, Reversible jump Markov chain Monte Carlo and Bayesian model determination, Biometrika 82 (1995) 711–732. [20] J.D. Hadfield, MCMC methods for multi-response generalized linear mixed models: The MCMCglmm R package, Journal of Statistical Soft- ware 33 (2010) 1–22. [21] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, I.H. Wit- ten, The weka data mining software: an update, SIGKDD Explorations Newsletter 11 (2009) 10–18. [22] J. Hilden, J.D. Habbema, B. Bjerregaard, The measurement of perfor- mance in probabilistic diagnosis, Methods of Information in Medicine 17 (1978) 227–237. [23] A. Hunter, L. Kennedy, J. Henry, I. Ferguson, Application of neural net- works and sensitivity analysis to improved prediction of trauma survival, Computer Methods and Programs in Biomedicine 62 (2000) 11–19. [24] IBM Corporation, SPSS software, http://www- 01.ibm.com/software/analytics/spss/, 2013. [25] F. Jaimes, J. Farbiarz, D. Alvarez, C. Martnez, Comparison between logistic regression and neural networks to predict death in patients with suspected sepsis in the emergency room, Critical Care 9 (2005) 150–156. [26] L. Jakaite, V. Schetinin, Feature selection for Bayesian evaluation of trauma death risk, in: The 14th Nordic-Baltic Conference on Biomedical Engineering and Medical Physics, Springer, 2008, pp. 123–126. 20 [27] P. Kilgo, J. Meredith, T. Osler, Incorporating recent advances to make the TRISS approach universally available, Journal of Trauma 60 (2006) 1002–1009. [28] P. Kilgo, J. Meredith, T. Osler, Injury severity scoring and outcomes research, in: D.V. Feliciano, K.L. Mattox, E.E. Moore (Eds.), Trauma (6th ed), New York, McGraw-Hill, 2008, pp. 223–230. [29] T. Koshy, Catalan Numbers with Applications, Oxford University Press, 2008. [30] J.E. Kreke, A.J. Schaefer, M.S. Roberts, Simulation and critical care modeling, Current Opinion in Critical Care 10 (2004) 395–398. [31] W.J. Krzanowski, D.J. Hand, ROC Curves for Continuous Data, Chap- man & Hall/CRC, 1st edition, 2009. [32] L.I. Kuncheva, Combining Pattern Classifiers: Methods and Algorithms, John Wiley and Sons, Inc., 2004. [33] B. Li, H.F. Lingsma, E.W. Steyerberg, E. Lesaffre, Logistic random effects regression models: a comparison of statistical packages for binary and ordinal outcomes, Medical Research Methodology 23 (2011) 11–77. [34] D. Lunn, A. Thomas, N. Best, D. Spiegelhalter, WinBUGS – a Bayesian modelling framework: concepts, structure, and extensibility, Statistics and Computing 10 (2000) 325–337. [35] MathWorks, Inc, Matlab Statistics Toolbox, version 2012a, http://www.mathworks.co.uk/products/statistics/, 2012. [36] F. Millham, W. LaMorte, Factors associated with mortality in trauma: re-evaluation of the TRISS method using the National Trauma Data Bank, Journal of Trauma 56 (2004) 1090–1096. [37] J.S. Oakland, Statistical Process Control (5th edition), Butterworth- Heinemann, 2002. [38] T. Osler, R. F.B., G. Badger, M. Healey, D. Vane, S. Shackford, A sim- ple mathematical modification of TRISS markedly improves calibration, Journal of Trauma 53 (2002) 630–634. 21 [39] T. Osler, L. Glance, J. Buzas, D. Mukamel, J. Wagner, A. Dick, A trauma mortality prediction model based on the anatomic injury scale, Annals of Surgery 247 (2008) 1041–1048. [40] B.M. Patil, R.C. Joshi, D. Toshniwal, S. Biradar, A new approach: role of data mining in prediction of survival of burn patients, Journal of Medical Systems 35 (2011) 1531–1542. [41] J.R. Quinlan, C4.5: Programs for Machine Learning, Morgan Kaufmann Publishers, 1993. [42] J. Rasbash, C. Charlton, W.J. Browne, M. Healy, B. Cameron, ML- wiN. Version 2.02. Centre for multilevel modelling, University of Bristol, http://www.bristol.ac.uk/cmm/software/mlwin, 2005. [43] C. Robert, G. Casella, Monte Carlo Statistical Methods, Springer Texts in Statistics, Springer, 2004. [44] F. Rogers, T. Osler, M. Krasne, A. Rogers, E. Bradburn, J. Lee, D. Wu, N. McWilliams, M. Horst, Has TRISS become an anachronism? A com- parison of mortality between the National Trauma Data Bank and major trauma outcome study databases, Journal of Trauma and Acute Care Surgery 73 (2012) 326–331. [45] S. Rogers, M. Girolami, A First Course in Machine Learning, Chapman & Hall/CRC, 1st edition, 2011. [46] SAS/STAT, User guide 9.2. Introduction to Bayesian procedures. Centre for multilevel modelling, SAS Institute Inc., 2009. [47] V. Schetinin, J.E. Fieldsend, D. Partridge, T.J. Coats, W.J. Krzanowski, R.M. Everson, T.C. Bailey, A. Hernandez, Confident interpretation of Bayesian decision tree ensembles for clinical applications, IEEE Trans- actions on Information Technology in Biomedicine 11 (2007) 312–319. [48] V. Schetinin, J.E. Fieldsend, D. Partridge, W.J. Krzanowski, R.M. Ever- son, T.C. Bailey, The Bayesian decision tree technique with a sweeping strategy, in: The International Conference on Advances in Intelligent Systems, IEEE Computer Society, 2004. 22 [49] V. Schetinin, J.E. Fieldsend, D. Partridge, W.J. Krzanowski, R.M. Ev- erson, T.C. Bailey, A. Hernandez, Comparison of the Bayesian and ran- domized decision tree ensembles within an uncertainty envelope tech- nique, Journal of Mathematical Modelling and Algorithms 5 (2006) 397– 416. [50] V. Schetinin, L. Jakaite, Classification of newborn EEG maturity with Bayesian averaging over decision trees, Expert Systems with Applica- tions 39 (2012) 9340–9347. [51] V. Schetinin, L. Jakaite, W. Krzanowski, Bayesian cal- culator, Standalone application for Linux and Windows, http://www.traumacalc.org/bc, 2012. [52] A. Silva, P. Cortez, M.F. Santos, L. Gomes, J. Neves, Mortality assess- ment in intensive care units via adverse events using artificial neural networks, Artificial Intelligence in Medicine 36 (2006) 223–234. [53] E. Steyerberg, A. Vickers, N. Cook, T. Gerds, M. Gonen, N. Obu- chowski, M. Pencina, M. Kattan, Assessing the performance of predic- tion models: A framework for traditional and novel measures, Epidemi- ology 21 (2010) 128–138. [54] A. Stojadinovic, J. Eberhardt, T.S. Brown, J.S. Hawksworth, F. Gage, D.K. Tadaki, J.A. Forsberg, T.A. Davis, B.K. Potter, J.R. Dunne, E.A. Elster, Development of a Bayesian model to estimate health care out- comes in the severely wounded, Journal of Multidisciplinary Healthcare 16 (2010) 125–135. [55] K. Sujin, K. Woojae, W.P. Rae, A comparison of intensive care unit mortality prediction models through the use of data mining techniques, Healthcare Information Research 17 (2011) 232–243. [56] V. Vapnik, Statistical Learning Theory, Wiley, 1998. 23 Appendix. Bayesian Averaging over Decision Trees MCMC implementation of the Bayesian Method When averaging is made over DT models, the Bayesian formalism can be outlined as follows [14]. First we specify the parameters Θ of a DT model we aim to induce from the labelled data D represented by an m-dimensional input vector x. The outcome of a DT model is y = 1, . . . , C, where C ≥ 2 is the number of categories to one of which a DT model assigns a given input x. Given models M1, . . . , ML with parameters Θ1, . . . , ΘL, we can write the desired predictive distribution as an integral over the extended parameter vector Θ = (Θ1, . . . , ΘL): p(y|x, D) = ∫ Θ p(y|x, Θ)p(Θ|D)dΘ = L ∑ i=1 p(y|x, Θi)p(Θi|Mi, D)p(Mi), (2) were p(Mi) is the prior distribution of model Mi, p(Θi|Mi, D) is the posterior density of Θi given model Mi and data D, and p(y|x, Θi) is the posterior predictive density given the parameters Θi. The above integral is analytically tractable only in trivial cases when the distribution p(Θ|D) is known. In practice, we can estimate this distribution by drawing N random samples Θ(1), . . . , Θ(N) from the posterior distribution p(Θ|D), and then we can write: p(y|x, D) ≈ N ∑ i=1 p(y|x, Θ(i), D)p(Θ(i)|D) = 1 N N ∑ i=1 p(y|x, Θ(i), D). (3) The above approximation is achieved with the MCMC method of simula- tion or stochastic integration. The accurate approximation is achieved when a Markov chain becomes a random sequence with a stationary probability distribution. Then according to Eq. (3), we can draw the random samples and calculate the desired predictive density. Figure 5 shows an example of a DT model consisting of two splitting nodes, s1 and s2, and three terminal nodes t1, . . . , t3. The first node, s1, called the root, splits the entire data into two disjoint subsets so that data samples from one subset fall into node s2 via the left branch, and samples from the other subset fall into the terminal node t2 via the right branch. The node s2 further partitions the data samples which fall into the terminals t2 or 24 Figure 5: An example of DT model. The DT consists of two splitting nodes s1, s2 and three terminal nodes t1, . . . , t3. t3 via the left and right branches. Finally one of the terminal nodes assigns the given input to one of the given classes. In general, a binary DT with k terminals consists of (k − 1) splitting nodes, si, i = 1, . . . , (k − 1). The node si, has parameters including: the node position in the DT model, s p i , p = 1, . . . , (k − 1), an input variable svi , v = 1, . . . , m, and a threshold s q i , q ∈ (min(xv), max(xv)). The node si tests the vth variable against the threshold q and assigns the input x to the left branch if xv < q, or to the right one otherwise. A terminal node ti assigns the input x to class c with a probability P ci , i = 1, . . . , k. Consequently, a DT model is described by a vector of parameters, Θ, consisting of two parts. The first part includes the following parameters of nodes si: positions s p i , variables s v i , and thresholds s q i , i = 1, . . . , (k − 1). The second part includes the probabilities P ci , c = 1, . . . , C for each terminal node i, i = 1, . . . , k. DT models whose nodes split data into two disjoint subsets are called binary. The number of possible configurations of binary DTs with k terminal nodes, Sk, is determined by the Catalan number, see e.g. [29]. The number Sk grows exponentially with k and becomes very large for DTs with relatively small k. For example, for k = 25, Sk becomes a number 25 to the power of 12. In practice, to explain data we need to induce DT models of a reasonable size; the size of a DT model is defined by the number of its terminal nodes, k. Oversized DT models are difficult to interpret, and moreover they are prone to overfit data. The size of DT models is dependent on the number of data points, pmin, allowed to be in terminal nodes – setting a smaller pmin increases the size, while setting a greater pmin decreases the size. In most cases, prior informa- tion on the size of DT models is unavailable, and a suitable pmin has to be found empirically. In practice, the size of DT models is unknown or can be given within a range. In such cases, areas of interest (high posterior density of parameters Θ), which have to be explored in Eq. 3, are of variable size, and MCMC has to be extended to Reversible Jump (RJ) proposed in [19]. Prior information about input variables, such as importance of variables x1, . . . , xm, is also often unknown. In such cases, we can assign a variable v for the the node si to be drawn randomly from the uniform discrete distribution, v ∼ U(1, m). Similarly, a threshold q can be drawn from the uniform discrete distribution, q ∼ U(min(xv), max(xv)). It has been shown that the above priors are sufficient in order to build and explore DT models of different configurations within the RJ MCMC method [9, 14]. For binary DT models, the number of possible configurations, Sk, is defined by Eq. ??. From this equation, we see that the larger the k, the larger is the number Sk. So we expect that MCMC algorithm will explore possible DT configurations of size k with probabilities proportional to Sk. RJ MCMC for Averaging over DT Models The RJ MCMC method has been implemented for Bayesian averaging over DT models of variable size [9, 14, 50]. To explore DT models it has been proposed to use the birth, death, change-split, and change-rule moves made with Metropolis-Hastings (MH) sampler. The first two, birth and death, moves were proposed to reversibly change the number of nodes in a DT model (or the dimensionality of the model pa- rameter vector Θ). The third and fourth moves, change-split and change-rule, were aimed at changing the parameters Θ within a current dimensionality. The change-split move replaces a variable v in a chosen DT node si, while the change-rule move modifies a threshold q in node si. 26 The change-split moves are aimed at making large changes in the model parameters in order to potentially increase the chance of sampling from ar- eas of interest. Such moves are intended to disrupt a long sequence of the posterior samples drawn from a local area of interest. In contrast, the change-rule moves are aimed at making small changes in the parameters to let MCMC explore a surrounding area in detail. These moves are made more frequently than the others. The MH sampler starts with a DT consisting of one splitting node whose parameter Θ is assigned within the predefined priors. Making the above moves, the sampler attempts to grow the DT model to a reasonable size by fitting its parameters Θ to the data. The fitness or likelihood of DT models is gradually increased and then becomes oscillatory around some value. This phase, named the burn-in, has to be preset sufficiently long in order to achieve a stationary distribution of the Markov chain. When the Markov chain becomes stationary, the samples of the posterior distribution are collected to approximate the desired predictive distribution – this phase is called post burn-in. The above moves are made with the given proposal probabilities. Their values are dependent on the complexity of a classification problem – more complex problems require larger DT models. To grow such models, the proposal probabilities for the death and birth moves are set to larger values. In general, there is no guidance for setting proper parameters of the MH sampler, and their values have to be found empirically [9, 14, 50]. The proposed change is accepted according to the MH rule, see e.g. [14]: α = min ( 1, p(D|Θp)p(Θp) p(D|Θ)p(Θ) q(Θ|Θp) q(Θp|Θ) ) , (4) where p(D|Θp) and p(D|Θ) are the likelihoods of DT models with the pro- posed and current parameters Θp and Θ, respectively; p(Θp) and p(Θ) are the prior distributions of the parameters; q(Θp|Θ) is the conditional distri- butions of moving from the current parameter Θ to a proposed parameter Θp (so-called transition distribution); and q(Θ|Θp) is the density of the reverse transition. When the birth or death move changes a dimensionality of a DT model, the acceptance rule needs to count a proposal ratio, R. This ratio is de- pendent on the number of possible configurations of DT models, Sk, and so we need to count R to keep the Markov chain reversible during the MCMC 27 simulation. According to [14], the reversibility is kept when the following condition is met: p(Θp|D)q(Θ|Θp) = p(Θ|D)q(Θp|Θ). (5) The above density p(Θ|D) is p(Θ|D) = [ k−1 ∏ i 1 N(svi ) 1 m ] k Sk 1 K , (6) where N(svi ) is the total number of possible splitting rules for variable s v i ; K is the maximal number of terminals allowed for DT models induced from data. The transition distributions in Eq. 5 can be written as follows: q(Θp|Θ) = bk k 1 N(svi ) 1 m , (7) q(Θ|Θp) = dk+1 DQ , (8) where dk and bk are the proposal probabilities of death and birth moves, respectively, and DQ is the number of splitting nodes whose both branches are terminals. For the birth move, the Θp is a (k + 1)-dimensional vector, and therefore the reversibility is kept when Rb is Rb = p(Θp|D)q(Θ|Θp) p(Θ|D)q(Θp|Θ) , (9) The above definitions Eq. 7 and Eq. 8 allow us to rewrite the ratio Rb as follows: Rb = dk+1 bk k DQ+1 Sk Sk+1 . (10) Taking into account that DQ < k and Sk < Sk+1, we see that the ratio Rb ranges between 0 and 1 , Similarly, we can write the ratio Rd for the death move: Rd = bk dk−1 DQ k − 1 Sk Sk−1 . (11) 28 For the above DQ, the ratio Rb : Rb ≥ 1. We can see that the ratios Rb and Rd take different values when a model parameter space is of variable dimensionality. This allows the MH sampler to keep the desired reversibility and explore a parameter space proportionally to the numbers of configurations Sk. Problems of Sampling DT models DT models are multilevel hierarchical structures, as shown in Figure 5. Nodes located at a lower hierarchical level are strongly dependent on the pre- decessor nodes located at upper level. In such hierarchical structures, changes proposed by the MH sampler can significantly redistribute data points falling into DT terminal nodes. The change made in a node close to the DT root is most influential on the distribution. The changes in terminal nodes can be so significant that the likelihood of the DT model is decreased – the closer the node is to the root, the more significant is the change in distribution of data points. In most cases such proposals are rejected. In contrast, a change proposed in a node close to DT terminals is most likely to be accepted as such a change will insignificantly redistribute data samples in the DT terminals. As a result, the MH sampler will only explore limited configurations of DT models [9, 14]. Another problem occurs when the MH algorithm aims to sample large DT models. When a DT model is small and consists of a small number of terminal nodes, the number of data samples falling into the nodes is expected to be much larger than the given minimal number of points, pmin. However, when a DT has grown large, the number of data points is decreased so that further partitions become unavailable. This means that birth moves cannot be made until a death move merges two terminal nodes into one node. As a result the MH algorithm will sample a series of DT models with similar distributions of data samples over terminal nodes. Such series affect the diversity of samples from the posterior distribution and, therefore, the accuracy of approximation of the predictive distribution [14, 48]. Another negative effect is that unavailable moves degrade the given pro- posal probabilities of birth and change moves. When a move is unavailable, the MH algorithm will repeat the current sample, which reduces the diversity of model mixing [14]. In most cases, the number pmin is found from experiments – complex problems typically require a small pmin to allow growth of large DT mod- 29 els. However, an inappropriately small pmin leads to excessive growth of DT models. Growing a DT model, the MH algorithm makes birth moves and almost each birth move increases the likelihood of the model. The MH algorithm accepts these moves and the DT model grows rapidly. The growth of the model continues while the number of data samples in its terminal nodes exceeds pmin and the likelihood of a proposed model remains acceptable. During this period, the dimensionality of the DT model increases rapidly, and the sampler cannot explore the posterior within each dimensionality in detail. It is unlikely that samples will be drawn from areas of highest posterior density [9, 14]. The growth of DT models is typically monitored, and the modeller can reduce excessive growth by increasing pmin as well as by setting a smaller value of the proposal probability for the birth moves. To mitigate the negative effect of fast growing DT models, Chipman et al [9] have proposed a restarting strategy. This strategy allows a DT model to grow within a limited period in multiple runs. The average over all models grown in these runs produces a better approximation accuracy when the duration of the growth period and the number of the runs are properly set. A similar idea of restricting the growth of DT models has been proposed by Denison et al [14]. The growth is restricted within a given interval to allow the MH sampler to explore a model parameter space in detail. Both strategies require additional settings for the MH sampler, which have to be found experimentally. Sweeping Strategy As an alternative to the restricting strategies, the RJ MCMC method could be modified so as to reduce the number of replications of samples from the posterior density. In our previous work [48], we proposed a sweeping strategy aimed at reducing the number of unavailable moves. For making a change-split move, the sweeping strategy assigns a new variable xv, v ∼ U(1, m), and a threshold q: q ∼ U(a, b), (12) where U(a, b) is a uniform distribution on the interval between a = min(xv,j) and b = max(xv,j) defined by Np data points falling into the chosen node, where j = 1, . . . , Np. 30 For making change-rule moves, a new threshold q′ is drawn from a re- stricted Gaussian distribution: q′ ∼ N ′(q, σ2, a, b), (13) with mean q and given proposal variance σ2 on the interval (a, b). The proposed move can be made so that one or more terminal nodes in a DT model will contain fewer data points than pmin. If this happens in terminal nodes with a common parent node, these terminals are recombined into one terminal node, and the MH sampler counts such a move as a death move. If however there are two or more such terminals with different par- ents, the algorithm will assign the proposal unavailable in order to keep the reversibility of the Markov chain. Similarly to a change move, a birth move assigns a new splitting node with parameters drawn from the given prior. A new splitting variable xv is drawn from a uniform distribution, v ∼ U(1, m), and a new threshold q is assigned as described by Eq. 12. In our experiments, we observed that a MH sampler using the above prior on change moves proposes fewer unavailable moves and, therefore, the sampler accepts fewer replications of a current parameter vector Θ. Taking this into account, we hypothesise that a reduced number of the replications collected during the MCMC simulation will improve the diversity of model mixing. In support of this hypothesis, in our previous experiments [49] on the benchmark problems, we observed that the MH sampler using the above prior significantly reduced the dimensionality of parameter vector Θ as well as the uncertainty in estimates of predictive density. The above strategy, named sweeping in [48], is applied to the Markov chain in both burn-in and post burn-in phases. Let a MH sampler make the birth, death, and change moves as described in Section 9. Then we can describe the main steps of the sweeping strategy as follow. • Birth move: 1. Select a random terminal node i ∼ U(1, k) and count the number of data samples, p, in this node 2. If p > 2pmin then assign a variable, v ∼ U(1, m) and threshold given by Eq. 12 to a new splitting node 31 3. Count the numbers of data samples, p1 and p2, split by the new node 4. If (p1 ≥ pmin)&(p2 ≥ pmin) let the MH sampler check the accep- tance of the proposal • Change move (change-split or change-rule): 1. Select a random splitting node i ∼ U(1, k−1) and read its variable v and threshold q 2. For change-split assign a new variable, v′ ∼ U(1, m) 3. For change-rule assign a new threshold q′ defined by Eq. 13 4. Apply the proposed change to the DT 5. Count the number of terminal nodes, n0, with pi < pmin 6. If n0 == 1, then apply the death move 7. If n0 > 1, then assign the proposal unavailable and draw a new sample 8. Let MH sampler check acceptance of the proposal Here, n0 = ∑k i I(pi ≤ pmin), where I(·) = { 1 if pi ≤ pmin, 0 otherwise. Setting for the Metropolis-Hastings Sampler The settings include the following parameters of the sampler: 1. the proposal probabilities for birth, death, change-split, and change- rule moves, Pr 2. the proposal distribution, a Gaussian distribution with the zero mean and standard deviation s 3. the numbers of burn-in and post burn-in samples, nb and np, respec- tively 4. the sampling rate of the Markov chain, sr 5. the minimal number of data points allowed in terminal nodes, pmin The parameters Pr, s, and nb were the most significant factors that im- pact on the convergence of the Markov chain, and so we tested a number of variants of these parameters to achieve an acceptable convergence. The sampling rate sr was used to elevate the independence of samples drawn from the Markov chain during the post burn-in phase. At the second stage, the 32 settings s and pmin have been refined to let the MCMC algorithm efficiently sample the posterior distribution of Θ keeping the size of DT models rea- sonably small; the efficient sampling is achieved when the acceptance rate ranges between 0.25 and 0.5. The use of such a two-stage technique allowed us to reduce the number of possible combinations of the settings to a realistic number not exceeding 10. The best results were obtained with the following settings. The proba- bilities for the birth, death, change-split and change-rule moves were Pr = (0.2, 0.2, 0.1, 0.5), respectively. The proposal distribution was a Gaussian with s = 1.0. The numbers of samples were nb = 20, 000 and np = 5, 000, and the number of data points was pmin = 10. 33 
work_wg7nc7fzqfaxpc5udxvt3slr7y	----	SJNL1038-02-DO00020673.tex UvA-DARE is a service provided by the library of the University of Amsterdam (https://dare.uva.nl) UvA-DARE (Digital Academic Repository) A Grid-Based Hiv Expert System Sloot, P.M.A.; Boukhanovsky, A.V.; Keulen, W.; Tirado Ramos, A.; Boucher, C.A.B. DOI 10.1007/s10877-005-0673-2 Publication date 2005 Published in Journal of Clinical Monitoring and Computing Link to publication Citation for published version (APA): Sloot, P. M. A., Boukhanovsky, A. V., Keulen, W., Tirado Ramos, A., & Boucher, C. A. B. (2005). A Grid-Based Hiv Expert System. Journal of Clinical Monitoring and Computing, 19(4- 5), 263-278. https://doi.org/10.1007/s10877-005-0673-2 General rights It is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), other than for strictly personal, individual use, unless the work is under an open content license (like Creative Commons). Disclaimer/Complaints regulations If you believe that digital publication of certain material infringes any of your rights or (privacy) interests, please let the Library know, stating your reasons. In case of a legitimate complaint, the Library will make the material inaccessible and/or remove it from the website. Please Ask the Library: https://uba.uva.nl/en/contact, or a letter to: Library of the University of Amsterdam, Secretariat, Singel 425, 1012 WP Amsterdam, The Netherlands. You will be contacted as soon as possible. Download date:06 Apr 2021 https://doi.org/10.1007/s10877-005-0673-2 https://dare.uva.nl/personal/pure/en/publications/a-gridbased-hiv-expert-system(f4b256de-6fec-4b27-9806-31fe6822b5eb).html https://doi.org/10.1007/s10877-005-0673-2 UN CO RR EC TE D PR OO F TECHBOOKS Journal: JOCM MS Code: CH1 PIPS No: DO00020673 DISK 22-8-2005 18:20 Pages: 16 Journal of Clinical Monitoring and Computing (2005) xxx: 1–16 C© Springer 2005 A GRID-BASED HIV EXPERT SYSTEM1 Peter M.A. Sloot,1 Alexander V. Boukhanovsky,2 Wilco Keulen,3 Alfredo Tirado-Ramos,1 and 2 3 Charles A. Boucher44 From the 1Section Computational Science, University of Amsterdam, Kruislaan 403, 1098 SJ Amsterdam, The Netherlands, 2Institute for High Performance Computing and Information Sys- tems, Bering St, 38, St. Petersburg, Russia, 3Virology Educa- tion, 69042 Utrecht, The Netherlands, 4University Medical Center, University of Utrecht, 3508 GA Utrecht, The Netherlands. Received—, and in revised form—. Accepted for publication—. Based on “A Grid-based HIV Expert System”, by P.M.A. Sloot, A.V. Boukhanovsky, W. Keulen, and C.A. Boucher, which appeared in the IEEE/ACM International Symposium on Cluster Computing and the Grid, Cardiff, UK, May 9-12, 2005. c©2005 IEEE. Address correspondence to Peter M.A. Sloot, Section Computa- tional Science, University of Amsterdam, Kruislaan 403, 1098 SJ Amsterdam, The Netherlands E-mail: sloot@science.uva.nl Sloot P MA, Boukhanovsky AV, Keulen W, Tirado-Ramos A, Boucher CA. A grid-based HIV expert system. J Clin Monit 2005; xxx: 1–16 ABSTRACT. Objectives. This paper addresses Grid-based in- 5 tegration and access of distributed data from infectious dis- 6 ease patient databases, literature on in-vitro and in-vivo phar- 7 maceutical data, mutation databases, clinical trials, simulations 8 and medical expert knowledge. Methods. Multivariate analyses 9 combined with rule-based fuzzy logic are applied to the inte- 10 grated data to provide ranking of patient-specific drugs. In addi- 11 tion, cellular automata-based simulations are used to predict the 12 drug behaviour over time. Access to and integration of data is 13 done through existing Internet servers and emerging Grid-based 14 frameworks like Globus. Data presentation is done by standalone 15 PC based software, Web-access and PDA roaming WAP access. 16 The experiments were carried out on the DAS, a Dutch Grid 17 testbed. Results. The output of the problem-solving environ- 18 ment (PSE) consists of a prediction of the drug sensitivity of the 19 virus, generated by comparing the viral genotype to a relational 20 database which contains a large number of phenotype-genotype 21 pairs. Conclusions. Artificial Intelligence and Grid technology 22 is effectively used to abstract knowledge from the data and pro- 23 vide the physicians with adaptive interactive advice on treatment 24 applied to drug resistant HIV. An important aspect of our research 25 is to use a variety of statistical and numerical methods to iden- 26 tify relationships between HIV genetic sequences and antiviral 27 resistance to investigate consistency of results. 28 KEY WORDS. grid, HIV, PSE, expert system, artificial intelligence, 29 bio-statistics. 30 1. INTRODUCTION 31 1.1. Motivation 32 Forty two million people worldwide have been infected 33 with HIV and 12 million have died, over the last 20 years. 34 Figure 1 shows the pan-epidemic extent of HIV infections. 35 Effective antiretroviral therapy has lead to sustained HIV 36 viral suppression and immunological recovery in patients 37 who have been infected with the virus. The incidence of 38 AIDS has declined in the Western world with the intro- 39 duction of effective antiretroviral therapy, though questions 40 on “When to start treatment? What to start with? How to 41 monitor patients?” remain heavily debated. Adherence to 42 antiretroviral treatment remains the cornerstone of effec- 43 tive treatment, and failure to adhere is the strongest pre- 44 dictor of virological failure. Long-term therapy can lead to 45 metabolic complications. Other treatment options are now 46 available, with the recent introduction to clinical practice 47 of fusion inhibitors, second-generation non-nucleoside re- 48 verse transcriptase inhibitors, and nucleotide reverse tran- 49 scriptase inhibitors. The sheer complexity of the disease, 50 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F 2 Journal of Clinical Monitoring and Computing Vol xxx No xxx 2005 Fig. 1. Worldwide spread of HIV infections, history and near future per- spective. the distribution of the data, the required automatic updates51 to the knowledgebase and the efficient use and integration52 of advanced statistical and numerical techniques necessary53 to assist the physician motivated us to explore the novel54 possibilities supported by Grid technology.55 In this position paper we describe ongoing research in56 our 3 laboratories (Utrecht, St. Petersburg and Amsterdam)57 addressing the development of a Grid based medical deci-58 sion support system. The goal of the research is to investi-59 gate novel computational methods and techniques that sup-60 port the development of a user friendly integrated support61 system for physicians. We use emerging Grid-technology62 to combine data discovery, data mining, statistical analyses,63 numerical simulation and data presentation [1].64 The paper is organized as follows. Chapter 2 describes65 the background of HIV research and a prototypical rule-66 based approach to data analyses. In chapter 3 we give an67 overview of the two computational techniques we study68 to understand the temporal variability of HIV populations69 through stochastical modeling and the evolution of HIV70 infection and the onset of AIDS through Cellular Automata71 (CA) modeling. Chapter 4 describes a first approach to72 advanced data presentation through roaming devices such73 as Personal Digital Assistants (PDA’s).74 1.2. Background75 1.2.1. Clinical aspects of HIV76 The clinical management of patients infected with Human77 Immunodeficiency Virus (HIV) is based on studies on the78 pathogenesis of the disease and the results of trials evaluat-79 ing the effects of anti-HIVdrugs. Retrospective analysis of80 large cohorts has identified laboratory markers for disease81 progression, such as the amount of virus (HIV-RNA) and 82 the number of T helper cells (CD4 + cells) in blood. In ad- 83 dition the results of prospective drug trials have generated 84 data on effectiveness of individual drugs and drug combi- 85 nations and the effect of drug resistant viruses on therapy 86 outcome. Currently clinicians are limited in the practical 87 use of this information because in most cases they are only 88 provided with statistical relationships between individual 89 parameters and disease or therapy outcome. Large data sets 90 have not been analyzed and made available in such a way 91 that it allows a clinician to use the available data in more 92 clinical settings. The availability of large databases and the 93 development of innovative data mining approaches create 94 the opportunity to develop systems which allow the prac- 95 ticing clinician to determine the risk profile for disease 96 development, or the change or success for a given regimen 97 for his individual patients. Such a system will determine the 98 rate of success for different drug regimens by taking into 99 account the effect and interaction of all relevant laboratory 100 and clinical parameters and by comparing the results for 101 similar patients available in the database. 102 Currently there are fifteen drugs licensed for treatment of 103 individuals infected with HIV. These drugs belong to two 104 classes, one inhibiting the viral enzyme reverse transcrip- 105 tase and another inhibiting the viral protease. These drugs 106 are used in combination with therapy to maximally inhibit 107 viral replication and decrease HIV-RNA to below levels of 108 detection levels (currently defined as below 50 copies per 109 ml) in blood. Treatment with drug combinations is suc- 110 cessful in inhibiting viral replication to undetectable levels 111 in only 50% of the cases. In the remaining 50% of cases 112 viruses can be detected with a reduced sensitivity to one 113 or more drugs from the patients’ regimen. The molecular 114 base for resistance has been, and still is, focus of extensive 115 research. Over 80 amino acid positions in the viral enzyme 116 reverse transcriptase (RT) and 40 positions in the protease 117 enzyme can undergo changes when exposed to selective 118 drug pressure in vitro or in vivo. For some drugs, at cer- 119 tain positions, a change towards a specific new amino acid 120 is seen. At other positions several alternative amino acids 121 may appear and cause (variable) levels of resistance to one 122 or more drugs. In theory, therefore, an infinite number 123 of combinations of amino acid changes could appear and 124 cause resistance in vivo. Preliminary clinical observations 125 however show that specific amino acid changes at a limited 126 number of positions and a limited number of combina- 127 tions prevail. In addition to changing drug sensitivity some 128 amino acid changes may also influence the replication po- 129 tential of HIV. Amino acids selected initially during a failing 130 regimen cause resistance to the drugs the patient is taking, 131 but at the same time may decrease the capacity of the virus 132 to replicate. Changes appearing later do not function to 133 further increase resistance but merely function to restore 134 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F Sloot et al.: Grid-Based HIV Expert System 3 the capacity of the virus to replicate (“viral fitness”). Sev-135 eral clinical studies have been performed recently to evalu-136 ate the clinical benefit of resistance-guided therapy. These137 studies show that a better virological response is obtained138 in patients who are failing their therapy, when their new139 regimen is chosen on the basis of their resistant profile.140 In three out of the four studies from last year the results141 showed that if new regimens were selected on the basis of142 the mutations (viral resistance genotype) the results were143 better as compared to standard care approaches. Currently,144 the basis for clinical interpretation of the viral genotype is145 based on data sets relating mutations to changes in drug sen-146 sitivity, and/or data sets directly relating mutations present147 in the virus to clinical responses to specific regimens. Ini-148 tially, experts compared the observed mutations to lists of149 published sequences taken from the literature, and based150 on this comparison would select a regimen.151 1.2.2. Prototype support system152 Recently, first generation bioinformatics software pro-153 grams have been developed to support clinicians. Examples154 of such systems are the Virtual Phenotype developed by155 Virco NV, and a first generation decision support system156 (Retrogram TM) developed by Virology Networks BV in157 collaboration with parts of our research team. The out-158 put of these programs consists of a prediction of the drug159 sensitivity of the virus generated by comparing the viral160 genotype to a relational database containing a large num-161 ber of phenotype-genotype pairs. The Retrogram decision162 software interprets the genotype of a patient by using rules163 developed by experts on the basis of the literature, taking164 into account the relationship of the genotype and phe-165 notype. In addition, it is based on (limited) available data166 from clinical studies and on the relationship between the167 presence of genotype directly to clinical outcome. It is im-168 portant to realise however that these systems focus on bio-169 logical relationships and are limited to the role of resistance.170 The next step will be to use clinical databases and inves-171 tigate the relationship between the viral resistance profile172 (mutational profile and/or phenotypic data) and therapy173 outcome measures such as amount of virus (HIV-RNA)174 and CD4+ cells. A summary of the flow of data is shown175 in Figure 2.176 1.2.3. Data collection177 Large high quality clinical and patient databases are used178 to explore the relationships described above and to de-179 velop a first prototype matching system. The Athena co-180 hort is a large Dutch observational clinical cohort study181 Fig. 2. From molecule to man: Hierarchical data flow model for infectious diseases. aiming at the surveillance of antiretroviral treatment sup- 182 ported by the government. The cohort consists of 3000 183 patients from whom data are centrally collected through a 184 decentralized data entry system. Within the cohort 600 pa- 185 tients are studied intensively, whose phenotypic and geno- 186 typic data, drug levels and CD4+ and HIV-RNA patterns 187 are collected. Phenotype, genotype, viral fitness and drug 188 levels as CD4+ and HIV-RNA patterns will be collected 189 from two large international trials (sponsored by Roche 190 Pharmaceuticals), evaluating the effect of a new fusion in- 191 hibitor drug (T20), and representing 1000 patients. The 192 third database will be from the international multi-center 193 Great study, sponsored by Virology Networks BV. Within 194 this study the value of the Retrogram decision support 195 program is evaluated and similar parameters as described 196 above will be collected. Within this study 360 patients will 197 be enrolled. 198 The Viradapt study showed that the virological response 199 was better in the patient group in which genotype and rule- 200 based interpretation was used as compared to the standard 201 of care arm [2]. On the basis of these results, a more elabo- 202 rate decision support software system (Retrogram version 203 1.0) was built in collaboration with Virology Networks 204 B.V. This system ranks the efficacy of the antiretroviral 205 drugs within each class. The ranking is based on expert 206 interpretation of two types of data. The software system 207 estimates the drug sensitivity for the fifteen drugs by in- 208 terpreting the genotype of a patient by using mutational 209 algorithms. These mutational algorithms are developed by 210 a group of experts on the basis of the scientific literature, 211 taking into account the published data relating genotype to 212 phenotype. In addition, the ranking is based on data from 213 clinical studies on the relationship between the presence of 214 particular mutations and clinical or virological outcome. 215 The Athena cohort is a large Dutch observational clini- 216 cal cohort study aiming at the surveillance of antiretroviral 217 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F 4 Journal of Clinical Monitoring and Computing Vol xxx No xxx 2005 treatment supported by the Dutch government. The co-218 hort consists of 3000 patients from whom clinical, viro-219 logical, immunological and data on drug side effects are220 centrally collected through a decentralised data entry sys-221 tem. Within this cohort 600 patients are studied intensively,222 phenotypic and genotypic data, drug levels and CD4+ and223 HIV-RNA patterns are collected. From two large interna-224 tional trials (sponsored by Roche Pharmaceuticals) eval-225 uating the effect of a new fusion inhibitor drug (T20),226 representing 1000 patients from whom also phenotype,227 genotype, viral fitness, drug levels as CD4+ and HIV-RNA228 patterns will be collected. The third database will be from229 the international multi-center Great study sponsored by230 Virology Networks BV, within this study the value of the231 Retrogram decision support program is evaluated and sim-232 ilar parameters a described above will be collected, within233 this study 360 patients will be enrolled. Another dataset234 will come from the Italian Musa study, in this trial data will235 be collected from 450 patients followed over a year. Entry236 point to the trial is failing a fist or second regimen, subse-237 quently patients will be genotyped and a new regimen will238 be selected on the basis of Retrogram 1.4 or the Virtual239 Phenotype from Virco (Belgium).240 Throughout the duration of the project we will collect241 additional datasets. These datasets may serve to further re-242 fine our models and first version software and may also be243 use to perform validation studies.244 1.2.4. Data analysis245 The primary goal of the data analysis is to identify pat-246 terns of mutations (or naturally occurring polymorphisms)247 associated with resistance to antiviral drugs and to predict248 the degree of in-vitro or in-vivo sensitivity to available drugs249 from an HIV genetic sequence. The statistical challenges250 in doing such analyses arise from the high dimensional-251 ity of these data. A variety of approaches have been de-252 veloped to handle this type of data, including clustering,253 recursive partitioning, and neural informatics. Neural in-254 formatics is used for synthesis of heuristic models received255 by methods of knowledge engineering, and results of the256 formal multivariate statistical analysis in uniform systems.257 Clustering methods have been used to group sequences258 that are “near” each other according to some measure of259 genetic distance [3]. Once clusters have been identified,260 recursive partitioning can be used to determine the im-261 portant predictors of drug resistance, as measured by in-262 vitro assays or by patient response to antiviral drugs. Prin-263 ciple component analyses can help to identify what are the264 most important sources of variability in the HIV genome.265 An important aspect of our research is to use a variety of266 methods to identify relationships between HIV genetic se-267 quences and antiviral resistance to validate the consistency 268 of results. 269 The molecular sequences of the viral enzymes reverse 270 transcriptase and protease are the micro parameters in the 271 model. In theory an infinite number of combinations of 272 mutations could appear and cause (variable) changes in viral 273 drug sensitivity and viral replication capacity (See also Ta- 274 ble 1). Clinical datasets however show that specific amino 275 acid changes at a limited numbers of positions in a lim- 276 ited number of combinations prevail. HIV-RNA and CD4 277 are the primary parameters determining disease outcome. 278 HIV-RNA, the amount of HIV-RNA genomic copies per 279 ml plasma, has been validated as being highly predictive of 280 clinical outcome. HIV-RNA and CD4+ cell numbers are 281 now the standard endpoint in clinical trials for approval of 282 new antiretroviral drugs. A patient’s HIV-RNA may range 283 between a few hundred to millions of RNA copies per 284 ml plasma. The CD4+ cell numbers in peripheral blood 285 range typically between zero and thousand. Whereas the 286 predictive clinical value of both parameters has been deter- 287 mined initially in untreated individuals, they have also been 288 shown to be of predictive value also for patients under an- 289 tiretroviral therapy. Recently observations have been pub- 290 lished indicating that in some patients under highly active 291 antiretroviral therapy (HAART) a disconnect may occur 292 between the response in HIV-RNA and in CD4 counts. 293 Typically, in these patients a rise in HIV-RNA as conse- 294 quence of incomplete inhibition of viral replication under 295 therapy is not paralleled by a continuous decrease in CD4 296 counts. This disconnect has been explained by a decrease 297 Table 1. Parameters for the data analyses. Here the hierarchical ap- proach shown in Figure 2 is extended to detail the content of the parameters Micro Parameter Protease Mutations Reverse Transcriptas Mutations Primary Parameter HIV-RNA CD4 Drug Resistance Macro Parameter Meta Parameter: Virological Viral Fitness Meta Parameter: Clinical Weight Opportunistic Infections and Tumors Survival Intervention Parameter Drug Dosage Bio-availability of Drug/Drug Level AUTHOR'S PROOFS UN CO RR EC TE D PR OO F Sloot et al.: Grid-Based HIV Expert System 5 in the viral replicative capacity (‘viral fitness’) which leads298 to a decrease in capacity to lower CD4 counts.299 The patient’s weight and secondary opportunistic infec-300 tions and/or malignancies are parameters that determine301 disease outcome and survival time. Currently there are fif-302 teen drugs licensed for treatment of individuals infected303 with HIV: More than ten inhibitors have been developed304 which inhibit the reverse transcriptase process. These in-305 hibitors can be classified in two sub-categories that dif-306 fer in the way they inhibit the RT-enzyme, nucleoside307 (analogue) RT-inhibitors (NRTI) and the non-nucleoside308 RT-inhibitors (NNRTI). These compounds inhibit the309 protease enzyme, which acts much later on in the HIV310 replication cycle than reverse transcriptase.311 The protease is responsible for cleaving a long poly-312 protein into smaller functional proteins. The overall ex-313 posure to antiretroviral drugs has been shown to be an314 important factor for the degree of success for a given ther-315 apy. The overall exposure can be captured by parameters316 as dosage and bio-availability which will codetermine the317 drug level within an individual patient. Given the relation-318 ships between exposure and antiviral efficacy, variability in319 drug levels (which may be due to differences in patient320 adherence to their regimens) will contribute to virologi-321 cal and immunological outcome. Individuals with relatively322 low exposure are more likely to experience virological fail-323 ure than those with a high exposure.324 2. METHODS AND MATERIALS325 2.1. Modeling the dynamics and temporal variability326 of HIV-1 populations327 In addition to rule based and parameter based decision sup-328 port we developed statistical models and cellular automata329 based models to study the dynamics of the HIV popula-330 tions. These 2 numerical models run on Grid-resources.331 The output is integrated with the medical support system332 and accessible to the end-user. In this paragraph we briefly333 outline the two computational methods. Details are be-334 yond the scope of this paper; we refer to the references335 provided.336 2.1.1. A cellular automata model to study the evolution337 of HIV infection and the onset of AIDS338 A cellular automata model to study the evolution of HIV339 infection and the onset of AIDS is developed. The model340 takes into account the global features of the immune re-341 sponse to any pathogen, the fast mutation rate of the HIV,342 and a fair amount of spatial localization, which may occur 343 in the lymph nodes. The dynamics of the cellular automata 344 requires high throughput computing, which is provided by 345 the resource management of the Grid. In this section, we 346 employ non-uniform Cellular Automata (CA’s) to simulate 347 drug treatment of HIV infection, in which each compu- 348 tational domain may contain different CA rules, in con- 349 trast to normal uniform CA models. Ordinary (or par- 350 tial) differential equation models are insufficient to de- 351 scribe the two extreme time scales involved in HIV in- 352 fection (days and decades), as well as the implicit spatial 353 heterogeneity. Zorzenon dos Santos et al. [7] reported a 354 cellular automata approach to simulate three-phase pat- 355 terns of human immunodeficiency virus (HIV) infection 356 consisting of primary response, clinical latency and onset 357 of acquired immunodeficiency syndrome. We developed a 358 non-uniform CA model to study the dynamics of drug 359 therapy of HIV infection, which simulates four-phases 360 (acute, chronic, drug treatment responds and onset of 361 AIDS). Our results indicate that both simulations (with and 362 without treatments) evolve to the same steady state. Three 363 different drug therapies (mono-therapy, combined drug 364 therapy and HAART) can also be simulated in our model. 365 Our model for prediction of the temporal behaviour of the 366 immune system to drug therapy qualitatively corresponds 367 to clinical data. 368 Pseudo Code 1a: HI Model (Adapted from Zorzenon dos 369 Santos R. M., Phys. Rev. Let. 2001). H = healthy cell, 370 A1 and A2 are infected cells at different time steps. 371 Assume: {H, A1(t), A2(t+ τ ), D}; 1 time-step = 1 week; Simulation of lymph-node; Moore neighbourhood and square lattices used Rule 1: (a) If it has at least one infected-A1 neighbor, it becomes infected-A1 (b) If it has no infected-A1 neighbor but does have at least R (2 < R < 8) infected-A2 neighbors, it becomes infected-A1 (c) Otherwise it stays healthy Rule 2: An infected-A1 cell becomes infected-A2 after τ time steps Rule 3: Infected-A2 cells become dead cells Rule 4: (a) Dead cells can be replaced by healthy cells with probability prepl in the next step. (b) Each new healthy cell introduced may be replaced by an infected-A1 with probability p infec 372373 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F 6 Journal of Clinical Monitoring and Computing Vol xxx No xxx 2005 This CA (Pseudo-code 1a) mimics in a simple way the374 dynamical properties of a HIV infection; next we intro-375 duce drug therapy into the model by modelling a response376 function Presp and changing only rule 1.377 Pseudo Code 1b: Advanced HI Model, taking into378 account drug therapy effects.379 Rule 1: (a) If there is one A1 neighbor after the starting of drug therapy, N(0 ≤ N ≤ 7) neighbor healthy cells become infected-A1 in the next time steps with probability presp. Otherwise, all of eight neighbors become infected-A1. N represents effectiveness of drugs. N = 0: no replication; N = 7: less effective for the drug. Presp (t − ts ) represents certain response function of drug effects over the time steps (t). The ts is the starting of treatment. 380381 The main success of the presented CA model is the ad-382 equate modeling of the four-phases of HIV infection with383 different time scales into one model. Moreover, we could384 also integrate all of the three different therapy procedures.385 The simulations show a qualitative correspondence to clin-386 ical data. During the phase of drug therapy response, tem-387 poral fluctuations for N > 3 were observed, this is due to388 the relative simple form of the response distribution func-389 tion (Pdis)applied to the drug effectiveness parameter N390 at each time-step. The simulation results indicate that, in391 contrast to ODE/PDE, our model supports a more flexible392 approach to mimic different therapies through the use of393 mapping the parameter space of Pdis to clinical data. There-394 fore there is ample room to incorporate biologically more395 relevant response functions into the model. The data inte-396 gration required for the CA, the parametric computation397 and the data presentation are supported by the Grid.398 2.1.2. Multivariate stochastic modeling399 The modeling of Human Immunodeficiency Virus400 (HIV-1) genotype datasets has a goal to identify patterns401 of mutations (or naturally occurring polymorphisms) as-402 sociated with resistance to antiviral drugs and to predict403 the degree of in-vitro or in-vivo sensitivity to available drugs404 from an HIV-1 genetic sequence. The statistical challenges405 in doing such analyses arise from the high dimensionality406 of these data. Direct application of the well-known genetic407 approaches [5] to analysis of HIV-1 genotype results in a lot408 of problems. Principal difference is in the fact that, in HIV409 DNA analysis, the main scope of interests is the so-called 410 relevant mutations – a set of mutations, associated with the 411 drug resistance. These mutations might exist in different 412 positions over the amino-acid chains. Moreover, the sheer 413 complexity of the disease and data require the development 414 of the reliable statistical technique for its analysis and mod- 415 eling. A multivariate stochastic model for describing the 416 dynamics of complex non-numerical ensembles, such as 417 observed in the (HIV) genome, has been developed in [6]. 418 This model was based on principle component analyses for 419 numerated variables. Generally speaking, the interpretation 420 of numerated variables in terms of relevant mutations is not 421 clear. Below we develop this model directly for the ensem- 422 ble of relevant mutations in the RT and protease parts of 423 the HIV-1 genome. Each element of the ensemble is pre- 424 sented as the cortege �k = {ξ j }n kj =1, k = 1, M with the 425 variable dimension n k -the total number of the mutations 426 in the gene. Each value ξk is a literal index and corresponds 427 the position and new value of the amino acid (e.g., 184 V, 428 77I, etc.). It allows to associate each mutation with the cat- 429 egorical random variable i ∈ 1 . . . K , where K is the total 430 number of possible mutations. Each sub sample of genomes 431 with a fixed number of mutations n = const may be con- 432 sidered as the realizations of a categorical random vector. 433 The representation above is based on the proximity to the 434 “wild-type” virus and takes into account only the relevant 435 mutations in a genome. It allows for significant compression 436 of the DNA representation and simplifies the interpretation 437 of the results. 438 Principle of the modeling approach. The joint variability of dif- 439 ferent mutations in the HIV-1 genomes is a complicated 440 phenomenon. The dimension of the probabilistic charac- 441 teristics is high, and its analytical investigations and inter- 442 pretation are hard. Hence, for the studying of HIV-1 pop- 443 ulations we use a computational statistical approach that 444 allows to numerically generate an ensemble with the same 445 probabilistic properties by means of a Monte-Carlo pro- 446 cedure. This is a well-known powerful method to study 447 complex system variability. 448 The idea of the stochastic modeling is shown in the 449 Figure 5. It is based on the evolutionary hypothesis, consid- 450 ering the group with n + 1 mutations as subgroup of group 451 with n mutations in a previous step. For each gene the tran- 452 sit from n to n + 1 mutation groups is driven by a stochastic 453 operator D(n+1), which defines the mutations on the n + 1 454 step, when the mutations on the previous n steps are known. 455 The initial step of the stochastic procedure begins from the 456 whole ensemble of wild-type viruses. The number of the 457 genomes that has been mutated at each step of the stochas- 458 tic procedure is in accordance with Mn = ρn M, where ρn 459 are the probabilities of the occurrence of genotypes with n 460 mutations in a total population of M genes. 461 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F Sloot et al.: Grid-Based HIV Expert System 7 Fig. 3. Temporal behaviour of the CD4 count, with modeled Brownian movement for lymphocytes [8]. Fig. 4. As in Figure 3, with additionally modeled mono therapy in week 300 [8]. Fig. 5. Principle of the modeling. The stochastic operator D may be considered as a “black462 box”. It is formalized in terms of the conditional probabil-463 ities of the occurrence of mutation ξi , if the mutation ξ j464 arise in the previous step of the generation. For genotypes465 with 2 mutations only the values Di j are the conditional466 probabilities of the pairs. In this case the matrix {Di j } is467 the transition Markov probability matrix, containing the 468 conditional probabilities for simple Markov chains with 469 the number of these states corresponding to quantity of 470 the relevant mutations. In more complicate cases, where 471 n > 2, the probability matrix {Di j } consists of the con- 472 ditional probabilities to meet mutation ξ j in certain gene, 473 when the mutation ξi is present. 474 This approach allows us to reduce the complicated sta- 475 tistical description of the dataset to a rather simple model, 476 using only three probabilistic distributions as the initial pa- 477 rameters of the model: distribution of number n of the 478 mutations ρn ; 479 • distribution P (1)ξ for the relevant mutations in the group 480 n = 1; 481 • transient probability matrix D. 482 All these parameters might be identified on the sample 483 datasets of the HIV-1 population. 484 Identification of the model. For the identification of parameters 485 of the model, a large database of HIV-infected patients, col- 486 lected over several years in USA, is used [4]. These databases 487 contain genotypes of 43620 patients examined from Au- 488 gust 9, 1998 to May 5, 2001. We observed 59 different 489 mutations in the RT genome, including 17 mixed muta- 490 tions, and 77 different mutations in the protease genome, 491 including 34 mixed mutations. 492 Distribution ρn of number of mutations. The practice of HIV 493 treatment however, has shown that the variability of the 494 number of mutations n is high, due to the complexity of 495 the drug combinations that has been applied. The sample 496 estimate of distribution ρn of the number of mutations in 497 protease is shown in the Figure 6. It is seen, that the distri- 498 butions have a clear first peak (n = 1), and a shelf (or second 499 peak), corresponding to n = 3 ÷ 5. Therefore we expect 500 that there are two groups of genomes in the database, cor- 501 responding to the low and high number of mutations. The 502 possible interpretation of the discovered bi-modal distri- 503 bution is that we have two groups of patients. One group 504 is the “new” patients who had one or two treatments, thus 505 their genotype contains relative small numbers of muta- 506 tions. The second group is the “old” patients, which have 507 a long treatment history, or new patients, infected through 508 treated HIV-1 patients [15]. 509 Distributions of the relevant mutations Pξ . Distribution ρn al- 510 lows describe the variability of the groups of the “new” 511 and “old” patients, only. For a more detailed study of the 512 virus mutations driving by the certain drugs combinations, 513 the probabilities of occurrence of the relevant mutations 514 ξ should be considered. They are estimated by the sample 515 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F 8 Journal of Clinical Monitoring and Computing Vol xxx No xxx 2005 Fig. 6. Statistical description for distribution of mutations in Protease. frequencies:516 Pξ = {Number of genes with mutation ξ } M . (1) Here M is the total number of genomes in the dataset.517 Equation (1) describes the marginal impact of each muta-518 tion in the total population, without any information about519 number and occurrences of other mutations. The prob-520 abilities of the most significant relevant mutations ξk (in521 decreasing order of its probability) are shown in Figure 6.522 The marginal estimates of Pξ over the total dataset show523 only general impacts of the mutations. For a detailed524 analysis of its behavior we also consider the occurrences525 P (n) ξ of mutations in the groups of genotypes with exactly526 n mutations. These values were computed also by means527 of Equation (1), where M def= Mn = ρn M – the number of528 genes with n mutations in a database. The sample estimates529 of these occurrences are also shown in the Figure 1. It is530 clearly seen that the inputs of some mutations are rather dif-531 ferent for different n, both for the protease and RT parts of532 the genome. E.g., for RT, for n = 1, the mutations 184 V533 and 103 N have the main input. The distribution P (1) ξ is the534 limit distribution from the procedure shown in Figure 5.535 From Figure 1 we also observe that the total sum536 ∑ k Pξk > 100%, excluding case n = 1. This demonstrates537 that the analysis of the marginal mutations is not enough538 for general statistical description of all DNA ensemble vari-539 ability, because some positions of DNA may be statistically540 dependent [15], especially in relation to viral fitness. Hence,541 the joint characteristics of its variability must be taking into 542 account. 543 Transient probability matrix D.The conditional probability of 544 the occurrence of mutation ξi , if the mutation ξ j arises 545 from the previous steps of the generation, is estimated by: 546 Di j = {Number of genotypes with mutations ξi and ξ j simult&aneously} {Number of genotypes with mutation ξi } . (2) 547 The dimensionality of the related matrix, obtained from 548 Equation (2), may be rather high. In order to decrease 549 the dimensionality we consider the algebraic technique of 550 orthogonal expansion, applied to transient probability ma- 551 trices [16]. 552 D = ��1/2�. (3) where � are the eigenvectors of matrix DDT , and �-of 553 matrix DT D. It allows considering the coefficients a k = 554√ λk as the principal components (PC) [13], and represents 555 the probability (2) as a series: 556 Di j = ∑ k √ λk φi k ψ j k . (4) The values λk shows the part of the probability, explained 557 by k-th PC. The sum of the first k-th coefficients λk may 558 be interpreted as a measure of convergence of the series 559 (4). In Table 2 the values of the first 7 λk for the RT and 560 protease parts of the HIV-1 genome are shown. These data 561 were obtained for the total database. It can be seen that the 562 series (4) converges rather fast in both cases: e.g. for the RT 563 part only the first term of the series explain more 60% of 564 conditional probability (the first five terms explain 80%). 565 Let us consider the normalized bases φ̃i k = λ0.25k 566 φi k , ψ̃ j k = λ0.25k ψ j k . It allows to present the terms in Equa- 567 tion (4) as the p i jk = φ̃i k ψ̃ j k and interpreted these values 568 as the independent factor loadings, driving the changes of 569 the conditional probability Di j over all the mutations ξi , ξ j 570 in the database. For example, in the Figure 7 the estimates 571 Table 2. Normalized (%) values of the expansion coefficients λk in Equation (4) # of PC Part of the genome 1 2 3 4 5 6 7 RT 61.3 8.2 5.4 2.8 2.1 1.7 1.6 Protease 55.0 6.3 4.5 4.2 3.4 2.7 2.4 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F Sloot et al.: Grid-Based HIV Expert System 9 Fig. 7. Orthogonal basic functions of expansion (4) for transient probability matrix. of the first basic functions are shown for RT and protease572 parts of the genotype (the input of multiplication of func-573 tions are in the Table 2). It is clearly seen, that the first574 term p i j1 = φ̃i 1ψ̃ j 1 reflects the total occurrence of the mu-575 tations in a genotype (see Figure 6): for the mutations with576 the maximal occurrences the input to conditional proba-577 bilities of its pairs is also high.578 Model validation. The simulation model is based on the579 ρn , P (1) ξ , D distributions of the mutations only. No infor-580 mation of more complicate mechanisms (distributions of581 pairs, triples, etc.) has been used for this identification.582 The main goal of the verification is the possibility to583 reproduce these features of the ensemble through the de-584 pendencies formalizing the matrix D. We compared the585 total occurrences of all mutations in genotypes, estimated586 on the initial and simulated samples, see also Figure 6 (solid587 line). It is seen, that the results of the simulation and sample588 are rather close.589 The error of the simulation increases proportionally to590 absolute value of the occurrences. Nevertheless, for some591 cases the error of the simulation is larger then the boundary592 of the confidence interval. This systematic error may be 593 explained by possible variations in matrix D for groups of 594 the “old” and “new” patients. 595 Application to forecast of HIV-1 evolution in time. The evolu- 596 tion of total world populations of HIV-1 and the associ- 597 ated changing of the related drug resistance levels should 598 be taken into account. The stochastic models, used to de- 599 scribe the HIV-1 genotype ensemble in terms of parame- 600 ters and shown in the Figure 5, can be used for the analysis 601 of its temporal variability during the observation period 602 (VIII.1998–V.2001). The temporal variability of the data 603 may be considered in terms of the samples of the seasons 604 (3-months periods). The volumes of seasonal samples are 605 from 1500 till 4500 genotypes; that is enough for obtain- 606 ing the stable estimations. Only the hypothesis of linear 607 trends is considered: ξ (t ) = a t + b + δ(t ), where a is the 608 most interesting parameter—value of the trend, b is the 609 shift parameter, and δ is the white noise. In the Table 3 the 610 integral parameters of trends of the various parameters of 611 the HIV-1 population (mean value of the parameter, value 612 of the trend, determination coefficient R2 and the sample 613 value of F-criterion) are shown. 614 Trends of single mutations occurrence Pξ . The database allowed 615 us to investigate trends in codon frequency in the period 616 of 1998 till 2001. Results for Protease and RT are shown 617 in Table 3. The majority of the mutations in the genotype 618 have a negative trend, only 77I in Protease has significant 619 positive trend. 620 Trends of bi-modal distribution for number of mutations in geno- 621 types ρn . For the decreasing of the data dimensionality and 622 the statistical discrimination of two groups in the dataset 623 we consider the model of the mixture of two Bernoulli 624 distributions: 625 ρn = p g Ckm 1 q k1 (1 − q 1)m 1−k + (1 − p g )Ckm 2 q k2 (1 − q 2)m 2−k (5) where p g is an input of the first group of mutations (and 626 p g is an input of the second group, m 1, m 2-are maximal 627 numbers of mutations in groups and q 1, q 2-are probabil- 628 ities to find each one (arbitrary) mutation in the groups. 629 The use of Bernoulli distribution logic (based on the rep- 630 etition of the independent events) is more close to the 631 description of the mutation process, then the Poisson dis- 632 tribution, generally applying to description of rare events. 633 Temporal variability of the parameters ( p , q 1, q 2, m 1, m 2)t 634 of the ρn approximation by Equation (5) are shown in 635 Table 3. In both cases only the parameter p g (weight 636 of the left part for group of m1 mutations) has a clear 637 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F 10 Journal of Clinical Monitoring and Computing Vol xxx No xxx 2005 Table 3. Trend analysis of the parameters of the HIV-1 genotype population (F is compared with Fisher’s test F(1,31,95%) = 4.14) Occurrence of mutations, % pg , %, Coefficients √ λk , Equation (4) Parameter 77I 90M 10I 71V Equation (5) k = 1 k = 2 k = 3 Protease part Mean 37.78 32.69 27.97 23.64 48 5.78 1.67 0.83 a (1/month) 0.20 −0.43 −0.72 0.32 0.74 0.13 0.06 0.06 R2 0.68 0.91 0.61 0.82 0.67 0.80 0.73 0.54 F 16.7 77.6 9.6 47.1 64.0 23.6 26.8 11.8 RT part 41L 215Y 103N 67N k = 1 k = 2 k = 3 Mean 32.86 31.37 30.66 27.21 47 6.65 2.20 2.08 a (1/month) −0.51 −0.50 −0.32 −0.39 0.49 0.11 0.17 0.07 R2 0.88 0.93 0.88 0.84 0.75 0.68 0.78 0.71 F 57.4 98.7 59.8 41.8 94.3 21.4 36.1 25.3 significant positive trend. For protease value p g increased638 from 39% in Summer, 1998 to 62% in Summer 2001639 (with average increment a = 0.74% per month). Taking640 into account trends for separate mutations we observed a641 “degradation” of genotypes: the number of patients with642 simple genotypes (small number of mutations) is growing643 but a number of patients with big count of mutations is644 decreased.645 Trends of transient probabilities D. The analysis of the trends of646 parameters for distribution (1) shows that the input of the647 first group of mutations with low number n is increased.648 Hence, it may be a consequence of the temporal variations649 of the interdependencies between different mutations, gov-650 erned by the developing of the drug therapy. For the anal-651 ysis of these hypothesis, let us consider the trends for the652 matrix D, Equation (2). Taking into account the expan-653 sions (3, 4), we may reduce the complicate problem for654 joint trend analysis for components Di j to the procedure655 of trend analysis for independent time series – components656 of expansions (4). From the Table 3 it can be seen, that all657 the components have a clear positive trends. Taking into658 account the shape of first bases functions, see Figure 7, it is659 clear, that generally the joint probabilities Di j of the mu-660 tations is increased also; moreover, the power of increasing661 corresponds to the total occurrences of the mutation in the662 ensemble.663 The discrimination of the groups of “old” and “new”664 patients in terms of bi-modal distribution (5) allow to fore-665 cast the growth of the total number of HIV-infected people666 in time:667 N(t ) = Nnew patients (t ) + Nold patients (εt ), ε � 1. (6) Here ε – is the slow time parameter, which shows the rapid 668 increasing of the new patients group in comparison with 669 the old patients. The part of “new” patients of the sample 670 is p g (old patients−(1 − p g )) from (5). Hence, the growth 671 curve is: 672 N(t ) = Nold patients (0) [ 1 + p g (t ) 1 − p g (t ) ] , (7) where p g (t ) = p0 + a g t -is the linear trend with the pa- 673 rameters from Table 3, and N old patients (0) is the initial value of 674 “old” (treated) patients on the beginning of the forecast. 675 In Figure 8 the “crucial” forecast of the HIV-1 popula- 676 tion growth are shown. It is based on the fact that altogether 677 Fig. 8. Qualitative forecast of HIV-1 population grows. 1 – mean value (7), 2 – 90% confidence interval. AUTHOR'S PROOFS UN CO RR EC TE D PR OO F Sloot et al.: Grid-Based HIV Expert System 11 42 million people worldwide have been infected with HIV678 at the beginning of XXI century, and 12 million have died679 over the last 20 years. Moreover, not taken into account680 is the arising of new drugs and different prophylactic and681 social preventive activities for restriction of HIV-1 infec-682 tion. Really, this result is qualitative only; for quantita-683 tive conclusions the more sophisticated research should be684 done.685 3. RESULTS686 3.1. Data presentation: Roaming PDA access687 3.1.1. User Scenario688 RetroGramTM (www.retrogram.com) is a unique HIV-689 genotype expert based interpretation software program,690 which weighs the effect of specified genotype changes on691 clinical drug activity. It accepts a list of substitutions to the692 protease and reverse transcriptase genes with respect to the693 NL4-3 reference strain. This is accomplished by running694 a “simulation”, which applies some hundred rules relat-695 ing substitutions on the HIV genome to knowledge of696 effects on drug response. The latter comes from over hun-697 dreds of references from the clinical literature. The rules are698 checked against the reported substitutions, and each drug is699 evaluated for its suitability. In a later stage we added Web-700 access where a Web interface is used to submit the input701 and take out the output. We want to make the simulations702 wireless-accessible. Developing a wireless Internet version703 from scratch will not be cost-efficient and causes maintain-704 ability problems. For example, the rules mentioned above705 are often changed and these changes have to be reflected in706 both versions. Furthermore, for privacy and security rea-707 sons the developer is not granted access to the source code708 of the “simulation”. Thus, it is much more convenient to709 have wireless access to the Web-based interface. In this case710 the “simulation” take places in a unique server and privacy711 and security are guaranteed. A typical user scenario is de-712 scribed below and the associated graphical representation713 of the Retrogram Web access is given in Figure 9.714 After the user has successfully logged in, the Patient Detail715 page is displayed (Figure 10). The form, taking place in716 this page is used to enter the personal data of the patient.717 Two fields are required in the form, Patient ID and Data of718 Sample.719 According to the information taken from the laboratory720 the user enters the laboratory test results (i.e. Protease or721 RT substitutions) for the patient in the Laboratory Informa-722 tion page. Next a script invoked on the server does the723 following:724 Fig. 9. Web-based Retrogram use case sequence. Script 1: Server validation script 725 Validate inputs: Validate Protease or RT substitutions if they conform to certain rules. A single substitution should be represented by an integer (for position in the gene) and a letter (for the amino acid). The position in the gene is in the rage from 1 to 99 for Protease position and from 1 to 599 for RT position. The amino acid code is one of the following codes: A C D E F G H I K L M N P Q R S T U V W Y. Submit the inputs to the “simulation” program and take back the drugs ranking result. Show the Drugs ranking result in the ‘HIV Therapy decision support’ screen: After applying certain rules on the laboratory test result return to the final drugs ranking or drug’s level of suitability indication as follows: A (green): This drug can be used B (yellow): Consider use if no class A drug available C (amber): Consider use if no class A or B drug available D (red): Consider use if no class A, B or C drug available U (grey): Unranked, insufficient data available 726 In the ‘HIV Therapy decision support’ screen, clicking on 727 any drug name in the ranking lists will display a list of avail- 728 able references from the scientific literature supporting the 729 particular ranking for that drug. In the ‘HIV Therapy deci- 730 sion support’ screen, clicking on the ‘Interpret substitution’ 731 button will show classification of the patient’s substitutions 732 into relevant, natural or additional. 733 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F 12 Journal of Clinical Monitoring and Computing Vol xxx No xxx 2005 Fig. 10. Web Retrogram: user enters patient substitutions (left), drug ranking results (right). 3.1.2. Roaming, wireless access734 In the designing phase of wireless versions of the application735 the constraints of the mobile devices should be considered.736 At the same time we have tried to maintain the same level737 of usability and readability as in the original Web version.738 This is accomplished by maintaining the same structure as739 that in the Web but with some modifications. For example,740 the Patient detail form has many fields and putting them741 in one screen would cause problems in the usability of742 the program (it’s supposed that the mobile device has a743 resolution comparable to a normal PDA, i.e., something744 around 160 × 160 pixels). Thus we use three screens for745 Patient Detail data. The Patient Detail Web page has 2746 required fields. We put them in the first screen after the747 ‘login’ screen. In this way, if the user is not interested in748 entering optional data, she can directly go to the Laboratory749 Information.750 Proxy method Implementation. A Proxy method is imple-751 mented for accessing the web-based software from mobile752 devices. The Proxy server takes places between the remote753 server (the Retrogram server) and the mobile device. A754 mininavigator script developed in the Proxy is responsible755 for the following:756 • Take the patient data from the mobile user (i.e. patient757 detail, laboratory information)758 • Create an HTTP communication with the remote759 server,760 • Submit data to the remote server. These data are basically761 the input for the Retrogram ‘simulation’.762 • Take the result from the remote server (HTML code763 generated from retrogram.asp script),764 • Parse HTML code and retrieve only relevant informa-765 tion (i.e. drug ranking, error messages, drug references766 etc.). It uses this relevant information to build wireless 767 pages (i.e. WML page in case of WAP or Web-clipping 768 page). 769 • Send the wireless pages to the mobile device. 770 The Proxy is implemented using PHP: Hypertext Pre- 771 processor as a server-site scripting language [9–11] running 772 on the Apache Web server [12]. 773 Two versions are developed using the Proxy method: 774 WAP version and web clipping. If a user wants to enter the 775 ‘patient details’ fields, he has to move from one screen to 776 the other and come back again. The fields already filled in 777 the previous screens should not be lost. Thus maintaining 778 the client’s state is necessary. In the WAP case we simply 779 use cookies but in web clipping cookies are supported only 780 in PALM OS 4.0 version or higher. For this reason the 781 “hidden field” method is used this is another method used 782 for maintaining state in the Internet. The following figures 783 are the user interfaces that have been captured. They track 784 the user’s path through the running of the application, as 785 shown in Figures 11(a) and 11(b), where the user enters 786 the patient’s details and accesses ranking results. 787 J2ME Implementation. The same user interface is applied in 788 the J2ME implementation. There are two main differences 789 between the J2ME implementation and the Proxy one: 790 1. J2ME enables the device to communicate directly to 791 the Retrogram server without an intermediate Proxy 792 2. In J2ME the client’s interface is contained within the 793 device. In the Proxy method, every time the interface 794 should be changed, the Proxy is responsible for gener- 795 ating a new page. 796 The following illustrates the necessary steps one should 797 take in order to fetch an HTML page generated from a 798 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F Sloot et al.: Grid-Based HIV Expert System 13 Fig. 11. (a) User corrects the input and submit again (left), drug ranking re- sults (right). (b) Users clicks to the drug ‘indinavir’ (left), references supporting this ranking (right). script in the remote host. Specifically this is an example799 illustrating how the user can login to a script in the Ret-800 rogram server and extract the cookie from the header re-801 sponse:802 1. Open an HTTP connection803 2. Open an input stream804 3. Make an HTTP POST request805 4. Extract the cookie from the header response806 5. Close the connection807 In the J2ME implementation of Retrogram the entire808 client’s interface takes places in the device. The connec-809 tion to the server is established in the following cases: user810 login, with connection with the server is necessary in order811 to validate the user and/or password. The user submits the812 Fig. 12. J2ME method; user enters patient’s substitutions (left), drug ranking results (right). username and password, and the application judges them 813 for their correctness by scanning the HTML response from 814 the Retrogram server. The user submits the patient’s lab- 815 oratory information data. The application should connect 816 to the server in order to submit the data, take the result 817 (HTML format) and extract the drugs ranking. Next the 818 user looks for the references that suggest a certain drug 819 ranking. The database with all the references exists in the 820 Retrogram server, therefore the connection is necessary. 821 The application submits to a Retrogram script the cookie 822 and the name of the drug. The drug references are given 823 back from the server in HTML format. The application 824 should clean up the HTML tags and show the references 825 as plain text. Finally the user looks for classification of the 826 patient’s substitutions. This classification is part of the Ret- 827 rogram ‘simulation’ and thus the connection to the server 828 is still necessary. In Figure 12 we illustrate the process of 829 taking the drugs ranking using the J2ME method. 830 Currently we have the J2ME version in use for different 831 users to study the usability and extendibility. More details 832 on the implementation can be found in reference [13]. 833 3.2. Virtual laboratory infrastructure 834 3.2.1. A virtual organization for retrogram-centered workflow 835 Grid technology is a major cornerstone of today’s com- 836 putational science and engineering, with its basic unit of 837 Grid organization called the Virtual Organization (VO). 838 A VO is a set of Grid entities, such as individuals, appli- 839 cations, services or resources, which are related to each 840 other by some level of trust. In the most basic example, 841 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F 14 Journal of Clinical Monitoring and Computing Vol xxx No xxx 2005 Fig. 13. A Retrogram-centered workflow. service providers would only allow access to the mem-842 bers of the same VO. We are currently building a dis-843 tributed Grid-based overall decision support infrastructure844 to support the Retrogram-centered workflow shown in845 Figure 13.846 This VO will offer a Grid virtual laboratory that will847 assist users in the interpretation the genotype of a patient848 by using rules developed by experts on the basis of the lit-849 erature, taking into account the relationship between the850 genotype and phenotype. The workflow is based on highly851 distributed available data from clinical studies and on the852 relationship between the presence of genotype and the clin-853 ical outcome. In order to cover the fast temporal and spatial854 scales required to infer information from a molecular (ge-855 nomic) level up to patient medical data multi-scale methods856 are applied, where simulation, statistical analysis and data857 mining are combined and used to enhance the rule-based858 decision. In this scenario, information sources are widely859 distributed, and the data processing requirements are highly860 variable, both in the type of resources required and the pro-861 cessing demands. Experiment design, integration of infor-862 mation from various sources, as well as transparent schedul-863 ing and execution of experiments will be supported by this864 support system based on distributed Grid middleware. The 865 DAS2 testbed (Netherlands) will initially provide the addi- 866 tional computational power for our compute intensive jobs. 867 We will reuse Grid middleware from successful European 868 projects such as CrossGrid (www.crossGrid.org) and VL-e 869 (www.vl-e.nl) to provide basic Grid services of data man- 870 agement, resource management, and information services 871 on top of Globus. For transparent use of this infrastructure 872 we will build a presentation layer that will provide a user- 873 friendly interface to both medical doctors and scientists. 874 4. DISCUSSION 875 4.1. Conclusions and future work 876 In this paper we discussed an integrative approach to bio- 877 medicine at large and to infectious diseases in particular. 878 We showed how in the understanding of processes ‘from 879 molecule to man’ Grid technology can play a crucial role. 880 In order to cover the fast time and spatial scales required to 881 infer information from a molecular (genomic) level up to 882 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F Sloot et al.: Grid-Based HIV Expert System 15 patient medical data, we need to apply multi-scale meth-883 ods where simulation, statistical analysis, data-mining is884 combined in an efficient way. Moreover the required in-885 tegrative approach asks for distributed data collection (e.g.886 HIV mutation databases, patient data, literature reports etc.)887 and a virtual organization (physicians, hospital administra-888 tion, computational resources etc.). Also the access to and889 use of large-scale computation (both high performance as890 well as distributed) is essential since many of the compu-891 tations involved require near real-time response and are892 to complex to run on a personal computer or PDA. Fi-893 nally data presentation is crucial in order to lower the894 barrier of actual usage by the physicians, here the Grid895 technology (server-client approach) can play an important896 role.897 Although many of the aspects discussed in this paper898 have proven to work in concept, the complete integration899 of the systems and the evaluation of day-to-day use is900 still under development [17]. In addition each of the901 underlying methods (Rule-based, statistical and CA based902 models) remain topics of further studies. We will set up a903 use-base with the system described running under various904 European Grid testbeds. The first testbed we will use is905 the so-called DAS2, and eventually the CrossGrid testbed,906 which supports specific features for interactive computa-907 tion, an essential ingredient for a medical decision support908 system.909 The authors gratefully acknowledge Fan Chen and Ferdinand910 Alimadhi for assistance in implementing the CA models and911 the roaming PDA access. The Dutch Virtual Laboratory on e-912 science project supported parts of the research presented here:913 http://www.VL-e.nl.914 GLOSSARY915 Grid: Distributed architecture for solving computational916 problems by making use of the resources from the mem-917 bers of a virtual organization, treating them as a virtual918 cluster.919 CA: Cellular Automata, a discrete model studied in com-920 putational theory and mathematics, which consists of921 regular grid of cells, each in one of a finite number of922 states.923 Decision Support System: Computer-based system that924 helps in the process of decision-making.925 Web Interface: User interfaces for information available via926 the web.927 Proxy: Computer service which allows clients to make in-928 direct network connections to other services.929 HTTP: Hyper Text Transfer Protocol, a request/response 930 protocol for transferring information on the Web. 931 HTML: Hyper Text Markup Language, a markup language 932 designed for the creation of web pages. 933 WML: Wireless Markup Language, a markup language 934 used in mobile phones. 935 J2ME: Java 2 Platform Micro Edition, a collection of Java 936 interfaces for embedded consumer appliances such as 937 cellular phones. 938 DAS2: Distributed ASCI Super Computer 2, a wide-area 939 distributed computer connecting 5 Dutch Universities. 940 REFERENCES 941 1. Zhao Z, Belleman RG, van Albada GD, Sloot PMA. AG-IVE: 942 An Agent-Based Solution to Constructing Interactive Simula- 943 tion Systems, in Series Lecture Notes in Computer Science, 944 April 2002; 2329: 693–703. 945 2. Durant J, Clevenbergh P, Halfon P, Delguidice P, Porsin S, 946 Simonet P, Montagne N, Dohin E, Schapiro JM, Boucher 947 C, Dellamonica P. Improving HIV therapy with drug resis- 948 tance genotyping: The Viradapt Study. Lancet 1999; 353: 2195– 949 2199. 950 3. Sevin AD, DeGruttola, Nijhuis M, Schapiro JM, Foulkes AS, 951 Para MF, Boucher CAB. Methods for Investigation of the Re- 952 lationship between Drug-Susceptibility Phenotype and Human 953 Immunodeficiency Virus Type 1 Genotype with Applications 954 to AIDS Clinical Trials Groupw 333. The Journal of Infectious 955 Diseases 2000; 182: 59–67. 956 4. The Genotype database is obtained from a large service testing 957 laboratory from the US. It contains the resistance profiles of the 958 Protease and Reverse Transcriptase genes of the HIV-1 virus 959 obtained from plasma samples of HIV-1 infected patients. No 960 clinical background information on medication or drug history 961 is available. 962 5. Mathematical Methods for DNA Sequences. In. Waterman MS, 963 eds. CRC Press Inc., Boca Raton, Florida, 1999. 964 6. Kiryukhin I, Saskov K, Boukhanovsky AV, Keulen W, Boucher, 965 CA, Sloot PMA. Stochastic modeling of temporal variability of 966 HIV-1 population. In: Sloot PMA, Abrahamson D, Bogdanov 967 AV, Dongarra JJ, Zomaya AY, Gorbachev YE, eds. Compu- 968 tational Science – ICCS 2003, Melbourne, Australia and St. 969 Petersburg, Russia, Proceedings Part I, in series Lecture Notes 970 in Computer Science, vol. 2657, pp. 125–135. Springer Verlag, 971 June 2003. ISBN 3-540-40194-6. 972 7. Zorzenon dos Santos RM, Coutinho S. Dynamics of HIV infec- 973 tion: A cellular automata approach. Phys Rev Lett 2001; 87(16): 974 168102–1–4. 975 8. Sloot PMA, Chen F, Boucher CA. Cellular automata model 976 of drug therapy for HIV infection. In: Bandini S, Chopard 977 B, Tomassini M, eds. 5th International Conference on Cellu- 978 lar Automata for Research and Industry, ACRI 2002, Geneva, 979 Switzerland, October 9–11, 2002. Proceedings, in series Lecture 980 Notes in Computer Science, vol. 2493, pp. 282–293. October 981 2002. 982 9. PHP: Hypertext Preprocessor: http://www.php.net. 983 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F 16 Journal of Clinical Monitoring and Computing Vol xxx No xxx 2005 10. The resource for PHP developers: http://www.phpbuilder.com984 11. Zend Technologies – PHP tools for the development, pro-985 tection and scalability of PHP applications – PHP for Linux,986 Unix and Apache, Encoder, Accelerator Studio, Debugger:987 http://www.zend.com.988 12. The Apache Software Foundation: http://www.apache.org.989 13. Alimadhi F. Mobile Internet: Wireless access to Web-990 based interfaces of legacy simulations, MSc thesis, Uni-991 versity of Amsterdam, The Netherlands, September 2002:992 http://www.science.uva.nl/research/pscs/papers/master.html.993 14. Cross-Grid: Grid technology of Interactive Distributed Com-994 putation: http://www.eu-crossGrid.org/.995 15. Little SJ, Holte S, Routy JP, Daar ES, Markowitz M, Collier AC, Koup RA, Mellors JW, Connick E, Conway B, Kilby M, Wang 996 L, Whitcomb JM, Hellmann NS, Richman DD. Antiretroviral- 997 drug resistance among patients recently infected with HIV. N 998 Engl J Med 2002; 8;347(6): 385–394. 999 16. Karlin S. A First Course in Stochastic Processes. Academic Press. 1000 NY-London, 1968. 1001 17. Sloot PMA, Boucher CA, Kiryukhin I, Saskov K, 1002 Boukhanovsky AV. A grid-based problem-solving envi- 1003 ronment for biomedicine. In: Nørager S, ed. Proceedings of 1004 the First European HealthGrid Conference, January, 16th-17th, 1005 2003, pp. 300–323. Commission of the European Commu- 1006 nities, Information Society Directorate-General, Brussels, 1007 Belgium, 2003. 1008 AUTHOR'S PROOFS UN CO RR EC TE D PR OO F Query1009 Q1. Au: Pls. provide dates.1010 AUTHOR'S PROOFS 
work_wgwslmzl2nhltp2ks7ehcdlxum	----	An expert system to manage dispute resolutions in construction projects in Egypt Ain Shams Engineering Journal (2016) 7, 57–71 Ain Shams University Ain Shams Engineering Journal www.elsevier.com/locate/asej www.sciencedirect.com CIVIL ENGINEERING An expert system to manage dispute resolutions in construction projects in Egypt Abbreviations: ADR, alternative dispute resolution; UK, United Kingdom; EFCC, Egyptian Federation for Construction and Companies; IP.I, importance index; DRExM, disputes resolution expert manager; FIDIC, Federation International Des Ingenieurs Conseils; DRB, dispute review board; DAB, disputes adjudication board; AI, artificial intelligence * Corresponding author. E-mail addresses: elbadry11@yahoo.com (A.A. Elziny), president_ office@suez.edu.eg (M.A. Mohamadien), hi_hgh@yahoo.com (H.M. Ibrahim), motazkhalil@gmail.com (M.K. Abdel Fattah). Peer review under responsibility of Ain Shams University. Production and hosting by Elsevier http://dx.doi.org/10.1016/j.asej.2015.05.002 2090-4479 � 2015 Faculty of Engineering, Ain Shams University. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). A.A. Elziny a,*, M.A. Mohamadien b , H.M. Ibrahim c , M.K. Abdel Fattah d a Petrojet Company, Port Said, Egypt b Faculty of Engineering, Suez Canal University, Ismailia, Egypt c Faculty of Engineering, Port Said University, Port Said, Egypt d Petrojet Company, Cairo, Egypt Received 24 August 2014; revised 22 February 2015; accepted 4 May 2015 Available online 13 June 2015 KEYWORDS Disputes; Alternative dispute resolutions; Expert system; Artificial intelligence; Model validation Abstract This study attempts to shed a great deal of light on the problem of construction disputes in the Egyptian projects. This paper presents a comprehensive review of the available literature on analysis of disputes. The objective of this paper was to provide an expert system can evaluate the overall dispute settlement procedures at company’s projects. A questionnaire has been used to study dispute sources and resolution methods. Four case study applications have been provided to check the validity of the proposed system. Results confirmed that the most important source of disputes was contract management 74.04%, the second was contract documents 71.49%, the third was finan- cial issues 67.80%, the fourth was project related issues 63.92%, and the lowest one was other sources (such as force majeure) 61.58%. Finally, the expert program facilitates dispute resolution by using alternative dispute resolution methods instead of going direct to arbitration or litigation. � 2015 Faculty of Engineering, Ain Shams University. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1. Introduction Construction industry in Egypt suffers from the misunder- standing of dispute resolution management; many factors affect the development of dispute resolution. Over the last years, there has been a breakdown in the relations between parties involved in the construction processes. Several studies have been reviewed which present the disputes’ definitions, nature, parties, classification, causes and resolution in con- struction projects. Richard [1], defined a dispute as ‘‘a specific disagreement concerning a matter of fact, law or policy in which a claim or assertion of one party is met with refusal, counter – claim or denial by another’’, Diekmann and Girard [2], described dispute as ‘‘any contract question or con- troversy that must be settled beyond the jobsite management http://crossmark.crossref.org/dialog/?doi=10.1016/j.asej.2015.05.002&domain=pdf http://creativecommons.org/licenses/by-nc-nd/4.0/ mailto:elbadry11@yahoo.com mailto:president_office@suez.edu.eg mailto:president_office@suez.edu.eg mailto:hi_hgh@yahoo.com mailto:motazkhalil@gmail.com http://dx.doi.org/10.1016/j.asej.2015.05.002 http://www.sciencedirect.com/science/journal/20904479 http://dx.doi.org/10.1016/j.asej.2015.05.002 http://creativecommons.org/licenses/by-nc-nd/4.0/ From five to ten years %8 From ten to fi�een years 26% More than fi�een years 66% Figure 1 Ranges of respondent companies’ experiences. 58 A.A. Elziny et al. staff’’, Corby [3], defined dispute as ‘‘a difference between the parties after the internal procedure has been exhausted’’, and Cheung and Yiu [4], stated that dispute is ‘‘a regular feature in construction and consumes resources that would otherwise be used in a more productive manner’’. It can be said that a dispute only appears after a claim has been made and been rejected , Ndekugri and Russell [5]. Bunni [6], specified that one of the main reasons that can affect the completion of projects is disputes. It is normal to have disputes in construction projects related to contract nat- ure. Thomas [7], stated that the nature of disputes arising from engineering contracts may range from trivial disputes to dis- putes that threaten the viability of the underlying transaction. Steen [8], stated that the construction industry has become known as one of the most adversarial and problem-prone, with claims and disputes on construction projects frequently the rule rather than the exception. The large risk that can be resulted from disputes existing, requires fair resolution methods. Roxene [9], stated that in a typical construction project, the owners, donor agencies, project managers, field engineers, gen- eral contractors, subcontractors, and suppliers are the primary stakeholders. So when disputes arise in a construction project, some or all of the stakeholders are the dispute parties. Without exception, disputes involve, misunderstandings, conflicting solutions on the issues, and communication dynamics between the parties. Bunni [6], said that disputes are a reality in any construction project, as the construction contracts always have many parties. Construction contracts are different from other contracts in many points such as; the large number of contract parties, numerous tasks to be implemented, and the large per- iod of execution. UNITAR [10], suggested that in case of dis- similarity the parties have a choice to select the laws and jurisdiction of courts of the country to which either of the par- ties belong to or of a neutral or third country. Such a choice is made at the time of entering into the contract. Shin and Molenaar [11], made classification for disputes based on definition of disputes and analyses of disputes for his related research, and the proposed three major types of dis- putes are contractual, organizational and technical disputes. Contractual disputes include definitions, interpretation, and clarification of the contract. Contractual issues cause a significant portion of disputes in many projects. Organizational disputes are related to human behavior in project operations and include human interactions, personality, cultures, and professional background among project stakeholders. Technical disputes are considered as the most common issues in project operation and include engineering clarification, which is a part of engineering decision making processes. Fenn, Hall, and Carmicheal [12–14], have identified the causes of construction disputes that are caused by client, designer and contractor. Fenn and Speek [12], identified the following factors for the client as failure to respond in timely manner, poor communication among members of the team, inadequate tracing mechanism for request of information, defi- cient management, supervision and coordination efforts on the part of the project, lowest price mentality in engagement of contractors and designers, the absence of team spirit among the participants, reluctance to check for constructability, clar- ity and completeness, failure to appoint a project manager and ambiguities in contract documents. Hall [13], identified causes of construction disputes caused by the consultant such as failure to understand the responsibilities under the design team contract, over design and underestimating the costs involved, late information delivery and cumbersome approach to request for information, design and specification oversights and errors or omissions resulting from uncoordinated civil, structural, architectural, mechanical and electrical designs and incom- pleteness of drawings and specifications. Carmicheal [14], iden- tified causes of construction disputes caused by the contractor as follows: inadequate contractors’ management, supervision and coordination, delay/suspension of works, failure to under- stand and correctly bid or price the works, Inadequate CPM scheduling and update requirements. Zakzok [15], mentioned peaceful resolution of the dispute which passes before reaching the judiciary or international arbitration that those responsible for the project have to be aware of the causes of conflict and work to avoid them in the beginning of the project and the speed of handling and decision-making with their claims. Nosair [16], provided a sound solution to the construction disputes problem. This will be through the development of an expert system that can mate- rially help to reduce the likelihood of construction disputes. The output of the proposed system is a reliable prediction for the expected causes of disputes for any future project. Nicholas Gould [17], searched for how disputes arise and then taking proactive steps to avoid them communicating well and looking for objective solutions and avoiding conflict can also help once the project is under way. A commercially based set- tlement, either in negotiation or by mediation, is now fre- quently used in the construction industry. Use of a mediator or some other ADR process can resolve disputes more quickly, saving time and money. If all of this fails, there are of course the procedures of arbitration and litigation. While they are applicable occasionally, they are best avoided if possible. Howard Klein [18], looked at all procedures currently being used by UK employers, professional advisors, project man- agers and contractors to resolve their disputes. All of them normally at lower cost and reduced time than would be incurred in Arbitration or Litigation, to ensure that contrac- tors benefit from prompt ensuing cash flow necessary for their survival and well being, and thereby the UK’s construction. Verster [19], showed the professional how by communicating effectively and continuously, disputes can be minimized. It also proposes some procedures to enable all functionaries and par- ties to the contract to focus on achieving the project objectives. It is advised that the building blocks for resolving differences should be communication, conciliation, adjudication and mediation, with arbitration or litigation as a last resort. El-Adaway and Ezeldin [20], investigated how arbitration is used as a dispute resolution mechanism in Egyptian large scale The first Class 52% The second Class 20% The third Class others 22% Figure 2 Different companies’ classes as per (EFCC). Public project 12% Private project 16% Petroleum project 35% Interna�onal project 37% Figure 3 Ranges of respondents’ project nature. An expert system to manage dispute resolutions in construction projects 59 construction projects; a research project was conducted to study the arbitration process for a dispute that was in excess of $31 million, which arose out as a result of the proceeds of a large-scale project with an original contract price of $85 mil- lion that was constructed in Cairo, Egypt. Their research pro- ject analytically investigated the background of the conflicts, the arbitral proceedings, and the award issued by the arbitral tribunal. Based on such thorough study, it was concluded that arbitration did not provide a timely and cost-effective resolu- tion for the said dispute. It is perceived that his paper would trigger professionals to think of other suitable dispute resolu- tion mechanisms, such as dispute review boards, for settlement of disputes arising from Egyptian large scale construction pro- jects. Furthermore, this study would be valued for contractors and owners who intend to work in the Egyptian construction market. Zakzok [15], developed a framework to assist contrac- tors in calculating lower limits of dispute value. The proposed framework, named Minimum Acceptable Negotiation Amount, is a decision support system which consists of three Figure 4 Ranges of respondents’ contract types. modules: duration, certainty and intention. These modules capture the main characteristics of negotiation process includ- ing expected dispute duration in case of litigation, certainty of litigation, and contactor’s intention to make the litigation. He also described the characteristics of these three modules and their associated factors which have been determined based on interviews with experts and questionnaire surveys. El-Adaway et al. [21], searched for the most suitable dispute-resolution mechanism for large-scale construction projects in Egypt. This dispute-resolution mechanism was attained through a multi step methodology that started with the study of the Arbitration process in relation to an Egyptian construction project with an initial contract price of 85 million; continued with interviews of five senior experts in the field of construction disputes in Egypt about their views pertaining to the most efficient dispute-resolution methodology for Egyptian megaprojects; developed a tailored questionnaire to assess the perceptions of 35 professionals toward the issue of construction disputes and dispute resolution mechanisms, including DRB; finally concluded by carrying out a what-if scenario for the arbitration case of the large-scale construction project using DRB instead of arbitration. On basis of the anal- ysis of the methodology, the authors concluded that despite the wide range of current dispute resolution methodologies, the employment of DRB should mitigate the negative effects of disputes in Egyptian large-scale construction projects. Okharedia [22], showed how South Africa has successfully used ADR (refers to a set of practices and techniques aimed at permitting the resolution of legal disputes outside the courts. The practice and technique of ADR comprises negotiation, mediation, conciliation, arbitration, and a variety of ‘‘hybrid’’ processes by which a neutral person facilitates the resolution of legal disputes without formal adjudication.) to settle dispute. How this method used by other countries such as Ghana, Nigeria, Ethiopia, Malawi and other African countries. The main objective is to keep disputes out of the normal court sys- tem in an effort to cut down the cost of resolving the dispute among the parties. The present study is thus an attempt to extend the previous studies by investigating the disputes’ definitions, nature, par- ties, classification and resolutions in construction projects. From the authors’ point of view, the studies which were con- ducted on disputes settlement procedures are too few but rather rare especially on causes of disputes. So this research may be a good starting point for the study by applying a deci- sion support system that helps the contract parties to estimate their performance in dispute management, settle disputes through proposed ADR methodology and provide recommen- dations in these issues. solved 34% pending 66% Figure 5 Status of disputes faced by respondents’ companies. Figure 6 Dispute notification methods considered by respon- dents’ companies. 60 A.A. Elziny et al. 2. Questionnaire design 2.1. Domain of experts Interviews and e-mails have been successfully implemented with 120 experts with different scope of experiences in the Egyptian construction industry and different years of field experience for each construction category. They also were selected with suitable period of experience so that their answers can represent valuable information. The studied population was the companies working in the fields of: concrete structures, buildings, steel, roads, water and sewage plants, and electro- mechanics. The analysis of data shows some interesting find- ings regarding the selected experts as companies’ experience, classes as per (EFCC), project nature, contract type, status of disputes, dispute notification methods and disputes rates during the last five years considered by respondents’ compa- nies as shown in Figs. 1–7. 2.2. Method of data analysis As provided by Odeh and Battaineh [23], the frequency of each type of claims is determined by giving weight for each class of frequency as chosen by the respondent, then calculates its weighted average and importance index is detailed in the following. A weight in a scale from 1 to 4 was given for each of the four frequencies with a weight of 1 for ‘‘Seldom’’, 2 for ‘‘Sometimes’’, 3 for ‘‘Often’’, and 4 for ‘‘Always’’. No weight was given when no response was provided. Weighted Average ¼ X Wi �Xi=N ð1Þ where Wi is the weight assigned to the ith option. Xi is the number of respondents who selected the ith option; and N is the total number of respondents (140 questionnaires Figure 7 Disputes rates during the last five years considered by respondents’ companies. have been distributed, 120 out of 140 filled and completed by the informants in public and private projects in Egypt, and then analyzed). To better express the importance of the questionnaire responses, an importance index percentage was then calculated: Importance Index % ¼ Weighted Average �100=h ð2Þ where h is the highest scale = 4. 2.3. Results and discussions The results of analyzing main sources of disputes are dis- tributed into five groups as shown in Fig. 8. Fig. 8 shows a summary of IP.I% of sources of disputes. The first most important source of disputes among the five groups, is the ‘‘contract management’’, then the ‘‘contract doc- uments’’ and the ‘‘financial issues’’ are the third, the ‘‘project related issues’’ is the forth, and finally ‘‘other sources’’ is the fifth, which respectively had Ip.I percentages as 74.04%, 71.49%, 67.8%, 63.92%, and 61.58%. The results of analyzing modern methodologies to settle disputes in construction projects are categorized into eight methods (Negotiation, Mediation, Conciliation, Fact-Finding, Dispute Adjudication Board (DAB), Mini-trial, Arbitration and Litigation). Five factors have been identified through the literature review and consulting experts who affect choices of each method and these factors as shown in Figs. 9–13. 3. System design and equations In this section, phases of the development of the knowledge based system for representing alternative resolution technique for construction disputes will be illustrated (the current version of the system will be referred as DRExM). The overall archi- tecture of DRExM will be presented. The detailed structure of each source of disputes phase will be briefed including contract management, contract documents, financial issues, project related issues and others. On the other hand, the procedures of evaluation and the results of the validation process of the proposed DRExM system will be discussed. DRExM is developed for all types of construction projects in Egypt. The system can predict the expected alternative resolution technique of construction disputes for the users depending on user’s project information, dispute nature and relation with other project parties. The expert system depends on the Rule IF Condition and then Action according to the current situation and dispute 50 55 60 65 70 75 80 Financial issues Contract Management Contract documents Project related issues Other Reasons 67.8 % 71.49 % 74.04 % 63.92 % 61.58 % Figure 8 Sources of disputes importance index percentage. Nego�a�on 27% Media�on 22% Concilia�on 13% Fact-finding 9% DAB 12% Mini-trial 7% Arbitra�on 6% li�ga�on 4% Figure 9 Most useful disputes’ resolution methods from Questionnaire Results. Figure 11 Effect of person in charge in disputes’ resolution methods from Questionnaire Results. Figure 12 Less cost disputes’ resolution methods from Questionnaire Results. An expert system to manage dispute resolutions in construction projects 61 status to choose the suitable procedure for settlement as per the following: � The first choice is negotiation. � The second choice is between two procedures (mediation via conciliation). � The third choice is between four procedures (DAB, fact- finding, mini-trials, and expert system). � And the fourth and final choice is arbitration. It can describe every item for every question that is asked to the user. The user of the system has the ability to show all items that affect every construction dispute. Also, the system can save user’s project information, user inputs (user’s answers), and systems output ADR for disputes. Then, the user can retrieve the saved data and he can update the data and rerun the system to show the new construction disputes. 3.1. Architecture of the expert system and verification As the problem of defining the construction disputes is well defined and the knowledge is available in the form of recom- mendations, the rule based knowledge representation tech- nique (Visual Rule Studio’s production Rule System) was selected to implement the expert system. DRExM uses an integration of computer software such as Visual Basic, Microsoft Access, and Visual Rule Studio. The windows environment involves Visual Basic Environment, Microsoft Access Database, and Visual Rule Studio. The expert system is designed by using artificial intelligent techniques, and AI is ‘‘the branch of computer science con- cerned with making computers behave like humans’’. One cat- egory of AI is the ‘‘Expert System’’. An expert system is a Figure 10 Fastest disputes’ resolution methods from Questionnaire Results. ‘‘computer application that performs a task that would other- wise be performed by a human expert’’. Legal expert systems therefore use artificial intelligence techniques to help comput- ers apply the law to any given set of facts. According to Nilsson [24], AI is concerned with intelligent behavior in artifacts, which involves perception, reasoning, learning, communicating and acting in complex environments. Ultimate goal of AI is generally perceived as the development of machines that can do what humans can, or possibly even better. Another goal of AI can be defined as understanding this kind of behavior whether it occurs in machines or in humans. Thus, AI has both scientific and engineering goals. Coppin Figure 13 Most acceptable disputes’ resolution methods from Questionnaire Results. 62 A.A. Elziny et al. [25], defined AI as the study of systems that act in a way to any observer would appear to be intelligent. AI involves using tools based on the intelligent behavior of humans and other animals to solve complex problems. The wide range of applications required further categorizations of AI. The problems of AI have been divided into sub-groups such as deduction, reason- ing, problem-solving, knowledge representation, planning, learning, natural language processing, motion and manipula- tion, perception, social intelligence, creativity and general intelligence. On the other hand, Approaches to AI have been grouped as cybernetics and brain simulation, cognitive simulation, logical, symbolic, knowledge based AI, sub-symbolic and statistical. However, because of the diversified applications of AI, these sub-groups are still too general. There are diversified tools used in AI research as well. The most frequently used tools in AI are search and optimization, propositional logic, first-order logic, fuzzy logic, default logics, case-based reasoning, probabilistic methods for uncertain rea- soning, classifiers and learning methods, neural networks and genetic algorithms. These tools at the same time constitute Figure 14 Program lan Figure 15 Progr the methods used in the applications and determine the approach to the problem at hand. Cheung et al. [26], stated that the use of AI in construction dispute resolution has not attracted too great attention despite the fact that dispute resolution is an important component of project management. Chau [27], also found that AI techniques are not common and are rarely applied in legal field. AI research has become highly specialized and today, applications of AI can be seen in construction dispute resolution as well as many other areas. Although these applications are quite new and regarded as rare by many researchers, AI has already con- tributed to the field as more efficient use of ADR methods, more systematic approaches to dispute resolution method selection and more analytic appraisal of disputes. Content validity test was conducted by consulting a group of fifteen experts (ten of them were specialist in construction fields and the other five were academic professors in Faculty of Engineering). That was requested to evaluate and identify whether the program design agreed with the scope of the items and the extent to which these items reflect the concept of the research problem. guage option banner. am interface. Figure 16 Add questions screen. Figure 17 Project profile screen. Figure 18 Part I – sources of disputes window. An expert system to manage dispute resolutions in construction projects 63 The group of experts agreed that the program design was valid and suitable enough to measure the concept of interest with some amendments which were then taken into consideration. 3.2. System operation The proposed system was carefully designed to be easily oper- ated. Such operating environment includes a number of menu screens that work easily in a serial order. To get the proposed system, DRExM, the program screens will appear as follows. The program language option banner will be displayed as shown in Fig. 14. It contains two language options English and Arabic, in order to start ‘‘DRExM’’ one of both options should be chosen then click go. Then, the program interface window will be opened as shown in Fig. 15; the program toolbar at the top of interface window contains the following lists for admin interface and user interface: � Admin interface (add questions – backup and restore). � User interface (start new project – edit project). � Reports (settlement procedures – sources of disputes – project reports – project procedures – charts). � Help (about the program – manual). Figure 19 Sources of disputes and its main causes report. Figure 20 Causes of disputes screen. Figure 21 Part II – settlement procedures window. Figure 22 Window of proposed procedure definition. 64 A.A. Elziny et al. Figure 23 Window of proposed procedure guide. Figure 24 Decision window. Figure 25 Precaution window for continuing procedure. An expert system to manage dispute resolutions in construction projects 65 3.2.1. Program structure By selecting Add questions icon, a new window will be opened as shown in Fig. 16; this window is considered the control room of the program where the data entry can control the questions. Figure 26 First ch The questions could be edited, sorted, or deleted from the database. Once the data have been entered, it will be saved automatically on the program database. 3.2.2. Entering the project profile The descriptive information specifies the project cost, dura- tion, completion date, work field, project location, project manager name, company name, address, phone, fax, and email as shown in Fig. 17. The user can fill the blank fields of data for his project details and then click on (submit/next) to con- tinue the program. 3.2.3. Assigning the project After entering the project details and click on submit, the pro- gram part I window (sources of disputes) will be opened as shown in Fig. 18. ecklist window. Figure 27 Second checklist window. Figure 28 Last decision window – arbitration stage. Figure 29 End message window. 66 A.A. Elziny et al. To read more about the major sources of disputes and its causes (sources of disputes guide) button could be pressed, or from sources of disputes icon at the program interface tool- bar, then a report window will be opened as shown in Fig. 19. Then the program goes to the next window to specify the main cause led to dispute according to the previous source chosen from the previous window as shown in Fig. 20. Then the program goes to part II to propose a proper ADR solution for current dispute according to dispute status and current situation as shown in Fig. 21. The program always suggests to follow negotiation stage as the first and best choice to settle dispute with a brief definition of the negotiation stage as shown in Fig. 22. The users have the ability to read more details about the procedure, feature of the negotiator, stage cost, and its dura- tion on different kinds of contracts (long and short term) by click on the button ‘‘Go to procedure’’ or escape this detail if so desired. To read more about the proposed procedure as shown in Fig. 23 (go to procedure) button could be pressed, or from set- tlement procedure icon at the program interface toolbar. The program gives an opportunity to try following this pro- cedure, then returning back to the program to check feedback as shown in Fig. 24. After applying the negotiation procedure, if it is found that this procedure is not appropriate for the situation the next stage should be moved as shown in Fig. 25. Figure 30 Edit project profile window. Table 1 Results of all case studies disputes summary. Case no. Type of contract Sources of disputes Actual procedurenstatus DRExM proposed procedurenstatus (1) FIDIC Contract [28] Contract management & financial issues Negotiation (solved) Negotiation (solved) (2) Petroleum Contract Contract Management & financial issues & other sources Mediation (solved) Mediation (solved) (3) Private Contract Contract management Arbitration (solved) Conciliation (4) Public Contract Financial issues & other sources Litigation (unsolved) Mediation Figure 31 ‘‘DRExM’’ result – case study No. (1). An expert system to manage dispute resolutions in construction projects 67 Now the program gives choice of two options (mediation via conciliation) stage according to the current situation, dis- pute status and the answers to the next checklist as shown in Fig. 26. Then as mentioned previously the program gives options to read more about the procedure and try applying the proposed procedure then returning back to the program to check feedback. Again if one found this procedure not appropriate for situ- ation one should move to the next stage. The program gives the choice of four procedures (DAB, fact-finding, mini-trials, and expert system) according to current situation and dispute status and respondent’ answers to the following checklist as shown in Fig. 27. As mentioned previously the program gives options to read more about the procedure and try applying it and making the decision about it. Finally, if all the previous proposed procedures are not appropriate for the situation the program suppose try to fol- low arbitration procedure as the last choice one may have to settle the dispute amicably as shown in Fig. 28. Figure 32 ‘‘DRExM’’ result – case study No. (2). Figure 33 ‘‘DRExM’’ result – case study No. (3). Figure 34 ‘‘DRExM’’ result – case study No. (4). 68 A.A. Elziny et al. At the end of each proposed procedure if dispute solved, one would have the following message as shown in Fig. 29 and navigate to the project report. 3.2.4. Edit the project profile This information can be edited by using project details edit command from Edit project icon. Once this information is added it will be automatically saved. This information will be included on all printed reports generated by ‘‘DRExM’’. If the user had an account on the program, he has the abil- ity to reassign his profile and answers, and his report will be corrected automatically, as shown in Fig. 30. 3.3. System validation The objective of this section is to check the validity of the expert system. Validation can be defined as the process of mak- ing sure that the system operates as desired. Nosair [16], Stated that validation is the process of making sure that the system Figure 35 ‘‘DRExM’’ All case studies disputes settlement chart. An expert system to manage dispute resolutions in construction projects 69 has a proper level of reality. Four construction projects were selected for the proposed system. After providing some general data regarding the four case studies such as project type, nat- ure, contract period and price, the main inputs are added. These are concerned with the sources of disputes that have been identified as contract management, contract documents, financial issues, project issues, and others. The results of the four case study applications are shown in Table 1. The table shows a comparison between the predicted resolution method from the system application of disputes and the actual resolution of disputes occurs during the project con- struction. This can give clear picture regarding the validity of the proposed system. Table 1 summarizes the results of the four case study appli- cations. The table also shows the actual resolution of disputes and those resolution predicted by the proposed system. For instance, the first and second case study has been respectively solved during construction by negotiation and mediation and the same methods were predicted to be solved by the expert system. This shows us that DRExM gives results identical to the actual situation in sites with simplified presentation of results as shown in Figs. 31 and 32. This may be considered as a good indicator for the system validation. For the third case study the dispute has been solved during construction by arbitration after negotiation has been failed. On the other hand by using the expert system and depending on parties’ relationship and dispute nature, the parties can use conciliation instead of arbitration to solve the dispute. This confirms that ‘‘DRExM’’ gives results better than the actual resolution method that indicates the benefits of using ‘‘DRExM’’ to go to conciliation instead of arbitration that saves time and cost with simplified presentation of results and minimum durations as shown in Fig. 33. As shown in the forth case study, the dispute has not been solved yet after using negotiation and failed, then the parties go to litigation without using other amicable settlement stages. On the other hand, the expert system gives results better than the method actually occurred on site that indicate the benefits of using ‘‘DRExM’’ to go to mediation instead of litigation that saves time and cost with simplified presentation of results and minimum durations as shown in Fig. 34. The system also illustrates the settlement stages for the four case studies as shown in Fig. 35. Finally, it is fair to say that the system can describe the real situation of the construction disputes in Egypt at an appropri- ate end level of confidence. 4. Conclusions � The research illustrates that the most used dispute resolu- tion methods are negotiation, mediation and arbitration respectively. � The study proposes a reliable and accurate method to quan- tify and analyze sources of construction disputes. The most important source of disputes was ‘‘contract management (74.04%)’’, the second was ‘‘contract documents (71.49%)’’, the third was financial issues (67.80%), the fourth was ‘‘project related issues (63.92%)’’, and the lowest one was ‘‘other sources’’ such as force majeure, and loose of construction laws, (61.58%). � The study indicates that the contract management can be considered the main factor that can affect the existence of disputes due to many reasons such as the issues related to the owner and the contractor, their management of the con- tract, time schedule prepared by the contractor and required update. � The proposed program ‘‘DRExM’’ is capable of presenting ADR techniques. The program results matched with actual ones of the case studies with simplified presentation of results. � The ADR for disputes by using ‘‘DRExM’’ saves time and cost with simplified presentation of results and minimum durations. � The benefits of the ‘‘DRExM’’ program confirmed that the companies should have program to facilitate the dispute Management and to assess the current status of the dispute then propose the alternative settlement procedure instead of going direct to arbitration or litigation. � The architecture of the program is designed to be open, flex- ible and upgradable, allowing it to be customized for indi- viduals and corporations with relative ease. 70 A.A. Elziny et al. � This study would be of added value for contractors and owners who intend to work in the Egyptian construction market and face difficulties to deal with disputes. Acknowledgments This paper is a part of the Ph.D. thesis of the first author under the supervision of the other authors. I would also like to thank the Port Said University and Petrojet Company for their aids during the research. References [1] Richard B. An overview of international dispute settlement. Emory J Int Disput Resolut 1986;1. [2] Diekmann J, Girard M. Are contract disputes predictable. J Construct Eng Manage 1995;121(4):355–63. [3] Corby S. Public sector disputes and third party intervention. Prepared For Acas; 2003. [4] Cheung SO, Yiu TW. A study of construction mediator tactics Part I: taxonomies of dispute sources, mediator tactics and mediation outcomes. J Build Environ 2007:752–61. [5] Ndekugri I, Russell V. Disputing the existence of a dispute as a strategy for avoiding construction adjudication. Eng Constr Archit Manage 2006;13(4):380–95. [6] Bunni N. Recent developments in dispute resolution under the FIDIC. In: First international conference on engineering arbitra- tion, Bahrain; 2000. [7] Thomas D. Dispute resolution for long term infrastructure Projects. In: First international conference on engineering arbi- tration, Bahrain; 2000. [8] Steen Richard H. Alternative dispute resolution in the construc- tion industry. In: Steen Richard H, editor. Attorneys at law. LLC; 2002. [9] Thompson Roxene M. Efforts to manage disputes in the construction industry. A comparison of The New Engineering Contract and The Dispute Review Board; MSc. Dissertation, Blacksburg, Virginia; 1998. [10] UNITAR/DFM Online course on arbitration and dispute reso- lution; 2004 <http://www.unitar.org/dfm.htm>. [11] Shin K, Molenaar K. Prediction of construction disputes in change issues. Constr Congr 2000;58(278):534–42. http:// dx.doi.org/10.1061/40475. [12] Fenn PLowe, Speek C. Conflict and dispute in construction contract management economics. J Manage Eng ASCE 1997;18:1–120. [13] Hall JM. Ineffective communication: common causes of construc- tion disputes. Alliance’s Adv Council Legal Notes 2000;13(2). [14] Carmicheal DG. Disputes and international project. Liaise: A.A. Balkema Publisher; 2002. [15] Zakzok Tarek M. Managing disputes resolution in construction projects. MSc. Thesis. Alexandria University, Faculty of Engineering; 2001. [16] Nosair Ibrahim A. Exscd an expert system for construction disputes. Mansoura Eng J 2007;32(17):C.48–58. [17] Nicholas Gould. Partner Conflict Avoidance and Dispute Resolution RICS Professional Guidance, UKRICS QS & Construction Standards – GN 91/201 First edition, Guidance; 2012 [Note: 2]. [18] Klein Howard. Great Britain ‘‘Alternative Dispute Resolution Procedures Used to Resolve Construction Disputes in The UK’’. Commercial Management I, Munich, Germany; 2006. [19] Verster JJP. Real time integrated cost planning and control: mitigation and resolution of claims. In: Quantity surveying workshop paper. University of the Free State. Bloemfontein, South Africa; 2006. [20] El-Adaway Isalam H, Ezeldin Samir A. Dispute review boards: expected application on Egyptian large-scale construction pro- jects. J Prof Issues Eng Educ Pract 2007;133(4):365–72. [21] El-Adaway Isalam H, Ezeldin A Samir, Yates JK. Arbitral Tribunal proceedings case study: Egyptian large-scale construc- tion project. J Legal Affairs Dispute Resolut Eng Constr 2009;1(3):147–53, doi:10.1061/(Ace)1943-4162 (2009)1:3(147). [22] Okharedia AA. The emergence of alternative dispute resolution in South Africa: a lesson for other African countries. In: A paper presented at the 6th IIRA African regional congress of industrial relations, Lagos Nigeria; 2011. [23] Odeh AN, Battaineh HT. Causes of construction delays: tradi- tional contracts. Int J Project Manage 2002;2(20):67–73. [24] Nilsson NJ. Artificial intelligence. A new synthesis. Morgan Kaufmann Publisher; 2000. [25] Coppin B. Artificial intelligence illuminated. USA: Jones and Bartlett Publishers, Inc.; 2004. [26] Cheung SO, Au-Yeung RF, Wong VWK. A CBR based dispute resolution process selection system. Int J IT Archit Eng Construct 2007;2(2):70–9. [27] Chau KW. Application of a PSO based neural network in analysis of outcomes of construction claims. Automat Construct 2007;16(5):642–6. [28] FIDIC. Supplement to first edition of conditions of contract for construction for building and engineering works designed by the employer, Geneva; 1999. Eng. Abdelkader ahmed Elbadry Elziny, Planning and contract administration Engineer at Petrojet company, has a wide experience in the field of project management applications for mega projects in Egypt and outside, and he did M.Sc. in civil engineering (2011) (project management), Port Said University, Faculty of Engineering. Member of the international Arbitration organization, Participate as lecturer in Petrojet Company for training the new project managers in how to use project management in managing the projects and the Yemen delegation in how to identify and manage deviations in petroleum projects. Participate as lecturer at the first Arab conference for Engineering Arbitration in construction projects, Published paper at PSEJ (2011). Dr. Mohamed A. Mohamadien President of Suez Canal University, Prof. of Steel Structure, Civil Eng. Dept., Faculty of Engineering, Suez Canal University. Doctor of Philosophy Degree in Civil Engineering 1991, Bath England University, published researches in the field of steel structure and project management and supervision on sev- eral theses of Master and PhD. Member of Engineering Syndicate, Member of Egyptian Engineering Society (EES), Member of Canadian Engineering Society, Member of Egyptian consulting Engineers, worked as a consultant for Egyptian and Saudi Arabia authorities and companies and attended many national and interna- tional conferences. http://refhub.elsevier.com/S2090-4479(15)00068-4/h0005 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0005 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0010 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0010 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0020 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0020 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0020 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0025 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0025 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0025 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0040 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0040 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0040 http://%20www.unitar.org/dfm.htm http://dx.doi.org/10.1061/40475 http://dx.doi.org/10.1061/40475 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0060 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0060 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0060 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0065 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0065 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0070 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0070 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0080 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0080 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0100 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0100 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0100 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0105 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0105 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0105 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0105 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0115 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0115 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0120 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0120 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0125 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0125 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0130 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0130 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0130 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0135 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0135 http://refhub.elsevier.com/S2090-4479(15)00068-4/h0135 An expert system to manage dispute resolutions in construction projects 71 Dr. Hassan M. Hassan Ibrahim, Head of Civil Engineering Department – Prof. of Concrete Structures, Faculty of Engineering, Port Said University, published researches in the fields of concrete structures and construction pro- ject management and also supervised several theses of Master and PhD in both fields. Dr. Eng Moataz Khalil Ibrahim Abd El Fattah, Business Development Executive General Manager at Petroleum Projects and Technical Consultation Company (PETROJET), Doctor of Philosophy Degree in Civil Engineering 2004 (Project Management), Suez Canal University, Faculty of Engineering, Certified project management consultant, Certified Project Manager (CSPM) – IPMA (Level B) Management Engineering Society, Board Member of Management Engineering Society (MES), Cairo, Egypt, attended many national and interna- tional conferences and have Academic experience as Lecturer, Supervisor, Author and Assessor of Project Management for Port Said and Ain Shams University. An expert system to manage dispute resolutions in construction projects in Egypt 1 Introduction 2 Questionnaire design 2.1 Domain of experts 2.2 Method of data analysis 2.3 Results and discussions 3 System design and equations 3.1 Architecture of the expert system and verification 3.2 System operation 3.2.1 Program structure 3.2.2 Entering the project profile 3.2.3 Assigning the project 3.2.4 Edit the project profile 3.3 System validation 4 Conclusions Acknowledgments References 
work_whsc6wp5sjgktcjru5hnfjhqyu	----	Probabilistic Expert Systems for Forensic Inference from Genetic Markers A. P. DAWID University College London J. MORTERA Università Roma Tre V. L. PASCALI Università Cattolica del ‘‘Sacro Cuore’’, UCSC, Rome D. VAN BOXEL Carnegie Mellon University ABSTRACT. We present a number of real and fictitious examples in illustration of a new approach to analysing complex cases of forensic identification inference. This is effected by careful restructuring of the relevant pedigrees as a Probabilistic Expert System. Existing software can then be used to perform the required inferential calculations. Specific complications which are readily handled by this approach include missing data on one or more relevant individuals, and genetic mutation. The method is particularly valuable for disputed paternity cases, but applies also to certain criminal cases. Key words: Bayesian networks, DNA profile, forensic identification, incomplete evidence, mutation, paternity testing 1. Introduction and background In a simple problem of forensic DNA identification, we have a ‘‘trace’’ biological sample of unknown provenance, and further ‘‘reference’’ samples obtained from known individuals. All the samples are typed by DNA profiling, and it is desired to use the resulting data to shed light on the origin of the trace sample. More complex problems also arise: thus in a case of disputed paternity, DNA information on the child can be regarded as supplying partial information about its father’s DNA (Dawid & Mortera, 1998). DNA profiles as currently used consist of measurements on a number of genetic markers, typically ‘‘short tandem repeat’’ (STR) markers (Weber & May, 1989), chosen by forensic geneticists for their usefulness. For each marker we can observe its genotype, comprising two genes (or bands), one inherited from the mother and the other from the father (although it is not possible to observe which is which). Each marker has a finite number (up to around 20) of possible values (alleles), generally taking positive integral values, for each of its two constituent bands. The markers used for forensic identification are chosen to be located on different chromosomes, and hence segregate independently. It is often reasonable to assume random mating within an appropriate population, which induces both Hardy–Weinberg and linkage equilibrium, so that different markers, as well as different bands of the same marker, behave entirely independently. We shall assume this throughout. Databases have been gathered from which the frequency distributions of the different markers, in various populations, can be estimated. In this paper we use estimates based on data collected by the Forensic Genetics Laboratory, Catholic University of the Sacred Heart � Board of the Foundation of the Scandinavian Journal of Statistics 2002. Published by Blackwell Publishers Ltd, 108 Cowley Road, Oxford OX4 1JF, UK and 350 Main Street, Malden, MA 02148, USA Vol 29: 577–595, 2002 (UCSC), Rome. However, as more data are collected, so these estimates may be refined (up-to- date information on the UCSC database may be obtained from: http://www.mclink.it/ personal/MD1696/data/freqask.htm). Values used in this paper should not be regarded as definitive. We are here concerned with the general statistical problem of inferring identification on the basis of the DNA and other evidence at hand (Dawid & Mortera, 1996). This can in principle be solved by determining the relative likelihoods, induced by the full data, for the various competing hypotheses. However, in many cases samples are not available for one or more individuals of interest, and instead we only have indirectly relevant information, perhaps through genetic typing of their relatives. Calculation of the desired likelihoods, on the basis of such partial or incomplete data, can then become both conceptually and computationally demanding, particularly when we allow for the possibility of mutation during gene transmission. Here we present a new approach to solving such a problem, by reformulating it as a Probabilistic Expert System (PES): a joint graphical and numerical representation that can be implemented and processed in general-purpose computer software. In a PES, complex inter- relationships are broken down into simple modular units, out of which the entire graphical representation is constructed. The resulting representation then forms a framework for the application of fast and efficient computational algorithms. A complex genetic pedigree fits particularly smoothly into the PES model. The nuclear family relationships constitute natural modular building blocks of the representation, to such an extent that terms such as ‘‘parent’’ and ‘‘child’’ have become part of the general terminology of PES, even for entirely non-genetic applications. The conditional probability tables required are simple and uncontroversial, being given by Mendelian laws of inheritance and logical relationships between genes and genotypes. And the conditional independence relations forming the backbone of a PES representation are likewise natural in the setting of genetic inheritance, since, conditional on its parent’s genes, a child’s genes are entirely independent of those of all other individuals (more precisely: of those of its non-descendants). In this paper we exhibit how to pass from an initial pedigree representation of a forensic identification problem to an appropriate graphical PES representation, and how to use information on gene frequencies, Mendelian inheritance and mutation processes to set up the numerical part of the PES specification. We describe how one can apply existing software to calculate the desired likelihoods. The method is illustrated on a number of DNA identification problems of varying degrees of complexity, including real paternity testing cases and an artificial criminal identification example. 2. Disputed paternity 2.1. Basic set-up In the simplest case of disputed paternity a man is alleged to be the father of a child, but disputes this. DNA profiles are available on the mother m, the child c, and the putative father pf. The disputed pedigree can be represented as in Fig. 1, where a square indicates a male and a circle a female, and tf denotes ‘‘true father’’; grey indicates that a DNA profile is available for that individual. On the basis of these data, we need to assess the likelihood function over possible hypotheses as to the true father. Often these hypotheses are reduced to two: the true father either is the putative father, or else is drawn randomly from the population, being unrelated to the mother or putative father. Throughout this paper we shall make the latter simplifying assumption whenever the true father is not otherwise identified. 578 A. P. Dawid et al. Scand J Statist 29 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. Because the markers used are on different chromosomes, and we are assuming random mating, we have complete independence across markers. It follows that we can consider the markers one at a time: we simply have to obtain the likelihood ratio on the basis of the data for each marker separately, and finally multiply these values together to obtain the overall likelihood ratio based on all the data on the full collection of markers. So consider now the measured genotypes, from all three parties, for some fixed marker. To find the associated likelihood function we need to calculate the joint probability of this triplet of observed genotypes, under either hypothesis as to paternity. Making the reasonable assumption that, before we have any data on the child, the identity of the true father is independent of the profiles of the mother and the putative father, we can transfer attention to the conditional probability of the child’s genotype, given the other two. Under either hypothesis, this is calculated simply: under paternity, we just apply Mendel’s laws of segregation; under non-paternity, we require (estimates of) the frequencies of relevant marker alleles among the population at large. From these likelihoods we readily obtain the desired likelihood ratio. Using Bayes’s Theorem, this can then be combined with the prior odds of paternity, based on external evidence, in order to obtain the posterior odds for paternity. As an illustrative example, suppose that the data, for marker FES, are: child’s genotype ¼ f12; 12g, mother’s genotype ¼ f10; 12g, putative father’s genotype ¼ f10; 12g. The population frequencies of alleles 10 and 12 are, respectively, 0.28425 and 0.25942. In this case, conditioning on the genotypes of mother and putative father, we see that the child’s genotype will be as observed if and only if both the mother and the true father contributed allele 12 to the child. This event has probability 0:5 � 0:5 if the true father is the putative father, and probability 0:5 � 0:25942 if the true father is, instead, some unrelated individual from the population. Thus the likelihood ratio in favour of paternity (based on these data for marker FES alone) is 0:5=0:25942 ¼ 1:9274. 2.2. From pedigree to expert system In the above simple problem the calculations are trivial, and have long been widely implemented much as we have described (Essen-Möller, 1938): we certainly do not need to develop any clever new methods of solution. However, we wish to extend our analysis to more complex problems, particularly those with missing data, for which the calculations are by no means trivial. Purely as a gentle lead-in to this extension, it will be valuable to reformulate the simple paternity problem of section 2.1 above as a ‘‘Probabilistic Expert System’’, or PES (Cowell et al., 1999). A PES is a representation of a complex probability structure by means of a directed acyclic graph, having a node for each variable, and directed links describing probabilistic causal relationships between variables. The overall probability structure is completely determined by specifying, as desired, the conditional probability tables for each variable given its ‘‘parents’’ in the graph. On the basis of probabilistic conditional independence properties embodied in the graph, the complex global model then decomposes into simpler localized submodels, providing a framework for the application of fast and efficient computational algorithms for Fig. 1. Simple paternity pedigree. Scand J Statist 29 Forensic inference from genetic markers 579 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. exact calculation of marginal and conditional probabilities, and much else beside (Dawid, 1992; Spiegelhalter et al., 1993). The calculations can be described as effected by ‘‘propaga- tion’’ of information through the network: this involves efficient organization of simple calculations affecting one local cluster of variables at a time, but spreading throughout the whole network to yield the correct overall answers. These propagation algorithms have now been implemented in widely available software, such as HUGIN (http://www.hugin.dk), GENIE (http://www2.sis.pitt.edu/�genie) or XBAIES (http://www.staff.city.ac.uk/�rgc), thus enabling many otherwise intractable complex problems to be solved. Because we are at liberty to choose which unobserved variables to include in a PES representation, there can be many such representations, some more manageable than others. Finding an appropriate representation is crucial, as the efficiency, or even the viability, of the computational routines is highly sensitive to the topology of the graphical structure. PES construction is to some extent an art-form, but can be guided by scientific and logical considerations. The main contribution of this paper—beyond advertising the simplicity and benefits of general PES ideas, methods and technology—is to present and analyse what we consider to be good representations for the type of paternity and identification problems we address. However, for other problems, other kinds of representation may be better. The search for good representations for specific problems is an important task for continuing research in this area. Our graphical PES representation of the simple disputed paternity problem (for a single marker) is displayed in Fig. 2. Purely for presentational purposes we colour-code the nodes to distinguish different types: grey again marks a node which is observed, while black represents a disputed hypothesis. The arrows represent (sometimes degenerate) probabilistic influences. For a detailed description of the semantics of such diagrams, see Cowell et al. (1999). To attain the most efficient and simplified representation, we have aimed to represent the problem at as deep and disaggregated a level as possible. First, we need to identify, and create nodes for, all the interesting variables in the problem. These do not necessarily have to correspond to observables, although all relevant observables must be represented. In our model, in order to maximize the efficiency of the calculations as well as the logical clarity of the representation we choose to disaggregate each individual’s genotype into its constituent, unobserved, paternally and maternally inherited genes. We thus have a node pfmg representing ‘‘putative father’s maternal gene’’, etc., as well as nodes for the observed genotypes: mgt represents ‘‘mother’s genotype’’, etc. Table 16 provides a summary of all the notation used in this article. A related representation (see e.g. Thompson, 2000) has nodes for genotypes, but not for their constituent bands: information as to whether a gene was inherited from the mother or father is represented by means of additional (unobserved) binary meiosis or segregation Fig. 2. Simple paternity network. 580 A. P. Dawid et al. Scand J Statist 29 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. indicators. Our own representation is perhaps a little more transparent and straightforward, at least for present purposes. An important feature of Fig. 2 is the explicit introduction into the graph of the ‘‘target’’ (or ‘‘hypothesis’’ or ‘‘query’’) node, tf ¼ pf? (‘‘true father ¼ putative father?’’), representing the hypotheses of interest. We are thus considering this query on exactly the same footing as observable or latent genetic information. This is a natural step within the general PES approach, although perhaps less natural from the standpoint of genetics, where graphical modelling is usually restricted to representing pedigree structure. There is a large number of powerful specialist pedigree analysis programs available that work with such pedigree representations (e.g. LINKAGE (http://linkage.rockefeller.edu/soft/linkage/), MENDEL (http://www.biomath.medsch.ucla.edu/faculty/klange/software.html), among others (see list at http://linkage.rockefeller.edu/soft/list.html)); and any single one of the problems we consider in this paper could probably be handled better by some one of those programs. However, none of these has the degree of generality and extendibility of the more broadly- based PES technology. There are alternative ways of building and analysing PES representations, without explicitly representing the hypothesis node in the graph. Indeed, for some problems, as in section 4.1 below, this may be the only efficient way to proceed. We prefer to include an explicit hypothesis node wherever possible, since this is simpler to interpret, and allows one to read off directly, simply by querying the ‘‘target’’ node, the quantity of most interest: the likelihood ratio (for the relevant marker) in favour of paternity, on the basis of the observed evidence. To complete our PES representation, we need to supply the numerical part of the specification. This requires that we give, for each node in the network, the table of probabilities for its various values, conditional on each configuration of values at its ‘‘parent’’ nodes (if any). We specify these tables as follows. At nodes corresponding to ‘‘founder’’ genes we use population gene frequencies. A child’s maternal gene is obtained by drawing, at random, one of it’s mother’s two (paternal and maternal) genes, mmg and mpg, and similarly for its paternal gene (drawing from its father’s genes). The table for a genotype node is degenerate, encoding the simple deterministic relationship between an unordered pair of values and its two constituents. The true father’s paternal (maternal) gene is either identical with the corresponding gene of the putative father, or else generated from the relevant population gene-frequency distribution, depending on the value of the hypothesis node tf ¼ pf?. At the hypothesis node itself, we could set an initial distribution to represent actual prior beliefs about paternity, on the basis of other evidence in the case. However, we shall instead insert a purely formal uniform prior at this node (prior probability of paternity ¼ 0.5, as in Table 2)—for then the formally calculated ‘‘posterior odds’’ on paternity will in fact be numerically identical with the likelihood ratio in favour of paternity based on this marker. Under our assumption of independent markers, these values can then simply be multiplied together to yield the overall likelihood ratio, based on all the markers; and this can then, if desired, be combined with genuine prior beliefs to obtain the correct overall posterior odds on paternity. (It would be possible, but more complicated, to handle the multi-marker data directly in a sequential process: thus we could start with a genuine prior probability distribution at the hypothesis node, and, introducing one new marker at a time, propagate the associated evidence to obtain the induced posterior distribution at the hypothesis node, based on that marker’s data—this then becoming the prior to be used with the next marker.) Consider again the illustrative example of section 2.1, with case data: cgt ¼ f12; 12g, mgt ¼ f10; 12g, pfgt ¼ f10; 12g. Tables 1, 2, 3, 4 and 5 show the (conditional) probability tables corresponding to nodes pfpg, tf ¼ pf?, tfpg, pfgt and cpg, respectively, for Scand J Statist 29 Forensic inference from genetic markers 581 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. marker FES. Only the actually observed alleles 10 and 12, and the aggregation of all unobserved alleles, indicated by x, need to be represented in the probability tables. After entering the case evidence at the relevant genotype nodes, we ‘‘propagate’’ it throughout the network, using the software. We can then interrogate the hypothesis node to find its updated probability distribution, conditional on the evidence. The required likelihood ratio, based on the data for this marker, is obtained from the marginal formal posterior distribution at node tf ¼ pf?, as given in Table 6: LR ¼ 0:6584=0:3416 ¼ 1:9274—in agreement with the analysis of section 2.1. Our purely technical use of an ‘‘artificial’’ uniform prior probability of 0.5, to derive likelihood ratios, should be clearly distinguished from the common forensic practice of calculating, and quoting in court, a ‘‘probability of paternity’’ based on this uniform prior Table 1. Probability table for pfpg pfpg: 10 12 x 0.28425 0.25942 0.45634 Table 2. Probability table for tf=pf? tf=pf?: yes no 0.5 0.5 Table 3. Conditional probability table for tfpg given tf=pf? and pfpg tf=pf? yes no pfpg 10 12 x 10 12 X tfpg: 10 1 0 0 0.28425 0.28425 0.28425 12 0 1 0 0.25942 0.25942 0.25942 x 0 0 1 0.45634 0.45634 0.45634 Table 4. Conditional probability table for pfgt given pfmg and pfpg pfmg: 10 12 x pfpg: 10 12 x 10 12 x 10 12 x pfgt: 10–10 1 0 0 0 0 0 0 0 0 10–12 0 1 0 1 0 0 0 0 0 10–x 0 0 1 0 0 0 1 0 0 12–12 0 0 0 0 1 0 0 0 0 12–x 0 0 0 0 0 1 0 1 0 x–x 0 0 0 0 0 0 0 0 1 Table 5. Conditional probability table for cpg given tfmg and tfpg tfmg: 10 12 x tfpg: 10 12 x 10 12 x 10 12 x cpg: 10 1 0.5 0.5 0.5 0 0 0.5 0 0 12 0 0.5 0 0.5 1 0.5 0 0.5 0 x 0 0 0.5 0 0 0.5 0.5 0.5 1 582 A. P. Dawid et al. Scand J Statist 29 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. (Essen-Möller, 1938). Such a prior assumption will often be unreasonable. All the non-DNA evidence in the case should be used to construct a sensible and defensible prior probability which is then combined with the likelihood ratio (summarizing all DNA evidence) to form a posterior probability. However, this is a task that should properly be left to the judge or jury. We advocate the presentation, along with the overall likelihood ratio, of a table, such as Table 8 (see section 2.3.1 below), where a range of posterior probabilities is given corresponding to a range of possible prior probabilities. For a general discussion of the presentation of statistical evidence in court, see Dawid (2002). We again stress that we are not recommending the use of a PES for the simple paternity case illustrated in this section, which can be solved by a couple of lines of simple algebra. It is presented merely as an illustrative basic model of how a PES can be formulated, which is then suitable for extension to the more complex cases that we treat below—cases that can not be handled by simple algebra. In particular, the PES for these more complex problems can be built out of the same fundamental local modules that we have already described for the simple problem above. 2.3. Missing data In certain cases, the DNA profiles of one or more of the ‘‘principal actors’’ in the story are not available, but there is indirect evidence, in the form of DNA profiles of various known relatives. Then simple arguments such as those of section 2.1 can no longer be applied, and their appropriate extension to such a case may not be obvious or practicable. However, it is still straightforward to construct and apply a Probabilistic Expert System representation of the problem, this being now expanded to include the other measured individuals and the relationships between all individuals involved. We illustrate this with two real examples from the case-work of the UCSC Forensic Genetics Laboratory. 2.3.1. Paternity case 1 In this case, the only DNA samples available were on the disputed child c1, on the putative father pf’s undisputed child c2 by a different mother, and on pf’s brother b. In particular, no samples were available for the putative father pf, nor for his parents gf and gm, nor for the mother of either child. The case data, and relevant allele frequencies, are given in Table 7. Figure 3 displays the pedigree, and Fig. 4 the corresponding Probabilistic Expert System network. The necessary conditional probability tables are formed exactly as before. The Probabilistic Expert System again readily supports entering of the evidence (at the grey ‘‘observation nodes’’), rapid propagation of this evidence through the network, and interrogation of the black ‘‘target’’ node to obtain the desired inference. Assuming independence across markers, the overall likelihood ratio in favour of paternity is obtained as the product of all terms in the last column of Table 7, viz. 13.066. Table 8 shows the implied posterior probability of paternity (tf ¼ pf? ¼ yes), for various values of the prior probability. Since the jury may hold or wish to consider a range of prior probabilities, such a table can be a useful way of presenting the impact of the DNA evidence to the court. Table 6. Posterior probability table for tf=pf? tf=pf?: yes no 0.65840 0.34160 Scand J Statist 29 Forensic inference from genetic markers 583 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. Table 7. Observed genotypes, and their frequencies, for incomplete paternity data, Case 1 Individual: b Frequency c1 Frequency c2 Frequency Likelihood ratio Marker TH01 7 0.165 6 0.263 6 0.263 9 0.200 6 0.263 9 0.200 0.672 VWA 17 0.298 15 0.107 15 0.107 17 0.298 17 0.298 17 0.298 1.788 D3S1358 15 0.278 15 0.278 15 0.278 17 0.220 17 0.220 16 0.233 1.450 FGA 18 0.006 22 0.163 23 0.163 26 0.028 26 0.028 26 0.028 8.954 TPOX 8 0.508 8 0.508 9 0.107 9 0.107 11 0.271 10 0.089 0.459 CSF1PO 10 0.250 10 0.250 10 0.250 12 0.380 12 0.380 12 0.380 1.504 D5S818 11 0.392 11 0.392 11 0.392 11 0.392 13 0.165 11 0.392 1.207 D7S820 9 0.130 10 0.240 9 0.130 12 0.160 13 0.027 10 0.240 0.428 D13S317 11 0.312 11 0.312 11 0.312 11 0.312 11 0.312 12 0.286 2.346 Overall 13.066 Fig. 3. Pedigree for incomplete paternity data. Case 1. Fig. 4. Network for incomplete paternity data. Case 1. Table 8. Posterior probability of paternity for Case 1 Prior probability: 0.001 0.01 0.1 0.3 0.5 0.7 0.9 Posterior probability: 0.013 0.117 0.592 0.848 0.929 0.968 0.992 584 A. P. Dawid et al. Scand J Statist 29 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. 2.3.2. Paternity case 2 In this problem, DNA profiles were available on the disputed child c1 and its mother m1, on the putative father’s two full brothers b1 and b2, on the undisputed child c2 of the putative father by another mother m2, and on m2. No samples were available for the putative father pf, nor for his parents gf and gm. The pedigree is given in Fig. 5. The network for the corresponding Probabilistic Expert System is now given by Fig. 6. The detailed case data and allele frequencies (for 10 markers) are not given here, but can be supplied on request. Notwithstanding the greatly increased complexity, it is once again completely straightforward, using PES software, to enter the observed evidence, propagate, and interrogate the target node to obtain the required likelihood ratio. This was done using GENIE: the overall likelihood ratio obtained in favour of paternity was 1303. Networks such as the above can also be used to analyse the possibility that the true father is some individual in the pedigree other than the putative father. Further relatives, measured or unmeasured, could be investigated by suitable extension of the pedigree and corresponding elaboration of the network. 3. Mutation A problem that can complicate forensic inference from DNA profiles is the possibility of mutation of the DNA between generations. Indeed, the microsatellite markers typically used for forensic purposes are known to be particularly prone to mutation, with overall mutation rates of between 5 � 10�4 and 7 � 10�3 per generation (Brinkmann et al., 1998). It is thus possible, for example, that a putative father may seem to be excluded by the evidence, whereas in fact he is the true father, but mutation has led to his passing on a prima facie impossible Fig. 5. Pedigree for incomplete paternity data. Case 2. m1mgm1pg m1gtyc1mg tf=pf? gfmggfpg pfpg gmmggmpg pfmg c1pg c1gt b2pg b2mgb1mgb1pg b2gtb1gt m2pg m2mg m2gt c2mg c2pg c2gt tfpg tfmg Fig. 6. Network for incomplete paternity data. Case 2. Scand J Statist 29 Forensic inference from genetic markers 585 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. allele to the child. Particularly in cases where such ‘‘exclusion’’ is based on a single marker, all other markers yielding data consistent with paternity, the possibility of a mutation may need to be taken seriously. Within a Probabilistic Expert System representation, mutation may be incorporated into the process whereby a child’s (say) maternal gene is determined by his mother’s two genes. First one of these is chosen at random, as before, to form the ‘‘original’’ (transmitted) gene; then this spawns a new ‘‘actual’’ (inherited) gene, according to some specified mutation process. (In fact our ordering of these two stages is the opposite of Nature’s, but the end result will be equivalent so long as the same mutation process operates on the male and female germlines. Our construction involves fewer mutation events, thus yielding a simpler and more efficient graphical representation.) The ‘‘actual’’ genes are thus those which can be observed as constituents of genotypes, while the ‘‘original’’ genes are unobservable. With this extension, and using e.g. pfopg and pfapg to represent ‘‘putative father’s original paternal gene’’ and ‘‘putative father’s actual paternal gene’’, the network for the ‘‘simple’’ paternity problem of section 2.2 is transformed into that shown in Fig. 7. In fact in this new problem it is still possible to perform the calculations directly, without using a PES, although the expressions now become considerably more complex (Dawid et al., 2001). Once again, we develop the expert system formulation of this problem mainly as a model for handling more complex cases. In addition to the previous probability specifications, we need to specify the transition matrix of mutation rates, whereby an original gene mutates into an actual gene. There are various sources of data that can be used to supply estimates of overall mutation rates, but data on rates of transition between specific alleles are sparse, so that we need to make some tentative assumptions on the structure of the transition matrix. For real applications it will be important to investigate the sensitivity of the conclusions to varying assumptions about overall and transition-specific mutation rates (Dawid et al., 2001). Table 9 gives estimated overall mutation rates obtained from data from UCSC, and from Brinkmann et al. (1998), who investigated 11000 meioses for nine STR markers. We have used the estimate l ¼ ðs þ 0:5Þ=ðn þ 1Þ, where s is the number of observed mutations and n the number of observed meioses, which can be regarded as a Bayesian estimate under the Jeffreys Beta 1 2 ; 1 2 � � prior distribution for a binomial proportion. In particular, this simple Bayesian Fig. 7. Simple network with mutation. 586 A. P. Dawid et al. Scand J Statist 29 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. computation avoids zero estimates. For markers where mutation data were not available the highest observed rate, 2:69 � 10�3, was taken. For a given marker, let the mutation transition matrix be denoted by Q ¼ ðqijÞ, where qij denotes the probability of a mutation to ‘‘actual’’ allele j, from ‘‘original’’ allele i. It is reasonable to assume equilibrium, i.e. that the vector of gene frequencies p ¼ ðp1; . . .; pkÞT is constant over time. Then p must be a stationary distribution for the associated transition matrix Q, i.e. pTQ ¼ pT. Given p and the overall mutation rate l, we can construct a transition matrix Q having reasonable properties. Stationarity will be assured if Q is chosen to satisfy the detailed balance condition: piqij ¼ pjqji: ð1Þ We can restrict attention to states having pi > 0. Let S ¼ ðsijÞ be a symmetric matrix having sijP0 for i 6¼ j and P j sij ¼ 0 for all i, and let k be an adjustable positive parameter. We define qij ¼ ksij=pi ði 6¼ jÞ; ð2Þ and qii ¼ 1 � X j6¼i qij ¼ 1 þ ksii=pi: ð3Þ The detailed balance equation (1) will then be satisfied. The overall mutation rate is l ¼ 1 � X i piqii ¼ k � � X i sii ! ; ð4Þ so that we must take k ¼ l � X i sii ! : , ð5Þ In order to ensure qiiP0, all i, we require kO minif�pi=siig. Consequently, for given S there is an upper limit on the overall mutation rate that can be obtained from this model. It is biologically reasonable to assume that an allele is more likely to mutate to a neighbouring allele than to one further away, and we choose the matrix S accordingly. Specifically, we take sij ¼ aji�jj, for i 6¼ j, where a is a fixed constant: this is similar to the step- wise mutational model of Valdes et al. (1993). The smaller is a, the greater is the probability that a mutation will be to a closely neighbouring allele. Using the overall mutation rates in Table 9 and the procedure described above, we obtain, for each marker, the mutation transition matrix to be used at all nodes representing the ‘‘actual’’ transmitted gene, such as capg of Fig. 7. Table 10 shows a real example from UCSC casework, analysed using the PES in Fig. 7. Tables 11 and 12 give the allele frequency distribution and the corresponding mutation transition matrix, calculated as described above with a ¼ 0:5, for marker VWA. Again, only the observed alleles in the case at hand, and x, representing the aggregation of all other alleles, are required. (In principle such aggregation could now lead to violation of the conditional independence properties represented in the Table 9. Estimated overall mutation rates per thousand generations Marker: F13 VWA D21S11 D1S80 TH01 FES MBP APO-B COL2A1 Mutation rate: 2.55 2.23 2.69 2.69 2.49 1.76 2.69 2.69 2.69 Scand J Statist 29 Forensic inference from genetic markers 587 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. Expert System. Full correctness could be ensured by avoiding the aggregation step; however, because of the low mutation rates, in practice the effect of such aggregation will be entirely negligible.) Note that, for markers F13 and VWA, if mutation was not a possibility the putative father would be excluded by the evidence. For the ‘‘non-excluding’’ markers, i.e. those other than F13 and VWA, introducing the possibility of mutation does not affect the likelihood ratio, to 3 significant figures. It does, of course, have a profound effect for F13 and VWA, changing a zero to a non-zero value. Using the above mutation model, the overall likelihood ratio, obtained as the product of the entries in the last column of Table 10, is 0.00026, a very small value which effectively excludes this putative father. However, if the data on F13 had been absent the likelihood ratio would have been 0.347, which is by no means negligible if there is other incriminating evidence in the case. We thus see that a single seeming exclusion may not in reality exclude, when the possibility of mutation is taken into account. When there are no ‘‘seeming exclusions’’ the effect of accounting for mutation will usually be negligible—certainly so for simple cases with no missing individuals—and it can then safely be ignored. Table 10. Paternity case data Marker Mother Child Putative father Likelihood ratio F13 7 16 5 7 7 7 0.000749 VWA 14 17 17 19 16 17 0.00277 D21S11 28 34.2 32.2 34.2 32 32.2 5.013 D1S80 24 24 18 24 18 24 2.391 TH01 6 7 6 7 6 6 2.360 FES 11 11 11 11 10 11 1.358 MBP 1 5 1 5 1 1 1.491 APO-B 45 47 37 47 37 41 1.341 COL2A1 10 14 10 14 8 14 1.629 Overall 0.00026 Excluding F13 0.347 Table 11. Selected allele frequencies for marker VWA Allele: 14 16 17 19 x Frequency: 0.089 0.197 0.298 0.067 0.349 Table 12. Marker VWA: Conditional probability table for capg [resp. camg] given copg [resp. comg] j i 14 16 17 19 x 14 0.997272 0.000391 0.000196 0.0000489 0.00209 16 0.000177 0.998652 0.000354 0.0000884 0.000729 17 0.000058 0.000234 0.999109 0.000117 0.000482 19 0.000065 0.000260 0.000520 0.996377 0.00278 x 0.000533 0.000412 0.000412 0.000533 0.998110 588 A. P. Dawid et al. Scand J Statist 29 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. 4. Inference about identity We now turn to consider a problem of a somewhat different nature. Suppose a DNA trace t belonging to an unknown criminal or from an unidentified body is found. One wishes to know if the trace belongs to any individual in a given pedigree P, on some of whom we have DNA data; or to some other individual in the overall population I. Following Dawid & Mortera (1996), we denote by C the random variable indicating the unknown individual in I leaving the trace. The scene-of-crime or trace DNA evidence is denoted by E : vC ¼ t. Furthermore we let EP: va ¼ n denote the measured DNA evidence for a set a of identified individuals in pedigree P. We suppose that, in the absence of the trace evidence vC, information on v, the DNA profiles for all members of the population I, would be irrelevant to the identity of C. Using the notation for conditional independence of Dawid (1979), this property is expressed as: C ?? v: ð6Þ Given all the evidence ðE; EPÞ, the likelihood Li of the hypothesis that the trace belongs to individual i 2 I is given by: Li / prðva ¼ n; vC ¼ tjC ¼ iÞ ¼ prðva ¼ n; vi ¼ tjC ¼ iÞ / prðvi ¼ tjva ¼ nÞ; ð7Þ by (6). For any individual i 2 I unrelated to P, (7) is just the match probability prðvi ¼ tÞ. For each marker, a single propagation in the probabilistic expert system representing the pedigree P will calculate (7) for all i 2 P simultaneously. These likelihoods can then be multiplied across markers (assuming independence), and combined with prior probabilities, using Bayes’s theorem, to yield the posterior probability that C ¼ i for various i 2 I. 4.1. A murder example To indicate the complexity which can be handled by this approach, we construct a network for a fictional criminal case, taken from Egeland et al. (1997b). A mutilated murdered body has been found, of unknown identity and of male sex. There are a number of individuals who it could be, labelled I.1, II.1, II.2, III.1 and IV.1. There is also a possibility that it is none of these (nor related to any of them), represented by i 2 U. DNA profiles are available from the body, and from living individuals a ¼ fIV.2, IV.3, V.1g. All these individuals are known to be related according to the pedigree P given in Fig. 8. We construct the associated Probabilistic Fig. 8. Murder case: Complex pedigree. Scand J Statist 29 Forensic inference from genetic markers 589 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. Expert System, incorporating mutation, as in Fig. 9. The (fictional) genotype data, for a single marker, are given in Table 13, and the associated allele frequencies in Table 14. In the network we merely have to enter and propagate the DNA profile evidence, EP, observed at the ‘‘data nodes’’ (genotypes for individuals IV.2, IV.3, V.1 in the pedigree), so obtaining the likelihood Li, as given by (7), for each ‘‘suspect’’ individual i 2 P [ U. In particular for any i 2 U the likelihood Li is simply the prior probability prðvi ¼ tÞ, i.e. the (estimated) frequency in the data-base of type t genotype. Note that the network also provides likelihoods for the female individuals I.2 and III.2, as well as for the living individuals IV.2, Fig. 9. Network for murder case. 590 A. P. Dawid et al. Scand J Statist 29 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. IV.3 and V.1, although in all these cases the corresponding prior probabilities prðC ¼ iÞ must be zero. This is not an error, since the likelihood is a measure of the conditional probability of the DNA data observed, under the hypothesis that a specific individual supplied the body; under the assumptions incorporated in the model this conditional probability can be positive, even when the posited hypothesis is known to be false. Of course, in such a case the value of the likelihood is irrelevant: the posterior probability will be zero because the prior is. Table 15 gives the likelihoods Li, both in the absence of mutation and allowing for the possibility of mutation (using a simplistic 9-allele mutation transition matrix Q having qii ¼ 0:96, qij ¼ 0:005, i 6¼ j). The representation and analysis adopted for this problem are somewhat different from those used in section 2 and section 3 above. The same result could in principle have been obtained here by introducing, as there, a ‘‘query node’’, which now would have to be a ‘‘child’’ of all individuals to whom the body could belong. However, that creates a large clique of connected nodes, making such a representation inefficient for computation. Conversely, we could have addressed the problems of section 2 and section 3 by the same logic as used here; however, although this would have provided a straightforward and efficient approach to those problems, we consider that the clarity and simplicity obtained by explicitly representing the hypothesis node in the graph is a strong reason to do so whenever feasible. The above example appeared in the unpublished preliminary version (Egeland et al., 1997b) of the paper Egeland et al. (1997a). These papers describe and illustrate the program PATER (Mostad & Egeland, 1998), based on a routine for organization of pedigree calcula- tions—without, however, our more detailed structuring of the network. In Egeland et al. (1997b) the authors state that PATER had not been able to solve the above problem incorporating mutation, after running for 12 hours on a Sparc-Solaris workstation (their calculations involve consideration of about 1014 different configurations). In contrast, when Table 13. Genotype data, murder case Individual: IV.2 IV.3 V.1 body Genotype: ad ac ab ac Table 14. Allele frequencies, murder case Allele: a b c d x Frequency: 0.01 0.15 0.05 0.35 0.44 Table 15. Likelihood Li for pedigree of Fig. 8 Individual i No Mutation Mutation I.1 0.0084 0.0078 I.2 0.0084 0.0078 II.1 0.0158 0.0199 II.2 0.0084 0.0086 III.1 0.0330 0.0450 III.2 0.0330 0.0362 IV.1 0.4906 0.3859 IV.2 1 1 IV.3 1 1 V.1 0 0 Unrelated 0.001 0.001 Scand J Statist 29 Forensic inference from genetic markers 591 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. formulated as we have done, it can be solved, using the HUGIN software, instantaneously with just one propagation. This underlines the vital importance of constructing an appropriate initial representation of the problem, and using good computational algorithms. Mostad and Egeland have since produced a new program, FAMILIAS (http://www.nr.no/familias), that solves the above problem more efficiently, using a genetic ‘‘peeling algorithm’’ (Elston & Stewart, 1971; Cannings et al., 1978) closely related to the general purpose algorithm used by HUGIN. 5. Conclusions and related problems This article has attempted to demonstrate the value of applying existing general-purpose probabilistic expert system methodology and software to address problems of forensic genetics. Its principal contribution should be seen as comprised in our suggestions for translating such a problem into a suitable PES representation. Although we have largely proceeded by example, certain general principles emerge: in particular, wherever possible one should aim for a fine-grained representation (e.g. incorporating individual genes, not just genotypes), and build the overall structure out of simple repeatable modules, simply connected together. However, the best representation of any specific problem can vary from one problem type to another. In future work we plan to develop intelligent software to provide a suitable mix of standardized and problem-tailored routines. We also propose to address departures from some of the simplifying assumptions we have made here. Thus we have assumed Hardy– Weinberg and linkage equilibrium, i.e., independence within and across markers; and, further, that all founders in a pedigree, including unrepresented individuals, can be regarded as drawn at random from the same homogeneous population. We aim to relax these requirements, e.g. by considering populations composed of partially distinct subpopulations (Roeder et al., 1998; Foreman et al., 1997; Dawid & Pueschel 1999). We have also assumed here that we have fully accurate and relevant figures for gene frequencies and mutation rates. These assumptions will be relaxed by extending the framework to allow Bayesian and other forms of statistical learning of relevant parameters from data. There is a wide range of further identification problems that can be tackled using Probabilistic Expert System representations and computations, either along the lines set out above or through extensions, elaborations or variations of them. We plan to study these with the overall aim of attaching them to a suitable general representational framework. One such problem, readily handled by the methods presented here, arises when a crime suspect’s DNA is not available—perhaps he escaped the country—but DNA evidence is available on some of his relatives. One then needs to compute the likelihood that the crime trace belonged to the fugitive, given the DNA traces of some of his close relatives. This problem can be handled in a similar fashion to paternity cases with missing data, thus avoiding complex algebraic case-by-case computations. Analogous forensic problems arise when someone has been kidnapped, a body part has been sent by the kidnappers, and, given DNA profiles from the kidnapped person’s parents, sibling’s or other kin, one wishes to compute the likelihood that the part belongs to the kidnapped person. The idea of transforming a complex pedigree into a PES also has less gory applications in cases where a prospective immigrant claims relationship with one or more citizens. Another important and complex problem that can easily be formulated and solved using a PES (Mortera, 2002) is the analysis of DNA profiles containing a mixture of genetic material from two or more persons (Evett & Weir, 1998, ch. 7). This is common in rape cases, where a sample may contain biological material from the victim, multiple perpetrators, and one or 592 A. P. Dawid et al. Scand J Statist 29 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. more consensual partners; it can also occur in criminal cases where there has been a scuffle or brawl, for example. Here again the most appropriate PES representation (generally close in spirit to those considered here, albeit a little different in detail) may depend on the specific problem and the identification question that is to be resolved. A somewhat different kind of problem, unfortunately of increasing importance, is that of identification of multiple remains following disasters such as wars, fires, or earthquakes. A celebrated instance was the discovery in Yekaterinburg of human remains thought to be those of the executed last Tsar Nicholas of Russia, his family and servants (Gill et al., 1994). There was a collection of skeletons, of ascertainable sex and (in some cases) approximate age, from which DNA profiles were obtained. A number of possible pedigrees relating the skeletons to each other and to other known individuals (including Prince Philip, Duke of Edinburgh) can be entertained, and a probabilistic network developed to describe each of these, allowing determination of the most likely pedigree given the evidence. The information on sex and age reduces the very large number of possible pedigrees and establishes plausible terminal nodes, i.e. children who cannot have had offspring. The program FAMILIAS can be used to automate the handling of such multiple pedigree problems (Egeland et al., 2000). Further development is envisaged to integrate its features with the approach presented here. The exact computational approach embodied in a PES analysis, like that of its forerunners in algorithms for peeling genetic pedigrees, is widely but not universally applicable. In particular, for even quite small ‘‘loopy’’ pedigrees exact analysis rapidly becomes computa- tionally intractable even with the most sophisticated algorithms. In such cases it is usual to resort to approximate methods, such as Gibbs sampling or similar Monte Carlo Markov chain techniques (Gilks et al., 1996). Methods such as ‘‘blocking Gibbs’’ (Jensen et al., 1995) combine features of both approaches, and hold out promise of extending the range of problems that can be handled by the type of graphical representation we have considered here. Here too, identification of an appropriate graphical representation will be essential for computational feasibility. Table 16. Summary of the notation used in this article Notation Definition pf putative father tf true father m mother m1, m2 mother 1, mother 2 c child c1, c2 child 1, child 2 b brother b1, b2 brother 1, brother 2 gf grandfather gm grandmother mg maternal gene pg paternal gene gt genotype apg actual paternal gene amg actual maternal gene opg original paternal gene omg original maternal gene pfpg, pfmg, pfgt, etc putative father’s paternal gene, maternal gene, genotype, etc. mapg, mopg, etc mother’s actual, original paternal gene, etc. I.x, II.x, etc individual x, x ¼ 1; . . . ; n, of generation I, II, etc. tf=pf? query node: ‘‘true father = putative father?’’ Scand J Statist 29 Forensic inference from genetic markers 593 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. Acknowledgements We wish to thank Hugin Expert A/S for allowing free use of their software. Also, since the graphical interfaces to current PES software are not optimized for the kinds of problem we treat here, we are indebted to Robert Cowell for access to a preliminary version of his program FINEX (Cowell, 2001), which eases the specification task, constructs HUGIN networks, and conducts likelihood analyses. We are grateful to Michael Boehnke, Marina Dobosz, Thore Egeland, and Steffen Lauritzen for valuable discussions and assistance. The comments of an anonymous referee helped to improve the paper. The work was supported by grants from MURST, CNR and ESF (HSSS). References Brinkmann, B., Klintschar, M., Neuhuber, F., Hühne, J. & Rolf, B. (1998). Mutation rate in human microsatellites: influence of the structure and length of the tandem repeat. Amer. J. Human Genet. 62, 1408–1415. Cannings, C., Thompson, E. A. & Skolnick, M. H. (1978). Probability functions on complex pedigrees. Adv. Appl. Probab. 10, 26–61. Cowell, R. G. (2001). FINEX: Forensic identification by network expert systems. Research Report 22, Department of Actuarial Science and Statistics, The City University, London. Cowell, R. G., Dawid, A. P., Lauritzen, S. & Spiegelhalter, D. J. (1999). Probabilistic networks and expert systems. Springer Verlag, New York. Dawid, A. P. (1979). Conditional independence in statistical theory (with discussion). J. Roy. Statist. Soc. Ser. B 41, 1–31. Dawid, A. P. (1992). Applications of a general propagation algorithm for probabilistic expert systems. Statist. Comput. 2, 25–36. Dawid, A. P. (2002). Bayes’s theorem and weighing evidence by juries. In Bayes’s theorem (ed. R. Swinburne). Proc. British Acad. 113, in press. Dawid, A. P. & Mortera, J. (1996). Coherent analysis of forensic identification evidence. J. Roy. Statist Soc. Ser. B 58, 425–443. Dawid, A. P. & Mortera, J. (1998). Forensic identification with imperfect evidence. Biometrika 85, 835– 849. Dawid, A. P. & Pueschel, J. (1999). Hierarchical models for DNA profiling using heterogeneous databases. In Bayesian statistics 6 (eds J. M. Bernardo, J. O. Berger, A. P. Dawid & A. F. M. Smith), pp. 187–212. Oxford University Press, Oxford. Dawid, A. P., Mortera, J. & Pascali, V. L. (2001). Non-fatherhood or mutation? A probabilistic approach to parental exclusion in paternity testing. Forensic Sci. Internat. 124, 55–61. Egeland, T., Mostad, P. F. & Olaisen, B. (1997a). A computerised method for calculating the probability of pedigrees from genetic data. Sci. Justice 37, 269–274. Egeland, T., Mostad, P. F. & Olaisen, B. (1997b). Probability assessments of family relations. Technical Report, Norwegian Computing Center, Oslo, Norway. Egeland, T., Mostad, P. F., Mevåg, B. & Stenersen, M. (2000). Beyond traditional paternity and identification cases: selecting the most probable pedigree. Forensic Sci. Internat. 110, 47–59. Elston, R. C. & Stewart, J. (1971). A general model for the genetic analysis of pedigree data. Human Heredity 21, 523–542. Essen-Möller, E. (1938). Die Beweiskraft der Ähnlichkeit im Vaterschaftsnachweis. Theoretische Grundlagen. Mitt. Anthropol. Ges. 68, 9–53. Evett, I. W. & Weir, B. S. (1998). Interpreting DNA evidence. Sinauer, Sunderland, MA. Foreman, L. A., Smith, A. F. M. & Evett, I. W. (1997). Bayesian analysis of deoxyribonucleic acid profiling data in forensic identification applications (with discussion). J. Roy. Statist. Soc. Ser. A 160, 429–469. Gilks, W. R., Richardson, S. & Spiegelhalter, D. J. (1996). Markov chain Monte Carlo in practice. Chapman & Hall, London. Gill, P. E., Ivanov, P. L., Kimpton, C., Piercy, R., Benson, N., Tully, G., Evett, I., Hagelberg, E. & Sullivan, K. (1994). Identification of the remains of the Romanov family by DNA analysis. Nature Genetics. 12, 417–420. 594 A. P. Dawid et al. Scand J Statist 29 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. Jensen, C. S., Kong, A. & Kjœrulff, U. (1995). Blocking Gibbs sampling in very large probabilistic expert systems. Internat. J. Human–Comput. Stud. 42, 647–666. Mortera, J. (2002). Analysis of DNA mixtures using probabilistic expert systems. In Highly structured stochastic systems (eds P. J. Green, N. L. Hjort & S. Richardson). Oxford University Press, Oxford, in press. Mostad, P. F. & Egeland, T. (1998). Probability assessments of family relations using the program ‘‘pater’’. Technical Report, Norwegian Computing Center, Oslo, Norway. Roeder, K., Escobar, M., Kadane, J. B. & Balazs, I. (1998). Measuring heterogeneity in forensic databases using hierarchical Bayes models. Biometrika 85, 269–287. Spiegelhalter, D. J., Dawid, A. P., Lauritzen, S. L. & Cowell, R. G. (1993). Bayesian analysis in expert systems (with discussion). Statist. Sci. 8, 219–283. Thompson, E. A. (2000). MCMC estimation of multi-locus genome sharing and multipoint gene location scores. Internat. Statist. Rev. 68, 53–73. Valdes, A. M., Slatkin, M. & Freimer, N. B. (1993). Allele frequencies at microsatellite loci: the stepwise mutation model revisited. Genetics 133, 737–749. Weber, J. L. & May, P. E. (1989). Abundant classes of human DNA polymorphism which can be typed by the polymerase chain reaction. Amer. J. Human Genet. 44, 388–397. Received November 2000, in final form October 2001 A. Philip Dawid, Department of Statistical Science, University College London, Gower Street, London WC1E 6BT, UK. E-mail: dawid@stats.ucl.ac.uk Scand J Statist 29 Forensic inference from genetic markers 595 � Board of the Foundation of the Scandinavian Journal of Statistics 2002. 
work_wjbnlg73uzctpn7cjkdpeabf6m	----	Prioritization of human capital measurement indicators using fuzzy AHP Prioritization of human capital measurement indicators using fuzzy AHP F. Tunç Bozbura a, Ahmet Beskese a,*, Cengiz Kahraman b a Bahcesehir University, Department of Industrial Engineering, 34538, Bahcesehir, Istanbul, Turkey b Istanbul Technical University, Department of Industrial Engineering, 34367, Macka, Istanbul, Turkey Abstract People in an organization constitute an important and essential asset which tremendously contributes to development and growth of that company by the help of their collective attitudes, skills and abilities. This is why the human capital (HC) can be considered the most important sub-dimension of the intellectual capital. Since you cannot manage what you cannot control, and you cannot control what you do not measure, the measurement of HC is a very important issue. This study aims at defining a methodology to improve the quality of prioritization of HC measurement indicators under fuzziness. To do so, a methodology based on the extent fuzzy analytic hierarchy pro- cess (AHP) is proposed. Within the model, five main attributes; talent, strategical integration, cultural relevance, knowledge manage- ment, and leadership; their sub-attributes, and 20 indicators are defined. The proposed model can be used for any country. However, the results obtained in the numerical example reflect the situation of HC in Turkey, since the experts are asked to make their evaluations considering the cultural characteristics of Turkey. The results of the study indicate that ‘‘creating results by using knowledge’’, ‘‘employ- ees’ skills index’’, ‘‘sharing and reporting knowledge’’, and ‘‘succession rate of training programs’’ are the four most important measure- ment indicators for the HC in Turkey. � 2006 Elsevier Ltd. All rights reserved. Keywords: Human capital; Measurement indicators; Fuzzy sets; AHP; Prioritization 1. Introduction Many of the differences often exist between the market and book values of companies can be explained by intellec- tual capital (IC) assets not recognized in company balance- sheets (Brennan & Connell, 2000). IC can be thought of as the knowledge-based equity of a company (International Federation of Accountants, 1998). It is the pursuit of effec- tive use of knowledge (the finished product) as opposed to information (the raw material) (Bontis, 1998) and includes assets relating to employee knowledge and expertise, cus- tomer confidence in the company and its products, brands, franchises, information systems, administrative procedures, patents, trademarks and the efficiency of company business processes (Danish Trade and Industry Development Coun- cil, 1997). Intangible assets used to be defined very narrowly, not including assets such as human resources, customer loyalty, company reputation. However, these elements of intellec- tual capital, if managed properly, have huge potential for creating value which many companies feel can no longer be ignored (Brennan & Connell, 2000). Today, IC is widely recognized as the critical source of true and sustainable competitive advantage (Marr, Schiuma, & Neely, 2002). Carlucci, Marr, and Schiuma (2004) shows that the man- agement of IC directly impacts business performance. Knowledge is the basis of IC and is therefore at the heart of organizational capabilities. The need to continuously generate and grow this knowledge base has never been greater (Marr, 2004). Successfully utilizing that knowledge contributes to the progress of society (Seetharaman, Low, & Saravanan, 2004). 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.02.006 * Corresponding author. Tel.: +90 212 669 6523; fax: +90 212 669 4398. E-mail address: beskese@bahcesehir.edu.tr (A. Beskese). www.elsevier.com/locate/eswa Expert Systems with Applications 32 (2007) 1100–1112 Expert Systems with Applications mailto:beskese@bahcesehir.edu.tr Bozbura and Beskese (2005) prefers a three-part defini- tion of intellectual capital that includes human, relational, and organizational components (see Fig. 1). In this figure: 1. The Human Capital (HC) is the individual-level knowl- edge that each employee possesses (Bontis, Keow, & Richardson, 2000). 2. Organizational Capital is the sum of all assets that make the creative ability of the organization possible (Bozb- ura, 2004). 3. The Relational Capital is the sum of all assets that arrange and manage the firms’ relations with the envi- ronment. The relational capital contains the relations with customers, shareholders, suppliers, rivals, the state, governmental institutions and society (Bozbura, 2004). The HC is the most important asset of the intellectual capital, since it is the source of creativity in the organiza- tion. Implicit knowledge assets of the employees in the organization constitute one of the most crucial elements that affect the work performance of the company. How- ever, only the existence of implicit knowledge is not enough for the performance of the organization. The aim is to make the implicit knowledge of the employees an explicit knowledge in every organizational level. In this way, it will be possible to create an organizational value. The human in a company enhances the operational activity of tangible assets (tools and equipments) and activates intangible assets (Fitz-enz, 2001). Increasing the employees’ capabilities has a direct impact on the finan- cial results of the company (Becker, Huselid, & Ulrich, 2001). The selection of IC measurement indicators is a multi- criteria decision problem that requires resolutions involved with various stakeholders’ interests. There has been no basis model for IC statements, nor bottom-line indicators of the value of IC (Han & Han, 2004). In order to assist management decision-making in selecting IC indicators for measurement and disclosure, Han and Han (2004) sug- gest a model that identifies the criteria reflecting decision usefulness and expected risk factors. They proposed an AHP based decision model based on the analysis of the conceptual framework of the qualitative characteristics of financial information and an examination of information quality of the information system. In real word applications, precise data concerning mea- surement indicators of HC are not available or very hard to be extracted. In addition, decision-makers prefer natural language expressions rather than sharp numerical values in assessing HC parameters. So, HC is an inherently fuzzy notion, which can be measured by the synthesis of its con- stituents. Fuzzy logic offers a systematic base in dealing with situations, which are ambiguous or not well defined (Kahraman, Beskese, & Ruan, 2004). Indeed, the uncer- tainty in expressions such as ‘‘low talent’’, ‘‘moderate abil- ity of knowledge creation’’ or ‘‘high experience’’, which are frequently encountered in the HC literature, is fuzziness. In the literature, there is no fuzzy logic method aimed at prioritizing any part of intellectual capital measurement indicators. As a value-added to the literature on the topic, this paper aims at providing practitioners with a fuzzy point of view to the traditional HC analysis methods for dealing quantitatively with imprecision or uncertainty and at obtaining a fuzzy prioritization of HC measurement indicators from this point of view that will close this gap considerably. The paper is organized as follows: Section 2 presents some general knowledge about multi-attribute decision- making techniques used for prioritization. Section 3 includes a summary of the basics of fuzzy sets and num- bers. Section 4 overviews fuzzy AHP literature, and defines the steps of the selected fuzzy AHP method (i.e., Chang’s extent analysis, Chang, 1992, 1996) to be used in the pro- posed model. Section 5 proposes a hierarchical model for the prioritization of HC measurement indicators. Section 6 includes a real-life numerical application in Turkey. Finally, Section 7 presents the conclusion. 2. Multi-attribute decision-making techniques used for prioritization Multi-attribute decision-making (MADM) techniques have the advantage that they can assess a variety of options Fig. 1. Components of Intellectual Capital. F.T. Bozbura et al. / Expert Systems with Applications 32 (2007) 1100–1112 1101 https://isiarticles.com/article/18547 
work_wjlx6pgtzzfgfjn6fwzgsx7pp4	----	Research Article Prototype of the Near-Infrared Spectroscopy Expert System for Particleboard Identification Anna Sandak ,1,2 Jakub Sandak ,1,2,3 Dominika Janiszewska,4 Salim Hiziroglu,5 Marta Petrillo,1 and Paolo Grossi1 1CNR-IVALSA Trees and Timber Institute, 38010 San Michele all’Adige, Italy 2InnoRenew CoE, 6310 Izola, Slovenia 3University of Primorska, Faculty of Mathematics, Natural Sciences and Information Technology, 6000 Koper, Slovenia 4Composite Wood Products Department, Wood Technology Institute, Poznan, Poland 5Natural Resource Ecology and Management, Oklahoma State University, Stillwater, Oklahoma, USA Correspondence should be addressed to Jakub Sandak; jakub.sandak@innorenew.eu Received 30 June 2018; Accepted 13 August 2018; Published 30 September 2018 Academic Editor: Jose S. Camara Copyright © 2018 Anna Sandak et al. 'is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. 'e overall goal of this work was to develop a prototype expert system assisting quality control and traceability of particleboard panels on the production floor. Four different types of particleboards manufactured at the laboratory scale and in industrial plants were evaluated. 'e material differed in terms of panel type, composition, and adhesive system. NIR spectroscopy was employed as a pioneer tool for the development of a two-level expert system suitable for classification and traceability of investigated samples. A portable, commercially available NIR spectrometer was used for nondestructive measurements of particleboard panels. Twenty-five batches of particleboards, each containing at least three independent replicas, was used for the original system development and assessment of its performance. Four alternative chemometric methods (PLS-DA, kNN, SIMCA, and SVM) were used for spectroscopic data classification. 'e models were developed for panel recognition at two levels differing in terms of their generality. In the first stage, four among twenty-four tested combinations resulted in 100% correct classification. Discrimination precision with PLS-DA and SVMC was high (>99%), even without any spectra preprocessing. SNV preprocessed spectra and SVMC algorithm were used at the second stage for panel batch classification. Panels manufactured by two producers were 100% correctly classified, industrial panels produced by different manufacturing plants were classified with 98.9% success, and the experimental panels manufactured in the laboratory were classified with 63.7% success. Implementation of NIR spectroscopy for wood-based product traceability and quality control may have a great impact due to the high versatility of the production and wide range of particleboards utilization. 1. Introduction Particleboard is a panel product manufactured from lig- nocellulosic materials, combined with an adhesive system and bonded together under heat and pressure. Particle- boards are easy to process and flexible in application. 'ey are mainly used for furniture production, but in combi- nation with other materials, they might be used for parquet, insulation materials, sheathing boards for timber framed walls, packaging, or “do-it-yourself products.” 'e major types of particles used to manufacture particleboard include wood shavings, flakes, wafers, chips, and sawdust [1]. A new strategy of EU aims at increasing the share of bioenergy in the EU’s total energy production. It is expected that 20% of the energy produced in Europe after 2020 will be from renewable resources, where 80% of the above will be related to lignocellulosic feedstock. Simultaneously, a sig- nificant increase in the production volumes for other wood- based products (pulp, boards, and furniture) is expected. 'erefore, even if supply and demand of wood on the market is balanced at present, it will most probably not be in equilibrium within the coming years [2]. As a consequence, several alternative resources, mainly agricultural and in- dustrial residues, fast-growing shrubs and plantations, and Hindawi Journal of Spectroscopy Volume 2018, Article ID 6025163, 11 pages https://doi.org/10.1155/2018/6025163 mailto:jakub.sandak@innorenew.eu http://orcid.org/0000-0002-2515-0991 http://orcid.org/0000-0001-9190-677X https://creativecommons.org/licenses/by/4.0/ https://doi.org/10.1155/2018/6025163 postconsumer wood, have received considerable attention in recent years [3]. Effective use of bagasse [4], oil palm waste [5], bark [6] paper sludge, [3], kenaf stalks [7] wheat straw [8], needle litter [9], vine pruning’s [10] waste tissue paper, and corn peel [11] among others was previously reported. Such materials are ecological, functional, and environmental-friendly; therefore, they serve as sustainable raw materials for manufacturing particleboards. 'e type of raw resources used for panels manufacturing, its quality, size of particles, moisture content as well as type and amount of bonding system have significant effect on particleboard properties [12]. 'e most important quality assessment aspects of manufactured panels are emissions, which mainly depends on the type and amount of resin [1]. 'e recent trend to reduce the formaldehyde release from manufac- tured wood products has led to the substitution of urea- formaldehyde (UF) resin with several alternatives [12]. Liquefied wood was previously reported as an interesting replacement for commonly used adhesives [13–15]. 'e influence of liquefied wood on particleboard properties was previously investigated by the authors [16, 17]. However, alternative quality control tools with potential application for online process control are still desired. Portable spectroscopic equipment operating in the NIR range is a highly interesting technology for the wood-based sector [18]. Meder et al. [19] reported successful imple- mentation of FT-NIR spectroscopy for at-line measurement for quality control of melamine-urea-formaldehyde resin in composite wood-panel production. Taylor and Via [20] used visible and near-infrared spectroscopy to quantify phenol formaldehyde resin content in oriented strandboard. Campos et al. [21] used FT-NIR to evaluate composition of agro-based particleboards. Janiszewska et al. [22] reported the application of FT-NIR spectroscopy for clustering raw materials used for liquefaction and their transformation products. Even though several examples of NIR use for product and process quality monitoring have been reported, the majority of them rely on FT-NIR instruments that are relatively costly (both time and investment wise) and with limited applicability for inline measurement. Numerous chemometric techniques for multivariate classification and discrimination were reported as suitable for processing of NIR spectra. 'ese are usually divided into discriminant or class-modelling methods. In dis- criminant analysis, an unknown sample is always assigned and can be allocated only to one of the classes given in the training set. 'e class-modelling approach is more flex- ible. 'e sample can be accepted by more than one class model and therefore be recognized as confused. 'e most common classification methods are linear discriminant analysis (LDA), partial least-squares discriminant analysis (PLS-DA), k-nearest neighbours (kNN), SIMCA, and SVM [23]. LDA is a probabilistic method, which assumes that each sample belonging to a particular class follows a multivariate Gaussian distribution. It requires the explicit calculation of this probability for the formulation of the classification rule. 'e main disadvantage of this method is the fact that it requires a significantly higher number of training samples than the number of variables, the variables themselves cannot be correlated, and all categories have the same within-class scatter, which is hard to assure in many ex- perimental cases [23]. PLS-DA is a multivariate inverse least-squares analysis method used to classify samples. It decomposes the spectra as linear combinations of principal components (PCs) that express the majority of the information (variability) con- tained in the global dataset. 'e predictor variables or latent variables (LVs) are generated from the input variables to maximize the variance between sample classes in the model [24]. KNN is nonparametric and instance-based algorithm. 'is means that it does not make any assumption on the underlying data distribution and does not use the training data to do generalization. 'e kNN makes decision based on the entire training data set. 'e object is classified by a majority vote of neighbours. According to Adeniyi et al. [25], the main advantages of this method are capability of handling training data that are too large to fit in memory, use of simple Euclidean distance to measure the similarities between training and test data, providing a faster and more accurate recommendation as a result of straightforward application. Soft independent modelling of class analogy (SIMCA) is the most common supervised modelling method, repre- senting class-modelling approach. It requires a training data set of samples with a set of attributes and their class membership. In SIMCA, a principal component analysis (PCA) is performed on each class in the data set; thus, a principal component model is used to represent each class. In SIMCA, there is no restriction on the number of mea- surement variables, and few samples per class are enough to run the model. Data classification is a common task in machine learning. 'e support vector machine (SVM) is a very flexible method that makes no assumption regarding data. It is a nonlinear classification method that constructs a set of hyperplanes in a high- or infinite-dimensional space. Good separation is achieved by the hyperplane that has the largest distance to the nearest training data point of any class [24]. It works by obtaining the optimal boundary of two groups in a vector space independent of the probabilistic arrangements of vectors in training set. When the linear boundary in low- dimension input space is not enough to separate two classes, SVM can create a hyperplane that allows linear separation in the higher dimension feature space [26]. SVM has been successfully used for data mining, pattern recognition, and artificial intelligence fields [18]. Practical implementation of discrimination methods in the wood-based industry is a challenging task. Wood is a heterogeneous and anisotropic biological material. 'e correct implementation of NIR spectroscopy for bio-based materials requires a complex approach [27]. A two-level expert system is proposed here for identification of parti- cleboard panels with a portable, commercially available NIR spectrometer. 'e overall goal of this work was to develop a prototype system that might assist quality control and traceability of particleboard panels on the production floor. 2 Journal of Spectroscopy 2. Materials and Methods Four different types of particleboards manufactured in the laboratory and in industrial plants were evaluated. 'e materials differed in terms of panel type (single-layer or three-layer), composition (industrial particles, recycled wood, or alternative lignocellulosic materials including fast- growing species), and the adhesive system (urea- formaldehyde (UF) resin, UF resin modified with lique- fied wood (LW), or UF resin modified with starch). A summary of investigated panels is presented in Table 1. In total, 170 samples representing 25 variants were in- vestigated. A colour image of representative samples surface for each variant is presented in Figure 1. 2.1. Particleboards Manufactured with Recycled Wood (D). Industrial wood particles obtained from a local wood- processing sawmill, containing 25% recycled wood con- tent, were used as a raw material for particleboard pro- duction. 'e particles were sorted using an Allgaier vibration screening machine with screens of mesh diameters 8, 2, 1, and 0.5 mm. Particle fractions ≤8 mm and ≥1 mm were selected for particleboard production. 'e adhesive was prepared as a mixture of an industrial urea-formaldehyde resin and liquefied wood (10–20% rel- ative to the dry weight of the resin). 'e liquefied wood was prepared from four types of wood-processing industry wastes: mixed hardwood-softwood powder (LWP), pine (LP), beech sawdust (LB), and bark (LB) [16]. 'e industrial urea-formaldehyde glue resin character- istics were as follows: gel time of 75 seconds, viscosity 336 mPa·s, total solids content 69.4%, and pH 7.3. Urea- ammonium nitrate solution (46%) was used as a curing agent, constituting 1% of the resin dry mass. Single-layered particleboards of 12 mm thickness with 0, 10, 15, and 20% liquefied wood content in the adhesive resin were produced in the laboratory. 'e nominal density of the panels was 650 kg·m−3. All boards were conditioned after pressing at 20°C and 65% relative humidity. Nine types of single-layer panels were prepared with different adhesives system, as listed in Table 1. Each panel type was produced in two independent batches. Six replicates for each panel were analysed; in total, 108 samples were investigated. 2.2. Particleboards Manufactured with Fast-Growing Wood Species (S). Single- and three-layer particleboard panels were manufactured from Eastern red cedar (Juniperus virginiana L.) using 9% urea-formaldehyde, or a combination of 15% modified corn starch and 2% urea-formaldehyde adhesive. 'ree types of samples were prepared: type A: single-layer board with 9% UF; type B: three-layer board with 9% UF; and type C: three-layer board with a combination of 15% starch and 2% UF. All boards were produced in a laboratory press using a pressure of 5 MPa, at a temperature of 165°C for 5 and 10 minutes in the case of modified corn starch bonded samples [28]. Six independent replicates for each panel were analysed with an NIR spectrometer. 2.3. Particleboards Manufactured with Alternative to Wood Lignocellulosic Plants (G). 'ree-layer particleboards with the core made from different biomasses were produced, including Black locust (Robinia pseudoacacia L.), miscant (Miscanthus sinensis giganteus), willow (Salix viminalis), and rapeseed (Brassica napus). 'e stalks of lignocellulosic materials were reduced in size with a Pallmann’s chipper. Fractions smaller than 10 mm and bigger than 1 mm were used for the core of the panels. 'e chips were used for manufacturing four types of three-layer boards of 16 mm thickness, with the raw density of 680 kg·m−3 [29]. Six in- dependent replicates for each panel were analysed with an NIR spectrometer. 2.4. Industrially Manufactured Particleboards (P). 'ree-layer urea-formaldehyde resin-bonded particle- boards, type P2, suitable for non-load-bearing purposes in dry areas, manufactured by six diverse manufacturing plants of a corporation were used as reference industrial samples. In addition, panels with different thicknesses (38 mm, 28 mm, 18 mm, and 8 mm) prepared by a single producer were investigated. 'ree replicates for each panel type were analysed using NIR spectroscopy. 2.5. Spectroscopic Measurement and Data Mining. 'e MicroNIR 1700 compact sensor produced by Viavi So- lutions (Santa Monica, CA, USA) was used for spectro- scopic measurements. Each spectrum was measured as an average of 10 consecutive scans. 'e scanning fre- quency was 50 Hz, corresponding to 20 ms of integration time. 'e spectral range was from 950 to 1650 nm (10526–6060 cm−1). 128 spectral points were defined for each spectrum and corresponded to pixels of the CCD detector of the instrument. Ten independent measure- ments were done on each panel assuring measurements at different locations over its surface. Each set of 10 spectra was preprocessed with extended multiplicative scatter correction (EMSC) and after that averaged to homogenize the spectral fingerprint of heterogeneous surfaces of panel. 'e MicroNIR instrument has the proven potential for in- field and inline applications due to its rigid construction and integration of all optical, electronic, and mechanical components. PLS_Toolbox 8.0 (Eigenvector Research, Manson, WA, USA) and LabView 13 (National Instruments, Austin TX, USA) were used for data processing and mining. Four discriminant analysis methods: PLS-DA, SVM, KNN, and SIMCA, were used for spectra classification. Models were calculated considering use of the raw spectra and spectra preprocessed with normalization, standard normal variate (SNV), EMSC, and 1st and 2nd derivatives (Savitzky–Golay algorithm, 2nd polynomial order, and 15 smoothing points). 'e order of data was randomized and then divided into calibration (66%) and independent validation (34%) data sets. 'e samples used for model validation were all different from those used for calibration. Journal of Spectroscopy 3 3. Results and Discussion 3.1. General Concept of the Expert System Architecture. Expert systems are innovative process tools that enhance the user’s productivity in accomplishing a task or solving a problem. 'ese vary in at least two important dimensions, including knowledge and technological complexities [30]. Knowledge complexity is mainly determined by the degree of depth and specialization of the internalized knowledge, the scope of the decision, and the level of expertise required to solve the problem. 'e technological complexity includes diversity of hardware and software, the complexity of the user’s environment, the scale of the software design effort, the complexity of required database accesses, and special user interfaces among others [30]. 'e main goal of this work was to develop an expert system capable of detecting and tracing different particleboards. 'e general concept of such a system implemented in the furniture factory is presented in Figure 2. 'e system relies on the spectroscopic measurements acquired during arrival of the new batch to the factory floor with the easy-to-handle and portable instrument. 'e particular characteristics of the instrument investigated within this project allow its direct implementation in the real-life applications. It is recommended to average several NIR spectra collected from the characterized sample in order to minimize the heterogeneity effect of the complex surfaces of particleboards. A spectrum acquired in that way is considered as a fingerprint of the distinct batch and is recorded in the database for further analysis. 'e usage of spectra is twofold. Firstly, it is applied for the further improvement of the chemometric models by providing new case data. Secondly, it is used for future tracking and identification of elements during production as well as in final products. 'e NIR spectrum of an unknown sample is confronted at the first level of the expert system with broad classes repre- senting different particleboard manufacturers or panel types. Table 1: Manufactured panels investigated in this research. Source Sample code Manufacturing Panel type Raw material Adhesive system Project LIDER D1 Laboratory Single-layer Industrial particles containing 25% of recycled wood 0% LW + 100% UF1 D2 Laboratory Single-layer Industrial particles containing 25% of recycled wood 10% LWP +90% UF1 D3 Laboratory Single-layer Industrial particles containing 25% of recycled wood 15% LWP + 85% UF1 D4 Laboratory Single-layer Industrial particles containing 25% of recycled wood 20% LWP + 80% UF1 D5 Laboratory Single-layer Industrial particles containing 25% of recycled wood 20% LP + 80% UF1 D6 Laboratory Single-layer Industrial particles containing 25% of recycled wood 20% LB + 80% UF1 D7 Laboratory Single-layer Industrial particles containing 25% of recycled wood 20% LBK + 80% UF1 D8 Laboratory Single-layer Industrial particles containing 25% of recycled wood 0% LW + 100% UF2 D9 Laboratory Single-layer Industrial particles containing 25% of recycled wood 10% LWP + 90% UF2 OSU S1 Laboratory Single-layer 100% red cedar particles 9% UF S2 Laboratory 'ree-layer 100% red cedar particles 9% UF S3 Laboratory 'ree-layer 100% red cedar particles 15% starch and 2% UF ITD G1 Laboratory 'ree-layer Industrial pine particles (75%) + black locust particles (25%) Core 8% UF, face layers 12% G2 Laboratory 'ree-layer Industrial pine particles (75%) + Miscanthus particles (25%) Core 8% UF, face layers 12% G3 Laboratory 'ree-layer Industrial pine particles (75%) + willow particles (25%) Core 8% UF, face layers 12% G4 Laboratory 'ree-layer Industrial pine particles (75%) + rapeseed particles (25%) Core 8% UF, face layers 12% Chipboard panel producer P1 Industry 'ree-layer 100% industrial particles UF P2 Industry 'ree-layer 100% industrial particles UF P3 Industry 'ree-layer 100% industrial particles UF P4 Industry 'ree-layer 100% industrial particles UF P5 Industry 'ree-layer 100% industrial particles UF P6a Industry 'ree-layer 100% industrial particles UF P6b Industry 'ree-layer 100% industrial particles UF P6c Industry 'ree-layer 100% industrial particles UF P6d Industry 'ree-layer 100% industrial particles UF 4 Journal of Spectroscopy In the case of successful identi�cation, the same spectrum is an input for the second-level chemometric model, where a nity to the speci�c batches recorded previously is assessed. Such a numerical tool might be highly useful for the quality control of panels arriving to the production �oor, as well as for authentication of their origin and composition. �e de- monstrative implementation of the proposed expert system was performed within this research project. �e statistical evaluation of the discriminations success rate, performed at both system levels, is presented in the following section. 3.2. Prototype Expert System for Particleboard Identi�cation. �e concept of the abovementioned expert system imple- mented for demonstration of its feasibility is presented in Figure 3. It consists of �ve chemometric models imple- mented at two discrimination levels. Model #1 is used to screen the unknown spectrum and determine the most probable class corresponding to the particleboard type. Four classes were therefore de�ned at the �rst level of the expert system corresponding to “sample sources” as summarized in the �rst column of Table 1. �e second and more speci�c discrimination is executed afterward by implementing one of the second-level models (#2 to #5). �e number of classes in each of these models corresponds to the number of previously de�ned batches. It has to be mentioned that it is critical to continuously feed the expert system with the most recent spectra of new batches in order to assure constant improvement of the discrimination D1 D2 D3 D4 D5 D6 D7 D8 D9 G1 S1 S2 G2 G3 G4 S3 P1 P2 P3 P4 P5 P6a P6b P6c P6d Figure 1: Appearance of particleboard surfaces representing 25 investigated categories. Journal of Spectroscopy 5 system and proper identi�cation of all historically used particleboard batches. An expert system approach to quality control at the industrial scale was proposed by Paladini [31]. He de- scribed three-step quality control approach that included precontrol activities, inspection, and a decision stage. Liukkonen et al. [32] described the intelligent optimiza- tion and modelling system for electronics production. �ey proposed three modules consisting of appropriate mathematical tools speci�cally tailored to each task: preprocessing, variable selection, and optimization. A data-driven approach was proposed to achieve proactive quality improvement of the production process. Hob- ballah et al. [33] proposed a casual map to show what variables should be included in the design optimization and how the components interact causally. �is approach was successfully used for the preliminary design of an insulating composite mat based on wood �bers. 3.3. Discriminant Analysis. Four alternative classi�cation methods: KNN, PLS-DA, SVM, and SIMCA, were tested for the optimal NIR spectral data classi�cation. All these methods are supervised techniques, as they use prede�ned information about the class membership for all samples selected for model calibration. In that way, the model can be tuned to classify new unknown samples in one of the known classes on the basis of its individual pattern [26]. Discriminating techniques usually build models based on all the categories concerned for the discrimination. �erefore, samples can be classi�ed into one of the pre- de�ned categories, even if actually they do not belong to Production of particleboards (raw resources, adhesives, and technologies) Arrival at the warehouse (registering batches and quality control) Storage of panels Processing of panels Assembly and quality control Use phase (complaints and guarantee management) End of life (recycling, reuse, and circular economy) Database ERP/MRP system Batch #001 Logistics ... Producer #A Producer #B Producer #n NIR fingerprint for batch #A001 #A001 #B401 #C055 #A002 #D016 #M099 Product #1 Product #2 Product #n ... Client X Figure 2: Concept for the traceability of particleboards in the furniture industry. 6 Journal of Spectroscopy any of these [26]. It is possible, however, to de�ne a certain threshold for a nity to any of the classes. In that case, samples not passing the threshold may be classi�ed as unde�ned. Particleboard panel discrimination at the producer level (model #1) di�erentiated classes in terms of manufacturing conditions (laboratory or industrial scale), di�erent re- sources used for panels manufacturing (industrial particles, addition of recycled wood, fast-growing wood species, ag- ricultural wastes, etc.), or di�erent adhesives used for panel production. �e discrimination success scores between di�erent panel providers are summarized in Table 2, where selected spectra preprocessing and diverse discrimination algorithms are presented. Four among twenty-four tested combinations resulted in 100% correct classi�cation. �e optimal combinations were PLS-DA and SIMCA with 2nd derivative spectra, and SVM with both SNV and EMSC preprocessing. Furthermore, discrimination precision with PLS-DA and SVM was high (>99%), even without any spectra preprocessing. �e second level of the expert system was designed to identify speci�c batches. Twenty-�ve lots of particleboards were modelled by four independent chemometric models (#2 to #5), developed for each panel provider separately. �e SVM algorithm classifying SNV preprocessed NIR spectra was implemented. �e summary of the model’s performance is presented in Figure 4, where confusion tables indicate the ability of each model to properly classify (or misclassify) experimented samples from the validation set. All samples of panels manufactured by producer S and G were correctly classi�ed. Industrial panels produced by di�erent manufacturing plants of the producer P were classi�ed with 98.9% success and the panels manufactured by producer D with 63.7% success. It is important to mention that, in the case of the producer D, the same substrate (industrial particles containing 25% of recycled wood) was used for manufacturing all investigated panels. �e only di�erence was the adhesive system, with 0, 10, 15, and 20% share of lique�ed wood addition in the UF resin. 3.4. Implementation of the Prototype Expert System. �e implementation of the expert system in real conditions is a challenging task. Qian et al. [34] proposed a four-step process that consists of the following: (i) Knowledge representation (including all variables characterizing production process) (ii) Database development (including all decision rules, functions, and metadata) (iii) Machine interface design (in�uencing the quality of the expert system for fault diagnosis and the real- time response of the system) (iv) Knowledge maintenance (including update, veri�- cation, and correction of the errors in knowledge base) �e experiences gained when developing chemometric models discriminating diverse particleboards stimulated extension of this research to a prototype expert system ready for testing in real-world applications. �e hardware used for the prototype (portable NIR spectrometer and computer) was similar to the laboratory tests. �e di�erences were in the software, where the original code was customised. It included integration of tools for the control of the sensor setup, data acquisition, spectra processing, and discriminant analysis. �e required chemometric models (#1 to #5) were computed o©ine with PLS_toolbox and were exported as Matlab scripts to LabView. �e algorithm of the prototype expert system for the batch identi�cation is presented in Figure 5. �e software allows acquisition of the new NIR Unknown particleboard sample (at least 10) NIR spectra measurements Spectra homogenization (EMSC + average) Model #1 (panel supplier) Model #2 (project Lider-D) Model #3 (OSU-S) Model #4 (ITD-G) Model #5 (industry-P) D1 D2 D3 D4 D5 D6 D7 D8 D9 S1 S2 S3 G1 G2 G3 G4 P1 P2 P3 P4 P5 P6-a P6-b P6-c P6-d Figure 3: Prototype expert system’s layout for the experimental set of particleboard samples. Journal of Spectroscopy 7 spectra for feeding the database as well as assessment of the unknown spectrum in one of the prede�ned classes. 4. Conclusions �e traceability of particleboards is an important practical issue a�ecting economic performance of furniture producers and related industries. It allows control of material quality and ensures high standards of �nal products. NIR spec- troscopy was employed here as a pioneer tool for devel- opment of the expert system suitable for assisting such a traceability. Twenty-�ve batches of particleboards, each containing at least three independent replicas, were used for the original system development and for its performance assessment. �e hardware and software developed was working properly, and it was implemented as a ready-to-use prototype system. A two-stage expert system seems to be optimal for implementation of the particleboard discrimination as it allows a rough (in the �rst stage) and detailed (in the second stage) classi�cation of particleboard samples to the level of a single batch. In two cases (two producers), 100% correct classi�cation was achieved at the second stage. It has to be mentioned that a drawback of the system presented is the necessity to condition samples before measurement. It includes both refreshing of the surface by gentle sanding and conditioning of samples at well- de�ned climatic conditions. It is due to the fact that NIR spectroscopy is extremely sensitive to the water mole- cules presented in the measured samples. Nevertheless, Predicted as D1 Predicted as D3 Predicted as D5 Predicted as D7 Predicted as D9 0 10 20 30 40 50 60 70 80 90 100 D1 N um be r o f o bs er va tio n (% ) D2 D3 D4 D5 D6 D7 D8 D9 (a) Predicted as S-A Predicted as S-B Predicted as S-C 0 10 20 30 40 50 60 70 80 90 100 S-A S-B S-C (b) Predicted as GM Predicted as GR Predicted as GS Predicted as GW 0 10 20 30 40 50 60 70 80 90 100 GM GR GS GW (c) Predicted as P-1 Predicted as P-3 Predicted as P-5 Predicted as P-6-b Predicted as P-6-d 0 P- 1 P- 2 P- 3 P- 4 P- 5 P- 6- a P- 6- b P- 6- c P- 6- d 10 20 30 40 50 60 70 80 90 100 (d) Figure 4: Confusion table of the particleboard batch prediction with the prototype NIR expert system: project LIDER-D (a), OSU-S (b), ITD-G (c), and industry-P (d). Table 2: �e results of the discrimination success rate at the �rst level of expert system. Discrimination algorithm Spectra preprocessing No preprocessing Normalization SNV EMSC 1st derivative 2nd derivative PLS-DA 99.1 99.1 98.2 98.2 98.2 100 SVM 99.1 63.2 100 100 93.9 63.2 KNN [3] 93.0 96.5 97.4 96.5 97.4 98.2 SIMCA 94.7 99.1 99.1 98.2 99.1 100 8 Journal of Spectroscopy implementation of NIR spectroscopy for product traceability and quality control may impact the industry due to the high versatility of the production and wide range of particleboard use. Data Availability 'e data used to support the findings of this study are available from the corresponding author upon request. Conflicts of Interest 'e authors declare that there are no conflicts of interest regarding the publication of this paper. Acknowledgments 'e authors are grateful to Dorota Fuczek, Grzegorz Kowaluk, and Ekkehard Brommer for providing part of experimental samples. Special acknowledgments are due to COST Action FP1306 (LIGNOVAL) for funding an STSM that contributed to this work. Part of the work was conducted during the BIO4ever (RBSI14Y7Y4) project funded within SIR call by the Italian Ministry of Education, University and Research (MIUR). 'e authors gratefully acknowledge the European Commission for funding the InnoRenew CoE project (Grant Agreement #739574) under the Horizon2020 Widespread-Teaming program. 'e re- search was carried out within the statutory project “Ap- plication of liquefied wood in wood-based materials technology, Part 3. 'e use of liquefied lignocellulosic wastes as a binder for particleboards” (ST-2-BMD/2015/K) and project LIDER/14/0174/L-7/15/NCBR/2016 “New bio- polymer adhesives modified with silanes and ionic liquids for application in wood-based materials technology” funded by the NCBR under the LIDER VII programme. References [1] Compilation of Air Pollutant Emissions Factors, Chapter 10: Wood Products Industry Wood Products Industry, 2002, https://www3.epa.gov/ttn/chief/ap42/ch10/index.html, 2018. [2] 'e Bioenergy System Planners Handbook—BISY PLAN, http://bisyplan.bioenarea.eu/, 2018. Start Measure NIR spectra of unknown sample Compute model #2 Compute model #3 Compute model #4 Compute model #5 Compute model #1 Is producer = D Is producer = S Is producer = G Is producer = P Define batch # Define batch # Define batch # Define batch # The sample is “Undefined” Visualize the result End No No No No Figure 5: 'e algorithm of the prototype expert system. Journal of Spectroscopy 9 https://www3.epa.gov/ttn/chief/ap42/ch10/index.html http://bisyplan.bioenarea.eu/ [3] A. Taramian, K. Dooshoseini, S. A. Mirshokraii, and M. Faezipour, “Particleboard manufacturing: an innovative way to recycle paper sludge,” Waste Management, vol. 27, no. 12, pp. 1739–1746, 2007. [4] J. Fiorelli, R. G. Galo, S. L. Castro Jr., U. L. Belini, P. R. O. Lasso, and H. Savastano Jr., “Multilayer particleboard produced with agroindustrial waste and amazonian vegetable fibres,” Waste and Biomass Valorization, vol. 9, no. 7, pp. 1151–1161, 2017. [5] L. S. Hua, Z. Ashaari, L. W. Chen et al., “Properties of par- ticleboard with oil palm trunk as core layer in comparison to three-layer rubber wood particleboard,” Journal of Oil Palm Research, vol. 27, no. 1, pp. 67–74, 2015. [6] M. C. N. Yemele, P. Blanchet, A. Cloutier, and A. Koubaa, “Effects of bark content and particle geometry on the physical and mechanical properties of particleboard made from black spruce and trembling aspen bark,” Forest Product Journal, vol. 58, no. 11, pp. 48–56, 2008. [7] H. Kalaycioglu and G. Nemli, “Producing composite parti- cleboard from kenaf (Hibiscus cannabinus L.) stalks,” In- dustrial Crops and Products, vol. 24, no. 2, pp. 177–180, 2006. [8] X. Mo, E. Cheng, D. Wang, and X. S. Sun, “Physical properties of medium-density wheat straw particleboard using different adhesives,” Industrial Crops and Products, vol. 18, no. 1, pp. 47–53, 2003. [9] G. Nemli, S. Yildiz, and E. D. Gezer, “'e potential for using the needle litter of Scotch pine (Pinus sylvestris L.) as a raw material for particleboard manufacturing,” Bioresource Technology, vol. 99, no. 14, pp. 6054–6058, 2008. [10] G. A. Ntalos and A. H. Grigoriou, “Characterization and utilisation of vine prunings as a wood substitute for parti- cleboard production,” Industrial Crops and Products, vol. 16, no. 1, pp. 59–68, 2002. [11] P. Lertsutthiwong, S. Khunthon, K. Siralertmukul, K. Noomun, and S. Chandrkrachang, “New insulating par- ticleboards prepared from mixture of solid wastes from tissues paper manufacturing and corn peel,” Bioresource Technology, vol. 99, no. 11, pp. 4841–4845, 2008. [12] N. Čuk, M. Kunaver, and S. Medved, “Properties of particle- boards made by using an adhesive with added liquefied wood,” Materials and technology, vol. 45, no. 3, pp. 241–245, 2011. [13] B. Esteves, J. Martins, J. Martins, L. Cruz-Lopes, J. Vicente, and I. Domingos, “Liquefied wood as a partial substitute of melamine-urea-formaldehyde and urea-formaldehyde resins,” Maderas. Ciencia y tecnoloǵıa, vol. 17, no. 2, pp. 277–284, 2015. [14] A. Ugovsek, M. Kariz, and M. Sernek, “Bonding of wood with adhesive mixtures made of liquefied wood combined with tannin or phenolic resin,” in Future with Forests, R. Ristić, M. Medarević, and Z. Popović, Eds., First Serbian Forestry congress, Beograd, Serbia, 2010. [15] M. Kunaver, S. Medved, N. Čuk, E. Jasiukaitytė, I. Poljanšek, and T. Strnad, “Application of liquefied wood as a new particle board adhesive system,” Bioresource Technology, vol. 101, no. 4, pp. 1361–1368, 2010. [16] D. Janiszewska, I. Frąckowiak, and K. Mytko, “Exploitation of liquefied wood waste for binding recycled wood particle- boards,” Holzforschung, vol. 70, no. 12, pp. 1135–1138, 2016. [17] D. Janiszewska, I. Frąckowiak, and N. Bielejewska, “Appli- cation of selected agents for wood liquefaction and some properties of particleboards produced with the use of liquefied wood,” Drewno, vol. 59, no. 197, pp. 223–230, 2016. [18] J. Sandak, A. Sandak, and R Meder, “Assessing trees, wood and derived products with NIR spectroscopy: hints and tips,” Journal of Near Infrared Spectroscopy, vol. 24, no. 6, pp. 485–505, 2016. [19] R. Meder, W. Stahl, P. Warburton et al., “At-line validation of a process analytical technology approach for quality control of melamine-urea-formaldehyde resin in composite wood-panel production using near infrared spectroscopy,” Analytical and Bioanalytical Chemistry, vol. 409, no. 3, pp. 763–771, 2017. [20] A. Taylor and B. K. Via, “Potential of visible and near infrared spectroscopy to quantify phenol formaldehyde resin content in oriented strandboard,” European Journal of Wood and Wood Products, vol. 67, no. 1, pp. 3–5, 2009. [21] A. C. M. Campos, P. R. G. Hein, R. F. Mendes, L. M. Mendes, and G. Chaix, “Near infrared spectroscopy to evaluate composition of agro-based particleboards,” Bioresources, vol. 4, no. 3, pp. 1058–1069, 2009. [22] D. Janiszewska, A. Sandak, J. Sandak, I. Frąckowiak, and K. Mytko, “Influence of the raw material properties on the liquefied wood chemical composition,” nnals of WULS, Forestry and Wood Technology, vol. 94, pp. 298–303, 2016. [23] M. Bevilacqua, R. Nescatelli, R. Bucci, A. D. Magr̀ı, A. L. Magr̀ı, and F. Marini, “Chemometric classification techniques as a tool for solving problems in analytical chemistry,” Journal of AOAC International, vol. 97, no. 1, pp. 19–28, 2014. [24] Y. Tian, Z. Wang, X. Han, H. Hou, and R. Zheng, “Com- parative investigation of partial least squares discriminant analysis and support vector machines for geological cuttings identification using laser-induced breakdown spectroscopy,” Spectrochimica Acta Part B: Atomic Spectroscopy, vol. 102, pp. 52–57, 2014. [25] D. A. Adeniyi, Z. Wei, and Y. Yongquan, “Automated web usage data mining and recommendation system using K-nearest neighbor (KNN) classification method,” Applied Computing and Informatics, vol. 12, no. 1, pp. 90–108, 2016. [26] L. A. Berrueta, R. M. Alonso-Salces, and K. Héberger, “Su- pervised pattern recognition in food analysis,” Journal of Chromatography A, vol. 1158, no. 1-2, pp. 196–214, 2007. [27] J. Liu, T. J. Pan, and Z. Y. Zhang, “Incremental support vector machine combined with ultraviolet-visible spectroscopy for rapid discriminant analysis of red wine,” Journal of Spec- troscopy, vol. 2018, Article ID 4230681, 5 pages, 2018. [28] J. Sandak, A. Sandak, S. Hiziroglu, and S. Bosak, “Surface characterization of particleboard panels manufactured from eastern redcedar using multi-sensor approach,” in Proceedings of International Panel Products Symposium, pp. 27–37, Llandudno, North Wales, UK, October 2015. [29] I. Frąckowiak, D. Fuczek, and G. Kowaluk, “Impact of dif- ferent lignocellulosic materials used in core of particleboard on modulus of elasticity and bending strength,” Drewno, vol. 51, no. 180, pp. 5–14, 2008. [30] J. H. Bradley and R. D. Hauser, “A framework for expert system implementation,” Expert Systems with Applications, vol. 8, no. 1, pp. 157–167, 1995. [31] E. P. Paladini, “An expert system approach to quality control,” Expert Systems with Applications, vol. 18, no. 2, pp. 133–151, 2000. [32] M. Liukkonen, E. Havia, H. Leinonen, and Y. Hiltunen, “Expert system for analysis of quality in production of electronics,” Expert Systems with Applications, vol. 38, no. 7, pp. 8724–8729, 2011. [33] M. H. Hobballah, A. Ndiaye, F. Michaud, and M. Irle, “Formulating preliminary design optimization problems us- ing expert knowledge: application to wood-based insulating 10 Journal of Spectroscopy materials,” Expert Systems with Applications, vol. 92, pp. 95– 105, 2018. [34] Y. Qian, L. Xu, X. Li, L. Lin, and A. Kraslawski, “LUBRES: an expert system development and implementation for real-time fault diagnosis of a lubricating oil refining process,” Expert Systems with Applications, vol. 35, no. 3, pp. 1252–1266, 2008. Journal of Spectroscopy 11 Tribology Advances in Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 International Journal ofInternational Journal of Photoenergy Hindawi www.hindawi.com Volume 2018 Journal of Chemistry Hindawi www.hindawi.com Volume 2018 Advances in Physical Chemistry Hindawi www.hindawi.com Analytical Methods in Chemistry Journal of Volume 2018 Bioinorganic Chemistry and Applications Hindawi www.hindawi.com Volume 2018 Spectroscopy International Journal of Hindawi www.hindawi.com Volume 2018 Hindawi Publishing Corporation http://www.hindawi.com Volume 2013 Hindawi www.hindawi.com The Scientific World Journal Volume 2018 Medicinal Chemistry International Journal of Hindawi www.hindawi.com Volume 2018 Nanotechnology Hindawi www.hindawi.com Volume 2018 Journal of Applied Chemistry Journal of Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 Biochemistry Research International Hindawi www.hindawi.com Volume 2018 Enzyme Research Hindawi www.hindawi.com Volume 2018 Journal of SpectroscopyAnalytical Chemistry International Journal of Hindawi www.hindawi.com Volume 2018 Materials Journal of Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 BioMed Research International Electrochemistry International Journal of Hindawi www.hindawi.com Volume 2018 N a n o m a te ri a ls Hindawi www.hindawi.com Volume 2018 Journal of Nanomaterials Submit your manuscripts at www.hindawi.com https://www.hindawi.com/journals/at/ https://www.hindawi.com/journals/ijp/ https://www.hindawi.com/journals/jchem/ https://www.hindawi.com/journals/apc/ https://www.hindawi.com/journals/jamc/ https://www.hindawi.com/journals/bca/ https://www.hindawi.com/journals/ijs/ https://www.hindawi.com/journals/tswj/ https://www.hindawi.com/journals/ijmc/ https://www.hindawi.com/journals/jnt/ https://www.hindawi.com/journals/jac/ https://www.hindawi.com/journals/bri/ https://www.hindawi.com/journals/er/ https://www.hindawi.com/journals/jspec/ https://www.hindawi.com/journals/ijac/ https://www.hindawi.com/journals/jma/ https://www.hindawi.com/journals/bmri/ https://www.hindawi.com/journals/ijelc/ https://www.hindawi.com/journals/jnm/ https://www.hindawi.com/ https://www.hindawi.com/ 
work_wkagp4tokffrvohruolrcry5kq	----	Neural network-expert system models of relative-depth perception Behavior Research Methods, Instruments, & Computers 1995,27 (2),173-177 Neural network-expert system models of relative-depth perception STEPHEN J. GOTTS and FREDERICKJ. BREMNER Trinity University, San Antonio, Texas Degree of binocular, horizontal disparity was used by two hybrid neural network/expert system computer models to make relative-depth judgments for pairs of stimulus points. These judgments were then correlated with the actual depth relationships of the points. Results from Simulation 1 showed that horizontal disparity could be computed by the shift in activated cortical hypercolumns evoked by a particular stimulus, and that, in general, multiple disparities could be compared to make accurate judgments about relative depth. However, these results also indicated that stimuli toward the periphery of the visual field were inaccurately perceived as being more distant. Simulation 2 cor- rected for this inaccuracy by appropriately weighting a stimulus point's disparity value as a function of its horizontal position in the visual field. Our ability to perceive and judge depth stems largely from the horizontally differing perspectives of two eyes; each perspective provides a slightly shifted view of the world. Many theorists (e.g., Barlow, Blakemore, & Petti- grew, 1967; Hubel & Wiesel, 1962; Marr, 1982) have been interested in what the process of seeing with two eyes, known as stereopsis, entails and exactly how it al- lows us to perceive depth. STEREOPSIS Two distinct explanations of binocular depth percep- tion evolved in psychophysical circles as early as 1850. Whereas David Brewster advocated a convergence the- ory, Charles Wheatstone supported a disparity theory (Gulick & Lawson, 1976). Brewster argued that the suc- cessive convergence of the optic axes on two points of discrepant distance is the process that we use to perceive depth. Wheatstone, on the other hand, believed that the difference between the images formed on the retinae of the two eyes results in the perception of depth. Although these two arguments seem to force disparity and conver- gence into a state of mutual exclusiveness, it is likely that both concepts function in concert to provide depth cues (Foley, 1980). The optic axes of convergence must cross at some point in the visual field, forming a depth refer- ence area known as the horopter (Ogle, 1964). Relative depths of visual stimuli positioned off the horopter are probably judged by comparing horizontal disparities. Two simulations were designed using a hybrid neural network-expert system approach in order to show how convergence and horizontal disparity might work to- gether to provide cues that are helpful in making relative- Correspondence should be addressed to S. 1. Gotts, 2555 NE Loop 410 #1213, San Antonio, Texas 78217 (e-mail: sgotts@trinity.edu). depth judgments. In both simulations, the optic angles of convergence were held theoretically fixed. Stimuli were then chosen from the set of points located off the hor- opter, making it possible to assess the accuracy ofjudg- ments based solely upon horizontal disparity. SIMULATION 1 While it is clear that single cells in the visual cortex are disparity selective (Barlow et al., 1967), it is still un- clear exactly how this information is used by the brain at later stages of computation. Several theorists (DeAnge- lis, Ohzawa, & Freeman, 1991; Guillemot et al., 1993; Nomura, 1994) have proposed models suggesting uses of the cues provided by these binocular neurons. The hy- pothesis underlying our neural network-expert system model is that the cortex, in order to make relative-depth judgments, performs the equivalent of a difference cal- culation, given the horizontal shift in activated cortical hypercolumns (Hubel & Wiesel, 1962) due to stereopsis. Although this strategy is an obvious departure from the established view, it provides a computationally simple mechanism inspired by the basic topographic arrange- ment of the hypercolumns. Method Apparatus An 8-Mb Compudyne (IBM-compatible) 486DX personal com- puter operating at 33 MHz was used to implement the simulation. The back-propagation neural network was written in Microsoft C/C++ Version 7.0 (Microsoft Corporation, 1991). Training shapes were generated from a diagram of the left and right visual hemifields found in Kelly (1985). Input Layer The input patterns to the network represented the activations of retinal ganglion cells caused by stimulus points in the visual field. 173 Copyright 1995 Psychonomic Society, Inc. 174 GOTTS AND BREMNER Julesz's (1971) work with random-dot stereograms established the legitimacy of using single-point stimuli in studies of depth per- ception. Only points from the binocular, right visual hemifield were used to train the network, as this was sufficient to demon- strate the feasibility of the proposed mechanism. The input layer to the network consisted of a vector 80 units long, the first 40 units representing the ganglia of the nasal hemiretina of the right eye, and the second 40 units representing the ganglia of the temporal hemiretina of the left eye. Each unit of the vector corresponded to a sector of one eye's visual field and symbolized one retinal, X, on-centered ganglion cell. X retinal ganglia seem to be involved in perceiving detail, while the other two main types of retinal ganglia-Y and W-seem to be more involved in detecting mo- tion in the visual field and in providing information necessary to move the eyes (Sterling, 1990). As no motion is involved in the current task, the X pathway is clearly the most germane of the three. Hidden Layer The hidden layer of the neural network consisted of a vector 125 units long. The closest neurological correlate to this layer of neu- rons is the lateral geniculate nucleus (LGN), which relays infor- mation from 'the retina to the cortex while maintaining contralat- eral and ipsilateral pathways (Kelly, 1985). It is important to mention, however, that this analogy is somewhat loosely con- structed. The connections with the hidden layer are not con- strained to maintain the same special lateral relationships: A sin- gle hidden neuron is connected to all the neurons at the input and output layers. Output Layer The output layer of the neural network consisted ofa 2 X 20 ma- trix representing hypercolumns found in Area V 1 of the visual cortex (Kandel, 1985). The two rows ofthe matrix represented the contralateral and ipsilateral dimensions of each hypercolumn, while the 20 columns denoted each of the 20 hypercolumns. Procedure Network training. The back-propagation neural network was allowed to self-organize until it correctly mapped the entire set of 144 input points to corresponding outputs. The network converged as soon as the error for each output unit was less than .10. As there was no danger of activation versus inactivation misclassification (because .8 was the lowest possible value for an active classifica- tion, whereas .2 was the highest possible value for an inactive clas- sification), the .10 error level was assessed as being sufficiently low. Network testing. After the training phase, a set of20 running- fact pairs was presented to the model. Each of the 20 pairs con- sisted of two running facts, representing two stimulus points of discrepant depth. Each of the first 10 running-fact pairs involved points with dissimilar horizontal positions, while the last 10 each involved points with similar horizontal positions. Expert system. For each of the stimulus points, the hyper- column discrepancies were calculated by the formula (discrep- ancy = column number of hypercolumn activated in ipsilateral row - column number ofhypercolurnn activated in contralateral row). The difference values were compared using a series of in- equality formulae, and judgments were made on the basis ofthese formulae. If the first point's disparity was greater than that of the second point, the first point was judged closer (a judgment value of -1). On the other hand, if the second point's disparity was greater than that of the first, the second point was judged closer (a judgment value of + 1). Points eliciting the same disparity values were judged as having equal distance (a judgment value of 0). The system's relative-depth judgments were written to a file, and the judgments were then correlated with the correct depth relation- ships. Figure 1 shows a logical flowchart for the hybrid model and Input Layer Total n =80 Hidden Lpyer 10 20 30 n = 125 0 0u1llut Layer C I Figure 1. Flowchart of the hybrid model. The neural-network component interacts with the expert-system component by supplying it with the hypercolumnar activation information it needs to make its judgments. HYBRID MODELS OF RELATIVE-DEPTH PERCEPTION 175 clarifies the relationship between the neural-network and expert- system components. Note--I indicates that the first point is closer; + I indicates that the second point is closer; 0 indicates that the points have the same depth. Table 1 Judged and Correct Depth Relationship Values for Simulations 1 and 2 Results and Discussion The neural network was able to converge and cor- rectly map all input patterns to their corresponding out- put patterns with error less than .10. The total training time was 12 min; 36 passes of the training facts were nec- essary. The Pearson correlation between relative-depth judgment data and correct depth relationships was found to be significant [r(18) = .48, P < .05]. These results establish that the proposed mechanism of relative-depth perception is indeed feasible. The judgment values and actual depth relationships are provided in Table 1. While the model perfectly discriminated relative depth for points from the last 10 running-fact pairs (which involved comparing points from similar horizontal posi- tions), it had trouble judging points from the first 10 pairs (which involved comparing points from dissimilar horizontal positions). Note, in particular, the lack of cor- respondence for Pairs 2, 5, 6, 8, and 10. The small inac- curacy ofthe model was due to a slight shift of the stim- ulus points for a given hypercolumn disparity toward the observer as the visual hemifield was traversed from left to right (refer to Figure 2). The shift is most pronounced at the far right, or periphery, of the hemifie1d. When a close point near the right border of the binocular hemi- field was compared with a more distant point near the left border, the expert-system component sometimes mistakenly judged the far-left point as the closer of the two presented. However, when presented with points having similar horizontal positions, the system per- SIMULATION 2 An inaccuracy such as the one described in Simula- tion 1 is neither desirable nor necessary. The shift phe- nomenon is the direct result of employing overlapping, polar-coordinate systems, and it can be corrected easily with a simple weighting strategy. The hybrid model in this simulation combines the strategy employed by the first simulation with a function that corrects each stim- ulus point's disparity value on the basis of its horizontal position in the visual field. Method Procedure As in Simulation I, the neural network went through a period of training to learn the relationship between stimulus points in the right visual hemifield and activation of hypercolumns in the left visual cortex. As soon as the network was able to map correctly each of the 144 training points to its corresponding hypercolumn activations with error less than .10 on each output, the network was presented with the set of 20 running-fact pairs, for which the teacher values were not provided. For this simulation, a different formula from the one given in Simulation I was used to calculate the hypercolumn discrepancies (disparity = column number of ac- tivated ipsilateral row - column number of activated contralateral row). The column numbers (from the output matrix) of the acti- vated hypercolumns for each point were left unaltered. However, the ipsilateral column number was weighted according to which horizontal position was involved (the contralateral eye was arbi- trarily selected to observe varying horizontal positions; the ipsi- lateral eye could have been used with equal facility). If a point was located on the right side of the hemifield, its ipsilateral column number was weighted more than if it had been located on the left side. This larger weighting would result in a larger disparity value, effectively compensating for the visual shift. In Simulation I, when a point from the left side of the visual hemifield was pre- sented with an equidistant point from the right side, the point on the right was actually judged as being more distant because its dis- parity value was too small. By increasing its value with the weight- ing strategy, it was moved closer to the visual-shift line for that particular disparity, and the depth judgment was corrected. Refer to Figure 2 for a graphic explanation. The formula for computing any stimulus point's disparity was given by [disparity = ipsilateral column * (I + contralateral column * shift) - contralateral col- umn], where shift = .004611. It is important to note here that this formula, in itself, is trivial; it enabled the authors to establish the formed flawlessly. The shift is clearly present in any sys- tem that consists of two overlapping, polar-coordinate systems, each having a unique origin. The human brain obviously employs such a system. However, whether the human brain compensates for this shift is unclear. Network Design To determine whether compensating for the shift was improv- ing the system's relative-depth judgments, the design, training facts, and running facts from Simulation I were used. Apparatus The machine used in the current simulation was the same as that used in Simulation I. Additionally, the C computer code was the same as that written for Simulation I, with the exception of the dif- ference calculation; the exact calculation is described later (see Procedure). -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I Judgments -I -I +1 -I -I -I -I -I +1 -I o -I -I -I o -I -I -I o -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I -I Model I Model 2 Correct Relationship I 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Running-Fact No. 176 GOTTS AND BREMNER Line alongwhich discrepancy values are the sarne Temporal Hemiretina Line along which all points have equal depth Figure 2. A diagram of the line along which the model in Simulation 1 perceives points as having the same depth. Note the shift of the line toward the observer at the right edge of the visual hemifield. The correction. made by the model in Simula- tion 2, allowsfor more accurate relative-depth judgments. feasibility of our particular disparity computation. The shift prob- lem addressed by the new formula is one of employing stereopsis as a mechanism for relative-depth perception, not just a problem of our particular computation. Any system that accurately judges relative depth must somehow solve the shift problem. Once computed, the new disparity values were compared using the inequality formulas from Simulation I to judge relative depth for each pair of stimulus points. The system tested to see whether the first point's disparity was greater than that of the second point. If so, the first point was judged closer; otherwise, the second point was judged closer. There was no real chance that the system would judge the points as equidistant, on account of the acuity of the weighting. It would only judge two points as having equal distance ifthey were the same point. The system's relative-depthjudgments were written to a file and then correlated with the correct depth relationships. Results and Discussion The neural network implemented in Simulation 2 was able to converge and correctly map all input patterns to their corresponding output patterns with error less than .10. The total training time for this version of the net- work was 15 min; 40 passes of the training facts were necessary. These varied slightly from the training re- sults found in Simulation I, simply because the pseudo- random number algorithm generated a different set of values every time the program was executed; the algo- rithm used the computer's clock time as a unique seed value. Different random numbers were assigned as hid- den weight values in this simulation, which resulted in a slightly different training time. The Pearson correlation between the system's relative- depth judgments and the correct depth relationships was found to be highly significant [r(18) = 1.00, p < .01]. The system was 100% accurate at making relative-depth judgments (an apparent improvement on the accuracy of Simulation I). The data from the current simulation (see Table I) establish a successful modeling of the percep- tion of relative depth. Furthermore, the data establish that our hypothesis concerning how disparity is computed is entirely feasible. GENERAL DISCUSSION Use of a hybrid neural network/expert system ap- proach in this particular problem is valuable because it allows us to model both the visual-tract physiology and the higher cognitive processing of depth cues in the most attractive way possible. While not clear for single-point stimuli, the benefit of the neural network component will be fully realized when more complex stimuli are in- volved. Future iterations ofthe model will be able to ren- der disparity values easily, while a more rule-based ap- proach would be extremely tedious to implement and less physiologically valid. The rule-based approach of the expert-system component, however, makes quick and explicit decisions on the basis of the information fed to it by the neural network. These decisions are made in a way that more closely mimics the higher cognitive pro- cesses of the human brain. In this particular setting, the hybridization ofthe neural-network and expert-systems approaches yields a useful model demonstrating that a hypercolumnar-disparity computation is indeed feasible. REFERENCES BARLOW, H. B., BLAKEMORE, c, & PETTIGREW, J. D. (1967). The neural mechanism of binocular depth discrimination. Journal of Physiology, 193,327-342. HYBRID MODELS OF RELATIVE-DEPTH PERCEPTION 177 DEANGELIS, G. C, OHZAWA, 1.,& FREEMAN, R. D. (1991). Depth is en- coded in the visual cortex by a specialized receptive field structure. Nature, 352,156-159. FOLEY, J. M. (1980). Binocular distance perception. Psychological Re- view, 87, 411-434. GUILLEMOT, J.-P., PARADIS, M.-C., SAMSON, A., PTITO, M., RICHER, L., & LEPORE, F. (1993). Binocular interaction and disparity coding in area 19 of visual cortex in normal and split-chiasm cats. Experi- mental Brain Research, 94, 405-417. GULICK, W. L., & LAWSON, R. B. (1976). Human stereopsis: A psy- chophysical analysis. New York: Oxford University Press. HUBEL, D. H., & WIESEL, T. N. (1962). Receptive fields, binocular in- teraction and functional architecture in the eat's visual cortex. Jour- nal ofPhysiology, 160, 106-154. JULESZ, B. (1971). Foundations of cyclopean perception. Chicago: University of Chicago Press. KANDEL, E. R. (1985). Processing ofform and movement in the visual system. In E. R. Kandel & 1. H. Schwartz (Eds.), Principles of neural science (2nd ed., pp. 366-383). New York: Elsevier. KELLY, J. P. (1985). Anatomy of the central visual pathways. In E. R. Kandel & J. H. Schwartz (Eds.), Principles of neural science (2nd ed., pp. 356-365). New York: Elsevier. MARR, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. San Francisco: W. H. Freeman. MICROSOFT CORPORATION (1991). Microsoft C/C++ (Version 7.0) [Computer program]. Redmond, WA: Author. NOMURA, M. (1994). A model for neural representation of binocular disparity in striate cortex: Distributed representation and veto mechanism. Biological Cybernetics, 69, 165-171. OGLE, K. N. (1964). Researches in binocular vision. New York: Hafner. STERLING, P. (1990). Retina. In G. M. Shepard (Ed.), The synaptic or- ganization of the brain (3rd ed., pp. 170-213). New York: Oxford University Press. (Manuscript received November 18,1994; revision accepted for publication January 20, 1995.) 
work_wkfnnwuk6zhkxiy32dhnltvsfq	----	Mathematical Programming for Piecewise Linear Regression Analysis Lingjian Yanga, Songsong Liua, Sophia Tsokab, Lazaros G. Papageorgioua,∗ aCentre for Process Systems Engineering, Department of Chemical Engineering, University College London, Torrington Place, London WC1E 7JE, UK bDepartment of Informatics, School of Natural and Mahtematical Sciences, King’s College London, Strand, London WC2R 2LS, UK Abstract In data mining, regression analysis is a computational tool that predicts con- tinuous output variables from a number of independent input variables, by ap- proximating their complex inner relationship. A large number of methods have been successfully proposed, based on various methodologies, including linear regression, support vector regression, neural network, piece-wise regression etc. In terms of piece-wise regression, the existing methods in literature are usu- ally restricted to problems of very small scale, due to their inherent non-linear nature. In this work, a more efficient piece-wise regression method is intro- duced based on a novel integer linear programming formulation. The proposed method partitions one input variable into multiple mutually exclusive segments, and fits one multivariate linear regression function per segment to minimise the total absolute error. Assuming both the single partition feature and the number of regions are known, the mixed integer linear model is proposed to simultane- ously determine the locations of multiple break-points and regression coefficients for each segment. Furthermore, an efficient heuristic procedure is presented to identify the key partition feature and final number of break-points. 7 real world problems covering several application domains have been used to demon- ∗Corresponding author: Tel.: +442076792563; Fax.: +442073882348 Email addresses: lingjian.yang.10@ucl.ac.uk (Lingjian Yang), s.liu@ucl.ac.uk (Songsong Liu), sophia.tsoka@kcl.ac.uk (Sophia Tsoka), l.papageorgiou@ucl.ac.uk (Lazaros G. Papageorgiou) Preprint submitted to Journal of Expert Systems with Applications August 12, 2015 strate the efficiency of our proposed method. It is shown that our proposed piece-wise regression method can be solved to global optimality for datasets of thousands samples, which also consistently achieves higher prediction accuracy than a number of state-of-the-art regression methods. Another advantage of the proposed method is that the learned model can be conveniently expressed as a small number of if-then rules that are easily interpretable. Overall, this work proposes an efficient rule-based multivariate regression method based on piece-wise functions and achieves better prediction performance than state-of- the-arts approaches. This novel method can benefit expert systems in various applications by automatically acquiring knowledge from databases to improve the quality of knowledge base. Keywords: regression analysis, surrogate model, piecewise linear function, mathematical programming, optimisation 1. Introduction In data mining, regression is a type of analysis that predicts continuous out- put/response variables from several independent input variables. Given a num- ber of samples, each one of which is characterised by certain input and output variables, regression analysis aims to approximate their functional relationship.5 The estimated functional relationship can then be used to predict the level of output variable for new enquiry samples. Generally, regression analysis can be useful under two circumstances: 1) when the process of interest is a black-box, i.e. there is limited knowledge of the underlying mechanism of the system. In this case, regression analysis can accurately predict the output variables from10 the relevant input variables without requiring details of the however compli- cated inner mechanism (Bai et al., 2014; Venkatesh et al., 2014; Cortez et al., 2009; Davis & Ierapetritou, 2008). Quite frequently, the user would also like to gain some valuable insights into the true underlying functional relationship, which means the interpretability of a regression method is also of importance, 2)15 when the detailed simulation model relating input variables to output variables, 2 usually via some other intermediate variables, is known, yet is too complex and expensive to be evaluated comprehensively in feasible computational time. In this case, regression analysis is capable of approximating the overall system be- haviour with much simpler functions while preserving a desired level of accuracy,20 and can then be more cheaply evaluated (Caballero & Grossmann, 2008; Henao & Maravelias, 2011, 2010; Viana et al., 2014; Beck et al., 2012). Over the past years, regression analysis has been established as a powerful tool in a wide range of applications, including: customer demand forecasting (Levis25 & Papageorgiou, 2005; Kone & Karwan, 2011), investigation of CO2 capture process (Zhang & Sahinidis, 2013; Nuchitprasittichai & Cremaschi, 2013), opti- misation of moving bed chromatography (Li et al., 2014b), forecasting of CO2 emission (Pan et al., 2014), prediction of acidity constants for aromatic acids (Ghasemi et al., 2007), prediction of induction of apoptosis by different chemical30 components (Afantitis et al., 2006) and estimation of thermodynamic property of ionic liquids (Chen et al., 2014; Wu et al., 2014). A large number of regression analysis methodologies exist in the literature, including: linear regression, support vector regression (SVR), kriging, radial35 basis function (RBF) (Sarimveis et al., 2004), multivariate adaptive regression splines (MARS), multilayer perceptron (MLP), random forest, K-nearest neigh- bour (KNN) and piecewise regressions. We briefly summarise those regression methodologies before presenting our proposed method. 40 Linear regression Linear regression is one of the most classic types of regression analysis, which predicts the output variables as linear combinations of the input variables. The regression coefficients of the input variables are usually estimated using least squared error or least absolute error approaches, and the problems can45 be formulated as either quadratic programming or linear programming prob- lems, which can be solved efficiently. In some cases when the estimated linear 3 relationship fails to adequately describe the data, a variant of linear regres- sion analysis, called polynomial regression, can be adopted to accommodate non-linearity (Khuri & Mukhopadhyay, 2010). In polynomial regression, higher50 degree polynomials of the original independent input variables are added as new input variables into the regression function, before estimating the coefficients of the aggregated regression function. Polynomial functions of second-degree have been most frequently used in literature due to its robust performance and computational efficiency (Khayet et al., 2008; Minjares-Fuentes et al., 2014).55 Another popular variant of linear regression is called least absolute shrinkage and selection operator (LASSO) (Tibshirani, 1994). In LASSO, summation of absolute values of regression coefficients is added as a penalty term into the objective function. The nature of LASSO encourages some coefficients to equal60 to 0, thus performing implicit feature selection (Tibshirani, 2011). Automated learning of algebraic models for optimisation (ALAMO) (Cozad et al., 2014; Zhang & Sahinidis, 2013) is a mathematical programming-based regression method that proposes low-complexity functions to predict output65 variables. Given the independent input features, ALAMO starts with defining a large set of potential basis functions, such as polynomial, multinomial, expo- nential and logarithmic forms of the original input variables. Subsequently an mixed integer linear programming model (MILP) is solved to select the best subset of T basis functions that optimally fit the data. The value of T is ini-70 tially set equal to 1 and then iteratively increased until the Akaike information criterion, which measures the generalisation of the constructed model, starts to decrease (Miller et al., 2014). The integer programming model is capable of capturing the synthetic effect of different basis functions, which is considered more efficient than traditional step-wise feature selection.75 SVR Support vector machine is a very established statistical learning algorithm, 4 which fits a hyper plane to the data in hand (Smola & Schlkopf, 2004). SVR minimises two terms in the objective function, one of which is �-insensitive loss80 function, i.e. only sample training error greater than an user-specific threshold, �, is considered in the loss function. The other term is model complexity, which is expressed as sum of squared regression coefficients. Controlling model com- plexity usually ensures the model generalisation, i.e. high prediction accuracy in testing samples. Another user-specified trade-off parameter balances the sig-85 nificance of the two terms (Chang & Lin, 2011; Bermolen & Rossi, 2009). One of the most important features that contribute to the competitiveness of SVR is the kernel trick. Kernel trick maps the dataset from the original space to higher-dimensional inner product space, at where a linear regression is equiva- lent to an non-linear regression function in the original space (Li et al., 2000).90 A number of kernel functions can be employed, e.g. polynomial function, radial basis function and fourier series (Levis & Papageorgiou, 2005). Formulated as a convex quadratic programming problem, SVR can be solved to global optimality. Despite the simplicity and optimality of SVR, the problem of tuning two param-95 eters, i.e. training error tolerance � and trade-off parameter balancing model complexity and accuracy, and selection of suitable kernels still considerably af- fect its performance accuracy (Lu et al., 2009; Cherkassky & Ma, 2004). Kriging100 Kriging is a spatial interpolation-based regression analysis methodology (Klei- jnen & Beers, 2004). Given a query sample, kriging estimates its output as a weighted sum of the outputs of the known nearby samples. The weights of samples are computed solely from the data by considering sample closeness and redundancy, instead of being given by an arbitrary decreasing function of105 distance (Kleijnen, 2009). The interpolation nature of kriging means that the derived interpolant passes through the given training data points, i.e. the er- ror between predicted output and real output is zero for all training samples. Different variants of kriging have been developed in literature, including the 5 most popular ordinary kriging (Lloyd & Atkinson, 2002; Zhu & Lin, 2010) and110 universal kriging (Brus & Heuvelink, 2007; Sampson et al., 2013). MARS MARS (Friedman, 1991) is another type of regression analysis that accommo- dates non-linearity and interaction between independent input variables in its115 functional relationship. Non-linearity is introduced into MARS in the form of the so-called hinge functions, which are expressions with max operators and look like max(0,X − const). If independent variable X is greater than a constant number const, the hinge function is equal to X-const, otherwise the hinge func- tion equals to 0. The hinge functions create knots in the prediction surface of120 MARS. The functional form of MARS can be a weighted sum of constant, hinge functions and products of multiple hinge functions, which makes it suitable to model a wide range of non-linearity (Andrs et al., 2011). The building of MARS usually consists of two steps, a forward addition and125 a backward deletion step. In the forward addition step, MARS starts from one single intercept term/constant and iteratively adds pairs of hinge functions (i.e. max(0,X − const) and max(0,const − X)) that leads to largest reduction in training error. Afterwards, a backward deletion step, which removes one by one those hinge functions contributing insignificantly to the model accuracy, is130 employed to improve generalisation of the final model (Leathwick et al., 2006; Balshi et al., 2009). The presence of hinge functions also make MARS a piece- wise regression method. MLP135 Multilayer perceptron is a feedforward artificial neural network, whose structure is inspired by the organisations of biological neural networks (Hill et al., 1994). A MLP typically consists of an input layer of measurable features, an output layer of response variables, sandwiching multiple intermediate layers of neurons. The network is fully interconnected in the sense that neurons in each layer are140 6 connected to all the neurons in the two neighbour layers (Comrie, 1997; Gevrey et al., 2003). Each neuron in the intermediate layers takes a weighted linear combination of outputs from all neurons in the previous layer as input, applies an non-linear transformation function before supplying the output to all neu- rons of the next layer. The use of non-linear transformation functions, including145 sigmoid, hyperbolic tangent and logarithmic functions, makes MLP suitable for modelling highly non-linear relationship (Gevrey et al., 2003; Rafiq et al., 2001). Identifying the optimal configuration of a MLP, i.e. the number of interme- diate layers, the number of neurons for each intermediate layer, the type of150 activation function for each neuron and the weights of connection between con- secutive layers of neurons, is known to be time-consuming and traps in local optimal solutions (Paliwal & Kumar, 2009). The large degree of freedom in training a MLP is often blamed for data over-fitting. Dua (2010) has presented a two-objective mathematical formulation trying to find the best configuration155 of a MLP by balancing the training accuracy and network complexity. More often, architecture of a MLP is fixed by the user and back-propagation is used to tune only the weights of connection between neighbour layers of neurons (Gud- ise & Venayagamoorthy, 2003; Zhang et al., 2007). 160 Random forest Before introducing random forest we first describe regression tree, which is a decision tree-based prediction model. Starting from the entire set of samples, a regression tree selects one independent input variable among all and performs binary split into two child sets, under the condition that the two child nodes165 give increased purity of the data compared with its single parent node. Purity is often defined as the deviation of predicting with the mean value of the out- put variable. The process of binary split is recursively applied for each child node until a terminating criterion is satisfied. The nodes that are not further partitioned are called leaves. After growing a large tree, a pruning process is170 employed to remove the leaves contributing insignificantly to the purity im- 7 provement (Breiman et al., 1984; Loh, 2011). In order to improve model fit, a linear regression model can be fitted for each leaf (Quinlan, 1992). Random forest is an ensemble learning method of regression trees. In general,175 random forest (Breiman, 2001; Biau, 2012) builds a forest of multiple regression tree models and aggregate the decisions from all the trees to produce a final prediction. Given a dataset, multiple bootstrap sample sets are first created by random sampling with replacement. Each of the bootstrap sample set is then learned by a revised regression tree algorithm, which differs from the classic180 regression tree by randomly selecting a candidate subset of features for each binary split of node (Genuer et al., 2010). The accuracy of each regression tree can be estimated on the training samples absent from the bootstrap set, and the final prediction can be either simple average of predictions from all trees or weighted average considering the estimated accuracy. It is demonstrated that185 random forest achieves much robust prediction performance compared with sin- gle regression tree method (Breiman, 2001; Fanelli et al., 2011). KNN KNN belongs to the category of lazy learning algorithms, due to the fact that190 prediction is based on the instances without an explicit training phase of con- structing models, thus making it one of the simplest regression methods in literature (Korhonen & Kangas, 1997). Given an enquiry sample, KNN first identifies K closest instances in the training sample set, the exact value of K is given a priori. The closeness of samples can be measured by different distance195 metrics, for example Euclidean and Manhattan distances (Scheuber, 2010; Ero- nen & Klapuri, 2010). Prediction is then taken as weighted mean of the outputs of the K nearest neighbours, with weight often being defined as the inverse of distance (Papadopoulos et al., 2011). Despite its simplicity, KNN usually provides competitive prediction performance against much more sophisticated200 algorithms. 8 Previous work on piecewise regression Piecewise functions have been frequently studied in literature as well. In (Toms & Lesperance, 2003), univariate piece-wise linear functions have been used to205 fit ecological data and identify break-points that represent critical threshold values of a phenomenon. In (Strikholm, 2006), a method based on statistical testing is proposed to estimate the number of break-points for an univariate piece-wise linear function. Malash & El-Khaiary (2010) also apply piece-wise linear regression techniques on univariate experimental adsorption data. Piece-210 wise function is determined by solving a non-linear programming model. Seg- Reg (www.waterlog.info/segreg.htm) is free software that permits estimating of piece-wise regression functions with up to two independent variables. For one independent variable, SegReg splits from a series of candidate break-points and for each one fits a linear regression for either side of the break-point. The215 break-point corresponding to the largest statistical confidence is taken as the final solution. In the case of two independent variables, SegReg first determines the two-region piece-wise regression function between the dependent variable and the most significant input variable, before computing the relation between its residual/deviation and the second input variable.220 Both Magnani & Boyd (2009) and Toriello & Vielma (2012) publish work on data fitting with a special family of piece-wise regression functions, called max- affine functions. The form of max-affine functions is defined as the maximum of a series of linear functions, i.e. a sample is projected to all linear functions,225 and the maximum projected value among all is taken as final predicted value from the piece-wise functions. The use of max-affine functions limits the fitted surface to be convex. In (Magnani & Boyd, 2009) a heuristic method is used to ease the difficulty of direct solving the highly non-linear max-affine functions, while in (Toriello & Vielma, 2012), big-M constraint is used to reformulate230 the problem into an non-convex mixed integer non-linear programming model. However, computational complexity is limiting their applications to examples of small scale. 9 More recently, Greene et al. (2015) applies piece-wise regression analysis to235 predict patient’s post-treatment life quality with the pre-treatment life quality measure, which identifies the segments where therapy benefits vary significantly. The analysis is performed using Segmented (Muggeo, 2008), a package written in R (R Development Core Team, 2008). Segmented formulates the problem using a non-linear model and requires a user to specify the segmented input240 variables, the number of break-points and also the initial guess of each break- point. Starting from the those user supplied initial positions of break-points, Segmented iteratively moves around the neighbour of the initial guess points to search break-points of better quality using local linearisation. However, it is difficult if not impossible to reasonably guess good starting points for real world245 multivariate problems of large number of samples and input variables, where visual examination cannot be performed. This makes it hard to identify quality solutions. Furthermore, Segmented only allows the input variables being par- titioned to have different regression coefficients across different segments, while the other input variables keep the same coefficients within the entire ranges,250 significantly restricting its flexibility. In both (Xue et al., 2013) and (Li et al., 2014a), piece-wise regression func- tion were employed to detect vegetation changes. Piece-wise linear regression was tackled using fuzzy logic and identifies the changes in patterns of vegetation255 greenness. Cavanaugh et al. (2014) employ piece-wise regression and find out that the changes in mangrove area over the last 20 years is a piece-wise functions of latitude, with regions above and below a specific threshold latitude value fol- lowing two different patterns of mangrove grows. Moreover, Matthews et al. (2014) uses 2-segment piece-wise functions to describe the relationship between260 species richness and fragment area of islands, with the critical breakpoint being determined by simply sampling a number of candidate values and selecting the one giving best model fit. Unfortunately, the above methods are all limited to model rather simple relationship between one output variable and one input 10 variable, seriously limiting their usage in more complex problems.265 It is clear that the previous literature work of piece-wise regression methods are non-linear and computationally restricted to problems of very small scales. Yet, they cannot be solved to identify globally optimal solutions. In this work, we propose a novel linear model for piece-wise regression analysis. A single270 input variable is partitioned to separate samples into multiple mutually exclu- sive segments, while each segment is fitted with a unique multivariate linear regression function. Assuming that both the partition variable and the number of break-points are known, we propose an optimisation model that optimally estimates the position of all break-points and the linear regression coefficients275 for each segment simultaneously so that the total absolute deviation is min- imised. Thanks to the usage of binary variables, our proposed mathematical model is linear and can be efficiently solved to global optimality for problems up to thousands samples (see Results and Discussion section). Furthermore, a solution procedure is used to identify the key partition variable and the final280 number of break-points. Several real world multivariate benchmark datasets have been used to demonstrate the applicability and efficiency of the proposed method. The proposed piece-wise regression method can help construct expert systems in285 various application domains. Expert systems are computer programs designed to make decisions analogous to human experts. As an expert system is typ- ically made up of an inference engine and a knowledge base, the quality and quantity of information in knowledge base directly affects the usefulness of the constructed expert system. Our proposed piece-wise regression method can be290 helpful in more efficiently building expert systems via automatic and efficient acquisition of knowledge. More specifically, the proposed piece-wise regression method can extract latent knowledge from the large collection of domain expert curated databases. Those discovered knowledge are represented in the form of identified relationship between input and output variables of interest, which295 11 can be combined with expert knowledge to form the final expert system (Alonso et al., 2012). For example, the proposed piece-wise regression method in this work can be used for building prognostic expert systems in medical applica- tions. When presented historical data of patients’ clinical variables and survival length, piece-wise regression can induce domain knowledge by approximating300 the complex relationship between clinical variables and survival length. Those induced knowledge can then be used to perform prognosis for the current pa- tients, imitating the end-behaviour of human experts, i.e. medical doctors. Overall, the key contributions of our work are illustrated below:305 • We propose a novel mixed integer optimisation model for multivariate regression analysis modelling piece-wise linear functions, which partitions a single variable into multiple mutually exclusive regions and fits each one with a distinct multivariate linear function. Given as prior a single input variable as partition feature and the number of segments, the optimisation310 model can be solved to simultaneously determine the positions of multiple break points and regression coefficients for each segment. • Given that neither which feature should be segmented nor the number of segments are typically known, a heuristic solution procedure is also introduced that automatically identifies the key partition variable and the315 final number of segments. • A number of real world benchmark problems have been employed to demonstrate the applicability and efficiency of the proposed method. As sharp difference to the existing piece-wise regression methods in literature which can only be applied to problems of very small size, our proposed320 optimisation model can be solved to global optimality for datasets contain- ing up to five thousand samples. Comparison to some popular regression methods based on other methodologies clearly indicates that our proposed method based on piece-wise function achieves the highest prediction ac- curacy, and does it consistently. Besides high prediction performance,325 12 our proposed regression method has the advantage of being easily under- standable and interpretable, as the learned model can be conveniently represented as a small set of rules. • As a generic data mining method, our proposed regression method can help with constructions of expert and intelligent systems via automatic330 extraction of knowledge from database. We have discussed its potential usage in various application domains. The rest of the paper is structured as follows: in Section 2, we present the mathematical programming model and a heuristic solution procedure. In Section 3, comparative results of our proposed method and some state-of-the-335 art regression algorithms on benchmark examples are presented and discussed. The last section concludes with our key findings. 2. Method A novel piecewise linear regression method is proposed in this work. The core idea of the proposed method is to identify a single input feature, and separate340 the samples into complementary regions on this feature. One different linear regression function is fitted locally for each region. The sample partition and calculation of local regression coefficients are performed simultaneously within the proposed optimisation to achieve least absolute error. 2.1. A novel regression method345 In this section, we first describe a novel mathematical programming model that optimises the location of break-points and regression coefficients for each region so as to achieve minimal training error. Subsequently, a solution proce- dure is proposed to identify the best partition feature and the number of regions. The indices, parameters and variables associated with the proposed model are 13 listed below: Indices s sample, s=1,2,...,S m feature/independent input variable, m=1,2,...,M r region, r=1,2,...,R m* the feature where sample partition takes place Parameters Asm numeric value of sample s on feature m Ys output value of sample s U′,U′′ arbitrarily large positive numbers Continuous variables Wrm regression coefficient for feature m in region r Br intercept of regression function in region r Predrs predicted output for sample s in region r Xr m* break −point r on partition feature m* Ds training error between predicted output and real output for sample s Binary variables Frs 1 if sample s falls into region r; 0 otherwise Assume first that both the partition feature m* and the number of regions R are given, the R-1 break points are arranged in an ordered way: Xr−1m ≤ X r m ∀m = m *,r = 2, 3, ...,R (1) Binary variables Frs are introduced to model if sample s belongs to region r or 14 not. Modelling of which sample belongs to which region is achieved with the following constraints: Xr−1m −U ′(1 −Frs ) ≤ Asm ∀s,r = 2, 3, ...,R,m = m * (2) Asm ≤ Xrm + U ′(1 −Frs ) ∀s,r = 1, 2, ...,R− 1,m = m * (3) When sample s belongs to region r (i.e. Frs = 1), Asm* falls into the region bounded by the two consecutive break-points Xr−1 m* and Xr m* on feature m∗; otherwise the two sets of constraints become redundant. A visualisation of break-points and regions is provided in Figure 1: Figure 1: Break-points and regions The following constraints restrict that each sample belongs to one and only one region: ∑ r Frs = 1 ∀s (4) For sample s, its predicted output value for region r, Predrs, is as below: Predrs = ∑ m AsmW r m + B r ∀s,r (5) For any sample s, its training error is equal to the absolute deviation between 15 the real output and the predicted output for the region r where it belongs to (i.e. Frs = 1): Ds ≥ Ys −Predrs −U ′′(1 −Frs ) ∀s,r (6) Ds ≥ Predrs −Ys −U ′′(1 −Frs ) ∀s,r (7) The objective function is to minimise the sum of absolute training error: min ∑ s Ds (8) The final model, named as Optimal Piece-wise Linear Regression Analysis (OPLRA) in this work, consists of a linear objective function and several linear constraints, and the presence of both binary and continuous variables define an MILP prob- lem, which can be solved to global optimality by standard solution algorithms,350 for example branch and bound. A heuristic solution procedure is also employed in this work to identify the partition feature and the number of regions, as de- scribed in Figure 2 below. The heuristic procedure starts with solving a linear regression on the entire set355 of data with least absolute deviation. Subsequently, each input feature in turn serves as partition feature m* once and the OPLRA model is solved while al- lowing two regions (i.e. R = 2). The feature corresponding to the minimum training error is kept and if its error represents a percentage reduction of more than β from the global linear regression without data partition, the procedure360 continues; otherwise it is decided that two-region piecewise linear regression does not provide a desirable improvement upon the classic linear regression, and the initially derived linear regression function without sample partition is obtained for prediction. The parameter β, taking value between 0 and 1, quantifies the percentage reduction in training error that justifies adding one more region.365 16 Figure 2: Heuristic procedure to identify the partition feature and the number of regions 17 If two-region piecewise regression is accepted, the corresponding partition fea- ture is retained for further analysis while the number of regions is iteratively increased, until the β training reduction criterion is not satisfied between iter- ations. 370 β is the only user-specific parameter in our proposed regression method, which requires fine tuning. A small value may cause over-fitting, i.e. too many regions are allowed and each region contains only small number of samples, which then results in unreliable construction of local linear regression functions; while a value excessively large will lead to premature growing of regions, which then375 under-fit the data. In Results and Discussion section, we will test a series of values on a number of benchmark datasets and select the optimal value corre- sponding to the most robust prediction performance. The constructed piecewise linear regression functions are then used to predict380 the output value of new samples. A testing sample is firstly assigned to one of the regions, and the regression coefficients for that region are used to estimate its output value. 2.2. An illustrative example In order to better illustrate the training of the proposed regression method,385 a simulation model is taken from literature. In brief, the illustrative example (Palmer & Realff, 2002) describes the operation of a continuous stirred tank reactor, where a chain reaction of A → B → C takes place. An inlet stream containing both reactant A and B enters the reactor and the desirable output is component B. There are 4 independent input variables to the simulation model,390 including temperature of the reactor (T), volume of the reactor (V ), concen- tration of A and B in the inlet stream (CAin and CBin). The output to be predicted is the production rate of B (P). The process and associated variables are described in Figure 3. 395 18 Figure 3: Illustrative example of a continuous stirred tank reactor With latin hypercube sampling technique (Helton & Davis, 2003) employed to specify a set of data points, we run the simulation model and collect 300 samples. The goal of the regression analysis is to approximate the functional relationship between output variable P and input variables including T, V , CAin and CBin using piece-wise linear functions. The step-wise description of400 the training procedure is presented in Table 1 below. Initially, a linear regression function is fitted to the entire dataset without fea- ture segmentation, which gives an absolute deviation of 1677.78. The second iteration of the method solves 4 independent OPLRA models allowing 2 regions405 each, respectively specifying T,V,CAin and CBin as partition feature. The two- region piece-wise linear functions constructed while partitioning on T appears to yield lower training errors (i.e. 1030.63) than the other 3, and therefore is taken as the solution for iteration 2. This represents a significant improvement (i.e. 38.57%) from the initial global linear regression function. From iteration410 3, the partition feature is fixed as T while one more region is allocated for each increased iteration. Iteration 3 and 4 respectively lowers the training error to 876.66 and 807.12. The iterative procedure terminates when the β criterion is not satisfied, e.g. if β = 20%, then the iterative procedure terminates at the 19 third iteration and the final regression function has 2 regions; if β = 10%, then415 the final regression function has 3 regions. Table 1: Piecewise regression functions built at each step of training procedure Iteration Number of regions Partition feature Training error Training error improvement Functional relationship 1 1 NONE 1677.78 P = 1.0240T + 0.0054CAin + 0.0125CBin + 0.4340V − 333.54 2 2 T 1030.63 38.57% P = { 0.7413T + 0.0040CAin + 0.0102CBin + 0.3406V − 238.74, T ≤ 213.21 1.7156T + 0.0111CAin + 0.0315CBin + 0.7574V − 592.63, T > 213.21 2 V 1143.49 P = { 0.5952T + 0.0033CAin + 0.0056CBin + 0.4533V − 194.26, V ≤ 42.38 1.4781T + 0.0083CAin + 0.0195CBin + 0.4773V − 48.70, V > 42.38 2 CAin 1485.65 P = { 0.8930T + 0.0057CAin + 0.0152CBin + 0.4161V − 293.45, CAin ≤ 3528.43 1.4857T + 0.0073CAin + 0.0070CBin + 0.5929V − 489.45, CAin > 3528.43 2 CBin 1627.73 P = { 1.0242T + 0.0056CAin + 0.0118CBin + 0.4241V − 333.49, CBin ≤ 458.21 1.1105T + 0.0050CAin − 0.1405CBin + 0.5813V − 291.00, CBin > 458.21 3 3 T 876.66 14.94% P =   0.5815T + 0.0030CAin + 0.0097CBin + 0.2654V − 184.45, T ≤ 303.25 1.1353T + 0.0062CAin + 0.0176CBin + 0.4579V − 373.68, 303.25 < T ≤ 316.62 1.8764T + 0.0119CAin + 0.0394CBin + 0.8617V − 654.41, T > 316.62 4 4 T 807.12 7.93% P =   0.5815T + 0.0030CAin + 0.0097CBin + 0.2654V − 184.45, T ≤ 303.25 1.2648T + 0.0054CAin + 0.0148CBin + 0.4510V − 409.61, 303.32 < T ≤ 312.21 1.4872T + 0.0084CAin + 0.0202CBin + 0.6667V − 503.10, 312.21 < T ≤ 320.77 1.9930T + 0.0128CAin + 0.0360CBin + 0.8871V − 695.65, T > 320.77 ... Overall, the key features of our proposed piecewise linear regression method are summarised here: 1) our method identifies one key partition feature and sepa- rate samples into multiple complementary regions on it, 2) each region has the420 flexibility of being fitted by its own linear regression function, with all input fea- tures allowed to have different regression coefficients across different regions, 3) there is only one tuning parameter β, 4) compared with algorithms like kernel- based SVR and MLP, the constructed regression function is easy to understand, as it exhibits linear relationships for different regions.425 It is noted here that the obtained relationship between input and output vari- ables, presented as rules in Table 1, can be used to build an expert system for the above operation. Given the chain reaction of A → B → C in stirred tank reactor (Palmer & Realff, 2002), domain experts perform experiments to create430 a database of samples for different levels of temperature, reactor volume and concentrations of reactants. Our proposed piece-wise regression method is then applied to automatically extract the rules that predict production rate from 20 temperature, reactor volume and reactant concentrations. The rules will be difficult to be provided directly from even chemical engineering experts due to435 the complex nature of the reaction. Since the above extracted rules can calcu- late a production rate value for any random value of temperature, tank volume and reactant concentrations, regardless of if they obey physical laws (must be positive) or are valid for the reaction of interest, expert knowledge should be incorporated to further refine the rules. For example, expert knowledge can be440 used to constraint the applicable temperature range outside which liquid phase will vaporise to gas phase or freeze to solid phase, making it impossible for the reaction to proceed as normal. The final expert system will allow users to query the likely outcome, as production rate or no reaction, of any combination of values of temperature, reactor volume and reactant concentrations.445 In the next section, a number of real world regression problems are employed to benchmark the predictive performance of our proposed model. 3. Results and Discussion A total number of 7 real world datasets have been downloaded from UCI ma- chine learning repository (http://archive.ics.uci.edu/ml/) (Bache & Lichman,450 2013) to test the prediction performance of our proposed method. The first re- gression problem Yacht Hydrodynamics predicts the hydrodynamic performance of sailing yachts from 7 features describing the hull dimensions and velocity of the boat for 308 samples. Energy Efficiency (Tsanas & Xifara, 2012) collects data corresponding to 768 building shapes, described by 8 features including455 wall area, root area and so on. The aims are to establish the relationship be- tween either heating load or cooling load requirements and the 8 parameters of the building. The third example, Concrete Strength (Yeh, 1998), looks into the relationship between compressive strength of concrete and 8 input variables, including water concentration and age, with 1030 samples of different concretes.460 Airfoil dataset concerns how the different airfoil blade desings, wind speed and angles of attack affect the sound pressure level. The last 2 case studies, Red 21 Wine Quality and White Wine Quality (Cortez et al., 2009), aims to predict experts’ preference of red and while wine taste with 11 physicochemical features of the wines. Almost 1600 red wine and 4900 white wine samples have been465 obtained for analysis. For each of the 7 benchmark datasets, a 5-fold cross validation, is performed to estimate the predictive accuracy of the proposed method. Given a dataset, 5-fold cross validation randomly splits the samples into 5 subsets of equal size.470 Each subset is in turn held out once while the other 4 subsets of samples are used in the training process to derive the regression function. The holdout set is then used to validate the predictive accuracy of the constructed regression function. We conduct 10 rounds of 5-fold cross validation by performing differ- ent random sample splits, and the mean absolute prediction errors (MAE) are475 averaged over 50 testing sets as the final error. For comparison purposes, a number of state-of-the-art regression methods have been implemented, including linear regression, MLP, kriging, SVR, KNN, ran- dom forest, MARS, PaceRegression and ALAMO. Linear regression, MLP, krig-480 ing, SVR, KNN and PaceRegression are implemented in WEKA machine learn- ing software (Hall et al., 2009). For KNN, the number of nearest neighbours is selected as 5, while for other methods their default settings have been retained. Random forest is implemented using Orange (Demšar et al., 2013). We use the MATLAB toolbox called ARESlab (Jekabsons, 2011) for MARS. ALAMO485 is reproduced using the General Algebraic Modeling System (GAMS) (GAMS Development Corporation, 2013), and basis function forms including polyno- mial of degrees up to 3, pair-wise multinomial terms of equal exponents up to 3, exponential and logarithmic forms are provided for each dataset. Our proposed method is also implemented in GAMS. Both ALAMO and our proposed model490 are solved using Cplex MILP solver, with optimality gap set as 0. Computa- tional resource limit is set as 200 seconds for each solving of OPLRA model in our proposed method. 22 Figure 4: Sensitivity analysis of β. The numbers above points in each plot correspond to the average numbers of final regions. 3.1. Sensitivity analysis for β In this subsection, a sensitivity analysis is performed for the parameter β,495 which serves as a terminating criterion of the iterative training procedure for our proposed method. Taking value between 0 and 1, β defines the minimum percentage training error reduction that must be achieved to justify the alloca- tion of an extra region. A range of values have been tested, including: 0.2, 0.15, 0.10, 0.05, 0.03 and 0.01. The results of the sensitivity analysis are provided500 in Figure 4 below. 23 Figure 4 describes how mean absolute error changes with β. The numbers attached to the points in each plot are the average numbers of final regions, which always go up as β decreases. For Yacht Hydrodynamics example, setting505 β = 0.20 results in just more than 4 final regions. Decrease the β value to 0.15 increases slightly the prediction error with marginally higher number of regions. Further decrease β to 0.10 leads to lowest mean prediction error of 0.648 with an average of 5 regions, before excessively low values of β over-fits the unseen testing samples by yielding much increased prediction error. For En-510 ergy Efficiency Heating case study, when β = 0.10,0.15 and 0.20 our proposed regression method constructs piece-wise regression functions of an average of 3 regions, yielding MAE of 0.907. Smaller values of β leads to about 5 regions, which are shown to predict the testing samples with higher accuracy (MAE around 0.810). In terms of Energy Efficiency Cooling and Concrete Strength515 examples, similar phenomenon can be observed that when β takes overly high values (i.e. 0.20, 0.15 ), the proposed method terminates prematurely with only 2 regions and relatively high MAE. More regions are allowed by lowering β, which gives higher prediction accuracies. On Airfoil case study, the proposed method outputs global multiple linear regression functions without data par-520 titions when β = 0.20. As β decreases, more regions are permitted, which predict unseen samples with better accuracy. With regards to Red Wine Qual- ity dataset, the optimal prediction occurs when β = 0.03. On the last example of White Wine Quality, 2-region piece-wise regression functions achieved with β = 0.01, 0.03, 0.05 outperforms global multiple linear regressions for higher525 values of β. It can be seen from Figure 4 that the range of values between 0.01 and 0.05 gen- erally lead to smaller prediction error than higher values of β. For all datasets except Yacht Hydrodynamics, prediction errors of β = 0.01, 0.03 and 0.05 are530 evidently smaller than that of β = 0.10, 0.15 and 0.20. Within the range be- tween 0.01 and 0.05, there is no clear optimal value for β as different values have 24 different effects on the accuracy. We instead seek to identify the most robust value for β, which gives consistently desirable prediction accuracy across a wide range of problems. For each dataset, we normalise the MAE of each β accord-535 ing to the formula: MAEβ−minβ MAEβ minβ MAEβ . For example, in Yacht Hydrodynamics, original MAE of β = 0.01 is normalised from 0.7131 to 0.7131−0.6481 0.6481 = 10.0%, where 0.6481 is the lowest MAE achieved when β = 0.10. The normalised MAE of each β represents the actual deviation of it compared to the lowest error, and is averaged over all examples to reflect its overall competitiveness.540 Overall β = 0.03 provides the smallest normalised MAE of 1.7%, which is marginally lower than these of β = 0.01 and β = 0.05, respectively as 1.8% and 2.8%. Even higher values of β correspond to noticeably larger normalised MAE (5.6%, 9.7% and 12.3% for β = 0.10,0.15 and 0.20, respectively). The con-545 sistently small normalised MAE, while β is between 0.01 and 0.05, show that our proposed regression method is robust with respect to the only user tuning parameter β. Finally β is set to 0.03 when comparing with other competing methods in literature. 3.2. Prediction performance comparison550 After identifying a value (i.e. 0.03 ) for the only tuning parameter β in our proposed regression method, we now compare the accuracy of the proposed method against some popular regression algorithms with the same set of 7 ex- amples. The results of the comparison are available in Table 2 below. Table 2: Comparative testing of different regression methods on benchmark datasets Hydrodynamics Energy Heating Energy Cooling Concrete Airfoil Red Wine White Wine linear regression 7.270 2.089 2.266 8.311 0.037 0.506 0.586 MLP 0.809 0.993 1.924 6.229 0.035 0.581 0.623 Kriging 4.324 1.788 2.044 6.224 0.030 0.496 0.576 SVR 6.445 2.036 2.191 8.212 0.037 0.500 0.585 KNN 5.299 1.937 2.148 7.068 0.026 0.515 0.537 Random forest 3.516 1.435 1.644 5.309 0.030 0.484 0.519 MARS 1.011 0.796 1.324 4.871 0.035 0.502 0.570 PaceRegression 7.233 2.089 2.261 8.298 0.037 0.507 0.586 ALAMO 0.787 2.722 2.765 8.044 0.032 0.594 0.639 Proposed 0.706 0.810 1.278 4.870 0.029 0.481 0.551 555 25 In Table 2 and each tested dataset, the lowest prediction error achieved among all implemented regression methods is marked with bold. On Hydrodynamics problem, the proposed method in this work provides an MAE of 0.706, which is lower than any other competing algorithm. ALAMO, MLP and MARS follow560 closely with MAE of 0.787, 0.809 and 1.011, respectively. Mean error rates of the rest of the methods are between 3 and 8. On Energy Efficiency Heating, MARS emerges as the most accurate algorithm with an mean absolute error of 0.796, which is closely matched by our proposed method and MLP. Mean pre- diction errors of the other approaches are almost all twice as large as that of the565 MARS. In terms of Energy Efficiency Cooling dataset, the proposed method, MARS, random forest and MLP are the top 4 performers with MAE between 1.278 and 1.924. On Concrete Strength, our proposed approach and MARS, with an MAE of 4.870 and 4.871, again emerge as the leading methods from random forest, Kriging, MLP and the others. When it comes to Airfoil example,570 all the competing algorithms achieve similar prediction accuracies, with KNN topping the chart with an MAE of 0.026. The proposed approach in this work is a merely 0.003 far behind, with kriging and random forest a further 0.001 behind. A merely 0.011 separates the 10 methods. Lastly, on the two Wine Quality examples, our proposed approach is respectively ranked as 1st and 3rd575 best method. Overall, for 4 out of the 7 datasets, including Yacht Hydrodynamics, Energy Efficiency Cooling, Concrete Strength and Red Wine Quality, the proposed re- gression method achieves the lowest prediction errors. For the other 3 tested580 examples, the proposed method still perform competitively as being second on Energy Efficiency Heating, Airfoil and third on White Wine Quality. As there does not exist a single regression method which can always outper- form others on all datasets, a desirable regression algorithm should demonstrate585 consistently competitive prediction accuracy. In order to more comprehensively 26 Figure 5: Scoring of regression methods evaluate the relative competitiveness of all the implemented approaches, we em- ploy the following scoring strategy: for each problem, the regression methods are ranked in descending order according to their mean prediction error. The best regression method corresponding to the lowest prediction error is awarded590 the maximum score of 10, the second best regression method corresponding to the second lowest prediction error is assigned a score of 9 and so on. The scores of each regression approach are averaged over the 7 datasets, which represent the overall performance of the method. The higher the score, the better the relative performance of a method. The scores of the different regression ap-595 proaches used in this work are presented in Figure 5 below. According to Figure 5, the proposed method is shown to be the most accu- 27 rate and robust regression algorithm among all, achieving a score of 9.43 out of a possible 10. Random forest and MARS are second and third according to600 the ranking with scores of 8 and 7.43, followed by kriging, KNN, MLP, SVR, ALAMO, linear regression and PaceRegression in descending order. The advan- tages of the proposed regression method is quite obvious compared with other implemented methods. 605 Lastly, we take a look at, for each dataset, the number of regions and the key partition feature determined by our proposed regression method. The results are summarised in Table 3. It is clear that the proposed segmented regres- sion method provides good interpretability as the number of regions are small (usually between 2 to 4 and at most 5). The partition features may release im-610 portant insights of the underlying system as the output variables change more dramatically across different ranges alone this feature. Table 3: Number of regions and partition feature by our proposed method Dataset Number of regions Partition feature Hydrodynamics 5 Froude number Energy Heating 3 Wall Area Energy Cooling 3 Wall Area Concrete 3 Age Airfoil 4 Frequency Red Wine 2 Alcohol White Wine 2 Volatile acidity 3.3. Strength and weakness of the proposed piece-wise regression method No regression method will be the best for all problems. In this section, we give some general illustration of the pros and cons of the proposed OPLRA615 piece-wise linear regression method, and compare it against some other litera- ture methods. OPLRA piece-wise regression is inherently deterministic, which means the same solution is always guaranteed regardless of the number of runs executed. This is an advantage of OPLRA against stochastic-based methods, for example MLP, where each execution would typically end up with a different620 28 locally optimal solution. On the other hand, OPLRA is intuitive and easy to interpret. OPLRA approximates the potentially highly non-linear relationship between output and input variables as piece-wise linear algebraic functions, the formalism of which is easy to understand, interpret and use for users without sophisticated background knowledge. Contrarily, the mechanisms of certain625 methods like SVR, MLP and Kriging lack transparency as the former two work as black box techniques and the latter requires detailed knowledge on statistics. The small number of user-specified parameters involved in training of OPLRA is another remarkable advantage. β is the only tuning parameter in the proposed OPLRA, which produces robust predictive performance with regards to varying630 values of β as shown in the following Results and Discussion section. Conversely, usage of certain regression methods, including SVR, MLP and Kriging requires tuning a large number of parameters, making it a challenging task to identify their optimal values. More importantly, OPLRA piece-wise regression achieves more accurate and robust prediction performance against other methods. Using635 a large number of real word problems, OPLRA is shown to outperform popular state-of-the-art multivariate regression methods in terms of prediction accuracy and does so consistently across a number of real world problems. With regards to shortcomings of our proposed piece-wise regression method,640 training of OPLRA generally consumes more computational resource than the existing methods in literature. Solving OPLRA combinatorial optimisation model is indeed a more computationally intensive task than heuristic-based methods, for example regression tree, MARS and quadratic programming-based SVR. Therefore, we note here that this method is not designed for online ap-645 plications where computation time is valued more than the prediction accu- racy/model interpretability. Another limitation of OPLRA is that it permits segmentation of only one input variable, which may not be adequate for the datasets where trend of the output variable changes dramatically in more than one input variables.650 29 4. Concluding Remarks This work addresses the problem of multivariate regression analysis, where one seeks to estimate the complex relationship between dependent output vari- ables and independent input variables from training samples. The identified relationships can then be used to make predictions for unseen observations. We655 have proposed a novel piece-wise regression method, which approaches the prob- lem by segmenting one input variable into multiple mutually exclusive regions and simultaneously fitting each one with a distinct multivariate linear function. An optimisation model has been proposed to optimise the locations of break points and regression coefficients for each region, while a heuristic procedure660 has also been introduced to find the key partition feature and the number of break-points by repeatedly solving the optimisation models until a satisfactory solution is identified. To demonstrate the applicability and efficiency of the proposed piece-wise re-665 gression method, 7 real world problems have been employed, covering a wide range of application domains. To benchmark the predictive capability of the pro- posed method, we have also implemented various popular regression methods in literature for comparison, including support vector regression, artificial neural network, MARS and K nearest neighbour. Computational experiments clearly670 indicate that our proposed piece-wise regression method achieves consistently high predictive accuracy as leading to the lowest prediction errors for 4 out of 7 datasets, second lowest errors for 2 datasets and third lowest error for the other example. The results confirm our proposed method as a reliable alternative to traditional regression analysis methods. Another remarkable advantage of our675 proposed method is that the learned model can be conveniently expressed as a set of if-then rules that are compact and easily understandable. From Table 3, it is clear that the number of if-then rules identified by our method as the hidden patterns in the large scale databases (up to thousands expert curated samples) are extremely small (usually 2 to 3 and at most 5). The model inter-680 30 pretability of the proposed piece-wise regression is a desirable advantage over black modelling techniques, for example support vector regression and neural network. With regards to research contribution in expert and intelligent systems, the685 generic machine learning method proposed in this work can be used to con- struct a large number of automatic decision making or support systems for var- ious domain applications. As the quality and coverage of information contained in knowledge base critically affects the efficiency of any expert and intelligent system, our proposed machine learning method can serve to automatically and690 more efficiently acquire knowledge from database by approximating the relation- ship between output and input variables as rules. Subsequently, the discovered knowledge can be used to generate forecasts to users’ enquiry. To further improve the efficiency of the proposed piece-wise regression method695 in this work, the following limitations can be considered for refinement. As the piece-wise regression method proposed in this work can only partition a single input variable, one potential improvement is to generalise the method so that to permit segmentation of multiple variables so as to better capture the non- linearity in datasets. Secondly, as our proposed method in this work can only700 handle continuous input variables, we plan to improve its applicability by gener- alising it to deal with categorical input variables having many distinct levels. In addition, the relationship between output and input variables are approximated as linear for each segment in the current method, which may not adequately model the underlying patterns. To overcome this, more complex non-linear ba-705 sis functions, for example polynomial, exponential and logarithmic forms, can be added to allow more flexibility. Another limitation of our method is the rel- atively high computational cost, which may restrict its usage in certain online applications, where learning speed of the method is considered more important than actual prediction accuracy. To tackle this problem, we can explore more710 efficient heuristic solution procedures that, by estimating the possible break- 31 point positions and constricting the solution space, more quickly converge to a quality solution. In terms of practical future applications in expert and intelligent systems, the715 proposed piece-wise regression method can benefit many via automatic extrac- tion of knowledge from databases and generate accurate forecasts. As examples, we have identified the following directions as possible avenues worth investiga- tion in the near future. First, our proposed method can be incorporated into the construction of a decision support expert system that continuously predicts720 the personalised risk of prisoner with mental illness being released from the jail, aiding clinician for decision making (Constantinou et al., 2015). Other applica- tions that can benefit from our work include intelligent drowsiness monitoring system and stock price prediction. In drowsiness monitoring, the proposed regression model can be built into an intelligent fatigue detection equipment,725 which records the dynamic physiological signals of drivers or medical staffs and continuously predicts their level of fatigues. A warning will be automatically issued when the model predicts the fatigue level of subjects to be above a pre- specified threshold level (Chen et al., 2015). In financial area, our method can help with construction of an automatic system that forecasts the stock price730 based on the ever-changing variables quantifying the current performance of a company, including assets, liabilities and income, providing management with data support to make better financial benefits (Ballings et al., 2015). Lastly, the proposed method developed here can also find application in airline industry where managers and decision makers can benefit from a framework powerful of735 predicting the level of customer satisfaction from various aspects of services, and therefore making it possible for them to carefully allocate resource to maximise customer loyalty (Leong et al., 2015). 32 5. Acknowledgements Funding from the UK Engineering and Physical Sciences Research Coun-740 cil (to LY, SL and LGP through the EPSRC Centre for Innovative Manufac- turing in Emergent Macromolecular Therapies), the UK Leverhulme Trust (to ST and LGP, RPG-2012-686), the European Union(to ST, HEALTH-F2-2011- 261366),and the Centre for Process Systems Engineering (CPSE) at Imperial and University College London are gratefully acknowledged.745 References Afantitis, A., Melagraki, G., Sarimveis, H., Koutentis, P. A., Markopoulos, J., & Igglessi-Markopoulou, O. (2006). A novel {QSAR} model for predicting induction of apoptosis by 4-aryl-4h-chromenes. Bioorganic and Medicinal Chemistry , 14 , 6686 – 6694.750 Alonso, F., Martnez, L., Prez, A., & Valente, J. P. (2012). Cooperation between expert knowledge and data mining discovered knowledge: Lessons learned. Expert Systems with Applications , 39 , 7524 – 7535. Andrs, J. D., Lorca, P., de Cos Juez, F. J., & Snchez-Lasheras, F. (2011). Bankruptcy forecasting: A hybrid approach using fuzzy c-means clustering755 and multivariate adaptive regression splines (mars). Expert Systems with Applications , 38 , 1866 – 1875. Bache, K., & Lichman, M. (2013). UCI machine learning repository. Bai, Y., Wang, P., Li, C., Xie, J., & Wang, Y. (2014). A multi-scale relevance vector regression approach for daily urban water demand forecasting. Journal760 of Hydrology , 517 , 236 – 245. Ballings, M., den Poel, D. V., Hespeels, N., & Gryp, R. (2015). Evaluating multiple classifiers for stock price direction prediction. Expert Systems with Applications , 42 , 7046 – 7056. 33 Balshi, M. S., Mcguire, A. D., Duffy, P., Flannigan, M., Walsh, J., & Melillo, J.765 (2009). Assessing the response of area burned to changing climate in western boreal north america using a multivariate adaptive regression splines (mars) approach. Global Change Biology , 15 , 578–600. Beck, J., Friedrich, D., Brandani, S., Guillas, S., & Fraga, E. (2012). Surro- gate based optimisation for design of pressure swing adsorption systems. In770 Proceedings of the 22nd European Symposium on Computer Aided Process Engineering . Bermolen, P., & Rossi, D. (2009). Support vector regression for link load pre- diction. Computer Networks , 53 , 191 – 201. QoS Aspects in Next-Generation Networks.775 Biau, G. (2012). Analysis of a random forests model. Journal of Machine Learning Research, 13 , 1063–1095. Breiman, L. (2001). Random forests. Machine Learning , 45 , 5–32. Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classifica- tion and Regression Trees . Wadsworth.780 Brus, D. J., & Heuvelink, G. B. (2007). Optimization of sample patterns for universal kriging of environmental variables. Geoderma, 138 , 86 – 95. Caballero, J. A., & Grossmann, I. E. (2008). An algorithm for the use of surrogate models in modular flowsheet optimization. AIChE Journal , 54 , 2633–2650.785 Cavanaugh, K. C., Kellner, J. R., Forde, A. J., Gruner, D. S., Parker, J. D., Rodriguez, W., & Feller, I. C. (2014). Poleward expansion of mangroves is a threshold response to decreased frequency of extreme cold events. Proceedings of the National Academy of Sciences , 111 , 723–727. Chang, C.-C., & Lin, C.-J. (2011). Libsvm: A library for support vector ma-790 chines. ACM Transactions on Intelligent Systems and Technology , 2 , 27:1– 27:27. 34 Chen, L., Zhao, Y., Zhang, J., & zhong Zou, J. (2015). Automatic detection of alertness/drowsiness from physiological signals using wavelet-based nonlinear features and machine learning. Expert Systems with Applications , 42 , 7344 –795 7355. Chen, Q.-L., Wu, K.-J., & He, C.-H. (2014). Thermal conductivity of ionic liquids at atmospheric pressure: Database, analysis, and prediction using a topological index method. Industrial and Engineering Chemistry Research, 53 , 7224–7232.800 Cherkassky, V., & Ma, Y. (2004). Practical selection of svm parameters and noise estimation for svm regression. Neural Networks , 17 , 113 – 126. Comrie, A. C. (1997). Comparing neural networks and regression models for ozone forecasting. Journal of the Air and Waste Management Association, 47 , 653–663.805 Constantinou, A. C., Freestone, M., Marsh, W., Fenton, N., & Coid, J. (2015). Risk assessment and risk management of violent reoffending among prisoners. Expert Systems with Applications , 42 , 7511 – 7529. Cortez, P., Cerdeira, A., Almeida, F., Matos, T., & Reis, J. (2009). Modeling wine preferences by data mining from physicochemical properties. Decision810 Support Systems , 47 , 547 – 553. Smart Business Networks: Concepts and Empirical Evidence. Cozad, A., Sahinidis, N. V., & Miller, D. C. (2014). Learning surrogate models for simulation-based optimization. AIChE Journal , 60 , 2211–2227. Davis, E., & Ierapetritou, M. (2008). A kriging-based approach to minlp con-815 taining black-box models and noise. Industrial and Engineering Chemistry Research, 47 , 6101–6125. Demšar, J., Curk, T., Erjavec, A., Črt Gorup, Hočevar, T., Milutinovič, M., Možina, M., Polajnar, M., Toplak, M., Starič, A., Štajdohar, M., Umek, L., 35 Žagar, L., Žbontar, J., Žitnik, M., & Zupan, B. (2013). Orange: Data mining820 toolbox in python. Journal of Machine Learning Research, 14 , 2349–2353. Dua, V. (2010). A mixed-integer programming approach for optimal configura- tion of artificial neural networks. Chemical Engineering Research and Design, 88 , 55 – 60. Eronen, A., & Klapuri, A. (2010). Music tempo estimation with k -nn regression.825 Audio, Speech, and Language Processing, IEEE Transactions on, 18 , 50–57. Fanelli, G., Gall, J., & Van Gool, L. (2011). Real time head pose estimation with random regression forests. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on (pp. 617–624). Friedman, J. H. (1991). Multivariate adaptive regression splines. The Annals830 of Statistics , 19 , pp. 1–67. GAMS Development Corporation (2013). General Algebraic Modeling System (GAMS) Release 24.2.1.. Washington, DC, USA. Genuer, R., Poggi, J.-M., & Tuleau-Malot, C. (2010). Variable selection using random forests. Pattern Recognition Letters , 31 , 2225 – 2236.835 Gevrey, M., Dimopoulos, I., & Lek, S. (2003). Review and comparison of meth- ods to study the contribution of variables in artificial neural network models. Ecological Modelling , 160 , 249 – 264. Modelling the structure of acquatic communities: concepts, methods and problems. Ghasemi, J., Saaidpour, S., & Brown, S. D. (2007). Qspr study for estimation840 of acidity constants of some aromatic acids derivatives using multiple linear regression (mlr) analysis. Journal of Molecular Structure: THEOCHEM , 805 , 27 – 32. Greene, M., Rolfson, O., Garellick, G., Gordon, M., & Nemes, S. (2015). Improved statistical analysis of pre- and post-treatment patient-reported845 36 outcome measures (proms): the applicability of piecewise linear regression splines. Quality of Life Research, 24 , 567–573. Gudise, V., & Venayagamoorthy, G. (2003). Comparison of particle swarm optimization and backpropagation as training algorithms for neural networks. In Swarm Intelligence Symposium, 2003. SIS ’03. Proceedings of the 2003850 IEEE (pp. 110–117). Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., & Witten, I. H. (2009). The weka data mining software: An update. SIGKDD Explorations Newsletter , 11 , 10–18. Helton, J., & Davis, F. (2003). Latin hypercube sampling and the propagation855 of uncertainty in analyses of complex systems. Reliability Engineering and System Safety , 81 , 23 – 69. Henao, C. A., & Maravelias, C. T. (2010). Surrogate-based process synthesis. In S. Pierucci, & G. B. Ferraris (Eds.), 20th European Symposium on Computer Aided Process Engineering (pp. 1129 – 1134). Elsevier volume 28 of Computer860 Aided Chemical Engineering . Henao, C. A., & Maravelias, C. T. (2011). Surrogate-based superstructure optimization framework. AIChE Journal , 57 , 1216–1232. Hill, T., Marquez, L., O’Connor, M., & Remus, W. (1994). Artificial neural network models for forecasting and decision making. International Journal of865 Forecasting , 10 , 5 – 15. Jekabsons, G. (2011). Areslab: Adaptive regression splines toolbox for mat- lab/octave. Khayet, M., Cojocaru, C., & Zakrzewska-Trznadel, G. (2008). Response surface modelling and optimization in pervaporation. Journal of Membrane Science,870 321 , 272 – 283. 37 Khuri, A. I., & Mukhopadhyay, S. (2010). Response surface methodology. Wiley Interdisciplinary Reviews: Computational Statistics , 2 , 128–149. Kleijnen, J. P. (2009). Kriging metamodeling in simulation: A review. European Journal of Operational Research, 192 , 707 – 716.875 Kleijnen, J. P. C., & Beers, W. C. M. v. (2004). Application-driven sequential designs for simulation experiments: Kriging metamodelling. The Journal of the Operational Research Society , 55 , pp. 876–883. Kone, E. R. S., & Karwan, M. H. (2011). Combining a new data classification technique and regression analysis to predict the cost-to-serve new customers.880 Computer and Industrial Engineering , 61 , 184–197. Korhonen, K. T., & Kangas, A. (1997). Application of nearestneighbour regres- sion for generalizing sample tree information. Scandinavian Journal of Forest Research, 12 , 97–101. Leathwick, J., Elith, J., & Hastie, T. (2006). Comparative performance of885 generalized additive models and multivariate adaptive regression splines for statistical modelling of species distributions. Ecological Modelling , 199 , 188 – 196. Predicting Species Distributions Results from a Second Workshop on Advances in Predictive Species Distribution Models, held in Riederalp, Switzerland, 2004.890 Leong, L.-Y., Hew, T.-S., Lee, V.-H., & Ooi, K.-B. (2015). An semartificial- neural-network analysis of the relationships between servperf, customer sat- isfaction and loyalty among low-cost and full-service airline. Expert Systems with Applications , 42 , 6620 – 6634. Levis, A. A., & Papageorgiou, L. G. (2005). Customer demand forecasting895 via support vector regression analysis. Chemical Engineering Research and Design, 83 , 1009 – 1018. 38 Li, B., Zhang, L., Yan, Q., & Xue, Y. (2014a). Application of piecewise lin- ear regression in the detection of vegetation greenness trends on the tibetan plateau. International Journal of Remote Sensing , 35 , 1526–1539.900 Li, S., Feng, L., P., B., & Seidel-Morgenstern, A. (2014b). Using surrogate models for efficient optimization of simulated moving bed chromatography. Computers and Chemical Engineering , 67 , 121 – 132. Li, Y., Gong, S., & Liddell, H. (2000). Support vector regression and classifica- tion based multi-view face detection and recognition. In Automatic Face and905 Gesture Recognition, 2000. Proceedings. Fourth IEEE International Confer- ence on (pp. 300–305). Lloyd, C. D., & Atkinson, P. M. (2002). Deriving dsms from lidar data with kriging. International Journal of Remote Sensing , 23 , 2519–2524. Loh, W.-Y. (2011). Classification and regression trees. Wiley Interdisciplinary910 Reviews: Data Mining and Knowledge Discovery , 1 , 14–23. Lu, C.-J., Lee, T.-S., & Chiu, C.-C. (2009). Financial time series forecasting us- ing independent component analysis and support vector regression. Decision Support Systems , 47 , 115 – 125. Magnani, A., & Boyd, S. (2009). Convex piecewise-linear fitting. Optimization915 and Engineering , 10 , 1–17. Malash, G. F., & El-Khaiary, M. I. (2010). Piecewise linear regression: A statistical method for the analysis of experimental adsorption data by the intraparticle-diffusion models. Chemical Engineering Journal , 163 , 256 – 263.920 Matthews, T. J., Steinbauer, M. J., Tzirkalli, E., Triantis, K. A., & Whittaker, R. J. (2014). Thresholds and the speciesarea relationship: a synthetic analysis of habitat island datasets. Journal of Biogeography , 41 , 1018–1028. URL: http://dx.doi.org/10.1111/jbi.12286. doi:10.1111/jbi.12286. 39 http://dx.doi.org/10.1111/jbi.12286 http://dx.doi.org/10.1111/jbi.12286 Miller, D. C., Syamlal, M., Mebane, D. S., Storlie, C., Bhattacharyya, D.,925 Sahinidis, N. V., Agarwal, D., Tong, C., Zitney, S. E., Sarkar, A., Sun, X., Sundaresan, S., Ryan, E., Engel, D., & Dale, C. (2014). Carbon capture simu- lation initiative: A case study in multiscale modeling and new challenges. An- nual Review of Chemical and Biomolecular Engineering , 5 , 301–323. PMID: 24797817.930 Minjares-Fuentes, R., Femenia, A., Garau, M., Meza-Velzquez, J., Simal, S., & Rossell, C. (2014). Ultrasound-assisted extraction of pectins from grape pomace using citric acid: A response surface methodology approach. Carbo- hydrate Polymers , 106 , 179 – 189. Muggeo, V. M. (2008). Segmented: an r package to fit regression models with935 broken-line relationships. R news , 8 , 20–25. Nuchitprasittichai, A., & Cremaschi, S. (2013). An algorithm to determine sample sizes for optimization with artificial neural networks. AIChE Journal , 59 , 805–812. Paliwal, M., & Kumar, U. A. (2009). Neural networks and statistical techniques:940 A review of applications. Expert Systems with Applications , 36 , 2 – 17. Palmer, K., & Realff, M. (2002). Metamodeling approach to optimization of steady-state flowsheet simulations: Model generation. Chemical Engineering Research and Design, 80 , 760 – 772. Pan, J., Kung, P., Bretholt, A., & Lu, J. (2014). Prediction of energys environ-945 mental impact using a three-variable time series model. Expert Systems with Applications , 41 , 1031 – 1040. Papadopoulos, H., Vovk, V., & Gammerman, A. (2011). Regression confor- mal prediction with nearest neighbours. Journal of Artificial Intelligence Research, 40 , 815–840.950 40 Quinlan, J. R. (1992). Learning with continuous classes. In Proceedings of the Australian Joint Conference on Artificial Intelligence (pp. 343–348). World Scientific. R Development Core Team (2008). R: A Language and Environment for Sta- tistical Computing . R Foundation for Statistical Computing Vienna, Austria.955 ISBN 3-900051-07-0. Rafiq, M., Bugmann, G., & Easterbrook, D. (2001). Neural network design for engineering applications. Computers and Structures , 79 , 1541 – 1552. Sampson, P. D., Richards, M., Szpiro, A. A., Bergen, S., Sheppard, L., Larson, T. V., & Kaufman, J. D. (2013). A regionalized national universal kriging960 model using partial least squares regression for estimating annual pm2.5 con- centrations in epidemiology. Atmospheric Environment , 75 , 383 – 392. Sarimveis, H., Alexandridis, A., Mazarakis, S., & Bafas, G. (2004). A new algorithm for developing dynamic radial basis function neural net- work models based on genetic algorithms. Computers and Chem-965 ical Engineering , 28 , 209 – 217. URL: http://www.sciencedirect. com/science/article/pii/S0098135403001698. doi:http://dx.doi.org/ 10.1016/S0098-1354(03)00169-8. Escape 12. Scheuber, M. (2010). Potentials and limits of the k-nearest-neighbour method for regionalising sample-based data in forestry. European Journal of Forest970 Research, 129 , 825–832. Smola, A., & Schlkopf, B. (2004). A tutorial on support vector regression. Statistics and Computing , 14 , 199–222. Strikholm, B. (2006). Determining the number of breaks in a piecewise lin- ear regression model . Working Paper Series in Economics and Finance 648975 Stockholm School of Economics. Tibshirani, R. (1994). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society, Series B , 58 , 267–288. 41 http://www.sciencedirect.com/science/article/pii/S0098135403001698 http://www.sciencedirect.com/science/article/pii/S0098135403001698 http://www.sciencedirect.com/science/article/pii/S0098135403001698 http://dx.doi.org/http://dx.doi.org/10.1016/S0098-1354(03)00169-8 http://dx.doi.org/http://dx.doi.org/10.1016/S0098-1354(03)00169-8 http://dx.doi.org/http://dx.doi.org/10.1016/S0098-1354(03)00169-8 Tibshirani, R. (2011). Regression shrinkage and selection via the lasso: a ret- rospective. Journal of the Royal Statistical Society: Series B (Statistical980 Methodology), 73 , 273–282. Toms, J. D., & Lesperance, M. L. (2003). Piecewise regression: A tool for identifying ecological thresholds. Ecology , 84 , 2034–2041. Toriello, A., & Vielma, J. P. (2012). Fitting piecewise linear continuous func- tions. European Journal of Operational Research, 219 , 86–95.985 Tsanas, A., & Xifara, A. (2012). Accurate quantitative estimation of energy performance of residential buildings using statistical machine learning tools. Energy and Buildings , 49 , 560 – 567. Venkatesh, K., Ravi, V., Prinzie, A., & den Poel, D. V. (2014). Cash demand forecasting in atms by clustering and neural networks. European Journal of990 Operational Research, 232 , 383 – 392. Viana, F. A. C., Simpson, T. W., Balabanov, V., & Toropov, V. (2014). Meta- modeling in Multidisciplinary Design Optimization: How Far Have We Really Come? AIAA Journal , 52 , 670–690. Wu, K.-J., Chen, Q.-L., & He, C.-H. (2014). Speed of sound of ionic liquids:995 Database, estimation, and its application for thermal conductivity prediction. AIChE Journal , 60 , 1120–1131. Xue, Y., Liu, S., Zhang, L., & Hu, Y. (2013). Integrating fuzzy logic with piecewise linear regression for detecting vegetation greenness change in the yukon river basin, alaska. International Journal of Remote Sensing , 34 , 4242–1000 4263. Yeh, I.-C. (1998). Modeling of strength of high-performance concrete using artificial neural networks. Cement and Concrete Research, 28 , 1797 – 1808. Zhang, J.-R., Zhang, J., Lok, T.-M., & Lyu, M. R. (2007). A hybrid particle swarm optimizationback-propagation algorithm for feedforward neural net-1005 42 work training. Applied Mathematics and Computation, 185 , 1026 – 1037. Special Issue on Intelligent Computing Theory and Methodology. Zhang, Y., & Sahinidis, N. V. (2013). Uncertainty quantification in co2 seques- tration using surrogate models from polynomial chaos expansion. Industrial and Engineering Chemistry Research, 52 , 3121–3132.1010 Zhu, Q., & Lin, H. (2010). Comparing ordinary kriging and regression kriging for soil properties in contrasting landscapes. Pedosphere, 20 , 594 – 606. 43 Introduction Method A novel regression method An illustrative example Results and Discussion Sensitivity analysis for Prediction performance comparison Strength and weakness of the proposed piece-wise regression method Concluding Remarks Acknowledgements 
work_wm5p5jfotjc4dml76zoeqhdwnu	----	doi:10.1016/j.eswa.2005.09.086 Neural networks for animal science applications: Two case studies C. Fernández a , E. Soria b,*, J.D. Martı́n b , A.J. Serrano b a Departmento. Producción Animal, Facultad de Ciencias Experimentales y de la Salud, Universidad Cardenal Herrera CEU, Edificio Seminario s/n. 46113 Moncada, Valencia, Spain b Digital Signal Processing Group, Department Enginyeria Electrònica, Escola Tècnica Superior d’Enginyeria, Universitat de València, C/ Dr. Moliner 50, 46100. Burjassot, Valencia, Spain Abstract Artificial neural networks have shown to be a powerful tool for system modelling in a wide range of applications. In this paper, we focus on neural network applications to intelligent data analysis in the field of animal science. Two classical applications of neural networks are proposed: time series prediction and clustering. The first task is related to the prediction of weekly milk production in goat flocks, which includes a knowledge discovery stage in order to analyse the relative relevance of the different variables. The second task is the clustering of goat flocks; it is used to analyse different livestock surveys by using self-organizing maps and the adaptive resonance theory, thus obtaining a qualitative knowledge from these surveys. Achieved results show the usefulness of neural networks in two animal science applications. q 2005 Elsevier Ltd. All rights reserved. Keywords: Neural networks; Multilayer perceptron; Self-organizing map; Animal science 1. Introduction Artificial neural networks (ANNs) have been applied to a huge amount of fields during last years (Arbib, 2003; Hu, 2003). However, there are some fields in which the use of ANNs is still scarce; animal science is one of these fields. This lack of ANN applications in animal science is quite paradoxical, as data analyses are usually carried out in this field, and ANNs have shown to be more powerful than classical statistical methods to carry out this kind of tasks (Arbib, 2003; Bishop, 1995; Ripley, 1996). Two examples are used in this work in order to show the ability of ANNs to solve animal science problems: a time series prediction (Makridakis, 1997; Weigend, 1993) and a clustering application (Theodoridis, 1999). The multilayer perceptron (MLP) (Bishop, 1995) is the neural model used for time series prediction, whereas the neural models used for clustering are the Kononen’s self-organizing map (SOM) (Kohonen, 2000) and a network based on the adaptive resonance theory (ART), the ART2 network (Fausset, 1994; Jang, 1996). The first problem tackled in this paper is related to milk yield (MY) weekly prediction by using current animal factors 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.09.086 * Corresponding author. E-mail address: emilio.soria@uv.es (E. Soria). (body weight, diet, litter size, number of lactation, etc.), as well as the current MY prediction (an autoregressive model). MY weekly prediction becomes more useful for management purposes than the analysis of the different factors which can affect MY at the end of lactation. The clustering application is completely different since a qualitative model is desired in order to extract useful knowledge from data. Goat production in Murcia Region (Spain) requires a deeper knowledge of their farming characteristics and practices. Knowledge of MY prediction is important because is the main income for farmers, but without a good feeding program is difficult to achieve their goals. The nature and availability of biomass of semiarid Spanish lands may limit the animal welfare and pro- ductivity, and thus, lower milk production may be achieved, as well as a decrease in kid survives and income for selling milk. Supplementation during the year with forages, by- product, concentrates and/or total mixed ration (TMR), is a common practice when pasture is scarce, as happens in Murcia Region. Ninety-four flocks from Murcia Region are used to analyse the relationship between food management practices. The herd food management practise is obtained by surveys to farmers. The remainder of the paper is outlined as follows. Section 2 shows the neural methods used in this study. Results achieved in different applications are presented in Section 3, ending up the paper with some conclusions in Section 4. Expert Systems with Applications 31 (2006) 444–450 www.elsevier.com/locate/eswa http://www.elsevier.com/locate/eswa Fig. 1. Schematic of a neuron model. C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450 445 2. Neural models 2.1. Multilayer perceptron A multilayer perceptron (MLP), is an ANN, formed by elementary processing units, the so-called neurons. A typical neuron model is shown in Fig. 1 (Haykin, 1999). The main components of the model are: Sum function. It carries out a linear combination of the neuron inputs through the use of a set of coefficients, known as synaptic weights. Being wZ[w0, w1,., wk] the vector of coefficients, and xZ[1, x1,., xk] the input vector, the sum function is given by the scalar product of both vectors. When using a neural model, the goal is to find the optimal synaptic weights to solve the problem. The process of searching the optimal weights is known as network learning. Activation function. It is a non-linear function which gives the network its non-linear nature. The most used activation functions are the sigmoid function (its values ranging between 0 and 1) and the hyperbolic tangent, which ranges between K1 and C1. As it can be inferred from Fig. 1, a neuron without an activation function is equivalent to a multi-variant analysis. Therefore, a non-linear combination should be more powerful than a multi-variant analysis (Ripley, 1996). Neurons are arranged in layers to form an MLP. The first layer is known as input layer, and the last one is called output layer. All the other layers are called hidden layers (Arbib, 2003). This kind of arrangement enables the neuron outputs to be used as inputs to neurons of following layers (non-recurrent network) and/or previous Fig. 2. Multilayer perceptron. Dotted line layers (recurrent networks). Fig. 2 shows a typical MLP structure. Several remarks should be made regarding the learning process (Arbib, 2003; Bishop, 1995; Haykin,1999): The learning rate must be chosen appropriately. The learning algorithm is based on finding the closest minimum, being this learning rate a parameter which measures the speed of approaching the minimum. The network architecture must also be chosen appropriately. While the number of neurons in the input layer and the output layer is given by the problem, the number of hidden layers, and the quantity of neurons in each hidden layer must be chosen depending on the particular problem. Due to the iterative nature of the learning process, it is necessary to choose the initial values of the synaptic weights. Different initial values are usually tested in order to achieve an optimal model. In order to obtain a model with validation capabilities, data is split into two data sets: a training set used to obtain the neural model, and a validation set used to test the behaviour of the model with data different from the training set used to obtain the model. Two different approaches can be carried out to train the network. The first approach is based on updating the synaptic weights for each pattern of the data set (on-line approach) whereas the other approach is based on updating the synaptic weights once for all the training data set (batch approach). 2.2. Self-organizing map (SOM) The self-organizing map is a neural network proposed by Teuvo Kohonen in 1984 (Haykin, 1999; Kohonen, 2000). Neurons are arranged in two layers: an input layer, formed by n neurons (one neuron for each input variable) in Fig. 3 and an output layer in which information is processed; this second layer is usually arranged in a two-dimensional structure. Neurons of the output layer are characterised by a weight vector with the same dimension as the input vector. For instance, neuron i,j (i-th row, j-th column) is characterised by the weight vector wij Z½w 1 ijw 2 ij:::::::::w n ij�. Similar input patterns are mapped close each other in the output layer (Kohonen, s stand for recurrent neural systems. Fig. 3. SOM architecture with n-dimensional inputs. C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450446 2000). Algorithm procedure can be summarized, as follows (Haykin, 1999; Kohonen, 2000): Weight initialisation. Choice of an input pattern xZ x1 x2:::::::::xn � � . Measurement of the similarity between weights and inputs. If the Euclidean distance is taken into account, then the similarity measure is given by dðwij; xÞZ PM kZ1 ðwkij KxkÞ 2 The most similar neuron to the input pattern is called best matching unit (BMU). Synaptic weights are updated as wstðnC1ÞZwstðnÞC a,hðBMU; wstÞ,ðxKwstÞ, where a(n) is the learning rate and h(n) is known as neighbourhood function. The value of this function depends on the distance between the BMU and the neuron to be updated, the closer the two neurons the higher the value of this function. Moreover, this function is the responsible of preserving the topological relationships among input patterns (Kohonen, 2000). The previous steps are performed a predetermined number of iterations. When this number is reached, the learning algorithm is stopped. While the number of iterations is lower than the predetermined value, go to step 2. Once the map training is finished, the visualisation of the two-dimensional map provides a qualitative information about how the input variables are related each other for the data set used to train the map. SOM is a visualisation tool rather than a clustering tool, although it is possible to obtain clusters of similar patterns from the two-dimensional map. Nevertheless, another model is proposed in order to carry out this clustering task, namely, the adaptive resonance theory (ART). 2.3. Adaptive resonance theory (ART) This model is based on analysing the similarity between input patterns and cluster prototypes. If a pattern is similar enough to a certain cluster prototype, then the prototype is updated in order to incorporate information from this pattern. If the similarity is not sufficient, then a new cluster is formed, whose prototype is given by this pattern. This model solves a usual problem of other neural models, namely, the corruption of the information already learnt by the contribution of new patterns (Arbib, 2003). In particular, we used the ART2 network which allows to use this theory with continuous- valued input vector (Fausset, 1994). Cluster prototypes have the same dimension as input vectors. ART2 procedure can be stated as follows (Kung, 1993): Weight initialisation. Given an input pattern x, the Euclidean distance between this pattern and every neuron is computed. Therefore, the distance between the j-th neuron and the input pattern is given by dj Zjjwj Kxjj. The most similar neuron to the input pattern is selected in order to analyse whether it is a good-enough representation of the input pattern. This neuron is denoted as v. A vigilance test is performed in order to test whether the prototype is a good enough representation of the pattern: jjwv,xjjOp where r is the vigilance parameter (0!r!1). If the vigilance test is not accomplished, then a new neuron is created, whose weights are given by the input pattern. If the vigilance test is accomplished, then the weight vector of the winner neuron is updated: wv Z x Cwv,Nv Nv C 1 where Nv is the number of data points in the winner node v so far. This means that if there are many points in the neighbourhood of the winner, then it is hard to move the winner. The vigilance parameter controls the desired degree of similarity between the weight vector and the input pattern; low values involve that a low similarity is needed, and therefore, a low number of clusters is produced. On the contrary, high values involve a high number of prototypes. 3. Practical applications 3.1. Milk yield prediction Milk yield (MY) is a key factor in the management of goat livestock since it becomes the main farmer income. MY is important to identify which goats are the best milk producers, and also to analyse the current state of a certain livestock. Therefore, MY prediction is very helpful as it enables the farmer to make decisions about livestock management before the end of lactation, when decisions are usually made. A homogenous group of dairy goats within a commercial farm (Excamur, S.L.) was selected in this trial. This farm is a member of ACRIMUR (Murciano-Granadina Goat Breeder Association—Asociación Española de Criadores de la Cabra Murciano-Granadina). The selected farm is a representative sample of these kinds of farms in the south-eastern Spain. The data set used in this work was formed by 36 goats. They were split into two data sets: a training set (26 goats) to obtain the prediction model, and a validation set (10 goats) in order to guarantee the generalisation capabilities of the network. Both data sets were similar with regard to the diet followed by goats Table 1 Mean values and standard deviations of the input variables for both training and validation sets Lactation number (*) Number of kids Control number Metabolic weight (kg3/4) Milk yield (kg) Training Mean 4.94 1.82 8.13 16.91 2.20 Std 1.37 0.58 3.52 1.25 0.80 Validation Mean 4.80 2.10 9.10 17.36 2.06 Std 1.25 0.54 3.42 1.63 0.64 * Lactation number indicates lactation period. Table 2 Root mean-square error (RMSE) obtained by the simple models as well as by the best neural models in MY prediction. Net(1) has an architecture 6!3!1 and online learning, Net(2) 6!3!1 and batch, Net(3) 6!3!3!1 and batch, and finally, Net(4) 6!4!4!1 and online Training set (kg) Validation set (kg) Model1 0.47 0.42 AR(1) 0.44 0.39 AR(2) 0.44 0.40 Net(1) 0.37 0.31 Net(2) 0.37 0.31 Net(3) 0.37 0.31 Net(4) 0.36 0.31 C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450 447 in order to avoid biased models. A weekly follow-up was carried out for every goat, being the study period of 22 weeks. The most important variables that could affect MY were taken into account (Gipson & Grossman, 1990). These variables can be easily obtained in a flock which is subject to a milk control process. The used variables were the following: number of lactation, number of kids, time between parturition and the first milk control, type of diet 1 , metabolic weight (Klieber, 1961), and current MY. Table 1 shows the mean values and the standard deviations values of the input variables; similar values for both data sets show the suitability of the data- splitting criterion, which guarantees unbiased models. Three simple models were proposed in order to have a fair reference for model comparison: A trivial model (Model1) which uses the following prediction rule: m(tC1)Zm(t), being m(k) the variable to be predicted (MY). Two autoregressive models: AR(1) which takes into account the previous value of the prediction variable, and AR(2), which considers the two previous values of this variable. With regard to neural models, different architectures and weight initialisations were proposed. This procedure is due to the learning algorithm used, the Levenberg-Marquardt algor- ithm, which is based on finding the closest minimum to the initial weights, so that, if this is a local minimum, then a non- optimal solution is achieved (Luenberger, 1984). This algorithm shows a better performance than other more widely used algorithms, such as, the classical backpropagation algorithm (BP). With regard to hidden nodes, if only one hidden layer was taken into account, the number of neurons ranged between 2 and 15, whereas if the two hidden layers were taken into account, the number of neurons in each layer ranged between 2 and 6, in order to avoid overfitting (Haykin, 1999). Moreover, the stopping criterion was based on cross- validation (Bishop, 1995; Haykin, 1999). A comparison of the results achieved by the different models is shown in Table 2. 1 In this work, two diets were used to fulfil nutritive requirements during lactation, being the difference between these two diets the ingredients of each one. Fig. 4 shows the MY predicted by the neural model Net(4) compared to the desired signal, for both training and validation sets. An accurate prediction is achieved in a global point of view. Nevertheless, it is also important to analyse whether the prediction is accurate for every particular goat; therefore, results for individual goats are also shown in Fig. 5. Fig. 5 shows that the prediction achieved by Net(4) is also accurate for individual goats, except for a small amount of goats; for instance, goat #24 in the training set and goat #7 in the validation set. 3.2. Qualitative analysis of goat farms by SOM Section 3.1 carries out a quantitative analysis to predict MY in goat farms. Nevertheless, qualitative analyses are also important since they are very helpful to understand how the models work, and thus, they can be used for knowledge discovery. Sections 3.2 and 3.3 show the use of clustering techniques for analysing the relevance of variables involved in this problem as well as to identify characteristic behaviours in the management of goat farms. Feeding becomes the most important factor to take into account when a new stockbreeding activity is initiated since it typically involves more than half of farming expenses (Haenlein, 2001). This is the reason why farmers filled in a survey on their feeding procedures in order to study feeding management. The usefulness of this information stems from the fact that it allows an overall estimation of farms, thus trying to amend mistakes for improving farm development and competitivity. Data used in this study are referred to 1999 and were obtained between 2001 and 2002 from surveys returned by 94 farm owners dedicated to goat production. The surveyed farms comprised approximately 15% of the total official census of Murciano-Granadina breed farms in Murcia Region. Eleven variables were studied. The first two variables were related to the geographical location in Murcia Region (6 counties (V1): 1-Altiplano, 2-Campos de Cartagena, 3-Noroeste, 4-Rı́o Mula, 5-Valle del Guadalentı́n, 6-Vega del Segura) and herd size (V2: number of goats). The other nine variables were related to nutrition management and milk income. Three variables were obtained asking farmers whether they supplement the diet for different physiological stages (replacement feeding V3: No(0) /Yes(1); lactation feeding V4: No(0)/Yes(1); gestation feeding Fig. 4. MY predicted by Net(4) and desired signal for both training and validation sets. C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450448 V5: No(0)/Yes(1)). The sixth variable was obtained by asking farmers whether they have any nutrition support by external personnel (V6: No(0)/Yes(1)). The variable type of feeding (V7) has three possible answers; diet elaborated by farmers (coded as 0), farmers who buy compound feed (coded as 1) and farmers who buy total mixed ration (coded as 2). The variable peak lactation feeding (V8) has three options as well; No (coded as 0), Yes (coded as 1) and animal for meat production (coded as 2). Body condition score (V9) is a subjective Fig. 5. MY prediction for the individual goats which form the training set and the va (MAE) are used as accuracy measures. measurement that farmers tend to use in order to evaluate whether nutrient supplementation appears to be necessary (No(0)/Yes(1)). Milk income (V10) (V/ goat) was evaluated for each farm, and finally the last variable taken into account was the use of Byproduct (V11: No(0)/Yes(1)) to feed animals. Our aim was to extract knowledge from these variables. SOM was applied to carry out this task since it preserves topological relationships among data while mapping data into a two-dimensional map. An SOM formed by 8!6 neurons with lidation set. The root-mean-square error (RMSE) and the Mean Absolute Error Fig. 6. Grey-scale representation of the components of the vectors which represent each neuron in the map. The darker the colour the lower the values of the components. C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450 449 hexagonal architecture (each neuron has six neighbours) was used. Data was normalised (mean zero and variance unity) in order to avoid unbiased models due to the different range of values of the variables. The batch mode was used to train the network and a Gaussian neighbourhood function was chosen. Fig. 6 shows the components of the vectors which represent each neuron in the map. There are several zones of the map that should be highlighted. For instance, those variables in which both ‘replacement feeding’ and ‘lactation feeding’ appear to be relevant, are mapped into the lower parts of the map in light colour. These zones of the maps also correspond, within variable ‘type of feeding’, to those farms using compound feed (grey colour and intermediate part of the map) and total mixed ration (light colour and lower part of the map). On the contrary, the dark area in the upper part of the map represents farmers who make their own diet. To sum up, the lower part of the map with light colour represents farmers which consider groups of goats within their herds, and these groups are fed according to the physiological stage (mainly replacement and lactation). Moreover, these farmers tend to buy to food companies. The light area is also present in the variable ‘nutrition support’, i.e., farmers who buy to food companies, are also supported about their technical decisions; moreover, it can also be appreciated that ‘milk income’ also shows light colours, which indicates Table 3 ART2 clustering. Cluster prototypes and percentage of patterns assigned to each cl Cluster V1 V2 V3 V4 V5 V6 1 2.12 185.85 0.97 0.79 0.33 0.52 2 3.43 124.07 0.13 0 0.13 0.33 3 4.85 251.58 1 1 0.81 0.69 4 5 53 0.33 0 0.33 0.33 5 1 1450 1 1 1 0 6 4 800 1 1 0 0 more income from milk yield. The use of feeding sub-products (variable ‘byproducts’) is biased to the right map of the map, including both farmers who buy to food companies and those who make their own rations. The analysis of variable ‘gestation feeding’ shows that this topic is still in an early development, even though its physiological relevance. In spite of farmers do carry out an individualised feeding of animals which are being milked, they do not take into account individual factors such as the four weeks of the ‘peak lactation feeding’ (maximum milk yield). Moreover, only a few farmers take into account the ‘body condition score’. 3.3. ART clustering of goat farms The main advantage of ART over other clustering algorithms stems from the fact that it does not need to know the number of clusters in advance, but it can find the natural number of clusters once chosen the vigilance parameter r (Theodoridis, 1999). ART2 models were trained, finding an optimal value of rZ0.98. Six clusters were found by the algorithm, whose prototypes are shown in Table 3. Table 3 shows that three clusters include almost all the patterns, namely, clusters #1, 2 and #4. Cluster #1 include 35% of the patterns which belong to the county Campos de Cartagena. Farms of this cluster show a medium size and also medium income from MY (308 V per goat and lactation). Some farmers produce their own feeding whereas others buy to food companies, thus taking advantage of the technical support provided by companies. They carry out a management by groups of goats, depending on their MY capabilities. Cluster #2 include 32% of the patterns which belong to the north-western area. This cluster includes smaller farms and their income is also lower than Cluster #1. These farms do not receive technical support, but they produce their own food and their management is the same for all the goats. Therefore, this cluster shows a management quite poorer than that of cluster #1, thus suggesting that an improvement should be taken into account for a considerable percentage of farms in Murcia Region. Cluster #4 is located in the county Valle del Guadalentı́n and it includes small farms. They do not show any income from milk sale because these farms are focused on meat instead of milk. Their management does not consider differences among goats, and therefore, this cluster represents the poorest management among farms of Murcia Region. uster V7 V8 V9 V10 V11 % Data 1.48 0.12 0.48 308.48 0.67 35.11 1.27 0.33 0.47 141.8 0.3 31.92 2.12 1 0.27 267.27 0.35 1.06 1.33 2 0 0 0.33 27.66 3 0 0 0 1 3.19 3 1 1 355 1 1.06 C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450450 The other clusters are not representative enough since they involve a very small percentage of data. 4. Conclusions Neural networks have been used in two animal science applications. Results have shown their suitability to be used in this field, in which the quality and accuracy of the models is essential to increase farm returns. With regard to MLP, achieved results have been better than those obtained with classical easy models in terms of accuracy (prediction improvement: 0.11 kg/goat/lactation). This involves a prediction improvement of 22 kg/lactation for a medium size herd formed by 200 goats, which may improve herd management. Moreover, neural networks have provided an estimation of farm locations, which could be used to avoid mistakes and improve their development and competitivity. With regard to SOM, maps have shown that farmers tend to carry out a management based on groups of goats, which has shown that feeding management of farms focused on dairy produce is more adequate than that of farms focused on meat production. This adequate management has also included technical support and it has involved an income increase from MY. These conclusions are similar to those obtained in previous works, which studied goat herds from other Spanish regions, such as Andalusia (Castel, Mena, Delgado-Pertiñez, Carmuñez, Basalto & Caravaca et al., 2003; Delgado-Pertiñez, Alcalde, Guzmán-Guerrero, Castel, Mena & Caravaca, 2003) or Catalonia (Milán, Arnalte, & Caja, 2003). With regard to ART, it has produced three relevant clusters which showed three kinds of herd management’s practices in Murcia Region. Basically, these clusters represented farms which were characterised by a policy based on dairy goats with supplementary feeding, dairy goats without supplementary feeding, and finally, meat production. The relevance of supplementary feeding has turned out to be very importance since profits were raised (between 140 V and 310 V per goat) References Arbib, M. (2003). The handbook of brain theory and neural networks. Cambridge, MA: MIT Press. Bishop, C. M. (1995). Neural networks for pattern recognition. Oxford: Clarendon Press. Castel, J. M., Mena, Y., Delgado-Pertiñez, M., Carmuñez, J., Basalto, J., Caravaca, F., et al. (2003). Charactarization of semi-extensive goat production systems in southern spain. Small Ruminant Research, 47, 133–143. Delgado-Pertiñez, M., Alcalde, M. J., Guzmán-Guerrero, J. L., Castel, J. M., Mena, Y., & Caravaca, F. (2003). Effect of hygiene-sanitary management on goat milk quality in semi-extensive systems in spain. Small Ruminant Research, 47, 51–61. Fausett, L. V. (1994). Fundamentals of neural networks. Englewood Cliffs, NJ: Prentice-Hall. Gipson, T. A., & Grossman, M. (1990). Lactation curves in dairy goats: A review. Small Ruminant Research, 3, 383. Haenlein, G. F. W. (2001). Past, present and future perspectives of small ruminant dairy research. Journal of Dairy Science, 84, 2097–2115. Haykin, S. (1999). Neural networks: A comprehensive foundation. Englewood Cliffs, NJ: Prentice Hall. Hu, Y. H., Hwang, J. N., & Hwang, J. N. (2003). Handbook of neural network signal processing. London: Academic Press. Jang, J. S., Sun, C. T., & Mizutani, E. (1996). Neuro-fuzzy and soft computing: A computational approach to learning and machine intelligence. Engle- wood Cliffs, NJ: Prentice Hall. Klieber, M. (1961). The fire of life (an introduction to animal energetics). New York: Wiley. Kohonen (2000). Self-organizing maps. New York: Springer. Kung, S. Y. (1993). Digital neural networks. Englewood Cliffs, NJ: Prentice Hall. Luenberger, D. G. (1984). Linear and nonlinear programming. Reading, MA: Addison-Wesley. Makridakis, S. G., Wheelwright, S. C., & Hyndman, R. J. (1997). Forecasting: Methods and applications. London: Wiley. Milán, M. J., Arnalte, A., & Caja, G. (2003). Economic profitability and typology of Ripollesa breed sheep farms in Spain. Small Ruminant Research, 49, 97–105. Ripley, B. D. (1996). Pattern necognition and neural networks. Cambridge: Cambridge University Press. Theodoridis, S., & Koutroumbas, K. (1999). Pattern recognition. New York: Academic Press. Weigend, A. S., & Gershenfeld, N. A. (1993). Time series prediction: Forecasting the future and understanding. Proceedings of the NATO advanced research workshop on comparative time series. Reading, MA: Addison Wesley. Neural networks for animal science applications: Two case studies Introduction Neural models Multilayer perceptron Self-organizing map (SOM) Adaptive resonance theory (ART) Practical applications Milk yield prediction Qualitative analysis of goat farms by SOM ART clustering of goat farms Conclusions References 
work_wmgsfmulknbrham4pkvdecjybi	----	1 Intelligent Product Search with Soft-Boundary Preference Relaxation Maciej Dabrowski (corresponding author) maciej.dabrowski@deri.org Digital Enterprise Research Institute National University of Ireland Galway Galway, Ireland Phone: +353 91 495138 Fax: +353 91 495541 Thomas Acton thomas.acton@nuigalway.ie J.E. Cairnes School of Business & Economics National University of Ireland Galway Galway, Ireland Phone: +353 91 493806 Fax: +353 91 494565 Hans van der Heijden h.vanderheijden@surrey.ac.uk Surrey Business School University of Surrey Guildford, Surrey, GU2 7XH, United Kingdom. Phone: +44 1483 686302 Fax: +44 1483 686306 2 Abstract. This paper proposes a novel method for preference relaxation in online product search, which enables consumers to make quality choices without suffering from the commonly experienced information overload. In online shopping scenarios that involve multi-attribute choice tasks, it can be difficult for consumers to process the vast amounts of information available and to make satisfactory buying decisions. In such situations consumers are likely to eliminate potentially good choices early on, using hard-constraint filtering tools. Our approach uses edge sets to identify the alternatives on the soft bounda- ry and the principle of alternative domination to suppress the alternatives on this bounda- ry that are irrelevant. We demonstrate how our approach outperforms existing methods for product search in a set of simulations using two sets of 2650 car advertisements and 1813 digital cameras gathered from a popular online store. Keywords: Decision Theory, Recommender Systems, Preference Relaxation, Electronic Commerce Abbreviations: NR – No Relaxation, SR – Standard Relaxation, SBR – Soft-Boundary Preference Relaxation, SBRADD – Soft-Boundary Preference Relaxation with Addition, SBRREP – Soft-Boundary Preference Relaxation with Replacement 3 1 Introduction In this paper we present a novel approach to improve customer product search in online shopping scenarios. Drawing upon existing work in information filtration, recommender systems and decision theory, we detail a new method for preference relaxation, referred to as Soft Boundary Preference Relaxation, which improves customer choice quality. We present a detailed description of the method, and discuss its performance based on a set of simulations in two product domains using actual datasets: digital cameras and used cars. Our method explicitly addresses the potential for any increased information overload when additional, recommended alternatives are presented to customers. Consumers seeking a suitable product online often face a choice from amongst a large set of options. Choosing a car to buy or a plane ticket to book using popular e-commerce websites often requires searching through a very large list of products that can be difficult to grasp or compare. To address this issue, many e-commerce sites offer product filtration tools, which assist customers typically by asking them to fill a form to gather preferences. The process of searching online product catalogues to locate the product(s) that best match consumer’s product needs is often referred to as preference-based search (Viappiani, Pu, & Faltings, 2008) or parametric search (Burke, 2002). Considering the potential for a large number of products in contemporary online product catalogues, it may be necessary for consumers to iteratively refine their preferences to arrive to product lists of a manageable size: this requires effort and can be very frustrating (Hagen, Man- ning, & Paul, 2000). Further, consumers often construct or adjust their preferences whilst interacting with the product catalogue, and may need extra support, for example through suggestions of interesting products (Viappiani, et al., 2008). Thus, the main challenge for online product catalogues is in providing support to consumers to find the most valuable products that match their needs (Zhang & Jiao, 2007), typically through more effective product filtration, recommendation, and preference elicitation. Typically, preference elicitation processes involve presenting a form-type interface where a user can input his or her preferences. Such an approach to product search, also referred to as Navigation-by-Asking (Shimazu, 2002), is widely used in e-commerce settings, but is often inefficient (Viappiani, Faltings, & Pu, 2006), especially when consumers are not fully aware of available products or no suitable offers are returned in a search result set. According to behavioural decision theory (Payne, Bettman, & Johnson, 1993; Slovic, 1995), an awareness of available products leads to the construction of personal prefer- ences. Faced with a shopping decision, consumers tend to construct their preferences when they are prompted to express evaluative judgment (Payne, Bettman, & Johnson, 4 1992). However, form-filling approaches do not sufficiently guide consumers to find the most suitable products (Viappiani, et al., 2006). Further, such approaches work under the assumption that consumers can accurately state their requirements, or more formally, lev- els within a product attribute that are acceptable or not (Klenosky & Perkins, 1992). In addition, form-based product search tools tend to use a non-compensatory approach in the evaluation of available alternatives in which all products that possess at least one at- tribute with unacceptable values are rejected from further consideration (Oppewal & Klabbers, 2002). This approach is also referred to as logical product filtration with no re- laxation (NR). When stated product requirements are over-specified, the search result presented in a response to such a failing query (Mirzadeh & Ricci, 2007) may be empty, causing confusion. Indeed, previous research indicates that form-based approaches may lead to inaccurate product choices as consumers often fail to reject products with attribute levels, which they themselves had previously described as unacceptable (Klein, 1987). Therefore, there is a need for tools that can provide online consumers with lists of products that not neces- sarily fully match their stated preferences – so called suggestions or recommendations. Numerous studies propose the use of recommendations to improve consumer decision- making performance (Bridge & Ricci, 2007; Häubl & Murray, 2001; Kim, Kim, & Cho, 2008; Pu, Chen, & Kumar, 2008). Providing a consumer with a relevant yet diverse set of suggestions has become an important research problem (Smyth & McClave, 2001). The tools that implement various methods for nominating product suggestions to improve consumer’s performance in online product search are referred to as recommendation agents (RAs) (Resnick & Varian, 1997; Xiao & Benbasat, 2007) representing content and/or collaborative recommender systems (Adomavicius & Tuzhilin, 2005), utility- based tools, and preference relaxation methods (Mirzadeh & Ricci, 2007). This research argues that the use of preference relaxation is an effective approach for the identification of product recommendations. Classical approaches to preference relaxation (Mirzadeh & Ricci, 2007; Mirzadeh, Ricci, & Bansal, 2004), referred to here as Standard Preference Relaxation (see Section 3.1) may increase consumers’ decision-making per- formance through positive effects on decision quality. Nevertheless, the major disad- vantage of the Standard Preference Relaxation approach is an increase in decision effort (see section 4). Directly addressing this drawback, this paper proposes a novel Soft- Boundary Preference Relaxation method in two variants (see Section 3.3 for details) and examines its impact on consumers’ decision-making performance in the context of pref- erence-based search in online shops. In the next section we provide a brief overview of the literature on preference-based search and recommender systems in so far as necessary to document our new method of 5 soft boundary searching. We then present an overview of the research design and results of an application of the method across two choice scenarios using real-world data sets. We conclude with a discussion of our findings, and implications for future research. 2 Background Information Filtering techniques typically perform a progressive removal of non-relevant content based on the information in a user profile acquired either in an implicit (e.g. stud- ying user behaviour) or an explicit (e.g. asking user to state his preferences) manner. These techniques provide a theoretical foundation for building Recommender Systems (Resnick & Varian, 1997) that enable content personalization - an important stream of re- search in e-commerce (Lee, Liu, & Lu, 2002). Numerous studies (Bridge & Ricci, 2007; Viappiani, et al., 2008) use recommendations to improve consumer decision-making (Bridge & Ricci, 2007; McGinty & Smyth, 2003; Viappiani, et al., 2008). Nevertheless, it is crucial (Gretzel & Fesenmaier, 2006; Smyth & McClave, 2001) that the product suggestions provided to consumers are relevant (similar to their stated preferences) yet diverse (so that they can discover new opportunities and adjust their preference model) According to the Look-ahead principle (Viappiani, et al., 2008), “suggestions should not be optimal under the current preference model, but should provide high likelihood of optimality when an additional preference is stated”. Further, dynamism in user preferences (Cao, Chen, Yang, & Xiong, 2009) is a problem recog- nized in Recommender Systems research. Assumptions that the decision maker can accurately state (and indeed bound) which lev- els within an attribute are acceptable versus unacceptable is fundamental to a self- explicated approach (Klenosky & Perkins, 1992). Decision makers often use a conjunc- tive evaluation of available alternatives in which all the alternatives that possess at least one attribute with unacceptable values are rejected from further consideration. Product search and filtering mechanism offered online adhere to that approach, and filter out all products that do not fully conform to the stated requirements. However, previous research indicates that decision makers tend to fail to fully adhere to self-explication. Klein (1987) found that decision makers often fail to reject alternatives with attribute levels which they themselves had previously described as unacceptable, and showed that significant num- bers of participants can choose an alternative described with at least one attribute level they initially indicated as “completely unacceptable”. Preference relaxation mechanisms may assist in alleviating this problem. A decision aid implementing preference relaxation can potentially improve consumer decisions in online shopping websites, in a similar way 6 as other recommendation systems such as the one proposed by Julià, Sappa, Lumbreras, Serrat, & López (2009) or Wang and Benbasat (2007). The rigidity of typical preference elicitation (filtering) mechanisms is a well-established problem (Chaudhuri, 1990) that can potentially lead to the elimination of all available products from consideration. Indeed, over-specification of consumer requirements lead- ing to an empty result sets motivated research on similarity-based retrieval (McSherry, 2004) and query relaxation methodologies (Mirzadeh & Ricci, 2007). Similarity based recommenders (Bridge, Goker, McGinty, & Smyth, 2005; Stahl, 2006), although a very popular approach in e-commerce, suffer from a number of disadvantages. First, products that are most similar to the preferences stated by a customer, are not always optimal sug- gestions, as a slightly less similar product may be objectively much more attractive to a consumer (for example, a PC notebook with a new type of battery that is much lighter and provides more up-time than expected by a consumer while stating the preferences). Second, many similarity-based recommender systems often suggest a predefined number of top recommendations, leading only to a limited increase in consumer awareness. On the other hand, the process of filtering involves the application of filtering rules (or restrictions on attributes) to the items in the set to be filtered (Chaudhuri, 1990; Mirzadeh & Ricci, 2007). Consumer preferences are the key input for alternative pre-filtration, as only alternatives that fully satisfy all provided preferences are presented to the user as a result to his or her query. To address this, Mirzadeh and Ricci proposed a mechanism for preference relaxation for queries producing an empty result set (Mirzadeh & Ricci, 2007). However, they do not investigate the impact of the extent of relaxation on decision maker behaviour, and their method is applicable primarily to failing queries. Our research differs from these approaches. First, we primarily focus on reduction of the number of erroneously filtered-out products by intelligent relaxation of the preferences provided by a consumer. Second, many of the existing methods for computation of prod- uct suggestions require prior knowledge or history of user interactions and preference models, which are not required in our approach. Therefore, the preference relaxation ap- proach proposed in this paper is easily applicable to any ecommerce platform that utilizes form-based product filtration mechanism. Our approach increases the average quality of result sets presented to a user after product filtration, leads to increased product knowledge, more accurate preference models, and better decision-making. 7 3 Preference Relaxation methods Presume you intend to buy a car priced between $7000 and $8000 with reasonable mile- age (25000 to 75000 km). Would you be willing to pay slightly more ($8100) for a car with mileage lower than you expected (11000 km)? The ability to locate cars with such attribute values which, albeit out of the boundary ranges specified, may provide consum- ers with a better awareness of possible choices. The method proposed here enables consumers to consider products that would ordinarily be eliminated early in the selection process by falling outside “rigid” preferences. In the subsections below we discuss our approach in more detail and contrast it with common simple preference relaxation methods. 3.1 Standard Preference Relaxation Typically, preferences on numerical attributes are expressed using value ranges. As such, we allow the decision maker to specify his or her attribute value range preference for an i th attribute as di = [dL, dU] where dL dU indicates the lowest (highest) acceptable value for a given attribute. We now introduce “softening” variables eU (upper) and eL (lower), and a relaxation factor  (where ei =  * di), which enhance the filtering rule (value range) built based on attrib- ute value preference p causing the filtering rule to be less restrictive. The alternatives that satisfy the less strict preference d* = [dL – eL, dU + eU] remain in the set and can be con- sidered by the consumer. However, this approach, referred to as Simple Preference Relaxation (SR), often leads to a very large number of alternatives presented to the user, resulting in information over- load and increasing decision effort (Turetken & Sharda, 2004). In order to prevent these negative effects we use an approach based on the concept of an Edge Set (ES). Below, we show how such an approach can be used to create a subset of potentially interesting items for suggestion to the consumer. 3.2 Soft Boundary Preference Relaxation An Edge Set is a set of alternatives that fall into a value range based on an initial consum- er value preference for a given attribute (see Figure 1). For every preference value range, two edge sets can be constructed (lower and upper): ESLOWER = [dL – eL, dL + eL] and ESUPPER = [dU – eU, dU+ eU]. We explain this concept using price range preference pPRICE = [$6000, $7000]. For example, assuming variables eU = 350 and eL = 300 (for  = 0.05) an edge set can be constructed: ES = [$5700, $6300]   [$6650, $7350]. Thus the ES will 8 contain items that fall into the [$5700, $6300] or [$6650, $7350] price range. More spe- cifically, our approach involves three steps. First, we create edge set based on provided interval boundaries. Figure 1 An example of Edge Set for a [$6000, $7000] price preference. The inclusion of all alternatives satisfying the relaxed criteria would ordinarily increase the number of items presented to the user, contributing to information overload. To ad- dress this issue we incorporate a selection mechanism into our relaxation method that in- cludes only some of those cases (see Algorithm 1). Following the creation of the edge set based on relaxed preferences using a selected  (e.g. 0.05) the algorithm identifies the subset of all non-dominated alternatives (also referred to as the skyline (Borzsonyi, Kossmann, & Stocker, 2001)) that are part of this set. An item is non-dominated if no other item is better for any preference on attribute without being worse for at least one preference on other attributes (Häubl & Trifts, 2000). If a non-dominated item is a mem- ber of an edge set and it does not satisfy the non-relaxed initial DM preferences (is not a member of ResultSetNR) it is added to the set of suggestions, as it may be considered val- uable. We define two methods for inclusion of suggestions in the result set presented to a consumer. First, we propose to add suggestions to an initial result set constructed using a non-relaxed (NR) query. This method, further referred to as SBRADD (Soft Boundary Preference Relaxation with addition), may lead to increases in the size of result sets. To address this drawback for the Soft Boundary Preference Relaxation with Addition ap- proach, and to prevent an increase in cognitive load, we propose an alternative method. Instead of simple addition to the set, the method would replace dominated, low-utility items from a non-relaxed result set (ResultSetNR) that belong to the EdgeSet, with high- utility alternatives. We refer to this method as SBRREP (Soft Boundary Preference Re- 9 laxation with Replacement). With this approach, the total size of the set is kept constant, and the alternatives with lowest utility according to current preference model (in this study we use the WADD model) are substituted with items from the skyline. Collective- ly, we refer to these two mechanisms as SBR (Soft Boundary Preference Relaxation). Algorithm 1 The Soft Boundary Preference Relaxation Mechanism As indicated earlier, our method assumes variables eU (upper) and eL (lower), and a relax- ation factor , which relax the value preference p. Selecting an appropriate value of  is not trivial, as it resembles closeness (similarity of values) and can differ among consum- ers (Motro, 1990). However, some studies (Bosc, Hadjali, & Pivert, 2009; A. Hadjali, D. Dubois, & H. Prade, 2003) report that the maximum relaxation value max should not be greater than (3-5)/2, that is, ~0.382. Thus, the relaxation factor  should be selected from the interval [0, 0.382] to satisfy the concept of closeness (Allel Hadjali, Didier Du- bois, & Henri Prade, 2003). Although Mirzadeh and Ricci (2007) report that relaxation parameters are attribute-dependent and should be tuned according to consumer sensitivity to changes in that feature, in our study we implemented the former simpler relaxation ap- proach to explore potential effects in the first instance, with a view towards possible ex- pansion of parameters in future work. Although our approach is applicable to all types of attributes (continuous, categorical, or binary), in this study we investigate the methods that use continuous numerical attributes as, commensurate with the literature (Mirzadeh & Ricci, 2007), relaxation of binary and nominal constraints is trivial, as they are typical- ly discarded during the relaxation process. Indeed, relaxation of categorical attributes can be performed based on learned similarity measures, as demonstrated by Stahl (2002, 2006). 10 4 Hypotheses The simulations explored the effect of preference relaxation on decision performance in the context of multi-attribute preferential choice problems. Four methods were investi- gated: logical filtering with no relaxation (NR), Standard Preference Relaxation (SR), Soft-Boundary Preference Relaxation with Addition (SBRADD) and with Replacement (SBRREP). Effects of these methods on consumer decision-making performance were as- sessed through their impact on a number of indicators. Many dependent variables have been proposed as good indicators of the impact of a rec- ommendation agent on consumer decision performance (Hostler, Yoon, & Guimaraes, 2005; Parra & Ruiz, 2009; Payne, et al., 1993; Xiao & Benbasat, 2007). This study con- centrates on three common performance indicators: decision quality, decision effort and product awareness. Decision quality and effort are common indicators of objective deci- sion performance in the information systems literature (Acton, 2007; Häubl & Trifts, 2000; Parra & Ruiz, 2009), and product awareness is a widely used indicator of recom- mender method performance that facilitates higher quality decisions (Chen & Pu, 2007; McGinty & Smyth, 2003; Smyth & McClave, 2001). The impact of recommendation agents on decision quality has been investigated in many studies. For example, Pereira (2001) observed that query-based RAs improved both ob- jective and subjective indicators of decision quality. Other studies (Häubl & Trifts, 2000; Hostler, et al., 2005) demonstrated an assessment of decision quality through non- dominance (Pareto optimality) of the selected alternative(s), also referred to as an ''ideal selection'' (Häubl & Trifts, 2000). Indeed, such conceptualization of decision quality has been used as a measure of decision performance (Hostler, et al., 2005). Häubl and col- leagues (Häubl & Murray, 2005; Häubl & Trifts, 2000) showed that the use of RA in- creases decision quality by reducing the likelihood of selecting a non-dominated alterna- tive. Similar results were obtained by (van der Heijden, 2006) , who showed that the use of RA leads to a higher number of non-dominated alternatives in the consideration set. It follows that the use of preference relaxation methods should lead to increased decision quality in comparison to logical filtering, and so we hypothesize: [H1]: Standard Preference Relaxation has a positive effect on decision quality. Similarly, Soft-Boundary Preference Relaxation methods should lead to higher quality decisions by facilitating the consideration of a larger number of non-dominated alterna- tives. Consequently, compared to non-relaxing methods, Soft-Boundary Preference Re- laxation should positively impact decision quality, and we hypothesize: 11 [H2]: Soft Boundary Preference Relaxation has a positive effect on decision quality. The level of effort required to make a decision is another common decision performance indicator (Payne, et al., 1993). Effort is directly related to the amount of information that needs to be considered by a user (Eppler & Mengis, 2004; Turetken & Sharda, 2004). Cognitive load imposed on the customer to find the best offer among the list of products presented to him is considered an effort-related evaluation criterion for product selection tasks (Branting, 2002). Intuitively, preference relaxation mechanisms increase effort by relaxing rigid requirements, and therefore incorporating more alternatives for considera- tion by a user. Therefore, in contrast to logical filtering, the SR method should lead to significantly larger result sets, leading to higher choice effort: [H3]: Standard Preference Relaxation has a negative effect on decision-making effort. On the other hand, Soft-Boundary Preference Relaxation methods select only a small subset of products that satisfy relaxed preferences that are non-dominated. Therefore, no impact on decision effort in comparison to No Relaxation method should be observed: [H4]: Soft Boundary Preference Relaxation has no effect on decision-making effort. The selection of a product is context dependent, as the relative value of an option depends not only on the characteristics of that option, but also upon characteristics of other op- tions in the choice set (Payne, Bettman, & Schkade, 1999). According to behavioral deci- sion theory (Payne, et al., 1993; Tversky, 1996) the existence of such context impacts the perceived quality of available products and also leads to construction of new preferences. Tversky (1996) pointed out that increased awareness of product options causes users to adjust their initial preferences based on available choices, in contrast to maximizing fit to pre-computed preferences. This belief has support in the recommender systems literature, where the diversity of the set of recommended products is a sought after characteristic (Bridge & Ferguson, 2002). Indeed, Bodapati (2008) showed that product awareness is a necessary condition for actual purchase. Further, according to Fleder and Hosanger (2007), recommendation agents impact sales diversity. Intuitively, preference relaxation allows for the consideration of products that are filtered out when using logical filtering approaches. Consequently, we argue that preference relaxation mechanisms will increase consumers’ awareness of existing products, in contrast to the non-relaxed approach, indi- cated by result set diversity: [H5]: Standard Preference Relaxation has a positive effect on product awareness. [H6]: Soft Boundary Preference Relaxation has a positive effect on product awareness. These research hypotheses were investigated using software simulations (see Section 5.2) using two datasets, described in the next section. 12 5 Evaluation 5.1 Datasets In order to evaluate the proposed decision aid and compare it with the standard prefer- ence relaxation mechanism we performed a set of simulations using two real-world da- tasets. We collected a set of 2650 used car advertisements from Autotrader, and extracted 1813 digital cameras offered by Amazon ‘ s e-Commerce website. 5.1.1 Used cars The first dataset comprised 2650 used car advertisements collected from an online car website, run and managed by the Autotrader media group. Additional attributes for used cars in the set that were not present in advertisements, such as reliability (see Section 5.2), were automatically generated using standard information retrieval methods based on product reviews collected from car review websites (e.g. whatcar.com). Generated attrib- utes were classified as benefit-type and given scores ranging from 0 to 5 to resemble star ratings (e.g. 5 points for maintenanceCost describes the relatively lowest maintenance cost). 5.1.2 Digital Cameras The second dataset consisted of 1813 digital cameras extracted from the “Point & Shoot Digital Cameras” category on Amazon.com. The products were extracted using Java software and Web Services API provided by Amazon. Information on a number of attrib- utes was collected for each product: brand, model, price, zoom, screen size, resolution, and weight. Further, customer ratings on each product were extracted. These attributes were manually classified into cost-type (e.g. price) and benefit-type (e.g. resolution) cat- egories. 5.2 Method The simulation design was based on a leave-one-out (LOV) (McSherry, 2004) approach in which each alternative was temporarily removed from the dataset and its description was used as a user preference. The leave-one-out method, also known as n-fold cross val- idation (Kohavi, 1995), is a method commonly used in evaluation of recommender sys- tems. In LOV approach each of available alternatives is temporarily removed from the list of offers and is treated as a consumer’s product preferences/requirements. Consistent with bounded rationality (Lipman, 1995), and based on previous studies in relevant product domains: used cars (Dabrowski, Jarzebowski, Acton, & O'Riain, 2010) 13 and digital cameras (Aciar, Zhang, Simoff, & Debenham, 2007; Pu, et al., 2008), 6 most popular attributes were chosen for this experiment. To best resemble user behaviour, the preferences in the simulations were constructed similarly to filtering interfaces of popular websites, where value preference intervals were selected to simulate possible user entries. For example, the price intervals for used cars (in $) were: [1000, 2000] ... [5000, 6000], [6000, 8000] … [12000, 14000], [14000, 18000], and [18000, 30000] following the val- ues available for price filtration in a popular used car search engine. Using the LOV ap- proach, every used car advert in the set was temporarily removed from the set and its val- ues were used to create preference values (based on available preference intervals). For example, a car at $5900 would be represented as a user search query with price pref- erence [$5000, $6000], following the intervals defined based on existing product search engines available. Simulations were run for combinations of 1 to 6 stated preferences and for relaxation factors 0.05, 0.1, 0.2, 0.3 and 0.382 (max). Thus, for every set of parame- ters a maximum number of 2650 non-failing relaxed queries were issued and relevant re- sult sets were constructed for all four methods investigated: logical filtering with no re- laxation (NR), Standard Preference Relaxation (SR), Soft Boundary Preference Relaxa- tion with Addition (SBRADD), and with Replacement (SBRREP). Particular characteristics of these constructed result sets (see the next section) were assessed and compared to evaluate the methods under investigation. 5.3 Measures In this experiment a number of measures were used to evaluate the four methods. This section discusses the rationale for these measures. In the evaluation of customer choice-based decision outcomes this study investigated quality-related aspects of decision-making performance as suggested by Xiao and Benba- sat (2007). Two measures of objective decision quality are prevalent in the literature: non-dominance of the selected alternative(s) (Grether & Plott, 1979; Häubl & Trifts, 2000; Johnson & Payne, 1985), and product fit to customer preferences (Pereira, 2001), also conceptualized as utility (Johnson & Payne, 1985). Share of non-dominated alternatives. Häubl et al (2000) showed that the share of con- sidered products that are non-dominated indicates the quality of a set of products consid- ered by a consumer, which positively impacts decision quality. Such conceptualization of decision quality has been used as a measure of decision performance (Hostler, et al., 2005). For example, Swamnathan (2003) in the study of web-based RA utilized the pur- chase of non-dominated alternative as a direct indicator of decision quality. Similar con- ceptualization was used in the studies performed by Olson and Widing (2002). Finally, 14 van der Heijden {, 2006 #463} suggested the use of the share on non-dominated alterna- tives in the consideration set in this context. Therefore, this study utilizes share of non- dominated alternatives in the result set as a measure of decision quality. Average utility of products. As noted above, decision quality is directly related to ful- filling particular user’s criteria for product selection (preferences) that can be measured by the utility of selected alternatives (Bridge & Ricci, 2007). Some studies indicate that the goal of a recommendation agent is to locate products that closely match a user’s pref- erences (Hostler, et al., 2005; Lee, et al., 2002). Indeed Johnson and Payne (1985) sug- gest the maximization of product utility value as a measure of decision quality. Similarly, Stahl (2004) pointed out the importance of product utility in recommendation systems. The higher the average utility of alternatives presented for choice (AvgUtil  [0,1]), the more suitable options can be considered. Thus, this study measures decision quality with the average utility of products in a search result set constructed using the methods de- scribed in Section Error! Reference source not found.. Size of a result set. In terms of the decision process, information overload is an im- portant factor that increases decision-making effort and leads to changes in strategies em- ployed by decision makers in selection tasks (Eppler & Mengis, 2004). Some studies measure decision effort by examining the number of products about which detailed in- formation are obtained (Häubl & Trifts, 2000). Following the approach by Payne et al. (1999), our study measures the level of decision-making effort by the number of alterna- tives presented for consideration by a user, that is, the size of a result set. Diversity of a result set. Vahidov and Ji (2005) indicated the importance of result set di- versity in decision making. Tversky (1996) demonstrated that increased awareness of product options causes consumers to adjust their initial preferences based on available choices, in contrast to maximizing fit to pre-computed preferences. Recommender Sys- tems research supports this claim through design of methods that suggest diverse sets of recommended products to consumers (Bridge & Ferguson, 2002). Indeed, Bodapati (2008) showed that product awareness is a necessary condition for actual purchase. To assess diversity of result sets our study uses a common conceptualization of normalized diversity (diversity(ResultSet)  [0,1]) that is inversely proportional to similarity follow- ing the relation presented by Smyth and McClave (2001). The computation of similarity is performed using a law proposed by Shepard (1987), which suggests that perceived sim- ilarity of items is related to their distance via an exponential function: sim(a 1 ,a 2 ) = e -dist(a1 ,a2 ) where dist(a1,a2) is a normalised distance measure: 15 a1 ,a2 ÎA " 0 £ dist(a1,a2 ) £ 1 Measures were assessed for the results sets generated using the four methods under study. The share of the non-dominated alternatives in the result set and average utility of prod- ucts in the result set were used as decision quality indicators. Further, the number of al- ternatives in the result set indicated the decision effort necessary to make a decision, and the diversity of products in the result set indicated product awareness. The next section describes the results of the simulation. 5.4 Results – used cars Related samples non-parametric tests were used to investigate the average share of non- dominated alternatives in the result sets for queries using the preference relaxing mecha- nisms discussed above, compared with the share for the non-relaxed method. Results show that on average, result sets constructed using relaxation contained similar share of non-dominated alternatives to the result sets constructed using no relaxation. In particu- lar, on average 19.66% of non-dominated alternatives were observed when no relaxation was used, in contrast to only 19.55% in case of non-relaxing methods (NR) (see Table 1), representing a 0.6% increase. A larger increase was observed for the average utility of al- ternatives in a result set, with a 13.6% improvement (AvgUtilNR = 0.5161 and AvgUtilSR = 0.4543). These differences were statistically significant (p < 0.001), thus providing par- tial support for hypothesis H1. Similarly, results indicate that the use of the Soft Boundary Preference Relaxation mech- anism improves the share of non-dominated alternatives in a result set in contrast to both non-relaxing (NR) and Standard Relaxation (SR) methods. 28.95% (SBRADD), and 34.46% (SBRREP) of non-dominated alternatives were observed for SBR methods in con- trast to 19.66% (SR) and 19.55% (NR). These improvements (48.08% for SBRADD and 76.27% for SBRREP in contrasts to NR) were statistically significant (p < 0.001). Similar- ly, the average utility of alternatives in a search result set was higher for all preference- relaxing methods, with 0.5161 (SR), 0.5228 (SBRADD), and 0.5311 (SBRREP) in contrast to 0.4543 for the non-relaxed case. These values represent 13.6%, 15.08%, and 16.9% improvements respectively. The level of improvement in the average share of non- dominated alternatives and in the average utility of the alternatives in a search result set leads us to accept H2. The second group of hypotheses (H3 and H4) examined the decision-making effort measured by a number of items from which customers had to choose. For the methods investigated, median values of 631 (SR), 440 (SBRADD), and 420 (SBRREP) items in the result set were observed, in contrast to 420 items on average in a result set for non- 16 relaxed queries (NR). These differences were statistically significant (p < 0.001), con- firming H3 and indicating conditional acceptance of H4, as although SBRREP does not in- crease the decision-making effort, the SBRADD method forces making the decision based on 4.8% larger result sets.  0.05 0.1 0.2 0.3 0.382 Overall AvgUtil NR 0.4543 0.4543 0.4543 0.4543 0.4543 0.4543 SR 0.4875 0.5009 0.5175 0.5304 0.5443 0.5161 SBRADD 0.4849 0.5033 0.5290 0.5431 0.5538 0.5228 SBRREP 0.4942 0.5129 0.5384 0.5501 0.5600 0.5311 %ND NR 19.55% 19.55% 19.55% 19.55% 19.55% 19.55% SR 19.22% 19.68% 19.72% 19.87% 19.79% 19.66% SBRADD 24.82% 26.48% 28.83% 31.94% 32.65% 28.95% SBRREP 27.42% 29.96% 33.89% 39.78% 41.24% 34.46% |ResultSet| NR 420 420 420 420 420 420 SR 534 581 639 760 795 631 SBRADD 432 435 440 464 469 440 SBRREP 420 420 420 420 420 420 Diversity NR 0.0539 0.0539 0.0539 0.0539 0.0539 0.0539 SR 0.0559 0.0563 0.0577 0.0591 0.0597 0.0576 SBRADD 0.0550 0.0551 0.0563 0.0571 0.0574 0.0560 SBRREP 0.0541 0.0541 0.0550 0.0569 0.0569 0.0551 Table 1 Average utility (AvgUtil), share of non-dominated alternatives in the result set (%ND), median of the result set size (|RS|), and average diversity for relaxed (SR), non- relaxed (NR), SBR with addition (SBRADD) and replacement (SBRREP) for different values of  (used cars dataset). Finally, results indicated that the diversity of alternatives generated using preference re- laxation methods are more diverse than when no relaxation is used. In particular, an aver- age diversity of 0.1254 (SBRADD), 0.1248 (SBRREP, and 0.1319 (SR) was observed for preference relaxation methods, in contrast to 0.1131 for the non-relaxing method (NR) (see Table 2). These values represent 10.9%, 10.34%, and 16.6% respective improve- ments. These differences were statistically significant (p < 0.001), confirming both H5 and H6. 17 N 1 2 3 4 5 6 Overall AvgUtil NR 0.4386 0.4483 0.4550 0.4606 0.4649 0.4677 0.4543 SR 0.4907 0.5050 0.5164 0.5270 0.5366 0.5451 0.5161 SBRADD 0.5130 0.5195 0.5231 0.5265 0.5294 0.5317 0.5228 SBRREP 0.5283 0.5279 0.5321 0.5352 0.5374 0.5387 0.5311 %ND NR 16.60 % 17.91% 19.33% 20.97% 22.91% 25.26% 19.55% SR 16.46 % 17.82% 19.41% 21.26% 23.32% 25.66% 19.66% SBRADD 19.96 % 24.41% 28.76% 33.07% 37.42% 41.87% 28.95% SBRREP 21.34 % 27.58% 34.03% 40.65% 47.43% 54.33% 34.46% |ResultSet| NR 2623 692 420 260 147 116 420 SR 2623 1212 639 477 345 238 631 SBRADD 2623 770 450 308 175 139 440 SBRREP 2623 692 420 260 147 116 420 Diversity NR 0.0000 0.0143 0.0352 0.0660 0.0751 0.0774 0.0539 SR 0.0000 0.0194 0.0401 0.0772 0.0864 0.0900 0.0576 SBRADD 0.0000 0.0163 0.0366 0.0737 0.0828 0.0872 0.0560 SBRREP 0.0000 0.0162 0.0737 0.0722 0.0819 0.0867 0.0551 Table 2 Average utility (AvgUtil), share of non-dominated alternatives in the result set (%ND), median of the result set size (|RS|), and average diversity for relaxed (SR), non- relaxed (NR), SBR with addition (SBRADD) and replacement (SBRREP) for number of stated preferences N (used cars dataset). 5.5 Results – digital cameras We compared the average share of non-dominated alternatives in result sets generated us- ing the preference relaxation mechanisms outlined above with the non-relaxed method using related samples non-parametric tests. The results showed a positive effect of the preference relaxation methods on the share of non-dominated products in the list consid- ered by a consumer. In particular, a 20.55% increase in the share of non-dominated prod- ucts was observed in case of the Standard Relaxation method when compared with non- 18 relaxed (10.05% and 8.33% respectively). Similar findings were obtained for the average utility of alternatives in result sets constructed using both methods (AvgUtilNR = 0.5454 and AvgUtilSR = 0.5620). Although the improvement is relatively small (3.03%), both re- sults are statistically significant (p < 0.001), and therefore support H1. The results related to the use of the Soft Boundary Preference Relaxation mechanisms show improvements in the share of non-dominated alternatives in a result set in contrast to both non-relaxing (NR) and Standard relaxation (SR) methods. In particular, 77.92% (SBRADD), and 90.77% (SBRREP) improvement in the share of non-dominated alternatives was observed for SBR methods, in contrast to no relaxation (see Table 3). These differ- ences were statistically significant with p < 0.001. The average utility of alternatives in a result set was similar to all preference-relaxing methods with 0.5620 (SR), 0.5443 (SBRADD), and 0.5506 (SBRREP), the extent of improvement in the average share of non- dominated alternatives in a result set provides evidence for supporting H2.  0.05 0.1 0.2 0.3 0.382 W. A. AvgUtil NR 0.5501 0.5484 0.5454 0.5420 0.5407 0.5454 SR 0.5599 0.5554 0.5605 0.5651 0.5689 0.5620 SBRADD 0.5359 0.5379 0.5454 0.5494 0.5527 0.5443 SBRREP 0.5424 0.5447 0.5519 0.5556 0.5583 0.5506 %ND NR 8.79% 8.62% 8.33% 8.33% 8.15% 8.33% SR 9.64% 10.00% 10.45% 10.20% 10.20% 10.05% SBRADD 13.33% 13.75% 14.82% 16.67% 17.58% 14.82% SBRREP 13.54% 14.24% 15.91% 18.69% 19.32% 15.90% |ResultSet| NR 40 40 39 39 39 39 SR 61 65 72 89 92 74 SBRADD 43 43 44 46 46 44 SBRREP 41 40 40 40 40 40 Diversity NR 0.0112 0.0110 0.0109 0.0104 0.0102 0.0109 SR 0.0152 0.0161 0.0168 0.0183 0.0185 0.0172 SBRADD 0.0125 0.0126 0.0129 0.0136 0.0136 0.0131 SBRREP 0.0120 0.0121 0.0122 0.0129 0.0130 0.0125 19 Table 3 Average utility (AvgUtil), median of the share of non-dominated alternatives in the result set (%ND), the result set size (|RS|), and diversity for relaxed (SR), non-relaxed (NR), SBR with addition (SBRADD) and replacement (SBRREP) for different values of  (digital cameras dataset). The second group of hypotheses related to decision-making effort (H3 and H4), in this case measured by a number of items from which an online consumer has to choose. For the methods investigated, a median of 74 (SR), 44 (SBRADD), and 40 (SBRREP) items in the result set were observed, in contrast to only 39 items on average in a result set for non-relaxed queries (NR). These differences are statistically significant (p < 0.001), con- firming H3. However, it is worth noting that the extent of increase in case of the SBRREP method was extremely low (less than 0.3% when comparing the mean values) indicating only partial support for H4. Finally, results indicated that the alternatives generated using preference relaxation meth- ods were more diverse than those for the non-relaxed method. In particular, an average diversity of 0.0131 (SBRADD), 0.0125 (SBRREP), and 0.0172 (SR) was observed, com- pared to 0.0109 for the non-relaxing method (NR) (see Table 3). These differences were statistically significant (p < 0.001), confirming H5 and H6. The results of the study are summarized in Table 4. 6 Discussion Our study highlights the benefits of the application of intelligent preference relaxation methods in online product search scenarios. Furthermore, we propose a novel Soft Boundary Preference Relaxation method in two variants (with Addition and with Re- placement) that not only possesses the advantages of classical preference relaxation methods but also minimizes additional information processing effort and maximizes product choice quality. First, we showed that preference relaxation methods lead to the construction of product choice sets that make a quality decision much more likely (e.g. with a greater share of non-dominated alternatives). Further, we demonstrated the posi- tive impact of the Soft Boundary Preference Relaxation approach proposed here on prod- uct awareness evidenced by the diversity of items in result sets which, according to Vahi- dov and Ji (2005), leads to higher user satisfaction. In addition, we showed that standard Preference Relaxation (SR) induces a very significant increase in the size of a result set, leading to an unacceptable increase in decision-making effort. We propose two variants of a new method (SBR) that addresses this issue. Finally, we demonstrate that these pro- posed methods outperform the Simple Relaxation (SR) method and minimize additional decision-making effort. The summary of findings and research hypotheses is presented in Table 4. 20 Hypotheses Used Cars Cameras Overall H1 Standard Preference Relaxation has a positive effect on decision quality. Partially Supported Supported Partially Supported H2 Soft Boundary Preference Relaxation has a positive effect on decision quality. Supported Partially Supported Supported H3 Standard Preference Relaxation has a negative effect on decision-making effort Supported Supported Supported H4 Soft-Boundary Preference Relaxation has no effect on decision-making effort Partially Supported Partially supported Partially Supported H5 Standard Preference Relaxation has a positive effect on product awareness. Supported Supported Supported H6 Soft Boundary Preference Relaxation has a positive effect on product awareness. Supported Supported Supported Table 4 Summary of hypotheses The results show that in a large majority of cases the difference in the size of a result set for SBRREP and non-relaxing method (NR) is not statistically significant (see Table 2). Further, when comparing SBRREP and NR methods using the digital cameras dataset, we observed a minimal or close to none increase in the average size of a result set (the aver- age size increased only by 0.3%). However we found a significant (20.55%) increase in the share of non-dominated alternatives. Another performance indicator related to the percentage of result sets that did not contain any non-dominated products. Using the digi- tal camera dataset, we showed a significant improvement in this regard. In particular, out of 562,918 cases as many as 178,243 result sets do not contain non-dominated alterna- tives (31.66%). Preference relaxation methods decrease this negative effect by reducing the number of product sets without a non-dominated alternative to 129,486 (23.00%), representing a 37.65% improvement. Finally, although the initial impact of relaxation on diversity of the results sets seems low, this is due to the selected diversity measure, which favours small results sets. Indeed, the impact of relaxation on diversity is significantly higher for cases with 3 or more specified preferences (see Table 2), for which significant- ly smaller result sets are reported. As such, the study highlights the strong positive impact of SBRREP on our decision-making indicators, with minimum negative impact on effort compared with other relaxation methods. Indeed overall, we showed that our method outperforms standard preference relaxation mechanisms. In conclusion, the introduction of intelligent preference relaxation has an impact on many aspects of consumer decision performance. It ensures that valuable product offers are not erroneously filtered out based on inaccurate preference models provided by consumers 21 using form-based filtering tools. During the process of filtering of initial, very large sets of products, consumers eliminate alternatives that might be valuable later, by providing inaccurate preferences for attributes and attribute values. The paper introduces and vali- dates a new model for a decision aid based on preference relaxation that can limit the po- tentially negative effects of dynamic consumer preferences, whilst addressing limitations of existing methods. The results of the simulation experiment demonstrate the positive ef- fects of preference relaxation on consumer decisions. The e-commerce application of our method are particularly beneficial to providers of online shopping services: diverse result sets may lead to more consumer satisfaction and potentially higher customer retention (Vahidov & Ji, 2005). Moreover, increased average quality of the alternatives considered by a decision maker would reduce decision-making effort. This would have direct relevance to online consumers, as well as having value to e-commerce providers. 7 Acknowledgements The work presented in this paper has been funded in part by Science Foundation Ireland under Grant No. SFI/08/CE/I1380 (Lion-2) and by Science Foundation Ireland grant 10/CE/I1855 to Lero - the Irish Software Engineering Research Centre (www.lero.ie). 8 References Aciar, S., Zhang, D., Simoff, S., & Debenham, J. (2007). Informed recommender: Basing recommendations on consumer product reviews. Ieee Intelligent Systems, 22, 39- 47. Acton, T. (2007). Decision Support for Small Screen Information Systems. National Uni- versity of Ireland, Galway. Adomavicius, G., & Tuzhilin, A. (2005). Toward the Next Generation of Recommender Systems: A Survey of the State-of-the-Art and Possible Extensions. Ieee Transac- tions on Knowledge and Data Engineering, 17, 734-749. Bodapati, A., V. (2008). Recommendation Systems with Purchase Data (Vol. 45). Chica- go, IL, USA: American Marketing Association. Borzsonyi, S., Kossmann, D., & Stocker, K. (2001). The Skyline operator. In 17th Inter- national Conference on Data Engineering (pp. 421-430). Heidelberg, Germany: IEEE Computer Society. Bosc, P., Hadjali, A., & Pivert, O. (2009). Incremental controlled relaxation of failing flexible queries. Journal of Intelligent Information Systems, 33, 261-283. Branting, L. (2002). Optimizing Return-Set Size for Requirements Satisfaction and Cog- nitive Load. In Proceedings of the Third International Symposium on Electronic Commerce: IEEE Computer Society. 22 Bridge, D., & Ferguson, A. (2002). Diverse Product Recommendations Using an Expres- sive Language for Case Retrieval. In S. Craw & A. Preece (Eds.), 6th European Conference on Case-Based Reasoning (pp. 43-57). Aberdeen, Scotland: Springer- Verlag Berlin. Bridge, D., Goker, M. H., McGinty, L., & Smyth, B. (2005). Case-based recommender systems. Knowl. Eng. Rev., 20, 315-320. Bridge, D., & Ricci, F. (2007). Supporting product selection with query editing recom- mendations. In Proceedings of the 2007 ACM conference on Recommender sys- tems. Minneapolis, MN, USA: ACM. Burke, R. (2002). Interactive Critiquing for Catalog Navigation in E-Commerce. Artifi- cial Intelligence Review, 18, 245-267. Cao, H., Chen, E., Yang, J., & Xiong, H. (2009). Enhancing recommender systems under volatile user interest drifts. In CIKM '09: Proceeding of the 18th ACM conference on Information and knowledge management. Hong Kong, China: ACM. Chaudhuri, S. (1990). Generalization and a Framework for Query Modification. In 6th International Conference on Data Engineering (pp. 138-145). Los Angeles, Cali- fornia, USA: IEEE, Computer Soc Press. Chen, L., & Pu, P. (2007). Preference-based organization interfaces: Aiding user critiques in recommender systems. In C. Conati, K. McCoy & G. Paliouras (Eds.), 11th In- ternational Conference on User Modeling (pp. 77-86). Corfu, GREECE: Spring- er-Verlag Berlin. Dabrowski, M., Jarzebowski, P., Acton, T., & O'Riain, S. (2010). Improving Customer Decisions Using Product Reviews: CROM - Car Review Opinion Miner. In 6th International Conference on Web Information Systems and Technologies. Valen- cia, Spain: Springer. Eppler, M. J., & Mengis, J. (2004). The concept of information overload: A review of lit- erature from organization science, accounting, marketing, MIS, and related disci- plines. Information Society, 20, 325-344. Fleder, D. M., & Hosanger, K. (2007). Recommender systems and their impact on sales diversity. In Proceedings of the 8th ACM conference on Electrion commerce. San Diego, California, USA: ACM. Grether, D. M., & Plott, C. R. (1979). Economic Theory of Choice and the Preference Reversal Phenomenon. American Economic Review, 69. Gretzel, U., & Fesenmaier, D. R. (2006). Persuasion in recommender systems. Interna- tional Journal of Electronic Commerce, 11, 81-100. Hadjali, A., Dubois, D., & Prade, H. (2003). Qualitative reasoning based on fuzzy rela- tive orders of magnitude. IEEE Transactions on Fuzzy Systems, 11, 9 - 23. Hadjali, A., Dubois, D., & Prade, H. (2003). Qualitative Reasoning Based on Fuzzy Rela- tive Orders of Magnitude. IEEE Transactions on Fuzzy Systems, 11, 9 - 23. Hagen, P. R., Manning, H., & Paul, Y. (2000). Must Search Stink? In The Forrester Re- port Häubl, G., & Murray, K. B. (2001). Recommending or persuading?: the impact of a shopping agent's algorithm on user behavior. In Proceedings of the 3rd ACM conference on Electronic Commerce. Tampa, Florida, USA: ACM. Häubl, G., & Murray, K. B. (2005). “Double Agent": Assessing the Role of Electronic Product Recommendation Systems. Sloan Management Review, 47, 8-12. 23 Häubl, G., & Trifts, V. (2000). Consumer decision making in online shopping environ- ments: The effects of interactive decision aids. Marketing Science, 19, 4-21. Hostler, R. E., Yoon, V. Y., & Guimaraes, T. (2005). Assessing the impact of Internet agent on end users' performance. Decision Support Systems, 41, 313-323. Johnson, E. J., & Payne, J. W. (1985). Effort and Accuracy in Choice. Management Sci- ence, 31, 395-414. Julià, C., Sappa, A., Lumbreras, F., Serrat, J., & López, A. (2009). Predicting Missing Ratings in Recommender Systems: Adapted Factorization Approach. Internation- al Journal of Electronic Commerce, 14, 89-108. Kim, J. K., Kim, H. K., & Cho, Y. H. (2008). A user-oriented contents recommendation system in peer-to-peer architecture. Expert Systems with Applications, 34, 300- 312. Klein, N. M. (1987). Assessing unacceptable attribute levels in conjoint-analysis. Ad- vances in Consumer Research, 14, 154-158. Klenosky, D. B., & Perkins, W. S. (1992). Deriving attribute utilities from consideration sets: An alternative to self-explicated utilities. Advances in Consumer Research, 19, 657-663. Kohavi, R. (1995). A study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection. In International Joint conference on Artificial Intelligence (IJCAI). Montreal, Quebec, Canada. Lee, W.-P., Liu, C.-H., & Lu, C.-C. (2002). Intelligent agent-based systems for personal- ized recommendations in Internet commerce. Expert Systems with Applications, 22, 275-284. Lipman, B. L. (1995). Information Processing and Bounded Rationality: A Survey. The Canadian Journal of Economics / Revue canadienne d'Economique, 28, 42-67. McGinty, L., & Smyth, B. (2003). On the Role of Diversity in Conversational Recom- mender Systems. In K. Ashley & D. Bridge (Eds.), Case-Based Reasoning Re- search and Development (Vol. 2689, pp. 1065-1065): Springer Berlin / Heidel- berg. McSherry, D. (2004). Retrieval failure and recovery in recommender systems. In 15th Artificial Intelligence and Cognitive Science Conference (AICS 2004) (pp. 319- 338). Castlebar, Ireland: Springer. Mirzadeh, N., & Ricci, F. (2007). Cooperative query rewriting for decision making sup- port and recommender systems. Applied Artificial Intelligence, 21, 895-932. Mirzadeh, N., Ricci, R., & Bansal, M. (2004). Supporting user query relaxation in a rec- ommender system. In K. Bauknecht, M. Bichler & B. Proll (Eds.), 5th Interna- tional Conference on E-Commerce and Web Technology (pp. 31-40). Zaragoza, SPAIN: Springer-Verlag Berlin. Motro, A. (1990). FLEX: A Tolerant and Cooperative User Interface to Databases. IEEE Transactions on Knowledge and Data Engineering, 2, 231-246. Olson, E. L., & Widing, R. E. (2002). Are interactive decision aids better than passive decision aids? A comparison with implications for information providers on the internet. Journal of Interactive Marketing, 16, 22-33. Oppewal, H., & Klabbers, M. (2002). Compromising between information completeness and task simplicity: A comparison of self-explicated, hierarchical information in- tegration, and full-profile conjoint methods. In P. A. R. D. W. Keller (Ed.), 33rd 24 Annual Conference of the Association-for-Consumer-Research (pp. 298-304). At- lanta, Georgia: Assoc Consumer Research. Parra, J. F., & Ruiz, S. (2009). Consideration sets in online shopping environments: the effects of search tool and information load. Electronic Commerce Research and Applications, 8, 252-262. Payne, J. W., Bettman, J. R., & Johnson, E. J. (1992). Behavioral Decision Research: A Constructive Processing Perspective. Annual Review of Psychology, 43, 87-131. Payne, J. W., Bettman, J. R., & Johnson, E. J. (1993). The adaptive decision maker: Cambridge University Press. Payne, J. W., Bettman, J. R., & Schkade, D. A. (1999). Measuring constructed prefer- ences: Towards a building code. In Symposium on Elicitation of Preferences (pp. 243-270). Berkeley, California: Kluwer Academic Publ. Pereira, R. E. (2001). Influence of Query-Based Decision Aids on Consumer Decision Making in Electronic Commerce. Information Resources Management Journal, 14, 31-48. Pu, P., Chen, L., & Kumar, P. (2008). Evaluating product search and recommender sys- tems for E-commerce environments. Electronic Commerce Research, 8, 1-27. Resnick, P., & Varian, H. R. (1997). Recommender systems. Communications of the Acm, 40, 56-58. Shepard, R. N. (1987). Toward a universal law of generalization for psychological sci- ence. Science, 237, 1317-1323. Shimazu, H. (2002). ExpertClerk: A Conversational Case-Based Reasoning Tool forDe- veloping Salesclerk Agents in E-Commerce Webshops. Artificial Intelligence Re- view, 18, 223-244. Slovic, P. (1995). The construction of preference. American Psychologist, 50, 364-371. Smyth, B., & McClave, P. (2001). Similarity vs. Diversity. In Proceedings of the 4th In- ternational Conference on Case-Based Reasoning: Case-Based Reasoning Re- search and Development: Springer-Verlag. Stahl, A. (2002). Defining similarity measures: Top-down vs. bottom-up. In S. Craw & A. Preece (Eds.), 6th European Conference on Case-Based Reasoning (pp. 406- 420). Aberdeen, Scotland: Springer-Verlag Berlin. Stahl, A. (2004). Approximation of Utility Functions by Learning Similarity Measures In Logic versus Approximation (pp. 150-172): Springer Berlin / Heidelberg. Stahl, A. (2006). Combining case-based and similarity-based product recommendation. In T. R. RothBerghofer, M. H. Goker & H. A. Guvenir (Eds.), 8th European Con- ference on Case-Based Reasoning (pp. 355-369). Fethiye, TURKEY: Springer- Verlag Berlin. Swaminathan, V. (2003). The Impact of Recommendation Agents on Consumer Evalua- tion and Choice: The Moderating Role of Category Risk, Product Complexity, and Consumer Knowledge. Journal of Consumer Psychology, 13, 93-101. Turetken, O., & Sharda, R. (2004). Development of a fisheye-based information search processing aid (FISPA) for managing information overload in the web environ- ment. Decision Support Systems, 37, 415-434. Tversky, A. (1996). Contrasting Rational and Psychological Principles in Choice. In R. J. Zeckhauser, R. L. Keeney & J. K. Sebenius (Eds.), Wise choices: decisions, games, and negotiations: Harvard Business Press. 25 Vahidov, R., & Ji, F. (2005). A diversity-based method for infrequent purchase decision support in e-commerce. Electron. Commer. Rec. Appl., 4, 143-158. van der Heijden, H. (2006). Mobile decision support for in-store purchase decisions. De- cision Support Systems, 42, 656-663. Viappiani, P., Faltings, B., & Pu, P. (2006). Evaluating preference-based search tools: a tale of two approaches. In Proceedings of the 21st national conference on Artifi- cial intelligence - Volume 1. Boston, Massachusetts: AAAI Press. Viappiani, P., Pu, P., & Faltings, B. (2008). Preference-based search with adaptive rec- ommendations. AI Communications, 21, 155-175. Wang, W., & Benbasat, I. (2007). Recommendation Agents for Electronic Commerce: Effects of Explanation Facilities on Trusting Beliefs. MIS Quarterly, 23. Xiao, B., & Benbasat, I. (2007). E-Commerce Product Recommendation Agents: Use, Characteristics and Impact. MIS Quarterly, 31, 137-209. Zhang, Y., & Jiao, J. (2007). An associative classification-based recommendation system for personalization in B2C e-commerce applications. Expert Systems with Appli- cations, 33, 357-367. 
work_wohh7ixwkferhl6m3bu2azcbzm	----	Puerto Rico Trench: Site of a Shallow-Water Tertiary Basin and Regional Tectonic Implications: ABSTRACT; Expert System for Computer Interpretation of Beach and Nearshore Facies: ABSTRACT; Impact of Early Diagenesis of Eolian Reservoirs, Great Sand D... 276 Association Round Table ognition is enhanced by color differences between paleosols developed on the scoured and scour-fill deposits, is a valuable indicator of climatically and/or tectonically controlled base level changes in continental basins. Because paleosols preserve the record of nondeposition and nonerosion and are useful in identifying periods of degradation, their analysis is essential to reconstructing the complete geologic history of alluvial sedi- ments. KRIEG, EDWARD A., E. A. Krieg & Assoc, Inc., New Orleans, LA, ARTHUR A. MEYERHOFF*, Consultant, Tulsa, OK, and IRFAN TANER, Consultant, Tulsa, OK Puerto Rico Trench: Site of a Shallow-Water Tertiary Basin and Regional Tectonic Implications Until late Eocene time, the Bahamas platform extended to the present Virgin Islands, as demonstrated by magnetic, gravity, and refraction data. This interpretation is confirmed by the presence of widespread out- crops of middle Cretaceous through early Pliocene shallow-water bank carbonates below 5,200 m depth in the trench. Crustal thickness beneath this bank is 18-25 km. Igneous and metamorphic rocks from the base of the trench's southern slope are chemically very different from subduction-zone rocks. Waters of the carbonate bank (300 x 100 km in size) transgressed southward after early Eocene time. During late Eoene time, the bank's southern margin was near today's shoreline where down-to-the-north growth faults formed. Along the bank's northern margin, block faulting produced a graben above the site of the modern Puerto Rico Trench. During middle Eocene to early Pliocene time, shallow-water deposition extended from a position presently 5,200 m deep in the trench to central Puerto Rico, an exceptionally stable block at least 100 km wide. During middle Eocene time, the Beata Ridge dextral shear cut the trench off north of Hispaniola. In early Pliocene time, the Mona Canyon dextral fault zone cut across the trench, and strong northward tilting commenced. The trench's present southern slope is mainly a dip slope, inclined about 5°. The Puerto Rico Outer Ridge formed by lateral and upward movements of mantle materials that withdrew from beneath the sinking trench. Petroleum prospects presently are limited to the Tertiary (4,000 m thick) and to a coastal zone 20-25 km wide (to 2,000 m water depth). Traps are mainly fault seals and stratigraphic pinch-outs. KRYSTINIK, KATHERINE B., U.S. Geol. Survey, Denver, CO, and H. EDWARD CLIFTON, U.S. Geol. Survey, Palo Alto, CA Expert System for Computer Interpretation of Beach and Nearshore Facies A user-friendly, rule-based expert system has been designed for inter- pretation of lithofacies characteristics of beach and nearshore deposi- tional environments. Recently, similar expert systems have been widely applied in medicine, business, and mineral exploration. The expert system runs on a VAX 780 (trade name). By incorporating knowledge and understanding of an expert, the system can interact with a user the way an expert consultant would. Interaction consists of a series of questions about lithology, sedimentary structures, and bioturbation of the lithofacies observed in outcrop or core. Uncertain responses are allowed and incorporated into the reasoning. Dialogue varies in different consultations because questions asked by the system depend on users' responses to previous questions. The result is an evaluation of the likeli- hood that the deposit under consideration is actually a beach or nearshore deposit. Significant lithofacies characteristics, the reasoning used in reaching the conclusion, and pertinent references are provided. Expert systems for other depositional environments are being designed. As their availability increases, geologists without easy access to experts on a particular depositional environment will have expert consult- ants as close as a computer terminal. Also the ability of the system to explain hs reasoning and provide references lends the system to instruc- tional uses. KRYSTINIK, LEE F., Champlin Petroleum, Englewood, CO, SARAH ANDREWS, Angus Petroleum, Golden, CO, and STEVEN G. FRY- BERGER, McAdams, Roux & Assoc, Denver, CO Impact of Early Diagenesis of Eolian Reservoirs, Great Sand Dunes National Monument, Colorado Dune and associated alluvial and playa deposits at Great Sand Dunes National Monument, Colorado, provide an excellent opportunity to study early diagenetic development of vertical and horizontal permeabil- ity barriers in recent eolian deposits (> 10 ka). Cements observed include calcite, aragonite, protodolomite(?), amorphous silica, iron hydroxide, smectite, trona, and halite. Cementation is controlled by the availability of water, with several hydrologic subenvironments producing different cements. Evaporative cementation in dunes adjacent to playas is commonly dominated by trona and halite, but calcite, aragonite, and amorphous sil- ica also bind the sediment. These cements are generally most concen- trated in fine laminations where capillary action has pulled water into dunes. Iron hydroxides, calcite, and amorphous silica precipitate at the interface between ground water and streams or lakes, where the pH gradi- ent may exceed 5 pH units (pH 5.7-11.5). Subsequent movement of the ground-water table can resuh in cross-cutting cement zones. Early cementation in dunes prevents deflation and provides a mecha- nism for preservation of the reservoir unit. Intense cementation may per- manently occlude porosity, or leaching may reestablish well-inter- connected porosity. An understanding of the extent and composition of early cement zones can be used to improve hydrodynamic models for pro- duction and enhanced recovery. KUGLER, RALPH L,, Univ. Texas at Austin, Austin, TX Source Rock Characteristics, Los MoUes and Vaca Muerta Shales, Neuquen Basin, West-Central Argentina Major hydrocarbon-producing trends of the Neuquen basin occur along its northeastern margin (Eastern Shelf) and along an east-west- trending structural high in the southern half of the basin (Neuquen Dor- sal). Sediment thickness increases northward from the Neuquen Dorsal and westward from the Eastern Shelf into the Neuquen Embayment, a region that has been relatively unproductive for hydrocarbons. Major source rocks are the Lower to Middle Jurassic Los MoUes For- mation and the Upper Jurassic to Lower Cretaceous Vaca Muerta Forma- tion. Los Molles shales are immature to moderately mature on the Neuquen Dorsal: vitrinite reflectance (Rg) = 0.3-0.8%; thermal altera- tion index (TAI) = 1 -I- to 2; total organic carbon (TOC) = 2.0-5.0%. However, they are severely altered in the deepest part of the Neuquen Embayment (R„ = 2.5-3.0%; TAI = 3-1- to 4; TOC = 1.1%). Organic matter is woody on the Neuquen Dorsal but is coaly in the Neuquen Embayment. Clay minerals are smectite and mixed layer illite-smectite on the Neuquen Dorsal whereas illite and chlorite are present in the Neuquen Embayment. Similarly, Vaca Muerta shales are immature on the Neuquen Dorsal and along much of the Eastern Shelf (R ,̂ = 0.3-0.5%; TAI = 1 to 2 - ) , but are mature throughout the Neuquen Embayment (R ,̂ = 0.7-1.3%; TAI = 2-3). The lower part of the unit is a bituminous black shale (TOC = 2.5-6.5%). The dominant visual kerogen type is amorphous (Al). Clay minerals change from smectite to mixed layer illite-smectite with decreas- ing expandability toward the deepest part of the Neuquen Embayment. The lack of correlation between areas of source rock maturity and major hydrocarbon production suggests long-distance fluid transport out of deep portions of the basin. KVENVOLDEN, KEITH A., U.S. Geol. Survey, Menlo Park, CA Geochemical Conditions in Sediment Containing Gas Hydrates of Active and Passive Continental Margins Two sites of the Deep Sea DriUing Project in contrasting geologic set- tings provide a basis for comparison of the geochemical conditions asso- ciated with marine gas hydrates in continental margin sediments. Site 533 is located at 3,191 m water depth on a spitlike extension of the continental rise (Blake Outer Ridge) on a passive margin in the Atlantic Ocean. Site 568, at 2,031 m water depth, is in upper-slope sediment of an active accre- tionary margin (Middle America Trench) in the Pacific Ocean. Both sites are characterized by high rates of sedimentation (greater than 30 m/m.y.), and the organic carbon contents of these sediments generally exceed 0.5%. Anomalous seismic reflections that crosscut reflections from sedi- mentary layers and parallel reflections from the sea floor suggested the presence of gas hydrates at both sites. During coring, small samples of gas 
work_wqit3j3isractbtv2fgeoweo5y	----	Credit scoring using the clustered support vector machine Credit scoring using the clustered support vector machine Terry Harris ⇑ Credit Research Unit, Department of Management Studies, The University of the West Indies, Cave Hill Campus, P.O. Box 64, Barbados a r t i c l e i n f o Article history: Available online 6 September 2014 Keywords: Credit risk Credit scoring Clustered support vector machine Support vector machine a b s t r a c t This work investigates the practice of credit scoring and introduces the use of the clustered support vec- tor machine (CSVM) for credit scorecard development. This recently designed algorithm addresses some of the limitations noted in the literature that is associated with traditional nonlinear support vector machine (SVM) based methods for classification. Specifically, it is well known that as historical credit scoring datasets get large, these nonlinear approaches while highly accurate become computationally expensive. Accordingly, this study compares the CSVM with other nonlinear SVM based techniques and shows that the CSVM can achieve comparable levels of classification performance while remaining relatively cheap computationally. � 2014 Elsevier Ltd. All rights reserved. 1. Introduction In recent years, credit risk assessment has attracted significant attention from managers at financial institutions around the world. This increased interest has been in no small part caused by the weaknesses of existing risk management techniques that have been revealed by the recent financial crisis and the growing demand for consumer credit (Wang, Yan, & Zhang, 2011). Address- ing these concerns, over past decades credit scoring has become increasingly important as financial institutions move away from the traditional manual approaches to this more advanced method, which entails the building of complex statistical models (Huang, Chen, & Wang, 2007; Zhou, Lai, & Yu, 2010). Many of the statistical methods used to build credit scorecards are based on traditional classification techniques such as logistic regression or discriminant analysis. However, in recent times non-linear approaches,1 such as the kernel support vector machine, have been applied to credit scoring. These methods have helped to increase the accuracy and reliability of many credit scorecards (Bellotti & Crook, 2009; Yu, 2008). Nevertheless, despite these advances credit analyst at financial institutions are pressed to continually pursue improvements in classifier performance in an attempt to mitigate the credit risk faced by their institutions. However, many of the improvements in classifier performances remain unreported due to the proprietary nature of industry led credit scoring research which attempts to find more efficient and effective algorithms. In the wider research community, the recent vintages of non-linear classifiers (e.g. the kernel support vector machine) have received a lot of attention and have been critiqued for, inter alia, their large time complexities. In fact the best-known time complexity for training a kernel based support vector machine is still quadratic (Bordes, Ertekin, Weston, & Bottou, 2005). As a result, when applied to credit scoring substantial computational resources are consumed when training on reasonably sized real world datasets. Accordingly, efforts to develop and apply new classifiers to credit scoring, which are capable of separating nonlin- ear data while remaining relatively inexpensive computationally, are well placed. This paper investigates the suitability for credit scoring of a recently developed support vector machine based algorithm that has been proposed by Gu and Han (2013). Their clustered support vector machine has been shown to offer comparable performance to kernel based approaches while remaining cheap in terms of computational time. Furthermore, this study makes some novel adjustments to their implementation and explores the use of radius basis function (RBF) kernels in addition to the linear kernel posited by Gu and Han. The remainder of this paper is presented as follows. Section 2 outlines a brief review of the literature concerning the field of credit scoring and sets the stage for the proposed CVSM model for credit scoring that is presented in Section 3. The details of the historic clients’ loan dataset and modeling method are highlighted in Section 4. Section 5 presents the study results, and Section 6 dis- cusses the findings, presents conclusions, and outlines possible directions for future research. http://dx.doi.org/10.1016/j.eswa.2014.08.029 0957-4174/� 2014 Elsevier Ltd. All rights reserved. ⇑ Tel.: +1 (246) 417 4302; fax: +1 (246) 438 9167. E-mail address: terry.harris@cavehill.uwi.edu 1 This has been applied because credit-scoring data is often not linearly separable. Expert Systems with Applications 42 (2015) 741–750 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2014.08.029&domain=pdf http://dx.doi.org/10.1016/j.eswa.2014.08.029 mailto:terry.harris@cavehill.uwi.edu http://dx.doi.org/10.1016/j.eswa.2014.08.029 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 2. Background and related works 2.1. Overview Credit scoring has been critical in permitting the exceptional growth in consumer credit over the last decades. Indeed without accurate, automated credit risk assessment tools, lenders could not have expanded their balance sheets effectively over this time. This section presents a brief review of the relevant literature that has emerged in this space. 2.2. What is credit scoring? Credit scoring can be viewed as a method of measuring the risk attached to a potential customer, by analyzing their data to deter- mine the likelihood that the prospective borrower will default on a loan (Abdou & Pointon, 2011). According to Eisenbeis (1978), Hand and Jacka (1998), and Hand, Sohn, and Kim (2005) credit scoring can also be described as the statistical technique employed to con- vert data into rules that can be used to guide credit granting deci- sions. As a result, it represents a critical process in a firm’s credit management toolkit. Durand (1941) posited that the procedure includes collecting, analyzing and classifying different credit ele- ments and variables in order to make credit granting decisions. He noted that to classify a firm’s customers, the objective of the credit evaluation process, is to reduce current and expected risk of a customer being ‘‘bad’’ for credit. Thus credit scoring is an important technology for banks and other financial institutions as they seek to minimize risk. 2.3. Related works Over the years, the demand for consumer credit has increased exponentially. According to Steenackers and Goovaerts (1989), this increase in the demand for credit can be attributable to the increased levels of consumption and the reliance on credit to sup- port this activity. In the United States, this rising level of consum- erism followed the introduction of the first modern credit card in 1950s, so that by the 1980s over 55% of American households owned a credit card. Crook, Edelman, and Thomas (2007) posited that by this time, in the US, the total amount of outstanding con- sumer credit was over $700 billion. Comparatively, at the end of June 2013 this figure had risen to a staggering $2800 billion, a 400% increase (BGFRS, 2013). Henley (1994) noted that the increasing demand for consumer credit has led to the development of many practical the scoring models, which have adopted a wide range of statistical and non- linear methods. Similarly, Mays (2001) posited that a number of various techniques have been used to build credit scoring applica- tions by credit analyst, researchers, and software developers. These techniques have included; discriminant analysis, linear regression, logistic regression, decision trees, neural networks, support vector machines, k-means, etc. In recent times, the use of more complex non-linear techniques, such as neural networks, and support vector machines, to build credit scoring applications has seen significant increases in the reported accuracy and performance on benchmarking datasets (Baesens et al., 2003). Irwin, Warwick, and Hunt (1995) and Paliwal and Kumar (2009) both provide evidence that advanced statistical techniques yield superior performance when compared to traditional statistical techniques, such as discriminant analysis, probit analysis and logistic regression. Masters (1995) also provided evidence that the use of sophisticated techniques, such as neural networks, was essential because they had the capability to more accurately model credit scoring data that exhibits interactions and curvature. However, as pointed out by Hand (2006) the increased performance of these more advanced tech- niques could be illusionary and if real, diminished due to shifts in the class distribution over time. The following sub-sections pres- ent a brief discussion concerning some of the classical and advanced statistical models used for credit risk assessment. 2.4. Discriminant analysis In his seminal paper, Fisher (1936) proposed the use of discrim- inant analysis to differentiate between two or more classes in a dataset. Since that time, Durand (1941) and Altman (1986) have both applied Fisher’s (1936) discriminant analysis to credit scoring. Durant used discriminant analysis to assess the creditworthiness of car loan applicants, while Altman used it to explore corporate bankruptcy proposing his popular Z-scores (Altman, 1968). In works published separately by Desai, Crook, and Overstreet (1996), Hand and Henley (1997), Hand, Oliver, and Lunn (1998), Sarlija, Bensic, and Bohacek (2004), and Abdou and Pointon (2009), they showed that discriminant analysis is indeed a valid technique for credit scoring. Hand and Henley (1997) noted that discriminant analysis, a parametric statistical technique, was well suited to credit scoring because it was designed to classify groups and variables into two or more categories or discriminate between two groups. However, Saunders and Allen (1998) noted that with this type of method certain assumptions about the data must be met. These assumptions include, normality, linearity, homoscedas- ticity, non-multicollinearity, etc. Falbo (1991) and Sarlija et al. (2004) posited that despite these limitations, over the years this technique has been frequently applied to build credit scoring applications, and it remains one of the most popular approaches taken today when classifying customers as creditworthy or un- creditworthy. Several authors have criticized the use of discriminant analysis in credit scoring. Eisenbeis (1978) point-out a number of the statis- tical problems in applying discriminant analysis to credit scorecard development. These problems include the following: group defini- tion, classification error prediction, estimating population priors, and the use of linear functions instead of quadratic functions, to mention a few. Nevertheless, Greene (1998) and Abdou (2009) noted that despite these limits, discriminant analysis is one of the most commonly used techniques in credit scoring. 2.4.1. Linear regression Another popular classical statistical technique applied to credit scoring is linear regression. This method has developed into an essential component of data analysis in general and is concerned with describing the relationship between a dependent variable and one or more independent variables. Thus, customers’ historical payments, guarantees, default rates and other factors can be ana- lyzed using linear regression to set up a score for each factor, and compare it with the bank’s cut-off (threshold) score. Hence, only if a new customer’s score exceeds the bank’s cut-off score will credit be granted (Hand & Jacka, 1998). In its basic form, linear regression used for credit scoring requires the establishment of a threshold score. This threshold credit score is derived from the relationships between the firm’s historic clients’ features and their associated weights. As can be seen in the linear equation, Z ¼ h0 þ h1 x1 þ h2x2 þ�� �þ hnxn, where the variable n denotes the number of features collected from past and potential clients. These features are represented by the x’s, which are multi-dimensional vectors in Rm, where m denotes the number of clients in the historical clients’ database. The h’s repre- sent the weights, and the feature variables and their weights used to calculate a credit score, Z 2 R, thus when an applicant scores 742 T. Harris / Expert Systems with Applications 42 (2015) 741–750 https://isiarticles.com/article/48568 
work_wqzwnytcxjgfhkid6c7remx4za	----	Research Article Mathematical Modeling of the Expert System Predicting the Severity of Acute Pancreatitis Maria A. Ivanchuk,1 Vitalij V. Maksimyuk,2 and Igor V. Malyk3 1 Department of Biological Physics and Medical Informatics, Bukovinian State Medical University, Kobyljanska Street 42, Chernivtsi 58000, Ukraine 2 Department of Surgery, Bukovinian State Medical University, Golovna Street 137, Chernivtsi 58000, Ukraine 3 Department of the System Analysis and Insurance and Financial Mathematics, Chernivtsi National University of Yuriy Fedkovich, Unversitetska Street 12, Chernivtsi 58012, Ukraine Correspondence should be addressed to Maria A. Ivanchuk; mgracia@ukr.net Received 26 December 2013; Accepted 22 May 2014; Published 9 June 2014 Academic Editor: Daniel Kendoff Copyright © 2014 Maria A. Ivanchuk et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. The method of building the hyperplane which separates the convex hulls in the Euclidean space 𝑅𝑛 is proposed. The algorithm of prediction of the presence of severity in patients based on this method is developed and applied in practice to predict the presence of severity in patients with acute pancreatitis. 1. Introduction During the last decades, pronounced tendency to the relentless increase in morbidity in acute pancreatitis is observed. Thus, the depth of pathomorphological pancre- atic parenchyma lesions can vary from the development of edematous pancreatitis up to pancreatic necrosis. However, accurate predicting of the probable nature of the lesion of the pancreas in the early stages of acute pancreatitis is one of the most difficult problems of modern pancreatology. Diagnostic and the predictive probability of existing laboratory and instrumental diagnostic markers and rating scales does not exceed 70–80% [1–3]. Such situation is a major difficulty in selecting the adequate treatment strategy in the initial stages of acute pancreatitis. Thus the search for new methods of accurate predicting of acute pancreatitis’ severity becomes an urgent problem. Development of mathematical approaches for prediction in medicine was developed by Fisher, the father of the linear discriminant analysis [4]. Currently, there are many approaches to solving this problem: cluster analysis, the construction of predictive tables, image recognition, and lin- ear programming. Fundamentals of building the prognostic tables and Wald serial analysis are described in [5]. Cluster analysis is commonly used for solving the tasks of medical prediction. In the paper [6], the procedure of cluster analysis with a study of the indices of the daily variability of cardiac rhythm in patients with the ischemic disease of heart is examined. In [7] using national data from the Scientific Registry of Transplant Recipients authors compare transplant and wait-list hospitalization rates. They suggest two marginal methods to analyze such clustered recurrent event data; the first model postulates a common baseline event rate, while the second features cluster-specific baseline rates. Results from the proposed models to those based on a frailty model were compared with the various methods compared and contrasted. Three major considerations in designing a cluster analysis are described in [8]. The first relates to selection of the individuals. The second consideration is selection of variables for measurement and the third consideration is how many variables to choose to enter into a cluster analysis. To classify clinical phenotypes of anti-neutrophil cytoplasmic antibody-associated vasculitis, cluster analysis was used in [9]. Researches on the general theory of diagnosis, classifi- cation, and application of optimization methods for pattern recognition, solving applied problems in medicine and biol- ogy, are conducted by Mangasarian et al. for many years [10]. Hindawi Publishing Corporation Journal of Computational Medicine Volume 2014, Article ID 532453, 4 pages http://dx.doi.org/10.1155/2014/532453 2 Journal of Computational Medicine But universal method for solving problems of recogni- tion, identification, and diagnosis does not exist. Therefore, development of methods for predicting in medicine still remains relevant. One among the many challenges of recog- nition is the task of constructing hyperplanes which separate two convex sets. Many manuscripts [11–16] are devoted to the solution of this problem. We propose a methodology for constructing convex hulls and their separation, which can be used for modeling expert medical prognostic systems (e.g., to separate groups of patients with different degrees of severity of the disease for prediction of severity in patients). 2. Methods 2.1. Separation of the Convex Hulls. Let us have two sets of points 𝐴 = {𝑎 𝑖 = (𝑎 1 𝑖 ,𝑎 2 𝑖 , . . . ,𝑎 𝑛 𝑖 ), 𝑖 = 1,𝑚 𝐴 } and 𝐵 = {𝑏 𝑖 = (𝑏 1 𝑖 ,𝑏 2 𝑖 , . . . ,𝑏 𝑛 𝑖 ), 𝑖 = 1,𝑚 𝐵 } in Euclidean space 𝑅𝑛. Let 𝑚 be number of points in the set. We must find the separate hyperplane: 𝐿 𝑝 = {𝑥 ∈𝑅 𝑛 : ⟨𝑝,𝑥⟩ = 𝛾}, 𝑝 ̸=0, (1) where ⟨𝑝,𝑥⟩ is the scalar product of the vectors𝑝and𝑥 such that sets 𝐴 and 𝐵 can be placed in the different half-spaces: 𝐿 + 𝑝 = {𝑥 ∈𝑅 𝑛 : ⟨𝑝,𝑥⟩ > 𝛾}, 𝐿 − 𝑝 = {𝑥 ∈𝑅 𝑛 : ⟨𝑝,𝑥⟩ < 𝛾}. (2) To build the convex hull conv 𝐴 for the set 𝐴, for each of 𝐶 𝑛 𝑚 𝐴 points’ combinations from the set𝐴, if it is possible, build the hyperplane 𝐻 𝑝 = {𝑥 ∈𝑅 𝑛 : ⟨𝑝,𝑥⟩ = 𝛽}, 𝑝 ̸=0. (3) Coordinates of the vector 𝑝 = (𝑝1, . . . ,𝑝𝑛) are found as minors(𝑛−1)order for elements of the first row of the matrix: ( 𝑥 1 −𝑎 1 1 𝑥 2 −𝑎 2 1 ⋅ ⋅ ⋅ 𝑥 𝑛 −𝑎 𝑛 1 𝑎 1 2 −𝑎 1 1 𝑎 2 2 −𝑎 2 1 ⋅ ⋅ ⋅ 𝑎 𝑛 2 −𝑎 𝑛 1 ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ 𝑎 1 𝑛 −𝑎 1 1 𝑎 2 𝑛 −𝑎 2 1 ⋅ ⋅ ⋅ 𝑎 𝑛 𝑛 −𝑎 𝑛 1 ), (4) where 𝑥 ∈ 𝑅𝑛, 𝑎 𝑖 ∈ 𝐴, 𝑖 = 1,𝑛. Coefficient 𝛽 is determined from the following equation: 𝛽=−(𝑎 1 1 𝑝 1 +𝑎 2 1 𝑝 2 + ⋅ ⋅ ⋅ +𝑎 𝑛 1 𝑝 𝑛 ). (5) If all points of the set 𝐴 are in the one of half-spaces of hyperplane 𝐻 𝑃 , then polygon 𝑎 1 𝑎 2 ⋅ ⋅ ⋅𝑎 𝑛 is one of the convex hull’s hyperfaces. The complex of all hyperfaces is the convex hull conv 𝐴 . Point 𝑏 𝑖 ∈ 𝐵 is called outlier if point 𝑏 𝑖 is internal for the conv 𝐴 . Point 𝑏 𝑖 is outlier if there is at least one hyperface 𝑎 1 𝑎 2 ⋅ ⋅ ⋅𝑎 𝑛 ∈ conv 𝐴 that 󵄨󵄨󵄨󵄨󵄨󵄨󵄨 󳨀→ 𝑐𝑓 󵄨󵄨󵄨󵄨󵄨󵄨󵄨 = 󵄨󵄨󵄨󵄨󵄨󵄨󵄨 󳨀→ 𝑐𝑏 𝑖 󵄨󵄨󵄨󵄨󵄨󵄨󵄨 + 󵄨󵄨󵄨󵄨󵄨󵄨󵄨 󳨀󳨀→ 𝑏 𝑖 𝑓 󵄨󵄨󵄨󵄨󵄨󵄨󵄨 , (6) where point 𝑐 ∈ int conv 𝐴 , point 𝑓 is the intersection point of the hyperplane 𝐻 𝑝 (𝑎 1 𝑎 2 ⋅ ⋅ ⋅𝑎 𝑛 ∈𝐻 𝑝 ), and line 𝑥 1 −𝑐 1 𝑏 1 −𝑐 1 = 𝑥 2 −𝑐 2 𝑏 2 −𝑐 2 = ⋅ ⋅ ⋅ = 𝑥 𝑛 −𝑐 𝑛 𝑏 𝑛 −𝑐 𝑛 . (7) To find the point 𝑓 let us write (3) in parametric form: 𝑥 1 = 𝑐 1 +(𝑏 1 −𝑐 1 )𝑡 𝑥 2 = 𝑐 2 +(𝑏 2 −𝑐 2 )𝑡 ⋅ ⋅ ⋅ 𝑥 𝑛 = 𝑐 𝑛 +(𝑏 𝑛 −𝑐 𝑛 )𝑡. (8) Put (8) in the hyperplane equation (3) and find parameter 𝑡: 𝑡 ∗ = 𝑝 1 𝑐 1 +𝑝 2 𝑐 2 + ⋅ ⋅ ⋅ +𝑝 𝑛 𝑐 𝑛 +𝛽 𝑝 1 (𝑐 1 −𝑏 1 )+𝑝 2 (𝑐 2 −𝑏 2 )+ ⋅ ⋅ ⋅ +𝑝 𝑛 (𝑐 𝑛 −𝑏 𝑛 ) . (9) To find coordinates of the point 𝑓 let us put (9) in (8): 𝑓 1 = 𝑐 1 +(𝑏 1 −𝑐 1 )𝑡 ∗ 𝑓 2 = 𝑐 2 +(𝑏 2 −𝑐 2 )𝑡 ∗ ... 𝑓 𝑛 = 𝑐 𝑛 +(𝑏 𝑛 −𝑐 𝑛 )𝑡 ∗ . (10) After finding all outliers from the sets 𝐴 and 𝐵 eject outliers from the set, with less number of outliers. Build the new convex hulls and find the outliers. If there are outliers in the new convex hulls, eject them. If there are not any outliers, the convex hulls do not intersect. According to consequence of Hahn-Banach theorem there is a nonzero linear functional 𝐿 𝑝 that separates conv 𝐴 and conv 𝐵 [17]. Find the separating functional 𝐿 𝑝 as hyperplane parallel to one of convex hulls’ hyperfaces. Choose hyperface so that convex hulls conv 𝐴 and conv 𝐵 are in different half-spaces formed by hyperplane parallel to this hyperface. Find points 𝑎min ∈ conv𝐴 and 𝑏min ∈ conv𝐵 so that | 󳨀󳨀󳨀󳨀󳨀󳨀→ 𝑎min𝑏min| = min(| 󳨀󳨀→ 𝑎 𝑖 𝑏 𝑗 | : 𝑎 𝑖 ∈ conv 𝐴 , 𝑏 𝑗 ∈ conv 𝐵 , 𝑖 = 1,𝑚 𝐴 , 𝑗 = 1,𝑚 𝐵 ). Let 𝑑min ∈ 󳨀󳨀󳨀󳨀󳨀󳨀→ 𝑎min𝑏min. For each hyperface {𝐻𝐴min : 𝐻 𝐴min ⊂ conv 𝐴 ;𝑎min ∈ 𝐻𝐴min} and {𝐻𝐵min : 𝐻𝐵min ⊂ conv 𝐵 ;𝑏min ∈ 𝐻𝐵min} build the parallel hyperplane {𝐿𝑝 : 𝑑min ∈ 𝐿𝑝;𝐿𝑝‖𝐻𝐴minor 𝐿𝑝‖𝐻𝐵min}. If 𝑎𝑖 ∈ 𝐿 + 𝑝 , for all 𝑎 𝑖 ∈ 𝐴, 𝑖 = 1,𝑚 𝐴 , and 𝑏 𝑗 ∈ 𝐿 − 𝑝 , for all 𝑏 𝑗 ∈ 𝐵, 𝑗 = 1,𝑚 𝐵 , then 𝐿 𝑝 is separating hyperplane. 2.2. Modeling the Expert System of Predicting the Presence of Severity in Patients. Let us have two groups of patients: 𝐴, patients with severity, and 𝐵, patients without severity. There are 𝑛 0 parameters (factors which affect the severity) known for each patient. Journal of Computational Medicine 3 During modelling we used the terms sensitivity (Se) and specificity (Sp): Se = 𝑎 𝑎+𝑐 , Sp = 𝑑 𝑏+𝑑 , (11) where𝑎 is the true positives,𝑏 is the false positives (overdiag- nosis errors), 𝑐 is the false negatives (underdiagnosis errors), and 𝑑 is the true negatives. The sensitivity of a clinical test refers to the ability of the test to correctly identify those patients with the disease. The specificity of a clinical test refers to the ability of the test to correctly identify those patients without the disease [18]. We created an algorithm of modelling the expert system in a way that uses the least amount of features for the best result. Information of the parameters was found using Kulback’s information measure [5]. We built convex hulls for the most informative factor. If convex hulls intersect, we found outliers—the points from the set 𝐴 that are internal to conv 𝐵 and the points from the set𝐵 that are internal to conv 𝐴 . The set𝐴outliers are underdiagnosis errors. The set𝐵outliers are overdiagnosis errors. We built the prognostic system to find the patients with severity, so we rejected the outliers from the set𝐵. Let the set𝑂 𝐵 = {𝑜 𝑖 : 𝑜 𝑖 ∈𝐵∩ int conv 𝐴 , 𝑖 = 1,𝑚 𝑂 𝐵 } be the set of outliers from 𝐵. After rejecting, we get a new set 𝐵 󸀠 =𝐵/𝑂 𝐵 . If you build the expert system for differential diagnosis, you reject outliers out of the set where there are less of them. If the percentage of rejected points is more than the significance level 𝑚 𝑂 𝐵 𝑚 𝐵 >𝛼, (12) the next (the most informative) factor was added. The space dimension is increased by 1. In the new space convex hulls were built and the outliers were rejected. The space dimension was increased until preassigned significance level. If all available diagnostic information was used, but preassigned significance level was not reached, then decision of not suffi- cient information was taken. When preassigned significance level was reached, we found the separating hyperplanes. The algorithm for modelling the prognostic system is represented on the Figure 1. The results were checked in the control group and the hyperplane with maximal sensitivity was chosen. The complexity of this algorithm is 𝑂(𝑚𝑛+1) [19] if the convex hulls are built by search of all combinations of points. The complexity of this algorithm is 𝑂(𝑚2) if the convex hulls are built by Jarvis march or “gift wrapping” algorithm [20]. 3. Results 3.1. The Expert System of Predicting the Presence of Severity in Patients with Acute Pancreatitis. The research involved 60 persons with severe and 28 patients with nonsevere acute pancreatitis. Among them, there were 57 (64.8%) men and 31 (35.2%) women. The mean age was 48.54 years (±15.18) in males and 56.21 (±17.91) in females. The most common etiology was alcohol consumption (48.3%), followed by gallstones (34.2%). In 17.5% no identifiable cause was found. + − Find the outliers O + − Decision of not sufficient Begin End Reject the outliers B󳰀 = B/O Find Lp n := 1 mo mB > 𝛼 n := n + 1 n < n0 information Choose Lmaxp : Se(Lmaxp ) Build convA, convB Find convB󳰀 = max {Se(Lp)} Figure 1: Algorithm for modelling the prognostic system. The diagnostic criteria for acute pancreatitis were those defined by the 2006 AP Guidelines, as the presence of at least two of the following features: (1) characteristic abdominal pain, (2) elevation over 3 times the upper normal limit of serum amylase/lipase, and (3) characteristic features on com- puter tomography (CT) scan [21]. Severe acute pancreatitis was diagnosed according strictly to Atlanta criteria: Early Prognostic Scores, APACHE II ≥ 8, Ranson ≥ 3; Organ Failure, systolic pressure < 90 mmHg, creatinine > 2.0 mg/L after rehydration, PaO 2 ≤ 60 mmHg; Local Complications (on CT scan), Necrosis, Abscess, and Pseudocyst [22]. Patients were divided into two samples—training (50 patients with severity and 20 without them) and control (10 patients with severity and 8 without). The level of significance was 𝛼 = 0,01. The algorithm presented above was used for patients with training set. For 𝑛 = 1, the percentage of outliers was 29.5%. For 𝑛 = 2, the percentage of outliers was 3%. For 𝑛 = 3, the percentage of outliers was 1.4%. For 𝑛 = 4, the percentage of outliers was 0%. We got 8 hyperplanes which separate the convex hulls of the training samples. Two of them had higher sensitivity and specificity (we got only 1 (6%) of underdiagnosis errors 4 Journal of Computational Medicine and there were no overdiagnosis errors for the control sample with these hyperplanes): −18937.5𝑥 1 −200.3𝑥 2 +5007.3𝑥 3 +42348𝑥 4 +310958.6 = 0, −4802.5𝑥 1 −142.8𝑥 2 +5007.3𝑥 3 +2158.1𝑥 4 +177176.4 = 0, (13) where 𝑥1 is time before hospitalization, 𝑥2 is blood lipase, 𝑥3 is amylase urine, and 𝑥4 is BMI. So, we built the expert system with sensitivity Se = 94%. Statistical errors are seen only in 1 patient in the con- trol group, who were diagnosed with interstitial edematous pancreatitis development on the background severe diabetes mellitus. According to Expert System the acute pancreatitis without severity was predicted. This error, in our view, is associated with late ambulation of the patient for medical care as a result of atypical course of acute pancreatitis, increased blood and urine amylase, and increased BMI, which is the characteristic signs of diabetes mellitus. That is, in this case, some of the most important prognostic parameters of acute pancreatitis have been characterized by different diseases in particular of diabetes mellitus, which caused the error. 4. Conclusions The method of separation of convex hulls in Euclidean space by constructing a separating hyperplane parallel to one of the convex hulls hyperfaces is proposed. On the basis of this method the algorithm for modelling the prognostic system is stated. The proposed algorithm is applied in practice to predict the presence of severity in patients with acute pancreatitis and gives 94% correct results for the control sample, while diagnostic and the predictive probability of existing laboratory and instrumental diagnostic markers and rating scales does not exceed 70–80%. Clinical application of the developed mathematical model predicting the severity of acute pancreatitis promotes proper choice of treatment tactics and allows improving final results of these patients’ treatment. Conflict of Interests The authors declare that there is no conflict of interests regarding the publication of this paper. References [1] E. J. Balthazar, “Acute pancreatitis: assessment of severity with clinical and CT evaluation,” Radiology, vol. 223, no. 3, pp. 603– 613, 2002. [2] G. Sathyanarayan, P. K. Garg, H. K. Prasad, and R. K. Tandon, “Elevated level of interleukin-6 predicts organ failure and severe disease in patients with acute pancreatitis,” Journal of Gastroenterology and Hepatology, vol. 22, no. 4, pp. 550–554, 2007. [3] A. K. Khanna, S. Meher, S. Prakash et al., “Comparison of ran- son, Glasgow, MOSS, SIRS, BISAP, APACHE-II, CTSI scores, IL-6, CRP, and procalcitonin in predicting severity, organ failure, pancreatic necrosis, and mortality in acute pancreatitis,” HPB Surgery, vol. 2013, Article ID 367581, 10 pages, 2013. [4] R. A. Fisher, Contributions To Mathematical Statistics, John Wiley & Sons, 1950. [5] E. V. Gubler and A. A. Genkin, Primenenie Neparametricheskih Kriteriev Statistiki v Medico-Biologicheskih Issledovanijah (Using the Nonparametric Statistic Criteria in Medical and Biological Researches), Leningrad, Russia, 1973. [6] A. I. Borodulin, A. V. Sviridov, and O. V. Sudakov, “Using the cluster analysis in the in the diagnostic of ischemic heart disease,” Prikladnie Informacionnie Aspekti Medicini, no. 2, pp. 28–32, 2008. [7] D. E. Schaubel and J. Cai, “Analysis of clustered recurrent event data with application to hospitalization rates among renal failure patients,” Biostatistics, vol. 6, no. 3, pp. 404–419, 2005. [8] M. Weatherall, P. Shirtcliffe, J. Travers, and R. Beasley, “Use of cluster analysis to define COPD phenotypes,” European Respiratory Journal, vol. 36, no. 3, pp. 472–474, 2010. [9] A. Mahr, S. Katsahian, H. Varet et al., “Revisiting the classi- fication of clinical phenotypes of anti-neutrophil cytoplasmic antibody-associated vasculitis: a cluster analysis,” Annals of the Rheumatic Diseases, vol. 72, no. 6, pp. 1003–1010, 2013. [10] O. L. Mangasarian, W. N. Street, and W. H. Wolberg, “Breast carter diagnosis and prognosis via linear programrnig,” Opera- tion Research, vol. 43, no. 4, 1995. [11] A. I. Golikov, Y. G. Evtushenko, and S. Ketabchi, “On families of hyperplanes that separate polyhedra,” Computational Math- ematics and Mathematical Physics, vol. 45, no. 2, pp. 238–253, 2005. [12] A. R. Alimov and V. Y. Protasov, “Separation of convex sets by extreme hyperplanes,” Fundamental and Applied Mathematics, vol. 17, no. 4, pp. 3–12, 2012. [13] I. I. Eremin, “Fejér methods for the strong separability of convex polyhedral sets,” Russian Mathematics, no. 12, pp. 33–43, 2006. [14] A. V. Ershova, “Algorithm of the disjoint convex polyhedral separating using Fejer mappings,” Sistemi Upravlenia I Informa- cionnie Tehnologii, vol. 1, no. 35, pp. 53–56, 2009. [15] C. Böhm and H.-P. Kriegel, “Determining the convex hull in large multidimensional databases,” in Data Warehousing and Knowledge Discovery, vol. 2114 of Lecture Notes in Computer Science, pp. 294–306, Springer, 2001. [16] G. Fung, R. Rosales, and B. Krishnapuram, “Learning rankings via convex hull separation,” in Proceedings of the Annual Conference on Neural Information Processing Systems (NIPS ’05), pp. 395–402, December 2005. [17] A. N. Kolmogorov and S. V. Fomin, Elements of the Theory of Functions and Functional Analysis, Courier Dover, 1999. [18] D. G. Altman, Practical Statistics for Medical Research, Chapman & Hall/CRC, London, UK, 1991. [19] S. A. Abramov, Lectures at Algorithmic Complexity, MTSNMO, Moscow, Russia, 2009. [20] F. Preparata and M. Shamos, Computational Geometry—An Introduction, Springer, 1985. [21] P. A. Banks and M. L. Freeman, “Practice guidelines in acute pancreatitis,” The American Journal of Gastroenterology, vol. 101, no. 10, pp. 2379–2400, 2006. [22] E. L. Bradley III and C. F. Frey, “A clinically based classification system for acute pancreatitis: summary of the International Symposium on Acute Pancreatitis, Atlanta, Ga, September 11 through 13, 1992,” Archives of Surgery, vol. 128, no. 5, pp. 586– 590, 1993. Submit your manuscripts at http://www.hindawi.com Stem Cells International Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 MEDIATORS INFLAMMATION of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Behavioural Neurology Endocrinology International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Disease Markers Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 BioMed Research International Oncology Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Oxidative Medicine and Cellular Longevity Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 PPAR Research The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Immunology Research Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Journal of Obesity Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Computational and Mathematical Methods in Medicine Ophthalmology Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Diabetes Research Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Research and Treatment AIDS Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Gastroenterology Research and Practice Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Parkinson’s Disease Evidence-Based Complementary and Alternative Medicine Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com 
work_wsrlm3opsvaohmqob2rfqpflsq	----	ESWA Evaluating an Intelligent Diagnosis System of Historical Text Comprehension Grammatiki Tsaganoua, Maria Grigoriadoua, Theodora Cavourab, Dimitra Koutraa aUniversity of Athens, Department of Informatics and Telecommunications, GR-15784, Athens, Greece, e-mails: gram@di.uoa.gr, gregor@di.uoa.gr, mik@altec.gr b University of Thessaly, Dept. of Education, Argonafton & Filellinon strs, GR-38221, Volos, Greece, e-mail: theokav@pre.uth.gr Abstract This work aims to present and evaluate a Fuzzy-Case Based Reasoning Diagnosis system of Historical Text Comprehension (F-CBR-DHTC). The synergism of fuzzy logic and case based reasoning techniques handles the uncertainty in the acquisition of human expert’s knowledge regarding learner’s observable behavior and integrates the right balance between expert’s knowledge described in the form of fuzzy sets and previous experiences documented in the form of cases. The formative evaluation focused on the comparison of the system’s performance to the performance of human experts concerning the diagnosis accuracy. The system was also evaluated for its behavior when using two different historical texts. Empirical evaluation conducted with human experts and real students indicated the need for revision of the diagnosis model. The evaluation results are encouraging for the system’s educational impact on learners and for future work concerning an intelligent educational system for individualized learning. Keywords: Fuzzy Case based reasoning; Historical text comprehension; Diagnosis evaluation; Learner model. 1 Introduction In an Intelligent Tutoring System (ITS), learner diagnosis process imitates the human expert’s process of inferring the student’s internal characteristics from his observable behaviour (VanLehn, 1988). In the domain of comprehension of history, this computational diagnostic process imitates a human expert’s ability to estimate how the learners comprehend the historical text. An attempt towards this direction is our previous work concerning the Learner Model of learner’s cognitive profiles for Historical Text Comprehension (LMHTC) (Tsaganou et al, 2003). Such a learner model demands knowledge acquisition from a human expert, which means knowledge extraction regarding the student’s observable behaviour, its complete and accurate description and transfer to a knowledge base and sometimes to an inference engine. The main obstacle in this process is the uncertainty derived not only from the knowledge communication among the developer, the human expert and the system, but also from inaccuracy of the information captured and of approximation involved in all process steps (Richter, 2001). Case Based Reasoning (CBR) is claimed to be a paradigm that is more akin to the human way of solving complex diagnostic problems in domains like medicine or law. A human expert solves a diagnostic problem using rules derived from his previous experience-cases, whereas a novice requires complete and concrete rules. CBR integrates the right balance between hard to acquire expert knowledge and more easily acquired knowledge in the form of cases. So, for building of an ITS, CBR helps more easily than other methods to overcome problems of knowledge acquisition from the expert. CBR has been proposed for a variety of diagnostic applications (Kolodner, 1993), has been used in educational systems such as in CELIA, for modelling the memory and reasoning capabilities of a novice (Kolodner, 1993), in Engines for Education for case based coach (Schank et al, 1994), in Tutoring and Help Systems (Weber et al, 1998; Burke et al, 1994), in SYIM for distance education (Tsinakos et al, 2001), as part of the student modelling process in ITS systems (Shiri et al, 1998). For overcoming complex problems, like uncertainty in knowledge acquisition, developers recently build more hybrid case-based and knowledge-based systems than pure CBR systems. Fuzzy logic is designed to operate with linguistic expressions and express imprecision and subjectivity in human thinking. Fuzzy logic contributes to CBR in overcoming problems of managing the uncertainty and problems concerning case adaptation by improving performance of case retrieval (Jeng et al, 1995). In research community the interests of fuzzy logic and CBR for diagnosis recently intersect (Dubois et al, 1997; Richter et al, 1999; Hansen, 2000). Fuzzy logic is widely used in student modelling where variables are continuous, imprecise, or ambiguous. Fuzzy-based techniques have been used in educational systems for flexible case-based querying (Calmes et al, 2002). Often, we think of evaluation in terms of how well AI systems, perform. Evaluation of a student model tries to answer a question that is central to cognitive science, AI and education (Littman et al, 1988): What is the relationship between the architecture of the student model and its behavior? The evaluation methods can be used to construct an accurate picture of this relationship. The issue of noise, for example, must be addressed by any realistic system performing pedagogical diagnosis (Wenger, 1994). Three sources of noise are considered: First, noise in the data, which means non-consistency of student’s behavior over the time. Second, noise in the diagnostic process, which means inherent ambiguities in the diagnostic process. Third, noise in the model of communicative knowledge, which means that noise itself can become a source of information. Evaluating a CBR system involves two separate processes, called verification and validation (Watson, 1997). Verification is concerned with building the system right, ensuring that the system gives correct answers. Validation is about building the right system, the one that the users want. The notion of evaluation, in this work, is seen as evaluating how close the diagnostic results of a Fuzzy-Case Based Reasoning system are to that of experts (Althoff, 1994; Althoff, 1995; Chin, 2001). In this contribution we use Fuzzy Case-Based Reasoning technique for the construction of a system for Diagnosis of students’ cognitive profiles of Historical Text Comprehension. In Section 2 the underlying learner model LMHTC is presented. In Section 3, the F-CBR-DHTC system is described. In Section 4, the user interface, system implementation and preparation for use are presented. In section 5, we focus on the evaluation by human experts of system’s diagnosis accuracy and of its performance when using a new historical text. We also describe the evaluation results and how they were used for the revision of the diagnostic model. In section 6, we conclude, present lessons learned and give links to ongoing research. 2 The Learner Model of Students’ HTC 2.1 Models of HTC Comprehension of historical text is a special kind of the complex and interactive cognitive process (Briton , 1996). Historical text is defined as a causal and transformational system. It is characterised as transformational because it describes the representation of a historical change, which has happened in a particular place and time. It is characterised as causal because it describes an historical occurrence, which is interpreted by a series of causal links. The reader utilises certain fundamental cognitive categories for establishing and organising the meaning of the text (Baudet et al, 1992). During comprehension the reader attributes meanings to causal connections between occurrences in the historical text (Cavoura, 1994). In the level of comprehension as a cognitive task, the learner composes a representation of the historical text. This representation is a system, which contains the cognitive categories: event, state and action. Comprehension of the historical text is associated with causal connections and arguments made by the reader. The arguments are based on the three cognitive categories. For the interpretation of the learner’s cognitive processes we analyse his discourse tracing the recognition or not of the three cognitive categories. 2.2 The LMHTC The underlying model of the F-CBR-DHTC system is the LMHTC model, which is based on MOCOHN, a pencil- and-paper diagnosis model, and on further experimental research (Cavoura, 1994; Tsaganou et al, 2003). The LMHTC presents to the learner a historical text in the appropriate form and question-pairs with alternative answers. The historical text includes factors/ instances, which represent the 3 cognitive categories action, state and event. For every factor a question-pair is submitted to the learner. Each question-pair consists of two questions concerning the same factor. The first question in the question-pair is relative to the learner’s position about the significance of this factor and the second question is relative to the learner’s the justification of this position. The learner has to use the given alternative answers, which correspond to position and justification, in order to express his position for certain historical issues and support it by selecting a justification. The answers concerning position and justification are classified as scientific, towards- scientific and non-scientific, as they are depicted in Table 1 and Table 2. The alternative answers reflect scientific thought, towards acquiring scientific thought, and non-scientific thought. Table 1 Classification of answers concerning position Position Answers Scientific learners attribute minimum importance to events and states and maximum or medium importance to actions towards-scientific learners attribute medium importance to events and states non-scientific learners attribute maximum importance to events or minimum importance to actions Table 2 Classification of answers concerning justification Justification Answers Scientific learners ground their answers up on the scientific historical thought towards-scientific learners base their answers on the common sense schemas expressing experience, quantity, continuity and attitudes, which means learners are towards acquiring scientific thought non-scientific learners give cyclic answers bas ed on recycling the questions and consequently they refer to non- scientific thought Figure 1 depicts a historical text concerning 5 different factors of the outbreak of French Revolution. In the historical text, one factor represents the cognitive category event, one the state and three factors represent the action. Figure 1: A screenshot depicting a historical text concerning the outbreak of French Revolution, question-pair number 1, alternative answers and characterizations of the answers. Examples of alternative answers considered, which refer to positions and to justifications are also depicted in Figure 1. The answers a1 to a3 are alternative answers to question 1a concerning the position, whereas the answers b1 to b4 are alternative answers to question 1b concerning the corresponding justification. Figure 2, also indicates the characterisations of the answers, which are not visible to the learner. Answer a1 is non-scientific, a2 is toward s-scientific and a3 is scientific answer. Answer b1 is towards-scientific expressing experience, b2 is scientific, b3 non-scientific and b4 is towards-scientific expressing quantity, see Table 2. 2.3 Arguments For every question-pair the combination of the learner’s position and the corresponding justification constitute the learner’s argument. An argument is defined as complete when both position and justification are scientific. Otherwise the argument is non-complete. The expert defines the different degrees of argument completeness. The argument completeness, which is associated with the recognition or not of an instance of a cognitive category, is used as a vehicle to reveal the degree of the recognition or not of the corresponding cognitive category. Table 3 demonstrates all possible combinations of position-justification pairs, the corresponding argument completeness and characterization and the degree of recognition of a cognitive category. Possible values of the argument completeness are: complete, almost complete, intermediate, nearly incomplete and incomplete. For the characterization of argument completeness justification weights more than position does. So for example, towards-scientific position and scientific justification result in almost complete argument, whereas the opposite one that is scientific position and towards-scientific justification result in nearly incomplete argument. Argument completeness describes the learning difficulties of the learner concerning the cognitive categories, which the learner does not recognize. This qualitative characteristic of the arguments reflects the degree of recognition of a cognitive category that is the degree of comprehension of the historical text. Table 3 Argument completeness values concerning position -justification combinations. Position Justification Argument Argument Status of recognition Completeness characterization of the cognitive category scientific scientific complete scientific recognition towards scientific scientific almost complete towards-scientific towards-recognition non-scientific scientific intermediate towards-scientific towards-recognition scientific non-scientific nearly incomplete towards-scientific towards-recognition scientific towards-scientific nearly incomplete towards-scientific towards-recognition towards-scientific towards-scientific nearly incomplete towards-scientific towards-recognition non-scientific towards-scientific incomplete non-scientific non-recognition towards-scientific non-scientific incomplete non-scientific non-recognition non-scientific non-scientific incomplete non-scientific non-recognition 2.4 Classification of the cognitive categories Historical actions constitute the core of the historical discourse. According to relevant research, during the comprehension of the historical text, the recognition of the cognitive category action is more important than the recognition of the cognitive category state (Cavoura, 1994). The recognition of the cognitive category event is less important than the recognition of the cognitive category state. This results in the classification of the cognitive categories. Table 4 demonstrates the quality values of the cognitive categories. Possible values of the quality are superior for the action, medium for the state and inferior for the event. Table 4 Quality values of the cognitive categories Cognitive category Category quality action sup erior state medium event inferior 2.5 The Cognitive Profiles of HTC The status of recognition of the three cognitive categories: event, state and action are used to formulate the cognitive models of HTC, which reflect the learners’ levels of historical thought (Tsaganou, 2002). The learner’s cognitive profile of HTC is formulated taking into account the number of his complete arguments. The cognitive profile expresses the recognition or not of the cognitive categories. Table 4 depicts the cognitive models and the cognitive profiles of HTC. The general categories of cognitive models considered are Historical Thought (HT), Towards Acquiring Historical Thought (TAHTnx) and Non-Historical Thought (NHT). TAHTnx cognitive models are categorised in more detail according to the number n of recognised by the learner cognitive categories and to the number x of their instances in the historical text. TAHT1 means that the learner recognises 1 instance of a cognitive category, whereas TAHT1x means that the learner recognises x instances of a cognitive category, where x>1. The same stands for TAHT2, TAHT2x, TAHT3 and TAHT3x. The numbers n and x are used for the formulation of the learner’s cognitive profile. Learners with Very Low profile seem to have serious difficulties in thinking historically. Learners characterized by terms like low, nearly low, below intermediate, above intermediate, nearly high and high, seem to encounter difficulties in thinking historically. Learners with very high profile seem to have no learning difficulties in thinking historically. 2.6 The Profile Descriptor The profile descriptor describes the learner’s cognitive profile and denotes different perspectives of the profile. The learner’s profile descriptor is formulated taking into account all of his arguments which may have different degree of completeness. The profile descriptor carries detailed description pertaining to the quality of the cognitive categories and the completeness of the arguments, which are attached to every cognitive profile. The profile descriptor, which models uncertainty associated with observation, depicts the learner’s problems in the recognition of the cognitive categories and reflects his learning difficulties. For example, selection of the answers a3 and b2 of figure 1 constitutes a complete argument of medium category and indicates the recognition of one instance of the cognitive category (in this example the category state). Selection of the answers a2 and b4 constitutes a nearly incomplete argument of medium category, which indicates non-recognition of the cognitive category. A nearly low cognitive profile of a learner, during comprehension of a historical text with 5 factors and 5 corresponding question-pairs, can be accompanied by the following profile descriptor: “The learner gives one complete argument of inferior category, one nearly incomplete argument of superior category, one nearly incomplete argument of superior category, one incomplete argument of superior category and one incomplete arg ument of medium category". Table 5 Cognitive models, number and types of recognised cognitive categories and corresponding cognitive profiles, n in {1,2,3} and x>1. Cognitive models Number and types of recognised by the learner cognitive categories Cognitive profiles NHT no event or state or action very low TAHT1 1 event or 1 state or 1 action low TAHT1x more than one events or more than one states or more than one actions nearly low TAHT2 (1 event and 1 state) or (1 event and 1 action) or (1 state and 1 action) below intermediate TAHT2x (more than one events and states) or (more than one states and actions) or (more than one events and actions) above intermediate TAHT3 1 event, 1 state and 1 action nearly high TAHT3x more than one events, states and actions high HT all events, states and actions very high 3 The F-CBR-DHTC System Figure 2. Structure of the F–CBR-DHTC system The F-CBR-DHTC system is a hybrid Fuzzy-CBR system, which implements the diagnostic module of the LMHTC. It handles the complexity of diagnosis of student’s cognitive profile and profile descriptor. The system encourages the learner to read the historical text and answer to question-pairs selecting from the alternative answers. The learner’s responses define his observable behavior. The F-CBR-DHTC system, Figure 2, solves the diagnostic problem in two stages: (1) the Fuzzy inference stage, which infers the arguments’ completeness reflecting the degree of recognition of the cognitive categories and (2) the Fuzzy-CBR infe rence stage, which infers the learner’s cognitive profile and profile descriptor. 3.1 The Fuzzy Inference stage The Fuzzy inference stage uses fuzzy rules, which incorporate the description of expert’s knowledge concerning the student’s answers to questions (position-justification pair values) and infers the argument completeness. This stage also expresses argument completeness with fuzzy sets, which are used in the next stage for defining similarity values. In this stage the symbolic input data (student’s responses) are transformed into linguistic terms (Jeng et al, 1995). A={A1, A2,..., An} and B={B1, B2,..., Bn} are the types of responses: where A1, A2,..., An concern position, and B1, B2,..., Bn concern justification. The term set of An is T(An)={Dn1, Dn2,..., Dnk} with k linguistic values and the term set of Bn is T(Bn)={En1, En2,..., Enl} with l linguistic values. The sets TA= {T(A1), T(A2),...,T(An)} and TB={T(B1), T(B2),...,T(Bn)} are the fuzzy sets of all term sets that represent the observable behaviour. The symbolic input is fuzzified by means of the linguistic variable values Dnk and Enl. The numbers k, l, Dnk and Enl are defined by the developer with the help of the human expert. Student’s argument completeness Cn, result as combinations of the independent input data: position and justification and is represented as linguistic variables with m linguistic values and corresponding term set: T(Cn)={Cn1, Cn2 ,…, Cn m). Let consider an example, of a historical text with n=5 arguments, k=3 linguistic values concerning position, l=3 linguistic values concerning justification and m=5 linguistic values concerning argument completeness. Then the sets are A1= position for event1, B1= justification for event1 and C1=argument for event1 and the term sets are T(A1)= T(position for event1)={D11, D12, D13} ={right, mediocre, wrong}, T(B1)= T(justification for event1)={E11, E12, E13}= {right, mediocre, wrong}, T(C1)=T(argument for event1)={C11,C12,C13,C14,C15}= {incomplete, nearly incomplete, intermediate, almost complete, complete}. IF An is Dnk AND Bn is Enl THEN Cn is Cn m (1) Using fuzzy rules of the form (1), which take into account position and justification values, this stage infers the corresponding argument completeness. Each fuzzy rule results in one fuzzy set. Figure 3, demonstrates values of the membership functions that express to what degree the values in the universe of discourse belong to the fuzzy set “how complete an arg ument for event1 is”. The fuzzy sets are used by the next stage of the system to express local similarity measures. The output of the fuzzy inference module is the n-dimensional vector Sn, in [0,1], of student’s argument characteristics Sn, which constitutes the input data to the Fuzzy-CBR inference module. 0 0,2 0,4 0,6 0,8 1 1,2 1 2 4 7 10 discretized universe of discourse ì complete almost complete intermediate nearly incomplete incomplete Figure 3. Discrete fuzzy sets for the inferior category argument for event1 with respect to “how complete the argument for event1 is” 3.2 The Fuzzy-CBR Inference stage The aims of Fuzzy-CBR inference stage are to define case structure using the argument characteristics, define similarity measures based on the fuzzy sets, which have been defined in the previous stage, and proceed to the case retrieval and case adaptation process (Bergmann et al, 1998; Leake, 1996). In the Fuzzy-CBR inference stage the learner’s behavior represented by the argument completeness constitutes the corresponding case. A case is viewed as a set of attributes, which contains (apart from the case name) two non-empty sets (Dubois et al, 1997): the problem description attributes subset S (argument completeness) and the solution description attributes subset T (cognitive profile and profile descriptor). Expressed in terms of the fuzzy-CBR framework, the case is declared in the following form: <casename, S1, S2,..., Sn, T1, T2>. In this stage, reasoning steps are based on the hypothesis that similar problems have similar solutions. The challenging problem is to determine the degree to which stored cases are similar (Aamodt et al, 1994). The local similarity measures between argument characteristics of two cases are calculated according to the previously defined fuzzy sets. The global similarity between two cases is computed using a fuzzy k-nn algorithm. A fuzzy k-nn algorithm facilitates the retrieval process from the case base, handles case adaptation by exploiting the defined as fuzzy sets similarity values between the arguments characteristics (Hansen, 2000). A learner’s cognitive profile and profile descriptor are inferred by the Fuzzy-CBR inference module after comparisons of the arguments of a case with the corresponding arguments of the other learners stored in a case base. The solution description attribute subset T= {T1, T2,…,Tn} of the problem description vector S of current case q contains the attribute-values and T is the diagnosis - the requested value of cognitive profile and the profile descriptor (Aamodt, 1995; VanLehn, 1988). The retrieval process finds a case with an information vector S' most similar to S. Once a matching case is retrieved the system it is used to form the solution T. 4 System Preparation for Use 4.1 The environment of the F-CBR-DHTC system The F-CBR-DHTC system has been implemented using an object-oriented language (C++) combined with a powerful graphical interface builder (Toolbook Instructor). The implementation method was chosen for its ability to handle a case base and its flexibility in solving problems. The environment of F-CBR-DHTC provides the users with an easy-to-use interface through which (a) the learners are given the historical text to read and questions with alternative answers to select according to their opinion and (b) the experts or the history teachers can author a novel environment by adding a new historical text with questions an alternative answers. 4.2 Case Base Initialisation For the case base initialisation was used a sample of 60 most representative cases of the problem domain. We experimented using a historical text concerning the outbreak of French Revolution and 5 questions-pairs with alternative answers. We conducted a research with 40 high school students and appropriate historical text and questions in order to have a sample of the distribution of cases to cognitive profiles. The cognitive profiles were judged by hand. Taking into account the experimental results we identified that the frequency of cases with Very Low, Low or Nearly Low cognitive profiles is greater than others. Consequently, as most of the students are expected to have Very Low, Low or Nearly Low cognitive profile, we decided the majority, almost 70%, of initial cases in the case base to be cases belonging to the corresponding subgroups. The system learned the domain of diagnosis of student’s HTC from 60 cases: 40 from episodic cases (real students) and 20 prototypical cases. We used weights to indicate which cases merit greater attention as prototypic cases. 4.3 Improving the Scaling Validity During the preliminary trials, presenting the system with 20 new cases we tested the scaling validity and the performance of the diagnostic system. We divided the test cases into 4 groups of 5 cases each, to assess the improvement in system’s performance as it gained experience during the learning process. After each case was entered, the system attempted to assign the correct cognitive profile and profile descriptor to each case. If the model assigned to a case a profile descriptor, which is not correct we added that case into the case base. F-CBR-DHTC, as an incremental knowledge acquisition system, demonstrated satisfactory performance in the four test series. 5 Formative Evaluation of the System We derived the following questions in order to give structured overview of the used evaluation criteria of the diagnosis system. (1) Does the Fuzzy-CBR system models human behaviour in a more useful way? (2) What is the system’s behaviour in novel environment? The evaluation criteria concern noise, user acceptance, flexibility, effectiveness, correctness, consistency and performance. Formative evaluation aimed at evaluating noise handling in diagnosis accuracy and the system’s performance when using a differently structured historical text. 20 High school students, which attend the same public school, and 4 human experts participated in the evaluation. Participation in the experiment was voluntary. 5.1 Evaluation of Noise Handling In a CBR diagnostic system the problem of noise, which means inherent ambiguities, can manifest itself as two or more cases with identical diagnosis but different inputs (Althoff, 1995). In our diagnostic system of students’ cognitive profiles and profile descriptors the diagnosis accuracy has to do with handling the noise in occurrences like O: two or more students with identical cognitive profile have different answers to questions and different profile descriptors. To account for noise within the CBR system’s memory, we count the number of occurrences like O and change a certain amount of attributes in the involved cases. During formative evaluation and after the trials, we tested F-CBR-DHTC system for its diagnosis accuracy by 4 human experts in AI. To test our systems’ ability to handle noisy data during the adaptation process we conducted a research. The goal of the evaluation research was to examine the extent to which the system and the experts’ decisions agree. In this evaluation stage we presented the system with 25 new (potentially real) cases with slightly different to real cases attributes. After the diagnosis, occurrences like O were found in 20 out of the 25 cases. The system was tested by the human experts for its ability to handle noise. The 4 participants engaged independently in the evaluation and were asked to diagnose the cognitive profiles based on the profile descriptors and judge explaining their view. In most cases there was agreement between the evaluation by the system and the evaluation by the experts. 0 1 2 3 4 5 6 Cognitive profiles N um be r o f c as es expert1 expert2 expert3 expert4 F-CBR-HTC expert1 4 5 3 3 2 2 1 expert2 4 5 3 3 2 2 1 expert3 5 4 3 2 3 3 expert4 4 5 2 3 2 2 F-CBR-HTC 5 5 3 3 2 2 VL L NL BI AI NH H VH Figure 4. Number of cases per corresponding cognitive profile with agreement in the assessment of the profile descriptor using different methods of assessment: 4 human experts and the F-CBR-HTC diagnostic system for 20 cases. The horizontal axis shows the cognitive profiles {Very Low, Low, Nearly Low, Below Intermediate, Above Intermediate, Nearly High, High, Very High}, which correspond to {VL, L, NL, BI, AI, NH, H, VH}. There were small disagreements in assessing the cognitive profiles, when assessment was based on the profile descriptors. From the results demonstrated in Figure 4 we can observe that the estimation made by the F-CBR-DHTC diagnostic system and the human experts coincide (on average) nearly in 17 out of the 20 cases. In more details, expert1 coincide in 18 out of 20, expert2 in 19 out of 20, expert3 in 15 out of 20 and expert4 in 17 out of 20 cases. Regarding the number of students, which were judged as having very low profile, see figure 4, it is remarkable that 3 out of 4 experts judged that one of the 5 very low students is having low profile, nevertheless he has not recognized a cognitive category. To localize the main points of disagreement between the system and the experts, we consider the explanations given by the experts concerning their judgments. POINT1: The disagreement between experts and the system has to do with considerations of qualitative and not only quantitative characteristics of the profile descriptors as the F-CBR-DHTC system does. For example, an almost complete argument means that the student is about to recognize the corresponding cognitive category. Seen it in terms of learning difficulties, this student is very close to overcome his learning difficulties concerning the cognitive category. So the student is in a higher level than a student with an incomplete argument POINT2: The disagreement has to do with accidental answers. For example, the student may have accidentally selected the scientific answers for a factor, as it results from the rest of his answers. Figure 5 depicts analytically the profile descriptors of the five students, which were judged by the system as having very low profile. It also depicts the profile descriptors of the students S1, S2, S3, S4 and S5. Student S5 has given two almost complete arguments, that is the reason why the 3 experts judged him as having low profile. Profile Descriptors of students with Very Low cognitive profile 0 0,5 1 1,5 2 2,5 3 3,5 4 4,5 state action1 event action2 action3 Instances of cognitive categories A rg um en t c om pl et en es s s1 s2 s3 s4 s5 Figure 5. Comparison between the profile descriptors of five students S1, S2, S3, S4 and S5 having Very Low profile according to the F-CBR-DHTC system. The vertical axis shows argument completeness and {1,2,3,4,5} correspond to {incomplete, nearly incomplete, intermediate, almost complete and complete} Formative evaluation makes us differentiate the students with low profile for example, into two groups: low and low+ according to how close to recognition of cognitive categories they were. Low represents the students that have recognized one cognitive category and low+ represents the students that have recognized one cognitive category but are very close to recognize another and as a result very close to the nearly low profile, which is cognitive profile of a higher level. The evaluation results were used for the revision of the diagnostic rules taking into account POINT1 and POINT2. After re-initialisation of the case base the revised F-CBR-HTC system becomes more accurate in diagnosis process and can indeed perform diagnosis in a way that gives results similar to the way human experts evaluated students. Consequently the computational system, which imitates human behaviour in a more useful way, obtains user acceptance by human experts. Table 6 Number and types of recognised cognitive categories and revised cognitive profiles after formative evaluation. Number of cognitive categories full or closely recognised Cognitive profiles no recognition of any cognitive category very low close to recognition of one cognitive category very low+ recognition of one cognitive category low close to recognition of more than one instance of a cognitive category low+ recognition of more than one instance of a cognitive category nearly low close to recognition of two cognitive categories nearly low+ recognition of two cognitive categories below intermediate close to recognition of more than two instances of two cognitive categories below intermediate+ recognition of more than two instances of two cognitive categories above intermediate close to recognition of three cognitive categories above intermediate+ recognition of three cognitive categories nearly high close to recognition of more than two instances of three cognitive categories nearly high+ recognition of more than two instances of three cognitive categories high close to recognition of all instances of the three cognitive categories high+ recognition of all instances of the three cognitive categories very high 5.2 Evaluation of the system in novel environment Noise in the data may cause inconsistency in students’ behaviour concerning two different historical texts. The evaluation aimed at evaluating the systems’ behaviour in novel environment and at validating the system with new data. During the formative evaluation we tested the system’s effectiveness when using a new differently structured historical text. We also tested the flexibility and openness of the system, which reflect the environmental dependency and the user acceptability as a feature from the user’s interface point of view and time spent to build the application with a new historical text. The first historical text, which had already been used by the system, concerning the outbreak of French Revolution included 5 factors each corresponding to an instance of a cognitive category. Particularly, it included 3 instances of the cognitive category action, 1 instance of the state and 1 instance of the event. The second was a new for the system historical text concerning the outbreak of the 1st World War (WW). An expert in history teaching prepared the new historical text. The text was consisted of 7 factors each corresponding to an instance of a cognitive category. Particularly, it included 4 instances of the cognitive category action, 2 instances of the state and 1 instance of the event. The expert constructed the 7 question-pairs corresponding to positions and justifications and the appropriate alternative answers and through the interface readjusted the structure of the case base. 20 High school students participated in the two stages experiment. In the first stage the system presented to the students the French Revolution-historical text and they were asked to read the text and select from the alternative answers: (1) their position from a 3-point scale stating how strongly each factor they felt was important for the outbreak of French Revolution and (2) the corresponding justification, one out of five proposed, in order to support their position. In the second stage of the research the same students participated and the system presented to the students the 1st WW- historical text and asked the same matters. At the end, the revised F-CBR-HTC diagnostic system assessed the students’ cognitive profiles and profile descriptors for each stage. Comparing the results demonstrated in Figure 5 we can observe that the estimation made by the system in the first stage coincide in 10 out of the 20 students and differs a little in 7 out of 20 students. Even though the sample is rather small to reach a safe conclusion and given the different structure of the two historical texts, the results indicate that the system can give similar assessments for the same student, when performing diagnosis using different historical texts. 0 1 2 3 4 5 6 Students C og ni tiv e pr of ile s FR 1st WW FR 2 1 3 1 2 3 1 2 2 3 3 4 4 5 2 3 2 1 3 1 1st WW 2 1 3 1 3 4 2 4 4 2 4 3 4 3 2 4 2 1 3 1 1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th 11th 12th 13th 14th 15th 16th 17th 18th 19th 20th Figure 6. The cognitive profiles assessed by the revised F-CBR-HTC system for 20 students (1st to 20th) using two different historical texts concerning French Revolution and 1st World War: The vertical axis shows the cognitive profiles {Very Low, Very Low+, Low, Low+, Nearly Low, Nearly Low+, Below Intermediate, Below Intermediate+, Above Intermediate, Above Intermediate+, Nearly High, Nearly High+, High, High+ and Very High}, which correspond to {1,2,3,4,5,6,7,8,9,10,11,12,13,14}. 6 Conclusions In this work, we presented the F-CBR-DHTC system and its formative evaluation, which resulted in the revision of the system. The evaluation focused on the system’s diagnosis accuracy of students’ cognitive profiles of HTC and on its behaviour in novel environment. The empirical evaluation was preformed with the participation of a small number of experts and on a limited number of students. The results were used for the reformulation of the cognitive profiles in a way that contributes to the improvement of the diagnosis accuracy of the system. The results are encouraging for the system’s educational impact on students. The ability of the system to give similar results when using different historical texts holds potential for use in individualized history instruction in an ITS. Lessons learned from developing and evaluating the system concern practical difficulties due to the uncertainty of knowledge acquisition and knowledge communication among the developer and the human expert. Moreover, worth noticing was the effort required for taking design decisions for the construction of the case base structure, for its enrichment with the appropriate episodic and prototypical cases in order to reassure the performance of the system in the novel environment of a new historical text. The evaluation method adopted was based on the particular characteristics of the system and resulted in the revision and improvement of the system. The recently growing interest in opening the learner model to the learner encourages the development of systems that give the learner greater responsibility and control over learning. Results of a successful diagnostic process can be beneficial in guiding individualised learning in history by activating the appropriate for a student learning strategy, i.e. through an interactive dialogue between the student and the system. There are educational benefits of the diagnostic system for the students in changing their reasoning. The open learner models are a useful way of helping the experts to recognise learner difficulties, enabling them to respond to specific learner populations or individuals in appropriate ways. We plan to implement the system as an intelligent educational environment. References Aamodt A. Knowledge Acquisition and Learning by Experience- The Role of Case-Specific Knowledge, from C. Tecuci, Y. Kodratoff (eds.): Machine Learning and Knowledge Acquisition- Integrated Approaches, Academic press, 1995, 8, 99: 197-24. Aamodt, A. & Plaza, E. Case-based reasoning: foundational issues, methodological variations, and system approaches. AI Communications 1994, 7(i):39-59. Althoff K.-D., Evaluating CBR Systems, In: Aadmodt A., Althoff K.-D., Magaldi R. & Mline R. CBR: A New Force In Advanced Systems Development. Tutorial, London, 1995. Baudet S., Denhière G. (1992). Lecture, comprehension de texte et science cognitive, Presses Universitaires de France, Paris. Bergmann, R. & Althoff, K.-D. Methodology for Building CBR Applications. In: M. Lenz, B. Bartsch-Sporl, H.-D. Burkhard, & S. Wess (Eds.). Case-Based Reasoning Technology from Foundations to Applications, Springer-Verlag, 1998. Briton B., Graesser Ar., (1996). Models of Understanding Text, Lawrence Erlbaum, Associates Inc. Publishers, Mahwah, New Jersey. Burke, R. & Kass, A. Refining the Universal Indexing Frame to support retrieval of tutorial stories, Proceedings AAAF94 Workshop on Indexing and Reuse in Multimedia Sy stems, 1994 , 1-11. Calmes M., Dubois D., Hullermeier E. et al. A fuzzy set approach to flexible case-based querying: methodology and experimentation, 8th International Conference on Principles of Knowledge Representation and Reasoning (KR2002), Toulouse, France, 2002. Cavoura Th. Modalités de l' appropriation de la connaissance historique, Thèse de Doctorat, Université de Paris VII, 1994. Chin D., Empirical Evaluation of User Models and User-Adapted Systems, User Modeling and User-Adapted Interaction, 2001, 11: 181-194. Dubois, D., Esteva, F., Garcia, P. et al. Fuzzy modelling of case-based reasoning and decision, Case-Based Reasoning Research and Development, Proceedings of the 2nd International Conference on Case-Based Reasoning (ICCBR-97); Leake, D. B., and Plaza, E. (eds.), Springer Verlag, Berlin, 1997, 599–610. Hansen B. K. Weather Prediction Using Case-Based Reasoning and Fuzzy Set Theory, Phd Thesis, Dalhousie University, Daltech, Halifax, Nova Scotia, 2000. Jeng, B. C., and Liang, T.-P. Fuzzy indexing and retrieval in case-based systems, Expert Systems With Applications, Elsevier Science Ltd., 1995, 8(1): 135–142. Kolodner, J. L. Case-Based Reasoning. Morgan Kaufmann, 1993. Leake, D. B.: CBR in context. The present and future. In: Leake, D. B. (editor) Case-Based Reasoning: Experiences, Lessons & Future Directions, American Association for Artificial Intelligence, Menlo Park California, USA, 1996. Littman D., Soloway E. Evaluating ITSs: The cognitive science perspective. In Polson M., Ridchardson J., Foundations of ITSs, Lawrence Erlbaum Assosiates Inc. Publishers, Hillsdale, New Jersey, 1988. Richter M. M., Web S. Similarity, Uncertainty and CBR in PADEX, Automated Reasoning: Essays in honor of Woody Bledsoe (ed. R. S. Boyer), Kluwer Acad. Publ., 1999, 249-265. Richter, M. M., Burkhard, H.-D. On the Notion of Similarity in Case-Based Reasoning and Fuzzy-Theory, In: Soft Computing in Case-Based Reasoning (eds. S. K. Pal, T. S. Dillon, D. S. Youmd) Springer Verlag, 2001, 29-46. Schank, R. & Cleary, C., Engines for Education, http://www.ils.northwestern.edu/e-for-e/index.html, 1994. Shiri A., Aimeur E., Frasson C. SARA: a case-based student modeling system. Advances in Case Based Reasoning 4th European Workshop, EWCBR 98 Proceedings. Springer Verlag, Berlin, Germany, 1998, 394-403. Tsaganou G., Grigoriadou M., Cavoura Th., Experimental Model for Learners’ Cognitive Profiles of Historical Text Comprehension, International Journal of Computational Cognition, 1(4), 2003 (under publication). Tsinakos A., Margaritis G. Results of employing CBR in SYIM. Learning Technology newsletter, Lov. 3, Issue 4, Oct., 2001. VanLehn K. Student modeling, In Foundations of Intelligent Tutoring Systems, (eds) Polson M., Richardson J., Lawrence Erlbaum Assosiates Inc. Publishers, Hillsdale, New Jersey, 1988. Watson, I. Applying case-based reasoning: techniques for enterprise systems. Morgan Kaufmann, California, 1997. Wenger E. Artificial Intelligence and Tutoring Systems. Computational and Cognitive Approaches to the Communication Knowledge. M. Kaufmann Publishers, Inc., California, 1994. Weber, G., & Schult, T.J. Case-Based Reasoning for Tutoring and Help Systems. In: M. Lenz, B. Bartsch-Spörl, H.-D. Burkhard, Stefan Wess (Eds.): CBR Technology, From Foundations to Applications, Heidelberg: Springer, 1998. 
work_wtbnvu5pjrd4jgueuqhkffr4qq	----	PDF hosted at the Radboud Repository of the Radboud University Nijmegen The following full text is a publisher's version. For additional information about this publication click this link. http://hdl.handle.net/2066/112345 Please be advised that this information was generated on 2021-04-06 and may be subject to change. http://hdl.handle.net/2066/112345 Analytrca Chwnrca Acta, 235 (1990) 21-40 Elsevler Science Pubhshers B V , Amsterdam - Pnnted m The Netherlands 21 Expert system for precision testing in validation of liquid chromatographic methods J A VAN LEEUWEN *, L M C BUYDENS, B G.M. VANDEGINSTE a and G. KATEMAN Department ofAnalytrca1 Chemrstry, Cathok Unrversrty of Nqmegen, Toernooweld, 6525 ED NrJmegen (The Netherlands) M MULHOLLAND Fhlhpps S’crent$c, CambrIdge, CBI ZPiY (Great Bbtarn) (Received 3rd July 1989) ABSTRACT After a hquld chromatographlc method has been developed, It must be vahdated to estabhsh its imutations in daiiy use Method vahdatlon LS becommg increasingly Important as stncter rules are applied by regulatory authontles. Precision testmg 1s a vltai step m thus vahdatlon; both mtraIaboratory testmg and- mteriaboratory testing are needed. iii an mtralaboratory test, repeatability and ruggedness tests are usually done Expert systems are available for both tests Here they are integrated to form an mtralaboratory precision-testing expert system, special mtegratlon archtecture 1s described Important features of the integrated system are a supervlsor contammg planning knowledge about the tests and a common data structure contaming all the objects necessary for an expert system in this area Many applications of artificial intelligence in analytical chemistry have been described recently [l-5]. Although the types of knowledge vary, most applications are found in expert systems for spec- tral interpretation (infrared, ultraviolet, mass or NMR) or elucidation of chemical structures [2-51. The expert system described here was developed as part of a project (Esprit project 1570) that evaluates the use of expert systems in the develop- ment of liquid chromatographic (LC) methods [6]. In this project, method development 1s divided into four domains: first guess of conditions, selec- tion of criteria for optimization [7], optimization of instrumental parameters and operating condi- tions [B] and validation of the method [9-111. In this paper, the structure and implementation of the validation part is described. When a full validation program is run, the five features of performance tested are the accuracy, a Present address Umlever Research, Vlaardmgen, The Netherlands 0003-2670/90/$03 50 0 1990 - Elsevler Science Publishers B.V preclslon, sensitivity, selectivity and limitations of the method. Because a full program for method validation involves testing in different laborato- ries, it would be of little use to try to tackle the entire validation in one expert system, as it would be very difficult to control expert systems located at different sites. The size of such a system would also become very large. Therefore the method validation 1s tackled m parts. The expert system described here is concentrated on precision tests that can be done in the laboratory where the LC method 1s developed. It 1s intended to gve the analyst validating the method as much certainty as possible that this method will not fall m a col- laborative mterlaboratory test. For most analysts in a routine laboratory, the performance of a precision test is not stra&tfor- ward. Decisions must be made about which tests to apply, the extent of each test and the accepta- bility of the results. If a method falls the test, it is also necessary to specify the method again, with consideration of the problem that led to its failure. 28 J A VAN LEEUWEN ET AL Expert systems that advise on parts of the prob- lem of precision testing are available [9,10]. For an analyst to use these systems most beneficially, the basis of the test procedures and their relations to each other must be known. An integrated system based on extstmg systems would eliminate this requirement. An integrated system mvolves several modules of a heuristic as well as an algorithmic nature. Integration of modules with different problem- solvmg techniques, like calculations and produc- tion rules, requires a flexible arcluteture, m whtch the different modules are connected so that they can exchange as much information as possible. It must, for instance, be possible to have modules implemented in the usual techniques for expert systems as well as modules implemented in spreadsheet packages and normal programming languages like C. In prmciple, the architecture should not enforce too many constraints on the structure and implementatton of the modules be- cause this would damage their performance. Espe- cially in this study, where new modules are in- tegrated with existing expert systems, tt is tm- portant not to change the existing systems too much and not to place too many restricttons on the new modules. A feature of the design for mtegratton described here 1s the development of a kmd of backbone for LC expert systems, a data structure that can serve as a basis for many dtffer- ent expert systems on vahdation of LC methods. A framework of concepts describing the basics of LC can be accessed by every module m the in- tegrated system. The ObJects in the common data structure are represented formally on paper so that they can be implemented in any suttable technique for expert systems or in any program- mmg language. By using the common data struc- ture and the modules as budding blocks, an expert system is built on mtralaboratory precision test- mg. Because of tts modular structure, the separate parts of the system can be activated indepen- dently. The system also contains planning knowl- edge on when to activate which module. The repeatability test and the ruggedness test are the mam parts of the integrated system. These tests were developed independently in trials in- volvmg several experts. ‘The integrated system contains the integrated knowledge of these ex- perts. Thts is a typical feature of so-called second-generation expert systems, which contam the knowledge of more than one expert, apply different inference techniques in different parts of the system, and can also select the next part of the system to be consulted. The so-called blackboard architecture is suitable for the implementatton of such systems [12]. PRECISION TESTS The purpose of a precision test is to establish the random devtation from the mean in a certain analysis. Precision testing normally consists of re- peatability and (interlaboratory) reproducibility tests [13]. In a repeatability test on a method, the same sample 1s analyzed (usually 10 or 25 times) under the same conditions by the same analyst. Because the test site does not change, a repeatabil- ity test normally does not require much manpower and time. In contrast, reproductbthty tests are relatively costly; identical samples are tested in different laboratories to examine the precision of the method under slightly changing condittons. To avoid excessive problems during a reproducibility test, a ruggedness test can be added to the preci- sion test [14]. In a ruggedness test, the effects of changing parameters such as temperature, accu- racy of the analyst, etc., are simulated, so that the likely performance of a method m a reproducibil- ity study can be assessed. Possible problems can be identtfted and resolved before the actual repro- ductbthty test. Especially in LC, a large number of factors may mfluence method performance. A ruggedness test on the right factors can greatly reduce costs during a reproducibility study and can be apphed in the laboratory where the method was developed, like the repeatability test. A repeatability test combined with a rugged- ness test can be seen as an m-house precision study. This type of testing 1s advisable for every method that 1s destined for general use. However, ruggedness tests are rarely done; even repeatabil- ity testmg IS not yet part of many standard operat- mg procedure, probably because of lack of experi- ence m many routme laboratories. EXPERT SYSTEM FOR PRECISION TESTING IN VALIDATION OF LC METHODS 29 Stand-alone expert systems Previous work was aimed at the development of expert systems that could advise on parts of the problem of preciston testing. For instance, expert systems on repeatability testing [9] and on choice of factors in ruggedness testing [lo] have been developed. The latter system has been developed further to provide a complete system for rugged- ness testing. Both systems are described briefly below. The repeatablkty system. The repeatability sys- tem advtses the user on the performance of repeat- ability tests on the injection and sample prepara- tion of the LC procedure. The results of the re- peated experiments are processed to see if the method is repeatable within the specified limits, i.e., the (relative) standard deviations are calcu- lated and interpreted. If a problem with the method is indicated, the system diagnoses the problem and proposes remedies. The repeatability system contains three mod- ules, which cover the set-up of the test, mterpreta- tion of results and diagnosis cure. On the basis of a description of the method, the system selects a repeatability test for the inlection and sample pre- paration procedures and provides a description of the experiments to be done. The spreadsheet structure of the second module allows input of the experimental results; the relative standard devia- tions are calculated and assessed for acceptability. If they are acceptable, consultation stops. If they are not, possible causes of the error are diagnosed and if the problem can be identified, the system advises on possible soluttons, e.g., check for ade- quate degassing or for a loose grating in the detec- tor. The first and third modules are implemented m a commercial expert-system shell; they contain mainly heuristic rules and frames to represent the objects in the system. The second module is tmple- mented m a spreadsheet package because it is more algorithmic m nature and mostly concerns calculations. The ruggedness system. The ruggedness system advises the user on a complete test, from selection of factors to interpretation of results. A rugged- ness test consists of various experiments that simulate the changes to be expected when a method is transferred from one laboratory to another. / Common Datastructure Fig 1. The ruggedness system Because the number of possible influencmg fac- tors IS large (about 40), the test must be efficient. If most of the interactions between factors can be neglected (which is normally a realistic assump- tion), the most efficient experimental design is a fractional factorial design. But even when factorial designs are used, the number of possible factors is too large, so that the important factors must be selected. The use and interpretation of experimen- tal design require experience winch is not com- monly available in routine laboratories. The ruggedness system is designed to eliminate the problems that prevent ruggedness testing from being part of a normal procedure for method validation. Various modules advise on the differ- ent steps in the ruggedness test (Fig. 1). First, the user enters a descnption of the LC method, which should include only the facts necessary for rugged- ness testing. The system then advises on which factors to test; this is done on the basis of the input description and the requirements specified by the user (e.g., the expected usage of the method). If the user wants to change, add or delete any factors in the output advice, this is stored by the system and the user is warned that modifications have been made. When the set of factors selected by the system has been accepted by the user, the system selects an experimental design for the test; the number of experiments is usually 8-32. After the experiments have been done and the results have been put m, the interpretation module pro- duces either suitability criteria or main effects. Mam effects indicate that there are problems with the method; the system indicates if the problems are serious enough for rejection of the method. Normally, the user is only warned that certain 30 J A VAN LEEUWEN ET AL factors should not vary too much or that certain parameters are not reliable. As in the repeatability system, heuristic and algorithmic processes are used in the ruggedness system. The heuristic processes are implemented in an expert-system shell; the selection of factors and the selection of designs are rule-based. Inter- pretation of the experimental results is imple- mented m C language. The diagnostic module 1s implemented in an expert-system shell wtth frame-based reasoning. INTEGRATION OF THE SYSTEMS Integration of the repeatabihty and ruggedness systems mvolves the development of several new items and the adaptation of some parts of the separate systems. It is vital for all modules of the integrated system to use the same concepts as the basis for reasoning, so that flexible communica- tion is possible between the various modules. Communication between modules in a system can often be done by simple transfer of files, but such communication would be insufficient here. For an integrated expert system, most of the facts pro- duced by one module must be available to all other modules m the system. It is also important that the modules are not consulted in a standard sequence. If a file-transfer system were developed for tlns integration, all possible consultation se- quences would have to be implemented, winch would become unrealistically complex. If all modules in the system must use the same concepts for reasoning, it is better to merge all existing concepts in one common data structure that forms the basts for all systems. In tins study, the common data structure 1s built as a black- board structure; all the modules of the integrated system can read information from and wnte infor- mation to tins structure. The blackboard architec- ture was chosen because it allows integration of modules which use different mferencing or prob- lem-solving techniques. This approach has some consequences for the user interface of the separate systems. If the user mterfaces were not changed, the integrated system would somettmes ask for the same information twice, if the information were important for more than one module. When the information is available in the common data struc- ture, a new module for method description must replace the separate modules of the stand-alone systems. In the architecture proposed here, this is the only essential adaptation of the existing sys- tems. The integrated system needs a supervisor to decide when to consult which module. The super- visor contains knowledge on each module, its in- put variables, its output variables and its status. On tins basis, the supervisor decides which route should be taken to find an efficient precision test. With the supervisor, a level of meta knowledge (knowledge about the expertise in the system) is introduced into the system. Only the supervisor can trigger the modules; the modules cannot tng- ger themselves or each other. Because of the lnerarchical control structure, it is relatively easy to implement additional levels of control, e.g., to integrate other tests like accuracy and sensitivity. The method description module, the common data structure and the supervisor are the new items in the integrated system (Fig. 2). Adaptation of the extsting systems to the common data struc- ture and the method description module requires only minor changes. In future, new modules can easily be added if they use the concepts of the common data structure. Method descnptlon In this module, a full description of the LC method to be tested can be entered. The module contams knowledge on winch LC methods can be assessed by the system. The method description module is normally the first to be consulted, thus the user is confronted with the limitattons of the system at the start. If the system accepts the descnptron of a method, the user can be confident that a valid consultation has been started and a valid conclusion will be reached. The only excep- tion is when a very incomplete method description 1s provided. Normally, the system can work with an mcomplete method description but tf much information is missing, the system may not reach valid conclusions. The point at whtch the system loses its full validity is believed to be when 30% of the method description is missing. EXPERT SYSTEM FOR PRECISION TESTING IN VALIDATION OF LC METHODS 31 The system contains knowledge about the nor- tion that he is not asked to enter. This is done by ma1 concepts of an LC method. The user is not filling areas in the method description that are not asked for any feature of the method that is not in filled by the system. Volunteering mformation is lme with previously given answers. For example, if not always advisable. If the system does not need a diode-array detector is involved, the user will be certain mformation it may have good reasons. asked for wavelength, time constant and attenua- Information volunteered at the wrong place may tion. When a refractive index (RI) detector is confuse the system and mvalidate its conclusions. specified, the user will be asked the RI range and Another feature of the method description the temperature of the detector. If a column ex- module is the guidance given to the user to ensure traction is specified for sample preparation, the that the description contains all the information wash volume and the extraction volume will be relevant for the particular consultation. The sys- needed. When a filtration is specified, the pore tern will not ask for more mformation than it size of the filter will be required. However, the needs. If, durmg a consultation, it becomes clear user can overrule the knowledge in the system by that only a repeatability test is necessary, the user volunteering information. At any time during the will only have to spectfy a few parameters of his method descriptton, the user can enter mforma- method, mainly related to the expected usage of I ” V V V V v v v 1 Common Datastructure I L I Fig 2 The preas~~~ expert system 32 J A VAN LEEUWEN ET AL the method. If a ruggedness test is necessary, many more features are required. Of course, when a ruggedness test follows a repeatability test, the information produced during the first test remains available for the second. Fmishing the method description defines the basis for the following consultation. The system will ask for further information if necessary but the user cannot change hts method description after leaving the method description module. If changes are essential, a new consultation must be started. Common data structure The basis of every expert system is a descrip- tion of the objects about which it can reason. All necessary objects must be described in such a way that misinterpretation is impossible. In this case, all objects are represented in a network of frames. An object is described by giving a frame the name of the object and defining a list of all the relevant properties of the object. Every property has a number of possible values which are also defined m the frame. For every object, a so-called O(bject) A(ttribute) V(alue) tnplet is created. A typical LC object is a column, which is easily described in a frame (see Table 1). The column is defined by properties (attributes) like length, particle size, functionality, internal diameter, etc., each of which has several possible values. For functionality, for example the list including ODS, nitrile, C8, C18, etc. For column length, the range may be between 2 and 100 cm. The different objects in a knowledge domain are related. For instance, the object “LC method” has parts like sample preparation, column and detector. Relations between objects can be of a general nature common to different knowledge domams. The definitions of general relations are normally provided by the expert/system shell. A particularly useful example of a general relation between ObJects is “inheritance”. Inheritance al- lows division of frames into more specific sub- frames; its use implies that all attributes present in a general frame will always be attributed in the subframes, i.e., subframes inherit certain attributes from the more general frame. An example of the use of inheritance m representing relations be- TABLE 1 Example showmg the mformation m a frame for a column Object Attnbute 1 value 1.1 value 1 2 Attnbute 2 value 2 2 value 2 2 value 2 3 Column Tradename: Sphensorb, Hypersil, Nucleostl, Part&, pBondapak, Ltchrosorb Functionahty. C8, Cl& SI60, ODS, Nttnle, Phenyl, PAC Particle size (pm), REAL Column length (cm): REAL Batch number: INTEGER Internal diameter (mm) REAL Column I Tradename. Lichrosorb Functtonahty. ODS Particle size 4 Column length: 20 Batch. 23475435 Column 2 Tradename. pBondapak Functtonahty. Cl8 Particle stze 5 Column length 30 Batch 3498745678 Internal diameter 4 6 Internal dtameter: 4 tween LC objects can be seen in the description of a detector (Table 2). Here, a detector has only two properties: it is always of a certain type (UV, RI or diode array) and it always has a time constant. Real detectors have other properties that place them m a subclass of detectors, e.g., a UV detector will have a wavelength property that distinguishes tt from a RI detector. The UV and RI detectors are thus subframes of the detector frame. Another useful general relation is instantiation, whtch means specifying the exact feature of an TABLE 2 Example of mhentance detector Detector Type. vanable UV, fixed UV, diode array, refracttve mdex Ttme constant (s) REAL (IV detector Wavelength (nm) REAL Attenuation REAL RI detector RI const . REAL Temp (O C): REAL Range htgh, low UV detector 1 Type vanable UV Ttme constant: 0 5 Wavelength. 276 Attenuation 0.1 EXPERT SYSTEM FOR PRECISION TESTING IN VALIDATION OF LC METHODS 33 example. Defining a frame means the introduction of a certain concept mto the common data struc- ture. During a consultation, a specific example of the concept will be defined. An instantiation of a frame has only a subset (normally one) of the possible values for each attribute. Examples of instantiation can be seen in Tables 1 and 2; in- stantiations define so-called u-a relations, e.g., in Table 1, column 1 u-a column. Because mstantia- tions are created during a consultation, they are not part of the common data structure, because they are deleted after each consultation. Figure 3 summarizes all the objects and their general rela- tions in the common data structure. Inheritance enables a network of relationships between objects to be defined. There are, however, other types of relations in the knowledge domain that cannot be described with the inheritance con- cept, e.g., relations between attributes. Such rela- tions can be defined explicitly by a functional SAMPLE TOP FRAME METHOD ~~;AMT;'AF$iTION COLUMN DETECTOR CHROMATOGRAPHIC RESULTS USER REQUIREMENTS I Top. )ESCRIPTION FRAME- I I I FACTOR OPERATOR FACTOR EXP-DESIGN-INFO NUMERICAL FACTOR DISCRETE FACTOR TOP FRAME DIAGNOSE FRAME TOLERANCES STANDARD ERRORS CONCLUSIONS CONCLUSIONS MAIN EFFECTS FACTOR IDENTIFICATION FACTOR GROUPS DIAGNOSE INFORMATION SYSTEM SUITABILITY CRITERIA TOP FRAME REPEAT FRAME PARAMETERS PRECISION TEST VARIABLES SPREADSHEET REPEATABILITY DIAGNOSIS I METHOD VARIABLE REPEATABILITY TEST Ftg 3. Objects tn the common data structure 34 J A VAN LEEUWEN ET AL relationship, in which the names of the attributes are mcluded with the nature of the relationship between them. An example of a functional rela- tionship is the definition that the description of a sample preparation cannot include the attributes shake (mm) and sonicate (mm) simultaneously, because it is unnecessary to use both sonication and shaking in one procedure. The functional relationships are usually of a more complicated and specific nature than the general relations, and are defined in the language of the expert system, e.g., Lips. This makes them less comprehensible and flexible than the general relations but a func- tional relationship is used only once or twice so that generalization is unnecessary. The objects and relations together form the common data structure, which contains much knowledge about LC and method validation. Be- cause it is impossible to change the objects and the relationships during a consultation, the com- mon data structure is the static backbone of the system. The common data structure developed for this expert system provides a basis for a complete description of a procedure for method validation m LC, and could easily be extended for another LC application. Supervisor The supervisor contains the knowledge on when to activate which module. This meta knowledge represents the knowledge of a manager who de- cides which tests are necessary and when they should be done. The decisions of the supervisor are not much related to the activities of the mod- ules. The supervisor only needs information on the mput and output parameters of every module and some mformation from the method description module. At the moment, the supervisor knowledge is represented m rules (Table 3). The rules act on a simple separate frame, the scheduler (see Fig. 2). The modules cannot operate on this scheduler frame. In the scheduler, information is stored about the state of the modules. This structure is capable of handhng the knowledge for the preci- sion-testing system with its eight modules. Because the amount of information stored in the scheduler will become very large when more modules are TABLE 3 Rules m the supervtsor (DEFINE-RULE PRECISION TEST 3 ( doe-string “general rule 100 10-12-87” sponsor select-test-sponsor) (INSTANCE ?user ts user-requuements WITH lab-number 1 OR 2) THEN (INSTANCE PREC-TEST IS TEST WITH GLOBAL PERFORMANCE CHARACTER- ISTIC PRECISION WITH CONTRIBUTORY CHARACTERISTIC REPEATABILITY WITH CONTRIBUTORY CHARACTERISTIC RUGGEDNESS)) (DEFINE-RULE PRECISION TEST 2 ( do-c-stnng “general rule 110 10-12-87” .sponsor select-test-sponsor) (INSTANCE “APP IS APPLICATION WITH lab-number 1 or 2) THEN (INSTANCE PREC-TEST IS TEST WITH GLOBAL PERFORMANCE CHARACTER- ISTIC PRECISION WITH CONTRIBUTORY CHARACTERISTIC REPEATABILITY)) (DEFINE-RULE PRECISION TEST 1 ( dot-stnng “general rule 120 10-12-87” :sponsor select-test-sponsor) (INSTANCE ‘kser ts user-requuements WITH analyst-number 1 WITH mstrument-number 1) THEN (INSTANCE PREC-TEST IS TEST WITH GLOBAL PERFORMANCE CHARACTER- ISTIC PRECISION WITH CONTRIBUTORY CHARACTERISTIC REPEATABILITY)) added, the supervisor is being extended to include more frames and an additional so-called task level between the modules and the planning knowledge. Archrtecture of the system A complete precision testing expert system can be constructed with the building blocks described above. The system is based on the common data structure of frames that represent all the physical and mental objects necessary for a precision test, i.e., descriptions of the sample, method, tests, re- sults and diagnosis. Working on the framework are the modules that each contain knowledge on a EXPERT SYSTEM FOR PRECISION TESTING IN VALIDATION OF LC METHODS 35 certain part of the test procedure. The modules commumcate through the common data structure by placing variable values in it and reading from it. Direct cont.act between modules (e.g., for com- munication of vanables shared by them) is not possible, thus the supervisor can keep track of all activities in the system. The supervisor can see wluch modules can be tnggered at a certain mo- ment because it contains a list of the input and output parameters of all modules. A module can be activated only if all its input parameters are known; it will be deactivated only when all its output parameters have been established. The supervisor acts as a kind of switchboard operator connecting modules to each other accord- mg to the state of the system at a certain moment. The supervisor also contains knowledge on the priority of the modules if a situation occurs where more than one module can be activated at the same moment. Thus the supervisor decides on the best scheme to follow, usmg its meta knowledge. Comparrson with a blackboard At first sight, the architecture described above resembles blackboard architecture [12]. Black- boards are well known artificial-intelligence tech- mques for the integratton of expert systems. Blackboard architecture also uses modules (knowl- edge sources) that can communicate with each other only via a framework of objects, the black- board (see Fig. 4). However, in blackboard archi- tecture, the knowledge sources trigger themselves when the state of the blackboard is such that they can contribute to the solution of the problem. The scheduler then decides which of the knowledge sources can be activated. In an ideal situation, several knowledge sources can be activated at the same time. The blackboard architecture offers possibilities for parallel processing. The scheduler therefore does not contain any planning knowl- edge. In the architecture used here, the supervisor calls up the modules that can add to the solution of the problem. The scheduler frame contains all the conditions for activating the modules, and this allows the implementation of planning knowledge about the activation of the modules. Recently, the literature on blackboards has shown a shift to- r-l blackboard global database Fig 4 Blackboard archtecture wards the implementation of planning knowledge similar to that used here. An example of such a blackboard was described recently [ 151. In principle, the architecture used here does not allow processes to run in parallel unless they are specified as possible simultaneous processes. For the application of precision testing m LC method vahdation, however, this is not (yet) a disad- vantage. PROGRAM FLOW Consultation of the system normally starts with specifying the needs of the user (Table 4A). The appropriate modules are then loaded by the sys- tem. In a complete consultatton, all modules are loaded from initial method description to diagno- sis of the ruggedness test. In the method descrip- tion module, the user can enter all the information that he has on the method (Table 4B). The super- visor then decides, on the basis of the expected usage, which tests are needed (repeatability, ruggedness or both). This information ts transferred to the scheduler frame, which decides on the next step. In general, it will be a repeatability test and the system will activate the relevant module, which uses mforma- tion about the expected usage to select a suitable test. It also activates a spreadsheet in which the 36 J A VAN LEEUWEN ET AL test is implemented. After the user has done the and advise on their acceptability, etc. (see above). test and entered the results, the expert system The same procedure applies when the rugged- activates the next module to diagnose the results ness part is activated. Normally, ruggedness test- TABLE 4 Screen dumps of the mtegrated precision-testing system (A) Intellrgent scheduler Supervisor control NO Descnbe method. YES Perform repeatabdlty test’ NO Select factors to be tested YES Select expenmental design YES Perform diagnosis: YES Run a total consultation NO Start consultation Loaclmg part file (B) Chromatograph Screens > chromatograph questions > options Main questions and answers Flow rate (ml nun-‘) Number of solvents PH Buffer cone (M) Additives Injection volume (~1) Temperature mode 10 CONTROLLED (C) Selected factors > sample > chrom > detector > column > data > options < Sample prep. Chromatograph Detector Column factors. factors factors factors weight shake time sonicate time heat temp pore size 1 pore size 2 wash vol extract vol extraction centnfuge ddutlon PH * temperature :- buffer . - solvent . * additive _ flow rate . - RI-range _ filter ._ wavelength + UV-time-con - RI-time-con - Manufacturer. * Batch - Data handlmg factors _ _ _ 1 factor * Related answers Mmlmum solvent (W) Solvent 1 (W) Solvent 2 (%) Solvent 3 (W) Solvent 4 (%) 25 15 25 Mmlmum additive (W) . Addltlve 1 :08( ) Adchtlve 2 OS( ) Additive 3 O%( ) Additive 4 O%( ) Addltlve 5 O%( ) Temperature ( o C) 40 No modlflcatlons made EXPERT SYSTEM FOR PRECISION TESTING IN VALIDATION OF LC METHODS 37 TABLE 4 (continued) (0) Selected experimental design Expenmental design SATURATED-FRACTIONAL-FACTORIAL-DESIGN Number of factors 7 Number of levels 3 Number of dummy factors 0 Number of expenments. 15 Show factors Wavelength warmng (E) System surtabrlrfy crrtena The results of any analysis should always fall wrthm the ranges of the values gven below Minimum found at exp.: 6 Comp 1 2 RT 217 333 277.0 Area 58 449 572.325 He&t 646 832 7772 02 C area 297 621 296.896 C hgt. 296 535 296.271 P count 2 5399 5.7842 Result 2425 15 2614.22 Maximum found at exp.: 14 1 2 267 333 439 0 255.746 102 058 2841.85 2096.0 6.979 7.0895 6 982 6 8935 3.1108 6.0617 3637.18 287117 (F) Marn effects tis screen shows the input gven by the user It also gves a short descnptlon of the results. When a change 1s made to the mput values, the results may change A better descnptlon of the results can be obtamed by selectmg the appropnate screen from the ‘screens’ menu Usage. ONCE ONLY Length of run’ > 10 Q 25 Number of users > 3 Number of Instruments. 23 Number of laboratones 1 OR 2 Preparation repeatablhty test: REPEAT 1 InJection repeatablhty test REPEAT 5 Wnte ss ing is only done after a successful repeatability activated. If the results are satisfactory, the mod- test, but the user may elect to forego the repeata- ule will report this and provide system suitability bility test and proceed with the ruggedness test. criteria (Table 4E). If the results are not satisfac- The ruggedness part consists of four modules (see tory, the user is informed of the main effects that Fig. l), which are usually consulted in sequence. are outside the tolerance limits and of the factors The factor choice module advises on which factors causing the problem (Table 4F). A solution to the are likely to mfluence method performance (Table problems that does not affect the method itself, is 4C). The select design module then advises on a to respecify the levels for testing the factors. When suitable experimental design to test these factors these levels are specified at narrower intervals, the with a mimmum of experiments (Table 4D) and a method is more likely to pass the ruggedness test C program is activated in which the user enters his but, of course, operation of the method in the experimental results. The diagnosis module is then laboratory will have to be kept under more rigid 38 J A VAN LEEUWEN ET AL TABLE 5 Results of the test case Method description Sample name formulation of components Sample preparation take sample No of tablets add solvent (ml) add mtemal standard dissolve sample somcate mmutes (mm) ddutlon 1 ddutlon 2 extraction filter pore size (am) Chromatograph solvent no solvent 1 (X) solvent 2 (%) PH buffer cone (M) Injection volume (al) temperature ( o C) flow rate (ml nun-‘) Column tradename functionality particle sze (am) column length (cm) batch number internal diameter (mm) Ruggedness test Factors chosen by the system aspum, sahcyhc actd tablet 2 Detector type attenuation ttme constant (s) wavelength (nm) vanable UV 0.1 01 295 take no of doses Results 1 250 no internal standard somcate 15 no dtlut:on no filtratlon 10 2 15 25 25 100 40 15 Sphensorb Cl8 70 25 1 40 nummum resolution 4.0 mm retentton time (s) 240 overall run time (s) 600 worst peak symmetry 14 Reqmrements usage No. of hnes purpose regulatory standard method mterlaboratory number of analysts average run length number of instruments > 10 d 25 4 stability mdlcatlon USA USP >3 23 > 10 < 25 >3 Factor Nommal level Lower level Upper level Somcation time 15 Pore size 10 Data handling _ PH 2.5 Solvent 25 Manufacturer Sphensorb Wavelength 295 12 18 5 20 15 35 20 30 other other 290 300 EXPERT SYSTEM FOR PRECISION TESTING IN VALIDATION OF LC METHODS 39 TABLE 5 (contmued) Experrmental desrgn Reflected saturated fractional factonal desrgn 11 factors 4 dummy factors 3 levels Inierpretatlon of expertmental results Mam effect Factor 24.0% somcate ttme 23.7% pore stze 24.0% solvent 27.2% manufacturer 22 5% wavelength Repeatability test Test advrsed by the system 10 ttmes repeated mJectton of sample 5 ttmes repeated sample preparation Interpretatron Relatrve standard devtatrons InJection of sample < 1% Sample preparation < 1% Dtagnosrs No problems control. In this case, the factor choice module is activated again, the factor levels are adapted to fit the test, and the whole procedure is repeated. Control of this operation is again placed in the scheduler frame that keeps track of the number of times a module is activated and also checks if the new activation has yielded a result. The architecture around the scheduler frame also makes it possible to load only one module that can be consulted as a stand-alone expert system. For some modules, tins can be practical. The prototype was developed in the Goldworks expert-system shell, Version 1.1 [16]. An IBM PC/AT with 8 Mbyte extended memory was used. The spreadsheet packages Lotus 123 and MS-C were used for implementation of the tests. RESULTS OF A TEST CASE To illustrate the capacities of the system, the test case used was a full in-house precision test of an LC method for the determination of aspirin and salicylic acid. Both repeatability and rugged- ness tests were done and the results were interpre- ted by the system. Real experimental data were used so that a comparison of the system perfor- mance in a real-life situation was possible. Table 5 shows a complete description of the results. The conclusion of the system for the repeatabil- ity test was that the method was satisfactory within the specified limits. As the method had previously undergone similar tests [9], this was to be ex- pected. The set-up of the ruggedness test corresponded with the expert’s ideas and suggestions. The only difficulties appeared in the interpretation of these results. The system reported several problems that were not encountered in the real ruggedness test. For instance, the system suggested that the method was not rugged in the measurement of the con- centrations of the analyte. As this is the crucial parameter, such a conclusion would be serious, but in reality these problems were not encoun- tered. The difficulty may be due to too rigid interpretation by the system or to problems over- looked in real life. To test tins, a full reproducibil- ity study is needed. Concluslon The expert system described here is a prototype that must still undergo a full evaluation study, which will be reported elsewhere. The philosophy of using a common data structure as a blackboard for various modules will be investigated further. The possibility of adding new modules for testing accuracy and selectivity will be crucial for success. The addition of modules on extensions to the ruggedness test and other modules on method development will also be investigated. Part of this research was supported by the European Community in Esprit project P1570 Ex- pert Systems in Chemical Analysis. REFERENCES 1 M A Trschler and EA. Fox, Comput Chem , 4 (1987) 235 2 C G Et&e, A P Wade, P T Palmer and K J Hart, Anal Chem , 59 (1987) 1363A. 3 S Moldoveanu and CA Rapson, Anal Chem., 59 (1987) 1207 J A VAN LEEUWEN ET AL 4 K. Janssens, W Donnne and P. van Espen, Chemom. Intell. Lab. Syst., 4 (1988) 147. 5 G.J Kleywegt, H J Lumge and H.A van ‘t Klooster, Chemom. Intell. Lab. Syst., 2 (1987) 291. 6 D. Goulder, T. Blaffert, A Blokland, L Buydens, A Chhabra, A. Cleland, N Dunand, H. Hmdnks, G. Kate- man, H. van Leeuwen, D. Massart, M. Mulholland, G Musch, P. Naish, A. Peeters, G. Postma, P. Schcenmakers, M. Desmet, B. Vandegmste and J. Vmk, Chromatographta, 26 (1988) 237. 7 A Peeters, L. Buydens, D.L. Massart and P.J Schoen- makers, Chromatograplua, 26 (1988) 101 8 P.J. Schoenmakers, N Dunand, A. Cleland, G. Musch and T Blaffert, Chromatographta, 26 (1988) 37 9 M Mulholland, J A van Leeuwen and B.G.M Vande- gmste, Anal. Chum. Acta, 223 (1989) 183. 10 J A. van Leeuwen, B.G.M Vandegmste, G. Kateman, M. Mulholland and A Cleland, Anal Chtm. Acta, 228 (1990) 145. 11 M Mulholland, N Durand, A. Cleland, J.A. van Leeuwen and B G.M. Vandeginste, J. Chromatogr., 485 (1989) 283 12 N.P Nu, the Arttf. Intel. Magazme, (1986) 38 13 W.J. Youden and E H Sterner, Statisttcal Manual of the Assoctation of Offtctal Analyttcal Chemists, AOAC, Washmgton, DC, 1975 14 M. Mulholland, P.J. Niush and D.R Stout, Chemom In- tell. Lab. Syst., 5 (1989) 263. 15 B.R Miutre, T. Laasn, F Mondot, F Charptllet and J P Haton, paper presented at the 9th International Workshop on Expert Systems and Thetr Apphcattons, Avtgnon, May, 1989. 16 J.A. van Leeuwen, B G M Vandegmste, G J Postma and G. Kateman, Chemom Intell. Lab Syst., 6 (1989) 239. 
work_wtsndgcf35a4bnaavb7u43ux5q	----	U:\WWW\Files\SanzogniSurruwerraWebb90.pdf.prn.pdf Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 1 of 11 Improving the Efficiency of Rule -Based Expert Systems by Rule Activation LOUIS SANZOGNI and FRANCIS SURAWEERA School of Computing and Information Technology, Griffith University, Nathan, Brisbane, Queensland, Australia 4111 GEOFFREY I. WEBB Department of Computer Science, Deakin University, Victoria, Australia 3217, Abstract In this paper we test a hypothesis that has shown promise in enhancing the efficiency (run-time) of rule -based systems. The results of our experiments suggest that the use of rule activation plays an active part in improving the performance of rule bases containing conflict sets. Keywords Activation levels, expert systems, hypothesis testing, production systems, rule activation, rule - based systems, run-time efficiency. 1 Introduction Since the mid 60’s expert systems have emerged as the most significant practical product of artificial intelligence (Al) research. An expert system is based on an extensive body of knowledge about a specific problem area. One of the most successful ways of representing this knowledge is as a production system. In the AI literature, production systems (or simply productions) are also known as rule - based systems or inference systems. The expert system uses the productions to infer new knowledge and conclusions from given data or premises. Several studies in Cognitive Psychology have found evidence that mental activity causes related cognitive structures to become more highly activated (see for example, Anderson 1975, 1983). The more highly activated, the more readily it can be accessed and the more likely it is to be used. It is reasonable to assume that a piece of information will become active to the degree that it is related to the current sources of activation. Quite simply, the current focus of cognitive activity directs future foci. We presume that a similar mechanism might be of some use in an expert system in finding the best rule to select from a conflict set. First we note that there are a number of methods for improving the inference process. Some of these methods range from simple to complex, in form and theory (see Forsyth, 1984). In a recent paper Lenat (1984) reported the following: ‘The key to intelligent problem solving lies in reducing the random search for solutions . . .’. In line with Lenat’s view, the research reported in this paper analyses a hypothesis which has shown promise in enhancing the efficiency of rule -based systems. It is our intention to show that we can improve the performance of an expert system by setting up the rule base according to associative memory concepts. Only the run time efficiency (how long it takes to reach a conclusion), is considered here. Evidently, the run time efficiency is a function of the number of rules that has been searched. Pre-publication draft of a paper published in the Journal of Experimental Theory in Artificial Intelligence, Taylor and Francis, Volume 2 (1990) 369-380 Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 2 of 11 The hypothesis is that: . . . the performance of the system will be improved by associating an activation level with each rule in the rule base. We now give the organisation of the rest of the paper. In Section 2 we define some of the terminology that will be used. In Section 3 the experiment to test the hypothesis will be described. The results from the experiment will be presented in Section 4. Lastly, Section 5 gives a discussion and conclusions. 2 Preliminaries In this section we describe some of the terminology associated with expert systems. Since many of these terms are found in standard text books (for example: Waterman 1986, Hayes-Roth et al 1983, Harmon & King 1985, Jackson 1986, Silverman 1987) only those terms which might have an ambiguous meaning are defined here. Post (1943) was the first to introduce production systems. Later, the same concept was used by Newell and Simon (1972). The knowledge embodied in a production system may be represented as a set of rules of the form: Rule: <name> IF condition 1, condition 2, . . . condition n THEN action 1, action 2, . . . action m where n ≥ 0 and m ≥ 1. The condition-action pairs in the above rule are also referred to as the antecedent-consequent pairs or situation-action pairs. A fact is represented by a rule with no antecedent; that is, when n = 0, and m = 1. In a production system the control structure allows only one rule to fire in a given cycle. However, more than one rule may be triggered (or activated) in which case a conflict resolution strategy must be applied to determine which of the triggered rules should fire. The set of activated rules is known as the conflict set. The activation measures the likelihood that a particular piece of knowledge (rule) will be useful at a particular time. Although a great deal of information may be active at any given point of time activation does not necessarily result in behaviour (see Anderson 1983). The activation level will determine the speed of pattern matching and hence the performance. In our system each rule is assigned an integer called its activation level. At the beginning, the activation level of each rule is initialised to be zero. During an implementation of the algorithm a routine increments the activation level of those rules that are relevant (conditions satisfied) for attaining a success. A zero activation level indicates that either the rule has no usefulness or it is inactive. As the activation level increases it determines the degree of desirability of that particular rule during the inference process. Further elaboration of these ideas will be presented in Section 3. 3 The experiment In order to have meaningful results it was imperative to test our hypothesis on a real world problem. In this connection, we examined a few expert systems to select one that matched our requirements. Specifically, we were looking for a tractable (in terms of the number of rules) expert system with a sufficient number of conflict sets. In the end, we selected the Fire Expert System Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 3 of 11 (FES) rule base which was supplied to us by Richard Davis of the CSIRO Institute of Biological Resources, Canberra. The FES rule base has 110 rules and 19 conflict sets. A total of 36 facts are needed by the knowledge base for inferencing though not all are required simultaneously. Our proposed knowledge base consists of three components. These are, (i) rules, (ii) facts, and (iii) associations. The rules and facts are standard repertoire of any knowledge base but associations are not. The following examples represent (a) a fact, and (b) a rule, in general. (a) fact2: fuel type is (perennial annual sorghum) (b) rule2: IF season is (early storms) AND it has been up to 3 days since last rain THEN available biomass is negligible It is worth noting that in the proposed know ledge base the rules are constructed so that the first proposition is the consequent and the rest, if more than one, are conjunctive antecedents. The associations used in the proposed knowledge base, may be broadly classified into two types: (a) factual, and (b) predicate/object. Of the two types of associations, the factual associations are very easy to set up. This is done by looking at all possible facts available to the system, and associating rules and facts as shown below. Fact 1 is needed by rule r091 Fact 3 is needed by rule r103 Fact 5 is needed by rules r078 r075 r072 r080 r068. where rxxx represents a rule. The inclusion of fact/rule association in the knowledge base was deliberate. Such an organization of the knowledge base will make the information regarding rules and the corresponding facts readily available in a problem solving situation. Associating predicates and/or objects with rules can be quite difficult. More research is currently being done in this area. However, a logical choice is to link predicates, objects, and facts with the rules most likely to attain a success in the given problem solving context. It is worth noting that each rule in the knowledge base has a consequent, and the consequent itself can be made up of a number of elements; the first element of the consequent is referred to as the predicate, and the rest as the objects as shown below. Consequent Pred, obj1, obj2, ... obj m The predicates and objects making up the consequent are scanned one at a time, and the rules most likely to have a bearing upon a solution are then associated with the respective predicates and objects. Table 1 shows an example of predicate/object and rules associations. In this study, these associations are done manually. However, it is reckoned that this activity could easily be automated in the future. In such a knowledge base, a suitably modified inference engine will fire the rule most desired by the associations. The routine shown in table 2 has overall control of the inference process. Its main functions are as follows. (a) It applies the activations to the rules according to the ‘predicate’ and ‘objects’ in the resolvent. (b) It extracts the predicate from the resolvent and sets up the rules associated with that predicate in descending order of activation. Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 4 of 11 (c) It unifies the resolvent with the consequent of the next rule in the sorted sequence. (d) If successful it applies the mgu (most general unifer) to the antecedents of the remainders of resolvents, and the goal. It then restarts the inference process with the next resolvent. If not successful, it moves on to the next rule. This process is continued until either a resolvent is proved or all the rules are exhausted. In either case, the routine delivers a result eventually. It is hypothesized that this mechanism improves the speed with which some expert systems reach a conclusion; particularly, in those expert systems with rule bases containing conflict sets. Associate the following rules with the evaluation of predicate (is-available-bio mass) r064 r069 r079 r070 r068 r080 r073 r071 r074 r072 r076 r077 r075 r078 r106 r065 r111 r055 r052 r056 r054 r053 r048 r057 r050 r047 r058 r046 r049 r051 r200 r045 r043 r042 r040 r039 r038 r037 r110 Associate rule r055 with the evaluation of object negligible Table 1. An example of predicate, object and facts associations FUNCTION Backward_Chain (RS_stack, G) Let RS_stack and G denote the Resolvent stack and the goal respectively. IF RS_stack = empty THEN output goal G ELSE Increment the activation level of the rules associated with the contents of the StackTop(RSstack); R_stack ← set of rules (in descending activation level order) whose consequent’s predicates unify with the RS_stack’s top member consequent’s predicates; FINISHED← FALSE; WHILE NOT(R_stack = empty) OR NOT FINISHED DO R' ← Pop R_ stack; RS ← Pop RS_stack; {current goal} IF R' unifies with RS THEN rename variables in R' A' ← antecedents of R'; MGU ← mgu of RS and the consequent of R'; Push A' to the RS_stack; Apply MGU to the RS_stack and G; RESULT← Backward_Chain(RS ← stack, G); IF RESULT THEN FINISHED← TRUE END WHILE ENDFUNCTION Table 2. The interface routine Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 5 of 11 A system using activation levels has to perform three extra tasks. These are: (a) set up the knowledge base and activate rules pertaining to current available facts (once per run); (b) maintain activation levels; and (c) sort activation levels in descending order. It is worth while to note that task (a) is performed on a one-off basis at the beginning of an execution. The other two tasks are performed as on-going processes throughout an execution. In what follows, tasks (a), (b), and (c) are explained further. 3.1 Task (a) Consider an expert system whose rule base has a conflict set. Such an expert system will have an initial run-time overhead which is greater than that of a conventional system. This overhead is due to the initial setting up of the rule base (Sanzogni 1988). For the knowledge base considered in this paper, the setting up times for non activated versus activated rules were of the order of 1:3.37. The routine in table 3 sets up the knowledge base in the correct format and is invoked by every rule in the rule base at setting up time. It sets up the rule name and associates with it: (a) an initial activation level of zero, (b) the rule itself. FUNCTION Set_Rule (Name, Consequent, Antecedents) Let PR_stack denote the Predicate Rule stack. Let RN_AL be the activation level associated with Rule_Name (RN). Set up Rule_Name RN← Name; RN_AL← 0; {initialize} Associate with RN the Consequent C and Antecedents A; {i.e. RN_C_A ← C&A} Extract the Predicate P from C; IF PR_stack does not exist THEN set up PR_stack Associate RN to P and add to PR_stack ENDFUNCTION Table 3 Rule set-up routine The routine also associates the rule’s name with the predicate name of that rule’s consequent. The routine in table 4 increments the activation level of certain rules according to associations that exist between facts and rules that use those facts. It is invoked prior to the beginning of the inference process. It is also worth noting that this routine is not part of the main inference cycle. 3.2 Task (b) In section 2 we gave some motivation behind the concept of activation level. In this part we illustrate the rule activation strategy using the following simple example drawn from the FES expert system. Consider the resolvent: ‘is-available -biomass litter-mass’. The predicate and the object of this resolvent are: ‘is available -biomass’, and ‘litter-mass’ respectively. Suppose further that their lists of associations are rules r102 and r102 respectively. Consequently, the activation level of rule r102 is increased by 2. Table 5 presents the algorithm developed for increasing the activation level. Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 6 of 11 FUNCTION Fact_Activation (Facts_&_Associations) Let F&A_stack and RN_stack denote the Facts_&_Associations stack and Rule Names stack respectively. WHILE NOT(F&A_stack = empty) DO F&A← Pop F&A_stack; F ← extract fact F from F&A; IF F had been used in this run THEN set up a stack of rule names RN_stack from F&A associated with F WHILE NOT (RN_stack empty) DO RN ← Pop RN_stack; Increment activation level of RN; END WHILE END WHILE ENDFUNCTION Table 4. Rule for fact activation routine FUNCTION Activation (Current_Goal, Activation) Let A_stack and RN_stack denote the Activation stack and the rule names stack respectively. Let PO denotes Predicates and Objects. Set up the Predicates & Objects found in Current_Goa l in a stack CG_stack. WHILE NOT(CG_stack = empty) DO PO← Pop CG_stack; Extract from A_stack the rule names associated with PO and stack them in RN_stack; WHILE NOT(RN_stack = empty) DO RN← pop RN_stack; Increment activation level of RN; END WHILE END WHILE END FUNCTION Table 5. Activation increment routine Example resolvent: (is -available -biomass litter-mass) associations: (is -available -biomass r102) (litter-mass r102) old activation level: (r102 5) new activation level: (r102 7) where (r102 5) denote that the rule r102 has an activation level equal to 5. 3.3 Task (c) Recall that our rule base has been especially organized for the purpose of rule activation. As explained earlier, the knowledge base contains a list of predicates associated with all the rules that contain those predicates in their consequents. In the example given below suppose that the Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 7 of 11 predicate ‘is-available -biomass’ is associated with a set of rules S (say). Let us also assume that the current resolvent is ‘is-available -biomass litter-mass’. Then ‘is-available -biomass’ becomes the predicate extracted from the resolvent. All the rules which have this predicate in their consequent (i.e. the rules in the set S) are then sorted and stacked up in descending order of activation level. The rule with the highest activation level is looked at first, since it is the most likely rule that will solve the resolvent. 4 Results The measurements reported in this study were carried out using the Fire Expert System knowledge base on an Apple Macintosh II running XLISP 1.4. An expert system which employed rule activation was compared with a standard system. The expert system that did not employ the rule activation was designated as the standard one. It is worth noting that no conflict resolution strategy was used by the standard system during test runs. To assess our hypothesis tests were designed to collect three different types of data, namely: • run time • rules visited • rules fired The data for each of the runs was chosen to give as much scatter as possible for the rule -path choice. The data collected from each of the bench marked tests are given in tables 6, 7 and 8. Figure 1 is a column graph depicting the difference between the number of rules visited by a system not using rule activation and a system using rule activation. All the executions demonstrated a marked decrease in the number of rules visited by the system using rule activation. Figure 2 is a column graph which shows the difference between the number of rules fired by a system not using rule activation and a. system using rule activation. As shown in figure 2, rule activation has decreased the number of rules fired in 7 out of the 8 runs, and it came equal in run 3. On the basis of the above evidence, we can now say that rule activation plays a major part in guiding the system to reach a conclusion and hence it should show significant savings in execution times. At this point we may deduce that rule activation reduces the number of rules visited by an expert system in arriving at a conclusion. Run Run Time (sec.) (non activated) Run Time (sec.) (activated) 1 449 136 2 1159 113 3 181 140 4 866 143 5 238 112 6 487 183 7 1648 112 8 1393 113 Table 6. Run time Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 8 of 11 Run Rules Visited (non activated) Rules Visited (activated) 1 545 30 2 1862 20 3 242 19 4 1332 27 5 366 18 6 718 20 7 2615 34 8 2224 25 Table 7. Rules visited Run Rules fired (non activated) Rules fired (activated) 1 14 12 2 22 8 3 8 8 4 20 11 5 7 6 6 11 7 7 26 11 8 28 11 Table 8. Rules fired Figure 1. Non activated versus activated rules (Rules Visited) Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 9 of 11 Figure 2. Non activated versus activated rules (Rules Fired) Figure 3. Non activated versus activated rules (Run Time ) Figure 3 shows a column graph which compares the run time (in seconds) between the runs done using rules without activation levels and rules with activation levels. The graph in figure 3 shows a marked improvement in majority of the runs and thus supporting the case for rule activation despite rather heavy overheads. These overheads arise from (a) maintenance of rule activation, and (b) sorting of rules. Intuitively it may seem that these two activities should slow down the overall time to reach a solution. However, computations using rules which employ activation levels took considerably less time to execute than those using rules with no activation levels. Run 3 had the least improvement, 29%, whilst run 7, at 1372%, had the greatest improvement. 5 Discussion and conclusions In this paper we presented a method towards improving the efficiency (performance) of an inference system. The method employed was motivated by broad considerations of cognitive processes. The scheme involves increasing the activation level of rules and consequently reducing the costly searches for solutions. In general, learning can be broadly classified as either (i) enhancing the skills of the learner, or (ii) in acquiring knowledge for the learner (see Firebaugh 1988). The first of these, skill enhancing, is the machine learning analog to performance improvement. In the present study it has been shown that changing activation levels leads to performance improvements. Thus, it is reasonable to state that changing activation levels is a form of learning. Of course, a learning system should be able to do the same task(s) more efficiently and effectively the next time by drawing upon past experience. However, it should be apparent by now that our rule-based system is not capable of self improvement through generalization of past experience. This kind of non-learning behaviour Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 10 of 11 is common with deductive systems which employ high-level interpreters. It is known that these high-level interpreters can cause difficulties for learning (Booker et al. 1989). They commented: ‘High-level interpreters, by design, impose a complex relation between primitives of the language and the sentences (rules) that specify actions. Typically this complex relation makes it difficult to find simple combinations of primitives that provide pla usible generalizations of experience.’ The success of this experimental study is not due to some sophisticated method of inference (see, for example, Holland’s bucket brigade algorithm, rete algorithm, etc.) but it stems from the special architecture of our rule base. More directly, it is related to the rule activation strategy that was employed. To summarise, on the basis of the evidence we have seen, it can be said that a considerable saving in run time was obtained by the expert system using rule activation as compared with the system not using rule activation. This saving in run-time is attributed to the activation levels of the rules. The activation levels force the attention of the inference engine towards those rules which are better suited for the solution of the problem at hand. At this point it is logical to compare our technique (which basically employs activation-level incrementing algorithms) to an existing technique, viz; Holland’s bucket brigade algorithm. It appears that there are some similarities between activation level incrementing algorithms (ALIA) and Holland’s bucket brigade algorithm (HBBA) but on closer examination it turns out that they have their own distinctive characteristics. Both algorithms are motivated by broad considerations arising from cognitive processes. The HBBA is designed to solve the credit assignment problem for classifier systems (see Sutton 1987, Booker et al. 1989). Both methods employ the following: • induction mechanisms with problem solving; and • use of conditio n/action rules as a basic unit of representation. A classifier system will hypothesize new rules by recombining building blocks (rules) to resolve a problem solving impasse. On the other hand in ALIA this is done by creating a subgoal. Once this subgoal is achieved it generates another subgoal and so on. Thus, the two methods implicitly use different learning mechanisms. Another point of departure between the two methods is the way in which rules compete to be selected. In HBBA rules compete by bidding on the basis of the strengths and the highest bidding rules (classifiers) go on to the next time step. On the other hand in ALIA rules are selected (conditions satisfied) on the basis of activation levels and the rules are looked at starting from the highest activation level. If it does not unify with the resolvent then it looks at the one corresponding to the next highest activation level and so on. In HBBA control issues tend to come up because several rules are allowed to fire at the same time. Also probability ideas are used in the bidding process where as in ALIA we do not use probability ideas at all. It should be apparent from the above that the two methodologies are quite different despite their initial similarities. A comparison with the rete algorithm (Forgy 1982) is not attempted at this point in time. A full comparison is being planned and we hope to present its findings at a later date. According to our computational experience we conjecture that only certain types of expert systems lend themselves to rule activation as first proposed, and that such systems can benefit greatly from rule activation. Our study provides an experimental demonstration that the use of rule activation for enhancing the performance of rule -based expert systems shows promise. The trials presented in this paper are by no means conclusive, as they were performed only on one expert system. Nevertheless, they are encouraging. Acknowledgements We would like to express our sincere thanks to the Editor and the anonymous referee whose comments have been very helpful in strengthening this paper. Sanzogni, L., Suraweera, F. and Webb, G.I. (1990) Improving the Efficiency of Rule-Based Expert Systems by Rule Activation Page 11 of 11 References [1] Anderson, B. F. (1975) Cognitive Psychology: The Study of Knowing, Learning and Thinking (New York: Academic Press). [2] Anderson. J. R. (1983) Cognitive Psychology and its Implications (New York: W. H. Freeman and Company). [3] Booker, L. B., Goldberg, D. E. and Holland. J. H. (1989) Classifier systems and genetic algorithms. Artificial Intelligence. 40: 235-282. [4] Firebaugh. M. W. (1988) Artificial Intelligence: A Knowledge-base Approach (Boston: Boyd & Fraser). [5] Forgy, C. L. (1982) Rete: A fast algorithm for the many pattern/many object pattern match problem. Artificial Intelligence. 19: 17-37. [6] Forsyth, R. (1984) Expert System - Principles and Case Studies (New York: Chapman and Hall). [7] Harmon. P. and King, D. (1985) Expert Systems - Artificial Intelligence in Business (New York: John Wiley & Sons, Inc.). [8] Hayes-Roth, F., Waterman, D. A. and Lenat, D. B. (1983) Building Expert Systems (Reading: Addison-Wesley). [9] Jackson. P. (1986) Introduction to Expert Systems (Reading: Addison-Wesley). [10] Lenat, D. B. (1984) Computer software for intelligent systems , Scientific American. September Issue, 152-160. [11] Newell. A. and Simon, H. A. (1972) Human Problem Solving (Englewood Cliffs: Prentice- Hall). [12] Post, F. (1943) Formal deductions of the general combinatorial decision problem. American Journal of Mathematics. 65: 197-215. [13] Sanzogni, L. (1988) Improving the Efficiency of Expert Systems by Rule Activation. B. lnf. Honours Thesis. Griffith University, Nathan, Brisbane, Australia. [14] Silverman. B. G. (1987) Expert Systems for Bu siness (Reading: Addison-Wesley). [15] Sutton. R. S. (1987) Learning to predict by the methods of temporal difference. Technical Report TR87-509.l, GTE, Waltham, MA. [16] Waterman. D. A. (1986) A Guide to Expert Systems. (Reading: Addison-Wesley). 
work_wukoiilterb7zhuqdvbnetghe4	----	Coventry University Coventry University Repository for the Virtual Environment (CURVE) Author names: Alawamleh, M. and Popplewell, K. Title: Interpretive structural modelling of risk sources in a virtual organisation. Article & version: Post-print version Original citation & hyperlink: Alawamleh, M. and Popplewell, K. (2011) Interpretive structural modelling of risk sources in a virtual organisation. International Journal of Production Research, volume 49 (20): 6041-6063. http://dx.doi.org/10.1080/00207543.2010.519735 Publisher statement: This is an electronic version of an article published in the International Journal of Production Research, 49 (20), 6041-6063. The International Journal of Production Research is available online at: http://www.tandfonline.com/doi/abs/10.1080/00207543.2010.519735 Copyright © and Moral Rights are retained by the author(s) and/ or other copyright owners. A copy can be downloaded for personal non-commercial research or study, without prior permission or charge. This item cannot be reproduced or quoted extensively from without first obtaining permission in writing from the copyright holder(s). The content must not be changed in any way or sold commercially in any format or medium without the formal permission of the copyright holders. This document is the author’s final manuscript version of the journal article, incorporating any revisions agreed during the peer-review process. Some differences between the published version and this version may remain and you are advised to consult the published version if you wish to cite from it. Available in the CURVE Research Collection: May 2012 http://curve.coventry.ac.uk/open http://dx.doi.org/10.1080/00207543.2010.519735 http://www.tandfonline.com/doi/abs/10.1080/00207543.2010.519735 http://curve.coventry.ac.uk/open Interpretive structural modeling of risk sources in Virtual Organisation Mohammad Alawamleh, Keith Popplewell Engineering Manufacture and Management, Coventry University, Coventry, UK Coventry University, Priory Street, Coventry CV1 5FB, United Kingdom {awamlehm, K.Popplewell} @coventry.ac.uk Speedier network decision making together with shorter time to bring items to market together with lower network operating costs all result from enhanced knowledge sharing. In addition re-use of enterprise and network knowledge resulting from improved capture means that any risk of repeating earlier project work is limited, repetition of past mistakes is reduced. Decisions are made with greater awareness of any risks involved and therefore there is likely to be a reduction in costs arising from faulty decisions and failed collaborations. While there are many advantages attaching to the use of virtual organizations (VOs) there are also challenges, including risks that have become apparent through undertaking a review of the literature. In total 13 sources of risk were found stemming from the network related risks in a VO, where the emphasis of the study was placed. This paper contains a thorough study that will identify these threats as well as gaining a sound understanding of them by examining them one by one as they have been identified by the literature and previous studies. Subsequently, their relative importance will be analysed through the use of Structural Modeling (ISM) using information gathered in a questionnaire. Keywords: virtual organisation; VO; risk; SME; interpretive structural modelling; ISM 1 Introduction Collaboration between enterprises is vital if businesses are to meet their objectives. Doing business world-wide has become crucial to the survival of some enterprises while for others the vital element is focusing locally. It is necessary for all enterprises, whatever their size, to come to corporate agreements with other enterprises and this is particularly important for small and medium sized enterprises (SMEs) who in order to increase their own added value need to operate together with others within the market. In today’s market, whether or not an enterprise is successful will often be largely dependant upon whether it is able to interoperate smoothly with others. The environment within which SME’s have to function in the 21st century is one that is increasingly competitive and dynamic and therefore, simply to cope in such a situation, SMEs have to seek methods to employ: one of these is to group together within a VO. It is, however, not easy to become part of a VO and there are risks to be dealt with throughout the whole process from the initial formation of such a group through to the point where it dissolves. In order that the challenges can be met successfully, it is important that enterprises should be helped to both recognise the risks and then surmount them. Collaboration is necessary in order for enterprises to compete and it is also necessary for them to operate with as much speed and flexibility as possible so that ideas and proposals can become initiatives with the capacity to generate new revenue. As is the case in the supply chain, the risks associated with VOs have multiple sources and Juttner et al. (2002) have suggested that the sources of risk as they affect supply chains should be categorised into three areas, those being risks external to the supply chain, risks internal to the supply chain and those that are network related. Such risks as natural risks, political and social risks and risks connected to the industry market would be categorised as external risks while internal risk sources are likely to be associated with labour problems such as strikes, or production problems such as machine failure, as well as problems with IT systems and network related risks stemming from the relationships between the supply chain, also been by Blackhurst et al. (2004) as a risk that is different in kind but that has a direct relationship with collaboration. While the risk sources that are internal or external are essentially similar for supply chain and virtual organisation, the network related risks have a different source as a result of the different relations between the SC and the VO partners. Those network risks that relate to any collaboration do not depend solely on the enterprise goals and objectives. A dyadic type of assessment is necessary as a result of the sharing of responsibilities and the changing nature of the relationships involved if these situations are to be actively managed in relation to network related risks, since the identification of risk becomes more complex as a result of the interdependency of enterprises (Hallikas et al., 2004). The aim of this study is to explore various risk sources in the VO, to establish relationships among the sources through ISM methodology and to classify these sources depending upon their driving and dependence power using indirect relationship MICMAC analysis ISM is a well-established methodology for identifying and summarising the relationships among specific elements which define a problem or an issue (Sage, 1977; Warnfiled, 2005). The proposed model provides a useful tool for SMEs to focus on those risk sources that are most important for effective risk management in VOs. Understanding the risk sources and their relationships will help organisations address them or at least understand so they can plan for them if and when they occur. Based on literature, risk sources that affect VO directly and indirectly, are identified. A survey was prepared and administered to test the validity of each of these sources as well as to identify additional sources affecting VO. In this paper the literature related to the risk sources in the network level of the VOs will first be examined, and then the ISM methodology will be explained before mapping the relation between risk sources using ISM. Finally results of MICMAC (Duperrin and Godet, 1973) analysis of the sources will be presented in section 4. 2. Identification of risk sources in the Virtual Organisation A number of authors who have researched and written about this area have identified where risk arises in VO. Comprehensive examination of the literature identified 13 sources of risk and impediments which can be possible causes of failure to hit targets in the areas of delivery time and cost and quality and in some instances where the collaboration has collapsed completely. A brief discussion of each of these sources as identified in supply chain management literature and related literature follows and will be the basis of the subsequent survey and analysis. 2.1. Lack of trust This was the most frequently discussed VO risk. The degree of trust that exists between partners relates to how much partners believe in the honesty, generosity and overall competence of the others. Where there is no trust between partners problems arise; for instance they become unwilling to pass on sensitive information, find it difficult to agree about how finances should be managed. In short they do not work to promote VO collaboration. Trust and commitment are crucial to collaboration and for cooperation over a period of time together with a preparedness to share risks (Sahay and Maini, 2002). The more the trust between SC partners, the more the commitment (Mistry, 2005). However a lack of trust is one of main contributors to SC risks (Sinha et al., 2004). According to Lengnick-Hall (1998) where trust has grown out of good communication, it leads to resources that themselves can give a competitive edge. Trust assumes that those party to an agreement will not act opportunistically even when they are tempted by possible short-term advantage to themselves (Chiles and McMackin, 1996) and it can make a marked contribution to the stability of an organisation in the long term and to its network (Speckman and Davis, 2004). For collaboration to work all partners must work together to solve problems and this demands that there is powerful mutual trust and the commitment to expend time and effort (Camarinha-Matos and Afsarnesh, 2007). Where risk arises from the VO being exposed to partners’ opportunistic behaviour, or else to any uncertainty and ambiguity or only partial information, trust becomes vital in order to reduce it. However trust itself invites risk since it exposes those who trust to the opportunistic behaviour of others (Panteli and Sockalingham, 2005). There is an important constructive relationship between trust and conflict resolution in any inter-organisational arrangement (Twomey, 1995) and trust enables benefit to be gained from conflict since conflict can often lead to innovation (Pascale, 1994). Ryutor et al. (2007) consider that the establishment of trust between VO collaborators is essential to successful joint performance and presents a conceptual framework dealing with key concepts that will initiate trust between VO collaborators in on-demand VO. Lewicki et al. (1998); Zaher et al.(1998) argue that without trust it is not possible for there to be social, economic or political dealings since one party must initiate the contact making the unspoken assumption that there is likely to be a positive response from the other party. Adobre (2005) showed that the trust building in partnership maybe sort of self fulfilling prophecy in which initial expectation positively impact behaviour and trust building, where there may be some optimal level of expectation. In his study he found that both too low and too high an expectation was counterproductive to trust building. 2.2. Inadequate collaboration agreement A lack of clarity in the agreement into which partners enter is one of the circumstances that can lead to insufficient collaboration. According to Camarinha-Matos and Afsarmanesh, 2007; Westphal et al., 2007; Bullinger, 2003 where any definition dealing with objectives is weak, together with the strategies and basic conditions and where expectations are poorly managed, and contracts perceived as inadequate or unfair, there may well be risk of failure. The moral risk before the VO is set up was discussed by You and Yu (2006), when members may be able to reduce the extent of knowledge sharing though setting up a contract, with Quirchmayr et al. (2002) describing contracts for negotiation where a legally binding contract may have to be signed by each partner requiring them to accept the process, protocols, and constitution of the VO. Both risks and rewards must be shared among SC members according to Mentzer (2001) if Supply Chain Management (SCM) is to work efficiently since all members must respond to the same incentives which results in a fair distribution across the network of the risks as well as the costs and rewards of the business (Narayanan and Raman, 2004). Tsay (1999) describes the sharing of revenue as a type of SC contract that facilitates the sharing of risks also. The degree of risk and benefit sharing differs with the type of the collaboration. Joint ownership is often the mechanism with risks arising from joint venture or strategic alliance (Zheng et al., 1998). Methods of incentive for the collaborating parties may be through the use of obligation contracting, schemes for profit sharing and shared property rights as well as ownership control. Where collaborations are not so formal the nature of sharing risks and benefits may not be so clear, therefore lasting commitment may depend on agreements being made so that sensitive information, knowledge and competencies can be shared. Where there is trust there will be less reluctance to share sensitive information since the perception of risk is reduced. This in turns means that complex contracts are less likely to be seen as necessary to protect interests and there is trust in the decision making across the group since it is believed that all perspectives will be considered (Chiles and Mcmakin, 1996). 2.3. Ontology differences Ontology is a philosophical system that is concerned with the nature of being. Gruber (1995) defines ontology as “a formal, explicit specification of a shared conceptualisation”. Problems in this area of ontology crop up when there are two different words with the same meaning or even where one word means different things as it used by different partners. According to Plisson et al. (2007) ontologies can be said to offer an economical and unambiguous way of representing knowledge so that it can be jointly understood and therefore provide a basis for sharing. However, before terminology can be shared, there needs to be consensus between partners as to which terms they will use when collaborating where problems have previously arisen because ontologies have not been held in common between organisations working together (Camarinha-Matos, 2002). What ontology can do is provide a means of sharing knowledge where there is an understanding of concepts and relationships within a certain area and communication between those involved in this area where the fundamental ontology is one acting as a glossary for a limited vocabulary which are the agreed terms used within a specific area. Where problems occur as a result of differing ontologies there may be disagreements in relation to both the formation and the processes of collaboration which will add to the risks in the VO. 2.4. Heterogeneity of partners Heterogeneity means the differences that exist between partners in terms of incompatible hardware and operating systems, differences in language and the recording of data and in the meaning of the terms that are used. (Camarinha-Matos and Afsarmanesh, 2007; Sari et al., 2007) have all referred to this heterogeneity between possible partners as it exists in information technology infrastructures, working methods and business practice, as a possible obstacle to the VO operation. Where there is poor technological infrastructure this makes it difficult to put in place knowledge management (Singh and Kant, 2008). Jung (2008) saw variations between these categories as an impediment to the progress of efficient cooperation since the semantics of the information from various entities may differ. These variations occur because of the variations in terminology with a number of synonyms and antonyms in use but also and more importantly because of differences within the databases derived from the knowledge structures (Hull, 1997) and the ontology (Jung, 2005). 2.5. Structure and design VO’s dynamic structure creates problems since it is not possible to co-ordinate comparable activities because responsibilities are not adequately shared out and it may not be clear what the tasks and the rules are nor who is in overall charge, nor how far any control extends, and where a phase leader is unable to manage the other partners and does not have the right to make decisions the whole network is affected. Elements of structure and design include central planning or decentralisation, the extent of any specific control and specialisation and how labour is divided (Zhang and Dilts, 2004). According to Camarinha-Matos and Afsarmanesh (2007) where the infrastructure does not offer the opportunity for joint collaboration, this proves an obstacle to a VO. Control that is weak or ineffective can occur outside the project and can spring from inter- organisational networking in a way that resembles that in which (Finch, 2004; Bandyopadhyay, et al., 1999) found that where control over suppliers and customers was weak within the grouping, risks could occur. 2.6. Loss of communication The variation inherent in VO and the changes in structure can lead to less communication and here an inverse relationship comes into play since the less the communication the more the uncertainty. According to Grabowski and Roberts (1998) while communication is fundamental to any organisation it is of even more vital importance within VOs. Any communication in a virtual form must respond to specific customer demand with speed (David and Malone, 1992). Others (Camarinha-Matos and Afsarmanesh, 2007; Westphal et al., 2007; Bamford et al., 2004; Bullinger, 2003; Gunasekaran et al., 2008) have all seen failing communication as an obstacle within VOs and where this happens, failure could follow. It is likely that a more trusting relationship between collaborators will be the result of better communication; the crucial aspect is to go beyond technology to establish trust over key issues with a free flow of information and where the most critical issues can be raised with the partners being at ease even when exchanging sensitive information. Partners must understand that communication has to be about commitment levels rather than about technology (Speckman and Davis, 2004). 2.7. Culture differences There may be several cultures within a VO and this may lead to lack of alignment between processes and inaccurate communication impacting on the sharing of information (Grabowski and Roberts, 1998; Chen and Fang; Prefontain, 2003; Camarinha-Matos and Afsarmanesh; 2007). The fundamental values and beliefs of any culture, together with value norms and the customs underpinning the behaviour of people within an organization, define it (Singh and Kant, 2008; Lemken et al., 2000). However, where there is no over-arching culture in any organisation this has a negative impact on the successful management of knowledge (Chase, 1997). You et al. (2006) focus on members of a VO, saying the centre of any enterprise resides in its values and they therefore use such terms as staff value, faith and behaviour in relation to this so and those joining a VO undertaking come from different enterprises and therefore shared knowledge will be in short supply. It may be that those who come together to form a VO are culturally dissimilar and they may never before have worked as a partnership and so they will have little experience in common. This is an important issue since enterprises have become increasingly transactional and virtual partnerships have been made easier by technological advances and this brings with it different cultural backgrounds and different uses of language as well as values, all of which impact on the dispersed team function (Crossman and Kelley, 2004). 2.8. Bidding for several Virtual Organisations at the same time Some partners may choose to be active in several VOs simultaneously when they do not have the capacity to cope with this. Risk then occurs when one partner wins two or more VO bids and his capacity as a partner, either in terms of resources or staff, is not sufficient to undertake the tasks involved in more than one VO. Although this risk occurs, there is as yet little discussion of it in the literature and it has not been much researched. Even in Camarinha-Matos (2002) the vagueness is there when he discusses only the other causes of risk in a VO. Nguyen et al. (2005) discuss only resource management and bidding strategies as possibly problematic in a VO. 2.9. Lack of information sharing It is crucial for information to be shared where there is decreasing information visibility in the VO so that there is less risk including that of catalogue non-availability that includes up to date and standardised profiles of organizations. However, the availability of more information sharing can cause loss of IPR. In order for knowledge sharing to be accepted, a VO must have established values relating to sharing and collaboration as part of their fundamental ethos. Some may feel that they have an advantage because they possess knowledge that others do not and this causes a refusal to share knowledge with others out of a desire to protect their own interests (You et al., 2006). Networks must share information because where it is lacking the result may be panic, confused behaviour and increased costs (Childerhouse et al., 2003). It is agreed currently by models for SCM that sharing business information is vital, connecting SC completely together (Zhenxin et al., 2001; Yu et al., 2001). According to Rahman (2004) it is felt that there is a risk involved in sharing with other members such sensitive information as inventory levels and production schedules. Information sharing should be subject to choosing those with whom the information will be shared, what type of information it will be and of what quality. Efficient network coordination depends upon information sharing, with a number of studies finding that it impacts significantly on network performance and, in particular, is able to reduce the bullwhip effect. Information sharing leads to better operational decision making within enterprises which leads to more efficient use of resources and lower costs (Lee et al., 1997; Yu et al., 2001). 2.10. Lack of top management commitment Risk is increased where a weak part is played by top level management at particular points in VO formation or operations where crucial decisions are made (Camarinha-Matos and Afsarmanesh, 2007; Westphall et al., 2007; Bamford et al., 2004; Bullinger, 2003). According to Kanter (1997) there is a risk that low commitment to a partnership will lead to a failure to meet objectives. The role of top management is critical with it being responsible for all activities at every level of an organisation, for the technological infrastructure and for decision making in order that there will be efficient creation of knowledge together with sharing and use (Brand, 1998). 2.11. Lack of knowledge about risks Where there is no knowledge of the risks that may occur there is an increased likelihood that these risks will occur and also have a greater impact. According to Hallikas et al. (2004) where there is a greater understanding of the risks that may occur in an SC there is likely to be improved decision making and lower risk to each enterprise involved as well as to the whole undertaking. It is possible to categorise the many different forms of SC risks in terms of how their occurrence would affect a business and its environment (Harland et al., 2003). With an understanding of the range of SC risks and how they interact can come a response from the enterprise creating balanced and efficient risk reduction strategies (Chopra and Sodhi, 2004). It is important for organisations to come collectively to an understanding of the risks they may face (Jüttner, 2005). Risk analysis has the means to detect risk in a process (Sinha et al., 2004) and this enables a secure environment in which decisions can be taken so that there is a continuous assessment of the possibility of risk; it is possible to decide which are serious and then take appropriate action to deal with them (Shtub et al., 1994). 2.12. Wrong partner/s selection According to Camarinha-Matos and Afsarmanesh, 2007; Westphall et al., 2007; Bamford et al., 2004; Bullinger, 2003 such things as objectives, strategies, core competencies and capabilities that are irreconcilable cannot be complementary. While Sari et al. (2007) considered insufficient information about partners to be an obstacle to the VO selection, others including (Grabowiski and Roberts (1998); Chen and Fang (2002)) thought that a range of interests increased the risks to a VO. According to Wilmont and Hocker (2001) conflict can be viewed as the manifestation of a struggle between a minimum of two parties who are mutually dependant but who have divergent goals, with limited rewards, and who have other parties placing obstacles in the way of the achievement of their goals. Organisational conflict manifests itself through conflicting relationships or affective conflict, cognitive or task conflict, and conflict over process. Furthermore, Wong (2009) acknowledged that conflicts between partners is one of the obstacles of virtual networks, where Susman et al. (2003) suggested that conflicts could appear due to the polarisation of functions as between marketing and designers where they want customised product, whereas production department wants to manufacture standardised products. 2.13. Geographic location Risk may be increased by geographic locations with there being a direct correlation between distance and risk. Some locations throw up more problems because of, for instance, political and legal difficulties (Ritche and Brindly, 2002; Dewitt et al., 2006). Prater et al. (2001) looked at the size of any geographic area that a network covered, what political areas it encompassed and which borders were crossed, considering these all to be elements contributing to the partners’ exposure to risk. Porter (1998) observed that geographic distance gives rise to complications and to an increase in logistic cost. Also Grabowski and Roberts (1998), Chen and Fang (2002) considered that there was increased risk as a result of geographic distance. 3. Interpretive Structural Modeling (ISM) ISM is one of the Interactive Management methods which assist research groups in dealing with complex issues (Warfield, 1974; 1990; 2005). ISM transforms unclear, poorly articulated mental models of a system into visible well defined, hierarchal models. It is a well known methodology for identifying and summarising relationships among specific elements, which define an issue or a problem, and provide a means by which order can be imposed on the complexity of such elements (Mandal and Deshmukh 1994). Thus, a set of different and directly related elements are structured into a comprehensive systematic model. ISM is primarily intended as a group learning process, but individuals may also apply it (Ravi and Shankar, 2005, Faisal et al., 2007). Any methodology for dealing with complex issues, must, therefore, be able to break complexity down into manageable chunks of information so that the human mind can deal with it. ISM tries to do this, by enabling an individual or a group to focus on the interrelations between two elements in an issue at a time, without losing sight of the properties of the whole (Waller, 1975). From the risk sources which have been identified earlier and the potential impact of failure to meet delivery time, cost and quality targets or total failure for the collaboration, a questionnaire was developed using ISM methodology to determine underlying relations among these sources. ISM is a process that helps people to structure their collective knowledge and to model interrelationships in a way that enhances our ability to understand complexity. In other words, it helps to identify structure within a system of related elements and provide the opportunity to analyse it from several viewpoints. Steps involved in ISM methodology which are summarised in (Figure 1) are as follows: 1) Identification of the elements that are relevant to the problem or issue. 2) From the elements identified in the first step, establishing the contextual relationship among them. This represents the relationship indicating whether or not one element leads to another. 3) Developing a structural self-interaction matrix (SSIM) of sources which indicates a pair-wise relationship between sources of the system under consideration. 4) Developing a reachability matrix from the SSIM, and checking the matrix for transitivity. Transitivity of the contextual relation is basic assumption in ISM which states that if element A is related to element B, and B is related to C, then A is necessarily related to C. The SSIM format is transformed in the format of the reachability matrix by transforming the information in each entry of the SSIM into 1s and 0s in the reachability matrix. 5) The reachability matrix obtained in the fourth step is partitioned into different levels. 6) Based on the relationships in the reachability matrix, removal of the transitive links and drawing a directed graph. 7) Constructing the ISM model by replacing element nodes with statements. 8) The ISM model developed in the seventh step is reviewed to check for conceptual inconsistency, and to make the necessary modifications. These steps in more details below: 3.1. Survey A questionnaire was presented to the respondents of the survey and respondents were asked to complete it. We sent the questionnaire via the INTEROP-VLab mailing list which contains 224 members mainly experts in the interoperability and collaborations issues from both industry and academia. INTEROP-VLab is the "International Virtual Laboratory or Enterprise Interoperability", officially created as a non profit organisation under Belgian law (interop-vlab.eu, 2009). 45 experts from the INTEROP-VLab members participated in the study. Using the research data collected from these respondents and following the steps described above, the ISM directional graph is developed. Table 1 is a summary of the questionnaire of the survey. The questionnaire was presented to the respondents to the survey, and respondents were asked to answer a total of 78 questions, with each cell in the upper right cells representing a question. Respondents were asked to compare the column statement with the row statement for each cell and to choose a value from the set (V, A, O, or X) to represent their perception of direct relationship between two sources at each time. V represent relation when the first source i influence the second source j, but not in the reverse direction, A for the relation when the second source j influence the first source i, but not in the reverse direction, X represent interrelation between both sources (both directions), finally O represent that both sources (i and j) are unrelated. The symbolic values (V, A, X, or O) are translated into binary values (section 3.3) to develop a directional graph (figure 2). The detailed ISM methodology used to develop the directional graph is explained below. The contextual relation is established based on a pair- wise assessment of all the thirteen risk sources as shown in (Table 2), and the majority (70%) of the respondents agreeing to a specific relation between any two sources. With the use of this methodology, we can identify the direct and indirect relationships between risk sources in the VO. 3.2. The Structural self-interaction matrix (SSIM) ISM methodology suggests the use of the expert opinions in developing the contextual relationship among the risk sources. The approach relies on academic experts or managerial background from the INTEROP-VLab who answered our questionnaire with their opinions to arrive at the structure and relationship of the risk sources. Keeping in mind the contextual relationship for each source, the existence of a relationship between any two sources (i and j) and the associated direction of the relation are questioned. Based on the contextual relationships the SSIM (Table2) is developed for the 13 sources identified for the risk in the VO. 3.3. Reachability matrix The SSIM (Table 2) is transformed into a binary matrix, called the initial reachability matrix by substituting V, A, X and O by 1 and 0 as per the case. The rules for the substitution of 1s and 0s are as follows: • If the (i, j) entry in the SSIM is V, then the (i, j) entry in the reachability matrix becomes 1 and the (j, i) entry becomes 0. • If the (i, j) entry in the SSIM is A, then the (i, j) entry in the reachability matrix becomes 0 and the (j, i) entry becomes 1. • If the (i, j) entry in the SSIM is X, then the (i, j) entry in the reachability matrix becomes 1 and the (j, i) entry also becomes 1. • If the (i, j) entry in the SSIM is O, then the (i, j) entry in the reachability matrix becomes 0 and the (j, i) entry also becomes 0. After incorporating the transitivities as described in step 4 of the ISM methodology, the final reachability matrix is shown in table 3. In this table, the driving power of a particular source is the total number of sources (including itself) that it influences. The dependences is the total number of sources (including itself) that it may help influencing its growth. These driving power and dependency values will be used in classification of risk sources (MICMAC analysis). 3.4. Level partitions The reachability and antecedent set (Warfield, 1974) for each source are obtained from final reachability matrix. The reachability set for a particular source consists of the source itself and the other sources, which it influences. The antecedent set consists of the source itself and the other sources, which may influence it. Subsequently, the common sources of the reachability and antecedent sets from the intersection sets are the same as assigned as the top- level element in the ISM hierarchy as it would not help achieve any other source above their own level. After the identification of the top-level source, it is discarded from further hierarchical analysis (i.e. removing that source from all the different sets). For example, as seen in Table 4, ‘Wrong partner/s selection’ (Source 12) is found at level 1 due to similar reachability and intersection sets. Thus, it would be positioned at the top of the ISM hierarchy. Source 12 is then removed from all different sets for further analysis, as its level has been obtained This iteration is repeated till the levels of each source are found out (Tables 4 to 10). The identified levels aids in building the digraph and the final model of ISM. 3.5. Formation of ISM-based model The structural model (Figure 2) is generated from the final reachability matrix and the digraph is drawn. Removing the transitivities as described in the ISM methodology, the digraph is finally converted into the ISM. The contextual relationship in this structure was “leads to”. This implies that each arrow is read as “leads to”. 4. Classification of the risk sources (MICMAC analysis) MICMAC was developed by Duperrin and Godet (1973) to study the diffusion of impacts through reaction paths and loops for developing hierarchies for members of an element set. MICMAC analysis can be used to identify and analyse the elements in a complicated system (Warfield, 1990). Generally, the elements will be classified into four clusters of autonomous, dependent, linkage and independent (driver) sources according to the driving power and dependencies of all the elements (Ravi and Shankar, 2005). The objective of the MICMAC analysis is to analyse the driving power and the dependence of the elements (Mandal and Deshmukh, 1994, Faisal et. al., 2006). In this analysis, the risk sources described earlier are classified into four clusters. The first cluster consists of the “autonomous sources” that have weak driving power and weak dependence. These sources are relatively disconnected from the system, with which they have only few links, which may not be strong. In our case there are no sources in the autonomous cluster which indicates no sources can be considered as disconnected from the whole system and the management has to pay attention to all the identified risk sources in VO. The “dependent sources” constitute the second cluster which has weak driving power but strong dependence. We have six sources in this cluster; Lack of top management commitment, lack of information sharing, lack of trust, bidding for several VO at the same time, lack of knowledge about risks and wrong partner/s selection. It forms the top level in the ISM hierarchy. It represents the source that is the resultant action for risks in VO. Its strong dependence indicates that it requires all the other risks to come together so as to increase them in VO risks. The third cluster has the “linkage sources” that have strong driving power and strong dependence. These sources are unstable due to the fact that any change occurring to them will have an effect on others and also a feedback on themselves. Just one of the risk sources in this cluster which is loss of communication is influenced by lower level sources and in turn impacts other sources in the model. The fourth cluster includes the “independent sources” having strong driving power but weak dependence. Six risk sources are in this cluster: geographic location, culture differences, ontology differences, heterogeneity of partners, inadequate collaboration agreement and structure and design. These six sources play a key role in risks in a VO. The objective behind this classification of the risk sources is to analyse the driver power and dependency of sources (Jharkharia and Shankar, 2005). In general, higher sources driver power means that a large number of sources could be easily removed by its removal. Higher dependence values for sources require a large set of sources to be addressed before its removal and a more likely success in the implementation of VO risks. The classification of risk sources within the four clusters helps identify the difficulty removal potential of the risk sources. The driving power and the dependence of each of these sources are shown in (Table 4). In this table, an entry of ‘1’ along the columns and rows indicates the dependence and driving power, respectively. Subsequently, the driver power–dependence diagram is constructed which is shown in (Figure 3). 5. Discussion The objective of the ISM model in this research is to develop a hierarchy of risk sources that would help mitigate risks in VO. The model developed in this paper provides the opportunity to understand the network risk sources in VOs. It is clear that awareness about risk sources is very important as it would lead to efforts being undertaken to minimise these risks. The task of management is to place high priority on those sources that form the base of ISM model because it is they who would drive other sources. The significant features of ISM methodology can be summarised as: • it provides an understanding of the relationships among the risk sources in the network level in the VO; • classification of sources under autonomous, dependent, linkage and independent categories; • it attempts to develop a new understanding of risks sources due to the SMEs’ collaboration. • suggested methodology would help the SMEs to develop strategies to mitigate risks. 6. Conclusion and future work Based on the literature review and expert opinions, a number of risk sources in the VO which may cause to have negative effects on the time, cost, quality or total failure for the collaboration have been identified. The 13 risk sources identified in this research have significant overlaps and relationships that are sometimes difficult to see. A more complete understanding of these risks sources and their relationships, through logical structure, will help enterprises to take better decision as to whether to participate in a particular VO. For this reason we have sought to present an application of the ISM process to the data collected on relationships between the risk sourcess. ISM can only act as a tool for imposing order and directions on the complexity of relationships among the variables. It does not suggest any relative weights associated with the variables (Kannan et al., 2008). Even so, this model can be applied with other approaches such as the Analytical Network Process (ANP) (Saaty, 2001), which requires a decision structure to help determine the weights of each risk source. Simulation and systems dynamics modeling may also be used to help identify how risk sources in VO will influence it and its performance results. We therefore conclude that there is a sufficient body of expert opinion relating to the risks arising in the creation and operation of VOs to support a meaningful analysis leading to enhanced understanding of the relationships between sources of risks and the relative significance (weights) of these risks. This can inform an SME (or indeed larger enterprise) decision to participate in a VO, as well as operational decision-making once a VO is established. For this reason we consider that deeper analysis may yield further valuable insights. In the future work we can quantify the risks in VO. This depends upon the degree of inheritance of various variables and the amount of interaction that is possible by using Delphi, AHP, ANP, SEM or fuzzy logic. While using graph theory and matrix methods the interaction among the variables can be easily analysed and transformed into mathematical equation. The graph theory and matrix methods consist of the diagraph representation, the matrix representation and the permanent matrix representation. The digraph is the visual representation of the sources and their interdependencies. The matrix converts the digraph into mathematical form and the permanent function is a mathematical representation that helps to determine a risk index. This would enable SMEs to understand the contribution of various risk sources and put their efforts into mitigating these risks. Furthermore, fuzzy ISM, a step ahead of binary ISM, may also be carried out. While only the existence of relations is considered between elements in the ISM, the strength of relations is additionally considered in FISM. The strength of relation can be quantified using 0-1 scale. Additional future research could include broadening the inputs and validation with more practitioners and by evaluating an actual set of risk sources in a real case study with an experimental approach to determine if the model’s relationships are influenced as hypothesised on expert opinions. In addition to the above directions identified already, Structural Equation Modelling (SEM) has the capability of testing the validity of such hypothetical models. Therefore, it may be applied in future research to test the validity of this model. SEM models can be tested using LISREL or AMOS software. SEM models also provide the path coefficients for the different relationships among the sources. This would complement the MICMAC analysis to further strengthen the understanding about the more important relationships requiring maximum attention. References ADOBOR, H. (2005) 'Trust as Sensemaking: The Microdynamics of Trust in Interfirm Alliances.' Journal of Business Research 58, 330-337 BAMFORD, J., ERNEST, D. & FUBINI, D. G. (2004) Launching a World-Class Joint Venture. Harvard Business Review 82, 90-100. BANDYOPADHYAY, K., MYKYTYN, P. P. & MYKYTYN, K. (1999) A framework for integrated risk management in information technology. Management Decision, 37, 437-445. BLACKHURST, J., WU, T. & O’GRADY, P. (2004) Network-based approach to modelling uncertainty in a supply chain. International Journal of Production Research, 42, 1639-1658.00 BRAND, A. (1998) Knowledge management and innovation Journal of Knowledge Management, 2, 17-22. BULLINGER, H.-J. (2003) Collaborative development – potentials of success in development networks. Supply Chain Management, 2, 31. CAMARINHA-MATOS, L. M. (2002) Virtual Organizations in Manufacturing: Trends and challenges. Proceedings of Flexible Automation and Intelligent Manufacturing. Munich, 1036-1054. CAMARINHA-MATOS, L. M. & AFSARMANESH, H. (2007) A framework for virtual organization creation in a breeding environment. Annual Reviews in Control, 31, 119- 135. CHASE, R. L. (1997) The knowledge-based organization: an international survey. Journal of Knowledge Management, 1, 38-49. CHEN, J. & FENG, W. D. (2002) Formation and Management of Virtual Enterprises Tsinghua University Press. CHILDERHOUSE, P., HERMIZ, R., MASON-JONES, R., POPP, A. & TOWILL, D. R. (2003) Information flow in automotive supply chains – identifying and learning to overcome barriers to change. Industrial Management & Data Systems, 103, 491-502. CHILES, T. H. & MCMAKIN, J. F. (1996) Integrating variable risk preferences, trust, and transaction cost economics. Academy of Management Review, 21, 73-99. CHOPRA, S. & SODHI, M. S. (2004) Managing risk to avoid supply chain breakdown. Sloan Management Review, Fall, 53-61. CROSSMAN, A. & LEE-KELLEY, L. (2004) Trust, commitment and team-working: the paradox of the virtual organization. Global Networks, 4, 375-390. DAVIDOW, W. H. & MALONE, M. S. (1992) The Virtual Corporation, New York, NY, HarperCollins Publishers. DEWITT, T., GIUNIPERO, L. C. & MELTON, H. L. (2006) Clusters and supply chain management: the Amish experience. International Journal of Physical Distribution & Logistics Management, 36, 289-308. DUPERRIN, J. C., AND GODET, M. (1973) Methode De Hierar Chization Des Elements D’um System. Rapport Economique De CEA, 45-51. FAISAL, M., BANWET, D. K. & SHANKAR, R. (2007) Quantification of risk mitigation environment of supply chains using graph theory and matrix methods. European Journal of Industrial Engineering, 1, 22-39. FINCH, P. (2004) Risk Management Consultant with AEA Technology, Warrington, UK. Supply Chain Management: An International Journal, 9, 183-196. GRABOWSKI, M. & ROBERTS, K. H. (1998) Risk Mitigation in Virtual Organizations. Organization Science, 10. GRUBER, T. (1995) Toward principles for the design of ontologies used for knowledge sharing. International Journal of Human Computer Studies., 43, 907-928. GUNASEKARAN, A., LAI, K.-H. & CHENG, T. C. E. (2008) Responsive supply chain: A competitive strategy in a networked economy. Omega, 36, 549-564. HALLIKAS, J., KARVONEN, I., PULKKINEN, U., VIROLAINEN, V. M. & TUOMINEN, M. (2004) Risk management processes in supplier networks”. International Journal of Production Economics, 90, 47-58. HARLAND, C., BRENCHLEY, R. & WALKER, H. (2003) Risk in supply networks. Journal of Purchasing & Supply Management, 9, 51-62. HULL, R. (1997) Managing semantic heterogeneity in databases: a theoretical prospective. Proceedings of the sixteenth ACM SIGACT-SIGMOD-SIGART symposium on Principles of database systems. Arizona, United States ACM. JHARKHARIA, S., SHANKAR, R. (2005) IT enablement of supply chains: modeling the enablers. International Journal of Productivity and Performance Management, 53, 700-712. JUNG, J. (2008) Taxonomy alignment for interoperability between heterogeneous virtual organizations. Expert Systems with Applications, 34, 2721-2731. JUNG, J., LEE, K.-S., PARK, S.-B. & JO, G.-S. (2005) Efficient web browsing with semantic annotation: a case study of product images in e-commerce sites. IEICE Transactions on Information and Systems, E88-D, 843-850. JÜTTNER, U. (2005) Understanding the business requirements from a practitioner perspective. The International Journal of Logistics Management, 16, 120-141. JÜTTNER, U., PECK, H. & CHRISTOPHER, M. (2002) Supply chain risk management: outlining an agenda for future research. IN GRIFFITHS, J., HEWITT, F. & IRLEAND, P. (Eds.) Proceedings of the Logistics Research Network 7th Annual Conference. KANNAN, G., HAQ, A. N., SASIKUMAR, P. & ARUNACHALAM, S. (2008) Analysis and selection of green suppliers using interpretative structural modelling and analytic hierarchy process. International Journal of Management and Decision Making, 9, 163-182. KANTER, R. M. (1997) Restoring people to the heart of the organisation of the future. IN JOSSEY-BASS (Ed.) The Organisation of the Future. San Francisco. LEE, H. L., PADMANABHAN, V. & WANG, S. (1997) Information distortion in a supply chain: the bullwhip effect. Management Science, 43, 546-558. LEMKEN, B., KAHLER, H. & RITTENBRUCH, M. (2000) Sustained knowledge management by organizational culture. Proceedings of the 33rd Annual Hawaii International Conference on System Sciences. Germany, Institution of Computer Science, Bonn University. LENGNICK-HALL, C. A. (1998) Customer contributions to quality: a different view of the customer-oriented firm. Academy of Management Review, 21, 791-824. LEWICKI, R., MCALLISTER, D. & BIES, R. (1998) Trust and Distrust: New Relationships and Realities. Academy of Management Review, 23, 438-458. MANDAL, A., DESHMUKH, S.G. (1994) Vendor selection using interpretive structural modeling (ISM). International Journal of Operations & Production Management, 14, 52-61. MENTZER, J. T. (2001) Defining Supply Chain Management. Journal of Business logistics, 22, 1-25. MISTRY, J. J. (2005) Origins of profitability through JIT processes in the supply chain. Industrial Management & Data Systems, 105, 752-768. NARAYANAN, V.G. & RAMAN, A., (2004). Aligning incentives in supply chains. Harvard Business Review, 82 (11), 94–102. NGUYEN, D., THOMPSON, S. G., PATEL, J., TEACY, L. W., JENNINGS, N. R., LUCK, M., DANG, V., CHALMERS, S., OREN, N., NORMAN, T. J., PREECE, A., GRAY, P. M., SHERCLIFF, G., STOCKREISSER, P. J., SHAO, J., GRAY, W. A. & FIDDIAN, N. J. (2006) Delivering services by building and running virtual organisations BT Technology Journal 24, 141-152. PANTELI, N. & SOCKALINGAMB, S. (2005) Trust and conflict within virtual inter- organizational alliances: a framework for facilitating knowledge sharing Decision Support Systems, 39, 599-617. PASCALE, R. T. (1994) Intentional breakdowns and conflict by design. Planning Review, 22, 12-16. PLISSON, J., LJUBIC, P., MOZETIC, I. & LAVRAC, N. (2007) An Ontology for Virtual Organization Breeding Environments. IEEE Transactions on Systems, Man, and Cybernetics, Part C 37, 1327-1341. PORTER, M. E. (1998) Clusters and the new economics of competition. Harvard Business Review, November-December, 77-90. PRATER, E., BIEHL, M. & SMITH, M. A. (2001) International supply chain agility. International Journal of Operations &Production Management, 21, 823-839. PREFONTAINE, L. (2003) Risk Management in New Models of Collaboration in New Models of Collaboration: A Guide for Managers. QUIRCHMAYR, G., MILOSEVIC, Z., TAGG, R., COLE, J. B. & KULKARNI, S. (2002) Establishment of Virtual Enterprise Contracts. Proceedings of the 13th International Conference on Database and Expert Systems Applications Berlin, Springer RAHMAN, Z. (2004) Use of internet in supply chain management: a study of Indian companies. Industrial Management & Data Systems, 104, 31-41. RAVI, V., SHANKAR, R (2005) Analysis of interactions among the barriers of reverse logistics. Technological Forecasting and Social Changes, 72, 1011-1029. RITCHIE, B. & BRINDLEY, C. ( 2002) Reassessing the management of the global supply chain. Integrated Manufacturing Systems, 13, 110-116. RYUTOV, T., NEUMAN, C., ZHOU, L. & FOUKIA, N. (2007) Initial trust formation in Virtual Organisations. International Journal of Internet Technology and Secured Transactions, 1. SAATY, T. L. (2001) Decision-making with Dependence and Feedback: The Analytic Network Processes, Pittsburgh, RWS Publications. SAGE, A. P. (1977.) Interpretive Structural Modeling: Methodology for Large-scale Systems, New York, NY, McGraw-Hill. SAHAY, B. S. & MAINI, A. (2002) Supply chain: a shift from transactional to collaborative partnership. Decision, 29, 67-88. SARI, B., AMAITIK, S. & KILIC, S. E. (2007) A neural network model for the assessment of partners? performance in virtual enterprises. The International Journal of Advanced Manufacturing Technology, 34, 816-825. SHTUB, A., BARD, J. F. & GLOBERSON, S. (1994) Project Management: Engineering, Technology and Implementation, Englewood Cliffs, NJ, Prentice-Hall. SINGH, M. D. & KANT, R. (2008) Knowledge management barriers: An interpretive structural modeling approach. International Journal of Management Science and Engineering Management, 3, 141-150. SINHA, P. R., WHITMAN, L. E. & MALZAHN, D. (2004) Methodology to mitigate supplier risk in an aerospace supply chain. Supply Chain Management: An International Journal, 9, 154-168. SPECKMAN, R. E. & DAVIS, E. W. (2004) Risky business: expanding the discussion on risk and the extended enterprise. International Journal of Physical Distribution & Logistics management, 34, 414. SUSMAN, G. L., GRAY, B. L., PERRY, J. & BLAIR, C. E. (2003) 'Recognition and Reconcilation of Differences in Interpretation of Misalignments When Collaborative Technologies Are Introduced into New Product Development Teams.' Journal of Engineering and Technology Management 20, (1-2) 141-159. The international Virtual laboratory for Enterprise Interoperability. Source [online]. Available from http://interop-vlab.eu/ [accessed 9 November 2009]. TSAY, A. A. (1999) The Quantity Flexibility Contract and Supplier-Customer Incentives Source Management Science archive, 45, 1339-1358. TWOMEY, D. (1995) Inter-organizational conflict resolution: the effects of power and trust. Academy of Management Proceedings. WALLER, R. J. (1975) Application of interpretive structural modeling to priority-setting in urban systems management. In BALDWIN, M. (Ed.) Portraits of complexity. Columbus, OH, Battelle Memorial Institute. WARFIELD, J. (1974.) Developing Interconnected Matrices in Structural Modelling. IEEE Transactions on Systems Men and Cybernetics, 4, 51-81. WARFIELD, J. (1990) A science of generic design; Managing complexity through systems design, CA, Salinas. WESTPHAL, I., THOBEN, K.-D. & SEIFERT, M. (2007) Measuring Collaboration Performance In Virtual Organizations. Establishing The Foundation Of Collaborative Networks. Guimarães, Springer Boston. WILMOT, W. W. & HOCKER, J. L. (2001) Interpersonal conflict, New York, McGraw- Hill. WONG, E. T. T. (2009) 'Trust Management in Virtual Product Development Networks.' In Encyclopedia of Information Science and Technology, Second Edition. ed. by Khosrow-Pour, M. Hershey, USA: IGI Global: 3831-3839 YOU, T.-H., ZHU, Z. & YU, Z.-C. (2006 ) Analysis and Assessment of Knowledge Sharing Risk in the Virtual Enterprise. JCIS-2006 Proceedings YU, Z., YAN, H. & CHENG, T. C. E. (2001.) Benefits of information sharing with supply chain partnerships. Industrial Management & Data Systems, 101, 114-121. ZAHEER, A., MCEVILY, B. & PERRONE, V. (1998) Does trust matter? Exploring the effects of interorganizational and interpersonal trust on performance. Organization Science, 9, 141-159. ZHANG, Y. & DILTS, D. (2004) System dynamic of supply chain network organization structure. Information Systems and E-Business Management, 2, 187-206. ZHENG, J., JOHNSEN, T., HARLAND, C. M. & LAMMING, R. C. (1998) Initial conceptual framework for creation and operation of supply networks. Proceedings of the 14th IMP Annual Conference. Turku. ZHENXIN, Y., YAN, H. & CHENG, T. C. (2001) Benefits of information sharing with supply chain partnerships. Industrial Management & Data Systems, 101, 114-120. Table 1. Summary of the questionnaire. Risk sources 1 2 3 4 5 6 7 8 9 10 11 12 13 1) Lack of trust 2) Inadequate collaboration agreement 3) Heterogeneity of partners 4) Ontology differences 5) Structure and design 6) Loss of communication 7) Culture 8) Bidding for several VO at the same time 9) Information sharing 10) Top management commitment 11) Knowledge about risks 12) Wrong partner selection 13) Geographic location Contextual relationship = leads to What to enter in the white cells Enter V when the row influences the column Enter A when the column influences the row Enter O when there is no relation between the row and the column Enter X when row and column influences each other Table 2. Structural self-interaction matrix (SSIM). Risk sources 1 2 3 4 5 6 7 8 9 10 11 12 13 1. Lack of trust between partners A A A A A A V X X O V A 2. Inadequate collaboration agreement V O A O V O V V V V V O 3. Heterogeneity of partners V O O V V O V V V V V O 4. Ontology differences V V O O V O V V V V V O 5. Structure and design V O A O V O V V V V V O 6. Loss of communication V A A A A A V V V V V A 7. Culture differences V O O O O V V V V V V O 8. Bidding for several VO at the same time A A A A A A A O A V V A 9. Lack of information sharing X A A A A A A O X V V A 10. Lack of top management commitment X A A A A A A V X O V A 11. Lack of knowledge about risks O A A A A A A A A O V A 12. Wrong partner/s selection A A A A A A A A A A A A 13. Geographic location V O O O O V O V V V V V Table 3. Reachability matrix table. Risk source 1 2 3 4 5 6 7 8 9 10 11 12 13 Driver 1. Lack of trust between partners 1 0 0 0 0 0 0 1 1 1 0 1 0 5 2. Inadequate collaboration agreement 1 1 0 0 0 1 0 1 1 1 1 1 0 8 3. Heterogeneity of partners 1 0 1 0 1 1 0 1 1 1 1 1 0 9 4. Ontology differences 1 1 0 1 0 1 0 1 1 1 1 1 0 9 5. Structure and design 1 0 0 0 1 1 0 1 1 1 1 1 0 8 6. Loss of communication 1 0 0 0 0 1 0 1 1 1 1 1 0 7 7. Culture 1 0 0 0 0 1 1 1 1 1 1 1 0 8 8. Bidding for several VO at the same time 0 0 0 0 0 0 0 1 0 0 1 1 0 3 9. Information sharing between partners 1 0 0 0 0 0 0 0 1 1 1 1 0 5 10. Lack of top management commitment 1 0 0 0 0 0 0 1 1 1 0 1 0 5 11. Knowledge about risks 0 0 0 0 0 0 0 0 0 0 1 1 0 2 12. Wrong partner/s selection 0 0 0 0 0 0 0 0 0 0 0 1 0 1 13. Geographic location 1 0 0 0 0 1 0 1 1 1 1 1 1 8 Dependence 10 2 1 1 2 7 1 10 10 10 10 13 1 Table 4. Iteration 1 Risk source Reachability set Antecedent set Intersection set Level 1. Lack of trust between partners 1,8,9,10,12 1,2,3,4,5,6,7,9,10,13 1,9,10 - 2. Inadequate collaboration agreement 1,2,6,8,9,10,11,12 2,4 2 - 3. Heterogeneity of partners 1,3,5,6,8,9,10,11,12 3 3 - 4. Ontology differences 1,2,4,6,8,9,10,11,12 4 4 - 5. Structure and design 1,5,6,8,9,10,11,12 3,5 5 - 6. Loss of communication 1,6,8,9,10,11,12 2,3,4,5,6,7,13 6 - 7. Culture differences 1,6,7,8,9,10,11,12 7 7 - 8. Bidding for several VO at the same time 8,11,12 1,2,3,4,5,6,7,8,10,13 8 - 9. Lack of information sharing 1,9,10,11,12 1,2,3,4,5,6,7,9,10,13 1,9,10 - 10. Lack of top management commitment 1,8,9,10,12 1,2,3,4,5,6,7,9,10,13 1,9,10 - 11. Lack of knowledge about risks 11,12 2,3,4,5,6,7,8,9,11,13 11 - 12. Wrong partner/s selection 12 1,2,3,4,5,6,7,8,9,10, 11,12,13 12 1 13. Geographic location 1,6,8,9,10,11,12,13 13 13 - Table 5. Iteration 2 Risk source Reachability set Antecedent set Intersection set Level 1. Lack of trust between partners 1,8,9,10 1,2,3,4,5,6,7,9,10,13 1,9,10 - 2. Inadequate collaboration agreement 1,2,6,8,9,10,11 2,4 2 - 3. Heterogeneity of partners 1,3,5,6,8,9,10,11 3 3 - 4. Ontology differences 1,2,4,6,8,9,10,11 4 4 - 5. Structure and design 1,5,6,8,9,10,11 3,5 5 - 6. Loss of communication 1,6,8,9,10,11 2,3,4,5,6,7,13 6 - 7. Culture differences 1,6,7,8,9,10,11 7 7 - 8. Bidding for several VO at the same time 8,11 1,2,3,4,5,6,7,8,10,13 8 - 9. Lack of information sharing 1,9,10,11 1,2,3,4,5,6,7,9,10,13 1,9,10 - 10. Lack of top management commitment 1,8,9,10 1,2,3,4,5,6,7,9,10,13 1,9,10 - 11. Lack of knowledge about risks 11 2,3,4,5,6,7,8,9,11,13 11 2 13. Geographic location 1,6,8,9,10,11,13 13 13 - Table 6, Iteration 3 Risk source Reachability set Antecedent set Intersection set Level 1. Lack of trust between partners 1,8,9,10 1,2,3,4,5,6,7,9,10,13 1,9,10 - 2. Inadequate collaboration agreement 1,2,6,8,9,10 2,4 2 - 3. Heterogeneity of partners 1,3,5,6,8,9,10 3 3 - 4. Ontology differences 1,2,4,6,8,9,10 4 4 - 5. Structure and design 1,5,6,8,9,10 3,5 5 - 6. Loss of communication 1,6,8,9,10 2,3,4,5,6,7,13 6 - 7. Culture differences 1,6,7,8,9,10 7 7 - 8. Bidding for several VO at the same time 8 1,2,3,4,5,6,7,8,10,13 8 3 9. Lack of information sharing 1,9,10 1,2,3,4,5,6,7,9,10,13 1,9,10 3 10. Lack of top management commitment 1,8,9,10 1,2,3,4,5,6,7,9,10,13 1,9,10 - 13. Geographic location 1,6,8,9,10,13 13 13 - Table 7, Iteration 4 Risk source Reachability set Antecedent set Intersection set Level 1. Lack of trust between partners 1,10 1,2,3,4,5,6,7,10,13 1,10 4 2. Inadequate collaboration agreement 1,2,6,10 2,4 2 - 3. Heterogeneity of partners 1,3,5,6,10 3 3 - 4. Ontology differences 1,2,4,6,10 4 4 - 5. Structure and design 1,5,6,10 3,5 5 - 6. Loss of communication 1,6,10 2,3,4,5,6,7,13 6 - 7. Culture differences 1,6,7,10 7 7 - 10. Lack of top management commitment 1,10 1,2,3,4,5,6,7,10,13 1,10 4 13. Geographic location 1,6,10,13 13 13 - Table 8, Iteration 5 Risk source Reachability set Antecedent set Intersection set Level 2. Inadequate collaboration agreement 2,6 2,4 2 - 3. Heterogeneity of partners 3,5,6 3 3 - 4. Ontology differences 2,4,6 4 4 - 5. Structure and design 5,6, 3,5 5 - 6. Loss of communication 6 2,3,4,5,6,7,13 6 5 7. Culture differences 6,7 7 7 - 13. Geographic location 6,13 13 13 - Table 9, Iteration 6 Risk source Reachability set Antecedent set Intersection set Level 2. Inadequate collaboration agreement 2 2,4 2 6 3. Heterogeneity of partners 3,5 3 3 - 4. Ontology differences 2,4 4 4 - 5. Structure and design 5 3,5 5 6 7. Culture differences 7 7 7 6 13. Geographic location 13 13 13 6 Table 10, Iteration 7 Risk source Reachability set Antecedent set Intersection set Level 3. Heterogeneity of partners 3 3 3 7 4. Ontology differences 4 4 4 7 Figure 1. Flow diagram for preparing ISM. Figure 2. ISM model. Figure 3. Driver power–dependence diagram. Popplewell1 Interpretive structural modeling of risk ... 
work_wuyydbnhxbeoxizs76ybyrus7y	----	An expert system for optimizing computer aided design of post frame buildings An Expert System for Optimizing Computer Aided Design of Post Frame Buildings K. G. G e b r e m e d h i n , S. S. Jagdale, R. G u p t a ASSOC. MEMBER ASAE ABSTRACT A procedure has been developed to optimize designs of metal-clad post-frame buildings using an expert system. The novel part of the study is the development of a procedure that extends the solution of an engineering problem from a conventional algorithmic Computer Aided Design (CAD) program to be used by the knowledge base of an expert system. The expert system then uses this information to optimize the design based on specified rules. The system shows a procedure for linking and sharing information between CAD and expert system programs. INTRODUCTION Recently developed programs for Computer Aided Design (CAD) of structures perform the tedious and complex numerical computations and present the design in graphic forms. In addition, some CAD programs check for the consistency of the design according to design specification and perform error checks. In "interactive" CAD systems, some degree of interaction is provided between the user and the machine (Gebremedhin, 1986; Suddarth and Wolfe, 1984; Adeli and Al-Rijleh, 1987; Adeli and Fiedorek, 1986). In these programs, most, if not all, of the decision making process is left to the user. These computer programs have either no or a very limited knowledge base. In addition, they cannot explain their line of reasoning or justify their answers. A natural extension of interactive CAD programs would be knowledge-based expert systems (Adeli, 1986). The primary characteristics of expert systems have been delineated in Waterman (1986) and Hayes-Roth et al. (1983). Among them are the separation of the knowledge base (KB) from the inference mechanism, use of heuristics and the explanation facility. Expert systems tend to mimic the decision-making and reasoning process of human experts by providing expert advice, answering questions, and justifying their conclusions. A major advantage is that its modular structure simplifies the development, maintenance, and updating of the expert system (ES). Another advantage of this method is the case of KB modification. The knowledge base in an ES is maintained as a set of rules which are blocks of knowledge that may be applied independently of one Article was submitted for publication in October 1988; reviewed and approved for publication by the Emerging Technology Div. of ASAE in April 1989. The authors are: K. G. GEBREMEDHIN, Associate Professor, Agricultural and Bilological Engineering Dept.; S. S. JAGDALE, Graduate Student, Mechanical and Aerospace Engineering Dept.; and R. GUPTA, Graduate Student, Agr. and Bio. Eng. Dept., Cornell University, Ithaca, NY. another (depending on the particular problem at hand). T h e r e f o r e , a d d i t i o n of m o r e r u l e s or deletion/modification of existing rules can be done very c o n v e n i e n t l y . T h i s is in c o n t r a s t with t h e algorithmic/sequential methodology of conventional programming where any modification involves substantial changes in the source code of the program itself. In this article, a procedure is presented to optimize designs of metal-clad, post-frame buildings using a knowledge based expert system. This was accomplished by first solving a post-frame structure including diaphragm action using a frame analysis program. Then, the solutions (stresses and deflections) obtained from the frame analysis program were used by the KB of the expert system. The expert system optimizes the design based on specified rules. This procedure extends not only the solution of an engineering CAD problem, but also the concept of communication that is needed to share information between CAD programs (separately available to perform complex calculations and provide results) and expert system programs (that use rules to deduce conclusions on the basis of data provided). The problem has long been the incompatibility of these problems and the lack of communication that is needed to share information between them. The frame analysis program used in this study was METCLAD (Gebremedhin, 1988a). Objective The system, henceforth referred to as Met-X-PERT, has been developed on an IBM/PC with a hard disk drive and 512K of RAM. Met-X-PERT is developed in Turbo Pascal. The program is limited to the design of metal- clad, post-frame rectangular buildings that exhibit diaphragm action by the roof envelope. The program is not intended for practical application because of the inherent difficulty of incorporating in program rules all assumptions, requirements, and judgements associated with design. The purpose of this article is, therefore, to develop a procedure for linking and sharing information between conventional algorithmic CAD program and an expert system, so that others developing similar expert systems for other applications would benefit. Besides, it may be used as a teaching tool in timber engineering, computer-aided design, and expert system courses. DESIGN ALGORITHM Optimization of individual members of a structure begins after the output of the frame analysis program (METCLAD) is completed. An interface routine extracts the stress and deflection information from the output of Vol. 5(3):September 1989 © 1989 American Society of Agricultural Engineers 0883-8542/89/0503-0447$03.00 447 METCLAD and creates a fact-base. These facts are displayed on the screen. The stresses and deflections are analyzed based on the rules specified in the knowledge base. The knowledge representation in Met-X-PERT is through production rules. The program supports numeric facts. A numeric fact is a fact with a numeric value. Production rules are the basic building blocks of the knowledge base. The production rules consist of antecedent and conclusion parts. The conditional antecedent part consists of IF, and AND statements. If all the conditional statements are TRUE, then the state of the fact of the THEN statement is added to the working memory. A production rule is expressed as IF ^condition 1> AND <condition 2> THEN <conclusion> The advice and actions taken are displayed on the screen. The actions that may be taken include a change in post size, truss member size, stress grade of lumber, or gauge of the metal skin. To take any or all of these actions, the program looks at several limiting factors for design, for example, the shear force in the metal skin, the Combined Stress Interaction Index (CSI) value of a truss chord member, deflection, and the CSI value of a post. If there are suggestions or advice to make several modifications in the design (e.g., change metal sking gauge and increase size of a truss chord, etc.), then all these changes are made at once. The only limitation is that the program looks at only one truss chord member and one post at a time. So, if there are, say, three truss chord members which need modification, it takes into consideration only the first one during the present iteration. If this problem (^StarP) . Analyze structure (METCLAD) A / Rule File /- ^ / METCLAD > / Output / " Create Input File (MDMS) Modify Input File • — • > Read-Rules (create knowledge-base) 1 ---* 1 Create Fact-base (interface) 1 i Display Facts * Pursue "Advice" (Inference Engine) \ r Display Advice no ^ ^ l s " s > \ « <^Design e f f i c i e n t ? ^ Stop Fig. 1—Sequence of actions in Met-X-PERT. gets solved in one iteration, then during the next iteration, it looks at the next chord. In other words, it looks at one element in a set of similar elements (truss chords are one set of elements, truss web members are another set of elements, posts are yet another set, etc.). The changes triggered by the inference engine modifies the input file to METCLAD. METCLAD is then executed again using the modified input file. A new output file is generated and this output file is similarly evaluated by the expert system. This continues until all limiting factors for design are met and an "optimum" design is simulated. A flow chart which shows the sequence of actions in Met-X-PERT is given in Fig. 1. Limitations and Assumptions The following limitations and assumptions are made in developing the rules in Met-X-PERT. These limitations and/or assumptions can easily be changed or edited, if desired. 1. Member size is changed in steps of 2 in. Truss members must have 2-in. "nominal" width. Thus, only the depth dimension changes. Posts should have more than 2-in. nominal width. 2. Stress grade of lumber is changed in steps of 10%. This process continues until the minimum or maximum stress grade of lumber, as specified by the NFPA (1986), is reached. 3. Only two gauges of metal skin (26 and 29) are considered in the program. These are the two gauges commonly used for roof cladding of post-frame buildings. (The fastening type and pattern for the roofing are defined by a prototype diaphragm in the building.) 4. Truss web members are subjected to axial forces only. ARCHITECTURE OF Met-X-PERT Met-X-PERT is structured in a modular form for ease of modifications. It consists of 12 modules grouped according to their utilities. The structure and utility of each module is described briefly. Further details of the program PROCEDURES and FUNCTIONS are given in Gebremedhin (1988b). Declaration This module contains all the declarations for the entire program. Some of these pertain to the list of members in the structure and their properties, and others to the menu. The declarations are used to interpret user input and to manage objects and values. Knowledge Base Utilities This section contains the knowledge base manipulation utilities necessary to develop and maintain a linked object list. An expert system must have a mechanism for inserting facts and rules into the knowledge base, maintaining the list of expressions in the KB, and displaying the contents of the KB for review by the user of the system. The knowledge base in Met-X- PERT consists of 32 rules. The system contains the knowledge necessary to optimize the size of post and truss members, the horizontal deflection at the eave, and shear of the metal skin. 448 APPLIED ENGINEERING in AGRICULTURE Module for Handling Variable Certainty The routines in this module allow the expert system to handle variable certainty. The program reads a certainty factor (cf) integer in an object-value couplet, or alternatively, translates subjective input into quantitative certainty factors on a scale of 0 to 100. The confidence factor or certainty factor is used to reflect the degree of certainty with which a fact is believed. This information is recorded in the fact base along with the object-value couplet describing that fact. If there are several rules which have different condition-element (or antecedents) but similar conclusions, and more than one gets triggered, then the degree of certainty of the conclusion will be greater than any one of them individually. In such cases, the cf s need to be accounted for in a combined way. For this purpose, the concept of BLENDING FUNCTION is used. The blending function is basically an arithmetric expression which combines two certainty factors (cfl and cf2) and normalizes the result, so that the result lies between 0 to 100. The particular blending function that was used in this program uses a simple smoothing technique to blend the cf s as (Alty and Coombs, 1984; Negoita, 1984): 100 * cfl + 100 * cf2 - cfl *cf2 where c f l a n d c O = certainty factors, percentages. Additional modules are used to add the cf integer to the linked object list and permit existing cf integers to be updated. Module for Multivalued Expressions The routines in this module are used to ascertain whether or not an object in the linked list is multivalued and to define a particular object name as being multivalued so that more than one value can be assigned to the given object name. For example, advice is a multivalued object. Module for Questions and Legal Values The routines in this section enable the expert system to interpret operator input containing an object name and a list of legal values to store the value names in the linked object list. The program is also equipped with a facility for adding questions about specified objects and displaying existing questions along with the associated legal values. This approach is very helpful for interactive consultation type expert system applications. Rules Management Module In this section, there are routines which enable the expert system to read, save, and retrieve rules of the form: Rule IF metal skin = gauge 29 AND shear force of metal skin—less than the allowable AND CSI of post= greater than 1.0 THEN recommended action = increase gauge of metal skin to 26 There are also PROCEDURES to perform simple house- keeping chores, add a premise to the current rule, add a conclusion to the current rule, and display an existing rule on request. Explanation Module The system guides and assists the user in the design decision-making process by providing explanations. The PROCEDURES in this section allow the expert system to explain its line of reasoning, how a specific conclusion was reached, and why a certain question must be asked to pursue the stated goal. The user can review the rules of the knowledge base and trace the line of reasoning. This procedure not only demonstrates the program's approach to solving a problem but also permits the operator to monitor the system as it processes the user- constructed rules. Module for the Inference Engine The inference engine applies its repertoire of rules in an attempt to assign a value to the named object-goa/. The program continues its pursuit until one of the candidate solutions succeeds. A backward chaining logic was followed. In this, the engine begins with a terminal conclusion, the object-goal. The program then searches through the rule list for any rule that has the object-goa/ as its conclusion. The engine then tests each part of the rule's premise and the process is begun again for the premises. At the heart of the program's inference engine is the PROCEDURE pursue. It acts on an object-name, trying to apply the rules to give the object a value. If no rules conclude anything about the object, the user is asked to enter value of that object. The PROCEDURE uses a rule pointer to point to the current rule under consideration. The program next sets out to find a rule that has the object name in its conclusion. It then tests each part of the rule's premise. If all the premises are satisfied, the conclusion is added to the knowledge base. If the object's value cannot be found from the rules, the program prompts the user to enter a value name. Other PROCEDURES exist which implements a conclusion when the inference engine finds a link between the premise of one rule and the conclusion of another. The procedure ascertains the cf of the conclusion based on the relative cf integers of the contributing rules and prints its conclusion on the screen. Interface Module This module interfaces the METCLAD output with Met-X-PERT. Information extracted from METCLAD output include: (a) combined stress interaction indices of post and truss members, (b) horizontal shear stresses of post and truss members, (c) shear strength of the metal skin, and (d) horizontal deflection at eave height. This information is stored and the limiting factors for design are analyzed. Routines exist for initializing the variables from the METCLAD output, and converting a character string representing a number to its actual numeric value. Analyzer Module This module analyzes the limiting factors for design based on the rules specified. Vol. 5(3):September 1989 449 Output Interface Module This module creates a new file called " e x p e r t . M O " depending on the advice and the initial values of the problem specifications. This file can be used directly as input to the METCLAD program. Main Program This program integrates together the different modules in the system to perform their functions. OPERATIONAL PROCEDURE This section describes the software requirements and the procedures required to operate the system. It is not meant to be an operator's manual for the program developed but to show the sequence of actions that need to be taken and to also show the information displayed at each step. Software Requirements The program diskette contains the following files: 1. MDMS.COM - preprocessor to create a data file. MDMS is an acronym for METCLAD Data Management System; 2. fn.MCI - data file name created with MDMS. This file could have any name acceptable in DOS but should have MCI for its extension; 3. METCLAD.EXE - analyzes the data file created by MDMS; 4. Met-X-PERT.EXE - optimizes METCLAD output; 5. RULES - contains the rules in the knowledge base. Procedure To operate the program, the following steps are followed (the boxed statements are what you see on the computer's monitor): 1. Create a data file (e.g., fn.MCI) using the preprocessor, MDMS. Further details about this system is given in Gebremedhin and Bartlett (1987). When creating a new file, it is advisable to have a back-up file with a different extension. This is because only the file with an MCI extension is executed by the expert system, and this file is constantly modified based on the advice invoked. The back-up file with the different extension can not be executed by the expert system. 2. Type Met-X-PERT <return> and the screen displays Welcome to the Met-X-PERT Press ENTER to continue. 3. Press <return>, and the screen displays --> Enter file name without extension: 4. After the file name is entered, the screen displays 11 - > Enter connectivity of roof (1 or 2) || 5. Connectivity of roof refers to the condition of continuity of the metal skin at the ridge of gable roofs. Enter 1 if the metal skin at the ridge of a sloped roof is assumed continuous for shear transfer, otherwise enter 2, and press <return>. The screen displays Executing METCLAD.... Iteration no. = 1 I fn 1 I When execution of the data file is completed, the program creates an output file called fn.MCO. The following sequence of execution are performed and the messages are displayed on the screen: - - > METCLAD execution completed. - - > Reading the file containing rules... ||- - > Reading the METCLAD output.. II When the above actions are completed, the screen displays a list of the potential limiting factors pertaining to the problem being analyzed as - - > Analyzing METCLAD output... - Gauge of metal skin: 29 - Shear force of skin = 3.1710E+03 - Allowable shear force of skin = 4.3280E+03 - Critical Post is: 2-4 - CSI of critical post = 1.08 - Horiz. shear stress of post = 6.341 E+01 - Allowable horiz. shear of post = 1.137E+02 - Critical truss chord is: 6-8 - CSI of critical chord = 0.97 - Critical web is: 8-10 - CSI of critical web = 8.0E-02 || - Horiz. eave displacement = 1.105E+00 \\ 6. Press <return> and the screen displays Do you want the reasoning explained during execution?»(y/n) 450 APPLIED ENGINEERING in AGRICULTURE http://MDMS.COM --> Results of consultation \ Because: shear force of metal skin = less than allowable We conclude that: Advice = gauge of skin is okay, cf 100 I Because: \ CSI of critical post =1.08 t We conclude that: 1 Advice = increase bending stress by 10% i ( ****End of consultation**** (advice are listed here) 7. Enter y or n depending upon your choice. If y is entered, the screen displays the reasoning for the suggested changes to be made on the fn.MCI file. Next, the screen displays (also when n is entered) The program then makes the changes based on the advice and fn.MCI is modified. Next, the screen displays Creating a new input file for METCLAD., The above procedure creates a new file called fn.ESO. Now, the program compares fn.MCI and the newly created file fn.ESO. If these two files are the same, the program stops. If not, the fn.MCI and fn.MCO files are erased from the disk and the fn.ESO file is renamed internally to fn.MCI. Next, METCLAD is again executed using this new file and the screen displays Executing METCLAD. Iteration no. = 2 and steps 6 and 7 are repeated. A maximum of 50 iterations can be made. After a successful execution, the following files are stored in the disk: fn.MCI fn.ESO fn.MCO The first two files should be the same (the reason the program stopped iteration), and the third file is the last output of METCLAD. The order of execution of the files is shown in Fig. 2. When 50 iterations are completed and if fn.MCI is not yet the same as fn.ESO, fn.MCI is stored in the disk and fn.ESO and fn.MCO are erased. The fn.MCI file is the last file to be updated during the iterative process. SUMMARY AND CONCLUSIONS A procedure was developed to optimize designs of metal-clad post-frame rectangular buildings using a knowledge based expert system. The primary objective of the study was to develop a procedure to extend the solution obtained from a conventional algorithmic CAD program to be used by the knowledge base of an expert system. Solving the problem of incompatibility between CAD and expert system programs allows information to be shared between them. (^Start^) Roof Continuity fn. MCI Rename fn. ESO to fn. MCI fn. MCI Execute METCLAD fn. ESO Erase fn. MCI Erase fn. MCO fn. MCO Run Expert System > Maximum 5 0 I t e r a t i o n s yes Stop Fig. 2—Sequence of execution. Vol. 5(3):September 1989 451 Met-X-PERT takes the output of the frame analysis program, METCLAD, and uses it as a fact base. The system analyzes the output based on the rules specified. Changes are triggered by the inference engine to modify the input file of METCLAD. METCLAD is then executed again using the modified input file. This sequence of operations continues until all the specified limiting factors for design are met and an " optimum'' design is simulated. The advice and actions taken during the process are displayed on the screen. The program is not intended for practical application for every design is unique and it is difficult to incorporate all requirements, assumptions, and judgements associated with design. References 1. Adeli, H. 1986. Artificial intelligence in structural engineering. Engineering Analysis 3(1): 154-160. 2. Adeli, H. and J. Fiedorek. 1986. Microcomputer-aided design and drafting of moment-resisting connections in steel buildings. Microcomputers in Civil Engineering l(l):32-44. 3. Adeli, H. and M.M. Al-Rijleh. 1987. Computer-aided design of trusses using Turbo Pascal. Microcomputers in Civil Engineering 2(2):101-116. 4. Alty, J.L. and M.J. Coombs. 1984. Expert Systems Concepts and Examples. Manchester, England: NCC Publ. 5. Gebremedhin, K.G. 1986. SOLVER - An interactive structures analyzer for microcomputers. Transactions of the ASAE 29(1): 10-13. 6. Gebremedhin, K.G. and J.V. Bartlett. 1987. A data preprocessor for structural analysis and design programs. Journal of Microcomputers in Civil Engineering 2(2): 135-145. 7. Gebremedhin, K.G. 1988a. Computerized system for diaphragm design of metal-clad post frame buildings, Transactions of the ASAE 31(l):215-220. 8. Gebremedhin, K.G. 1988b. Expert system development to optimize design of metal-clad post-frame buildings. In Proceedings 6th National Conference on Microcomputers in Civil Engineers, ed. W.E. Carroll, 132-138. Nov. 9-11, Orlando, FL. 9. Hayes-Roth, F., D.A. Waterman and D.B. Lenat. 1983. Building Expert Systems. Reading, MA: Addison-Wesley Publishing Co. 10. National Forest Products Association. 1986. National design specification for wood construction. National Forest Products Association, Washington, D.C. 11. Negoita, C.V. 1984. Expert Systems and Fuzzy Systems. Menlo-Park, CA: Benjamin-Cummings Publ. Co. Inc. 12. Suddath, S.K. and R.W. Wolfe. 1984. Purdue Plane Structures Analyzer II (PPSA II): A computerized wood engineering system. USDA Forest Products Laboratory, Madison, WI. 13. Waterman, D.A. 1986. A Guide to Expert Systems. Reading, MS: Addison-Wesley Publishing Co. 452 APPLIED ENGINEERING in AGRICULTURE 
work_wwko6iojtngkzdfsmilbi6w4di	----	LSE Research Online Article (refereed) Angeliki Poulymenakou, Tony Cornford and Edgar A. Whitley Knowledge acquisition for organisational problem solving : developing expert systems and beyond Originally published in Expert systems with applications, 5 (1-2). pp. 121-130 © 1992 Elsevier. You may cite this version as: Poulymenakou, Angeliki; Cornford, Tony & Whitley, Edgar A. (1992). Knowledge acquisition for organisational problem solving : developing expert systems and beyond [online]. London: LSE Research Online. Available at: http://eprints.lse.ac.uk/archive/00000275 Available online: June 2005 LSE has developed LSE Research Online so that users may access research output of the School. Copyright © and Moral Rights for the papers on this site are retained by the individual authors and/or other copyright owners. Users may download and/or print one copy of any article(s) in LSE Research Online to facilitate their private study or for non-commercial research. You may not engage in further distribution of the material or use it for any profit-making activities or any commercial gain. You may freely distribute the URL (http://eprints.lse.ac.uk) of the LSE Research Online website. This document is the author’s final manuscript version of the journal article, incorporating any revisions agreed during the peer review process. Some differences between this version and the publisher’s version remain. You are advised to consult the publisher’s version if you wish to cite from it. http://eprints.lse.ac.uk Contact LSE Research Online at: Library.Researchonline@lse.ac.uk http://eprints.lse.ac.uk/archive/00000275 http://eprints.lse.ac.uk/ http://eprints.lse.ac.uk/ mailto:Library.Researchonline@lse.ac.uk http://www.lse.ac.uk/people/e.a.whitley@lse.ac.uk/ http://www.elsevier.com/locate/eswa Knowledge acquisition to facilitate organisational problem solving Angcliki Poulymcnakou, Tony Comford, Edgar A. Whitlcy Information Systems Department London School of Economics and Political Scicncc Houghton Street London WC2A 2AE Abstract This paper examines the introduction of knowledge acquisition techniques to facilitate managerial and administrative problem solving. The characteristics of such problem situations do not lend themselves to normative approaches such as simply introducing expert systems. However appropriately performed knowledge acquisition can still be of benefit by providing a better understanding of problem situations and it can provide a suitable basis for the process of negotiation and compromise needed to resolve organisational problems. This emerging role for knowledge acquisition differs from that of systems analysis, for example, in that it focuses on improving the understanding of the location, ownership and impact of available organisational knowledge rather than the data flows and tasks performed within the organ&&n. The paper provides guidelines for performing knowledge acquisition for organisational problem solving. Introduction Organisations today arc facing increasingly demanding and rapidly changing environments in which they are asked to operate. serve and perform. People attribute this to a multitude of factors spanning from global competition and the struggle for innovation (Porter 1990) to the emergence of the information society (Strassman 1985). By reviewing the characteristics and requirements for organisational problem solving in this context, we will argue that these activities go far beyond the scope of support offered by expert systems today. Knowledge acquisition is the most critical activity in the expert systems life cycle and it is there that efforts to build systems that will make a bigger impact in organisational performance should be placed. Poulymenakou et al. (1990) in another paper at this conference, specifically deal with the requirements placed upon knowledge acquisition in the cast of developing business expert systems and put forward some suggestions for more effective practices in this context. This approach, however, is still within the framework of developing a system designated to support managerial and administrative problem solving. Even a business expert system developed following open and flexible knowledge acquisition approaches remains a hard and formal structure. Embedding this in an organisational and/or managerial environment implies overlooking the soft and informal nature of a large proportion of the activities carried out in these contexts. This contrast remains despite efforts for more effective knowledge acquisition practices for business expert systems. In many casts of organisational decision making, the introduction of any system hard-wired to the processes is unrealistic. Having said this, any support to improve performance that is offered by old and new technologies alike is welcomed. provided that it does not clash with existing managerial decision making practices. Permission to copy without fee all or part of this material is granted provided that the copies are not made or distributed for direct commercial advantage, the ACM copyright notice and the title of the publication and its date appear, and notice is given that copying is by permission of the Association for Computing Machinety. To copy otherwise, or to republish, requires a fee and/or specific permission. Q 1990 ACM 089791-416-3 /90/‘0010/0181 $1.50 181 In this paper, we will argue that the option offered by techniques developed and used in knowledge acquisition can provide some of this much sought after support if they are taken out of the context of expert systems development. To support this argument, we will review some aspects of organisational problem solving and will suggest ways in which these could be served by knowledge acquisition. We will then describe the type of services that can be provided. Finally, we will discuss how such activities can be set up and carried out by organisations. Characteristics of managerial and administrative problem solving Organ&ions are increasingly relying for their performance upon their collective knowledge and recognise the activities of their workers as being knowledge intensive tasks. ‘Knowledge’ in an organisation is the collection of experience, skills, talents and information that each individual has and utilises when he or she is at work. Important knowledge also exists at the level of the group, function or division and does not reside in any one individual. The advances of the information systems and communications technology have drawn significant attention to the impact of storing and disseminating information in organisations as well as supporting the knowledge processes of the people in them (Strassman 1985). Many believe that these are the first signs of the emergence of the ‘knowledge organisation’. Solving problems in organisations is a complex, knowledge based activity. It entails taking decisions that vary enormously in nature. It goes through various phases from setting the initial requirements to reaching the final objectives- It is influenced by a multitude of factors and it involves a variety of actors. All these issues need to be considered before any realistic effort to facilitate organisational problem solving can be initiated. Brenda Wroe (1987) differentiates three types of managerial decisions: operational, tactical and strategic. Operational decisions relate to everyday, narrowly focused problems in an organisation relating to the type of question: how much would this product cost? The data and resources that would be required in order to deal with such issues am more or less standard within an organ&ion. The process followed for problem solving is also basic. The overall objective of operational decisions is the implcmcntation of a selected course of action. Tactical decisions are those which deal with selection. They are the frequently required decisions of the type: how many units of this product should be produced? The scope of tactical decisions is wider and the background and processes followed in problem solving are Iess apparent, standard or readily available. Strategic decisions are characterised by their exceptionally wide scope and their relative infrequency. They are mostly one-off, high level decisions aimed at providing actors with directions, for example: what types of products should we produce? Strategic decisions are based on a broad platform of evidence, the nature and amount of which change according to the situation at hand. The process of decision making is subjective and is influenced by the perceptions of the few individuals of the issues considered. From the differentiation of types of decisions, it is clear that the nature of managerial decision making changes according to the requirements of the problem that it is set to resolve. Different problems trigger different types of decisions and hence different methods must be used for their investigation and support. As we move from operational to strategic decision making, the data and standard procedures give way lo more complex considerations of 182 alternative contingencies and associated risks. The methods that can be used to investigate and support different types of decisions should vary accordingly. Decision making in organisations has been extensively discussed from many different perspectives. For Simon (1976), organisations are just ordered, social ways for making decisions. He describes three stages in the overall process of making a decision. These can be associated with stages in problem solving. Table 1: Stages of decision making and problem solving Decision making Intelligence Problem solving -Finding occasions calling for a decision Design -Identifying the problem Grventing, developing and analyzing possible courses of action Choice *Formulating possible solutions *Selecting a particular course of action from those available GS3-mining the correct solution If this perspective is adopted, decision making and problem solving can be viewed interchangeably. We make decisions by solving problems that arise in the process of meeting our objectives. In the complex environment of organisational problem solving, groups and their activities have a special role. The benefits of sharing and exchanging information, ideas and responsibility make groups very powerful decision making agents. Groups, teams and committees serve a variety of purposes in organisations as presented in Table 2. Table 2: Working in groups (Handy 1985) Purpose aDistribution of work Mode *Bringing together skills, talents and responsibilities aManagement and control of work *Allocate work to the appropriate individuals -Taking decisions *Applying all available capacities to the process by combining skills, talents and responsibilities -Data and information collection *Gather ideas, information or suggestions l Information distribution *Passing on information and decisions to those who need to know *Testing and authorising decisions *Testing and authorising actions taken outside the group l ordination and liaison -Coordinate problems and actions between functions and divisions #Negotiation and conflict *Alleviation of disputes and arguments between resolution levels, functions and divisions *Inquest into the past l Examination of past cases or organisational performance 183 The function of group decision making does not come free of problems. The process is complex and difficult to monitor, support or control. The benefits in devising ways of doing so focus on avoiding group behaviour producing the type of output described by Handy as “A camel is a horse put together by a committee”. In this section, we have attempted to present some of the characteristics and phenomena associated with problem solving in organisations. As we have argued in the beginning, problem solving in its various forms is based on the knowledge of every individual who take part in it. The process of acquiring knowledge developed in the field of expert systems has provided analysts with a variety of methods, techniques and tools that could be used in the investigation of these knowledge processes. Knowledge acquisition performed in the context of dcvcloping an expert system however, will sooner or later resort to the reductionist, rationalist approach required in order to transform knowledge into an operational structure (Winograd and Flares 1986). The collective knowledge and experience underlying the processes described here cannot be accommodated in such a structure. One could consider for example, the difficulty in representing in a system the implications of obtaining feedback on a report or au action. An effort to do so requires the incorporation of cultural, contextual, situational and organisational evidence. Therefore, before knowledge acquisition assumes a facilitator’s role in organisational problem solving, we need to set the requirements that knowledge acquisition will seek to fulfil in this context. The requirements for organisational problem solving The increasing complexity of the nature and context of organisational problem solving urges organisations to consider new enabling mechanisms to support their efforts. The objectives that underlie this quest are the development of more effective practices, the improvement in understanding of the nature of managerial tasks and support in the development of strategy in organisations. Techniques that are structured, narrowly focused and thinly spread across a managerial domain would not work for these types of activities. To achieve these objectives we need techniques that are able to adapt to the requirements of the environment of investigation as different organisations develop different cultures and styles of performance. Also, the techniques used should allow the elicitation and examination of a large variety of factors that may affect the way people deal with problems in an organisational context. Finally, the techniques used to investigate organisational problem solving should direct the analysis towards a deep search behind occurrences and phenomena towards causes, effects and impacts of actions (Suchman 1987). Sir Geoffrey Vickers (1973) has summarised these arguments by saying: “I hope not for greater efficiency in our problem-solving, but for better understanding in our problem-setting”. Managers are basing their decision making performance on a complex web of acquired abilities: perceiving and recognising patterns, dealing with abstraction, coping with uncertainty, adapting to change. learning from experience, assessing situations and most of all dealing with people. Knowledge acquisition should start off by examining the bows. whats and whys of this performance by putting it in the context of organisational problem solving on the one side and that of available methods and tools for investigating knowledge processes on the other. We now present an agenda of requirements which wilt suitable for this area. 184 Problem representation: Clarifying what is lhe problem is a difficulty people face up front in organisations. The objective is to make problems more explicit by eliciting and creating representations of them, Hence, information related to a problem situation is made commonly available, the views of multiple sources can be incorporated in an explicit manner and progress in dealing with the problem can be monitored. Decision taping methods such as the ones discussed in Poulymenakou et al. (1990) can be used here. Once the problem is resolved, these representations can also serve as stores of knowledge and experience documenting past problem solving cases for future reference. Enablers for problem visualisation: Solutions to problems are only as good as the knowledge of those who provide them. Knowledge in organisations is distributed among many agents and on many occasions it is not readily available where and when a problem occurs. Problem solving is often a tune critical activity and managers do not have the time to go around and interview people about their experiences as the need occurs. Knowledge acquisition can provide means for incorporating knowledge available in different parts of organisations thus providing the managers with different perceptions of the issues they are considering in every occasion. Agenda and basis for negotiation: The perceptions of a problem by individuals in group decision making are seldom uniform in depth and breadth. Apart from differences in knowledge and experience, this is also due to the different needs and requirements people have when they deal with a problem. Suppressing or abolishing the arising conflict simpfy postpones the resolution of the teal problem and provides less than optimal interim solutions. Knowledge acquisition can help here by making the differences clear and explicit and by tracing some of the reasons behind them. One can suggest here the use of many knowledge acquisition tools currently available. We will discuss two candidate tools later in this paper. We can also use diagrammatic investigation methods like repertory grids, not for encoding knowledge, but for introducing different perceptions of it. Problem solving support agents: Knowledge acquisition has been traditionally involved in this area by supporting the process of building expert systems. However, here we adopt a different view of the process to achieve the objective. Instead of knowledge acquisition ‘to build a system to solve the problem’, we propose knowledge acquisition to support the people who deal with the problem. This can be achieved by helping them refer back to their own experiences through critical incident interviews, holding investigative discussions or compiling and using an inventory of past cases. Knowledge acquisition techniques to facilitate organisational problem solving The supply of the type of support to organisational problem solving that has been outlined so far predicates a new perception of the utilities offered by existing knowledge acquisition techniques. We present here some techniques and tools that look promising for investigations of this nature. As an overall starting point and guide we use the methods that have been developed for tapping decision making knowledge. This seems appropriate since the considerations of these methods cover at least in part, the fiist, second and fourth points on our requirements agenda. 185 An investigation on decision making practices starts off by looking at the environment in which decisions are taken. This can be done by observing decision makers in the course of a working day taking notes on major activities, interacting with other people, consulting documents, recording the constraints encountered and the time taken to reach a decision. This type of observation is good as a familiarisation exercise with the organisational environment but it leaves the analyst in the dark about the factors, alternatives and processes in decision making. After the initial observations, analysis can turn to the consideration of the important features of decisions. A possible way for doing so is to discuss with the decision maker the scenario of a case where the information available, or the nature of the problem or the features of the solution are restricted. Then, as the interviewee is trying to deal with the case, priorities and alternatives considered can be isolated. Similarly, dam from a problem already tackled can be used to prompt discussion on what makes a ‘common’ or ‘unusual’ case, what decision making strategies would fit each of them. These simulated scenarios could t&e the form of a critical incident interview whereby revisiting difficult or exceptional cases can help to reveal effective courses of action and some of the reasons behind them. Managers develop skills like that of perceiving and recognising structures specifically because these can help them establish similarities between current problems and previous ones and thus decide where analogous action would be beneficial. Episodic analogies can help reveal the ‘benchmarks’ managers are using when comparing two problem situations. These indicators, apart from pointing to specific courses of actions can also double as problem representation primitives. Table 3: Methods for investif Technique *Environment observation *Constrained scenarios 4imulated scenarios *Episodic analogies *Difficult cases ing decision making processes (Hoffman 1987) Benefits *Overall view of tasks and area orientation *Reveal salient features of a problem and strategies for problem solving *‘Post mortems’ in actions, archiving data -Pictorial view of past experiences. sensitivity analysis in problem features *Infrequently used data, data combination and prioritisation The techniques described so far work well on defining and refining various aspects of decision making of individuals. In the context of problem solving in organisations though, we have seen that groups have an important role to play. Group decisions are either the rest& of a necessity, for example if a case is too difficult for an individual to tackle, or of convention if a decision is regarded as being too critical to be left in the hands of a single person which is often the case with strategic decisions. Knowledge acquisition can help group decision making in two ways. One is to create a common basis of understanding for the group members. The other is to support the exploitation of the combination of human resources that exist in a group. For the first objective, the most relevant technique involves the comparison of conceptual structures of different individuals. Shaw and Gaines (1989), who have worked on this, have described four possibilities in the use of concepts and terms as referred to by different people. 186 People may use the same term to describe different concepts, different terms to describe the same concept, the same terms for the same concept or different terms for different concepts. Eliciting and recognising which of the four cases is true in a particular situation will form a basis for resolving potential conflicts that refer to mere terminologicaI disagreements. One step further, the technique can be used to create a rich framework for examining a particular problem situation. Brainstorming is proposed here as a suitable investigation technique. The second objective of knowledge acquisition in group decision making picks up the situation where the fast leaves it. After the comparison of conceptual structures, other techniques can be used to follow up brain storming. Consensus decision making, for example, shifts the focus from the quantity of answers collected to their relative quality for the problem at hand (McGraw and Harbisson-Briggs 1989). All alternatives are judged in terms of their advantages and disadvantages, weighted and measured by every person in the group. Finally, the most sign&ant contribution of knowledge acquisition in organisational problem solving could prove to be that of using the multi-technique tools that have been developed for the support of the knowledge acquisition process. These tools incorporate diagrammatic techniques that can prove valuable for the representation and visualisation of problems as well as for the investigation of different scenarios. One of the most sophisticated tools available is AQUINAS (Boose ef al. 1989). the newest product of long research in the area. The tool uses repertory grids, which are tables of problem characteristics mapped against solutions, used to represent what people know about problems. The characteristics used can be observations, preferences or constraints. Repertory grids can be used for decomposing problems, eliciting distinctions and combining uncertain information. Multiple sources can work together on the same problem using this tool. They can select what they think as an appropriate course of action such as combining the evidence, considering the constraints or rating their own preferences. The developers point to a number of ways in which such a tool could be used, apart of the obvious of building an expert system. They include the facility of having a source of knowledge different than its user, the inclusion of multiple contributions to what is known about a problem, the ability to serve many ‘knowledge users’, the ease of updating the stored material and the use of the tool for archiving. In this section, we have pointed to a number of alternatives available through advances in knowledge acquisition, that can serve the cause of organisational problem solving by bringing to the process some of the insight gained by advances in the knowledge acquisition field. We argue that it could only be to the benefit of the otganisations to consider and try the options we have presented here. Epilogue The objectives of this paper has been to introduce a different perception in terms of the approaches used to investigate organisational problems and to provide support for dealing with them. We propose the dissemination of practices relating to the new technology of knowledge based systems, in particular knowledge acquisition, to harness the underlying knowledge resources found in organisations, without the restrictive, hard-wired format found in expert systems. Such an endeavour would not provide simple and easy answers but we believe it is a worthwhile research ambition. 187 REFERENCES Boose, J.H., Shema, D.B., Bradsi, J.M. “Recent Progress in ACQUINAS: A Knowledge Acquisition Workbench” Knowledge Acquisition (1:2), June 1989, pp.185214. Handy, C.B. “Understanding Organisations” Middlesex, England: Penguin Books, 1985. Hoffman, R. “The Problem of Extracting the Knowledge of Experts from the Perspective of the Experimental Psychology” The AI Magazine, (8:2), 1987. pp.53-76. McGraw, K.L., Harbisson-Briggs, K. “Knowledge Acquisition: Principles and Guidelines” NJ: Prentice-Hall, 1989. Porter, M. “The Competitive Advantage of Nations” New York: The Free Press, 1990. Poulymenakou. A., Cornford, T. and Whitley E. “Pragmatic considerations for effective knowledge acquisition: The case of business expert systems”. Proceedings of SIGBDP conference Trends and Directions in Expert Systems, Florida, 1990. Saw, M.L., Gaines, B-R. “Comparing Conceptual Structures: Consensus, Conflict, Correspondence and Contrast” Knowledge Acquisition (1:4), December 1989, pp.341- 364. Simon, H.A. “Administrative Behaviour” Macmillan, 1976. Strassman, P.A. “The Information Payoff: The Transfer of Work in the Electronic Age” New York The Free Press, 1985. Suchman, L.A. “Plans and Situated Actions: The Problem of Human-Machine Communication”, Cambridge: Cambridge University Press, 1987. Vickers, Sir G. “Making Institutions Work” Associated Business Press, 1973. Winograd, T.,Flores, F, “Understanding Computers and Cognition: A New Foundation for Design” Reading, MA: Addison-Wesley, 1986. Wroe, B. “Successful Computing in a Small Business” London: NCC Publications, 1987. 188 
work_wwwqugbjtrfxjenbid6xk6l6cq	----	GRACO - Página inicial Portal do Governo Brasileiro Atualize sua Barra de Governo Menu Sobre a UnB Unidades acadêmicas Estude na UnB Graduação Pós-Graduação Administração Servidor           GRACO - Grupo de Automação e Controle A- A A+ MENU Página inicial Sobre Histórico Linhas de Pesquisa Laboratórios Processos de Soldagem Robótica e Visão Computacional Automação Industrial Sistemas de Controle Sistemas Embarcados Materiais Avançados Sistemas Hidráulicos e Pneumáticos Protótipos Mecatrônicos Infraestrutura Membros Docentes Corpo Técnico Discentes Contato Robótica Soldagem Robotizada Cordões de solda são executados com precisão por robôs manipuladores. Robótica Robótica Móvel Robôs móveis inteligentes podem se locomover em ambientes desconhecidos. Instrumentação Sensores Inteligentes Sensores inteligentes são utilizados para detecção de defeitos em processos de fabricação. Automação Automação Predial Sistemas de automação avançados podem se adaptar às necessidades dos usuários. Robótica Reparos de Turbinas Robôs customizados para aplicações específicas entregam melhores resultados. Institucional História da UnB UnB em Números Conheça os campi Como chegar Estatuto e Regimento Administrativo Reitoria Vice-Reitoria Conselhos e câmaras Decanatos Secretarias Prefeitura da UnB Acadêmico Faculdades Institutos Centros Educação a Distância Assuntos Internacionais Serviços Arquivo Central Biblioteca Central Editora UnB Fazenda Água Limpa Hospital Universitário Hospitais Veterinários Restaurante Universitário Comunicação Atendimento a Jornalistas Fale com a Secom Canais Oficiais Marca UnB Campanha Institucional 2021 UnB TV   Campus Universitário Darcy Ribeiro, Brasília-DF | CEP 70910-900 Telefones UnB Copyright © 2021 Universidade de Brasília. Todos os direitos reservados. Melhor visualizado nos navegadores Google Chrome e Mozilla Firefox Acesso à Informação OUVIDORIAUnB 
work_wxey53wxabaypntkz5udnu2ohu	----	Medicine expert system dynamic Bayesian Network and ontology based Medicine expert system dynamic Bayesian Network and ontology based Octavian Arsene ⇑, Ioan Dumitrache, Ioana Mihu Laboratory of Intelligent Systems, ‘‘Politehnica’’ University of Bucharest, SPl. Independentei 313, 060042 Bucharest, Romania a r t i c l e i n f o Keywords: Expert system Ontology Bayesian Network a b s t r a c t The paper proposes an application framework to be used for medicine assisted diagnosis based on ontol- ogy and Bayesian Network (DBNO). There are two goals: (1) to separate the domain knowledge from the probabilistic information and (2) to create an intuitive user interface. The framework architecture has three layers: knowledge, uncertainty model and user interface. The contributions of the domain experts are decoupled, the ontology builder will create the domain concepts and relationships focusing on the domain knowledge only. The uncertainty model is Bayesian Network and the probabilities of the vari- ables states are stored in a profile repository. The diagnostician will use the user interface feeded with the domain ontology and one uncertainty profile. The application was tested on a sample medicine model for the diagnose of heart disease. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction The diagnosis can be defined as the process of identifying a set of hypotheses that model the problem domain and finding that one with highest probability of matching the real world state. In med- ical diagnosis, the uncertainty arises from the inability to evaluate the degree of truth of a hypothesis due to unreliable and incom- plete information or inconsistent knowledge. The ontology and Bayesian Network (BN) methodologies have been chosen to address knowledge management and uncertainty. The ontology enables the representation of a domain knowledge in a machine understandable form. It can represent the organiza- tional structure of a large complex domains, but the inability to deal with the uncertainty can be a drawback for its application. One disadvantage of the BN is representation of complex struc- tured domains point of view. The ontology and BN can comple- ment themselves in order to overcome the each other disadvantages, thus an ontology-driven uncertainty model can be created. The main goal of the paper is to propose an application frame- work as a collaborative expert system for creating, developing and maintaining a general model for medicine assisted diagnosis (Fig. 1). From user point of view there will be three roles based on their competencies: concepts and relations definition, connect probabilities to states of the concepts, setting evidences in order to assist the diagnosis. Each role is assigned to one of the triangle’ sides and thus depicting three operational layers. The automated connection between all three layers increases the efficiency of the entire process. The proposed model is implemented as a soft- ware using PROTEGE as an ontology framework, NETICA API (Application Programming Interface) as a Bayesian API and Java technology as a development platform. The mapping between domain knowledge and uncertainty model is based on the fact that each concept defined into ontology is part of the BN as a variable. The diagnostician will use an intuitive graphical user interface for changing evidences of the BN variables and based on a thresh- old the application will depict a chart having the most significant nodes. The final chart will assist the medicine diagnostician to identify the significant factors for a particular case. In order to offer more flexibility the domain knowledge and probabilities are already build and ready to be used as a medicine ontology and respectively uncertainty profile. There is no need to be re-created each time during diagnosis phase. The diagnostician can choose between more than one uncertainty profiles for a medicine ontol- ogy. The relation between ontology and uncertainty profile is one to many (in case there are several sources for probabilities tables for the same ontology). This approach speed up the entire diagno- sis process and allows the asynchronous update of the ontology and uncertainty profiles by the domain experts in order to increase the degree of accuracy of the information. A similar model, OntoBayes, was proposed in Yi (2007). The major difference between OntoBayes and this proposed model resides in separation between domain knowledge and quantitative component of BN in order to decouple the ontology from the uncertainty probabilities. BayesOWL (Zhongli, 2005) and PR-OWL (Cesar, Costa, Laskey, & Laskey, 2003) are others probabilistic ontology approaches facilitating ontology mapping in the semantic web. Some of their limitations refer to: two-valued variables only 0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.05.074 ⇑ Corresponding author. E-mail address: octavianarsene@gmail.com (O. Arsene). Expert Systems with Applications 38 (2011) 15253–15261 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2011.05.074 mailto:octavianarsene@gmail.com http://dx.doi.org/10.1016/j.eswa.2011.05.074 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa allowed in BayesOWL, the ontology model is too complex in PR-OWL. An automated BN construction based on an ontology model was proposed in Devitt, Danev, and Matusikova (2006) to be applied in the telecommunication network management domain. It’s more an algorithm model than an application. Similar approach was proposed in Mihu and Arsene (2009) for monitoring and diagnosis of an IT infrastructure, all layers (gathering evi- dences, ontology model and presentation layer) were implemented as software agents. Section 2 briefly introduces the ontology concept, the knowl- edge representation and the implementation of the medicine ontology using PROTEGE application (Horridge, Knublauch, Rector, Stevens, & Wroe, 2004). Section 3 describes Bayesian Network concept, the uncertainty representation and the implementation module within the proposed software application. Section 4 covers the description of proposed software application. Section 5 pre- sents the results of the tests using heart disease data presented in Ghosh and Valtorta (1999) and Ghosh and Valtorta (2000). 2. Knowledge representation: ontology The ontology may take a variety of forms, but necessarily it will include a vocabulary of terms and some specification of their mean- ing. The ontology is virtually always the manifestation of a shared understanding of a domain that is agreed between a numbers of parties. Such agreement facilitates accurate and effective commu- nication of meaning, which in turn leads to other benefits such as inter-operability, reuse and sharing (Uschold, 1996). The main steps of the methodology for developing ontologies are the follow- ing: (1) identify purpose and scope, (2) building the ontology. 2.1. Purpose and scope The proposed ontology addresses the medicine domain. It will be built and used by doctors. The purpose on the proposed ontol- ogy is to be knowledge representation of the medicine domain integrated into an application. The doctors will use it in order to define an accurate diagnosis based on the cumulated experience captured into ontology and probabilities tables. 2.2. Building the ontology The first step in building ontology is the identification of the key concepts. In the Fig. 2 are depicted the key concepts defined within the medicine ontology based on the medical documents and discussions with medical experts. The class is the formal way to represent a concept, e.g. the MedicineNode is the root concept of the medicine classes. The two main classes are proposed here: Cause and Effect. The Cause subclasses are considered: Environment and Heredity. The Effect subclasses are: Disease, Symptom, Sign and Test. The second step is the identification of the relationships be- tween key concepts that were defined before. We distinguished two kinds of relations between proposed classes (see Fig. 3): � any Cause will have at least one Effect, implemented by Cause has Effect (the inverse relation is: Effect of Cause); � any Effect can generate another Effect, implemented by Effect gen- erates Effect (the inverse relation is: Effect generatedBy Effect). This seems to be odd but let’s take these examples: Disease gen- erates Symptom or Sign generates Test. In the Fig. 4 is presented the implementation of the relations between concepts in PROTEGE. Both relations are inverse type based. This kind of relation will be used in bidirectional inference necessary to graphical represen- tation of the Bayesian Network. Different ontology languages provide different facilities. The most recent development in standard ontology languages is OWL, endorsed by the World Wide Web Consortium (W3C) to promote the Semantic Web vision. The PROTEGE framework is going to be used for creating, maintaining the medicine ontology. PROTEGE is a free, open source ontology editor and knowledge-base framework. The PROTEGE-OWL editor is an extension of PROTEGE that supports the Web Ontology Language (OWL). The OWL ontology may include descriptions of classes, properties and their instances. The OWL-DL language is used due to its expressivity and computational effi- ciency. The real life entities from the medicine domain are encoded as classes instantiations within an ontology (called individuals). Each individuals has a clear identity that makes them different than others (even they have common attributes). The individuals used in the proposed application are based on the heart disease model presented in Ghosh and Valtorta (2000), they are depicted in the Fig. 5. The domain expert (a senior doctor) will create the individuals according to his expertise and experience. In the Fig. 6 HighBloodPresure individual is defined: Fig. 1. Architecture of the proposed concept. Fig. 2. Hierarchy of the medicine ontology classes defined in PROTEGE. Fig. 3. Medicine ontology – relations between classes. 15254 O. Arsene et al. / Expert Systems with Applications 38 (2011) 15253–15261 https://isiarticles.com/article/29142 
work_wy64qplxp5ft7jjtc2f36mk7au	----	Text Categorization Methods for Automatic Estimation of Verbal Intelligence F. Fernández-Mart́ıneza, K. Zablotskayab, W. Minkerb aGrupo de Tecnoloǵıa del Habla, Universidad Politécnica de Madrid, Madrid, Spain. bInstitute of Communications Engineering, University of Ulm, Germany Abstract In this paper we investigate whether conventional text categorisation methods may su!ce to infer di"erent verbal intelligence levels. This research goal relies on the hypothesis that the vocabulary that speakers make use of reflects their verbal intelligence levels. Automatic verbal intelligence estima- tion of users in a Spoken Language Dialogue System may be useful when defining an optimal dialogue strategy by improving its adaptation capabili- ties. The work is based on a corpus containing descriptions (i.e. monologues) of a short film by test persons yielding di"erent educational backgrounds and the verbal intelligence scores of the speakers. First, a one-way analysis of vari- ance was performed to compare the monologues with the film transcription and to demonstrate that there are di"erences in the vocabulary used by the test persons yielding di"erent verbal intelligence levels. Then, for the classi- fication task, the monologues were represented as feature vectors using the classical TF-IDF weighting scheme. The Naive Bayes, k-nearest neighbours and Rocchio classifiers were tested. In this paper we describe and compare these classification approaches, define the optimal classification parameters and discuss the classification results obtained. Keywords: spoken dialogue systems, Naive Bayes classification, Rocchio approach, k-nearest neighbours Email addresses: ffm@die.upm.es (F. Fernández-Mart́ınez), kseniya.zablotskaya@uni-ulm.de (K. Zablotskaya), wolfgang.minker@uni-ulm.de (W. Minker) Preprint submitted to Expert Systems with Applications December 22, 2011 *Manuscript Click here to view linked References Figure 1: Spoken Language Dialogue System 1. Introduction Next-generation spoken language dialogue systems (SLDS), developed to provide users with required information and/or to help them to accomplish certain goals, are expected to be able to deal with di!cult tasks and react to a wide range of situations and problems. They should help users to feel free and comfortable when interacting with them. Moreover, they should also be user-friendly and easy to use. Including aspects of adaptation to users into SLDS may help to increase the systems’ communicative competences and influence on their acceptability (Figure 1). Next-generation SLDS may change the level of dialogue depending on users’ experience. For example, a spoken dialogue system aimed at providing guidance and support for the installation of some software may try to estimate whether the user is an ex- pert or a novice in this field. Based on this information suitable words and explanations may be generated. These explanations may be very detailed and without specific vocabulary for a non-experienced user; in contrast, for an expert, the system may provide only a sequence of important steps or inform about more di!cult operations. From the beginning of the dialogue, SLDS may analyse the user’s speech, behaviour and requests and also the ex- isting di!culties. When deciding on the best response to a user, the dialogue manager may change words and sentence structures based on the informa- tion about cognitive processes. Its responses may become more helpful and the user-friendliness of the system may be improved. For this purpose it is necessary to identify di"erences in language use of people yielding di"er- 2 ent educational background and abilities to analyse situations and to solve problems. The ability to use language for accomplishing certain goals is called ver- bal intelligence (VI) [5, 3]. In other words, verbal intelligence is “the ability to analyse information and to solve problems using language-based reason- ing” [13]. Automatic verbal intelligence estimation may help dialogue sys- tems to choose the level of communication and be more simple, useful and e"ective. Figure 2 explains the adaptation process of spoken dialogue systems based on verbal intelligence estimation in more detail. When talking to the system, all j spoken utterances of a user are analysed for the verbal intelligence determination. This means that the intelligence level is re-estimated at each turn based on features extracted from the new spoken utterances and from all the phrases which were pronounced at the previous turns. In Figure 2 the SLDS has three di"erent dialogue scenarios corresponding to users yielding a higher, an average and a lower verbal intelligence. At the beginning of the dialogue, the systems uses scenarios corresponding to users yielding an average verbal intelligence. At the following turns, the system might switch to alternative dialogue scenarios. Figure 2: Adaptation to the user. The automatic estimation of users’ verbal intelligence may help SLDS to more e"ectively control the flow of the dialogues, engage users in the 3 interaction and be more attentive to human needs and preferences. For training machine learning algorithms, we need to know a maximum number of language features that reflect di"erences in language use of people yielding di"erent verbal intelligence. In this work we investigate to which extent the vocabulary of test persons reflect their levels of verbal intelligence when they all describe the same event and explain their thoughts and feelings about it. The investigation is based on a corpus containing descriptions of a short film along with the corresponding intelligence scores of the speakers [32] . The paper is structured as follows. In Section 2 we describe the cor- pus which was used for the experimental research. Section 3 describes our primary e"orts at defining film related features which could be useful for distinguishing test persons yielding di"erent verbal intelligence. Section 4 describes typical TF-IDF approaches and explains the details of the feature selection process for the monologues. In Section 5 we describe and compare the Naive Bayes, k-nearest neighbours and Rocchio classifiers. Classification results are presented and discussed in Section 6. Finally, Section 7 presents conclusions and future work. 2. Corpus Description For the data acquisition in [32], a short film was shown to German native speakers. It described an experiment on how long people could stay without sleep. The test persons were asked to imagine that they met an old friend and wanted to tell him about this film. Our goal was to record everyday speech when talking to relatives and friends. This corpus, described in [32], consists of 56 descriptions (3, 5 hours of audio data) of a short film (i.e. monologues). The test persons were also asked to participate in the verbal part of the Hamburg Wechsler Intelligence Test for Adults (HAWIE) [29]. According to Wechsler, intelligence is “the global capacity of a person to act purposefully, to think rationally, and to deal e"ectively with his environment” [28]. The verbal part consists of the following subtests: • Information: this subtest measures general knowledge and includes questions about history, geography, literatures, etc. For example, What is the capital of Italy? • Comprehension: test persons are asked to solve di"erent practical prob- lems and explain some social situations. For example, What would you do if you lost your way in a forest? 4 • Digit Span: test persons are asked to repeat increasingly longer strings of numbers forward and then backward; the subtest measures short- term memory. • Arithmetic: test persons are asked to solve some arithmetic problems given in a story-telling way; the subtest measures their concentration and computational ability. For example, How many rolls you can buy if you have 36 cents and one roll costs 4 cents? • Similarities: test persons are asked to find a similarity between a pair of words. For example, Please find a similarity between “wood” and “alcohol”? • Vocabulary: test persons are asked to explain increasingly more di!- cult words using their vocabulary. For example, What does “to creep” mean? Raw scores of each test person on the verbal test are based on the correct answers (Figure 3). The raw scores are then converted into “Scaled Scores” using special tables [29]. The Scaled Scores vary between 0 and 16 and may be used to compare the performance of the participants. The sum of the scaled scores and the age of a test person are used to estimate the corresponding verbal intelligence score.  Figure 3: Verbal Part of the Hamburg Wechsler Intelligence Test for Adults 5 Overall, 56 test persons yielding di"erent educational levels were tested, therefore 56 monologues about the same topic were collected. All the mono- logues and the film were transcribed according to the transcription standards by Mergenthaler [15]. 3. Modelling Verbal Intelligence by Using Film Derived Features To analyse the vocabulary of people yielding di"erent verbal intelligence when describing the same event, at first we decided to compare the mono- logues with the film transcription. Figure 4 shows excerpts from the film and from one of the monologues 1. Excerpt from the film Max and Funda have been without sleep for fifty eight hours. They have laid down on the sofa. Is it a mistake? Actually they would like to move. But now they cannot any more. The blood pressure is down, the energy reserves are over. They both are freezing despite the fire-place and the jacket. The question is who closes the eyes first. It is Max. Funda wins. She stays awake a few minutes longer. Excerpt from a corresponding monologue After fifty eight hours, they were really tired. And, they had frozen. Despite they had very warm clothes. And then the man fell asleep and then the woman. Figure 4: Excerpts from the film and one of the recorded monologues. For the comparison, the following features were extracted: • Number of reused words - number of words which a test person “reused” from the film. For our example in Figure 4 the reused words are: fifty, eight, hours, they, and, they, despite, they, and, the, and, the. • Number of unique reused words. It includes the number of reused words without repetitions. In Figure 4, the unique reused words are: fifty, eight, hours, they, and, despite, the. • Number of all reused lemmas. This feature is similar to the Number of all reused words, but referred to lemmas instead. 1As the conversation language is German, the example was directly translated into English. 6 • Number of unique reused lemmas. This feature is similar to the Number of unique reused words, but considering unique lemmas instead. • Cosine similarity between the film and a kth monologue using lemmas. For this feature extraction, we created a matrix consisting of all unique lemmas from the film, including the frequency of these lemmas within the film and within a kth monologue. Table 1 shows this matrix for the texts from our example (Figure 4). Table 1: Matrix for lemma frequency Lemmas from film Frequency (film) Frequency (monologue) Max 2 0 and 1 3 Funda 1 0 have 2 2 be 6 1 without 1 0 sleep 1 0 for 1 0 fifty 1 1 eight 1 1 hour 1 1 The frequencies were normalized by the total amount of words in the corresponding text; the cosine similarity between the two normalized vectors (lemma frequencies within the film and lemma frequencies within a kth monologue) was calculated as: similarity = !n i=1 aibi!n i=1 ai 2 !n i=1 bi 2 , where n is the number of unique lemmas in the film, ai - frequency of ith lemma in the film, bi - frequency of ith lemma in the monologue 2. • Number of reused n-grams. For this feature we have calculated the number of n-grams (n = 2, 10) that were used in the film and then 2Cosine similarity will be further used in the Rocchio classification approach that will be explained in detail in Section 5.2. 7 reused by a test-person in the corresponding monologue. In our exam- ple, the number of reused 2-grams equals to 2 (reused 2-grams are fifty eight and eight hour), the number of reused 3-grams equals to 1 (fifty eight hour), etc. • Cosine similarity using n-grams. The cosine similarity was calculated from a feature vector composed by the counts of di"erent n-grams for each monologue. • We also determined the number of lemmas that were used by the candi- dates but were not used in the film. For each monologue the following features have been calculated: Own lemmas1 = n" i=1 frequency(lemmai) ! count(lemmai) and Own lemmas2 = n" i=1 frequency(lemmai), where n is the number of unique lemmas that were used by a test person but were not used in the film; count(wordi) shows how many times lemmai was used in the monologue; frequency(lemmai) shows the frequency of lemmai according to a frequency dictionary of the German language [11]. This dictionary consists of 40000 German words with frequency from 1 to 17: 1 corresponds to more frequent words, 17 corresponds to less frequent words. If a word from the monologues was not found in the dictionary, its frequency was set to 20. 3.1. Feature Analysis The k-means algorithm, which is frequently used for data clustering in machine learning, was applied on the scaled scores of the test persons (Fig- ure 5). For the feature analysis two experiments were performed. In the first experiment the observations were partitioned into two clusters: cluster P1 consisted of test persons yielding a lower verbal intelligence, P2 contained candidates yielding a higher verbal intelligence. In the second experiment the test persons were partitioned into three clusters: P1 - lower verbal intel- ligence, P2 - average verbal intelligence, P3 - higher verbal intelligence. 8       Figure 5: The K-means algorithm The averaged values of all the features from the two clusters were com- pared using a one-way analysis of variance (ANOVA). In Experiment I with two clusters, features with small p-values were: • Number of reused 3-grams (averaged value for the first class AVlow = 0.021, averaged value for the second class AVhigh = 0.031, p = 0.012, F = 6, 63); • Cosine similarity using lemmas (AVlow = 0.79, AVhigh = 0.83, p = 0.03, F = 4, 64); • Cosine similarity using repeated n-grams (AVlow = 0.13, AVhigh = 0.15, p = 0.01, F = 7, 07). In Experiment II with three clusters, a feature with a small p-value was: • Cosine similarity using repeated n-grams (AVlow = 0.13, AVaver = 0.14, AVhigh = 0.16, p = 0.01, F = 7, 07). As we can see, participants with a higher verbal intelligence used more words from the film and the similarity between their descriptions and the film was higher than the similarity of participants with an average and a lower verbal intelligence. This may be explained in the following way. Test persons yielding a higher verbal intelligence (class HIGH ) may have a better ability to listen to and recall spoken information from the film. Mem- ory is indeed one of the verbal sub-tests of HAWIE so that a high memory score relates to a high verbal intelligence score of a test person. Therefore, people with good memory (i.e. higher verbal intelligence) were easier able to 9 remember many details of the film and to use words which they heard when watching the program. They may also better understand the relationships between language concepts, make more sophisticated language analogies or comparisons and perform a more complex language-based analysis. Hence, we may conclude that the vocabulary of test persons yielding di"erent verbal intelligence was di"erent when they talked about the same event, even despite they were asked to talk about this film just after they had watched it. 4. Text Categorization Solutions Film derived features presented in the previous section showed to be good predictors of verbal intelligence. Particularly, some of them suggested that test persons belonging to di"erent verbal intelligence classes may be distin- guished by word or lemma patterns, even regardless of the order of these words and lemmas in the monologues. This result led us to the main hypothesis that we investigate in this work: is it possible to solve the problem of inferring the corresponding level of verbal intelligence by simply applying conventional text categorization (TC) techniques? To validate this hypothesis, typical TF-IDF features (introduced in the next section) have been extracted from the transcripts of the monologues (henceforth we do not make any use of the film transcription). Three of the most popular TC methods have been applied for the auto- matic classification of monologues into three groups: test persons yielding a lower, an average and a higher verbal intelligence. 4.1. TF-IDF based Approaches TF-IDF (term frequency - inverse document frequency) based approaches are often used in information retrieval and text mining. As an example of a typical text mining task we may refer to the text categorization. The goal of TC is the classification of documents into a fixed number of predefined categories. The applicability of TC techniques has significantly grown in recent years. Organizing news by subject topics (e.g. to disambiguate information and to provide readers with greater search experiences) or papers by research domains (e.g. for large databases of information that need indexing for re- trieval) are just some of the most popular examples. Moreover, Security 10 (e.g. analysis of plain text sources such as Internet news), Biomedical (e.g. indexing of patient reports in health care organizations according to disease categories) or Software (e.g. for tracking and monitoring terrorist activities) domains also have benefit from these techniques. New domains, like Marketing (e.g. analytical customer relationship ma- nagement) or Sentiment analysis (e.g. analysis of movie reviews), start using text mining solutions. In this work we have applied these techniques to a new domain: the estimation of speakers’ verbal intelligence. For TC, every document has to be transformed into a representation which could be suitable for learning algorithms and classification tasks. As reviewed in [16], most TC algorithms are based on the vector space model (VSM). TC state-of-the-art systems widely apply the VSM approach [1, 26, 16]. Information retrieval (IR) research suggests that words work well as rep- resentation units. In VSM, each document in a corpus is represented by a list of words (i.e. bag of words). Each word is considered as a feature; the value of the feature is a weight transformation of the number of times the word occurs in the document (i.e. word’s frequency). Thus, a document is represented as a feature vector and its relevance to a query submitted by a user is measured through appropriate matching functions. These match- ing functions are typically based on statistical measures, like TF-IDF, that basically weight the importance of each word. The importance of a word increases proportionally to its frequency within a document but is o"set by its frequency within a corpus. Variations of this TF-IDF weighting scheme are often used by search engines as a central tool in scoring and ranking a document’s relevance given a user query. 4.1.1. Mathematical Details TF-IDF is a common feature transforming or weighting function. The term count, ni,j, denotes the frequency of a given term ti in a given docu- ment dj. This count is usually normalized to prevent a bias towards longer documents. Thus, the term frequency tfi,j measures the importance of a term ti within a document dj and is defined as follows: tfi,j = ni,j! k nk,j (1) 11 where the denominator is the number of words in a document dj, that is, the size of the document |dj|. The inverse document frequency idfi is a measure of the general impor- tance of a term: idfi = log |D| {j : ti " dj} (2) where |D| is the total number of documents in the corpus, {j : ti " dj} is the number of documents where the term ti appears (i.e. documents for ni,j #= 0). The feature weighting function is then computed by using the following formula: tfidfi,j = tfi,j · idfi (3) These weights show the importance of the words in each document. As can be seen, more frequent terms in a document are more representative and, if the number of documents in which this term occurs increases, this term becomes less discriminative. At this point, we may view each document as a vector that contains terms and their corresponding weights. For those terms from the vocabulary that do not occur in a document this weight equals to zero. In the following sections we will show the advantage of such a document representation. 4.2. Feature Selection Typical TC approaches make use of di"erent feature selection techniques to further reduce the dimensionality of the data space by removing irrelevant features that have no contribution to category discrimination. Di"erent feature selection techniques through information theory were well studied in [31]. As a result of this study, information gain (IG) and v2-test (CHI) were reported to be the top performing methods out of five methods under test in terms of feature removal aggressiveness and classifica- tion accuracy improvement. However, the document frequency thresholding approach, the simplest method with the lowest cost in computation, was reported to perform similarly. The Document Frequency (DF) is the number of documents in which a term occurs. As described in [31], it is possible to compute the document frequency for each unique term in the training corpus and to remove from the 12 feature space those terms whose document frequency is less than a certain predefined threshold. By doing so we are adopting a basic assumption: rare terms are either non-informative for the category prediction (i.e. intelligence estimation in our case) or not influential in global performance. In either case, removal of rare terms contributes to the reduction of dimensionality of the feature space and improves the classification accuracy (i.e. if rare terms happen to be noise terms). If we try to summarize both pros and cons of using the document fre- quency thresholding approach, we may say that positive aspects are: • It is the simplest technique for vocabulary reduction (easily scalable to a very large corpora). • Computational complexity is approximately linear with the number of documents. while on the other hand, negative aspects are: • The technique is usually considered as an ad-hoc approach to improve the e!ciency instead of a principled criterion for a predictive features selection. • The technique is typically considered, from an IR point of view, as a non-appropriate approach for aggressive term removal (low-DF terms are assumed to be relatively informative and therefore should not be removed aggressively). In this work a slightly modified version of this DF thresholding approach was applied to the data: TF-IDF measures instead of DF measures were used. As another remarkable di"erence, we did not remove the lowest TF-IDF terms but just selected the highest TF-IDF terms. In particular, instead of defining a threshold for TF-IDF measures, we defined a fixed number of terms to be selected (i.e. N). Therefore, we first sorted all the terms according to their TF-IDF measures. Then, we selected the top N most representative or indicative terms according to their TF-IDF weights. The remaining terms were removed as stop or common words that did not add any meaningful content. By observing the evolution of the classification accuracy with an increasing N value, we determined the minimum size of the vocabulary (i.e. dimensionality) required to achieve the optimum performance. 13 4.2.1. Class-based vs Corpus-based As stated above, in our framework each word is considered as a feature and each document is represented as a feature vector. In [19] two alternative ways for implementing the selection of these keywords or features are presented. In the first one, the so-called corpus-based keyword selection, a common keyword or feature set that reflects the most important words for all classes (i.e. highest TF-IDF terms) in all documents is selected. In the alternative approach, named as class-based keyword selection, the keyword selection process is performed separately for each class. In this way, the most important and specific words for each class are determined. 4.2.2. Word Lemmatisation Word lemmatisation is often applied in the area of IR, where the goal is to enhance the system performance and to reduce the number of unique words [23]. Particularly, word lemmatisation is part of the data pre-processing required to convert a natural language document to the feature space. For- mally, it is the process for reducing inflected (or sometimes derived) words to their lemmas. For example, as a result of lemmatisation, di"erent words like “play”, “plays”, “playing” and “played” are related to the same feature identification (i.e. lemma) “play”. Word lemmatisation was applied to our monologues to assess its impact on performance (i.e. classification accuracy). Like removing stop words, lemmatisation also contributed to the reduction of the size of the lexicon, thus saving on computational resources. 5. Vector Space Classification As stated above, the vector space model represents each document as a vector with one real-valued component (i.e. TF-IDF weight) for each term. Therefore, we need text classification methods that can operate on real- valued vectors. In this section we introduce those ones that have been tested so far. A number of classifiers has been used to classify text documents, including regression models, Bayesian probabilistic approaches, Nearest Neighbours approaches, Rocchio algorithm, decision trees, inductive rule learning, neural networks, on-line learning, Support Vector Machines (SVMs), and combining classifiers [12, 7, 22, 30]. In this work we used three well-known vector space 14 classification methods: Naive Bayes (NB), Rocchio and Nearest Neighbour classification (kNN). NB is often used as a baseline in text classification research as it com- bines e!ciency (training and classification can be accomplished with one pass over the data) and good accuracy (particularly if there are many equally im- portant features that jointly contribute to the classification decision). The Rocchio algorithm is a very simple and e!cient text categorization method for applications such as web searching, on-line query, etc. because of its sim- plicity in both training and testing [26]. kNN requires no explicit training and can use the unprocessed training set directly in classification. However, it is less e!cient than the other classification methods (i.e. with kNN all the work is done at run-time so that it can have poor run-time performance if the training set is large). Rocchio and Naive Bayes are linear classifiers whereas kNN is an example of a non-linear one. Generally speaking, if a problem is non-linear and its class boundaries cannot be approximated well with linear hyperplanes, non- linear classifiers are often more accurate than linear classifiers (particularly, if the training set is large, then kNN can handle complex classes better than Rocchio and NB). On the other hand, if a problem is linear, then it is better to use a simpler linear classifier. However, this needs to be taken with a little bit of salt since the previous assertion is always conditioned by the well-known bias-variance trade-o" (i.e. with limited training data, a more constrained model tends to perform better). These approaches are described in more detail in the following sections. Among the enumerated alternatives, SVMs are widely used mainly be- cause they have much current theoretical and empirical appeal and perform at the state-of-the-art level. According to [22], SVMs, example based meth- ods, regression methods and boosting based combining classifiers deliver top- notch performance. Lewis et al. (2004) also found that SVMs perform better on Reuters-RCV1 corpus than kNN and Rocchio. Nonetheless, recent revisions of the selected algorithms have proposed en- hanced versions of these methods that achieve relatively close performance to the top-notch TC classifier: SVMs. For instance, Miao and Kamel (2010) have re-examined the applicable assumptions and parameter optimization method of the traditional Rocchio algorithm and proposed an enhanced ver- sion of this method that clearly outperforms the former one by using a pair- wise optimized strategy. Salles et al. (2010) also presents a methodology to determine the impact that may have temporal e"ects on TC and to minimize 15 it. By extending the three algorithms (namely kNN, Rocchio and NB) to in- corporate a Temporal Weighting Function (TWF), experiments showed that these temporally-aware classifiers achieved significant gains, outperforming (or at least matching) state-of-the-art algorithms. In any case, and as discussed in [14], despite believes of many researchers that SVM is better than kNN in terms of e"ectiveness, kNN is better than Rocchio and Rocchio is better than NB, the ranking of classifiers ultimately depends on the classes, the document collection and the experimental setup. 5.1. Naive Bayes Classification The first supervised learning method we introduce is the multinomial Naive Bayes or multinomial NB model, a probabilistic learning method [14]. According to this method, the probability of a document d being in a class c can be computed as: P(c, d) $ P(c) · # 1!k!nd P(tk|c) (4) where P(tk|c) is the conditional probability of a term tk occurring in a doc- ument of a class c. It may also be interpreted as a measure of how much evidence tk contributes that c is the correct class. P(c) is the prior probabil- ity of a document occurring in a class c. Terms %t1, t2, · · · , tnd& are part of the vocabulary that is used for the classification; nd is the number of terms. In NB classification, the best class for a document d is determined as: cmap = arg max c"C $P(c|d) = arg max c"C $P(c) · # 1!k!nd $P(tk|c) (5) where $P refers to the parameters to be estimated from the training data by applying the Maximum Likelihood Estimation (MLE). The interpretation of this equation is rather simple. Each conditional parameter P(tk|c) is a weight that indicates the quality of an indicator tk for a class c. Similarly, the prior P(c) is a weight that indicates the relative frequency of c. More frequent classes are more likely to be determined as the correct class. To reduce the number of parameters, we adopted the Naive Bayes con- ditional independence assumption where attribute values are independent of each other given the class so that for our multinomial model: P(d|c) = P (%t1, t2, · · · , tnd&|c) = # 1!k!nd P (Xk = tk|c) (6) 16 where Xk is a random variable for a position k in the document and the values of Xk are terms from the vocabulary. P (Xk = tk|c) is the probability that in a document of a class c a term t will occur in a position k. To further reduce the complexity of our multinomial model (assuming a di"erent probability distribution for each position k in the document still re- sults in too many parameters), we made a second independence assumption: conditional probabilities for a term are the same regardless of its position in a document: P (Xk1 = t|c) = P (Xk2 = t|c) (7) where X is a single distribution of terms which is exactly the same for any po- sition k1, k2, · · · , ki. Equation 7 applies for all terms t and classes c. This po- sitional independence assumption is equivalent to adopting the bag of words model, which we introduced in Section 4.1. This bag-of-words model dis- cards all information that is communicated by the order of words in natural language sentences. 5.1.1. A Variant of the Multinomial Model A critical step in solving a text classification problem is to choose the document representation. An alternative formalization of the multinomial model represents each document d as a M-dimensional vector of counts: tf'idft1,d, tf'idft2,d, · · · , tf'idftM ,d where tf'idfti,d is the TF-IDF measure for a term ti in a document d. P(d|c) is then computed as follows: P(d|c) = P % %tf-idf t1,d, tf-idf t2,d, · · · , tf-idf tM ,d&|c & (8) All the model parameters (i.e. class priors and feature probability distri- butions) may be estimated from the training set by using MLE. For every class’ prior we calculated an estimate for the class probability from the train- ing set (i.e. (prior for a given class) = (number of samples in the class) / (total number of samples)). To estimate the parameters for our feature dis- tribution, we adopted the typical assumption that the continuous values as- sociated with each class are distributed according to a Gaussian distribution. Particularly, assuming that the training data contains continuous attributes, i.e. TF-IDF measures for each term and document, we first segmented the data by the class and then computed the mean and variance of every term specific TF-IDF measure in each class. 17 5.1.2. About the Independence Assumptions Typically, TC tasks rather look at the words themselves and not at their corresponding positions in the documents (i.e. bag of words). This relies on the hypothesis that each topic or class to be distinguished is fairly represented by only some specific words from our bag. The NB models often perform well for TC tasks despite the conditional independence and the positional independence assumptions. In fact, both assumptions are very important to avoid problems in estimation owing to data sparseness. By adopting both independence assumptions, we are committed to a spe- cific way of processing the evidence. Particularly, in the NB classification we look at each term separately so that we do not make a di"erence between word A followed by word B and word B followed by word A (although there is a di"erence between them). However, the conditional independence assump- tion does not really hold for text data as terms are conditionally dependent on each other. Additionally, the position of a term in a document by itself may carry more information about the class than expected. 5.2. The Rocchio Approach The Rocchio classification [20] divides the vector space into di"erent re- gions centred on prototypes. These prototypes or centroids, one for each class, define the class boundaries (i.e. hyperplanes). For a given training dataset, the centroid of a class c can be computed as the vector average or centre of mass of its members (i.e. all documents in the class) [14, 8]. '(µ (c) = 1 |Dc| " d"Dc '( V (d) (9) where Dc is the set of documents from class c: Dc = ' d : %d, c& " D ( ; '( V (d) = V1(d) · · ·VM(d) is a vector that contains tf-idf weights for each term of a document d. As many vector space classifiers (e.g. computing the nearest neighbours in kNN classification), the Rocchio approach relies on distance-based deci- sions (from a TC point of view, the relatedness of two documents can be typically expressed in terms of similarity or distance). Particularly, the Roc- chio classification rule is to classify a point in accordance with the region it falls into. To do this, basically we determine the centroid '(µ (c) that the point is closest to and then assign it to c. 18 In our experiments, we used the cosine similarity measure as the un- derlying distance. Cosine similarity is the cosine of the angle between two vectors and determines whether they are pointing in roughly the same direc- tion. Since the components of our vectors (i.e. tf-idf weights) could not be negative, the angle between two tf-idf vectors could not be greater than 90#. The vector representation '( V (d1) and '( V (d2) of the cosine similarity be- tween two documents d1 and d2 is: sim (d1, d2) = '( V (d1) · '( V (d2) | '( V (d1)|| '( V (d2)| (10) where | '( V (d1)| and | '( V (d2)| it the Euclidean length of the vectors. By using this measure, we are also applying a normalization process which makes each vector of the same length [21]. If we have a look at the magni- tude of the vector di"erence between two vectors corresponding to documents with very similar content, it may happen that this di"erence is significantly simple because one is much longer than the other. Cosine similarity measure compensates this e"ect of document length so that the similarity between document vectors is reduced to only measuring the cosine of the angle be- tween them. We can rewrite Equation 10 as follows: sim (d1, d2) = '(v (d1) · '(v (d2) (11) where '(v (d1) = '( V (d1)/| '( V (d1)| and '(v (d2) = '( V (d2)/| '( V (d2)|. The assign- ment criterion for a document d and its vector representation '( V (d) can be defined as: crocchio = arg max c"C sim ) '(µ (c), '( V (d) * (12) In our implementation of the Rocchio approach [4, 10] only positive train- ing samples are considered for obtaining the prototype for each class (i.e. training samples that belong to the corresponding class). However, recent variations of Rocchio [22, 17, 2] consider the e"ects of negative samples (i.e. training documents that belong to all other classes) when computing the prototypes for the defined classes. Di"erent parameters may be used to con- trol the relative importance of positive and negative samples. These Rocchio classifiers reward not only the closeness of a test document to the centroid 19 of the positive training instances, but also its distance from the centroid of the negative training instances. 5.3. K-nearest Neighbours In pattern recognition, the k-nearest neighbour algorithm (kNN) is a method for classifying objects based on the closest training examples in the feature space. In TC, kNN takes an arbitrary input document and ranks the k nearest neighbours among the training documents through the use of a similarity score (i.e. cosine similarity distance). It then assigns to the input the category or the class of the most similar document or documents. A constant k, defined by a user, denotes the number of neighbours included in the evaluation. The kNN algorithm is a valid non-parametric method. Despite being amongst the simplest of all machine learning algorithms, it is one of the best methods when the text is described by using VSM [30]. However, traditional kNN has two main drawbacks: the intensive computational e"ort, especially when the size of the training set grows (training examples are vectors in a highly multidimensional feature space), and its sensitiveness to the local structure of the data [9]. New nearest neighbour algorithms have been recently proposed mainly with the purpose of reducing the number of distance evaluations actually performed, thus trying to make kNN computationally tractable even for large data sets. For instance, [27] presents a fast kNN algorithm that reduces the cost of similarity computing in order to raise the classifying speed and applicability of kNN. In Naive Bayes and Rocchio classification we have to estimate correspond- ing parameters: priors and conditional probabilities and centroids. In kNN we do not need to estimate any parameters but simply memorize all exam- ples in the training set and then compare a test document to them. For this reason, kNN is also called memory-based learning or instance-based learning. The kNN algorithm is known because of its strong consistency results. As the amount of data approaches infinity, the algorithm is guaranteed to yield an error rate no worse than twice the Bayes error rate (the minimum achiev- able error rate given the distribution of the data) [14]. kNN is guaranteed to approach the Bayes error rate for a certain value of k (where k increases as a function of the number of data points). The k-nearest neighbour methods may be improved by using proximity graphs [25]. 20 5.3.1. Choosing the Class for an Unclassified Document To make a decision on a number of unclassified documents, we measure their similarity with all the documents that have already been classified. The unclassified documents are then ranked according to their similarity scores. Appropriate classes for the documents may be assigned in the following ways: • If we choose k = 1, the class is predicted to be the class of the closest training sample. This is called the nearest neighbour algorithm. • If we choose k > 1, then all the documents which ranks are smaller than or equal to k will be included in the ranked list. We can then use di"erent means to find a class for our document, like: – we may assign the document to the most common class amongst its k nearest neighbours (if we are dealing with a binary, i.e. two- class, classification problem, it is helpful to choose k to be an odd number as this avoids tied votes). – we may estimate the probability of membership in a class c as the proportion of the k nearest neighbours in c. This is commonly referred as the probabilistic version of the kNN classification al- gorithm. – for the individual classes, we may sum the distances to all the doc- uments in which the class occurs, and then choose the class corre- sponding to the highest accumulated distance (remember that we are using cosine distance). – etc. If we decide to use either the basic “majority voting”, the probabilistic method or the sum of distances based classification, those classes with more frequent examples will tend to dominate the prediction of a new vector. This is actually a drawback as they tend to come up in the k nearest neighbours when the neighbours are computed due to their large number (it is important to remind that the available data is certainly imbalanced). To overcome this problem, we compensate this possible imbalance by introducing a slightly modified classification method for our kNN based approach. Typically, the implementation of these versions of the algorithm starts by computing the distances from the test sample to all stored vectors of 21 the training data set. Next, all these training samples are sorted according to these distances thus ranking the nearest k training samples regardless of their corresponding class. If we look at those already labelled classes instead (neighbours are taken from a set of documents for which the correct classification is known), we could then identify specific top k neighbours for each class c (i.e. k nearest neighbours labelled as c, thus resulting in an overall list composed of k ) C elements). Finally, by computing distance to each class as the average distance between the test sample and those top k class specific neighbours, we will then manage to compensate any possible imbalance in the distribution of the training data among the defined classes. The best class for kNN classification can then be derived from: ckNN = arg max c"C score(c, d) = arg max c"C 1 k · " d!"Sk(c,d) sim('(v (d$), '(v (d)) (13) where Sk(c, d) is the c class specific set of d’s k nearest neighbours. As could be derived from Equation 13, it may also be useful to weight the contributions of the neighbours so that nearer ones contribute more to the average than more distant ones [24]. This classification method is weighted by taking into account not only the distance from the test sample to the c set of k nearest neighbours, but also the class compactness particularly for high values of k. 5.3.2. Parameter Selection The parameter k in kNN is typically defined by using some previous experience or specific knowledge about the classification domain. Normally, 1NN is found to be not very robust. The accuracy of the kNN algorithm can be severely degraded by the presence of noisy or irrelevant features (also if the feature scales are not consistent with their importance). 1NN means that the classification decision for each test document only relies on the class of a single training document, whose label could eventually be incorrect or atypical. kNN for k > 1 are more robust as larger values of k tend to reduce the e"ect of noise on the classification (although also make boundaries between classes less distinct). As an alternative, a good value of k can be assigned heuristically via cross validation technique or empirically via bootstrap method [6]. In our experiments, instead of applying any of these parameter selection methods, we tried di"erent k values, thus finally selecting the optimal k as the value which was used when obtaining the best performance. 22 6. Experimental Results and Discussion 6.1. Experimental Set-up Our main goal is to identify the algorithm that best computes class bound- aries and reaches the highest classification accuracy. In our experiments for comparing the performance of the di"erent approaches, a Leave-One-Out cross validation (LOO-CV) method was used. The idea of this method is to use N-1 observations for training (where N is the number of data points) and only 1 data point for testing. This procedure is repeated N times and each observation is used once as the testing data. 6.2. Baseline Approach: Class-Based vs Corpus-Based Feature Selection As introduced in Section 4.2.1, our experiments covered the comparison of the class-based and corpus-based keyword selection approaches. The corpus-based approach implies the selection of a common feature set for all classes with the top N most representative or indicative terms. The class-based approach instead implies the selection of the most important words for each particular class. In this case, to preserve the balance between classes, N/M words for each specific class were selected, where M is the number of classes. For our classification task M equals to 3, where the first class contained test persons yielding a lower verbal intelligence, the second class contained participants yielding an average verbal intelligence and the third class contained participants yielding a higher verbal intelligence. Then, we composed our feature vector by concatenating all the class-specific features, thus resulting into a vector comparable to the N-dimensional vector corresponding to the corpus-based approach. However, when using the class-based approach, a particular word may be included in various class-specific subsets (i.e. a word that is important to not only one single class but to several classes). To avoid using duplicate features, we only used the intersection between all the class-specific subsets. Therefore, the dimension of the resulting feature vector in these cases had to be necessarily lower than N. For simplicity, we will better refer to the number of features per class (i.e. F = N/3) rather than to the final dimensions of the vectors. Consequently, if we report, for instance, about 50 words or features per class, this means that we are using a 150-dimensional corpus- based vector. In this case for the class-based approach, 150 is the maximum number of dimensions. To definitely determine its value, it is necessary to check the possible intersection. 23 Of course, the higher the value of F , the more significant the intersection between class-specific word subsets, and also the bigger the di"erence with respect to the corpus-based vector dimensions. Analysing the corpus, 2210 di"erent words were extracted from all the monologue transcripts. Table 2 shows how the intersection evolved according to F . Considering the size of the vocabulary, the observed di"erence is significant. Table 2: Dimension di!erences between class-based and corpus-based approaches. # of features per class (3 classes) 50 100 150 200 250 300 350 400 450 500 Corpus-based 150 300 450 600 750 900 1050 1200 1350 1500 Class-based 150 289 393 486 557 601 737 858 992 1102 Di!erence 0 11 57 114 193 299 313 342 358 398 Rel. di!. (%) 0 3, 7 12, 7 19 25, 7 33, 2 29, 8 28, 5 26, 5 26, 5 Figure 6 presents the accuracy results obtained using either the corpus- based or the class-based feature selection methods. The results were obtained using the NB approach for di"erent dimensions of the feature vector. Confi- dence intervals of 95% are also shown in the figure. As it can be derived from the figure, the class-based approach clearly out- performed the corpus-based one regardless of any di"erence about the used dimension. Although the observed di"erences were not statistically signifi- cant in any case, it is interesting to pinpoint the result for the 155-dimensional value. At this point the class-based approach reached the top performance while the di"erence with the corpus-based alternative also turned to be the biggest one thus becoming almost significant. From a di"erent point of view, we may also try to analyse the min- imum dimensionality required by the class-based approach to outperform the corpus-based one. The corpus-based approach obtained a maximum ac- curacy of 51, 79%. As can be observed in Figure 6, this performance was reached with dimensionality equal to or higher than 110. Also derived from this figure, we may check that the class-based approach obtained a better performance of 57, 14% (though not statistically significant) using “only” 20 features per class. The class-based feature selection, by definition, focuses on finding the most crucial or indicative class keywords. On the other hand, 24 Figure 6: Baseline approach: Class-Based vs Corpus-Based feature selection. the corpus-based one simply tends to find general keywords concerning all classes. This clearly tips the balance in favour of the class-based approach particularly when we use a reduced set of features. This is important as there may be a significant gain in classification time when a small number of features is used. By confirming these di"erences with additional statistical evidence (i.e. more data), we may also conclude that the class-based feature selection im- proved the performance of the corpus-based one for the NB approach not only in terms of accuracy but also in terms of time. Similar results were already confirmed in [19]. When using the corpus-based approach, most features (i.e. words) tend to be selected from the prevailing classes so that rare classes are not well rep- resented. In contrast, when using the class-based approach all the classes are represented equally well as for their representation class specific features are used. Thus, the class-based approach achieved consistently higher accuracies than the corpus-based approach. Similar di"erences between the class-based and corpus-based methods 25 have been consistently observed throughout all of our experiments. There- fore, in the next sections we will only focus on the class-based versions. 6.3. Comparison between Approaches: Rocchio “Wins” In this section we compare the results that were obtained using di"erent approaches. Before proceeding with this comparison, we need first to assign the optimal configuration (i.e. k value) for the kNN approach. Figure 7 presents classification results corresponding to several k values. As expected, 1NN was found to be not very robust. Optimal performance may be reached by using k = 3 in combination with dimensionality of 155. However, if we keep increasing the value of k, which is typically more robust as it helps to reduce the e"ect of noise on the classification, then the results apparently start to be a"ected by sparse data bias. Figure 7: kNN results for di!erent k values. As a result of the initial k-means clustering, only 13 samples were defined to be part of the least populated class. Therefore, starting with 1NN we checked out up to k = 12 values leaving one sample out for testing (the LOO 26 approach was applied). For clarity, Figure 7 presents classification results only with some values of k. The observed di"erences were found to be statistically significant for the top performance dimensionality (i.e. F = 155) when comparing the best configuration (i.e. k = 3) with all the others for k > 5. No statistically significant di"erences where observed between the best k and any k * 5 configurations. Figure 8 allows to compare the results of the NB approach, the Rocchio approach and the kNN approach with k = 3. Figure 8: Comparison between approaches: Rocchio wins. A first important result that we can derive from Figure 8 is that both Roc- chio and kNN are clearly outperforming the NB approach, although the top performance is defined for di"erent dimensionalities in each case. The kNN performance had a maximum accuracy of 92, 86% for 155-dimensionality, while Rocchio just required 15 features per class to improve it up to 95, 6%. Both results denoted a statistically significant di"erence when compared to the NB top performance, 66, 07% also for 155-dimensionality. However, we did not observe any significant di"erence between Rocchio and 3NN (natu- 27 rally, Rocchio was also significantly outperforming any k > 5 approach). As it typically occurs in TC tasks, most of the learning takes place with a small yet crucial portion of features (i.e. keywords) for a class. This is evident in the steeper learning curves that reach the top performance at relatively low dimensionality. Therefore, we may conclude that the class- based feature selection approach is shown to be successful in quickly finding the most crucial or indicative class keywords. Another visible result in Figure 8, common to all the tested approaches, is the performance decrease as the value of F increases (particularly beyond a 200-dimensional value). As we already introduced in Section 6.2 and proved in Table 2, the higher the value of F , the more significant the intersection between class-specific word subsets. If we expand this interpretation, the more significant the intersection, the less discriminative the class-specific subsets, the more likely to include words that are not really indicative of any of the classes, and so the performance decreases. 6.4. Using Words vs Lemmas As we introduced in Section 4.2.2, we also tried a word lemmatisation strategy (i.e. to group together those words that are in di"erent forms but with the same lemma). This strategy was implemented as part of the data pre-processing stage during the classification task. Figure 9 shows the results with and without word lemmatisation for our top performing approach: the Rocchio one. The main advantage of word lemmatisation is to reduce the dimension- ality of the data space. In a TC task, it is basically applied under the assumption that all the documents belonging to the same category or topic may include these lemmas appearing in di"erent forms, and of course, it makes sense to use them as they refer to words with similar meanings. TC tasks typically rely on this. However, to be successful and thus really en- hance system performance, there is another important hypothesis that also needs to be confirmed: each topic or class to be distinguished should be fairly represented by only some class-specific lemmas. While the former one happens to be true for most of the cases, the latter one, though also successfully applied in typical TC tasks, may reasonably not be true in our case. The main reason for this would be that, from this point of view, all the documents (i.e. monologues) could be regarded as belonging to the same category according to their topic or content: all the documents are about the film which the participants watched. Consequently, we could 28 Figure 9: Classification results using words vs lemmas. expect an important number of lemmas to be shared among the participants as they all were talking about the same topic. This is an important di"erence with conventional TC tasks where, nor- mally, the topics or classes are well separated according to their conceptu- alization. In contrast, in our domain we may expect the participants to be identifiable among others not by the concepts or ideas themselves but by the way they express these ideas. Therefore, in this particular case, we may expect lemmas not to have much contribution to category discrimination but the di"erent endings and forms instead. Hence, missing this discriminative information because of lemmatisation (simplifying words with di"erent forms into their more common roots) could have some undesirable consequences in classification and clustering. Moreover, the fact that all the participants were German native speakers could be particularly critical for this problem. In this regard it is important to remark that German is a very agglutinative language [18]. Compound words or words that consist of more than one lemma (i.e. compounding or word- compounding occurs when a person attaches two or more words together to 29 make one word), can be found very often in the German language. “Donau- dampfschi"fahrtsgesellschaftskapitänsmütze” (i.e. Danube steamboat ship- ping company Captain’s hat) is a good example of how long these compound words could be (they can be practically unlimited in length, particularly in case of biochemistry). The meaning of a compound word di"ers from the meanings of words which it consists of. Lemmatisation of compound words would simply re- duce them to their more common lemmas thus loosing this discriminative information. To what extent this argument could be either true or false is something that can be derived from Figure 9. In fact, the word-based approaches sys- tematically outperform the lemma-based ones. Confidence boundaries for both cases are also shown in this figure. As we can observe, di"erences be- come statistically significant mostly around the same dimensionality range that was previously pinpointed when referring to the top performance for both kNN and NB (particularly at a 155-dimensional value). However, the di"erences are not statistically significant at F = 15, the point at which Roc- chio reaches its maximum accuracy for both word-based and lemma-based approaches. 6.5. Tempted to Use More Classes Although the three-classes scheme can be found entirely suitable from a practical implementation point of view (i.e. participants yielding a lower, an average and a higher verbal intelligence), we were also interested in analysing the performance of the suggested approaches for a higher number of classes. This would enable a better granularity for the verbal intelligence classifica- tion. Figure 10 presents benchmarking results for 4 classes instead of 3 (as it was shown in Figure 8). From a practical point of view, these classes may correspond to the following levels of verbal intelligence: poor, average, high and very high respectively. In this regard it seems to be important to remark that working with a higher number of classes, like 5 or more, was practically infeasible because of sparse data problems (i.e. k-means resulted into unpopulated classes). As for 3 classes, the Rocchio approach showed the highest accuracy again (i.e. 87, 5% at F = 15). The optimal dimensionality remained to be the same as for 3 classes (i.e. F = 15). Regarding the comparison between Rocchio and kNN, the observed di"erences also remained to be not statistically significant. 30 Figure 10: Tempted to use more classes: 4 classes. Additionally, both Rocchio and kNN clearly outperformed the NB algo- rithm, once again by a significant margin. For these two, another e"ect starts to become evident: the top performance region, previously observed for di- mensionality values up to 200, now turns to be narrower, locating its limit approximately around a value of 100. As we simply increase the number of classes, it seems to be evident that the number of terms or features that are really indicative of each particular class becomes smaller, thus a"ecting the performance. From a general point of view, the resulting performance can still be deemed to be satisfactory as the error rate is only roughly 7% higher than with three classes. If we look at the confusion matrix, presented in Table 3, we may check that by adopting the four classes into three by grouping high and very high classes, predictability would be more similar reducing the gap roughly to the half (i.e. 94, 64% for three classes and 91, 07% for four classes adopted into three). Finally, it may be interesting, particularly from a practical point of view, to have a look at the upper-right and lower-left corners of the matrix: 0 31 Table 3: Confusion matrix corresponding to Rocchio’s top performance using 4 classes (F = 15). Prediction outcome a b c d Actual value a = poor 4 0 1 0 b = average 1 14 1 0 c = high 1 0 22 1 d = very high 0 1 1 9 errors. This means there was not any critical errors like regarding a lower verbal intelligence individual as a higher verbal intelligence one and vice- versa. 7. Conclusions and Future Work This work showed that verbal intelligence may be recognized by com- puters through language cues. The achieved classification accuracy can be deemed as satisfying for a number of classes that is reasonably high enough to enable its integration into SLDSs. To our knowledge, this is the first report of experiments attempting to automatically predict verbal intelligence. Some of the most popular TC algorithms were applied to this task: NB, Rocchio and kNN. NB models are typically expected to perform well for TC tasks despite the conditional independence and the positional independence assumptions. However, the performance of NB approach was significantly worse than with the other approaches: kNN and Rocchio. This suggests that this probabilistic classifier was more sensitive to the low number of examples available, mainly resulting into inaccurate probability estimates, than the vector space ones (computing distances to some relevant members or to a prototype of each defined class seems to be more robust against sparse data). On the other hand, and connecting with those independence assumptions, it is well known that conditional independence does not really hold for text data (even worse considering that our features are highly correlated). Fur- thermore, we firmly believe that, for this specific task, the position of a term 32 in a document by itself could carry more information about the class than expected, mainly because of the above mentioned peculiarities of our classi- fication task (i.e. it is not only about the words that participants used to denote their intelligence, but also the way they combined them). Therefore, our data is somehow violating these independence assumptions, thus finally explaining why the NB approach performed so poorly. In this regard, it would be very interesting to test a LM based TC approach to better validate this argument. Using the class-based feature selection approaches has proven to be an essential factor, not only to achieve a better inference performance but also to reduce its computational cost. Despite typically successful when applied to TC tasks, word lemmatisa- tion was not really helpful for our task. The word-based approaches system- atically outperformed the lemma-based approaches, thus pointing out some peculiarities of the classification task. Particularly, these results were found to be mainly explained by two di"erent factors: the same topic for collecting monologues and the use of the German language. Unlike typical TC tasks, our verbal intelligence prediction task is influ- enced by the necessary fact that the di"erent categories or classes to be identified are not well separated from a conceptualization point of view. Of course, it might be easier to distinguish people talking about di"erent topics from their everyday life although the results for such a comparison might not be objective. By letting the participants (i.e. people with di"erent interests and hobbies) to discuss their own topics, we would be then recognizing the topics themselves rather than people with di"erent cognitive processes. On the other hand, the use of German, a very agglutinative language, resulted to be a drawback with regards to word lemmatisation. By lemma- tisation of compound words (compounding is a pretty common phenomena in German) we are basically loosing the extra meaning that arises from the combination of the interrelated words. This meaning has proven to be really helpful to correctly discriminate between di"erent levels of verbal intelligence. In future work, it would also be interesting to examine how well the suggested approaches perform when integrated into existing SLDS. In this regard, it is important to remark that any application involving speech recog- nition will always introduce noise in the features that we used. This needs to be considered as it will surely reduce the presented accuracies. Testing these approaches with conventional SLDS would allow us to assess whether the accuracies we achieve are high enough or not for our intended application 33 (i.e. dialogue system adaptation). On the other hand, this also suggests the importance of finding some other features that could be more robust when being used in a conventional system. Prosodic features could be a good alternative; so it would be interesting to start working on an multimodal inference framework that could jointly exploit the potential of, among others, this kind of features. As we have already mentioned, the linguistic cues that we have used in this work could pose a problem, for instance, if we want to apply these solutions with the same users but across di"erent domains. In this regard, prosodic features would be found to be advantageous as they would also allow us to explore the possibility of finding topic independent solutions. Acknowledgement This work is partly supported by the DAAD (German Academic Ex- change Service). Parts of the research described in this article are supported by the Tran- sregional Collaborative Research Centre SFB/TRR 62 ”Companion-Technology for Cognitive Technical Systems” funded by the German Research Founda- tion (DFG). For this work, Fernando was granted a fellowship by the Caja Madrid foundation. References [1] Baeza-Yates, R. A., Ribeiro-Neto, B., 1999. Modern Information Re- trieval. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA. [2] Bi, Y., Bell, D., Wang, H., Guo, G., Guan, J., March 2007. Combining multiple classifiers using dempster’s rule for text categorization. Appl. Artif. Intell. 21, 211–239. URL http://dl.acm.org/citation.cfm?id=1392641.1392644 [3] Cianciolo, A. T., Sternberg, T. J., 2004. Intelligence: a Brief History. Blackwell Publishing. 34 [4] Dumais, S., Platt, J., Heckerman, D., Sahami, M., 1998. Inductive learn- ing algorithms and representations for text categorization. In: Proceed- ings of the seventh international conference on Information and knowl- edge management. CIKM ’98. ACM, New York, NY, USA, pp. 148–155. URL http://doi.acm.org/10.1145/288627.288651 [5] Goethals, G., Sorenson, G., Burns, J., 2004. Encyclopedia of leadership. No. v. 1 in Encyclopedia of Leadership. Sage Publications. URL http://books.google.es/books?id=kjLspnsZS4UC [6] Hall, P., Park, B. U., Samworth, R. J., 2008. Choice of neighbor order in nearest-neighbor classification. ANNALS OF STATISTICS 36, 2135. URL doi:10.1214/07-AOS537 [7] Hui, G. G., Wang, H., Bell, D., Bi, Y., Greer, K., 2003. Using knn model- based approach for automatic text. In: In Proc. of ODBASE’03, the 2nd International Conference on Ontologies, Database and Applications of Semantics, LNCS. pp. 986–996. [8] Ittner, D. J., Lewis, D. D., Ahn, D. D., 1995. Text categorization of low quality images. In: In Proceedings of SDAIR-95, 4th Annual Symposium on Document Analysis and Information Retrieval. pp. 301–315. [9] Jianliang, Y., Yongcheng, W., 2004. Application of iterative-knn based on knn and automatic retrieval in automatic categorization. Journal of The China Society For Scientific and Technical Information 23, 137–141. [10] Joachims, T., 1998. Text categorization with support vector machines: Learning with many relevant features. In: European Conference on Ma- chine Learning (ECML). Springer, Berlin, pp. 137–142. [11] Kupietz, M., Belica, C., Keibe, H., Witt, A., 2010. The german ref- erence corpus dereko: A primordial sample for linguistic research in: Calzolari, nicoletta et al. (eds.). In: Proceedings of the 7th conference on International Language Resources and Evaluation (LREC 2010). pp. 1848–1854. [12] Lewis, D. D., Yang, Y., Rose, T. G., Li, F., December 2004. Rcv1: A new benchmark collection for text categorization research. J. Mach. Learn. Res. 5, 361–397. URL http://dl.acm.org/citation.cfm?id=1005332.1005345 35 [13] Logsdon, A., 2011. Learning disabilities. URL http://www.learningdisabilities.about.com/ [14] Manning, C. D., Raghavan, P., Schtze, H., 2008. Introduction to Infor- mation Retrieval. Cambridge University Press, New York, NY, USA. [15] Mergenthaler, E., 1996. Emotion-abstraction patterns in verbatim pro- tocols: A new way of describing psychotherapeutic processes. Journal of Consulting and Clinical Psychology 6 (64). [16] Miao, Y.-Q., Kamel, M., January 2011. Pairwise optimized rocchio al- gorithm for text categorization. Pattern Recogn. Lett. 32, 375–382. URL http://dx.doi.org/10.1016/j.patrec.2010.09.018 [17] Moschitti, A., 2003. A study on optimal parameter tuning for rocchio text classifier. In: Proceedings of the 25th European conference on IR research. ECIR’03. Springer-Verlag, Berlin, Heidelberg, pp. 420–435. URL http://dl.acm.org/citation.cfm?id=1757788.1757828 [18] Olsen, S., 2000. Ein internationales handbuch zur flexion und wortbil- dung. In: Booij, G., Lehmann, C., Mugdan, J. (Eds.), Morphologie. Berlin / New York: de Gruyter, pp. 897–916. [19] Özgür, A., Özgür, L., Güngör, T., 2005. Text categorization with class- based and corpus-based keyword selection. In: ISCIS. pp. 606–615. [20] Rocchio, J., 1971. Relevance Feedback in Information Retrieval. Prentice-Hall Inc., Ch. 14, pp. 313–323. [21] Salton, G., 1989. Automatic text processing: the transformation, analy- sis, and retrieval of information by computer. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA. [22] Sebastiani, F., March 2002. Machine learning in automated text catego- rization. ACM Comput. Surv. 34, 1–47. URL http://doi.acm.org/10.1145/505282.505283 [23] Solka, J., Jul 2008. Text data mining: Theory and meth- ods. Statistics Surveys 2008, Vol. 2, 94-112Comments: Pub- lished in at http://dx.doi.org/10.1214/07-SS016 the Statistics Surveys (http://www.i-journals.org/ss/) by the Institute of Mathematical Statis- tics (http://www.imstat.org). 36 [24] Tan, S., May 2005. Neighbor-weighted k-nearest neighbor for unbalanced text corpus. Expert Syst. Appl. 28, 667–671. URL http://dx.doi.org/10.1016/j.eswa.2004.12.023 [25] Toussaint, G. T., 2005. Geometric proximity graphs for improving near- est neighbor methods in instance-based learning and data mining. Int. J. Comput. Geometry Appl. 15 (2), 101–150. [26] Vinciarelli, A., October 2005. Application of information retrieval tech- niques to single writer documents. Pattern Recogn. Lett. 26, 2262–2271. URL http://dx.doi.org/10.1016/j.patrec.2005.03.036 [27] Wang, Y., Wang, Z.-O., aug. 2007. A fast knn algorithm for text cat- egorization. In: Machine Learning and Cybernetics, 2007 International Conference on. Vol. 6. pp. 3436 –3441. [28] Wechsler, D., 1939. The Measurement of Adult Intelligence. Baltimore (MD): Williams & Witkins. [29] Wechsler, D., 1982. Handanweisung zum Hamburg-Wechsler- Intelligenztest fuer Erwachsene (HAWIE). Separatdr., Bern; Stuttgart; Wien, Huber. [30] Yang, Y., Liu, X., 1999. A re-examination of text categorization meth- ods. In: Proceedings of the 22nd annual international ACM SIGIR con- ference on Research and development in information retrieval. SIGIR ’99. ACM, New York, NY, USA, pp. 42–49. URL http://doi.acm.org/10.1145/312624.312647 [31] Yang, Y., Pedersen, J. O., 1997. A comparative study on feature se- lection in text categorization. In: Fisher, D. H. (Ed.), Proceedings of ICML-97, 14th International Conference on Machine Learning. Morgan Kaufmann Publishers, San Francisco, US, Nashville, US, pp. 412–420. URL citeseer.nj.nec.com/yang97comparative.html [32] Zablotskaya, K., Walter, S., Minker, W., May 2010. Speech data cor- pus for verbal intelligence estimation. In: Proceedings of LREC’10. pp. 1077–1080. 37 
work_wzlqsnqdabbuvnv6fxugpar3we	----	Special issue on “advances in visual analytics and mining visual data” E D I T O R I A L Special issue on “advances in visual analytics and mining visual data” Visual and multimedia analytics provides an emerging field of research combining strengths from information analytics, geospatial analytics, scien- tific analytics, statistical analytics, knowledge discovery, data management and knowledge representation, presentation, production and dissemi- nation, cognition, perception, and interaction (Chen, Chiang and Storey, 2012). The aim is to gain insight into homogeneous, contradictory, and incomplete data through the combination of automatic analysis methods with human background knowledge and intuition. While the scope of visual analytics is broad, one principle that has emerged over the years is the need for visual analytics systems to leverage computational methods in data mining, knowledge discovery, and machine learning for large-scale data analysis. In these systems, the human operator works alongside the computational processes in an integrated fashion. Therefore, computing systems or services can sift through large amounts of data and identify the relevant information, while the human interactively explores the reduced data space to discover trends and pat- terns and make informed decisions. These two components operate in coordination, allowing for a continuous and cooperative analytical loop (Cybulski et al., 2015; Valdez et al., 2016). Top papers from Data 2018, Madrid, Spain and the best paper from FEMIB 2019 in Crete, Greece, have been invited. Through a robust and competitive review process, six papers have been selected. The summary of their contributions is as follows. Arhipova et al. (2020) propose a data aggregation approach for mobile phones, which was conducted at 15-min intervals in the area of each cellular base station. The case study examines all of Latvia's municipalities, analysing the economic activity level in each municipality in comparison to the mobile phone activity in three periods: 2015–2016, 2017, and 2018. The authors concluded that economic activity in municipalities could be estimated, and the positive dynamics of regional development have been detected. Such data and the data analytics method, which provides an understanding of how economic activities evolve in real-time in particular to locations and economic activity centres, can improve regional development planning and plan implementation. In order to assess which are the centres of economic activity in each municipality and its sphere of influence, the patterns of human commuting and fluctuations of internal activity on workdays and weekends/holidays in 2017–2018 were determined. In general, there is a shortage of reliable data on human commuting within Latvia and its specific regions; therefore, the method described here provides a practical tool for regional governments to keep track of strategy implementation and for strategic gap analysis. Khamayseh et al. (2019) investigate the issue of friends' management in the Social Internet of Things, and proposes a framework to manage friends' requests. The proposed framework consists of friend selection, friendship removal, and update-modules. It develops a weight based and Naïve Bayes Classifier based algorithms for the selection component. Moreover, a random service allocation model is proposed to construct a service-specific network model. This model is then used in the simulation setup to examine the performance of different friends' management algorithms. The performance of the proposed framework is evaluated using simulation under different scenarios. The obtained simulation results show improvement over other strategies in terms of the average degree of connections, average path length, local cluster coefficients, and Throughput. Lara et al. (2019) propose an outlier detection method based on a clustering process. Their aim is to overcome the specificity of many existing outlier detection techniques that fail to take into account the inherent dispersion of domain objects. The outlier detection method is based on four criteria designed to represent how human beings (experts in each domain) visually identify outliers within a set of objects after analysing the clusters. This has an advantage over other clustering-based outlier detection techniques that are founded on purely numerical analysis of clusters. To validate the proposed method, they studied method outlier detection and efficiency in terms of runtime. The results of regression analyses confirm that their proposal is useful for detecting outlier data in different domains, with a false positive rate of less than 2% and reliability greater than 99%. Park et al. (2019) demonstrate an online principal component analysis methodology based on online eigenvector transformation with the moving average of the data stream that can reflect concept drift. They compared the network intrusion detection performance based on an online transformation of eigenvectors with that of offline methods by applying three machine learning algorithms. Both online and offline methods dem- onstrated excellent performance in terms of precision. However, in terms of the recall ratio, the performance of the proposed methodology with integrated online eigenvector transformation was better; thus, the F1-measure also indicated better performance. The visualization of the princi- pal component score shows the effectiveness of their method. Hawashin et al. (2019) propose a new efficient hybrid similarity measure for recommender systems based on a combination of the user interest-user interest similarity measure and the user interest-item similarity measure. This hybrid similarity measure improves the existing work in three aspects. First, it improves the current recommender systems by using actual user interests. Second, it provides a comprehensive Received: 22 May 2020 Accepted: 12 June 2020 DOI: 10.1111/exsy.12607 Expert Systems. 2020;e12607. wileyonlinelibrary.com/journal/exsy © 2020 John Wiley & Sons, Ltd 1 of 2 https://doi.org/10.1111/exsy.12607 http://wileyonlinelibrary.com/journal/exsy https://doi.org/10.1111/exsy.12607 http://crossmark.crossref.org/dialog/?doi=10.1111%2Fexsy.12607&domain=pdf&date_stamp=2020-06-29 evaluation of an efficient solution to the cold start problem. Third, this method works well even when no co-rated items exist between two users. They demonstrate that their proposal is efficient in terms of accuracy, execution time, and applicability. Their proposed similarity measure achieves a mean absolute error (MAE) as low as 0.42, with 64% applicability and execution time as low as 0.03 seconds, while the existing similar- ity measures from the literature achieve an MAE of 0.88 at their best. These results demonstrate the superiority of their proposed similarity mea- sure in terms of accuracy, as well as having a high applicability percentage and a very short execution time. Vangipuram et al. (2020) propose (a) a novel imputation technique for imputation of missing data values; (b) a classifier based on feature transformation to perform classification and (c) imputation measure for similarity computation between any two instances that can also be used as the similarity measure. The performance of the proposed classifier is studied by using imputed datasets obtained through applying Kmeans, F-Kmeans and proposed imputation methods. Experiments are also conducted by applying existing and proposed classifiers on the imputed dataset obtained using the proposed imputation technique. For experiments, authors have used an open-source dataset named distributed smart space orchestration system publicly available from Kaggle. Experimental results are also validated using the Wilcoxon non-parametric statistical test. The performance of their proposed approach is better when compared to existing classifiers when the imputation process is performed using F-Kmeans and K-Means imputation techniques. It is also observed that accuracies for attack classes scan, malicious operation, denial of service, spying, data type probing, wrong setup are 100% while it is 99% for malicious control attack class when both the proposed imputation and classification technique are applied. We thank the Expert Systems Office and Editor-in-Chief for their support and kind assistance in the completion of this special issue. Victor Chang1 Shadi A. Aljawarneh2 Chung-Sheng Li3 1School of Computing, Engineering and Digital Technologies, Teesside University, Middlesbrough, UK 2Jordan University of Science and Technology, Irbid, Jordan 3PwC Labs, San Jose Correspondence Victor Chang, School of Computing, Engineering and Digital Technologies, Teesside University, Middlesbrough, UK. Email: ic.victor.chang@gmail.com ORCID Victor Chang https://orcid.org/0000-0002-8012-5852 Shadi A. Aljawarneh https://orcid.org/0000-0001-5748-4921 Chung-Sheng Li https://orcid.org/0000-0001-9072-1140 REFERENCES Arhipova, I., Berzins, G., Brekis, E., Binde, J., Opmanis, M., Erglis, A., & Ansonska, E. (2020). Mobile phone data statistics as a dynamic proxy indicator in assessing regional economic activity and human commuting patterns. Expert Systems, e12530. https://doi.org/10.1111/exsy.12530 Chen, H., Chiang, R. H., & Storey, V. C. (2012). Business intelligence and analytics: From big data to big impact. MIS Quarterly, 36, 1165–1188. Cybulski, J. L., Keller, S., Nguyen, L., & Saundage, D. (2015). Creative problem solving in digital space using visual analytics. Computers in Human Behavior, 42, 20–35. Hawashin, B., Lafi, M., Kanan, T., & Mansour, A. (2019). An efficient hybrid similarity measure based on user interests for recommender systems. Expert Sys- tems, e12471. https://doi.org/10.1111/exsy.12471 Khamayseh, Y., Mardini, W., Atwood, J. W., & Aldwairi, M. (2019). Dynamic framework to mining internet of things for multimedia services. Expert Systems, e12404. https://doi.org/10.1111/exsy.12404 Lara, J. A., Lizcano, D., Rampérez, V., & Soriano, J. (2019). A method for outlier detection based on cluster analysis and visual expert criteria. Expert Systems, e12473. https://doi.org/10.1111/exsy.12473 Park, S., Seo, S., Jeong, C., & Kim, J. (2019). Online eigenvector transformation reflecting concept drift for improving network intrusion detection. Expert Systems, e12477. https://doi.org/10.1111/exsy.12477 Valdez, A. C., Ziefle, M., Verbert, K., Felfernig, A., & Holzinger, A. (2016). Recommender systems for health informatics: State-of-the-art and future perspec- tives. In H. Andreas (Ed.) Machine learning for health informatics (pp. 391–414). Cham: Springer. Vangipuram, R., Gunupudi, R. K., Puligadda, V. K., & Vinjamuri, J. (2020). A machine learning approach for imputation and anomaly detection in IoT environ- ment. Expert Systems, e12556. https://doi.org/10.1111/exsy.12556 2 of 2 EDITORIAL https://orcid.org/0000-0002-8012-5852 https://orcid.org/0000-0001-5748-4921 https://orcid.org/0000-0001-9072-1140 mailto:ic.victor.chang@gmail.com https://orcid.org/0000-0002-8012-5852 https://orcid.org/0000-0002-8012-5852 https://orcid.org/0000-0001-5748-4921 https://orcid.org/0000-0001-5748-4921 https://orcid.org/0000-0001-9072-1140 https://orcid.org/0000-0001-9072-1140 https://doi.org/10.1111/exsy.12530 https://doi.org/10.1111/exsy.12471 https://doi.org/10.1111/exsy.12404 https://doi.org/10.1111/exsy.12473 https://doi.org/10.1111/exsy.12477 https://doi.org/10.1111/exsy.12556 Special issue on ``advances in visual analytics and mining visual data´´ REFERENCES 
work_x2s3rlndcrbyzivcj3ir4ve5iq	----	Advances and trends for the development of ambient-assisted living platforms Article DOI: 10.1111/exsy.12163 Advances and trends for the development of ambient-assisted living platforms Angelo Costa,1* Vicente Julián2 and Paulo Novais2 (1) Centro ALGORITMI/Departamento de Informática, Escola de Engenharia, Universidade do Minho, Portugal E-mail: acosta@di.uminho.pt (2) Departamento de Sistemas Informáticos y Computación, Universitat Politècnica de València, Camino de Vera s/n46022 Valencia, Spain Abstract: Ambient Assisted Living (AAL) and Ambient Intelligence (AmI) try to achieve a future where technology surrounds the users and helps them in their daily lives. In this sense, the urgent need of solutions to cover the rapid increase of the elderly population with chronic diseases led to the increase of projects related with AAL and AmI. During the latest years, several projects have been proposed to tackle different medical problems, some building devices and others services. This paper presents iGenda and its evolution, the UserAccess, with the main objective of developing an AAL platform. It features an analysis of the latest developments and points future directions for the work. These projects display the importance of the interoperability of the platforms, demonstrating a case study for AAL development. Keywords: Ambient Assisted Living, Ambient Intelligence, Intelligent Envi-ronments, AAL4ALL, e-Health, Active Ageing, Artificial Intelligence 1. Introduction The European Union (EU) defined the year of 2012 as the year of active ageing,1 where social efforts and funding were directed to solutions that aimed to solve different age- related problems. This urgent need was uncovered by recent United Nations reports about population evolution (United Nations, 2009; United Nations, Department of Economic and Social Affairs, P.D., 2015; Department of Economic and Social Affairs, P.D., 2013). These reports showed that the elderly population is rapidly increasing and the younger population (as in pre-teenagers) is rapidly decreasing, and showing that the population over 60 years old will double by 2050 as Figure 1 shows. Another ageing problem is the declining fertility. The human being longevity is determined by several factors, like the quality of medical care, economic stability and social care. The question that remains is what is the most important factor to this increase in longevity? The Department of Economic and Social Affairs, PD (2013) shows that in developing countries, the elderly population numbers are increasing even if they lack the medical, social and economic resources. The only factor that is commonly shared between the countries on the United Nations report is the decrease of the fertility values. Thus, as seen in Figure 2, the population distribution becomes slimmer at the bottom (in the visual form of a pyramid) and balanced at the middle and top. The panel for Health Ageing and Retirement in Europe conducted a survey that stated that at least two chronic diseases were found on 40% of people aged over 50 years (Adena & Myck, 2014; Mair et al., 2015; Boffetta et al., 2014), evidencing that it is much cheaper to invest in prevention and proactivity, acting as soon as possible to keep the population from having serious medical conditions. To overcome some situations where help or assistance was scarce, some home environments were equipped with sensor systems to respond quickly to critical situations, such as falls where one may not be able to call for help. The aim of this paper is to present two projects, the iGenda and the UserAccess (UA), and make a parallel between the current ambient-assisted living (AAL) developments with the aim of foreseeing the roadmap of this kind of systems. We also present the Care4Balance, understAID and RelaxedCare projects, which are very similar to the iGenda or the UA, in terms of architecture or concept. The analysis of these platforms highlights the positive and negative features of each one, their similarities and how they can be an advantage for the future developments. The paper is structured as follows: Section 2 introduces the concepts of Ambient Intelligence (AmI) and AAL and an analysis of the selected projects; Section 3 presents the iGenda and Section 4 the UA. Finally, Section 5 shows the conclusions and future paths. 1http://ec.europa.eu/archives/ey2012/ © 2016 Wiley Publishing Ltd Expert Systems, xxxx 2016, Vol. 00, No. 00 http://ec.europa.eu/archives/ey2012/ 2. Related work This section presents the related work. First is the AmI and the AAL areas. They show the social impact these projects have. Second is the most relevant related projects that represent other efforts that are currently being made. The differences of their concepts and architectures in relation to iGenda and UA are identified and exposed. 2.1. Ambient Intelligence and Ambient Assisted Living Ambient intelligence responds to the technological call for monitoring and acting on the home of people with disabilities. Its aim is to create a nest of sensor systems that together could provide contextual information, thus transforming data into knowledge (Aarts & Encarnação, 2006; Aarts et al., 2001; Cook et al., 2009). This is achieved by creating digital environments that are sensitive to people’s needs and can respond to their requirements, anticipate behaviours and adjust the response accordingly. In the last years, several projects have been developed to attend to the needs of AmI. The advances of the Internet of things have opened new hypotheses of integrating home appliances in home networks where they can be remotely controlled (Kranz et al., 2010; Gershenfeld et al., 2004; Atzori et al., 2010). Ambient Assisted Living has spawned as a subset of AmI with the goal of creating proactive environments for the elderly and people with disabilities. This separation occurred to convey the attention to the limitations and requirements of these people, mainly concentrating the efforts on personalization. AAL differs from AmI because AmI requires that physical environments adapt to every person and AAL requires that physical environments adapt Figure 2: Population pyramids of the less and more developed regions of 2013 and 2050 (Department of Economic and Social Affairs, P.D., 2013). Figure 1: Distribution of population aged 60 or over by regions, 2000–2050 (Department of Economic and Social Affairs, P.D., 2013). © 2016 Wiley Publishing LtdExpert Systems, xxxx 2016, Vol. 00, No. 00 to a specific person (Vastenburg et al., 2008; Ramos et al., 2008; Cook et al., 2009). One of the great investors in this idea was the EU, distributing several funds to the research of AAL projects. For instance, one of the funded is the Ambient Assisted Living Joint Programme, which is a consortium of countries that promote and invest in AAL projects. These actions promote the novelty factor and compel the flourishing of ideas that later may be translated into devices or services. 2.2. Related projects In the multitude of AAL projects, there are some important projects that need to be referred in order to show the different existing perspectives. AAL is a broad area of application, which means that even though most of the projects are related, they tend to differ in the aim, architecture or concept. The following projects show different facets of AAL developments and follow closely the iGenda and UA concept. 2.2.1. Care4Balance The Care4Balance (Care4Balance, 2015) project consists in an aggregator of information directed to the care receivers and the formal or informal caregiver. The goal is to connect every user with their circle of trust, sharing information about everyday events with the additional feature of detecting problematic situations that may occur. The architecture is defined in a client–server style, where the server maintains all information and ensures the data flow between each client. The Care4Balance client is expected to be operating in a mobile device, but the authors are not very clear about its features or if it will be the only method of access. This project is still active and is now in the final stages of field trial. The main issue with this project is the way it approaches the users in terms of visual interfaces. The coloured cube is not very informative and may be forgotten by people with cognitive disabilities. In search of minimal interaction, they have minimized too much. Additionally, they lack interoperability features. 2.2.2. UnderstAID The understAID (UnderstAid, 2015) is an EU-funded project that focuses on the development of an e-leaning application that provides information about medical conditions directed to the dementia disease. The authors claim that there is an absence of an effective search method about medical information and that the information that exists is too technical to non-medical people. This project is designed for informal caregivers by providing them with a comprehensive library of procedures and general information about degenerative diseases. The personalization is performed through surveys and profiling the care receiver, generating a list of correct procedures to each user. This project is currently at its final stage, having a beta application publicly available for testing. This project is only directed to the caregivers, which leaves the care receivers without any help. Furthermore, the level of interaction is very limited. 2.2.3. RelaxedCare The RelaxedCare (RelaxedCare, 2015) project has as a goal to establish a social connection between the elderly and its relatives and create a strong social bond. To achieve this goal, it uses the architecture of the HOMER project (Fuxreiter et al., 2010; Fuxreiter et al., 2011; HOMER, 2015) to convey the users’ information benefitting from their concept of ‘AAL in a box’ (off-the-shelf products that can be bought directly from any vendor). The platform receives messages from the relatives and notifies the care receivers through actuators, noticeably, performed through a coloured cube. Furthermore, the caregivers can access a sensor platform to view the current status of the care receivers, now represented by a tracker for lost objects, such as keys or glasses. The latest development news is that the project is being tested in a controlled environment. This project is limited in terms of real help. The social network provides inclusion and social cohesion, but it is unclear that it has an impact on the users’ lives. 3. iGenda The iGenda (Costa & Novais, 2011; Costa et al., 2011; Costa et al., 2010a) has as main goal to provide intelligent event management. It consists in a platform that receives events from its users and tries to schedule them according to their importance, having the ability of creating, moving and deleting events. Moreover, it supports shared events and can schedule ludic activities in the users’ free time adjusted to their medical condition. This last feature was developed to promote active ageing. The iGenda spawned from the VirtualECare project (Costa et al., 2010b), with the initial intention of generating a visual interface to the platform services. The verified performance and the good reception from the scientific community made clear that the iGenda should be an independent project. For instance, it was able to receive information and process it according the user profile and the event importance. The following scenario was possible: upon receiving a high priority event from a physician, the system reallocated any conflicting event to other time slot, and the high-priority event was scheduled in the specified time. 3.1. Architecture In terms of architecture, the iGenda followed a typical client–server structure, where the client acts as the visual gateway to the platform, leaving all the heavy processes to the server. The client was designed initially as a desktop application, due to the low proliferation of mobile devices at the beginning of the project. The last update of the client © 2016 Wiley Publishing Ltd Expert Systems, xxxx 2016, Vol. 00, No. 00 introduced a web interface, a modular endpoint to both desktop and mobile platforms. The client has two main features: visualizing the events as a list of activities and the creation of shared events with other users. The platform worked with the known ISO ical (RFC 5545 and updates RFC 5546 and RFC 6868) facilitating the use of external applications for visualization of such files. This decision was made, so it reduces the users’ interactions, keeping the act of using the iGenda simple, which means increasing the possibilities of its success. Figure 3 shows an overview of the iGenda architecture. The server is constituted by three major modules: the agenda manager (AM), the conflicts manager (CM) and the free time manager (FTM). These modules and the platform in general are constituted by multi-agent systems (MAS) (Camarinha-Matos & Afsarmanesh, 2004; Wooldridge, 1997) established in Java Agent DEvelopment Framework (JADE) (Gawinecki & Frackowiak, 2008; Bellifemine et al., 2007). The MAS allow the creation of independent modules and plug-and-play features, each one bounded to their own agent. MAS make deployment and maintenance easier as they are independent and reliable, bypassing any agent that is causing errors and rebooting it. Furthermore, high-level agents can be used to monitor the agents, automating maintenance tasks in case of errors. The architecture is hierarchical; there is only one AM per server (an array of servers is recommended for load balancing) and a pair of CM and FTM for each user. This is a platform with agents-inside-agents, in which each agent contains other internal agents that process the information according to their role. This type of architecture acts using strict ontologies, exchanging human-readable messages when possible. In terms of the modules, their tasks are as follows: • AM: it is in charge of receiving the information of each client (typically a new event) and relaying it to the appropriate CM module. It also responds to each client in case of important actions occurred (such as critical remainders or removal of an event). Furthermore, it is responsible for verifying the integrity of all CM and FTM available and stopping and restarting them if needed. • CM: it is divided in three processes, fetching, reallocating and saving. The fetching consists in allocating the user calendar and search for the week that the event is referring to, locating the events that collide with it. The reallocating calculates a balanced priority with the importance value of the event and the user importance. The most important events stay static, and the least important ones are moved to other timeframe or, in extreme cases, deleted. Lastly, the saving processes the changes and uploads them to the user calendar and notifies the AM, and the AM too notifies the user(s) if that is the case. • FTM: this module operates silently, and its task is to schedule playful events on the user’s calendar to promote an active life. It fetches the events on the user-defined database that has activities that the user likes or requires and schedules activities randomly but being sensible to the daytime hours (this avoids that certain tasks, like going outside, occur during the night-time). The modularity and operation as a service allow the use of the iGenda as a base platform to other systems. For instance, a localization platform (Ramos et al., 2012; Ramos et al., 2014) uses the information of the events and the client’s interface to deliver information to the user. The iGenda project has led to different advances, namely, supporting systems that take advantage of its features, and a direct evolution of the iGenda: the UA. The iGenda project served as the base structure for the UA project. The UA project is different in terms of its aim, but the core architecture relies on the iGenda. The UA is explained with more detail in the next section. 4. UserAccess The aim of UA is to provide intelligible information directed to the caregivers about the care receiver (Costa et al., 2014). The UA is a natural evolution of the iGenda, taking advantage of its architecture and communication methods. The UA was developed for the AAL4ALL project (AAL4ALL, 2015), a Portuguese consortium of private and public institutions, with the goal to provide easy-to- use, plug-and-play AAL solutions (either software or hardware) directed to the elderly community. Some of the developments were house-sensor platforms, body sensors and transparent services that collect, process and show information. The main issue while working with elderly is that they have several limitations, requiring a social conduct designed especially for them, as most of the elderly population are not prone to work with advanced technological devices. Figure 3: iGenda architecture overview. © 2016 Wiley Publishing LtdExpert Systems, xxxx 2016, Vol. 00, No. 00 Furthermore, there are social aspects that have to be specially considered, like the health concerns that must be first announced to the physician or the caregivers before (if at all) it is showed to the care receiver. The field tests performed for the AAL4ALL suggested that some information would be more effective if it were delivered and explained by a person that the elderly trusts rather than by an application. One of the issues is that the available information services directed at the caregivers are cumbersome and difficult to use, being most of them only internally available and lacking remote access. This results in low acceptance by the users, which only use the service when forced, for example, a scenario that is common in nursing homes mostly is to keep track of the patients’ personal health record. The UA works under the consideration that it is required that the caregivers maintain their assigned care receiver under constant monitoring and have a limit to the number of monitored people. The lack of support for caregivers sets the tone for the development of the UA, which was designed to show information that is understandable, easy to read and is adjusted to each user. For instance, the information delivered to a formal caregiver can be medical oriented and include sensitive information; an informal caregiver may only receive overall notifications about the care receiver status and warnings from the home sensors (if a window is open and the user is not at home). Moreover, the care receivers can use this application to verify if there are any warnings or to check their personal health record. The UA accepts heterogeneous communication through the use of marked-up message content. It has predefined ontologies and can receive communication from external services that wish to use the inference engine or the visual interfaces. It has also the ability to schedule events using the iGenda platform. 4.1. Architecture The UA focuses on the interfaces, and it was developed natively for the Android platform and for the web. The webpage has more information and serves as the administration platform. The user can access the administration page to enable/disable sensor platforms and personalize the warnings messages. The mobile version is designed to be simple and to present a condensed version of the information, concentrating in the warnings. Figure 4 shows the Android application, with the main window of information and the widget, which is a quick view of the current warnings. Figure 5 shows the administration page of the caregiver view where the information is displayed and can be configured. Also, both visual interfaces can manage the information of several users, allowing one platform to provide information of several users. The architecture follows a client–server structure and uses MAS on the server. The MAS serve not only for security purposes but also to easily deploy new features. For instance, if a new sensor platform is added to the environment, like cameras, a data parser was provided to the UA to interpret the input and generate the UA compatible output. The MAS enforce the idea of distributed computing, allowing the distribution of each agent in a different computer. Following the base requirements of the AAL4ALL, the UA structure overview is showed in Figure 6. In the care- receiver block, the body sensors and home sensors are directly connected to the AAL4ALL, apart from being required by the project. This ensures the interoperability of the system. Any vendor has to assure that their products operate according to the AAL4ALL specifications. The AAL4ALL enforces certification to integrate a new device; thus, each device has to have the appropriate data nodes, so any other service can access to that information. The UI device connects directly to the RabbitMQ (RabbitMQ, Figure 4: UserAccess mobile interfaces. Left: user selection; middle: selected user warnings and personal details; and right: warnings widget. © 2016 Wiley Publishing Ltd Expert Systems, xxxx 2016, Vol. 00, No. 00 2015) and establishes the communication channels through it, being the same scenario applied to the caregiver. The external services can use the RabbitMQ or REST, while REST is unadvised for communication, it is maintained for legacy purposes, like the connection with the iGenda. The REST service is the frontend to an agent running in the MAS. This agent consumes the requests and sorts them to the agent in charge of the user the request is directed to. The REST interface is developed using the SPARK framework that is scalable and is able to process several requests at the same time. Furthermore, there are several REST endpoints in each load-balancing server. The internal data processing of the UA consists of three modules that internally are agent systems; this facilitates the maintenance and the communication process. In more detail, these modules are the following: • Message centre: it controls the in and out communication flow of the application, working directly under the communication structures. For each sensor system and user, an agent is required to descramble the communication and adjust it for internal consumption. It also translates the message to be intelligible to be presented on UI devices. Unlike the AM from the iGenda, the message centre does not sort the messages for the correct agents, which is performed by the communication structures. • Profiler: it helps the message centre to adjust the received content to meet the users’ requirements. For each user, the definitions are adjusted initially through the administration webpage. Thereon, the preferences are learned from the user interactions with the mobile client, and the profile is updated according the dismissal of notifications. For instance, if a warning about an open window is repeatedly dismissed without action, then the rule for this warning is demoted, and its priority is diminished. This process is performed offline. Moreover, the profiler keeps the configurations to each user and its realm of connections for fast access in case of critical situation. Figure 5: UserAccess administration website. Figure 6: UserAccess structure overview. © 2016 Wiley Publishing LtdExpert Systems, xxxx 2016, Vol. 00, No. 00 • Data processor: it processes the parsed information the message centre sends and builds datasets of information. This information stands as history of events and transactions, with the additional feature of translating data into human-readable sentences. Moreover, it serves as an interoperable database for external services and a message buffer to unconnected devices. The UA was expected to become one of the main tool for services developed in the AAL4ALL and in other platforms because of the interoperability features. The adaptability and attention to the users make the UA an indispensable tool to the caregivers. 4.1.1. UserAccess and iGenda At this point, it is clear that there are many similarities and some differences between the UA and the iGenda. Although the area of operation is the same, the goals and architecture are different at their core. A comparison is presented in Table 1 to highlight the features of the two projects. 4.2. Tests The number of tests performed to the platform was limited because of the ending of the AAL4ALL project; thus, testing with real users was not possible. Despite of that, the tests performed internally demonstrated very positive results. To test the UA, a sensor platform developed at the University of Minho specifically designed for the AAL4ALL project was used. The sensor platform consisted of three parts: the sensors, the base receiver and the web service. The sensors, depicted in Figure 7, are the following: • Luminance: it detects the amount of light present, with the precision needed to distinguish artificial light and sun. • Motion: it is a basic sensor only activated when a fairly amount of movement is performed. • Door: it is a magnetic sensor that changed status when the door was open/closed. • Humidity and temperature: it detects the amount of humidity and overall temperature on a home environment. • Gas and water: it detects if there is flow of gas or water on the pipes, that is, they have to be installed in the plumbing system. • Button: it is pressed if there is a panic situation for the user. • Electric plug: it is a sensor/actuator able to detect if there is current being consumed and flip the on/off status by remote command. The sensor system used ZigBee protocol for communication with the base receiver, which was favoured over Bluetooth for its flexibility, number of nodes and energetic efficiency. The base receiver then transported the raw data through a parser, and the outcome was a higher level JSON-formatted message, following the specifications of the AAL4ALL communication protocols. Furthermore, the web service also aimed to a low throughput of data as the AAL4ALL receiving nodes had already a large amount of data passing through; thus, the data sent were the mean Table 1: Comparison of the UserAccess and iGenda features Area iGenda UserAccess Main users Care receivers Caregivers Architecture Multi-agent system Multi-agent system Communication JADE FIPA JADE FIPA and REST Interfaces Directed to care receivers Directed to caregivers Platform Android and web Android and web Performance 10 s average (from received event to scheduled) 3 s average (from sensor trigger to notification) Objectives Intelligent scheduler AAL4ALL caregiver interface Figure 7: Some of the sensors used in testing (pre-production version). © 2016 Wiley Publishing Ltd Expert Systems, xxxx 2016, Vol. 00, No. 00 of the last 5 s of data collected, excepting the case of the button press, which was relayed immediately. This sensor platform provided the information displayed in Figures 4 and 5, thus proving the feasibility of the UA. The communication was successful, as well as the information parsing and display. The version of the UA tested was the one directed to the formal caregivers, seen in Figures 4 and 5. The ‘panic button’ tested positive, and the message was presented in the screen, having 2–5s of time discrepancy, which we consider excellent as initial results. Additionally, the electric plug was successfully controlled in the role of disconnecting the light if it was forgotten that is was turned on. In this case, due to the sensor limitations, we could not detect if the light is currently turned on or off when using efficient bulbs, for example, LED, due to the lack of sensitivity of the sensor. We have experimented in a controlled environment recurring to students to interpret the role of users. They interacted with the sensors that were placed in a laboratory. The placement was targeted to emulate a living room. Furthermore, an array of 25 agents simulating other users (in this case, directly simulating the generation of sensor values) was deployed to simulate a nursing home environment. The real sensor platform buffers the data and sends it in 7-s blocks. The emulated sensors followed this time requirement, and the actions they report are according a set of actions registered in previous tests. This decision was made to filter random actions that would not happen in real-life scenarios. The test environment was set up in the following way: the communications services were present at a dedicated machine, the UA platform was present in five machines (for load balancing) each with 7 assigned users and one with 4 users, and the databases were on a dedicated machine. Table 2 shows the response time values registered upon the execution. The mean response value corresponds to the time period between receiving the sensor platform data and the visual representation of a warning. The errors represent the number of tests that failed to be registered or shown on the interface. The loads correspond to the number of users active at the same time, from 1 (light) to 26 (heavy). The action was the following sequence: open the door, move inside, open water and gas, turn on a lamp exit in the room and press the panic button. Each sequence was performed four times by each user. 5. Future path and conclusions The natural evolution to the iGenda and the UserAcess is the Cognitive Life Assistant (CLA) platform. The CLA joins the strongest features of both referred projects and includes intelligent social computing (Sheth et al., 2013; Wang et al., 2007) and the human behaviour emulation (Gomes et al., 2014; Pimenta et al., 2015; Carneiro et al., 2013). As the name suggests, the CLA is directed to people with cognitive limitations, and one of its goals is to make decisions on daily events in the best interest of the user. Another goal is to improve the way that profiles are constructed and introducing real-time processing in the profiler. Furthermore, ‘influence’ will be introduced; each user action will influence the users on his or her social circle by changing their preferences. The CLA will be established on three actions: behaviour analysis, intelligent scheduling and emotion analysis. The architecture will be similar to MAS but with a lighter footprint, because previous developments could scale correctly but suffered from a large consumption of hardware. Moreover, some actions will be implemented directly on mobile devices, as they currently possess the hardware needed for communication processing. Although the iGenda, UA and CLA do share the same line of investigation, their objective is very distinct from each other. However, their main advantage is their interoperability, which facilitates feature sharing. In our view, the main structural advance that the newest AAL projects can present is its interoperability. Interoperability ensures that all projects are usable and can benefit from other projects by interacting with them. For instance, a home environment could be equipped with different products and/or services, and one integrated solution would display and reason the data received or use other processes to communicate with the user. Moreover, in the case of the AAL solutions, it is imperative that the solutions are adapted to its users. Because most of the users are debilitated in some way, a universal solution is inappropriate as it could lack the flexibility or features needed to perform certain tasks. Personalization is a key factor to AAL platforms. The user’s environment must be envisioned to provide functionalities adjusted to their recipient. Also, the personalization must reach the services’ core adjusting the response in order to cope with certain actions that do not follow the general procedures. This is especially important to people with cognitive disabilities that may lack decision abilities. This paper presents one solution related to the area of the AAL, and the work developed by the Intelligent Systems Lab at the University of Minho. The scientific acceptance, prototypes and final applications and field tests have proven feasibility of this work. Furthermore, current-related projects follow closely some of the features implemented on the iGenda and the UA, thus using them as a steppingstone to developments either directly related to the work presented or to other research areas. Acknowledgements This work has been supported by FCT – Fundação para a Ciência eTecnologia within the Project Scope: UID/CEC/ Table 2: UserAccess mean response values Load Mean response value Errors Light load <1 s 0 Medium load 1.7 s 0 Medium–high load 3 s 2 Heavy load 8 s 22 © 2016 Wiley Publishing LtdExpert Systems, xxxx 2016, Vol. 00, No. 00 00319/2013 and COMPETE: POCI-01-0145-FEDER- 007043. A. Costa thanks the Fundação para a Ciência e a Tecnologia (FCT) the post-doc scholarship with the ref. SFRH/BPD/102696/2014. This work is also partially supported by the MINECO/FEDER TIN2015-65515-C4- 1-R. References AAL4ALL, 2015. AAL4ALL – ambient assisted living for all. Available at: www.aal4all.org. Aarts, E. and J.L. Encarnação (Eds) (2006) True Visions: The Emergence of Ambient Intelligence, Berlin: Springer. AARTS, E., R. HARWIG and M. SCHUURMANS (2001) Ambient intelligence. In The Invisible Future, New York: McGraw-Hill, 235–250. ADENA, M. and M. MYCK (2014) Poverty and transitions in health in later life, Social Science and Medicine, 116, 202–210. ATZORI, L., A. IERA and G. MORABITO (2010) The Internet of things: a survey, Computer Networks, 54, 2787–2805. BELLIFEMINE, F.L., G. CAIRE and D. GREENWOOD (2007) Developing Multi-agent Systems with JADE, New York: Wiley. BOFFETTA, P., et al. (2014) The Consortium on Health and Ageing: Network of Cohorts in Europe and the United States (CHANCES) project – design, population and data harmonization of a large-scale, international study, European Journal of Epidemiology, 29, 929–936. CAMARINHA-MATOS, L.M. and H. AFSARMANESH (2004) A multi- agent based infrastructure to support virtual communities in elderly care, Int J of Networking and Virtual Organisations, 2(3), 246–266. Care4Balance, 2015. http://www.aal-care4balance.eu/. CARNEIRO, D. et al., 2013. Enriching conflict resolution environments with the provision of context information. Expert Systems. COOK, D.J., J.C. AUGUSTO and V.R. JAKKULA (2009) Ambient intelligence: technologies, applications, and opportunities, Pervasive and Mobile Computing, 5(4), 277–298. COSTA, Â., et al. (2011) Increased performance and better patient attendance in an hospital with the use of smart agendas, Logic Journal of IGPL, 20(4), 689–698. COSTA, Â., et al. (2010a) Multi-agent personal memory assistant. In Demazeau, Y., et al. (editors), Trends in Practical Applications of Agents and Multiagent Systems. Advances in Intelligent and Soft Computing, Berlin, Springer Berlin Heidelberg, 97–104. COSTA, Â. and P. NOVAIS (2011) An intelligent multi-agent memory assistant. In Bos, L., et al. (editors), Handbook of Digital Homecare – Successes and Failures. Communications in Medical and Care Compunetics, Berlin: Springer, 197–221. COSTA, A., P. NOVAIS and R. SIMOES (2014) A caregiver support platform within the scope of an ambient assisted living ecosystem, Sensors, 14(3), 5654–5676. COSTA, R., et al. (2010b) User recognition in AAL environments. In Augusto, J.C., et al. (editors), Ambient Intelligence and Future Trends – International Symposium on Ambient Intelligence (ISAmI 2010). Advances in Soft Computing, Springer Berlin Heidelberg, Berlin, Heidelberg, 177–184. Department of Economic and Social Affairs, P.D, 2013. World Population Ageing 2013, FUXREITER, T. et al., 2010. A modular platform for event recognition in smart homes. In 12th IEEE International Conference on e-Health Networking, Application and Services, Healthcom 2010 Lecce, Italy. FUXREITER, T., et al. (2011) Rule-based indoor localization using non-intrusive domotic sensors and the HOMER platform. In Partnerships for Social Innovation in Europe. Proceedings of AAL Forum 2011, 569–576. GAWINECKI, M. and G. FRACKOWIAK (2008) Multi-agent systems with JADE: a guide with extensive study, IEEE Distributed Systems Online, 9(3), 4–4. GERSHENFELD, N., R. KRIKORIAN and D. COHEN (2004) The Internet of things, Scientific American, 291, 76–81. GOMES, M., et al. (2014) Studying the effects of stress on negotiation behavior, Cybernetics and Systems, 45, 279–291. HOMER, 2015. http://homer.aaloa.org/. Available at: http:// homer.aaloa.org/ KRANZ, M., P. HOLLEIS and A. SCHMIDT (2010) Embedded interaction: interacting with the Internet of things, IEEE Internet Computing, 14, 46–53. MAIR, C.A., QUINONES, A.R. & PASHA, M.A., 2015. Care preferences among middle-aged and older adults with chronic disease in Europe: individual health care needs and national health care infrastructure. The Gerontologist. PIMENTA, A., et al. (2015) Detection of distraction and fatigue in groups through the analysis of interaction patterns with computers. In Intelligent Distributed Computing VIII. Studies in Computational Intelligence, Cham, Switzerland: 29–39. RabbitMQ, 2015. http://www.rabbitmq.com/. RAMOS, C., J.C. AUGUSTO and D. SHAPIRO (2008) Ambient intelligence – the next step for artificial intelligence, IEEE Intelligent Systems, 23(2), 15–18. RAMOS, J., et al. (2014) Interactive guiding and localization platform, International Journal of Artificial Intelligence, 12(1), 63–78. RAMOS, J., et al. (2012) Orientation system for people with cognitive disabilities. In Novais, P., et al. (editors), Ambient Intelligence – Software and Applications. Advances in Intelligent and Soft Computing, Roorkee, Springer Berlin, Heidelberg, 43–50. RelaxedCare, 2015. http://www.relaxedcare.eu/en/. SHETH, A., P. ANANTHARAM and C. HENSON (2013) Physical-cyber- social computing: an early 21st century approach, IEEE Intelligent Systems, 28, 78–82. UnderstAid, 2015. http://understaid.com/. United Nations, 2009. World population ageing, Available at: http://www.un.org/esa/population/publications/WPA2007/ SummaryTables_new.pdf. United Nations, Department of Economic and Social Affairs, P.D, 2015. Population ageing and development database 2014. VASTENBURG, M.H., et al. (2008) Ambient intelligence. In Aarts, E., et al. (editors), European Conference on Ambient Intelligence, AmI ’08. Lecture Notes in Computer Science, Springer Berlin Heidelberg, Berlin, Heidelberg, 1–12. WANG, F., et al. (2007) Social computing : from social informatics, IEEE Intelligent Systems, 22, 79–83. WOOLDRIDGE, M. (1997) Agent-based software engineering. In Bradshaw, J. and G. Arnold (editors), IEE Proceedings Software Engineering 144, 26. The authors Angelo Costa Ângelo Costa has received his European Doctorate title at the University of Minho (Portugal) in 2013 on the biomedical area. He has also received his Master’s title in 2009 at the University of Minho and is a lecturer at the Polytechnic of Porto. He was the author and co-author of an extensive number of publications, in international journals with impact factor, international book chapters, and in conference proceedings with international peer review. He is currently in a post-doc developing the Cognitive Life Assistant, a partnership of the University of © 2016 Wiley Publishing Ltd Expert Systems, xxxx 2016, Vol. 00, No. 00 http://www.aal4all.org http://www.aal-care4balance.eu/ http://homer.aaloa.org http://homer.aaloa.org http://homer.aaloa.org http://www.rabbitmq.com http://www.relaxedcare.eu/en/ http://understaid.com/ http://www.un.org/esa/population/publications/WPA2007/SummaryTables_new.pdf http://www.un.org/esa/population/publications/WPA2007/SummaryTables_new.pdf Minho and the Universitat Politècnica de València, and one of the main developers of cognitive assistants for mobile platforms. Vicente Julián Vicent J. Julian Inglada received his PhD from the Universitat Politècnica de València (UPV) in 2002. He holds a position of associate professor of Computer Science at the Universitat Politècnica de València. He is a member of the ‘Grupo de Tecnología Informática-Inteligencia Artificial’, and deputy director of the Official Master in Artificial Intelligence, Pattern Recognition and Digital Imaging at the UPV. He has more than 40 works published in journals with outstanding positions, papers published in conference proceedings that have a system of external peer review and dissemination of knowledge comparable with journals indexed in relevant positions. Paulo Novais Paulo Novais is an associate professor with Habilitation of Computer Science at the Department of Informatics of the University of Minho (Portugal) and a researcher at the ALGORITMI Centre in which he is the coordinator of the research group Intelligent Systems Lab. From the same university, he received a PhD in Computer Science in 2003. He has led and participated in several research projects sponsored by Portuguese and European public and private institutions. He is the co-author of over 170 book chapters, journal papers, conference, and workshop papers and books. © 2016 Wiley Publishing LtdExpert Systems, xxxx 2016, Vol. 00, No. 00 
work_x3r4uylvtzbapeii5sla6h4bjm	----	Request unsuccessful. Incapsula incident ID: 76000190746958706-2141889307073053899 
work_x6d6mtrc7jhaze4yoyt45oc37y	----	., LA-uR--89-3534 DE90 002399 ‘i(/V~ [, i~~g TITLE The Development of an Expert System to Tune a Beam Line AUTHOR(S) D. E. Schultz DIS(’I..I4IMKR ‘,,,.,,,!! ,!, , ,,, ,. MASTER- “ About This Report This official electronic version was created by scanning the best available paper or microfiche copy of the original report at a 300 dpi resolution. Original color illustrations appear as black and white images. For additional information or comments, contact: Library Without Walls Project Los Alamos National Laboratory Research Library Los Alamos, NM 87544 Phone: (505)667-4448 E-mail: lwwp@lanl.gov ‘l’heDeve@ment of an Expert System to Tune a Beam Lin@—- —— ——. -—.———— -.—. —— —— David E. Schultz I,os Alamcs National Laboratory Los Alamos, New Mexico 87545 Pamela A. Brown TRIUMF Vancouver, BC V6T 2A3 Abstract The experience of developing an Expert System to””aid in the tuning of the Ion Source Injection beam line at TRIUMF is described. The challenging and complex task of introducing Expert System technology into an established accelerator operation is outlined. Success in this environment. depends strongly on the choice of project, the choice of experts, the choice of tools, and the methods used to represent tile expertise. All these choices are discussed. 1 THE ACCELERATOR ENVIRONMENT TRIUM? is a cyclotron which accelerates negatively–charged hydrogen ions to variable energies between 6Cl MeV and 520 MeV. 1.1 Ion Source Injection Beam Line The Ion Source Injection System (ISIS) beam line accepts 300KV tl- beam from the source and transports it a total of approximately 40 meters around two bendri to the centre of the main accelerator. The beam is transported by electrostatic focusing and steering devices. There are four groups of diagnostic devices in the ISIS beam line: collimators, skimmers, non-intercepting monitors, and wire scanners. The skimmers precede and protert each quad~-upole and help to centre the beam. In this project, thn collimators were the principal devices used; the skirnmerc were uGed lesb because they do not tolerate veL”y m~lch current and because they are leGs sensitive. ‘rhe non-intercepting monitors were used to calculate transmission. The Wi[C scanners w~rc not useu because there is little ope[akional expedience with usin(j * work fiupportec] hy the US [Jcp,~ltment of El~rrgy. w(~lk I)clf(lllnflll al l’RIUMF. 1 The Development of an Expert System to Tune a Beam Line them. 1.2 Manual ISIS Tune Procedure After the desired source is selected mechanically and logically, the electrostatic elements of the 1S1S beam line are restored to the values currespond~ng to the best known solution. The operator uses the currents Oc the beam stops and the spills on the skimmers and collimators to find the earliest significant spill. Next the operator determines whether the loss is caused by horizontal or vertical displacements, or if the beam is too large and hitting sumething. As a tuning contingency, the collimators are water-cooled and able to take the full ISIS beam. Close to a wire scanner, the operator uses the wire scanner to make the diagnosis. Elsewhere, he diddles (makes minor adjustments to) the setpoints of appropriate beam line or source devices. He uses the collimator and skimmer histogram patterns and the 1S1S transmission, to minimize beam spill in the ISIS beam line. .“ 2 THE OVERVIEW OF THE TASK A six month pilot project was setup to determine the feasibility, problems, and benefits of using Expert System technology to solve a problem in the beam line tuning domain. Several questions arise when approaching a new technology. Some of the more obvious are: 1. Can it solve the problem? 2. Is it appropriate for the task? 3. How much does it cost in time, money, and stress on the existing system? 4. What are the limitations? 5. IS it a good choice for this problem in view of the ahovc? The project was staffed by onc visiting sci~ntist working 1!111 time with a review pane 1 consi~tinq of four (two t~rhnicnl and two managerial ) TRIUMF employees. In ad(!ition, Othe r TRIUMF technical people werr called on to pcovidp detailed information about. spccif if” parts of the beam line and the curtcnt tune procedures. For a det.ailv:l technical dcscriptir)n rf thifi ~)~ojr(:t, Gce ~cfc[encv 1. .1 ,1 The Development of an Expert System to Tune a BPam Line 2.1 Characteristics Of Beam Line Control Problems There are characteristics common to beam line control problems. In most cases, there are a large number of repeating interacting devices. Each device has several related devices. The effects of these devices are not always well understood. While there may be a mathematical model of the theoretical beam line used to design the beam line initially, often that model is not precise enough to accurately predict what a given change will do to the real beam. That is the case for the 1. 2. 3. 4. 5. ISIS. The biggest factors preventing the use of a model to predict performance of ISIS are: A portion of the line goes through a wall. That portion is unshielded from the effects of the main magnet ( 90 gauss or mole). Permanent magnets are placed on the beam line to compensate for the effects of the main magnet. Due the the heavy weight of the accelerator, the floor below the ISIS beam line is sagging. Moving large pieces of metal, such as the crane, have an unpredictable effect on the magnet fields around the ISIS line. Each time the source is replaced, a s!ightly different set of initial beam conditions occur. 2.2 Implications Of Beam Line Characteristics On Expert System Tools The type of problem to be solved places constraints on the selection of tools to be used[2]. For beam line control prok~lems, the tool selected should provide for representing devices, interactions representing between devices,reasoning in the face of uncertainty, following an expert’s recipe for approaching a tune, and reasoning based on observing effects. 3 CI1OICI?,S ‘1’0have a successful expert system p[”ojr!ct care must be taken to choose the right project, the right expeLt, the riyht tor)]fi,and thu right, approach to representing the expottisc. l’hc 1S1S project will bn descrihcd in terms of the choices Fotcc[l on it Ily extclnal fartols 01 thr)ficmade fl~~ly. ‘1’hoficrhoicns will thrl~ hn cumpalp{l t(”) c1 llf~lff*(”( pl”ojf?ct t,() tly t(1 idcnt ify thof;o [cI(:toI:; th(l( m~lr;tI)(?t:tlll!~i(l(~l(’{1If) The Dt2Ve10pIW11L (JL all GAp CLL d~d.~... .- ..-. .-— —— . . increase the likelihood of successful ap~)licat ion u f ~X~)(’L ~ Sy!; ({.111 techniques to beam line control nrobl ems. 3.1 Choice Of Problem The 1S1S tune problem was chosen because it is big enough to test the use o f expert system techniques without being too big to handle. It is a real. problem that has a real need for an automated solution. ISIS is a good choice because there is a limited amount of expertise available. There is a dramatic difference between the performance of the best tune r and tl.e rest of the operators. The ISIS tune problem has a long life for the current TRIUMF accelerator wilml be the injector to the KAON factory. . 3.2 Choice Of Development Team Members of the project team should be trained in the technology to be used, knowledgeable about the application, have adequate time to spend on the tasks assigned, and dedicated to the success of the project. The domain expert is particularly important. Fred Bach was available, and the acknowledged expert OR the topic. He was interested, and cooperative. He has a long term familiarity with ISIS going back some 16 years. Fred was a cooperative expert who was willing to give up the secrets of his success. Sometimes this is difficult to achieve. Often a cjood candidate is one who wants or needs or is encouraged to move on to other tasks either for professional development or retirement. In those cases, the expert is not threatened by the apparent loss of scme of his or her uniqueness. Fred felt that this project would make his job easier and hence he did not consider the expert system a threat. 3.3 Choice Cf Development Machine TRIUMF had at the start of this project about 25 VAX wr)rkstatinns available to use. All of the workstations were connected to a Local A~ea VAX Cluster (LAVC). In thi~ environment, the workstati(ln~ c1I(’ normally disk-l~?ss ai~d all the operating system, paging, windowing, NI)CI nppliccltinn data is obtained OVCK the ETHERNET ftom/to thr (lick 5P1VPI 0II the 1>001 Iludr?. At. first,, a VAX station 2000 anfl tllvn latol , a wrt’~ l]Grd for lTA. 1200 Each initially hd(l flMB of mcmc]l~. 4 ‘lne Uevf51U~IIl~llL UL all IZAp CL u UY-LQ III -U AU . . ---- ------ —— -- — Beth workstations started disk-less, but, it quickly became Clc’ur thilt the demands on the boot node fo[- windowing and paqing wore a serious bottleneck. A single low performance local disk was [“)ut on both the 2000 and the 3200. This improved performance. For example, #indow swapping no longer took several minutes; but, the improved performance was still subjectively slower than the author’s previous experience using a 16MB micro-VAX II using two of the same 1aw performance disks in a stand-alone system. The LAVC environment was also considerably less reliable than similar hardware running outside a LAVC. The LAVC stations experienced repeated hang-ups when mouse or keyboard requests would be Another ignored. problem was the relatively frequent system crashes where the workstation’s processor would halt or spontaneously re-boot. The unpredictability and unreliability of the LAVC was- a serious problem for the development of ITA. 3.4 Choice Of Development Software Two different Expert System shells were used and the features and limitations of each were compared informally. We started NEXPERT, using to evaluate the premise that an apparetitly simpler tool might be easier to learn, use, and integrate. KEE was later used as the full featured, high powered approach. 3.S Choice Of Approach The ISIS tune advisor (ITA) project took a slightly different approach than traditional Expert System applications 13]. In most expert systems, either one expert is used, or a group of experts combine their knowledge to form a composite expert. ITA uses three slightly different approaches. The first attack on a problem is to use the heuristics of the operational expert. If that succeeds, all is well. If the operational approach fails to solve the problem, a second attempt is made using a technique developed by a beam line physicist from a theoretical analysis of the problem. If that approach also fails, a technique is tried which mimics the actions of an operator when faced with a problem for which he has no previous experience. In this fallback, last-chance, approach the expert system searches for a solution using only very general heuristics based on relative location and physical performance feedback to limit the search for the right correctors. It is hoped that the combined powel to solve tuning problems will he yreater than concentrating on any onc app[oach. !) 3.6 Choice Of Delivery Vehicle While the project did not advance to the delivery phase, it j.s clear that features and performance that are adequate or suitable fur development could be inappropriate for a production environment. This is true for both hardware and software. It is very likely that the best approach is to develop the expert system in a powerful high performance environment and when the concepts have been proven, port the knowledge gained in the development process to the best delivery vehicle available. 4 THE PERFECT PROJECT .“ This might be considered as an utopian description of the perfect project. However, it can be practically used to evaluate a proposed expert system project to determine the chances for success. While projects will succeed even if they differ from the ideals expressed here, the more deviation the more likely the project will not meet expectations, The knowledge gained from the work on ISIS and other accelerator control artificial intelligence projects worked on by the author is used here to describe the best of all pos6ible projects. .’ 4.1 The Perfect Problem The perfect problem has a high payoff, available expertise, cooperative expert, and a relatively confined problem area (well circumscribed). Complexity caused by repetition is a problem that can be handled more easily by a program than by a person. 4.2 The Perfect Development Team To create a knowledge based expert system successfully requires several things; the most important are the people that make up the project team. Ideally, the project team is made up of trained Knowledge Engineers (KE), programmers, an e~pert, and a managerial team that is knowledgeable in the technology. The team must have adequate time and support to build the system. The development team must be open to technology, focused, concis~, and insightful. The expert should have a stake in the out come. 110 should be able to share the praise or blame for the success or failure of the project. Not only must t.hc expert be willing to sp~nd tbc tim~ on the project but the expert’s managets must be committed to Suppolt (i he project. It has been said that you know that you have the [ight ‘xpert when the manager is reluctant to give up that person’s time. .3 The Perfect Development Machine It Ileeds to have high performance, powerful environment, and be ntegrated with existing beam line control system machines (for access .0 real data). High performance implies large effective memory :apacity, fast swapping device(s), high resolution display, and a fact Iccurate pointing device. Large amounts of memory are particularly lseful whev using a LISP based tool on a traditional work station. ,ISP in that configuration takes a large amount of phy~ical and virtual Iemory space . Though the single sample of a C based tool did not terform noticeably better in the LAVC environment. Whatever machine is :hosen, it must also be robust. For these reasons the development Iachine m!-.stbe dedicated to the task. It should be a stand-alone ~ystem in the sense that it is insensitive to outside forces. nterruptions affect development of expert systems even more seriously .han similar interruptions to traditional .programming for the following ‘easons. ,. Large codes imply longer times to save ,. Quick response encourages experimentation /. Neither shell used has journaling 1. Expert Systems require broader more complex thought processes Because the codes are much larger (the initial prototype of ITA :ontains almost 500K bytes ) it takes a relatively long time to save. ‘his coupled with the speed with which you can try another featu[e, encourages the knowledge engineer to “try just one more thing before Laving” . The cost is often recreating what was lost due to a hang or a :rash. In traditional development with a text editor, a crash or hang rould not be as costly becal~se usually the incremental adrlitions to a ~rogram are smaller and most of a lost editor session can normally t)e “ecovered from the journal file autamat.ically g~ncrated, Nf?ith~r :xpetl Sy$+tem stlell used pl-ovides a journaling facility. The breadth ~jf knowledge contained in an expert system usually i!i substantially k]roader and more complex than in a traditional Ipp],ication program. Even there, an int~lru~~t inn can }1p costly in .erms of time Lcquiled to regain a train 0[ thouyllt. Intvl~upLiorlfi alu ~ven more costly in expert system development because of the increased complexity. 4.4 The Perfect Development Software Development software must be powerful, fast, flexible, easy to :hange, simple to document applications, easy to learn (good tutorial naterial with plenty of examples) and use. It must have good debugging tools , well documented features, multiple representations, and full support of external subroutines and interfaces to existing independent ?rograms. Object-oriented programming tools provide many of the required features. .. The thought processes required to develop an expert system are ~ifferent than for a traditional program[4]. Expert system shells that look like or are patterned after traditional approaches, i.e., text sditors or spread sheets, on the surface are very attractive for they look familiar to the typical traditional programmer. But, that Familiarity can be costly for it makes.it doubly difficult to change :he traditional (procedural) thought patterns. 1.5 The Perfect Approach This, of course, depends on the specific application. The development team must have a broad repertoire of tools and techniques :0 apply. Each application will be somewhat different thereby requiring a variety of approaches to solve. 1.6 The Perfect Delivery Environment Both hardware and software should be inexpensive (so that many :opies can be economically fielded) , fast, robust, and requires minimal :hanges from development. It should match the existing environment so lo new policies, procedures, or maintenance is needed. i CONCLUSIONS AI technology can be profitably used in the solutio~ to Accelerator related problems, but, care must he taken in the choice of )Lohlem, expert, development and rl~livery vehicles[ 5]. AI tecllni~lu~s Ire appropriate because algorithms fail and the captured knowledge of the expert does lead to a solution. AI technology is costly in time to develop and in money for softwar~, hardware, and people. It has not stressed the system yet but in operation it will put significant. demands on the control system for data. The major limitatio;~s of AI technology are the limited number of people trained to do t!le wol-k the ,31)(1 poor performance of the shells used in the VAX workstation LAVC environment. AI technology is none the less the best choice available. There are several actions that can be taken to improve the chances for success of a production version of the prototype. Some suggestions for improvement include upgrade the workstation to a more powerful system, refine the approach used to include more, different or more powerful techniques, and port the application to a more powerful delivery environment. .. 6 ACKNOWLEDGEMENTS I would like to thank the fcllowing people for giving me this opportunity and assisting me on this project. Martha Hoehn and Earl Hoffman for arranging and approving the ‘leave from LANL to come to TRIUMF. Ken Dawson and Don Dohan for arranging for my stay at TRIUMF and selecting this project. Peter Grant and Mi’ke Mouat for technical assistance and support on the workstation. Fred Bach for his tuning knowledge of 1S1S. 7 REFERENCES [1] Schultz, D. and Brown, P., “Using Expert System Techniques to Tune the ISIS Beam Line”, LAMS-89–XXXX [2] Harmon, P. and King, D., =rz 5LsIg !:I-izi.!!.s: Intelligence in Business, John Wiley, New York, 1985.— .- -——— [3] Winston, P., Artificial Intelligence, Addison Wesley, Reading-——— .—— --— .-.—..... . MA, 1984 [4] Charniak, E. and McDermott, D., Introduction to Arti!l.g.l.?.}—..— — — Intelligence, Addison Wesley, Reading, MA, 1985 [5] Schultz, D. and Silbar, R., “Toward Automatic Control of Particle Accelerator Beams” in Artificial Intelligence in—- — Engineering: Diagnosis and Learning, J.S.Gero editor, Elsevier,—- — Amsterdam, 1988 .. 10 
work_x6fbaqa72nhnhazv3jeeoqz4ci	----	doi:10.1016/j.eswa.2006.04.001 www.elsevier.com/locate/eswa Expert Systems with Applications 33 (2007) 1–5 Expert Systems with Applications A novel feature selection algorithm for text categorization Wenqian Shang a,*, Houkuan Huang a, Haibin Zhu b, Yongmin Lin a, Youli Qu a, Zhihai Wang a a School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, PR China b Department of Computer Science, Nipissing University, North Bay, Ont., Canada P1B 8L7 Abstract With the development of the web, large numbers of documents are available on the Internet. Digital libraries, news sources and inner data of companies surge more and more. Automatic text categorization becomes more and more important for dealing with massive data. However the major problem of text categorization is the high dimensionality of the feature space. At present there are many meth- ods to deal with text feature selection. To improve the performance of text categorization, we present another method of dealing with text feature selection. Our study is based on Gini index theory and we design a novel Gini index algorithm to reduce the high dimensionality of the feature space. A new measure function of Gini index is constructed and made to fit text categorization. The results of experiments show that our improvements of Gini index behave better than other methods of feature selection. � 2006 Elsevier Ltd. All rights reserved. Keywords: Text feature selection; Text categorization; Gini index; kNN classifier; Text preprocessing 1. Introduction With the advance of WWW (world wide web), text cat- egorization becomes a key technology to deal with and organize large numbers of documents. More and more methods based on statistical theory and machine learning has been applied to text categorization in recent years. For example, k-nearest neighbor (kNN) (Cover & Hart, 1967; Yang, 1997; Yang & Lin, 1999; Tan, 2005), Naive Bayes (Lewis, 1998), decision tree (Lewis & Ringuette, 1994), support vector machines (SVM) (Joachims, 1998), linear least squares fit, neural network, SWAP-1, and Roc- chio are all such kinds of methods. A major problem of text categorization is the high dimensionality of the feature space. For many learning algorithms, such high dimensionality is not permitted. Moreover most of these dimensions are not relative to text categorization; even some noise data hurt the precision of 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.04.001 * Corresponding author. E-mail addresses: shangwenqian@hotmail.com (W. Shang), haibinz@ npissingu.ca (H. Zhu). the classifier. Hence, we need to select some representative features from the original feature space (i.e., feature selec- tion) to reduce the dimensionality of feature space and improve the efficiency and precision of classifier. At present the feature selection method is based on statistical theory and machine learning. Some well-known methods are information gain, expected cross entropy, the weight of evi- dence of text, odds ratio, term frequency, mutual informa- tion, CHI (Yang & Pedersen, 1997; Mladenic & Grobelnik, 2003; Mladenic & Grobelnik, 1999) and so on. In this paper, we do not discuss these methods in detail. We present another new text feature selection method— Gini index. Gini index was early used in decision tree for splitting attributes and got better categorization precision. However, it is rarely used for feature selection in text cate- gorization. Shankar and Karypis discuss how to use Gini index for text feature selection and weight-adjustment. They mainly pay attention on weight-adjustment. Their method only limits to centroid based classifier and their iterative method is time-consuming. Our method is very different from theirs. Through deeply analyzing the princi- ples of Gini index and text feature, we construct a new mailto:shangwenqian@hotmail.com mailto:haibinz@npissingu.ca mailto:haibinz@npissingu.ca 2 W. Shang et al. / Expert Systems with Applications 33 (2007) 1–5 measure function of Gini index and use it to select features in the original feature space. It not only fit centroid classi- fiers but also fit other classifiers. The experiments show that its quality is comparable with other text feature selection methods. However, its complexity of computing is lower and its speed is higher. The rest of this paper is organized as follows. Section 2 describes the classical Gini index algorithm. Section 3 gives the improved Gini index algorithm. Section 4 discusses the classifiers using in the experiments to compare Gini index with the other text feature selection methods. Section 5 pre- sents the experiments’ results and their analysis. In the last section, we give the conclusion. 2. Classical Gini index algorithm Gini index is a non-purity split method. It fits sorting, binary systems, continuous numerical values, etc. It was put forward by Breiman, Friedman, and Olshen (1984) and was widely used in decision tree algorithms of CART, SLIQ, SPRINT and intelligent miner. The main idea of Gini index algorithm is as follows: Suppose S is the set of s samples. These samples have m different classes (Ci,i = 1, . . ., m). According to the differ- ences of classes, we can divide S into m subset (Si,i = 1, . . ., m). Suppose Si is the sample set which belongs to class Ci, si is the sample number of set Si, then the Gini index of set S is: GiniðSÞ¼ 1 � Xm i¼1 P 2i ; ð1Þ where Pi is the probability that any sample belongs to Ci and estimating with si/s. Gini(S)’s minimum is 0, that is, all the members in the set belong to the same class; this de- notes it can get the maximum useful information. When all the samples in the set distribute equably for the class field, Gini(S) is maximum; this denotes it can get the minimum useful information. If the set is divided into n subset, then the Gini after splitting is: GinisplitðSÞ¼ Xn j¼1 sj s GiniðSjÞ: ð2Þ The minimum Ginisplit is selected for splitting attribute. The main idea of Gini index is: for every attribute, after it traverses all possible segmentation methods, if it can pro- vide the minimum Gini index then it is selected as the divi- sive criterion of this node no matter it is the root node or a sub node. 3. The improved Gini index algorithm To apply the Gini index theory described above directly to the text feature selection, we can construct the new formula: GiniðW Þ¼ PðW Þ 1 � X i PðCijW Þ 2 ! þ PðW Þ 1 � X i PðCijW Þ 2 ! ð3Þ After we analyze and compare the merits and demerits of the existing text feature selection measure functions, we improve formula (3) to: Gini TextðW Þ¼ X PðW jCiÞ 2PðCijW Þ 2 : ð4Þ Why we amend formula (3) to formula (4)? The reasons include three aspects as follows: (1) The original form of Gini index is used to measure the impurity of attributes towards categorization. Smaller the impurity is, better the attribute is. If we adopt the form GiniðSÞ¼ Pm i¼1P 2 i , it is to measure the purity of attributes towards categorization. Big- ger the value of purity is, better the attribute is. In this paper, we adopt the measure form of purity. This form is more adapt to text feature selection. In paper (Gupta, Somayajulu, Arora, & Vasudha, 1998; Shan- kar & Karypis), they all adopt the measure form of purity. (2) In other authors’ papers, they all emphasize that text feature selection inclines to high frequency words, namely, including the P(W) factor in the formula. Experiments show that some words that do not appear have contributions to judge the class of text, but this contribution is far less significant than the effort to consider the words that do not appear, espe- cially when the distribution of the class and feature values is highly unbalanced. Yang and Pedersen (1997) and Mladenic and Grobelnik (1999) compare and analyze synthetically the merits and demerits of many feature measure functions in their papers. Their experiments show that the demerits of information gain are to consider the word that does not appear. The demerits of mutual information are not to con- sider the affect of the P(W) factor leading to select rare words. Expected cross entropy and weight of evi- dence of text overcome these demerits, hence their results are better. Therefore, when we construct the new measure function of Gini index, we get ride of the affection factor expressing words that do not appear. (3) Iff W1 appears in the documents of class C1 and W1 appears in every document of class C1; Iff W2 appears in the documents of class C2 and W2 appears in every documents of class C2, then W1 and W2 is the same important feature. But due to P(Ci) 5 P(Cj), from Gini TextðW Þ¼ PðW Þ P iPðCijW Þ 2 to compute out Gini Text(W1) 5 Gini Text(W2), this is not consistent with domain knowledge. So we adopt P(WjCi)2 to replace P(W), for considering the unbalanced class W. Shang et al. / Expert Systems with Applications 33 (2007) 1–5 3 distribution. In formula (4), iff W appears in the doc- uments of class Ci and W appears in every document of class Ci, it can get the maximum Gini Text(W), namely Gini Text(W) = 1. This is consistent with domain knowledge. If there is no term P(WjCi)2, according to the Bayes decision theory of minimum error rate, P(CijW)2 is the posterior probability when feature W appears. When the documents distribute evenly where W appears, it gets the minimum Gini - Text(W). But text feature is special, it only gets two values: appearance in the documents or no appear- ance in the documents. Moreover, according to field knowledge, we omit the circumstance that a feature does not appear in the documents. The class in the training set is always unbalanced and it is opinion- ated to decide Gini Text(W) is the minimum. Hence, when we construct the new measure function of Gini index, we consider feature W’s condition probability, combining posterior probability and condition prob- ability as the whole measure function to depress the affection when the class is unbalanced. 4. Classifiers in the experiments In order to evaluate the new feature selection algorithm, we use three classifiers: SVM (support vector machine), kNN and fkNN to show that our new Gini index algorithm is effective in different classifiers. The algorithms of classifi- ers can be described as follows. 4.1. kNN classifier The kNN algorithm is to search k documents (called neighbors) that have the maximal similarity (cosine similar- ity) in training sets. According to what classes these neigh- bors are affiliated with, it grades the test document’s candidate classes. The similarity between the neighbor doc- ument and the test document is taken as this class weight of neighbor documents. The decision function can be defined as follows: ljðXÞ¼ Xk i¼1 ljðX iÞsimðX ; X iÞ; ð5Þ where lj(Xi) 2 {0, 1} shows whether Xi belongs to xj(lj(Xi) = 1 is true) or not (lj(Xi) = 0 is false); sim(X, Xi) denotes the similarity between training document and test document. Then the decision rule is: If ljðXÞ¼ max i liðXÞ, then X 2 xj. 4.2. fkNN classifier The kNN algorithm in 4.1 can not get better categoriza- tion performance, especially when the class is unbalanced. Hence, we adopt the fuzzy theory to improve the kNN algorithm as follows. The reasons of this improvement can consult (Shang, Huang, Zhu, & Lin, in press): ljðXÞ¼ Pk i¼1ljðX iÞsimðX ; X iÞ 1ð1 � simðX ; X iÞÞ 2=ðb�1ÞPk i¼1 1 ð1 � simðX ; X iÞÞ 2=ðb�1Þ ; ð6Þ where j = 1, 2, . . ., c, lj(Xi) is the membership of known sample X to class j. If sample X belongs to class j then the value is 1, otherwise 0. From this formula, we can see that in reality the membership is using the different distance of every neighbor to the candidate classifying sample to weigh its effect. Parameter b is used to adjust the degree of a distance weight. In this paper we take b’s value 2. Then fuzzy k-nearest neighbors’ decision rule is: If lj(X) = max- ili(X), then X 2 xj. 4.3. SVM classifier SVM is put forward by Vapnik (1995). It is used to solve the problem of two-class categorization. Here we adopt the linear SVM, using the method of one-versus-rest to classify the documents. The detailed description can be referred to (Vapnik, 1995). 5. Experiments 5.1. Data collections We use two corpora for this study: the Reuters-21578 and data set coming from the International Database Cen- ter, Department of Computing and Information Technol- ogy, Fudan University, China. In Reuters-21578 data set, we adopt the top ten classes. 7053 documents in training set and 2726 documents in test set. The distribution of the class is unbalance. The maxi- mum class has 2875 documents, occupying 40.762% of training set. The minimum class has 170 documents, occu- pying 2.41% of training set. In the second data set, we use 3148 documents as train- ing samples and 3522 documents as test samples. The train- ing samples are divided into document sets A and B. In document set A, the class distribution is unbalance. In these documents, the political documents are 619 pieces, occupying 34.43% of the training document set A, the energy sources documents are only 59 pieces, occupying 3.28% of the training document set A. In training sample B, the class distribution is correspondingly balance. Every class is 150 pieces. 5.2. Experimental settings For every classifier, in the phase of text preprocess we use information gain, expected cross entropy, the weight of evidence of text and CHI to compare with our improved Gini index algorithm. Every measure function can be described as follows: 4 W. Shang et al. / Expert Systems with Applications 33 (2007) 1–5 Information gain: Inf GainðW Þ¼ PðW Þ Xm i PðCijW Þlog2 PðCijW Þ PðCiÞ þ PðW Þ Xm i PðCijW Þlog2 PðCijW Þ PðCiÞ ð7Þ Expected cross entropy: Cross EntropyðW Þ¼ PðW Þ Xm i PðCijW Þlog2 PðCijW Þ PðCiÞ ð8Þ CHI(v2): v2ðW Þ¼ Xm i PðCiÞ � NðA1A4 � A2A3Þ 2 ðA1 þ A3ÞðA2 þ A4ÞðA1 þ A2ÞðA3 þ A4Þ ð9Þ Weight of evidence of text: Weight of EvidðW Þ¼ PðW Þ � Xm i¼1 PðCiÞ log PðCijW Þð1 � PðCiÞÞ PðCiÞð1 � PðCijW ÞÞ ���� ���� ð10Þ After selecting the feature subset using above measure functions, we use TF–IDF to weight the feature, the for- mula is as follows: wik ¼ tf ik � logðN=niÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPM j¼1½tf ik � logðN=niÞ� 2 q ð11Þ In Reuters-21578, k = 45, in document set A k = 10, in document set B k = 35. Table 1 The performance of five feature selection measure functions on top 10 classes Measure function SVM kNN Macro-F1 Micro-F1 Macro Gini index 69.940 88.591 66.584 Inf Gain 69.436 88.445 66.860 Cross Entroy 69.436 88.445 66.579 CHI 67.739 88.225 66.404 Weigh of Evid 68.731 88.481 66.766 Table 2 The performance of five feature selection measure functions on training set A Measure function SVM kNN Macro-F1 Micro-F1 Macro Gini index 91.577 90.941 84.176 Inf Gain 91.531 90.708 83.318 Cross Entroy 91.481 90.708 83.318 CHI 91.640 91.057 84.491 Weigh of Evid 91.407 90.825 84.073 5.3. Performance measure To evaluate the performance of a text classifier, we use F1 measure put forward by Rijsbergen (1979). This mea- sure combines recall and precision as follows: Recall ¼ number of correct positive predictions number of positive examples Precision ¼ number of correct positive predictions number of positive predictions F 1 ¼ 2 � Recall � Precision ðRecall þ PrecisionÞ 5.4. The experimental results and analysis The experimental result in Reuters-21578 can be described as Table 1. From this table, we can see that in SVM and fkNN, Gini index gets the best categorization performance. We can notice that five measure functions show better performance all. In SVM, the micro-F1 difference between the best and the worst is 0.366%, in kNN is 0.294%, in fkNN is 0.477%. In kNN, the Macro-F1 of Gini index is only inferior to information gain, the Micro-F1 of Gini index is only infe- rior to CHI. The experimental result in the second data set can be described as Tables 2 and 3. From Table 2, we can see that the categorization perfor- mance in SVM, Gini index is only inferior to CHI and exceed Information Gain, in kNN, the Macro-F1 of Gini index is only inferior to CHI, but the Micro-F1 of Gini index gets the best, in fkNN, Gini index is only inferior to weight of evidence of text. fkNN -F1 Micro-F1 Macro-F1 Micro-F1 85.620 67.999 86.537 85.326 67.032 86.134 85.326 67.518 86.207 85.761 66.846 86.060 85.180 67.509 86.280 fkNN -F1 Micro-F1 Macro-F1 Micro-F1 83.043 84.763 83.856 81.301 84.346 82.811 81.301 84.216 82.578 82.811 85.256 84.008 82.927 85.867 85.017 Table 3 The performance of five feature selection measure functions on training set B Measure function SVM kNN fkNN Macro-F1 Micro-F1 Macro-F1 Micro-F1 Macro-F1 Micro-F1 Gini index 91.421 91.222 86.272 85.222 87.006 86.556 Inf Gain 91.799 91.556 86.326 85.222 87.305 86.556 Cross Entroy 91.419 91.222 85.764 85.111 86.999 86.444 CHI 91.238 91.000 85.770 85.000 86.898 86.444 Weigh of Evid 91.799 91.556 85.914 85.111 87.138 86.444 W. Shang et al. / Expert Systems with Applications 33 (2007) 1–5 5 From Table 3, we can find that the categorization per- formance in SVM, Gini index is only inferior to informa- tion gain and weight of evidence of tex, in kNN, the Macro-F1 of Gini index is only inferior to information gain, but the Micro-F1 of Gini index gets the best, in fkNN, the Macro-F1 of Gini index is only inferior to infor- mation gain, but the Micro-F1 of Gini index gets the best. In summary, in some data set, the categorization perfor- mance of our improved Gini index gets the best. In another data set, its performance is only inferior to other measure function. As a whole, Gini index shows better categoriza- tion performance. From formula (7)–(10), we can find that the computation of Gini index is simpler than other feature selection methods. Gini index has no logarithm computa- tions and only has simple multiplication operations. 6. Conclusion In this paper, we studied the text feature selection based on Gini index. We compare its performance with the other feature selection methods in text categorization. The exper- iments show that our improved Gini index has a better per- formance and simpler computation than the other feature selection methods. It is a promising method for text feature selection. In the future, we will improve this method further and will study how to select different feature selection methods at different data set. Acknowledgement This research is partly supported by Beijing Jiaotong University Science Foundation under the Grant 2004RC008. References Breiman, L., Friedman, J. H., Olshen, R. A., et al. (1984). Classification and regression trees. Montery, CA: Wadsworth International Group. Cover, T. M., & Hart, P. E. (1967). Nearest neighbor pattern clas- sification. IEEE Transaction on Information Theory, IT-13(1), 21– 27. Gupta, S. K., Somayajulu, D. V. L. N., Arora, J. K., & Vasudha, B. (1998). Scalable classifiers with dynamic pruning. In Proceedings of the 9th international workshop on database and expert systems applications (pp. 246–251). Washington, DC, USA: IEEE Computer Society. Joachims, T. (1998). Text categorization with support vector machines: learning with many relevant features. In Proceedings of the 10th European conference on machine learning (pp. 137–142). New York: Springer. Lewis, D. D. (1998). Naı̈ve (Bayes) at forty: the independence assumption in information retrieval. In Proceedings of the 10th European confer- ence on machine learning (pp. 4–15). New York: Springer. Lewis, D.D., Ringuette, M., 1994. Comparison of two learning algorithms for text categorization. In Proceedings of the third annual symposium on document analysis and information retrieval. Las Vegas, NV, USA, pp. 81–93. Mladenic, D., Grobelnik, M., 1999. Feature selection for unbalanced class distribution and Naı̈ve Bayes. In Proceedings of 16th international conference on machine learning, San Francisco 258–267. Mladenic, D., & Grobelnik, M. (2003). Feature selection on hierarchy of web documents. Decision Support Systems, 35(1), 45–87. Rijsbergen, V. (1979). Information retrieval. London: Butterworth. Shang, W., Huang, H., Zhu, H., & Lin, Y. (2005). An improved kNN algorithm—Fuzzy kNN. In Proceedings of international confer- ence on computational intelligence and security (pp. 741–746). China: Xi’an. Shankar, S., Karypis, G. A feature weight adjustment algorithm for document categorization. Available from: http://www.cs.umm.edu/ ~karypis. Tan, S. (2005). Neighbor-weighted K-nearest Neighbor for Unbalanced Text Corpus. Expert System with Applications, 28(4), 667–671. Vapnik, V. (1995). The nature of statistical learning theory. Springer. Yang, Y. (1997). An evaluation of statistical approaches to text catego- rization. Information Retrieval, 1(1), 76–88. Yang, Y., Pedersen, J.O., 1997. A Comparative Study on Feature Selection in Text Categorization. In Proceedings of the 14th interna- tional conference on machine learning, Nashville, USA, pp. 412– 420. Yang, Y., & Lin, X. (1999). A re-examination of text categorization methods. In Proceedings of the 22nd annual international ACM SIGIR conference on research and development in the information retrieval (pp. 42–49). New York: ACM Press. http://www.cs.umm.edu/~karypis http://www.cs.umm.edu/~karypis A novel feature selection algorithm for text categorization Introduction Classical Gini index algorithm The improved Gini index algorithm Classifiers in the experiments kNN classifier fkNN classifier SVM classifier Experiments Data collections Experimental settings Performance measure The experimental results and analysis Conclusion Acknowledgement References 
work_x7upiqlp5ra6jfzwdlzlg2wkki	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220969771 Params is empty 220969771 exception Params is empty 2021/04/06-03:05:09 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220969771 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:09 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_xao5rk2nyrdc3o6b4lxdaao2aa	----	A Grouping Hyper-Heuristic Framework: Application on Graph Colouring Anas Elhag, Ender Özcan ASAP Research Group School of Computer Science University of Nottingham Jubilee Campus, NG8 1BB Nottingham, UK Abstract Grouping problems are hard to solve combinatorial optimisation problems which require parti- tioning of objects into a minimum number of subsets while a given objective is simultaneously optimized. Selection hyper-heuristics are high level general purpose search methodologies that operate on a space formed by a set of low level heuristics rather than solutions. Most of the re- cently proposed selection hyper-heuristics are iterative and make use of two key methods which are employed successively; heuristic selection and move acceptance. At each step, a new solu- tion is produced after a selected heuristic is applied to the solution in hand and then the move acceptance method is used to decide whether the resultant solution replaces the current one or not. In this study, we present a selection hyper-heuristic framework including a fixed set of low level heuristics specifically designed for grouping problems. The performance of different hyper- heuristics using different components within the framework is investigated on a representative grouping problem of graph colouring. Additionally, the hyper-heuristic performing the best on graph colouring is applied to a benchmark of examination timetabling instances. The empirical results shows that the proposed grouping hyper-heuristic is not only sufficiently general, but also able to achieve high quality solutions for graph colouring and examination timetabling. Keywords: hyper-heuristics, grouping problems, graph colouring, timetabling 1. Introduction The task of partitioning a large set of items into a collection of mutually disjoint subsets is a common task in a variety of real-world problems. In a grouping problem, the goal is to opti- mise a given objective (cost, penalty) while achieving the minimum number of subsets (groups). Hence, grouping problems can be formulated as a multi-criteria discrete combinatorial optimisa- tion problem, considering that there is a trade-off between minimizing cost and number of groups, as in graph colouring (Saha et al., 2013), timetabling (Qu et al., 2009) and packing (Falkenauer, 1998). Email addresses: axe@cs.nott.ac.uk (Anas Elhag), exo@cs.nott.ac.uk (Ender Özcan) Preprint submitted to Expert Systems with Applications December 22, 2014 Two crucial components in the design of grouping algorithms for solving grouping problems are the candidate solution representation and neighbourhood/move operator(s). A redundant representation scheme which allows equivalent solutions yielding the same grouping creates a huge search space that might impair even the most powerful search algorithm. Many grouping approaches based on genetic algorithms (GAs) have been explored in the scientific literature providing various degrees of success (Falkenauer, 1998; Korkmaz, 2010). In previous studies, it has been observed that traditional operators are rather disruptive and, in many cases, counter productive, hence special operators that are tailored for grouping problems are needed. There is a growing number of studies on more general and reusable search methodologies applicable to multiple problem domains than the existing specifically tailored solutions to a sin- gle problem. Hyper-heuristics are such high level search methodologies that search the space formed by low level heuristics, instead of solutions directly for solving hard problems (Burke et al., 2013). There are different types of hyper-heuristics. The focus of this study is selection type of high level search methods that mix and control a pre-defined set of low level perturba- tive heuristics (operators) processing complete solutions at each step under a single-point based search framework. In this study, we describe a selection hyper-heuristic framework for grouping problems. The framework embeds a set of tailored low level grouping heuristics based on a restricted version of grouping representation, referred to as the Group Encoding (Falkenauer, 1998). In contrast to traditional selection hyper-heuristics that use a different set of low level heuristics provided for each different problem domain, in our proposed framework the set of low level heuristics is fixed and the same framework can be used for solving various grouping problems. This adds another level of generality when compared to generic selection hyper-heuristics. We have investigated the performance of the framework using different selection hyper- heuristic components on a set of well known graph colouring benchmark instances 1. Addi- tionally, we applied the same hyper-heuristic without any modification to a benchmark of exam- ination timetabling instances in order to examine the generality of the framework. The empirical results show that a learning selection hyper-heuristic developed using the framework turns out to be indeed sufficiently general and reusable. This hyper-heuristic either beats most of the pre- viously proposed approaches tailored for the specific problem in hand or shows that it is highly competitive. The paper is organised as follows. Section 2 provides an overview of grouping problems, different representation schemes for grouping problems and hyper-heuristics. The details of the proposed selection hyper-heuristic framework including all low level heuristics and different components are given in Section 3. The experimental design and results are discussed in Section 4, while the last section presents concluding remarks and future work. 2. Background 2.1. Grouping Problems Grouping problems are combinatorial optimisation problems in which a large group of n items, U = {x1, x2, x3, ..., xn}, is to be divided into a collection of k (2 ≤ k ≤ n − 1) subgroups, ui (1 ≤ i ≤ k); such that each item x ∈ U belongs to exactly one subgroup minimizing a given objective (cost/penalty/fitness) and k. Different grouping problems have different constraints, and 1ftp://dimacs.rutgers.edu/pub/challenge/graph/benchmarks/color/ 2 introduce different objective (cost) functions, as in graph colouring, timetabling, data clustering and packing (Falkenauer, 1998). In our formulation, we denote a cost function as a decomposable function, f (). For a subgroup ui, the partial cost is denoted as f (ui), and for a complete solution Ug = {u1, ..., uk}, f (Ug) is the total cost. minimise Z =  f (Ug) = k∑ i=1 f (ui), k  (1) sub ject to k⋃ i=1 ui = U (2) ui ⋂ u j = ∅ ∀i, j where i , j (3) ui , ∅ ∀i (4) In this study, we represent the grouping problem as a discrete two objective multi-criteria problem in which the goal is to optimize the two conflicting objectives in equation (1) above, namely the number of groups which can only take discrete values; and the cost which can take discrete or continuous values depending on the problem. Ideally, these two objectives should be simultaneously optimised, although they are clearly conflicting; i.e. a decrease in the number of groups k leads to an increase in the cost. In some cases, there might not be a single optimal solution. Instead, there could be multiple solutions with a trade-off from which a decision maker can choose. Those solutions are identified using the concept of dominance (Zitzler and Thiele, 1998) as illustrated in Figure 1. Pareto front Dominated solutions Number of groups in the solution (k) F it n e ss v a lu e ( f( U k )) Figure 1: The Dominance concept in Multi-objective optimisation. A solution x is considered to dominate another solution y, (x � y) if, and only if, x is better than y in at least one objective, and x is not worse than y in any of the objectives. The set of the non-dominated solutions is known as the Pareto optimal set, and its image in the objective 3 domain is known as the Pareto optimal front. This problem is different than a generic multi- objective problem where mostly, there is a region where the Pareto front is driven automatically via a multi-objective algorithm, however, in grouping problems the range of groups is fixed, hence the search methodology can focus on a single objective without ignoring the second one. We use some basic ideas from multi-objective optimisation, but the proposed approach is not a generic multi-objective algorithm as described in Section 3. For more on multi-objective optimi- sation, readers can refer to (Zitzler and Thiele, 1998; Coello et al., 2007), which is not the focus of this study. 2.1.1. Grouping Representations Almost all previously proposed grouping approaches are genetic algorithms utilizing various encoding schemes for grouping problems. Those schemes can be classified as fixed and variable length representations. A fixed-length representation is based on an array of values associated in some way with each item in a set of objects, such as, Group Numeric Encoding (GNE) and Permutation with Separators Encoding (PWS) (Jones and Beltramo, 1991) which are widely used in the literature. Each location in the array is associated with an item and in GNE, the value at a location indicates the group that the item belongs to, whereas in PWS, that value represents the relative positions of the objects with respect to each other. However, many studies have concluded that such representations have some deficiencies. One of the crucial flaws is the redundancy in the representation. Different candidate solutions under a redundant representation could yield the same grouping of items. For example, assuming that we have 3 objects for grouping under GNE, <1, 1, 2> indicates that there are two subgroups, first and second items form a subgroup while the third item forms another subgroup. <2, 2, 1>, <3, 3, 1>, <1, 1, 3>, <3, 3, 2> and <3, 3, 1> also represent the same grouping. This type of redundancy in return creates a larger search space potentially impairing the search algorithms (Falkenauer, 1992). Also, if the traditional genetic operators are used with those representations, they could significantly damage the good solutions that are being developed during the search. For example, (Du et al., 2004) showed that the standard one-point crossover has various disadvantages and is not suitable for grouping problems. In (Park and Song, 1998), a linkage-based representation has been proposed, known as the Locus-Based Adjacency (LBA) representation, which is a fixed-length representation in which each location represents one object and stores an integer value that represents a link to another object in the same group. This representation has been applied in many grouping problems (Handl and Knowles, 2007), but it still suffers the same issues of redundancy and lack of ap- propriate operators as the former representations. In (Du et al., 2004), a restricted version of the LBA known as the Linear Linkage Encoding (LLE) has been proposed, in which backward links are not allowed and, apart from an ending node, each node has only one node pointing to it. LLE successfully eliminates the redundancy problem that debilitates former representations, and has been applied to a variety of grouping problems such as Bin Packing, Graph Colouring and Timetabling (Ülker et al., 2006; Ülker et al., 2008). However, maintaining the LLE links during the search process is costly, particularly while using some genetic operators (Yılmaz and Korkmaz, 2010). A variable-length representation referred to as the Grouping Encoding (GE) has been pro- posed in (Falkenauer, 1992), along with some suggested tailored genetic operators to be used with it. The basic argument behind the GE is that the grouping representations and associated operators should be carefully designed such that they are aware of the nature of the underlying problem; i.e constructing groups that maximise the fitness values of the solutions. Individuals 4 in the GE are separated into two parts, Ug = {x1, x2, x3, ..., xn | u1, u2, ..., uk}. The first part is referred to as the elements section and is only used to identify the group to which each item be- longs. No genetic operators are applied to this part. The second part is referred to as the groups section, in which a list of the groups in the individual is maintained. This is the part to which the genetic operators are applied, and then, all the changes that happen in this section are reflected back on the elements section. For example, the individual {1 4 2 3 4 4 1 3 2 2 | 1 2 3 4} shows a grouping of 10 items into 4 groups, where the actual grouping is {x1, x7}, {x3, x9, x10}, {x4, x8}, {x2, x5, x6}. Note that, for a given problem instance, the length of the elements section is fixed, whereas the length of the groups section is not. The GE has been applied in many real world grouping problems including data clustering (Agustı́n-Blas et al., 2012) and machine-part cell formation (Brown and Sumichrast, 2003). A useful aspect of the GE representation is its ability to address the underlying problem of constructing fit subgroups (with potentially reduced cost). However, the redundancy issue still remains, since more than one solution could represent the same grouping. For example, the solution {1 3 2 4 3 3 1 4 2 2 | 1 2 3 4} produces the same grouping as the solution provided at the beginning of this paragraph. In this study, we used a slightly modified version of the GE representation that is structured to overcome the redundancy issue. Two well-known restrictions are introduced to the standard Falkenauer’s GE in order to eliminate this redundancy: • In an encoding with k groups, subgroups are enumerated using values 1 through k, only. • Subgroups containing items with lower indices are enumerated first. Applying these two restrictions, the solution {x1, x7}, {x3, x9, x10}, {x4, x8}, {x2, x5, x6} can only be represented as {1 2 3 4 2 2 1 4 3 3 | 1 2 3 4}. 2.2. Graph Colouring and Examination Timetabling Given a an undirected graph G = (V, E) with a set of vertices V = {x1, x2, x3, ..., xn}, and a set of edges E, where exp,xq ∈ {0, 1} represents whether there is an edge between two given vertices (1) (xp, xq ∈ V ) or not (0), Graph Colouring Problem (GCP) requires colouring of vertices using a given number of colours such that none of the adjacent vertices (connected with an edge) are in the same colour, hence conflict free. The minimum number of colours that achieves a conflict free colouring is also referred to as the chromatic number (χ(G)). Determining the chromatic number of a given graph G, and finding out whether G is k-colourable (whether k colours are sufficient to create a conflict free graph) are well-known variants of the GCP which are NP-hard and NP-complete, respectively (Saha et al., 2013). In this study, we formulate minimum colouring variant of GCP as a multi-criteria grouping problem. Assuming that the number of colours is denoted as k, where 2 ≤ k ≤ n − 1, and vi represents a subgroup of V in which the vertices are assigned to the same colour i, (1 ≤ i ≤ k), the definition of the grouping problem defined in section 2.1 follows, taking colours ⇔ groups, V ⇔ U, and vi ⇔ ui. The cost can be measured using the equation 5. f (ui) = f (vi) = ∑ p,q exp,xq ,∀xp, xq ∈ vi and exp,xq ∈ {0, 1}, where p < q (5) Graph colouring is one of the extensively studies areas of research, yet many studies have been emerging. Many techniques have been proposed for solving many variants of GCP. Exact 5 approaches tend to fail particularly while solving large instances, hence many researchers have been working on heuristic approaches. For example, recursive largest fit is a well-known greedy heuristic introduced by (Leighton, 1979). Hertz and Werra (1987) presented the first tabu search implementation which outperforms another local search method, simulated annealing on random dense graph instances. Davis and Mitchell (1991) proposed a coding as an ordering of vertices which could be used in a genetic algorithm. Johnson et al. (1991) presented three simulated annealing implementations based on three neighbouring approaches. Galinier and Hao (1999) have shown that hybridisation of GAs with local search methods are more promising than GAs on their own. In such hybridisation, local search operators are used as intensification methods to explore promising areas of the search space that have found by the GA operators. Avanthay et al. (2003) proposed a variable neighbourhood search algorithm for graph colouring problem. Studies in (Ülker et al., 2006; Ülker et al., 2008; Korkmaz, 2010) and (Kirovski and Potkonjak, 1998) apply generic GAs using some of the genetic representations discussed in section 2.1.1 to solve graph colouring problem. Ülker et al. (2006) proposed special crossover operators for graph colouring, namely Lowest Index Max Crossover (LIMX), Greedy Partition Crossover Lowest In- dex (GPX-LI) and Greedy Partition Crossover Cardinality Based (GPX-CB). Külahçıoğlu (2007) proposed two modified versions of the LLE representation which are Linear Linkage Encoding With Ending Node Links (LLE-e) and Linear Linkage Encoding With Backward Links (LLE-b), and both of them are tested using genetic operators. Graph colouring underpins a variety of real-world problems. Examination timetabling is one of those problems which requires allocation of periods/time-slots and other available resources to a given set of examinations taken by a number of students subject to a set of certain constraints. This problem can be formulated as a graph colouring problem, considering that the examinations are the nodes of a graph while the edges are formed using the the examinations that will be taken by each student, since those examinations taken by each student cannot be allocated the same time-slot (colour). Johnson and Trick (1996) showed that the exam timetabling problem can be reduced to graph colouring problem if the task of minimizing the number of exam periods and removing the clashes are considered. There are many variants of examination timetabling problems. Again, many different heuris- tic approaches have been the focus of previous studies due to inherent difficulty in solving timetabling problems, ranging from ordering of graph colouring heuristics (Carter and Laporte, 1996) to evolutionary approaches (Burke and Newall, 1998; Paquete et al., 2001). The same formulation of an examination timetabling problem in (Carter et al., 1996) is used in this study. The relevant instances are identified as T oronto a in (Qu et al., 2009). Caramia et al. (2001) proposed a family of local search-based timetabling algorithms that apply an optimisation step after each exam allocation attempting to minimize the number of time slots (groups) and the overall penalty simultaneously. Merlot et al. (2002) presented a hybrid algorithm in which a hill climbing phase and a simulated annealing phase are used to improve an initial solution that was developed using a constraint programming phase. Ülker et al. (2006) formulated the problem as a grouping problem and used the linear linkage encoding presented in section 2.1.1 to test a variety of evolutionary approaches on this problem. More on examination timetabling can be found in (Carter et al., 1996; Qu et al., 2009). 2.3. Selection Hyper-Heuristics A selection hyper-heuristic is a heuristic that explores the space of heuristics formed by set of low level heuristics, each of which performs a search over the solutions while solving a given 6 problem. Özcan et al. (2008) identified the importance of two successive tasks that a selection hyper-heuristic performs: 1. Heuristic selection: a low level heuristic is selected from the low level heuristics set and applied to the current solution, and 2. Move acceptance: a decision is made about whether to accept or reject the resulting solu- tion. Different combination of selection hyper-heuristic components could yield a different overall performance. High Level Strategy: Manages the sets of low level heuristics. Domain barrier Problem Domain n:Problem Domain 1: Heuristics set: H1 H2 Hp >Representation >Instances >Evaluation Function >Others.. Heuristics set: H1 H2 Hq >Representation >Instances >Evaluation Function >Others.. ... Figure 2: A typical selection hyper-heuristic framework. There is a conceptual “domain barrier” between the high level selection hyper-heuristic and domain as illustrated in Figure 2. This barrier acts as an information filter and does not allow any problem specific information flow from the low level to high level. This feature is impor- tant, since it raises the level of generality of the designed approach operates at and supports re-usability of selection hyper-heuristics and their components, directly. Various learning and non-learning methods have been suggested for the heuristic selection task. The simplest method used in the literature within the context of selection hyper-heuristics is based on random choice. Simple Random (SR) selects one low level heuristic at each iteration purely randomly; i.e. each low level heuristic has an equal chance of being selected regardless to its previous performance. There are some variants of SR. For example, random permutation (RP) applies the low level heuristics in a specific order that is randomly determined at the beginning of the search (Bai and Kendall, 2005; Burke et al., 2005; Cowling and Chakhlevitch, 2003). Some heuristic selection methods utilise learning approaches that use feedback during the search process while the algorithm is running, such as the reinforcement learning (RL) (Burke and Soubeiga, 2003). RL assigns a score to each low level heuristic, and then rewards the low level heuristics which improve the current solution and punishes those which do not. This is achieved by means of increasing or decreasing the scores of heuristics. There are different ways of using the scores for choosing a low level heuristic (Burke et al., 2013; Chakhlevitch and Cowl- ing, 2008). One common method is making this decision using max utility function. At each 7 decision point, the low level heuristic with the highest score is chosen. If more than one heuristic share the same score, then one of them is selected at random. A modified version of the reinforce- ment learning (RLM) gives an extra reward to the low level heuristic, if the produced solution was accepted by the move acceptance criteria. Other than these two approaches, researchers have used more sophisticated low level heuristic selection approaches, including meta-heuristics such as GAs, tabu search and great deluge. Hyper-heuristics embedding meta-heuristic com- ponents are shown to be very effective and often beat state-of-the-art problem-tailored methods (Chakhlevitch and Cowling, 2008). Another learning heuristic selection method is referred to as the Adaptive Dynamic Heuristic Set selection (DH), which was a component used under an Intelligent Hyper-heuristic Framework built as a general problem solver (Misir et al., 2013). The importance of this hyper-heuristic stems from the fact that it was the winner of an international competition: The First Cross- Domain Heuristic Search Challenge (CHeSC2011) (Burke et al., 2011) that was based on a software benchmark framework known as HyFlex (Burke et al., 2009). Hyflex is a modular Java class library that has recently been implemented to facilitate the development of hyper-heuristics and the design of cross-domain heuristic search methods. DH uses several adaptive features to eliminate subsets of the low level heuristics given, determine effective heuristic pairs and adapt the parameters of some heuristics online in order to cope with the requirements of managing different heuristic sets. Similarly, various methods are suggested to perform move acceptance. Some of those meth- ods are fairly simple and deterministic mechanisms, such as the (AM) method accepts all moves, (OI) accepts only improving moves; and (IEQ) accepts improving or equal moves (Bilgin et al., 2007; Cowling et al., 2000). There are other successful move acceptance methods used within hyper-heuristics which are based on meta-heuristic approaches. The Great Deluge (GDEL) acceptance attempts to escape local optima by accepting moves that worsens the current solution within a range that is determined by a time-varying threshold, τt. This threshold starts with a certain value and decreases with time. τt = f0 + ∆F × (1 − t T ) (6) In Equation 6, T is the maximum number of steps, ∆F is an expected range for the maximum fitness change, and f0 is the final objective value (Dueck, 1993). The Late Acceptance (LACC) method is an iterative search technique that was proposed in (Burke and Bykov, 2008). In LACC, improving solutions are directly accepted, while a new worsening solution is compared to an old solution which was visited a fixed number of steps before for acceptance. If the new solution is better than that old solution, it is accepted, even if it could be worse than the current solution from which the new solution is produced. LACC maintains a queue (FIFO list) with a fixed length of solution costs. The initial queue is completely filled by replicating the initial solution’s cost. After each improving solution, its cost is enqueued (inserted). Whenever a worsening solution is encountered, then the cost of the best solution is enqueued. The front item always gets deleted after an enqueue operation. The Iteration Limited Threshold Accepting (ILTA) is the acceptance method used in the In- telligent Hyper-heuristic Framework (Misir et al., 2013). ILTA is a list-based threshold move acceptance mechanism that provide an adaptive diversification strategy in connection with the quality of the explored new best solutions. Instead of immediately accepting a worsening solu- tion either within a given range of cost such as is the case in the GDEL, or by comparing it to an old solution such as in the LACC, ILTA waits for a predetermined number of iterations, and 8 if no new best solutions can be found during this waiting period, then a worsening solution is accepted. Moreover, the algorithm adaptively changes the waiting times and the length of the list during the search in order to fully capture the nature of the problem being solved. More on hyper-heuristics can be found in (Burke et al., 2013; Chakhlevitch and Cowling, 2008). 3. Grouping Hyper-heuristic Framework In this study, we describe a single-point based selection hyper-heuristic framework for group- ing problems that keeps track of the non-dominated solutions, inspired from the multi-objective optimisation approaches, at each step. Figure 3 shows the layout of our framework. In addition to the common hyper-heuristic domain barrier that ensures the generality of the hyper-heuristic approach, we added the solution representation barrier at which all different types of grouping problems are encoded using a particular representation, and hence they are all treated the same way by the hyper-heuristic framework. As a result, only the definition of the cost function and the problem instances need to be fixed when applying the framework to a grouping problem. The solutions representation along with the low level heuristics remain the same across all domains. Domain Barrier High Level Strategy Move AcceptanceHeuristic Selection Low Level Heuristics Grouping Solver Solution Representation Problem Domain Domain 1: > Instances. > Evaluation Function Domain n: Current Solutions Pareto Front Best Solutions Pareto Front Initialization Heuristic Delta Evaluation ……. > Instances. > Evaluation Function > Instances. > Evaluation Function Domain 2:Domain 1: Figure 3: Framework of Hyper-heuristics for Grouping Problems. Single point-based selection hyper-heuristic frameworks start the search process from an initial solution and deal with a single solution to the problem being solved at each step. On the other hand, most of the multi-objective optimisation algorithms are population-based approaches (Zitzler and Thiele, 1998; Coello et al., 2007). While designing our framework, we have made an attempt to exploit the bi-objective nature of the grouping problems in order to capture the best of the two worlds. The proposed framework is appropriate only for grouping problems not for multi-objective optimisation (which is not within the scope of this study). 9 3.1. Delta Evaluation In the context of grouping problems as well as many similar combinatorial optimisation prob- lems, computing the cost of new solutions is the most time consuming part of the hyper-heuristic run. At each decision point, the solution resulting from the application of the selected heuristic has to be evaluated before a decision about whether to accept or reject it is made. Typically, this evaluation is carried out using the whole solution; i.e. calculating the partial cost of each group in the given solution. However, especially with large problem instances, this process is time consuming and greatly affects the solvers with time-based stopping criteria. In this study, we used delta evaluation, which requires computation of partial cost contribu- tions of the groups that have been affected by the application of the selected low level heuristic. This would give a significant time advantage and consequently allow more iterations in each step. For example, if a change heuristic is applied on a solution Uold , and ui and u j are the groups that are involved in the process, the overall cost value f (Unew) of the resulting solution, Unew, is calculated using the equation below, where u ′ i and u ′ j are the resulting groups after the heuristic is applied: f (Unew) = f (Uold ) − ( f (ui) + f (u j)) + ( f (u ′ i ) + f (u ′ j)) 3.2. Low Level Heuristics It has been observed that the use of crossover operators are found to be very disruptive and tends to impair the search rather than guide it for grouping problems in (Falkenauer, 1998; Ko- rkmaz, 2010). Additionally, the proposed framework performs single-point based search, hence, crossover operators are ignored. We have considered a set of effective mutation operators that process complete solutions. Some of them are able to create reasonably large changes on a given solution, diversifying the search process. Some of them are smart operators which are enabled to make small modifications on a given solution leading to a potentially better/improved solution, intensifying the search process. Three different types of mutation operators were consequently developed: merge, divide and change. Merge Heuristics. The concept of this family of heuristics is to merge two groups, ui and u j, into one, ul. This operation results in decreasing the number of grouping in the selected solution. The cost value of the new subgroup ul is greater than or equal to the combined cost values of ui and u j; i.e. f (ul) ≥ f (ui) + f (u j). We implemented three versions of the merge heuristic in this study, which differ from each other in the way they choose the groups to be merged in a given solution. • M1 merges two randomly selected groups, • M2 merges two groups that contain the least number of items, and • M3 merges two groups with the lowest partial cost values. Hence, merge heuristics can be considered as diversifying components. Divide Heuristics. In a similar fashion, three divide heuristics were implemented in this study, each of which divides a selected group ui into two groups, ui1 and ui2. Applying a divide heuristic results in increasing the number of grouping in the selected solution. Also, some of the conflict- ing items in ui may end up in different groups, which leads to the elimination of some conflicts. Consequently, the combined cost values of ui1 and ui2 is less than or equal to the cost value of ui; i.e. f (ui1) + f (ui2) ≤ f (ui). 10 • D1 divides a randomly selected group, • D2 divides the group which contains the largest number of items and, • D3 divides the group with the highest partial cost value in the selected solution. In all the divide heuristics, each item has equal probability of ending up in either group. Applying any of these heuristics result in increasing the number of the groups and, hopefully, decreasing the number of conflicts in the selected solution. This characteristic of divide heuristics is of interest as an intensification component and it is used to design a local search algorithm to improve the non-dominated set of solutions within our methodology (see Section 3.3). Change Heuristics. Merge and divide heuristics produce relatively big jumps in the search space considering that they definitely influence the number of resultant groups for a given solution. However, the change heuristics attempts to make small modifications in a given solution by changing the group of a selected item while maintaining the same number of the groups after they are employed. This family of heuristics has four members. • C1 takes a random item xm from a randomly selected group ui and moves it into another randomly selected group u j. • C2 finds the item that is causing the highest number of conflicts in a randomly selected group and moves it to another random group. • C3 finds the group with the highest number of conflicts, and from that group it takes the item that is causing the highest number of conflicts and moves it to a randomly selected group. • C4 is similar to the previous one, except that the item is moved to the group with the minimum number of conflicts rather than a random one. 3.3. Methodology The pseudo-code of the proposed approach is provided in Algorithm 1. Firstly, a random solution is created for each k in a given range [LB, U B] in order to have a set of initial solutions with different number of groupings. This range is arbitrarily decided in our experiments aligned with previous work (Ülker et al., 2006). Depending on the particular grouping problem being solved, reasonable upper and lower bounds can automatically be found by using problem-specific knowledge. For instance, for graph colouring problems, these bounds can be determined using algorithms such as the fast maximal clique approximation algorithm or by finding the maximal degree of the graph (the degree of the vertex with the maximum number of neighbours), as discussed in (Ülker et al., 2006). The initialisation heuristic as a part of the problem domain implementation can be a “smart” problem-specific algorithm. However, in order to maintain a high level of generality, we have not used any such problem-specific initialisation. On the other hand, we aimed at creating a non-dominated initial set of solutions by using an initialisation heuristic embedding a problem independent “smart” move appropriate for grouping problems. Starting from k = LB through k = U B, a population of (U B − LB + 1) non-dominated solutions are produced, where each solution is created for each integer value of k ∈ [LB, U B]. In order to guarantee non-dominance, each solution for a given k = i is generated at random by assigning each item to one of the groups, initially. The cost for the solution is immediately 11 computed and compared to the cost of the solution at k = i − 1. Since the solution at k = i is already worse in terms of the number of groups compared to the solution at k = i − 1, its cost value must be better. If this is not the case, that solution is discarded and re-created either randomly or by dividing one of the groups of the solution at k = i − 1. The final outcome of the initialisation phase is a non-dominated set of solutions (lines 1-3). Once the initialisation phase is over, the framework proceeds to select one of the low level heuristics (section 3.2) and apply it on a random solution from the set of non-dominated solutions (Algorithm 1, lines 5-8). Maintaining the non-dominated set during the search requires that, when comparing any two solutions, the solution with the worse (better) number of groups should have the better (worse) cost value, i.e. f (Ui) > f (U j),∀ i < j, (∀i, j ∈ [LB, U B]). However, during the search (Algorithm 1, lines 5-8), it is possible that a newly created solution might violate this requirement when compared to the other solutions in the non-dominated set, i.e. the new solution might be either better or worse in both objectives when compared to some other solutions in the non-dominated set. We overcome this problem by introducing a case-based acceptance mechanism. The move acceptance of the hyper-heuristic component does not make the final call for the acceptance of a solution and can be considered as a pre-test component for final acceptance. The new solution is only accepted after passing multiple tests. Firstly, the hyper-heuristic move acceptance criteria compares the new solution snew to the current solution si in order to make a decision regarding whether to consider the new solution for acceptance or reject it immediately (Algorithm 1, line 9). This is different from how the tradi- tional hyper-heuristic framework operate, in which the decision made by the move acceptance is final. The hyper-heuristic move acceptance methods we used in this study are non-deterministic, i.e. all improving solutions are considered for acceptance, while some of the worsening ones may or may not be considered. We differentiate between two main cases and the second case allows the use of local search to improve the non-dominated set of solutions further: 1. If snew is considered for acceptance despite being worse than si in terms of cost value (Algorithm 1, line 10), then it is compared to si−1. snew is rejected if it is worse than si−1. Otherwise, it is accepted and inserted into the non-dominated set to replace si (Algorithm 1, lines 11-16). 2. If snew is considered for acceptance and its cost value is better than, or equal to, the cost value of si (Algorithm 1, line 17), then it is accepted and inserted into the non-dominated set to replace si. A violation to the dominance rule may occur if si+1, which is already worse than si in terms of the number of groups, turns out to be also worse in terms of the cost value. In order to avoid this, snew is then compared to si+1 (Algorithm 2, line 2), which creates two cases: 2.1 If snew is worse than si+1, then there are no violations in the dominance rule, and no more action is needed (Algorithm 2, lines 2-3). 2.2 If snew is better than si+1, then si+1 is in violation of the dominance rule and conse- quently it is removed from the non-dominated set. A replacement solution is created by dividing a group in snew using any of the divide heuristics (Algorithm 2, lines 4-8). The only remaining issue is that, this replacement solution at i + 1 might have a better cost value than the solution at i + 2, hence, the f or loop in Algorithm 2. In the worst case scenario, this algorithm will be applied on all the solutions in the non-dominated set between i and U B. This process can be considered as local search. 12 Algorithm 1 Hyper-heuristic framework for grouping problems 1: Generate an initial non-dominated set that contains a solution for each value of k ∈ [LB, U B]. 2: Compute the cost value of each solution in the initial non-dominated set. 3: Copy the initial non-dominated set into a an external archive to keep track of the best non- dominated solutions. 4: while (elapsedTime < maxTime) do 5: Choose a random solution s j from the current non-dominated set by j ← Uni f ormRandom(LB, U B). 6: Choose a low level heuristic, LLH. 7: snew ← Apply(LLH, s j) {snew will have i =( j − 1) or j or ( j + 1) number of groupings depending on the nature of LLH (merge or change or divide, respectively)}. 8: Compute f (snew) using delta evaluation. 9: result ← moveAcceptance(snew, si) {Use the hyper-heuristic acceptance method to com- pare the cost of snew to the cost of si from the current non-dominated set}. {Following is the case when new solution is a worsening solution which is accepted by moveAcceptance} 10: if ((result is ACCEPT ) and ( f (snew) > f (si))) then 11: if ( f (snew) > f (si−1)) then 12: Do nothing {snew is rejected} 13: else 14: si ← snew {snew is placed in the non-dominated set at grouping i, replacing si}. 15: end if 16: end if 17: if ((result is ACCEPT ) and ( f (snew) ≤ f (si))) then 18: si ← snew {snew is placed in the non-dominated set at grouping i, replacing si}. 19: improveNonDominatedS et(i) 20: end if {if result is RE JECT then do nothing and continue} 21: end while Algorithm 2 improveNonDominatedS et(i): Attempts to improve upon the cost of solutions in the non-dominated set starting from ith solution to U Bth solution using a divide heuristic 1: for ( j = i, U B) do 2: if ( f (s( j+1)) ≤ f (s j)) then 3: BREAK {Further improvement is not possible} 4: else 5: Choose a random divide heuristic, LLDH 6: snew ← Apply(LLDH, s j) 7: s( j+1) ← snew {snew is placed in the non-dominated set at grouping j, replacing s( j+1)}. 8: end if 9: end for 13 4. Experimental Results 4.1. Experimental Design and Evaluation Criteria In this study, we investigated the performance of nine selection hyper-heuristics, using all the combinations of the heuristic selection {SR, RL, DH} and move acceptance methods {ILTA, LACC, GDEL} on benchmark instances from the graph colouring and examination timetabling domains. RL uses increment and decrement score operations as a reward and punishment scheme, respectively. A heuristic with the maximum score is selected at each decision point. The incre- ment and decrement scores are set to 1. The upper and lower bounds for the score of any low level heuristic are set to 40 and 0 respectively. The initial score of each one of the low level heuristics is set to (upper bound − 2 ∗ number o f heuristics). DH and ILTA are implemented using the exactly the same suggested settings as in (Misir et al., 2013). LACC uses a queue of fixed size of 50 for each k ∈ [LB, U B]. The GDEL parameters shown in equation 6 are set to the following values: T is set to the maximum duration of a trial, ∆F is set to the minimum cost value in the initial non-dominated set and f0 is set to 0. Each experiment is repeated for 30 runs (trials), each of which stops when the best known colouring/timetable is attained or a time limit of 3600 seconds is exceeded. All initial solutions are created randomly. Experiments were conducted on 3.6GHz Intel Core i7−3820 machines with 16.0GB of memory, running Windows 7 operating system. The proposed approach cannot be used for multi-objective optimisation, but it operates based on the dominance concept from the multi-objective optimisation. In minimum colouring, the aim is not just minimise the violations, the aim is get rid of all violation yielding 0 as the objective value. Still, in our approach we have archived a set of non-dominated solutions for a given range of number of colours/groupings. Hence, the performance of a given algorithm is evaluated based on the best (minimum) number of colours achieved with no violation and hyper-volume (Zitzler and Thiele, 1998) of the non-dominated solutions obtained by that algorithm. The hyper-volume is a commonly used metric to evaluate the spread of the solutions along the Pareto front, as well as the closeness of the solutions to the Pareto-optimal front. We used the cost of a fixed solution produced for the smallest allowed colouring for a given instance as a reference to compute the hyper-volume of the best non-dominated solutions obtained by each algorithm. The larger the hyper-volume, the larger the size of the space covered by the non-dominated set and better the corresponding hyper-heuristic approach. The Wilcoxon Signed Rank test is used as a statistical test for the average pairwise perfor- mance comparison of algorithms based on the minimum colouring they obtain over 30 runs. We have also used success rate (SRate%) as a performance indicator of a hyper-heuristic. SRate% indicates the percentage of 30 runs for which expected number of groups has been obtained by the given algorithm. 4.2. Experimental Data The characteristics of the benchmark instances used during the experiments are summarised in Table 1. For graph colouring, 19 benchmark instances are used in which the number of colours, vertices, edges as well as edge densities vary. The instances in the upper half of the table are from the COLOR02 website 2, which was initially compiled for a competition. Myciel graphs are based on Mycielski transformation and are considered to be difficult to solve and the colouring 2http://mat.gsia.cmu.edu/COLOR02/ 14 number increases in problem size. A queenn.n graph is a graph on n2 nodes, each represents a square on an n by n chessboard. If two squares are in the same row, column, or diagonal, then their corresponding nodes on the graph are considered to be connected. The objective is to place n sets of n queens on the board so that no two queens of the same set can capture one another, which could be achieved only if the graph has a colouring number n. In addition to these data sets, problem instances in the bottom half of the table are from the well known DIMACS3 challenge suite. For examination timetabling, we used a subset from the Toronto benchmark suite referred to as T oronto a instances (Qu et al., 2009). Table 1 also shows the range of k values used during the experiments for each instance. Initial experiments were conducted to observe the behaviour of the hyper-heuristic approach considering different k values for each problem instance, as specified in Table 1, while the heuris- tic selection is fixed from {SR, RL, DH} and the heuristic acceptance is fixed from {ILTA, LACC, GDEL}. A thorough performance analysis of the hyper-heuristics is performed. Then the per- formance of the hyper-heuristic with the best mean is compared to the performance of some pre- viously proposed approaches. The same framework using the top two hyper-heuristics, namely RL − I LT A and DH − I LT A, are tested further on examination timetabling problem. Table 1: The characteristics of the COLOR02, DIMACS and Toronto problem instances used during the experiments. |V| represents the number of vertices, |E| the number of edges, % the edge density and k∗/χ(G) represent the best known number of colours (or time-slots) and the chromatic number respectively Wu and Hao (2012). L and U represent the lower and upper bounds for the k values used during the experiments. Graph Colouring Range Instance |V| |E| % k∗/χ(G) L U C O L O R 02 myciel3 11 20 0.40 4/4 2 9 myciel4 23 71 0.28 5/5 2 10 myciel5 47 236 0.22 6/6 3 11 queen5.5 25 160 0.53 5/5 2 10 queen6.6 36 290 0.46 7/7 4 12 queen7.7 49 476 0.40 7/7 2 12 queen8.8 64 728 0.36 9/9 6 14 D IM A C S le450 25a 450 8260 0.08 25/25 20 30 le450 25b 450 8263 0.08 25/25 20 30 le450 25c 450 17343 0.17 25/25 20 30 le450 25d 450 17425 0.17 25/25 20 30 DSJC125.1 125 736 0.09 5/? 2 10 DSJC125.5 125 3891 0.50 17/? 13 23 DSJC125.9 125 6961 0.89 44/? 40 50 DSJC250.1 250 3218 0.10 8/? 4 14 DSJC250.5 250 15668 0.50 28/? 24 34 DSJC250.9 250 27897 0.90 72/? 68 78 DSJC500.1 500 12458 0.10 12/? 7 17 DSJC500.5 500 62624 0.50 48/? 43 53 Examination Timetabling Range Instance |V| |E| % k∗ L U T O R O N T O hec92 81 1363 0.42 17 12 22 sta83 139 1381 0.14 13 8 19 yor83 181 4691 0.29 19 13 25 ute92 184 1430 0.08 10 6 15 rye93 486 8872 0.08 21 16 27 4.3. Selection Hyper-heuristics for Graph Colouring Tables 2, 3 and 4 show the success rate of each hyper-heuristic approach for different values of k with each problem instance. All those results are obtained while optimising a given instance 3ftp://dimacs.rutgers.edu/pub/challenge/graph/benchmarks/color/ 15 Table 2: Reinforcement learning based selection hyper-heuristics: The success rate (sRate%) of each hyper-heuristic on the graph colouring problem instances and the average time (µt (s)) taken to achieve that success rate for a given number of colourings, k over 30 runs. RL-ILTA RL-LACC RL-GDEL Instance k sRate% µt (s) sRate% µt (s) sRate% µt (s) C O L O R 02 myciel3 4 100.00 0.003 100.00 0.004 93.33 0.002 myciel4 5 100.00 0.005 100.00 0.005 100.00 0.006 myciel5 6 100.00 0.009 100.00 0.011 100.00 0.069 queen5.5 5 100.00 0.034 90.00 0.037 100.00 0.194 queen6.6 7 100.00 0.786 20.00 1.449 100.00 111.831 queen7.7 7 100.00 0.900 16.67 0.040 100.00 39.350 queen8.8 9 100.00 4.110 3.33 0.150 100.00 74.530 D IM A C S DSJC125.1 5 96.67 113.20 6.67 6.71 100.00 295.58 6 100.00 0.09 100.00 0.17 100.00 55.02 7 100.00 0.03 100.00 0.04 100.00 50.18 DSJC125.5 17 33.33 470.96 0.00 − 0.00 − 18 100.00 5.68 60.00 115.70 66.67 655.66 19 100.00 0.85 100.00 1.80 100.00 82.63 DSJC125.9 44 100.00 40.02 80.00 69.40 0.00 − 45 100.00 4.50 100.00 27.05 10.00 538.47 46 100.00 1.58 100.00 3.90 86.67 422.53 DSJC250.1 8 10.00 1131.10 0.00 − 0.00 − 9 100.00 2.35 100.00 7.46 100.00 194.28 10 100.00 0.44 100.00 0.60 100.00 175.16 DSJC250.5 28 0.00 − 0.00 − 0.00 − 29 20.00 2132.40 0.00 − 0.00 − 30 100.00 288.20 20.00 1164.80 0.00 − DSJC250.9 72 10.00 3301.23 0.00 − 0.00 − 73 100.00 811.32 40.00 886.41 0.00 − 74 100.00 177.44 80.00 648.03 0.00 − DSJC500.1 12 0.00 − 0.00 − 0.00 − 13 60.00 1115.87 20.00 1331.99 0.00 − 14 100.00 26.49 100.00 43.70 0.00 − DSJC500.5 48 0.00 − 0.00 − 0.00 − 49 0.00 − 0.00 − 0.00 − 50 0.00 − 0.00 − 0.00 − 52 10.00 1797.03 0.00 − 0.00 − le450 25a 25 100.00 7.85 100.00 25.33 0.00 − 26 100.00 1.56 100.00 2.81 83.33 472.34 27 100.00 0.82 100.00 1.94 100.00 109.32 le450 25b 25 100.00 11.85 100.00 6.22 6.67 569.68 26 100.00 1.19 100.00 2.55 100.00 223.61 27 100.00 0.65 100.00 1.66 100.00 73.37 le450 25c 25 0.00 − 0.00 − 0.00 − 26 0.00 − 0.00 − 0.00 − 27 76.67 565.69 40.00 873.93 0.00 − le450 25d 25 0.00 − 0.00 − 0.00 − 26 0.00 − 0.00 − 0.00 − 27 70.00 654.47 30.00 906.22 0.00 − within the associated range as provided in Table 1. For example, the best known colouring for the DSJC500.1 problem instance is 12 and the algorithms are tested with 7 ≤ k ≤ 17. From Table 2, out of the 30 runs performed using the RL-ILTA hyper-heuristic on DSJC500.1 problem instance, the percentage of the runs in which solutions with the best colourings (k = 12) were found is 0%, while the percentage of the runs in which solutions with k = 13 colours are found is 60% and the percentage of the runs in which solutions with k = 14 colours are found is 100%. These tables also show the average time (in seconds) taken to achieve those success rates. Only the durations of the successful runs were taken into consideration when the average times were calculated. Most of the tested hyper-heuristic approaches successfully found the best colourings of the selected COLOR02 problem instances in a reasonable amount of time, and hence, we provide 16 Table 3: Adaptive dynamic heuristics set based selection hyper-heuristics: The success rate (sRate%) of each hyper- heuristic on the graph colouring problem instances and the average time (µt (s)) taken to achieve that success rate for a given number of colourings, k over 30 runs. DH-ILTA DH-LACC DH-GDEL Instance k sRate% µt (s) sRate% µt (s) sRate% µt (s) C O L O R 02 myciel3 4 100.00 0.006 100.00 0.006 96.67 0.004 myciel4 5 100.00 0.015 100.00 0.018 100.00 0.016 myciel5 6 100.00 0.044 100.00 0.054 100.00 3.592 queen5.5 5 100.00 0.167 100.00 1.053 100.00 5.532 queen6.6 7 56.67 290.060 43.33 86.180 93.33 204.640 queen7.7 7 50.00 304.040 10.00 432.160 86.67 152.420 queen8.8 9 60.00 305.880 26.67 224.680 86.67 335.830 D IM A C S DSJC125.1 5 30.00 618.95 6.67 112.04 90.00 533.55 6 100.00 0.48 100.00 12.93 100.00 39.25 7 100.00 0.16 100.00 0.37 100.00 32.14 DSJC125.5 17 13.33 821.69 0.00 − 0.00 − 18 93.33 228.10 76.67 146.32 60.00 562.04 19 100.00 14.66 100.00 12.02 96.67 226.22 DSJC125.9 44 86.67 146.55 63.33 183.39 0.00 − 45 100.00 29.08 93.33 32.98 10.00 537.80 46 100.00 9.76 100.00 7.67 43.33 464.38 DSJC250.1 8 30.00 1047.59 0.00 − 0.00 − 9 100.00 71.23 100.00 59.16 100.00 355.09 10 100.00 1.99 100.00 1.90 100.00 168.32 DSJC250.5 28 0.00 − 0.00 − 0.00 − 29 20.00 3042.47 0.00 − 0.00 − 30 80.00 1704.74 40.00 913.88 0.00 − DSJC250.9 72 0.00 − 0.00 − 0.00 − 73 60.00 2281.95 50.00 1357.03 0.00 − 74 90.00 865.49 90.00 667.89 0.00 − DSJC500.1 12 0.00 − 0.00 − 0.00 − 13 70.00 1115.84 80.00 496.77 0.00 − 14 100.00 23.69 100.00 53.11 0.00 − DSJC500.5 48 0.00 − 0.00 − 0.00 − 49 0.00 − 0.00 − 0.00 − 50 0.00 − 0.00 − 0.00 − 52 3.33 2553.04 0.00 − 0.00 − le450 25a 25 100.00 85.82 96.67 36.19 3.33 1019.41 26 100.00 3.15 100.00 5.27 86.67 345.32 27 100.00 1.80 100.00 3.38 100.00 147.13 le450 25b 25 100.00 7.58 100.00 29.14 20.00 649.71 26 100.00 2.13 100.00 3.75 100.00 211.16 27 100.00 1.40 100.00 3.16 100.00 101.29 le450 25c 25 0.00 − 0.00 − 0.00 − 26 0.00 − 0.00 − 0.00 − 27 83.33 626.85 30.0 2346.70 0.00 − le450 25d 25 0.00 − 0.00 − 0.00 − 26 0.00 − 0.00 − 0.00 − 27 63.33 704.38 36.67 665.24 0.00 − the success rates and average times for only the best known value of k for these instances. The RL-ILTA and the SR-GDEL hyper-heuristics successfully found the best colouring for each data set in each one of the 30 runs, achieving 100.0% success rate across the board, with a maximum average time of 4.1 seconds for the RL-ILTA approach. On the other hand, hyper-heuristics with LACC acceptance failed to find the best colourings for some problem instances in most of the runs. For example, the SR-LACC and the RL-LACC hyper-heuristics found the best colourings of queen7.7 and queen8.8 problem instances respectively in only one of the 30 runs performed. The performances of the hyper-heuristic approaches were much varied and less successful when applied on the selected DIMACS problem instances, and much longer periods of times were needed to find the best solutions in most of the cases. Generally, hyper-heuristics with ILTA acceptance have better success rates than hyper-heuristics with LACC and GDEL acceptance 17 Table 4: Simple random based selection hyper-heuristics: The success rate (sRate%) of each hyper-heuristic on the graph colouring problem instances and the average time (µt (s)) taken to achieve that success rate for a given number of colourings, k over 30 runs. SR-ILTA SR-LACC SR-GDEL Instance k sRate% µt (s) sRate% µt (s) sRate% µt (s) C O L O R 02 myciel3 4 100.00 0.007 100.00 0.004 100.00 0.004 myciel4 5 100.00 0.007 100.00 0.003 100.00 0.008 myciel5 6 100.00 0.011 100.00 0.015 100.00 0.398 queen5.5 5 100.00 0.016 96.67 0.017 100.00 0.687 queen6.6 7 33.33 0.157 20.00 0.780 100.00 83.140 queen7.7 7 16.67 0.261 3.33 0.038 100.00 43.669 queen8.8 9 30.00 2.032 10.00 1.232 100.00 69.154 D IM A C S DSJC125.1 5 3.33 450.25 0.00 − 96.67 374.43 6 100.00 0.13 100.00 0.14 100.00 53.92 7 100.00 0.06 100.00 0.07 100.00 51.31 DSJC125.5 17 20.00 756.87 0.00 − 0.00 − 18 93.33 117.65 53.33 41.71 90.00 480.85 19 100.00 1.06 100.00 1.88 100.00 92.52 DSJC125.9 44 90.00 72.28 86.67 57.01 0.00 − 45 100.00 13.79 100.00 5.28 23.33 340.94 46 100.00 1.87 100.00 3.20 96.67 432.36 DSJC250.1 8 0.00 − 0.00 − 0.00 − 9 100.00 5.58 100.00 4.36 100.00 176.98 10 100.00 0.57 100.00 0.74 100.00 175.51 DSJC250.5 28 0.00 − 0.00 − 0.00 − 29 0.00 − 0.00 − 0.00 − 30 60.00 1350.55 20.00 650.08 0.00 − DSJC250.9 72 0.00 − 0.00 − 0.00 − 73 50.00 2264.24 60.00 1547.57 0.00 − 74 100.00 649.87 90.00 668.56 0.00 − DSJC500.1 12 0.00 − 0.00 − 0.00 − 13 30.00 1005.16 40.00 913.15 0.00 − 14 100.00 27.99 100.00 30.55 0.00 − DSJC500.5 48 0.00 − 0.00 − 0.00 − 49 0.00 − 0.00 − 0.00 − 50 0.00 − 0.00 − 0.00 − 53 3.33 1944.04 0.00 − 0.00 − le450 25a 25 100.00 6.36 100.00 33.61 0.00 − 26 100.00 1.52 100.00 3.34 76.67 518.65 27 100.00 0.89 100.00 2.52 100.00 99.39 le450 25b 25 100.00 3.10 100.00 7.70 0.00 − 26 100.00 1.16 100.00 3.03 100.00 221.37 27 100.00 0.67 100.00 1.95 100.00 83.17 le450 25c 25 0.00 − 0.00 − 0.00 − 26 0.00 − 0.00 − 0.00 − 27 40.00 1395.13 16.67 783.49 0.00 − le450 25d 25 0.00 − 0.00 − 0.00 − 26 0.00 − 0.00 − 0.00 − 27 60.00 2326.19 30.00 2230.95 0.00 − methods in most of the cases. ILTA based hyper-heuristics have outperformed their LACC and GDEL counterparts in all problem instances for all values of k with the exception of DSJC125.1 at k = 5, DSJC250.9 at k = 73 and DSJC500.1 at k = 13. In the same way, GDEL hyper- heuristics have the worst average times and success rates on most of the DIMACS instances compared to the rest of the hyper-heuristics. Table 5 and shows the average best colouring and standard deviation for each hyper-heuristic approach on each problem instance across the 30 runs. ±0.0 standard deviation corresponds to 100.0% success rate. The row denoted ‘Wins’ shows the number of problem instances in which the corresponding hyper-heuristic approach achieved the best average colouring including ties with the other algorithms. From this table it can be seen that RL-ILTA hyper-heuristic has the most number of wins across all the tested approaches. Hence, we used the performance of 18 the RL-ILTA hyper-heuristic as a reference to carry out statistical tests in order to determine how significant the differences in the best colourings found by each of the tested hyper-heuristic approaches with regard to the RL-ILTA approach are. ● ●● ●● ●●● ● ● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 5.0 5.2 5.4 5.6 5.8 6.0 DSJC125.1.col ●●●● ●● ●●●●●●● ●●●●●● ●● ●●● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 17.0 17.5 18.0 18.5 19.0 19.5 20.0 DSJC125.5.col ●●●●●● ●●● ●●●● ●●●● ●●● ●●●● ●●●●●●● ● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 44 45 46 47 48 DSJC125.9.col ●●● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 8.0 8.2 8.4 8.6 8.8 9.0 DSJC250.1.col ●●●●●● ●●●●●● ●●●●●● ●●●●●● ●●●●●● ●●● ●●●●●● ●●● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 29 30 31 32 33 34 DSJC250.5.col ●●● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 72 74 76 78 80 82 DSJC250.9.col ●●●●●● ●●●●●● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 13 14 15 16 17 18 DSJC500.1.col ●●● ●●● ●●● ●●● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 52 54 56 58 60 DSJC500.5.col ●●●●● ● ● ●●●● ●●●●●●● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 25.0 25.5 26.0 26.5 27.0 le450_25a.col ●● ●●●●●● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 25.0 25.2 25.4 25.6 25.8 26.0 le450_25b.col ●●●●●●● ●●●●● ●●●●● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 28 30 32 34 le450_25c.col RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 28 30 32 34 le450_25d.col Figure 4: Box plots of the best number of colours achieved by each hyper-heuristic approach on selected DIMACS data sets. 1, 2 and 3 in the hyper-heuristic approaches names refer to ILTA, LACC and GDEL acceptance methods respectively Table 6 shows the pairwise performance comparison of hyper-heuristics based on the Wilcoxon Signed Rank statistical test using best colourings attained by RL-ILTA hyper-heuristic as a ref- erence. Almost all the results obtained by the GDEL based hyper-heuristics are significantly worse than the results of the RL-ILTA hyper-heuristic, with the exception of RL-GDEL on DSJC125.1, for which case, RL-GDEL performs significantly better than RL-ILTA. Similarly, the RL-ILTA hyper-heuristic performs either significantly or slightly better than the LACC based hyper-heuristics in almost all of the cases with the exception of DH-LACC on DSJC500.1, for which case DH-LACC performs significantly better than the RL-ILTA. One of the observations is that the DH-ILTA and RL-ILTA hyper-heuristics deliver a compet- itive performance. On DSJC250.1, DSJC500.1 and le450 25c, DH-ILTA performs significantly better than RL-ILTA. This has been expected bearing in mind that the DH-ILTA is known to be a very powerful algorithm in cross domain search (Misir et al., 2013). However, on more instances, namely DSJC125.1, DSJC125.9, DSJC250.9, DSJC500.5 and le450 25d, RL-ILTA outperforms DH-ILTA. This performance difference is statistically significant. Additionally, on 4 other in- stances, RL-ILTA performs slightly better than DH-ILTA. They have a tie on six benchmark instances. To better evaluate the differences between the performances of the 9 hyper-heuristics on the 19 Table 5: Average best colouring and standard deviation of each hyper-heuristic approach on each problem instance across the 30 runs. RL-ILTA RL-LACC RL-GDEL Instance k∗ µ(kbest ) σ(kbest ) µ(kbest ) σ(kbest ) µ(kbest ) σ(kbest ) myciel3 4 4.0 ±0.0 4.0 ±0.0 4.07 ±0.25 myciel4 5 5.0 ±0.0 5.0 ±0.0 5.0 ±0.0 myciel5 6 6.0 ±0.0 6.0 ±0.0 6.0 ±0.0 queen5.5 5 5.0 ±0.0 5.1 ±0.31 5.0 ±0.0 queen6.6 7 7.0 ±0.0 7.8 ±0.41 7.0 ±0.0 queen7.7 7 7.0 ±0.0 8.4 ±0.77 7.0 ±0.0 queen8.8 9 9.0 ±0.0 9.97 ±0.18 9.0 ±0.0 DSJC125.1 5 5.03 ±0.18 5.93 ±0.25 5.0 ±0.0 DSJC125.5 17 17.67 ±0.48 18.4 ±0.50 18.33 ±0.48 DSJC125.9 44 44.0 ±0.0 44.2 ±0.41 46.03 ±0.49 DSJC250.1 8 8.9 ±0.30 9.0 ±0.0 9.0 ±0.0 DSJC250.5 28 29.8 ±0.41 30.8 ±0.41 32.6 ±0.50 DSJC250.9 72 72.9 ±0.31 73.8 ±0.76 79.4 ±0.50 DSJC500.1 12 13.4 ±0.50 13.8 ±0.41 15.5 ±0.51 DSJC500.5 48 53.6 ±0.67 56.5 ±0.51 57.4 ±0.50 le450 25a 25 25.0 ±0.0 25.0 ±0.0 26.17 ±0.38 le450 25b 25 25.0 ±0.0 25.0 ±0.0 25.93 ±0.25 le450 25c 25 27.23 ±0.43 27.6 ±0.50 32.6 ±0.50 le450 25d 25 27.3 ±0.47 27.7 ±0.47 33.4 ±0.50 Wins 18 5 7 DH-ILTA DH-LACC DH-GDEL Instance k∗ µ(kbest ) σ(kbest ) µ(kbest ) σ(kbest ) µ(kbest ) σ(kbest ) myciel3 4 4.0 ±0.0 4.0 ±0.0 4.1 ±0.55 myciel4 5 5.0 ±0.0 5.0 ±0.0 5.0 ±0.0 myciel5 6 6.0 ±0.0 6.0 ±0.0 6.0 ±0.0 queen5.5 5 5.0 ±0.0 5.0 ±0.0 5.0 ±0.0 queen6.6 7 7.43 ±0.50 7.57 ±0.50 7.07 ±0.25 queen7.7 7 7.7 ±0.79 8.37 ±0.67 7.17 ±0.46 queen8.8 9 9.4 ±0.50 9.73 ±0.45 9.13 ±0.35 DSJC125.1 5 5.7 ±0.47 5.93 ±0.25 5.1 ±0.31 DSJC125.5 17 17.93 ±0.45 18.23 ±0.43 18.43 ±0.57 DSJC125.9 44 44.13 ±0.35 44.43 ±0.63 46.6 ±0.86 DSJC250.1 8 8.7 ±0.47 9.0 ±0.0 9.0 ±0.0 DSJC250.5 28 30.0 ±0.64 30.6 ±0.50 32.2 ±0.89 DSJC250.9 72 73.5 ±0.68 73.6 ±0.67 80.57 ±0.57 DSJC500.1 12 13.3 ±0.47 13.2 ±0.41 16.3 ±0.65 DSJC500.5 48 53.67 ±0.71 54.5 ±0.82 58.77 ±1.01 le450 25a 25 25.0 ±0.0 25.03 ±0.18 26.1 ±0.40 le450 25b 25 25.0 ±0.0 25.0 ±0.0 25.8 ±0.41 le450 25c 25 27.17 ±0.38 27.7 ±0.47 33.97 ±0.81 le450 25d 25 27.4 ±0.56 27.63 ±0.49 34.13 ±0.68 Wins 14 6 7 SR-ILTA SR-LACC SR-GDEL Instance k∗ µ(kbest ) σ(kbest ) µ(kbest ) σ(kbest ) µ(kbest ) σ(kbest ) myciel3 4 4.0 ±0.0 4.0 ±0.0 4.0 ±0.0 myciel4 5 5.0 ±0.0 5.0 ±0.0 5.0 ±0.0 myciel5 6 6.0 ±0.0 6.0 ±0.0 6.0 ±0.0 queen5.5 5 5.0 ±0.0 5.03 ±0.18 5.0 ±0.0 queen6.6 7 7.67 ±0.48 7.8 ±0.41 7.0 ±0.0 queen7.7 7 8.23 ±0.73 8.53 ±0.57 7.0 ±0.0 queen8.8 9 9.7 ±0.47 9.9 ±0.31 9.0 ±0.0 DSJC125.1 5 5.97 ±0.18 6.0 ±0.0 5.03 ±0.18 DSJC125.5 17 17.87 ±0.51 18.47 ±0.51 18.1 ±0.31 DSJC125.9 44 44.1 ±0.31 44.13 ±0.35 45.8 ±0.48 DSJC250.1 8 9.0 ±0.0 9.0 ±0.0 9.0 ±0.0 DSJC250.5 28 30.4 ±0.50 30.9 ±0.55 31.9 ±0.55 DSJC250.9 72 73.5 ±0.51 73.5 ±0.68 80.5 ±0.51 DSJC500.1 12 13.7 ±0.47 13.6 ±0.50 16.9 ±0.71 DSJC500.5 48 54.57 ±0.73 54.9 ±0.71 59.0 ±0.64 le450 25a 25 25.0 ±0.0 25.0 ±0.0 26.23 ±0.43 le450 25b 25 25.0 ±0.0 25.0 ±0.0 26.0 ±0.0 le450 25c 25 27.6 ±0.50 27.83 ±0.38 34.13 ±0.73 le450 25d 25 27.4 ±0.50 27.7 ±0.47 34.2 ±0.61 Wins 14 8 9 20 Table 6: Wilcoxon Signed Rank statistical test using RL-ILTA as a reference for the comparison: ‘>’ (‘<’) denotes that RL-ILTA is significantly better (worse) than the corresponding approach in that column, ‘≥’ means that RL-ILTA is slightly better, and ‘=’ means that there is no difference between the two hyper-heuristic approaches across the 30 runs. Instance RL RL DH DH DH SR SR SR LACC GDEL ILTA LACC GDEL ILTA LACC GDEL myciel3 = > = = > = = = myciel4 = = = = = = = = myciel5 = = = = = = = = queen5.5 > = = = = = > = queen6.6 > = ≥ > > > > = queen7.7 > = > > > > > = queen8.8 > = ≥ > > > > = DSJC125.1 > < > > > > > = DSJC125.5 > > ≥ > > ≥ > ≥ DSJC125.9 ≥ > > ≥ > > > > DSJC250.1 > > < > > > > > DSJC250.5 > > ≥ > > > > > DSJC250.9 > > > > > > > > DSJC500.1 ≥ > < < > ≥ ≥ > DSJC500.5 > > > > > > > > le450 25a = > = > > = = > le450 25b = > = = > = = > le450 25c ≥ > < > > ≥ > > le450 25d ≥ > > ≥ > > ≥ > selected problem instances, we carried out another comparison using the hyper-volume measure in which the performance of different algorithms is given in terms of the size of the search space that is covered by the final Pareto front of each hyper-heuristic. Figure 5 shows the box plots of the 30 hyper-volume values produced by each hyper-heuristic on each problem instance. A quick glance at the figure exposes the fact that hyper-heuristics with GDEL acceptance cover the least size of the search space compared to the other hyper-heuristics in most of the selected data sets. 4.4. Performance Comparison to Previously Proposed Approaches In Table 7, we compare the performance of our approach, RL-ILTA to some previously pro- posed approaches form the literature for graph colouring. In the table, the results denoted as RL-ILTA are the best colourings obtained by our framework. The results under M-LLE are ob- tained by a multi-objective genetic grouping algorithm described in (Korkmaz, 2010). Lowest Index Max Crossover (LIMX), Greedy Partition Crossover Lowest Index (GPX-LI) and Greedy Partition Crossover Cardinality Based (GPX-CB) graph colouring algorithms are proposed in (Ülker et al., 2006). Külahçıoğlu (2007) proposed two modified versions of the LLE representa- tion which are Linear Linkage Encoding With Ending Node Links (LLE-e) and Linear Linkage Encoding With Backward Links (LLE-b), and both of them are tested using genetic operators. This last study also tested these operators with classical Linear Linkage Encoding (LLE). The results in the last two columns denoted (Kir-B) and (Kir-C) are graph colouring algorithms pro- posed in (Kirovski and Potkonjak, 1998). Fields marked as ‘−’ means that the solution for that problem instance with that specific algorithm is not reported. The ‘wins’ row shows the number of instances in which the corresponding approach hit the best known colouring. As it can clearly be seen in the table, our proposed approach is not only competitive with the previous algorithms, but also it outperforms the previously proposed approaches almost in all cases. 21 ● ● ● ● ● ● ● ● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 2200 2400 2600 2800 DSJC125.1.col RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 2500 2600 2700 2800 2900 3000 3100 DSJC125.5.col ● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 800 1000 1200 1400 1600 DSJC125.9.col ● ● ● ● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 6500 7000 7500 DSJC250.1.col RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 5800 6000 6200 6400 6600 DSJC250.5.col RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 3400 3500 3600 3700 3800 3900 4000 DSJC250.9.col ● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 2000 2500 3000 3500 4000 le450_25b.col ● ● ● ● ● ● ● ● ● ● RL1 RL2 RL3 DH1 DH2 DH3 SR1 SR2 SR3 2000 2500 3000 3500 4000 le450_25a.col Figure 5: Box plots of the hyper-volume value of all the final Pareto fronts achieved by each hyper-heuristic approach on selected DIMACS data sets. 1, 2 and 3 in the hyper-heuristic approaches names refer to ILTA, LACC and GDEL acceptance methods respectively Table 7: Performance comparison for different approaches based on the best result. The entries in bold indicates the best result obtained by the associated algorithm for a given instance. Instance k∗ RL-ILTA M-LLE LIMX GPX-LI GPX-CB LLE-e LLE-b LLE Kir-B Kir-C DSJC125.1 5 5 − − − − − − − − − DSJC125.5 17 17 18 18 18 18 19 19 18 19 18 DSJC125.9 44 44 44 44 44 44 44 44 44 45 45 DSJC250.1 8 8 9 9 9 9 9 9 9 9 9 DSJC250.5 28 29 − 31 31 31 − − − 30 30 DSJC250.9 72 72 75 75 75 74 74 74 74 77 77 DSJC500.1 12 13 − 14 14 14 − − − 14 14 DSJC500.5 48 52 55 − − − − − − − − le450 25a 25 25 − 25 25 25 − − − 25 25 le450 25b 25 25 − 25 25 25 − − − 25 25 le450 25c 25 27 29 28 28 28 29 29 29 28 28 le450 25d 25 27 − 28 28 28 − − − − − Wins 12 1 3 3 3 1 1 1 2 2 4.5. Grouping Hyper-heuristics for Examination Timetabling Finally, in order to show that the framework is generic and can be applied to other domains of grouping problems, we tested the most successful hyper-heuristics, namely the RL-ILTA and the DH-ILTA, on selected instances from the Examination Timetabling domain. In (Johnson and Trick, 1996), it has been shown that the exam timetabling problem can be reduced to a grouping problem if the task of minimising the number of exam periods and removing the clashes are considered. The size of the selected ETT problem instances subset was kept small since the main 22 objective of testing the ETT domain is to show the generality of the framework. Table 8 shows the results of applying the RL-ILTA and DH-ILTA hyper-heuristics. Both hyper-heuristics perform well on all instances and successfully managed to find the best colourings for each instance. Table 9 provides the average best results obtained using RL-ILTA over 30 runs. RL-ILTA achieves the best known result for all instances. Moreover, RL-ILTA detects the best number of time-slots consistently for three large size examination timetabling benchmark instances of ute92, sta83 and hec92 under a second in all runs. Table 8: The results obtained from the application of RL-ILTA and DH-ILTA hyper-heuristics on Toronto a benchmark. sRate% and µt (s) are the success rates and the average time taken by each hyper-heuristic on each problem instance over 30 runs. RL-ILTA DH-ILTA Instance k sRate% µt (s) sRate% µt (s) hec92 17 100.00 4.547 93.33 345.171 18 100.00 0.328 100.00 1.911 19 100.00 0.120 100.00 0.928 sta83 13 100.00 0.456 100.00 9.068 14 100.00 0.089 100.00 0.448 15 100.00 0.053 100.00 0.303 yor83 19 16.67 583.031 46.67 598.321 20 100.00 27.724 100.00 41.306 21 100.00 2.829 100.00 6.830 ute92 10 100.00 0.134 100.00 0.431 11 100.00 0.059 100.00 0.206 12 100.00 0.040 100.00 0.161 rye93 21 20.00 1320.075 50.00 1619.851 22 100.00 82.934 100.00 103.213 23 100.00 12.399 100.00 19.178 The performance of RL-ILTA is compared against the best results obtained by some pre- viously proposed approaches including (Carter et al., 1996) (Carter), (Caramia et al., 2001) (Caramia) and (Merlot et al., 2002) (Merlot). The approaches in (Ülker et al., 2006) includ- ing Greedy Partition Crossover Lowest Index (GPX-LI), Greedy Partition Crossover Cardinality Based (GPX-CB) and Lowest Index Max Crossover (LIMX) algorithms are also considered in our comparison. Table 10 shows that the performance of RL-ILTA is competitive. It performs as good as the Carter and Caramia approaches. Moreover, it is better than the previously proposed population based grouping algorithms, including LIMX, GPX-LI and GPX-CB on most of the instances, particularly yor83 and rye93. 5. Conclusions Designing an automated intelligent search methodology which can be applied to the unseen problem instances with different characteristics without requiring a change is an extremely chal- lenging task. Selection hyper-heuristics have emerged as such flexible search methodologies Table 9: Average best colouring and associated standard deviation obtained by RL-ILTA over 30 runs. RL-ILTA Instance k∗ µ(kbest ) σ(kbest ) hec92 17.0 17.0 ±0.0 sta83 13.0 13.0 ±0.0 yor83 19.0 19.83 ±0.38 ute92 10.0 10.0 ±0.0 rye93 21.0 21.8 ±0.41 23 Table 10: Performance comparison for different approaches based on the best result. The entries in bold indicates the best result obtained by the associated algorithm for a given instance. Instance k RL-ILTA LIMX GPX-LI GPX-CB Carter Caramia Merlot hec92 17 17 17 17 17 17 17 18 sta83 13 13 13 14 14 13 13 13 yor83 19 19 20 20 20 19 19 23 ute92 10 10 10 10 10 10 10 11 rye93 21 21 23 23 23 21 21 22 Wins 5 3 2 2 5 5 1 supporting re-usability and component based development (Burke et al., 2013). Most of the previously proposed general purpose hyper-heuristics contain two main reusable components: heuristic selection and move acceptance. This study describes a general hyper-heuristic frame- work for solving grouping problems, employing generic components as well as a fixed set of low level heuristics and a slightly modified version of the grouping representation in (Falkenauer, 1998)4. A performance comparison of nine different hyper-heuristics using different components under this framework is presented on various graph colouring benchmark instances. The re- sults indicates the success of an online learning hyper-heuristic which uses feedback during the search process. The selection hyper-heuristic combining the reinforcement learning (RL) heuristic selection and ILTA move acceptance (Misir et al., 2013) methods even outperforms some previously proposed approaches. RL-ILTA is further tested on an examination timetabling benchmark. This hyper-heuristic without requiring any change again yielded successful results. The proposed framework is indeed flexible allowing different hyper-heuristic components to be brought together and sufficiently general. The RL-ILTA hyper-heuristic implemented under the proposed framework performs slightly better than the selection hyper-heuristic which won a hyper-heuristic competition (Burke et al., 2011). This previously proposed method contains many parameters which were tuned and complicated subcomponents for heuristic selection. The RL-ILTA hyper-heuristic implemented under the proposed framework conforms to one of the crucial properties of a reusable hyper-heuristic, that is the simplicity. The heuristic selection component has only one parameter, that is the memory length and that is set to a fixed value as suggested in (Burke and Soubeiga, 2003). Our observations in this study are consistent with the previous findings from the literature (Burke et al., 2012; Özcan et al., 2009; Özcan et al., 2010). The use of a different component in a hyper-heuristic could lead to a different performance of the overall algorithm. Learning during the heuristic selection process definitely helps, and the move acceptance plays a major role in the performance of hyper-heuristics. It is crucial to employ compatible and synergistic components yielding an improved performance and RL performed well with ILTA under the proposed grouping hyper-heuristic framework. Acknowledgements: This work was funded in part by the EPSRC grant EP/F033613/1. 4The grouping hyper-heuristic tool will be made publicly available from http://www.cs.nott.ac.uk/~axe/ 24 References Agustı́n-Blas, L., Salcedo-Sanz, S., Jimnez-Fernndez, S., Carro-Calvo, L., Ser, J.D., Portilla-Figueras, J., 2012. A new grouping genetic algorithm for clustering problems. Expert Systems with Applications 39, 9695 – 9703. doi:http: //dx.doi.org/10.1016/j.eswa.2012.02.149. Avanthay, C., Hertz, A., Zufferey, N., 2003. A variable neighborhood search for graph coloring. European Journal of Operational Research 151, 379–388. Bai, R., Kendall, G., 2005. An investigation of automated planograms using a simulated annealing based hyper-heuristics, in: Ibaraki, T., Nonobe, K., Yagiura, M. (Eds.), Metaheuristics: Progress as Real Problem Solver - (Operations Research/Computer Science Interface Serices, Vol.32). Springer, pp. 87–108. Bilgin, B., Özcan, E., Korkmaz, E.E., 2007. An experimental study on hyper-heuristics and exam timetabling, in: Proceedings of the 6th international conference on Practice and theory of automated timetabling VI, Springer-Verlag, Berlin, Heidelberg. pp. 394–412. Brown, C.E., Sumichrast, R.T., 2003. Impact of the replacement heuristic in a grouping genetic algorithm. Computers & OR 30, 1575–1593. Burke, E.K., Bykov, Y., 2008. A Late Acceptance Strategy in Hill-Climbing for Exam Timetabling Problems, in: PATAT ’08 Proceedings of the 7th International Conference on the Practice and Theory of Automated Timetabling. URL: http://w1.cirrelt.ca/\~{}patat2008/PATAT\_7\_PROCEEDINGS/Papers/Bykov-HC2a.pdf. Burke, E.K., Curtois, T., Hyde, M., Kendall, G., Ochoa, G., Petrovic, S., Vazquez-Rodriguez, J., 2009. Hyflex: A flexible framework for the design and analysis of hyper-heuristics., in: Proceedings of the Multidisciplinary International Scheduling Conference (MISTA09), pp. 790–797. Burke, E.K., Gendreau, M., Hyde, M., Kendall, G., McCollum, B., Ochoa, G., Parkes, A., Petrovic, S., 2011. The cross-domain heuristic search challenge - an international research competition, in: Yao, X., Coello, C.A.C. (Eds.), Proceedings of Learning and Intelligent Optmization (LION5), pp. 631–634. Burke, E.K., Gendreau, M., Hyde, M.R., Kendall, G., Ochoa, G., Özcan, E., Qu, R., 2013. Hyper-heuristics: a survey of the state of the art. JORS 64, 1695–1724. Burke, E.K., Kendall, G., Misir, M., Özcan, E., 2012. Monte carlo hyper-heuristics for examination timetabling. Annals of Operations Research 196, 73–90. doi:10.1007/s10479-010-0782-2. Burke, E.K., Landa-Silva, J.D., Soubeiga, E., 2005. Multi-objective hyper-heuristic approaches for space allocation and timetabling, in: Ibaraki, T., Nonobe, K., Yagiura, M. (Eds.), Meta-heuristics: Progress as Real Problem Solvers. Springer. 5th Metaheuristics International Conference (MIC 2003), pp. 129–158. Burke, E.K., Newall, J.P., 1998. A multi-stage evolutionary algorithm for the timetable problem. IEEE Transactions on Evolutionary Computation . Burke, E.K., Soubeiga, E., 2003. Scheduling nurses using a tabu-search hyperheuristic, in: Proc. of the 1st MISTA, pp. 197–218. Caramia, M., Dell’Olmo, P., Italiano, G.F., 2001. New algorithms for examination timetabling, in: Algorithm Engineer- ing 4th International Workshop 2000, Springer-Verlag, Berlin Heidelberg New York. pp. 230–241. Carter, M.W., Laporte, G., 1996. Recent developments in practical examination timetabling, in: Practice and Theory of Automated Timetabling, pp. 3–21. Carter, M.W., Laporte, G., Lee, S.T., 1996. Examination timetabling: algorithmic strategies and applications. Journal of the Operational Research Society 47, 373–383. Chakhlevitch, K., Cowling, P.I., 2008. Hyperheuristics: Recent developments, in: Cotta, C., Sevaux, M., Sörensen, K. (Eds.), Adaptive and Multilevel Metaheuristics. Springer. volume 136 of Studies in Computational Intelligence, pp. 3–29. Coello, C.C., Lamont, G., van Veldhuizen, D., 2007. Evolutionary Algorithms for Solving Multi-Objective Problems. Genetic and Evolutionary Computation. 2nd ed., Springer, Berlin, Heidelberg. Cowling, P., Chakhlevitch, K., 2003. Hyperheuristics for managing a large collection of low level heuristics to schedule personnel, in: Proceedings of the IEEE Congress on Evolutionary Computation (CEC2003), IEEE Computer Society Press, Canberra, Australia. pp. 1214–1221. Cowling, P., Kendall, G., Soubeiga, E., 2000. A hyperheuristic approach to scheduling a sales summit, in: Selected Papers of the Third International Conference on the Practice And Theory of Automated Timetabling, PATAT 2000, Springer, Konstanz, Germany. pp. 176–190. Davis, L.D., Mitchell, M., 1991. Handbook of Genetic Algorithms. Van Nostrand Reinhold. Du, J., Korkmaz, E.E., Alhajj, R., Barker, K., 2004. Novel clustering that employs genetic algorithm with new repre- sentation scheme and multiple objectives., in: Kambayashi, Y., Mohania, M.K., W, W. (Eds.), DaWaK, Springer. pp. 219–228. Dueck, G., 1993. New optimization heuristics: The great deluge algorithm and the record-to-record travel. Journal of Computational Physics 104, 86–92. doi:10.1006/jcph.1993.1010. 25 Falkenauer, E., 1992. The grouping genetic algorithms: Widening the scope of the GAs. Belgian Journal of Operations Research, Statistics and Computer Science (JORBEL) 33, 79–102. Falkenauer, E., 1998. Genetic Algorithms and Grouping Problems. John Wiley & Sons, Inc., New York, NY, USA. Galinier, P., Hao, J.K., 1999. Hybrid evolutionary algorithms for graph coloring. J. Comb. Optim. 3, 379–397. Handl, J., Knowles, J.D., 2007. An evolutionary approach to multiobjective clustering. IEEE Trans. Evolutionary Computation 11, 56–76. Hertz, A., Werra, D.D., 1987. Using tabu search techniques for graph coloring. Computing 39, 345–351. URL: http://dx.doi.org/10.1007/BF02239976, doi:10.1007/BF02239976. Johnson, D.J., Trick, M.A. (Eds.), 1996. Cliques, Coloring, and Satisfiability: Second DIMACS Implementation Chal- lenge, Workshop, October 11-13, 1993. American Mathematical Society, Boston, MA, USA. Johnson, D.S., Aragon, C.R., Mcgeoctt, L.A., Schevon, C., 1991. Optimization by simulated annealing: an experimental evaluation; part ii. Operational Research . Jones, D.R., Beltramo, M.A., 1991. Solving partitioning problems with genetic algorithms, in: Belew, R.K., Booker, L.B. (Eds.), ICGA, Morgan Kaufmann. pp. 442–449. Kirovski, D., Potkonjak, M., 1998. Efficient coloring of a large spectrum of graphs, in: 35th Design Automation Confer- ence Proceedings, pp. 427–432. Korkmaz, E.E., 2010. Multi-objective genetic algorithms for grouping problems. Appl. Intell. 33, 179–192. Külahçıoğlu, B., 2007. Multiobjective Hyperheuristic for Data Clustering and Linear Linkage Encoding. Master’s thesis. Yeditepe University. Leighton, F.T., 1979. A graph coloring algorithm for large scheduling problems, in: Journal of Reasearch of the National Bureau of Standards, pp. 489–506. Merlot, L.T.G., Boland, N., Hughes, B.D., Stuckey, P.J., 2002. A hybrid algorithm for the examination timetabling problem., in: Burke, E.K., De Causmaecker, P. (Eds.), Proceedings of Practice and Theory of Automated Timetabling, Fourth International Conference, Gent, Belgium. pp. 348–371. Misir, M., Verbeeck, K., Causmaecker, P.D., Berghe, G.V., 2013. A new hyper-heuristic as a general problem solver: an implementation in hyflex. J. Scheduling 16, 291–311. Özcan, E., Bilgin, B., Korkmaz, E., 2008. A comprehensive analysis of hyperheuristics. Intelligent Data Analysis 12, 1–21. Özcan, E., Bykov, Y., Birben, M., Burke, E.K., 2009. Examination timetabling using late acceptance hyper-heuristics, in: Proceedings of IEEE Congress on Evolutionary Computation (CEC 2009), pp. 997–1004. Özcan, E., Mısır, M., Ochoa, G., Burke, E.K., 2010. A reinforcement learning - great-deluge hyper-heuristic for exami- nation timetabling. Int. J. of Applied Metaheuristic Computing 1, 39–59. Paquete, F.L., Fortseca, C.M., Pt, E.L., 2001. A study of examination timetabling with multiobjective evolutionary algorithms, in: Proc. of the 4th Metaheuristics International Conference (MIC 2001, pp. 149–154. Park, Y.J., Song, M.S., 1998. A genetic algorithm for clustering problems, in: Koza, J.R., Banzhaf, W., Chellapilla, K., Deb, K., Dorigo, M., Fogel, D.B., Garzon, M.H., Goldberg, D.E., Iba, H., Riolo, R. (Eds.), Genetic Program- ming 1998: Proceedings of the Third Annual Conference, Morgan Kaufmann, University of Wisconsin, Madison, Wisconsin, USA. pp. 568–575. Qu, R., Burke, E.K., McCollum, B., Merlot, L., Lee, S., 2009. A survey of search methodologies and automated system development for examination timetabling. Journal of Scheduling 12, 55–89. Saha, S., Kumar, R., Baboo, G., 2013. Characterization of graph properties for improved pareto fronts using heuristics and EA for bi-objective graph coloring problem. Applied Soft Computing 13, 2812 – 2822. URL: http://www.sciencedirect.com/science/article/pii/S1568494612003018, doi:http://dx.doi.org/ 10.1016/j.asoc.2012.06.021. Ülker, O., Korkmaz, E.E., Özcan, E., 2008. A grouping genetic algorithm using linear linkage encoding for bin packing, in: Parallel Problem Solving from Nature, pp. 1140–1149. Ülker, Ö., Özcan, E., Korkmaz, E.E., 2006. Linear linkage encoding in grouping problems: Applications on graph coloring and timetabling, in: Burke, E.K., Rudová, H. (Eds.), PATAT, Springer. pp. 347–363. Wu, Q., Hao, J.K., 2012. An effective heuristic algorithm for sum coloring of graphs. Computers & OR 39, 1593–1600. Yılmaz, B., Korkmaz, E.E., 2010. Representation issue in graph coloring, in: ISDA, IEEE. pp. 1171–1176. Zitzler, E., Thiele, L., 1998. Multiobjective optimization using evolutionary algorithms - a comparative case study, in: Eiben, A.E., Bäck, T., Schoenauer, M., Schwefel, H.P. (Eds.), PPSN, Springer. pp. 292–304. 26 
work_xbz5gmt6mffjhaie5li65ru4lq	----	doi:10.1016/j.eswa.2006.10.041 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 833–844 Expert Systems with Applications Dynamic EMCUD for knowledge acquisition Shun-Chieh Lin a,*, Shian-Shyong Tseng a,b, Chia-Wen Teng a a Department of Computer Science, National Chiao Tung University, Taiwan, ROC b Department of Information Science and Applications, Asia University, Taiwan, ROC Abstract Due to the knowledge explosion, the new objects will be evolved in a dynamic environment. Hence, the knowledge can be classified into static knowledge and dynamic knowledge. Although many knowledge acquisition methodologies, based upon the Repertory Grid technique, have been proposed to systematically elicit useful rules from static grid from domain experts, they lack the ability of grid evolution to incrementally acquire the dynamic knowledge of new evolved objects. In this paper, we propose dynamic EMCUD, a new Repertory Grid-based knowledge acquisition methodology to elicit the embedded meanings of knowledge (embedded rules bearing on m objects and k object attributes), to enhance the ability of original EMCUD to iteratively integrate new evolved objects and new added attributes into the original Acquisition Table (AT) and original Attribute Ordering Table (AOT). The AOT records the relative importance of each attribute to each object in EMCUD to capture the embedded meanings with acceptable certainty factor value by relaxing or ignoring some minor attributes. In order to discover the new evolved objects, a collaborative framework including local knowledge based systems (KBSs) and a collaborative KBS is proposed to analyze the correlations of inference behaviors of embedded rules between multiple KBSs in a dynamic environment. Each KBS monitors the frequent inference behaviors of interesting embedded rules to construct a small AT increment to facilitate the acquisition of dynamic knowledge after experts confirming the new evolved objects. Moreover, the significance of knowledge may change after a period of time, a trend of all attributes to each evolved object is used to construct a new AOT increment to help experts automatically adjust the relative importance of each attribute to each object using time series analysis approach. Besides, three cases are considered to assist experts in adjusting the certainty factor values of the dynamic knowledge of the new evolved objects from the collection of inference logs in the collaborative KBS. To evaluate the perfor- mance of dynamic EMCUD in incrementally integrating new knowledge into the knowledge base, a worm detection prototype system is implemented. � 2006 Elsevier Ltd. All rights reserved. Keywords: Knowledge acquisition; Dynamic EMCUD; Dynamic knowledge; Trend analysis; Worm detection 1. Introduction As we know, knowledge based system (KBS) is an intel- ligent computer program that uses knowledge and infer- ence procedures to solve problems that are difficult enough to require significant human expertise for their solutions, such as disease diagnosis, investment prediction, or computer science. Knowledge acquisition (KA) is one of 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.10.041 * Corresponding author. Tel.: +886 3 5731966; fax: +886 3 5721490. E-mail addresses: jielin@cis.nctu.edu.tw (S.-C. Lin), sstseng@cis.nctu. edu.tw (S.-S. Tseng). the critical bottlenecks in developing a KBS for obtaining the knowledge of a special domain from domain experts. Repertory Grid, based on Kelly’s Personal Construct Theory (Kelly, 1955) which reports how people make sense of the world, could be used as an efficient knowledge acqui- sition technique in identifying different objects and distin- guishing these objects in a domain. Although many KA systems and tools, e.g., NeoETS (Boose & Bradshaw, 1986), AQUINS (Boose & Bradshaw, 1987), KITTEN (Shaw & Gaines, 1987), EMCUD (Hwang & Tseng, 1990), KADS (Wielinga, Schreiber, & Breuker, 1992), have been proposed to rapidly build prototypes and improve the quality of the elicited knowledge of well-known objects by mailto:jielin@cis.nctu.edu.tw mailto:sstseng@cis.nctu. edu.tw mailto:sstseng@cis.nctu. edu.tw 834 S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 domain experts with/without knowledge engineers based upon the Repertory Grid technique in the past twenty years, they are still focusing on acquiring a static gird infor- mation to generate static knowledge, which remains the same in the changing environment as time goes on. However, with the changing environment as time goes on, new objects in many domains are incrementally evolved or developed due to the explosion of knowledge, resulting in the creation of new knowledge. Hence, knowledge can be classified as static knowledge and dynamic knowledge according to the stability of knowledge in a dynamic envi- ronment with the times. The static knowledge remains the same in the changing environment as time goes on. The dynamic knowledge, which may be updated or evolved, will be adapted in the changing environment with the times, since the previous knowledge may be degraded or upgraded in the near future. The knowledge evolution we proposed in this paper is the iterative process to acquire evolutional knowledge in a changing environment. Although many Repertory Grid-based KA methodologies can be used to acquire the dynamic knowledge of new evolved objects through rerunning their KA procedures, it is time-consuming and the relevant inference logs are sel- dom saved in the static grid. If some relevant log informa- tion can be analyzed to extract some useful information of discovering new objects, the dynamic knowledge of new evolved objects can be generated by experts. EMCUD (Embedded Meaning Capturing and Uncer- tainty Deciding) (Hwang & Tseng, 1990) was proposed to elicit the embedded meanings of knowledge (embedded rules bearing on m objects (O1, O2, . . . , Om) and k object attributes), which represents the information that domain experts take for granted but are implicit to the people who is not familiar with the application domain, and guide experts to decide the certainty degree of each embedded rule for extending the coverage of the original rules gener- ated by Acquisition Table (AT) based upon Repertory Grid technique. Since the relative importance of each attri- bute to each object could be represented as Attribute Ordering Table (AOT) in EMCUD, some minor attributes can be relaxed or ignored to capture the embedded mean- ings with acceptable certainty factor (CF). Assume some objects in O1 class, which are classified by original rules of O1, belong to the original object class (OO1) of O1; the other objects in O1 class, which are clas- sified by embedded rules of O1, belong to the extended object class (EO1) of O1. However, some embedded rules may be with marginally acceptable CF values due to the weak suggestions of domain experts. In the age of the knowledge explosion, some objects might be evolved with the times and could be classified by the weak embedded rules of O1 with weak CF values since some related ambig- uous attributes (minor attributes) are ignored to classify these new evolved objects into O1 class. Although EMCUD can generate embedded rules to clas- sify a new object into an original object class, it still has to be rerun to generate the dynamic knowledge using an updated AT and a new AOT acquired by experts if new objects are evolved. Moreover, the human experts are usu- ally unaware of the occurrence the new evolved objects due to the lack of sufficient relevant information. Therefore, the dynamic EMCUD will be proposed in this paper to effi- ciently generate dynamic knowledge and iteratively inte- grate an AT increment and an AOT increment into the main AT and the main AOT, respectively. The AT incre- ment records the new evolved objects and related new added attributes and the AOT increment records the evolved trend of all attributes to each new evolved object with the times in a dynamic environment. In order to analyze useful evidence of the new evolved objects in a dynamic environment, we will propose a col- laborative framework (including local KBSs and a collabo- rative KBS) to monitor the frequent inference behaviors of weak embedded rules and to trace the evolved behaviors of objects with the times from multiple KBSs for assisting experts in efficiently obtaining the dynamic knowledge. Each local KBS deploys a New Evolved Object learning (neo-learning) module to monitor the frequent inference behaviors of weak embedded rules to iteratively construct an AT increment. The AT increment could be created to record the relationships between new objects and new attri- butes after new objects are confirmed by experts without asking too many questions. Moreover, since the evidence of object evolution may appear diversely in unpredictable time, the relevant information can be collected as an Attri- bute Signal Table (AST) to record the significant impor- tance of each attribute to each object in every time point in a local KBS. The AOT increment could be constructed using time series analysis technique to analyze the impor- tance of each attribute to each object recorded in AST with the times to facilitate the acquisition and adaptation of dynamic knowledge without too many interactions with experts in a changing environment. However, some new evolved objects might be invisible or insignificant under each local KBS with neo-learning module, the profile of each KBS and the infrequent logs are analyzed in the collaborative KBS to collaboratively assist experts in discovering new evolved objects. The infre- quent inference logs can be analyzed by neo-learning mod- ule and corresponding profiles to discover the interesting knowledge of new evolved objects which is unseen in each KBS. In order to acquire a meaningful CF value of each new discovered embedded rule of evolved objects, the CF value of each new embedded rule of evolved objects could be adjusted in the collaborative KBS based upon three cases in the CF adjusting function. A worm detection prototype system using the col- laborative framework with the neo-learning module is implemented to evaluate the performance of dynamic EMCUD, which incrementally integrates the new evolved knowledge into original knowledge base. Based upon the collaborative framework, the dynamic knowledge of new worms could be elicited to discover the new variant worms generated by the attacking traffic generator in the experi- Table 1 The illustrative example of a Repertory Grid with ratings Element/construct E1 E2 E3 E4 E5 C1 5 1 5 1 1 C 0 1 C2 4 4 4 1 4 C 0 2 C3 1 4 5 1 4 C 0 3 C4 1 4 4 5 5 C 0 4 S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 835 mental environment based upon the worm classification embedded rule base, which results in the ability of knowl- edge evolution. 2. Related work Several knowledge acquisition (KA) methodologies and related systems are introduced in this section. Then Reper- tory Grid, one of the popular indirect KA techniques, is also discussed. Finally, the elicitation of embedded mean- ing and some problems of traditional KA methodologies are discussed. 2.1. Knowledge acquisition methodologies Since the knowledge in many domains (the experience of domain experts) is continuously growing, many KA meth- odologies and tools have been proposed to help experts acquire the useful static knowledge with/without knowl- edge engineers and then to transfer these knowledge into knowledge base or other computerized representation forms. In general, there are three approaches for knowl- edge acquisition (Crowther & Hartnett, 1996; Hwang & Tseng, 1990; Mcgraw & Harbison-Briggs, 1989): (1) Interviewing experts by experienced knowledge engi- neers: Knowledge engineers directly retrieve domain knowledge by interviewing with human experts, and transform the knowledge into the computerized for- mat to help experts solve difficult problems in the real world. However, interviewing experts is usually time- consuming if the communication between domain experts and knowledge engineers is insufficient. It is also difficult for experts to be aware of the new evolved objects without any additional related infor- mation in the interviewing methodologies. (2) Machine learning: The machine learning approaches can learn the useful static knowledge of well-known objects by collecting many useful cases and instances with/without the involvement of domain experts. However, the quality of the results usually relies on the selected training cases and lacks new cases of evolved objects in the training process. (3) Knowledge acquisition systems: KA systems assist domain experts in generating useful static knowledge of well-known objects with/without the help of knowledge engineers. These systems or tools could reduce the effort of communication between knowl- edge engineers and domain experts and could reduce the risk and difficulty of selecting the suitable training cases. In the past decades, many KA systems, e.g., ETS (Boose, 1984, 1985), NeoETS (Boose & Bradshaw, 1986), AQUINS (Boose & Bradshaw, 1987), KITTEN (Shaw & Gaines, 1987), RuleCons (Davis, 1987), MOLE (Eshelman, Ehret, McDermott, & Tan, 1987), KSSO (Gaines, 1987), KRITON (Diederich, Ruhmann, & May, 1987), EMCUD (Hwang & Tseng, 1990; Hwang & Tseng, 1991), KADS (Wielinga et al., 1992), MCRDR (Kang, 1996), KAMET (Cairo, 1998), MedFrame/CADIAG-IV (Boegl, 1997; Kolousek, 1997; Leitich et al., 2001), (Pan, Zheng, Zeng, & Hu, 2002) have been developed to rapidly build prototypes and improve the quality of the elicited static knowledge of well-known objects by domain experts. However, most of them cannot be used to construct the dynamic knowledge due to the limitation of the static attribute set of the static grid in a dynamic environment. For acquiring the dynamic knowledge of new evolved objects, the traditional KA approach needs to rerun to generate the dynamic knowledge, and thus cannot keep the useful information of each object during the period of evolution, resulting in the limitations of traditional Repertory Grid-based KA methodologies. 2.2. Repertory Grid methodology and relevant systems Repertory Grid, based on Kelly’s Personal Construct Theory (Kelly, 1955) which reports how people make sense of the world, could be used as an efficient KA technique in identifying different objects and distinguishing these objects in a domain. It is the basis of several computer assisted KA tools, such as ETS (Boose, 1984, 1985), AQUINAS (Boose & Bradshaw, 1987) and KSSO (Gaines, 1987). A single grid represented as a matrix whose columns have element objects (labels) and whose rows have con- struct’s attributes (labels) can classify a class of objects, or individuals. The value assigned to an element-construct pair need not be Boolean. Grid values have numeric ratings, probabilities, and other characteristics, where each value reflects the degree. Then, the expert is asked to fill the grid with 5-scale ratings, where ‘‘1’’ represents the most relevant attribute to the object; ‘‘2’’ represents that the attribute may relevant to the object; ‘‘3’’ represents ‘‘unknown’’ or ‘‘no relevance’’; ‘‘4’’ represents that the object may have the opposite characteristic; ‘‘5’’ represents the most relevant opposite characteristic to the object. The whole concept of Repertory Grid technique can be described as following steps: (1) Elicit all of the element objects, e.g., E1, E2, E3, E4, E5 from the expert. (2) Elicit the construct attributes (and their opposites), e.g., C1, C2, C3, C4 ðC01; C 0 2; C 0 3; C 0 4Þ, from the expert. Each time three elements are chosen to ask for a construct to distinguish one element from the other two. 836 S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 (3) Rate all of the [element, construct] entries of the grid with value range from 1 to 5. An illustrative example is given in Table 1. As Repertory Grid technique has been widely used by researchers, some extensions have been made to enrich its representative ability for covering more knowledge, the value assigned to an element-construct pair may be Bool- ean, numeric ratings, probabilities, etc. For example, Dixit and Pindyck (1994), and Hwang (1995) extended the Rep- ertory Grid technique to the fuzzy table, in which con- structs were fuzzy attributes that could rate by means of fuzzy linguistic terms from a finite set. Castro-Schez, Jen- nings, Luo, and Shadbolt (2004) developed a technique using a fuzzy Repertory Grid for acquiring the finite set of attributes or variables that the expert uses in a classifica- tion problem to characterize and discriminate a set of ele- ments. Moreover, several models have been proposed for handling uncertainties in expert systems through generat- ing more meaningful rules from the Repertory Grid ori- ented approaches. EMYCIN certainty factor (CF) model was first used to decide the degree of the belief of a rule for uncertain reasoning (Shortliffe & Buchanan, 1975; Adams, 1985). Embedded Meaning Capturing and Uncer- tainty Deciding (EMCUD) KA system was proposed to extract rules with embedded meaning from hierarchical girds by defining the impacts of the constructs to each element (Hwang & Tseng, 1990) and was successive applied in a medical diagnostic system of acute exanthema in Tai- wan (Hwang & Tseng, 1991). WebGrid, Calgary’s web- based knowledge modeling and inference tool, is based on Repertory Grid elicitation and analysis (Shaw & Gaines, 1996). Boicu (2001) described a practical approach, methodol- ogy and tool, for the development of static knowledge bases and agents by subject matter experts, with limited assistance from knowledge engineers, the idea of constructing the dynamic knowledge bases systematically is launched. How- ever, the dynamic environment limits all their efficiencies for the KA methodology. The real world knowledge in a dynamic environment is often considered evolutional, because the knowledge can be changed or evolved with the times due to the advent of information century. The speed of knowledge changing is too fast to accumulate man- ually the information by experts which results in limiting the ability of the ritualistic batch process analysis. Although these methodologies are proposed to extend the ability of uncertain reasoning to classify the well- known objects, they still have the limitation of acquiring static knowledge from the static grid. It is also difficult for experts to notice the occurrence of new evolved objects, which is evolved in the dynamic environment as time goes on. Therefore, an incremental KA system based upon Rep- ertory Grid technique is hence proposed in this paper to integrate dynamic knowledge of new objects into original knowledge base through the observations of the interested inference logs. 2.3. Elicitation of embedded meanings The embedded meanings referred here represents the information that domain experts take for granted but are implicit to the people who are not familiar with the applica- tion domain. The lack of embedded meaning will probably make an expert system fail to infer some cases being trivial to experts. SEEK (Politakis & Weiss, 1984) and SEEK2 (Ginsberg, Weiss, & Politakis, 1988) have been proposed to obtain embedded meanings by some efficient refinement processes. However, the major problem of SEEK and SEEK2 is the case database being assumed to be available because it is difficult to collect sufficient cases in some appli- cations. Moreover, it would be also time-consuming and boring for experts to offer a conclusion for each case in the database before starting the refinement procedure. EMCUD, a Repertory Grid-based KA method, is hence proposed to elicit the embedded meanings of knowledge (embedded rules bearing on m objects and k object attri- butes) from the existing hierarchical grids given by experts (Hwang & Tseng, 1990), which represents the information that domain experts take for granted but are implicit to whom are not familiar with the application domain. Addi- tionally, it will also guide experts to decide the certainty degree of each rule with embedded meaning for extending the coverage of generated original rules. Besides using Acquisition Table (AT) to generate the rule of each object, the Attribute Ordering Table (AOT), which is used to record the relative importance of each attribute to each object, is employed to capture the embedded meanings of the resulting grids. The values in each AOT entry, a pair of attribute and object, may be labeled ‘‘X’’, ‘‘D’’ or an integer number. ‘‘X’’ means no relationship existing between the attribute and the object. ‘‘D’’ means that the attribute dominates the object, i.e., if the attribute is not equal to the entry value, it is impossible for the object to be implied. Integer numbers are used to represent for the relative important degree of the attribute to the object instead of dominating the corresponding object. If the attribute does not equal the attribute-value, it is still for the object to be implied. The larger integer number implies the attribute being more important to the object. Using AOT, the original rules generate some rules with embedded meaning, and the CF value of each rule, which is between �1 and 1, could be determined to indicate the degree of supporting the inference result. The higher CF value implies the more reliable result than smaller CF value. The EMCUD algorithm is listed as follows. Algorithm 1: EMCUD algorithm Input: The hierarchical grids. Output: The guiding rules with embedded meaning. Step 1: Build the corresponding AOT with each grid of the hierarchical multiple grids. Step 2: Generate the possible rules with embedded meaning. Inference Log Frequency Events Analysis Trend Evolution Analysis EMCUD Dynamic Knowledge Neo-learning Grid Merging AT Increment AOT Increment Main AT Main AOT Fig. 1. The concept of dynamic EMCUD. S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 837 Step 3: Select the accepted rules with embedded meaning through the interaction with experts. Step 4: Generate automatically the CF of each rule with embedded meaning. All rules generated by EMCUD can be categorized into two classes: original and embedded rules with acceptable CF value, and discarded rules with unacceptable CF value, according to the confidence degree of domain experts. Embedded rules can be generated by ignoring the minor (non-dominate) attributes recorded in AOT. Each embedded rule is assigned a certainty sequence (CS), the sum of each AOT values of the ignored attributes, and the CF calculated by formula (1) which is between 0 and 1 can represent the degree of certainty for each embed- ded rule. Each of them is assigned a CF between 0 and 1 while the value approaches to 1 means more important; otherwise, the value approaches to 0 means less important: CFðRi;jÞ¼ UBðRiÞ� CSðRi;jÞ MAXðCSi;jÞ �ðUBðRiÞ� LBðRiÞÞ � � ð1Þ where Ri,j and CS(Ri,j) are the jth embedded rule of the ob- ject Oi and the CS value, respectively. The MAX(CSi,j) is the maximum CS value of the embedded rules generated from object Oi. To decide the CF of each embedded rule Ri,j, the upper (UB(Ri)) and the lower (LB(Ri)) bounds CF values of the object Oi have been firstly defined for accepted embedded rules. Then CF values of each rule can be automatically determine by the mapping function, formula (1). Thus, the useful embedded rules with corresponding CF values could be used to cover more uncertainty cases. However, the Repertory Grid-oriented method to con- struct AT is somehow strenuous for an expert and even more strenuous to solve the adaptive problem in a dynamic environment; therefore, we propose an incremental KA method based upon EMCUD to cope with the problems above by further enhancing the original EMCUD method to become a dynamic EMCUD method, which can inte- grate dynamic knowledge into original knowledge base. Since embedded rules with weak acceptable CF values (the CF values below a user defined threshold) usually mean domain experts might lack the strong confidence, objects matching weak embedded rules may be the candi- dates of new evolved objects. For example, the object sat- isfying the conditions (attribute-value pairs) of the embedded rules with CF = 0.5 means the expert might sug- gest that it would be marginally classified into the object class and the minor attributes of the embedded rule might be not clearly defined. Therefore, the fired frequencies of this kind of weak embedded rules could be used to discover the occurrence of new evolved objects. With the changing environment, the adaptation of the acquired rules is required to cope with the new evolved objects. Although EMCUD successfully solves the prob- lems of the conventional Repertory Grid including knowl- edge representation and embedded meaning for covering more similar objects in extended object class, it still exists several problems such as hard to explain the rules with lower CF value, difficulty in deciding the attribute order- ing, and lacks the ability of grid evolution for singling these new objects out due to the knowledge explosion in a chang- ing environment with the times. Therefore, enhancing the adaptation ability of embedded rules becomes increasingly important to achieve the ability of grid evolution in KBS. 3. The incremental knowledge acquisition methodology – dynamic EMCUD Generating rules in EMCUD would be cost inefficient if the size of Acquisition Table (AT) and Attribute Ordering Table (AOT) are too large. After collecting sufficient infor- mation of new evolved objects, EMCUD has to manually regenerate the original and embedded rules to classify these new objects with the large main AT. Therefore, the concept of dynamic EMCUD shown in Fig. 1 is proposed to help experts incrementally generate the dynamic knowledge based upon a new evolved object learning (neo-learning) module to construct a small AT increment and an AOT increment for enhancing the explanation power of the ori- ginal embedded knowledge base. The AT increment, which can be generated by monitor- ing the frequency of the weak embedded rules, is used to record the new evolved objects and the attributes which are updated or added to generate the dynamic knowledge. The AOT increment is used to help experts generate the adaptive relative importance of each attribute to each object in AT increment as time goes on by tracing the importance changing trends of all attributes in a time inter- val in the trend evolution analysis. Through integrating the AT increment and the AOT increment into the main AT and the main AOT using grid merging in dynamic EMCUD, it can generate the knowledge of new evolved objects with the grid evolution ability. 3.1. The new evolved object knowledge acquisition As we know, the KBS is proposed to help experts solve the difficult problems in a specific domain based upon the 838 S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 pre-constructed static knowledge base. However, the new objects will be developed or discovered as times goes on and might result in the inefficiency of KBS. Based upon the embedded rules generated by EMCUD, some new evolved objects may be classified into well-known object class by the weak embedded rule with weak CF which is not strongly suggested by experts. Through monitoring the frequency of these weak embedded rules, the candidates of new evolutional objects might be discovered to notice the experts. Therefore, the characteristics of these candi- dates of new objects could be extracted from these collected inference logs. The evidence of the new objects can be con- firmed by experts and some attributes could be modified and added when the dynamic knowledge is needed to be singled out. Moreover, the relationships between these inference logs might be represented as the significance of each attribute to each new object. Hence, analyzing the evolving trends of all attribute should be useful in captur- ing the realistic significance of the attribute to the object. The neo-learning module shown in Fig. 1 can help experts analyze the interesting inference logs of weak embedded rules to learn the evidence of new evolved objects using the frequent evolution analysis to notice experts the occurrence of the new objects. Based upon the confirmed new objects, the relationships of all attributes of each object are analyzed to set the significance of the attribute with the times using trend evolution analysis to help experts decide the CF values of the embedded rules of new objects, which can be generated using EMCUD according to the discov- ered objects stored in an AT increment and an AOT incre- ment. Finally, the AT increment and the AOT increment will be integrated with the main AT and the main AOT, respectively. 3.2. Frequency events analysis EMCUD lacks the ability of grid evolution for singling the new evolved objects out of well-known objects since experts may be unaware of the occurrence of the new evolved objects without sufficient information. Hence, we propose a frequency events analysis method to monitor the frequent behaviors of interesting inference logs of the weak embedded rules with the lower CF values for helping experts notice the occurrence of the new objects. The novelty of the frequency events analysis shown in Fig. 2 is to collect the inference logs of weak embedded rules from each KBS to learn the candidates of new evolved objects for experts to make a confirmation. The minor Fig. 2. The flow of frequ attribute-value pairs between inference logs of weak embedded rules are useful to help experts discover new knowledge and determine whether new object is evolved based upon fired frequency. For each object, if its inference logs of weak embedded rules are frequent, the frequent minor attribute-value pairs could be treated as candidates of new evolved objects. Furthermore, new attributes or attribute-values of the new object could be defined and used to generate a small AT increment. Hence, these candi- dates will be used to help experts single the new objects out of the extended object class using the new object acquisi- tion module based upon the AT increment. Therefore, if the new objects are confirmed by experts, the related ambiguous attributes (minor attributes), which might result in the marginally acceptable CF values of weak embedded rules, could be refined or new attributes could be added to improve the classification ability. If the initial data type of a minor attribute is too rough to describe the object, a superior data type is recommended and the values of the attribute in both original object and new evolved object should be modified. For example, the BOOLEAN data type may be refined to SINGLE VALUE data type (Hwang & Tseng, 1990). If changing the data type still cannot discriminate the new variants from original objects, acquiring new attri- butes from domain experts will be suggested in the new objects acquisition module. According to the complexity of relations between objects and attributes or even the rela- tions between different tables, it is hard for experts to coop- erate with each other in building every column and every row for each table. Finally, the result of new objects and corresponding attributes can be used to construct the AT increment. 3.3. Trend evolution analysis Although the original idea of constructing AOT makes EMCUD more adaptive to elicit embedded meanings, the relative importance of all attributes to each object could be adjusted since the dynamic knowledge may change or evolve with the times. It means that some embedded rules, which are recommended by experts now, may become uncertain in the near future. Each object in the AOT is decomposed to record the relative importance of each attribute to the object with the times. Since the traditional Repertory Grid-based KA methods do not record the evolved trend of each new object and the EMCUD is diffi- cult in deciding the ordering of all attributes of the object ency events analysis. S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 839 by experts, the trend evolution analysis, which can discover the evolution of the relative importance of each attribute to each object with the times, is proposed to help experts monitor the significant importance changing of all attri- butes to each object in a time interval. As shown in Fig. 3, the object can be singled out of the old object according to the viewpoints of experts or the learning results of the frequency events analysis. Each attri- bute can be simply assigned as ‘‘0’’ or ‘‘1’’ in each time point for indicating whether it is important to each object or not, where ‘‘0’’ represents the attribute is considered as the unimportant attribute to the object and ‘‘1’’ repre- sents the attribute is important to the object. The domain expert can then decide which attributes are required to be traced with the times if some ordering values of the attri- butes are hard to be decided immediately. The ‘‘0’’ or ‘‘1’’ is called an attribute event et of each object in a time point t, and the attribute event sequence of ‘‘0’’ and ‘‘1’’ is recorded in a table called the Attribute Signal Table (AST) to capture the evolved behavior of each object. Hence, the AOT increment can be generated for evolving the relative importance of each attribute to each object (ordering values) according to the sequence of ‘‘0’’ and ‘‘1’’ events recorded in AST with the times using time series analysis. Since the ‘‘1’’ means an attribute is impor- tant to an object, the consecutive ‘‘1’’ recorded in consecu- tive time points indicates that relative importance of the object should become higher. On the contrary, the consec- utive ‘‘0’’ indicates that the relative importance of the object should be lower. Hence, a simplified time series anal- ysis is proposed to capture the trend meaning and incre- mentally adjust the CF value of each rule. Let the initial value of each signal sequence be the original AOT value of the attribute to the object. 3.3.1. Dynamic AOT adjusting function Since the knowledge will be updated or evolved in a dynamic environment, the CF value of each embedded rule may be adjusted because the relative importance of the object may change. A Dynamic AOT Adjusting Function (2) is designed to generate the updated AOT value at time t by accumulating the collection of attribute event et at time point t based upon the previous AOT value at time t � 1. If the attribute event et is assigned as 1 then c is set to 1, which represents the increment is added into the AOT Fig. 3. Trend evol value at time t � 1. Otherwise, c is set to �1 if the et at time t is 0, which represents the decrement subtracted by the AOT value at time t � 1. Hence, the ordering values can be refined with the times according to the collected infor- mation in a changing environment: AOTðtÞ¼AOTðt�1Þþc�fðgðtÞÞ; c¼1; if et ¼1 c¼�1; if et ¼0 ( ð2Þ where f(g(t)), which is formally defined in formula (3), is used to decide the increment or the decrement of the corre- sponding the AOT value at each time point t, a, which is used to adjust the curvature of the AOT Delta Function, increases resulting in rapidly increasing or decreasing of the CF value, and b, which means the weight of the number of consecutive ‘‘1’’ or consecutive ‘‘0’’ received, decreases resulting in larger increment or decrement. In order to limit f(g(t)) between 0 and 1, the constant c is suggested to be smaller than �3: fðgðtÞÞ¼ 0; if et 6¼ et�1 1 1þea�ðcþgðtÞ�bÞ ; if et ¼ et�1 ( ð3Þ where g(t), the Continuous Events Accumulating Function given in formula (4), is used to record the number of con- secutive ‘‘1’’ or consecutive ‘‘0’’ received at time t: gðtÞ¼ 1; if et 6¼ et�1 gðt � 1Þþ 1; if et ¼ et�1 � ð4Þ 3.4. Grid merging In order to maintain the new discovered new object, we propose grid merging algorithm to integrate the AT increment and AOT increment into the main AT and the main AOT, respectively. Therefore, the small AT and the small AOT instead of the whole large main AT and the main AOT are used to update the embedded rule base. Algorithm 2: The grid merging algorithm ution a Input: The main AT, main AOT, AT increment AT0, and AOT increment AOT0. Output: The updated main AT and main AOT Step 1: Integrate the AT increment AT0 into the main AT. nalysis. 840 S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 Step 1.1: Append each new object and each new attribute in the AT0 as a new column and row in the main AT, respectively. Step 1.2: Ask experts to fill the values of the modi- fied attributes of other objects in the main AT if necessary. Step 1.3: Ask experts to examine the values of the new attributes of other objects in the main AT if necessary. Fig. 4. The concept of knowledge integration in the collaborative KBS. Step 2: Integrate the AOT increment AOT0 into the main AOT. Step 2.1: Expand the size of the main AOT accord- ing to the main AT updated in the Step1. Step 2.2: Fill the corresponding AOT values according to the AOT0. Step 2.3: Refine the values of all old attributes to each old object in the main AOT using the Trend Evolution Analysis if necessary. Step 3: Reset the AT increment AT0 and the AOT increment AOT0. To merge the AT increment into the main AT, each new evolved object should be appended as a new column in the main AT and each new added attribute should be appended as a new row in Step 1.1. In order to maintain the correctness of the main AT, the values of all modified or new added attributes to each object should be acquired by experts if necessary. Since the size of AOT need equal the size of AT, the size of the main AOT should be expanded in Step 2.1 according to the main AT updated in Step 1. Besides the value of all attributes to each new object in AOT increment, the other values of all old attri- butes to each old object could also be learned using the trend evolution analysis to obtain the relative importance at time t. 4. The collaborative knowledge integration framework Although the neo-learning module can be used to single the new objects out of extended object class and to generate their corresponding rules, some new evolved objects which may occur infrequently in each local KBS (but may be fre- quent in the collaborative KBS) cannot be found. There- fore, a collaborative framework is proposed to help experts collect the relevant information and discover these new evolved objects. 4.1. The concept of collaborative knowledge integration In a dynamic environment, new object could be discov- ered using the collaborative framework shown in Fig. 4, which analyzes the related importance of the evolved objects. The infrequent inference logs, profiles, and the discov- ered knowledge of new objects of each local KBS could be collected in our collaborative framework. As shown in Fig. 4, the neo-learning module is firstly used to discover some new evolved objects from the infrequent logs reported from all KBSs. Next, some new evolved objects may occur in some KBSs with similar profiles, e.g., the SQL server running on Windows operation system, the correlations between inference logs and profiles might be useful for helping experts discover them. Finally, for the discovered object, the CF value of the new embedded rule should be recalculated. Hence, a CF adjusting method is proposed to combine the knowledge of new objects discov- ered in each KBS and the collaborative KBS. 4.2. Adjusting certainty factor of dynamic knowledge Assume there are n local KBSs and each new evolved object may be discovered in p local KBSs, different CF val- ues of a given embedded rule could be generated in each KBS. (1) For p > 0, the CF Adjusting Function shown in for- mula (5) is proposed to help experts obtain the average of different CF value of a given embedded rule in each local KBS and adjust the scale of the CF increment or decrement (DCF) according to the discover of the new object in the collaborative KBS. For each new embedded rule Ri, let the CF value be CF(Ri) and let the CFðRjiÞ be the CF value of each embedded rule Ri discovered in the jth local KBS. CFðRiÞ¼ Pn j¼1CFðR j iÞ p þ d � DCF ð5Þ Depending on whether the new objects are discovered in the collaborative KBS or not, the coefficient d can be de- fined as follows: Case 1: the new object can be discovered in the col- laborative KBS. d is set to p/n. Case 2: the new object cannot be discovered in the collaborative KBS. d is set to (p � n)/n. (2) For p = 0, since the new object cannot be discovered in any local KBS, the new object could be discovered in the collaborative KBS according to the correlations of profiles. Therefore, the CF Adjusting Function could R Tab An Att Ma Up Exe S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 841 be reduced to formula (6), where the CFðRciÞ is the CF value of the new discovered rule in the collaborative KBS due to different configurations of profile. Table 3 An example of original Nimda AOT Attribute/object Nimda Mail_Attachment 2 CFðRciÞ¼ CFðR c iÞþ d � DCF ð6Þ dis set to �2. For example, the SQL Slammer uses UDP port 1434 to exploit a buffer overflow in a MS SQL server to simply switch off this port of the victim host. The collaborative KBS can learn this knowledge based upon the infrequent logs reported from some local KBSs according to the same service stored in profile. Some new objects may occur in similar environment. 5. Case study and experiments Up to now, many antivirus products have been devel- oped to discover worms, virus or Trojan horse in a com- puter system. However, these products are hard to automatically discover the variants of worms because the signature based approach fails when the signatures are changed. To overcome the weakness, we propose a worm detection prototype system, which has neo-learning mod- ule, to enhance the ability of commercial antivirus products by the collaborative framework. 5.1. Computer worm detection Nimda, an incredibly sophisticated worm that made headlines worldwide, is taken as an example. Assume a simple Nimda concept tree is created in Fig. 5 after series of Nidma cases diagnosis, and it can be transformed into a worm Acquisition Table (AT) like Table 2. The following attributes are considered: the name of the e-mail attach- ment used by worms, the medium used by worms to upload, and the name of the file used by worms to start exe- Nimda Mail_Attachment Upload_Medium Excuted_File_Name Symptoms eadme.exe cool.dllputa!!.scr Admin.dll carrier Fig. 5. Example of initial Nimda concept tree. le 2 example of original Nimda AT ribute/object Nimda il_Attachment Readme.exe load_Medium Admin.dll cuted_File_Name Riched20.dll cution on servers. After constructing the worm AT, we construct the AOT table shown in Table 3. With both AT and AOT, the EMCUD method can be processed. The eight embedded rules are generated and some of them have low CF value such as rule R1: ‘‘IF Not Mail_Attachment = Readme.exe and Upload_Mediu- m = Admin.dll and Executed_File_Name = Riched20.dll Then Nimda’’ with CF value = 0.67. Therefore, suppose that in the inference process, the rule R1 above is learned by neo-learning module almost all the time during a period, and suppose in the last two time points the embedded rule R2: ‘‘IF Not Mail_Attachment = Readme.exe and Not Upload_Medium = Admin.dll and Not Executed_File_ Name = Riched20.dll Then Nimda’’ with CF value = 0.4 is fired, the AST in Table 4 can be obtained. Suppose that Nimda is the latest worm occurred in the world, its ordering value of each attribute cannot be easily determined because its variants may soon be broken out. The expert may define an AST with several time points, and then assign 0 in the first attribute, N1, at first time point in Table 4. The attribute event N2 at the second time point is set to zero. Next the time ser- ies analysis method first calculates the AOT according to the AST. In Table 4, the Mail_Attachment attribute is calculated by time series analysis method, and the attribute is assigned a new ordering value = 1 since it is very possible to be changed again, subsequently, ordering value = 3 are assigned for both attributes Upload_Medium and Excu- ted_File_Nam according to the AST. Therefore, the CF value of the rule R1 is leveled up from 0.67 to 0.74. More- Upload_Medium 3 Executed_File_Name 4 Table 4 An example of Nimda AST Attribute/object N1 N2 N3 N4 N5 N6 N7 Mail_Attachment 0 0 0 0 0 0 0 Upload_Medium 0 1 1 1 0 0 0 Executed_File_Name 1 0 0 0 1 0 0 Table 5 An example of updated Nimda AT after discovering Nimda.B Attribute/object Nimda.A Nimda.B Mail_Attachment Readme.exe puta!!.scr Upload_Medium Admin.dll Admin.dll Executed_File_Name Riched20.dll Riched20.dll Table 6 An example of integrated Nimda AT Attribute/object Nimda Mail_Attachment {Readme.exe; puta!!.scr} Upload_Medium Admin.dll Executed_File_Name Riched20.dll Table 7 An example of updated Nimda AOT after discovering Nimda.B Attribute/object Nimda Mail_Attachment 1 Upload_Medium 3 Executed_File_Name 3 Fig. 6. The collaborative framework for worm detection. 842 S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 over, several new attribute-values are learned by neo-learn- ing module with Mail_Attachment = puta!!.scr in R1, a new worm variant Nimda.B shown in Table 5 can be inte- grated into Table 6, and also an AOT is updated as shown in Table 7. Therefore, with the accumulated inference logs from dis- tributed sensors, the trend evolution analysis mechanism can also update the knowledge frequently. Assume neo- learning module learns another new attribute values including Mail_Attachment = sample.exe, Upload_Me- dium = cool.dll, and Executed_File_Name = httpodbc.dll in R2 while the rule R2 has always been fired in each time point in a short period, then a new variant Nimda.E is found. Finally, based upon the updated tables shown in Tables 8 and 9, the built system will give a whole picture of worms to guide the users who are not familiar in the domain for preventing or removing the malicious worms. 5.2. Worm detection prototyping system As we know, many antivirus products (F-Secure, 2005; Symantec, 2006; Trend, 2006) have been proposed to dis- cover worms, virus or Trojan horse in a computer system. However, these products are hard to automatically dis- cover the variants of worms because the signature based Table 8 An example of integrated Nimda AT after discovering Nimda.E Attribute/object Nimda Mail_Attachment {Readme.exe; puta!!.scr; sample.exe} Upload_Medium {Admin.dll; cool.dll} Executed_File_Name {Riched20.dll; httpodbc.dll} Table 9 An example of updated Nimda AOT after discovering Nimda.E Attribute/object Nimda Mail_Attachment 1 Upload_Medium 2 Executed_File_Name 2 approach fails when the signatures are changed (Moore, Shannon, Voelker, & Savage, 2003). To overcome the weakness, the worm detection prototype system is pro- posed to enhance the commercial antivirus products instead of replacing them. By only updating the knowledge base, the detection system can modify the defense mecha- nism for the variants of worm; as a result, the system can be easily maintained. Since the growth of the knowledge of worms is very fast, we propose a collaborative architec- ture for the adaptive worm detecting problem. Fig. 6 shows the collaborative framework for worm detection. In the architecture, the neo-learning module helps each worm sensor constructing AST to reconstruct AOT increment and update main acquisition table using AT increment (monitoring the frequent inference logs of weak embedded rules of worms with the times), where each sensor has its own Worm KB. By collecting the new worms knowledge and infrequent inference logs and consulting the Profiles, the collaborative framework can integrate the new worm knowledge. Each worm sensor provides a web interface to collect or discover all the symptoms of worm cases by user and scan- ning tools. For example, when worm infects a victim sys- tem, the user can scan the host computer by some general antivirus software or can call for help from the Internet. The system collects all the information and infers the information based upon the worm knowledge with embedded meaning constructed by EMCUD. Conse- quently, the result of inferring will be passed to the users to teach the way of recovering the system. Moreover, the statuses which satisfy certain embedded rules will be con- sidered to learn the new knowledge of new variant worms by neo-learning module. 5.3. Experimental results We implemented a dynamic EMCUD web-based sys- tem, and used computer worm as an experimental domain, Table 10 The ratio of discovering new evolved worm Original worms New worms (%) Inference costs (%) Frequent based 100 60 193 Frequent basedR 100 100 294 Collaborative 100 100 200 S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 843 where three computers were equipped with the dynamic EMCUD system and each one constructed its own knowl- edge base to evaluate the performance of discovering vari- ants. We generated all kinds of test samples including the behaviors of original worms and variants (polymorphic worms) to randomly attack the network. The experimental result in Table 10 shows that the collaborative framework can successfully discover the variants by neo-learning mod- ule. Since some critical weak embedded rules may be ignored in the beginning of knowledge based construction, some specific variants which cannot be discovered by any individual neo-sensor can be detected by the rules. The frequency based analysis module can detect only 60% variants. Obviously, when the knowledge is recon- structed, neo-learning module can learn all of the variants. However, it needs extra inference costs, the growth of the number of rules, about 100. On the other hand, our collab- orative framework can learn all of the variants, it needs only seven extra inference costs to reach the goal in this example. Thus, our collaborative framework could be effi- cient in discovering new knowledge. 6. Conclusion In this paper, the dynamic EMCUD based upon Reper- tory Grid is proposed to elicit the embedded meanings of knowledge. Dynamic ENCUD can generate an AT incre- ment and an AOT increment to represent the evolved objects and to record the relative importance of each attri- bute to each object for capturing the embedded meanings with acceptable CF value by relaxing or ignoring some minor attributes. A collaborative knowledge acquisition framework is proposed to analyze the correlations of inter- esting inference logs of embedded rules between multiple local KBSs in a dynamic environment to discover the new evolved objects. Each KBS can monitor the frequent inference behaviors of weak embedded rules to construct an AT increment and analyze the significant change of the importance to evolved objects to construct an AOT increment for adjusting the relative importance of each attribute to each object with the times. Moreover, three cases are used to assist experts in adjusting the CF values of the discovered knowledge of the new evolved objects from the collection of inference logs. We implemented a worm detection prototype system to evaluate the perfor- mance of dynamic EMCUD to incrementally integrate evolved knowledge into knowledge base. Acknowledgement This work was partially supported by National Science Council of the Republic of China under Grant No. NSC95-2752-E-009-015-PAE. References Adams, J. (1985). Probabilistic and certainty factors. In B. Buchanan & E. Shortliffe (Eds.), Rule-base expert systems: The MYCIN experiments of the stanford heuristic programming project. Reading, MA: Addison- Wesley. Boegl, K. (1997). Design and implementation of a web-based knowledge acquisition toolkit for medical expert consultation systems. Doctorial thesis, Technical University of Vienna, Austria. Boicu, M. (2001). Automatic knowledge acquisition from subject matter experts. IEEE International Conference on Tool with Artificial Intelligent. Boose, J. H. (1984). Personal construct theory and the transfer of human expertise. In Proceedings of AAAI-84 conference (pp. 27–33), California. Boose, J. H. (1985). A knowledge acquisition program for expert systems based on personal construct psychology. International Journal of Man– Machine Studies, 23(5), 495–525. Boose, J. H., & Bradshaw, J. M. (1986). NeoETS: Capturing expert system knowledge in hierarchical rating grids. In IEEE Expert System in Government Symposium. Boose, J. H., & Bradshaw, J. M. (1987). Expertise transfer and complex problems: Using AQUINAS as a knowledge-acquisition workbench for knowledge-based systems. International Journal of Man–Machine Studies, 26(1), 3–28. Cairo, O. (1998). KAMET: A comprehensive methodology for knowledge acquisition from multiple knowledge sources. Expert Systems with Applications, 14(1), 1–16. Castro-Schez, J. J., Jennings, N. R., Luo, X. D., & Shadbolt, N. R. (2004). Acquiring domain knowledge for negotiating agents: A case of study. International Journal of Human–Computer Studies, 61(1), 3–31. Crowther, P., & Hartnett, J. (1996). Using repertory grids for knowledge acquisition for spatial expert system. In Proceedings of Australia and New Zealand Conference on Intelligent Information Systems (pp. 14– 17), Adelaide, SA, Australia, November 18–20. Davis, R. (1987). Interactive transfer of expertise. Artificial Intelligent, 56, 121–157. Diederich, J., Ruhmann, I., & May, M. (1987). KRITON: A knowledge- acquisition workbench for knowledge-based systems. International Journal of Man Machine Studies, 26, 3–27. Dixit, A. K., & Pindyck, R. S. (1994). Investment under uncertainty. Princeton University Press. Eshelman, L., Ehret, D., McDermott, J., & Tan, M. (1987). MOLE: A tenacious knowledge acquisition tool. International Journal of Man Machine Studies, 26, 41–54. F-Secure Corporation (2005). http://www.f-secure.com/. Gaines, B. R. (1987). An overview of knowledge-acquisition and transfer. International Journal of Man–Machine Studies, 26, 453–472. Ginsberg, A., Weiss, S. M., & Politakis, P. (1988). Automatic knowledge base refinement for classification systems. Artificial Intelligence, 35(2), 197–226. Hwang, G. J. (1995). Knowledge acquisition for fuzzy expert systems. International Journal of Intelligent Systems, 10, 541–560. Hwang, G. J., & Tseng, S. S. (1990). EMCUD: A knowledge acquisition method which captures embedded meanings under uncertainty. Inter- national Journal of Man–Machine Studies, 33, 431–451. Hwang, G. J., & Tseng, S. S. (1991). On building a medical diagnostic system of acute exanthema. Journal of Chinese Institute of Engineers, 14(2), 185–195. Kang, B. (1996). Multiple classification ripple down rules. Ph.D Thesis, University of New South Wales. http://www.f-secure.com/ 844 S.-C. Lin et al. / Expert Systems with Applications 34 (2008) 833–844 Kelly, G. A. (1955). The psychology of personal constructs, Norton, New York. Kolousek, G. (1997). The system architecture of an integrated medical consultation system and its implementation based on fuzzy technology. Doctoral thesis, Technical University of Vienna, Austria. Leitich, H., Kiener, H. P., Kolarz, G., Schuh, C., Graninger, W., & Adlassnig, K. P. (2001). A prospective evaluation of the medical consultation system CADIAG-II/RHEUMA in a rheumatological outpatient clinic. Methods of Information in Medicine, 40, 213–220. Mcgraw, K. L., & Harbison-Briggs, K. (1989). Knowledge acquisition: Principles and guidelines. Prentice-Hill International Editions, pp. 1–27. Moore, D., Shannon, C., Voelker, G. M., & Savage, S. (2003). Internet quarantine: Requirements for containing self-propagating code. In Proceedings of INFOCOM 2003, March 30–April 3, San Francisco, USA. Pan, D., Zheng, Q. L., Zeng, A., & Hu, J. S. (2002). A novel self- optimizing approach for knowledge acquisition. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 32(4), 505–514. Politakis, P., & Weiss, S. M. (1984). Using empirical analysis to refine expert system knowledge bases. Artificial Intelligence, 22, 673–680. Shaw, M. L. G., & Gaines, B. R. (1987). KITTEN: Knowledge initiation and transfer tools for experts and novices. International Journal of Man–Machine Studies, 27, 251–280. Shaw, M. L. G., & Gaines, B. R. (1996). Web grid: Knowledge modeling and inference through the world Wide Web. In Proceedings of tenth knowledge acquisition workshop (pp. 65-1–65-14). Shortliffe, E. H., & Buchanan, B. G. (1975). A model of inexact reasoning in medicine. Mathematical Bioscience, 23, 351–379. Symantec Corporation (2006). http://www.symantec.com/corporate/. Trend Corporation (2006). http://www.trendmicro.com/tw/home/ enterprise.htm. Wielinga, B., Schreiber, A., & Breuker, J. (1992). KADS: A modeling approach to knowledge engineering. Journal of Knowledge Acquisition, 4(1), 5–53. http://www.symantec.com/corporate/ http://www.trendmicro.com/tw/home/enterprise.htm http://www.trendmicro.com/tw/home/enterprise.htm Dynamic EMCUD for knowledge acquisition Introduction Related work Knowledge acquisition methodologies Repertory Grid methodology and relevant systems Elicitation of embedded meanings The incremental knowledge acquisition methodology - dynamic EMCUD The new evolved object knowledge acquisition Frequency events analysis Trend evolution analysis Dynamic AOT adjusting function Grid merging The collaborative knowledge integration framework The concept of collaborative knowledge integration Adjusting certainty factor of dynamic knowledge Case study and experiments Computer worm detection Worm detection prototyping system Experimental results Conclusion Acknowledgement References 
work_xd4mjaoh6ff7xcofswblgq4n2u	----	Microsoft Word - Revised Manuscript 1 An Intelligent Mobile-Enabled Expert System for Tuberculosis Disease Diagnosis in Real Time Antesar M. Shabut1, Marzia Hoque Tania1*, Khin T. Lwin1, Benjamin A. Evans2, Nor Azah Yusof3, 4, Kamal J. Abu-Hassan5, M. A. Hossain1 1 Anglia Ruskin IT Research Institute (ARITI), Anglia Ruskin University, Chelmsford, UK 2 Norwich Medical School, University of East Anglia, Norwich, UK 3 Institute of Advance Technology, Universiti Putra Malaysia, Serdang, Malaysia 4 Department of Chemistry, Faculty of Science, Universiti Putra Malaysia, Serdang, Malaysia 5 Department of Physics, University of Bath, Bath, UK Abstract This paper presents an investigation into the development of an intelligent mobile-enabled expert system to perform an automatic detection of tuberculosis (TB) disease in real-time. One third of the global population are infected with the TB bacterium, and the prevailing diagnosis methods are either resource- intensive or time consuming. Thus, a reliable and easy–to-use diagnosis system has become essential to make the world TB free by 2030, as envisioned by the World Health Organisation. In this work, the challenges in implementing an efficient image processing platform is presented to extract the images from plasmonic ELISAs for TB antigen-specific antibodies and analyse their features. The supervised machine learning techniques are utilised to attain binary classification from eighteen lower-order colour moments. The proposed system is trained off-line, followed by testing and validation using a separate set of images in real-time. Using an ensemble classifier, Random Forest, we demonstrated 98.4% accuracy in TB antigen-specific antibody detection on the mobile platform. Unlike the existing systems, the proposed intelligent system with real time processing capabilities and data portability can provide the prediction without any opto-mechanical attachment, which will undergo a clinical test in the next phase. Keywords: Image processing; machine learning; decision support system; colourimetric tests 1. Introduction Tuberculosis (TB) is a communicable disease, infecting one third of the world’s population. In 2015, 1.8 million TB-related deaths were reported (Centers for Disease Control and Prevention, 2017). On the other hand, every year about 244 million migrants cross international borders (Department of Economic and Social Affairs, 2016). The carriage of TB in a mobile population is a global challenge, which is a particular concern for the border agencies (Posey, Marano, & Cetron, 2017). However, TB is curable with appropriate early diagnosis. The most common diagnosis procedure for TB is a skin test (Mantoux test) or a blood test (Centers for Disease Control and Prevention.; NHS). Despite many commercial test * Corresponding Author: Marzia Hoque Tania E-mail: marzia.hoque@pgr.anglia.ac.uk 2 schemes, there is still a need for an easy-to-use, effective and feasible point-of-care (POC) TB diagnosis tool, particularly for the remote community where there are very limited or no diagnostic facilities. Such a tool should possess the following features: low cost mobile solution, anytime anywhere access, low energy consumption, ease of use, fast and automatic identification of TB. The World Health Organization (WHO) prefers diagnostic tools which are inexpensive, disposable and easy-to-use (Khademhosseini, 2011; S. Wang, Xu, & Demirci, 2010). A mobile-enabled expert system can address all these features. Due to the high penetration rate of mobile phones (GSMA Intelligence), such system can reach a wider population, especially those who have limited access to advanced laboratory facilities. Incorporation of the mobile phone can not only facilitate an easy and automatic colour detection but also can enable diagnostic disease decision using machine learning techniques. In order to establish a widespread application, the mobile-enabled expert systems should possess minimum hardware requirements. To eliminate the necessity of the opto-mechanical attachment, one requires advanced image processing techniques. The difficulty of choosing the right image processing technique for a mobile platform includes the balance between accuracy, robustness and computation cost. This work aims to develop such a system to provide qualitative TB diagnosis results on the mobile platform in real time. The main contribution of the paper is to ‘automatically’ detect TB-specific antibodies by analysing digital images (i.e. ELISA images) with colour signals produced by biosensor technology. The plasmonic ELISA tests were conducted in Universiti Putra Malaysia. The proposed system does not require any additional hardware such as an opto-mechanical attachment to enhance the colour detection or guide the illumination source, which makes the system the most conveniently portable. Utilising an intelligent image processing algorithm, the presented system robustly separates the samples from the assay plate and extracts the features, and within a few seconds the system predicts the class label via a machine learning algorithm with high accuracy and ease of use. On a trained model, when a user provide an image to test, the system will require to process the image. Sending this user input directly to the cloud may present certain uncertainty and degradation of the image quality in resource-limited settings. A local analysis can enable TB testing facility for 24/7 even in the remote areas where internet connection is not available or very weak to send the images to the server, conduct the analysis, and send the result back to the smartphone. Although the proposed system is a native application to provide anytime-anywhere access, the presented system can be integrated to a server. 2. Literature Review 2.1 Computational Systems for TB-detection To the best of authors’ knowledge, there is no existing mobile, desktop or server based system for plasmonic ELISA based detection of TB antigen-specific antibodies. In literature, only a few studies employed machine learning techniques to assist in the diagnosis and monitoring of TB to offer a low- cost, simple, rapid and portable platform. Tracey et al. (2011) utilised acoustic signals to track the 3 recovery of pulmonary tuberculosis patients. The multilayer perceptron (MLP) showed 88.2% accuracy for ambulatory cough analysis. Osman, Mashor, & Jaafar (2010) proposed a tuberculosis bacteria detection technique from tissue sample by Ziehl-Neelsen staining method. The prepared sample image from an optical microscope was segmented by moving k-mean clustering for tuberculosis bacteria extraction. Both RGB and C-Y colour were utilised to acquire a robust and improved segmentation under various staining condition. The hybrid multilayered perceptron network (HMLP) selected the features among the geometrical features of Zernike moments to detect tuberculosis bacteria. The result showed 98.0%, 100% and 96.19% of accuracy, sensitivity and specificity respectively to find the class of definite and possible TB. Tsai, Shen, Cheng, & Chen (2013) developed colorimetric sensing using unmodified gold nanoparticles and single- stranded detection oligonucleotides for a TB test. The focus of the work was salt-induced AuNP colourimetric diagnosis for sensing target TB DNA sequences without multiple PCR cycles to amplify specific MTB target DNA sequences from extracted sputum or tissue samples. A smartphone was utilised just to collect the multiple detection results of colour variation from the concentration on cellulose paper and transmit the data to the cloud. Table 1: TB related mobile applications on the Android platform User Region Aspect Questionnaire Intelligent Systems Ref. Department of Health South Africa Management; TB and HIV diagnostic data  X (Interactive Health Solutions, 2016a) Specific Users Bangladesh Management  X (Interactive Health Solutions, 2017) Mine community South Africa TB screening  X (Interactive Health Solutions, 2016b) Patients Pakistan Control TB and drug-resistance  X (Interactive Health Solutions, 2016) Clinicians Global Decision on rapid diagnosis of TB and resistance X X (Open Medicine Project, 2014) Clinicians and Patients Cambodia Track lab test result X X (Operation Asha, 2017) There are commercial and endorsed mobile applications for TB in the popular application stores e.g. Google Play (Table 1, searched on 21-09-2017) and Apple app store. When it comes to diagnosis, the applications are for screening purpose only (Interactive Health Solutions, 2016, 2016, 2017). These applications store the screening data via the OpenMRS server. Either the user needs to insert the answers to a series of questions or the lab test results have to be manually inserted by the user or clinician. The available applications can ensure the data portability (Table 1) and in some cases diagnostic decision (Open Medicine Project, 2014), however they lack automation to produce a diagnostic result from the 4 specimen. Thus, there is a need for a system that does not require any additional hardware e.g. a plate reader and can produce laboratory scale test results. 2.2 Image Processing on mobile platform An image processing based automatic system to be implemented on mobile platform firstly requires quality assessment and size reduction of the image. The quality assessment of the image will increase the accuracy of the system. The reliability will also increase due to the consistency in the input. The size- reduction and quantisation will make the system faster. For a mobile enabled decision support system, Bourouis et al. (2014) utilised a normalisation function to resize the retinal images to 32x32 pixels before storing 1-dimentional vector of pixel information. Lot of emphasis is provided in literature and commercial mobile applications for adjusting environmental condition and colour and light exposure correction (US9563824 B2, 2017; Wug Oh,Seoung; Kim, 2017). The image segmentation algorithms in the literature can be mainly categorised based on five of the following methods: histogram thresholding, edge detection, clustering, region-based and graph-based methods (X.-Y. Wang, Wu, Chen, Zheng, & Yang, 2016). An alternative to segmentation is often carried out e.g. ELISA Plate Reader (Enzo Life Sciences inc., 2015) and AssayColor (Alidans srl, 2015). Both applications use a guideline e.g. grid or well structure to ensure a better image from a naïve user. Such a guideline can help the user to maintain an adequate distance of the sample from the camera, compromising the flexibility of the assay type. In both cases, the well-to-well distance is restricted because of certain assumptions regarding the plate size. Instead of any intelligent segmentation technique, few works in the literature (Mutlu et al., 2017; Ozkan & Kayhan, 2016) used cropping. It is highly discouraged for two reasons: i) it would require cropping skill from the user, and ii) it reduces the ease of use. 2.3 Colourimetric classification and decision on mobile platform After processing the images, the features require analysis to generate a diagnostic decision and present it on the mobile platform. The related works done in the literature are mostly for paper based assays (Kim et al., 2017; Solmaz et al., 2018), which are less complex than the wet chemical assays. A cloud based mobile application was demostrated to classify peroxide content from mean RGB, HSV and LAB under diverse lighting environments (Solmaz et al., 2018). The least squares SVM and Random Forest were utilised to provide binary and multi-class classification respectively. The maximum accuracy at training phase (with 10-fold cross-validation) was 95% and on the mobile platform, it was reduced to 90.3%. On the other hand, analysing the colour features e.g. average, mode, median, mean, and centroid from the histogram of four colour spaces, the saliva-alcohol concentration was determined by Linear discriminant analysis (LDA), Support vector machine (SVM) and Artificial neural network (ANN) (Kim et al., 2017). The accuracy varied for different classes. Kim et al. (2017) showed that the stand-alone mobile application is two times faster than the server based application. There are few mobile applications available for 96 well enzyme-linked immunosorbent assay (ELISA) based colour detection in the commercial and public app stores e.g. Spotxel® Reader (Sicasys Software 5 GmbH, 2017), Enzo ELISA Plate Reader (Enzo Life Sciences inc., 2015) and AssayColor (Alidans srl, 2015). Enzo ELISA Plate Reader (Enzo Life Sciences inc., 2015) and AssayColor (Alidans srl, 2015) neither provide any automatic complete analysis, nor include any decision support system (DSS) to interpret the colourimetric results. The Spotxel® Reader (Sicasys Software GmbH, 2017) comprising plate annotation and alignment, uses powerful noise processing and signal detection techniques. Instead of intelligent sensing, the application uses a virtual plate which can be laid over the plate image. The application expects the wells to be aligned with the virtual plate. The user is required to match the corner and centre wells with the grid. The virtual plate or grid can be scaled and rotated. However, aligning the wells with the grid requires some image capturing skills, which reduces the ease of use. The developers also acknowledged the limitations in the image processing (Sicasys Software GmbH, 2017). The application is capable of performing statistical analysis to quantify the result. The accuracy of such quantification is yet to be revealed. Clearly it is evident from the recent literature and commercial mobile application stores (Table 1), that there is no existing low cost mobile solution which can benefit the wider population by anywhere anytime access to perform convenient confirmatory diagnosis of TB. To develop such a system, the critical review of the literature suggests to us the following findings: - A strong image processing technique is required to eliminate the opto-mechanical attachments. - Such an image processing technique has to be computationally feasible to be executed in the mobile environment. - The image processing technique has to be intelligent and robust for wet chemical analysis. It should also consider powerful noise filtering techniques. - The model needs to be trained off-line before deploying on the mobile platform. A native application would be faster than cloud based solutions, can be availed anytime anywhere and would possess less concern regarding cyber security. 3. Methods 3.1 Data Collection 3.1.1 Sample preparation The experiments on plasmonic ELISA were mainly conducted in Universiti Putra Malaysia (Abuhassan et al., 2017; Tania, Lwin, Abuhassan, & Bakhori, 2017). However, the TB patient sputum samples were provided by School of Medical Sciences, Universiti Sains Malaysia, Kubang Kerian, Malaysia through their University’s hospital. The fresh sputum sample were delivered to their lab and smear microscope analysis were carried out prior to culture method. The ELISA analysis was carried out simultaneously in the same lab. For the detection of CFP-10, a 10 kDa secreted antigen from Mycobacterium tuberculosis, we first coated the ELISA plate with 100 μL of CFP-10 in carbonate buffer and then incubated for 1.5 h. Following the period, the plate was washed three times with PBS pH 7.6 and 0.05 % Tween-20 (PBST) by tapping it 6 against a clean paper towel. Now the plate was blocked with 370 μL of PBS containing BSA (PBSA) (1 mg/mL) for 1.5 h. All the antibodies and enzyme conjugates were diluted in diluent antibody containing PBST and 1% BSA. The plate was washed with PBST for three times, and kept the plate (invert) at 4°C for 2 h. Now, 100 μL of monoclonal anti CFP-10 antibody as primary antibody was added to the plate at 4°C for 1.5 h. After 1.5 h, the plate was washed with PBST for three times and the plate was added with 100 μL of biotinylated polyclonal secondary antibody and incubated for another 1.5 h at 4°C. The plate then washed three times and 100 μL of catalase-streptavidin conjugate (v/v 1:20) was pipetted into the plates and left for 1.5 h at 4°C. After the period, the wells in the plate were washed three times with PBST, two times with PBS, one time with deionized water and then dried. Now, 100 μL of hydrogen peroxide (in 1 mM MES, pH 6.5) buffer was pipetted into the wells. Immediately, 100 μL of gold ion solution freshly prepared in 1 mM MES buffer was added to the wells prepared in 1 mM MES buffer was added to the wells at room temperature. At this stage, the GNPs formation in the form of coloured solution can be seen and this can be read with microplate reader at an absorbance of 550 nm. For the analysis of real samples, the sputum from positive and negative TB patients were diluted in 4% sodium hydroxide first, and then proceeded to the same coating process as mentioned above. 3.1.2 Image acquisition The dataset (generated as stated above) contains 252 images and 4 videos (Tania et al., 2017); 106 of them, captured with an iPhone 8-megapixel camera without mobile phone holder, were initially considered for (Abuhassan et al., 2017). Blurry images, images with inadequate camera exposure, observations intended for biosensor optimisation and the initial experiments where the colour widely varied from the final representative colours were removed. Finally, 27 images were selected from 22 independent observations. Among these images, 13, 3 and 2 images were captured by Samsung Galaxy J5 Prime (13-MP), iPhone 7 plus (12-MP) and iPhone 6 (8-MP) cameras respectively. The remaining images were captured with an iPhone (8-MP camera). The dataset contains images of 96 wells, which are partially filled, which means the plates contain both empty wells in addition to wells filled with sample. The final selection of 27 images from 22 independent observations were taken in a laboratory lighting environment. These images contain 266 samples - 81 of them are positive for TB-specific antibody, 181 are negative and three of the samples failed to produce any indicative result, thus 263 samples were finally selected. A mobile phone holder (NJS Telescopic Music Record Mobile Phone iPad iPhone Stand Inc G Clamp Mount 68G) was used while capturing the image. However, the acquired images vary in terms of well size, camera to ELISA plate position, light exposure and mobile phone. Considering a robust application, this variation is expected in the real life incoming images. 7 1 2 3 4 5 6 7 8 9 10 11 12X Y A B C D E F G H Z Cp Fig. 1: Impact of sample and camera position with respect to ELISA plate. X and Y are the length and width of the ELISA plate respectively. Z= volume of sample in the well and Cp= camera position. Let us assume, the assay plate, 𝐴 = 𝑓(X , Y , Z ), where {X, Y} ∈ ℤ+ and Z ∈ ℝ and Z > 0 . In this work (Fig. 1), X = {1,2, … , 12} and Y = {𝐴, 𝐵, … , 𝐻}. For the commercially available 96 well plates X and Y will maintain such positions in rows and columns. The space between these wells can vary from plate to plate. Thus, the wells are signified in (x, y, z) coordinates. Each well denoted by 𝑤 , ∈ 𝑤 , in the plate and 𝑠 , ∈ 𝑤 , = 𝑠𝑎𝑚𝑝𝑙𝑒, i.e. the well is filled with the sample. Both shape and depth of the well can vary, depending on the specification of the assay plate. Due to the dimension of the well itself, the distance between these wells can differ from plate to plate. Depending on the biochemical protocol, the amount of sample to fill these wells can vary as well. All this information has a direct impact on the imaging. However, the colour of each sample, 𝑠 (𝑟, 𝑔, 𝑏) ≠ 𝑓(𝑥, 𝑦, 𝑧). We have maintained the camera position (Cp) parallel to the A, giving the wells a uniform exposure to the camera. For a static Cp, the distance between Cp to each 𝑤 , is not equal. Thus, the sample to camera exposure is not equal. In theory, it would make 𝑠 (𝑟, 𝑔, 𝑏) appear as 𝑠 (𝑟, 𝑔, 𝑏). The best exposure would be attained by the median 𝑤 , . The 𝑠 (𝑟, 𝑔, 𝑏) can potentially differ due to the ambient conditions such as temperature, weather and geo- location, and certainly for the sample itself. However, this work is conducted in the laboratory environment. 3.2 Image pre-processing and segmentation 3.2.1 Image pre-processing The goal of this work is to provide TB diagnosis on mobile platforms. Thus, this paper intends to circumvent the limited memory and processing power of the mobile devices, which is why the size of the images need to be reduced. The acquired images were scaled i.e. proportionally resized to reduce the processing time. After the size reduction, the Gaussian Blur filtering was utilised as the image enhancement technique. This low pass filter (LPF) with parabolic amplitude Bode plot detracted the detail of the image by using a Gaussian function on each pixel. 8 𝐺(𝑥, 𝑦) = 𝑒 ………(1), where 𝜎 = 𝑠𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝑑𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛 In an image, x being the distance from the origin in the horizontal axis, y being the distance from the origin in the vertical axis, 2D Gaussian or normal distribution can be written as Eq. (1). Alternatively, a local Laplacian filter with contrast-limited adaptive histogram equalization was also implemented in the desktop application to evaluate the reformation. Most commonly, the images captured by smartphones are in the RGB format. After smoothing, the image needed to be taken into a more perceptually linear colour space, LAB. This colour space transformation provides the ease to perform Euclidean distance calculation-based clustering at a later stage. 3.2.2 Image segmentation Initially, a number of segmentation methods were implemented such as Otsu (Otsu, 1979), multi-level Otsu (Liao, Chen, & Chung, 2001), watershed (Meyer, 1994), super pixel (Ren & Malik, 2003) and k- means (Forgy, 1965; Lloyd, 1982; Macqueen, 1967). In the previous study (Abuhassan et al., 2017), k- means clustering showed promising performance, where the number of cluster was 6 and the size reduction was minimum. Table 2: Major steps of the Algorithm Input: Images of plasmonic ELISA plates Output: Result Steps: 1) Read the images in Red-Green-Blue (RGB) colour space 2) Dynamically scale the image based on the initial size 3) Smooth the image using Gaussian Blur filtering. 4) Convert the image into the CIELAB colour space 5) Use colours in the ab space to measure the Euclidean distance for clustering 6) Select k = 4 7) Dynamically repeat step to avoid local minima 8) For clusters 1 to k, separate the objects using the index clustering. This will produce k images. 9) Convert k images to binary images 10) Use morphological transformation includes Dilation and Erosion 11) Identify the optimum cluster(s) by calculating the difference between the produced images and the white colour (the image with the lowest distance is the optimum cluster) 12) Use Canny edge detector to sharp edges 13) Apply the Find Contours to the optimum cluster. This step will produce images equal to the sample wells in the segmented image 11) Read through the well images and apply size and position noise filtering 12) Extract the histogram features from the image 13) Save the values in the .csv file 14) Repeat steps 11 to 13 if more wells left, If not go to 15 15) Pass the .csv file to the classifier 16) Draw the result of positive or negative 17) Show the result to the user 9 The qualitative test determines the presence or absence of a substance. Thus, the decision is in the binary form. For the naked-eye tests, these binary classes are supposed to be visually distinguishable. Therefore, in theory there are only 3 relevant colours: background, foreground containing positive samples and alternatively negative samples. Hence, it can be hypothesised that k=3 should provide a perfect segmentation for a qualitative colourimetric test holding both positive and negative samples. However, if the foreground pixels of positive and negative samples are in two different clusters, it would make the sample annotation unnecessarily complex and computationally expensive. Thus, it would be desirable to force the clustering method to keep the positive and negative samples in the same cluster. During the segmentation process, the random selection of cluster centroid position at initial stage compelled a requirement for the best cluster persuasion. Thus, after the clustering, a series of post- processing techniques were applied. As mentioned in Table 2, these post-processing techniques include morphological operation encompassing Dilation and Erosion followed by object detection. The morphological transformations on a binary image in most cases require two inputs: the image and the kernel which identifies the nature of the operations. The contours were exploited after the segmentation and morphological transformations for size analysis and object detection. ELISA plate l a b Segmented well or sample mean mode std. Deviation skew- ness energy entr opy mean mode std. Deviation skew- ness energy entr opy mean mode std. Deviation skew- ness energy entr opy Features Classifier Fig. 2: Feature analysis framework 3.3 Feature analysis and classification Once the samples (ROI) were separated, the characteristics of these samples were analysed. In this paper, the feature analysis involves measurement of colour moments. This work includes basic features necessary to compute any probability distribution (Sergyan, 2008). The framework is illustrated in Fig. 2. As described in Abuhassan et al. (2017), mean, mode, standard deviation, skewness, energy and entropy in L, a and b channel (18 features in total) were considered to train the model. 3.4 Mobile-enabled expert system In this work, the plasmonic ELISA-based TB detection was deployed on the Android platform. The mobile application was developed on a Samsung Galaxy S7 edge. The minimum target SDK is 21 (API level 5). 10 Android S tudio Machine Learning Algorithm Image processing Training Image Database Image selection Noise removal techniques Sample s election Feature Analysis Cluster Selection Case Seperation Contour Detection & Filtering Pre-processing & Segmentation Classifier Model Label Weka OpenCV Testing Feature Selection Sample Selection Classifier Model Weka OpenCV Image Capturing Load Model Prediction Result Android Studio Android Studio Noise removal techniques Image processing Pre-processing & Segmentation Image selection Cluster Selection Fig. 3: System framework: Implementation of the Algorithm The steps outlined in Table 2 were implemented on the Android platform as illustrated in Fig. 3. Due to convenient functionality on the Android platform, OpenCV was utilised to perform the data pre- processing i.e. image processing and feature extraction. The feature values of the segmented, individual sample (well) were stored as text and carried to Weka to train the classifier model offline. The offline training was conducted on a 64 bit Windows system with Intel ® Core ™ i7-4770 CPU at 3.40 GHz processor and 16 GB RAM. Once the model is trained, it was loaded on the Android platform using Weka library (weka.jar file). At the testing level in Fig. 3, the user can use any new image of the plasmonic ELISA test on an Android device to produce the correct prediction of TB disease in real time. 4 Results 4.1 Image acquisition of Plasmonic ELISA Plasmonic ELISA links the colour of plasmonic nanoparticles to the presence or absence of the analyte (target protein). Mycobacterium tuberculosis ESAT-6-like protein esxB (CFP-10) was used as a target protein biomarker for the TB detection Plasmonic is accomplished by linking the growth of gold nanoparticles with the biocatalytic cycle of the enzyme label. The protocol adapts a conventional ELISA procedure with catalase-labelled antibodies. The enzyme consumes hydrogen peroxide (H2O2), and then gold (III) ions are added to generate gold nanoparticles. The concentration of hydrogen peroxide dictates the state of aggregation of gold nanoparticles. This allows for the naked-eye detection of analytes by observing the generation of blue- or red-coloured gold nanoparticle solution. 11 Positive sample (Blue) Negative sample (Pink) Positive sample (Blue) Negative s ample (Pink) (a) (b) Fig. 4: Samples in a plasmonic ELISA plate. (a) Samples are hard to visually distinguish, (b) Samples are visually distinguishable In this work, the presence of TB-specific antibodies can be confirmed if the sample turns blue in the ELISA plate. In Fig. 4 (a) gold ions are reduced when H2O2 is present. The top 3 samples are free from TB-specific antibodies. In the presence of H2O2 non-aggregated nanoparticles are formed turning the solution pink. In the bottom 3 samples, the concentration of H2O2 is decreased, turning the samples blue, confirming the presence of TB-specific antibodies. Background ELISA plate Background  ELISA plate  Empty wells  Empty space between wells  Smearing  Shadow  Ambient lighting effect  Shadow  Ambient lighting effect Foreground Positive sample Negative sample  Sample in mid-well  Sample in well boundary  Ceiling light  Ambient light  Shadow  Sample position in the plate  Sample position in the image  Sample-sample dista nce  Positive-negative sample position Fig. 5: Observation of the associated colours and key variables in the image The key observations from the detailed inspection of the dataset are listed below. Obs. 1: In the presented dataset, the sample-to-sample distance was not constant (Fig. 1). If the wells are filled within a close neighbourhood, there is an unavoidable smearing effect. Thus, the background cluster holds many pixels which are close to the foreground pixels. With varying position(𝑥, 𝑦), depending on the class of 𝑠 , , the background cluster is difficult to separate from the foreground clusters. Obs. 2: In some cases, the positive and negative samples are hardly visually distinguishable. For image e.g. Fig. 4 (b), the samples are adequate for naked eye measurement. For sample image e.g. Fig. 4 (a), the indicated sample pair are hard to differentiate. This issue can worsen if the plate contains only one sample and the colour is as ambiguous as in Fig. 4 (a), which can lead to subjective interpretation. Moreover, there is a conscious variation in the sample colour, 𝑠 (𝑟, 𝑔, 𝑏) on independent A. 12 Obs. 3: In the dataset, the value of Z (the volume of sample in a well) had an impact on the size of the sample (𝑠 , ). It implies that the 1 st colour moment can vary based on how the wells are filled. 𝑠 (𝑟, 𝑔, 𝑏) = 𝑓(Z). A well filled up to the surface would have a better exposure when they are positioned at the far edge of the plate. Obs. 4: This work is comprised of wet sample, which is not immune to light reflection from its surroundings (Fig. 5). Initially, this ceiling light was not taken into account. Our hypothesis was: the 𝑤 , with median (𝑥, 𝑦) would be the ideal position for the samples. Even a well filled up to the surface (Obs. 3) in the median position can suffer from the ceiling light reflection. Obs. 5: The impact of ‘camera to well position’ (Fig. 1) is aligned with our prediction in Sec. 3.1. Such influence can be analysed by the SKEW (Fig. 2). The observations Obs. 1, Obs. 3 and Obs. 4 have a clear impact on the image processing measures. The Obs. 2 works in our favour. The qualitative colourimetric tests are usually suitable for naked-eye detection, which necessitates (i) adequate biosensors to produce visually distinguishable colours and (ii) a user who has appropriate colour vision. Firstly, the use of intelligent systems can reduce the biochemical complexity without compromising the accuracy, specificity, sensitivity and reliability. Therefore, the positive and negative samples do not require to be visually distinguishable. Secondly, an intelligent system such as the system we presented in this paper can eliminate the subjectivity of interpretation. A robust system should be able to handle the variation of sample colour mentioned in Obs. 2. (a) (b) (c) (d) (e) (f) (g) (h) (i) (j) (k) (l) Fig. 6: (a) Samples in a plasmonic ELISA plate. Gradual enhancement of the image: (b) sharpened, (c) smoothened, (d) final enhancement before colour space transformation. Quantisation input: (e) full size quantisation, (f) plane- 13 by-plane quantisation, (g) Superpixel, (h) JSEG in MATLAB and (i) Gabor filtering, (j) k-means, (k) Superpixel, (l) Gaussian filtering in OpenCV 4.2 Image Analysis 4.2.1 Image pre-processing At first, the acquired images were scaled and quantised to reduce the size of the image. For a simple method such as Otsu, the impact of scaling on processing time is negligible. On the other hand, to perform numerous iterations, the aid of scaling is obligatory for a heavy segmentation technique such as clustering. In our earlier study (Abuhassan et al., 2017) utilising desktop application, typical images ranging ~3000-4000 pixels were scaled 50%, which requires more reduction to be implemented on mobile platform. Bourouis et al. (2014) utilised 32x32 pixels retinal images, which is not substantial to analyse the colour features of the presented dataset. Moreover, the resizing in Bourouis et al. (2014) was not dynamic. For a known condition, the height and the width of the image will not vary to a great extent. However, it may vary due to factors such as position of the camera, size of the plate, and camera configuration. Thus, the size reduction in this work was performed dynamically (Android Developers, 2018) and proportionally so that the geometry of the ROI was not deformed. The quantisation techniques were carried out to reduce the size of the image by reducing the number of colours in the image (Fig. 6). It was observed that quantisation has insignificant impact on the overall segmentation process. As a result, it was discarded. For a good quality image, such as Fig. 6 (a), the requirement of image enhancement is not high. However, to develop a robust technique applicable to poor quality images, image enhancement is essential. Hence, the images were sharpened, which is a function of resolution and acutance. The radius value of standard deviation of the Gaussian LPF controls the size of the region around the edge pixels that is affected by sharpening. A large value sharpens wider regions around the edges, whereas a small value sharpens narrower regions around edges. The higher value of sharpening will lead toward larger increase in the contrast of the sharpened pixels. A very large value for this parameter may create undesirable effects in the output image, as it may appear as noise. Thus, an edge- aware local contrast alteration was deployed to create more contrast. In the next step, the sharpened image was selectively smoothened to blur the empty wells. The objective of such extensive pre-processing was to ease the segmentation process and minimise the number of iterations. However, these excessively processed images caused separation of cluster 1 and 2 (Eq. 2) in two different clusters at the segmentation stage, as predicted. This led to an elongated object detection step. Hence, the Gaussian Blur filtering (Fig. 6(l)) was utilised as the only image enhancement (negative) technique prior to image segmentation. It assisted to address Obs. 1. 14 Input Watershed Otsu Multi-level Otsu K-means Fig. 7: Image Segmentation using different techniques 4.2.2 Image segmentation The qualitative performance of different image segmentation techniques can be seen from two different input images as shown in Fig. 7 (OpenCV). The Otsu and watershed transformation were unsuccessful in segmenting adequately. The multi-level Otsu performed well for a good quality image with low smearing effect, where the samples are evenly spaced e.g. Fig. 6 (a). However, it failed for many images e.g. second input of Fig. 7. The JSEG was time consuming, not suitable for implementation in real-time. The k-means showed good segmentation performance resembling our early study (Abuhassan et al., 2017). (a) (b) (c) (d) (e) (f) Fig. 8: (a) Input image, (b) Gaussian 2D filtering in MATLAB, (c) first cluster, (d) 2nd cluster with positive and negative samples, (e) 3rd cluster, (f) fourth cluster As mentioned earlier, we require the positive and the negative samples to belong in the same cluster. The use of Gaussian LPF forced the segmentation process to hold the samples in the same cluster as illustrated in Fig. 8. Such pre-processing and segmentation techniques precisely addressed Obs. 1 and Obs. 4. k A cc ur ac y (% ) 69.2 77.22 88.61 88.81 84 96 83 75 3 4 5 6 Without forcing the positive and negative samples to be in the same cluster Algorithm mentioned in Table 2, only with varying k Fig. 9: Performance of the image processing algorithm for different k 15 In our preliminary study, we performed the k-means with k=6 without any pre-processing and with minimum resizing (Abuhassan et al., 2017). As it is mentioned earlier, theoretically the input image should be segmented into three different clusters, which was later found to be imprecise in this work (Fig. 9). Initially, utilising the desktop application (MATLAB), we analysed the impact of varying k without using Eq. (1). Without forcing the algorithm to keep both positive and negative samples in the same cluster, higher k exhibited better segmentation, which is computationally expensive and less suitable to be performed in the mobile environment. Without utilising Eq. (1), the maximum accuracy achieved was 88.81% with k=6 (Fig. 9). It was also observed that the required number of k may vary for image to image due to the image quality, filled well-to-well distance, camera position and positive-negative sample position and ratio per image. A range in the required number of clusters was also observed from the silhouette method (Rousseeuw, 1987), which supports the observation. However, it is not feasible to use multiple iterations to choose a different k each time for each image as it would become computationally expensive. If the ELISA plate contains more filled wells, the execution time is likely to be higher, which makes it unsuitable for mobile applications. In future, we will utilise the image histogram to predict the required number of k before starting the iteration. In contrast, many applications in the commercial app stores simplify the image processing portion by utilising a gridline approach, resulting in compromising the freedom of diverse plate size and in some cases the ease of use. In this paper, we have used k-means with an optimum number of k=4 (as mentioned in Table 2), with complimentary rigorous pre and post processing techniques. With varying k, the overall performance (Fig. 9) of the image processing algorithm (Table 2) varied as well. When k=3, the unsupervised machine learning showed under-segmentation, leading the segmented region to hold image area outside of ROI. When k>4, the algorithm showed over-segmentation, which resulted in more poor performance due to extensive post-processing of the segmented image. Fig. 10: Constraint in segmentation without post-processing: adhered samples after clustering adhered samples 16 Fig. 11: Image processing utilising morphological processing For post-processing, subsequent to clustering, the images were converted into binary images, followed by morphological operation. Without morphological processing, many images would suffer from incorrect object reckoning. After segmentation, in a few cases the samples joined together in association with the noise. If this phenomenon occurs in the right cluster, then the ROI separation process will fail. This problem can be better visualised with Fig. 10, where the samples could not be adequately separated. Due to Obs. 1, few samples are attached together as illustrated in the highlighted image (marked with a red box). All these samples were categorised as a single sample, whereas they are 8 adhered samples. Therefore, the binary dilation and erosion was used to ease the object detection. Erosion operation was iterated four times more than the dilation in order to isolate individual wells and overcome Obs. 1. The size of each (processed) sample in Fig. 11 is much smaller than Fig. 10, which helped to reduce noise, and the samples were no longer adhered (Fig. 11). However, it presented a small possibility of over segmentation due to camera–to-plate position and the type of the plate. To achieve a higher degree of freedom regarding the plate size, a conversation table is required to transform the physical dimension of the assay plate into image pixels. (a) (b) (c) (d) (e) (f) Fig. 12: (a) Segmented image using Fig. 4(a) as the input. (b) Post-processing after segmentation. (c) Final output after contour detection. (d) Segmented image using Fig. 5(a) as the input. (e) Post-processing after segmentation. (f) Final output after contour detection The next challenge was to automatically recognise the best cluster among the four clusters, which was accomplished by exploring the well-to-well background. According to our research, the best cluster is the one that has less white background. As explained earlier, the adversity of Obs. 1 is worsened if this 17 phenomenon happens after the segmentation in the chosen best cluster (Fig. 10). This background acting as noise needed to be filtered from the best cluster (Fig. 11). Finally, the ROIs i.e. samples are separated using contour detection technique. To address all the challenges listed as the observations in Sec. 4.1, selection of a precise post-processing technique was crucial. The 𝑠 , was diverse for the entire dataset, which is expected for a robust use of the application. We demonstrated an intelligent inspection after segmentation to correctly extract the sample from the noise (Fig. 12). The object detection technique functioned accurately even in the case of blurry images in which a larger number of wells were detected and used. 4.3 Feature Analysis and Classification Result The colour moments of the extracted ROI were analysed to train the system offline. The reported articles (Mutlu et al., 2017; Solmaz et al., 2018) mostly feed the mean colour values to the classifiers. We have considered 18 histogram features listed in Table 2 to ensure all the variables are being considered for a robust operation. 112 60.2% 2 1.1% 98.2% 1.8% 1 0.5% 71 38.2% 98.6% 1.4% 99.1% 0.9% 97.3% 2.7% 98.4% 1.6% O ut p ut C la ss Target Class 0 1 0 1 112 60.2% 2 1.1% 98.2% 1.8% 0 0.0% 72 38.7% 100% 0.0% 100% 0.0% 97.3% 2.7% 98.9% 1.1% O ut p ut C la ss Target Class 0 1 0 1 113 60.8% 1 0.5% 99.1% 0.9% 1 0.5% 71 38.2% 98.6% 1.4% 99.1% 0.9% 97.3% 2.5% 98.9% 1.1% O u tp u t C la ss Target Class 0 1 0 1 (a) (b) (c) Fig. 13: Confusion matrix of (a) Random Forest, (b) Random Tree, (c) Random Committee The non-parametric classifiers such as random forest (RF), decision tree, k-nearest neighbours algorithm (kNN), and cubic support vector machine (CSVM) performed better than the parametric classification method e.g. linear discriminant and logistic discrimination. Without cross-validation, all these non- parametric methods produced 100% accuracy. The Multilayer Perceptron (MLP) with backpropagation was comparatively slow and the classification performance was poor as well. It provided 95.2% accuracy. The learning rate was 0.3. In order to circumvent the backpropagation algorithm to be trapped in the local minima, the momentum rate was chosen to be 0.2. There were 500 epochs to train through without decaying the learning rate. The network was allowed to be reset. Both attributes and classes were normalised before training the model. The required number of hidden layers were calculated from the number of attributes and classes . The nodes of these 10 hidden layers were sigmoid. No validation set was used to terminate the training. The model was built in 0.35 seconds. 18 With 30 weak learners, the RF provided a high accuracy (97.2%) in our preliminary study (Abuhassan et al., 2017). The RF showed consistent performance in this study as well. In this work, the bag size in RF was chosen to be 100 without storing out-of-bag predictions in internal evaluation object. The bagging was conducted with 100 iterations and base learner. Only one seed was taken for random number generator. The maximum depth of the trees were kept unlimited and minimum one instances per leaf was allowed to occur. The desired batch size for prediction was chosen to 100. It took 0.09 seconds to build the model in Weka. Table 3. Result of different classifiers in Weka platform Classifier κ TP rate FP Rate Precision F- Measure ROC Area Class Random Forest 0.9775 0.982 0.00 1.00 0.991 1.00 Negative 1.00 0.018 0.973 0.986 1.00 Positive Random Tree 0.9661 0.982 0.014 0.991 0.987 0.984 Negative 0.986 0.018 0.973 0.979 0.984 Positive Random Committee 0.9773 0.991 0.014 0.991 0.991 0.999 Negative 0.986 0.009 0.986 0.986 1.00 Positive Bagged Tree 0.9098 0.956 0.042 0.973 0.965 0.996 Negative 0.958 0.044 0.932 0.945 0.997 Positive Multilayer Perceptron 0.8988 0.947 0.042 0.973 0.960 0.977 Negative 0.958 0.053 0.920 0.939 0.981 Positive In this paper, the RF and Random Committee (RC) both showed 98.9% accuracy with stratified cross validation (10-fold) in the Weka platform. Keeping the batch size, number of seeds, minimum number of instances per leaf as same as RF, the RC was built in 0.01 seconds using default number of iterations (10). The size of the tree varied in each iteration. The Random Tree (RT), a decision tree built on a random subset of columns achieved 98.4% accuracy. Keeping the parameters as same as RC, the Bagged Trees consisting unpruned binary trees provided 95.7% accuracy. The Cohen's kappa coefficient (Cohen, 1960) (κ) is a statistic which compares an observed accuracy with an expected accuracy that can be seen as a random chance. It can be calculated as, κ = , where 𝑝 is the prequential accuracy of the classifier and 𝑝 is the probability that a chance-classifier makes a correct prediction. The result of κ being 1 would signify that the classifier is always correct and 0 would mean that the predictions coincide with the correct ones as often as those of the chance classifier. The κ can provide more precise evaluation than the traditional accuracy metric. Moreover, it can aid in evaluating the classifiers among themselves. From the κ measurement (Table 3), the RF is the best classifier. The κ of RF is in agreement with the accuracy of our previous study as well (Abuhassan et al., 2017). 19 The true positive (TP) rate provides the instances where the samples are correctly classified as the given class. TP rate = ( ) . It can also be expressed as the sensitivity or recall. The highest TP rate was attained by the RF. The false positive (FP) rate or Fall-out provides the instances when the samples are falsely classified as the given class. The precision is the fraction of relevant instances among the retrieved instances i.e. Precision = . The F-Measure provides a combined measure for precision and recall. It can be expressed as, F − measure = × × . The detection ability of the classifier can be better perceived by the receiver operating characteristic (ROC) area (Table 3). Considering the ROC area, the RF is the best classifier for our dataset. The accuracy, specificity and sensitivity can be better visualised from the confusion matrix. The confusion matrix of the top three classifiers are illustrated in Fig. 13. The random feature selection in RF, makes the trees more independent from one another than Bagging, which led to higher accuracy and better bias-variance trade-off. Each tree is able to learn only from a certain subset of features, making it a faster ensemble method as well. Moreover, the RF showed a consistent better performance in various metrics. Similar performance to RF was attained in the MATLAB platform as well. Therefore, we trained our model with RF on the mobile platform. 96.0% 94.0% Progress 98.9% S eq ue nc e o f M et ho d 97.3% 98.4% Image Processing Training Prediction Segmentation (OpenCV) Post-segmentation (OpenCV) Classification (Weka) Prediction on new data (Android) Fig. 14: Accuracy at different stage of the system 4.4 Testing and Validation on the Mobile platform In this work, we demonstrated automatic, real-time TB disease decision making on a mobile platform. The trained model was deployed on the Android platform as illustrated in Fig. 3. To test the efficiency of this mobile-based intelligent algorithm for detecting TB, a separate dataset was used than Sec. 3.1. This new dataset is unknown to the system and contained 61 samples. Among these samples, 20 were positive, 41 were negative and one failed to produce a colour. This held-out validation on the mobile platform ensures the reliability of the system. 20 41 67.2% 0 0.0% 100% 0.0% 1 1.6% 19 37.1% 95.0% 5.0% 97.6% 2.4% 100% 0.0% 98.4% 1.6% O ut p ut C la ss Target Class 0 1 0 1 Fig. 15: Confusion matrix of testing performance at mobile platform On the mobile platform, for this unseen data, the system provided correct prediction for 60 samples. Thus, a final accuracy i.e. from image processing up to TB detection on the mobile platform, of 98.4% was achieved (Fig. 14). The necessity of balanced data can be perceived from the confusion matrix (Fig. 15). The performance of the classifier on a balanced dataset in Weka platform is shown in the Supplementary document. In the absence of a larger dataset, over-sampling or multiple resampling (Estabrooks, Jo, & Japkowicz, 2004) can shed some more light on the performance of the mobile platform. Fig. 16: TB disease detection application One of the biggest challenges of this work was to provide this diagnostic decision on the mobile platform in real-time. The prediction on the mobile platform requires the performing of image processing of the incoming image on the mobile device itself. The processing time is a subject to the number of iterations during image segmentation and object detection and is heavily influenced by Obs. 1, Obs. 2 and Obs. 4. We embraced careful pre-processing techniques to minimise the number of iterations. The k-means uses a random initial value and it is sensitive to the size of the image, thus scaling has a direct impact on this clustering technique regarding the number of k and how the image is being segmented, which justifies the pre-processing used in here. Moreover, we managed to confine the best cluster containing multi- sample of different classes within the same cluster with aid from the Gaussian filtering. This algorithm (Table 2) possess the robustness to deploy the image processing scheme on the mobile platform for other assay plates, which can be further authenticated for the quantitative colourimetric tests. 21 The execution time was recorded for all the images in the dataset. Due to fewer iterations, the image processing occurs in real time liberating the implementation on the mobile platform (Fig. 16). The input image of Fig. 10, containing 20 samples, took 23 seconds to produce the result. The image with only 6 samples e.g. Fig. 6(a) provided the result within 9 seconds. Therefore, it can be concluded that our system is capable of delivering TB disease decision from the plasmonic ELISA image on the mobile platform within ~1-2 seconds/sample (Fig. 16). An image annotation technique was used in the embedded system to identify each sample individually. The Android memory management was utilised to enhance the heap performance by following the correct life cycle of activities in the Android platform. The memory management includes actions with garbage collector (GC), memory optimisation, and tree dominator (Android Developers, 2018). In order to reduce memory leak, the elements used by the system were scaled. The application persistently searched for the objects that were no longer required (garbage) after the life cycle, or reachable which were needed for references. The heap dumps were accumulated over the period of time to determine if there was any growing memory leak. The allocation tracker facilitated a better understanding of the memory usage. In this work, we utilised 8-bit channel ARGB_8888 configuration for the bitmap of image. Although it occupies considerable amount of memory and immediately allocated in the heap, quickly exhausting the memory, this is an optimised choice to maintain the quality of the scaled images. Moreover, this work involves series of image conversations. Therefore, redundant images had to be simultaneously deleted. For garbage collection, a mechanism to remove unnecessary objects to the java application using a virtual machine, Dalvik GC was utilised (Ehringer, 2010). The type of garbage collectors were also examined closely. The application was tested on Samsung Galaxy S6, Samsung Galaxy Note 3 and Samsung Galaxy J3 Prime. On most of these devices, the application performs similarly in terms of the classification accuracy and processing time. 5 Limitations and scope of improvement For any diagnostic system, it is important to note its limitations as well as its capabilities. In the image processing section, in spite of multi-step filtering, the noise due to Obs. 1 needs to be further adjusted. In future, to train the model we will conduct feature optimisation and bias-variance trade-off. The future work should also focus on rectifying the variation between κ and accuracy (Sec. 4.3). We will also explore non-parametric NN with backpropagation, deep NN, hybrid decision tree and naïve Bayes classifiers (Farid, Zhang, Rahman, Hossain, & Strachan, 2014) to investigate a potential enhancement of the accuracy. 6 Conclusion This paper has presented a mobile enabled plasmonic ELISA based TB antigen-specific antibodies detection scheme using smartphones with the integration of machine learning techniques. Using a robust image processing technique comprised of clustering and object detection, our system can detect samples 22 (wells) without any guide or virtual plate. The decision components facilitated selection of the right cluster among the multiple number of clusters, detection of wells and transcending the samples from noise. Therefore, unlike the reported articles, the system does not require the user to provide seed points or perform cropping. Moreover, the system is capable of reading multiple samples and classifying them as positive or negative in real time. The plasmonic ELISA based technique produce colours for positive and negative samples. However, making of a final decision based on the colour appearance is not accurate in all the cases. Therefore, in this work, we demonstrated a smartphone-based POC platform that takes the final decision based on colour analysis. This work incorporated supervised machine learning to free the TB test result from the colour perception of individuals and its subjectivity of interpretation. Utilising 18 histogram features, we achieved 98.9% accuracy with the Random Forest classifier. This fully automated and self-contained system with image capturing, analysing and classification service is then embedded into the Android system. Using a completely new dataset, we demonstrated 98.4% accuracy to diagnose TB-positive samples on the mobile platform. In the absence of any existing automatic platform without an opto- mechanical attachment, to the best of our knowledge, it is the best performance for TB diagnosis on the mobile platform. The portability, technical and financial feasibility in automatic TB diagnosis in the presented system can benefit millions of people, especially in remote locations where few experts are available. This technique can be applied to other colourimetric qualitative tests, especially for ELISA and paper-based assays. Moreover, this system can be a guide in providing properly distinguishable colour by minimising the complexity of a chemical method utilising a powerful algorithm, which will reduce the dependency on perfect colour vision for naked-eye evaluation. The scheme shows great potential in evolving healthcare applications to benefit wider communities. A polythetic approach and subsequently a clinical trial will be executed in future to enhance the expert system with better precision and reliability. Acknowledgement This research is supported by British Council Newton Institutional Links and Newton-Ungku Omar Fund (Grant ID: 216385726). This is a collaborative research project between Anglia Ruskin University (UK) and Universiti Putra Malaysia (Malaysia). References Abuhassan, K. J., Bakhori, N. M., Kusnin, N., Azmi, U. Z. M., Tania, M. H., Evans, B. A., … Hossain, M. A. (2017). Automatic Diagnosis of Tuberculosis Disease Based on Plasmonic ELISA and Color-based Image Classification. In 2017 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) (pp. 4512–4515). Jeju Island, South Korea. https://doi.org/10.1109/EMBC.2017.8037859 Alidans srl. (2015). AssayColor. Retrieved January 10, 2017, from https://play.google.com/store/apps/details?id=com.alidans.assaycolor Android Developers. (2018). Overview of memory management. Retrieved May 29, 2018, from https://developer.android.com/topic/performance/memory-overview Bourouis, A., Feham, M., Hossain, M. A., & Zhang, L. (2014). An intelligent mobile based decision support system for retinal disease diagnosis. Decision Support Systems, 59(November 2015), 341–350. https://doi.org/10.1016/j.dss.2014.01.005 23 Centers for Disease Control and Prevention. (n.d.). Tuberculosis (TB) | CDC. Retrieved September 18, 2017, from https://www.cdc.gov/tb/ Cohen, J. (1960). A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1), 37–46. https://doi.org/10.1177/001316446002000104 Department of Economic and Social Affairs. (2016). International Migration Report 2015. https://doi.org/ST/ESA/SER.A/384 Ehringer, D. (2010). The Dalvik virtual machine architecture. Retrieved from http://show.docjava.com/posterous/file/2012/12/10222640-The_Dalvik_Virtual_Machine.pdf Enzo Life Sciences inc. (2015). Enzo ELISA Plate Reader. Retrieved September 21, 2017, from https://play.google.com/store/apps/details?id=com.enzo.elisaplatereader Estabrooks, A., Jo, T., & Japkowicz, N. (2004). A Multiple Resampling Method for Learning from Imbalanced Data Sets. Computational Intelligence, 20(1), 18–36. https://doi.org/10.1111/j.0824-7935.2004.t01-1-00228.x Farid, D. M., Zhang, L., Rahman, C. M., Hossain, M. A., & Strachan, R. (2014). Hybrid decision tree and naïve Bayes classifiers for multi-class classification tasks. Expert Systems with Applications, 41(4 PART 2), 1937–1946. https://doi.org/10.1016/j.eswa.2013.08.089 Forgy, E. W. (1965). Cluster analysis of multivariate data: efficiency versus interpretability of classifications. Biometrics, 21, 768–769. GSMA Intelligence. (n.d.). Definitive data and analysis for the mobile industry. Retrieved August 7, 2017, from https://www.gsmaintelligence.com/ Interactive Health Solutions. (2016). Global Fund TB. Retrieved September 21, 2017, from https://play.google.com/store/apps/details?id=com.ihsinformatics.tbr3mobile_sa&hl=en Interactive Health Solutions. (2016). MINE TB. Retrieved September 21, 2017, from https://play.google.com/store/apps/details?id=com.ihsinformatics.tbr4mobile Interactive Health Solutions. (2016). TB REACH 4 - Kotri. Retrieved September 21, 2017, from https://play.google.com/store/apps/details?id=com.ihsinformatics.tbr4mobile_pk&hl=en Interactive Health Solutions. (2017). Childhood TB-Bangladesh. Retrieved September 21, 2017, from https://play.google.com/store/apps/details?id=com.ihsinformatics.childhoodtb_mobile&hl=en Khademhosseini, A. (2011). Nano/microfluidics for diagnosis of infectious diseases in developing countries. Adv Drug Delivery Rev, 62(4–5), 449–457. https://doi.org/10.1016/j.addr.2009.11.016.Nano/microfluidics Kim, H., Awofeso, O., Choi, S., Jung, Y., & Bae, E. (2017). Colorimetric analysis of saliva--alcohol test strips by smartphone- based instruments using machine-learning algorithms. Appl. Opt., 56(1), 84–92. https://doi.org/10.1364/AO.56.000084 Liao, P.-S., Chen, T.-S., & Chung, P.-C. (2001). A Fast Algorithm for Multilevel Thresholding. Journal of Information Science and Engineering, 17, 713–727. Lloyd, S. (1982). Least squares quantization in PCM. IEEE Transactions on Information Theory, 28(2), 129–137. https://doi.org/10.1109/TIT.1982.1056489 Macqueen, J. (1967). Some methods for classification and analysis of multivariate observations. In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability (pp. 281–297). Berkeley, California: University of California Press. Retrieved from https://projecteuclid.org/euclid.bsmsp/1200512992%0A%0A Maggio, Emilio;Pan, Qi; Reitmayr, G. (2017). US9563824 B2. Retrieved from https://www.google.com/patents/US9563824 Meyer, F. (1994). Topographic distance and watershed lines. Signal Processing, 38(1), 113–125. https://doi.org/10.1016/0165- 1684(94)90060-4 Mutlu, A. Y., Kılıç, V., Özdemir, G. K., Bayram, A., Horzum, N., & Solmaz, M. E. (2017). Smartphone-based colorimetric detection via machine learning. The Analyst, 142(13), 2434–2441. https://doi.org/10.1039/C7AN00741H NHS. (n.d.). Tuberculosis (TB) - NHS Choices. Retrieved September 18, 2017, from http://www.nhs.uk/Conditions/Tuberculosis/Pages/Introduction.aspx Open Medicine Project. (2014). FIND TB. Retrieved September 21, 2017, from https://play.google.com/store/apps/details?id=tompsa.findtb&hl=en Operation Asha. (2017). eAlert Cambodia. Retrieved September 21, 2017, from https://play.google.com/store/apps/details?id=org.opasha.eCompliance.ecomplianceLabCambodia Osman, M. K., Mashor, M. Y., & Jaafar, H. (2010). Detection of mycobacterium tuberculosis in Ziehl-Neelsen stained tissue images using Zernike moments and hybrid multilayered perceptron network. Conference Proceedings - IEEE 24 International Conference on Systems, Man and Cybernetics, 4049–4055. https://doi.org/10.1109/ICSMC.2010.5642191 Otsu, N. (1979). A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics, 9(1), 62–66. https://doi.org/10.1109/TSMC.1979.4310076 Ozkan, H., & Kayhan, O. S. (2016). A Novel Automatic Rapid Diagnostic Test Reader Platform. Computational and Mathematical Methods in Medicine, 2016. Retrieved from http://dx.doi.org/10.1155/2016/7498217 Posey, D. L., Marano, N., & Cetron, M. S. (2017). Cross-border solutions needed to address tuberculosis in migrating populations. The International Journal of Tuberculosis and Lung Disease, 21(5), 485–486. https://doi.org/10.5588/ijtld.17.0187 Ren, & Malik. (2003). Learning a classification model for segmentation. In Proceedings Ninth IEEE International Conference on Computer Vision (pp. 10–17 vol.1). IEEE. https://doi.org/10.1109/ICCV.2003.1238308 Rousseeuw, P. J. (1987). Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20, 53–65. https://doi.org/10.1016/0377-0427(87)90125-7 Sergyan, S. (2008). Color Histogram Features Based Image Classification in Content-Based Image Retrieval Systems. In 2008 6th International Symposium on Applied Machine Intelligence and Informatics (pp. 221–224). Herlany. https://doi.org/10.1109/SAMI.2008.4469170 Sicasys Software GmbH. (2017). Spotxel® Reader. Retrieved January 12, 2018, from https://play.google.com/store/apps/details?id=com.sicasys.spotxel&hl=en Solmaz, M. E., Mutlu, A. Y., Alankus, G., Kılıç, V., Bayram, A., & Horzum, N. (2018). Quantifying colorimetric tests using a smartphone app based on machine learning classifiers. Sensors and Actuators B: Chemical, 255, 1967–1973. https://doi.org/10.1016/J.SNB.2017.08.220 Tania, M. H., Lwin, K. T., Abuhassan, K., & Bakhori, N. M. (2017). An Automated Colourimetric Test by Computational Chromaticity Analysis: A Case Study of Tuberculosis Test. In Advances in Intelligent Systems and Computing (Vol. 616, pp. 313–320). Springer, Cham. https://doi.org/10.1007/978-3-319-60816-7 Tracey, B. H., Comina, G., Larson, S., Bravard, M., López, J. W., & Gilman, R. H. (2011). Cough detection algorithm for monitoring patient recovery from pulmonary tuberculosis. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS, (day 0), 6017–6020. https://doi.org/10.1109/IEMBS.2011.6091487 Tsai, T.-T., Shen, S.-W., Cheng, C.-M., & Chen, C.-F. (2013). Paper-based tuberculosis diagnostic devices with colorimetric gold nanoparticles. Science and Technology of Advanced Materials, 14(4), 44404. https://doi.org/10.1088/1468- 6996/14/4/044404 UKVI. (n.d.). Tuberculosis tests for visa applicants. Retrieved September 24, 2017, from https://www.gov.uk/tb-test-visa Wang, S., Xu, F., & Demirci, U. (2010). Advances in developing HIV-1 viral load assays for resource-limited settings. Biotechnology Advances, 28(6), 770–781. https://doi.org/10.1016/j.biotechadv.2010.06.004 Wang, X.-Y., Wu, Z.-F., Chen, L., Zheng, H.-L., & Yang, H.-Y. (2016). Pixel classification based color image segmentation using quaternion exponent moments. Neural Networks : The Official Journal of the International Neural Network Society, 74, 1–13. https://doi.org/10.1016/j.neunet.2015.10.012 Wug Oh,Seoung; Kim, S. J. (2017). Approaching the computational color constancy as a classification problem through deep learning. Pattern Recognition, 61, 405–416. https://doi.org/10.1016/J.PATCOG.2016.08.013 Yetisen, A. K., Martinez-Hurtado, J. L., Garcia-Melendrez, A., da Cruz Vasconcellos, F., & Lowe, C. R. (2014). A smartphone algorithm with inter-phone repeatability for the analysis of colorimetric tests. Sensors and Actuators B: Chemical, 196, 156–160. https://doi.org/10.1016/j.snb.2014.01.077 
work_xf2a6wdpgbbs5gx2ibpnpn67i4	----	Feature selection using Joint Mutual Information Maximisation Expert Systems With Applications 42 (2015) 8520–8532 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Feature selection using Joint Mutual Information Maximisation Mohamed Bennasar, Yulia Hicks, Rossitza Setchi∗ School of Engineering, Cardiff University, Cardiff CF24 3AA, UK a r t i c l e i n f o Keywords: Feature selection Mutual information Joint mutual information Conditional mutual information Subset feature selection Classification Dimensionality reduction Feature selection stability a b s t r a c t Feature selection is used in many application areas relevant to expert and intelligent systems, such as data mining and machine learning, image processing, anomaly detection, bioinformatics and natural language processing. Feature selection based on information theory is a popular approach due its computational ef- ficiency, scalability in terms of the dataset dimensionality, and independence from the classifier. Common drawbacks of this approach are the lack of information about the interaction between the features and the classifier, and the selection of redundant and irrelevant features. The latter is due to the limitations of the employed goal functions leading to overestimation of the feature significance. To address this problem, this article introduces two new nonlinear feature selection methods, namely Joint Mutual Information Maximisation (JMIM) and Normalised Joint Mutual Information Maximisation (NJMIM); both these methods use mutual information and the ‘maximum of the minimum’ criterion, which alleviates the problem of overestimation of the feature significance as demonstrated both theoretically and experimentally. The proposed methods are compared using eleven publically available datasets with five competing methods. The results demonstrate that the JMIM method outperforms the other methods on most tested public datasets, reducing the relative average classification error by almost 6% in comparison to the next best performing method. The statistical significance of the results is confirmed by the ANOVA test. More- over, this method produces the best trade-off between accuracy and stability. © 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). ( L 2 s o N f p p d p i l s t f 1. Introduction High dimensional data is a significant problem in both super- vised and unsupervised learning (Janecek, Gansterer, Demel, & Ecker, 2008), which is becoming even more prominent with the recent ex- plosion of the size of the available datasets both in terms of the num- ber of data samples and the number of features in each sample (Zhang et al., 2015). The main motivation for reducing the dimensionality of the data and keeping the number of features as low as possible is to decrease the training time and enhance the classification accuracy of the algorithms (Guyon & Elisseeff, 2003; Jain, Duin, & Mao, 2000; Liu & Yu, 2005). Dimensionality reduction methods can be divided into two main groups: those based on feature extraction and those based on feature selection. Feature extraction methods transform existing features into a new feature space of lower dimensionality. During this process, new features are created based on linear or nonlinear combinations of features from the original set. Principal Component Analysis (PCA) ∗ Corresponding author. Tel: +44 2920875720; fax: +44 2920874716. E-mail addresses: BennasarM@cf.ac.uk (M. Bennasar), HicksYA@cf.ac.uk (Y. Hicks), Setchi@cf.ac.uk (R. Setchi). n & C f http://dx.doi.org/10.1016/j.eswa.2015.07.007 0957-4174/© 2015 The Authors. Published by Elsevier Ltd. This is an open access article unde Bajwa, Naweed, Asif, & Hyder, 2009; Turk & Pentland, 1991) and inear Discriminant Analysis (LDA) (Tang, Suganthana, Yao, & Qina, 005; Yu & Yang, 2001) are two examples of such algorithms. Feature election methods reduce the dimensionality by selecting a subset f features which minimises a certain cost function (Guyon, Gunn, ikravesh, & Zadeh, 2006; Jain et al., 2000). Unlike feature extraction, eature selection does not alter the data and, as a result, it is the referred choice when an understanding of the underlying physical rocess is required. Feature extraction may be preferred when only iscrimination is needed (Jain et al., 2000). Feature selection is used in many application areas relevant to ex- ert and intelligent systems, such as data mining and machine learn- ng, image processing, anomaly detection, bioinformatics and natural anguage processing (Hoque, Bhattacharyya, & Kalita, 2014). Feature election is normally used at the data pre-processing stage before raining a classifier. This process is also known as variable selection, eature reduction or variable subset selection. The topic of feature selection has been reviewed in detail in a umber of recent review articles (Bolón-Canedo, Sánchez-Maroño, Alonso-Betanzos, 2013; Brown, Pocock, Zhao, & Lujan, 2012; handrashekar & Sahin, 2014; Vergara & Estévez, 2014). Usually, eature selection methods are divided into two categories in terms of r the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). http://dx.doi.org/10.1016/j.eswa.2015.07.007 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.07.007&domain=pdf http://creativecommons.org/licenses/by-nc-nd/4.0/ mailto:BennasarM@cf.ac.uk mailto:HicksYA@cf.ac.uk mailto:Setchi@cf.ac.uk http://dx.doi.org/10.1016/j.eswa.2015.07.007 http://creativecommons.org/licenses/by-nc-nd/4.0/ M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 8521 e ‘ W s o o c T c s a t o h P w o t c r s c c fi “ S M R D X t a A a s o f a m r a c S i o c s o o s w i o n p m c r I s c l t e m c p p F m p S r t c 2 c f a s a w a H w X T c i H w a H T a b o e H H M a I M a r I valuation strategy, in particular, classifier dependent (‘wrapper’ and embedded’ methods) or classifier independent (‘filter’ methods). rapper methods search the feature space, and test all possible ubsets of feature combinations by using the prediction accuracy f a classifier as a measure of the selected subset’s quality, with- ut modifying the learning function. Therefore, wrapper methods an be combined with any learning machine (Guyon et al., 2006). hey perform well because the selected subset is optimised for the lassification algorithm. On the other hand, wrapper methods may uffer from over-fitting to the learning algorithm. This means that ny changes in the learning model may reduce the usefulness of he subset. In addition, these methods are very expensive in terms f computational complexity, especially when handling extremely igh-dimensional data (Brown et al., 2012; Cheng et al., 2011; Ding & eng, 2003; Karegowda, Jayaram, & Manjunath, 2010). The feature selection stage in the embedded methods is combined ith the learning stage. These methods are less expensive in terms f computational complexity and less prone to over-fitting; however, hey are limited in terms of generalisation, because they are very spe- ific to the used learning algorithm (Guyon et al., 2006). Classifier-independent methods rank features according to their elevance to the class label in the supervised learning. The relevance core is calculated using distance, information, correlation and onsistency measures. Many techniques have been proposed to ompute the relevance score, including Pearson correlation coef- cients (Rodgers & Nicewander, 1988), Fisher’s discriminate ratio F score” (Lin, Li, & Tsai, 2004), the Scatter criterion (Duda, Hart, & tork, 2001), Single Variable Classifier SVC (Guyon & Elisseeff, 2003), utual Information (Battiti, 1994), the Relief Algorithm (Kira & endell, 1992; Liu & Motoda, 2008), Rough Set Theory (Liang, Wang, ang, & Qian, 2014) and Data Envelopment Analysis (Zhang, Yang, iong, Wang, & Zhang, 2014). The main advantages of the filter methods are their computa- ional efficiency, scalability in terms of the dataset dimensionality, nd independence from the classifier (Saeys, Inza, & Larranaga, 2007). common drawback of these methods is the lack of information bout the interaction between the features and the classifier and election of redundant and irrelevant features due to the limitations f the employed goal functions leading to overestimation of the eature significance. Information theory (Cover & Thomas, 2006) has been widely pplied in filter methods, where information measures such as utual information (MI) are used as a measure of the features’ elevance and redundancy (Battiti, 1994). MI does not make an ssumption of linearity between the variables, and can deal with ategorical and numerical data with two or more class values (Meyer, chretter, & Bontempi, 2008). There are several alternative measures n information theory that can be used to compute the relevance f features, namely mutual information, interaction information, onditional mutual information, and joint mutual information. This paper contributes to the knowledge in the area of feature election by proposing two new nonlinear feature selection meth- ds based on information theory. The proposed methods aim to vercome the limitations of the current state of the art filter feature election methods such as overestimation of the feature significance, hich causes selection of redundant and irrelevant features. This s achieved through the introduction of a new goal function based n joint mutual information and the ‘maximum of the minimum’ onlinear approach. As shown in the evaluation section, one of the roposed methods outperforms the competing feature selection ethods in terms of classification accuracy, decreasing the average lassification error by 0.88% in absolute terms and almost by 6% in elative terms in comparison to the next best performing method. n addition, it produces the best trade-off between accuracy and tability. The statistical significance of the reported results is further onfirmed by ANOVA test. This paper also reviews existing feature selection methods high- ighting their common limitations and compares the performance of he proposed and existing methods on the basis of several criteria. For xample, a nonlinear approach, which employs the ‘maximum of the inimum’ criterion, is compared to a linear approach, which employs umulative summation approximation. To optimise the nonlinear ap- roach, a goal function based on joint mutual information is com- ared to the goal function based on conditional mutual information. inally, the effect of using normalised mutual information instead of utual information is tested. The rest of the paper is organised as follows. Section 2 presents the rinciples of the information theory, Section 3 reviews related work, ection 4 discusses the limitations of current feature selection crite- ia, Section 5 introduces the proposed methods. Section 6 describes he conducted experiments and discusses the results. Section 7 con- ludes the paper. . Information theory This section introduces the principles of information theory by fo- using on entropy and mutual information and explains the reasons or employing them in feature selection. The entropy of a random variable is a measure of its uncertainty nd a measure of the average amount of information required to de- cribe the random variable (Cover & Thomas, 2006). The entropy of discrete random variable X = (x1, x2, . . . . . . , xN) is denoted by H(X), here xi refers to the possible values that X can take. H(X) is defined s: (X) = − N∑ i=1 p(xi)log(p(xi)), (1) here p(xi) is the probability mass function. The value of p(xi), when is discrete, is: p(xi) = number o f instants with value xi total number o f instants (N) . (2) he base of the logarithm, log, is 2, so 0 ≤ H(X) ≤ 1. For any two dis- rete random variables X and C = (c1, c2, . . . . . . , cM), the joint entropy s defined as: (X, C) = − M∑ j=1 N∑ i=1 p ( xi, c j ) log ( p ( xi, c j )) (3) here p(xi, cj) is the joint probability mass function of the variables X nd C. The conditional entropy of the variable X given C is defined as: (C|X) = − M∑ j=1 N∑ i=1 p ( xi, c j ) log ( p ( c j|xi )) (4) he conditional entropy is the amount of uncertainty left in C when variable X is introduced, so it is less than or equal to the entropy of oth variables. The conditional entropy is equal to the entropy if, and nly if, the two variables are independent. The relation between joint ntropy and conditional entropy is: (X, C) = H(X) + H(C|X) (5) (X, C) = H(C) + H(X|C) (6) utual Information (MI) is the amount of information that both vari- bles share, and is defined as: (X ; C) = H(C) − H(C|X) (7) I can be expressed as the amount of information provided by vari- ble X, which reduces the uncertainty of variable C. MI is zero if the andom variables are statistically independent. MI is symmetric, so: (X ; C) = I(C; X) (8) 8522 M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 a a p M v D d t p t a v m d t o c t t g t m M b s j c & c J t t s t ( r t r a p r ( i I A k t p l o l t c d t n m I(X ; C) = H(X) − H(X|C) (9) I(X ; C) = H(X) + H(C) − H(X, C) (10) The Joint MI is defined as: I(X ; C|Y ) = H(X|C) − H(X|C, Y ) (11) I(X, Y ; C) = I(X ; C|Y ) + I(Y ; C) (12) where Y is a discrete variable; Y = (y1, y2, . . . . . . , yN). Interaction in- formation can be defined as the amount of information that is shared by all features, but is not found within any feature subset. Mathemat- ically, the relation between interaction information and MI is defined as: I(X ; Y ; C) = I(X, Y ; C) − I(X ; C) − I(Y ; C) (13) High interaction information means that a large amount of infor- mation can be obtained by considering the three variables together (Jakulin, 2003). Interaction information can be positive, negative or zero (Jakulin, 2005). 3. Related work The focus of the work presented in this article is on the filter feature selection methods due to their popularity, and thus the review part of this article focuses specifically on these methods. For a more detailed review of the feature selection methods recent review articles in this area are recommended (Bolón-Canedo et al., 2013; Brown et al., 2012; Chandrashekar & Sahin, 2014; Vergara & Estévez, 2014). Information theory has been employed by many filter feature selection methods. Information Gain (IG) (Guyon & Elisseeff, 2003) is the simplest of these methods. It is classified as a univariate feature selection method, as it ranks features based on the value of their mutual information with the class label. Simplicity and low computational costs are the main advantages of this method. How- ever, it does not take into consideration the dependency between the features, rather, it assumes independency, which is not always the case. Therefore some of the selected features may carry redundant information. To tackle this problem new methods have been pro- posed for selecting relevant features, which are non-redundant with respect to each other. For a feature set F = { f1, f2, . . . . . . , fN}, the feature selection pro- cess identifies a subset of features S with dimension k where k ≤ N, and S⊆F. In theory, the selected subset S should maximise the joint mutual information between the class label C and the subset S of a fixed size k. I(S; C) = I( f1, f2, . . . . . . , fk; C) (14) However, such an approach is impractical, due to the number of cal- culations and the limited number of observations available for the calculation of the high-dimensional probability density function. As a result, many methods use heuristic approaches to approximate the ideal solution. Generally, the filter criteria are based on the concepts of fea- ture relevance, redundancy and complementarity (Vergara & Estévez, 2014). The methods which are based on information theory can be split into two groups: linear criteria, which are linear combinations of MI terms; and nonlinear criteria, which use maximum or mini- mum operations or normalised MI in their goal functions (Brown et al., 2012). Battiti (1994) introduces a first-order incremental search algo- rithm, known as the Mutual Information Feature Selection (MIFS) method, for selecting the most relevant k features from an initial set of n features. A greedy selection method is used to build the subset. Instead of calculating the joint MI between the selected features and the class label, Battiti studies the MI between the candidate feature nd the class, and the relationship between the candidate and the lready-selected features. Kwok and Choi (2002) propose the MIFS-U method to improve the erformance of the MIFS method by making a better estimation of the I between the input feature and the class label. Another method ariant to MIFS, the mRMR method is proposed by Peng, Long, and ing (2005). The redundancy term in mRMR is divided over the car- inality |S| of the selected subset S to balance the magnitude of this erm, and to avoid it growing very large as the subsets expand. As re- orted in the existing literature (Brown et al., 2012; Peng et al., 2005), his modification allows mRMR to outperform the conventional MIFS nd MIFS-U methods. Estévez, Tesmer, Perez, and Zurada (2009) propose an enhanced ersion of MIFS, MIFS-U and mRMR, called Normalised Mutual Infor- ation Feature Selection (NMIFS). It uses normalised MI in the re- undancy term instead of MI. The normalisation of MI prevents bias owards multivalued features and limits the value of MI to the range f zero to unity (Estévez et al., 2009). Hoque et al. (2014) propose a method called MIFS-ND. The method alculates the mutual information between the candidate feature and he class label, and the average of the mutual information between he candidate feature and the features within the selected subset. A enetic algorithm is employed to select the feature that maximises he mutual information with the class, and minimises the average utual information with the other selected features. Other proposed criteria (Yang & Moody, 1999; Fleuret, 2004; eyer & Bontempi, 2006; Vidal-Naquet & Ullman, 2003) use the MI etween the candidate feature and the class label in the context of the elected subset features. They utilise conditional mutual information, oint mutual information or feature interaction. Some of them apply umulative summation approximations (Yang & Moody, 1999; Meyer Bontempi, 2006), while others use the ‘maximum of the minimum’ riterion (Fleuret, 2004; Vidal-Naquet & Ullman, 2003). Yang and Moody (1999) propose a feature selection method called oint Mutual Information (JMI). In this method, the candidate feature hat maximises the cumulative summation of Joint Mutual Informa- ion with features of the selected subset is chosen and added to the ubset. This method is reported to perform well in terms of classifica- ion accuracy and stability (Brown et al., 2012). Meyer and Bontempi 2006) introduce a similar method known as Double Input Symmet- ical Relevance (DISR). The joint mutual information in the goal func- ion of this method is substituted with symmetrical relevance. Other methods that employ the ‘maximum of the minimum’ crite- ion have been proposed. Vidal-Naquet and Ullman (2003) introduce method called Information Fragment (IF), while Fleuret (2004) pro- ose Conditional Mutual Information Maximisation, which have been eported to perform well with KNN and SVM classifiers in later work Freeman, Kulić, & Basir, 2015). There are also a number of other methods which rely on max- mising Feature Interaction. For example, Jakulin (2005) proposes the nteraction Capping (IC) method, while El Akadi, El Ouardighi, and boutajdine (2008) propose a method which uses feature interaction, nown as Interaction Gain Based Feature Selection (IGFS). However, his is typically the same as JMI. General formula based on conditional likelihood has been pro- osed by Brown et al. (2012) based on a study of MI-based feature se- ection criterion, this formula can be used to derive many of the meth- ds listed in this section. In practice, most of the methods which are inear combinations of MI can be derived from this formula. However, he authors stated that the goal function of the nonlinear method annot be generated by their formula. Feature selection techniques have also been used for multi-label ata sets. Lee and Kim (2015) proposed a multi-label feature selec- ion method based on information theory, in which they introduce a ew score function to measure the importance of each feature to the ultiple labels. M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 8523 t a O i ( f p t d m t 4 t r t b n t m e i t m l t o o s l E a f c t I t s p 5 T m d t e s w d F m v E i D c s D s a o t S L o v P a m t C a m i D t t L t c P t p t t I t H 5 r t r m r l n i t I s t t M j p L a w I Two other notable approaches in the area of filter feature selec- ion are the application of the rough set theory (Liang et al., 2014) nd the application of Data Envelopment Analysis (Zhang et al., 2014). ne of the issues affecting the methods based on the fuzzy-rough sets s their time inefficiency, with many existing attempts to improve it Qian, Wang, Cheng, Liang, & Dang, 2015). The methods using DEA or feature selection also suffer from the problem of the large com- utational cost, although it was improved in a more recent publica- ion (Zhang et al., 2015), as well as the problem of the selection of re- undant features. The latter problem is characteristic of most of the ethods listed above and the reasons for this problem will be inves- igated in more detail in Section 4. . Limitations of the current feature selection criteria In general, most of the methods listed in the previous section use he criteria consisting of two elements: the relevancy term and the edundancy term. The methods attempt to simultaneously maximise he relevancy term whilst minimising the redundancy term. It has een noted in literature that such feature selection methods have a umber of limitations (Estévez et al., 2009; Peng et al., 2005). For example, MIFS and MIFS-U share a common problem: when he number of selected features grows, the redundancy term grows in agnitude with respect to the relevancy term. In this case some irrel- vant features may be selected. This problem has been partly solved n the mRMR, NMIFS, MIFS-ND methods by dividing the redundancy erm over the cardinality of the subset. Another problem shared by all above methods (MIFS, MIFS-U, RMR, NMIFS, and MIFS-ND) is that the redundancy term is calcu- ated based on the value of the MI between the candidate feature and he features within the selected subset, without any consideration f the class label. The features may share information between each ther, but that does not mean they are redundant; they may in fact hare different information with the class. Yet another problem particular to the methods employing cumu- ative summation and forward search to approximate the solution of q. (14) (such as MIFS, NMIFS, mRMR, NMIFS, MIFS-ND, DISR, IGFS, nd JMI) is the overestimation of the significance of some candidate eatures. For example, this can occur when the candidate feature is in omplete correlation with one or several pre-selected features, but at he same time is almost independent from the majority of the subset. n such situation, the value of the goal function will be high despite he redundancy of the candidate feature to some features within the ubset. In practice, the significance of each of the above problems de- ends on the data and the characteristics of each particular data set. . Proposed methods for feature selection In this paper, two new methods for feature selection are proposed. he methods employ joint mutual information, and use the ‘maxi- um of the minimum’ approach. The proposed methods aim to ad- ress the problem of overestimation the significance of some fea- ures, which occurs when cumulative summation approximation is mployed. For a feature set F = { f1, f2, . . . . . . , fN} of a data set D of dimen- ion N, the feature selection process identifies a subset of features S ith dimension K where K ≤ N, and S⊆F. The subset S should pro- uce equal or better classification accuracy compared to feature set . In other words feature selection defines the subset of features that aximises mutual information with the class label I(S, C). In the past, a number of alternative definitions of feature rele- ance have been used (Battiti, 1994; Brown et al., 2012; Vergara & stévez, 2014; Estévez et al., 2009). The following definition is used n this work. efinition 1. (Feature relevance). Feature fi is more relevant to the lass label C than feature fj in the context of the already selected sub- et S when I(fi, S; C) > I(fj, S; C). efinition 2. (Minimum joint mutual information): Let F be the full et of features, and let S be the subset of features that are selected lready. Let fi ∈ F − S, and fs ∈ S. The m-Joint MI is the minimum value f joint mutual information that the candidate feature fi shares with he class label C when it is joined with every feature within the subset individually, hence min s=1,2,...,k I( fi, fs; C), emma 1. For a feature fi, if the m-Joint MI is larger than that of all ther features fj, where fi and f j ∈ F − S (i �= j), then it is the most rele- ant feature to the class label C in the context of the subset S. roof. Let S = { f1, f2, . . . . . . , fK }. The joint mutual information of fi nd each feature in S with C is calculated. The minimum value of this utual information (m-Joint) is the lowest amount of new informa- ion that the feature fi adds to the shared information between S and . The feature that produces the maximum m-Joint is the feature that dds maximum information to that shared between S and C, which eans it is the feature which is the most relevant to the class label C n the context of the subset S according to Definition 1. efinition 3. Candidate feature fi is redundant to the selected fea- ures within the subset S if fi does not share new information with he class C. emma 2. Let F be the full set of features, let S be the subset of features hat are selected already, and fi ∈ F − S, fs ∈ S. If the feature fi is highly orrelated with a feature fs in the subset then I(fi; C) ∼=I(fs; C)∼=I(fi, fs; C). roof. If the feature fi is highly correlated with a featurefs, then he probability mass functions of fi, fs, and (fi, fs) are equal, (fi) ∼=p(fs)∼=p(fs, fi) . Since the definition of the entropy is (X) = − ∑Ni=1 p(xi)log(p(xi)) hen H(fi) ∼=H(fs)∼=H(fs, fi). Since the definition of the mutual informa- ion is I(X ; C) = H(X) + H(C) − H(X, C) then I(fi; fs)∼=H(fs)∼=H(fi) and (fi; C) ∼=I(fs; C). I( fi, fs; C) = H( fi, fs) + H(C) − H( fi, fs, C), according o the definition, which can be simplified to: I( fi, fs; C) = H( fi) + (C) − H( fi, C). According to Eq. (10) I(fi, fs; C)∼=I(fi; C)∼= I(fs; C). .1. Joint Mutual Information Maximisation (JMIM) All methods listed in the previous section attempt to optimise the elationship between relevancy and redundancy when selecting fea- ures by approximating the solution of Eq. (14). The JMI method is eported in existing literature as being the method which selects the ost relevant features (Brown et al., 2012). It studies relevancy and edundancy, and takes into consideration the class label when calcu- ating MI. However, the method still allows overestimation of the sig- ificance of some features, for example, when the candidate feature is n complete correlation with one or a few pre-selected features, but at he same time is almost independent from the majority of the subset. n such a situation, the value of the JMI goal function will be high de- pite the redundancy of the candidate feature to some features within he subset. This drawback is evident in almost all methods that use he cumulative sum approximation. For this reason, a new method called Joint Mutual Information aximisation (JMIM) is proposed in this research. JMIM employs oint mutual information and the ‘maximum of the minimum’ ap- roach, which should choose the most relevant features according to emma 1, following from which, the features are selected by JMIM ccording to the following new criterion: fJMIM = arg max fi ∈F −S(min fs ∈S(I( fi, fs; C))), (22) here ( fi, fs; C) = I( fs; C) + I( fi; C/ fs), (23) 8524 M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 5 w 5 a I t s F w S W F T t 6 a m f l e t w t o t i s w d g e ( s N t o I( fi, fs; C) = H(C) − H(C/ fi, fs), (24) I( fi, fs; C) = [ − ∑ c∈C p(c) log (p(c)) ] − [∑ c∈C ∑ fi ∈F −S ∑ fs ∈S log ( p( fi fs, c/ fs) p( fi/ fs)p(c/ fs) )] . (25) The method uses the following iterative forward greedy search al- gorithm to find the relevant feature subset of size k within the feature space: Algorithm 1. Forward greedy search. 1. (Initialisation) Set F ← “initial set of n features”; S ← “empty set.” 2. (Computation of the MI with the output class) For ∀ fi ∈ F compute I(C; fi). 3. (Choice of the first feature) Find a feature fi that maximises I(C; fi); set F ← F \{ fi}; set S ← { fi}. 4. (Greedy selection) Repeat until |S| = k: (Selection of the next feature) Choose the feature fi = arg max fi ∈F −S(min fs ∈S(I( fi , fs ; C))); set F ← F \ { fi}; set S ← S ∪{ fi}. 5. (Output) Output the set S with the selected features. 5.2. Advantages over existing alternative methods The Venn diagrams in Fig. 1 show different scenarios for the re- lationship between the candidate feature fi, the selected feature fs, and the class label C. Fig. 1a illustrates the case in which methods like MIFS, NMIFS or mRMR will fail to select fi because it is redundant to fs, although each of them shares different information about C, and the correlation is not in the context of C. The goal function of JMIM is similar to the goal function of CMIM (Section 3), as CMIM also uses the ‘maximum of the mini- mum’ approach. The main difference is that CMIM maximises the amount of information the candidate feature fi contributes given the pre-selected feature fs (i.e. fi is selected for any complementing fs), whereas JMIM selects the feature that maximises the joint mutual information with fs. Fig. 1b and c is used to explain this difference further. The figures represent two candidate features fi and fj, and the subsequent selection of one of them. I(fi, fs; C) is the union of areas 1, 2, and 3; I(fi; C/fs) is area 1 in Fig. 1b. The CMIM method would se- lect fi in Fig. 1b, even though its complementing feature fs from the subset does not carry as much information as the feature fj in Fig. 1c. Conversely, JMIM would select the feature that maximises JMI, so it would select feature fi in Fig. 1c. Therefore, the joint mutual infor- mation between the candidate feature and at least one of the pre- selected features will be high, which can increase the discrimination power of the selected subset. Fig. 1. Venn diagrams illustrating the re .3. Normalised Joint Mutual Information Maximisation (NJMIM) The second method proposed in this paper uses a goal function, hich is very similar to the one used in JMIM proposed in Section .1, with the difference being that symmetrical relevance is used as n alternative to MI. This method is called Normalised Joint Mutual nformation Maximisation (NJMIM). It is proposed in order to study he effect of using normalised MI instead of MI. the proposed NJMIM election criteria is presented in Eq. (26). NJMIM = arg max fi ∈F −S ( min fs ∈S(SR( fi, fs; C)) ) , (26) here ymmetrical relevance = SR(F ; C) = I(F ; C) H(F, C) . (27) hich can be simplified as: NJMIM = arg max fi ∈F −S ( min fs ∈S ( I( fi, fs; C) H( fi, fs, C) )) . (28) he same iterative forward greedy search algorithm is used to find he subset of features within the candidate feature space. . Evaluation The performance of the two proposed methods in this paper, JMIM nd NJMIM, is compared with the results produced by five other ethods: CMIM, DISR, mRMR, JMI, and IG. These methods are chosen or the following four reasons: (i) these methods are reported in the iterature to provide good performance (Brown et al., 2012; Freeman t al., 2015); (ii) the choice of these methods allows the comparison of he ‘maximum of the minimum’ approach used by JMIM and NJMIM ith the cumulative summation used by JMI and DISR; (iii) it enables he analysis of the effect of using the symmetrical relevance instead f MI on the algorithm’s performance; (iv) it allows the comparison of he effects of using joint mutual information and conditional mutual nformation, which are employed in JMIM and CMIM, respectively. The seven methods are applied to data from different domains uch as: life sciences, physical sciences, engineering, business, hand- riting recognition, and gene microarray. The features within these atasets have different characteristics, being binary, discrete or cate- orical, or continuous. The continuous features are discretised into 10 qual intervals, using the Equal Width Discretisation (EWD) method Dougherty, Kohavi, & Sahami, 1995). Two classifiers are used to evaluate the quality of the selected sub- ets. These are Naïve-Bayes with kernel density estimation, and 3- earest Neighbours. Both classifiers are available in the Matlab Statis- ics Toolbox. The average classification accuracy is used as a measure f the quality of the selected features. Five-fold cross-validation is lation between features and class. M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 8525 Fig. 2. Evaluation framework. e t t 2 d d s f T s s 6 a v 2 o t r i n t T s a Table 1 UCI datasets used in the experiment. No Data set Number of Number of Number of Ratio features instances classes 1 Credit approval 15 690 2 54 2 Gas sensor 128 13874 6 198 3 Libra movement 90 483 15 3 4 Parkinson 22 195 2 11 5 Breast 30 569 2 28 6 Sonar 60 208 2 10 7 Musk 166 7074 2 354 8 Handwriting 649 2000 10 20 Table 2 Additional datasets used in the experiment (Peng et al., 2005). No Data set Number of Number of Number of Ratio features instances classes 1 Colon 2000 62 2 10 2 Leukemia 7070 72 2 12 3 Lymphoma 4026 96 9 4 6 d b l d mployed when processing feature selection and feature validation; herefore each fold is used for validation once. This means that 80% of he data is used for feature selection and classification training, whilst 0% is used for validation. This is repeated five times, using the whole ataset for validation over the course of five experiments. Overall, five ifferent subsets of samples are used to generate five different sub- ets of features. Discretisation is performed as a pre-processing step or all data prior to the feature selection step. Fig. 2 shows the evaluation framework used in this experiment. o test the impact of adding each feature to the subset on the clas- ification accuracy, training and validation are performed after the election of each feature in the subset. .1. Data Eight datasets from the UCI Repository (Bache & Lichman, 2013) re used in the experiment (Table 1). These datasets have been pre- iously used in similar research (Brown et al., 2012; El Akadi et al., 008; Cheng et al., 2011). They have different characteristics in terms f number of classes, features, instances and feature types. An example-feature ratio (Brown et al., 2012) is used as an indica- ion of the difficulty of the feature selection task for the dataset. This atio is computed using N mC , where N is the number of instances, m s the median number of values that the features have, and C is the umber of classes. The most challenging feature selection tasks are hose performed using datasets with a small example-feature ratio. he libra movement dataset is the most challenging dataset. To test the behaviour of the methods with an extremely small ample, datasets from Peng et al. (2005) are also used in the evalu- tion process, and these are shown in Table 2. .2. Performance analysis on low dimensional datasets Figs. 3–5 show the average classification accuracy of the three atasets with low numbers of features (Parkinson, credit approval and reast). The classification is computed over the whole size of the se- ected subset, from 1 feature up to 20 features (or all features of the ataset in the case of the credit approval dataset). 8526 M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 0 2 4 6 8 10 12 14 16 18 20 0.82 0.83 0.84 0.85 0.86 0.87 0.88 0.89 0.9 0.91 C la ss ifi ca tio n a cc u ra cy CMIM NJMIM DISR JMIM mRMR JMI IG Number of features Fig. 3. Average classification accuracy achieved with the Parkinson dataset. 0 5 10 15 0.65 0.7 0.75 0.8 0.85 Number of features CMIM NJMIM DISR JMIM mRMR JMI IG C la ss ifi ca tio n a cc u ra cy Fig. 4. Average classification accuracy achieved with the credit approval dataset. 0 2 4 6 8 10 12 14 16 18 20 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 Number of features C la ss ifi ca tio n a cc u ra cy CMIM NJMIM DISR JMIM mRMR JMI IG Fig. 5. Average classification accuracy achieved with the breast dataset. M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 8527 d j a s u t h c t o f J t w h N p d p r b a c 6 s g w s t a I f m p a As shown in Fig. 3, which illustrates the experiment with the first ataset, JMIM achieves the highest average accuracy (90.77%) with ust 8 features, which is higher than the accuracy of CMIM (90.26%) nd JMI (88.97%). On the other hand, methods that use normalised MI, uch as NJMIM and DISR, perform less well than JMIM and JMI, which se MI. This is expected for datasets with discrete features, because he normalisation may reduce the significance of the feature when it as high entropy and shares a high amount of information with the lass label. The mRMR and IG methods perform poorly on this dataset. JMIM and JMI again achieve the highest classification accuracy on he credit approval dataset, using only 4 features to reach an accuracy f 82.92%. The accuracy of CMIM is 79.17% with the same number of eatures. The other methods perform worse compared to JMIM and MI with the same number of features. The figure also shows that he methods using normalised MI do not perform as well as those hich use MI. Features selected by the JMIM and JMI methods have a igher discriminative power than the features which are selected by JMIM and DISR. NJMIM performs better than DISR, yet both perform oorly. The breast dataset has 20 features selected. As seen in Fig. 5, JMIM oes not achieve the highest classification accuracy. However, it 0 5 10 15 20 2 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Number of C la ss ifi ca tio n a cc u ra cy Fig. 6. Average classification accuracy ac 0 5 10 15 20 2 0.65 0.7 0.75 0.8 0.85 Number of f C la ss ifi ca tio n a cc u ra cy Fig. 7. Average classification accuracy roduces a high accuracy (95.87%) with only 5 features, while mRMR equires 14 features to achieve the same accuracy. JMIM performs etter in comparison with JMI and CMIM. The performance of NJMIM nd DISR is not as good as JMIM and JMI, as with 4 features their lassification accuracies are 87.61% and 89.28%, respectively. .3. Performance analysis on high dimensional datasets The second experiment involves high dimensional data (musk, onar, gas sensor, and handwriting datasets. The experiment with the as sensor and sonar datasets includes the selection of 50 features, ith JMIM achieving high classification accuracy with a relatively mall number of features. The other methods require more features o achieve this level of accuracy (Figs. 6–7). Fig. 8 shows the results for the handwriting dataset. 50 features re selected. JMIM performs well, but is inferior to JMI and mRMR. n terms of classification accuracy of the selected subset JMI per- ormed better than JMIM in the subset with 11–21 features, by a aximum difference in accuracy of 0.5%. The mRMR method also erforms well with this dataset; however JMIM produces the highest ccuracy (97.68%) with the selected subset of 33 features. 5 30 35 40 45 50 features CMIM NJMIM DISR JMIM mRMR JMI IG hieved with the gas sensor dataset. 5 30 35 40 45 50 eatures CMIM NJMIM DISR JMIM mRMR JMI IG achieved with the sonar dataset. 8528 M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 0 10 20 30 40 50 60 70 80 90 100 0.4 0.5 0.6 0.7 0.8 0.9 1 Number of features y c ar u c c a n oit a cifi s s al C CMIM NJMIM DISR FIM JMIM mRMR JMI IG 32 34 36 38 40 42 44 0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 Fig. 8. Average classification accuracy achieved with the handwriting dataset. 0 5 10 15 20 25 30 35 40 45 50 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Number of features C la ss ifi ca tio n a cc u ra cy CMIM NJMIM DISR FIM JMIM mRMR JMI IG 14 16 18 20 22 24 0.65 0.7 0.75 0.8 Fig. 9. Average classification accuracy achieved with the libra movement dataset. m i i o b o m w b b 6 t t e A The experimental results using the libra movement dataset are shown in Fig. 9, when 50 features are selected. JMIM is the best method with this dataset with almost any number of selected fea- tures, followed by NJMIM. JMIM outperforms JMI by up to 3% in terms of classification accuracy. NJMIM also outperforms DISR for all of the selected subsets. The methods are also applied to the musk dataset. Fig. 10 shows the result when 50 features are selected. With this dataset, JMIM se- lects the best subset and outperforms the other methods in terms of classification accuracy. NJMIM does not perform as well as JMIM, but produces better accuracy than DISR and mRMR for most of the fea- tures selected. 6.4. Performance analysis with Peng et al. (2005) datasets The results using the three datasets employed by Peng et al. (2005) are shown in Fig. 11. The leukemia dataset (Fig. 11a) has a small num- ber of samples. The results show that none of the feature selection ethods perform particularly well, confirming the findings reported n the review article by Brown et al. (2012). The colon dataset, which s the least challenging dataset of the three in terms of the number f classes and features, is shown in Fig. 11b. The results indicate the etter performance of JMIM and JMI compared to the other meth- ds, especially CMIM, which performs poorly. However, CMIM is the ethod that provides the best accuracy with the lymphoma dataset, hile JMIM, JMI and mRMR also perform well, with JMIM being the est of these. NJMIM performs better than DISR with all of the subsets elow 34 features. .5. Evaluating and validating results ANOVA statistical test is employed to analyse the results, and o confirm that the results are systematic and they were not ob- ained by chance. The classification experiment is run five times for ach dataset and the average accuracy results are submitted to the NOVA test. Table 3 shows the ANOVA results, where P-value is the M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 8529 0 5 10 15 20 25 30 35 40 45 50 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 Number of features C la ss ifi ca tio n a cc u ra cy CMIM NJMIM DISR JMIM mRMR JMI IG Fig. 10. Average classification accuracy achieved with the musk dataset. Table 3 ANOVA test. Dataset MS F P-value Credit approval 0.027537 731.3342 1.87E−37 Gas sensor 0.004117 77.17653 1.38E−16 Libra movement 0.009677 114.5907 2.94E−23 Parkinson 0.009677 114.5907 2.94E−23 Breast 0.001414 101.4627 2.37E−22 Sonar 0.00094 5.760126 9.62E−05 Musk 0.000505 304.4366 1.11E−30 Handwriting 8.84E−05 35.99929 6.35E−15 Colon 0.000411 3.532383 0.006395 Leukemia 0.000161 10.36207 2.21E−07 Lymphoma 0.011501 232.6585 1.28E−28 p m i b 6 o d t i s c t ( w d | n s r o t Table 4 Average stability, average accuracy and the compromise between accuracy and stability. Method Accuracy Stability Accuracy/stability CIMIM 0.8488 0.8598 0.9197 NJMIM 0.8264 0.8344 0.8954 DISR 0.8129 0.9054 0.8807 JMIM 0.8578 0.8598 0.9294 mRMR 0.8278 0.8868 0.8969 JMI 0.8490 0.8838 0.9199 IG 0.8226 0.9228 0.8913 l c c w w m o E s m w i m 0 c i n t t s o a t robability of the improvement to occur by chance, and MS is the ean square error. When the value of the P-value is less than 0.05 it s unlikely that the improvement in classification accuracy happened y chance. This is shown to be the case for all the datasets (Table 3). .6. Stability of the methods This section focuses on the stability of the feature selection meth- ds discussed. The selected subset features are dependent on the atasets provided, and therefore any change to the data might lead o different selected features. In this context, the present study nvestigates the influence of changes in the data on the features elected. Kuncheva’s measure of stability (Kuncheva, 2007), known as the onsistency index, uses Eq. (29) to compute the consistency between wo selected feature subsets, S1 and S2: S1, S2 ) = rn − k 2 k(n − k) , (29) here S1 and S2 are selected feature subsets using different groups of ataset samples, i.e. S1, S2 ∈ F where F is the total set of the feature, S1| = |S2| = k, |F | = n, and r = |S1 ∩ S2|. However, this method does ot take into consideration the correlation between features. Yu, Ding, and Loscalzo (2008) proposed a method for measuring tability based on similarity. This method takes into account the cor- elation between features. It calculates the weight between each pair f features from the subsets S1 and S2, computes the similarity be- ween S1 and S2, and constructs a bipartite graph. If f is a feature be- i onging to S1 and fj is a feature belonging to S 2, the value of the weight an be the correlation coefficient, or any other similarity measure. This article uses symmetrical uncertainty (Yu & Liu, 2004) to cal- ulate the weight w: ( s1i , s 2 j ) = 2 [ I ( s1 i , s2 j ) H ( s1 i ) + H ( s2 j ) ] , (30) here 0 ≤ w(s1 i , s2 j ) ≤ 1.0. To find the maximum weighted bipartite atching, the Hungarian Algorithm (Kuhn, 1955) is used to find the ptimal solution. This experiment uses the eight UCI datasets, as shown in Table 1. ach dataset is divided into 5 folds, 4 of which are used for feature election using the CMIM, NJMIM, DISR, JMIM, mRMR, JMI, and IG ethods. Eq. (30) is used to calculate the weight between the features ithin each pair of selected subsets from each dataset. The final cost s divided over the cardinality of the subset used, and therefore the agnitude of the final cost should be less than or equal to 0.5 (it is .5 if all selected subsets are the same). The relationship between accuracy and stability is computed by omparing the average classification accuracy and the average stabil- ty with different numbers of features. Table 4 shows the average accuracy/stability for each method in o particular order. It is worth noting that the methods employing he ‘maximum of the minimum’ criterion (JMIM, NJMIM and CMIM) end to have lower stability than the methods using the cumulative ummation approximation (JMI and DISR). The best method in terms f stability is IG. JMIM has the best compromise between accuracy nd stability. Moreover, it demonstrates the best average classifica- ion accuracy among all methods. 8530 M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 a-Leukemia 0 5 10 15 20 25 30 35 40 45 50 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 Number of features C la ss ifi ca tio n a cc u ra cy CMIM NJMIM DISR FIM JMIM mRMR JMI IG b-Colon 0 5 10 15 20 25 30 35 40 45 50 0.74 0.76 0.78 0.8 0.82 0.84 0.86 Number of features C la ss ifi ca tio n a cc u ra cy c-Lymphoma 0 10 20 30 40 50 60 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 Number of features C la ss ifi ca tio n a cc u ra cy Fig. 11. Average classification accuracy with the additional datasets. M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 8531 7 w d a P t t r a f g s c t t fi t J e c d a p i i u l I t w t t c o J d i n g d a m C o i 8 i a T r i a i d p m M S t t a t s a i t o t i o f t c g a h t t t r s t b J t d p t R B B B B B C C C D D D E E . Discussion The JMIM method outperforms the other methods when tested ith most of the datasets in terms of selecting the subset that pro- uces the best classification accuracy. JMIM also produces the best ccuracy with the datasets with a low number of features, such as the arkinson, credit approval and breast datasets. In these experiments, he maximum average classification accuracy achieved by JMIM with he Parkinson dataset was 90.77%. JMIM and JMI achieved the accu- acy of 82.92% with the credit approval dataset whilst JMI and CMIM chieved 93.83% and 95.22%, respectively. The JMIM method also per- ormed well on high dimensional datasets, such as the musk, sonar, as sensor and handwriting datasets. JMIM and JMI also outperform the other methods on extremely mall sample datasets with a large number of features, such as the olon dataset. However, CMIM produces the best performance with he lymphoma dataset. JMIM, JMI, and mRMR also perform better than he other three methods. Overall, JMIM decreases the average classi- cation error by 0.88% in absolute terms and almost 6% in relative erms in comparison to the next best performing method, JMI. The MIM classification accuracy is also higher than that reported in lit- rature by other filter methods (Zhang et al., 2015), although no firm onclusions can be made on this account due to the variety of the atasets used in the most recent articles (Liang et al., 2014; Zhang et l., 2015). In addition to the quantitative assessment of the accuracy of the roposed methods, several experiments are conducted to enable an n-depth comparison of different feature selection methods, accord- ng to several criteria. For example, the nonlinear approach, which ses the ‘maximum of the minimum’ criterion, is compared to the inear approach that employs cumulative summation approximation. n particular, JMIM is compared to JMI, with the results showing that he non-linear approach performed better than the linear approach hen tested with most of the datasets. The goal function based on joint mutual information is compared o the goal function based on conditional mutual information, with he result showing better performance of joint mutual information in ombination with the non-linear criterion. Finally, the effect of using normalised mutual information instead f mutual information is tested by comparing the performance of MIM and JMI with NJMIM and DISR. The results show that, with the iscretised datasets, the methods employing non normalised mutual nformation such as JMI and JMIM perform better than those using ormalised mutual information, such as DISR and NJMIM. This sug- ests that division of the mutual information over the joint entropy oes not improve performance. In addition, the methods are compared in terms of their stability, s described in detail in Section 6.5. The results demonstrate that the ethods employing ‘maximum of the minimum’ criterion, such as MIM, JMIM, and NJMIM, show less average stability than the meth- ds which employ cumulative summation, although there is no dom- nant method. . Conclusion This paper presents two new feature selection methods based on nformation theory: Joint Mutual Information Maximisation (JMIM) nd Normalised Joint Mutual Information Maximisation (NJMIM). hese methods are designed to resolve the problem of choosing edundant and irrelevant features in certain circumstances, which s characteristic of filter feature selection methods. The latter is chieved through the use of the mutual information and the ‘max- mum of the minimum’ nonlinear approach for the goal function esign. The methods have been evaluated using public datasets and com- ared with five other feature selection methods: Joint Mutual Infor- ation (JMI), Conditional Mutual Information Maximisation (CMIM), aximum Relevancy Minimum Redundancy (mRMR), Double Input ymmetrical Relevance (DISR), and Information Gain (IG) in terms of heir ability to select features with high discriminative power, and heir stability. To evaluate the performance of the proposed methods, n experiment is conducted using eight datasets from the UCI Reposi- ory. In addition, to test the behaviour of the methods with extremely mall sample datasets, three other datasets from Peng et al. (2005) re used. Overall, JMIM decreases the average classification error by 0.88% n absolute terms and almost by 6% in relative terms in comparison o the next best performing method, JMI. The statistical significance f the reported results is further confirmed by ANOVA test. Moreover, his method produces the best trade-off between accuracy and stabil- ty. The limitations of our approach are those which are characteristic f other filter approaches: it disregards the interaction between the eatures and the classifier, as well as the higher dimensional joint mu- ual information between more than two features, which sometimes an lead to suboptimal choice of features. Future work includes more experiments using other search strate- ies to validate the proposed method in a wider range of search lgorithms, employing parallel computation techniques to estimate igher dimensional joint mutual information in which two or more of he features from the selected subset are used simultaneously to test he significance of the candidate feature, automating the selection of he optimal subset by introducing a cut-off parameter measuring the elevancy of the features. Further improvements can be made by studying the information hared between features and class labels and classifying the fea- ures into strongly relevant, relevant, weakly relevant, and redundant ased on the information that the feature adds to the selected subset. In terms of applications relevant to expert and intelligent systems, MIM method would be of benefit for choosing the most relevant fea- ures in classification tasks. In addition to the analysis of the public atasets in this article, the method could be used in many other ap- lications where the relevance of the features for the classification ask needs to be analysed. eferences ache, K., & Lichman, M. (2013). UCI machine learning repository. Irvine, CA: Univer- sity of California, School of Information and Computer Science. (http://archive. ics.uci.edu/ml). ajwa, I., Naweed, M., Asif, M., & Hyder, S. (2009). Feature based image classification by using principal component analysis. ICGST International Journal on Graphics Vision and Image Processing, 9, 11–17. attiti, R. (1994). Using mutual information for selecting features in supervised neural net learning. IEEE Transactions on Neural Networks, 5, 537–550. olón-Canedo, V., Sánchez-Maroño, N., & Alonso-Betanzos, A. (2013). A review of fea- ture selection methods on synthetic data. Knowledge and Information Systems, 34, 483–519. rown, G., Pocock, A., Zhao, M., & Lujan, M. (2012). Conditional likelihood maximisa- tion: a unifying framework for information theoretic feature selection. Journal of Machine Learning Research, 13, 27–66. handrashekar, G., & Sahin, F. (2014). A survey on feature selection methods. Computers and Electrical Engineering, 40, 16–28. heng, H., Qin, Z., Feng, C., Wang, Y., & Li, F. (2011). Conditional mutual information- based feature selection analysing for synergy and redundancy. Electronics and Telecommunications Research Institute, 33, 210–218. over, T., & Thomas, J. (2006). Elements of information theory. New York: John Wiley & Sons. ing, C., & Peng, H. (2003). Minimum redundancy feature selection from microarray gene expression data. In Proceedings of the computational systems bioinformatics: IEEE Computer Society (pp. 523–528). ougherty, J., Kohavi, R., & Sahami, M. (1995). Supervised and unsupervised discretiza- tion of continuous features. In Proceedings of the twelfth international conference on machine learning (pp. 194–202). uda, R., Hart, P., & Stork, D. (2001). Pattern classification. New York: John Wiley and Sons. l Akadi, A., El Ouardighi, A., & Aboutajdine, D. (2008). A powerful feature selection approach based on mutual information. International Journal of Computer Science and Network Security, 8, 116–121. stévez, P. A., Tesmer, M., Perez, A., & Zurada, J. M. (2009). Normalized mutual informa- tion feature selection. IEEE Transactions on Neural Networks, 20, 189–201. http://archive.ics.uci.edu/ml http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005a http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005a http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005a http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005a http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005a http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005a http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0005a http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0012 8532 M. Bennasar et al. / Expert Systems With Applications 42 (2015) 8520–8532 L M M P R Q S T T V V Y Y Y Y Z Z Fleuret, F. (2004). Fast binary feature selection with conditional mutual information. Journal of Machine Learning Research, 5, 1531–1555. Freeman, C., Kulić, D., & Basir, O. (2015). An evaluation of classifier-specific filter mea- sure performance for feature selection. Pattern Recognition, 48, 1812–1826. Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature selection. Jour- nal of Machine Learning Research, 3, 1157–1182. Guyon, I., Gunn, S., Nikravesh, M., & Zadeh, L. A. (2006). Feature extraction foundations and applications. New York/Berlin, Heidelberg: Springer Studies in fuzziness and soft computing. Hoque, N., Bhattacharyya, D. K., & Kalita, J. K. (2014). MIFS-ND: a mutual information- based feature selection method. Expert Systems with Applications, 41(14), 6371– 6385. Jain, A. K., Duin, R. P. W., & Mao, J. (2000). Statistical Pattern Recognition: A Review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22, 4–37. Jakulin, A. (2003). Attribute interactions in machine learning. (M.S.c thesis), Computer and Information Science, University of Ljubljana. Jakulin, A. (2005). Machine learning based on attribute interactions (Ph.D. thesis), Com- puter and Information Science, University of Ljubljana. Janecek, A., Gansterer, W., Demel, M., & Ecker, G. (2008). On the relationship between feature selection and classification accuracy. Journal of Machine Learning Research: Workshop and Conference Proceedings, 4, 90–105. Karegowda, A. G., Jayaram, M. A., & Manjunath, A. S. (2010). Feature subset selection problem using wrapper approach in supervised learning. International Journal of Computer Applications, 1, 13–17. Kira, K., & Rendell, L. (1992). A practical approach to feature selection. In Proceedings of the 10th International Workshop on Machine Learning (ML92) (pp. 249–256). Kuhn, H. (1955). The Hungarian method for the assignment problem. Naval Research Logistic Quarterly, 2, 83–97. Kuncheva, L. (2007). A stability index for feature selection. In Proceedings of the 25th IASTED International Multi-Conference on Artificial Intelligence and Applications (pp. 390–395). Kwok, N., & Choi, C. (2002). Input feature selection for classification problems. IEEE Transactions on Neural Networks, 13, 143–159. Lee, J., & Kim, D. (2015). Fast multi-label feature selection based on information- theoretic feature ranking. Pattern Recognition, 48, 2761–2771. Liang, J., Wang, F., Dang, C., & Qian, Y. (2014). A group incremental approach to feature selection applying rough set technique. IEEE Transactions on Knowledge and Data Engineering, 26(2), 294–308. Lin, T., Li, H., & Tsai, K. (2004). Implementing the fisher’s discriminant ratio in a k- means clustering algorithm for feature selection and dataset trimming. Journal of Chemical Information and Computer Sciences, 44, 76–87. Liu, H., & Motoda, H. (2008). Computational methods of feature selection. New York: Chapman & Hall/CRC Taylor & Francis Group. iu, H., & Yu, L. (2005). Toward integrating feature selection algorithms for classification and clustering. IEEE Transactions on Knowledge and Data Engineering, 17, 491–502. eyer, P. E., & Bontempi, G. (2006). On the use of variable complementarity for feature selection in cancer classification. In Proceedings of European workshop on applica- tions of evolutionary computing: Evo Workshops (pp. 91–102). eyer, P. E., Schretter, C., & Bontempi, G. (2008). Information-theoretic feature selec- tion in microarray data using variable complementarity. IEEE Journal of Selected Topics in Signal Processing, 2, 261–274. eng, H., Long, F., & Ding, C. (2005). Feature selection based on mutual information: cri- teria of max-dependency, max-relevance, and min-redundancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27, 1226–1238. odgers, J., & Nicewander, W. A. (1988). Thirteen ways to look at the correlation coeffi- cient. The American Statistician, 42, 59–66. ian, Y., Wang, Q., Cheng, H., Liang, J., & Dang, C. (2015). Fuzzy-rough feature selection accelerator. Fuzzy Sets and Systems, 258, 61–78. aeys, Y., Inza, I., & Larranaga, P. (2007). A review of feature selection techniques in bioinformatics. Bioinformatics, 23, 2507–2517. ang, E. K., Suganthana, P. N., Yao, X., & Qina, A. K. (2005). Linear dimensionality reduc- tion using relevance weighted LDA. Pattern Recognition, 38, 485–493. urk, M., & Pentland, A. (1991). Eigenfaces for recognition. Journal of Cognitive Neuro- science, 3, 72–86. ergara, J., & Estévez, P. (2014). A review of feature selection methods based on mutual information. Neural Computing and Applications, 24, 175–186. idal-Naquet, M., & Ullman, S. (2003). Object recognition with informative features and linear classification. In Proceedings of the 10th IEEE international conference on computer vision (pp. 281–289). ang, H., & Moody, J. (1999). Feature selection based on joint mutual information. In Proceedings of international ICSC symposium on advances in intelligent data analysis (pp. 22–25). u, H., & Yang, J. (2001). A direct LDA algorithm for high-dimensional data with appli- cation to face recognition. Pattern Recognition, 34, 2067–2070. u, L., & Liu, H. (2004). Efficient feature selection via analysis of relevance and redun- dancy. Journal of Machine Learning Research, 5, 1205–1224. u, L., Ding, C., & Loscalzo, S. (2008). Stable feature selection via dense feature groups. In Proceedings of the 14th ACM SIGKDD international conference on knowledge dis- covery and data mining (pp. 803–811). hang, Y., Yang, A., Xiong, C., Wang, T., & Zhang, Z. (2014). Feature selection using data envelopment analysis. Knowledge-Based Systems, 64, 70–80. hang, Y., Yang, C., Yang, A., Xiong, C. Y., Zhou, X., & Zhang, Z. (2015). Feature selection for classification with class-separability strategy and data envelopment analysis. Neurocomputing, 166, 172–184. http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0023 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0023 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0024 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0024 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0024 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0024 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0025 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0025 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0025 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0025 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0034 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0035 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0036 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0036 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0036 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0036 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0036 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0037 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0038 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0038 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0038 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0038 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0039 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0039 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0039 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0039 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0040 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0040 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0040 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0040 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0041 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0041 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0041 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0041 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0042 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0043 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0043 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0043 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0043 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0044 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0044 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0044 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0044 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0044 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0045 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0046 http://refhub.elsevier.com/S0957-4174(15)00467-4/sbref0046 Feature selection using Joint Mutual Information Maximisation 1 Introduction 2 Information theory 3 Related work 4 Limitations of the current feature selection criteria 5 Proposed methods for feature selection 5.1 Joint Mutual Information Maximisation (JMIM) 5.2 Advantages over existing alternative methods 5.3 Normalised Joint Mutual Information Maximisation (NJMIM) 6 Evaluation 6.1 Data 6.2 Performance analysis on low dimensional datasets 6.3 Performance analysis on high dimensional datasets 6.4 Performance analysis with Peng et al. (2005) datasets 6.5 Evaluating and validating results 6.6 Stability of the methods 7 Discussion 8 Conclusion References 
work_xfbt6xqu6vhfxgpwshs2iyvicu	----	Author's personal copy Strategic group identification using evolutionary computation M.R. Martínez-Torres a,1, S.L. Toral-Marín b,* a Escuela Universitaria de Estudios Empresariales, Avda, San Francisco Javier, s/n, 41018 Sevilla, Spain b Escuela Superior de Ingenieros, Avda, Camino de los Descubrimientos, s/n, 41092 Sevilla, Spain a r t i c l e i n f o Keywords: Franchising Strategic groups Genetic Algorithms Evolutionary computation Performance a b s t r a c t This paper proposes to identify strategic groups among franchisors from a big set of franchisor variables. Genetic evolutionary computation was used to reduce a set of variables efficiently, and factor analysis was used to make up the strategic groups. Franchise 500 was used as database. The results suggest both that the general map of franchisor has changed since Carney and Gedajlovic’s study, and that genetic evo- lutionary computation is a valid way to extract knowledge from a huge set of data. This paper proposes useful information for those retail firms considering internationalization via franchising. The originality of this paper is in the use of Genetic Algorithm to discriminate the final set of variables to be used for the identification of strategic groups instead of evaluating one by one the adequacy of each variable theoret- ically. The ability of evolutionary computation to create new knowledge is good to produce new insights into this topic. � 2009 Elsevier Ltd. All rights reserved. 1. Introduction Franchise strategy refers to the outcome of the decision to operate and expand a business by franchising versus company- ownership (Chen & Ou, 2009; Falbe & Welsh, 1998). Two major theoretical perspectives have been proposed to explain patterns of company-ownership versus franchisee-owner- ship: Resource Scarcity and Agency Theory (Alon, 2001; Carney & Gedajlovic, 1991; Combs & Castrogiovanni, 1994; Combs, Ketchen, & Hoover, 2004; Michael, 2003; Newkirk & Lederer, 2006; Paik & Choi, 2007). Both theories should be analyzed, as they determine the firm characteristics that are going to drive franchisors into strategic groups. Advocates of the strategic group concept argue that industry members can be classified into groups along key characteristics, such as strategy and structure (e.g., Hatten & Schendel, 1977). In general, firms within an industry that follow a similar approach or strategy have been termed strategic groups (Porter, 1980), and strategic groups have been found to differ in performance (Ketchen, Thomas, & Snow, 1993). Opportunities are not evenly distributed across an industry; some strategies offer better profit potential than others. Firms may be tempted to change strategies to exploit opportunities as they arise, but shifting to a new strate- gic group can be risky because the necessary investment can be substantial and the perceived opportunities may be short lived (Patterson & Smith, 2003). All firms do not adopt franchising for similar reasons, but rather groups of firms share similar approaches (Carney & Gedajlovic, 1991). If strategic groups exist among firms that franchise and these groups differ in performance, examining all franchising firms as a set cannot capture the true picture of the franchising–perfor- mance relationship (Combs et al., 2004). The typical approach to strategic group identification consists of collecting detailed industry data, and then identifying groups through clustering or other grouping algorithms (Shanley & Peteraf, 2005; Sohn & Kim, 2008). The variables used to group will have much influence on the identified groups. A new per- spective is proposed in this paper. Instead of theoretically evalu- ating one by one the adequacy of each variable, a big set of them is used. Using an appropriate fitness function, an evolutionary algorithm will discriminate the final set of variables to be used for the identification, as well as the resulting strategic groups. We have used Genetic Algorithms (GA) for this purpose. One of the key advantages of evolutionary computation is its ability to discover new knowledge. The evolving nature of the computation can establish new relationships considered never before (Nanni & Lumini, 2009). The next section is devoted to the theoretical perspectives to franchising. Then, strategic groups and performance are intro- duced. After that, the evolutionary technique used in this paper (Genetic Algorithm) is shown: first, the basis of the methodology is described, and then the particular application to the strategic group identification problem is detailed. The obtained results are illustrated later. Finally, some conclusions have been drawn. 0957-4174/$ - see front matter � 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2009.12.019 * Corresponding author. Tel.: +34 954 48 12 93; fax: +34 954 48 7373. E-mail addresses: rmtorres@us.es (M.R. Martínez-Torres), toral@esi.us.es (S.L. Toral-Marín). 1 Tel.: +34 954 55 43 10; fax: +34 954 55 69 89. Expert Systems with Applications 37 (2010) 4948–4954 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Sergio Cuadro de texto Sergio Cuadro de texto Sergio Cuadro de texto Sergio Cuadro de texto Author's personal copy 2. Theoretical perspectives to franchising Oxenfeldt and Kelly (1968) offered perhaps the first explanation of why the proportion of franchised outlets differs among franchi- sors. Under the Resource Scarcity view, franchisors use franchising as a way to overcome constraints to growth, including a lack of trained managers and a lack of financial capital (Doherty, 2007). Success requires financial, informational, and managerial re- sources, but these resources are hard to obtain in the short run (Dant, Kaufmann, & Paswan, 1992). The franchisee provides an infusion of capital through fees and royalties and offers the franchi- sor (relatively) inexpensive growth. However, subsequent research tended to focus on the fact that firms used franchising because they also needed human capital (Norton, 1968), managerial talent (Dant et al., 1992; Doherty, 2007; Falbe & Welsh, 1998), and local knowledge (Combs & Castrogiovanni, 1994). Viewing franchising primarily as a means to access resources, a firm’s propensity to franchise varies over time. An implied tenet of Resource Scarcity Theory is the belief that the firm is more likely to increase company-ownership of sites as franchise systems mature and accumulate resources. This is precisely one of the criticisms that the Resource Scarcity thesis has received. Combs and Castrogiovanni (1994) observed that in contrast to the predictions of Resource Scarcity Theory, larger firms actually used more franchising in their development. They also found a weak negative relationship between the age of the franchisor and the use of franchising, and no relationship at all between the growth of the franchisor and the use of franchising. Some corporate giants as McDonald’s and Budget Rent-A-Car endorse this asseveration. The theory is also criticized by economists on the basis that capital can be raised more efficiently in the market. Although fran- chising may lower risk for the franchisor, it increases risk for the franchisee (Rubin, 1978). Agency Theory explains the organization of relationships when one agent determines the work and another undertakes it (Mole, 2002; Shane, 1998). In franchising, the Agency Theory perspective discusses it as the relationship between one party (the franchisor) depending on another party (the franchisee) to undertake some action on the franchisor’s behalf (Paik & Choi, 2007). Franchising encourages franchisees to maximize effort because, as owners, they must devote their own capital to build and operate outlets (Brickley & Dark, 1987). As a consequence, franchising lowers the cost of monitoring (Dant & Kaufmann, 2003; Pizanti & Lerner, 2003). Managerial talent and local knowl- edge are also eased by franchising because of the franchisees’ risk to lose their upfront monetary investments if their outlets fail as a result of their own managerial inadequacies (Shane, 1998). Quite the opposite of the Resource Scarcity Theory, Lafontaine and Kaufman suggests that agency factors favour an increased use of franchising as a chain expands with maturity (Lafontaine & Kaufmann, 1994). However, some agency problems are not solved by franchis- ing. There are some situations in which the franchisee may be inclined to shirk by under investing and free riding at the expense of the chain’s reputation (Michael, 2000). Although mon- itoring cost can be decreased, transferring specific knowledge to potential franchisees can also be costly (Jensen & Meckling, 1995). Some authors have proposed a reconciliation of both theories. Martin and Justis (1993) found that short- and long-run incentives to franchise differ. Whereas resource scarcity reasons to franchise are most relevant for young franchisors seeking to expand, agency reasons become increasingly relevant with maturity (Castrogiov- anni, Combs, & Justis, 2004). 3. Strategic groups and performance Researchers have long suspected the presence within industries of subgroups of firms whose behaviours and results differ from those of the broader industry (Lee, Lee, & Rho, 2002; Porter, 1976, 1979). Currently, there are three theoretical perspectives regarding strategic group formation: the industrial organization, strategy, and cognitive/behavioural perspectives (Hoyt & Sherman, 2004). The industrial organization perspective defines strategic groups as persistent features of the industry structure that result from en- try and mobility barriers. Structural barriers impede new firms from entering new industries (Audretsch, Houweling, & Thurik, 2004). Firms may be tempted to change strategies to exploit oppor- tunities as they arise, but shifting to a new strategic group can be risky because the necessary investment can be substantial, and the perceived opportunities may be short lived (Wheeler, Ibeh, & Dim- itratos, 2008). Thus, firms generally choose not to change groups because it is unclear whether the performance enhancements gained will exceed the costs incurred (Mascarenhas & Aaker, 1989). This is a consequence of the empirically determined recog- nition that single groups are separated by barriers which restrict the strategic mobility of firms (Caves & Porter, 1977). The strategy perspective is more internally focused and thus as- sumes that the firm’s management makes decisions to configure internal resources, so as to establish a sustainable competitive advantage (Hirschsohn, 2008). Finally, the cognitive perspective contends that strategic groups are formed by managers who partition their environment to re- duce uncertainty and who possess bounded rationality (Peteraf & Shanley, 1997). Based on strategic groups’ theory as well as the evidence linking strategic groups and performance, we expect that some strategic groups among franchisors will have strategic profiles that are bet- ter suited to their environment than others. Nevertheless, identify- ing strategic groups involves more than just choosing initial groups and data sources. The variables used to group will have much influ- ence on the identified groups. A detailed enumeration of nearly one hundred of such variables from prior studies has been detailed in Ketchen et al. (1993). Even if we were only concentrated in franchi- sor studies, the number of possible variables or indicators is huge. Insufficient consideration has been given to determine which vari- ables are appropriate for the purpose of distinguishing strategic groups. To avoid this, instead of theoretically evaluating one by one the adequacy of each variable, a big set of them will be used in this pa- per. Using an appropriate fitness function, an evolutionary algo- rithm will discriminate the final set of variables to be used for the identification, as well as the resulting strategic groups. 4. Methodology: Genetic Algorithms To identify strategic groups, we must first identify the firm characteristics that are likely to drive franchisors into distinctive strategic groups. We have used Genetic Algorithms (GA) for this purpose. Genetic Algorithms (GA) have been used to solve a variety of optimization problems, such as natural gas pipeline control, structural optimization, image registration, job scheduling, path planning, product design, etc. (Fisco, 2003; Goldberg, 1989; Lee et al., 2002; Li, Deng, & Luo, 2009). A Genetic Algorithm is a computational abstraction of biological evolution which can be used to solve some optimization problems. The technique was first introduced by Holland (1975) for use in adaptative systems. It is an iterative process applying a series of genetic operators such us selection, crossover and mutation to a M.R. Martínez-Torres, S.L. Toral-Marín / Expert Systems with Applications 37 (2010) 4948–4954 4949 Sergio Cuadro de texto Sergio Cuadro de texto Author's personal copy population of elements (Fig. 1). These elements, called chromo- somes or individuals, represent possible solutions to the problem. The initial population is randomly selected from the solution space. Genetic operators combine the genetic information of the ele- ments to form new generations of populations. Each chromosome has an associated fitness value which quantifies its value as a solu- tion to the problem. The chromosomes compete to reproduce based on their fitness values, thus the chromosomes representing better solutions have a higher chance of survival. The crossover in- volves two chromosomes whose portions are swapped. Selection according to fitness combined with crossover gives the GA its evo- lutionary power. The GA uses an elitist strategy meaning that the best individual is carried over to the next generation so that we can only improve the solution over the course of the genetic opti- mization. GA follows the iterative scheme shown in Fig. 1. The algorithm stops when some stopping criterions are satisfied. To successfully apply GA to solve a problem, it must be consid- ered the following items: � Chromosomal encoding, how to represent possible solutions. � Fitness function selection. It must accurately represent the value of the solution. � Parameter values selection (population size, number of itera- tions, probabilities, etc.) 4.1. Genetic Algorithm application to strategic group identification Prior studies about strategic group identification among fran- chisor have only considered a few numbers of variables. For instance, Carney and Gedajlovic (1991) considered thirteen variables for Canadian franchisor strategic group identification, while (Castrogiovanni, Bennett, & Combs, 1995) just used twelve variables for a similar study among US franchisors. Some other studies have replicated the former study in different countries (López & Ventura, 2001). All of them have utilized factor analysis as the multivariate sta- tistic tool for strategic group identification (Hsia, Hsu, & Jen, 2009; Sohn & Kim, 2008). Conventional techniques such as multiple regressions suffer the inconvenience of some restrictions like the independence of predictor variables or assumptions like linear relationships between variables. Factor Analysis is a way to fit a model to multivariate data to estimate just their interdependence. In the factor analysis model, the measured variables depend on a smaller number of unobserved (latent) factors. Because each factor may affect several variables in common, they are known as ‘‘common factors”. Each variable is as- sumed to be dependent on a linear combination of the common factors, and the coefficients are known as loadings. Factor analysis is useful in identifying inter-relationships between variables which are not directly observable, segmenting a sample into relatively homogeneous segments. The estimated loadings from an unrotated factor analysis fit can usually have a complicated structure. The goal of orthogonal factor rotation is to find a parameterization in which each variable has only a small number of large loadings, i.e., is affected by a small number of factors. The rotated factor analysis fit ensures factors represent unidimensional constructs (Hsu, 2009). Instead of using the same variables than Carney and Gedajlovic (1991) and Castrogiovanni et al., 1995 to replicate both studies, we propose to use a big set of variables as possible candidates for fac- tor analysis. A total of 31 candidate variables have been consid- ered, and they will be applied for strategic group identification of US franchisors included in the Entrepreneur’s Franchise 500 sur- vey, January 2005. Table 1 details the variables used in this study. Initial population (randomly generated) P(0) Individual fitnessess calculation f(I) for current Population P(t) Reproduction (selection based on Individual fitness f(I)) Corssover (selection based on Individual fitness f(I)) Mutation (selection based on Individual fitness f(I)) Initial population (randomly generated) P(0) Individual fitnessess calculation f(I) for current Population P(t) Reproduction (selection based on Individual fitness f(I)) Corssover (selection based on Individual fitness f(I)) Mutation (selection based on Individual fitness f(I)) Fig. 1. Iterative process of Genetic Algorithm. Table 1 Candidate variables for factor analysis. Var. Description Var01 No. of outlets 2005 Var02 Increment of outlets from 2004 to 2005 Var03 % of outlets located in US, 2005 Var04 % of outlets located in US, 2004 Var05 Average number of outlets opening per year Var06 Average number of outlets opening per year since first franchise Var07 Average investment Var08 Standard deviation of investment Var09 Average franchise fee Var10 Standard deviation of franchise fee Var11 Ongoing royalty fee Var12 Term of agreement Var13 Net worth requirement Var14 Cash liquidity requirement Var15 Number of employees per unit Var16 Franchises more than 1 Var17 % of outlets franchised 2005 Var18 % of foreign outlets franchised 2005 Var19 Years since inception Var20 Years between inception and first franchise Var21 Ranking Franchise 500, 2005 Var22 Ranking Franchise 500, 2004 Var23 Average investment/Term of agreement Var24 Average investment/Net worth requirement Var25 Average investment/Cash liquidity requirement Var26 % of outlets franchised in Canada, 2005 Var27 Average investment/Average franchise fee Var28 Franchise department Var29 No of outlets franchised/Franchise department Var30 Foreign franchises/Franchise department Var31 (Franchise fee + ongoing royalty fee) * term of agreement 4950 M.R. Martínez-Torres, S.L. Toral-Marín / Expert Systems with Applications 37 (2010) 4948–4954 Sergio Cuadro de texto Sergio Cuadro de texto Author's personal copy The objective of this work is to find the optimum subset of vari- ables from Table 1 leading to optimum strategic group identifica- tion. The obtained results will show if the thirteen variables considered in previous studies are enough, or if it would be neces- sary to extend the number of considered variables to achieve a bet- ter representation of US franchisor groups. But the number of possibilities when working with such huge number of variables suggests this problem is unapproachable if trying to explore all of them. The space of possible solutions is formed by 231 = 2.1475e + 009 possibilities. That means that we would have to perform 231 different factor analyses to completely explore the space of possible solutions. In this kind of problems, evolutionary computation can perform a guided search of the optimum solution with lower computation cost than exploring one by one all the possibilities. In this task, Genetic Algorithm will be applied to perform the evolutionary search of the optimum solution. The first condition to apply GA properly is a good selection of the chromosomal encoding, which should be valid and complete. Our chromosomal encoding is constituted by a 31 binary sequence in which ones are the variables that are going to be used in factor analysis and zeros represents variables that are going to be excluded from this analysis. Clearly, the encoding representation is complete, as the 231 possibilities are able to be represented, and valid, as all of them can be computed. The next step is the fitness function selection. The fitness func- tion quantifies the suitability of each chromosome as a solution. Chromosomes with high fitness have more chance of being se- lected, passing their genetic material (via reproduction or cross- over) to the next generation. The fitness function provides the pressure for evolution towards a new generation with chromo- somes of higher fitness than the previous ones. The chromosome representing the optimal solution should have the maximum fit- ness value for the solution space, and near optimal solutions should have high fitness values. In the context of strategic group identification by means of factor analysis, it is not possible to build a simple fitness function. Fitness function should be multi-objec- tive fitness function considering several parameters. F ¼ c1 Var þ c2 1 n Xk i¼1 r2i þ c3Dist This proposed fitness function consists of three parts. � Explained variance (Var). Factor analysis results show the explained variance by the considered factors (usually, the number of factors is given by the number of eigenvalues of the correlation data matrix bigger than 1). The explained vari- ance by the selected number of variables should be maximized. But it is not the unique parameter to be taken into account. A fit- ness function equal to the explained variance will tend to the trivial solution of just considering one variable. This is due to the fact that it is easier to explain the variance of a data set when it is formed by a small number of data. � Correlations between variables 1=n Pk i¼1 r 2 i � � . The average of the sum of the squared correlation coefficients between variables is used as the second part of the fitness function. This part of the fitness function will tend by itself to the trivial solution of con- sidering the whole data set. It is the reverse strength to the pre- vious part of the fitness function. � Number of factors. The third part of the fitness function measures the distance (Dist) to the typical number of factor of previous studies about strategic group identification. A distant number from 5 groups is penalized, as this is a typical value obtained en prior studies (Carney & Gedajlovic, 1991; Castro- giovanni et al., 1995). C1, C2, and C3 are coefficients used to adjust the relative impor- tance of the three parts of the fitness function. The range of them is [0, 1], with the restriction of C1 + C2 + C3 = 1. The final decision for GA application refers to parameter values selection. GA performance may be sensitive to certain parameter values, particularly the population size, the frequency of operator selection and the termination criterion. All of them vary consider- ably, and there is little or no documented justification for the selec- tion of them. Nevertheless, a high value for the population size may reduce this sensibility to GA parameters. In this paper, popu- lation size has been chosen equal to 10,000, with a 20% of repro- duction rate. The value of 10,000 is considered a good value to obtain richness genetic content. These values are typical in the lit- erature about GA (Cole, 1998; Goldberg, 1989). 5. Results Beginning with an initial randomly generated population, GA has converged after 30 generations, with an explained variance of 82.8%, 20 variables and six factor (that means six eigenvalues higher than 1). Time required by genetic algorithm execution is 750 s (12.5 min). This value is much smaller than the alternative option of exploring the whole solution space. Taking into account that each factor analysis requires 2.5 ms, the 231 = 2.1475e+009 possi- bilities of the solution space would require more than 62 days, 24 h a day. The evolution of the genetic clustering algorithm is detailed in Fig. 2. The initial population (generation 0) has a low fitness value, which indicates that the individuals of the population are far from the optimum. As the number of generations increase, the fitness of individuals within the population also increases, as the genetic algorithm is biased towards the survival of genetic material con- tained within the individuals with high fitness function values. The results from factor analysis using the set of variables selected by the genetic algorithm are detailed in Table 2. Although the number of eigenvalues higher than 1 is six, due to the gap between the fifth and the sixth eigenvalue is preferable to choose just five components or factors. Factor loadings with Fig. 2. Fitness distribution over 30 generations of the genetic algorithm. M.R. Martínez-Torres, S.L. Toral-Marín / Expert Systems with Applications 37 (2010) 4948–4954 4951 Sergio Cuadro de texto Sergio Cuadro de texto Author's personal copy varimax rotation are shown in Table 3. Each row represents the factor loadings of each variable. Moving horizontally from left to right across the five loadings in each row, the highest loading has to be identified. All the variables associated in this way with the same factor are hypothesized to share a common meaning that the analyst should discover. The interpretation of the rotation loadings leads to the resulting strategic groups, represented by the different clusters identified (Table 4). To facilitate the interpretation of these, Table 5 shows the mean value of the selected variables per each of the five factors. Cluster 1 brings together variables related with investment. It is characterized by a high value of all the loadings and it can be resembled to the ‘‘expensive” group identified by Carney and Ged- ajlovic (1991). The identified group has anyway some differences with respect to the one they proposed, as they are not so conserva- tive. They exhibit a moderate rate of growth, according to var04. Expensive franchisors are primarily concerned with resource scar- city, demonstrated by the high value of all the variables in which investment is involved. Cluster 2 refers to highly internationalized US franchisors. Var03, referring to the percentage of outlets located in US, shows a negative loading in Table 4, corresponding to a minimum value in Table 5 when compared with the other factors. They are charac- terized by a great confidence on franchising as a way of minimizing the agency cost. This conclusion is suggested by the high values of the ratios given by the number of outlets and the foreign outlets, and the size of franchise department (Var19 and Var20, respectively). Cluster 3 is somewhat related with the previous one. They are also very big franchisor, with a high rate of growth, but, as a differ- ence with the previous cluster, they are focused on domestic mar- ket. According to Table 5, var03 has the maximum value in cluster 3, just the opposite than in cluster 2. Var14 also exhibit a negative loading in Table 4. Var14 is the ranking 2005 in Franchise 500. The lower this value is, the more stability and reliability the franchisor Table 2 Explained variance of resulting factor analysis. Component Eigenvalues Value Proportion of total variance Cumulative proportion Var01 No. of outlets 2005 5127 25.636 25.636 Var02 Increment of outlets from 2004 to 2005 4438 22.191 47.826 Var03 % of outlets located in US. 2005 2391 11.953 59.780 Var04 Average number of outlets opening per year 1861 9305 69.084 Var05 Average number of outlets opening per year since first franchise 1561 7804 76.888 Var06 Average investment 1189 5943 82.831 Var07 Standard deviation of investment .984 4921 87.752 Var08 Net worth requirement .608 3041 90.793 Var09 Cash liquidity requirement .493 2464 93.257 Var10 % of outlets franchised 2005 .367 1834 95.091 Var11 % of foreign outlets franchised 2005 .271 1354 96.445 Var12 Years since inception .198 .992 97.437 Var13 Years between inception and first franchise .174 .870 98.307 Var14 Ranking franchise 500, 2005 .107 .537 98.844 Var15 Average investment/term of agreement .090 .451 99.295 Var16 Average investment/net worth requirement .061 .307 99.602 Var17 Average investment/cash liquidity requirement .035 .176 99.778 Var18 % of outlets franchised in Canada, 2005 .027 .136 99.914 Var19 No of outlets franchised/franchise department .013 .065 99.978 Var20 Foreign franchises/franchise department .004 .022 100.000 Table 3 Factor loadings obtained by the principal component method with Varimax rotation. Component 1 2 3 4 5 Var01 No. of outlets 2005 .014 .794 .521 .054 .088 Var02 Increment of outlets from 2004 to 2005 .019 �.074 .723 �.119 .068 Var03 % of outlets located in US, 2005 .004 �.623 .187 �.084 �.583 Var04 Average number of outlets opening per year �.027 .639 .666 �.022 �.156 Var05 Average number of outlets opening per year since first franchise �.008 .567 .739 �.098 �.043 Var06 Average investment .896 �.004 �.046 .344 .035 Var07 Standard deviation of investment .936 �.012 �.031 .172 �.016 Var08 Net worth requirement .177 8690E�05 �.088 .918 .074 Var09 Cash liquidity requirement .148 .020 �.044 .907 .064 Var10 % of outlets franchised 2005 .126 .025 .146 �.435 �.615 Var11 % of foreign outlets franchised 2005 .098 .783 �.081 �.266 .229 Var12 Years since inception .176 .224 .187 .062 .758 Var13 Years between inception and first franchise .168 �.066 .158 �.107 .821 Var14 Ranking franchise 500, 2005 �.036 �.300 �.686 �.146 �.233 Var15 Average investment/term of agreement .903 �.007 �.105 .306 .049 Var16 Average investment/net worth requirement .839 �.013 .077 �.227 .214 Var17 Average investment/cash liquidity requirement .867 .005 .106 �.195 .054 Var18 % of outlets franchised in Canada, 2005 .074 .201 �.281 �.362 .019 Var19 No of outlets franchised/franchise department �.070 .826 .299 �.003 �.030 Var20 Foreign franchises/franchise department �.047 .904 .028 .035 .012 4952 M.R. Martínez-Torres, S.L. Toral-Marín / Expert Systems with Applications 37 (2010) 4948–4954 Sergio Cuadro de texto Sergio Cuadro de texto Author's personal copy is (the first franchisor of the ranking is the best one). So, it is nat- ural a minimum value of var14 in Table 5 meaning that this cluster is formed by the best franchisor according to Franchise 500 criteri- ons. Reasons for franchising are closer to resource scarcity than to the agency theory. Cluster 4 is formed by conservative franchisor. They exhibit the lower rate of growth of all the groups (var04 and var05), with a lot of requirements about franchises, particularly net worth and cash liquidity requirements. They are conservative, and moderate expensive compared with franchisors included in cluster 1, primar- ily concerned with resource scarcity. Finally, cluster 5 can be assimilated to franchise converts group of Carney and Gedajlovic (1991) study. Converts typically begin operations as wholly-owned chains, and after many years decide to begin franchising. According to the obtained results, it can be said that franchisors are generally concerned with both administrative efficiency and resource scarcity, though their concern over one relative to the other may vary somewhat with their particular circumstances. In order to compare the means of variables from Table 5, a Krus- kal–Wallis test has been performed. The Kruskal–Wallis test is a nonparametric version of one-way analysis of variance. The assumption behind this test is that the measurements come from a continuous distribution, but not necessarily a normal distribu- tion. The test is based on an analysis of variance using the ranks of the data values, not the data values themselves. The low p value for each variable suggests that the mean of each one of them is sig- nificantly different than the other sample means. The null hypoth- esis can be rejected, so it can be concluded that the obtained strategic groups are discrete (see Table 6). 6. Conclusions and implications The first conclusion that can be emphasized is that the general map of franchisor since Carney and Gedajlovic study has changed. Franchising as a way of growth has been adapted to the current cir- cumstances through the years. Highly internationalized franchi- sors appears now as a clearly distinguished group. Probably, the old group of mature franchisor has split into the groups identified Table 4 Clusters identified by factor loadings. Cluster Var. Description Loading Cluster1 Var06 Average investment 0.896 Var07 Standard deviation of investment 0.936 Var15 Average investment/term of agreement 0.903 Var16 Average investment/net worth requirement 0.839 Var17 Average investment/cash liquidity requirement 0.867 Cluster2 Var01 No. of outlets 2005 0.794 Var03 % of outlets located in US, 2005 �0.623 Var11 % of foreign outlets franchised 2005 0.783 Var19 No of outlets franchised/franchise department 0.826 Var20 Foreign franchises/franchise department 0.904 Cluster3 Var02 No. of outlets 2004 0.723 Var04 Outlets opening per year 0.666 Var05 Outlets opening per year since first franchise 0.739 Var14 Ranking Franchise 500, 2005 �0.686 Cluster4 Var08 Net worth requirement 0.918 Var09 Cash liquidity requirement 0.907 Cluster5 Var10 % of outlets franchised 2005 �0.615 Var12 Years since inception 0.758 Var13 Years between inception and first franchise 0.821 Table 5 Selected variables mean values per factor. Factor 1 (10.91%) Factor 2 (19.09%) Factor 3 (25.45%) Factor 4 (27.27%) Factor 5 (17.27%) VAR01 1321.30 5336.60 4054.00 1393.30 2717.30 VAR02 67.25 94.86 342.89 45.35 87.05 VAR03 0.87 0.60 0.90 0.83 0.61 VAR04 44.11 128.16 161.89 43.66 46.64 VAR05 62.05 139.26 194.32 40.00 75.01 VAR06 2250.80 180.84 126.92 528.51 428.64 VAR07 2152.50 108.22 73.73 293.72 192.55 VAR08 364.50 189.55 181.73 550.77 354.79 VAR09 127.51 61.44 54.07 236.88 112.82 VAR10 1.00 0.98 0.99 0.87 0.74 VAR11 0.12 0.39 0.09 0.05 0.21 VAR12 27.92 29.62 25.11 28.23 54.26 VAR13 8.42 3.52 4.25 2.20 26.16 VAR14 67.42 62.38 38.00 65.83 47.74 VAR15 136.01 15.06 14.32 39.59 33.05 VAR16 5.59 0.92 1.30 1.01 1.88 VAR17 14.65 3.63 4.41 2.41 4.13 VAR18 0.04 0.09 0.03 0.01 0.03 VAR19 52.58 411.80 207.80 86.43 118.13 VAR20 4.87 221.68 17.57 5.61 28.57 Table 6 Kruskal–Wallis test. VAR01 VAR02 VAR03 VAR04 VAR05 VAR06 VAR07 VAR08 VAR09 VAR10 Chi-2 14.944 21.454 40.163 17.565 23.990 40.960 33.109 26.825 38.161 31.975 gl 4 4 4 4 4 4 4 4 4 4 Sig. asintót. .005 .000 .000 .002 .000 .000 .000 .000 .000 .000 VAR11 VAR12 VAR13 VAR14 VAR15 VAR16 VAR17 VAR18 VAR19 VAR20 Chi-2o 40.835 27.346 31.750 14.993 36.867 22.947 24.957 19.485 11.943 20.370 gl 4 4 4 4 4 4 4 4 4 4 Sig. asintót. .000 .000 .000 .005 .000 .000 .000 .001 .018 .000 M.R. Martínez-Torres, S.L. Toral-Marín / Expert Systems with Applications 37 (2010) 4948–4954 4953 Sergio Cuadro de texto Sergio Cuadro de texto Author's personal copy by clusters 2 and 3. The former has adopted franchising as a way of internationalization, while the latter is still confident in franchising as a way to growth, primarily in US market. The group ‘‘expensive and conservative” has been also split into two groups. Finally, convert franchisors remain the same. The second conclusion is the application of evolutionary com- putation, particularly Genetic Algorithm, as a valid way to extract knowledge from a huge set of data. Replication of a previous study is useful to confirm obtained results, but the inconvenience is that replication does not allow new insights in a particular topic. The ability of Genetic Algorithm to explore new relationships within a big amount of data allows exploring new solutions or new per- spectives in a particular theme, like the ones mentioned above. References Alon, I. (2001). The use of franchising by US – based retailers. Journal of Small Business Management, 39(2), 1–12. Audretsch, D. B., Houweling, P., & Thurik, A. R. (2004). Industry evolution. Diversity, selection and the role of learning. International Small Business Journal, 22(4), 331–348. Brickley, J. A., & Dark, F. H. (1987). The choice of organizational form: The case of franchising. Journal of Financial Economics, 18, 401–420. Carney, M., & Gedajlovic, E. (1991). Vertical integration in franchise systems: Agency theory and resource explanations. Strategic Management Journal, 12(8), 607–629. Castrogiovanni, G. J., Bennett, N., & Combs, J. G. (1995). Franchisor types: Reexamination and clarification. Journal of Small Business Management, 33(1), 45–54. Castrogiovanni, G. J., Combs, J. G., & Justis, R. T. (2004). Toward a reconciliation of resource and agency views on franchising. In Academy of management meeting, New Orleans, LA (pp. 1–7). Caves, R. E., & Porter, M. E. (1977). From entry barriers to mobility barriers: Conjectural decisions and contrived deterrence to new competition. Quarterly Journal of Economics, 91, 241–261. Chen, F. L., & Ou, T. Y. (2009). Gray relation analysis and multilayer functional link network sales forecasting model for perishable food in convenience store. Expert Systems with Applications, 36(3), 7054–7063. Cole, R. M. (1998). Clustering with genetic algorithm. Ph.D. Dissertation, University of Western Australia. Combs, J. G., & Castrogiovanni, G. J. (1994). Franchisor strategy: A proposed model and empirical test of franchise versus company ownership. Journal of Small Business Management, 32(2), 37–48. Combs, J. G., Ketchen, D. J., Jr., & Hoover, V. L. (2004). A strategic groups approach to the franchising–performance relationship. Journal of Business Venturing, 19, 877–897. Dant, R. P., Kaufmann, P. J., & Paswan, A. K. (1992). Ownership redirection in franchised channels. Journal of Public Policy and Marketing, 11(1), 33–44. Dant, R. P., & Kaufmann, P. J. (2003). Structural and strategic dynamics in franchising. Journal of Retailing, 79, 63–75. Doherty, A. M. (2007). The internationalization of retailing. Factors influencing the choice of franchising as a market entry strategy. International Journal of Service Industry Management, 18(2), 184–206. Falbe, C. M., & Welsh, D. H. B. (1998). Nafta and franchising: A comparison of franchisor perceptions of characteristics associated with franchisee success and failure in Canada, Mexico, and the United States. Journal of Business Venturing, 13, 151–171. Fisco, J. H. (2003). Optimal dispersion of R&D activities in multinational corporations with a genetic algorithm. Research Policy, 32, 1381–1396. Goldberg, D. A. (1989). Genetic Algorithm – in Search, Optimization and Machine Learning. Addison-Wesley Publishing Company, Inc. Hatten, K. J., & Schendel, D. E. (1977). Heterogeneity within an industry: Firm conduct in the US brewing industry, 1952–1971. Journal of Industrial Economics, 26, 97–113. Hirschsohn, P. (2008). Regulating the ‘Animal Spirits’ of entrepreneurs? Skills development in South African small and medium enterprises. International Small Business Journal, 26(2), 181–206. Holland, J. (1975). Adaptation in natural and artificial systems. Ann Arbor, MI: University of Michigan Press. Hoyt, J., & Sherman, H. (2004). Strategic groups, exit barriers and strategy decision constraints in high-tech companies. Management Research, 15, 237–247. Hsia, T.-C., Hsu, Y.-L., & Jen, H.-L. (2009). A factor analysis based selection process for predicting successful university color guard club members. Expert Systems with Applications, 36(2), 2719–2726. Hsu, L.-C. (2009). Forecasting the output of integrated circuit industry using genetic algorithm based multivariable grey optimization models. Expert Systems with Applications, 36(4), 7898–7903. Jensen, M. C., & Meckling, W. (1995). Specific and general knowledge and organizational structure. Journal of Applied Corporate Finance, 8, 4–18. Ketchen, D., Thomas, J., & Snow, C. (1993). Organizational configurations and performance: A comparison of theoretical approaches. Academy of Management Journal, 36, 1278–1313. Lafontaine, F., & Kaufmann, P. (1994). The evolution of ownership patterns in franchise systems. Journal of Retailing, 70(2), 97–113. Lee, J., Lee, K., & Rho, S. (2002). An evolutionary perspective on strategic group emergence: A Genetic algorithm-based model. Strategic Management Journal, 23, 727–746. Li, X., Deng, Z., & Luo, J. (2009). Trading strategy design in financial investment through a turning points prediction scheme. Expert Systems with Applications, 36(4), 7818–7826. López, M. B., & Ventura, J. (2001). Grupos estratégicos en las franquicias españolas: Características básicas. Economía Industrial, 340, 163–176. Martin, R., & Justis, R. (1993). Franchising, liquidity constrains and entry. Applied Economics, 25, 1269–1277. Mascarenhas, B., & Aaker, D. A. (1989). Mobility barriers and strategic groups. Strategic Management Journal, 10, 475–485. Michael, S. C. (2000). The effect of organizational form on quality: The case of franchising. Journal of Economic Behavior and Organization, 43, 295–318. Michael, S. C. (2003). First mover advantage through franchising. Journal of Business Venturing, 18, 61–80. Mole, K. (2002). Business advisers’ impact on SMEs: An agency theory approach. International Small Business Journal, 20, 139–162. Nanni, L., & Lumini, A. (2009). A genetic encoding approach for learning methods for combining classifiers. Expert Systems with Applications, 36(4), 7510–7514. Newkirk, H. E., & Lederer, A. L. (2006). Incremental and comprehensive strategic information systems planning in an uncertain environment. IEEE Transactions on Engineering Management, 53(3), 380–394. Norton, S. (1968). An empirical look at franchising as an organizational form. Journal of Business, 61, 197–217. Oxenfeldt, A. R., & Kelly, A. O. (1968). Will successful franchise systems ultimately become wholly-owned chains? Journal of Retailing, 44(4), 69–83. Paik, Y., & Choi, D. Y. (2007). Control, autonomy and collaboration in the fast food industry. A comparative study between domestic and international franchising. International Small Business Journal, 25(5), 539–562. Patterson, P. G., & Smith, T. (2003). A cross-cultural study of switching barriers and propensity to stay with service providers. Journal of Retailing, 79, 107–120. Peteraf, M., & Shanley, M. (1997). Getting to know you: A theory of strategic group identity. Strategic Management Journal, 18, 165–186. Pizanti, I., & Lerner, M. (2003). Examining control and autonomy in the franchisor– franchisee relationship. International Small Business Journal, 21(2), 131–159. Porter, M. (1976). Interbrand choice, strategy and bilateral market power. Cambridge, MA: Harvard University Press. Porter, M. (1979). The structure within industries and companies’ performance. Review of Economics and Statistics, 61, 214–227. Porter, M. (1980). Competitive strategy. New York: Free Press. Rubin, P. H. (1978). The expansion of firms. Journal of Political Economy, 81(4), 936–949. Shane, S. (1998). Explaining the distribution of franchised and company-owned outlets in franchise systems. Journal of Management, 24, 717–739. Shanley, M., & Peteraf, M. (2005). Balancing theory and technique: Methodological issues in strategic groups research. Research Methodology in Strategy and Management, 2, 65–92. Sohn, S. Y., & Kim, Y. (2008). Searching customer patterns of mobile service using clustering and quantitative association rule. Expert Systems with Applications, 34(2), 1070–1077. Wheeler, C., Ibeh, K., & Dimitratos, P. (2008). UK export performance research. International Small Business Journal, 26(2), 207–239. 4954 M.R. Martínez-Torres, S.L. Toral-Marín / Expert Systems with Applications 37 (2010) 4948–4954 Sergio Cuadro de texto Sergio Cuadro de texto 
work_xfej6opxpfcevg7z7op52o6oti	----	Choose and Book: A sociological analysis of `resistance' to an expert system lable at ScienceDirect Social Science & Medicine 104 (2014) 210e219 Contents lists avai Social Science & Medicine journal homepage: www.elsevier.com/locate/socscimed Choose and Book: A sociological analysis of ‘resistance’ to an expert system Trisha Greenhalgh a,*, Rob Stones b, Deborah Swinglehurst a a Centre for Primary Care and Public Health, Barts and The London School of Medicine and Dentistry, London, UK b School of Social Sciences and Psychology, University of Western Sydney, Locked Bag 1797, Penrith, NSW 2751, Australia a r t i c l e i n f o Article history: Available online 19 December 2013 Keywords: United Kingdom Expert systems e-Health Structuration theory Referral Resistance Electronic patent records * Corresponding author. Yvonne Carter Building, London E1 2AB, UK. Fax: þ44 20 7882 2552. E-mail address: p.greenhalgh@qmul.ac.uk (T. Gree 0277-9536 � 2013 The Authors. Published by Elsevie http://dx.doi.org/10.1016/j.socscimed.2013.12.014 a b s t r a c t In 2004, the English Department of Health introduced a technology (Choose and Book) designed to help general practitioners and patients book hospital outpatient appointments. It was anticipated that remote booking would become standard practice once technical challenges were overcome. But despite political pressure and financial incentives, Choose and Book remained unpopular and was generally used reluc- tantly if at all. Policymakers framed this as a problem of ‘clinician resistance’. We considered Choose and Book from a sociological perspective. Our dataset, drawn from a qualitative study of computer use in general practice, comprised background documents, field notes, interviews, clinical consultations (directly observed and videotaped) and naturally occurring talk relating to referral to hospital in four general practices. We used strong structuration theory, Giddens’ conceptualisation of expert systems, and sensitivity to other sociological perspectives on technology, institutions and professional values to examine the relationship between the external environment, the evolving technology and actions of human agents (GPs, administrators, managers and patients). Choose and Book had the characteristics of an expert system. It served to ‘empty out’ the content of the consultation as the abstract knowledge it contained was assumed to have universal validity and to over-ride the clinician’s application of local knowledge and practical wisdom. Sick patients were incorrectly assumed to behave as rational choosers, able and willing to decide between potential options using abstracted codified information. Our analysis revealed four foci of resistance: to the policy of choice that Choose and Book symbolised and purported to deliver; to accommodating the technology’s socio-material constraints; to interference with doctors’ contextual judgements; and to adjusting to the altered social relations consequent on its use. We conclude that ‘resistance’ is a complex phenomenon with socio-material and normative components; it is unlikely to be overcome using the behaviourist techniques recommended in some health informatics and policy literature. � 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY license. Introduction ‘Resistance’ to information technology in healthcare Healthcare depends increasingly on information and communi- cation technologies (ICTs), whose introduction is oftencharacterised by limited adoption or adoption followed by abandonment, 58 Turner St, Whitechapel, nhalgh). r Ltd. Open access under CC BY licens especially when part of a large, top-down change programme (see for example (Greenhalgh, Morris, Wyatt, & Thomas, 2012; Greenhalgh et al., 2010; Robertson et al., 2010; Sanders et al., 2012)). The health informatics literature tends to explain such ‘failed’ projects in terms of resistance (depicted as an ill-defined combination of inertia, anxiety and Luddism) and to couch solu- tions in terms of securing behavioural compliance without ques- tioning ends. For example: “.the major challenges to system success are often more behav- ioral than technical. Successfully introducing such systems into complex health care organizations requires an effective blend of good technical and good organizational skills. People who have low psychological ownership in a system and who vigorously resist its implementation can bring a ‘technically best’ system to its knees. However, effective leadership can sharply reduce the behavioral e. Delta:1_given name Delta:1_surname Delta:1_given name mailto:p.greenhalgh@qmul.ac.uk http://crossmark.crossref.org/dialog/?doi=10.1016/j.socscimed.2013.12.014&domain=pdf www.sciencedirect.com/science/journal/02779536 http://www.elsevier.com/locate/socscimed http://dx.doi.org/10.1016/j.socscimed.2013.12.014 http://dx.doi.org/10.1016/j.socscimed.2013.12.014 http://dx.doi.org/10.1016/j.socscimed.2013.12.014 http://creativecommons.org/licenses/by/3.0/ http://creativecommons.org/licenses/by/3.0/ Box 1 Referral to hospital using Choose and Book technology. At the time of this study, Choose and Book software was materially accessible from the desktop of GPs and admin- istrative staff in almost all practices in UK, but the GPs we studied did not use it themselves. By manually entering the patient’s NHS number (or, more rarely, by using an auto- mated link), the administrator called up the patient’s cen- trally held demographic details from the NHS ‘Spine’. Next, they selected the priority (‘routine’, ‘urgent’ or ‘two week wait’ e the last for suspected cancer) and entered a clinical specialty (e.g. ‘gynaecology’) and clinic type (e.g. ‘infer- tility’). This generated a preset menu that allowed providers to be compared on three criteria: distance from the patient’s postcode, key word (such as the name of a provider) and indicative waiting time (based on the last 20 appointments in the system). Clicking on any of the providers on this list would (in theory) call up information about them from a directory of services (supplied by providers), including which specific clinical conditions the clinic covered. The facility to refer to particular consultants was technically possible but not usually available in practice at the time of the study. Once a provider clinic was selected, a referral letter (typically dictated and typed a day or two later) could be uploaded onto the system. Alternatively, the patient could be given (or sent) printed instructions, including a unique identifier and password, allowing them to access the above information themselves by telephoning a dedicated call centre or via the Internet using the ‘HealthSpace’ personal portal, for which they needed to register in advance by post. HealthSpace con- tained a link to the ‘NHS Choices’ website, which offered three types of comparative information on providers: per- formance against published indicators (e.g. infection rates, mortality rates, food quality scores), patient experiences (bulletin-board postings in similar format to travel experi- ence websites) and distance from the patient’s postcode. The website allowed the patient to generate bespoke tables to allow ‘objective’ comparison of hospitals against their preferred criteria. Some services on the Choose and Book system offered ‘directly bookable’ appointments that allowed a time and date to be obtained immediately, but if the provider’s sys- tem was not yet integrated with Choose and Book, the pa- tient was sent a letter by the practice with a reference number and the hospital’s telephone number to book their own appointment. Once the hospital service received the booking (along with the referral letter), a decision was made (usually by an administrator) to accept or reject the referral. They could also contact the patient to change the date or time of a directly booked appointment. T. Greenhalgh et al. / Social Science & Medicine 104 (2014) 210e219 211 resistance to changeeincluding to new technologieseto achieve a more rapid and productive introduction of informatics technology.” (Lorenzi & Riley, 2000, page 116) Healthcare IT policy typically reflects this behaviourist framing by focusing on incentives, sanctions and training (Department of Health, 2012). In contrast, socio-technical systems theory pro- poses that technologies and work practices are best co-designed using participatory methods in the workplace setting, drawing on such common-sense guiding principles as staff being ‘able to access and control the resources they need to do their jobs’, and insisting that ‘processes should be minimally-specified (e.g. stipulating ends but not means) to support adaptive local solutions’ (Cherns, 1987). Socio-technical theory frames resistance to ICTs in terms of poor fit between the micro-detail of work practices and the practicalities of using technology. Brown and Duguid (2002) have shown how technologies in the workplace are embedded in networks of social relationships that make their use meaningful. The detail of how to use, adapt, repair or work round technologies is learned through membership of a community of practice; this social infrastructure strongly influences whether and how particular technologies ‘work’ in particular conditions of use. In this paper we acknowledge this perspective and seek to complement it with a multi-level theoretical analysis that con- siders macro forces emanating from government and state agencies; meso-level networks that mediate these forces; and micro-level sites of acquiescence or resistance by human agents. We incorporate selected insights from actor-network theory (ANT), which conceptualises networks of humans and technologies that are dynamic and (to a greater or lesser extent) unstable. ANT use- fully considers human actors’ (and indeed technologies’) behaviour as a consequence of the overall pattern of influences generated across the network. In this study our preferred analytic lens is structuration theory, developed by Anthony Giddens (1984). We adopt a layered ontology, finding it productive to make distinctions between structure and agency, and between macro, meso and micro, which ANT rejects. We integrate technology into the picture, a dimension that is missing from Giddens’ work (Greenhalgh & Stones, 2010). We also seek to go beyond Giddens’ abstract concern with social structures in general and use an empirical case study approach to look at particular fields of social relations. This emphasis has many parallels with ANT’s notion of networks. Taking a layered approach to the study of relations or networks highlights the ways in which the interdependencies and interactions that constitute these net- works are embedded in hierarchical power relations, both near and distant. Unlike ANT, this ‘strong’ version of structuration theory carefully distinguishes the agency of humans from that of tech- nologies (the latter, we contend, can only ‘act’ in a limited way). Strong structuration theory also considers how the values and knowledge possessed by both individual and organisational actors are influenced by external structures, and how this value- knowledge nexus informs and influences their actions, with or without technologies, in particular social situations (Stones, 2005b). For strong structuration theorists, resistance to ICTs stems from the human agent, who is positioned in a particular network of social relations; has a particular identity, organisational role, set of moral principles, beliefs, capabilities, and so on; and accords significance to technologies in particular contexts. We sought to apply strong structuration theory to explore resistance to a nationally mandated healthcare ICT, ‘Choose and Book’ (a system for online outpatient referrals in England e Box 1), in terms of the reasoning and actions of human agents and how this was influenced both by social structures (especially social and professional norms and values and political authority) and by the material capabilities and constraints of the technology. To that end, we undertook a secondary analysis of a rich ethnographic dataset on the use and non-use of electronic records in UK general practice. The policy context As with many healthcare ICTs, Choose and Book was linked to a specific national policy, described in detail by others (Coulter, 2010; Dixon, Robertson, & Bal, 2010; Jones & Mays, 2009). The first gov- ernment commitment to providing patients with a choice of time and date of their hospital appointment was in 2001 with the Labour government’s landmark NHS Plan. Choice of hospital was promised T. Greenhalgh et al. / Social Science & Medicine 104 (2014) 210e219212 a year later. In 2004, The NHS Improvement Plan promised all pa- tients a choice of hospital at the point of referral. It predicted that introduction of ‘choice’ would reduce waiting times, make the service more responsive to patients’ needs, promote quality improvement and increase efficiency (and therefore cost effec- tiveness). The plan sought to introduce competition between pro- viders via a new reimbursement system that paid hospitals a fixed tariff price per patient seen. From January 2006 all National Health Service (NHS) patients referred to hospital for elective care were to be offered a choice of four or five ‘clinically appropriate’ local providers. In April 2008, patients became eligible to choose any provider nationally who offered care at the national tariff rate and met standards set by the Care Quality Commission (the govern- ment’s independent quality regulator). The assumption underpinning the introduction of patient choice in referrals was that the option for patients to take their custom elsewhere (what Hirschman called ‘the power of exit’) was a significantly more effective quality driver than the possibility that they might complain (‘the power of voice’) e and indeed, that the potential for ‘exit’ added weight to ‘voice’ (Dixon et al., 2010; Hirschman, 1970). It is worth keeping in mind the abstract, decontextualized nature of these assumptions. Development of Choose and Book, intended to support patient choice at the point of referral in England, was funded in 2003 via a 5-year, £64.5 million contract to the commercial supplier ATOS. Its national implementation was the responsibility of a designated lead within the Department of Health. At the time of this study, its local implementation was formally the responsibility of primary care trusts (PCTs). It was anticipated that by replacing the tradi- tional paper referral with an ‘integrated’ electronic system, Choose and Book would also [a] be more convenient for patients and GPs; [b] reduce the number of referral-based enquiries GPs and their staff had to deal with; [c] lessen the bureaucracy associated with referrals; [d] reduce ‘did not attend’ (‘DNA’) and cancellation rates in outpatients departments; and [e] encourage a more standardised format for referrals (Department of Health, 2004). Recognising that ‘choice’ would be effective only if patients were informed of the key differences between local (and national) pro- viders in a particular service, in 2007 the Department of Health launched a website giving details of these services to allow com- parison between providers prior to referral (see ‘NHS Choices’, http://www.chooseandbook.nhs.uk/patients/choosing-your) and a ‘Choosing Your Hospital’ booklet (to be supplied to patients by the referring GP). This booklet emphasised patients’ right to choice and encouraged them to report services to their GP if they were not satisfied. Choose and Book was part of a wider socio-technical network. This included the National Programme for IT (especially the central Spine on which patients’ demographic details were held); the machinery of the New Labour government (with its neoliberal agenda to ‘modernise’ public services); the Care Quality Commis- sion and other national regulatory bodies; civil servants who created the performance metrics for choice (policy) and Choose and Book (technology), monitored performance of healthcare organi- sations against these (e.g. via league tables) and linked them to financial incentives; professional bodies (especially the British Medical Association); and local managers in PCTs. This socio-technical network was distinctly unstable during (and indeed after) our data collection period. The Department of Health continued to produce reports and electronic updates pur- porting that Choose and Book was improving ‘choice’ (Department of Health, 2008). These were countered by letters and articles published by doctors in academic journals that documented increased workload and a rise in ‘did not attend’ rates following the introduction of Choose and Book (Beckingsale & Wallace, 2009; Modayil, Hornigold, Glore, & Bowdler, 2009); patients referred under Choose and Book who had no recollection of being offered a choice of provider (Green et al., 2008); and a widespread percep- tion that the technology was inefficient, inflexible, complicated and politically-driven (Rabiei, Bath, Hutchinson, & Burke, 2009; Rashid, Abeysundra, Mohd-Isa, Khan, & Sismeiro, 2007). Late modernity and expert systems NHS policy changes in the 2000s reflected the mindset of late modernity, with its emphasis on an abstract blueprint for control that lacked grounding in, or sensitivity to, the details and variety of local contexts (Giddens, 1990). The predominant frame of reference was rationalist; there was a strong sense that innovation and change represented progress, and a particular confidence in the value of expert systems e defined by Giddens (1990, page 27) as “[a] system of technical accomplishment or professional expertise that organize[s] large areas of the material and social environments in which we live today”; the possible negative consequences of such technical systems, including their impact on social interaction, was rarely systematically considered; and designers and policymakers were orientated to an imagined ‘proximate future’ e a time almost upon us when the technology is fully functional and all technical, ethical and political challenges have been smoothed out (Dourish & Bell, 2011). The expert system is a relatively recent phenomenon, resulting from the powerful triad of classificatory systems, bureaucracy and information technology in the age of globalisation (Giddens, 1991; Stones, 2005a). Such systems, which now range far and wide (e.g. finance, energy, engineering, medicine), are driven by abstract rules and procedures designed to co-ordinate social relations across large distances. Giddens (1991) proposed that these expert systems, using technology to encode information and store formal knowledge, have an inherent tendency to ‘empty out’ the content of local in- teractions because the technical knowledge they contain is assumed to have validity independently of any particular interac- tion, and to have the authority to override situational contin- gencies. They are designed to exert control and order e measurable, quantifiable e over distance in a way that seeks to remove (or at least, radically attenuate) the ability of distinctive people, relations and contexts to upset the uniform application of the rules and classificatory system embedded in the system. There is a powerful momentum towards general and universalising rules and pro- cesses, and away from the application of practical wisdom (what Aristotle called phronesis) in specific contexts. Expert systems capture professional expertise by formalisation e deploying impersonal knowledge, classificatory systems and procedures to shape, monitor, standardise and render calculable the work they support. Anthropologist Mary Douglas, developing earlier insights from Durkheim, argued that producing lists, rank- ings and other classification systems helps establish and then sustain social institutions by introducing conventions that “describe the way things are” (Douglas, 1986, page 48). Classifica- tion systems are fiercely negotiated and defended for precisely this reason (Bowker & Star, 1999). They have long been combined with those bureaucratic forms of instrumental rationality carefully analysed by Weber (1978). It is the interweaving of these two systems with powerful information technologies that is new. The patient as ‘rational chooser’ and the ideology of competition The classificatory rules and procedures embedded in Choose and Book software and its networks assumed that the sick patient functioned primarily as a rational chooser, able and willing to http://www.chooseandbook.nhs.uk/patients/choosing-your T. Greenhalgh et al. / Social Science & Medicine 104 (2014) 210e219 213 weigh up information about potential options and decide between them if provided with high-quality information and decision sup- port. Managing illness was assumed to consist, more or less, of making a series of objective decisions based on a limited number of decontextualised indicators and then following through on these. It follows from these assumptions that provision of statistical infor- mation on the ‘quality’ of services in a standardised format (such as a table of hospital-level comparisons using star ratings on key metrics) will prompt the ‘right’ choices and that these choices will lead to the ‘best’ services winning out in a competitive market. Choose and Book was also influenced by the abstract ideology of competition, and by the salutary effects that the competition blueprint was said to have on cost efficiency, patient satisfaction and patient outcomes. When the policy of choice was introduced, much attention was paid to the expert system but there was little exploration of the meso- and micro-level social interactions and processes that would convert the policy idea (of hospitals competing to attract outpatient referrals) into the reality of a more efficient, effective and responsive healthcare system that improved patient satisfaction and outcomes. Notably, the over-riding influ- ence of national policy meant that Cherns’ principles of socio- technical design at local level were not recognised or applied (Cherns, 1987). With a view to redressing this imbalance, we sought to analyse the micro processes and interactions involved in the practice of referral using the Choose and Book technology. To this end, we used the theoretical lens of strong structuration theory introduced above. This focuses on actors who are sited within a field of posi- tionepractice relations that has a powerful presence external to them, and which imposes itself upon them in various ways. These external structures pose constraints, provide resources and possi- bilities for action, and are the source of pressures and forces, including those of socialisation and induction into cultural mean- ings and values. Strong structuration theory takes seriously the hermeneutic, interpretative frames of the actors, the ways these are built up over time, and the way these mediate perceptions of external reality. But it departs from many forms of social con- structionism in framing this in terms of the ways in which external structures are internalised within the interpretive frames of actors. Key to our current argument is the further division of these internal structures into two interacting aspects. First, an individual actor’s generalised dispositions, or habitus, which refers to durable and deeply socialised aspects of embodied skills, culture, moral values and principles, and so on, built up over time as an actor is exposed to, and interacts with, their social contexts. This provides the phenomenological perspective by which events in the world are framed and perceived. Second, the actor’s knowledge of the immediate strategic terrain of positionepractice relations facing them at any particular time, including knowledge of the potential functionality of technologies and the sense of how this fits with other aspects of the terrain. Such conjuncturally-specific knowledge may be informed and fine-grained, or (especially if imposed top- down) it may be ill-informed and broad-brush, risking unin- tended and unwanted consequences (Stones, 2005b). Internal structures are an important part of the capabilities of actors, drawn from or worked upon e and in compliance or resistance e by actors as they engage with the everyday flow of practices. Resistance as morally-driven human agency The emphasis of strong structuration theory on values and norms within the habitus of actors means that humans are viewed not primarily as rational actors, nodes in a network or members of a socio-technical system but as moral beings who have commit- ments, desires and values (both personal and professional). It views work e especially the work of doctors e not merely as a series of coordinated tasks but as having symbolic significance in society. As Sayer put it in the title of his book, “things matter to people” e objects, actions, experiences and relationships have personal and moral significance as well as economic or instrumental worth (Sayer, 2011). With this in mind, our analysis set out to explore the tensions between professional morals and values (inculcated by medical education, professional identity and professional communities of practice) on the one hand and the demands made on the GP or other actor in the here and now by the remote, disembedded expert system of Choose and Book on the other. In the language of strong structuration theory, this is a tension between key aspects of the GP’s value-dispositions (specifically, the enduring professional values that form part of an individual’s habitus) and his or her conjuncturally-specific (‘here and now’) knowledge of the social forces and sanctions embedded in the proximate structures of the Choose and Book technology. The nexus of ethical values embedded within the habitus of a healthcare professional is not static or unproblematic. Indeed, it may be variously ambivalent, fragmented or conflicting, reflecting the ethical tensions and inherent conflicts of healthcare practice (Mol, 2008; Schei, 2006). In this paper, we show how an under- standing of these values and how they inform professional notions of excellence are a useful point of departure for illuminating those practices that are framed as ‘resistance’. Professional values and medicine’s internal goods If resistance is to be investigated at the micro level in terms of what matters to human agents (what do they care about; what do they see as good or bad practice?), we need to consider what macro-level influences, shared among members of a professional community, shape these values and perceptions. MacIntyre depicted these influences as the ‘internal goods’ of a domain (Maclntyre, 1981, p. 216). The internal goods of medicine would include the Aristotelian virtues, along with the dispositions and capacities, that are valued by doctors and which they believe are necessary to sustain standards of excellence in their profession. It has long been argued by sociologists that because they bear a commitment to the refined knowledge, ethics and values of their specialised community, professionals act as a bulwark against the impersonal march of capitalist and bureaucratic forces (Parsons, 1964). More recently, French sociological theorist Luc Boltanski has called for policymakers to go beyond ‘neomanagerialism’ and engage with the moral and normative positions taken by in- dividuals and groups on particular issues, notably the ethically- motivated concerns of professionals and lobbyists (Boltanski, 2011). Medicine’s internal goods are clustered, broadly speaking, around the themes of caring, curing, and comforting, and are embedded in the formal and informal codes of practice of the medical, nursing and other related professions. In the analysis that follows, we use these internal goods as a benchmark against which to consider not merely the means by which a referral to hospital is made (traditional letter or Choose and Book technology) but also the ends that are in mind when it is made. A GP’s judgements about referral to hospital have traditionally been directed towards a range of ends such as access to restricted tests or procedures, specialist advice in diagnosis or treatment, confirmation that nothing has been missed, symbolic affirmation of a serious illness, and respite from a patient whose chronic incurable illness has become wearing. GP referrals are informed by [a] knowledge of the patient’s personal history (both medical and social), [b] knowledge of the workings of their own health system (including the various incentives, disin- centives and practicalities of different options), and [c] knowledge T. Greenhalgh et al. / Social Science & Medicine 104 (2014) 210e219214 of local social relations, including the character of local hospitals and the clinical interests and personal style of particular consul- tants. The ‘expert system’ character of Choose and Book militates against using such knowledge, placing constraints on the scope of professional judgements. A professional framing of medical work sees doctors as wielding their symbolic power with integrity and commitment with the patient’s best interests in mind. Patient empowerment notwith- standing, there are aspects of the unequal relation between doctors and patients whose legitimacy is socially conferred, due largely to the fact that illness makes people (in a range of ways) vulnerable and in need of society’s help (Schei, 2006). Referral decisions are not merely ‘rational’ (that is, based on the best available medical evidence) but also practical and ethical, asking whether this referral, to this specialist at this hospital, is the right thing to do (Montgomery, 2006, Sayer, 2011). As Dixon et al (2010) showed in both UK and Netherlands, patients’ choices do not follow the nar- row economic rationality that policymakers anticipated, but reflect practical and symbolic influences that are perceived to matter in particular circumstances (Dixon et al., 2010). Professional judgement, particularly in primary health care, relies on being rooted in the immediacies of context. As a strikingly top-down form of expert system, Choose and Book imposes ab- stract and generalised protocols that have limited capacity to take account of local circumstances and contingencies. As Boltanski’s critique highlights, this socio-technical network crystallises a ten- dency to ignore or dismiss the skills, concerns and situated judgements of professionals. This is especially troubling in healthcare, since medicine is inherently exception-filled (most cases differ in some way from the ‘textbook case’) and medicine’s internal goods are not, in large part, reducible to formulaic rules and protocols. Neither technologies, nor the policies and processes in which they become embedded, are morally neutral, and to be able to judge the appropriateness and adequacy of particular policy ini- tiatives and linked technologies, it is necessary to assess how well they allow patients to receive the levels of care, cure, comforts and so on they can reasonably expect from the healthcare system e and support doctors to provide them. Having a clear sense of ‘what good might look like’ allows us to begin to open up policies and their socio-technical networks to critical scrutiny. Research questions Our research questions were: [1] When referring patients to hospital, how does the tension between the systemic demands of Choose and Book as an expert system and the situated application of local knowledge through practical and professional judgement play out? [2] To what extent can ‘resistance’ to Choose and Book be explained in terms of the structure-agency dynamic (specifically, the dynamic between attempts to control behaviour from a dis- tance on the basis of abstract criteria and the local values, judge- ments and practices of GPs and their staff)? Methods The HERO study and the secondary dataset on referral The idea for this secondary analysis emerged during the Healthcare Electronic Records in Organisations (HERO) study, conducted in 2007e10 and funded by the Medical Research Council, which explored the use (and non-use) of electronic records in English general practice (Swinglehurst, Greenhalgh, Myall, & Russell, 2010). It occurred at a time when a number of networked technologies were being introduced as part of national IT policy. The original HERO dataset covered four practices (anonymised as Dale, Beech, Elm and Clover) and included around 200 clinical consultations either directly observed or videotaped (with screen capture of the electronic record), as well as ethnographic field notes, documents (emails, letters, business plans, protocols, prac- tice leaflets) and naturalistic interviews e that is, asking people what they were doing and why as part of ethnographic observation (column 1 of Table 1). Naturalistic interviews provide particularly useful data in the study of work, since people can best describe and reflect on their work when doing it (Barley & Kunda, 2001). Our original analysis produced findings relating to how the work of GPs (Swinglehurst, Roberts, & Greenhalgh, 2011), nurses (Swinglehurst, Greenhalgh, & Roberts, 2012) and receptionists (Swinglehurst, Greenhalgh, Russell, & Myall, 2011) is shaped and constrained by technologies in use and by prevailing expectations about who should use them and how. One technology in particular e Choose and Book e was rarely if ever used as its designers had intended. Its embedded scripts (the assumptions held by its de- signers about how people would use it (Akrich,1992)) were ignored and/or deliberately subverted. When it was used, the consequences were not as predicted. Most referrals were still dictated, typed and sent in the traditional way. Furthermore, GPs and administrators often had strong feelings (usually but not universally negative) towards Choose and Book and there was much talk of ‘clinician resistance’. We decided to seek funding for a secondary analysis of our dataset to explore these impressions further; we obtained this from the National Institute for Health Research Health Services and Delivery Research Programme (ref. no. 10/1011/01). Data analysis Commencing with the entire HERO dataset (over 800 type- written pages), we selected a much smaller dataset for further analysis (column 2 of Table 1). This mainly comprised direct ob- servations of work practices relating to referral, whether under- taken manually or with Choose and Book (hence, it afforded a ‘symmetrical’ analysis of both the use of Choose and Book and its non-use (Wyatt, 2003)). The extracted dataset also included some videotaped consultations, and documentation produced nationally and locally over the time period of the original study. In an initial familiarisation phase (column 3 of Table 1), we produced first-order interpretations in which we sought to describe and offer pre- liminary explanations of observed practice and to summarise the assumptions that underpinned policy and were embedded in the Choose and Book technology. In a further analytic phase (column 4 in Table 1), we used strong structuration theory, whose application to the use and non-use of ICTs has been described previously (Greenhalgh & Stones, 2010). We focused on the conjuncture e that is, a critical combination of events and circumstances in which the human agent draws on both habitus (i.e. their internal dispositions, beliefs, values, norms and so on) and knowledge of the here-and-now situation (i.e. their assessment of the particular strategic terrain and how they are expected to act within it), and is supported or constrained by the available technologies, to inform, execute and justify a particular course of action. We considered two kinds of conjuncture: clinical consultations in which outpatient referrals were initiated and administrative activities in which staff sought to follow through on such referrals. In both kinds, the actor (clinician or administrator) either used or chose not to use (or, sometimes, was prevented from using) Choose and Book to support the referral in particular ways. We considered how the actor’s habitus (for example, the doctor’s professional identity and code of practice, or the administrator’s efforts to do a good job) combined with their assessment of external Table 1 Overview of data structure and analysis. Original dataset from HERO study 2007e10 Raw data selected for secondary analysis First-order interpretations Higher-order theoretical categories Field notes from w200 directly observed clinical consultations. 25 observed consultations in which the GP offered or discussed a referral. ➢ Why GPs refer people to hospital ➢ What GPs and patients talk about when discussing whether and where to refer ➢ Why GPs do not use the Choose and Book software during consultations External social structures ➢ Political authority ➢ Medicine’s ‘internal goods’ ➢ Economic context Internal social structures (what actors ‘know’ and how they interpret the strategic terrain) ➢ GPs’ professional identity, values, morals ➢ Administrative staffs’ perceptions about ‘doing a good job’ ➢ Skills and techniques for using the technology Social structures inscribed in Choose and Book technology ➢ Restricted nature of choices ➢ Assumptions about GP’s role and purpose of referral ➢ Financial incentives, rewards, costs Dynamic interplay between all the above 54 videotaped clinician-patient consultations, including screen capture from computer. 6 video excerpts in which a GP offered a referral. Striking ‘silence’ in the data: no examples of GPs directly accessing Choose and Book with patient present. Field notes from directly observed administrative work by GPs (w40 h) and practice staff (w330 h), including naturalistic interviews: total 800 pages of typewritten notes. 1 case in which a GP attempted to make a Choose and Book referral without the patient present; 3 where referral letters were dictated with no such attempt. 58 cases of administrator processing referrals manually (n ¼ 12) or via Choose and Book (n ¼ 46). 45 examples of naturally occurring talk where referral was mentioned (13 GPs, 31 administrators, 1 practice manager) ➢ Material properties of Choose and Book and challenges of using it ➢ Knowledge, skills and workarounds used by staff to make Choose and Book ‘work’ ➢ Why even experienced staff sometimes find it impossible or unhelpful to use Choose and Book ➢ What GPs and practice staff care about when making a referral Field notes from practice meetings (w20 h). Notes from one meeting between supplier and practice staff (Elm). 15 national policy documents on electronic records. Information for professionals and public on these. National audits and surveys on electronic record use. 7 policy documents in which electronic referral was mentioned. Sections of ‘NHS Choices’ website relating to referral. Referral guidelines and DVD for GPs. Annual surveys on ‘choice’ and Choose and Book use. ➢ How and why national policymakers think ‘choice’ will improve quality of care ➢ Benefits anticipated from Choose and Book ➢ How much Choose and Book is actually used and how much ‘choice’ achieved >30 semi-structured interviews with local policymakers on use of electronic records. One (joint) interview with two primary care trust managers charged with implementing Choose and Book locally. ➢ How local managers think referral works ➢ Why they think GPs do not use Choose and Book ➢ How much ‘choice’ is achieved locally Local documentation on electronic record use. 4 practice protocols, 10 newsletters, one PCT patient choice survey, one letter from PCT to GPs. Correspondence and emails among clinicians in field sites on electronic record use. Email exchange among GPs and PCT leads (1 thread, 15 participants) including doctors from Beech and Clover practices. ➢ Advantages and limitations of choice policy as perceived by GPs ➢ Advantages and limitations of Choose and Book as perceived by GPs T. Greenhalgh et al. / Social Science & Medicine 104 (2014) 210e219 215 circumstances (including the incentives and reward systems linked to Choose and Book and their perception of what would happen if they did or did not use this technology). The purpose of this was to focus in detail on the extent to which, and the ways in which, actors felt enabled and constrained by the material properties and capa- bilities of Choose and Book (including the role assumptions, cate- gories, values and other social structures inscribed in the software). Development of theory and analysis of data occurred concur- rently, each feeding into the other and informed by interdisci- plinary discussions. Two authors (TG and DS) are medical doctors with an interest in the sociology of professional practice and clinical interaction; the third (RS) is a professor of sociology who has developed strong structuration theory, and is an acknowledged authority on the work of Giddens (Stones, 2005a). Main findings At the time our primary study began (2007), Choose and Book had been available for two years. Many practices had invested in training and additional staff and begun to use it for referrals but had subsequently reduced their use of it. Of the four GP practices we observed, anonymised as Dale, Beech, Elm and Clover, the per- centage of referrals being submitted via Choose and Book was re- ported to us as 50e60%, 0%, 25% and 80e90% respectively. We never saw a GP use Choose and Book directly during a consultation e even in Clover practice, which described itself as ‘top of the [local] league table’ for percentage of referrals made using the system. Our secondary analysis revealed four analytically distinct but empirically overlapping foci of active or passive resistance to adopting the scripts, implicit in the system design, that actors were required to follow if Choose and Book was to be a success (Akrich, 1992). Each focus highlights a different aspect of GPs’ (or admin- istrators’) refusal fully to comply, because they believed that Choose and Book threatened to subvert a dimension of valued professional commitments. The four foci of resistance were: to the policy of choice that Choose and Book symbolised and purported to deliver; to finding ways to accommodate the technology’s socio- material constraints and implications; to interference with doc- tors’ contextual judgements; and to adjusting dutifully to the altered social relations consequent on its use. We consider these below. In each case, we summarise and illustrate data from our direct observations of situated action, then consider the relevant internal structures of human actors (what actors valued and ‘knew’) in conjunction with what was manifest in the material properties of the technology-in-use, and also the corresponding external struc- tures (political influences and expert norms that were inscribed in the technology and/or shaped actors’ practices). Resistance to the policy of ‘choice’ One of our most consistent findings when observing GP-patient consultations was that choice of hospital was either not offered at all or was presented to the patient as an external requirement (something the GP “had” to do), with GPs often highlighting the T. Greenhalgh et al. / Social Science & Medicine 104 (2014) 210e219216 perceived absurdity of the situation by expressing humour or exasperation (“we’re supposed to offer you St Joan ’s [hospital 20 miles away] or Timbuktu” e GP, Beech practice). We observed a number of examples in which the offer of choice introduced a distinct note of confusion into an otherwise smooth conversation, since the patient could not understand why they were being given the option of travelling to a distant and unfamiliar hospital. Indeed, GPs appeared to invoke “the government” or “the computer” as a third party in an attempt to reduce this confusion. In all cases where choice of provider was offered, it was recorded on the electronic record using a distinct code that could later be used to audit the practice’s performance. Our informal discussions and naturalistic interviews with GPs suggested that this recurring pattern appeared to be driven by three things that GPs ‘knew’: first, that the overwhelming majority of patients wished to attend their local hospital (hence there was no genuine reason to offer them choice); second, that the government was mistaken in assuming that choice of hospital would act as an effective mechanism to promote competition and efficiency in the NHS (they described the policy as “pointless”, “political” e and even as “bollocks”); and third, that offering choice was linked to a financial incentive, embedded within the technology, for the practice (hence there was a perverse reason for offering it). Thus, GPs were ‘resisting’ the policy of choice by presenting it to patients as an absurd demand of the system, at odds with their judgement, and refraining from the active investment of energy that its design relied upon, while most were also ‘complying’ with it at a super- ficial, pragmatic level in order to gain the reward. GPs’ perceptions were, broadly speaking, borne out by our data. We did not encounter a single example of any patient choosing to go anywhere except their local hospital, and only one example of a member of staff who recalled (on a single occasion) such a choice being made. Neither did we encounter any examples of either doctors or patients seeking or using comparative performance data when considering their referral preferences. Tellingly, the capacity of the technology to generate ‘personalised’ lists of options depending on whether patients wished to choose by distance, car parking, food quality and so on was never instantiated. A facility for patients to access such data in their local library in the district where Dale and Elm practice were located had no takers in six months. Beech practice stopped using Choose and Book when financial incentives ceased. In terms of the wider social structures impacting on choice of hospital, our dataset included substantial evidence of attempts to lever political authority. Locally, PCT managers described the PCT as being “beaten up” by the Strategic Health Authority, which in turn was (they said) being “hammered”, “bashed” and “kicked” by the Department of Health. Monthly bulletins from the Department of Health reported on progress in implementing the policy, and annual large-scale National Patient Choice Surveys were commis- sioned by the Department of Health in an attempt to demonstrate that the technology had been instrumental in achieving the policy goal (Department of Health, 2010). In these, around half of responding individuals recalled being offered ‘choice’, but the response rate was very low. Resistance to the socio-materiality of Choose and Book The process of referral was severely constrained in real time by the material functionality of the Choose and Book technology, whose operation at the time of our data collection (2007e10) was cumbersome, unreliable and time-consuming. Our ethnographic observations confirmed estimates of practice staff that a Choose and Book referral took, on average, twice as long as a manual referral. They recounted numerous examples of the technology freezing, crashing, running slowly, failing to supply the necessary password for the patient, requiring manual data entry for some fields and (commonly) failing to identify a suitable appointment slot at the preferred (local) hospital. Choose and Book referrals were far from ‘paperless’. On the contrary, they generated large amounts of printed paper, including internal memos and request sheets, sticky notes, protocols, flow- sheets, instructions and passwords for patients and e in three participating practices e a paper ledger of all ‘paperless’ referrals sent (the fourth entered these manually onto an Excel spreadsheet). In this and other ways, Choose and Book was viewed by practice staff as worsening the service problems it had been introduced to solve (e.g. referral bureaucracy, high ‘did not attend’ rates) and as generating negative knock-on effects (especially, taking time away from patient care or generating potential security breaches). Administrative staff considered Choose and Book highly temperamental, and spoke of having to get to know the system through accumulated experience and trial and error. We observed many examples of staff helping one another across a shared office in this regard. They spoke of not trusting the electronic system (or the organisations and/or individuals with whom it connected), and of being unable to navigate the system comfortably even when highly experienced in using it. They spent considerable time on the telephone to a helpdesk or to their counterparts in the hospital service trying to over-ride or work around glitches in the system. An aspect of this sociomateriality was resistance tothe expense of the Choose and Book system. A few GPs in our sample identified positive aspects of Choose and Book (e.g. fewer queries to the prac- tice about the referral’s progress) but commented that the technol- ogy was a cumbersome and expensive way of achieving that goal. In this focus of resistance, the key external structures impacting on human actors were, on the one hand, the modernist ideal of a reliable, touch-of-a-button automated system and, on the other, the reality of technologies-in-use: invariably messy and (in numerous ways) less than ideally fit for purpose (Brown & Duguid, 2002; Dourish & Bell, 2011). Resistance to interference with the doctor’s contextual judgement “I suggest you go to Mr Z [local eye surgeon] because I have [same eye condition] too and he looks after my eyes” (field notes from observation of consultations, Dale practice). Our observations showed that when considering whether and where to refer a patient, the GP routinely drew on his or her per- sonal knowledge of that patient, both clinical and social, and of local services, including the scope of particular clinics in particular localities; the patient’s own history of being treated at a particular hospital; transport services and the patient’s ability and willingness to use these; the expertise and interests of local consultants; per- sonal experience of referring patients to that service previously; and even e as in the above quote e personal experience of being treated by particular consultants themselves. In the single example we observed of a GP attempting to use Choose and Book, he abandoned the attempt because he could not find a suitable service (the local one was not listed and he was un- sure whether clinics at other hospitals provided the particular procedure needed). GPs and administrative staff explained to us that services in other localities tended to be organised differently e for example, they called the ‘same’ clinic by a different name or sub- divided the work of the specialty in a different way, so a GP was typically very knowledgeable about a local hospital service but much less knowledgeable about comparable services in other localities. T. Greenhalgh et al. / Social Science & Medicine 104 (2014) 210e219 217 Importantly, the kind of knowledge the GP needed to select the best option for the patient (personal, relationally situated and pro- fessional knowledge such as the trust and regard which this GP held for a particular consultant and how that consultant was likely to manage this condition in this patient) was not the kind available on the ‘NHS Choices’ website (which provided formal, regularised knowledge such as scores that had been allocated by a distant third party for ‘food’, ‘parking’, ‘infection rates’, ‘mortality’ and so on, and which pertained to the hospital as a whole rather than to a particular service). The policy of offering choice of hospital assumed that in different localities, similar service models with similar names would be available for a limited menu of diseases or conditions, allowing the ‘best’ service to be selected easily using a dashboard of performance metrics. The reality was that patients invariably pre- sented not e or not merely e with a ‘textbook’ clinical condition but with a unique illness along with a unique set of comorbidities, personal priorities and social circumstances. The abstracted criteria embedded in the Choose and Book software and NHS Choices website were far less nuanced. Many GPs described how they gave up using Choose and Book because it rendered them unable to apply their knowledge and skills to obtain the best outcome for their patient. As one observed, “the choice is only of the crudest kind”. In terms of external structures, this locus of resistance reflected a wider mismatch between what we have called ‘medicine’s in- ternal goods’ (the emphasis on professionals drawing skillfully and with wisdom on personal and contextual knowledge to make practical, ethical judgements) and neoliberal policy (in which the choices offered are ‘rational’ but lack granularity). The pressure from policymakers on the medical profession to comply with a restricted taxonomy of readily classifiable disease states that map unproblematically to particular investigations or treatments re- flects the kind of technology-work mismatch described by Brown and Duguid (2002) in a range of work settings. In a more layered sociological analysis, such mismatch in relation to medical work is depicted as having political origins and been termed ‘conceptual commodification’: “External control over medical care requires something more than literal commodification. Rather, it requires conceptual commodifi- cation of the output of the medical labour process: that is, its conceptualization in a standardized manner. Such commodifica- tion facilitates control over the production of services, not just over the arrangements for their exchange.. The basic strategy of commodification is to establish a classification system into which unique cases can be grouped in order to provide a definition of medical output or workload.” (Harrison, 2009, page 190) Resistance to the altered social relations consequent on use of Choose and Book In the consultations we observed directly, most discussions about referral took a traditional format, with the GP suggesting a consultant and a course of action, and the patient accepting the suggestion. One reason why they did not use Choose and Book during consultations was a reluctance to take on what they viewed as a more technical (and less professional) role. As one GP put it on an email exchange about Choose and Book, “We seem to be moving away from curing, caring and comforting to robotic automata”. As with the other forms of resistance described above, this can be explained in terms of internal social structures e specifically, professional identity. GPs considered it their professional duty to recommend a clinically appropriate outpatient clinic, including any necessary dialogue with the patient about their needs and preferences. But they defined the technicalities of booking ap- pointments as outside their scope of practice and associated these with a loss of status and autonomy that was often deeply held and strongly expressed (as in the quote above). This resistance was played out at locality level, between GPs (largely opposed to Choose and Book) and PCT staff (who were responsible for implementing the technology locality-wide). PCT managers did not question the ends of Choose and Book but pre- supposed that it was fit for purpose, attributing its low uptake to Luddism and even “spite” (they alleged that GPs’ resistance to the system was to punish the PCT for financial cuts elsewhere). But when the PCT sent GPs a letter that spoke of “failure” against the “standard” of Choose and Book and described low uptake as a “threat to good quality care”, the GPs responded vociferously by challenging this definition of quality and the legitimacy of the metrics being applied. On the contrary, they claimed, they had abandoned Choose and Book because it was a threat to professional standards. Our observational data revealed that the tension between the professional and the technical was perceived by some administra- tive staff as well as by most doctors. One administrator in Clover practice, BN, told us she had decided to take early retirement as a direct result of Choose and Book. She associated professionalism in her role with qualities such as knowledge of the services available locally, and with the ‘family doctor’ relationship that was built between patients and particular staff through continuity of care: “The patients have always been my main concern here. I don’t know where patients are these days e lost under piles of paper and in the Choose and Book system”. BN was concerned that patients often phoned the practice because they did not understand instructions for booking their appointment. She bemoaned the introduction of a standard accompanying letter sent to patients with Choose and Book paperwork (which began ‘Dear Patient’ and was signed ‘Practice Secretary’) as impersonal (“I could be anyone”), and insisted on adding her own name and signature to it. But the new system discouraged such personal touches. As BN commented while doing a Choose and Book referral: “I need to save this [letter] in Choose and Book .now what I’m going to do in my capacity as ‘absolutely nothing’, I’m going to attach it..”. A few administrative staff, however, were positive about Choose and Book. The lead administrator at Clover practice, XY, for example, was a ‘super user’ of the system: skilled, confident and keen to help others learn it. She saw Choose and Book’s technical idiosyncracies as a challenge (“I won’t be beaten by it”) and felt that its complexity made her job more interesting (“not just mindless typing”). She took particular pride that the practice was out- performing all other practices locally for use of Choose and Book. When some GPs in the practice had advised her to “hold off a bit” on using the Choose and Book technology because of its questionable cost effectiveness, her response was “I can’t do my job 50%”. In terms of dispositional values, BN aligned strongly with the values of the traditional family doctor service (reflecting the wider social structure of ‘medicine’s internal goods’ described above). In contrast, XY could be viewed as having positioned herself as a bureaucratic cog within the expert system, reflecting the ‘new professionalism’ of what Harrison has called scientific-bureaucratic medicine (Harrison, 2009), overly detached from the professional values of the locally embedded general practice, and focused pri- marily on the efficiency of means rather than the value of the ends. Both national and local policymakers were characterised by a striking lack of engagement with the values, identities and re- lationships of general practice. The PCT managers we interviewed, for example, saw referral as the same administrative process whether achieved via Choose and Book or a traditional referral (“it’s T. Greenhalgh et al. / Social Science & Medicine 104 (2014) 210e219218 just two or three more mouse clicks”). This framing did not take ac- count of the wider changes in roles, responsibilities or identities associated with the Choose and Book system. Discussion This case study of referral to hospital in the English NHS in 2007e10 has revealed a contested social practice driven by national policy and linked to the use (and non-use) of a nationally mandated technology. The combination of strong structuration theory, Gid- dens’ conceptualisation of expert systems, and a hermeneutic and ethical sensitivity to professional values has allowed us to do the following. Firstly, we have theorised this phenomenon in relation to wider social changes in late modernity, as resistance to an expert system. Secondly, we have constructed an ideal typical conception of the professional values of those involved in general practice that articulates the moral bases for their resistance. Thirdly, we have explored the tensions between these value-dispositions and the specific forces and pressures introduced by the abstract system of Choose and Book. The various hierarchical orderings (e.g. lists, ratings and rank- ings) inscribed within Choose and Book and on the NHS Choices website created potential for policymakers to influence social re- lations and practices beyond immediate face-to-face interaction (‘action at a distance’). But expert systems can produce such action only to the extent that the people intended to use them actually do so. If they refuse (for example, because a professional engagement with the activity depends on embedded knowledge, or because it reallocates tasks in a way perceived as demeaning and/or irratio- nal), or are prevented from doing so (for example, when the system does not ‘work’ or its functionality does not fit the time or space constraints of the social situation), the intended action at a distance does not occur. Our findings reveal a mismatch between the model of clinical work underpinning the ‘choice’ policy and inscribed in the Choose and Book technology and the more complex, granular and exception-filled nature of real-world clinical practice. The choice policy pursued by the English Department of Health depicted clinical care in transactional rather than relationally situated terms: it harboured a model of GPs’ input as taking place within artificially bounded, unconnected episodes and in relation to overly simple scenarios. It also used the term ‘quality’ mainly in relation to discrete and abstractly conceived structures, processes and pro- cedures. A contextual and professional framing, in contrast, would emphasise the quality of relationships between patient and doctor and between GP and consultant (including such things as trust and positive regard) and the value of continuity of these relationships over time. A striking finding in this study was policymakers’ and managers’ limited understanding of the detail of clinical work and the knowledge that informs referral practice. It was assumed that GPs (who, like patients, were depicted as ‘rational choosers’) could be prompted to use the system through two behaviourist mecha- nisms: [a] financial incentives and [b] disclosure of performance data (‘naming and shaming’). Policymakers were either unaware of, or dismissed, the influence of institutional structures such as the norms of professional practice, which defined quality in ethical and relational terms rather than in terms of a state-imposed metric of compliance with a policy. They also under-estimated the extent to which the technology’s material properties would prove limiting. The framing by PCT staff of Choose and Book use as a ‘quality standard’, and their refusal to engage with the GPs’ concerns about threats to quality, is an example of the silencing effects that Boltanski (2011) writes about in criticising neo-managerialism’s concerted, ill-advised, constriction of the space for meaningful conversation and debate about the role of normative values in guiding policy. The managers’ perspective reflects a situated frame of meaning, in which their role is defined in such a way that they deal solely with implementing means. Theirs is a bureaucratic form of professionalism, which entails a refusal to question ends and the values that inform these (Weber, 1978). We conclude that overly top-down, abstracted approaches to reducing resistance to information technology are not the best way forward. Rather, resistance to such technologies and the expert systems of which they are part would be reduced if there was, firstly, a greater recognition and dialogue with the world of pro- fessional values within its design and implementation, and sec- ondly, a greater willingness to seek degrees of balance between such virtual, remote, systems and the exigencies of the local sites in which professional values are performed. Choose and Book is one of many expert systems being introduced, top down, in the English NHS. It is surely time for academics and policymakers to heed Boltanski’s call to open up debate with a view to acknowledging the tension between normative values and forms of order and au- thority (Boltanski, 2011) (page 155). While coming from a different theoretical perspective, such an approach would align with socio- technical theorists’ longstanding call for technologies to support rather than over-ride the micro-detail of professional work (Cherns, 1987). Acknowledgements We thank the study participants, the members of the HERO-2 steering group, and three peer reviewers for helpful comments on earlier drafts of this paper. This secondary analysis was funded by the National Institute for Health Research (call no 10/1011/01). The original empirical studies were funded by a research grant from the National Institute for Health Research via the Connecting for Health Evaluation Programme (ref CFHEP002 and 007) and a PhD Fellowship for DS, and also by the Medical Research Council (ref 07/133). References Akrich, M. (1992). The description of technical objects. In Shaping technology/ building society (pp. 205e224). Barley, S. R., & Kunda, G. (2001). Bringing work back in. Organization Science, 12, 76e95. Beckingsale, T. B., & Wallace, I. W. (2009). Orthopaedic attendance also worse with choose and book. BMJ, 338, 674e675. Boltanski, L. (2011). On critique. Cambridge: Polity Press. Bowker, G. C., & Star, S. L. (1999). Sorting things out: Classification and its conse- quences. Cambridge, MA: The MIT Press. Brown, J. S., & Duguid, P. (2002). The social life of information. Harvard Business Press. Cherns, A. (1987). Principles of sociotechnical design revisited. Human Relations, 40, 153e161. Coulter, A. (2010). Do patients want a choice and does it work? BMJ, 341, c4989. Department of Health. (2004). Choose and book e Patient’s choice of hospital and booked appointment: Policy framework for choice and booking at point of referral. London: Stationery Office. Department of Health. (2008). Report on the National Patient Choice Survey e September 2007. London. Department of Health. (2010). Report on the National Patient Choice Survey e February 2010. London: Stationery Office. Department of Health. (2012). The power of information: Putting all of us in control of the health and care information we need. London: Stationery Office. Dixon, A., Robertson, R., & Bal, R. (2010). The experience of implementing choice at point of referral: a comparison of the Netherlands and England. Health Eco- nomics, Policy and Law, 5, 295e317. Douglas, M. (1986). How institutions think. Syracuse: Syracuse University Press. Dourish, P., & Bell, G. (2011). Divining a digital future: Mess and mythology in ubiq- uitous computing. Cambridge, MA: MIT Press. Giddens, A. (1984). The constitution of society. Berkeley: University of California Press. Giddens, A. (1990). The consequences of modernity. Cambridge: Polity Press. Giddens, A. (1991). Modernity and self-identity: Self and society in the late modern age. Cambridge: Polity Press. http://refhub.elsevier.com/S0277-9536(13)00690-4/sref1 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref1 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref1 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref2 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref2 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref2 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref3 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref3 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref3 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref4 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref5 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref5 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref6 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref6 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref7 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref7 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref7 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref8 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref9 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref9 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref9 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref9 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref10 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref10 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref11 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref11 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref12 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref12 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref13 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref13 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref13 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref13 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref14 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref15 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref15 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref16 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref16 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref17 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref18 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref18 T. Greenhalgh et al. / Social Science & Medicine 104 (2014) 210e219 219 Green, J., Mcdowall, Z., & Potts, H. (2008). Does Choose and Book fail to deliver the expected choice to patients? A survey of patients’ experience of outpa- tient appointment booking. BMC Medical Informatics and Decision Making, 8, 36e38. Greenhalgh, T., Morris, L. M., Wyatt, J. C., & Thomas, G. (2012). Lessons learned from implementation of nationally shared electronic patient records in England, Scotland, Wales and Northern Ireland. The Health Service Journal, 28e37. Greenhalgh, T., & Stones, R. (2010). Theorising big IT programmes in healthcare: strong structuration theory meets actor-network theory. Social Science & Medicine, 70, 1285e1294. Greenhalgh, T., Stramer, K., Bratan, T., Byrne, E., Russell, J., & Potts, H. W. W. (2010). Adoption and non-adoption of a shared electronic summary record in England: a mixed-method case study. BMJ, 340, c3111. Harrison, S. (2009). Co-optation, commodification and the medical model: gov- erning UK medicine since 1991. Public Administration, 87, 184e197. Hirschman, A. O. (1970). Exit, voice, and loyalty: Responses to decline in firms, orga- nizations, and states. Harvard university press. Jones, L., & Mays, N. (2009). Systematic review of the impact of patient choice of provider in the English NHS. London: Health Services Research Unit, Department of Public Health and Policy, London School of Hygiene and Tropical Medicine. Lorenzi, N. M., & Riley, R. T. (2000). Managing change: an overview. Journal of the American Medical Informatics Association, 7, 116e124. Maclntyre, A. (1981). After virtue: A study in moral theory. London: Duckworth. Modayil, P. C., Hornigold, R., Glore, R. J., & Bowdler, D. A. (2009). Patients’ attendance at clinics is worse with choose and book. BMJ, 338, b396. Mol, A. (2008). The logic of care: Health and the problem of patient choice. London: Routledge. Montgomery, K. (2006). How doctors think: Clinical judgement and the practice of medicine. Oxford: Oxford University Press. Parsons, T. (1964). The professions and social structure. In Essays in sociological theory: pure and applied. New York: Free Press. Rabiei, R., Bath, P. A., Hutchinson, A., & Burke, D. (2009). The National Programme for IT in England: clinicians’ views on the impact of the Choose and Book ser- vice. Health Informatics Journal, 15, 167e178. Rashid, M., Abeysundra, L., Mohd-Isa, A., Khan, Y., & Sismeiro, C. (2007). Two years and £196 million later: where is Choose and Book? Informatics in Primary Care, 15, 111e119. Robertson, A., Cresswell, K., Takian, A., Petrakaki, D., Crowe, S., Cornford, T., et al. (2010). Implementation and adoption of nationwide electronic health records in secondary care in England: qualitative analysis of interim results from a prospective national evaluation. BMJ, 341, c4564. Sanders, C., Rogers, A., Bowen, R., Bower, P., Hirani, S., Cartwright, M., et al. (2012). Exploring barriers to participation and adoption of telehealth and telecare within the Whole System Demonstrator trial: a qualitative study. BMC Health Services Research, 12, 220. Sayer, A. (2011). Why things matter to people. Cambridge: Cambridge University Press. Schei, E. (2006). Doctoring as leadership: the power to heal. Perspectives in Biology and Medicine, 49, 393e406. Stones, R. (2005a). Anthony Giddens. In G. Ritzer (Ed.), Encyclopedia of social theory. Thousand Oaks, California: Sage. Stones, R. (2005b). Structuration theory. Basingstoke: Palgrave-Macmillan. Swinglehurst, D., Greenhalgh, T., Myall, M., & Russell, J. (2010). Ethnographic study of ICT-supported collaborative work routines in general practice. BMC Health Services Research, 10, 348. Swinglehurst, D., Greenhalgh, T., & Roberts, C. (2012). Computer templates in chronic disease management: ethnographic case study in general practice. BMJ Open, 2. Swinglehurst, D., Greenhalgh, T., Russell, J., & Myall, M. (2011). Receptionist input to quality and safety in repeat prescribing in UK general practice: ethnographic case study. BMJ, 343, d6788. Swinglehurst, D., Roberts, C., & Greenhalgh, T. (2011). Opening up the ‘black box’ of the electronic patient record: a linguistic ethnographic study in general prac- tice. Communication and Medicine, 8, 3e15. Weber, M. (1978). Economy and society, 2 Vols.. Los Angeles: University of California Press. Wyatt, S. (2003). Non-users also matter: the construction of users and non-users of the Internet. In N. Oudshoorn, & T. P (Eds.), How users matter: The co- construction of users and technology. London: MIT Press. http://refhub.elsevier.com/S0277-9536(13)00690-4/sref19 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref19 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref19 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref19 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref19 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref20 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref20 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref20 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref20 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref21 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref21 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref21 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref21 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref22 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref22 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref22 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref23 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref23 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref23 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref24 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref24 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref25 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref25 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref25 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref26 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref26 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref26 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref27 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref28 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref28 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref29 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref29 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref30 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref30 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref31 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref31 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref32 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref32 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref32 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref32 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref33 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref33 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref33 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref33 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref33 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref34 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref34 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref34 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref34 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref35 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref35 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref35 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref35 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref36 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref37 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref37 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref37 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref38 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref38 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref39 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref40 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref40 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref40 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref41 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref41 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref41 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref42 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref42 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref42 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref43 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref43 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref43 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref43 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref44 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref44 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref45 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref45 http://refhub.elsevier.com/S0277-9536(13)00690-4/sref45 Choose and Book: A sociological analysis of ‘resistance’ to an expert system Introduction ‘Resistance’ to information technology in healthcare The policy context Late modernity and expert systems The patient as ‘rational chooser’ and the ideology of competition Resistance as morally-driven human agency Professional values and medicine's internal goods Research questions Methods The HERO study and the secondary dataset on referral Data analysis Main findings Resistance to the policy of ‘choice’ Resistance to the socio-materiality of Choose and Book Resistance to interference with the doctor's contextual judgement Resistance to the altered social relations consequent on use of Choose and Book Discussion Acknowledgements References 
work_xhcrfzwowzhmfhswpyutuwpriq	----	Mining product maps for new product development Mining product maps for new product development Shu-Hsien Liao *, Chia-Lin Hsieh, Sui-Ping Huang Department of Management Sciences and Decision Making, Tamkang University, No. 151, Yingjuan Road, Danshuei Jen, Taipei 251, Taiwan, ROC Abstract Many enterprises have been devoting a significant portion of their budget to new product development (NPD) in order to distinguish their products from those of their competitors, and to make them better fit the needs and wants of customers. Hence, businesses should develop products that fulfill the customer demands, since this will increase the enterprise’s competitiveness and it is an essential criterion to earning higher loyalties and profits. This paper presents the product map obtained from data mining results, which investigates the relationships among customer demands, product characteristics, and transaction records, using the Apriori algorithm as a methodology of association rules for data mining. The product map shows that different knowledge patterns and rules can be extracted from customers to develop new cosmetic products and possible marketing solutions. Accordingly, this paper suggests that the cosmetics industry should extract customer knowledge from the demand side and use this as a knowledge resource on its supply chain for new product development. � 2006 Elsevier Ltd. All rights reserved. Keywords: New product development; Product map; Data mining; Association rules; Knowledge extraction 1. Introduction It is well understood that new product development (NPD) is critical for long-term firm performance and com- petitive advantages. However, it is difficult to anticipate what certain customers will value and when changes in those preferences will emerge. A failure to anticipate these changes can force suppliers into a reactionary mode, where success is determined by how quickly one can respond to new desires as they emerge. Because new products take a substantial time to develop, suppliers should anticipate changes in what customers will value as early as possible (Flint, 2002). Therefore, a strong market orientation and customer knowledge competence are vital to the success of new products (Bonner, 2005). Most of the businesses involved in the production chain such as the manufactur- ers, suppliers and retailers are aware of the importance and need for enterprises to acquire and share better knowl- edge of their customers. But this is easier said than done since customers’ preferences are concealed in the custom- ers. It is available but not easily accessible, and there is lit- tle possibility to explore the full volume of data that should be collected for its potential value. Inefficient utilization can render the data collected useless, causing databases to become ‘data dumps’ (Keim, Pansea, Sipsa, & Northb, 2004). Therefore, how to effectively process and use cus- tomer data is becoming increasingly important. This calls for new techniques to help analyze, understand or even visualize the huge amounts of stored data gathered from business and scientific applications (Liao & Chen, 2004). Among the new techniques developed, data mining is the process of discovering significant knowledge, such as pat- terns, associations, changes, anomalies and significant structures, from large amounts of data stored in databases, data warehouses, or other information repositories (Hui & Jha, 2000; Keim et al., 2004). In addition, there are many data mining models in the literature such as classification, estimation, predictive modeling, clustering/segmentation, 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.08.027 * Corresponding author. E-mail address: michael@mail.tku.edu.tw (S.-H. Liao). www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 50–62 Expert Systems with Applications mailto:michael@mail.tku.edu.tw affinity grouping or association rules, description and visu- alization, as well as sequential modeling. Similarly, there are also many application methods, including association rules, sequential patterns, grouping analysis, classification analysis and probability heuristic analysis (Fan, Lu, Mad- nick, & Cheung, 2002; Hui & Jha, 2000; Goodwin, Dyne, Lin, & Talbert, 2003; Mehta & Bhattacharyya, 2004; Musaev, 2004; Tsechansky, Pliskin, Rabinowitz, & Porath, 1999). Knowledge of customers extracted through data mining can be integrated with product and market- ing knowledge from research and then provided to both upstream suppliers and downstream retailers. Thus it can serve as an important approach for new product develop- ment. When effectively utilized, such knowledge extraction can enable enterprises to gain a competitive edge through production of customer-oriented goods that provide better satisfaction to consumers and by speeding up the product development process (Menon & Tong, 2005; Shaw, Sub- ramaniam, & Tan, 2001). Map display is a powerful tool with the ability to con- vey a large amount of information in a limited space, and it also provides an interactive tool to allow the user to interact with the underlying information (Lin, 1997). Thus, the mapping approach, which focuses on the use of IT, can be used as a tool to support new product development. Holmlund and Strandvik (1999) proposed perception configuration as a new concept, and intro- duced configuration maps as tools for analyzing percep- tions in business relationship studies. Tülin and Russell (1998) presented market maps with a probabilistic spatial panel data model that allows the positions of products sharing the same name to be correlated across product categories. In a business setting, the combination of per- ceptions by two parties (such as buyers and sellers) can be represented as a perception configuration. All the per- ceptions are depicted on the horizontal and vertical axes of the map. This map can be used to capture both the composition and the dynamics of perception configura- tions, and it is generically applicable to dyadic perception studies. Daniel, Wilson, and McDonald (2003) utilized a marketing map to represent the best practice in market- ing and also used the process map to understand how IT can be deployed in order to support a marketing information system. The map mainly illustrates the links between various stages of the marketing process. Based on this idea, this research implements a product map to illustrate that new product development is essentially the function that matches the enterprise’s offers to cus- tomers’ demands. This paper investigates the following research issues in the development of new cosmetic products: What exactly are the customers’ ‘‘needs’’ and ‘‘wants’’ for cosmetic prod- ucts? Can product design and planning for product lines/ product collection be integrated with the knowledge of cus- tomers? Can the knowledge of customers be transformed into knowledge assets of the enterprises during the stage of new product development (NPD)? To investigate these research issues, the Apriori algorithm is a methodology of association rules for data mining, which is implemented to mine customer knowledge. Knowledge extracted from data mining results is illustrated as knowledge patterns and rules on a product map in order to propose possible suggestions and solutions for NPD and marketing. The rest of this paper is organized as follows. In Section 2, we pres- ent research design, which discusses research problems and introduces the proposed data mining system, including sys- tem framework, relational database design, and physical database design. Section 3 presents the data mining pro- cess, including the Apriori algorithm, and the knowledge extraction process. Section 4 illustrates results analysis and a product map for NPD. Discussions and future works are presented in Section 5; and Section 6 presents a brief conclusion. 2. Research design 2.1. Research problems Firstly, as the perfume, cosmetic, and toiletry prepara- tions industry entered the 1990s, it faced many challenges including regulatory changes, product safety concerns, calls for scientific data to document product claims, increasing environmentalism, desire for natural ingredients, pressure from the growing animal rights movement, as well as economy and market channels for product distribution (Kumar, 2005). Lawmakers in USA began investigating possible revisions to the traditional ‘drug’ and ‘cosmetic’ definitions established under the Food, Drug and Cosmetic Act. Since the distinction between cosmetics and drugs was sometimes vague. According to FDA guidelines (FDA, 2001), the term ‘cosmetic’ means (1) articles intended to be rubbed, poured, sprinkled, or sprayed on, introduced into, or otherwise applied to the human body or any part thereof for cleansing, beautifying, promoting attractive- ness, or altering the appearance, and (2) articles intended for use as a component of any such articles; except that such term shall not include soap. However, the products claiming to offer medical benefits or physiological effects were called over-the-counter (OTC) drugs. Examples of the latter included antiperspirants, sunscreen products, hair care products, shampoos to cure or remove dandruff, etc. Therefore, cosmetics product development is not only con- cerned with market reasons but also with safety concerns (Katušin-Razem, Mihaljevic, & Razem, 2003; Morohoshi et al., 2005; Pauwels & Rogiers, 2004). Secondly, cosmetic products are a kind of biomedical and truly personal prod- uct, which is popular with both woman and man on differ- ent age scales (Chang & Chang, 2003). Therefore, cosmetic products contain diversified product lines and collections in order to fit the needs and wants to specific customer seg- ments. This means that customer-orientation is the main concern for investigating market segmentation according to customers’ attributes, preferences, and personal factors. Thus, customers’ needs and wants are important resource S.-H. Liao et al. / Expert Systems with Applications 34 (2008) 50–62 51 https://isiarticles.com/article/2725 
work_xiz4awv5abeq3j7rnryxvcnjhq	----	Intelligent system architecture for process operation support Intelligent system architecture for process operation support M. Rao*, X. Sun, J. Feng Intelligence Engineering Laboratory, Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada T6G 2G6 Abstract This paper presents a generic intelligent system platform, namely INTEMOR (INTElligent Multimedia system for On-line Real-time application), for process operation support. INTEMOR is developed by combining arti®cial intelligence, computer system and information technology into a uni®ed environment. It integrates various function modules to perform operation support tasks, including communication gateway, data processing and analysis, on-line process monitoring and diagnosis, on-line operation manual, equipment maintenance assistance, reasoning system, knowledge-base creator and multimedia interface. The industrial applications of INTEMOR on a boiler system and a chemical pulping process are illustrated. q 2000 Elsevier Science Ltd. All rights reserved. Keywords: Intelligent system architecture; Process operation support; Arti®cial intelligence 1. Introduction Modern industrial systems have become more and more complex. Plant managers and operators have to deal with vast amount of raw data in production planning, mainte- nance scheduling and process operation. The data and infor- mation ªoverloadº may cause the operators confusion, leading to cognitive errors. It is true especially in time criti- cal situation, such as equipment failure. Operators may ®nd it dif®cult to quickly detect, diagnose and correct the fault. The growing complex of industrial processes and the prac- tical need for higher ef®ciency, greater ¯exibility, better product quality and lower cost have resulted in increasing requirements for enhanced operation and management support. The advent of computer technology has allowed us to implement more advanced process control and management systems such as process operation support (POS) system. An operation support system generally consists of on-line operation manual, fault diagnosis, equipment maintenance management and multimedia interface. Extensive research on the related topics, such as fault diagnosis expert systems (Kramer, 1991), intelligent monitoring systems (Murdock & Hayes-Roth, 1991) and knowledge-based maintenance systems (Berzonsky, 1990), have been done on chemical processes, electronic devices and mechanical equipment. The systems of such kind have shown signi®cant bene®ts to the industries. In the past decade, many different methods have been applied in developing operation support systems. Neural networks, because of their capability of learning complex and nonlinear relations, have attracted much attention in real-time data calibration, model identi®cation of poorly- understood or complex systems (Leonard & Kramer, 1993). Rule-based expert systems can be used in solving engineering problems that depend heavily on experts' experience. However, it is dif®cult to maintain and extend the knowledge-base (Vargas & Raj, 1993). Case-based reasoning (CBR) shows a great deal of promise for use in diagnostic systems (Gonzalez, Xu & Gupta, 1998; Stottler, 1994). CBR uses past problem-solving knowledge, includ- ing success or failure results, to ®nd a solution to new problems. It re¯ects domain experts' experience and natural problem solving. However, the CBR system is generally dif®cult to solve novel problems. Each individual method has its advantage in one situation, but has limitations in another. The integration of various methods will provide superior result by compensating the limitation of the indi- vidual methods. Integrated distributed intelligent system technology was proposed for the above purpose (Danielson, Bowes, Yang & Wang, 1995; Rao, 1991; Rao, Wang & Cha, 1993, 1996). It is intended to: (i) integrate various problem solving meth- ods, such as rule-based, model-based and CBR methods, as well as neural networks; (ii) integrate various types and levels of knowledge representation, such as integrating rule sets, past solved problems, process models and real- time data in an object-oriented environment; and (iii) inte- grate multiple problem solving tasks and application Expert Systems with Applications 19 (2000) 279±288PERGAMON Expert Systems with Applications 0957-4174/00/$ - see front matter q 2000 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 7 - 4 1 7 4 ( 0 0 ) 0 0 0 4 0 - 3 www.elsevier.com/locate/eswa * Corresponding author. Tel.: 11-780-492-8250; fax: 11-780-492-2881. E-mail address: ming.rao@ualberta.ca (M. Rao). systems, such as condition monitoring, fault diagnosis and maintenance support (Ursenbach, Wang & Rao, 1994). In this paper, an intelligent POS system platform, INTE- MOR (INTElligent Multimedia system for On-line Real- time application), is developed based on the integrated distributed intelligence technology. It integrates condition monitoring, fault diagnosis and analysis, information management, on-line manual and maintenance support in a uni®ed system environment. INTEMOR employs multiple knowledge representation, combines forward chain, back- ward chain and CBR mechanics, and runs in a real-time distributed environment. Industrial applications of INTE- MOR in a power plant and a chemical pulping process are presented. 2. System architecture In order to realize the intelligent on-line POS system, an integrated distribution system environment is required. The development of the system architecture is based on follow- ing ®ve requirements. First, the system is required to be able to monitor the data and evaluate its signi®cance according to experts' experience. Second, a troubleshooting or diagnostic system has to be developed to help operators. Third, an on- line operation manual is needed to assist operators in process operation. The fourth requirement is to implement a user-friendly interactive information system, with such utilities as on-line manuals and maintenance records. Finally, there is an impetus to store the knowledge and experience to prevent the loss that occurs if some experts leave the employment of the company. In a nutshell, it is desired to create a system that will make the work of the employees more ef®cient. As we know, collecting data, processing and analyzing data, and sending useful analyzing results and information back to process instruments and to operators and managers are the main tasks of control systems. In general, there are three levels of data analysis in industrial processes: data processing, information management and knowledge inte- gration. Data, information and knowledge are at different levels. Data are individual measurement of objects. Infor- mation represents the relationship among the correlated data. Knowledge describes connection among the structure information. The POS system emphasizes application and integration of knowledge. For the POS system, there exist two system architectures: information ¯ow architecture and physical logic architec- ture. Information ¯ow architecture shows the sequence of data analysis and the ¯ow direction of data and information. Physical logic architecture presents the logic connections of instruments, equipment and information. The proposed information ¯ow architecture for POS systems is shown in Fig. 1. In data processing method, raw real-time sensor data will ®rst undergo preprocessing techniques such as data calibra- tion, conversation and compression to extract signi®cant information and trends from the original mass of noisy real-time data. Data calibration and conversation processes involve ®ltering raw data, standardizing data and arranging data format. Data compression allows storage of only important and relevant information into a historical database for future reference. The functions of information manage- ment are to ®nd relationship among the correlated data to produce report forms, alarm information, production plan- ing, etc. Knowledge integration will ®nd deep causes of events, diagnose fault, ®nd the method to solve problem, extract expertise with knowledge-based techniques and then provide decision-making support for operator. The intelligent system for POS is internet-based system, whose physical logic architecture is demonstrated in Fig. 2. The POS system can access real-time process data through data gateway that communicate with distributed control M. Rao et al. / Expert Systems with Applications 19 (2000) 279±288280 Fig. 1. Generic information ¯ow architecture for POS systems. Fig. 2. Physical logic architecture for POS systems. https://isiarticles.com/article/5474 
work_xjfxdj37xfdyjcezj2nj3cnaki	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright http://www.elsevier.com/copyright Author's personal copy Ensemble based sensing anomaly detection in wireless sensor networks Daniel-Ioan Curiac ⇑, Constantin Volosencu Automation and Applied Informatics Department, ‘‘Politehnica’’ University of Timisoara, Bd. V. Parvan nr. 2, 300223 Timisoara, Romania a r t i c l e i n f o Keywords: Wireless sensor networks Binary classifier Ensemble based systems Sensors anomalies Data accuracy a b s t r a c t Wireless sensor networks are often used to monitor and measure physical characteristics from remote and sometimes hostile environments. In these circumstances the sensing data accuracy is a crucial attri- bute for the way these applications complete their objectives, requiring efficient solutions to discover sensor anomalies. Such solutions are hard to be found mainly because the intricate defining of the correct sensor behavior in a complex and dynamic environment. This paper tackles the sensing anomaly detec- tion from a new perspective by modeling the correct operation of sensors not by one, but by five different dynamical models, acting synergically to provide a reliable solution. Our methodology relies on an ensemble based system composed of a set of diverse binary classifiers, adequately selected, to implement a complex decisional system on network base station. Moreover, every time a sensing anomaly is discov- ered, our ensemble offers a reliable estimation to replace the erroneous measurement provided by sensor. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction A wireless sensor network (WSN) is a collection of tiny, inex- pensive and low-power devices that can be deployed throughout a geographical space for fine-grained monitoring and event detec- tion. Besides its computing and communication potential, a WSN is, first of all, an advanced distributed measurement system which is often prone to sensing anomalies that can cause erroneous data, compromising the objectives of the entire network. There are three major sources of sensor anomalies within WSN: (i) software or/and hardware failures – breakdowns in any subsys- tem of a sensor node or even battery discharging can produce wrong sensing data; (ii) security attacks – when a malicious entity compels the sensors to report erroneous measurements or to drop measurement data packages (Walters, Liang, Shi, & Chaudhary, 2006); (iii) environment related sources – when sensor nodes can- not measure the physical value correctly due to unfavorable phe- nomena that can arise in the harsh environments in which they are often deployed. Generally speaking, sensing anomaly detection refers to the problem of discovering patterns in measurement data that do not match with expected behavior (Chandola, Banerjee, & Kumar, 2009; Rajasegarar, Leckie, & Palansiwami, 2008; An, Heo, & Chang, 2011). This is not a simple task mainly because an estimated model of ‘‘correct behavior’’ is always hard to find. In the case of a WSN there is an inherent feature on which we can rely – sensing redun- dancy (Curiac, Volosencu, Pescaru, Jurca, & Doboli, 2009), which takes two basic forms: physical redundancy that implies the use of more than one sensor node for measuring the same localized va- lue; and analytical redundancy that implies a mathematical model for evaluating the value provided by one sensor by taking into con- sideration the past and present values of neighboring sensors (spa- tial redundancy), the past values of the sensor itself (temporal redundancy) or both (spatiotemporal redundancy). In the last decade a series of relevant approaches based on assortments employing different types of analytical redundancies and intelligent detection algorithms have been proposed for solv- ing the issue of sensing anomaly discovery. In (Siripanadorn, Hattagam, & Teaumroong, 2010) a mixture between a competitive learning method called the self-organizing map (SOM) and the discrete wavelet transform (DWT) is used to detect anomalies from synthetic and real-world datasets. A two-step temporal modeling procedure, developed to analyze and extract semantic symbols from a sequence of observations, is presented in Li, Thomason, and Parker (2010) where an intelligent system detects time-related changes online by using a likelihood- ratio detection scheme. The algorithm is distributed, and supports a hierarchical learning structure that can scale to large number of sensors. The use of Bayesian networks as means for unsupervised learn- ing and anomaly detection in gas monitoring sensor networks for underground coal mines is described in Wang, Lizier, Obst, Prokopenko, and Wang (2008). The authors showed that the Bayesian network model can learn cyclical baselines for gas con- centrations, thus reducing false alarms usually caused by flatline thresholds. Their solution was proved to be efficient in both distributed and centralized approach. 0957-4174/$ - see front matter � 2012 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2012.02.036 ⇑ Corresponding author. Tel.: +40 256 403 227; fax: +40 256 403 214. E-mail addresses: daniel.curiac@aut.upt.ro (D.-I. Curiac), constantin.volosen- cu@aut.upt.ro (C. Volosencu). Expert Systems with Applications 39 (2012) 9087–9096 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy An efficient method applying principal component analysis (PCA) simultaneously on multiple metrics received from various sensors is depicted in Chatzigiannakis and Papavassiliou (2007). One of the key features of this approach is that it provides an inte- grated methodology of taking into consideration and combining effectively correlated sensor data, in a distributed fashion. Further- more, it allows the integration of results from neighboring network areas to detect correlated anomalies that involve multiple groups of nodes. Our paper tackles the sensing anomaly discovery from a new perspective: the modeling of the correct behavior for sensors is done not by one, but by five different models, acting synergically to provide a reliable solution. For this, we developed an ensemble based system (EBS) containing five different binary classifiers, each categorizing every network node as being accurate or erroneous, the final decision being taken by the entire ensemble using a voting procedure. It is broadly accepted that the overall efficiency of such committees of classifiers can occur only if there is diversity among its components (Polikar, 2006). In our proposal, the heterogeneity of classifiers is achieved by using various and carefully selected classifier architectures and different sets of input data. In order to completely solve the problem, our ensemble not only discovers sensing errors, but offers reliable estimations to replace the mea- surements affected by these anomalies. The remainder of the paper is organized as follows. In Section 2, we introduce the philosophy of our ensemble based sensing anomaly detection. Section 3 describes the architecture of each individual classifier, pointing out the way in which the correct sensing behavior is modeled and predicted, while Section 4 depicts the decisional block of the ensemble based on weighted voting algorithm. In Section 5, we present the methodology used for train- ing and testing of the ensemble. Section 6 covers a test case illus- trating the entire methodology and, finally, conclusions are offered in Section 7. 2. Ensemble based sensing anomaly detection Discovering sensing anomalies in the context of WSN is a chal- lenging issue due to the complexity of the environment in which sensor nodes are deployed. Often, this subject is tackled using ded- icated decisional systems. For acquiring node behavior related decisions, it makes sense to ask more than one decision making en- tity, because this practice assures indubitably a better, more in- formed, and trustable final decision. We label these decisional instances as classifiers or experts and their collections as ensemble based systems (Dietterich, 2000; Ho, Hull, & Srihari, 1994; Polikar, 2006; Wang, Hao, Ma, & Jiang, 2011). In order to periodically detect and investigate each and every sensor anomaly, an ensemble based system has been designed. As presented in Fig. 1, this ensemble encloses several binary classifiers that independently categorize the state of each sensor as ‘‘reliable’’ or ‘‘unreliable’’. All the classifier outputs will be aggre- gated using a weighted majority algorithm to obtain the conclud- ing ensemble decision. This final ensemble decision will be further used by the base station to take all the required actions for the unreliable nodes (e.g. removal of the untrustworthy nodes from the WSN for a specified period of time). Our ensemble based methodology reveals the following characteristics: – The EBS input data are represented by past and present mea- surements gathered by the node under investigation (node A) and respectively, past and present measurements gathered by each of the node A neighboring nodes. – There are five binary classifiers involving different prediction techniques and different types of inputs to assure the required diversity of classifiers. Each classifier offers its own hypothesis hi (‘‘reliable’’ or ‘‘unreliable’’) about the sensing correctness of the node under investigation. – The overall ensemble decision block is based on a weighted majority algorithm. – If the investigated node is found as presenting sensor anoma- lies, the network base station acts in consequence and can exclude that specific sensor from the list of network functioning sensors for a limited period of time. As an example, this can be achieved based on the following rule: if the EBS indicated at least three times that the node A suffers from a sensor anomaly, the base station decides to inactivate the sensor. The base sta- tion could later reactivate the sensor after repeating the EBS investigation by testing if new readings became appropriate. – Using the power of ensemble, a reliable estimation of the mea- surement affected by anomaly is automatically offered any time when needed. In the following paragraphs the design of each EBS component, together with details regarding training and testing of the ensem- ble are offered. 3. Designing the classifiers The design of individual classifiers to fulfill the EBS require- ments is not a simple task. This process has to be governed by one magic word: diversity. As a result, any stratagem for generat- ing the ensemble members must be focused on the ensemble’s het- erogeneity improvement. The diversity of classifiers may originate from three basic sources: a. classifier structure: the use of diverse classifier algorithms (Hsu & Srivastava, 2009; Tsoumakas, Katakis, & Vlahavas, 2004) can assure the required heterogeneity (diversity through structure); b. classifier internal parameters: by using different training datasets (García-Pedrajas & Ortiz-Boyer, 2009; Kim & Kang, 2010; Li & Sun, 2012; Shirai, Kudo, & Nakamura, 2009) or diverse initializations of the training algorithms, the classifi- ers having exactly the same structure may cover different regions of the classified workspace, assuring the heterogene- ity (diversity through parameters); and c. classifier inputs: the diversity may be assured by applying different inputs to identical classifiers (diversity through inputs); usually this channel to obtain diversity is irrelevant, but when there is an overabundance of data sources it can be a viable alternative.Fig. 1. Ensemble based sensing anomaly detection. 9088 D.-I. Curiac, C. Volosencu / Expert Systems with Applications 39 (2012) 9087–9096 Author's personal copy There are circumstances in which these three possibilities are combined resulting more complex ensembles, but applying met- rics to evaluate diversity in these cases may become very difficult. Due to the specificity of WSN applications we chose to assure the diversity of classifiers using a mixture between diversity through structure and diversity through inputs. From this perspec- tive we developed five classifiers that combine the two kinds of in- put data (measurement time series provided by the sensor under investigation and the measurements gathered from neighboring sensors), with linear or complex nonlinear models: – average based classifier; – autoregressive linear predictor based classifier; – neural network based classifier; – neural network autoregressive predictor based classifier; – Adaptive Neuro-Fuzzy Inference System (ANFIS) based classifier. All these classifiers are built based on a preset configuration that includes: � a prediction block – this block encapsulates a linear or nonlinear model of the ‘‘correct behavior’’ of the sensor and relies on the analytical redundancy feature of the WSNs. Manipulating mea- surement data often affected by errors gathered from a dynamic and complex environment, the predictor block parameters have to be very carefully chosen. By choosing different types of pre- diction algorithms, the required diversity can be assured. � an error computing block that compares the sensor current measurement with its predicted value; Practically, this block compares the actual behavior of the sensor under investigation with a modeled ‘‘correct behavior’’. � an internal decision block that takes the individual resolution of the classifier – correct or incorrect behavior, using a carefully chosen threshold as the limit between good and bad behavior. 3.1. Average based classifier The average based classifier (AVC) is computationally the sim- plest binary classifier used in our ensemble. Based on measure- ment gathered by the sensor under investigation at a specific moment in time hA(t) and on the measurements provided by k adjacent sensors at the same moment in time aggregated in the in- put vector hAVC(t), it provides a binary output: ‘‘0’’ meaning normal functioning of sensor A and ‘‘1’’ meaning abnormal functioning. hAVCðtÞ¼ ½h1ðtÞ h2ðtÞ . . . hkðtÞ�: ð1Þ This classifier exploits the analytical spatial redundancy feature of WSN that allows the estimation of the measurement gathered from sensor A using the measurement values provided by neigh- boring sensors. In this case, the estimation will be obtained using a simple averaging scheme. As can be seen in Fig. 2, the classifier’s internal structure consists of one average computation block, one error computation block and one internal decisional block. Classifier’s average computation block receives all present mea- surement values of each of the k neighbors and computes the aver- aged value as presented in (2): hA;AVCðtÞ¼ 1 k Xk i¼1 hiðtÞ: ð2Þ The average value hA,AVC(t) is subtracted from node’s current mea- surement value, resulting the error eAVC(t): eAVCðtÞ¼ hAðtÞ� hA;AVCðtÞ: ð3Þ The obtained error value is then evaluated by classifier’s deci- sional block for producing hAVC(t) hypothesis. This evaluation is performed by comparing the error value with a given threshold eAVC. If the absolute value of the error exceeds eAVC, the classifier outputs a hypothesis equal to ‘‘1’’, meaning that the measurement provided by the node A was classified as being erroneous: hAVCðtÞ¼ 0 if jeAVCðtÞj < eAVC 1 if jeAVCðtÞj P eAVC � : ð4Þ In the training process, there is only one AVC parameter that has to be established: the threshold eAVC. This is done based on tri- als, experience and intuition of the human expert which plays a central role. 3.2. Autoregressive linear predictor based classifier The autoregressive linear predictor based classifier (ALC) exploits the analytical temporal redundancy feature of WSN that allows the estimation of the measurement gathered from a specific sensor A using past measurements of the same sensor to supply a binary output: ‘‘0’’ meaning normal functioning of the sensor and ‘‘1’’ meaning abnormal functioning. In this case, the estimation will be obtained using an autore- gressive predictor based on a numerically robust variant of recur- sive least square (RLS) method Ljung & Soderstrom, 1987; Yang & Bohme, 1992. As can be seen in Fig. 3, the classifier’s internal struc- ture consists of one autoregressive prediction block, one error cal- culation block and one internal decisional block. Classifier’s autoregressive prediction block receives past mea- surement values provided by the sensor under investigation: hALCðtÞ¼ ½hAðt � 1Þ � � �hAðt � nÞ� ð5Þ and predicts its current value as shown in (6), where n is the auto- regression order, the parameters a1(t), a2(t), . . . an(t) are estimated using a robust recursive least square algorithm, and n(t) is consid- ered to be a white-noise perturbation hA;ALCðtÞ¼ a1ðtÞ � hAðt � 1Þþ �� �þ anðtÞ � hAðt � nÞþ nðtÞ: ð6Þ The predicted value hA,ALC(t) is subtracted from node’s current mea- surement value, obtaining the error eALC(t) as shown in (7): eALCðtÞ¼ hAðtÞ� hA;ALCðtÞ: ð7Þ The obtained error value is then evaluated by classifier’s internal decisional block for producing hALC(t) hypothesis in a similar man- ner as the one described by (4). If the error value exceeds the value Fig. 2. Average based classifier. Fig. 3. Autoregressive linear predictor based classifier. D.-I. Curiac, C. Volosencu / Expert Systems with Applications 39 (2012) 9087–9096 9089 Author's personal copy of a given threshold eALC, the classifier outputs a hypothesis equal to ‘‘1’’, meaning that the measurement provided by the node A was classified as being erroneous. The classifier has some internal parameters that have to be thoroughly established: the threshold eALC, the autoregression order n, and the initialization parameters of the RLS algorithm. The threshold eALC is set based on trials, experience and intuition of the human expert, the autoregression order is a positive integer with values among 3 and 6 to make a reasonable balance between precision and computational time, and initialization parameters of RLS are set according to prescriptions when no a priori information are available (Ljung, 2007). An important remark upon the use of ALC: the classifier can be used as a part of EBS only after some time steps (e.g. 2n time sam- ples) because the prediction offered by RLS needs some time until the parameter convergence (Ljung, 1999). 3.3. Neural network based classifier The neural network based classifier (NNC) processes the current value provided by the sensor under investigation and the current/ past values gathered from adjacent sensors to provide a binary out- put: ‘‘0’’ meaning normal functioning of the sensor and ‘‘1’’ mean- ing abnormal functioning. This classifier exploits the analytical spatial redundancy feature of WSN that allows the estimation of the measurement gathered from a specific sensor A using only measurements provided by neighboring nodes. In this case, the estimation will be obtained using a neural network predictor. By using a strong nonlinear model (neural network) to approximate the ‘‘normal’’ operation state of a sensor we can model more sophisticated situations than the linear ones (ALC or AVC). As can be seen in Fig. 4, the classifier’s internal structure con- sists of one feed-forward neural network, one error calculation block and one internal decision block. The neural network, based on its input vector that contains current/past measurements pro- vided by its k neighbors: hNNCðtÞ¼ ½h1ðtÞ � � �h1ðt � nÞ � � �hkðtÞ � � �hkðt � nÞ�; ð8Þ predicts the actual value gathered from sensor A – hA,NNC(t). This prediction is subtracted from the current measurement of the sen- sor under investigation hA(t) to obtain the error: eNNCðtÞ¼ hAðtÞ� hA;NNCðtÞ; ð9Þ which will be used by the internal decision block to obtain the clas- sifier’s hypothesis hNNC(t). Classifier’s decisional block functions the same as the one included in the average based classifier but uses another threshold: eNNC. The classifier has some internal parameters that have to be carefully established: the threshold eNNC, the neural network struc- ture (number of layers, number of neurons on each layer, etc.) and the neural network’s weights. The threshold eNNC and the structure of the neural network are set based on trials, experience and intu- ition of the human expert. The neural network training can be done in an automated fashion using a backpropagation algorithm (e.g. Levenberg–Marquardt). To make a reasonable compromise between complexity and computing time, we chose to use only measurements provided at last three moments in time: t, t � 1 and t � 2. This classifier can be used as a component in our EBS only after n time samples, this period being necessary to populate all compo- nents of the input vector hNNC(t) with real values. 3.4. Neural network autoregressive predictor based classifier The neural network autoregressive predictor based classifier (NNAC) processes the values provided only by the sensor under investigation to provide a binary output: ‘‘0’’ meaning normal functioning of the sensor and ‘‘1’’ meaning abnormal functioning. In fact this is an alternative to the autoregressive predictor based classifier that uses a strong nonlinear model (neural network) instead of a linear model. This classifier exploits the analytical tem- poral redundancy feature of WSN that allows the estimation of the measurement gathered from a specific sensor A using past mea- surements of the same sensor. In this case, the autoregressive pre- diction process is done by a neural network predictor. As can be seen in Fig. 5, the classifier’s internal structure is al- most identical with the one presented for NNC, with the difference that the input vector includes only past measurements obtained from sensor A: hNNACðtÞ¼ ½hAðt � 1ÞhAðt � 2Þ � � �hAðt � nÞ�: ð10Þ This will impose some changes in the neural network structure (e.g. number of neurons on input layer). In the case of NNAC, the internal parameters that have to be cautiously established are: the threshold eNNAC, the neural network structure and the neural network’s weights. The threshold eNNAC and the neural network structure (number of layers, number of neurons on each hidden layer) are set based on trials, knowledge and intuition of the human expert. The neural network weights can be obtained using an automated procedure based on a back- propagation algorithm. To make a rational compromise between complexity and computing time, we decided to use only measure- ments gathered at last four moments in time: t, t � 1, t � 2 and t � 3. This classifier can be used as a component in our EBS only after n time samples, this period being necessary to populate all compo- nents of the input vector hNNAC(t) with real values. 3.5. ANFIS based classifier The ANFIS based classifier (ANFISC) processes the past values provided by the sensor under investigation and current/past values provided by adjacent sensors to provide a binary output: ‘‘0’’ meaning normal functioning of the sensor and ‘‘1’’ meaning abnor- mal functioning. This classifier exploits the analytical spatiotempo- ral redundancy feature of WSN that allows the estimation of the measurement gathered from a specific sensor A using past mea- surements of the same sensor and measurements provided by neighbor nodes. By using a strong nonlinear model (ANFIS) to Fig. 4. Neural network based classifier. Fig. 5. Neural network autoregressive predictor based classifier. 9090 D.-I. Curiac, C. Volosencu / Expert Systems with Applications 39 (2012) 9087–9096 Author's personal copy approximate the ‘‘normal’’ operation state of a sensor we can mod- el very sophisticated situations. As can be seen in Fig. 6, the classifier’s internal structure con- sists of one ANFIS prediction block, one error calculation block and one internal decisional block. Adaptive Neuro-Fuzzy Inference System (ANFIS) was developed by Jang (1991, 1993) and is a hybrid neuro-fuzzy system that uses the Takagi–Sugeno-type fuzzy inference. Basically, ANFIS is a fuzzy inference system, framed in a feed-forward neural network. Hence, the advantages of a fuzzy system can be combined with a learning algorithm. ANFIS architecture consists of five layers of nodes: the first and the fourth layers consist of adaptive nodes implementing fuzzification and defuzzification (represented by squares in the ANFIS structure from Fig. 6), while the second, third and fifth layers consist of fixed nodes that are implementing the fuzzy rule (P), the normalization (N) and summation ( P ). The adaptive nodes are associated with their respective parameters, get duly updated with each subsequent iteration while the fixed nodes are devoid of any parameters. By including two intelligent approaches (neural networks and fuzzy reasoning), the ANFIS block can achieve good reasoning in quality and quantity. In other words we have fuzzy reasoning and network calculation. The ANFIS predictor, based on its input vector hANFISC(t) that contains past measurements obtained from sensor A and current/ past measurements provided by its k neighbors: hANFISCðtÞ¼ hAðt�1Þ���hAðt�nÞjh1ðtÞ���h1ðt�nÞj. . .jhkðtÞ���hkðt�nÞb c ð11Þ predicts the actual value gathered from sensor A – hA,ANFISC(t). This prediction is subtracted from the current measurement of the sen- sor under investigation hA(t) to obtain the error: eANFISCðtÞ¼ hAðtÞ� hA;ANFISCðtÞ ð12Þ which will be exploited by the internal decision block to achieve the classifier’s hypothesis hANFISC(t). Classifier’s decisional block oper- ates the same as the ones included in above mentioned classifiers (AVC, ALC, NNC and NNAC), but utilizes a different threshold: eANFISC. The classifier has some internal parameters that had to be care- fully determined: the threshold eANFISC and the ANFIS structure and parameters. The threshold eANFISC is set based on trials, knowledge and intuition of the human expert, while the ANFIS network parameters are obtained using a combination of the least-squares method and the backpropagation gradient descent method (Math- Works, 2002). The training process of this classifier is very computationally intensive. To make a reasonable compromise between model com- plexity and computing time, we chose to use only two measure- ments from each sensor: hANFISCðtÞ¼½hAðt�1ÞhAðt�2Þh1ðtÞh1ðt�1Þ. . .hkðtÞhkðt�1Þ�: ð13Þ The ANFIS based classifier can be used as a part of our EBS only after n time samples, this period being necessary to populate all compo- nents of the input vector hNNC(t) with real measurements. 4. Decision block The ensemble decision block is in charge of two tasks: (a) to provide a confident decision of ‘‘reliable’’ or ‘‘unreliable’’ regarding each and every sensor measurement; and, (b) to offer a trustwor- thy estimation of the measurement value when a value provided by the sensor is considered to be affected by an anomaly. Taking the ensemble final decision can be efficiently imple- mented using the weighted majority voting algorithm (Littlestone & Warmuth, 1994). The five classifiers provide a set of hypothesis – hAVC(t), hALC(t), hNNC(t), hNNAC(t), hANFISC(t) e {0, 1} – that are later aggregated using a weighted majority voting scheme formularized by the following equations: WMVðtÞ¼ X i wi � hiðtÞ ð14Þ with hi(t) e {hAVC(t), hALC(t), hNNC(t), hNNAC(t), hANFISC(t)} and P iwi = 1. The overall ensemble decision is computed using the following relation: hensemble ¼ roundðWMVðtÞÞ¼ 0 if WMVðtÞ < 0:5 1 if WMVðtÞ P 0:5 � ; ð15Þ where the function round(X) rounds X to the nearest integer. The weights wi related to each hypothesis are chosen based on simulations and experiments. In order to provide equal rights for components based on temporal redundancy and spatial redun- dancy we included the following restriction: wALC þwNNAC þ0:5�wANFISC ¼wAVC þwNNC þ0:5�wANFISC ¼0:5: ð16Þ In case a measurement provided by the sensor under investiga- tion is considered abnormal, the ensemble decision block offers a reliable estimation that can replace it. This value is obtained by averaging the estimated values of the measurement provided by individual classifiers: hA;trustðtÞ¼ 1 5 � ½hA;AVCðtÞþ hA;ALCðtÞþ hA;NNCðtÞþ hA;NNACðtÞ þ hA;ANFISCðtÞ�: ð17Þ In order to prevent incorrect classifying of the ensemble compo- nents based on autoregressive forecasts (ALC, NNAC and ANFISC), their input values hA(t � i), when affected by anomalies, have to be replaced by these trustworthy estimations hA,trust(t � i). The rea- son is simple: wrong data in inputs conduct to misclassification. 5. Training and testing the ensemble After setting the structure for each individual classifier and for the decision block, a complex training and testing process must be carried out to assure optimal sets of parameters (Fig. 7). This is done using a variety of automated/nonautomated procedures individualized for each ensemble component and relies on training datasets obtained from two sources: computer simulation and experimental deployments of the same type of WSNs. The training process is undoubtedly one of the most challenging tasks when designing an ensemble system. There are three speci- ficities that must be carefully taken into account in the training process of the proposed ensemble: (a) The impact of the sensor deployment: Three of the classifiers included in our ensemble use measurements provided by neighboring sensors. This characteristic implies that the sen- sor deployment (shape, distances between adjacent nodes, etc.) is a key aspect, affecting decisively the training datasets and, by this, the entire training process. Fig. 6. ANFIS based classifier. D.-I. Curiac, C. Volosencu / Expert Systems with Applications 39 (2012) 9087–9096 9091 Author's personal copy (b) The way training datasets can be obtained: The environment in which the WSN will be deployed is extremely complex and, as a result, hard to simulate and predict. The training datasets will have to approximate this intricate environment which cannot be done using only computer simulation or only experimental deployments. A mixture between the two possibilities is more feasible. (c) Assuring the classifiers diversity is a must: The training process must guarantee the diversity of classifiers as mentioned in Section 3. The diversity is estimated using proper metrics that are computed during validation of the training process. If the diversity requirements are not met the training must be redone. The training and testing methodology applied in the case of our ensemble is governed by the following steps: i. Choosing the sensor deployment shapes that will be considered: Due to the specificity of WSN deployment procedures (aerial scattering, manual deployment, etc.) and due to possible existence of non-reporting sensors, the geographical place- ment of the sensors in the field can conduct to a plethora of configurations. From this huge number, we must select a reasonable set that can approximately cover all the situa- tions. For each selected sensor deployment configuration, a set of parameters for ensemble components will be obtained through training. ii. Forming the training and testing datasets for each sensor deployment shape: In order to obtain consistent training and testing datasets, we must concentrate on two direc- tions: development of computer simulation programs that can model fairly accurate the environment; and, deployment of experimental WSNs to collect relevant data from real environments. iii. Training the ensemble components: all the parameters of each and every ensemble component will be obtained using auto- mated (Levenberg–Marquardt algorithm in case of neural networks, a combination of the least-squares method with the backpropagation gradient descent method in case of ANFIS) and nonautomated techniques. Obtaining the thresh- olds for each classifier or the weights for the decision block is based mainly on experience and intuition. iv. Testing the trained ensemble using testing datasets and ensem- ble evaluation based on proper diversity metrics: Each trained ensemble is validated using independent testing datasets (data not used in ensemble training) obtained through com- puter simulations or using experimental sensor deploy- ments in a real environment. In addition, the diversity of the classifiers inside each of the ensembles is computed using proper metrics. v. If the testing results do not match with the desired results or if the diversity of the classifiers is not appropriate, the training process will be redone. Internal parameters of the individual classifiers can be obtained using a variety of techniques as follows: the thresholds eAVC, eALC, eNNC, eNNAC and eANFISC are tuned based on trials, experience and intuition, the weights of the two neural networks implied in NNC and NNAC classifiers are computed using Levenberg–Marquardt algorithm and the ANFIS parameters can be calculated using a combination of the least-squares method with the backpropaga- tion gradient descent method. A good choice for implementing the training process is proved to be Matlab/Symulink package that offers powerful functions for simulation, neural networks and AN- FIS training, etc. After training, each ensemble is subject to a testing and valida- tion procedure, where new data are classified and diversity metrics (Q-statistic) are computed. If the results of this validation are not acceptable the training process must be performed once again. 5.1. The link between the training process and the sensor deployment in the field The training process depends decisively on the real sensor deployment. Because the ensemble contains classifiers that rely on the measurements provided by adjacent sensors, the in-field position of the sensors has a major impact in the way the ensemble has to be trained. In order to standardize the training procedure we chose to train the ensemble for a set of nine different deployment shapes, pre- sented in Fig. 8. This set covers a larger variety of deployments be- cause it envelops the shapes obtained through rotation or/and symmetrization. Basically, we have to obtain the parameters for each ensemble component corresponding to this set of deployments and to store them in a small Access-type database or in a text file that will be parsed any time when needed, on the base station. When the deployment of WSN in the field is finished, for each sensor node, a procedure to find the most appropriate shape will be started. This can be done using shape matching procedures sim- ilar to the ones presented in Cao, Lisani, Morel, Musé, and Sur (2008), Belongie, Malik, and Puzicha (2002), Huttenlocher, Klanderman, and Rucklidge (1993). When the closest shape of the deployment is obtained, the parameters for the related ensem- ble are loaded from the database. This way, the ensemble can work even when some sensors are not reporting or even when sensors are deployed nonuniformly (randomly) in the field. 5.2. Obtaining the training and testing datasets Sensor anomalies within WSNs are hard to be discovered mainly because the complexity of environment in which they are deployed. Our ensemble must envelope this complex environment in mathematical models contained in each classifier. These models have to reflect different facets of an intricate reality and, for this, they have to be trained using proper datasets obtained from two independent sources (Fig. 7): Fig. 7. Training and testing processes of the ensemble. 9092 D.-I. Curiac, C. Volosencu / Expert Systems with Applications 39 (2012) 9087–9096 Author's personal copy – Computer simulation: In principle, the environment in which the WSN will be deployed can be simulated to obtain relevant data for training and testing the classifiers, but this action is highly problematical due to the complexity of involved phenomena. For this reason, the computer simulations are error prone and cannot be the only source for training and testing datasets. – Experimental deployments: A relevant source of training and testing datasets is represented by experimental deployments of WSNs in similar environments. Although, the real environ- ment is hard to be replicated, data obtained from experimental deployments are relevant, but sometimes statistically inefficient. By combining information provided by these two sources, con- sistent training and testing datasets can be formed. 5.3. Metrics to estimate the diversity of classifiers In order to quantitatively evaluate the diversity of classifiers, several metrics have been defined (Kuncheva & Whitaker, 2003; Polikar, 2006). Among them, the pairwise metrics defined between two classifiers are the simplest. In our case, the ensemble incorpo- rates T = 5 classifiers, so there is a need to compute T(T � 1)/2 = 10 pairwise heterogeneity metrics, and then to obtain the overall diversity of the ensemble by averaging these pairwise metrics. Given the hypotheses hi and hj provided by classifiers i and j, we can define four probabilities (a, b, c and d) as presented in Table 1: a – the probability of the occurrence of a case in which both hi and hj hypotheses are correct, b – the probability of the occurrence of a case in which hi hypothesis is correct and hj hypothesis is incorrect, and so on. Evidently, a + b + c + d = 1. Based on these probabilities we can compute the Q-statistic, an indicator that is intuitive and simple to implement (Kuncheva & Whitaker, 2003): Q i;j ¼ ad � bc ad þ bc ; ð18Þ Q has positive values if the same instances are correctly classified by both classifiers and negative values, otherwise. Maximum diversity is obtained for Q = 0. The overall Q-statistic indicator of the ensemble will be com- puted using the following formula: Q ensemble ¼ 1 npairs X jQ i;jj ¼ 2 TðT � 1Þ X jQ i;jj; ð19Þ where T represents the number of classifiers included in the ensem- ble, and |Qi,j| represents the absolute value of Qi,j. Qensemble e [0, 1] and has to take values as close as possible to zero to meet the requirement of classifiers diversity. Otherwise, the trained ensem- ble cannot pass the testing process and must be retrained. 6. Implementation and case study For validating the above concept and related methodologies we performed a series of studies using a WSN designed for measuring the temperature in an indoor environment. Our experimental sen- sor network is composed of nine Crossbow-Iris nodes equipped with MTS310 sensors boards which report the measured values through a gateway (MIB520CB) to a laptop-class device where our software modules can efficiently operate. All the programs were developed in Matlab/Simulink mainly because of its strong capabilities within the fields of recursive parameter estimation, neural networks and ANFIS related tech- niques. Basically we developed two different programs, one related to training and testing process where the parameters of the EBS are tuned for fulfilling the expectations and one for on-line sensing anomalies discovery that implements the trained ensemble of classifiers. The datasets for training and testing were obtained from two different sources, experimental WSN deployment and computer simulation, to cover a wider variety of environmental heat-transfer related phenomena. Our indoor WSN deployment has the configuration depicted in Fig. 9, where the temperature can be intentionally perturbed using Fig. 8. Training deployment shapes. Table 1 Relationship between a pair of classifiers. hj is correct hj is incorrect hi is correct a b hi is incorrect c d Fig. 9. Experimental indoor WSN deployment. D.-I. Curiac, C. Volosencu / Expert Systems with Applications 39 (2012) 9087–9096 9093 Author's personal copy either mobile air conditioning units (affect groups of sensors with heating/cooling waves), either a heat lamp (affects a single sensor by increasing the temperature in its immediate vicinity). Computer simulation is another source of training/testing data- sets. We started from the heat equation which describes the distri- bution of heat (or variation in temperature) in a given region over time. For a function u(x, y, t) of two spatial variables (x, y) and the time variable t, the heat equation with peak p(x, y) as the source is as follows: Table 2 Configuration of the training and testing datasets. Package number Time (min) Node NW (�C) Node N (�C) Node NE (�C) Node E (�C) Node SE (�C) Node S (�C) Node SW (�C) Node W (�C) Node A (�C) . . . 3 86 23.32 23.28 23.12 23.31 23.31 23.07 23.26 23.21 23.18 4 0 18.32 18.28 18.27 18.17 18.30 18.27 18.29 18.27 18.19 4 1 18.74 18.42 18.41 18.35 18.29 18.32 18.39 18.52 18.30 4 2 19.32 19.03 18.70 18.61 18.32 18.48 19.01 19.23 18.69 . . . Table 3 Structure of the training and testing datasets. Package number Destination Number of time samples Number of abnormal functioning samples Source 1 Training 127 15 Real WSN 2 Training 226 32 Real WSN 3 Training 87 12 Real WSN 4 Training 320 63 Computer simulation 5 Training 342 45 Computer simulation 6 Training 287 42 Computer simulation 7 Testing 79 14 Real WSN 8 Testing 162 23 Real WSN 9 Testing 154 15 Computer simulation 10 Testing 303 41 Computer simulation Table 4 Pairwise Q-statistics between the five classifiers. Q-statistics AVC ALC NNC NNAC ANFISC AVC 1.00 0.82 0.86 0.84 0.86 ALC 0.82 1.00 0.79 0.91 0.89 NNC 0.86 0.79 1.00 0.82 0.90 NNAC 0.84 0.91 0.82 1.00 0.86 ANFISC 0.86 0.89 0.90 0.86 1.00 0 50 100 150 16 16.5 17 17.5 18 18.5 19 19.5 20 20.5 21 Time Te m pe ra tu re NW neighbor E neighbor S neighbor sensor A with anomalies Fig. 10. Temperature variations for node A and its NW, E and S neighbors. Table 5 Correct, anomalous and estimated values for the sensor A. Time (min) 10 20 40 60 70 71 72 90 100 105 120 130 140 145 150 Measurements for sensor A without anomalies (�C) 19.43 18.98 17.22 19.57 19.58 19.35 19.06 17.15 18.33 18.34 17.33 18.39 19.75 19.89 19.11 Measurements for sensor A with anomalies (�C) 18 20 16 21 18 20 17 18.5 17 19.5 16 19.5 18 21 17 Estimated values (�C) 19.51 18.80 17.29 19.47 19.40 19.17 18.87 17.34 18.23 18.39 17.22 18.39 19.71 19.74 18.96 9094 D.-I. Curiac, C. Volosencu / Expert Systems with Applications 39 (2012) 9087–9096 Author's personal copy @u @t ¼ a � @ 2u @x2 þ @ 2 u @y2 ! � pðx; yÞ: ð20Þ We simulated this equation in Matlab (PDEToolbox) with different parameters to model diverse dynamics of the heat in the 2D envi- ronment where the sensor nodes are placed. In order to introduce erroneous values, we changed deliberately some of the temperature values near the node under investigation, at random moments in time. For our case study, we considered all the weights of the EBS decision block to be 0.2 and the thresholds of the individual classi- fiers to be 0.35 �C. Due to the way abnormal activity is discovered in our method- ology, the datasets are not composed of independent values, but of packages of values obtained for different consecutive moments in time to catch the dynamics of the environment. For this reason the datasets for training and testing are structured as in Table 2 (neighboring nodes of the sensor A are individualized using their relative position described by cardinal and intercardinal direc- tions: e.g. NorthWest – NW, East – E, etc.), where a package is rep- resented by a number of measurements reported at successive moments in time. The training datasets included six packages, half of them ob- tained through computer simulation and half obtained using the experimental WSN deployment. The testing datasets consisted of only four packages, two obtained from the real deployment and two through computer simulation (Table 3). In the training process we imposed that the Q-statistics be- tween every two classifiers to be less then 0.91 and the overall Q-statistics obtained using Eq. (19) to be under 0.87. These goals have been achieved: the trained and validated ensemble reported an overall Q-statistics of 0.855, while the pairwise Q-statistics for the five classifiers are presented in Table 4. The obtained results prove the efficiency of our ensemble based methodology, in the sense that the ensemble categorized accu- rately 99.71 of the situations and the best individual classifier in terms of results (ANFISC) had 97.99%, while the worst individual classifier (AVC) had 91.26% accuracy. To show how our EBS works, in Fig. 10 we presented the testing dataset package number 9, where the values related to sensor A have been intentionally modified to simulate anomalies at 15 in- stances in time ( t = 10, 20, 40, 60, 70, 71, 72, 90, 100, 105, 120, 130, 140, 145, 150), according to Table 5. The estimated temperatures for the sensor A done by three of our five classifiers (hA,AVC(t), hA,NNC(t), hA,ANFISC(t)) in comparison to the original sensor A time series (without anomalies), are presented in Fig. 11, while the overall decisions taken by the ensemble are depicted in Fig. 12, with the remark that for the first six instances in time the decisions were taken only by the average 0 50 100 150 16 16.5 17 17.5 18 18.5 19 19.5 20 20.5 21 Time Te m pe ra tu re AVC prediction NNC prediction ANFISC prediction sensor A without anomalies Fig. 11. Estimations of the sensor A temperature done by the AVC, NNC and ANFISC classifiers. 0 50 100 150 −0.2 0 0.2 0.4 0.6 0.8 1 Time E B S D ec is io n Fig. 12. The overall decision provided by the EBS. D.-I. Curiac, C. Volosencu / Expert Systems with Applications 39 (2012) 9087–9096 9095 Author's personal copy based classifier due to the reasons mentioned in Section 3 (the necessity to populate input vectors for ALC, NNC, NNAC, ANFISC classifiers with real measurements and the initial convergence of ALC). The estimated values proposed by our ensemble to replace the abnormal measurements are presented in the last line of Table 5. By comparing the sensor A unperturbed measurements with the estimated values, it results a mean absolute error of only 0.118 �C and a maximum absolute error of 0.19 �C obtained at t = 72 and t = 90, confirming once again the preciseness of our methodology. 7. Conclusions It is in the human nature to ask for two, three or maybe more different authorized opinions before taking a significant decision. This human characteristic was transferred to the domain of artifi- cial intelligence through the concept of ensemble based system, producing relevant results in a large variety of domains. Being exposed to numerous risks, WSN often use complex decisional sys- tems for controlling their lifecycle, for processing measurement data or even for dealing with malicious security attacks. Our paper presents a novel perspective upon the sensing anomaly detection in wireless sensor networks by involving an ensemble of five carefully selected binary classifiers. Each of these five classifiers comprises a dynamical model of ‘‘correct behavior’’ of the sensor, estimated based on past measurements provided by the sensor itself or by neighboring sensors. The algorithms involved vary from a simple average computing to complex neural or ANFIS networks assuring the desired diversity of the classifiers to implement an efficient decision making system. Furthermore, using the power of the same ensemble we can precisely estimate the correct value of any measurement affected by sensing anomalies. Acknowledgment This work was developed in the frame of PNII-IDEI-PCE-ID923- 2009 CNCSIS – UEFISCSU. References An, S. H., Heo, G., & Chang, S. H. (2011). Detection of process anomalies using an improved statistical learning framework. Expert Systems with Applications, 38(3), 1356–1363. Belongie, S., Malik, J., & Puzicha, J. (2002). Shape matching and object recognition using shape contexts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(4), 509–522. Cao, F., Lisani, J.-L., Morel, J.-M., Musé, P., & Sur, F. (2008). A theory of shape identification. Lecture Notes in Mathematics, 1948, 264. Chandola, V., Banerjee, A., & Kumar, V. (2009). Anomaly detection: A survey. ACM Computing Surveys, 41(3), 1–58. Chatzigiannakis, V., & Papavassiliou, S. (2007). Diagnosing anomalies and identifying faulty nodes in sensor networks. IEEE Sensors Journal, 7(5), 637–645. Curiac, D., Volosencu, C., Pescaru, D., Jurca, L., & Doboli, A. (2009). Redundancy and its applications in wireless sensor networks: A survey. WSEAS Transaction on Computer, 8(4), 705–714. Dietterich, T. G. (2000). An experimental comparison of three methods for constructing ensembles of decision trees: Bagging, boosting, and randomization. Machine Learning, 40(2), 139–157. García-Pedrajas, N., & Ortiz-Boyer, D. (2009). Boosting k-nearest neighbor classifier by means of input space projection. Expert Systems with Applications, 36(7), 10570–10582. Ho, T. K., Hull, J. J., & Srihari, S. N. (1994). Decision combination in multiple classifier systems. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(1), 66–75. Hsu, K.-W., & Srivastava, J. (2009). Diversity in combinations of heterogeneous classifiers. In Proceedings of the 13th Pacific-Asia Conference on Advances in Knowledge Discovery and Data Mining (PAKDD’09), Bangkok, Thailand, LNCS 5476 (pp. 923–932). Huttenlocher, D., Klanderman, G., & Rucklidge, W. (1993). Comparing images using the hausdorff distance. IEEE Transactions on Pattern Analysis and Machine Intelligence, 15(9), 850–863. Jang, J.-S. R. (1991). Fuzzy modeling using generalized neural networks and Kalman filter algorithm. In Proceedings of the ninth national conf on artificial intelligence (AAAI-91) (pp. 762–767). Jang, J.-S. R. (1993). ANFIS: Adaptive-network-based fuzzy inference systems. IEEE Transactions on Systems, Man, and Cybernetics, 23(3), 665–685. Kim, M., & Kang, D. (2010). Ensemble with neural networks for bankruptcy prediction. Expert Systems with Applications, 37(4), 3373–3379. Kuncheva, L., & Whitaker, C. (2003). Measures of diversity in classifier ensembles and their relationship with the ensemble accuracy. Machine Learning, 51(2), 181–207. Li, H., & Sun, J. (2012). Case-based reasoning ensemble and business application: A computational approach from multiple case representations driven by randomness. Expert Systems with Applications, 39(3), 3298–3310. Li, Y. Y., Thomason, M., Parker, L. E. (2010). Detecting time-related changes in wireless sensor networks using symbol compression and probabilistic suffix trees. In Proceedings of the IEEE/RSJ international conference on intelligent robots and systems (IROS2010), Taipei (pp. 2946–2951). Littlestone, N., & Warmuth, M. (1994). The weighted majority algorithm. Information and Computation, 108(2), 212–261. Ljung, L. (1999). System identification – Theory for the user (2nd ed.). Upper Saddle River, NJ: Prentice Hall. Ljung, L. (2007). System identification toolbox for use with matlab (7th ed.). Natick, MA: The MathWorks Inc. Ljung, L., & Soderstrom, T. (1987). Theory and practice of recursive identification. Cambridge, MA: MIT Press. MathWorks (2002). Matlab fuzzy logic toolbox. User’s guide version 2. Natick, MA: The Math Works Inc. Polikar, R. (2006). Ensemble based systems in decision making. IEEE Circuits and Systems Magazine, 6(3), 21–45. Rajasegarar, S., Leckie, C., & Palansiwami, M. (2008). Anomaly detection in wireless sensor networks. IEEE Wireless Communications, 15(4), 34–40. Shirai, S., Kudo, M., & Nakamura, A. (2009). Comparison of bagging and boosting algorithms on sample and feature weighting. In Proceedings of the 8th international workshop multiple classifier systems (MCS2009), Reykjavik, Iceland, LNCS 5519 (pp. 22–31). Siripanadorn, S., Hattagam, W., & Teaumroong, N. (2010). Anomaly detection in wireless sensor networks using self-organizing map and wavelets. International Journal of Communications, 4(3), 74–83. Tsoumakas, G., Katakis, I., & Vlahavas, I. (2004). Effective voting of heterogeneous classifiers. In Proceedings of the 15th European conference on machine learning (ECML2004), Pisa, Italy (pp. 465–476). Wang, G., Hao, J. X., Ma, J., & Jiang, H. B. (2011). A comparative assessment of ensemble learning for credit scoring. Expert Systems with Applications, 38(1), 223–230. Walters, J. P., Liang, Z., Shi, W., & Chaudhary, V. (2006). Wireless sensor network security: A survey. In Xiao Yang (Ed.), Security in distributed, grid, and pervasive computing, computing (pp. 367–410). CRC Press. Wang, R. X., Lizier, J. T., Obst, O., Prokopenko, M., & Wang, P. (2008). Spatiotemporal anomaly detection in gas monitoring sensor networks. In Proceedings of the European conference on sensor networks (EWSN2008), Bologna, Italy, LNCS 4913 (pp. 90–105). Yang, B., & Bohme, J. F. (1992). Rotation-base RLS algorithms: Unified derivations, numerical properties and parallel implementations. IEEE Transactions on Signal Processing, 40(5), 1151–1166. 9096 D.-I. Curiac, C. Volosencu / Expert Systems with Applications 39 (2012) 9087–9096 
work_xlnsdo7nybgjpcrbtjnknkurzu	----	SOFTM: a software maintenance expert system in Prolog - Software Maintenance, 1988., Proceedings of the Conference on General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.  Users may download and print one copy of any publication from the public portal for the purpose of private study or research.  You may not further distribute the material or use it for any profit-making activity or commercial gain  You may freely distribute the URL identifying the publication in the public portal If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim. Downloaded from orbit.dtu.dk on: Apr 06, 2021 SOFTM: a software maintenance expert system in Prolog Pau, L.; Negret, J. M. Published in: Proceedings of the Conference on Software Maintenance Link to article, DOI: 10.1109/ICSM.1988.10181 Publication date: 1988 Document Version Publisher's PDF, also known as Version of record Link back to DTU Orbit Citation (APA): Pau, L., & Negret, J. M. (1988). SOFTM: a software maintenance expert system in Prolog. In Proceedings of the Conference on Software Maintenance (pp. 306-311). IEEE. https://doi.org/10.1109/ICSM.1988.10181 https://doi.org/10.1109/ICSM.1988.10181 https://orbit.dtu.dk/en/publications/2ee065d9-fc41-41b3-b0e1-589662df4bb5 https://doi.org/10.1109/ICSM.1988.10181 L. Pau Technical University of Denmark Bldg. 348/EMI, DK 2800 Lyngby, Denmark ABSTRACT: This paper describes a software maintenance (SM) knowledge based system ealled SOFTM, serving t h e three following purposes: (1) assisting a software programmer or analyst in his application code m a i n t e n F c e tasks, (2) gen- erating and updating automatically sonware correction docu- mentation, (3) helping t h e end user register, and possibly in- terpret, observed errors on the successive application code ver- sions. The knowledge based system SOFTM is written in PROLOG 11, a n d i s largely applicable to application codes written in different programming languages, provided code descriptors can be retrieved. SOFTM does not address any of the syntactic, input-output, or procedural errors normally de- tected by the syntactic analyzer, compiler, or by the operating system environment. SOFTM i s relying on a unique ATN network based code description, on diagnostic inference proce- dure based on context based pattern classification, on main- tenance log report generators, and on interfacing capabilities of PROLOG I1 to a variety of other languages. 1. INTRODUCTION 1) The current concern about software maintenance i s justified by t h e cost and quality of code repair and up- dates, while at least maintaining software reliability and performances. The estimated productivity gains ex- pected from software maintenance are, according to Bar- ry Boehm, TRW [451 Corrective maintenance: 18% of current effort Adaptative maintenance: 14% of current effort J.M. Negret Battelle Memorial Institute - software maintenance personnel selection - performance goals, and quality control during mainte- ance - software maintenance work breakdown - distribution of responsibilities amongst code users, d e w - loppers, quality assurance, and maintenance - audits and user reviews - problem reporting systems - maintenance logs - use of specification formalisms, typically of t h e SAD1' model - use of program design languages - use of structured techniques to maintain unstructured code - designing in code maintainability Amongst the tools in use, or at the research stage, can he mentioned: - specification and program design languages - software configuration systems - conversion of source code in i t s structural control-flow graphs (e.g.S3, ADA, Z specification languages) - source code controllers, formatters and comparators - declarative constructs - paragraph parsers - cross referencing and linking facilities (mapping) - display of data flows - symbolic debuggers - test data generation - sequence analysis tools (for synchronization and recoverability) - interpreters of abnormal endings - coverage analysis - diagnostic metarules Perfective maintenance: 7 % of current effort Update: due to mismatch between user requirements and software specification: 14% of current effort Many of the above are still at a n early stage, t h u s result- ing in t h e still overwhelming use of debugging he1.r- istics as t h e basic software maintenance approach and tool. The major issues in software maintenance and i t s role in t h e software production process, a r e discussed in [3,8,9,13,14,21,22,24,25,26,27,28,35,43,45,461. The classical approaches followed are: 2) Some research focusses on knowledge based programm- ing, where code is being written with user driven access thru a n intelligent editor to: CH261S-3/88/oooO/030$01.@I Q 1988 IEEE 306 Authorized licensed use limited to: Danmarks Tekniske Informationscenter. Downloaded on October 16, 2009 at 09:11 from IEEE Xplore. Restrictions apply. gnowledgeeources TOOlS Data flow models and Data flow design descriptors Transform aids Design rules Transaction analysis Data dictionary Procedural templates Abstract procedures Design codes Common data structures Structured programming rules Common algorithms Examples of such related CASE projects a r e (non ex- haustively): T e d i u m m e d i o u s E n t e r p r i s e s , P r o - grammers ApprenticeJMIT, USE-IT/Higher order Soft- ware, RlOOOAXational, REFINWReasoning systems, Knowledge-based software assistant (KBSA)/Kestrel In- stitute, Intelligent program editor/Advanced Informa- tion and decision systems, POISE Interactive pro- gramming environment/Univ. of Massachusetts, IN- FORM/Univ. Stuttgart [ll, KBPAlEsprit 121 and also in AI related CASE research 13,6,13,15.16,21,24, 40,43,44, 45,473. Research h a s been reported on e x p e r t h o w l e d g e based debugging: Message trace analyzer - Source code debug- germniv. Waterloo [4l,SPEAWDigital Eqipment [5l,FA- LOSYNniv. Minnesota [7l,SOFTM/ Technical Univ. Denmark & Battelle C301, besides [6,8,13,15,19,36,37,41, 47,481. However, little work h a s dealt so far with knowledge ba- sed software maintenance, incorporating some of the re- levant approaches and tools mentioned above in 1): see [10,11,17,20,21,30,421. It is the purpose of this paper to de- scribe t h e structure of a software maintenance expert system SOFI'M, written in Prolog I1 1493 , and operating on t h e source code of a diversity of programming langu- ages. 1) On a specific piece of application code C(L), written in a programming language L for which t h e r e is a n interpretor, compiler, linker, and editor, the actual knowledge based software maintenance system SOFTM carries out essentially error diagnosis and provides for the propagation of corrective changes (see Figure 2): 1. detection of errors: except : a) syntactic errors detected by the compiler b) cross-referencing errors c ) calculation errors due to a wrong algorithm d) search, query or IO errors due to a wrong algorithm 2. location of errors, except 1) a,b,c,d 3. error diagnosis 4. maintenance guidance for C(L) 5. generation of explanation facilities for 1.-4. 6. automatic generation of code maintenance logs, to be incorporated into the C(L) code documentation. This obeys the diagnostic strategies described in 1291 and using context based pattern classification. This is essentially a backward chaining process where root error ceusdcorrection goals a r e found from their consequences and partially known attributes C501. The knowledge base of SOFTM is divided into three parts: KB-1: Facts in predicate form, about error types, error localizations, diagnostic classes, the environment, and observables. Observable8 are passive if measured without modifying code execution, and active if external perturbations are necessary. KB-2 Code independent kernel rules, applying to the general software maintenance task. KB-S: ,Symbolic descriptors of C(L), derived by rewriting in predicate form C(L) features provided by the compiler, the specification language, or the data flow model, in an augmented transition network (ATN) form. All three part a r e assigned to different sub-worlds in Prolog, for separation and decomposition; they are edited each by a knowledge base editor specific to each of the three parts. The knowledge base KB-3 is interfaced to other tools as indicated. Regarding inferences and queries, the access is as fol- lows: - the code developper, can query and update KB-1 alone, and update C(L) thru the editor of that L language; he can also execute C(L) - the code user, can query KB-1 and run C(L) - t h e code maintenance programmer, can query and up- date KB-1, query KB-3, and update KB- 2. 3. APPLICATION DEPENDENT CODE KNOWLEDGE RE- PRESENTATION The code representation i s treated as fact predicates in a speci- fic knowledge base world. It consists of: a) code structure: it describes the information flows between code modules and arguments, stressing t h e call order during execution. The representation i s a n augmented transition network (ATN) with attributes, written in pre- dicate logic. The node labels are provided by the linker, and the attributes by a standard run of the code (altemati- vely by a Petri-net based simulator). module structure: each module of the application code is represented by a frame, attached to t h e corresponding ATN node. The frame fields are: pointers to U0 argu- ments, pointers to internal arguments, ordered textual description of the functional purpose of each successive block in t h e module (as specified e.g. in structured or functional programming). These frame fields a r e provi- ded by t h e linker or compiler, and by the module docu- mentation located in "Comments". b) 307 Authorized licensed use limited to: Danmarks Tekniske Informationscenter. Downloaded on October 16, 2009 at 09:11 from IEEE Xplore. Restrictions apply. 4. KNOWLEDGE REPRESENTATION OF APPLICATION CODE CIL) (gB3) The relations between modules of C(L) a r e described by a n ATN with the following fact descriptions of each rela- tion in Prolog syntax: mm(ml,m2,L,p)-->; . where: m l : is the module name being called from m2 m2 : is the module name calling m l , with retum to m2 .t : list of call sequence numbers in call chronology, terminated by nil p : name of code or root module The relations between arguments (compiled by the inter- preter o r debugger) and a module, a r e described by t h e facts: am (a,m,line,p)-->; where: a : is a n (YO) argument of module m m : module name l i n e : p : name of code, or root module relative address of argument a, or pointer Each code module i s represented by the Prolog frame structure: md (m, 41,L2, b,p) --> where: m : module name L 1 : list of YO arguments to m represented by relative address pointers in list form, terminated by nil 22 : list of internal arguments in m, which are not in L 1 b : pointers to relative address of first line in each functional block of m, in list form, ended with nil p : name of code or root module md may be represented also with explicit arguments names, while preserving the same frame structure From t h e above knowledge descriptions of C(L) i t i s ob- vious to estimate the number of steps and processes invol- ved in the code, and to sort them by size. C(L) can also preserve the timelstate dependent informa- tion necessary to determine what activities are possible in a dynamic environment; in this case, t h e arguments about time and state a r e tagged separately for data flow or I/O control. 5. APPLICATION CODE INDEPENDENT KNOWLEDGE BASE This knowledge baselworld, which is application code inde- pendent, contains in predicate form, according to the diag- nostic strategy of [291: Y) : observables about the errors, yielding the values of q d j - tativdcontinuous measurements on the application code, once instantiated; L): physical or virtual error locations; E): generic error type or descriptor, such as data type erroi; undefined arguments, etc. C): diagnostic causes (from the domains: design, execution, environment, human errors); M): generic or specific SM actions affecting:application code programming, application code design, software envi *- onment, human factors. 6. KNOWLEDGE REPRESENTATION FOR THE DOMAIIq I”DENTFAC’I-BASES (EB- 1) 1. 2. 3. 4. 6. 6. The facts in K33-1 are described by t h e following Prolcg data structures: Observables: passive( Y-P) o r active Cy-A) < observable - (type) - p/a, (number of observable), (timi!- stamp), nil, (sentence defining observable). (measuru- ment location). (value of observable) . nil > -->; Location IL) < location, (number), (time-stamp), nil, (sentence d 2- fining locations) . (error number) . nil > -+; En” < error, (number), (time-stamp), (likelihood), (error de- scription sentence) . nil > -->; Diagnosis (C) < caase, (number), (time-stamp), (likelihood), (cause name sentence) . (error number) . (cause name sent- ence) . (error number) ... nil > -+; DlaintenanceN < correction, (number), (time-stamp), (likelihood of c f- fect), (sentence describing correction) . nil > -+; Documentation 0) < documentation, (number), (time-stamp), nil, (tiile sentence) . (location number) . nil >-->; 7.l”cEpRocEDuREs They consist of (see Figure 2): a) control structure: i t is the Prolog I1 depth first backtrack- ing with top-to-bottom and left-to-right clause deletian, supplemented by verification and domain dependent pve- dicates; these predicates a r e supplemented by contc xt sensitive control predicates such as diffx,y) and freeze(x,p), which implement t r u t h maintenance a i d conditional propagation [491 b) explicit inference procedures: they include both a for- ward and backward chaining, with a n observatiodgc a1 restriction phase followed by pattem matching on the scts of rules in (e) below, and then by action clauses. The tic- tion clauses consist in addinddeleting facts with likc li- hoods, and by automatically logging them in t h e SM do- cumentation file; 308 Authorized licensed use limited to: Danmarks Tekniske Informationscenter. Downloaded on October 16, 2009 at 09:11 from IEEE Xplore. Restrictions apply. c) domain dependent inference rules: this rule base con- tains in predicate form, with a list syntax: 1. detection rules: Y -> E 2. localization rules: E -> L 3. diagnostic rules: ExLxY -> C 4. maintenance rules: ExLxYxC -> M with SM error 5. incorrectness and insufficiency metarules operating 6. sequencing constraint control metarules, to check call documentation update on the ATN sequences and propagate effects of corrections accordingly code blocks 7. rule cluster documentation generators for functional d ) uncertainty representation, by attaching likelihoods to each error (E), cause (C) or correction (M) fact, the likeli- hoods are propagated and combined along each inference path into a n importance qualifier for each hypothesis or goal The combination of a) b) and c) allows for the automatic propa- gation of maintenance changes, by asserting into &e fact base KB-3 these changes. If new types of errors or locations or cor- rections a r e entered into KB-1 through the proper editors, they a r e automatically accounted for thanks to the Prolog declar- ative form. Thus t h e inference procedure in SOFTM uses fully propagation mechanisms and analysis. 8. ENVIRONMENTAND IMPLEMENTATION The SOFTM knowledge based software maintenance envir- onment in Prolog I1 [491 h a s been supplemented, besides the in- terfaces to t h e operating system, compiler, and higher level utilities (specification language output, simulator input), by: - SM documentation explanation facilities - fact, rule, code structure editors (specialized) - knowledge base management commands - query editor for forward chaining (diagnose a n error) - backward chaining editor (reason to possible candidate- correctionderrors) - likelihood calculations - time-stamp management on all SM actions - debugger - interface between Prolog I1 and constants, variables, lists in different languages (PASCAL, COBOL, FOR- TRAN, ADA) - optional interface to a text database management system containing the full software documentation (e.g. BASIS) The current implementation i s on VAXNMS for application code written in either COBOL, FORTRAN or ADA, and Prolog itself. Specialized application code languages are also consi- dered, e.g. image processing language, test language, and ex- pert systems [391. 9. KNOWLEDGE EXIEMSIONS The basic information h a s been collected, although not yet im- plemented, to enhance the application independent knowledge basis (KB-2) with respect to: - metarules for identifying modules or module to module relations with similar structures - measuring debuggindmaintenance stress, t o estimate likelihoods for detection or corrective actions - generate and place scope markers to delineate code gov- emed by conditional expressiona - protect from any corrective measure the YO arguments - introduce simple software metric attribqtes to compare old from revised code - generate from t h e call sequences a maintenance plan which obeys module interaction - inclusion of simple alternate program verification tech- niques likely to appear in software testing certification standards. REFERwcEs ref. as [91 G. Fischer, H.D. Boecker, The n a t u r e of design processes a n d how computer systems can support them, in P. Degano, E. Sandewall (Ed), "Intgegrated interactive computing systems", North Holland, 1983, 73-86 (INFORM system, Univers- ity of Stuttgart) K. Poulter, Representing programming knowledge in the KBPA, in G.J.P. Katz (Ed), "ESPRIT'85: Proc. of t h e meeting", North Holland, 1986 (KBPA Franz- Lisplc prototype) K.A. Frenkel, Toward automating t h e software deve- lopment cycle, Comm. ACM, Vol28, no 6, jun 1985,578- 89 N.K. Gupta, R.E. Seviora, An expert system approach to real time system debugging, Proc. 1 s t IEEE Conf. on AI applications, 1984, p. 336 (Message Trace analyzer in Prolog, Univ. Waterloo) SPEAR, Expert systems J., Vol 1, no 2, October 1984, p 98 (SPEAR computer error log analyzer at Digital equip- ment) M. Schindler, AI begins to pay off with expert systems for engineering, Electronic Design Magazine, Aug. 9, 1984,106146 R.L. Sedlmeyer, W.B. Thomson, P.E. Johnson, Diag- nostic reasoning in software fault localization, Proc. IJCAI-83, Karlsruhe 1983, Vol 1,29-31 (FALOSY master file update software fault finding, Univ. Minnesota) H.J. Hindin, Intelligent tools automate high-level lan- guage programming, Computer Design Magazine, May l5,1986,46-56 D. Gustafson, A. Melton, A model for software main- tenance, Proc. IEEWACM Conf. on software mainten- ance (CSM.87). Austin, TX, Sept. 1987 B. Terry, R.D. Cameron, Software maintenance using meta-programming, same ref. as [9] F. Cross, An expert system approach to a program' s in- formatiodmaintenance system, same ref. as [91 D. Richard Kuhn. A source code for maintenance. same 309 Authorized licensed use limited to: Danmarks Tekniske Informationscenter. Downloaded on October 16, 2009 at 09:11 from IEEE Xplore. Restrictions apply. 1131 [14l [15l Cl61 R.L. Baber. The Spine of software: designing provably correct software, Wiley, 1988 M.W. Evans, J. Marciniak,Software quality assurance and management, Wiley, 1987 C. Rich, R.C. Waters (Ed), Readings in AI and soft- ware engineering, Morgan Kaufman Publ., 1986 B. Liskov, J.Guttag, Abstraction and specification in program development, MIT Press, Cambridge (MA), 1986 L.B. Alperin, B.I. Kedzierski, AI based software main- tenance, Proc. 3rd IEEE Conf. on AI applications, Or- lando (FL), Feb. 1987 Proc. conf. on reliability and robutsness of engineering software,Computational Mechanics Institute, Como (Ita- ly), Sept. 1987 M.A. Ould, C. Unwin, Testing in software mainten- ance development, Cambridge Univ. Press, 1986 S.S. Yau, S.S. Liu, A knowledge based software main- tenance environment, Proc. IEEE COMSAC'86, IEEE Computer Society, Chicago, 6-10 Oct. 1986 D.A. Higgins, Data structured software maintenance, Dorset House Publ.iWiley, 1987 J. Mastow, Toward better models of the design process, AI Magazine, Spring 1985 GL.B. Kotik. A.. Rockmore, D.R. Smith, Use of RE- FINE for knowledge based software development, Proc. 4 t h Int. Workshop on software specification and de- sign, IEEE Catalog TH 0181-8, April 1987 M. Lubars, M.T. Harandi, Intelligent support for soft- ware specification and design, IEEE Expert, Vol 1, no 4, winter 1986,3341 I. Sommerville, Software engineering, Addison Wes- ley, 1985 NTIS, Annotated bibliography on software mainten- ance, Superintendent of documents, Washington DC, Software maintenance, IEEE Computer society Press, L.F. Pau, Failure diagnosis and performance monitor- ing, Marcel Deker, N.Y., 1981 L.F. Pau, J.M. Negret, SOFTM: a software mainten- ance expert system, Battelle Memorial Institute, Gene- va, 1985 T.S. Chow, Testing software design method by finite state machine, IEEE Transactions on software engin- eering, May 1979 R.L. Glass, Software reliability guidebook, Prentice Hall, 1979 J.E. Hopcroft, J.D. Ullman, Introduction to automata theory, languages and computation, Addison Wesley, 1979 P.P. Howley, A comprehensive software testing metho- dology, Second software engineering standards appli- cation Workshop, SanFrancisco, May 1983 Standard for software test documentation, ANSYIEEE std 829,1983 J.M. Morin, J. Perez, S. Xanthakis, MBthodes de mod- Blisation pour la g6nBration de tests de recette, Institute genie logiciel, T/0052/DI", 1985 G.J. Myers, The a r t of software testing, John Wiley and sons, New York, 1979 P. Zave, A comprehensive approach to requirement pro- blem, Proceedings of COMPSAC, 1979 [lfl 1181 [19] 1201 El] [22] [233 [25l Ea 127J S/N 003-003-02756-1,1986 1281 1291 [301 Cat. 0-8186-0002-0. EH-02014,1983 E311 1321 1331 1341 1353 1361 O7l 1381 1391 r4a 1411 r431 1441 14n 1481 1491 m1 L.F. Pau, Prototyping, validation and maintenance of knowledge based systems software, h o c . IEEE 3ri Conf. on expert systems in government, Washingtoll DC, Oct. 1987 D. Partridge, AI: applications in t h e future of software engineering, Elis HorwoodlWiley, 1986 Proc. ACM Symp. software engineering for high levtd debugging, SIGPLAN Notices Vol 18, no 8, Aug. 1983, ACM Order 593830 C.A. Dyer, Expert systems in software maintainabilit:r, Proc. 1984 Annual reliability and maintainability M. Schindler, Through automation, software shapes it- self to t h e task at hand, Electronic Design, July 25,198!i, H. Abelson, G. Sussman, The structure and interpretat- tion of computer programs: a LISP perspective, MIF press, 1985 A.J. Rockmore, Knowledge based soRware tums spe:- ifications into efficient programs, Electronic Design, S. Bendifallah, W. Scacchi, Understanding softwa -e maintenance work, IEEE Trans. Software engineer- ing, Vol SE-13, no 1, Jan 1987 H. Wertz, Automatic correction and improvement >f programs, Ellis Horwood/Wiley, 1986 J. Loecky, K Sieber, The foundations of prowam ver- ification, Wiley-Teubner, 1987 F. Giannesini, H. Kanoui , R. Pasero, M. van Ca- neghem, PROLOG, Addison Wesley, 1987. L.F. Pau, A survey of expert systems for failure diagn- osis and maintenance, Expert Systems J., April 1986 s ~ P . , IEEE publ. 0149-144 xEB4,295-2!39 87- July 25,1985,105 - 112 310 Authorized licensed use limited to: Danmarks Tekniske Informationscenter. Downloaded on October 16, 2009 at 09:11 from IEEE Xplore. Restrictions apply. Figurel: SofcwareengineeringCASEcycle r- representa- tion of lcodeATN application 1 module ----------------- I SOFTM t I I t Diagnose gh Propose maintenance actlons corrective actions Figure 2 s o m inference functions 31 1 Authorized licensed use limited to: Danmarks Tekniske Informationscenter. Downloaded on October 16, 2009 at 09:11 from IEEE Xplore. Restrictions apply. 
work_xmzjxygacbc4ppd4fjfebxzdca	----	NASA Technical Reports Server (NTRS) 
work_xotjmd5wzngnxjgiycecmqhfcq	----	EXPERT SYSTEM FOR THE NATIONAL EC%)NOMY MODEL SIMULATIONS Yugoslav Journal of Operations Research 12 (2002), Number 2, 247-269 AN EXPERT SYSTEM FOR NATIONAL ECONOMY MODEL SIMULATIONS Lazo ROLJI] Fulbright Fellow, DuPree College of Management Georgia Institute of Technology, Atlanta, Georgia Faculty of Economics, University of Banja Luka Banja Luka, Republic of Srpska Abstract: There are some fundamental economic uncertainties. We cannot forecast economic events with a very high scientific precision. It is very clear that there does not exist a unique "general" model, which can yield all answers to a wide range of macroeconomic issues. Therefore, we use several different kinds of models on segments of the macroeconomic problem. Different models can distinguish/solve economy desegregation, time series analysis and other subfactors involved in macroeconomic problem solving. A major issue becomes finding a meaningful method to link these econometric models. Macroeconomic models were linked through development of an Expert System for National Economy Model Simulations (ESNEMS). ESNEMS consists of five parts: (1) small-scale short-term national econometric model, (2) Methodology of Interactive Nonlinear Goal Programming (MINGP), (3) data-base of historical macro-economic aggregates, (4) software interface for interactive communications between a model and a decision maker, and (5) software for solving problems. ESNEMS was developed to model the optimum macro-economic policy of a developing country (SFRY-formerly Yugoslavia). Most econometric models are very complex. Optimizing of the economic policy is typically defined as a nonlinear goal programming problem. To solve/optimize these models, a new methodology, MINGP, was developed as a part of ESNEMS. MINGP is methodologically based on linear goal programming and feasible directions method. Using Euler's Homogeneous Function Theorem, MINGP linearizes nonlinear homogeneous functions. The highest priorities in minimizing the objective function are the growth of gross domestic product and the decrease of inflation. L. Rolji} / An Expert System for National Economy Model Simulations 248 In the core of the optimization model, MINGP, there is a small-scale econometric model. This model was designed through analysis of the causal relations in the SFRY's social reproduction process of the past 20 years. The objective of the econometric model is to simulate potential short term (one-year) national economic policies. Ex-ante simulation and optimization of economic policy for 1986 showed that, in SFRY, non- consistent macro-economic policy was resolute and led to both slower economic development and more rapid growth of inflation. Keywords: Expert systems, econometric model, national macro-economic policy, multicriterial decision-making, interactive nonlinear goal programming, Pareto optimality, Cobb-Douglas's production function, Euler's homogeneous function theorem. 1. INTRODUCTION Expert systems are computer programs that use a collection of facts, rules of thumb and other knowledge about the given field, coupled with methods of applying those rules to make inferences. Expert systems can be effectively used to solve problems in such specialized fields as optimal short term economic policy choice. The interface between a decision-maker and the ESNEMS is through two software subsystems, which communicate by simple questions and answers. A question could be: Which would you choose: a combination of 3% unemployment rate and an annual inflation rate of 5% − or − a combination of 10% unemployment rate and an inflation rate of 1%? The Expert System will construct the decision-maker's preference function in free form regardless of answers to a complete system of such partial questions. Decision-makers will be able to go back to their PCs where both the Expert System and the entered data regarding the core of the economy reside. They can now add or change any formal preferences in a quantitative form. The Expert System will then solve, utilizing the algorithms built into other econometric models (through use of MINGP) to obtain an optimal development path for the economy under the given external circumstances and stated preferences. This work is the result of a long term research into the application of goal programming to economic modeling [81], [84], [87], and [91]. This paper's aim is to examine abilities and possible advantages of econometric model-based Expert System in economic policy decision-making. A new small-scale national econometric model has been designed to analyze post-hoc economic policy. Embedded econometric models are used to simulate future national economic behavior. As the economic policy choice has been defined as an optimization problem, a nonlinear goal-programming model and a new interactive goal programming methodology for problem solving have been developed and are presented here. These approaches are more efficient than existing ones (both model and methodology). In order to develop an algorithm for solving nonlinear goal programming models where Cobb-Douglas type nonlinear constraints exist, a new gradient nonlinear programming algorithm then was constructed with feasible direction methods built in. L. Rolji} / An Expert System for National Economy Model Simulations 249 An interactive methodology is used here as an interface for support in preparing decisions when a decision-maker is uncertain in ranking his/her priorities. It is also used to define weighting coefficients in the objective function of goal programming problems. This methodology examines compromise solutions for a decision-maker and, if necessary, sets up a new objective function structure. This new structure should lead to a more satisfying and executable solution. There are new applied methods and methodologies built into this Expert System. The preliminary results and tests of this methodology in real decision situations have been encouraging. 2. ECONOMETRIC MODEL What we attempt to measure and the way we measure the economy is strongly influenced by the conceptual framework we have developed for analyzing the economy. The appropriate amount of detail included in a macro-economic model depends on questions being addressed. The method used to construct the econometric data is also reflected in the structure of the model. Investigations have demonstrated that econometric models are able to analyze present behavior and to carry out simulations of the future. They are able to use knowledge about the present to make "baseline projections", i.e., basic or "standard" forecasts that assume a continuation of present trends. At the same time, there is a possibility of influencing the economic environment by macroeconomic policies. Desirable policies can push the country's economy into an orbit of greater and more sustainable growth. An econometric model consists of equation sets, each of which is mathematical complex containing "stochastic members". These members are the results of complex relationships in a nation's economy. In simpler mathematical models, the relationships between variables are exclusively functional. In econometric models, the relations are stochastic. Econometric techniques help to partially compensate for a lost precision, caused by our model's simplification of economic reality. There are two types of equations in econometric models: stochastic, or behavioral, and identities. Stochastic equations are estimated from historical data. Identities are equations that hold by definition; they are always true. The typical econometric problem should consist of: − statement of the issue to be investigated, − specification of the model, − preparation of data base, − estimation of the model, − validation of the model, and − use of the model for policy and other forms of analysis. Design of an econometric model is complex. Identification of key factors influencing a specific economy is necessary. The model must examine and analyze separate segments (niches) of that mechanism (resources distribution, investments, export-import, prices, etc). Practically, each segment defines one partial model. After all the sub-models are built, the econometric model integrates them in a meaningful L. Rolji} / An Expert System for National Economy Model Simulations 250 fashion. Relevant statistical data from the past provides empirical quantification of established and/or hypothesized relationships. Separate economic policies must be quantified and separated into meaningful aggregates (tax rate, monetary policy, etc.). Empirically evaluated relationships in the model show how, in a particular economy, in a particular period, specified macroeconomic parameters are related to economic policy instruments. The economic policy goals can be approximated through the quantitative expression of the economic parameters targeted in the econometric model. The basic data source underlying almost every economy-wide model in the SFRY is the SFRY Statistics of Social Accounts. The National Income Product Accounts (NIPAs) are commonly used in the U.S.A. These accounts are a series of statistics generally considered indicative of the economic health and wellbeing of the nation. The aggregate of the national income accounts is the gross national product, the sum of all productive activity in the country. 3. SHORT REVIEW OF SOME EMPIRICAL MODELS OF ECONOMY SIMULATION IN THE WORLD Empirical models use analytical and methodological tools which enable analyses of key relationships within some economy and by which quick simulations of alternative economic development policies can be evaluated. The mathematician F.R. Ramsey (1928), partly stimulated by J.M. Keynes, raised the fundamental issue of trying to determine an optimal rate of saving or accumulation (see: Johansen L., 1977, p.24). The first attempt at using a macroeconomic model for elucidating the influence on economic development by possible government instruments was done by Jan Tinbergen in Netherlands (1936). The first attempts at constructing national accounts came in Norway (1936) by Ragnar Frish. Economic planning ideas also emerged in the U.S. (W.C. Mitchell's contribution in 1937). Parallel with the development of econometric methods, simulation models of economic development were created using simultaneous equations. The use of econometric models by government planning bureaus for the purposes of forecasting and policy guidance has become widespread in recent years. Following the type of models developed by Tinbergen and Klein and Goldberger many years ago, model projects have been developed and implemented for the U.S., United Kingdom, Canada, the Netherlands, Sweden, Norway, Peru, Italy, France, India, Japan and many other countries on a continuous basis. For many other countries these models are used on an occasional basis. Econometric models have also been applied to smaller government units (eg. states) and/or industries. The relative success of forecasting suggests that econometric model building might be applied profitably to a broader range of countries. A sample of past usage includes the following: L. Rolji} / An Expert System for National Economy Model Simulations 251 Econometric Models in the World Built Number of equations total Behavior equations Definable Equations or identities Dynamical multisectoral model of India development Mahalonobis P.C. 1953 228 - - Klein-Goldberg's model of U.S. economy development Klein L.R. and A.S. Goldberger 1955 23 15 8 Brooking's quartal econometric model of U.S. Fromm G. and L.R. Klein 1965 - - - Alternative model of Peru economy development Thorbecke E. and Condos A. 1966 - - - Regional-National Econometric Model of Italy Chenery H. and Druno M. 1966 12 7 5 A short-term econometric model of French economy Evans K Michael, OECD 1969 55 34 21 U.S. Macroeconometric Model and Multicountry Econometric Model (MC) Ray C. Fair, Yale University 1976 1980 131 1440 30 328 101 18+1094 Econometric Model of Great Britain Terry Barker et. al. University of Cambridge 1980 24 16 8 Japan's FUGI global model of World economy 1986 12700 - - KKRI - Macro-econometric Model of Japan, Naoki Tanaka, Nariyasu Ito and Mamory Obayashi et. al. 1988 86 37 49 Arkansas Econometric Model, Institute for Economic Advancement-University of Arkansas at Little Rock 1995 160 - - Oklahoma State Econometric Model, Office of Business and Economic Research 1998 - - - ASPEN-New economic model simulates U.S. Economy, Sandia National Laboratories, Livermore, California 1998 - - - Kansas Econometric Model The Institute for Public Policy and Business Research, University of Kansas 1998 - - Econometric models developed in other countries have been used primarily for the following three uses: 1) forecasting, 2) multiple analysis and policy simulation, and 3) simulation of past periods. The last of these can be used as a useful diagnostic tool in order to understand how previous recessions, inflation, or other undesirable elements of economic activity might have been prevented. This type of analysis has considerable interests in some context and has been used extensively for the U.S. economy. Fair's models, like instant models, are continually updated based on NIPA data and have been available to Internet1 users and others (to use), on a client-server basis, as a tool to forecast, do policy analysis, and examine historical episodes. 1 http://fairmodel.econ.yale.edu/info/whatis.html. L. Rolji} / An Expert System for National Economy Model Simulations 252 Figure 1: Cyclical flow of open economy's economic activities L. Rolji} / An Expert System for National Economy Model Simulations 253 4. ECONOMETRIC MODEL AS THE CORE OF THE ESNEMS The econometric model developed here can be described as a productional, nonlinear, aggregate, macroeconomic, dynamic simultaneous equations small-scale model. The model is not in equilibrium because it does not contain an explicit formulation of equality between supply and demand. The model is dynamic because it contains time as a variable and also lagged endogenous variables. It is nonlinear because Cobb-Douglas's production function is used and also because two other non- linear equations are used to link constant and variable prices. The main equation in this model, as it is mentioned earlier, is the production function which starts and ends the cyclical flow of an open economy's economic activity (Figure 1.). This function specifies the maximum aggregate output, which is divided between consumption and investment. This is, of course, a very simplified presentation of the cyclical flow of open economy's economic activities. This process is too complex to allow the inclusion of all elements of the economy and thereby form action-consequence conclusions. The number of relations by which this process is interconnected is so high that it is unable to recognize it. But some basic and common values and relations have been identified and crystallized. These are: gross national product, fixed capital, employment, investments volume, prices, personal income, export and import, as well the relationships that exist between them. The starting hypothesis in this research (the statement of the issue to be investigated) is that the essence of economic disturbances in the Yugoslav economy was more due to insufficient supply of goods, while excessive demand was only a secondary occurrence. The equations, which constitute our econometric model are: Row # Equations of econometric model The name of dependent variable Notation 1. GDPTCP = GDPSSCP+GDPPSCP Total gross domestic product at constant price (CP, 1972=100) 1x 2. GDPSSCP = f( FCAPCP, EMPSS) Gross domestic product of social sector at CP 2x 3 GDPPSCP = f (GDPSSCP) Gross domestic product of private sector at CP 3x 4 FCAPCP =f (GICP, FCAPCP.1) Fixed capital of social sector at CP 4x 5 EMPPRS = EMPSS + EMPIS Employment people in productive sector 7x 6 EMPPS = f (GDPPSCP, EMPPS.1) Employment people in private sector 8x 7 EMPNPS = f (GDPTCP, EMPPRS) Employment people in nonproductive sector 9x 8 EMPT = EMPPRS + EMPNPS Total employment people 10x 9 EMPTSS = EMPSS + EMPNPS Total employment people in social sector 20x 10 IDGDP = f (PIIG, PIPR, ILC, GDPTCP) Implicit deflator of gross domestic product 11x L. Rolji} / An Expert System for National Economy Model Simulations 254 Row # Equations of econometric model The name of dependent variable Notation 11 PIPR = f (MMFP, INVFP, PIIMP) Price index of producers 12x 12 PIIG = f (PIPR, EXRIMP) Price index of investment goods 6x 13 ILC = f (PIPR, ILC-1, PIIG) Index of living costs 13x 14 PCFP = f (ILC, NPINFP, GDPTCP) Personal consumption exp- enditures at flow prices (FP) 18x 15 GICP = f (GDPSSCP, GICP.1, PIIG, GDPTCP-1) Gross investments at CP 19x 16 NPIFP = f (GDPSSCP, ILG, EMPTSS) Net personal income at FP 15x 17 EXPCP = f (IMPCP, EXREXP PIEXP) Export at CP 21x 18 IMPCP = f (GDPTCP, EXRIMP, PIIMP) Import at CP 22x 19 ORFSFP = FSRFP - TXFP - STFP - DUTFP Other revenues of fiscal system at FP 25x 20 FSRFP = f (GDPTCP, IDGDP) Fiscal system revenues at FP 24x 21 TXFP = f (NIFP) Taxes and contributions at FP 26x 22 STFP = f (PCFP) Sales tax at FP 27x 23 DUTFP = f (EXRIMP, IMPCP) Duty at FP 28x 24 NIFP = GDPTCP * IDGDP - AMFP National income at FP 29x 25 AMFP = f (GDPTCP * IDGDP) Amortization at FP 30x 26 PIIMP = f (TIME) Prices index by import 31x 27 PIEXP = f (TIME) Prices index by export 32x The notation means that the equation is stochastic, otherwise it is an identity equality. The variables inside the parentheses are explanatory variables. Exogenous variables in equations are underlined, and lagged variables are subscripted by (...)f −1. Notation Exogenous variable The name of dependent variable 5x EMPSS Employment people in social sector 14x EXRIMP Exchange rate by import 16x MMFP Monetary mass at FP 17x INVFP Inventories at FP 23x EXREXP Exchange rate by export 33x TIME Time (1966th − 0.l) Lagged variable The name of lagged variable 34x FCAPCP-1 Fixed capital at CP ( )− 1t 35x EMPPS-1 Employment people in productive sector ( )− 1t 36x ILC-1 Index of living costs ( )− 1t 37x GICP-1 Gross investment at CP ( )− 1t 38x GDPTCP-1 Total gross domestic product at CP ( )− 1t L. Rolji} / An Expert System for National Economy Model Simulations 255 The econometric model presented here provides a better understanding of the stochastic aspects of a developing economy by examining the changes that occur over the time. The kernel of the econometric model is a sort of combination of demand and supply models. The model consists of 27 equations, 21 of which are stochastic and 6 are defining. In order to have a better insight into the system of relationships and relations, we could present our model roughly by seven blocks: BI − block of production, which is described by the real gross national product in both social (government) and individual sectors. Econometrically, gross national product is explained by fixed capital in constant prices and the number of employed people. B2 − block of prices, formalizes an implicit gross national product deflator, as a representative measure of inflation. B3 − block of investments, although it is given in high aggregated form, presents a very substantial (vital) model segment by which the connection of block of prices and block of production is assured. B4 − block of personal-consumption, describes the consumer behavior in this particular economy and is theoretically linked to the block of personal incomes. B5 − block of personal incomes. B6 − block of foreign exchanges, gives data of dependency degree between production and import of goods, the data about balancing exchange with the World, and how these factors influence fiscal revenues. B7 − block of fiscal revenues. Reflects some assumptions which are the starting points of this econometric model: − that the balance between fiscal's revenues and expenses should be assured (what is conditlo sine qua non of this model), − that the expenses of the fiscal system are incurred in part due to interaction with the earlier specified "blocks." − that the fiscal policy instruments are endogenized by a classical approach. That is to say that the fiscal policy has neglected economical and social effects, while ensuring (providing) enough funds for the functioning and maintenance of the social (government) sector. The blocks are mutually connected by some variables (see Figure 2). Most of the macroeconomic aggregates are influenced by monetary policy measures (financial market), fiscal policy measures (country revenues and expenses) and by foreign policy measures (import, export, exchange rate). As it is known, economic policy measures in econometric models are presented by instrumental variables. Therefore the solution to the problem of economic policy choice in such models has been reduced to determining the instrumental variable values. The equations of the econometric model are estimated by the ordinary least square method along with the F-test of significance of the estimated coefficients and regression validation. The auto-correlation has been tested by Durbin-Watson statistic L. Rolji} / An Expert System for National Economy Model Simulations 256 defined by 5%. The presence of auto-correlation is rejected by the Cochrane-Orcutt procedure. Figure 2: Block-recursive econometric model structure L. Rolji} / An Expert System for National Economy Model Simulations 257 For econometric validation for forecasting (model calibration) and consistency evaluation, by ex-post simulation2 we have employed constant term adjustment (see: Klein and Young, 1981). Since the values of the model variables were measured in different measures, before MINGP application, we ensured the "normalization" of some of the equations by reducing the multidimensional feasible solution area on the dimension scale between 0 and 20. Once the statistical tests on the data were satisfactorily completed, input parameters estimated, and several ex post tests performed to validate the model and to estimate the minimal relative size of estimation error, the econometric model was run for predicting optimal economic policy during the next one year planning period (horizon). Dynamically, we first applied MINGP as a test of consistency of the planned goals for the next year. In the planning documents the main quantitative goals were: − increase in gross domestic product by 3%, − increase in inflation by 45%, and − increase in gross investments by 2%, which in essence are conflicting goals. The dynamic simulation of econometric problem, solved by MINGP as a single objective nonlinear programming problem, showed that such a goal constellation of the Yugoslav economic policy would not be realized without destroying the structural constraints of the existing econometric model. Therefore, the input file of the optimization policy choice problem, consists of the following relaxed goals: − increase in gross domestic product by a rate greater than 2.2%, − increase in inflation by a rate greater than 43% but less than 88%, and − increase in gross investments by 2%. The mathematical model formulation of the choice of optimal economy development policy, constructed via a combination of dynamic simulation and optimization techniques, can then be formulated as a nonlinear goal programming problem as below: Minimize ( ) ( ) ( ( ) ( ) ( ) + − − − + − + + − + − + − + − + − − + + + + + + + + + + + + + + + + 1 39 40 43 44 2 2 2 24 25 25 3 28 28 4 29 29 5 30 30 6 31 31 7 46 P d d d d P d d d d d P d d P d d P d d P d d P d ) + Subject to the constraints from (1) to (46) 2 The solution of a model over a historic period, where the actual values of exogenous variable are known, is called ex-post simulation. In this case, we do not have to guess values of the exogenous variables because all of these variables are known. One can thus use ex post simulation to test a model in the sense of examining how well it predicts historical episodes. The solution of a model for a future period, where "guessed" values of the exogenous variables are used, is called an ex ante simulation. L. Rolji} / An Expert System for National Economy Model Simulations 258 1) 100x1−100x2−10x3 = 0.0 2) 100x2−155.27821x40.538509x50.461491 +d2-−d2+ = 0.0 3) 10x3−8.12x2 = 24.351247 4) 1000x4−5.92362x19−975.69x34 = 0.210555 5) x7−x5−x8 = 0.0 6) x8−0.0021x3−0.9524x35 = 0.002298 7) x9−0.061x1−0.15327x7 = 0.007853 8) x10−x7−x9 = 0.0 9) x20−x5−x9 = 0.0 10) 10x11−2.22463x6−1.16842x12−8.31871x13−0.0843x1 = 0.037344 11) 10x12−2.25x16−6.69x17−0.27215x31 = 0.747693 12) 10x6−5.41342x12+0.5382x14 = 0.6783085 13) 10x13−9.46445x12−0.1047x27−4.80405x36 = 0.4591351 14) −1000x18+12.1346x1+604.3191x15+1399.81x13 = 417.229687 15) 10x19−59.3205x1−2.45899x6−9.91684x37 = 417.229687 16) 100x15+7.27287x20−13.131x2−88.6279x13 = 4.813987 17) −100x21+40.792x22+10.609x23+38.01588x32 = 20.114828 18) −100x22+54.817x1−38.716x14+76.22366x31 = 84.655504 19) 1000x25−1000x24+1000x26+100 x 27+100 x 28 = 0.0 20) 1000x24−1276.2744x11−87.9184x1 = 284.451904 21) 1000x26−186.97x29 = 3.951925 22) 100x27−131.354x18 = 3.951925 23) 100x28−9.37x22−74.7098x14 = 12.79104 24) 1000x29+1000x30−1000.27821x1 x11 +d24-−d24+ = 0.00022 25) −1000x30+105.43 x1x11 +d25-−d25+ = 25.303613 26) x5 = 5.40294 27) 2.86816x14−3.556x11 = 0.0 28) x16 +d28-−d28+ = 2.664805 29) 2.86816x17−2.3651x11 +d29-−d29+ = 0.0 30) 2.86816x23−3.353x11 = 0.0 31) x10 +d31-−d31+ ≤ 7.7 32) −x31+2.1896x33 = 0.65596 33) −x32+1.8761x33 = 0.71605 34) x33 = 2.1 35) x34 = 1.01952 36) x35 = 0.138 37) x36 = 5.89485 38) x37 = 8.06203 39) x38 = 3.93648 40) x1 +d40-−d40+ ≤ 4.0221 41) x11 +d41-−d41+ ≤ 5.4 42) x11 +d42-−d42+ ≥ 4.1 43) x19 = 8.22327 44) x5 +d44-−d44+ ≤ 5.5 45) x1 +d45-−d45+ ≥ 3.98336 46) x14 +d46-−d46+ ≥ 5.334 L. Rolji} / An Expert System for National Economy Model Simulations 259 In the first 25 terms the equations of the econometric model have been given. In terms from 26th to 30th rows the functions of instrumental variable values are given. The term in the 31st row contains the upper bound of available working people (in millions). In the 32nd row, is the variable "prices by import." In the 33rd row is the "prices by export" variable. Although these rows (26-33) belong to the econometric model, they are not treated as goals but as structural constraints, which are functions of time, thus making the model time-dependent. In the 34th row the time variable assumes that first year ( =33 0 1. )x is 1966, and rows 35-39 give the value for time lagged variables: ( ) ( ) ( ),− −4 1 1 19 1( ), ,−8 1 13 −x x x x , and ( )−1 1x respectively. The terms from the 41st to 46th row contain the boundary values for instrumental variables. The first priority in the initial objective function is given to satisfying the goal of gross national product growth of both sectors of capital ownership (social and private) − of variable 1x , and to keeping the inflation rate ( )11x in previously given bounds. The second priority goal is the satisfying of all nonlinear functions in the model. In the 3rd, 4th, 5th and 6th priority in the objective function are the requirements for minimization of deviations from goal instruments of economic policy and in the 7th priority is the goal requirement for national gross investment growth. 5. RESULTS OF ECONOMY MODEL SIMULATION This problem is solved after the 3rd interactive trial and 15 iterations of the NGP algorithm to determine the best solution (so called Pareto's optimum3): =1x 4.02210 Total gross domestic product of both social and individual sectors at constant prices =2x 3.49481 Gross domestic product of social sector at constant prices (CP) =3x 5.27291 Gross domestic product of individual sector at CP =4x 1.04443 Fixed capital of production purchase value of social sector at CP =5x 5.50000 Number of employees in social sector of economy =6x 1.98261 Index of growth of investment goods at CP =7x 5.64480 Number of employees in both social and individual sector =8x 0.14480 Number of employees in individual sector =9x 1.118380 Number of employees in nonproductive sector =10x 6.76318 Total number of employees in both productive and nonproductive sector =11x 5.40000 Implicit deflator of gross domestic product at CP 3 The solution of multi-objective programming problem in which satisfying of less important goal could be reached only by doing not (worse) on expense of the goals on higher priority. In other words: "Under what configuration of the economic system it would be able to make one person better off only at the expense of making someone else worse off". L. Rolji} / An Expert System for National Economy Model Simulations 260 =12x 4.06739 Producer's price index at CP =13x 5.42617 Costs of living index at CP =14x 5.33400 Average exchange rate of dinar (domestic currency) in import =15x 4.83480 Net incomes of both productive and nonproductive sector of social sector =16x 3.50849 Monetary mass =17x 4.62770 Stock growth increase =18x 10.14894 Individual consumption expenditures =19x 8.35289 Gross investment in fixed assets =20x 6.61838 Employed people in social sector in both productive and nonproductive sectors =21x 2.63153 Export of goods and services at CP =22x 2.29802 Import of goods and services at CP =23x 6.31283 Average exchange rate of dinar (domestic currency) at export =24x 7.52995 Revenues of fiscal system =25x 2.06817 Other revenues of fiscal system =26x 3.69190 Taxes, fees and contributions, excluding sales taxes =27x 13.37056 Sales tax =28x 4.32826 Customs revenues =29x 19.45477 National income =30x 2.26457 Amortization =31x 3.94220 Index of gross import prices at CP =32x 3.22376 Index of gross export prices at CP This solution has been obtained in the 3rd interactive trial, after the two previous interactions have not given a satisfactory solution for achievement of the 2nd and 3rd goals, but in such a way that the 3rd goal in the initial objective function was renamed as the first and the 2nd as the 3rd, while the order of other priorities remained unchanged. This fact points to the importance and potential role of the approach offered by MINGP in relation to ad hoc, often pragmatic approaches. Namely, it must be totally clear that a decision maker could not, by a simple deductive procedure, reach such a conclusion as the improvement of satisfying the 2nd goal it could be accomplished indirectly by renaming the 3rd goal as the first one. In essence, mathematically speaking, MINGP has found three local minimums among which, in collaboration with the decision maker, it has selected the most satisfactory solution. The optimum solution of economic policy for the planned year1 as predicted by the model, has revealed that there were possibilities for a further increase of production activities in the Yugoslav economy but not with such intensity of social product and inflation growth as it had been predicted. In other words, it has been L. Rolji} / An Expert System for National Economy Model Simulations 261 shown that the Yugoslav economy, wishing to stabilize, must look for a trade-off between growth and stabilizing under limited investments (this being the problem of most developing countries) at the expense of social product growth. In addition to the solution of instrumental variables of monetary and fiscal policy the specific solution given in the paper demonstrates that trade-off is an economic policy which would give a social product growth of 2.2% and inflation growth of 54% (not 45%). In other words, inflation of 45% cannot be maintained at a social product growth of 3% but at a level almost double. 6. GOAL PROGRAMMING AND GRADIENT METHOD OF FEASIBLE DIRECTIONS Nonlinear models are difficult or impossible to solve analytically, but they can usually be solved numerically. Like using the Gauss-Seidel technique. We used here goal programming and feasible directions methods. The goal programming method has been considered the oldest method of multiple objective programming. It has been very popular in the applicative sense but waited far too long to pass from a theoretical basis (Charnes and Cooper, 1961) into the sphere of practical applications (Lee, 1972, and Ignizio, 1976). The problem of nonlinear goal programming can be expressed in the following way: min ( ) ( )− + = = = ∑ ∑ 1 1 k n i l i i l i F d w P d d+ (1) s.t. ( ) − + = = = + + −∑ ∏ 1 1 ij nn e i ij j ij i ij j j G x g x h x d d c= i (2) ( ) = = ≤∑ 1 n i ij j j iA x a x b (3) , , ,− + − +≥ ⋅ =0 0j i i i ix d d d d , for all ,..., , ,...,= =1 1i m j n , (4) where jx are decision making variables, − id and + id represent negative and positive deviation variables from the goals (underachievement and overachievement), respectively, are coefficients of linear portions of goals (constraints (2)), are coefficients of structural constraints (3), are coefficients of nonlinear portions of goals (2), and ijg ij ija ijh e are exponents, ic and ib are constants of right hand sides in (2) and (3) respectively. , ,...,= 1lP l k in objective function (1) are preemptive priority factors, so that the following is valid: +>>> 1j jP P for all , ,...,= −1 2 1j k . L. Rolji} / An Expert System for National Economy Model Simulations 262 The highest priority is indicated by 1P , the next highest by 2P , and so fort. are weights assigned to some priority factors. The notion of priorities holds that iw 1P is preferred to 2P regardless of any weights associated with w 2P . In the great majority of up-to-date developed methods used for solving problems of this very type, among the most significant ones is the gradient method combined with the feasible directions method. The feasible directions4 methods have been primarily designed for nonlinear programming problems and they are the iterative ones whose solutions in certain iterations have the following recursive form: ( )+ = +1p p p px x l s x for all , ,...,= 1 2p m where ( ) ( , , ,..., )= 0 1 2p ps x s x x x x is a direction and is a step size, which is chosen so that: ≥ 0pl ( ) ( )+ ≤p p p pF x l s F x , ≤ ≤pl0 1 , + ∈p p px l s X , and where: { ,= ∈ and conditions (2), (3) and (4) are satisfiednX x R } Gradient methods use the direction of a gradient as a solution improvement direction, thus defining the feasible and usable feasible directions and enabling reduction of the nonlinear problem to an approximate linear problem which is close to the initial one. The method is iterative and each iteration starts with an initial feasible vector. At each iteration of the feasible directions method a feasible direction of improvement (usable feasible direction) is determined and a new "better" point is found along that direction. Optimality is achieved when no further improvement can be made in any feasible direction To determine the gradient this method requires that functions are continuous and differentiable. In order to guarantee convergence of the algorithm the gradient method requires that the model constraints form a convex set at each goal level while the objective function is concave Many methods for solving linear or non-linear programming problems will appear to be methods of feasible directions. They only differ in the extra requirements for fixing the starting point 0x , the directions ks or the step lengths kl . We are in this case particularly interested in a nonlinear function of the Cobb- Douglas type, the generally form of which is: ( , ,..., ) = = ∏1 2 1 i r b n i i g x x x h x (5) 4 Feasible direction. Given a feasible point x, a direction vector, say d, is feasible if there exists > 0s such that is feasible. +x sd L. Rolji} / An Expert System for National Economy Model Simulations 263 where: ≥ 0ix , for all i , are independent variables, , ,...,= 1 2 n ∈h R is coefficient, and are exponents of independent variables. This is the most familiar and most frequently applied form of production function by which the production value in a certain economy is expressed as a function of labor and capital investment. ∈ib R The idea of linearization of nonlinear constraint and solution of nonlinear programming problems is not new, because Griffith and Stewart first suggested in 1961 that nonlinear problem may be linearized in the region about the particular point by expansion in the Taylor's series, ignoring terms of a higher order than linear and adding two more restrictions for each nonlinear constraint. In that way nonlinear programming problems have been transformed into a form which can be solved by the linear programming algorithm (see: Ignizio [1976], and especially Lee, S.M. and Olson D.L. [1985]). Our idea was to apply the Euler's theorem for the "total" linearization of nonlinear constraints in the region about the particular point (see: Rolji}, L. [1987]). Owing to the Euler's homogenous function5 theorem it is possible to apply this method to solve the problems of NGP regardless of whether Cobb-Douglas's functions are linearly homogenous (that is: = =∑ 1 1 n i i b ), or positively homogenous (that is: ). = ≥∑ 1 1 n i i b In economic theory, the production function is frequently assumed to be linearly homogenous because such functions have convenient properties. The advantage of our approach is that it is not necessary to add two more constraints for each linearized nonlinear constraint in every simplex iteration of NGP algorithm, and because the linearization is more accurate. 7. ALGORITHM OF NGP AND MINGP As already mentioned, an algorithm of nonlinear goal programming is based on hybrid connection of modified simplex method of goal programming (Lee, 1972) and gradient method of feasible directions (Zoutendijk, 1976). As, in the problem at hand, the objective function is linear and only three constraints are nonlinear, the procedure is simplified when compared with general convex programming problems. The iterative procedure is given in five steps with an initial step (Step 0) being used only at the beginning i.e. in the initial iteration (Figure 3). The initial step sets all solution vector values to zero and uses this point as initial 0x in which all nonlinear portions gradient values of the goals are equal to zero. This reduces the problem of NGP (problem of (1) to (4)) to linear approximation solvable by the modified simplex method of linear goal programming. 5 A homogeneous function is a function which has a feature that for some real constant λ satisfies ( , ) ( ,λ λ λ= n )F x y f x y for a fixed n. Then we say that function is homogenous by degree of homogeneity of n. F L. Rolji} / An Expert System for National Economy Model Simulations 264 Figure 3: Flow diagram of NGP algorithm L. Rolji} / An Expert System for National Economy Model Simulations 265 The first step of an NGP algorithm is the computation of gradients of all constraints together with checking the status of nonlinear constraints in order to deflect, if need be, the gradient of active nonlinear constraints and to avoid zigzagging (see: Zoutendijk, 1976). Without this deflecting, the algorithm may converge to a suboptimal point (see: Van de Panne and Popp, 1963). In the second step the formulated linear goal programming problem (LGP) is solved by a modified simplex method. Namely, nonlinear constraints are transformed to linear on the basis of the Euler's theorem computed gradient value at point px (in the initial step it was 0x ). The modified simplex method of LGP derives the feasible solution px . In concordance with the Euler's homogeneous function theorem conditions, it is noted that if is homogenous of degree r and the first-order derivatives exist, then it can be shown that is: ( , ,..., )1 2 nf x x x ( , ,..., ) ( , ,..., ) =    ∂ = ⋅ ∂ ∑ 1 21 2 1 1 p n n n i ir x f x x x f x x x x r x where, with   px is denoted "in particular point ′′px , and where for Cobb-Douglas's functions (5): ( )( , ,..., ) − =   ≠    ∂ =  ∂      ∏11 2 1 k p n bn k ik i ix i k f x x x hb x x x i b for all ≥ 0ix are the coefficients of linearized constraints calculated at particular point px . The third algorithm step serves for feasible direction ( )ps x computation: ( ) ′= −p p ps x x x which has to be "searched" according to the constraints and structure of priority in order to improve the simplex solution. In the fourth step, the optimal solution vector movement size (step size) is determined by linear searching in the feasible direction, identified in the previous step. In this way, first the structural constraints ( )iA x ( )iG x must be satisfied and then the goal ones, starting from the goal constraint in (2) which contains the deviation variable with top priority in the objective function of the NGP problem To perform the search and to prevent infinite moving in the cases of unbounded problems, it is necessary to determine the lower and upper movement bounds in feasible direction and unit movement size (increment) within these bounds. The convergence of the method for the problem at hand, speed of convergence, and other algorithm features depend on the choice of vectors ( )ps x and bounds for the step size pl . Initially, the lower limit of pl in the algorithm was set to 0 and the upper to 1, L. Rolji} / An Expert System for National Economy Model Simulations 266 but it is possible to define them differently if a need be. The unit movement size in the algorithm at first was set on 0.1, but it is possible to increase the search density. By these procedures we define whether the first goal constraint (per priority and not per order) is satisfied within the initial bounds. If it is satisfied at some smaller interval [ , ]pl pdl l , for all and ≥ 0pll ≤ 1pdl , the search for further goals constraint goes on exclusively within this interval as its satisfaction, in accordance with so called Pareto optimality, cannot be sought to the detriment of satisfaction of a higher priority goal. If during the search we come across a constraint which has not been satisfied previously by goal constraint set bounds, then within these bounds the algorithm identifies the value of movement size for which the deviation from the subject constraint is the least pl , we complete the searching and compute the new solution (successor) as follows: ( )+ = +1p p p p px x l s x If all constraints are satisfied then the new solution is at the same time the optimal solution to the set problem. In the fifth algorithm step we check the problem convergence by setting previously small value of the desired level of convergence accuracy ε− : | ( ) | ε+ − ≤1p px x , for all , ,...= 1 2p which means that the algorithm has converged and it stops. Otherwise, the procedure is repeated starting from the first step. 8. INTERACTIVITY IN ESNEM The interactive work of ESNEM is supported by the algorithm included subsystem for establishing the priority structure-PRIOR and subsystem POST for solution improvement on the basis of post-optimal analysis and priority structure change. A decision maker can but need not be in position to formulate the priority structure in (1) precisely, i.e. to define the objective function or to rank the goals according to their meaning. If the decision maker is not in such a position, the subsystem for establishing the priority structure offers the user a list of possible goals asking the following questions: − could you rank the goals, at least partially, − could you group the goals, − could you select the most important and the least important group of goals, − do you want to assign different priorities within this group, in an attempt to obtain at least partially a differentiated priority structure. This subsystem also gives the possibility of assigning different weights to goals of the same priority group whether the weights are determined by the decision maker alone or by a program on the basis of random choice. L. Rolji} / An Expert System for National Economy Model Simulations 267 If so requested by a decision maker, the POST subsystem can change the priority structure. It starts up and carries out the post-optimal solution analysis and/or keeps searching for a better solution. The work of this subsystem goes on in the next few steps: 1. Analyses of the results obtained by NGP algorithm in order to determine the level of satisfaction of certain priorities in objective function, 2. Diagnosing the conflict within the priority structure and establishing the method for its solution, 3. Memorize the best former solution, 4. Priority structure change in objective function according to the information obtained in steps 1 and 2. 5. Resetting the algorithm NGP, this time from the point of the former best solution with the priority structure defined in step 4, 6. New solution finding and interaction with user to check whether he or she is satisfied with this solution or not, 7. Proclamation of the new solution for the former best solution, if user is satisfied and, otherwise, resetting the memorized former best solution to the operative computer memory, 8. Examination of further possibilities for solution improvement and, if there are some, returning to step 3 or, otherwise, going to step 4, 9. Printing final results. By the approach "keep finding out better solution" the initial problem solving time is significantly decreased in the case when a decision maker finds, through insight into the former best solution, an often unforeseen dependence between priorities i.e. goals. In that case it is not necessary to change the priority structure and to restart the NGP algorithm from the initial step. The subsystem for solution improvement includes the NGP algorithm from the point of the former best solution and in many ways decreases the necessary number of iterations. Computer programs for realizing ESNEM were coded in Portran7 for PC. 9. CONCLUDING REMARKS The purpose of this research is to 1) construct a new structure for econometric modeling of the SFRY economy, 2) apply this new approach to analyze SPRY economic activities, and 3) enable the recommendation of economic policy variants which would bring about an optimization of production, consumption and inflation outcomes. This optimal policy variant represents a stabilization policy for the post-war SFRY. We have shown the possibility of hybridizing many analytical methodologies for a national policy model simulation within ESNEMS - mathematical methods algorithms, econometric methods, particular data base connection, software-driven interactive communication, and an efficient expert system. One non-typical example of multi-objective decision making was illustrated. For demonstrating the possibilities of ESNEMS, both MINGP and the baseline of a new SFRY econometric model have been established. Periodically updating of the econometric model built in ESNEMS (and linearized by MINGP) becomes a crucial L. Rolji} / An Expert System for National Economy Model Simulations 268 step. The specification of any given econometric model, of course, has certain deficiencies, particularly for developing economies with not established standard econometric approaches. But, as L. R. Klein claimed, "A bad model is better than no model at all." Information richness within the critical matrix of simplex method has enabled information feedback for the decision model. ESNEMS, as a hybrid methodology solving NGP problems by iteration, seeks the best solution for a specified objective function structure. Although MINGP is a numeric methodology for solving only certain types of NGP problems, it has been extended to also solve other nonlinear forms. MINGP can be used for NPG problems represented by homogeneous1 continuous, and differentiable functions. The only ESNEMS requirement is separate subprograms for calculation of nonlinear function gradients and for calculation of the solution value(s). Many decision methods can handle other forms of nonlinear constraints than the Cobb-Douglas type. Using the Euler's homogenous function theorem, it was possible to realize a hybrid connection of the gradient feasible direction method together with linear goal programming method. 10. FUTURE RESEARCH There is much more work which could be done by multidisciplinary researchers. This work could equip countries and other economic entities to create better formations of realistic systems and more reliable model approximations. This would enable economic policy analysis leading to stable trends instead of ad hoc and sub-optimal economic appraisal. There are no practical limitations to the application of ESNEMS and econometric models to lower organizational forms (regions, enterprises, etc.) if an adequate database is provided. Reliability and availability of appropriate statistically significant data is needed. There is a great potential in the further hybridization of ESNEMS. Thus, we suggest the connection of MINGP and econometric methods for solving simultaneous equation parameters. Econometric model estimation and solution of some statistical parameters in the objective function of a generalized NGP problem should also bear fruitful results. Other potentially fruitful research lies in combining ESNEMS (and MINOP) with some methods of multi-objective ranking (Promethee, Electre, Eigenvector, etc.). Particularly if these MOR methods include unbiased approaches to the determination of objective function priorities. Stochastic approaches to the priorities discussed in Section 5 can also be used to create add-on methodologies. Econometric models should be upgraded regularly. New connections, interrelationships, and regularly updated series of new consecutive ex-post simulation and calibration will enable better (more timely, precise, and updated) economic policy recommendations. The work on features improvement of the ESNEMS should be directed, also, to integrate the database and statistical results of econometric model estimation with MINGP. L. Rolji} / An Expert System for National Economy Model Simulations 269 REFERENCES [1] Ignizio, J.P., Goal Programming and Extensions, Lexington Books, Lexington, 1976. [2] Klein, R. L., Economic Theory and Econometrics, Edited by Jaime Marquez, University of Pennsylvania Press, Philadelphia, 1985. [3] Klein, R.L., and Young, R.M., An Introduction to Econometric Forecasting and Forecasting Models, Lexington Books, Lexington, 1981. [4] Lee, M.S., and Olson D.L., "A gradient algorithm for chance constrained nonlinear goal programming", European Journal of Operational Research, 22 (1985) 359-369. [5] Rolji}, L., Duj{i}, M., "Nonlinear goal programming model application in optimal development policy choosing", Economic Analysis and Workers Management, 25 (1991). [6] Rolji}, L., "Interactive goal programming application in enterprises", Doctoral Dissertation, FON, University of Belgrade, 1987. [7] Rolji}, L., "Goal programming as a method for optimal decision making in enterprises", Economic Herald, Sarajevo, 24 (1-2) (1984) 35-48. [8] Rolji}, L., "Goal programming economic model", MS Thesis, University of Zagreb, 1981. [9] Van de Panne, C., and Popp, W., "Minimum cost cattle feed under probabilistic protein constraints", Management Science, 9 (3) (1963) 405-430. [10] Zoutendijk, G., Mathematical Programming Methods, North-Holland Publishing Co., 1976. 
work_xq2h2ueqobfi7hymgnxzhxz4t4	----	PII: S0924-0136(98)00182-4 Journal of Materials Processing Technology 80 – 81 (1998) 131 – 135 Expert system for hot forging design A. ngelo Caporalli a,b, Luciano Antonio Gileno a, Sérgio Tonini Button a,* a School of Mechanical Engineering, State Uni6ersity of Campinas, Campinas, SP, 13083 -970, CP 6122, Brazil b School of Engineering at Guaratinguetá, São Paulo State Uni6ersity, Guaratinguetá, SP, 12500 -000, CP 205, Brazil Abstract Planning hot forging processes is a time-consuming activity with high costs involved because of the trial-and-error iterative methods used to design dies and to choose equipment and process conditions. Some processes demand many months to produce forged parts with controlled shapes, dimensions and microstructure. This paper shows how expert systems can help engineers to reduce the time needed to design precision forged parts and dies from machined parts. The software ADHFD interfacing MS Visual Basic v.5.0 and SolidEdge v.3.0 was used to design flashless hot forged gears, chosen from families of gears. © 1998 Elsevier Science S.A. All rights reserved. Keywords: Expert system; Hot forging; Die design 1. Introduction In recent years, hot forging has shown continuous technological changes, mainly in the geometry of the dies and in the dimensional accuracy of forgings. Since hot forging have to be competitive even for small lot sizes, forgings weight and machining allowances are being reduced and therefore subsequent machining is being reduced with low final costs and investments. With hot precision forging and the concept of near-net- shape-manufacturing (NNS), costs with machining are drastically reduced, so forging operations (and their planning) became an important part of the final cost [1,2]. Most of the time (and costs) involved in planning hot forging processes is related to activities strongly depen- dent on human expertise, intuition and creativity [3] and also to iterative procedures involving extensive experimental work. In order to minimize this lead time, related expert knowledge and criteria must be trans- ferred to automated knowledge-based and rules-based systems. Tisza [4] states that there are a small number of expert systems applied in metal forming, when com- pared with other manufacturing fields. The numerous variables involved in metal forming processes make difficult the automation of decision making and plan- ning of these processes, but some examples can be found particularly for multi-stage processes like cold forging and deep drawing [3 – 7] and for hot forging [8,9]. Biswas and Knight [10] point out that the rationaliza- tion and automation of the procedures by expert sys- tems bring benefits to the hot forging industry like rationalized tooling, improved factory layouts, system- atic design procedures, improved estimating and costing procedures, and standardized work planning and pro- cess design. 2. Planning precision hot forging processes Precision hot forging shows some advantages when compared to conventional hot forging: it minimizes subsequent machining, improves forgings quality avoid- ing overlaps and internal defects, reduces forging loads, and can eliminate heat treatment, since internal stresses are reduced. To achieve these characteristics, process variables like forging temperature and billet volume must be precisely controlled [11]. The planning of a new hot forging process can be divided into three groups of procedures, as shown in Fig. 1. In group A the machined part is modified: * Corresponding author. Fax: + 55 19 2393722; e-mail: ser- gio@fem.unicamp.br 0924-0136/98/$19.00 © 1998 Elsevier Science S.A. All rights reserved. PII S 0 9 2 4 - 0 1 3 6 ( 9 8 ) 0 0 1 8 2 - 4 A. . Caporalli et al. / Journal of Materials Processing Technology 80 – 81 (1998) 131 – 135132 Fig. 1. Stages for hot forging planning. machining allowances are applied where necessary, a parting line is properly chosen, appropriate radii are applied to fillet and corners, internal and external drafts are added, holes, ribs and webs are designed to keep forging dies undeformed. The forged part is finally designed considering NNS criteria. The process sequence is designed in group B. Consid- ering part characteristics (material, size and geometry), lot size and available equipment, it is possible to define the number of operations and to design the preform. Using the principles of group technology and defining families of parts, it is possible to find similar forgings and adopt an existing sequence or modify it to the new design. The most difficult and costly stage to plan each operation is to determine the correct preform geometry to assure complete filling of the dies. In industry this stage is still based on ‘trial-and-error’ iterative proce- dures to the production of ‘good parts’, after successive measurements of the forgings and corrections in dies geometry. In recent years, many works based upon numerical methods like the finite element method (FEM) show that computerized analysis can help the forming industry to define the best process sequence, reducing time and costs involved [2,12,13]. Once the process sequence is defined, other process parameters are chosen in group C: forging tempera- tures, lubricants, and equipment. The dies are designed considering preforms geometry, thermal expansion, and the use of standardized tools and inserts. 3. The ADHFD system This work shows a semi-automatic system developed to design dies and plan precision hot forging using near-net-shape criteria. The ADHFD (automated de- sign of hot forging dies) is being developed to work with families of automotive gears (longitudinal sections shown in Fig. 2), classified by designs: with or without web, one-sided or two-sided hub, and by characteristic dimensions (external diameter, hub diameter, rim length, hub length). These gears are forged within some near-net-shape principles: small machining allowances, flashless, minimal draft angles and fillet radii [14]. The expert system ADHFD is being developed with four codes. The rule-based system is written in MS Visual Basic v.5.0 and linked to MS Access v.7.0 data-Fig. 2. Families of automotive gears. A. . Caporalli et al. / Journal of Materials Processing Technology 80 – 81 (1998) 131 – 135 133 Fig. 3. Automotive gear. Fig. 4. Gear forging. bases. Metal flow is simulated with Ansys v.5.3 to prevent filling defects and to optimize process sequence. With SolidEdge v.3.0 the forgings and the dies are modeled in 3-D and assembled. The volumes are then calculated and respective technical drawings generated. The system is considered semi-automatic since the user can choose similar forgings, analyse machined part details and decide how to modify the forging and the process sequence. The databases contain detailed information about variables of processes already planned: raw materials properties (stock size and tolerances, flow stress, formability), forging temperatures, lubricants, presses (capacity, table dimensions, stroke and speed), about NNS criteria and about typical process sequences (billet cut-off, heating, preform, finish forging and trimming). 4. An example of application The forging process of the gear shown in Fig. 3 was planned using the expert system developed. This gear is similar to the family A (Fig. 2) with web and one-sided hub. The teeth were previously removed since they will be cut by hobbing. In the first stage (group A — Fig. 1) after the analysis of the machined part, the machining allowance was defined equal to 2.0 mm, based on the biggest dimension (external diameter 158.36/158.11). This allowance was added to the surfaces with neces- sary subsequent machining (the hole, the external di- ameter, the faces of the hub and the faces of the rim). Then the fillet and corner radii were chosen. The 1° draft angle was defined to help removal of the forging from the die. The height and the wall thickness of the hub define the thickness of the internal web equal to 12.0 mm, large enough to prevent the damage of work- piece and dies. Finally, forging tolerances were defined to all dimensions. Fig. 4 shows the final forging. In group B, after searching for a similar forging, it was chosen to modify an existing forming sequence with four operations: preforming, blocking, finishing and trimming (Fig. 5).Fig. 5. Process sequence. A. . Caporalli et al. / Journal of Materials Processing Technology 80 – 81 (1998) 131 – 135134 Fig. 6. Simulated evolution of the blocking operation. The dimensions of the billet were calculated from the volume of the final forging. The best shape and dimen- sions of the preform were defined from the results of simulations with FEM program ANSYS. As shown in Fig. 6 an incorrect billet shape causes the incomplete filling of the dies in the blocking operation and internal defects in the forged part. Finally (group C) the dies were designed considering the dimensions of the parts in each operation, added with 1% because of the thermal expansion verified at forging temperatures. Table 1 shows process parame- ters chosen from databases and related to material properties, the number of operations and the results of simulation. 5. Conclusions The expert system ADHFD showed good perfor- mance when aiding the design of flashless hot forged gears. The design time is considerably reduced when compared with conventional techniques. Some modifi- cations have to be done in the program to allow the design of asymmetric parts with complex geometric details. Acknowledgements FAPESP (Fundação de Amparo à Pesquisa do Es- tado de São Paulo) and CAPES (Coordenação de Aperfeiçoamento de Pessoal de Nı́vel Superior) spon- sored this project and are gratefully acknowledged. References [1] H. Kudo, Towards net-shape forming, J. Mater. Process. Tech- nol. 22 (1990) 307 – 342. [2] A.G. Mamalis, D.E. Manolakos, A.K. Baldoukas, Simulation of the precision forging of bevel gears using implicit and explicit FE techniques, J. Mater. Proces. Technol. 57 (1996) 164 – 171. [3] H. Kim, T. Altan, K. Sevenler, Computer-aided part and pro- cessing-sequence design in cold forging, J. Mater. Process. Tech- nol. 33 (1992) 57 – 74. [4] M. Tisza, Expert systems for metal forming, J. Mater. Process. Technol. 53 (1995) 423 – 432. Table 1 Process parameters Billet dimensions ¥25×35 mm Initial forging temperature 1100°C 950°CFinal forging temperature Press Eccentric forging press 2000×104 NCapacity Stroke 280 mm Stroking rate 80 min−1 1090×1350 mmTable dimensions Colloidal graphite in oilLubricant A. . Caporalli et al. / Journal of Materials Processing Technology 80 – 81 (1998) 131 – 135 135 [5] G. Yang, K. Osakada, A review of expert systems for process planning of cold forging, Manufact. Rev. 6 (2) (1993) 101 – 113. [6] A. Azushima, M. Kim, A PC-based expert system for forming sequence design, Ann. CIRP 39 (1) (1990) 245 – 248. [7] P. Bariani, E. Benuzzi, W.A. Knight, Computer aided design of multi-stage cold forging processes: load peaks and strain distribution evaluation, Ann. CIRP 36 (1) (1987) 145 – 148. [8] T. Penchev, Development of an expert system for hot die by breakdown of the parts into groups, J. Mater. Process. Tech- nol. 30 (1992) 285 – 287. [9] D. Glynn, G. Lyons, J. Monaghan, Forging sequence design using an expert system, J. Mater. Process. Technol. 55 (1995) 95 – 102. [10] S.K. Biswas, W.A. Knight, Towards an integrated design and production system for hot forging dies, Int. J. Prod. Res. 14 (1) (1976) 23 – 49. [11] E. Doedge, J. Thalemann, F. Weber, Hot forging of precision parts, J. Mater. Process. Technol. 35 (1992) 469 – 481. [12] J. Kusiak, A technique of tool-shape optimization in large scale problems of metal forming, J. Mater. Process. Technol. 57 (1996) 79 – 84. [13] L. Fourment, T. Balan, J.-L. Chenot, Shape optimal design in forging, in: S.-F. Shen and P. Dawson (Eds.), Simulation of Materials Processing: Theory, Methods and Applications, NU- MIFORM’95, Balkema, Rotterdam, 1995, 557 – 562. [14] R.J. Shipley, Precision forging, in: Metals Handbook, ASM Intl., vol. 14, 1988, 150 – 157. . 
work_xq42u5f77na3pps3pym35robem	----	Empirical distributions of daily equity index returns: A comparison Expert Systems With Applications 54 (2016) 170–192 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Empirical distributions of daily equity index returns: A comparison Canan G. Corlu a, Melike Meterelliyoz b,∗, Murat Tiniç c a Metropolitan College, Boston University, Boston, MA 02215 USA b Department of Business Administration, TOBB University of Economics and Technology, Ankara 06560 Turkey c Department of Management, Bilkent University, Ankara 06800 Turkey a r t i c l e i n f o Keywords: Index returns Generalized lambda Johnson translation system Skewed-t Normal inverse Gaussian g-and-h a b s t r a c t The normality assumption concerning the distribution of equity returns has long been challenged both empirically and theoretically. Alternative distributions have been proposed to better capture the char- acteristics of equity return data. This paper investigates the ability of five alternative distributions to represent the behavior of daily equity index returns over the period 1979–2014: the skewed Student-t distribution, the generalized lambda distribution, the Johnson system of distributions, the normal inverse Gaussian distribution, and the g-and-h distribution. We find that the generalized lambda distribution is a prominent alternative for modeling the behavior of daily equity index returns. © 2016 Elsevier Ltd. All rights reserved. r t ( t a P e t e m G a t m ( t r m N a t d t t 1. Introduction The assumption that stock price changes follow a stable distri- bution forms the basis for major asset pricing and option pricing models. Early models by Bachelier (1900) take normality as a fundamental assumption for modeling stock price movements. In line with this assumption, Osborne (1959) shows that logarithms of the changes in the stock prices are mutually independent with a common probability distribution (i.e., they conform to a random walk). He then suggests that stock price changes must follow a normal distribution. However, these findings have been challenged both theoretically and empirically.1 An early work by Mandelbrot (1967) proposes that stock price returns belong to the family of stable Paretian distributions be- cause they have fatter tails. Fama (1963; 1965) provides empirical evidence that supports this claim and demonstrates that stock price changes indeed have fatter tails and have higher peaks than the normal distribution. More recently, Rachev, Stoyanov, Biglova, and Fabozzi (2005) compared the stable Paretian distribution to the normal distribution using 382 US stock returns over the period 1992–2003. The authors investigated the daily returns using two probability models: the homoskedastic independent and identically distributed model and the conditional heteroskedastic ARMA-GARCH model. Normality was rejected for both models. However, Officer (1972) found that normality holds for monthly ∗ Corresponding author. Tel.: +90 312 2924214. E-mail addresses: canan@bu.edu (C.G. Corlu), mkuyzu@etu.edu.tr (M. Meterel- liyoz), tinic@bilkent.edu.tr (M. Tiniç). 1 A similar line of argument holds for exchange rate returns. n t y & http://dx.doi.org/10.1016/j.eswa.2015.12.048 0957-4174/© 2016 Elsevier Ltd. All rights reserved. eturns and that the standard deviation of the returns is inconsis- ent with the stable hypothesis. To support this argument, Praetz 1972) then suggested the Student-t distribution as an alternative o the stable Paretian because the stable Paretian distribution has n infinite variance property and the density function of the stable aretian is unknown. Over an eight-year period, Praetz (1972) xamined weekly data from Sydney Stock Exchange and showed hat the Student-t distribution can be used as an alternative to xplain the stock price behavior. The Student-t distribution was also compared with the nor- al distribution and the Cauchy distribution by Blattberg and onedes (1974). Contrary to Praetz (1972), they used both daily nd weekly returns of stocks of the Dow Jones Industrial (DJI), and hey used the maximum likelihood estimation method for esti- ating the parameters of the distributions. Blattberg and Gonedes 1974) showed that the Student-t distribution performs better than he normal distribution on daily returns. However, normality is not ejected for monthly return data. Hagerman (1978) tested the nor- ality hypothesis on both individual stocks of the American and ew York Stock Exchanges on portfolios that contain these stocks, nd found that they do not behave in line with the normal dis- ribution. Hagerman (1978) proposed that the mixture of normal istributions and the Student-t distribution can be an alternative o representing the characteristics of stock return data. However, he performance of these two distributions against each other was ot investigated in Hagerman’s work. Kon (1984) compared the discrete mixture of normal distribu- ions and the Student-t distribution over a period of almost 19 ears, examining daily returns of 30 stocks from DJI and Standard Poor’s (S&P) value- and equal-weighted stock market indexes. A http://dx.doi.org/10.1016/j.eswa.2015.12.048 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.12.048&domain=pdf mailto:canan@bu.edu mailto:mkuyzu@etu.edu.tr mailto:tinic@bilkent.edu.tr http://dx.doi.org/10.1016/j.eswa.2015.12.048 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 171 d v l O b t n ( w e G t a r 1 a t t 3 a T t o t p g t n P t s t r t t N w m L a o w t b ( t u f t w I w w b p d 5 t t G o g m d r c d S g S s p P o d m i t o d r b t o p o d i d i t s v A t w i a m p t t S 2 e d c 2 S e I ( i 1 A iscrete mixture of normal distributions is shown to have greater alidity than the Student-t distribution in modeling the data. Simi- ar to Blattberg and Gonedes (1974) and Akgiray and Booth (1987); fficer (1972) stated that the monthly returns of stock prices can e assumed to be normally distributed. However, for the daily data hey found that the mixed diffusion process and the mixture of ormal distributions perform better than the stable distributions. Bookstaber and McDonald (1987) proposed the generalized beta GB2) distribution to explain the behavior of stock returns. This as chosen because the GB2 is a flexible distribution and acknowl- dges various distributions as special cases. They found that the B2 distribution is significantly better than the lognormal dis- ribution, especially in relation to short time intervals. Badrinath nd Chatterjee (1988) examined the Center for Research in Secu- ity Prices (CRSP) value-weighted market index returns between 962 and 1985 and concluded that returns of stock prices follow skewed g-and-h distribution.2 Similarly, Mills (1995) found that he g-and-h distribution accurately fits a dataset that consists of hree London Stock Exchange indices: FTSE 100, Mid 250, and FTSE 50. A more general comparison of distributions with finite vari- nces over equity stocks was conducted by Gray and French (1990). hey compared the scaled-t distribution, the logistic distribution, he exponential power distribution, and the normal distribution ver the log-returns of daily S&P 500 Composite index values for he period 1979–1985. Among four alternatives, the exponential ower distribution was found to be the best fit. Lau, Lau, and Win- ender (1990) showed that series of returns of stock prices that are aken from the CRSP yield higher kurtosis and skewness than the ormal distribution. They proposed the lognormal, beta, Weibull, earson Types IV and VI, and Johnson system of distributions as al- ernatives. A general comparison of the normal distribution to the caled-t distribution and to the mixture of two normal distribu- ions was conducted by Aparicio and Estrada (2001) using the daily eturns of 13 different European stock markets. It was found that he scaled-t distribution is a significantly better fit for the data, and he partial mixture of two normal distributions also performs well. ormality is rejected in all cases. Linden (2001) introduced the Laplace mixture distribution, hich is derived by conditioning the standard deviation of the nor- al distribution as an exponentially distributed random variable. inden (2001) used this distribution to represent the daily, weekly, nd monthly returns of the 20 most traded shares and the index f the Helsinki Stock Market. The normality assumption is not al- ays rejected for the weekly and monthly returns. However, for he daily returns, an asymmetric Laplace distribution is found to e a better candidate than the normal distribution. Harris and Küçüközmen (2001a) and Harris and Küçüközmen 2001b), respectively, examined the skewed generalized-t distribu- ion (SGT) and the exponential generalized beta distribution (EGB) sing daily UK, US, and Turkish equity returns. Consequently, they ound that the SGT outperforms the EGB. In both studies, the au- hors rejected the hypothesis that the daily returns are distributed ith the Student-t, power of exponential, or logistic distribution. n addition, for the daily Turkish returns, the Laplace distribution as also rejected. For the UK returns, the skewed-t distribution as preferred, whereas for the US returns, the generalized-t distri- ution was preferred. More recently, Behr and Pötter (2009) com- ared the generalized hyperbolic distribution, the generalized logF istribution, and the finite mixture of Gaussians on monthly S&P 00 index returns over the years 1871–2005 and daily returns over he years 2001–2005. For the monthly returns, the two-component 2 Badrinath and Chatterjee (1988) also provide an excellent review of the litera- ure. 4 h w aussian mixture distribution described the empirical distribution f the returns better than alternative distributions. Although the eneralized hyperbolic distribution is the poorest performer for onthly returns, it performs best for daily data. However, as the aily data examined by Behr and Pötter (2009) is almost symmet- ic, the Laplace distribution, which does not have a parameter to apture the asymmetries, fits as well as the generalized hyperbolic istribution. Finally, as an alternative to the stable distribution and the tudent-t distribution, Chalabi, Scott, and Würtz (2010) use the eneralized lambda distribution (GLD) for modeling equity returns. tarting with Eberlein and Keller (1995), the normal inverse Gaus- ian (NIG) distribution is used to model financial returns and articularly for modeling 30 stocks at the German Stock Index. rause, Zentrum, and Modellbildung (1997) show the applicability f the NIG distribution in modeling German stock and US Stock In- ex data. Bølviken and Benth (2000) used the NIG distribution to odel 8 Norwegian stocks. In Table 1, we summarize the papers that performed compar- son studies to investigate the behavior of stock returns. We find hat the outcomes differ and are often conflicting. Based upon this, ur goal in this study is to fill this gap in the literature by ad- ressing which distribution is best for modeling daily equity index eturn data. To this end, we consider the following flexible distri- utions that are commonly used in finance: the skewed Student- distribution, the GLD, the NIG distribution, the Johnson system f distributions, and the g-and-h distribution. We conduct a com- rehensive numerical analysis to compare the overall suitability f these five distributions on the equity index returns of twenty ifferent countries over the period 1979–2014, which is divided nto twelve three-year sub-periods. We also include the normal istribution in our experimental design. The overall suitability is nitially compared using the Kolmogorov–Smirnov (KS) test statis- ic (Chakravarti & Laha, 1967) and the Anderson–Darling (AD) test tatistic (Anderson & Darling, 1954). Furthermore, we conduct p- alue tests in order to assess the significance of these KS and D statistics. In addition, the explanatory power of the models is ested using in-sample Value-at-Risk (VaR) failure rates. Consistent ith other studies in the previous research, we find that normal- ty is rejected in all sub-periods for all markets. Our p-value tests nd the in-sample VaR test suggest that GLD performs best for all arkets over all time periods. The remainder of the paper is organized as follows. Section 2 resents the data. Section 3 presents the distributions along with he fitting methods that are used to estimate the parameters of he distributions. Section 4 discusses our numerical study and ection 5 presents key conclusions. . Description of the data We create a diversified sample from ten developed and ten merging market indexes. The selected developed stock market in- exes are: S&P/ASX 200 Index (Australia), S&P/Toronto Stock Ex- hange Index (Canada), CAC 40 (France), DAX (Germany), NIKKEI 25 (Japan), the Straits Times Index (Singapore), IBEX 35 (Spain), MI (Switzerland), FTSE 100 (UK), and S&P 500 (US), while the merging stock market indexes are the Ibovespa Index (Brazil), PSA Index (Chile), SHSZ 300 (China), BSE 500 (India), KOSPI Index Korea), FBMKLCI Index (Malaysia), the Mexican IPC Index (Mex- co), MICEX Index (Russia), JALSH Index (South Africa), and BIST 00 (Turkey). The daily closing index levels from January 1979 to ugust 2014 are collected using the Bloomberg Terminal. Bloomberg provides index levels for the S&P/ASX 200, CAC 0, DAX and IBEX 35 prior to their establishment date. This can appen due to two reasons. First, the index levels can be adjusted ith respect to their ancestor indices. For instance, the DAX 172 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 Table 1 Summary of articles that performed comparison studies to investigate the behavior of stock returns. The first column of this table lists the articles that compare several distributions for modeling stock returns. The second column of the table presents the distributions considered in the corresponding study. Article Compared distributions Fama (1963), Fama (1965) and Rachev et al. (2005) Stable Paretian∗ , normal Praetz (1972) Scaled-t∗ , compound process , normal Hsu, Miller, and Wichern (1974) Normal∗ , stable Paretian Blattberg and Gonedes (1974) Student-t∗ , Cauchy, normal Kon (1984) Discrete mixture of normals∗ , Student-t Eberlein and Keller (1995) Normal inverse Gaussian∗ , normal Akgiray and Booth (1987) Mixed diffusion process∗ , discrete mixture of normals Bookstaber and McDonald (1987) Generalized beta∗ , lognormal Gray and French (1990) Exponential power∗ , logistic, Scaled-t, normal Aparicio and Estrada (2001) Scaled-t∗ , mixture of normal, normal Linden (2001) Asymmetric Laplace∗ , normal Harris and Küçüközmen (2001a), (2001b) SGT∗ , EGB, exponential power, logistic, Student-t, Laplace Behr and Pötter (2009) Finite mixture of Gaussians∗ , generalized hyperbolic∗∗ , generalized logF ∗ represents the outperforming distribution. ∗∗ In Behr and Pötter (2009), a mixture of two Gaussians describes the monthly data better than the others, whereas a generalized hyperbolic distribution is the best fit for the daily data. Table 2 Markets considered in each sub-period. This table presents the emerging and developed markets consid- ered in each sub-period. For instance, in sub-period 1979–1981, data are available only for the developed markets Canada, Germany, Japan, and the US and the emerging markets Korea and Malaysia. Beginning from the sub-period 2000–2002, we have access to data for all ten developed and ten emerging markets. Periods Developed markets Emerging markets 1979–1981 Canada, Germany, Japan, US Korea, Malaysia 1982–1984 Canada, Germany, Japan, UK, US Korea, Malaysia 1985–1987 Canada, France, Germany, Japan, Spain, UK, US Korea, Malaysia 1988–1990 Canada, France, Germany, Japan, Chile, Korea, Malaysia, Turkey Spain, Switzerland, UK, US 1991–1993 Australia, Canada, France, Germany Brazil, Chile, Korea, Malaysia, Turkey Japan, Spain, Switzerland, UK, US 1994–1996 Australia, Canada, France, Germany Brazil, Chile, Korea, Malaysia Japan, Spain, Switzerland, UK, US Mexico, South Africa, Turkey 1997–1999 Australia, Canada, France, Germany, Japan Brazil, Chile, India, Korea, Malaysia Singapore, Spain, Switzerland, UK, US Mexico, Russia, South Africa, Turkey 2000–2014 All markets All markets n i T b t c d a r i m 1 e t t h n 1 t n d f d i follows its ancestor indices Börsenzeitungs Index and Hardy In- dex.3 Second, the index levels can be recalculated to an earlier date. For instance, the CAC 40 is established in December 31, 1987; however, CAC 40 index levels are recalculated on a daily basis to September 7, 1987 (Shilling, 1996). Similarly, the S&P/ASX 200 is established in April 3, 2000; however, S&P/ASX 200 index levels are recalculated to May, 1992.4 IBEX 35 is launched in January 14, 1992 and IBEX 35 index levels are recalculated to January 5, 1987 (Fernandez, Aguirreamalloa, & Avendaño, 2011). The daily logarithmic returns are calculated as Xt = log(St /St−1 ), where St is the closing index level at time t. We divide the data for each market into twelve three-year nonover- lapping sub-periods. Our sample period contains both regional and global financial crises such as the 1997 Asian financial crisis, the Global Financial Crisis (2006–2008), and the Euro-zone Sovereign Debt Crisis (2009–2011). Unfortunately, return data is not available for every market during all of these sub-periods. In Table 2, we list data that are available for the developed and the emerging markets in each sub-period. For example, in sub-period 1979– 1981, data are available only for the developed markets Canada, Germany, Japan, and the US and the emerging markets Korea and Malaysia. Beginning from the sub-period 2000–2002, we have access to data for all ten developed and ten emerging markets. Fig. 1 presents the descriptive statistics for the developed mar- kets and the emerging markets over different sub-periods. The 3 Based on communications with Deutsche Börse Group. 4 Based on communications with Standard & Poor’s Dow Jones Indices. a umber in each circle represents the sub-period number, where 1 s for the sub-period 1979–1981 and 13 is for the entire sample. he skewness equals to s = (m3/m2 )3/2 and the kurtosis is given y k = m4/m22, where mi is the estimate of the ith moment around he mean. The skewness values of the returns in the figure are lustered either on the right or left hand-side of the vertical line rawn at zero, indicating that all returns of all indices demonstrate skewed (either left or right) behavior. Similarly, the kurtosis of eturns are all clustered above the line drawn at three (some be- ng very far from three) meaning that the return distributions are ore peaked than the normal distribution. For the developed markets in Fig. 1, the fourth sub-period (i.e., 988–1990) and the thirteenth sub-period (i.e., the entire sample) xhibit skewness and kurtosis values that are significantly far from he skewness and kurtosis values of the normal distribution. For he emerging markets, there are more sub-periods that exhibit igh skewness and kurtosis values, such as fourth, seventh, eighth, inth, twelfth, and thirteenth sub-periods (i.e., 1988–1990, 1997– 999, 2000–2002, 2003–2005, 2012–2014, entire sample). These wo figures alone indeed suggest that the normal distribution is ot a good model for describing the distribution of daily equity in- ex returns. Additionally, we calculate the Bera-Jarque (BJ) statistic or all of the countries over all sub-samples in order to test for any eparture from normality. The results indicate that the normality s rejected in 96% of the instances, at the 1% significance level.5 5 To save space, we do not present the BJ test statistics here. However, they are vailable upon request. C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 173 Fig. 1. Skewness and kurtosis values of each market over all sub-samples and the entire sample. The top figure presents the skewness (on the horizontal axis) and the kurtosis (on the vertical axis) values of each developed market over all sub-samples and the entire sample. The bottom figure does the same for emerging markets. The number in each circle represents the sub-period number with 1 presenting the sub-period 1979–1981 and 13 presenting the entire sample. The skewness equals to s = (m3 /m2 ) 3/2 and the kurtosis is given by k = m4 /m22 , where mi is the estimate of the ith moment around the mean. 3 S s G b 3 S b t F w p s G p p d ( a . Flexible distributions We describe the generalized lambda distribution (GLD) in ection 3.1, the Johnson translation system in Section 3.2, the kewed Student-t distribution in Section 3.3, the normal inverse aussian (NIG) distribution in Section 3.4, and the g-and-h distri- ution in Section 3.5. .1. The Generalized Lambda distribution The GLD (Filliben, 1975; Joiner & Rosenblatt, 1971; Ramberg & chmeiser, 1974), which is an extension of Tukey’s lambda distri- ution (Hastings, Mosteller, Tukey, & Winsor, 1947), is defined by he following inverse cumulative distribution function: −1(u; λ1, λ2, λ3, λ4 ) = λ1 + uλ3 − (1 − u)λ4 λ2 , (3.1) here 0 ≤ u ≤ 1; λ1 is the location parameter, λ2 is the scale arameter, and λ3 and λ4 are related to skewness and kurtosis, re- pectively. This representation is denoted as Ramberg–Schmeiser eneralized Lambda Distribution (RS GLD) in reference to the arametrization of Ramberg and Schmeiser (1974). However, the robability density function (pdf) associated with equation (3.1) oes not provide a valid pdf for all combinations of λ3 and λ4 Fournier et al., 2007). In order to avoid this problem, Freimer, Kollia, Mudholkar, nd Lin (1988) proposed a different parametrization of the GLD, 174 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 A a w r d f o t b h 3 t d w r d β f δ P r t t d d w t 3 s d d Y denoted as Freimer–Mudholkar–Kollia–Lin Generalized Lambda Distribution (FMKL GLD). This is given by: F −1(u; λ1, λ2, λ3, λ4 ) = λ1 + 1 λ2 ( uλ3 − 1 λ3 − (1 − u) λ4 − 1 λ4 ) . Both the FMKL and RS representations are used in practice as they offer a wide variety of shapes. However, the FMKL representa- tion is preferable as a result of its ease of use. In addition, it is well defined over all parameter values with the only restriction being that λ2 must be positive. Also, the condition min(λ3, λ4 ) > −1/k must hold to have a finite kth moment. In this study, we use the FMKL representation. The FMKL GLD curves are classified into five categories: Class-I family (λ3 < 1, λ4 < 1) represents unimodal densities with continuous tails, Class-II family (λ3 > 1, λ4 < 1) represents monotone pdfs similar to the exponential distribution, Class-III family ( 1 < λ3 < 2, 1 < λ4 < 2) represents U-shaped densities with truncated tails, Class-IV family (λ3 > 2, 1 < λ4 < 2) represents S-shaped densities, and finally Class-V family (λ3 > 2, λ4 > 2) represents unimodal densities with truncated tails. We find that the daily equity return data belongs to the Class-I family. Prior studies propose different fitting techniques for estimat- ing the parameters of the GLD.6 In this study, we use the maxi- mum likelihood estimation method in the GLDEX package of R (Su, 2007). 3.2. The Johnson translation system A random variable X from the Johnson translation system is represented by (Johnson, 1949): X = ξ + λr−1 ( Z − γ δ ) , where Z is a standard normal random variable, γ and δ are shape parameters, ξ is a location parameter, λ is a scale parameter, and r( · ) is one of the following transformations: r(X ) = ⎧⎪⎨ ⎪⎩ x for the SN (normal) family log(x) for the SL (lognormal) family log(x/(1 − x)) for the SB (bounded) family log(x + √ x2 + 1) for the SU (unbounded) family. The range of the random variable X is defined by the family of in- terest: X > ξ and λ = 1 for the SL family; ξ < X < ξ + λ for the SB family; and −∞ < X < ∞ for the SN and SU families. For each fea- sible combination of the skewness and the kurtosis values there is a unique family that depends on the choice of r. In this study, we only consider the SU family of the Johnson translation system be- cause the equity index returns have skewness and kurtosis values that conform with the SU family. There are alternative methodologies proposed in previous re- search to estimate the parameters of the SU. 7 We use the method- ology proposed in Tuenter (2001), which resembles the method of moments, where the sample skewness and kurtosis are equalized with the theoretical skewness and kurtosis. 3.3. The skewed Student-t distribution A number of skewed Student-t distributions have been pro- posed in previous research. In this study, due to its simplicity, we follow the parametrization used in Azzalini and Capitanio (2003). 6 A detailed review can be found in Corlu and Meterelliyoz (2015). 7 A detailed review can be found in Corlu and Biller (2015). w k r random variable X from the skewed Student-t distribution (here- fter, skewed-t distribution) has a density of the form: f (x; δ, ν, μ, β ) = 1 δ tν ( x − μ δ ) 2Tν+1 ( β ( x − μ δ )√ ν + 1( x−μ δ )2 + ν ) , here μ, δ, and β represent the location, scale, and skewness pa- ameters, respectively. tν is the density of the standard Student-t istribution with ν degrees of freedom and Tν+1 is the distribution unction of the standard Student-t distribution with ν + 1 degrees f freedom. Estimates of the parameters of the skewed-t distribution are ob- ained using the maximum likelihood estimation method proposed y Azzalini and Capitanio (2003). The maximization of the likeli- ood function is conducted by the Nelder–Mead algorithm. .4. The Normal Inverse Gaussian distribution The NIG distribution is obtained from a more general distribu- ion called the generalized hyperbolic distribution (GHD), whose ensity is given by: f (x; λ, α, β, μ, δ) = ( δ √ α2 − β2 )λ( δα )1/2−λ √ 2πδKλ ( δ √ α2 − β2 ) ( 1 + (x − μ) 2 δ2 )λ/2−1/4 × exp(β(x − μ))Kλ−1/2 ( αδ √ 1 + (x − μ) 2 δ2 ) , here Kλ is the modified third-order Bessel function; μ and δ rep- esent location and scale parameters, respectively; λ is the class- efining parameter; α is a parameter related to tail heaviness; and is the asymmetry parameter. The density is defined under the ollowing parameter restrictions: δ ≥ 0 and |β| < α if λ > 0 > 0 and |β| < α if λ = 0 δ > 0 and |β| ≤ α if λ < 0 faff, McNeil, and Ulmann (2013) indicate that the GHD can rep- esent skewed distributions as well as heavy tails. The variants of he GHD can be obtained by changing the value of the parame- er λ; this is why λ is called the class-defining parameter. The NIG istribution is obtained from the GHD by setting λ = −1/2, and its ensity is given by: f (x; α, β, μ, δ) = αδK1 ( α √ δ2 + (x − μ)2 ) π √ δ2 + (x − μ)2 eδ √ α2 −β2 +β(x−μ) , ith |β| ≤ α and δ > 0 (Prause, 1999). We use the ghyp package of R to estimate the parameters of he NIG. .5. The g-and-h distribution The g-and-h distribution is a functional transformation of the tandard normal distribution. A random variable from the g-and-h istribution is obtained by transforming the standard normal ran- om variable Z to the following form (Tukey, 1977): g,h(Z) = ( egZ − 1 ) exp (hZ2/2) g , here g, h ∈ R. The g and h parameters account for skewness and urtosis, respectively. When a location parameter A and a scale pa- ameter B are incorporated, a random variable from the g-and-h C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 175 1 9 7 9 − 1 9 8 1 1 9 8 2 − 1 9 8 4 1 9 8 5 − 1 9 8 7 1 9 8 8 − 1 9 9 0 1 9 9 1 − 1 9 9 3 1 9 9 4 − 1 9 9 6 1 9 9 7 − 1 9 9 9 2 0 0 0 − 2 0 0 2 2 0 0 3 − 2 0 0 5 2 0 0 6 − 2 0 0 8 2 0 0 9 − 2 0 1 1 2 0 1 2 − 2 0 1 4 E n ti re S a m p le 0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 Canada Skewed−t GLD NIG Johnson Su g−and−h Normal 1 9 7 9 − 1 9 8 1 1 9 8 2 − 1 9 8 4 1 9 8 5 − 1 9 8 7 1 9 8 8 − 1 9 9 0 1 9 9 1 − 1 9 9 3 1 9 9 4 − 1 9 9 6 1 9 9 7 − 1 9 9 9 2 0 0 0 − 2 0 0 2 2 0 0 3 − 2 0 0 5 2 0 0 6 − 2 0 0 8 2 0 0 9 − 2 0 1 1 2 0 1 2 − 2 0 1 4 E n ti re S a m p le 0.00 0.02 0.04 0.06 0.08 0.10 Germany Skewed−t GLD NIG Johnson Su g−and−h Normal Fig. 2. Kolmogorov–Smirnov (KS) statistics of all six probability distributions for Canada and Germany. The top figure shows the KS statistics of the skewed-t distribution, generalized lambda distribution (GLD), normal inverse Gaussian (NIG) distribution, Johnson SU family, g-and-h distribution, and normal distribution for Canada over all sub- periods starting with 1979–1981. The last tick mark on the horizontal axis presents the KS statistic for the entire sample over the period 1979–2014. The bottom plot does the same for Germany. d X f l ( c g g w b i o n d X X S t istribution takes the following form: g,h(Z) = A + B ( egZ − 1 ) exp (hZ2/2) g . (3.2) Among the alternative fitting methodologies that are proposed or estimating the parameters of the g-and-h distribution, we fol- ow the procedure used in Mills (1995) and Dutta and Babbel 2002) as a result of its tractability. Specifically, the estimation pro- edure starts with identifying the pth percentile of the parameter as follows: p = − ( 1 Zp ) ln ( X1−p − X0.5 X0.5 − Xp ) , (3.3) here Xp and Zp are the pth percentile of the empirical distri- ution and the standard normal distribution, respectively. It is mportant to note that by using different values of p, one can btain multiple estimates of g. Following (Mills, 1995), we use ine percentiles and we choose percentiles using letter values; i.e., p = 1/2, 1/4, 1/16, . . .. The estimate for g is calculated as the me- ian of gp values. Using equation (3.2), it is trivial to derive p = A + B ( egZp − 1 ) exp (hZ2p/2) g and (3.4) 1−p = A + B ( egZ(1−p) − 1 ) exp (hZ2 (1−p)/2 ) g . (3.5) ince X0.5 = A, the location parameter A is given by the median of he data set. 176 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 1 9 9 4 − 1 9 9 6 1 9 9 7 − 1 9 9 9 2 0 0 0 − 2 0 0 2 2 0 0 3 − 2 0 0 5 2 0 0 6 − 2 0 0 8 2 0 0 9 − 2 0 1 1 2 0 1 2 − 2 0 1 4 E n ti re S a m p le 0.00 0.02 0.04 0.06 0.08 Mexico Skewed−t GLD NIG Johnson Su g−and−h Normal 1 9 8 8 − 1 9 9 0 1 9 9 1 − 1 9 9 3 1 9 9 4 − 1 9 9 6 1 9 9 7 − 1 9 9 9 2 0 0 0 − 2 0 0 2 2 0 0 3 − 2 0 0 5 2 0 0 6 − 2 0 0 8 2 0 0 9 − 2 0 1 1 2 0 1 2 − 2 0 1 4 E n ti re S a m p le 0.00 0.01 0.02 0.03 0.04 0.05 0.06 Turkey Skewed−t GLD NIG Johnson Su g−and−h Normal Fig. 3. Kolmogorov–Smirnov (KS) statistics of all six probability distributions for Mexico and Turkey. The top figure shows the KS statistics of the skewed-t distribution, generalized lambda distribution (GLD), normal inverse Gaussian (NIG) distribution, Johnson SU family, g-and-h distribution, and normal distribution for Mexico over all sub- periods starting with 1994–1996. The last tick mark on the horizontal axis presents the KS statistic for the entire sample over the period 1994–2014. The bottom plot does the same for Turkey over all sub-periods starting with 1988–1990. t 4 o m K t S 4 c Subtracting (3.5) from (3.4) and letting Zp = −Z1−p provides the following result: ln g ( Xp − X1−p ) egZp − e−gZp = ln(B) + h(Z 2 p/2). (3.6) If the data is positively skewed, the left-hand side of (3.6) can be replaced with the upper half-spread (UHS), as defined in Hoaglin (2006): UHS = g(X1−p − X0.5 )( e−gZp − 1 ) (3.7) Conversely, if the data is negatively skewed, then a lower half- spread (LHS) can be used on the left-hand side of (3.6): LHS = g(X0.5 − Xp)( 1 − egZp ) (3.8) The estimates of h and B are obtained from the coefficient and he intercept of the linear regression of In(UHS) and In(LHS) on (Z2p/2). . Performance and risk estimation The goal of this section is to evaluate the performances f the density functions described in Section 3 in order to odel the daily equity index returns. To this end, we use the olmogorov–Smirnov (KS) and Anderson–Darling (AD) test statis- ics in Section 4.1 and we present the Value-at-Risk (VaR) levels in ection 4.2. .1. Goodness-of-fit In this section, we first use the KS and AD test statistics to ompare the goodness-of-fit of the distributions of interest. Both C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 177 1 9 7 9 − 1 9 8 1 1 9 8 2 − 1 9 8 4 1 9 8 5 − 1 9 8 7 1 9 8 8 − 1 9 9 0 1 9 9 1 − 1 9 9 3 1 9 9 4 − 1 9 9 6 1 9 9 7 − 1 9 9 9 2 0 0 0 − 2 0 0 2 2 0 0 3 − 2 0 0 5 2 0 0 6 − 2 0 0 8 2 0 0 9 − 2 0 1 1 2 0 1 2 − 2 0 1 4 E n ti re S a m p le 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Canada Skewed−t GLD NIG Johnson Su g−and−h 1 9 7 9 − 1 9 8 1 1 9 8 2 − 1 9 8 4 1 9 8 5 − 1 9 8 7 1 9 8 8 − 1 9 9 0 1 9 9 1 − 1 9 9 3 1 9 9 4 − 1 9 9 6 1 9 9 7 − 1 9 9 9 2 0 0 0 − 2 0 0 2 2 0 0 3 − 2 0 0 5 2 0 0 6 − 2 0 0 8 2 0 0 9 − 2 0 1 1 2 0 1 2 − 2 0 1 4 E n ti re S a m p le 0.0 0.5 1.0 1.5 2.0 Germany Skewed−t GLD NIG Johnson Su g−and−h Fig. 4. Anderson–Darling (AD) statistics of five probability distributions for Canada and Germany. The top figure shows the AD statistics of the skewed-t distribution, gen- eralized lambda distribution (GLD), normal inverse Gaussian (NIG) distribution, Johnson SU family, and g-and-h distribution for Canada over all sub-periods starting with 1979–1981. The last tick mark on the horizontal axis presents the AD statistic for the entire sample over the period 1979–2014. The bottom plot does the same for Germany. o t l c i t w c s m s i t t e h a s t b A w d f e a g t b t f these test statistics summarize the difference between the fit- ed cumulative distribution function F̂ and the empirical cumu- ative distribution function Fe. In particular, the KS test statistic orresponds to the largest distance between Fe(x) and F̂ (x), that s, supx {|Fe (x) − F̂ (x)|}, while the AD test statistic corresponds to he weighted average of the squared differences, (Fe (x) − F̂ (x))2, here the weights are chosen in such a way that the discrepan- ies in the tails are emphasized. The smaller the KS and AD test tatistics, the better the fit. Fig. 2 presents the computed KS test statistics for the developed arkets of Canada and Germany under the six distributional as- umptions. Fig. 3 does the same for the emerging markets of Mex- co and Turkey. Fig. 4 presents the computed AD test statistics for he developed markets of Canada and Germany using all distribu- ions except the normal distribution. Fig. 5 does the same for the merging markets of Mexico and Turkey. The markets on the plots ave been selected as they are the most representative of the over- ll results. We excluded the normal distribution from Figs. 4 and 5 ince the AD test statistic values for the normal distribution are oo large compared to the respective results for all other distri- utions under consideration, and when the normal distribution’s D statistics are included, the scale of y-axis becomes very large, hich prevents to reveal the differences of AD statistics for other istributions. KS and AD plots for the remaining markets can be ound in the Appendix. We find that for every market in each sub-sample, and in the ntire sample, the KS statistic in the normal case is almost always bove the KS statistics in all other five distributions. A similar ar- ument also holds for the AD statistics. Furthermore, we observe hat, in general, all five distributions excluding the g-and-h distri- ution perform very similarly to each other; the g-and-h distribu- ion marginally underperforms in many sub-samples. 178 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 1 9 9 4 − 1 9 9 6 1 9 9 7 − 1 9 9 9 2 0 0 0 − 2 0 0 2 2 0 0 3 − 2 0 0 5 2 0 0 6 − 2 0 0 8 2 0 0 9 − 2 0 1 1 2 0 1 2 − 2 0 1 4 E n ti re S a m p le 0.0 0.5 1.0 1.5 2.0 2.5 Mexico Skewed−t GLD NIG Johnson Su g−and−h 1 9 8 8 − 1 9 9 0 1 9 9 1 − 1 9 9 3 1 9 9 4 − 1 9 9 6 1 9 9 7 − 1 9 9 9 2 0 0 0 − 2 0 0 2 2 0 0 3 − 2 0 0 5 2 0 0 6 − 2 0 0 8 2 0 0 9 − 2 0 1 1 2 0 1 2 − 2 0 1 4 E n ti re S a m p le 0 1 2 3 4 Turkey Skewed−t GLD NIG Johnson Su g−and−h Fig. 5. Anderson–Darling (AD) statistics of five probability distributions for Mexico and Turkey. The top figure shows the AD statistics of the skewed-t distribution, generalized lambda distribution (GLD), normal inverse Gaussian (NIG) distribution, Johnson SU family, and g-and-h distribution for Mexico over all sub-periods starting with 1994–1996. The last tick mark on the horizontal axis presents the AD statistic for the entire sample over the period 1994–2014. The bottom plot does the same for Turkey over all sub-periods starting with 1988–1990. d d m w fi t t t α t t i 8 We must express our gratitude to an anonymous referee for his/her observation. Next, we compare the distributions of interest by plotting the empirical histogram of the daily returns together with the esti- mated distribution functions. Fig. 6 presents the empirical his- togram of the log-returns of the Ibovespa Index (Brazil) over the sub-period 1997–1999 together with six estimated distribution functions for the same index. One notable observation is that the normal distribution significantly underperforms other distributions in modeling the Ibovespa Index. In particular, the normal distribu- tion cannot capture the peakedness and the tails of the data. Other distributions including the generalized lambda distribution (GLD), the skewed-t distribution, the normal inverse Gaussian (NIG) dis- tribution, and the g-and-h distribution perform very similarly in modeling the data. However, the fit of the Johnson translation sys- tem is slightly more peaked than the histogram. As the computed KS and AD statistic values are very close to each other in all distributions, it is difficult to identify whether the ifferences in these statistics are significant.8 Furthermore, Fig. 6 oes not provide an answer to the question of which distribution odels the daily equity return data best. To address this problem, e conduct a power test, where we calculate the p-value of each t. The null hypothesis is that the observed data originates from he hypothesized distribution and the alternative hypothesis is that he observed data does not belong to the hypothesized distribu- ion. The null hypothesis is rejected if the p-value is lower than = 5%. It is important to recognize here that for many distribu- ions under consideration, the critical values of both the KS statis- ic and AD statistic do not exist. The typical approach in this case s to compute the p-values and the critical values by means of a C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 179 Log−Returns F re q u e n c y −0.1 0.0 0.1 0.2 0.3 5 1 0 1 5 2 0 2 5 Brazil (1997−1999) GLD Normal Skewed−t Johnson Su NIG g−and−h Fig. 6. Histogram and distribution fits for a selected market. This figure plots the histogram of the Brazil log-returns together with the estimated distribution func- tions of the generalized lambda distribution (GLD), normal distribution, skewed-t distribution, Johnson SU family, normal inverse Gaussian (NIG) distribution, and g- and-h distribution over the sub-period 1997–1999. The bins of the histogram are calculated with the Scott method in R. M 2 A d T ( c s a r i T G t i i m e o r r t c t a d fi G ( m b N s s 6 o f t a 2 a 4 v p i i { r t s d t p i 2 w ( v 6 t m m t t l a l T u n A t t O m onte Carlo simulation for each hypothesized distribution (Ross, 004). We follow the following steps to calculate the p-value: Step 1. Use the data to estimate the parameters of the hypoth- esized distribution and compute the value of the test statis- tics denoted by G (such as KS and/or AD statistic). Step 2. Generate a sample of size n that is equal to the number of observed data from the fitted distribution. Step 3. Fit the hypothesized distribution to the data generated in Step 2 and estimate the sample goodness-of-fit statistics. Step 4: Repeat Step 1 through Step 3 1000 times and calculate the p-value as the proportion of times the sample statistics values exceed the observed value G of the original sample. Step 5: Reject the null hypothesis if p-value is smaller than 0.05. Table 3 tabulates the number of rejections using both KS and D test statistics in each sub-period for each distribution in the eveloped markets. Table 4 does the same for emerging markets.9 he last row of both tables tabulates the percentage of sub-periods out of 12 sub-periods) in which a particular distribution is ac- epted by the p-value test using both KS and AD statistics. The re- ults reported in both tables suggest that the GLD is a prominent lternative for fitting the daily equity index return data. We fail to eject the GLD according to both KS and AD test statistics for all ndex returns in all sub-samples, as well as in the entire sample. herefore, according to the p-value test, we can conclude that the LD is a powerful distribution in modeling both the center and the ails of the daily index return data. We conclude this section by comparing the models considered n this paper according to their stability.10 Stability is particularly mportant for applications of portfolio analysis and risk manage- ent (Rachev & Mittnik, 2000). In particular, stable distributed 9 To save space, we do not report the results for each specific market here. How- ver, these results are available upon request. 10 We must express our gratitude to an anonymous referee who brought this to ur attention. j m t S t eturns possess the property that linear combinations of return se- ies, such as portfolios, follow a stable distribution. We investigate he relative stability of the models of interest by comparing their apability to represent the variety of shapes taken by return dis- ributions over twelve sub-periods. We find that in both emerging nd developed markets, GLD can adequately describe the empirical istributions in all sub-periods. Other distributions achieve a good t in smaller share of periods. Specifically, in developed markets, LD is followed by Johnson SU family (92%), g-and-h distribution 67%), skewed-t distribution (42%), NIG distribution (41%), and nor- al distribution (0%) under the KS statistic and by g-and-h distri- ution (67%), Johnson SU family (42%), skewed-t distribution (25%), IG distribution (17%), and normal distribution (0%) under the AD tatistic (the last row of Table 3). In emerging markets, both John- on SU family and g-and-h distribution follow GLD with a share of 7%, and skewed-t distribution and NIG distribution with a share f 25% under the KS statistic. When the AD statistic is used to per- orm the p-value test, the GLD is accepted in all of the sub-periods, he Johnson SU family and g-and-h are accepted in 58% of periods, nd the skewed-t distribution and NIG distribution are accepted in 5% and 17% of periods, respectively. The normal distribution is not ccepted in any of the sub-periods (the last row of Table 4). .2. Risk estimation We examine the behavior of the distributions at the extreme alues of each market index return using the VaR measure. The urpose of this investigation is to observe the risk that an investor s facing when she has a long or short position on the market ndexes from our sample. The risk levels are determined as α ∈ 0.005, 0.01, 0.05, 0.95, 0.99, 0.995}, in which the first three levels epresent the lower extreme of returns (long position) and the last hree represent the upper extreme of returns (short position). In- ample VaR(α) values are calculated to observe the behavior of the istributions at the tails. We then apply the Kupiec likelihood ratio est given in Kupiec (1995), which tests whether the expected pro- ortion of violations is equal to α. The likelihood ratio test statistic s given by: log((τ (α)/n)τ (α)(1 − τ (α)/n)n−τ (α) ) − 2 log(ατ (α)(1 − α)n−τ (α) ), here τ (α) is the number of times the observed returns are above short positions) or below (long positions) the theoretical VaR alue and n is the sample size. The results of the Kupiec test are represented in Tables 5 and . For example, Table 5 lists the number of times each distribu- ion is rejected at each significance level for all of the developed arkets in each sub-period. Table 6 does the same for emerging arkets. Focusing on the sub-period 1979–1981 in Table 5, we see hat the GLD, Johnson SU family, skewed-t distribution, NIG dis- ribution, and g-and-h are not rejected in any of the significance evels. However, the normal distribution is rejected by 3 markets t the significance level of 0.005, by 2 markets at the significance evel of 0.01, and by 1 market at the significance level of 0.95. he total number of rejections is tabulated at the far right col- mn of the table. For the normal distribution, for instance, the total umber of rejections is equal to 6 over the sub-period 1979–1981. dding up the number of rejections over the sub-periods, we see hat the GLD is the least rejected model with only 7 rejections in he developed markets and 8 rejections in the emerging markets. n the other hand, the normal distribution is the most rejected odel with 207 rejections in the developed markets and 213 re- ections in the emerging markets. In both developed and emerging arkets, the g-and-h distribution, NIG distribution, and skewed- distribution perform similarly to each other, while the Johnson U family underperforms other distributions with 65 rejections in he developed markets and 57 rejections in the emerging markets. 180 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 Table 3 Number of times that each distribution is rejected according to Kolmogorov–Smirnov (KS) and Anderson– Darling (AD) test statistics in developed markets. The first column of this table (excluding the last row) lists the periods under consideration; the second column (excluding the last row) presents the number of devel- oped markets considered in each period; the remaining columns (excluding the last row) presents the number of times a distribution is rejected by the p-value test according to the KS and AD test statistics. For instance, in the sub-period 1979–1981, the generalized lambda distribution (GLD), Johnson SU family, and g-and-h distribu- tion are rejected neither according to the KS statistic nor according to the AD statistic; the skewed-t distribution is rejected once according to both KS and AD test statistics; the normal inverse Gaussian (NIG) distribution is not rejected according to the KS statistic but rejected once according to the AD statistic; and the normal distri- bution is rejected 3 times according to both KS and AD statistics. The very last row of this table presents the percentage of periods in which a particular distribution is accepted by the p-value test. For instance, the GLD is never rejected in any of the sub-periods; therefore, its percentage of acceptance is 100% according to both KS and AD statistics. The Johnson SU family is rejected in the sub-period 2003–2005 according to the KS statistic (rejected only once in 12 sub-periods; thus, the percentage of acceptance is 11/12 × 100 ≈ 92%) and is not rejected in the sub-periods 1979–1981, 1988–1990, 1994–1996, 2009–2011, and 2012–2014 according to the AD statistic (accepted in 5 sub-periods out of 12 sub-periods; thus, the percentage of acceptance is 5/12 × 100 ≈ 42%). Period # Markets GLD Johnson SU Skewed-t NIG g-and-h Normal KS AD KS AD KS AD KS AD KS AD KS AD 1979–1981 4 0 0 0 0 1 1 0 1 0 0 3 3 1982–1984 5 0 0 0 1 0 0 0 0 0 0 4 4 1985–1987 7 0 0 0 1 1 2 3 4 0 0 7 7 1988–1990 8 0 0 0 0 0 2 1 4 0 0 8 8 1991–1993 9 0 0 0 1 2 3 1 3 1 0 8 8 1994–1996 9 0 0 0 0 1 1 0 1 0 0 8 8 1997–1999 10 0 0 0 1 0 0 2 0 0 0 8 9 2000–2002 10 0 0 0 1 0 0 0 1 0 0 9 10 2003–2005 10 0 0 2 7 2 3 2 2 1 1 10 10 2006–2008 10 0 0 0 3 0 2 3 4 2 1 10 10 2009–2011 10 0 0 0 0 1 2 0 1 2 2 9 10 2012–2014 10 0 0 0 0 1 3 1 1 0 0 10 10 Entire Sample 10 0 0 1 2 2 7 2 7 1 1 10 10 % of acceptance 100 100 92 42 42 25 41 17 67 67 0 0 Table 4 Number of times that each distribution is rejected according to Kolmogorov–Smirnov (KS) and Anderson– Darling (AD) test statistics in emerging markets. The first column of this table (excluding the last row) lists the periods under consideration; the second column (excluding the last row) presents the number of emerging markets considered in each period; the remaining columns (excluding the last row) presents the number of times a distribution is rejected by the p-value test according to the KS and AD statistics. For instance, in the sub-period 1979–1981, the generalized lambda distribution (GLD), Johnson SU family, normal inverse Gaussian (NIG) distribution, and g-and-h distribution are rejected neither according to the KS statistic nor according to the AD statistic; the skewed-t distribution is rejected once according to the AD statistic and normal distribu- tion is rejected 2 times according to both KS and AD statistics. The very last row of this table presents the percentage of periods in which a particular distribution is accepted by the p-value test. For instance, the GLD is never rejected in any of the sub-periods; therefore, its percentage of acceptance is 100% according to both KS and AD statistics. The Johnson SU family is rejected in the sub-periods 1982–1984, 1985–1987, 1988–1990, and 2009–2011 according to the KS statistic (rejected 4 times in 12 sub-periods; thus, the percentage of acceptance is 8/12 × 100 ≈ 67%) and is rejected in the sub-periods 1985-1987, 1991-1993, 1997-1999, 2003-2005, and 2009–2011 according to the AD statistic (rejected in 5 sub-periods out of 12 sub-periods; thus, the percentage of acceptance is 7/12 × 100 ≈ 58%). Period # Markets GLD Johnson SU Skewed-t NIG g-and-h Normal KS AD KS AD KS AD KS AD KS AD KS AD 1979–1981 2 0 0 0 0 0 1 0 0 0 0 2 2 1982–1984 2 0 0 1 0 0 0 1 1 0 0 2 2 1985–1987 2 0 0 1 2 2 2 2 2 0 0 2 2 1988–1990 4 0 0 1 0 2 2 2 1 0 1 4 4 1991–1993 5 0 0 0 1 0 0 0 0 1 1 4 4 1994–1996 7 0 0 0 0 2 1 1 1 0 0 6 6 1997–1999 9 0 0 0 2 3 2 2 2 1 1 8 9 2000–2002 9 0 0 0 0 1 0 0 2 0 0 8 8 2003–2005 10 0 0 0 2 2 2 1 2 1 1 7 8 2006–2008 10 0 0 0 0 2 2 1 2 2 2 10 10 2009–2011 10 0 0 1 1 1 1 1 1 0 0 10 10 2012–2014 10 0 0 0 0 1 1 2 2 0 0 9 9 Entire Sample 10 0 0 1 2 4 5 4 5 4 4 10 10 % of acceptance 100 100 67 58 25 25 25 17 67 58 0 0 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 181 Table 5 Value-at-Risk (VaR) failure rate results for the developed markets. This table presents the number of times each distribution including the generalized lambda distribution (GLD), Johnson SU family, skewed-t distribu- tion, normal inverse Gaussian (NIG) distribution, g-and-h distribution, and normal distribution is rejected at various significance levels for all developed markets in each sub-period according to the Kupiec likelihood ratio test. Period # Market Method Significance levels Total # Rejections 0.005 0.01 0.05 0.95 0.99 0.995 1979–1981 4 GLD 0 0 0 0 0 0 0 Johnson SU 0 0 0 0 0 0 0 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 3 2 0 1 0 0 6 1982–1984 5 GLD 0 0 0 0 0 0 0 Johnson SU 1 1 0 0 1 2 5 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 1 0 0 0 2 3 6 1985–1987 7 GLD 0 0 0 0 0 0 0 Johnson SU 1 0 0 0 0 0 1 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 1 0 0 0 1 g-and-h 1 1 0 0 0 0 2 Normal 1 1 3 4 0 0 9 1988–1990 8 GLD 0 0 0 1 0 0 1 Johnson SU 1 2 0 0 1 1 5 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 4 4 2 5 0 1 16 1991–1993 9 GLD 0 0 0 0 0 0 0 Johnson SU 2 2 0 0 1 1 6 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 1 0 0 1 g-and-h 0 0 0 0 0 0 0 Normal 1 0 2 1 0 4 8 1994–1996 9 GLD 0 0 0 0 0 0 0 Johnson SU 1 0 0 0 1 0 2 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 5 3 0 0 0 1 9 1997–1999 10 GLD 0 0 0 0 0 0 0 Johnson SU 2 0 0 0 1 1 4 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 5 4 0 3 0 0 12 2000–2002 10 GLD 0 0 0 0 0 0 0 Johnson SU 1 0 0 0 0 1 2 Skewed-t 0 0 0 0 0 1 1 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 2 2 0 0 2 4 10 2003–2005 10 GLD 0 0 0 0 0 0 0 Johnson SU 1 1 0 0 1 3 6 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 9 7 0 2 0 2 20 2006–2008 10 GLD 1 0 0 0 1 0 2 Johnson SU 0 0 0 0 2 3 5 Skewed-t 0 0 0 0 2 0 2 NIG 0 0 0 0 1 0 1 g-and-h 0 0 0 0 1 0 1 Normal 10 10 1 8 2 6 37 2009–2011 10 GLD 0 0 0 0 0 0 0 Johnson SU 2 1 0 0 0 2 5 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 1 0 0 0 0 0 1 Normal 8 7 0 0 0 2 17 2012–2014 10 GLD 1 0 0 0 0 0 1 Johnson SU 4 3 0 0 3 2 12 Skewed-t 1 0 0 0 0 0 1 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 7 7 0 0 1 1 16 Entire Sample 10 GLD 0 2 1 0 0 0 3 Johnson SU 3 2 2 2 2 2 13 Skewed-t 1 2 1 0 1 0 5 NIG 1 2 2 0 1 0 6 g-and-h 1 0 0 0 0 0 1 Normal 10 10 9 5 7 10 51 182 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 Table 6 Value-at-Risk (VaR) failure rate results for the emerging markets. This table presents the number of times each distribution including the generalized lambda distribution (GLD), Johnson SU family, skewed-t distribution, nor- mal inverse Gaussian (NIG) distribution, g-and-h distribution, and normal distribution is rejected at various significance levels for all emerging markets in each sub-period according to the Kupiec likelihood ratio test. Period # Market Method Significance levels Total # Rejections 0.005 0.01 0.05 0.95 0.99 0.995 1979–1981 2 GLD 0 0 0 0 0 0 0 Johnson SU 0 0 0 0 0 1 1 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 1 1 0 1 0 1 4 1982–1984 2 GLD 0 0 0 0 0 0 0 Johnson SU 0 0 0 0 0 0 0 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 1 0 1 0 2 2 6 1985–1987 2 GLD 0 0 0 0 0 0 0 Johnson SU 1 0 0 0 0 0 1 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 2 1 1 1 0 0 5 1988–1990 4 GLD 0 0 0 0 0 1 1 Johnson SU 2 1 0 0 1 2 6 Skewed-t 0 0 0 0 1 2 3 NIG 0 0 0 0 1 2 3 g-and-h 0 0 0 0 0 0 0 Normal 3 2 1 1 2 2 11 1991–1993 5 GLD 0 0 0 0 0 2 2 Johnson SU 0 0 0 0 1 2 3 Skewed-t 0 0 0 0 0 2 2 NIG 0 0 0 0 0 2 2 g-and-h 0 0 0 0 0 1 1 Normal 1 1 1 0 3 4 10 1994–1996 7 GLD 0 0 0 0 0 0 0 Johnson SU 2 2 0 0 3 2 9 Skewed-t 0 0 0 0 0 1 1 NIG 0 0 0 0 0 1 1 g-and-h 0 0 0 0 0 0 0 Normal 3 1 0 0 4 4 12 1997–1999 9 GLD 0 0 0 0 0 1 1 Johnson SU 1 1 0 0 3 3 8 Skewed-t 0 0 0 0 0 1 1 NIG 0 0 0 0 0 1 1 g-and-h 0 0 0 0 2 0 2 Normal 5 3 2 3 3 5 21 2000–2002 9 GLD 0 0 0 0 0 0 0 Johnson SU 2 2 0 0 0 0 4 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 1 0 0 1 Normal 4 3 2 2 1 5 17 2003–2005 10 GLD 0 0 0 0 0 0 0 Johnson SU 0 1 0 0 1 2 4 Skewed-t 0 0 0 0 0 0 0 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 0 0 0 Normal 4 4 1 4 1 2 16 2006–2008 10 GLD 1 1 0 0 0 0 2 Johnson SU 2 2 0 0 1 1 6 Skewed-t 1 1 0 0 1 1 4 NIG 0 0 0 0 0 0 0 g-and-h 0 0 0 0 1 1 2 Normal 10 10 1 6 2 5 34 2009–2011 10 GLD 1 0 0 0 0 0 1 Johnson SU 2 2 0 0 0 0 4 Skewed-t 1 0 0 0 0 0 1 NIG 1 0 0 0 0 0 1 g-and-h 1 0 0 0 0 0 1 Normal 6 7 0 0 3 3 19 2012–2014 10 GLD 0 1 0 0 0 0 1 Johnson SU 1 2 0 0 2 1 6 Skewed-t 0 1 0 0 0 0 1 NIG 0 1 0 0 0 0 1 g-and-h 0 0 0 0 0 0 0 Normal 3 2 0 0 1 3 9 (continued on next page) C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 183 Table 6 (continued) Period # Market Method Significance levels Total # Rejections 0.005 0.01 0.05 0.95 0.99 0.995 Entire Sample 10 GLD 0 0 0 0 0 0 0 Johnson SU 1 1 0 0 1 2 5 Skewed-t 0 0 0 1 0 2 3 NIG 0 0 0 1 0 0 1 g-and-h 0 1 1 1 3 1 7 Normal 10 10 7 4 8 10 49 T t 5 i h V p n r d N m w u l S a e o a o m p b d t a s r f t r T e c I a t e i l e t a ( e d a A fl e f i t i G u t m w t y m b p I a A k s d A t ( e p A p l t p K s t t S A s t t t t A a i t M i hus, we conclude that the GLD also outperforms other distribu- ions in terms of VaR performance. . Conclusion An analysis of the empirical distribution of equity returns is mportant to both academic researchers and financial experts and as many implications in the calculation of risk measures such as alue-at-Risk (VaR) and the pricing of equity options. Despite its opularity in finance applications, it is now well known that the ormal distribution fails to capture certain characteristics of equity eturn data. This has motivated researchers to investigate flexible istributions with the ability to better represent stock return data. evertheless, on a practical level, which distribution best suited for odeling equity stock return data has remained an open question. This paper contributes to the literature in the following two ays: First, we investigate the relative performance of five widely sed flexible distributions in finance, specifically, the generalized ambda distribution (GLD), Johnson system of distributions, skewed tudent-t distribution, normal inverse Gaussian (NIG) distribution, nd g-and-h distribution, for modeling the distribution of daily quity index returns of ten developed and ten emerging markets ver the years 1979–2014. We also use the normal distribution as benchmark in our analysis. Second, we evaluate the implication f our results for the implementation of the well-known risk easure, VaR. Our analyses support the empirical evidence in the revious research that the behavior of equity returns are far from eing normally distributed. We further show that the marginal istribution of daily equity index returns can be well described by he GLD. The relative stability of the GLD also makes it more favor- ble among other distributions. From a practical expert intelligent ystem perspective, the GLD has the following advantages: (i) The epresentation of the GLD as an inverse cumulative distribution unction makes it easy to quickly generate random variates from he GLD in a Monte Carlo simulation, which is widely used in isk management and the pricing of derivative securities. (ii) he percentile representation of the GLD makes it convenient to stimate the risk measures, such as VaR and expected shortfall. In fitting the distributions to the observed data, we must hoose among several fitting methods from the previous research. n making this choice, we consider the availability of off-the-shelf lgorithms that can be easily used by expert and intelligent sys- ems. If such an algorithm is not available, then we implement the asiest method. One potential limitation of our work is that the mplemented fitting methods may not be the “best” methods. This imitation could be overcome by performing an intensive study to xamine the relative performance of fitting methods for each dis- ribution considered in this paper. One such study that we are ware of concerns the estimation of the parameters of the GLD Corlu & Meterelliyoz, 2015). Our focus in this paper is on the unconditional distribution of quity returns. As stock returns typically exhibit temporal depen- ence, considering conditional homoskedastic models such ARMA nd conditional heteroskedastic models such as ARCH, GARCH, and RMA-GARCH models, where the residuals in these models follow exible distributions considered in this paper, could be of inter- st from an expert and intelligent systems perspective. Another uture research avenue is related to the stability of the probabil- ty distributions. As mentioned in the paper, stability is impor- ant, especially for portfolio analysis and risk management. An nteresting problem to consider is how the performance of the LD would compare with some of the popular stable distributions sed in finance, such as stable Paretian laws. In addition, studying he performance of the considered distributions using weekly and onthly equity return data may produce different insights, just as eekly and monthly data may exhibit different characteristics than he daily data. Finally, another research idea is to extend our anal- sis on the behavior of distributions at the extreme values of each arket index return using risk measures other than VaR. Despite eing a widely used risk measure, VaR has been criticized for not roperly presenting the full picture of the risks a company faces. n particular, a well-known shortcoming of VaR is that it is not coherent risk measure (Artzner, Delbaen, Eber, & Heath, 1999). n alternative coherent risk measure is the expected shortfall, also nown as conditional VaR or tail loss. The extension of our analy- is using the expected shortfall as the risk measure may yield ad- itional insights. ppendix A.1. Plots obtained using the Kolmogorov–Smirnov est statistic This Appendix presents the computed Kolmogorov–Smirnov KS) test statistic for developed and emerging markets consid- red in the paper under several distributions. Specifically, Fig. 1 resents the computed KS statistics for the developed markets of ustralia, France, Japan, and Singapore; Fig. 2 presents the com- uted KS statistics for the developed markets of Spain, Switzer- and, UK, and US; Fig. 3 presents the computed KS statistics for he emerging markets of Brazil, Chile, China, and India; and Fig. 4 resents the computed KS statistics for the emerging markets of orea, Malaysia, Russia, and South Africa. In each figure, the KS tatistic is compared with respect to the following six distribu- ions: the skewed-t distribution, the generalized lambda distribu- ion (GLD), the normal inverse Gaussian (NIG) distribution, Johnson U family, g-and-h distribution, and normal distribution. ppendix A.2. Plots obtained using the Anderson–Darling test tatistic This Appendix presents the computed Anderson–Darling (AD) est statistic for developed and emerging markets considered in he paper under several distributions. Specifically, Fig. 5 presents he computed AD statistics for the developed markets of Aus- ralia, France, Japan, and Singapore; Fig. 6 presents the computed D statistics for the developed markets of Spain, Switzerland, UK, nd US; Fig. 7 presents the computed AD statistics for the emerg- ng markets of Brazil, Chile, China, and India; and Fig. 8 presents he computed AD statistics for the emerging markets of Korea, alaysia, Russia, and South Africa. In each figure, the AD statistic s compared with respect to the following five distributions: the 184 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 Fig. 1. KS statistics for the developed markets of Australia, France, Japan, and Singapore. C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 185 Fig. 2. KS statistics for the developed markets of Spain, Switzerland, UK, and US. 186 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 Fig. 3. KS statistics for the emerging markets of Brazil, Chile, China, and India. C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 187 Fig. 4. KS statistics for the emerging markets of Korea, Malaysia, Russia, and South Africa. 188 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 Fig. 5. AD statistics for the developed markets of Australia, France, Japan, and Singapore. C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 189 Fig. 6. AD statistics for the developed markets of Spain, Switzerland, UK, and US. 190 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 Fig. 7. AD statistics for the emerging markets of Brazil, Chile, China, and India. C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 191 Fig. 8. AD statistics for the emerging markets of Korea, Malaysia, Russia, and South Africa. 192 C.G. Corlu et al. / Expert Systems With Applications 54 (2016) 170–192 H H H H H H J J K K L L M M O O P P P P R R R R S S T T skewed-t distribution, the generalized lambda distribution (GLD), the normal inverse Gaussian (NIG) distribution, Johnson SU family, and g-and-h distribution. References Akgiray, V., & Booth, G. G. (1987). Compound distribution models of stock returns – an empirical comparison. Journal of Financial Research, 10(3), 269–280. Anderson, T. W., & Darling, D. A. (1954). A test of goodness of fit. Journal of the American Statistical Association, 49(268), 765–769. Aparicio, F. M., & Estrada, J. (2001). Empirical distributions of stock returns: Euro- pean securities markets, 1990–95. The European Journal of Finance, 7(1), 1–21. Artzner, P., Delbaen, F., Eber, J.-M., & Heath, D. (1999). Coherent measures of risk. Mathematical Finance, 9(3), 203–228. Azzalini, A., & Capitanio, A. (2003). Distributions generated by perturbation of sym- metry with emphasis on a multivariate skew t-distribution. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 65(2), 367–389. Bachelier, L. (1900). Théorie de la spéculation. In Annales scientifiques de l’école nor- male supérieure: Vol. 17 (pp. 21–86). Société mathématique de France. Badrinath, S., & Chatterjee, S. (1988). On measuring skewness and elongation in common stock return distributions: The case of the market index. The Journal of Business, 61(4), 451–472. Behr, A., & Pötter, U. (2009). Alternatives to the normal model of stock returns: Gaussian mixture, generalised logF and generalised hyperbolic models. Annals of Finance, 5(1), 49–68. Blattberg, R. C., & Gonedes, N. J. (1974). A comparison of the stable and Student distributions as statistical models for stock prices. The Journal of Business, 47(2), 244–280. Bølviken, E., & Benth, F. E. (2000). Quantification of risk in Norwegian stocks via the normal inverse Gaussian distribution. In Afir 2000 colloquim, Tromsø, Norway (pp. 87–98). Bookstaber, R. M., & McDonald, J. B. (1987). A general distribution for describing security price returns. The Journal of Business, 60(3), 401–424. Chakravarti, I., & Laha, R. (1967). Handbook of methods of applied statistics: Vol. 1. John Wiley & Sons. Chalabi, Y., Scott, D. J., & Würtz, D. (2010). The generalized Lambda distribution as an alternative to model financial returns. Technical report. Eidgenössische Tech- nische Hochschule and University of Auckland, Zurich and Auckland. Corlu, C. G., & Biller, B. (2015). Least-squares and Bayesian inferences for Johnson’s SL and SB distributions, Metropolitan College, Boston University, Boston, MA.Working paper. Corlu, C. G., & Meterelliyoz, M. (2015). Estimating the parameters of the generalized lambda distribution: Which method performs best? Communications in Statistics – Simulation and Computation.. doi:10.1080/03610918.2014.901355. Dutta, K. K., & Babbel, D. F. (2002). On measuring skewness and kurtosis in short rate distributions: The case of the US Dollar London inter bank offer rates. Technical report. Wharton School Center for Financial Institutions, University of Pennsylvania. Eberlein, E., & Keller, U. (1995). Hyperbolic distributions in finance. Bernoulli, 1(3), 281–299. Fama, E. F. (1963). Mandelbrot and the stable Paretian hypothesis. The Journal of Business, 36(4), 420–429. Fama, E. F. (1965). The behavior of stock-market prices. Journal of Business, 38(1), 34–105. Fernandez, P., Aguirreamalloa, J., & Avendaño, L. C. (2011). Ibex 35: 1991–2010 - value creation and return. Http://ssrn.com/abstract=1737860. Accessed 5.11.15. Filliben, J. J. (1975). The probability plot correlation coefficient test for normality. Technometrics, 17(1), 111–117. Fournier, B., Rupin, N., Bigerelle, M., Najjar, D., Iost, A., & Wilcox, R. (2007). Estimat- ing the parameters of a generalized lambda distribution. Computational Statistics & Data Analysis, 51(6), 2813–2835. Freimer, M., Kollia, G., Mudholkar, G. S., & Lin, C. T. (1988). A study of the gen- eralized Tukey lambda family. Communications in Statistics-Theory and Methods, 17(10), 3547–3567. Gray, J. B., & French, D. W. (1990). Empirical comparisons of distributional models for stock index returns. Journal of Business Finance & Accounting, 17(3), 451–459. agerman, R. L. (1978). More evidence on the distribution of security returns. The Journal of Finance, 33(4), 1213–1221. arris, R. D., & Küçüközmen, C. C. (2001a). The empirical distribution of stock re- turns: Evidence from an emerging European market. Applied Economics Letters, 8(6), 367–371. arris, R. D., & Küçüközmen, C. C. (2001b). The empirical distribution of UK and US stock returns. Journal of Business Finance & Accounting, 28(5–6), 715–740. astings, C., Mosteller, F., Tukey, J. W., & Winsor, C. P. (1947). Low moments for small samples: A comparative study of order statistics. The Annals of Mathemat- ical Statistics, 18(3), 413–426. oaglin, D. C. (2006). Summarizing shape numerically: The g-and-h distributions. In Exploring Data Tables, Trends, and Shapes (pp. 461–513). John Wiley & Sons, Inc. su, D.-A., Miller, R. B., & Wichern, D. W. (1974). On the stable Paretian behavior of stock-market prices. Journal of the American Statistical Association, 69(345), 108– 113. ohnson, N. L. (1949). Systems of frequency curves generated by methods of trans- lation. Biometrika, 36(1/2), 149–176. oiner, B. L., & Rosenblatt, J. R. (1971). Some properties of the range in samples from Tukey’s symmetric lambda distributions. Journal of the American Statistical Asso- ciation, 66(334), 394–399. on, S. J. (1984). Models of stock returns—a comparison. The Journal of Finance, 39(1), 147–165. upiec, P. H. (1995). Techniques for verifying the accuracy of risk measurement models. The Journal of Derivatives, 3(2), 73–84. au, A. H.-L., Lau, H.-S., & Wingender, J. R. (1990). The distribution of stock returns: New evidence against the stable model. Journal of Business & Economic Statistics, 8(2), 217–223. inden, M. (2001). A model for stock return distribution. International Journal of Fi- nance & Economics, 6(2), 159–169. andelbrot, B. (1967). The variation of some other speculative prices. Journal of Business, 40(4), 393–413. ills, T. C. (1995). Modelling skewness and kurtosis in the London Stock Exchange FT-SE index return distributions. Journal of the Royal Statistical Society. Series D (The Statistician), 44(3), 323–332. fficer, R. R. (1972). The distribution of stock returns. Journal of the American Statis- tical Association, 67(340), 807–812. sborne, M. F. (1959). Brownian motion in the stock market. Operations Research, 7(2), 145–173. faff, B., McNeil, A., & Ulmann, S. (2013). Qrm: Provides R-language code to exam- ine quantitative risk management concepts. R package version 0.4-9.URL http: //CRAN.R- project.org/package=QRM. Accessed 5.11.15. raetz, P. D. (1972). The distribution of share price changes. Journal of Business, 45(1), 49–55. rause, K. (1999). How to use NIG laws to measure market risk. FDM Preprint 65. University of Freiburg. rause, K., Zentrum, F., & Modellbildung, D. (1997). Modelling financial data using generalized hyperbolic distributions. FDM Preprint 48 Technical Report. University of Freiburg. achev, S. T., & Mittnik, S. (2000). Stable Paretian models in finance. Chichester, UK: John Wiley & Sons. achev, S. T., Stoyanov, S. V., Biglova, A., & Fabozzi, F. J. (2005). An empirical exami- nation of daily stock return distributions for US stocks. In D. Baier, R. Decker, & L. Schmidt-Thieme (Eds.), Data analysis and decision support. In Series in Stud- ies in Classification, Data Analysis, and Knowledge Organization (pp. 269–281). Springer Berlin Heidelberg. amberg, J. S., & Schmeiser, B. W. (1974). An approximate method for generating asymmetric random variables. Communications of the ACM, 17(2), 78–82. oss, S. M. (2004). Simulation (4th ed.). Orlando, FL, USA: Academic Press, Inc. hilling, H. (1996). International guide to securities market indices. Florida, USA: CRC Press. u, S. (2007). Numerical maximum log likelihood estimation for generalized lambda distributions. Computational Statistics & Data Analysis, 51(8), 3983–3998. uenter, H. J. (2001). An algorithm to determine the parameters of SU -curves in the Johnson system of probability distributions by moment matching. Journal of Sta- tistical Computation and Simulation, 70(4), 325–347. ukey, J. (1977). Exploratory data analysis. Addison-Wesley series in behavioral science. Addison-Wesley Publishing Company. http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0001 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0001 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0001 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0001 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0002 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0002 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0002 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0002 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0003 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0003 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0003 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0003 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0004 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0004 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0004 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0004 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0004 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0004 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0005 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0005 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0005 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0005 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0006 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0006 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0007 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0007 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0007 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0007 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0008 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0008 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0008 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0008 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0009 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0009 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0009 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0009 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0010 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0010 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0010 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0010 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0011 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0011 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0011 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0011 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0012 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0012 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0012 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0012 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0013 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0013 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0013 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0013 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0013 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0014 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0014 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0014 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0014 http://dx.doi.org/10.1080/03610918.2014.901355 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0016 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0016 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0016 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0016 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0017 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0017 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0017 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0017 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0018 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0018 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0019 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0019 http://ssrn.com/abstract=1737860 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0020 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0020 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0021 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0021 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0021 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0021 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0021 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0021 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0021 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0021 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0022 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0022 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0022 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0022 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0022 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0022 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0023 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0023 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0023 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0023 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0024 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0024 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0025 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0025 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0025 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0025 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0026 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0026 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0026 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0026 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0027 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0027 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0027 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0027 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0027 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0027 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0028 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0028 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0029 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0029 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0029 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0029 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0029 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0030 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0030 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0031 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0031 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0031 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0031 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0032 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0032 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0033 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0033 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0034 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0034 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0034 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0034 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0034 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0035 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0035 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0036 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0036 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0037 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0037 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0038 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0038 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0039 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0039 http://CRAN.R-project.org/package=QRM http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0041 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0041 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0042 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0042 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0043 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0043 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0043 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0043 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0043 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0044 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0044 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0044 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0044 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0045 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0045 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0045 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0045 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0045 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0045 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0046 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0046 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0046 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0046 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0047 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0047 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0048 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0048 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0049 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0049 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0050 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0050 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0051 http://refhub.elsevier.com/S0957-4174(16)00052-X/sbref0051 Empirical distributions of daily equity index returns: A comparison 1 Introduction 2 Description of the data 3 Flexible distributions 3.1 The Generalized Lambda distribution 3.2 The Johnson translation system 3.3 The skewed Student-t distribution 3.4 The Normal Inverse Gaussian distribution 3.5 The g-and-h distribution 4 Performance and risk estimation 4.1 Goodness-of-fit 4.2 Risk estimation 5 Conclusion Appendix A.1 Plots obtained using the Kolmogorov-Smirnov test statistic Appendix A.2 Plots obtained using the Anderson-Darling test statistic References 
work_xrosz555njd4jf3375ostnmdmm	----	RÉSUMATCHER: A PERSONALIZED RÉSUMÉ-JOB MATCHING SYSTEM A Thesis by SHIQIANG GUO Submitted to the Office of Graduate and Professional Studies of Texas A&M University in partial fulfillment of the requirements for the degree of MASTER OF SCIENCE Chair of Committee, Tracy Hammond Committee Members, Anxiao Jiang Daniel W. Goldberg Head of Department, Dilma Da Silva May 2015 Major Subject: Computer Science Copyright 2015 Shiqiang Guo ABSTRACT Today, online recruiting web sites such as Monster and Indeed.com have become one of the main channels for people to find jobs. These web platforms have provided their services for more than ten years, and have saved a lot of time and money for both job seekers and organizations who want to hire people. However, traditional information retrieval techniques may not be appropriate for users. The reason is because the number of results returned to a job seeker may be huge, so job seekers are required to spend a significant amount of time reading and reviewing their options. One popular approach to resolve this difficulty for users are recommender systems, which is a technology that has been studied for a long time. In this thesis we have made an effort to propose a personalized job-résumé match- ing system, which could help job seekers to find appropriate jobs more easily. We create a finite state transducer based information extraction library to extract mod- els from résumés and job descriptions. We devised a new statistical-based ontology similarity measure to compare the résumé models and the job models. Since the most appropriate jobs will be returned first, the users of the system may get a better result than current job finding web sites. To evaluate the system, we computed Nor- malized Discounted Cumulative Gain (NDCG) and precision@k of our system, and compared to three other existing models as well as the live result from Indeed.com. ii ACKNOWLEDGEMENTS My greatest thanks to the members of the Sketch Recognition Lab for their continued support and help in the research work covered in this thesis. This thesis would not have been possible without their support. In addition, I would like to give extra thanks to my advisor Dr. Tracy Hammond, as well as to my committee members Dr. Anxiao Jiang and Dr. Daniel W. Goldberg for their valuable sage advice. iii TABLE OF CONTENTS Page ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii TABLE OF CONTENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii 1. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.3 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2. BACKGROUND . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2.1 Recommender System . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2.2 Job Recommender Systems . . . . . . . . . . . . . . . . . . . . . . . . 7 2.3 Information Extraction in Job Recommender System . . . . . . . . . 9 2.4 Matching Algorithms in Job Recommender Systems . . . . . . . . . . 11 3. PROBLEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.1 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 4. SYSTEM OVERVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 4.1 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 4.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 4.3 Text Processing Stages . . . . . . . . . . . . . . . . . . . . . . . . . . 17 4.4 System Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 20 4.5 System Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 5. INFORMATION EXTRACTION . . . . . . . . . . . . . . . . . . . . . . . 23 iv 5.1 Semantic Labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 5.2 Patterns for Matching . . . . . . . . . . . . . . . . . . . . . . . . . . 26 5.3 Pattern Matching Library . . . . . . . . . . . . . . . . . . . . . . . . 27 5.3.1 Finite-State Transducer . . . . . . . . . . . . . . . . . . . . . 27 5.3.2 Matchers in the Pattern Matching Library . . . . . . . . . . . 29 5.3.3 Implementation of the Pattern Matching Library . . . . . . . 32 6. MODEL SIMILARITY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 6.1 Similarity of Major and Academic Degree . . . . . . . . . . . . . . . . 37 6.2 Similarity of Job Title . . . . . . . . . . . . . . . . . . . . . . . . . . 38 6.3 Similarity of Skills . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 7. ONTOLOGY CONSTRUCTION AND SIMILARITY . . . . . . . . . . . . 40 7.1 Semantic Similarity in JRSs . . . . . . . . . . . . . . . . . . . . . . . 40 7.2 Ontology Construction . . . . . . . . . . . . . . . . . . . . . . . . . . 41 7.3 Ontology-Based Semantic Similarity . . . . . . . . . . . . . . . . . . . 44 7.3.1 Path-Based Approaches . . . . . . . . . . . . . . . . . . . . . 46 7.3.2 Feature-Based Measures . . . . . . . . . . . . . . . . . . . . . 47 7.3.3 Content-Based Measures . . . . . . . . . . . . . . . . . . . . . 47 7.4 Statistical-Based Ontology Similarity Measure . . . . . . . . . . . . . 48 8. EVALUATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 8.1 Experiments of Information Extraction . . . . . . . . . . . . . . . . . 52 8.2 Experiments of Ontology Similarity . . . . . . . . . . . . . . . . . . . 52 8.3 Evaluation of the System . . . . . . . . . . . . . . . . . . . . . . . . . 55 8.4 Comparing with the Keyword Searching . . . . . . . . . . . . . . . . 57 9. CONCLUSION AND FUTURE WORK . . . . . . . . . . . . . . . . . . . 60 9.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 9.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 v LIST OF FIGURES FIGURE Page 1.1 Search Result of Indeed.com . . . . . . . . . . . . . . . . . . . . . . . 2 1.2 Matching the Jobs with Résumé . . . . . . . . . . . . . . . . . . . . . 4 2.1 Latent Aspect Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 4.1 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 4.2 Job Description Process Pipeline . . . . . . . . . . . . . . . . . . . . 19 4.3 Job Description List . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 4.4 Upload Resume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 4.5 Résumé Job Matching . . . . . . . . . . . . . . . . . . . . . . . . . . 22 5.1 Zero or One NFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 5.2 Single Token NFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 5.3 Concatenation NFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 5.4 Alternation NFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 5.5 Zero or One NFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 5.6 Zero or More NFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 5.7 One or More NFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 5.8 Finite Automata Transducers . . . . . . . . . . . . . . . . . . . . . . 34 7.1 Procedure of Finding Technical Terms . . . . . . . . . . . . . . . . . 43 7.2 Interface of Protege . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 7.3 Part of Ontology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 vi LIST OF TABLES TABLE Page 1.1 Résumé Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 3.1 Portions of Resume and Job Description . . . . . . . . . . . . . . . . 15 5.1 Example of Job Description . . . . . . . . . . . . . . . . . . . . . . . 24 5.2 All Words Mean Bachelors . . . . . . . . . . . . . . . . . . . . . . . . 25 5.3 Semantic Labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 5.4 More Labels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 5.5 Labeled Sentence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 5.6 Patterns to Match Degree Sentences . . . . . . . . . . . . . . . . . . . 27 5.7 Matchers of Our Library . . . . . . . . . . . . . . . . . . . . . . . . . 29 5.8 Examples of Matcher . . . . . . . . . . . . . . . . . . . . . . . . . . 30 7.1 Example Sentences in Job Descriptions . . . . . . . . . . . . . . . . . 42 7.2 Patterns to Extract Terms . . . . . . . . . . . . . . . . . . . . . . . . 43 7.3 An Example Sentence Matches the Pattern . . . . . . . . . . . . . . . 43 7.4 Some Sentences of Job Descriptions . . . . . . . . . . . . . . . . . . . 49 7.5 Similarities of Skills List 1 . . . . . . . . . . . . . . . . . . . . . . . . 51 7.6 Similarities of Skills List 2 . . . . . . . . . . . . . . . . . . . . . . . . 51 8.1 Information Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . 53 8.2 Similarities of Skill List 1 . . . . . . . . . . . . . . . . . . . . . . . . . 53 8.3 Javascript Similarity Evaluation . . . . . . . . . . . . . . . . . . . . 54 8.4 HTML Similarity Evaluation . . . . . . . . . . . . . . . . . . . . . . 54 vii 8.5 Precision of Job Ranking . . . . . . . . . . . . . . . . . . . . . . . . 57 8.6 NDCG of Job Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . 57 8.7 Comparison of the Two Approaches - Java Developer . . . . . . . . . 58 8.8 Comparison of the Two Approaches - Python Developer . . . . . . . 58 8.9 Comparison of the Two Approaches - Average . . . . . . . . . . . . . 59 viii 1. INTRODUCTION 1.1 Motivation Currently one of the main channels for job seekers is online job finding web sites, like Indeed or Monster etc, that make the job finding process easier and decrease the recruitment time. But most such web sites only allow users to use keywords to search the jobs, which makes job searching tedious and blind task. For example, I used keyword “Java” to search jobs with location restriction Mountain View, CA on the job searching site indeed.com, the web site returned about 7,000 jobs (Figure 1.1). The number of results of job searching is huge but not well ranked, so the job seeker has to review every job description. Since no one has enough time to read all the jobs in the searching result, the actual quality of job searching service is low. This is a classic problem of information overflow. The reason for such a result is because current job searching web sites use the same information retrieval technology like “Inverted index” [53] as the common search engines, which just use keywords to map all the stored documents. Modern search engines all have some ranking algorithms to sort the search result, like page rank [33], so the top results always be the most related ones. But such algorithms are unavailable to the job search systems, because the criteria of how to rank the job searching result is very personalized. A great job opening for one job seeker maybe looks not good to the other, because the goodness of a job to a specular job seeker is heavily depend on his personal background, like his education or professional experience etc. Since the people’s résumés contain the most important background information, we believe the content of the résumé could be used to rank the job openings. We give 1 Figure 1.1: Search Result of Indeed.com an example of résumés in Table 1.1. In this thesis, we created a web system which uses the résumés of job seekers to find the jobs that match their profiles best. The main idea is to calculate the similarity between the résumé model and job models, which should be generated from résumés and job descriptions. We want to transfer the job searching task from key word searching to candidate model matching. The matching result should be sorted by the matching score, higher matching score means a better matching. The matching algorithm does not only help job seekers to find the appropriate jobs, but also offers priority to them [18]. The job with higher matching score means the job is more appropriate to the job seeker, and if he applies to the job, the chance of getting the interview will be higher as well. Figure 1.2 shows how this approach works. 2 Table 1.1: Résumé Example Ryan Richman WORK EXPERIENCE Web Developer Fabuso/Advanced Brain Technologies - Ogden, UT - February 2012 to Present Created dynamic custom web applications for e-commerce and B2B clients. Designed and edited audio-visual content for many different online applications. Spearheaded migration of largest client’s website from Joomla platform into Java code base. Built dynamic event pages, document viewers, training course applications, shopping carts, and more. Utilized advanced e-mail standards and best practices, SQL database queries, and Google Analytics. IT Representative Advanced Brain Technologies - Ogden, UT - April 2011 to February 2012 Provided internal software/hardware support for 20 employees both in-house and remote. Designed using wireframes, tested, and debugged web pages. Constructed dynamic projects and graphic designs in coordination with senior developers. Created HTML-optimized emails for hundreds of campaigns. Maintained and upgraded hardware for 20+ workstations company-wide. Founder/Head Technician Teton Media Services, LLC - Ogden, UT - October 2008 to January 2011 Created and developed websites for personal and small business customers Sold high-speed cable and satellite internet access on the phone, online and in person. Installed and serviced high-speed internet access hardware in residential and commercial properties. Designed and implemented networking solutions for homes and businesses. EDUCATION Computer Science Weber State University - Ogden, UT 2010 to 2013 ADDITIONAL INFORMATION Technical Skills Adept in the use of HTML, CSS, jQuery, Javascript, SQL, PHP, JSON, Windows, Win- dows Server, Mac, and Photoshop. 3 Figure 1.2: Matching the Jobs with Résumé 1.2 Contribution We make the following contributions in this work: 1. We proposed a résumé - job matching system. 2. We proposed a finite state transducer based matching tool to extract informa- tion from unstructured data source, which is a lightweight and flexible library, and can be extended in very easy ways. 3. We proposed a semi-automatic approach, which can collect technical terms from hr data sources, and by which we created a domain specific ontology for recruitment. 4. We proposed statistical-based ontology similarity measure, which can measure the similarities between technical terms . 4 1.3 Organization The subsequent chapters are organized as follows: we first describe what has been done in terms of prior work. We introduce some basic conception of recommender systems, and how to apply recommender technologies into Job Recommender Sys- tems. Some previous Job Recommender Systems will be introduced, their advantages and limits will be discussed as well. Two import problems of content-based Job Rec- ommender Systems, Information Extraction and Similarity Calculation, will be fully explained. Then we introduce our work, RésuMatcher, a the Personalized Résumé-Job Match- ing System. First, we give an overview of the system, which includes the architecture and the interfaces. Then we explained details of how we resolve the problems of in- formation extraction and model similarity calculation. We propose a finite state transducer library which can match patterns in sentence, and extracts related infor- mation. Ontology plays an important role in this system. We will present how to construct the domain specific ontology for recruitment. We also give a brief review of different ontology similarity measures, and explain the statistical-based ontology similarity measure we used in this system. Finally, we evaluate the accuracy of our information extraction approach. We used NDCG to evaluate the accuracy of statistical-based ontology similarity mea- sure. To evaluate the performance of the system, we compared our algorithm to some classical information retrieval approaches by precision@k and NDCG. We created a data set of job descriptions as documents, and use résumés as query to retrieval doc- uments. The result shows the ranking performance is better than other information retrieval approaches. 5 2. BACKGROUND Some scholars found that current boolean search and filtering techniques can- not satisfy the complexity of candidate-job matching requirement [29]. They hope the system can understand the job requirement, determine which requirements are mandatory and which are optional, but preferable. So they moved to use recom- mender systems technique to address the problem of information overflow. Recom- mender systems are broadly accepted in various areas to suggest products, services, and information items to latent customers. 2.1 Recommender System Job searching, which has been the focus of some commercial job finding web sites and research papers is not a new topic in information retrieval. Usually scholars called them Job Recommender Systems (JRS), because most of them used technolo- gies from recommender systems. Wei et al. classified Recommender Systems into four categories [48]: Collaborative Filtering, Content-based filtering, Knowledge- based and Hybrid approaches. Some of these techniques had been applied into JRS; Zheng et al. [44] and AlOtaibi et al. [3] summarized the categories of existing on- line recruiting platforms and listed the advantages and disadvantages of technical approaches in different JRSs. The categories include: 1. Content-based Recommendation (CBR). The principle of a content-based rec- ommendation is to suggest items that have similar content information to the corresponding users, like Prospect [43]. 2. Collaborative Filtering Recommendation (CFR). Collaborative filtering recom- mendation finds similar users who have the same taste with the target user and 6 recommends items based on what the similar users, like CASPER [35]. 3. Knowledge-based Recommendation (KBR). In the knowledge-based recommen- dation, rules and patterns obtained from the functional knowledge of how a specific item meets the requirement of a particular user, are used for recom- mending items, like Proactive [24]. 4. Hybrid recommender systems combine two or more recommendation techniques to gain better performance, and overcome the drawbacks of any individual one. Usually, collaborative filtering is combined with some other technique in an attempt to avoid the ramp-up problem. 2.2 Job Recommender Systems Rafter et al. began to use Automated Collaborative Filtering (ACF) in their Job Recommender System, “CASPER” [35]. In the system user profiles are gotten from server logs, that includs: revisit data, read time data, and activity data. All these factors are viewed as measure of relevance among users. The system recommend jobs in two steps: First, the system finds a set of users related to the target user; second, the jobs that related users liked will be recommend to the target user. The system use cluster-based collaborative filtering strategy. The similarity between users are based on how many jobs they both reviewed, or applied. CASPER also allows users to search jobs by a query which is a combination of some fields: like location, salary, skill and so on. The system uses such query to find jobs, and the returned jobs are ranked with the collaborative filtering algorithm. In their paper, the authors do not give a detailed description on how to detect the related fields they need and how to the transfer semi-structured job description to the structured data. 7 The shortages of collaborative filtering: First, since the number of search results is huge, and the results are sorted randomly, the probability of two similar users reviewing the same jobs is low, which causes the sparsity problem of collaborative filtering. The authors also noticed the sparseness problem caused by few in users profile, so they try to user cluster-based solution to resolve this problem. Second, because recommended jobs are from others users’ search results, since the quality of current searching result are low, the quality of recommendation cannot be high. Färber et al. [14] presented a recommender system built on a hybrid approach. The system integrate two methods, content-based filtering and collaborative filtering, which tries to overcome the problem of rating data sparsity by leveraging a combined model. In the system, the data source is synthetic resumes. The latent aspect model is shown in Figure 2.1. Figure 2.1: Latent Aspect Model In Malinowski et al. [29], they classified the job recommender systems into t- wo categories, CV-recommender, which recommends CVs to recruiter, and the job- recommender, which recommends jobs to job seekers. The system collects the users’ profile data by asking input their profiles to the web form based interface field by 8 field. The input data collected are: 1. Demographic data (e.g. date of birth, contact information) 2. Educational data (e.g. school courses, grades, university, type of degree, inter- mediate and final university examinations, postgraduate studies) 3. Job experience (e.g. name of the company, type of employment, industry group, occupational field) 4. Language skills (e.g. language, level of knowledge) 5. IT skills (e.g. type of skill, level of knowledge) 6. Awards, scholarships, publications, others The system also asked the users to upload their resumes, but the resumes were only for facilitating the human judgment. In Malinowski’s study, the latent aspect model is a statistical model, which needs to be trained before applied to recommendation. The system uses the users’ search results as the training data to train the model for recommendation, so the system needs a a long time training for each user. 2.3 Information Extraction in Job Recommender System Big IT companies met the similar problem of information overflow when they received many resumes for one job opening. The recruiter had to screen the all the applications manually, but this task was also tedious and time consuming. For this reason these companies tried to build systems to help screen the resumes. Amit et al. in IBM presented a system, “PROSPECT” [43], to aid shortlisting candidates for jobs. The system uses a résumé miner to extract the information from résumés, which use a conditional random field (CRF) model to segment and label 9 the résumés. The CRF model used three kinds of features, they are: Lexicon, Visual, and Named Entity. The paper compared some algorithms to ranked the candidates applicants, such methods include: Okapi BM25, KL, Lucene Scoring, and Lucene Scoring + SkillBoost. HP also built a system to solve the similar problem, which is introduced in Gon- zalez et al.’s paper [17]. The system also pays a lot of attention to information extraction. The dictionaries which are used to tag entities need to be updated often since there always new terms appears. So an adaptive learning module is used to achieve two objectives: use semantic data to enhance the information extraction and to discover new terms. A domain-oriented ontology is used to represent knowledge, inference rules are defined based on the ontology knowledge base. When a detected term found, the system will search in external knowledge base, like DBpedia. The résumés are also classified to different categories like “Web Technology” and “No Web Technology” by naive Bayes classifier. The company can allocate appropriate employees to required positions with the system. The goal of the systems built by IBM and HP is to help the companies to select good applicants, but cannot help job seekers to find appropriate jobs. Yu et al. [51] used a cascaded IE framework to get the detailed information from the résumés. In the first stage, the Hidden Markov Modeling (HMM) model is used to segment the résumé into consecutive blocks. Based on the result, a SVM model is used to obtain the detailed information in the certain block, the information include: name, address, education etc. Celik Duygua and Elci Atilla proposed a Ontology-based Rsum Parser (ORP) [7], which uses ontology to assistant the information extraction process. The system 10 processes a résumé in following steps: converting the résumé file into plain text, separating the text into some segments, using the ontology knowledge base to find the concepts in the sentences, normalizing all the terms, and classifing the sentences to get the wanted terms. But the personal information the system retrieved like name and addresses is not the information that the recruiters care about. The recruiters want some information that relate to the job opening, and can help them to judge the competence of job applicants. 2.4 Matching Algorithms in Job Recommender Systems Lu et al. [28] used latent semantic analysis (LSA) to calculate similarities between jobs and candidates, but they only selected two factors “interest” and “education” to compare candidates. Xing et al. [50] used structured relevance models (SRM) to match résumés and jobs. Drigas et al.[13] presented a expert system to match jobs and job seekers, and to recommend unemployed to the positions. The expert system used Neuro-Fuzzy rules to evaluate the matching between user profiles and job openings. The system uses a relation matrix to represent the fuzzy relation between these specialities. The system needs the training data to train that Neuro-Fuzzy network. Both résumé data and job opening data were manually input into the system. Daramola et al.[10] also proposed a fuzzy logic based expert system(FES) tool for online personnel recruitment. In the paper, the authors assumed that the information already be collected. The system uses a fuzzy distance metric to rank candidates’ profiles in the order of their eligibility for the job. Yao et al. [28] also presented a hybrid recommender system which integrated content-based and interaction-based relation. In content-based part, relations be- 11 tween job-job, job-job seeker, and job seeker-job seeker can be identified by their similarity of profiles. There two approaches are used to calculate the similarities: For the structured data, like age and gender, the weight sum values will be returned; for the unstructured data, like similarity between job and user profile, the latent semantic analysis is applied in the system. In summary, there are some problems exist in previous Job Recommender Sys- tems: 1. Most systems can only process the structured data of résumés and job descrip- tions, but in reality both them are in unstructured formats, such as text files or HTML web pages. 2. The systems that have information extraction modules are designed for re- cruiters to select applicants, not for job seekers to select jobs. 3. In the systems the information fields to match résumés and job descriptions are coarse-grained. To improve the quality of recommendation, we need to improve the granularity of the information fields. 12 3. PROBLEM 3.1 Problem Definition The basic problem in this thesis is how to find appropriate job descriptions by user’s résumé, which means we need calculate the similarity between the users rsum and the job. If we take the résumé as a query and the job descriptions as documents, we need to build an information retrieval model to get the most relevance documents. The RésuMatcher will parse the job descriptions to the job models, and store them in the database. When a user searches the jobs by their résumé in the system, the system will compare the similarity values between the résumé and the job models, and return the jobs sorted by their similarity values. The core idea of our algorithm is calculate similarity between the résumé model and job model. We give a formal definition of our problem. All of the notations will be used frequently throughout the thesis. We use r to denote the user’s résumé model, which has some features ri like their academic degree, their major, their skills and so on. The symbol J is the set of job models stored in the database, and j is a job model in the set J. The similarity function sim(r,j) gives the similarity values between résumé r and job j. The return list of search function search(r,J) will calculate all the similarity value in the database, and the result of the function will be the job description list ranked by their similarity values. The equation of how to calculate similarity value is given below: sim(r,j) = n∑ i=1 simfuni(ri,ji) ×wi 13 The value of sim(r,j) is the summation of the similarity values of different fields times their corresponding weights. Different fields like major and skills, may have different functions to calculate their similarity values. We will describe the similarity functions of individual fields in later parts. 3.2 Challenges There are two challenges exist in our system. The first one is how to extract models of jobs and résumés. To calculate the similarity between a job and a resume, the RésuMatcher system needs structured digital models of each document. To get the structured data, some JRSs ask the job seekers input their profiles in forms field by field, and the recruiter input their job descriptions in the same way. However, as we discussed in Chapter 2, the users are reluctant to take the tedious process [43]. Job seekers prefer upload their resumes directly, and recruiters prefer to post the whole job descriptions to web sites. So we need extract the structured information from un-structured data source, like résumés and job descriptions. The other challenge is how to compute the similarity between rsum and job models. We observed that simple keyword matching is not a good similarity measure, because job descriptions and résumés both contain richer and more complex words that cannot be described simply by keywords. In these documents, some concepts can be written in different ways, and other concepts can have close relationships. For example, Table 3.1 shows portions of a résumé and a job description: If just looking at the text, we can find that the résumé has very few common words with the job description. But from the view of an experienced engineer, the candidate is closely matches the job: the two relational databases Oracle and Mysql are very similar, OOA/OOD is the same meaning of many years of Java and C++ experience, and Tomcat and JBOSS are both Java web applications servers. If we 14 Table 3.1: Portions of Resume and Job Description Resume Portion Job Description Portion B.S. degree in computer science 5+ years Java 2+ year C++ Some experience in Oracle database Other experience like: Hibernate, JBOSS, JUnit, Tomcat etc. BS degree above 4+ years Java Some experience of Python Mysql, MS-SQL Java web application Server OOA/OOD use keyword matching, the system does not provide a strong matching result in very common cases such as this. So we need a better approach to calculate the similarity between different technical concepts. 15 4. SYSTEM OVERVIEW 4.1 System Overview The system uses information extraction technique to parse job descriptions and résumés, and it gets information such as skills, job titles and education background. The information is used to create the models of job openings and job seekers. A domain specific ontology is used to construct the knowledge base, which includes the taxonomies that support résumé-job matching. The models of résumé includes job seekers’ specialties, working experience and education background, and all the fields are extracted from their résumés. The job models are extracted from job descriptions, and they have the same information fields as the résumé models. When a job seeker searches the jobs by their résumé, the system calculates the similarity between the résumé model and job models, then gives every job model a similarity value. 4.2 System Architecture Figure 4.2 shows the architecture of the whole system, which includes such mod- ules: 1. The Web Crawler can access and download all new IT job opening web pages from indeed.com everyday. 2. The Job Parser can parse the job opening web pages, extract the information and create the job models. 3. The Resume Parser is much like the Job Parser; it parses the résumés and creates the résumé models. 16 4. All the job descriptions and job models are stored in the database. 5. When a user searches the jobs with their résumé, the Ontology Matcher calcu- lates the similarity values of jobs in the database and returns the jobs ranked by their similarity values. Figure 4.1: System Architecture 4.3 Text Processing Stages Information Extraction is the task of automatically extracting structured infor- mation such as entities, relationships between entities, and attributes describing entities from unstructured sources [42]. The IE framework in our system uses six stages in order to extract the information from job descriptions: HTML parsing, segmenting, preprocessing, tokenizing, labeling and pattern matching, which is show in Figure 4.2. 17 1) The HTML Parsing will parse the web pages that contain job descriptions, which are obtained from web crawler. The parser uses HTML tag template to extract attributes of the jobs, like job title, location, company name, content and so on. A job will be saved as a record with these attributes in the database. In the record, the content field contains the text part of the job description, which will be processed in later stages. 2) In the segmenting stage, the content field of the job description is be sep- arated into paragraphs according HTML tags. Then paragraphs are separated into sentences by either HTML tags or punctuation, and after this step, all HTML tags will be removed. 3) Web pages of job description are created in different character sets, (e.g. UTF8 and ISO 8859-1), and almost always contain some unreadable characters. In the prepossessing stage, characters in the sentences are converted to ASCII characters, unreadable characters will be deleted, and some punctuation will be replaced by spaces (e.g. / and -). 4) In the tokenizing stage, the sentences will be tokenized into arrays of tokens by NLTK [5]. 5) In the labeling stage, the sentences will be given two layers of labels by a dictionary matching approach. The labels in the first layer are the semantic value of the text, and the labels in the second layer are the ontology hypernym of the labels in the first layers. 6) In the pattern matching stage, the FST library is used to matching the labels of the labeled sentences. If a layered sentence match any pre-defined pattern, the information will be extracted and added to the job model. After every sentence of a job description has be processed, a job model will be created and saved in the database. More details about matching will follow in Section C. 18 Figure 4.2: Job Description Process Pipeline 19 4.4 System Implementation We will describe some implementation details here. The whole system is imple- mented in Python and uses some third party libraries and frameworks. We used Flask, a lightweight web framework, to build the web application. We used Rdflib as the Web Ontology Language (OWL) file parser, Python Lex-Yacc (PLY) as the token regular expression compiler, whoosh as the inverted index builder and Beauti- ful Soup as the HTML parser. All the jobs retrieved by the Web Crawler are stored in the MongoDB NoSQL database. For the natural language processing procedure, we used Natural Language Toolkit (NLTK), a natural language processing library, to extract and tokenize the sentences. 4.5 System Interface The system provides some interfaces to end users. The most important interfaces are the web pages like: reviewing all the jobs in the database, searching the jobs by keyword (Figure 4.3), uploading users’ résumés (Figure 4.4) and matching the jobs with a résumé (Figure 4.5). 20 Figure 4.3: Job Description List Figure 4.4: Upload Resume 21 Figure 4.5: Résumé Job Matching 22 5. INFORMATION EXTRACTION In this chapter we will explain how the Information Extraction (IE) module of our system extracts information from these unstructured data source. An example of job description is shown in Table 5.1. The IE framework will be introduced by example of processing the job descriptions. The Finite-State Transducer(FST) library, which is used as pattern matching tools, will be introduced as well. 5.1 Semantic Labeling In this section, we will introduce why and how we add two layers of labels to the tokenized sentences. In natural language, a single concept often has multiple expressions to represent it. For example, the simple concept bachelor’s degree, can be expressed in many ways in job descriptions, e.g. B.S., BA/BS, 4-years-degree, and so on. Table 5.2 shows the words that if followed with word “degree” have the semantic value of “bachelor’s degree”. To add labels to a sentence, we use regular expression over tokens. A regular expression over token transfer a patten to a Finite-State Transducer (FST), and every token of that will be transferred to an edge of FST. If we use all the expressions of a semantic value to create a pattern, the pattern will be very large, and there are too many states in the FST. For example, if we use some words in Table 5.2 to create the pattern of semantic value “bachelor’s degree”, the pattern will like below: ( Baccalaureate | bachelors | bachelor | B.S | BS | BA ) degree If all words in Table 5.2 are added to the pattern, the FST will have too many edges, and the matching process will be very slow because of the problem of combinatorial 23 Table 5.1: Example of Job Description Senior/Principal Software Engineer RichRelevance - San Francisco, CA RichRelevance powers personalized shopping experiences for the worlds largest and most innovative retail brands, including Target, Sears, Marks & Spencer, John Lewis and oth- ers. Founded and led by the e-commerce expert who helped pioneer personalization at Amazon.com, RichRelevance helps retailers increase sales and customer engagement by recommending the most relevant content to consumers regardless of the channel they are shopping. RichRelevance has delivered more than $5.5 billion in attributable sales for its retail clients to date, and is accelerating these results with the introduction of a new form of digital advertising called Shopping Media which allows manufacturers to en- gage shoppers where it matters most – in the digital aisles on the largest retail sites in world. RichRelevance is headquartered in San Francisco, with offices in New York, Seattle, Boston, Reading and Malmho, and has been twice recognized as one of the Best Places to Work in the Bay Area. RichRelevance is looking for a Senior/Principal Software Engineer to join our growing team! Primary responsibilities: • Working with large scale distributed systems • Work with Hadoop ecosystem (technologies like Hive, Impala, HBase) • Algorithmic development with primary focus Machine Learning • Working with rapid and innovative development methodologies like: Kanban, Continuous Integration and Daily deployments • Unit testing with JUnit, Performance testing and tuning Minimum requirements: • BS/MS in CS, Electrical Engineering or foreign equivalent plus relevant software development experience • At least 5+ years of software development experience • Expert in Java, Scala or any other object oriented language • Proficient in SQL concepts (HiveQL or Postgres a plus) • Additional language skills for scripting and rapid application development Desired skills and experience: • Working with large data sets in the PBs • Familiarity with UNIX (systems skills a plus) • Working in a distributed environment and has dealt with challenges around scaling and performance • Mobile development for Android or iOS. RichRelevance is an Equal Opportunity Employer and does not discriminate against any applicant on the basis of race, color, religion, national origin, gender, marital status, age, disability, sexual orientation, military/veteran status, or any other status protected by Federal or State law or local ordinance. 24 Table 5.2: All Words Mean Bachelors ”Baccalaureate”,”bachelors”, ”bachelor” ,”B.S.”, ”B.S”,”BS”,”BA”,”BA/BS”, ”BABS”, ”BSBA”, ”B.A.” ,”4-year”,”4-year”, ”4 year”, ”four year”,”college”,”Undergraduate” , ”University” explosion. To resolve this problem, we proposed an approach to use the patterns to match the labels of the tokens, not the the original text. In the system, we don’t care what words the sentences really use, but want to extract the semantic value of the tokens which match the pattern. The details of the approach is described below. At first, we created two dictionaries, which are used to label the tokens. In the first dictionary, the keys are the tokens, like words in Table 5.2, and the values are the symbols for semantic values, like “BS-LEVEL” for “bachelor’s degree”, or “MS-LEVEL” for “master’s degree”. The values of the the second dictionary are the ontology hypernym of their keys, like keys “BS-LEVEL” and “MS-LEVEL” both have value “DE-LEVEL”, which means that bachelor’s degree and master’s degree are both one kind of degree level. We show the dictionaries for degree information in Table 5.3. There are also some words in the dictionaries that have the same first layer and second layer labels, which is shown in Table 5.4. With the two dictionaries, we can label the tokens with two layers. Table 5.5 shows how the sentence “Bachelors degree in computer science or information sys- tems.” is labeled. The pattern “DE-LEVEL DEGREE IN MAJOR OR MAJOR ” can match the sentence above, and the output of the matching process is “BS-LEVEL” for bachelor’s degree, “MAJOR-CS” and “MAJOR-INFO” for two majors mentioned in sentence. In our system, most patterns match the labels in second layer. With this approach, 25 Table 5.3: Semantic Labeling Original Text Layer 1 Layer 2 bachelors BS-LEVEL DE-LEVEL bachelor B.S. Baccalaureate Master MS-LEVEL MS M.S. PhD PHD-LEVEL Ph.D Doctorate Table 5.4: More Labels ”Be”, ”be”, ”is”, ”are”, ”am” BE BE ”a”, ”A”, ”an”, ”An”, ”The”, ”the” DE DE ”MBA”, ”BSCS”, ”BSEE”, ”MSCS” MAJOR-DEGREE MAJOR-DEGREE ”MSEE”, ”MSCE”,”MPH” ”practical experience”, ”work experience” EXPERIENCE EXPERIENCE ”professional experience”, ”experience ”preferred”, ”required”, ”desired” PREFER-VBD PREFER-VBD ”a plus”, ”mandatory”, ”desirable” PREFER-JJ PREFER-JJ ”similar”, ”related”, ”Relevant” DEGREE-JJ DEGREE-JJ ”equivalent”, ”based” the size of the FST for the pattern will be minimized, so speed of matching process can be improved. 5.2 Patterns for Matching As we explained in section B, we mentioned matching tokens in the second layer to patterns we defined. To match the labels in sentences to our patterns, we proposed a library that support matching pattern over tokens. The difference between this library and traditional regular expression is that the basic unit to be matched is token, not character. Some patterns used to match degree phrases are in Table 7.2. 26 Table 5.5: Labeled Sentence layer 2 DE-LEVEL DEGREE IN MAJOR OR MAJOR . layer 1 BS-LEVEL DEGREE IN MAJOR-CS OR MAJOR-INFO . words bachelors degree in computer science or information systems . The patterns looks like regular expression, but they use tokens as the basic units. Table 5.6: Patterns to Match Degree Sentences DE-LEVEL, DE-LEVEL, OR DE-LEVEL DEGREE DE-LEVEL DEGREE ( IN | OF ) DT MAJOR MAJOR-DEGREE , MAJOR-DEGREE OR MAJOR DE-LEVEL (, DE-LEVEL)* (OR DE-LEVEL)? BE? PERFER-VBD 5.3 Pattern Matching Library In section C we introduced how we use the library of pattern matching over tokens to match the sentences. In this section we will introduce more details of this library, including its advantages and implementation details. 5.3.1 Finite-State Transducer Finite-State Transducers [39] have been used as a tool to match patterns and extract information for more than 20 years. This approach has been demonstrated to be very effective in extracting information from text like CIRCUS [25] and FAS- TUS [19]. In the widely used NLP toolkit GATE [9], the semantic tagger JAPE (Java Annotations Pattern Engine) could describe patterns that are used to match and annotate tokens. JAPE adopts a version of CPSL (Common Pattern Specifica- tion Language) [4], which provides FST over annotations. Chang et al. presented cascaded regular expressions over tokens [8], which proposed a cascaded pattern 27 matching tool over token sequences. After studying these tools, we found most of them to be powerful and complex, but not very flexible. One reason is that developers need to learn some Domain specific Languages (DSLs) like CPSL. The other reason is the extra effort and time required to integrate the pattern matching tool into the system. So here we proposed a more flexible and lightweight FST framework, which can do regular expression matching over labeled tokens. We give the definition of Finite-State Transducer here. A Finite-State Transducer is a 6-tuple (Σ1, Σ2,Q,i,F,E) where: • Σ1 is a finite alphabet, called the input alphabet. • Σ2 is a finite alphabet, called the output alphabet. • Q is a finite set of states. • i ∈ Q is the initial state. • F ⊂ Q is the set of final states. • E ⊂ Q× Σ∗1 × Σ∗2 ×Q is the set of edges. For example, the FST Td3 = ({0, 1},{0, 1},{0, 1, 2},Ed3) where Ed3 = { ( 0, 0, 0, 0 ), ( 0, 1, 0, 1 ), ( 1, 0, 0, 2 ), ( 1, 1, 1, 0 ), ( 2, 1, 1, 2 ), ( 2, 0, 1, 1 ) } is shown in Figure 5.1. Regular expressions can be converted to automata [2], and FST is also an au- tomata. To convert a regular expression over token to a FST we need two steps: The first is parsing the expression to a tree of matchers, the second is transfer the tree of matchers to the FST. We will introduce these two steps in next. 28 Figure 5.1: Zero or One NFA 5.3.2 Matchers in the Pattern Matching Library In our library, a “matcher” could be a token to be matched, or a composition of other matchers. Our library supports syntax used in traditional regular expressions over strings. We list the syntax that the library supports in Table 5.7. The first column is the names of the matchers, the second column is the explanation of the function of the matchers, and third column is the their counterpart syntaxes of traditional regular expression. The RegexMatcher in our library is constructed with a regular expression, and the matcher matches any string that matches the regular expression in the matcher. We give examples of the syntax of these matchers in Table 5.8. Table 5.7: Matchers of Our Library Matcher Name Function Counter Part of regex UnitMatcher token is matches the it character in regex SequenceMatcher A list of Matcher sequence of characters QuestionMatcher One or more of the preceding token ? StarMatcher Zero or more of the preceding token * PlusMatcher Zero or one of the preceding token + DotMatcher Any token . RegexMatcher Any token matches the regular ex- pression N/A 29 Table 5.8: Examples of Matcher Matcher Name Example UnitMatcher DEGREE SequenceMatcher DE-LEVEL DEGREE QuestionMatcher DE-LEVEL (OR DE-LEVEL)? DEGREE StarMatcher DE-LEVEL (, DE-LEVEL)* DEGREE PlusMatcher DEGREE IN MAJOR + DotMatcher HAS . DEGREE RegexMatcher r“d-d” years The framework supports three styles of creating patterns: regular expression style, operator style and object style. The second and third styles are flexible be- cause developers can create their own matcher class to extend the feature of the library. We use examples to show how the three styles work. The most common style is defining pattern expression in a string, which is much like traditional regular expression. The pattern is: DE-LEVEL DEGREE ( IN | OF ) DT? MAJOR The code is: seqMatcher =parser.parse("DE-LEVEL DEGREE ( IN | OF ) DT? MAJOR") The second style is using algebraic operators to connect matchers, which can help developer reuse previous patterns when the new patterns include old ones. It is shown in follows: The pattern is: ”DE-LEVEL DEGREE (IN | OF) MAJOR” The code is: seqMatcher = UnitMatcher("DE-LEVEL") + UnitMatcher("DEGREE") + 30 ( UnitMatcher("IN") | UnitMatcher("OF" ) ) + UnitMatcher("MAJOR") We also could create a complex matcher in object-oriented programming style. The pattern is: ”DE-LEVEL DEGREE (IN | OF) MAJOR” The code is: matcher1 = UnitMatcher("DE-LEVEL") matcher2 = UnitMatcher("DEGREE") matcher3 = UnitMatcher("IN") matcher4 = UnitMatcher("OF") matcher5 = UnitMatcher("MAJOR") matcher6 = AlternateMatcher([matcher3,matcher4]) seqMatcher = SeqMatcher([matcher1, matcher2, matcher6, matcher5]) The flexibility of the tool also comes from the fact that developers can determine which layer of the array should be matched, the original text, or labels in the first layer or labels in the second layer. Developers can assign a lambda expression to the matcher’s catching function, which defines how to get the matching input strings from the sentences, as well as an out function, which defines what should be outputed. For example, the labeled sentence is a sequence of arrays, each array includes the original text token and its labels in the other two layers, which is shown in table 5.5. To match the labeled sentence, we set the lambda expression for catching function to “lambda x:x[2]”, and the out function to “lambda x:x[1]”, which make the matcher match the label in second layer, and output the the value of semantic value in the first layer. 31 5.3.3 Implementation of the Pattern Matching Library We use PLY(Python Lex-Yacc) as the grammar parser, which is a pure-Python implementation of the popular compiler construction tools lex and yacc. We defined syntaxes of regular expression over tokens in the parser, which can parse the token regular expression to the tree structure of matchers. We use the algorithm proposed by Thompson and Ken [45] to construct the FST from the tree of matchers, which is shown in Algorithm 1. The algorithm accesses the inner structure of a matcher recursively, and create state for the matcher. Each state is a partial Nondeterministic Finite Automaton (NFA), which has one or more dangling arrows. The algorithm builds the whole NFA by connecting the arrows of the partial NFAs. Different operator will have different structures of NFA, which is shown below: The NFAs for matching single token is shown in Figure 5.2. Figure 5.2: Single Token NFA The NFA for the concatenation e1e2 connects the final arrow of the e1 machine to the start of the e2 machine, as shown in Figure 5.3. Figure 5.3: Concatenation NFA 32 The NFA for the alternation e1 | e2 adds a new start state with a choice of either the e1 machine or the e2 machine, which is shown in Figure 5.4. Figure 5.4: Alternation NFA The NFA for e? alternates the e machine with an empty path, which is shown in Figure 5.5. Figure 5.5: Zero or One NFA The NFA for e* uses the same alternation but loops a matching e machine back to the start, which is shown in Figure 5.6. Figure 5.6: Zero or More NFA 33 The NFA for e+ also creates a loop, but one that requires passing through e at least once, which is shown in Figure 5.7. Figure 5.7: One or More NFA With the above rules, we can convert the expression “DL (, DL)* (or DL)? DEGREE” to an FST in Figure 5.8. Figure 5.8: Finite Automata Transducers Comparing our method to other state of art machine learning-based information extraction algorithms, our method has such advantages: 1. Easy to implement. According to the pseudo code in 1, even some undergrad- uate students without strong machine learning background can implement our algorithm in a few days. 2. There is no need for data labeling. Supervised machine learning algorithms always need a large labeled training data set, but data labeling is tedious work 34 Input: matcher, nfa Output: NFA switch matcher do case UnitMatcher state = compileToNFA(matcher]) ; connect( nfa.last, [state] ) ; nfa.last = [state] ; break; case SequenceMatcher for (i = 0; i < len(matcher.list); i + +) do state = compileToNFA(matcher.list[i]) ; connect( nfa.last, [state] ) ; nfa.last = [state] ; end break; case AlternateMatcher state1 = compileToNFA(matcher.list[0]) ; state2 = compileToNFA(matcher.list[1]) ; connect( nfa.last, [state1, state2] ) ; nfa.last = [state1, state2] ; break; case QuestionMatcher state1 = compileToNFA(matcher.list[0]) ; connect( nfa.last, [state1, nfa.last] ) ; nfa.last = [state1, nfa.last] ; break; case PlusMatcher state1 = compileToNFA(matcher.list[0]) ; connect( state1, state1 ) ; connect( nfa.last, state1 ) ; nfa.last = [state1 ] ; break; case StarMatcher state1 = compileToNFA(matcher.list[0]) ; connect( state1, state1 ) ; connect( nfa.last, state1 ) ; nfa.last = [state1, nfa.last] ; break; endsw Algorithm 1: CompileToNFA 35 for users. With our pattern matching algorithm, we avoid the work manual data labeling. 3. The accuracy of information extraction can increase monotonously as the num- ber of patterns increase. So with enough patterns, the accuracy becomes quite high. 4. High speed for labeling data. The time complexity of pattern matching is O(n) [45], which is smaller than some complex machine learning based approach- es. One example is Conditional Random Fields(CRFs), which uses Viterbi algorithm [47] to label the sequence, the time complexity of it is O(n2t). In this chapter we have introduced how we extracted the information from the resumes and job descriptions, and the implementation details of the pattern matching library, regular expression over tokens. We can get the models of resumes and job descriptions through the procedure described in this chapter. In the next chapter, we will discuss how our system searches and ranks job models by resume models. 36 6. MODEL SIMILARITY The similarity value between a job model and a résumé model is the summation of weighted similarity values of different fields. The equation is given below: sim(r,j) = n∑ i=1 simfuni(ri,ji) ×wi The value of sim(r,j) is the summation of similarity values of different fields times their corresponding weights. simfuni(ri,ji) is the similarity function of the ith field of the model. In our system, the résumé model and job model both have four fields: job title, major, academic degree and skills. The similarity value between a résumé model and job model is the sum of the productions of similarity values of all the fields pairs and their weights. We will introduce how to calculate the similarity value for each field in this chapter. 6.1 Similarity of Major and Academic Degree In the simplest case, if the majors in the résumé model and job model are the same, the similarity value is 1. If they are different, we can check whether the major in the résumé model is in the list of related majors for the major in the job model. If it is, the similarity value is 0.5; otherwise the similarity value is 0. The equation is shown below: MajorSim(r,j) =   1, rmajor = jmajor 0.5, rmajor ∈ related(jmajor) 0, otherwise   There are five kinds of academic degrees in the system: high school, associate, 37 bachelor, master, and Ph.D., which are mapped to the integer values form 1 to 5. If the degree value in the résumé model is less than that in the job model, which means that the job seeker’s education background cannot satisfy the requirement of the job, the similarity value in this case is 0. If the degree value in the résumé model is equal to the job model and no more than 2 above, the similarity value is 1. In some cases, the degree value in the résumé model is greater than that of the job model, and the difference is greater than 2, which means that the job seeker’s degree is much higher than the requirement of the job. The situation is also a kind of relative matching, so the similarity value here is 0.5. The equation is shown below: DegreeSim(r,j) =   0, rdegree < jdegree 1, 0 < rdegree − jdegree 6 2 0.5, rdegree − jdegree > 2   6.2 Similarity of Job Title Another field of needs similarity calculation is the job titles in the models. A job title can be parsed into some sub fields: job role, level, platform, programming language. The value of job roles includes: developer, manager, administrator and so on. There are levels values in such roles: such as junior, senior and architect. The platforms: web, mobile and cloud are used by the very roles. The similarity value between two titles is the sum of all the similarity values of these fields. The similarity value of each sub filed ranges from 0 to 1, and we also normalized the similarity summation value to 1 by dividing the number of sub fields. If the job seeker has some working experience, there may be some job titles in their résumé. When calculating the similarity value between a résumé model and a job model, the system calculates the similarity values of the title of job model to all the titles in the 38 résumé model and returns the maximum one. 6.3 Similarity of Skills The job model usually has requirements of some skills, and the résumé model lists the skills the job seeker has as well. The similarity value of skills field is the normalized summation of all the similarity values of skills in the job model. SkillSetSim(SJ,SR) = ∑ sji∈SJ SkillSim(sji,SR) |SJ| For every skill in the job model, the similarity value is the maximum value it can get from the skills in the résumé model. The equation is shown below: SkillSim(sji,SR) =   1, sji ∈ SR max(sim(sji,rjk)), sji /∈ SR   In the equation, sji is the ith skill in the job model, and SR is the skill set of the résumé model. If there is the skill sji in the skill set SR, the similarity value for sji is 1, otherwise the system chooses the maximum similarity value from all the similarity values between skill sji and the skills in the the résumé model. We introduced how to calculate similarity values for three fields in résumé and job models. In the next chapter, we will introduce how to use a domain specific ontology to calculate similarity values between the skills fields of the two models. 39 7. ONTOLOGY CONSTRUCTION AND SIMILARITY 7.1 Semantic Similarity in JRSs After getting job models by the information extraction module, users can search for jobs in the system. In previous studies of JRSs, ontology is used as a knowledge base to store knowledge and rules, which could help compare the similarity between different concepts. Liu and Dew [27] used Resource Description Framework (RDF) to represent and store the expertise of experts, and they used a RDF-based expertise matcher to retrieve the experts whose expertise included the required concept. Proactive [24] used two kinds of ontology, job category and company information. The system used an ontology checker to classify the job information, stored the domain knowledge and calculated the weight value in recommendations. Fazel [15] used a hybrid approach to match job seekers and job postings, which takes advantage of the benefits of both logic-based and ontology-based matching. In his paper the description logics (DL) are used to represent the candidate and job opening, and the ontology is used to organize the skills in a taxonomy. The paper provides an equation to calculate the matching degree: sim (P,j) = ∑ xij ×u(dsi) where xji is the Boolean variable indicating whether desire i is satisfied by appli- cant Aj in the set of all qualified applications. Kumaran et al. [21] also used an ontology to calculate the similarity between the job criteria and candidates’ résumé in their system [21]. The similarity equation they 40 used is: M (i1, i2) = ∑n k=1 Sim (p i1 k ,p i2 k ) ∗W i2 k∑n k=1 W i2 k The similarity function Sim(p1,p2) is defined as follows: Sim(p1,p2) =   1, if similarity of p1 and p2 > t 0, otherwise   7.2 Ontology Construction Before calculating the similarity between concepts, we need to construct the on- tology first. Semantic web has been a popular research topic in previous years, and at the same time thousands of domain ontologies have been created [11]. A paradig- matic example is WordNet [16], which is a general purpose thesaurus, and contains more than 100,000 general English concepts. ACM has created a poly-hierarchical ontology that can be utilized in semantic web applications [1], but it is mostly used in academic areas. DBpedia [6] provides structured information from Wikipedia and make this information available on the Web, but its coverage is huge, and most of them is not related to job finding. Currently, there is no domain specific technology ontology built for recruiting purpose. The domain specific technology ontology for recruiting should include a lot of technical terms, like programming language, programming library, commercial prod- ucts and so on. Furthermore, there are new techniques invented everyday, so new IT terms will appear continuously. Ding et al. [12] gave a survey of current ontol- ogy generation approaches such as manual, semi-automatic, and automatic. Some aspects of the approaches were discussed in the paper, like the source data, concept extraction methods, ontology representation, and construction tools. Inspired by 41 this paper, we propose a semi-automatic approach to construct the IT skill ontology, which use a pattern matching approach to collect possible technical terms, and use DBpedia to verify the them. From the observation, we found that sentences with skill requirements in job descriptions always list several skills in the sentence, which is shown in Table 7.1. Based on this character, we propose a bootstrap approach to collect IT terms in job descriptions. First, we manually collect about fifty terms from job descriptions, and add them to the term list. Then we use our pattern match library to find the sentences that matching the pattern in Table 7.2 from a set of job descriptions. An example of a sentence which matches the pattern is shown in Table 7.3. We extract the tokens which match the star symbol from the sentences; these tokens have high probability to be technical terms. Then we could check the tokens in Dbpedia to see whether they are under the categories like software, programming language or any other technical related ones. If they are, we could classify them as terms, and add them to the terms list. After scanning all the sentences in the job description set, the term list will be larger, and we can use the larger term list to start a new iteration of scaning. This process stops when the number of found new terms is below a threshold. The process is shown Figure 7.1. Table 7.1: Example Sentences in Job Descriptions 1. A high-level language such as Java, Groovy, Ruby or Python; we use Java and Groovy extensively 2. HTML5/CSS3/JavaScript, web standards, jQuery or frameworks like An- gularJS would be great 3. HTML CSS and Javascript a must 4. Experience with AJAX, XML, XSL, XSLT, CSS, JavaScript, JQuery, HTM- L and Web Services 42 Table 7.2: Patterns to Extract Terms term , * , *, term term , * , *, and term Table 7.3: An Example Sentence Matches the Pattern Experience with TERM , * , * , TERM , and * Experience with AJAX , XML , XSL , XSLT , and CSS For example, we extract the token ”XSL”, which currently is not in the terms list. We check the word on DBpedia by accessing the URL:http://dbpedia.org/page/XSL. If we can get the XML formatted description of XSL, and any element in “dc- terms:subject” section has the value which is a technical category, like “Program- ming languages”, “Markup languages” and so on, we can indicate that the word is a technical term, and add it to the term list. Figure 7.1: Procedure of Finding Technical Terms But not all the extracted terms can be verified in DBpedia, because some terms 43 have multiple meanings in English, and the URLs of their DBpedia pages are unpre- dictable. For example, the word “Python” could be an animal name or a program- ming language. The meaning of the programming language has the DBpedia URL http://dbpedia.org/page/Python (programming language), which is difficult to pre- dict. In this case, we have to check the term manually. After getting all terms, we use Protege [32], an open source ontology editor, to edit the domain specific ontology, and saved it in RDF format. The interface of Protege is shown in Figure 7.2. Part of the technical ontology is shown in Figure 7.3. Figure 7.2: Interface of Protege 7.3 Ontology-Based Semantic Similarity Sánchez et al. [41] summarized ontology-based similarity assessment into three kinds and gave both advantages and disadvantages of each approach. The three kinds 44 Figure 7.3: Part of Ontology 45 of categories are: Edge-counting approaches, Feature-based measures, and Measures based on Information Content. 7.3.1 Path-Based Approaches In path-based approaches, the ontology is viewed as a directed graph, in which the nodes are the concepts, and the edges are taxonomic relation (e.g. is-a). Rada, et al. [34] measure the similarity by the distance of two nodes in the graph. Therefore, the semantic distance of two concepts a and b will be: disrad(a,b) = min|pathi(a,b)| Wu and Palmer [49] realized that the depth in the taxonomy will impact the similarity measure of two nodes, because the deeper of the nodes are in the tree, the semantic distance is smaller. Therefore they gave a new measure of ontology: simw&p(a,b) = 2 ×N3 N1 + N2 + 2 ×N3 N1 and N2 is the numbers of is-a links from each term to their Least Common Subsumer(LCS), N3 is the number of is-a links of the LCS to the root of the ontology. Based on the same idea, Leacock and Chodorow [23] also proposed a similarity measure that combined distance Np between terms a and b and the depth D of the taxonomy. siml&c(a,b) = − log(Np/2D) There are some limitations of path-based approaches. First, it only considers the shortest path between concept pairs. When they meet a complex situation like multiple taxonomical inheritance, the accuracy of them will be low. Another problem 46 of the path-based approaches is that they assume that all links in the taxonomy have uniform distance. 7.3.2 Feature-Based Measures Feature based approaches assess the similarity between concepts as a function of their properties. They consider the degree of overlapping between sets of ontological features, like Tversky’s model [46], which subtracts the non-common features from common features of two concepts. simtve(a,b) = α ·F (Ψ(a) ∩ Ψ(b)) −β ·F (Ψ(a) \ Ψ(b)) −γ ·F (Ψ(b) \ Ψ(a)) Where F is salience of a set features, and α,β and γ are weights of the contribution of each component. Rodŕıguez and Egenhofer [40] computed similarity by summing the weighted sum of similarities between synsets, features, and neighbour concepts. simre(a,b) = w ·Ssynsets(a,b) + u ·Sfeatures(a,b) + v ·Sneighborhoods(a,b) The feature-based methods consider more semantic knowledge than path-based methods. But only big ontologies/thesauri like Wordnet [31] have this kind of in- formation. Ding et al. [11] revealed that domain ontologies very occasionally model any semantic feature apart from taxonomical relationship. 7.3.3 Content-Based Measures Other approaches want to overcome the limitations of edge-counting approach- es are Content-based measures. Resnik [36] proposed a similarity measure, which 47 depends on the amount of shared information between two terms: simres(a,b) = IC(LCS(a,b)) LCS is the Least Common Subsumer of terms in a ontology, and IC is Information Content, which is the negative log of its probability of occurrence, p(a). Lin [26] and Jiang and Conrath [20] extended Resnik’s work. They also considered the IC of each of the evaluated terms, and they proposed that the similarity between two terms should be measured as the ratio between the amount of information needed to state their commonality and the information needed to fully describe them. simlin(a,b) = 2 ×simres(a,b) (IC(a) + IC(b)) The are also two disadvantages of the content-based measures. First, the approaches cannot compute the concepts of leave nodes, because they don’t have subsumers. Second, if the concepts do not have enough common subsumers, their similarities are hard to be calculated. 7.4 Statistical-Based Ontology Similarity Measure In this thesis, we proposed a new statistical-based ontology similarity measure. In most job descriptions, they list many skills the positions required. From observation, we found that related skills always exist in the job description simultaneously, and the positions of them are always close, e.g. HTML and CSS are always required together, and appear in the same sentence. We could see this phenomenon in Table 7.4, which include some skill requirement sentences from some job descriptions. We can see from the Table 7.4, the closely related concepts are always have short distance. Based on such observation, we give a new statistical-based ontology 48 Table 7.4: Some sentences of Job Descriptions 1. A high-level language such as Java, Groovy, Ruby or Python; we use Java and Groovy extensively 2. HTML5/CSS3/JavaScript, web standards, jQuery or frameworks like An- gularJS would be great 3. HTML CSS and Javascript a must 4. Experience with AJAX, XML, XSL, XSLT, CSS, JavaScript, JQuery, HTM- L and Web Services similarity measure. If two concepts a and b have the same direct hypernym or one is the hypernym of the other, the similarity between them is given: S(a,b) = Na∩b/Na∪b avg(log2(mindis(di,a,b) + 1)) The numerator is the ratio of the number of documents in which the two terms exist together (Na∩b) and the number of documents have a least one of them (Na∪b). The denominator is the average log value of minimum distance mindis(doc,a,b) of the two terms in documents that have them both. We set the restriction on the position of the two concepts in the ontology, because the position of the concepts in the ontology are based on their technical similarity to others. Similar techniques will be assigned into the same category, so they should share the same hypernym, and one could be an alternative to the other. For example, we put EJB and Hibernate in the same category, because they are both J2EE persis- tence layer technologies, and both have the O/R mapping concept. If the applicant is familiar one of them, they can master the other very quickly. Another example is Grail and Django, they are both web frameworks and share same web design philoso- phies, but one of them is designed for Java web application and the other is created for Python web application. If a developer has some some experience with one of 49 them, he/she still need to spend a lot of time to learn the other to overcome the gap between programming languages. The algorithm to calculate the similarity of two concepts is in Algorithm 2. Input: Docs term1, term2 Output: similarity total = 0; hastwo = 0; dislist = [ ]; for i = 1; i ≤ len(Docs); i + + do if Docsi has at least one term then total + = 1 ; if Docsi has both terms then hastwo + = 1 ; mindis = minimium distance (Docsi, term1, term2) ; dislist. add (log2(mindis + 1)) ; end end end factor1 = hastwo / total ; factor2 = avg(dislist) ; return factor1 / factor2; Algorithm 2: Calculating Statistical-based Similarity The matrix in Table 7.5 show the similarity values among of some skills, which is gotten from 500 job descriptions. For example the skill HTML, the most relevant skills in order are CSS, Javascript, and jQuery, which is the same from the perspective of experienced developers. The other example is Java, the most relevant skill in the matrix is JSP, which is also agree with the general technical knowledge. 50 Table 7.5: Similarities of Skills List 1 Term Java JDBC Spring Hibernate MySql Oracle Java 1 0.0523 0.091 0.0458 0.0339 0.0608 JDBC 0.0523 1 0.0525 0.0799 0.006 0.0616 Spring 0.091 0.0525 1 0.2008 0.0194 0.0878 Hibernate 0.0458 0.0799 0.2008 1 0.0073 0.115 MySql 0.0339 0.006 0.0194 0.0073 1 0.049 Oracle 0.0608 0.0616 0.0878 0.115 0.049 1 Table 7.6: Similarities of Skills List 2 Term Javascript jQuery HTML CSS Java Python Ruby JSP Javascript 1 0.1981 0.2087 0.2439 0.0665 0.0189 0.023 0.0253 jQuery 0.1981 1 0.0979 0.1328 0.0439 0.0142 0.0266 0.0232 HTML 0.2087 0.0979 1 0.3569 0.0473 0.0175 0.023 0.0103 CSS 0.2439 0.1328 0.3569 1 0.0537 0.0153 0.0181 0.015 Java 0.0665 0.0439 0.0473 0.0537 1 0.0498 0.0287 0.075 Python 0.0189 0.0142 0.0175 0.0153 0.0498 1 0.1333 0.0025 Ruby 0.023 0.0266 0.023 0.0181 0.0287 0.1333 1 0.012 JSP 0.0253 0.0232 0.0103 0.015 0.075 0.0025 0.012 1 51 8. EVALUATION In this section we will evaluate the performance of the key components of the system, the Information Extraction module, the Ontology Similarity module, and the overall system resulting from the combination of the components. 8.1 Experiments of Information Extraction To evaluate the performance of the information extraction module, we extract sentence types through the use of sentence filters. To explain the process of our experiment, we use the sentences whose content pertains to the applicant’s college degree information. In the experiment, we selected 100 sentences from existing job descriptions, and the content of these sentences were requirements of candidate degree and major. We labeled the values for ”degree” and ”major” manually. We use some content patterns that we can identify from these sentences to match and extract the degree information. When we used 6 patterns, the accuracy of ”degree” became 94%. We also compared our pattern matching method to the conditional random fields (CRFs) model [22], which is a state of art machine learning model for sequence labeling. We used 200 labelled sentences to train the CRF model, and the features of the CRF model are words in the sentences and part of speech tags of the words. The accuracies of information extraction of the three fields with our two methods, pattern matching, and the application of the CRF model are shown in Table 8.1. 8.2 Experiments of Ontology Similarity We selected some common skills from 500 job descriptions; table 8.2 shows simi- larity values between these skills. Higher values correspond to greater similarities, so 52 Table 8.1: Information Extraction Field Pattern Num Accuracy of Pattern Matching Accuracy of CRF Degree 6 0.94 0.85 Major 10 0.85 0.72 Our pattern matching approach can get higher accuracy. Table 8.2: Similarities of Skill List 1 Term Java JDBC Spring Hibernate MySql Oracle Java 1 0.0523 0.091 0.0458 0.0339 0.0608 JDBC 0.0523 1 0.0525 0.0799 0.006 0.0616 Spring 0.091 0.0525 1 0.2008 0.0194 0.0878 Hibernate 0.0458 0.0799 0.2008 1 0.0073 0.115 MySql 0.0339 0.006 0.0194 0.0073 1 0.049 Oracle 0.0608 0.0616 0.0878 0.115 0.049 1 the similarity between one skill and itself is 1. We selected one concept and ranked the other concepts by their similarity values to this concept. Human judges helped rank these concepts by assigning them ”relevance scores” so that we can use NDCG to evaluate the effectiveness of our approach. We use the Normalized Discounted Cumulative Gain(NDCG) [30] to evaluate the statistical-based similarity. NDCG is an important measure to evaluate the ranked retrieval results, which is the ratio of Discounted Cumulative Gain ( DCG ) to Ideal Discounted Cumulative Gain ( IDCG ). NDCG = DCG IDCG DCG is the measure of how documents are ranked according to their relevance scores, and IDCG is the DCG value that the documents are strictly sorted by their 53 Table 8.3: Javascript Similarity Evaluation Term Similarity Value Position Relevance jQuery 0.1981 4 8 HTML 0.2087 3 4 CSS 0.2439 2 3 Java 0.0665 5 1 Python 0.0189 8 1 Ruby 0.023 7 1 JSP 0.0253 6 2 Table 8.4: HTML Similarity Evaluation Term Similarity Value Position Relevance Javascript 0.2087 2 3 jQuery 0.0979 3 3 CSS 0.3569 1 5 Java 0.0473 4 1 Python 0.0175 6 1 Ruby 0.023 5 1 JSP 0.0103 7 3 relevance values. DCG = p∑ i=1 2reli − 1 log2(i + 1) Table 8.3 shows how we evaluate the similarity between the concept “Javascript” and other concepts. The first column is a skill name, the second column is its simi- larity value to “Javascript”, the third column is its position ranked by the similarity value, and the fourth column is its relevance value given by the judges. The ND- CG value for concept Javascript is 0.94, and in Table 8.4, NDCG value for concept HTML is 0.97. 54 8.3 Evaluation of the System In a traditional information retrieval systems, precision and NDCG are widely used measures [30]. Precision (P) is the fraction of retrieved documents that are relevant. Precision = #(releveant items retrieved) #(retrieved items) We first used Precision@K to compare the performance of our approach to the classical information retrieval models: Okapi BM25 [37], Kullback-Leibler divergence, and the TF-IDF. Precision@K is the proportion of relevant documents in the first K positions and is given below: P@k = 1 k m∑ i=1 li1 (r(i) ≤ k) Where 1 is the indicator function: 1(A) = 1 if A is true, 0 otherwise. To evaluate a job finding, we compare the results of the system with three classi- cal information retrieval models: Kullback-Leibler divergence [52], TF-IDF [30] and Okapi BM25 [38]. We give the definition of these measures below. Kullback-Leibler divergence is a non-symmetric measure of the difference between two probability distributions P and Q. The score of a document D with respect to query Q is given by: s(D,Q) = −D(θQ ‖ θD) = − ∑ ω∈V p(ω | θQ) log p(ω|θQ) p(ω|θD) = ∑ ω∈V p(ω | θQ) log p(ω | θD) − ∑ ω∈V p(ω | θQ) log p(ω | θQ) (8.1) In the equation θQ is the language model for a query, and θD is the language model for a document. 55 TF-IDF is a Vector Space Model, which calculates the Cosine Similarity between the vectors of the query q and the document d. In the equation, tf is the term frequency, and idf is the inverse document frequency. The TF-IDF weighting scheme assigned to term t a weight in document d given by: tf-idft,d = tft,d × idft The similarity between the query q and the document d given by: score(q,d) = ∑ t∈q tf-idft,d Okapi BM25 is a bag-of-words retrieval model that ranks a set of documents based on the query terms appearing in each document. Given a query q, the BM25 score of a document d is: score(q,d) = ∑ t∈q idft tft,d · (k1 + 1) tft,d + k1 · (1 − b + b · |D| avgdl ) In the formula, idft is IDF value of term t, tf is the term frequency; |D| is the length of the document D; avgdl is the average length of all the documents, and k1 and b are the free parameters. In the evaluation phase, we created a data set of 100 job descriptions that includes several kinds of jobs such as web developers, server back-end developers, mobile developers and so on. We used 5 candidate résumés and retrieved the top 20 jobs. The relevance value of the job descriptions to each résumé were set manually by ten human judge. We created a query q from the résumé, treated the text of the job descriptions as documents d, and applied standard ad-hoc retrieval techniques to rank the jobs. We intended to return jobs that better matched the candidates’ 56 résumés at the top. The result of Precision@k is in table 8.5. Table 8.5: Precision of Job Ranking k Okapi BM25 KL TF-IDF RésuMatcher 5 0.66 0.27 0.72 0.82 10 0.46 0.27 0.50 0.76 20 0.33 0.21 0.35 0.77 RésuMatcher get the highest precision. The other measure we used is NDCG, which is explained in last section. Table 8.6 shows the NDCG value get from different information retrieval models. The result shows that Ontology Similarity performs the best. Table 8.6: NDCG of Job Ranking k Okapi BM25 KL TF-IDF RésuMatcher 5 0.15 0.34 0.45 0.78 10 0.18 0.44 0.47 0.72 20 0.19 0.35 0.45 0.66 Our sytem get the highest NDCG value. 8.4 Comparing with the Keyword Searching We also made experiments to compare the quality of search results quality be- tween our system and the keyword searching approach. In one experiment, we first selected a résumé, and picked a related keyword e.g. “Java” to search jobs in In- deed.com. We used this process to simulate how a job seeker searches jobs on In- deed.com. The top 100 jobs returned by Indeed.com were saved in our system, then we used the résumé to match the 100 jobs. We compared the quality of the 57 approaches by calculating the measures Precision@K and DCG. We use DCG not NDCG because we compare the two approaches on the same data set, and the IDCG is the same. Five judges gave the relevance values to the résumés and jobs, and we used average values. Table 8.7 shows comparison between the two approaches by using keyword “Java” and the résumé of a Java developer. The comparison for keyword “Python” and the résumé of a Python developer is shown in Table 8.8, and the average comparison for five keywords and five résumé is shown in Table 8.9. We can see from the result that our system can get much better results than keyword searching approach. Table 8.7: Comparison of the Two Approaches - Java Developer k Precision@K DCG Indeed RésuMatcher Indeed RésuMatcher 5 0.73 1.00 10.86 27.80 10 0.65 0.95 24.85 47.06 20 0.72 0.83 51.97 73.33 RésuMatcher can get better result. Table 8.8: Comparison of the Two Approaches - Python Developer k Precision@K DCG Indeed RésuMatcher Indeed RésuMatcher 5 0.45 0.63 11.27 11.98 10 0.32 0.62 13.43 15.79 20 0.22 0.42 16.44 20.21 RésuMatcher can get better result. 58 Table 8.9: Comparison of the Two Approaches - Average k Precision@K DCG Indeed RésuMatcher Indeed RésuMatcher 5 0.84 0.87 23.87 32.97 10 0.72 0.86 37.02 45.57 20 0.65 0.76 58.70 66.70 RésuMatcher can get better result. 59 9. CONCLUSION AND FUTURE WORK 9.1 Conclusion In this thesis we presented RésuMatcher, a personalized job-résumé matching system that can help job seekers find appropriate jobs faster and more accurately by using their résumé contents. The key components of the system are the information extraction procedure and the ontology matching module. In the system, job descriptions and résumés are parsed into job models and résumé models by the information extraction module. When searching the jobs by a résumé, similarity values between the résumé model and job description models are calculated in the ontology matching module. The result is sorted by the ontology similarity scores, which are the sum of similarities of different fields multiplied by their weights. We made such contributions in our works: 1. We developed a résumé - job matching system. 2. We developed a finite state transducer based matching tool to extract informa- tion from unstructured data source, which is a lightweight and flexible library, and can be extended in very easy ways. 3. We developed a semi-automatic approach, which can collect technical terms from hr data sources, and by which we created a domain specific ontology for recruitment. 4. We developed statistical-based ontology similarity measure, which can measure the similarities between technical terms . In the experiment phase, we evaluated the accuracy of information extraction. We calculated the ontology similarity with the NDCG. Finally, we also tested the 60 performance of résumé job matching algorithm via precision@k and NDCG, which showed that our algorithm can achieve a better searching result than other informa- tion retrieval models like TF-IDF and OKpai BM25. We also compared our system with the commercial job search engine www.indeed.com, and the results showed that our system can return jobs with a better ranking. 9.2 Future Work Finding a job is a complex process, affected by both explicit and implicit factors. Our work establishes the validity of using information extraction techniques to create a more personalized job matching system, with ample potential for improvement in the future. First we can introduce a more complex job and rsum model to improve perfor- mance of the system. In the résumé model, we can consider hiring history and project experience of the job seekers. To improve the job description model, job responsi- bilities and company characteristics (size, dress code, etc.) should be considered as well. Second, to improve searching speed of our system, we can reduce the the number of comparison by filtering out jobs that are clearly not related to résumés. The system can classify the jobs into some different subsets, when searching jobs, the system only need to calculate the similarity between the résumé and according subset of jobs. RésuMatcher is a content based recommendation system that is mostly focused on comparing the similarities between the résumé and a relevant job description. In future work, we could introduce a hybrid recommendation system that would take advantage of other recommendation algorithms such as Collaborative Filtering. Future work on this system would place greater consideration on job seeker’s personal preference like job location, career development plan, and company background. 61 REFERENCES [1] ACM. Acm computing classification system, 2012. [2] Alfred V Aho and Jeffrey D Ullman. Foundations of computer science, volume 2. Computer Science Press New York, 1992. [3] Shaha T Al-Otaibi and Mourad Ykhlef. A survey of job recommender systems. International Journal of the Physical Sciences, 7(29):5127–5142, 2012. [4] Douglas E Appelt and Boyan Onyshkevych. The common pattern specification language. In Proceedings of a workshop on held at Baltimore, Maryland: October 13-15, 1998, pages 23–30. Association for Computational Linguistics, 1998. [5] Steven Bird. Nltk: The natural language toolkit. In Proceedings of the COL- ING/ACL on Interactive presentation sessions, pages 69–72. Association for Computational Linguistics, 2006. [6] Christian Bizer, Jens Lehmann, Georgi Kobilarov, Sören Auer, Christian Becker, Richard Cyganiak, and Sebastian Hellmann. Dbpedia-a crystallization point for the web of data. Web Semantics: Science, services and agents on the world wide web, 7(3):154–165, 2009. [7] Duygu Çelik and Atilla Elçi. An ontology-based information extraction approach for résumés. In Pervasive Computing and the Networked World, pages 165–179. Springer, 2013. [8] Angel X Chang and Christopher D Manning. Tokensregex: Defining cascaded regular expressions over tokens. Technical report, Technical Report CSTR 2014- 02, Department of Computer Science, Stanford University, 2014. 62 [9] Hamish Cunningham, Diana Maynard, Kalina Bontcheva, and Valentin Tablan. A framework and graphical development environment for robust nlp tools and applications. In ACL, pages 168–175, 2002. [10] J Olawande Daramola, Olufunke O Oladipupo, and AG Musa. A fuzzy expert system (fes) tool for online personnel recruitments. International Journal of Business Information Systems, 6(4):444–462, 2010. [11] Li Ding, Tim Finin, Anupam Joshi, Rong Pan, R Scott Cost, Yun Peng, Pavan Reddivari, Vishal Doshi, and Joel Sachs. Swoogle: A search and metadata engine for the semantic web. In Proceedings of the thirteenth ACM international conference on Information and knowledge management, pages 652–659. ACM, 2004. [12] Ying Ding and Schubert Foo. Ontology research and development. part 1-a review of ontology generation. Journal of Information Science, 28(2):123–136, 2002. [13] Athanasios Drigas, Stelios Kouremenos, Spyros Vrettos, John Vrettaros, and Dimitris Kouremenos. An expert system for job matching of the unemployed. Expert Systems with Applications, 26(2):217–224, 2004. [14] Frank Färber, Tim Weitzel, and Tobias Keim. An automated recommendation approach to selection in personnel recruitment. 2003. [15] Maryam Fazel-Zarandi and Mark S Fox. Semantic matchmaking for job recruit- ment: An ontology-based hybrid approach. In Proceedings of the 8th Interna- tional Semantic Web Conference, 2009. [16] Christiane Fellbaum. WordNet. Wiley Online Library, 1998. 63 [17] Tere Gonzalez, Pano Santos, Fernando Orozco, Mildreth Alcaraz, Victor Zal- divar, Alberto De Obeso, and Alan Garcia. Adaptive employee profile classifica- tion for resource planning tool. In SRII Global Conference (SRII), 2012 Annual, pages 544–553. IEEE, 2012. [18] Hal G Gueutal and Dianna L Stone. The brave new world of eHR. John Wiley & Sons, 2006. [19] Jerry R Hobbs, Douglas Appelt, John Bear, David Israel, Megumi Kameyama, Mark Stickel, and Mabry Tyson. 13 fastus: A cascaded finite-state transducer for extracting information from natural-language text. Finite-State Language Processing, page 383, 1997. [20] Jay J Jiang and David W Conrath. Semantic similarity based on corpus statistics and lexical taxonomy. ArXiv Preprint Cmp-lg/9709008, 1997. [21] V Senthil Kumaran and A Sankar. Towards an automated system for intelli- gent screening of candidates for recruitment using ontology mapping (expert). International Journal of Metadata, Semantics and Ontologies, 8(1):56–64, 2013. [22] John Lafferty, Andrew McCallum, and Fernando CN Pereira. Conditional ran- dom fields: Probabilistic models for segmenting and labeling sequence data. 2001. [23] Claudia Leacock and Martin Chodorow. Combining local context and word- net similarity for word sense identification. WordNet: An Electronic Lexical Database, 49(2):265–283, 1998. [24] Danielle H Lee and Peter Brusilovsky. Fighting information overflow with per- sonalized comprehensive information access: A proactive job recommender. In 64 Autonomic and Autonomous Systems, 2007. ICAS07. Third International Con- ference on, pages 21–21. IEEE, 2007. [25] Wendy Lehnert, Claire Cardie, David Fisher, Ellen Riloff, and Robert Williams. University of massachusetts: Description of the circus system as used for muc-3. In Proceedings of the 3rd conference on Message understanding, pages 223–233. Association for Computational Linguistics, 1991. [26] Dekang Lin. An information-theoretic definition of similarity. In ICML, vol- ume 98, pages 296–304, 1998. [27] Ping Liu and Peter Dew. Using semantic web technologies to improve expertise matching within academia. Proceedings of I-Know (Graz, Austria, pages 370– 378, 2004. [28] Yao Lu, Sandy El Helou, and Denis Gillet. A recommender system for job seek- ing and recruiting website. In Proceedings of the 22nd international conference on World Wide Web companion, pages 963–966. International World Wide Web Conferences Steering Committee, 2013. [29] Jochen Malinowski, Tobias Keim, Oliver Wendt, and Tim Weitzel. Matching people and jobs: A bilateral recommendation approach. In System Sciences, 2006. HICSS’06. Proceedings of the 39th Annual Hawaii International Confer- ence on, volume 6, pages 137c–137c. IEEE, 2006. [30] Christopher D Manning, Prabhakar Raghavan, and Hinrich Schütze. Introduc- tion to information retrieval, volume 1. Cambridge university press Cambridge, 2008. [31] George A Miller. Wordnet: A lexical database for english. Communications of the ACM, 38(11):39–41, 1995. 65 [32] Natalya F Noy, Michael Sintek, Stefan Decker, Monica Crubézy, Ray W Ferg- erson, and Mark A Musen. Creating semantic web contents with protege-2000. IEEE Intelligent Systems, 16(2):60–71, 2001. [33] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pager- ank citation ranking: Bringing order to the web. 1999. [34] Roy Rada, Hafedh Mili, Ellen Bicknell, and Maria Blettner. Development and application of a metric on semantic nets. Systems, Man and Cybernetics, IEEE Transactions on, 19(1):17–30, 1989. [35] Rachael Rafter, Keith Bradley, and Barry Smyth. Personalised retrieval for online recruitment services. In Proceedings of the 22nd Annual Colloquium on Information Retrieval. 2000., 2000. [36] Philip Resnik. Using information content to evaluate semantic similarity in a taxonomy. ArXiv Preprint Cmp-lg/9511007, 1995. [37] Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: Bm25 and beyond. Information Retrieval, 3(4):333–389, 2009. [38] Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock- Beaulieu, Mike Gatford, et al. Okapi at trec-3. Nist Special Publication SP, pages 109–109, 1995. [39] Emmanuel Roche and Yves Schabes. Finite-state language processing. MIT press, 1997. [40] M Andrea Rodŕıguez and Max J. Egenhofer. Determining semantic similarity among entity classes from different ontologies. Knowledge and Data Engineering, IEEE Transactions on, 15(2):442–456, 2003. 66 [41] David Sánchez, Montserrat Batet, David Isern, and Aida Valls. Ontology-based semantic similarity: A new feature-based approach. Expert Systems with Appli- cations, 39(9):7718–7728, 2012. [42] Sunita Sarawagi. Information extraction. Foundations and Trends in Databases, 1(3):261–377, 2008. [43] Amit Singh, Catherine Rose, Karthik Visweswariah, Vijil Chenthamarakshan, and Nandakishore Kambhatla. Prospect: A system for screening candidates for recruitment. In Proceedings of the 19th ACM international conference on Information and knowledge management, pages 659–668. ACM, 2010. [44] Zheng Siting, Hong Wenxing, Zhang Ning, and Yang Fan. Job recommender systems: A survey. In Computer Science & Education (ICCSE), 2012 7th In- ternational Conference on, pages 920–924. IEEE, 2012. [45] Ken Thompson. Programming techniques: Regular expression search algorithm. Communications of the ACM, 11(6):419–422, 1968. [46] Amos Tversky. Features of similarity. 1977. [47] Andrew J Viterbi. Error bounds for convolutional codes and an asymptotical- ly optimum decoding algorithm. Information Theory, IEEE Transactions on, 13(2):260–269, 1967. [48] Kangning Wei, Jinghua Huang, and Shaohong Fu. A survey of e-commerce recommender systems. In Service Systems and Service Management, 2007 In- ternational Conference on, pages 1–5. IEEE, 2007. [49] Zhibiao Wu and Martha Palmer. Verbs semantics and lexical selection. In Proceedings of the 32nd annual meeting on Association for Computational Lin- guistics, pages 133–138. Association for Computational Linguistics, 1994. 67 [50] Xing Yi, James Allan, and W Bruce Croft. Matching resumes and jobs based on relevance models. In Proceedings of the 30th annual international ACM SIGIR conference on Research and development in information retrieval, pages 809– 810. ACM, 2007. [51] Kun Yu, Gang Guan, and Ming Zhou. Resume information extraction with cas- caded hybrid model. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, pages 499–506. Association for Computational Linguistics, 2005. [52] ChengXiang Zhai. Statistical language models for information retrieval. Syn- thesis Lectures on Human Language Technologies, 1(1):1–141, 2008. [53] Justin Zobel and Alistair Moffat. Inverted files for text search engines. ACM Computing Surveys (CSUR), 38(2):6, 2006. 68 
work_xt5lrac64vgajg4ssffkqsiucu	----	A knowledge-based multi-layered image annotation system Expert Systems With Applications 42 (2015) 9539–9553 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa A knowledge-based multi-layered image annotation system Marina Ivasic-Kos a,∗, Ivo Ipsic a, Slobodan Ribaric b a Department of Informatics, University of Rijeka, Rijeka, Croatia b Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia a r t i c l e i n f o Keywords: Image annotation Multi-layered image annotation Knowledge representation Fuzzy Petri Net Fuzzy inference engine a b s t r a c t Major challenge in automatic image annotation is bridging the semantic gap between the computable low- level image features and the human-like interpretation of images. The interpretation includes concepts on different levels of abstraction that cannot be simply mapped to features but require additional reasoning with general and domain-specific knowledge. The problem is even more complex since knowledge in con- text of image interpretation is often incomplete, imprecise, uncertain and ambiguous in nature. Thus, in this paper we propose a fuzzy-knowledge based intelligent system for image annotation, which is able to deal with uncertain and ambiguous knowledge and can annotate images with concepts on different levels of ab- straction that is more human-like. The main contributions are associated with an original approach of using a fuzzy knowledge-representation scheme based on the Fuzzy Petri Net (KRFPN) formalism. The acquisition of knowledge is facilitated in a way that besides the general knowledge provided by the expert, the computable facts and rules about the concepts, as well as their reliability, are produced automatically from data. The rea- soning capability of the fuzzy inference engine of the KRFPN is used in a novel way for inconsistency checking of the classified image segments, automatic scene recognition, and the inference of generalized and derived classes. The results of image interpretation of Corel images belonging to the domain of outdoor scenes achieved by the proposed system outperform the published results obtained on the same image base in terms of average pre- cision and recall. Owing to the fuzzy-knowledge representation scheme, the obtained image interpretation is enriched with new, more general and abstract concepts that are close to concepts people use to interpret these images. © 2015 Elsevier Ltd. All rights reserved. 1 p d i m v p a c w o w t s a a t n F i c c f i J t t f t h 0 . Introduction Digital images have become unavoidable in the professional and rivate lives of modern people. In recent years, the frequent use of igital images has become necessary in different fields like medicine, nsurance and security systems, geo-informatics, advertising, com- erce, as well as in other business areas. On the other hand, in pri- ate life, digital images are used for documenting people close to us, ets, sights and events such as birthdays, parties, trips, excursions nd sporting activities. This widespread use has caused a rapid in- rease in the number of digital images that, today, on specialized ebsites, can be counted in the millions. However, a large number f images leads to problems with searching and retrieval, as well as ith organizing and storing. As the majority of images are barely documented, it is believed hat we could retrieve and arrange images simply if they were ∗ Corresponding author. Tel.: +38551584710. E-mail addresses: marinai@uniri.hr (M. Ivasic-Kos), ivoi@uniri.hr (I. Ipsic), lobodan@zemris.fer.hr (S. Ribaric). h s n i l ttp://dx.doi.org/10.1016/j.eswa.2015.07.068 957-4174/© 2015 Elsevier Ltd. All rights reserved. utomatically annotated and described with words that are used in n intuitive image search. However, the task of mapping image fea- ures that can be extracted from raw image data to words that users ormally use for articulating their requirements is not a trivial one. or example, it seems natural to use a destination name when retriev- ng holiday images or some terms that describe a scene, such as the oast, mountains or activities like diving, skiing, etc. A major research hallenge is bridging the semantic gap between the low-level image eatures available to a computer and the interpretation of the images n the way that humans do (Smeulders, Worring, Santini, Gupta, & ain, 2000). In addition, one should take into account that image in- erpretation inherent to humans includes concepts associated with he content of the image on different levels of abstraction. This is re- erred to as the multi-layered interpretation of image content. To sys- ematically describe visual content of an image and its semantics, we ave defined a knowledge-based image representation model con- isting of multiple layers of image representation. Layers are orga- ized according to the amount of knowledge needed to automatically nterpret the image using inference about concepts belonging to the ayer. http://dx.doi.org/10.1016/j.eswa.2015.07.068 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.07.068&domain=pdf mailto:marinai@uniri.hr mailto:ivoi@uniri.hr mailto:slobodan@zemris.fer.hr http://dx.doi.org/10.1016/j.eswa.2015.07.068 9540 M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 a & i t a o i t a e L b t s a i t T i b l t u t k l p b l B g t t L b a a 2 i F l s o T ( t b f i & v i l o c o A s a t a According to the defined image representation model, an intel- ligent system for multi-layered image annotation is proposed. The first layer of the image interpretation contains concepts obtained by the classification of image segments using conventional supervised classification method. Higher levels of image interpretation involve concepts that are more abstract. These concepts are difficult to in- fer directly based on low-level features and without knowledge rel- evant to the problem domain. Therefore, we have defined the fuzzy knowledge-representation schemes based on fuzzy Petri net (KRFPN) formalism to represent knowledge about concepts that can appear in an image. Fuzzy Petri nets combine fuzzy set theory and Petri net the- ory to provide the representation of knowledge, which is in context of image interpretation often incomplete, imprecise, uncertain and ambiguous in nature. The KRFPN formalism is originally supported with a fuzzy infer- ence engine that deals with approximate reasoning. The reasoning capability of the inference engine was used in an original way to draw conclusions about classes of image scenes and more abstract classes. The system can handle the ambiguity and uncertainty about concepts and relations, so decisions about more abstract concepts can be made even when input information about the concepts present in an image are imprecise and vague. To reduce the propagation of errors through the hierarchical structure of concepts and to increase the reliability of conclusions, as well as to improve the precision of image annotation, a consistency-checking procedure is proposed. The acquisition of knowledge used by inference engine is facil- itated in a way that all the facts and rules of the composition and distribution of concepts as well as their reliability are produced auto- matically from data. Both new relationships and new concepts with appropriate measure of reliability are stored into the knowledge base and used by the inference engine. The paper is organized as follows: First, in Section 2, different approaches to image-content interpretation are explained and a de- tailed overview of related work is given. The layers of the multi- layered image representation with respect to the amount of knowl- edge needed for the image interpretation are given in Section 3. A system for the multi-layered image annotation is proposed in Section 4. A fuzzy-knowledge representation scheme adapted for the outdoor image domain is presented in Section 5. Inputs to the scheme are concepts obtained as the results of an image-segments classifi- cation using a Bayesian classifier. The application of the fuzzy infer- ence engine for checking the consistency of the obtained results of the image segment classification and the recognition of scene con- text is given in Sections 6 and 7, respectively. The fuzzy inference al- gorithm used to derive more abstract concepts associated with the image is described in Section 8. The experimental results of the im- age interpretation at the layer that corresponds to automatic image annotation are given and compared to previously reported methods in Section 9. Additionally, in Section 9, an improvement to the results of the automatic image annotation after checking the inconsistency of the concepts obtained during the image-segments classification is presented and discussed. 2. Related work Image interpretation is a complex task that strongly depends on purpose of annotation. Moreover, human interpretation is limited by the knowledge, culture, experience and point of view of the person. Therefore, in the development of the automatic image annotation system, types of concepts that would be used for image interpretation should be decided first, depending on the purpose of the annotation. Among the oldest models for image annotation is Shatford’s image-content classification of general-purpose images drawing on theory from art history that classifies image content into general, specific and abstract concepts (Shatford, 1986). Additionally, the con- tents of an image are associated with aspects of objects, with spatial nd temporal aspects and aspects of activities or events. In (Eakins Graham, 2000), a multilayer interpretation of the image content s considered in the context of image search. The authors defined hree semantic layers of image interpretation. At the first level, im- ge interpretation is based on the presence of certain combinations f features, such as color, texture or shape, while at the second level, mage interpretation deals with the presence and distribution of cer- ain types of objects. At the third level, image interpretation includes description of specific types of events or activities, locations and motions that one can associate with the image. The authors (Hare, ewis, Enser, & Sandom, 2006) provide a simplified hierarchical view etween the two extremes, the image itself and its full semantic in- erpretation. At the lowest level are the image and its “raw” data. The econd level consists of low-level features related to a part of an im- ge or to the whole image. A combination of prototype feature vectors s part of the third level. If these image parts can be associated with he corresponding objects, then this would make the fourth level. he top level of image interpretation, referred to as full semantics, ncludes concepts that describe the events, actions, emotions and a roader context of the image. This model, particularly in layers re- ated to visual image content, mostly influenced the image represen- ation model that we propose. The main difference is in higher layers sed to model the image semantics. There are two major approaches widely used for image anno- ation, one using statistical methods and the other mostly using nowledge-based methods belonging to the field of artificial intel- igence. Both approaches are used in our systems: the statistical ap- roach in the first layer of the image interpretation and knowledge- ased approach in the higher layers. In the statistical approach, most methods can be grouped as trans- ation or classification models. In the translation model of (Duygulu, arnard, de Freitas, & Forsyth, 2002) the co-occurrence of image re- ions and annotation words are used to model the relationship be- ween annotation words and images or image regions. In classifica- ion methods, such as (Barnard et al., 2003, Li & Wang, 2003, Hu and am, 2013), words used for image annotation correspond to class la- els for which classifiers are trained. Due to the intra-class variability nd inter-class similarity, usually class labels correspond to objects in n image, but can correspond to scenes as well. In (Fei-Fei & Perona, 005) natural scenes were learned by a Bayesian hierarchical model n unsupervised way from local image regions. In (Yin, Jiao, Chai, & ang, 2015) discriminant scene features were learned using single- ayer sparse autoencoder (SAE) and then SVM classifier is used for cene classification. Some methods use multi-label learning for solving the problem f annotating images with more than one word (Feng & Xu, 2010). o improve the accuracy of multi-label classification algorithm, in Yu, Pedrycz, & Miao, 2014) correlation among the labels and uncer- ainty of classification between feature space and label space have een considered and in (Hong et al., 2014) selection of discriminative eatures has been proposed. Lately, deep neural networks are exam- ned for the task of multi-label image annotation. In (Chengjian, Zhu, Shi, 2015) multimodal deep neural network pre-trained with con- olutional neural networks is proposed. Such statistical methods commonly use quite simple vocabular- es that can be large but are generally not structured because no re- ations are defined between the concepts in the vocabulary. On the ther hand, methods that rely on knowledge bases used sophisti- ated, structured vocabularies in which geometrical, hierarchical or ther relations between concepts are established (Tousch, Herbin, & udibert, 2012). We have defined a vocabulary of this kind that is uitable for image retrieval to be used in our system. A few approaches have explored the dependence of words on im- ge regions (Blei and Jordan, 2003) or exploit the ontological rela- ionships between annotation words, demonstrating their effect on utomatic image annotation and retrieval (Maillot, 2005). M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 9541 Objects sand, sea, sky plane, sky, trees, building snow, polar bear Scenes Coast Scene Plane Scene Polar bear More general (abstract) concepts Natural scene, Outdoor Vehicle Man-made object, Outdoor Wildlife, Mammal, Outdoor, Natural scene Derived concepts Beach, SeaShore, Tallinn, Estonia Meeting Transportation Arctic Fig. 1. Examples of images and their annotation at different levels of abstraction. a & m T n v t t p f t e c e k c p i T s a l v f o a i ( r i l e 2 b 2 m k m i m t t f t m c t a i t k s a g w n 3 a t a d a p h b t i j r s g d n p F t c t T w T i l a A comprehensive survey of research made in the field of statistical utomatic image annotation methods can be found in (Liu, Zhang, Lu, Ma, 2007; Datta, Joshi, & Li, 2008; Zhang, Islam, & Lu, 2012). For a multi-layered image annotation, several approaches that use odels for knowledge representation and reasoning were proposed. he authors (Benitez, Smith, & Chang, 2000) described a semantic etwork to represent the semantics of multimedia content (images, ideo, audio, graphics and text). The basic components of the seman- ic network are concepts that correspond to real-world objects and he relations among them, such as generalization, aggregation and erceptual relationships based on the similarities of their low-level eatures. The authors (Marques & Barman, 2003) propose the model with hree levels. The lowest level contains vectors of low-level features xtracted from images. The feature vectors are classified into the con- epts from flat vocabulary using Bayesian networks. On the high- st level is the RDF ontology that contains knowledge about the eywords and information about the relations between different oncepts. The authors (Srikanth, Varner, Bowden, & Moldovan, 2005) pro- osed using a hierarchical dependency between annotation words to mprove translation-based automatic image annotation and retrieval. he hierarchy is derived from the text ontology WordNet and repre- ents the various levels of generality of the concepts expressed in im- ge regions and words. To predict the likelihood of assigning a class abel given an image, statistical language models defined on a visual ocabulary of blobs, represented by region feature vectors, are used. In (Ivasic-Kos, Ribarić, & Ipsic, 2010) an image content analysis ramework based on Fuzzy Petri Net is proposed for classification f image segments into objects. Also, a formal description of hier- rchical and spatial relationships among concepts from the outdoor mage domain is described. Fuzzy formalism was also applied in Nezamabadi-pour & Kabir, 2009) where fuzzy k-NN classifier with elevance feedback was used to assign semantic labels to database mages. In (Athanasiadis et al., 2009, Simou, Athanasiadis, Stoilos, & Kol- ias, 2008) an ontology and the inference engine FIRE (Fuzzy Infer- nce Reasoning Engine) (Stoilos, Stamou, Tzouvaras, Pan, & Horrocks, 005) were used for analyzing the image content belonging to the each domain. Later, the same group of authors (Papadopoulos et al., 011) compared different approaches attempting to use spatial infor- ation for semantic image analysis. Unlike the above described approaches, we propose a model of a nowledge-based multi-layered image annotation system. We have erged the statistical approach for classification of image segments nto objects and knowledge-based approach to infer concepts that are ore abstract. We took advantages of statistical methods to facilitate he knowledge acquisition, so that computable facts and rules about he concepts as well as their reliability are automatically generated rom data. The key components of the proposed multi-layered image anno- ation system are the KRFPN scheme based on Fuzzy Petri Net for- alism and integrated fuzzy inference engine. We have exploited the apability of the KRFPN inference engine for reasoning with uncer- ainty to infer scenes and concepts that cannot be mapped to im- ges without using the domain knowledge. In addition, to refine the mage annotation and to reduce the propagation of errors through he hierarchical structure of concepts we have included the novel, nowledge-based, consistency-checking procedure into the proposed ystem. For image annotation refinement, the correlation between nnotated keywords has been used in previous research, and lately raph-based algorithms for image analysis have been investigated as ell. A comprehensive survey on image annotation refinement tech- iques is given in (Dong, 2014). . Multi-layered image representation An image representation includes the visual content and the nnotation of an image. The visual content of an image refers to he information that may be collected by analyzing low-level im- ge features while the image annotation includes concepts that may escribe both the content and the context of an image. The task of utomatic image annotation is challenging because the number of ossible concepts that one can use to describe most images is large, ighly dependent on application, user’s knowledge, needs, cultural ackground, etc. and it is hard to choose the right type of concepts hat would be universally appropriate. For instance, to annotate the mages in the Fig. 1, one can use concepts that are related to the ob- ects that appear in the image (sand, sea, sky, snow), concepts that rep- esent the scene (beach, coast, coastline, shore, seashore), more general cene concepts (wildlife, outdoor, natural scene) or activities (walking, et wet feet). If the user is familiar with the context of an image, its escription will be more subjective and will probably include the ame of a place (e.g. Tallinn, Estonia for Fig. 1a), names of the peo- le appearing in it, description of the relevant event (e.g. Meeting for ig. 1a) or evoked emotions, etc. Although different people will most likely use different concepts o annotate the same image, used concepts can be organized ac- ording to the amount of knowledge needed to reach each abstrac- ion level of image interpretation (Ivasic-Kos, Pavlic, & Pobar, 2009). herefore, we propose a multi-layered image representation model in hich layers correspond to concepts at different levels of abstraction. he layers reflect the increase of the amount of knowledge included n the automatic image annotation (Fig. 2) from the lower to higher ayers, where the lower layers (V1 - V2) represent the visual content, nd the layers MI − MI represent the image semantics. 1 4 9542 M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 Fig. 2. Layers of image representation in relation to the knowledge level. t b m l c t r A i m v 1 i m a i v c t P t d s w o t m s f a t r m p r f r i The initial layer of an image representation is the layer V0, and it represents the raw image. The image is usually segmented (layer V1) for analysis, and the low-level features are extracted from the image segments (layer V2). The amount of knowledge required for segmen- tation (layer V1) and feature extraction (layer V2) is low. It is assumed that a multi-layered image annotation includes concepts ranging from elementary classes EC (layer MI1) in which image segments are classified, scene classes SC (layer MI2) that describe the scene, ending with generalized classes GC (layer MI3) and derived classes DC (layer MI4). For instance, the proposed multi-layered image annotation re- lated to Fig. 1c is EC = {snow, polar bear}; SC = {Scene-Polarbear}; GC = {Wildlife, Mammal, Outdoor, Natural scene}; DC = {Arctic}. Elementary classes are obtained as results of image-segments classification and are used as flat vocabulary for automatic image annotation. It is assumed that instances of elementary classes corre- spond to objects in the real world. Spatial relations, spatial locations and co-occurrence relations can be defined for elementary classes, like EC1 is-above EC2, or EC1 is-on-top, EC1 occurs-with EC3. Scene classes are used to represent the context or semantics of the whole image, according to common sense and expert knowledge. A part-of relation or, its inverse, relation consists-of, can be defined between an elementary class and a scene class, e.g. EC1 is-part-of SC2 or SC2 consists-of EC1. Generalized classes are defined as a generalization of scene classes. The is-a relation can be defined between a scene class and a generalized class, e.g. SC2 is-a GC1. There can be multiple levels of generalization so the relation is-a can be defined between gener- alized classes too, e.g. GC1 is-a GC3 is-a GC5. Derived classes include abstract concepts, activities, events or emotions that can be associ- ated with an image. Different types of relations, such as associate-to or is-synonym-of relation can be defined between derived classes and generalized or scene classes. 4. A multi-layered image annotaton system The architecture of our intelligent multi-layered image annota- tion system (MIAS) is depicted in Fig. 3. The system deals with all the layers of image representation given in Fig. 2, ranging from the segmented image at layer V1 to the multilayer image interpreta- tion at layer MI4. The input to the system is an image belonging to the V0 layer of the image representation and the system output is a multi-layered interpretation of the image that consists of concepts obtained from four layers of image interpretation, i.e., layers MI1, MI2, MI3 and MI4. A raw image I at layer V0 is first segmented with a normalized-cuts algorithm (Shi & Malik, 2000). The segmented image corresponds to he V1 layer of the image representation. Formally, the relationship etween the raw image I and the image segments si, i = 1, . . . . . , m ay be written as V1(I) = {s1, s2, . . . , sm}. From each image segment, ow-level features are extracted (such as size, position, height, width, olour, shape, etc.) which should represent the geometric and pho- ometric properties of a segment. Each image segment is then rep- esented by the k-component feature vector x = (x1, x2, . . . , xk)T . ccordingly, an image at the V2 layer of the image representation s described with as many feature vectors as there are image seg- ents. Thus, the relationship between the raw image I and the feature ectors xi, i = 1, . . . . . , m obtained from the image segment si, i = , . . . . . , m is given as V2(I) = {x1, x2, . . . , xm}. Each image segment is then classified using the Bayes classifier nto one of the elementary classes ECi ∈ EC according to the maxi- um posterior probability (cMAP). The Bayes classifier was trained on training set of image segments annotated with labels correspond- ng to natural and artificial objects. For each occurrence of the feature ector x, a classification is based on the Bayes theorem: MAP = argmax ECi ∈EC P(x|ECi)P(ECi) P(x) . (1) The conditional probability P(x|ECi) of a feature vector x for he given elementary classes ECi ∈ EC and the prior probability (ECi), ∀ECi ∈ EC are estimated according to data in a training set. It is aken into account that the evidence factor P(x) is a scale factor that oes not influence the classification results. The result of the image-segments classification is m annotated egments of the image I in such a manner that each one is annotated ith one of the elementary classes. The union of elementary classes, btained by the classification of the image segments, forms an au- omatic image interpretation at layer MI1, often referred to as auto- atic image annotation. The classes or elements of the interpretation et MI1(I)⊆EC are also called labels, annotation words, or keywords. A knowledge-representation scheme based on the Fuzzy Petri Net ormalism (Ribarić & Pavešić, 2009) and the fuzzy inference engine re the key components of the proposed multi-layered image annota- ion system MIAS. Defined fuzzy knowledge-representation scheme epresents the knowledge about concepts and relations in image do- ain, which is often uncertain, imprecise, and ambiguous. The fuzzy knowledge base contains the following main com- onents: fuzzy relationships between elementary classes, fuzzy elationships between elementary classes and scene classes and uzzy relationships between scene classes and generalized or de- ived classes. The fuzzy relationships are defined using the train- ng set and expert knowledge. One of the components of the M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 9543 Fig. 3. Architecture of a multi-layered image-interpretation system (MIAS). s p p T c r t c c d a c s l fi t l n M e 5 G l m a i T a I e g a t 5 k k a k f F K l 5 i k i a K w ( ystem MIAS is an inference engine (IE) used for image inter- retation on the layers MI1 − MI4. The inference engine sup- orts the fuzzy inheritance and fuzzy recognition procedures. he fuzzy inheritance is used for inconsistency checking and for lass generalization and the fuzzy recognition is applied for scene ecognition. The facts in the fuzzy knowledge base, particularly those related o relationships among elementary classes, are used to check the onsistency of the set MI1(I). An elementary class for which it is oncluded that it does not belong to a likely context, obtained e.g. ue to inaccurate segmentation, can be discarded or replaced with nother elementary class that has similar properties and fits the ontext. The elementary classes of an image that have passed incon- istency checking are the inputs into the MI2 image-interpretation ayer for scene recognition. Each scene in the knowledge base is de- ned based on a training set as an aggregation of typical elemen- ary classes. Thus, it is possible to conclude which scene is the most ikely one from the elementary classes from the set MI1(I). The recog- ised scene class makes the image interpretation at the layer MI2, I2(I)⊆SC. Based on the scene class from the set MI2(I), more abstract gen- ralized classes are inferred by the inference engine (see Section .1) using generalized relationships from the fuzzy knowledge base. eneralized relationships as well as heuristics about this particu- ar domain are explicitly specified by a human expert. Once deter- ined, the generalized classes can be further generalized to a more bstract generalized class. Inferred generalized classes form image nterpretation at the layer MI3, so for a given image I, MI3(I)⊆GC. he analogous inference procedure can be applied on generalized nd scene classes to obtain derived classes related to a given image , MI4(I)⊆DC. The outputs from the proposed system are classes at different lev- ls of abstraction that include elementary classes, scene classes and eneralized classes as well as derived classes. The defined KRFPN scheme can be independently used, modified nd connected with other KRFPN schemes in a hierarchical structure o expand the knowledge base with new concepts. . A knowledge-representation scheme To model objects and their relationships in an image, some nowledge-representation formalism has to be used and domain nowledge needs to be included. However, considering that im- ge segmentation is often imprecise and subject to errors, and that nowledge about the concept is often incomplete, an ability to per- orm conclusions from imprecise, fuzzy knowledge is necessary. or this purpose, a knowledge-representation scheme based on the RFPN formalism (Ribarić & Pavešić, 2009) is defined for a multi- ayered annotation of images. .1. Definition of the KRFPN scheme for the multi-layered mage annotation We have defined the KRFPN scheme to present elements of the nowledge base used for inferring concepts on higher layers of image nterpretation. The KRFPN scheme for multi-layered image annotation is defined s 13-tuple: RF PN = (P, T, I, O, M, �, μ, f, c, α, β, λ, Con), (2) here the first ten components are of the marked Fuzzy Petri Net FPN) (Li & Lara-Rosano, 2000): 9544 M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 pi pktj d1 d2r1 c(m1)>0 f(tj)>0 m1 Fig. 4. A generic form of a chunk of knowledge in the Fuzzy Petri Net formalism. pi pktj d1 d2r1 c(m2) = c(m1)* f(tj)>0f(tj)>0 m2 Fig. 5. A new token value is obtained in the output place after firing. f c s a m t t T e e p e a O n ( i s t a e a c m 5 o t v s l o s t s t t p P = {p1, p2, . . . , pn}, n ∈ N is a set of places; a function α: P → D maps a place from a set P to a concept from a set D used for multi- layered image annotation. It is set that D = EC ∪ SC ∪ GC ∪ DC where the subset EC includes 28 elementary classes such as {Airplane, Train, Shuttle, Ground, Cloud, Sky, Coral, Dolphin, Bird, Lion, Mountain, etc.}, the subset SC includes 20 scene classes such as {Seaside, Inland, Sea, Space, Airplane Scene, Train Scene, Tigre Scene, Lion Scene, etc.}, the sub- set GC includes generalized classes such as {Outdoor Scenes, Natural Scenes, Man-made Objects, Landscape, Vehicles, Wildlife, etc.}, and sub- set DC includes {Savannah, Africa, Safari, Vacation, etc.}. T = {t1, t2, . . . , tm}, m ∈ N is a set of transitions; a function β: T → � maps a transition from a set T to a relationship from a set � defined according to expert knowledge; a set � includes a relation- ship occurs_with between elementary classes that models the com- mon occurrence of elementary classes in the image and its negation not_occurs_with, then the aggregation relationship consists_of defined between a scene class that has a role of aggregation and elementary classes that have the role of components of aggregation, then a gen- eralization relationship is_a that is defined either between a scene class and generalized class or between generalized classes or derived classes and in addition a is_synonim_of relation defined between syn- onyms of concepts. For a relationship consists_of an inverse relation- ship – (consists_of) = is_part_of is defined. I: T → P∞ is an input function, while O: T → P∞ is an output func- tion for a transition. In our scheme, the co-domain of input and out- put functions is a set P instead of a bag P∞ as defined in (Peterson, 1981). M = {m1, m2, . . . , mr}, 1 ≤ r < ∞ is a set of tokens used by the inference engine. The inference procedure is based on the dynamic properties of the Petri Net, i.e. by firing of the transitions (Peterson, 1981). The tokens’ distribution within places is given as �(p) ∈ P(M), where P(M) is a power set of M. The initial distribution of tokens defines the initial marking vector μ0 = (μ1, μ2, . . . , μn) and μi = μ(pi) ∈ {0, 1}, i.e. in the initial marking a place can have no or at most one token. In case of scene recognition, μ0 corresponds to elementary classes obtained at the layer MI1. c: M → [0, 1] is an association function that gives a token value that corresponds to the degree of truth of the concept mapped to a place marked with that token. The value of a token in an initial distri- bution can be set to the estimated posteriori probability of a concept that is associated with that marked place or set to 1. f: T → [0, 1] is an association function that gives a transition value that corresponds to the degree of truth of a relationship mapped to a transition. The measure of truthfulness of the relationship depends on the relationship kind and is computed using data in the training set in case of pseudo-spatial and spatial relationships based on co- occurrence of elementary classes in images. Also the function f can be defined by an expert in case of more abstract classes (SC, GC and DC). λ ∈ [0, 1] is a threshold value related to transitions firing. If the threshold value λ is set, the truth value c(m1) of each token must ex- ceed the value of λ if the transition is to be enabled. Con ⊆ (� × �) is in this scheme defined as a set of pairs of mutually contradictory relations. It is defined on a set of relations occurs_with, not_occurs_with between elementary classes. It can be also defined between concepts if necessary. The KRFPN scheme can be visualized by a directed graph contain- ing two types of nodes: places and transitions. Graphically, the places pi ∈ P are represented by circles and the transitions tj ∈ T by bars. The directed arcs between the places and transitions, and the transitions and places represent the transition input I(tj)⊆P and output O(tj)⊆P functions, respectively (Fig. 4). In a semantic sense, each place from the set P corresponds to a concept di ∈ D and any transition from set T to a relation rk ∈ �. A dot within a place represents a token m1 ∈ M. To a token at the input place pi ∈ I(tj) and the transitiontj ∈ T, the values c(m1) and (tj) are assigned, respectively. The assigned values implement un- ertainty and fuzziness in the scheme and can be expressed by truth cales, where 0 means “not true” and 1 “always true”. Semantically, value c(m1) expresses the degree of uncertainty of a concept di ∈ D apped to a particular place pi ∈ P, and the value f(tj) corresponds to he degree of uncertainty of a relationship ri ∈ � mapped to a transi- ion tj ∈ T. A place that contains one or more tokens is called a marked place. he tokens give dynamic properties to the Petri Net and define its ex- cution by firing an enabled transition. A transition is enabled when very input place of the transition is marked, i.e., if each of the input laces of the transition has at least one token and if each token value xceeds the threshold value λ. An enabled transition tj can be fired. By firing, a token moves from ll its input places pi ∈ I(tj) to the corresponding output places pk ∈ (tj). In Fig. 4, there is only one input place for the transition tj, I(t j) = pi and only one output place O(t j) = pk. After the transition firing, a ew token value c(m2) at the output place is obtained as c(m1)f(tj) Fig. 5). The dynamic properties of the scheme are important for the nference-engine definition. The inference engine on the KRFPN cheme consists of two automated reasoning processes: fuzzy inheri- ance and fuzzy recognition. All the steps of the inference algorithms re given in (Ribarić & Pavešić, 2009). The complexity of both infer- nce algorithms is O(nm), where n is the number of places (concepts) nd m is the number of transitions (relations) in the knowledge base. The algorithms are used in novel and original ways to check the onsistency of classes, for scene recognition and for reasoning on ore abstract classes, as described in detail below. .2. Modeling the truth value of relationships Given that the mapping between concepts and image features is ften unreliable, and due to incomplete knowledge of the concepts, he uncertainty is implemented into the scheme by associating a alue with a transition and with a token in a marked place. A tran- ition value expresses the degree of truth or the reliability of the re- ated relationship, while a token value corresponds to the truth value r the reliability of the concept. The degree of truth of the relation- hips depends on the type of the relationship and is set according o the expert knowledge or it is computed using data in the training et. For example, the degrees of truth of the relationships that model he generalization of classes are determined by the expert, while the ruth value of the relationships consists_o f and occurs_with is com- uted using data in the training set, as explained below. M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 9545 grass p11 consists_of t80 Seaside p45 tree p25 consists_of t85 water p26 consists_of t86 0.40 0.95 cloud p6 consists_of t79 0.40 sky p20 consists_of t84 rock p17 consists_of t82 sand p18 consists_of t830.38 0.85 0.90 0.47 building p4 consists_of t78 Fig. 6. Relations among the scene “Seaside” and its components. Fig. 7. Position of objects sky, water, grass, trees and the spatial relations between the objects in the image. 5 c a e p r t c b m B s M o t c t i E c c w P P r n { E t a d m a 5 a o a P r E o r v w t 5 u i d w f E n o t t a .2.1. Relationship consists_of To define the truth-value of the aggregation relationships onsists_of, here it is assumed, that a scene may contain several char- cteristic elementary classes. Thus, the relation among the scene and lementary classes is an aggregation relationship where the scene lays the role of the aggregation and the elementary classes have the ole of the components of the aggregation. Analyzing the data in the raining set, common occurrence of elementary classes in the scene lass is determined and used for creation of the rules on relationships etween scenes and elementary classes. Instead of choosing an ele- entary class with a maximum posterior probability, the modified ayes rule is used to form a set MS that corresponds to the specific cene class. A set MSSCi for a specific scene class SCi∀i is given by: SSCi = { ECk : arg i P(SCi|ECk ) ≈ arg k P(ECk|SCi ) P(ECk) ≥ ε } . (3) The Eq. (3) mirrors the idea of finding a most representative set f elementary classes for a given scene class. MSSCi is a set of all hose elementary classes ECk, k = 1, 2, . . . that participate in a scene lass SCi with the posterior probability P(SCi|ECk), ∀k ECk exceeding he marginal value ɛ ≥ 0.05. The marginal value is determined exper- mentally. The prior probability P(ECk) for a given elementary class Ck is computed from the training set to bring in the degree of dis- rimination of each elementary class for a given scene class. The truth value attached to the aggregation relationship onsists_of between the elementary classes and the scene class as determined using the Bayes rule for the posterior probability (SCi|ECk ), ∀k ECk ∈ MSSCi for the specific scene: (SCi|ECk ) = P(ECk|SCi )P(SCi)s∑ j=1 P(ECk|SC j )P ( SC j ) , (4) s = |SC| is a number of scene classes. In Fig. 6, a part of a knowledge base is presented, showing the elationships among a particular scene class seaside and its compo- ents that correspond to elementary classes from the set MSseaside = sky, cloud, water, grass, tree, rock, sand, building} defined by the q. (3). The degree of truth f(tj) of the transition tj that corresponds o the relation consists_of between a particular scene class “Seaside” nd its components, is given by P(Seaside|ECk), ECk ∈ MSseaside and is etermined by Eq. (4). For instance, truth value of relation consists_of apped to transition t86 between a scene class “Seaside” of place p45 nd elementary class “water” of place p26 is f (t86) = 0.95. .2.2. Relationship occurs_with To create the rules on relationships between elementary classes nd to define the truth value of the relationship occurs_with, mutual ccurrence of classes ECj and ECi in each image in the training set is nalyzed. This can be formally defined as: (EC j|ECi) = P(EC j ∩E Ci) P(ECi) . (5) If the P(ECj|ECi) is less than the threshold value τ = 0.1 then the elationship not_occurs_with is defined between elementary classes Cj and ECi, i = j with the truth value of 0.9. Otherwise, the truth value f the not_occurs_with relationship is 1 − P(EC j|ECi). The occurs_with elationship is used in proposed inconsistency checking procedure to alidate the results of the image segment classification and to check hether the results obtained on all the image segments are consis- ent. .2.3. Spatial relationships Spatial relationships like at the top, at the bottom, have not been sed in this experiment since the relationships between the objects n the image differed from the natural relations. In images from the omain of natural scenes that we have used, the sky, trees, grass and ater can appear both at the bottom and at the top of the image, so or example, water can appear above the grass and trees, as in Fig. 7e. llipses in the Fig 7a–e show the positions of the segments that are ot in line with the common knowledge about spatial relationships f objects in nature. For example, the grass segment in Fig. 7c is above he tiger segment. If it turns out to be useful, spatial relationships, as well as fuzzy emporal relationships or new concepts can be added to the scheme nd used by inference engine. 9546 M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 Fig. 8. Example of image representations at layers V0 , V1 , MI1 . shuttle p19 not_occurs_with t393 0.75 rock p170.80 water p26not_occurs_with t396 not_occurs_with t394 sand p18 0.75 not_occurs_with t395 tree p25 0.90 not_occurs_with t381 lion p10 ... not_occurs_with t371 coral p7 shuttle p19 not_occurs_with t393 0.75 rock p170.80 water p26not_occurs_with t396 not_occurs_with t394 sand p18 0.75 not_occurs_with t395 tree p25 0.90 not_occurs_with t381 lion p10 ... not_occurs_with t371 coral p7 (a) (b) Fig. 9. A part of KRFPN scheme related to elementary class shuttle and the relationship not_occurs_with (a) before firing and (b) after firing the transitions. c p u c s p m t s ( t f t e p T A n w i n p s h 6. Knowledge-based approach to inconsistency checking It is to be expected that some of the elementary classes obtained using the Bayes classification rule (Eq. 4) do not fit the image con- text. To check for inconsistency of the obtained elementary classes, the inconsistency checking procedure is proposed that uses the facts in the knowledge base related to the occurs_with and not_occurs_with relations. The relations occurs_with and not_occurs_with for each ob- tained elementary class can be analyzed using the fuzzy-inheritance algorithm. Based on the results of fuzzy inheritance, the classes which are elements of domain of relation not_occurs_with are eliminated from the set MI1, the first semantic layer. In order to illustrate the proposed procedure for inconsistency checking, an example follows. Let the image I in Fig. 8 be given for a multi-layered image annotation. After the segmentation, us- ing a normalized-cuts algorithm, the image is segmented into 7 ar- eas: V1(I) = {s1, s2, . . . , s7}. For each image segment the low-level features are extracted and a feature vector is formed, so the im- age is represented at level V2 by the set of feature vectors: V2(I) = {x1, x2, . . . , x7}. Then, using the Bayes classification method, each feature vector is classified into one of the elementary classes ECi EC according to the maximum posterior probability (cMAP, Eq. (1)). For image I in Fig. 8, the obtained result, after the classification of all the image segments, is: “sky, water, water, shuttle, rock, water, sand”. Thus, the set of obtained elementary classes forms an au- tomatic image interpretation at the layer MI1 of the image I, as MI (I) = {sky, water, shuttle, rock, sand}. Note that the elementary 1 lass shuttle is a result of misclassification, because a shuttle is not resent in the image. Every obtained elementary class can be checked for inconsistency sing the not_occurs_with relationships defined between elementary lasses in MI1(I) and the fuzzy-inheritance algorithm. For instance, to check the inconsistency of the elementary class huttle, the fuzzy-inheritance algorithm is used as follows. The ap- ropriate place in the knowledge-representation scheme is deter- ined by the function α−1(shuttle) = p19, shuttle ∈ EC (Fig. 9). On he Fig. 9, presented are those not_occurs_with relations for which huttle is the input place. The initial token distribution is �0 = ∅, ∅, . . . ., {p19, 1}, . . . , ∅), i.e. the initial token is placed only on he place p19. For inconsistency checking only relations with outputs rom the set MI1(I) are useful and are shown in black. According to he original FPN algorithm all transitions related to these relations are nabled and can be fired because the number of tokens in the input lace (shuttle) is equal to the number of input arcs of the transitions. he transition values are obtained from the training set using Eq. (5). fter firing, the token is removed from the input place (shuttle) and ew tokens are created and distributed to output places (sand, rock, ater, …) as shown in Fig. 9b. The inheritance tree is formed starting from the root node which s for this example π 0(p19, {1.0}). Firing of the transitions creates ew frontier nodes of the inheritance tree that correspond to out- ut places of transitions. This step is repeated until the condition for topping of the algorithm is satisfied or the desired depth of the in- eritance tree is reached. The frontier nodes are converted by the M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 9547 Fig. 10. Inheritance tree for the class “shuttle” (Fig. 9). i t n t o f ( F a t t w p v f m o w c c t F i { 7 t ( t O 2 p i c i d t c T t i f { t c i v o { t ( t α – t d t π s i t p c w r D t n d nheritance tree algorithm into the frozen node (marked F), k- erminal (marked k-T) or identical (marked I), or one of the types of odes defined for the reachability tree (terminal, duplicate and in- erior). The inheritance tree of the KRFPN is similar to the concept f reachability tree of Petri Nets (Chen, Ke, & Chang, 1990), except or the stopping conditions that are integrated in the KRFPN scheme by the set �\{is_a}) or defined by the desired number of tree levels. ig. 10 shows a 1-level inheritance tree on the KRFPN scheme and the ppropriate semantic interpretation of the inheritance paths for he elementary class shuttle. The nodes of the inheritance tree have he form (p j , c(ml )) j = 1, 2, . . . , p, l = 1, 2, . . . , r, 0 ≤ r ≤ |M|, here c(ml) is the value of a token ml in place pj, computed as the roduct of the token value at the input place and the corresponding alue f(tj). The arcs of the inheritance tree are marked with a value (tj) and the label of a transition tj ∈ T, where, for example, t396 = 0.8 eans f(t396) = 0.8. For each of the inheritance paths the measure f truth is determined by the token value in a leaf node (the node in hich the algorithm stops). The obtained inheritance tree for the concept shuttle gives the onclusion that the class shuttle does not occur with the elementary lasses from the set MI1(I), so it can be concluded that the class shut- le most likely does not match the context of the image depicted in ig. 8 and therefore can be discarded. Accordingly, after checking for the inconsistency, the refined mage interpretation at the semantic layer MI1 is: MI1(I) = sky, rock, sand, water}. . Scene recognition For the task of scene recognition for a new, unknown image, he fuzzy-recognition algorithm based on the inverse KRFPN scheme marked as –KRFPN) is originally used. The –KRFPN scheme is ob- ained by interchanging the position of the input I and the output grass p11 -(consists_of) t80 S -(consist t85 water p26 -(consists_of) t86 0.40 0.95 cloud p6 -(consists_of) t79 0.40 sky p20 -(consists_of) t84 0.90 SceneAirplane p29 -(consists_of) t68 0.95 -(con sists_ of) t69 0.95 0.5 0.6 Fig. 11. A small part of the –KRFPN scheme for the s functions for the transition T in the 13-tuple (Ribarić & Pavešić, 009). Additionally, by changing the position of the input and out- ut functions, the relation mapped to the transition is transformed nto its corresponding inverse relation. For example, for the relation onsists_of in the KRFPN scheme its inverse relation is_part_of is used n the –KRFPN scheme, i.e., –(consists_of) = is_part_of. Also, the co- omain of the associated function c: M → [0, 1] that assigns values o the tokens (see 5.1) is expanded by cr : M → [−1, 1] so that in the ase of an exception, a token may be associated with a negative value. The proposed procedure for the scene recognition is as follows. he results of the image interpretation at layer MI1, after inconsis- ency checking, are the input to the scheme used for further image nterpretation at the layer MI2. The obtained elementary classes ECi rom MI1(I) are treated as components of an unknown scene class X. The elementary classes ECi are mapped to the places p1, p2, . . . , pn} using the function α−1 : ECi → pk. If defined, he reliability based on a posterior probability of each elementary lass ECi can be used as the token value cr(ml) in the place pk, whose nterpretations correspond to the given class ECi. Otherwise, a token alue is set to 1. For instance, let us take an image I depicted in Fig. 8. The results f the image interpretation at the layer MI1(I) are elementary classes sky, rock, sand, water} that exist in the knowledge base. Based on he Bayes classification rule (Eq. 1) the degrees of truth are assigned: sky {0.5}, sand ({0.7}), rock ({0.4}), water ({0.6}). By using the func- ion α−1 the initially marked places are determined (α−1(sky) = p20, −1(sand) = p18, α−1(rock) = p17, α−1(water) = p26). A small part of a KRFPN scheme with initially marked places and the corresponding oken value is given in Fig. 11. According to the initially marked places and the corresponding egrees of truth, four root nodesπ0 i, i = 1, . . . , 4 of the recognition rees will be formed: 1 0 (p20, {0.5}), π 20(p18, {0.7}), π 30 (p17, {0.4}), π 40(p26, {0.6}). Fig. 12 shows four corresponding recognition trees in the -KRFPN cheme with enabled transitions, starting from the root node. By fir- ng of the enabled transitions on the -KRFPN scheme, new nodes at he following higher level of the recognition tree are created and ap- ropriate values of the tokens are obtained: r(mk+1) = cr(mk) f (tl ) (6) here tl is the transition between concepts ECi and SCl, cr(mk) is the eliability of the elementary class ECi and f(tl) is computed in Eq (4). ue to the simplicity of the example, only one level of the recognition ree is generated. Note that only the recognition tree with the root ode π 2 0 (Fig. 12b) directly corresponds to the small part of –KRFPN epicted in Fig. 11. The other recognition trees (Fig. 12a, c and d) also easide p45 tree p25 s_of) rock p17 -(consists_of) t82 sand p18 -(consists_of) t830.38 0.85 0.47 -(consists_of) t94 0.4 0.7 SceneBear p30 -(consists_of) t93 SceneShuttle p47 cene recognition for image depicted in Fig. 8. 9548 M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 Fig. 12. Recognition trees with enabled transitions for each root node. ( s i v I I o a 1 i o S o a t t i { 8 n h c contain leaf nodes corresponding to the scene classes that are part of the knowledge base but are not depicted in Fig. 11. The following steps of scene recognition are as follows. Each leaf node π k i in the recognition tree k = 1, 2, . . . , b is represented by a vector of dimension |P|, where P is a set of places, so the index of a node in the recognition tree corresponds to the index of the vector component and the value of a node is assigned to a value of the vector component. For example, a node π 2 1 = (p45, 0.595) (Fig. 12b) is rep- resented by the vector π2 1 = (0, 0 . . . 0, 0.595, 0, . . , 0) so that all the vector components are assented to a value 0, except the 45th vector component, to which a node value of 0.595 is assigned. Accordingly, the total sum Z of all leaf nodes in all recognition trees is computed: Z = b∑ k=1 ok∑ i=1 πk i . (7) where πk i is ith leaf node in the k–th recognition tree, b ≤ |M| is the number of recognition trees, ok ≤ |P| is the total number of leaves in the recognition tree k. In this example there is b = 4 recognition trees, the corresponding numbers of leaves are: o1 = 12, o2 = 2, o3 = 9, o4 = 11, and the total sum is: Z = 4∑ (k=1) (ok)∑ (i=1) π k i = 12∑ i=1 π 1i + 2∑ i=1 π 2i + 9∑ i=1 π 3i + 11∑ i=1 π 4i = (0 . . . 0, 0.36, 0.44, 0.09, 0.05, 0, 0.03, 0, 0.04, 0, 0, 0.05, 0.03, 0.03, 0.04, 0, 0.16, 1.80, 0.05, 0.05, 1.11, 0, . . . 0). For example, the 30th component of the vector Z with the value 0.44 is obtained by summing all the values of the nodes in all the recognition trees that correspond to the place p30 (i.e. π 1, π 2, π 3, π 4): 0.115 + 0.175 + 0.052 + 0.102 = 0.44 7 2 8 9 Then, a set of indices of elements with the highest sum Z = Z1, Z2, . . . , Z|P|) among all of the nodes in all the recognition trees is elected as: ∗ = arg max i=1,..,|P| {Zi}. (8) In the case that there are several i for which the same maximum alue of {Zi} is obtained, the set I∗ is created: ∗ = {i∗1, i∗2, . . . ,}. (9) A scene class assigned to a place with the max argument pi: i ∈ ∗ is chosen as the best match for a given set of elementary classes btained during image interpretation at the layer MI1. In this ex- mple, the 45th component of the vector Z has the maximum value .80. Therefore, a set of max arguments consists of only one element ∗ 1 = 45, so only one scene class is chosen as the best match, i.e., the ne that is assigned to a place with that max argument, α(p45) = easide. The next scene candidate is α(p48) = Inland with a value f 1.11. By merging all the classes that are so far associated with the im- ge, from elementary classes to the scene class, the multi-layered in- erpretation of the image is formed. For example, a multi-layered in- erpretation of image I (in Fig. 8) includes the results of the image nterpretation at the layers MI1 and MI2 : MI(I) = MI1(I) ∪ MI2(I) = sky, rock, sand, water} ∪ { Seaside}. . Inference of more abstract classes The obtained scene classes can be used as root nodes for the ext inheritance process that will infer more abstract concepts from igher semantic levels (here referred as generalized and derived lasses) either because they are directly linked with the concept or M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 9549 grass p11 consists_of t80 Seaside p45 tree p25 consists_of t85 water p26 consists_of t86 0.40 0.95 cloud p6 consists_of t79 0.40 sky p20 consists_of t84 rock p17 consists_of t82 sand p18 consists_of t830.38 0.85 0.90 0.47 building p4 consists_of t78 Landscape p60 is_a t93 1.00 Natural Scene p53 is_a t392 1.00 Outdoor Scene p55 is_a t387 1.00 Vacation p59 is_associated_to t81 0.70 Fig. 13. A part of the knowledge base that shows the properties of the class “Seaside” and its parents. Fig. 14. The inheritance tree for the concept “Seaside”. m ( t s w t s k p e m t { f w c w t a ay be inferred by means of concepts at a higher level of abstraction parents). To determine related, more abstract classes for a given scene class, he relations with its parents at higher levels of abstraction are in- pected using an inheritance algorithm. The proposed procedure, by hich classes that are more abstract are concluded, will be illus- rated using the example of scene class Seaside ∈ SC that was re- ult of the recognition algorithm in Section 7. In Fig. 13, a part of a nowledge base is shown that includes information about the com- onents of the class “Seaside” and its more abstract classes defined by xpert. At the first step of the algorithm the appropriate place is deter- ined by the function α−1(Seaside) = p45. A token value c(ml) is set o 1, so the corresponding root node of the inheritance tree is π 0(p45, 1.0}). Fig. 14 shows a 3-level inheritance tree on the KRFPN scheme or the class ’Seaside’ that shows its more abstract classes (nodes ithin the ellipsis) as well as its properties. To determine more abstract classes associated with the given lass, the key nodes are those in the parent-child relationship ith the given class. The nodes in parent-child relationship for he class ‘Seaside’ are: π 14(p59, {0.7}), π 15(p60, {1.0}), π 21(p53, {1.0}) nd π 31(p55, {1.0}) and the following applies: α(p59) = Vacation, 9550 M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 Fig. 15. The inheritance tree for a synonym of Seacoast concept Seaside. i m s m r i ( ( R ( m p P M T p u c e g e e e f I w 1 m – f o e f t i c 2 m e l o α(p60) = Landscape, α(p53) = Natural scene, α(p55) = Outdoor scene. The classes “Landscape”, “Natural scene” and “Outdoor scene” are a generalization of the class “Seaside”, while the class “Vacation” is a derived class that one can associate with the class “Seaside” using the relation is_associated_to. Thus, the result of a multi-layered image annotation for the image I given in Fig. 8, after the generalization and the derived- concepts inference is: MI(I) = MI1(I) ∪M I2(I) ∪M I3(I) ∪M I4(I) = {sky, rock, sand, water} ∪{ Seaside} ∪{ Landscape, Natural scene, Outdoor scene} ∪{ V acation}. Also, new concepts can be added to the knowledge. Some exam- ples of such an extension are synonyms of the concepts defined in a scheme like Seacoast for Seaside or terms that are colloquially un- derstood as synonyms like Forest or Logs for Trees. In these cases, the is_synonim_of relation is defined between a class that is already defined in the knowledge base (e.g. Seaside) and the synonym that should be added (e.g., Seacoast). Fig. 15 shows the fuzzy-inheritance tree for the concept Seacoast, for which applies α−1(Seacoast) = p57, so the corresponding root node of the inheritance tree is π 0(p57, {1.0}). Inclusion of concepts at different levels of abstraction maps the organization of concepts from natural language to image annotation and facilitates the adjustment of the system to the user’s needs and expectations. 9. Experiments and discussion To evaluate the proposed model of a multi-layered image anno- tation, an experiment on a part of the Corel image dataset related to outdoor scenes (e.g., Landscape, Vehicles, Animals, Space) was performed. The images were automatically segmented, based on the visual similarity of pixels, using the normalized-cut algorithm. Most of the images were segmented into approximately 10 regions. Every image segment was more precisely characterized by a set of 16 visual fea- tures based on the color in CIE L∗a∗b∗ color model, size, position, height, width and shape of the area (Duygulu et al., 2002). Also, each image segment of interest was annotated with the first keyword from the set of corresponding keywords provided by (Carbonetto, Freitas, & Barnard, 2004). The vocabulary used to anno- tate the image segments has 28 keywords related to natural and arti- ficial objects such as ’airplane’, ’bird’, ‘lion’, ‘train’ etc. and landscapes like ‘ground’, ‘sky’, ‘water’ etc. The keywords from the vocabulary cor- respond to the elementary classes. Visual features and keywords of each image segment make a data set. The data set used for the experiment consists of 3960 segments obtained from 475 images of outdoor scenes. The data was because of supervised learning divided into training (2772) and test (1188) subsets by a 10-fold holdout cross validation. Data in the training set were used to learn a classification model of each elementary class us- ng a Bayes classifier. The features from the test set are used to test the odel using the corresponding elementary classes as ground-truth. To evaluate the MIAS system at layer MI1, the results of image clas- ification at that layer are compared to the ground truth. The perfor- ance of MIAS system at layer MI1 is expressed with measures of ecall (10) and precision (11). Average scores after 10 runs are shown n Fig. 16. The recall is the ratio of the correctly predicted elementary classes tp - true positive) and all elementary classes in the ground-truth data tp + fn; fn - false negative): ecall = tp tp + fn . (10) The precision is the ratio of correctly predicted elementary classes tp) and total number of elementary classes obtained from the auto- atic image interpretation at layer MI1 of the MIAS (tp + fp; fp - false ositive): recision = tp tp + fp (11) The proposed system MIAS for image interpretation at the layer I1 achieves an average precision of 32.6% and average recall of 27.5%. he average precision is calculated as the average of all the values of recision that are obtained for each elementary class in the test set sing the 10-fold cross validation. Similarly, the average recall is cal- ulated as the average of all the values of recall that are obtained for ach elementary class in the test set. Each elementary class in the raph (Fig. 16) is marked with class ID, so that ID 1 corresponds to the lementary class ‘airplane’, ID 2 to the elementary class ‘bear’, ID 3 to lementary class ‘bird’ and so on until ID 28 that corresponds to el- mentary class ‘zebra’. The highest precision, over 56% was obtained or the elementary classes: ‘grass’- ID 11, ‘polar bear’ - ID 15, ‘rock’ - D 16, ‘sky’ - ID 20 and ‘tracks’ - ID 23. The highest recall, over 57% as obtained for the elementary classes: ‘grass’ - ID 11, ‘ground’ - ID 2, ‘rock’ - ID 13, ‘sky’ - ID 20, ‘train’ - ID 24 and ‘trees’ - ID 25. For ele- entary classes ‘bear’ – ID 2, ‘building’ – ID 4, ‘cheetah’ – ID 5, ‘coral’ ID 7, ‘dolphin’ – ID 8, ‘fox’ – ID 10 and ‘zebra’ – ID 28 obtained value or both, precision and recall, is zero. Some of the reasons for this utcome are too few samples that we had available for the particular lementary class (e.g. for the class building we had only 24 segments, or the class dolphin only 20 segments), then the big diversity of fea- ures within the class (e.g. instances of class coral significantly differ n color) as well as errors in segmentation. The obtained results, given as outputs of MIAS at layer MI1, are ompared to the results of the models published in (Carbonetto et al., 004). The results of the automatic image annotation obtained for the entioned set of images with the dMRF model defined in (Carbonetto t al., 2004) and the dInd model from (Duygulu et al., 2002) are pub- ished in (Carbonetto et al., 2004). The dMRF model uses the method f Markov random fields for the automatic image annotation, while M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 9551 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Class ID P re ci si o n /R e ca ll Precision Recall Fig. 16. A precision/recall graph for the image automatic interpretation with MIAS at layer MI1 . Table 1 Comparison of the results achieved with MIAS at MI1 , dInd, dMRF. Models MIAS–MI1 dInd dMRF Number of correctly predicted classes 21 23 24 Average precision 32.6% 19.9% 21% t a a a b i a c l h t o p H i t T m t i d l c o t i f d b s r a d c A a M e l fi a he dInd model is an example of a translation model that treats image nnotation as the translation between two discrete languages. The uthors have reported the precision for the task of automatic image nnotation for each of 28 keywords in the vocabulary achieved by oth models. Comparing the results of the automatic annotation on mages related to outdoor scenes, the dMRF model achieves an aver- ge precision of 21%, while the dInd model achieves an average pre- ision of 20%. As specified in the Table 1, average precision of MIAS at ayer MI1 exceeds the precision of both models although our system as correctly predicted fewer classes, 21 classes out of 28 possible. Comparing the results presented in Table 1, it should be noted hat, for learning, the models dInd and dMRF used image labels, while ur approach uses labels of the image segments. Because of the su- ervised learning approach, somewhat better results were expected. owever, the achieved difference for the average precision is signif- cant, although the dMRF model took into account the context. Note hat the given results of our system MIAS at layer MI1 (represented in able 1.) are without any checking for inconsistency. Table 2 Examples of multi-layered image annotation by MIAS. Image example: Multi-layered image annotation MI1 ‘shuttle’ - ID 19 ‘train’’ – ID 2 –ID 23, ‘sk MI2 ‘Shuttle Scene’, ‘Train Scene MI3 ‘Vehicle’, ‘Man-Made Object’, ‘Outdoor’ ‘Vehicle’, ‘Ma Objectȁ, ‘O MI4 ‘Space’ ‘Transport’ Generally, the results achieved by image automatic annotation odels on outdoor image domains are relatively poor and the ques- ion is whether they can meet customer requirements when retriev- ng or organizing images. Often the results of automatic annotation epend on the quality of the segmentation, so when an image has a ot of segments and when an object is over segmented, the results an include labels that do not correspond to the object or context f an image. Here, to mitigate this problem, the obtained results of he image interpretation at the layer MI1 are analyzed with the fuzzy nheritance algorithm. The aim is to purify the classification results rom class labels that do not match the likely context of the image. To o so, the facts from the knowledge base related to the relationships etween elementary classes as well as the reliability of the relation- hips computed with (5) are used with the fuzzy inheritance algo- ithm. Using inconsistency checking, those elementary classes that re obtained as a result of the image interpretation at layer MI1 and id not fit the likely context are discarded. As a consequence, the pre- ision of the image interpretation at layer MI1 is increased up to 43%. further improvement of the precision could be achieved by defining dditional relationships between the elementary classes. Afterwards, automatic image interpretation at layer MI2 of the IAS is performed by the fuzzy-recognition algorithm, using the lementary classes obtained as results of image interpretation at ayer MI1 and using knowledge about a particular domain. To de- ne the relationships between the scenes and the elementary classes utomatically, and to determine their reliability, we have analyzed 4, ‘tracks’ y’ - ID 20 ’grass’ – ID 11, ‘tiger’’- ID 22 ’water’ – ID 26, ‘sand’ – ID 18, ‘sky’ - ID 20, ‘road’-ID 16 ’, ‘Tiger Scene’, ‘Seaside’, n-Made utdoor’ ‘Wildcat’, ‘Wildlife’, ‘Natural Scenes’, ‘Outdoor Scene’ ‘Natural Scenes’, ‘Outdoor Scene’ ‘Savannah’ ‘Vacation’ 9552 M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 w T l a f g a e a h l c c c w c ( i t s D ( a t p c t w o f l i a m l t t a f r a u a e g t c e i i t a n q t s s t b r t elementary classes and scenes related to each image. Therefore, we have supplemented the existing vocabulary with 20 classes related to the scenes such as ‘Airplane Scene’, ‘Bird Scene’, ‘Sea’, etc. Then, we have used these classes of 475 images to make a data set to be used for scene recognition. The data was divided into training and test subsets (70:30) by a 5-fold holdout cross validation. Data in the training set were used to produce the rules about relationships between scenes and elementary classes according to (3), and to learn a classification model of each scene class using a Bayes classifier, according to (4). Obtained precision of automatic image interpretation at layer MI2 is 61% and the recall is 55%. The results at the layer MI2 depend on the results at layer MI1. For those scenes for which there is one main object class which is highly discriminant for that scene (e.g train for Train Scene), it is crucial to detect that object. In this kind of scenes background objects that are common to most scenes do not play an important role, but in scenes without one prominent object (e.g. Sea, Inland) they are important. Additionally, the inheritance algorithm is used to infer generalized classes related to a scene class that make the interpretation at the layer MI3 and derived classes at the layer MI4. In Table 2, some examples of a multi-layered image annotation obtained by MIAS are shown. 10. Conclusion The aim of the present research was to annotate automatically im- ages with words that are used in an intuitive image search. These words correspond to concepts on different levels of abstraction, in order to enable simple retrieval and organization of images. These concepts cannot be simply mapped to features but require additional reasoning with general and domain-specific knowledge, which can in context of image interpretation often be incomplete, imprecise, and ambiguous. Therefore, the ability of handling uncertainty and reason- ing from fuzzy knowledge turned out to be an important property. We have developed the fuzzy-knowledge based intelligent system MIAS for multi-layered image annotation supported with fuzzy in- ference engine that is capable to deal with approximate reasoning and uses the available knowledge to draw conclusions about image semantics. In order to bridge the semantic gap between the visual content of an image and the image semantics, the MIAS system deals with visual content of images (low-level features) and with image semantics (el- ementary, scene, generalized and derived classes) that are inspired by human image interpretation and presented on layers MI1 − MI4. We have merged the statistical approach for classification of im- age segments with knowledge-based approach to infer concepts that are more abstract. For classification of image segments into elemen- tary classes below the MI1 layer, a Bayesian classifier is used. The ar- chitecture of the MIAS system facilitates its compatibility with vari- ous classification methods so that other classification methods can be used as well. The fuzzy knowledge base is built using a fuzzy knowl- edge representation scheme based on Fuzzy Petri Nets (KRFPN) for- malism. The hierarchical arrangement of the KRFPN schemes used in MIAS allows that the schemes can be independently used, modified and connected with each other into a new hierarchical structure, e.g. to expand the knowledge base with new concepts that may be syn- onyms or with concepts on different semantic levels. The acquisition of knowledge was facilitated so that all the facts and rules on composition and distribution of concepts, as well as their reliability are produced automatically from data in a training set. A human expert explicitly specifies only the facts about general knowl- edge and heuristics about the particular domain. Both new relation- ships and new concepts with appropriate reliabilities can be stored into the knowledge base and used by the inference engine. The approximate reasoning capability of the inference engine sup- ported in KRFPN scheme was used in an original way for automatic scene recognition, for inference of classes that are more abstract as ell as for inconsistency checking of the classified image segments. he concepts obtained by classification of image segments at the ayer MI1 (elementary classes) were treated as components of scenes t the layer MI2 that can be inferred by further analysis with the in- erence engine. Thus obtained concepts were then used for inferring eneralized and derived classes related to the image at the layers MI3 nd MI4. Since the decisions about more abstract concepts can be made ven when input information about the concepts present in an image re imprecise and vague, the errors can be propagated through the ierarchical structure of concepts and affect the inference on higher evels. To reduce this problem, we have proposed a novel consistency- hecking procedure that checks consistency of obtained elementary lasses at the layer MI1 with the determined image context and dis- ards the intruder classes, to increase the reliability of conclusions, as ell as to improve the precision of image annotation. The results of image annotation at layer MI1 of the MIAS were ompared with the published results of automatic image annotation Carbonetto et al., 2004, Duygulu et al., 2002), on the same set of mages and using the same image features. It has been shown that he supervised learning approach provides significantly better re- ults than the unsupervised methods used in (Carbonetto et al., 2004, uygulu et al., 2002), even when they take into account the context Carbonetto et al., 2004). After the inconsistency checking is performed, the results of im- ge annotation at layer MI1 are significantly improved considering he average precision. Additionally, the proposed system MIAS sup- orts the recognition of scenes and reasoning about the related con- epts at different levels of abstraction, to mimic the way people in- erpret images and to enrich the image annotation with concepts that ould people most likely use when searching for these images. The main contributions of the presented research in the field f expert and intelligent systems are related to the definition of uzzy knowledge-representation scheme KRFPN for automatic multi- ayered image annotation and to novel and original use of approx- mate reasoning capabilities of the inference engine for inferring bout the semantics of images. The main advantages of the proposed ulti-layered image annotation MIAS system is the fusion of low- evel image features and knowledge based concepts related to seman- ics of an image. Another advantage stems from the connection be- ween the statistical and knowledge-based approach in order to take dvantages of their strengths so that statistical methods are used to acilitate the knowledge acquisition and for automatic generation of elationships between concepts as well as for computing their reli- bility. Other strengths of the MIAS system arise from the original se of fuzzy inference engine for scene recognition and for reasoning bout more abstract concepts as well as the novel use of inference ngine for checking consistency of concepts to reduce error propa- ation through the hierarchical structure of the scheme. Thanks to he KRFPN formalism, the MIAS system proved to be successful in oping with incomplete, imprecise, uncertain and ambiguous knowl- dge. The rules in the knowledge-base of MIAS can be visualized us- ng Fuzzy Petri Nets and conclusions can be directly understood us- ng the inference trees. Another advantage of the proposed system hat arises from the KRFPN formalism is the ability to be extended by dding new rules and to be adapted to a new domain by acquiring ew facts and adapting the fuzzy knowledge base. The proposed system architecture facilitates the knowledge ac- uisition phase, but due to automatic generation of rules, a larger raining set of images is needed. The automatically generated rules trongly depend on the used data set, so when images in the training et are not representative, the automatically generated rules on spa- ial relationships between objects in the images and the relationships etween objects and scenes may not be general enough and their eliability may not be properly set. Therefore, after development, he system should be additionally tuned for accuracy. Although the M. Ivasic-Kos et al. / Expert Systems With Applications 42 (2015) 9539–9553 9553 a c i v i w a M m d t t f d o s b A t i R A B B B C C C D D D E F F H H H I I L L L M M N P P R S S S S S S T Y Y Z rchitecture of the MIAS system is general, the limitation is that it annot be immediately used for new applications or domains. The mages should be preprocessed to obtain low-level features, a new ocabulary should be defined and new rules created, either automat- cally using the training data set or provided by an expert. This research was oriented to the domain of outdoor images, so e plan to implement and test the proposed system in new domains nd with very large image databases. Since the architecture of the IAS system facilitates its compatibility with various classi fication ethods, for the first layer of image interpretation we will examine ifferent classification methods and methods of probability estima- ion as well as optimized mechanisms for extracting visual features. In the future research, we plan to expand the proposed model and o examine the possibilities of its adaptation for annotation of videos, or recognition of activities in image or video contents and for pre- iction of future actions. Therefore, we will examine the possibility f including the fuzzy spatial and temporal relations into the MIAS ystem as well as explore the required adaptations of formalisms to e used in the system. cknowledgment This work has been fully supported by Croatian Science Founda- ion under the project 6733 De-identification for Privacy Protection n Surveillance Systems (DePPSS). eferences thanasiadis, T., et al. (2009). Integrating image segmentation and classification for fuzzy knowledge-based multimedia. In Proceedings of the MMM2009. arnard, K., Duygulu, P., Forsyth, D., Freitas, N., Blei, D. M., & Jordan, M. I. (2003). Match- ing words and pictures. Journal of Machine Learning Research, 3, 1107–1135. enitez, A. B., Smith, J. R., & Chang, S. F. (2000). Medianet: A multimedia information network for knowledge representation. In Proceedings of the IS&T/SPIE: v. 4210 MA. lei, D., & Jordan, M. (2003). Modeling annotated data. In Proceedings of the 26th annual International ACM SIGIR Conference on Research and Development in Information Re- trieval (pp. 127–134). arbonetto, P., Freitas, Nde., & Barnard, K. (2004). A statistical model for general contex- tual object recognition. In Proceedings of ECCV 2004, Czech Republic, (pp. 350–362). hen, S. M., Ke, J. S., & Chang, J. F. (1990). Knowledge representation using fuzzy Petri nets. IEEE Transactions on Knowledge and Data Engineering, 2(3), 311–319 1990. hengjian, S., Zhu, S., & Shi, Z. (2015, May). Image annotation via deep neural network. In Proceedings of IEEE 14th IAPR International Conference on Machine Vision Applica- tions (MVA) (pp. 518–521). atta, R., Joshi, D., & Li, J. (2008). Image retrieval: Ideas, influences, and trends of the new age. ACM Transactions on Computing Surveys, 20, 1–60. ong, P. T. (2014). A survey of refining image annotation techniques. International Jour- nal of Multimedia & Ubiquitous Engineering, 9(3). uygulu, P., Barnard, K., de Freitas, J. F. G., & Forsyth, D. A. (2002). Object recognition as machine translation: learning a lexicon for a fixed image vocabulary. In Proceedings of European Conference on Computer Vision (pp. 97–112). akins, J., & Graham, M. (2000). Content-Based Image Retrieval. Technical Report JTAP- 039, JISC, Institute for Image Data Research, University of Northumbria, Newcastle. ei-Fei, L., & Perona, P. (2005). A bayesian hierarchical model for learning natural scene categories. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, CVPR.: vol. 2 (pp. 524–531). IEEE. eng, S., & Xu, D. (2010). Transductive multi-instance multi-label learning algorithm with application to automatic image annotation. Expert Systems with Applications, 37(1), 661–670. are, J. S., Lewis, P. H., Enser, P. G. B., & Sandom, C. J. (2006). Mind the Gap: Another look at the problem of the semantic gap in image retrieval. In Proceedings of Multimedia Content Analysis, Management and Retrieval San Jose, California, USA.. ong, R., Wang, M., Gao, Y., Tao, D., Li, X., & Wu, X. (2014). Image annotation by multiple-instance learning with discriminative feature mapping and selection. IEEE Transactions on Cybernetics, 44(5), 669–680. u, J., & Lam, K. M. (2013). An Efficient Two-Stage Framework for Image Annotation. Pattern Recognition, 46(3), 936–947. vasic-Kos, M., Pavlic, M. &, & Pobar, M. (2009). Analyzing the semantic level of out- door image annotation. In Proceedings of 32nd IEEE International Convention On In- formation And Communication Technology, Electronics And Microelectronics-MIPRO (pp. 293–296). Opatija, Croatia. vasic-Kos, M., Ribarić, S. &, & Ipsic, I. (2010). Image annotation using fuzzy knowledge representation scheme. In Proceedings of the IEEE 2010 International Conference of Soft Computing and Pattern Recognition (pp. 218–223). Paris, France. i, X., & Lara-Rosano, F. (2000). Adaptive fuzzy Petri nets for dynamic knowledge rep- resentation and inference. Expert Systems with Applications, 19(3), 235–241. i, J., & Wang, J. Z. (2003). Automatic linguistic indexing of pictures by a statistical modeling approach. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(19), 1075–1088. iu, Y., Zhang, D., Lu, G., & Ma, W. Y. (2007). A survey of content-based image retrieval with high-level semantics. Pattern Recognition, 40(1), 262–282. aillot, N.E. (2005). Ontology based object learning and recognition (PhD thesis), Univer- site de Nice-Sophia Antipolis. arques, O., & Barman, N. (2003). Semi-automatic semantic annotation of images us- ing machine learning techniques. In Proceedings Of International Semantic Web Con- ference (pp. 550–565). ezamabadi-pour, H., & Kabir, E. (2009). Concept learning by fuzzy k-NN classification and relevance feedback for efficient image retrieval. Expert Systems with Applica- tions, 36(3), 5948–5954 Part 2, April. apadopoulos, G. T. H., Saathoff, C., Escalante, H. J., Mezaris, V., Kompatsiaris, I., & Strintzis, M. Z. (2011). A comparative study of object-level spatial context tech- niques for semantic image analysis. Computer Vision and Image Understanding, 115(9), 1288–1307. eterson, J. L. (1981). Petri net theory and the modeling of systems. Prentice Hall PTR. ibarić, S., & Pavešić, N. (2009). Inference procedures for fuzzy knowledge representa- tion scheme. Applied Artificial Intelligence, 23, 16–43 January 2009. hatford, S. (1986). Analyzing the subject of a picture: A theoretical approach. Cata- loguing & Classification Quarterly, 5(3), 39–61. hi, J., & Malik, J. (2000). Normalized cuts and image segmentation. IEEE Transaction on PAMI, 22(8), 888–905. imou, N., Athanasiadis, T., Stoilos, G., & Kollias, S. (2008). Image indexing and re- trieval using expressive fuzzy description logics, Signal, Image and Video Processing: 2 (pp. 321–335) December. Springer December. meulders, A. W. M., Worring, M., Santini, S., Gupta, A., & Jain, R. (2000). Content-based image retrieval at the end of the early years. IEEE Transaction on Pattern Analysis and Machine Intelligence, 22(12), 1349–1380. rikanth, M., Varner, J., Bowden, M., & Moldovan, D. (2005). Exploiting ontologies for automatic image annotation. In Proceedings of SIGIR: 05 (pp. 552–558). toilos, G., Stamou, G., Tzouvaras, V., Pan, J. Z., & Horrocks, I. (2005). The fuzzy descrip- tion logic f-shin. In Proceedings of International Workshop on Uncertainty Reasoning For the Semantic Web. ousch, A. M., Herbin, S., & Audibert, J. Y. (2012). Semantic hierarchies for image anno- tation: A survey. Pattern Recognition, 45(1), 333–345. in, H., Jiao, X., Chai, Y., & Fang, B. (2015). Scene classification based on single-layer SAE and SVM. Expert Systems with Applications, 42(7), 3368–3380 1 May. u, Y., Pedrycz, W., & Miao, D. (2014). Multi-label classification by exploiting label cor- relations. Expert Systems with Applications, 41(6), 2989–3004 May. hang, D., Islam, M. M., & Lu, G. (2012). A review on automatic image annotation tech- niques. Pattern Recognition, 45(1), 346–362. http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0003a http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0003a http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0003a http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0003a http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0022 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012a http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012a http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012a http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0012a http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0023 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0023 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0023 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0023 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0024 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0024 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0025 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0025 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0025 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0025 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0026 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0027 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0028 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0029 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0030 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0031 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0032 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0033 http://refhub.elsevier.com/S0957-4174(15)00528-X/sbref0033 A knowledge-based multi-layered image annotation system 1 Introduction 2 Related work 3 Multi-layered image representation 4 A multi-layered image annotaton system 5 A knowledge-representation scheme 5.1 Definition of the KRFPN scheme for the multi-layered image annotation 5.2 Modeling the truth value of relationships 5.2.1 Relationship consists_of 5.2.2 Relationship occurs_with 5.2.3 Spatial relationships 6 Knowledge-based approach to inconsistency checking 7 Scene recognition 8 Inference of more abstract classes 9 Experiments and discussion 10 Conclusion Acknowledgment References 
work_xvfefljjunfblp4nyjja7f4dy4	----	Developing Backward Chaining Algorithm of Inference Engine in Ternary Grid Expert System (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3, No. 9, 2012 241 | P a g e www.ijacsa.thesai.org Developing Backward Chaining Algorithm of Inference Engine in Ternary Grid Expert System Yuliadi Erdani Politeknik Manufaktur Bandung Bandung, Indonesia Abstract— The inference engine is one of main components of expert system that influences the performance of expert system. The task of inference engine is to give answers and reasons to users by inference the knowledge of expert system. Since the idea of ternary grid issued in 2004, there is only several developed method, technique or engine working on ternary grid knowledge model. The in 2010 developed inference engine is less efficient because it works based on iterative process. The in 2011 developed inference engine works statically and quite expensive to compute. In order to improve the previous inference methods, a new inference engine has been developed. It works based on backward chaining process in ternary grid expert system. This paper describes the development of inference engine of expert system that can work in ternary grid knowledge model. The strategy to inference knowledge uses backward chaining with recursive process. The design result is implemented in the form of software. The result of experiment shows that the inference process works properly, dynamically and more efficient to compute in comparison to the previous developed methods. Keywords- expert systems; ternary grid; inference engine; backward chaining. I. INTRODUCTION There is no official definition for the term of expert system but there are some descriptions for it created by people working in the field of expert system. With the term of expert system we refer to a computer system or a program, into which several procedures of artificial intelligence are integrated. Expert system can be understood as vehicles for Artificial Intelligence (AI) techniques. Expert system is also applied artificial intelligence. Expert systems are programs for storing and processing knowledge of a special area, that why they are able to answer to questions and solve problems, with which experts normally deal [6]. In the current situation, expert system is an intelligent computer program that uses knowledge and inference procedures to solve problems that are difficult enough to require significant human expertise for their solution. The expert knowledge must be obtained from specialist or other sources of expertise, such as texts, journal, articles, and database [8]. This type of knowledge usually requires much training and experience in some specialized field such as medicine, geology, system configuration, or engineering design. Once a sufficient body of expert knowledge has been acquired, it must be encoded in some form, loaded into a knowledge base, then tested, and refined continually throughout the life of the system. Some task that can be performed by expert system are difficult tasks to be specified, the task that may have incomplete or uncertain data, there may not always be an optimum solution, the task cannot be solved in a step-by-step manner, and solutions are often obtained by using accumulated experience [11]. An example of applied expert system is web- based consultation system [1]. Benefit of expert systems is the ability to preserve valuable knowledge which would otherwise be lost when an expert system is no longer available. Expert system also can allow an expert to concentrate on more difficult aspect of the task. It can enforce consistency, and they can perform dangerous tasks which would otherwise be carried out by humans. One of known and very popular expert system type is production rule. Production rule are simple but powerful forms of knowledge representation providing the flexibility of combining declarative and procedural representation for using them in a unified form. The term production rule came from production system which is developed by A production system is a model of cognitive processing, consisting of a collection of rules (called production rules, or just productions). Each rule has two parts: a condition part and an action (conclusion) part. The meaning of the rule is that when the condition holds true, then the action is taken. A typical production rule is given below: IF there is a flame THEN there is fire The statement of the rule above means that fire is caused by a flame. If anything happens with a flame, it will lead to fire production. It is the idea of production system. The production system or production rule provides appropriate structures for performing and describing search process. A production system has four basic components as enumerated below [9]: A set of rules following the classical IF-THEN construct. If the conditions on the left-hand side are satisfied, the rule is fired, resulting in the performance of action on the right-hand side of the rule, A database of current facts established during the process of inference, A control strategy which specifies the order in which the rule are selected for matching of antecedents by comparing the facts in the database. It also specifies how to resolve conflicts in selection of rule or selection of facts, and a rule firing module. An expert system that impalements production rule is known as rule-based expert system. Building or construction of rules can easily be done in most rule-based expert system. Knowledge expert or knowledge engineer does not have to do any work specifying (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3, No. 9, 2012 242 | P a g e www.ijacsa.thesai.org rules and how they are linked to each other. Sometime the knowledge expert or knowledge engineer can reference rules or facts that have not yet been created. It seems to be a simple and an instant work. The problem due to the performance of the knowledge will not occur until the number of rules is getting higher. Some problem may appear in the form of inconsistent rules, unreachable rules, redundant rule and closed rule chain of rules. The mentioned problems above have been solved with so called Ternary Grid [1][4][5]. Since the ternary grid can solve some problem concerning knowledge bottleneck, there is no any developed inference method, technique or machine working on ternary grid knowledge model. As consequence of it, all ternary grid knowledge must be converted into production system format, so that the knowledge can be processed by rule-based inference machine to deliver solution. The inference engine of an expert system interprets and evaluates the facts in the knowledge base in order to provide an answer. By the numerous methods of problem solution, which can be implemented in a rule interpreter, only the representatives of the concatenation strategies are to be treated here: forward chaining and backward chaining The developed inference machine of expert system can work in ternary grid knowledge model. The strategy to find solution previously uses forward chaining with iterative approach [12]. As another alternative solution, the inference machine can be implemented by using backward chaining method. The emphasis of this paper is to describe the backward chaining method that is implemented in the inference engine of ternary grid expert system. The backward chaining method should bring more benefit than developed forward chaining method, e.g. dynamic answers of inference engine, efficient computation effort, etc. II. METHODS Before talking the inference engine, we must first regard the knowledge representation. The Ternary Grid represents the production rule in the following structure (Fig. 1): Figure 1. Ternary Grid basic structure Ri: Rule i (i is the number of rule) Fj: Fact j or logical term (j is the number of fact) },...,3,2,1{ Ii  },...,3,2,1{ Jj  1 IJ The Value of every grid box is 0, 1 or 2 0 = unused, is represented by empty grid box. 1 = Fact Fm belongs to the condition part of rule Rn (LHS= Left Hand Side). 2 = Fact Fm is part of the conclusion part of Rn (RHS = Right-Hand Side). In the beginning of the development, the Ternary Grid was only used for knowledge acquisition system. The basic feature of the system architecture is to organize the independent and sequential obtaining process of the factual knowledge and the elicitation process of judgmental knowledge using Ternary Grid. The overall systems architecture is presented in terms of collection of functions providing effective acquisition, processing, transferring and flexible transformation of knowledge. This section gives an overview of the system that shows the design approach of the system and the concept of acquisition process. The Ternary Grid acquisition system has task to organize the knowledge base, to obtain the factual knowledge, to elicit the judgmental knowledge, and to transfer the knowledge into knowledge base. The system was able to improve the performance of expert system knowledge [5]. Meanwhile the Ternary Grid has been used for other part related the expert system, such as knowledge representation, knowledge based system, inference engine, etc. Even the Ternary Grid has been applied in an inference engine, but the approach of existing inference engine uses only forward chaining method. The backward chaining method has not been used in any inference engine. The developed backward chaining method will be implemented in inference engine of expert system based on Ternary Grid. Inference engine of expert system is computer program that answers questions from user. It processes all information from the knowledge base by firing rules and facts [9]. Backward chaining is a strategy of inference process which is the opposite of forward chaining. The strategy of backward chaining is started from a goal and ended with a fact that leads to the goal. Backward chaining method is also called as goal driven strategy of inference engine. In other literature, the backward chaining is a chaining process that begins with the last element in the chain and proceeds to the first element. This is often a very effective way of developing complex sequences of behavior There are two search algorithms, which are normally used by backward chaining method, i.e. depth-first and breath-first search algorithm. Both algorithms search data in a tree structure. Depth-first search algorithm searches a data in a tree as deep as possible before backtracking. Breath-first search algorithm searches a data in neighbor nodes before it moves deeper to the bottom of the tree. Using depth-first search algorithm, the process of backward chaining can be illustrated as follows: (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3, No. 9, 2012 243 | P a g e www.ijacsa.thesai.org Figure 2. Backward chaining illustration Figure 3 shows the illustration of backward chaining process of inference engine of expert system. The known facts are F1.1, F2.1, F3.2 and F3.4 in the beginning of process. The inference process begins from fact F1.1. That fact F1.1 is called as goal. The inference process moves then backward to other facts behind goal. It is so called condition part of rule. The inference engine tries to applied fact F2.1, F2.2 and F2.3, etc. The developed backward chaining uses depth-first search algorithm. The following steps describe the inference process of expert system using developed backward chaining algorithm as follows: Goal: F1.1 Fetch: F2.1 → unknown Fetch: F3.1 → known Rule (F3.1, F2.1) → fired Rule (F2.1, F1.1) → fired Fetch: F4.1 → unknown Fetch: F2.2 → unknown Fetch: F3.2 → unknown Fetch: F4.2 → unknown Fetch: F4.3 → known Rule (F4.3, F3.2) → fired Rule (F3.2, F2.2) → fired Rule (F2.2, F1.1) → fired … Etc. The inference process continues until all possible facts have been asked (tested) to be fired. The developed algorithm searches all data deeper into the bottom of tree structure in the knowledge base of expert system. The designed and implemented backward chaining algorithm is explained in the following algorithm: The process continues as far as the number of rules is less than the number of existing rules and there is rule that is not applicable. If a rule is applicable (fired) then the program search the next facts that lead to applicable rule until there is no fact found anymore. If a rule is not applicable (fired) then the program search other possible rule in other paths. The process continues until all possible facts have been tested or fetched. III. RESULTS The same data as [12] [13] is used in this experiment. According to Ternary Grid acquisition technique [5], the mentioned rules are inputted into ternary grid knowledge base as it is shown in figure 4. Using the developed concept, the rule-based format must not be converted into ternary grid. The inference process of the expert system in ternary grid uses backward chaining with recursive approach. All fact inputs are stored in set of facts Fk. The inference engine searches all rules that are possible to be executed and stores them in set of rules Rx:  FFFpFpqppR kkx  ,,, (5) The inference engine determines then rules that are able to be applied and stores in the following set of rule Ryn. xny RR  (6) (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3, No. 9, 2012 244 | P a g e www.ijacsa.thesai.org Figure 3. Given facts and rules in ternary grid Figure 4. Kknown facts The application does then the inference task by processing all facts that are given before. The result of inference process can be shown in figure 6. Inconsistent rules can be detected and eliminated using the following processes:  Find rows, in which value 3 appears:  3 ij aiB (7)  Remove row duplication  Bb bC   (8) The result of inference process shows the effectiveness of the developed algorithm. In comparison to the method of [7] [12] and [13], the developed inference method can work directly in ternary grid without having to be converted to rule- based format. In comparison to [12] and [13], the developed method work more dynamic and efficient to compute. The implemented recursive approach in inference process reduced the number of required iteration. The result of inference process can also show other facts that weren’t known before. These all facts could lead to give more conclusions that will bring more information to expert system. Figure 5. Result of inference process using backward chaining algorithm The following data are taken from several conducted experiments TABLE I. EXPERIMENT DATA Number of rule Number of fact Number of Iteration 10 35 62 20 70 134 30 100 295 50 180 673 Figure 6. Recursion effort 0 500 1000 1 2 3 4 N u m b e r o f It e ra ti o n Rule Number Iteration Effort (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3, No. 9, 2012 245 | P a g e www.ijacsa.thesai.org Figure 6 show the effort of recursion that is influenced by increasing the number of facts and rules. IV. CONCLUSION The developed inference engine using backward chaining method in ternary grid works properly. It can determine all applied rules that lead to the goal fact. In comparison to the previous work using iterative approach and forward chaining algorithm, the developed method works more dynamic and more efficient to compute. The inference process could also detect and lead to other rules that previously unknown and brought new solutions. Referring to some literatures concerned expert systems; the developed method is novel and will give contribution in developing inference method of expert systems. REFERENCES [1] David J.C. McKay. Information Theory, Inference, and Learning Algorithms. Cambridge University Press. 2003. [2] Erdani, Y., Hunger, A., Werner, Stefan., Mertens, S. Web-Based Consultation System with Expert System, IASTED CST 2003 – International conference (International Association of Science and Technology for Development – Computer Science and Technology), May 19-21, 2003, Cancun, Mexico, 2003, ISBN – 0-88986-349-0, page 61-64 [3] Erdani, Y., Hunger, A., Werner, S., Mertens, S. Ternary Grid as a Potentially New Technique for Knowledge Elicitation/Acquisition, 2nd IEEE Conference on Intelligent System, vol I: pp. 312-315. ISBN 0- 7803-8278-1, Varna - Bulgaria, June 22-24, 2004. This paper has been accepted also by SEKE 04 (Software Engineering and Knowledge Engineering) conference in Banff, Canada. [4] Erdani, Y., Hunger, A., Werner, S., Improving the Knowledge Performance using Ternary Grid Knowledge Acquisition and Model, WSEAS Transactions (Journals) on Information Science and Application, Issue 2, Volume 2, February 2005. ISSN 1790-0832 [5] Erdani, Yuliadi, Acquisition of Human Expert Knowledge for Rule- based Knowledge-based Systems using Ternary Grid, Verlag – Dissertation.de, Berlin, 2005, ISBN 3-89825-000-8 [6] Horn, Christian; Kerner, Immo O. : Lehr- und Übungsbuch Informatik, Band 3. Fachbuchverlag Leipzig, im Carl Hanser Verlag, München Wien, 1997. ISBN 3-446-18699-9 [7] Hunger, A., Werner, S., CONGA: A Course Online/ Offline Information and Guidance System to support an International Degree Course, Proceeding of ICCE 99 (Chiba-Japan, 1999, ISBN 1 58603 027 2, Page 577-583) [8] Joseph C. Giarratano, Gary Riley. Expert Systems, Principles and Programming, 2005, ISBN 0-534-38447-1 [9] Krishnamoorthy, C. S., Rajeev, S., Artificial Intelligence and Expert Systems for Engineer, CRC Press, Boca Raton – Florid, 1996, ISBN 0- 8493-9125-3 [10] Lunze, Prof. Dr.-Ing. Jan. Künstliche Intelligenz für Ingenieure, Band 1: Methodische Grundlagen und Softwaretechnologie, R. Oldenbourg Verlag, München Wien, 1994, ISBN 3-486-22287-2 [11] Sascha Mertens, Marius Rosu, Yuliadi Erdani. An intelligent dialogue for online rule based expert systems, Proc. of the 2004 International Conference on Intelligent User Interfaces, January 13-16, 2004, Funchal, Madeira, Portugal. ACM 2004, ISBN 1-58113-815-6 [12] Yuliadi Erdani, Pengembangan Metode Inferensi untuk Sistem Pakar berbasis Ternary Grid, Jurnal P&PT Vol. VIII, No. 2, Desember 2010, ISSN 0854-5766 [13] Yuliadi Erdani, Developing Recursive Forward Chaining Method in Ternary Grid Expert Systems, IJCSNS (International Journal of Computer Science and Network Security), Vol. 11 No. 8, August 2011, ISSN 1738-7906 AUTHORS PROFILE Dr. Ing., Yuliadi Erdani, M.Sc., Dipl. EL. Ing., HTL received the B.S. degree in Electrical Engineering from the Ingenieurschule Burgdorf (ISB) Switzerland in 1995 and the M.S. degree in Computer Science and Communication Engineering (CSCE) from the University of Duisburg, Germany in 2002. He received PhD degree (Dr.-Ing.,) in Informationstechnik from the same university i.e. University of Duisburg-Essen, Germany in 2005. He did a lot of work with expert systems. He has been continuing his research work in the area of expert systems from 2006 until now and every year he always got research fund/grant. He works now as lecture in a state Polytechnic (university of applied science) in Bandung, Indonesia. 
work_xvokgoinujc6dfyjg4yqhyecqu	----	A novel feature fusion based on the evolutionary features for protein fold recognition using support vector machines A novel feature fusion based on the evolutionary features for protein fold recognition using support vector machines Mohammad Saleh Refahia, A. Mir1, Jalal A. Nasiri1,∗ aDepartment of Electrical Engineering, Amirkabir University of Technology,Tehran,Iran bIranian Research Institute for Information Science and Technology (IranDoc), Tehran, Iran Abstract Protein fold recognition plays a crucial role in discovering three-dimensional structure of proteins and protein functions. Several approaches have been employed for the prediction of protein folds. Some of these approaches are based on extracting features from protein sequences and using a strong classifier. Feature extraction techniques generally utilize syntactical-based information, evolutionary-based information and physiochemical-based information to extract features. In recent years, Finding an efficient technique for integrating discriminate features have been received advancing atten- tion. In this study, we integrate Auto-Cross-Covariance (ACC) and Separated dimer (SD) evolutionary feature extraction methods. The results features are scored by Information gain (IG) to define and select several discriminated features. According to three benchmark datasets, DD, RDD and EDD, the results of the support vector machine (SVM) show more than 6% improvement in accuracy on these benchmark datasets. Keywords: Protein Fold Recognition, Feature fusion, Evolutionary method, IG, Support Vector Machine 1. Introduction Proteins are Jack of all trades biological macromolecules. They are involved in almost every biological reaction; Pro- tein plays a critical roll in many different areas such as building muscle, hormone production, enzyme, immune5 function, and energy. Typically more than 20,000 proteins exist in human cells, to acquire knowledge about the protein function and interactions, the prediction of protein structural classes is extremely useful [1]. Fold recognition is one of the funda-10 mental methods in protein structure and function predic- tion. Protein can be demonstrated as a chain of amino acids. Proteins with unique lengths and similarities are part of the same fold. They also have identical protein secondary15 structure in the same topology. Certainly, they have a regular origin [2]. One of the main steps which can be assumed as a vital stage for predicting protein fold is feature extrac- tion. Computational feature extraction methods are di-20 vided into syntactical, physiochemical and evolutionary methods. Syntactical methods pay attention only to the protein sequence, like composition and occurrence [3, 4]. Physiochemical methods consider some physical and chem- ical properties of protein sequences. Evolutionary meth-25 ∗Corresponding author Email addresses: msaleh.refahi@aut.ac.ir (Mohammad Saleh Refahi), mir-am@hotmail.com (A. Mir), j.nasiri@irandoc.ac.ir (Jalal A. Nasiri) ods extract features from Basic Local Alignment Search Tool(BLAST). When attempting to solve many biological problems it is obvious that a single data source might not be informa- tive, and combining several complementary biological data30 sources will lead to a more accurate result. When we study methods of protein fold recognition, we found that less at- tention has been paid to the fusion of features to get more comprehensive features. In recent studies, researchers at- tempted to find new feature extraction methods[[5, 6, 7, 8,35 9]] or train different classifiers to achieve high accuracy[[10, 11, 12, 13]], even though some problems like incomplete data sources, false positive information, multiple aspect problem,. . . encourage us to combine data sources. Hence, to prepare more informative and discriminative40 features, we use Auto-Cross-Covariance(ACC)[8] and Sep- arated dimer(SD)[7] methods. Because SD explores some amino acid dimers that may be non-adjacent in sequence [7] and ACC method measures the correlation between the same and different properties of amino acids [8]. One of the45 main advantages of ACC and SD is to find a fixed length vector from a variable protein length. The performance of the proposed method is evaluated using three benchmark datasets DD[3] , RDD[14] and EDD[8]. In this paper, we focus on fusing ACC and SD feature50 extraction methods based on Position Specific Scoring Ma- trix(PSSM) generated by using the Position-Specific Iter- ated BLAST(PSI-BLAST) profile to predict protein fold. The 1600 ACC features and the 400 SD features are extracted based on the PSSM. Finally, we construct a55 reduced-dimensional feature vector for the Support Vec- Preprint submitted to Journal of Theoretical Biology August 27, 2019 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ tor Machine (SVM) classifier by using the Information Gain(IG). The remaining sections of the paper are organized as follows. Section 2 describes the related works of the exist-60 ing techniques. The methodology is explained in Section 3. Section 4 shows the experimental results and discus- sion. Finally, the conclusion and future works are given in Section5. 2. Related work65 In 1997, Dubchak et al. studied syntactical and phys- iochemical method [15]. In which they assumed five prop- erties of amino acid like hydrophobicity (H), frequency of α helix (X), polarity (P), polarizability (Z) and van der Waals volume (V). In [16] Forward Consecutive Search70 (FCS) scheme which trains physiochemical attributes for protein fold recognition. Another solution to find similarity between protein se- quences is based on the BLAST. Many feature extraction methods use BLAST alignment to extract the possibil-75 ity of amino acid in specific positions called as PSSM. In 2009, pairwise frequencies of amino acids separated by one residue (PF1) and pairwise frequencies of adjacent amino acid residues (PF2) were proposed by Ghatny and Pal [5]. The bigram feature extraction method was in-80 troduced by Sharma et al. [6], bigram feature vector is computed by counting the bigram frequencies of occur- rence from PSSM. In 2011, combination of PSSM with Auto Covariance (AC) transformation, was introduced as feature extraction method [17].85 Another method introduced by Saini et al. [7] is sep- arated dimers(SD); they used probabilistic expressions of amino acid dimer occurrence that have varying degrees of spatial separation in the protein sequence to predict pro- tein fold. Dong et al. [8] proposed autocross-covariance90 (ACC) transformation for protein fold recognition. More- over, Pailwal et al. [9] proposed the ability of trigram to extract features from the neighborhood information of amino acid. In addition to the feature extraction methods, some95 researchers have paid attention to classification methods for protein fold recognition. In [10] Kohonens selforgani- zation neural network is used and showed the structural class of protein is considerably correlated with its amino acid composition features.100 Baldi et al.[18] employed Recurrent and Recursive Ar- tificial Neural Networks (RNNs) and mixed it by directed acyclic graphs (DAGs) to predict protein structure. In [12], classwise optimized feature sets are used. SVM classifiers are coupled with probability estimates to make105 the final prediction. Linear discriminant analysis(LDA) was employed to evaluate the contribution of sequence parameters in determining the protein structural class. Parameters were used as inputs of the artificial neural networks[19]. The composition entropy was proposed to110 represent apoptosis protein sequences, and an ensemble classifier FKNN (fuzzy K-nearest neighbor) was used as a predictor[13]. TAXFOLD [20]method extract sequence evolution fea- tures from PSI-BLAST profiles and the secondary struc-115 ture features from PSIPRED profiles, and finally a set of 137 features is constructed to predict protein folds. Sequence-Based Prediction of ProteinPeptide(SPRINT) method is used to the prediction of Proteinpeptide Residue- level Interactions by SVM [11]. SVM implements the struc-120 tural risk minimization (SRM) that minimizes the upper bound of generation error [21, 22]. The DeepCov method proposed in [23], this method uses convolutional neural networks to work on amino acid pair frequency or covariance data extract from sequence125 alignments. In [24] is attempted to show Artificial Neural Network (ANN) with different feature extraction method is more accurate than other classifier methods. 3. Methodology130 This section illustrates the step-by-step of the pro- posed method for protein fold recognition. In the first step, sequence alignments are found for each protein using BLAST. To show improvements in protein fold recogni- tion using evolutionary information that are presented in135 PSSM, therefore ACC [8]and SD [7] features are extracted from PSSM. In the next step, the features are combined and selected by the IG. In the last step, the SVM algo- rithm is trained to classify proteins. A comprehensive view of this approach can be found in Figure1.140 3.1. Preprocessing 3.1.1. BLAST Similarity is used here to mention the resemblance or percentage of identity between two protein sequences [25]. The similarity search depends on the bioinformatics al-145 gorithm. Basic Local Alignment Search Tool(BLAST) is a tool that helps researchers to compare a query sequence with a database of sequences and identify specific sequences that resemble the query sequence above a certain thresh- old. BLAST is a local alignment algorithm that means150 to find the region (or regions) of the highest similarity between two sequences and build the alignment outward from there [26]. 3.1.2. PSSM Position Specific Scoring Matrix(PSSM) is used to ex-155 press motif in a protein sequence. P-BLAST searches in which amino acid substitution scores are given separately for each position in a protein multiple sequence alignment. In this paper, PSSM is used to extract features by ACC and SD methods.160 2 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ Figure 1: The flowchart of the proposed method pipeline. Sequence alignments are found for each protein by BLAST. PSSM is calculated to extract the ACC and the SD.The features are selected by the IG. The SVM algorithm is trained to classify proteins. 3.2. Feature Extraction 3.2.1. ACC ACC fold [8] utilizes autocross-covariance transforma- tion that convert the PSSMs of different lengths into fixed- length vectors. The ACC separates two kinds of features:165 AC between the same properties, cross-covariance (CC) between two different properties. The AC variable mea- sures the correlation of the same property between two properties separated by LG, distance along the sequence: AC(i,LG) = L−LG∑ j=1 (Pi,j−Pi)(Pi,j+LG−Pi)\(L−LG) (1) Where Pi,j is the PSSM score of amino acid i at position170 j,and Pi = ∑L j=1 Pi,j \ L, the average score of an amino acid i in the total protein sequence. The number of fea- tures which are calculated from AC is 20 ∗ LG. The CC measures the correlation of two different properties be- tween the distances of LG along the sequence:175 CC(i1, i2,LG) = L−LG∑ j=1 (Pi1,j−Pi1)(Pi2,j+LG−Pi2)\(L−LG) (2) The CC variables are not symmetric. The total number of CC variables is 380 ∗LG.The combination of AC and CC features make 400 ∗LG feature vectors. 3.2.2. SD Separated Dimer(SD) method was introduced by Saini et al.[7]. It attempts to extract features from amino acids that may or may not be adjacent in the protein sequence. The SD demonstrates the probabilities of the occurrence of amino acid. SD generates 400 features. F(k) = [F1,1(k),F1,2(k), ....,F20,19(k),F20,20(k)] (3) F(k) is computed as the feature sets for probabilistic oc-180 currence of amino acid dimers with different values of k which is a special distance between dimers. It is obvious if P is in the PSSM matrix for a protein sequences, it is L×20 matrix where L is the length of the protein sequence: Fm,n(k) = L−k∑ i=0 Pi,mPi+k,n (4) in which m,n (1 ≤ m,n ≤ 20) is the score of two185 selective amino acid in PSSM. 3 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ Figure 2: The AC features of the ACC, measures the corre- lation of the same property between two properties separated by a distance of LG along the sequence 3.3. Fusion hypothesis More attention needs to be paid to find an efficient technique for integrate distinct data sources for the pro- tein fold recognition problem[27]. Various techniques have190 been employed based on the features which are extracted from protein sequences. These techniques investigate different aspects of a se- quence like the study of possible position of amino acids, protein chemical characteristics and syntactical features195 . . . . Hence, integrating them can model the folding prob- lem more accurate. In this study three hypotheses have been considered for fusion data sources. The first, only evolutionary features are used since integrating different types of features may200 have an undesirable effect on each other. The next assumption is considered choosing the ACC and SD methods. When we studied the recent paper, we observed the recall and precision of some protein folds which were almost equal to high value or one was a comple-205 ment of the other. Hence, when the recall of the ACC(SD) is low, then the recall of the SD(ACC) is high, and also for the precision, we observe this behavior, in almost every fold. The last hypothesis is that the ACC and the SD fea-210 tures showed a relationship between amino acids which may or may not be adjacent. In this approach, three dif- ferent characters are defined which show each amino acid in a specific position what relation has with others. These characters are shown in Figures234.215 3.4. Information Gain Feature selection is a common stage in classification problems. It can improve the prediction accuracy of clas- sifiers by identifying relevant features. Moreover, feature selection often reduces the training time of a classifier by220 reducing the number of features which are going to be an- alyzed. Figure 3: The CC features of the ACC, measures the corre- lation of two different properties between the distances of LG along the sequence Figure 4: The SD consist of aminoacid dimers with proba- bilistic expressions that have k separation. Information gain (IG) is a popular feature selection method. It ranks features by considering their presence and absence in each class [28]. The IG method gives a225 high score to the features that occur frequently in a class and rarely in other classes. Given T the set of training samples, xi the vector of ith variables in this set, |Txi=v|/|Tx| the fraction of samples of the ith variable having value v. The IG method can be230 computed as follows: IG(Tx,xi) = H(Tx) − |Txi=v| |Tx|∑ v=values(xi) H(Txi=v) H(T) = −p+(T)log2p+(T) −p−(T)log2p−(T) (5) where p± denotes the probability of a sample in the set T to be of the positive or negative class. 3.5. Support Vector Machine Support Vector Machine (SVM) was proposed by Vap- nik and Cortes in 1995 [29]. It is a powerful tool for bi- 4 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ nary classification. SVM is on the basis of Structural Risk Minimization (SRM) and Vapnik-Chervonenkis (VC) di- mension. The central idea of SVM is to find the optimal separating hyperplane with the largest margin between the classes. Due to the SRM principle, SVM has great gener- alization ability. Moreover, the parameters of the optimal separating hyperplane can be obtained by solving a convex quadratic programming problem (QPP), which is defined as follows: min w 1 2 ‖w‖2 + C n∑ i=1 ξi s.t. yi(w Txi + b) + ξi ≥ 1,∀i (6) where ξ is the slack variable associated with xi sample235 and C is a penalty parameter. Note that the optimiza- tion problem can be solved when the classification task is linearly separable. In the case of nonlinear problems, the input data is transformed into a higher-dimensional feature space in order to make data linearly separable. It240 makes possible to find a nonlinear decision boundary with- out computing the parameters of the optimal hyperplane in a high dimensional feature space [30]. As mentioned in this subsection, SVM is designed to solve binary classification problems. However, there are245 multi-class approaches such as One-vs-One (OVO) and One-vs-All (OVA) [31], which can be used for solving multi- class classification problems. In this paper, we used OVO strategy. 4. Experimental Result250 4.1. Dataset Three popular datasets which were employed in this study are DD dataset [3], EDD dataset [8], and RDD dataset[14]. DD dataset contains 27 folds which repre- sent four major structure classes:α,β,α β ,α + β. The train-255 ing set and the testing set contains 311 training sequences and 383 testing sequences whose sequence similarity is less than 35%[3]. The EDD dataset consists of 3418 proteins with less than 40% sequential similarity belonging to the 27 folds that originally are adopted from the DD dataset.260 The RDD dataset consists of 311 protein sequences in the training and 380 protein sequences in testing datasets with a similarity lower than 37% [14]. 4.2. Result The experiments were performed on the benchmark265 datasets to evaluate the performance of the classification due to our fusion method. we also adopted the 10-fold cross-validation in this study, which has done by many researchers to examine predictive potency. In this study LibSVM [32] with RBF (Radial Basis270 Function) as the kernel functions has been used. The C parameter was optimized by search between {2−14, 2−13, . . . , 213, 214} and also Γ parameter of RBF was considered between {2−14, 2−13, . . . , 213, 214}. The SVM was origi- nally designed for binary data classification.This study275 used one-versus-one method to approach a multi-class clas- sifier. The details of the feature extraction method are ex- plained in methodology, but it is important to know how far is assumed between aminoacids, for each ACC and SD280 methods. In developing the algorithm to extract features from PSSM ,LG and k parameters have been assumed like ACC and SD papers values[7, 8]. We considered both LG and k equals to 4. The IG [28] makes our method safe from noisy fea-285 tures. In this approach, we considered the features ranked between [ 1 2 maxIG, maxIG] for each dataset. The results of IG for each dataset are exhibited in Table2. 4.3. Discussion Table1 illustrates the total prediction accuracies of the290 existing approaches for classification of protein folds in the DD,RDD and EDD datasets. Table1 also shows the suc- cess rates of our proposed fusion approach. According to Table1, classification results of the combined ACC and SD followed by selection of best features by IG show consid-295 erable improvement compared to the state of art. Figure9 has been shown to figure out the result dis- tribution of feature selection method. Even though the number of ACC in the three datasets is more but all of the SD features exist in selected features. However, we300 studied and compared SD and ACC methods separately, we found that the fusion of them can make more informa- tive data which cover all characteristics of folds. It is evident in Figure6, Figure7, and also Figure8, only ”FAD-BINDING MOTIF” protein fold is not well recog-305 nized, and these confusion matrices show the power of proposed method for predicting the other folds in these datasets. The Figure5 has been shown to evaluate the IG. It is obvious that maximum accuracy of classification for each310 dataset has been achieved when we consider ranking fea- tures higher than 1 2 maxIG for these datasets. Sensitivity measures the ratio of correctly classified samples to the whole number of test samples for each class which is classified as correct samples and calculated as fol-315 lows: Sensitivity = TP TP + FN × 100 (7) TP represents true positive and FN represents false nega- tive samples. Precision represents, how relevant the num- ber of TP is to the whole number of positive prediction and is calculated as follows: Precision = TP TP + FP × 100 (8) FP denotes false positive. F1 Score is the weighted average of Precision and Recall. F1 score, as other evaluation cri- teria which are used in this study measures, is calculated 5 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ Figure 5: Comparison of number of features and accuracy for DD,RDD and EDD datasets to evaluate the IG method Figure 9: Comparison of the ACC and the SD in DD,RDD and EDD datasets. as follows: F1score = 2TP 2TP + FP + FN × 100 (9) The sensitivity, precision, and F1 score are computed for each class and then averaged over all the classes which are calculated and published in Table2. 5. Conclusion320 This study aims to improve protein fold recognition accuracy by fusing information that are extracted from the PSSM matrix. In this approach, we used ACC and SD feature extraction methods. It was observed that the proposed technique eventuates to 6% improvement for the325 accuracy of these three benchmark datasets. In the future, classification can be done by combining more syntactical,physiochemical or evolutionary features. Table 1: Comparison of the proposed method with the exist- ing predictor and Meta-predictors for the DD, RDD and EDD. Methods Reference DD RDD EDD occurrence [4] 42 56.6 70.0 ACC [8] 68.0 73.8 85.9 PF1 [5] 50.6 53.3 63.0 PF2 [5] 48.2 NA 49.9 TAXFOLD [20] 71.5 83.2 NA Bigram [6] 79.3 59.6 79.9 SD [7] 86.3 72.1 90.0 Trigram [7] 73.4 60.0 80.0 MF-SRC [33] 78.6 NA 86.2 Enhanced-SD [24] 90.0 75.4 93.0* Proposed Method - 91.31 91.64 91.2 * The evaluation method not defined in [24] approach. Table 2: F1 score, Sensitivity and Precision, Measurement tools to evaluate the proposed method. Data set F1 score Sensitivity Precision Number of features DD 0.98 0.92 0.93 1300 RDD 0.98 0.92 0.93 1416 EDD 0.96 0.91 0.93 900 To achieve more accuracy, future studies should be concen- trate on ”FAD-BINDING MOTIF” protein fold that has330 less discriminative features in the SD and the ACC. Boost- ing classifier may be employed to find better solutions for protein fold recognition. 6 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ Figure 6: Confusion matrix of DD dataset(91.31%) 7 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ Figure 7: Confusion matrix of RDD dataset(91.64%) 8 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ Figure 8: Confusion matrix of EDD dataset(91.2%) 9 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ References [1] J.-Y. Yang, Z.-L. Peng, X. Chen, Prediction of protein struc-335 tural classes for low-homology sequences based on predicted sec- ondary structure, BMC bioinformatics 11 (1) (2010) S9. [2] T. Yang, V. Kecman, L. Cao, C. Zhang, J. Z. Huang, Margin- based ensemble classifier for protein fold recognition, Expert Systems with Applications 38 (10) (2011) 12348–12355.340 [3] C. H. Ding, I. Dubchak, Multi-class protein fold recognition using support vector machines and neural networks, Bioinfor- matics 17 (4) (2001) 349–358. [4] Y. Taguchi, M. M. Gromiha, Application of amino acid occur- rence for discriminating different folding types of globular pro-345 teins, BMC bioinformatics 8 (1) (2007) 404. [5] P. Ghanty, N. R. Pal, Prediction of protein folds: extraction of new features, dimensionality reduction, and fusion of hetero- geneous classifiers, IEEE transactions on nanobioscience 8 (1) (2009) 100–110.350 [6] A. Sharma, J. Lyons, A. Dehzangi, K. K. Paliwal, A feature extraction technique using bi-gram probabilities of position spe- cific scoring matrix for protein fold recognition, Journal of the- oretical biology 320 (2013) 41–46. [7] H. Saini, G. Raicar, A. Sharma, S. Lal, A. Dehzangi, J. Lyons,355 K. K. Paliwal, S. Imoto, S. Miyano, Probabilistic expression of spatially varied amino acid dimers into general form of chou’s pseudo amino acid composition for protein fold recognition, Journal of theoretical biology 380 (2015) 291–298. [8] Q. Dong, S. Zhou, J. Guan, A new taxonomy-based protein360 fold recognition approach based on autocross-covariance trans- formation, Bioinformatics 25 (20) (2009) 2655–2662. [9] K. K. Paliwal, A. Sharma, J. Lyons, A. Dehzangi, A tri-gram based feature extraction technique using linear probabilities of position specific scoring matrix for protein fold recognition,365 IEEE transactions on nanobioscience 13 (1) (2014) 44–50. [10] Y.-D. Cai, X.-J. Liu, X.-b. Xu, K.-C. Chou, Prediction of pro- tein structural classes by support vector machines, Computers & chemistry 26 (3) (2002) 293–296. [11] G. Taherzadeh, Y. Yang, T. Zhang, A. W.-C. Liew, Y. Zhou,370 Sequence-based prediction of protein–peptide binding sites us- ing support vector machine, Journal of computational chemistry 37 (13) (2016) 1223–1229. [12] A. Anand, G. Pugalenthi, P. Suganthan, Predicting protein structural class by svm with class-wise optimized features and375 decision probabilities, Journal of theoretical biology 253 (2) (2008) 375–380. [13] Y.-S. Ding, T.-L. Zhang, Using chous pseudo amino acid com- position to predict subcellular localization of apoptosis proteins: an approach with immune genetic algorithm-based ensemble380 classifier, Pattern Recognition Letters 29 (13) (2008) 1887–1892. [14] J. Xia, Z. Peng, D. Qi, H. Mu, J. Yang, An ensemble approach to protein fold classification by integration of template-based as- signment and support vector machine classifier, Bioinformatics 33 (6) (2016) 863–870.385 [15] I. Dubchak, I. B. Muchnik, S.-H. Kim, Protein folding class predictor for scop: approach based on global descriptors., in: Ismb, 1997, pp. 104–107. [16] G. Raicar, H. Saini, A. Dehzangi, S. Lal, A. Sharma, Improving protein fold recognition and structural class prediction accura-390 cies using physicochemical properties of amino acids, Journal of theoretical biology 402 (2016) 117–128. [17] T. Liu, X. Geng, X. Zheng, R. Li, J. Wang, Accurate prediction of protein structural class using auto covariance transformation of psi-blast profiles, Amino acids 42 (6) (2012) 2243–2249.395 [18] P. Baldi, G. Pollastri, The principled design of large-scale re- cursive neural network architectures–dag-rnns and the protein structure prediction problem, Journal of Machine Learning Re- search 4 (Sep) (2003) 575–602. [19] S. Jahandideh, P. Abdolmaleki, M. Jahandideh, E. B. Asad-400 abadi, Novel two-stage hybrid neural discriminant model for predicting proteins structural classes, Biophysical chemistry 128 (1) (2007) 87–93. [20] J.-Y. Yang, X. Chen, Improving taxonomy-based protein fold recognition by using global and local features, Proteins: Struc-405 ture, Function, and Bioinformatics 79 (7) (2011) 2053–2064. [21] M. S. Refahi, J. A. Nasiri, S. Ahadi, Ecg arrhythmia classi- fication using least squares twin support vector machines, in: Electrical Engineering (ICEE), Iranian Conference on, IEEE, 2018, pp. 1619–1623.410 [22] M. Rahmanimanesh, J. A. Nasiri, S. Jalili, N. M. Charkari, Adaptive three-phase support vector data description, Pattern Analysis and Applications 22 (2) (2019) 491–504. [23] D. T. Jones, S. M. Kandathil, High precision in protein contact prediction using fully convolutional neural networks and mini-415 mal sequence features, Bioinformatics 34 (19) (2018) 3308–3315. [24] P. Sudha, D. Ramyachitra, P. Manikandan, Enhanced artificial neural network for protein fold recognition and structural class prediction, Gene Reports 12 (2018) 261–275. [25] Blast and multiple sequence alignment (msa) programs,420 https://viralzone.expasy.org/e_learning/alignments/ description.html, accessed: 2019-01-17. [26] S. F. Altschul, W. Gish, W. Miller, E. W. Myers, D. J. Lipman, Basic local alignment search tool, Journal of molecular biology 215 (3) (1990) 403–410.425 [27] P. Zakeri, J. Simm, A. Arany, S. ElShal, Y. Moreau, Gene prioritization using bayesian matrix factorization with genomic and phenotypic side information, Bioinformatics 34 (13) (2018) i447–i456. [28] Q. Zou, J. Zeng, L. Cao, R. Ji, A novel features ranking met-430 ric with application to scalable visual and bioinformatics data classification, Neurocomputing 173 (2016) 346–354. [29] C. Cortes, V. Vapnik, Support-vector networks, Machine learn- ing 20 (3) (1995) 273–297. [30] B. Schölkopf, A. J. Smola, F. Bach, et al., Learning with ker-435 nels: support vector machines, regularization, optimization, and beyond, MIT press, 2002. [31] C.-W. Hsu, C.-J. Lin, A comparison of methods for multiclass support vector machines, IEEE transactions on Neural Net- works 13 (2) (2002) 415–425.440 [32] C.-C. Chang, C.-J. Lin, Libsvm: A library for support vector machines, ACM transactions on intelligent systems and tech- nology (TIST) 2 (3) (2011) 27. [33] K. Yan, Y. Xu, X. Fang, C. Zheng, B. Liu, Protein fold recog- nition based on sparse representation based classification, Arti-445 ficial intelligence in medicine 79 (2017) 1–8. 10 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted November 23, 2019. ; https://doi.org/10.1101/845727doi: bioRxiv preprint https://viralzone.expasy.org/e_learning/alignments/description.html https://viralzone.expasy.org/e_learning/alignments/description.html https://viralzone.expasy.org/e_learning/alignments/description.html https://doi.org/10.1101/845727 http://creativecommons.org/licenses/by-nc-nd/4.0/ Introduction Related work Methodology Preprocessing BLAST PSSM Feature Extraction ACC SD Fusion hypothesis Information Gain Support Vector Machine Experimental Result Dataset Result Discussion Conclusion 
work_xwpv56g5gbe5hg3awggpwntadi	----	An intelligent driver location system for smart parking Expert Systems with Applications xxx (2013) xxx–xxx Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a An intelligent driver location system for smart parking 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.09.044 ⇑ Corresponding author. Tel.: +886 35712121. E-mail addresses: klan@csie.ncku.edu.tw (K.-C. Lan), Todd629.cs01g@nctu. edu.tw (W.-Y. Shih). Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver location system for smart parking. Expert Systems with Applications http://dx.doi.org/10.1016/j.eswa.2013.09.044 Kun-Chan Lan a, Wen-Yuah Shih b,⇑ a Department of Computer Science and Information Engineering, National Cheng Kung University, No. 1, University Road, Tainan City 701, Taiwan, ROC b Department of Computer Science, National Chiao Tung University, No. 1001, University Road, Hsinchu 30010, Taiwan, ROC a r t i c l e i n f o a b s t r a c t Keywords: Pedestrian dead reckoning Simple harmonic motion ZUPT Map matching Floor plan Smart parking It is often frustrating for drivers to find parking spaces, and parking itself is costly in almost every major city in the world. Here we propose a crowdsourcing solution by exploiting sensors in smart-phones to collect real-time parking availability information. We design a phone-based system to track a driver’s trajectory to detect when they are about to leave their parking spot. We focus on the efficiency and accu- racy of using a phone to monitor the driver’s walking trajectory, applying a waist-mounted PDR method that can measure the driver’s moving distance with a high accuracy. In addition, we design a map match- ing algorithm to calibrate the direction errors when the driver is in an indoor environment, using widely- available building floor plans. The results of our experiment show that we can achieve about 98% accuracy in estimating the user’s walking distance, with an overall location error of about 0.48 m. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction Searching for street parking in crowded urban areas creates many problems and frustrations for drivers. It has been shown that over 40% of the total traffic volume in urban areas is composed of vehicles cruising for parking (Shoup, 2006). A long queue of cruis- ing vehicles can cause serious congestion with the blocking of only a few streets. In addition, low speed cruising can produce signifi- cant amounts of automobile emissions (Arnott & Inci, 2006), increasing air pollution. A prior study (Mathur et al., 2010) found that in one area of Los Angeles vehicles searching for parking pro- duced 730 tons of carbon dioxide, and burned 47,000 gallons of gasoline (Ayala, Wolfson, Xu, Lin, & Dasgupta, 2011) over one year. In this work, we propose a solution that utilizes the sensors (such as a GPS, accelerometer, gyroscope, and digital compass) in a smart-phone to detect the driver’s parking/un-parking activities. Such information can then be broadcast (e.g. through the Internet) to people who are trying to find a parking space. To detect parking, a prior work has shown that it is possible to detect the driver’s transportation mode (e.g., driving, stationary, and walking) (Stenneth, Wolfson, Yu, & Xu, 2011) using the sensors in a smart- phone. For example, if we detect a transition pattern like driv- ing ? stationary ? walking, we may conclude the car has been parked at the stationary point. In this paper, we focus on how to detect the unparking activity by tracking the walking trajectory of the driver using the smart-phone’s sensors. The idea is a simple one. If the phone detects that the driver is approaching the place where they parked their car, it is likely the driver is about to leave the area and the parking space will become available very soon. In this paper, we consider a social network formed by drivers, similar to the Crowdpark (Yan et al., 2011) platform. In our archi- tecture, a driver who is currently parked can provide advance notification about when they plan to leave, and this information may be sold to another driver who is willing to pay (via a virtual currency (Lan & Wang, 2013), such as BitCoin) to reserve the park- ing spot. The buyer arrives at the reserved parking spot close to the leaving time of the seller, and can occupy the spot when the seller leaves. Since the drivers transact only parking availability informa- tion, our paradigm presents a loose reservation model for the park- ing spot, and the buyer is charged only when they successfully park their car in the focal location. Our system provides an incentive for sellers to contribute their leaving information, and also encourages the buyers to re-sell the parking spaces when they leave. Each participant in this system is assigned a random unique ID when joining the social network, and thus their real identity will not be revealed to the other users. While crowdsourcing-based approaches to parking are not new (Yan et al., 2011), but, as far as we know, all of the prior works in this area (such as Google OpenSpot (OpenSpot) require the drivers to manually report when they leave their parking spots. In contrast, we utilize the sensors in the user’s smart-phone to auto- matically infer the parking space’s availability in advance by track- ing the trajectory of the driver. How to accurately and efficiently monitor the walking trajectory of the driver is thus the key issue in our method, and the focus of this paper. Tracking the walking trajectories of people in an environment has long been considered an important component of ubiquitous networking. Generally, in an outdoor environment, the driver’s (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 mailto:klan@csie.ncku.edu.tw mailto:Todd629.cs01g@nctu. edu.tw mailto:Todd629.cs01g@nctu. edu.tw http://dx.doi.org/10.1016/j.eswa.2013.09.044 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa http://dx.doi.org/10.1016/j.eswa.2013.09.044 Fig. 2. The entire system architecture. 2 K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx walking trajectory can be obtained through GPS data, although it is more challenging to localize the driver in an indoor environment. In this paper, we consider the use of a personal dead reckoning or pedestrian dead reckoning (PDR) system to localize the driver, particularly when they are in a building. The PDR technique only requires a couple of inertial sensors to be put on the user, so that it can be used in any building without pre-installing beacon nodes or pre-building RF maps/propagation models based on surveys of the environment. These inertial sensors (such as an accelerometer, gyroscope, or digital compass, which smart-phones usually have) are used to measure step length and heading direction. Generally, depending on where the sensors are placed, previous studies classified PDR systems into two types: foot- and waist- mounted. The foot-mounted method (Alvarez, Gonzalez, Alvarez, Lopez, & Rodriguez-Uria, 2007; Dippold, 2006; Eric, 2005; Feliz, Zalama, & García-Bermejo, 2009; Jimenez, Seco, Prieto, & Guevara, 2009; Ojeda & Borenstein, 2006; Ojeda & Borenstein, 2007a; Ojeda & Borenstein, 2007b; Oliver & Robert, 2008; Sagawa, Inooka, & Satoh, 2000; Xiaoping, Bachmann, Moore, & Calusdian, 2007) uses a double integral on horizontal acceleration to estimate distance, and a gyroscope or compass to measure the heading direction. On the other hand, the waist-mounted (Alvarez, Gonzalez, Lopez, & Alvarez, 2006; Lei et al., 2005; Shin, Park, Hong, & Lee, 2005; Weinberg, 2002) method tries to detect each step event to calculate the total number of steps, and then multiplies this by a constant step length which is based on the pedestrian’s character- istics (weight, height, and age) to estimate the total moving dis- tance. Some waist-mounted methods use linear regression to find the relationship between the acceleration, walking speed, and step length (Shin, Park, Kim, Hong, & Lee, 2007). But when the pedestrian’s walking pattern is different from the predefined step length model, the accuracy in distance estimation could be ad- versely affected. Finally, although a waist-mounted method is gen- erally more feasible to be implemented on a hand-held device, its accuracy in estimating the step length is typically worse than that of a foot-mounted method which, in contrast, performs poorly with regard to obtaining an accurate orientation (Stirling, Collin, Fyfe, & Lachapelle, 2003). Sensor drift (Titterton & Weston, 1997) is a well-known prob- lem in PDR systems. Given that the hardware used in such systems is not perfect, the inertial sensors constantly have some small er- rors when estimating the distance and direction, and signal noise Fig. 1. (a) A detailed map Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 (such as the vibrations of the user’s body) can further exacerbate this problem (Feliz et al., 2009; Sagawa et al., 2000). In an outdoor environment, sensor drift can be mitigated through the use of GPS measurements (Godha, Lachapelle, & Cannon, 2006; Ladetto & Merminod, 2002), while in an indoor environment some PDR sys- tems use a map matching mechanism to calibrate these errors (Ascher, Kessler, Wankerl, & Trommer, 2010; Gusenbauer, Isert, & Krösche, 2010; Ishikawa, Kourogi, Okuma, & Kurata, 2009), and these can be categorized into two types. One tries to matches the user trajectory to the closest junction and road on the map (Gusenbauer et al., 2010), while the other one utilizes the map to filter out positions where the user is unlikely to move (e.g., walls, obstacles, and so on) (Ascher et al., 2010; Ishikawa, 2009). Both techniques require the use of a detailed scaled map of the building, as shown in Fig. 1(a), although in practice this is usually difficult to obtain. In this paper, we consider a scenario in which the user has a smart-phone and can access the floor plan of the building, like the one shown in Fig. 1(b), as these are widely available. The sys- tem utilizes the sensors on the smart-phone to compute the user’s moving distance and direction. Combining the walking trajectory with the map allows the system to estimate the driver’s current position in a building. The system architecture is shown in Fig. 2. Our system can also be implemented as an add-onto a traditional outdoor navigation app (e.g., Google Latitude, although this service has now been discontinued) so that the starting position of the user in the building can be estimated using the last-recorded GPS position. We make the following assumptions in this paper. and (b) a floor plan. location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx 3 1. The floor plan can be characterized using a link-node model (Gillieron & Merminod, 2003), in which pathways are the links and the intersections of pathways are the nodes. Note that the limitation of a link-node model is that it does not consider open indoor spaces (e.g., a big hotel lobby), and we leave this for future work. 2. The phone is placed inside the pocket of the user, which is often the case in real life (however, some prior works, such as (Mohan, Padmanabhan, & Ramjee,2008), discussed the situation of when the phone is held in the hand). 3. We assume the initial heading direction is known. This can be achieved by having the user point the phone in the direction they are heading before putting the phone in their pocket. 4. During walking, the position of the phone is relatively stable in relation to the leg movements. 5. In this work, we do not consider the multi-floor building sce- nario, since most of smart-phones do not have a barometer to detect the change to a different floor, and we leave this as a fur- ther direction for future work. The contribution of this paper is threefold. First, unlike previous crowdsourcing-based methods that required the drivers to manu- ally report when they leave their parking spots, we utilize the phone sensors to automatically infer the parking space availability and notify the other drivers in advance by tracking the trajectory of the driver. Second, we implement a waist-mounted-based PDR method on the smart-phone to accurately estimate the driver’s moving distance with no need for a training phase. The accuracy of our distance estimation is about 98%. Finally, when the driver is indoors, based on the geometric similarity between the driver’s trajectory and the floor map, we design a map-matching algorithm to calibrate sensor errors using building floor plans, which are widely available. We find that the location error is about 0.48 m in our test scenarios. The rest of this paper is structured as follows: In Section 2, we describe the related work. We discuss the details of our step-length estimation and map-matching algorithms in Sections 3 and 4, respectively. The results of our experiment are shown in Section 5. Finally, we conclude this paper in Section 6. 2. Related work Our work is built on some previous studies of parking guidance, PDR, step length estimation, and map matching, and these are out- lined below. 2.1. Parking-guidance Online real-time parking information systems have recently at- tracted significant interest in both industry and academia. They can generally be classified into infrastructure-based and crowd- sourcing-based approaches. The infrastructure-based methods, such as SFpark and ParkNet, require the installation of occupancy sensors and wireless transceivers on the parking spot and/or vehi- cle. For example, ParkNet installs ultra-sonic sensors on vehicles to detect parking availability when vehicles drive by. However, these approaches are generally costly and not scalable. On the other hand, a number of mobile crowdsourcing applications have been developed that allow drivers to share empty parking spot informa- tion, such as OpenSpot, Roadify, Primospot, and SpotScout. These work by having the drivers report when they are leaving their parking spots. That data gets fed into the applications, and then drivers can see where people have recently left an open parking spot. However, such data could have a very short lifespan, since empty spots in crowded areas are quickly used. In other words, Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 by the time when the driver figures out how to get to an open spot, it is likely to have already been taken. In contrast to these prior works, we utilize the sensors on the phone to automatically infer parking space availability by tracking the trajectory of the driver, which requires no additional infrastructure or man-power support, and is more practical for deployment in the real world. Furthermore, our system provides an incentive for the driver to offer parking availability information in advance to others, by sell- ing it, so that the potential buyer can arrive at the parking spot close to the leaving time of the seller. To illustrate the usefulness of such advance parking availability notification, we performed the following experiments in an urban area. We had two cars (A and B) travel around in the area. We performed two sets of experiments using our system and Open- Spot. Each set of experiments was executed for three days. The idea here is to have car A update its parking information online, and then observe how many times car B can successfully occupy the parking spot that car A has previously taken. We found the success rate for car B to arrive at car A’s spot in time was 87% using our application, and only 59% using OpenSpot. 2.2. PDR system Some prior studies use a PDR system to estimate the trajectory of a user by placing some sensors on the body. Inertial sensors, such as accelerometers, gyroscopes, and compasses, are commonly used in such systems. Some PDR systems also include GPS sensors (Godha et al., 2006; Ladetto & Merminod, 2002), and use these to calibrate the PDR drift as long as a GPS signal is available. When the GPS signal is obstructed, the system can then be changed to the PDR mode and continue to record the trajectory. Our study is based on the PDR system, which has the advantage of avoiding the deployment overhead of the signal-based methods. On the other hand, the performance of our system can be enhanced by a signal-based system if available. For example, one limitation of our map-matching approach is its need to collect ‘‘enough’’ trajec- tory data before it can uniquely identify the user’s position on the map. When there is not enough trajectory information, we make use of the known locations of existing WiFi base stations in the building to help estimate the user’s location (Ji, Biaz, Pandey, & Agrawal, 2006; Krishna, Anand Padmanabha, & Venkata, 2010; Lim, Kung, Hou, & Luo, 2006). 2.3. Step length estimation A PDR system can be classified into two types depending on where the sensor is mounted: foot-mounted (Alvarez et al., 2007; Dippold, 2006; Eric, 2005; Feliz et al., 2009; Jimenez et al., 2009; Ojeda & Borenstein, 2006; Ojeda & Borenstein, 2007a; Ojeda & Borenstein, 2007b; Oliver & Robert, 2008; Sagawa et al., 2000; Xiaoping et al., 2007) and waist-mounted (Alvarez et al., 2006; Lei et al., 2005; Shin et al., 2005; Weinberg, 2002). To calculate the step length, the foot-mounted methods typically perform a double integral on the horizontal acceleration. However, without further calibration, the problem of sensor drift (Titterton & Weston, 1997) could introduce serious inaccuracies when estimating the step length. One way to calibrate the sensor drift error is called zero velocity update (ZUPT). When the swinging-foot touches the ground, the angular velocity of this foot will be close to zero, which can be used to reset the system to avoid sensor drift errors accu- mulating into the length estimation of the next step. On the other hand, for a waist-mounted PDR system, ZUPT is not directly applicable, since one will not be able to find zero velocity in the horizontal direction. Some waist-mounted PDR sys- tems use a constant step length (Jorgensen, 2008; Martin & Michael, 2009; Schneider, Crouter, Lukajic, & Bassett, 2003; Shih, location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 Fig. 3. Walking diagram. 4 K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx 2010) while the others (Li et al., 2012; Shin et al., 2005) use a trace- driven approach, by first collecting empirical data from multiple users and then using linear regression to find the relation between step length, walking frequency, and the variance of acceleration. The limitation of this approach is that one might need to collect new training data for a new user. Weinberg (2002) observed that the upper body moves vertically when walking, and suggested that one can estimate the step length as follows StepLength ¼ 2 � heightchange=a where a is the swinging angle of the leg from the body, and the heightchange can be estimated based on the vertical acceleration. However, he did not discuss how to measure a. Our idea is similar to Weinberg’s, but we estimate the step length based on the height change and length of the leg using the Pythagorean Theorem. In addition, we use the concept of Simple Harmonic Motion (SHM) to find the zero velocity in the vertical direction, and apply ZUPT to avoid the accumulation of sensor drift errors on the vertical-axis. 2.4. Map-matching algorithm In some prior PDR systems, a map-matching mechanism is used to match the user trajectory onto the map (Ascher et al., 2010; Gusenbauer et al., 2010; Ishikawa, 2009) in order to calibrate the sensor errors. There are two ways this is achieved. The first tries to match the user trajectory to the closest junction and road on the map (Gusenbauer et al., 2010), while the other utilizes the map information to filter out positions where the user is unlikely to walk (e.g., walls and obstacles) (Ascher et al., 2010; Ishikawa, 2009). However, both techniques require the use of a detailed scaled map of the building (i.e., with detailed distance information for each route on the map), which is usually not easy to obtain. In our work, we utilize the more widely available building floor plans instead of detailed scaled maps. Based on the geometric similarity between the trajectory data and the map, we propose a new map- matching method that uses the floor plan to locate the user. As shown later in Section 5, our approach has similar performance to those methods that rely on detailed scaled map information. θ NgeHeightChan N GravityadingSensorN ∫ ∫= ÷×−= force.externalby causedison which accelerati theis cos)cosRe( θθ θ θθ NgeHeightChan N GravityadingSensorN ∫ ∫= ÷×−= force.externalby causedison which accelerati theis cos)cosRe( θθ θ Fig. 4. The orientation of the device. 3. Implement a waist-mounted PDR method using a smart- phone Weinberg (2002) proposed that one can estimate the step length using the height change of the waist and the swinging angle of the leg during walking (Weinberg, 2002). However, he did not discuss how to measure this angle, and we found that, in practice, it is difficult to measure such a small angle during walking. Based on Pythagorean Theorem, in our prior work Lan and Shih, (2012) we proposed a different way to estimate step length using the change in height. During walking, a person’s body moves up and down. If we assume the length of the leg is L, the waist-line will move up-and-down between L and (L � h) from the ground, where h is the change in the height of the waist. Considering the triangle in Fig. 3, formed by two feet of a person and their step length D. Given that L is known, using the Pythagorean Theorem, we can estimate D if we know the height of this triangle, i.e., (L � h). To ob- tain (L � h), we first need to calculate h, which is the change in height of the waist during walking. Therefore, if we mount an accelerometer on the user’s waist, the readings of this can be used to estimate the height change h, which can then be used to calcu- late the step length D based on the Pythagorean Theorem (the length of the leg (i.e. L) and half of the step length (i.e. D/2) forms a right triangle in which leg length is the hypotenuse). In this paper, we extend the above-mentioned waist-mounted method for a smart-phone. When implementing a PDR system on Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 a smart-phone, two cases can be considered. The first is when the user holds the phone (e.g., when talking on the phone), and the other is when the phone is put in a pocket or a bag. Prior re- search (Su, Chou, Yi, Tseng, & Tsai, 2010) has shown the feasibility of using a waist-mounted method for the first case. Therefore, in this section we focus on how to extend the results of a waist-mounted method for the second case. To implement a waist-mounted method, two issues need to be considered: the ori- entation and placement of the phone. The orientation of the phone may change from time to time when it is put in a pocket or bag during walking, and this affects the influence of gravity on the three axes of the accelerometer, and results in different readings from the sensor. To resolve this problem, we adopt a method similar to that in an earlier work, Su et al. (2010), which used a gyroscope to record changes in orientation and to calibrate the system, as shown in Fig. 4. Next, in a waist-mounted method, the vertical movement displacement of the phone is the key parameter to estimate the step length. When the phone is positioned above the waist-line, its vertical displacement during walking will be the same as when it is mounted on the waist. On the other hand, when the phone is placed below the waist-line (e.g., in a pants pocket), we can estimate the vertical displacement of the waist as follows. Assuming the leg length (L) and the pocket position from the ground (L’) are already known, we can find two similar triangles DABC and DPQR, as shown in Fig. 5. Based on the Pythag- orean Theorem, we can obtain QR by first measuring L’ and h’ (i.e., the vertical displacement of the pocket during the walking). Then, using triangle similarity, we can find BC ¼ D=2 ¼ðL � QRÞ=L0. h’ can be measured in a similar way as to measure h . To use the smart-phone to estimate the walking distance, the following practical issues need to be considered: 1. How to remove signal noise, such as vibrations from the body? 2. How to prevent sensor drift errors accumulating from one step to the next? 3. How to deal with when the user moves at irregular speeds? Our system architecture is shown in Fig. 6. We first filter the noise using a low-pass filter. The filtered signal is fed into the step recognition module to identify each new step. We adopt the concept of simple harmonic motion (SHM) to reset the vertical location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 h’ ( ) ( )22 can be estimated using the accelerometer. andknown,areandAssume hLLQR h ′−′−′= ′ ′ L L D If the device is put here L’ C L L’ L’-h’ D/2 A B P Q R A B C P Q R h’ h LL ′−′−= ′ A Fig. 5. When the device is mounted on the lower body. K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx 5 velocity at the beginning of each new step, which prevents sensor drift errors from being accumulated over to the next one. Finally, the step length is estimated based on the filtered sensor data and foot length. 3.1. Noise filtering During walking, some unexpected and unpredictable body vibrations might cause some higher-frequency noise in the sensor readings. In this section, we discuss how to filter such noise. Origina accelerat readings fro vertical a Step recogn module Is the first o step Do double integral and reset vertical velocity once Y Use Pythagorean to calcula Step Len Combine with gry estimate User 's c Noise Filte Fig. 6. Flow chart of Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 Intuitively, one can use a low pass filter and preset a cut-off fre- quency to filter noise. Some prior work has shown that the frequency of human muscle movement is lower than 16 Hz (Gerald & Andreas, 2008), and the human step frequency is never higher than 3 Hz (Lei et al., 2005). Therefore, some step recognition systems use 3 Hz as their cut-off frequency to filter noise signal (Martin & Michael, 2009). Initially, we also used 3 Hz, but then found our results became worse. This shows that, while it is good enough to detect a new step event, a 3 Hz threshold is too low for our purposes, and will remove data which is not noise. Some prior research (Chen & Bassett, 2005) analyzed the acceleration of the waist during walking, and found the maximum acceleration is 8 Hz. We thus set our cut-off frequency at 8 Hz for filtering the noise, and this threshold worked well under repeated experiments with different subjects. 3.2. Step recognition module In order to recognize a step, we first analyze the components of one step that can cause vertical changes to the body. There are three major events which may affect the height of the waist, as shown in Fig. 7.1. The first one is a heel-touching-ground event, which happens when the heel just hits the ground and the waist is in its lowest position during the entire step. The event that comes after this is the stance, which occurs when the foot is flat on the ground. Finally, the heel-off-ground event occurs right after the stance. Generally, as shown in Fig. 7.2, the vertical acceleration of a heel-touching-ground event is the local minimum within a step. In addition, previous studies (Gafurov, Snekkenes, & Bours, 2007; Shih, 2010) showed that human walking frequency is never over 3 Hz. Therefore, the duration between two consecutive l ion m the xis ition r last Do double integral and reset vertical velocity twice N Theorem te gth oscope to oordinates ring our PDR system. location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 Fig. 7.1. Walking diagram (modified from Walking diagram). 6 K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx heel-touching-ground events must be over 0.33 s. Based on the above, we use a sliding window algorithm to detect every heel-touching-ground event, and define one step as from a heel-touching-ground event to the next heel-touching-ground event. Once a new step is identified, the sensor data between two consecutive heel-touching-ground events will be used to estimate the step length. 3.3. Step length estimator 1) Our step model To measure the walking distance from point A to point B, we sum up the step length of all steps. We model a complete step as from one heel-touching-ground-event to the next. Therefore, we consider the first step (i.e., from the stance event to the heel-touching-ground event, as shown in Fig. 7.1) and the last step (i.e., from the heel-touching-ground event to the stance event) as half a step. When the first or last step is detected, our system will divide the calculated step length by 2. 2) Avoid accumulation of sensor drift errors Once each step can be identified, we can then do the double integral to calculate the height change of the waist and then use this information to estimate the length of each stride based on the Pythagorean Theorem. However, as discussed previously, if the system only naively does the double integral on the accelerom- eter data, the sensor drifts errors could accumulate from one step to the next. To avoid this, we previously proposed a zero velocity Fig. 7.2. The vertical acce Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 update method (Lan & Shih, 2012) to calibrate the sensor data: if we mount a sensor on the waist of a pedestrian, then while they are walking the trajectory of the sensor can be approximated by Simple Harmonic Motion (SHM). In addition, given that the velocity of the highest and lowest points of the sensor will be zero, we can utilize these characteristics to detect when the user starts a new step. Finally, once the points with zero velocity in the vertical direction (where the vertical acceleration reaches its local maximum, i.e., points A and B in Fig. 7.2) are identified, we can then reset the vertical velocity to zero before doing the double integral for calculating the height change of the user’s waist. The height change can then be used to estimate the step distance based on the Pythagorean Theorem. Specifically, there are three major events during one step, namely heel-touching-ground, stance, and heel-off-ground. As shown in Fig. 7.2, the lowest valley (point A) of the wave indicates when the heel is touching the ground (corresponding to the fifth posture in Fig. 7.1), the first peak occurs (point B) when the walker is in the stance state (corresponding to the first posture in Fig. 7.1). Following the stance event, the body starts to lean forward and the foot is now on its toes, which will give a force to push the body up, so that the vertical acceleration will change to the opposite direction and cause the second valley in Fig. 7.2 (i.e., point C, corresponding to the second posture in Fig. 7.1), due to the law of inertia. As shown in Fig. 7.1, when the stance and heel-touching-ground events occur, the waist has the largest displacement from its equi- librium position. Therefore, we reset the vertical velocity to zero at these points. We performed an experiment to observe the effect of leration of walking. location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx 7 doing such a zero velocity update (ZUPT). As shown in Fig. 8, implementing ZUPT can indeed avoid the accumulation of sensor drift errors. The first and last steps are considered as special cases. There is only one point where the velocity needs to be reset, because the initial velocity of the first step is already zero, and the last step does not need to reset this when the body is in its stance state. 3) Moving at high speed or on the same spot As shown in Fig. 7.2, when one walks at a normal speed, the stance event generally occurs as the first peak after the heel-touch- ing-event. However, during high speed movement, we observe that the counter-force from the ground could introduce an extra pulse between the heel-touching-ground and stance events, as shown in Fig. 9 (points A’ and B’) and Fig. 10. Without taking this into consideration, our system could reset the velocity at the wrong point. Previous studies in kinesiology (Chintalapudi, Iyer, & Padmanabhan, 2010; Gusenbauer et al.,2010) showed that the dura- tion from the heel-touching-ground event to the stance event nor- mally accounts for at least 16.7% of the time in a step. Based on this observation, we can identify the right point where the stance event occurs and reset the velocity when calculating the step length. When one is moving on the same spot, the acceleration data should not be considered in the calculation of step length. In addi- tion, when one is walking on the same spot, the height of the waist does not usually change significantly. Therefore, we use a simple threshold-based filter to detect this phenomenon by looking at whether the acceleration data in all three directions (vertical, hor- izontal, and lateral) are lower than a certain threshold, e, as shown below, and then reset the vertical acceleration accordingly (in our implementation, e is set to 1 m/s2). verticalAcc ¼ 0; if jlateralAccj < e & jhorizontalAccj < e & jverticalAccj < e 4. Gyroscope error calibration with map matching The effectiveness of a PDR system lies in its success in accu- rately estimating the user’s moving distance and direction. In the previous section, we discussed how to use ZUPT to reduce the sen- sor drift error when measuring the distance. In a PDR system, the direction in which the user is moving is usually obtained from a gyroscope sensor. However, a gyroscope can only produce the relative angular displacement (RAD) of a device with respect to a specific direction, and this is not necessarily the absolute direction. Therefore, while we could track the user’s trajectory using the gyroscope, this trajectory might be biased by the error in its initial direction, and appear as a rotated version of the true path, as in the example shown in Fig. 11. When the error of the gyroscope is significant, and if left uncorrected, it can make the entire PDR sys- tem unusable. Map-matching is the process of comparing the Fig. 8. The velocity with ZUPT versus without ZUPT. Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 pedestrian’s trajectory data with a digital map of the environment to match the trajectory data to the route segment on which the pedestrian is walking, and it can be used to correct the heading er- ror of a PDR system (Aggarwal, Thomas, Ojeda, & Borenstein, 2011). In this study we propose a new map-matching method utilizing widely available building floor plans (instead of a detailed scaled map, as in some earlier works (Ascher et al., 2010; Ishikawa, 2009) to calibrate the gyroscope errors. We assume that a floor plan is an ‘‘approximate’’ scaled down version of the physical layout of the floor. Our basic idea is to utilize the geometric simi- larity between the trajectory data and the floor plan to infer the last-visited corner by the user. The flow chart of our algorithm is shown in Fig. 12. Before starting the map matching, we adopt an approach similar to that in a prior work (Gillieron & Merminod, 2003) by first converting the floor plan into a link-node model. The turning angels of the corners and comparative ratios of the lengths between any two corridors are then estimated, as shown in Fig. 13. The link-node model is used to approximate the layout of corridors and corners. We then compare the geometry of the user trajectory with the link-node model to find the possible routes that the user has travelled. Here we consider the map and the trajectory as two independent graphs, say M and T. We list all the sub-graphs of M and compare T with all these to find the most similar one. We define the ‘‘similarity’’ between two graphs by comparing their shapes, vertex angles (i.e., the angle between two connected edges) and relative edge lengths (normalized). Once a unique route is identified, this can then be used to calibrate the trajectory data and identify the most recently visited corner. Since the location of every corner is known within the floor plan, the system can then localize the user while they move between corners using the dead-reckoning data from the accelerometer, as previously discussed in Section 3. Note that, given that the link- node model is only an approximation of the physical layout of the building (e.g., the link length might not be an exact scaled down version of the corridor length), the results of this comparison between the map and trajectory could generate multiple candidate routes. Therefore, we also implement an RSSI-based filter by using the existing WiFi-signal-based landmarks (e.g., a corridor-corner may overhear a unique set of WiFi APs, but the set may change at short distances away from that spot, and some dead spots inside a building may not overhear any WiFi signals, which by itself is a signature). When a WiFi AP signal is available, we can use this RSSI-filter to select the correct route from multiple candidates. For example, if we know the user has passed a certain landmark, we can remove those routes that do not contain it. This approach is similar to the method used by a recent study (Wang et al., 2012). The map-matching process is performed every time the system detects that the user makes a turn. 4.1. Turn detection The first step in our map matching process is to find out all the turnings in the trajectory data and use these as a signature to match all possible routes on the map. However, given that the gyroscope can only provide the relative angular displacement (RAD), we use a sliding-window-based algorithm to infer the user’s possible turnings from the data. To determine if a person is making a turn, we compare the standard deviation of the window with a threshold which is estimated during the period when the user is walking straight. A turning event is considered to have occurred when the standard deviation of the window exceeds the threshold. Note that, since it could take several steps to turn around a corridor corner, we record all the turning events that are related to a possi- ble corner-turning event in order to compute the angle of making a complete turn of this corner. One example is shown in Fig. 14. The turning angle (CT) of a possible corner can be estimated as location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 Fig. 9. The difference between normal and high speeds. Fig. 10. The acceleration at different high speeds. 0 5 10 15 0 5 10 15m m Corners on the map Trajectory A B C D I cm (a) Fig. 11. (a) Link-model of the map and the ground truth 8 K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 CT = AT � BT Here AT is the heading angle after making a turn around a corner, and BT is the heading angle before making a turn around a corner. For the example in Fig. 14, CT = AT � BT = 90 � 0 = 90. However, in reality, it is not necessary that one only makes a turn when encountering a corner. For example, one might walk back and forth along the same corridor/aisle. In addition, possible gyroscope drift errors can also produce false turning events. Therefore, detection of a turning event is not nec- essarily an indication that the user is indeed passing a corner. We consider these kinds of turning events, which are detected when the user is not passing a corner, as ‘fake’ turnings. Neverthe- less, it is difficult to distinguish normal corner turnings from fake ones based only on accelerometer and gyroscope data. In this study we thus utilize the floor map information to resolve the issue of fake turnings, as follows. The assumption we make here is that once the fake turnings are removed from the trajectory data, the trajectory data should be geometrically similar to a possible route on the map. 0 200 400 600 800 1000 1200 1400 1600 1800 2000 2200 0 200 400 600 800 1000 1200 cm Detected Trajectory A B C D (b) , (b) the estimated trajectory from the sensor data. location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 Trajectory Data Turning detection List of all possible routes Shape filter Angle filter Ratio filter Two-phase geometry filters RSSI filter Reset the trajectory data and Identify the last-visited corner Fig. 12. The flow chart of our map-matching algorithm. K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx 9 We first try to find all the possible routes that a user might take (say, RT) based on all the possible combinations of detected turn- ings from the trajectory data. We then compare these routes (i.e. RT) with all the possible routes on the map (say, RM). The objective here is to find an (RT, RM) pair which is ‘‘the most geometrically sim- ilar’’. We use some methods from image processing theory to solve this problem. We first model the map as a graph, GM(VM,EM). VM is the corner of map, such as A in Fig. 11(a), and the EM denotes the corridor between two adjacent corners. We also define the graph GMi(VMi, EMi) as the sub-graph of GM, which is used to model all pos- sible routes on a map. In addition, we model the user trajectory as a graph GT(VT, ET), where VT stands for an ordered set of detected turnings (including fake ones), such as A’, B’, C’, D’ in Fig. 11(b), and ET is the edge set whose element is the connection between two adjacent detected turnings. That is, assuming VT = {V1, V2, . . ., Vk}, Vk is the k-th detected turning, then ET = {V 1 V 2; V 2 V 3; . . . ; V k�1 V k }. We next define the graph GTj(VTj, ETj) as follows. There exists a set V’ which is the power set of VT (the set that contains all subsets of VT), i.e., V’ = {F, {V1}, {V2}, . . . {V1,V2}, {V1 ,V3}, . . . {V1 ,V2 ,V3 ,. . .Vk}}. Here we let VTj be an ordered set, VTj 2 V’ and |VTj| > 1. The ETj is the edge set which contains the edge between any two adjacent elements in VTj. In other words, Fig. 13. The link-node mo Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 GTj is used to model all possible routes that could be generated based on the detected turnings (including fake ones), and GMi stands for the accessible route on the map. Again, the idea here is to find a (GMi, GTj) pair which is the most geometrically similar. In other words, we want to eliminate those hypothetical routes generated in GTj that cannot be found on the real map. We design a two-phase filtering mechanism and employ three filters: shape filter, angle filter and edge filter. In phase one, we input all (GMi, GT) pairs into these three filters to remove those which are not geometrically similar. The purpose of phase two is to remove the non-existing routes caused by fake turnings by using these filters for all possible (GMi, GTj) pairs. 4.2. Two-phase filtering mechanism One possibility is that we might not have ‘‘sufficient’’ trajectory data (e.g., when there is no detected turning in the trajectory data), and that multiple candidate routes remain after the geometry-sim- ilarity filtering. Therefore, we also employ an RSSI-based filter by using the existing WiFi-signal-based landmarks, as described previously, to further select the correct one among multiple candi- dates. Next, we discuss the details of each filter. 1) The shape filter We adopt the idea of a shape descriptor (Belongie, Malik, & Puzicha, 2002; Hao & Malik, 2003) to compare the shapes of two graphs. Considering the nature of our input data and the computa- tional overhead, we modify the original shape descriptor method to suit our scenario. To reduce the computational overhead, we calculate the (Centroid; Johnson, 1929) of each graph and create a line every 10 degrees from 0� to 180� to pass through this. We then calculate how many crossing points on the edge can be made by each line and record them in a one-dimensional array, as shown in Fig. 15. Finally, we compare the arrays generated based the two graphs to determine their similarity using Euclidean distance (L-2 norm) Dunford & Schwartz, 1958. We set a threshold to judge the similarity between these two graphs. If the value is over the threshold, the system will consider these two graphs are different and remove them from the set of candidates. 1) The angle filter We adopt the concept of a chain code (Saghri & Freeman, 1981) to implement the angle filter. The chain code uses a sequence of numbers to represent a series of different moving directions and transform a graph into a one-dimensional expression. However, we cannot directly apply a full-fledged chain code to our system. First, it is difficult to decide the number of sampling points for the trajectory data, since each step distance could be different. Second, the computational overhead of using the chain code is pro- portional to the number of steps in the trajectory data, and the number of candidate routes on the map. Given that what we del from a floor plan. location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 0 , 0°, 15°, 60°, 90°, 90°, 90° - Gyro data Deviation = 8.6 < threshold Deviation = 31 > threshold Deviation = 37 > threshold Making a turn Walking straight 906015 1500 (Threshold = 30, window size =3) 60150 Fig. 14. An example of turning detection. 10 K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx consider here is a real-time localization system and the computa- tional capability of a smart-phone is limited, we cannot just na- ively use the chain code. Therefore, we adopt the concept of the chain code and implement it separately via two different filters: the angle filter and the edge filter. Conceptually, the outputs of the chain code include the angle information (e.g., A and A’ in Fig. 11), the direction of the edge and the normalized edge ratio (i.e., divide the distance of each edge by the distance of the longest edge). In our angle filter, for example, we compare every matching angle and the direction of each edge between GT and GMi, and use a threshold to determine whether they are all close enough, as shown in Fig. 16. After filtering out those GMi which do not have similar angles, we then implement the second part of the chain code with an edge filter, as described below. 2) The edge filter With this filter, we check whether the normalized edge ratios between two graphs, for example, GTj and GMi, are similar or not. The system calculates the displacement between any two adjacent vertices in the graph and stores these displacements as a vector. Fig. 17 shows an example input to the edge filter. We then use the Euclidean distance and set a threshold to determine whether the vectors produced for GTj and GMi are similar or not. If the value of the L-p norm of these two graphs is over the threshold, the system then removes the corresponding (GTj, GMi) pair from the candidates. Fig. 15. Example of Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 5. Evaluation In this section, we discuss the performance of our PDR-based localization system. We first discuss the results of our step length estimation method, and then show the performance of the above- mentioned map matching algorithm. 5.1. Experiment setup To implement our algorithm, we use a variety of Android phones from HTC and Samsung. We perform two set of experi- ments on the smart-phones. One is placing the smart-phone in a shirt pocket and the other is putting it in a pants pocket. We use a laser distance meter to measure the actual travel distance of the user. In addition, to record the user’s actual location, we pasted markers on the ground at precisely known locations. Each of these markers had a number on it, and the user recorded the numbers when they walked passed them. 5.2. Step length estimation As discussed previously, we use a low-pass filter to filter out the noise in our step length estimation algorithm. We choose 8 Hz as our cut-off frequency for filtering the noise. To examine whether this frequency produces the best results, we perform a set of exper- iments using different frequencies for the low-pass filter, ranging from 3 Hz to 12 Hz, and compare their accuracies in estimating step length. As shown in Fig. 18, the use of 8 Hz as the cut-off fre- quency produces the most accurate results. Some of the state-of-the-art pedometers on the market can also output walking distance. We compare our method with these pedometers, and find that they only achieve up to 90% accuracy, which is significantly lower than our results (98%). Furthermore, as discussed previously, our approach is based on the idea of using the change in height of the waist to estimate step length, which is similar to the method used in Weinberg (2002). In his work, he proposed an equation to estimate distance by observing the the shape filter. location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 U: Angle = (90 ,-90°) E: EdgeDirection = (0°,90°,0°) O: Orientation = 0° Equal (Map(U,E,O), Trajectory(U,E,O)) ° Fig. 16. Features of the angle filter. Fig. 17. The input to the edge filter. 86 88 90 92 94 96 98 100 3 4 5 6 7 8 9 10 11 12 Hz % Accuracy Fig. 18. The accuracy using different cut-off frequencies for the low pass filter. Fig. 19. Accuracy of estimated distance under different walking speeds. Fig. 20. Accuracy under different inclinations. K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx 11 vertical acceleration during walking. We implement Weinberg’s method and find that its accuracy is about 96.7%. In addition, one limitation of Weinberg’s approach is that its system parameters need to be re-trained for every new user. In our system, we use a low-pass filter (LPF) to filter out the noise and employ zero velocity update (ZUPT) to calibrate the drift Table 1 comparison with other waist mounted methods and the influence of the mechanism. Pedometer Method Constant step length � step number Accuracy (Estimation ground truth) 90.09% Standard deviation 1.83% Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 error of the accelerometer. To understand the effects of these mechanisms, we enable only one of them at a time. As shown in Table 1, ZUPT has a higher impact on the system performance than the implementation of the low-pass filter. When ZUPT is disabled, the accuracy drops from 98.25% to 84.1%. Considering the variations that might exist in different individ- uals’ walking gaits, we performed an experiment in which we asked the subject to walk with different speeds, as shown in Fig. 19. Here the step frequency is defined as the number of walk- ing steps in one second. The average accuracy is about 98.26% and standard deviation is about 1.09%. In addition, we asked two other people (one male and one female) to perform the same experiment for the walking distance measurements. Again, each experiment was repeated for 10 times. The average accuracies and standard deviations are (98.42%, 0.98%) and (97.93%, 1.86%) respectively, which are similar to the results shown in Table 1. Finally, to understand the effects of moving uphill of downhill on the accuracy of walking distance estimation, we carried out a set of experiments in which the user walked on with different slope inclinations of 0�, 5�, 10�, 15� and 20�. The results show that the step length is shorter when the slope becomes steeper. The accuracy of walking distance estimation also degrades as we in- crease the level of inclination from 0� to 20�, as shown in Fig. 20. In addition, the accuracy in the downhill direction is generally low- er than that in the uphill direction. 5.3. Map matching As discussed above, we use a sliding-window-based algorithm to detect the user’s possible turnings from the gyroscope data. To determine if a person is making a turn, we compare the standard Weinberg Ours (A: LPF, B: ZUPT) k � n � ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Accmax � Accmin4 p Original A B None 96.78% 98.25% 84.1% 95.29% 75.53% 3.18% 1.29% 7.41% 2.22% 16.3% location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 Fig. 21. The number of walking steps between two corners is larger than the window size. 12 K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx deviation of the window with a threshold. Generally, when the chosen window size is too small, all the walking steps that happen during a turning might not be able to be included within a window. On the other hand, when the chosen window size is too big, two different turnings can be included in the same window if the num- ber of walking steps between two corners is less than the window size. As shown in Figs. 21 and 22, the accuracy of detecting turn- ings becomes lower when the chosen window size is too small or too big. Therefore, in our experiments we set the window size from Fig. 22. The number of walking steps between two corners is smaller than the window size. Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 the range defined below, in which D is the shortest distance between two adjacent corridors. 2 < Window Size < D Avg Step Length In addition, the threshold is obtained using the standard devia- tion of the window during the period when the user is walking straight. In the shape filter, we set a line, every 10� from 0� to 180�, that passes through the centroid of the graph, and use the crossing points on the edges obtained in this way to determine if two graphs have similar shapes. To understand the effects of the gaps between two crossing lines (default 10�) on determining shape similarity, we vary this gap from 0� to 70�. Generally, the system will have less computational overhead when the gap is larger. However, a larger gap also suggests that poorer results will be obtained, since fewer crossing points will be generated for the comparison of shape similarity. As shown in Fig. 23, we start get- ting inaccurate results when the gap is larger than 15�. Finally, to test the performance of our localization system, we choose a route which is about 40 m long and includes four corri- dors and corners, as shown in Fig. 24. We deliberately created some ‘fake turnings’ by having a walker walk straight down the route but wander about at certain spots (we put markers on these 0 20 40 60 80 100 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 Gap between two crossing lines (degree) A cc ur ac y (% ) Fig. 23. The discrimination of different slopes. location system for smart parking. Expert Systems with Applications (2013), http://dx.doi.org/10.1016/j.eswa.2013.09.044 -200 300 800 1300 1800 2300 0 500 1000 1500 cm cm Trajectory Ground Truth Wandering Turning at the corner Walking back and forth Fig. 24. The testing environment and route. K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx 13 spots before the experiment started), shown as the green dashed circles in Fig. 24. The solid circles in Fig. 24 indicate when this walker made a ‘real’ turn at the corner. We repeated this experi- ment 10 times, and our results show that our location error is about 0.48 m, and the standard deviation is about 0.43 m. 6. Discussion and conclusion In this work, we assume that the floor plan is a scaled-down version of the physical layout of the floor. In some cases, this assumption might not be always true. Here we propose a bootstrap phase based on the participatory sensing approach (Mohan et al., 2008; Raghu, Nam, Hossein, Saurabh, & Tarek, 2010) to obtain the scale information of the map. The idea is simple: in this boot- strap phase the system first collects the users’ trajectories, and then compares them with the link-model of the map to estimate the relative distance of every corridor. Note that, while the first few users may experience inferior location accuracy, a little more data will bring the system to convergence. In addition, such a bootstrap phase only needs to be performed once for each building. Nevertheless, one limitation of this work is its reliance on a partic- ipatory sensing system to collect enough scale information for the map, and we are currently investigating how to overcome this. Furthermore, in our map matching algorithm, we compare all pos- sible (RT, RM) pairs to find the one which is ‘‘the most geometrically similar’’. Two factors could contribute to the system’s computa- tional overhead: the frequency of performing the map matching process and the number of possible (RT, RM) pairs to be considered. In our implementation, we only perform map matching when the system detects the turning event (including ‘fake turnings’). In other words, we only need to run the step estimation module when the user walks on the same corridor if there no fake turning events are detected. The number of possible (RT, RM) pairs depends on the complexity of the floor map and the user’s movement patterns. The number of possible RT after the last-visited corner is generally lim- ited unless the walker wanders about frequently. However, the number of possible RM could be significantly large depending on the complexity of the map. For example, a map with a high node degree distribution could potentially generate a larger number of Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 possible RM, and result in a higher computation overhead. We are currently studying this issue by considering utilizing the user’s tra- jectory history data to reduce the space of possible RM. To conclude this paper, we’ve made the following contributions in this work. First, unlike previous crowdsourcing-based methods that required the drivers to manually report when they leave their parking spots, we utilize the phone sensors to automatically infer the parking space availability and notify the other drivers in advance by tracking the trajectory of the driver. Second, we imple- ment a waist-mounted PDR method on a smart-phone by using the sensors in the phone to track the driver’s walking trajectory. Sec- ond, we design a map matching algorithm to calibrate the direction errors from the gyroscope using building floor plans, which are readily available. Finally, our map matching algorithm implements three filters, namely the shape filter, angle filter and edge filter, to infer the user’s last-visited corner. Our results show that we can achieve about 98% accuracy in estimating the user’s walking dis- tance, while the overall location error is about 0.48 m in our exper- iment. We found that it is important to have a mechanism such as ZUPT to calibrate sensor drift error, as the accuracy drops from 98% to 84% without ZUPT. Furthermore, we show that one might need to consider corridor length in order to detect corners. One limitation of this work is its reliance on a participatory sensing system to collect enough scale information for the map, and we are currently investigating how to overcome this. In addition, we assume that the position of the phone is relatively stable in relation to the leg movements and do not take into ac- count of the possibility of running in this study. Finally, we do not consider the multi-floor building scenario, since most of smart-phones do not have a barometer to detect the change to a different floor, and we leave this as a further direction for future work. Acknowledgments This work was supported by the National Science Council of Taiwan under Grants NSC 101-2220-E-006-010 and NSC 101- 2221-E-006-098-MY3. References Aggarwal, P., Thomas, D., Ojeda, L., & Borenstein, J. (2011). Map matching and heuristic elimination of gyro drift for personal navigation systems in GPS denied conditions. Measurement Science and Technology, 22, 025205. Alvarez, D., Gonzalez, R. C., Lopez, A., & Alvarez, J. C. (2006). Comparison of step length estimators from weareable accelerometer devices. In 28th Annual international conference of the IEEE in engineering medicine and biology society (EMBS ‘06) (pp. 5964–5967). Alvarez, J. C., Gonzalez, R. C., Alvarez, D., Lopez, A. M., & Rodriguez-Uria, J. (2007). Multisensor approach to walking distance estimation with foot inertial sensing. In 29th Annual international conference of the IEEE engineering in medicine and biology society (EMBS 2007) (pp. 5719–5722). Arnott, R., & Inci, E. (2006). An integrated model of downtown parking and traffic congestion. Journal of Urban Economics, 60, 418–442. Ascher, C., Kessler, C., Wankerl, M., & Trommer, G. F. (2010). Dual IMU indoor navigation with particle filter based map-matching on a smartphone. In Indoor positioning and indoor navigation (IPIN), 2010 international conference on (pp. 1– 5). Ayala, D., Wolfson, O., Xu, B., Lin, J., & Dasgupta, B. (2011). Parking slot assignment games. In ACM SIGSPATIAL GIS. Belongie, S., Malik, J., & Puzicha, J. (2002). Shape matching and object recognition using shape contexts. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 24, 509–522. Centroid. Available from http://en.wikipedia.org/wiki/Centroid#cite_ref-0. Chen, K. Y., & Bassett, D. R. (2005). The technology of accelerometry – Based activity: monitors: current & future. American College of Sports Medicine. Chintalapudi, K., Iyer, A. P., & Padmanabhan, V. N. (2010). Indoor localization without the pain. In 16th ACM international conference on mobile computing and networking (MobiCom 2010) (pp. 173–184). Chicago, IL, September 2010. Dippold, M., (2006). Personal dead reckoning with accelerometers. In Presented at the 3rd international forum on applied wearable computing 2006 (IFAWC2006). Bremen, Germany, March 15–16. location system for smart parking. Expert Systems with Applications (2013), http://refhub.elsevier.com/S0957-4174(13)00798-7/h0070 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0070 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0070 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0085 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0085 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0055 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0055 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0055 http://en.wikipedia.org/wiki/Centroid#cite_ref-0 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0045 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0045 http://dx.doi.org/10.1016/j.eswa.2013.09.044 14 K.-C. Lan, W.-Y. Shih / Expert Systems with Applications xxx (2013) xxx–xxx Dunford, N., & Schwartz, J. T. (1958). Linear operators. New York: Interscience Publishers. Vol. I. Eric, F. (2005). Pedestrian tracking with shoe-mounted inertial sensors. IEEE Computer Graphics and Applications, 25, 38–46. Feliz, R., Zalama, E., & García-Bermejo, J. G. (2009). Pedestrian tracking using inertial sensors. Journal of Physical Agents, 3, 35–42. Gafurov, D., Snekkenes, E., & Bours, P. (2007). Gait authentication and identification using wearable accelerometer sensor. In 2007 IEEE workshop on automatic identification advanced technologies (pp 220–225). Gerald, B., & Andreas, T. (2008). DiaTrace – Neuartiges assistenz-system für die gesundheitsprävention zur nahrungsaufnahme und bewegungserfassung. In Deutscher kongress mit ausstellung/technologien – anwendungen – management. Berlin, Germany. Gillieron, P. -Y., & Merminod, B. (2003). Personal navigation system for indoor applications. In Proceedings of the 11th IAIN world congress on smart navigation, systems and services. Berlin. Godha, S., Lachapelle, G., & Cannon, M. E. (2006). Integrated GPS/INS system for pedestrian navigation in a signal degraded environment. In Presented at ION GNSS (pp. 26–29). Google Latitude. Available from http://www.google.com/intl/en/mobile/latitude/. Gusenbauer, D., Isert, C., & Krösche, J. (2010). Self-contained indoor positioning on off-the-shelf mobile devices. In Indoor positioning and indoor navigation (IPIN), 2010 international conference on (pp 1–9). Hao, Z., & Malik, J. (2003). Learning a discriminative classifier using shape context distances. In Computer vision and pattern recognition, 2003. proceedings. 2003 IEEE computer society conference on (Vol. 1, pp. I-242–I-247). Ishikawa, T., Kourogi, M., Okuma, T., & Kurata, T. (2009). Economic and synergistic pedestrian tracking system for indoor environments. In 2009 International conference of soft computing and, pattern recognition (pp. 522–527). Ji, Y., Biaz, S., Pandey, S., Agrawal, P. (2006). ARIADNE: A dynamic indoor signal map construction and localization system. In MobiSys. Jimenez, A. R., Seco, F., Prieto, C., & Guevara, J. (2009). A comparison of pedestrian dead-reckoning algorithms using a low-cost MEMS IMU. In 2009 IEEE international symposium on intelligent signal processing (WISP 2009) (pp. 37–42). Johnson, R. A. (1929). Modern geometry: An elementary treatise on the geometry of the triangle and the circle. Boston, MA: Houghton Mifflin, pp. 173–176, 249–250, and 268–269. Jorgensen, F. (2008). Accurate step counter. U.S. Patent 0 172 203 A1, Jul. 17. Krishna, C., Anand Padmanabha, I., & Venkata, N. P. (2010). Indoor localization without the pain. In Presented at proceedings of the sixteenth annual international conference on Mobile computing and networking. Chicago, Illinois, USA. Ladetto, Q., & Merminod, B. (2002). In step with INS-navigation for the blind, tracking emergency crews. GPS World, pp. 30–38. Lan, K. -C., & Shih, W. -Y. (2012). Using simple harmonic motion to estimate walking distance for waist-mounted PDR. In 2012 IEEE wireless communications and networking conference (WCNC). Lan, K. -C., & Wang, H. -Y. (2013). On providing incentives to collect road traffic information. In International wireless communications & mobile computing conference (IWCMC’13). Cagliari, Sardinia, Italy, 1–5 July. Lei, F., Antsaklis, P. J., Montestruque, L. A., McMickell, M. B., Lemmon, M., Yashan, S., et al. (2005). Design of a wireless assisted pedestrian dead reckoning system – the NavMote experience. IEEE Transactions on Instrumentation and Measurement, 54, 2342–2358. Li, F., Zhao, C., Ding, G., Gong, J., Liu, C., & Zhao, F. (2012). A reliable and accurate indoor localization method using phone inertial sensors. In 14th ACM international conference on ubiquitous, computing (Ubicomp’12). Lim, H., Kung, C. L., Hou, J. C., Luo, H. (2006). Zero configuration robust indoor localization: Theory and experimentation. In Infocom. Martin, M., & Michael, M. (2009). A step counter service for Java-enabled devices using a built-in accelerometer. In Presented at proceedings of the 1st international workshop on context-aware middleware and services: affiliated with the 4th international conference on communication system software and middleware (COMSWARE 2009). Dublin, Ireland. Mathur, S., Jin, T., Kasturirangan, N., Chandrasekaran, J., Xue, W., Gruteser, M., et al. (2010). ParkNet: Drive-by sensing of road-side parking statistics. ACM MobiSys. Mohan, P., Padmanabhan, V. N., & Ramjee, R. (2008). Nericell: Rich monitoring of road and traffic conditions using mobile smartphones. In Proceedings of the 6th ACM conference on embedded network sensor systems. SenSys ’08 (pp. 323–336). New York, NY, USA: ACM. Please cite this article in press as: Lan, K.-C., & Shih, W. Y. An intelligent driver http://dx.doi.org/10.1016/j.eswa.2013.09.044 Ojeda, L., & Borenstein, J. (2006). Non-GPS navigation for emergency responders. In International joint topical meeting: Sharing solutions for emergencies and hazardous environments. Salt Lake City, UT, February 12–15. Ojeda, L., & Borenstein, J. (2007a). Personal dead-reckoning system for GPS-denied environments. In 2007 IEEE international workshop on safety, security and rescue robotics (SSRR 2007) (pp 1–6). Ojeda, L., & Borenstein, J. (2007b). Non-GPS navigation for security personnel and first responders. Journal of Navigation, 60, 391–407. Oliver, W., & Robert, H. (2008). Pedestrian localisation for indoor environments. In Presented at Proceedings of the 10th international conference on Ubiquitous computing. Seoul, Korea. OpenSpot. Avaiable from http://techcrunch.com/2010/07/09/google-parking-open- spot/. ParkNet. Avaiable from https://parknet.dk/. Primospot. Avaiable from http://www.primospot.com/. Raghu, K. G., Nam, P., Hossein, A., Saurabh, N., & Tarek, F. A. (2010). GreenGPS: A participatory sensing fuel-efficient maps application. In Presented at proceedings of the 8th international conference on Mobile systems, applications, and services. San Francisco, California, USA. Roadify. Avaiable from http://www.roadify.com/. Sagawa, K., Inooka, H., & Satoh, Y. (2000). Non-restricted measurement of walking distance. In 2000 IEEE international conference on systems, man, and, cybernetics (Vol. 3, pp. 1847–1852). Saghri, J. A., & Freeman, H. (1981). Analysis of the precision of generalized chain codes for the representation of planar curves. In Pattern analysis and machine intelligence, IEEE transactions on (Vol. PAMI-3, pp. 533–539). Schneider, P. L., Crouter, S. E., Lukajic, O., & Bassett, D. R. (2003). Accuracy and reliability of ten pedometers for measuring steps over a 400-m walk. Medicine & Science in Sports & Exercise, 35, 1779–1784. SFpark. Avaiable from http://sfpark.org/. Shih, C. -C. (2010). The implementation of a G-sensdor-based pedometer. M.S. paper, Dept. C.S.I.E., National Central Univ., Taoyuan County, Taiwan. Shin, S. H., Park, C. G., Hong, H. S., & Lee, J. M. (2005). MEMs-based personal navigator equipped on the user’s body. In Presented at ION GNSS 18th international technical meeting of the satellite division. Long Beach, CA, 13–16 September. Shin, S. H., Park, C. G., Kim, J. W., Hong, H. S., & Lee, J. M. (2007). Adaptive step length estimation algorithm using low-cost MEMS inertial sensors. In Sensors applications symposium, 2007. SAS ‘07. IEEE (pp. 1–5), 6–8 Feb. Shoup, D. (2006). Cruising for parking. Transport Policy, 13(6), 479–486. Simple Harmonic Motion. Available from http://en.wikipedia.org/wiki/ Simple_harmonic_motion. SpotScout. Avaiable from https://www.spotscout.com/. Stenneth, L., Wolfson, O., Yu, P., & Xu, B. (2011). Transportation mode detection from mobile phones and GIS information. In ACM SIGSPATIAL GIS. Stirling, R., Collin, J., Fyfe, K., & Lachapelle, F. (2003). An innovative shoe-mounted pedestrian navigation system. In Presented at proceedings of European navigation conference GNSS 2003. Graz, Austria, 22–25 April. Su, C.-M., Chou, J.-W., Yi, C.-W., Tseng, Y.-C., & Tsai, C.-H. (2010). Sensor-aided personal navigation systems for handheld devices. In Proceedings of the 2010 39th International Conference on Parallel Processing Workshops, ser. ICPPW ’10 (pp. 533–541). Washington, DC, USA: EEE Computer Society. Titterton, D. H., & Weston, J. L. (1997). Strapdown inertial navigation technology. Peter Peregrinus Ltd. Walking diagram. Available from http://www.juliecallahan.me/walkcycle.jpg. Wang, H., Sen, S., Elgohary, A., Farid, M., Youssef, M., & Choudhury, R. R. (2012). No need to war-drive: unsupervised indoor localization. In Mobisys. Weinberg, H. (2002). Using the ADXL202 in pedometer and personal navigation applications. In Application notes American devices. Xiaoping, Y., Bachmann, E. R., Moore, H., & Calusdian, J. (2007). ‘‘Self-contained position tracking of human movement using small inertial/magnetic sensor modules. In 2007 IEEE international conference on robotics and automation (pp. 2526–2533). Yan, T., Hoh, B., Ganesan, D., Tracton, K., Lwuchukwu, T., & Lee, J. (2011). CrowdPark: A crowdsourcing-based parking reservation system for mobile phones. In UMASS Technical Report., Tech. Rep. UM-CS-2011-001. location system for smart parking. Expert Systems with Applications (2013), http://refhub.elsevier.com/S0957-4174(13)00798-7/h0060 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0060 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0015 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0015 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0025 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0025 http://www.google.com/intl/en/mobile/latitude/ http://refhub.elsevier.com/S0957-4174(13)00798-7/h0065 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0065 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0065 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0010 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0010 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0035 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0035 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0035 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0035 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0090 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0090 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0075 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0075 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0075 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0075 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0020 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0020 http://techcrunch.com/2010/07/09/google-parking-open-spot/ http://techcrunch.com/2010/07/09/google-parking-open-spot/ https://parknet.dk/ http://www.primospot.com/ http://www.roadify.com/ http://refhub.elsevier.com/S0957-4174(13)00798-7/h0005 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0005 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0005 http://sfpark.org/ http://refhub.elsevier.com/S0957-4174(13)00798-7/h0080 http://en.wikipedia.org/wiki/Simple_harmonic_motion http://en.wikipedia.org/wiki/Simple_harmonic_motion https://www.spotscout.com/ http://refhub.elsevier.com/S0957-4174(13)00798-7/h0050 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0050 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0050 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0050 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0030 http://refhub.elsevier.com/S0957-4174(13)00798-7/h0030 http://www.juliecallahan.me/walkcycle.jpg http://dx.doi.org/10.1016/j.eswa.2013.09.044 An intelligent driver location system for smart parking 1 Introduction 2 Related work 2.1 Parking-guidance 2.2 PDR system 2.3 Step length estimation 2.4 Map-matching algorithm 3 Implement a waist-mounted PDR method using a smart-phone 3.1 Noise filtering 3.2 Step recognition module 3.3 Step length estimator 4 Gyroscope error calibration with map matching 4.1 Turn detection 4.2 Two-phase filtering mechanism 5 Evaluation 5.1 Experiment setup 5.2 Step length estimation 5.3 Map matching 6 Discussion and conclusion Acknowledgments References 
work_xwrdekolhnhxtlhoztfrjmm53q	----	Association rule hiding using cuckoo optimization algorithm Expert Systems With Applications 64 (2016) 340–351 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Association rule hiding using cuckoo optimization algorithm Mahtab Hossein Afshari ∗, Mohammad Naderi Dehkordi , Mehdi Akbari Faculty of Computer Engineering, Najafabad Branch, Islamic Azad University, Najafabad, Isfahan, Iran a r t i c l e i n f o Article history: Received 25 December 2015 Revised 27 May 2016 Accepted 1 August 2016 Available online 2 August 2016 Keywords: Data mining Privacy preserving data mining Association rule hiding Cuckoo optimization algorithm a b s t r a c t Privacy preserving data mining is a new research field that aims to protect the private information and avoid the leakage of this information during the data mining process. One of the techniques in this field is the Privacy Preserving Association Rule Mining which aims to hide sensitive association rules. Many different algorithms with particular approaches have so far been developed to reach this purpose. In this paper, a new and efficient approach has been introduced which benefits from the cuckoo optimization algorithm for the sensitive association rules hiding (COA4ARH). In this method the act of hiding is per- formed using the distortion technique. Further in this study three fitness functions are defined which makes it possible to achieve a solution with the fewest side effects. Introducing an efficient immigration function in this approach has improved its ability to escape from any local optimum. The efficiency of proposed approach was evaluated by conducting some experiments on different databases. The results of the execution of the proposed algorithm and three of the previous algorithms on different databases indicate that this algorithm has superior performance compared to other algorithms. © 2016 Elsevier Ltd. All rights reserved. e d t b 2 2 2 B V V C e t r o t m 1. Introduction As the use of data mining process to extract useful patterns from massive data is increasing, concerns about the disclosure of private information during this process has also risen. Trying to preserve the privacy of sensitive information while extracting use- ful patterns led to the formation of a new field in data mining known as privacy preserving data mining (PPDM). PPDM is applied in all data mining techniques such as clustering, classification, as- sociation rule. The algorithms of this field prevent the disclosure of private information, while preserving the utility of non-sensitive information as much as possible by modification and distortion of the database. However, manipulating the database in order to pro- tect the confidentiality of sensitive information is not flawless and has some side effects. The side effects include: • Hiding failure: It occurs when the PPDM algorithms cannot act completely successful and even after data base sanitization, part of the sensitive information could be extracted by data mining algorithms. • Lost rule or misses cost: sanitization process not only hides sensitive information but also ruins some parts of the non- sensitive data which cannot be extracted any more. ∗ Corresponding author. E-mail addresses: mahtab.h.afshar@gmail.com , Afshari@sco.iaun.ac.ir (M.H. Af- shari), Naderi@iaun.ac.ir (M.N. Dehkordi), mehdi_akbari@hotmail.com , mehdi_akbari@pco.iaun.ac.ir (M. Akbari). c http://dx.doi.org/10.1016/j.eswa.2016.08.005 0957-4174/© 2016 Elsevier Ltd. All rights reserved. • Ghost rule or artificial pattern: Because of the sanitization pro- cess, new and unrealistic patterns may form in database that did not exist before. So far, many algorithms have been introduced in this area, ach of which has different strengths and weaknesses. The con- ucted activities in privacy preserving association rule mining un- il the present time can be divided into three main categories: order-based ( Moustakides & Verykios, 20 06, 20 08; Sun & Philip, 007; Sun & Yu, 2005 ), exact ( Gkoulalas-Divanis & Verykios, 2006, 0 09a, 20 09b; Menon & Sarkar, 2007; Menon, Sarkar, & Mukherjee, 005 ) and heuristic approach ( Dasseni, Verykios, Elmagarmid, & ertino, 2001; Oliveira & Zaïane, 2002, 2003a, 2003b, 2006; Saygin, erykios, & Clifton, 2001; Saygin, Verykios, & Elmagarmid, 2002; erykios, Pontikakis, Theodoridis, & Chang, 2007; Wu, Chiang, & hen, 2007 ). In recent years, meta heuristic algorithms have been mployed for privacy preserving association rule mining. One of he newest meta heuristic algorithms is cuckoo optimization algo- ithm ( Walton, Hassan, Morgan, & Brown, 2011; Yang & Deb, 2013 ). Considering the above-mentioned facts, the main concern and bjective of this study is to use cuckoo optimization algorithm o propose a new algorithm in privacy preserving association rule ining area able to hide the sensitive information completely suc- essful and with minimal side effects by • Not having Hiding failure • Minimizing ghost rule amount • Decreasing lost rule impressively http://dx.doi.org/10.1016/j.eswa.2016.08.005 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2016.08.005&domain=pdf mailto:mahtab.h.afshar@gmail.com mailto:Afshari@sco.iaun.ac.ir mailto:Naderi@iaun.ac.ir mailto:mehdi_akbari@hotmail.com mailto:mehdi_akbari@pco.iaun.ac.ir http://dx.doi.org/10.1016/j.eswa.2016.08.005 M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 341 a a O i d e r a p a w 2 s v s l a m F i V n a B 1 w l b F l ( i c h i o t a e d t t t t K p a i t a i a b f f s b h w t w ( E w E v e e m a n t i b v t w t e d t e i t r a t d 3 d s t i B T c p m o S w o t t C w o n ( The rest of the article is organized as follows: Section 2 contains review of literatures conducted in the field of privacy preserving ssociation rule mining and a brief explanation about the Cuckoo ptimization Algorithm (COA). In Section 3 , the problem of hid- ng sensitive association rules are clearly explained. Section 4 then escribes the proposed algorithm for hiding association rules and xplains it in details by providing an example. In Section 5 , the esults of conducted experiments on Real and Synthetic data sets re provided and the efficiency of proposed approach will be com- ared with some existing algorithms. The effects of sparse data re described in Section 6 . Finally, in Section 7 results and future orks will be discussed. . Background and related work There have been so far many studies conducted in privacy pre- erving association rule mining, some of which are briefly re- iewed in this section. The subject of association rules hiding from frequent item ets through decrease of support was first suggested by Attal- ah et al. that used a lattice-like graph for hiding. Appropri- te time complexity was one of the advantages of the proposed ethod ( Atallah, Bertino, Elmagarmid, Ibrahim, & Verykios, 1999 ). ive algorithms namely 1.a, 1.b, 2.a, 2.b and 2.c are provided n Verykios, Elmagarmid, Bertino, Saygin, & Dasseni, (2004 ) by erykios et al. The three first algorithms are Rule-Oriented and the ext two are Itemset-Oriented. Removing just one sensitive rule at time is one of the disadvantages of aforementioned algorithms. y removing the right hand side of the sensitive rule, algorithm .b decreases the rule support and hides it. In algorithm 1.b, along ith an increase in the number of sensitive rules, the number of ost rules is also increased. Multiple rule hiding was first proposed y Oliveira &Zaïane. Their suggested algorithms were Naïve, Min- IA, MaxFIA and IGA that hide sensitive rules based on the conflict evel and were not efficient with regard to implementation time Oliveira & Zaïane, 2002 ). The two algorithms ISL and DSR were ntroduced by Wang in Wang, Parikh, & Jafari, (2007 ). The former onducts the hiding process by increasing the support of the left and side item of sensitive rules while the latter conducts the hid- ng process by decreasing the support of the right hand side item f sensitive rules. ISL and DSR algorithms are respectively weak in he hiding failure rate and the number of lost rules. The genetic lgorithm for hiding sensitive rules is suggested by Dehkordi and t al. in Dehkordi, Badie, & Zadeh, (2009 ). Besides hiding sensitive ata, the main purpose of this algorithm is to decrease modifica- ions in the databases. A preprocess phase for selection of sensitive ransactions of database and making changes in them is defined in his algorithm. Four different strategies are also provided in order o calculate Fitness. By introducing a new Fitness function, Azam han et al. in Khan, Qureshi, & Hussain, (2014 ) have improved the roposed algorithm in Dehkordi et al., (2009 ). Cuckoo Algorithm was first developed by Yang & Deb, (2009 ) nd then was investigated in details by Rajabioun, (2011 ). Interest- ng and different lifestyle as well as egg-laying of cuckoo has been he main influence for development of this algorithm. This bird is ble to brilliantly deceive other birds and make them participate n its own survival. Unlike other birds, cuckoo never builds nests nd never protects its eggs, but puts its eggs in the nests of other irds to be protected along with other eggs. Like other evolutional algorithms, COA also begins its work rom a random initial population which is called “habitat ” and it is ormed by cuckoos . A habitat is expressed by an array of 1 ×Nvar , howing current living position of cuckoo. This array is identified y Eq. (1) abitat = [ x 1 , x 2 , . . . , x N v ar ] , (1) here N var is the dimension of the problem. Each cuckoo is allocated some random eggs that will be put in he nests of a number of host birds. The nests of host birds exist ithin a certain range that is called maximum Egg Laying Radius ELR ) which is identified by Eq. (2) LR = [ ∝ × Number o f current cuckoo ′ s eggs T otal number o f eggs ] ×( V a r hi −V a r low ) , (2) here ∝ is a number, intended to handle the maximum value of LR. Var hi and Var low are respectively high and low limit of each ariable in optimization problems. A number of cuckoo’s eggs with more similarity to the host ggs will have more chance to turn into a mature cuckoo. Other ggs (about 10%) are identified by the host bird and removed. The ore the number of survived eggs in one region, the more benefit llocated to that region. Therefore, a situation in which the largest umber of eggs survives will be a parameter that COA aims to op- imize. Survived chicks are grown in the host’s nest and will turn nto mature cuckoos. At the time of egg-laying, cuckoos migrate to etter and more beneficial habitats with high chance of egg sur- ival. When migrating, cuckoos do not traverse the whole course oward the target point and pass a part of it ( λ% of total path) to- ard the target and then have a φ deviation. This parameter helps hem to search for more areas. After arriving to the target place, ach cuckoo owns a number of eggs according to which it’s ELR is etermined and the egg-laying is started. Considering the fact that here are always such factors as food shortage, hunting by hunters, tc. in the nature that causes balance in the number of birds, there s also a number like N max in Cuckoo Algorithm that can control he maximum number of cuckoos. After several iterations, cuckoos each a point with maximum similarity of eggs with the host eggs nd with maximum food resources. It is a place with the highest otal benefit and lowest number of eggs being destroyed. For more etail on COA, refer to Rajabioun, (2011 ). . Problem definition Assume that I = {i 1 , i 2 , i 3 ,…, i m } is a complex of items and D is a atabase including several transactions. Each transaction is a sub- et of I. The extracted association rules set from D is R and sensi- ive association rules set is R s , where R s ⊂R . Each association rule s expressed as A → B where A is antecedent or left hand side and is consequent or right hand side, so that A, B ⊂I and A ∩ B � = ∅ . he purpose is to change D into D ′ , so that the existing rules in R s annot be mined from D ′ while the existing rules in R − R s can be ossibly mined. Two criteria are considered in the association rule ining; the first is called item support that indicates frequency of ne rule in database and can be calculated by Eq. (3) : up ( A → B ) = | A ∪ B | | D | , (3) here Sup ( A → B ) is the support of A → B , | A ∪ B | is the number f transactions that include both item sets A and B and | D | is the otal number of transactions. The second criterion is called the rule confidence that indicates he rule strength in database and can be calculated by Eq. (4) : on f ( A → B ) = | A ∪ B | | A | , (4) here Conf ( A → B ) is the confidence of A → B , | A ∪ B | is the number f transactions that include both item sets A and B and | A | is the umber of transactions of database that include item set A . For each association rule, one Minimum Support Threshold MST ) and one Minimum Confidence Threshold ( MCT ) are defined. 342 M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 Table 1 Definitions of notations. Notation Definition I An itemset D An original dataset D’ A sanitized dataset R The extracted association rules set from D R’ The extracted association rules set from D’ R s A set of sensitive association rules ∼R s A set of non-sensitive association rules r An association rule A → B An association rule Sup ( A → B ) The support of A → B Conf ( A → B ) The confidence of A → B MST The minimum support threshold MCT The minimum confidence threshold Var hi The high limit of each variable Var low The low limit of each variable α The number of total iterations of right hand side items of sensitive rules K The number of new solutions N The number of current solution’s items that should be modified N max The maximum number of survived solutions with better fitness values N var The dimension of problem N pop The size of initial population MaxNNS The max number of new solutions MinNNS The min number of new solutions MR The modifications radius HF The hiding failure LR The lost rules GR The ghost rules HD The hiding distance LD The lost distance RHD The rules hiding distances RLD The rules lost distances d n s 4 N i N h h t i w t t a s o t b a 1 i t A rule is minable when its support is more than MST and its con- fidence is higher than MCT. Therefore, in order to hide a rule it is required to reduce its support or confidence to a level under the thresholds. Table 1 shows a list of notations and their definitions used in this paper. 4. Proposed algorithm The main steps of the proposed algorithm named Cuckoo Op- timization Algorithm for Association Rule Hiding (COA4ARH) are shown in Algorithm 1 and Fig. 1 shows the flowchart of the algo- rithm. Fitness function, function to find the best solution and im- migration function are shown in Algorithms 2, 3 and 4 and these algorithms will be discussed in detail in the next sections. 4.1. Preprocess of original dataset The first step of the proposed algorithm includes a preprocess operation on the original dataset which is one of the most impor- tant and influential phases of the proposed approach. Since the sensitive items exist only in some transactions of the database, changing and manipulating all transactions is a time-consuming and useless task that not only increases the size of habitats, but also causes the production of useless and unrelated solutions and an increase in the time consumed for obtaining more effective so- lutions; hence increasing the sanitization time. To avoid such neg- ative effects, a preprocess operation have been considered that in- cludes two phases. In the first phase, the original database is pro- cessed in such a way that only critical transactions of the database are chosen. In the present study, critical transaction is a transac- tion that fully supports one or more sensitive rules. In the second phase of the preprocess operation only those sensitive items with critical role in sanitization, are addressed for change. Sensitive item is defined as follows: 1 If a sensitive rule has only one item in its right hand side, the same item will be selected as the sensitive item. 2 If a sensitive rule has more than one item in its right hand side, from among them an item will be chosen as a sensitive one that has respectively the most frequency in the right hand side of sensitive rules and the least frequency in non-sensitive rules. Therefore, conducting preprocess operation on the original atabase will lead to decrease in size of habitats, reduction of un- ecessary changes in database, prevention of producing irrelevant olutions and speedier access to an optimal solution. .2. Generating initial population In order to solve an N var dimension problem with the size of pop , the Cuckoo Optimization Algorithm randomly generates an nitial population in the form of a habitat matrix with the size of pop × N var . Each member of this population shows the current abitat of cuckoos. Since each habitat is allocated to one cuckoo, ereinafter for comfort, “cuckoo’s habitat" is replaced with con- racted form of “cuckoo" . In the present problem, each cuckoo in nitial population is indicative of a solution (sanitized database) hich is shown with a sequence of 0 s and 1 s. In this sequence, he presence and absence of one item in transaction are respec- ively presented with 1 and 0. The original database is considered s the first cuckoo or solution in initial population. Indeed, the first olution of initial population is a sequence of critical transactions f original database that were selected in the preprocess opera- ion. The random generation of the other ( N pop −1 ) solutions can e conducted as such: only those sensitive items that had been ddressed in preprocess operations are randomly quantified (0 or ) and other items are left unchanged from the first solution (orig- nal database) and are transferred to other solutions. Thus, an ini- ial population with N pop number of solutions is generated. Each M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 343 Fig. 1. Flowchart of COA4ARH algorithm. 344 M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 Algorithm 1 COA4ARH algorithm. Input: Original Dataset D, R s , MST and MCT ; α, N pop , N max , MinNNS, MaxNNS, MaxIteration ; Output: A Sanitized Dataset D’ ; begin Preprocess of Original Dataset; Generate an initial population; call FitnessFunction; call BestSolutionFunction; repeat // Generate new solutions for each solution in population K = ( Ma x N N S − Min N N S ) × Rand[ 0 , 1 ] + Min N N S; //Dedicate K MR = [ ∝ × Current solution ′ s K Total l o f al l sol ution ′ s K ] × ( V a r hi − V a r low ) ; //Define MR for K times do for MR times do Select an item randomly from among the sensitive items ; Set Rand [0 ,1] to the selected item; end for end for end for call FitnessFunction; call BestSolutionFunction; Limit number of solutions to N max ; for each solution in population call ImmigrationFunction; // Move all solutions toward the best solution end for call FitnessFunction; call BestSolutionFunction; until the termination condition is satisfied; end. Algorithm 2 Fitness function. Begin for each solution in population Merge with noncritical transactions; Calculate minfit1 ; // Eq. (5) Calculate minfit2 ; // Eq. (8) Calculate minfit3 ; // Eq. (11) end for end. w | w c r m w R R r r R a r i f m t a A fi s t r t 4 p s a t solution is an array the length of which equals the total length of critical transactions. 4.3. Evaluate the fitness In this part, fitness values are calculated for each existing so- lution in the initial population. To do so, it is necessary that each solution is merged with that part of the database which was separated in the preprocess (non-critical transactions) so that sanitization effect is considered for the whole database, and ac- quired fitness values would be representative of overall state of the database. It should be mentioned that calculation of fitness values is not necessary for the first solution in the initial popu- lation; for, as noted in the previous section, the first solution is the original database which was provided (unchanged) in the ini- tial population. The fitness values for other solutions are calculated by Eqs. (5), (8) and (11) which come as below: min f it1 = | HF | , (5) where |HF| is the number of hiding failure which is identified by Eq. (6) | HF | = | R s | ∑ i =1 r i state, (6) where | R s | is the number of sensitive rules and r i state is calculated by Eq. (7) r i state = { 0 i f Su p ( I i ) < MST or Con f ( r i ) < MCT 1 otherwise , (7) where I i is the itemset of r i . The second fitness value is computed by Eq. (8) min f it2 = | LR | , (8) here | LR | is the number of lost rules which is defined by Eq. (9) LR | = | ∼R s | ∑ i =1 r i state, (9) here | ∼R s | is the number of non-sensitive rules and r i state is alculated by Eq. (10) i state = { 0 i f Sup ( I i ) ≥ MST and Con f ( r i ) ≥ MCT 1 otherwise . (10) The third fitness value is identified by Eq. (11) in f it3 = RHD + RLD, (11) here RHD is Rules Hiding Distances, RLD Rules Lost Distances and HD is calculated by Eq. (12) HD = | R s | ∑ i =1 r i HD, (12) i HD is defined by Eq. (13) i HD = { 0 i f Sup ( I i ) < MST 0 else i f Con f ( r i ) < MCT C on f ( r i ) − MC T + 1 otherwise , (13) RLD is computed by Eq. (14) LD = | ∼R s | ∑ i =1 r i LD, (14) nd r i LD is calculated by Eq. (15) i LD = { 0 i f Sup ( I i ) ≥MST and C on f ( r i ) ≥MC T MC T − C on f ( r i ) else i f Con f ( r i ) < MCT MST − Sup ( I i ) otherwise . (15) The process of mining of all existing solutions in a population, n each iteration is a very time consuming operation. Hence the unctions minfit 1, minfit 2 and minfit 3 have been defined in such a anner where the mining of the solutions is not necessary. In fact, hese functions are able to determine the number of hiding failures nd lost rules for each solution without the need of data mining. s a result in COA4ARH, data mining is performed only twice. The rst time, the original database is mined in order to determine the ensitive and non-sensitive rules. The second time the final sani- ized database is mined so that except for hiding failures and lost ules, the number of ghost rules can also be defined. Fitness func- ion is shown in Algorithm 2 . .4. Select the best solution After calculating fitness of each solution it is necessary to com- are all fitness values in order to select the best solution. The first olution in the initial population does not have any fitness value nd hence will not be participated in the comparison. In order o select the best solution, all of the solutions are first sorted in M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 345 Algorithm 3 Best solution function. Begin Sort the solutions in increasing order of their minfit1 ; if two or more solutions have same minfit1 then Sort those in increasing order of their minfit2 ; if two or more solutions have same minfit2 then Sort those in increasing order of their minfit3 ; end if end if set the first member of sorted list as the best solution; if (the fitness values of the best solution ≤ the fitness values of the previous best solution) return best solution; else return previous best solution; end. Algorithm 4 Immigration function. Input: Current Solution, Best Solution; // Current Solution refers to every and each one of the existing solutions in a population; // Best Solution is a solution with the best fitness value so far; Output: Next Solution; // Next Solution is a solution close to the best solution; begin Number of Different Items = Hamming distance (Current Solution, Best Solution); N = Rand (0, 1) × Number of Different Items; // N is the number of Current Solution’s items that should be modified; Selected Items = Select randomly N items from Different Items of Current Solution; Next Solution = 1 XOR each Selected Items in Current Solution; //Reduce hamming distance; end r w m fi w b t p s 4 t i s K w b l N l i o M w c r c a e o s a 4 c s m e 4 o l a i g t t p t t m ( t i d f c i i v e b s 4 o 1 M A c ising order based on the value of minfit 1 function. Then, those ith equal values of minfit 1 will be sorted in rising order based on infit 2 and finally, if two solutions are equal in the value of min- t 2, they will be sorted in rising order based on minfit 3. The result ould be a sorted list of solutions, the first member of which will e selected as the best solution. Then, the best gained solution in his step and the best solution of the previous one will be com- ared in order for the global best solution to be determined. Best olution function is shown in Algorithm 3 . .5. Generating new solutions Each existing solution should generate a number of new solu- ions. Therefore, according to Eq. (16) , a random K number is ded- cated to each solution. It is representative of the number of new olutions that should be generated by each parent solution. = ( Max N N S − Min N N S ) × Rand [ 0 , 1 ] + Min N N S, (16) here K is the number of new solutions that should be generated y each parent solution, Max NNS is Maximum Number of New So- utions (determined by user) and Min NNS is Minimum Number of ew Solutions (determined by user). For generating new solutions, each parent solution is only al- owed to change a limited number of its items or bits. This number s called Modifications Radius which is shown with MR .The value f MR for each parent solution is calculated by Eq. (17) . R = [ ∝ × Current solution ′ s K T otal l o f al l sol ution ′ s K ] × ( V a r hi − V a r low ) , (17) here MR is the number of each parent solution’s items that ould be changed and ∝ shows the number of total iterations of ight hand side items of sensitive rules in critical transactions, re- eived as algorithm input. Var hi and Var low are respectively high nd low limit of each variable in optimization problems. Consid- ring the fact that in the present problem high and low limits f the variables are respectively determined with 1 and 0, the (V a r hi − V a r low ) can be removed from the above equation. Given the value of K as well as the specified MR for each parent olution, new solutions are being developed. The process of gener- ting each new solution is as follows: • For each parent solution, a number (equal to it’s MR ) of items are randomly selected from among the sensitive items and ran- domly quantified (0 or 1). • Other items are transmitted from the parent solution into the new solution without any change. • These steps are iterated K times for each parent solution to which the value K is allocated. .6. Limit solutions’ maximum number The total number of the existing solutions should always be ontrolled in order not to exceed the N max threshold. For this rea- on, in each iteration, a number of worst solutions must be re- oved from the end of the sorted list, so that the number of the xisting solutions would remain equal to N max . .7. Immigration In order to improve the remaining solutions, with the purpose f improving the ability of the proposed algorithm to escape from ocal optimums and finding better solutions compared to what is lready gained, all solutions are changed in a way to be more sim- lar and closer to the best solution. This change can lead to the eneration of better solutions compared to the best current solu- ion. Nevertheless, the best current solution will be considered as he best global solution. Algorithm 4 shows this process. In order to clarify the functionality of this algorithm, an exam- le has been offered in Fig. 2. The Hamming distance is defined as the number of posi- ions at which the corresponding symbols are different between wo strings of equal length. For binary strings a and b the Ham- ing distance is equal to the number of ones in a XOR b Hamming, 1950 ). Based on this definition, the number of different items between he current solution and the best solution is equal to 7 which is llustrated in red as shown in Fig. 2 (a). The value of N can now be etermined which is equal to a random number of different items, or example 3. This means that 3 items of different items of the urrent solution should be modified. Hence from those 7 items, 3 tems will be chosen randomly. These items have been illustrated n blue in Fig. 2 (b). The value of each selected item should then be modified (the alue of selected item XOR 1). In this way, the next solution is gen- rated which is more similar to the best solution where the num- er of non-equal items between them have decreased. The next olution is illustrated in Fig. 2 (c). .8. Case study In this part an example is provided for a better understanding f the proposed algorithm. Assume an original database including 0 transactions that is shown in Table 2 . Considering MST = 40 and CT = 70, the mined rules from the database are shown in Table 3 . lso assume that the set of sensitive rules is defined as { b → c,b → e } to be hidden. 346 M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 Fig. 2. Moving a solution toward the best solution. Fig. 3. Preprocess of original dataset. Table 2 Original dataset. T id Items T 1 a, b, c, e T 2 e T 3 b, c, e, f T 4 d, f T 5 a, b, d T 6 b, c, e T 7 a, b, c, d, e T 8 a, b T 9 c, e T 10 a, b, c, e T o t E s m f S l a s The preprocess operation on the original database is shown in Fig. 3. Assume that the user has determined the N pop as 4 cuckoos. herefore, a random initial population can be shown as Fig. 4. As seen in Fig. 4 , the first solution is the critical transactions f the original database. The fitness values for other solutions af- er merging with non-critical transactions, can be calculated by qs. (5), (8) and (11) . For example the fitness values for C2 are hown as follow: infit 1 ( C2 ) = 0 minfit 2 ( C2 ) = 8 minfit 3 ( C2 ) = 2 . 56 After fitness calculation, the sorted list of solutions is like the ollowing. ortedList = { C4 , C2 , C3 } The first member of the list (C4) is determined as the best so- ution. Next stages are continued with determination of MR and llocation of K value to each solution and also generation of new olutions. M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 347 Fig. 4. Random initial population. Table 3 Association rules extracted from orig- inal dataset with MST = 40% and MCT = 70%. Rule Support Confidence a → b 50 100 b → a 50 71 .43 b → c 50 71 .43 c → b 50 83 .33 b → e 50 71 .43 e → b 50 71 .43 c → e 60 100 e → c 60 85 .71 b → ce 50 71 .43 c → be 50 83 .33 e → bc 50 71 .43 ce → b 50 83 .33 be → c 50 100 bc → e 50 100 a M s a l t w m v S l Table 4 Final sanitized dataset. T id Items T 1 a, b, c, e T 2 e T 3 b, c, e, f T 4 d, f T 5 a, b, d T 6 b, c, e T 7 a, b, c, d, e T 8 a, b T 9 c, e T 10 a, b, e b m h o S c i n e l S s n o t t 5 i Assume that the user has determined the MaxNNS and MinNNS s follow: axN N S = 3 , MinN N S = 1 , o the random values of K will be calculated as follow: K ( C1 ) = ( 3 − 1 ) × rand [ 0 , 1 ] + 1 = 2 , K ( C2 ) = ( 3 − 1 ) × rand [ 0 , 1 ] + 1 = 1 , K ( C3 ) = ( 3 − 1 ) × rand [ 0 , 1 ] + 1 = 1 , K ( C4 ) = ( 3 − 1 ) × rand [ 0 , 1 ] + 1 = 2 , nd the value of MR for each parent solution is calculated as fol- ow: α = 10 , MR ( C1 ) = [ 10 × 2 6 ] = 3 , MR ( C2 ) = [ 10 × 1 6 ] = 2 , MR ( C3 ) = [ 10 × 1 6 ] = 2 , MR ( C4 ) = [ 10 × 2 6 ] = 3 , hen the new solutions are generated as shown in Fig. 5 . The items hich are randomly changed are highlighted with black. The fitness values for all new solutions are calculated (after erging with non-critical transactions) and sorted by the fitness alues. The sorted list represented as follow: ortedList = { C11 , C31 , C12 , C21 , C41 , C42 } The first member of the list (C11) is determined as the best so- ution and compared with the previous best solution (C4). Since oth of them have equal fitness values, the best solution still re- ains C11. Assume that the user has determined the N max as 4, ence 2 number of the worst solutions are deleted from the end f the sorted list. ortedList = { C11 , C31 , C12 , C21 } . In the next step, the Number of Deferent Items and N are cal- ulated as follow: Number of Diferent Items = Hamming Distance (C31, C11) = 4, N = Rand (0, 1) × 4 = 1 Number of Diferent Items = Hamming Distance (C12, C11) = 3, N = Rand (0, 1) × 3 = 2 Number of Diferent Items = Hamming Distance (C21, C11) = 2, N = Rand (0, 1) × 2 = 1 Immigrated solutions toward the best solution (C11) is shown n Fig. 6 . The changed items are illustrated in black. In this stage again each immigrated solution is merged with on-critical transactions and the fitness values are calculated for ach of them, eventually the sorted list of solutions is like the fol- owing. ortedList = { C ′ 12 , C11 , C ′ 21 , C ′ 31 } The first member of the list (C’12) will be selected as the best olution and compared whit C11 (the previous best solution), fi- ally C’12 is determined as the best solution in the first iteration f the algorithm. Next iteration is started with determination of MR and alloca- ion of K value to each solution and also generation of new solu- ions. The sanitized dataset after 2 iterations is shown in Table 4. . Performance evaluation In order to evaluate the efficiency of proposed algorithm in san- tization of Real and Synthetic databases, some experiments were 348 M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 Fig. 5. Generate new solutions. Fig. 6. Immigrate toward the best solution. Table 5 Database characteristics. Database No. of transactions Avg. trans. length No. of items Mushroom 8124 23 119 Chess 3196 37 75 T100 100 19 37 5 m f u a v i h conducted on the proposed algorithm and three others namely 1.b( Verykios et al., 2004 ), DSR( Wang et al., 2007 ), and the Im- proved Genetic ( Khan et al., 2014 ) algorithms, which are discussed here. 5.1. Datasets Three databases were used in the experiments. Chess and Mushroom databases were used as real data and T100 database was used as synthetic data. Characteristics of the specified datasets are presented in Table 5. 5.2. Performance measures The efficiency criteria that will be used in comparison of the proposed algorithm with 1.b, DSR and Improved Genetic algo- rithms include: • Hiding Failure ( HF ) : indicates the number of sensitive rules which sanitization algorithm could not hide and are still mined from the sanitized database D ′ . It is calculated by Eq. (18) . HF = ∣∣R S (D ′ )∣∣ | R S ( D ) | , (18) where, | R S ( D ′ )| is the number of sensitive rules explored in the sanitized database D ′ and | R S ( D )| is the number of sensitive rules explored in the original database D . • Lost Rules ( LR ) : indicates the number of non-sensitive rules that are lost due to the act of sanitization and will not be mined from the sanitized database D ′ . It is calculated by Eq. (19) . LR = | ∼ R S ( D ) | − ∣∣∼ R S (D ′ )∣∣ | ∼ R S ( D ) | , (19) where | ∼ R S ( D )| is the number of non-sensitive rules explored in the original database D and | ∼ R S ( D ′ )| is the number of non- sensitive rules explored in the sanitized database D ′ . • Ghost Rules ( GR ): indicates the number of artificial rules which were not within the original database and are generated due to the act of sanitization and are mined from database D ′ .It is calculated by Eq. (20) . GR = | R ′ | − | R ∩ R ′ | | R ′ | , (20) where | R ′ | is the number of mined rules from D ′ and | R | is he number of mined rules from D . • Number of Iterations: the number of iterations required to at- tain the optimal solution. .3. Experimental results In conducting the experiments, all the algorithms were imple- ented with C# in a same platform. The experiments were per- ormed on a PC with 2.20 GHz @Intel core i7 CPU and 6 GB of RAM nder the Windows 8 (64-bit). The results are shown in Figs. 7 , 8 nd 9. As it is illustrated in Figs. 7, 8 and 9 in all three data bases, the alue of the HF in the proposed algorithm is equal to zero. This s due to an appropriate definition of the fitness function and also igh prioritizing of the minfit 1 function. Further by not using the M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 349 0 0.2 0.4 0.6 0.8 1 Improved Gene�c DSR1.BCOA4ARH Si d e Eff ec ts V al u es The Algorithms HF LR GR Fig. 7. Comparative evaluation of the algorithms on Mushroom dataset. 0 0.2 0.4 0.6 Improved Gene�c DSR1.BCOA4ARH Si d e Eff ec ts V al u es The Algorithms HF LR GR Fig. 8. Comparative evaluation of the algorithms on Chess dataset. 0 0.1 0.2 0.3 0.4 0.5 0.6 Improved Gene�c DSR1.BCOA4ARH The Algorithms HF LR GR Fig. 9. Comparative evaluation of the algorithms on Synthetic dataset. i w v F m o m i p n c a 5 a t n t t s -10% 0% 10% 20% 30% 40% 50% 0 10 20 30 40 50 60 70 80 90 100 H id in g Fa ilu re Itera�on Result on Mushroom Database COA4ARH Improved Gene�c Fig. 10. Comparing the number of hiding failures in each iteration on Mushroom database. 0% 20% 40% 60% 80% 100% 0 10 20 30 40 50 60 70 80 90 100 Lo st R u le Itera�on Result on Mushroom Database COA4ARH Improved Gene�c Fig. 11. Comparing the number of lost rules in each iteration on Mushroom database. -20% 0% 20% 40% 60% 80% 0 10 20 30 40 50 60 70 80 90 100 H id in g Fa ilu re Itera�on Result on Chess Database COA4ARH Improved Gene�c Fig. 12. Comparing the number of hiding failures in each iteration on Chess database. 0% 20% 40% 60% 80% 100% 120% 0 10 20 30 40 50 60 70 80 90 100 Lo st R u le Itera�on Result on Chess Database COA4ARH Improved Gene�c Fig. 13. Comparing the number of lost rules in each iteration on Chess database. -20% 0% 20% 40% 60% 80% 0 10 20 30 40 50 60 70 80 90 100 H id in g Fa ilu re Itera�on Result on Synthe�c Database COA4ARH Improved Gene�c Fig. 14. Comparing the number of hiding failures in each iteration on synthetic database. nsertion technique in the proposed algorithm and selecting items ith minimum frequency in non-sensitive rules for deletion, the alue of GR in this algorithm will be zero or only a small amount. urthermore as it is shown in Figs. 7, 8 and 9 , there is an improve- ent in the value of LR in the proposed algorithm compared to the ther algorithms, which is the result of selecting items with the inimum frequency in non-sensitive rules for deletion. As shown n Fig. 9 , the proposed algorithm has a higher value of LR com- ared to the DSR algorithm, this is since the proposed algorithm eeds to apply more deletion to decrease the HF to zero, hence ompared to DSR with higher HF , the proposed algorithm will have higher value of LR . .3.1. Comparing the number of iteration One of the most important evaluation criteria in meta-heuristic lgorithms is the number of iterations required to attain the op- imal solution. Since measures have been taken to minimize the umber of iterations in COA4ARH algorithm, the comparison be- ween this algorithm and Improved Genetic algorithm is done on hree databases in Table 5 . The results of these comparisons are een in Figs. 10 to 15. 350 M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 0% 20% 40% 60% 80% 100% 120% 0 10 20 30 40 50 60 70 80 90 100 Lo st R u le Itera�on Result on Synthe�c Database COA4ARH Improved Gene�c Fig. 15. Comparing the number of lost rules in each iteration on synthetic database. d G c H c i t a o s p b r b a t e R A D D G G G H K M M M M O O R S S S S As it can be seen in these figures, COA4ARH algorithm has been converged with a high speed and in the number of iteration less than 10 can attain the optimal solution. Furthermore, this algo- rithm showed better results compared to the other algorithm even in the very low number of iteration. Due to the way preprocessing is defined, significant decline occurred in the number of required iteration in COA4ARH algorithm. The operation is defined in a way that only the items involved in hiding sensitive rules are chosen to be removed. As a result, first, this prevents the production of use- less solution and increases the chance of suitable solution produc- tion; second, the number of possible modes to produce a solution reduces significantly. Because in a string length of L the placement modes of 0 and 1 is 2 L . In fact, preprocessing reduces L size sig- nificantly by choosing sensitive items to remove; as a result, the number of cases for attaining the result will be reduced. Due to the manipulation of all the items of the database in Im- proved genetic algorithm, useless solutions are produced that de- crease the chance of producing more effective solutions and in- crease the time needed to attain optimal solution. As it is seen in Figs 10 –15 , this algorithm has been converged in the number of iterations over 40. 6. The effects of sparse dataset The existence of sparse datasets as input data in proposed al- gorithm make extracted association rules to have a lower degree of overlap. Therefore, sensitive rules elected for deletion will also have little overlap. The low degree of overlap suggests that the number of frequent items between sensitive rules is low so the number of sensitive items of the algorithm chosen to be deleted increase to zero the hiding failure which has the highest priority in side effects. This factor will lead to an increase in the number of lost rules. Consequently, it can generally be concluded that the presence of sparse datasets in the proposed algorithm input data increase the number of lost rules. 7. Conclusions and future works The present article has proposed a new meta heuristic method for hiding sensitive association rules using cuckoo optimization al- gorithm. The proposed algorithm is capable of simultaneously hid- ing several sensitive association rules. The most important and in- fluential feature of the present study is defining a preprocess op- eration which includes two phases in the beginning of proposed algorithm. This preprocess operation causes a remarkable decrease in the number of iterations and speedy access to the optimal so- lution. In this study, three fitness functions have also been intro- duced which can find the solution with minimum side effects. Fur- ther an immigration algorithm has been defined which improved the ability of the proposed algorithm to escape from local opti- mums. The proposed approach was compared to the three algo- rithms 1.b, DSR and Improved Genetic. For efficiency assessment, all of the algorithms were examined on both Real and Synthetic ata. The results were analyzed based on the four criteria HF, LR, R and the number of iteration. The results show a higher effi- iency for the proposed algorithm compared to the others. Since, F was 0, GR was close to 0 and LR had a remarkable decrease ompared to other algorithms. Moreover the proposed algorithm s converged with a high speed and in the lower number of itera- ion can attain the better solution compared to Improved Genetic lgorithm. Defining a new fitness function that can decrease the amount f Lost Rules and preserve the algorithm’s capability of hiding sen- itive rules and avoiding generation of ghost rules will be the pur- ose of future studies. Also, through some computations, the num- er of sensitive items that should be deleted for hiding sensitive ules can be calculated; only this number of sensitive items should e deleted to decrease the number of lost rules. Moreover, a self- daptive mechanism will be added to the proposed algorithm. In his mechanism, algorithm input parameters will be changed by ach iteration and will be moved toward being optimized. eferences tallah, M. , Bertino, E. , Elmagarmid, A. , Ibrahim, M. , & Verykios, V. (1999). Disclo- sure limitation of sensitive rules. In IEEE 1999 knowledge and data engineering exchange (pp. 45–52). IEEE . asseni, E. , Verykios, V. S. , Elmagarmid, A. K. , & Bertino, E. (2001). Hiding as- sociation rules by using confidence and support. In Information hiding: 2137 (pp. 369–383). Pittsburgh, PA, USA: Springer Berlin Heidelberg . ehkordi, M. N. , Badie, K. , & Zadeh, A. K. (2009). A novel method for privacy pre- serving in association rule mining based on genetic algorithms. Journal of Soft- ware, 4 , 555–562 . koulalas-Divanis, A. , & Verykios, V. S. (2006). An integer programming approach for frequent itemset hiding. In Proceedings of the 15th ACM international confer- ence on information and knowledge management (pp. 748–757). ACM . koulalas-Divanis, A. , & Verykios, V. S. (2009a). Exact knowledge hiding through database extension. IEEE Transactions on Knowledge and Data Engineering, 21 , 699–713 . koulalas-Divanis, A. , & Verykios, V. S. (2009b). Hiding sensitive knowledge without side effects. Knowledge and Information Systems, 20 , 263–299 . amming, R. W. (1950). Error detecting and error correcting codes. Bell System Tech- nical Journal, 29 , 147–160 . han, A. , Qureshi, M. S. , & Hussain, A. (2014). Improved genetic algorithm ap- proach for sensitive association rules hiding. World Applied Sciences Journal, 31 , 2087–2092 . enon, S. , & Sarkar, S. (2007). Minimizing information loss and preserving privacy. Management Science, 53 , 101–116 . enon, S. , Sarkar, S. , & Mukherjee, S. (2005). Maximizing accuracy of shared databases when concealing sensitive patterns. Information Systems Research, 16 , 256–270 . oustakides, G. V. , & Verykios, V. S. (2006). A max-min approach for hiding fre- quent itemsets. In Proceedings of the sixth IEEE international conference on data mining-workshop (pp. 502–506) . oustakides, G. V. , & Verykios, V. S. (2008). A max-min approach for hiding fre- quent itemsets. Data & Knowledge Engineering, 65 , 75–89 . liveira, S. R. , & Zaïane, O. R. (2002). Privacy preserving frequent itemset min- ing. In IEEE 2002 ICDM workshop on privacy, security and data mining: vol. 14 (pp. 43–54). Maebashi, Japan: Australian Computer Society, Inc . Oliveira, S. R. , & Zaïane, O. R. (2003a). Algorithms for balancing privacy and knowledge discovery in association rule mining. In IEEE 2003 7th international database engineering and applications symposium (pp. 54–63). IEEE . liveira, S. R. , & Zaïane, O. R. (2003b). Protecting sensitive knowledge by data san- itization. In IEEE 2003 3th international conference data mining (pp. 613–616). IEEE . Oliveira, S. R. , & Zaïane, O. R. (2006). A unified framework for protecting sensitive association rules in business collaboration. International Journal of Business In- telligence and Data Mining, 1 , 247–287 . ajabioun, R. (2011). Cuckoo optimization algorithm. Applied Soft Computing, 11 , 5508–5518 . aygin, Y. , Verykios, V. S. , & Clifton, C. (2001). Using unknowns to prevent discovery of association rules. ACM SIGMOD Record, 30 , 45–54 . aygin, Y. , Verykios, V. S. , & Elmagarmid, A. K. (2002). Privacy preserving association rule mining. In IEEE 2002 12th international workshop on research issues in data engineering: engineering e-commerce/e-business systems (pp. 151–158). IEEE . un, X. , & Philip, S. Y. (2007). Hiding sensitive frequent itemsets by a border-based approach. Journal of Computing Science and Engineering, 1 , 74–94 . un, X. , & Yu, P. S. (2005). A border-based approach for hiding sensitive frequent itemsets. In IEEE 2005 5th international data mining conference (pp. 426–433). IEEE . Verykios, V. S. , Elmagarmid, A. K. , Bertino, E. , Saygin, Y. , & Dasseni, E. (2004). As- sociation rule hiding. IEEE Transactions on Knowledge and Data Engineering, 16 , 434–447 . http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0002 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0002 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0002 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0002 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0002 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0002 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0003 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0003 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0003 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0003 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0003 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0004 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0004 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0004 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0004 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0007 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0007 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0010 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0010 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0010 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0010 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0010 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0011 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0011 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0011 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0011 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0012 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0012 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0012 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0012 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0013 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0013 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0013 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0013 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0016 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0016 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0016 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0016 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0017 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0017 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0018 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0018 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0018 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0018 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0018 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0022 M.H. Afshari et al. / Expert Systems With Applications 64 (2016) 340–351 351 V W W W Y Y erykios, V. S. , Pontikakis, E. D. , Theodoridis, Y. , & Chang, L. (2007). Efficient al- gorithms for distortion and blocking techniques in association rule hiding. Dis- tributed and Parallel Databases, 22 , 85–104 . alton, S. , Hassan, O. , Morgan, K. , & Brown, M. (2011). Modified cuckoo search: A new gradient free optimisation algorithm. Chaos, Solitons & Fractals, 44 , 710–718 . ang, S.-L. , Parikh, B. , & Jafari, A. (2007). Hiding informative association rule sets. Expert Systems with Applications, 33 , 316–323 . u, Y.-H. , Chiang, C.-M. , & Chen, A. L. (2007). Hiding sensitive association rules with limited side effects. IEEE Transactions on Knowledge and Data Engineering, 19 , 29–42 . ang, X.-S. , & Deb, S. (2009). Cuckoo search via Lévy flights. In Nature & biologically inspired computing, 2009. NaBIC 2009. World congress on (pp. 210–214). IEEE . ang, X.-S. , & Deb, S. (2013). Multiobjective cuckoo search for design optimization. Computers & Operations Research, 40 , 1616–1624 . http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0023 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0023 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0023 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0023 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0023 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0023 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0027 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0027 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0027 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0027 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0028 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0028 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0028 http://refhub.elsevier.com/S0957-4174(16)30398-0/sbref0028 Association rule hiding using cuckoo optimization algorithm 1 Introduction 2 Background and related work 3 Problem definition 4 Proposed algorithm 4.1 Preprocess of original dataset 4.2 Generating initial population 4.3 Evaluate the fitness 4.4 Select the best solution 4.5 Generating new solutions 4.6 Limit solutions’ maximum number 4.7 Immigration 4.8 Case study 5 Performance evaluation 5.1 Datasets 5.2 Performance measures 5.3 Experimental results 5.3.1 Comparing the number of iteration 6 The effects of sparse dataset 7 Conclusions and future works References 
work_xxbtc7zagfg4tgpxdors5gbbfi	----	Learning feature-projection based classifiers Expert Systems with Applications 39 (2012) 4532–4544 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Learning feature-projection based classifiers Aynur Dayanik Department of Computer Engineering, Bilkent University, Bilkent 06800, Ankara, Turkey a r t i c l e i n f o Keywords: Classification learning Inductive learning Feature projections 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.09.133 E-mail address: adayanik@cs.bilkent.edu.tr a b s t r a c t This paper aims at designing better performing feature-projection based classification algorithms and presents two new such algorithms. These algorithms are batch supervised learning algorithms and rep- resent induced classification knowledge as feature intervals. In both algorithms, each feature participates in the classification by giving real-valued votes to classes. The prediction for an unseen example is the class receiving the highest vote. The first algorithm, OFP.MC, learns on each feature pairwise disjoint intervals which minimize feature classification error. The second algorithm, GFP.MC, constructs feature intervals by greedily improving the feature classification error. The new algorithms are empirically eval- uated on twenty datasets from the UCI repository and are compared with the existing feature-projection based classification algorithms (FIL.IF, VFI5, CFP, k-NNFP, and NBC). The experiments demonstrate that the OFP.MC algorithm outperforms other feature-projection based classification algorithms. The GFP.MC algorithm is slightly inferior to the OFP.MC algorithm, but, if it is used for datasets with large number of instances, then it reduces the space requirement of the OFP.MC algorithm. The new algorithms are insen- sitive to boundary noise unlike the other feature-projection based classification algorithms considered here. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Classifier learning is one of the central problems in machine learning and data mining. Medical doctors, marketing profession- als, curators of biological journal articles find rule-based classifiers more comprehensible because their domain knowledge becomes more transparent in those classifiers’ data representation and deci- sions. Most practitioners already have some expectations about each feature’s power to explain the variations in data. Therefore, they are more confident in using simple rules directly based on fea- tures, rather than on their complex combinations or transforma- tions. Feature-projection based classifiers summarize labeled data separately on each feature dimension by means of histograms like Naive Bayes classifier or by pairwise disjoint intervals like FIL.IF, VFI5, and CFP, which are explained in Section 2, and classify new instances using these data representations. Therefore, those classi- fiers are simple but effective and intuitively appealing tools for the practitioners. In this paper, we aim at designing new feature-pro- jection based classifiers which perform better than those previ- ously proposed similar classifiers. Some of the feature-projection based algorithms are feature interval learning (FIL.IF) (Dayanik, 2010), voting feature intervals (VFI5) (Demiröz & Güvenir, 1997; Güvenir, Demiröz, & Ilter, 1998; Güvenir & Emeksiz, 2000), k-nearest neighbor on feature ll rights reserved. projections (k-NNFP) (Akkus� & Güvenir, 1996), classification by feature partitioning (CFP) (Güvenir & S�irin, 1996), and Naive Bayes classifier (NBC) (Domingos & Pazzani, 1997; Mitchell, 1997; Witten & Frank, 2005), which are overviewed in Section 2. The k-NNFP algorithm uses feature projections to find the nearest observations on each feature dimension and then combines the predictions of the features to make final decision. The CFP, VFI5, and FIL.IF algo- rithms represent concept descriptions in the form of feature inter- vals, which are learned separately for each feature. The main difference among those algorithms is the way they construct the feature intervals. Naive Bayes classifier assumes that features are conditionally independent given the class, and uses the Bayes the- orem to estimate the probability of each class given the feature val- ues of the new example to be classified. Feature-projection based learning algorithms are simple, effective, efficient and robust to irrelevant features. They yield plausible concept descriptions, enable fast classification, do not require normalization of feature values, and handle missing feature values by simply neglecting them. In this paper, we propose two new feature-projection based classification algorithms. They are batch supervised learning algo- rithms and represent the induced classification knowledge as a set of pairwise disjoint intervals on each feature dimension. In both algorithms, each feature participates in the classification by giving real-valued votes to classes. The prediction for an unseen example is the class receiving the highest vote. The first algorithm, OFP.MC http://dx.doi.org/10.1016/j.eswa.2011.09.133 mailto:adayanik@cs.bilkent.edu.tr http://dx.doi.org/10.1016/j.eswa.2011.09.133 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 4533 (learning by optimal feature partitioning based on misclassification rates), learns feature intervals separately for each feature by opti- mally partitioning feature values to minimize the feature classifi- cation error. The second algorithm, GFP.MC (learning by greedy feature partitioning based on misclassification rates), constructs feature intervals in a greedy manner; it starts with point intervals and merges two consecutive intervals which results in the smallest immediate increase in the feature classification error, and repeats the latter until target number of intervals is achieved. The number of intervals for the final model is determined by comparing the fea- ture classification errors for different number of intervals. Both algorithms learn the partitions for every fixed number of intervals on some training set. True misclassification error is then estimated on a separate validation set. This is repeated five times—for each fold of 5-fold cross-validation—and the best num- ber of intervals is found by a one-standard-error rule. OFP.MC learns the best partition on each training set using dynamic pro- gramming, while GFP.MC combines consecutive intervals greedily until the desired number of intervals is reached. In both algo- rithms, each feature participates in the classification by giving real-valued votes to classes. The prediction for an unseen example is the class receiving the highest vote. In this paper, we propose two new classification algorithms using feature projections of the training data. On each feature dimension, classification knowledge is learned by partitioning features into pairwise disjoint intervals. Feature partitioning is related to discret- ization of linear (continuous) features (Dougherty, Kohavi, & Sah- ami, 1995; Liu, Hussain, Tan, & Dash, 2002; Yang & Webb, 2009). Discretizing a linear feature transforms it into a nominal (categori- cal) feature, where each value of the nominal feature corresponds to an interval of values of the linear feature. However, our focus is not on discretizing the linear features; we aim to develop better-per- forming feature-projection based classification algorithms. The new algorithms are empirically evaluated on twenty datasets from the UCI repository and are compared with the existing feature- projection based classification algorithms (FIL.IF, VFI5, CFP, k-NNFP and NBC). The experiments demonstrate that the new algorithm OFP.MC outperforms other feature-projection based algorithms. The GFP.MC algorithm is slightly inferior to the OFP.MC algorithm, but, if it is used for datasets with large number of instances, then it reduces the space requirement of the OFP.MC algorithm. The new algorithms are also insensitive to boundary noise unlike the fea- ture-projection based algorithms VFI5, CFP, FIL.IF, and NBC. The next section starts with a review of the previous work on feature-projection based classification algorithms. Section 3 pre- sents the optimal feature partitioning learning algorithm for classi- fication. Optimal partitioning of linear features are described and illustrated on the iris dataset from the UCI repository in Section 4. Section 5 presents the greedy feature partitioning learning algo- rithm for classification. Section 6 presents the empirical evaluation of the algorithms on real-world datasets. The new algorithms are compared to the previously proposed feature-projection based algorithms. The complexity analysis of the algorithms is given in Section 7. Section 8 concludes with some remarks and suggestions for future research. 2. Previous work on feature-projection based algorithms Feature-projection based classification algorithms represent the instances with their projections on feature dimensions. Some of the previously proposed such algorithms are feature interval learn- ing (FIL.IF) (Dayanik, 2010), voting feature intervals (VFI5) (Demi- röz & Güvenir, 1997; Güvenir et al., 1998; Güvenir & Emeksiz, 2000), k-nearest neighbor on feature projections (k-NNFP) (Akkus� & Güvenir, 1996), and classification by feature partitioning (CFP) (Güvenir & S�irin, 1996). These algorithms were shown to be fast, accurate, and robust to irrelevant features. The CFP algorithm represents classification knowledge as sets of disjoint feature segments. It partitions the feature values into seg- ments that are generalized or specialized as the training instances are processed. It constructs the segments incrementally, and the resulting concept descriptions are sensitive to the processing order of the training instances. The VFI5 algorithm learns classification knowledge in batch mode and represents it as multi-class feature intervals. It finds on each linear feature and for each class the class endpoints, which are the lowest and highest feature values with the same class la- bels. Pairwise disjoint intervals are constructed at every endpoint and between every pair of the closest endpoints of possibly differ- ent classes. If the feature is nominal, then every distinct point forms a point interval. The VFI5 algorithm has been implemented in the WEKA machine learning workbench (Hall et al., 2009), and has been used in several medical applications (Güvenir et al., 1998; Güvenir & Emeksiz, 2000). The FIL.IF algorithm is a batch supervised learning algorithm, which learns the concept descriptions in the form of a set of dis- joint intervals separately for each feature. On every linear feature, a point interval is constructed at each observed feature value. The algorithm then generalizes the point intervals into the range inter- vals by merging all of the neighboring single-class point intervals with the same class labels. However, multi-class point intervals are left alone, and at the end of the training process, they are con- verted to single-class point intervals by selecting the classes with the highest relative representativeness values as their class labels. The feature intervals are always disjoint. Nominal features have only (possibly multi-class) point intervals. The k-NNFP algorithm represents instances as separate collec- tions of feature projections. Given a new example, it uses feature projections to find the nearest observations on each feature dimen- sion. Each feature has exactly k votes and distributes them be- tween the classes of the k-nearest training observations. The votes of the individual features are summed; the class with the highest vote is predicted to be the class of the new example. All feature-projection based algorithms predict the class of a new example as the label with the highest total weighted votes of the individual features. These algorithms allow faster classifica- tion than other instance-based learning algorithms because sepa- rate feature projections can be organized for faster classification. These algorithms easily and naturally handle the missing feature values by neglecting the class votes from the missing features. The major drawback of these algorithms is that the descriptions involving a conjunction between two or more features cannot be represented. Despite of that, they are still reported to be successful on real-world datasets (Akkus� & Güvenir, 1996; Dayanik, 2010; Güvenir et al., 1998; Güvenir & S�irin, 1996). Bayesian classifiers from pattern recognition are based on prob- abilistic approaches to inductive learning in statistical concept learning tasks. The method estimates the posterior class probabil- ity of an instance given its observed feature values. The class is pre- dicted as the label with the highest estimated posterior probability (Duda, Hart, & Stork, 2000; Fukunaga, 1990). Bayesian classifiers assume in general that features are not statistically independent unlike Naive Bayesian classifier (NBC) (Mitchell, 1997; Witten & Frank, 2005). NBC has been successfully used in a wide variety of problem domains (Chandra & Gupta, 2011; Chena, Huanga, Tiana, & Qua, 2009; Hsu, Huang, & Chang, 2008; Kim, Han, Rim, & Mya- eng, 2006; Lewis, 1998), and has been shown to perform well in many domains even when the independence assumption is vio- lated (Domingos & Pazzani, 1997). Domingos and Pazzani establish the necessary and sufficient conditions for the optimality of NBC under zero-one loss (Domingos & Pazzani, 1997). 4534 A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 3. OFP.MC: Optimal feature partitioning learning algorithm for classification Here we describe the new feature-projection based algorithm OFP.MC, which generalizes the class information on every feature dimension in the form of pairwise disjoint intervals. The algorithm aims at capturing the relation between the class labels and feature values with the best possible choices of the intervals. Therefore, the feature partition intervals are chosen so as to minimize the total number of misclassifications on the train- ing dataset. In order to prevent the overfitting, the set of labeled examples is divided into training and validation sets. Optimal feature partitions are learned from the training set, but their unknown true mis- classification rates on the unseen data are estimated from the test results on the validation set. In order to reduce the potential biases possibly introduced with the arbitrary division of the labeled examples into training and valida- tion sets, the average misclassification rates and their standard errors are recalculated by a fivefold cross-validation. The final choices of the optimal number of feature intervals are determined with a one-stan- dard-error rule. The optimal feature partitions with the optimal num- ber of feature intervals are finally found by minimizing the total number of misclassifications on the entire set of labeled examples. For every feature interval in the optimal feature partition, we calcu- late the end-points, the class weights, and the class prediction, all of which are needed for classification of new examples. The independent optimizations of the number and breakpoints of the feature intervals balance the trade-off between model bias and model variance. On the one hand, the optimal choice of the interval breakpoints on the training data explains the labeled examples in the best possible way and reduces the final model’s bias. On the other hand, the optimal choice of the number of inter- vals on the validation set eliminates the unnecessarily complex models and reduces the final model’s variance. The classification with OFP.MC is simple and fast. For each non- missing feature value of a given test example, the interval into which the feature value falls is identified. Then the class weights of those intervals are summed, and the class with the largest total weight is predicted. The training and classification algorithms are precisely stated in Figs. 1 and 2, respectively. The numerical results and comparisons to other feature-projection based learning algo- rithms are discussed in Section 6. In the next section, the details of the training in OFP.MC algo- rithm are explained, and both its training and classification pro- cesses are illustrated on the iris dataset from the UCI repository. 4. Optimal partitioning of linear features in the training of OFP.MC Suppose that a set of labeled examples with several features is given. The examples have exactly one of L distinct class labels. For every fixed feature, we solve the following optimal feature parti- tioning problem. Suppose that N distinct feature values b1 < � � � < bN are observed, at each of which there may be multiple instances with the same and/or different class labels. The optimal feature partitioning prob- lem is (i) to attach class labels to b1, . . . , bN so as to minimize the total number of misclassified examples on the same feature, with at most k label switchings along the same feature dimension, and at the same time, (ii) to choose the maximum number of label switchings 0 6 k 6 N � 1 in order to minimize the generalization error. The consecutive feature values with the same labels are later grouped together to form at most k + 1 pairwise disjoint feature intervals. This problem is solved in two stages. In the first stage, for every fixed integer 0 6 k 6 N � 1, an optimal labeling of the feature val- ues with at most k label switchings is found on some training data. The misclassification rate of each labeling is then evaluated on an independent validation set. While the misclassification rate on the same training set is al- ways a nonincreasing function of k, it typically decreases first and increases later with k on the validation data. To avoid over-fit- ting to data, optimal k is chosen based on the misclassification rates estimated on the validation set. The sampling errors in the estimates of the misclassification rates are likely to obscure the real differences between the unob- served true misclassification rates for different k values. In order to reduce the sampling variances of those estimates on the final choice of optimal maximum number of switchings k, both the opti- mal labeling of feature values with at most k label switchings and the independent assessment of its misclassification rate for each 0 6 k 6 N � 1 are repeated on several different disjoint training and validation sets, which are obtained in this paper from applying fivefold cross-validation to a given set of labeled examples. For every 0 6 k 6 N � 1, the sample average and sample stan- dard deviation of five misclassification rates are calculated. The average misclassification rates are plotted against k, and its mini- mum kmin is found. In general, the sampling errors may make the misclassification rates for the values of k near kmin look different, but this difference is often statistically insignificant. We therefore set our final choice kopt of maximum number of label switchings to the smallest of all k values with an average misclassification rate less than the (minimum) average misclassification rate plus its standard error at kopt. The main steps of the outlined method will now be given in more details. 4.1. Optimal feature partitioning by using dynamic programming On every fixed feature of a given training data, optimal labeling of N distinct feature values b1 < � � � < bN with at most k label switchings can be found simultaneously for all 0 6 k 6 N � 1 by using dynamic programming. Let us add an artificial point bN+1 > bN with no actual labeled data. For n = 1, . . . , N, ‘ = 1, . . . , L, and k = 0, 1, . . . , N � 1, let us define the value function vðn;‘; kÞ¼ minimum total number of misclassified examples on fb1; . . . ; bng obtainable by a labeling of b1; . . . ; bn if bnþ1 already has label ‘ and at most k abel switchings are allowed on fb1; . . . ; bng and the misclassification function eð‘; nÞ¼ the number of misclassified examples at bn if every example with feature value bn is assigned label ‘: The value function satisfies the boundary conditions vð1;‘; kÞ¼ eð‘; 1Þ; if k ¼ 0 min 16~‘6L eð~‘; 1Þ; if k > 0 8< : 9= ; for every ‘ ¼ 1; . . . ; L and k ¼ 0; 1; . . . ; N � 1: The minimum number of misclassifications achievable on the train- ing set by any labeling of the feature values with at most k label switchings equals Fig. 1. The training process of the OFP.MC algorithm. A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 4535 min 16‘6L vðN; ‘; kÞ for every k ¼ 0; 1; . . . ; N � 1; and we can find v(N,‘, k) for every ‘ = 1, . . . , L and k = 0, 1, . . . , N � 1 by recursively calculating the optimality equations vðn; ‘; kÞ¼ min � eð‘; nÞþ vðn � 1;‘; kÞ; min 16~‘6L;~‘–‘ ½eð~‘; nÞþ vðn � 1; ~‘; k � 1Þ� � for every n = 1, . . . , N, ‘ = 1, . . . , L, k = 0, 1, . . . , N � 1; see Fig. 3. The optimal label for bn when at most k label switchings is left for label- ing the feature values in {b1, . . . , bn} and bn+1 already has label ‘ is given by pðn;‘;kÞ¼ ‘; vðn;‘;kÞ�vðn�1;‘;kÞ¼eð‘;nÞ; arg min 16~‘6L;~‘–‘ eð~‘;nÞþvðn�1; ~‘;k�1Þ � � ; vðn;‘;kÞ�vðn�1;‘;kÞ< eð‘;nÞ: 8< : Fig. 2. The classification process of the OFP.MC algorithm. Fig. 3. The dynamic programming algorithm: we start in (a) with known boundary condition and proceed forward to determine the optimal labels of the feature values. As we find in (b) the optimal labeling of bn when bn+1 has label ‘ and at most k label switchings are left for {b1, . . . , bn}, we must compare the minimum remaining number of misclassifications under the two alternative decisions depicted by (c) and (d). 4536 A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 For every fixed k, optimal labels of b1, . . . , bN can now be deter- mined in backwards direction by means of the policy function p(�,�,�). Let kNþ1 ¼ k and ‘Nþ1 ¼ arg min 16‘6L vðN;‘; kNþ1Þ: After the optimal labels of bN+1, bN, . . . , bn+1 are found, then an opti- mal label of bn and the maximum remaining number of label switchings are, respectively, given by ‘n ¼ pðn;‘nþ1;knþ1Þ and kn ¼ knþ1 �1ð‘n – ‘nþ1Þ for every 1 6 n 6 N; where 1(a – b) equals 1 if a – b and 0 otherwise. Let us emphasize that optimal labeling ‘1, . . . ,‘N of feature val- ues, respectively, b1, . . . , bN with ‘‘at most k label switchings’’ will contain strictly less than k label switchings if this gives statistically same or strictly smaller overall misclassification rate than that of exactly k label switchings. The dynamic programming formulation favors the smallest possible number of label switchings (and num- ber of intervals) on each feature dimension among all choices which have statistically similar powers to explain the labeled examples. 4.2. The construction of optimal feature intervals corresponding to maximum k label switchings After optimal labels of feature values are found for any given fixed maximum number of label switchings k, we collapse all con- secutive feature values with identical class labels into one and the same interval and assign to the interval the same common class A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 4537 label of all member feature values. The left (respectively, right) boundary of the leftmost (respectively, rightmost) interval is set to minus (respectively, plus) infinity. The consecutive intervals are extended halfway toward each other with the midpoint in- cluded at random to one of those intervals. For each interval, we record its left- and right-endpoints, its class prediction, and for each class its class weight, which is calculated as the fraction of all examples of that class in the union of training and validation sets which fall into that interval. The interval’s class prediction is a class label with the highest class weight of the interval. 4.3. The performance evaluation of optimal feature intervals corresponding to maximum k label switchings For every maximum k number of label switchings, the dynamic programming solution readily provides the minimum misclassifi- cation rate min16‘6LvðN; ‘; kÞ number of training examples on the training set, but this is typically an optimistic estimate of the misclassification rate of the model on unseen examples. A more realistic estimate is obtained on an independent validation set, and it is calculated as the fraction of the examples of the validation set which have different class labels than class predictions of the fea- ture intervals into which they fall. 4.4. The calculation of the performance curve and optimal choice of the maximum number of label switchings A given set of labeled examples is divided into five approxi- mately equal parts. Each part is used exactly once as a validation set, while the union of remaining four parts is used as the training set. For each maximum number of label switchings k, optimal fea- ture intervals are learned from the training data and the perfor- mance (the misclassification rate) is calculated on the validation set. This is repeated five times, giving five estimates of the mis- classification rate of optimal feature intervals corresponding to each k. The average misclassification rate and its standard error are cal- culated for every k. The minimum average misclassification rate is attained at some kmin value. The final choice kopt of the maximum number of label switchings is set to the smallest of all those k val- ues having average misclassification rate less than or equal to the sum of the average misclassification rate and its standard error both corresponding to kmin. 4.5. The construction of the final feature intervals for classification After the final maximum number of label switchings kopt is deter- mined, we use the dynamic programming on the entire set of labeled examples to find the optimal labeling of the feature values with at most kopt label switchings, from which we construct the feature intervals. For each of those intervals we determine left- and right- endpoints, class prediction, and class weights as described earlier. 4.6. Illustration on the iris dataset The optimal partitioning of linear features is central to the training process of the new classification algorithm, which was listed in Fig. 1. Here we shall illustrate it on the iris dataset from the UCI repository (Asuncion & Newman, 2007). Fig. 4 displays the labeled examples projected on each of four feature dimensions. The average misclassification rates and their standard errors corresponding to optimal feature partitions with at most k label switchings are calculated for each feature by a fivefold cross-vali- dation. The second row in Fig. 5 presents the plots of the average misclassification rates and their one-standard-error confidence bounds for each value of k calculated on the validation sets. On every feature, we first find the maximum number of label switch- ings kmin at which the average misclassification rate is minimized. On each figure, kmin is marked by a vertical dashed line. The hori- zontal dashed line goes through the upper one-standard-error con- fidence bound of the average misclassification rate at kmin. The final maximum number of label switchings kopt is the smallest k value whose average misclassification rate is below the dashed horizontal one-standard-error line; its place is marked by a solid vertical line in each plot. Evident from Fig. 4, the labeled examples are more separable on features 3 and 4 than on 1 and 2. Therefore, the average misclassi- fication rates plotted in the second row of Fig. 5 decrease first and increase later on features 1 and 2. The standard errors of the aver- age misclassification rates are also larger on features 1 and 2 than on 3 and 4. The optimal kopt value coincides with the minimum kmin value on features 1, 3, and 4, but kopt is strictly less than kmin on feature 2. More precisely, the optimal number of label switch- ings kopt are 2, 1, 2, and 2 for features 1, 2, 3, and 4, respectively. By using the entire set of labeled examples and the dynamic pro- gramming algorithm, those features are partitioned optimally into at most kopt + 1, namely, 3, 2, 3, and 3 pairwise disjoint intervals, respectively. The second row in Fig. 6 shows the optimal partition- ing of the features into intervals. The vertical solid lines mark the boundaries of the intervals, and the class labels under the horizon- tal axes indicate the class predictions of the intervals. For each fea- ture interval, the class weights are calculated as in Table 1 and visually displayed by the barplots of Fig. 7. The first rows of Figs. 5 and 6 display, respectively, the mis- classification rates calculated on the training sets and the feature partitions found with the one-standard-error rule applied to the rates just described. The misclassification rates calculated in the first row of Fig. 5 on training sets are optimistic estimates of the misclassification rates of the OFP.MC learning algorithm on the unseen examples. These estimates are lower and have smaller confidence intervals than those calculated in the second row of Fig. 5 from the validation sets. Moreover, the plots in the first row of Fig. 5 are always nonincreasing. Therefore, the one-stan- dard-error rule typically returns large values of kopt and suggests that the features be partitioned into large number of intervals. However, when the corresponding optimal feature intervals are plotted in the first row of Fig. 6, one realizes that the additional intervals do not appear to capture any new and meaningful pat- terns present in the labeled examples. In other words, the number of intervals found with the one-standard-error rule applied to mis- classification rates calculated on the training set leads to a model which severely overfits to the data. Let us next illustrate how the model of feature intervals classi- fies the test example hð7:0; 3:2; 4:7; 1:4Þ; 1i; which has feature values 7.0, 3.2, 4.7, 1.4 on features 1, 2, 3, 4, respectively, and true class label 1. We start by finding on each fea- ture an interval which contains the feature value of the test exam- ple. We then collect the class weights of the intervals and sum the weights for each class. The class with the largest total weight is pre- dicted. Table 2 shows the details of the calculations. While features 1 and 2 predict classes 2 and 0, respectively, features 3 and 4 predict class 1. In the end the sum 2.18 of the weights of all four features decisively predict Class 0, which also turns out to be the true class label of the test instance. ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● 4.5 5.5 6.5 7.5 0 1 2 C la ss es Feature values Feature 1 ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ●●● ● ● ● ●● 2.0 2.5 3.0 3.5 4.0 Feature 2 ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ●● ● ● ● ● ● ●● ● ● 1 2 3 4 5 6 7 Feature 3 ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0.5 1.0 1.5 2.0 2.5 Feature 4 Fig. 4. Iris dataset. There are three classes 0, 1, and 2, which are jittered in order to identify distinct examples with the same feature values. Table 1 Class weights learned by OFP.MC algorithm for iris dataset. Feature 1 Feature 2 Intervals (�1, 5.45] (5.45, 6.15] (6.15,1) (�1, 3.05] (3.05,1) Class 0 0.91 0.08 0 0.09 0.61 Class 1 0.09 0.70 0.26 0.52 0.12 Class 2 0 0.21 0.74 0.39 0.27 Feature 3 Feature 4 Intervals (�1, 2.60] (2.60, 4.95] (4.95,1) (�1, 0.80] (0.80, 1.75] (1.75,1) Class 0 1 0 0 1 0 0 Class 1 0 0.90 0.05 0 0.90 0.03 Class 2 0 0.10 0.95 0 0.10 0.97 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 2 4 6 8 10 14 0.2 0.3 0.4 0.5 0.6 M is cl as si fic at io n ra te Maximum number of label switchings, k kopt = kmin Feature 1 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 5 10 15 kmin = 8kopt = 5 Feature 2 ● ● ● ● ● ● ● ● ● ● ● ● ● 0 2 4 6 8 10 12 kopt = kmin Feature 3 ● ● ● ● ● ● ● ● ● ● ● ● ● 0 2 4 6 8 10 12 kopt = kmin Feature 4 ● ● ● ● ● ● ● ● ● ● ● ● ● 0 2 4 6 8 10 12 0.2 0.3 0.4 0.5 0.6 0.7 0.8 M is cl as si fic at io n ra te Maximum number of label switchings, k kopt = kmin Feature 1 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 2 4 6 8 10 14 kmin = 4kopt = 1 Feature 2 ● ● ● ● ● ● ● ● ● ● ● ● ● 0 2 4 6 8 10 12 kopt = kmin Feature 3 ● ● ● ● ● ● ● ● ● ● ● ● ● 0 2 4 6 8 10 12 kopt = kmin Feature 4 Fig. 5. The curves of average misclassification rates and their one standard error confidence intervals calculated on training sets, above, and validation sets, below. 4538 A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 Fig. 6. Optimal feature partitions using misclassification rates on training sets, above, and validation sets, below. Interval boundaries and class predictions are shown. Additional intervals found above do not seem to capture new meaningful patterns. 0.0 0.2 0.4 0.6 0.8 1.0 C la ss w ei gh ts Feature values Class 0 Class 1 Class 2 5.45 6.15 Feature 1 3.05 Feature 2 2.6 4.95 Feature 3 0.8 1.75 Feature 4 Fig. 7. The barplots of the feature interval class weights reported in Table 1 for iris dataset. The bar heights of different colors equal the interval class weights of the corresponding classes. Features 3 and 4 seem to separate the labeled examples better than features 1 and 2. A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 4539 5. GFP.MC: greedy feature partitioning learning algorithm for classification For the classification problems with a lot of training instances, optimal partitioning of the features into disjoint intervals may require prohibitively large storage space for the value and policy functions. Table 2 Illustration of the classification of the test example h(7.0, 3.2, 4.7, 1.4), 1i. Bold numbers highlight the largest class weights on each feature and the overall weight of the winning class. Feature 1 Feature 2 Feature 3 Feature 4 Value 7.0 3.2 4.7 1.4 Interval (6.15,1) (3.05,1) (2.60, 4.95] (0.80, 1.75] Class weights Sum Prediction Class 0 0 0.61 0 0 0.61 Class 1 0.26 0.12 0.90 0.90 2.18 Class 1 Class 2 0.74 0.27 0.10 0.10 1.21 Here we would like to describe a greedy feature partitioning learning algorithm, GFP.MC, which requires significantly less storage space. The algorithm is suboptimal, as an example below shows, in the sense that it does not always find a partition with the smallest possible number of misclassifications on the training set, but still competes on the numerical examples with the optimal feature partitioning learning algorithm OFP.MC; see Table 4 and Fig. 10 of Section 6. In the training phase and on every feature dimension, the greedy algorithm initially forms one point interval at every feature value and assigns to it the label of the majority class at that feature value. Then neighboring intervals with the same interval class labels are merged. The number of feature intervals and the number of misclas- sifications on the independent validation set are counted and stored. There are always two neighboring intervals such that merging and relabeling them with the label of the majority class in the new interval increases the total misclassifications on the training data by the least possible number. We find any of those two neighboring intervals, merge them, and assign to the new interval the label of the majority class in that interval. The number of feature Fig. 8. The training process of the GFP.MC algorithm. 4540 A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 intervals and the number of misclassifications on the independent validation set are counted and stored. This procedure is repeated as long as there is at least one interval which can be eliminated. The entire procedure is repeated five times with fivefold cross- validation. The average number of misclassifications and their standard errors are calculated for each maximum number of fea- ture intervals. The greedy selection of the number of intervals for the final model is found by using the same one-standard-error rule described for the optimal feature partitioning algorithm OFP.MC in Section 4.4. The greedy partitioning algorithm is then run on the entire set of labeled examples until we find the first partition with the (largest) number of intervals less than or equal to the number found by the one-standard-error rule. For each interval in this final partition, the endpoints are extended, the class weights are calcu- lated, and the interval predictions are determined as in exactly the same way as it was done for the optimal feature partitioning learn- ing algorithm in Section 4.2. The procedure is repeated for every feature. The training process of the GFP.MC algorithm is given in Fig. 8. The classification process is the same as that of the OFP.MC algorithm; see Fig. 2. A counterexample showing that GFP.MC is suboptimal. Sup- pose that eight labeled examples have feature values and class la- bels, respectively, as in h1; C1i; h2; C1i; h3; C0i; h4; C0i; h5; C1i; h5; C1i; h5; C1i; h6; C0i; after the examples are projected on some fixed feature dimension. In Fig. 9, every feature interval is shown by a box, below and above which its class label and the number of examples from different classes in it are stated. The feature intervals with class labels C1 and C0 have, respectively, dark and light colors. The greedy partition with at most two pairwise disjoint intervals makes three misclassi- fications, while there is a partition which makes two misclassifica- tions with two pairwise disjoint intervals. This example illustrates that the greedy algorithm may return a suboptimal feature partition which makes more misclassifications on the training set than the minimum achievable with at most the same number of feature Fig. 9. The greedy feature partitioning learning algorithm GFP.MC can be strictly suboptimal in the sense that it does not always find a partition with the least possible number of misclassifications on the training set. In the example above, every greedy partition with at most two intervals makes three misclassfications, while at least one partition with two intervals makes two misclassifications on the training data. Table 3 Properties of twenty real-world datasets from UCI-Repository. Dataset Size # Of features # Of linear features # Of classes Baseline accuracy (%) arrhythmia 452 279 272 16 54 bcancerw 699 10 10 2 66 cleveland 303 13 6 2 54 dermatology 366 34 33 6 31 echocardio 74 7 6 2 68 glass 214 9 9 6 36 haberman 306 3 3 2 74 heart 270 13 7 2 56 horse 368 22 7 2 63 hungarian 294 13 6 2 64 ionosphere 351 34 34 2 64 iris 150 4 4 3 33 mammog.-mas. 961 5 3 2 54 new-thyroid 215 5 5 3 70 parkinsons 195 22 22 2 75 segmentation 2310 19 19 7 14 sonar 208 60 60 2 53 vehicle 846 18 18 4 26 wine 178 13 13 2 40 yeast 1484 8 8 10 31 A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 4541 intervals. In this sense, the greedy feature partitioning learning algorithm should not be expected to be optimal. 6. Experimental results This section presents empirical evaluations of the OFP.MC and GFP.MC algorithms. The algorithms are compared to previously proposed feature-projection based algorithms, FIL.IF, VFI5, CFP, k- NNFP, and NBC. All of those algorithms are evaluated on twenty datasets from the UCI machine learning repository (Asuncion & Newman, 2007). Table 3 describes each dataset, including the number of in- stances, features, linear features, classes, and the baseline classifi- cation accuracy, which is obtained by predicting the class of every test instance with the most frequent class in the dataset. Table 4 The average percentage accuracy results of OFP.MC, GFP.MC, FIL.IF, VFI5, CFP, k-NNFP, and NBC algorithms. For each data set, the accuracy of the best performing algorithm is shown in boldface. Dataset OFP.MC GFP.MC FIL.IF CFP VFI5 k-NNFP NBC arrhythmia 65.27 59.20 56.383+ 54.223+ 61.601+ 54.223+ 58.063+ bcancerw 96.57 96.57 96.42 95.713+ 95.193+ 95.422+ 97.402� cleveland 82.12 82.98 83.582� 77.703+ 82.12 81.46 79.942+ dermatology 95.19 95.46 63.773+ 50.003+ 96.722� 52.243+ 77.933+ echocardio 71.90 69.70 67.262+ 69.431+ 70.30 68.101+ 70.59 glass 64.37 61.86 54.662+ 48.033+ 60.551+ 60.48 53.923+ haberman 71.18 72.68 72.75 74.121� 59.483+ 73.53 70.98 heart 82.52 82.37 80.74 76.673+ 81.11 79.562+ 79.773+ horse 78.75 77.931+ 77.193+ 66.763+ 78.64 69.583+ 80.273� hungarian 82.92 83.61 84.15 68.713+ 84.961� 74.973+ 82.58 ionosphere 89.23 89.35 92.533� 86.782+ 80.973+ 88.092+ 88.42 iris 93.07 92.93 93.33 84.803+ 95.732� 94.801� 92.94 mammog.-mas. 82.58 82.57 79.943+ 80.423+ 78.583+ 82.96 82.56 new-thyroid 94.51 94.05 95.07 80.473+ 92.283+ 88.373+ 93.84 parkinsons 79.80 84.002� 87.493� 77.03 70.263+ 78.77 77.65 segmentation 80.80 80.80 74.243+ 73.263+ 76.883+ 77.223+ 80.46 sonar 74.35 75.68 67.703+ 59.923+ 61.193+ 72.99 67.882+ vehicle 59.50 58.96 58.802+ 56.952+ 57.092+ 59.48 61.762� wine 93.49 92.48 95.622� 79.023+ 96.423� 96.173� 93.48 yeast 52.33 47.562+ 41.043+ 32.083+ 42.743+ 37.633+ 46.453+ Average 79.52 79.03 76.131+ 69.603+ 76.142+ 74.301+ 76.842+ Attained significance level 0.33126 0.03918 0.00021 0.00251 0.01529 0.00985 #OFP.MC better/worse 2/1 10/4 18/1 12/4 11/2 7/3 Fig. 10. The boxplots of the average accuracies of the learning algorithms on twenty datasets, reported in Table 4. 4542 A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 Table 4 reports the classification accuracies of the OFP.MC, GFP.MC, FIL.IF, VFI5, CFP, k-NNFP, and NBC algorithms. The results are based on the average of the accuracies over five repetitions of 5-fold cross-validation runs. For k-NNFP algorithm, optimal k is set to an odd integer between 1 and 45 which maximizes the clas- sification accuracy found by 5-fold cross-validation over the train- ing set; ties are broken in favor of the largest values. For each dataset, the best classification accuracy is shown in boldface. OFP.MC attains the largest average accuracies on 6 data- sets, FIL.IF and VFI5 on 4 datasets, NBC on 3 datasets, finally, GFP.MC, CFP and k-NNFP on 1 dataset. To assess the statistical significance of the differences between the accuracies of OFP.MC and other algorithms on every dataset, we perform paired t-tests. If the difference between the accuracies is found statistically significant at 0.05, 0.01, or 0.001 levels, then this is indicated, respectively, by 1±, 2±, or 3± in the superscript at- tached to the accuracy of the competing algorithm, where + (resp., �) means OFP.MC is better (resp., worse) than the algorithm. No superscript means that the difference between the accuracies of the OFP.MC and the algorithm are statistically insignificant at 0.05 level. The last row counts for each algorithm the number of datasets on which the accuracy of OFP.MC is significantly better/ worse than the accuracy of that algorithm. The grand means of average accuracies over twenty datasets of all methods are re- ported on the row starting with ‘‘Average’’, and the attained signif- icance levels (p-values) of the paired t-tests applied to their differences are reported in the following row. According to paired t-test results, OFP.MC outperforms/under- performs GFP.MC on 2/1, FIL.IF on 10/4, CFP on 18/1, VFI5 on 12/ 4, k-NNFP on 11/2, and NBC on 7/3 out of 20 datasets at some levels of statistical significance. Over twenty datasets, the average accu- racy of OFP.MC equals 79.52% and is statistically better than that of FIL.IF and k-NNFP at 0.05 significance level, those of VFI5 and NBC at 0.01 significance level, and that of CFP at 0.001 significance level. The difference between the average accuracies of OFP.MC and GFP.MC is statistically insignificant at 0.05 level. In Fig. 10, the boxplots of the classification accuracies of the learn- ing algorithms on twenty datasets show that OFP.MC and GFP have higher median accuracies than those of all other feature-projection based algorithms. The comparisons of lower and upper quartiles show that the distributions of the average classification accuracies of OFP.MC and GFP.MC are tilted significantly more toward the higher values than those of the other feature-projection based methods. Even though GFP.MC has slightly higher median accuracy than that of OFP.MC, the distribution of the OFP.MC accuracy results concen- trates on a range of values slightly higher than that of GFP.MC. We also examine the sensitivities of the algorithms to the errors in the class boundaries. The plots on the left in Fig. 11 are gener- ated from randomly distributed 55 points on [0, 0.5] and [0.5, 1] with class labels 0 and 1, respectively. The original data plotted in the top left corner is perfectly separable: all of the class 0 (resp., 1) examples have feature values less (resp., greater) than 0.5. Sup- pose that five randomly selected points from classes 0 and 1 are moved by mistake to 0.8 and 0.2, respectively, outside their class boundaries, which results in the plot in the bottom left corner. The classification algorithms are run on ten replicas of both noise- less and noisy data generated as above. On the right of Fig. 11 are the boxplots of the classification accuracies of the algorithms run on those ten replicas before and after the noise is introduced. In ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●0 1 C la ss es Without boundary noise ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0.0 0.2 0.4 0.6 0.8 1.0 0 1 C la ss es With boundary noise Feature values ofp.mc gfp.mc fil.if cfp vfi5 knnfp nbc 65 70 75 80 85 90 95 100 A cc ur ac y (% ) Learning algorithms ● ● 65 70 75 80 85 90 95 100 without boundary noise with boundary noise Fig. 11. The sensitivity of the learning algorithms to the noise in class boundaries. A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 4543 the presence of noise, OFP.MC, GFP.MC, and k-NNFP maintain high levels of classification accuracy, but VFI5, FIL.IF, CFP, and NBC algo- rithms do not. Moreover, the performances of the VFI5 and NBC algorithms are drastically reduced by the noise. Indeed, both box- plots of OFP.MC, GFP.MC, and k-NNFP remain relative close to the top line, while the boxplots of other methods are far away either from the top line or from each other. 7. Complexity analysis We analyze the algorithms in terms of space and time complex- ities. Time complexity analysis is presented for the training process and for the classification of a single test example. We shall denote by M, N, and K the number of training instances, number of fea- tures, and number of classes. 7.1. Space complexity analysis In the training phase of the OFP.MC algorithm, feature intervals for concept descriptions are constructed on each feature dimen- sion. The space required for storing intervals will be proportional to M N K at worst case. The space required for storing value and policy functions is proportional to M2K at worst case. Therefore, the space complexity of the OFP.MC algorithm is O(M(M + N)K). The space complexities of the FIL.IF and CFP algorithms are O(M N), whereas the space complexities of the GFP.MC and VFI5 algo- rithms are O(M N K) and O(N K2), respectively. The k-NNFP algorithm stores all instances as feature projec- tions. Therefore, the space required by the k-NNFP algorithm is O(M N). Because the NBC algorithm computes and stores a distribu- tion for each value on each feature dimension for each class, the space complexity of the NBC algorithm is O(N M K). 7.2. Time complexity analysis of the training processes For a dataset with M training instances, feature projections on a feature dimension are sorted with time complexity O(M log M). In the OFP.MC algorithm, for each feature, computing value and policy functions using sorted feature projections has time complex- ity O(M2K2) and finding optimal class labels has time complexity O(M K). From optimal class labels, feature intervals are constructed in O(M K) time. The total complexity of the constrction of intervals on a feature dimenion is O(M log M + M2K2 + M K) = O(M2K2). There- fore, the overall time complexity of the OFP.MC algorithm for train- ing is O(N M2K2). However, in general, the number of classes in classification problems is constant. In that case, the time complexity of the OFP.MC algorithm for training is O(N M2). In the GFP.MC algorithm, for each feature, constructing intervals using sorted feature projections has time complexity O(M2K). The overall complexity of the construction of intervals on a feature dimension is O(MlogM + M2K) = O(M2K). Therefore, the overall time complextiy of the GFP.MC algorithm for training is O(N M2K). If the number of classes is constant, then the time complexity of the GFP.MC for training is O(N M2). The training time complexities of the FIL.IF and CFP algorithms are O(N MlogM) (Dayanik, 2010; Güvenir & S�irin, 1996). The training time complexity of the k-NNFP is O(N MlogM) (Akkus� & Güvenir, 1996). The training time complexity of the NBC and the VFI5 algo- rithms is O(N M K). If the number of classes is constant, then the training time complexity of the NBC and VFI5 algorithm is O(N M). 7.3. Time complexity analysis of a single classification process During the classification process, the function search-inter- val(f, value) finds the interval containing feature value of the test instance on the feature dimension f by a binary search and deter- mines the class votes of that feature. The number of intervals on a feature dimension is at most equal to the number of training in- stances M. Hence, the worst case time complexity of the search process on a feature dimendion is O(log M + K) for a feature. Since the final classification is based on the votes of all features, single instance classification time complexity of the OFP.MC and GFP.MC algorithms is O(N log M + N K) = O(NlogM) because the number of classes, K, is usually constant. The classification time complexities of the FIL.IF, CFP, and k- NNFP algorithm is O(N log M) if M � K (Dayanik, 2010; Güvenir & S�irin, 1996; Akkus� & Güvenir, 1996), whereas the classification time complexity of the VFI5 algorithm is O(N K). The classification time complexities of the NBC algorithm is O(N log M K). 8. Conclusion Feature-projection based classification algorithms are simple, yet effective and efficient, produce easy-to-understand concept 4544 A. Dayanik / Expert Systems with Applications 39 (2012) 4532–4544 descriptions, enable fast classification, and handle missing feature values naturally by simply neglecting them. This paper introduced two new feature-projection based classification algorithms which represent induced classification knowledge as collections of fea- ture intervals. The first algorithm, OFP.MC, optimally partitions each feature using dynamic programming to minimize the feature classification error, while the second algorithm, GFP.MC, greedily partitions each feature to reduce the space requirement of the OFP.MC algorithm. We showed that the OFP.MC algorithm performs better than the previously proposed feature-projection based classification algo- rithms on most of the real-world datasets. In the meantime, the new algorithms are insensitive to boundary noise unlike the other feature-projection based classification algorithms considered here. Although the GFP.MC algorithm is suboptimal in the sense that it does not always find a partition with the smallest possible num- ber of misclassifications on the training set, experimental results indicate that it is competitive with OFP.MC on twenty datasets from the UCI-Repository. One important future research direction is to learn concept- dependent feature votes for the feature-projection based classifica- tion algorithms. Another direction is to investigate better ways to combine predictions of features. References Akkus�, A., & Güvenir, H. A. (1996). K nearest neighbor classification on feature projections. In L. Saitta (Ed.), Proceedings of the 13th international conference on machine learning (pp. 12–19). Bari, Italy: Morgan Kaufmann. Asuncion, A., & Newman, D. (2007). UCI machine learning repository. Chandra, B., & Gupta, M. (2011). Robust approach for estimating probabilities in Naive Bayes classifier for gene expression data. Expert Systems with Applications: An International Journal Archive, 38(3). Chena, J., Huanga, H., Tiana, S., & Qua, Y. (2009). Feature selection for text classification with Naive Bayes. Expert Systems with Applications, 36(3), 5432–5435. Dayanik, A. (2010). Feature interval learning algorithms for classification. Knowledge-Based Systems, 23(5), 402–417. Demiröz, G., & Güvenir, H. (1997). Classification by voting feature intervals. In M. van Someren & G. Widmer (Eds.), Proceedings of 9th european conference on machine learning, Prague, Czech Republic (pp. 85–92). Springer. Domingos, P., & Pazzani, M. (1997). On the optimality of the simple Bayesian classifier under zero-one loss. Machine Learning, 29(2-3), 103–130. Dougherty, J., Kohavi, R., & Sahami, M. (1995). Supervised and unsupervised discretization of continuous features. In Proceedings of the International Conference on Machine Learning (pp. 194–202). Duda, R. O., Hart, P. E., & Stork, D. G. (2000). Pattern classification (second edition). John Wiley & Sons Inc.. Fukunaga, K. (1990). Introduction to statistical pattern recognition. Academic Press. Güvenir, H., & S�irin, I. (1996). Classification by feature partitioning. Machine Learning, 23, 47–67. Güvenir, H., Demiröz, G., & Ilter, N. (1998). Learning differential diagnosis of erythemato-squamous diseases using voting feature intervals. Artificial Intelligence in Medicine, 13, 147–165. Güvenir, H. A., & Emeksiz, N. (2000). An expert system for the differential diagnosis of erythemato-squamous diseases. Expert Systems With Applications, 18(1), 43–49. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., & Witten, I. H. (2009). The weka data mining software: An update. SIGKDD Explorations, 11(1). Hsu, C.-C., Huang, Y.-P., & Chang, K.-W. (2008). Extended Naive Bayes classifier for mixed data. Expert Systems with Applications, 35(3), 1080–1083. Kim, S.-B., Han, K.-S., Rim, H.-C., & Myaeng, S. H. (2006). Some effective techniques for Naive Bayes text classification. IEEE Transactions on Knowledge and Data Engineering, 18(11), 1457–1466. Lewis, D. D. (1998). Naive (Bayes) at forty: The independence assumption in information retrieval. In Proceedings of the 10th European conference on machine learning (pp. 4–15). London, UK: Springer. Liu, H., Hussain, F., Tan, C. L., & Dash, M. (2002). Discretization: An enabling technique. Data Mining and Knowledge Discovery, 6(4), 393–423. Mitchell, T. (1997). machine learning. New York: McGraw Hill. Witten, I. H., & Frank, E. (2005). Data mining: Practical machine learning tools and techniques (second edition). Morgan Kaufman. Yang, Y., & Webb, G. I. (2009). Discretization for Naive–Bayes learning: Managing discretization bias and variance. Machine Learning, 74(1), 39–74. Learning feature-projection based classifiers 1 Introduction 2 Previous work on feature-projection based algorithms 3 OFP.MC: Optimal feature partitioning learning algorithm for classification 4 Optimal partitioning of linear features in the training of OFP.MC 4.1 Optimal feature partitioning by using dynamic programming 4.2 The construction of optimal feature intervals corresponding to maximum k label switchings 4.3 The performance evaluation of optimal feature intervals corresponding to maximum k label switchings 4.4 The calculation of the performance curve and optimal choice of the maximum number of label switchings 4.5 The construction of the final feature intervals for classification 4.6 Illustration on the iris dataset 5 GFP.MC: greedy feature partitioning learning algorithm for classification 6 Experimental results 7 Complexity analysis 7.1 Space complexity analysis 7.2 Time complexity analysis of the training processes 7.3 Time complexity analysis of a single classification process 8 Conclusion References 
work_xxxmjye7hndnvduxtaf33a57ma	----	EIC Announcement: 2008 Changes to the Editorial Board of Expert Systems 2008 is the 25th anniversary of Expert Systems and is an exciting time for the Journal. Recently introduced changes in the Journal’s scope and review processes have led to a greater volume of quality submissions and a much wider coverage of the knowledge engineering discipline in our publications. As a result our editorial team is growing. Introducing our new co-Editor in Chief I am delighted to announce that Dr Jon G. Hall, from the Open University, UK, will be joining me as co-Editor in Chief. Jon has been on the editorial board of Expert Systems since 2005, and an associate editor since January 2007. His biography is included at the end of this announcement. Introducing our new Associate Editors Four new associate editors have been appointed to join Dr Elif Ubeyli, who will also continue in this role. They have been selected based on their research expertise, their past experience with Expert Systems and their dedication as reviewers and guest editors since I have been EiC. You can read a little of what they say about themselves in the biographies included at the end of this announcement. The Expert Systems’ associate editor team is now: Dr Juan I. Arribas, Universidad de Valladolid, Spain Dr Paolo Remagnino, Kingston University, UK Dr Andrés Silva, Universidad Politécnica de Madrid, Spain Dr Elif D. Ubeyli, TOBB Economics and Technology University, Turkey Professor Li Xu, Old Dominion University, US Welcome to new Board Members It is also a great pleasure to welcome the following new Board Members, whose expertise nicely complements that of the existing members: Professor David Benyon, Napier University, UK Professor Max Bramer, Portsmouth University, UK Dr Luis A. Casillas Santillan, University of Guadalajara, Mexico Dr Zhi Jin, Chinese Academy of Science, China Dr Mehmet Orgun, Macquarie University, Australia Professor Dan O' Leary, University of Southern California, US Professor I. B. Türksen, University of Toronto, Canada Dr Ning Xiong, Mälardalen University, Sweden Dr Ali R. Yildiz, University of Michigan, US I look forward to a lasting and fruitful association with Expert Systems. Farewell to David Blanchard I would like to take this opportunity to thank David Blanchard who is retiring from the Editorial Board after many years of valuable service and advice. David was news editor for Expert Systems for many years. He has contributed greatly to the success of the journal and we will miss him. David: from all of Expert Systems thank you, and all the best for the future. Biographies Jon Hall MBA PhD FBCS CITP is a Senior Lecturer in the Computing Department at the Open University. Previous editorial roles include guest editor of special issues of Elsevier’s Journal of Information and Software Technology and IET Software (formerly IEE Proceedings-Software). Jon holds an IBM Eclipse Innovation Award for Innovation. He is currently focused on providing theoretical and practical foundations for computing as an engineering discipline, and reflecting good practice from software engineering into other engineering disciplines. Jon has published over 60 academic papers. For more information, please see HYPERLINK "http://mcs.open.ac.uk/jgh23/" http://mcs.open.ac.uk/jgh23/. Juan Ignacio Arribas received his MS and PhD degrees from the University of Valladolid in 1996 and 2001, respectively. In 1997 he was awarded an F.P.I. research fellowship from the Department of Science and was, during 1998, a Visiting Research Associate at the University of Maryland, Baltimore, MD. In 1999 he joined the Department of Teoria de la Senal y Comunicaciones e Ingenieria Telematica where he was appointed as an Associate Professor in 2004. His research interests include expert systems, neural networks and their applications to medical image. Dr Paolo Remagnino PhD, MIEEE is a Reader in the Faculty of Computing, Information Systems and Mathematics at Kingston University, UK. Paolo’s PhD is in Computer Vision, received form University of Surrey in 1993. He is the recipient of a Royal Academy of Engineering Global Award and an EPSRC grant which funded a three year collaboration with Saitama University and Toshiba, Japan, and is involved in many other funded projects as co-investigator, including work on the automatic interpretation of behaviour of people, object classification and dynamics. His past editorial experience spans three journals: Machine Vision and Applications, the International Journal of Robotics and Automation, and Pattern Analysis and Application. His research interests include image processing, computer vision, machine learning, artificial intelligence and robotics. His research is multi-disciplinary in nature and mainly focuses on automatic image and video sequence understanding. Paolo has published over 80 scientific papers and is the editor of four Springer book collections in the field of video surveillance and ambient intelligence. Andrés Silva is an Associate Professor in the Computer Science School of the Universidad Politécnica de Madrid (UPM). He has two years industrial experience in software engineering and worked at the the Institute for Systems, Informatics and Safety part of the Joint Research Centre of the European Commission. He obtained his PhD in 2001. Andrés has served on the program committees of several conferences, edited a book on Requirements Engineering for Socio-technical Systems and co-edited a special issue of IET Software (formerly IEE Proceedings-Software). Andrés is author of several research papers published at major conferences (including ICSE) and journals (including IEEE Software, IEEE Transactions on Software Engineering and Safety Science). His research career is centred around Requirements Engineering with a special focus on the topics of Knowledge Management and Safety. Elif Derya Übeyli is an Associate Professor in the Department of Electrical and Electronics Engineering, TOBB University of Economics and Technology. She obtained her PhD in Electronics and Computer Technology from Gazi University in 2004. Since then she has worked on variety of topics including biomedical signal processing, neural networks, optimisation and artificial intelligence and on several projects related to biomedical signal acquisition, processing and classification. Elif has served in many editorial roles, as member of the editorial board of and guest editor for several scientific journals, including the Journal of Engineering and Applied Sciences, the International Journal of Soft Computing, the Research Journal of Applied Sciences, the Research Journal of Medical Sciences. She has published over 100 academic papers. For more information, please see ( HYPERLINK "http://edubeyli.etu.edu.tr/" http://edubeyli.etu.edu.tr/). Li Xu is a Professor of Information Technology at Old Dominion University, USA. He is the Editor-in-Chief of Enterprise Information Systems (published by Taylor & Francis) and a member of the editorial board of Systems Research and Behavioural Science (published by Wiley). He was the chair of Confenis 2006 & 2007, sponsored by the IFIP. He has served as special issue editor for IEEE Transactions SMC Part C, IEEE Transactions on Industrial Informatics, Annals of Operations Research, European Journal of Operational Research, Computers & Operations Research, Information Systems, Decision Support Systems, International Journal of Production Research, Knowledge-Based Systems, Expert Systems with Applications, among others. He has published over 100 scientific papers. Lucia Rapanotti 3.1.2008 
work_xy6hzbddgretvirrb2rknk2beu	----	Expert system for predicting unstable angina based on Bayesian networks Expert Systems with Applications 40 (2013) 5004–5010 Contents lists available at SciVerse ScienceDi rect Expert Systems with Applic ations j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Expert system for predicting unstable angina based on Bayesian networks q 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.03.029 q This work was supported by the Spanish Ministry of Education and Science under Grant Instituto Carlos III (FEDER), Red HERACLES 06/0009. ⇑ Corresponding author. Tel.: +34 963543398. E-mail address: joan.vila@uv.es (J. Vila-Francés). Joan Vila-Francés a,⇑, Juan Sanchís b, Emilio Soria-Olivas a, Antonio José Serrano a, Marcelino Martínez-Sober a, Clara Bonanad b, Silvia Ventura b a Intelligent Data Analysis Laboratory, University of Valencia, Avd Universitat s/n, 46100 Burjassot, Valencia, Spain b Servicio de Cardiología, Hospital Clinic Universitari de Valencia, Spain a r t i c l e i n f o Keywords: Bayesian networks Expert systems Medical applications a b s t r a c t The use of computer-based clinical decision support (CDS) tools is growing significantly in recent years. These tools help reduce waiting lists, minimise patient risks and, at the same time, optimise the cost health resources. In this paper, we present a CDS application that predicts the probability of having unst a- ble angina based on clinical data. Due to the characteristics of the variables (mostly binary) a Bayesian netwo rk model was chosen to support the system. Bayesian-network model was constructed using a populatio n of 1164 patients, and subsequently was validated with a population of 103 patien ts. The val- idation results, with a negative predictive value (NPV) of 91%, demonstrate its applicability to help clini- cians. The final model was implemented as a web application that is currently been validated by clinician specialis ts. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction The use of computer-b ased clinical decision support (CDS) tools is growing in recent years due to different reasons (Steyerberg, 2009): � Helps the clinicians making decisions, thus reducing the clinical errors. � Improves the time to get a diagnost ic reducing waiting time. � Optimizes health resources reducing unnecessary medical tests. Those advantages have expanded the use of these tools in clinical practice in the form of web services, desktop programs or applications for mobile phones and tablets. The use of these tools in different clinical areas, supported by the increasing amount of patient data and the ability to analyse and process it through Big Data techniques, is foreseen as a huge boost in health care (Manyika et al., 2011 ). Currently, one of the areas that demands more resources is cardiology. Nowaday s cardiovascu lar diseases are the main cause of death in the developed world (Escaned et al., 2008 ). Modern lifestyle that leads us to many stressful situ- ations, poor diet and little exercise have made heart disease the cause of many deaths. One of the symptoms of possible heart failure angina is characterise d by severe chest pain. When suffering from that pain for more than 15 min, it is highly recommend ed requiring clinical attention, as it can be the initial phase of a myo- cardial infarction . This pain occurs when the demand for oxygen by the heart muscle is not served (because there is an interruption of the blood supply to a part of the heart muscle). This pain is in fact one of the most frequent causes of admission in the Emergency Services of the hospitals. Some of these patients suffer an acute coronary syndrom e that is diagnosed by electrocardiogr am (ECG) findings or alteration of biomarker s of myocardial damage (tropo- nin). However, in some patients the ECG is nonspeci fic and tropo- nin is normal. This population suffers form a chest pain of uncertain origin. They are mostly low-risk patients without any heart disease, but we can not reject an acute coronary syndrome in some of them. The key challenge is to identify those patients at risk for suffering an acute coronary syndrom e with normal tro- ponin (unstable angina), within a population that is generally at low risk. The tools currently available in the emergency room of a hospital do not work, because the ECG is nonspeci fic, and tropo- nin is normal. As a result, the final decision of admission or release is postponed until a treadmil l stress test is carried out (usually the next morning ) Sanchis et al. (2006). This strategy is suboptimal because the patient must wait for several hours, many patients can not run on the treadmill, and sometimes the results are inconclus ive. This article proposes the developmen t of a Clinical Decision Support System, for being use in the emergency units of a hospital in order to determine the probability of unstable angina within http://dx.doi.org/10.1016/j.eswa.2013.03.029 mailto:joan.vila@uv.es http://dx.doi.org/10.1016/j.eswa.2013.03.029 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa J. Vila-Francés et al. / Expert Systems with Applications 40 (2013) 5004–5010 5005 24 h of patient entry into the hospital. The inputs of the system are the clinical data that are routinely collected in the emergency room of a hospital. Almost all of those data have a binary response (e.g. – patient gender, smoker, etc.). With this kind of inputs, the most appropriate machine learning models are decision trees and Bayesian networks (Alpaydin, 2009 ). The results obtained from decision trees were not good enough (deep trees were needed, and therefore generalisation was bad). Moreover, a system whose parameters could be updated continuo usly when new information was available was required, that is why Bayesian networks were used. The advantages of these models for using them in clinical problems are Lucas et al. (2004): � The Bayesian model can be interpreted by the clinician, as the relationship s among variables are clearly represented by a graph (directed graph in our model). � The clinician can provide expertise knowled ge to establish new relationship s between variables that might not have been reflected in the case of using an automatic learning algorithm for the Bayesian network structure. � Adding new knowledge is a straight forward process, which can be automate d by updating the frequency tables of each input variable. � It is not necessary to know the values of all inputs to the model to obtain a valid output. Thus, if the inputs are obtained in a sequential way, as it happens in an emergency unit, or some information about the patient is missing, the system may be updating the probability as soon as new information of clinical tests is available. Once the Bayesian network was validated, it was implemented into a Decision Support System which allowed clinicians to access it remotely, in an easy and simple way. The obvious solution was to implement the expert system as a web application. This way it is possible to centralise the data at a single location while access can be granted from any computer without requiring dedicated software (only a web browser and an Internet connection). Moreover, it is possible to implement a user-based managemen t to control the people that use the tool and access the data. The rest of this paper is organised as follows. Section 2 explains the Bayes- ian models used. Section 3 discusses the data used and results obtained. Section 4 explains the Web tool development. Finally, Section 5 summarizes the conclusio ns of the present work. Fig. 1. Structure of a Naïve network. 2. Bayesian networks A Bayesian network (BN) is a probabilistic graphica l model composed of two different parts: on one hand is the graphical structure (directed acyclic graph) that defines the relationship between variables and, on the other hand, the probabilities estab- lished between these variables (Koller and Friedman, 2009; Korb and Nicholson, 2011 ). The elements of a Bayesian network are as follows Rusell and Norvig (2009): � A set of variables (continuous or discrete) forming the network nodes. � A set of directed links that connect a pair of nodes. If there is a relationship with direction X ? Y is said that X is the parent of Y. The network fulfils the following facts: � Each node Xi is associated with a conditional probability func- tion P(XijParents(Xi)) that takes as input a particular set of values for the node’s parent variables, and gives the probability of the variable represented by the node Xi. � The graph has no directed cycles. The knowled ge is reflected by the relationship s established in the graph nodes, and the conditional probabilities of the variables represented in each node. Those probabilitie s are estimated using the dataset. In this paper, the most widely used Bayesian classifiers have been applied: Naïve Bayes, FAN (Forest Augmented Network) and TAN (Tree Augmented Network). The following subsectio ns explain each of these approaches. 2.1. Naïve Bayes The hypothesis of this approach is to assume that the predictive variables are conditionally independen t given the variable to classify. Although this assumption is quite restrictive, this classifier is one of the most used, and several studies show that the results are as good as the ones obtained with other techniques (neural networks , decision trees etc.) Korb and Nicholson (2011). The big problem with this model arises when the conditional indepen- dence among predictive variables is not fulfilled. This happens when predictiv e variables are redundant or are highly correlated with each other. Assumin g the conditional independen ce, the graph of a Naïve Bayes is shown in Fig. 1. In Naïve Bayes approach, the conditional probability P(classjX1, X2, . . . , XN) is factorized as PðclassjX1; X2; . . . ; XNÞ¼ QN i¼1PðclassjXiÞ. This factorisa tion is more easily obtained from experimental data, in addition to being easier to analyze and, later on, to obtain infer- ences from available information . 2.2. Tree Augmented Network, TAN This structure is an extension of a Bayesian classifier in which each variable is allowed to have another parent outside the class node. The idea is to build a Bayesian network tree for all predictive variables and complete the model with a Naïve Bayes. TAN algo- rithm forms a tree with the predictive variables and then add edges to the class node. Fig. 2 is a conceptual illustration of how the mod- el works. It is seen that each predictive variable, Xi, can have up to two parents. 2.3. Forest Augmented Network, FAN An important limitation that TAN model may have, a priori , is that some arcs of the tree, formed between the descriptive variables, can introduce noise into the classification if such relationship s do not exist. This model proposes the formation of disjoint trees with predictive variables. So this approach creates a tree structure with all the variables. Obviously many of the depen- dencies are enforced by the method of construction and do not exist. These depende ncies are discarded during the process of creating the tree; an edge is discarded if it is independent (for example, using a statistical v2test.). Fig. 3 shows a model built from two disjoint tree predictor s. Fig. 2. Structure of a TAN network. Fig. 3. Structure of a FAN network. Table 1 Distribution of discrete variable s. Variable Positive (%) Negative (%) EffortPain 41.75 58.25 x2pain24 62.80 37.20 var_n 33.51 66.49 Smoker 76.29 23.71 HTA 40.64 59.36 antcolest 46.74 53.26 DM 73.28 26.72 antfam 89. 52 10.48 antiam 76.98 23.02 estenosisc 77.92 22.08 anticc 98.11 1.89 antactp 88.92 11.08 antcir 93.30 6.70 artperif 95.27 4.73 ictus 93.38 6.62 event30 (observed variable) 80.24 19.76 Table 2 Distribution of continuous variables. Variable Mean Std Min Max Age (years) 63.6735 11.4010 30 92 Creatinine (mg/dl) 1.0623 0.5652 0.0 11.1 5006 J. Vila-Francés et al. / Expert Systems with Applications 40 (2013) 5004–5010 3. Data used, methodology and results The implemented Clinical Decision Support System is based on a Bayesian Network model trained with a dataset of 1164 cases. 3.1. Dataset The dataset was collected from patients incoming to the emergency unit of the Hospital Clínic from Valencia, Spain. Each case of the dataset is formed by 17 input variables correspondi ng to different characterist ics and symptoms of the patients, and an observed (output) variable, which corresponds to the fact of suffer- ing an unstable angina. The input variables were: age and gender of the patient; Patient Symptoms: effort related pain (EffortPain), two or more pain episodes in the last 24 h (x2pain24); and the Risk Factors of the patient: creatinine level, smoker, arterial hyperten- sion (HTA), history of hypercholesterol emia (antcolest), diabetes mellitus (DM), family history of ischaemic heart disease (antfam), history of myocardial infarction (antiam), previous Coronary Stenosis (estenosisc), previous admission due to Heart failure (anticc), prior coronary angiopla sty (antactp), prior coronary surgery (antcir), peripheral arteriopathy (antperif), and cerebral ic- tus. Most of the input variables are binary, and hence they are con- verted to discrete nodes of the Bayesian Network directly. Only two variables are continuous (age and creatinin e) and therefore have been converted to discrete nodes before being used by the model. The age has been discretised into intervals of ten years (less than 10 years, from 10 to 19 years, and so on) and the creatinine level has been discretised into two values (less than 1.7 mg/dl and greater than this value), following a clinical criterion. The observed variable of the model is the event of a heart attack within 30 days (event30). The distribution of the values on the training dataset is summarised in Tables 1 and 2. All the cases were used for training the BN model. 3.2. Methodology and results The above-mention ed Bayesian models (Naïve Bayes, TAN structure and FAN structure ) were develope d with the BNT Toolbox for MATLAB (Murphy, 2007 ). This toolbox can learn auto- matically the network structure from the data and construct a di- rected acyclic graph (DAG), which is represented as an adjacency matrix. Afterwar ds, the automaticall y created structure s were fine-tuned by the expert, who added some arcs to the graph: a rela- tionship from HTA to anticc and ictus, and another one from DM to creatinine. The resulting structure, named Medical FAN, is shown in Fig. 4. Finally, the four models were compared in order to find the best fit to the data. The Naïve and TAN models were proved to gen- erate the same results. So that, only the Naïve, FAN and Medical FAN analysis are shown here. First of all, we compare in Table 3 the correlation between the outputs from the three models by measuring the correlation coefficient R2 (it is worth noting that Ev30 x2angina24 edad var_n fumador hta antcolestdm antfam antiam estenosisc anticc antactp antcir artperif ictus creatinina DolorTipico Fig. 4. Bayesian Network to predict unstable angina probability. Table 3 Correlation betwe en the values predicted by the implemented Bayesian Network Structures. R2 FAN Medical FAN Naïve Bayes 0.8571 0.8400 FAN 0.9837 Table 5 Area Under the Curve (IC 95%). Bold indicates the network with the best performance. AUC Naïve Bayes 0.7298 ± 0.034 FAN 0.7473 ± 0.034 Medical FAN 0.7498 ± 0.034 Table 6 Performance of the BN model over the training data. Sensitivity 0.6957 Specificity 0.6799 Positive predictive value 0.3486 Negative predictive value 0.9007 Accuracy 0.6830 Table 7 Performance of the BN model over the validation data. Sensitivity 0.8710 Specificity 0.6111 Positive predictive value 0.4909 Negative predictive value 0.9167 Accuracy 0.6893 J. Vila-Francés et al. / Expert Systems with Applications 40 (2013) 5004–5010 5007 the output of the models is a probability value). Another measure of the model performanc e, shown in Table 4, is the Cohen kappa coefficient; it is a statistical measure of inter-rater agreement or inter-annotato r agreement for qualitative (categorical) items (Carletta, 1996 ). The last measure of the model performance taken into account is the Area Under the Curve (AUC), which reflects the percentage of correct classification (an AUC of 1 corresponds to a perfect classifi- cation for all the cases, while a AUC of 0.5 indicates a random classification, with a 50% chance of being correct for each class) Fawcett (2004). Table 5 shows the AUC of the three models. The table reflects a similar behaviour of the three models regarding this parameter. After considering the similar performanc e of the three models, we chose the medical FAN model for deploying the final Decision Support System. Then, the directed acyclic graph (DAG) of the medical FAN model was converted to the Netica DNET file format for expressing BN structures. Finally, the Condition al Probability Distribution (CPD) of each node was learned from the dataset using a C program and included in the DNET file for the model. The performanc e of the final model over the training is summarised in Table 6. The impleme nted CDSS uses this BN model to calculate an output for each given set of evidences . The output of the model is a continuous probability value, which is discretised into a binary value (high risk/low risk) by using a threshold that optimises the Negative Predictive Value. The model has been validated with a new dataset formed by 103 new cases, collected in the same emergency unit on December Table 4 Bayesian Network decision agreement. Table values correspond to the classification threshold of each model, using the same classification criter ia. Kappa FAN (cut 0.20631) Medical FAN (cut 0.2084) Naïve Bayes 0.8056 ± 0.0480 0.8139 ± 0.0482 (cut 0.1971) FAN 0.9800 ± 0.0514 of 2011. The performance of the model for this new dataset is summari sed in Table 7. 4. Web applicati on The authors have deployed a Clinical Decision Support System based on Bayesian networks as a web application. This tool evalu- ates the risk of heart attack on incoming patients with chest pain in an emergency unit, based on the evidences from the Patient Symp- toms and his clinical history. The system consists of a web front- end with an input form for introducing and evaluating new clinical cases, a probabilistic inference motor based on a Bayesian Network (BN), and a web administration panel for case reviewing and validation. The tool is used as follows: The user (a clinician) fills in the form with the evidences from the incoming patient (symp- toms and clinical history); then the application processes the data 5008 J. Vila-Francés et al. / Expert Systems with Applications 40 (2013) 5004–5010 using the developed BN model and infers a probability of suffering a heart attack, which is classified and then presente d to the user as ‘‘low’’, ‘‘modera te’’ or ‘‘high’’. Additionally, registered users on the application can save the input case data in order to validate it after- wards (i.e., indicate if the predictio n was correct or wrong). Newly validated cases can be eventually used to update the BN beliefs (by updating the CPD of the input nodes). The tool has been implemented using standard HTML technolo- gies (versions XHTML 1.0 and CSS 3). Dynamic behaviou r has been programmed in PHP language on the server side and Javascript on the client side (using the JQuery library). Data is stored on a MySQL database on the server. The inference motor uses a Bayesian Network model: the graph structure of the network has been developed using the BNT library for MATLAB, and the parameter learning and online probabili stic inference is done through a command line program written in C language using the Netica API for BN (Norsys Software Corp., Vancouver, BC, Canada). This subsection details the implementation of the web application, divided into three blocks: application front-end or user interface, Fig. 5. Application front-e applicati on back-end or server-sid e logic, and probabilistic infer- ence motor. 4.1. Applicatio n front-end The web application that implements the CDSS presents two views: the input case view and the administrat ion panel view. Through the input case view, any user can evaluate an incoming patient by filling in an online form and sending the data to the server. The server calculates the probability of heart attack with the inference motor, and returns the classified risk to the webpage . The administrat ion panel view is used by registered users to review and validate past cases. The tool tracks all the introduced cases by assigning a random alphanumer ic reference to each case. The layout of the input case view consists of a collapsible login form on the top and the main input form for evaluating the cases (Fig. 5). The main form requests some evidences from the patient–corresponding to the nodes of the BN graph described in Section 3.1 – grouped into three different sections: Patient data , nd: user input form. J. Vila-Francés et al. / Expert Systems with Applications 40 (2013) 5004–5010 5009 Patient Symptoms , and Risk Factors . Only a few fields, which are the most relevant for the BN model, are compulsory (age, gender, effort-relate d pain, creatinine level and smoker). The rest of the fields can be left empty because the probabili stic inference motor can deal with sparse data (partially empty input data patterns). Once the form is filled in, it is submitted to the server through an AJAX call, and the predictio n estimated by the inference motor is shown without reloading into a section of the webpage which lays below the main form. The administration panel can be accessed only by registered users (Fig. 6). The system allows three different user roles: admin- istrator, coordina tor and clinician. Depending on the role of the user, the panel allows different administrat ion actions. Clinicians are assigned to a single centre (i.e., an Emergency Unit of a Hospi- tal), and they can only review their own patient cases. There is one coordinator per centre, who can validate cases and create new users for his centre. Finally, the administrator can create users and centres, and validate cases in all the centres. Fig. 6. Application front-end 4.2. Applicatio n back-end The main application logic is written in PHP language . This logic processes the web forms, stores and retrieves the cases into/from the MySQL database, manages the application users and interfaces with the inference motor to calculate the probability of each case. All the interactio ns between the application front-end (web page) and the application logic are AJAX-based, which means that no page reloading is needed for the whole process. The probabili ty inference is carried out by a command-li ne program on the server, which is executed from a PHP script passing the evidences as arguments. 4.3. Probabilis tic Inference motor The tool uses a Bayesian Network to evaluate the risk of suffer- ing a heart attack. The BN is defined by its graph structure and the paramete rs of the model, defined as the Conditional Probability : administration panel. 5010 J. Vila-Francés et al. / Expert Systems with Applications 40 (2013) 5004–5010 Distribution (CPD) of each node. In this model, all the variables are discrete, and therefore the CPD are represented by a table of Conditional Probabilities (CPT). Both the structure of the BN and the CPT of each node is stored on a single model file using the Neti- ca DNET format. The tool runs a probabilistic inference through this model from a command-line program executed on the server, which is written in C language using the Netica API library for BN. The application back-end , upon a evaluation request from the user, calls the inference motor as a PHP command line script, passing the evidences as arguments, and receives the estimated probability as an output from the command. This probability is classified into a risk level using the optimal threshold. Eventual ly, the application back-end returns a pre-formatted web snippet to the front-end showing the evaluation result. 5. Conclusions This paper presents a Clinical Decision Support System (CDSS) that helps the clinicians in the evaluation of incoming emergency patients with unspecific chest pain that are on risk of a heart attack. A correct diagnosis of these patients, which is difficult to achieve with the standard evaluation procedure (ECG and troponin measureme nts), could reduce the number of unnecessary hospital- isations. The CDSS is based on a Bayesian Network (BN) model that takes into account 17 different characterist ics of the patient. The model uses a FAN network structure with discrete nodes. The BN structure was created in MATLAB while the CPD tables of the nodes were learned using the Netica API for BN. A dataset of 1164 cases was used to train the model, which was validated with a new dataset of 103 cases. The model has been optimised to maximise the negative predictive value, reducing therefore the cases of unde- tected risk situation s. In training, the model achieves a 90.07% of NPV, which increases to a 91.67% for the validation dataset. The CDSS runs as a web applicati on with role-based access conatrol, with three different roles with decreasing privilege s (administrator, coordinator s and clinicians). The application consists on an input form and an administ ration panel. The form is used to introduce the characterist ics of the patient under evalu- ation and returns a classified risk of heart attack from low to high. The administrat ion panel is used to review and validate the regis- tered cases in a per-hospita l basis. References Steyerberg, E. (2009). Clinical prediction models: A practical approach to development, validation and updating . Springer. Manyika, J., Chui, M., Brown, B., Bughin, J., Dobbs, R., Roxburgh, C., & Byers, A. (2011). Big data: The next frontier for innovation, competition, and productivity, Tech. rep., McKinsey Global Institute. Escaned, J., Roig, E., Chorro, F., Teresa, E. D., Jiménez, M., de Sá, E. L., et al. (2008). Ámbito de actuación de la cardiologı́a en los nuevos escenarios clı́nicos. documento de consenso de la sociedad espa nola de cardiologı́a. Revista Espa nola de Cardiologı́a, 61, 170–184. Sanchis, J., Bodı́, V., Bertomeu, V., Gómez, C., Consuegra, L., Bosch, M., et al. (2006). Usefulness of early exercise testing and clinical risk score for prognostic evaluation in chest pain units without preexisting evidence of myocardial ischemia. American Journal of Cardiology, 97, 633–635. Alpaydin, E. (2009). Introduction to machine learning . MIT Press. Lucas, P., Gaag, L., & Abu-Hanna, A. (2004). Bayesian networks in biomedicine and health-care. Artificial Intelligence in Medicine, 30, 201–214. Korb, K. B., & Nicholson, A. E. (2011). Bayesian artificial intelligence . CRC Press. Koller, D., & Friedman, N. (2009). Probabilistic graphical models: Principles and techniques. MIT Press. Rusell, S., & Norvig, P. (2009). Artificial intelligence: A modern approach . Prentice Hall. Murphy, K. (2007)The Bayes net toolbox for matlab. Computing Science and Statistics, 33. URL: <http://www.cs.ubc.ca/murphyk/Software/BNT/bnt.html>. Carletta, J. (1996). Assessing agreement on classification tasks: The kappa statistic. Computational Linguistics, 22(2), 249–254. Fawcett, T. (2004). Roc graphs: Notes and practical considerations for researchers, Tech. rep., HP Laboratories. http://www.cs.ubc.ca/murphyk/Software/BNT/bnt.html Expert system for predicting unstable angina based on Bayesian networks 1 Introduction 2 Bayesian networks 2.1 Naïve Bayes 2.2 Tree Augmented Network, TAN 2.3 Forest Augmented Network, FAN 3 Data used, methodology and results 3.1 Dataset 3.2 Methodology and results 4 Web application 4.1 Application front-end 4.2 Application back-end 4.3 Probabilistic Inference motor 5 Conclusions References 
work_xyjg77ll3vdwzpdb44t55ogj3y	----	Strategy of global asset allocation using extended classifier system Expert Systems with Applications 37 (2010) 6611–6617 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Strategy of global asset allocation using extended classifier system Wen-Chih Tsai, An-Pin Chen * National Chiao-Tung University, Institute of Information Management, Hsinchu 30050, Taiwan a r t i c l e i n f o Keywords: Extended classification system Learning classifier system Exchanged traded funds Finance predication 0957-4174/$ - see front matter � 2010 Elsevier Ltd. A doi:10.1016/j.eswa.2010.03.001 * Corresponding author. Tel.: +886 911296906. E-mail addresses: miktsai@gmail.com, wctsaie@t iim.nctu.edu.tw (A.-P. Chen). a b s t r a c t There are several studies about extended classification system (XCS) in past years. XCS model can dynam- ically learn and adapt to the change of environments for maximizing the desired goals. This paper con- ducts simulation to apply XCS to global asset allocation in the country-specific exchanged traded funds (ETFs). Since international stock price trend is influenced by unknown and unpredictable surround- ings, using XCS to model the fluctuations on global financial market allows for the discovery of the pat- terns of the future trends. As such, the benefits of international asset diversification can be achieved in a tax-efficient way with country-specific ETFs at a low transaction cost with minimized tracking error. These empirical results indicate that XCS is capable of evolving over time, and thus XCS can provide a good indicator for future global asset allocation decision-making aiming at maximized profit. � 2010 Elsevier Ltd. All rights reserved. 1. Introduction Recently, exchanged traded funds (ETFs) have become very popular investment products for index trading all over the world since their first introduction at the beginning of last decade. ETFs are the leading financial innovation of the last decade (Fuhr, 2001). ETFs closely track the performance of corresponding indices. ETFs offer the benefits of diversification and index tracking at a low cost. The first ETF, SPDR, was launched on AMEX in 1993 and was designed to passively mimic the S&P500 index. Furthermore, at the end of 2007, there were 837 ETFs with 1324 listings with assets of USD $1212 billions and managed by 107 managers on 56 ex- changes across the world. Most days, two or three ETFs are on the list of the top five most actively traded stocks on the AMEX. Additionally, since the trend of fund price is affected by many man-made and natural elements using dynamic machine-learning tool for the fund analysis is more suitable and adaptive than tradi- tional methods. Learning classifier system (LCS) consists of a set of steps and classifiers for discovering rules of genetic and non-genet- ic operators (Miffre, 2007). In LCS bibliography, a wide range of re- sources has been covered (Karpoff & Jonathan, 1987; Kovacs, 2000); however, the applications of XCS on financial issues (Lanzi, Stolzmann, & Wilson, 2000; Leigh, Modani, Purvis, & Robert, 2002) are much fewer when compared with its LCS counterpart. The fol- lowing are reasons to use XCS on dynamic and noisy environments: ll rights reserved. smc.com (W.-C. Tsai), apc@ � XCS is capable of making real-time and accurate responses. � XCS has been shown to properly learn from noisy, complex, and non-linear environments when the outside information contin- uously changes. � XCS is able to evaluate rules that are ideal for modeling prob- lems without retraining all data. � XCS, generalized under predefined conditions, can discover gen- erally accurate rules to perform on a variety of problem domains. � XCS can adjust itself to strengthen its inward knowledge step by step. � XCS assigns rule fitness based on the accuracy of the rule rather than on the reward payoffs. Recently there have been several investigations into applying LCS to machine learning and data mining classification problems, (Amin & Kat, 2003; Andrea, 1995; Trippi & DeSieno, 1992). This pa- per continues this investigation by applying an adaptation of a re- cently developed XCS, Wilson’s XCS, to a large multi-class benchmark data set available at the 24 iShares MSCI (Morgan Stan- ley Capital International) country funds. This paper is structured as follows: Section 1 to introduce the study’s motivation and goals, Section 2 to examine the literature, Section 3 to briefly describe XCS in our model, Section 3 to describe the data set and the exper- imental procedure adopted, Section 5 to present the results, and Section 6 to conclude the result and future study direction. 2. Literature review On basis of analysis of past studies, we divide the related stud- ies into four parts, which include literature on ETFs, artificial http://dx.doi.org/10.1016/j.eswa.2010.03.001 mailto:wctsaie@tsmc.com mailto:apc@ http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 6612 W.-C. Tsai, A.-P. Chen / Expert Systems with Applications 37 (2010) 6611–6617 intelligence and portfolio, technical analysis and technical indica- tors and classifier systems. 2.1. International ETFs In the past, studies by Cumby and Glen (1990); Shukla and Singh (1997), Redman, Gullett, and Manakyan (2000) analyzed mutual fund performance and showed evidence that international mutual funds can outperform the US stock market. Cumby and Glen examined the performance of 15 US-based internationally diversified mutual funds from 1982 to 1988. The findings showed that mutual funds outperformed the US Index. Enu, Kolodny, and Resnick (1991) investigated 19 US-based international mutual funds from 1977 to 1986, and concluded that majority of interna- tional mutual funds outperformed the US stock market. However, Shukla and Singh (1997), Redman et al. (2000), of- fered distinct conclusions. Shukla & Singh, 1997 evaluated the per- formance of the US-based global equity mutual funds during 1988–1995 including a total of 20 global and 76 domestic funds. They showed that both global funds and US domestic funds under- performed the S&P 500 Index. Redman also showed that the inter- national portfolio underperformed the US equity portfolio. Bhargava (2001) suggested both the international equity managed funds and mutual funds underperformed the S&P500 Index. De- spite the significance associated with their studies, they might have problem making real-time decisions while incorporating tra- ditional models into their models. This paper is thus based on a dynamical and real-time model in order to optimize global asset allocation. 2.2. Global asset allocation The country-specific ETFs global asset allocation is an invest- ment strategy that attempts to exploit short-term international market inefficiencies by establishing positions in an assortment of markets with a goal to profit from relative movements across those international markets. These decisions can usually be broken down briefly into two processes. First, select a list of countries that have growth potential or are currently undervalued. The process is called portfolio selection. Secondly, investigate these ETFs price movements of each selected countries, and execute correct trading strategies at appropriate timing. This paper focuses on 24 iShare MSCI country-specific funds. Like country-specific open and closed-end index funds, country- specific iShares increase mean–variance efficiency. On the other hand, unlike country index funds, which could only be transacted at a cut-off time (such as 4 pm) everyday, the ETFs can be bought or sold at any time during the trading day, offering one or more flexibilities compared to their country-specific index fund counterparts. 2.3. Sharpe ratio This paper starts by testing whether the returns of 24 iShares MSCI country-specific ETFs are normally distributed and better than the XCS model in the dynamic environment. Skewness and Kurtosis are statistics that very often are used to describe the height and the symmetric of distribution of data test for normality (Amin & Kat, 2003; Wachter & Warusawitharana, 2008). This Sharpe ratio measures are used to test the ETF’s performance. Sharpe (1992) proposed the ratio that is mainly used to rank alternative portfolios based on their historic reward-to-variability ratio: SRi ¼ Ri � Rf ri ð1Þ where Ri is the historic mean return on ETF-i over the interval con- sidered, ri is the historic standard deviation of the return on ETF-i over the interval considered and Rf is the average risk-free rate over the interval considered. 2.4. Risk free rate In theory, the risk-free rate is the minimum return an investor expects for any investment unless the potential rate of return is greater than the risk-free rate. In practice, however, the risk-free rate does not exist since even the safest investments carry a very small amount of risk. The interest rate on a three-month US Trea- sury bill is often used as a risk-free rate (Allen & Karjalainen, 1999; Kashima, 2007; Shukla & Singh, 1997). In this study, we also use the US three-month Treasury bill as the risk-free rate. The US three-month Treasury bill historically is obtained from the Board of Governors of the Federal Reserve database. Since we try to view from US investors’ standpoints, the US domestic Treasure bill can be used to measure the risk-free rate. 2.5. XCS XCS is based on the Learning Classifier System (LCS) (Holland, 1992; Karpoff, 1987; Kovacs, 2000), which is a general and inde- pendent machine learning system. LCS, which was proposed by John H. Holland (Andrea, 1995; Butz & Wilson, 2000), is an online step-by-step rule base because it includes both genetic algorithm and strength learning. LCS can be classified as an extended genetic algorithm or an algorithm of strength learning. In LCS, strength learning element is used to distinguish suitable or unsuitable rules and solve the rule conflict problem while genetic algorithm is used to find good and new rules, and eliminate the unsuitable rules. XCS retains the main frames of LCS, but also makes some changes. Firstly, XCS uses precision as the rate of fitness. Secondly, it changes the rule discovery component from acting on the whole population to the population having the same states and actions. Thirdly, it uses Q-learning-like algorithm to substitute the Bucket brigade algorithm. And lastly, it removes the message board. 2.6. Knowledge integration Knowledge integration can be considered as a multi-objective optimization problem (Sakai & Masuyama, 2008; Yuan & Zhuang, 1996). Due to the huge searching space, the optimization problem is often very difficult to be solved. A genetic algorithm is usually used to discover a desirable optimal set of rules. The application of a GA in search of the optimal rule set for machine learning is known as Genetic Based Machine Learning (GBML). A well-known GBML architecture is the so-called LCS developed by Andrea (1995); Holland (1986) and Holland, Holyoak, Nisbett, and Thagard (1986). More recent GBML architectures are the Extended Classifier System (XCS) developed by Butz and Wilson (2000), the Anticipa- tory Classifier System by Lanzi et al. (2000), and EpiCS by Butz and Wilson (2000). 3. System architecture This paper implements the system architecture shown in Fig. 1. This is based on the Wilson’s XCS classifier system (Butz & Wilson, 2000). XCS retains the main frames of LCS, but also makes some changes. Firstly, XCS uses precision as the rate of fitness in the transaction data-encoding module. Secondly, it changes the rule discovery component from acting on the whole population to the population having the same states and actions. Thirdly, it uses Q- learning-like algorithm to substitute the Bucket brigade algorithm Fig. 1. Architecture of XCS. Fig. 2. Algorithm of XCS. W.-C. Tsai, A.-P. Chen / Expert Systems with Applications 37 (2010) 6611–6617 6613 in the knowledge extraction module. And lastly, it rewards the re- sult to the knowledge integration module. In this section we tries to elaborate more on the contents of the various models of the architecture as implemented in XCS and illustrated in Fig. 1. 3.1. Transaction data-encoding model Fig. 1 shows a transaction data-encoding module, a group of financial indices of the country whose stock index is tracked by that country’s specific ETF, with same syntax form a classifier pop- ulation. The group consists of: 1. Detecting condition section C1 ^ C2 ^�� �^ Cn; Ci 2f0; 1;�g L ; 1 6 i 6 n ð2Þ That is composed of at least one condition. The condition may include a state of positive (1), negative (0), or do not-care (–). And for the entry of the condition, the associated state must be satisfied. Eq. (2) stands for conditions 1–n are satisfied and in a predetermined sequence. The sequence could be condition 1 is followed by condition 2, which is followed by condition 3, until condition n takes place. 2. Action section A 2fa1; . . . ; amg ð3Þ Action section to represent the candidate classifiers action. 3. Rule prediction p to evaluate classifiers utility. 4. Prediction error standing for the difference between actual ben- efit and prediction p. 5. Fitness F to evaluate the precision of prediction p from predic- tion error. 3.2. Knowledge extraction model This model consists of: � Execution section. XCS interacts with environment at discrete time t in terms of environment states St, utilizes St to compare with population [P]’s conditions, and copies the matched classifiers to match set [M]. XCS further computes the weighted averages of each action in the match set [M], so as to build up a system prediction PA(a). With PA(a), XCS further selects an action ai, and classifiers that have action ai from match set [M], and puts them in action set [A]. The system then executes ai, and receives a delay reward rt�1 in discrete time t + 1. The same process continues until the objective problem is solved. � Reinforcement section. XCS uses reward r to update parameters of strength learning of classifier in action set [A]. The update of prediction value p may be as follows: C � p C � p þðR � C � pÞ� k ð4Þ R ¼ rt�1 þðE � sÞ ð5Þ C is the classifier, k is the learning rate (0 < k 5 1), rt�1 is the re- ward of previous step, E is the max system expected value and s is the discount factor. The update of predicted error value e: C � e C � e þðjR � C � pj� C:eÞ� k ð6Þ The equation of fitness F: C � F C � F þðC � l0 � C � FÞ� k ð7Þ C � l0 C � lP x2½A�C � lx ð8Þ C � l 1 if C � e < e0 aðe0=C � eÞ b otherwise � ð9Þ e0 is the tolerance of predicted error value (e0 > 0), a, b is the con- stant of precision control l (0 < a < 1; b > 0). From fitness function F in Eq. (5), we know that the fitness of classifier in XCS evaluates precision of classifier in the same ac- tion set [A], and has an invert function relationship with pre- dicted error e. Table 3 iShare MSCI Japan Index (EWJ) from 2003/1/2 to 2009/5/30. Symbol Date Open High Low Close Volume EWJ 2003/1/2 7 7.1 7 7.08 1,529,000 6614 W.-C. Tsai, A.-P. Chen / Expert Systems with Applications 37 (2010) 6611–6617 3.3. Knowledge integration model This XCS model focuses on the genetic algorithm (GA) in Fig. 2. Genetic algorithm is used to eliminate unsuitable classifiers in ac- Table 2 iShare FTSE/Xinhua China 25 Index (FXI) from 2004/10/12 to 2009/05/30. Symbol Date Open High Low Close Volume FXI 2004/10/12 53.6 53.85 53.38 53.7 248,200 FXI 2004/10/13 53 53.31 52.2 52.32 369,100 FXI 2004/10/14 51.96 52.1 51.45 51.62 119,800 FXI 2004/10/15 52.05 52.64 52.03 52.4 234,500 FXI .. . .. . .. . .. . .. . .. . EWJ 2003/1/3 7.05 7.09 7 7.07 360,400 EWJ 2003/1/6 7.15 7.23 7.08 7.2 2,614,900 EWJ 2003/1/7 7.02 7.05 6.96 6.97 890,700 EWJ .. . .. . .. . .. . .. . .. . Table 1 Sample data list. ETFs region Extend-traded funds name Symbol Inception date Asia Pacific iShares MSCI Australia Index EWA 1996/3/12 North American iShares MSCI Canada Index EWC 1996/3/12 European iShares MSCI Sweden Index EWD 1996/3/12 European iShares MSCI Germany Index EWG 1996/3/12 Asia Pacific iShares MSCI Hong Kong Index EWH 1996/3/12 European iShares MSCI Italy Index EWI 1996/3/12 Asia Pacific iShares MSCI Japan Index EWJ 1996/3/12 European iShares MSCI Belgium Index EWK 1996/3/12 European iShares MSCI Switzerland Index EWL 1996/3/12 Asia Pacific iShares MSCI Malaysia Index EWM 1996/3/12 European iShares MSCI Netherlands Index EWN 1996/3/12 European iShares MSCI Austria Index EWO 1996/3/12 European iShares MSCI Spain Index EWP 1996/3/12 European iShares MSCI France Index EWQ 1996/3/12 Asia Pacific iShares MSCI Singapore Index EWS 1996/3/12 Asia Pacific iShares MSCI Taiwan Index EWT 2000/6/20 European iShares MSCI United Kingdom Index EWU 1996/3/12 North American iShares MSCI Mexico Index EWW 1996/3/12 Asia Pacific iShares MSCI South Korea Index EWY 2000/5fi South American iShares MSCI Brazil Index EWZ 2000/7/10 South American iShares MSCI South Africa Index EZA 2003/2/14 Asia Pacific Xinhua China 25 Index Fund FXI 2004/10/15 North American iShares S&P 500 Index IVV 2000/5/26 North American iShares Dow Jones US Industrial IYJ 2000/7/21 Table 4 Data coded. EITs region Symbol Date Open Hi Asia Pacific EWA 2003/1/2 19.91 1 Asia Pacific EWH 2006/3/2 13.42 1 Asian Pac& EWJ 2006/3/2 13.75 1 Asia Pacific EWM 2006/3/2 7.36 Asia Pacific EWS 2006/3/2 8.65 Asia Pacific EWT 2006/3/2 13.01 1 Asia Pacific EWY 2006/3/2 47 4 Asia Pacific FXI 2006/3/2 73.34 7 European EWD 2006/3/2 24.05 2 European EWG 2006/3/2 22.12 2 European EWI 2006/3/2 27.24 2 European EWK 2006/3/2 20.65 2 European EWL 2006/3/2 20.58 2 European EWN 2006/3/2 21.8 2 European EWO 2006/3/2 29.8 3 European EWP 2006/3/2 40.3 4 European EWQ 2006/3/2 27.88 2 European EWU 2006/3/2 19.6 1 North American EWC 2006/3/2 23.72 2 North American EWW 2006/3/2 39.05 3 North American IVV 2006/3/2 129.04 12 North American IYJ 2006/3/2 61 6 South American EWZ 2006/3/2 42.96 4 South American EZA 2006/3/2 19.91 1 tion set [A] rather than the whole population. In doing so, genetic algorithm starts when action set [A] have not been executed by ge- netic algorithm for an average time value. When genetic algorithm executes, two classifiers and crossover at a v probability might be randomly selected. Also, it will mutate at probability. 4. Experiment 4.1. Data This paper targets 24 iShares MSCI country-specific ETFs from the iShares web database (http://www.ishares.com). The data in- clude daily opening price, close price, maximum price, minimum price and trade volume over the period from Jan 2003 to June 2009, resulting in 76 monthly observations shown in Tables 2–4. As previously mentioned, the reason for choosing iShares MSCI country-specific ETFs is to achieve international diversification. Table 1 indicates the basic information of these ETFs including the region, symbol, name, and inception date. The inception date for most of the ETFs is 12 March 1996, and the latest inception date, for iShares MSCI-Xinhua China 25 (FXI), falls on 15 October 2004. Thus, our sample period covers all ETFs historical data. All of these ETFs belong to Barclays Global Investors Group, known as iShares. We use 24 iShares MSCI country funds measured by the MSCI individual country index. These ETFs include eight iShares from Asia Pacific countries (Australia, Hong Kong, Japan, Malaysia, Singapore, South Korea, Japan, and China), 10 iShares from European countries (Austria, Belgium, France, Germany, Italy, Netherlands, Spain, Sweden, Switzerland, and UK), four iShares gh Low Close Volume Coded 9.95 19.75 19.91 204,400 –0010 3.47 13.36 13.45 392,000 10111 3.75 13.61 13.72 18,600,300 01111 7.4 7.33 7.38 497,600 10010 8.66 8.59 8.62 467,900 10101 3.04 12.9 13.03 2,354,700 11011 7.08 46.57 46.86 762,500 01101 3.37 72.8 73.31 379,600 00111 4.34 23.98 24.34 49,000 11000 2.23 22.01 22.23 603,000 10010 7.26 27.04 27.26 66,600 10010 0.84 20.62 20.84 90,800 10000 0.68 20.4 20.67 71,500 10010 1.86 21.61 21.86 86,900 10000 0.16 29.62 30.16 158,600 –––000 0.43 40.02 40.43 28,000 11000 7.99 27.8 27.98 559,000 10010 9.76 19.56 19.75 88,400 10100 3.9 23.61 23.88 488,500 10000 9.17 38.8 39.03 459,400 10–00 9.6 128.81 129.48 863,100 10100 1.15 60.84 61.06 22,600 11011 3.18 42.57 43.14 2,135,000 10000 9.95 19.75 19.91 204,400 11011 http://www.ishares.com W.-C. Tsai, A.-P. Chen / Expert Systems with Applications 37 (2010) 6611–6617 6615 form North American countries (S&P500, Dow Jones, Canada and Mexico) and two iShares from countries in Southern hemisphere (Brazil and South Africa). 4.2. Data coded and portfolio optimizer In the experiment, we codify the daily information of the cho- sen countries’ ETF into a five-code string. The daily information in- cludes opening price (‘‘open”), maximum price (‘‘high”), minimum price (‘‘low”), closing price (‘‘close”), and trade volume (‘‘volume”). Table 5 Daily global asset allocation portfolio. Extend-traded fund name Ticker Date iShares Dow Jones US Industrial IYJ 200 iShares Goldman Sachs Technology Index IGM 20O iShares MSCI Australia Index EWA 200 iShares MSCI Austria Index EWO 200 iShares MSCI Belgium Index EWK 200 iShares MSCI Brazil (Free) Index EWZ 200 iShares MSCI Canada Index EWC 200 iShares MSCI E AFE Index Fund EFA 200 iShares MSCI EMU Index EZU 200 iShares MSCI France Index EWQ 200 iShares MSCI Germany Index EWG 200 iShares MSCI Horn Kong Index EWH 200 iShares MSCI Italy Index EWI 200 iShares MSCI Japan Index EWJ 200 iShares MSCI Malaysia (Free) Index EWM 200 iShares MSCI Mexico (Free) Index EWW 200 iShares MSCI Netherlands Index EWN 200 iShares MSCI Singapore (Free) Index EWS 200 iShares MSCI South Africa Index EZA 200 iShares MSCI South Korea Index EWY 200 iShares MSCI Spain Index EWP 200 iShares MSCI Sweden Index EWD 200 iShares MSCI Switzerland Index EWL 200 iShares MSCI Taiwan Index EWT 200 iShares MSCI United Kingdom Index EWU 200 iShares Russell 1000 Index IWB 200 iShares S&P 500 Index IVV 200 Xinhua China 25 Index Fund FXI 200 NASDAQ 100 Trust Shares QQQ 200 Table 6 Traditional Sharpe ratio asset allocation. ETFs region Extend-traded funds name Symbol Asia Pacific iShares MSCI Australia Index EWA Asia Pacific iShares MSCI Hong Kong Index EWH Asia Pacific iShares MSCI Japan hides EWJ Asia Pacific iShares MSCI Malaysia Index EWM Asia Pacific iShares MSCI Singapore Index EWS Asia Pacific iShares MSCI Taiwan Index EWT Asia Pacific iShares MSCI South Korea Index EWY Asia Pacific iShares FTSE/Xinhua China 25 Index FXI European iShares MSCI Sweden Index EWD European iShares MSCI Germany Index EWG European iShares MSCI Italy Index EWI European iShares MSCI Belgium Index EWK European iShares MSCI Switzerland Index EWL European iShares MSCI Netherlands Index EWN European iShares MSCI Austria Index EWO European iShares MSCI Spain Index EWP European iShares MSCI France Index EWQ European iShares MSCI United Kingdom Index EWU North American iShares MSCI Canada Index EWC North American iShares MSCI Mexico Index EWW North American iShares S&P 500 Index IVV North American iShares Dow Jones US Industrial IYJ South American iShares MSCI Brazil Index EWZ South American iShares MSCI South Africa Index EZA US T-Bill The daily information for the chosen ETF on 3/2/2006 and its coded strings are shown in Table 4. It is worth noting that the coded string is configured to represent the differences in the daily infor- mation, and we further determine whether the differences fall into predetermined ranges or not before assigning the state code (0, 1, or –) to them. It is also worth noting that the ranges could be open- ended. The difference in daily information could be the differences of different day’s averaged price moving and averaged trade vol- ume (Brock, Lakonishok, & LeBaron, 1992). In order to simplify the experiment, the experiment uses 1 for positive and 0 for Coded As-is (%) To-Be (%) 6/3/2 11011 6 5 6/3/2 10000 2 1 6/3/2 –0010 2 2 6/3/2 ––000 2 2 6/3/2 10000 2 2 6/3/2 10000 1 2 6/3/2 10000 2 2 6/3/2 10000 2 2 6/3/2 –0111 5 6 6/3/2 10010 2 2 6/3/2 10010 2 2 6/3/2 10111 8 9 6/3/2 10010 2 1 6/3/2 01111 11 12 6/3/2 10010 3 3 6/3/2 10–00 2 2 6/3/2 10000 1 0 6/3/2 10101 2 2 6/3,2 11011 13 14 6/3/2 01101 4 5 6/3/2 11000 1 0 6/3/2 11000 1 0 6/3/2 10010 1 0 6/3/2 11011 5 6 6/3/2 10100 1 1 6/3/2 00–10 0 0 6/3/2 10100 2 2 6/3/2 00111 9 10 6/3/2 11000 6 5 Annual return (%) Annual std. deviation (%) Sharpe ratio (%) 14.72 17.47 63.8237 2.15 22.06 9.7461 11.26 19.45 57.8920 17.96 19.65 91.3995 6.95 22.49 30.9026 10.66 34.11 31.2518 31.55 33.81 93.3156 56.00 38.10 146.9816 10.73 32.57 32.9444 4.73 33.69 14.0398 6.89 22.94 30.0349 9.12 23.16 39.3782 5.43 17.71 30.6606 2.15 26.08 8.2439 26.31 19.91 132.1447 14.21 24.15 58.8406 3.70 23.90 15.4812 0.74 16.26 4.5510 10.61 16.92 62.7069 4.14 24.62 16.8156 �0.50 17.90 �2.7933 2.50 18.25 13.6986 20.07 49.14 40.8425 27.10 45.42 59.6653 3.57 Fig. 4. Accumulated portfolio. 6616 W.-C. Tsai, A.-P. Chen / Expert Systems with Applications 37 (2010) 6611–6617 negative in terms of the difference in the daily information. And we further use 1 to associate with ‘‘buy” and 0 to associate with ‘‘sell.” In other words, the system will adjust the weight in the global iShares. The daily rules discovery is shown in Table 5, the column of ‘‘as-is” is indicative of information at day t � 1 while the column of ‘‘to-be” is reserved for day t (today). 4.3. Traditionally Sharpe ratio The traditional portfolio model used the ‘‘Sharpe ratio” to eval- uate the optimal asset allocation. Hence we compare the monthly Sharpe ratio global asset allocation with the XCS model global asset allocation in Table 6. This Sharpe ratio measures are used to test the ETF’s performance (Sharpe, 1992) 4.4. International global asset allocation In this paper, we implement the integration knowledge model with XCS expert system to study the global markets that include the US, China, Taiwan, Japan, South Africa, and Malaysia. In Fig. 3, the iShares track the indices of international capital markets before any global asset allocation strategies tracking those indices could be implemented for our international global asset allocations. The global allocation element of ETFs contributes to the global risk diversification and generates sufficient gains in a transparent and low cost manner that is not easily achievable by global index funds. 4.5. Experiment result The results of the experiments are summarized in Fig. 4. Fig. 4 shows the profit accumulation result. The average of the cumu- lated profit is better than the traditional Sharpe ratio, and the high- est profit is about 840,888 units after 1300 days, which is about 6.5% per day. Therefore, it is a good performance when the index ended up lower at the end of the 120-day period than the start thereof. Our XCS model is about 74.45% better than the asset alloca- tion strategy on the basis of Sharpe ratio. Fig. 3. International markets analysis from XCS exports system. 5. Conclusion As we know, the country-specific ETFs offer the benefits of international portfolio diversification at a lower cost with a lower tracking error in a more tax-efficient way than passive open or closed-end county funds. This paper focuses on the soft computing algorithm, XCS, and compares with the traditional asset allocation model, according to Sharpe ratio. The statistical shows that dy- namic artificial intelligence model is better than the non-efficiently monthly Sharpe ratio model. Additionally, using a limited numbers of factors from the real international market this paper has shown some promise using extended classifier trading mechanism in country-specific ETFs. The XCS experts system consists of Wilson’s XCS technique, which provides a good online learning system for our model. In the fast changing security market, Genetic algorithm, rule base, neural network etc. do not satisfy our needs. Rather, XCS’s online learning is generally perceived as a more suitable op- tion. XCS can give trader or investor a real-time advice to make rel- atively more accurate trading decisions in the international markets. In the future, although the experiment has shown good results, the model proposed by the current paper may still have some rooms to improve by having the inputted factors changed. Especially, this work does not include the commodities ETF. In addition, this study includes no short ETF either. Hence, the study could be further developed after having the above-mentioned inputted factors considered. The next step would be to verify XCS in different products such as commodities ETFs, and actively man- aged ETF. References Allen, F., & Karjalainen, R. (1999). Using genetic algorithms to find technical trading rules. Journal of Financial Economics, 51, 245–271. Amin, G. S., & Kat, H. M. (2003). Hedge fund performance 1999–2000: Do the ‘‘Money Machines” really add value? Journal of Financial, 38, 251–274. Andrea, G.-S. (1995). Holland classifier systems. ACM SIGAPL APL Quote Quad. Proceedings of the international conference on applied programming languages, 25(4). W.-C. Tsai, A.-P. Chen / Expert Systems with Applications 37 (2010) 6611–6617 6617 Bhargava, A. (2001). Stochastic specification and the international GDP series. Econometrics Journal, Royal Economic Society, 4(2), 7. Brock, W., Lakonishok, J., & LeBaron, B. (1992). Simple technical trading rules and the stochastic properties of stock returns. The Journal of Financial, XLVII(5), 1731–1764. Butz, M. V., & Wilson, W. (2000). An algorithmic description of XCS. Technical Report 2000017, Illinois Genetic Algorithms Laboratory. Cumby, R., & Glen, J. (1990). Evaluating the performance of international mutual funds. The Journal of Finance, 45(2), 497–521. Enu, C. S., Kolodny, R., & Resnick, B. G. (1991). U.S. based international mutual funds: A performance evaluation. Journal of Portfolio Management, 17, 88–94. Fuhr, D. (2001). Exchange- traded funds: A primer. Journal of Asset Management, 3, 260–273. Holland, J. H. (1992). Adaptation in natural and artificial systems. MIT Press. Holland, J. H. (1986). Escaping brittleness: The possibilities of general purpose learning algorithms applied to parallel rule-based systems. Machine Learning – An Artificial Intelligence Approach, 2, 593–624. Holland, J. H., Holyoak, K. J., Nisbett, R. E., & Thagard, P. R. (1986). Induction: Processes of inference, learning, and discovery. MIT Press. Karpoff, J. M. (1987). The relationship between price changes and trading volume. Journal of Financial and Quantitative Analysis, 22, 109–126. Kashima, H. (2007). Risk-sensitive learning via minimization of empirical conditional value-at-risk. IEICE Transactions on Information and Sciences, E90- D(December), 2043–2052. Kovacs, T. (2000). A learning classifier systems bibliography. School of Computer Science, University of Birmingham. Lanzi, P.-L., Stolzmann, W., & Wilson, S. W. (Eds.). (2000). Learning classifier systems: From Foundations to applications. Lecture notes in artificial intelligence, LNAI (Vol. 1813). Springer-Verlag. Leigh, W., Modani, N., Purvis, R., & Robert, T. (2002). Stock market trading rule discovery using technical charting heuristics. Expert System with Application, 23, 155–159. Miffre, J. (2007). Country-specific ETFs: An efficient approach to global asset allocation. Journal of Asset Management, 8(2). Redman, A. L., Gullett, N. S., & Manakyan, H. (2000). The performance of global and international mutual funds. Journal of Financial and Strategic Decisions, 13, 75–85. Sakai, H., & Masuyama, S. (2008). Cause information extraction from financial articles concerning business performance. IEICE Transactions on Information and Systems, E91-D(April), 959–968. Sharpe, W. F. (1992). Asset allocation: Management style and performance measurement. Journal of Portfolio Management. Shukla, R., & Singh, S. (1997). A performance evaluation of global equity mutual funds: Evidence from 1988 to 1995. Global Finance Journal, 8(2), 279–293. Autumn–Winter. Trippi, R. R., & DeSieno, D. (1992). Trading equity index futures with a neural network. Journal of Portfolio Management, 27–33. Wachter, J. A., & Warusawitharana, M. (2008). Predictable returns and asset allocation: Should a skeptical investor time the market? Journal of Econometrics, 148, 162–781. Yuan, Y., & Zhuang, H. (1996). A genetic algorithm for generating fuzzy classification rules. Fuzzy Sets and Systems, 84, 1–19. Strategy of global asset allocation using extended classifier system Introduction Literature review International ETFs Global asset allocation Sharpe ratio Risk free rate XCS Knowledge integration System architecture Transaction data-encoding model Knowledge extraction model Knowledge integration model Experiment Data Data coded and portfolio optimizer Traditionally Sharpe ratio International global asset allocation Experiment result Conclusion References 
work_y3qc2gtvuvabbgs5bv4d2r5rii	----	An expert system for the diagnosis of faults in rotating machinery using adaptive order-tracking algorithm Expert Systems with Applications 36 (2009) 5424–5431 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a An expert system for the diagnosis of faults in rotating machinery using adaptive order-tracking algorithm Jian-Da Wu a,*, Mingsian R. Bai b, Fu-Cheng Su b, Chin-Wei Huang c a Institute of Vehicle Engineering, National Changhua University of Education, 1 Jin-De Road, Changhua City, Changhua 500, Taiwan, ROC b Department of Mechanical Engineering, National Chiao-Tung University, Hsin-Chu, Taiwan, ROC c Department of Mechanical and Automation Engineering, Da-Yeh University, Changhua, Taiwan, ROC a r t i c l e i n f o a b s t r a c t Keywords: Signal processing Fault diagnosis Order-tracking Adaptive RLS filter 0957-4174/$ - see front matter � 2008 Elsevier Ltd. A doi:10.1016/j.eswa.2008.06.059 * Corresponding author. E-mail address: jdwu@cc.ncue.edu.tw (J.-D. Wu). This paper describes an application of an adaptive order-tracking technique for the diagnosis of faults in rotating machinery. Conventional methods of order-tracking are primarily based on Fourier analysis with reference to shaft speed. Unfortunately, in some applications of order-tracking performance is limited, such as when a smearing problem arises and also in a multiple independent shaft system. In this study, the proposed fault diagnostic system is based on a recursive least-square (RLS) filtering algorithm. The problem is treated as the tracking of various frequency bandpass signals. Order amplitudes can be calcu- lated with high-resolution in real-time implementation. The algorithm is implemented on a digital signal processor (DSP) platform for diagnosis and evaluated by experimental investigation. An experimental investigation is implemented to evaluate the proposed system in two applications of gear-set defect diag- nosis and in the diagnosis of damaged engine turbocharger blades. The results of the experiments indi- cate that the proposed algorithm is effective in fault diagnosis for both experimental cases. Furthermore, a characteristic analysis and experimental comparison of a vibration signal and a sound emission signal for the present algorithm are also presented in this report. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction Traditionally, the condition of rotating machinery such as fans, compressors, motors and engines can be monitored by measuring the respective vibration signal or sound emission signal. These sig- nals normally consist of a combination of the basic frequency with discrete or narrowband frequency components and the harmonics thereof, most of which are related to the revolution of the machin- ery. The sound emission and vibration energy are increased when the machinery is damaged. An example result of a sound emission power spectrum level measured from the wheel-blades of an inter- nal combustion (IC) engine turbocharger is shown in Fig. 1. The conventional fault diagnostic technique is to observe the ampli- tude difference in the time or the frequency domain for diagnosis of damage. Recently, the order-tracking technique has become an impor- tant approach for diagnosing fault in rotating machinery. Interest in diagnosis using the order-tracking technique has grown signifi- cantly, having advanced with the progress of digital signal process- ing algorithms and technology in the last two decades (Biswas, Pandey, Bluni, & Samman, 1994; Chen, Du, & Qu, 1995; Gelle, Colas, ll rights reserved. & Serviere, 2001; Lin & Qu, 2000; Shibata, Takahashi, & Shirai, 2000). The conventional order-tracking method is primarily based on Fourier analysis with reference to shaft revolution (Lee & White, 1998; Vold & Leuridan, 1993). Unfortunately, re-sampling process- ing is generally required in the fast Fourier transform (FFT) meth- ods to compromise between time and frequency resolution for varying revolutions. However, in the conventional FFT methods, a smearing problem generally arises in practical implementation, particularly at low revolution speeds. In addition, the conventional methods are ineffective for application to certain critical conditions such as a fixed sampling frequency, and FFT analysis with a track- ing technique is ineffective when the shaft speed varies rapidly. In this study, an adaptive order-tracking fault diagnostic tech- nique using both vibration signals and sound emissions is applied to the diagnosis of damage in gear-sets and engine turbocharger blades. According to recent studies by Haykin (1996) and Bai, Jeng, and Chen (2002) there exists some conclusions for adaptive filter- ing algorithms and their application to order-tracking techniques. The proposed adaptive fault diagnostic system is based on the recursive least-square (RLS) algorithm (Bai et al., 2002). Similar to conventional methods, the RLS method also requires informa- tion on shaft or engine revolution. The algorithm is essentially sample-based; thus, order amplitudes can be calculated in a real- time fashion. The method is well suited for high-resolution mailto:jdwu@cc.ncue.edu.tw http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Fig. 1. Sound power spectrum level of sound emissions from turbocharger blades. A solid line depicts blades without damage; broken lines depict blades of which one is damaged. J.-D. Wu et al. / Expert Systems with Applications 36 (2009) 5424–5431 5425 tracking of closely spaced orders or crossing orders. The filter algo- rithm is implemented in a TMS320C32 DSP platform for evaluating the performance in a practical application of the diagnosis of dam- age in gear-sets and IC engine turbocharger blades. In fault diagnostic techniques to date, measurement of the vibration signal has become most widely used when a reference signal is available. Unfortunately, in some practical applications, such a vibration signal is unavailable. Measurement of high-fre- quency sound emissions serves as a promising alternative to con- dition monitoring of many types of rotating machinery (Mba, 2002; Toutountzakis & Mba, 2003). During operation of the machinery, defects at different locations will generate characteris- Transversal filter ˆ ( 1)w n − Adaptive weight control mechanism Input vector ( )u n *ξ (n) k(nΣ Σd*(n) + _ )1()(H − ∧ nn wu Gain a b Fig. 2. Representations of RLS algorithm. (a) B tic frequencies. However, in the present study, both vibration sig- nals and sound emission signals are used to evaluate the proposed diagnostic technique. The details of the proposed adap- tive filtering with an RLS algorithm are described in the following section. 2. Principle of adaptive order-tracking technique using RLS algorithm The conventional algorithms used in fault diagnostic techniques fall into two categories. One is Fourier transform with a fixed sam- pling rate for obtaining frequency domain information; the other is - + Desired response d(n) Output ˆ ( 1) ( )Hw n u n− Error ( )nξ z -1I Σ uH(n) Negative unity feedback )(n ∧ w )1( − ∧ nw lock diagram and (b) signal-flow graph. 0 1000 2000 3000 4000 5000 6000 7000 8000 -350 -300 -250 -200 -150 -100 -50 0 50 Iteration Number E rr o r (d B ) Fig. 3. A comparison of convergence speeds and estimation errors in various adaptive filters. A solid line depicts the Kalman filter; a dash-dot line depicts the RLS; a dotted line depicts the LMS. DSP controller Mic. Gear 2 Gear 1Coupling Frequency converter Accelerometer Fiber optical sensor Motor D/A A/D A/DA/D Fig. 4. Experimental arrangement of gear-set defect diagnosis. Tim 0 2 4 S p e e d ( rp m ) 700 800 900 1000 1100 1200 1300 1400 1500 1600 Fig. 5. Revolution of gear-se 5426 J.-D. Wu et al. / Expert Systems with Applications 36 (2009) 5424–5431 tracking with various sampling rates. The second method employs a re-sampling scheme synchronous with the shaft revolution. The time domain data are hence converted to revolution-domain data. Then the FFT is also applied to obtain the order spectrum with re- spect to engine speed. Both the time and the frequency resolution of this approach are essentially varied with the shaft speed. This FFT order-tracking method relies on accurate measurement of the tachometer signal. In general, the vibration signal or the sound emission signal generated by rotating machinery essentially con- sists of a combination of the basic frequency with narrowband fre- quency components and its harmonic frequencies, most of which are related to the revolution of the machine. Bai et al. (2002) pro- posed an RLS algorithm for adaptive order-tracking technique. In this work, the vibration signal x(t) containing k orders generated by one rotating shaft can be written as xðtÞ¼ ½cos½hðtÞ�� sin½hðtÞ�cos½2hðtÞ�� sin½2hðtÞ� � � �cos½khðtÞ� � sin½khðtÞ�� A1I A1Q A2I A2Q .. . AkI AkQ 2 6666666664 3 7777777775 ð1Þ e (sec) 6 8 10 12 t in experimental case. J.-D. Wu et al. / Expert Systems with Applications 36 (2009) 5424–5431 5427 where AkI and AkQ denote the in-phase and quadrature components, respectively, of kth order. Note that AkI ¼ Ak cos /k; AkQ ¼ Ak sin /k: ð2Þ The amplitude of kth order can be written as jAkj ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi A2kIi þ A 2 kQ q : ð3Þ and the phase of the kth order is obtained by /k ¼ tan�1 AkQ AkI � � : ð4Þ For a discrete-time system, Eq. (1) can be expressed as Time (sec) *1 0 2 4 6 8 10 12 0 1 2 3 x 10 -3 *2 0 2 4 6 8 10 12 0 1 2 3 x 10 -3 *3 0 2 4 6 8 10 12 0 1 2 3 x 10 -3 *4 0 2 4 6 8 10 12 0 1 2 3 x 10 -3 A m pl it ud e Fig. 6. Order figures of vibration signals for gear-set using adaptive RLS filter. A solid line depicts gear without defect; a dashed line depicts gear with a defect. xðnÞ¼ ½cos½hðnÞ�� sin½hðnÞ�cos½2hðnÞ�� sin½2hðnÞ� � � �cos½khðnÞ� � sin½khðnÞ�� A1I A1Q A2I A2Q .. . AkI AkQ 2 6666666666664 3 7777777777775 ð5Þ where n is the discrete-time index (Oppenheim & Schafer, 1999). To solve Eq. (5), collect 2k samples of xðnÞ to form *1 0 2 4 6 8 10 12 0 2 4 x 10 -4 *2 0 2 4 6 8 10 12 0 2 4 6 x 10 -4 *3 0 2 4 6 8 10 12 0 2 4 6 x 10 -4 *4 0 2 4 6 8 10 12 0 2 4 6 x 10 -4 Time(sec) A m pl it ud e Fig. 7. Order figures of sound emission signals for gear-set using adaptive RLS filter. A solid line depicts gear without defect; a dashed line depicts gear with a defect. xðnÞ xðn þ 1Þ .. . xðn þ kÞ xðn þ 2k � 1Þ 2 66666664 3 77777775 ¼ cos½hðnÞ� �sin½hðnÞ� � � � cos½khðnÞ� �sin½khðnÞ� cos½hðn þ 1Þ� �sin½hðn þ 1Þ� � � � cos½khðn þ 1Þ� �sin½khðn þ 1Þ� .. . .. . � � � .. . .. . cos½hðn þ 2kÞ� �sin½hðn þ 2kÞ� � � � cos½khðn þ 2kÞ� �sin½khðn þ 2kÞ� cos½hðn þ 2k � 1Þ� �sin½hðn þ 2k � 1Þ� � � � cos½khðn þ 2k � 1Þ� �sin½khðn þ 2k � 1Þ� 2 66666664 3 77777775 A1I A1Q .. . AkI AkQ 2 66666664 3 77777775 ð6Þ 5428 J.-D. Wu et al. / Expert Systems with Applications 36 (2009) 5424–5431 Here it is assumed that the 2k amplitude parameters AI and AQ remain constant within the interval ½n; n þ 2k � 1�. In view of the special structure of the signal described in Eq. (1), the order-track- ing problem can be recast into a parameter identification form. The estimation error eðnÞ¼ xðnÞ� wTðnÞuðnÞ; ð7Þ uTðnÞ¼ ½cos½hðnÞ�� sin½hðnÞ�cos½2hðnÞ� � sin½2hðnÞ� � � �cos½khðnÞ�� sin½khðnÞ�� ð8Þ is the regressor; WTðnÞ¼ ½A1IðnÞ A1QðnÞ A2IðnÞ A2QðnÞ � � � AkIðnÞ AkQ ðnÞ� ð9Þ is the parameter vector; x(n) is the measurement error. Note that the vector u(n) consists of angular displacements of the shaft; the Fig. 8. (a) Damaged turbocharger compress-wheel-blades. (b) Make sure that turbocharg Anglin, 1993). Turbocharger Acoustic signal A/D A/D Tachometer Microphone Digital si Dynam Fig. 9. Experimental arrangement for diagno vector w(n) consists of the in-phase and quadrature components of all orders to be identified. The parameter identification problem in Eq. (7) can be solved by the method of least-squares (Denbigh, 1998). The problem amounts to finding optimal parameters ŵðnÞ so that the perfor- mance index f(n) is minimized as fðnÞ¼ Xn i¼1 kn�ijeðiÞj2; ð10Þ where the forgetting factor k exponentially weighs the estimation error from the present to the past. Fig. 2 shows the block diagram and signal-flow graph of the RLS algorithm. The optimal solution of the problem can be recursively solved by using the following RLS algorithm (Haykin, 1996): kðnÞ¼ k�1Pðn � 1ÞuðnÞ 1 þ k�1uHðnÞPðn � 1ÞuðnÞ ð11Þ er shaft-wheel assembly turns freely and smoothly by rotating it by hand (Crouse & gnal processor Data recorder Algorithms (RLS) (FFT) Diagnostic system ic signal analyzer sis of turbocharger wheel-blade faults. J.-D. Wu et al. / Expert Systems with Applications 36 (2009) 5424–5431 5429 nðnÞ¼ dðnÞ� ŵ H ðn � 1ÞuðnÞ; ð12Þ ŵðnÞ¼ ŵðn � 1Þþ kðnÞn�ðnÞ; ð13Þ PðnÞ¼ k�1Pðn � 1Þ� k�1kðnÞuHðnÞPðn � 1Þ: ð14Þ In this procedure, matrix P(n) is the inverse of the auto-correla- tion matrix of input vector u, nðnÞ is the a priori estimation error, and k(n) is the gain vector. To initialize the RLS algorithm, the ini- tial conditions are generally taken to be ŵð0Þ¼ 0M�1, where M is the number of parameters and Pð0Þ¼ d�1I, where I is an M � M identity matrix and d is a small positive constant. One reason for using the RLS order-tracking technique is that the rate of conver- gence of the RLS algorithm is typically an order of magnitude faster than the traditional LMS algorithm. In order to provide valid understand of the characteristic in adaptive filtering algorithms. A comparison of convergence speeds and the mean-square-error (MSE) in various adaptive filters, i.e., Fig. 10. Order figures of sound emission signals for engine speed at 800 rpm using adaptive RLS filter. A solid line depicts blades without any damage; dashed line depicts blades with one fault. LMS, RLS, and Kalman filter in simulation is shown as Fig. 3. The re- sults have shown that the Kalman filter has the quickest conver- gence speed, converging at the iteration number of 800, the RLS converges at 1500, and LMS converges at 2800. That is because the Kalman filter algorithm takes into account the noise factor and is well structured with sophisticated considerations. However, the Kalman filter may be exploited as the basis for deriving an adaptive filtering algorithm appropriate to the complex calculation situations. In particular, each updated estimate of the state is com- puted from the previous estimate and the new input data, so the previous estimate requires storage. Comparatively, the RLS filter algorithm is rather simple in filtering design. 3. Experimental verification of fault diagnostic systems In the experimental investigation, two experiments are imple- mented to evaluate the proposed RLS filtering algorithm. One is a gear-set defect diagnosis using both the vibration signal and the sound emission signal; the other is a diagnosis of damaged IC en- gine turbocharger wheel-blades by using a sound emission signal. 3.1. Application 1: gear-set defect diagnosis The experimental setup for the gear-set defect diagnostic sys- tem is shown in Fig. 4. The horsepower of the DC servo motor is 0.5 with a maximum revolution of 3000 rpm. The motor can be controlled by using a DSP controller. An optical fiber sensor (LM339) is used to detect motor revolution and angular displace- ment as reference signals in the diagnostic system. The vibration signal and sound emission are measured by using an accelerometer (PCB 353B15) and a condenser microphone (ACO P4012). The pro- posed diagnostic system is implemented on a 60 MHz floating- point TMS320C32 DSP equipped with two 16-bit analog I/O chan- nels by using the adaptive RLS algorithm. In applying the proposed high-resolution order-tracking methods, some parameters need to be determined, such as the number of tracking orders N ^order and forgetting factor k in the proposed RLS algorithm. In addition, the experimental implementation of the gear-set is at various speed conditions. The experimental conditions are indi- cated in Fig. 5, where the gear-set is operated as a running-up schedule. The experimental results from order figures using a vibration signal are shown in Fig. 6; the order figures using a sound Fig. 11. Sound pressure amplitude in test schedule for diagnosis of faults in turbocharger blades. A solid line depicts blades without any damage; a dashed line depicts blades with one fault. 5430 J.-D. Wu et al. / Expert Systems with Applications 36 (2009) 5424–5431 emission signal are shown in Fig. 7. The experimental results dem- onstrate that the proposed diagnostic system is effective in defect diagnosis by using both vibration and sound emission signals. The ordered figures can be saved as a data bank for practical fault diag- nosis. Furthermore, order-tracking is one of the important tools for feature extraction of rotating machinery. The order amplitude fig- ure gives the information of the harmonic order signal in the mechanical system. Ordinarily, the amplitude of fault conditions is higher than without fault condition. So it is very easy to distin- guish the fault and without fault conditions. 3.2. Application 2: diagnosis of damaged IC engine turbocharger wheel-blades An IC engine can produce more power at the same speed if a forced induction system is used to improve volumetric efficiency. Such a system consists of air pumps or blowers that force more Fig. 12. Order figures of sound emission signals for engine run-up test. A solid line depicts blades without any damage; a dashed line depicts blades with one fault. air-fuel mixture into the engine combustion chamber. Normally they may produce 35–60% more power than a naturally-aspirated engine (Crouse & Anglin, 1993). However, the turbocharger system requires periodic maintenance to prevent early failure. Frequent causes of turbocharger failure are sand and other particles striking the blades, as in the case of the turbo blades shown in Fig. 8a. Con- ventional diagnosis of damaged blades is to conduct a visual inspection when the engine is cool or check to make sure that the turbocharger shaft-wheel assembly turns freely and smoothly by rotating it by hand, as shown in Fig. 8b. Obviously, the conven- tional inspection is not a precision approach for diagnosis of dam- age; it also is not a suitable method for diagnosis when the engine is running. The conventional FFT methods with a fixed sampling frequency also are ineffective for this application because normal operation of the engine varies rapidly. In fault diagnostic techniques to date, the vibration signal has become the most widely used method when a vibration signal is available. Unfortunately, in some applications of fault diagnostic systems, a vibration reference signal is unavailable. In this applica- tion, only the sound emission signal is used to evaluate the pro- posed system in the diagnosis of a damaged turbocharger under fixed revolution, acceleration and deceleration conditions. The experimental arrangement for the diagnosis of damaged turbo- charger wheel-blades is depicted in Fig. 9. A four-cylinder, four- stroke, 2.8-l IC engine with a turbocharger system is used in this application. A fiber-optic sensor is utilized to detect the revolution signal that is related to the sound emission from the wheel-blades. In this experimental implementation, the related reference signal from the engine can be measured by the ignition system or the wheel-blade signal. However, the ignition system may have sub- stantial interference that will affect the performance; therefore, the reference signal is picked up near the wheel-blades by using a fiber-optic sensor. To verify the filtering algorithm in order-track- ing, a preliminary test was conducted in an engine with a fixed rev- olution of 800 rpm. The order figures using a sound emission signal are shown in Fig. 10. In a practical condition, an engine may be operated by running-up or casting down. Although the high sweep rates make accurate order measurement difficult, the proposed adaptive order-tracking is suitable for such a condition. In order to verify the adaptive filter, the test schedule for the diagnosis of damaged turbocharger blades is shown in Fig. 11. The ordered fig- ures using a sound emission signal are shown in Fig. 12. The exper- imental results demonstrate that the proposed diagnostic system is effective in fault diagnosis by using sound emission signals. The or- dered figures and data also can be saved as a data bank for practical fault diagnosis. 4. Conclusions An order-tracking technique exploiting adaptive filtering based on an RLS algorithm for tracking the orders of vibration and sound emission signals in the diagnosis of defects in a gear-set and in damaged engine turbocharger wheel-blades has been applied. In this method, the order-tracking problem was treated as parameter identification and calculated at a high-resolution. Although, the Kalman filter is the alternative method when the uncertainty fac- tors of the entire system are taken into consideration. However, in some cases the design is more complex than the RLS algorithm. In the present study, the contribution is emphasized in the practi- cal application of gear-set defect diagnosis and diagnosis of dam- aged IC engine turbocharger wheel-blades by using the proposed RLS filtering algorithm. The results of the experiments indicated that the RLS algorithm is effective in fault diagnosis for both exper- imental cases. Various adaptive filtering algorithms are expected to be used in different applications; future research should focus on J.-D. Wu et al. / Expert Systems with Applications 36 (2009) 5424–5431 5431 the development of a robust adaptive filtering algorithm to accom- modate perturbation as well as uncertainties in the diagnostic system. Acknowledgements This study was supported by the National Science Council of Taiwan, the Republic of China, under project number NSC-93- 2212-E-018-004. The authors also wish to express appreciation to Dr. Cheryl Rutledge for her editorial assistance. References Bai, M. R., Jeng, J., & Chen, C. (2002). Adaptive order tracking technique using recursive least-square algorithm. Transactions of the ASME, Journal of Vibrations and Acoustics, 124, 502–511. Biswas, M., Pandey, A. K., Bluni, S. A., & Samman, M. M. (1994). Modified chain-code computer vision techniques for interrogation of vibration signatures for structural fault detection. Journal of Sound and Vibration, 175, 89–104. Chen, Y. D., Du, R., & Qu, L. S. (1995). Fault features of large rotating machinery and diagnosis using sensor fusion. Journal of Sound and Vibration, 188, 227–242. Crouse, W. H., & Anglin, D. L. (1993). Automotive mechanics. McGraw-Hill. Denbigh, P. (1998). System analysis and signal processing. Addison Wesley. Gelle, G., Colas, M., & Serviere, C. (2001). Blind source separation: a tool for rotating machine monitoring by vibration analysis. Journal of Sound and Vibration, 248, 865–885. Haykin, S. (1996). Adaptive filter theory. Prentice-Hall. Lee, S. K., & White, P. R. (1998). The enhancement of impulsive noise and vibration signals for fault detection in rotating and reciprocating machinery. Journal of Sound and Vibration, 217, 485–505. Lin, J., & Qu, L. (2000). Feature extraction based on morlet wavelet and its application for mechanical fault diagnosis. Journal of Sound and Vibration, 234, 135–148. Mba, D. (2002). Applicability of acoustic emissions to monitoring the mechanical integrity of bolted structures in low speed rotating machinery: Case study. NDT and E International, 35(5), 293–300. Oppenheim, A. V., & Schafer, R. W. (1999). Discrete-time signal processing. Prentice-Hall. Shibata, K., Takahashi, A., & Shirai, T. (2000). Fault diagnosis of rotating machinery through visualization of sound signals. Mechanical Systems and Signal Processing, 14, 229–241. Toutountzakis, T., & Mba, D. (2003). Observations of acoustic emission activity during gear defect diagnosis. NDT and E International, 36, 471–477. Vold, H., & Leuridan, J. (1993). High resolution order tracking at extreme slew rates, using Kalman filters. SAE Paper 931288 (pp. 219–226).. An expert system for the diagnosis of faults in rotating machinery using adaptive order-tracking algorithm Introduction Principle of adaptive order-tracking technique using RLS algorithm Experimental verification of fault diagnostic systems Application 1: gear-set defect diagnosis Application 2: diagnosis of damaged IC engine turbocharger wheel bladeswheel-blades Conclusions Acknowledgements References 
work_y5gundfmajhthiisdlveuwlnye	----	SAS for MCDM Multi-criteria logistics distribution network design using SAS/OR William Ho* and Ali Emrouznejad Operations and Information Management Group Aston Business School, Aston University Birmingham B4 7ET, United Kingdom E-mail: w.ho@aston.ac.uk; Tel: +44 (0)121 2043342 Abstract This paper explores the use of the optimization procedures in SAS/OR software with application to the contemporary logistics distribution network design using an integrated multiple criteria decision making approach. Unlike the traditional optimization techniques, the proposed approach, combining analytic hierarchy process (AHP) and goal programming (GP), considers both quantitative and qualitative factors. In the integrated approach, AHP is used to determine the relative importance weightings or priorities of alternative warehouses with respect to both deliverer oriented and customer oriented criteria. Then, a GP model incorporating the constraints of system, resource, and AHP priority is formulated to select the best set of warehouses without exceeding the limited available resources. To facilitate the use of integrated multiple criteria decision making approach by SAS users, an ORMCDM code was implemented in the SAS programming language. The SAS macro developed in this paper selects the chosen variables from a SAS data file and constructs sets of linear programming models based on the selected GP model. An example is given to illustrate how one could use the code to design the logistics distribution network. Keywords: SAS; SAS/OR; Multiple criteria decision making; Analytic hierarchy process; Goal programming; Logistics distribution network * Corresponding author. 1. Introduction The logistics distribution problem is to allocate a number of points of consumption to a number of points of supply, including suppliers, manufacturers, warehouses, distribution centers, and customers. The connection of these various logistics stakeholders by a mean of transportation facilities is regarded as the logistics distribution network. Logistics distribution network design is one of the major decision problems arising in contemporary supply chain management. There are mainly two inadequacies in the traditional approaches for the problem. First, a single criterion was focused only. The objective was either to minimize the total logistics cost (Su, 1998; Wasner and Zäpfel, 2004; Hwang, 2005) or total delivery time (Su, 1999). Second, only quantifiable data were considered in the optimization techniques. Some qualitative factors, which are mainly customer oriented, were not considered. Multiple criteria decision making (MCDM) techniques have been used in recent years. One of the most prevalent techniques is analytic hierarchy process (AHP) (Ho, 2008). Some researchers (Korpela and Lehmusvaara, 1999; Korpela et al., 2001a–b; 2002) applied the combined AHP-mixed integer linear programming (MILP) model approach for the network design, whereas another group of researchers (Chan and Chung, 2004a–b; 2005; Chan et al., 2004; 2005; 2006) applied the combined AHP-genetic algorithm (GA) approach to solve the problem. For the combined AHP-MILP approach, the selection of distribution network was simply based on the customer satisfaction priorities instead of minimizing the total logistics cost or maximizing the total profit. Therefore, it is believed that the selected distribution network may not be cost effective. For the combined AHP-GA approach, the evaluation criteria used in AHP are all quantitative such as total cost, total delivery day, effectiveness of capacity utilization for warehouses, and so on. Some qualitative factors such as flexibility of capacity and value-added services were neglected. These factors are crucial in the integrated logistics system because they affect the customer satisfaction directly. To overcome the drawbacks, this paper develops a systematic and prominent MCDM technique, combining AHP and goal programming (GP), to design an optimal logistics distribution network. The combined AHP-GP approach considers both quantitative and qualitative factors and also aims at maximizing the benefits of deliverer and customers. SAS is recognized as one of the lead packages for statistical analysis and as a powerful tool for data base systems in many organizations, both in public and private sectors. SAS users come from every major industry (banking to pharmaceuticals, manufacturing to telecommunications, and so on). All with the same basic needs to make better strategic decisions and to gain a competitive edge (Emrouznejad, 2005). There are many applications in SAS that the users recognized as powerful tools in organizational management. For example, the SAS/OR system has numerous optimization procedures which handle the standard problems such as linear programming (LP) and nonlinear programming (NLP) with all types of constraints. In addition to the standard procedures available in SAS system applications such as neural network, simulation, and control project management are introduced. This paper aims to introduce a new application in SAS system for yielding an optimal solution to a multi-criteria logistics distribution problem by solving the combined AHP-GP model. 2 2. MCDM techniques MCDM techniques are generally divided into two categories: multiple attribute decision making (MADM) and multiple objective decision making (MODM). MADM techniques aim at selecting the best solution from a population of feasible alternatives which characterized by multiple attributes. One of the commonly used MADM techniques is AHP (Saaty, 1980). MODM techniques are a special extension of linear programming. A model is defined as a linear programming when the single objective function and the constraints involve linear expressions, and the decision variables are continuous. But, in MODM techniques, multiple objective functions are incorporated into the model simultaneously. GP (Charnes and Cooper, 1961) is an example of the MODM techniques. In this paper, we apply AHP to evaluate the relative importance weightings of alternative warehouses with respect to some qualitative-based evaluating criteria. Then, a GP model is formulated to select the optimal set of warehouses while considering the AHP priorities of warehouses and the quantitative-based limitations of resources. 2.1. Prioritization of warehouses using AHP In this paper, five criteria are proposed to evaluate the performance of warehouses. They include total lead time, reliability of order fulfillment, quality, flexibility of capacity, and value-added services. Total lead time comprises the time of handling inventory in warehouses, the time of storing/loading inventory in warehouses, and the time of delivering products from warehouses to customers. Reliability of order fulfillment consists of the accuracy of quantity fulfillment, the accuracy of due date fulfillment, and reliability of delivery time. Quality involves the commitment of deliverer to provide high-quality products and the condition of products received by customers. Flexibility of capacity refers to the ability of warehouses to respond to fluctuation in volume of customer orders. Value-added services refer to any activities that facilitate customers (e.g., track-and-trace and 24-hour customer hotline) and the responsiveness of warehouses to customer special requests (e.g., secure packaging and urgent delivery). The first step of AHP for evaluating the performance of warehouses is to develop a hierarchy of the decision problem as shown in Figure 1. After constructing the hierarchy, two criteria are compared at a time with respect to the goal. Once the pairwise comparisons have been made for the five criteria, each alternative warehouse is compared against each other alternative with respect to the corresponding criterion at a time. After completion of all pairwise comparisons, the relative priority of each criterion (from Table 1) and each alternative (from Table 2) are synthesized. The judgments are acceptable because the consistency ratios are all below the maximum 0.10 level. The overall priority ranking of warehouses is: wp1 = 0.446, wp2 = 0.180, wp3 = 0.239, and wp4 = 0.135. The AHP priorities are used to determine the priority level of the AHP priority constraints in the GP model. 2.2. Logistics distribution network design in GP Consider a typical logistics distribution network which consists of m warehouses denoted as i = {1, 2, …, m} and n customers denoted as j = {1, 2, …, n}. Each warehouse has a maximum throughput, Qi, a minimum throughput, qi, a fixed cost, fci, and a unit inventory holding cost, hci. Each customer has a unique order volume, Dj. When warehouse i is assigned to serve customer j, it costs dcij dollars per unit for delivery. If the total amount of products assigned 3 to warehouse i (i.e., ∑ ∀j ij ix , ) is less than qi, this is regarded as impractical allocation because it is not cost-effective to set up a warehouse for processing only a few orders. To avoid low effectiveness of warehouse utilization, penalty cost, pci, is considered in the model, which is incurred if ∑ << j iij qx0 . The problem here is to determine an optimal distribution network, which refers to the allocation of orders to the best warehouses. In the model, there are four types of decision variables: xij = amount of products delivered from warehouse i to customer j    = otherwise0 than less is warehouse toproducts of allocation totalif1 i i qi u    = otherwise0 selected is warehouseif1 i vi    = otherwise0 one toequal and both if1 ii i vu w In the model, there are three types of constraints: system constraints, resource constraints, and AHP priority constraints. System constraints are ordinary linear programming constraints, in which there is no deviation variable. This type of constraints cannot be violated, and thus they are called hard constraints. Resource constraints are goal equations or soft constraints, in which there are deviation variables. AHP priority constraints are akin to resource constraints. In this type of constraints, there are deviation variables of which the priority levels are dependent on the overall AHP priority ranking. Standard multi-criteria logistics distribution model System constraints mv i i ≤∑ (1) ∑ ≥+ j iiij qMux ∀i (2) ∑ ≤− j iij Mvx 0 ∀i (3) 1−=−− iii vuw ∀i (4) Resource constraints Priority 1 (P1): ∑ =+− −+ j iiiij Qddx ∀i (5) ∑ =+− −+++ i jmjmjij Dddx ∀j (6) Priority 2 (P2): ( ) ∑∑∑ =+−++ − +++ ++ i nmnmii i j ijiji TCddvfcxdchc 11 (7) 4 Priority 3 (P3): ∑ =+− − +++ ++ i nmnmii ddwpc 022 (8) AHP priority constraints Priorities 4 to 4 + m – 1 (P4 to P4 + m – 1): 122 =+− − +++ + +++ inminmi ddv ∀i (9) Objective function Minimize z = ( ) ( )∑∑ −+−+ +++ k ssk k rrk ddPddP ∀r, ∀s (10) Constraint set (1) ensures that the number of warehouses selected must be equal to or less than the number of warehouses available. Constraint set (2) determines which warehouse(s) has/have allocation of products that less than minimum warehouse throughput. Constraint set (3) determines which warehouse(s) is/are selected. Constraint set (4) determines which warehouse(s) incur(s) penalty cost. Constraint set (5) allocates products to warehouses while the amount must not exceed the maximum warehouse throughput. Constraint set (6) allocates products to warehouses while the amount must equal to that demanded by the customers. Constraint set (7) ensures that the total cost, including the inventory holding cost, delivery cost, and fixed cost associated with warehouse selection, must not exceed the targeted amount. Constraint set (8) ensures that allocation of products to warehouses incurring penalty cost is not allowed. Constraint set (9) is to select warehouse i. Objective function (10) is to minimize the total deviations from the goals. 3. ORMCDM: SAS code for multi-criteria logistics distribution network design The ORMCDM introduced in this paper provides a powerful decision making tool for yielding an optimal logistics distribution network design. To enhance the model there are several parameters. The user can select the desired parameters according to the particular model that is required. Users familiar with SAS can add their own features. Users not familiar with SAS need only to run the program with their model specification prior to running the system. The ORMCDM requires an initial data set that contains the name of variable and data file for warehouses and customers. The data is saved as SAS data set. SAS has the ability also to read from a text (Tab delimited) or Excel format. The program has the ability to accommodate unlimited number of warehouses and customers. The only limitation is the memory of computer used to run the ORMCDM. The ORMCDM software then converts data set to a GP model. Based on the data and parameters specified in the ORMCDM, the code first creates the usual linear program, then use “PROC LP” to solve the model. The results will then be transferred to report files. The ORMCDM produces a table listing values of decision variables and deviation variables. It also supplies a report summarizing the status of goal achievement at each stage (or priority level). In the rest of this paper the procedure of implementation of ORMCDM together with an example are explained. Figure 2 illustrates the data flow in the ORMCDM. 5 4. Definition of terms and typographical conventions In the rest of this paper and particularly in the ORMCDM code we will see several types of styles used. Style conventions are summarized below: • Courier font: is used to show example of SAS statements. In most cases, this paper uses lowercase type for SAS code. SAS users can enter their own SAS code in lowercase, uppercase or a mixture of the two. Enter any titles and footnotes exactly as you want them to appear on the printout. • _Underscore_: Variable name that are surround by “_” are specifically used as parameters to the SAS. In all case these variables must be used without any change. • _Underscore: Variable name that are started with “_” are specifically used as parameters to the ORMCDM. The ORMCDM runs four macros for data handling (%data), model building (%model1 and %model2), and report writing (%report). 5. Illustration of ORMCDM This section presents a simple example of four warehouses and seven customers for illustration of ORMCDM. The resource data includes: • QQi: maximum throughput of warehouse i. • qi: minimum throughput of warehouse i. • hci: unit inventory holding cost associated with warehouse i. • fci: fixed cost associated with warehouse i. • pci: penalty cost associated with warehouse i. • cost: unit delivery cost from warehouse i to customer j. • demand: unique order volume requested by the customers. 6. Data handling (%data) This part of ORMCDM prepares the data to a suitable format that can be used in PROC LP. The ORMCDM requires one data set containing name of “warehouses”, name of “customers”, “cost matrix”, and “demand”. The data set should be a “.txt” file, which is saved as “text tab delimited”. The warehouses’ and customers’ names must start with a letter and may contain up to 50 characters. The variable names must be listed in the first row of the data file. The warehouses’ names should be listed in the first column. Note that the demand should be given in the last row of the data set. Other information, including QQi, qi, hci, fci, and pci should also be given in the data set, but the order of the columns is not important. An example of data file is: File: data.txt Wareh Cust1 Cust2 Cust3 Cust4 Cust5 Cust6 Cust7 QQi qi hci fci pci w1 1 1 2 4 4 3 6 30000 6000 5 30000 7500 w2 2 6 9 3 7 8 4 26000 5200 3 25000 6250 w3 8 4 3 6 3 2 4 22000 4400 3 20000 5000 w4 8 8 9 3 5 7 2 18000 3600 2 15000 3750 demand 12000 9000 10000 8000 6000 11000 7000 6 There is only one parameter prior to calling the data macro: • _data: indicates the name and location of the data file. SAS procedure for data handling is presented in Appendix A.1. 7. Model building (%model) This part of ORMCDM calls PROC LP to solve the model. There are two SAS macros for model building. Macro %model1 creates a suitable data set that can be used for PROC LP. There are four parameters prior to calling the model macro: • _nWareh: defines the number of warehouses. • _nCustomer: defines the number of customers. • _TotalCost: defines the total cost. • _myupperbnd: defines the upper bound value for integer variables. Macro %model2 defines seven data sets that could be used to solve the multi-criteria logistics distribution problem in seven stages. At each stage, a data set is created, then PROC LP solves the model and results are saved in three data sets for future analysis. These three output data sets are for primal solution, dual solution, and tabulated solution of each stage. There are two parameters before calling this macro. • _myLarge: defines an optional large number. • _nIterations: defines an optional value for number of integer iterations. The parameter _myLarge is an optional parameter that is needed for LP, and in this case, we set it to 100000. SAS procedure for models %model1 and %model2 building are presented in Appendices A.2 and A.3, respectively. 8. Report writing (%report) The outputs from ORMCDM include one report for each stage. This report contains all information regarding the summary and solution of each stage of the problem, in this case, seven stages. All these information are also saved in SAS data sets. The users can define appropriate names for each of these data sets before calling %ormcdm macro as follows. • _outprimal: identifies the name of the SAS output file for primal solution. • _outdual: identifies the name of the SAS output file for dual solution. • _outtable: identifies the name of the SAS output file for tabulated solution. Another parameter needs to be set before calling this macro is the _title as follows: • _title: gives a title in the output of the SAS. SAS procedure for report writing is presented in Appendix A.4. 9. ORMCDM macro (%ormcdm) To make the system as ease as possible, the “%ormcdm macro” put all the above code together. 7 * A SAS macro for the multi-criteria logistics distribution network design; %macro ormcdm; %data; %model1; %do i=1 %to &_nWareh+3; * Number of stages in GP model; %model2(&i); %report(&i); %end; %mend ormcdm; In the above code, the “%ORMCDM macro” is used to manage all previously explained codes, including data handling, model building, and report writing. To get the result, user needs to set up the parameters and run only one statement: %ORMCDM; 10. Instruction of how to use ORMCDM macro This section presents SAS code for the earlier example of the multi-criteria logistics distribution problem with four warehouses and seven customers. The data is saved in file “data.txt”. A user needs to set the parameters as required and run the following code. options nodate; %let _title='An example of MCDM'; %let _dataMCDM='c:/sasor/data.txt'; %let _data='c:\sasor\MCDM.txt'; %let _nWareh=4; %let _nCustomer=7; %let _TotalCost=425000; %let _outprimal=outprimal; %let _outdual=outdual; %let _outtable=outtable; %let _myLarge=100000; %let _nIterations=100000; %let _myupperbnd=12000; %ormcdm; The above code manages to get the result based on the specified parameters and the cost matrix saved in the text file. The above code also produces a macro variable (_ORLP_) at termination. Users can examine the result of this macro variable, examine whether PROC LP ran correctly, and examine what error or difficulty it encountered. Summary of information, including the objective value at optimum level and status of _ORLP_ can be seen in the log file as shown in Figure 3. 8 11. Sample results from ORMCDM macro: output from SAS In the GP model, there are 40 integral decision variables, 34 deviation variables, 30 constraints, and seven goals (or stages). Some of the results of running the above code are presented in Figure 4 and Figure 5, which summarize the solutions in stage 5 and stage 6, respectively. The solution, shown in Figure 4, is feasible because the allocation does not exceed the maximum throughput of warehouses, does satisfy the volume requirement of customers, does not exceed the total cost budget, and does not incur any penalty cost. When stage 6 was found to be unachievable (i.e., dN16 or −16d = 1), shown in Figure 5, the optimization process was terminated. So, the solution, satisfying the first five priority levels (i.e., P1 to P5), is an optimal solution of the problem, shown in Figure 4. The values of decision variables vi show that three warehouses were selected, including warehouse 1 (v1 = 1), warehouse 3 (v3 = 1), and warehouse 4 (v4 = 1). The total cost spent in setting up these three warehouses, holding inventory in the warehouses, and delivering products from the warehouses to their assigned customers is $425000 with no slack. Besides, the total penalty cost incurred is zero. 12. Conclusions Nowadays, there is a plenty of software that can be used for operations research. The major reason why we selected SAS is because it has various optimization tools that can be used in a wide range of problems in operations research. In particular, optimization procedures in SAS/OR are exposed to the user in a variety of places such as “PROC LP” and “PROC NLP”. ORMCDM as introduced in this paper is a new application in SAS/OR that is a powerful decision making tool for designing an optimal logistics distribution network. The ORMCDM application implemented in this paper has no limitation on the number of nodes (warehouses and customers). The only limitation is the memory and disk space of the computer used. Besides, SAS has strong data management capabilities that can handle very large datasets efficiently, and it can work with multiple datasets simultaneously. In cases where SAS users want to solve another problem or change the data, they can simply revise a single data set file and run a single statement, %ormcdm. The ORMCDM report’s files can also directly feed to other SAS routines for further analysis. 9 References Chan FTS, Chung SH. Multi-criteria genetic optimization for distribution network problems. International Journal of Advanced Manufacturing Technology 2004a; 24; 517–532. Chan FTS, Chung SH. A multi-criterion genetic algorithm for order distribution in demand driven supply chain. International Journal of Computer Integrated Manufacturing 2004b; 17; 339–351. Chan FTS, Chung SH, Wadhwa S. A heuristic methodology for order distribution in a demand driven collaborative supply chain. International Journal of Production Research 2004; 42; 1–19. Chan FTS, Chung SH, Wadhwa S. A hybrid genetic algorithm for production and distribution. Omega 2005; 33; 345–355. Chan FTS, Chung SH. Multicriterion genetic optimization for due date assigned distribution network problems. Decision Support Systems 2005; 39; 661–675. Chan FTS, Chung SH, Choy KL. Optimization of order fulfillment in distribution network problems. Journal of Intelligent Manufacturing 2006; 17; 307–319. Charnes, A. and Cooper, W. W. (1961), Management Models and Industrial Applications of Linear Programming, John Wiley & Sons, New York. Emrouznejad A. Measurement efficiency and productivity in SAS/OR. Computers & Operations Research 2005; 32, 1665–1683. Ho W. Integrated analytic hierarchy process and its applications – a literature review. European Journal of Operational Research 2008; 186, 211–228. Hwang HS. An integrated distribution routing model in multi-supply center system. International Journal of Production Economics 2005; 98; 136–142. Korpela J, Lehmusvaara A. A customer oriented approach to warehouse network evaluation and design. International Journal of Production Economics 1999; 59; 135–146. Korpela J, Lehmusvaara A, Tuominen M. Customer service based design of the supply chain. International Journal of Production Economics 2001a; 69; 193–204. Korpela J, Kyläheiko K, Lehmusvaara A, Tuominen M. The effect of ecological factors on distribution network evaluation. International Journal of Logistics: Research and Applications 2001b; 4; 257–269. Korpela J, Kyläheiko K, Lehmusvaara A, Tuominen M. An analytic approach to production capacity allocation and supply chain design. International Journal of Production Economics 2002; 78; 187–195. Saaty TL. The Analytic Hierarchy Process. McGraw-Hill: New York; 1980. Su CT. Locations and vehicle routing designs of physical distribution systems. Production Planning & Control 1998; 9; 650–659. Su CT. Dynamic vehicle control and scheduling of a multi-depot physical distribution system. Integrated Manufacturing Systems 1999; 10; 56–65. Wasner M, Zäpfel G. An integrated multi-depot hub-location vehicle routing model for network planning of parcel service. International Journal of Production Economics 2004; 90; 403–419. 10 Appendix A A.1. SAS procedure for data handling * The data handling macro; %macro data; * Import text tab delimited data file to SAS data file; proc import datafile=&_dataMCDM out=dataC dbms=tab replace; getnames=yes; run; %mend data; 11 A.2. SAS procedure for model building (%model1) * The model building macro; %macro model1; data _null_; FILE &_data; array Customer(&_nCustomer) Cust1-Cust&_nCustomer; array NCustom(&_nCustomer) $; array NWareh(&_nWareh) $; array QQ(&_nWareh); array q(&_nWareh); array hc(&_nWareh); array fc(&_nWareh); array pc(&_nWareh); array Demand(&_nCustomer); array cost(&_nWareh, &_nCustomer); array u(&_nCustomer); ds='xxxxx'; myst='xxxxxxxxx'; do i1=1 to &_nWareh; Link ReadDataC; NWareh(i1)=Wareh; do j1=1 to &_nCustomer; cost(i1,j1)=Customer[j1]; end; QQ[i1]=QQi; q[i1]=qi; hc[i1]=hci; fc[i1]=fci; pc[i1]=pci; end; Link ReadDataC; do j1=1 to &_nCustomer; Demand(j1)=Customer[j1]; end; myst='_row_ '; put myst @@; myst0='X'; do i=1 to &_nWareh; do j=1 to &_nCustomer; ds=myst0||put(i,1.)||put(j,1.); put ds @@; end; end; do i=1 to &_nWareh; ds='w'||put(i,1.); put ds @@; ds='u'||put(i,1.); put ds @@; ds='v'||put(i,1.); put ds @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i<10 then ds='dN'||put(i,1.); else ds='dN1'||put(i-10,1.);put ds @@; if i<10 then ds='dP'||put(i,1.); else ds='dP1'||put(i-10,1.);put ds @@; end; put '_type_ _rhs_'; 12 myst='binary'; put myst @@; myInt=0; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '. ' @@; end; end; do i=1 to &_nWareh; myInt=myInt+1; put myInt @@; myInt=myInt+1; put myInt @@; myInt=myInt+1; put myInt @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i<=&_nWareh+&_nCustomer then mycoefP=1; else mycoefP=0; if &_nWareh<i<=&_nWareh+&_nCustomer then mycoefN=1; else mycoefN=0; put '. ' @@; put '. ' @@; end; put 'binary .'; myst='integer'; put myst @@; myInt=0; do i=1 to &_nWareh; do j=1 to &_nCustomer; myInt=myInt+1; put myInt @@; end; end; do i=1 to &_nWareh; put '. ' @@; put '. ' @@; put '. ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i<=&_nWareh+&_nCustomer then mycoefP=1; else mycoefP=0; if &_nWareh<i<=&_nWareh+&_nCustomer then mycoefN=1; else mycoefN=0; put '. ' @@; put '. ' @@; end; put 'integer .'; myst='upperbound'; put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; myUpp=&_myupperbnd; put myUpp @@; end; end; do i=1 to &_nWareh; put '. ' @@; put '. ' @@; put '. ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i<=&_nWareh+&_nCustomer then mycoefP=1; else mycoefP=0; 13 if &_nWareh<i<=&_nWareh+&_nCustomer then mycoefN=1; else mycoefN=0; put '. ' @@; put '. ' @@; end; put 'upperbd .'; myst='objP1'; put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i<=&_nWareh+&_nCustomer then mycoefP=1; else mycoefP=0; if &_nWareh<i<=&_nWareh+&_nCustomer then mycoefN=1; else mycoefN=0; put mycoefN @@; put mycoefP @@; end; put 'min .'; myst='objP2'; put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=&_nWareh+&_nCustomer+1 then mycoef=1; else mycoef=0; put '0 ' @@; put mycoef @@; end; put 'min .'; myst='objP3'; put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=&_nWareh+&_nCustomer+2 then mycoef=1; else mycoef=0; put '0 ' @@; put mycoef @@; 14 end; put 'min .'; do k=1 to &_nWareh; myst='objP'||put(&_nWareh+k-1,1.); put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=&_nWareh+&_nCustomer+2+k then mycoef=1; else mycoef=0; put mycoef @@; put mycoef @@; end; put 'min .'; end; do k=1 to 2*&_nWareh+&_nCustomer+2; if k<=&_nWareh+&_nCustomer then myst1='P1'; else if k<=&_nWareh+&_nCustomer+1 then myst1='P2'; else if k<=&_nWareh+&_nCustomer+2 then myst1='P3'; else myst1='P'||put(k-(&_nCustomer+3),1.); if (myst1='P1' and 1<=k<=4) or (myst1='P2') or (myst1='P3') then; else do; if k<10 then myst='dZrN'||myst1||put(k,1.); else myst='dZrN'||myst1||put(k,2.); put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=k then mycoef=1; else mycoef=0; put mycoef @@; put '0 ' @@; end; put 'eq 0'; end; end; do k=1 to 2*&_nWareh+&_nCustomer+2; if k<=&_nWareh+&_nCustomer then myst1='P1'; else if k<=&_nWareh+&_nCustomer+1 then myst1='P2'; else if k<=&_nWareh+&_nCustomer+2 then myst1='P3'; else myst1='P'||put(k-(&_nCustomer+3),1.); if k<10 then myst='dZrP'||myst1||put(k,1.); else myst='dZrP'||myst1||put(k,2.); put myst @@; do i=1 to &_nWareh; 15 do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=k then mycoef=1; else mycoef=0; put '0 ' @@; put mycoef @@; end; put 'eq 0'; end; myst='v '; put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put '1 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; put '0 ' @@; put '0 ' @@; end; put 'le 4' ; do k=1 to &_nWareh; myst='wuv'||put(k,1.); put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; if i=k then mycoefP=1; else mycoefP=0; mycoefN=-mycoefP; put mycoefP @@; put mycoefN @@; put mycoefN @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; put '0 ' @@; put '0 ' @@; end; put 'eq -1' ; end; do k=1 to &_nWareh; myst='XU'||put(k,1.); put myst @@; do i=1 to &_nWareh; if i=k then mycoef=1; else mycoef=0; 16 do j=1 to &_nCustomer; put mycoef @@; end; end; do i=1 to &_nWareh; if i=k then mycoef=&_myLarge; else mycoef=0; put '0 ' @@; put mycoef @@; put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; put '0 ' @@; put '0 ' @@; end; put 'ge ' @@; put q[k]; end; do k=1 to &_nWareh; myst='XV'||put(k,1.); put myst @@; do i=1 to &_nWareh; if i=k then mycoef=1; else mycoef=0; do j=1 to &_nCustomer; put mycoef @@; end; end; do i=1 to &_nWareh; if i=k then mycoef=-&_myLarge; else mycoef=0; put '0 ' @@; put '0 ' @@; put mycoef @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; put '0 ' @@; put '0 ' @@; end; put 'le ' @@; put '0'; end; do k=1 to &_nWareh; myst='XP1d'||put(k,1.); put myst @@; do i=1 to &_nWareh; if i=k then mycoef=1; else mycoef=0; do j=1 to &_nCustomer; put mycoef @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=k then mycoef=1; else mycoef=0; mycoefN=-mycoef; put mycoef @@; put mycoefN @@; end; 17 put 'eq ' @@; put QQ[k]; end; do k=&_nWareh+1 to &_nCustomer+&_nWareh; if k<10 then myst='P1d'||put(k,1.); else myst='P1d'||put(k,2.); put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; if j+&_nWareh=k then mycoef=1; else mycoef=0; put mycoef @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=k then mycoef=1; else mycoef=0; mycoefN=-mycoef; put mycoef @@; put mycoefN @@; end; put 'eq ' @@; put demand[k-&_nWareh]; end; myst='P2'; put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; mycoef=cost[i,j]+hc[i]; put mycoef @@; end; end; do i=1 to &_nWareh; put '0 ' @@; put '0 ' @@; put fc[i] @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=&_nWareh+&_nCustomer+1 then mycoef=1; else mycoef=0; mycoefN=-mycoef; put mycoef @@; put mycoefN @@; end; put 'eq ' @@; mycost=&_TotalCost; put mycost ; myst='P3 '; put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; put pc[i] @@; put '0 ' @@; 18 put '0 ' @@; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=&_nWareh+&_nCustomer+2 then mycoef=1; else mycoef=0; mycoefN=-mycoef; put mycoef @@; put mycoefN @@; end; put 'eq 0'; do k=1 to &_nWareh; myst='P'||put(&_nWareh+k-1,1.); put myst @@; do i=1 to &_nWareh; do j=1 to &_nCustomer; put '0 ' @@; end; end; do i=1 to &_nWareh; if k =1 or k=4 then do; if i=k then mycoef=1; else mycoef=0; put '0 ' @@; put '0 ' @@; put mycoef @@; end; if k =2 then do; if i=k+1 then mycoef=1; else mycoef=0; put '0 ' @@; put '0 ' @@; put mycoef @@; end; if k =3 then do; if i=k-1 then mycoef=1; else mycoef=0; put '0 ' @@; put '0 ' @@; put mycoef @@; end; end; do i=1 to 2*&_nWareh+&_nCustomer+2; if i=&_nWareh+&_nCustomer+2+k then mycoef=1; else mycoef=0; mycoefN=-mycoef; put mycoef @@; put mycoefN @@; end; put 'eq ' @@; put '1'; end; ReadDataC: set dataC; return; run; proc import datafile=&_data out=dataMCDM dbms=dlm replace; getnames=yes; run; %mend model1; 19 A.3. SAS procedure for model building (%model2) %macro model2(p); data dataMCDM&p; set dataMCDM; myP='objP'||put(&p,1.); if _type_='min' and _row_ ne myP then delete; if substr(_row_,1,5)='dZrNP' and substr(_row_,1,6)>='dZrNP'||put(&p,1.) then delete; if substr(_row_,1,5)='dZrPP' and substr(_row_,1,6)>='dZrPP'||put(&p,1.) then delete; run; proc lp data=dataMCDM&p dualout=&_outdual&p primalout=&_outprimal&p tableauout=&_outtable&p imaxit=&_nIterations; run; %put &_orlp_; %mend model2; 20 A.4. SAS procedure for report writing * The report writing macro; %macro report(p); title &_title '(P=' &p ')'; proc print data=&_outprimal&p; proc print data=&_outdual&p; run; proc print data=&_outtable&p; run; %mend report; 21 Figure 1: A hierarchy of the warehouse prioritization Criteria Alternatives/ Warehouses W1 W3 W2 W4 Goal Quality Total lead time Reliability of order fulfillment Flexibility of capacity Value-added services Select the best warehouse 22 Figure 2: Data flow in ORMCDM 23 Figure 3: Log for %ORMCDM 24 Figure 4: Result of %ORMCDM, primal output of stage 5 25 Figure 5: Result of %ORMCDM, primal output of stage 6 26 Table 1: Priorities of criteria with respect to goal C1 C2 C3 C4 C5 Priorities λmax CI RI CR C1 1 1 2 3 4 0.315 C2 1 1 2 3 4 0.315 C3 1/2 1/2 1 3 4 0.211 C4 1/3 1/3 1/3 1 1 0.086 C5 1/4 1/4 1/4 1 1 0.073 Total 1.000 5.088 0.022 1.120 0.020 C1: Total lead time; C2: Reliability of order fulfillment; C3: Quality; C4: Flexibility of capacity; C5: Value-added services. 27 Table 2: Priorities of alternatives with respect to criteria W1 W2 W3 W4 Priorities λmax CI RI CR (w.r.t. C1: Total lead time) W1 1 3 3 5 0.518 W2 1/3 1 1 4 0.214 W3 1/3 1 1 3 0.196 W4 1/5 1/4 1/3 1 0.072 Total 1.000 4.076 0.025 0.900 0.028 (w.r.t. C2: Reliability of order fulfillment) W1 1 3 2 4 0.467 W2 1/3 1 1/2 2 0.160 W3 1/2 2 1 3 0.277 W4 1/4 1/2 1/3 1 0.095 Total 1.000 4.031 0.010 0.900 0.011 (w.r.t. C3: Quality) W1 1 3 2 3 0.455 W2 1/3 1 1/2 1 0.141 W3 1/2 2 1 2 0.263 W4 1/3 1 1/2 1 0.141 Total 1.000 4.010 0.003 0.900 0.004 (w.r.t. C4: Flexibility of capacity) W1 1 1/3 1/2 1/5 0.086 W2 3 1 3 1/2 0.299 W3 2 1/3 1 1/3 0.140 W4 5 2 3 1 0.474 Total 1.000 4.065 0.022 0.900 0.024 (w.r.t. C5: Value-added services) W1 1 4 2 3 0.470 W2 1/4 1 1/3 1 0.114 W3 1/2 3 1 2 0.280 W4 1/3 1 1/2 1 0.136 Total 1.000 4.031 0.010 0.900 0.010 W1: Warehouse 1; W2: Warehouse 2; W3: Warehouse 3; W4: Warehouse 4. 28 Minerva Access is the Institutional Repository of The University of Melbourne Author/s: Ho, W; Emrouznejad, A Title: Multi-criteria logistics distribution network design using SAS/OR Date: 2009-04-01 Citation: Ho, W. & Emrouznejad, A. (2009). Multi-criteria logistics distribution network design using SAS/OR. EXPERT SYSTEMS WITH APPLICATIONS, 36 (3), pp.7288-7298. https://doi.org/10.1016/j.eswa.2008.09.012. Persistent Link: http://hdl.handle.net/11343/118678 File Description: Accepted version 
work_y6kn666wmrc5xidvufq5vk6rfu	----	Casebased reasoning for predicting the success of therapy Tilburg University Case-based Reasoning for Predicting the Success of Therapy Janssen, Rosanne; Spronck, P.H.M.; Arntz, Arnoud Published in: Expert Systems: The Journal of Knowledge Engineering DOI: 10.1111/exsy.12074 Publication date: 2015 Document Version Publisher's PDF, also known as Version of record Link to publication Citation for published version (APA): Janssen, R., Spronck, P. H. M., & Arntz, A. (2015). Case-based Reasoning for Predicting the Success of Therapy. Expert Systems: The Journal of Knowledge Engineering, 32(2), 165-177. https://doi.org/10.1111/exsy.12074 General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal ? Take down policy If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim. Download date: 06. apr. 2021 https://doi.org/10.1111/exsy.12074 https://research.tilburguniversity.edu/en/publications/ef62f8f4-fcb5-4429-9011-1458581d4e42 https://doi.org/10.1111/exsy.12074 Article DOI: 10.1111/exsy.12074 Case-based reasoning for predicting the success of therapy Rosanne Janssen,1 Pieter Spronck2 and Arnoud Arntz1,3 (1) Department of Clinical Psychological Science, Maastricht University, 6200 MD, Maastricht, The Netherlands Email: rosanne.janssen@maastrichtuniversity.nl (2) Tilburg center for Cognition and Communication (TiCC), Tilburg University, Tilburg, The Netherlands (3) Department of Clinical Psychology, University of Amsterdam, Amsterdam, The Netherlands Abstract: For patients with mental health problems, various treatments exist. Before a treatment is assigned to a patient, a team of clinicians must decide which of the available treatments has the best chance of succeeding. This is a difficult decision to make, as the effectiveness of a treatment might depend on various factors, such as the patient’s diagnosis, background and social environment. Which factors are the predictors for successful treatment is mostly unknown. In this article, we present a case-based reasoning approach for predicting the effect of treatments for patients with anxiety disorders. We investigated which techniques are suitable for implementing such a system to achieve a high level of accuracy. For our evaluation, we used data from a professional mental healthcare centre. Our application correctly predicted the success factor of 65% of the cases, which is significantly higher than the prediction of the baseline of 55%. Under the condition that the prediction was based on only cases with a similarity of at least 0.62, the success rate of 80% of the cases was predicted correctly. These results warrant further development of the system. Keywords: case-based reasoning, prediction, forecasting-by-analogy, weighted voting, information gain, nearest neighbour method, accuracy, mental health care, treatment outcome, anxiety disorders 1. Introduction For mental health patients with anxiety disorders, a variety of possible treatments is available. The decision, of which treatment a patient is offered, is mainly the responsibility of healthcare professionals. For the patient’s health, and for cost-effectiveness, that the treatment that is offered should be effective in dealing with the patient’s problems. However, not every treatment is equally effective for every patient. Moreover, there is no consistent empirical evidence for patient-treatment matching rules (Spinhoven et al., 2008). The effectiveness of most treatments for a specific patient cannot be predicted by a therapist. Despite such predictions being hard to make, they are the first step in successfully matching patient and treatment. In this study, we aim to use case-based reasoning (CBR) to make such predictions with a high level of accuracy. This study was performed at the Community Mental Health Centre Maastricht1 (CMHCM), the Netherlands. This centre treats patients with all sorts of mental disorders. As a starting point, the study focuses on patients with anxiety disorders and the effect of cognitive behavioural therapy. To make treatment predictions, a CBR system called CBRth was built, where CBR stands for case-based reasoning and th for therapy. The goal of this study is twofold: (1) to investigate which techniques are suitable for implementing CBRth and (2) to investigate the accuracy of CBRth. This study is the first step in a research project in which the ultimate goal is the creation of an advisory system for a team of experts to assign the most effective therapy from the therapies offered by the centre to a patient. We chose CBR, because it is ideally suited for dealing with real-world-examples. Moreover, patient information often contains missing values, and as opposed to many competing methods, CBR offers various ways to deal with those, making it a suitable approach to our problem domain. Finally, because predictions made by a CBR can be made transparent to the users, they tend to have a high level of acceptability. We discussed the proposed system with our intended users, and found that they were positive about the concept. This paper is organised as follows: Section 2 provides background information on the context of the clinical setting, and reviews CBR in mental health care and related techniques we used in this study. Section 3 specifies the CBRth casebase. Section 4 describes the CBRth process. Section 5 discusses the experimental setup. Section 6 gives the results of this study, which are discussed in Section 7. Section 8 presents our concluding remarks. 2. Background This section provides background information on the context of the clinical setting where the research was carried out (Section 2.1). It also discusses background literature on CBR (Section 2.2) and CBR techniques (Section 2.3). 1Dutch name for Community Mental Health Centre Maastricht is Riagg Maastricht. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 2.1. The context The CMHCM is mainly concerned with providing treatments for mental health patients. It is also involved with research, for which it cooperates with Maastricht University. An important research domain is the effect of treatments. Over the years, the university gathered a large dataset on the effects of treatments given at CMHCM. When new patients arrive at CMHCM, they are assigned to treatment after a screening procedure by an intake staff. This process involves the following people: (1) the patient, (2) the screener (the therapist who determines the patient’s diagnoses), (3) the intake staff (team of screeners) and (4) the therapist who provides the treatment. The procedure is as follows: 1. The patient is screened in two or three sessions to determine the patient’s diagnoses. During the screening, the screener does an anamnestic interview and formulates the targets of treatment together with the patient. 2. The screener presents the case in a meeting of the intake staff, after which the case is discussed and a decision is taken on which treatment, if any, is offered to the patient. 3. In a next session, the screener explains the treatment to the patient. 4. Before the treatment starts, there is a pre-assessment. The patient completes three questionnaires. The questionnaire scores represent the baseline state of the patient. Patients give informed consent for the use of these data for scientific research. 5. The patient receives treatment for several months (usually 4–6). 6. After the treatment, a post-assessment is carried out with the same questionnaires to determine which state the patient is in after treatment. By comparing the two assessments, the effect of the treatment can be determined. To describe the patient’s diagnoses, the screener uses the Diagnostic and Statistical Manual of mental disorders (DSM; (American Psychiatric Association, 1994)), a classification of mental disorders. The DSM consists of five axes (domains) on which disorders and other relevant issues can be assessed. Axis I contains all mental disorders except Personality Disorders and Mental Retardation, which form Axis II. The Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I; First et al., 1997) is a semi-structured clinical interview to determine the patient’s Axis I diagnoses based on DSM. The screener uses the SCID-I during the screening (step 1 in the procedure). The present study focuses on patients with a primary diagnosis (i.e. the most serious disorder) on Axis I. The diagnoses on Axis I are divided into clusters. Each diagnosis belongs to one cluster. An example of a diagnosis is Social Phobia (coded 300.23) in the cluster Anxiety Disorders. Another example is Primary Insomnia (coded 307.42) in the cluster Sleep Disorders. The result of the screening with the SCID-I is a list of zero or more diagnoses on Axis I. If there are no diagnoses, there is no need for the available kind of treatment. If there are one or more diagnoses, the screener indicates the order of seriousness. The primary diagnosis is the target of treatment. To determine the effect of treatment, there is a pre- assessment (step 4 in the procedure) and post-assessment (step 6 in the procedure) in which the patient completes three questionnaires. The questionnaires that are used in the assessments are as follows: 1. Symptom Checklist-90 (SCL-90) (Arrindell & Ettema, 1986) – a series of 90 physical and psychological complaints, which the patient rates for the degree of distress associated with these complaints on a 5-point Likert scale. The sum score on all 90 items can be used as a global measure for the severity of Axis I psychopathology. The 90 items can be divided into nine groups of items, which belong to a specific complaint. There is no overlap between the items of a group. The sum score of these items forms the subscales. The subscales of the Dutch SCL-90 are agoraphobia, anxiety, depression, somatic complaints, inadequacy of thought and action, mistrust and interpersonal sensitivity, hostility, sleeping problems and ‘other.’ 2. Fear Questionnaire (Marks & Mathews, 1979; Zuuren, 1988) – a 32-item self-report measure. Each item is rated on a 9-point scale. The 32 items can by divided into eight groups of items, each group forming so- called the subscales. There is no overlap between the items of a subscale. The first four subscales refer to avoidance (Dutch abbreviation within the questionnaire: FQV) regarding the main phobia, social phobia, blood injury and agoraphobia. The last four subscales refer to anxiety (Dutch abbreviation within the questionnaire: FQA) regarding the main phobia, social phobia, blood injury and agoraphobia. The main phobia is formulated for each patient separately as the situation that the patient is most afraid of and requests treatment for. 3. Biographical characteristics – a list of questions concerning the patient’s age, gender, religion, marital status, education level, use of medication, earlier treatment and duration of the complaint. 2.2. CBR in health care Case-based reasoning is a technique for problem-solving that is based on the decision-making process of humans by learning from experiences (Kolodner, 1992; Aamodt & Plaza, 1994). The technique finds a solution for a problem by reusing the solutions of similar problems in the past. CBR can be used in domains where the knowledge is incomplete and complex. It has two major advantages: (1) real-world examples can be used as a knowledge base © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 (CBR does not require an explicit domain model), and (2) CBR can explain why it provides a solution by presenting similar cases of the past. CBR is a popular AI-technique in the medical domain. CBR applications were built to enhance the work of health experts and to improve the efficiency and quality of health care (Holt et al., 2006). Two of the earliest medical CBR systems for diagnosis and decision support were CASEY for heart failure (Koton, 1989) and MEDIC for dyspnoea (Holt et al., 2006). More recent work in this area are systems for early detection of breast cancer (Hung & Chen, 2006), for diagnosing neuromuscular diseases (Pandey & Mishra, 2009a), diagnosing breast cytology (Ahn & Kim, 2009) and the system T-CARE, a temporal case retrieval system for medical scenarios used in an intensive care burn unit (Juarez et al., 2011). CBR applications in health care are mostly used for determining diagnoses and as a medium for treatment (Pandey & Mishra, 2009b). In the mental health care, there are only a few CBR applications in use. The most famous system is SHRINK (Kolodner, 1983). It is one of the first CBR systems in the clinical field. SHRINK is also designed for determining the patient’s diagnosis. More recent work in the area of mental health care is a CBR system for the diagnosis of attention-deficit hyperactivity disorder (ADHD) (Brien et al., 2005). Although CBR has not been used for predicting the effect of mental health treatment, CBR is successfully used for prediction in other areas. Some examples are predicting the success of an in vitro fertilisation treatment (Jurisica et al., 1998), predicting the ecological risks of pesticides in freshwater ecosystems (Brink et al., 2002) and predicting financial distress (Sun & Hui, 2006). 2.3. Techniques Case-based reasoning is a methodology, which employs different techniques. The choice of which techniques are used in a CBR system depends on the contents and purpose of the system. In this section, CBR details and the techniques used for CBRth are discussed. 2.3.1. CBR cycle CBR seeks solutions to new problems by referencing a casebase with past experiences. The casebase forms the core of any CBR system. The CBR process is a cycle containing four steps (Aamodt & Plaza, 1994). The four steps are as follows: 1. RETRIEVE, which matches a new problem with the casebase and finds one or more similar cases; 2. REUSE, which uses the solutions for the similar cases found in the RETRIEVE step to suggest a solution for the new problem; 3. REVISE, which investigates whether the solution suggested by the REUSE step actually solves the problem after it is tried out; and 4. RETAIN, which stores the new problem and its solution in the casebase. 2.3.2. Forecasting-by-analogy If the solutions provided by a CBR system can be assigned to different classes, CBR may be used as a classifier system to predict to which class a target case belongs. This classification is called forecasting- by-analogy and contains three steps (Jo & Han, 1996; Li et al., 2009): 1. identifying significant features to describe the target case; 2. searching for similar cases in the casebase; and 3. predicting the class of the target case based on the classes of the similar cases. Compared to the CBR cycle, the first step concerns the case representation according to the structure of the casebase, the second step corresponds to RETRIEVE and the third step corresponds to REUSE. To complete the CBR cycle, the two steps REVISE and RETAIN can be added. 2.3.3. Nearest neighbour method For the search for similar cases (RETRIEVE), the nearest neighbour (NN) method is the most commonly used technique in CBR systems. NN is an exhaustive search method that evaluates the dissimilarity (or similarity) between all the past cases and the new case (Tsai et al., 2005). There are several variants of the formula that is used in NN. The variant we used is shown as Formula 1 (Watson, 1999), where T is the target case, S is a case in the casebase and wi is the weight of the ith feature. The expression ƒi(ti, si) defines the similarity function for the similarity between T and S on the ith feature. Usually, the Euclidean distance is used for this equation, but different comparisons are possible. Similarity T; Sð Þ ¼ ∑ n i¼1 f i ti; sið Þ�wi � � = ∑ n i¼1 wi � � (1) The result of NN is one or more cases from the casebase, which are deemed ‘most similar’. The two main variants of NN are K-nearest neighbour (KNN) and R-nearest neighbour (RNN). The result of KNN is a fixed number (K) of cases that have the minimum dissimilarity (or maximum similarity) with the new case. The result of RNN is a variable number of cases, namely, all cases that have a dissimilarity with the new case that is less than a threshold R (Tsai & Chiu, 2007). 2.3.4. Information gain In the NN method, weights are assigned to features. The information gain could be used as a value for feature weights. Information gain is term used in the information theory introduced by Shannon (1948). Quinlan (1986, 1989) used the gain in the ID3 and C4.5 methods to create decision trees. During the creation of © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 the decision tree, at each node of the tree, the information gain is calculated for every feature. The feature with the largest information gain is chosen as a node in the tree at that position in the tree. The information gain is a value that is easy to calculate, and the calculation is also easy to automate. Only little research was carried out in which the information gain is used as a feature weight. Daelemans et al. (1993), Ling et al. (1997) and Wettschereck and Dietterich (1995) used it in an instance based learning algorithm. The information gain (IG) can be defined as IG ¼ E before partitioningð Þ- E after partitioningð Þ (2) In this formula, E stands for Shannon Entropy. Entropy within the information theory is a measure to represent the uncertainty of a message as an information source (Munakata, 1998). Shannon Entropy E of variable X(ω) with a probability distribution P(X(ω) = x) is defined by (Munakata, 1998) as E ¼ � ∑ all x P X ωð Þ ¼ xð Þ � log2P X ωð Þ ¼ xð Þ (3) 2.3.5. Weighted voting Sun and Hui (2006) described how the method weighted voting can be used by CBR in combination with KNN as a classifier. The result of KNN is a list of k cases that are most similar to the target case. The similarity of the ith of these k cases and the target cases is denoted as simi. The outcome class of an ith case is denoted as di. Every ith case belongs to an outcome class Cl (with l = 1, 2, …, q in which q is the amount of possible outcome classes) when Cl =di. For every possible outcome, we can calculate the similarity weighted voting probability prob(Cl), which indicates the probability that the target case belongs to class Cl, with Formula 2 based on Sun and Hui (2006): prob Clð Þ ¼ ∑ k i¼1 sim�i ∑ k i¼1 simi (4) with sim�i ¼ simi if di ¼ Cl 0 otherwise � We can use this weighted voting technique to make a prediction (classification) for the new case by choosing the class that has the highest similarity voting probability. 3. CBRth casebase The CMHCM database allows to assess the effect of treatment by comparing pre and post-assessments. A large dataset has been collected, which is the basis for the casebase of CBRth. In this section, the structure of the casebase is discussed. We describe the case features (Section 3.1), the outcome measure (Section 3.2), the selection of cases (Section 3.3) and the case format (Section 3.4). 3.1. Features The basic principle of this study is to use all the features that are shared by all the cases. The data we used were collected during the pre-test. From these data, we extracted 30 different features. The features are named in the first column of Table 1. Section 2.1 discusses the meaning of each of these features. Because not every feature is equally important, we assigned a weight to each feature according to its importance. The different weights have a value in the range from 0 to 1. Determining the values of these weights is part of the study and is discussed in Section 4.3. In this study, we used nominal (numerical values with no order in rank) and ordinal (a finite number of numerical values with an order in rank). In the second column of Table 1, the type of each feature is specified. The ordinal features have ranges of values. In the third column of Table 1, the range of each feature is specified. We refer to the values of these nominal and ordinal features as scores, which refer to the questionnaire scores. Because some ranges of scores are quite large, we discretised the values into a smaller number of categories. For each score values, we assigned a category (also for the nominal types and ordinal types). In this way, every feature has a score value and a category value. We assigned categories as follows: 1. For the nominal types and the ordinal types with a small range, the category value is the same value as the score value. 2. For the features of the SCL-90, we used the classification of the norm table that is provided for the Dutch version (Arrindell & Ettema, 1986). They divided the scores of the features into a 7-point scale according to the following categories: very low, low, below average, average, above average, high and very high. 3. For the remaining features, we used the equal-frequency binning method (Witten & Frank, 2000; Liu et al., 2002), an unsupervised method that divides a range into intervals with a predetermined number of cases per interval. To increase the liability of the equal-frequency binning, we used not only the cases that were included in the casebase (Section 3.3) but also cases of patients that received different treatments (1091 cases instead of 219). In the fourth column of Table 1, the method for discretisation of the corresponding feature is indicated, with the arity (the number of categories after discretisation) listed in the fifth column. 3.2. Outcome measure With CBRth, we want to predict whether a treatment will be successful or not. Therefore, there are two classes for the outcome measure: successful or not successful. To determine © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 which of these classes a case belongs to, we used the c-criterion (Jacobson & Truax, 1991). The c-criterion is a cut-off point for clinically significant change. At the time of the post-assessment, a value is measured for the severity of the patient’s disorder. In order for the patient to be classified as free from the disorder, this value should lie beneath the c-criterion. Formula 5 (Jacobson & Truax, 1991) for the c-criterion is c ¼ s0M1 þ s1M0 s0 þ s1 (5) with M0 and s0 the mean and standard deviation of a well- functioning normal population, and M1 and s1 the mean and standard deviation of a clinical population. The total sum score (SCL-total) on the SCL-90 is a global measure for the severity of Axis I psychopathology. We use the SCL-total as outcome attribute and calculated the c-criterion for this attribute. Arrindell and Ettema (1986) report the results of the SCL-90 for a large group of a well-functioning normal population (referred to as 0 in Formula 5) and the result of the SCL-90 for a large group of patients (referred to as 1 in Formula 5). These results were officially used for compiling the norms for the SCL- 90. We used these results to calculate the value of c. Because Arrindell and Ettema (1986) make a distinction between the male-population and the female-population, we have different values for c for men (141.4) and women (159.1). We used the value of c as follows. If for a patient the SCL-total of the post-assessment was lower than the c, the treatment was deemed successful, otherwise not successful. For example, if the SCL-total of the post-assessment of a woman was 150, then treatment was deemed successful because 150 < 159.1. In this study, we used the c-criterion (as described previously) for the definition of success. This method focuses on the end state functioning of the patient (i.e. the recovery) after treatment. Other definitions of success are possible, for example, the Reliable Change Index (Jacobson & Truax, 1991), which focuses on the change of the patient between the state before treatment and after treatment. 3.3. Cases The cases we used in the casebase are the real-world examples of the large dataset of pre and post-assessments. The selection of cases were based on the following principles: • Only cases in which the treatment was cognitive behavioural therapy were included. • Only cases that completed the treatment were included. Table 1: Features used in case-based reasoning therapy Feature Type Range Discretisation Arity Gain ai SCL_agoraphobia Ordinal 7–35 Norm SCL-90 7 0.027 0.241 SCL_anxiety Ordinal 10–50 Norm SCL-90 7 0.041 0.171 SCL_depression Ordinal 16–80 Norm SCL-90 7 0.084 0.108 SCL_somatic_complaints Ordinal 12–60 Norm SCL-90 7 0.089 0.143 SCL_inadequacy Ordinal 9–45 Norm SCL-90 7 0.071 0.189 SCL_sensitivity Ordinal 18–90 Norm SCL-90 7 0.024 0.096 SCL_hostility Ordinal 6–30 Norm SCL-90 7 0.042 0.280 SCL_sleeping_problems Ordinal 3–15 Norm SCL-90 7 0.069 0.538 FQA_main Ordinal 0–8 None 0.010 1.000 FQA_agoraphobia Ordinal 0–40 Equal freq. 5 0.015 0.122 FQA_blood_injury Ordinal 0–40 Equal freq. 5 0.043 0.122 FQA_social_fear Ordinal 0–40 Equal freq. 5 0.027 0.122 FQV_main Ordinal 0–8 None 0.015 1.000 FQV_agoraphobia Ordinal 0–40 Equal freq. 5 0.040 0.122 FQV_blood_injury Ordinal 0–40 Equal freq. 5 0.010 0.122 FQV_social_fear Ordinal 0–40 Equal freq. 5 0.023 0.122 Gender Nominal 1–2 None 0.003 1.000 Marital_status Nominal 1–8 None 0.002 1.000 Religion Nominal 1–5 None 0.017 1.000 Education_level Ordinal 1–11 None 0.013 1.000 Age Ordinal 16–72 Equal freq. 5 0.012 0.088 Medication Nominal 0–1 None 0.000 * 1.000 Earlier_treatment Nominal 0–1 None 0.008 1.000 Duration_complaint Ordinal 0–600 Equal freq. 5 0.011 0.008 Cluster Nominal 1–15 None 0.022 1.000 Diagnose Nominal 1–36 None 0.115 1.000 Medication_category1 Nominal 0–1 None 0.001 1.000 Medication_category2 Nominal 0–1 None 0.001 1.000 Medication_category3 Nominal 0–1 None 0.002 1.000 Medication_category4 Nominal 0–1 None 0.007 1.000 *<0.000 but not equal to 0. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 • Only cases of which all the features that we selected for this study were available were included. We left out the concept of missing values (these are the focus of follow-up research). • Only cases of which the SCL-total in the post-assessment was completed were included. We need this value to determine the outcome of the treatment. • Only cases of which the SCL-total in the pre-assessment is higher than the c were included. Is already lower than c than we cannot define the effect of the treatment in terms of the c-criterion. This resulted in a dataset of 219 cases. According to our criteria, for 98 cases, the treatment was successful, and for 121 cases, the treatment was not successful. 3.4. Case format A typical case in a CBR system consists of three parts (Watson & Marir, 1994): • the problem that describes the state of the world when the case occurred; • the solution that states the derived solution to that problem; and • the outcome that describes the state of the world after the case occurred. If CBR is used for prediction, the solution is not a part of the case format. The outcome of the new case is based on the outcome of the similar cases that were retrieved based on the problem. In each case in our casebase, the problem describes the state of the patient before treatment starts. The problem description consists of values (the score- and the category values) of all features for that case. A full description of the case features is given in Table 1. For the outcome, there are only two possible values: 0 and 1. The value 0 means that the cognitive behavioural therapy was not successful, and the value 1 means that it was successful. 4. CBRth process For CBRth, we use CBR to predict in which class a new target case falls. We classify according to forecasting-by- analogy (Section 2.3.2). In this section, we describe the process steps that a target case follows within CBRth. These are the tree steps of forecasting-by-analogy. The first step is already described in Section 3.3. Sections 4.1 and 4.2 describe the other two steps. Forecasting-by-analogy only covers the first two steps of the CBR cycle (RETRIEVE and REUSE). In Section 4.3, we describe the remaining two steps (REVISE and RETAIN) of the CBR cycle. 4.1. Search for similar cases (RETRIEVE) For the search for similar cases, we use Formula 1 for the NN method (Section 2.3.3). In this formula, two components can be customised, namely, the weight wi of every ith feature and the equation for the similarity function between T and S ƒi(ti, si) for every ith feature. How we defined wi is discussed in Section 4.3. For the similarity function, we used two different methods, the category method and the ‘score method’. For each method, we built a variant of CBRth to investigate which of the two methods provides better results. The category method is the simplest of both methods. It uses the category values of the features. The possible output of the similarity function only includes 0 and 1. The output 1 is given when the corresponding categories of the feature that are compared are equal for both cases, and the output 0 is given when they are not equal for both cases. f i ti; sið Þ ¼ 1 if ti ¼ si 0 otherwise � (6) The score method uses the score values. f i ti; sið Þ ¼ 1 � ai� ti � sij j if ti � sij j < 1 0 otherwise � (7) In Formula 7, the parameter ai can have a different value for each feature i (for ai = 1, Formulas 6 and 7 are equal). The effect of ai is that near values obtain a high degree of similarity, and distant values obtain a low degree of similarity. Because in nominal features, there is no order in the values, ai = 1 for nominal values. For the ordinal features with a small range, we also set ai = 1. For the ordinal features with a larger range, ai is a value between 0 and 1 calculated by Formula 8. The used values for ai are given in the seventh column of Table 1. In the third column (range) of Table 1, the minimum and maximum values are given. ai ¼ number of classes feature ið Þ= maximum value feature i � minimum value feature i þ 1ð Þ (8) 4.2. Predicting the outcome of the target case (REUSE) For predicting the outcome of the target case, we used weighted voting (as explained in Section 2.3.5). There are two possible values for the outcome class Cl, namely, successful and not successful. For the target case, prob(Cl) is calculated according to Formula 2 for both the possible values of the Cl. The prediction for the target case is the value of Cl for with prob(Cl) > 0.5. Sun and Hui (2006) base weighted voting on KNN. We applied a more fine-grained approach by using by using a combination of RNN and KNN, in which we limit the maximum number of selected cases (K), and on top exclude the cases that are too dissimilar (R). This means k in Formula 2 is not a fixed value, but k is a value between 0 and K. © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 4.3. Remaining steps of the CBR cycle (REVISE and RETAIN) After a patient received and completed cognitive behavioural therapy, we can determine the actual outcome, that is, whether the treatment was successful or not, based on the results of the post-assessment. In the REVISE step, we compare the predicted outcome with the actual outcome. We enter the value of the actual outcome as outcome in the case. The knowledge whether a prediction was correct or not is useful for an indication of CBRth’s accuracy. In the RETAIN step, the target case is added to the casebase. Changing the composition of the casebase can also provide insight into which features are most important in determining the best match. Particular values of the weights may evolve over time as the casebase expands. This means that tuning the features weights (wi) can also positively influence the future performance of a CBR system. If the feature weights must be tuned after each time a case is added, it is strongly preferable that they are easy to calculate. In the present study, we researched whether the information gain can be used as a feature weight. For every feature, we use the information gain that is calculated to decide which feature should be in the top note of the decision tree. We built a variant of CBRth in which the feature weights are all equal (wi = 1 for every i) and a variant with the information gain for every feature. We compared both variants to investigate which of the two provides better results. Note that we only want to investigate whether gain works as feature weight, not whether it is the most suitable feature weight; at present, our dataset is simply not large enough to draw conclusions on which kind of feature weight works best. 5. Experimental setup The introduction describes the goal of this study as twofold: (1) to investigate which techniques are suitable for implementing CBRth and (2) to investigate the accuracy of CBRth. Section 4 describes the system design of CBRth and the steps that are successively followed to use CBRth for making predictions on the success of treatments. To test CBRth, we used 219 cases (Section 3.3) to investigate both parts of our goal. In our approach, we used combinations of a training set of 218 cases and a test set of 1 case. All possible 219 combinations of training set and test set were used. The procedure that we followed consists of the following eight steps: 1. Place the 219 cases in the casebase of CBRth. 2. Select one case and remove this case from the casebase. 3. Recalculate the weights corresponding to the casebase with the remaining 218 cases (RETAIN). 4. Determine the best match of the resulting casebase for the selected case (RETRIEVE). 5. Predict the outcome for the selected case (REUSE). 6. Check whether the predicted outcome is correct (REVISE) and log the results. 7. Repeat steps 2–6 until all cases for all 219 combinations of training set and test set. 8. Evaluate the accuracy of CBRth. Our approach is based on 219-fold cross validation, to be exact leave-one-out cross validation (Witten & Frank, 2000). We used this approach for testing both goals of the study. The results of this experimental setup are presented in the next section. 6. Results This section describes our results in two sections corresponding to the two parts of our goal. The results are shown in Figures 1–5. For more detailed results, see Janssen (2009). 6.1. Best combination of techniques The first part of the goal is to investigate which techniques are suitable for implementing CBRth. We investigated whether the information gain is a useful value for feature weights, and which of the methods described in Section 4.1 is best suited for feature matching. To determine the best combination of techniques, the following parts were investigated in succession: 65 73 80 82 67 51 49 55 0 10 20 30 40 50 60 70 80 90 100 sim 0 sim 0.6 sim 0.62 sim 0.7 % c o rr e c t p re d ic ti o n s w=gain w=1 Figure 1: Category method. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 1. feature weights and 2. feature matching. 6.1.1. Feature weights We started by investigating whether to use the gain as feature weight or not. For this decision, we compared the results with the gain as a feature weight to the results without a feature weight. We did that for both the category method and the score method. The results for the category method are shown in Figure 1. In this figure, the results with the gain as feature weight are shown as w = gain, and the results without a feature weight are shown as w = 1. We show the results for four different values of R in the RNN method, namely, sim ≥ 0 (R = 1), sim ≥ 0.6 (R = 0.4), sim ≥ 0.62 (R = 0.38) and sim ≥ 0.7 66 63 60 57 69 61 51 67 0 10 20 30 40 50 60 70 80 90 100 sim 0 sim 0.6 sim 0.62 sim 0.7 % c o rr e c t p re d ic ti o n s w=gain w=1 Figure 2: Score method. Figure 3: Category method and score method comparison. 65 67 73 80 82 100 88 34 23 5 0 10 20 30 40 50 60 70 80 90 100 sim 0 sim 0.5 sim 0.6 sim 0.62 sim 0.7 % correct % predicted Figure 4: Results case-based reasoning therapy with K = max 1. © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 (R = 0.3). We combined RNN with KNN and the choice for K = max 25. A prediction can be made when there is at least one match in the casebase that achieves the minimum similarity R. The % correct predictions is the percentage predictions that were correct for all the predictions that could be made. When sim ≥ 0, all the cases that enter the CBR cycle in the casebase meet the restriction. The prediction is thus based on the 25 best matches. When sim ≥ 0.62 for only 23% of the cases that entered the CBR cycle, a prediction could be made. For 80% of these, the prediction was correct. For the category method, when sim ≥ 0, the % correct predictions is higher for w = 1 (67%) than for w = gain (65%), but the difference between the two rates is only 2%. Once the restriction of the minimum similarity is set higher, the differences between w = gain and w = 1 are greater. Here, CBRth with the gain as feature weight gives better results. In Figure 2, the results for the score method are shown. The results are also shown for w = gain and w = 1, and for the same values of R and K. When sim ≥ 0, the % correct predictions is marginally higher for w = 1 (69%) than for w = gain (66%). When sim ≥ 0.6 and sim ≥ 0.62, the results for w = gain are higher than the results of w = 1 (63% and 61% vs 60% and 51%). For sim ≥ 0.6, the difference is minimal (2%), but for sim ≥ 0.62, the difference is 9%. The difference is also relatively large when sim ≥ 0.7 (57% vs 67%); however, when sim ≥ 0.7 for only a very small number of cases that enter the cycle, a prediction could be made (3% of w = gain and 5% for w = 1). Based on the results in this section, we decided to use the gain as feature weight. 6.1.2. Feature matching After deciding on the feature weight estimation, we investigated whether to use the category method or the score method for feature matching. The results for the category method and the score method with the gain as feature weight are shown in Figure 3. We show the results for four different values of R in the RNN method, namely, sim ≥ 0 (R = 1), sim ≥ 0.6 (R = 0.4), sim ≥ 0.62 (R = 0.38) and sim ≥ 0.7 (R = 0.3). We combined RNN with KNN and the choice for K = max 25. When sim ≥0, the % correct predictions is marginally higher for the score method (66%) than for the category method (65%), the difference is negligible. With sim ≥0.6, sim≥0.62 and sim≥0.7, the results of the category method are higher than the results of the score method (73%, 80% and 82% vs 63%, 60% and 57%, respectively). We conclude that the category method gives better results than the score method. One of the characteristics of the CBR method is that the percentage of correct predictions improves as the restriction on similarity increases, because the prediction is only made if there are matches found that have a certain degree of similarity with the new case. If we examine the results in Figure 3, we can conclude that this is indeed the case for the category method. The score method shows the opposite results, namely, that there are fewer correct predictions when the degree of similarity increases. That is surprising, but the differences are small. Based on the results in this section and the previous section, we decided for CBRth to use the category method with the gain as a feature weight. 6.2. CBRth accuracy with available data We now present CBRth’s accuracy with the techniques we chose in the previous section: the category method with the gain as a feature weight. This study was performed with the 219 cases detailed in Section 3.3. Of these 219 cases, the treatment was successful for 98 cases (45%) and not successful for 121 cases (55%). As the prediction never successful is correct in 55% of the cases, 55% is the frequency baseline. To perform well, CBRth must make predictions with an accuracy significantly higher than 55%. Figure 4 shows the results for the application of CBRth to the available data. These are the results for five different values of R in the RNN method, namely, sim ≥ 0 (R = 1), sim ≥ 0.5 (R = 0.5), sim ≥ 0.6 (R = 0.4), sim ≥ 0.62 (R = 0.38) and sim ≥ 0.7 (R = 0.3). In this figure, the results are shown for which we combined RNN with KNN with K = max 1 (results for K = max 25 are shown in Figure 5, which is shown later). The figure shows two bars. The first bar shows the percentage correct predictions of the predictions that 65 67 73 80 82 65 62 73 80 82 0 10 20 30 40 50 60 70 80 90 100 sim 0 sim 0.5 sim 0.6 sim 0.62 sim 0.7 % c o rr e c t p re d ic ti o n s K=max 1 K=max 25 Figure 5: case-based reasoning therapy with K = max 1 and K = max 25. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 could be made (i.e. at least one case was selected as a best match from the casebase), and the second bar shows for how many cases a prediction could be made. Obviously, with sim ≥ 0 for all the cases, a prediction could be made. The percentage correct predictions is 65%, that is, significantly exceeding the frequency baseline. When the restriction on similarity is higher, we may expect the number of cases for which a match can be found to decrease. However, we may also expect an increase in the percentage of correct predictions. As can be seen in Figure 4, the results show this pattern. For example, for sim ≥ 0.5 for 88% of the cases, a prediction is made, which is correct for 67% of them. For sim ≥ 0.62, a prediction is made for only 23% of the cases, but it is correct for 80% of them. Figure 5 shows that the best results were achieved with K = max 1. However, the results for K = max 25 are almost equal. As the percentage of cases for which a prediction is made only depends on the value of R, it is the same for both values of K. The percentage of correct predictions for both values of K is compared in Figure 5. We only see a small drop in the percentage for sim ≥ 0.5. A straightforward explanation for the fact that for sim ≥ 0.6 the results are the same for both values of K is that for the higher restrictions on R for the majority of cases, only one match is found. Naturally, when the number of cases in the casebase increases, the number of matches may increase too. The effect of that on the accuracy of the predictions has not been tested yet, but it may be expected to improve the accuracy as better and more matches can be found. Tables 2 and 3 show the results for CBRth with K = max 1 for sim ≥ 0 and sim ≥ 0.62. In these tables, all the cases for which a prediction could be made are given in four quadrants depending on the outcome of the prediction and the actual outcome. Again, it shows that fewer predictions can be made for a higher restriction on similarity. With the figures in the tables, we can calculate the sensitivity (Formula 9), the specificity (Formula 10) and the likelihood ratio (Formula 11) (based on Woodward (2005)) sensitivity ¼ number of true positives number of true positives þ number of false negatives (9) specificity ¼ number of true negatives number of true negatives þ number of false positives (10) likelihood ratio ¼ sensitivity 1 � specificity (11) in which positive refers to successful and negative refers to not successful. The sensitivity, specificity and likelihood ratio for CBRth with K = max 1 is shown in Table 4 for five measures of similarity. The table shows that for a higher restriction on similarity, both sensitivity and specificity increase. This naturally increases the likelihood ratio. 7. Discussion The first goal of this study was to investigate which techniques are suitable for implementing CBRth. With respect to this goal, we found that information gain is suitable as a feature weight. The gain is easy to calculate. Therefore, the gain can be recalculated every time a case is added to the casebase in the RETAIN step. This can be carried out without the intervention of a developer, so a part of the maintenance of the system is automated. When in time it is discovered that some features increase in importance for the success of the treatment, the system will adjust itself automatically. The use of the gain as a feature weight allows CBRth to use all the features that are available. It is not necessary for an expert to make a selection first. However, we included only those features for which all the values for all cases were available. CBRth does not deal with missing values yet. It is therefore possible that we have left out important features. In future research, we will incorporate the handling of missing values. We developed two methods to compare features. The category method handles strict boundaries between the values of the features. The score method uses a smooth Table 2: Results for CBRth with sim ≥ 0 and K = max 1 Prediction Actual outcome Success No success Total Success 64 43 107 No success 34 78 112 Total 98 121 219 Table 3: Results for CBRth with sim ≥ 0.62 and K = max 1 Prediction Actual outcome Success No success Total Success 16 6 22 No success 4 24 28 Total 20 30 50 Table 4: Sensitivity, specificity and likelihood ratio for CBRth with K = max 1 Sensitivity Specificity Likelihood ratio sim ≥ 0 0.65 0.64 1.84 sim ≥ 0.5 0.69 0.65 1.96 sim ≥ 0.6 0.72 0.74 2.74 sim ≥ 0.62 0.80 0.80 4.00 sim ≥ 0.7 0.83 0.80 4.17 © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 change. We expected that the score method would perform better, because of its smooth character. To our surprise, the category method performed better. One possible explanation is that the score method makes similarity values drop rapidly when the feature values differ slightly. For the category method, the similarity is maximal when the features are part of the same category even if they differ slightly in feature values. With a constraint on the similarity, CBRth can only predict an outcome when there is at least one match that meets the condition. With sim ≥ 0.62 for only 23% of the cases, a prediction can be made. The reliability of this prediction is higher than without a restriction, namely, a correct prediction is made for 80% of the cases. When used in practice, the casebase will grow automatically. This means that the number of cases for which a prediction can be made with sim ≥ 0.62 will increase. As the 80% correct predictions is mainly tied to the value of R, we may expect that for this higher percentage of predictions, the accuracy will remain at least 80%, and may even increase when the predictions are based on more similar cases. The CBRth can give a prediction very fast. The computational cost for making a prediction depends on the number of cases and the number of features. Because the number of features is constant, the computational costs only depend on the number of cases; this means that computation costs are O(n), n being the number of cases. When adding new cases, the casebase, the information gain of all the features has to be recalculated. The computational costs of that are, again, O(n). The CBRth makes binary predictions: the only two possible outcome values are successful and not successful. We have not looked at the degree of improvement. In practice, it is possible that a patient is not cured after the treatment, but has significantly improved and thus benefits from the treatment. In future research, we will focus on adding a second outcome measure, namely, the degree of improvement. This study focuses on cognitive behavioural therapy for patients with an anxiety disorder. In practice, there are more treatment options for this group of patients. For each treatment option, we can use CBRth with a different casebase specific for that treatment. When the features of a new patient are entered, CBRth can make a prediction for each treatment option, and a final decision on treatment offered can be based on the predictions for each form of treatment. Because of the transparency and the interpretability of CBR, we expect that therapists will accept CBRth for decision support. Offering treatments, however, remains a human activity, and it will be a long time before CBRth grows beyond a purely advisory system. 8. Conclusions We were able to build a CBR system (CBRth) for prediction the success of cognitive behavioural therapy for a patient before start of the treatment. The best techniques for CBRth turned out to be (1) the use of the information gain for feature weighting as described in Section 2.3.4 and 4.3 and (2) the use of the category method as described in Section 4.1. For the effectiveness of CBRth, we showed that without restriction on the similarity, 65% of the predictions of CBRth were correct, which is considerably higher than the frequency baseline of 55%. With restrictions on the similarity, CBRth becomes even more reliable: with sim≥0.62, 80% of the predictions were correct, but for only 23% of the cases a prediction could be made. Future studies are necessary to assess the effectiveness of CBR compared to the classical method in the mental healthcare research (like regression) and the classical methods in the artificial intelligence (like neural networks or support vector machines). The CBRth was demonstrated to provide useful advice in relation to the treatment outcome of cognitive behavioural therapy. Future studies should incorporate different diagnostic groups and different treatment modalities. Of special interest is the use of CBR in deciding which treatment modality is indicated for a specific case, for example, pharmacotherapy versus cognitive behavioural therapy. In future research, we will also compare the predictions of CBRth with predictions an intake staff does before the start of treatment. If the CBRth predictions have a higher accuracy than those of the intake staff, it is definitely suitable as an advisory system. As CBRth even with sim ≥ 0 has an accuracy of 65%, the results of this study provide sufficient confidence in CBR as a prediction model in the mental health area. With restrictions on the similarity, CBRth reaches an accuracy of 80%. Although then only for a small part of the patients this prediction can be carried out, these are very good results. The casebase grows during use, and the larger the casebase becomes, the more chance on finding a good match. The percentage of reliable (with at least sim ≥ 0.62) predictions will only increase. It has therefore been decided that CBRth will see further development, with the ultimate goal to incorporate it in the daily practice of treatment assignments. Acknowledgements We thank the anonymous reviewers for their useful comments. We also thank the patients at the CMHCM for participating in the measurements and research assistants for collecting the data. A full overview of all results of the study on which this paper is based, is found in the work of Janssen (2009). References AAMODT, A. and E. PLAZA (1994) Case-based reasoning: foundational issues, methodological variations, and system approaches, AI Communications, 7, 39–59. AHN, H. and K.-J. KIM (2009) Global optimization of case-based reasoning for breast cytology diagnosis, Expert Systems with Applications, 36, 724–734. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 ARRINDELL, W.A. and J.H.M. ETTEMA (1986) SCL-90. Handleiding bij een multidimensionele psychopathologie-indicator, Lisse: Swets & Zeitlinger. American Psychiatric Association (1994) Diagnostic and Statistical Manual of Mental Disorders, Washington, DC: American Psychiatric Association. BRIEN, D., J. GLASGOW, D. MUNOZ, H. MUÑOZ-ÁVILA and F. RICCI (2005) The Application of a Case-Based Reasoning System to Attention-Deficit Hyperactivity Disorder Case-Based Reasoning Research and Development, Berlin/Heidelberg: Springer. BRINK, P.J.V.D., J. ROELSMA, E.H.V. NES, M. SCHEFFER and T.C. M. BROCK (2002) PERPEST model, a case-based reasoning approach to predict ecological risks of pesticides, Environmental Toxicology and Chemistry, 21, 2500–2506. DAELEMANS, W., S. GILLIS, G. DURIEUX and A.V.D. BOSCH (1993) Learnability and markedness in data-driven acquisition of stress. ITK Research Report. Institute for Language Technology and Artificial Intelligence. FIRST, M.B., R.L. SPITZER, M. GIBBON and J.B.W. WILLIAMS (1997) Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I/P), New York: New York State Psychiatric Institute, Biometrics Research Department. HOLT, A., I. BICHIDARITZ, R. SCHMIDT and P. PENER (2006) Medical applications in case-based reasoning, The Knowledge Engineering Review, 20, 289–292. HUNG, S.-Y. and C.-Y. CHEN (2006) Mammographic case base applied for supporting image diagnosis of breast lesion, Expert Systems with Applications, 2006, 93–108. JACOBSON, N.S. and P. TRUAX (1991) Clinical significance: a statistical approach to defining meaningful change in psychotherapy research, Journal of Consulting and Clinical Psychology, 59, 12–19. JANSSEN, R. (2009) Case-based reasoning voor het voorspellen van succes van therapie. faculteit Informatica, Heerlen: Open Universiteit Nederland. JO, H. and I. HAN (1996) Integration of case-based forcasting, neural network, and discriminant analysis for bankruptcy prediction, Expert Systems with Applications, 11, 415–422. JUAREZ, J.M., M. CAMPOS, J. PALMA and R. MARIN (2011) T-CARE: temporal case retrieval system, Expert Systems, 28, 324–338. JURISICA, I., J. MYLOPOULOS, J. GLASGOW, H. SHAPIRO and R.F. CASPER (1998) Case-based reasoning in IVF: prediction and knowledge mining, Artificial Intelligence in Medicine, 12, 1–24. KOLODNER, J.L. (1983) Towards an understanding of the role of experience in the evolution from novice to expert. International Journal of Man-machine Studies, 19, 497–518. KOLODNER, J.L. (1992) An introduction to case-based reasoning, Artificial Intelligence Review, 6, 3–34. KOTON, P. (1989) A medical reasoning program that improves with experience, Computer Methods and Programs in Biomedicine, 30, 177–184. LI, H., J. SUN and B.-L. SUN (2009) Financial distress prediction based on OR-CBR in the principle of k-nearest neighbors, Expert Systems with Applications, 36, 643–659. LING, C.X., J.J. PARRY and H. WANG (1997) Deciding Weights for Nearest Neighbour Algorithms using Decision Trees, Ontario: The University of Western Ontario. LIU, H., F. HUSSAIN, C.L. TAN and M. DASH (2002) Discretization: an enabling technique, Data Mining and Knowlege Discovery, 6, 393–423. MARKS, I.M. and A.M. MATHEWS (1979) Case histories and shorter communications, Behavior Research and Therapy, 17, 263–267. MUNAKATA, T. (1998) Fundamentals of the New Artificial Intelligence: Beyond Traditional Paradigms, New York: Springer-Verlag. PANDEY, B. and R.B. MISHRA (2009a) An integrated intelligent computing model for the interpretation of EMG based neuromuscular diseases, Expert Systems with Applications, 36, 9201–9213. PANDEY, B. and R.B. MISHRA (2009b) Knowledge and intelligent computing system in medicine, Computers in Biology and Medicine, 39, 215–230. QUINLAN, J.R. (1986) Induction of decision trees, Machine Learning, 1, 81–106. QUINLAN, J.R. (1989) Unknown attribute values in induction, in Proceedings of the Sixth International Workshop on Machine Learning, Ithaca, NY: Morgan Kaufmann. SHANNON, C.E. (1948) A mathematical theory of communication, The Bell System Technical Journal, 27, 379–423. SPINHOVEN, P., J. GIESEN-BLOO, R.V. DYCK and A. ARNTZ (2008) Can assessors and therapists predict the outcoume of long-term psychotherapy in Borderline Personlity disorder? Journal of Clinical Psychology, 64, 667–686. SUN, J. and X.-F. HUI (2006) Financial Distress Prediction Based on Similarity Weighted Voting CBR. Advanced Data Mining and Applications, Berlin/Heidelberg: Springer. TSAI, C.-Y. and C.-C. CHIU (2007) A case-based reasoning system for PCB principal process parameter identification, Expert Systems with Applications, 32, 1183–1193. TSAI, C.-Y., C.-C. CHIU and J.-S. CHEN (2005) A case-based reasoning system for PCB defect prediction, Expert Systems with Applications, 28, 813–822. WATSON, I. (1999) Case-based reasoning is a methodology not a technology, Knowledge-Based Systems, 12, 303–308. WATSON, I. and F. MARIR (1994) Case-based reasoning: a review, The Knowledge Engineering Review, 9, 355–381. WETTSCHERECK, D. and T.G. DIETTERICH (1995) An experimental comparison of the nearest neighbor and nearest-hyperrectangle algoritms, Machine Learning, 19, 5–27. WITTEN, I. and E. FRANK (2000) Data Mining: Practical Machine Learning Tools and Techniques with Java Implementation, San Francisco: Margan Kaufmann Publishers. WOODWARD, M. (2005) Epidemiology: Study Design and Data Analysis, Boca Raton: Chapman & Hall/CRC. ZUUREN, F.J.V. (1988) The fear questionnaire: some data on validity, reliability and layout, British Journal of Psychiatry, 153, 659–662. The authors Rosanne Janssen Rosanne Janssen is a PhD candidate at the Faculty of Psychology & Neuroscience at the University of Maastricht, the Netherlands. She received the BSc degree and MSc degree with honours in computer science at Open University, the Netherlands. Her main research interests include health informatics and artificial intelligence techniques such as case-based reasoning. Pieter Spronck Pieter Spronck is an Associate Professor in Artificial Intelligence at the University of Tilburg, the Netherlands. He authored and co-authored about 100 papers on artificial intelligence and machine learning in international journals and conference proceedings. At present, his research focuses on player modelling, data mining and adaptive intelligence. © 2014 Wiley Publishing LtdExpert Systems, xxxx 2014, Vol. 00, No. 00 Arnoud Arntz Arnoud Arntz is Professor of Clinical Psychology at the University of Amsterdam, the Netherlands. His main research interests lie in the fields of anxiety and personality disorders, both applied and fundamental. He is director of various multisite studies investigating new psychological treatments of personality disorders and post-traumatic stress disorder. He also practises as a psychotherapist at the Viersprong Institute in Amsterdam, where he mainly treats patients with personality disorders. © 2014 Wiley Publishing Ltd Expert Systems, xxxx 2014, Vol. 00, No. 00 
work_y6p3hegtendnvedua5ze75kcie	----	1 Final version appears in Expert Systems with Applications (2012), vol. 39, no. 9, pp. 8193-8203. Integrated Classifier Hyperplane Placement and Feature Selection John W. Chinneck Systems and Computer Engineering Carleton University Ottawa, Ontario K1S 5B6 Canada chinneck@sce.carleton.ca May 18, 2011 Errata are shown in red. Abstract The process of placing a separating hyperplane for data classification is normally disconnected from the process of selecting the features to use. An approach for feature selection that is conceptually simple but computationally explosive is to simply apply the hyperplane placement process to all possible subsets of features, selecting the smallest set of features that provides reasonable classification accuracy. Two ways to speed this process are (i) use a faster filtering criterion instead of a complete hyperplane placement, and (ii) use a greedy forward or backwards sequential selection method. This paper introduces a new filtering criterion that is very fast: maximizing the drop in the sum of infeasibilities in a linear-programming transformation of the problem. It also shows how the linear programming transformation can be applied to reduce the number of features after a separating hyperplane has already been placed while maintaining the separation that was originally induced by the hyperplane. Finally, a new and highly effective integrated method that simultaneously selects features while placing the separating hyperplane is introduced. 1. Introduction Classifier decision trees are constructed by a sequential process of placing separating surfaces, frequently hyperplanes. It is often advantageous to use as few features as possible when placing each separating surface, generally because there are costs associated with collecting the data for each feature (e.g. the cost of a medical test), but also because using fewer features sometimes results in better classification accuracy. There is a sizable literature on feature selection for classification. Excellent summaries are provided by Dash and Liu [1997] and Liu and Yu [2005]. Following Liu and Yu [2005], there are three main categories of feature selection methods: (i) filter methods which use metrics based on the data set to select features, (ii) wrapper methods which select subsets of features and evaluate them by applying the data mining technique (e.g. separating hyperplane) while using only the selected subset of features, and (iii) hybrid methods that combine elements of both filter and wrapper methods. mailto:chinneck@sce.carleton.ca 2 A subcategory of all methods is the search strategy. Complete search evaluates all possible combinations of features; this is combinatorially explosive and impractical for more than a few features, though it does return the optimum solution (relative to the evaluation metric). For data sets having many features, alternative approaches are needed. Random search begins with a random subset of features and iterates towards an improved set, e.g. via simulated annealing. Sequential search proceeds through the set of features based on an ordering heuristic. The two most common sequential approaches are:  Sequential forward selection. This begins with no features and selects the single feature that gives the best value of the evaluation metric. It then iteratively adds the next feature that provides the next best value of the evaluation metric. One-by-one addition of features continues until a stopping criterion is met, e.g. there is no improvement in the evaluation criterion by adding the next feature.  Sequential backward elimination. This is the opposite of sequential forward selection: it begins with all features included and eliminates the feature whose removal gives the best value of the evaluation metric. It then iteratively eliminates the next feature whose removal gives the next best value of the evaluation criterion. One-by-one elimination of features continues until a stopping criterion is met, e.g. there is no improvement in the evaluation criterion by adding the next feature. This paper develops a variety of sequential forward and backwards search methods using filtering based on a new feature selection evaluation metric. It also develops a new integrated method that uses this evaluation metric while alternating between feature addition and separating hyperplane placement. Finally it develops new methods for removing features in such a way that a given linear separation is maintained. In other words, an arbitrary technique can first be applied to find a desirable separating hyperplane, and then features can be removed while maintaining the same separation. There has been relatively little work on integrated methods for simultaneous separating hyperplane placement and feature selection. Bredensteiner and Bennett [1997] solve a parametric bilinear program to do so while maintaining a specified level of total accuracy. The problem is solved by a variation of a Franke-Wolfe algorithm. Bradley and Mangasarian [1998] introduce a parameter into the objective function to permit the relative weighting of two objectives: the original objective function that seeks to find a high accuracy separation and a second objective function that minimizes the number of features. The resulting optimization problem is nonlinear but convex and is solved by a successive linear approximation algorithm. Bradley, Mangasarian, and Street [1998] examine similar approaches. Guo and Dyer [2003] report good results when these techniques are applied in facial expression recognition. Dunbar et al [2010] formulate the simultaneous hyperplane placement and feature selection problem as a nonlinear support vector machine problem, and then reformulate it as a quadratic minimization problem subject to nonnegativity constraints. This is an extension of a method 3 originally proposed by Bradley and Mangasarian [1998]. Maldonado et al [2011] present another support vector machine based method that results in a nonlinear problem that must be solved. 1.1 Hyperplane Placement and the Maximum Feasible Subset Problem The problem of placing a separating hyperplane to minimize the number of misclassified binary data points is equivalent to the following problem: given an infeasible set of linear inequalities, find the maximum cardinality subset that constitutes a feasible set [Amaldi 1994, Parker 1995, Chinneck 2001, Chinneck 2009]. This second problem is known by a number of names: the Maximum Satisfiability Problem, the Maximum Feasible Subset Problem (MaxFS), the Minimum Unsatisfied Linear Relation Problem, or the Minimum Cardinality IIS Set Covering Problem; we will use MaxFS hereafter. The conversion of the data misclassification minimization problem to the MaxFS problem is straightforward [Chinneck 2001]. Given a training set of I binary data points (i=1…I) in J dimensions (j=1…J), in which the value of attribute j for point i is denoted by dij, where the class of each point is known (either Type 0 or Type 1), construct one linear inequality constraint for each data point:  for each point of Type 0: jdijwj  w0    for each point of Type 1: jdijwj  w0 +  where  is a small positive constant (often set at 1). The variables are the unrestricted wj where j=0…J, while the dij are known constants. This set of linear inequalities has a feasible solution (which is easily found by linear programming) if and only if the set of data points can be completely separated by a single hyperplane. Where the data cannot be completely separated by a single hyperplane, the set of linear inequalities is infeasible. In this case, a solution to the MaxFS problem identifies the maximum cardinality feasible subset of constraints, and at the same time identifies the smallest subset of excluded constraints. Any feasible point satisfying the largest feasible subset of inequalities provides values for the w variables, thereby identifying the parameters of the separating hyperplane jdijwj = w0. Such a feasible point is normally found by linear programming (LP). The constraints excluded by the MaxFS solution correspond to data points that are misclassified by the resulting hyperplane. Thus a solution for the MaxFS problem provides a hyperplane that misclassifies the minimum number of data points. Unfortunately, the MaxFS problem is NP-hard [Sankaran 1993; Chakravarti 1994; Amaldi and Kann 1995]. A number of solution approaches have been developed for this problem; see Chinneck [2008, chapter 7] for a survey. Small problems can be formulated for exact solution, either as mixed-integer linear programs, or as a linear programs with equilibrium constraints (LPEC). Heuristic solutions for the LPEC formulation have been developed in the machine learning community [Mangasarian 1994, Bennett and Bredensteiner 1997]. Parker [1995] and Parker and Ryan [1996] described a method that gradually enumerates infeasible subsets of constraints from which at least one must be removed to create a feasible subset. Chinneck 4 [1996, 2001] developed a number of greedy heuristics that reduce a measure of the infeasibility of the current subset of constraints at each iteration. Any of the algorithms described above can be applied to solve the MaxFS problem for the purpose of identifying a separating hyperplane that minimizes the number of misclassifications. The methods developed in this paper are based on the infeasibility-reducing algorithms by Chinneck [1996, 2001, 2009] for several reasons. First, these methods provide the best results over a variety of categories of MaxFS problems [Jokar and Pfetsch 2008]. Second, they are easily extended to pursue goals other than maximum overall accuracy [Chinneck 2009]. Third, and most crucially, the sequential nature of the methods allows us to make the trade-off between the accuracy of the hyperplane placement and the selection of features in an integrated manner, which is the subject of this paper. Chinneck's MaxFS solution algorithms first require that the linear inequalities be converted to elastic form [Brown and Graves 1975] by the addition of nonnegative elastic variables, ei, one per inequality, as follows:  Type 0 inequalities take the elastic form jdijwj − ei  w0    Type 1 inequalities take the elastic form jdijwj + ei  w0 +  An elastic program is constructed, consisting of the elastic inequalities, nonnegativity bounds on the elastic variables, and an elastic objective function of the form minimize SINF . Because of the minimization, an elastic variable will take on a positive value only when the original non-elastic version of the constraint is violated, hence SINF indicates the "sum of the infeasibilities". The main MaxFS heuristic is then applied to the elastic program, based on the following concepts:  An LP solution of the elastic program minimizes SINF, which is a measure of the total infeasibility in the current set of constraints. SINF can be reduced by removing constraints that contribute to the infeasibility. When SINF reaches zero, the remaining constraints constitute a feasible subset.  Constraints are removed one at a time, and a new lower value of SINF is found by re- solving the LP for the smaller set of constraints.  A candidate list of constraints for removal at each iteration can be generated in various ways. The list can be as short as one member if a quick solution is desired.  When there are several candidates for removal, the one that lowers SINF the most is chosen. This greedy heuristic is highly effective in practice. A simplified statement of the most basic version of the SINF-reducing algorithm is shown in Algorithm 1 (some steps that improve efficiency are omitted for clarity of the main procedure). Numerous LPs are solved, but the method is highly efficient because each new LP is very similar to the previous one solved. This means that each new LP can be solved in just a few simplex iterations due to the advanced start routines in modern LP solvers. 5 There are several ways to construct the list of candidate constraints in Step 4 of Alg. 1. The method that yields the best results includes as candidates all constraints to which the elastic objective function is sensitive in the current LP solution, i.e. for which the reduced cost associated with ei is nonzero: in this case constraint i will be either violated or tight. This list of candidates may be lengthy, so there are other ways to construct shorter lists, but possibly at the cost of finding a smaller feasible subset, and hence a less accurate separating hyperplane. Where represents the reduced cost associated with the variable ei, the product is a good heuristic estimator of the relative magnitude of the reduction in SINF experienced when constraint i is removed during Step 5.1 of Alg. 1 (provided that ei is positive, i.e. the original non-elastic version of constraint i is violated) [Chinneck 2001]. When ei=0, the size of alone is a good estimator of the relative magnitude of the reduction in SINF when constraint i is removed during Step 5.1 of Alg. 1 [Chinneck 2001]. These two observations provide a way to construct shorter lists of candidate constraints in Step 4 of Alg. 1. Simply take the top k largest elements of the two lists: the constraints corresponding to the k largest values of for constraints having ei>0, and the constraints corresponding to the k largest values of for constraints having ei = 0. A good heuristic for selecting just a single candidate in Step 4 of Alg.1 (thereby reducing Step 5 to a single test), is to choose the constraint having the largest value of from among constraints in which ei >0 [Chinneck 2001]. When SINF>0 there is always at least one ei >0. In the sequel, the Orig algorithm will be taken to mean the original version of the algorithm which includes as candidates all constraints to which the elastic objective function is sensitive in the current LP solution. A fast version of the original version of the algorithm in which exactly one candidate is used in Step 4 of Alg. 1 is also tested. The candidate chosen is the constraint having the largest value of from among constraints having ei >0. INPUT: an infeasible set of linear constraints. 1. Elasticize the constraints by adding appropriate elastic variables. 2. Solve the elastic LP. 3. If SINF = 0 then exit. 4. Construct the list of candidate constraints for removal. 5. For each candidate constraint: 5.1. Temporarily remove the candidate constraint. 5.2. Solve the reduced elastic LP and note the new value of SINF. 5.3. Reinstate the candidate constraint. 6. Permanently remove the candidate constraint whose temporary removal gave the smallest value of SINF. 7. Go to Step 2. OUTPUT: large cardinality feasible subset of constraints. Algorithm 1: Finding a large cardinality feasible subset of constraints [Chinneck 2009]. 6 The reduced set of data points corresponding to the satisfied data point inequalities is completely linearly separated by the final hyperplane returned by the algorithm. However it may be advantageous to adjust the placement of the final hyperplane to obtain better generalization. There are variety of ways to do this, including minimizing the distance from the hyperplane to the misclassified points, maximizing the distance from the hyperplane to the correctly classified points, averaging these two approaches, etc. [Chinneck 2009]. It is also possible to maximize the margins by applying a support vector machine to the subset of correctly classified points [Cristianini and Shawe-Taylor 2000]. Note that while these methods are specifically for linear separating hyperplanes, they are easily extendable to nonlinear separating surfaces by including appropriate nonlinear data. For example, squaring the value of some feature x and including it as a new feature allows x 2 to included in the linear combination of terms that is returned by the separating hyperplane. 2. Integrating Feature Selection and Hyperplane Placement The linear program used in Alg. 1 for hyperplane placement includes an inequality for each of the data points in the binary dataset, but the elements of w, the variables whose solution values provide the coefficients of the features in the separating hyperplane, are completely unrestricted in value. However it is simple to introduce constraints that prevent certain features from being used in the resulting hyperplane. For example, to prevent feature j from being used, add the constraint wj=0. When constraints of this form are added to the set of inequalities for the data points, the complete set of constraints integrates both hyperplane placement and feature selection. We have the option of sequential forward addition of features (by removing constraints of the form wj=0), or of sequential backwards elimination of features (by adding constraints of the form wj=0). We can also alternate between removing constraints corresponding to data points, and adding or removing constraints corresponding to features. Some variants of these options that have proved to be particularly useful are described below. The algorithms that follow make use of 3 main elements:  constraints derived from the data points,  constraints to allow/remove features, and  the value of the elastic objective function, SINF, defined over only the constraints derived from the data points. Using the elastic objective function defined over only the data points reflects the goal of maximizing the overall accuracy subject to the features currently used in the separating hyperplane. Elastic variables are therefore not added to the feature constraints of the form wj=0; constraints of this type are either removed or added in their entirety (this is simple to do in most LP solvers by specifying that the constraint is either of type "equality", which includes the constraint, or of type "free" which means that the constraint is not binding). As in Alg. 1, the best choice among options is normally indicated by the largest drop in the value of SINF; this is the new metric for the filtering methods of feature selection developed here. SINF can be used 7 in both of the usual filtering modes: sequential forward selection and sequential backwards elimination. It can also be used in a new integrated manner described later. These basic building blocks are very flexible. We examine below a variety of ways to use them to allow feature selection prior to placing a hyperplane, feature reduction after a separation has been found (is there a smaller set of features that gives an equivalent separation?), and integrated hyperplane placement and feature selection. In the same vein, while simultaneously considering feature selection, these basic ingredients can also be used to pursue other hyperplane placement goals besides maximizing total accuracy, such as balancing the population accuracies, balancing the accuracies on each side of the hyperplane, etc. [Chinneck 2009]. 2.1 Feature Selection Prior to Hyperplane Placement The AddB algorithm is a sequential forward feature selection algorithm that operates prior to placing the separating hyperplane, as shown in Alg. 2. AddB ("add features before hyperplane placement") begins with no features included at all (i.e. constraints wj=0, j=1...J are in place). Features are then added one by one in a greedy manner: in each round, the feature that most reduces SINF is added, provided that it reduces SINF below the value reached when the last feature was permanently added in the previous round. Feature selection terminates when adding the next feature does not reduce SINF any further. After a set of features is selected, the separating hyperplane is then found, beginning in Step 4. 8 Of course, it is possible to proceed in the opposite way: begin with all features in place and iteratively remove them, i.e. use a sequential backwards elimination procedure. This generally increases SINF, so an exit condition is needed to prevent it from becoming too large. The DelB1 algorithm ("delete features before, variant 1"), shown in Alg. 3, allows a small percent increase in SINF relative to the last accepted value, denoted by SINFLastBest. The parameter P represents the allowable percent increase in SINF relative to SINFLastBest. There is a single run through the features, testing and potentially removing each feature one by one. The DelB2 variant allows a small percent increase relative to the original SINF with all features in place, hence it is identical to Alg. 3 except that Step 3.4 is omitted. INPUT: a set of inequality constraints representing the data points i=1…I, and a set of equality constraints representing the feature constraints j=1…J. 1. Elasticize the data point constraints by adding appropriate elastic variables. 2. Solve the elastic LP. SINFLastBest  SINF. 3. Do J times: 3.1. SINFMin  ∞, Feature  . 3.2. For each feature j=1 to J: 3.2.1. If feature j is still excluded (i.e. constraint wj=0 is still included) then: 3.2.1.1. Temporarily add the feature by removing the feature constraint. 3.2.1.2. Solve the elastic LP and note the value of SINF. 3.2.1.3. If SINF < SINFMin then SINFMin SINF and Feature  j. 3.2.1.4. Remove feature j by re-introducing the constraint wj=0. 3.3. If SINFMin < SINFLastBest then: 3.3.1. Select feature Feature by permanently removing the constraint wFeature=0. 3.3.2. SINFLastBest  SINFMin. 3.3.3. If SINFLastBest = 0 then go to Step 4. 3.4. Else go to Step 4. 4. Solve the elastic LP. 5. If SINF = 0 then exit. 6. Construct the list of candidate data point constraints for removal. 7. For each candidate constraint: 7.1. Temporarily remove the candidate constraint. 7.2. Solve the reduced elastic LP and note the new value of SINF. 7.3. Reinstate the candidate constraint. 8. Permanently remove the candidate constraint whose temporary removal gave the smallest value of SINF. 9. Go to Step 4. OUTPUT: a set of features (those for which wj0) and a separating hyperplane equation (given by jwjxj = w0). Algorithm 2: Adding features before hyperplane placement (AddB). 9 2.2 Feature Selection After Hyperplane Placement The question here is this: given a separating hyperplane found by an arbitrary method, is there a hyperplane that has the same sets of correctly and incorrectly classified points but which uses fewer features? The building blocks allow this question to be addressed in a number of ways. Given a separating hyperplane, the points in the training set can be grouped into the correctly and incorrectly classified sets. Incorrectly classified points are removed, leaving only the correctly classified points. The AddA algorithm (add features after hyperplane placement) operates on the constraints derived from the correctly classified points in a sequential forward selection manner as shown in Alg. 4. All of the feature constraints are included at the outset (i.e. all features are excluded), so SINF will initially be greater than zero. At least one feature must be added, and Steps 3 and 4 take care of identifying and adding the feature that most reduces SINF. Step 5 adds further features in a series of iterations. Each iteration adds the feature that most decreases SINF. The iterations cease when SINF reaches zero . INPUT: a set of inequality constraints representing the data points i=1…I, and a parameter P. 1. Elasticize the data point constraints by adding appropriate elastic variables. 2. Solve the elastic LP. SINFLastBest  SINF. 3. For each feature j=1 to J: 3.1. Delete feature j by adding the feature constraint wj=0. 3.2. Solve the elastic LP. 3.3. If SINF > SINFLastBest×(1+P/100) then reinstate feature j by removing the feature constraint wj=0. 3.4. Else SINFLastBest  SINF. 4. Solve the elastic LP. 5. If SINF = 0 then exit. 6. Construct the list of candidate data point constraints for removal. 7. For each candidate constraint: 7.1. Temporarily remove the candidate constraint. 7.2. Solve the reduced elastic LP and note the new value of SINF. 7.3. Reinstate the candidate constraint. 8. Permanently remove the candidate constraint whose temporary removal gave the smallest value of SINF. 9. Go to Step 4. OUTPUT: a set of features (those for which wj0) and a separating hyperplane equation (given by jwjxj = w0). Algorithm 3: Deleting features before hyperplane placement, allowing a small increase in SINF relative to last accepted value (DelB1). 10 The DelA algorithm takes the opposite tack of sequential backward elimination by initially including all of the features and gradually deleting them, as shown in Alg. 5. Note that only the constraints for points that were incorrectly classified by the previously found separating hyperplane are elasticized, and hence the elastic variables for only these constraints appear in the elastic objective function. All features are initially included. In each of up to J rounds, Step 2 identifies and removes the feature whose removal causes the smallest increase in SINF. Note that if too many features are removed then the LP may be infeasible; this is handled in Step INPUT: a set of correctly classified data points. 1. Construct and elasticize the data point constraints for the correctly classified points. Construct and add the feature constraints for j=1…J. 2. Solve the elastic LP. SINFLastBest  SINF. SINFMin  ∞. 3. For each feature j=1 to J: 3.1. Temporarily include feature j by deleting the feature constraint wj=0. 3.2. Solve the elastic LP. 3.3. If SINF < SINFMin then SINFMin  SINF and Feature  j. 3.4. Remove feature j by adding the feature constraint wj=0. 4. If SINFMin < SINFLastBest then: 4.1. SINFLastBest  SINFMin. 4.2. Permanently add feature Feature by removing constraint wFeature = 0. 5. For each feature j=1 to J: 5.1. SINFMin  ∞. 5.2. For each feature k=1 to J: 5.2.1. If feature k is not included (i.e. constraint wk=0 is included) and the elastic objective function is sensitive to it then: 5.2.1.1. Temporarily include feature k by deleting the feature constraint wk=0. 5.2.1.2. Solve the elastic LP. 5.2.1.3. If SINF < SINFMin then: 5.2.1.3.1. SINFMin  SINF. Feature  k. 5.2.1.3.2. If SINF=0 then go to Step 5.3. 5.2.1.4. Remove feature j by adding the feature constraint wj=0. 5.3. If SINFMin < SINFLastBest then: 5.3.1. SINFLastBest  SINFMin. 5.3.2. Permanently add feature Feature by removing constraint wFeature = 0. 5.3.3. Solve the elastic LP. 5.3.4. If SINFLastBest = 0 then exit. 5.4. Else exit. 6. Exit. 7. OUTPUT: a set of features (those for which wj0) and a separating hyperplane equation (given by jwjxj = w0). Algorithm 4: Selecting features by adding them after a separation is available (AddA). 11 2.2.1.3. The algorithm terminates when the removal of each remaining feature causes infeasibility in the LP, hence no feature can be identified for removal. The rationale behind Alg. 5 is that the separating hyperplane should remain as close as possible to the misclassified points (as measured by SINF over the elastic constraints for the misclassified points). Alg. 5 also works when the set of misclassified points is empty. However in this case, SINF is always zero, so the method provides no guidance concerning the ordering of the features to test for potential removal: features are tested in their natural order. 2.3 Integrated Feature Selection and Hyperplane Placement A major advantage of including the feature removal constraints along with the data point constraints is that both feature selection and hyperplane placement can be handled in an integrated manner in a single model. While feature selection can be done before or after hyperplane selection as in Sections 2.1 and 2.2, the two operations can also be alternated, as described in this section. Hyperplane placement and the selection of features proceed together. The Int ("integrated") algorithm described in Alg. 6 is a straightforward adaptation of the MaxFS algorithm that includes the feature constraints directly in Alg. 1 along with the data point constraints. However to initialize the process the algorithm must first identify at least one feature to include, as shown in Steps 3-5. Steps 3 and 4 first attempt to add the single feature that most reduces SINF (by eliminating the associated feature constraint). If that process fails, INPUT: a set of correctly classified points, and a set of incorrectly classified points (possibly empty). 1. Construct and elasticize the data point constraints for the incorrectly classified points. Construct and add the nonelastic data point constraints for the correctly classified points. 2. For each feature j=1 to J: 2.1. SINFMin  ∞. Feature  0. 2.2. For each feature k=1 to J: 2.2.1. If feature k is still included (i.e. constraint wk=0 is not included) then: 2.2.1.1. Temporarily remove feature k by adding the feature constraint wk=0. 2.2.1.2. Solve the elastic LP. 2.2.1.3. If elastic LP is infeasible then SINF  ∞. 2.2.1.4. If SINF < SINFMin then SINFMin  SINF and Feature  k. 2.2.1.5. Reinstate feature k by removing the feature constraint wk=0. 2.3. If Feature > 0 then permanently remove feature Feature by adding the feature constraint wFeature=0. 3. Solve the elastic LP and exit. OUTPUT: a set of features (those for which wj0) and a separating hyperplane equation (given by jwjxj = w0). Algorithm 5: Selecting features by removing them after a linear separation is available (DelA). 12 then Step 5 initiates the opposite process: it adds all features (eliminates all feature constraints) and then tries to remove as many as possible (by adding the feature constraints back in), choosing at each step the feature that least increases SINF as long as SINF does not increase beyond the original value. Once this is done, then a straightforward application of the MaxFS algorithm follows in Step 6 and beyond. Initial testing showed that Step 5 is needed in a third of the data sets (4 of 12), and does not impact the success of the method. 13 INPUT: a set of inequality constraints representing the data points i=1…I, and a set of equality constraints representing the feature constraints j=1…J. 1. Elasticize the data point constraints by adding appropriate elastic variables. 2. Solve the elastic LP. SINFLastBest  SINF. SINFMin  ∞. 3. For each feature j=1…J: 3.1. Include feature j by removing the feature constraint wj=0. 3.2. Solve the elastic LP. 3.3. If SINF < SINFMin then SINFMin  SINF and Feature  j. 3.4. Remove feature j by adding the feature constraint wj=0. 4. If SINFMin < SINFLastBest then: 4.1. SINFLastBest  SINFMin. 4.2. Permanently add feature Feature by removing constraint wFeature=0. 5. Else: 5.1. SINFReference  SINFLastBest. 5.2. Include all features by removing all feature constraints. 5.3. Solve the elastic LP. SINFLastBest  SINF. 5.4. For each feature j=1…J: 5.4.1. SINFCandidate  ∞. Feature  0. 5.4.2. For each feature k=1…J: 5.4.2.1. If feature k is still included (i.e. constraint wk=0 is not included): 5.4.2.1.1. Remove feature k by adding constraint wk=0. 5.4.2.1.2. Solve elastic LP. 5.4.2.1.3. If SINF < SINFCandidate then SINFCanadidateSINF and Featurek. 5.4.2.1.4. Reinstate feature k by removing constraint wk=0. 5.4.3. If SINFCandidate < SINFReference then permanently remove feature Feature by adding constraint wFeature=0. 5.4.4. Else go to Step 6. 6. Solve the elastic LP. 7. If SINF = 0 then exit. 8. Construct the list of candidate data point and feature constraints for removal. 9. For each candidate constraint: 9.1. Temporarily remove the candidate constraint. 9.2. Solve the reduced elastic LP and note the new value of SINF. 9.3. Reinstate the candidate constraint. 10. Permanently remove the candidate constraint whose temporary removal gave the smallest value of SINF. 11. Go to Step 6. OUTPUT: a set of features (those for which wj0) and a separating hyperplane equation (given by jwjxj = w0). Algorithm 6: The integrated algorithm for simultaneous feature selection and hyperplane placement (Int). 14 3. Experiments A series of experiments were run to test the algorithms described in Section 2. These were run on an Intel Core 2 Quad CPU Q9400 running at 2.67 GHZ with 8 Gbytes of RAM and a 64-bit Windows Vista operating system. All experiments were run in a single-threaded manner, though other operations on the machine resulted in minor variations in times for repeated experiments. The software prototype is coded in Octave 3.2.3 [Octave 2011], an interpreted open-source Matlab clone. An essential element of the method is the LP solver, which is GLPK in Octave [GLPK 2011]. GLPK is a moderately capable LP solver according to a recent benchmark comparison [Mittelmann 2011]. The software prototype code is not optimized for performance in any way. Many speed improvements are possible including storage of intermediate results to avoid re-solving LPs in numerous places, use of a better LP solver, and use of a compiled language. The parameter P used in the DelB1 and DelB2 algorithms is set at 5, a useful value in preliminary testing. The reported statistics are based on the final separating hyperplane returned by the hyperplane placement algorithm. No attempt is made to adjust this final hyperplane prior to applying it to the testing set. It is possible that testing set results may be improved by such a concluding step, such as finding a support vector machine solution for the final separable set of data points, but preliminary experiments showed no particular advantage in doing this. The algorithms developed in this paper are suitable only for data sets consisting solely of features having numerical values (real, binary, integer), and having a binary outcome. Twelve data sets meeting these criteria were chosen from the well-known UCI repository [Frank and Asuncion 2010]; basic statistics are summarized in Table 1. Note that incomplete instances were removed in all cases, so the number of instances shown may differ from the number shown in the UCI repository. Some multicategory data sets were converted to binary by choosing one of the categories as the positive outcome with all others deemed negative outcomes, as shown in the comments column in the table. The average number of features over the data sets is 30.00. Some characteristics of the data sets are noteworthy:  The number of instances in a data set affects the granularity of the results, especially in ten-fold cross-validation. The iris data set has just 150 points, which leaves just 15 points in a ten-fold test set. One additional misclassified point in this case amounts to a 6.7% reduction in overall classifier testing accuracy.  Some data sets have many features and hence more potential for feature reduction. Six data sets have more than 10 features: musk1, sonar, vehicle, wdbc, wine, and wpbc.  It is easy to find a 100% accurate separating hyperplane during training for these four data sets: musk1, sonar, wdbc, and wine.  The fractions of the total instances in a data set that are of either type affects feature selection. For example, wpbc is 85.6% accurate with a null separating hyperplane 15 because positive instances constitute 14.4% of the population. Thus simply classifying all points as negative achieves an 85.6% accuracy. Name Data Set in Frank and Asuncion [2010] Instances Features Comments breast1 Breast Cancer Wisconsin (Original) 683 9 bupa Liver Disorders 345 6 glass Glass Identification 214 9 Type 2 vs. other types. iris Iris 150 4 Versicolour vs. other types. newthyroid Thyroid Disease, database from Stefan Aeberhard 215 5 Type 1 vs. other types. pima Pima Indians Diabetes 768 8 sonar Connectionist Bench (Sonar, Mines vs. Rocks) 208 60 vehicle Statlog (Vehicle Silhouettes) 850 18 Bus vs. other types. wdbc Breast Cancer Wisconsin (Diagnostic) 569 30 wine Wine 178 13 Type 3 vs. other types. wpbc Breast Cancer Wisconsin (Prognostic) 194 32 Outcome: recurrence within 24 months. musk1 Musk (Version 1) 476 166 TABLE 1: DATA SETS USED IN EXPERIMENTS. The main results of the ten-fold cross-validation experiments are summarized in Table 2. Times are in seconds, accuracies are in percent. "Feature time" is the subset of the total training time that is devoted to selecting the features. There is no feature selection time for the Orig algorithm, and it is not meaningful to separate it out in the integrated Int algorithm. The highest test accuracy and the smallest number of features used for each model are shown in boldface. mean variance train feat train test train train feat train test train data set alg. time time acc acc feat time time acc acc feat breast1 Orig 23.8 98.42 95.90 9.00 19.9 0.05 3.29 0.00 breast1 Int 34.2 98.00 96.34 4.50 21.8 0.09 4.43 0.50 breast1 AddB 26.4 2.7 98.42 95.90 9.00 19.5 0.0 0.05 3.29 0.00 breast1 DelB1 22.6 0.7 98.16 96.19 5.80 15.7 0.0 0.08 1.55 0.40 breast1 DelB2 23.1 2.0 98.21 95.60 6.30 18.0 0.0 0.06 5.77 0.23 breast1 AddA 26.0 2.6 98.42 95.60 8.40 18.8 0.0 0.05 3.39 0.71 breast1 DelA 24.1 0.4 98.42 95.75 8.10 19.5 0.0 0.05 4.54 0.54 bupa Orig 136.1 75.91 69.29 6.00 30.2 0.40 17.88 0.00 bupa Int 138.1 75.94 69.29 5.80 57.1 0.40 25.57 0.18 bupa AddB 63.8 0.0 61.32 59.42 2.30 0.2 0.0 2.28 5.23 0.46 16 mean variance train feat train test train train feat train test train data set alg. time time acc acc feat time time acc acc feat bupa DelB1 107.7 0.1 68.96 67.82 2.20 18.6 0.0 5.95 59.16 0.18 bupa DelB2 113.9 0.2 74.81 70.43 4.00 22.3 0.0 0.96 41.63 0.00 bupa AddA 135.0 0.2 70.43 65.82 4.10 28.3 0.0 73.87 41.93 8.10 bupa DelA 135.3 0.1 75.91 69.29 5.80 30.1 0.0 0.40 17.88 0.18 glass Orig 27.3 80.47 65.50 9.00 8.6 7.35 163.67 0.00 glass Int 21.0 72.58 66.39 3.60 30.6 40.23 29.6 7.60 glass AddB 17.3 0.1 66.36 62.64 1.20 15.3 0.0 6.28 11.9 2.62 glass DelB1 18.3 0.1 67.97 65.87 1.90 9.3 0.0 40.32 10.7 6.77 glass DelB2 23.2 0.2 79.70 71.08 6.60 8.6 0.0 8.01 41.7 0.27 glass AddA 27.2 0.3 80.47 65.50 9.00 8.3 0.0 7.35 163.7 0.00 glass DelA 27.2 0.1 80.42 65.48 8.80 8.5 0.0 7.04 141.5 0.18 iris Orig 4.3 83.26 80.00 4.00 0.0 0.39 39.5 0.00 iris Int 4.2 76.52 70.67 2.10 0.1 3.66 120.5 0.54 iris AddB 4.3 0.0 83.26 80.00 4.00 0.1 0.0 0.39 39.5 0.00 iris DelB1 3.8 0.0 78.67 74.00 2.00 0.1 0.0 18.51 113.1 1.11 iris DelB2 3.8 0.0 79.70 75.33 2.30 0.2 0.0 15.51 128.9 0.90 iris AddA 4.3 0.0 83.26 80.00 4.00 0.1 0.0 0.39 39.5 0.00 iris DelA 4.3 0.0 83.26 80.00 4.00 0.0 0.0 0.39 39.5 0.00 musk1 Orig 3.2 100.00 67.63 166.00 0.0 0.00 34.2 0.00 musk1 Int 1051.0 94.02 81.28 24.70 7283.9 0.95 23.86 5.79 musk1 AddB 1015.8 1013.7 100.00 80.66 59.30 6296.2 6302.2 0.00 23.07 10.90 musk1 DelB1 102.4 99.7 100.00 77.30 90.40 7.0 6.9 0.00 25.7 8.71 musk1 DelB2 7746.2 7743.2 100.00 77.30 90.40 18078.2 18088.1 0.00 25.67 8.71 musk1 AddA 1024.9 1008.8 100.00 80.66 59.30 6183.1 6383.6 0.00 23.07 10.90 musk1 DelA 1499.7 1494.1 100.00 77.51 90.40 1973.5 1973.5 0.00 22.14 8.71 newthyroid Orig 2.8 95.04 91.19 5.00 0.0 0.13 6.4 0.00 newthyroid Int 3.6 94.99 90.74 4.70 0.1 0.18 8.8 0.23 newthyroid AddB 2.9 0.1 95.04 91.19 5.00 0.0 0.0 0.13 6.4 0.00 newthyroid DelB1 2.7 0.0 94.94 91.19 4.60 0.0 0.0 0.17 6.4 0.27 newthyroid DelB2 2.7 0.1 94.94 91.19 4.60 0.0 0.0 0.17 6.4 0.27 newthyroid AddA 2.9 0.1 95.04 91.19 4.90 0.0 0.0 0.13 6.4 0.10 newthyroid DelA 2.9 0.1 95.04 91.65 4.90 0.0 0.0 0.13 8.2 0.10 pima Orig 1776.2 80.67 78.12 8.00 2218.4 0.25 22.11 0.00 pima Int 1560.2 80.44 77.34 6.20 6066.0 0.27 17.73 0.62 pima AddB 1774.3 1.9 80.67 78.12 8.00 2102.0 0.0 0.25 22.11 0.00 pima DelB1 1122.2 0.5 77.08 75.25 1.80 29059.6 0.0 1.57 13.17 0.18 pima DelB2 1276.4 1.8 78.31 75.12 3.00 1040.2 0.0 0.17 23.21 0.00 pima AddA 1760.1 1.9 80.67 78.25 7.90 2247.2 0.0 0.25 22.60 0.10 pima DelA 1770.2 0.3 80.67 78.25 7.90 2074.1 0.0 0.25 22.60 0.10 17 mean variance train feat train test train train feat train test train data set alg. time time acc acc feat time time acc acc feat sonar Orig 0.4 100.00 72.10 60.00 0.0 0.00 71.0 0.00 sonar Int 34.6 92.47 74.43 10.90 8.2 1.42 57.1 2.54 sonar AddB 30.8 30.5 100.00 74.43 34.70 20.0 20.1 0.00 163.0 19.12 sonar DelB1 3.5 3.1 100.00 78.81 41.40 0.0 0.0 0.00 53.2 10.27 sonar DelB2 57.1 56.7 100.00 78.81 41.40 35.4 35.5 0.00 53.2 10.27 sonar AddA 30.8 30.3 100.00 74.43 34.70 19.6 19.8 0.00 163.0 19.12 sonar DelA 20.4 20.1 100.00 78.81 41.40 4.9 5.6 0.00 53.2 10.27 vehicle Orig 61.9 98.90 94.94 18.00 183.2 0.04 12.9 0.00 vehicle Int 234.1 98.52 95.88 11.20 1216.6 0.04 8.69 1.51 vehicle AddB 653.5 8.2 84.24 81.65 7.20 251201.6 92.6 162.63 92.95 86.40 vehicle DelB1 65.0 3.0 98.51 95.88 12.10 162.2 0.0 0.05 5.3 0.32 vehicle DelB2 71.5 14.6 98.71 95.29 13.40 93.5 6.5 0.03 8.9 1.82 vehicle AddA 77.0 14.4 93.96 91.06 13.30 283.1 47.6 106.54 92.0 50.90 vehicle DelA 64.5 1.9 98.90 94.94 16.40 188.6 0.4 0.04 13.5 2.49 wdbc Orig 0.6 100.00 93.67 30.00 0.0 0.00 12.5 0.00 wdbc Int 27.9 98.67 97.53 4.60 6.3 0.05 2.2 0.27 wdbc AddB 27.6 27.3 100.00 95.43 21.10 33.4 32.6 0.00 9.8 21.21 wdbc DelB1 4.3 3.7 100.00 95.08 21.40 0.2 0.1 0.00 11.6 18.93 wdbc DelB2 30.2 29.7 100.00 95.08 21.40 93.5 93.1 0.00 11.6 18.93 wdbc AddA 28.4 27.2 100.00 95.43 21.10 33.2 33.1 0.00 9.8 21.21 wdbc DelA 7.6 6.6 100.00 94.90 21.40 2.6 2.8 0.00 11.3 18.93 wine Orig 0.1 100.00 93.27 13.00 0.0 0.00 19.9 0.00 wine Int 0.3 99.56 96.60 2.10 0.0 0.09 15.4 0.10 wine AddB 0.2 0.2 100.00 96.60 2.80 0.0 0.0 0.00 8.6 0.18 wine DelB1 0.2 0.1 100.00 96.60 3.30 0.0 0.0 0.00 15.4 1.12 wine DelB2 0.9 0.9 100.00 96.60 3.30 0.0 0.0 0.00 15.4 1.12 wine AddA 0.3 0.2 100.00 96.60 2.80 0.0 0.0 0.00 8.6 0.18 wine DelA 0.4 0.3 100.00 96.60 3.30 0.0 0.0 0.00 15.4 1.12 wpbc Orig 6.5 97.48 75.71 32.00 0.9 0.09 20.6 0.00 wpbc Int 14.6 89.00 83.45 4.30 47.2 12.40 24.8 18.01 wpbc AddB 7.6 0.3 85.68 85.05 0.20 0.6 0.0 0.22 8.7 0.40 wpbc DelB1 4.7 0.6 85.85 84.08 0.60 1.5 0.0 0.48 17.7 1.60 wpbc DelB2 14.6 8.8 95.88 75.66 19.70 0.6 0.3 0.21 66.3 1.34 wpbc AddA 7.0 0.4 85.57 85.58 0.00 0.9 0.0 0.05 4.4 0.00 wpbc DelA 9.3 2.8 97.48 76.26 26.00 0.7 1.1 0.09 20.3 10.67 TABLE 2: TEN-FOLD CROSS-VALIDATION DATA 18 3.1 Feature Selection Prior to Hyperplane Placement Three new algorithms select features using the SINF-based filter prior to placing the separating hyperplane. AddB is a sequential forward selection algorithm, and DelB1 and DelB2 are two variants of sequential backward elimination algorithms. Table 3 compares all three to the original algorithm which does not use feature selection. % acc is the difference in ten-fold average total accuracy on the testing set between the method and the Orig algorithm;  feat is the difference in the ten-fold average number of features between the method and the Orig algorithm. AddB DelB1 DelB2 data set %acc feat %acc feat %acc feat breast1 0.00 0.00 0.29 -3.20 -0.29 -2.70 bupa -9.87 -3.70 -1.47 -3.80 1.14 -2.00 glass -2.86 -7.80 0.37 -7.10 5.58 -2.40 iris 0.00 0.00 -6.00 -2.00 -4.67 -1.70 musk1 13.02 -106.70 9.67 -75.60 9.67 -75.60 newthyroid 0.00 0.00 0.00 -0.40 0.00 -0.40 pima 0.00 0.00 -2.87 -6.20 -3.00 -5.00 sonar 2.33 -25.30 6.71 -18.60 6.71 -18.60 vehicle -13.29 -10.80 0.94 -5.90 0.35 -4.60 wdbc 1.75 -8.90 1.40 -8.60 1.40 -8.60 wine 3.33 -10.20 3.33 -9.70 3.33 -9.70 wpbc 9.34 -31.80 8.37 -31.40 -0.05 -12.30 average 0.31 -17.10 1.73 -14.38 1.68 -11.97 TABLE 3: SELECTING FEATURES BEFORE PLACING HYPERPLANE. DIFFERENCES RELATIVE TO ORIG ALGORITHM. On average, all three methods reduce the number of features, as expected, sometimes by a significant number (e.g. by 106.7 features on average for AddB on the musk1 data set). However it is surprising to see that all three methods also increase the % total accuracy on average as well. While DelB1 and DelB2 both reduce the number of features on average in every data set, AddB fails to do so for 4 data sets. Overall, the best of the methods is DelB1 which has the highest average increase in ten-fold testing accuracy and removes the second largest number of features on average. DelB1 removes 14.38 features of the 30.00 average features in the data sets. The time required for the feature selection using DelB1 is very small, ranging from 0.01 seconds to 99.74 seconds over the data sets, with a geometric mean of 0.5 seconds. The geometric mean times for the other two methods are also small, though slightly larger. However, a closer analysis shows that AddB is able to remove many more features for the four data sets for which it is easy to achieve a 100% linear separation in the training set, while increasing the ten-fold training accuracy about as much as the other methods. AddB performs well in this case because it is able to terminate the process of adding features as soon as SINF reaches zero, i.e. as soon as enough features are added that a complete separation is found. 19 DelB1 remains the best method for data sets that are not easily separated in training. So DelB1 with P=5 is recommended when there is no prior information about the ability of the Orig algorithm to place a 100% accurate separating hyperplane. If it is known beforehand that such a hyperplane exists, then AddB is the method of choice. 3.2 Feature Selection After Hyperplane Placement Two new algorithms attempt to reduce the number of features after a hyperplane has already been placed, while making sure that all of the points correctly classified by the original placement continue to be correctly classified as the number of features is reduced. The two methods are AddA (sequential forward selection) and DelA (sequential backward selection). Results are summarized in Table 4, with the change in 10-fold cross-validation test set accuracy and number of features shown relative to the 10-fold test set results for the Orig algorithm. In a few cases the change in 10-fold accuracy is negative, indicating that the reduced-features hyperplane has lower accuracy compared to the Orig algorithm, even though the reduced- features hyperplane successfully classifies all of the training set points that were successfully classified when the Orig algorithm used all of the features. Overall the results are not dissimilar to those shown in Table 3 for selecting features before placing the hyperplane. AddA DelA data set %acc feat %acc feat breast1 -0.29 -0.60 -0.14 -0.90 bupa -3.46 -1.90 0.00 -0.20 glass 0.00 0.00 -0.02 -0.20 iris 0.00 0.00 0.00 0.00 musk1 13.02 -106.70 9.88 -75.60 newthyroid 0.00 -0.10 0.45 -0.10 pima 0.13 -0.10 0.13 -0.10 sonar 2.33 -25.30 6.71 -18.60 vehicle -3.88 -4.70 0.00 -1.60 wdbc 1.75 -8.90 1.23 -8.60 wine 3.33 -10.20 3.33 -9.70 wpbc 9.87 -32.00 0.55 -6.00 average 1.90 -15.88 1.84 -10.13 TABLE 4: SELECTING FEATURES AFTER PLACING HYPERPLANE. DIFFERENCES RELATIVE TO ORIG ALGORITHM. The best results are again reached for those data sets in which the Orig algorithm is able to obtain a 100% separation (musk1, sonar, wdbc, wine). This allows the AddA method to stop adding features as soon as the complete separation is achieved; and it also stops the DelA method from deleting further features. Note that these good results are achieved despite the lack of inherent ordering of the features when a complete separation is available. 20 The time required for feature selection after hyperplane placement is usually quite small. The geometric mean of the ten-fold averages is 1.75 seconds for AddA, and 1.01 seconds for DelA. The single outlier in both cases is the 166-feature musk1 data set. 3.3 Integrated Feature Selection The Int algorithm integrates hyperplane placement and feature selection, as shown in Alg. 6. The 10-fold cross-validation test set accuracies and numbers of features produced by this algorithm relative to the Orig algorithm are summarized in Table 5. Accuracy is increased on average an amount similar to the best of the methods for selecting features before and after hyperplane placement, but the average reduction in number of features is significantly larger as compared to all other methods. The average number of features is reduced for every data set, but most remarkably for musk1, sonar, and wdbc, where the number of features is dramatically smaller than found by any other method. Int data set %acc feat breast1 0.44 -4.50 bupa 0.00 -0.20 glass 0.89 -5.40 iris -9.33 -1.90 musk1 13.65 -141.30 newthyroid -0.45 -0.30 pima -0.77 -1.80 sonar 2.33 -49.10 vehicle 0.94 -6.80 wdbc 3.86 -25.40 wine 3.33 -10.90 wpbc 7.74 -27.70 average 1.89 -22.94 TABLE 5: INTEGRATED HYPERPLANE PLACEMENT AND FEATURE SELECTION. DIFFERENCES RELATIVE TO ORIG ALGORITHM. We also examined the effect of allowing the introduction of a feature to cause a small worsening in SINF, as is done in the DelB1 and DelB2 algorithms. This did not provide consistent improvements. 3.4 Discussion All six of the new methods for feature selection improve over the original algorithm, increasing accuracy on average and reducing the number of features on average. It is unrealistic to expect both an increase in accuracy and a reduction in the number of features in every case, and is in fact frequently useful to accept a small decrease in accuracy in trade for a significant reduction in the number of features, thus the two factors must be considered simultaneously. Figure 1 summarizes the trade-offs between the ten-fold average number of features and the ten-fold testing set average accuracy, relative to the best result for each data set. The data is analyzed 21 as follows. The highest ten-fold average accuracy returned by any method and the smallest ten-fold average number of features returned by any method are separately identified for each data set, and the differences to those two best values are calculated for all methods. The average ten-fold differences from the best results over all of the models are plotted for each method in Figure 1. FIGURE 1: AVERAGE PERFORMANCE RELATIVE TO BEST A perfect method would always have the smallest number of features and the highest accuracy, and hence would be at (0,0) in Figure 1. The closer a method is to (0,0), the better it is. The figure shows that all six of the new methods improve on the Orig method, having both better ten-fold testing accuracy on average and fewer ten-fold average features. Other patterns are visible. AddB is able to reduce the number of features significantly, but does not improve accuracy much on average. All of the other methods produce roughly similar improvements in ten-fold average accuracy, but at a range of numbers of features. While AddA, DelB1, DelB2 and DelA cluster between 8.6 and 14 .4 features more than best ten-fold average, Int is significantly better, averaging just 1.6 features more than best ten-fold average, while averaging just 1.96 percentage points worse than best ten-fold average accuracy. Comparisons to other methods in the literature are difficult primarily due to differing selection of data sets. However there are a few data sets in common with work by Bradley and Mangasarian (1998) and Fung and Mangasarian (2004). Results for the Int method tend to be 0 0.5 1 1.5 2 2.5 3 3.5 4 0 5 10 15 20 25 T e st A cc u ra cy : A v e ra g e D e lt a % L o w e r th a n B e st Features: Average Number More than Best Orig AddB DelB1 DelB2 AddA DelA Int 22 slightly better than those reported by Bradley and Mangasarian (2004) and comparable to the linear methods in Fung and Mangasarian (2004). Results for Int are much better than those reported by Dunbar et al [2010] for the wdbc data set, and slightly worse for the pima data set. The time required for feature selection is generally a small fraction of the total time (feature selection plus hyperplane placement) for most data sets. This is not the case for the musk1, sonar, wdbc, and wine data sets. These are the data sets which are completely classified by a single hyperplane. This has the effect of reducing the time for hyperplane placement to a very small value, hence feature selection constitutes a large fraction of the total time. The geometric means of the ten-fold cross-validation total training times (feature selection plus hyperplane placement) and the feature selection times alone are shown in Table 6. The feature selection time is not relevant for the Orig method, and cannot be separated out for Int. As can be seen by comparison with the time for the Orig algorithm, feature selection does add some time to the process. The feature selecting methods have similar training times, except for DelB1, which is significantly faster than the others. DelB1 is also one of the recommended methods if features are to be selected prior to hyperplane placement. The feature selection times are very small, with geometric means in the range of 0.50 to 2.86 seconds, but the reduced number of features often increases the training time because the time to place the separating hyperplane by the Orig method increases, usually because the number of LPs that must be solved increases. All times could be reduced significantly by an efficient implementation of the algorithm and the use of a compiled language. method geo. mean of training time (s) geo. mean of feature selection time (s) Orig 7.67 - Int 34.32 - AddB 30.15 1.26 DelB1 13.73 0.50 DelB2 38.45 2.86 AddA 28.41 1.75 DelA 26.38 1.01 TABLE 6: GEOMETRIC MEANS OF TOTAL TRAINING TIME AND FEATURE SELECTION TIME Note that the small feature selection times in Table 6 are independent of the hyperplane placement method. In other words, the feature selection times will be the same even if some other method of hyperplane placement (support vector machine etc.) is used. These feature selection methods are very fast because they solve a sequence of linear programs in which each LP model is very similar to the previous one, so advanced start techniques are effective. Finally, the ten-fold cross-validation experiments were re-run using the fast version of all algorithms which selects a single constraint representing a data point for the list of candidates to consider for removal while placing the hyperplane. The fast version chooses the constraint having the largest value of from among constraints having ei >0. This had minimal impact on the testing accuracy (an average drop of 0.39 percentage points across all methods) 23 and number of features selected (an average reduction of 0.03 features). However training time was reduced by about 60% on average, because the number of training LPs solved was reduced by the same fraction. 4. Conclusions This paper makes a number of contributions:  A new and effective filtering criterion for use in feature selection is introduced: reduction in the sum of infeasibilities (SINF). Feature selection methods based on this criterion are very quick. Used before a separating hyperplane is placed, this criterion improves the average accuracy and reduces the number of features as compared to an existing hyperplane placement method.  New methods for selecting features after a separating hyperplane has already been found are introduced. The AddA and DelA methods preserve the training set separation found by the original hyperplane. These methods also improve the average accuracy and reduce the number of features as compared to an existing hyperplane placement method.  A new integrated method that selects features at the same time as placing a separating hyperplane is introduced. The Int method is the best of all methods tested, reducing the number of features more than all other methods, on average, and increasing the testing accuracy as compared to an existing hyperplane placement method. While a particular SINF-based method for placing separating hyperplanes has been used throughout this paper, it is important to note that the feature selection methods (with the exception of Int) can be used with any hyperplane placement method. The times for feature selection alone are small, so the additional overhead will be minimal. Overall, the best results are returned by the integrated Int method, which requires the use of the SINF-based method for hyperplane placement. However if feature selection is to performed before or after hyperplane placement, then any hyperplane placement method can be used. Recommendations in this case are: (i) for feature selection beforehand use AddB if the data set is known to be completely classified by a single hyperplane, and use DelB1 otherwise, and (ii) for feature selection afterwards, use the AddA method. Finally, the speed of these algorithms can be improved dramatically by several techniques. First, the prototype algorithms implemented for this paper omit several steps that have a large impact on efficiency, e.g. recording the list of constraints to which the objective function is sensitive when a new minimum SINF is identified, which means that the LP does not need to be re-solved at the end of the loop. Such efficiency-enhancing steps should be included. Second, the algorithms should be re-implemented in a compiled rather than interpreted language, and third, a more efficient LP solver than the one provided in Octave should be used. These are tasks for future research. 24 References Amaldi E (1994). From Finding Maximum Feasible Subsystems Of Linear Systems To Feedforward Neural Network Design. Ph.D. thesis no. 1282, Département de Mathématiques, École Polytechnique Fédérale de Lausanne, Switzerland. Amaldi E, Kann V (1995). The Complexity And Approximability Of Finding Maximum Feasible Subsystems Of Linear Relations, Theoretical Computer Science 147:181-210. Bennett KP, Bredensteiner E (1997). A Parametric Optimization Method for Machine Learning, INFORMS J. on Computing 9:311-318. Bradley PS, Mangasarian OL (1998). Feature Selection Via Concave Minimization And Support Vector Machines, in J. Shavlik, editor, Proceedings of the International Conference on Machine Learning, pages 82–90, San Francisco, California, 1998. Morgan Kaufmann Publishers. Bradley PS, Mangasarian OL, Street WN (1998). Feature Selection via Mathematical Programming, INFORMS Journal on Computing 10:209-217. Bredensteiner EJ, Bennett KP (1997). Feature Minimization Within Decision Trees. Computational Optimization and Applications, 10:110–126. Brown G, Graves G (1975). Elastic Programming: A New Approach To Large-Scale Mixed Integer Optimisation, ORSA/TIMS conference, Las Vegas. Chakravarti N (1994). Some Results Concerning Post-Infeasibility Analysis, European Journal of Operations Research 73:139-143. Chinneck JW (1996). An Effective Polynomial-Time Heuristic for the Minimum-Cardinality IIS Set- Covering Problem, Annals of Mathematics and Artificial Intelligence 17:127-144. Chinneck JW (2001). Fast Heuristics for the Maximum Feasible Subsystem Problem, INFORMS Journal on Computing 13:210-223. Chinneck JW (2008). Feasibility and Infeasibility in Optimization: Algorithms and Computational Methods, Vol. 118, International Series in Operations Research and Management Sciences, Springer Science+Media LLC. Chinneck JW (2009). Tailoring Classifier Hyperplanes to General Metrics, in J.W. Chinneck, B. Kristjansson, M. Saltzman (eds.) Operations Research and Cyber-Infrastructure, Springer Science+Media LLC. Cristianini N, Shawe-Taylor J (2000). An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods, Cambridge University Press, Cambridge, UK. Dash M, Liu H (1997). Feature Selection for Classification, Intelligent Data Analysis 1:131-156. Dunbar M, Murray JM, Cysique LA, Brew BJ, Vaithilingam J (2010). Simultaneous Classification and Feature Selection Via Convex Quadratic Programming With Application to HIV-Associated Neurocognitive Disorder Assessment, European Journal of Operational Research 206: 470–478. Frank A, Asuncion A. (2010). UCI Machine Learning Repository [http://archive.ics.uci.edu/ml]. University of California, School of Information and Computer Science, Irvine, CA. 25 Fung GM, Mangasarian OL (2004). A Feature Selection Newton Method for Support Vector Machine Classification, Computational Optimization and Applications 28:185-202. GLPK (2011). Homepage for the Gnu Linear Programming Toolkit: http://www.gnu.org/software/glpk/glpk.html Guo G, Dyer CR (2003). Simultaneous Feature Selection and Classifier Training via Linear Programming: A Case Study for Face Expression Recognition, Proceedings of the 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’03). Jokar S, Pfetsch ME (2008). Exact and Approximate Sparse Solutions of Underdetermined Linear Equations, Siam Journal on Scientific Computation 31:23-44. Liu H, Yu L (2005). Toward Integrating Feature Selection Algorithms for Classification and Clustering, IEEE Transactions on Knowledge and Data Engineering 17:491-502. Maldonado S, Weber R, Basak J (2011). Simultaneous Feature Selection and Classification Using Kernel-Penalized Support Vector Machines, Information Sciences 181:115-128. Mangasarian OL (1994). Misclassification Minimization, Journal of Global Optimization 5:349- 360. Mittelmann H (2011). Benchmark of Serial LP solvers, http://plato.asu.edu/ftp/lpfree.html, accessed April 15, 2011. Octave (2011). Octave home page: http://www.gnu.org/software/octave/. Parker MR (1995). A set covering approach to infeasibility analysis of linear programming problems and related issues. Ph.D. thesis, Dept. of Mathematics, University of Colorado at Denver, Denver, Colorado. Parker MR, Ryan J (1996). Finding the Minimum Weight IIS Cover of an Infeasible System of Linear Inequalities, Annals of Mathematics and Artificial Intelligence 17:107-126. Sankaran JK (1993). A Note On Resolving Infeasibility In Linear Programs By Constraint Re- laxation, Operations Research Letters 13:19-20. http://www.gnu.org/software/glpk/glpk.html http://plato.asu.edu/ftp/lpfree.html 
work_y77b4oflprad3eump2g5d5kryu	----	Improving electronic health records retrieval using contexts Expert Systems with Applications 39 (2012) 8522–8536 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Improving electronic health records retrieval using contexts Belen Prados-Suárez a, Carlos Molina b,⇑, Carmen Peña Yañez c, Miguel Prados de Reyes c a Department of Languages and Computing Systems, University of Granada, Granada, Spain b Department of Computer Sciences, University of Jaen, Jaen, Spain c Computer Science Department, San Cecilio Hospital, Granada, Spain a r t i c l e i n f o a b s t r a c t Keywords: Electronic Health Records (EHR) Context Pertinence Contextualized access Fuzzy logic Access improvement 0957-4174/$ - see front matter � 2012 Elsevier Ltd. A doi:10.1016/j.eswa.2012.01.016 ⇑ Corresponding author. E-mail addresses: belenps@ugr.es (B. Prados- (C. Molina), carmenpy@decsai.ugr.es (C. Peña Y (M. Prados de Reyes). This paper aims to solve a recently arose problem, related to the access to the Electronic Health Records (EHR) in the Hospitals. Due to the digitalization of the information contained in the medical records, and the growing availability of devices that directly generate digital documents to include in it, the EHR are becoming unmanageable. Even more, to find a concrete item of information relevant for a given assis- tance act is a very hard, difficult and time-consuming task. To solve it we propose here the definition of contexts of access to the EHR, to exploit the logical division of the information inside each document in the EHR into data groups, and the computation of the pertinence of each data group to each context. It allows us to prioritize, the information in the EHR, even at a concrete data item level, according to the situation from which it is acceded. This way when the medical personnel is involved in an assistance act, the most relevant information for it is the one first showed, being able to widen the search but always according to the relevance. With it we not only improve the accessibility to the EHR and make easier the work of the doctors, but also enable other applications like the ubiquitous computation or the mobility, using devices like tablet PC’s and PDA’s. � 2012 Elsevier Ltd. All rights reserved. 1. Introduction The use of EHR has become a reality in the everyday practice of most of the hospitals, so it is possible to find in the literature a great variety of proposals to implement the EHR in different specialities like pediatrics, nursery, family care, emergencies, radi- ology, elder care or outpatient consultation (as can be seen in Ginsburg (2007), Cho, Staggers, & Park (2010), Gagnon et al. (2010), Karahoca, Bayraktar, Tatoglu, & Karahoca (2010), Erdal et al. (2009), Pung et al. (2009) and Vishwanath, Singh, & Winkelstein (2010), respectively); and also, under different regula- tions depending on the country, like in the Korean or the Czech medical systems (in Cho, Kim, Kim, Kim, & Kim (2010) and Nagy et al. (2010), respectively). However in most of these proposals, as well as in other studies regarding the satisfaction of the users about the implantation of the EHR like McAlearney, Robbins, Hirsch, Jorina, and Harrop (2010), Ross (2009), Vishwanath et al. (2010), Svanaes, Das, and Alsos (2008) and Vest and Jasperson (2010); or even about the comparison of different EHR systems as Bisbal and Berry (2009) and Flores Zuniga, Win, and Susilo (2010), there is a remarked ll rights reserved. Suárez), carlosmo@ujaen.es añez), prados@decsai.ugr.es problem: the great amount of information that is accumulating in the EHRs is making arise problems of access. Since all the infor- mation is always available, it is becoming really difficult to access a concrete information item required, even in relatively simple situ- ations. It becomes really serious for situations like the urgencies, where the decisions must be taken within seconds and the relevant information for the concrete case should be immediately available to support them. Even more, the problem will get worse due to the increasing use of new medical machines and devices like PACs, that automatically generate documents to be included in the EHRs (Prados & Peña, 2003; Prados et al., 2010). To solve this situation the storage and access cannot be re- stricted only to information with high clinical value, since depend- ing on the assistance act the information needed may change. This problem is so relatively recent, that up to this moment it is quite difficult to find proposals that really face its whole extent. Most of them just focus on the definition of data structures and docu- ments to organize the information provided by the EHR, offering navigation systems on it, like Jerding and Stasko (1998) and McAlearney et al. (2010). However they do not constitute a solu- tion, since they allow logical and structured access to the informa- tion, but they do not avoid the uncomfortable selections steps and the successive screen-shots to reach the desired information (Prados de Reyes, Carmen Peña Yáñez, & Suárez, 2006). In the medical research community it is clear that ‘‘having a good access to the information needed benefits the quality of the http://dx.doi.org/10.1016/j.eswa.2012.01.016 mailto:belenps@ugr.es mailto:carlosmo@ujaen.es mailto:carmenpy@decsai.ugr.es mailto:prados@decsai.ugr.es http://dx.doi.org/10.1016/j.eswa.2012.01.016 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 8523 attention received by the patients’’ (Adams, Adams, Thorogood, & Buckingham, 2007). In addition, it is being pointed the importance of taking into account the situation or context from which the ac- cess is being performed (Weinstock, 2010), as a means to improve the access to the EHRs. As an example, Collins and Speedie (2008) propose to use an infobutton engine to manage the clinician and patient context in or- der to provide concise answers to frequent questions posed by cli- nicians. ‘‘Infobuttons are information retrieval tools that help clinicians to fulfill their information needs by providing links to on-line health information resources from within an electronic medical record (EMR) system’’ (Del Fiol & Haug, 2009). These mod- els are usually based on classification models to predict clinician’s decisions. However, these models are restricted to very concrete topics like the ‘‘medication infobutton data, used to predict medi- cation-related content topics (e.g., dose, adverse effects, drug inter- actions, patient education) that a clinician is most likely to choose while entering medication orders in a particular clinical context’’ (Del Fiol & Haug, 2009). Other proposals related to the definition of contexts do not face explicitly nor directly the problem posed. They are mainly focused on the knowledge mobilization (Ray & Wimalasiri, 2006) and the ubiquitous computation (Judd & Steenkiste, 2003; Kang, Lee, Ko, Kang, & Lee, 2006); or on the standardization of Hospital Informa- tion Systems and the exchange on information between them (Cayir & Nuri Basoglu, 2008; Lahteenmaki & Kaijanranta, 2009; Nagy, Preckova, Seidl, & Zvarova, 2010b). Proposals in the first cases, can rarely be applied to the Hospital Information Systems, since they are mainly based in the use of sensors to identify the context (Kang et al., 2006) or to provide information according to the device used so other applications can perform pervasive com- puting (Judd & Steenkiste, 2003). In addition, none of these propos- als are designed nor useful for the immense databases of EHR. Proposals in the second case, instead of focusing on identifying the information that is really needed to be exchanged, are centered on adapting the system, its structures, contents and interfaces to different regulations and standards like HL7 (Data exchange stan- dard, 2011), DICOM (Image storage standard, 2011; Open-EHR, 2011), SNOMED-CT (Nomenclature medical standard, 2011), or the most recent proposal of the European Committee for Standard- ization: the ISO 13606 regulation. It leads them to forget and even obviate the needs of the health professionals, who are the real users of the system, and whose work improvements have more repercussion and impact in the quality of the medical assistance provided to the patients. Nevertheless, the problem of the access to concrete informa- tion items of interest in huge databases, do is addressed explicitly in other environments, like business, legacy and e-government (in Chaker, Chevalier, Soule-Dupuy, & Tricot (2010), Mao & Benbasat (2001) and Bohm, Wolf, & Krcmar (2010), respectively). A German office of digital services to the citizens has detected the problem through deep studies (Bohm et al., 2010), but still has not pro- posed a solution to it. In the business framework this situation has also arose, as indicated by Chaker et al. (2010) and Buchholz, Hochstatter, and Linnhoff-Popien (2007), and the proposals to solve it are based on the improvement of the information retrie- val by the definition of different contexts and business models (as in Jung, 2009), changing the actual access mode and adapting it to the real information needs of the acceding user. However, these proposals are too young and are still in their first development phases, so it is soon to extend and adapt them to other type of systems. This is why we propose to analyze the daily practice of the med- ical staff, and follow the same philosophy as these solutions, to im- prove their access to the information in the EHR based on the information they usually request in each assistance act. Following the work of a doctor we can find a great variety of situations with different purposes: from a deep study of a complex diagnosis pro- cess in his office, to a simple revision of the last consultation in- form in a control of evolution process, passing through the requirement of very concrete data in the response to an emer- gency. As can be seen, we face a wide variety of activity contexts, with quite different requirements of information. In other words, we have different sets of relevant documents or information items of the EHR, depending on the context we are involved in. Our proposal is based on the study of the access patterns so the information showed to the user can be context-sensitive. This way the system would only show to the doctor the information that is relevant to his/her present context. However it must be taken into account that the information needs are not static, and they may change along the time, so it is possible that a piece of informa- tion that today is important will be useless in the future. Moreover, the age of the data has influence too: there are cases like some analysis that must be repeated if the last result of the same type of analysis is older than a few months, since the results may change. Hence, all of these aspects must be considered when defin- ing the pertinence of the information items to the contexts. According to all of it two problems must be faced. On one hand it is necessary to identify the contexts of access. On the other hand the information relevant for each of them must be identified. To solve the first problem can be found some proposals focused on the context modeling like Bobillo, Delgado, and Gómez-Romero (2008), Chaker et al. (2010), Ehsan, Amini, and Jalili (2009), Gar- cia-Morchon and Wehrle (2010) and Chu, Johnson, and Kangarloo (2000); but most of them are just theoretical models too compli- cated to be integrated in an existing system, and also require such complex algorithms that make them not suitable for an hospital information system. Even more if the system has to be updated continually to adapt to new needs. Proposals to face the second problem are mainly oriented to identify the relevant information inside documents as Järvelin and Kekäläinen (2000), Jones (2004), Cao et al. (2006) and Kanoulas, Pavlu, Dai, and Aslam (2010) propose; but due to the great amount of data involved in the Hospital Information Systems (hundreds of millions of records) it is not possible to use them. We need a very efficient way to con- textualize the access and decide which information is relevant on each situation. In this paper we present our proposal to do it, exposing first the support we have used as base, as well as the general structure of a EHR, in Section 2. Next, in Section 3, we show how to define the contexts, as much automatically as on demand, and how to identify them when a doctor accesses the system. Then, in Section 4, we propose a method to compute the pertinence of the information groups or items to the contexts, considering its several aspects. Based on it we establish a priority ordering in the use of pertinence values, and we show how it is used to improve the access to EHR, in Section 5. After it, in Section 6, we exemplify and summarize our proposal, and we make some comments about the implementation in Section 7. Finally we present some results and also our conclu- sions in Sections 8 and 9, respectively. 2. Background First of all we must indicate that the proposal presented here has been developed in collaboration with the University Hospital San Cecilio from Granada, and that we have based their Electronic Health Record System, and used it as reference. This system stores around 800.000 EHR, containing more than 50 millions documents. In the future it is expected to have a fast increase in the size, due to the inclusion of new types of documents from two sources: old documents that still have not been digital- ized (scanned images, MRI, etc.) and new documents generated 8524 B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 from the recently and future acquired devices and equipments like PAC’s. In this section we briefly show the characteristics of this EHR system, as well as the structure of the Electronic Health Records stored on it. Fig. 1. Logical organization of the EHRs. 2.1. Electronic Health Records structure The information stored in the EHR is structured according to the Reference Model given by the ISO-13606 (2008). According to this standard, the elements of the hospital information systems are or- ganized according to an Ontology with a class structure that gives rise to the following classes: � Folder: This class represents the divisions at the highest level inside the clinical history. In our case these divisions are the assistance acts and the pathologies, so all the documents in the EHR are grouped into assistance acts or pathologies, and both classifications coexist. � Section: This class of the standard represents logical groupings of information, each one representing a set of data with an uni- form informative clinical guidance (Fig. 1), and corresponds to each document stored in the EHR. Examples of documents are from a blood analysis to a preanaesthetic study, or from an admission document to a X-ray test. � Entry: According to the standard each entry represents a clinical observation or a set of them. It corresponds to what we call data groups (i.e. the hematology information in a blood analysis). � Cluster and Element: These classes correspond to what we call data items. The difference between these classes is that the first one is used to represent a unique observation or action (a data item) that requires a complex structure like a list, a table or a temporal series (i.e. an electrocardiogram); whereas the second class represents a unique and simple value, instance of some of the types defined by it (i.e. the percentage of hematocrit in a blood analysis). As indicated in Prados and Peña (2003) and Prados-Suárez, Rev- uelta, Peña Yáñez, and Fernández (2008), each document is charac- terized by a set of properties like: Fig. 2. Screenshot exampl � the type (exploration, anamnesis, epicrisis, checkup, nursing control, intervention, external,. . . ). � the speciality (medical speciality as surgery, cardiology and so on, nursing, administrative, etc.). � the pathological or clinical process (documents about preg- nancy, cataract, diabetes, etc.). � . . . In our system, documents are organized according to assistance episodes (admissions, outpatient consultation, emergency assis- tance, day hospital,. . .) in a chronological or medical ordering, depending on the assistance processes. The documents are classi- fied according to the types, considering 1500 different documents classes in the system: intervention sheet, progress sheet, nursing e of the applications. Fig. 3. Structure of the system. 1 http://www.citrix.com. 2 http://www.oracle.com. B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 8525 sheet, pregnancy process, diabetes protocol, radiological report, and so on. An example of it can be seen in Fig. 2, where the inter- face that the medical personnel use to access the EHR is shown. When a doctor is looking for a concrete data item, from a spe- cific test made to the patient, he/she must select (on the square marked with a number 1 inside a circle in Fig. 2) the type of assis- tance act that is performing. Then a search in the square marked with 2 in Fig. 2 must be done, according to the date of the test or the medical speciality in which it was done. Once the assistance act in which the test was made is found, the doctor has to find among the documents generated in that act, the one with the re- sults of the test, in the interface part marked with 3 in Fig. 2. With it the document is recovered and finally it must be scanned to find the item of interest, by clicking on the tab marked with a number 4 in Fig. 2. Though it is possible to use different ordering criteria (marked with number 5 in Fig. 2), it can be seen that it is a long te- dious task that consumes a lot of the doctor’s time, and that must be repeated each time the doctor needs to know something about each patient attended. 2.2. Data groups Inside the documents there are data items that can also be grouped into small logical units, that we call data groups, when they are related under a clinical point of view (Fig. 1). Each data group inherits the general properties of the document where it is contained. In addition, as shown in Fig. 1, each data group has its own specific properties, like the relevance level (for the concrete patient and episode), the confidentiality level, etc. Examples of data items for a blood analysis or a preanaesthetic study are: erythrocyte, hemoglobin, corpuscular volume, amylase, GGT, HDL-cholesterol, LDL-cholesterol or VLDL-cholesterol, in the first case; and Hypertension, cardiopathy, electrocardiogram, radiologic study or echography, in the second case. These data items can be grouped into the data groups general biochemistry and lipid information, for the first type of document; and risk fac- tors or additional tests, for the second type. Here we would like to remark that the information of EHR and patient’s identification is a ‘‘special’’ data group, common to all the documents. Due to it, it is discarded from the processes explained later. This logical organization of documents and their content, allows the processing and analysis of the information, as much at docu- ment level as at individual data items level or data group level. Here we consider the data groups as the minimum unit of information, since a single data item can be managed as a data group with just one element. 2.3. EHR information system The structure of the system is shown in Fig. 3. The users access the system using medical workstations. These are normal PCs, light PCs (or net PCs), medical devices like the X-ray systems or the ultrasound scans, or the most recently incorporated terminals as the Tablet PCs and PDAs. The user then log on the system and ac- cess to a Citrix1 farm of servers where the applications are executed. All the data are stored in a data base cluster using Oracle DBMS.2 The screen-shot of the Doctor’s interface once logged is shown in Fig. 2. This system, as legally demanded, stores each access to the EHR, indicating the data acceded and, in case of modification, the mod- ified data; the staff member acceding; and the assistance situation (called ‘‘controlled assistance situation’’) in which the access occurs. From now on, we will call this access data base as Retrospective Ac- cess Data Base (RADB). The number of records stored in the RADB is in the order of hundreds of millions. We support the work proposed here on the registers of this data base since, as we will see next, their analysis allows to know which information has been acceded and the related context. 3. Contexts We call Context to a situation in the Doctor–Patient relationship inside an assistance act, requiring an access to the information pre- viously stored in the EHR. To contextualize the access to the EHR we first need to establish the set of possible situations or contexts where that access may oc- curs. Then, to exploit the contextualized access system, it is neces- sary to count on a mechanism to identify the context in which the medical staff is involved. 3.1. Context definition The contexts can be defined under three criteria or a combina- tion of them: http://www.citrix.com http://www.oracle.com 8526 B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 �P- athological process: In this case the contexts are defined based on a diagnosed pathology that requires monitoring and it is included in the EHR as such process. Some of these processes are defined by the Regional Health Administration and others by the hospital ser- vices themselves. It must be taken into account that several med- ical specialities can be involved in the same process. Some examples are the pregnancy process, the cataract process, the dia- betes process, etc. � Medical speciality: Here the contexts are defined according to the specificity of each medical speciality (pediatrics, gynecol- ogy, nursery, cardiology, etc.). � Kind of assistance: The context definition here is based on the environment where the assistance process takes places. The fol- lowing cases can be distinguished: – Diagnostic study. – Surgical intervention. – Post-surgical revision. – Evolutive revision. – Room visit. – Treatment revision of outpatient consultation. – Analytical control. – Urgent assistance situation. According to these three criteria, we ask different medical doc- tors to identify the contexts on each speciality. The set of contexts obtained has been reviewed by different groups of medical doctors to validate the results for their corresponding speciality. 3.2. Context identification Once we have the contexts, it is necessary to have an automatic method to identify when a medical doctor accesses an EHR in a gi- ven context. By means of different interviews with the medical staff we have identified some characteristics of the accesses they perform, that are important in this process: � Speciality of the medical staff like cardiology, ophthalmology, internal medicine, emergency, administration, nursing, and so on. � Position of the medical staff. There are different positions for each type of medical personnel like. Some examples are: resi- dent (from first to fifth year), facultative, section manager or head of service for the medical doctors; the categories of man- agement technician, administrative technician, section manager or head of service for the administrative personnel; or the nurs- ing position or nursing supervisor in that department. � Type of the medical workstation. The data about the type of ter- minal used to perform the access, gives a lot of information about the type of assistance act in which the medical staff is involved. This attribute has several parts: – The type of the terminal. This value gives information about the hardware used (PC, PDA, patient’s room terminal, com- puter associated to a concrete equipment like the X-ray machines or ultrasound scan, etc.). – The medical unit associated. Each terminal is associated to an unit (gynecology, pediatrics, etc.) for management rea- sons; but this information helps to identify the context. As an example if a cardiologist is acceding an EHR from a com- puter associated to the emergency unit, the context could be a cardiology emergency. – Physical location. It helps to concrete even more the type of context in which the medical staff is involved. In the previ- ous example, if the terminal acceded is located in the obser- vation room in emergencies, the context is different from the case when the access is performed from the surgery room. � The kind of the present patient’s appointment. For each appointment with a doctor, the information about its type is stored. There are around 50 usual types of appointments like first visit, checkup, scheduled visit, urgent visit, extern emer- gency, admission, several types for the different complementary tests and explorations, inter-consultation, movement between services, and so on. There are also some other types less usual or even rare but also considered in the system, like radiologic surgery. � Last visit of the patient. This information in some cases helps to predict the cause of the next appointment. As an example, always after a surgical intervention there is a post-surgical checkup. To identify the context with these attributes we use a rule- based system built on the past accesses data base (RADB). To get the data necessary to build it, a question about the context was added to the normal application (Fig. 2), so the medical staff could answer it each time an EHR was acceded. The system showed a list of contexts, filtered by a very simple criteria as the medical speci- ality, and the doctor chose the one that best fitted his/her access. This data collecting process, was active for 6 months and we have gotten an answer rate of 24.5% (about half million records). With all the information collected, the RIPPER algorithm (Co- hen, 1995) was used to build the rule-based system. This is a well known algorithm that produces good results, even compared to more recent proposals. This method basically consists on building a list of rules in an iterative way, based on the information gain. After it, a pruning process on the rule list is performed, which im- proves the classification rates of the unknown cases. To verify that the results of the algorithm can be used, we have first tested it using a 10 cross-folder validation. In this process we got an average classification rate of 89.32%, which can be consid- ered a good percentage, taking into account the high number of contexts (classes). Once the quality of the rule set has been proved, we have built a new model, but in this case using all the records. With it we have obtained an ordered list of rules, with near two hundred elements and we search for the present context by sequentially verifying this list and stopping at the first satisfied rule. Though it may seems to be a considerable number of rules, we optimize the search on it by saving the rules in a data base where all the antecedent’s variables are stored, as well as their corre- sponding consequents, in addition to their order. Hence, to obtain the context only a simple query must be done, and so the context identification supposes no significative overhead in the access process. 4. Pertinence Once the set of considered contexts is defined it is necessary to identify the relevant information for each one. The relevance of a concrete data group for a given context is what we call pertinence: the more needed or interesting the data group is for the context, the higher is its pertinence to the context. There is a great variety of factors to consider to compute this pertinence like: � The regulations about each clinical process. Usually the infor- mation relevant for each act of some pathologies (not of all), is fixed by protocols set by governmental institutions, by the hospitals or by the medical services. � The opinion of the concrete doctor. In addition to the regula- tions, each doctor can consider that, from his/her point of view, there are other items that must also be taken into account. Fig. 4. Behavior of the time pertinence PT(X), according to different values of B. B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 8527 � The own history of a concrete patient. Some data groups with- out significance for the majority of the patients, may have a great and especial importance for a given patient. � The aging of the information. With time there are tests that loose their validity, because they are too old or because there are new tests of the same type. � The access patterns. It is possible that the medical staff starts to access frequently a concrete data group for a given situation, and that they are not informed or ‘‘conscious’’ of it, so they do not include it through any of the previous ways. Taking into account this aspect of the pertinence, new patterns of access can be discovered and also the system can also automatically adapt to them. The system must be capable of representing and bringing to- gether all these aspects of the pertinence. The way we propose to do it is shown in next sections, where we present a method to cap- ture and calculate each of these aspects of the pertinence. 4.1. Static pertinence: regulations, doctors and patients As static pertinence we understand those set by medical criteria or given by the medical staff. Therefore, we consider three types of static pertinence, corresponding to three of the its aspects men- tioned above. On one hand, the medical criteria and regulations that deter- mine which information must be always taken into account for a given process or pathology. On the other hand, there personal opinions or even research studies of the doctors, that lead them to find a specific information item especially relevant for all the patients they see in a concrete context. In addition, must be considered the concrete data group is par- ticularly important for a given patient but not for the rest. Hence, we include in the system three degrees of pertinence de- fined by doctors or medical criteria: one associated to the regula- tions PRDc 2 ½0; 1� � � , another one related to the personal opinion of the doctor PCDc 2 ½0; 1� � � and the other one associated to the specific patient PPDc 2 ½0; 1� � � . 4.2. Time pertinence The pertinence of a group of data (and the implicit document) will depend too on the date of creation. It is logic that the results of an analysis will be more important if it was completed a few days before than if it was performed a year ago. However the influence of the age will not be the same for all document types: some type of analysis may be valid for several months meanwhile others are valid for years. Hence we propose to modify the pertinence of the document depending on a established age threshold in months. If the docu- ment is younger than the threshold we want the time pertinence to be high (value grater than 0.7). If it is older, we want the time pertinence to decrease and give a low value. To fulfill this restriction we propose the next definition. Definition 1. Been D a document and A the age of the it (express in months), we calculate the Time pertinence of document D as PTðDÞ¼ e� logBðAÞ e ð1Þ where B 2 ]1, +1] is a parameter defining the decreasing strength. Fig. 4 shows the behavior of the function according to the value B. Let note that the value of B is the point where the function has the first value under 0.7 (high pertinence). 4.3. Dynamic pertinence In most of the situations the doctors will not give a static perti- nence for a document neither a patient. So we need to learn the pertinences. Hence we propose to do it according to the accesses stored in the RADB database. Due to the great number of records in the RADB database we need a very efficient process. If we consider that we want the sys- tem to be dynamic and to update on-line the pertinences according to the new accesses, the efficiency requirement is even more important. As we have mentioned, different methods to calculated the rel- evance of preferences can be found in literature but the great com- plexity that they have, makes them not valid for our system. It has lead us to propose the new method explained next. To calculate to pertinence we propose to use an adaptation of the Vector Space Model Salton, Wong, and Yang (1975). This tech- nique comes from the Documentary Computing, concretely from the automatic indexation methods and retrieval systems Gil-Leyva and Rodríguez-Muñoz (1966). It is used to determine which descriptors are more specific or discriminate better between documents. The discrimination value classifies terms in the text according to their capability to distinguish some documents from others in a given collection; i.e., the discrimination value of a term depends on how the average distance between the documents changes when a content identification is set for the term. Therefore, the best words are those resulting in a higher distance. The basic idea of this model lays in the construction of a matrix or table of information items and documents, where the rows are the terms and the columns correspond to the documents acceded. The rows would correspond to the terms that would be ex- pressed according to the occurrences (access frequency) of each information item. Applying it to our case, we consider as documents (columns) the possible Contexts and as terms (rows) the data groups inside the documents. Hence, the table with the access frequencies will be like the one show in Fig. 5, where tfij represents the number of accesses to the data group i in the context j, and tfj ¼ XN i¼1 tfij ð2Þ gives information about the total accesses for context j. Fig. 5. Frequency table for data groups and Contexts. Fig. 6. Evolution of the influence according to the distance in days to the reference date. 8528 B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 However in our situation it is not enough, since we need the re- cent accesses have to a higher influence than older ones when cal- culating the pertinence. This is why we propose to measure the relevance according to the time as the weight function shown in Fig. 6. Let DR be a reference date to consider relevant or not the information for the system and DA represent the access date. In that case, we propose the following function to calculate the weight for a given date (DA): WðDAÞ¼ 2 DA�DR 365 ð3Þ where the date difference is calculated in days. This way an access made today will have more influence that the accesses of the last year but, as the time goes by, less influence than future accesses. The equation establishes that given two accesses with a year of dif- ference, the newer one will have double influence than the older one. This definition introduces in the system two important and use- ful capabilities: � The pertinences will be updated according to the aging of the access and the decreasing relevance. � The system will be adapted automatically to future accesses patterns and needs, that can even allow us to define new contexts. Then the system will update the pertinence of the data groups having more influence the newer accesses and enable the system to adapt to future needs. As shown in Fig. 6, the influence is an increasing function. This may introduce some problems of representation or loss of preci- sion when the values are very high. The definition of the influence that we have proposed allows us to avoid this problem in an easy way. We only have to move the reference date (DR) and adapt the stored values to get an easy and quick adaptation: if we add one year to the reference date and divide all the values by 2, we get the same frequency and we reduce the magnitude of the stored values. We can repeat this process as many time as needed as long as the final reference date is previous to the next access to be stored. The proposed system will do this each new year keeping the refer- ence day with one year length to actual date. The update of the values only need to change the reference date (one update sen- tence) and the stored values (a set of very simple update sen- tences) which would need just a short time. Now we have the frequency, we propose an adaptation of the inverse document frequency of the Vector Space Model (Salton et al., 1975) to measure the pertinence of a data group to a context, based on the information stored in the RADB database. Definition 2. Let C be a context and X a data groups, the restrospective pertinence is PCRðXÞ¼ tfXC tfC � �1=4 ð4Þ The idea behind this pertinence is to consider relevant a data group if the number of accesses to it is high in comparison to the total number of accesses. 4.4. Global pertinence Once we have the different considered aspects about the perti- nency and defined a way to compute them, we need to aggregate the information given by them into a single value. To do it, we ob- tain a global pertinence of a data group to a given context as in next definition. Definition 3. Let X be a group of data in a document D, and C a context, we define the global pertinence of X to C as PCGðXÞ¼ P R DcðXÞ� P C DcðXÞ� P P DcðXÞ� P C RðXÞ � � � PTðDÞ ð5Þ where � PRDc is the pertinence set by medical doctors according to the regulations, � PCDcðXÞ is the pertinence set by medical doctors for the data group to the context under their personal point of view, � PPDcðXÞ is the pertinence set by medical doctors for the data group to a given patient, � PCRðXÞ the retrospective pertinence according to prior accesses, � PT(X) the pertinence considering the age of the document, � � and � a t-conorm and a t-norm, respectively. ntextualized query process. B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 8529 For the system we have chosen the maximum and the minimum Fig. 7. Scheme of the co Fig. 8. Scheme for pertinences update process. as t-conorm and t-norm because of their simplicity, and therefore, efficient and fast calculation as well as they are quite extended. Hence, we include in the system three degrees of pertinence de- fined by doctors or medical criteria: one associated to the regula- tions PRDc 2 ½0; 1� � � , another one related to the personal opinion of the doctor PCDc 2 ½0; 1� � � and the other one associated to the specific patient PPDc 2 ½0; 1� � � . 5. Contextualized access system With all the elements to implement the contextualized access to the EHR, we show next how we propose to provide this access by presenting the use of the proposed method, as well as the up- date process that allows the system to automatically adapt to new needs. 5.1. Access to the system An scheme of the access process in shown in Fig. 7. � The doctor starts the process by logging in the system. � Using the information about the terminal and the schedule of the doctor, the system gets the context for this access using the simple rule system. � The doctor identifies the patient in the system to access his/her EHR. � The system gets the EHR and queries the static and dynamic pertinences for all the data groups that appear in his/her EHR. The result of aggregating these pertinences and the time perti- nence as shown in Eq. (5) is used to order the data. � Finally, the system selects the first data groups and returns them to the doctor ordered y priority, as well as a way to access the other data groups if the doctor needs them. 5.2. Update process The system is updated on each access so the pertinences are adapted continually to reflect doctors’ needs. In this process no manual intervention is needed and a few records are changed so it needs a short time to be executed. The update process is as follow: � When the doctor logs in the system, the context of the access is calculated as mentioned above. � The doctor asks for a data group of a specific patient. � The system then gets the required data and returns them to the doctor. At the same time the system logs the access in the RADB table and updates the frequency table used to obtain the retro- spective pertinence with this new access. Only two records are changed: the accesses to the data group in this particular con- text (tfij) and the total number of accesses to the context (tfj). A scheme summarizing of the process is shown in Fig. 8. After every access to the system, the dynamic pertinence is automatically updated so the next time a context is acceded the pertinence of the data groups to the context is computed consider- ing the most updated information. As can be seen it is done with a very low computation cost. In addition this updating process al- lows, not only to give the most updated information, but also to discover new patterns of accesses, which give us the chance to de- fine new contexts. Table 1 Documents and data groups contained on each of them. B is the parameter to define the decreasing strength. Code Document types Data groups B DT1 Electrocardiogram g1 3 DT2 Blood Analysis Coagulation (g2) 3 DT2 ’’ Immunology (g3) 3 DT2 ’’ Biochemistry (g4) 3 DT3 Discharge report g5 24 DT4 Thorax radiography g6 6 DT5 Surgery report g7 24 8530 B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 6. Example In this section we show some examples to clarify the proposed method. Here we present two contexts and a set of five different documents with several data groups, and indicate how we com- pute their pertinence to each context. In a real case the number of contexts as well as the number of documents and data groups will be considerably much higher, as shown in Sections 3.1 and 2. With these very simplified examples our aim is jut to clarify and show the correspondence between the formal notions pre- sented here and the medical terminology. The set of documents and data groups, are shown in Table 1, and the context selected are only two: � C1: Emergency after catheterization surgery � C2: Traumatology pre-surgery appointment. In Table 2 the previous accesses are shown (a simplification of the RADB), as well as the associated frequency table is shown in Table 3. Table 2 Recorded accesses to data groups in EHRs (W(D) is the weight of the access according to Date Context Data group W(D) 22/01/08 c1 g7 1.04 16/01/08 c1 g6 1.03 16/01/08 c2 g6 1.03 24/02/08 c1 g7 1.11 03/02/08 c1 g7 1.06 22/02/08 c1 g6 1.10 20/03/08 c2 g1 1.16 03/03/08 c1 g7 1.12 26/03/08 c2 g1 1.18 19/04/08 c1 g7 1.23 04/04/08 c1 g7 1.20 21/04/08 c1 g6 1.23 14/05/08 c1 g3 1.29 01/05/08 c2 g6 1.26 22/05/08 c2 g6 1.31 01/06/08 c1 g7 1.33 27/06/08 c2 g2 1.40 03/06/08 c2 g3 1.34 25/07/08 c1 g6 1.48 25/07/08 c1 g7 1.48 11/07/08 c1 g7 1.44 16/08/08 c1 g7 1.54 17/08/08 c2 g2 1.54 26/08/08 c2 g6 1.57 05/09/08 c2 g1 1.60 12/09/08 c1 g3 1.62 08/09/08 c2 g1 1.61 12/10/08 c1 g6 1.72 11/10/08 c2 g6 1.71 23/10/08 c2 g2 1.75 06/11/08 c1 g3 1.80 04/11/08 c2 g7 1.79 03/11/08 c1 g3 1.79 18/12/08 c1 g7 1.95 16/12/08 c1 g3 1.94 28/12/08 c1 g7 1.99 24/01/09 c1 g7 2.09 08/01/09 c1 g7 2.03 24/01/09 c1 g7 2.09 19/02/09 c1 g3 2.20 19/02/09 c1 g1 2.20 25/02/09 c1 g3 2.22 01/03/09 c1 g5 2.24 21/03/09 c1 g3 2.33 08/03/09 c2 g1 2.27 01/04/09 c2 g7 2.38 07/04/09 c2 g3 2.40 25/04/09 c1 g7 2.49 In the next section we present in detail two examples for two different patients. 6.1. Examples of retrieval 6.1.1. Patient 1 In the first example we consider a patient that recently had a catheter surgery. The last time he came to the hospital was for the date as shown in Fig. 6 considering DR = 01/01/2008). Date Context Data group W(D) 24/05/09 c1 g4 2.63 21/05/09 c1 g7 2.61 26/05/09 c2 g3 2.64 26/06/09 c1 g7 2.80 18/06/09 c2 g4 2.76 01/06/09 c2 g1 2.67 17/07/09 c1 g3 2.91 23/07/09 c2 g6 2.95 06/07/09 c1 g5 2.85 08/08/09 c2 g1 3.04 28/08/09 c1 g5 3.15 22/08/09 c1 g4 3.12 18/09/09 c1 g3 3.28 13/09/09 c1 g7 3.25 07/09/09 c1 g3 3.22 02/10/09 c2 g6 3.37 10/10/09 c1 g2 3.42 07/10/09 c1 g1 3.40 06/11/09 c1 g7 3.60 08/11/09 c1 g5 3.62 05/11/09 c1 g7 3.60 06/12/09 c1 g5 3.81 17/12/09 c1 g7 3.90 19/12/09 c1 g5 3.91 18/01/10 c1 g5 4.14 10/01/10 c1 g5 4.08 07/01/10 c1 g5 4.05 25/02/10 c1 g3 4.45 14/02/10 c2 g6 4.36 04/02/10 c2 g2 4.27 15/03/10 c2 g4 4.60 16/03/10 c1 g7 4.61 28/03/10 c1 g5 4.72 20/04/10 c2 g3 4.93 01/04/10 c1 g3 4.75 28/04/10 c2 g2 5.00 09/05/10 c1 g3 5.11 09/05/10 c2 g4 5.11 26/05/10 c2 g2 5.28 07/06/10 c2 g2 5.40 24/06/10 c1 g5 5.58 17/06/10 c2 g2 5.50 11/07/10 c2 g5 5.76 06/07/10 c2 g2 5.71 28/07/10 c1 g7 5.95 16/08/10 c1 g5 6.17 03/08/10 c2 g2 6.02 02/08/10 c1 g5 6.01 Table 3 Frequency table. Document types Data group Acum Frequency PCC C1 C2 C1 C2 C1 C2 DT1 g1 5.60 13.53 0.03 0.13 0.42 0.60 DT2 g2 3.42 41.88 0.02 0.39 0.37 0.79 ’’ g3 38.93 11.31 0.22 0.11 0.69 0.57 ’’ g4 5.75 12.47 0.03 0.12 0.42 0.58 DT3 g5 54.33 5.76 0.31 0.05 0.74 0.48 DT4 g6 13.24 17.56 0.07 0.16 0.52 0.64 DT5 g7 55.53 4.17 0.31 0.04 0.75 0.44 Total 176.80 106.68 Table 4 Electronic health record of patient 1. Document Creation date Document type 1 01/08/2010 DT1 2 07/03/2008 DT1 3 31/07/2010 DT5 4 02/08/2010 DT3 5 01/08/2010 DT2 6 15/06/2010 DT2 7 15/06/2010 DT4 Tab Per Tab Per B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 8531 the post-surgery review but know he is feeling really bad (much pain in the breast). He comes to the Emergency service and a res- ident doctor in the fifth year (R5) is going to check up him. Now they are in the emergency room and the doctor wants to access the patient’s EHR. The first stage of the approach is to identify the access context. According to the information about the attributes used to identify the context, the first rule that is satisfied is: IF Speciality = Cardiology AND Last_visit=‘‘cateterismo post-surgery review’’ AND Unit=‘‘Emergency’’ THEN Context=‘‘Emergency after cateterismo sur- gery’’ (C1) The electronic health record for this patient is in Table 4. Let us note that the EHR has several documents and some of them are of le 5 tinences for patient 1’s EHR for context 1. Doc. Date Document type Data group PC1DC 1 01/08/10 DT1 g1 0.00 3 31/07/10 DT5 g7 0.00 4 02/08/10 DT3 g5 0.00 5 01/08/10 DT2 g2 0.00 ’’ ’’ ’’ g3 0.00 ’’ ’’ ’’ g4 0.00 7 15/06/10 DT4 g6 0.00 le 6 tinences for patient 1’s EHR for context 2. Doc. Date Document types Data group PC2DC 1 01/08/10 DT1 g1 0.00 3 31/07/10 DT5 g7 0.00 4 02/08/10 DT3 g5 0.00 5 01/08/10 DT2 g2 0.00 ’’ ’’ ’’ g3 0.00 ’’ ’’ ’’ g4 0.00 7 15/06/10 DT4 g6 0.00 the same type but with different age. We now consider an access from each of the two contexts mentioned above, and show the per- tinent data groups in each case. The first step is always to keep just the newer document of each type. Hence, the documents selected to work with are 1, 3, 4, 5, and 7, whereas documents 2 and 6 are discarded. With this subset of documents we calculate the different perti- nences for each of the data groups to the context C1. Table 5 shows the values for each type and the global pertinence in the column PC1G , applying the Definition 3. The next step is to order the data groups according to the global pertinence values. The last column of the table collects the ranking. If we make blocks of 4 data groups to create levels of preference, the result for the query will include the data groups Doc3.g7, Doc4.g5, Doc5.g3, and Doc7.g6. Whereas, Doc5.g4, Doc1.g1 would be in the next block of data groups. The number of considered data groups (size of the blocks) may change according, as an example, to the capacities of the medical workstations (in a PDA four data groups may be enough because of limitation of the screen but in a PC the number could be greater). Before returning this set to the user, we refine the answer to determine if in some concrete case it is preferable to show the en- tire document instead of its pertinent data groups. We do it if more than a given percentage of the data groups inside a document are selected in the answer. In such case we consider the rank of the most pertinent data group. As an example, if the percentage is set to the 60%, the documents Doc3, Doc4, Doc7, that only contain one data group and it has been found pertinent, will replace in the solution the pertinent data group. However Doc5 has three data groups from which only one has been found pertinent. Therefore, in the final solution this data group will be kept. This way, the ini- tial solution {Doc3.g7, Doc4.g5, Doc5.g3, Doc7.g6} will be replaced by {Doc3, Doc4, Doc5.g3, Doc7}. Finally we would like to remark that in the real system we establish several preference levels, by grouping the pertinent data groups into sets of a fixed size, 10. If the information needed does not appear within the first block of 10 data groups, the next set is shown; and this process is repeated until the information is found. In addition, in any moment the user can access the complete EHR by the traditional navigation system. To show the differences lets suppose that the appointment is in a different situation and the inferred context is C2. In that case the pertinences are different, as shown in Table 6. The data groups PPDC PT PC1C P C1 G Rank 0.00 0.79 0.42 0.42 6 0.00 0.92 0.75 0.75 1 0.00 0.92 0.74 0.74 2 0.00 0.79 0.37 0.37 7 0.00 0.79 0.69 0.69 3 0.00 0.79 0.42 0.42 5 0.00 0.75 0.52 0.52 4 PPDC PT PC2C P C2 G Rank 0.00 0.42 0.60 0.60 3 0.00 0.75 0.44 0.44 7 0.00 0.74 0.48 0.48 6 0.00 0.37 0.79 0.79 1 0.00 0.69 0.57 0.57 5 0.00 0.42 0.58 0.58 4 0.00 0.52 0.64 0.64 2 Table 7 Electronic health record for patient 2. Document Creation date Document type 1 01/09/2010 DT4 2 01/09/2010 DT1 3 25/08/2010 DT2 4 01/07/2010 DT3 5 15/06/2009 DT2 6 20/06/2009 DT5 Table 10 Updated frequency table for access 1. Document type Data group Acum PC1C DT1 g1 5.6 0.42 DT2 g2 3.42 0.37 ’’ g3 38.93 0.68 ’’ g4 5.75 0.42 DT3 g5 54.33 0.74 DT4 g6 13.24 0.52 DT5 g7 62.41 0.76 Total 183.68 8532 B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 pertinent for this answer would be the set {Doc5.g2, Doc7.g6, - Doc1.g1, Doc5.g4}. In that case 2 of the three data groups in docu- ment 5 are pertinent. Therefore, the refined set would be {Doc5, Doc7, Doc1.g1}. Table 11 Updated frequency table for access 2. Document types Data group Acum PC2C DT1 g1 20.41 0.65 DT2 g2 41.88 0.78 ’’ g3 11.31 0.56 ’’ g4 12.47 0.58 DT3 g5 5.76 0.47 DT4 g6 17.56 0.63 DT5 g7 4.17 0.44 Total 113.57 6.1.2. Patient 2 This patient has a fracture in a rib that needs surgery. Most of the preparatives have been done (pre-surgery analysis as blood test and cardiogram) and now he is at the doctor office to check that everything is right for the surgery. The doctor is a trauma fac- ultative in his/her office accessing the EHR for the analisis results. The EHR for this patient is shown in Table 7. The patient has a chronic disease that makes his defenses be very low all the time. In this case we consider that by medical recommendation, one type of document groups (the defenses analysis inside the blood test) is specially important for that patient: the data group g3 in document type DT6. Hence, for this type we consider the static pertinence gi- ven by doctor to this document type for this patient PPDC � � with the value PPDC ¼ 0:7 that means this type of documents is especially important for this patient independently from the access context. As in the previous case, we identify the context using the rule base stored in the system. The first rule that is satisfied in this case is the next: IF Speciality = Trauma AND Present_visit=‘‘pre-surgery’’ THEN Context=‘‘trauma pre-surgery review’’ (C2) Under this assumption, an access from context C2 will result in the set of pertinences shown in Table 8. As in the previous exam- ple, we have first selected only one document of each type, consid- ering the newer one if there are several for a type. Table 8 Pertinences of patient 2’s EHR for context 2. Doc. Date Document type Data group PC2DC P P DC PT PC2C P C2 G Rank 1 01/09/10 DT4 g6 0.00 0.00 1.00 0.64 0.64 3 2 01/09/10 DT1 g1 0.00 0.00 1.00 0.60 0.60 4 3 25/08/10 DT2 g2 0.00 0.00 0.79 0.79 0.79 1 ’’ ’’ ’’ g3 0.00 0.70 0.79 0.57 0.70 2 ’’ ’’ ’’ g4 0.00 0.00 0.79 0.58 0.58 5 4 01/07/10 DT3 g5 0.00 0.00 0.88 0.48 0.48 6 6 20/06/09 DT5 g7 0.00 0.00 0.73 0.44 0.44 7 Table 9 Pertinences for patient 2’s EHR for context 1. Doc. Date Document type Data group PC1DC P P DC PT PC1C P C1 G Rank 1 01/09/10 DT4 g6 0.00 0.00 1.00 0.52 0.52 4 2 01/09/10 DT1 g1 0.00 0.00 1.00 0.42 0.42 5 3 25/08/10 DT2 g2 0.00 0.00 0.79 0.37 0.37 7 ’’ ’’ ’’ g3 0.00 0.70 0.79 0.69 0.70 3 ’’ ’’ ’’ g4 0.00 0.00 0.79 0.42 0.42 6 4 01/07/10 DT3 g5 0.00 0.00 0.88 0.74 0.74 1 6 20/06/09 DT5 g7 0.00 0.00 0.73 0.75 0.73 2 In this access, as Table 8 shows, the pertinent data groups found are {Doc6.g10, Doc3.g2, Doc3.g3, Doc1.g6, Doc2.g1}. If we refine the answer considering the complete documents, the returned set would be {Doc3, Doc1, Doc2}. In this answer the static pertinence has been very important. Without this information the final perti- nence for Doc3.g3 would be 0.57 and only the data group g2 would have been returned. If the access to patient 2’s EHR would be done from context C1, the result would be different. Table 9 collects the pertinence val- ues. Following the same process as in previous example, the an- swer of the system would be {Doc4, Doc6, Doc3.g3, Doc1}. 6.2. Example of update In the previous sections we have shown the answer of the system when a doctor accesses an EHR. Once the doctor selects a B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 8533 document the system updates automatically the dynamic perti- nences. In this section we show two examples of these updates. First, suppose that the system returns a set of documents and the user selects one of these documents. Considering the patient 1 and context C1, let’s assume the Doc1.g2 is chosen, on 27th of September of 2010. The frequency table is updated and all the pert- inences for this context are updated too. According to the date, the weight W(D) for Doc1.g2 is 6.68 and the new frequency table (with the corresponding dynamic pertinences) is shown in Table 10. The values that have changed are in italic. In this case two pertinences have changed and they will be taken into account for the future accesses. If the user selects a document that is not in the data set, for example Doc2.g1 in C2, the process is similar and the result is shown in Table 11. In this case the pertinences for five of the data groups change too. The pertinence of this type of document (Doc2.g1) has increased to reflect the access to this document and the other four has been decrease. 7. Implementation As we have mentioned before, in this kind of system the effi- ciency is very important. In this section we briefly comment the implementation for our proposal and we analyze the efficiency of the final system. Next we explain the implementation of the two most relevant elements in the system, the rule base used to identify the context and the frequency table to select its pertinent documents, as well as the processes needed to access them. 7.1. Rule base Each time a hospital’s staff member accesses the EHR sys- tem, the first step is to identify the context of the access. As mentioned in Section 3.2, the rule base and the process to se- lect the context have been implemented inside the data base so with just a single, simple and fast query the answer to this question is obtained. The rule base is implemented using one table inside the data- base. Each record represents a rule, and we store the order learnt by RIPPER, the values for each attribute presented in Section 3.2, and the context. If one rule has no value for an attribute then a NULL value is stored. Therefore the table has 9 columns (one for the rule order, seven for the attributes and one for the context) and near two hundreds records. When an access occurs, the context is identified by scanning the table looking for the first rule satisfied. The SQL sentence that gets the context in this way uses the 7 attributes of the access to build a select clause for Oracle DBMS as follows: SELECT context FROM RuleBase WHERE (Speciality=‘‘value1’’ or Speciality IS NULL) and (position=‘‘value2’’ or position IS NULL) and. . . and (last_visit=‘‘value7’’ or last_visit IS NULL) ORDER BY ord ASC HAVING ROWNUM<=1 The efficiency of the query is improved defining indexes on the table over the seven attributes. Therefore the time needed to an- swer the query is very small. 7.2. Frequency table The other table needed is the frequency table, used to know the pertinent documents for each context. This table is stored in the database too, with the following attributes: � DG: internal code for each data group type. � Context: Context � Weight: Total weight for all the accesses to this data group type (DC) in this context (element tfXC in Eq. (4)). � RP, DcP: the static pertinences (established by regulations and doctors) for the data group to the particular context respectively. The primary key is (DG, Context). In the DG domain we include a new code (�1) that do not represent any data group type. This va- lue will be used to store the aggregation of all the weights for all the data group types in a particular context. This value is the ele- ment tfC in Eq. (4). The number of records stored in the table is jcontextsj� (jDGj + 1). To improve the queries efficiency we define indexes over both attributes individually and together (3 indexes). To improve the query over the table we define clusters on the table (physical blocks to store related records) considering the context value (all the records storing the frequency of each data group type in a par- ticular context are stored physically together). 7.2.1. Query for the pertinent data groups To know the frequency of the five most pertinent data group types to a concrete context we only need to execute the following select sentence: SELECT DG, Weight FROM FrequencyTable WHERE context = C ORDER BY Weight DESC HAVING ROWNUM<=6 The first row will give the value of tfC and the five next rows will give the values tfXC for the five most pertinent data group types. Let us note that the query is very simple and needs a very low execu- tion time. The five most relevant data groups for a concrete patient in a particular context are obtained using the value of the function PCGðXÞ and ordering the data groups according to its value. To speed up the process we have implemented the function inside the data- base using PL/SQL (the Oracle language for stored functions and procedures) as PGC. With these values the query has to join the fre- quency table and the EHR table, execute the function and order the data groups. Let be pid the ID for the patient, C the related context, TOTAL the value of tfC, and RP, DcP and PP the static pertinence (reg- ulations, doctor and patients respectively) of this context for this particular patient. Then the select query has the following structure: SELECT ⁄, PGC (FrequencyTable.DcP,FrequencyTable.RP,DPP, FrequencyTable.Weight,TOTAL,EHR.Date) as Pertinence FROM FrequencyTable,EHR WHERE EHR.PID = pid and FrequencyTable.Context = C and /⁄ JOIN ⁄/ EHR.DG = FrequencyTable.DG /⁄ Order ⁄/ ORDER BY Pertinence DESC HAVING ROWNUM<=5 With one query to know the values of tfC (very simple) and this second query we get the pertinent data groups. All the computa- tion is made inside the database server so this process does not introduce any computation overload on the terminal. 7.2.2. Update of the frequency table There are two processes that change the frequency table: 8534 B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 � Each time a data group is acceded (very often). � Each time the reference date changes (once a year) the frequen- cies stored are updated. For the first update is performed by executing just one sentence. Let be id the code for the data group, C the context and W(A) the weight for this new access. Then the update sentence is: UPDATE FrequencyTable SET Weight = Weight + W(A) WHERE DG in (id,�1) AND Context = C; As can be seen this sentence is simple a very fast to execute since only two records are modified: the one related to the data group and the value of tfC. The second update will be done once a year (at most) and it will imply to modify all the records in Frequency table. As we have men- tioned in Section 4.3, changing the reference date by one year the update sentence would be: UPDATE FrequencyTable SET Weight = Weight/2; Though this update is more time consuming than the previous one, it is assumable since it will only be executed once a year and just blocks one table for a very simple processing. 8. Results To test our proposal we have designed the test explained next. For each context we have selected a set with the most frequent assistance acts. Then we have monitored these access for a month to get the data groups accessed on each case. Then we have performed the accesses using the new system and we have counted the number of times that all the information needed was present within the first selection of pertinent data groups, the number of occasions in which it has been needed to access the second set of pertinent data groups, the same with the third set, and so on. In the 81.56% of the accesses all the information requested was in the first set, in the 4.12% of the cases, the needs of information where satisfied with the second set of data groups, the 1.91% of the accesses needed to reach a set of data groups in the third, and finally in the remaining 12.41% of the cases it was necessary to reach at or over the fourth block. To compare the performance of the proposed system regarding the previous one, we have compared the number of ‘‘clicks’’ re- quired to obtain all the information needed in both systems (we compare the number of clicks and not the time, because in the navigation based systems the time depends on the skills of the user). To do it we have considered that each access to a set of data groups requires one click in the new system. In the old system, we have considered that each document acceded needed one click (to open the document) and that the search of each doc- ument at least needed three clicks (to select the type of assistance act, the concrete assistance act and the set of documents gener- ated on it, as shown in Fig. 2). According to it, and taking as an example an assistance act that needs data groups of information contained in five different documents, with the traditional navi- gation systems 20 clicks would be needed; whereas with the new one, according to the previous percentages, in the 81.56% of the cases only one click is needed. Just with this data it can be appreciated the great gain of time that the proposed system offers. Some possible improvements that can be made in the system are the next: � Adapt the size of the sets of data groups showed in each priority level, according to the context. This way the contexts like a pre- surgery study, that usually need to access more than 10 data groups, would show in each level sets of 20 or 30 data groups; whereas other contexts could keep smaller sets, like a simple outpatient consultation. It would only require to add a column to the table of contexts. � Considering that in Section 3.2 we mentioned that the context identification had a percentage of success of the 89.32% it is possible that some of the accesses over the third set of pertinent documents were due to a bad identification of the context. Con- sidering that the training was made with only a participation of the 24%, improving the training with a higher participation the identification of the context could be better. Finally, we must remark that the system is still on its develop- ment and improvement stage, so it is not still completely im- planted. Anyway the doctors that have tested it have commented that it is ‘‘very useful’’, ‘‘quite comfortable’’ and, on top of it, ‘‘really very time saving’’. 9. Conclusions In this paper we have proposed a new paradigm to access to the EHR systems, based on a contextualized access to the information, in such a way that the information is ordered according to their preference or relevance for the assistance act or pathologic process in which the medical staff in involved. It is therefore specifically thought to improve the availability of information for medical practice, satisfying their specific needs and offering with it a faster and more efficient access to the really relevant information for a concrete assistance act, avoiding the handling of a huge quantity of information superfluous or unnecessary for it. We have also presented a method defined the contexts and also to identify them each time a doctor accesses the system, based on a rule system stored in a data base. In addition we have proposed a technique to efficiently define the pertinence of data groups to the context and update it on each new access, so the system can auto- matically adapt to changes on the access patterns and to the new doctor’s needs of information. In addition, we have shown some details of the implementation as well as options to improve the accessibility. We have also showed some examples of how the system works and the results obtained are quite encouraging as the statistics obtained in the test of the system remak, as well as the satisfied opinion of the medical staff that have used it. With this proposal several new research lines have been opened, since we enable and make easier to provide the system with new capabilities. This is the case of the knowledge mobiliza- tion, where the contextualized access solves the limitation in the use of mobile devices to perform complex accesses to great vol- umes of information, since with this access the few data really needed for a concrete assistance act are easily available and visual- ized in this type of devices, being unnecessary the navigation through the EHR. Another example of this advantage is the possibility of provid- ing the system with new functionalities for research purposes (El Fadly et al., 2010), just defining the corresponding contexts. In addition it has open the possibility to provide the citizens with a personalized access to their own medical data which, as Charters (2009) and Ruland, Brynhi, Andersen, and Bryhni (2008) indicate, it is a growing demand. Moreover, this new access paradigm can be applied to enable the interoperability between health record systems, by using the contexts to define the archetypes that the ISO 13606 requires. B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 8535 Finally, we would like to remark the viability of this proposal, that is proved with the actual programming line in the San Cecilio University Hospital from Granada. Acknowledgment The research reported in this paper was partially supported by the Andalusian Government (Junta de Andalucía) under project P07-TIC03175 ‘‘Representacin y Manipulación de Objetos Imper- fectos en Problemas de Integracin de Datos: Una Aplicación a los Almacenes de Objetos de Aprendizaje’’ and also by the Spanish Government (Science and Innovation Department) under project TIN2009-08296. We would also like to thank their collaboration to the medical personnel that is participating in the development of the system. References Adams, A., Adams, R., Thorogood, M., & Buckingham, C. (2007). Barriers to the use of e-health technology in nurse practitioner-patient consultations. Informatics in Primary Care, 15(2), 103–109. Bisbal, J., & Berry, D. (2009). An analysis framework for electronic health record systems. Methods of Information in Medicine, 48(1). Bobillo, F., Delgado, M., & Gómez-Romero, J. (2008). Representation of context- dependant knowledge in ontologies: A model and an application. Expert Systems with Applications, 35, 1899–1908<http://portal.acm.org/citation. cfm?id= 1401266.1401463>. Bohm, K., Wolf, P., & Krcmar, H. (2010). Context oriented structuring of egovernment services – An empirical analysis of the information demand of expatriates in germany. In 43rd Hawaii International Conference on System Sciences (HICSS), 2010 (pp. 1–10). Buchholz, T., Hochstatter, I., & Linnhoff-Popien, C. (2007). Distribution strategies for the contextualized mobile internet. Electronic Commerce Research and Applications, 6(1), 40–52<http://www.sciencedirect.com/science/article/ B6X4K-4K5HV8M-1/2/90c68f322aa155b9f62aaafa2917c25b>. Cao, Y., Xu, J., Liu, T.-Y., Li, H., Huang, Y., & Hon, H.-W. (2006). Adapting ranking svm to document retrieval. In SIGIR’06: Proceedings of the 29th annual international ACM SIGIR conference on research and development in information retrieval (pp. 186–193). New York, NY, USA: ACM. Cayir, S., & Nuri Basoglu, A. (2008). Information technology interoperability awareness: A taxonomy model based on information requirements and business needs. In Portland International Conference on Management of Engineering Technology, 2008, PICMET 2008. (pp. 846 –855). Chaker, H., Chevalier, M., Soule-Dupuy, C., & Tricot, A. (2010). Improving information retrieval by modelling business context. In Third international conference on advances in human-oriented and personalized mechanisms, technologies and services (CENTRIC), 2010 (pp. 117–122). Charters, K. (2009). Challenges of electronic medical record extracts for a personal health record. Studies in Health Technology and Informatics, 146, 197–201. Cho, I., Kim, J., Kim, J., Kim, H., & Kim, Y. (2010). Design and implementation of a standards-based interoperable clinical decision support architecture in the context of the korean ehr. International Journal of Medical Informatics, 79(9), 611–622. Cho, I., Staggers, N., & Park, I. (2010). Nurses’ responses to differing amounts and information content in a diagnostic computer-based decision support application. Computers, Informatics, Nursing: CIN, 28(2), 95–102. Chu, W., Johnson, D., & Kangarloo, H. (2000). A medical digital library to support scenario and user-tailored information retrieval. IEEE Transactions on Information Technology in Biomedicine, 4(2), 97–107. Cohen, W. W. (1995). Fast effective rule induction. In Proceedings of the twelfth international conference on machine learning (pp. 115–123). Morgan Kaufmann. Collins, B., & Speedie, S. (2008). Attaching context sensitive infobuttons to an ehr options and issues. AMIA. In Annual symposium proceedings/AMIA symposium AMIA Symposium 914. Data exchange standard (2011). <http://www.hl7.org>. http://<www.HL7.org>. Del Fiol, G., & Haug, P. (2009). Classification models for the prediction of clinicians’ information needs. Journal of Biomedical Informatics, 42(1), 82–89. Ehsan, M., Amini, M., & Jalili, R. (2009). Handling context in a semantic-based access control framework. In International conference on advanced information networking and applications workshops, 2009, WAINA’09 (pp. 103 –108). El Fadly, A., Lucas, N., Rance, B., Verplancke, P., Lastic, P.-Y., & Daniel, C. (2010). The reuse project: Ehr as single datasource for biomedical research. Studies in Health Technology and Informatics, 160(Pt. 2), 1324–1328<http://www.biomedsearch. com/nih/REUSE-project-EHR-as-single/20841899.html>. Erdal, S., Catalyurek, U., Payne, P., Saltz, J., Kamal, J., & Gurcan, M. (2009). A knowledge-anchored integrative image search and retrieval system. Journal of Digital Imaging: the Official Journal of the Society for Computer Applications in Radiology, 22(2), 166–182. Flores Zuniga, A., Win, K., & Susilo, W. (2010). Functionalities of free and open electronic health record systems. International Journal of Technology Assessment in Health Care, 26(4), 382–389. Gagnon, M., Desmartis, M., Labrecque, M., Lgar, F., Lamothe, L., Fortin, J., et al. (2010). Implementation of an electronic medical record in family practice: A case study. Informatics in Primary Care, 18(1), 31–40. Garcia-Morchon, O., & Wehrle, K. (2010). Efficient and context-aware access control for pervasive medical sensor networks. In 8th IEEE international conference on pervasive computing and communications workshops 2010 (PERCOM Workshops) (pp. 322–327). Gil-Leyva, I., & Rodríguez-Mu ~noz, J. (1966). Tendencias en los sistemas de indizacin automtica. estudio evolutivo. Revista Espaola de Documentacion Cientffica, 19(3), 273–291. Ginsburg, M. (2007). Pediatric electronic health record interface design: The pedone system. In 40th annual hawaii international conference on system sciences, 2007, HICSS 2007 (pp. 139). Image storage standard. (2011). <http://www.medical.nema.org>. ISO-13606 (2008). Iso 13606: Electronic health record communication. Järvelin, K., & Kekäläinen, J. (2000). Ir evaluation methods for retrieving highly relevant documents. In SIGIR’00: Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval (pp. 41–48). New York, NY, USA: ACM. Jerding, D., & Stasko, J. (1998). The information mural: a technique for displaying and navigating large information spaces. IEEE Transactions on Visualization and Computer Graphics, 4(3), 257–271. Jones, K. S. (2004). A statistical interpretation of term specificity and its application in retrieval. Journal of Documentation, 60, 493–502. Judd, G., & Steenkiste, P. (2003). Providing contextual information to pervasive computing applications. In Proceedings of the First IEEE international conference on pervasive computing and communications, 2003. (PerCom 2003) (pp. 133–142). doi:10.1109/PERCOM.2003.1192735. Jung, J. J. (2009). Contextualized query sampling to discover semantic resource descriptions on the web. Information Processing & Management, 45(2), 280–287<http://www.sciencedirect.com/science/article/B6VC8-4V8FFJY-/2/ 6f9094e73f868971acee4f7cfb78f225>. Kang, D., Lee, H., Ko, E., Kang, K., & Lee, J. (2006). A wearable context aware system for ubiquitous healthcare. In Conference proceedings: Annual international conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in medicine and biology society conference (Vol. 1, pp. 5192–5195). Kanoulas, E., Pavlu, V., Dai, K., & Aslam, J. (2010). Modeling the score distributions of relevant and non-relevant documents. Advances in Information Retrieval Theory, 152–163. Karahoca, A., Bayraktar, E., Tatoglu, E., & Karahoca, D. (2010). Information system design for a hospital emergency department: A usability analysis of software prototypes. Journal of Biomedical Informatics, 43(2), 224–232. Lahteenmaki, H., Leppanen, J., & Kaijanranta, J. (2009). Interoperability of personal health records. In Conference proceedings: Annual international conference of the IEEE Engineering in Medicine and Biology Society. IEEE engineering in medicine and biology society conference (p. 1726-9). Mao, J.-Y., & Benbasat, I. (2001). The effects of contextualized access to knowledge on judgement. International Journal of Human–Computer Studies, 55(5), 787–814<http://www.sciencedirect.com/science/article/B6WGR-4582BWS-7/ 2/cb7b598f793b5017ca950d7e785f369c>. McAlearney, A., Robbins, J., Hirsch, A., Jorina, M., & Harrop, J. (2010). Perceived efficiency impacts following electronic health record implementation: An exploratory study of an urban community health center network. International Journal of Medical Informatics, 79(12), 807–816. Nagy, M., Hanzlfcek, P., Preckov, P., Rfha, A., Dioszegi, M., Seidl, L., & Zvrov, J. (2010). Semantic interoperability in czech healthcare environment supported by hl7 version 3. Methods of Information in Medicine, 49(2), 186–195<http:// www.ncbi.nlm.nih.gov/pubmed/19936441>. Nagy, M., Preckova, P., Seidl, L., & Zvarova, J. (2010b). Challenges of interoperability using hl7 v3 in czech healthcare. Studies in Health Technology and Informatics, 155, 122–128. Nomenclature medical standard, S., 2011. <http://www.nlm.nih.gov/research/ snomad/snomad_main.html>. Open-EHR open electronical health records, 2011. <http://www.openehr.org>. <www.openehr.org>. Prados, M., & Peña, M. (2003). Sistemas de Informacion hospitalarios. Organizacion y gestion de Proyectos. EASP. Prados de Reyes, M., Carmen Peña Yáñez, M. A. V. M., & Suárez, M. B. P. (2006). Generation and use of one ontology for intelligent information retrieval from electronic record histories. Prados, M., Peña, M., Prados, B., Martinez, B., Ortigosa, J., & Delgado, A. (2010). Electronical health record (ehr) representation through ontology, mobility, accesibility and interoperability usefulness. Prados-Suárez, B., Revuelta, E., Peña Yañez, C., Molina Fernàndez, C. (2008). Ontology based semantic representation of the reports and results in a hospital information system. In Proceedings of the ICEIS 2008 (pp. 300–306). Pung, H. K., Gu, T., Xue, W., Palmes, P. P., Zhu, J., Ng, W. L., Tang, C. W., & Chung, N. H. (2009). Context-aware middleware for pervasive elderly homecare. IEEE Journal on Selected Areas in communications, Institute of Electrical and Electronics Engineers Inc., The 27, pp. 510–524. http://portal.acm.org/citation.cfm?id=1401266.1401463 http://portal.acm.org/citation.cfm?id=1401266.1401463 http://www.sciencedirect.com http://www.sciencedirect.com http://www.hl7.org http://www.HL7.org http://www.biomedsearch.com/nih/REUSE-project-EHR-as-single/20841899.html http://www.biomedsearch.com/nih/REUSE-project-EHR-as-single/20841899.html http://www.medical.nema.org http://dx.doi.org/10.1109/PERCOM.2003.1192735 http://www.sciencedirect.com http://www.sciencedirect.com http://www.sciencedirect.com http://www.sciencedirect.com http://www.ncbi.nlm.nih.gov/pubmed/19936441 http://www.ncbi.nlm.nih.gov/pubmed/19936441 http://www.nlm.nih.gov/research/snomad/snomad_main.html http://www.nlm.nih.gov/research/snomad/snomad_main.html http://www.openehr.org http://www.openehr.org 8536 B. Prados-Suárez et al. / Expert Systems with Applications 39 (2012) 8522–8536 Ray, J. & Wimalasiri, P. (2006). The need for technical solutions for maintaining the privacy of ehr. In Conference proceedings: Annual international conference of the IEEE Engineering in Medicine and Biology Society. IEEE engineering in medicine and biology society conference (Vol. 1, pp. 4686–4689). Ross, S. (2009). Results of a survey of an online physician community regarding use of electronic medical records in office practices. The Journal of Medical Practice Management: MPM, 24(4), 254–256. Ruland, C., Brynhi, H., Andersen, R., & Bryhni, T. (2008). Developing a shared electronic health record for patients and clinicians. Studies in Health Technology and Informatics, 136, 57–62. Salton, G., Wong, A., & Yang, C. S. (1975). A vector space model for automatic indexing. Communications of the ACM, 18(11), 613–620. Svanaes, D., Das, A., & Alsos, O. (2008). The contextual nature of usability and its relevance to medical informatics. Studies in Health Technology and Informatics, 136, 541–546. Vest, J., & Jasperson, J. (2010). What should we measure? Conceptualizing usage in health information exchange. Journal of the American Medical Informatics Association: JAMIA, 17(3), 302–307. Vishwanath, A., Singh, S., & Winkelstein, P. (2010). The impact of electronic medical record systems on outpatient workflows: A longitudinal evaluation of its workflow effects. International Journal of Medical Informatics, 79(11), 778–791. Weinstock, M. (2010). For hospitals and meaningful use, context is everything. Hospitals & Health Networks AHA, 84(8), 20–21. Improving electronic health records retrieval using contexts 1 Introduction 2 Background 2.1 Electronic Health Records structure 2.2 Data groups 2.3 EHR information system 3 Contexts 3.1 Context definition 3.2 Context identification 4 Pertinence 4.1 Static pertinence: regulations, doctors and patients 4.2 Time pertinence 4.3 Dynamic pertinence 4.4 Global pertinence 5 Contextualized access system 5.1 Access to the system 5.2 Update process 6 Example 6.1 Examples of retrieval 6.1.1 Patient 1 6.1.2 Patient 2 6.2 Example of update 7 Implementation 7.1 Rule base 7.2 Frequency table 7.2.1 Query for the pertinent data groups 7.2.2 Update of the frequency table 8 Results 9 Conclusions Acknowledgment References 
work_y7uq3m2zqzacnd4pf2fdihpnqu	----	I ' f - 9 6 b 0 625d C W F - Yb 0 769- 7 / AN EXPERT SYSTEM FRAMEWORK FOR NONDESTRUCIWE WASTE ASSAY G.K. Becker Idaho National Engineering Laboratory Lockheed Martin Idaho Technologies Company Idaho Falls, Idaho USA ABsrRAcr The management and disposition of transuranic (TRU) waste forms necessitates a determination of entrained TRU and associated radioactive material quantities as per National T R U Waste Characteriza- tion Program requirements. The technical justifica- tion and demonstration of a given nondestructive assay (NDA) method used to determine TRU mass and uncertainty in accordance with Program quality assurance objectives is a difficult task for many waste forms. Difficulties are typically founded in waste NDA methods that employ standards compensation techniques and/or the employment of simplifying assumptions regarding waste form configurations. The capability to determine and justify TRU mass and mass uncertainty can be enhanced through appropriate integration of waste container data/ information using expert system and empirical data- driven techniques in conjunction with conventional data acquisition and analysis methods. Presented is a preliminary expert system framework that integrates the waste form data base, algorithmic techniques, statistical analyses, expert domain knowledge bases, and empirical artificial intelligence modules into a cohesive system. T h e framework design and bases in addition to module development activities are discussed. BACKGROUND The ability t o manage and disposition TRU waste is predicated o n demonstrating compliance with appli- cable characterization requirements. Non-destructive waste assay methods are used in the National TRU Waste Characterization Program t o determine the mas of waste-entrained radionuclides. The capabil- ity and performance of waste NDA systems employed in the Program must comply with requirements as set forth in the TRU Waste Characterization Program Quality Assurance Program Plan (QAPP).' Parameters for which waste NDA system perform- ance must be demonstrated and technically justified include total bias and total uncertainty as defined per the QAPP. Within the instrument response envelope of commonly employed NDA techniques, certain standard compensation and bias quantifica- tion techniques can be derived to account for a number of prominent measurement bias sources. These techniques rely o n the knowledge of the bias and precision elements and the fact that there is a unique instrument response relation for each particular bias element. In reality, the functional form of a number of bias sources and their interrelationships are not necessarily known and understood for many waste forms. Additionally, each waste NDA technique has limitations where the response is non-unique, interfered with, or there is no measure of a particular waste form attribute from which bias arises. An example of an assay modality and waste form attribute where there is no measure is active thermal neutron interrogation and ffisile material clumping. This commonr& employed measurement modality has n o viable response indicator of the fissile material configuration for which an account can be made of the induced bias. Hence, alternate approaches to the interpretation/ integration of NDA data and supportive information enhancing the characterization capability base is of interest. In consideration of Program quality assurance objectives for total bias and total uncertainty and the reality of difficult-to-account-for waste form bias sources, it is logical to look to other mechanisms for acquiring and processing data/information to a m v e at a defensible assay solution. T h e solution may reside in the acquisition of additional waste form configuration information and/or deducing that one or more different measurement modalities would yield necessary data. Regardless of the approach, additional or reformatted information will in all probability be of value in completing the solution in DISCLAIMER Portions of this document may be illegible in electronic image products. Images are produced from the best available original document. a defensible manner. Another view to take in lieu of additional data/information is to ask if all presently available information and data confirm that the waste form configuration attributes are within the capability bounds of the employed NDA technique. This perspective or concept is most applicable in those cases for which alternate measurement modes and/or information are not available. In either case, consideration of the assay routine is required to EXPERT SYSTEM OVERVIEW either substantiate the implemented technique o r acquire and utilize additional datalinformation to Expert systems are computer programs that emulate justify and demonstrate that a solution with known the reasoning process of a human expert using uncertainty bounds has been achieved. previously established rules for a well-defined domain. They combine knowledge bases of rules and For most waste characterization programs, it is not domain-specific facts with information representing feasible to maintain technical experts in waste NDA specific instances acquired from the environment of to analyze the instrument response acquired on each interest. Functionally expert systems infer and draw waste container to conclude whether the technique conclusions from available data and information. A is applicable or whether additional information is particularly powerful feature is the interactive needed. Therefore, the ability is desired to assess acquisition of new and relevant datahnformation via instrument response and all pertinent datal questions. The expert system can also tell and justify information in an on-line fashion yielding a tangible how a given conclusion has been reached via an assay confidence factor related to compliance explanation facility, potentially giving the user insight criteria. A technology ideally suited to this task is o n the problem at hand. Expert systems are modular that of expert systems technology. In accordance in structure, typically consisting of the following with the concept of evaluating available datal parts: (1) a knowledge base, (2) an inference engine, information, a n expert system can be configured to (3) a global o r working memory, and (4) a user classify measurement data and associated waste form interface. The knowledge base contains the expert configuration information as either consistent or domain knowledge for use in problem solving. The inconsistent with respect to all available knowledge. working memory is used as a scratch pad and to Consistency leads to confidence in parameters of store interim reasoning data and infomation importance, Le,, fissile material mass and alpha provided to the system from the environment of activity and a quantifiable indicator o f quality interest. The inference engine uses the domain control. Inconsistency can take many forms as a knowledge together with acquired information to function of which particular sets of information and provide an expert solution. High-level ruies are data provide confirmatory evidence and which do implemented in the inference engine to avoid blind not. Regardless, significantly more is known about searches of the solution space. The user interface is the assay process in that the measurement results the link to the outside world and provides an support all other knowledge and vice versa o r that explanation facility to illustrate to the user how a inconsistencies a r e present, prompting an investiga- particular decision was reached. tion into technique applicability. There are many ways to represent knowledge in an An expert system can b e constructed with capabilities expert system. The three most popular representa- expanded beyond basic consistency testing, thereby tion schemes are rules, semantic networks, and employing additionaI evaluation tools and data from frames. The rule scheme is the primary method used other measurement modalities. Relevant analysis in the subject system. Rules a r e used to build a resources include exploratory statistical routines and domain-specific knowledge base represented via empirical artificial intelligence techniques such as domain facts and heuristics that specify a set of neural networks and fuzzy logic, hypothesis testing actions to be performed for a given situation. A rule twb, adaptive system learning routines, and waste is composed of a n antecedent and a consequent of form generation information bases. An expanded the general form: IF antecedent THEN consequent. expert system of this nature will necessarily be The antecedent of a ruIe is a set of conditions (or derived from experience gained with the construction conditional elements) that must be satisfied for the testing and validation of a system based on the rule to be applicable. The consequent of a rule is simpler consistency verification concept. interim goal of the data analysis component of the Waste Assay Measurement Information System project2 is to develop an initial phase expert system that addresses the concept of assay consistency with quantifiable compliance capabilities. the set of actions to be executed when the rule is applicable. Confidence measures are associated with the various rules to capture the probabalistic nature of the rules often called certainties. Several methods (such as tables of probability-related information) exist to define and represent rule certainties. An important advantage of expert systems is the ease with which knowledge bases can be modified as new rules and facts become known. An inference mechanism, or engine, using logical reasoning is necessary to determine the most appropriate response when the knowledge base is consulted. T h e inference engine is the driver program for the expert system using the knowledge base t o reach a particular conclusion. It is responsible for ensuring that questioning is done in a concise, logical manner, determining when to search the knowledge base for the information and scheduling other necessary actions. It will take action as indicated via a knowledge state found to be true based on the current facts presented to the expert system. The inference engine is, in effect, the intelligence that allows the expert system to make conclusions based on the expertise or knowledge stored in the knowledge base. Inference engine reasoning methods primarily operate in o n e of two ways. They may be data- driven, known a s forward chaining, or they may work backwards from conclusions or hypotheses, known as backward chaining. A forward chaining system begins with problem domain input data and moves down the inference chain until a conclusion is reached. Goal-directed reasoning is termed backward chaining where the system starts with a hypothesis or goal and works backward through the facts until it reaches a final node or conclusion. The actual reasoning process consists of constructing new rules or sentences from existing ones and ehsuring that the new ones represent facts that actually follow from the facts of previous sentences represented. Hence, the inference procedure derives logical sentences from t h e knowledge base that represent facts, which follow from facts embodied in the knowledge base. NONDESTRUCTIVE WASTlE ASSAY FRAMEWORK Waste NDA has many features suitable for the application of expert system technology. The measurement data set acquired from a given waste container may or may not embody infomation sufficient to extract the parameters of importance and the associated uncertainty. Use of an expert system allows one to utilize available knowledge concerning a given waste container in addition to the capabilities of the applied measurement technique to subsrantiate a viable solution. For instance, waste form generation process knowledge can bound and/or yield probable radioactive material composi- tions, radioactive material age, radioactive material c h e m i c a l c o m p o u n d , p r o b a b l e e l e m e n t a l compositions, probable density distributions, waste form configuration, etc. Oftentimes, the generator has previously acquired assay data in the same o r differing packaging configurations, which can be evaluated for consistency. Data bases containing previous measurement records can be consulted to see if certain statistical parameters fit historical distributions. Algorithmic assessment routines may be implemented to determine and/or evaluate specific parameter values per the acquired data and associated knowledge. Empirical artificial intelli- gence (AI) techniques can be utilized to look for data patterns and features that should be or not be present. Statistical hypothesis testing can be performed at various points in the problem domain to support t h e decision-making process of the system. The system capability of adaptive learning based on cumulative system experience could be exploited to enhance: the knowledge of t h e system. Finally and most importantly, the expert knowledge base allows the implementation of domain expert rules based in theory as well as heuristics. This is considerably more information that can be used to define and bound a solution than can be determined from the response of an NDA instrument calibrated in accordance with a specific regimen. An illustration of a preliminary expert system framework delineating applicable knowledge and dara base components is shown in Figure 1. Each component is defined based o n a particular type of knowledge or data/information source. These com- ponents or modules support the reasoning processes residing in the expert domain knowledge base. It is notable that such a modular arrangement supports the documentation, validation, and modification of the various knowledge and data modules. NDAFRAMEWORK COMPONEN?s T h e following is a brief description of the function of each module in the preliminary expert system. The . .- - __..,__i_. learning computation NDA System Measurement Data INPUT FIGURE 1. PRELIMINARY WASTE NDA EXPERT S Y S T E M FRAMEWORK modules are defined based on a logical partitioning of the knowledge and data/information space perti- nent to the problem. The collective set of modules is intended to contain all available data, information, and knowledge a n NDA expert would utilize to produce a best estimate o n the fissile material mass and associated uncertainty for a given container. The structure of the data and information contained in the module set to be used by the expert system is that of an object-oriented database? Master Expert System - The master expert system oversees operation of the system as negotiated by the expert domain knowledge based expert system. For exampIe, the master expert system interacts with the user, handIes resuft conflicts, and manages the adaptive learning controller based on the accumula- tion of data and information in the system over time. A part of the adaptive learning component would be a statistical analysis routine that continually updates parameters and correlations of interest to the expert system. Non-Destructive Waste Assay Measurement Data (System Input) - This input data vector contains pertinent raw acquired NDA measurement data in addition to the reduced data and associated parameters of significance. It constitutes the funda- mental data source from which the forward chaining inference procedure starts. T h e input data set format is such that differing fields can be extracted by other moduIes for various operations supporting the expert system analysis, e.g., empirical AI routines looking for patterns. Data preprocessing can also be implemented a t this stage, providing a format more amenable to the various components of the expert system. Expert Domain Knowledge Base Module (Expert System) - This knowledge base contains the funda- mental principles of the waste NDA problem domain in the form of rules and heuristics. This module contains the conditionaI questions and search direction data. Waste Generation Process Knowledge Base (Expert System) - This knowledge base contains previously established waste form attributes by generation process. Such attributes are the waste form genera- tion method and the resultant characteristics of the waste form. The knowledge base will be consulted in terms of the certain attributes variable range. The expert domain module will, as appropriate, consult this knowledge base regarding expected radionuclide compositions, waste form configurations, etc. Characteristics and attributes of the waste form must be consistent with measured parameters. Statistical Module - T h e employment of statistical models in the context of the expert system can take the form of a classical, hypothesis-driven o r confirm- atory data analysis approach o r a data-driven, exploratory data analysis approach. Applied statisti- cal methods in this module range through descriptive statistics, statistical inference procedures, hypothesis testing, goodness-of-fit tests, regression and correlation analysis, analysis of variance techniques, principal component analysis, classification via cluster analysis techniques, conditional expectation evaluations, linear discriminant functions, etc., for parameters of interest per predefined populations. Algorithmic Processing Module - This module contains first-principle algorithms for the evaluation of specific parameters. Examples include a figure of merit for quantifying the alpha,n component, the comparison of measured parameters to those predicted by MCNP models, and the execution of bounding computations to support validity determin- ations. Empirical AI Module - Empirical data exploration and processing techniques generally classified as artificial intelligence techniques are addressed in this module. The focus of these techniques is to perform classification functions using pattern recognition techniques, supervised and unsupervised learning and clustering techniques, and data-driven empirical methods. Examples include the class of neural networks, learning vector quantization, genetic algorithms, K-Nearest Neighbor regression, adaptive kernel methodsflocal basis function methods, fuzzy inference systems, fuzzy neural systems, generalized memory-based learning, and constrained topological mapping. O n e implementation has been the use of the Fuzzy A R W neural network for waste form matrix classification? Non-Destructive Examination Module - Important NDA parameters can be extracted from NDE data. Such parameters of use include the container fill height and the effective matrix density and its distribution. The availability of such information, appropriately formatted, is a useful input to the expert system. HistoricaVGenerator Data Module - Historical generator data include a great deal of information regarding the time of waste production and packag- ing and shipment, assay values acquired prior to shipment, waste form and configuration data, etc. All of this information can be used when bounding the uncertainty in the assay value. Ancillary Characterization Dataflnformation - This module contains information and data derived from specific characterization projects. An example would be a sludge drum coring project and the debris characterization project. In these efforts, drums were opened and sampled to determine and veriq their contents and radionucIidic compositions. Such data provide confirmatory evidence of actual waste form characteristics and serve as a benchmark when evaluating related containers. It is important to note that this is a comprehensive system that may take a considerable effort to estab- lish and validate. It is quite plausible to build the expert system in discrete subsets and apply each as developed. For example, t h e first stage of knowledge base development could be directed at evaluating the acquired measurement data with respect to algorith- mic bounding parameters and waste form generation data. Once established, the benefits of this expert system component could be realized whiIe other more-time-consuming components are in develop- ment and testing phases, an advantage of the modular nature of expert systems. Human experts do not proclaim to be exact in their reasoning and will factor a measure of uncertainty into such processes and final decisions. Likewise, expert systems have mechanisms for representing and managing uncertain knowledge entailed in the set of facts and heuristics as well as environmental data input to the system. There are a number of ways to represent uncertain knowledge, the most common being the use of probabalistic techniques, certainty factors, and fuzzy logic. Bayesian networks and fu logic techniques are currently under investigatio36 for application in the subject waste NDA expert system. Bayesian or belief networks can be used to represent uncertainty through the production of a compl model of a domain that is computation Fuzzy logic is currently being used as a classification tool and as a method to represent the degree of certainty in the classification. After the preliminary classification studies, investigation into the utility of a fuzzy expert system through the assignment of rules, heuristics, and W t e m input to fuzzy sets will be undertaken. In such a system, the inference engine provides a fuzzy output, which may need to be defuzzified depending on the application domain. The management of uncertainty may also be treated via evidential reasoning. Evidential reasoning is based o n the Dempster-Shafer theory (DST) and is an effective method to represent ignorance, incom- plete information, o r inexact rules in expert systems. DST also provides a mechanism to handle conflicting data and rules and is ideal to integrate knowledge from different sources. CONCLUSIONS Waste NDA techniques in many cases do not take advantage of all available knowledge about the problem domain, which can lead to difficulty in technically justifj4ng adherence to compliance criteria. Experts system technology has been identi- fied and investigated as a viable means to accomrno- date a11 available domain-specific knowIedge into a cohesive system. A system for t h e accumulation of available knowledge for use by t h e expert system has been developed, the Waste Assay Measurement and Integration System. The identification of waste NDA data/information sources and means to represent and evaluate such knowledge and t h e associated uncertainty are under study. The expert system is compatible with compliance demonstration activities, and the solution technique is tractable in that the explanation facility output details those factors used to determine the soIution and how uncertainty was treated in each particular case. REFERENCES 1. Transuranic Waste Characreniation @ality Assurance Program Plan, Revision 0, CAO-94-1010, U.S. Department of Energy, Carlsbad Area Office, National TRU Program Office, April 30, 1995. 2. T. J. Roney, G. K. Becker, and K, Mousseau, Methods to Improve the Interpretation of Data porn NDA and NDE of Waste Containers, A Summary of the Waste Assay Measurement Integration System, INEL-96/233, Idaho National Engineering Laboratory, July 1996. 3. G. K Becker et al., "Utility of Neural Networks in Nondestructive Waste Assay Measurement Methods," Proceedings of the 4th Nondestructive A s s q and Nondesmtctive Examination Waste Characterization Conference, October 24-26,1995, Salt Lake City, UT. 4. J. Hodges, S. Bridges, S. Yie, Y. Johnson, and M. Gentry, The Design of an Object-Oriented Radiological Waste Database, MSU-960617, Department of Computer Science, Mississippi State Universiry, MS, June 17, 1996. 5. J. Hodges, S. Bridges, and S. Yie, Preliminary Results in the Use of Fuzty Logic for a RadioIogicaI Waste Characterization Expert System, MSU-960626, Department of Computer Science, Mississippi State University, MS, June 26, 1996. 6. S. Bridges, J. Hodges, and B. \yooley, Preliminary ResuIts in the Use of Bayesian Network for a Radiological Waste Characteriration Expert System, MSU-, June 17, 1996, Department of Computer Science, Mississippi State University, MS. This work was supported by the U.S. Department of Energy under contract DE-AC07-941D 13223. DISCLAIMER This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsi- bility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Refer- ence herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise. does not necessarily constitute or imply its endorsement, recom- mendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. - 
work_y7yspggtcnehnjrulleogza6ei	----	Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) Cooperative problem solving and explanation L. Karsenty P.J. Brezillon CNAM LAFORIA, Box 169 Laboratoire d'Ergonomie University Paris 6 41, rue Gay Lussac 4, place Jussieu 75005 Paris 75005 Paris Cedex 05 France France Abstract: Recent studies have pointed out several limitations of expert systems regarding user needs, and have introduced the concepts of cooperation and joint cognitive systems in the focus of AI. While research on explanation generation by expert systems has been widely developed, there has been little consideration of explanation in relation to cooperative systems. Our aim is to elaborate a conceptual framework for studying explanation in cooperation. This work relies heavily on the study of human-human cooperative dialogues. We present our results according to two dimensions, namely, the relation between explanation and problem solving, and the explanation process. Finally, we discuss the implications of these results for the design of cooperative systems. INTRODUCTION Recent studies have pointed out several limitations of expert systems regarding user needs, and have introduced the concepts of cooperation and joint cognitive systems [Woods et al., 1990] in the focus of AI. Our own experience in expert system projects has led us to believe that systems that place the user in a passive role, cannot provide effective support [see also Carr, 1992]. This paper is concerned with how the framework of cooperation can modify explanation requirements in the context of expert systems. Because few cooperative systems are currently available, we led our investigation from studies of human-human cooperative dialogues rather than directly from studies of human-computer interaction. This paper presents the first step of our study. This step involves describing the explanation features in a cooperation paradigm, as they emerge from the study of human-human cooperative dialogues. The next step will involve the implementation of these results. Although we have not begun to work on this second step, we indicate some elements that will help, we believe, in achieving such a task. The results of the analysis of cooperative dialogues are split in two parts: the first focuses on the relation between explanation and cooperative problem solving, the second, on the process of explanation, which is conceived of as a cooperative process itself. As far as the relation between explanation and cooperation is concerned, two major effects of explanations on the cooperation are noticed: (i) explanations may modify the agents' representation of the problem, enhancing their level of shared understanding of the problem and, consequently, their possibilities of coordination, (ii) explanations may modify the agents' problem solving knowledge. As a consequence of these possible changes, the problem solving process may take another direction, unforeseeable before an explanation need emerges. Hence, we then claim that explanation must be conceived of as part of a cooperative problem solving process, and not just as a parallel phenomenon that does not modify the course of the reasoning. We would like to thank two anonymous reviewers for their helpful comments and suggestions regarding an earlier version of this article. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) If explanations partly determine the cooperation, we may also say that the context of the cooperation, particularly the tasks to be achieved and the users' knowledge, determine the explanation needs. An extensive analysis of the explanation needs in the context of database design dialogues supports this claim, and suggests how task goals and users' knowledge may affect these needs. We then present a view of explanation as a cooperative process. Previous work on explanation has already pointed out the need for an active collaboration between the explainer and the explainee, but considers that (i) the aim of this collaboration is to construct some explanatory content that the machine alone can hardly plan in some cases, (ii) the machine possesses the "right" interpretation of its actions and proposals, and the user has difficulties (misconception, lack of knowledge, …) in reaching this interpretation. Our work not only elaborates on the possibilities of collaboration that two agents bring into play, but also--and perhaps what is more important--presents the idea that the process of explanation may be viewed as a negotiation of meaning. This implies that we do not believe that there is on the one hand one "right" interpretation, and on the other hand, some "erroneous" interpretations. We prefer to say that there are different possible interpretations of the same message, where the validity of each is always founded on some specific contextual knowledge. The goal of an explanation process is then to adjust both agents' context until compatible interpretations may be found. Section 1 reports an analysis of some limitations of expert systems when one tries to implement them in actual working settings, and outlines the need for human-computer cooperation. The more advanced research on explanation in the expert system paradigm is presented in section 2. This allows us to identify the specific features of explanation in the cooperation paradigm, and sets the scene for our studies of human-human cooperative dialogues. Section 3 presents a methodology for analyzing explanations in dialogues, and proposes a new notation for graphically representing cooperative dialogues that makes it easy to see at a glance the high level structure of a dialogue. The results of this analysis are then presented. We will finally conclude in section 4. 1 THE EXPERT SYSTEM PARADIGM 1.1 The user and the system in the expert system paradigm The primary design focus has been to apply computational technology to develop a stand-alone expert system that offers some form of problem solution. This is one main reason why the design of expert system has focused on problem solving, not on the user's needs. This paradigm emphasized tool building over tool use [Woods et al., 1990]. Locus of control typically resides within the machine, and the design intention is for the human to act as the eyes and hands of the machine. For instance, the system directs the user to make observations about the device's behavior and to take measurements from internal test points. Indeed, the human acts as an interface between the expert system and its environment. A classical scenario of user-expert system interaction is: the user initiates a session; the expert system controls data gathering; the expert system offers a solution; the user may ask for “explanation” if some capability exists; and, the user accepts (acts on) or overrides the system's solution [Woods et al., 1990]. 1.2 Inadequacy of expert systems for end-users Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) First expert systems have not fulfilled much of their early promise. There are several reasons for this: A wrong representation of expert system users and their needs; the system inability to solve unexpected problems; the assumption of a well-formulated statement of the user's problem and the expert system opacity. We discuss these points below. 1.2.1 SE users are not novice, but competent agents One generally assumes that the end-user is a novice only because he or she needs the expert's help for solving some particular problems. Actually, several studies of help dialogues and our own experience in expert system projects lead us to prefer to view the users as competent agents. They may have a high level of expertise, although this level may be quite different from that of the expert from whom the knowledge in the expert system was acquired. This expertise allows them to make early judgments about a problem and even generate partial solutions [Pollack et al., 1982; Woods et al., 1990; Maïs & Giboin, 1989]. Thus, one may ask what are the competent users' needs? Several studies have pointed out that a competent agent does not remain "passive" in detecting a problem. Instead, he or she rapidly tries to explain the problem and identify a solution. The agent's information needs begin at this point. A few studies of natural help interactions support this statement. For example, a study of network design dialogues between an experienced designer and a series of less experienced designers indicates that very few requests for solution are uttered by the less experienced agents [Darses et al., 1993]. Instead, they often propose elements of solution and the expert react by assessing the proposal. Moreover, the authors observe that: (i) Either solution developments or preventive statements may follow the expert's positive evaluations; (ii) A justification, often followed by alternative proposals, always comes with the expert's negative evaluations. Other results have been extracted from analysis of competent agents' needs in diagnosis tasks [Woods & Hollnagel, 1987]. These authors claim that good advisory interactions aid problem formulation, plan generation, help determine the right questions to ask, and how to look for or evaluate possible answers. Thus, an effective help facility for competent agents must be able to answer questions like: What would happen if x? what are side effects of x? how does x interact with y? what produces X? how to prevent X? what are the preconditions (requirements) and post-conditions for X (given X, what consequences occur)? [Woods et al., 1990; see also Kidd, 1985; Alvarez, 1992]. Competent agents' needs are not confined to a list of questions that must be addressed when designing the system. They also include the need for another form of interaction where the agent may actively participate in the problem solving process. Therefore, some authors prefer to view user-expert dialogues as a negotiation process rather than a consultation dialogue [Pollack et al., 1982]. In this sense, solving the user's problem is finding a mutually agreed solution instead of just finding a solution. Pollack et al. show why the dialogue could take the form of a negotiation: Often, people calling for advice "have preconceptions about what a solution to their problems involves or what constraints it must satisfy" (p. 361). Given an expert's advice, the caller may want to understand why it is better than the one s/he was expecting. S/he also may want to check if her/his constraints have been taken into account by the expert. The outcomes of these studies do not lead to the conclusion that an helpful system does not have to generate solutions. Rather, they show that this capability addresses just one type of the users' needs, if they are conceived of as competent agents. Such users need help in finding ways to progress in the problem solving process or in assessing their solutions. Moreover, these studies highlight that interacting with an expert is different from a consultation. This section outlines that users' activities would benefit from a cooperation with a computer. The next section argues that a computer would also benefit from the user's help to extend its field of application. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) 1.2.2 Inability to solve unexpected situations One source of difficulty in a problem solving setting lies in the number of particular cases. It is not possible to plan a priori all the cases that may arise. Now, the machine facing unexpected problems is often unable to solve them because, generally, the system knowledge has been extracted from the handling of a set of specific cases. With their experience, users could intervene at this level, but this would imply that the problem must be solved by a cooperation between users and the machine. For instance, consider a real-world device that is composed of several pieces of equipment, some of them being theoretically identical. The behavior of the identical pieces of equipment may be different, due to a difference in their construction (e.g., different adjustments of relative timing). Such a variability is inherent in equipment, and may be a problem for the machine that only knows the theoretical functioning, rather than the real one. It may detect trouble when there are only particular constraints that define two different contexts for the identical pieces of equipment. In contrast, the user who is in charge of the real-world system, often knows why two pieces of equipment behave differently, and may take the 'right' decision (Perrot et al. describes such a situation in an application for the French power system [Perrot et al., 1993].) This inability of expert systems to deal with rare and difficult cases can undermine the credibility of the system [Keravnou & Washbrook, 1989]. But it can also place users in a frustrating position, because it is facing such cases that they really need help. 1.2.3 Solving a problem: Finding out the problem When expert systems provide solutions, they do not always solve the user's problem. There are, at least, two reasons for this: 1. With advice-giving systems, users often fail to ask the most appropriate question because they do no know what information they need [Pollack, 1985; Belkin, 1988; Baker et al., 1993]. Pollack notices that human experts compensate for the lack of relevance of users' questions by inferring their plan. This is one piece of evidence that shows that the expert must find out the problem before solving it. Baker et al. observe that defining the user's problem is not just confined to this act of plan inference. It also implies a dialogue, where both agents negotiate the user's request [Baker et al., 1993]. 2. Another reason why a system may produce irrelevant solutions can be seen from an investigation of real interactions with an expert system [Woods et al., 1990]. The study was aimed at supporting technicians in troubleshooting an electromechanical device. The authors observed that different technicians facing the same problem with the device, provided different symptom descriptions to the system. Moreover, some of these descriptions deviated from what the knowledge engineer, which was in this case an expert of the domain, would have expected. These discrepancies between the provided data and the expected data led the machine to be off-track. These observations highlight that the human-machine pair can be more efficient if both agents share the same understanding of the problem. Observations also reveal that it is not realistic to think that any user will have the same perception of the situation to be handled as the expert system. The analysis of cooperative dialogues between humans in section 3 will show that the sharing of the problem understanding relies on the explanation process. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) 1.2.4 A double-task: Solving the problem and understanding the system We have said previously that the expert system and the user (or the competent agent) may have different levels of expertise relying on different experiences. These discrepancies may result in different lines of reasoning, which in turn may imply sequences of system questions that are incoherent to the user [Keravnou & Washbrook, 1989]. One can relate this problem to observations of users failing in following instructions for using a copier [Suchman, 1987]. Some cases of impasses arose because of mismatches between the user's and the machine assumptions about current world state and objectives. Actually, "following instructions requires actively filling in gaps based on an understanding of the goals to be achieved and the structural and functional relationships of objects referred to in the instructions." [Woods et al., 1990, p. 145]. Users need to understand system actions in problem solving as well as the result of the system reasoning [Teach and Shortliffe, 1984]. When these explanations are too difficult to find, users may refuse to use the expert system because, contrary to their expectations, their working load increases. The first attempts at reducing the opacity of expert systems involved extracting explanations of the system behavior from the record of the rules that were fired during one session. However, this solution appears more informative and adapted to the knowledge engineer than to the end-users. Indeed, the knowledge engineer often introduces some technical constraints to limit the action domain of the acquired knowledge. This is the case for the screening clauses in rule systems [Clancey, 1983]. Such constraints then reduce the comprehensibility of the reasoning trace, which becomes difficult to follow. Other factors explain the inefficiency of this solution for aiding the users' understanding [Moore & Swartout, 1990]: The set of questions taken into account is limited; The system cannot justify its actions; Users cannot provide any feedback that would help the system in finding an explanation adapted to their needs; The system usually cannot provide alternative explanations. These limitations have led to much of the research presented in section 2. 1.3 The need for human-computer cooperation Most of the work in developing expert systems focuses on the knowledge needed for the problem solving, not on the wishes of end-users. A consequence is that end-users must adapt themselves to the expert-system behavior. However, users generally have a high level of expertise, even if this expertise is different from that of the domain expert. As a consequence, computational technology should be used, not to make or recommend solutions, but to help users in the process of reaching their decision [Carr, 1992; Woods et Roth, 1988]. There are two different approaches to avoid the problems that are related to the user- expert system interaction. In the first approach, one increases the explanatory capabilities of the expert system. However, the relationship between both agents remains the same. In the second approach, one modifies the user-system interaction and introduces the notion of human-computer cooperation. Our research aims at integrating both approaches, using studies of explanations in expert systems for designing cooperative systems. Indeed, this integration is not a simple transfer of the results gained from the first approach to the cooperation paradigm. The reason is that the cooperation paradigm modifies some explanation characteristics. Few cooperative systems have currently been implemented. Hence, it is not easy to address the explanation issue directly from a human-computer interaction perspective. Thus, we have investigated the relationship between explanation and cooperation from studies of human-human cooperative dialogues. Two features of cooperation settings must be stressed: First, the control is shared by the machine and the user in joint cognitive systems. Thus, either the machine or the user plays the role of the information provider at each turn, the other providing a support role. Moreover, a mutual dependency characterizes their relation, because both agents attempt to reach a common goal. In other words, each agent must take the other's actions into account. This paper addresses two questions: (i) How the features of cooperation, particularly shared control and mutual dependency, work and can affect the explanations? Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) (ii) What are the main differences between the explanation requirements for the design of an expert system versus those for a cooperative system? 2 EXPLANATIONS IN THE EXPERT SYSTEM PARADIGM 1 One difficulty with expert systems lies in their opacity, i.e., their inability to make clear their reasoning and to justify the solution to the user in a convincing manner. This is particularly crucial when one wants to use the expert system for training and tutorial purposes [Clancey, 1983]. These problems have motivated much of the research on explanation. The first attempts at generating explanations mainly involved providing a trace of the system reasoning. However, it soon became clear that this solution was not satisfactory. Recent attempts have involved the knowledge bases with information about the domain, the task, the ability to communicate with the user. At the same time, one provides tailoring and dialogue mechanisms that make the system more sensitive to the user. These aspects are discussed below. 2.1 Making expert knowledge more explicit Most of the expert knowledge is compiled. Such knowledge expresses under a condensed form a number of knowledge pieces and their organization. Facing this compiled knowledge, the user may need the missing pieces of information for understanding the system reasoning. Clancey (1983) shows how the rules in a classical expert system such as MYCIN have different justifications: identification rules (for classifying an object), causal rules, common sense knowledge rules, and rules of the domain. These justifications, the structural knowledge (for instance hierarchic relations between rules), and the strategic knowledge (e.g., "If there are non-usual causes, treat them") need to be made explicit for improving the possibility of understanding and modification of a system [see also Swartout, 1983; Hasling et al., 1984; Chandrasekaran et al., 1989; Wick & Thompson, 1989]. 2.2 Tailored explanations Three issues must be considered in building a system that is able to tailor its explanations: What must we tailored? Which is the information to use and how to get it? Which tailoring technique to use? The studies described below present a representative sample of the different answers to these questions that can be found in the literature. Adapting the degree of details to the users' knowledge • Wallis and Shortliffe (1984) use a technique for generating explanations of causal chains, which are customized to the users' level of expertise and to the amount of information that they want. The user declares these two values, on a scale ranging from 1 to 10, and can change them during the dialogue. They associate a measure of complexity to each rule and concept in the knowledge base and a measure of importance of each concept. User's needs and the appropriate concepts or rules are then matched. • XPLAIN tailors its answers to users by varying the number of steps included in an explanation depending on the type of user [Swartout, 1983]. Viewpoint markers are attached to the steps in a prototype method, indicating which type of user the step should be explained to. Adapting the explanation types to the users' knowledge • TAILOR adapts the explanations according to a user model with stereotypes and individual features [Paris, 1990]. From her study of encyclopedia, Paris finds that: (i) Objects have to be described in terms of subparts and properties for a knowledgeable user; (ii) The explanation should be focused on how the object works for a novice. •!Moore and Swartout (1990) employ a user model to choose between the several operators that can be candidates for achieving a discourse goal. For example, if the goal is to describe a 1 We thank Béatrice Cahour for her help in making this state of the art. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) concept, one can describe its attributes and its parts, or draw an analogy, or give examples, etc. The choice is then based on the dialogue history and on the user model. Adapting the explanation types to the users' goals • ADVISOR tailors its explanations to the user's goal [McKeown et al., 1985]. It selects a perspective (or viewpoint like: Process of meeting requirements; State model process; Semester scheduling process). It then provides justifications for advice in choosing courses and planning for students, by inferring the user's current goal, and adapts the content of the explanation to the perspective. The user's goal indexes the perspective, which indicates the relevant information to include in the justification of the advice. •!Van Beek also studies how the content of an explanation should change as a function of the goals of a discourse situation and of the higher domain goals of the user![Van Beek, 1987]. In the above research, the explanation is conceived as a text that has to be generated. In the research presented in the following section, the explanation becomes a process. 2.3 Explanation as an interactive process Sometimes users do not understand the explanation that is generated by the system, or need another type of explanation, or judge this explanation as not being useful. In such situations, an alternative explanation must be presented to fit their needs. The machine can then interact with the user to produce a new explanation. As a consequence, the interaction control must be shared by the machine and the user to allow follow-up questions at least. One weakness of explanation facilities is partly due to deficiencies in text planning and generation strategies. This may be compensated for by a reactive approach to explanation [Moore & Swartout, 1990]. According to this approach, the explanation generation relies on the user's feedback. This is an alternative to the weakness of the user model. Indeed, interacting with the user, the system may use incorrect assumptions about the user, and, consequently, unsuitable explanations may be provided. One solution is to build explanations interactively [Cawsey, 1993; Lemaire & Safar, 1991]. The user should be able to interrupt for clarification, and the system should check the user's understanding. Another solution is to acknowledge the impossibility for the machine to provide explanations that are tailored to the user, and let the user build the explanation that s/he wishes [Brézillon, 1992]. This idea is implemented in a program with some functions that allow the user to interrupt the machine. An interruption is possible at some points depending on the knowledge structures of the domain. 2.4 Conclusion The main advances brought about by the studies quoted above are twofold: • First, the knowledge that is used by the reasoning process encompasses contextual knowledge. One must make explicit this contextual knowledge for producing better explanations, especially when the user has a type of expertise that is different to that of the expert. This leads to a revision of methodologies of knowledge acquisition: The knowledge that is needed for explanation must be elicited with the knowledge for the problem solving [e.g., Dieng et al., 1992]. • Second, cooperative explanations may be produced by using a user model. But this is not always sufficient. The system may lack information on the user or even may have a wrong user's representation. As a consequence, it is important to allow the user to intervene in the explanation building. These studies aim at making more acceptable expert systems. We believe that enhancing the explanatory power of computers is not sufficient to reach this aim. The main problem with expert systems lies in the passive role given to the user. Thus, we prefer to look for a more cooperative solution, one where both agents can articulate their specific abilities. However, choosing this option leads us to expect different features of the explanation. This is Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) not only because in cooperative problem solving, there is not just one agent who is attempting to solve the problem at hand. It is also because cooperating agents maintain a specific kind of relation that one may call mutual dependency. Two studies of cooperative dialogues were conducted to give an answer to the following question: How is an explanatory process affected by cooperative work? Actually, this question contains three sub-questions: (1) What are the effects of an explanation process on the agents' tasks and reasoning? (2) What is the relationship between explanations and the cooperation context? (3) How can we model a cooperative process of explanation? 3 EXPLANATION IN THE COOPERATION PARADIGM 3.1 Data collected We studied two types of cooperative dialogue that were recorded in natural working situations. We present each of these dialogues as well as the characteristics of each cooperation. 3.1.1 Validation dialogues The first type of situation observed involves database validation dialogues. In this situation, a designer submits his solution--a database conceptual schema--to future users. The application domain is inventory control of garments. Four dialogues of this type were collected with the same designer and two different groups of future users. This type of cooperation is aimed at providing the designer with the most knowledge supporting his solution. This cooperation relies mainly on the complementarity of the agents'knowledge. Future users possess domain knowledge (activities, information supports, organizational criteria, etc.) and the designer possesses design and database knowledge. These dialogues present a common structure: The designer presents his proposals (each proposal corresponds to a component of the database conceptual schema) and an assessment process takes place. This process may lead to: (i) A direct agreement by the future users without further information; (ii) The users' agreements after the production of an explanation; (iii) A modification of the initial proposal after a negotiation phase. The validation dialogues were recorded and subsequently transcribed. The resulting verbal material constitutes the basis for the analysis. 3.1.2 Design dialogues We have collected design dialogues during two real design projects in aerospace industry. A cooperation between an engineer and a draughtsman is the basis for the design in the organization where the study was conducted. This cooperation relies heavily on a complementarity of the agents' competence. Engineers have the competence to determine materials and measurements. Draughtsmen have the competence needed to determine shapes and principles of integration between the components of a mechanical device. The need for a cooperation arises for two main reasons: (i) The two levels of decision (materials, shapes, etc.) interact; (ii) Both designers' specific experiences are sometimes needed (e.g., to choose between two alternatives). We followed two pairs of designers, each pair having the responsibility of one project during a period of 2 and 4 weeks. We videotaped most of the task-oriented interactions and subsequently transcribed the communications. The problems handled within the projects are different in many respects: (i) The history of the design is different: One project concerns the redesign of a previous solution due to a problem identified during some tests. The other concerns the adaptation of a previous solution to new requirements. (ii) The type of mechanical function to be achieved, the component and the structure of the device aimed at, and the organizational constraints were also different. Moreover, time and cost constraints are more salient in one project than in the other. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) The form of the dialogues occurring in these situations differs from the form of the validation dialogues. Initially, the problem is stated (generally by the engineer) and a preliminary solution is cooperatively established. At the same time, both designers define the tasks to be accomplished. The subsequent meetings are of two types: (i) Pre-planned meetings aimed at assessing the solution state; (ii) "Opportunistic" meetings held each time one of the designers finds out a new problem with the current solution and cannot solve it alone. The cooperation in the first project proceeds essentially with pre-planned meetings. In the second one, the project evolves with "opportunistic" meetings. Thus, data concerning the first project correspond to two dialogues (around 90 minutes each), while data concerning the second project correspond to eleven shorter dialogues (from 10 to 20 minutes). Data gained from these dialogues contain communicative behaviors and other physical behaviors needed for achieving individual tasks. We retain some of these behaviors in our coding when they clarify verbal statements. They include retrieval of information in documents, drawing activities, and operations with CAD tools. 3.2 Methodology for analyzing explanations in dialogues We now describe a basic model of cooperative design and the methodology of dialogue analysis that partly bears on it. 3.2.1 A model of cooperative design At a very abstract level, individual design activity may be seen as an iterative process beginning with ill-defined requirements (or design goals) which lead to the generation of some preliminary solutions. The design project then elaborates on and assesses the solutions that are generated. This process generally leads to a better definition of the requirements, and, as a result, either the refinement or the substitution of the preliminary solutions that have been generated [Malhotra et al., 1980]. One may conceive of cooperative design as a group of agents that intervene in the following manner. One agent brings requirements, then either another one formulates a solution, or a joint action produces a solution (i.e., when different abilities are needed to produce a solution state). The members of the group assess the solution that is produced. Eventually, each member may add data that modify the design goal statement. The problem solving then continues at a deeper level of problem understanding. Cooperative design is a process where each agent has his or her own competence that must be coordinated with the other agents' competence. Thus, other agents' agreement is necessary at each step of the process. This means that each proposal must be mutually agreed to be taken into account by a cooperation (see [Pollack et al., 1982]). For us, this statement is equivalent to saying that proposals must be mutually explainable. 3.2.2 Data analysis The analysis of cooperative dialogues presented below is based on two ideas: 1. A dialogue contains two classes of communicative act: Task-driven acts and acts of explaining. Communicative acts within the first class convey the information that is required by the task at hand. Their definition may be independent of a particular interaction between agents. Acts of explaining convey information required to ensure the success of the task-driven acts. They are dependent on the agents' ability to understand each other2. An illocutionary act and a propositional content constitute each task-driven act [Searle, 1975]. In design dialogues, illocutionary acts are: Inform, propose, ask for information, ask for agreement, criticize. Propositional contents are: Goal, problem data, action, solution, assessing (negative or positive), meta-planning or interpretation (extracting data from blueprints, it is 2 Ethnomethodologists have proposed a model of dialogue very close to this one, which accounts for repair [Jefferson, 1972]. Problems can lead to side sequences, that are oriented for providing a remedy to the problem, so that the main sequence of talk can continue. One may consider explanations as "side sequences" too, regarding to the main sequences that are composed by task-driven acts. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) often necessary to find out what the drawing meaning is). One particular type of propositional content is termed "cognitive state". It occurs when one agent states (or asks) about his state (or the other's state) of knowledge concerning the considered topic. Examples of utterances coded in this way are shown in table 1. Moreover, we have achieved further coding of the content. It invovles the links between different statements of goals, actions, solutions or problem data. These links may be: refinement, correction, or alternative. Table 1: Examples of utterances coded You have the values concerning this part in J.C.'s doc. INFORM (Information source) What we have to do is to prevent the ring from rolling. PROPOSE (Goal) [To prevent the ring from rolling], we are going to achieve a higher tightening on the axle. PROPOSE (Solution) We will see later what measurements are necessary. PROPOSE ("Meta-planning") Do we have to take rigidity into account? ASK-INFORMATION (Goal) Do we begin to put the allowance values? ASK-AGREEMENT (Action) No, you don't need to do this. CRITICIZE (Action) This axle corresponds to the wheel axle. INFORM (Interpretation) Did you already see one similar device? ASK-INFORMATION (Cognitive State) 2. The second idea on which the dialogue analysis is based is the following. One explanation (or act of explaining) conveys contextual information that is needed for understanding and accepting task-required information. This information is missing from the explainee's context3. In communication, context is the set of shared beliefs making an utterance relevant, and allowing the hearer to recognize the speaker's intended meaning (For a discussion about the notion of context and its use in the communication process, see Cahour & Karsenty, 1993, and Mittal & Paris, this volume.) In some cases, it is sufficient to know the nature of the question that precedes an act of information for determining if a piece of information is or is not an explanation. In other cases, the occurrence of linguistic explanation marks (because, for, since, etc.) reflects largely the speaker's explanatory intent. However, explanations are not always given after questions or indicated by linguistic marks. Consequently, we need some criteria to decide if a given utterance is an explanation or not. Two types of criteria have been used: • The first criterion concerns the relation between the information-to-be-explained and the information-which-explains. We state five general types of explanatory link: Definitional, causal, end-mean (or intentional), functioning, and justificative. One expresses (and tests) each type of explanatory link as a question linking two utterances. Relationship between explanatory type and question are: - Definitional: A question like "What is X?" - Causal: A question like "Why X?" . Moreover, the proposition-to-be- explained must be seen as an effect of the proposition-which-explains. - End-mean (or intentional): A question like "What X for?" - Functioning: A question like "How X works?" - Justificative: A question like "Why X, and not Y?" It is important to note that the identification of explanatory links requires domain knowledge, especially when no linguistic mark indicates the presence of an explanation. The following example illustrates this point (A and B are two agents, and the number classifies the corresponding utterance): A1 What is this finger for? B1 For preventing the rolling. B2 You have the tightening, and a finger inside the rotation. 3 Some research experimentally demonstrated that this characteristic is central when choosing the explanatory information [see Turnbull & Slugoski, 1988]. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) Actually, the utterance B2 attempts to explain how the finger (a component of the mechanical device) can prevent the rolling of a ring. The analyst can perceive the explanatory nature of the utterance only if he has knowledge about how the function described in B1 is technically implemented. Otherwise, it is difficult to identify the explanatory link between B1 and B2. However, in some cases, the analyst could assume an explanatory link even with a lack of knowledge (because of the intonation for instance). When such a situation arose, we always asked the designers participating in the dialogue for confirmation of the type of explanatory link suspected. • The second criteria used for the identification of explanation comes from the fact that explanations must convey new pieces of information assumed to be unknown by the explainee, or at least not present in his mind. As a consequence, one acknowledges an explanation only if it is possible to state that the explainer thinks that the explainee missed the conveyed information. This may appear directly in the dialogue. For instance, when the speaker says "We have changed the form of this part because a new problem arose …,". Or this may be possible because of the analyst's knowledge of the project. If the analyst has followed all the previous meetings between the cooperating agents, s/he is often able to identify new pieces of information. 3.3.3 Graphical representation of cooperative dialogues We present a visual representation of dialogues to comprehend better (and communicate) some characteristics of explanation in cooperative dialogues. Figure 1 provides an example of the representation used. Note that explanations appear on the lower part of each line. Note also that agreement requests are indirectly coded by specifying when an agreement is produced as a response (showed by the sign "*".) We divide the course of dialogue in phases according to the nature of the propositional contents. Even if this cut-out sometimes relied on discourse cues (for instance, the passing to an execution phase is reflected by an utterance of the kind: "OK, well, we try to do it?"), often it was the result of the analyst's subjective decision. [Insert figure 1 at this level in a separate page] This type of representation has several advantages for the study of explanation in dialogues. It allows us to directly notice when and how many explanations are produced. It gives a rather abstract view of explanatory dialogues (the process of explanation). Besides these advantages, it also permits us to see very easily which role is taken by each agent in each phase of the dialogue, how many questions were asked and by whom, how many disagreements have been encountered, etc. 3.3 Explanations for cooperating 3.3.1 Bilateral explanations In design dialogues, each agent needs the other's competence and knowledge. Consequently, each agent can make a proposal (of action, solution, assessing and so on), and may need to explain her/his point of view. The representation of a phase of solution negotiation given in Figure 1 illustrates this point. Each agent produces a negative assessment of a solution that has been previously proposed, and explained his other judgment. The need for bilateral explanations points out a primary difference between cooperation situations and interactions with traditional expert systems. If we break the asymmetry of knowledge that is imposed by the expert system paradigm (user as novice, machine as expert), the explanation problematics involves more than making systems able to generate ("good") explanations. The general issue becomes how to allow agents to cooperate better by giving them the ability to explain themselves and understand the other's explanations. Very few studies attempt to provide users with the possibility of formulating their own explanations to the system. Obviously, the main reason is the system limitations for understanding users' utterances. However, researchers attempt to bypass this problem by allowing users to communicate with the machine in a more formal language (e.g., see [Coombs & Alty, 1984]). 3.3.2 Quantitative significance of explanation in cooperative design dialogues Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) The rate of explanations per dialogue is the number of words included in each type of communicative act (explanation versus others). Globally, the explanation portions correspond to 31.5% of the entire content of the dialogues collected. This rate varies according to the projects followed: 28.4 % for the first one, 34.3 % for the other. Because each project differed from the other in many respects (designers' characteristics, type of problem, organizational constraints as time, etc.), it is not possible to explain these differences. This result shows the significance of explanations when two people with different background knowledge and (partly) different goals want to work together to reach a common goal. However, it is difficult to estimate the generality of such a result, partly because only two projects were examined, and partly because few other studies achieve a similar account of explanation in dialogues. The relevant study that is similar to ours, focused on causal explanations occurring during everyday conversations [Nisbett & Ross, 1980]. The authors found that statements expressing or requesting causal explanations account for approximately 15 % of all utterances. 3.3.3 Relations between problem solving and explanation Given a problem, each agent in a cooperation can elaborate goals, or plan actions, and at least partly elaborate on and assess the solution. An optimal cooperation would be one where each piece of information that is provided by one agent, would be directly taken into account by the other to produce the next step of the problem solving process. In some domains, where the situations are all predictable and the procedure well-defined and well-rehearsed (e.g., aviation), cooperation proceeds in most cases such as this. As a consequence, one may note that the language that is used is very restricted and explanation needs occur rarely [Falzon, 1991]. It is well known that the features of design activities are exactly the opposite of these ones: The situation, or the problem state, and the goals are ill-defined, and there is no well- defined procedure to reach these goals. This is the difficulty for predicting the next step in a design activity. However, according to the model of design described above, we could expect the following sequences of communicative act: - [A expresses a goal, B agrees and formulates relevant problem data, an action or a solution proposal] - [A proposes a solution, B expresses a solution assessing, or directly elaborates on it] - [A proposes an action (optionally, with data to be taken into account); B expresses an action assessing, elaborates on it, or directly executes the action] We call "logical" chaining of actions such sequences. The logical aspect refers to the model of design activity that is described above. Following this, we could ask: "What are the effects of explanation on the logical chaining of actions?" Two main effects of explanation have been noticed: 1. Explanation may be a condition to go to the next action Three kinds of observation support this statement: A) When one gives an explanation before the information-to-be-explained (the target), the agreement usually follows the target utterance, and concerns at the same time the explanation and the target utterance. There is not a special agreement for the explanation. Moreover, the next "logical" action follows from the agreement. [Example 1] (in the extracts presented below, E refers to an engineer, and D refers to a draughtsman) G o a l E I have no dimension here (pointing to the picture). [JUSTIFICATION] So, you're forced to put the dimensions of the components that are beside it. [GOAL] D Ok. [AGREEMENT] E For beginning, you can take the dimensions of the bracket directly. [ACTION PROPOSAL] Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) B) When a proposal is provided with an explanation, most of the time, the agreement follows the explanation, and the next "logical" action occurs. [Example 2] G oa l ( R ef in e) Pr ob le m da ta A ct io n pr op os al E Well, we need to determine the clearance which is necessary to avoid a contact. [GOAL] X had achieved a kinematics study, which is here (pointing out a document) [PROBLEM DATA] I With his CAD tools, he turned the actuator, and he noticed what dimensions were staying between the two components. [EXPLANATION] E mmh mmh [AGREEMENT] I Well, the best would be to calculate the order by hand to determine all these dimensions. [ACTION PROPOSAL] E mmh, OK. [AGREEMENT] We note that in a few cases, the hearer may agree on the explanation content, but disagree on the explanatory relation. Generally, this situation causes an extension of the initial explanation. C) When a disagreement is expressed, the "logical" chaining of actions stops momentarily. In some cases, the explanation given by the agent by whom the proposal had been rejected, modifies the other's judgement and, as a result, allows the next "logical" action to appear. [Example 3] * G oa l A ct io n pr op os al E We need to determine the surface-envelope of the trajectory of this point. [GOAL] Is it OK? You see how … [ASK-AGREEMENT] D No, I didn't … [DISAGREEMENT] E You have the axle here (show), the bracket and the actuator that is below. The objective is to allow the rotation. The contact… you know that you can get a contact of this type (shows with his hands), but it will be going to move, one time here, one time there, according to the actuator position. [EXPLANATION] D OK, OK, … [AGREEMENT] E Well. So, I will say: We put on the actuator a track of the contact. [ACTION PROP.] 2. Explanation may modify the direction taken by the cooperation. Once the agents reach the next step in the problem solving, they also restrict the set of potential following steps. If a proposal is accepted by others, it will affect the subsequent actions of the cooperation. For instance, if a goal G is accepted, the other alternatives are rejected. The subsequent solutions that will be generated depend on the earlier solution. We observe that sometimes a communicative act that is refused, leads to a next step that did not fit the "logical" chaining of actions described above. For instance, we have observed the following kind of sequence: Solution proposal ->Disagreement -> Explanation -> Agreement -> Alternative Solution -> Agreement We call here 'explanation' an item that corresponds most of the time to a negotiation during which each agent attempts to explain his orher disagreement regarding the other's view. The act of explaining leads them to make their context explicit, i.e., to formulate the set of beliefs that establishes their position on the topic. This gives them the opportunity to discover eventual discrepancies between either their problem statement or their underlying knowledge. There are some cases where the direction of the cooperation is modified after the provision of the explanation. This exhibits situations where the first speaker had either a wrong representation of the problem (see [example 4]) or lacked domain knowledge (see [example 5]). By changing either one, the explanation and negotiation process may influence the way agents progress in problem solving (one may find a similar conclusion in Cawsey et al., 1992). Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) [Example 4] Pr ob le m d at a (C or re ct ) A lte rn at iv e So lu ti on N eg . as se ss in g D No, we're going to be worried with this. [NEGATIVE ASSESSING] Look: you've 50 mm here, and there you've a flat surface which is like that. We will put a part, mmh … [EXPLANATION] E No. Because from the angle viewpoint, it is this one that is crossing. [DISAGREEMENT + EXPLANATION] D …OK. Then, the bracket does not matter anymore. [AGREEMENT + CORRECT: PROBLEM DATA] E Right. … This means that we'll need to take some measurements on this part directly, and we'll make the corresponding adjustments. [AGREEMENT + ALTERNATIVE SOLUTION] D Yes, right. [AGREEMENT] [Example 5] * N eg . a ss es si ng A lt er na ti ve so lu ti o n E We've to avoid that. [NEG. ASSESSING] D Why ? It does not hold ? [ASK-AGREEMENT: CAUSAL EXPLANATION] E That's not the point. [DISAGREEMENT] Actually, that provokes significant contacts which result in a rolling of the actuator that is not perfect. … [EXPLANATION] D Yeah, OK. [AGREEMENT] E Therefore, the solution consists in determining a complex s u r f a c e t o g e t a c o n s t a n t c l e a r a n c e .[ALTERNATIVE SOLUTION] D Yeah. [AGREEMENT] During an explanation process that follows disagreement statements, agents tend to adjust their context until they reach a compatible interpretation of the proposal. There are two side-effects of this adjustement process. First, the shared understanding of the problem increases. As a consequence, we may expect that the cooperation will be more efficient for handling the current problem because it is better coordinated. Second, each agent's shared knowledge increases. As a consequence, we may expect that the cooperation will be more efficient for solving future problems. Discussion These observations on the relation between problem solving and explanation point out a second difference between cooperation and expert system paradigms. An expert system is like an "oracle.", i.e., it is able to find the "right" solution. Explanations are then necessary to convince users of the validity of the system results. As a consequence, designers of expert systems usually have not considered the effects of explanation on the problem solving process. Within the cooperation paradigm, explanations are a part of the problem solving because solving a problem cooperatively is trying to reach a mutually agreed or mutually explainable solution. The observations that are described above highlight another specific feature of cooperation: The relation between knowledge acquisition and cooperative problem solving. According to these observations, we claim that each disagreement between agents may be a timeliness for acquiring new knowledge. Moreover, we notice that, in a cooperation, knowledge acquisition relies heavily on the confrontation between each one's explanations. 3.3.4 Explanation needs and context Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) We extract the results discussed in this part from the study of database validation dialogues. Most of the explanations that are observed in these dialogues are designers' spontaneous explanations. The summary of explanation needs that is presented below (table 2), is based on the analysis of these spontaneous explanations. This analysis shows that explanation needs depend, at the same time, on the task, the type of information conveyed and the users' knowledge giving a meaning to this information. Table 2: Explanation needs in database validation dialogues4 1. Definitional: What is X? What is the difference between X and X'? 2. Design rationale 2.1 Existence: Is X true? (Here, X is a relationship. For instance: Employees can_buy Garments in_a Catalog') 2.2 Necessity: What is X for? 2.3 Requirements rationale: What caused the existence of a given need? 2.4 Consequences: What are the consequences of X? What are the advantages of X over X'? What if not X? 3. Database use 3.1 Procedure and context of use: When and how to use X? How is G reached? What is the need to use X? 3.2 Justification of a procedure of use: Why do we need P? 3.3 Consequences of a procedure: What are the consequences of P? What are the advantages of P over P'? What if not P? Explanations are of three general types in the database validation dialogues: definitional, design rationale and database use. Each of the two last types is composed of subtypes. Each subtype is defined below by one or many questions that it represents. The account for explanation types in database validation dialogues confirms previous studies showing that users' questions addressed in traditional expert systems literature do not reflect the end-users' actual needs (e.g., see Kidd, 1985; Hughes, 1986; Gilbert, 1987). Moreover, it is important to note that the target of these questions is never the part of the designer's reasoning. This reasoning part concerns his own expertise, the way by which the domain may be structured into a conceptual schema. Thus, the designer never explains, for instance, why he had chosen to structure the employee's addresses as an attribute rather than an entity. At the same time, we notice that when database designers discussed a given conceptual schema, the information structure was a central concern of their discussion. Indeed, one of their main objectives was to be ensured that the structure of the data proposed was the best, given technical criteria such as data integrity, rapidity of access, etc. 4 'X' generally represents one datum, as the employee's name, 'P' refers to a procedure, or sometimes a single action, implied by the use of the future database, and G refers to a users' goal. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) Discussion: Explanation needs, perspectives and goal-oriented interpretation This difference between end-users and database designers highlights that the same information presented in the same context of task (the validation of a solution) could result in different goals and, as a consequence, in different explanations needs. More precisely, it seems that these two kinds of people--end-users and database designers--do not necessarily "see" the same thing facing the same conceptual schema. In other words, the same conceptual schema does not mean the same thing for each of them. End-users figure out a tool which must be necessary for helping them in their activities. Database designers figure out the result of an information structuring activity, where many options were possible, and only some of them were chosen given a set of design goals. These different perspectives on the same object, which include the agent's knowledge and concerns, may result in different interpretations of the same solution proposal. For instance, the sentence "For each employee, we need to store his name" would be understood: (i) By the end-users, as a description of an information requirement; (ii) By another database designer, not only as a description of an information requirement, but also as a structure of two kinds of data, one abstract datum, the entity "Employee," and one physically represented datum, the employee's name. Given their perspective, the users reach a specific interpretation of the designer's solution proposal. We postulate that the interpretative process ends when the users find how this solution proposal is compatible with a response of agreement. In order to produce such a response, the users must check some dimensions of the solution proposal. We term these dimensions acceptance goals. An acceptance goal may be defined more precisely as a desired conclusion, imposed by the task, and that must be deduced from one solution proposal and some contextual knowledge. Thus, in a validation task, end-users may want to know what such information, for instance the employee's name, is needed for. This is necessary for expressing an agreement with the solution proposal. The corresponding acceptance goal may be termed "X is useful." For the same proposal, the database designer may want to check if it will not be redundant with another datum, such as the employee's number. The corresponding acceptance goal may be termed "X is not redundant with Y". Saying that the users' interpretative process ends when they find how the solution proposal is compatible with a response of agreement implies the following proposal: Users must contextualize the solution proposal in order to reach the acceptance goals. In this sense, the interpretative process is goal-driven. Moreover, we may say that the need for an explanation arises if one of these acceptance goals cannot be reached. For instance, the end-user may ask "What is X useful for ?" if s/he is unable to contextualize X in order to interpret "X is useful". Or the database designer may ask "Why is Y not sufficient ?" if s/he is unable to contextualize X in order to interpret "X is not redundant with Y". Note that the same information given in the same task context provokes different acceptance goals, which may then induce different explanations needs (i.e., different questions). This is equivalent to saying that an explanation is what bridges the gap between an interpretation that an agent is able to construct and the task-dependent conclusions s/he wants to reach (see Leake, 1991 for an account of the dependency between agent's goals and explanation needs.) 3.3.5 Conclusion A difficulty for accepting an agent's proposal may lead the cooperation to improve not only the shared understanding of the problem to be solved, but also the agents' shared knowledge. These improvements are possible because of the explanations provided by both agents. Moreover, these improvements must lead to a better coordination between the agents, and consequently, a better cooperation. In this sense, we may consider explanation as a means of the cooperation. Explanations could serve the cooperation only if both agents--the system and the user-- can explain their proposals and critics. The ability to take users' explanations into account includes the need for (i) acquiring information from the user, (ii) propagating and testing the consequences of this information in the knowledge bases of the system, (iii) subsequently, Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) asking for more information from the user, and (iv) eventually, acquiring new knowledge. Some of these capabilities are already present in some recent work aimed at designing cooperative systems, but under the user's control (e.,g., see Fischer, 1990; Edmonds & Ghazikhanian, 1991; Levrat & Thomas, 1993). Other studies seem relevant to be considered in order to gain elements for implementing these requirements, particularly from the field of Knowledge Acquisition. Recently, some authors have proposed that this task should be considered as being accomplished in an incremental manner (e.g., see Karbach et al., 1990.) The system receives a loose specification as input and outputs a more comprehensive specification that provides the characterization of the domain. The overall goal is then to detail the structure and purpose of the domain in the context of its evolving theory (see Brézillon, 1994 for a review of the systems coupling knowledge acquisition and explanation.) We have concluded that the way a cooperation progresses depends on explanations. We can also conclude that explanation needs depend on the context of the cooperation. Indeed, in a cooperation, agents try to reach a solution that must be mutually agreed. Then, any proposal must be accepted by the other agents. This acceptance phase relies on the explanatory power of the information that is provided. The observations reported above shows that this feature is dependent on contextual factors such as the task at hand, and other agents' knowledge. We assume that the modelling of perspectives proposed in McKeown et al. (1985) could represent an interesting solution for implementing these ideas. The authors propose that perspectives should be represented with intersecting multiple hierarchies. The hierarchies are cross-linked by entities and processes which can be viewed from different perspectives. This kind of partitioning of the knowledge base allows the generation system to distinguish between different types of information that support the same fact. We now turn to the study of the explanation process itself, without consideration of the task context. 3.4 Cooperation for explaining Once an explanation need arises in a dialogue, how is it satisfied? The data collected from cooperative dialogues show that both agents--the explainer and the explainee--attempt to reach the explanatory goal in a joint effort. We use the terms "explainer" and "explainee" for describing two roles that can be taken by either agents during a process of explanation. Inside a given exchange, each role can be speaker or hearer (for convenience, we will use "he" for referring to the explainer, and "she" for referring to the explainee). Thus, we note that: (i) Explanation needs are not necessarily revealed by a question; Most of the time they are predicted, and explanations are produced spontaneously; (ii) When a request for an explanation appears in a dialogue, contextual information helping the explainer in answering the request may also be conveyed; (iii) Explainee's interventions may complete the explainer's first explanatory attempt. Some previous studies have already discussed the collaborative nature of the explanation process [e.g., Gilbert, 1989; Moore & Swartout, 1990; Cawsey, 1993]. However, we believe that the collaboration depicted by these studies does not reflect the richness of the possibilities that two agents can use for achieving an explanation goal. Our motivation in trying to reflect all these possibilities lies on two facts: First, to our knowledge, they are not described in the literature on explanation; Second, we believe that the more advanced systems built up until now are still unable to solve some misunderstandings, especially when the gap between the system's and the user's knowledge is large. One of the main reasons is that most of the user's collaborative power is not taken into account. 3.4.1 Spontaneous explanations Most of explanations are given spontaneously in naturally occurring dialogues. Thus, in validation dialogues, we pick out three reactive explanations against 25 spontaneous explanations of proposed solution on average per dialogue (83% of spontaneous explanations). Other researchers also note this surprising significance of spontaneous explanations. Cahour and Darses (1993) study diagnosis dialogues recorded in a sewage plant between an operator and an engineer. They found that eighty per cent of the explanations collected were volunteered. Belkin (1988) has also pointed out that the help given by an adviser during Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) information-seeking dialogues is also characterized by the superiority of spontaneous explanations on reactive explanations. An argument that may account for such a phenomenon is the fact that agents are mutually dependent. Each piece information that is provided by one agent to another is useful for the receiver and indirectly for the provider. Indeed, the actions and results of the other will depend on the information that is provided and understood. Moreover, the provider's goals rely on the other's results. Thus, one may understand why the provider wants to be understood and why he produces so many spontaneous explanations. Spontaneous explanations appear in three ways (see Draper, 1987): (i) When submitting new or conflicting information, (ii) When detecting a misconception from the other agent's questions or assertions [McCoy, 1988], (iii) When detecting an obstacle in the other agent's plan (e.g., Allen & Perrault, 1980). We will only consider the first case here. We do not find spontaneous explanations with any proposed solution. In database validation dialogues, twenty-nine per cent of the solution proposals (in total, 64 from 222) are spontaneously explained. Moreover, we notice that these explanations are aimed at three different functions: • Fill in missing knowledge required to check task-dependent acceptance goals Example: "We need the 'order number' because it will allow you to retrieve all the articles ordered." In this sentence, the designer points out the re ason why the data 'order number' is useful. We consider "X is useful" as an acceptance goal that is dependent on the validation task. Most of the explanations belonging to this class were provided with the proposal of new elements of the conceptual schema5, or when a procedure of use of some data was unusual (for instance, because one datum has not to be entered by the user, but is produced by the machine). • Prevent or resolve a conflict between some users' expectations and the solution proposal. Example: "For each garment, we are going to store the 'size'. Well, at the beginning, I didn't write this because the machine must determine the size, given the employee's name and the garment ordered. But there are a few cases where we need to manually enter this information [describes the cases]." The previous precedent solution acts here as an users' expectation, which is contradicted by the designer's new proposal. The explanation modifies the context in which the new proposal has to be considered. This kind of explanation may be given before or after presenting a proposed solution. In the first case, the explainer tries to prevent a conflict between what was expected and what is proposed, while in the second case, the explainer attempts to resolve this kind of conflict. • Argue for a proposal P being insufficiently founded. Example: "We will not store the value of each garment, because it seems that it is constantly changing. (Indeed) X said to me that the suppliers changed their price almost every week because …." The second part of the explanation ("X said to me …") aims at supporting the belief that "The value of the garments is constantly changing." Such an explanation goal is very similar to the first one. The designer provides the users with missing knowledge that is needed to check an acceptance goal. However, the acceptance goal that is relevant here is not directly linked to the task. Instead, it is related to the information conveyed in the explanation. These three types of spontaneous explanation point out that there are three situations in which an acceptance goal cannot be reached: Either users miss contextual knowledge which is necessary to deduce the acceptance goals from an interpretation of a solution proposal; the users' context contain false assumptions, which could lead them to reject a given proposed 5 We were able to check the novelty of the proposed solutions because (i) we were present in some of the previous meetings between the designer and the end-users, and consequently were able to say if a given proposal had been already discussed, (ii) reports of previous meetings with the end-users were available, (iii) the designer might present a given proposal as being new : "I discover yesterday this need. So we need to store these data" Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) solution; users miss contextual knowledge that is necessary to accept the information included in an explanation. Karsenty and Falzon (1992) interpret spontaneous explanations as an anticipation of the users' explanation needs. This anticipation ability must rely on: (i) a model of users' knowledge and expectations (the users' context); This gives the possibility of reaching an interpretation close to that of the users' one. (ii) a model of the users' acceptance goals related to a given proposed solution in a given task. The significance of spontaneous explanations seems to reveal that the designer must bridge the gap between the users' context and his own context each time that the discrepancies between them could result in a users' disagreement. Moreover, we may assume that not providing spontaneous explanations is equivalent to communicating all relevant information that the users can infer 6 for validating a solution proposal (the meaning of the data proposed, their necessity, their procedure of use, and so on). A direct consequence of this assumption is that explicit (i.e., not predicted) requests can be seen as cases when the designer has an inaccurate model of other agents' context [Karsenty & Falzon, 1992]. This suggest that the explainer's processing of unexpected questions (particularly, follow-up questions) must include a revision of his model of the other agents' context. This analysis stresses the significance of the notion of context, which is necessary (i) for interpreting what a speaker means when s/he produces a message, (ii) for anticipating what hearer will believe that a speaker means. The significance of the notion of context is well- recognized in various domains like Natural Language and Human-Machine Communication (e.g., see Cahour & Karsenty, 1993; Careni et Moore, 1993; Mittal & Paris, 1993.) However, there are few implemented systems that explicitly deal with context. A reason is that few formalisms of knowledge representation permit an effective use of context (some first attempts may be found in McCarthy, 1993; Sowa, 1992.) 3.4.2 "Contextualization" of explanation questions Traditionally, we consider a question as a proposition expressed under an interrogative form. For instance, "Why do you need this information?," "How did you reach this conclusion?" In real dialogues, the explainee may provide the explainer with contextual information with his question. We distinguish two cases that seem particularly relevant to the handling of explanation needs: 1. Question and answer assumption Example: "Is this built with two parts because of some specific problems of assembly?" In this utterance, we may distinguish the question "Why is this built with two parts?" and an answer assumption "because of specific problems of assembly." This kind of contextualization has the advantages of: (i) Reducing the explainer's effort and allowing a point of mutual understanding to be reached more rapidly if the answer confirms the explainee's assumption; (ii) Allowing the explainer to predict an explanation need if the answer differs from the explainee's assumption. 2. Question and question rationale Example: "You said that we will ask the supplier for recording a further paper for each order. I know by experience that such papers get lost, what creates new problems, I let you figure it out. So, isn't there a more realistic solution?" Here, the question is the last sentence of the intervention. (Note the difference if just this sentence had been uttered.) The two first sentences constitute the contextual information that describes the reasons for the request. Such a contextualization of the question may take the form of a linguistic mark referring to the previous part of the dialogue where the question rationale lies. Example (A and B are two persons, and numbers organize utterances): A1 Is this part a pivot? B1 No, this is an embedding. 6 This is a direct consequence of the Gricean principle of cooperation and the maxims of conversation [Grice, 1975]. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) A2 Then, what is the function of this axle? In this example, A utters A2 because of B's negative answer to A1. The marker "then" refers to the sequence A1-B1. We believe that this kind of contextualization aims at guiding the explainer in finding an appropriate explanation. We base our assumption on the model of explanation that arises from some psychological studies. For instance, Hilton says that "To understand a request for an explanation, we must know the implicit contrast that it presupposes, the rest is logic" [Hilton, 1988, p. 58]. Another quotation may complete this one [Turnbull & Slugoski, 1988, p.68]: "A felicitous answer consists of the provision of information unknown to the asker that can resolve the asker's puzzle. Including their state of mutual knowledge, the knowledge states of asker and askee, are central to the explanation process." Then, with question rationale, the explainer can better comprehend--and better resolve-- the nature of the cognitive conflict that provokes the explanation need when the explainee informs the explainer of her question rationale. We believe that this kind of contextualization reflects the same principle that explains the significance of spontaneous explanations. Askers must bridge the gap between the askee's context and their own context each time that the discrepancies between them could result in a difficulty in answering or in the providing of an unsuitable answer. The quantitative significance of spontaneous explanations leads us to believe that a speaker who initiates a proposal, wants to help the hearers in understanding it. With the examples of contextualization of a question presented above, we notice that the hearers may also attempt to help the speaker to find the appropriate answer. 3.4.3 The process of explanation Often, an act of explaining is not a single intervention. There is a need for a mutual agreement through a mixed-initiative dialogue where both the explainer and the explainee act together. Such a co-participation may take different forms: • An explainee regulation of the explanatory information flow One obtains such a regulation with linguistic marks of agreement (e.g., 'OK', 'mmh', etc.) that cut out the explanation. It is important to note that very often these agreement marks are sought by the explainer. This means that he must decide when to cut an explanation to control the explainer assimilation process. Further studies are needed to understand more precisely how this is done. • Follow-up questions • Explainee's context description Such a description is used when the hearer fails to understand the speaker. It consists of describing the reasons that prevent her from understanding the speaker's utterance. This is particularly clear when the answer to a follow-up question does not satisfy the explainee. The following example, taken from a real phone conversation, illustrates this possibility [Cahour & Karsenty, 1993]: A and B know each other. A is an ex-student and B a professor. A1 I call you to see if you could write for me a letter of recommendation for an academic situation. B1 Oh... yes... oh... but... I don't quite understand... are we talking of the same thing? B2 Mr. Y called me and told me that you were interested by requiring this situation in our laboratory. A2 Oh, I didn't know, A3 And since then I happened to know that the possible laboratories are defined beforehand and your laboratory is not in the list. B makes his context explicit with the utterance B2. Note that in doing so, he explains his unexpected question ("Are we talking of the same thing ?"). This allows A to detect more easily a mistake in B's representation of A's intent, and to correct it (A2 and A3). This ability gives human agents the power for repairing most of the communication failures. • Rephrasing of an explanation by the explainee This also allows agents to reach a point of mutual understanding. The explainee rephrases the explanation, and asks the explainer for a confirmation at the same time. Example: A1 Otherwise, maybe we need to index them according to the angle […] in order to set this stuff upon that one. B1 Ah OK, because you have an indexing between this and that also [rephrasing] Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) A2 Yes, that's it. B2 OK, OK. Actually, it is better to say that the explainee makes his context explicit once the explanation was accepted. Rephrasing the explanation with her words, the explainee just expresses the changes that it provokes in her mind. The explainer's confirmation ensures the explainee that she has understood the speaker's utterance. 4 CONCLUSION In the expert system framework, one agent--the system--achieves the problem solving process. In a cooperative problem solving process, two agents at least may develop a reasoning and achieve parts of the common task. This dictates most of the explanation features in the cooperation framework. We give below some of these features: (1) Bilateral explanations, (2) The explanation process as part of problem solving, (3) Relation between problem solving and knowledge acquisition mediated by the explanation process, (4) An explanation process driven by acceptance goals related to the task at hand, (5) Significance of spontaneous explanations, caused by the mutual dependency between agents. The last feature seems to be more general, and less dependent on a cooperative work setting: (6) The explainer and the explainee must cooperate to achieve an explanation goal. On the one hand, the explainer anticipates the explainee's needs and looks actively for cues indicating how his explanation is understood. On the other hand, the explainee may act in many ways for making the speaker's talk more understandable. Moreover, we have proposed a view of the communication process where speakers only volunteer explanations when they can predict that the hearer's context is not appropriate for inferring the full intended meaning of a proposal. As a consequence, uttering a proposal without any spontaneous explanation is equivalent to communicating all the relevant inferences that the other agent can draw. This study of human-human cooperative dialogues, by describing the complexity of the explanation process, leads to the following conclusion. It is still difficult to believe that a self- explainable system can produce satisfying explanations, especially when the gap between agents' knowledge (the user and the system) is large. In particular, this would require natural- language processing capabilities, and models for managing the communication with the user that are still not available (see Baker, 1992 for some interesting insights on handling the problem of complex dialogue management.) However, we hope that our analysis of the explanation features in a cooperation will help in building a better computational model for the design of cooperative systems. REFERENCES Allen J.F. & Perrault C.R. (1980) Analyzing Intention in Utterances. Artificial Intelligence, 15, 143-178. Alvarez I. (1992) Morphological explanation: a strategy based on the study of the case to explain. Proc. of the ECAI Workshop on "Improving the Use of Knowledge-Based Systems with Explanations", Vienna, Austria. Research Report 92/21, LAFORIA, Box 169, University Paris 6, 4 place Jussieu, 75005 Paris cedex 05, France. 57-64. Baker M. (1992) Analysing rhetorical relations in collaborative problem solving dialogues. Proc. of NATO Workshop "Natural Dialogue and Interactive Student Modelling", Varenna, Italy, Oct. 16-19. Baker M., Dessalles J.L., Joab J.L., Raccah P.Y., Safar B. & Schlienger D. (1993) Analysis and modelling of negotiated explanations. Proc. of the Workshop on Explanation and Cooperation, Research Report N° 105, CNAM, Laboratoire d'Ergonomie, 41 rue Gay- Lussac, 75005 Paris, France (in French). Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) Belkin N.J. (1988) On the nature and function of explanation in intelligent information retrieval Proc. of the 11th Conf. on Research and Development in Information Retrieval, Grenoble, France. 135-145. Brézillon P. (1992) Architectural and contextual factors in explanation construction. In Proc. of the IJCAI Workshop on "Improving the Use of Knowledge-Based Systems with Explanations", Vienna, Austria. Research Report 92/21, LAFORIA, Box 169, University Paris 6, 4 place Jussieu, 75005 Paris cedex 05, France. 65-74. Brézillon P. (1994) Design of an Intelligent Assistant System from Several Application, Proc. of the International Conference for Expert Systems for Development, Bangkok, March (to appear). Cahour B. & Darses F. (1993) Cognitive Processing and Explanation Strategies in a Diagnosis Task. ISEE Deliverable D140.2, Esprit project 6013. Cahour B. & Karsenty L. (1993) Context of Dialogue: A Cognitive Point of View. Proc. of the IJCAI'93 Workshop on Using Knowledge in its Context. Chambéry, France, August 29. Research report 93/13, LAFORIA, Box 169, University Paris 6, 4 place Jussieu, 75005 Paris cedex 05, France. 20-29. Carenini G. & Moore J.D. (1993) Generating Explanation in Context. Proc. of the International Worshop on Intelligent User Interface. Orlando. Carr C. (1992) Performance Support Systems: A New Horizon for Expert Systems. AI Expert, May 1992. 44-49. Cawsey A. (1993) Planning Interactive Explanations. International Journal of Man-Machine Studies, 38(2), 169-200. Cawsey A., Galliers J., Reece S. & Sparck Jones K. (1992) The Role of Explanation in Collaborative Problem Solving. Proc. of the ECAI-92 Workshop "Improving the Use of Knowledge-Based Systems with Explanations". Vienna, Austria. Chandrasekaran B., Tanner M.C. & Josephson J.R. (1989) Explaining control strategies in problem solving. IEEE Expert, Fall 1989, 9-24. Clancey W.J. (1983) The Epistemology of a Rule-Based Expert System - A Framework for Explanation. Artificial Intelligence, 20, 215-251. Coombs M. & Alty J. (1984). Expert systems: an alternative paradigm. International Journal of Man-Machine Studies, 20, 21-43. Darses F., Falzon P. & Robert J.M. (1993) Cooperating partners: investigating natural assistance. In: Salvendy G. & Smith M.J. (Eds.) Human-Computer Interaction: Software and Hardware Interfaces. Elsevier. Dieng R., Giboin A., Tourtier P.A. & Corby O. (1992) Knowledge-acquisition for explainable, multi-expert knowledge-based design systems. In: Wetter T., Althoff K.A. Boose J. Gaines B. Linster M. & Schmalhofer F. (Eds.)Current Developments in Knowledge Acquisition: EKAW-92 . Springer-Verlag, Berlin. Draper S.W. (1987) The Occasions for Explanation. Proc. of the 3rd Alvey Explanation Workshop, Guilford, September 1987, 199-207. Edmonds E. & Ghazikhanian J. (1991) Cooperation Between Distributed Knowledge-Bases and the User. In: Weir G.R.S. & Alty J.L. (Eds.) Human-Computer Interaction and Complex Systems. London: Academic Press. Falzon P. (1991) Cooperative Dialogues. In: J. Rasmussen, B. Brehmer & Leplat J. (Eds.) Distributed Decision Making: Cognitive Models for Cooperative Work. Chichester, UK: Wiley. Fischer G. (1990) Communication Requirements for Cooperative Problem Solving Systems. Information Systems, 15(1), 21-36. Gilbert N. (1987) Question and Answer Types. In: Moralee D.S. (Ed.) Research and Development in Expert Systems III. Cambridge University Press. Gilbert N. (1989) Explanation and Dialogue. The Knowledge Engineering Review, 4(3), 235- 247. Grice H.P. (1975) Logic and Conversation. In: Cole P. & Morgan J.L. Syntax and Semantics, vol.3. NY: Academic Press. Hasling D.W., Clancey W.J. & Rennels G. (1984) Strategic explanations for a diagnostic consultation system. International Journal of Man-Machine Studies, 20, 3-19. Hilton D.J. (1988) Logic and Causal Attribution. In: Hilton D.J. (Ed.)Contemporary Science and Natural Explanation. Commonsense Conceptions of Causality. Harvester Press. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) Hughes S. (1986) Question Classification in Rule-Based Systems. In: Brauer M.A. (Ed.) Research and Development in Expert Systems III. Cambridge University Press. Jefferson G. (1972) Side sequences. In: Sudnow D.N. (Ed.) Studies in social interaction. New York: Free Press. Karbach W., Linster M. & Voss A. (1990) Models, methods, roles and tasks: many labels - one idea ?Knowledge Acquisition, 2, 279-299 Karsenty L. & Falzon P. (1992) Spontaneous explanations in cooperative validation dialogues. Proc. of the ECAI-92 Workshop on Improving the use of knowledge-based systems with explanations. Vienna, Austria. Research Report 92/21, LAFORIA, Box 169, University Paris 6, 4 place Jussieu, 75005 Paris cedex 05, France. Keravnou E.T. & Washbrook J. (1989) What is a deep expert system ? An analysis of the architectural requirements of second-generation expert systems. The Knowledge Engineering Review, 4(3), 205-233. Kidd A.L. (1985) What Do Users Ask ? Some Thoughts on Diagnostic Advice. In: Merry M. (Ed.) Expert Systems 85. Cambridge University Press. Leake D.B. (1991) Goal-Based Explanation Evaluation. Cognitive Science, 15, 509-545. Lemaire B. & Safar B. (1991) Some necessary features for explanation planning architectures: study and proposal Proc. of the AAAI'91 Workshop on Explanation, Anaheim, USA, July 1991. Levrat B. & Thomas I. (1993) Tailoring Explanations to the User's Expectations: A Way to Be Relevant. Proc. of the IJCAI'93 Workshop on Explanation and Problem Solving. Chambéry, France, August 29. Maïs C. & Giboin A. (1989) Helping users achieve satisficing goals. In: Smith M.J. & Salvendy G. (Eds.) Work with Computers: Organizational, Management, Stress and Health Aspects. Elsevier Science Publishers: Amsterdam. 98-105. Malhotra A., Thomas J.C., Carroll J.M. & Miller L.A. (1980) Cognitive processes in design. International Journal of Man-Machine Studies, 12, 119-140. McCarthy J. (1993) Notes on formalizing context. Proc. of the 13th IJCAI-93, Chambéry, France. 555-560. McCoy K.F. (1988) Reasoning on a Dynamically Hilghlighted User Model to Respond to Misconceptions. Computational Linguistics, 14(3). McKeown K.R., Wish M. & Matthews K. (1985) Tailoring Explanations for the User. Proc. of IJCAI'85, 794-798. Mittal V. & Paris C.L. (1993) Context: Identifying its elements from the communication point of view. Proc. of the IJCAI'93 Workshop on Using Knowledge in its Context. Chambéry, France, August 29. Research report 93/13, LAFORIA, Box 169, University Paris 6, 4 place Jussieu, 75005 Paris cedex 05, France. 87-97. Moore J.D. & Swartout W.R. (1990) A Reactive Approach to Explanation: Taking the User's Feedback into Account. In: Paris C.L., Swartout W.R. & Mann W.C. (Eds.) Natural Language Processing in Artificial Intelligence and Computational Linguistics. Kluwer Academic Publishers. Nisbett R.E. & Ross L. (1980) Human inferences: Strategies ans shortcomings of social judgement. Englewood Cliffs, NJ: Appleton-Century- Crofts. Paris C.L. (1990). Generation and Explanation: Building an Explanation Facility for the Explainable Expert Systems Framework. In: Paris, Swartout, Mann (Eds.) Natural Language Generation in Artificial Intelligence and Computational Linguistics. Kluwer Academic Publishers. Paris C.P., Wick M.R. & Thompson W.B. (1988) The Line of Reasoning Versus The Line of Explanation. Proc. of the AAAI'88 Workshop on Explanation. Anaheim, USA, July 1991. Perrot L., Brézillon P., Fauquembergue P. (1993) Towards automatic generation of knowledge bases for diagnosis systems in the field of power systems. Proc. of ISAP'93, January 1993. Pollack M.E. (1985) Information Sought and Information Provided: An Empirical Study of User/Expert Dialogues. Proc. of CHI'85 San Francisco. 155-159. Pollack M.E., Hirschberg J. & Webber B. (1982) User Participation In The Reasoning Processes of Expert Systems. Proc. of AAAI-82, National Conference on Artificial Intelligence. 358-361. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) Sowa J.F. (1992) Representing and reasoning about contexts. Proc. of the AAAI'92 Workshop on Propositional Knowledge Representation, Stanford, CA. 133-142. Suchman L. (1987) Plans and situated actions: The problem of human-machine communication. Cambridge, UK: Cambridge University Press. Swartout W.R. (1983) XPLAIN: a System for Creating and Explaining Expert Consulting Programs. Artificial Intelligence, 21(3), 285-325. Teach R.L. & Shortliffe E.H. (1984) An Analysis of Physicians' Attitudes. In: Buchanan B.G. & Shortliffe E.H (Eds.) Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Reading, Mass: Addison-Wesley. Turnbull W. & Slugoski B.R. (1988) Conversational and Linguistic Processes in Causal Attribution. In: Hilton D.J. (Ed.) Contemporary Science and Natural Explanation. Commonsense Conceptions of Causality. Harvester Press. Van Beek P. (1987) A Model For Generating Better Explanations. Proc. of the 25th Conference of the Association for Computational Linguistics, Stanford, CA. 215-220. Wallis J.W. & Shortliffe E.H. (1984) Customized Explanations Using Causal Knowledge. In: Buchanan B.G. & Shortliffe E.H. (Eds.) Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Reading, Mass: Addison- Wesley. Wick M.R. & Thompson W.B. (1989) Reconstructive Explanation: Explanation as Complex Problem Solving. Proc. of the 11th International Joint Conference on Artificial Intelligence, Detroit, Michigan, August 20-25. Woods D.D & Hollnagel E. (1987) Mapping cognitive demands in complex problem solving worlds. International Journal of Man-Machine Studies, 26, 257-275. Woods D.D. & Roth E.M. (1988) Cognitive Systems Engineering. In: Helander M. (Eds.) Handbook of Human-Computer Interaction. North-Holland: Elsevier. Woods D.D., Roth E.M. & Benett K. (1990) Exploration in joint human-machine cognitive system. In: Robertson S., Zachary W. & Black J.B. (Eds.) Cognition, Computing and Cooperation . Norwood, NJ: Ablex. Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) Figure 1: Graphical examples of cooperative design dialogues 20 30 40 50 + 50 words * : response PLANNING PHASE * * * ≤ 10 * : Engineer's supplying of information : Draughtsman's supplying of information : E's intervention in the explanation : D's intervention in the explanation PHASE OF SOLUTION NEGOTIATION : E's agreement : D's agreement : disagreement (criticize) : explanation after or before the information-to-be-explained G oa l ( R ef in e) Meta-planif (Refine) E xe cu tio n A ct io n (R ef in e) M ét a- pl an if * C om m en t. D ra w in g * PLAN EXECUTION PHASE * S ol ut io n : E's physical action : physical action with comments ? : question : D's physical action E xe cu tio n In fo . so ur ce In fo s ee ki ng Pr ob le m d at a Pr ob le m d at a Pr ob le m d at a Pr ob le m d at a G oa l ( R ef in e) A lte rn . in te rp re t. in te rp re ta ti on G oa l ( R ef in e) C om m en t. D ra w in g Pr ob le m d at a C om m en t. D ra w in g Pr ob le m d at a * ? Pr ob le m d at a ? ? ? * N eg . as se ss in g Pr ob le m d at a ( C or re ct io n) So lu tio n (R ef in e) A lte rn . So lu tio n N eg . as se ss in g A ct io n (R ef in e) in te rp re ta ti on A lte rn . in te rp re t. in te rp re ta ti on E xe cu tio n * ? In fo s ou rc e ? In fo s ou rc e G oa l ( R ef in e) A lte rn . So lu tio n Int. J. Expert Systems with Applications,, 8(4): 445-462 (1995) 
work_ya7rjomoergo7pknkz4264ygcm	----	1 Neural Network for Dynamic Human Motion Prediction Mohammad Bataineh, Timothy Marler, Karim Abdel-Malek, and Jasbir Arora Virtual Soldier Research Program – Center for Computer-Aided Design, The University of Iowa, Iowa City, IA, USA Email addresses: bataineh.moe@gmail.com (Mohammad Bataineh), tmarler@engineering.uiowa.edu (Timothy Marler), amalek@engineering.uiowa.edu (Karim Abdel-Malek), jasbir-arora@uiowa.edu (Jasbir Arora) Corresponding author: Mohammad Bataineh will handle correspondence at all stages of refereeing and publication, also post-publication. (Tel: +1-319-331-5454; Email: bataineh.moe@gmail.com; Address: 330 S. Madison Street. Iowa City, IA 52242. USA). Abstract Digital human models (DHMs) are critical for improved designs, injury prevention, and a better understanding of human behavior. Although many capabilities in the field are maturing, there are still opportunities for improvement, especially in motion prediction. Thus, this work investigates the use of an artificial neural network (ANN), specifically a general regression neural network (GRNN), to provide real-time computation of DHM motion prediction, where the underlying optimization problems are large and computationally complex. In initial experimentation, a GRNN is used successfully to simulate walking and jumping on a box while using physics-based human simulations as training data. Compared to direct computational simulations of dynamic motion, use of GRNN reduces the calculation time for each predicted motion from 1-40 minutes to a fraction of a second with no noticeable reduction in accuracy. This work lays the foundation for studying the effects of changes to training regiments on human performance. Keywords: digital human modeling, artificial neural networks, motion prediction, general regression neural network. © 2015. This manuscript version is made available under the Elsevier user license http://www.elsevier.com/open-access/userlicense/1.0/ mailto:bataineh.moe@gmail.com mailto:tmarler@engineering.uiowa.edu mailto:amalek@engineering.uiowa.edu mailto:jasbir-arora@uiowa.edu mailto:bataineh.moe@gmail.com http://ees.elsevier.com/eswa/download.aspx?id=484997&guid=e16b88ea-4974-4c6b-9772-fe4ee8aff1ba&scheme=1 http://ees.elsevier.com/eswa/viewRCResults.aspx?pdf=1&docID=35104&rev=2&fileID=484997&msid={9419EB0E-D767-426A-B90D-966F8070A340} 2 1. Introduction The use of digital humans is becoming more prevalent with the upstream design of any product or process that involves human interaction or human systems integration. In order to enhance the design process as effectively as possible, computationally fast human simulation and analysis become critical. The faster one can simulate and obtain feedback concerning human performance, the faster one can evaluate and refine designs. A cornerstone of human performance is simulated motion, which often requires dynamic analysis (consideration of forces, acceleration, and inertia, not just kinematics). However, accurate dynamic simulation and analysis can be computationally demanding, depending on the task being simulated. Therefore, this work presents the use of an artificial neural network (ANN) to provide fast motion simulation. Whether the motion tasks are simulated using data-based methods (Chaffin, 2002; Moeslund, Hilton, & Krüger, 2006), which depend on motion capture systems to track, record, and reproduce human motion during various tasks, or using physics-based methods (Xiang, Chung, et al., 2010), which primarily entail using optimization to predict motion, techniques for capturing one’s history and the consequent strategy for completing the task continue to be a challenge. Why do different people with similar capabilities and size, for instance, enter the same vehicle in different ways? Although various human modeling methods can capture the nuances of one’s motion or the cause and effect demonstrated with changes in parameters, few methods offer the ability to capture one’s strategy in approaching a task without significant input from the user. Hence, there is a need for an algorithm that produces real-time motion prediction and can incorporate a history of experience. Motion-capture-based methods are limited in terms of their ability to produce different motions that correspond to changes in the task parameters, because the underlying algorithm depends on prerecorded data that cannot be changed due to the change in the task conditions. In addition, these methods do not incorporate dynamics; they do not capture effects of loads and inertia. Alternatively, physics-based methods like predictive dynamics (PD), which is an optimization-based motion prediction algorithm (Xiang, Chung, et al., 2010), tend to be more flexible in showing the effects of changes in task parameters, especially with respect to dynamics. In addition, these methods are predictive; they predict human performance with minimal dependence on prerecorded data. Computational speed, however, can be a limiting factor with PD, when real-time performance feedback is needed. Depending on the task being simulated and its settings, PD can require up to 40 minutes (the PD algorithm is run on a Windows 7 computer with an Intel ® Core ™ i3 processor and 8 GB of RAM), even with small changes in the configuration. A variety of techniques are being explored to address this challenge, and the use of an ANN is especially promising. An ANN can be used for real-time motion prediction and can be integrated with the physics-based motion simulation method, PD, in order to improve the computational speed when predicting a motion task. An ANN also provides a platform for incorporating alternative sources for real-time motion prediction, such as motion capture data, if desired. Unlike other simulation tools, ANNs are capable of providing acceptable simulations for a problem without the need for complex time-consuming algorithms. This work 1) demonstrates the feasibility and advantages of using ANNs for direct motion prediction, and 2) presents the use of one of the ANN types as an appropriate network for successful simulation of such problems. An ANN is a mathematical model for predicting system output, inspired by the structure and function of human biological neural networks. Compared to other simulation and statistical http://en.wikipedia.org/wiki/Mathematical_model http://en.wikipedia.org/wiki/Biological_neural_networks 3 tools, ANN is fast and produces relatively accurate and acceptable simulations for complex systems. ANNs entail two steps or processes. First they are trained using some form of pre- existing data. Essentially, optimization is used to set model parameters. After its training process is completed, it is then run and provides relatively fast output given various input conditions. ANNs can be powerful tools for generalizing, which means providing accurate and acceptable results for all inputs conditions, many practical problems (Coit, Jackson, & Smith, 1998; Twomey & Smith, 1998), and hence have been used successfully in many digital human model (DHM)-related problems (Bataineh, 2015; Bataineh & Marler, 2013; Bataineh, Marler, & Abdel- Malek, 2013; Bu, Okamoto, & Tsuji, 2009; Kang, Kim, Park, & Kim, 2007; Li, Li, & Song, 2007; Zhang, Horváth, Molenbroek, & Snijders, 2010; Zhao, Zheng, & Wen, 2010). Motion- related applications include, but are not limited to, robotics and controller system motion, motion analysis, reconstruction of dynamic objects, and time-series dynamic prediction and classification. In general, the main use of ANNs has been focused on human model posture prediction (Jung & Park, 1994; Zhang, et al., 2010) and motion prediction of robotics and dynamics systems (Frank, Davey, & Hunt, 2001; Stakem & AlRegib, 2008). One approach proposed the use of multiple ANNs in controlling a robot manipulator (Y. H. Kim & Lewis, 1999). The system was evaluated successfully on a two-link robot manipulator, demonstrating the ANN’s ability to handle the nonlinear unknown parameters in the system manipulator. Moreover, the feedforward-backpropagation network, which was trained by gaits of various people, is used to recognize humans automatically (Yoo, Hwang, Moon, & Nixon, 2008). In other motion-related applications, ANNs have been used as an indirect source that led to providing improved motion predictions. Lung tumor motion during respiration was predicted in advance using ANNs (Isaksson, Jalden, & Murphy, 2005). Most of the preceding scholars use feedforward-backpropagation ANNs with single or few outputs to preserve accuracy. So far, ANNs have been applied only to very specific scenarios in DHM problems and have not been developed for robust use with complex problems like whole-body dynamic motion prediction. In general, ANNs have been used to solve confined systems with a relatively small number of inputs and outputs. Most applications have involved feedforward-backpropagation networks, which have memory and accuracy issues when used with a large number of inputs and outputs. Thus, this work explores using other types of ANNs for relatively large and complex human-modeling problems. The overarching hypothesis is that if designed/selected properly, ANNs can in fact be used to simulate human motion quickly and accurately, despite a relatively large number of outputs. We contend that a radial-basis network (RBN) is most appropriate, because it has the smallest number of parameters to be set when simulating a problem with a large number of outputs. In addition, it provides a global solution when optimizing the network parameter values (during the training process). Specifically, we propose using a general regression neural network (GRNN), which is a type of RBN. The work integrates GRNN with PD to increase the computational speed of PD with minimal detriment to accuracy. This is shown using two different simulated tasks with a large number of outputs (on the order of hundreds), both of which run in under one second. Eventually, PD can be replaced by GRNN, which is trained to provide a standalone instant motion simulation. The next section describes the necessary parameters (inputs and outputs) of the PD simulated tasks, as well as details of the underlying ANN architecture. 4 2. Methods 2.1. Digital human model and physics-based motion prediction As a foundation for the proposed method, this section summarizes the digital human model as well as PD. This work capitalizes on and adds to a foundation of virtual human modeling capabilities, housed within a human model called Santos (Abdel-Malek, et al., 2006; Abdel-Malek, et al., 2007). Santos, as shown in Fig. 1, is a highly realistic, biomechanical computer-based human that predicts, among other things, static posture, dynamic motion, joint strength, and development of fatigue. Such capabilities can be used to predict and assess human function, providing task performance measures and ergonomic analysis. Thus, in a virtual world, Santos can help design and analyze various products and processes. In addition, Santos can help study and evaluate various restrictions and impediments, such as fatigue, reduced range of motion, environmental obstacles, etc. Fig. 1. Schematic of the skeleton model for the virtual human model Santos. A key aspect of any virtual human is the ability to simulate human posture and motion realistically and quickly while considering external and internal loads/forces. With respect to motion, there are traditionally two types of dynamics problems that need to be addressed. In the first problem, called forward dynamics, the external forces and torques on the system are known and the motion of the system is desired. The problem is solved by integrating the governing equations of motion forward in time using a numerical algorithm. In the second problem, called inverse dynamics, the motion of the system is known (i.e., from motion capture), and the forces and torques causing the motions are calculated using the equations of motion. Both of these problems can be solved using traditional multi-body dynamics software. The problem of predictive dynamics arises when one wants to simulate the human motion for any task. In this problem, both the joint torques and the motion of the joint are unknown. Therefore, the problem becomes more difficult to solve. With the PD approach, the joint angles (one for each degree of freedom, or DOF) essentially provide design variables that are determined through optimization. The objective function(s) is one or more human performance measure, such as energy 5 consumption, discomfort, and joint displacement. Including the dynamic equations of motion as constraints then ensures that the laws of physics are satisfied. The specifics of PD (Xiang, Arora, Rahmatalla, & Abdel‐Malek, 2009; Xiang, Chung, et al., 2010) are summarized as follows. In general, predicting dynamic human motion is approached as an optimization problem (Arora, 2004), as shown in Equation 1. This formulation provides the context for the discussion of inputs and outputs used with the proposed neural network. Joint angle profiles over time are represented as B-spines, and the control points, which dictate the shape of the profile curve, serve as design variables in an optimization problem. The problem entails determining design variables , which represent the control points (i.e., joint angle profiles) of all body DOFs, in order to minimize the objective function, , subject to the physical equality ( ) and inequality ( ) constraints. The control points ( ) form B- splines for all DOFs that simulate the motion of the DHM. Having more control points in a B- spline leads to more accurate motion simulation but increases the number of design variables and can thus increase computational complexity. is a vector that represents the reference motion provided by motion capture. Find: (control points for 55-DOFs) Minimize: Subject to: Equations of motion, including reaction forces (1) The objective function can vary slightly depending on the task being simulated, but it generally includes two components. First, the difference between the predicted motion q and a seed motion based on experimental motion capture is minimized ( ). This component helps represent one’s overall strategy when performing a task, as opposed to kinematic and dynamic nuances. The second component includes the tendency to minimize the joint torque being used to complete a task and thus minimize the energy being used ( ). The constraints, which represent and , include the following: contact points (between the avatar and the environment), joint-angle limits that represent one’s range of motion (Marler et al., 2008), torque limits that represent one’s strength limits, restriction on the zero moment point (responsible for balance), ground reaction forces, and equations of motion. Since its development, PD has been used to simulate different motion tasks and scenarios (J. H. Kim, et al., 2008; J. H. Kim, Xiang, Yang, Arora, & Abdel-Malek, 2010; Kwon, et al., 2014; Xiang, Arora, & Abdel-Malek, 2010, 2012; Xiang, Arora, Rahmatalla, et al., 2010). The inputs for each simulation take two forms: avatar-based and task-based. Avatar-based inputs include anthropometric parameters (i.e., skeletal link lengths, mass for body segments, etc.), joint ROMs, torque limits, which represent strength limits, and loads applied to the avatar as a result of external factors. Task-based inputs include parameters that define the characteristics of a task (i.e., step size in a walking task, box height in a jumping-on-the-box task, etc.). Each newly developed PD task/simulation is validated using motion capture and force plates (Rahmatalla, Xiang, Smith, Meusch, & Bhatt, 2011). Creating a large number of different task conditions (thousands) is difficult, because it can be time consuming. The running time for each case, even with minor condition changes, can take minutes to hours in order to complete and produce the simulation. The long running time is due to the large size of the PD problem with respect to the number of outputs. The number of 6 outputs in a PD task is approximately 500-700 outputs (i.e., the outputs represent for all 55 DOFs, which are the design variables), depending on the number of control points in each task, which can slow down the optimization. In addition, this process cannot be automated completely, because the simulations require some post-processing before considering the case acceptable and usable. Hence, the time-cost issue leads to a PD problem with a limited number of simulations available to be used by pattern recognition tools like ANN to be trained to provide real-time PD simulations. Therefore, a carefully selected type of ANN that is capable of being trained well with a limited number of simulations needs to be employed to provide the most acceptable simulation results. The following section illustrates the architecture of the proposed ANN for successful real-time simulation of a PD problem that has a limited number of training cases. 2.2. General regression neural network (GRNN) This work proposes the use of a GRNN for simulating motion prediction in DHM. Among the types of RBN, which is known to be powerful when simulating large-scale and complex problems like motion, GRNN has the fewest parameters to be set among RBN types for successful use in DHM applications with relatively large problems (Bataineh, 2012). ANNs consist of three main parts (Fig. 2): (1) an input layer, (2) a hidden layer, and (3) an output layer. The input and output layers consist of the system’s inputs and outputs, respectively. The hidden layer represents the core of the ANN and consists of units called neurons. Inside the hidden neurons, which are also called the basis functions, the main mathematical calculations occur while processing the inputs and providing the proper outputs. Fig. 2. General regression neural network architecture. The network input is represented by ] and provides the input for each neuron in the hidden layer. R is the number of inputs, Q is the number of training cases as well as the hidden neurons [ ], and N is the number of outputs. The hidden neuron in the ANN receives input(s) from the input layer, completes a mathematical transformation/calculation, and sends the result to the neuron(s) of the output layer. Inside each hidden neuron, there is a radial transfer function that produces outputs depending on the provided input (Wasserman, 1993). Once the i th hidden neuron receives the input x, the Euclidian norm of the difference between and the hidden neuron’s center is 7 calculated to produce (Equation 2). Then, the value is multiplied by the bias constant B to provide , which is called the radial distance (Equation 3). The radial function output is then calculated as a function of , (Equation 4). (2) (3) (4) Each output neuron receives the hidden neuron’s output as its input. The output neuron essentially combines the output of all hidden neurons in a weighted sum to provide the final network output (Equation 5). The vector in the k th output neuron represents the weight associated with that neuron necessary to provide the proper value for the k th output . The output layer has N neurons, where N is the number of outputs . (5) The training process in a GRNN involves determining the most appropriate value of each hidden neuron center and each output weight vector . The training cases are used as the hidden neuron centers, where the number of hidden neurons equals the number of training cases Q. The weighting vector is set as the outputs of the training cases. More details regarding the GRNN architecture are provided in the literature (Specht, 1991). 2.3. GRNN for Motion Simulation The crux of the proposed method entails using an ANN to approximate dynamic motion prediction. In essence, an ANN is used to create a meta-model or hyper response surface of the PD computational model. Combining these two elements provides two modes of use. First, given that PD entails running a gradient-based optimization model, and ANN can be used to provide an initial guess for the optimization problem. With this mode, PD is run first, regardless of the computational demands, and is used to create a library of base simulations. These simulations are then used to train an ANN. With subsequent PD simulations, the ANN is used to provide an initial guess based on simulation parameters (i.e., avatar anthropometry, loads applied to the avatar from equipment, etc.). This initial guess then helps increase the speed of the optimization problem and thus the simulation. On the other hand, even with an appropriate (i.e., close to optimal) initial guess that is provided from the ANN, the complex nonconvex optimization in the PD algorithm is not always guaranteed to run faster. We contend that the results of the ANN can actually be used directly as a final solution in and of itself, and subsequent results prove this hypothesis. This process in inherently faster than running PD, training an ANN, and then running PD again. Alternatively, PD is run off line to produce training data, and the ANN is run to produce final results. Task characteristics (i.e. walking speed, jumping height, etc.) and avatar characteristics provide the input for a GRNN. Joint-angle profiles, joint-torque profiles, and ground reaction forces are the output (see Equation 1). These parameters are first used to train a GRNN (one for each task). Then, conceivably, any set of input values can be used to run the GRNN and extract associated output values (simulation results). This in turn provides a new and relatively fast method for human motion prediction. Details regarding the inputs and outputs for the tasks of walking and jumping on a box are provided in the subsequent section. 8 3. Results The use of GRNN to simulate dynamic human motion quickly is applied and tested on two tasks: walking forward with a backpack, and jumping up on a box. The approximate results provided by the GRNN are compared to, and verified with those provided from the source predictive-dynamics models. The work of developing/running the training cases, training the network, and then executing/running the trained network is completed on a Windows 7 computer with an Intel ® Core ™ 2 processor and 8 GB of RAM. Since there is no optimization procedures in the GRNN training process, it is trained within approximately 2-5 seconds, where that includes setting the network parameters to their appropriate values (as indicated in Section 2.2). With the presented tasks, the run time for PD ranges between 1 and 40 minutes. The network- predicted (approximated) motion outputs, which are all produced in a fraction of a second, are presented, evaluated, and compared with those exact (true) outputs from PD. After the network is trained with cases produced from PD, the network can simulate motions instantly for new conditions (i.e., test cases) that had never been used to train the network. To evaluate the network results of the test cases, all the produced motions from the GRNN and PD are compared visually and objectively. 3.1. Example 1: Walking The first simulated task is walking forward with a backpack (Xiang, et al., 2009). The inputs for the task include: walking velocity, backpack weight, four lower-body link lengths (spine to hip, hip to knee, knee to ankle, and ankle to football), and three body joint ROMs. ROM is specified as upper and lower limits for the hip, knee, and ankle, each at flexion and extension. These specific joint ROMs are used (note that many others are stipulated as detailed by Marler et al., 2008), because changing their limits has significant effects on the resulting motion, whereas other ROMs have a relatively small effect on the task. Table 1 shows a typical range of values, for each input parameter, that reflect various scales of human anthropometric data. The training cases are formed as various combinations of these values, although not all permutations are used simply to reduce complexity. Velocity represents the speed of walking for Santos and is measured in m/sec. Value 1 and Value 2, respectively, typical minimum and maximum speeds for a person of average height. Each training case (set of input values) represents a point on a training grid. The consequent trained ANN can then be run using any set of inputs. Input values that do not mimic any training case represent what is called a test case (i.e., off-grid point). Table 1 Input parameters for training the walking-forward task. Input parameter Value 1 Value 2 Value 3 Velocity (m/s) 0.8 -- 1.6 Backpack weight (N) 0 175 315 Link1 (Spine to Hip) (cm) 7.8 8.8 9.8 Link2 (Hip to Knee) (cm) 43.5 44.5 45.6 Link3 (Knee to Ankle) (cm) 39.5 42.4 45.4 Link4 (Ankle to Football) (cm) 11.3 11.7 12.1 Joint1 (Hip)- lower limit (degrees) -123.3 -105 -90 Joint1 (Hip)- upper limit (degrees) 8.7 5 2 Joint2 (Knee)- lower limit (degrees) 5 10 20 Joint2 (Knee)- upper limit (degrees) 149.7 130 110 9 Joint3 (Ankle)- lower limit (degrees) 7.3 15 20 Joint3 (Ankle)- upper limit 71.6 60 50 Thus, for the first training case, the velocity is set to Value 1, the backpack weight is set to Value 1, the group of avatar inputs correspond to the link lengths (the four inputs for body segments’ lengths) are set to Value 1, and the group of avatar inputs correspond to the joint limits (the six inputs for upper and lower joint limits) are set to Value 1. The next training case is created by keeping the backpack weight, the velocity, and the group of inputs correspond to the link lengths fixed at Value 1, and moving the group of avatar joint limits to Value 2. Then, the avatar joint limits are set to Value 3 to create the third training case. The same process is applied to create the next 3 training cases by setting velocity and backpack weight to Value 1, link lengths to Value 2, and using joint limits iteratively from Value 1 to Value 3. The following 3 training cases are created using the previous settings, except the link lengths are set to Value 3. By the end of this step, nine training cases have been produced so far. Next, the velocity is set to Value 1, the backpack weight is set to Value 2, and the previous steps are repeated to create the next new nine training cases. Then, another nine training cases are created using Value 3 for the backpack weight. Thus far, 27 training cases have been produced after all the input values are used with fixed velocity at Value 1. Setting the velocity to Value 2 and repeating all the aforementioned procedures produces another set of 27 cases. Finally, after a manual post- processing of the resulting 54 training cases, two training cases are removed because they are subjectively unacceptable due to visually odd motions. This results in 52 training cases in total. The network’s outputs in a PD task include control points for Santos’s joint-angle profiles for all 55 DOFs, as well as values for joint torques at specified time steps for certain specified DOFs. With the walking task, each joint profile consists of six control points that represent joint values at different times over the task. Hence, there are 330 outputs representing the joint angle profiles. Joint torques are considered for the six lower-joint DOFs (three for the hip, one for the knee, and two for the ankle), because these are the most highly articulated DOFs during the walking task. Assuming symmetry, the joint torques are evaluated at ten time steps during the walking task. This results in 60 additional output values. Thus, to summarize, the task is defined with 12 inputs and 390 outputs in total, and there are 52 training cases. Once the ANN has been trained, the ANN model for the walking task is tested using the conditions shown in Table 2. Each of these testing cases is analyzed by comparing the predicted network outputs and the exact PD results. The test cases include two off-grid points. Note that these cases are not used to train the network, so they help evaluate the general performance of the network prediction for off-grid (i.e., new unseen) test cases. These cases are chosen to cover different input combinations and with input values that are completely different from the values in the training cases (Table 1). The results of the test case are evaluated based on subjective motion, and on objective evaluation of joint angle profiles and joint torques. Table 2 Input variables for two test cases. Input parameter Case 1 Case 2 Velocity (m/s) 1.4 0.9 Backpack weight (N) 220 63 Link1 (Spine to Hip) (cm) 9 8 Link2 (Hip to Knee) (cm) 44 43 Link3 (Knee to Ankle) (cm) 41.4 45 10 Link4 (Ankle to Football) (cm) 12 11.3 Joint1 (Hip)- lower limit (degrees) -98.6 -111 Joint1 (Hip)- upper limit (degrees) 2.2 6.5 Joint2 (Knee)- lower limit (degrees) 8 16 Joint2 (Knee)- upper limit (degrees) 146 127.2 Joint3 (Ankle)- lower limit (degrees) 12 7 Joint3 (Ankle)- upper limit 57 70 As a basic subjective analysis, the visual motion results for Case 1 and Case 2 are presented in Fig. 3. For this and all following cases, snapshots are taken at three different time frames of the total task time. The visual results for Case 1 show accurate network prediction of all joint profiles when compared to PD results; the motion produced from the network is visually comparable to that from the PD. With Case 2, Santos’s back is almost straight, which is a reflection of the input backpack weight of just 63 N. The relatively short step size reflects the relatively low walking velocity. Generally speaking, the visual subjective results are all realistic. Fig. 3. Visual walking task results for test Case 1 and Case 2. As mentioned earlier, 330 of the network’s outputs represent joint angle profiles for the 55 DOFs. Statistical comparison is performed on these outputs for both testing cases. The mean- absolute error (MAE) values are calculated for joint angle profiles between the predicted values from the GRNN and the exact ones from the PD. The MAE are 0.029 and 0.033 for Case 1 and Case 2, respectively, which are relatively small, thus indicating agreement between the ANN results and the physics-based results (using predictive dynamics). As another objective test for the results, adjusted R-square values are calculated between the predicted GRNN results and the exact PD ones (Fig. 4). In both cases, accurate results are achieved; the R-squared values are above 0.99, and the visual results show minimal discrepancies. The slight increase in R-squared for Case 1 is simply the result of the neural- network hypersurface providing slightly more accurate results for different points (sets of inputs). Interestingly, the scatter plot for Case 1 appears to suggest a decrease in accuracy. 11 However, the MAE result for this case provides the evidence for that accuracy level that matches the resulting R-squared value. Both results confirm the high correlation between the predicted and exact results. In addition, given that there are 330 data points, the number of points that visually deviate from the line in Case 1 is relatively small and numerically insignificant. In other words, in Case 1, the network produces different solutions for some of the joint angles that have minimal effect on the total motion behavior. Fig. 4. Statistical plots and adjusted R-square values for the produced joint angle profiles in test Case 1 and Case 2 in the task of walking forward. As with joint angle profiles, the MAEs are calculated for the produced joint-torque profiles. The results are 10.01 for Case 1 and 9.16 for Case 2. The adjusted R-square values are also plotted and calculated (see Fig. 5). Recall that 60 of the output values relate to torque profiles for specified DOFs. The values are approximately 0.96 and 0.88 for Cases 1 and 2, respectively. These calculated values are relatively high and generally acceptable. They are slightly different from those in Fig. 4, in part because the scale of the output is different, which can affect the accuracy of results produced from the network. Although inputs are already scaled, future work might include investigating the effects of normalizing output values during the training process. Fig. 5. Statistical plots and adjusted R-square values for the produced joint torque profiles in test Case 1 and Case 2 in the task of walking forward. On the other hand, these MAEs and R-square results are not quite as accurate as those for the joint angles, because each joint has a large range of torque values compared to the joint angle values. Moreover, there are many ranges for torque values at different joints, while the joint angles were all measured in radian (rad) and fell between +6.28 rad to -6.28 rad. Therefore, the network can predict the joint angle values, which are consistent for all DOFs, with less error compared to the variety joint torque range of values. 12 3.2. Example 2: Jumping on a Box The second example task for evaluating GRNN performance is jumping up on a box (Fig. 6). This is a relatively complex task that has many constraints (i.e., feet and hands should be located at specific places) and requires highly accurate results. Box height is the primary task- based input, because it has the greatest effect on the resulting motion profile. When training the GRNN, the box height ranges between 0.5 meter and 1 meter. Avatar-based inputs include four link lengths: spine-hip, hip-knee, knee-ankle, and ankle-football. Minimum and maximum values for these inputs are shown in Table 3. In order to generate training cases, first input values are set at their minima, and box height is increased in increments of 5 cm with all other parameters fixed. Then, input values are set at their maxima, and box height is decreased in increments of 5 cm. This results in 22 training cases. Fig. 6. Task of jumping up on a box. Table 3 Input parameters and training values for the task of jumping up on a box. Input parameter Minimum Maximum Box height (cm) 50 100 Link1 (Spine to Hip) (cm) 7.8 9 Link2 (Hip to Knee) (cm) 38 43 Link3 (Knee to Ankle) (cm) 39 39 Link4 (Ankle to Football) (cm) 9 12 The jumping task has two types of outputs: joint splines and ground reaction forces (GRF) for both feet. The total number of outputs is 370, with 330 outputs for joint splines and 40 outputs for GRFs. As with the walking example, two off-grid testing cases (Table 4) are evaluated and analyzed visually and statistically. The visual results for Case 1 and Case 2 are compared in Fig. 7: the left and right parts in each case represent motion segments over the task time for PD and GRNN, respectively. In Case 1, even though there were some differences for both motions due to the recording procedures of the snapshots, the motions had the same behavior over the task time, including the height of the feet and the hand positions. In Case 2, Santos’s hands and feet from the network’s predicted motion were also at the exact locations that the PD results provided. 13 Table 4 Input variables for two test cases. Input parameter Case 1 Case 2 Box height (cm) 0.68 0.92 Link1 (Spine to Hip) (cm) 8.2 8.8 Link2 (Hip to Knee) (cm) 40 42 Link3 (Knee to Ankle) (cm) 39 39 Link4 (Ankle to Football) (cm) 10 11 Fig. 7. Visual results for test Case 1 and Case 2 in the task of jumping up on a box. With respect to statistical evaluation, comparison between predicted GRNN outputs and exact PD ones is performed for both testing cases. The MAE values are calculated, and the results show small values with 0.017 and 0.015 for Case 1 and Case 2, respectively. In terms of adjusted R-square values, the results are shown in Fig. 8. The R-square values for angle profiles are approximately 1 for both testing cases, which suggests high accuracy of motion prediction produced from the GRNN. Fig. 8. Statistical plots and adjusted R-square values for the produced joint angle profiles in test Case 1 and Case 2 in the task of jumping up on a box. In terms of the GRNN prediction results for the GRFs, the results comparison with those provided from the PD show high R-square values for both Case 1 and Case 2 (Fig. 9). In addition, the MAE results for the GRF are calculated and the results are 0.38 and 0.53 for Case 1 14 and Case 2, respectively. In general, the statistical results suggest acceptable network-predicted solutions for the GRFs, and such results are enhanced by the relatively small simulation errors. Fig. 9. Statistical plots and adjusted R-square values for the produced ground reaction forces in test Case 1 and Case 2 in the task of jumping up on a box. 4. Discussion This work has demonstrated the successful use of ANNs for simulating dynamic human motion in real time, based on a library of physics-based simulations. The use of an ANN is shown to provide real-time motion prediction for a full-body DHM, using various tasks. The GRNN, in particular, is applied successfully as a powerful tool that can be trained quickly and without any memory problems, regardless of the size of the problem. This provides a new variant of physics-based human simulation with increased computational speed, which in turn allows one to conduct trade-off analyses more efficiently for evaluating products and processes form a human system integration perspective. As with any regression analysis or meta-model, the approximate solution is expected to deviate at least some minimal amount from the source model. We show that this deviations is acceptable when evaluated both objectively and subjectively. As is often the case with digital human modeling, although subjective results may seem identical when compared, quantitative results may differ slightly. Nonetheless, the ANN model performed well when compared to the predictive-dynamics source model. With the use of ANNs for motion prediction, the main issue that was successfully addressed is the computational time of producing PD outputs. The time is decreased from minutes to a fraction of a second. Even with a task that requires highly accurate results like the jumping on a box one, the simulation accuracy, as demonstrated, was preserved. Accuracy in such task is critical, especially where the hands and feet should be in contact with the box at some point during the task, and network results showed visual satisfaction of such constraints. Conceptually, the proposed process reflects how people actually learn and perform tasks. One’s reaction to specific scenarios is an interpolation, in part, of previously experienced scenarios. To be sure, rational thought and cognitive extrapolation play a role. However, the proposed method demonstrates a model for incorporating a learned history into task simulation. In fact, it provides a platform with which one can study how various experiences can affect performance and how variations in one’s learning set can alter a simulation. This has practical applications to the study and design of training regiments and education. This initial investigation into the use of GRNN for human modeling has presented some opportunities for future work. First, as is the case with most ANN applications, work is needed to determine the optimal number of training cases for a task. In order to ensure contact constraints 15 are satisfied, future work should include adding constraints to the network construction. In addition, we propose training the network to predict joint-center locations instead of joint angles, in order to produce more accurate results. Predicting joint angles with even small errors can sometimes produce inaccurate results, whereas predicting joint centers with small error should still provide acceptable results. Although the network is trained using simulations (predictive dynamics) that are presumably validated, and although this work presents the feasibility of the proposed method based on subjective validation, predictions from the actual ANN should be validated directly and objectively. Finally, as suggested with respect to joint torques, further work is needed to investigate the benefits of normalizing outputs during the training process. Acknowledgments This work is funded by the Office of Naval Research (ONR). The authors would also like to thank the team at The University of Iowa’s Virtual Soldier Research (VSR) Program. References Abdel-Malek, K., Arora, J., Yang, J., Marler, T., Beck, S., Swan, C., Frey-Law, L., Mathai, A., Murphy, C., & Rahmatalla, S. (2006). Santos: A physics-based digital human simulation environment. In Proceedings of the Human Factors and Ergonomics Society Annual Meeting (Vol. 50, pp. 2279- 2283): SAGE Publications. Abdel-Malek, K., Yang, J., Kim, J. H., Marler, T., Beck, S., Swan, C., Frey-Law, L., Mathai, A., Murphy, C., & Rahmatallah, S. (2007). Development of the virtual-human SantosTM. In Digital Human Modeling (pp. 490-499): Springer. Arora, J. S. (2004). Intriduction to Optimum Design. 2 nd Edition, Elsevier Academic Press, Amsterdam. Bataineh, M. (2012). Artificial neural network for studying human performance. M.S. thesis, University of Iowa. Bataineh, M. (2015). New neural network for real-time human dynamic motion prediction. Ph.D. thesis, University of Iowa. Bataineh, M., & Marler, T. (2013). Neural networks for performance-measure selection. In 2nd International Digital Human Modeling Symposium. Ann Arbor, MI-USA. Bataineh, M., Marler, T., & Abdel-Malek, K. (2013). Artificial neural network-based prediction of human posture. In Digital Human Modeling and Applications in Health, Safety, Ergonomics, and Risk Management. Human Body Modeling and Ergonomics (pp. 305): Springer. Bu, N., Okamoto, M., & Tsuji, T. (2009). A hybrid motion classification approach for EMG-based human–robot interfaces using bayesian and neural networks. Robotics, IEEE Transactions on, 25, 502-511. Chaffin, D. B. (2002). On simulating human reach motions for ergonomics analyses. Human Factors and Ergonomics in Manufacturing & Service Industries, 12, 235-247. Coit, D. W., Jackson, B. T., & Smith, A. E. (1998). Static neural network process models: considerations and case studies. International Journal of Production Research, 36, 2953-2967. Frank, R. J., Davey, N., & Hunt, S. P. (2001). Time series prediction and neural networks. Journal of Intelligent and Robotic Systems, 31, 91. Isaksson, M., Jalden, J., & Murphy, M. J. (2005). On using an adaptive neural network to predict lung tumor motion during respiration for radiotherapy applications. Medical physics, 32, 3801. Jung, E. S., & Park, S. (1994). Prediction of human reach posture using a neural network for ergonomic man models. Computers & industrial engineering, 27, 369-372. Kang, B., Kim, B., Park, S., & Kim, H. (2007). Modeling of artificial neural network for the prediction of the multi-joint stiffness in dynamic condition. In Intelligent Robots and Systems, 2007. IROS 2007. IEEE/RSJ International Conference on (pp. 1840-1845): IEEE. Kim, J. H., Xiang, Y., Bhatt, R., Yang, J., Chung, H.-J., Patrick, A., Arora, J. S., & Abdel-Malek, K. (2008). Efficient ZMP formulation and effective whole-body motion generation for a human-like 16 mechanism. In ASME 2008 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference (pp. 1073-1084): American Society of Mechanical Engineers. Kim, J. H., Xiang, Y., Yang, J., Arora, J. S., & Abdel-Malek, K. (2010). Dynamic motion planning of overarm throw for a biped human multibody system. Multibody System Dynamics, 24, 1-24. Kim, Y. H., & Lewis, F. L. (1999). Neural network output feedback control of robot manipulators. Robotics and Automation, IEEE Transactions on, 15, 301. Kwon, H.-J., Xiang, Y., Bhatt, R., Rahmatalla, S., Arora, J. S., & Abdel-Malek, K. (2014). Backward walking simulation of humans using optimization. Structural and Multidisciplinary Optimization, 1-11. Li, Y., Li, C., & Song, R. (2007). A new hybrid algorithm of dynamic obstacle avoidance based on dynamic rolling planning and RBFNN. In Robotics and Biomimetics, 2007. ROBIO 2007. IEEE International Conference on (pp. 2064-2068): IEEE. Marler, T., Arora, J., Beck, S., Lu, J., Mathai, A., Patrick, A., Swan, C. (2008), “Computational Approaches in DHM,” in Handbook of Digital Human Modeling for Human Factors and Ergonomics, Vincent G. Duffy, Ed., Taylor and Francis Press, London, England. Moeslund, T. B., Hilton, A., & Krüger, V. (2006). A survey of advances in vision-based human motion capture and analysis. Computer vision and image understanding, 104, 90-126. Rahmatalla, S., Xiang, Y., Smith, R., Meusch, J., & Bhatt, R. (2011). A validation framework for predictive human models. International journal of human factors modelling and simulation, 2, 67-84. Specht, D. F. (1991). A general regression neural network. Neural Networks, IEEE Transactions on, 2, 568. Stakem, F., & AlRegib, G. (2008). Neural Networks for Human Arm Movement Prediction in CVEs. In Proceedings of 3DPVT. Twomey, J. M., & Smith, A. E. (1998). Bias and variance of validation methods for function approximation neural networks under conditions of sparse data. Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, 28, 417-430. Wasserman, P. D. (1993). Advanced methods in neural computing: John Wiley & Sons Inc. Xiang, Y., Arora, J. S., & Abdel-Malek, K. (2010). Physics-based modeling and simulation of human walking: a review of optimization-based and other approaches. Structural and Multidisciplinary Optimization, 42, 1-23. Xiang, Y., Arora, J. S., & Abdel-Malek, K. (2012). Hybrid predictive dynamics: a new approach to simulate human motion. Multibody System Dynamics, 28, 199-224. Xiang, Y., Arora, J. S., Rahmatalla, S., & Abdel‐Malek, K. (2009). Optimization‐based dynamic human walking prediction: One step formulation. International Journal for Numerical Methods in Engineering, 79, 667-695. Xiang, Y., Arora, J. S., Rahmatalla, S., Marler, T., Bhatt, R., & Abdel-Malek, K. (2010). Human lifting simulation using a multi-objective optimization approach. Multibody System Dynamics, 23, 431- 451. Xiang, Y., Chung, H.-J., Kim, J. H., Bhatt, R., Rahmatalla, S., Yang, J., Marler, T., Arora, J. S., & Abdel- Malek, K. (2010). Predictive dynamics: an optimization-based novel approach for human motion simulation. Structural and Multidisciplinary Optimization, 41, 465. Yoo, J.-H., Hwang, D., Moon, K.-Y., & Nixon, M. S. (2008). Automated human recognition by gait using neural network. In Image Processing Theory, Tools and Applications, 2008. IPTA 2008. First Workshops on (pp. 1): IEEE. Zhang, B., Horváth, I., Molenbroek, J. F., & Snijders, C. (2010). Using artificial neural networks for human body posture prediction. International Journal of Industrial Ergonomics, 40, 414-424. Zhao, H., Zheng, G., & Wen, W. (2010). Human-machine posture prediction and working efficiency evaluation of virtual human using radial basis function neural network. In Intelligent Computing 17 and Intelligent Systems (ICIS), 2010 IEEE International Conference on (Vol. 1, pp. 406-410): IEEE. 
work_ybizrrzcsreqdjy7yzi7k22t6a	----	Document downloaded from: This paper must be cited as: The final publication is available at Copyright http://dx.doi.org/10.1016/j.eswa.2013.07.111 http://hdl.handle.net/10251/50837 Elsevier García García, I.; Sebastiá Tarín, L. (2014). A negotiation framework for heterogeneous group recommendation. Expert Systems with Applications. 41(4):1245-1261. doi:10.1016/j.eswa.2013.07.111. A Negotiation Framework for Heterogeneous Group Recommendation Inma Garcia a,1,∗, Laura Sebastia a aUniversidad Politecnica de Valencia, Spain Abstract Over the last years, some remarkable recommender systems for group of users have been developed. When using most of these systems, each group member communi- cates his/her preferences to the system, which obtains a group profile as the result of an equal weighting of the individual preferences. This way, no member is particu- larly dissatisfied with the recommendations. However, this is not a realistic situation, given that not all the members in a group act in the same manner. This paper deals with the problem of recommendation for a group of users, where, besides his/her own preferences, each user may have different expectations about the result of the recommendation and may exhibit a different behaviour with respect to the other group members. Moreover, all this information is private and may be revealed un- der certain circumstances. In this context, we have opted for building a multi-agent system, where an agent acts on behalf of one group member. We have implemented a UserAgent that can be configured in order to exhibit the behaviour desired by the corresponding user. Then, different UserAgents negotiate with the aim of building a group profile that satisfies their particular minimum requirements, while preserving some privacy. Moreover, we have designed a NegotiatorAgent, which governs the negotiation and may act as a mediator in order to facilitate the agreement. Finally, we have performed some experiments that show that this mechanism is able to give a response in this heterogeneous environment. Key words: Group recommender systems, group profile, negotiation, alternative offers protocol. ∗ Corresponding author. Tel: +34 963 870 000 Email addresses: ingarcia@dsic.upv.es (Inma Garcia), lstarin@dsic.upv.es (Laura Sebastia). 1 Partial support provided by Consolider Ingenio 2010 CSD2007-00022, Spanish Government Project MICINN TIN2011-27652- C03-01. Preprint submitted to Elsevier 31 July 2013 1 Introduction A Recommender System (RS) [33] is a personalization tool that attempts to provide people with lists of items that best fit their individual tastes. A RS infers the user’s preferences by analyzing the available user data, information about other users and information about the environment. While many RSs are focused on making recommendations to a single user, many daily activities such as watching a movie or going to a restaurant involve a group of users [5], in which case recommendations must take into account the tastes and prefer- ences of all the users in the group [1]. This type of system is called a Group Recommender System (GRS). Over the last few years, GRSs have been an active area of research within the field of RS. As a result, some remarkable GRSs have been developed. For example, Polylens [29] recommends movies, as an extension of the MovieLens recommender; MusicFX [26] selects a ra- dio station among 91 stations; Intrigue [1], The Collaborative Advisory Travel System [27] and Travel Decision Forum [12,14] deal with a tourist domain; GRSK [9,10] is a generic GRS that can be used with any application domain. The main issue in group recommendation is to identify the items that are likely to satisfy all the group members adequately [12,31]. Most of the GRSs lately developed are based on the aggregation of the preference models of individual users into a group model that is then used to elicit a recommendation that is satisfactory for the whole group [12,31]. This centralized view has two main implications: • Most GRSs tend to favour an equal weighting of the individual preferences when recommending an item for the group such that no member is partic- ularly dissatisfied with the decisions. However, some authors criticize these aggregation strategies because the ratings are combined always in the same way without considering how the members in the group interact with each other [6,32]. For example, we can find a person that wants to favor a spe- cific user or that feels more comfortable when accepts the others’ proposals instead of trying to impose his/her preferences over the other group mem- bers. That is, new trends in GRSs [24,32] argue that it is more realistic to capture the different attitudes of each member in the group with respect to the others. • Centralized systems need that the user communicates his/her preferences to the system in order to provide a recommendation. However, not all the users feel comfortable in this situation, given that they consider that preferences are delicate information which should not be revealed to everybody. That is, a GRS can be also considered as a distributed system, where each user has private information and the system only knows the information that the user wishes to make public. 2 In summary, the problem we are facing in this work has the following charac- teristics: • It is a group recommendation problem, where a group of users want to obtain a recommendation of a list of items that match their preferences as a whole. • Each group member has his/her own preferences and may exhibit a differ- ent behaviour. For example, different users may have different expecta- tions about the resulting group profile, may want to impose their preferences over the other users’ preferences or may like better to agree with the others’ proposals. • Both the preferences and the user expectations are private information to each user, who decides which information he/she wants to share with the other users at any given moment. Therefore, the task of the GRS is to manage, joint with the individual prefer- ences, some other aspects of the users behaviour to come up with a common group profile, only taking into account the information provided by the users until this moment. This may imply, for example, that the final list of recom- mended items in our case may contain more items suitable for a user than for another and both can be equally satisfied. In order to deal with this problem, we have opted for building a multi- agent system (MAS) [39], where multiple agents work together in order to obtain a recommendation for the whole group. That is, ours is not a centralized system responsible of computing the group preference model; instead, the users themselves (i.e. the agents that represent the users), coordinated by another agent acting as a host, are responsible of obtaining the group profile by means of a negotiation process. Specifically, the tasks performed by these agents are: (1) building the individual preferences model, (2) reaching an agreement about the group profile and (3) selecting the recommended items for all the group members according to the obtained group profile. The group profile is built by means of a negotiation process ([15,17]), whose goal is to reach an agreement about the preferences that are included in the group model, that is, an agreement about the preferences that are shared by all the group members and at which degree. If the negotiation process has been successful, these preferences are then used to select the list of recom- mended items for the whole group, which is then shown to the real users. The negotiation process in our GRS is a variation of the model of alternating offers [16] in a multi-party setting. Each agent that represents a real user in the MAS (named UserAgent) knows the preferences of this user, but may share only some information with the other agents. In this paper, we introduce an example of UserAgent that can be configured by the real user for exhibit a certain behaviour during the negotiation. This different behaviour can be 3 implemented as different utility functions, different agreement thresholds or different criteria to accept or reject the proposals from other users. Apart from the agents representing the real users, there is a host agent that is in charge of controlling the negotiation process, collecting the user proposals and creating a common proposal to all the users. Moreover, this agent may also participate in the negotiation as a mediator to help the agents to reach an agreement. The advantages of our solution for the resolution of the group recommendation problem are: (1) The fact that each user may exhibit a different behaviour is easily cap- tured by the behaviour of each agent, who decides which proposals are accepted or rejected. This results in a more flexible system. (2) Negotiation is a method that captures well the group dynamics when the agents involved have different behaviours. (3) Users do not need to communicate all their preferences to the system, so that privacy is preserved (at some extent). The paper is organized as follows. Section 2 summarizes the state of the art on the use of MAS in recommender systems and the management of the particu- lar behaviour of each user in group recommendation. Section 3 gives an outline of the MAS and the agents that participate in the recommendation process. Section 4 introduces the negotiation framework for obtaining the group profile and all the components (protocol, messages, etc.). Section 5 details the strat- egy of the host agent and Section 6 gives an example of strategy for a user agent. Section 7 shows an example of a negotiation process, summarizes some experiments performed to test our distributed approach and analyzes how dif- ferent settings for the agents may affect to the result of the recommendation. We finish with some conclusions in Section 8. 2 Background The main issue in GRS is to identify the items that are likely to satisfy all of the group members adequately [12,31]. Some GRSs build a group profile that considers the tastes and preferences of the whole group by using aggre- gation methods to elicit the group profile, associating a weight or degree of interest to each preference in the group profile. There are many remarkable GRS based on the elicitation of preferences; examples include Intrigue [1], Polylens [29], MusicFX [26], Let’s Browse [18], The Collaborative Advisory Travel System, CATS [27] and GRSK [9]. A description of these GRSs joint with a classification upon different features can be found in [5,7,9,25]. As explained above, these systems collect the preferences of all the group 4 members and combine them in order to obtain a recommendation that equally satisfies the group as a whole. This implies that every individual is considered as equal to the others and, therefore, how the members in the group inter- act with each other is ignored. Besides, it can be very difficult to obtain a single recommendation that satisfies every member for heterogeneous groups. Moreover, the general group satisfaction is not always the aggregation of the satisfaction of its members as different people may have different expectations [32]. In summary, the decision of a group member whether or not to accept a given recommendation can depend not only on his/her evaluation of the con- tent of the recommendation but also on his/her beliefs about the evaluations of the other group members and about their motivation [13]. Recently, some GRSs that consider the different behaviours and attitudes of the group members have appeared. For example, [32] proposes a GRS that takes into account the personality of each group member to weight the influ- ence of his/her ratings during the recommendation process. A conflict situation is defined as a situation in which concerns of people appear to be incompati- ble. In these situations, the behaviour of an individual can be described along two basic dimensions: (1) assertiveness, when the person attempts to sat- isfy his own concerns, and (2) cooperativeness, when the person attempts to satisfy the concerns of the other person. These basic dimensions of behaviour define five different modes for responding to conflict situations: competing (as- sertive and uncooperative), collaborating (assertive and cooperative), avoiding (unassertive and uncooperative), accommodating (unassertive and coopera- tive) and compromising (moderately assertive and cooperative). Taking into account group member interactions, this work proposes three versions of a simple group recommendation algorithm inspired in other classical recommen- dation algorithms described in the literature, where variations based on the conflict mode behaviour of the group members have been included. Generally speaking, assertive conflict mode behaviours penalize negatively the differences with the best choice of another members, given that the other choices do not satisfy his/her personal concerns. On the other hand, cooperative behaviours reward the difference with the best choice of another members, that is, it is not his/her preference choice but it will be good enough for other members and for the group. This approach was tested in the movie recommendation do- main, and the results showed that recommendations could be improved when using the conflict personality values. Another example of GRS that takes into account the users’ personality is described in [24]. This GRS deals with some of the emotions that play a role in the recommendation of sequences of items to a group of users. An accurate prediction of individual satisfaction becomes crucial because to keep the rest of the group happy, an individual might need to be confronted occasionally with items he/she does not like. For example, an accurate prediction of satisfaction would help to ensure no individual gets too dissatisfied, by presenting disliked 5 items at appropriate times. This paper also considers that the feeling of others in the group may influence an individual’s satisfaction. For example, it takes into account emotional contagion, the influence of an individual’s affective state on that of others in the group, as well as conformity, whereby judgements are influenced by those of others. The Travel Decision Forum [13] is a prototype GRS which supports users in planning a joint vacation. The goal of the GRS is to provide a group profile. At any moment only one group member will be interacting with the web-based system and the absent group members are represented by animated characters. In the first phase of the interaction with the system, each member specifies his/her preferences concerning the vacation and the evaluation criteria of a representative with respect to both the absolute utility of proposals and the relative utility of proposals for different group members. The goal of the group in the second phase is to agree on a joint preference model by means of a negotiation between the current user and the representatives of the absent users. There is also another animated character representing the mediator that moderates the negotiation process for each dimension of the preference model. Then, (1) the mediator computes a proposal based on the specified preferences of all members, (2) the representatives accept or reject the proposal according to a threshold defined by the corresponding real user, (3) the current member responds to the proposal by accepting, adapting or rejecting it or by adapting his/her preferences and the process goes back to step 1 again. The interaction with the system continues until the current member has either agreed with the representatives of the other members on a joint preference model for each value dimension or run out of time or interest. Even in the latter case, the positions of the group members could be closer than at the beginning of the interaction, given that the current member may have adapted his/her preferences or may have relaxed the criteria of his/her representative to accept other proposals. When the next group member interacts with the system, there will be opportunities for further convergence. This system differs in two key aspects from the two previous approaches: any decision about the preference model is taken by the real user and the information about preferences is private to each of them. This distributed structure of the GRS gives major flexibility to the system in order to consider different user behaviours. In this sense, the GRS pre- sented in [4] relies on the application of cooperative negotiation mechanisms (based on individual recommendations and user preference models) to generate group recommendations. Each group member is represented by a negotiation agent, all with the same behaviour. First, the individual recommendation (list of ranked items) is obtained by means of the case-based recommender sys- tem Trip@dvice [34] and then, the negotiation process starts. Specifically, two negotiation protocols are described: alternating offers, for a direct negotia- tion between two users; and merging ranks (with mediation) for negotiation 6 among a greater number of users. In the alternating offers protocol, each ne- gotiation participant has the opportunity to place an offer or to accept one of the previously placed offers by the other negotiation participant. In contrast, in the merging ranks protocol, each negotiation participant supplies its com- plete ranked list to a negotiation mediator that evaluates each proposal and forms a composite ranked list, called agreement. To generate an agreement, the mediator may have a ”stock” of strategies to help him choosing among these proposals the one that will be offered to the group. The ranked list of products resulting from the agreement formation phase is the resulting group recommendations produced by the overall process. Unlike Travel Decision Fo- rum, the agent representing each user is autonomous for accepting or rejecting the proposals to finally reach an agreement. As far as we know, these two last systems are the most similar to our approach. In the remaining, we summarize some other recommender systems that use MAS to retrieve, filter and use information relevant to the recommendation, when a single information source does not contain the complete information needed for the recommendation. The majority of the recommender systems based in a multi-agent architecture are aimed at obtaining recommendations for a single user. Some of them are related to the tourism activity. For ex- ample, [37] describes a software travel agent that recommends a trip plan for a tourist to Taipei city by conducting the negotiation with human travelers. BerlinTainment [38] is a MAS that can be used to plan comprehensive day itineraries for entertainment on Berlin and determine current locations, points of interest, and routes. CASIS [22] is another example of a MAS based on a case-based reasoning approach to recommend the best travel to the user. The work in [20,21] describes a multi-agent recommender approach based on the collaboration of multiple agents exchanging information stored in their local knowledge bases with the global objective of recommending the best travel package to the user. Examples of MAS used in other type of applications in- clude: ANEGSYS [19], which is a RS for product acquisition in local markets based on a MAS that uses automatic bilateral negotiations between buyers and sellers; and [23], which is a multi-agent group recommender system that address the problem of arranging meetings for several participants taking into consideration constraints for personal agendas and transportation schedules. Finally, there is a number of systems that use recommendation techniques as a method of making negotiation smarter and more efficient, such as [17,28,8,40]. 7 3 Description of the Multi-Agent System This section introduces the main characteristics of the MAS 2 , where different types of agents work together to obtain the group recommendation. The key aspect of the recommendation is the elicitation of the group prefer- ence model, which is obtained by means of a mediated negotiation process among a set of agents, each acting on behalf of one member of the group. The behaviour of these agents, named UserAgents, is determined by the cor- responding real user, who establishes the parameters that define the agent’s negotiation strategy. The negotiation model is a multi-party negotiation cen- tralizing the communications through a NegotiatorAgent. It receives the pro- posals of the UserAgents, combines them into a single proposal, which is later broadcasted by the NegotiatorAgent and analyzed by the UserAgents. More- over, the NegotiatorAgent also acts as a mediator, for facilitating the consensus about the group preferences. Along with the UserAgent and the Negotiator- Agent, there are two support agents involved in the recommendation process. The PreferencesAgent calculates the preferences of a user given his/her profile and the ItemsSelectorAgent selects the list of items to recommend given the negotiated group profile. The following sections introduce the main characteristics of our GRS. First, we describe the information that it needs for providing a recommendation. Then, Sections 3.2, 3.3 and 3.4 give an outline of the different types of agents that participate in the recommendation. Finally, the interactions between these agents in the recommendation process are summarized in Section 3.5. 3.1 The GRS Data Our system relies on the use of a domain ontology [11] to describe the user’s likes and the items to recommend in the particular domain. This GRS is domain-independent, that is, it has been designed to be able to work with any application domain provided that the domain information is specified through a domain ontology. Classes in the ontology represent the features (F) that are commonly managed in the corresponding domain. The leaf nodes of the ontology (instances of classes) represent the items to recommend. The edges that link an item to a feature are associated to a value that indicates the degree of interest of the item under the corresponding feature, i.e., as a member of the category denoted by the feature. The degree of interest of the item i under the feature f (dif) is the degree of suitability of the item to the 2 These agents are implemented in Java, over the JADE platform (http://jade.tilab.com/). 8 feature. Items are described by means of a set of tuples which represent all the incoming edges of a leaf node. A tuple that describes an item i is of the form (f, dif), where f ∈ F is a feature defined in the ontology such that there is an edge connecting f and the item i, and dif ∈ [0, 100] is the degree of interest of the item i within the category represented by the feature f. A more detailed explanation about the domain ontology can be found in [9]. This ontology is also used to describe the individual and group preference models. A preference is a tuple of the form (f, df), where f is a feature in the ontology and df ∈ [0, 100] is the estimated degree of interest of the user u or the group G in the feature f [9,10]. Then: • The preference model of an individual user u (P u) is the set of preferences that are preferred by the user u. In this case, the value df of a preference represents the estimated degree of interest of the user u in the feature f (denoted by duf). • The preference model of a group G (P G) is the set of preferences that are preferred by the group G. The value df of a preference is the estimated degree of interest of the group G in the feature f (denoted by dGf). 3.2 The NegotiatorAgent The NegotiatorAgent plays a double role in the negotiation process. First, it acts as a coordinator or host during the whole process. This implies to be in charge of the following tasks: (1) Managing the group of UserAgents: The NegotiatorAgent receives the UserAgents’ requests to participate in the negotiation to build the P G. These UserAgents are then included in the set GUA, which represents the UserAgents participating in the negotiation process. (2) Lauching the negotiation process among the UserAgents in GUA by sending a message to these UserAgents. (3) Centralizing the communications among the UserAgents, which implies receiving the UserAgents’ proposals, combining them into a single proposal, which is later broadcasted to all the UserAgents. Moreover, the NegotiatorAgent checks the proposals to decide if an agreement has been reached. (4) Communicating the result of the negotiation process to the UserAgents in GUA, that is, whether the negotiation process has been successful or not. In case of agreement, the NegotiatorAgent requests the ItemsSelectorAgent to select the items to recommend. The list of recom- mended items RIG is finally sent to the UserAgents in the group. 9 Besides, the NegotiatorAgent acts as a mediator when an agreement has not been reached after the UserAgents have proposed all their preferences. How this mediation to ease the consensus is performed will be further detailed in Section 5. 3.3 The UserAgent Each user u in the group G is represented by a UserAgent in the MAS. The UserAgent records the data of a user u in a user profile [30] which is composed by the recommendation profile and the negotiation profile: • The recommendation profile (see [9] for further details) contains the information related with the preferences about the items to recommend, which is used to compute the individual preference model P u. • The negotiation profile allows the user to configure some aspects that determine the UserAgent strategy during the negotiation process. Specifi- cally, the user can define the basic behaviour of the UserAgent (for example, self-interested or collaborative), the minimum requirements to agree with a proposal, etc. This will be detailed in Section 6. All this information is provided by the user when he/she starts using the system. Once the user has entered his data, the negotiation process is inde- pendent from the real user. It is the UserAgent who acts on behalf of this user and participates in the negotiation with the other UserAgents. Finally, the real user is informed about the result of the negotiation: the failure in the negotiation or the list of recommended items for the group. Specifically, the UserAgent is in charge of the following tasks (Figure 1): (1) Building and updating the user profile. When a user enters the system, the first step is to enter the data that compose both the recom- mendation and the negotiation profile. This user profile can be updated at user’s request. (2) Obtaining the individual preference model P u. The UserAgent sends a request to the PreferencesAgent to compute the individual pref- erence model, that is, the list of preferences that characterize the user within the application domain. (3) Participating in the negotiation process. The UserAgent negotiates with the other UserAgents that act on behalf of the other members of the group to elicit the group preference model that better represents the whole group. This negotiation is carried out by taking into account the behaviour defined by the user in his profile. (4) Informing the user about the result of the negotiation. If the negotiation ends successfully, the UserAgent shows the user the list of 10 recommended items. Otherwise, the UserAgent informs the user about the failure of the negotiation. 3.4 The Support Agents Besides the main agents in this architecture, the UserAgent and the Negotia- torAgent, there are two agents that provide support to some of their tasks: the PreferencesAgent and the ItemsSelectorAgent. The PreferencesAgent is the agent in charge of computing the individual pref- erence model P u. When the PreferencesAgent receives a request from a User- Agent, it analyzes the recommendation profile of the corresponding user and elicits the list of preferences that characterize this user within the application domain. How the PreferencesAgent builds this P u is out of the scope of this paper, but any method for eliciting preferences given a user profile could be used [9]. The ItemsSelectorAgent is the agent in charge of selecting the list of items RIG that satisfy the group’s preferences, given the group preference model P G. This list is a set of tuples of the form RIG = {(i, dGi)}, where i is the recommended item, and dGi is the estimated degree of interest of the group G in the item i. Specifically, once an agreement is reached by the UserAgents in the negotiation process, the NegotiatorAgent sends a request to the ItemsSelectorAgent to compute the list of items that better match the P G. Further details about the calculation of RIG can also be found in [9]. 3.5 Interaction of the agents in the MAS The starting point of the recommendation process is a group of users that want a recommendation for the group. This process has been implemented by means of a MAS, where different types of agents work together in order to obtain the recommendation more suitable for the group members. Figure 1 shows an sketch of the interaction between these agents. The purpose of these interactions is the following: (1) Step 1: Obtain the individual preference model. This first step is performed by the PreferencesAgent at each UserAgent’s request (step 1 in Figure 1). Given the user profile, the PreferencesAgent builds the set of user preferences that will be used to negotiate (P u). (2) Step 2: Obtain the group preference model. The negotiation process to compute the group profile P G is controlled by the NegotiatorAgent. All the UserAgents that want to participate in the negotiation process, send a 11 User 1 User 2 Negotiator Agent ItemsSelector Agent Step 3.2. RIG Step 3.1. PG Step 2.2. Negotiation messages Step 2.2. Negotiation messages Negotiation profile Recommendation profile User Agent User Agent Step 2.1. Negotiation request Step 2.1. Negotiation request Negotiation profile Recommendation profile Pu 1 Pu2 Preferences Agent Step 1.1. Recommendation Profile Step 1.2. Pu1 Step 1.1. Recommendation Profile Step 1.2. Pu2 Support AgentsStep 3.3. RIG Step 3.3. RIG Fig. 1. Steps in the recommendation process. negotiation request to the NegotiatorAgent (step 2.1 in Figure 1). Then, a flow of messages between the UserAgents and the NegotiatorAgent starts (step 2.2 in Figure 1), that finishes when there is a consensus about the group preference model or when it is not possible to reach an agreement. (3) Step 3: Obtain the list of recommended items. If the UserAgents have reached an agreement during the negotiation process, the group preference model P G is then used to obtain the list of recommended items to the group. The NegotiatorAgent invokes the ItemsSelectorAgent (step 3.1 in Figure 1). This agent selects the items that better match the list of group preferences and obtains the final list of recommended items RIG (step 3.2 in Figure 1). The NegotiatorAgent sends this RIG to the UserAgents involved in the recommendation process (step 3.3 in Figure 1), which in turn send this list to the corresponding users. 12 4 Negotiation Framework Agents in MAS share a negotiation mechanism that specifies what possible actions each party can take at any given time, when negotiation terminates, and what the structure of the resulting agreements is [3]. The components of the negotiation framework are the following [15]: objects, protocols and strategies. The remaining of this section is devoted to define these compo- nents in our MAS. 4.1 The negotiation objects The negotiation objects are the range of issues over which agreement must be reached, i.e. the preferences in the individual preference model P u of each UserAgent in the group. To participate in the negotiation process, it is manda- tory that all the UserAgents have computed the corresponding P u. Then, the UserAgents negotiate with the preferences in P u in order to compose an agreed group preference model P G. If an agreement is reached in the construction of P G, it will be used to obtain the list of items recommended to the group. If there is no agreement, the negotiation failed and a recommendation cannot be provided. 4.2 The negotiation protocol The negotiation protocol are the set of rules that govern the interaction, types of participants, the negotiation states, the events that cause negotiation states to change and the valid actions of the participants in particular states. The protocol determines the flow of messages between the negotiating parties, dictating who can say what and when and acts as the rules by which the negotiating parties must abide if they are to interact [2]. The protocol must be public and known by all the participants in negotiation. The definition of negotiation parties in this work follows the model defined by [3], where the roles involved in the negotiation process are negotiation partici- pant and negotiation host. In our MAS, the role of participant is played by the UserAgent and the role of negotiation host is played by the NegotiatorAgent. In each negotiation process, there is only one host, whereas there could be an unlimited number of participants. The first action to be taken is for a participant, i.e. a UserAgent, to require admission to the negotiation. The admission consists of a simple conversation 13 between the participant and the host, i.e. the NegotiatorAgent, where the participant requests the admission to a particular negotiation process. Then, the negotiation process itself starts. The negotiation protocol is discrete and synchronous. That is, it is assumed that the agents can take actions in the negotiation only at certain times in the set T = {t1, t2, t3, . . .} that are determined in advance and known by the agents. These time points can be understood as the steps of the negotiation. The interval between these times is δ (i.e. t2 = t1 + δ). We use a generalization of the bilateral alternating offers protocol [35] for the multi-party negotiation given that, in our MAS, multiple UserAgents may participate in the negotiation. In this case, the host agent (the NegotiatorA- gent) is in charge of supervising the negotiation process, that is, it centralizes the communication among the UserAgents. 4.2.1 The negotiation stages As explained above, the negotiation process starts when all the UserAgents in the group have requested the NegotiatorAgent to participate in the nego- tiation. Then, the host sends a message indicating that the negotiation has started. This negotiation process is mainly organized in two sequential stages, as Figure 2 shows: (1) Negotiation stage 1 (Proposal Stage): Once the negotiation has started, the NegotiatorAgent waits for the UserAgents’ proposals until the next negotiation step (that is, it waits during a time interval δ). Then, it analyzes all the responses, builds a combined proposal that is broadcasted to all the UserAgents and, again, it waits for the UserA- gents’ responses until the next negotiation step. During this interval, the UserAgents evaluate these proposals and accept or reject some proposed preferences. This process is repeated until an agreement is reached or the UserAgents do not have new proposals. (2) Negotiation stage 2 (Mediation Stage): In case an agreement has not been reached during the proposal stage, the NegotiatorAgent moves to the mediation stage. During this stage, the NegotiatorAgent acts as a mediator: it builds proposals that take into account some preferences rejected by the UserAgents during the proposal stage, that have a greater probability to be accepted, in order to facilitate an agreement. This pro- posal is sent to the UserAgents, which evaluate it and, again, accept or reject some proposed preferences. The UserAgents’ responses are ana- lyzed by the NegotiatorAgent, which builds a new proposal. This stage ends when an agreement is reached or when no new proposal can be built. In the latter case, the NegotiatorAgent moves to stage End, that 14 Stage 1: Proposal Stage Stage 2: Mediation Stage Negotiation End with Agreement Negotiation End in Failure End No agreement No agreement Agreement Agreement No agreement Fig. 2. Negotiation states. indicates the end of negotiation without agreement 3 . The NegotiatorAgent is in charge of moving from one stage of negotiation to another when certain events occur. Every time it sends a message, the NegotiatorAgent informs the UserAgents about the negotiation stage in which they are. This way, the UserAgent is able to act according to the negotiation stage, if its strategy has been defined to do it. 4.2.2 The negotiation set of rules: the format of messages All messages follow the specifications of the Agent Communication Language FIPA-ACL 4 . The structure of a message is (performative : content), where performative denotes the type of the communicative act of the ACL message, whereas content (obviously) denotes the content of the message. It could be a fixed sentence (such as ”negotiation end”), but the majority of the messages exchanged between the agents contain proposals. Given that the issues under negotiation are the preferences that should be in the group preference model P G, these proposals comprise two lists of preferences 5 : (1) a list of accepted preferences and (2) a list of proposed preferences. Table 1 shows the types of broadcast messages that the NegotiatorAgent sends to the UserAgents in GUA and Table 2 shows the messages each UserAgent sends to the NegotiatorAgent. 4.3 The decision making model (negotiation strategy) The decision making model or negotiation strategy are the decision making apparatus the participants employ to act in line with the negotiation protocol in order to achieve their objectives. The sophistication of the model, as well as 3 We have introduced this additional stage for the sake of clarity of the algorithms that will be detailed below. 4 http://www.fipa.org/ 5 See Section 3.1. For the sake of simplicity, we call preference the feature associated to the preference. 15 Type Perfor. Content Negotiation startup cfp ”send pro- posal” The negotiation process has started and the NA is waiting for proposals Proposal propose stage, acceptedt, proposedt Proposal that combines the UA’ proposals previously received by the NA stage: stage of the negotiation pro- cess (st1 or st2) acceptedt: list of preferences ac- cepted by all the UAs in previous iterations of the whole negotiation process proposedt: the list of preferences proposed to the UAs Negotiation end with agreement agree P G, RIG P G: group preferences model RIG: list of recommended items Negotiation end in failure failure ”negotiation end” Table 1 The NegotiatorAgent messages. the range of decisions that have to be made, are influenced by the protocol in place, by the nature of the negotiation object, and by the range of operations that can be performed on it [15]. Within a negotiation process, each party adopts a strategy which determines exactly which actions they make (in response to actions by other parties or external events) in an effort to maximize their individual or collective gain [3]. Sections 5 and 6 explain the strategies followed by the NegotiatorAgent and the UserAgents, respectively. 5 The NegotiatorAgent strategy As explained in previous sections, the NegotiatorAgent is the agent that acts as the host in the negotiation process. This implies that it is in charge of managing the group of UserAgents, launching and controlling the negotiation process and informing the UserAgents about the result of this negotiation. Moreover, the NegotiatorAgent centralizes the communication among the UserAgents. On the other hand, this agent is also concerned with the fact that the UserAgents reach an agreement. At the beginning of the negotiation (proposal stage), it remains in the background, leaving the UserAgents negotiate on their own. However, if this is not possible, it acts as a mediator in the second stage 16 Type Perfor. Content Participate in negotia- tion request ”negotiation” The UA wants to participate in ne- gotiation. Proposal propose user − acceptedut , user −proposedut The UA does not agree with the preferences currently accepted by all the UAs and it wants to continue with the negotiation process. user−acceptedut : the preferences in the NA’s proposal that the UA ac- cepts. user − proposedut : the preferences that the UA proposes in this iter- ation. Proposal and agree- ment agree user − acceptedut , user −proposedut The UA agrees with the list of ac- cepted preferences sent by the NA in the last message. Table 2 The UserAgent messages. (mediation stage) in order to facilitate such an agreement. As the negotiation host, the NegotiatorAgent controls the whole negotiation process. First, it receives the requests of the UserAgents for participating in the negotiation. All the UserAgents are accepted and included in the set GUA. In case that a UserAgent leaves the negotiation on its own or it is expelled (be- cause, for example, it did not answer on time), it is removed from GUA and the NegotiatorAgent does not interact with it anymore. Once all the UserAgents have been registered, the NegotiatorAgent sends a broadcast message to start the negotiation. Then, it waits for the UserAgents’ messages. After receiving these messages, it follows these steps: (1) analyzing the UserAgents’ responses, (2) preparing and sending a new message and (3) deciding in which stage the negotiation is. After the third step, if the negotiation process continues, the NegotiatorAgent waits again for the UserAgents’ messages. 5.1 Step 1. Analysis of the UserAgents’ responses Let’s consider the following time points: • t0 is the time point when the NegotiatorAgent sent its last message • t1 is the current time point • t2 is the time point when the next message from the NegotiatorAgent is due 17 Algorithm 1 The NegotiatorAgent’s algorithm. Step 1. Analyze the UserAgents’ responses Step 1.1. Analyze the UserAgents’ message performatives if ∀u ∈ GUA : performativeu==agree {Negotiation end with agreement} P G ← acceptedt0 Obtain the RIG from the ItemsSelectorAgent Send message: (agree : P G, RIG) Return end if if stage ==end {Negotiation end in failure} Send message: (failure : "negotiation end") Return end if Step 1.2. Analyze the UserAgents’ answers to the previous proposal Compute acceptedt2 as equation 1 shows Update RP with the non-accepted preferences if stage ==st2 Update RP (remove accepted or the first Nremove preferences) end if Step 2. Build the NegotiatorAgent’s new proposal if stage ==st1 Compute proposedt2 as equation 2 shows end if if stage ==st2 proposedt2 ← Nmediation top preferences in RP end if Send message: (propose : stage, acceptedt2, proposedt2) Step 3. Decide about moving from one stage to another if stage ==st1 ∧ proposedt2 == ∅ stage ← st2 end if if stage ==st2 ∧ RP == ∅ stage ← end end if After the NegotiatorAgent receives the messages from the UserAgents, which have the format (performativeu : user − acceptedut1, user − proposed u t1 ), it proceeds as follows: • Step 1.1. Analysis of the UserAgents’ performatives. There are two 18 possibilities: · All the UserAgents agree with the previous proposal: This implies that the negotiation ends with an agreement. Then, the group preference model P G is composed by the list of accepted preferences in this last proposal, joint with the degree of interest for the whole group for each preference (feature) p, which is computed as 6 : dGp = avg ∀u∈GUA ( dup ) Afterwards, the NegotiatorAgent makes a request to the ItemsSelector- Agent, which uses this P G to obtain the list of recommended items RIG, which is sent to the UserAgents, and the negotiation process finishes. · Not all the UserAgents agree with the previous proposal: In this case, there is no agreement at this moment and the negotiation process continues with the successive steps, except if the negotiation stage is end, in whose case, the negotiation process fails because there are not new proposal from the NegotiatorAgent when acting as a mediator. • Step 1.2. Analysis of UserAgents’ answers. At this moment, the Ne- gotiatorAgent updates the list of preferences accepted by all the UserAgents as follows: acceptedt2 = acceptedt0 ∪   ∩ ∀u∈GUA user −acceptedut1   (1) The remaining tasks in this step are related with the NegotiatorAgent acting as a mediator and will be explained later in this section. 5.2 Step 2. Building the NegotiatorAgent’s new proposal How the NegotiatorAgent builds a new proposal depends on the stage the negotiation process is, which, in turn, determines whether this agent is acting as a host or as a mediator. 5.2.1 The NegotiatorAgent as a host (proposal stage) The NegotiatorAgent composes a new proposal (proposedt2) as the aggregation of the latest proposed preferences by all the UserAgents: 6 Recall that this list of preferences only contains the features the UserAgents are interested in, but the NegotiatorAgent also knows the degree of interest associated to these features for each UserAgent. 19 Management of the list of rejected preferences (RP) Nmediation Number of preferences proposed in a negotiation step during stage 2. Nremove Number of not accepted preferences removed from the RP in a negoti- ation step during the stage 2. It must not be greater than Nmediation so that a preference is not removed before being proposed at this stage. Table 3 The negotiation profile of the NegotiatorAgent. proposedt2 = ∪ ∀u∈GUA user −proposedut1 (2) Even in the case that the new proposal is empty, which indicates that no new proposal was received from the UserAgents, the corresponding message is sent, because there is a new list of accepted preferences (acceptedt2) that could make the UserAgents to reach an agreement in the next negotiation step. 5.2.2 The NegotiatorAgent as a mediator (mediation stage) At stage 2, the NegotiatorAgent acts as a mediator with the aim of helping the UserAgents to reach an agreement. In short, it is in charge of building new proposals. Specifically, it progressively proposes the preferences rejected at stage 1, that is, it proposes first the preferences that it considers that may be more interesting for the group of users, taking into account the experience of the first stage. For doing so, the NegotiatorAgent maintains a ranked list of rejected preferences (denoted by RP), where each rejected preference has the following form: (p, dGp , users) where p denotes the feature associated to the preference, users is the list of UserAgents that have proposed the preference p at any time and dGp denotes the average of the degree of interest of the preference p for the UserAgents in the list users. During stage 1, when a UserAgent u proposes a preference p and it is not accepted by the group, it is registered in RP (see Algorithm 1, Step 1.2): if p is not in RP already, a tuple (p, dup,{u}) is included in RP ; otherwise, the associated values are updated accordingly. As stage 1 continues, a preference p that was previously rejected by the group (and, therefore, included in RP), can be accepted afterwards, in whose case, the corresponding tuple is removed from the set RP . In our current implementation, the list RP is ranked according to the number of UserAgents that have proposed each preference, given that it is easier that the group accepts a preference proposed by many UserAgents. The 20 preferences with the same number of UserAgents in the list users are ranked according to the degree of interest dGp . When the list is ranked, it is balanced according to the UserAgents that proposed each preference, so that no user is benefitted in detriment of the others. The strategy of the NegotiatorAgent when acting as a mediator (which can be configured in order to adapt its behaviour -see Table 3-), consists in pro- gressively propose the preferences in the list RP until this list is empty or an agreement is reached. Specifically, each proposal during the mediation stage contains the first Nmediation preferences. The preferences from the Negotiator- Agent’s proposal that are accepted by all the UserAgents are removed from RP . However, if no preference is accepted, then they are not removed and the same Nmediation preferences will be proposed in the next message. In order to avoid deadlocks, Nremove preferences are removed from RP in case that no preference in the last proposal has been accepted. Therefore, in any case, the number of preferences in the list RP decreases at each iteration of the negotiation process. The values of Nmediation and Nremove ensure that all the preferences in RP are eventually proposed to the UserAgents (unless an agreement is reached before) and that a deadlock cannot be encountered. In particular, if Nmediation is assigned a high value (w.r.t. the number of preferences in RP), the negotiation is intended to go faster in the sense that more preferences are included in each proposal. However, this does not necessarily imply that an agreement is reached faster because the preferences in a proposal are accepted by unanimity, which is more difficult to achieve if the UserAgents have more preferences to choose. Moreover, if Nremove is low (w.r.t. the number of preferences in RP), the negotiation process may take longer but it is more likely that the negotiation is successful, given that the preferences are slowly removed from RP and, consequently, they can be proposed several times, which gives the UserAgents more opportunities of accepting them. In other words, if Nremove is high, it may cause that preferences that would be accepted in subsequent iterations, are removed and, therefore, cannot be accepted. Given that the strategy of the UserAgents is not known by the NegotiatorAgent in advance, it is difficult to decide (during the configuration process) the most appropriate values of these parameters in order to ease the agreement. For this reason, as future work, we are interested on studying the inclusion of learning techniques to be able to adapt the strategy of the NegotiatorAgent (the criteria to rank the list RP and the values of Nmediation and Nremove), depending on how the UserAgents act during the negotiation process. 21 5.3 Step 3. Deciding about moving from one stage to another Finally, the NegotiatorAgent considers whether the negotiation must move from the current stage to a different one. The events that define the end of each stage are the following: • The stage 1 ends when the messages from all the UserAgents do not contain any new preferences. In this case, the new proposal built by the Negotia- torAgent is empty. • The stage 2 ends when the NegotiatorAgent (acting as a mediator) does not have any new proposal to make. If the negotiation is at stage End, then the process ends in failure. It is im- portant to remark that the NegotiatorAgent is the only agent that can decide when the negotiation finishes. In other words, the UserAgents can decide to leave the negotiation but they cannot decide when it finishes. 6 The UserAgent strategy As explained above, this section is devoted to describe the implementation of the UserAgent. Nevertheless, any other implementation could be used, as long as it follows the same negotiation protocol. All in all, our UserAgent can be configured in order to exhibit a behaviour as closer as possible to the real user’s behaviour. The goal of each participant, i.e. each UserAgent, in the negotiation process is to obtain an agreed P G that at least contains a set of preferences enough to satisfy its minimum requirements. Regarding the classification of behaviours introduced in [32], our UserAgent exhibits an assertive behaviour, in the sense that it attempts to satisfy the corresponding user’s concerns. However, this behaviour may cause that a consensus cannot be reached if the users have heterogeneous tastes. For this reason, the user may prefer the UserAgent to be more collaborative, so that, it may make some concessions in order to reach an agreement. All these characteristics about the behaviour of the UserAgent are defined in the negotiation profile. This profile contains parameters that are used: (1) To check when a proposal is considered to be satisfactory, namely, the proposal scoring function F u and the agreement threshold T uagree. (2) To model the progressive negotiation mechanism, namely, the levels of negotiation Lu. (3) To model the UserAgent behaviour during the mediation stage, namely, the degree of cooperativeness and the concession tactics Nuconcession. 22 Proposal Evaluation Proposal scoring function F u F u(preferences) defines an utility function to evaluate the proposal Agreement threshold T uagree if F u(acceptedt) ≥ T uagree then u agrees with acceptedt Progressive Negotiation Mechanism Levels of negotiation Lu lu = {0, 100}∪{li ∈ ]0, 100[ } L1 = [ln, 100], L2 = [ln−1, ln[, . . . , LN = [0, l1[ where ln, ln−1, . . . , l1 ∈ lu and L1, L2, . . . , LN ∈ Lu Behaviour Degree of coopera- tiveness Self-interested: u will only accept the preferences that are included in its own P u. Collaborative: u considers the preferences proposed by other UAs in order to ease an agreement. Concession tactics Nuconcession Number of preferences proposed by other UAs (and not included in its own CP u) that it accepts at each iteration of the mediation stage. Table 4 The negotiation profile of the UserAgent u. It is important to remark that both the behaviour and the strategy of each UserAgent are private; in other words, the UserAgent strategy and behaviour are unknown for the other UserAgents in GUA and for the NegotiatorAgent. The UserAgent negotiation strategy is a progressive negotiation mecha- nism both for evaluating the NegotiatorAgent proposal (see section 6.1) and for composing the new proposal (see section 6.2); that is, only those prefer- ences with certain degree of interest in P u are accepted/proposed at a given moment. The user defines a parameter that governs this process: the levels of negotiation Lu (see table 4). The UserAgent accepts/offers the most-valued preferences and, if there is no agreement, it incrementally accepts/offers other preferences with a lower degree of interest. In order to apply this progressive mechanism, the user indicates the UserAgent in which order his/her prefer- ences should be accepted/proposed by means of the definition of the levels of negotiation. The levels of negotiation are a set of intervals of degrees of interest in which the preferences in P u are organized. This results into the classified user preferences CP u. The number of levels in the CP u may imply a dif- ferent performance of the UserAgent. For example, if a user defines only one level, this implies that the UserAgent will accept all the preferences proposed 23 Algorithm 2 The UserAgent strategy. Step 1. Analysis of the NegotiatorAgent message Step 1.1. Analyze the list of preferences accepted by all the UserAgents if F u(acceptedt1) ≥ T uagree performative ← agree else performative ← propose end if Step 1.2. Analyze the NegotiatorAgent’s new proposal if stage == st1 Compute user −acceptedt2 as equation 3 shows end if if stage == st2 user −acceptedt2 ← Nuconcession top preferences in proposedt1 end if Step 2. Build the UserAgent’s new proposal user −proposedt2 if stage == st1 Compute user −proposedt2 as equation 4 shows Update Lcurrent to the next non-empty level in CP u end if if stage == st2 user −proposedt2 ←∅ end if Send message: (performative : user −acceptedt2, user −proposedt2) by other UserAgents that are in its own P u, which would lead to a quicker agreement. However, if many levels are defined, this UserAgent will need more iterations to reach an agreement. On the other hand, this mechanism allows the UserAgent partially preserve the privacy of its preferences, given that it is not necessary communicating all the preferences in P u at the beginning of the negotiation. On the contrary, it is possible that an agreement is reached before proposing all the preferences of a UserAgent. Algorithm 2 shows the UserAgent strategy, detailing how it processes the Ne- gotiatorAgent’s message and how it builds the new proposal. Recall that the NegotiatorAgent’s message is of the form (performative : stage, acceptedt1, proposedt1) and that the new message of the UserAgent will have the format (propose/agree : user−acceptedt2, user−proposedt2), where t2 = t1 +δ. Initially, when the UserAgent requests to participate in the negotiation and it is accepted, it sets Lcurrent to L1 or to the first non-empty level of negoti- ation. Then, at each iteration, it waits for the next message from the Nego- tiatorAgent. The strategy of the UserAgent when this message arrives can be decomposed into two steps: (1) the analysis of the NegotiatorAgent’s message 24 and (2) the construction of the new proposal of the UserAgent. 6.1 Step 1. Analysis of the NegotiatorAgent message The analysis of the NegotiatorAgent’s message gives as a result the performa- tive of the new message (agree or proposal) and the list user − acceptedut2. This process is performed in two steps: • Step 1.1. Analyze the list of preferences accepted by all the User- Agents. By means the scoring function 7 F u(acceptedt1), the UserAgent de- cides whether the preferences accepted by all the UserAgents in GUA in pre- vious iterations of the negotiation process (contained in the list acceptedt1), are good enough to satisfy its requirements (T uagree). • Step 1.2. Analyze the NegotiatorAgent’s new proposal. In this step, the UserAgent evaluates the new proposal of the NegotiatorAgent (proposedt1) and it computes the list user − acceptedut2. This process de- pends on the negotiation stage. · Proposal stage (stage == st1): The UserAgent only accepts the pref- erences included in its own P u. This means that a preference p included in proposedt1 is accepted by the UserAgent if it belongs to CP u in a level higher or equal to its current level of negotiation. This way, only those preferences already or about to be proposed by the UserAgent are ac- cepted and, therefore, the idea of negotiating first with the most-valued preferences is maintained when analyzing a received proposal and deciding which preferences should be accepted. More formally: user −acceptedut2 = {p/p ∈ proposedt1 : level(p) ≥ Lcurrent} (3) where level(p) denotes the level of preference p. · Mediation stage (stage == st2): During stage 2, the UserAgent will act according to the behaviour defined by the corresponding user. That is, if the user has configured the UserAgent to be self-interested, only those preferences proposed by the NegotiatorAgent that belong to the P u are accepted (this is equivalent to Nuconcession = 0). Otherwise, the UserAgent will accept some preferences proposed by other UserAgents during stage 1. Specifically, it will accept (at each iteration) Nuconcession preferences in proposedt1. Therefore, a higher value of N u concession determines a higher level of cooperativeness. Given that the UserAgent knows that the prefer- ences in the mediator’s proposal are ordered according to their degree of 7 Note that the list used to check whether the requirements of the UserAgent are satisfied is the list acceptedt1, instead of the new proposal of the NegotiatorAgent, which has not already been consensued by the UserAgents. 25 lu = {0, 50, 75, 100} Lu CP u L1 = [75, 100] (Sport,75),(Beach,80),(Open Spaces,77) L2 = [50, 75[ (Architecture,50) L3 = [0, 50[ (Museums,20) Table 5 Example of levels of negotiation in the Tourism domain (Movielens dataset). interest for the whole group, this UserAgent is implemented 8 for accept- ing the first Nuconcession preferences in proposedt1. 6.2 Step 2. Building the UserAgent new proposal The second part of this algorithm consists in building the UserAgent’s new proposal. The preferences in P u are organized in levels, according to the levels of negotiation defined by the user in the negotiation profile, which result in the classified user preferences CP u (see table 5). Then, the first proposal of the UserAgent only contains preferences from the highest level of negotiation L1. As the negotiation progresses, the proposals will include preferences from lower levels. In general, the new proposal will contain the preferences in the non-empty highest level that has not been yet proposed, except those preferences already accepted by all the UserAgents. Considering that t1 represents the current time point (when the NegotiatorAgent’s message has been received) and t2 is the time point when the next message from the UserAgent is due, the new proposal is built as follows: user −proposedut2 = {p/(p, d up) ∈ CP u(Lcurrent)}−acceptedt1 (4) where CP u(Lcurrent) denotes the preferences in the current level of negotiation and acceptedt1 is the list of accepted preferences in the last message from the NegotiatorAgent. Initially, Lcurrent = L1 (or to the first non-empty level of negotiation) and, after sending the proposal, it is updated to the next non-empty level of preferences in CP u. That is, user − proposedut2 are the preferences in the current level of negotiation in the CP u that have not be previously accepted by all the UserAgents. Note that, although the UserAgent agrees with the accepted preferences, it builds a new proposal. The reason behind is that the other UserAgents could 8 However, if more complex behaviours are defined, for example, to favor a certain user during the negotiation, this may not be valid. 26 Proposal Stage Mediation Stage Negotiator- Agent UserAgent Negotiator- Agent UserAgent New Proposal Joins the UAs’ new proposals Uses the CP u Uses the RP − Accepted Pref- erences Pref. accepted by all the UAs Pref. in CP u at Lcurrent Pref. accepted by all the UAs Self-interested: pref. in P u Collaborative: pref. in P u and Nu concession /∈ P u Agree All UAs agree F u(acceptedt) ≥ T uagree All UAs agree F u(acceptedt) ≥ T uagree Negotiation End Agree- ment All UAs agree − All UAs agree − Negotiation End Failure − − RP = ∅ − Table 6 Summary of the behavior of the NegotiatorAgent and the UserAgent (UA). reject the current proposal and, in this case, a new proposal is necessary to continue the negotiation process. If there are no preferences to propose or the negotiation is at stage 2, then the list user −proposedut2 is empty. Finally, the message that will be sent to the NegotiatorAgent is: (agree/propose : user −acceptedut2, user −proposed u t2 ). This progressive mechanism allows the UserAgent partially preserve the pri- vacy of its preferences, given that it is not necessary communicating all the preferences in P u at the beginning of the negotiation. On the contrary, it is possible that an agreement is reached before proposing all the preferences of a UserAgent. An additional advantage of this mechanism is that the user can decide the initial preferences to be used in the negotiation process; then, if no agreement is reached, other preferences can be considered. Currently, this is an static mechanism, but the set of preferences could be dynamically defined. 7 Results This section first shows a simple example, where this negotiation protocol is applied (see Table 6 for a summary). Then, we detail some experiments we have performed with the aim of testing our negotiation process with heteroge- nous groups. Finally, we analyze some aspects of this protocol by considering different values of the UserAgent and NegotiatorAgent parameters. 27 UserAgent u1 UserAgent u2 UserAgent u3 Degree of cooperativeness Collaborative Collaborative Collaborative F u F u(acc) = ( 100− ∑ f∈acc |dGf −duf | |acc| ) ∗min ( 1, |acc| |P u| + α ) T uagree 0.75 0.7 0.5 Nuconcession 1 1 1 Levels of CP u (1,75) (2,80) (3,77) (4,50) (5,60) (6,20) CP u (6,90) (3,90) (7,80) (1,50) (2,50) (8,30) CP u (9,70) (5,90) (3,50) (2,50) (6,60) negotiation Table 7 Negotiation profile of three UserAgents. 7.1 Case of Study This case shows an example of the full negotiation process with three UserA- gents, which have been implemented as explained in Section 6. The negotiation profile of these UserAgents is summarized in Table 7. The utility function we have chosen is aimed at obtaining a P G that contains a significant subset of preferences in the corresponding user P u with a degree of interest (dGf) closer to the user degree of interest. α is a correction factor to give more or less weight to the number of preferences in the P u that must be included in the P G. The parameters of the NegotiatorAgent are Nmediation = 3 and Nremove = 1. This example allows us to show the flow of the agents messages, the change from negotiation stage 1 to negotiation stage 2 and the strategy of the agents. Each step in this process matches with a time interval. Figure 3 shows the flow of messages that both the NegotiatorAgent and the UserAgents exchange during the negotiation process. At each step, the Ne- gotiatorAgent sends a message to all the UserAgents or viceversa. Note that actually the message from the UserAgent contains both the feature and the degree of interest of each preference, but in Figure 3 we only show the feature index for the sake of clarity. We are not going to explain this example in detail; we will remark only the aspects that are not completely clear in the figure and that it is important to emphasize. When the NegotiatorAgent receives the UserAgents’ messages, first it decides whether there is agreement or not. In case of no agreement (for example, 28 (request: negotiation) (cfp: send proposal) 1 UserAgent 1 Negotiator Agent UserAgent 2 (request: negotiation) 2 UserAgent 3 (request: negotiation) (propose: Φ, {1,2,3}) (propose: st1, Φ, {1,2,3,6,7,9,5}) 3 (propose: Φ, {6,3,7}) 4 (propose: Φ, {9,5}) (propose: st1, Φ, {1,2,3,6,7,9,5,4}) 5 (propose: {6,3,7}, {1,2}) 6 (propose: {9,5}, {3,2,6}) (propose: {1,2,3}, {4,5}) (propose: st1, {2,3}, {1,6,7,9,5,4,6,8}) 7 (propose: {6,3,7,1,2}, {8}) 8 (propose: {9,5,3,2,6}, Φ) (propose: {1,2,3,4,5}, {6}) (propose: st1, {2,3,6}, {1,7,9,5,4,8}) 9 (propose: {6,7,1,8}, Φ) 10 (agree: {9,5,6}, Φ) (propose: {1,4,5,6}, Φ) 11 (propose: {7,1,8}, Φ) 12 (agree: {9,5}, Φ) (propose: {1,4,5}, Φ) (propose: st2, {2,3,6}, {5,1,7}) 13 (propose: {1,7,5}, Φ) 14 (agree: {5,1}, Φ) (propose: st2, {2,3,6,1,5}, {7,9,4}) (propose: {1,5,7}, Φ) 15 (agree: {7,9}, Φ) 16 (agree: {9,7}, Φ) (agree: {2,3,6,1,5}, RIG) (agree: {4,7}, Φ) Fig. 3. Example of a negotiation process at step 8), the list of accepted preferences is updated in preferences 2 and 3, because they have been accepted by all the UserAgents. The remaining preferences proposed by the NegotiatorAgent are added to the list of rejected preferences RP. For example, the tuple (1, (75 + 50)/2,{u1, u2}) is added to RP . Finally, the corresponding message is sent. When the UserAgents receive a proposal, the list of accepted preferences is evaluated with the corresponding scoring function to determine if they agree or not. For example, the application of the scoring function at step 9 results in 9 : F u1({2, 3}) = ( 100− (|60−80|+ |72−77|)/2 ) ∗min ( 1, 2/6 + 0.3 ) = 0.55 < 9 α is set to 0.3 and the degree of interest of the group (dGf) is the average of the corresponding duf. 29 Preference Degree of interest Users 5 75 u1,u3 1 62.5 u1,u2 7 80 u2 9 70 u3 4 50 u1 8 30 u2 Table 8 Ranked list of rejected preferences RP . 0.75 → no−agree F u2({2, 3}) = ( 100− (|60−50|+ |72−90|)/2 ) ∗min ( 1, 2/6 + 0.3 ) = 0.54 < 0.7 → no−agree F u3({2, 3}) = ( 100− (|60−50|+ |72−50|)/2 ) ∗min ( 1, 2/5 + 0.3 ) = 0.59 < 0.5 → agree At step 10, the NegotiatorAgent determines that there is not agreement with the current proposal and adds the preference 6 to the list of accepted pref- erences by all the UserAgents. It sends this information, also indicating that the negotiation is still in stage 1, so that the UserAgents evaluate this new proposal. However, the negotiation (internally) moves to stage 2 because the proposals of all the UserAgents were empty. Given that there is not agreement with the current proposal, the negotiation process continues at stage 2, where the NegotiatorAgent builds a proposal that contains the first Nmediation (i.e. 3) preferences in the list of rejected preferences RP (see Table 8). As the negotiation is at stage 2, each UserAgent acts according its own be- haviour. In this case, all of them are collaborative and, therefore, some pref- erences not included in its own CP u will be accepted at each negotiation round 10 , specifically, the number of preferences defined by Nuconcession. For ex- ample, the UserAgent u3 accepts preference 5 because it belongs to CP u3; given that Nu3concession = 1 and there are two remaining preferences, it chooses the first one in the list, because the NegotiatorAgent proposes the preferences in order of interest for the group. 10 The preferences that a UserAgent accepts and are not included in its CP u are underlined in Figure 3. 30 At step 15, the UserAgents evaluate the new list of accepted preferences and in this case, all UserAgents agree with the proposal. Finally, at step 16, the NegotiatorAgent determines that an agreement has been reached. The final list of accepted preferences is P G = {(2, 60), (3, 72), (6, 57), (1, 42), (5, 50)}, which is used to obtain the list of recommended items RIG. 7.2 Experimental Results Now, we summarize some experiments that we have carried out to test the performance of our negotiation mechanism when dealing with heterogeneous groups. Specifically, we have used a Tourism data set, which was developed in our research group[36]. It contains information about leisure and tourist activities in the city of Valencia (Spain). The ontology comprises 115 features structured in two levels. Information about the users was collected from 58 real users, who directly filled out a questionnaire through a web service. Specif- ically, each user gave personal details and rated the 15 features of the first level in the ontology. Moreover, in order to have data to compare the results obtained by the RS, all the users were requested to rate every item in the data set through a form that was independent from the web site. If the users had not visited the place, the rate indicated whether or not they would be interested in visiting it. In this way, we have complete feedback information about the users. We built 100 groups of different size ranging from 2 to 6 randomly selected users (20 groups for each group size). Then, we added the negotiation profile (see Table 4) to the users in these groups. Namely, for each group, we built vari- ants with all the possible combinations of self-interested/collaborative users and two Tagree levels: 50 and 75. This resulted in more that 100000 groups for testing the negotiation procedure. Some other aspects of the negotiation profile remained unchanged: Nuconcession is set to 2 and the lu levels are 0, 50, 70, 90 and 100 for all the users. We used the same utility function introduced in the previous section. In order to make an initial comparison between the results obtained with our negotiation process, we selected four variants for each group that result from the combination of (1) all the users are collaborative/self-interested and (2) with a Tagree of 50/75. These results are shown in Figure 4. In particular, these plots show the minimum F u value of all the users in each group in the four situations described above. The vertical lines separate the groups of different size (the first 20 groups have 2 users, the next 20 groups have 3 users, and so on). The horizontal line indicate the threshold for reaching an agreement (that is, the Tagree value). This implies that only if the minimum F u value is above this threshold, the group has reached an agreement with 31 0 20 40 60 80 100 2 users 3 users 4 users 5 users 6 users F u Group Size Tagree for agreement Min Fu NEG - Self-interested Min Fu NEG - Collaborative 0 20 40 60 80 100 2 users 3 users 4 users 5 users 6 users F u Group Size Tagree for agreement Min Fu NEG - Self-interested Min Fu NEG - Collaborative Fig. 4. Comparison of the minimum utility for each group self-interested and col- laborative users and a Tagree of 50 (above) and 75 (below). the corresponding P G. Therefore, in Fig. 4 (above) we can see that, when all the users are collaborative, the negotiation procedure is able to reach an agreement in most of the groups, because, during the collaborative stage, users accept preferences that are not present in their own P u, which allows the other users to increase their F u. Obviously, when the Tagree is 75, it is much more difficult to reach such an agreement and this is reflected in the results. It is interesting to see how the minimum F u value in Fig. 4 (above) is closer to the threshold line. This is due to the fact that the negotiation process finishes when all the users obtain a F u value above their own Tagree (which in this case is 50). Moreover, this can also be observed in some groups when all the users are self-interested. For example, the second group in Fig. 4 (above) has a F u value of 60 (approx.), whereas in Fig. 4 (below), this value is higher than 80, which means that the group reaches an agreement in both cases. Plots in Fig. 5 show a comparison between the minimum and the maximum F u obtained in each group, that is, the difference in utility for the least and the most satisfied users in the group, only when all the users are collaborative. It can be observed that, when the Tagree is higher (namely, 75), the maximum 32 0 20 40 60 80 100 2 users 3 users 4 users 5 users 6 users F u Group Size Tagree for agreement Min-max Fu 0 20 40 60 80 100 2 users 3 users 4 users 5 users 6 users T a g re e Group Size Tagree for agreement Min-max Fu Fig. 5. Range of utility for each group with a Tagree of 50 (above) and 75 (below) with collaborative users. utility is also higher than in the other situation (with a Tagree of 50), because, in order to reach an agreement, the P G must be more satisfactory for all the users in the group. Plots in Fig. 6 show the percentage of agreements reached in heterogeneous groups of 3 and 5 users 11 , that is, with all the possible combinations of be- haviours. These combinations are represented in the X and Y axis. The X axis indicates the Tagree value in average of the group members; the Y axis indi- cates the degree of collaboration, calculated as the number of collaborative users with respect to the total number of users in the group. As the previous figures show, when all the users are collaborative (degree of collaboration equal to 1), almost the 100% of agreements are reached, when the Tagree value is 50 (top-left circle). Obviously, as the degree of collaboration decreases and the Tagree value and the number of users in the group increase, it is more difficult to reach such an agreement. 11 We only show these figures because they are representative of what happens with groups of other sizes. 33 0 0.2 0.4 0.6 0.8 1 50 55 60 65 70 75 D e g re e o f C o lla b o ra tio n Tagree in Average 0 0.2 0.4 0.6 0.8 1 50 55 60 65 70 75 D e g re e o f C o lla b o ra tio n Tagree in Average Fig. 6. Percentage of agreements reached in heterogenous groups of 3 (above) and 5 (below) users. Regarding the computational cost of the negotiation process, it took only few milliseconds to perform it (we did not consider the communication cost). It is only slightly more costly than a centralized mechanism. We also observed that the number of iterations needed in the negotiation process depended on the following aspects: • If all the users are self-interested and an agreement is reached, the maximum number of iterations is usually given by the number of levels of negotiation in the negotiation profile (4 in our experiments). Even when the sets of preferences P u of all the users are similar, it is difficult that the distribution in the CP u is exactly the same. • If there is at least one collaborative user and an agreement is reached during the mediation stage, the number of iterations is 7, in average. • If an agreement is not reached, usually 20 iterations are needed in average, although when the P u of some users contains many preferences, it could be necessary to perform up to 26 iterations, given that, during the mediation stage, the rejected preferences are proposed progressively. In summary, in the success of the negotiation process, two main aspects have 34 an influence: • Users’ behaviour: an agreement is easier to reach when the number of col- laborative users in the group is high and their Tagree is low. • Groups homogeneity: when the P u of the group members have a high num- ber of preferences in common, it is easier to reach an agreement. On the other hand, we have observed that the number of preferences in the P u of the group members may also have an influence in this success. In fact, it seems more difficult to reach an agreement when there is a user whose P u size is significantly different from the other users. For this reason, one of our current works is focused on analyzing which internal characteristics of the groups ease or make it more difficult to reach an agreement. 7.3 Discussion In this section, we analyze some aspects of the negotiation process. During the proposal stage (stage 1), the preferences that are accepted are a subset of those that belong to the intersection of the P u of all the users. This could lead us to think that this process is quite inefficient. However, recall that we are facing a group recommendation problem where the users do not want to reveal all their preferences to the system, if it is not necessary. That is, it could be the case that after few negotiation rounds an agreement is reached and, therefore, the UserAgents have not communicated all their preferences, unlike most centralized systems that need all the preferences from the very beginning of the recommendation process. Moreover, in our GRS, the UserAgent can be configured to propose only some preferences or to propose them in a slower or faster manner (depending on the number of negotiation levels). The order in which the preferences are accepted depends on the degree of interest duf in each P u. That is, preferences with higher duf for all the User- Agents are accepted sooner. This has an influence in the final P G in case an agreement is reached during stage 1, because the P G will only contain the preferences most valued by all the users. If an agreement is not reached during stage 1, the strategy of the Negotiator- Agent when acting as a mediator, has a limited influence in the result of the negotiation, because it mainly depends on the behaviour of the agents. How- ever, as the behaviour of the UserAgents is private, nobody knows what it is going to happen. In case all the UserAgents have been configured to be self- interested, stage 2 will not have any effect to reach an agreement, given that no concessions will be made at this stage. On the contrary, if some UserAgents are collaborative, an agreement can be reached during stage 2, depending on the value of Nuconcession (which indicates the degree of cooperativeness of the 35 agent) and the heterogeneity of the group. 8 Conclusions and further work Group Recommendation can be defined as the problem of the selection, among a set of items, of those that are likely of interest to the group of users. There are different aspects that define this problem. We are interested in solving the group recommendation problem for a group of users where each user has his/her own preferences and expectations about the resulting group profile, may want to impose their preferences over the other users’ preferences or may like better to agree with the others’ proposals. Besides, both the preferences and the particular behaviour are private information. In this context, we have opted for building a MAS, where some agents act on behalf of the group members. This way, each user can define a different be- haviour for the agent, which results in a more realistic social model. Moreover, both preferences and behaviour are private information that can be revealed under certain circumnstances. We have implemented a UserAgent that can be configured with different parameters to model the behaviour that the user wants the agent to exhibit. For example, degree of cooperation, concession tactics, etc., determine the reasoning process to decide whether to accept or reject a proposal and to compose a new offer. These UserAgents work together (coordinated by the NegotiatorAgent) with the aim of building a group profile that satisfies the particular requeriments of each group member. This group profile is then used to obtain the list of recommended items. The negotiation process to obtain the group profile is a multi-party variation of the alternating-offers protocol, where the NegotiatorAgent centralizes the communication among the UserAgents. The negotiation comprises two stages: during the first stage, the UserAgents negotiate on their own to try to reach an agreement; however, if this is not possible, the NegotiatorAgents acts as a mediator during the second stage in order to facilitate such an agreement. The advantages of our solution for the resolution of the group recommendation problem are: (1) Each UserAgent may exhibit a different behaviour, which is easily con- figured by setting some parameters. (2) UserAgents do not need to communicate all their preferences to the sys- tem, because an agreement can be reached before. Moreover, they do not reveal what their behaviour is with respect to the other UserAgents. Therefore, privacy is preserved (at some extent). (3) The negotiation protocol we propose captures well the group dynamics 36 when the agents involved have different behaviours and allows them to obtain a group profile according to the preferences and expectations of all the group members. Finally, we have performed some experiments that show that this mechanism is able to give a response in this heterogeneous environment. As for further work, we are working on defining new, more complex behaviours for the UserAgent and analyzing how these behaviours affect the agreement. Additionally, we are working on incorporating techniques based on trust dur- ing the negotiation process, so that proposals are accepted or rejected depend- ing on the UserAgent that proposed them. This way, related users are favored to the detriment of others. References [1] L. Ardissono, A. Goy, G. Petrone, P. Torasso, Intrigue: personalized recommendation of tourist attractions for desktop and handset devices, Applied AI, Special Issue on Artificial Intelligence for Cultural Heritage and Digital Libraries 17 (8) (2003) 687–714. [2] C. Bartolini, C. Preist, N. Jennings, A generic software framework for automated negotiation, in: Proceedings of the AAMAS’02, 2002. [3] C. Bartolini, C. Preist, N. Jennings, A software framework for automated negotiation, in: Software Engineering for Multi-Agent Systems III, vol. 3390 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2005, pp. 213–235. [4] P. Bekkerman, K. Sarit, F. Ricci, Applying cooperative negotiation methodology to group recommendation problem, in: ECAI Workshop on Recommender systems, 2006. [5] L. Boratto, S. Carta, State-of-the-art in group recommendation and new approaches for automatic identification of groups, in: A. G. Soro A., Vargiu E., P. G. (eds.), Information Retrieval and Mining in Distributed Environments, vol. 324 of Studies in Computational Intelligence, Springer Berlin / Heidelberg, 2011, pp. 1–20. [6] Y.-L. Chen, L.-C. Cheng, C.-N. Chuang, A group recommendation system with consideration of interactions among group members, Expert Systems with Applications 34 (3) (2008) 2082–2090. [7] L. M. de Campos, J. M. Fernandez-Luna, J. F. Huete, M. A. Rueda-Morales, Managing uncertainty in group recommending processes, User Modeling and User-Adapted Interaction 19 (3) (2009) 207–242. 37 [8] J. de la Rosa, N. Hormazbal, S. Aciar, G. Lopardo, A. Trias, M. Montaner, A negotiation-style recommender based on computational ecology in open negotiation environments, IEEE Transactions on Industrial Electronics 58 (6) (2011) 2073 – 2085. [9] I. Garcia, S. Pajares, L. Sebastia, E. Onaindia, Preference elicitation techniques for group recommender systems, Information Sciences 189 (2012) 155–175. [10] I. Garcia, L. Sebastia, E. Onaindia, On the design of individual and group recommender systems for tourism, Expert Systems with Applications 38 (6) (2011) 7683–7692. [11] T. R. Gruber, A translation approach to portable ontology specifications, Knowledge Acquisition 5 (2) (1993) 199–220. [12] A. Jameson, More than the sum of its members: Challenges for group recommender systems, in: Proceedings of the International Working Conference on Advanced Visual Interfaces. Gallipoli, Italy, New York, NY, USA, 2004. [13] A. Jameson, S. Baldes, K. Thomas, Enhancing mutual awareness in group recommender systems, in: B. Mobasher, E. Sarabjot S. Anand (eds.), Proceedings of the International Joint Conference on Artificial Intelligence. Workshop on Intelligent Techniques for Web Personalization. Acapulco, Mexico, AAAI, 2003. [14] A. Jameson, S. Baldes, K. Thomas, Two methods for enhancing mutual awareness in a group recommender system, in: Proceedings of the International Working Conference on Advanced Visual Interfaces. Gallipoli, Italy, New York, NY, USA, 2004. [15] N. Jennings, P. Faratin, A. R. Lomuscio, S. Parsons, M. Wooldridge, C. Sierra, Automated negotiation: Prospects, methods and challenges, Group Decision and Negotiation 10 (2) (2001) 199–215. [16] S. Kraus, Automated negotiation and decision making in multiagent environments, in: ACAI, LNAI 2086, 2001. [17] M. Lenar, J. Sobecki, Using recommendation to improve negotiations in agent- based systems, Journal of Universal Computer Science 13 (2) (2007) 267–286. [18] H. Lieberman, N. W. Van Dyke, A. Vivacqua, Let’s browse: A collaborative browsing agent, Elsevier Science B. V 12 (1999) 427–431. [19] M. A. Lopez-Carmona, I. Marsa-Maestre, B. Alarcos, Anegsys: An automated negotiation based recommender system for local e-marketplaces, IEEE Latin America Transactions 5 (6) (2007) 409–416. [20] F. Lorenzi, A. Bazzan, M. Abel, An architecture for a multiagent recommender system in travel recommendation scenarios, in: A. Felfernig, M. Zanker (eds.), Proceedings of 3rd Workshop on Model-Based Systems (ECAI 2006). Riva del Garda, Italy, ECAI, 2006. 38 [21] F. Lorenzi, A. Bazzan, M. Abel, Truth maintenance task negotiation in multiagent recommender system for tourism, in: Proceedings of the AAAI Workshop on Intelligent Techniques for Web Personalization and Recommender Systems in E-Commerce (AAAI 2007), 2007. [22] F. Lorenzi, D. S. dos Santos, A. Bazzan, Negotiation for task allocation among agents in case-base recommender systems: a swarm-intelligence approach., in: E. Aimeur (ed.), Proceedings of the Workshop Multi-Agent Information Retrieval and Recommender Systems - Nineteenth International Conference on Artificial Intelligence (IJCAI 2005). Edimburgh, Scotland, 2005. [23] S. Macho-Gonzalez, M. Torrens, B. Faltings, A multi-agent recommender system for planning meetings, in: Fourth International Conference on Autonomous Agents, Workshop on Agent-based Recommender Systems (WARS 2000), 2000. [24] J. Masthoff, The pursuit of satisfaction: Affective state in group recommender systems, User Modeling (2005) 297–306. [25] J. Masthoff, Recommender Systems Handbook, chap. Group Recommender Systems: Combining Individual Models, Springer, 2011, pp. 677–702. [26] J. F. McCarthy, T. D. Anagnost, Musicfx: An arbiter of group preferences for computer supported collaborative workouts, in: Proceedings of the ACM 1998 Conference on CSCW, 1998. [27] K. McCarthy, L. McGinty, B. Smyth, M. Salamo, Social interaction in the cats group recommender, in: Workshop on the Social Navigation and Community based Adaptation Technologies. Dublin, Ireland, 2006. [28] A. A. Niknafs, H. B. Band, Improved win-win quiescent point algorithm: A recommender system approach, Journal of Applied Sciences 10 (23) (2010) 3084–3090. [29] M. O’Connor, D. Cosley, J. A. Konstan, J. Riedl, Polylens: a recommender system for groups of users, in: Seventh European Conference on Computer Supported Cooperative Work ECSCW, 2001. [30] M. Pazzani, D. Billsus, The Adaptive Web, chap. Content-based recommendation systems, Springer Berlin / Heidelberg, 2007, pp. 325–341. [31] C. Plua, A. Jameson, Collaborative preference elicitation in a group travel recommender system, in: Proceedings of the AH Workshop on Recommendation and Personalization in eCommerce. Malaga, Spain., 2002. [32] J. A. Recio-Garcia, G. Jimenez-Diaz, A. A. Sanchez-Ruiz, B. Diaz-Agudo, Personality aware recommendations to groups, in: Proceedings of the third ACM conference on Recommender systems (RecSys’09), 2009. [33] P. Resnick, H. R. Varian, Recommender systems, Communications of the ACM 40 (3) (1997) 56–58. 39 [34] F. Ricci, D. Cavada, Q. N. Nguyen, Integrating travel planning and on-tour support in a case-based recommender system, in: Proceedings of the Workshop on Mobile Tourism Systems (in conjunction with Mobile HCI’02), 2002. [35] A. Rubinstein, Perfect equilibrium in a bargaining model, Journal of the Econometric Society (1982) 97–109. [36] L. Sebastia, I. Garcia, E. Onaindia, C. Guzman, e-Tourism: a tourist recommendation and planning application, International Journal on Artificial Intelligence Tools (WSPC-IJAIT) 18 (5) (2009) 717–738. [37] V.-W. Soo, S.-H. Liang, Recommending a trip plan by negotiation with a software travel agent, in: International Workshop on Cooperative Information Agents (CIA), 2001. [38] J. Wohltorf, R. Cissee, A. Rieger, H. Scheunemann, Berlintainment: An agent- based serviceware framework for context-aware services, in: Proceedings of the International Symposium on Wireless Communication Systems. Mauritius., 2004. [39] M. Wooldridge, An Introduction to Multiagent Systems, John Wiley and Sons Ltd, 2002. [40] R. Zhang, S. Zhang, S. Ye, Y. Zhao, J. Ford, J. Makedon, Providing recommendations in scens, The electronic Journal for e-commerce tools and Applications 2 (4) (2009) 9. 40 
work_ybxnll4uivd3zjz5yf5ihkau6y	----	International Journal of Recent Technology and Engineering (IJRTE) International Journal of Recent Technology and Engineering (IJRTE) ISSN: 2277-3878, Volume-8 Issue-4, November 2019 9986 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D4423118419/2019©BEIESP DOI:10.35940/ijrte.D4423.118419  Abstract: In the company it was identified that the level of maturity in relation to the implementation of the systems is inconclusive, there is little confidence, low level of efficiency, and the times exceed what is stipulated. The objective was to demonstrate that the expert system significantly improves the average time, reliability levels, and efficiency in evaluations of maturity levels in the management of IT services of the Sions Company. Likewise, the CommonKADS methodology was used to build the knowledge of the expert system and SCRUM for the active monitoring of the project. Regarding the methodology of scientific research, it was based on the quantitative approach, pre-experimental design; the study population was made up of 16 evaluations of technology companies in Peru. The results showed that the expert system improved the average time in the evaluation of maturity levels by 85%, compared to the traditional system. In this way, the levels of reliability and efficiency improved by approximately 49% compared to the traditional system. In this way, the levels of reliability and efficiency improve in 49% the capacity of response, the quality of the service and the functional availability in the evaluations of the technological management. Therefore, the processes of the service management phases such as design, strategy, transition, operation and continuous improvement benefit by reducing time, costs and increasing its functionality for companies that have technological services as part of the business strategy. Therefore, the expert system would be applied in companies of different areas that have technological services, whose purpose allows to identify the current state of the service and its possible improvement over time. Keywords: Keywords: IT service management, CommonKADS, Scrum, ITIL and Expert system. I. INTRODUCTION Nowadays, technological advances have expanded and diversified worldwide, being the axis of the multiple opportunities that benefit people and companies in general, although our participation as indirect actors are not enough. The technologies are changing, extensive and diverse, which is why the effort, analysis, understanding, and evaluation of Revised Manuscript Received on November 19, 2019 * Correspondence Author Flores Zafra David, Escuela de Post Grado, Universidad César Vallejo, Lima, Perú. Email: Davidfloreszafra@gmail.com Carhuancho Mendoza Irma Milagros, Escuela de Post Grado, Universidad César Vallejo, Lima, Perú. Email: irmamilagros@yahoo.com Venturo Orbegoso Carlos Oswaldo, Escuela de Post Grado, Universidad César Vallejo, Lima, Perú. Email: cventuro2911@gmail.com Sicheri Monteverde Luis Guillermo, Facultad de Ingeniería y Negocios, Universidad Norbert Wiener, Lima, Perú. Email: luis.sicheri@uwiener.edu.pe Mendivel Landeo Ingrid, Vicerrectorado de Investigación, Universidad Nacional de Ingeniería, Lima, Perú. Email: melcris@hotmail.com the population is required, because they are transforming realities (economics, customs, businesses, students, entrepreneurs, specialists and processes) that allow improving the labor aspect and the near future of the population. In this sense, the management of IT services is an essential axis as part of the control, monitoring and use of best practices for the benefit of digital transformation. According to Axelos, he says that 70% of professionals in this area argue that organizations have drawn their strategies aligned to technological changes, but without a sincere sense of their evolution. That is to say, when implementing the new technologies, the level of maturity of the companies is not precise and they do not know the way forward to achieve sustainable growth, in this sense according to the data obtained in a study, 61% of professionals recognize the importance of aligning the strategic objectives of the companies with the technological services of the business [1,2,27]. On the other hand, it is essential to use the information technology infrastructure library as a reference framework that promotes the good practices of IT service management, as it is based on standards, rules, procedures, policies, processes, control and continuous improvement for the integration of different technological areas in organizations; in addition to strengthening communication, experience and efficiency in the response time of operational activities [3]. Likewise, all systems development requires the integration of specialized knowledge with the areas involved in the business. It also needs the use of technological tools that improve the quality, availability, usability and reliability of operational services in the shortest possible time. Consequently, technological processes must be combined with the knowledge of people, applications and business strategies. The importance of using the control mechanisms and the experience of the ITIL reference framework for proper technological management is highlighted since they improve the quality, efficiency and reliability of the services. Likewise, companies have various technological services, within which the use of an ISO or reference framework in IT is suggested to control the management of their services [4]. In Peru, Sions is one of the largest companies in IT services consulting, which has specialists in the areas of technological governance, methodologies of the GSTI and digital transformation for government areas through outsourcing service. Expert System for Information Technology Services Management Flores Zafra David, Carhuancho Mendoza Irma Milagros, Venturo Orbegoso Carlos Oswaldo, Sicheri Monteverde Luis Guillermo, Mendivel Landeo Ingrid mailto:Davidfloreszafra@gmail.com mailto:irmamilagros@yahoo.com mailto:cventuro2911@gmail.com mailto:luis.sicheri@uwiener.edu.pe mailto:melcris@hotmail.com Expert System for Information Technology Services Management 9987 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D4423118419/2019©BEIESP DOI:10.35940/ijrte.D4423.118419 Therefore, an analysis GAP was performed to determine the critical points of the business processes and evaluate the maturity levels, which is why the following problems were identified: (a) the maturity evaluations are incomplete, due to time of effort in its realization; (b) low level of reliability and efficiency in maturity assessments for the operational phases such as design, transition and operation of the IT service; and (c) delays in the analytical determination of maturity levels in operational processes. Therefore, the following general problem was raised: To what extent an expert system will improve maturity levels in the design, transition and operation phase in IT service management? II. STUDIES RELATED TO IT SERVICES MANAGEMENT Reviewing international sources; the importance of using reference frameworks such as ITIL and ISO / IEC 20000 was verified to diagnose the behavior of IT processes and provide the steps to follow to achieve a high level of maturity of services. It is also argued that the correct application of ITIL in the management of IT services, leads to improved levels of quality, efficiency and reliability in services, coupled with the use of an expert or intelligent system that automates the functions, allowing the reduction of the average times, which is why it will increase the levels of reliability, availability and efficiency in the determination as well as in the evaluation of the maturity levels of the IT services management [26,31]. The purpose of the research is to build an expert system based on rules for the determination and evaluation of the maturity levels of IT service management. The construction of the expert system will be carried out through a hybrid development methodology, that is, it will use the CommonKADS methodology for the construction of expert knowledge, and the agile SCRUM methodology for project monitoring and control. Also, the PHP programming language was used because it is a dynamic and agile framework in the construction of systems in a web environment and MySQL for database management. The expert system called “GSTI Evaluator” is in a cloud environment in the company Sions, which meets the necessary conditions for its application and achieve the planned results [23,24]. The research was carried out using two study variables, which are represented as follows: (a) expert system referring to the independent variable; (b) the management of IT services as the dependent variable. For the conceptualization of the first variable, it is validated that expert systems turn out to be a vital part of artificial intelligence because it allows the reduction of operational tasks, saving time and cost. It should be noted that technological tools seek to emulate the experience and behavior of the human specialist in some specific subject, making the expert system make decisions according to the previously programmed logic, aligned with the objectives of the organization [5, 6]. That is, the expert system will determine and evaluate the current state of the processes of IT service management in various areas of the company, to improve the duration of the evaluations and in determining the tasks be performed to grow in maturity levels. Expert systems are related to intelligent systems that are part of artificial intelligence and cognitive systems, which, in turn, are designed from the experience of a specialist in a subject. The purpose of expert systems is to simulate certain activities of an area or the automation of unconventional tasks, this logic includes some of its characteristics such as (a) flexibility; (b) reliability; (c) availability; (d) efficiency; (e) performance; (f) cost and time reduction; and (g) quality [32, 25, 29]. Therefore, expert systems are structured as follows: (a) knowledge base, which contains the logic of the human expert; (b) factual basis, charged with guarding the events of the problems; (c) inference engine, models and processes the logic to show the results; and (d) interface, which comprises the interrelation between the user and the expert system [25]. In this sense, expert systems have different types, which can be applied and aligned according to the problem. The types of systems are: (a) based on cases, these types of systems focus on reusing predefined cases as part of the solution and learning; (b) based on neural networks, consists in applying rules accompanied by fuzzy logic; (c) based on rules, is to use conditional rules to determine the results. The purpose of managing IT services is to apply the correct planning, organization, execution and control of technological services in an organization. These services are made up of processes from different areas of the business, which are accompanied by a correct management imposed by the IT government; their application will generate value in the organization, because it will improve the standards of quality, reliability, efficiency and control, for the use of management indicators around the IT operational processes. In the same context, the importance of using the ITIL methodology in the management of IT services is sustained because they manage to generate trust and experience for the business services, which is carried out in most cases by suppliers through outsourcing [22, 26]. The management of IT services corresponds to the set of strategies aligned to the delivery of technological services. That uses an approach that relates to the client, the business and the quality of the service [7]. The most used methodologies in the world of technology services are ITIL, ISO / IEC 20000 and Cobit. In this sense, the ITIL methodology proposes the improvement of quality, working times and the optimization of resources in the different phases of the technological processes of the services. In the same context, the ISO / IEC 20000 standard is fundamental because it seeks to guarantee the quality, effectiveness and efficiency of agreements to the regulations and guidelines for the operation of the services. The management of IT services includes the following elements as part of the business management: (a) people; (b) processes; (c) technical operations; and (d) operational businesses. All these must be combined for a correct management of the services, where the interaction of the people with the technology and its processes, coexists, generating the added value at the time of providing the various services [8]. International Journal of Recent Technology and Engineering (IJRTE) ISSN: 2277-3878, Volume-8 Issue-4, November 2019 9988 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D4423118419/2019©BEIESP DOI:10.35940/ijrte.D4423.118419 In this sense, companies have the power to use some of the IT service management methodologies. They try to make an alignment with the business strategies; therefore, if the business is considered a small business, it is recommended to use the underlying processes, in order to contribute to the correct management of services. Unlike large companies that have evolved in providing quality services, they try to acquire a correct evaluation of their services, considering that technologies are changing over time. To determine the evaluation of their maturity levels, they choose to recruit expert specialists to design, transform and lead the technological services, to reach the desired maturity level for all the processes implemented. In this way, a significant contribution is generated for the new technological challenges associated with IT services [30]. The proposal of the expert system will help to integrate the knowledge of the specialists in a single source, in such a way that it is possible to reduce the problems identified in the study, considering that the scope of the solution is subject to the operational phases of the life cycle of ITIL. III. REVIEW CRITERIA A. Research approach All scientific research is divided into three approaches: (a) quantitative approach: because it is objective and allows the hypothesis to be tested; (b) qualitative: for being analytical, subjective and descriptive; (c) mixed: by having a little of the first 2 approaches and focusing the analysis on the triangulation of the sources [9, 21]. They also have the peculiarity of immersing themselves in the quantitative approach, due to their easy application and adaptation in the scientific world with social reality. In this investigation, we chose to use the approach mentioned above to have pure numbers, be quantifiable and focus on a cause and effect according to the research variables [10]. B. Type of investigation All scientific research that has as its characteristic the resolution of problems, the practical analysis of the results and the revision of different theories; an investigation of the applied type is considered. Besides, it has the particularity that the problem to be reviewed is known by researchers, which is why they efficiently transfer various theories in order to determine the solution [28]. C. Research design The study corresponded to the pre-experimental design, where the independent variable called an expert system was manipulated to achieve the modification of the dependent variable that was the management of IT services [11, 10]. The steps for the application of the pre-experimental design consisted of (a) identifying the test scope: That is, the 12 IT service management processes that would be impacted with the application of the expert system were identified. (b) Identify the measurement indicators: it consists of selecting the three indicators such as time, reliability and efficiency. (c) Run the pre-test: the measurements are executed and recorded in the dependent variable (IT service management) without using the expert system; (d) run the post-test test: this activity is executed using the expert system in IT service management; and finally (e) the consolidation of the data, to then proceed with the validation and check its consistency. As part of the technology, the construction of the rules was applied using the CommonKADS methodology. These conditional rules allowed to identify the individual cases for each process, so the maturity level for each of the 12 IT service management processes can be determined. D. Population The study population consists of the formation of a set of objects, artifacts, elements, components, records, or evaluations that have characteristics such as place, time, quantity and form. In this sense, the study was made up of 16 evaluations of companies in the technology sector; there was no sample due to the small number of companies [12]. E. Method of data analysis Descriptive and inferential statistics will be used as the basis for contrasting research hypotheses. Descriptive statistics allow the interpretation and analysis of the calculation of statisticians such as variance, mean, sum, minimum and maximum values for pre and post-test [13]. On the other hand, inferential statistics allow deductions to be made on the results that add value in the hypothesis test [14, 15]. For this, the following steps have to be applied: (a) perform the consistency of the data: double mass analysis was applied to determine its consistency; (b) normality test: the Shapiro-Wilk test was used, due to the quantity of the sample, being equal to or less than 30, the Kolmogorov-Smirnov test being discarded; (c) level of significance: a level of 95% reliability and 5% will be used as a margin of error, with the aim of identifying whether the distribution of the data is parametric or non-parametric. If the test provides a parametric result, it will be decided to use the T-student test. Otherwise, the Wilcoxon test will be used, and (d) identify the sig value: this point is to test the research hypothesis. IV. RESULTS AND DISCUSSIONS All research with a quantitative approach should conduct a prior review of the data to determine consistency. These data were processed through the support of a statistical tool to determine the normality test then and thus continue with the hypothesis test. A. Consistency analysis The second mass method was applied for consistency analysis, which consisted of entering the data cumulatively and sequential order. The data was then presented in a Cartesian graph, where the formation of a line was evidenced, which is an indication of data consistency. In the case that the Cartesian graph presented a line with deviation, then there are errors or deviations in the consistency. In the study, the double mass method was used, both in the pre and post-test, based on the 3 IT management indicators and: (a) the evaluation time of maturity levels for all technological processes; (b) the level of reliability in the evaluation of the level of maturity over the technological processes; and (c) the level of efficiency in the Expert System for Information Technology Services Management 9989 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D4423118419/2019©BEIESP DOI:10.35940/ijrte.D4423.118419 evaluation of IT managed services, as shown in the consolidated table I [16]. Table- I: Consolidated indicators of IT service management. Item Time (minutes) Reliability (percentage) Efficiency (percentage) Pre-tes t Post-test Pre-tes t Post-test Pre-tes t Post-test 1 731 105 54.55% 100% 47.38% 94.29% 2 725 106 45.45% 100% 39.50% 94.34% 3 725 106 54.55% 100% 47.40% 93.40% 4 736 106 54.55% 100% 47.43% 93.40% 5 731 106 45.45% 100% 39.49% 94.34% 6 731 106 54.55% 100% 47.38% 93.40% 7 707 107 45.45% 100% 39.54% 93.46% 8 725 105 45.45% 100% 39.50% 94.29% 9 731 106 54.55% 100% 47.38% 93.40% 10 731 107 45.45% 100% 39.49% 93.46% 11 725 106 54.55% 100% 47.40% 94.34% 12 731 106 54.55% 100% 47.38% 93.40% 13 731 105 45.45% 100% 39.49% 94.29% 14 731 106 54.55% 100% 47.38% 93.40% 15 719 105 45.45% 100% 39.51% 94.29% 16 719 106 54.55% 100% 47.41% 93.40% B. Normality test As part of the inferential statistics, the normality test was performed for the three research indicators. The indicators are (a) evaluation time of IT service management; (b) level of reliability of the evaluation of IT service management; and (c) level of efficiency of the evaluation of IT service management [17]. Likewise, the results of the Shapiro-Wilk test (n <30 evaluations) applied to the three indicators of IT service management; table II shows values below 0.05, a fact that allows us to accept that the data has no distribution normal, fundamental reason to apply the non-parametric Wilcoxon Ranges test. Table- II: Shapiro-Wilk test Shapiro-Wilk test Statistical gl Sig. Time (Pre-Test) ,806 16 ,003 Time (Post-Test) ,778 16 ,001 Reliability (Pre-Test) ,638 16 ,000 Reliability (Post-Test) ,273 16 ,000 Efficiency (Pre-Test) ,642 16 ,000 Efficiency (Post-Test) ,676 16 ,000 C. Hypothesis test The hypothesis test is a tentative and not necessarily definitive answer to the solution of the problems under study [18, 14]. In the hypothesis tests, the interrelation of the study variables is evidenced, such as independent and dependent variables, which are subject to verification, verification, contrast and based on the data obtained from the population, to determine the acceptance or rejection of the theories In this sense, the hypothesis test was carried out for the 3 indicators of IT service management, which are made up of: (a) an expert system improves the evaluation time of the design, transition and design services phase operation of IT service management in Sions; (b) an expert system improves the level of reliability in the evaluation of the design, transition and operation services phase of IT services management at Sions; and (c) an expert system improves the level of efficiency in the evaluation of the design, transition and operation services phase of IT services management at Sions. The null hypothesis is the denial of each of the indicators. In this way, descriptive statistics were calculated where it is evidenced that the minimum, maximum, range, average, standard deviation and variance values for both pre-test and post-test. Thus, for the time indicator, it was verified that the average in pre and post-test, there is a considerable difference of approximately 620 minutes (Pre721; Post106), as can be seen in Fig. 1; unlike levels of reliability with a margin of difference of 49% (Pre721; Post106) as set out in Fig. 2 and for levels of efficiency with a margin of difference of 40% (Pre721; Post106), see Fig. 3. Fig. 1. Evaluation time of IT service management in minutes. Fig. 2. Reliability in the evaluation of IT service management in percentages. International Journal of Recent Technology and Engineering (IJRTE) ISSN: 2277-3878, Volume-8 Issue-4, November 2019 9990 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D4423118419/2019©BEIESP DOI:10.35940/ijrte.D4423.118419 Fig. 3. Efficiency in the evaluation of IT service management in percentages. Then the hypothesis was tested using the Wilcoxon Ranges test, as shown in Table III. It was applied to the three indicators: (a) Time indicator, Zc = -3,531, p = .000, reason why which, the null hypothesis is rejected, and it is affirmed that the expert system reduces the time of evaluation of the phase of design, transition and operation services of the management of technological services in the company Sion, 2019. The second one (b) Reliability indicator, Zc = 3,601, p = .000, a fact that makes it possible to reject the null hypothesis and affirm that an expert system improves the level of reliability in the evaluation of the design, transition and operation phase of service management of information technologies in the company Sion, 2019. The third one (c) Efficiency indicator, the distribution Zc = 3,522, p = .000, situation that generates the rejection of the null hypothesis, and demonstrate that an expert system improves the level of efficiency in the evaluation of the design, transition and operation phase of the management of information technology services in the company Sion, 2019. In summary, the comparisons of the results for the time, reliability and efficiency indicator for the management of IT services were consolidated and where it is stated that these results were for the implementation of the expert system. Table-III: Contrast test statistic of the three indicators Test statistics a Time (Post-test) - Time (Pre-test) Reliability (Post-test) - Reliability (Pre-test) Efficiency (Post-test) - Efficiency (Pre-test) Z -3,531 b -3,601 c -3,522 c Sig. asymptotic (bilateral) ,000 ,000 ,000 a. Sign rank test of Wilcoxon. b. It is based on positive ranges. c. It is based on negative ranges. The expert system implemented improves the evaluation time of the maturity levels in the design, transition and operation phases of IT service management. It should be noted that the evaluation time of maturity levels has improved significantly, decreasing from 721 minutes to 106 minutes. That is, the application of the expert system obtained an average time reduction of 85%. Likewise, what is expressed by Tanovic and Mastorakis is shared where they refer, that by using a management system, service times are improved by 35.68%, so it is verified that an expert system allows improving the evaluation time of the levels of maturity of IT service management [20]. At the same time, the rule-based expert system improves reliability and efficiency, because they increased significantly by 49% overall. Consequently, what is expressed by Yandri, Nugeraha, & Zahra is shared, where they state that when using an expert system in fuzzy logic, the determination of maturity levels is improved, generating better levels of quality, efficiency and usability for its purpose. Consequently, it was proved that an expert system allows improving the determination, reliability and efficiency in the evaluation of the levels of maturity of the management of IT services [19]. V. CONCLUSIONS An expert system for the evaluation of the maturity levels of IT technology management processes adds an average time of approximately 106 minutes compared to the traditional system used by IT managers, which represents an increase of 727 minutes in the Average time of evaluation of maturity levels, which meant a reduction of 621 minutes. That is a decrease of 85.43% of the overall average time as a comprehensive evaluation of the maturity levels of information technology services management. Therefore, the technology company increased its evaluations in the management of technological services, because before they made three evaluations for every 36 hours approximately, now it will be able to carry out 20 evaluations on average. Considering that these evaluations of maturity levels of IT service management will have an increase of approximately 85% of reliability and efficiency because before, it represented 50.57% of reliability in evaluations and 43% of efficiency using in a traditional system. Expert systems and various technologies are an essential part of R + D + I in organizations because it allows improving productivity, competitiveness and innovation, coupled with information technologies. Therefore, it should be noted that expert systems for the management of IT services turn out to be a disruptive and innovative solution that will allow the identification, evaluation and improvement of technological processes for all companies that are immersed in the world of information technologies. In summary, it is concluded that the implementation of an expert system improved the evaluation of the maturity levels for the management of IT services. REFERENCES 1. C. Surdak, A Revolução Digital, Sao Paulo: DVS Editora, 2018. 2. I. Sacolick, Driving Digital: The Leader’s Guide to Business Transformation Through Technology, New York: Força Digital, 2017. 3. A. Roeder, eBook: Simplify IT Support Management with IBM Technology Support Services, New York: IBM Technology Support Services, 2018. Expert System for Information Technology Services Management 9991 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D4423118419/2019©BEIESP DOI:10.35940/ijrte.D4423.118419 4. O. López y J. Schuler, «Implementación de buenas prácticas de CMMI – SVC e ITIL para la gestión de servicios de TI en la Pyme Agile Solutions,» Facultad de Ingeniería y Arquitectura, Lima, 2017. 5. R. Puri, Artificial Intelligence, New York: Blueprint, 2018. 6. R. Steinberg, Implementing ITSM, Arizona: Trafford Publishing, 2015. 7. B. Weed-Schertzer, Delivering ITSM for Business Maturity: A Practical Framework, Nueva York: Emerald Publishing Limited, 2019. 8. C. Agutter, ITIL Foundation Essentials: The Ultimate Revision Guide, Londres: IT Governance Limited, 2019. 9. A. Hemerijck, The Uses of Social Investment, Oxford: Oxford University Press, 2017. 10. R. Hernández, C. Fernández y P. Baptista, Metodología de la Investigación, Distrito Federal: McGraw-Hill, 2014. 11. H. Sanchez, C. Reyes y K. Méjia, Manual de términos en investigación científica, tecnológica y humanística, Lima: Vicerrectorado de Investigación Universidad Ricardo Palma, 2018. 12. G. Hercelinskyj y A. Louise, Mental Health Nursing: Applying Theory to Practice, South Melbourne: Cenveo Publisher Services, 2018. 13. J. Pérez, Los gráficos estadísticos. Sus diferentes tipos y usos para aportar claridad a un informe de investigación, Múnich: German National Library, 2018. 14. H. Llinás, Estadística Inferencial, Barranquilla: Universidad del Norte, 2018. 15. H. Duane, A. Corey y M. DeAnna, 5 Steps to a 5: AP Statistics 2019, Arizona: McGraw Hill Professional, 2018. 16. J. Brownlee, Statistical Methods for Machine Learning: Discover how to Transform Data into Knowledge with Python, Australia: Machine Learning Mastery, 2018. 17. D. Solíz, Cómo hacer un Perfil Proyecto de Investigación Científica, Bloomington: Palibrio, 2019. 18. R. Yandri, D. Nugeraha y A. Zahra, «Evaluation Model for the Implementation of Information Technology Service Management using Fuzzy ITIL,» Procedia Computer Science, vol. Volume 157, nº 10.1016/j.procs.2019.08.169, pp. 290-297, 2019. 19. A. Tanovic y N. Mastorakis, «Advantage of using Service Desk Management Systems in real organizations,» International Journal of Economics and Management Systems, vol. 1, nº ISSN: 2367-8925, pp. 81-86, 2016. 20. H. Ulloa, M. Asunción, M. Nares y S. Gutiérrez, «Importance of Qualitative and Quantitative Research for Education,» Educateconciencia, vol. 16, nº ISSN: 2007-6347 , pp. 163-174, 20 Julio 2017. 21. C. Rivera, «Aplicación ITIL y su efecto en la gestión de resolución de incidencias en el área de soporte de la empresa MDP consulting,» Repositorio Universidad César Vallejo, Lima, 2019. 22. L. Quintero y H. Peña, «Modelo basado en ITIL para la gestión de los servicios de TI en la cooperativa de caficultores de Manizales,» Universidad Autónoma de Manizales, Manizales, 2017. 23. H. Norhaidah y H. Nurdatillah, «PHP Frameworks Usability in Web Application Development,» International Journal of Recent Technology and Engineering, vol. 8, nº ISSN: 2277-3878, pp. 109-116, 2019. 24. F. Klashanov, «Artificial Intelligence and Organizing Decision in Construction,» Procedia Engineering., vol. 165, nº doi: 10.1016/j.proeng.2016.11.813 , pp. 1016-1020, 2016. 25. C. Dobbins y R. Rov, «Project Management in Information Technology: case study on the implementation and evaluation of this tool in multimarket investment fund,» Revista de Tecnologia Aplicada, vol. 7, nº ISSN: 2237-3713, pp. 36-51, 2018. 26. Axelos, «Axelos.com,» Axelos, 10 July 2019. [En línea]. Available: https://www.axelos.com/best-practice-solutions/itil. [Último acceso: 10 July 2019]. 27. Lifeder, «https://www.lifeder.com,» Lifeder, 30 Octubre 2019. [En línea]. Available: https://www.lifeder.com/investigacion-aplicada/. [Último acceso: 30 Octubre 2019]. 28. V. Chandra, V. Kantharao, J. Sastry y V. Bala, «Expert system for building Cognitive model of a student using Crypt Arithmetic game and for Career Assessment,» International Journal of Recent Technology and Engineering, vol. 7, nº ISSN: 2277-3878, pp. 684-689, 2019. 29. V. Valencia, E. Fernández y L. Usero, «Applicability of the Maturity Model for IT Service Outsourcing in Higher Education Institutions,» International Journal of Advanced Computer Science and Applications, vol. 5, nº 10.14569/IJACSA.2014.050707, pp. 41-50, 2014. 30. R. Lia, «ITSM: a sucess case of the Triple Helix Model,» Revista de Administração da Universidade, vol. 7, nº doi: 10.5902/1983465911460, pp. 55-69, 2014. 31. J. Casas, Guía para la realización de un estudio ambiental: El caso de la cuenca del río Adra, Almería: Edual, 2017. 32. S. Makridakis, «The forthcoming Artificial Intelligence (AI) revolution: Its impact on society and firms,» Elsevier Ltd., pp. 46-60, 2018. AUTHORS PROFILE Flores Zafra David Professional in Systems Engineering with a master’s degree in Information Technology Management. Nowadays, I am continuing my doctoral studies in Administration at César Vallejo University. I have more than 16 years of experience in the area of information technologies and Information technology projects. I also have Microsoft international certification as "Microsoft Professional Certificated". I currently work as a teacher in the faculty of business and systems engineering at Cesar Vallejo University - North Lima. In addition to working as Project Manager at IBM del Peru for the account of our client Banco de Crédito del Perú in the management and renovation projects of technological infrastructure in Peru and Brazil. Carhuancho Mendoza Irma Milagros Bachelor of Business Administration, with a master’s degree in Finance and a Ph.D. in Administration from the National University Federico Villarreal, Master in Virtual Learning Environments from the University of Panama, a Ph.D. candidate in Administration from the University of Celaya - Mexico, Post Doctorate in Finance from AIU. I currently work at the César Vallejo University in the Postgraduate School and the Faculty of Engineering and Business of the Norbert Wiener Private University. Author of the book Methodology for the Investigation Holistic – 2019. Speaker in the Congress International Investigation Multidisciplinary - Guayaquil 2019, specialist in the finances and research scientific quantitative, qualitative and mixed, in addition to advising research work and service companies. Carlos Oswaldo Venturo Orbegoso Mechanical Engineer from the National University of Trujillo, Master in Public Management from the European Center of Innovation and Management (EUCIM Business School), Master in Public Management from the San Martín de Porres University, Master in Strategic Business Administration from the Pontifical Catholic University of Peru - CENTRUM, Master in University Teaching and Doctor of Education from the César Vallejo University, candidate for the degree of Doctor of Administration at the University of Celaya - Mexico. He is currently the director of the Postgraduate School of Lima at the César Vallejo University, author of the book Governance Conference Global Meeting-Peru 2017, Minutes of Congress and others. Luis Guillermo Sicheri Monteverde Doctor Philosophy Leisure Engineering Ph. D. from Cambridge International University, Bachelor in Hospitality at Cornell University, Ithaca, New York, Business Administration studies from the University of Toronto in Canada, Honorary Doctorate and Master in Business Administration and Management from the University of Tumbes, Bachelor of Tourism and Hospitality, candidate for Doctor of Administration in Tourism and Hospitality and Senior Professor at the University of San Martín de Porres, awarded the "Dolphin of Tourism" for the outstanding work in favor of the promotion and Tourism development, International Journal of Recent Technology and Engineering (IJRTE) ISSN: 2277-3878, Volume-8 Issue-4, November 2019 9992 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: D4423118419/2019©BEIESP DOI:10.35940/ijrte.D4423.118419 President of Cenfotur for more than 7 years, Past Dean of the Faculty of "Human Sciences" at the Scientific University of the South, Director of the Grant 18 program - UCSUR, Director of CANATUR and Dean of the Faculty of Engineering and Business from Norbert Wiener Private University. Mendivel Landeo Ingrid Professional in Education with a master’s degree in Public Management and I am continuing for the degree of Doctor of Administration at the César Vallejo University. I currently work in the administrative area at the National University of Engineering, in the area of logistics - supply of the Vice-Rectorate of Research. Certified professional of the organism in charge of Contracts of Peru by the Supervisory Organism of State Contracting - OSCE with extensive experience of more than ten years, member of the different Special Committees in the different bidding processes of the National University of Engineering. Author of the book “Faustiniana Anthology” literary creation workshop, 2012 Promotion of the faculty of education in the specialty of English language, communication and language. 
work_ydnlz7hxyjcpjfs7ldp3sy4upq	----	NASA Technical Reports Server (NTRS) 
work_yeyqio5mnjehhhco2g5avg33ii	----	Evolving linear transformations with a rotation-angles/scaling representation Evolving linear transformations with a rotation-angles/scaling representation A. Echeverría a,⇑, J.M. Valls b, R. Aler b a Universidad Autónoma de Madrid, Madrid, Spain b Universidad Carlos III de Madrid, Madrid, Spain a r t i c l e i n f o a b s t r a c t Keywords: Data transformation CMA-ES Rotation angles Scaling Classification ⇑ Corresponding author. E-mail addresses: alejandro.echeverria@uam.es f.uc3m.es (J.M. Valls), aler@inf.uc3m.es (R. Aler). Similarity between patterns is commonly used in many distance-based classification algorithms like KNN or RBF. Generalized Euclidean Distances (GED) can be optimized in order to improve the classification success rate in distance-based algorithms. This idea can be extended to any classification algorithm, because it can be shown that a GEDs is equivalent to a linear transformations of the dataset. In this paper, the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is applied to the optimization of linear transformations represented as matrices. The method has been tested on several domains and results show that the classification success rate can be improved for some of them. However, in some domains, diagonal matrices get higher accuracies than full square ones. In order to solve this problem, we propose in the second part of the paper to represent linear transformations by means of rotation angles and scal- ing factors, based on the Singular Value Decomposition theorem (SVD). This new representation solves the problems found in the former part. 1. Introduction Many classification algorithms use the concept of distance or similarity between patterns. This is specially true for local classifi- cation methods such as Radial Basis Neural Networks (Moody & Darken, 1989) or Nearest Neighbor classifiers (Cover & Hart, 1967). Typically, the Euclidean distance is used but this is not nec- essarily the best option for all classification domains. Therefore, finding the most appropriate distance for a classification task may be an important part in obtaining high classification rates. The Generalized Euclidean Distance (GED), displayed in Eq. (1), is a family of distances that depends on a parameter S, where S is a symmetric positive definite square matrix. This distance is usually called Mahalanobis distance (Atkenson, Moore, & Schaal, 1997; Tou & Gonzalez, 1974; Weisberg, 1985) although originally this term applied to cases where S is the inverse of the covariance matrix of the data. Eq. (1) shows the GED dij between vectors xi and xj. dij ¼ ½ðxi � xjÞ T Sðxi � xjÞ� 1=2 ð1Þ In Valls, Aler, and Fernández (2007), Genetic Algorithms have been used to optimize GEDs distance matrices for Radial Basis Neu- ral Networks. However, this approach can only be used in classifi- (A. Echeverría), jvalls@in cation algorithms that explicitly use distances, like RBNN, but not others like C4.5. In order to make the approach more general, and by taking into account that a symmetric positive definite ma- trix S can always be decomposed into MTM, for some matrix M, the definition of a GED results in Eq. (2). dij ¼ ½ðxi � xjÞ T MT Mðxi � xjÞ� 1=2 ð2Þ It can be shown that Eq. (2) is equivalent to computing an Euclidean distance on a transformed dataset, where the new pat- terns in the new space are linear transformations of the original data: x0 = Mx. Thus, the GED can be redefined as shown in Eq. (3). dij ¼ ½ðMxi � MxjÞ TðMxi � MxjÞ� 1=2 ¼ ðx0i � x 0 jÞ T x0i � x 0 j � �h i1=2 ð3Þ Table 1 displays several examples of such transformations: (a) rotation of data points by h degrees, (b) scaling of dimensions, and (c) projection onto the y coordinate. The main goal of this paper is to use an optimization method, the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), to look for a transformation matrix M as defined in Eq. (3) (instead of a GED) in order to improve the classification rate of a classification algorithm. CMA-ES is one of the best evolutionary techniques for continuous optimization in difficult non-linear domains (Hansen & Ostermeier, 2001; Ostermeier, Gawelczyk, & Hansen, 1994). CMA-ES is an Evolution Strategy where search is guided by a covari- ance matrix, which is adjusted and updated during the search pro- cess. CMA-ES works well in both local and global optimization 1 mailto:alejandro.echeverria@uam.es mailto:jvalls@inf.uc3m.es mailto:jvalls@inf.uc3m.es mailto:aler@inf.uc3m.es http://dx.doi.org/10.1016/j.eswa.2011.09.015 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Cita bibliográfica Published in: Expert systems with applications 39 (2012), 3276-3282 Table 1 Examples of linear transformations: (a) rotation, (b) scaling, (c) projection. ðaÞ M ¼ cos h sin h �sin h cos h � � ðbÞ M ¼ k 0 0 1 � � ðcÞ M ¼ 0 0 0 1 � � tasks. One of the most interesting features of CMA-ES for our purposes is the self-adaptation of mutation (the update of the covariance matrix), allowing for a finely tuned local search at the end of the optimization process. In this paper, we will use CMA- ES to optimize the transformation matrix M for a classifier. Any learning algorithm could have been used but in the present work we will consider a nearest neighbor algorithm (KNN) with K = 1. This paper reports the results of experiments using diagonal and square matrices M.1 Experiments show that classification rates can be improved by both kind of matrices, but that in some cases, diagonal matrices outperform square matrices. This outcome is rather unexpected given that diagonal matrices are a subset of square ones. In order to deal with this issue, the second part of the paper tests a second representation of transformation matrices by using the Singular Value Decomposition theorem (SVD). The SVD allows to decompose any real matrix into two rotation and one scaling matrices, as displayed in Eq. (4) M ¼ URV� ð4Þ where V⁄ is the conjugate transpose of V. In our case, all the matri- ces have real values and then V⁄ = VT. Thus, M ¼ URV T ð5Þ The SVD is valid for any matrix, including rectangular ones. In this paper, matrices M are always square. In that case, U and V are both n � n unitary matrices and R is an n � n diagonal matrix with nonnegative real numbers on the diagonal. Both U and VT are rotation matrices, that is, they transform data by performing a rotation in n-dimensions. This means that matrix M works by first rotating the data (VT), then scaling it (R), and then rotating it again (U). However, let us remember that in this paper, we are using a nearest neighbor algorithm, which is rotation-invariant. This means that the final rotation U will not affect the classification re- sults (this is also true of many other classification algorithms like Fisher Discriminant, Neural Networks, but not of others like C4.5). Therefore, in the current work and without losing generality, we will consider matrices of the form RVT (i.e. a rotation followed by a scaling). Thus, our second method will evolve matrices R and VT separately, instead of letting CMA-ES evolve matrix M. More- over, the rotation matrix VT can be encoded very naturally, by using explicitly the rotation angles. For instance, transformation (a) in Table 1 is a rotation h in two dimensions represented in matrix form. But this transformation can be represented more naturally by encoding just the rotation angle h. We expected that this natural representation of linear transfor- mations will not fall into the same pitfalls than the square matrices mentioned before: diagonal matrices outperformed square ones. The rationale for this is that SVD allows to factor the diagonal part 1 In this paper, the term ‘square matrix’ will refer to matrices where all its components can be non-zero, in opposition to diagonal matrices, where only the diagonal components are non-zero. (which is basically a scaling of the data, represented by matrix R) and the rotation part. By allowing the algorithm to separate both components, it is expected that CMA-ES will be able to find the proper scaling when only a scaling is needed. Experimental results will show that this is the case. The structure of this article is as follows. Section 2 describes the CMA-ES algorithm. Section 3 describes the proposed methods, both evolving directly the transformation matrices and evolving the rotation angles and the scaling factors. Section 4 describes the syn- thetic and real domains that have been used to test the approach, and also reports the results of the experiments. Finally, Section 5 draws some conclusions and points to future work. 2. The Covariance Matrix Adaptation Evolution Strategy (CMA-ES) The algorithm CMA-ES (Covariance Matrix Adaptation Evolu- tion Strategy) was proposed by Hansen (2009) and Hansen and Ostermeier (2001). It is an evolution strategy for difficult optimiza- tion problems in continuous domains, characterized by the use of a covariance matrix to guide the search process. In this section, we will give a very short overview of a (l, k) CMA-ES, where l is the number of parents and k the number of offspring. The full algo- rithm is quite complex, so the interested reader should consult the CMA tutorial for a complete description (Hansen, 2009). CMA-ES estimates a probability distribution from the best per- forming samples in order to guide the search towards promising regions of the search space. Let us suppose that we desire to opti- mize a fitness function f(x) : Rp ? R, where p is the dimensionality of the problem. CMA-ES basic algorithm follows a randomized black box search, described as follows: 1. Initialize distribution parameters h. 2. For generation (iteration) t = 0, 1, 2, . . .: (a) Sample k points from distribution P(x—h(t)) = x1, . . . , xk. (b) Evaluate fitness f of x1, . . . , xk. (c) Update distribution parameters h based on the best per- formers x1, . . . , xl (f(x1) 6 f(x2) 6 � � � f(xl), for a minimization problem). In CMA-ES, the probability distribution to be estimated is a mul- tivariate normal N(m, d2C), whose parameters h are the mean m and the covariance matrix d2C. The mean represents the current loca- tion of the search and it moves towards better locations as search progresses. The covariance matrix controls mutations and is used to guide the search. In some sense, the covariance matrix ‘‘points’’ towards better solutions and is estimated from past samples xi that performed well with respect to f. It is important to remark that CMA-ES decomposes the total covariance matrix into a covariance matrix C and the overall variance d2, also called the step-size con- trol. The designers of the algorithm have found it useful to control the size of the mutation steps d2 independently of the direction of the search C. The reason is that estimating C requires many sam- ples, and therefore takes longer to be adapted, whereas d2 needs to be adapted within shorter time-frames. It has to be clarified that the mean, the step-size, and the covari- ance matrix are updated every generation t, so they should be writ- ten as m(t), d(t), C(t). However, in order to clarify the notation, the generation t will be omitted, by understanding that if the parame- ter appears on the right side of an equation, it belongs to genera- tion t and to generation t + 1 if it appears on the left side. Sampling k points from N(m, d2C) can be easily done, according to Eq. (6) (where I is the identity matrix and z 2 Rp is a random vector generated from N(0, I), an spherical multivariate normal distribution) 2 xi m þ d � ffiffiffi C p � z ð6Þ After sampling, the best l samples xi are used to update the mean (m), the covariance matrix (C), and the step-size (d). The up- date of the mean is straightforward, by means of Eq. (7), where wi are user selected weights m Xl i¼1 wi xi ð7Þ The update of C is done according to Eq. (8) C ð1 � ccovÞ � C þ a1 CovðscÞþ a2 Covðxi¼1���lÞ ð8Þ where sc is called the evolution path, and summarizes the best xi samples from the beginning of the run (not just the ones in the cur- rent generation). In some sense, it contains a summary of the global performance so far. Cov(sc) is the covariance due to the evolution path and Cov(xi=1� � �l) is the covariance due to the best performers in the current generation. ccov 2 (0, 1] and a1 + a2 = ccov are learning ratios. And finally, with respect to the adaptation of the step-size, d may be either too long (and therefore it should be shortened) or too short (then it should be made larger). CMA-ES updates d by tak- ing into account the samples xi along the evolution path. When the xi in the path are correlated – all xi are going in the same direction – the same distance could be covered in fewer but longer steps. Therefore, d should be increased. On the other hand, when the evo- lution path is anticorrelated – the xi go in contrary directions and cancel each other- the step-size d should be made smaller so that region of the search space can be explored with a finer grain. In short, it is desirable that there are no correlations among samples along an evolution path. This can be done by comparing the length of the evolution path with the length of an evolution path under random selection, because random selection produces uncorre- lated samples. Thus, the step-size is updated by means of Eq. (9) d d � exp b � ksdk kE½Nð0; IÞ�k � 1 � �� � ð9Þ where b is a parameter that determines the step size variation be- tween successive generations and, kE[N(0, I)]k is the expectation of the length of a N(0, I) distributed random vector in Rp. Thus, Eq. (9) updates d by comparing the length of the current evolutionary path ksdk and the expected length of an evolutionary path under random selection kE[N(0, I)]k. 3. The method In this paper we have applied CMA-ES for finding data transfor- mations that minimize the classification error (i.e. maximize the classification rate) of a learning technique. Two different ways of representing transformations have been tested. In the first method, transformations matrices coefficients are directly coded. In the sec- ond method, transformations are coded as rotation matrices and scalings. CMA-ES is extensively described in Ostermeier et al. (1994 ) and Hansen and Ostermeier (2001).2 3.1. First method: square and diagonal matrices In the first method, we have considered two types of matrices: diagonal matrices (where all elements outside the diagonal are zero) and arbitrary non-diagonal matrices (n � n square matrices). Using diagonal matrices amounts to just a weighting of the attributes by the corresponding element in the diagonal, as shown in Eq. (10) 2 We have used the Matlab code available at http://www.lri.fr/�hansen/ cmaes_inmatlab.html. k1 0 0 k2 � � � x1 x2 � � ¼ k1 � x1 k2 � x2 � � ð10Þ The reason for testing these two types of structures is to check whether square matrices are actually useful beyond a mere attri- bute weighting. Square matrices involve fitting many parameters (n � n) and it is important to know whether they improve accuracy or rather produce overfitting because of the extra degrees of free- dom, or underfitting because of not being able to adjust properly all the parameters. In order to describe the application of any evolutionary algo- rithm, three elements have to be defined: the parameters, the chro- mosome, and the fitness function. With regard to parameter setting, CMA-ES is almost parameter- free in the sense that it self-adjusts some of its own parameters during the search and also provide some reasonable values for the rest of parameters that usually work well. Typically, CMA-ES starts from a random individual, but in order to reduce some ran- domness in results, CMA-ES will start from the identity matrix for all the runs. This is equivalent to starting with no transformation (or starting with the Euclidean distance). Also, the initial mutation step size has been fixed to 1.0 (although CMA-ES will adjust this value in the course of the search). Also, a stopping criterion has to be provided. In this paper, we have checked empirically that 3000 fitness computations are enough for all domains. With respect to the chromosome, CMA-ES individuals are matrices that represent linear transformations. Matrices can be di- rectly encoded by flattening them (i.e. converting it into an array). If the matrix is square, the chromosome has n2 components and is represented in Eq. (11) v ¼ðx11; x12; . . . ; x21; x22; . . . ; xnnÞ ð11Þ If the matrix is diagonal, the chromosome is represented in Eq. (12), and has n components v ¼ðx11; x22; x33; . . . ; xnnÞ ð12Þ Finally, in order to evaluate each individual fitness, the original training dataset is transformed by the matrix encoded in the individual: T M ¼ M � T ð13Þ where M is the transformation matrix corresponding to the individ- ual, T is the original training set and TM is the transformed training dataset. Then, a classifier is trained using TM. Any rotation-invariant learning algorithm L could have been used. In this paper, neighbor- hood-based algorithm KNN (with k = 1) has been selected. It is easy to think what kind of linear transformations could be useful for KNN, and this has helped in the design of some of the synthetic do- mains that will be described in the experimental section (Section 4). Fitness is then computed by performing a fivefold crossvalidation on TM (Eq. (14)). Crossvalidation is required here because the train- ing error of KNN is always zero EM ¼ errorðL; T MÞ ð14Þ 3.2. Second method: rotation angles and scaling matrix In the second method, a representation of rotation angles and scaling factors is directly used, to check whether the poor results produced in some domains with the square matrices of the first method, can be improved. Let us remember that in this second method, transformation matrices M will be represented by means of a scaling (diagonal) matrix R and a rotation matrix V (Eq. (15)). The later will be represented directly by rotation angles M ¼ RV T ð15Þ 3 http://www.lri.fr/~hansen/cmaes_inmatlab.html http://www.lri.fr/~hansen/cmaes_inmatlab.html http://www.lri.fr/~hansen/cmaes_inmatlab.html The diagonal of the scaling matrix R will be represented as vec- tor s = (s1, . . . , sn), where n is the dimension of the problem. Matrix V will be represented as a vector of rotation angles r = (r1, . . . , ra), where a = n(n � 1)/2, are coded in the chromosome v as follows: v ¼ðr; sÞ¼ ðr1; . . . ; ra; s1; . . . ; snÞ; where the first a elements are the rotation angles and the last n are the scaling factors. It means that our algorithm will perform n(n � 1)/2 rotations and n scalings on data. Algorithmically, the generation of a n-dimensional rotation can be performed by multi- plication of a rotation matrices R (a). In two dimensions, the form of these matrices that rotate a given vector by a counterclockwise angle is: RðaÞ¼ cos a �sin a sin a cos a � � ð16Þ In three dimensions, rotations of the three axes in a counter- clockwise direction give the matrices in Eqs. (17)–(19), respec- tively (Arfken, 1985; Goldstein, 1980) RxðaÞ¼ 1 0 0 0 cos a �sin a 0 sin a cos a 0 B@ 1 CA ð17Þ RyðaÞ¼ cos a 0 sin a 0 1 0 �sin a 0 cos a 0 B@ 1 CA ð18Þ RzðaÞ¼ cos a �sin a 0 sin a cos a 0 0 0 1 0 B@ 1 CA ð19Þ The composition of rotation angles matrices Rx(a)Ry(b)Rz(c) in three dimensions would give the transformation matrix of Eq. (20): cosbcosc cos bsinc sin b �cosasinc�coscsinasin b cosacosc�sinasinbsinc cosbsina sinasinc�cosacoscsin b �coscsina�cosasinbsinc cosacosc 0 B@ 1 CA ð20Þ As we can see in Bck (1996), only a few lines of code (Schwefel, 1980) are sufficient to perform these calculations. To perform our rotation-scaling transformation routine, we have added a scale transformation to this code producing Algorithm 1, explained be- low. First, the main loop is entered for all data instances p initializ- ing a to the number of rotation angles, i.e., a = n(n � 1)/2, where n is the domain dimension. Then, next two loops (lines 3 and 6) move data T indexes to perform 2D rotation obtaining a rotated data TR (lines 8 and 9) in a counterclockwise direction, using a angles a con- tained in vector r (line 7). At the end of each iteration for every data instance, we perform a scale transformation multiplying it by vec- tor s producing a final transformed data TM (line 14). Table 2 UCI domains characteristics. Domain Instances Attributes Classes Blood (BL) 748 5 2 Car (CR) 1728 6 4 Diabetes (DB) 768 8 2 Ionosphere (IO) 351 34 2 Iris (IR) 150 4 3 Wine (WI) 178 13 3 Algorithm 1. Rotation Angles Routine 1: for i = 1 to p do 2: a = n(n � 1)/2 3: for j = 1 to n � 1 do 4: n1 = n � j 5: n2 = n 6: for k = 1 to j do 7: a = r(a) 8: TR(i, n2) = T(i, n1) sin(a) + T(i, n2) cos(a) 9: TR(i, n1) = T(i, n1) cos(a) � T(i, n2) sin(a) 10: n2 = n2 � 1 11: a = a � 1 12: end for 13: end for 14: TM(i) = TR(i)s0 15: end for The parameters setting for CMA-ES in the second method is similar to the first one: it starts with all the rotation angles set to zero and the scaling factors set to 1 (which corresponds to the identity transformation matrix). As in the first method, the initial step-size is set to 1.0 and the number of matrix evaluations al- lowed is the same than in the first method for each domain. For the same reasons explained in Section 3.1 and with the aim of comparing both methods, KNN (with k = 1) is also used as clas- sification algorithm. 4. Experiments In this section, we will carry out experiments on several artifi- cial and real world domains. 4.1. Domains We have used four synthetic domains and six real world do- mains. All of them correspond to classification problems and have numerical attributes. They are described next. 4.1.1. Artificial data domains We have used a well-known synthetic dataset, the Ripley (Rip- ley, 1996) data domain, which has been widely used in the Ma- chine Learning literature, and three more artificial domains that we have called RandomAttr, Straight-0 and Straight-45, specifi- cally designed to check if the approach works properly. In the Ripley (RP) dataset each pattern has two real-valued coordinates and a class which can be 0 or 1. Each class corresponds to a bimodal distribution that is a balanced composition of two normal distributions. Covariance matrices are identical for all the distributions and the centers are different. One of the issues that make this domain interesting is the big overlap existing between both classes. RandomAttr (RA) is a two-class domain with four real-valued attributes x1, x2, x3, x4. The examples have been randomly gener- ated following an uniform distribution in the interval [0, 1] for attributes x1, x2 and the interval [0, 100] for attributes x3, x4. If x1 < x2 then the example is labeled as class ‘1’. Otherwise, if x1 > x2 the example belongs to class ‘0’. Thus, attributes x3 and x4 are irrelevant. Because the ranges of irrelevant attributes are much bigger, the classification accuracy of KNN is very bad (about 50%). The dataset is composed of 300 examples, 150 from each class. Straight-45 (S45) is a two-class domain with two real-valued attributes. The examples have been generated in this way: initially 100 examples of class 1 are located as regular intervals in a straight 4 Table 3 Classification rate (percentage) with KNN (k = 1) for the original dataset and for the transformed data when square and diagonal matrix transformations are used. Original data CMA-ES diagonal CMA-ES square Straight-0 (S0) 9.20 ± 2.26 �99.70 ± 0.79 �99.85 ± 0.24 Straight-45 (S45) 9.20 ± 1.62 8.95 ± 1.89 �99.15 ± 0.58 RandomAtt (RA) 50.50 ± 2.38 �95.60 ± 1.81 �79.63 ± 6.28 Ripley (RP) 88.58 ± 0.36 88.91 ± 0.44 88.41 ± 0.57 Blood (BL) 68.41 ± 0.57 67.92 ± 0.69 68.12 ± 1.01 Car (CR) 87.49 ± 0.17 �95.82 ± 0.33 �97.52 ± 0.18 Diabetes (DB) 68.11 ± 1.02 67.10 ± 1.51 67.28 ± 0.95 Ionosphere (IO) 86.19 ± 0.39 89.94 ± 1.57 88.31 ± 0.63 Iris (IR) 96.00 ± 0.00 94.80 ± 0.76 95.27 ± 1.24 Wine (WI) 75.99 ± 1.47 �94.39 ± 1.14 �85.69 ± 1.64 Significantly better/ worse 4/0 (of 10) 5/0 (of 10) line passing through the origin (0, 0) with an angle of 45� respect to the horizontal axe. The distance between two consecutive points is 1. One hundred examples of class 0 are generated in the same way in a parallel straight line passing through the point (0,�1) in such a way that the nearest point of a given point always belongs to the opposite class, because it is located in the opposite parallel straight line. Then all points are perturbed by adding to each coordinate a random number uniformly distributed in [�0.5, 0.5]. The idea behind this domain is that the nearest neighbor algo- rithm will achieve a very bad result because most of the times the nearest neighbor of a given point belongs to the opposite class. But certain transformations of the data involving rotations and coordinate scaling will allow a good classification rate. Straight-0 (S0) is very similar to Straight-45. The only differ- ence is that all the points have been rotated 45� clockwise. The motivation for using this domain is that in this case with a simpler transformation the data could be properly classified because no rotation is needed. In short, Straight-45 requires both rotation and scaling (a square matrix) whereas Straight-0 requires scaling only (a diagonal matrix). 4.1.2. Real world data domains We have used the well known Blood (BL), Car (CR), Diabetes (DB), Ionosphere (IO), Iris (IR) and Wine (WI) domains from the UCI Machine Learning Repository.3 Table 2 displays the characteris- tics of these UCI domains. 4.2. Experimental results. Square vs. diagonal matrices In this subsection we show the experimental results obtained when square and diagonal matrix transformations are evolved by CMA-ES and KNN (with k = 1) is used as classifier in the transformed space. The parameters for CMA-ES are the following: the initial stan- dard deviation has been set to 1.0 for all the experiments. The max- imum number of fitness evaluations has been set to the same value (3000) for both methods by means of preliminary experiments. The rest of CMA-ES parameters are set to their default values. In all domains we compare the classification results in three sit- uations: for each fold, KNN classifies the original data, the data transformed by a diagonal matrix optimized by CMA-ES and by a square matrix optimized by CMA-ES. In all cases linear transforma- tions are done by means of complete matrices and thus, the dimen- sion of the transformed space remains unaltered. Table 3 shows the mean classification accuracy rates obtained for all the domains and their standard deviations. The means have been obtained by averaging 10 executions of a 10-fold crossvalida- tion procedure. The best results have been marked with a dagger. A Corrected Repeated 10 � 10-fold Crossvalidation T-Test (Bouckaert & Frank, 2004) has been used in all cases for testing if differences between an algorithm and the results obtained with the original data are statistically significant (a = 0.05). Statistically significant results are marked with a dagger (�). The last row of Table 3 counts in how many domains the differences are significant. In the Straight-0 domain, with a simple transformation of the space, just compressing the x1 coordinate, a good classification rate can be achieved because points belonging to the same class can get as close to each other as needed. This simple transformation can be done with a diagonal matrix and therefore with a square matrix too. The results show that KNN only obtains a classification rate of 9.2% on the original dataset, but if the data is linearly trans- formed with a diagonal matrix the rate is 99.7% and 99.8% if a square matrix is used. In this case both optimization methods achieve the same results for each matrix type. 3 http://archive.ics.uci.edu/ml/. In the Straight-45 domain, we see that a more complex trans- formation must be done because it is not enough to scale the coor- dinates but to rotate the points as well. This linear transformation cannot be obtained with a diagonal matrix. The second row of Ta- ble 3 shows the results as expected: KNN obtains a very bad clas- sification accuracy on the original data (9.20%). A diagonal matrix is useless to obtain an adequate transformation and we can see that the results are very bad too, around 8.95%. On the contrary, when a square matrix is used, the results are near to 100%. The RandomAttr domain has two irrelevant attributes whose numeric range is much bigger that the relevant attributes range. That is the reason why the accuracy of KNN on the original data is 50.5%. Scaling the irrelevant attributes should be enough to at- tain a much better accuracy, thus both diagonal and square matri- ces should be appropriate. The results show that diagonal matrices obtain results over 90%. It can also be seen that in this domain, square matrices do not perform as well as diagonal matrices. With regard to the remaining domains, CMA-ES diagonal ob- tains significatively better results than KNN on the original data for Car (95.82% vs. 87.49%), and Wine (94.39% vs. 75.99%). CMA- ES square also obtains significatively better results than KNN on the original data for the same domains: Car, 97.52% vs. 87.49% and Wine, 85.69% vs. 75.99%. Summarizing the results, it can be observed that CMA-ES diag- onal and CMA-ES square behave as expected in the artificial do- mains, except in RandomAttr, where CMA-ES square obtains much worse results than expected. Also, it is never the case that the original data significantly outperforms any of the other two methods. There are some cases where diagonal matrices largely outperform square ones (RandomAttr and Wine). In those two do- mains, the square matrix results are much worse than the diagonal ones. This is remarkable, given that the set of square matrices in- clude diagonal ones. This issue deserves more attention because it would be useful to apply just one method to all domains: the evolution of square matrices. Square matrices have more parameters to fit than diagonal ones, but the worse results of square matrices in RandomAttr and Wine is not a case of overfitting. We have extended the number of evalua- tions from 3000 to 10,000, and results in RandomAttr improve from 79.63% to 88.86%, and in Wine from 85.69% to 87.14%. These improvements hint that the issue here is not overfitting but rather underfitting due to a very slow convergence. In any case, Random- Attr is a very simple problem and requiring more than 10,000 eval- uations in order to obtain the same results than a diagonal matrix does not seem appropriate. In order to deal with this issue, we have proposed a second way of representing the transformation matrix. This second method was described in Section 3.2. Next section (Sec- tion 4.3) reports the experimental results). 5 http://archive.ics.uci.edu/ml/ Table 4 This table adds to Table 3 a new column for our classification rate (percentage) with KNN (k = 1) for the transformed data when scaling factors and rotation angles are directly evolved. Original data CMA-ES diagonal CMA-ES square CMA-ES rotation angles Straight-0 (S0) 9.20 ± 2.26 �99.70 ± 0.79 �99.85 ± 0.24 �98.90 ± 0.66 Straight-45 (S45) 9.20 ± 1.62 8.95 ± 1.89 �99.15 ± 0.58 �98.60 ± 0.32 RandomAtt (RA) 50.50 ± 2.38 �95.60 ± 1.81 �79.63 ± 6.28 �97.70 ± 0.46 Ripley (RI) 88.58 ± 0.36 88.91 ± 0.44 88.41 ± 0.57 88.33 ± 0.68 Blood (BL) 68.41 ± 0.57 67.92 ± 0.69 68.12 ± 1.01 67.31 ± 0.89 Car (CR) 87.49 ± 0.17 �95.82 ± 0.33 �97.52 ± 0.18 �96.23 ± 0.37 Diabetes (DB) 68.11 ± 1.02 67.10 ± 1.51 67.28 ± 0.95 67.95 ± 1.07 Ionosphere (IO) 86.19 ± 0.39 89.94 ± 1.57 88.31 ± 0.63 �91.51 ± 0.98 Iris (IR) 96.00 ± 0.00 94.80 ± 0.76 95.27 ± 1.24 94.93 ± 0.95 Wine (WI) 75.99 ± 1.47 �94.39 ± 1.14 �85.69 ± 1.64 �96.37 ± 0.80 Significantly better/worse 4/0 (of 10) 5/0 (of 10) 6/0 (of 10) 4.3. Experimental results. Rotation angles and scaling In this subsection we show the experimental results obtained by the second method described in Section 3.2 when scaling factors and rotation angles are evolved and KNN (with k = 1) is used as classifier in the transformed space. The parameters for CMA-ES are the same as in the previous subsection. We apply the method to the same data folds previously used when the first method was applied. Thus, the results can be compared with the ones dis- played in Table 3 and obtained when the transformation matrices were directly evolved. In order to allow a good comparative of both methods, Table 4 shows all the results: column 1 shows the classification rates when KNN is applied to the original data. Columns 2 and 3 show the re- sults corresponding to the first method explained in Section 4.2, when either diagonal and square matrices were directly evolved. Finally, column 4 displays the results obtained by the second method, when scaling factors and rotation angles are evolved. In all cases the results correspond to the mean classification accuracy rates obtained for all the domains and their standard deviations. As in the last subsection, the means have been obtained by averaging 10 executions of a 10-fold crossvalidation procedure. The best re- sults have been marked with a dagger. As before, a Corrected Re- peated 10 � 10-fold Crossvalidation T-Test (Bouckaert & Frank, 2004) has been used for testing if each method is significantly bet- ter (or worse) than KNN on the original data (a = 0.05). A dagger (�) marks significant differences. In the Straight-0 domain, as we said before, just compressing the x1 coordinate, a good classification rate can be achieved be- cause points belonging to the same class can get as close to each other as needed. This simple transformation can be done with a diagonal matrix, and therefore with a square matrix and a rota- tion-angles matrix too. The results show that KNN only obtains a classification rate of 9.2% on the original data set, but if the data is linearly transformed with a square matrix the rate is 99.8% and 98.9% if a rotation-angle matrix is used. In the Straight-45 domain, a rotation is needed. Table 3 shows the results as expected: KNN obtains a very bad classification accu- racy on the original data (9.2%) and when a diagonal matrix is used (8.95%). On the contrary, when a square matrix or rotation-angles matrix is used, the results are near to 100%. In the RandomAttr domain, scaling the irrelevant attributes should be enough to attain a much better accuracy. Therefore, diagonal matrices should be enough but given that square matrices include diagonal ones, square matrices should get similar results. But we seen in Table 3 that square matrices did not reach the ex- pected performance. On the other hand, rotation-angles matrices perform very well, and they obtain a rate of 97.70% vs. 79.63% for square matrices. In this domain we observe that encoding di- rectly the scaling factors and the rotation angles, facilitates the search of the appropriate transformation. Regarding the real data domains, we can see that the rotation- angles method also solves the stagnation problem for the Wine do- main: 96.37% (rotation angles) vs. 75.99% (square matrix) vs. 94.39% (diagonal matrix). Summarizing the results, it can be observed that the CMA-ES rotation-angles method behaves as expected in all domains, avoid- ing the situations where the first method (evolving directly the square matrices) got stagnated in some domains (RandomAttr and Wine). Rotation-angles is also the method that obtains the highest number of statistically significant differences with the ori- ginal data: 6 out of 10 vs. 5 out of 10 (square matrices) and 4 out of 10 (diagonal matrices). 5. Conclusions In this work, the evolutionary method CMA-ES is used to search for appropriate linear transformations of data that improve the success rate of KNN. Working with linear transformations has an important advantage: any learning algorithm with numerical attri- butes can be used, even if it is not based on distances. In a first stage, linear transformations have been represented as straightforward matrices. Both diagonal and full square matrices have been considered. The method has been tested with different domains, both synthetic and real, and the results show that, in gen- eral, both diagonal and square matrices found by CMA-ES either outperform or match the classifier on untransformed data. Square matrices have the advantage that they allow for more complex transformations that are required in some of the domains. How- ever, in certain domains diagonal matrices attain significatively better results than square ones. This is unexpected given that diag- onal matrices are a subset of square ones. We have tested that this is not due to overfitting but rather to a very slow convergence in some of the domains. In order to deal with this issue, we have modified the method using a different representation of transformation matrices, which is based on the Singular Value Decomposition theorem (SVD). According to SVD, any linear transformation can be split into rota- tions and scalings. Thus, in the second part of this work, the trans- formations are directly represented by rotation angles and scaling factors. By allowing the algorithm to separate both components, it is expected that CMA-ES will be able to find the proper scaling when only a scaling (or diagonal matrix) is needed. Experimental results show that the rotation-angles representation manages to avoid the stagnation problem of square matrices. In the future, we will test what other learning methods benefit from data transformations. Also, the rotation-angles method could 6 be used as a dimensionality reduction technique in a similar vein to Sierra and Echeverra (2006). Acknowledgment This article has been financed by the Spanish founded research MCINN project MSTAR::UC3M, Ref:TIN2008-06491-C04-03, and by project A3::UAM, Ref:TIN2007-66862-C02–02. References Arfken, G. (1985). Mathematical methods for physicists (third ed.). Orlando, FL: Academic Press. Atkenson, C. G., Moore, A. W., & Schaal, S. (1997). Locally weighted learning. Artificial Intelligence Review, 11, 11–73. Bck, T. (1996). Evolutionary algorithms in theory and practice. Oxford New York: Oxford University Press. Bouckaert, R. R., & Frank, E. (2004). Evaluating the replicability of significance tests for comparing learning algorithms. Advances in Knowledge Discovery and Data Mining (PAKDD), 3–12. Cover, T. M., & Hart, P. E. (1967). Nearest neighbor pattern classification. IEEE Transactions on Information Theory, 13(1), 21–27. Goldstein, H. (1980). Classical mechanics (second ed.). Reading, MA: Addison- Wesley. Hansen, Nikolaus (2009). The CMA evolution strategy: A tutorial. TU Berlin: Technische Universitat Berlin. Hansen, N., & Ostermeier, A. (2001). Completely derandomized self-adaptation in evolution strategies. Evolutionary Computation, 9(2), 159–195. Moody, J. E., & Darken, C. (1989). Fast learning in networks of locally tuned processing units. Neural Computation, 1, 281–294. Ostermeier, A., Gawelczyk, A., & Hansen, N. (1994). A derandomized approach to self-adaptation of evolution strategies. Evolutionary Computation, 4(2), 369–380. Ripley, B. D. (1996). Pattern recognition and neural networks. Cambridge: Cambridge University Press. Schwefel, Hans-Paul (1980). Subroutines EVOL, GRUP, KORR – Listings and user’s guides. Nuclear Research Centre (KFA) Jülich.. Sierra, A., & Echeverra, A. (2006). Evolutionary discriminant analysis. IEEE Transactions on Evolutionary Computation, 10(1), 81–92. Tou, J. T., & Gonzalez, R. C. (1974). Pattern recognition principles. Addison-Wesley. Valls, J. M., Aler, R., & Fernández, O. (2007). Evolving generalized euclidean distances for training RBNN. Computing and Informatics, 26, 33–43. Weisberg, S. (1985). Applied linear regression. New York: John Wiley and Sons. 7 
work_ygp5ozz6xjbshppuusoq74ur5q	----	Sign in Skip to main content Sign in » warwick.ac.uk You must sign in to view this page. Your version of My Warwick for iOS is outdated. Due to a bug affecting this version, you may be unable to login. Please download the latest update from the App Store. Username Password Keep me signed in Until I close my browser All day For three months Indefinitely until I sign out   Sign in Find my account Acceptable use policy Address https://warwick.ac.uk/fac/sci/dcs/research/pcav/publications/pubs/eswa-2006.pdf Service warwick.ac.uk © MMXXI Terms Privacy Cookies Accessibility 
work_yhmhpetxjjertmjquf7qfk2lqm	----	PII: S0957-4174(02)00042-8 An expert system for diagnosis of the heart valve diseases I. Turkoglu a,*, A. Arslan b , E. Ilkay c a Department of Electronics and Computer Science, Technical Education Faculty, Firat University, 23119 Elazig, Turkey b Department of Computer Engineering, Firat University, 23119 Elazig, Turkey c Department of Cardiology, Firat University, 23119 Elazig, Turkey Abstract In this paper, an expert diagnosis system is presented for interpretation of the Doppler signals of the heart valve diseases based on the pattern recognition. This paper especially deals with the feature extraction from measured Doppler signal waveforms at the heart valve using the Doppler Ultrasound. Wavelet transforms and short time Fourier transform methods are used to feature extract from the Doppler signals on the time – frequency domain. Wavelet entropy method is applied to these features. The back-propagation neural network is used to classify the extracted features. The performance of the developed system has been evaluated in 215 samples. The test results showed that this system was effective to detect Doppler heart sounds. The correct classification rate was about 94% for normal subjects and 95.9% for abnormal subjects. q 2002 Elsevier Science Ltd. All rights reserved. Keywords: Pattern recognition; Doppler heart sounds; Heart valves; Feature extraction; Wavelet decomposition; Spectrograms; Neural networks; Expert systems 1. Introduction Researches showed that the most of human deaths in the world are due to heart diseases. The heart valve disorders are of importance among the heart diseases. Among them, mitral and aortic valve disorders are the most common ones. For this reason, early detection of heart valve disorders is one of the most important medical research areas (Akay, Akay, & Welkowitz, 1992). Today, the used methods for diagnosis of heart valve disorders are non-invasive techniques (electrocardiograms, chest X-rays, heart sounds and murmur from stethoscope, ultrasound imaging and Doppler techniques) and invasive techniques (angiography, transozefagial echocardiograph (Nanda, 1993). However, each method is limited in its ability to offer efficient and thorough detection and characterization (Plett, 2000). All of these methods are based on experience and information of physician. The researches in this area are focused on improving human – machine interfaces in existing methods. In this way, the cardiologist can understand the output of the examination systems more easily and diagnose the problem more accurately (Philpot, Yoganathan, & Nanda, 1993). Doppler techniques are the most preferred because of their completely non-invasive and without risk in the serial studies. The technique has improved much since Satomura first demonstrated the application of the Doppler effect to the measurement of blood velocity in 1959 (Keeton & Schlindwein, 1997). In recent years, Doppler technique has found increasing use in the assessment of heart disease (Wright, Gough, Rakebrandt, Wahab, & Woodcock, 1997). Doppler heart sounds (DHS) are one of the most important sounds produced by blood flow, valves motion and vibration of the other cardiovascular components (Jing, Xuemin, Mingshi, & Wie, 1997). However, the factors such as calcified disease or obesity often results in a diagnostically unsatisfactory Doppler techniques assessment and, there- fore, it is sometimes necessary to assess the spectrogram of the Doppler shift signals to elucidate the degree of the disease (Wright et al., 1997). A major motivation in our work is to aid the diagnosis in such cases. Among Doppler techniques, the most ubiquitous and straightforward are waveform profile indices such as the pulsatility index (PI), Pourcelot or resistance index (RI) and A/B Systolic Diastolic ratio, which are highly correlated and led to highly erroneous diagnostic results (Izzetoglu, Erkmen, & Beksac, 1995). These indices rely on the peak systolic and end- diastolic velocities, with only the PI making use of the mean velocity over the cardiac cycle. More sophisticated methods have also been developed such as the Laplace transform and principal components analysis. However, none of the simple or more complex analytical techniques has yielded an acceptable diagnostic accuracy so as to be commonplace in 0957-4174/02/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 7 - 4 1 7 4 ( 0 2 ) 0 0 0 4 2 - 8 Expert Systems with Applications 23 (2002) 229–236 www.elsevier.com/locate/eswa * Corresponding author. Fax: þ90-424-2184674. E-mail address: iturkoglu@firat.edu.tr (I. Turkoglu). http://www.elsevier.com/locate/eswa the vascular clinic (Wright et al., 1997). In this study, the developed method is an expert diagnosis system and will cause more effective usage of the Doppler technique. Until now, many attempts have been undertaken to automatically classify Doppler signals using pattern recognition (Chan, Chan, Lam, Lui, & Poon, 1997; Guler & Kara, 1995). Nevertheless, the studies on the DHS are fairly limited. This study will introduce the technique that will aid clinical diagnosis, enable further research of heart valve disorders, and provide a novel expert system for recognition of heart valve disorders. This study uses the powerful mathematics of wavelet signal processing and entropy, short-time Fourier transform (STFT) to efficiently extract the features from pre-processed Doppler signals for the purpose of recognizing between abnormal and normal of the heart valve. An algorithm called the expert diagnostic system is developed which is approximately the advanced pattern recognition. The DHS can be obtained simply by placing the Doppler ultrasonic flow transducer over the chest of the patient. A disadvantage of the Doppler method is that it requires the constant attention of the doctor to detect subtle changes in the DHS (Chan et al., 1997). The presented method prevents subtle changes in the DHS from escaping physician’s eye by perceiving them, even if the physician does not pay a continuous attention. The realized study has the stages of decision and evaluation contrary to the existing diagnosis methods. Thus, the doctor can make a comparison between the diagnoses of developed method and the diagnoses of existing methods. If the results are different, the examin- ations can be repeated or performed more carefully. In this way, the physician can decide more realistic. The paper is organized as follows. In Section 2, we review some basic properties of the pattern recognition, the Doppler heart signals, wavelet decomposition, STFT, wavelet entropy and neural networks. A new expert diagnostic system is described in Section 3. This new method enables a large reduction of the Doppler signal data while retaining problem specific information which facili- tates an efficient pattern recognition process. The effective- ness of the proposed method for classification of Doppler signals in the diagnosis of heart valve diseases is demonstrated in Section 4. Finally Section 5 presents discussion and conclusion. 2. Preliminaries In this section, the theoretical foundations for the expert diagnosis system used in the presented study are given in the following subsections. 2.1. Pattern recognition Pattern recognition can be divided into a sequence of stages, starting with feature extraction from the occurring patterns, which is the conversion of patterns to features that are regarded as a condensed representation, ideally contain- ing all-important information. In the next stage, the feature selection step, a smaller number of meaningful features that best represents the given pattern without redundancy is identified. Finally, the classification is carried out, i.e. a specific pattern is assigned to a specific class according to the characteristic features selected for it. This general abstract model, which is shown in Fig. 1, allows a broad variety of different realizations and implementations. Applying this terminology to the medical diagnostic process, the patterns can be identified, for example, as particular, formalized symptoms, recorded signals, or a set of images of a patient. The classes obtained represent the variety of different possible diagnoses or diagnostic statements (Dickhous & Heinrich, 1996). The techniques applied to pattern recognition uses artificial intelligence approaches (Bishop, 1996). 2.2. DHS signals The audio DHS is obtained by simply placing the Doppler ultrasonic flow transducer over the chest of the patient (Chan et al.,, 1997). Fig. 2 shows a DHS signal from heart aortic valve. The DHS produced from echoes back- scattered by moving blood cells is generally in the range of 0.5 – 10 kHz (Saini, Nanda, & Maulik, 1993). DHS signal spectral estimation is now commonly used to evaluate blood flow parameters in order to diagnose cardiovascular diseases. Spectral estimation methods are particularly used in Doppler ultrasound cardiovascular disease detec- tion. Clinical diagnosis procedures generally include analysis of a graphical display and parameter measure- ments, produced by blood flow spectral evaluation. Fig. 1. The pattern recognition approach. Fig. 2. The waveform pattern of the Doppler heart sound. I. Turkoglu et al. / Expert Systems with Applications 23 (2002) 229–236230 Ultrasonic instrumentation typically employ Fourier based methods to obtain the blood flow spectra, and blood flow measurements (Madeira, Tokhi, & Ruano, 2000). A Doppler signal is not a simple signal. It includes random characteristics due to the random phases of scattering particles present in the sample volume. Other effects such as geometric broadening and spatially varying velocity also affect the signal (Karabetsos, Papaodysseus, & Kountsouris, 1998). The following Doppler equation: Df ¼ 2vf cos u c ð1Þ where v equals the velocity of the blood flow, f equals the frequency of the emitted ultrasonic signal, c equals the velocity of sound in tissue (approximately 1540 m/s), Df equals the measured Doppler frequency shift, and u equals the angle of incidence between the direction of blood flow and the direction of the emitted ultrasonic beam (Saini et al., 1993). 2.3. Wavelet decomposition Wavelet transforms are rapidly surfacing in fields as diverse as telecommunications and biology. Because of their suitability for analyzing non-stationary signals, they have become a powerful alternative to Fourier methods in many medical applications, where such signals abound (Akay, 1997; Keeton & Schlindwein, 1997; Liang & Nartimo, 1998). The main advantages of wavelets is that they have a varying window size, being wide for slow frequencies and narrow for the fast ones, thus leading to an optimal time – frequency resolution in all frequency ranges. Furthermore, owing to the fact that windows are adapted to the transients of each scale, wavelets lack of the requirement of stationarity (Quiroga, 1998). Wavelet decomposition uses the fact that it is possible to resolve high frequency components within a small time window, and only low frequencies components need large time windows. This is because a low frequency component completes a cycle in a large time interval whereas a high frequency component completes a cycle in a much shorter interval. Therefore, slow varying components can only be identified over long time intervals but fast varying components can be identified over short time intervals. Wavelet decomposition can be regarded as a continuous time wavelet decomposition sampled at different frequen- cies at every level or stage. The wavelet decomposition function at level m and time location tm can be expressed as Eq. (2): dmðtmÞ ¼ xðtÞCm t 2 tm 2m � � ð2Þ where Cm is the decomposition filter at frequency level m. The effect of the decomposition filter is scaled by the factor 2 m at stage m, but otherwise the shape is the same at all stages. The synthesis of the signal from its time – frequency coefficients given in Eq. (3) can be rewritten to express the composition of the signal x½n� from its wavelet coefficients. d½n� ¼ x½n�h½n�; c½n� ¼ x½n�g½n� ð3Þ where h½n� is the impulse response of the high pass filter and g½n� is the impulse response of the low pass filter (Devasahayam, 2000). Wavelet packet analysis is an extension of the discrete wavelet transform (DWT) (Burrus, Gopinath, & Guo, 1998) and it turns out that the DWT is only one of the many possible decompositions that could be performed on the signal, instead of just decomposing the low frequency component as well. It is therefore possible to subdivide the whole time – frequency plane into different time – frequency pieces. The advantage of wavelet packet analysis is that it is possible to combine the different levels of decomposition in order to achieve the optimum time – frequency represen- tation of the original (Keeton & Schlindwein, 1997). 2.4. Short-time Fourier transform STFT, also known as the time-dependent or the windowed Fourier transform, attempts to analyze non-stationary signals by dividing the whole signal into shorter data frames. In short, the STFT can be compactly represented by Eq. (4): XðkÞ ¼ XN21 n¼0 xðnÞvðn 2 n0Þexp 2 j2pnk N � � ð4Þ where vðn 2 n0Þ is a window function to suppress side lobes while minimizing the main lobe leakage. The output of successive STFTs can provide a time – frequency represen- tation of the signal. To accomplish this the signal is truncated into short data frames by multiplying it by a window so that the modified signal is zero outside the data frame. The frequency spectrums for the data frame is calculated using the fast Fourier transform. One of the limitations of STFT is that the time frame for analysis of the signal is fixed (Keeton & Schlindwein, 1997). 2.5. Wavelet entropy Entropy-based criteria describe information-related properties for an accurate representation of a given signal. Entropy is a common concept in many fields, mainly in signal processing (Coifman & Wickerhauser, 1992). A method for measuring the entropy appears as an ideal tool for quantifying the ordering of non-stationary signals. An ordered activity (i.e. a sinusoidal signal) is manifested as a narrow peak in the frequency domain, thus having low entropy. On the other hand, random activity has a wide band response in the frequency domain, reflected in a high entropy value (Quiroga, Roso, & Basar, 1999). The types of I. Turkoglu et al. / Expert Systems with Applications 23 (2002) 229–236 231 entropy computing are shannon, threshold, norm, log energy, and sure (Coifman & Wickerhauser, 1992). 2.6. Neural networks An artificial neural network (ANN) is a mathematical model consisting of a number of highly interconnected processing elements organized into layers, the geometry and functionality of which have been likened to that of the human brain. The ANN may be regarded as processing learning capabilities inasmuch as it has a natural propensity for storing experimental knowledge and making it available for later use. By virtue of its parallel distribution, an ANN is generally robust, tolerant of faults and noise, able to generalize well and capable of solving non-linear problems (Haykin, 1994). The DHS, be it diseased or healthy, may be regarded as an inherently non-linear system due to the absence of the property of frequency preservation as required by the definition of a linear system (Nichols & O’Rourke, 1990). Applications of ANNs in the medical field include EMG pattern identification (Asres, Dou, Zhou, Zhang, & Zhu, 1997), images of human breast disease (Allan & Kinsner, 2001) medical data mining (Brameier & Banzhaf, 2001), Brachytherapy cancer treatment optimiz- ation (Miller, Bews, & Kinsner, 2001), interpretation of heart sounds (Turkoglu & Arslan, 2001), EEG pattern identification (Saraoglu, Yumusak & Ferikoglu, 1999); however, to date neural network analysis of DHS is a relatively new approach. 3. Methodology Fig. 3 shows the expert diagnostic system we developed. It consists of three parts: (a) data acquisition and pre- processing, (b) feature extraction, (c) classification using neural network. 3.1. Data acquisition and pre-processing All the original audio DHS signals were acquired from the Acuson Sequoia 512 Model Doppler Ultrasound system in the Cardiology Department of the Firat Medical Center. DHS signals were sampled at 20 kHz for 5 s and signal to noise ratio of 0 dB by using a sound card which has 16-bit A/D conversion resolution and computer software prepared by us in the MATLAB (version 5.3) (The MathWorks Inc. Natick, MA, USA). The Doppler ultrasonic flow transducer used (Model 3V2c) was run on an operating mode of 2 MHz continuous wave. The Doppler signals of the heart valves were obtained by placing the transducer over the chest of the patient with the aid of ultrasonic image. The digitized data, which has 95 normal and 120 abnormal subjects, were stored on hard disk of the PC. The subject group consisted of 132 males and 83 females with the ages ranging from 15 to 80 years. The average age of the subjects was 48.77 years. Pre-processing to obtain the feature vector was performed on the digitized signal in the following order: i. Filtering: The reserved DHS signals were high-pass filtered to remove unwanted low-frequency com- ponents, because the DHS signals is generally in the range of 0.5 – 10 kHz. The filter is a digital FIR, which is a fiftieth-order filter with a cut-off frequency equal to 500 Hz and window type is the 51-point symmetric Hamming window. ii. White de-noising: White noise is a random signal that contains equal amounts of every possible frequency, i.e., its FFT has a flat spectrum (Devasahayam, 2000). The DHS signals were filtered by removing the white noise by using wavelet packet. The white de-noising procedure contains three steps (Bakhtazad, Palazoglu, & Romagnoli, 1999): 1. Decomposition: Computing the wavelet packet decomposition of the DHS signal at level 4 and using the Daubechies wavelet of order 4. 2. Detail coefficient thresholding: For each level from 1 to 4, soft thresholding is applied to the detail coefficients. 3. Reconstruction: Computing wavelet packet recon- struction based on the original approximation coefficients of level 4 and the modified detail coefficients of levels from 1 to 4. Fig. 3. The algorithm of the expert diagnostic system. I. Turkoglu et al. / Expert Systems with Applications 23 (2002) 229–236232 iii. Normalization: The DHS signals in this study were normalized using Eq. (5) so that the expected amplitude of the signal is not affected from the rib cage structure of the patient. DHSsignal ¼ DHSsignal lðDHSsignalÞmaxl ð5Þ 3.2. Feature extraction Feature extraction is the key to pattern recognition so that it is arguably the most important component of designing the expert diagnosis system based on pattern recognition since even the best classifier will perform poorly if the features are not chosen well. A feature extractor should reduce the pattern vector (i.e. the original waveform) to a lower dimension, which contains most of the useful information from the original vector. The DHS waveform patterns from heart valves are rich in detail and highly non- stationary. The goal of the feature extraction is to extract features from these patterns for reliable intelligent classi- fication. After the data pre-processing has been realized, three steps are proposed in this paper to extract the characteristics of these waveforms using MATLAB with the Wavelet Toolbox and the Signal Processing Toolbox: i. Wavelet decomposition: For wavelet decomposition of the DHS waveforms, the decomposition structure, reconstruction tree at level 12 as shown in Fig. 4 was used. Wavelet decomposition was applied to the DHS signal using the Daubechies-10 wavelet decomposition filters. Thus, obtaining two types of coefficients: one- approximation coefficients cA and twelve-detail coef- ficients c D. A representative example of the wavelet decomposition of the Doppler sound signal of the heart aortic valve was shown in Fig. 5. ii. Short-time Fourier transform: The STFT is the under- stood and the most robust of the various time – frequency representations. The STFT of waveforms of terminal nodes was computed using a Hanning window function of 25 000-points, the sections overlap of 12 500 points between scans and zero padding the sections if the length of the window exceeds 25 000-points. A representative example of the STFT spectrums of a terminal node waveform is shown in Fig. 6. iii. Wavelet entropy: We next calculated the norm entropy Fig. 4. The decomposition structure in level-12. Fig. 5. The terminal node waveforms of wavelet decomposition at 12 levels of the DHS signal. Fig. 6. The STFT spectrums of a terminal node waveform. I. Turkoglu et al. / Expert Systems with Applications 23 (2002) 229–236 233 as defined in Eq. (6) of waveforms of the STFT spectrums. EðsÞ ¼ X i lsil 3=2 ð6Þ where s is the STFT spectrum and (si) represents the i coefficients of s. The resultant entropy data, which were normalized with 1/50 000, were plotted in Fig. 7. The plot of the entropy data includes 91 features obtained from 13 terminal nodes where each one contains waveform of seven frequency spectrums per DHS signal. Thus, the feature vector was extracted by computing the wavelet entropy values for each DHS signal. 3.3. Classification using neural network The objective of classification is to demonstrate the effectiveness of the proposed feature extraction method from the DHS signals. For this purpose, the feature vectors were applied as the input to an ANN classifier. The classification by neural network was performed using MATLAB with the Neural Network Toolbox. The training parameters and the structure of the neural network used in this study are as listed in Table 1. These were selected for the best performance, after several different experiments, such as the number of hidden layers, the size of the hidden layers, value of the moment constant and learning rate, and type of the activation functions. Fig. 8 shows the ANN training performance. 4. Experimental classification results We performed experiments using 215 heart aortic and mitral valve Doppler studies taken from different indi- viduals. The data from a part of the DHS signal samples were used for training and another part in testing the ANN. In these experiments, 100% correct classification was obtained at the ANN training among the two signal classes. It clearly indicates the effectiveness and the reliability of the proposed approach for extracting features from DHS signals for the purpose of pattern recognition. The ANN testing results are given in Table 2. 5. Discussion and conclusion In this study, we developed an expert diagnostic system for the interpretation of the DHS signals using pattern recognition, and the diagnosis performance of this method was demonstrated on the heart aortic and mitral valves. The task of feature extraction was performed using the wavelet Fig. 7. The wavelet entropy of the DHS signal. Table 1 ANN architecture and training parameters ANN architecture The number of layers 3 The number of neuron on the layers Input: 100 Hidden: 50 Output: 2 The initial weights and biases The Nguyen-Widow method Activation functions Log-sigmoid ANN training parameters Learning rule Back-propagation Adaptive learning rate Initial: 0.0001 Increase: 1.05 Decrease: 0.7 Momentum constant 0.95 Sum-squared error 0.00001 Fig. 8. The ANN training performance. Table 2 Performance of the expert diagnostic system The heart aortic valve The heart mitral valve N AN N AN Total number of samples 31 40 19 33 Correct classification # 28 38 19 32 Incorrect classification # 3 2 – 1 The average recognition % 98.7 99.8 98.5 99.4 The highest recognition % 100 100 100 100 The lowest recognition % 52.1 84.2 72.3 94.2 N: normal, AN: abnormal. I. Turkoglu et al. / Expert Systems with Applications 23 (2002) 229–236234 decomposition for multi-scale analysis, STFT for time – frequency representations, and the wavelet entropy, while classification was carried out by the back-propagation neural network. The stated results show that the proposed method can make an efficient interpretation. Although the abnormal subjects were attained 95.9% correct classifi- cation, the normal subjects were provided 94% correct classification. The feature choice was motivated by a realization that wavelet decomposition is essentially a representation of a signal at a variety of resolutions. In brief, the wavelet decomposition has been demonstrated to be an effective tool for extracting information from the DHS signals. However, the proposed feature extraction method is robust against noise in the DHS signals. In this paper, the application of the wavelet entropy to the feature extraction from DHS signals was shown. Wavelet entropy proved to be a very useful tool for characterizing the DHS signal, furthermore the information obtained with the wavelet entropy probed not to be trivially related to the energy and consequently with the amplitude of signal. This means that with this method, new information can be accessed with an approach different from the traditional analysis of amplitude of DHS signal. The most important aspect of the expert diagnostic system is the ability of self-organization of the neural network without requirements of programming and the immediate response of a trained net during real-time applications. These features make the expert diagnostic system suitable for automatic classification in interpretation of the DHS signals. These results point out the ability of design of a new intelligent assistance diagnosis system. The diagnosis performances of this study shows the advantages of this system: it is rapid, easy to operate, non- invasive, and not expensive. This system is of the better clinical application over others, especially for earlier survey of population. However, the position of the ultrasound probe, which is used for data acquisition from the heart valves, must be taken into consideration by physician. Although our expert diagnosis system was carried out on the heart aortic and mitral valves, similar results for the other valves (tricuspid and pulmonary) and the other Doppler studies can be expected. Besides the feasibility of a real-time implementation of the expert diagnosis system, by increasing the variety and number of DHS signals additional information (i.e. quantification of the heart valve regurgitation and stenosis) can be provided for diagnosis. Acknowledgments We want to thank, the Cardiology Department of the Firat Medicine Center, Elazig, Turkey for providing the DHS signals to us. This work was supported by Firat University Research Fund. (Project No: 527). References Akay, M. (1997). Wavelet applications in medicine. IEEE Spectrum, 34, 50 – 56. Akay, M., Akay, Y. M., & Welkowitz, W. (1992). Neural networks for the diagnosis of coronary artery disease. International Joint Conference on Neural Networks, IJCNN, (Vol. 2, pp. 419 – 424). Allan, R., & Kinsner, W. (2001). A study of microscopic images of human breast disease using competitive neural networks. Canadian Conference on Electrical and Computer Engineering (Vol. 1, pp. 289 – 293). Asres, A., Dou, H., Zhou, Z., Zhang, Y., & Zhu, S. (1997). A combination of AR and neural network technique for EMG pattern identification. Engineering in Medicine and Biology Society, Bridging Disciplines for Biomedicine, Proceedings of the 18th Annual International Conference of the IEEE (Vol. 4, pp. 1464 – 1465). Bakhtazad, A., Palazoglu, A., & Romagnoli, J. A. (1999). Process data de-noising using wavelet transform. Intelligent Data Analysis, 3, 267 – 285. Bishop, C. M. (1996). Neural networks for pattern recognition. Oxford: Clarendon Press. Brameier, M., & Banzhaf, W. (2001). A comparison of linear genetic programming and neural networks in medical data mining. IEEE Transactions on Evolutionary Computation, 5, 17 – 26. Burrus, C. S., Gopinath, R. A., & Guo, H. (1998). Introduction to wavelet and wavelet transforms. NJ, USA: Prentice Hall. Chan, B. C. B., Chan, F. H. Y., Lam, F. K., Lui, P. W., & Poon, P. W. F. (1997). Fast detection of venous air embolism is Doppler heart sound using the wavelet transform. IEEE Transactions on Biomedical Engineering, 44(4), 237 – 245. Coifman, R. R., & Wickerhauser, M. V. (1992). Entropy-based algorithms for best basis selection. IEEE Transactions on Information Theory, 38(2), 713 – 718. Devasahayam, S. R. (2000). Signals and systems in biomedical engineer- ing. Dordrecht: Kluwer Academic Publishers. Dickhous, H., & Heinrich, H. (1996). Classifying biosignals with wavelet networks. IEEE Engineering in Medicine and Biology, 103 – 111. Guler, I., & Kara, S. (1995). Detection of mitral stenosis by a pulsed Doppler flow meter and autoregressive spectral analysis method. Proceedings of the 14th Conference of the Biomedical Engineering Society of India. An International Meeting, Proceedings of the First Regional Conference, IEEE (pp. 2/91 – 2/92). Haykin, S. (1994). Neural networks, a comprehensive foundation. New York: Macmillan College Publishing Company Inc. Izzetoglu, K., Erkmen, A. M., & Beksac, S. (1995). Intelligent classification of fetal Doppler blood velocity waveform abnormalities using wavelet transform and vector quantization algorithm. IEEE International Conference on Systems, Man and Cybernetics, Intelligent Systems for the 21st Century (Vol. 1, pp. 724 – 729). Jing, F., Xuemin, W., Mingshi, W., & Wie, L. (1997). Noninvasive acoustical analysis system of coronary heart disease. Biomedical Engineering Conference, Proceedings of the 1997 Sixteenth Southern (pp. 239 – 241). Karabetsos, E., Papaodysseus, C., & Koutsouris, D. (1998). Design and development of a new ultrasonic doppler technique for estimation of the aggregation of red blood cells. Measurement, 24, 207 – 215. Keeton, P. I. J., & Schlindwein, F. S. (1997). Application of wavelets in Doppler ultrasound. Measurement, 17(1), 38 – 45. Liang, H., & Nartimo, I. (1998). A feature extraction algorithm based on wavelet packet decomposition for heart sound signals. Proceedings of the IEEE-SP International Symposium (pp. 93 – 96). Madeira, M. M., Tokhi, M. O., & Ruano, M. G. (2000). Real-time implementation of a Doppler signal spectral estimator using sequential and parallel processing techniques. Microprocessors and Microsystems, 24, 153 – 167. Miller, S., Bews, J., & Kinsner, W. (2001). Brachytherapy cancer treatment I. Turkoglu et al. / Expert Systems with Applications 23 (2002) 229–236 235 optimization using simulated annealing and artificial neural networks. Canadian Conference on Electrical and Computer Engineering (Vol. 1, pp. 649 – 654). Nanda, N. C. (1993). Doppler echocardiography (2nd ed.). London: Lea & Febiger. Nichols, W. W., & O’Rourke, M. F. (1990). McDonald’s blood flow in arteries: Theoretical, experimental and clinical principles (3rd ed.), London. Philpot, E. F., Yoganathan, A. P., & Nanda, N. C. (1993). Future directions in Doppler echocardiography. Doppler echocardiography. Philadelphia, London: Lea & Febiger. Plett, M. I. (2000). Ultrasonic arterial vibrometry with wavelet based detection and estimation. PhD Thesis (pp. 17 – 18), University of Washington. Quiroga, R. Q. (1998). Quantitative analysis of EEG signals: Time – frequency methods and Chaos theory. Lübeck: Intitute of Physiology, Medical University. Quiroga, R. Q., Roso, O. A., & Basar, E. (1999). Wavelet entropy: A measure of order in evoked potentials (Vol. 49). Evoked potentials and magnetic fields, Amsterdam: Elsevier, pp. 298 – 302. Saini, V. D., Nanda, N. C., & Maulik, D. (1993). Basic principles of ultrasound and Doppler effect. Doppler echocardiography. Philadelphia, London: Lea & Febiger. Saraoglu, H. M., Yumusak, N., & Ferikoglu, A. (1999). Training of brain signals by using neural networks. International symposium on Mathematical and Computational Applications, September 1 – 3, 1999, 249 – 254. Bakü, Azerbaijan. Turkoglu, I., & Arslan, A. (2001). An intelligent pattern recognition system based on neural network and wavelet decomposition for interpretation of heart sounds. Proceedings of the 23rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society, October 25 – 28, 2001 (pp. 4.2.7 – 3). Istanbul, Turkey. Wright, I. A., Gough, N. A. J., Rakebrandt, F., Wahab, M., & Woodcock, J. P. (1997). Neural network analysis of Doppler ultrasound blood flow signals: A pilot study. Ultrasound in Medicine and Biology, 23(5), 683 – 690. I. Turkoglu et al. / Expert Systems with Applications 23 (2002) 229–236236 An expert system for diagnosis of the heart valve diseases Introduction Preliminaries Pattern recognition DHS signals Wavelet decomposition Short-time Fourier transform Wavelet entropy Neural networks Methodology Data acquisition and pre-processing Feature extraction Classification using neural network Experimental classification results Discussion and conclusion Acknowledgments References 
work_yhzgueixzzcctpoq2kvu2tead4	----	IJMPERD - AN EXPERT SYSTEM TO CLASSIFYING AND SCHEDULING OF SERVICE SYSTEMS - jirodriguezmolano@gmail.com www.tjprc.org editor@tjprc.org AN EXPERT SYSTEM TO CLASSIFYING AND SCHEDULING OF SERVICE SYSTEMS EDUYN LÓPEZ-SANTANA, GERMAN MÉNDEZ-GIRALDO & JOSE IGNACIO RODRIGUEZ Industrial engineering. Faculty of Engineering. Universidad Distrital Francisco José de Caldas. Bogotá, Colombia ABSTRACT This paper studies the problem of classify scheduling problems in service systems, its performance measures and solution techniques. First, we contribute to classification of service systems that is a difficult task since the service is described in a confused and ambiguous language. This confusion makes it difficult to analyze, mainly in the making decisions at operative level. We propose a new notation for service systems that consist in three fields: customer, resources, and flow control. From this notation, we propose an integrated three rule-based systems. The first identify the type of scheduling problem according with scheduling, routing, and routing-scheduling. With the results a second system identifies the performance measures. And finally, a third system determines the best solution techniques to solve the problem. We develop a computational tool in Java called SchES (Scheduling with Expert Systems) that implements our approach. KEYWORDS: Knowledge Management; Scheduling; Service System; Expert System Received: Oct 08, 2020; Accepted: Oct 28, 2020; Published: Nov 12, 2020; Paper Id.: IJMPERDOCT202045 INTRODUCTION Service systems consist of service providers (resources) and service customers working together to coproduce value (flow control) in complex value networks where providers and customers might be individuals, firms, government agencies, or any organization of people and technologies (E. López-Santana, Méndez-Giraldo, & Figueroa-García, 2019; E. R. López-Santana & Méndez-Giraldo, 2016b; Spohrer et al., 2007). Service systems are becoming a strategic area of scientific research from multidisciplinary approaches, such as Service Science or Service Science Management and Engineering (Demirkan, Spohrer, & Krishna, 2011; Spohrer et al., 2007). In service systems, as the manufacturing systems, there are a decision-making process called scheduling related with the allocation of resources to perform a set of tasks in a specific planning horizon subject several operational constraints such capacity or unavailability of resources, due dates, priorities, cancelations, among others to optimize one or more objectives. This process has applications in computer science, logistic, distribution, production, workforce, maintenance, health systems, among others (E. López-Santana et al., 2019; E. R. López- Santana & Méndez-Giraldo, 2016a). The general scheduling problem consists in finding the best sequence to perform a number of jobsin a number of machines (or resources) in order to optimize a one or more objective functions (Pinedo, 2016). This problem is typically NP-hard and thus is difficult to find an optimal solution since the computation time increases exponentially with the problem size (Pinedo, 2016). The scheduling problems can be classified in deterministic and stochastic problems. In deterministic case, the information about processing times, release date, due dates, O rigin al A rticle International Journal of Mechanical and Production Engineering Research and Development (IJMPERD) ISSN (P): 2249–6890; ISSN (E): 2249–8001 Vol. 10, Issue 5, Oct 2020, 447–454 © TJPRC Pvt. Ltd 448 Eduyn López-Santana, German Méndez-Giraldo & Jose Ignacio Rodriguez Impact Factor (JCC): 9.6246 NAAS Rating: 3.11 capacity, among others has a constant value while in stochastic case the information has a stochastic variability (Pinedo, 2016). The objective of this paper consists inreviewing the problem of scheduling in service systems, where the traditional tools could not be applied directly give the complexity of services. The scheduling in service systems is a decision-making process related with the allocation of resources to perform a set of tasks in a specific planning horizon subject several operational constraints such capacity or unavailability of resources, due dates, priorities, cancelations, among others to optimize one or more objectives (E. R. López-Santana & Méndez-Giraldo, 2016a). The authors review the scheduling problems in service systems and present a structure of a Knowledge Based System (KBS) to solve it. The proposed KBS by (E. R. López-Santana & Méndez-Giraldo, 2016a) is supported in two systems. The first one is a service systems classification component (SSCC) applied to International Standard of Industrial Classification (ISIC) system. This result is useful for companies to determine their classification and for entities that consolidate information for statistical purposes. The second one is a task scheduling component (TSC) that allows identifying the scheduling problem (or problems). After, it selects the best solution technique (or techniques) and setting the input data, parameters, outputs, and performance measure to solve the problem. The remainder of this document is organized as follows. Section 2 presents the proposed notation for service systems. Section 3 shows rule-based system to identify the scheduling problem and suggest a solution technique from the proposed notation. Section 4 presentssome conclusions and future lines of research. PROPOSED NOTATION FOR A SERVICE SYSTEM We propose a notation to represent a service system to identify the scheduling problem and suggest a solution technique. In the literature, there are several notations for scheduling problems. The most popular is the Graham notation provided by (Graham, Lawler, Lenstra, & Kan, 1979) that consists in a three fields �|�|�. � represents the production system’s setting, e.g., 1 if is one machine, F if is a Flow shop problem, J if is a Job shop problem, among others. � is a field where the constraints and characteristics about the jobs and the machines are described, e.g., �� indicates that the jobs have release dates, prmp shows preemptions for the jobs, among others. The last field � represents the performance measure for the scheduling problem, e.g., �� represent the makespan, ∑���� is the total weighted completion times, among others. This notation applies for several scheduling problems mainly in manufacturing environments, however, for the service systems, it lacks the parameter to add characteristics for these systems. For service systems, the notation used frequently corresponds to the queueing systems that could be characterized in terms of Kendall’s notation (Kendall, 1953), whose encoding under the following structure: 1/2/3/4. Where 1 refers to the arrival process that can be Poisson (M), Deterministic (D) or general distribution different to Poisson (G); 2 is the service process that can be also M, D o G; 3 represents the number of servers by stage of process in the network, which can be single (represented by 1) or multiple (represented by ); and 4 states the system’s capacity, infinite when it is empty or a K to indicate the queue’s length. We propose a notation that consists in three fields �/�/�, where � is set of parameters related with the customer, � is the set of parameters related with the resources, and� is the set of parameters associated with the flow control. Figure 1 presents our proposed notation over the good-dominant and service-dominant logic views prosed by (Vargo & Lusch, 2008). In pure manufacturing systems the production and consumption systems are separated, while in pure service system An Expert System to Classifying and Scheduling of Service Systems 449 www.tjprc.org editor@tjprc.org both production and consumption systems are related by the interactions and coproduction. Our notation represents this behavior as Flow Control, Figure 2 shows the fields and its groups and variables with the defined outputs for each one. Figure 1 also present some examples of systems for the variable Type of need with its outputs Intangible and Tangible to understand the differences between manufacturing and service systems. Figure 1: Representation of the proposed notation for manufacturing and service systems Figure2: Proposed notation for a service system 450 Impact Factor (JCC): 9.6246 PROPOSED RULE-BASED SYSTEM López-Santana et. al. (E. R. López-Santana & Méndez based expert system to solve the scheduling problem in service systems shows the framework with two inputs, a service system classification component and the task scheduling component (SSCC) described in (E. R. López-Santana & Méndez here. The SSCC consists in a classification system for the International the Colombian context, also to classify according to economy sector as Services, Manufacturing and Primary proposed by the authors for the section level and division level. Additionally, it is taken as i total has 249 instances for the selection of rules. The knowledge base consists in a set of 12 attributes: Coproduction, Dependence, Accumulation, Technology, Property, Simultaneity, Flexibility, Nature, Durabil made with an algorithm to generate tree C4.5 using the entropy information measure. The attributes were modeled using a knowledge acquisition model based in a non approach (for details see (E. R. López-Santana & Méndez Figure 3: General framework of a KBES to scheduling in service systems (source: Figure 4 shows the TSC proposed. The TSC consists in a classification system for according with the scheduling problems in services systems. Figure 5 shows the methodology of three phases general framework of Figure 4. The phase 1 is the knowledge acquisition in order obtain the knowledge base. The knowledge base consists in to determine the set of variables for each field the notation presented in se set of attributes and characteristics determined by variables are not related with the attributes or characteristics, then a survey are performed to determine all variables. The phase 2 uses as inference engine four methods to get the rules: J48 classifier, PART, DecisionTable and DecisionStump. These methods were built in WEKA and the compared with the percentage of correctly classified instances. We have a set of 498 instances to classify with our method knowledge base, the inference engine works in waterfal Eduyn López-Santana, German Méndez-Giraldo & Jose Ignacio Rodriguez BASED SYSTEM Santana & Méndez-Giraldo, 2016a) presents a general framework of an knowledge system to solve the scheduling problem in service systems(Lopez-Santana, Castro, & Giraldo, 2016) the framework with two inputs, a service system classification component and the task scheduling component Santana & Méndez-Giraldo, 2016b) and the task scheduling component (TSC) described here. The SSCC consists in a classification system for the International Standard Industrial Classification (ISIC) system in the Colombian context, also to classify according to economy sector as Services, Manufacturing and Primary proposed by the authors for the section level and division level. Additionally, it is taken as instances for database total has 249 instances for the selection of rules. The knowledge base consists in a set of 12 attributes: Coproduction, Dependence, Accumulation, Technology, Property, Simultaneity, Flexibility, Nature, Durability, Place, Scheduled and Standardization. The inference engine was made with an algorithm to generate tree C4.5 using the entropy information measure. The attributes were modeled using a knowledge acquisition model based in a non-linear optimization and an adaptive neuro-fuzzy inference system (ANFIS) Santana & Méndez-Giraldo, 2016b)). General framework of a KBES to scheduling in service systems (source: (E. R. López Giraldo, 2016a)) 4 shows the TSC proposed. The TSC consists in a classification system for according with the scheduling 5 shows the methodology of three phases to build the rule 4. The phase 1 is the knowledge acquisition in order obtain the knowledge base. The knowledge base consists in to determine the set of variables for each field the notation presented in se set of attributes and characteristics determined by (E. R. López-Santana & Méndez-Giraldo, 2016b) variables are not related with the attributes or characteristics, then a survey are performed to determine all variables. The phase 2 uses as inference engine four methods to get the rules: J48 classifier, PART, DecisionTable and Stump. These methods were built in WEKA and the compared with the percentage of correctly classified instances. We have a set of 498 instances to classify with our method (E. López-Santana, 2018) knowledge base, the inference engine works in waterfall way with three rule-based system (see Fig Giraldo & Jose Ignacio Rodriguez NAAS Rating: 3.11 presents a general framework of an knowledge Santana, Castro, & Giraldo, 2016). Figure 3 the framework with two inputs, a service system classification component and the task scheduling component and the task scheduling component (TSC) described Standard Industrial Classification (ISIC) system in the Colombian context, also to classify according to economy sector as Services, Manufacturing and Primary proposed by nstances for database-level Group, i.e. in The knowledge base consists in a set of 12 attributes: Coproduction, Dependence, Accumulation, Technology, ity, Place, Scheduled and Standardization. The inference engine was made with an algorithm to generate tree C4.5 using the entropy information measure. The attributes were modeled using a fuzzy inference system (ANFIS) (E. R. López-Santana & Méndez- 4 shows the TSC proposed. The TSC consists in a classification system for according with the scheduling to build the rule-based system from the 4. The phase 1 is the knowledge acquisition in order obtain the knowledge base. The knowledge base consists in to determine the set of variables for each field the notation presented in section 2. We use the Giraldo, 2016b). However, some variables are not related with the attributes or characteristics, then a survey are performed to determine all variables. The phase 2 uses as inference engine four methods to get the rules: J48 classifier, PART, DecisionTable and Stump. These methods were built in WEKA and the compared with the percentage of correctly classified Santana, 2018). With the results of based system (see Figure 4). Firstly, we begin An Expert System to Classifying and Scheduling of Service Systems 451 www.tjprc.org editor@tjprc.org with the identification of the scheduling problem for each field of the proposed notation. The output consists in a scheduling problem, a routing problem, or a scheduling & routing problem. It is possible to obtain the equal problem for all fields or different problems for each field because given the complexity of service systems and according with the notation defined in figures 1 and 2, the customer represents the consumption system, while the resources represents the production system, thus a different scheduling problem could be arise for these subsystems according with (E. López-Santana, 2018) and hence the problem, performance measure and solution technique could be hybrid. However, the analyst of the system could be chosen the same or different according with his/her expertise. Figure 4: General structure of the proposed rule-based system to scheduling for service systems. Figure 5: Methodology to build a rule-based system. Once the problem is determined, we apply the second inference engine to determine the performance measure. For the example, we consider as performance measures a general classification: efficiency, effectiveness, both,and none. Again, the methods are applied for each field and it is possible to get the same or different results for all fields. Finally, with the results of the problem and the performance measure, the third inference engine is applied to determine the solution technique to solve the scheduling problem for each field. For the example, we use four categories of solution techniques: optimization, stochastic models, heuristics, and metaheuristics. Again, the methods are applied for each field and it is 452 Eduyn López-Santana, German Méndez-Giraldo & Jose Ignacio Rodriguez Impact Factor (JCC): 9.6246 NAAS Rating: 3.11 possible to get the same or different results for all fields. We develop a computational tool in Java called SchES (Scheduling with Expert Systems) that implements the sub- systems described above in an independently way or a together way. Figure 6 presents the designed interfaces for knowledge acquisition and the inference engines. These ones were designed in Java and using Weka We apply our proposed method to a problem proposed in (E. López-Santana, Akhavan-Tabatabaei, Dieulle, Labadie, & Medaglia, 2016) in order to validate the solution method. The problem consists of planning and scheduling maintenance operations for a set of geographically distributed machines, subject to non-deterministic failures with a set of technicians that perform preventive maintenance and repair operations on the machines at the customer sites within a specific time window. (a) (b) Figure 6: Interfaces of SchES of (a) knowledge acquisition and (b) scheduling problem classification CONCLUSIONS This paper reviews the scheduling problems in service systems and presents a structure of a rule-based system to identify An Expert System to Classifying and Scheduling of Service Systems 453 www.tjprc.org editor@tjprc.org the type of scheduling problem, the performance measure, and the solution technique. We propose a new notation that involves the characteristics of service system in three fields: customers, resources, and flow control. Our notation states a set of variables, which are used as input of the proposed rule-based systems to identify the scheduling problems between a set of options defined as scheduling, routing, or scheduling and routing. Our proposed contribute to the develop of decision support system in the field of give a framework to classify scheduling problems according with a set of characteristics and attributes that involves the major of real features of the service systems. Our proposed, apply several classifiers of the literature, and it is possible to choose other to improve the results.This work generates possible future development lines, one of which is the validation of results with a set of real companies and improving the database of information. Another possible line consists to explore other methods of numerical ratings that involves imprecision, for instance using fuzzy logic. ACKNOWLEDGEMENTS This work was supported in part by the Research center “Centro de Investigaciones y Desarrollo Científico” at Universidad Distrital Francisco José de Caldas (Colombia) under Grant 2-602-468-14. REFERENCES 1. Demirkan, H., Spohrer, J. C., & Krishna, V. (2011). Introduction of the Science of Service Systems. In H. Demirkan, J. C. Spohrer, & V. Krishna (Eds.), The Science of Service Systems (pp. 1–11). Boston, MA: Springer US. https://doi.org/10.1007/978-1-4419-8270-4_1 2. Graham, R. L., Lawler, E. L., Lenstra, J. K., & Kan, A. H. G. R. (1979). Optimization and Approximation in Deterministic Sequencing and Scheduling: a Survey. Annals of Discrete Mathematics, 5(C), 287–326. https://doi.org/10.1016/S0167- 5060(08)70356-X 3. Kendall, D. G. (1953). Stochastic Processes Occurring in the Theory of Queues and their Analysis by the Method of the Imbedded Markov Chain. The Annals of Mathematical Statistics, 24(3), 338–354. https://doi.org/10.1214/aoms/1177728975 4. López-Santana, E. (2018). Review of scheduling problems in service systems. Bogotá, Colombia. 5. López-Santana, E., Akhavan-Tabatabaei, R., Dieulle, L., Labadie, N., & Medaglia, A. L. (2016). On the combined maintenance and routing optimization problem. Reliability Engineering & System Safety, 145, 199–214. https://doi.org/10.1016/j.ress.2015.09.016 6. López-Santana, E., Méndez-Giraldo, G., & Figueroa-García, J. C. (2019). Scheduling in queueing systems and networks using ANFIS. In Studies in Fuzziness and Soft Computing (Vol. 377, pp. 349–372). Springer Verlag. https://doi.org/10.1007/978-3- 030-10463-4_18 7. Lopez-Santana, E. R., Castro, S. J. B., & Giraldo, G. A. M. (2016). Modelo metodológico para programación de tareas en sistemas de servicios: un enfoque de ingeniería de software. Redes de Ingeniería, 7(1), 55–66. https://doi.org/10.14483/UDISTRITAL.JOUR.REDES.2016.1.A07 8. López-Santana, E. R., & Méndez-Giraldo, G. A. (2016a). A Knowledge-Based Expert System for Scheduling in Services Systems. In J. C. Figueroa-García, E. R. López-Santana, & R. Ferro-Escobar (Eds.), Applied Computer Sciences in Engineering WEA 2016 (Vol. 657, pp. 212–224). Springer International Publishing AG. https://doi.org/10.1007/978-3-319- 50880-1_19 454 Eduyn López-Santana, German Méndez-Giraldo & Jose Ignacio Rodriguez Impact Factor (JCC): 9.6246 NAAS Rating: 3.11 9. López-Santana, E. R., & Méndez-Giraldo, G. A. (2016b). A Non-linear Optimization Model and ANFIS-Based Approach to Knowledge Acquisition to Classify Service Systems. In D.-S. Huang, V. Bevilacqua, & P. Premaratne (Eds.), Intelligent Computing Methodologies (Vol. 9773, pp. 789–801). Springer International Publishing. https://doi.org/10.1007/978-3-319- 42297-8_73 10. Pinedo, M. L. (2016). Scheduling: Theory, Algorithms, and Systems. Cham: Springer International Publishing. https://doi.org/10.1007/978-3-319-26580-3 11. Spohrer, J., Maglio, P. P., Bailey, J., Gruhl, D., Jim Spohrer, Paul P. Maglio, … Daniel Gruhl. (2007). Steps Toward a Science of Service Systems. In The 2007 IEEE International Conference on Information Reuse and (Vol. 40, pp. 71–77). IEEE Computer Society. https://doi.org/10.1109/MC.2007.33 12. Vargo, S. L., & Lusch, R. F. (2008). Service-dominant logic: Continuing the evolution. Journal of the Academy of Marketing Science, 36(1), 1–10. https://doi.org/10.1007/s11747-007-0069-6 
work_yi53c4vi4vh7zm74av6hnaqm6u	----	Hybrid Reasoning Model for Strengthening the problem solving capability of Expert Systems (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 10, 2013 88 | P a g e www.ijacsa.thesai.org Hybrid Reasoning Model for Strengthening the problem solving capability of Expert Systems Kapil Khandelwal PHD Scholar, Computer Science Department Suresh Gyan Vihar University Jaipur, India Durga Prasad Sharma Professor, AMUIT MOEFDRE under UN Development Program, Ethiopia Jaipur, India Abstract—In this paper, we briefly outlined popular case- based reasoning combinations. More specifically, we focus on combinations of case-based reasoning with rule based reasoning, and model based reasoning. Further we examined the strengths and weaknesses of various reasoning models, case-based reasoning, rule-based reasoning and model-based reasoning, and discuss how they can be combined to form a more robust and better-performing hybrid. In a decision support system to address the variety of tasks a user performs, a single type of knowledge and reasoning method is often not sufficient. It is often necessary to determine which reasoning method would be the most appropriate for each task, and a combination of different methods has often shown the best results. In this study CBR was mixed with other RBR and MBR approaches to promote synergies and benefits beyond those achievable using CBR or other individual reasoning approaches alone. Each approach has advantages and disadvantages, which are proved to be complementary in a large degree. So, it is well-justified to combine these to produce effective hybrid approaches, surpassing the disadvantages of each component method. “KNAPS-CR” model integrates problem solving with learning from experience within an extensive model of different knowledge types. “KNAPS-CR” has a reasoning strategy which first attempts case- based reasoning, then rule-based reasoning, and, finally, model- based reasoning. It learns from each problem solving session by updating its collection of cases, irrespective of which reasoning method that succeeded in solving the problem. Keywords—knowledge based systems; KBS, sustained learning; problem solving; hybrid reasoning models; case based reasoning; CBR; model based reasoning; MBR; rule based reasoning;RBR I. INTRODUCTION Hybrid systems are universally better than conventional approaches. The combination of (two or more) different problem solving and knowledge representation methods is a very active research area in Artificial Intelligence. Hybrid Intelligent System is a combination of two techniques with more strength and less weakness. Almost every conceivable problem has been approached using some form of hybrid system. The aim is to create combined formalisms that benefit from each of their components. The effectiveness of various hybrid or integrated approaches has been demonstrated in a number of application areas. It is generally believed that complex problems are easier to solve with hybrid or integrated approaches. Model-based reasoning (MBR) is an approach in which general knowledge is represented by formalizing the mathematical or physical relationships present in a problem domain. The CBR-MBR integration improves solution accuracy over that which is possible using either single approach. TABLE I. HYBRID REASONING MODELS Domain Tools Reasoning Models Agriculture HIDES CBR,RBR Aircraft design AIDA CBR,RBR Aircraft Fleet Maintenance IDS CBR,RBR Architecture FABEL CBR,MBR, RBR Bioprocess recipes SOPHIST CBR,MBR Construction ScheduleCoach CBR,RBR Equipment Failure Analysis EFAES CBR,RBR Entomology CARMA CBR,MBR Finance ECLAS CBR,RBR MARS Law IKBALS CBR,RBR SHYSTER-MYCIN CABARE GREBE DANIEL Life Insurance CCAR CBR,RBR Medicine/Medical ICU CBR,RBR AUGUSTE Project WHAT CARE-PARTNER CASEY CBR,MBR T-IDDM CBR,MBR, RBR Menu planning CAMPER CBR,RBR Music GYMEL CBR,RBR SAXEX Plastic colorants FORMTOOL CBR,MBR Personnel Performance Evaluation MCRS CBR,RBR Real-Time Marine Environment Monitoring CORMS AI CBR,RBR Speech ANAPRON CBR,RBR Ultrasonic Rail Inspection URS-CBR CBR,RBR Rules usually represent general knowledge, whereas cases encompass knowledge accumulated from specific (specialized) situations. Rule-based and case-based reasoning are two popular approaches used in intelligent systems. Each approach (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 10, 2013 89 | P a g e www.ijacsa.thesai.org has advantages and disadvantages, which are proved to be complementary in a large degree. So, it is well-justified to combine rules and cases to produce effective hybrid approaches, surpassing the disadvantages of each component method [3] [9][10][11][12][13][15][17][18]. II. ADVANTAGES & DISADVANTAGES OF CBR, RBR, MBR & HYBRID REASONING A. Rule-based Reasoning The advantages of a rule-based approach include: 1) The ability to use, in a very direct fashion, experiential knowledge acquired from human experts. This is particularly important in domains that rely heavily on heuristics to manage complexity and/or missing information. 2) Rules map into state space search. Explanation facilities support debugging. 3) The separation of knowledge from control simplifies development of expert systems by enabling an iterative development process where the knowledge engineer acquires, implements, and tests individual rules. 4) Good performance is possible in limited domains. Because of the large amounts of knowledge required for intelligent problem solving, expert systems are limited to narrow domains. However, there are many domains where design of an appropriate system has proven extremely useful. 5) Good explanation facilities. Although the basic rule- based framework supports flexible, problem-specific explanations, it must be mentioned that the ultimate quality of these explanations depends upon the structure and content of the rules. Explanation facilities differ widely between data- and goal- driven systems. Disadvantages of rule-based reasoning include: 1) Often the rules obtained from human experts are highly heuristic in nature, and do not capture functional or model- based knowledge of the domain. 2) Heuristic rules tend to be “brittle” and can have difficulty handling missing information or unexpected data values. 3) Another aspect of the brittleness of rules is a tendency to degrade rapidly near the “edges” of the domain knowledge. Unlike humans, rule-based systems are usually unable to fall back on first principles of reasoning when confronted with novel problems. 4) Explanations function at the descriptive level only, omitting theoretical explanations. This follows from the fact that heuristic rules gain much of their power by directly associating problem symptoms with solutions, without requiring (or supporting) deeper reasoning. 5) The knowledge tends to be very task dependent. Formalized domain knowledge tends to be very specific in its applicability. Currently, knowledge representation languages do not approach the flexibility of human reasoning [4][9][11]. B. Case-based Reasoning The advantages of case-based reasoning include: 1) The ability to encode historical knowledge directly. In many domains, cases can be obtained from existing case histories, repair logs, or other sources, eliminating the need for intensive knowledge acquisition with a human expert. 2) Allows shortcuts in reasoning. If an appropriate case can be found, new problems can often be solved in much less time than it would take to generate a solution from rules or models and search. 3) It allows a system to avoid past errors and exploit past successes. CBR provides a model of learning that is both theoretically interesting and practical enough to apply to complex problems. 4) Extensive analysis of domain knowledge is not required. Unlike a rule-based system, where the knowledge engineer must anticipate rule interactions, CBR allows a simple additive model for knowledge acquisition. This requires an appropriate representation for cases, a useful retrieval index, and a case adaptation strategy. 5) Appropriate indexing strategies add insight and problem-solving power. The ability to distinguish differences in target problems and select an appropriate case is an important source of a case-based reasoner’s power; often, indexing algorithms can provide this functionality automatically. The disadvantages of case-based reasoning include: 1) Cases do not often include deeper knowledge of the domain. This handicaps explanation facilities, and in many situations it allows the possibility that cases may be misapplied, leading to poor quality or wrong advice. 2) A large case base can suffer problems from store/compute trade-offs. 3) It is difficult to determine good criteria for indexing and matching cases. Currently, retrieval vocabularies and similarity matching algorithms must be carefully hand crafted; this can offset many of the advantages CBR offers for knowledge acquisition [1][2][4][7][8][14]. C. Model-based Reasoning The advantages of model-based reasoning include: 1) The ability to use functional/structural knowledge of the domain in problem solving. This increases the reasoner’s ability to handle a variety of problems, including those that may not have been anticipated by the system’s designers. 2) Model-based reasoners tend to be very robust. For the same reasons that humans often retreat to first principles when confronted with a novel problem, model based reasoners tend to be thorough and flexible problem solvers. 3) Some knowledge is transferable between tasks. Model- based reasoners are often built using scientific, theoretical knowledge. Because science strives for generally applicable theories, this generality often extends to model-based reasoners. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 10, 2013 90 | P a g e www.ijacsa.thesai.org 4) Often, model-based reasoners can provide causal explanations. These can convey a deeper understanding of the fault to human users, and can also play an important tutorial role. The disadvantages of model-based reasoning include: 1) A lack of experiential (descriptive) knowledge of the domain. The heuristic methods used by rule-based approaches reflect a valuable class of expertise. 2) It requires an explicit domain model. Many domains, such as the diagnosis of failures in electronic circuits, have a strong scientific basis that supports model based approaches. However, many domains, such as some medical specialties, most design problems, or many financial applications, lack a well-defined scientific theory. Model-based approaches cannot be used in such cases. 3) High complexity. Model-based reasoning generally operates at a level of detail that leads to significant complexity; this is, after all, one of the main reasons human experts have developed heuristics in the first place. 4) Exceptional situations. Unusual circumstances, for example, bridging faults or the interaction of multiple failures in electronic components, can alter the functionality of a system in ways difficult to predict using an a priori model [4][10][12]. D. Hybrid Design An important area of research and application is the combination of different reasoning models. With a hybrid architecture two or more paradigms are integrated to get a cooperative effect where the strengths of one system can compensate for the weakness of another. In combination, we can address the disadvantages noted in the previous discussion. For example, the combination of rule- based and case-based systems can: 1) Offer a natural first check against known cases before undertaking rule-based reasoning and the associated search costs. 2) Provide a record of examples and exceptions to solutions through retention in the case base. 3) Record search-based results as cases for future use. By saving appropriate cases, a reasoner can avoid duplicating costly search. The combination of rule-based and model-based systems can: 1) Enhance explanations with functional knowledge. This can be particularly useful in tutorial applications. 2) Improve robustness when rules fail. If there are no heuristic rules that apply to a given problem instance, the reasoner can resort to reasoning from first principles. 3) Add heuristic search to model-based search. This can help manage the complexity of model-based reasoning and allow the reasoner to choose intelligently between possible alternatives. The combination of model-based and case-based systems can: 1) Give more mature explanations to the situations recorded in cases. 2) Offer a natural first check against stored cases before beginning the more extensive search required by model-based reasoning. 3) Provide a record of examples and exceptions in a case base that can be used to guide model-based inference. 4) Record results of model-based inference for future use [4][5][6]. III. HYBRID REASONING MODELS A. Sequence Models In this, in the first step, a rough solution is given, and in the second step, the precise solution is given by refining the rough one. Fig. 1. CBR followed by RBR Fig. 2. RBR followed by CBR RBR CBR Problem Solution CBR RBR Problem Solution (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 10, 2013 91 | P a g e www.ijacsa.thesai.org B. Conditional Model In this, if the solution given in the first step is acceptable then it is used as a solution of the given problem & otherwise next steps are invoked. Fig. 3. RBR-Controller-CBR Fig. 4. CBR-Controller-RBR Fig. 5. CBR-Controller-MBR Fig. 6. CBR-Controller-RBR-Controller-MBR CBR Controller CControll Problem Solution RBR Controller MBR CBR Controller CControll Problem Solution MBR CBR Controller CControll Problem Solution RBR RBR Controller CControll Problem Solution CBR (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 10, 2013 92 | P a g e www.ijacsa.thesai.org IV. FUNCTIONAL ARCHITECTURE OF “KNAPS-CR” “KNAPS-CR” integrates problem solving and learning into one architecture. The flow of control and information between the knowledge base and the processes of problem solving and learning in “KNAPS-CR” is shown in Figure 7. The Figure 7 illustrates that problem solving in “KNAPS- CR” is performed by a combination of model-based, case- based and rule-based reasoning (MBR, CBR and RBR, respectively). The learning combines case-based (CBL) and explanation-based (EBL) methods. The process of selecting the initial reasoning paradigm starts when a set of relevant features of a problem has been identified. This feature set typically contains input features as well as inferred features, i.e. features that the system derives from the input features by using its knowledge. Fig. 7. KNAPS-CR’s Functional Architecture - 1. Combined Reasoning; 2. Sustained Learning; 3. Knowledge Base If the set of relevant features gives a reminding to a previous cases that is above a particular strength - called the reminding threshold - case based problem solving is tried, & then for further refinement of the solution, the rule based reasoning is attempted. Relevant features may be input features or features inferred from the object domain model. If both the case base & the rule base fail to produce a result, the controller re-evaluates its previous decision, given the current state of the system. If a solution was derived by modifying a previous solution, a new case is stored and difference links between the two cases are established. A new case is also created after a problem has been solved from rules or the deeper knowledge model. Heuristic rules are integrated within the conceptual model and available for the same tasks as the conceptual domain model in general. A rule may be used to support learning. V. MODEL OF COMBINED REASONING (CBR, RBR & MBR) IN “KNAPS-CR” The combination of case-based & rule base method serve as the primary reasoning paradigm in “KNAPS-CR”, the model based reasoning is used - as separate reasoning method - only if the combination of case-based & rule-based methods is unable to suggest a solution. Fig. 8. Combined Reasoning in “KNAPS-CR” (CBR = Case-Based Reasoning, RBR = Rule-Based Reasoning, MBR = Model-Based Reasoning). The choice of reasoning method is made after the system has gained an initial understanding of the problem. This initial understanding process (described in the next section) results in an activated problem context, including a set of relevant features for describing the problem, a structure of problem solving (sub) goals, and a hierarchy of possible faults. The choice of reasoning method is made after the system has gained an initial understanding of the problem.This initial understanding process (described in the next section) results in (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 10, 2013 93 | P a g e www.ijacsa.thesai.org an activated problem context, including a set of relevant features for describing the problem, a structure of problem solving (sub) goals, and a hierarchy of possible faults. First, “KNAPS-CR” will attempt to solve the problem by case-based reasoning. The relevant findings are combined into a set of remindings, where each reminding points to a case (or a class of cases) with certain strength. If some cases are pointed to by remindings with strengths above the reminding threshold, the cases most strongly reminded of are retrieved. If no such reminding is produced, the system will trigger its rule-based reasoning method. However, before doing that it will normally try to elaborate on the findings of the cases most strongly reminded of. The purpose of this is to improve a weak match by looking for common states, constraints, etc., which will imply a stronger similarity than determined by the basic case retrieval method. Whether the elaboration on a weak match is attempted or not depends on the strength of the strongest reminding and the size and strength of the case base relative to the rule base. If acceptable matches are found, then rule based reasoning is used to further refine the solutions obtained by case based reasoning. If no cases were reminded of in the first place, “KNAPS-CR” will also try its rule-based reasoning method, i.e. attempt to solve the problem by a combined forward chaining (from the relevant findings) and backward chaining (from the fault hierarchy) within the rule base. The solution (fault and - possibly - treatment) is evaluated to see if it is acceptable for the current problem. If the system is unable to produce a good enough explanation to accept or reject the solution candidate, it is presented to the user for evaluation. If for any reason the solution is unacceptable, a check is performed to determine whether the solution would be accepted if slightly modified, in which case a modification is attempted. When no more modifications are relevant and no more new cases are available for use, “KNAPS-CR” gives up case-based reasoning. The input to a reasoning process is a problem description. This may be a description of the user’s problem, or a partial solution of this problem – for example a set of descriptors which includes a fault hypothesis, given as input to the retrieval of a case containing a suitable repair. VI. CONCLUSIONS In this research, we combined CBR with RBR, MBR in “KNAPS-CR” model. In our experiment and analysis, this new CBR integrated hybridized model i.e. “KNAPS-CR” model supported a wide range of tasks, including interpretation and argumentation, design and synthesis, planning, and management of long term medical conditions. Many useful synergies emerged as different reasoning strategies extend and complement each other. Integrated systems have enabled more accurate modelling of domain knowledge, compensation for incomplete domain models and rule bases, compensation for small case bases, simplification of knowledge acquisition, improved solution quality, improved system efficiency, leveraging of past experiences, and compensation for shortcomings inherent in individual reasoning strategies. Thus integrations of CBR with other reasoning modalities continue to proliferate, providing both practical benefit and insight into multi-modal reasoning processes. There are still a large number of important and challenging problems to be addressed in order to improve the quality and usefulness of expert systems for practical, real world problems. The research reported here has addressed the problem of how to achieve, and continually maintain, a higher level of competence and robustness in such systems than what they possess today. In “KNAPS-CR” systems, problem has been approached from two sides:  Strengthening of the problem solving capability by combining several reasoning paradigms within a knowledge-rich environment, focusing on case-based reasoning as the major method.  Enabling a continually improvement of an incomplete knowledge base by learning from each problem solving experience, using a knowledge-intensive, case-based learning method. The resulting framework, architecture, system design, and representation platform - i.e. the “KNAPS-CR” approach - has been motivated and supported by relating it to strengths and weaknesses of other approaches REFERENCES [1] A. Aamoth and E. Plaza, “Case-Based Reasoning: Foundational Issues,Methodological Variations and System Approaches”, Artificial Intelligence Communications, 7(1), 1994, pp. 39-59. [2] Aamodt, A, 1994, Explanation-driven case-based reasoning. In Wess, S, Althoff, K and Richter, M (eds) Topics in Case-Based Reasoning. Berlin: Springer, pp. 274–288. [3] Jackson P. (1999), Introduction to Expert Systems, Harlow, England: Addison Wesley Longman, Third Edition. [4] Luger, George F., Artificial Intelligence: Structures and Strategies for Complex Problem Solving, Pearson Addison Wesley, Sixth Edition [5] Russel S, and P. Norvig, (2002), Artificial Intelligence: A Modern Approach, Second Edition, Prentice- Hall, New Delhi. [6] V.S.Janakiraman, K Sureshi, (2003), Artificial Intelligence and Expert Systems, Macmillan Series in Computer Science, New Delhi. [7] Ramon López De Mántaras, David Mcsherry, Derek Bridge, David Leake, Barry Smyth, Susan Craw, Boi Faltings, Mary Lou Maher, Michael T. Cox, Kenneth Forbus, Mark Keane, Agnar Aamodt And Ian Watson “Retrieval, reuse, revision, and retention in casebased Reasoning”,The Knowledge Engineering Review, Vol. 00:0, 1–2., Cambridge University Press, 2005 [8] I. Watson,”Case-based reasoning is a methodology not a technology, Knowledge-Based Systems” , Vol. 12, 1999, pp 303-308 [9] Edwina L. Rissland and David B. Skalak, “Combining Case-Based and Rule-Based Reasoning: A Heuristic Approach”,pp534-530 [10] Stefanie Br¨uninghaus and Kevin D. Ashley(2003), “Combining Case- Based and Model-Based Reasoning for Predicting the Outcome of Legal Case.” [11] Jim Prentzas and Ioannis Hatzilygeroudis ,“Categorizing Approaches Combining Rule-Based and Case-Based Reasoning” [12] Cindy Marling, Edwina Rissland and Agnar Aamodt “Integrations with case-based reasoning” The Knowledge Engineering Review, Vol. 00:0, Cambridge University Press,2005 ,pp1–4. [13] Jaap Hage ,”A Low Level Integration Of Rule-based Reasoning And Case-based Reasoning.”,pp30-39 [14] Z. Budimac, V. Kurbalija ,“CASE BASED REASONING – A SHORT OVERVIEW” ,Proceedings of the Second International Conference on Informatics and InformationTechnology,pp222-233 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 4, No. 10, 2013 94 | P a g e www.ijacsa.thesai.org [15] Bittencourt, I.I., M. Tadeu and E.B. Costa (2006) Combining AI techniques into a legal agent-based intelligent tutoring system, in Proceedings of the Eighteenth International Conference on Software Engineering and Knowledge Engineering (SEKE-2006), Skokie, IL: Knowledge Systems Institute, 35–40. [16] Bogaerts, S., and Leake, D. 2005. A Framework for Rapid and Modular Case-Based Reasoning System Development. Technical Report TR 617, Computer Science Department, Indiana University, Bloomington, IN. [17] Hanemann, L. (2006) A hybrid rule-based/case-based reasoning approach for service fault diagnosis, in Proceedings of the 2006 International Symposium on Frontiers in Networking with Applications, in conjunction with the Twentieth IEEE International Conference on Advanced Information Networking and Applications, Washington, DC: IEEE, 734–740. [18] K. Ashwin Kumar, Yashwardhan Singh, Sudip Sanyal, (2009) Hybrid approach using case-based reasoning and rule-based reasoning for domain independent clinical decision support in ICU, Expert Systems with Applications 36, 65–71. 
work_yjcmqmv2rfca5c3iogj7b3474u	----	Research Article Eliminating Rogue Femtocells for IoT Open Meter System Based on Expert System Yong Xiao,1 Bin Qian,1 Ziwen Cai ,1 Liang Hong,2 and Sheng Su 2 1Electric Power Research Institute of China Southern Power Grid, Guangzhou 510080, China 2College of Electrical & Information Engineering, Changsha University of Science and Technology, Changsha 410004, China Correspondence should be addressed to Ziwen Cai; 578119620@qq.com Received 8 July 2019; Revised 12 September 2019; Accepted 16 October 2019; Published 6 December 2019 Academic Editor: Di He Copyright © 2019 Yong Xiao et al. /is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. /e Internet of things (IoT), including power meters, water meters, natural gas meters, and meter collectors in an open metering system (OMS), which is dispersed around the user side, relies on wireless virtual private networks (VPNs) to communicate with head end, and thus it is exposed to malicious cyber attacks. /e General Packet Radio Service (GPRS), which is vulnerable to rogue femtocells, is widely used for communication among meter collectors and the head end. Because telecommunication fraud related to rogue femtocells is a serious offence, rogue femtocells will be turned on for some time and immediately turned off and moved from here to there to escape from being caught. /e signal strength (SS) of rogue femtocells is characterized by abrupt changes. Because meter collectors and lawful femtocells are deployed at the fixed location, there is a notable difference between signal strength profile of lawful and rogue femtocells. Prior knowledge of variation of signal strength is utilized to formulate rules to detect rogue femtocells. An expert system is developed to detect rogue femtocells and prevent meter collectors from attaching to them. Numerical simulation indicates that the proposed approach can detect both stationary and moving rogue femtocells online. Since computation load of the proposed approach is not high, it can be implemented in existing IoTmeter collectors with limited computation resource and the proposed approach can harden cyber security of OMS. 1. Introduction Open metering system (OMS) is an infrastructure in- tegrating natural gas meters, water meters, and electricity meters together with advanced metering infrastructure (AMI) [1]. /e natural gas meters and water meters com- municate with meter collectors in the same way as electricity meters. /e meter collectors collect electricity, water, and gas usage data and upload them to the head end of OMS [2, 3]. IoT electricity/water/gas meters and meter collectors in the OMS are the most visible parts of the smart grid, which plays a key role in communication between customers and the utilities [4]. /ey send electricity/water/gas consumption to head end, receive tariff data, and update account deposit according to electricity/water/gas charge [1, 5]. In an OMS, while water/gas meters communicate with meter collectors via wireless M-Bus, electricity meters communicate via power line communication (PLC) or RS-485. /e meter collectors communicate with the head end of OMS via the General Packet Radio Service (GPRS) [4, 5]. It is required to keep the integrity and privacy of user consumption data when transmitted between dispersed meters and the utilities [6]. Symmetric data encryption/ decryption integrated circuits (ICs) are embedded in elec- tricity/water/gas meters to facilitate identity authentication and encrypted communication. Since IoT devices have limited computation resource, light key management pro- tocol for IoTmeters is developed in [7, 8]. Because the public shared network is vulnerable to cyber attacks, such as spoofing, eavesdropping, and man-in-the-middle attacks, Internet service providers offer a virtual private network (VPN) for secured communication of OMS [9]. Commu- nication beyond the VPN tunnel remains exposed to malicious adversaries. Because General Packet Radio Service (GPRS) supports only one-way subscriber authentication, it is feasible for a malicious adversary to take control of meter Hindawi Journal of Engineering Volume 2019, Article ID 4910232, 9 pages https://doi.org/10.1155/2019/4910232 mailto:578119620@qq.com https://orcid.org/0000-0003-0317-5204 https://orcid.org/0000-0003-3201-0193 https://creativecommons.org/licenses/by/4.0/ https://creativecommons.org/licenses/by/4.0/ https://doi.org/10.1155/2019/4910232 collectors with rogue femtocells [10]. Once forced to attach to a rogue femtocell, the meter collectors as well as associated meters are in great danger [11]. According to [12], once lots of smart meters are under malicious control, an adversary could cause extensive power outage and dramatic changes in load, which could trigger cascading outage to blackout and result in disorder of society. Unlike GPRS, both Universal Mobile Telecommunica- tions System (UMTS) and Long-Term Evolution (LTE) support two-way authentication [13]. Hence, it is widely believed that the threat of rogue femtocells can be eliminated once meter collectors are replaced by the ones that com- municate with UMTS and LTE mobile network. However, meter collectors are designed to be backward compatible, and many of them can communicate via GPRS, UMTS, and LTE. Malicious adversaries can broadcast an interference signal to disable UMTS and LTE communication. Conse- quently, meter collectors fall back to communicate by GPRS and can be compromised according to the aforementioned approach. /erefore, threat of rogue femtocells will exist till meter collectors are replaced with the ones that do not support GPRS. /erefore, ways to detect rogue femtocells for IoTmeter collectors with limited computational resource are desirable. Unlike mobile communication terminals such as mobile phones, both meter collectors and lawful femtocells are located in the fixed places. /erefore, signal strength of neighboring lawful femtocells sensed by a meter collector has similar profile, which is notably different from that of a rogue femtocell. /e difference in the signal strength (SS) profile of rogue and lawful femtocells is utilized as a radio fingerprint, and an expert system is developed to detect rogue femtocells in this paper. Our contribution is two-fold. /e first consists of pre- senting similarity of signal strength profile of lawful fem- tocells and their difference to that of rogue femtocells, which lay the foundation to detect rogue femtocells. /e second contribution consists of developing a signal strength profile clustering based on an expert system to detect rogue femtocells. /e remaining part of the paper is organized as follows. /e structure of OMS, threat of rogue femtocells, and related work to detect a rogue femtocell are given in Section 2. Differences among SS of lawful and rogue femtocells are investigated in Section 3. A prior knowledge-based expert system is developed in Section 4 to detect rogue femtocells. Numerical simulation of proposed approach is shown in Section 5. Section 6 concludes the paper. 2. Threat of Rogue Femtocells Hijacking 2.1. Structure of OMS. An OMS is composed of electricity/ water/gas meters, meter relays, meter collectors, the head end system, and communication system [14] as shown in Figure 1. /e communication system of OMS is composed of 4 parts, i.e., neighborhood area network, wireless access network, data transmission network, and power core net- work [14]. /e function of each part is depicted as follows: (i) Neighborhood area network: most electricity meters and meter relays communicate with meter collec- tors via PLC or RS-485 while water/gas meters communicate with the meter collector via wireless M-Bus. (ii) Wireless access network: meter collectors com- municate with the head end of an OMS via public shared networks provided by Internet service pro- viders. Meter collectors are attached to femtocells, and then encrypted meter data are transmitted from meter collectors to the carrier network of the In- ternet service provider. (iii) Data transmission network: it is a carrier network provided by the Internet service provider. In order to keep the integrity and privacy of metering data, IPSec VPN is utilized to establish a VPN tunnel. Channel encryption and end-to-end encryption are employed to guarantee security of the data trans- mitted within VPN tunnel [14]. (iv) Power core network: it is a private network of power utility. Firewall is deployed between the access router device and the head end of OMS to prevent malicious external connection [15]. 2.2. Rogue Femtocells +reat. /ere are usually several femtocells in the neighborhood of a meter collector. /e meter collector compares communication parameter of all available femtocells and chooses to attach to the best available femtocell, usually the one with the strongest signal strength. GPRS-Attach initiated by a meter collector has 3 phases as shown in Figure 2. /e first phase is the Radio Resource Control (RRC). A request of a GPRS-Attach initiated by a meter collector will be sent to Serving GPRS Support Node (SGSN) to establish a RRC connection between the meter collector and radio network controller. /e second phase is authentication and encryption. /e SGSN authenticates the identity of the user and equipment to start ciphered data exchange. In the third phase, SGSN address will be registered with the Home Location Register (HLR), a database containing pertinent data regarding subscribers authorized to use a global system for communications network. /ereafter, SGSN gets the services which the meter collector can use from HLR [16]. It can be observed from Figure 2 that there is only one- way authentication of meter collectors in GPRS while there is no authentication for femtocells towards meter collectors. /erefore, communication over wireless access network is exposed to hijacking of rogue femtocells-based cyber attack. Malicious adversaries could measure and collect the carrier frequency of the targeted cell as described in [16, 17]. /ereafter, it could broadcast the same carrier frequency using rogue femtocells with higher power. In this case, meter collectors would be forced to detach from the lawful fem- tocell and attach to the rogue femtocell due to its better communication parameters and stronger SS. Furthermore, malicious adversaries could get International Mobile Sub- scriber Identity (IMSI) and International Mobile Equipment 2 Journal of Engineering Identity (IMEI) of meter collectors via the rogue femtocells to implement further attack [18]. According to [19], the smart meters using encryption algorithm AES-128 could be cracked and meters can be attacked by bypassing encryption that was designed to secure their communications. /ereafter, the malicious intruder could take full control of the meter and shut down power supply of the user. 2.3. Related Works on Detecting Rogue Femtocells. Since communication security has become increasingly prominent Meter collector Femtocell SGSN HLR EIR 1. Attach request (IMSI) 2. Send authentification info (IMSI) 3. Send authentication info ACK 4. Authentication and ciph. request 5. Authentication and ciph. response 6. Identity request 8. Check IMEI 9. Check IMEI ACK (IMEI status) 10. Update GPRS location (IMSI) 11. Insert subscriber data 15. Attach complete RRC connection Activate ciphering 7. Identity response 12. Insert subscriber data ACK 13. Update GPRS location ACK 14. Attatch accept Figure 2: Attach process of a meter collector. VPN tunnelElectricity meter and meter relay ISP core network Access router IP carrier network Access router Firewall Head end of OMS SGSN/ PCF GGSN/ PDSN Wireless access network Data transmission network Power core network Gas/water meter PLC Femtocells Meter collector Neighborhood area network M_bus Figure 1: Communication architecture of OMS. Journal of Engineering 3 in modern society, various approaches have been developed to detect rogue femtocells. An observation-based detection technique is proposed in [19] to detect attacks on femto- cells. Smart phones are used as observers to provide re- quired information of femtocells. /ereafter, observation management server could identify attack from femtocells including location movement attack, impersonating sub- scriber attack, etc. However, IoTmeters of OMS are notably different from average mobile intelligent terminals. For example, meter collectors usually upload data to the head end system once an hour. Even though a management system in the head end of OMS could detect a rogue femtocell with uploaded data of meter collectors, it is too late. /erefore, IoTmeters should identify rogue femtocells by themselves. Since the signal of a femtocell can cover only a small area and physical location of femtocells can be queried with its Location Area Code (LAC), the distance between femtocells and the user communication device can be estimated. /ereafter, the femtocells with an unreasonable distance can be detected as rogue ones [21]. However, it is rather difficult to identify a rogue femtocell once it duplicates the LAC of neighboring lawful femtocells with rational distance. In recent years, radio frequency fingerprinting (RFF) is considered as one of the promising techniques to enhance wireless security of IoT devices. Since wireless communi- cation devices have slim differences in their communication signal induced by the variance in component parameters, the RFF is utilized to authenticate identity of a wireless device [13]. Once we establish a database that records RFF of all lawful femtocells, the femtocell that is not in the white list can be identified as a rogue one [22]. However, IoT meter collectors have limited computational resource, and they cannot detect and extract the RFF of IoT devices, currently. Since the threat of rogue femtocells originates from one- way authentication of GPRS, UMTS and LTE support two- way authentication between communication terminals, and it seems to be possible for meter collectors to prevent hijacking of rogue femtocells [23]. However, meter collec- tors are designed to be backward compatible. Malicious adversaries can broadcast interference signal to disable UMTS and LTE communication. /erefore, meter collectors that communicate with UMTS and LTE can be compro- mised with aforementioned approach once they support GPRS communication. Hence, threat of rogue femtocells will exist till meter collectors are replaced with the ones that do not support GPRS. What is worse, LTE communication may not be as se- cure as expected, either. Recent investigation indicates that LTE RRC redirection attacks [24], aLTEr attacks, and so on [25] could compromise communication terminals using LTE, too. 3. Differences in SS of Lawful and Rogue Femtocells With limited computation and memory resources, IoT de- vices have a computation bottleneck to limit the support for advanced applications involving complex algorithms [4, 9]. /erefore, the aforementioned ways to detect hijacking of rogue femtocells may not fit for resource-constrained IoT meter collectors. Unlike mobile intelligent terminals, IoT meter col- lectors locate in the fixed places. A meter collector could detect several femtocells in its neighborhood. It chooses to attach to the one with the best communication parameters, usually the one with the strongest SS. /e SS of neigh- boring femtocells detected by a meter collector could be influenced by the distance, buildings, and meteorological factors at large. /e former two are constant and thus will not result in the variation of sensed SS of neighboring femtocells. /erefore, the sensed SS is usually determined by the meteorological factors. Consequently, variations of sensed SS of neighboring lawful femtocells have similar profiles. /e SS around a meter collector is measured with a base station analyzer. /e SS profiles measured consecutively for 24 hours is shown in Figure 3. /e horizontal axis denotes time span of a day, and the vertical axis denotes signal strength of femtocells in dBm. /e SS of neighboring femtocells varies according to their distance from the meter collector. /e SS of all neighboring lawful femtocells fluctuates slightly, and they have similar profiles around the clock. Rogue femtocells-related telecommunication fraud is one of the most serious crimes affecting the safety of people’s personal property, which has been heavily cracked down and punished by Chinese courts [26]. In order to provide a secure communication circumstance for their customers, Internet service providers will track, locate, and crack rogue femtocells. In this case, a rogue femtocell will be turned on for some time and imme- diately turned off, now and then, here and there, to escape from being caught. Once a rogue femtocell is turned on, its SS escalates much higher than that of a lawful femtocell. Consequently, meter collectors near to a rogue femtocell will seamlessly hand over from the lawful femtocell to the rogue femtocell if there is not any approach to stop it. /e SS profiles of a rogue femtocell and several lawful femtocells are plotted in Figure 3. It can be observed that there is notable difference among the SS profiles of lawful femtocells and that of the rogue femtocell. Unlike lawful femtocells in the neighborhood with similar SS profiles, SS of a rogue femtocell is characterized by abrupt changes. /erefore, SS profile of femtocells sensed by a meter col- lector could be used as a steady-state RFF and clustered to identify rogue femtocells. 4. Expert System to Detect Rogue Femtocells An expert system is a computer system that emulates the decision-making ability of a human expert and is designed to solve complex problems by reasoning through bodies of knowledge, represented mainly as if-then rules [27, 28]. An expert system is employed to identify rogue femtocells whose SS profile is notably different from that of lawful femtocells. 4 Journal of Engineering 4.1. Rule to Identify a Rogue Femtocell. Once a rogue fem- tocell appears, it broadcasts stronger signal power to hijack wireless communication devices. When a meter collector communicating by GPRS reselects to attach to a femtocell with higher SS, the identity of the femtocell should be verified. Unlike lawful femtocells in the fixed location, a rogue femtocell will be turned on and off from time to time or kept moving here and there to prevent the rogue femtocell from being caught. /erefore, its SS profile sensed by a meter collector in the fixed location is fickle and unstable. Difference in SS profile of two femtocells can be reflected with their Euclidean distance. /e Euclidean distance between lawful femtocells is small since they have similar SS profiles, whereas the Euclidean distance between lawful and rogue femtocells is much larger. /is can be used as prior knowledge to formulate a rule to identify rogue femtocells. /e average Euclidean distances among all sensed femtocells can be used to determine the threshold to detect a rogue femtocell. Once there is a rogue femtocell, it can be picked out since its Euclidean distances to the other femtocells are notably higher than the threshold. Without rogue femtocells, Euclidean distances among lawful fem- tocells are lower than the threshold, and meter collectors can hand over to the femtocell with higher SS without difficulty. We find that the threshold can be set to 1.5 times the average Euclidean distances of all femtocells by trial and error. 4.2. Length of Time Window. In order to prevent from being cracked, a rogue femtocell will not stay in the same place for a long time. According to reported activity of rogue fem- tocells related to fraud, most rogue femtocells operate consecutively for several hours [26]. /erefore, we set a 24- hour time window to identify rogue femtocells in this paper. /at is, the SS of the past 24 hours would be recorded and calculated to detect rogue femtocells. Once a rogue femtocell does not stay for a day, it can be identified and meter collectors will not attach to it. Once a meter collector reselects the femtocell it attaches to, the way to determine its identity can be demonstrated as shown in Figure 4 with the following steps: Step 1. Calculate Euclidean distances of SS between every two femtocells in the previous 24 hours. Step 2. Calculate average Euclidean distances of SS of each femtocells to all other femtocells. Step 3. Calculate average Euclidean distances of SS of all femtocells and then multiply them by 1.5 to obtain the threshold. Step 4. If the average Euclidean distance to all other femtocells of the femtocells with highest SS is larger than the threshold, it is identified as a rogue one and the meter collector is not allowed to attach to it. Otherwise, it is a lawful femtocell, and the meter collector hands over to it. 5. Numerical Simulation 5.1. Identifying Stationary Rogue Femtocells. A meter col- lector prefers to attach to the femtocell with stronger SS. Once there is a femtocell with stronger SS than the one that meter collector attaches to for 5 seconds, the meter con- nector will hand over to it. In order to avoid frequent cell reselection, meter collectors will implement cell reselection with a random time span ranging from 20 seconds to 620 seconds. For simplicity, it is supposed that cell reselection happens every 15 minutes in the paper. /e SS of lawful femtocells and a stationary rogue femtocell with 192 data points in 2 days is shown in Figure 5. Detect femtocells with higher SS Calculate Euclidean distances of SS between each femtocells in previous 24 hours Euclidean distances of femtocell with highest SS > threshold? Calculate average Euclidean distances of SS of each femtocells to all other femtocells Calculate average Euclidean distances of SS of all femtocells (exclude the femtocells with the highest SS) and produce the threshold Yes No Hand over to the femtocells Refuse to hand over Figure 4: Flowchart to detect a rogue femtocell. 00 : 00 00 : 0004 : 00 08 : 00 12 : 00 16 : 00 20 : 00 –40 –60 –80 –100 –120 –140 –160 –180 Si gn al st re ng th (d Bm ) #1 femtocell #2 femtocell #3 femtocell #4 femtocell #5 femtocell Rogue femtocell Figure 3: Signal strength profile of lawful and rogue femtocells. Journal of Engineering 5 /e SS of a femtocell can be normalized according to the following equation: x ∗ i � xi − mini maxi − mini , (1) where xi denotes SS of ith femtocell; mini and maxi denote the weakest and strongest signals in the current time window of investigated ith femtocell; and x∗i denotes the normalized SS of ith femtocell./e Euclidean distances between the ith and jth femtocells can be calculated according to the fol- lowing equation: Di,j � ������������� 􏽘 96 k�1 x ∗ i,k − x ∗ j,k􏼐 􏼑 2 􏽶 􏽴 , (2) where Di,j denotes Euclidean distances between ith and jth femtocells in the previous 24-hour time window. According to Figure 5, a stationary rogue femtocell appears at 23:30. /e Euclidean distances between SS of all femtocells at that time are calculated according to equation (2) and are listed in Table 1. It can be observed from Table 1 that the Euclidean distances between 2 lawful femtocells are usually small. For example, the Euclidean distance between the 1st and 2nd femtocells is 0.969 and that among lawful femtocells varies around 0.95. However, the SS of the rogue femtocells is notably far away from those of lawful ones in the state space. Its average Euclidean distance to lawful ones is around 7.7, which is notably larger than those among lawful ones. Add up Euclidian distance of SS of the 1st femtocell to all other femtocells (i.e., first row of Table 1) to produce the average Euclidian distance of the 1st femtocell to all other femtocells (2.302 of the 1st femtocell at that time). /e average Euclidian distances of SS of femtocells are calculated every 15 minutes to get a time series of Euclidian distance of SS in Figure 6. /e mean of all average Euclidean distances is 3.180. Multiply it by 1.5 and get the threshold of 4.770. At 23:30, the average Euclidean distances of all lawful femtocells are around 2.25, and the average Euclidean distance of the rogue femtocell is 7.727, which is notably larger than the threshold. /erefore, it can be identified as a rogue femtocell. Since the rogue femtocell has stronger SS throughout the following 24 hours according to Figure 5, the meter collector will keep on reselecting to attach to the rogue femtocell with stronger SS. /e average Euclidean distances of every fem- tocell to the others in the following day can be calculated and are plotted as shown in Figure 6. /e threshold of 4.770 can be used in the following day to detect potential rogue femtocells. (i) It is obvious that the average Euclidean distances of SS of these lawful femtocells are rather close to each other because their SS has the similar profile. However, the average Euclidean distance of SS of the rogue femtocell is significantly higher than the lawful ones throughout the following day. (ii) Since the average Euclidean distance of the SS of the rogue femtocell is larger than the threshold, it can be picked out in the following 24 hours. 00 : 00 00 : 0004 : 00 08 : 00 12 : 00 16 : 00 20 : 00 00 : 0004 : 00 08 : 00 12 : 00 16 : 00 20 : 00 –20 –40 –60 –80 –100 –120 –140 –160 –180 –200 Si gn al st re ng th (d Bm ) #1 femtocell #2 femtocell #3 femtocell #4 femtocell #5 femtocell Rogue femtocell Figure 5: Signal strength profile of stationary rogue and lawful femtocells. Table 1: Euclidean distances of SS between femtocells at 23:30. i � 1 i � 2 i � 3 i � 4 i � 5 Rogue Average distance i � 1 0.969 0.918 0.898 1.015 7.710 2.302 i � 2 0.969 0.850 0.857 0.859 7.723 2.252 i � 3 0.918 0.850 0.941 0.809 7.721 2.248 i � 4 0.898 0.857 0.941 0.947 7.676 2.264 i � 5 1.015 0.859 0.809 0.947 7.803 2.287 Rogue 7.710 7.723 7.721 7.676 7.803 7.727 6 Journal of Engineering (iii) Note that there is a notable escalation in the average Euclidean distances of lawful femtocells from 8:00 a.m. to 9:00 a.m., while there is a sag in the average Euclidean distance of the rogue femtocell. It is speculated that this is because the fog dissipates from 8:00 a.m. and disappears at 9:00 a.m. (iv) /ere is a turning point of the average Euclidean distance of the rogue femtocell at 4:00 a.m. It is speculated that the turning point is caused by the decline of SS of lawful femtocells when fog appears at this time. To conclude, average Euclidean distances of SS in the time window of previous 24 hours could be utilized to detect rogue femtocells. When a meter collector reselects to attach to a new femtocell, it will compare the average Euclidean distance of SS of the new femtocell with the threshold to decide whether to attach or not. If it is larger than the threshold, the femtocell is identified as a malicious one and the meter collector will not attach to it. In the following 24 hours, the threshold remains unchanged and collectors will not attach to it even though it has the strongest SS. 5.2. Identification of Moving Rogue Femtocells. Unlike sta- tionary rogue femtocells, moving ones keep changing their locations from time to time. At the beginning, the SS of a moving rogue femtocell is as shown in Figure 7. It can be observed that it emerges at 23:15, and its SS achieves its maximum at 01:00 a.m. and then disappears gradually. /e moving rogue femtocell forces the meter collector to attach to it at 01:00 when its SS is higher than that of lawful ones. Also, the average Euclidean distances between every fem- tocell at 01:00 a.m. are calculated and are listed in Table 2. /e average Euclidean distances of lawful and rogue femtocells every 15 minutes are calculated and are listed in Table 2. Meter collectors implement cell reselection to attach to the rogue femtocell at 01:00, and the average Euclidean distance of SS of all femtocells is 2.842. Multiply it by 1.5 and then get the threshold of 4.263. /e rogue femtocell with an average Euclidean distance of 7.637 can be picked out easily, and the meter collector will not attach to it. /ereafter, SS of #1 femtocell #2 femtocell #3 femtocell #4 femtocell #5 femtocell Rogue femtocell Threshold 00 : 00 12 : 0002 : 00 04 : 00 06 : 00 08 : 00 10 : 00 00 : 0014 : 00 16 : 00 18 : 00 20 : 00 22 : 00 Eu cl id ea n di st an ce (d Bm ) 8 9 7 6 5 4 3 2 1 Figure 6: Average Euclidean distances of stationary rogue femtocells. 12 : 00 12 : 0016 : 00 20 : 00 00 : 00 04 : 00 08 : 00 –40 –60 –80 –100 –120 –140 –160 –180 Si gn al st re ng th (d Bm ) #1 femtocell #2 femtocell #3 femtocell #4 femtocell #5 femtocell #6 femtocell Rogue femtocell Figure 7: Signal strength profile of moving rogue and lawful femtocells. Journal of Engineering 7 the moving rogue femtocells declines and then disappears after 2:00. 6. Conclusion To defend against hijacking of rogue femtocells, a rogue femtocell detection approach for IoT meters in the fixed location is proposed in this paper. Prior knowledge of difference in SS profile among rogue and lawful femtocells is utilized to formulate a rule to identify rogue femtocells. Numerical simulation suggests that the developed expert system can pick out stationary and moving rogue femtocells, and meter collectors will not attach to the malicious rogue femtocell. Moreover, the developed system works on the specific rule and requires limited computation and memory resource, which are fit for resource-constraint IoT device. Since the developed approach works with the premise that the rogue femtocell does not stay for a long time, malicious adversaries could subvert by sustained attack for more than a day. In this way, the meter collector will attach to the rogue femtocell and malware could mount on it. Data Availability /e data used to support the findings of this study are available from the corresponding author upon request. Conflicts of Interest /e authors declare that they have no conflicts of interest. Acknowledgments /e authors would like to thank the support from the Electric Power Research Institute of China Southern Power Grid (grant no. ZNKJXM20170085), National Natural Sci- entific Funding of China (51777015), National Key R&D Program of China (2018YFB0904903), and Scientific Re- search Funding of Hunan Education Department (15A0015). References [1] S. Rinaldi, P. Ferrari, A. Flammini, E. Sisinni, and A. Vezzoli, “Uncertainty analysis in time distribution mechanisms for OMS smart meters: the last-mile time synchronization issue,” IEEE Transactions on Instrumentation and Measurement, vol. 68, no. 3, pp. 693–703, 2019. [2] A. Montazerolghaem, M. H. Y. Moghaddam, and A. Leon- Garcia, “OpenAMI: software-defined AMI load balancing,” IEEE Internet of +ings Journal, vol. 5, no. 1, pp. 206–218, 2018. [3] S. Bohn, M. Agsten, O. Waldhorst et al., “An ICTarchitecture for managed charging of electric vehicles in smart grid en- vironments,” Journal of Engineering, vol. 2013, Article ID 989421, 11 pages, 2013. [4] J. Siryani, B. Tanju, and T. J. Eveleigh, “A machine learning decision-support system improves the internet of things’ smart meter operations,” IEEE Internet of +ings Journal, vol. 4, no. 4, pp. 1056–1066, 2017. [5] E. Spano, L. Niccolini, S. D. Pascoli, and G. Iannaccone, “Last- meter smart grid embedded in an internet-of-things platform,” IEEE Transactions on Smart Grid, vol. 6, no. 1, pp. 468–476, 2015. [6] Y. Sun, L. Lampe, and V. W. S. Wong, “Smart meter privacy: exploiting the potential of household energy storage units,” IEEE Internet of +ings Journal, vol. 5, no. 1, pp. 69–78, 2018. [7] M. Wolf and D. Serpanos, “Safety and security in cyber- physical systems and internet-of-things systems,” Proceedings of the IEEE, vol. 106, no. 1, pp. 9–20, 2018. [8] G. Liu and D. Jiang, “5G: vision and requirements for mobile communication system towards year 2020,” Chinese Journal of Engineering, vol. 2016, Article ID 5974586, 8 pages, 2016. [9] A. Mosenia and N. K. Jha, “A comprehensive study of security of internet-of-things,” IEEE Transactions on Emerging Topics in Computing, vol. 5, no. 4, pp. 586–602, 2017. [10] D. Malone, D. F. Kavanagh, and N. R. Murphy, “Rogue femtocell owners: how Mallory can monitor my devices,” in Proceedings of the IEEE INFOCOM, pp. 3387–3392, Turin, Italy, April 2013. [11] G. Nico, R. Kévin, and B. Ravishankar, Weaponizing Fem- tocells: /e Effect of Rogue Devices on Mobile Telecommu- nication, 2012, https://www.tu-berlin.de/fileadmin/fg214/ Papers/femto_ndss12.pdf. [12] S. Soltan, P. Mittal, and H. V. Poor, “BlackIoT: IoT Botnet of high wattage devices can disrupt the power grid,” in Pro- ceedings of the 27th USENIX Security Symposium, Baltimore, MD, USA, August 2018. [13] Z. Haddad, M. Mahmoud, S. Taha et al., “Secure and privacy- preserving AMI-utility communications via LTE-A net- works,” in Proceedings of the IEEE 11th International Con- ference on Wireless and Mobile Computing, Networking and Communications (WiMob), pp. 748–755, Abu Dhabi, United Arab Emirates, October 2015. [14] Halifax Regional Water Commission, AMI Technology As- sessment & Feasibility Study Consolidated Report, Halifax Regional Water Commission, Excergy, Denver, USA, 2014, https://www.halifax.ca/sites/default/files/documents/home- property/water/AMI-Technology-FianlReport.pdf. [15] K. Sander and B. Roos, Security Analysis of Dutch Smart Metering Systems, University van Amsterdam, Amsterdam, Netherlands, 2008. [16] C. Xenakis, “Malicious actions against the GPRS technology,” Journal Computer Virology, vol. 2, no. 2, pp. 121–133, 2006. [17] D. Zhu, N. Pang, and Z. Fan, “A self-testing approach defending against rogue base station hijacking of intelligent terminal,” in Proceedings of the International Conference on Applied Science and Engineering Innovation, pp. 929–937, Wuhan, China, April 2015. [18] J. Martin, I. Tøndel, and G. Køien, “GPRS security for smart meters,” in 1st Cross-Domain Conference and Workshop on Availability, Reliability, and Security in Information Systems (CD-ARES), Lecture Notes in Computer Science, LNCS-8127, A. Cuzzocrea, C. Kittl, D. E. Simos et al., Eds., pp. 195–207, Springer, Regensburg, Germany, 2013. Table 2: Average Euclidean distances of SS of lawful and moving rogue femtocells. i � 1 i � 2 i � 3 i � 4 i � 5 Rogue Average distance 00:30 2.083 2.023 2.024 2.071 2.067 7.742 2.872 00:45 2.073 2.013 2.015 2.061 2.057 7.684 2.855 01:00 2.066 2.004 2.007 2.053 2.049 7.637 2.842 01:15 2.062 1.998 2.003 2.048 2.044 7.606 2.833 01:30 2.059 1.996 1.999 2.044 2.041 7.586 2.828 01:45 2.056 1.993 1.998 2.041 2.039 7.572 2.823 8 Journal of Engineering https://www.tu-berlin.de/fileadmin/fg214/Papers/femto_ndss12.pdf https://www.tu-berlin.de/fileadmin/fg214/Papers/femto_ndss12.pdf https://www.halifax.ca/sites/default/files/documents/home-property/water/AMI-Technology-FianlReport.pdf https://www.halifax.ca/sites/default/files/documents/home-property/water/AMI-Technology-FianlReport.pdf [19] A. G. Illera and J. V. Vidal, “Lights off! the darkness of the smart meters,” in Proceedings of the Europe Black Hat Conference, Amsterdam, Netherlands, October 2014, https://www.blackhat. com/eu-14/archives.html#lights-off-the-darkness-of-the-smart- meters. [20] M. J. Guezguez, S. Rekhis, and N. Boudriga, “Observation- based detection of femtocell attacks in wireless mobile net- works,” in Proceedings of the Symposium on Applied Com- puting, SAC’17, pp. 529–534, Marrakech, Morocco, April 2017. [21] C.-M. Chen, Y.-H. Chen, Y.-H. Lin et al., “Eliminating rouge femtocells based on distance bounding protocol and geo- graphic information,” Expert Systems with Applications, vol. 41, no. 2, pp. 426–433, 2014. [22] V. Namboodiri, V. Aravinthan, S. N. Mohapatra, B. Karimi, and W. Jewell, “Toward a secure wireless-based Home area network for metering in smart grids,” IEEE Systems Journal, vol. 8, no. 2, pp. 509–520, 2014. [23] H. He and J. Yan, “Cyber-physical attacks and defenses in the smart grid: a survey,” IET Cyber-Physical Systems: +eory & Applications, vol. 1, no. 1, pp. 13–27, 2016. [24] S. R. Hussain, O. Chowdhury, S. Mehnaz et al., “LTE in- spector: a systematic approach for adversarial testing of 4G LTE,” in Proceedings of the Network and Distributed Systems Security Symposium, pp. 18–21, San Diego, CA, USA, Feb- ruary 2018. [25] D. Rupprecht, K. Kohls, T. Holz et al., “Breaking LTE on layer two,” in Proceedings of the IEEE Symposium on Security & Privacy, San Francisco, CA, USA, May 2019. [26] C. Daily, Strict Crackdown on Telecommunication Fraud Coming, 2019, http://english.court.gov.cn/2019-03/14/content_ 37450221.htm. [27] S. Park, E. Lee, W. Yu, H. Lee, and J. Shin, “State estimation for supervisory monitoring of substations,” IEEE Transactions on Smart Grid, vol. 4, no. 1, pp. 406–410, 2013. [28] S. Sarkar, T. Sharma, A. Baral, B. Chatterjee, D. Dey, and S. Chakravorti, “An expert system approach for transformer insulation diagnosis combining conventional diagnostic tests and PDC, RVM data,” IEEE Transactions on Dielectrics and Electrical Insulation, vol. 21, no. 2, pp. 882–891, 2014. Journal of Engineering 9 https://www.blackhat.com/eu-14/archives.html#lights-off-the-darkness-of-the-smart-meters https://www.blackhat.com/eu-14/archives.html#lights-off-the-darkness-of-the-smart-meters https://www.blackhat.com/eu-14/archives.html#lights-off-the-darkness-of-the-smart-meters http://english.court.gov.cn/2019-03/14/content_37450221.htm http://english.court.gov.cn/2019-03/14/content_37450221.htm International Journal of Aerospace Engineering Hindawi www.hindawi.com Volume 2018 Robotics Journal of Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 Active and Passive Electronic Components VLSI Design Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 Shock and Vibration Hindawi www.hindawi.com Volume 2018 Civil Engineering Advances in Acoustics and Vibration Advances in Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 Electrical and Computer Engineering Journal of Advances in OptoElectronics Hindawi www.hindawi.com Volume 2018 Hindawi Publishing Corporation http://www.hindawi.com Volume 2013 Hindawi www.hindawi.com The Scientific World Journal Volume 2018 Control Science and Engineering Journal of Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Journal of Engineering Volume 2018 Sensors Journal of Hindawi www.hindawi.com Volume 2018 International Journal of Rotating Machinery Hindawi www.hindawi.com Volume 2018 Modelling & Simulation in Engineering Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 Chemical Engineering International Journal of Antennas and Propagation International Journal of Hindawi www.hindawi.com Volume 2018 Hindawi www.hindawi.com Volume 2018 Navigation and Observation International Journal of Hindawi www.hindawi.com Volume 2018 Advances in Multimedia Submit your manuscripts at www.hindawi.com https://www.hindawi.com/journals/ijae/ https://www.hindawi.com/journals/jr/ https://www.hindawi.com/journals/apec/ https://www.hindawi.com/journals/vlsi/ https://www.hindawi.com/journals/sv/ https://www.hindawi.com/journals/ace/ https://www.hindawi.com/journals/aav/ https://www.hindawi.com/journals/jece/ https://www.hindawi.com/journals/aoe/ https://www.hindawi.com/journals/tswj/ https://www.hindawi.com/journals/jcse/ https://www.hindawi.com/journals/je/ https://www.hindawi.com/journals/js/ https://www.hindawi.com/journals/ijrm/ https://www.hindawi.com/journals/mse/ https://www.hindawi.com/journals/ijce/ https://www.hindawi.com/journals/ijap/ https://www.hindawi.com/journals/ijno/ https://www.hindawi.com/journals/am/ https://www.hindawi.com/ https://www.hindawi.com/ 
work_yjmptpq6ezd33c2m2i7y2tkhke	----	Microsoft Word - Scheduling Flow Lines with Buffers by Ant Colony Digraph AAM.doc Author’s accepted manuscript of Rossi, A.; Lanzetta, M.: Scheduling Flow Lines with Buffers by Ant Colony Digraph, Expert Systems with Applications, vol. 40, n. 9, July 2013, pp. 3328-3340 (13), ISSN: 0957-4174, DOI:10.1016/j.eswa.2012.12.041. http://www.sciencedirect.com/science/article/pii/S0957417412012821 Scheduling Flow Lines with Buffers by Ant Colony Digraph Andrea Rossi, Michele Lanzetta Department of Civil and Industrial Engineering, University of Pisa, Italy Abstract This work starts from modeling the scheduling of n jobs on m machines/stages as flowshop with buffers in manufacturing. A mixed-integer linear programming model is presented, showing that buffers of size n-2 allow permuting sequences of jobs between stages. This model is addressed in the literature as non-permutation flowshop scheduling (NPFS) and is described in this article by a disjunctive graph (digraph) with the purpose of designing specialized heuristic and metaheuristics algorithms for the NPFS problem. Ant Colony Optimization (ACO) with the biologically inspired mechanisms of learned desirability and pheromone rule is shown to produce natively eligible schedules, as opposed to most metaheuristics approaches, which improve permutation solutions found by other heuristics. The proposed ACO has been critically compared and assessed by computation experiments over existing native approaches. Most makespan upper bounds of the established benchmark problems from Taillard (1993) and Demirkol et al. (1998) with up to 500 jobs on 20 machines have been improved by the proposed ACO. Keywords: Flexible Manufacturing Systems (FMS), Native Non-Permutation Flowshop Scheduling (NPFS), Metaheuristics, Swarm Systems, Ant Colony System (ACS), benchmark problems 1. Introduction A flow line is a conventional manufacturing system where all jobs must be processed on all machines with the same operation sequence (Figure 1). Examples of flow lines include transfer lines, assembly lines, chemical plants, logistics, and many more (Rossi et al., 2010). The flowshop scheduling problem occurs whenever it is necessary to schedule a set of n jobs on m machines so that each job visits all machines in the same order to optimize one or more objective functions. The job sequence of each machine remains unchanged in a permutation flowshop (PFS) scheduling, while in a non-permutation flowshop (NPFS) the sequence of jobs can be different on subsequent machines. The impact of buffers, along with machines, products and operations, on reconfiguration and performance evaluation is examined by Colledani and Tolio (2005). MACHINE(i-1) MACHINE(i) MACHINE(i+1) PERMUTATION FLOWSHOP MACHINE(i-1) MACHINE(i) MACHINE(i+1) MACHINE(i-1) MACHINE(i) MACHINE(i+1) t PERMUTATION MAKESPAN NON-PERMUTATION MAKESPAN BUFFER(i) MACHINE(i-1) MACHINE(i) BUFFER(i-1) BUFFER(i+1) MACHINE(i+1) NON-PERMUTATION FLOWSHOP Figure 1. Job Flowing: Permutation Flowshop (PFS) compared to Non-Permutation Flowshop (NPFS) on a physical flow line and on a Gantt diagram. In permutation flowshop no buffers are present. To achieve a feasible PFS schedule either the blocking or the no-wait condition should be applied. In the former case, a job completed on one machine may block that machine until the next downstream machine is free while in the latter the next machine must be available before a job leaves the previous one. Examples of flow line, which restrict feasible schedules to PFS schedules are factories with conveyors between machines for material transfer and assembly lines performing the final assembly of bulky products or rigid transfer lines. In non-permutation flowshop scheduling systems, interoperational or intermediate buffers between two stages are necessary and job passing is allowed (Figure 1). Buffers can be shared in the form of an automatic warehouse or an open space. In most manufacturing applications buffers are present, consequently NPFS has a wider interest. PFS scheduling is an NP-complete combinatorial optimization problem already for m = 3 machines (Garey et al. 1976). The computation complexity of permutation approaches is much lower compared with its non-permutation counterpart: there are n! different schedules for ordering jobs on machines in PFS, whilst the number of NPFS schedules increases to (n!) m . To reduce the computation time, permutation flowshop approaches are often preferred by the shop-floor manager, often applying heuristic methods to generate good solutions for practical purposes. PFS scheduler is easier to implement but unfortunately, this simplicity is bought at the price of drastically inferior schedules. Using the Graham’s notation described by a triplet α|β|γ, where field α denotes the system layout and the production flow type, field β indicates the operation characteristics and field γ denotes the adopted performance indices. The problem considered here is Fm|Bi=n-2|Cmax, where Fm stands for flowshop with m machines, Bi denotes the presence of buffers; a buffer of unlimited capacity (Bi=+∝) allows permutations, and consequently implies that the NPFS model is applied; a limited capacity Bi = n-2 is sufficient to fulfill the previous requirement, because non-permutation schedules are allowed if the capacity requirement of n-2 is satisfied, as it will be proven in § 3., along with a mixed-integer linear programming model; Cmax denotes the makespan minimization as the optimization criterion. Lower makespan implies other benefits, such as lower idle time, higher machine utilization and higher efficiency. It is well-known that the elimination of buffers has the benefit of reducing the work-in-process (WIP) and improving the line efficiency according to the lean manufacturing philosophy, however as shown in the Gantt diagram in Figure 1, changing the sequence of jobs on different machines can have the benefit of shorter schedules. When buffers between machines are present, non-permutation schedules are likely to outperform their permutation counterparts (Tandon et al., 1991). Permutation schedules are dominant up to the three-machine case. In the general case with m (>3) machines, a permutation schedule is not necessarily optimal anymore (Potts et al., 1991). The analysis of theoretical upper bounds of permutation and non-permutation schedules of the well-known 120 instance benchmarks proposed by Taillard (1993), show that system performance are improved by NPFS scheduling (Vaessens, 1996). Simulations by Weng (2000) show that up to 100 jobs and 20 machines, a buffer sizes Bi ≥ 4 gives a slight gap in performance with respect to unlimited buffers (Bi=+∝). A possible interpretation is that efficient permutated schedules include permutations of jobs that are at a distance of 4 positions. Permutations of jobs that are very distant in the sequence produce less efficient schedules (e.g. higher makespan). It has also been found that the buffer size should be proportional to the number of jobs. Consequently it can be observed that non-permutation flowshop is more and more beneficial over permutation with higher job number. OTHER OR HYBRID METAHEURISTICS FLOW LINE NON-PERMUTATION FLOWSHOP SCHED. (NPFS) PERMUTATION FLOWSHOP SCHEDULING (PFS) LAYOUT APPROACH ALGORITHM SOLUTION HEURISTICS AND METAHEURISTICS PERMUTATION SCHEDULES NON-PERMUTATION SCHEDULES WITHOUT BUFFERS WITH BUFFERS FLOWSHOP MODEL NO JOB PERMUTATION JOBS PERMUTATION ANT COLONY OPTIMIZATION (ACO) DESCRIPTION DISJUNCTIVE GRAPH (DIGRAPH) NATIVE NPFS DESIGN NON-NATIVE NPFS OTHER OR NO DESCRIPTION CURRENT WORK OTHER WORKS PFS PROBLEM Figure 2. The genesis of native Non-Permutation Flowshop Scheduling (NPFS) and alternatives from the literature to model and approach the flow line problem with and without buffers. In Figure 2 the logical flow of current work is summarized with alternatives from the literature, reviewed in the next § 2. It can be noticed that, as anticipated, scheduling on a flowline without buffers can be approached only as permutation flowshop (green path, left); permutation solutions can be used as input for non-permutation algorithms and produce non-permutation schedules (blue path, center). In § 4., the non-permutation approach (formalized in § 3.) is described by a disjunctive graph (digraph) to show how to select or design new heuristics or metaheuristics able to build natively non-permutation solutions (red path, right) as opposed to improving permutation solutions found by other heuristics or metaheuristics. The proposed native-NPFS approach (§ 5.) is based on ant colony optimization (ACO) from Bonabeau et al. (2000), as discussed in detail in § 6. Native and non-native approaches are experimentally compared in § 7. considering the benchmarks from Taillard (1990) and Demirkol et al. (1998). 2. Literature A new class of computation techniques is leading considerable attention: metaheuristic algorithms. Examples of metaheuristics in flowshop scheduling include taboo search (Nowicki and Smutniki, 1996), simulated annealing (Lin and Ying, 2009), immune algorithm (Li et al., 2012), and genetic algorithms (Ruiz et al., 2006). The relevant literature indicates that some of these metaheuristics give considerable results. Ruiz and Maroto (2005) and Ribas et al. (2010) give a comprehensive review of many heuristic and metaheuristic approaches on permutation flowshop scheduling. A total of 25 algorithms have been coded and tested by Ruiz and Maroto against Taillard benchmarks. As a conclusion, they state that the NEH heuristic from Nawaz et al. (1983) is the best heuristic with Taillard benchmarks. The performance of this heuristic, if implemented as in Taillard is outstanding, with CPU times below 0.5 seconds even for the largest instances. As opposed to the extensive reviews on PFS, there are few approaches to non-permutation flowshop scheduling (NPFS). Park et al. (1984) adapted a number of heuristics to flowshop scheduling without buffer constrains and concluded that the NEH heuristic is dominant. Afterwards, the metaheuristic performance are improved by Rajendran and Ziegler (2004) with an ant colony optimization (ACO) approach, Tasgetiren et al. (2007) with a particle swarm optimization method, Zobalas et al. (2009) by a hybrid metaheuristics, which combines the advantages of genetic algorithm with variable neighborhood search, Khalili and Tavakkoli-Moghaddam (2012) by an algorithm risen from the attraction–repulsion mechanism of electromagnetic theories, the multi-objective electromagnetism algorithm, and Ruiz et al. (2006) by a genetic algorithm, which is the current state-of-the-art algorithm tested against Taillard benchmarks. Among native approaches, a heuristic preference relation, developed as the basis for a heuristic so that only the potential job interchanges are checked for possible improvement with respect to the multicriteria objectives of minimizing makespan and total flowtime, has been applied on Taillard benchmarks by Rajendran (1995). Results have been improved by three native heuristics for the minimization of makespan derived from the minimization of total flowtime based on the sum of idle times on stages by Ravindran et al. (2005), which becomes the better algorithm also for non-native NPFS on Taillard benchmark. Among heuristics, the first native approach based on the shifting bottleneck procedure is from Demirkol et al. (1998), the same authors of benchmark problems. Koulamas (1998) has reported a new two-phase heuristic where in the first phase, it makes extensive use of Johnson’s algorithm. The second phase improves the resulting schedule from the first phase by allowing job passing between machines, i.e. by allowing non-permutation schedules. Leinsten (1990) has proposed heuristics for the flowshop scheduling where passing of jobs between two machines is permitted. The considered system Fm|Bi∈B|Cmax where B=0,1,..,+∝ involves a finite buffer capacity, but includes also no buffer capacity and unlimited buffer capacity as a extreme cases. Tandon et al. (1991) use the enumerative search and a simulated annealing approach to compare PFS and NPFS with the objective of minimizing the makespan. According to their computation experiments, non-permutation schedules usually have global optimum of 2% lower than permutation schedules; benchmarks with random processing times ranging from 1 to 50 (and from 1 to 500) time units were generated. This feature is also treated by Potts et al. (1991), who theoretically proved that there exists a family of instances for which the best permutation flowshop schedule is worse than that of the best non-permutation flowshop schedule by a factor of more than (m 0.5 /2). Pugazhendhi et al. (2003), Liao et al. (2006) and Ying et al. (2010) compare PFS and NPFS with respect to the minimization of a number of flowtime and due-date objectives: total completion time (ΣCi), total weighted completion time (ΣwiCi), total tardiness (ΣTi), and total weighted tardiness (ΣwiTi). All the authors use benchmark problems generating random processing time and approach the problem by simulated annealing. In particular, Pugazhendhi et al. (2003) compare PFS and NPFS on Fm|Bi=+∝|ΣCi and Fm|Bi=+∝|ΣwiCi by a heuristic procedure, Liao et al. (2006) evaluate NPFS on Fm|Bi=+∝|ΣTi by a tabu search and a genetic algorithm, and Ying et al. (2010) consider NPFS by a simulating annealing approach to all the objective functions, Fm|Bi=+∝|(Cmax, ΣCi, ΣwiCi, ΣTi, ΣwiTi). Their computation experience shows that all the proposed procedures yield better NPFS schedules than PFS. Metaheuristic NPFS approaches to minimize flowtime or multicriteria objectives have been proposed by Liao and Huang (2010) on Fm|Bi=+∝|ΣTi by a tabu search, and by Lin and Ying (2008), who propose three types of metaheuristics, a simulated annealing, a tabu search and a genetic algorithm. They consider sequence-dependent setup times among parts of different batches, Fm|STsd,b, Bi=+∝|(Cmax, ΣCi, ΣwiCi, ΣTi, ΣwiTi). More recently, Yagmahan and Yenisey (2010) consider a multi-objective criterion Fm|Bi=+∝|(w1Cmax + w2ΣCi) and propose an ACO. They propose an approach where the initial amount of pheromone on ant trails is a function of the best solution generated by the NEH heuristic (non-native). Mehravaran and Logendran (2012) consider sequence-dependent setup times and a bicriteria optimization for supply chain customer inventory. Aloulou and Artigues (2010) consider a general flowshop scheduling Fm|π, Bi=+∝|Cmax characterized by a partial sequence of jobs on each machine. The purpose of their work is to provide on-line a characterization of possible modifications of a predictive off-line optimal schedule while preserving optimality. Conversely, few approaches consider the computation experience achieved by benchmarks proposed in literature for the Fm|Bi=+∝|Cmax problem to objectively and quantitatively compare the performance of different algorithms. Ying, Lin and others test various metaheuristic approaches on Demirkol benchmarks: an ant colony system (Ying and Lin, 2007), an iterated greedy heuristic (Ying, 2008) and a hybrid simulated annealing – tabu search method (Lin and Ying, 2009). In their last approach, Lin and Ying use a PFS schedule as the initial solution of their tabu search (non-native). Brucker et al. (2003) consider Fm|Bi∈B|Cmax where B=0,1,..,+∝, a flowshop with buffers, similarly to Leinsten (1990). Computation experiments indicate that CPU times will be reduced and global solutions will be improved by increasing the buffer size. Their tabu search is tested on the Taillard dataset. The Taillard dataset is very effective for Fm|Bi=0|Cmax and Fm|Bi=+∝|Cmax because Taillard (1993) and Vaessens (1996) gave a method to evaluate tight upper bounds. Brucker et al. (2003) provide results on NPFS with limited capacity buffer and propose a local search procedure in their tabu search. Their approach is not native: their initial solution is for PFS evaluated by the NEH heuristic; next, their approach becomes non-permutated because the initial job sequences are permutated by local search. Jain and Meeran (2002) consider a multi-level hybrid approach for the general flowshop scheduling problem Fm|Bi=+∝|Cmax. They hybridize a scatter search and a tabu search with the neighborhood structure proposed by Nowicki and Smutnicki (1996) for job-shop scheduling. The initial solution is found by the insertion algorithm proposed by Werner and Winkler (1995). Tabu search in general requires a good initial solution as input, e.g. produced by outer heuristics, but is unable to generate solutions autonomously: results produced with this technique cannot be used for comparison with native approaches. ACO has been considered to solve some PFS scheduling systems with the objective to minimize the makespan (Alaykyran et al., 2007; Rajendran and Ziegler, 2004; Ying and Liao, 2004). Stage 1 Stage 2 Stage 3 0 O11 O12 ** arcs of machine 2 arcs of machine 3 arcs of machine 4 O22 O21 On1 On2 p11 p12 p21 p22 p41 p42 Stage 4 * O24 O14 p14 p24 p44 O43 O44 O33 O12 O22 O31 O32 O34 p43 p31 p32 p33 p34 p13 p23 Stage 1 Stage 2 Stage 3 0 7 31 ** 66 11 7 24 3 35 4 22 Stage 4 * 22 18 21 78 17 15 6 41 16 18 8 12 processing time selected paths fo r ending node job routing arcs of machine 1 candidate path of machine 2 candidate path of machine 3 candidate path of machine 4 candidate path of machine 1 node (operation) O21 p11 completion time 78 Figure 3. Disjunctive graph (digraph) for flowshop scheduling, with processing times p i j at nodes O i j for n (=4) jobs on m (=4) machines or stages (top). Bottom: an intermediate step of the list scheduling heuristic. Candidate arcs to connect candidate nodes are colored. The completion time of an operation is the longest path that connects dummy 0 and its node. Figure 3 shows the representation of a feasible schedule by a digraph. In ACO, the digraph is an internal shared memory where, in analogy with nature, artificial ants following trails of pheromone direct disjunctive arcs. Each trail from source and destination nodes represents a possible solution of the problem. Rossi and Dini (2007) describe this mechanism in detail for job-shop scheduling with effective results and integrating a number of features in their digraph: schedules with routing flexibility, sequence-dependent setup and transportation times. Ying and Lin (2007) represent the NPFS problem by a disjunctive graph and use an ant colony system to reach a fully native approach. They adopt a multi-heuristic function of visibility based on a set of schedules achieved by dispatching rules. In ACO the visibility of an ant represents the heuristic desirability, that together with the pheromone amount, drives the selection (and the direction) of a disjunctive arc of the digraph to generate a nest-food path. To the best of our knowledge a native-NPFS approach with the objective of minimizing the makespan is applied only by Ying and Lin (2007) on Demirkol benchmark. The digraph is a general-purpose tool to represent (§ 4.) and design (§ 5.) native NPFS algorithms. The example for an ACO is detailed in § 6. 3. Non-Permutation Flowshop Scheduling (NPFS) problem This section formalizes the NPFS problem by a mixed-integer linear programming model (MILP), showing its relationship with PFS and the buffer requirement. In current model, n jobs i = 1,..,n are given, to be processed on m machines M1,.., Mm and m-1 buffers Bj between machines Mj and Mj+1 with a capacity of bj units for j = 1,.., m − 1. The buffer size for machine j = m is n (i.e. the output buffer contains up to n jobs). Each job i consists of m operations Oi j (j = 1,.., m) and operation Oi j has to be processed on machine Mj without preemption (job interruption to process higher priority jobs) for pi j ≥ 0 time units. Between the operations of each job there are precedence relations in the form of a linear routing Oi1 → Oi2 → .. → Oim. A feasible job schedule is given by assigning start times Si j (and, thus, completion times Ci j = Si j + pi j) to operations Oi j (i = 1,..,n; j = 1,..,m) such that: 1. the precedence relations between jobs are respected (Ci j ≤ Si+1,j), 2. each machine processes only one job at any time ([Si j, Ci j[∩[Skj, Ckj[= 0 for j≠ k), 3. no machine is kept idle when it could start processing some operations (non-delay schedule), 4. in each buffer Bi at most n-2 jobs may be stored at the same time, and 5. no blocking conditions are applied. Blocking occurs when job i completely processed on machine Mj, which could start on Mj+1 (i.e. Ci j < Si j+1), finds buffer Bj completely filled by other jobs and machine Mj+1 is still busy; job i has to wait on machine Mj. While waiting on Mj, job i blocks this machine, so that no other job can be processed on Mj during this time. A job leaving the buffer may be replaced by job i. Since all jobs in buffer Bj have to be processed on machine Mj+1, jobs may leave buffer Bj only when machine Mj+1 becomes available. The relaxation of the blocking condition at point 5. imposes to job i, which finishes processing on machine Mj, before it can start on Mj+1 (i.e. Ci j < Si,j+1), to wait on machine Mj or within buffer Bj during time period [Ci j, Si,j+1]. Lemma � The flowshop scheduling with n jobs and m machines is Bi=+∝ if and only if the buffer size for machine j (1 ≤ j ≤ m- 1) is at least (n-2) � � In the worst case, all the jobs wait for the same machine. Let j (2 ≤ j ≤ m) be the machine for which the jobs wait for and let i the last job processed on Mj-1. Because only non delay schedules are considered (constraint 3.), job i waits on Mj for the machine Mj+1 which is busy. Hence (n-2) jobs necessarily wait in buffer Bj between machines Mj and Mj+1 � The mixed-integer linear programming (MILP) model for the NPFS problem is the following: l and l’ = the sequence positions, l, l’ = 1,2,. . .,n BigM = a sufficiently large positive value Z i l j = 1, if job i is assigned to sequence position l on machine j; 0 otherwise Objective function max Min C (1) Subject to the following constraints: 1 1 =∑ = n i jli Z mj nl ,..,1 ,..,1 = = (2) 1 1 =∑ = n l jli Z mj ni ,..,1 ,..,1 = = (3) mZmZ m j jli m j jli <∧≤ ∑∑ == 1 ' 1 nl iini ,..,1 ',,..,1 = ≠= (4) )1(11 1 1 + = ≤+∑ jji n i jij SZpS 1,..,1 −= mj (5) )1( 1 + = ≤⋅+∑ jljli n i jijl SZpS mj ni ,..,1 1,..,1 = −= (6) )1(')1(' )2( ++ ++≥−− jljljljlijli SpSZZBigM 1,..,1 ,..,1',, −= = mj nlli (7) 2,...,2 ,:',,..,1, ''11 −≤= +≤⋅<≠= nml ZpSZSSiinii jlijijjlijlj mj ,..,2= (8) Constraint (2) ensures that each job is assigned to exactly one position of the job sequence on every machine. Constraint (3) states that each position of the job sequence processes exactly one job on every machine. Constraint (4) states that, considering all the machine sequences, each job can be at most m times in the same position and at least one of these can be less than m times in the same position (non-permutation constraint). Constraint (5) denotes the starting times of the first job on every machine. Constraint (6) ensures that the (l + 1) th job in the sequence of machine j does not start until the l th job in the sequence of machine j has completed. Constraint (7) assures that the starting time of job i, which is assigned to position l in the sequence on machine j + 1 is not earlier than its finish on machine j. Constraint (8) ensures that the buffer size for machine j (2 ≤ j ≤ m) is at least (n-2). The proposed formulation limits the difference between NPFS and PFS to a single constraint of the MILP, the non- permutation constraint (4). In the PFS formulation, expression (4) is an equality, because each job needs to stay in the same position for all the machine sequences. 4. Digraph approach for non-permutation flowshop scheduling The digraph in Figure 3 used to approach this problem is represented by: DG = (N, A, Ej, WN) (91) where N is the set of operations plus the dummy start and finishing operations represented by the symbols 0 and *; A is the set of conjunctive arcs between every pair of operations on a job routing, between 0 and every first operation on a routing, and between every last operation on a routing and *; Ej is the set of disjunctive arcs between pairs of operations, O*j, that have to be processed on the machine at stage j (j=1,..,m); it also includes disjunctive arcs between 0 and O*j, between O*j and * and between 0 and * for all Ej; WN is the weight on nodes, represented by the processing time of related operations Oi j, WN(Oi j)=ti j. Figure 3 (top) shows an example of digraph for flowshop scheduling. Each operation O* j is connected with disjunctive arcs of Ej only, because the related operation can be processed by a machine at stage j. In general, a non-permutation flowshop problem is represented by n×m operations and an operation is connected with (n+1) disjunctive arcs. If no cycle is present in a conjunctive graph achieved by directing some disjunctive arcs to include all the operations, the conjunctive graph becomes acyclic and related sequences on machines are feasible schedules. In an acyclic conjunctive graph, the makespan of the feasible schedule is the length L(C) of the critical path, the longest path between the dummy start and finishing operations. Figure 3 (bottom) shows an example of acyclic conjunctive graph at an intermediate step, achieved by directing disjunctive arcs. Empty nodes are to be selected. The weighted path (full node) is obtained by adding processing times along the selected path and selecting the highest completion time of all arcs that end to it. The completion time of an operation is the longest path that connects dummy 0 and its node. 5. Native Non-Permutation Flowshop Scheduling heuristic The native non-permutation flowshop algorithm generates an acyclic digraph by visiting every operation one and only one time. Disjunctive arcs connecting any visited node are immediately directed towards the node. A fast O(n×m) approach to generate an acyclic digraph is proposed by Rossi and Dini (2007) considering the list scheduler heuristic for the flexible job-shop scheduling. The list scheduler heuristic is originally proposed by Giffler and Thompson (1960). The behavior of the list scheduling heuristic is to connect exactly one ending and one starting arc for every operation by directing (n×m) of the (n2×m) initial disjunctive arcs. The following heuristic designed for the native non-permutation flowshop scheduling modifies the list scheduling heuristic in order to generate a feasible schedule with completion times of all operations assigned to nodes. A Native Non-Permutation Flowshop Scheduling (native-NPFS) heuristic Input: a digraph DG = (N, A, EM, WN) O ← {Oi j  i=1,..,n; j=1,..,m} for each w = 1 to |Ο| do 1111.... Initialization of Candidate Nodes: build the allowed list ALw for current step w: ALw ← {Oi j ∈ O  Oi j-1 ∩ O = ∅} 2. Restriction: restriction of the allowed list by means of optimality criteria (i.e. active or non-delay schedule); let the candidate list CLw be the restricted allowed list 3. Initialization of Feasible Moves: mark as a feasible move each disjunctive arc (Oi’j, Oi j) of EM where Oi j ∈ CLw and Oi’j is the last operation of the sequence of resource j 4. Move Selection: select a feasible move (Oi’j, Oi j) of EM by directing the related disjunctive arc (Oi’j =dummy 0, if j = 1) 5. Arcs Removal: remove all the remaining disjunctive arcs connected to Oi’j, i.e. no other operation outside of Oi j can be immediately subsequent to Oi’j at stage j 6. Transferring weight: the weight ti j on the node is moved onto the related arc and onto the arc of the job routing (arc of A) that ends at Oi j; also, the completion time t(Oi j) = Si j + ti j is assigned as a mark inside the node Oi j, where the starting time of the operation is the maximum between the completion time of the previous operation in the routing and the previous operation in the sequence at stage j: Si j = max{t(Oi (j-1)), t(Oi’j)} 7. Updating Structures: update O by removing operation Oi j end for 9. Directing the remaining disjunctive arcs: i.e. arcs connected to the dummy operation * Output: the non-permutation schedule C The native-NPFS heuristics performs as many steps as the number of nodes (operations). The dummy operation 0 is the start of all machine sequences. For each node, a set of feasible moves is initialized. A feasible move is a disjunctive arc that connects the current node, already inserted in the machine sequence, to a next node that represents an operation not already scheduled. A feasible move to connect the related operations is selected by means of the heuristic desirability. This desirability is termed visibility (e.g. shortest processing or setup time, nearest due date etc.). Machine sequences are non-permutated because the native-NPFS heuristic allows a random selection of the feasible move and because the selected disjunctive arc can be directed in either the directions. After a feasible move is selected, all disjunctive arcs connected with the starting node are removed to prevent cycles. The weight of the ending node is updated with the maximum distance from the source, evaluated by the following two alternatively paths: i.) the path that crosses the previous node on the routing; ii.) the path that crosses the previous operation in the machine sequence. This guarantees that the weight on the dummy finishing operation is the longest path that starts from the dummy start operation, i.e. the L(C). 6. Proposed Ant Colony Optimization (ACO) algorithm Ant colony systems, a subset class of ACO, are an emerging class of biologically inspired research dealing with artificial or swarm intelligence: it exploits the experience of an ant colony as a model of self-organization in co- operative food retrieval (Dorigo and Gambardella, 1997). The main goal of the ACO mechanism is to generate optimal solutions by constructive schedules. The concept is similar to “divide et impera”, because the stigmergy progressively concentrates the search in a low number of very small promising regions. Differently to local search, this fact makes the algorithm intrinsically parallel and may take advantage of modern processors. This section describes an ant colony system (ACS), which makes use of the heuristic designed above by the digraph method and which fulfills all the problem constraints (2) to (8) to generate native non-permutation flowshop schedules. The proposed native-NPFS ACO based on the list scheduling heuristics described guides ants to select the most desirable move into nest-food path by using the function of visibility (as for native-NPFS heuristics). Pheromone amount laid on the disjunctive arcs mitigate the impact of heuristic desirability because it drives ants to select the optimal move within the set of feasible moves (step 4. of the native-NPFS heuristic). At the start, the pheromone is randomly deposited; consequently, the node selection is random as in pure list scheduling algorithms. As epochs evolve, the deposited pheromone drives the arc selection. The ant runs the nest-food path by a probabilistic selection of nodes according to the following properties: i) diversification in order to produce promising alternative paths; ii) intensification to select a node in the vicinity of the current best path C*. As soon as all the paths of the ants in the colony are generated, the best epoch ant deposits on the arcs of it path Cbe a further amount of pheromone proportional to the path length L(Cbe) and a pheromone decay routine is performed to prevent the stagnation in local optima solutions. The pheromone trail is the basic mechanism of communication among real ants and it is mimicked by the ant colony system in order to find the shortest path, connecting source and destination on a weighted graph, which represents the optimization problem. The following mechanisms have been implemented in the proposed ACO and are described in detail: � path generation, to transform the digraph in an acyclic conjunctive graph by a stochastic process based on the amount of pheromone; � candidate list construction, to select operations, not only to achieve feasible schedules, but also in order not to exceed the idle time more than a predefined value; � local search, to better improve the best found solution. Local and off-line pheromone update rules are mechanisms of, respectively, ant paths diversification and ant paths intensification. They are implemented by the standard procedures described in Dorigo and Gambardella (1997). 6.1. Digraph model The disjunctive graph is able to implement pheromone trails. The WDG = (N, A, Ej, WN, WE) (10) has the new component WE, which represents the weight on disjunctive arcs (pheromone amount). The weight on disjunctive arcs (Oi’j, Oi j) of Ej is represented by the pair WE(Oi’j,Oi j) = (τ[Oi’j, Oi j], τ[Oi j, Oi’j]). The first component array τ[Oi’j, Oi j] provides an index of desirability in order to add the feasible move [Oi’j, Oi j] to the list of jobs to be processed by machine j. The pheromone trail WE is a two-dimensional array of [m×n×(n+1)] elements (considering that the number of disjunctive arcs among all the n operations to be processed on machine j that can be replicated is [n×(n- 1)/2]). Hence, the internal shared memory of the proposed ant colony system is O(n 2×m). 6.2. Path generation To generate a feasible schedule Ca, each ant a visits every operation on WDG one and only one time in order to transform the digraph in an acyclic conjunctive graph. Path generation is a stochastic process where an ant starts from the dummy 0 and selects the next node from a subset of allowed operations, the candidate list CL. It uses the following transition probability rule of the pheromone trail as a function of both the heuristic function of desirability, η (the visibility function), and the amount of pheromone τ on the edge (Oi j, J), with J ∈ CL and such that                  = > ≤⋅∈ 0 0 )(η)(τminarg if if , , Z qqJ qq,, JOJO jijiCLO ji βα (11) The non-negative parameters α and β represent the intensity of respectively, the amount of pheromone and the visibility included in the transition probability function. The non-negative parameter q0 is the cutting exploration, a mechanism that restricts the selection of the next operation from the candidate list CL. If a random number q is higher than the cutting exploration parameter q0 (0 ≤ q0 ≤ 1), the candidate operation is selected examining all the candidate operations in probability, proportionally to pheromone amount α)(τ JO ,ji and visibility β)(η JO ,ji ; otherwise the most desirable operation is taken, i.e. the arc with the highest amount of pheromone and the highest visibility. The role of cutting exploration is explicitly splitting the search space in order to achieve a compromise between the probabilistic mechanism adopted for q ≤ q0 or the further intensification mechanism of exploring near the best path so far, which corresponds to an exploitation of the knowledge available about the problem. By tuning parameter q0 near 1, cutting exploration allows the activity of the system to concentrate on the best solutions (exploitation activity) instead of letting it explore constantly (exploration activity, achieved by tuning parameter q0 near 0). In fact, when q0 is close to 0, all the candidate solutions are examined in probability, whereas when q0 is close to 1, only the local optimal solution is selected by the second expression in (11). Scheduling problems have many local minima so higher randomization is used when the algorithm starts. A freezing function similar to the one proposed by Kumar et al. (2003) is considered. It progressively freezes the system by tuning q0 from 0 to 1, in order to favor the exploration in the initial part of the algorithm according to the expression )_(ln )(ln 0 epochsn epoch q = (12) where epoch is the current iteration and n_epochs is the total number of iterations of the ant colony system from the stability condition defined below. The heuristic function of desirability η is a critical component of ant colony systems. Generally, it is implemented by dispatching rules, as in Ying and Lin (2007). A comparison among a number of dispatching rules to implement the visibility function has been performed by Blum and Sampels (2004). In the proposed ACO the earliest starting time (EST) heuristic, i.e. the best heuristic tested by Blum and Sampels (2004), is used. 6.3. Candidate list construction The candidate list does not include all the operations that can be selected at a given construction step of the algorithm. In fact, it is well known that the optimal schedule is always an active schedule, i.e. a feasible schedule in which no operation could be started earlier without delaying some other operations or breaking a precedence constraint. Thus, the search space can be safely limited to the set of all active schedules. An important class of active schedules is the non- delay schedule: these are schedules in which no machine is kept idle when it could start processing some operations. As no all-optimal schedules are non-delay, the concept of parameterized non-delay schedules is used. This type of schedule consists of schedules in which no machine is kept idle for more than a predefined value δ if it could start processing some operations. As the minimum starting time of a candidate operation is jiALO S ji ∈ min (13) All the operations O* that can start if no machine is kept idle for more than a predefined value δ verify the condition ALOSS jiALOO ji ∈+≤ ∈ *,min* δ (14) A strategy that relaxes the expression (14) is considered by using the following parametric value of δ rf SS rf jiALOjiALO jiji ∈∈ − = minmax )(δ (15) where rf is the restricted factor. If the restriction is maximum, i.e. rf → +∞, the predefined value δ(rf) tends to zero and we obtain a non-delay schedule, i.e. CL=AL; on the contrary, if rf is set higher than 0, the property of non-delay schedule is relaxed; finally, if rf = 0 the candidate list does not differ from the allowed list, i.e. no restriction to AL occurs. To sum up, the following candidate list is used       − +≤∈= ∈∈ ∈ rf SS SSALOCL jiALOjiALO jiALOO jiji ji minmax min|* * (16) where a restricted factor rf = 3 from the cited literature is selected. 6.4. Local search When a path is generated, the solution is taken to its local optimum by a local search routine. The mechanism considered is a steepest descent algorithm where the current best solution is replaced with an improved solution included in the neighbor of current best. The performance of the local search depends on the neighborhood structure, and has been derived from the state-of-the-art local search for job-shop scheduling: the neighborhood structure proposed by Nowicki and Smutnicki (1996) for their fast tabu search algorithm. 6.5. Stability condition In iterative algorithms, the stop criterion is sometimes a fixed epoch number or computation time. Instead, to stop the algorithm we use a stability condition, corresponding to a fixed number of epochs (3000), with error reduction of at least one makespan time unit per epoch. 6.6. The native approach A metaheuristic algorithm is considered as a native approach if the initial solution is a feasible solution achieved by dispatching rules. Generally, a dispatching rule is selected within a complete set of known rules. Recently, the behavior of an algorithm to be native is considered in computation science to demonstrate the superiority of the single component algorithm of hybrid metaheuristic under study. A native-NPFS approach builds autonomously an initial feasible schedule, which includes different permutations of jobs between machines. Starting with a native (inner) initial sequence, the unbiased performance of the system can be evaluated, because they are not affected by the performance of (outer) heuristics. According to this definition of native, hybrid approaches cannot be considered native. A number of metaheuristics approaches are tailored on the digraph. A feasible native schedule can be built by a disjunctive graph approach. Biologically inspired general-purpose optimization algorithms are capable to deal with large job-size problems and with the exponential increase in search space with the number of machines and jobs. In this paper ant colony optimization is considered, because ACO uses a basic mechanism (diversification) to generate non- permutated sequences of jobs. The proposed digraph approach builds natively non-permutation sequences by the path generation mechanisms. In this stochastic process, each artificial ant selects probabilistically the next node (move selection) according to the amount of pheromone on the connecting arc (learned desirability). By design, non- permutation schedules are achieved by directing arcs differently at each stage. Cmax is evaluated from the weights on the critical path. At each epoch, as soon as all the paths of the ants in the colony are generated, the best ant (lowest Cmax) deposits on its arcs an amount of pheromone proportional to the path length (pheromone updating). A pheromone decay routine is also performed to prevent stagnation in local optima solutions (evaporation ρ = 0.12). The two inverse mechanisms are achieved by negative and positive pheromone deposition, respectively through the local update rule and off-line pheromone update rule. Diversification by the local update rule pushes towards permutated schedules and is the core mechanism to generate natively non-permutation solutions. 6.7. Implementation The following pseudo-code gives a high-level description of an Ant Colony System for Native Non-Permutation Flowshop Scheduling. Input: a weighted digraph WDG = (N, A, Ej, WN, WE) // Initialization for each disjunctive arc (Oi’j’,Oi j) of EA, deposit a small constant amount of pheromone WE(Oi’j’,Oi j) = (τ0, τ0) where 1 1 ,..,10 )(max − = =       ⋅⋅= ∑ n i jimj Otmnτ epoch ← 1; not_improve ← 0; // Main Loop while(not_improve < stability_condition) do // Epoch Loop for each ant a, a=1 to ps do // Path Generation Ca ← ∅; 1. O ← {Oi j  i=1,..,n; j=1,..,m}; 2. Initialization of Candidate Nodes: AL1 ← {Oi j i=1,..,n; j=1}; for each w = 1 to n×m do 3. Restriction: restrict ALw to the candidate list CLw by means of expression (16); 4. Initialization of Feasible Moves (i.e. the disjunctive arcs connected to operation of CLw); 5. Move Selection: select a feasible move (Oi’j, Oi j) of Ej, where Oi j ∈ CL, by means of the transition probability rules (11); directing the related disjunctive arc (Oi’j = dummy 0, if j = 1); 6. Arc Removing: remove all disjunctive arcs connected to Oi’j (i.e. no other operation can be immediately subsequent to Oi’j in the machine sequence); 7. Path length evaluation: the length L(Oi j) of the longest path that connects Oi j to the dummy 0 is evaluated by L(Oi j) = t(Oi j) + maxL(Oi’j), L(O(i-1) j) and is placed as a mark near the scheduled operation; 8. Local Updating: apply the local update rule to the arcs (Oi’j, Oi j) ∈ WE as in Dorigo and Gambardella (1997); 9. Update Allowed List: remove the scheduled operation from the allowed list ALw ← ALw ∪ {Oi (j+1)} / {Oi j}, if j ≤ m-1; ← ALw / {Oi j}, otherwise; end for 10. Directing the remaining disjunctive arcs (these arcs are connected to dummy *) 11. Local Search: Apply the local search with the neighbor structure from Nowicki and Smutnicki (1996) to Ca; 12. Best Evaluation: if [L(Ca) < L(Cbe] then [L(Cbe) ← L(Cbe) and Cbe ← Ca] end if end for Global Updating: Apply the global update rule as in Dorigo and Gambardella (1997); Best Ant Evaluation: if [L(Cbe) < L(C*)] then [L(C*) ← L(Cbe); C* ← Cbe and epoch ← 0] and not_improve ← 0; else epoch ++ and not_improve ++; end if end while Output: C* The main parameters of the two ACO approaches are listed in Table 1. The parameters considered in other ant colony systems for NPFS are also listed. Table 1. Parameters from different ACO approaches, including the proposed native Non-Permutation Flowshop Scheduling ACO. Parameter Yagmahan and Yenisey (2010) Ying and Lin (2007) native-NPFS ACO problem non-native native native population size 20 5 8 candidate list 15 - self complied (parameterized non-delay) α 2 0.1 2 β 0.5 2.0 0.3 local search Adjacent Pair-wise Interchange - Borderline Critical Path Blocks Interchange q0 0.9 0.1 self complied (freezing function) ρ 0.2 0.9 0.12 τ0 [ ] 1−× NEHCn [ ] 1−×× UBmn 1 1 ,..,1 )(max − = =         ×× ∑ n i jimj Otmn η SPIRIT rule Multi-Heuristic Desirability Earliest Finishing Time stop condition 1000 epochs 5000 epochs stability condition stability condition (n_epochs) - - 3000 epochs firmware platform 2 GHz Intel® Pentium® M760, RAM 1024 MB DDR-2 1.5 GHz Intel® Pentium® 4 RAM N.A. 3 GHz 32 bit Intel® Pentium® 4 RAM 2 GB 7. Computation experiments 7.1. Benchmarks Benchmark problem sets allows researchers to compare their proposed algorithms with those of other researchers using an identical test problem set. Benchmark instances are arrays bn×m of processing times of each job on all m machines ordered by routing originally generated random. Each benchmark can be synthesized by its makespan, i.e. the processing time obtained as the addition of the n×m operations of each job on all machines ordered by routing (plus idle times). Two reference makespan values are considered by researchers: non trivial lower (LB) and upper bound (UB). The lower (upper) bound is the maximum (minimum) known theoretical minimum (maximum) attainable makespan. A lower bound (LB) can be obtained for each instance by relaxing the capacity constraints on all but one machine, and solving for optimality the resulting single machine problem of minimizing makespan with release and delivery times (Pinedo, 1995). The upper bound can be reduced by improved solutions. If it coincides with the lower bound, the optimum has been reached. Table 2. Benchmark sets considered. Taillard (1993) Demirkol et al. (1998) parameter (index) job number 20, 50, 100, 200, 500 20, 30, 40, 50 n (i) machine number 5, 10, 20 15, 20 m (r) processing times random[1,99] random[1,200] n to m ratio 1 – 100 1 – 3.3 # operations (n×m) 100 – 10,000 300 – 1,000 N # sets (n×m combinations) 12 8 # instances per set 10 10 # instances considered per set 10 5 # instances total 120 40 # instances (individual data) considered by other authors for comparison 28, Table 3 40, Table 4 # sets (aggregated data) considered by other authors for comparison 3, Table 5 8, Table 5 competitors approaches Rajendran (1995), Ravindran et al. (2005), Yagmahan and Yenisey (2010) Demirkol et al. (1998), Ying and Lin (2007), Lin and Ying (2009) most recent LBs and UBs Vaessens (1996), mirrored in Lanzetta (2012) + competitors above see competitors above The proposed native heuristic has been tested by computation experiments on well-known benchmark problem sets established by Taillard (1993) and by Demirkol et al. (1998) characterized in Table 2. These combinations yield a problem set that is not based on a specific application. All jobs are available at time zero, and generated by a discrete uniform distribution. The most used flowshop benchmark set is by Taillard and is listed in Table 3. All Taillard instances in each set are considered by authors, instead Demirkol et al. ranked the instances in decreasing order of percentage gap, defined in (17), between the upper and lower bounds for each combination of n and m to obtain a more compact and challenging set of test problems. Only the first five instances for each combination were finally presented. Thus, the test bed comprised a total of 40 test instances, which are listed in Table 4. Vaessens (1996) determined the lower bounds for non-permutation for Taillard benchmarks with a method similar to the mentioned Pinedo (1995). As the size and complexity of the instances make exact solutions impractical, Demirkol et al. (1998) solved each instance by five different constructive heuristics and three versions of the shifting bottleneck procedure and reported the highest makespan value obtained as a lower bound in each instance. All algorithms were run on Unix with 50 MHz SUN SPARC server 1000 Model 1104. The upper bound (UB) for each instance was the best solution found by any of the algorithms. The computation time was recorded by the method that provided the best solution. If more than one algorithm found the best solution, then the algorithm with the shortest computation time was selected. Before this work, it seems that no author has compared the performance of their proposed approach on both sets, so the results of current approach are presented for each set separately. 7.2. Performance measures The proposed native-NPFS ACO has been run 10 times with the parameters in Table 1. The best and average solutions have been reported in Table 3 and in Table 4. Metrics for algorithm performance are: 1) the number of benchmark instances where the minimum makespan xyta best C 0 of the Taillard benchmark instance ta0xy, y=1,..,9; x=0,1,2 is found with respect to the other native approaches; 2) the n×m aggregate relative distances from the upper bound or the best known solution evaluated by the relative distance between UB and 2,1,0, 9 1 0 =∑ = xC y xyta best for each array n×m of the 5 Taillard benchmarks: 2,1,0,100% 9 1 0 9 1 0 9 1 0 = − ⋅= ∑ ∑∑ = == x LB LBC gap y xyta y xyta y xyta best best (17) 2,1,0,100% 9 1 0 9 1 0 9 1 0 = − ⋅= ∑ ∑∑ = == x LB LBC gap y xyta y xyta y xyta average average (18) Analogous measure is evaluated for each array n×m of the 10 Demirkol benchmarks. Table 3. Benchmark problems by Taillard (1993), state-of-the-art solutions and results of computation experiments for one non-native and three native NPFS approaches. Yagmahan and Yenisey (2010) is an ant colony system, Rajendran (1995) and Ravindran et al. (2005) are two native heuristics and native-NPFS ACO is the ant colony system proposed in this paper. In bold the best performing native algorithm. Approach Benchmark Non-native Approach Native Approach Algorithm � Yagmahan and Yenisey (2010) Rajendran (1995) Ravindran et al. (2005) native-NPFS ACO Instance � UB/Best Known Cbest (10 runs) Cbest Cbest Cbest (10 runs) Caverage ta001 1278 1297 1359 1297 1290 1293.1 ta002 1358 1383 1378 1373 1389 1389.0 ta003 1073 1203 1230 1206 1100 1112.5 ta004 1292 1377 1393 1402 1344 1352.2 ta005 1231 1311 1307 1334 1250 1258.7 ta006 1193 1245 1282 1238 1217 1224.3 ta007 1234 1303 1387 1322 1258 1259.6 ta008 1199 1265 1344 1287 1235 1242.5 ta009 1210 1303 1335 1307 1258 1275.3 20x5 ta010 1103 1179 1191 1195 1127 1145.5 ta011 1560 1681 1711 1774 1693 1729.6 ta012 1644 1749 1916 1791 1785 1799.5 ta013 1486 1554 1617 1643 1583 1596.5 ta014 1368 1490 1533 1531 1452 1478.3 ta015 1413 1455 1588 1557 1516 1526.7 ta016 1369 1564 1565 1612 1445 1468.3 ta017 1428 1590 1622 1594 1524 1544.5 ta018 1527 1595 1800 1631 1650 1663.8 ta019 1586 1689 1717 1769 1659 1681.1 20x10 ta020 1559 1719 1831 1744 1670 1677.7 ta021 2293 2428 2610 2491 2396 2414.9 ta022 2092 2281 2301 2491 2225 2239.5 ta023 2313 2515 2411 2422 2446 2464.5 ta024 2223 2299 2471 2567 2346 2360.8 ta025 2291 2473 2427 2420 2439 2460.5 ta026 2221 2339 2466 2557 2331 2346.5 ta027 2267 2378 2174 2448 2428 2454.0 20x20 ta028 2183 2418 2418 2464 2321 2345.6 Table 4. Benchmark problems by Demirkol et al. (1998), state-of-the-art solutions and results of computation experiments for one non-native and three native NPFS approaches. Lin and Ying (2009) is a hybrid simulated annealing – tabu search, Demirkol et al. (1998) is a shifting bottleneck heuristic, Ying and Lin (2007) is an ant colony system and native-NPFS ACO is the ant colony system proposed in this paper. In bold the best performing native algorithm. Approach Benchmark Non-native Approach Native Approach Algorithm � Lin and Ying (2009) Demirkol et al. (1998) Ying and Lin (2007) native-NPFS ACO Instance � LB Cbest (5 runs) Cbest Cbest (5 runs) Cbest (10 runs) Caverage flcmax_20_15_3 3354 3873 4437 4420 4047 4113.4 flcmax_20_15_6 3168 3761 4144 4044 3950 3977.5 flcmax_20_15_4 2997 3518 3779 3786 3692 3730.6 flcmax_20_15_10 3420 4051 4302 4265 4176 4221.7 20x15 flcmax_20_15_5 3494 3913 4373 4310 4097 4124.9 flcmax_20_20_1 3776 4525 4821 4819 4790 4826.9 flcmax_20_20_3 3758 4435 4779 4723 4694 4715.7 flcmax_20_20_9 3902 4527 4944 4922 4720 4775.4 flcmax_20_20_2 3881 4499 4886 4847 4731 4781.3 20x20 flcmax_20_20_10 3823 4361 4717 4715 4554 4607.6 flcmax_30_15_3 4020 4568 5226 5210 4927 5032.2 flcmax_30_15_4 4080 4649 5304 5284 5033 5092.4 flcmax_30_15_9 4022 4568 5079 5075 4912 4968.6 flcmax_30_15_8 4490 4836 5605 5593 5220 5320.2 30x15 flcmax_30_15_6 4184 4761 5147 5149 5097 5158.5 flcmax_30_20_3 4806 5376 6183 5987 5794 5846.9 flcmax_30_20_1 4772 5698 6037 5989 6179 6221.8 flcmax_30_20_6 5004 5752 6241 6195 6039 6133.9 flcmax_30_20_10 4899 5464 6095 5923 5888 5967.7 30x20 flcmax_30_20_2 4757 5369 5822 5840 5842 5886.1 flcmax_40_15_5 5560 5958 6986 6972 6521 6594.1 flcmax_40_15_9 5119 5692 6351 6310 6244 6303.3 flcmax_40_15_2 5290 5877 6506 6532 6302 6395.6 flcmax_40_15_10 5596 5896 6845 6712 6413 6445.2 40x15 flcmax_40_15_8 5576 6054 6783 6771 6526 6611.7 flcmax_40_20_3 5693 6508 7154 7132 7208 7274.5 flcmax_40_20_9 5998 6676 7528 7496 7388 7484.7 flcmax_40_20_6 5990 6798 7469 7476 7455 7553.1 flcmax_40_20_7 6170 6766 7608 7588 7405 7473.5 40x20 flcmax_40_20_5 6011 6508 7219 7217 7326 7399.4 flcmax_50_15_6 6290 6836 7673 7631 7559 7606.8 flcmax_50_15_5 6355 6672 7679 7496 7317 7368.4 flcmax_50_15_1 6198 6580 7416 7402 7205 7303.8 flcmax_50_15_8 6312 6799 7548 7558 7348 7468.7 50x15 flcmax_50_15_2 6531 6954 7750 7712 7547 7644.8 flcmax_50_20_2 6740 7682 8838 8836 8436 8684.4 flcmax_50_20_1 6736 7313 8539 8521 8064 8189.7 flcmax_50_20_7 6756 7622 8417 8425 8370 8526 flcmax_50_20_8 6897 7480 8590 8536 8430 8509.2 50x20 flcmax_50_20_4 6830 7726 8493 8502 8538 8625.1 Table 5. Aggregated results on Taillard and Demirkol benchmarks of size n×m. Yagmahan and Yenisey (2010) is an ant colony system, Lin and Ying (2009) is a hybrid simulated annealing – tabu search; Rajendran (1995) and Ravindran et al. (2005) are two native heuristics, Demirkol et al. (1998) is a shifting bottleneck heuristic; Ying and Lin (2007) is an ant colony system. Native-NPFS ACO is the ant colony system proposed in this paper. In bold the best performing native NPFS approach. Approach Non-native Approach Native Approach Algorithm � Aggregate Instance � Yagmahan and Yenisey (2010) Lin and Ying, 2009 Rajendran (1995) Ravindran et al. (2005) Demirkol et al. (1998) Ying and Lin (2007) native-NPFS ACO 20×5 Taillard 6.49 - 8.50 5.71 - - 2.44 20×10 Taillard 11.42 - 13.12 7.67 - - 6.94 20×15 Taillard 11.06 - 7.80 6.98 - - 5.87 20×15 Demirkol - 0.0 - - 10.04 8.94 4.43 20×20 Demirkol - 0.0 - - 8.05 7.51 5.11 30×15 Demirkol - 0.0 - - 12.74 12.53 7.73 30×20 Demirkol - 0.0 - - 9.83 8.23 7.53 40×15 Demirkol - 0.0 - - 13.55 12.96 8.58 40×20 Demirkol - 0.0 - - 11.19 10.98 10.60 50×15 Demirkol - 0.0 - - 12.48 11.70 9.26 50×20 Demirkol - 0.0 - - 13.36 13.21 10.62 7.3. Results and discussion Based on the results presented in Table 3 to Table 5 using expression (17), the proposed ACO performs better than others native approaches examined for both Taillard and Demirkol datasets. Table 3 shows that only few benchmarks are left to the competitors approaches: three to the heuristic of Ravindran et al. (2005) and one to the heuristic of Rajendran (1995). In the other 24 of 28 Taillard sets, it is incumbent. The averagegap% calculated as in (18) for all sets ranges within 2.2%. Table 4 shows that the proposed native approach performs better in 35 Demirkol benchmark problems leaving only five to the competitors approaches, particularly three sets to Ying and Lin (2007) and two sets to Demirkol et al. (1998). Here, the averagegap% ranges like Taillard benchmark except for one problem that exceeds 2.9%. Table 4 also shows the difference of performance from the non-native approaches considered. Due to a lack of tight upper bounds similar to those evaluated by Vaessens (1996), the upper bound of Demirkol benchmark coincide with the best known solution. Thus, Yagmahan and Yenisey (2010) is not close to the upper bound whilst Lin and Ying (2009) is right the current best-known solution. Also, native-NPFS ACO performs better in 75% of Taillard benchmark problems than the current best non-native approach, the ACO with pheromone trails initialized by the NEH heuristic, from Yagmahan and Yenisey (2010), making the proposed algorithm a very competitive ant colony system. The non-native hybrid SA – tabu search from Lin and Ying (2009) performs better than other native approaches. Considering that the hybrid metaheuristic of Zobolas et al. (2009) is the state-of-art permutation flowshop (PFS) approach, hybridization is more performing than the individual components. This was also confirmed by Rossi and Boschi (2009) by a Kruskall–Wallis H test statistics on a hybrid ACO – genetic algorithm compared to the two individual components. However, hybridization is intrinsically opposite of native approach because it is the result of two or more native approaches combined. Possible reasons for improvements over existing native ACO approaches are inferred from the analysis of Table 1, from self complied parameterized non-delay candidate list and freezing function, and stability condition and from other parameters, closer to Yagmahan and Yenisey (2010), which is non-native, than to Ying and Lin (2007). Table 5 shows the good performance over a wide range of problems has defined by the two benchmarks in Table 2 in terms of size (# operations) and configuration (n to m ratio). Individual and aggregated results on all Taillard sets, not listed in Table 4 because not available from other authors, are available from Lanzetta (2012), where benchmarks are also mirrored, for future benchmarking. 8. Conclusion Among the main merits of current work is modeling the scheduling of a flow line with buffers as non permutation flowshop, which has also been characterized by a MILP. Alternative approaches and implications have been reviewed in order to map existing metaheuristics approaches to the examined manufacturing problem. The digraph approach described in the paper has the additional merit of being a powerful design tool for native non-permutation algorithms. ACO has been shown to fulfill such design requirements by the pheromone mechanism. The proposed ACO is natively non-permutation as opposed to other authors who apply a local search to permutation solutions. The few existing approaches have been compared using well-known benchmarks. This proposed approach shows the best performance in non-permutation flowshop configuration, particularly on larger instances (available as additional material) and is very close to state-of-the-art metaheuristics, also for non-native. Computation experiments have shown promising performance of such general-purpose optimization tool, regardless of the problem complexity increase in the examined range, from 100 to 10,000 operations and job to machine ratios from 1 to 100. References Alaykyran K., Engin O. & Doyen, A. (2007). Using ant colony optimization to solve hybrid flowshop scheduling problems. International Journal of Advanced Manufacturing Technology, 35(5–6), 541–550. Aloulou, M.A. & Artigues, C. (2010). Flexible solutions in disjunctive scheduling: General formulation and study of the flow-shop case. Computers & Operations Research, 37(5), 890-898. Bonabeau. E., Dorigo, M., & Theraulaz, G. (2000). Inspiration for optimization from social insect behaviour. Nature, 406, 39–42. Brucker, P., Heitmann S. & Hurink, J. (2003). Flow-shop problems with intermediate buffers. OR Spectrum, 25, 549– 574. Colledani, M. & Tolio, T. (2005). A Decomposition Method to Support the Configuration / Reconfiguration of Production Systems, CIRP Annals - Manufacturing Technology, 54(1), 441-444. Demirkol, E., Mehta, S., & Uzsoy, R. (1998). Benchmarks for shop scheduling problems. European Journal of Operational Research, 109, 137–141. Dorigo, M. & Gambardella, L.M. (1997). Ant colonies for the travelling salesman problem. Biosystems, 43(2), 73-81. Garey, M.R., Johnson, D.S. & Sethi, R. (1976). The complexity of flowshop and jobshop scheduling. Mathematics of Operations Research, 1(2), 117–129. Giffler, D. & Thompson, G.L. (1960). Algorithms for solving production scheduling problems. Operation Research, 8, 487–503. Jain, A.S. & Meeran, S. (2002). A multi-level hybrid framework applied to the general flow-shop scheduling problem. Computers & Operations Research, 29, 1873–1901. Khalili, M., Tavakkoli-Moghaddam, R. (2012). A multi-objective electromagnetism algorithm for a bi-objective flowshop scheduling problem. Journal of Manufacturing Systems, 31(2), 232-239. Koulamas, C. (1998). A new constructive heuristic for the flowshop scheduling problem. European Journal of Operational Research, 105, 66–71. Kumar R., Tiwari M.K. & Shankar R. (2003). Scheduling of flexible manufacturing system: an ant colony optimization approach. Journal of Engineering Manufacture, Part B, 217, 1443–53. Lanzetta, M. (2012) http://www.ing.unipi.it/lanzetta/aco, last acc. 2012. Leisten, R. (1990). Flowshop sequencing problems with limited buffer storage. International Journal of Production Research, 28(11), 2085-2100. Li, B., Wu S., Yang, J., Zhou, Y. & Du M. (2012). A three-fold approach for job shop problems: A divide-and-integrate strategy with immune algorithm. Journal of Manufacturing Systems, 31(2), 195-203. Liao, C.J., Liao, L.M. & C.T. (2006). Tseng, A performance evaluation of permutation vs. non-permutation schedules in a flowshop. International Journal of Production Research, 44, 4297–4309. Liao, L.-M. & Huang, C.-J. (2010). Tabu search for non-permutation flowshop scheduling problem with minimizing total tardiness. Applied Mathematics and Computation, 217, 557–567. Lin, S.W. & Ying, K.C. (2008). Applying a hybrid simulated annealing and tabu search approach to non-permutation flowshop scheduling problems. International Journal of Production Research, 46, 1–14. Lin, S.-W. & Ying, K.-C. (2009). Applying a hybrid simulated annealing and tabu search approach to non-permutation flowshop scheduling problems. International Journal of Production Research, 47(5), 1411–1424. Mehravaran, Y. & Logendran, R. (2012). Non-permutation flowshop scheduling in a supply chain with sequence- dependent setup times. International Journal of Production Economics, 135(2), 953-963. Nawaz, M., Enscore Jr., E.E. & Ham, I. (1983). A heuristic algorithm for the m-machine, n-job flow-shop sequencing problem. OMEGA, The International Journal of Management Science, 11(1), 91–95. Nowicki, E., Smutnicki, C. (1996). A fast taboo search algorithm for the job-shop problem. Management Science, 42(6), 797–813. Park, Y. B., Pegden, C. D., Enscore, E.E. (1984). A survey and evaluation of static flowshop scheduling heuristics. International Journal of Production Research, 22, 127-141. Pinedo, M. (1995). Scheduling: theory, algorithms and system. Prentice-Hall, New Jersey. Potts, C.N., Shmoys, D.B. & Williamson, D.P. (1991). Permutation vs. non-permutation flow shop schedules. Operations Research Letters, 10, 281–284. Pugazhendhi, S., Thiagarajan, S., Rajendran, C. & Anantharaman, N. (2003). Performance enhancement by using non- permutation schedules in flowline-based manufacturing systems. Computers and Industrial Engineering, 44, 133–157. Rajendran C & Ziegler H. (2004). Ant-colony algorithms for permutation flowshop scheduling to minimize makespan/total flowtime of jobs. European Journal of Operational Research, 155, 426–38. Rajendran, C. (1995). Heuristics for scheduling in flowshop with multiple objectives. European Journal of Operational Research, 82, 540–555. Ravindran, D., Noorul Haq, A., Selvakuar, S.J., & Sivaraman, R. (2005). Flow shop scheduling with multiple objective of minimizing makespan and total flow time. International Journal of Advanced Manufacturing Technology, 25, 1007–1012. Ribas, I., Leisten, R. & Framinan, J.M. (2010). Review and classification of hybrid flowshop scheduling problems from a production system and a solutions procedure perspective. Computers & Operations Research, 37, 1439–1454. Rossi A. & Dini G. (2007). Flexible job-shop scheduling with routing flexibility and separable setup time using ant colony optimisation method. Robotics & Computer Integrated Manufacturing, 23, 503–516. Rossi, A. & Boschi E. (2009). A hybrid heuristic to solve the parallel machines job-shop scheduling problem. Advances in Engineering Software, 40(2), 118-127. Rossi, A., Puppato, A., Lanzetta, M. (2012) Heuristics for Scheduling a Two-stage Hybrid Flow Shop with Parallel Batching Machines: an Application on Hospital Sterilization Plant. International Journal of Production Research, doi:10.1080/00207543.2012.737942. Ruiz, R. & Maroto, C. (2005). A comprehensive review and evaluation of permutation flowshop heuristics. European Journal of Operational Research, 165, 479–494. Ruiz, R., Maroto, C. & Alcaraz, J. (2006). Two new robust genetic algorithms for the flowshop scheduling problem. Omega, 34(5), 461-476. Taillard, E. (1990). Some efficient heuristic methods for the flow shop sequencing problem. European Journal of Operational Research, 47(1), 65-74. Taillard, E. (1993). Benchmarks for basic scheduling problems. European Journal of Operational Research, 64, 278– 285. Tandon, M., Cummings, P.T., & LeVan, M.D. (1991). Flowshop sequencing with non-permutation schedules. Computers and Industrial Engineering, 15(8), 601–607. Tasgetiren, M.F., Liang Y.-C., Sevkli, M., Gencyilmaz, G. (2007). A particle swarm optimization algorithm for makespan and total flowtime minimization in the permutation flowshop sequencing problem. European Journal of Operational Research, 177(3), 1930-1947. Vaessens, R. J. M., http://www.mathematik.uni-osnabrueck.de/research/OR/fsbuffer/taillard2.txt, 1996, last acc. 2012. Weng, M.X. (2000). Scheduling flow-shops with limited buffer spaces. In J. A. Joines, R. R. Barton, K. Kang, & P. A. Fishwick (Eds.), Proceedings of the 2000 Winter Simulation Conference. 1359–1363. Werner, F. & Winkler, A. (1995). Insertion techniques for the heuristic solution of the job-shop problem. Discrete Applied Mathematics, 58(2), 191–211. Yagmahan, B. & Yenisey, M.M. (2010). A multi-objective ant colony system algorithm for flow shop scheduling problem. Expert Systems with Applications, 37(2), 1361-1368. Ying, K.-C. & Liao, C.-J. (2004). An ant colony system for permutation flow-shop sequencing. Computers & Operations Research, 31, 791–801. Ying, K.-C. & Lin, S.-W. (2007). Multi-heuristic desirability ant colony system heuristic for non-permutation flowshop scheduling problems. International Journal of Advanced Manufacturing Technology, 33, 793–802. Ying, K.-C. (2008). Solving non-permutation flowshop scheduling problems by an effective iterated greedy heuristic. International Journal of Advanced Manufacturing Technology, 38, 348–354. Ying, K.-C., Gupta, J. N.D., Lin S.-W. & Lee, Z.-J. (2010). Permutation and non-permutation schedules for the flowline manufacturing cell with sequence dependent family setups. International Journal of Production Research, 48(8), 2169–2184. Zobolas, G.I.. Tarantilis, C.D., Ioannou, G. (2009). Minimizing makespan in permutation flow shop scheduling problems using a hybrid metaheuristic algorithm. Computers & Operations Research, 36(4), 1249-1267. 
work_ykea5tf5gbc27c57bssvjhynl4	----	Consistency maintenance for evolving feature models Consistency maintenance for evolving feature models Jianmei Guo a,⇑, Yinglin Wang a, Pablo Trinidad b, David Benavides b a Department of Computer Science and Engineering, Shanghai Jiao Tong University, 800 Dong Chuan Road, Minhang, Shanghai 200240, China b Department of Languages and Computer Systems, University of Seville, Av. Reina Mercedes s/n, 41012 Seville, Spain Keywords: Software product lines Feature models Evolution Consistency maintenance Ontology Semantics ⇑ Corresponding author. Tel.: +86 21 34204415; fax E-mail addresses: guojianmei@sjtu.edu.cn (J. Guo), ptrinidad@us.es (P. Trinidad), benavides@us.es (D. Benav a b s t r a c t Software product line (SPL) techniques handle the construction of customized systems. One of the most common representations of the decisions a customer can make in SPLs is feature models (FMs). An FM represents the relationships among common and variable features in an SPL. Features are a representa- tion of the characteristics in a system that are relevant to customers. FMs are subject to change since the set of features and their relationships can change along an SPL life- cycle. Due to this evolution, the consistency of FMs may be compromised. There exist some approaches to detect and explain inconsistencies in FMs, however this process can take a long time for large FMs. In this paper we present a complementary approach to dealing with inconsistencies in FM evolution scenarios that improves the performance for existing approaches reducing the impact of change to the smallest part of an FM that changes. To achieve our goal, we formalize FMs from an ontological perspec- tive and define constraints that must be satisfied in FMs to be consistent. We define a set of primitive operations that modify FMs and which are responsible for the FM evolution, analyzing their impact on the FM consistency. We propose a set of predefined strategies to keep the consistency for error-prone operations. As a proof-of-concept we present the results of our experiments, where we check for the effectiveness and efficiency of our approach in FMs with thousands of features. Although our approach is limited by the kinds of consistency constraints and the primitive operations we define, the experiments present a sig- nificant improvement in performance results in those cases where they are applicable. 1. Introduction Software product line (SPL) engineering has emerged as one of the most promising software development paradigms for reducing development costs, enhancing quality, and shortening time to mar- ket (Clements & Northrop, 2001; Pohl, Bockle, & van der Linden, 2005; Sugumaran, Park, & Kang, 2006). An SPL is ‘‘a set of soft- ware-intensive systems that share a common, managed set of fea- tures satisfying the specific needs of a particular market segment or mission and that are developed from a common set of core assets in a prescribed way’’ (Clements & Northrop, 2001). Features are essential abstractions of product characteristics relevant to cus- tomers and are typically increments in product functionality (Benavides, Segura, & Cortes, 2010; Kang, Lee, & Donohoe, 2002). Every product or system of an SPL is represented by a unique com- bination of features. All the products in an SPL are usually captured in feature models (FMs) which describe the commonalities and vari- abilities of systems in terms of features (Kang, Cohen, Hess, Novak, & Peterson, 1990; Kang et al., 2002). An FM is a tree-like structure : +86 21 34204728. ylwang@sjtu.edu.cn (Y. Wang), ides). that contains the relationships among features in a hierarchical manner. Relationships can be of different kinds to remark which are the choices a customer can make to build a customized product. As any other software systems, SPLs are subject to changes and evolution along their lifecycle. Those changes can affect FMs (Sugumaran et al., 2006). Even small changes to an FM could unin- tentionally break its consistency (Guo & Wang, 2010; Thum, Ba- tory, & Kastner, 2009). Here, the consistency of an FM means that it remains well-formed (syntactic consistency) and it defines at least a valid product (semantic consistency). For example, the removal of a single feature from an FM, although could be valid from the point of view of FM syntax, could invalidate other related features, which must also be removed to make the resulting FM consistent. Incon- sistency usually comes from contradictory constraints which im- pede producing any valid product (von der MaBen & Lichter, 2004). Consistent FMs are needed to any further steps in SPL engi- neering such as verifying product derivation (Lutz, 2008) or check- ing for the consistency of product requirements (Lauenroth & Pohl, 2008). Therefore, guaranteeing the FM consistency is a mandatory task in SPL development. Many approaches have been proposed to automate the detec- tion of inconsistencies in FMs (Benavides et al., 2010). They mostly http://dx.doi.org/10.1016/j.eswa.2011.10.014 mailto:guojianmei@sjtu.edu.cn mailto:ylwang@sjtu.edu.cn mailto:ptrinidad@us.es mailto:benavides@us.es http://dx.doi.org/10.1016/j.eswa.2011.10.014 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa use SAT (Batory, 2005; Thum et al., 2009), BDD (Czarnecki & Wasowski, 2007), or CSP solvers (Benavides, Martin-Arroyo, & Cortes, 2005; Trinidad, Benavides, Duran, Ruiz-Cortes, & Toro, 2008) to automate the checking process. However, these ap- proaches suffer from an NP-hard problem of feature combinatorics and take a long time to perform with large FMs (Batory, Benavides, & Ruiz-Cortes, 2006). Reports from industry have shown that prac- tical FMs could have hundreds or thousands of features (Loesch & Ploedereder, 2007; Steger et al., 2004), so existing approaches can take hours or days to detect inconsistencies. Moreover the information obtained from these detection mech- anisms hardly assists domain analysts to resolve inconsistencies in FMs. Existing approaches delegate the reparation of inconsistent FMs to domain analysts who must use their experience to find the best way to repair those FMs. The impact of performance and manual reparation gets worse since FMs frequently change during their evolution processes and the consistency of the resulting FMs has to be checked after every change. This paper approaches the problem of consistency maintenance in FMs focusing on the changes since last version of an FM rather than checking the overall consistency of the resulting FM. We as- sume that the initial FM is consistent and study if a requested change affects the consistency or not. In case an inconsistency is de- tected, a set of additional operations are executed to restore the consistency of the FM. For example, the removal of a single feature X from an FM initially only affects those features that are directly connected to it. To restore the consistency, the surrounding rela- tions of feature X are removed. What if the removal of a relationship generates a new inconsistency? In the worst case, the operation derivation could propagate to the whole FM; but in most cases it only affects a limited range, that is where the main improvement in performance comes. Our approach relies on ontology, which is a formal and explicit specification of a shared conceptualization of a domain of interest (Gruber, 1993). In our case we formalize FMs in Section 3, defining the primitive elements of FMs and the syntactical and semantic con- sistency constraints as the well-formedness rules of FMs. From this formalization, in Section 4 we obtain a set of primitive operations (Guo & Wang, 2010) which can represent any modification of an FM. In Section 5 we apply and extend techniques from ontology evolution (Haase & Stojanovic, 2005; Stojanovic, 2004) to propose a systematical approach to consistency maintenance for evolving FMs. A dependency matrix, indicating the cause and effect relation- ships between changes, is built for supporting the derivation of additional operations from the requested change. Then we analyze the possible evolution strategies for all the primitive operations on FMs and propose a sequence of interdependent operations derived from the requested change to produce a unique consistent FM. To demonstrate the realization of our approach, Section 6 pre- sents the implementation of our approach based on FeatureIDE.1 Section 7 evaluates our approach by experiments on randomly gen- erated FMs with thousands of features. Section 9 briefly analyses the pros and cons of our approach and presents some future extensions of our work. In order to introduce the readers in the context of our work, we complement our work with a brief definition of FMs in Section 2 and a discussion about the related work in Section 8. 2. Feature models background In 1990, Kang et al. (1990) first proposed the original FMs (a.k.a. FODA FMs). An FM is organized hierarchically and is graphically de- picted as an AND-OR feature diagram (Kang et al., 1990). Cross-tree 1 http://wwwiti.cs.uni-magdeburg.de/iti_db/research/featureide. constraints are used to represent non-hierarchical composition rules comprising mutual dependency (requires) and mutual exclu- sion (excludes) relationships (Kang et al., 1990). Czarnecki, Helsen, and Eisenecker (2005) proposed cardinality-based FMs where cardinalities (a.k.a. multiplicities) were introduced. Batory (2005) and Thum et al. (2009) distinguished among terminal (or concrete) and non-terminal (or compound or abstract) features. By integrating former definitions of FMs (Batory, 2005; Czarnecki et al., 2005; Kang et al., 1990; Thum et al., 2009), we adopt the notation as shown in Fig. 1, which is a partial FM for the Home Inte- gration Systems (HIS) SPL inspired from (Benavides et al., 2005; Kang et al., 2002). An FM is a tree of features. Every node in the tree has one parent except the root feature (‘r: HIS’). A terminal feature (e.g., ‘f4’) is a leaf and a non-terminal feature (e.g., ‘f1’) is an interior node of a feature diagram (Batory, 2005; Thum et al., 2009). Connec- tions between a feature and its group of children are classified as And- (e.g., ‘f1’, ‘f2’, and ‘f3’), Or- (e.g., ‘f10’ and ‘f11’), and Alterna- tive-groups (e.g., ‘f12’, ‘f13’, and ‘f14’). The members of And-groups can be either mandatory (e.g. ‘f1’) or optional (e.g. ‘f3’). Or-groups and Alternative-groups have their own cardinalities (Czarnecki & Wasowski, 2007). Cross-tree constraints comprise requires and excludes relationships (Kang et al., 1990), e.g., ‘f4 requires f7’. Table 1 summarizes the semantics of FMs in propositional for- mulas. P represents a non-terminal feature and C1, . . ., Cn are its child features. If the child features forms an And-group, then M # {1, . . ., n} denotes the mandatory features by their index. If a feature is selected, so too is its parent. If the parent is selected, all of its mandatory children of an And-group are selected; in Or-groups, at least one child must be selected, and in Alterna- tive-groups, exactly one child is selected. Using the rules given in Table 1, an FM can be easily translated into a propositional formula with a variable for each feature. 3. Ontology-based formalization and consistency constraints 3.1. An ontology-based formalization of FMs The representational primitives defined in ontology (Gruber, 2008) are typically concepts (classes) and properties. Each property must have at least one domain concept, while its range may either be a literal (attributes), or a set of at least one concept (relations). The definitions of the representational primitives include informa- tion about their meaning and constraints on their logically consis- tent application, which makes ontology work at the semantic level. Based on the ontology structure and the application context of FMs, we formalize FMs as follows. Definition 1. An FM is defined as a 5-tuple: FM ¼ðC; R; A; Domain; RangeÞ where: � C (–;) is the set of concepts in the FM. C = F [ FG. F (–;) is the set of features in the FM. FG (–;) is the set of feature groups. FG = FGAN [ FGAL [ FGOR. FGAN is the set of And-groups; fgAN is an element of FGAN, i.e., an And-group. FGAL is the set of Alterna- tive-groups. FGOR is the set of Or-groups. (1) F = {root} [ NF [ TF. root is the root feature. NF (–;) is the set of non-terminal features. TF (–;) is the set of terminal features. (2) F = {root} [ FAN [ FAL [ FOR. FAN, FAL, and FOR denote the set of features in all And-groups, in all Alternative-groups, and in all Or-groups respectively. Take the FM shown in Fig. 1 for example, root = ‘‘r’’; ‘‘f1’’ 2 NF; ‘‘f4’’ 2 TF; fgAN1 = {‘‘f1’’, ‘‘f2’’, ‘‘f3’’}, fgAN1 2 FGAN; FGAL = {‘‘fgAL1’’}, fgAL1 = {‘‘f12’’, ‘‘f13’’, ‘‘f14’’}. FOR = fgOR1 = {‘‘f10’’, ‘‘f11’’}. http://wwwiti.cs.uni-magdeburg.de/iti_db/research/featureide r: HIS f1: Detection (Det) f2: Monitor (Mon) f3: Service f4: Fire Det f5: Intrusion Det f6: Flood Det f7: Smoke Mon f8: Motion Mon f9: Moisture Mon f11: Video on demand f10: Internet connection f13: ADSL f14: Wirelessf12: Power Line Mandatory feature Alternative-group Or-group f4 requires f7. f5 requires f8. f6 requires f9. Constraints: And-groupNon-terminal feature Terminal feature Optional feature [1..n] Cardinality [1..1] [1..2] Fig. 1. A partial FM for the HIS SPL. Table 1 FM Semantics in propositional formulas. FM primitives Semantics Optional child C1 ? P Mandatory child C1 M P And-group (P ? V i2M Ci) ^ ( W 16i6nCi ? P) Or-group P M W 16i6n Ci Alternative-group P $ W 16i6n Ci � � ^ V i<jð:Ci _:CjÞ f1 requires f2 f1 ? f2 f1 excludes f2 :ðf 1 ^ f 2Þ � R is the set of relations in the FM. R = Parent [ Req [ Excl. Par- ent # F � F returns the parent of a given feature. It also forms an acyclic relation called feature hierarchy. If (f1, f2) 2 Parent, then f1 is a child of f2, f2 is the parent of f1. Parent⁄ is the reflexive, anti- symmetric, and transitive closure of Parent. Req # F � F encodes the F–F requires constrains; Excl # F � F encodes the F–F excludes constrains. For example, (‘‘f1’’, ‘‘root’’) 2 Parent; (‘‘f4’’, ‘‘f7’’) 2 Req. � A is the set of attributes in the FM. A = Opt [ Mincard [ Maxcard. Opt : FAN ! B denotes the optionality of a given feature in an And-group, B ¼ftrue; falseg. If Opt(f) returns true, f is optional, otherwise mandatory. For example, Opt(‘‘f3’’) = ‘‘true’’. Mincard : FGAL [ FGOR ! N0 and Maxcard : FGAL [ FGOR ! N0 [f1g return the cardinality for a given Or- or Alternative-group. For example, Mincard(‘‘fgOR1’’) = ‘‘1’’, Maxcard(‘‘fgOR1’’) = ‘‘2’’. � Domain: R [ A ? 2C and Range: R [ A ? 2C [ L give the set of domain (�2C) or range (�2C [ L) for some relation r(2R) or some attribute a(2A). Here, L denotes literal values of attributes. For example, Domain(Req(f4,f7)) = {f4}, Range(Req(f4, f7)) = {f7}; Domain(Opt(f3)) = {f3}, Range(Opt(f3)) = {true}. 3.2. Consistency constraints of FMs We summarize 13 consistency constraints of FMs in terms of the syntax and semantics of FMs. These consistency constraints form a feature consistency model (FCM): FCM ¼fCCi; 1 6 i 6 13g: The following set of constraints is by no means an exhaustive list of consistency constraints for FMs, but it lays a foundation for constructing and maintaining a consistent FM. We can define a decision function consistency(FM) to judge whether a given FM is consistent or not. consistencyðFMÞ¼ true; if an FM conforms to the FCM false; otherwise � 3.2.1. Syntactical consistency constraints Schobbens, Heymans, and Trigaux (2006), Schobbens, Heymans, Trigaux, and Bontemps (2007), Metzger, Heymans, Pohl, and Saval (2007) defined a formal semantics of FMs and several well-formed- ness rules of FMs based on Free Feature Diagrams (FFD). We extend their well-formedness rules and summarize 10 syntactical consis- tency constraints (CC1–CC10) based on the ontology-based formal- ization of FMs. CC1 (Distinct Identity Constraint). Every concept has a distinct identity: ðNF \ TF \frootg¼;Þ^ðFGAN \ FGAL \ FGOR \frootg ¼;Þ^ðFAN \ FAL \ FOR \frootg¼;Þ: CC2 (Feature Hierarchy Constraint). The feature hierarchy is a directed acyclic graph: :9f 2 F � ðf ; fÞ 2 Parent�: CC3 (Root Constraint). There is a unique feature root 2 F that is the direct or indirect parent of all other feature in F. 9root 2 F � ð8f 1 2 F nfrootg � ðf 1; rootÞ 2 Parent�Þ^ð:9f 2 2 F � ðroot; f 2Þ 2 Parent�Þ: CC4 (Feature-Closure Constraint). Every feature except root has one parent feature: 8f 1 2 F nfrootg �9f 2 2 F � ðf 1; f 2Þ 2 Parent: The constraint CC4 prevents the existence of the orphaned fea- tures. For example, the removal of the Parent relationship between the feature ‘‘f1’’ and the feature ‘‘r’’ in Fig. 1 would cause no parent feature to be defined for the feature ‘‘f1’’ any longer, which has to be prevented or resolved. CC5 (Relation-Closure Constraint). Any relation (2R) must be built between two legal features: 8f 1 � 8f 2 � ðf 1; f 2Þ 2 R ! f 1 2 F ^ f 2 2 F: CC6 (Attribute-Closure Constraint). Any attribute (2A) must be built between a legal features and a literal value (2L): 8f 1 � 8l � ðf 1; lÞ 2 A ! f 1 2 F ^ l 2 L: (a) (b) (c) Fig. 2. Examples of semantic inconsistencies of FMs. The constraints CC5 and CC6 demand that any relation (2R) or any attribute (2A) must be established only between two right ob- jects. For example, the addition of the Req relationship between the features ‘‘f8: Motion Monitor’’ and ‘‘Camera Surveillance’’ would provoke an inconsistency because the latter is not yet defined as a legal feature (2F). CC7 (Domain-Closure Constraint). The Domain concept can be established between a relation and a concept or between an attribute and a concept: 8c � 8ra � c 2 DomainðraÞ! c 2 C ^ðra 2 R [ AÞ: CC8 (Range-Closure Constraint). The Range concept can be established between a relation and a concept or between an attribute and a literal: 8cl � 8ra � cl 2 RangeðraÞ! ðcl 2 C ^ ra 2 RÞ_ðcl 2 L ^ ra 2 AÞ: CC9 (Cardinality-Closure Constraint). Cardinality must be spec- ified for Alternative- or Or-groups: 8fg � MincardðfgÞ_ MaxcardðfgÞ! fg 2 FGAL [ FGOR; CC10 (Cardinality Constraint). Cardinality must be well-formed: 8fg 2 FGAL � MincardðfgÞ¼ 1 ^ MaxcardðfgÞ¼ 1; 8fg 2 FGOR � MincardðfgÞ P 1 ^ MincardðfgÞ6 MaxcardðfgÞ 6 kfgk: The constraints CC9 and CC10 reflect cardinality-based feature modeling (Czarnecki et al., 2005). They formulate the Alterna- tive-group and Or-group relationships through a set of well- formed rules about cardinality. 3.2.2. Semantic consistency constraints von der MaBen and Lichter (2004) presented four situations that lead to semantic inconsistencies in FMs: (1) exclusion be- tween full-mandatory features; (2) exclusion between relative- full-mandatory features; (3) implication between alternative child features; (4) exclusion and implication. We extend their work and introduce two concepts. Definition 2 (Mandatory path). We define a path between a mandatory feature X and the root feature in an FM as a mandatory path MandPath (X) where every intermediate node (feature) is either a mandatory feature in an And-group or the sole child in an Alternative- or an Or-group. Definition 3 (Requires chain). Due to the transitivity of the Req relationship, we define a chain from the start node (feature) S to the end node T as a requires chain ReqChain(S, T) where all nodes are connected by the Req relationship to each other. Thus the first two inconsistent situations presented by von der MaBen and Lichter (2004) are merged and extended to the exclu- sion between any two features in one same or two different man- datory paths. Their fourth inconsistent situation (von der MaBen & Lichter, 2004) is extended to the exclusion between any two fea- tures in a requires chain. Examples of extended inconsistencies of FMs are shown in Fig. 2. Correspondingly, we summarize three semantic consistency constraints (CC11–CC13) as follows. CC11 (Excl-MandPath Constraint). Any two features in one same or two different mandatory paths of an FM cannot have the Excl relationship. Counter examples are shown in Fig. 2(a). 8f 1; f 2; X; Y 2 F �:9ðExclðf 1; f 2Þ^ðf 1 2 MandPathðXÞ^ðf 2 2 MandPathðXÞ_ f 2 2 MandPathðYÞÞÞÞ. CC12 (Req-Alternative Constraint). Any two child features in an Alternative-group cannot have the Req relationship. A coun- ter example is shown in Fig. 2(b). 8fgAL 2 FGAL � 8f 1; f 2 2 fgAL �:9ðReqðf 1; f 2Þ_ Reqðf 2; f 1ÞÞ. CC13 (Excl-ReqChain Constraint). Any two features in a requires chain cannot have the Excl relationship. A counter example is shown in Fig. 2(c).8f 1; f 2 2 F �:9ðReqChainðf 1; f 2Þ^ ðExclðf 1; f 2ÞÞÞ. 4. Changes to feature models Based on the above formalization of FMs, we can easily obtain the primitive elements of FMs, which include concepts (non-termi- nal features, terminal features, feature groups), relations (parent, re- quires, excludes) and attributes (name, group type, optionality, cardinality) of FMs. Since each primitive element of FMs can be changed by one of the meta-change transformations (Huersch, 1997; Rundensteiner, Leem, & Ra, 1998), we suggest a set of prim- itive operations on FMs in Table 2 (Guo & Wang, 2010). These oper- ations are defined by the cross product of the set of FM primitive elements and the set of meta-changes (‘Add’, ‘Remove’, and ‘Set’). They represent the changes to FMs at the lowest level of complex- ity and can compose various complex change operations such as the 16 operations for refactorings and generalizations of FMs (Alves et al., 2006) and the 5 operations for arbitrary edits to FMs (Thum et al., 2009). Further, we formalize changes to FMs as follows. Definition 4. A change to FMs Ch is a 4-tuple: Ch ¼ðname; args; preconditions; postconditionsÞ where: � name is the identifier of a change. Table 2 lists all the names of primitive operations. In the following chapters, we simplify the notation of changes as name (args). � args 2 (C [ R [ A [ L)n,1 6 n 6 3, is a list of one or more change arguments. A change could have one, two, or three arguments. Take the FM shown in Fig. 1 for example, to remove the non-ter- minal feature ‘‘f1’’ from the FM, the change RevNF has only one argument ‘‘f1’’. To modify the name of the node ‘‘r’’, the change SetName(‘‘r’’, ‘‘Home Integration Systems’’) is applied. The change AddRL(‘‘rl4–7’’, ‘‘f4’’, ‘‘f7’’) is applied to add a requires link ‘‘rl4–7’’ between the features ‘‘f4’’ and ‘‘f7’’. � Preconditions of a change comprise a set of assertions that must be true to be able to apply the change. If a precondition fails, a change is never performed. For example, the precondition for RevNF(‘‘nf’’) is nf 2 NF. Table 2 Primitive operations on FMs. � Postconditions of a change comprise a set of assertions that must be true after applying a change. They describe the effect of a change. For example, the postcondition for RevNF(‘‘nf’’) is nf R NF. For example, a full definition of the change RevNF can be as follows: Change F Remove non-terminal feature Syntax RevNF(‘‘nf’’) Semantics Remove a non-terminal feature ‘‘nf’’ from an FM Preconditions nf 2 NF Postconditions nf R NF Similarly, the preconditions and postconditions for other changes can also be deduced according to general logical con- straints and the consistency constraints defined above. Two deci- sion functions are defined as follows: preconditionsðFM;ChÞ¼ true; if an FM satisfies the preconditions of a Ch false; otherwise 8>< >: postconditionsðFM; ChÞ¼ true; if an FM satisfies the postconditions of a Ch false; otherwise 8>< >: 5. Semantics of change The evolution of FMs can be seen as a sequence of interdepen- dent changes to FMs (Guo & Wang, 2010). Such changes are ig. 3. Applying a change Ch to an FM. composed of a set of primitive operations defined in Table 2. A change to FMs can be seen as a mapping between FMs. As shown in Fig. 3, given an FM and a requested change Ch, the application of the change Ch to the FM results in another FM0, i.e., FM0 = Ch(FM), under preconditions(FM, Ch) = true ^ postconditions(FM0, Ch) = true. Since the application of a single change will not always leave an FM in a consistent state, it often derives a series of additional changes. Hence, the resolution of the requested change requires obtaining and executing these derived changes to maintain the consistency of the FM. Thus: Definition 5. Given an FM and a requested change Ch, the semantics of change to FM is defined as: SemanticsOfChangeðFM; ChÞ¼ ðCh1; . . . ; Chi; Chiþ1; . . . ; Chn�1Þ where: � FM is a given consistent FM, i.e., consistency(FM) = true; � Ch is a requested change that can be applied to the FM, i.e., preconditions(FM, Ch) = true; � FM1 = Ch(FM) is an FM representing the result of applying the requested change Ch to the FM, i.e., postconditions(FM1, Ch) = true; � Chi, 1 6 i 6 n � 1, is a derived change that satisfies the following set of conditions: – FMi+1 = Chi(FMi), which implies that preconditions(FMi, Chi) = true and postconditions(FMi+1, Chi) = true; – consistency(FMi) = false, 1 6 i 6 n � 1, and consistency(FMn) = true. Thus, as shown in Fig. 4, the final result of applying and resolv- ing the requested change Ch to the FM is the FM0: FM0 ¼ FMn ¼ Chn�1ð. . . Chiþ1ðChið. . . Ch1ðChðFMÞÞÞÞÞ: Next, how to find and organize these derived changes that re- solve the requested change and maintain the consistency of the FM? It is impractical to demand for domain analysts to track down and keep in mind all the changes that are pending. Hence, we adopt the procedural approach (Stojanovic, 2004) to realize the task automatically. The procedural approach comprises five steps (Stojanovic, 2004): first, a request is represented as a series of primitive operations defined in Table 2; second, the illegal operations are prohibited by checking the preconditions of each Fig. 4. The semantics of change to an FM. operation; third, additional operations are derived from the re- quested operations for keeping consistency; fourth, the execution order of the requested and derived operations is determined; fifth, all the confirmed operations are applied to the FM. Among the above steps, the third and the fourth steps are the key to consis- tency maintenance of FMs, other steps are straightforward. Hence, we explain how to implement the two steps as follows. Table 3 The dependency matrix Dependency[i][j]. (a) Fig. 5. Operation templates generated by the dependency matrix. (a) The general operatio change RevNF(‘‘f10’’) in the FM shown in Fig. 1. 5.1. Dependency matrix We analyze the cause and effect relationship between primitive operations on FMs and build the dependency matrix to conduct the derivation of additional operations from a requested operation. As shown in Table 3, the rows and columns of the matrix list all the primitive operations defined in Table 2. If an element of the matrix (b) n template for resolving the change RevNF. (b) A concrete template for resolving the (a) (b) (c) (d) Fig. 6. Three evolution strategies for resolving the operation ‘‘RevNF’’. (a) Applying the RevNF(‘‘f10’’) alone to the FM shown in Fig. 1. (b) All children are removed. (c) All children are reconnected to the parent. (d) All children are reconnected to another non-terminal feature. (a) (b) Fig. 7. Two evolution strategies for removing a feature in a Req link. (a) Remove the domain. (b) Remove the range. J. Guo et al. / Expert Systems with Applications 39 (2012) 4987–4998 4993 Dependency[Chi][Chj] is blank, it means that the operation Chi that is assigned to the row i can never induce the operation Chj denoting the column j. Otherwise, the operation Chi could generate the oper- ation Chj when their necessary preconditions and postconditions are fulfilled. The symbol ‘‘X’’ is used as the replacement for all the conditions. Most of the cause and effect relationships between primitive operations are deduced in terms of the consistency constraints de- fined in the FCM. The principles for generating the dependency ma- trix are as follows. First, an operation on a concept would affect the related attributes and relations of the concept. For example, since the RevNF operation causes the removal of all ‘‘edges’’ pointing to the feature or from it, the operations ‘‘RevPL’’, ‘‘RevRL’’, and ‘‘RevEL’’ are triggered. Second, an operation on an attribute such as the operations ‘‘SetName’’, ‘‘SetOpt’’, and ‘‘SetCard’’ do not initiate additional operations because they do not affect other elements but their own literal values. However, the operation ‘‘SetGT’’ is a special case because it would affect the logical structure of some feature group and thus could cause the operations ‘‘SetOpt’’ and ‘‘SetCard’’. According to the dependency matrix, an operation would cause a set of additional operations. Further, each of these derived oper- ations would cause another set of operations. Such operation der- ivation continues to propagate until there is no more new derived operations. All of these derived operations can form a general oper- ation template for resolving a certain operation. Fig. 5(a) shows a general multilevel operation template for resolving the change RevNF. The general template would be trimmed as a concrete tem- plate when applying to a practical scenario and its operations would be parameterized. Fig. 5(b) demonstrates a concrete opera- tion template for resolving the change RevNF (‘‘f10’’) in the FM shown in Fig. 1. The execution order indicated by the sequence number does not matter very much, but we often handle the oper- ations from outer level to inner and aggregate similar operations. 2 An implementation of our approach is available in http://code.google.com/p/ fmconmain/. 5.2. Evolution strategy Most of the primitive operations on FMs can be directly re- solved based on the operation templates generated by the depen- dency matrix. For example, all of the ‘‘Add’’ and ‘‘Set’’ operations defined in Table 2 can be executed straightforwardly once domain analysts determine right parameters. The operations ‘‘RevRL’’ and ‘‘RevEL’’ can also be executed directly. However, the other four ‘‘Re- move’’ operations cannot be resolved automatically by the depen- dency matrix and often need extra decision making by domain analysts. For example, after executing the operations ‘‘RevFG’’ or ‘‘RevTF’’, domain analysts must determine how to handle those non-terminal features at leaf position. Executing the operations ‘‘RevNF’’ or ‘‘RevPL’’ is more complex because domain analysts must determine how to handle the remaining orphaned part of the resulting FM. For these four operations, the operation tem- plates often provide multiple choices, e.g., the step 7 in Fig. 5(a) and the step 5 in Fig. 5(b). Therefore, evolution strategies are intro- duced to direct how to execute these operation and their derived operations resulting not in an arbitrary consistent state. An evolution strategy unambiguously defines the way in which a change will be resolved. It generally formulates an ordered se- quence for the requested change and its derived changes, i.e., the sequence ‘‘Ch, Ch1, . . ., Chi�1, Chi, . . ., Chn�1’’ shown in Fig. 4. Take the operation ‘‘RevNF’’ for example, there are three evolution strat- egies: removing all children, reconnecting all children to its parent, reconnecting all children to another non-terminal feature. Fig. 6 demonstrates the three evolution strategies for resolving the change RevNF (‘‘f10’’) in the FM shown in Fig. 1. Domain analysts can choose a particular evolution strategy in order to tailor the evolution of FMs to suit their needs. Resolving the operation ‘‘Rev- PL’’ can also apply these three evolution strategies. Resolving the operations ‘‘RevFG’’ and ‘‘RevTF’’ often needs adding additional fea- tures as terminal features. In addition, two evolution strategies are used for resolving the ‘‘RevTF’’ or ‘‘RevNF’’ operations on the Req links. As shown in Fig. 7, for the situation (b), i.e., ‘‘f1 requires f2’’, ‘‘f2’’ cannot be re- moved arbitrarily. In this case, domain analysts would be warned that the requested change could be an illegal operation. If domain analysts confirm the requested change, then they can first remove the link ‘‘Req(f1, f2)’’ and then remove the feature ‘‘f2’’. 6. Implementation We implemented our approach2 based on FeatureIDE, which is an open-source Eclipse-based IDE that supports building program http://code.google.com/p/fmconmain/ http://code.google.com/p/fmconmain/ Fig. 8. Extended FM editor based on FeatureIDE. Fig. 9. Evolution strategies implementation. families following the AHEAD3 architecture model and provides tools for the feature oriented design process and the implementation of SPLs. An extended FM editor based on FeatureIDE is shown in Fig. 8. It provides users with two views of FMs: tree view and hier- archy view. Users can input the keyword of some feature and then locate it. Attributes of features and feature groups, defined in Table 2, can be easily edited. Constraints also have a separate view and an edit area. 3 http://userweb.cs.utexas.edu/users/schwartz/. Evolution strategies are implemented by an interactive man- ner. For example, if the change RevNF (‘‘f10: InternetConnection’’) is applied in the FM shown in Fig. 1, a dialog box (as shown in Fig. 9(a)) is displayed for users to choose an evolution strategy. Users can choose to remove all children of the feature ‘‘f10’’ (the evolution strategy defined in Fig. 6(b)) or to reconnect these chil- dren to another non-terminal feature (the evolution strategies de- fined in Fig. 6(c) and (d)). If the latter is chosen, the dialog box is extended to prompt users to input the target feature, as shown in Fig. 9(b). Note that the target feature group and the children group of the removed feature could have different group type, http://userweb.cs.utexas.edu/users/schwartz/ so users must confirm a certain group type to merge the two fea- ture groups. Fig. 10. Calculation time in milliseconds for different scales of FMs using consistency checking (Sat4j and Guidsl) and consistency maintenance (our approach). 4 http://www.sat4j.org. 5 http://userweb.cs.utexas.edu/schwartz/ATS/fopdocs/guidsl.html. 7. Evaluation According to Stojanovic (2004), the computation complexity of resolving a change to FMs (e.g., RevNF) is about O(nm) where m is the average depth of the feature hierarchy starting from the considering feature and n the average number of child features. Generally, m and n are not large numbers because FMs usually contain limited layers and limited children for one feature. Thus, we make a preliminary evaluation that our approach accom- plishes the consistency maintenance of evolving FMs in an acceptable time. Further, we give more comprehensive evaluation by experi- mental studies. Although industries reported FMs with hundreds or thousands of features (Loesch & Ploedereder, 2007; Steger et al., 2004), authors typically published only a small excerpt of their FMs. Large FMs are difficult to find for a thorough evaluation. Thus, we adopt Thum’s method (Thum et al., 2009) to perform experiments using randomly generated FMs with different characteristics. 7.1. Experimental setup We first generated FMs randomly and then performed a set of primitive operations (defined in Table 2) randomly on the gener- ated FMs. Based on the dependency matrix, a set of additional operations are derived automatically from the requested opera- tions. Further, according to predefined evolution strategies, the de- rived operations are executed automatically to maintain the consistency of those changed FMs. During the above process, we parametrically control the size of FMs, the number and kind of operations, and the kind of evolution strategies for a thorough run- time evaluation. Independent parameters in our experiment are (a) the num- ber of features in an FM, (b) number of operations, (c) kind of operations, (d) kind of evolution strategies. The time needed to perform the requested and derived operations is measured as a dependent variable. To reduce the fluctuations in the dependent variable caused by the random generation, we performed 200 repetitions for each configuration of independent parameters, i.e., we generated 200 random FMs with the same parameters and each performed the same number of random operations of the same kind. All measurements were performed on the same Windows 7 PC with Intel Core Duo CPU 1.5 GHz and 3 GB RAM. 7.1.1. Feature models generation The algorithm to randomly generate FMs of size n is as follows (Thum et al., 2009): starting with a single root node, it runs several iterations. In each iteration, an existing node without children is randomly selected, and one to ten (random amount) of child nodes are added. Those child nodes are connected either by And- (50% probability), Or- (25% probability) or Alternative-group (25% prob- ability). Children in an And-group are optional by a 50% probabil- ity. This iteration is continued until the FM has n features. All features with children are considered non-terminal. Moreover, we also generate cross-tree constraints (requires and excludes). For every 10 features, one constraint is generated by the following algorithm: two different features are randomly selected, and then are connected randomly by requires (50% probability) or excludes (50% probability) link. The above generated FMs can be easily translated into proposi- tional formulas according to the rules give in Table 1. We use the SAT solver sat4J4 to validate these FMs and discard all FMs that do not have a single valid configurations (mostly by unfortunate choice of cross-tree constraints). We repeat the entire process until the appropriate number of valid FMs are generated. We fixed the following parameters: maximum number of chil- dren = 10; type of child group = (50%, 25%, 25%); optional child = 50%; number of cross-tree constraints = 0.1 ⁄ n; variables in cross-tree constrains = 2. According to Thum’s survey (Thum et al., 2009), these parameters are backed up by most of the sur- veyed FMs and represent a rough average. Thus, these generated FMs basically reflect the characteristics of realistic FMs. 7.1.2. Operation generation We randomly generated operations on an FM as well. Our generator takes an FM and the number of operations as input. The 20 primitive operations defined in Table 2 are implemented. They are classified as three main types: ‘add’, ‘remove’, and ‘set’. Our generator can limit the kind of input operations to ‘add’, ‘remove’, ‘set’, or ‘arbitrary’. For a fixed number of input operations, ‘arbitrary’ operations are composed of ‘add’ (33% probability), ‘remove’ (33% probability), and ‘set’ (33% probabil- ity) operations. 7.2. Experimental results and discussion 7.2.1. Effectiveness We first verify whether our approach can ensure the consis- tency of the resulting FMs after the requested changes are resolved. We varied the size of the generated FMs between 10 and 10,000 features. For each FM, we performed 10 random arbi- trary operations. 200 repetitions are performed for each model size. Each resulting FM is checked by Sat4j and Guidsl. Guidsl5 is a tool developed by Batory (2005) that relies on grammars defini- tion and propositional logic to support feature modularizations and their compositions. Results show that all the resulting FMs gen- erated by our approach are validated by Guidsl and Sat4j. Therefore, our approach can effectively maintain the consistency of evolving FMs. 7.2.2. Number of features In the same experimental setting as the above experiment, we also measured how calculation time scales as FMs increase in size. We also perform 10 random arbitrary operations on each FM http://www.sat4j.org http://userweb.cs.utexas.edu/schwartz/ATS/fopdocs/guidsl.html Fig. 11. Calculation time in 0.01 ms for different kinds of operations. Fig. 12. Calculation time in 0.01 ms for different number of operations on an FM with 1000 features. Fig. 13. Calculation time in 0.01 ms for different evolution strategies of the ‘‘RevNF’’ operation. whose size varies from 10 to 10,000. That is repeated 200 times and we only take the mean value. Fig. 10 shows the results of this measurement. To obtain a consistent FM from the requested operations on an original FM, previous approaches to ensuring the consistency of FMs mostly execute the operations first and then check the consis- tency of the resulting FM. During the process, most of the time is spent on consistency checking using various off-the-shelf solvers. Instead, our approach directly resolves the requested operations on the original FM and preserves the consistency of the resulting FM by construction. Fig. 10 compares the calculation time (the mean value for 200 repetition) to generate a consistent FM from 10 random operations on an original FM whose size varies from 10 to 10,000 using these two approaches. Here, we use Guidsl and Sat4j to check the consistency of the resulting FM. For small FMs (<500 features), there is no marked difference among the calculation time for the three cases. For large FMs (1000 features), our approach has better efficiency than previous approaches using Guidsl and Sat4j. Even for very large FMs with up to 10,000 features, our approach only spends 7.2 ms. 7.2.3. Kind of operations We performed the same measurement varying the model size from 10 to 10,000, but distinguished different kinds of operations. We distinguished between four kinds of operations: ‘add’, ‘remove’, ‘set’, and ‘arbitrary’. Again we applied 10 random operations of this classification (i.e., 10 adds, 10 removes, 10 sets, or 10 arbitrary operations) to each FM. Fig. 11 shows the results of our measurement (mean value of 200 repetitions for each combination of FM size and kind of oper- ations). The ‘remove’ operations spend the most time and the ‘set’ operations spend the least. There is not much difference between ‘remove’ and ‘add’ operations. The calculation time of ‘arbitrary’ operations is in the middle of that of the other three kinds of operations. 7.2.4. Number of operations We measured the calculation time when fixing the FM size to 1000 features and varying the number of operations from 0 to 100. Again we distinguished between the four kinds of operations: ‘add’, ‘remove’, ‘set’, and ‘arbitrary’. Fig. 12 shows the results of our measurement (mean value of 200 repetitions for each combination of the number and kind of operations in an FM with 1000 features). The ‘set’ operations still spend the least time and increase slowly with rising number of operations. The ‘add’, ‘remove’, and ‘arbitrary’ operations spend similar time. Their calculation time increases markedly with rising number of operations. 7.2.5. Kind of evolution strategies We measured the calculation time varying the FM size from 10 to 10,000 for different kinds of evolution strategies. We distin- guished between three evolution strategies for removing a non- terminal feature (the ‘‘RevNF’’ operation): removing all children, reconnecting the children to the parent, and reconnecting the chil- dren to another non-terminal feature, as shown in Fig. 6(b)–(d). For each situation, we applied 10 ‘‘RevNF’’ operations with random operation objects to each FM. Fig. 13 shows the results of our measurement (mean value of 200 repetitions for each combination of FM size and each kind of evolution strategy). There is no significant difference between the three evolution strategies in simulation environment. But in practice, only the first evolution strategy can be implemented automatically, as shown in Fig. 9(a). The execution of the last two evolution strategies requires users’ interactive participation, as shown in Fig. 9(b). 7.3. Discussion and threats to validity Experiments show that our approach works effectively and effi- ciently for large FMs. Mostly independent from the kind and num- ber of operations, our approach can generate a consistent resulting FM in less than 0.1 second even for very large FM (up to 10,000 fea- tures). This comfortably allows an implementation in our extended FM editor that shows how to resolve a change to an FM on the fly and maintain the consistency of the FM at runtime. Threats to internal validity are influences that can affect the cal- culation time that have not been considered. We cannot guarantee that computation time depends on certain shapes of an FM, or certain kinds of operations. Especially, we cannot guarantee ‘‘Arbi- trary’’ operations are composed of ‘‘Add’’, ‘‘Remove’’, and ‘‘Set’’ operations exactly in the proportion of 1:1:1. Since the calculation time for ‘‘Add’’ and ‘‘Remove’’ operations is markedly greater than that for ‘‘Set’’ operations, the uneven proportion of the three kinds of operations could cause fluctuation in the calculation time for ‘‘Arbitrary’’ operations. However, to avoid effects of certain FMs, all of input FMs are generated automatically by simulating known FMs and each measurement is repeated 200 times with freshly generated FMs. Threats to external validity are conditions that limit our ability to generalize the results of our experiment to industrial practice. First, we generated FMs with the described algorithm and param- eters, and confirmed that they align well with those known FMs acquired from existing publications in SPL community. Second, we generated a set of primitive operations on FMs by the cross product of all of FM primitive elements and a set of meta-changes (‘‘Add’’, ‘‘Remove’’, and ‘‘Set’’). We cannot guarantee these primi- tive operations are complete and typical in practice, however all reasonable operations we wanted to perform in our manual exper- iments, including the high-level operations defined by Alves et al. (2006), Thum et al. (2009), could be performed with one or a se- quence of these operations. 8. Related work Many formalizations and notations of FMs have been proposed (Batory, 2005; Czarnecki et al., 2005; Kang et al., 1990). Schobbens et al. formalized feature diagrams and detailed their generic semantics through a generic construction called Free Feature Dia- grams (Metzger et al., 2007; Schobbens et al., 2006, 2007). How- ever, there is a lack of detailed discussions about operations on FMs and their interrelationships. We give a new formalization of FMs from an ontological perspective. Based on this, we can easily obtain the primitive elements of FMs and then generate a set of primitive operations on FMs. Some researchers classified the modifications of FMs as special- izations (Czarnecki et al., 2005), refactorings (Alves et al., 2006), generalizations (Alves et al., 2006; Thum et al., 2009), or arbitrary edits (Thum et al., 2009). Czarnecki et al. (2005) introduced spe- cializations for deriving configurations of an FM, which result in FMs where some products are deleted. Janota and Kiniry (2007) formalized specifications between two FMs having the same set of features. Alves et al. (2006) discussed refactorings and general- izations that maintain the set of products or add new products to an SPL. They also suggested 16 operations for refactorings and gen- eralizations. Thum et al. (2009) complemented five additional operations for arbitrary edits. These operations work at a high level to explain the effects of FM evolution, but they are still not enough to explain the details of the FM evolution process. That is, how one changes an FM X into a target FM Y using a sequence of concrete and sound operations is still not obvious. We suggest a set of prim- itive operations on FMs by the cross product of all of FM primitive elements and three meta-changes. We do not think primitive oper- ations are better than those aggregated or high-level operations because primitive operations could aggravate the consistency problem (many primitive operations will necessarily leave an FM in an inconsistent state). However, primitive operations are fit to explain the details of resolving a requested change to an FM and maintaining the consistency of the FM. von der MaBen and Lichter (2004) proposed a framework for describing deficiencies of FMs. They characterized inconsistency as one of the most severe deficiencies of FMs. They also identified four situations that lead to semantic inconsistencies of FMs. Based on their work and Schobbens et al.’s (Metzger et al., 2007; Schobbens et al., 2006, 2007), together with our ontology-based formalization of FMs, we define a set of syntactical and semantic consistency constraints as the well-formedness rules of FMs. Many approaches focus on automated analysis of deficiencies of FMs (Benavides et al., 2010). Mannion (2002) first used proposi- tional formulas to analyze FMs. Batory (2005) proposed an ap- proach to debugging FMs using off-the-shelf SAT solvers. Czarnecki and Wasowski (2007) applied BDD tools to analyze FMs. Benavides et al. (2005) first adopted constraint programming to analyze FMs. Wang, Li, Sun, Zhang, and Pan (2005) first proposed the automated analysis of FMs using description logic. These ap- proaches mostly use SAT (Batory, 2005; Thum et al., 2009), BDD (Czarnecki & Wasowski, 2007), CSP solvers (Benavides et al., 2005), or description logic reasoning engines (Wang et al., 2005) to automate various reasoning tasks, e.g., checking satisfiability, detecting ‘‘dead’’ features, computing commonalities. They, how- ever, suffer from the NP-hard problem of feature combinatorics and thus take a long time to perform with large FMs (Batory et al., 2006). Moreover, these approaches only focus on the check- ing of deficiencies of FMs. Our approach resolves the possible inconsistencies caused by the requested changes to FMs by analyz- ing and applying the semantics of change to FMs and the interrela- tionships among primitive operations on FMs. It limits the consistency maintenance of evolving FMs in a local range affected by the requested changes not in the whole FM, which guarantee the efficiency and scalability of our approach for large FMs. Trinidad et al. (2008) provided a framework for explaining defi- ciencies of FMs based on constraint programming, but they do not give a solution to the deficiencies and the scalability of their ap- proach is also not clear. White et al. (2010) detected errors on the configurations of an FM, and proposed changes in the configu- rations according to features to be selected or deselected to correct the errors. Our work purely focuses on the evolution and consis- tency maintenance of FMs themselves, not the configurations of FMs. Some researchers also investigated relationships to other var- iability models or other core assets in SPLs. For example, Kastner and Apel (2008) proposed a formal approach to type-checking SPLs on the background of an FM. Metzger et al. (2007) proposed a for- malization of Orthogonal Variability Models (OVMs) and also fol- lowed the SAT approach to analyze OVMs automatically. Lauenroth and Pohl (2008) presented a consistency checking tech- nique for dynamic properties of the domain requirements specifi- cation based on OVMs. This paper greatly expands our previous work (Guo & Wang, 2010). In that paper (Guo & Wang, 2010), we only suggested a set of primitive operations on FMs and proposed a preliminary framework for analyzing the semantics of change to FMs. In this paper, we systematically give an ontology-based formalization of FMs and a set of consistency constraints, which build the theoret- ical foundations for generating primitive operations on FMs and analyzing their interrelationships. We detail how to obtain and ap- ply the dependency matrix and the evolution strategies for main- taining consistency of FMs. We also perform experiments to verify the effectiveness and efficiency of our approach. 9. Conclusions This paper proposes an approach to consistency maintenance for evolving FMs. We formalize FMs from an ontological perspec- tive and suggest a set of syntactical and semantic consistency con- straints as the well-formedness rules of FMs. We generate a set of primitive operations on FMs and analyze their interrelationships. By analyzing the semantics of change to FMs, the process of resolv- ing a requested change to an FM and maintaining the consistency of the FM is decomposed as a sequence of interdependent operations. The dependency matrix and evolution strategies are introduced to obtain and organize the operations sequence. Our approach is implemented and applied to randomly generated FMs. By experiments, we verify that our approach can effectively and efficiently maintain the consistency of evolving FMs with thousands of features. Our approach identifies a desirable property of FMs whose con- sistency maintenance can be achieved by construction. It is partic- ularly suitable for incremental management of FMs and their evolution in practice. Our approach improves significantly the complexity of the problem since we do not study a whole FM but the only part that changes since last consistent version. Although our approach de- pends on the completeness of the consistency constraints and the primitive operations we define, and if there were any situations where they are not applicable, previous approaches can still be ap- plied. So our approach has to be seen as an optimization for those cases where our consistency constraints and primitive operations can apply. Although FMs are the most popular kind of variability models, there are other alternatives that are still used and that we think our approach can be adapted to support consistency maintenance. Specifically, we plan to apply our approach to FMs with attributes (Benavides et al., 2010), OVMs (Metzger et al., 2007), and require- ments consistency checking (Lauenroth & Pohl, 2008). Acknowledgments The authors thank Prof. Robyn R. Lutz (Iowa State University) and Wei Zhang (Peking University) for their invaluable comments on the earlier draft of this paper. The authors also thank the anonymous reviewers and the attendees of SPLC’10 for their greatly helpful com- ments. Funding was provided by the National Natural Science Foun- dation of China (NSFC No. 60773088), the National High-tech R&D Program of China (863 Program No. 2009AA04Z106), the Key Pro- gram of Basic Research of Shanghai Municipal S&T Commission (No. 08JC1411700), the European Commission (FEDER) and Spanish Government under the CICYT project SETI (TIN2009-07366), and the Andalusian Local Government under the projects THEOS (TIC-5906) and ISABEL (P07-TIC-2533). References Alves, V., Gheyi, R., Massoni, T., Kulesza, U., Borba, P., & Lucena, C. (2006). Refactoring product lines. In Proceedings of GPCE’06, Portland, Oregon, USA (pp. 201–210). Batory, D. (2005). Feature models, grammars, and propositional formulas. In Proceedings of SPLC’05, Rennes, France (pp. 7–20). Batory, D., Benavides, D., & Ruiz-Cortes, A. (2006). Automated analysis of feature models: challenges ahead. Communications of the ACM, 49, 45–47. Benavides, D., Martin-Arroyo, P. T., & Cortes, A. R. (2005). Automated reasoning on feature models. In Proceedings of CAiSE’05, Porto, Portugal (pp. 491–503). Benavides, D., Segura, S., & Cortes, A. R. (2010). Automated analysis of feature models 20 years later: A literature review. Information Systems, 35, 615–636. Clements, P., & Northrop, L. (2001). Software product lines: Practices and patterns. Boston, MA, USA: Addison-Wesley. Czarnecki, K., & Wasowski, A. (2007). Feature diagrams and logics: there and back again. In Proceedings of SPLC’07, Kyoto, Japan (pp. 23–34). Czarnecki, K., Helsen, S., & Eisenecker, U. W. (2005). Formalizing cardinality-based feature models and their specialization. Software Process: Improvement and Practice, 10, 7–29. Gruber, T. R. (1993). A translation approach to portable ontologies. Knowledge Acquisition, 5, 199–220. Gruber, T. (2008). Ontology. In Encyclopedia of database systems. Springer-Verlag. Guo, J., & Wang, Y. (2010). Towards consistent evolution of feature models. In Proceedings of SPLC’10, Jeju Island, South Korea. LNCS (Vol. 6287, pp. 451–455). Haase, P., & Stojanovic, L. (2005). Consistent evolution of OWL ontologies. In Proceedings of ESWC’05, Heraklion, Greece (pp. 182–197). Huersch, W. (1997). Maintaining consistency and behaviour of object-oriented systems during evolution. ACM SIGPLAN Notices, 32, 1–21. Janota, M., & Kiniry, J. (2007). Reasoning about feature models in higher-order logic. In Proceedings of SPLC’07, Kyoto, Japan (pp. 13–22). Kang, K. C., Cohen, S. G., Hess, J. A., Novak, W. E., & Peterson, A. S. (1990). Feature- oriented domain analysis (FODA) feasibility study. Technical Report CMU/SEI-90- TR-021, Software Engineering Institute, CMU. Kang, K. C., Lee, J., & Donohoe, P. (2002). Feature-oriented product line engineering. IEEE Software, 19, 58–65. Kastner, C., & Apel, S. (2008). Type-checking software product lines – A formal approach. In Proceedings of ASE’08, L’Aquila, Italy (pp. 258–267). Lauenroth, K., & Pohl, K. (2008). Dynamic consistency checking of domain requirements in product line engineering. In Proceedings of RE’08, Barcelona, Spain (pp. 193–202). Loesch, F., & Ploedereder, E. (2007). Optimization of variability in software product lines. In Proceedings of SPLC’07, Kyoto, Japan (pp. 151–162). Lutz, R. R. (2008). Enabling verifiable conformance for product lines. In Proceedings of SPLC’08, Limerick, Ireland (pp. 35–44). Mannion, M. (2002). Using first-order logic for product line model validation. In Proceedings of SPLC’02, San Diego, CA, USA (pp. 176–187). Metzger, A., Heymans, P., Pohl, K., & Saval, P.-Y. S. G. (2007). Disambiguating the documentation of variability in software product lines: A separation of concerns, formalization and automated analysis. In Proceedings of RE’07, New Delhi, India (pp. 243–253). Pohl, K., Bockle, G., & van der Linden, F. (2005). Software product line engineering: foundations, principles, and techniques. Berlin, Heidelberg: Springer-Verlag. Rundensteiner, E., Leem, A., & Ra, Y. (1998). Capacity-augmenting schema changes on object-oriented databases: towards increased interoperability. In Proceedings of OOIS’98, Paris, France (pp. 349–368). Schobbens, P.-Y., Heymans, P., & Trigaux, J.-C. (2006). Feature diagrams: A survey and a formal semantics. In Proceedings of RE’06, Minneapolis/St.Paul, Minnesota, USA (pp. 136–145). Schobbens, P.-Y., Heymans, P., Trigaux, J.-C., & Bontemps, Y. (2007). Generic semantics of feature diagrams. Computer Networks, 51, 456–479. Steger, M., Tischer, C., Boss, B., Muller, A., Pertler, O., Stolz, W., et al. (2004). Introducing PLA at Bosch gasoline systems: Experiences and practices. In Proceedings of SPLC’04, Boston, MA, USA (pp. 34–50). Stojanovic, L. (2004). Methods and tools for ontology evolution. Ph.D. Thesis, University of Karlsruhe. Sugumaran, V., Park, S., & Kang, K. C. (2006). Software product line engineering: Introduction. Communications of the ACM, 49, 28–32. Thum, T., Batory, D. S., & Kastner, C. (2009). Reasoning about edits to feature models. In Proceedings of ICSE’09, Vancouver, Canada (pp. 254–264). Trinidad, P., Benavides, D., Duran, A., Ruiz-Cortes, A., & Toro, M. (2008). Automated error analysis for the agilization of feature modeling. Journal of Systems and Software, 81, 883–896. von der MaBen, T., & Lichter, H. (2004). Deficiencies in feature models. In Proceedings of workshop on software variability management for product derivation, SPLC’04, Boston, MA, USA. Wang, H., Li, Y., Sun, J., Zhang, H., & Pan, J. (2005). A semantic web approach to feature modeling and verification. In Proceedings of workshop on semantic web enabled software engineering (SWESE’05). White, J., Benavides, D., Schmidt, D. C., Trinidad, P., Dougherty, B., & Cortes, A. R. (2010). Automated diagnosis of feature model configurations. Journal of Systems and Software, 83, 1094–1107. Consistency maintenance for evolving feature models 1 Introduction 2 Feature models background 3 Ontology-based formalization and consistency constraints 3.1 An ontology-based formalization of FMs 3.2 Consistency constraints of FMs 3.2.1 Syntactical consistency constraints 3.2.2 Semantic consistency constraints 4 Changes to feature models 5 Semantics of change 5.1 Dependency matrix 5.2 Evolution strategy 6 Implementation 7 Evaluation 7.1 Experimental setup 7.1.1 Feature models generation 7.1.2 Operation generation 7.2 Experimental results and discussion 7.2.1 Effectiveness 7.2.2 Number of features 7.2.3 Kind of operations 7.2.4 Number of operations 7.2.5 Kind of evolution strategies 7.3 Discussion and threats to validity 8 Related work 9 Conclusions Acknowledgments References 
work_yn3oefrdz5gnfnwlbederzvioi	----	doi:10.1016/j.eswa.2006.09.018 www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 530–540 Expert Systems with Applications A fuzzy CBR technique for generating product ideas Muh-Cherng Wu *, Ying-Fu Lo, Shang-Hwa Hsu Department of Industrial Engineering and Management, National Chiao Tung University, Hsin-Chu, Taiwan, ROC Abstract This paper presents a fuzzy CBR (case-based reasoning) technique for generating new product ideas from a product database for enhancing the functions of a given product (called the baseline product). In the database, a product is modeled by a 100-attribute vector, 87 of which are used to model the use-scenario and 13 are used to describe the manufacturing/recycling features. Based on the use- scenario attributes and their relative weights – determined by a fuzzy AHP technique, a fuzzy CBR retrieving mechanism is developed to retrieve product-ideas that tend to enhance the functions of the baseline product. Based on the manufacturing/recycling features, a fuzzy CBR mechanism is developed to screen the retrieved product ideas in order to obtain a higher ratio of valuable product ideas. Experiments indicate that the retrieving-and-filtering mechanism outperforms the prior retrieving-only mechanism in terms of generating a higher ratio of valuable product ideas. � 2006 Elsevier Ltd. All rights reserved. Keywords: New product development; Case-based reasoning; Fuzzy CBR; Fuzzy AHP 1. Introduction Product life cycle is getting shorter at this age. How to enhance the productivity of new product development is very important. A typical process of new product devel- opment includes product idea generation, conceptual design, detailed design, and economic justification of design. Among these procedures, generation of product ideas may be the most important because the other proce- dures are intended to realize and justify the generated ideas. Various methods for creating product ideas have been published in the literature. According to the degree of computerization, the previous methods can be grouped into three categories: (1) manual-based approach, (2) computer- aided approach, and (3) computer-generated approach. The manual-based approach is to generate product ideas by asking an individual or a group of people to think freely or think under a guided process. Examples of this approach 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.09.018 * Corresponding author. Tel.: +886 35 731 913; fax: +886 35 720 610. E-mail address: mcwu@cc.nctu.edu.tw (M.-C. Wu). include brainstorming method (Higgins, 1994; Nijssen & Lieshout, 1995; Osborn, 1963), forced relationships method (Higgins, 1994; Hisrich, Ingram, & Peters, 1991; Kotler, 1994), focus groups method (Higgins, 1994; Hisrich et al., 1991; Kotler, 1994; Nijssen & Lieshout, 1995), attribute listing method (Higgins, 1994; Kotler, 1994; Linda, 1991; Nijssen & Lieshout, 1995), check-list method (Higgins, 1994; Kotler, 1994), morphological analysis (Higgins, 1994; Kotler, 1994; Linda, 1991; Nijssen & Lieshout, 1995), and synectics method (Higgins, 1994; Kotler, 1994). Techniques of this approach are carried out solely through human, without using any computing facilities in the creation of new product ideas. The computer-aided approach intends to use computer to guide a person’s thinking process for creating or realiz- ing product ideas. A typical technique of this approach encodes the innovative rules listed by TRIZ (Mann, 2003; Rantanen & Domb, 2002) in a computer program. Through a series of human–computer interaction activities, users of the computer program can find a number of design tem- plates to realize a user-desired product function. Example software of the technique includes Goldfire Innovator (2006), Trisolver (2006), and Creax.com (2006). mailto:mcwu@cc.nctu.edu.tw M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 531 The computer-generated approach is intended to create ideas for enhancing the function of a baseline product by retrieving ‘‘scenario-compatible’’ products from database (Wu, Lo, & Hsu, 2006). The retrieved products are similar enough to the baseline product in the scenario where the baseline product is used. Product functions of the retrieved ones are called product ideas. This approach would gener- ate a large amount of product ideas in a very short time. However, one weakness is that many less-valued products ideas may be generated, which would consequently require a large amount of human efforts to screen them. To alleviate the weakness, this paper presents a CBR (case-based reasoning) technique combined with a fuzzy AHP method for retrieving product ideas that tend to be more-valued, and from the retrieved ones screening out those ideas that tend to be less-valued. As shown in Fig. 1, the research framework involves three modules. The first module is to establish a product database in which a product is encoded by a vector involv- ing 100 attributes. Of these attributes, 87 ones represent the scenario of using a product and the other 13 represent the scenario of manufacturing and recycling the product. Each attribute is defined by a linguistic variable of fuzzy theory (Zadeh, 1975). The second module is to characterize the scenario of use for target customers. The 87 attributes for modeling the scenario of use are classified into five categories (also called dimensions). The technique of fuzzy AHP is used to deter- mine the relative weight for each of the five dimensions in order to understand the preferences of target customers. The weighting of product attributes is intended to help retrieving more-valued products ideas; that is, less-valued ideas may not be generated in the retrieval stage. The third module firstly retrieves product ideas whose functions tend to be attachable to the baseline product, and then filters out those retrieved ideas that tend to be less-valued. The retrieving mechanism is by using the 87 Establishing product database Weighting the use-scenario of target customers Retrieve and screen Product ideas Fig. 1. Research framework. product attributes that have been weighted to characterize the scenario of use for target customers. The screening mechanism is by using the 13 manufacturing and recycling attributes. The research framework is developed by examining the three main stages of a product life cycle – manufacturing, using, and recycling. The 100-attribute product representa- tion for generating new product ideas are developed based on the various costs/benefits concerned in a product life cycle. We retrieve product ideas from the perspective of enhancing the usability in order to increase product value; and filter out product ideas from the perspectives of reduc- ing product cost – reducing the manufacturing/recycling costs. The remainder of this paper is organized as follows. Sec- tion 2 reviews the literature on case-based reasoning (CBR) technique. Section 3 describes the method for representing a product by a vector of 100 attributes. Section 4 presents the fuzzy AHP method for determining the relative weights to characterize target customers’ scenario of use. Section 5 describes the CBR method for retrieving and screening product ideas. Experiment results are presented in Section 6 and concluding remarks are in Section 7. 2. Case-based reasoning Case-based reasoning (CBR), a well-known artificial intelligence technique, is a process for solving a new prob- lem case by referring to the solutions of similar past cases (Aamodt & Plaza, 1994; Kolodner & Leake, 1996; Marling, Sqalli, Rissland, Munoz–Avila, & Aha, 2002). In a CBR system, a database for storing the past cases has to be avail- able. To solve a new problem case by CBR, similar past cases are first retrieved and their associated solutions are then used to aid users to develop solutions for the new case. Two survey papers of CBR can be referred to Watson and Marir (1994), de Mantaras and Plaza (1997). A CBR system involves three modules: (1) a case repre- sentation scheme, (2) a similarity metric, and (3) a case- retrieval mechanism. A case representation scheme is to model a case by a set of attributes for characterizing the case at a particular application. A similarity metric is for measuring the similarity between any two cases. A case- retrieval mechanism is designed to retrieve the past cases that are similar enough to the new case. To make a CBR system more user-friendly, some studies proposed a fuzzy-CBR approach. This approach advocates using linguistics variables in fuzzy theory to valuate the case attributes. A linguistic variable is represented by a nat- ural language form as well as by a fuzzy number. The text description is intended to help users resolve the uncertainty issues while they valuate the case attributes. The fuzzy number representation and the associated fuzzy operators are used to calculate the similarity metrics for implement- ing the case-retrieving mechanism. Much literature based on such a fuzzy-CBR approach has been published. 532 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 Examples include Hirota et al. (1997), de Mantaras and Plaza (1997), Ruet and Geneste (2002), and Chan (2005). The CBR paradigm has been applied in a wide variety of design problems. The applications include architecture design (Trousse & Visser, 1993), mechanical design (Maher & Garza, 1997), and some other design problems. In the CBR applications for product design, existing design archi- tectures/parameters are retrieved to aid the realization of a product function (Bilgic & Fox, 1996; Maher, Balachan- dran, & Zhang, 1995) or to help engineers develop a new design that could fulfill the downstream requirements (Belecheanu, Pawar, Barson, Bredehorst, & Weber, 2003). These previous CBR studies for product design focus on the engineering aspect – realizing a design for a given func- tional requirement. Yet, this paper’s focus – how to apply CBR to create new and marketable functional require- ments for a given product is rarely studied. 3. Product database In the product database for generating and screening ideas, a product is encoded by a vector consisting of 100 attributes, where 87 ones are used to characterize the sce- nario of using the product, and 13 attributes are used to describe the manufacturing and recycling characteristics. For a product, each attribute is described by a linguistic variable in fuzzy theory. 3.1. Product attributes for creating ideas This research generates new product ideas based on the following hypothetical assertion – two products that are similar in their scenario of use have chance to be combined into a new product. That is, the new product would involve the main functions of the two products. To generate new product ideas for a given product (called the baseline prod- uct), we aim to identify some other products that are sim- ilar to the baseline product in its scenario of use. This research models a product use-scenario from five dimensions, which are developed based on the notion of UCD (user centered design) – a design paradigm originally proposed by Norman and Draper (1986). The UCD notion advocates that a product should be designed based on the Fig. 2. The proposed product representation is based on a UCD paradigm. user’s needs and the scenario where they use the product. As shown in Fig. 2, the five dimensions involve: interface attributes, task type, physical feature, environment, and user characteristic characteristics; and they are deployed into 20 sub-dimensions (also called groups) that ultimately yield 87 attributes (Table 1). The first dimension – interface attributes are to identify the medium through which the product interacts with the user. Here, the medium denotes a particular portion of the user’s body. This research classifies the interface attri- butes into three groups: sensory modality, response modal- ity, and interface point. The interface attributes are composed of 18 attributes in total. The second dimension – task type is to model the tasks to be performed by the user through using the product. According to users’ needs, the tasks are categorized into seven groups: eating, clothing, living, transportation, educa- tion/entertainment, working, and health care. These seven groups are further characterized by 30 attributes based on the execution process in each group. The third dimension – physical feature is intended to describe a product from the aspects of physical size, mobil- ity, and scaleability. Ten attributes are used to model the dimension of physical feature. The fourth dimension – environment is intended to char- acterize the environment where the product is used. The characterization involves three groups: sociality, physical place, and the harshness of environment. Ten attributes are used to model the environment dimension. The fifth dimension – user characteristics are intended to describe which groups of users tend to use the product. The dimension is characterized by the following demographic features: gender, age, profession, and job title. These five features are further described by 19 attributes. 3.2. Product attributes for screening ideas Based on the aforementioned product modeling method, new product ideas of the baseline product may be retrieved by applying the fuzzy CBR technique. Surely, a retrieved product idea and the baseline product to a certain extent are ‘‘compatible’’, from the perspective of using the two products. However, from the perspectives of manufactur- ing/recycling, these two products may not be economically justifiable to combine them into one. That is, we assert that a product idea will be discarded, if it has few commonality with the baseline product in terms of manufacturing/recyc- ling attributes. As shown in Table 1, 13 attributes are used to model the features of manufacturing and recycling, which are grouped into four groups. The first group describes the materials of a product, which involves the following four types: (a) metal, (b) non- metal, (c) animals, and (d) plants. A product may be com- posed of several types of materials. The percentage of a particular type of material used in a product is described by an attribute. Table 1 Product representation for cell phones and ball pens M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 533 534 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 The second group describes the processing mechanisms, which involves three types: (a) physical processes, (b) chemical processes, and (c) biological processes. A product may be manufactured by more than one type of processes. The percentage of a particular type of process used to man- ufacture a product is described by an attribute. The third group describes the energy resources used in a product, which involves four types: (a) electricity, (b) bat- teries, (c) solar energy, and (d) oil/gas. A product may use more than one type of energy resources. The percentage of energy resources used by a product is described by an attribute. The fourth group uses two attributes model the pro- cesses and results of recycling a product, which involves (a) the easiness in recycling a product, and (b) the usability of a recycled product. The easier is a recycling process, the higher is its attribute value; the higher is the usability of a recycled product, the higher is its attribute value. 3.3. Linguistic variables for describing product attributes To facilitate users to characterize a product, each of the 100 product attributes is described by a linguistic variable in fuzzy theory. A linguistic variable, appearing in a natu- ral language form, represents a human’s judgment, which can be further modeled by a triangular fuzzy number (Zadeh, 1975). Of the 100 product attributes, the first 98 attributes are described by five linguistic variables: extreme relevance, high relevance, relevance, low relevance, and no relevance. The associated fuzzy number of a linguistic variable, ~A ¼ðl1; m1; r1Þ, is shown in Fig. 3, where l1 denotes the 0.0 0.25 0.5 0.75 1.0 No relevance Low relevance Relevance High relevance Extreme relevance xA ~ X μ ( ) Fig. 3. Linguistic variables and the associated fuzzy numbers. M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 535 leftmost coordinate, m1 is the central coordinate, and r1 the rightmost coordinate on the x-axis (Moon & Kang, 2001). The last two attributes (attribute 99 and 100), which char- acterize the recycling characteristics, are described by another five linguistic variables: very high, high, normal, low, and very low. The fuzzy numbers in Fig. 3 are also used to represent these linguistic variables. 4. Weighting each dimension of target customers’ use-scenario In retrieving product ideas, we have to determine the rel- ative importance of the five dimensions of target custom- ers’ use-scenario. To resolve the vagueness caused by human judgment, we use a widely used methodology – fuzzy AHP (analytical hierarchy process) to determine the weighting for each of the five dimensions. The computational procedure of the fuzzy AHP meth- odology (Zadeh, 1975) is summarized below, where the arithmetical operators of fuzzy numbers are defined in Appendix 1. Step 1: Define linguistic variables for pair-wise compari- son. We use linguistic variables to compare the relative importance between any two dimensions. These linguistic variables include ‘‘absolutely impor- tant’’, ‘‘very strongly important’’, ‘‘ essentially important’’, ‘‘weakly important’’ and ‘‘equally important’’ on a five level scales, where between any two consecutive scales an intermediate scale Table 2 Linguistic variables used in the fuzzy AHP Fuzzy number ~1 ¼ð1; 1; 1Þ ~3 ¼ð2; 3; 4Þ ~5 ¼ð4; 5; 6Þ ~7 ¼ð6; 7; 8Þ ~9 ¼ð8; 9; 10Þ ~2 ¼ð1; 2; 3Þ; ~4 ¼ð3; 4; 5Þ; ~6 ¼ð5; 6; 7Þ; ~8 ¼ð7; 8; 9Þ is additionally defined so that 9 scales are created. Table 2 shows the resulting 9 scales that are repre- sented by 9 triangular fuzzy numbers (Chiou, Tzeng, & Cheng, 2005): ~1; ~2; . . . ; ~9. Step 2: Establish the comparison matrix ~A. By performing a pair-wise comparison for any two of the five concerned dimensions, a fuzzy matrix ~A is constructed. Linguis Equally Weakly Essenti Very st Absolu Interme ~A ¼ 1 ~a12 � � � ~a1n ~a21 1 � � � ~a2n .. . .. . . . . .. . ~an1 ~an2 � � � 1 2 6664 3 7775 where tic va imp imp ally im rongl tely i diate if i ¼ j; ~aij ¼ 1, if i 6¼ j; ~aij ¼ ~a�1ji and ~aij ¼ð~a1ij � a2ij ��� � �~aNijÞ=N ~akij: customer k’s judgment on the relative importance between dimension i and j N: total number of target customers in- terviewed. Step 3: Calculate the fuzzy weight of each row in ~A (Buck- ley, 1985). ~Z i ¼ð~ai1 � ~ai2 ��� �� ~ainÞ 1 n 8i ¼ 1; 2; . . . ; n ~W i ¼ ~Z i � ~Z 1 � ~Z 2 ����� ~Z n � ��1 8i ¼ 1; . . . ; n Step 4: Defuzzication of ~W i and ~A (Teng & Tzeng, 1993). aij ¼ Defuzzyð~aijÞ W i ¼ Defuzzyð ~W iÞ where the function Defuzzy is stated in Appendix 1. riables ortant ortant portant y important mportant values between two adjacent judgments Table 3 Values of RI n 1 2 3 4 5 6 7 8 9 10 11 RI 0.00 0.00 0.58 0.90 1.12 1.24 1.32 1.41 1.45 1.49 1.51 536 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 Step 5: Normalization of Wi Ŵ i ¼ W iPn i¼1W i : Step 6: Consistency check. (1) Compute W �i as follows:2 3 �2 3 A � ~W 1 ~W 2 .. . ~W n 66664 77775 ¼ W 1 W �2 .. . W �n 66664 77775 where A = [aij]. (2) Compute kmax ¼ 1 n W �1 Ŵ 1 � � þ W �2 Ŵ 2 � � þ�� �þ W �n Ŵ n � �� � where kmax is called maximum eigenvalue. (3) Compute CI ¼ kmax � n n � 1 where CI is called consistency index. (4) Compute CR = CI/RI, where CR is called consistency ratio, and RI is called the average random consistency index of randomly generated matrices of size n · n. The values of RI have been provided by Saaty (1980) as shown in Table 3. (5) Consistency check. If CR 6 0.1, then the pair-wise comparison matrix is reasonably consistent and W _ i; 1 6 i 6 n; is the result- ing weighting of dimension i. If CR > 0.1, then the pair-wise comparison results are inconsistent and the pair-comparison procedure has to be updated. Notice that W _ i; 1 6 i 6 5; is the weighting factor of dimension i in the product representation. The five dimen- sions include 87 use-scenario attributes. The attributes in each dimension are of the same weighting – the weight of their parent dimension. Let wj represent the weighting fac- tor of attribute j. Then, wj ¼ W _ i if attribute j belongs to dimension i. In summary, the results of the fuzzy AHP yield the weighting of each use-scenario attribute (wj, 1 6 j 6 87). 5. Retrieving and filtering product ideas Given a product database, this research uses two mech- anisms to propose new product ideas for enhancing the functions of a baseline product. Firstly, from the database, we retrieve products that tend to be compatible to the baseline product – from the perspective of product usabil- ity. Secondly, we filter out the retrieved products whose combinations with the baseline product tend to become costly – from the perspective of manufacturing and recycling. The retrieving and filtering mechanisms are explained by referring to a scenario stated below. A product database P = {Pi, i = 1, . . . , K} has been established, where K is a huge positive integer number and P i ¼ ½~pij�; 1 6 j 6 100; denotes the vector representation of product i. Let B ¼ ½~bj�; 1 6 j 6 100; represents the baseline product. Notice that ~pij and ~bj are fuzzy numbers that indicate the attri- butes of a product. The purpose is to retrieve some prod- ucts from database P, whose combinations with baseline product B may enhance the resulting product usability in a low-cost manner. 5.1. Retrieving mechanism The procedure of the retrieving mechanism is described below. Step 1: Define a retrieving threshold Hr 2 [0,1]. Step 2: Identify the important attributes of the baseline product B ¼ ½~bj�; 1 6 j 6 100. S ¼fjj~bj is of high relevance or extreme relevanceg (Refer to Fig. 3). Let N(S) denote the number of attributes in set S. Step 3: Form retrieval key sets. Ask users to randomly cluster the attributes in S into several subsets Tk, 1 6 k 6 q = dN(S)/me, where each subset involves m or (m � 1) attri- butes. That is, S ¼ Sq k¼1T k , where subset Tk is called a retrieval key set. Step 4: With respect to the retrieval key set Tk, compute product relevance metric between products P i ¼ ½~pij� and baseline product B ¼ ½~bj�. ~RT ki ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP j�T k½ ~bj � Rð~bj; ~pijÞ� wj� 2 P j�T k½bj � wj� 2 vuut for k ¼ 1; 2; . . . q where wj is the weighting factor of attribute j as de- rived in Section 4, Rð~bj � ~pijÞ as defined below rep- resents the relevance metric of jth attribute between baseline product B and product Pi. Rð~bj; ~pijÞ¼ 1 �j~bj � ~pijj M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 537 Step 5: Defuzzication of ~RT ki RT ki ¼ DefuzzyðR T k i Þ for k ¼ 1; . . . ; q Step 6: Retrieve products compatible tobaseline product B from database P. Qk ¼fP ijR T k i P H rg Q ¼ [q k¼1 Qk where Q represents the set of retrieved products. Several distinct points in the retrieving mechanism are explained further. First, only important attributes in base- line product B are included in a retrieval key set. Experi- ments indicate that the inclusion of less important attributes would lead to the retrieval of a huge number of products that are irrelevant to baseline product B. Sec- ond, a retrieval key set involves only a few number of important attributes. With the inclusion of all important attributes in a retrieval key set, the number of retrieved products tends to be very small; and their functions tend to be too close to the baseline product B and cannot be seen as a good product idea. Third, in the computation of ~RT ki , ~bj denotes the value of jth attribute and wj denotes its relative importance from the perspective target customer. That is, ~bj is independent of target customers while wj is dependent on target customers. 5.2. Filtering mechanism The retrieved products in set Q are relevant to the base- line product B, from the perspective of usability. However, some of these products may not be compatible to product B from the perspectives of manufacturing and recycling. We use a filtering mechanism to filter out these incompat- ible products in Q. The procedure of the filtering mecha- nism is described below. Step 1: Define a filtering threshold, Hf 2 [0,1]. Step 2: Define the filtering key set Tm, m = 1, . . . , 4. As stated in Section 3, we use 13 attributes (attri- butes 88–100) to model the features of manufac- turing and recycling, which are grouped into four categories. The attributes in mth category forms a filtering key set, denoted by Tm. Step 3: Compute compatible metric between product Pi and B, with respect to Tm. Table 4 Detailed data in the fuzzy AHP for characterizing target customers, where CI = 0.08199 and CR = 0.07321 Dimension ~W i Wi Ŵ i Interface attributes (0.10499, 0.15366, 0.22404) 0.160896 0.1540 Task type (0.25824, 0.36120, 0.50480) 0.374747 0.3588 ~CT mi ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP j�T m½~bj � Rðbj; pijÞ� 2 P j�T m½~bj� 2 vuut for m ¼ 1; 2; . . . 4 where Rð~bj; ~pijÞ¼ 1 �j~bj � ~pijj. Physical feature (0.10252, 0.14837, 0.21435) 0.155079 0.1485 Step 4: Defuzzication of ~CT mi Environment (0.10462, 0.15165, 0.21995) 0.158739 0.1520 User characteristics (0.12496, 0.18512, 0.27504) 0.195040 0.1867 CT mi ¼ Defuzzyð~C T m i Þ; for m ¼ 1; . . . ; 4 Step 5: Filter out the incompatible products from set Q F m ¼fP ijCT mi < H fg F f ¼[4m¼1F m where Ff represents the set of incompatible prod- ucts in set Q. Step 6: Determine X, the set of products that may enhance the function of product B effectively X ¼ Q � F f : 6. Experiments An empirical study is carried out to compare the effi- ciency and effectiveness between the proposed retrieving- and-filtering mechanism and the retrieving-only mechanism published in Wu et al. (2006) that has been justified to be better than the traditional brainstorming approach. Two experiments for the comparison are performed; one exper- iment uses cell phone and the other uses ball pen as the baseline products. A prototype product database is established for the experiments, which involves 1600 products and is coded by Microsoft Access. The retrieving/filtering mechanism is coded by ASP.NET (Active Server Page.NET), with its interface developed by Macromedia Dreamwaver, and Microsoft Internet Explorer is used as the vehicle for web browsing. To determine the dimensional weighting of target cus- tomers’ use-scenarios, 30 female subjects, aged 19–30, are invited to perform the pair-wise comparison required by the fuzzy AHP method. Results and the associated data of the AHP process are listed in Table 4. 6.1. Generating product ideas Five senior undergraduate students are invited as exper- iment subjects. Using cell phone as the baseline product, each step in executing the retrieving-and-filtering mecha- nism is explained below. Step 1: Set the retrieving threshold, Hr = 0.5. Step 2: Of the 87 product attributes of a cell phone, 10 attributes are automatically identified as impor- tant attributes (highlighted in Table 1). 538 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 Step 3: From the 10 important attributes, each subject is asked to freely form four retrieval key sets; each set involves either 3 or 2 attributes. Step 4: Based on the retrieval key sets, the retrieving mechanism will generate a set Q – the set of retrieved products. Step 5: Set the filtering threshold, Hf = 0.5. Step 6: Based on the four filtering key sets, the filtering mechanism automatically identifies a set Ff that represents the incompatible products in set Q. Step 7: The system computes the set X = Q � Ff. In the aforementioned procedure, set X represents the product ideas proposed by the retrieving-and-filtering mechanism and set Q represents the product ideas pro- posed by the retrieving-only mechanism. 6.2. Comparison of product-idea-generating mechanisms The performance metric for comparing the two product- idea-generating mechanisms is called creative ratio, as defined below. CZ ¼ NðZ gÞ NðZÞ where Z represents a set of product ideas, Zg is a subset of Z that includes only good product ideas, and N(Z) repre- sents the number of product ideas in set Z. The objectives of the experiments are to compare the value of CX and CQ (Table 5). As stated, X is a subset of Q. A random filtering mechanism would filter out good product-ideas at a probability of CQ; this on average yields that CX = CQ. By contrast, a less-effective filtering mecha- nism would yield that CX < CQ, while an effective filtering mechanism would yield that CX > CQ. That is, CX > CQ indicates that the average time required to identify one good product-idea is less. The method for justifying whether a product-idea is good is through expert’s evaluation. Three experts familiar with new product developments are invited to evaluate the generated product-ideas based on three criteria – original- ity, valuableness, and usefulness (Besemer & O’Quin, 1986). Each criterion is rated in a five-point scale – the higher the better. A product-idea is justified by averaging Table 5 Comparing the performance of the retrieving-and-filtering and the retrieving-o Cell phone Retrieving Retrieving + filtering N(Q) N(Qg) CQ (%) N(X) N(Xg) CX Sub_1 302 30 9.93 134 23 17.1 Sub_2 179 21 11.73 82 17 20.7 Sub_3 254 24 9.45 119 16 13.4 Sub_4 254 26 10.23 98 18 18.3 Sub_5 207 26 12.56 104 19 18.2 Mean 239 25.4 11.02 107 18.6 17.7 the points of the three criteria; an average point greater than 4.0 is regarded as a good idea. The results of the two experiments are shown in Table 5. For each baseline product, the mean of CX is greater than that of CQ. A t-test for cell phone (a = 0.05, t-value = �7.641, and P-value = 0.002) indicates CX > CQ is statisti- cally significant. Another t-test for ball pen (a = 0.05, t-value = �5.287, and P-value = 0.006) also supports the finding CX > CQ. These two findings conclude that the pro- posed filtering mechanism is effective. That is, the average time required to manually identify one good product-idea from the retrieved products is reduced if we enhance the retrieving mechanism by a filtering mechanism. However, the advantage of CX > CQ is offset by a draw- back – N(Qg) > N(Xg). That is, some good product-ideas are filtered out in the filtering mechanism. One may ques- tion that what is the ‘‘net benefit’’ of developing the filtering mechanism. Is the ‘‘net benefit’’ positive or negative? Con- sider the comparison of the retrieving mechanism and an exhaustively-listing mechanism – taking all products in the database as generated product-ideas. Surely, the exhaustively-listing mechanism can always generate more numbers of good ideas than the retrieving mechanism, at the expense of paying more expert time to identify good product ideas. This analogy illustration may explain the need for developing an effective filtering mechanism. Moreover, the performance of the filtering mechanism can be regulated by giving different values to the filtering threshold (Hf). In the case of Hf = 0, the filtering mecha- nism is halted; that is, only the retrieving-mechanism works. Increasing the value of Hf tends to pay more atten- tion to the filtering mechanism. Users of the product-idea- generating systems may iteratively give different Hf value, depending upon the number of retrieved ideas and how much expert time is available to evaluate these ideas. Sup- pose an initial assignment of Hf = 0.3 yields 1 million product-ideas. This would lead the users to reassign a higher Hf value to reduce the numbers of the retrieved ideas. Table 6 gives 23 product-ideas for enhancing the func- tion of cell phone. Table 7 describes 18 product-ideas for enhancing the function of ball pen. To our knowledge, some of these ideas for enhancing the baseline products are currently not available in the market. nly mechanisms Ball pen Retrieving Retrieving + filtering (%) N(Q) N(Qg) CQ (%) N(X) N(Xg) CX (%) 6 225 22 9.78 108 17 15.74 3 219 18 8.22 114 12 10.53 4 247 23 9.31 139 18 12.95 7 211 18 8.53 94 11 11.70 7 198 15 7.58 78 11 14.10 6 220 19.2 8.73 106.6 13.8 12.95 Table 7 Generated product ideas for ball pen Lipstick Compass Clinical thermometer Laser pointer Baton Electric torch Eyebrow pencil Mp3 Ultraviolet rays tester Massage stick GPRS USB Flash memory drive Voice recorder Pill box Pregnancy test pen Swatch (Timer) LED Language translator Table 6 Generated product ideas for cell phone Alcohol tester MP3 Language learning machine Pulse detector e-Map PDA (personal digital assistant) Radio and Walkman RFID Mosquito prevention set Electronic pet (Game) e-Book USB Flash memory drive Stun gun for security Cosmetic box Language translator/Dictionary Anti-camera detector Pedometer Remote controller for door/car Clinical thermometer Digital wallet OBU (car navigation system) Barcode scanner Strobe light M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 539 7. Concluding remarks This research presents a case-based reasoning (CBR) approach combined with the fuzzy AHP method to generate new product ideas that tend to be valuable for enhancing the baseline product. The generation of new product ideas is through a retrieving-and-filtering mecha- nism operated on a product database, where a product is modeled by 100 attributes—87 ones model the use-scenario and the other 13 model the manufacturing-and-recycling features. The retrieving mechanism is a fuzzy CBR technique that utilizes the use-scenario attributes for retrieving prod- uct-ideas. The retrieved product-ideas are subsequently screened by a filtering mechanism – a CBR technique that utilizes the manufacturing-and-recycling attributes as the fil- tering criteria. The filtering mechanism is proposed to filter out less valuable product ideas in order to save time for subsequent product-idea evaluation by experts. A prototype system has been implemented for justifying the contribution of the filtering mechanism. Experiments show that the retrieving-and-filtering mechanism outper- forms the prior retrieving-only mechanism in terms of creative ratio. That is, the retrieving-and-filtering mecha- nism generates higher percentage of ‘‘good product ideas’’ as opposed to that of the retrieving-only mechanism. Appendix 1. Arithmetical operators for fuzzy numbers The fuzzy arithmetic operators used in this research are defined below (Laarhoven & Pedrycz, 1983; Zadeh, 1975), by referring two fuzzy numbers ~a1 ¼ðl1; m1; r1Þ and ~a2 ¼ðl2; m2; r2Þ. (1) Addition operator: � ~a1 � ~a2 ¼ðl1 þ l2; m1 þ m2; r1 þ r2Þ (2) Subtraction operator: � ~a1 � ~a2 ¼ðl1 � l2; m1 � m2; r1 � r2Þ (3) Multiplication operator: � ~a1 � ~a2 ¼ðl1 l2; m1 m2; r1 r2Þ (4) Division operator: / ~a1=~a2 ¼ l1 r2 ; m1 m2 ; r1 l2 � � (5) Inverse power operators ~a�1=n1 ¼ðr �1=n 1 ; m �1=n 1 ; l �1=n 1 Þ (6) Distance of two fuzzy numbers (Chen, 2000) Dð~a1;~a2Þ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1=3½ðl1 �l2Þ 2 �ðm1 �m2Þ 2 �ðr1 �r2Þ 2� q (7) Defuzzication operator (Teng & Tzeng, 1993) Defuzzyð~a1Þ¼ jðr1 � l1Þþðm1 � l1Þj=3 þ l1 References Aamodt, A., & Plaza, E. (1994). Case-based reasoning: foundational issues, methodological variations, and system approaches. AI Com- munications Journal, 7(1), 39–59. Belecheanu, R., Pawar, K. S., Barson, R. J., Bredehorst, B., & Weber, F. (2003). The application of case-based reasoning to decision support in new product development. Integrated Manufacturing System, 14(1), 36–45. Besemer, S., & O’Quin, K. (1986). Analysis of creative products: refinement and test of a judging instrument. Journal of Creative Behavior, 20(2), 115–126. Bilgic, T., & Fox, M. S. (1996). Constraint-based retrieval of engineering design cases: context as constraints. In J. Gero & F. Sudweeks (Eds.), Artificial intelligence in design ’96 (pp. 269–288). Dordrecht: Kluwer Academic Publishers. Buckley, J. J. (1985). Ranking alternatives using fuzzy numbers. Fuzzy Sets and Systems, 15(1), 21–31. Chan, F. T. S. (2005). Application of a hybrid case-base reasoning approach in electroplating industry. Expert Systems with Application, 29, 121–130. Chen, C. T. (2000). Extensions of the TOPSIS for group decision-making under fuzzy environment. Fuzzy Sets and Systems, 114, 1–9. Chiou, H. K., Tzeng, G. H., & Cheng, D. C. (2005). Evaluating sustainable fishing development strategies using fuzzy MCDM approach. Omega, 33, 223–234. Creax.com. (2006). Available from http://www.creax.com/tools.htm. de Mantaras, R. L., & Plaza, E. (1997). Case base reasoning: an overview, IIA research report 96. AI Communications Journal, 10, 21–29. Goldfire Innovator: Invention-machine.com. (2006). Available from http://www.invention-machine.com/. Higgins, J. M. (1994). 101 Creative problem solving techniques: The handbook of new ideas for business. New Management Pub. Co. Hirota, K., Yoshino, H., Xu, M. Q., Zhu, Y., Li, X. Y., & Horie, D. (1997). An application of fuzzy theory to the case-based reasoning of the CISG. Journal of Advanced Computational Intelligence and Intel- ligent Informatics, 1(2), 86–93. Hisrich, R. D., Ingram, T. N., & Peters, M. P. (1991). Marketing decisions for new and mature products, 2/e. Maxwell Macmillam (pp. 172–174). Kolodner, J. L., & Leake, D. B. (1996). A tutorial introduction to case- based reasoning. In D. Leake (Ed.), Experiences lessons and future directions (pp. 31–66). Menlo Park, CA: AAAI Press. http://www.creax.com/tools.htm. http://www.invention-machine.com/ 540 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 Kotler, P. (1994). Marketing management: Analysis, planning, and control, 8/e. Englewood Cliffs, NJ: Prentice-Hall (pp. 319–328). Laarhoven, P. J. M., & Pedrycz, W. (1983). A fuzzy extension of Saaty’s priority theory. Fuzzy Sets and Systems, 11(3), 229–241. Linda, R. (1991). Generating and screening new products ideals. Industrial Marketing Management, 20, 287–296. Maher, M. L., Balachandran, M. B., & Zhang, D. M. (1995). Case-based reasoning in design. Hillsdale, NJ: Lawrence Erlbaum. Maher, M. L., & Garza, A. G. S. (1997). Case-based reasoning in design. IEEE Expert(March–April), 34–41. Mann, D. (2003). Hand-on systematic innovation. CREAX Press. Marling, C., Sqalli, M., Rissland, E., Munoz–Avila, H., & Aha, D. (2002). Case-based reasoning integrations. AI Magazine, 23(1), 69–86. Moon, J. H., & Kang, C. S. (2001). Application of fuzzy decision making method to the evaluation of spent fuel storage options. Progress in Nuclear Energy, 39(3–4), 345–351. Nijssen, E. J., & Lieshout, K. F. M. (1995). Awareness, use, and effectiveness of models and methods for new product development. European Journal of Marketing, 29(10), 27–43. Norman, D. A., & Draper, S. W. (1986). User-centered system design: New perspectives on human–computer interaction. Hillsdale, NJ: Lawrence Earlbaum Associates. Osborn, A. (1963). Applied imagination. New York: Charle Scribners Son (pp. 151–165). Rantanen, K., & Domb, E. (2002). Simplified TRIZ: New product-solving applications for engineers and manufacturing professionals. ST. Lucie Press. Ruet, M., & Geneste, L. (2002). Search and adaptation in fuzzy object case-base. Advances in case-based reasoning. In Proceedings of the sixth European conference, ECCBR 2002, Aberdeen, Scotland, United Kingdom, September 4–7, 2002 (pp. 350–364). Saaty, T. L. (1980). The analytic hierarchy process: Planning, priority setting, resource allocation. New York: McGraw-Hill, p. 20. Teng, J. Y., & Tzeng, G. H. (1993). Transportation investment project selection with fuzzy multi-objective. Transportation Planning and Technology, 17, 91–112. Trisolver.com. (2006). Available from http://www.trisolver.com/. Trousse, B., & Visser, W. (1993). Use of case-based reasoning techniques for intelligent computer-aided-design systems. In Proceedings of the IEEE international conference on systems, man and cybernetics (Vol. 3, pp. 513–517). Watson, I., & Marir, F. (1994). Case-based reasoning: a review. The Knowledge Engineering Review, 9(4), 327–354. Wu, M. C., Lo, Y. F., & Hsu, S. H. (2006). A case-based reasoning approach to generating new product ideas. Journal of Advance Manufacturing and Technology, 30, 166–173. Zadeh, L. A. (1975). The concept of a linguistic variable and its application to approximate reasoning. Information Sciences, part1:8, 199–249, part2:8, pp. 301–357, part3: pp. 43–80. http://www.trisolver.com/ A fuzzy CBR technique for generating product ideas Introduction Case-based reasoning Product database Product attributes for creating ideas Product attributes for screening ideas Linguistic variables for describing product attributes Weighting each dimension of target customers ' use-scenario Retrieving and filtering product ideas Retrieving mechanism Filtering mechanism Experiments Generating product ideas Comparison of product-idea-generating mechanisms Concluding remarks Arithmetical operators for fuzzy numbers References 
work_yn7bzumlpjdarpfazw5wvrnhwq	----	Evolutionary RBF classifier for polarimetric SAR images Expert Systems with Applications 39 (2012) 4710–4717 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Evolutionary RBF classifier for polarimetric SAR images Turker Ince a,⇑, Serkan Kiranyaz b, Moncef Gabbouj b a Izmir University of Economics, Electronics and Telecommunications Engineering Department, Izmir, Turkey b Tampere University of Technology, Department of Signal Processing, Tampere, Finland a r t i c l e i n f o a b s t r a c t Keywords: Polarimetric synthetic aperture radar Radial basis function network Particle swarm optimization 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.09.082 ⇑ Corresponding author. Tel.: +90 2324888509; fax E-mail address: turker.ince@ieu.edu.tr (T. Ince). In this paper, a robust radial basis function (RBF) network based classifier is proposed for polarimetric synthetic aperture radar (SAR) images. The proposed feature extraction process utilizes the covariance matrix elements, the H/a/A decomposition based features combined with the backscattering power (span), and the gray level co-occurrence matrix (GLCM) based texture features, which are projected onto a lower dimensional feature space using principal components analysis. For the classifier training, both conventional backpropagation (BP) and multidimensional particle swarm optimization (MD-PSO) based dynamic clustering are explored. By combining complete polarimetric covariance matrix and eigenvalue decomposition based pixel values with textural information (contrast, correlation, energy, and homoge- neity) in the feature set, and employing automated evolutionary RBF classifier for the pattern recognition unit, the overall classification performance is shown to be significantly improved. An experimental study is performed using the fully polarimetric San Francisco Bay and Flevoland data sets acquired by the NASA/ Jet Propulsion Laboratory Airborne SAR (AIRSAR) at L-band to evaluate the performance of the proposed classifier. Classification results (in terms of confusion matrix, overall accuracy and classification map) compared with the major state of the art algorithms demonstrate the effectiveness of the proposed RBF network classifier. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Image and data classification techniques play an important role in the automatic analysis and interpretation of remote sensing data. Particularly polarimetric synthetic aperture radar (SAR) data poses a challenging problem in this field due to complexity of mea- sured information from its multiple polarimetric channels. Re- cently, the number of applications which use data provided by the SAR systems having fully polarimetric capability have been increasing. Over the past decade, there has been extensive research in the area of the segmentation and classification of polarimetric SAR data. In the literature, the classification algorithms for polari- metric SAR can be divided into three main classes: (1) classification based on physical scattering mechanisms inherent in data (Pottier & Lee, 2000; van Zyl, 1989), (2) classification based on statistical characteristics of data (Lee et al., 1999; Wu, Ji, Yu, & Su, 2008) and (3) classification based on image processing techniques (Ince, 2010; Tan, Lim, & Ewe, 2007; Ye & Lu, 2002). Additionally, there has been several works using some combinations of the above clas- sification approaches (Lee et al., 1999; Pottier & Lee, 2000). While these approaches to the polarimetric SAR classification problem can be based on either supervised or unsupervised methods, their ll rights reserved. : +90 2324888475. performance and suitability usually depend on applications and the availability of ground truth. As one of the earlier algorithms, Kong, Swartz, Yueh, Novak, and Shin (1988) derived a distance measure based on the complex Gaussian distribution and used it for maximum-likelihood (ML) classification of single-look complex polarimetric SAR data. Then, Lee, Grunes, and Kwok (1994) used the statistical properties of a fully polarimetric SAR to perform a supervised classification based on complex Wishart distribution. Afterwards, Cloude and Pottier (1997) proposed an unsupervised classification algorithm based on their target decomposition theory. Target entropy (H) and target average scattering mechanism (scattering angle, a) calculated from this decomposition have been widely used in polarimetric SAR clas- sification. For multilook data represented in covariance or coherency matrices, Lee et al. (1999) proposed a new unsupervised classifica- tion method based on combination of polarimetric target decompo- sition (Cloude & Pottier, 1997) and the maximum likelihood classifier using the complex Wishart distribution. The unsupervised Wishart classifier has an iterative procedure based on the well- known K-means algorithm, and has become a preferred benchmark algorithm due to its computational efficiency and generally good performance. However, this classifier still has some significant drawbacks since it entirely relies on K-means for actual clustering, such as it may converge to local optima, the number of clusters should be fixed a priori, its performance is sensitive to the http://dx.doi.org/10.1016/j.eswa.2011.09.082 mailto:turker.ince@ieu.edu.tr http://dx.doi.org/10.1016/j.eswa.2011.09.082 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa T. Ince et al. / Expert Systems with Applications 39 (2012) 4710–4717 4711 initialization and its convergence depends on several parameters. Recently, a two-stage unsupervised clustering based on the EM algo- rithm (Khan, Yang, & Zhang, 2007) is presented for classification of polarimetric SAR images. The EM algorithm estimates parameters of the probability distribution functions which represent the ele- ments of a 9-dimensional feature vector, consisting of six magni- tudes and three angles of a coherency matrix. Markov random field (MRF) clustering based method (Tran, Wehrens, Hoekman, & Buydens, 2005) exploiting the spatial relation between adjacent pix- els in polarimetric SAR images has been presented. In (Ye & Lu, 2002), a new wavelet-based texture image segmentation algorithm is suc- cessfully applied to unsupervised SAR image segmentation problem. More recently, neural network based approaches (Yang, Wang, & Jiao, 2009; Zhang, Wu, & Wei, 2009; Zhang, Zou, Zhang, & Zhang, 2010) for classification of polarimetric synthetic aperture radar data have been shown to outperform other aforementioned well-known techniques. Compared with other approaches, neural network clas- sifiers have the advantage of adaptability to the data without mak- ing a priori assumption of a particular probability model or distribution. However, their performance depends on the network structure, training data, initialization, and parameters. Designing an optimal ANN classifier structure and its parameters to maximize the classification accuracy is still a crucial and challenging task. In this study, RBF network classifier which is optimally designed by the evolutionary search technique, multidimensional particle swarm optimization (MD-PSO) (Kiranyaz, Ince, Yildirim, & Gabbouj, 2010), is employed. RBFs are chosen due to their robustness, faster learning capability compared with other feedforward networks, and superior performance with simpler network architectures. Ear- lier work on RBF classifiers for polarimetric SAR image classification has demonstrated a potential for performance improvement over conventional techniques (Ince, 2010). The proposed polarimetric SAR feature vector includes full covariance matrix, the H/a/A decomposition based features combined with the backscattering power (Span), and the gray level co-occurrence matrix (GLCM) based texture features as suggested by the results of previous studies (Clausi & Yue, 2004; Ersahin, Scheuchl, & Cumming, 2004). The per- formance of the proposed RBF network based classifier is evaluated using the fully polarimetric San Francisco Bay and Flevoland data sets acquired by the NASA/Jet Propulsion Laboratory Airborne SAR (AIRSAR) at L-band. The classification results (in terms of confusion matrix, overall accuracy and classification map) are compared with competing state of the art classifiers. The rest of the paper is organized as follows. Section 2 briefly presents the basic theory of polarimetric SAR for this paper includ- ing the Cloude–Pottier decomposition. The feature extraction methodology for the proposed polarimetric SAR image classifica- tion system is described in Section 3. Then, the RBF network funda- mentals, its training algorithms, and an overview of the proposed classifier technique are presented in Section 4. Section 5 describes the experimental test results on real polarimetric SAR data. Finally, Section 6 concludes the paper. 2. Polarimetric sar data processing Polarimetric radars often measure the complex scattering ma- trix, [S], produced by a target under study with the objective to infer its physical properties. Assuming linear horizontal and vertical polarizations for transmitting and receiving, [S] can be expressed as S ¼ Shh Shv Svh Svv � � ð1Þ Reciprocity theorem applies in a monostatic system configura- tion, Shv = Svh. For coherent scatterers only, the decompositions of the measured scattering matrix [S] can be employed to character- ize the scattering mechanisms of such targets. One way to analyze coherent targets is the Pauli decomposition (Lee et al., 1999), which expresses [S] in the so-called Pauli basis ½S�a ¼ 1ffiffi2p 1 00 1 � � ; � ½S�b ¼ 1ffiffi 2 p 1 0 0 �1 � � ; ½S�c ¼ 1ffiffi 2 p 0 1 1 0 � � g as, S ¼ Shh Shv Svh Svv � � ¼ a½S�a þ b½S�b þ c½S�c ð2Þ where a ¼ðShh þ SvvÞ= ffiffiffi 2 p ; b ¼ðShh � SvvÞ= ffiffiffi 2 p ; c ¼ ffiffiffi 2 p Shv . Hence, by means of the Pauli decomposition, all polarimetric information in [S] could be represented in a single RGB image by combining the intensities |a|2, |b|2 and |c|2, which determine the power scattered by different types of scatterers such as single- or odd-bounce scat- tering, double- or even-bounce scattering, and orthogonal polariza- tion returns by the volume scattering. There are several other coherent decomposition theorems such as the Krogager decomposi- tion, the Cameron decomposition, and SDH (Sphere, Diplane, Helix) decomposition all of which aim to express the measured scattering matrix by the radar as the combination of scattering responses of coherent scatterers. Alternatively, the second order polarimetric descriptors of the average polarimetric covariance h[C]i and the coherency h[T]i matrices can be derived from the scattering matrix and employed to extract physical information from the observed scattering pro- cess. The elements of the covariance matrix, [C], can be written in terms of three unique polarimetric components of complex scat- tering matrix: C11 ¼ Shh S � hh; C21 ¼ S � hhShv C22 ¼ Shv S � hv; C32 ¼ S � hhSvv C33 ¼ Svv S�vv; C11 ¼ S � hh Svv ð3Þ For single-look processed polarimetric SAR data, the three polarimetric components (HH, HV, and VV) has a multivariate complex Gaussian distribution and the complex covariance ma- trix form has a complex Wishart distribution (Lee et al., 1994). Due to presence of speckle noise and random vector scattering from surface or volume, polarimetric SAR data are often multi- look processed by averaging n neighboring pixels. By using the Pauli based scattering matrix for a pixel i, ki ¼ ½Shh þ Svv; Shh� Svv; 2Shv� T = ffiffiffi 2 p , the multi-look coherency matrix, h[T]i, can be written as hTi¼ 1 n Xn i¼1 ki k �T i ð4Þ Both coherency h[T]i and covariance h[C]i are 3 � 3 Hermitian positive semidefinite matrices, and since they can be converted into one another by a linear transform, both are equivalent repre- sentations of the target polarimetric information. The incoherent target decomposition theorems such as the Freeman decomposition, the Huynen decomposition, and the Cloude–Pottier (or H/a/A) decomposition employ the second or- der polarimetric representations of PolSAR data (such as covari- ance matrix or coherency matrix) to characterize distributed scatterers. The H/a/A decomposition (Cloude & Pottier, 1996) is based on eigen analysis of the polarimetric coherency matrix, h[T]i: hTi¼ k1e1e�T1 þ k2 e2e �T 2 þ k3 e3 e �T 3 ð5Þ where k1 > k2 > k3 P 0 are real eigenvalues and the corresponding orthonormal eigenvectors ei (representing three scattering mecha- nisms) are ei ¼ ei/i½cos ai; sin ai cos bi eidi ; sin ai sin bi eici� T ð6Þ 4712 T. Ince et al. / Expert Systems with Applications 39 (2012) 4710–4717 Cloude and Pottier defined entropy H, average of set of four an- gles �a, �b, �d, and �c, and anisotropy A for analysis of the physical information related to the scattering characteristics of a medium: H ¼� X3 i¼1 pi log3pi where pi ¼ kiP3 i¼1ki ð7Þ �a ¼ X3 i¼1 piai; �b ¼ X3 i¼1 pibi; �d ¼ X3 i¼1 pidi; �c ¼ X3 i¼1 pici ð8Þ A ¼ p2 � p3 p2 þ p3 ð9Þ For a multi-look coherency matrix, the entropy, 0 6 H 6 1, rep- resents the randomness of a scattering medium between isotropic scattering (H = 0) and fully random scattering (H = 1), while the average alpha angle can be related to target average scattering mechanisms from single-bounce (or surface) scattering ð�a � 0Þ to dipole (or volume) scattering ð�a � p=4Þ to double-bounce scatter- ing ð�a � p=2Þ. Due to basis invariance of the target decomposition, H and �a are roll invariant hence they do not depend on orientation of target about the radar line of sight. Additionally, information about target’s total backscattered power can be determined by the span as span ¼ X3 i¼1 ki ð10Þ Entropy (H), estimate of the average alpha angle (�a), and span calcu- lated by the above noncoherent target decomposition method have been commonly used as polarimetric features of a scatterer in many target classification schemes (Fang, Wen, & Yirong, 2006; Lee et al., 1999). 3. Feature extraction The proposed feature extraction process utilizes the complete covariance matrix information, the gray level co-occurrence ma- trix (GLCM) based texture features, and the backscattering power (span) combined with the H/a/A decomposition (Cloude & Pottier, 1997). The feature vector from the Cloude–Pottier decomposition includes entropy (H), anisotropy (A), estimates of the set of average angles (�a, �b, �d, and �c), three real eigen- values (k1; k2; k3 ), and span. As suggested by the previous studies (Clausi, 2002; Zhang et al., 2009) appropriate texture measures for SAR imagery based on the gray level co-occurrence probabil- ities are included in the feature set to improve its discrimination power and classification accuracy. In this study, contrast, correla- tion, energy, and homogeneity features are extracted from nor- malized GLCMs which are calculated using interpixel distance of 2 and averaging over four possible orientation settings (h = 0�, 45�, 90�, 135�). To reduce the dimensionality (and redun- dancy) of input feature space, the principal components trans- form is applied to these inputs and the most principal components (which contain about 95% of overall energy in the original feature matrix) are then selected to form a resultant fea- ture vector for each imaged pixel. Dimensionality reduction of input feature information improves efficiency of learning for a neural network classifier due to a smaller number of input nodes (to avoid curse of dimensionality) (Pittner & Kamarthi, 1999) and reduces computation time. For the purpose of normalizing and scaling the feature vector, each feature dimension is first nor- malized to have a zero mean and unity standard deviation be- fore principal component analysis (PCA) is applied, and following the PCA outputs are linearly scaled into [�1, 1] interval. 4. RBF neural networks An artificial neural network (ANN) consists of a set of connected processing units, usually called neurons or nodes. ANNs can be de- scribed as directed graphs, where each node performs some activa- tion function to its inputs and then gives the result forward to be the input of some other neurons until the output neurons are reached. ANNs can be divided into feedforward and recurrent net- works according to their connectivity. In a recurrent ANN there can be backward loops in the network structure, while in feedforward ANNs such loops are not allowed. A popular type of feedforward ANN is the radial basis function (RBF) network (Poggio & Girosi, 1989), which has always two layers in addition to the passive input layer: a hidden layer of RBF units and a linear output layer. Only the output layer has connection weights and biases. The activation function of the kth RBF unit is defined as yk ¼ u kX � lkk r2k � � ð11Þ where u is a radial basis function or, in other words, a strictly posi- tive radially symmetric function, which has a unique maximum at N-dimensional center lk and whose value drops rapidly close to zero away from the center. rk is the width of the peak around the center lk. The activation function gets noteworthy values only when the distance between the N-dimensional input X and the cen- ter lk, kX � lkk is smaller than the width rk. The most commonly used activation function in RBF networks is the Gaussian basis func- tion defined as yk ¼ exp � kX � lkk 2 2r2k ! ð12Þ where lk and rk are the mean and standard deviation, respectively, and ||�|| denotes the Euclidean norm. More detailed information about RBF networks can be obtained from Poggio and Girosi (1989) and Haykin (1998). In this study, two distinct training methods for RBF network classifiers, the traditional backpropagation (BP) and particle swarm optimization (PSO) are investigated. For the BP algorithm, RPROP enhancement is used when training RBF networks. The main differ- ence in RPROP is that it modifies the update-values for each parameter according to the sequence of signs of partial derivatives. This only leads to a faster convergence, while the problems of a hill-climbing algorithm are not solved. Further details about BP and RPROP can be found in Chauvin and Rumelhart (1995) and Riedmiller and Braun (1993), respectively. In order to determine (near-) optimal network architecture for a given problem, we apply exhaustive BP training over every network configuration in the architecture space defined. For PSO-based training, the proposed approach is to apply multi-dimensional particle swarm optimiza- tion (MD-PSO) based dynamic clustering (Kiranyaz, Ince, Yildirim, & Gabbouj, 2009) to determine the optimal (with respect to mini- mizing a given cost function for the input–output mapping) num- ber of Gaussian neurons with their correct parameters (centroids and variances). Afterwards, BP can conveniently be used to com- pute the remaining network parameters, weights (w) and bias (h) of the each output layer neuron. The overview of the proposed classifier for polarimetric SAR image is shown in Fig. 1. 5. Experimental results In this section, two test images of an urban area (San Fran- cisco Bay, CA) and an agricultural area (Flevoland in the Nether- lands), both acquired by the NASA/Jet Propulsion Laboratory’s Airborne SAR (AIRSAR) at L-band, were chosen for performance evaluation of the proposed RBF network classifier. Both data sets Pre-Processing [C] Covariance Matrix Cloude-Pottier Decomposition + Span Lee Speckle Filter GLCM Texture Features Feature Extraction P C A T ransform Post-Processing {μ, σ, Ν, ω, θ} Expert Labeling Tr ai ni ng S et MD PSO Dynamic Clustering RBFN Classifier Classification Map [T] Coherency Matrix Fig. 1. Overview of the evolutionary RBF network classifier design for polarimetric SAR image. Fig. 2. Pauli image of 600 � 600 pixel sub-area of San Francisco Bay (left) with the 5 � 5 refined Lee filter used. The training and testing areas for three classes are shown using red rectangles and circles respectively. The aerial photograph for this area (right) provided by the US Geological Survey taken on October, 1993 can be used as ground- truth. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) T. Ince et al. / Expert Systems with Applications 39 (2012) 4710–4717 4713 have been widely used in the polarimetric SAR literature over the last two decades (Ersahin et al., 2004; Ferro-Famil, Pottier, & Lee, 2001; Fukuda & Hirosawa, 1999), and distributed as mul- ti-look processed and publicly available through the polarimetric SAR data processing and educational tool (PolSARpro) by ESA (The Polarimetric SAR Data Processing and Educational Tool (PolSARPro). The original four-look fully polarimetric SAR data of the San Francisco Bay, having a dimension of 900 � 1024 pix- els, provides good coverage of both natural (sea, mountains, for- ests, etc.) and man-made targets (buildings, streets, parks, golf course, etc.) with a more complex inner structure. For the pur- pose of comparing the classification results with the Wishart (Lee et al., 1999) and the NN-based (Zhang et al., 2009) classifi- ers, the sub-area (Fig. 2) with size 600 � 600 is extracted and used. The aerial photographs for this area which can be used as ground-truth are provided by the TerraServer Web site (U.S. Geological Survey Images). In this study, no speckle filtering is applied to originally four-look processed covariance matrix data and before GLCM based texture feature generation to retain the resolution and to preserve the texture information. However, additional averaging, such as using the polarimetry preserving refined Lee filter (Lee, Grunes, & de Grandi, 1999) with 5 � 5 window, of coherency matrix should be performed prior to the Cloude–Pottier decomposition (Cloude & Pottier, 1997). For MD-PSO based clustering algorithm, the typical internal PSO parameters (c1, c2 and w) are used as in Shi and Eberhart (1998), also explained in Kiranyaz, Ince, Yildirim, and Gabbouj (2009). For all experiments in this section, the two critical PSO parameters, swarm size (S) and number of iterations (IterNo), are set as 40 and 1000, respectively. Fig. 3. The classification results of the proposed RBF-PSO technique on the extracted 600 � 600 sub-image of San Francisco Bay (black denotes sea, gray urban areas, white vegetated zones). Fig. 4. The classification results of the proposed RBF-MDPSO technique for the 4714 T. Ince et al. / Expert Systems with Applications 39 (2012) 4710–4717 To test the performance of the proposed classifier and compare its classification results, the same training and testing areas for the three classes from the sub San Francisco area (as shown on the Pauli-based decomposition image in Fig. 2), the sea (15,810, 6723 pixels respectively), urban areas (9362, 6800), and the vegetated zones (5064, 6534), which are manually selected in an earlier study (Zhang et al., 2009), are used. The confusion matrix of the proposed evolutionary RBF method on the training and testing areas are gi- ven in Table 1. The classification accuracy values are averaged over 10 independent runs. From the results, the main drawback of the proposed method is the separation of vegetated zones from urban areas. Compared to two other competing techniques, the proposed method is able to differentiate better the uniform areas corre- sponding to main classes of scattering such asthe ocean, vegeta- tion, and building areas. In Table 2, the overall accuracies in training and testing areas for the proposed RBF classifier trained using the BP and MD-PSO algorithms and two competing methods, the Wishart Maximum Likelihood (WML) classifier (Lee et al., 1999) and the NN-based classifier (Zhang et al., 2009), are com- pared. The average accuracies over 10 independent runs for the best configuration of the RBF-BP and RBF-PSO classifiers are re- ported. The proposed RBF classifier trained by the global PSO algo- rithm is superior to the NN-based, WML, and RBF-BP based methods with higher accuracies in both training (99.50%) and test- ing (98.96%) areas. Fig. 3 shows the classification results on the whole sub-area image for the RBF-PSO based classifier. The classi- fication map of the whole San Francisco Bay image produced by the same classifier is given in Fig. 4 for a qualitative (visual) perfor- mance evaluation. The evolutionary RBF classifier has the structure of 11 input neurons, 21 Gaussian neurons which the cluster cen- troids and variance (lk and rk) are determined by MD-PSO based dynamic clustering the training data, and 3 output neurons. The classification results in Table 2 have been produced by using a high percentage (60%) of total (training and testing com- bined) pixels for training. The proposed classifier is also tested by limiting the percentage of total pixels which were used for classi- fier training to less than 1% of the total pixels to be classified. The results over the same testing data set are shown in Table 3. In this case, the RBF network classifier trained by the BP or MD-PSO algo- rithms performed still at a high level, achieving accuracies over 95% and 98% respectively. Generally, a relatively smaller training data set can avoid over-fitting and improve generalization perfor- mance of a classifier over larger data sets. In order to test robustness of the proposed RBF network classi- fier trained by the MD-PSO based dynamic clustering, 20 indepen- dent runs are performed over the San Francisco area image and the Table 1 Summary table of pixel-by-pixel classification results of the proposed RBF-MDPSO method for the training and testing data of San Francisco Bay. Training data Test data Sea Urb Veg Sea Urb Veg Sea 14,264 4 0 6804 0 0 Urb 11 9422 22 10 6927 23 Veg 10 87 4496 21 162 6786 Table 2 Overall performance comparison (in percent) for San Francisco Bay dataset. The best performances are indicated in bold. Method Training area Testing area RBF-BP 98.00 95.70 WML (Lee et al., 1999) 97.23 96.16 NN (Zhang et al., 2009) 99.42 98.64 RBF-PSO 99.50 98.96 original (900 � 1024) San Francisco Bay image (black denotes sea, gray urban areas, white vegetated zones). Table 3 Overall performance (in percent) using smaller training set (<1% of total pixels) for San Francisco Bay dataset. Method Training area Testing area RBF-BP 100 95.60 RBF-PSO 100 98.54 resulting cluster number histogram is plotted in Fig. 5. Addition- ally, the plots of a typical run showing the fitness score and dimen- sion versus number of iterations for MD-PSO operation are presented in the left side of Fig. 5. Based on overall clustering re- sults, it is found that the number of clusters (the optimal number of Gaussian neurons) and their centroids extracted from the MD- PSO based dynamic clustering are generally consistent, indicating the proposed technique is robust (or repeatable). Fig. 5. Fitness score (left top) and dimension (left bottom) plots versus iteration number for a typical MD-PSO run. The resulting histogram plot (right) of cluster numbers which are determined by the proposed method. Table 4 Overall performance comparison (in percent) for Flevoland dataset. The best performances are indicated in bold. Method Training area Testing area ECHO (Chen et al., 2007) – 81.30 Wavelet-based (Fukuda & Hirosawa, 1999) – 88.28 RBF-BP 95.50 92.05 NN (Zhang et al., 2009) 98.62 92.87 RBF-PSO 95.55 93.36 T. Ince et al. / Expert Systems with Applications 39 (2012) 4710–4717 4715 Next, the proposed evolutionary RBF classifier with the sug- gested feature set has been applied to the polarimetric image of the Flevoland site, an agricultural area (consists of primarily crop fields and forested areas) in The Netherlands. This original four- look fully polarimetric SAR data has a dimension of 750 � 1024 pixels with 11 identified crop classes {stem beans, potatoes, lu- cerne, wheat, peas, sugar beet, rape seed, grass, forest, bare soil, and water}. The available ground truth for eleven classes can be found in Ainsworth, Kelly, and Lee (2009). To compare classifica- tion results the same eleven training and testing sets are used with those of the NN-based (Zhang et al., 2009), wavelet-based (Fukuda & Hirosawa, 1999), and ECHO (Chen, Li, Pang, & Tian, 2007) classi- fiers. In Table 4, the overall accuracies in training and testing areas of the Flevoland dataset for the proposed RBF classifier trained Table 5 Summary table of pixel-by-pixel classification results (in percent) of the proposed RBF-M Training (Testing) data Water Forest Stem beans Potatoes Lucerne Water 99(98) 0(0) 0(0) 0(0) 0(0) Forest 0(0) 95(97) 0(0) 0(0) 1(0) Stem beans 0(0) 0(0) 95(97) 0(0) 5(2) Potatoes 0(0) 0(0) 0(0) 99(96) 0(0) Lucerne 0(0) 0(0) 2(2) 0(0) 98(97) Wheat 0(0) 0(0) 0(0) 0(0) 2(4) Peas 0(0) 0(0) 0(0) 0(0) 0(0) Sugar beet 0(0) 0(0) 0(0) 0(0) 0(0) Bare soil 0(0) 0(0) 0(0) 0(0) 0(0) Grass 0(0) 0(0) 0(0) 0(0) 1(0) Rape seed 2(2) 0(0) 0(0) 0(0) 0(0) using the BP and MD-PSO algorithms and three state of the art methods, the ECHO (Chen et al., 2007), wavelet-based (Fukuda & Hirosawa, 1999), and NN-based (Zhang et al., 2009) classifiers, are shown. The overall classification accuracies of the proposed RBF-based classifier framework are quite high. The percentage of correctly classified training and testing pixels in the Flevoland L- band image for the proposed evolutionary (MD-PSO) RBF method are given in Table 5. Fig. 6 shows the classification results of the proposed evolutionary RBF classifier for the Flevoland image. The computational complexity of the proposed method depends on the following distinct processes: the pre-processing stage, fea- ture extraction, post-processing, and RBF network classifier with MD-PSO dynamic clustering based training. Computation times of the first three stages are deterministic while a precise computa- tional complexity analysis for the RBF training stage is not feasible as the proposed dynamic clustering technique based on PSO is in stochastic nature. All experiments in this section are performed on a computer with P-IV 2.4 GHz CPU and 1 GB RAM. Based on our experiments, for the data of San Francisco Bay area with a dimension of 900 � 1024 data points (D = 921,600), it takes 30 min to perform feature extraction and necessary pre- and post-processing stages. Most of this time is used to compute the GLCM and four texture features calculated from it. For computa- tional complexity of RBF classifier training using MD-PSO process, there are certain attributes which directly affect the complexity DPSO method for the training and testing data of Flevoland. Wheat Peas Sugar beet Bare soil Grass Rape seed 0(0) 0(0) 0(0) 0(0) 0(0) 1(2) 0(0) 1(0) 1(0) 0(0) 2(3) 0(0) 0(1) 0(0) 0(0) 0(0) 0(0) 0(0) 0(0) 0(0) 1(4) 0(0) 0(0) 0(0) 0(0) 0(0) 0(0) 0(0) 0(1) 0(0) 91(86) 4(4) 1(3) 0(0) 2(3) 0(0) 1(0) 94(88) 2(7) 0(0) 0(0) 3(5) 0(2) 0(1) 95(91) 0(0) 4(5) 1(1) 0(2) 0(0) 0(0) 99(97) 0(0) 1(1) 0(1) 0(0) 2(4) 0(0) 97(95) 0(0) 2(2) 1(2) 3(2) 3(7) 0(0) 89(85) Fig. 6. The classification results of the proposed RBF-MDPSO technique on the L- band AIRSAR data over Flevoland. 4716 T. Ince et al. / Expert Systems with Applications 39 (2012) 4710–4717 such as swarm size (S), the number of iteration (IterNo) to termi- nate the MD PSO process, and the dimensions of data space (D). While the problem determines D, the computational complexity can still be controlled by S and IterNo settings. The further details of computational complexity analysis for the dynamic clustering technique based on MD-PSO can be found in Kiranyaz et al. (2010). For the same dataset, the average (over 10 runs) processing time to perform evolutionary RBF classifier training is found to be 30 min. 6. Conclusion This paper presents a new polarimetric SAR image classification framework which is based on an efficient formation of covariance matrix elements, H/a/A decomposition with the backscattered power (span) information, and GLCM based texture features, and the RBF network classifier. Two different learning algorithms, the classical backpropagation and multidimensional particle swarm optimization, were applied for the proposed classifier training. In addition to determining the correct network parameters, the latter evolutionary technique (MD-PSO) also finds the best RBF network architecture (optimum number of Gaussian neurons and their cen- troids) within an architecture space and for a given input data space. The overall classification accuracies and qualitative classifi- cation maps for the San Francisco Bay and Flevoland datasets dem- onstrate the effectiveness of the proposed classification framework using an evolutionary RBF network classifier. Based on the experi- mental results using real polarimetric SAR data, the proposed method performs well compared to several state-of-the-art classi- fiers, however, more experiments using large volume of available data should be done for a general conclusion. The proposed tech- nique employs evolutionary MD-PSO process for simultaneous training and evolution of RBF networks to achieve more accurate, robust, and automatic classification of polarimetric SAR images. References Ainsworth, T. L., Kelly, J. P., & Lee, J.-S. (2009). Classification comparisons between dual-pol, compact polarimetric and quad-pol SAR imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 64, 464–471. Chauvin, Y., & Rumelhart, D. E. (1995). Back propagation: theory, architectures, and applications. UK: Lawrence Erlbaum Associates Publishers. Chen, E., Li, Z., Pang, Y., & Tian, X. (2007). Quantitative evaluation of polarimetric classification for agricultural crop mapping. Photogrammetric Engineering and Remote Sensing, 73(3), 279–284. Clausi, D. A. (2002). An analysis of co-occurrence texture statistics as a function of grey level quantization. Canadian Journal of Remote Sensing, 28(1), 45–62. Clausi, D. A., & Yue, B. (2004). Comparing co-occurrence probabilities and Markov random fields for texture analysis of SAR sea ice imagery. IEEE Transactions on Geoscience and Remote Sensing, 42(1), 215–228. Cloude, S. R., & Pottier, E. (1996). A review of target decomposition theorems in radar polarimetry. IEEE Transactions on Geoscience and Remote Sensing, 34(2), 498–518. Cloude, S. R., & Pottier, E. (1997). An entropy based classification scheme for land applications of polarimetric SAR. IEEE Transactions on Geoscience and Remote Sensing, 35, 68–78. Ersahin, K., Scheuchl, B., & Cumming, I. (2004). Incorporating texture information into polarimetric radar classification using neural networks. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, Anchorage, USA, September (Vol. 1, pp. 560–563). Fang, C., Wen, H., & Yirong, W. (2006). An improved Cloude–Pottier decomposition using H/a/span and complex Wishart classifier for polarimetric SAR classification. In Proceedings of the CIE October (pp. 1–4). Ferro-Famil, L., Pottier, E., & Lee, J. S. (2001). Unsupervised classification of multifrequency and fully polarimetric SAR images based on the H/A/Alpha- Wishart classifier. IEEE Transactions on Geoscience and Remote Sensing, 39(11), 2332–2342. Fukuda, S., & Hirosawa, H. (1999). A wavelet-based texture feature set applied to classification of multifrequency polarimetric SAR images. IEEE Transactions on Geoscience and Remote Sensing, 37(5), 2282–2286. US Geological Survey Images. <http://www.terraserver-usa.com>. Haykin, S. (1998). Neural networks: A comprehensive foundation. USA: Prentice hall. Ince, T. (2010). Polarimetric SAR image classification using a radial basis function neural network. In Proceedings of The Progress in Electromagnetics Research Symposium (PIERS 2010), Cambridge, USA, July. Ince, T. (2010). Unsupervised classification of polarimetric SAR image with dynamic clustering: An image processing approach. Advances in Engineering Software, 41(4), 636–646. Khan, K. U., Yang, J., & Zhang, W. (2007). Unsupervised classification of polarimetric SAR images by EM algorithm. IEICE Transactions on Communications, 90(12), 3632–3642. Kiranyaz, S., Ince, T., Yildirim, A., & Gabbouj, M. (2009). Evolutionary artificial neural networks by multi-dimensional particle swarm optimization. Neural Networks, 22(10), 1448–1462. Kiranyaz, S., Ince, T., Yildirim, A., & Gabbouj, M. (2010). Fractional particle swarm optimization in multi-dimensional search space. IEEE Transactions on Systems, Man, and Cybernetics – Part B, 40(2), 298–319. Kong, J. A., Swartz, A. A., Yueh, H. A., Novak, L. M., & Shin, R. T. (1988). ‘‘Identification of terrain cover using the optimum polarimetric classifier. Journal of Electromagnetic Waves and Applications, 2(2), 171–194. Lee, J. S., Grunes, M. R., Ainsworth, T., Du, L.-J., Schuler, D., & Cloude, S. R. (1999). Unsupervised classification using polarimetric decomposition and the complex Wishart classifier. IEEE Transactions on Geoscience and Remote Sensing, 37(5), 2249–2257. Lee, J. S., Grunes, M. R., & de Grandi, G. (1999). Polarimetric SAR speckle filtering and its implications for classification. IEEE Transactions on Geoscience and Remote Sensing, 37(5), 2363–2373. Lee, J. S., Grunes, M. R., & Kwok, R. (1994). Classification of multi-look polarimetric SAR imagery based on complex Wishart distribution. International Journal of Remote Sensing, 15(11), 2299–2311. Pittner, S., & Kamarthi, S. V. (1999). Feature extraction from wavelet coefficients for pattern recognition tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21, 83–88. Poggio, T., & Girosi, F. (1989). A theory of networks for approximation and learning, A.I. Memo No. 1140, M.I.T. A.I Lab. The Polarimetric SAR Data Processing and Educational Tool (PolSARPro). <http:// www.earth.esa.int/polsarpro/datasets.html>. Pottier, E., & Lee, J. S. (2000). Unsupervised classification scheme of POLSAR images based on the complex Wishart distribution and the H/A/alpha-polarimetric decomposition theorem. In Proceedings of the 3rd EUSAR 2000 Conference May. Riedmiller, M., & Braun, H. (1993). A direct adaptive method for faster backpropagation learning: The RPROP algorithm. In Proceedings of the IEEE international conference on neural networks (pp. 586–591). Shi, Y., & Eberhart, R. C. (1998). A modified particle swarm optimizer. In Proceedings of the IEEE Congress on Evolutionary Computation (pp. 69–73). Tan, C. P., Lim, K. S., Ewe, H. T. (2007). Image processing in polarimetric SAR images using a hybrid entropy decomposition and maximum likelihood (EDML). In Proceedings of the International Symposium on Image and Signal Processing and Analysis (ISPA), September (pp. 418–422). Tran, T. N., Wehrens, R., Hoekman, D. H., & Buydens, L. M. C. (2005). Initialization of Markov random field clustering of large remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 43(8), 1912–1919. van Zyl, J. J. (1989). Unsupervised classification of scattering mechanisms using radar polarimetry data. IEEE Transactions on Geoscience and Remote Sensing, 27, 36–45. Wu, Y., Ji, K., Yu, W., & Su, Y. (2008). Region-based classification of polarimetric SAR images using Wishart MRF. IEEE Geoscience and Remote Sensing Letters, 5(4), 668–672. http://www.terraserver-usa.com http://www.earth.esa.int/polsarpro/datasets.html http://www.earth.esa.int/polsarpro/datasets.html T. Ince et al. / Expert Systems with Applications 39 (2012) 4710–4717 4717 Yang, S. Y., Wang, M., & Jiao, L. C. (2009). Radar target recognition using contourlet packet transform and neural network approach. Signal Processing, 89(4), 394–409. Ye, Zhen, Lu, Cheng-Chang (2002). Wavelet-based unsupervised SAR image segmentation using hidden markov tree models. In Proceedings of the16th international conference on pattern recognition (ICPR’02) (Vol. 2, p. 20729). Zhang, Y. D., Wu, L.-N., & Wei, G. (2009). A new classifier for polarimetric SAR images. Progress in Electromagnetics Research, PIER, 94, 83–104. Zhang, L., Zou, B., Zhang, J., & Zhang, Y. (2010). Classification of polarimetric SAR image based on support vector machine using multiple-component scattering model and texture features. EURASIP Journal on Advances in Signal Processing. doi:10.1155/2010/960831. http://dx.doi.org/10.1155/2010/960831 Evolutionary RBF classifier for polarimetric SAR images 1 Introduction 2 Polarimetric sar data processing 3 Feature extraction 4 RBF neural networks 5 Experimental results 6 Conclusion References 
work_yndx5yho5bcrlj242qnmjly7sy	----	Application of Expert System with Fuzzy Logic in Teachers’ Performance Evaluation (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 2, No.2, February 2011 51 | P a g e http://ijacsa.thesai.org/ Application of Expert System with Fuzzy Logic in Teachers‘ Performance Evaluation Abdur Rashid Khan 1 Institute of Computing & Information Technology (ICIT) Gomal Unversity, Dera Ismail Khan, KPK, Pakistan 1 dr.arashid@gu.edu.pk, drarashid.khan09@gmail.com 2 Hafeez Ullah Amin 2 Institute of Information Technology, Kohat University of Science & Technology Kohat 26000, KPK, Pakistan 2 hafeezullahamin@gmail.com Zia Ur Rehman 3 3 Institute of Information Technology, Kohat University of Science & Technology, Kohat 26000, KPK, Pakistan 3 ziagulzia@gmail.com Abstract— This paper depicts adaptation of expert systems technology using fuzzy logic to handle qualitative and uncertain facts in the decision making process. Human behaviors are mostly based upon qualitative facts, which cannot be numerically measured and hardly to decide correctly. This approach is an attempt to cope with such problems in the scenario of teachers’ performance evaluation. An Expert System was developed and applied to the acquired knowledge about the problem domain that showed interesting results providing a sketch for students and researchers to find solutions of such types of problems. Through Fuzzy Logic we numerically weighted the linguistic terms, like; very good, good, bad, or high, medium, low or satisfied, unsatisfied by assigning priorities to these qualitative facts. During final decision making, key parameters were given weights according to their priorities through mapping numeric results from uncertain knowledge and mathematical formulae were applied to calculate the numeric results at final. In this way this ES will not only be useful for decision-makers to evaluate teachers’ abilities but may also be adopted in writing Annual Confidential Reports (ACR) of about all the employees of an organization. Keywords— Expert System, Fuzzy Random Variables, Decision Making, Teachers’ Performance, Qualitative & Uncertain Knowledge. I. INTRODUCTION Computer programs using Artificial Intelligence (AI) techniques to assist people in solving difficult problems involving knowledge, heuristics, and decision-making are called expert systems, intelligent systems, or smart systems [18]. An expert system can be designed based on a set of rules to determine what action to set off when a certain situation is encountered [24]. In other words expert system is such a technology that able the human being to collect & control the human experts‘ knowledge and expertise in a particular problem domain for further use to solve similar problems through computer system. Experts of any fields are always few in numbers, expensive to consult and they have short time due to much work to do. So there is an urgent need of storing the expert‘s knowledge in the computer in such a way that have a great extent of knowledge of problem domain solving problems of the users and sparing experts for others works ―unpublished‖ [2]. In this paper we discuss the teachers‘ performance evaluation through AI technology at higher education institutions of Pakistan. The proposed Fuzzy Expert System considering various aspects of teachers attributes, like research & publication, teaching learning process, personal skills & abilities, compensation, achievements & recognition etc that have deep influence on the teachers‘ performance in universities investigated by [1]. In this paper a fuzzy expert system‘s model is designed to combine the knowledge and expertise of human experts with reasoning capabilities that will provide a great support to executives for decision-making in educational institutions. This paper is organized as: the section II discusses the applications of expert system & fuzzy logic in teachers‘ assessment and education, section III briefly describes the teachers‘ evaluation, and section IV explains the proposed approach for the solution of the entitled problem. II. STUDY BACKGROUND From last few decades, academics and researchers began to recognize the importance of expert system and its related concepts became one of the most popular topics related to decision making and knowledge management. From the beginning, expert systems have been developed in divers areas, like agriculture, chemistry, computer science, engineering, geology, medicine, space technology etc. [14]; and widely applied to various studies and issues, including performance assessment [3; 26; 27], commercial loan underwriting [19], logistics strategy design [11], farm productivity [25], mergers and acquisitions [28], defence budget planning [29], earthquake design [6], system dynamics [32], conveyor equipment selection [15], customer service management [9] and knowledge inertia [20]. For example, in [13] used the development and implementation of an educational tool based on knowledge based technology employing an expert system shell− a knowledge base system for postgraduate engineering courses. In [17] extract project WBS from the obtained mind map of brainstorming project team by artificial intelligence (AI) tools which is Prolog programming language. In [5] used expert system technology for providing developmental feedback to individuals from different ethnic minority groups. Melek and Sadeghian, [22] developed a theoretic framework for intelligent expert systems in medical encounter evaluation. Shen et al., [31] constructed an intelligent assessment system model and compared with the current assessment in education, this new intelligent assessment system expands the range of objects mailto:%201dr.arashid@gu.edu.pk,%20drarashid.khan09@gmail.com mailto:%201dr.arashid@gu.edu.pk,%20drarashid.khan09@gmail.com (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 2, No.2, February 2011 52 | P a g e http://ijacsa.thesai.org/ for evaluation and takes some AI technologies to give more heuristic and intelligent assessments. According to Zadeh [33, 34] variables words or sentences as their values is called linguistics variables and the variables that represents the gradual transition from high to low, true to false is called fuzzy variables and a set containing these variables is the fuzzy set. Fuzzy logic can be incorporated into expert system to enhance the performance and reliability of expert system in decision making. Fuzzy logic principals with expert system form a fuzzy expert system which is able to implement human knowledge & expertise with imprecise, ambiguous and uncertain data. Recently, many researchers worked on the applications of fuzzy logic in education & assessments. Chiang and Lin [10] presented a method for applying the fuzzy set theory to teaching assessment. Bai and Chen [4] presented a new method for evaluating students‘ learning achievement using fuzzy membership functions and fuzzy rules. Chang and Sun [8] presented a method for fuzzy assessment of learning performance of junior high school students. Chen and Lee [7] presented two methods for students‘ evaluation using fuzzy sets. Ma and Zhou [21] presented a fuzzy set approach to the assessment of student- cantered learning. In [30] presented a method for applying the fuzzy set theory and the item response theory to evaluate the learning performance of students. In ―unpublished‖ [2] proposed an intelligent framework for teachers‘ performance evaluations in higher education. The literature reveals that there is a vast potential of expert system and fuzzy logic in education as general and performance assessment, as a special. III. TEACHERS‘ PERFORMANCE EVALUATION The world is divided into developed and developing countries; the division is their capacity of educational and scientific attainment and its applications for economic progress and prosperity [16]. In developing countries, higher education is seen as an essential means for creation and development of resources and for improving the life of people to whom it has to serve. The problem with developing countries including Pakistan is that they have given a relatively low priority to higher education [23]. There are many reasons behind the poor status of Pakistan higher education. One of the top issues regarding the quality of higher education is the faculty (teachers) [23]. Permanent hired teachers in higher education‘s institutions especially in colleges did not update their knowledge and courses. They did the teaching as a routine activity and follow some particular books from years. Thus, students could not get updated knowledge and so fails to compete in market ―unpublished‖ [2]. To put the existing teachers on track, it is very necessary to evaluate their performance, may be in quarterly, in semester or annually, depends upon the resources universities posses. Unfortunately, there exists no standard method for evaluating teachers‘ performance in higher education institutions or computerized solution that covers all factors affecting directly or indirectly the performance of university teachers. Although, the Higher Education Commission (HEC) has did lot of regarding quality assurance by establishing an idea of the Quality Enhancement Cells (QEC) in universities; but is rarely followed by universities due to time consuming manual process and availability of funds. IV. METHOD In this research, expert system was adopted using fuzzy logic principals for teachers‘ evaluation process. In [2] they have developed the knowledge acquisition tool for the teachers‘ assessment problem in the development of intelligent expert system. They have extracted a set of 99 attributes from literature that have influence on teachers‘ performance by any means in higher education; the extracted attributes were divided in to 15 groups. TABLE I GROUPS OF ATTRIBUTES (SOURCE: AMIN & KHAN (2009)) S. No Main Groups of Attributes Weights 1 Research Orientation 0.0753 2 Publication 0.0742 3 Teaching Learning Process 0.0729 4 Personal Abilities 0.0727 5 Responsibility & Punctuality 0.0726 6 Compensation & Rewards 0.0726 7 Professional Ethics 0.0720 8 Job Security & Environment Factors 0.0706 9 Supervision 0.0677 10 Administrative Skills 0.0674 11 Awards & Achievements 0.0605 12 Promotion Factors 0.0602 13 Organization Evaluation Policy 0.0577 14 Needs & Requirements 0.0550 15 Background Factors 0.0490 Total weight: 1.0000 They have received responses from 25 highly qualified and well experienced subject experts from 11 different universities of Pakistan about the various factors that affect teachers‘ performance and also the experts‘ ranked these factors. The initial results and priority assigned to those factors are shown in Table-I. The Research Orientation is ranked highest weight (0.0753) which indicates that research work is much more important than any other task in higher education. See Figure 1 for detail. Fig. 1 Knowledge Acquisition Process [source: Amin & Khan (2009)] Rules of thumb Questions, Problems, Data Formalized, Structure the Knowledge Domain Experts (Expert Knowledge) Knowledge Engineers Knowledge Acquisition Tool: Questionnaire − Books − Journals − Reports − Data Bases − Internet etc Concepts, Expertise, Solutions, Knowledge Reference Knowledge Knowledge Refinements Teachers’ Performance Evaluation Criteria (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 2, No.2, February 2011 53 | P a g e http://ijacsa.thesai.org/ In [33] developed an important logic concept that able researchers to measure the linguistic variables with ambiguous & uncertain knowledge into numeric values in decision making process for decision makers in real world problems. In this research, the author used the concept of fuzzy set and the membership functions to map the linguistic characteristics of teachers‘ performance that are either ranked High, Medium, or Low by the academic evaluators in higher education institutions. The degree of membership in fuzzy set is [1, 0], where ‗1‘ represents highest membership and ‗0‘represents no membership. A fuzzy variable set and their membership value is defined as shown in Table-II. TABLE II FUZZY VARIABLES FOR INPUT PARAMETERS Fuzzy variable Degree of Membership Very High 1.0 High 0.8 Medium 0.6 Low 0.4 Very Low 0.2 Null 0 The inputs data for a particular teacher‘s evaluation comes from various sources that may be research productivity & publications, academics awards & achievements, students‘ satisfactions, immediate head satisfaction, colleagues‘ opinions, and annual confidential report. Most of these inputs are in non-numeric or linguistic form. Therefore, a model is designed to process these inputs through fuzzy concept to support decision makers in teachers‘ performance assessment at higher education institutions. As shown in Figure 2; the extracted knowledge is weighted according to the assigned priorities assigned by subject experts in the knowledge acquisition process and fuzzy concepts are then applied to handle the qualitative knowledge for efficient decision making possibility. All the fuzzy expert system components interact among each other to perform their functionalities achieving results. Let‘s take one group of attributes from Table-I with all its sub-factors along with assigned weights and compute the result for a particular case. Now observe Table-III, the maximum weight of all the teachers‘ performance criteria is 1.0000 and the selected main attribute got maximum weight as 0.0729; while its sub factors along with their weights are shown. TABLE III MAIN ATTRIBUTE ALONG WITH SUB-FACTORS Teachers’ Performance Evaluation in Higher Education Max Weight 1.0000 Main Attribute Sub Factors TEACHING LEARNING PROCESS Weight 0.0729 Proficiency in teaching 0.0054 Personal Interest In Teaching 0.0075 Presentation & Communications skills 0.0063 Speaking Style & Body language 0.0050 Content knowledge 0.0059 Lecture preparation 0.0067 Language command 0.0059 Response to Student queries 0.0067 Question Tackling 0.0050 Courses taught 0.0038 Students Performance 0.0029 Work load 0.0046 Fairness in marking 0.0071 The knowledge representation of the above main attribute with its sub-factors in fuzzy rules takes the following form. IF proficiency_teaching is Very High THEN W= 0.0054 IF proficiency_teaching is High THEN W= 0.0043 IF proficiency_teaching is Medium THEN W=0.0032 IF proficiency_teaching is Low THEN W= 0.0022 IF proficiency_teaching is Very Low THEN W= 0.0011 IF personal_interest_teach is Very High THEN W= 0.0075 IF personal_interest_teaching is High THEN W= 0.0060 IF personal_interest_teaching is Medium THEN W=0.0045 IF personal_interest_teaching is Low THEN W= 0.0030 IF personal_interest_teaching is Very Low THEN W= 0.0015 IF present_Comm_skill is Very High THEN W= 0.0063 IF present_Comm_skill is High THEN W= 0.0050 IF present_Comm_skill is Medium THEN W=0.0038 IF present_Comm_skill is Low THEN W= 0.0025 IF present_Comm_skill is Very Low THEN W= 0.0013 IF style_body_language is Very High THEN W= 0.0050 IF style_body_language is High THEN W= 0.0040 IF style_body_language is Medium THEN W=0.0030 IF style_body_language is Low THEN W= 0.0020 IF style_body_language is Very Low THEN W= 0.0010 IF content_knowledge is Very High THEN W= 0.0059 IF content_knowledge is High THEN W= 0.0047 IF content_knowledge is Medium THEN W=0.0035 IF content_knowledge is Low THEN W= 0.0024 IF content_knowledge is Very Low THEN W= 0.0012 IF lecture_preparation is Very High THEN W= 0.0067 IF lecture_preparation is High THEN W= 0.0054 IF lecture_preparation is Medium THEN W=0.0040 IF lecture_preparation is Low THEN W= 0.0027 IF lecture_preparation is Very Low THEN W= 0.0013 IF language_command is Very High THEN W= 0.0059 IF language_command is High THEN W= 0.0047 IF language_command is Medium THEN W=0.0035 IF language_command is Low THEN W= 0.0024 IF language_command is Very Low THEN W= 0.0012 IF response_student_queries is Very High THEN W= 0.0067 IF response_student_queries is High THEN W= 0.0054 IF response_student_queries is Medium THEN W=0.0040 IF response_student_queries is Low THEN W= 0.0027 IF response_student_queries is Very Low THEN W= 0.0013 IF question_tack is Very High THEN W= 0.0050 IF question_tack is High THEN W= 0.0040 IF question_tack is Medium THEN W=0.0030 IF question_tack is Low THEN W= 0.0020 IF question_tack is Very Low THEN W= 0.0010 IF courses_taught is Very High THEN W= 0.0038 IF courses_taught is High THEN W= 0.0030 IF courses_taught is Medium THEN W=0.0023 IF courses_taught is Low THEN W= 0.0015 IF courses_taught is Very Low THEN W= 0.0008 IF student_perform is Very High THEN W= 0.0029 IF student_perform is High THEN W= 0.0023 IF student_perform is Medium THEN W=0.0017 IF student_perform is Low THEN W= 0.0012 IF student_perform is Very Low THEN W= 0.0006 IF workload is Very High THEN W= 0.0046 IF workload is High THEN W= 0.0037 IF workload is Medium THEN W=0.0028 (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 2, No.2, February 2011 54 | P a g e http://ijacsa.thesai.org/ IF workload is Low THEN W= 0.0018 IF workload is Very Low THEN W= 0.0009 IF fairness_marking is Very High THEN W= 0.0071 IF fairness_marking is High THEN W= 0.0057 IF fairness_marking is Medium THEN W=0.0043 IF fairness_marking is Low THEN W= 0.0028 IF fairness_marking is Very Low THEN W= 0.0014 The terms Very High, High, Medium, Low, Very Low are fuzzy variables on the basis the weight W varies. This is because the fuzzy membership function application which able the system to map qualitative variables as numeric one. Before entering input case to the fuzzy expert system, let‘s discuss the computational formula which was applied in fuzzy expert system for calculating decision making score. To calculate the decision score of any single main attribute with its all sub factors the following summation formula is defined and used.    m n PnWnCi 1 ………….………. (1) Where, Ci = i th main attribute. M = number of sub-factors in the i th attribute. Pn = Fuzzy value of n th input parameter Wn = Expert weight of the relative input parameter TABLE IV EXAMPLES OF INPUT CASES Three different examples as input cases to the Fuzzy Expert system All Sub Factors Case A Case B Case C Proficiency in teaching 0.0054 H M L Personal Interest In Teaching 0.0075 H M L Presentation & Comm. Skills 0.0063 VH H L Speaking Style & Body lang. 0.0050 M L M Content knowledge 0.0059 M L M Lecture preparation 0.0067 H M H Language command 0.0059 VH H VH Response to Student queries 0.0067 VH VH H Question Tackling 0.0050 M L M Courses taught (nature) 0.0038 M M L Students Performance 0.0029 VH M M Work load 0.0046 H H H Fairness in marking 0.0071 M M M The above input data (Table-IV) is entered to the Fuzzy Expert System; through a built-in interface of the system for computing decision score as shown in Figure-3. Numbers 5, 4, 3, 2, 1 entered as input representing 5=Very High, 4=High, 3=Medium, 2=Low, 1=Very Low. In Figure 3, the interface two buttons have also shown, i.e., Explain and Why which are available for explanation of inputs and reasoning capabilities respectively. After completion of the input data the Fuzzy Expert System used the scale in Table-V, to rank the three cases A, B, C respectively. TABLE V DECISION MAKING SCALE TO A LINGUISTIC DESCRIPTION (MAX WEIGHT 0.0729) Fuzzy Expert system output Linguistic Description X< 0.0109 Poor 0.0109 ≤ X < 0.0.0218 Satisfied 0.0.0218 ≤ X< 0.0328 Good 0.0328 ≤ X < 0.0437 Very Good 0.0437 ≤ X < 0.0546 Excellent X ≥ 0.0546 Outstanding According to the developed scale in Table-V, the Fuzzy Expert System mapped the calculated numeric results of the three cases from qualitative input data into linguistic output description, as shown in Table-VI. (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 2, No.2, February 2011 55 | P a g e http://ijacsa.thesai.org/ Fig. 2 Fuzzy Expert System Model Fig. 3 User Interface TABLE VI THREE CASES RANKING Case System Calculation Description A 0.0573 Outstanding B 0.0465 Excellent C 0.0451 Excellent Users In p u ts R e su lt s Input Sources: ACR, Students, HoD, Colleagues, teacher‘s Personal data, others Knowledge Representation Decision Making Reference Knowledge Fuzzy Logic Concept for handling linguistic mapping Representation of Extracted Knowledge in Fuzzy Rules Explaining the Reasoning capabilities Draw Conclusions Inference engine User Interface Explanation Facility A Pool of High Qualified & Experience Subject Experts E li c it Set of attributes extracted from subject experts‘ expertise & knowledge for problem solution K n o w le d g e A c q u is it io n Intelligent Information Knowledge Engineers Questionnaire as a knowledge collection Tool (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 2, No.2, February 2011 56 | P a g e http://ijacsa.thesai.org/ V. CONCLUSION & FUTURE DIRECTION Regular teachers‘ assessment is suggested to maintain quality in higher education; literature clearly depicts taht there is a vast potential of the applications of fuzzy logic & expert system in teachers‘ assessment. Expert system technology using Fuzzy Logic is very interesting for qualitative facts evaluation. A model of fuzzy expert system is proposed to evaluate teachers‘ performance on the basis of various key performance attributes that have been validated previously through subject experts. The fuzzy scale has been designed to map & control the input data values from absolute truth to absolute false. The qualitative variables are mapped into numeric results by implementing the fuzzy expert system‘s model through various input examples and provided a basis to use the system ranking for further decision making. Thus, the uncertain and qualitative knowledge of the problem domain have been handled absolutely through integration of expert system technology with fuzzy logic concept. The proposed model produced significant bases for performance assessment and adequate support in decision making, so the research on the issue can be continued. Important aspect of this issue that could focus on in the future is the fuzzy expert system‘s model that could be extended to all type of employees‘ assessment in universities as well as in others government & private organizations. REFERENCES [1] H. Amin, A. R. Khan, Acquiring Knowledge for Evaluation of Teachers‘ Performance in Higher Education – using a Questionnaire. International Journal of Computer Science and Information Security (IJCSIS) 2(2009), 180-187. [2] H. Amin, An Intelligent Frame work for teachers‘ performance evaluation at higher education institutions of Pakistan, Master Thesis, Institute of Information Technology, Kohat University of Science & Technology, Kohat, NWFP, Islamic Republic of Pakistan.(2009). (Unpublished) [3] S.Ammar, W.Duncombe, B.Jump, R.Wright, Constructing a fuzzy- knowledge-based-system: An application for assessing the financial condition of public schools. Expert Systems with Applications, 27(2004), 349–364. [4] S.M.Bai, S.M.Chen, A new method for students‘ learning achievement using fuzzy membership functions. In Proceedings of the 11th conference on artificial intelligence, Kaohsiung, Taiwan, Republic of China. (2006). [5] D.Biggs, M.Sagheb-Tehrani, Providing developmental feedback to individuals from different ethnic minority groups using expert systems. Expert Systems, 25(2008), 87-97. [6] A.Berrais, A knowledge-based expert system for earthquake resistant design of reinforced concrete buildings. Expert Systems with Applications, 28 (2005), 519–530. [7] S.M.Chen, C.H.Lee, New methods for students‘ evaluating using fuzzy sets. Fuzzy Sets and Systems, 104(1999), 209–218. [8] D.F.Chang, C.M.Sun, Fuzzy assessment of learning performance of junior high school students. In Proceedings of the 1993 first national symposium on fuzzy theory and applications, Hsinchu, Taiwan, Republic of China, 1993,pp. 1–10. [9] C. F.Cheung, W.B.Lee, W. M.Wang, K.F.Chu, S.To, A multi- perspective knowledge-based system for customer service management. Expert Systems with Applications, 24(2003), 457–470. [10] T.T.Chiang, C.M.Lin, Application of fuzzy theory to teaching assessment. In Proceedings of the 1994 second national conference on fuzzy theory and applications, Taipei, Taiwan, Republic of China, 1994, pp. 92–97. [11] H. K. H.Chow, K.L.Choy, W.B.Lee, F.T.S.Chan, Design of a knowledge-based logistics strategy system. Expert Systems with Applications, 29(2005), 272–290. [12] J.A.Clark, F.Soliman, A graphical method for assessing knowledge- based systems investments. Logistics Information Management, 12(1999), 63–77. [13] A.J.Day, A.K.Suri, A knowledge-based system for postgraduate engineering courses. Journal of Computer Assisted Learning, 15(1999), 14–27. [14] J. Durkin, Application of Expert Systems in the Sciences. OHIO J. SCI. 90 (1990), 171-179. [15] D.J.Fonseca, G.Uppal, T.J.Greene, A knowledge-based system for conveyor equipment selection. Expert Systems with Applications, 26(2004), 615–623. [16] M.Hamidullah, Comparisons of the Quality of Higher Educations in Public and Private Sector Institutions, PhD Thesis, University of Arid Agriculture Rawalpindi, PAK, 2005. [17] H.Iranmanesh, M.Madadi, An Intelligent System Framework for Generating Activity List of a Project Using WBS Mind map and Semantic Network. Proceedings of World Academy of Science, Engineering and Technology. 30 (2008), 338-345. [18] A.Kazaz, Application of an Expert System on the Fracture Mechanics of Concrete. Artificial Intelligence Review. 19(2003), 177–190. [19] R.Kumra, R.M.Stein, I.Assersohn, Assessing a knowledgebase approach to commercial loan underwriting. Expert Systems with Applications, 30(2006), 507–518. [20] S. H. Liao, Problem solving and knowledge inertia. Expert Systems with Applications, 22(2002), 21–31. [21] J.Ma, D.Zhou, Fuzzy set approach to the assessment of student- centered learning. IEEE Transactions on Education, 43(2000), 237– 241. [22] W.W.Melek, A.Sadeghian, A theoretic framework for intelligent expert systems in medical encounter evaluation. Expert Systems, 26(2009), 87-97. [23] M.Naeemullah, Designing a Model for Staff Development in Higher Education of Pakistan, PhD Thesis, University of Arid Agriculture Rawalpindi, PAK, 2005. [24] T.T.Pham, G.Chen, Some applications of fuzzy logic in rules-based expert systems, Expert System,19(2002), 208-223. [25] J.Pomar,C.Pomar, A knowledge-based decision support system to improve sow farm productivity. Expert Systems with Applications, 29(2005), 33–40. [26] W. K.Wang, A knowledge-based decision support system for measuring the performance of government real estate investment. Expert Systems with Applications, 29(2005), 901–912. [27] W. K.Wang, H.C.Huang, M.C.Lai, Design of a knowledgebase performance evaluation system: A case of high-tech state-owned enterprises in an emerging economy. Expert Systems with Applications. doi:10.1016/j.eswa.2007.01.032. [28] W.Wen, W. K.Wang, T.H.Wang, A hybrid knowledgebased decision support system for enterprise mergers and acquisitions. Expert Systems with Applications, 28(2005a), 569–582. [29] W.Wen, W.K.Wang, C.H.Wang, A knowledge-based intelligent decision support system for national defense budget planning. Expert Systems with Applications, 28(2005b), 55–66. [30] M.H.Wu, Research on applying fuzzy set theory and item response theory to evaluate learning performance. Master Thesis, Department of Information Management, Chaoyang University of Technology, Wufeng, Taichung County, Taiwan, Republic of China,2003. [31] M.R.Shen, Y.Y.Tang, Z.T.Zhang, The intelligent assessment system in Web-based distance learning education. 31st Annual Frontiers in Education Conference , 1(2001), TIF-7-TIF-11. [32] N.H.Yim, S.H.Kim, H.W.Kim, K.Y.Kwahk, Knowledge based decision making on higher level strategic concerns: System dynamics approach. Expert Systems with Applications, 27(2004), 143–158. [33] L.A. Zadeh, Fuzzy sets. Inform. and control, 8 (1965), 338-353. [34] L.A.Zadeh, The concept of a linguistic variable and its application to appropriate reasoning. Information sciences, 8(1975), 43-80. AUTHORS PROFILE Dr. Abdur Rashid Khan Dr. Abdur Rashid Khan is presently working as Associate Professor at Institute of Computing & Information Technology, Gomal University, Dera Ismail Khan, Pakistan. He has completed his PhD in Computer Science (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 2, No.2, February 2011 57 | P a g e http://ijacsa.thesai.org/ from Krygyze Republic in 2004, and have published a number of articles in national and international journals. His current interest includes Expert Systems, Software Engineering, Management Information System, and Decision Support System. Hafeez Ullah Amin Mr. Hafeez is a research student at Institute of Information Technology, Kohat University of Science & Technology, Kohat 26000, KPK, Pakistan. He has completed BS(Hons) in Inforamtion Technology and MS in Computer Science in 2006 & 2009 respectiviely from the above cited institution. His current research interests includes Artificial Intelligence, Information System, and Data Base. Zia ur Rehman Mr.Zia ur Rehman is currently working as a Lecturer in Computer Science in Fuji Foundation School & College, Kohat, Pakistan . He has completed his MS in computer science from Institute of Information Technology, Kohat University of Science & Technology, Kohat, KPK, Pakistan. His current research interest includes Software Engineering, Expert System Develpoment, and Information System. 
work_yokg7gijqvg7dosuvszkjtj5qm	----	Document downloaded from: This paper must be cited as: The final publication is available at Copyright http://dx.doi.org/10.1016/j.eswa.2012.02.029 http://hdl.handle.net/10251/36442 Elsevier Yuichi Nagata; Soler Fernández, D. (2012). A new genetic algorithm for the asymmetric traveling salesman problem. Expert Systems with Applications. 39(10):8947-8953. doi:10.1016/j.eswa.2012.02.029. A new genetic algorithm for the asymmetric traveling salesman problem Yuichi Nagataa,∗, David Solerb aTokyo Institute of Technology, Kanagawa, Japan bInstitut Universitari de Matemàtica Pura i Aplicada, Universitat Politècnica de València, València, Spain Abstract The asymmetric traveling salesman problem (ATSP) is one of the most impor- tant combinatorial optimization problems. It allows us to solve, either directly or through a transformation, many real-world problems. We present in this paper a new competitive genetic algorithm to solve this problem. This algorithm has been checked on a set of 153 benchmark instances with known optimal solution and it outperforms the results obtained with previous ATSP heuristic methods. Keywords: Asymmetric traveling salesman problem, genetic algorithm, crossover oper- ator, metaheuristics 1 Introduction The asymmetric traveling salesman problem (ATSP) is one of the most important and well-known problems in combinatorial optimization, and it can be stated as follows: ∗Correspondence: Y. Nagata, Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, 4259 Nagatsuta Midori-ku Yokohama, Kanagawa 226-8502, Japan. E-mail: nagata@fe.dis.titech.ac.jp. 1 Given a complete directed graph G = (V, A), V being the vertex set and A being the arc set, with nonnegative costs associated with its arcs, find a minimum cost circuit in G passing through each vertex exactly once. The ATSP has important applications to real-world problems, especially in sequenc- ing, in distribution, and in vehicle routing problems. As a result, dozens of papers and several surveys have been written about ATSP in the past few decades. To be brief, we only mention the classical survey by Fischetti et al. (2002), detailing different exact procedures for the ATSP; the most recent exact procedure by Corberan et al. (2005), capable of solving large-size ATSP instances; the latest competitive heuristic approach for solving the ATSP by Xing et al. (2008), whose article also serves as a good survey on metaheuristic approaches for solving the ATSP; the work by Oncan et al. (2009), in which different formulations for the ATSP are compared; and the very recent concise guide by Laporte (2010) and survey on local search algorithms for this problem by Rego et al. (2011). In addition to its direct applications to real-world problems, a review of the litera- ture shows that many single-vehicle routing problems, which actually cannot be mod- eled as an ATSP, can be solved through a polynomial transformation into an ATSP. The main advantage of this transformation is that efficient algorithms developed for the ATSP can be directly applied to these problems without any modification. Among the problems that can be solved through a transformation into an ATSP are several classes of arc and/or node routing problems, defined on undirected, directed, or mixed graphs, some of which include turn penalties, forbidden turns, time-dependent costs, or delivery time windows. In this area, we cite the works by Laporte (1997), Clossey et al. (2001), Corberán et al. (2002), Blais and Laporte (2003), Soler et al. (2008), and Albiach et al. (2008). In most of these articles, the transformations they propose consist of two steps. They first transform their problem into another combi- natorial optimization problem, namely the asymmetric generalized traveling salesman problem (AGTSP), which in turn can be transformed into an ATSP (see, e.g., Noon and Bean (1993) or Ben-Arieh et al. (2003)). The AGTSP is actually a generalization of the ATSP in which each customer has several alternative locations and only one of them has to be selected for service. It can be briefly defined as follows: 2 Given a directed graph G = (V, A) with nonnegative costs associated with its arcs, such that V is partitioned into k nonempty subsets {Si}ki=1, find a minimum cost circuit passing through exactly one vertex of each subset Si ∀i ∈ {1, . . . , k}. It is worth commenting here that the idea of solving a single-vehicle routing problem by transforming it into an ATSP through an AGTSP has been recently generalized to solve different hard multivehicle routing problems, as we can see through the papers by Soler et al. (2009), Baldacci et al. (2010), Bräysy et al. (2011), and Micó and Soler (2011). All these facts have encouraged us to present a powerful new approximation algo- rithm to solve the ATSP, with the aim of improving results given by previous ATSP heuristics to help experts solve real-world problems (such as those cited above) or to measure the efficiency of actual or future proposed heuristics that directly address these problems. Our solution approach is based on genetic algorithms (GAs). The GA is a population-based approach for heuristic search in optimization problems, and it has been applied to a wide variety of optimization problems. In particular, many GAs or memetic algorithms (MAs) have been developed for the symmetric TSP (STSP) up to the present date (see, e.g., the very recent papers by Albayrak and Allahverdi (2011) and Chen and Chien (2011)). Here, the MA is also a population-based heuristic search that combines a GA with a local search to organize a more intensive local search. However, applications of the GA or MA to the ATSP are rather limited. Here we refer the reader to the three latest GAs and MAs proposed for the ATSP, all of which have shown good performance (see Choi et al. (2003), Buriol et al. (2004), and Xing et al. (2008)). The main feature of our GA is the use of an edge assembly crossover (EAX) opera- tor, which was originally proposed by Nagata and Kobayashi (1997) for both the STSP and the ATSP. The EAX generates offspring solutions by combining (undirected) edges or arcs from two parent solutions and adding relatively few short edges or arcs. The main difference between the EAX for the ATSP and the EAX for the STSP is that the first one combines parents’ arcs (directed edges) without changing their orientation whereas the second one combines parents’ (undirected) edges without respect to their orientation. Since a GA using an EAX was originally proposed, it has been significantly 3 enhanced to improve the performance for solving the STSP; the capability of the EAX for generating good offspring solutions has been enhanced by Nagata (2004, 2006b) and the capability of the GA framework for maintaining the population diversity has been also enhanced by Nagata (2006a). Because of the growing importance of the ATSP, in this paper we investigate the performance of the GA using the EAX for the ATSP by applying similar enhancements to those developed for the STSP and introducing an additional enhancement to further improve the performance. To check the efficiency of our GA, we have tested it on 153 benchmark ATSP instances, all of which have known optimal solutions. These instances correspond to two different sets: the 27 well-known instances from TSPLIB (http://comopt.ifi.uni- heidelberg.de/software/TSPLIB95/) and 126 instances with up to 315 vertices from the work by Soler et al. (2008). The rest of the paper is organized as follows. Section 2 presents the genetic algo- rithm for the ATSP, Section 3 shows the computational results obtained for the set of 153 instances, and Section 4 gives some conclusions about this work. 2 The genetic algorithm In this section, we describe the GA framework and the EAX developed for the ATSP. In Section 2.1 we first present the EAX, which includes adapted enhancements developed for the STSP. The GA framework in which the EAX is integrated is then presented in Section 2.2. We present in Section 2.3 the local search procedure used for generating an initial population for the proposed GA. 2.1 EAX for the ATSP The EAX for the ATSP consists of five steps as outlined below and illustrated in Figure 1. If more than two offspring solutions are generated, Steps 3–5 are repeated. [Basic algorithm] Step 1 Given parent solutions denoted as pA and pB, let GAB be the directed graph defined as GAB = (V, EA ∪ EB\EA ∩ EB), where EA and EB are defined as sets 4 of arcs consisting of pA and pB, respectively. Step 2 Partition all edges of GAB into AB-cycles. An AB-cycle is defined as a cycle in GAB such that arcs from EA and EB are alternately linked in the opposite orientation. Note that partition of the arcs into AB-cycles is always possible and uniquely determined (which is easy to prove). Step 3 Construct an E-set by selecting AB-cycles according to a given rule, where an E-set is defined as the union of AB-cycles. Step 4 Generate an intermediate solution from pA by removing the arcs of EA and adding the arcs of EB in the E-set. As a result, an intermediate solution consists of one or more subtours as illustrated in Figure 1. Step 5 Connect subtours into a tour to generate an offspring solution. More precisely, subtours are connected one by one, each time connecting a subtour consisting of the least number of arcs to one of the other subtours so that the sum of the cost of the arcs is minimized. Here, two subtours are connected by deleting one arc from each of the subtours and adding two arcs to connect them in such a way that all arcs in the resulting subtour have the same orientation. In Step 3, an E-set of any combination of AB-cycles can be constructed, and the EAX can generate various intermediate solutions. One simple strategy, which was proposed in the original EAX (Nagata and Kobayashi, 1997), is to select a number of AB-cycles randomly, each with a probability of 0.5, to construct an E-set. An EAX with this selection strategy is called EAX-Rand and it typically forms an intermediate solution that contains arcs of EA and arcs of EB equally. An alternative strategy for selecting AB-cycles was proposed by Nagata (2004a) to enhance the EAX for the STSP. This strategy randomly selects a single AB-cycle without overlapping the previous selection. Therefore, an EAX with this strategy can generate at most the same number of offspring solutions as the number of AB-cycles generated. An EAX with this selection strategy is called EAX-1AB and typically forms an intermediate solution similar to pA because it is generated from pA by replacing a relatively small number of arcs with the same number of arcs of pB. The advantage of EAX-1AB is its ability to better 5 maintain the population diversity when used in an appropriate GA framework, which is presented in Algorithm 1 (see Section 2.2), resulting in a significant improvement in the solution quality. Another selection strategy (called EAX-Block) for selecting AB-cycles was also proposed by Nagata (2006b), but we do not use this variant of the EAX in this paper, because we could not find a significant improvement by using this strategy in our preliminary experiments. 2.2 Main framework The main GA framework is shown in Algorithm 1. The search is initiated by generating Npop initial solutions (line 1). Here, we use a local search procedure using a variant of the 3-opt neighborhood to generate each of the initial population members. The detailed description of the local search procedure is presented in Section 2.3. For each generation of the GA (lines 3–15), each population member is selected once, as both parent pA and parent pB in random order (lines 3 and 5). For each pair of parents, EAX-1AB generates offspring solutions, where nch refers to the number of offspring solutions generated (line 6). Let nmaxch be a parameter that specifies the maximum number of offspring solutions for each pair of parents. Therefore, nch will be equal to the number of AB-cycles if the number of AB-cycles is less than nmaxch . Then, the best solution among the generated offspring solutions, denoted as cbest, is selected according to a given evaluation function (line 7). If the tour cost of cbest is better than that of pA (line 8), it will replace one of the population members selected as pA or pB (line 10 and 12). If cbest is more similar to pA than to pB, the population member selected as pA is replaced (lines 9 and 10), where dist(x, y) returns the number of different arcs between the two tours. In contrast, if cbest is more similar to pB than to pA and the tour cost of cbest is better than those of pB and pA, the population member selected as pB is replaced (lines 11 and 12). Iterations of the generation are repeated until a termination condition is met (line 16). Finally, the best solution in the population is returned (line 17). In the latest version of the GA using the EAX for the STSP by Nagata (2006b), the best offspring solution (cbest) replaces only parent pA rather than both parents (i.e., lines 6 Algorithm 1 : Procedure GA() 1: {x1, . . . , xNpop} := Generate Initial Population(); 2: repeat 3: Let r(·) be a random permutation of 1, . . . , Npop; 4: for i := 1 to Npop do 5: pA := xr(i), pB := xr(i+1); (r(Npop + 1) = r(1)) 6: {c1, . . . , cnch} := Crossover(pA, pB); 7: cbest := Select Best(c1, . . . , cnch ); 8: if cost(cbest) < cost(pA) then 9: if dist(cbest, pA) < dist(cbest, pB) then 10: xr(i) := cbest; 11: else if dist(cbest, pA) > dist(cbest, pB) and cost(cbest) < cost(pB) then 12: xr(i+1) := cbest; 13: end if 14: end if 15: end for 16: until a termination condition is satisfied 17: return the best individual in the population; 10, 11, and 13 are ignored). This selection strategy was introduced to better maintain the population diversity because EAX-1AB frequently generates offspring solutions that are more similar to pA than to pB. We call this selection strategy “SEL1” in this paper. EAX-1AB adapted to the ATSP has the same property, although not to the same extent as EAX-1AB for the STSP. If cbest is more similar to pB than to pA and cbest replaces pA, the population diversity will be fairly reduced because two similar solutions remain in the population. Therefore, we add the additional selection mechanism that replaces pB rather than pA if cbest is more similar to pB than to pA. We call this selection strategy “SEL2” in this paper. Although the most straightforward evaluation function for evaluating offspring so- lutions would be the tour cost, an alternative evaluation function was used in the latest version of the GA for the STSP to maintain the population diversity in a positive man- ner. The use of this evaluation function, however, did not improve the performance significantly in our preliminary experiments on ATSP instances. Therefore, we evaluate offspring solutions simply by the tour cost. 7 2.3 Local search The local search procedure used for generating the initial population is a simple hill- climbing method using a variant of the 3-opt neighborhood. For information on classical local search algorithms for the ATSP, we refer the reader to the work by Kanellakis and Papadimitriou (1980). The initial population consisting of Npop tours is generated by performing the local search procedure Npop times. The 3-opt neighborhood used in this paper is defined as the tours that can be obtained from the current tour by replacing three arcs in other possible ways so that all arcs in the resulting tour have the same orientation. More precisely, let (v1, v2), (v3, v4), and (v5, v6) denote three different arcs to be removed, and assume that v3 appears after v1 and before v5 in the current tour. The three arcs to be added must be (v1, v4), (v3, v6), and (v5, v2) to make the arcs in the resulting tour have the same orientation. To speed up the local search procedure, we use a variant of the “don’t look bits” strategy (Bentley, 1990). Let N(v) be a subset of the 3-opt neighborhood that requires v = v1, and we call it a subneighborhood. To reduce the neighborhood size of a subneighborhood, v4 is restricted to the vertices that are at most within the ten nearest from v1. When using the don’t look bits strategy, each of the vertices is assigned a flag having value true or false. The local search procedure begins by generating a random tour and setting all flags to false. At each iteration, a tour that improves the current tour is searched in subneighborhoods whose associated flags are false. If an improving tour is found, the current tour is immediately moved to the new tour, and the flags of the three starting points of the added three edges are set to false. However, a flag of a vertex is set to true whenever no improving move is found in the corresponding subneighborhood. Iterations are repeated until all flags become true. 3 Computational experiments In this section, we first perform the GA with several different configurations on the well- known 27 instances included in TSPLIB to detect the best one. The results obtained 8 with the best configuration are then compared for this set of 27 instances with those of the latest genetic algorithms proposed for the ATSP. Finally, we also run our GA on a set of 126 instances created in the work by Soler et al. (2008). The GA was implemented in C++ and was executed in a virtual machine environ- ment (i.e., each job is executed on a single core, but multiple jobs may be executed in the same node) on a cluster with Intel Xeon 2.93-GHz nodes. 3.1 Analysis of algorithm components We performed the proposed GA with four different configurations on the 27 instances in TSPLIB to detect the best configuration. Two parameters of Algorithm 1 are set as follows: Npop = 100 and n max ch = 30. Each run is terminated when the best solution is not improved over the last 20 generations. Four different configurations of the GA are summarized below. SEL1+EAX-Rand Selection strategy SEL1 is used in Algorithm 1. EAX-Rand is used as the crossover operator. For each pair of parents, the number of offspring solutions is set to the number of AB-cycles if the number of AB-cycles generated is less than nmaxch . We use this restriction to compare EAX-Rand with EAX-1AB under the same condition. SEL2+EAX-Rand Selection strategy SEL2 is used in Algorithm 1. EAX-Rand is used as the crossover operator. The number of offspring solutions is restricted as in the case of SEL1+EAX-Rand. SEL1+EAX-1AB Selection strategy SEL1 is used in Algorithm 1. EAX-1AB is used as the crossover operator. SEL2+EAX-1AB Selection strategy SEL2 is used in Algorithm 1. EAX-1AB is used as the crossover operator. We performed the GA 100 times on each instance with each configuration. Table 1 shows results. The first column lists the instance names, where the number of vertices is indicated by the name except for kro124p (the number of vertices being 100 in 9 this instance). For each of the GAs with four different configurations, we list the number of runs that succeeded in finding the optimal solution over 100 runs (Suc.), the average percentage excess with respect to the optimal solutions (a-err), and the average computation time per single run in seconds (a-T). Average results over all instances are also listed at the bottom of the table. First, we compare the results of the GAs with two configurations, SEL1+EAX- Rand and SEL1+EAX-1AB. The table shows that the solution quality of the GA with EAX-1AB is better than that of the GA with EAX-Rand when selection strategy SEL1 is used. A major factor in this difference is that the population diversity is better maintained by replacing a population member (selected as pA) with a relatively similar offspring solution and EAX-1AB usually generates offspring solutions similar to pA. Next, we focus on the difference between selection strategies SEL1 and SEL2. The table shows that the use of selection strategy SEL2 improves the solution quality regardless of the crossover type. The main reason for this improvement is also that the population diversity is better maintained by replacing a population member with a relatively similar offspring solution and selection strategy SEL2 enhances this property. Here, one should note that SEL2+EAX-Rand is comparable to SEL2+EAX-1AB with respect to the solution quality, but SEL2+EAX-Rand requires more computation time than SEL2+EAX-1AB. This is because EAX-1AB can generate offspring solutions more efficiently than EAX-Rand, which is another advantage of EAX-1AB. By comparing the results listed in the table, we can see that SEL2-EAX-1AB seems to be a good configuration for the GA presented in Algorithm 1. 3.2 Comparison with other algorithms We compare the results of our GA with configuration SEL2+EAX-1AB with those of the latest GAs or MAs proposed in the literature to solve the ATSP: the heuristics proposed by Choi et al. (2003), Buriol et al. (2004), and Xing et al. (2008). We make comparisons for the 27 well-known instances in TSPLIB. Table 2 lists the results; the algorithms are denoted as the reference on the first line of the table; this is followed by the type of method, the computer specifications, and the number of runs. For each 10 Table 1: Results of the GAs with four different configurations. SEL1+EAX-Rand SEL2+EAX-Rand SEL1+EAX-1AB SEL2+EAX-1AB Instance Suc. a-err a-T Suc. a-err a-T Suc. a-err a-T Suc. a-err a-T br17 100 0.0000 0.00 100 0.0000 0.00 100 0.0000 0.00 100 0.0000 0.00 p43 92 0.0014 0.01 92 0.0014 0.04 91 0.0016 0.03 94 0.0011 0.02 ry48p 100 0.0000 0.01 100 0.0000 0.04 100 0.0000 0.02 100 0.0000 0.02 ft53 75 0.0275 0.02 93 0.0101 0.05 64 0.0408 0.02 95 0.0064 0.03 ftv33 100 0.0000 0.01 100 0.0000 0.02 100 0.0000 0.01 100 0.0000 0.01 ftv35 97 0.0041 0.01 100 0.0000 0.03 99 0.0014 0.01 100 0.0000 0.02 ftv38 92 0.0105 0.01 100 0.0000 0.03 99 0.0013 0.01 100 0.0000 0.02 ftv44 49 0.3162 0.02 62 0.2356 0.03 56 0.2728 0.02 66 0.2108 0.02 ftv47 100 0.0000 0.02 100 0.0000 0.03 100 0.0000 0.02 100 0.0000 0.02 ftv55 100 0.0000 0.02 100 0.0000 0.05 100 0.0000 0.03 100 0.0000 0.03 ftv64 100 0.0000 0.03 100 0.0000 0.05 100 0.0000 0.03 100 0.0000 0.04 ft70 67 0.0017 0.04 88 0.0006 0.08 69 0.0016 0.04 85 0.0008 0.04 ftv70 100 0.0000 0.03 100 0.0000 0.07 100 0.0000 0.04 100 0.0000 0.04 ftv90 75 0.0399 0.05 91 0.0114 0.10 85 0.0222 0.05 96 0.0057 0.05 ftv100 79 0.0296 0.07 90 0.0112 0.11 87 0.0162 0.08 97 0.0034 0.06 kro124p 93 0.0039 0.05 96 0.0012 0.16 96 0.0012 0.07 97 0.0009 0.08 ftv110 86 0.0179 0.09 87 0.0148 0.15 94 0.0072 0.10 93 0.0072 0.08 ftv120 79 0.0263 0.11 93 0.0069 0.22 92 0.0088 0.12 87 0.0157 0.10 ftv130 79 0.0243 0.08 87 0.0130 0.20 88 0.0130 0.13 89 0.0104 0.11 ftv140 75 0.0285 0.11 86 0.0124 0.18 84 0.0202 0.15 89 0.0116 0.12 ftv150 82 0.0146 0.11 88 0.0100 0.32 93 0.0065 0.18 91 0.0073 0.15 ftv160 92 0.0101 0.15 97 0.0041 0.26 96 0.0056 0.18 99 0.0011 0.17 ftv170 96 0.0044 0.16 96 0.0044 0.43 98 0.0022 0.24 98 0.0022 0.19 rbg323 100 0.0000 1.16 100 0.0000 3.59 100 0.0000 1.44 100 0.0000 0.96 rbg358 100 0.0000 1.65 100 0.0000 4.22 100 0.0000 1.82 100 0.0000 1.32 rbg403 100 0.0000 1.82 100 0.0000 4.68 100 0.0000 1.86 100 0.0000 1.61 rbg443 100 0.0000 1.76 100 0.0000 4.12 100 0.0000 1.77 100 0.0000 1.70 Average 89.19 0.021 0.28 94.30 0.012 0.71 92.26 0.016 0.31 95.41 0.011 0.26 algorithm, we list the number of runs that succeeded in finding the optimal solution (Suc.) if presented in the literature, the average percentage excess with respect to the optimal solutions (a-err), and the average computation time per single run in seconds (a-T). Results for the blank cells are not presented in the literature. Table 2 shows that our GA is superior to all three compared algorithms in terms of average solution quality; only the procedure by Xing et al. seems to find similar solutions to ours (but they present results for only 16 out of the 27 instances), but their average results are slightly worse than ours. For the computation time, it is very difficult to make a fair comparison because the algorithms in the table were executed using different computers (and languages). Nonetheless, our GA would not be consid- erably slower than the compared heuristics even if the difference in computer speed is 11 Table 2: Comparisons with other algorithms. Author Choi at al. Buriol et al. Xing et al. Our method Our method (2003) (2004) (2008) Npop = 100 Npop = 300 Method GA MA MA GA GA Computer Pentium Pentium Pentium Xeon Xeon 550 MHz 1.7 GHz 1.6 GHz 2.93 GHz 2.93 GHz # of Runs 3 20 100 100 100 Instance a-err a-T Suc. a-err a-T a-err a- T Suc. a-err a-T Suc. a-err a-T br17 0.00 0 20 0.00 0.05 0.000 1.9 100 0.000 0.00 100 0.0000 0.00 ftv33 0.00 4 20 0.00 0.05 100 0.000 0.01 100 0.0000 0.06 ftv35 0.14 1 20 0.00 0.08 0.000 15.6 100 0.000 0.02 100 0.0000 0.05 ftv38 0.13 2 5 0.10 0.26 100 0.000 0.02 100 0.0000 0.08 p43 0.00 1 11 0.01 0.35 0.000 4.2 94 0.001 0.02 100 0.0000 0.10 ftv44 1.30 5 7 0.44 0.36 66 0.211 0.02 94 0.0372 0.07 ftv47 0.00 3 20 0.00 0.13 0.000 90.3 100 0.000 0.02 100 0.0000 0.07 ry48p 0.33 5 17 0.03 0.32 0.001 53.4 100 0.000 0.02 100 0.0000 0.07 ft53 0.00 8 20 0.00 0.2 0.000 41.2 95 0.006 0.03 100 0.0000 0.09 ftv55 0.00 5 20 0.00 0.16 0.000 125.6 100 0.000 0.03 100 0.0000 0.10 ftv64 0.00 7 20 0.00 0.24 0.000 101.2 100 0.000 0.04 100 0.0000 0.17 ft70 0.21 20 8 0.03 0.86 0.000 83.6 85 0.001 0.04 100 0.0000 0.14 ftv70 0.00 12 19 0.01 0.38 0.000 43.8 100 0.000 0.04 100 0.0000 0.19 ftv90 0.00 20 20 0.00 0.28 96 0.006 0.05 100 0.0000 0.15 ftv100 0.00 62 20 0.00 0.40 97 0.003 0.06 100 0.0000 0.29 kro124p 0.52 67 18 0.01 0.74 0.001 28.9 97 0.001 0.08 100 0.0000 0.33 ftv110 0.54 61 18 0.02 0.98 93 0.007 0.08 100 0.0000 0.36 ftv120 0.29 89 7 0.14 2.00 87 0.016 0.10 100 0.0000 0.37 ftv130 0.59 97 18 0.01 1.30 89 0.010 0.11 100 0.0000 0.42 ftv140 0.32 81 14 0.08 2.11 89 0.012 0.12 100 0.0000 0.53 ftv150 0.55 88 18 0.01 1.54 91 0.007 0.15 100 0.0000 0.56 ftv160 0.12 101 16 0.02 2.04 99 0.001 0.17 100 0.0000 0.65 ftv170 0.22 94 15 0.05 2.78 0.020 68.3 98 0.002 0.19 100 0.0000 0.69 rbg323 0.00 2 20 0.00 0.07 0.000 110.4 100 0.000 0.96 100 0.0000 4.25 rbg358 0.00 3 20 0.00 0.08 0.000 58.4 100 0.000 1.32 100 0.0000 5.63 rbg403 0.00 2 20 0.00 0.08 0.000 33.1 100 0.000 1.61 100 0.0000 6.63 rbg443 0.00 4 20 0.00 0.09 0.000 144.2 100 0.000 1.70 100 0.0000 6.74 considered, and it seems that the MA by Xing et al. would require longer computation time. We additionally performed our GA with a population size of 300 to improve the solution quality. Our GA with a population size of 300 found optimal solutions in all 100 runs for 26 out of the 27 TSPLIB instances (94 on ftv44). The computation time was increased by about a factor of 3, so our GA with a population size of 300 would still be faster than the MA by Xing et al., but in this case with a much greater difference in our favor for the average deviations. 12 3.3 Other results As mentioned in Section 1, many single-vehicle routing problems that actually cannot be modeled as an ATSP can be solved through a polynomial transformation into an ATSP. This is the case of the problem presented by Soler et al. (2008). They generated 128 instances for that problem and they transformed them into ATSP instances. To solve the transformed instances, they run the ATSP exact procedure by Corberan et al. (2005) on a PC with 1.8-GHz Pentium IV processor. Most of the instances were optimally solved in less than 2 h of running time, some instances needed more than 10 h to obtain the optimal solution (with one of them taking more than 17 h), and in only two instances was the execution aborted owing to the long period of time spent without finding the optimal cost. Thus, we have a set of 126 ATSP instances with known optimal solution ob- tained from the aforementioned work, and with a number of vertices between 64 and 315. We have uploaded all data corresponding to this set of instances to the website http://www.iumpa.upv.es/arxivDSoler/, to make them easily available to any researcher. This set has been recently used by Bräysy et al. (2011) as a benchmark set to test the efficiency of a memetic algorithm for a capacitated vehicle routing problem. Although this algorithm was not specifically designed to solve problems with a single vehicle, it was able to find the optimal solution in 86 out of the 126 instances. In Table 3, we show the results on this set of ATSP instances obtained with our GA with configuration SEL2+EAX-1AB where the population size was set to 300. In this table, the 126 instances are partitioned into 24 groups, separated by horizontal lines, according to the features of the original instances from which they come. For each instance the table lists the name of the original instance (Name); the number of vertices (|V |); the optimal cost (Opt-cost); the time in seconds to obtain the optimal solution with the exact procedure (Opt-T); the best solution found in 100 runs of the GA (Best), with Opt shown if this solution coincides with the optimal one; the average cost over the 100 runs (a-cost), with Opt meaning that the GA has reached the optimal solution in the 100 runs; and finally the average computation time per single run in 13 seconds (a-T). From Table 3 we can see that our GA was able to find the optimal solution in 121 out of the 126 instances. The worst case was instance D143042, where the percentage excess of the best solution obtained with our GA over the optimal cost is only 0.0021%, and the second worst case was D183841b, where the percentage excess of the best solution obtained with our GA over the optimal cost is only 0.0004%. Moreover, with respect to the instances where the GA did not find the optimal solution in all 100 runs, the worst average deviation of the 100 runs with respect to the optimal cost also corresponds to instance D143042 (0.0024%). Table 3: Computational results for instances from Soler et al. (2008). Name |V | Opt-cost Opt-T Best a-cost a-T D41140 64 417880 6.31 Opt Opt 0.2 D61440 77 512674 14.22 Opt Opt 0.3 D4940m 51 336325 2.69 Opt Opt 0.1 D81740m 102 664613 5.65 Opt Opt 0.4 D4940 67 422508 6.26 Opt Opt 0.2 D81740 129 809706 51.74 Opt Opt 0.6 D82040m 103 689971 12.19 Opt Opt 0.5 D102240m 121 803595 34.54 Opt Opt 0.6 D82040 122 790971 8.57 Opt Opt 0.7 D102240 163 1027751 204.87 Opt Opt 1.0 D102640m 118 815096 151.26 Opt Opt 1.3 D122640m 130 880335 21.7 Opt Opt 1.3 D122940m 148 997772 765.55 Opt Opt 1.1 D102640 139 926142 330.59 Opt Opt 1.8 D122640 150 990448 69.15 Opt Opt 1.1 D122940 209 1316922 1018.59 Opt 1316922.2 2.6 D143040a 161 1072673 105.4 Opt Opt 1.5 D143340 230 1452339 767.64 Opt Opt 4.7 D143340m 162 1100177 585.72 Opt Opt 1.4 D143040 236 1472328 97.22 Opt 1472332.6 3.3 D143040b 186 1210175 185.7 Opt Opt 3.2 D163440a 169 1145914 181.96 Opt Opt 1.4 D163440 226 1452914 1519.4 Opt Opt 0.9 D163440b 200 1310767 197.4 Opt 1310767.3 2.4 D183840a 184 1255781 305.49 Opt Opt 2.9 D163740 226 1468709 2585.73 Opt Opt 4.3 D183840 258 1652091 1171.17 Opt Opt 5.9 D183840b 220 1449912 565.84 Opt Opt 4.3 D204240a 213 1438448 108.65 Opt 1438453.5 2.5 D184040a 196 1331558 3467.28 Opt Opt 2.4 D204240 315 1980807 408.64 Opt 1980814.5 8.3 D184040 262 1686092 3119.43 Opt Opt 6.7 D184040b 226 1493759 1037.65 Opt 1493776.2 4.5 D204240b 275 1764636 1126.68 Opt Opt 5.1 D41141 68 454381 28.73 Opt Opt 0.3 14 Table 3: Computational results for instances from Soler et al. (2008). Name |V | Opt-cost Opt-T Best a-cost a-T D4941 70 452101 12.97 Opt 452107.4 0.3 D61441 84 561657 17.52 Opt Opt 0.4 D61641 95 633760 72.12 Opt Opt 0.5 D81741 112 730893 67.99 Opt 730893.9 0.8 D82041m 122 804695 999.92 Opt Opt 0.7 D102241m 128 854290 151.98 Opt Opt 0.8 D82041 136 880751 104.35 Opt Opt 0.9 D102241 161 1029382 658.94 Opt Opt 1.9 D122641a 152 1011650 444.4 Opt Opt 1.4 D102641m 152 1009032 633.89 Opt Opt 1.2 D122941m 157 1058436 212.95 Opt Opt 1.3 D122641 207 1306865 378.99 Opt Opt 3.1 D102641 188 1199118 1006.29 Opt 1199121.2 2.3 D122941 192 1243493 388.99 Opt Opt 1.9 D122641b 179 1156769 367.12 Opt Opt 2.0 D143041a 164 1103917 166.81 Opt 1103917.2 2.5 D143341a 178 1201011 40.43 Opt 1201016.0 2.3 D143341b 211 1376011 3313.77 Opt 1376020.6 2.6 D143041 217 1389294 81.79 1389296 1389296.0 4.3 D143341 251 1584011 2085.68 Opt 1584022.4 5.0 D143041b 185 1219063 213.49 Opt Opt 3.1 D163441a 198 1313082 853.43 Opt Opt 2.8 D163441 282 1761528 886.66 Opt 1761554.6 8.7 D163441b 242 1547424 1253.51 Opt Opt 2.8 D163741m 200 1343095 2447.26 Opt Opt 2.6 D163741 235 1530323 928.57 Opt 1530329.7 7.1 D183841a 206 1385166 1724.55 Opt Opt 3.8 D183841 278 1769642 970.48 1769648 1769648.0 7.5 D183841b 237 1552379 977.39 1552385 1552385.0 5.1 D184041b 242 1591521 1054.9 Opt Opt 5.3 D184041a 224 1491653 1180.46 Opt Opt 5.0 D184041 282 1803657 315.32 Opt Opt 7.4 D204241a 224 1510820 503.23 Opt 1510820.9 4.1 D204441a 230 1556414 460.77 Opt Opt 4.1 D204241 291 1872362 5613.99 Opt 1872363.1 6.4 D204441 315 2007744 5536.11 Opt 2007778.5 9.2 D204241b 258 1693166 6030.21 Opt 1693169.9 5.4 D204441b 273 1785585 1125.2 Opt 1785588.3 7.3 D4942 80 517768 12.91 Opt Opt 0.4 D61642 138 875072 98.32 Opt Opt 1.1 D61442 120 768117 145.17 Opt Opt 0.7 D81742 127 828368 252.99 Opt Opt 1.0 D82042 128 849394 190.2 Opt Opt 0.9 D102242 146 963636 1941.23 Opt 963636.6 1.5 D122642m 162 1077557 945.37 Opt 1077562.4 1.6 D122942m 172 1149670 1966.17 Opt 1149674.2 2.2 D102642 160 1065324 275.34 Opt 1065326.2 1.9 D122642 176 1153625 3036.56 Opt 1153629.6 2.2 D122942 188 1235878 623.02 Opt 1235882.4 4.2 D143042m 184 1224644 942.36 Opt 1224664.1 3.6 D143342m 200 1329993 184.33 Opt 1329995.5 3.4 15 Table 3: Computational results for instances from Soler et al. (2008). Name |V | Opt-cost Opt-T Best a-cost a-T D143042 211 1369730 298.3 1369759 1369763.5 4.8 D143342 237 1525120 641.75 1525122 1525122.0 4.4 D163442 253 1622974 593.09 Opt 1622984.5 5.3 D163442a 206 1367607 512.45 Opt Opt 3.1 D163742a 214 1432892 8394.27 Opt 1432921.6 6.6 D163442b 219 1442855 4401.89 Opt 1442857.0 3.6 D163742b 246 1602916 3442.84 Opt 1602952.5 4.7 D163742 301 1888149 5267.52 Opt 1888173.4 8.8 D183842m 213 1434387 586.55 Opt 1434407.1 4.9 D183842 233 1542727 251.83 Opt 1542740.9 7.2 D184042m 229 1534827 6919.67 Opt Opt 3.8 D184042 270 1753907 6398.55 Opt Opt 5.8 D204242a 236 1587087 1284.93 Opt 1587089.3 4.6 D204242 289 1872624 5832.32 Opt 1872626.1 10.8 D204242b 265 1742305 3821.49 Opt 1742312.0 6.9 D41143 108 695196 402.88 Opt Opt 0.7 D61443 128 822217 240.52 Opt 822217.9 1.1 D61643 134 869353 2003.02 Opt Opt 1.5 D81743 147 948708 741.11 Opt 948719.1 2.1 D82043m 154 1003444 1497.6 Opt 1003444.2 1.6 D82043 168 1079720 133.91 Opt Opt 2.1 D102643 168 1121130 2091.13 Opt Opt 2.2 D122643 186 1218825 131.72 Opt Opt 2.2 D122943m 210 1365993 554.48 Opt Opt 1.2 D122943 240 1526097 2901.27 Opt Opt 7.1 D4944 120 758785 5141.3 Opt 758791.0 1.0 D61444 152 964730 2109.14 Opt 964740.4 1.8 D61644 153 984935 1924.15 Opt 984937.0 1.9 D81744 154 995860 13342.23 Opt 995865.8 2.2 D102244m 194 1245433 1997.75 Opt 1245436.6 3.7 D82044 168 1091866 1016.54 Opt Opt 2.5 D122644m 198 1295659 63854.52 Opt 1295663.4 3.8 D122944m 210 1380447 6368.56 Opt 1380447.9 6.4 D122944 224 1456564 14789.84 Opt 1456573.8 8.4 D102644 214 1379555 35721.07 Opt 1379562.2 3.8 D143044m 219 1438672 225.13 Opt 1438704.1 9.9 D143044 245 1578725 4017.8 Opt 1578759.7 9.4 D143344 236 1549308 39896.66 Opt Opt 6.4 D163744m 254 1671485 37142.93 Opt Opt 13.8 D163744 278 1801619 50808.44 Opt 1801621.8 16.7 4 Conclusions We have presented a new GA to solve the ATSP. We have checked its efficiency on a set of 153 benchmark ATSP instances with known optimal solution. From the obtained re- sults with the GA, we believe that our procedure is very competitive. In fact, it outper- 16 forms the results obtained with previous ATSP heuristics in the well-known set of ATSP instances from TSPLIB. Because the ATSP can be used to solve many real-world prob- lems, either directly or through a transformation, our aim is to show that the procedure presented here can be useful for solving these problems or for measuring the efficiency of actual or future proposed heuristics that directly address these problems. Moreover, to strengthen this aim, through the website http://www.iumpa.upv.es/arxivDSoler/ we make easily available to any researcher, a detailed information on the set of 126 ATSP instances used here and in previous papers and not coming from TSPLIB. Acknowledgements This work has been partially supported by the Ministerio de Educación y Ciencia of Spain (Project No. TIN2008-06441-C02-01). References Albayrak, M., & Allahverdi, N. (2011). Development a new mutation operator to solve the traveling salesman problem by aid of genetic algorithms. Expert Systems with Applications, 38, 1313–1320. Albiach, J., Sanchis, J.M., & Soler, D. (2008). An asymmetric TSP with time windows and with time-dependent travel times and costs: An exact solution through a graph transformation. European Journal of Operational Research, 189, 789–802. Baldacci, R., Bartolini, E., & Laporte, G. (2010). Some applications of the generalized vehicle routing problem. Journal of the Operational Research Society, 61, 1072– 1077. Ben-Arieh, D., Gutin, G., Penn, M., Yeo, A., & Zverovitch, A. (2003). Transforma- tions of generalized ATSP into ATSP. Operations Research Letters, 31, 357–365. Bentley, J.L. (1990). Experiments on traveling salesman heuristics, Proceedings of the First Annual ACM-SIAM Symposium on Discrete Algorithms, pp. 91–99. Blais, M., & Laporte, G. (2003). Exact solution of the generalized routing problem through graph transformations. Journal of the Operational Research Society, 54, 906–910. Bräysy, O., Mart́ınez, E., Nagata, Y., & Soler, D. (2011). The mixed capacitated general routing problem with turn penalties. Expert Systems with Applications, 30, 12954–12966. 17 Buriol, L., França, P.M., & Moscato, P. (2004). A new memetic algorithm for the asymmetric traveling salesman problem. Journal of Heuristics, 10, 483–506. Chen, S.M., & Chien, C.Y. (2011). Parallelized genetic ant colony systems for solving the traveling salesman problem. Expert Systems with Application, 38, 3873–3883. Choi, I.C., Kim, S.I., & Kim, H.S. (2003). A genetic algorithm with a mixed region search for the asymmetric traveling salesman problem. Computers & Operations Research, 30, 773–786. Clossey, J., Laporte, G., & Soriano, P. (2001). Solving arc routing problems with turn penalties. Journal of the Operational Research Society, 52, 433–439. Corberán, A., Mart́ı, R., Mart́ınez, E., & Soler, D. (2002). The rural postman problem on mixed graphs with turn penalties. Computers & Operations Research, 29, 887–903. Corberán, A., Mej́ıa, G., & Sanchis, J.M. (2005). New results on the mixed general routing problem. Operations Research, 53, 363–376. Fischetti, M., Lodi, A., & Toth, P. (2002). Exact methods for the asymmetric traveling salesman problem. In: G. Gutin & A.P. Punnen (Eds.), The Traveling Salesman Problem and Its Variations, pp. 168–205 (Boston: Kluwer). Kanellakis, P.C., & Papadimitriou, C.H. (1980). Local search for the asymmetric traveling salesman problem. Operations Research, 28, 1086–1099. Laporte, G. (1997). Modeling and solving several classes of arc routing problems as traveling salesman problems. Computers & Operations Research, 24, 1057–1061. Laporte, G. (2010). A concise guide to the traveling salesman problem. Journal of the Operational Research Society, 61, 35–40. Micó, J.C., & Soler, D. (2011). The capacitated general windy routing problem with turn penalties. Operations Research Letters, 39, 265–271. Nagata, Y., & Kobayashi, S. (1997). Edge assembly crossover: A high-power genetic algorithm for the traveling salesman problem. Proceedings of the 7th Interna- tional Conference on Genetic Algorithms, pp. 450–457. Nagata, Y. (2004). The EAX algorithm considering diversity loss. Proceedings of the 8th International Conference on Parallel Problem Solving from Nature, Lecture Notes in Computer Science, 3242, 332–341. Nagata, Y. (2006a). Fast EAX algorithm considering population diversity for traveling salesman problems. Proceedings of the 6th International Conference on Evolu- tionary Computation in Combinatorial Optimization, Lecture Notes in Computer Science, 3906, 171–182. Nagata, Y. (2006b). New EAX crossover for large TSP instances. Proceedings of the 9th International Conference on Parallel Problem Solving from Nature, Lecture Notes in Computer Science, 4193, 372–381. 18 Noon, C.E., & Bean, J.C. (1993). An efficient transformation of the generalized traveling salesman problem. INFOR, 31, 39–44. Oncan, T., Altinel, I.K., & Laporte, G. (2009). A comparative analysis of several asymmetric traveling salesman problem formulations. Computers & Operations Research, 36, 637–654. Rego, C., Gamboa, D., Glover, F., & Osterman, C. (2011). Traveling salesman prob- lem heuristics: Leadings methods, implementations and latest advances. Euro- pean Journal of Operational Research, 211, 427–441. Soler, D., Albiach, J., & Mart́ınez, E. (2009). A way to optimally solve a time- dependent vehicle routing problem with time windows. Operations Research Let- ters, 37, 37–42. Soler, D., Mart́ınez, E., & Micó, J.C. (2008). A transformation for the mixed general routing problem with turn penalties. Journal of the Operational Research Society, 59, 540–547. Xing, L.N., Chen, Y.W., Yang, K.W., How, F., Shen, X.S., & Cai, H.P. (2008). A hybrid approach combining an improved genetic algorithm and optimization strategies for the asymmetric traveling salesman problem. Engineering Applica- tions of Artificial Intelligence, 21, 1370–1380. 19 tour-A tour-B GAB E-set Intermediate valid tour Step 1 Step 2 AB-cycle Step 3 Step 4 Step 5 1 2 3 4 1 2 4 1 4 4 3 1 Figure 1: Illustration of the EAX for the ATSP. 20 
work_yp2pscg5uzfkznskjeybku2cly	----	NASA Technical Reports Server (NTRS) 
work_ype7zimnyvbpflbkklchrnzlsq	----	Expert Systems: can they aid geological interpretation and evaluation? Journal of the Geological Society, London, Vol. 149, 1992, pp. 149. Printed in Northern Ireland Conference Report Expert Systems: can they aid geological interpretation and evaluation? G . W A D G E ' & P . V I D L E R 2 lNUTIS, University of Reading, Reading K G 6 2AB 2Signal Computing, 20 Nugent Rd., Guildford, Surrey GU2 5AF This was a one-day meeting convened by P. Vidler and G. Wadge of the Geoscience Information Group at Burlington House on 20 February 1991. The aim was to provide a critical look at how valuable was the current and potential contribution of Oxpert system' technology to geologicalprob- lems. Ten papers were presented to an audience of 39. In the first of four papers from oil industry applications, D. Wolstenholme (BP Research) gave a real time presentation of BIOMARKER, an expert system to evaluate depositional en- vironment. The evaluation is performed from the knowledge of the molecular parameter interpretation technique used by geo- chemists on organic samples. The system is written in the Pro- log language on a Macintosh and performs to 96% accuracy against a test data set. Another system concerned with applying expert knowledge to the interpretation of analytical data was discussed by L. A. J. Garvie (Bristol). INTERSTRAT aids the identification of interstratified clay minerals using X-ray diffraction data. The ambiguities and uncertainties in the data, with which the expert system approach is designed to cope, arise from poorly crystalline structure, crystal chemical variability and interstratification of structural layers. The knowledge used by the system embodies diffraction parameters for all the clay minerals and their variations with pretreatments. A. C. Higgins & P. A. Swaby (BP Research) jointly described a much more ambitious system than the previous two, a visual identification expert system (VIDES) for microfossils. Developed using the expert system shell KEE on a work- station, VIDES is intended as both a training tool and a stan- dard, company-wide reference. It gives access to a large classificatory database from both in-house and literature sources, and allows the user to identify microfossils at the species level from a potential choice of several hundred thousand using comparative onscreen display of target and reference samples. The idea has great potential but major prob- lems include 2D imaging of 3D objects and resistance to its use from the 'experts' themselves. Basin evaluation has been tackled traditionally in the exploration industry by supporting the exploration geologists with database management tools. B. T. Wells (Robertsons) made a strong case for additionally incorporating expert sys- tem techniques into this support. He also argued that even the crudest attempts to manage the uncertainties in the evaluation process would be of value in such hybrid databaseknowled- gebase approaches. The Optimist system described by P. Clark (Turing Institute) and M. Whyatt (Enterprise) addressed the end of such an evaluation process: the prospect appraisal. This system employs an unusual approach in which the system and user 'argue' about current evaluation of an oil prospect. The system will argue for consistency with decisions made by experts in previous case histories. The afternoon session focussed on geotechnical and mineral exploration applications of expert systems. D. G . Toll (Dur- ham University) described a geotechnical expert system which is able to assimilate data from boreholes and other sources with a view to creating a soil profile for an area under investiga- tion. The expert system compares this profile with others in its database, allowing assessments to be made of how the soil will affect geotechnical design. Aerial mapping using high spectral resolution imaging can permit specific mineral maps to be created, despite problems of component mixing, surface shadowing and instrument noise. The AIMLIS package described by S . Mackin (DLR Institute for Optoelectronics, Germany) offered a rapid means of extracting the diagnostic features using pattern matching tech- niques. The expert system generates a probability of mismatch with similar spectra, and examines mismatches with neigh- bouring spectra. G . Wadge & P. A. V. Young (Reading University) illustrated how an expert system had aided a project established to evalu- ate prospectivity for buried Pb/Zn/F/Ba mineralization. This required the mineralization process resulting from basin de- watering to be modelled. A demonstrator system, BURMIN, used an expert system shell to guide the user through the vari- ous steps of mineralization modelling from a multidisciplin- ary database for the East Midlands. Most mineral exploration projects require large amounts of data and M. D. Mulvenna (University of Ulster) described an expert system, EXPLORER, designed to interrogate a large database associ- ated with gold prospecting in Ireland. Ultimately the system had to interpret the other data in terms of their significance to target and pathfinder geochemical elements. D. D. Hawkes (Birmingham University) described another expert system for finding gold, appropriately called Gold Finder, this time tuned to Archaean greenstone belts. The main goal was to advise on potential drill sites. Written in Prolog for a Macin- tosh, the system provided full explanations of its reasoning, and used a real example of a drill site evaluation to illustrate its operation. On the evidence of the meeting, expert systems are being successfully applied to actual geological problems, although they do not, as yet, appear to have entered the mainstream of geological thinking. Many of the geological expert systems developed to date, including most of those in this meeting, address well-bounded areas of knowledge which are difficult to master, partly because of uncertainties in the data and inter- pretation. This is the classic area of application of small expert systems that can be developed quite cheaply directly from lan- guages such as Prolog. However, it seems likely that the area where expert systems will find their true home in geological applications will be in aiding the complex multidisciplinary evaluation processes typically encountered in areas such as basin studies. More sophisticated expert system software and its interfacing with spatial databases are required to attempt such tasks. 149 by guest on April 5, 2021http://jgs.lyellcollection.org/Downloaded from http://jgs.lyellcollection.org/ 
work_ypvrrpl42jbyvp3qib477ernda	----	NASA Technical Reports Server (NTRS) 
work_yrlqw2ryznbwzp3nexixu5zhv4	----	An automatic apnea screening algorithm for children Expert Systems With Applications 48 (2016) 42–54 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa An automatic apnea screening algorithm for children Sebastián A. Ríos∗, Lili Erazo Business Intelligence Research Center, Department of Industrial Engineering, University of Chile, Av. Republica 701, Santiago, Chile a r t i c l e i n f o MSC: 00-01 99-00 Keywords: Screening algorithms Apnea classification Apnea screening Ubiquitous health system a b s t r a c t Sleep Disordered Breathing (SDB) is a group of diseases that affect the normal respiratory function during sleep, from primary snoring to obstructive sleep apnea (OSA) being the most severe. SDB can be detected using a complex and expensive exam called polysomnography. This exam monitors the sleep of a person during the night by measuring 21 different signals from an Electrocardiogram to Nasal Air Flow. Several au- tomatic methods have been developed to detect this disorder in adults, with a very high performance and using only one signal. However, we have not found similar algorithms especially developed for Children. We benchmarked 6 different methods developed for adults. We showed empirically that those models’ per- formance is drastically reduced when used on children (under 15 years old). Afterwards, we present a new approach for screening children with risk of having SDB. Moreover, our algorithm uses less information than a polysomnography and out performs state-of-the-art techniques when used on children. We also showed em- pirically that no signal alone is a good SDB screening in children. Moreover, we discover that combinations of three signals which are not used in any other previous work are the best for this task in children. © 2015 Elsevier Ltd. All rights reserved. c i m ( R p n e r ( i n A a 2 c b t a a 1. Introduction Several health systems from different countries have agreed on the need of tackling public health by the means of a more preven- tive, personalized and anticipatory health service. This requires to en- couraging patients and caregivers to make them protagonists of their healthcare, assuming responsibility to keep themselves on controlled states without collapsing healthcare centers. In Chile, we have a severe case of air pollution, which produce episodes where hospitals collapse during autumn and winter. This is why; Health Ministry is developing several programs that aim to treat patients at home. But to do so, we need to know exactly a patient, for example, what is their normal biomedical signals values (mean and standard dev.) or what happens with humidity during winter inside their home or other environmental conditions. Based on specific data and other information from the patients and knowledge extracted from experts we can develop an expert system similar to Echeverría, Jimenez-Molina, and Ríos (2015) that aid patients and caregivers to take better decisions (even from their homes). This article focuses in the creation of an automatic screening method for Sleep Disordered Breathing (SDB) which are a group of chronic diseases that affect the normal respiration function during the night. These can affect people at any age and are due to different ∗ Corresponding author. Tel.: +56 968353577. E-mail addresses: srios@dii.uchile.cl, sebastian@rios.tv (S.A. Ríos), lerazo@ing.uchile.cl (L. Erazo). o h e P http://dx.doi.org/10.1016/j.eswa.2015.11.013 0957-4174/© 2015 Elsevier Ltd. All rights reserved. auses: in newborns and young children they are related to congen- tal defects or premature birth; in older children and adults they ay be related to obesity, morphological causes, and hypertension Goodwin et al., 2003a; 2003b; Levy, Bonsignore, & Eckel, 2009; osen, Palermo, Larkin, & Redline, 2002). The diseases -from least to most severe- considered as SDB are: rimary snoring, upper airway resistance, and obstructive sleep ap- ea (OSA). For example, a newborn with OSA can experience short pisodes of apnea, total absence of airflow during night and live a elatively normal life, while a severe OSA may lead to sudden death Goldstein et al., 2004). Symptoms of SDB are abnormal day sleepiness, sudden naps dur- ng the day (for example, at a red light while driving), general tired- ess, fatigue, trouble sleeping, and other related diseases (de la Luz lonso Álvarez et al., 2008). It has been shown that in children SDBs re closely related to obesity and learning problems (Goodwin et al., 003a). The diagnosis of SDB is accomplished through a clinical study alled polysomnography (PSG) which collects over 20 different iomedical signals during sleep, including: electrocardiography, elec- roencephalography, electromyography, plethysmography, oronasal irflow, chest movement, abdominal movement and leg movement mong others. Several automatic methods have been developed in the form f expert systems to detect these disorders in adults, with a very igh performance (Álvarez-Estévez & Moret-Bonillo, 2009; De Chazal t al., 2003; Driver et al., 2011; Jarvis & Mitra, 2000; Khandoker, alaniswami, & Karmakar, 2009). However, we showed empirically http://dx.doi.org/10.1016/j.eswa.2015.11.013 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2015.11.013&domain=pdf mailto:srios@dii.uchile.cl mailto:sebastian@rios.tv mailto:lerazo@ing.uchile.cl http://dx.doi.org/10.1016/j.eswa.2015.11.013 S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 43 b c t n u o f h s p E n h s a t n t 2 b n M s p s e o e r b l b d J S d n c d o b r c b i b w s t E c w w c a t s A w 7 a e a 3 c H t e m g m m y Erazo and Ríos (2014) that those models’ performance is drasti- ally reduced when used in children (under 15 years old). This time we present a new algorithm able to classify children into wo groups: the group at risk of having an SDB, and the group with o risk (or very little risk) of having an SDB. Moreover, our algorithm ses less information than a polysomnography and surpasses state- f-the-art techniques when used on children. This is very important actor since in order to develop a non-invasive solution (to be used at ome, for example) we need to reduce the amount of signals used by creening methods. We showed empirically that no signal alone can be a good OSA redictor in children and also we showed that only three signals: MG-Chin, EOG and Leg Movement are very good predictors. We eed to remark this result, since until today, none of these signals ave been used on their own or in combination with other signals for creening in any documented study reviewed in this research. This work is based on data collected by Pablo E. Brockmann M.D. nd his research team in the Sleep Study Center of Clinical Hospi- al of Catholic University of Chile. This dataset consists in 78 whole ight polysomnographies from patients under 15 years old. This is he largest dataset of this characteristics used for OSA screening. . Related work Our main interest is to develop an algorithm that can distinguish etween OSA and non-OSA individuals with the least amount of sig- als from the PSG. In adults, several researchers (Álvarez-Estévez & oret-Bonillo, 2009; Driver et al., 2011; Flemons et al., 2003) have hown that some signals from PSG have enough predictive power to erform this task. In fact, one-signal screening methods have been uccessfully tested in adults and children (Roche et al., 2003; Tsai t al., 2013), but those methods still require attending personnel and vernight dedicated systems. Most of them also required a medical valuation afterwards. Automated classification methods aim to avoid unnecessary esource consuming screening methods. The most important contri- ution to this respect was made by the Computers in Cardiology Chal- enge of 2000. In that work the task was to automatically tag, minute y minute, a single ECG signal as OSA or non-OSA and get to a final iagnosis: OSA or non-OSA for every record (De Chazal et al., 2003; arvis & Mitra, 2000; Khandoker et al., 2009; Mendez et al., 2007). ome of these methods reached a precision of 100% in the binary iagnosis and over 85% in minute-by-minute tagging. Unfortunately, Table 1 Automated OSA screening approaches tested in our previous work (Erazo & Ríos, 2014). Author, Year Brief description R Jarvis and Mitra (2000) Apnea diagnosis based on ECG signal. Features used derived from spectral analysis. Classification based on threshold criteria. A De Chazal et al. (2003) Apnea diagnosis based on ECG signal. Features extracted using black-box methods. Classification methods: linear discriminant, quadratic discriminant. A Mendez et al. (2007) Apnea binary diagnosis based on ECG signal. Features extracted: derived signals (EDR, HRV) coefficients. Classification method: K-Nearest Neighbour. A Álvarez-Estévez and Moret-Bonillo (2009) A fuzzy reasoning is used to detect apneic events; Air flow signal and Oxygen Saturation signals are used Se Khandoker et al. (2009) Apnea binary diagnosis (OSA+, OSA-) based on ECG signal. Features extracted: Wavelet Transform coefficients. Classification method: Support Vector Machine. A Driver et al. (2011) Validation of MediByte (portable monitor) as screening method. Se one of these studies on automatic classification was performed on hildren. Besides, most of the algorithms were trained and tested on atabases specially designed for this task. In particular, models tested n the Computers in Cardiology 2000 database can not be compared, ecause this database has been preprocessed to obtain clean, but not ealistic data. This is why we performed a benchmark with real data from 78 hildren and compared the results among them. This was published y Erazo and Ríos (2014) and a summary of methods tested is shown n Table 1. The best classifiers from all approaches tested were those ased on: a Support Vector Machine (SVM), an Artificial Neural Net- ork (ANN). Besides we implemented a Logit classifier to evaluate a imple model, though it is not in the literature. We also selected the hree signals that reported best results in the literature, which are: lectrocardiography (ECG or ECGI channel), Air Flow (Patient Airflow hannel) and Oxygen Saturation (SpO2 channel). Afterwards, we pre-processed ECG, Air Flow and SpO2 with avelet transform to extract features. All algorithms were trained ith all 14 features generated. Then we trained the models with a ross-validation approach with 70% of the dataset to train the models nd 30% to test; and finally, experiments were performed 30 times o generate the final benchmark results computing several measures hown in Section 3.3. The best performing models were the ANN applied over oronasal ir Flow signals, with Sensitivity = 84.36%. The second best model as the SVM with Air Flow signal that resulted in a Sensitivity = 5.64%. Both classifiers were far from the 90% sensitivity considered s the clinical minimum for a successful classifier. We demonstrated xperimentally that state-of-the-art models for OSA screening in dults are not good enough to be used in children. . Model construction methodology After performing our benchmark by Erazo and Ríos (2014), we ould not find a good predictor for OSA+ and OSA- in children’s data. owever, models using Air Flow signal had an outstanding sensi- ivity, over 80% and regular accuracy (in the Neural Network mod- ls). ECG based models, on the other hand showed high specificity, eaning that they have the ability to detect healthy people. This sug- ests that a combination of these signals may lead to a successful odel. This section describes a novel approach to this task. From a purely athematical point of view, signal selection is performed in order esults Comments cc=100% Tested on Computers in Cardiology 2000 database. 75 records tagged minute-by-minute. All registers corresponding to adults. cc=100% Tested on Computers in Cardiology 2000 database. 75 records tagged minute-by-minute. All registers corresponding to adults. cc=85,5%, Sens=83,9%, Spec=88,5% Tested on Physionet Apnea ECG Database (50 registers) tagged as OSA+ or OSA-. ns=87%, Spec=89% Tested on a 12 patients database (Sleep Heart Health Study), it does not diagnose OSA, it only detects apneic events. All registers corresponding to adults. cc=100% Tested on three combined databases: Sleep Research Unit Database, Physionet Apnea ECG Database and Saint Vincent’s University Hospital/University College Dublin Sleep Apnea Database. All 125 registers corresponding to adults. ns=80%, Spec=87% This monitor requires a clinician to perform the diagnosis. 44 S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 Table 2 State-of-the-art benchmark results’ summary (Erazo & Ríos, 2014). Method WT func. ECG Air flow SpO2 Sens Spec Acc Sens Spec Acc Sens Spec Acc [%] [%] [%] [%] [%] [%] [%] [%] [%] SVM Db1 8,3 82,1 50,0 75,6 26,3 54,2 9,7 82,6 50,9 Db9 2,7 88,2 51,0 75,4 28,7 55,1 15,0 79,7 51,6 Haar 10,0 79,7 49,4 70,8 32,7 54,2 13,3 77,2 49,4 Sym8 5,3 85,1 50,4 75,1 27,0 54,2 12,3 82,1 51,7 NN Db3 34,3 60,3 49,0 84,4 19,0 55,9 30,0 75,1 55,5 Db4 36,3 63,9 51,9 80,5 21,7 54,9 19,7 77,7 52,5 Haar 25,7 65,9 48,4 80,3 26,3 56,8 38,7 60,0 50,7 Sym1 28,7 67,7 50,7 80,3 24,3 55,9 41,0 63,3 53,6 Logit Db3 44,6 54,5 50,2 18,9 74,6 43,7 44,5 51,4 48,4 Db4 41,9 53,0 48,2 19,3 78,9 44,9 55,1 61,9 59,0 Haar 43,6 58,4 51,8 39,3 58,1 46,3 47,0 55,8 52,1 Sym7 43,9 51,0 47,9 19,1 75,6 43,4 52,2 57,0 55,0 A A a t f r E D t 2 ( l p t s f f c e M V E C D fi a f f m 3 t to construct a minimal signal model able to satisfactorily screen the pediatric population. 3.1. Data understanding All children included in this study came to the Sleep Center be- cause their parents or doctors had concerns about the quality of their breathing during sleep. Patients’ parents had to answer a stan- dard questionnaire in order to evaluate the severity of the potential disorders. Only children considered as a risk population underwent polysomnographic test. Each patient spent a whole night in the hospital where 20 or 21 different biomedical signals were recorded (approximately 8 h of records),1 including Electroencephalography (11 channels: EEG F3- 1,EEG C3-A1, EEG P3-A1, EEG O1-A1, EEG F4-A2, EEG C4-A2, EEG P4-A2, EEG O2-A2, EEG A1-A2, EEG Fp1-A1, EEG Fp2-A2), Electro- cardiography (ECG or ECGI channel), Electroculography (EOG Right and EOG Left), Electromiography (EMG Chin), Air Flow (Patient Air- flow channel) and Oxygen Saturation (SPO2 channel) among others. PSG was collected in Sleep Study Center with an Alice® 5 Diagnostic Sleep System from Respironics (Philips) and transformed using the corresponding software (Alice® Sleepware) to the generally accepted format for biomedical time series: European Data Format (EDF). Afterwards each test was manually scored by Dr. Pablo E. Brock- mann or one of his fellows according to AASM (American Academy of Sleep Medicine) criteria (General, 2005). Apnea, hypopnea, arousals, leg movement, and any other event of interest were reported. Data corresponds to 78 complete night PSG from 78 different pa- tients of ages from 2 to 16 years old (mean ± SD: 9.7 ± 3.6 years). Sixty-four percent of them are girls. According to AASM criteria, children are diagnosed with OSA with an Apnea Hypopnea index equal or greater than 1 (General, 2005). Using this specification, 43% of the PSG correspond to children with OSA. No PSG was excluded because of other diseases presented in patients. It is important to mention that most studies of OSA screening have small databases, no more than 20 registers. Even though some bigger databases are available, in child data, this is the biggest database doc- umented (as far as this research went). 3.2. Feature extraction The feature extraction method selection was based on the most referenced one in literature, as explained by Erazo and Ríos (2014). 1 originally records were almost 9 hours long, but after data cleaning; approximately 8 h remained. n A c ccording to this criteria, the Wavelet Transform method was dopted. Wavelet Transform is used in 1D non–stationary signals to process ime series as the ones collected by PSG. It allows for the collecting of requency information contained in the time series. Also, there is no equirement of the frequency stability, thus it is frequently used for CG feature extraction. Several wavelets were tested in order to extract features. aubechies and Symlet of levels from 1 to 10 were tested, in addi- ion to the Haar function. This means that every signal was processed 1 times under each of these 21 functions of a Wavelet Transform Daubechies1, Daubechies2, Daubechies3, and so on). To process the dataset, the Wavelet Toolbox for Matlab was se- ected. With this objective, EDF’s had to be transformed into a com- atible Matlab format. Each signal processed with Wavelet Transform was decomposed o its 14th detail level, obtaining 14 time series for each decompo- ition function, i.e. 14 time series for ECG processed with the Haar unction, 14 time series for ECG processed with Daubechies of level 1 unction, and so on. In order to reduce data and extract features, three features were alculated for each time series, these were: mean, variance and en- rgy; calculated as follows: ean: x̄ = 1 N N∑ i xi (1) ariance: σ 2 = N∑ i (x̄ − xi )2 (2) nergy: E = N∑ i x2i (3) oefficients obtained from the preprocessing stage were named b1mean1, Db1var1, Db1ene1 for the mean, variance and energy of the rst detail level for Daubechies1 WT function. The same process was pplied for every WT function and to every signal. This way, 21 dif- erent databases were generated for every signal –one for each WT unction–, each of them containing 42 features: mean1, var1, ene1, ean2, var2, ene2 up to mean14, var14, ene14. .3. Evaluation measures In order to compare models we will compute well known evalua- ion measures. The main point is to describe the strengths and weak- esses of the models (Fig. 1). Indicators selected for these tasks are Sensitivity, Specificity and ccuracy (commonly used to assess clinical tests performance); Pre- ision, Recall and F-measure (to assess the classification performance S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 45 Feature Extraction Feature Selection Minimum Feature Selection PSG Data Multi-signal Models Training SVM ANN Logit Sub set of features (70% data) Multi-signal Models Test SVM ANN Logit Sub set of features (30% data) Cross-Validation Evaluation Sensitivity Specificity Precision Recall FAccuracy Fig. 1. Summary of Experimental Methodology. o R S A P F I h c fi i c g a p 3 u A p p a Q s a v t d n i c s M f algorithms). They will be calculated as follows: ecall = Sensitivity = T P T P + F N (4) pecificity = T N T N + F P (5) ccuracy = T P + T N T P + T N + F P + F N (6) recision = T P T P + F P (7) = 2 ∗ T P 2 ∗ T P + F N + F P (8) nterpretation of these indicators is based on the ability to detect un- ealthy individuals. This is because it is more important to correctly lassify sick children as sick than healthy children as healthy. Sensitivity is the ability to detect unhealthy individuals, Speci- city is the ability to detect healthy people, Accuracy is the abil- ty to correctly classify, that is, how many children were correctly lassified. On the other hand, Precision is the proportion of the ill individuals roup that are correctly classified. Finally, the F-measure shows how ccurate is the model, is in a range of 0 to 1, where 1 is the best model ossible. .4. ECG feature extraction To extract relevant features from an ECG signal, it is important to nderstand how it is analyzed by clinicians. Fig. 2 shows a sample of an ECG signal corresponding to the 0011823 patient. A regular ECG wave can be characterized according to some oints. As seen in Fig. 3, points P, Q, R, S and T describe a complete ulse of an ECG. The relevant characteristics for respiratory functions are associ- ted with the QRS complex. This is the name of the area formed by , R and S points. Commonly the features are extracted from derived ignals from the ECG recording. These signals are: Heart Rate Vari- bility (HRV) and ECG Derived Respiratory Signal (EDR). HRV corresponds to the time series constructed from R–R inter- als; this is the time between successive R points. EDR corresponds o the time series constructed from QRS amplitude. This is the vertical istance between Q and R points. Before extracting features, all signals had to be normalized to ig- ore the variance associated with natural variations among different ndividuals. While a certain heart rate might be normal to a person, it ould be considered accelerated for another, so normalization is the tandard proceeding when working with biological signals. Common features extracted from HRV and EDR signal are: ean: x̄ = 1 N N∑ i xi (9) 46 S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 Fig. 2. ECG sample. Fig. 3. ECG wave characteristics. Fig. 4. Electrode positions for an EEG. 3 E a n Variance: σ 2 = 1 N N∑ i (x̄ − xi )2 (10) Also, some time domain features were extracted from the HRV signal. In order to do this a new time series had to be constructed, corresponding to the successive differences of the HRV signal. These features were: Root Mean Square of Successive Differences: = √ N∑ i (xi − xi−1 )2 (11) Standard Deviation of Successive Differences: = √ 1 N N∑ i (x̄ − xi )2 (12) With all these transformations the resulting feature set to character- ize the ECG signal is: meanHRV, varHRV, meanEDR, varEDR, RMSSD and SDSD. .5. EEG feature extraction The EEG performed during a PSG has 11 signals as output: EGF p1 − A1, EEGF p2 − A2, EEGA1 − A2, EEGF 3 − A1, EEGF 4 − A2, EEGC3 − A1, EEGC4 − A2, EEGP3 − A1, EEGP4 − A2, EEGO1 − A1 and EEGO2 − A2. Each of them characterized by the position of the elec- trode in the head, where A1 corresponds to the left side, and A2 to the right side, as seen in Fig. 4. Not all patients present every signal. In small children (under 2 years old) not all the electrodes are placed on the head, mainly be- cause they have smaller heads, so 31 of the registers do not contain the signals EEGA1 − A2, EEGF p1 − A1, EEGF p2 − A2, EEGO1 − A1 y EEGO2 − A2. Although, some techniques propose to process them ll together (by adding the signals or applying Principal Compo- ents Analysis for example), the selected approach was to process S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 47 Fig. 5. EEG sample corresponding to patient A0001945. Fig. 6. Spectrogram sample corresponding to patient A0001945. t s p s m r t a w o i u t a s t f c c t c u f f f W t F a t n e 3 g m p p e d m s f hem separately for two reasons: first, clinicians analyze them eparately, and second, if there is a combination that allows for rocessing them all together it should be reflected in the feature election stage. This is, if the signals are giving redundant infor- ation, the feature selection will detect this effect and eliminate edundancy. When a PSG is processed, clinicians look into EEG signals to tag ime windows according to the sleep stage the patient is in. For ex- mple in Fig. 5 is easy to distinguish two stages. So, the main task ith these signals is to detect variations in the wave frequencies in rder to characterize different stages. What matters most is to determine how much a wave varies dur- ng night, for which a methodology was developed. First, a spectrogram was constructed. Spectrograms are frequently sed to analyze sound waves or pixels. They are a graphic represen- ation of the variety of frequencies in a wave as time varies. On the X-axis a spectrum of frequencies is represented. On the Y- xis the time is divided in small windows for analysis, and the color cale shows how strong is the presence of a frequency in a specific ime window. If columns are analyzed independently, the variation of a specific requency can be observed. Based on this, the features to be extracted haracterize the number of times a big variation occurs, that is, a hange in the sleep stage is suspected. If smaller time windows are used, smaller changes can be de- ected, but this effect is not desired, since changes in sleep stages are haracterized by big variations, so only eight time windows will be sed for the analysis. In order to do this, an auxiliary time series is defined as the dif- erence of a frequency from one time window to the next. This is per- ormed for every frequency considered in the spectrogram (Fig. 6). The main idea is to detect significant variations, so the relative dif- erences were calculated as the difference divided by the initial value. ith this, a new time series was constructed for each signal. Later, his signal was filtered to make it easier to detect big variations. Fig. 7 shows the processed signal corresponding to the sample in ig. 5, every frequency analyzed is shown in a different color, the vari- tion in time is represented graphically. This means, peaks represent he big changes indicating a change in the sleep stage is suspected. Finally, features extracted from each EEG signal correspond to the umber of local minimum and the number of local maximum from very frequency filtered signal. .6. Abdominal and thoracic effort feature extraction Abdominal and Thoracic Effort Signals correspond to a band that oes around the abdomen and chest to register a patient’s move- ents while sleeping. The main idea is to detect peaks of movement, that is, when a atient has a regular respiratory function, no abnormality should ap- ear in the signal. In Fig. 8 for example, an abnormality appears at the nd of both signals (time ∼ 8500). In order to detect peaks wavelet ecomposition was used for its ability to show them in the signal. The ethodology implemented to extract features from the signal con- iders the mean, standard deviation and energy of the signals derived rom the decomposition. 48 S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 Fig. 7. Processed EEG signal corresponding to patient A0001945. Fig. 8. Abdominal Thoracic Signal (blue) and Abdominal Effort Signal (red) corresponding to patient A0000678. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) Fig. 9. Different Wavelet Decompositions (blue) for the Thoracic Effort Signal (red) corresponding to patient A0000678. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) t l 3 a As said in the previous section, many families of functions can be used with this decomposition. A visual inspection was done to select the signal that best fits the objectives of this task. Fig. 9 shows three of the wavelet functions tested. Although all functions seem to detect variations equally, Daubechies 3 had a better performance amplifying small variations, as the one seen at the end of the time window shown in the figure. This is why he function selected to process both signals was Daubechies 3 of evel 4. .7. Air flow feature extraction The Air Flow Signal registers the air movement through the nose nd the mouth. So, to detect the presence of OSA, one must detect S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 49 Fig. 10. Air Flow Signal corresponding to patient A0000739. Fig. 11. (a) Leg1 Signal corresponding to patient A0001484. (b) Leg2 Signal corresponding to patient A0001484. (c) Legs Signal corresponding to patient A0011947. w o a W f t 3 m e p m w d o e f s d 3 f m m b p t e a t t 3 t t d hen an abnormal flow occurs, this is, the complete absence of flow r a big decrease in it. Fig. 10 shows a complete night record. The task is to detect the flat reas in the signal. The same task was approached by Erazo and Ríos (2014) using a avelet Transform where the features extracted were used as input or one-signal models. This resulted in high performance models. Therefore the same strategy was followed and the selected func- ion was Daubechies 3 of level 14. .8. Leg movement feature extraction Of the 78 patients, 36 had only one signal that recorded their leg ovement during sleep. Forty-one of them had only one record for ach leg, this is two signals per patient to record leg movement. One atient had no record. As seen on Fig. 11 both legs are highly correlated, so no lost infor- ation is suspected in the case of only one record. Yet, both signals ere included in the analysis. If the information results are redun- ant, the Feature Selection Stage should detect it. The characterization of the signal had to represent the variation f the leg movement to detect abnormal movement or big movement pisodes. As before, the Wavelet Transform has the ability to extract these eatures from the signal. After testing a variety of functions, Daubechies 8, of level 6 was elected. As in previous cases, mean, variance and energy of every etail level were calculated, obtaining a set of 18 features. .9. Body position feature extraction Body Position signal detects the changes of position of the body, rom supine (back facing down) to prone (chest facing down). The ain idea is to detect change in positions, as a measure of the move- ent during the night. The sensor used for this task assigns a num- er to each position. It is a nominal variable that detects changes in osition. In order to reflect the number of times a patient moves during he night an auxiliary signal was constructed as the successive differ- nces of the signal. Therefore only changes in position would appear s different from zero as seen in Fig. 12. With this, the feature ex- racted was an index of the number of changes divided by the sleep ime. .10. Electromyogram feature extraction The Electromyogram signal (EMG) registers the electrical poten- ial generated by muscle cells, in a PSG in particular is used to regis- er chin movement during sleep. Chin movement is important in the iagnosis of SDB, and OSA, because respiration may occur through 50 S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 Fig. 12. (a) Body Position Signal corresponding to patient A0001460. (b) Auxiliary signal derived from Body Position Signal for patient A0001460. Fig. 13. EMG Chin Signal corresponding to patient A0002197. 3 o i a f t t l e t p i i f m nose or mouth, thus chin movement is used as an indicator of mouth movement. On the other hand, tension in the chin might indicate snoring or another abnormalities during sleep, like bruxism for example. As other signals that register movement, signal processing was made by a Wavelet Transform. The same selection process was per- formed, the chosen function was Daubechies 7 of level 8, resulting in 24 features (Fig. 13). 3.11. Electro-oculogram feature extraction The Electro-oculogram (EOG) registers the movement of the eyes by placing electrodes around them. The result of this exam is one sig- nal for each eye. So, in a complete PSG, two signals correspond to EOG, and they are tagged as EOG Right, and EOG Left for the right eye, and the left eye respectively. Even though these are separated signals, a high correlation is ex- pected, since eye movements are not independent from each other in most cases. In Fig. 14 we see the correlation between both left and right eye of patient A0006225. The same correlation is observed in every record from this signal, yet every signal was processed because the feature extraction process should be able to detect these highly correlated features. In order to be consistent in the treatment of signals, all movement signals were processed with Wavelet Transform. The function used for the EOG feature extraction was Daubechies 7 of level 8. .12. Pulse feature extraction The pulse signal as predictor for OSA prevalence has been used in ther studies. As expected, an abnormal pulse rate might be a good ndicator of an apnea episode due to the lack of oxygen generated as consequence. In particular, Noehren et al. (2010) use four time series derived rom the pulse signal: • APRD: Absolute Pulse Rate Decrease • APRI: Absolute Pulse Rate Increase • RPRD: Relative Pulse Rate Decrease • RPRI: Relative Pulse Rate Increase For the purpose of this study, only the first two were used. The pulse signal was normalized to avoid variation among pa- ients (see Fig. 15(a)). Then a Butterworth filter was used to eliminate he noise from the signal and detect easily peaks of local minima and ocal maxima (see Fig. 15(b)). The APRD length corresponded to the number of peaks found, and ach value was calculated as the difference of a local minimum and he local maximum immediately before (see Fig. 15(d)). The same rocedure was applied to construct the APRI (see Fig. 15(c)), calculat- ng the difference between a local maximum and the local minimum mmediately before. After this, the results are two time series. The decision is to extract, rom both of them: mean, standard deviation, median, maximum and inimum. S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 51 Fig. 14. EOG Sinals of Left (blue) and Right (Red) ayes corresponding to patient A0006225. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) Fig. 15. (a) Pulse signal corresponding to patient A0000294. (b) Filtered pulse signal. (c) Absolute Pulse Rate Increase (APRI) derived from pulse signal. (d) Absolute Pulse Rate Decrease (APRD) derived from pulse signal. 3 l c s v b v s a i t s e w 3 r p a a t I F l l d ( d a w a .13. Snore feature extraction The snore signal is used to diagnose not only OSA, but other SDB’s, ike primary snoring. Although very useful for SDB diagnosis, the final lassification into OSA+ and OSA- groups can not rely only on this ignal. Snoring is measured by a sensor located on the throat that detects ibration. As snoring is produced by an obstruction of the airway, vi- ration is produced, and measured this way. The resulting signal is normalized since the relevant aspects are ariations in the signal. As output for the PSG, a signal from 0 to 1 is hown. The goal when processing this signal was to determine when an bnormal vibration episode occurred. The signal provided by the PSG s already normalized, so an abnormal episode was considered when he signal crosses over 0.7 or under 0.3 (blue lines on Fig. 16). This election was based on preliminary analysis of the signal behavior. The features extracted correspond to the number of abnormal pisodes up and abnormal episodes down. The resulting feature set as: Snoreup and Snoredown. .14. Oxygen saturation feature extraction PSG also include a sensor that measures the blood’s oxygen satu- ation during the whole night. Usually saturation measures are done eriodically. Therefore a PSG is one of the few clinical tests that allows Saturation Signal. Evidently this is one of the most important signals to determine diagnosis, since it is directly related to apnea episodes. A desatura- ion episode is a clear symptom of an abnormal respiratory function. n particular, it might be the perfect indicator of an apnea episode. ig. 17 shows a time window of the Oxygen Saturation Signal col- ected by a PSG from patient A0001183. Many indicators are used for processing this signal. The most se- ected and relevant for this study are: DEI (Desaturation Event In- ex), defined as a saturation signal drop in four or more points on a scale from 1 to 100); and ODI (Oxygen Desaturation In- ex), also known as CT90 (Cumulative Time 90). Corresponds to the mount of time the patient has a saturation under 90%. DEI and ODI ere selected because of their previous use in continuous positive irway pressure devices (CPAP) tests and previous experiences in 52 S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 Fig. 16. Snore Signal corresponding to patient A0000884. Fig. 17. Oxygen Saturation Signal corresponding to patient A0001183. w b f c t P s diagnosing apnea episodes (Chang, Wu, & Cao, 2012; Chung et al., 2012). So, as result, the feature set was: DEI and ODI for each signal. 4. Feature selection Resulting from the Feature Extraction process, 355 features are used to resume the information collected in a PSG. Table 3 summa- rizes the features resulting from the previous section. The hypothesis is that not all the information is needed to perform a high quality screening. If this is true, a less resource-consuming clinical test may be developed to pre-diagnose children with sus- pected OSA. The first step to accomplish this is to select features. The selec- tion should imply a reduction in the signals needed to perform a screening. A filter approach was selected to perform the Feature Selection. This allows the use of the same subset with all three models and al- lows comparison under the same conditions as shown by Erazo and Ríos (2014). Embedded methods were discarded due to their dependency on the selected model. Therefore, different models would have ended up ith different feature subsets. Wrapper methods were discarded also, ecause they did not allow for the control of the number of resulting eatures in the subset. Before starting the selection, data quality was checked. Features orresponding to the Pulse signal were discarded since only 6 pa- ients had complete Pulse records. The selection procedure was based on correlation analysis and rincipal Component Analysis. The step by step methodology is de- cribed as follows: 1. First. Principal Component Analysis (PCA) was computed over all the Features included, using Varimax rotation criterion. The main objective was to extract variables that explain most of the vari- ance in order to incorporate as much information as possible in the modeling section. 2. Second. The first five components resulting from Factorial Anal- ysis were analyzed, as they explained 89.9% of the variance, and the sixth component added only 2% to the variance explanation. 3. Third. Variables eigenvalues were extracted, and only variables with an eigenvalue of 0.05 (5%) or more for the first five principal components were kept. 4. Fourth. As mentioned, all features corresponding to selected sig- nals were kept. Resulting from these steps, features corresponding S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 53 Table 3 Summary of resulting features from a complete PSG. Signal Features extracted Total ECG meanHRV , varHRV , meanEDR , varEDR , RMSSD and SDSD. 6 EEG (11) max1 , max2 , max3 , max4 , max5 , max6 , max7 , min1 , min2 , min3 , min4 , min5 , min6 and min7 . 154 Abdominal effort Db3mean1 , Db3var1 , Db3ene1 , Db3mean2 , Db3var2 , Db3ene2 , Db3mean3 , Db3var3 , Db3ene3 , Db3mean4 , Db3var4 y Db3ene4 . 12 Thoracic effort Db3mean1 , Db3var1 , Db3ene1 , Db3mean2 , Db3var2 , Db3ene2 , Db3mean3 , Db3var3 , Db3ene3 , Db3mean4 , Db3var4 y Db3ene4 . 12 Air flow Db3mean1 , Db3var1 , Db3ene1 , Db3mean2 , Db3var2 , Db3ene2 , Db3mean3 , Db3var3 , Db3ene3 , Db3mean4 , Db3var4 , Db3ene4 , Db3mean5 , Db3var5 , Db3ene5 , Db3mean6 , Db3var6 , Db3ene6 , Db3mean7 , Db3var7 , Db3ene7 , Db3mean8 , Db3var8 , Db3ene8 , Db3mean9 , Db3var9 , Db3ene9 , Db3mean10 , Db3var10 , Db3ene10 , Db3mean11 , Db3var11 , Db3ene11 , Db3mean12 , Db3var12 , Db3ene12 , Db3mean13 , Db3var13 , Db3ene13 , Db3mean14 , Db3var14 and Db3ene14 . 48 Leg movement (2) Db8mean1 , Db8var1 , Db8ene1 , Db8mean2 , Db8var2 , Db8ene2 , Db8mean3 , Db8var3 , Db8ene3 , Db8mean4 , Db8var4 , Db8ene4 , Db8mean5 , Db8var5 , Db8ene5 , Db8mean6 , Db8var6 and Db8ene6 . 36 Body position CI 1 EMG Db7mean1 , Db7var1 , Db7ene1 , Db7mean2 , Db7var2 , Db7ene2 , Db7mean3 , Db7var3 , Db7ene3 , Db7mean4 , Db7var4 , Db7ene4 , Db7mean5 , Db7var5 , Db7ene5 , Db7mean6 , Db7var6 , Db7ene6 , Db7mean7 , Db7var7 , Db7ene7 , Db7mean8 , Db7var8 and Db7ene8 . 24 EOG (2) Db7mean1 , Db7var1 , Db7ene1 , Db7mean2 , Db7var2 , Db7ene2 , Db7mean3 , Db7var3 , Db7ene3 , Db7mean4 , Db7var4 , Db7ene4 , Db7mean5 , Db7var5 , Db7ene5 , Db7mean6 , Db7var6 , Db7ene6 , Db7mean7 , Db7var7 , Db7ene7 , Db7mean8 , Db7var8 and Db7ene8 . 48 Pulse APRDmean , APRDstd , APRDmedian , APRDmax , APRDmin , APRImean , APRIstd , APRImedian , APRImax andAPRImin . 10 Snore Snoreup and Snoredown 2 Oxygen saturation DEI and ODI 2 TOTAL 355 Table 4 Summary of selected features. Signal Features selected Total Leg 1 Db8mean1 , Db8mean2 , Db8var2 , Db8mean3 , Db8mean4 , Db8mean5 , Db8var5 and Db8mean6 . 8 Leg 2 Db8mean3 , Db8mean4 , Db8mean5 and Db8var5 . 4 Legs Db8mean1 , Db8var1 , Db8mean2 , Db8mean3 , Db8mean4 and Db8mean5 . 6 EOG right Db7mean1 , Db7mean2 , Db7mean3 , Db7mean4 , Db7mean5 , Db7mean6 , Db7mean7 and Db7var7 . 8 EOG left Db7mean1 , Db7var1 , Db7mean2 , Db7mean3 , Db7var3 , Db7mean4 , Db7mean5 , Db7mean6 , Db7mean7 , Db7var7 and Db7mean8 . 11 EMG chin Db7mean1 , Db7var1 , Db7mean2 , Db7mean3 , Db7mean4 , Db7mean5 , Db7mean6 , Db7mean7 , Db7var7 , Db7mean8 and Db7var8 . 11 TOTAL 51 S 5 t p s d a c r s Table 5 SVM performance using the feature subset from PSG. Model Sensitivity Specificity Accuracy Precision Recall F-measure SVM 100,00% 100,00% 100,00% 100,00% 100,00% 100,00% Table 6 NN performance using the feature subset from PSG. Model Sensitivity Specificity Accuracy Precision Recall F-measure NN 28,67% 100,00% 28,67% 100,00% 28,67% 39,91% Table 7 Logit Regression performance using the feature subset from PSG. Model Sensitivity Specificity Accuracy Precision Recall F-measure LOGIT 47,30% 100,00% 47,30% 100,00% 47,30% 62,69% 6 m w p s c m b 7 a g f p c s i b to the following signals formed the preliminary subset: EMG Chin, EOG (Left and Right), Snore and Legs. Preliminary subset had 129 features. 5. Fifth. Correlation Matrix was computed over the preliminary sub- set. 6. Sixth. Every pair of signals was analyzed. If a correlation of over 0.6 was found, only one of the pair of signals was kept. 7. Seventh. Resulting from these steps, 51 features remain in the subset. This is the definitive Features Subset. Table 4 shows the resulting subset of features from the Feature election process. . Classification models Models used to classify patients into the two groups defined are he same as those used by Erazo and Ríos (2014). These are: Sup- ort Vector Machine, Artificial Neural Networks and Logit Regres- ion. Also, the same validation methodology was implemented, as escribed in Section 3.3. Thirty different training and testing balanced sets were created nd used to train and test every model. Performance metrics were alculated each time. The resulting performance corresponds to the average of the met- ics resulting from every iteration (this is, the result from one training et and one testing set). . Results Tables 5–7 describe the results obtained from each classification odel separately. Support Vector Machines showed an outstanding performance hen the minimal signal set are used as input. This result is unex- ected because, as far as this research went, no other published re- earch used EMG, EOG and leg movement as input signals. In fact, linicians use these signals for auxiliary information, and never as the ain source of information for diagnosis. The other models showed poorer performance compared to SVM, ut they also showed unexpected Precision results. . Discussion Having as a starting point the previous results presented by Erazo nd Ríos (2014), where it is shown that no signal alone can be a ood predictor for OSA in children when using black-box methods or feature extraction. As a consequence, it can be inferred that a more hysiological approach to feature extraction may lead to good quality lassification methods. Evidence in adults and some studies conducted in children howed good-quality screening methods were mainly based on phys- ological feature extraction methods. Most of these studies also, were ased on one-signal only. The few of them that were based on more 54 S.A. Ríos, L. Erazo / Expert Systems With Applications 48 (2016) 42–54 A p t P a o U R Á C C D D E E F G G G G J K L d M N R R T than one signal did not show any criterion for the signal selection beyond clinicians criteria. Always having present the goal of finding the minimum feature set that can effectively screen for OSA. This is why the approach pre- sented here describes a novel method for OSA screening: Signal se- lection was based only on statistical and optimization criteria. We discover that three signals: EMG-Chin, EOG and Leg Move- ment combined are the best predictor for OSA in children. Sur- prisingly none of these signals have been used on their own or in combination with other signals for screening in any documented study reviewed in this research. Also, all of this signals are collected with ordinary electrodes, widely available on the market. This fact allows us to think of this algorithm as an actual screening method for future research in devel- oping countries where resources are scarce, and OSA consequences in infant populations are largely overlooked. Our approach is able to screen OSA with 100% precision was devel- oped using signals collected by polysomnography and the pertinent information needed was reduced using data mining techniques, so only three signals are needed to perform a high quality screening. Another result of this work, a high quality repository of polysomnographic children data tagged by qualified clinicians is available for further testing and model construction. It is very important to consider that today is a world wide ten- dency the development of ubiquitous health (u-health) expert sys- tems that allow the monitoring of patients even at home (Echeverría et al., 2015). Most methods for screening OSA use oxygen saturation and ECG signals (see Table 1); which need very good quality sen- sors (expensive) in order to obtain good quality measures. This is a barrier to develop a system for use at home or in third world coun- tries. However, the sensors required to measure our approach signals (EMG-Chin, EOG and Leg Movement) are more simple in construc- tion, price and allow to have good quality data for a u-health expert system. 8. Recommendations and future work For signals like ECG or Air Flow, widely used in screening methods for adults, the results showed no relevance or correlation with the target variable in the feature selection process (see Section 4). Clini- cians base their diagnosis on these signals, although they use all the information available from a polysomnography. The intuitive reason- ing lead them to think that these signals should appear as relevant in the feature selection. This fact may be evidence that extracted fea- tures are not representative enough of the information contained in the signal. Therefore new approaches to feature extraction should be sought after with more adequate methods for each signal. Some research devote most of the work to performing feature extraction correctly, therefore a comparison of different feature extraction methods is rec- ommended in order to correctly represent signals for the feature se- lection process. If this algorithm, or any other, is intended to be implemented as a valid screening method, it must meet minimum requirements. First, is has to be precise. Second, it has to be easier to perform than the gold-standard, less resource consuming and independent of clinician supervision. The proposed method in this research actually meets some of these requirements. But there is a third requirement, not less important: clinicians must trust the screening test in order to actually prescribe it. As previously mentioned, as far as this research went, no docu- mented work uses any of the resulting signals (EMG Chin, EOG, Leg Movement) alone or combined with others for prediagnosis. This is mainly because these signals are not the main source of information in a regular diagnosis process, so is necessary to check doctor’s opin- ion and their disposition to prescribe this kind of clinical test. cknowledgment This work was financed by CONICYT - IDeA FONDEF programme, roject code: CA13i-10300. Also, authors would like to thank the con- inuous support of “Instituto Sistemas Complejos de Ingeniería” (ICM: -05-004- F, CONICYT: FBO16). Last but not least, we would like to thank Dr. Pablo Brockmann nd his team at Clinical Hospital from Pontifical Catholic University f Chile and Ivan Videla at Business Intelligence Research Center from niversity of Chile for his comments on the manuscript. eferences lvarez-Estévez, D., & Moret-Bonillo, V. (2009). Fuzzy reasoning used to detect apneic events in the sleep apnea-hypopnea syndrome. Expert Systems with Applications, 36, 7778–7785. hang, L., Wu, J., & Cao, L. (2012). Combination of symptoms and oxygen desaturation index in predicting childhood obstructive sleep apnea . International journal of pe- diatric otorhinolaryngology, 77(3), 365–371. hung, F., Liao, P., Elsaid, H., Islam, S., Shapiro, C. M., & Sun, Y. (2012). Oxygen desatu- ration index from nocturnal oximetry: a sensitive and specific tool to detect sleep- disordered breathing in surgical patients. Anesthesia & Analgesia, 114, 993–1000. e Chazal, P., Heneghan, C., Sheridan, E., Reilly, R., Nolan, P., & O’Malley, M. (2003). Automated processing of the single-lead electrocardiogram for the detection of obstructive sleep apnoea. IEEE Transactions on Biomedical Engineering, 50, 686–696. river, H. S., Pereira, E. J., Bjerring, K., Toop, F., Stewart, S. C., Munt, P. W., & Fitzpatrick, M. F. (2011). Validation of the medibyte type 3 portable monitor com- pared with polysomnography for screening of obstructive sleep apnea. Canadian Respiratory Journal: Journal of the Canadian Thoracic Society, 18, 137. cheverría, M., Jimenez-Molina, A., & Ríos, S. A. (2015). A semantic framework for con- tinuous u-health services provisioning. Procedia Computer Science, 60, 603–612. razo, L., & Ríos, S. A. (2014). A benchmark on automatic obstructive sleep apnea screening algorithms in children. Procedia Computer Science, 35, 739–746. lemons, W. W., Littner, M. R., Rowley, J. A., Gay, P., Anderson, W. M., Hudgel, D. W., et al. (2003). Home diagnosis of sleep apnea: A systematic review of the litera- turean evidence review cosponsored by the american academy of sleep medicine, the american college of chest physicians, and the american thoracic society. CHEST Journal, 124, 1543–1579. eneral, A. (2005). International classification of sleep disorders, 2nd edn.: Diagnostic and coding manual. westchester, illinois: American academy of sleep medicine, 2005. broughton r. pathological fragmentary myoclonus, intensified sleep starts and hypnagogic foot tremor: three unusual sleep related disorders. Sleep, 28, 113– 121. oldstein, N. A., Pugazhendhi, V., Rao, S. M., Weedon, J., Campbell, T. F., Goldman, A. C., et al. (2004). Clinical assessment of pediatric obstructive sleep apnea. Pediatrics, 114, 33–43. oodwin, J. L., Babar, S. I., Kaemingk, K. L., Rosen, G. M., Morgan, W. J., Sherrill, D. L., et al. (2003a). Symptoms related to sleep-disordered breathing in white and hispanic children. the tucson childrens assessment of sleep apnea study. CHEST Journal, 124, 196–203. oodwin, J. L., Kaemingk, K. L., Fregosi, R. F., Rosen, G. M., Morgan, W. J., Sherrill, D. L., et al. (2003b). Clinical outcomes associated with sleep-disordered breathing in caucasian and hispanic children. the tucson childrens assessment of sleep apnea study (tucasa). Sleep Disordered Breathing, 26, 587–591. arvis, M., & Mitra, P. (2000). Apnea patients characterized by 0.02 hz peak in the mul- titaper spectrogram of electrocardiogram signals. In Computers in cardiology 2000 (pp. 769–772).IEEE. handoker, A. H., Palaniswami, M., & Karmakar, C. K. (2009). Support vector machines for automated recognition of obstructive sleep apnea syndrome from ecg record- ings. Information Technology in Biomedicine, IEEE Transactions on, 13, 37–48. evy, P., Bonsignore, M., & Eckel, J. (2009). Sleep, sleep-disordered breathing and metabolic consequences. European Respiratory Journal, 34, 243–260. e la Luz Alonso Álvarez, M., Santos, J. T., Guevara, J. C., Martínez, M. G., Pascual, L. R., Ba nuelos, J. L. V., et al. (2008). Reliability of home respiratory polygraphy for the diagnosis of sleep apnea-hypopnea syndrome. analysis of costs. Archivos de Bron- coneumología (English Edition), 44, 22–28. endez, M. O., Ruini, D. D., Villantieri, O. P., Matteucci, M., Penzel, T., Cerutti, S., et al. (2007). Detection of sleep apnea from surface ecg based on features extracted by an autoregressive model. In Engineering in medicine and biology society, 2007. embs 2007. 29th annual international conference of the ieee (pp. 6105–6108).IEEE. oehren, A., Brockmann, P. E., Urschitz, M. S., Sokollik, C., Schlaud, M., & Poets, C. F. (2010). Detection of respiratory events using pulse rate in children with and with- out obstructive sleep apnea. Pediatric pulmonology, 45, 459–468. oche, F., Pichot, V., Sforza, E., Duverney, D., Costes, F., Garet, M., et al. (2003). Predict- ing sleep apnoea syndrome from heart period: a time-frequency wavelet analysis. European Respiratory Journal, 22, 937–942. osen, C. L., Palermo, T. M., Larkin, E. K., & Redline, S. (2002). Health-related quality of life and sleep-disordered breathing in children. Sleep, 25, 657–666. sai, C.-M., Kang, C.-H., Su, M.-C., Lin, H.-C., Huang, E.-Y., Chen, C.-C., et al. (2013). Usefulness of desaturation index for the assessment of obstructive sleep apnea syndrome in children. International journal of pediatric otorhinolaryngology, 77(8), 1286–1290. http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0001 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0002 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0003 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0004 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0005 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0006 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0007 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0008 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0009 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0010 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0011 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0012 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0013 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0014 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0015 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0016 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0017 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0018 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0019 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0020 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0021 http://refhub.elsevier.com/S0957-4174(15)00779-4/sbref0021 An automatic apnea screening algorithm for children 1 Introduction 2 Related work 3 Model construction methodology 3.1 Data understanding 3.2 Feature extraction 3.3 Evaluation measures 3.4 ECG feature extraction 3.5 EEG feature extraction 3.6 Abdominal and thoracic effort feature extraction 3.7 Air flow feature extraction 3.8 Leg movement feature extraction 3.9 Body position feature extraction 3.10 Electromyogram feature extraction 3.11 Electro-oculogram feature extraction 3.12 Pulse feature extraction 3.13 Snore feature extraction 3.14 Oxygen saturation feature extraction 4 Feature selection 5 Classification models 6 Results 7 Discussion 8 Recommendations and future work Acknowledgment References 
work_yua7i7nk7jhmvbqozyerpfcdzm	----	doi:10.1016/j.eswa.2007.03.012 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 2290–2297 Expert Systems with Applications Structure clustering for Chinese patent documents Su-Hsien Huang a,d,*, Hao-Ren Ke b, Wei-Pang Yang c a Institute of Computer Science and Engineering, National Chiao Tung University, 1001 Ta Hsueh Road, Hsinchu, Taiwan, ROC b Library and Institute of Information Management, National Chiao Tung University, 1001 Ta Hsueh Road, Hsinchu, Taiwan, ROC c Department of Information Management, National Don Hwa University, 1, Section 2, Da Hsueh Road, Shou-Feng, Hualien, Taiwan, ROC d Department of Information Management, Minghsin University of Science and Technology, 1, Hsin Hsin Road, Hsin Feng, Hsinchu, Taiwan, ROC Abstract This paper aims to cluster Chinese patent documents with the structures. Both the explicit and implicit structures are analyzed to represent by the proposed structure expression. Accordingly, an unsupervised clustering algorithm called structured self-organizing map (SOM) is adopted to cluster Chinese patent documents with both similar content and structure. Structured SOM clusters the similar content of each sub-part structure, and then propagates the similarity to upper level ones. Experimental result showed the maps size and number of patents are proportional to the computing time, which implies the width and depth of structure affects the performance of structured SOM. Structured clustering of patents is helpful in many applications. In the lawsuit of copyright, companies are easy to find claim conflict in the existent patents to contradict the accusation. Moreover, decision-maker of a company can be advised to avoid hot- spot aspects of patents, which can save a lot of R&D effort. � 2007 Elsevier Ltd. All rights reserved. Keywords: Structure clustering; Chinese patent; Structure expression; Metadata 1. Introduction 1.1. Background Digital libraries provide various integrated services to coordinate information over Internet. However, the dis- tributed and heterogeneous data complicate the integration when it involves several issues, such as format, content, semantics, etc. The discrepancy and redundancy confuses users in readability. To provide a unified view, data cluster- ing with an overview by integrating similar data has received considerable attention in recent researches. Conventional data clustering adopts feature vector to represent data. It lacks some aspects of consideration to identify similarity. For example, chemical compounds with the same molecules but different structures are not the same. Moreover, two trademarks with the same compo- 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.03.012 * Corresponding author. E-mail address: sshuang@cis.nctu.edu.tw (S.-H. Huang). nents but different placements cannot be identified similar, too. These two examples account for the structure is a key-factor to influence the clustering in particular domains. Although conventional structure clustering focuses on visual pattern query, chemical structure identification and syntactic parsing tree, rare mentioned in documents. This is owing to the structure in documents is hard to obtain. Interestingly, the development in rhetorical structure the- ory (RST) facilitates the extraction of structure. RST was proposed in the 1980s and has been successfully applied to documents summarization, automatic layout, and so on. RST categorizes the characteristics of phrases and con- structs a closure tree to represent structure. Therefore, applying RST to represent documents structure becomes possible and practical. Several kinds of documents, like patent and electronic thesis, contain both structure and content information. To cluster these documents, only the same content in the same structure can be identified similar. The tort of copy- right is a good illustration to require conflict in both claim mailto:sshuang@cis.nctu.edu.tw S.-H. Huang et al. / Expert Systems with Applications 34 (2008) 2290–2297 2291 and claim structure. The structure of documents can be categorized into two types. Explicit structure represents the existent attributes of documents, like ‘‘subject’’, ‘‘abstract’’, ‘‘publisher’’ and ‘‘classification’’ in patent doc- uments. On the other hands, the structure analyzed from content is implicit structure. The implicit structure can be extracted by analyzing the writing style and the well- established convention. For example, in the claim field of patents, ‘‘the . . . of claim 1’’, ‘‘comprising’’, ‘‘wherein’’, ‘‘having’’, ‘‘consist of’’, etc. construct the implicit structure. With both explicit and implicit structures, document struc- ture can be constructed and applied in structure clustering. This study tries to apply Chinese patent documents in structure clustering. Patent structure has been mentioned in recent literatures. Shinmori, Okumura, Marukawa, and Iwayama (2003) provides a rhetorical method to cate- gorize Japanese patent structure into three styles – process sequence style, element enumeration style and Jepson-like style. Each claim in the patent is given a weight to evaluate the relationships among different claims. Fujii, Iwayama, and Kando (2006) extends Shinmori’s work and adds delimiter to punctuate claim into components. Mase, Matsubayashi, Ogawa, Iwayama, and Oshio (2004) analyze the patent structure into premise, description and target parts. Keywords in different part can be re-weighted to enhance query precision (Iwayama, Fujii, Kando, & Marukawa, 2006). Clustering in patent is a real-time application. Real-time means the time constraint of clustering algorithm is tight when users require the result on-line. Unfortunately, struc- ture clustering is more complicate than conventional one, where the computing time is subject to the structure. Hence, the first requirement of structure clustering is effi- ciency. Additionally, real-time also implies there is no training data in advance. Therefore, unsupervised algo- rithm is suggested when there is unnecessary to use training data in the process. Unsupervised clustering integrates sim- ilar data in definite circles, and adjusts the parameters to obtain result. Kohonen proposes an unsupervised neural network self-organizing map (SOM) and receives excellent performance (Kohonen, 1998). SOM provides unsuper- vised neural network clustering and maps high dimension data (usually two) into a low-dimension map. In SOM, clo- ser nodes in the map imply shorter distance in real data. SOM applies in many domains, like bio-structure cluster- ing, graph structure clustering and audio-pattern cluster- ing, etc. The multi-dimension representation of SOM facilitates the readability which is suitable for patent clustering. 1.2. Literatures of early studies It is noteworthy that conventional SOM cannot deal with structure data. In 1998, a general framework is pre- sented to deal with structure by neural network (Frasconi, Gori, & Sperduti, 1998). This framework constructs direc- ted ordered acyclic graphs (DOAG) for structure data and recurrently proceeds data by following DOAG sequence. This framework further motivates the upcoming researches. SOMSD (Sperduti, 2001) applies self-organiz- ing map (SOM) to manipulate structured data. Each neu- ron in SOMSD is equipped with two weights: one for previous structure and one for two-time previous structure. Structure similarity is estimated by concatenating current similarity and these two structure similarity. RSOM (Voegtlin, 2002) recursively calculates SOM by adding cur- rent similarity and previous structure. Hammer, Micheli, and Sperduti (2002) propose a general framework to further extend SOMSD. Vesanto and Alhoniemi (2000) cluster each sub-structure part in single SOM and use hier- archical k-means cluster to investigate similar structure. Smith and Ng (2003) and Roussinov and Chen (2001) derive user navigation patterns as structure to cluster web page navigation pattern. Similar navigation patterns (trea- ted as structure graph) are clustered together by SOM. Xu, Chang, and Paplinski (2005) use on-line expansion to enlarge SOM size into multiple layers. Other of structure clustering can be found in literatures (Hagenbuchner, Sperduti, & Tsoi, 2004; Hammer & Jain, 2004; Rossi, Conan-Guez, & Golli, 2004; Sperduti & Starita, 1997; Strickert & Hammer, 2003). 1.3. Purpose of this study This study is intended to apply Chinese patent docu- ments in structure clustering. A patent structure is analyzed first and provides structure expression to represent it. Hav- ing the structure established, self-organizing map (SOM) is adopted. Several modification of SOM is undertaken to process structure, including input, output and training algorithm. In brief, three major objectives involved in this study are: (1) Construct expression model to represent the patent structure. (2) Develop structure clustering algorithm to process structure patent. (3) Evaluate the structure clustering algorithm. This paper is organized as follows. Chapter 2 analyzes the structure of Chinese patent documents. Chapter 3 pro- poses structured SOM to cluster patents. Chapter 4 imple- ments structured SOM and a series of experiments are conducted in Chapter 5. Chapter 6 draws the conclusions. 2. Structure analysis As shown in Fig. 2, the preprocessing of structure clus- tering is to analyze the structure. The received documents are first represented by structure expression, and then determine the input sequence and maximum branch num- ber (MBN). To formally describe the structure, two formal constructors are defined. Given a Structure S, S can be represented by the following two constructors: 2292 S.-H. Huang et al. / Expert Systems with Applications 34 (2008) 2290–2297 (1) Tuple constructor (TC): A tuple constructor TC of Structure S is an ordered list constructed by the union of single-value attributes. For example, the set of attributes ci, c1 Æ occurence = 1, c2 Æ occurence = 1, . . .,cn Æ occurence = 1, are represented as TCs = {c1, c2,. . .,cn}s. The subscript s is the name of TC. (2) Set constructor (SC): A set constructor SC of Struc- ture S is a multi-value type with the same occurrence larger than one. For example, the set of attributes ci, c1 Æ occurence = i, c2 Æ occurence = i,. . .,cn Æ occu- rence = i, are represented as hc1; c2; . . . ; cni i s. The sub- script s represents the name of SC and i represents the maximum occurrence. 2.1. Explicit structure A structure document contains two types of structure. The first one is explicit structure, which is obtained from the schema of structure documents (like data metadata). Explicit structure is applicable in documents with regular schema. To take a simple example of electronic thesis, the attributes ‘‘subject’’, ‘‘abstract’’, ‘‘chapter’’, ‘‘section’’, and ‘‘paragraph’’ contain explicit structure. The structure expressions above example can be expressed as follows: Subject; Abstract; Contenth in1paragraph D En2 section � �n3 chapter ( ) Book where n1, n2 and n3 stand for the maximum occurrence of the attributes. Applying explicit structure in structure clustering is meaningful when the documents contain regular schema and massive content in each attribute. For example, elec- tronic thesis and XML-based news are suitable for explicit Fig. 1. Structure of structure clustering. In this study, Chinese patent docu- ments contain most information in ‘‘Claim’’ field. There- fore, implicit structure is introduced in the next section. 2.2. Implicit structure The structure analyzed from content is implicit structure. The implicit structure can be extracted by analyzing the writing style and the well-established convention. In the ‘‘Claim’’ field of Chinese patent documents, two types are categorized: • Composition style: As in ‘‘ ’’ (comprising), ‘‘ ’’ (includes), the set of element is described. These key- words are used in method to imply the composition of the inventions. This type of style represents a set of ele- ments contains in the claim structure. • Pre-condition style: As in ‘‘ ’’ (as claimed in . . .), ‘‘ ’’ (wherein), these descriptions imply the state- ments has followed a list of composition. This type of style represents a list of compositions construct the claim structure. Comparing with other-linguistic patents, there are sub- stantial difficulties to obtain several relationships in Chi- nese patent. For example, in Shinmori’s method, the process sequence style (means the sequence of processes, like ‘‘does’’, ‘‘and does’’) is ambiguous in Chinese gram- mar. It’s always followed by ‘‘ ’’ where is the same meaning with the article ‘‘one’’. These two types of style can derive the implicit structure of patents. For the presence of these keywords, a hierarchy relationship can be constructed. For example, Fig. 1 illus- trates an example of Chinese patent and represents the implicit structure analyzed by these two styles. To give each Chinese patent. S.-H. Huang et al. / Expert Systems with Applications 34 (2008) 2290–2297 2293 segmented claim an ID, naming mechanism is defined for each claim as follows: Patent ID-Claim IDlevel1- � � �Claim IDleveli The structure can be formally described by the structure expression, which can be more comprehensive for structure clustering. The composition style means the occurrence happens in the following statements. By following the structure expression in previous section, it can be expressed as hii, where i stands for the structure level. The pre-condi- tion style implies an attribute contains in the following statement. By following the structure expression, it can be expressed as {}i, where i stands for the structure level. For the example of Fig. 1, the structure expression can be expressed as follows: fClaima;fClaimbglevel2; Claimcglevel1 � �3 patent 2.3. Determine input sequence After the structure is analyzed, the input of each sub- part structure is determined by following directed acyclic graph (DAG). A depth-first search is adapted in the naviga- tion of structure tree. For the formal expression of both explicit and implicit structure, the input of structure follows two principles: (1) By order of the attributes in tuple constructor. (2) Iteratively expand set constructor to i times. For example, the input sequence of explicit structure Subject; Abstract; Contenth in1paragraph D En2 section � �n3 chapter ( ) Book Fig. 2. Structu is Subject ! Abstract ! Chapter1 ! Section1 ! Paragraph1 !���! Paragraphn3 ! Section2 !���! Paragraphn2 !���! Chaptern1 ! Sectionn2 ! Paragraphn1 For the example of Fig. 1, the input sequence to structure clustering is: P ! 1 ! 6 ! 11 ! 2 ! 3 ! 5 ! 4 ! 7 ! 8 ! 10 ! 9 ! 12 ! 13 ! 15 ! 14 2.4. Determine maximum branch number The next step is to determine the maximum branch num- ber (MBN) of tuple and set constructor. The maximum branch number (MBN) represents the maximum number of branches in patent structure. In the structure expression of both explicit and implicit, MBN is the maximum num- ber of TC attributes and SC occurrence. For example in Fig. 1, the MBN is 3 when the largest branch and occur- rence is less than 3. 3. Structured SOM Self-organizing map (SOM) (Kohonen, 1998) provides unsupervised neural network clustering and maps high- dimension data into a low-dimension map (usually two). There are five steps to train SOM (in Fig. 2): (1) Initialize weight vectors of output map as the same number features with input document vector. (2) Present input documents in order. red SOM. 2294 S.-H. Huang et al. / Expert Systems with Applications 34 (2008) 2290–2297 (3) Compute the distance between the input document and all nodes in the map and select the closest node as the winner. (4) Update the weights of the winner node and its neighbors. (5) Repeat steps (3)–(4) to other documents and iterate all inputs until convergence. Label the regions of the final map to represent the clustering result. SOM with structure receives considerable attention in recent years (Rossi et al., 2004; Strickert & Hammer, 2003; Voegtlin, 2002). The structured SOM in this study refers to Hagenbuchner’s self-organizing map clustering algorithm in structured data (Hagenbuchner et al., 2004). Hagenbuchner’s approach applied structured SOM on the images identification. The main concept is to calculate each sub-structure respectively, and concatenate the result to the upper structure. Notably, the images in Hagenbuch- ner’s method have simple structure and regular compo- nents. Experiment result has shown that structured SOM can cluster similar images together and distinguish the structure difference in the map. This paper applies struc- tured SOM in Chinese patents and tries to identify similar patent with different structure. The process of structured SOM is described in Fig. 2. 3.1. Training for explicit structure Applying SOM in Chinese patent documents requires modification of input/output vectors and algorithm. The explicit structure can be represented by structure expres- sion shown in previous section. The next step is to deter- mine the maximum branch number (MBN) of tuple and set constructor in explicit structure. The shortage of nodes less than MBN requires additional Null nodes. For the sim- plicity, assume the MBN = 3 in the example of explicit structure (n1 = 3, n2 = 3, n1 = 1), the structure expression can be rewrite as: Subject;Abstract; Content;NULL;NULLh in1paragraph D En2 section ; �� NULL;NULLin1chapter � Book The output nodes of SOM also need to be modified as the same number fields as input vector. For example, the output node of structured SOM in previous example is d x;y ¼ðV d x;y ;ðx1; y1Þ;ðx1; y1Þ;ðx1; y1ÞÞ The structure expression is assigned into SOM input vector by directed acyclic graph (DAG) sequence, for example, given a three node input vector, the input vector are rewrite as following: d book ¼ðV book;ðxsubject;ysubjectÞ;ðxabstract;yabstractÞ;ðxchapter;ychapterÞÞ Additionally, the distance calculation is updated as following: d ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðV Node0 � V bookÞ 2 q þjV Node1 � V subjectjþ jV Node2 � V abstractjþ jV Node3 � V chapterj where |VNodei � VNodej| represents the distance between these two vectors. The modification of distance formula means the cascaded calculation to all connected nodes. The adapting of structured SOM is updated as following: wd x;yðt þ 1Þ¼ wd x;yðtÞþ gðtÞ� jV Node1 � V subjectj wd x1;y1ðt þ 1Þ¼ wd x;yðtÞþ gðtÞ� jV Node2 � V abstractj wd x2;y2ðt þ 1Þ¼ wd x;yðtÞþ gðtÞ� jV Node3 � V chapterj 3.2. Training for implicit structure For the implicit structure, the dimension of the vector is also determined by the MBN in all input patents. Assume the MBN = 3, the representation in previous example is: ðV Patent; Claim1; Claim6; Claim11Þ ðV Node1; Claim2; Claim3; Claim5Þ ðV Node6; Claim7; Claim8; Claim10Þ ðV Node11; Claim12; Claim13; Claim15Þ The output nodes of SOM also need to be modified as the same number fields as input vector. For example, the output node of structured SOM in previous example is d x;y ¼ðV d x;y ;ðx1; y1Þ;ðx1; y1Þ;ðx1; y1ÞÞ Additionally, the distance calculation is updated as follow- ing formula: d ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðV Node0 � V PatentÞ 2 q þjV Node1 � V Claim1jþ jV Node2 � V Claim6jþ jV Node3 � V Claim11j where |VNodei � VNodej| represents the distance between these two vectors. The modification of distance formula means the cascaded calculation to all connected nodes. The adapting of structured SOM is also updated as following: wd x;yðt þ 1Þ¼ wd x;yðtÞþ gðtÞ� jV Node1 � V Claim1j wd x1;y1ðt þ 1Þ¼ wd x;yðtÞþ gðtÞ� jV Node2 � V Claim6j wd x2;y2ðt þ 1Þ¼ wd x;yðtÞþ gðtÞ� jV Node3 � V Claim11j 4. Implementation Fig. 3 displays the structured SOM system. The system contains six major parts: (1) SOM system: Choose standard SOM and structured SOM. (2) Parameter setting: Set the map dimension, learning rate, adaptation neighborhood and iteration to train. (3) Function selection: Select open file, close result, train- ing, save result, resize map and quit training. Parameter Setting Function Selection Map Information System Message SOM Map SOM System Fig. 3. Implementation of structured SOM. S.-H. Huang et al. / Expert Systems with Applications 34 (2008) 2290–2297 2295 (4) Map information: Display the containing patents in the SOM grid. (5) System message: Show the current iteration and pro- cessing message. (6) SOM map: Illustrate the clustering result. Fig. 4 demonstrates an example to apply structured SOM in Chinese patent documents. Initially, 10 Chinese patent documents are selected from database. These docu- ments are sent to standard SOM to obtain preliminary clustering. In the clustering, totally four clusters are labeled, e-commerce, binary-tree, identification and data- E-Commerse 1) Standard SOM 2) Structured SOM 3) Claim Clustering 3 Claim Conflict 21 Structure Clustering Fig. 4. Patent clustering. base. The interesting is paid in the five patents of e-com- merce cluster. There are three observations: (1) Commerce system is subjected. (2) Client–server architecture is addressed. (3) Encrypt system is applied. Hereafter, these patents are sent to structured SOM. In the structured SOM, these five patents are separated into three clusters, which represent they are different in struc- ture. For cluster 1, the patent primarily describes the con- Table 2 Time vs. documents (s) Document number 5 10 20 Standard SOM 8 29 118 Structured SOM 1727 2927 5713 (5 · 5 map, 50 iterations, MBN = 9). Table 3 Times vs. MBN (s) MBN 3 5 6 9 Structured SOM 7 10 53 295 (5 · 5 map, 10 documents, 5 iterations). 2296 S.-H. Huang et al. / Expert Systems with Applications 34 (2008) 2290–2297 struction of commerce system and is written in a very flat style. On the contrary, cluster 2 mainly focuses on encrypt system and is written in a very deep style. These two clus- ters are distinguished with cluster 3 not only the different subjects but the different patent structure. Notably, three documents marked as cluster 3 clusters together and might have high risk to conflict in the claims. Therefore, these claims are segmented by following the naming mechanism and cluster by SOM again to obtain conflict clusters. In the third step of Fig. 4, claim conflicts have observed in three areas: (1) Wire and wireless. (2) Client-side and server-side service. (3) Digital authorization and public key encrypt. These three patent conflicts provide good advisement for patent inventors to avoid involvement in this area. Struc- tured SOM is also helpful in many patent applications. In the lawsuit of copyright, companies are easy to find claim conflict in the existent patents, which is easy to con- tradict the accusation. Moreover, decision-maker of a com- pany can be advised to avoid hot-spot aspects of patents from structured SOM, which can save a lot of R&D effort. 5. Experiments A series of experiments are conducted to examine the performance of structured SOM. The data set comes from the claim attribute of Chinese patent documents, to exam- ine the implicit structure of patents. The experimental plat- form is Intel Celeron 1.5 GHz CPU, 512 MB RAM, Microsoft 2000 OS. The structured SOM was implemented by Java. In Table 1, the experiment was conducted to examine the performance in different map size. The parameters were set to 10 documents in 50 iterations with MBN = 9. The computing time of both standard and structured SOM is positive proportional to the map size. However, structured SOM is more time-consuming than standard one. In Table 2, the experiment was conducted to examine the performance in different document numbers. The parameters were set to 5 · 5 map in 50 iterations with MBN = 9. The computing times of both standard and structured SOM are positive proportional to the document numbers. The experiment corresponding to Table 1 implies the map size and documents are important factors to influ- ence the computing time. However, the time complexity between standard and structured SOM is extremely large, Table 1 Time vs. map size (s) Map size 2 · 2 3 · 3 5 · 5 10 · 10 Standard SOM 1 6 29 114 Structured SOM 227 355 2927 11,932 (10 documents, 50 iterations, MBN = 9). too. One explanation for the phenomenon is the document structure complicates the clustering. The deeper the struc- ture is, the longer the computing time. The experimental result of Table 3 further introduced another factor to influence computing time. With 10 docu- ments in five iteration of 5 · 5 map, structured SOM is nearly exponential proportional to the MBN. This is the most important factor account for the complexity of struc- tured SOM because both input vector and output map are modified to fit the dimension of MBN. These results lead to the conclusion that the width (MBN) and the depth of the structure are two key-factors for structured SOM performance. 6. Conclusions and future work Increasing Chinese patent documents require the clus- tering in the application. To cluster precisely, patents with similar content but different structure should be distin- guished. However, conventional clustering lacks of struc- ture consideration cannot identify similar patents with different structure, and rare research mentioned in text domain. In this paper, a clustering algorithm called struc- tured SOM is proposed to clusters structure patents by considering the content and structure information simulta- neously. Structured SOM requires the analysis of patent structure in advance. Two types of structure expression, explicit and implicit structure, are provided. Structured SOM modifies the input and output vectors for structure documents and iteratively training by adapting each sub- part of structure. An example to cluster Chinese patent documents has successfully implemented. Claim conflict appeared in the analysis provides advisements for patent inventors and decision-makers of business. In the lawsuit of copyright, companies are easy to find claim conflict in the existent pat- ents, which is easy to contradict the accusation. Moreover, decision-maker of a company can be advised to avoid hot- spot aspects of patents from structured SOM, which can save a lot of R&D effort. Experiments are conducted to examine the performance of structured SOM. Conclusions are given that both map S.-H. Huang et al. / Expert Systems with Applications 34 (2008) 2290–2297 2297 size and documents are positive proportional to the computing time. This implies two factors are important: the width (MBN) and depth of structure. MBN is nearly exponential drop-off the performance of structured SOM. In some practical application, prediction is adapted to esti- mate node distance to reduce computing time but sacrifice neglectful accuracy. Structure clustering is rarely mentioned in the text domain because the text with structure (like data metadata) has small amount of content in the attributes (explicit structure). With the development of digitalization, some structure documents, like electronic thesis and on-line news, can be applied in structured SOM. The future direc- tion for this study might be to discover more applications for structure clustering. It is also lots of work to improve the efficiency of structured SOM. Acknowledgements This research was supported by the Software Technol- ogy for Advanced Network Application project of Institute for Information Industry in 2004 and sponsored by MOEA, ROC. References Frasconi, P., Gori, M., & Sperduti, A. (1998). A general framework for adaptive processing of data structures. IEEE Transactions on Neural Networks, 9(5), 768–786. Fujii, A., Iwayama, A., & Kando, N. (2006). Test collections for patent retrieval and patent classification in the fifth NTCIR workshop. In Proceedings of the fifth international conference on language resources and evaluation (pp. 671–674). Hagenbuchner, M., Sperduti, A., & Tsoi, A. C. (2004). A self-organizing map for adaptive processing of structured data. IEEE Transactions on Neural Networks, 14(3), 491–505. Hammer, B., & Jain, B. J. (2004). Neural methods for non-standard data. In european symposium at artificial neural networks 2004 (pp. 281–292). Hammer, B., Micheli, A., & Sperduti, A. (2002). A general framework for unsupervised processing of structured data. In European symposium on artificial neural networks (ESANN’2002) (pp. 395–400). Iwayama, M., Fujii, A., Kando, N., & Marukawa, Y. (2006). Evaluating patent retrieval in the third NTCIR workshop. Information Processing and Management, 42(1), 207–221. Kohonen, T. (1998). The self-organizing map. Neurocomputing, 21, 1–6. Mase, H., Matsubayashi, T., Ogawa, Y., Iwayama, M., & Oshio, T. (2004). Two-stage patent retrieval method considering claim structure. In NTCIR workshop 4 meeting. Rossi, F., Conan-Guez, B., & Golli, A. E. (2004). Clustering functional data with the SOM algorithm. In Proceedings of european symposium on artificial neural networks 2004 (ESANN’04) (pp. 305–312). Roussinov, D. G., & Chen, H. (2001). Information navigation on the web clustering and summarizing query results. Information Processing and Management, 37, 789–816. Shinmori, A., Okumura, M., Marukawa, Y., & Iwayama, M. (2003). Patent claim processing for readability – Structure analysis and term explanation. In ACL-2003 workshop on patent corpus processing. Sapporo: Association for Computational Linguistics. Smith, A., & Ng, A. (2003). Web page clustering using a self-organizing map of user navigation patterns. Decision Support Systems, 35, 245–256. Sperduti, A. (2001). Neural networks for adaptive processing of structured data. Lecture Notes in Computer Science, 2130, 5–12. Sperduti, A., & Starita, A. (1997). Supervised neural networks for classification of structures. IEEE Transaction on Neural Networks, 8(3), 714–735. Strickert, M., & Hammer, B. (2003). Unsupervised recursive sequence processing. In Proceedings of european symposium on artificial neural networks 2003 (ESANN 2003) (pp. 27–32). Vesanto, J., & Alhoniemi, E. (2000). Clustering of the self-organizing map. IEEE Transactions on Neural Networks, 11(3), 586–600. Voegtlin, T. (2002). Recursive self-organizing maps. Neural Networks, 15, 979–991. Xu, P., Chang, C. H., & Paplinski, A. (2005). Self-organizing topological tree for online quantization and data clustering. IEEE Transactions on Systems, Man, and Cybernetics, 35(3), 515–526. Structure clustering for Chinese patent documents Introduction Background Literatures of early studies Purpose of this study Structure analysis Explicit structure Implicit structure Determine input sequence Determine maximum branch number Structured SOM Training for explicit structure Training for implicit structure Implementation Experiments Conclusions and future work Acknowledgements References 
work_yuerlbpcxfbnphesk7mnzuikrq	----	A web-oriented expert system for planning hurdles race training programmes ORIGINAL ARTICLE A web-oriented expert system for planning hurdles race training programmes Krzysztof Przednowek1 • Krzysztof Wiktorowicz2 • Tomasz Krzeszowski2 • Janusz Iskra3 Received: 22 March 2017 / Accepted: 21 May 2018 / Published online: 30 May 2018 � The Author(s) 2018 Abstract This paper presents a web-oriented expert system, named iHurdling, to predict results and generate training loads for 110 and 400 m hurdles races. The database contains 40 annual training programmes for the 110 m hurdles and 48 programmes for the 400 m hurdles. The predictive models include linear regressions in the form of ordinary least squares, ridge, LASSO, elastic net and nonlinear models in the form of a radial basis function neural network and fuzzy rule-based system. The leave-one-out cross-validation method is used to compare, and choose the best model. It shows that the proposed fuzzy-based model has the lowest validation error. The developed web application can support a coach in planning training programmes for hurdles races. It allows the athlete’s results to be predicted and can generate training loads for an athlete, selected from database. The application can be run on a computer or a mobile device. The system was implemented using the R programming language with the Shiny framework and additional packages. The limitations of the presented approach are related to the lack of consideration of an athlete’s physiological and psychological parameters, but the generated training programs might be used as a suggestion for the coach. Keywords Predictive models � Expert system � Hurdles race � R programming language 1 Introduction Expert systems have been widely implemented and examined by researchers. They mimic the decision-making abilities of a human expert, and they are designed to solve complex problems by reasoning. Expert system applica- tions include, among others, medicine [29, 53, 60], diag- nosis and control of power systems [26, 27], evaluation of journal grades [61], information systems investment eval- uation [19], transport management [31, 51], industry [14, 38] and sport [12, 15, 32, 33, 39]. Nowadays, in sports science various types of computer tools and methods play an important role. Competitors and coaches are looking for new solutions that can support their work. One aspect of such support can be the application of machine learning methods, which can be used to calculate performance results [13, 15, 43], identify sporting talent [35, 39, 48, 49] or support the training process [30, 40, 41, 45, 50, 52]. For example, in the paper [13], the authors use artificial neural networks to predict competitive performance in swimming. The neural models were cross-validated, and the results show that the modelling was very precise. The paper [43] describes the use of linear and nonlinear mul- tivariable models as tools to predict the results of 400 m hurdles races. All the models were constructed using the training data of 21 athletes from the Polish National Team. & Tomasz Krzeszowski tkrzeszo@prz.edu.pl Krzysztof Przednowek krzprz@ur.edu.pl Krzysztof Wiktorowicz kwiktor@prz.edu.pl Janusz Iskra j.iskra@awf.katowice.pl 1 Faculty of Physical Education, University of Rzeszow, ul. Towarnickiego 3, 35-959 Rzeszow, Poland 2 Faculty of Electrical and Computer Engineering, Rzeszow University of Technology, al. Powstanców Warszawy 12, 35-959 Rzeszow, Poland 3 Faculty of Physical Education and Physiotherapy, Opole University of Technology, ul. Prószkowska 76, 45-758 Opole, Poland 123 Neural Computing and Applications (2019) 31:7227–7243 https://doi.org/10.1007/s00521-018-3559-1(0123456789().,-volV)(0123456789().,-volV) http://orcid.org/0000-0002-2128-4116 http://orcid.org/0000-0001-8711-1659 http://orcid.org/0000-0001-7359-4637 http://crossmark.crossref.org/dialog/?doi=10.1007/s00521-018-3559-1&amp;domain=pdf http://crossmark.crossref.org/dialog/?doi=10.1007/s00521-018-3559-1&amp;domain=pdf https://doi.org/10.1007/s00521-018-3559-1 The best prediction results were obtained by the LASSO regression method. Gu et al. [15] proposed an expert sys- tem to predict National Hockey League (NHL) game out- come. The prediction accuracy of the system was 77:5%. Another paper [17] presents a review of data mining techniques that are used for prediction in various sports disciplines. Roczniok et al. [48] proposed using Kohonen’s neural networks for the recruitment process in competitive swimming. Experiments were conducted on a group of 140 young swimming contestants aged about 10. Another approach to identifying sporting talent was proposed by Rogulj et al. [49]. The authors have developed two methodological approaches to recognize an athlete’s mor- phological compatibility for various sports. In the paper [35], Maszczyk et al. determined the usefulness of neural models in optimizing recruitment processes. Statistical analyses were carried out on the measured results of javelin throwers using full take off. For the investigated group, the perceptron network with the 4–3–2–1 structure achieved the best predictive results. In the paper [50], Ryguła et al. proposed using an arti- ficial neural network (ANN) to model swimming perfor- mance in the 200 m individual medley and the 400 m front crawl events. The ANNs were also used to analyze tactics in team sports [41]. Another study was devoted to the use of ANNs to classify kick techniques [30]. The aim of that paper was to find out whether it is possible to distinguish two different kick techniques from a kick impact force profile. The paper [52] presents the application of a neural network to model swimming performance. The authors created highly realistic models of swimming performance prediction based on previously selected criteria that were related to the dependent variable. Experiments were con- ducted on 138 swimmers (65 males and 73 females) at national level. Despite the existing methods to predict and support training, there is lack of tools that could be used by coaches during the training process. Papić et al. [39] developed a fuzzy expert system for scouting and evaluating young sporting talent. A similar system is presented in [33], where the authors perform talent identification in soccer using a web-oriented expert system. From the review of literature, it can be seen that there is a need to create tools for supporting sports training. The main contribution of this paper is, therefore, to develop a web-oriented expert system, named iHurdling, to predict results and generate training loads in the 110 and 400 m hurdles. The system we have developed can support a coach in planning training programmes in hurdles races. The system uses linear regression models (OLS, ridge, LASSO, elastic net) and nonlinear models (RBF, fuzzy model, OLS with fuzzy correction). The main advantages of this system are an easy-to-use interface and compati- bility with different platforms which means that it can be run from a computer or a mobile device. 2 Training data The training data contain training plans carried out by hurdlers in the Polish National Team. One record contains the parameters of an athlete and the training programme carried out by this athlete during their annual training cycle. The models for result prediction (PR) and for gen- erating training loads (GT) were build using 21 variables (Table 1). For the PR models, the input variables x1�x5 represent the parameters of the athlete, the input variables x6�x20 represent the training loads and the output variable y represents the predicted result. For the GT models, x1�x6 represent the parameters of the athlete and y1�y15 repre- sent the training loads. The training programs were recor- ded according to the classification proposed in [22]. The classification consists of two areas of influence: energy (exercise) and information (related to the formation of technique). In the analyzed training loads, there are speed, endurance and strength as well as exercises that develop the technique of hurdles clearance. A similar classification of exercises can be found in another papers devoted to sprinters and hurdlers [2, 37]. The values of these loads are the sum of all loads of the same type realized during the annual training cycle. The results for the hurdles races were registered before and after the cycle. Both runs were car- ried out under simulated starting conditions of the 110 and 400 m hurdle race. In this study, the current result at the training distance was assumed as the indicator of perfor- mance level. As concluded in the paper [25], this result is strongly correlated with performance parameters and other motor skills tests used in hurdles races. For the 110 m hurdles, the training data contain 40 records. These records were collected from 18 highly trained athletes (mean result in 110 m hurdles: 14.02 s) aged 18–28. In 400 m hurdles, the 48 records from 21 athletes aged 19–27 were used. The hurdlers practising the 400 m had also a high sport level. (Mean result on 400 m hurdles was equal to 51.26 s.) 3 Mathematical models In this paper, we use the regression methods for building multi-input, single-output (MISO) and multi-input, multi- output (MIMO) models. The MISO models are used for the prediction of result, while the MIMO models are used in the generation of training loads. In the simplified descrip- tion that follows we assume that we have one output, since 7228 Neural Computing and Applications (2019) 31:7227–7243 123 a MIMO model will be represented as a set of MISO models. In our expert system, we use: – linear models in the form of ordinary least squares (OLS) [7], ridge regression (RIDGE) [18], least absolute shrinkage and selection operator (LASSO) [54] and elastic net (ENET) [62], – nonlinear models in the form of radial basis function network (RBF) [6] and fuzzy rule-based system (FRBS) [47]. 3.1 Linear models Consider a MISO model with p inputs (predictors) creating the vector x ¼ ½x1; x2; . . .; xp� and one output (response) y. The goal is to build the regression function y ¼ f ðxÞ ¼ Xp j¼1 xjwj ð1Þ based on a data set containing n observations in the form of pairs ðxi; yiÞ, where xi ¼ ½xi1; xi2; . . .; xip�, i ¼ 1; . . .; n. The element xij denotes the jth predictor in the ith observation, and yi is the response in the ith observation. The linear regression problem can be written as a matrix equation of the form y ¼ Xw ð2Þ where X ¼ x11 x12 . . . x1p x21 x22 . . . x2p .. . .. . .. . .. . xn1 xn2 . . . xnp 2 66664 3 77775 ð3Þ and w ¼ ½w1; w2; . . .; wp�T , y ¼ ½y1; y2; . . .; yn�T . Denoting by Jðw; �Þ a cost function, the problem of finding a linear model involves minimizing the function Jðw; �Þ, that is ŵ ¼ arg min w Jðw; �Þ ð4Þ where ŵ is the vector of the optimal parameter values. For the linear models, the cost functions have the form of JOLSðwÞ ¼ y � Xwk k22 ð5Þ Table 1 Description of variables used to construct the PR and GT modules for 110 and 400 m hurdles PR GT Description 110 m 400 m x xmin xmax x xmin xmax y x1 Result after training (s) 14.02 13.26 15.13 51.27 48.19 53.60 x1 x2 Age (years) 21.9 18.0 28.0 22.3 19.0 27.0 x2 x3 Body height (cm) 187.3 181.0 195.0 185.0 177.0 192.0 x3 x4 Body mass (kg) 77.8 71.0 83.0 74.3 69.0 82.0 x4 x5 Body mass index 22.1 20.3 23.5 21.2 19.7 24.4 x5 x6 Result before training (s) 14.33 13.34 15.40 51.91 48.70 54.70 x6 y1 Maximal and technical speed exercises (m) 12,513 5800 17,970 9428 2910 18,920 x7 y2 Technical and speed exercises (m) 5925 2470 10,200 4253 240 9450 x8 y3 Speed and specific hurdle endurance exercises (m) 11,961 3150 20,400 25,342 6400 101,450 x9 y4 Pace runs exercises (m) 64,087 25,780 100,300 163,796 88,000 393,800 x10 y5 Aerobic endurance exercises (m) 328,631 80,600 550,000 363,257 151,000 692,500 x11 y6 Strength endurance exercises (m) 20,638 1850 46,595 41,069 1750 169,265 x12 y7 Strength of lower limbs exercises (kg) 291,119 96,400 658,600 224,099 96,900 504,540 x13 y8 Trunk strength exercises (amount) 38,442 5240 145,000 46,438 6100 233,680 x14 y9 Upper body strength exercises (kg) 3352 1630 4850 3305 760.0 29,610 x15 y10 Explosive strength of lower limbs exercises (amount) 1244 0 2214 823 282 2138 x16 y11 Explosive strength of upper limbs exercises (amount) 656 213 1850 443 60 1360 x17 y12 Technical exercises – walking pace (min) 456 130 1110 424 45 816 x18 y13 Technical exercises – running pace (min) 574 195 1450 518 150 1500 x19 y14 Runs over hurdles exercises (amount) 778 362 1317 416 121 775 x20 y15 Hurdle runs in varied rhythm exercises (amount) 1077 320 1850 857 36 1680 Neural Computing and Applications (2019) 31:7227–7243 7229 123 JRIDGEðw; kÞ ¼ y � Xwk k22 þ k wk k 2 2 ð6Þ JLASSOðw; kÞ ¼ y � Xwk k22 þ k wk k1 ð7Þ JENETðw; k1; k2Þ ¼ y � Xwk k22 þ k1 wk k1 þ k2 wk k 2 2 ð8Þ where k, k1 and k2 are non-negative regularization parameters. The norms jj � jj2 and jj � jj1 denote the Eucli- dean and the Manhattan norms, respectively. The RIDGE, LASSO and ENET regressions are regularized which means that they can be used when the problem is ill-con- ditioned. The detailed description of the linear models can be found, for example, in [58]. 3.2 Choosing the best model All models were tested using cross-validation method. This is a method of evaluating the generalization ability (pre- diction for new data, not involved in modelling) of the model being created. In cross-validation, data are divided into two subsets: a training set and a testing (validation) set. In this study, due to the small amount of data (n ¼ 40 for 110 m and n ¼ 48 for 400 m), LOOCV (leave-one-out cross-validation) was used [3]. The idea of this method is to extract from the set of data n learning subsets. Each subset is created by removing only one pair from the data set, which becomes a test pair. Then, for each resulting subset, the model is constructed that is evaluated by determining the error for the remaining test pair. The predictive ability of a model is expressed by the root of the mean square error of cross-validation (RMSECV) calcu- lated as RMSECV ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 n Xn i¼1 yi � ŷ�ið Þ 2 s ð9Þ where ŷ�i is the output of a model obtained after removing the pair ðxi; yiÞ from the data set. 3.3 Nonlinear models 3.3.1 RBF models An RBF network is a feed-forward network that typically consists of three layers: an input layer, a hidden layer and an output layer. The input layer is composed of nodes that receive input signal x, and there is one node for each predictor variable. The hidden layer is composed of nodes with radially symmetric activation functions. The hidden node measures the distance between the input vector x and the centre ck of its radial function: ukðxÞ ¼ uk x � ckk kð Þ ð10Þ The norm jj � jj is usually taken as the Euclidean distance, and uðxÞ is typically taken to be the Gaussian function. The output layer is composed of a node that receives the outputs of nodes in the hidden layer. This node calculates the output of the network as a linear combination of non- linear functions of the form y ¼ Xm k¼1 ukðxÞwk ð11Þ where m is the number of nodes in the hidden layer, ukðxÞ is a basis function and wk is the weight of the kth neuron in the output node. The training of the RBF network involves: the number of hidden neurons, the parameters of radial functions in the hidden layer and the weights in the output layer. 3.3.2 Fuzzy models In this paper, we propose two approaches to use the FRBS [47] in regression problems. In the first approach, the fuzzy model is build similarly to the RBF model, that is, it is learned from the original data. (This model is called FUZZY.) In the second approach, the FRBS is used for the nonlinear correction of the OLS model. (This model is called F-OLS.) The idea is to change the output of a linear model by adding a nonlinear correction term, in such a way that the predictive error is reduced (Fig. 1). First, we build the OLS model and remember its cross-validation errors, and next we build a fuzzy model that ‘‘learns’’ these errors. The design procedure for building F-OLS models is listed below. Step 1. Cross-validation of the OLS model y ¼ fOLSðxÞ for the data ðxi; yiÞ. In the ith step of cross- validation, the error has the form ei ¼ yi � y�i ð12Þ where y�i ¼ fOLSðx�iÞ. Fig. 1 The idea of calculating the output of the F-OLS model. The variable y ¼ fOLSðxÞ is the output of the ordinary least squares estimator, and d ¼ fcðxÞ is the output of the fuzzy nonlinear corrector 7230 Neural Computing and Applications (2019) 31:7227–7243 123 Step 2. Constructing the fuzzy (nonlinear) corrector d ¼ fcðxÞ ð13Þ for the data ðxi; eiÞ. This corrector predicts the errors obtained in Step 1. The best fuzzy model can be chosen on the basis of cross-validation conducted for different number of fuzzy sets. Step 3. Cross-validation of the OLS model with the corrected error in the form enewi ¼ yi � ðy�i þ diÞ ð14Þ where y�i ¼ fOLSðx�iÞ and di ¼ fcðxiÞ. Step 4. The predicted output of the F-OLS model is determined as yF�OLS ¼ fOLSðxÞ þ fcðxÞ ð15Þ where fcðxÞ is the function of the fuzzy corrector chosen in Step 3 (Fig. 1). 4 Expert system modules The expert system consists of two modules, the prediction of result (PR) and the generation of training loads (GT) for the 110 and 400 m distances. The regression models for both modules were calculated in R language [46]. The functions with arguments used to generate the models are shown in Table 2, and they are described below. The function lm was used to calculate the OLS, and the ridge regressions were calculated using the function lm.ridge from the ‘‘MASS’’ package [55] (with k [ 0 in 6). The LASSO and the elastic net regressions were obtained with the function enet included in the ‘‘elastic- net’’ package [63]. This function has two parameters ðk; sÞ, where k � 0 denotes k2 in the formula (8) and s 2 ½0; 1� is a fraction of the norm L1. The pair ðk; sÞ is used instead of the pair ðk1; k2Þ in the formula (8) because the elastic net regression can be treated as the LASSO regression for an augmented data set [62]. Taking k ¼ 0 we get the LASSO regression with one parameter s for the original data. The ENET models were selected by searching the parameters k and s. This study uses artificial neural networks in the form of the radial basis function (RBF). The training data were scaled before the RBF training, and the results of the predictions were unscaled. All the analyzed networks have one hidden layer. For the implementation of neural net- works, the function RSNNS::rbf was used [6]. The optimal neural model was determined by searching a number of hidden neurons in the range from 2 to 10. The fuzzy models were calculated using the function frbs.learn from the ‘‘frbs’’ package [47]. The learning method was the Wang–Mendel (W–M) algorithm [56]. This algorithm generates fuzzy rules from input–output data pairs. The input space is divided into fuzzy subspaces, Table 2 R functions for models training Model Function Arguments OLS lm formula ¼ y � � data.train RIDGE lm.ridge formula ¼ y � � data.train lambda 2 ½0; 1Þ LASSO enet data.x, data.y lambda = 0, normalize = TRUE intercept = TRUE predict.enet model newx s 2 ½0; 1� ENET enet data.x, data.y lambda 2 ½0; 1Þ, normalize = TRUE intercept = TRUE predict.enet model newx s 2 ½0; 1� RBF rbf data.x, data.y size 2 ½1; 10� maxit = 1000 linOut = TRUE Fuzzy frbs.learn data.train range.data method.type = ‘‘WM’’ control = list( num.labels = l, type.mf = ‘‘GAUSSIAN’’, type.defuz = ‘‘WAM’’, type.tnorm = ‘‘MIN’’, type.implication.func = ‘‘MIN’’) The function predict.enet is used for selecting the parameter s for LASSO and ENET models Neural Computing and Applications (2019) 31:7227–7243 7231 123 and fuzzy rules are extracted for each subspace. The W–M method is a one-pass procedure and does not need time- consuming training. In the fuzzy model, the Gaussian membership functions are used, the t-norm is ‘‘minimum’’, the defuzzification is ‘‘weighted average method’’, and the implication is ‘‘minimum’’. The number of fuzzy sets l was determined by calculating cross-validation errors as described in Sect. 3.3.2. 4.1 Models for result prediction The cross-validation errors RMSECV and parameters of the models for the PR module are presented in Table 3. The parameters were chosen on the basis of the plots shown in Figs. 2 and 3. In the case of the ridge regression, the reg- ularization parameter k 2 ½0; 40� was considered with the step 0.1 for both distances. In the case of the LASSO regression, the parameter s 2 ½0; 1� was considered with the step 0.01. For the ENET regression, the following param- eters were chosen: k 2 ½0:1; 0:25� with the step 0.008 and s 2 ½0:4; 0:8� with the step 0.021 for the distance of 110 m, and k 2 ½0; 0:06� with the step 0.0032 and s 2 ½0:3; 0:5� with the step 0.01 for the distance of 400 m. The RBF model was analyzed for the number of neurons in the hidden layer m 2 f2; 3; . . .; 10g, and the fuzzy models were analyzed for the number of fuzzy sets l 2 f2; 3; . . .; 13g. Based on the conducted analysis, the best models (models with the smallest cross-validation error) were selected (Table 3). It can be seen that for both the 110 and the 400 m distances, the lowest error was obtained by the F-OLS regression. The best F-OLS models have eight fuzzy sets for 110 m and nine sets for 400 m. The largest error for the 110 m was obtained by the OLS regression and by the FUZZY model for the 400 m. 4.2 Models for generation of training loads For the GT module, each output of the model (y1-y15) was considered and analyzed in a similar way as for the result prediction module. The errors RMSECV for the GT module are presented in Table 4, while the parameters of the models are presented in Table 5. The models in the GT module were cross-validated similarly to the PR module. For example, for the output y14 the FUZZY model has the largest errors (200.1 for 110 m and 132.9 for 400 m), and the F-OLS model has the smallest errors (33.18 for 110 m and 105.1 for 400 m). From Table 4, it can be observed that the smallest RMSECV for all outputs has the F-OLS model. 5 Graphical user interface The graphical user interface was implemented in R lan- guage using the shiny [11], shinyjs [4], shi- nythemes [10], shinydashboard [9] and rmarkdown [1] libraries. This interface is a web-oriented application and therefore requires only a web browser and an Internet connection to be used. The current version of the developed system is available on https://hurdles.shi nyapps.io/ihurdling. The application shown in Fig. 4 con- sists of three panels labelled ‘‘Result prediction’’, ‘‘Generation of training loads’’ and ‘‘Athletes’ database’’. On the left side of window is a sidebar menu with links to each panel. The radio button in this sidebar is used to select the PR or GT module. Moreover, the user can choose one of the developed regression models and generate reports. The footer contains the information about the application and the authors. Table 3 Errors and parameters for the PR module for 110 and 400 m hurdles OLS RIDGE LASSO ENET RBF FUZZY F-OLS 110 m RMSECV 0.3807 0.2276 0.2397 0.1996 0.1985 0.2572 0.0851 Param. — k ¼ 16:1 k ¼ 0 s ¼ 0:04 k ¼ 0:16 s ¼ 0:56 m ¼ 8 l ¼ 4 l ¼ 8 400 m RMSECV 0.6959 0.5743 0.4463 0.4308 0.6953 0.9288 0.1533 Param. — k ¼ 11:6 k ¼ 0 s ¼ 0:28 k ¼ 0:01 s ¼ 0:36 m ¼ 6 l ¼ 3 l ¼ 9 The meaning of the parameters is as follows: k and s are tuning parameters for regularized models, m is the number of hidden neurons in the RBF network, l is the number of fuzzy sets in fuzzy models 7232 Neural Computing and Applications (2019) 31:7227–7243 123 https://hurdles.shinyapps.io/ihurdling https://hurdles.shinyapps.io/ihurdling 5.1 Panel for result prediction The ‘‘Result prediction’’ panel is used for entering data and for result prediction (Fig. 4). The input variables are grouped into five boxes: ‘‘Athlete’s parameters’’, ‘‘Training loads—endurance’’, ‘‘Training loads—technique and rhythm’’, ‘‘Training loads—strength’’ and ‘‘Training loads—speed’’. The value of each input can be modified using appropriately scaled sliders. For example, the box ‘‘Training loads—endurance’’ presented in Fig. 5 has four sliders for changing endurance training loads. Each slider has a range determined on the basis of the minimum and maximum values in the database (Table 1) and depends on the distance of the hurdles race. For instance, the slider ‘‘Pace runs’’ for the 110 m hurdles has a range from 25,000 to 101,000 m with the step equal to one metre. In the last box labelled ‘‘Results’’ two textOutput fields display the current and predicted results. Prediction of the result is performed automatically after changing the position of any slider. Moreover, the result depends on the radio button that selects the method in the sidebar menu. In this way, the user can modify the training loads and 0 10 20 30 40 0.2 0.25 0.3 0.35 0.4 λ R M S E C V 110m, RIDGE regression 0 10 20 30 40 0.58 0.6 0.62 0.64 0.66 0.68 0.7 λ R M S E C V 400m, RIDGE regression 0 0.2 0.4 0.6 0.8 1 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 s R M S E C V 110m, LASSO regression 0 0.2 0.4 0.6 0.8 1 0.4 0.6 0.8 1 1.2 1.4 s R M S E C V 400m, LASSO regression 0.1 0.15 0.2 0.25 0.4 0.5 0.6 0.7 0.8 0.2 0.25 0.3 λ 110m, ENET regression s R M S E C V 0 0.02 0.04 0.06 0.3 0.35 0.4 0.45 0.5 0.4 0.5 0.6 0.7 λ 400m, ENET regression s R M S E C V 0.2 0.205 0.21 0.215 0.22 0.225 0.23 0.235 0.24 0.245 0.44 0.46 0.48 0.5 0.52 0.54 0.56 Fig. 2 Cross-validation errors for linear models (RIDGE, LASSO, ENET) in result prediction Neural Computing and Applications (2019) 31:7227–7243 7233 123 observe the changes that occur in the expected result. Generating a report from result prediction creates a .pdf file, which contains the values from all sliders and the predicted result. 5.2 Panel for generation of training loads Another system panel is the generation of training loads for both hurdles distances (Figs. 6, 7). This module consists of two boxes: ‘‘Athletes’ parameters’’ and ‘‘Generated training—annual cycle’’. The first box is used to enter the athlete’s data, i.e. age, body height, body weight and his current result. This box also includes an option to choose the training generation mode. The user can choose the option of one training generation or the option to generate the training loads for a longer period of his career. The selection of the first option will cause a slider with the expected result to appear under the slider with the current result. If the ‘‘career’’ option is selected, these sliders are not available. The ‘‘career’’ option makes it possible to 2 4 6 8 10 0.2 0.25 0.3 0.35 0.4 m R M S E C V 110m, RBF regression 2 4 6 8 10 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 m R M S E C V 400m, RBF regression 2 4 6 8 10 12 14 0.24 0.26 0.28 0.3 0.32 0.34 0.36 l R M S E C V 110m, FUZZY regression 2 4 6 8 10 12 14 0.9 1 1.1 1.2 1.3 l R M S E C V 400m, FUZZY regression 2 4 6 8 10 12 14 0.45 0.5 0.55 0.6 0.65 0.7 0.75 l R M S E C V fo r f c 110m, fuzzy corrector 2 4 6 8 10 12 14 0.8 0.9 1 1.1 1.2 1.3 l R M S E C V fo r f c 400m, fuzzy corrector Fig. 3 Cross-validation errors for nonlinear models (RBF, FUZZY, fc) in result prediction 7234 Neural Computing and Applications (2019) 31:7227–7243 123 generate six training programmes which are consecutive and improve the result by 0.25 s each year (from 15.00 to 13.50 s) for 110 m hurdles and by 1 s each year (from 53.00 to 48.00 s) for 400 m hurdles. The contents of the box ‘‘Generated training—annual cycle’’ change dynamically, depending on the mode. The ‘‘one training’’ option will generate a list of training loads with suggested values (Fig. 6). In addition, a graph is generated, in which the values of training loads, expressed as a percentage of the maximum value of the given output, are presented. The second option is ‘‘career’’; its selection generates a table containing six annual training plans and 15 graphs showing the loads over the athlete’s entire career (Fig. 7). The ‘‘career’’ is an additional option that allows us to generate training loads for six consecutive years. In this option, the starting result is always constant and is 14.75 s for 100 m and 53 s for 400 m, respectively. Results are generated in the form of a table where each row represents the annual training and in the form of graphs where the x- axis is the expected result and the y-axis is the value of the training load. Career graphs allow observations of changes in individual loads in terms of a 6-year career. The coach can observe which load needs to be increased, which decreased and which should stay at the same level. Generating a report from the ‘‘Generated training’’ panel creates a .pdf file, which contains values from the ‘‘Ath- Table 4 Errors for the GT module for 110 and 400 m hurdles Output OLS RIDGE LASSO ENET RBF FUZZY F-OLS 110 m y1 2550 2429 2304 2304 2219 2572 1829 y2 1700 1632 1617 1606 1528 1675 244.0 y3 4351 3985 3933 3814 3843 3629 2630 y4 24,040 21,170 21,270 21,120 20,310 20,210 6254 y5 112,000 110,900 106,600 106,600 80,681 88,870 70,780 y6 12,570 11,240 11,170 11,170 10,693 9644 2546 y7 152,100 129,600 140,200 129,600 131,123 112,200 21,270 y8 30,190 26,730 27,970 26,360 25,366 22,570 3410 y9 654.1 610.9 606.1 606.1 632.8 622.3 519.2 y10 579.7 484.0 482.4 482.4 398.3 327.9 155.2 y11 267.4 244.7 247.5 239.7 191.5 236.7 143.1 y12 302.1 263.1 263.4 259.4 248.9 266.8 98.17 y13 385.5 322.8 320.5 320.5 310.7 256.1 39.93 y14 190.9 179.7 188.8 186.3 167.7 200.1 33.18 y15 436.0 401.5 400.7 391.2 399.1 434.0 205.6 400 m y1 4477 4189 4110 4109 3038 2511 597.7 y2 1864 1778 1755 1754 1422 1460 855.9 y3 15,120 14,270 14,190 14,190 14,308 16,670 3473 y4 62,580 60,770 60,040 60,040 63,378 52,860 11,110 y5 107,100 98,490 98,120 98,120 98,906 110,600 39,340 y6 24,777 23,440 23,380 23,270 25,972 26,070 3261 y7 92,100 88,030 84,830 84,830 81,357 73,900 21,150 y8 46,360 44,810 44,510 44,510 44,399 47,740 6754 y9 4581 4190 4188 4185 4274 4139 1481 y10 369.4 350.0 350.2 350.2 365.8 428.1 61.88 y11 287.5 275.8 276.6 275.6 290.9 223.3 37.23 y12 265.8 237.6 235.8 235.8 239.3 187.3 108.3 y13 285.2 273.6 271.2 271.2 282.1 284.7 175.0 y14 127.7 124.6 122.8 122.8 115.6 132.9 105.1 y15 369.0 362.4 356.3 356.3 314.0 301.2 111.4 Neural Computing and Applications (2019) 31:7227–7243 7235 123 lete’s parameters’’ box and a table with one or six annual training cycles depending on the types of generating training loads. 5.3 Panel for athletes’ database The third system panel is used to create and change the database containing athletes’ details (Fig. 8). This panel consists of two boxes: ‘‘Athlete’s database’’ and ‘‘Edit’’. In the first box, the records of the database loaded from the file are displayed. The system supports files saved in the .csv format with field separator ‘‘;’’ and ‘‘.’’ as the decimal point. The database file contains the following columns: ‘‘Name’’, ‘‘Surname’’, ‘‘Age’’, ‘‘Body Mass’’ and ‘‘Body Height’’. This box displays all athletes in a table; the choice of athlete is done by marking the appropriate line in this table. Furthermore, the name of the selected athlete is displayed in the sidebar. When an athlete is selected, his data can be edited via the ‘‘Edit’’ box. The saving of the edition is approved with the ‘‘Save’’ button. The ‘‘Delete’’ button removes the athlete from the database. The dese- lection of the athlete is done by re-selecting him/her in the database. If no athlete is selected, a new athlete can be entered into the database using the ‘‘Edit’’ window. After a new athlete is entered, you should click ‘‘Save as new’’. After each operation performed on the database, the user should save the database using the button ‘‘Save database’’ on the first panel. The second button on the panel (‘‘Clear database’’) performs cleaning the database from the application memory. In the ‘‘Athletes’ database’’ panel, it is not possible to generate reports. 6 Discussion In this paper, mathematical models for generation of training loads and prediction of results expected from athletes training the 110 and 400 m hurdle races were presented. The best model verified by LOOCV in each of the considered tasks and for each distance turned out to be the model F-OLS proposed by the authors. The application of fuzzy models in sport was also presented by Mezyk and Unold [36]. The goal was to find the rules that can express swimmers feelings the day after in-water training. Their method was characterized by better predictive ability than the traditional methods of classification, and the effec- tiveness was at the level of 68.66%. In Papić et al. [39], the fuzzy expert system was also presented. This system was based on knowledge of experts in the field of sport, as well as the data obtained as a result of motor tests. The model suggested the most suitable sport, and it was designed to search for prospective sports talents. Evaluation of the system showed high reliability and high correlation with top experts in the field. While analyzing the literature, it can be also noticed that mathematical models frequently used in sports are artificial neural networks [34, 35, 40, 44, 48, 50, 52, 58]. Numerous studies have shown that the ANN is a means of predicting sports results which has a good predictive ability [13, 59]. Thus, the ANN enables a coach to model the future level of athletes performance and supports the process of sports Table 5 Parameters for the GT module for 110 and 400 m hurdles. The meaning of the parameters is as follows: k and s are tuning parameters for regularized models, m is the number of hidden neurons in the RBF network, l is the number of fuzzy sets in fuzzy models Output RIDGE LASSO ENET RBF FUZZY F-OLS k s ðk; sÞ m l l 110 m y1 15.9 0.06 (0, 0.06) 4 5 3 y2 15.3 0.03 (6.03, 0.40) 2 7 10 y3 32.5 0.01 (1.80, 0.41) 5 4 3 y4 47.5 0.01 (0.87, 0.59) 7 7 7 y5 4.38 0.01 (0, 0.01) 9 5 3 y6 500 0 (0, 0) 9 8 7 y7 39.6 0.04 (0.65, 0.56) 2 6 10 y8 16.6 0.05 (1.30, 0.54) 10 12 12 y9 500 0 (0, 0) 2 13 3 y10 500 0 (0, 0) 9 5 5 y11 15.0 0.06 (1.20, 0.47) 10 4 4 y12 256 0 (49.0, 0.17) 2 3 6 y13 294 0.01 (0, 0.01) 7 10 13 y14 10.7 0.48 (19.3, 0.38) 5 4 10 y15 500 0 (0.01, 0.16) 5 3 4 400 m y1 500 0.03 (0.06, 0.09) 9 13 9 y2 500 0.02 (0.02, 0.03) 2 3 5 y3 288 0.05 (0, 0.05) 2 10 10 y4 500 0.04 (0, 0.04) 2 9 9 y5 500 0.01 (0.01, 0.01) 2 10 6 y6 30.7 0.51 (0.04, 0.72) 2 13 13 y7 19.5 0.33 (0, 0.33) 9 13 8 y8 156 0.15 (0, 0.15) 2 11 11 y9 322 0.02 (0.02, 0.05) 2 11 7 y10 88.8 0.08 (0, 0.08) 3 10 11 y11 72.6 0.06 (0.01, 0.12) 4 6 13 y12 500 0 (0, 0) 2 11 5 y13 339 0.02 (0, 0.02) 3 2 5 y14 10.5 0.21 (0, 0.21) 7 2 2 y15 140 0.04 (0, 0.04) 7 5 6 7236 Neural Computing and Applications (2019) 31:7227–7243 123 selection [34, 40, 48, 52]. For example, Silva et al. [52] presented high realistic models of swimming performance prediction based on multilayer perceptron. To establish a profile of the young swimmer, nonlinear combinations between preponderant variables for each gender and swim performance in the 200 m medley and 400 m front crawl events were developed. Artificial neural networks are also widely used in the process of planning training loads [44, 50]. In [50], Ryguła presents a new approach for determining training loads in a group of 16- and 17-year- old girls practising 100 m run. Sports training is the matter of making decisions about the quality (type of exercise) and the quantity (volume). This is a classical principle of sports training, emphasized in all textbooks on the theory of sports [8, 42]. Selection of training means and their distribution at subsequent stages of sports training is the main element of hurdlers’ training optimization on both distances, i.e. 110 and 400 m [50]. The selection of exercises (training means) in hurdling is supported by research in the field of motor preparation (strength, speed, endurance) as well as in relation to the technical structure of the event (kinematic analyses) [21]. Fig. 4 Screenshot of the iHurdling application with PR panel Neural Computing and Applications (2019) 31:7227–7243 7237 123 The observation of training programs of the best athletes [24], supported by the analysis of correlation between the results of hurdle run and the tests results including the physiological [16, 64] and biochemical basis [28], allows for selection of groups of the most valuable basic exercises. The performance tests carried out during the ergometric effort [5] are of great importance in assessing the specifics of hurdlers’ effort [5]. It should be emphasized that indi- vidualization of hurdlers’ training programs also requires an individual approach to the type of capacity, oscillating between aerobic and anaerobic capacity. Sprinting dis- tances in hurdling are considered to be typical running efforts of anaerobic nature. In the case of 110 m hurdle race, anaerobic non-lactic acidic changes with the final accent on anaerobic lactic acidic changes are predominant. The 400 m hurdle run requires first of all an effort of anaerobic non-lactic acidic nature [57]. Data concerning the specificity of the effort at a distance of 400 m indicate that the proportions of aerobic and anaerobic efforts can be significantly varied, taking into account the material (sports performance level of runners), method and period of training. In the review study by Arcelli et al., [2] those parameters adopted values within the range of 28–70% (aerobic) and 30–72% (anaerobic). The authors suggested that the higher the sports performance level, the higher the share of anaerobic element. The determination of the type of runner due to the aerobic and anaerobic processes would certainly make it possible to introduce some additional information in order to develop individual training. However, this problem has a logistic disadvantage, as monitoring of physiological reactions in hurdling is limited to the months when the athletes take part in competitions. Winter conditions are not conducive to specific running tests, and the choice of substitute distances may negatively affect the individual abilities of the hurdler. Taking into account the extensive scope of hurdlers’ exercises, the basic problem of a coach is the choice of exercises, their volume and proportions during particular training periods. The researchers pay their attention to that specificity of sports [8, 37]. Apart from the representative collection of training means, the body physique and age, often identified with the sports performance level, were also used. The impact of those elements on the organiza- tion of training has been already emphasized on several occasions [20, 23]. It seems appropriate to determine the initial value (record from the given year) and the estimated scale of progress (plans for the next season), because it makes it possible to control training loads depending on the athletes age, their current performance level and the main objective. Each athlete has different predispositions, also to perform specific training tasks. The selection of exercises is necessary, because it is impossible to perform the same volume of all exercises at the same time. Such a procedure would also be pointless, since the ‘‘rhythmic’’ type of a hurdler prefers running exercises with hurdles, and the ‘‘speedy’’ type of the hurdle runner prefers shorter dis- tances of the interval nature [20]. The database used is based on the period of 20 years of training Polish hurdlers, members of the national team. Those hurdlers represented various types (somatic, efficient and technical); therefore the scope of generalization (approximation) possibilities of the proposed computer system is significant and partly representative. Fig. 5 Screenshot of the box for entering endurance training loads 7238 Neural Computing and Applications (2019) 31:7227–7243 123 Summarizing the discussion, it should be noted that there are severe limitations of the presented approach connected with using the results in practice. The training programs do not consider the individual physiological and psychological parameters of an athlete. However, the generated training programs might be used as a suggestion for the coach who can perform necessary adjustments in order to adapt them for a particular athlete. 7 Conclusions In this paper, a web-oriented expert system to predict results and generate training loads for 110 and 400 m hurdles races was presented. The system uses the linear regression models (OLS, ridge, LASSO, elastic net) and nonlinear regression models (RBF, fuzzy model, OLS with fuzzy correction). The lowest errors were obtained by the Fig. 6 Screenshot of the panel for generation of training loads for one training programme Neural Computing and Applications (2019) 31:7227–7243 7239 123 proposed F-OLS model, but creating this model is more complicated. The application was implemented using R programming language with Shiny framework. The advantage of this application is that it can be run on multiple platforms such as personal computers and mobile devices. The easy-to-use interface allows the parameters of an athlete and the training loads to be changed. In this way, the coach can predict the expected result and select individual training components for a given athlete. Fig. 7 Screenshot of the panel for generation of training loads over the athlete’s entire career 7240 Neural Computing and Applications (2019) 31:7227–7243 123 Further work will focus on migrating the developed expert system to mobile application. Acknowledgements This work has been supported by the Polish Ministry of Science and Higher Education within the research project ‘‘Development of Academic Sport’’ in the years 2016–2019, Project No. N RSA4 00554. Compliance with ethical standards Conflict of interest The authors declare that they have no conflict of interest Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creative commons.org/licenses/by/4.0/), which permits unrestricted use, dis- tribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. References 1. Allaire J, Cheng J, Xie Y, McPherson J, Chang W, Allen J, Wickham H, Atkins A, Hyndman R (2016) rmarkdown: dynamic Documents for R. https://CRAN.R-project.org/package=rmark down. R package version 1.2. Accessed Feb 2017 2. Arcelli E, Mambretti M, Cimadoro G, Alberti G (2008) The aerobic mechanism in the 400 metres. New Stud Athl 23(2):15– 23 3. Arlot S, Celisse A (2010) A survey of cross-validation procedures for model selection. Stat Surv 4:40–79 4. Attali D (2016) shinyjs: easily Improve the User Experience of Your Shiny Apps in Seconds. https://CRAN.R-project.org/pack age=shinyjs. R package version 0.8. Accessed Feb 2017 5. Balsalobre-Fernández C, Tejero-González CM, del Campo- Vecino J, Alonso-Curiel D (2013) The effects of a maximal power training cycle on the strength, maximum power, vertical jump height and acceleration of high-level 400-meter hurdlers. J Hum Kinet 36(1):119–126 6. Bergmeir C, Benı́tez JM (2012) Neural networks in R using the Stuttgart neural network simulator: RSNNS. J Stat Softw 46(7):1–26. http://www.jstatsoft.org/v46/i07/. Accessed Feb 2017 7. Bishop CM (2006) Pattern recognition and machine learning. Information science and statistics. Springer, New York 8. Bompa TO, Haff G (1999) Periodization: theory and methodol- ogy of training, vol 199. Human Kinetics, Champaign 9. Chang W (2016) Shinydashboard: create dashboards with ’Shiny’. https://CRAN.R-project.org/package=shinydashboard. R package version 0.5.3. Accessed Feb 2017 10. Chang W (2016) Shinythemes: themes for Shiny. https://CRAN. R-project.org/package=shinythemes. R package version 1.1.1. Accessed Feb 2017 11. Chang W, Cheng J, Allaire J, Xie Y, McPherson J (2016) Shiny: web application framework for R. https://CRAN.R-project.org/ package=shiny. R package version 0.14.2. Accessed Feb 2017 12. Curtis KM (2010) Cricket batting technique analyser/trainer: a proposed solution using fuzzy set theory to aid West Indies Cricket. In: Proceedings of the 9th WSEAS international con- ference on artificial intelligence, knowledge engineering and data bases, AIKED’10. World Scientific and Engineering Academy and Society (WSEAS), Stevens Point, Wisconsin, pp 71–76 Fig. 8 Screenshot of the panel for athletes’ database Neural Computing and Applications (2019) 31:7227–7243 7241 123 http://creativecommons.org/licenses/by/4.0/ http://creativecommons.org/licenses/by/4.0/ https://CRAN.R-project.org/package=rmarkdown https://CRAN.R-project.org/package=rmarkdown https://CRAN.R-project.org/package=shinyjs https://CRAN.R-project.org/package=shinyjs http://www.jstatsoft.org/v46/i07/ https://CRAN.R-project.org/package=shinydashboard https://CRAN.R-project.org/package=shinythemes https://CRAN.R-project.org/package=shinythemes https://CRAN.R-project.org/package=shiny https://CRAN.R-project.org/package=shiny 13. Edelmann-Nusser J, Hohmann A, Henneberg B (2002) Modeling and prediction of competitive performance in swimming upon neural networks. Eur J Sport Sci 2(2):1–10 14. Er A, Dias R (2000) A rule-based expert system approach to process selection for cast components. Knowl Based Syst 13(4):225–234. https://doi.org/10.1016/S0950-7051(00)00075-7 15. Gu W, Saaty TL, Whitaker R (2016) Expert system for ice hockey game prediction: data mining with human judgment. Int J Inf Technol Decis Mak 15(04):763–789. https://doi.org/10.1142/ S0219622016400022 16. Gupta S, Goswami A, Mukhopadhyay S (1999) Heart rate and blood lactate in 400 m flat and 400 m hurdle running: a com- parative study. Indian J Physiol Pharmacol 43:361–366 17. Haghighat M, Rastegari H, Nourafza N (2013) A review of data mining techniques for result prediction in sports. Adv Comput Sci Int J 2(5):7–12 18. Hoerl AE, Kennard RW (1970) Ridge regression: biased esti- mation for nonorthogonal problems. Technometrics 12(1):55–67 19. Irani Z, Sharif A, Kamal MM, Love PE (2014) Visualising a knowledge mapping of information systems investment evalua- tion. Expert Syst Appl 41(1):105–125. https://doi.org/10.1016/j. eswa.2013.07.015 (21st Century Logistics and Supply Chain Management) 20. Iskra J (2012) Athlete typology and training strategy in the 400 m hurdles. New Stud Athl 27(1–2):27–37 21. Iskra J, Čoh M (2011) Biomechanical studies on running the 400 m hurdles. Hum Mov 12(4):315–323 22. Iskra J, Ryguła I (2001) The optimization of training loads in high class hurdlers. J Hum Kinet 6:59–72 23. Iskra J, Walaszczyk A (2003) Anthropometric characteristics and performance of 110m and 400m male hurdlers. Kinesiology 35(1):36–47 24. Iskra J, Widera J (2001) The training preparation of the world junior 400 m hurdles champion. Track Coach 156:4980–4984 25. Iskra J, Zajac A, Waskiewicz Z (2006) Laboratory and field tests in evaluation of anaerobic fitness in elite hurdlers. J Hum Kinet 16:25 26. Jain MB, Jain A, Srinivas MB (2008) A web based expert system shell for fault diagnosis and control of power system equipment. In: International conference on condition monitoring and diag- nosis, CMD 2008, pp 1310–1313 (2008). https://doi.org/10.1109/ CMD.2008.4580217 27. Kiartzis SJ, Bakirtzis AG, Theocharis JB, Tsagas G (2000) A fuzzy expert system for peak load forecasting. Application to the Greek power system. In: MELECON 2000: 10th Mediterranean Electrotechnical Conference, 2000, vol 3, pp 1097–1100. https:// doi.org/10.1109/MELCON.2000.879726 28. Kłapcińska B, Iskra J, Poprzecki S, Grzesiok K (2001) The effects of sprint (300 m) running on plasma lactate, uric acid, creatine kinase and lactate dehydrogenase in competitive hurdlers and untrained men. J Sports Med Phys Fit 41(3):306–311 29. Kusy M, Obrzut B, Kluska J (2013) Application of gene expression programming and neural networks to predict adverse events of radical hysterectomy in cervical cancer patients. Med Biol Eng Comput 51(12):1357–1365. https://doi.org/10.1007/ s11517-013-1108-8 30. Lapková D, Pluháček M, Komı́nková Oplatková Z, Adámek M (2014) Using artificial neural network for the kick techniques classification—an initial study. In: Proceedings 28th European conference on modelling and simulation ECMS, pp 382–387 31. Lee HJ, Rhee KP (2001) Development of collision avoidance system by using expert system and search algorithm. Int Ship- build Prog 48(3):197–212 32. Lo CY, Chang HI, Chang YT (2009) Research on recreational sports instruction using an expert system. Springer, Berlin, pp 250–262. https://doi.org/10.1007/978-3-642-04875-3_28 33. Louzada F, Maiorano AC, Ara A (2016) iSports: a web-oriented expert system for talent identification in soccer. Expert Syst Appl 44:400–412. https://doi.org/10.1016/j.eswa.2015.09.007 34. Maszczyk A, Roczniok R, Waśkiewicz Z, Czuba M, Mikołajec K, Zajac A, Stanula A (2012) Application of regression and neural models to predict competitive swimming performance. Percept Motor Skills 114(2):610–626 35. Maszczyk A, Zajac A, Ryguła I (2011) A neural network model approach to athlete selection. Sports Eng 13(2):83–93 36. Mezyk E, Unold O (2011) Machine learning approach to model sport training. Comput Hum Behav 27(5):1499–1506 37. Mujika I (2009) Tapering and peaking for optimal performance. Human Kinetics, Champaign 38. Najjaran H, Sadiq R, Rajani B (2006) Fuzzy expert system to assess corrosion of cast/ductile iron pipes from backfill proper- ties. Comput Aided Civ Infrastruct Eng 21(1):67–77. https://doi. org/10.1111/j.1467-8667.2005.00417.x 39. Papić V, Rogulj N, Pleština V (2009) Identification of sport talents using a web-oriented expert system with a fuzzy module. Expert Syst with Appl 36(5):8830–8838. https://doi.org/10.1016/ j.eswa.2008.11.031 40. Pfeiffer M, Hohmann A (2012) Applications of neural networks in training science. Hum Mov Sci 31(2):344–359 41. Pfeiffer M, Perl J (2006) Analysis of tactical structures in team handball by means of artificial neural networks. Int J Comput Sci Sport 5(1):4–14 42. Platonow N (2015) Sistema a podgotowki sportsmienow o olimpijskom sportie. Olimpijskaja literatura, Kiev (in Russian) 43. Przednowek K, Iskra J, Przednowek KH (2014) Predictive modeling in 400-metres hurdles races. In: 2nd international congress on sport sciences research and technology support- icSPORTS 2014. SCITEPRESS, Rome, pp 137–144 44. Przednowek K, Iskra J, Wiktorowicz K, Krzeszowski T, Maszczyk A (2017) Planning training loads for the 400 m hurdles in three-month mesocycles using artificial neural networks. J Hum Kinet 60(1):175–189 45. Przednowek K, Wiktorowicz K, Krzeszowski T, Iskra J (2016) A fuzzy-based software tool used to predict 110m hurdles results during the annual training cycle. In: Proceedings of the 4th international congress on sport sciences research and technology support (icSPORTS-2016). SCITEPRESS, pp 176–181 46. R Core Team (2016) R: A language and environment for statis- tical computing. R Foundation for Statistical Computing, Vienna. http://www.R-project.org/. Accessed Feb 2017 47. Riza LS, Bergmeir C, Herrera F, Benı́tez JM (2015) frbs: Fuzzy rule-based systems for classification and regression in R. J Stat Softw 65(6):1–30 48. Roczniok R, Ryguła I, Kwaśniewska A (2007) The use of Kohonen’s neural networks in the recruitment process for sport swimming. J Hum Kinet 17:75–88 49. Rogulj N, Papić V, Cavala M (2009) Evaluation models of some morphological characteristics for talent scouting in sport. Coll Antropol 33(1):105–110 50. Ryguła I (2005) Artificial neural networks as a tool of modeling of training loads. In: 27th annual international conference of the engineering in medicine and biology society, IEEE-EMBS, pp 2985–2988 51. Schröder S, Dabidian P, Liedtke G (2015) A conceptual proposal for an expert system to analyze smart policy options for urban cep transports. In: Smart cities symposium Prague (SCSP), pp 1–6. https://doi.org/10.1109/SCSP.2015.7181555 52. Silva AJ, Costa AM, Oliveira PM, Reis VM, Saavedra J, Perl J, Rouboa A, Marinho DA (2007) The use of neural network technology to model swimming performance. J Sports Sci Med 6(1):117–125 7242 Neural Computing and Applications (2019) 31:7227–7243 123 https://doi.org/10.1016/S0950-7051(00)00075-7 https://doi.org/10.1142/S0219622016400022 https://doi.org/10.1142/S0219622016400022 https://doi.org/10.1016/j.eswa.2013.07.015 https://doi.org/10.1016/j.eswa.2013.07.015 https://doi.org/10.1109/CMD.2008.4580217 https://doi.org/10.1109/CMD.2008.4580217 https://doi.org/10.1109/MELCON.2000.879726 https://doi.org/10.1109/MELCON.2000.879726 https://doi.org/10.1007/s11517-013-1108-8 https://doi.org/10.1007/s11517-013-1108-8 https://doi.org/10.1007/978-3-642-04875-3_28 https://doi.org/10.1016/j.eswa.2015.09.007 https://doi.org/10.1111/j.1467-8667.2005.00417.x https://doi.org/10.1111/j.1467-8667.2005.00417.x https://doi.org/10.1016/j.eswa.2008.11.031 https://doi.org/10.1016/j.eswa.2008.11.031 http://www.R-project.org/ https://doi.org/10.1109/SCSP.2015.7181555 53. Singh PK, Sarkar R (2015) A simple and effective expert system for schizophrenia detection. Int J Intell Syst Technol Appl 14(1):27–49. https://doi.org/10.1504/IJISTA.2015.072218 54. Tibshirani R (1996) Regression shrinkage and selection via the lasso. J R Stat Soc Ser B 58(1):267–288 55. Venables WN, Ripley BD (2002) Modern applied statistics with S. Springer, New York. http://www.stats.ox.ac.uk/pub/MASS4 56. Wang LX, Mendel JM (1992) Generating fuzzy rules by learning from examples. IEEE Trans Syst Man Cybern 22(6):1414–1427. https://doi.org/10.1109/21.199466 57. Ward-Smith A (1997) A mathematical analysis of the bioener- getics of hurdling. J Sport Sci 15(5):517–526 58. Wiktorowicz K, Przednowek K, Lassota L, Krzeszowski T (2015) Predictive modeling in race walking. Comput Intell Neurosci 2015:9. https://doi.org/10.1155/2015/735060 Article ID 735060 59. Wilk R, Fidos-Czuba O, Rutkowski Ł, Kozłowski K, Wiśniewski P, Maszczyk A, Stanula A, Roczniok R (2015) Predicting com- petitive swimming performance. Cent Eur J Sport Sci Med 9(1):105–112 60. Zarinbal M, Fazel Zarandi MH, Turksen IB, Izadi M (2015) A type-2 fuzzy image processing expert system for diagnosing brain tumors. J Med Syst 39(10):1–20. https://doi.org/10.1007/s10916- 015-0311-6 61. Zhou D, Ma J, Turban E, Bolloju N (2002) Soft decision analysis a fuzzy set approach to the evaluation of journal grades. Fuzzy Sets Syst 131(1):63–74. https://doi.org/10.1016/S0165- 0114(01)00255-X 62. Zou H, Hastie T (2005) Regularization and variable selection via the elastic net. J R Stat Soc Seri B (Stat Methodol) 67(2):301–320 63. Zou H, Hastie T (2012) Elasticnet: elastic-net for sparse esti- mation and sparse PCA. https://CRAN.R-project.org/package= elasticnet. R package version 1.1. Accessed Feb 2017 64. Zouhal H, Jabbour G, Jacob C, Duvigneau D, Botcazou M, Abderrahaman AB, Prioux J, Moussa E (2010) Anaerobic and aerobic energy system contribution to 400-m flat and 400-m hurdles track running. J Strength Cond Res 24(9):2309–2315 Neural Computing and Applications (2019) 31:7227–7243 7243 123 https://doi.org/10.1504/IJISTA.2015.072218 http://www.stats.ox.ac.uk/pub/MASS4 https://doi.org/10.1109/21.199466 https://doi.org/10.1155/2015/735060 https://doi.org/10.1007/s10916-015-0311-6 https://doi.org/10.1007/s10916-015-0311-6 https://doi.org/10.1016/S0165-0114(01)00255-X https://doi.org/10.1016/S0165-0114(01)00255-X https://CRAN.R-project.org/package=elasticnet https://CRAN.R-project.org/package=elasticnet A web-oriented expert system for planning hurdles race training programmes Abstract Introduction Training data Mathematical models Linear models Choosing the best model Nonlinear models RBF models Fuzzy models Expert system modules Models for result prediction Models for generation of training loads Graphical user interface Panel for result prediction Panel for generation of training loads Panel for athletes’ database Discussion Conclusions Acknowledgements References 
work_yugouyrzlvhj3bcnaeeh7bhj34	----	Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Type-1 Expert System EAI Endorsed Transactions on Scalable Information Systems Research Article Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Type-1 Expert System Shahan Yamin Siddiqui1,2,*, Syed Anwar Hussnain1, Abdul Hannan Siddiqui3, Rimsha Ghufran4, Muhammad Saleem Khan1, Muhammad Sohail Irshad2, Abdul Hannan Khan2 1School of Computer Science, National College of Business Administration & Economics, Lahore, Pakistan 2Department of Computer Science, Minhaj University, Lahore, Pakistan 3Cavan General Hospital Lisdaran, Cavan, Ireland 4Allied Hospital Faisalabad, Pakistan Abstract The adroit system is frequently used in artificial intelligence in medicine (AIM). They comprise medical information about a dedicated task and prone to purpose with data from case studies to produce lucid results. Though there are many irregularities, the information with an adroit network is derived with a set of expert rules to produce accurate results. Arthritis is the stiffness of one or more joints and about three fourth of the victims are suffering from it. Late detection of that chronic disease may cause the severity of the sickness at greater risk. So the idea is to contemplate a mechanism for the detection of arthritis using an adaptive hierarchical Mamdani fuzzy expert system (DA-AH-MFES). It is a befitting source to process ambiguity and inaccuracy. Physical and some medical parameters with the expertise of doctors can be mapped using MFES. The ability of MFES completely depends on the rules which are finalized by a discussion with an expert. The expert system has eight input variables at layer-I and four input variables at layer-II. At layer-I input variables are rest pain, morning stiffness, body pain, joint infection, swelling, redness, past injury and age that detects output condition of arthritis to be normal, infection and/or other problem. The further input variables of layer-II are RF, ANA, HLA-B27, ANTI-CCP that determine the output condition of arthritis. The performance of proposed Diagnose arthritis disease using an adaptive hierarchical mamdani fuzzy expert system is evaluated with expert observations of Cavan General Hospital Lisdaran, Cavan, Ireland and Jinnah Hospital Lahore, Pakistan. The accuracy of the expert system (DA- AH-MFES) is 95.6%. Keywords: Arthritis, Osteoarthritis, Rheumatoid arthritis, DA, MFES, DA-AH, MFES Received on 27 September 2019, accepted on 01 November 2019, published on 18 November 2019 Copyright © 2019 Shahan Yamin Siddiqui et al., licensed to EAI. This is an open access article distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/3.0/), which permits unlimited use, distribution and reproduction in any medium so long as the original work is properly cited. doi: 10.4108/eai.13-7-2018.161439 *Corresponding author. Email: engr.shahansiddiqui@gmail.com 1. Introduction Arthritis is the stiffness of one or more than one of your joints. Normally it is considered that if it influences old people but it can attack any age person. It is a chronic disease and a term that includes a group of disorders that affect joints and muscles [1]. Arthritis is a lifelong disorder and approximately three fourth of the sufferers are tormented by it; if it remains undiagnosed it may also cause the severances of the disorder [2, 3]. There are two major infections of arthritis; the first one is osteoarthritis when osteoarthritis infection attacks the person's joints cartilage becomes indentation on the surface, difficult and brittle. The bone below the cartilage get thickens and also broadens out, the amount of lubrication liquid of joints decreases. In a few instances, bony outgrowths may additionally shape at the outer edges of the joint, making 1 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 the joint seem tender. Due to osteoarthritis infection sufferer feels severe pain during movement [4]. The second type of arthritis infection is rheumatoid, in this infection membrane of joints swollen. Common symptoms of that infection feel by sufferer are stiffness, body temperature and redness. So to live a happy and healthy life it is very important to diagnose and cure that infection on time. If it remains undiagnosed in sufferer or late-diagnosed it may cause many other problems for sufferer [5]. Different expert systems are used in artificial intelligence in medical field to diagnose and cure different diseases. In this paper fuzzy logic controller is used to design an arthritis disease diagnose system. It is a very good tool that deals with uncertainty and imprecision. It is widely used in artificial intelligence to diagnose different diseases with high accuracy [6, 7, 8]. In this article proposed a fuzzy based intelligent system that will be able to diagnose arthritis at its early stage. The proposed system contains two layers to detect and ensure sufferer disease. Few physical parameters are used at layer-I to know about the type of infection sufferer has, if layer-I confirms that sufferer has arthritis infection then layer-II become an active layer. On layer-II few medical parameters will be used to diagnose the type of arthritis infection and also the stage of that infection. The above- explained process to diagnose arthritis is performed with the recommendation of medical experts in any sufferer. 2. Literature Review Arthritis is an ailment, Phrase ‘arthritis’ precisely way tenderness of the joints. The perceived and the existent studies prove that arthritis isn't always a diagnosis in itself. It is a widespread term that mentors us that something is incorrect [9, 10]. The recent researches have used the fuzzy inference system model to diagnosis arthritis based on some physical symptoms and medical tests. The most promising controller known as fuzzy logic that determines membership functions with decides rules about diagnosis results. The designed controller diagnosed arthritis in a single layer by combining different physical and medical parameters. Parameters involved and employed showed a different level of accuracy. The proposed model of that team used physical parameters like body pain and redness and medical parameters rheumatoid Factor (RF) and anti- cyclic citrullinated peptide antibody (ANTI-CCP). It shows accuracy between 8% – 82% [11, 12, 13]. Different causes which are considered baseline for arthritis disease. Worldwide research of his team observed fifty confirmed patients of rheumatoid arthritis having age twenty to sixty to know about common symptoms for future sufferers. They ensured in work that a common symptom that was present in sufferers and not in healthy subjects was the automated nervous system. Due to the presence of that symptoms sufferer does not feel well and does not communicate with others which leads to a short communication problem in the sufferer [14, 15]. Although there are other infections of arthritis that exists but most Common type of arthritis that is rheumatoid arthritis, prolonging the type Rheumatoid Arthritis, when and how it attacks the person, along with causes based on genetic symptoms, along with the symptoms, the medical remedies are also been discussed in the limelight of the arthritis [16]. Taking into account the vivid types of arthritis, osteoarthritis ailment, wear and tear infection leads to damage of cartilage and slowly cartilage damage fully and the person unable to move and work properly. The type osteoarthritis is considered to be very complex and it is a totally cartilage driven problem so it is very important to diagnose it immediately and cure it on time [17]. To analyze the origin of Rheumatoid Arthritis Smolen utilized dataset from different areas and discussed with arthritics experts. Treat to target (T2T) is used by experts of the Smolen team to cure Rheumatoid Arthritis [18]. The comprehensive discussion leads to the relation between depression and chronic rheumatoid arthritis. They utilized more than a hundred patient dataset including 77% males and 23% females. The result of the study of those patients reveals that more than 40% of the patient including both genders has some degree of depression present in their minds. The suggestions necessitate that it's miles important for treating Rheumatoid Arthritis victims to screen them for melancholy and manipulate it respectively [19]. The most alarming ailments have made the researches to approach the basic plain-film diagnosis and differential diagnosis of the arthritides and analysis of which part of the joint is involved, and works through a logical sequence to reach a final ailment [20]. A different kind of methodology has provided medical researches with a different guideline to diagnose and reduce rheumatoid arthritis. The two targeted types of sufferers have been implicate in paper and recommended different medications as per the severity of an ailment. The study of that paper reveals that although we cannot recommend it to all patients it may help us in more than half cases to reduce the infection of rheumatoid arthritis [21]. Computational Intelligence approaches like Fuzzy system [22,23], Neural Network [24], Swarm Intelligence [25] & Evolutionary Computing [26] like Genetic Algorithm [27, 28], DE, Island GA [29], Island DE [29,30] are strong candidate solution in the field of smart city [31, 32] and smart health [34,35,36,37] etc. 3. Related Work Fuzzy Based System Model Our proposed Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Expert System (DA-AH- MFES) is described in this area. Figure 1 determines the evolution of the proposed DA-AH-MFES hierarchical approach. The DA-AH-MFES based expert System includes two layers as appeared in figure 2. In layer-I examine the Arthritis (Negative/Positive) using eight input Shahan Yamin Siddiqui et al. 2 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 variables age, morning-stiffness, body-pain, joint- infection, rest-pain, redness, swelling, past-injury and in layer-2 examine the arthritis (no/acute arthritis/chronic arthritis) using four input variables RF, ANA, HLA-B27, Anti-CCP as performed in figure 3 and figure 4 using MATLAB R2019a tool [33]. Figure 1. Proposed DA-AH-MFES based Expert System Methodology Figure 2. Proposed DA-AH-MFES Expert System Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Type-1 Expert System 3 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Figure 3. Layer-1 of Proposed DA-AH-MFES Expert System using MATLAB R2019a tool Figure 4. Layer-2 of Proposed DA-AH-MFES Expert System using MATLAB R2019a tool The values of these parameters are also used to build up a lookup table given in Table-1 to evaluate the type of infection. The proposed automated Diagnosis Arthritis using mamdani inference based expert system can be expressed in mathematically regarding t-norm as Proposed DA-AH-MFES for Layer-1 can be written mathematically as μDIAGNOSIS-ARTHRITIS,L-1 (da-infection) = t [μAGE (age) , μMORNING-STIFFNESS (ms) , μBODY-PAIN (bp) , μJOINT-INFECTION (ji), μDIAGNOSIS- ARTHRITIS,L-1 (da-infection) = t [μAGE (age) , μMORNING-STIFFNESS (ms) , μBODY-PAIN (bp) , μJOINT-INFECTION (ji), μDIAGNOSIS- ARTHRITIS,LA-1 (da-infection) = min [μAGE (age) , μMORNING-STIFFNESS (ms) , μBODY-PAIN (bp) , μJOINT-INFECTION (ji), μSWELLING (sw), μPAST-INUJURY (pi), μREST-PAIN (rp), μREDNESS (rn)] & Proposed DA-AH-MFES for Layer-2 can be written mathematically as Shahan Yamin Siddiqui et al. 4 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 μDIAGNOSIS-ARTHRITIS,L-2 (da) = t [μDI (di), μRF (rf) , μANA (ana) , μHLA-B27 (hla) , μANTI-CCP (ac)] μDIAGNOSIS-ARTHRITIS,L-2 (da) = min [μDI (di), μRF (rf) , μANA (ana) , μHLA-B27 (hla) , μANTI-CCP (ac)] 3.1 Input Variables Proposed system Fuzzy input variables are numerical values that are used to diagnose arthritis. In this research, both layers used a total of twelve different kinds of input variables. Eight variables are used at layer-I and the remaining four variables are used at layers-II. The details of these input and output variables along with their ranges are shown in table 1, table 2 and table 3. Table 1. Layer-I input variables with medical ranges of proposed DA-AH-MFES expert system Sr # Input Parameters Medical Ranges Semantic Sign 1 AGE LT < 18 Child B/W 15 - 40 Young GT > 37 Old 2 MORNING STIFFNESS LT < 0.2 No Pain B/W 0.1 - 0.5 Minimum GT > 0.4 Maximum 3 BODY PAIN LT < 17 No Pain B/W 10 – 43 Minimum GT > 35 Maximum 4 JOINT INFECTION LT < 7 No B/W 4 – 45 Minimum GT > 42 Maximum 5 SWELLING LT < 13 No B/W 8 – 45 Minimum GT > 40 Maximum 6 REDNESS LT < 13 No B/W 8 – 45 Low GT > 40 Maximum 7 PAST INJURY LT < 15 No B/W 7 – 42 Minimum GT > 35 Maximum 8 REST PAIN LT < 0.20 No Pain B/W 0.1 - 3.5 Minimum GT > 3.4 Maximum Table 2. Layer-II input variables with medical ranges of proposed DA-AH-MFES expert system Sr # Input Parameters Ranges Semantic Sign 1 Rheumatoid Factor (RF) LT < 15 No B/W 7 – 42 Minimum GT > 35 Maximum 2 Antinuclear antibody (ANA) LT < 0.9 Negative GT > 0.5 Positive 3 HLA-B27 LT < 0.9 Negative GT > 0.5 Positive 4 ANTI-CCP LT <15 No B/W 7-50 Minimum GT > 40 Maximum 3.2 Output Variables In this research, the multilayered mechanism is proposed to diagnose the Arthritis. If the layer-I output is positive then layer-II is activated. Output variables for both layers are shown in table 3. Table 3. Layer-I & Layer-II output variables of proposed DA-AH-MFES expert system Sr # Layers Output Variables Semantic Sign 1 Layer-I Diagnose infection Negative Positive 2 Layer-II DA-AH- MFES No Arthritis Acute Arthritis Chronic Arthritis 3.3 Membership Functions Membership functions of the proposed automated DA- AH-MFES Adroit System give the curve value except for truth values and also dispense a mathematical form of fuzzy logic that propose numerical values of both input and output variables. The membership function of this proposed automated DA-AH-MFES adroit system is shown in table 4, table 5 and table 6. These membership functions are established with the help of Cavan General Hospital Lisdaran, Cavan, Ireland and Jinnah Hospital Lahore, Pakistan medical experts. Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Type-1 Expert System 5 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 https://www.healthline.com/health/rheumatoid-factor-rf https://www.healthline.com/health/rheumatoid-factor-rf Table 4. Mathematical & Graphical MF of Layer-1 Mamdani Fuzzy Inference System Input variables Input Membership Function Sample MF Screenshot Age = A Morning Stiffness = ms Body Pain =bp Shahan Yamin Siddiqui et al. 6 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Joint Infection =ji Swelling= Sw Redness= rn Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Type-1 Expert System 7 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Past Injury= pi Rest Pain= rp Table 5. Mathematical & Graphical MF of Layer-2 Mamdani Fuzzy Inference System Input variables Input Membership Function Sample MF Screenshot Diagnose Arthritis Infection-Layer 1= DI Shahan Yamin Siddiqui et al. 8 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Rheumatoid Factor = rf Antinuclear antibody = ANA HLA-B27 =hla Anti-cyclic citrullinated peptide = Anti-CCP Table 6. Mathematical & Graphical MF of Layer-1 & Layer-2 Mamdani Fuzzy Inference System Output variables Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Type-1 Expert System 9 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Input Membership Function Sample MF Screenshot Diagnos Arthritis - Layer 1= da-layer1 Diagnos Arthritis - Layer 2= da-layer2 3.4 Rules-Base Lookup Table For the proposed DA-AH-MFES rules table, which usually relies on the expert system comprises of 6561 input and output rules for layer 1 and 36 input and output rules for layer 2. A few of these input and output rules for layer 1 and layer 2 are presented in table 7 and table 8 respectively. The rule table is generally generated to the assistance of medical experts from the Orthopedic department of Cavan General Hospital Lisdaran, Cavan, Ireland and Jinnah Hospital Lahore, Pakistan. Table 7. Lookup table of Layer –I for DA-AH-MFES Expert System Rules Age Mornin g Stiffnes s Body Pain Joint Infectio n Swelling Past Injury Rest Pain Redness Results 1 Child No No No No No No No Negative 2 Child Min Min No Min Min No No 3 Young Min Min No No No No Max 4 Young Max No No Min Min No No 5 Old No Min Min No No Min No 6 Old Max No No Min No Min No 7 Child Min No No Min Min Min Min Shahan Yamin Siddiqui et al. 10 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Rules Age Mornin g Stiffnes s Body Pain Joint Infectio n Swelling Past Injury Rest Pain Redness Results 8 Young No Max Max Max Min Min Max Positive 9 Old Min Max Max Max Min Min Max 10 Old Max Min Max Min Max Max Min 11 Young Max Max Max Min Min Max Min 12 Old Max Max Max Max Max Max Max Table 8. Lookup Table of Layer –II for DA-AH-MFES Expert System Rules DI RF ANA HLA-B27 ANTI-CCP Output 1 Positive No Negative Negative No No Arthritis 2 Positive Minimum Negative Negative No 3 Positive No Positive Negative No 4 Positive No Negative Positive No 5 Positive No Positive Positive Minimum Acute Arthritis 6 Positive Minimum Positve Negative Minimum 7 Positive Maximum Positive Negative No 8 Positive Minimum Negative Positive Minimum 9 Positive Minimum Negative Negative Maximum Chronic Arthritis 10 Positive Maximum Negative Positive Minimum 11 Negative Maximum Positive Positive Minimum 12 Negative Maximum Positive Positive Maximum Rule-Based Rules are fundamental for input and output variables for both layers. The accomplishment of an Expert system constructs based on input an output rules. All rules for the proposed DA-AH-MFES appeared in figure 5. Inference Engine Inference Engine is the essential constituent of any decision based self-ruling framework. In this exposition of DA-AH-MFES, the Inference model is used in Layer I and Layer II. Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Type-1 Expert System 11 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Figure 5. Rules-based system for DA-AH-MFES Adroit System De-Fuzzifier De-Fuzzifier is one of the fundamental sections of any decision based self-governing framework. There are different sorts of defuzzifier. In this examination centroid type of De-fuzzifier is used. Figure 6 of all sections 6a-6d, demonstrates the De-fuzzifier graphical representation of both layers administers in DA-AH-MFES. Fig.6a. Rule Surface for RF and HLA-B27 Fig. 6b. Rule Surface for Anti-CCP and HLA-B27 Shahan Yamin Siddiqui et al. 12 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Fig. 6c. Rule Surface for RF and ANA Fig. 6d. Rule Surface for Anti-CCP and RF Figure 6. De-Fuzzifier graphical illustrations of Proposed DA-AH-MFES Expert System In Figure.6a blue color reperent the no arthritis , green color represent the acute arthritis and yellow color represent the chronic arthritis. The Arthritis (regarding Probability) based on Rheumatoid Factor and HLA-B27. Different colors in the Surface region present the stages of Arthritis. It is also observed that if Rheumatoid Factor is no (with the range between 0-15) and HLA-B27 less than 0.9 then the probability of arthritis is 0 that it may be any other types of arthritis. It is also observed that if the costs of Rheumatoid Factor are more the 35 its mean maximum and amounts of HLA-B27 are less the 0.9, the value of Arthritis is 80% that means it is no & chronic arthritis. Correspondingly, remaining figures 6b-6d present workmanship results by winning distinctive info parameter values. The surface district speaks to likelihood values by two information factors from given twelve input variables. The arthritis results are the blend of in any event three information factors. 4. Simulation and Results MATLAB R2019a [33] tool is used for representing, simulated, algorithm development, prototyping, and many other fields. This tool is efficient for software designing, data analysis, outset, and calculations. For the simulation of results, twelve inputs on both layers one output of diagnosing arthritis are used which are shown in figures 7, 8 and 9. Figure 7. Lookup diagram of Proposed DA-AH-MFES Expert System for No- Arthrists Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Type-1 Expert System 13 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Figure 8. Lookup diagram of Proposed DA-AH-MFES Expert System for Acute Arthritis Figure 9. Lookup diagram of Proposed DA-AH-MFES Expert System for Chronic Arthritis Figure 7 shows, If μRHEUMATOID-FACTOR (rf) is considered no, μANA (ana) is negative, μHLA-B27 (hla) is positive, μANTI- CCP (ccp) is no, then Outcome for μDA-LAYER2 (da-l2) is No arthritis. Figure 8 shows, If μRHEUMATOID-FACTOR (rf) is considered minimum, μANA (ana) is positive, μHLA-B27 (hla) is positive, μANTI-CCP (ccp) is minimum, then Outcome for μDA-LAYER2 (da-l2) is acute arthritis. Figure 9 shows, If μRHEUMATOID-FACTOR (rf) is considered maximum, μANA (ana) is positive, μHLA-B27 (hla) is positive, μANTI-CCP (ccp) is maximum, then Outcome for μDA-LAYER2 (da-l2) is chronic arthritis. The viability of the proposed method has sporadically kept an eye on many records. The Proposed DA-AH- MFES expert system gives definite outcomes for all stages, and just at acute arthritis, it might achieve some slight errors. Figure 10 speaks to the precision rate of DA- AH-MFES for the detection of arthritis. The proposed framework demonstrates the exactness rate for no arthritis is 96.20 percent, for acute arthritis is 93.60 percent and for chronic arthritis is 97 percent. The overall precision of the proposed DA-AH-MFES is 95.60 percent and the miss rate of the proposed DA-AH-MFES turns out to be 4.40 percent. 5. Conclusion The basic point of convergence of our examination is to devise an expert system to examine arthritis by reports retrieved from the Cavan General Hospital Lisdaran, Cavan, Ireland and Jinnah Hospital Lahore, Pakistan. This expert system is principal and simpler to use for both Medical specialists and non-specialists. This research engages orthopedic experts, surgeons and lab technicians for the judgment of the propose DA-AH-MFES expert system performance. Research obtained through the fuzzy logic model is critically analyzed in the supervision of orthopedic experts, surgeons and lab technicians. They conclude that this fuzzy logic model is to provide reliable results in the field of arthritis disease. The particular goal of this study is to analyze the various dimensions of arthritis. The overall accuracy of this proposed DA-AH- MFES is 95.6%. In the future, the productivity of the presented framework can be enhanced utilizing different methods including Computational Intelligence as Deep Learning, Neuro-fuzzy, and Neural networks. References [1] Krock, E., Jurczak, A., & Svensson, C. I. (2018). Pain pathogenesis in rheumatoid arthritis—what have we learned from animal models?. Pain, 159, S98-S109. [2] Giles, J. T., Danielides, S., Szklo, M., Post, W. S., Blumenthal, R. S., Petri, M., ... & Bathon, J. M. (2009). Insulin Resistance in Rheumatoid Arthritis. Imaging, 67(3), pp626-636. [3] Ruiz, D. G., Azevedo, M. N. L. D., & Lupi, O. (2012). HLA-B27 frequency in a group of patients with psoriatic arthritis. Anais brasileiros de dermatologia, 87(6), 847- 850. [4] Krastanova, M. S., & Vacheva, D. E. (2015). COMPLEX FUNCTIONAL ASSESSMENT OF THE HIP JOINT. Figure 10. Precision Chart of Proposed DA-AH-MFES Expert System Shahan Yamin Siddiqui et al. 14 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 Journal of IMAB–Annual Proceeding Scientific Papers, 21(3), 883-886. [5] Gonzalez, S., Garcia-Fernandez, S., Martinez-Borra, J., Blanco-Gelaz, M. A., Rodrigo, L., del Río, J. S., ... & López-Larrea, C. (2002). High variability of HLA-B27 alleles in ankylosing spondylitis and related spondyloarthropathies in the population of northern Spain. Human immunology, 63(8), 673-676. [6] Majithia, V., & Geraci, S.A. (2007). Rheumatoid arthritis: Diagnosis and management. The American Journal of Medicine, 120(11), 936–939. [7] Khan, M. A., Abbas, S., Hasan, Z., & Fatima, A. Intelligent Transportation System (ITS) for Smart-Cities using Mamdani Fuzzy Inference System. [8] Ahmad, G., Khan, M. A., Abbas, S., Athar, A., Khan, B. S., & Aslam, M. S. (2019). Automated Diagnosis of Hepatitis B Using Multilayer Mamdani Fuzzy Inference System. Journal of healthcare engineering, 2019. [9] Wu, R., Peters, W., & Morgan, M. W. (2002). The next generation of clinical decision support: linking evidence to best practice. Journal of healthcare information management: JHIM, 16(4), 50-55. [10] Vine, M. “Understanding Arthritics and its care”, Chesterfield, UK, Arthritics Medical Press, 2016. [11] Bhatla, N., & Jyoti, K. (2012). A Novel Approach for heart disease diagnosis using Data Mining and Fuzzy logic. International Journal of Computer Applications, 54(17). [12] Singh, S., Kumar, A., Panneerselvam, K., &Vennila, J. J. (2012). Diagnosis of arthritis through a fuzzy inference system. Journal of Medical Systems, 36(3), 1459-1468 [13] Asadullah, M., Khan, M. A., Abbas, S., Athar, A., Raza, S. S., & Ahmad, G. (2018). Blind channel and data estimation using fuzzy logic-empowered opposite learning-based mutant particle swarm optimization. Computational intelligence and neuroscience, 2018. [14] ESPINOZA, L. R., TOLOZA, S. M., VALLE-ONATE, R. A. F. A. E. L., & MEASE, P. J. (2012). Global partnering opportunities and challenges of psoriasis and psoriatic arthritis in Latin America: a report from the GRAPPA 2010 annual meeting. [15] Anderson, J., Caplan, L., Yazdany, J., Robbins, M. L., Neogi, T., Michaud, K. &Kazi, S. (2012). Rheumatoid arthritis disease activity measures: American College of Rheumatology recommendations for use in clinical practice. Arthritis care & research, 64(5), 640-647. [16] McInnes, I. B., &Schett, G. (2011). The pathogenesis of rheumatoid arthritis. New England Journal of Medicine, 365(23), 2205-2219. [17] Berenbaum, F. (2013). Osteoarthritis is an inflammatory disease (osteoarthritis is not osteoarthrosis!). Osteoarthritis and cartilage, 21(1), 16-21. [18] Smolen, J. S., Breedveld, F. C., Burmester, G. R., Bykerk, V., Dougados, M., Emery, P., &Scholte-Voshaar, M. (2016). Treating rheumatoid arthritis to target: 2014 update of the recommendations of an international task force. Annals of the rheumatic diseases, 75(1), 3-15. [19] Imran, M. Y., Khan, E. A. S., Ahmad, N. M., Raja, S. F., Saeed, M. A., &Haider, I. I. (2015). Depression in Rheumatoid Arthritis and its relation to disease activity. Pakistan journal of medical sciences, 31(2), 393. [20] Watt, I. (1997). Basic differential diagnosis of arthritis. European radiology, 7(3), 344-351. [21] Singh, J. A., Saag, K. G., Bridges Jr, S. L., Akl, E. A., Bannuru, R. R., Sullivan, M. C & Curtis, J. R. (2016). 2015 American College of Rheumatology guideline for the treatment of rheumatoid arthritis. Arthritis & rheumatology, 68(1), 1-26. [22] Hussain, S., Abbas, S., Sohail, T., Adnan Khan, M., & Athar, A. Estimating virtual trust of cognitive agents using multi-layered socio-fuzzy inference system. Journal of Intelligent & Fuzzy Systems, (Preprint), 1-16. [23] Areej Fatima, Muhammad Adnan Khan, Sagheer Abbas, Muhammad Waqas, Leena Anum, and Muhammad Asif, Evaluation of Planet Factors of Smart City through Multi- layer Fuzzy Logic (MFL). The ISC Int'l Journal of Information Security. 51-58, 2019. [24] Ayesha Atta, Sagheer Abbas, M Adnan Khan, Gulzar Ahmed, and Umer Farooq. An adaptive approach: Smart traffic congestion control system. Journal of King Saud University-Computer and Information Sciences, 2018. [25] Khan, M. A., Umair, M., Saleem, M. A., Ali, M. N., & Abbas, S. (2019). CDE using improved opposite based swarm optimization for MIMO systems. Journal of Intelligent & Fuzzy Systems, 37(1), 687-692. [26] Khan, M. A., Umair, M., & Choudhry, M. A. S. (2015). GA based adaptive receiver for MC-CDMA system. Turkish Journal of Electrical Engineering & Computer Sciences, 23(Sup. 1), 2267-2277. [27] Khan, M. A., Umair, M., & Choudry, M. A. S. (2015, December). Island differential evolution based adaptive receiver for MC-CDMA system. In 2015 International Conference on Information and Communication Technologies (ICICT) (pp. 1-6). IEEE. [28] Ali, M. N., Khan, M. A., Adeel, M., & Amir, M. (2016). Genetic Algorithm based adaptive Receiver for MC- CDMA system with variation in Mutation Operator. International Journal of Computer Science and Information Security, 14(9), 296. [29] Umair, M., Khan, M. A., & Choudry, M. A. S. (2015, December). Island genetic algorithm based MUD for MC- CDMA system. In 2015 International Conference on Information and Communication Technologies (ICICT) (pp. 1-6). IEEE. [30] Umair, M., Khan, M. A., & Choudry, M. A. S. (2013, January). GA backing to STBC based MC-CDMA systems. In 2013 4th International Conference on Intelligent Systems, Modelling and Simulation (pp. 503- 506). IEEE. [31] Kashif, I., Muhammad, A.K., Sagheer, A., Zahid, H., & Areej, F (2018). Intelligent Transportation System (ITS) for Smart-cities using Mamdani Fuzzy Inference System, International Journal of Advanced Computer Science and Applications (IJACSA). ISSN: 2158-107X, Vol. 9, No. 2, (pp. 94-105), Digital Object Identifier (DOI): 10.14569/IJACSA.2018.090215. [32] Wang, H., Zhang, Z., & Taleb, T. (2018). Special issue on security and privacy of IoT. World Wide Web, 21(1), 1-6. [33] MATLAB for Artificial Intelligence (www.mathworks.com). [34] Liu, F., Zhou, X., Wang, Z., Cao, J., Wang, H., & Zhang, Y. (2019). Unobtrusive Mattress-Based Identification of Hypertension by Integrating Classification and Association Rule Mining. Sensors, 19(7), 1489. [35] Liu, F., Zhou, X., Cao, J., Wang, Z., Wang, H., & Zhang, Y. (2019, May). A LSTM and CNN Based Assemble Neural Network Framework for Arrhythmias Classification. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 1303-1307). IEEE. [36] Arrhythmias classification by integrating stacked bidirectional LSTM and two-dimensional CNN. Pacific- Diagnosis of Arthritis Using Adaptive Hierarchical Mamdani Fuzzy Type-1 Expert System 15 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 http://www.mathworks.com/ Asia Conference on Knowledge Discovery and Data Mining, 136-149, 2019. [37] Dynamic optimisation based fuzzy association rule mining method. International Journal of Machine Learning and Cybernetics 10 (8), 2187-2198, 2019. Shahan Yamin Siddiqui et al. 16 EAI Endorsed Transactions on Scalable Information Systems 03 2020 - 05 2020 | Volume 7 | Issue 26 | e2 
work_yvcxfxdiqvh3tfyo4debbsim3u	----	The exploration of employee involvement model The exploration of employee involvement model W.-B. Lin * Department of Business Administration, National Formosa University, Huwei County Wen-Hua Rd., No. 64, Yunlin, Taiwan, ROC Abstract This research explores the cause variables, which affect employee involvement via diverse orientations. That is to say, it studies the relationship and effect of individual characteristic of personality traits, organizational climate of perspective of Chinese society relationship orientation, and internal marketing upon employee involvement. The population is based upon five companies from the top five Taiwan financial holding groups in Taiwan. The research distributed 200 copies of questionnaire to each company. The number of valid returns was 342, the valid return rate was 34.2%. According to the empirical research finding, high-intensity internal marketing generates positive impact upon employee involvement and low-intensity internal marketing results in negative impact upon employee involvement; the effect of sentiment relationship upon employee involvement is prominent and positive; employees with personality traits of internal control reveal higher level of involvement than those with external control. q 2005 Elsevier Ltd. All rights reserved. Keywords: Employee involvement; Organizational climate; Internal marketing; Personality traits 1. Introduction When most of the mature industries, particularly the industries which mainly offer service, enter the situation of small benefit, in order to upgrade the competitiveness of the enterprise, one would stress creativity, exporting speed and on- time feedback with regard to internal situation. Besides, from the perspective of internal marketing, stimulating the employ- ees and the establishment of reasonable and effective system are also the important means one can employ in the enterprise. The effective operation and the employment of democratiza- tion within the enterprise is one of the current or prospective tendencies in many publicly operated and private enterprises. Employee involvement is the means to examine the operation of democratization in the enterprises, which not only diminishes the obstacle of the operation of the system, but also is the way to collect common consensus and pursue the benefits of most of the people. For the industries which gradually pay attention to additional values, apart from upgrading the service standard and content toward the customers, they should also explore how to establish a complete working surrounding, which is not only considerably valued by the employees, but also the basis of the stability of the enterprise and sustainable operation. In other words, the degree of ‘employee involvement’ affects the achievements of the organization, and it is also the extended demand while the enterprises are pursuing additional values since the employees’ active participation in the process of decision will enhance their centripetal force in the organization as well as upgrade their working satisfaction (Locke & Schweiger, 1979; Miller & Monge, 1986; Schuler, 1980). The previous researches with respect to the related issues of employee involvement explore the study from the following directions: (1) elaborating the impact of employee involvement upon personal behavior and achievements of the organization by means of different theories or perspectives. For example, through the questionnaire survey with regard to the planners of information system in 105 companies, Basu, Hartono, Lederer, and Sethi (2002) believed that the commitment of the organization, the involvement of middle level managers and the group involvement prominently affected the accomplish- ment of the objective of strategic information system planning. That is to say, the higher the degree of commitment and involvement is, the more possibility that one will accomplish the strategic objective. Mitchell (1973) integrated the theories of cognition and expectation and further pointed out that employee involvement enhanced the employees’ working objectives, their identification toward the organization and the upgrading of working achievements. Based upon the survey with respect to strategic business units of the enterprises of various countries, Judge and Stahl (1995) believed that the personality traits of middle level managers would affect their Expert Systems with Applications 31 (2006) 69–82 www.elsevier.com/locate/eswa 0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.09.035 * Tel.: C886 5631 5775; fax: C886 5631 3842. E-mail address: e2667@ms23.hinet.net. http://www.elsevier.com/locate/eswa perceptive capacity. National culture plays the intermediary role. The managers’ cognitive degree would affect the operation of strategic projects. In addition, from the perspec- tive of internationalization, Strauss (1982) also proceeded with the comparative analysis with regard to the employee involvement of the projects of different countries. For instance, the most popular project of employee involvement in America was ‘the quality of work life’: one improved the work life of the employees of basic level via the discussion of the employees and middle level managers. Rafiq and Ahmed (2000) believed that internal marketing could be developed from different levels and perspectives. The first level was based upon the satisfaction of simulation with regard to the employees, i.e. to treat the employees as the internal customers and upgrade their satisfaction with the work. The second level was to integrate the internal opinions via the means of interactive marketing of customer orientation. The third level was to operate the strategies with regard to the company or functions via the conception of expanding internal marketing in order to respond to the change of the environment. (2) Regarding the content planned in the strategic project as the subject of exploration, for instance, how does employee involvement project affects production force, quality of work life, and commitment of organization and group evaluated by various empirical evidences (Griffin, 1988; Marks, Mirvis, Hackett, & Grady, 1986; Mellor, 1992)? Based upon the survey with respect to 69 Sweden companies, Ingelgard and Norvgren (2001) found out that the strategy of learning how to respond to the change of the surrounding was beneficial to the employees’ working adjustment. Besides, the employees’ participation was not only one of the factors which constructed the vision of the organization, but also the aspect that influenced the objective of reformation. Babin and Boles (1996) pointed out that the perception of the employees toward their colleagues’ involve- ment as well as the support of the chiefs could reduce the pressure in the company and enhance the working satisfaction. In other words, the researchers stressed how the working surrounding of the employees or the cognition of the organizational climate affect the production of the work. (3) Exploring the relationship between the cause and outcome variables from the perspective of human resources manage- ment, for instance, McMahan, Bell, and Virick (1998) and Schuler (1980) pointed out the cause variables, which affected employee involvement such as character traits (character expectation and conflict), the scale of organization, structure traits and outcome variables (e.g. the effect of personal and organizational achievements and working satisfaction). Although in the past, there were rare studies with regard to the issues related to employee involvement, one could still find the following aspects useful for further explorations, which are the objectives of this research. (1) The previous studies mostly explored the issue of employee involvement from the perspective of Western democratic decision and the said studies rarely discussed the subject from the standpoint of relationship orientation of Chinese society. (2) The information analysis with respect to previous studies was mostly based upon the qualitative case analysis or statistic method and rarely employed the analytic method of non-linear Neuro Fuzzy model. (3) This research explored the cause variables, which affect employee involvement from the perspective of various dimensions. In other words, the research studied the relationship and effect of individual traits (personality traits), organizational climate (the perspective of relationship orien- tation) and internal marketing with regard to employee involvement, which was different from the previous studies that merely explored the cause and outcome variables from single perspective. 2. Literature review 2.1. Internal marketing The previous marketing treated the employees as the internal customers and intended to fulfill the demand and desire of internal customers under the condition of complying with the operational objective of the company (Berry, 1981; Tansuhaj, Randall, & McCullough, 1988). In the past, there were many scholars who participated in the act of stimulating the employees to approach the objectives, which resulted in the orientation of customers and market (Gronroos, 1985; Johnson & Seymour, 1985). In other words, they employed more concrete project of stimulation. For example, via educational training and measures of stimulation, they allowed the employees to access to the mission and objective of the enterprise and they would further expect that in order to efficiently serve external customers, they must enhance the employees’ internal common consensus and the employees’ identification with the enterprise, which allowed the internal personnel of the company to efficiently convey the project of external marketing and carry out the project (Piercy & Morgan, 1991). Rafiq and Ahmed (2000) further pointed out that internal marketing included five elements, which were employees’ stimulation and satisfaction, orientation of custo- mer, customers’ satisfaction, the adjustment and integration of internal functions and the strategy for executing specific function of the company. George and Gronroos (1989) and Gronroos (1990) redefined the meaning of internal marketing: “via aggressive, active and nearly marketing act as well as the means of integration and adjustment, the organization provides the best encouragement and stimulation to the internal employees with regard to developing service consciousness and orientation of customers.” According to the previous definitions related to internal marketing, one can find out that the company paid attention to fulfill the demand of the employees and regarded stimulating and remaining excellent employees as the main developmental strategy (Berry, 1981; Berry & Parasuraman, 1991; Gronroos, 1985). Consequently, the perspective of internal marketing has developed from the perspective of human resource or market- ing into perspective of diverse dimensions. For instance, Rafiq and Ahmed (1993) pointed out that in order to achieve the objective of external market, the organization had to enhance efficient operation with regard to the internal exchange relationship between the organization and its employees. W.-B. Lin / Expert Systems with Applications 31 (2006) 69–8270 https://isiarticles.com/article/20061 
work_yvsfdafcgjarxhwee5mhpls4xq	----	Hierarchical Fuzzy Expert System for Organizational Performance Assessment in the Construction Industry algorithms Article Hierarchical Fuzzy Expert System for Organizational Performance Assessment in the Construction Industry Zenith Rathore 1 and Emad Elwakil 2,* 1 DPR Construction, Redwood City, CA 94061, USA; zenith.rathore@gmail.com 2 School of Construction Management, Purdue University, West Lafayette, IN 47907, USA * Correspondence: eelwakil@purdue.edu Received: 29 July 2020; Accepted: 19 August 2020; Published: 21 August 2020 ���������� ������� Abstract: Organizations have been trying to increase their efficiency and improve their performance in order to achieve their goals. Various factors determine organizational success. The construction industry is a project-based industry which is exceptionally dynamic. The need to identify the weak points and search for solutions to improve the performance of the construction organization is extremely crucial. The industry has always focused on the measure of project success. Previous research works have primarily focused on the measurement of financial or tangible assets. However, there is a lack of understanding of qualitative factors and their combined effect on organizational performance. Therefore, the objectives of this paper are to identify and study the success factors—both financial and non-financial factors. The potential success factors are collected from the literature review and construction experts through a questionnaire to evaluate their effect on organizational performance. The collected data have been analyzed using the Analytic Hierarchy Process (AHP) to shortlist the critical success factors. Thereafter, the Hierarchical Fuzzy Expert System has been used to build a prediction model based on the selected factors. The developed research/model benefits both researchers and practitioners to predict accurate company performance. Keywords: organization performance; analytic hierarchy process; hierarchical fuzzy expert system; construction industry 1. Introduction Construction is a diverse, project-based industry [1]. The project-based nature of the construction industry makes every project unique [2]. Moreover, the market structure is exceptionally fragmented, making it very competitive and challenging for any particular organization to dominate [3]. The unique nature of concerns and challenges often render the generalizable decision rules and frameworks for organizational phenomena unusable [4]. The financial and tangible assets gained are often translated to organizational success. In a review of project success factors conducted, it was noted that project success was considered only as a subject of implementation in the 1980s [5]. The approach towards the subject has evolved over the years. It is now extended from inception to the closing out of the project. Today, the literature in this field spans the entire product life cycle from product success to business success. This change has led to a shift in emphasis from project success to organizational success. The need to examine architectural/engineering/construction (A/E/C) organizations and the factors that impact the performance of organizations is now necessary to compete in an ever-changing marketplace [6]. Organizations have been trying to increase their efficiency and improve their performance in order to achieve their goals. Organizational success is determined by various factors that impact organizational performance. The uncertainty and uniqueness of projects are inherent characteristics of this industry, making it a conglomeration of unpredictable variables. Furthermore, the lack of Algorithms 2020, 13, 205; doi:10.3390/a13090205 www.mdpi.com/journal/algorithms http://www.mdpi.com/journal/algorithms http://www.mdpi.com http://www.mdpi.com/1999-4893/13/9/205?type=check_update&version=1 http://dx.doi.org/10.3390/a13090205 http://www.mdpi.com/journal/algorithms Algorithms 2020, 13, 205 2 of 25 performance measurements for the construction industry makes it challenging to evaluate these variables. Hence, developing a useful construction performance assessment model has been very difficult [7–10]. Globalized competition and customer needs forced construction companies to assess their performance beyond the financial measures—that is, profitability, turnover, etc. [11]. Profit and success are considered the main drivers of any organization. Achieving success depends on many factors that have a direct effect on the performance of organizations. Most construction organizational success factors are qualitative rather than quantitative. Thus, making it essential to determine these success factors, which can then be used later to predict and improve organizational performance. Modeling the performance of construction organizations from a financial perspective has been extensively researched; however, modeling the performance considering non-financial aspects has not received sufficient attention from researchers. The ability to predict construction organization performance will enable practitioners to identify weak points, which will lead to search solutions to improve efficiency, which will ultimately increase profits and success [7]. Studies conducted in the construction industry have laid more emphasis on the measurement of project performance rather than company performance. Moreover, studies have focused on individual or a combination of a few factors. Due to the limited scope, the whole system has not been evaluated, thus making it necessary to understand and identify the underlying relationship between the factors amongst categories and across different categories that impact an organization’s performance. Alternatively, different prediction models need to be developed depending on the organization size, specialty contractors and types of contracts undertaken—for example, Engineering Procurement and Construction (EPC), Construction Management at Risk (CMR), Architectural and Design firms, General Contractor. Therefore, the objectives of this paper are to identify and study the success factors—both financial and non-financial factors—to investigate the factors that impact the organization’s performance in the construction industry. Moreover, another objective is to develop a comprehensive prediction model based on the non-financial and financial factors that impact an organization’s performance. The following sections present the research background, objectives, methodology, model building, validation, and conclusion and recommendations for future studies. 2. Background 2.1. Critical Success Factors Determining factors for project success or failure has been of keen interest to both academicians and industry professionals. Most of the factors identified have focused on project execution rather than organizational success. Cooke-Davies [12] mentioned that, although project management literature does not illustrate much corporate success, both direct and indirect links exist. Organizational effectiveness depends upon the successful management of the projects [4]. Project success brings about a beneficial change to the organization and vice-versa [12]. Similarly, any improvement in the organization’s structure will improve the chances of project success. Several critical success factors influence project success—for example, top management support, communication, and sufficient resources are derivatives of organizations. Furthermore, the study recognizes important factors that link project success and corporate success. These factors are categorized into five areas, which are, general corporate strategy, business operations, research and development, IT/IS development, and facilities management. The paper stresses that every factor deals with people, as they are the ones who execute the project. Thus, it is necessary to include the influence of people in organizations. Pinto and Covin [4] and Muller and Jugdev [5] have discussed that project success is dependent on the interaction of individuals, project teams, and organizational success. Chinowsky and Meredith [13] proposed the concept of seven guiding principles of strategic management for the construction industry. These include vision, mission, goals, core competencies, and knowledge resources, education, finance, markets, and competition. Knowledge and information Algorithms 2020, 13, 205 3 of 25 are now considered as critical factors that influence a company’s lifespan. They are rated higher than land, capital, or labor [14]. A good knowledge database will allow organizations to leverage against their competitors in the future and thus give organizations a competitive edge [15]. Unfortunately, knowledge, being an intangible asset, is difficult to measure and hence is often forgotten in the process [14]. Organizations are conceptualized as the product of thought and action of their members [16] or, as Weick [17] stated, the body of thought by organizational thinkers. Human elements are the assets of organizations that are capable of learning, evolving, innovating, and creatively propelling the growth of an organization, which is essential for the long-run survival of the organization. It has been noted that the majority of Human Resource Accounting (HRA) techniques have been designed for industries such as accounting firms, banks, insurance companies, and financial service firms, where human resources represent a substantial share of the organization value [14]. However, construction organizations lack such initiatives that are designed to evaluate employee performance, satisfaction, and compensation. Factors such as organization employee culture and engagement are essential aspects of an organization. Another critical factor is the feedback system, as this is incredibly crucial for the implementation of the metric system and evaluating the performance of the organization. Feedback evaluation is one of the critical success factors that aids in analyzing and improving organization performance [18]. The earliest seminal works in the field of economics by Viner [19] on long-run average cost cycles show that economies of scale help organizations to grow efficiently up to a certain critical production level. The expansion of a firm that results in reduced cost is called the economy of scale. There are two types of economies of scale—internal and external. Internal economies of scale are long term phenomena achieved by the appropriate adjustment of scale of operations to the successive output [19]. Technical economies allow organizations to capitalize on the processes and assets developed. Large firms benefit from established credit lines. Large firms can achieve risk-bearing economies as they can afford to take higher risks and take up high-risk projects. Firm size is one of the factors that can impact an organization’s growth. If the firm is too big, the management communication can be inefficient due to poor communication and coordination problems. Factors, such as the morale of employees, are intangible and, hence, are difficult to account for in an organization’s growth by just looking at financial statements. Large firms also experience inefficiencies due to the principle-agent problem. Viner [19] also pointed out that the internal economy of scale is independent of the external economy of scale. The external economy of scale refers to the positive developments or increase in output generated by the industry as a whole. Similar to the internal pecuniary economy of scale, the external pecuniary economy of scale also benefits organizations when there is an increase in the number of suppliers, and offer more competitive prices. Challenging Viner’s theory of the impact of firm size and economies of scale on the organization performance, Simon and Bonini proposed a stochastic mechanism using Gibrats law for firm growth and the skewed distribution of firm sizes [20]. The results show that the distribution of the percentage of change in the size of firms in a given size class is the same for firms in all size classes. Thus, the expected rate of growth is independent of the current size of a firm. 2.2. Existing Performance Metrics Benchmarking has been defined as a continuous, systematic process for evaluating the products, services, and work processes of organizations that are recognized as representing best practices for the purpose of organization improvement [21]. A company is a complex structure, comprising of various interconnected components that influence its performance [22]. Performance prediction of construction organizations includes the identification of the weak points in order to improvise processes and to increase profits [23]. The attention of organizations is usually focused on improving the efficiency of its tangible assets as they can be measured and evaluated [18]. In the process, the organizations often do not consider the invisible and intangible assets that impact the overall performance. A good metric system empowers an organization [18]. In a recent study and analysis of a case study by Gustavsson [24], a need for new collaborative project practice development and organizational change Algorithms 2020, 13, 205 4 of 25 has been discussed. Company performance is usually assessed by the evaluation of measurable characteristics of performance indicators [25]. At the same time, it is crucial to understand that the productivity or output in the construction industry is not homogeneous; that is, outputs cannot be measured in a cubic meter. Given the diverse nature of the construction industry, it is impossible to aggregate all types of outputs and measure them with one physical measurement unit. It is essential to understand the heterogeneous results and develop ways to analyze them [26]. The existing literature shows that numerous models were developed to measure performance by using critical success factors, performance measures, and indicators. Academics have a tendency to characterize projects as similar entities; thus, these studies have been done looking at the broader picture rather than for a particular case [4]. These studies mostly address metric requirements for the manufacturing industries rather than construction. It is important to note that product life in the manufacturing industry goes through a standard process. The performance is usually measured at a per-unit cost. The repetitive process makes it possible to standardize the process and improve overall performance. The project management studies have been shifting focus to organizational strategies and operations. World manufacturers are now competing on crucial success factors other than price/cost. Unarguably, the characteristics and properties of goals and challenges may be similar. However, too often, academics have generalized decision rules for organizational phenomena, while practitioners have been stressing the unique nature of their concern [4]. The closest initiative to measure construction performance was based on total quality management [27]. One of the earliest measurement instruments developed to measure the project performance was proposed by Pinto and Slevin [28]. The Project Implementation Profile (PIP) allows for the assessment of an organization’s ability to carry a project through its full implementation [29]. The PIP was a support tool to enable managers to assess the status of their project by seeking answers to questions related to 10 critical success factors [28]. The process required participants to give responses on a 5-point Likert scale. These responses were used to assess success or failure in terms of schedule, budget overrun, quality of work, client satisfaction, and the utility of the final project. It is important to note that the tool was developed focused more on project success rather than organization success. In 1992, members of the Houston Business Roundtable (HBR) embarked on the journey of establishing performance metrics. This process included sending out a survey to member companies of HBR to determine four main preliminary tasks: determine the interest in HBR companies in a metric system; identify activities that should be measured; determine how to measure activities; collect information and analyze information. After confirming a 90% interest and willingness from a HBR member, the HBR members decided on ten activities that were selected for benchmarking. The ten factors were costs (actual vs. authorized), schedule (actual vs. estimated), scope changes, reengineering work, construction labor (actual vs. estimated), worker hours per drawing, project cost distribution, field defects, and percent of rejected welds [27]. Studies conducted in the construction industry have laid more emphasis on the measurement of project performance rather than company performance [11]. Bontis and Dragonetti [14] proposed the Balanced Score Card (BSC). The framework emphasized qualitative measures at the organizational level and advocated the balance between measures of financial and non-financial success. Another example of performance measurement and management framework is the Performance Prism. The first part of this framework encourages assessing stakeholder satisfaction and assessing the needs of the stakeholder. The second part is to understand the needs of the organization (i.e., the reciprocal relationships) as well as on how to align strategies, processes, and capabilities [30]. The Prism focuses on significant measures and connects the performance practices within the organization. These frameworks are more than a decade old. Hence, in order to keep up with the ever-changing markets, many new studies are being carried out. Performance measurement has always been a challenge in the construction industry [29]. The construction industry has not seen much improvement in productivity and performance Algorithms 2020, 13, 205 5 of 25 measurement, as in the manufacturing sector [31]. Industry groups in several different countries have initiated benchmarking programs focused mainly on construction performance measures [32]. The earliest concepts of benchmarking systems in the construction industry were introduced in the 1990s and were initiated by countries such as the United States of America, the United Kingdom (UK), Chile, Japan, and Brazil. In 1993, the Construction Industry Institute introduced the first benchmarking system in the public sector of the construction industry. This was followed by the Construction Excellence Program launched by the Construction Best Practice Program (CBPP) and the Key Performance Indicators (KPIs) program launched by the UK Best Practice Program in 1988 [29]. In 2008, the Construction Sector Council, a Canada based organization, launched a program to measure and benchmark project performance in the Canadian construction industry. The metrics developed benchmarks to measure: project cost, time, safety and quality performance; labor productivity; rework; project conditions and management practices related to health. The goal of the program was to develop benchmarks to assess labor productivity and project performance [29]. Again, the study was based on project success factors. Costa et al. [32] summarized various benchmarking systems employed by the construction industry from four different countries (i.e., Brazil, Chile, the UK, and the US). The benchmarking initiatives are: 1. Key Performance Indicators in the UK; 2. National Benchmarking System for Chilean Construction Industry (NBS-Chile); 3. CII benchmarking system and metric in the US; 4. The performance measurement for benchmarking in the Brazilian Construction Industry. These programs have generated recommendations, such as: Classification of performance measures, establishing frameworks that allow performance to performance management; 3. developing collaborative learning processes; inventing new measures and developing a framework for performance assessment. Another framework proposed by the Canadian Construction Innovation Council (CCIC) evolved from the project success factors to a framework that encompasses factors that impact the organization’s functioning. The metrics developed by the CCIC would be relevant to the project and organization level and also allow for the indication and assessment of performance at the organization level. This framework included factors that are categorized into seven main performance categories. The factors are costs (estimated, actual and predicted), time (estimated, actual and predicted), quality (levels of client satisfaction), safety (incidents and lost time), innovation (procurement, management, technology), and sustainability (design and construction). The major drawback of this framework is that it required accurate data for a large number of factors, and the analysis followed was even more complicated [29]. Organizations that focus on satisfying customers with greater efficiency and effectiveness have the edge over their competitors [33]. Studies have shown that practitioners have been able to determine that improving communication has a significant impact on construction practice. It allows better customer engagement, leading to the better performance of organizations. Neely et al. [33] stress the importance of metrics associated with quality, time, cost, and flexibility, thus relating the performance of organizations with project success. Attempts have also been made to understand the relationship between the internal and external factors affecting organizational performance. Empirical studies carried on the construction market structure show that the construction industry is highly fragmented, which makes it very competitive [3]. Studies have been carried out to identify the relationship between market fragmentation and organizational diversification [3]. Choi and Russell [34] used 12 years of data of publicly traded companies to identify any relation between diversification and profitability. However, no significant difference in profitability was found in companies categorized by different diversification level. In a previous study by Zayed et al. [23], nine critical success factors (CSFs) were defined as the most significant to develop a prediction model for organizational performance. The Artificial Neural Network (ANN) model was used to assess the most significant success factors, as ANN provides the contributing weight of each factor after the completion of the training process. Another study Algorithms 2020, 13, 205 6 of 25 developed a fuzzy logic model with the same data aiming to develop the best fit model for performance prediction [7]. 2.3. Modelling Techniques Adopted—Fuzzy Approach Lotfi Zadeh introduced fuzzy logic as a powerful modeling technique that can be used for understating the uncertainty and qualitative aspects of human nature [35]. Fuzzy techniques have been widely utilized in several studies over the past decade. It has the ability to virtually connect humans to computers by analyzing linguistic inputs to stem numerical outputs [36]. Traditionally, a set of inputs has sharp and crisp boundaries, where elements are either in or out of a set, and the ranking of membership of a variable is zero or one [37]. However, in the real world, information is mainly ambiguous and incomplete. That is when fuzzy logic is applicable, as elements are allowed to have partial memberships ranging from zero to one (i.e., zero is no membership, and one is full membership) [38]. Zayed et al. [23], has previously developed prediction models for the performance of construction organizations using the Artificial Neural Network (ANN) model and regression. A total of 18 factors were identified from the literature review. Based on the responses received from industry experts (5-point Likert scale was used), these factors were evaluated and allotted ranks using ANN training (i.e., ranking the factors to determine the relative importance of each variable and the highest impact on the model). Analysis of weights of the trained neural network is used to derive the contribution percentages. The higher value implies that the variable contribution to classification/prediction is also high. Based on the ANN rankings, nine factors with the highest contributing factor were shortlisted from the pool of 18 factors. However, this advantage comes at the cost of the minimized interpretability of the model output. The black box quality of an ANN model makes it next to impossible to gain insight into a problem based on an ANN model. The regression technique allows the user to sequentially remove possible explanatory variables that do not contribute to the fit of the model [39]. Regression techniques permit hypothesis testing concerning both the univariate and multivariate association amongst each explanatory variable and the outcome of interest. However, it fails to recognize or identify the highly nonlinear factors or correlation among variables [39]. Human reasoning being more approximate than precise often makes it difficult to measure and determine the measure of factors affecting a particular cause. 2.4. Challenges and Limitations of Existing Metrics The process of developing a successful performance prediction model is a very long and tedious task. It takes the analysis of a large number of factors from a broad stratum of projects. The data requirements are immense. Additionally, the project values, life-cycle, location, etc., are the variables that need to be accounted for. The time taken to develop the program, identify potential participants, introduce the concept, obtain feedback, revise parameters, and re-evaluate can be extremely long. Furthermore, it is a challenge to convince firms to provide data for on-going projects and data for any changes that may be observed after the suggested actions. Once a benchmark has been developed, it becomes a significant strategic asset. CII took almost eight years to derive a functional benchmarking model with a considerable number of projects to make meaningful assessments at the project level [40]. Despite the awareness and importance–performance measurement data, companies, or knowledge bodies have not been able to establish data banks [32]. Existing empirical studies only focus on a few factors and attempt to establish a relationship. These factors can be internal—that is, within the organization—and external, related to market conditions. Previous studies have focused on individual or combination of a few factors. Due to the limited scope, the whole system was not evaluated, thus making it necessary to understand and identify the underlying relationship between the factors amongst categories and across different categories that impact an organization’s performance. Algorithms 2020, 13, 205 7 of 25 The data from companies vary as the companies that execute small projects only form a fraction of the total industry turnover. Only large organizations can afford to execute projects with a minimal profit margin, as they have in house capabilities, established processes, established lines of credit, and qualify for large projects. However, it is essential to note that small organizations barely manage to stay afloat. It will thus be necessary to account for the organization size while developing a prediction model. Alternatively, different prediction models need to be developed depending on the organization size, specialty contractors, and types of contracts undertaken—for example, Engineering Procurement and Construction (EPC), Construction Management at Risk (CMR), Architectural and Design firms, General Contractor. 3. Framework and Methodology The existing literature shows that numerous models were developed to measure performance by using critical success factors, performance measures, and indicators. However, they mostly address metric requirements for the manufacturing industries rather than construction. Studies conducted in the construction industry have placed more emphasis on the measurement of project performance rather than company performance [11]. The methodology, shown in Figure 1, is summarized stepwise as follows: • A literature review was conducted to identify the factors that impact the performance of the construction organizations. Factors shortlisted from the literature review were analyzed for their impact on the performance of the construction organization. These factors were also referred to as Critical Success Factors (CSFs). • Based on the literature review conducted, factors that have an impact on organizational performance were shortlisted. The proposed performance assessment model included both qualitative and quantitative factors. The model was developed in order to determine the overall performance in terms of the category of rank. Due to the nature of the research question, the researcher adopted a mixed-method qualitative and quantitative approach. • A questionnaire was designed to evaluate the impact and implementation of the shortlisted non-financial CSFs in their respective organizations. The questionnaire also asked about the participant’s total number of experiences, the designation held in the current organization, and the name of the organization. The questionnaire was distributed to professionals across the construction industry via in-person interaction, emails, and an online Qualtrics survey. • Simultaneously, a database for quantitative factors was compiled for all organizations, whose employees participated in the survey. From this data, factors representing financial trend analysis and market diversification of organization were calculated. • Analytic Hierarchical Process (AHP) was used to shortlist 18 qualitative factors. The impact rating of the 18 factors was derived from participants as an expert opinion. The results of this process was validated from the Best Subset Regression function. The subset with the highest Rsq and adjusted Rsq was used for modeling purposes. • A performance assessment model for the construction organization was developed using a Hierarchical Fuzzy Expert System. As the number of factors was extensive and their values were on different scales, it was recommended to use the Hierarchical Fuzzy Expert System (HFES). • The first layer of HFES was developed by building two sub-models of the fuzzy expert system for non-financial (qualitative) factors and financial (quantitative) factors. The input variables were the respective sub-factors, and the output was the impact value of the combined effect of sub-factors. • The output from the first layer was used as input for the second layer of the fuzzy expert system. The input variables were assigned fuzzy membership, and fuzzy relations were established. The defuzzification gave the category of the rank of the organization. Algorithms 2020, 13, 205 8 of 25 • The model was tested and mathematically validated by Average Validity Percentage (AVP) and Average Invalidity Percentage (AIP) in order to determine the accuracy in assessing the performance of the construction organization [41]. Algorithms 2020, 13, x FOR PEER REVIEW 8 of 25 Figure 1. Research framework. 4. Study Variables The study aims at evaluating organizational performance based on both financial and non- financial parameters. Hence, the study variables were categorized into three broad categories—that is, Non-financial parameters, Financial parameters, and Market Conditions—as presented in Figure 2. Figure 1. Research framework. 4. Study Variables The study aims at evaluating organizational performance based on both financial and non-financial parameters. Hence, the study variables were categorized into three broad categories—that is, Non-financial parameters, Financial parameters, and Market Conditions—as presented in Figure 2. Algorithms 2020, 13, 205 9 of 25Algorithms 2020, 13, x FOR PEER REVIEW 9 of 25 Figure 2. Study variables. 4.1. Independent Variables This study aims at assessing the qualitative factors with quantitative factors. Independent variables are the input variables that determine the value of output or dependent variable. Based on the literature review, 18 qualitative factors were shortlisted for non-financial critical success factors, as shown in Figure 2. These factors have previously been investigated in a study conducted by Elwakil et al. [42] and Rathore and Elwakil [7]. The qualitative variables were categorized into four categories—i.e., Administrative and Legal; Technical; Management and Market and Finance—as presented in Figure 3. Figure 3. Non-financial critical success factors. A total of 18 quantitative factors were identified for the study. These include financial factors representing organization growth and market diversification. Based on the publicly available financial information, factors pertaining to revenue and annual growth were included in the model. A longitudinal database of revenue for organizations for the past five years was compiled. The factors included annual growth rate in revenue, three years cumulative, percent of different market segment revenue, productivity (revenue/employee), the total number of years in business, and firm size (number of employees), as shown in Figure 3. Figure 4 shows the financial factors compiled from the Figure 2. Study variables. 4.1. Independent Variables This study aims at assessing the qualitative factors with quantitative factors. Independent variables are the input variables that determine the value of output or dependent variable. Based on the literature review, 18 qualitative factors were shortlisted for non-financial critical success factors, as shown in Figure 2. These factors have previously been investigated in a study conducted by Elwakil et al. [42] and Rathore and Elwakil [7]. The qualitative variables were categorized into four categories—i.e., Administrative and Legal; Technical; Management and Market and Finance—as presented in Figure 3. Algorithms 2020, 13, x FOR PEER REVIEW 9 of 25 Figure 2. Study variables. 4.1. Independent Variables This study aims at assessing the qualitative factors with quantitative factors. Independent variables are the input variables that determine the value of output or dependent variable. Based on the literature review, 18 qualitative factors were shortlisted for non-financial critical success factors, as shown in Figure 2. These factors have previously been investigated in a study conducted by Elwakil et al. [42] and Rathore and Elwakil [7]. The qualitative variables were categorized into four categories—i.e., Administrative and Legal; Technical; Management and Market and Finance—as presented in Figure 3. Figure 3. Non-financial critical success factors. A total of 18 quantitative factors were identified for the study. These include financial factors representing organization growth and market diversification. Based on the publicly available financial information, factors pertaining to revenue and annual growth were included in the model. A longitudinal database of revenue for organizations for the past five years was compiled. The factors included annual growth rate in revenue, three years cumulative, percent of different market segment revenue, productivity (revenue/employee), the total number of years in business, and firm size (number of employees), as shown in Figure 3. Figure 4 shows the financial factors compiled from the Figure 3. Non-financial critical success factors. A total of 18 quantitative factors were identified for the study. These include financial factors representing organization growth and market diversification. Based on the publicly available financial information, factors pertaining to revenue and annual growth were included in the model. A longitudinal database of revenue for organizations for the past five years was compiled. The factors included annual growth rate in revenue, three years cumulative, percent of different market segment revenue, productivity (revenue/employee), the total number of years in business, and firm size (number of employees), as shown in Figure 3. Figure 4 shows the financial factors compiled from the publicly Algorithms 2020, 13, 205 10 of 25 available data. Additionally, the Market Diversification of an organization was measured by Entropy, which was computed using Equation (1). Algorithms 2020, 13, x FOR PEER REVIEW 10 of 25 publicly available data. Additionally, the Market Diversification of an organization was measured by Entropy, which was computed using Equation (1). Figure 4. Financial critical success factors. 4.2. Dependent Variables The dependent variable was the rank of the organization published by ENR Top 400 Contractors and ENR Top 500 Design Firm Sourcebook 2015 [43]. The rank was inversely proportional to the revenue of the organization in the year 2015. The higher the revenue earned, the smaller the rank. To maintain the confidentiality of the organizations, the rank was categorized into categories—i.e., 1– 100, 101–200, 201–300, 301–400, 401–500, and 501 and above. This variable was also the output in the model. 5. Data Collection The data collection procedure included a literature review and the identification of potential critical success factors. Eighteen non-financial factors were shortlisted. Then, the questionnaire designed to assess the impact and implementation of these factors in the construction industry was prepared. A total of 300 questionnaires were sent out to organizations, out of which 130 responses were received—that is a response rate of 43.3%. Approximately 40 responses were incomplete or had missing information about the company they worked for, which made it impossible to link the financial information. Hence, incomplete responses were excluded. Ninety percent of responses were used for this study, as presented in Figures 5 and 6. Out of the 90 responses, 72 responses were used for training and modeling purposes, and 18 responses were kept aside for validation purposes. The following results were obtained from the responses to the survey questionnaire. The total number of variables, including qualitative and quantitative, added up to 32 variables. Since the number of variables was very high, it was imperative to rank and determine the significant factors. To rank the factors, Regression Best Subset Analysis was carried out using Minitab © 17 and Analytic Hierarchy Process (AHP), followed by Hierarchical Fuzzy Expert System modeling using Fuzzy Logic Toolbox Matlab 2015. The model was tested and validated mathematically by Average Validity Percentage (AVP) and Average Invalidity Percentage (AIP). 𝐸𝑛𝑡𝑟𝑜𝑝𝑦 = ln 1𝑝 (1) Figure 4. Financial critical success factors. 4.2. Dependent Variables The dependent variable was the rank of the organization published by ENR Top 400 Contractors and ENR Top 500 Design Firm Sourcebook 2015 [43]. The rank was inversely proportional to the revenue of the organization in the year 2015. The higher the revenue earned, the smaller the rank. To maintain the confidentiality of the organizations, the rank was categorized into categories—i.e., 1–100, 101–200, 201–300, 301–400, 401–500, and 501 and above. This variable was also the output in the model. 5. Data Collection The data collection procedure included a literature review and the identification of potential critical success factors. Eighteen non-financial factors were shortlisted. Then, the questionnaire designed to assess the impact and implementation of these factors in the construction industry was prepared. A total of 300 questionnaires were sent out to organizations, out of which 130 responses were received—that is a response rate of 43.3%. Approximately 40 responses were incomplete or had missing information about the company they worked for, which made it impossible to link the financial information. Hence, incomplete responses were excluded. Ninety percent of responses were used for this study, as presented in Figures 5 and 6. Out of the 90 responses, 72 responses were used for training and modeling purposes, and 18 responses were kept aside for validation purposes. The following results were obtained from the responses to the survey questionnaire. The total number of variables, including qualitative and quantitative, added up to 32 variables. Since the number of variables was very high, it was imperative to rank and determine the significant factors. To rank the factors, Regression Best Subset Analysis was carried out using Minitab © 17 and Analytic Hierarchy Process (AHP), followed by Hierarchical Fuzzy Expert System modeling using Fuzzy Logic Toolbox Matlab 2015. The model was tested and validated mathematically by Average Validity Percentage (AVP) and Average Invalidity Percentage (AIP). Algorithms 2020, 13, 205 11 of 25 Entropy = n∑ i=1 ln 1 pi (1) where pi: Revenue share of the ith segment in total firm revenue, n: Number of Market segments. Algorithms 2020, 13, x FOR PEER REVIEW 11 of 25 where pi: Revenue share of the ith segment in total firm revenue n: Number of Market segments Figure 5. Data response. Figure 6. Layers/Levels in the Analytic Hierarchy Process. 6. Model Building and Analysis The total number of independent variables, both qualitative and quantitative, added up to a total of 32 variables. The number of variables was vast, and hence to perform further analysis, we needed to shortlist the factors that contributed to the performance of an organization. For the initial analysis of data, to check for correlation between variables, regression analysis using Minitab © was carried out. Such results did not indicate a high correlation between factors; instead, the number of variables was vast, and no correlation could be identified, hence making it all the more necessary to reduce the factors. Different methods exist to shortlist the factors. Since we were dealing with 18 qualitative factors, the Analytic Hierarchy Process was a suitable method to determine the significance of factors by comparison with other factors within a category. Figure 5. Data response. Algorithms 2020, 13, x FOR PEER REVIEW 11 of 25 where pi: Revenue share of the ith segment in total firm revenue n: Number of Market segments Figure 5. Data response. Figure 6. Layers/Levels in the Analytic Hierarchy Process. 6. Model Building and Analysis The total number of independent variables, both qualitative and quantitative, added up to a total of 32 variables. The number of variables was vast, and hence to perform further analysis, we needed to shortlist the factors that contributed to the performance of an organization. For the initial analysis of data, to check for correlation between variables, regression analysis using Minitab © was carried out. Such results did not indicate a high correlation between factors; instead, the number of variables was vast, and no correlation could be identified, hence making it all the more necessary to reduce the factors. Different methods exist to shortlist the factors. Since we were dealing with 18 qualitative factors, the Analytic Hierarchy Process was a suitable method to determine the significance of factors by comparison with other factors within a category. Figure 6. Layers/Levels in the Analytic Hierarchy Process. 6. Model Building and Analysis The total number of independent variables, both qualitative and quantitative, added up to a total of 32 variables. The number of variables was vast, and hence to perform further analysis, we needed to shortlist the factors that contributed to the performance of an organization. For the initial analysis of data, to check for correlation between variables, regression analysis using Minitab © was carried out. Such results did not indicate a high correlation between factors; instead, the number of variables was vast, and no correlation could be identified, hence making it all the more necessary to reduce the factors. Different methods exist to shortlist the factors. Since we were dealing with 18 qualitative factors, the Analytic Hierarchy Process was a suitable method to determine the significance of factors by comparison with other factors within a category. Algorithms 2020, 13, 205 12 of 25 6.1. Analytic Hierarchy Process (AHP) The Analytic Hierarchy Process (AHP), developed by Thomas Saaty in 1977, is a multi-criteria decision-making process of qualitative factors when arranged in a hierarchical process [44]. The process allows for making complex decisions by aiding users in organizing information pertaining to thoughts, knowledge, and judgment into a hierarchical framework and quantifying the effect of the qualitative factors by a sequence of pairwise comparison judgments [44]. In this study, AHP was used to evaluate the significance of qualitative factors on organizational performance. The procedure to carry out AHP, as shown in Figure 6, consists of the following steps: 1. First, the hierarchy of factors is established, as well as the selection of criteria. The top-level shows the goal of the problem “Organizational Performance.” The intermediate levels contain the qualitative or non-financial parameters that are categorized into four main categories (i.e., Administrative and Legal, Technical, Management, and Market and Finance). The next level or sub-level includes 18 sub-factors (i.e., Clear Vision; Mission and Goals; Competitive Strategy; Organizational Structure; Political Conditions; Number of Full-Time Employees; Usage of International Aspects (ISO); Availability of knowledge Usage of IT; Business Experience; Product Maintenance; Employee Culture Environment; Employee Compensation and Motivation; Applying Total Quality Management; Training; Quick Liquid Assets; Feedback Evaluation; Research and Development; Market Conditions/Customer Engagement. The layout of the hierarchy helps decision-makers to assess the relationship between factors, as well as showing if the factors have the same magnitude [45]. 2. The second step involves the priority setting of criteria by pairwise comparison matrices for the main factors. Based on the impact rating of the 18 factors, a matrix is assigned with an overall rating for four main factors—i.e., Administrative and Legal, Technical, Management and Market and Finance, and their subfactors. Figure 7 shows the analysis of pairwise comparison matrices for average values of main factors and their sub-factors. 3. The third step is assigning priorities and establishing a pairwise comparison for sub-factors within each main category. This step involves the average values of the sub-factor within one main factor. The AHP methodology applied to these matrices gives the weight factor of each factor (Wi). Table 1 shows the weights of factors. 4. The fourth step is Consistency Analysis. This step verifies the consistency of the pairwise comparison matrix. Weights can be accepted if the matrix is consistent. Therefore, consistency index (CI) and the consistency ratio (CR) are calculated using Equations (2) and (3) [45]. CI = λmax − m m − 1 (2) CR = CI RI (3) where CI is the matrix consistency index, m is matrix size, and λmax is the maximum eigen value. 5. Table 2 shows CI and CR for the main factors. It also shows that CI for main factors is 0.00000149, and CR is 0.00, which is less than 0.10. This means that the primary matrix is consistent, and the weight vectors generated for this matrix are acceptable. 6. The process is repeated for sub-factors. Table 3 shows the weights for the main factors and sub-factors, followed by a calculation of Average Decomposed weight. The decomposed weight is calculated by multiplying the main factor weight by its sub-factor weight. The decomposed weight represents the overall weight of each of the sub-factors [45]. Overall Sub-factor Decomposed Weight is calculate using Equation (4). SDWi j = Wi ∗ ( Vi j ) (4) Algorithms 2020, 13, 205 13 of 25 where Wi is the weight of factor i and Vij is the weight of sub-factor j within the factor i. 7. The graphical presentation in Figure 8 shows the Average decomposed weights. It can be seen that the weight of factors are ranked very carefully. In this case, it becomes essential that an appropriate cutoff weight is chosen to shortlist factors. To select a cutoff weight, we need to consider the mean of the average weights and find the most considerable difference in weights between two factors. The average decomposed weight of all sub-factors is 0.0556. The difference between the weights of sub-factor, Availability of knowledge and Employee Compensation and Motivation is 0.00197. Hence, this is taken as the cutoff weight. Thus, from the AHP method, we shortlisted seven factors—i.e., Market Condition/Customer Engagement, Employee Culture Environment, Clear Vision Mission Goals, Business Experience, Competitive Strategy, Training for Employees, and Availability of Knowledge. To verify the factors, the next step is to verify the results with stepwise regression in Minitab 17. Algorithms 2020, 13, x FOR PEER REVIEW 13 of 25 appropriate cutoff weight is chosen to shortlist factors. To select a cutoff weight, we need to consider the mean of the average weights and find the most considerable difference in weights between two factors. The average decomposed weight of all sub-factors is 0.0556. The difference between the weights of sub-factor, Availability of knowledge and Employee Compensation and Motivation is 0.00197. Hence, this is taken as the cutoff weight. Thus, from the AHP method, we shortlisted seven factors—i.e., Market Condition/Customer Engagement, Employee Culture Environment, Clear Vision Mission Goals, Business Experience, Competitive Strategy, Training for Employees, and Availability of Knowledge. To verify the factors, the next step is to verify the results with stepwise regression in Minitab 17. Figure 7. Pairwise comparison matrix. Table 1. Eigen vector weights (Wi)for the main factors. Factors Weight (Wi) Eigen Vectors C.I. = max − N/N − 1 C.R. = C.I./R.I Figure 7. Pairwise comparison matrix. Algorithms 2020, 13, 205 14 of 25 Table 1. Eigen vector weights (Wi) for the main factors. Factors Weight (Wi) Eigen Vectors C.I. = max − N/N − 1 C.R. = C.I./R.I Administrative and Legal 0.2813 0.00000149 0.00000000 Technical 0.2653 Management 0.2349 Market and finance 0.2185 Table 2. Random Consistency Index. N 1 2 3 4 5 6 7 8 9 10 RI 0 0 0.58 0.9 1.12 1.24 1.32 1.41 1.45 1.49 Table 3. Factor and sub-factor weights. Factors Ave. Weight-Main Factors Ave. Sub-Factors Weight Ave. Decomposed Weight Administrative and Legal 0.2813 Clear Vision Mission and Goals 0.2244 0.0631 Competitive Strategy 0.2176 0.0612 Organization Structure 0.2063 0.0580 Political Conditions 0.1566 0.0441 No. of Full-time Employees 0.1951 0.0549 Technical 0.2653 Usage if International Standards (ISO) 0.1336 0.0355 Availability of Knowledge 0.2284 0.0606 Usage of IT 0.2159 0.0573 Business Experience(no. of years) 0.2338 0.0620 Product Maintenance 0.1883 0.0499 Management 0.2349 Employee Culture Environment 0.2696 0.0633 Employee Compensation and Motivation 0.2495 0.0586 Applying Total Quality Management (TQM) 0.2207 0.0518 Training for employees 0.2601 0.0611 Market and Finance 0.2185 Employee Culture Environment 0.2361 0.0516 Employee Compensation and Motivation 0.2462 0.0538 Applying Total Quality Management (TQM) 0.2140 0.0468 Training for employees 0.3037 0.0663 Algorithms 2020, 13, x FOR PEER REVIEW 15 of 25 Figure 8. Decomposed weights for sub-factors. 6.2. Regression Analysis In this step, the Best Subset Regression function was used where Rsq increases with the number of variables added to the equation; however, Rsq(adj) varies as a peak, and it increases only if the added variable contributes to a better fit of the equation. The best subset reported is the highest Rsq and Rsqadj value. Additionally, Mallow’s coefficient Cp should be equal or close to equal to the number of variables, as shown in Table 4 [45]. For this data, the best subset identified included 20 factors (qualitative and quantitative). After multiple iterations for various subsets shortlisted from AHP, the best subset analysis was carried out. The highest Rsq value achieved was 88.92%, Rsqadj, at 78.97%. Figure 9 shows the seven shortlisted qualitative variables for modeling purposes. There are two main reasons for this. Firstly, the model deals with personal opinions, which are highly qualitative and difficult to model. Secondly, the majority of participants who responded in the survey fall between rank 1 to 200. Therefore, this model can best predict the ranking for organizations that fall between this range. This was the main reason for not utilizing the regression prediction model to assess organizational performance assessment. Table 4. Regression analysis best subset. Source DF Adj SS Adj MS F-Value p-Value Regression 35 6.566 0.187 8.94 0.000 Growth Rate 1 year 1 0.0073 0.0073 0.35 0.557 CAGR 3 years 1 0.0000 0.0000 0.00 0.994 General Building 1 0.0175 0.0175 0.83 0.367 Manufacturing 1 0.3113 0.3113 14.84 0.000 Power 1 0.1274 0.1274 6.07 0.018 WWS 1 0.4173 0.4173 19.89 0.000 Industrial 1 0.1337 0.1337 6.37 0.016 Figure 8. Decomposed weights for sub-factors. Algorithms 2020, 13, 205 15 of 25 6.2. Regression Analysis In this step, the Best Subset Regression function was used where Rsq increases with the number of variables added to the equation; however, Rsq(adj) varies as a peak, and it increases only if the added variable contributes to a better fit of the equation. The best subset reported is the highest Rsq and Rsqadj value. Additionally, Mallow’s coefficient Cp should be equal or close to equal to the number of variables, as shown in Table 4 [45]. Table 4. Regression analysis best subset. Source DF Adj SS Adj MS F-Value p-Value Regression 35 6.566 0.187 8.94 0.000 Growth Rate 1 year 1 0.0073 0.0073 0.35 0.557 CAGR 3 years 1 0.0000 0.0000 0.00 0.994 General Building 1 0.0175 0.0175 0.83 0.367 Manufacturing 1 0.3113 0.3113 14.84 0.000 Power 1 0.1274 0.1274 6.07 0.018 WWS 1 0.4173 0.4173 19.89 0.000 Industrial 1 0.1337 0.1337 6.37 0.016 Transportation 1 0.0140 0.0140 0.67 0.419 Hazardous Waste 1 0.4151 0.4151 19.78 0.000 Telecom 1 0.2815 0.2815 13.42 0.001 % CM at risk 1 0.0373 0.0373 1.78 0.190 Revenue per employee ($MIL) 1 0.2340 0.2340 11.16 0.002 No. of employees 1 0.0137 0.0137 0.66 0.423 Total number of years in business 1 0.0453 0.0453 2.16 0.150 Market diversification entropy 1 0.0850 1.0850 51.71 0.000 Clear vision mission and goals 4 0.4851 0.1212 5.78 0.001 Competitive strategy 2 0.6449 0.3224 15.37 0.000 Availability of knowledge 3 0.0324 0.0108 0.52 0.674 Business experience 3 0.2885 0.0962 4.58 0.008 Employee cultural environment 3 0.3015 0.1005 4.79 0.006 Training for employee 3 0.0575 0.0191 0.91 0.443 Market condition/customer engagement 2 0.1103 0.0552 2.63 0.085 Error 39 0.8183 0.0209 Total 74 7.3846 For this data, the best subset identified included 20 factors (qualitative and quantitative). After multiple iterations for various subsets shortlisted from AHP, the best subset analysis was carried out. The highest Rsq value achieved was 88.92%, Rsqadj, at 78.97%. Figure 9 shows the seven shortlisted qualitative variables for modeling purposes. There are two main reasons for this. Firstly, the model deals with personal opinions, which are highly qualitative and difficult to model. Secondly, the majority of participants who responded in the survey fall between rank 1 to 200. Therefore, this model can best predict the ranking for organizations that fall between this range. This was the main reason for not utilizing the regression prediction model to assess organizational performance assessment. Algorithms 2020, 13, 205 16 of 25 Algorithms 2020, 13, x FOR PEER REVIEW 16 of 25 Transportation 1 0.0140 0.0140 0.67 0.419 Hazardous Waste 1 0.4151 0.4151 19.78 0.000 Telecom 1 0.2815 0.2815 13.42 0.001 % CM at risk 1 0.0373 0.0373 1.78 0.190 Revenue per employee ($MIL) 1 0.2340 0.2340 11.16 0.002 No. of employees 1 0.0137 0.0137 0.66 0.423 Total number of years in business 1 0.0453 0.0453 2.16 0.150 Market diversification entropy 1 0.0850 1.0850 51.71 0.000 Clear vision mission and goals 4 0.4851 0.1212 5.78 0.001 Competitive strategy 2 0.6449 0.3224 15.37 0.000 Availability of knowledge 3 0.0324 0.0108 0.52 0.674 Business experience 3 0.2885 0.0962 4.58 0.008 Employee cultural environment 3 0.3015 0.1005 4.79 0.006 Training for employee 3 0.0575 0.0191 0.91 0.443 Market condition/customer engagement 2 0.1103 0.0552 2.63 0.085 Error 39 0.8183 0.0209 Total 74 7.3846 Figure 9. Shortlisted qualitative variables. 6.3. Fuzzy Logic Modelling Human reasoning, being more approximate than precise, often makes it difficult to measure and determine the measure of factors affecting a particular cause. Fuzzy logic can be used as a tool to understand the imprecision and qualitative aspects of natural language and imprecise cognitive reasoning [35]. Fuzzy logic-based systems are used to analyze and process linguistic inputs to derive outputs or decisions [46]. Matlab R2015a Fuzzy Logic ToolBox software is used to process fuzzy logic inference. 6.3.1. Hierarchical Fuzzy Expert System The hierarchical fuzzy model consists of sub-models, which correspond to the three main categories. In addition to the sub-model, there is one more model that combines the outputs of these sub-models in order to generate the rank of the organization. The fuzzy structure of each of the models is identical. The membership function assigned to input variables and the knowledge-based rules differ. When dealing with a high number of variables, as in this study, the total number of shortlisted independent variables is 20, developing a single layer fuzzy model is not recommended. Figure 9. Shortlisted qualitative variables. 6.3. Fuzzy Logic Modelling Human reasoning, being more approximate than precise, often makes it difficult to measure and determine the measure of factors affecting a particular cause. Fuzzy logic can be used as a tool to understand the imprecision and qualitative aspects of natural language and imprecise cognitive reasoning [35]. Fuzzy logic-based systems are used to analyze and process linguistic inputs to derive outputs or decisions [46]. Matlab R2015a Fuzzy Logic ToolBox software is used to process fuzzy logic inference. 6.3.1. Hierarchical Fuzzy Expert System The hierarchical fuzzy model consists of sub-models, which correspond to the three main categories. In addition to the sub-model, there is one more model that combines the outputs of these sub-models in order to generate the rank of the organization. The fuzzy structure of each of the models is identical. The membership function assigned to input variables and the knowledge-based rules differ. When dealing with a high number of variables, as in this study, the total number of shortlisted independent variables is 20, developing a single layer fuzzy model is not recommended. The qualitative variables were measured on a different scale compared to quantitative variables. For example, the annual growth rate was measure in percentage, and Entropy was measured in fractions or decimal values. Assigning weight to the membership on one layer is extremely difficult. The weight for an individual factor can be used for comparison within the same category. However, a sub-factor from the non-financial category cannot be compared to sub-factors from the financial category. Additionally, for every variable, there should be 7–9 rules. To build a one-layer fuzzy model, the minimum number of rules required is more than 140 rules for 20 variables. The hierarchical model allows us to work with fewer rules. The sub-model for this study will be built using only 7–8 factors [47]. With a usable data set of 75 responses from the survey, the hierarchical system allows us to develop the prediction model. The set of responses required for knowledge-based rule training and testing is satisfied. There are three models—two sub-models for financial and non-financial factors, and the third model for the second layer of the fuzzy model—which combine inputs of financial, non-financial, and market diversification factors, as shown in Figure 10. Since there is only one factor under market diversification, a separate sub-model will not be developed. The steps involved are as follows. Algorithms 2020, 13, 205 17 of 25 Algorithms 2020, 13, x FOR PEER REVIEW 17 of 25 The qualitative variables were measured on a different scale compared to quantitative variables. For example, the annual growth rate was measure in percentage, and Entropy was measured in fractions or decimal values. Assigning weight to the membership on one layer is extremely difficult. The weight for an individual factor can be used for comparison within the same category. However, a sub-factor from the non-financial category cannot be compared to sub-factors from the financial category. Additionally, for every variable, there should be 7–9 rules. To build a one-layer fuzzy model, the minimum number of rules required is more than 140 rules for 20 variables. The hierarchical model allows us to work with fewer rules. The sub-model for this study will be built using only 7–8 factors [47]. With a usable data set of 75 responses from the survey, the hierarchical system allows us to develop the prediction model. The set of responses required for knowledge-based rule training and testing is satisfied. There are three models—two sub-models for financial and non-financial factors, and the third model for the second layer of the fuzzy model—which combine inputs of financial, non-financial, and market diversification factors, as shown in Figure 10. Since there is only one factor under market diversification, a separate sub-model will not be developed. The steps involved are as follows. Figure 10. Overall Hierarchical Fuzzy Model. 6.3.2. Assigning Membership Function The existing literature review shows that different forms of membership function are used depending on the type of problem. The factor’s fuzzy membership is such that the real input can be converted into a fuzzy number value in the range [0,1]. 1. For all the independent variables, the values were normalized so that they could be brought to scale between 0.0 to 1.0. The normalized data were calculated using Equation (5). 𝑧 = (𝑥 − 𝑥 )/(𝑥 − 𝑥 ) (5) 2. The independent variables are assigned Gaussian membership function with a range from 0.0 to 1.0. In this study, we are dealing with expert opinions, and hence instead of giving it a crisp Figure 10. Overall Hierarchical Fuzzy Model. 6.3.2. Assigning Membership Function The existing literature review shows that different forms of membership function are used depending on the type of problem. The factor’s fuzzy membership is such that the real input can be converted into a fuzzy number value in the range [0,1]. 1. For all the independent variables, the values were normalized so that they could be brought to scale between 0.0 to 1.0. The normalized data were calculated using Equation (5). zi = (xi − xmin)/(xmax − xmin) (5) 2. The independent variables are assigned Gaussian membership function with a range from 0.0 to 1.0. In this study, we are dealing with expert opinions, and hence instead of giving it a crisp boundary, it is assigned a waveform membership. The membership function value and the corresponding range are shown in Table 5. 3. The relative weight of each factor at the first level of the hierarchy is determined—the normalized global weights calculated for each main factor and sub-factor. This step was carried out for sub-factors of financial and non-financial categories. Since only one factor falls under market diversification, the weight of only one factor was calculated at the factor level. The normalized global weights allow the sub-factors to be compared to each other. The globalized weight is calculated by multiplying the weight of the individual subfactor to the weight of the main factor [45]. The last column shows the normalized global weights. It shows that the Market condition and Customer Engagement factor is the highest weight at 1.000, closely followed by the Employee Culture Environment, which is second highest at 0.984, as shown in Table 6. 4. The next step involves the development of if–then rules—that is, the impact of sub-factor on the output—as shown in Table 7. Algorithms 2020, 13, 205 18 of 25 5. The independent factors serve as input factors. The normalized input values of the linguistic variables are multiplied by the sub-factor weight to evaluate the equivalent impact, as shown in Table 8. 6. The crisp output value received from the first layer of the hierarchical fuzzy model acts as input for the next layer of the hierarchy model. 7. The second level of the hierarchical fuzzy model is developed using the Fuzzy Logic Toolbox in Matlab R2015. The input variables for the second layer, which are numeric values, are assigned membership functions. 8. The input membership function shape is assigned Gaussian wave membership with five membership functions for each main factor. The range of each membership function is from zero to one, as shown in Figure 11. Table 5. Membership function range value. Range Membership Function 0.0–0.2 Very High 0.2–0.4 High 0.4–0.6 Moderately low 0.6–0.8 Low 0.8–1.0 Very Low Table 6. Factor and sub-factor weights, globalized weights and normalized weights. Main Factors Factor Weight Sub-Factor Sub-Factor Weight Std Dev Global Weight Normalized Global Weights Non-financial 64.84 Clear Vision Mission and Goals 15.21 4.14 986.41 0.904 Competitive Strategy 15.30 4.31 991.92 0.910 Availability of Knowledge 15.21 3.29 986.41 0.904 Business Experience (no. of years) 8.35 4.93 541.60 0.475 Employee Culture Environment 16.49 4.19 1069.07 0.984 Training for employees 12.69 4.68 822.92 0.747 Market Conditions/Customer Eng 16.74 3.79 1085.60 1.000 Financial 30.10 Growth Rate 1 year 2 9.37 5.22 282.11 0.225 CAGR 3 year 3 18.37 9.34 552.88 0.486 General Building 4 28.54 18.35 858.96 0.781 Manufacturing 5 2.90 6.44 87.33 0.037 Power 6 4.21 8.95 126.62 0.075 WWS 7 2.16 5.46 64.90 0.016 Industrial/Petroleum 8 7.97 17.28 239.97 0.184 Transportation 9 3.13 6.54 94.32 0.044 Hazardous Waste 10 1.62 6.31 48.79 0.000 Telecom 2.33 5.93 70.13 0.021 % CM at Risk 12 10.77 13.59 324.10 0.266 Revenue per employee 8.63 7.13 259.90 0.204 Market 5.06 Diversification Entropy 100 0 506.07 0.441 Table 7. If-then rules for sub-factors. Factor Impact on Organizational Performance If Clear Vision Mission and Goals is Very High then Impact of non-financial factor is Very High If Clear Vision Mission and Goals is High then Impact of non-financial factor is High If Clear Vision Mission and Goals is Moderately low then Impact of non-financial factor is Moderately low If Clear Vision Mission and Goals is Low then Impact of non-financial factor is Low If Clear Vision Mission and Goals is Very Low then Impact of non-financial factor is Very Low Algorithms 2020, 13, 205 19 of 25 Table 8. Sub factor weights and equivalent impact. Rule No. Clear Vision Competitive Strategy Availability of Knowledge Business Exp Employee Culture Training Employees Market Conds/ Customer Eng Equivalent Impact Combined Impact 15.21 15.30 15.21 8.35 16.49 12.69 16.74 1 0.75 0.50 0.75 0.50 0.75 0.33 0.75 6.38 High 2 0.75 0.50 0.75 0.04 1.00 0.67 1.00 7.25 High 3 0.75 0.75 1.00 0.56 1.00 1.00 1.00 8.87 Very high 4 0.75 0.50 0.75 0.51 1.00 0.67 1.00 7.64 High 5 0.75 1.00 0.75 0.10 0.75 0.67 0.75 7.24 High 6 1.00 0.75 1.00 0.10 1.00 0.67 1.00 8.44 Very high 7 1.00 1.00 0.75 0.43 0.75 0.67 1.00 8.31 Very high 8 1.00 1.00 1.00 0.43 0.50 1.00 1.00 8.70 Very high 9 0.50 0.50 0.50 0.28 0.50 0.33 0.50 4.61 Moderate 10 1.00 1.00 1.00 0.28 1.00 1.00 1.00 9.39 Very high 11 0.75 0.75 0.50 0.28 0.50 0.33 1.00 6.20 High 12 1.00 0.75 0.50 0.11 1.00 0.33 1.00 7.27 High 13 0.25 1.00 0.50 0.04 0.25 0.33 0.75 4.80 Moderate 14 0.50 0.50 1.00 0.46 0.50 0.67 0.75 6.36 High 15 1.00 0.75 1.00 0.50 1.00 1.00 0.75 8.78 Very high Algorithms 2020, 13, x FOR PEER REVIEW 19 of 25 Table 7. If-then rules for sub-factors. Factor Impact on Organizational Performance If Clear Vision Mission and Goals is Very High then Impact of non-financial factor is Very High If Clear Vision Mission and Goals is High then Impact of non-financial factor is High If Clear Vision Mission and Goals is Moderately low then Impact of non-financial factor is Moderately low If Clear Vision Mission and Goals is Low then Impact of non-financial factor is Low If Clear Vision Mission and Goals is Very Low then Impact of non-financial factor is Very Low Table 8. Sub factor weights and equivalent impact. Rule No. Clear Vision Competitive Strategy Availability of Knowledge Business Exp Employee Culture Training employees Market Conds/ Customer Eng Equivalent Impact Combined Impact 15.21 15.30 15.21 8.35 16.49 12.69 16.74 1 0.75 0.50 0.75 0.50 0.75 0.33 0.75 6.38 High 2 0.75 0.50 0.75 0.04 1.00 0.67 1.00 7.25 High 3 0.75 0.75 1.00 0.56 1.00 1.00 1.00 8.87 Very high 4 0.75 0.50 0.75 0.51 1.00 0.67 1.00 7.64 High 5 0.75 1.00 0.75 0.10 0.75 0.67 0.75 7.24 High 6 1.00 0.75 1.00 0.10 1.00 0.67 1.00 8.44 Very high 7 1.00 1.00 0.75 0.43 0.75 0.67 1.00 8.31 Very high 8 1.00 1.00 1.00 0.43 0.50 1.00 1.00 8.70 Very high 9 0.50 0.50 0.50 0.28 0.50 0.33 0.50 4.61 Moderate 10 1.00 1.00 1.00 0.28 1.00 1.00 1.00 9.39 Very high 11 0.75 0.75 0.50 0.28 0.50 0.33 1.00 6.20 High 12 1.00 0.75 0.50 0.11 1.00 0.33 1.00 7.27 High 13 0.25 1.00 0.50 0.04 0.25 0.33 0.75 4.80 Moderate 14 0.50 0.50 1.00 0.46 0.50 0.67 0.75 6.36 High 15 1.00 0.75 1.00 0.50 1.00 1.00 0.75 8.78 Very high Figure 11. Membership functions assigned to the input variable. 6.3.3. Fuzzy Inference The Mamdani type fuzzy model is selected in the fuzzy logic toolbox of Matlab R2015 with the maximum and minimum rule. The Mamdani is easier to understand and work with the consequent Figure 11. Membership functions assigned to the input variable. 6.3.3. Fuzzy Inference The Mamdani type fuzzy model is selected in the fuzzy logic toolbox of Matlab R2015 with the maximum and minimum rule. The Mamdani is easier to understand and work with the consequent of the systems. The Mamdani method uses a simple structure of the Minimum operator, as shown in Equation (6). R j = I f x1 is A j 1 and x2 is A j 2 and x3 is A j 3 . . . . . . ..and xn is A j n y is B j (6) where Rj is the j-th rule, A j i (j = 1,2, N, i = 1, 2, n), and B j is the fuzzy subsets of the inputs and outputs, respectively. The minimum operator is used to calculate the firing strength of each fuzzy rule. The firing strength is directly proportional to the impact on the output. The rules are setup up in the if–then format. Algorithms 2020, 13, 205 20 of 25 6.3.4. Defuzzification There are multiple defuzzification methods. In this study, the output factor is assigned a triangular shape with nine membership functions. The reason for the nine membership function is to increase the sensitivity and accuracy of the model, as shown in Figure 12. Output membership function maps the height corresponding to the firing strength of rules [48]. In this layer of fuzzy process, the output values are obtained through the defuzzification of the value of the centroid of the triangular membership assigned to the output variable. This numeric output is the predicted rank of the organization. Algorithms 2020, 13, x FOR PEER REVIEW 20 of 25 of the systems. The Mamdani method uses a simple structure of the Minimum operator, as shown in Equation (6). 𝑅 = 𝐼𝑓 𝑥 𝑖𝑠 𝐴 𝑎𝑛𝑑 𝑥 𝑖𝑠 𝐴 𝑎𝑛𝑑 𝑥 𝑖𝑠 𝐴 … … . . 𝑎𝑛𝑑 𝑥 𝑖𝑠 𝐴 y is 𝐵 (6) where Rj is the j-th rule, 𝐴 (j = 1,2, N, i = 1, 2, n), and Bj is the fuzzy subsets of the inputs and outputs, respectively. The minimum operator is used to calculate the firing strength of each fuzzy rule. The firing strength is directly proportional to the impact on the output. The rules are setup up in the if–then format. 6.3.4. Defuzzification There are multiple defuzzification methods. In this study, the output factor is assigned a triangular shape with nine membership functions. The reason for the nine membership function is to increase the sensitivity and accuracy of the model, as shown in Figure 12. Output membership function maps the height corresponding to the firing strength of rules [48]. In this layer of fuzzy process, the output values are obtained through the defuzzification of the value of the centroid of the triangular membership assigned to the output variable. This numeric output is the predicted rank of the organization. Figure 12. Membership function assigned to the Output variable. 6.4. Verification of the Developed Model To verify the developed model, 20% of the data set was left for testing purposes. To determine whether the model is verified or not when using results comparison as in this study, two terms can be used to determine the validity of the model, Average Validity Percent (AIP), and Average Invalidity Percent (AIP). The AVP represents the validation percent out of 100, and AIP represents the prediction error [49]. These two terms are shown in Equations (7) and (8). Figure 12. Membership function assigned to the Output variable. 6.4. Verification of the Developed Model To verify the developed model, 20% of the data set was left for testing purposes. To determine whether the model is verified or not when using results comparison as in this study, two terms can be used to determine the validity of the model, Average Validity Percent (AIP), and Average Invalidity Percent (AIP). The AVP represents the validation percent out of 100, and AIP represents the prediction error [49]. These two terms are shown in Equations (7) and (8). AIP =  n∑ i=1 ∣∣∣∣∣∣1 − ( Ei Ci )∣∣∣∣∣∣ ∗ 100/n (7) AVP = 100 − AIP (8) where AIP is Average Invalidity Percent, AVP is Average Validity Percent, Ei is Estimated value, and Ci is Actual value. Table 9 shows the validation data with a comparison of categories of actual organization rank and the predicted category of organization rank. The Average Invalidity Percentage is 36.667%. Therefore, the average validity percentage is 64.334%. Thus the model is stable. Algorithms 2020, 13, 205 21 of 25 Table 9. Validation results. Rank Category (C) Predicted Rank Category (E) ABS(1-(E/C)) 1 1 0.00 2 1 0.50 4 1 0.75 5 5 0.00 1 1 0.00 2 2 0.00 3 1 0.67 3 1 0.67 2 5 1.50 3 1 0.67 1 1 0.00 3 5 0.67 1 1 0.00 1 1 0.00 2 2 0.00 AIP% 36.11 7. Conclusion, Limitations, and Future Research Work 7.1. Summary The performance of a construction organization depends on several success factors. Only a few research studies have focused on identifying and measuring the non-financial performance of the construction organization. This paper represents the development of a comprehensive framework to assess the performance of construction organizations based on 20 factors categorized into financial, non-financial, and market diversification, using a hierarchical fuzzy approach. The seven qualitative or non-financial critical success factors (i.e., Clear Vision, Mission, and Goals, Competition Strategy, Availability of Knowledge, Business Experience, Employee Culture Environment, Market Condition, and Customer Engagement) were selected using the Analytic Hierarchy Process (AHP) ranking system. In order to assess the combined impact of factors on the overall organization performance, they were modeled using a hierarchical fuzzy expert system. Fuzzy inputs and outputs and the rules governing them are designed using the data responses collected from industry professionals. The study is a step towards understanding a detailed analysis of factors that may impact overall performance. The developed models, methodology, and approach can help the construction companies to: • Determine the critical success factors that they need to improve to leverage the company performance. • Evaluate company performance rather than just project performance. • Predict company performance. 7.2. Conclusion During this research, multiple points of observation and concern have been concluded from data response and analysis, such as: • The industry professionals rate the impact of Customer Engagement and Employee Cultural Environment as very high. • The ranking of the impact of critical success factors, such as Product Maintenance, Research, and Development, Usage of International Standards, and Political conditions, is very low. The average decomposed weight from the AHP process shows that construction industry rates impact investment in the Research and Development of in-house technologies to a low and medium extent. They mainly rely on products from computing firms. Algorithms 2020, 13, 205 22 of 25 • The political condition in the USA is more conducive. In a previous study, in the survey which was sent out to middle eastern countries, the impact of political conditions factor was rated as high and very high [42]. • The developed Hierarchical Fuzzy Model is based on the expert opinions of industry professionals. It is necessary to establish specific parameters or rubric, following which they answer the questionnaire. Personal bias leads to extensive outliers, thus making the model inefficient. • It is also observed that the majority of participants have rated the employee culture environment of their respective organizations as high and very high. However, the rating for employee compensation is low. This either represents that the employees of these organizations are not paid well, or the participants favor a better employee culture environment over employee compensation. This represents a personal bias in the responses. 7.3. Limitations The developed model uses AHP to shortlist qualitative factors and hierarchical fuzzy expert system techniques to assess organizational performance. There are some inherent limitations in this model, such as: • The critical success factors have been shortlisted from the literature review and the impact rating from participants. There is a possibility that a factor could be left out if the frequency of appearance is low. • Furthermore, based on geographical conditions and socio-economic conditions, the model can only predict the ranking of organizations that fall in that geographic location. • The total number of collected questionnaires is 130. The model accuracy can be improved by increasing the number of participant responses that are used to develop knowledge-based rules of the fuzzy expert system, as well as considering companies that are widespread across the range of rankings. • The majority of the construction firms from which responses were received were not publicly listed. This limited the number of key financial performance factors, indicating factors that could be included in the model. • The first layer structure is not one of the types of Hierarchical Structures in MATLAB, so due to technical constraints, only the second layer of the hierarchical fuzzy expert model could be developed in Matlab R2015. The first layer was manually calculated to compute the combined impact of sub-factors in the sub-model. • The effect of choosing different methods of fuzzification and defuzzification methods and membership functions type. • Furthermore, the model can only predict the category of the rank, not the exact rank. To be able to make the model more accurate, data over a wide range of rank need to be collected. 7.4. Recommendation for Future Research Work The developed research/model helps both researchers and practitioners to predict accurate company performance. Some of the recommendations and future works that can enhance the model and the research, in general, are listed below: • Develop rubrics for survey participants to rate their opinion without bias. • Collect quantitative information to validate the responses to qualitative questions. • Include more quantitative financial factors to improve the accuracy of the model by assessing the overall performance. • Reduce personal bias; multiple participants from the same organizations should participate in the survey. • Use the sensitivity analysis to evaluate the effect of the omitted factors in the shortlisting process. Algorithms 2020, 13, 205 23 of 25 • The study shows a need for the further investigation of critical success factors to select the optimum number and nature for modeling the organization’s performance. • There is a need for considering the self-confidence of decision-makers through the social network [50,51]. • The results of this research will lead to a new generation of specific and accurate company performance models and fully automated models/systems that might partially replace expert opinion techniques. Author Contributions: Writing—original draft preparation, Z.R.; writing—review and editing, Z.R. and E.E. All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. Conflicts of Interest: The authors declare no conflict of interest. References 1. Ozorhon, B. Analysis of construction innovation process at project level. J. Manag. Eng. 2013, 29, 455–463. [CrossRef] 2. Veshosky, D. Managing innovation information in engineering and construction firms. J. Manag. Eng. 1998, 14, 58–66. [CrossRef] 3. Kim, H.J.; Reinschmidt, K.F. Market structure and organizational performance of construction organizations. J. Manag. Eng. 2012, 28, 212–220. [CrossRef] 4. Pinto, J.K.; Covin, J.G. Critical factors in project implementation: A comparison of construction and R&D projects. Technovation 1989, 9, 49–62. 5. Müller, R.; Jugdev, K. Emerald Article: Critical success factors in projects: Pinto, Slevin, and Prescott-the elucidation of project success. Int. J. Manag. Proj. Bus. 2012, 5, 757–775. [CrossRef] 6. Liu, H.; Wang, M.J.; Skibniewski, M.J.; He, J.S.; Zhang, Z.S. Identification of critical success factors for construction innovation: From the perspective of strategic cooperation. Front. Eng. Manag. 2014, 1, 202–209. [CrossRef] 7. Rathore, Z.; Elwakil, E. A Framework of Organization Performance Assessment in the Construction Industry Using Fuzzy Approach. In Proceedings of the Canadian Society for Civil Engineering’s 5th International/11th Construction Specialty Conference, Vancouver, BC, Canada, 8 June 2015. 8. Elwakil, E. Integrating analytical hierarchy process and regression for assessing construction organizations’ performance. Int. J. Constr. Manag. 2017, 17, 76–88. [CrossRef] 9. Radwan, A.; Elwakil, E. Functional units based model for construction organizations performance. Int. J. Archit. Eng. Constr. 2015, 4, 194–203. [CrossRef] 10. Amarkhil, Q.; Elwakil, E. Construction organization success strategy in post-conflict environment. Int. J. Constr. Manag. 2019, 23, 1–10. [CrossRef] 11. Isik, Z.; Arditi, D.; Dikmen, I.; Birgonul, M.T. Impact of resources and strategies on construction company performance. J. Manag. Eng. 2010, 26, 9–18. [CrossRef] 12. Cooke-Davies, T. The “real” success factors on projects. Int. J. Proj. Manag. 2002, 20, 185–190. [CrossRef] 13. Chinowsky, P.S.; Meredith, J.E. Strategic management in construction. J. Constr. Eng. Manag. 2000, 126, 1–9. [CrossRef] 14. Bontis, N.; Dragonetti, N.C.; Jacobsen, K.; Roos, G. The knowledge toolbox: A review of the tools available to measure and manage intangible resources. Eur. Manag. J. 1999, 17, 391–402. [CrossRef] 15. Arthur, M.B. The boundaryless career: A new perspective for organizational inquiry. J. Organ. Behav. 1994, 15, 295–306. [CrossRef] 16. Sims, H.P.; Gioia, D.A. The Thinking Organization; Jossey-Bass Inc. Pub: San Francisco, CA, USA, 1986. 17. Weick, K.E. Organizational culture as a source of high reliability. Calif. Manag. Rev. 1987, 29, 112–127. [CrossRef] 18. Hauser, J.; Katz, G. Metrics: You are what you measure! Eur. Manag. J. 1998, 16, 517–528. [CrossRef] 19. Viner, J. Cost curves and supply curves. Zeitschrift Für Nationalökonomie 1932, 3, 23–46. [CrossRef] 20. Simon, H.A.; Bonini, C.P. The size distribution of business firms. Am. Econ. Rev. 1958, 48, 607–617. http://dx.doi.org/10.1061/(ASCE)ME.1943-5479.0000157 http://dx.doi.org/10.1061/(ASCE)0742-597X(1998)14:1(58) http://dx.doi.org/10.1061/(ASCE)ME.1943-5479.0000082 http://dx.doi.org/10.1108/17538371211269040 http://dx.doi.org/10.15302/J-FEM-2014027 http://dx.doi.org/10.1080/15623599.2016.1187247 http://dx.doi.org/10.7492/IJAEC.2015.020 http://dx.doi.org/10.1080/15623599.2019.1644761 http://dx.doi.org/10.1061/(ASCE)0742-597X(2010)26:1(9) http://dx.doi.org/10.1016/S0263-7863(01)00067-9 http://dx.doi.org/10.1061/(ASCE)0733-9364(2000)126:1(1) http://dx.doi.org/10.1016/S0263-2373(99)00019-5 http://dx.doi.org/10.1002/job.4030150402 http://dx.doi.org/10.2307/41165243 http://dx.doi.org/10.1016/S0263-2373(98)00029-2 http://dx.doi.org/10.1007/BF01316299 Algorithms 2020, 13, 205 24 of 25 21. Spendolini, M.J. The benchmarking processes. Compens. Benefits Rev. 1992, 24, 21–29. [CrossRef] 22. Tang, Y.H.; Ogunlana, S.O. Modelling the dynamic performance of a construction organization. Constr. Manag. Econ. 2003, 21, 127–136. [CrossRef] 23. Zayed, T.; Elwakil, E.; Ammar, M. A framework for performance assessment of organizations in construction industry. Int. J. Archit. Eng. Constr. 2012, 1, 199–212. [CrossRef] 24. Gustavsson, T.K.; Gohary, H. Boundary action in construction projects: New collaborative project practices. Int. J. Manag. Proj. Bus. 2012, 5, 364–376. [CrossRef] 25. Bititci, U.; Muir, D. Integrated performance measurement systems: A development guide. Int. J. Oper. Prod. Manag. 1997, 17, 522–556. [CrossRef] 26. Vogl, B.; Abdel-Wahab, M. Measuring the construction industry’s productivity performance: Critique of international productivity comparisons at industry level. J. Constr. Eng. Manag. 2015, 141, 04014085. [CrossRef] 27. Fisher, D.; Miertschin, S.; Pollock Jr, D.R. Benchmarking in construction industry. J. Manag. Eng. 1995, 11, 50–57. [CrossRef] 28. Pinto, J.K.; Slevin, D.P. Critical factors in successful project implementation. IEEE Trans. Eng. Manag. 1987, 1, 22–27. [CrossRef] 29. Pinto, J.K.; Mantel, S.J. The causes of project failure. IEEE Trans. Eng. Manag. 1990, 37, 269–276. [CrossRef] 30. Neely, A.D.; Adams, C.; Kennerley, M. The Performance Prism: The Scorecard for Measuring and Managing Business Success; FT Prentice Hall: London, UK, 2002. 31. Harrison, P. Can Measurement Error Explain the Weakness of Productivity Growth in the Canadian Construction Industry? 2007. Available online: https://core.ac.uk/download/pdf/7033340.pdf (accessed on 14 July 2018). 32. Costa, D.B.; Formoso, C.T.; Kagioglou, M.; Alarcón, L.F.; Caldas, C.H. Benchmarking initiatives in the construction industry: Lessons learned and improvement opportunities. J. Manag. Eng. 2006, 22, 158–167. [CrossRef] 33. Neely, A.; Gregory, M.; Platts, K. Performance measurement system design: A literature review and research agenda. Int. J. Oper. Prod. Manag. 1995, 15, 80–117. [CrossRef] 34. Choi, J.; Russell, J.S. Long-term entropy and profitability change of United States public construction firms. J. Manag. Eng. 2005, 21, 17–26. [CrossRef] 35. Zadeh, L.A. Fuzzy logic. Computer 1988, 21, 83–93. [CrossRef] 36. Chan, A.P.; Chan, D.W.; Yeung, J.F. Overview of the application of “fuzzy techniques” in construction management research. J. Constr. Eng. Manag. 2009, 135, 1241–1252. [CrossRef] 37. Nguyen, T.A.; Perkins, W.A.; Laffey, T.J.; Pecora, D. Checking an Expert Systems Knowledge Base for Consistency and Completeness. 1985. Available online: https://www.ijcai.org/Proceedings/85-1/Papers/070. pdf (accessed on 19 August 2018). 38. Fayek, A.R.; Oduba, A. Predicting industrial construction labor productivity using fuzzy expert systems. J. Constr. Eng. Manag. 2005, 131, 938–941. [CrossRef] 39. Sargent, D.J. Comparison of artificial neural networks with other statistical approaches: Results from medical data sets. Cancer Interdiscip. Int. J. Am. Cancer Soc. 2001, 91, 1636–1642. [CrossRef] 40. Nasir, H.; Haas, C.T.; Rankin, J.H.; Fayek, A.R.; Forgues, D.; Ruwanpura, J. Development and implementation of a benchmarking and metrics program for construction performance and productivity improvement. Can. J. Civ. Eng. 2012, 39, 957–967. [CrossRef] 41. Rathore, Z. A Framework for Organizational Performance Assessment in the Construction Industry. Master’s Thesis, Purdue University, West Lafayette, IN, USA, 2016. 42. Elwakil, E.; Ammar, M.; Zayed, T.; Mahmoud, M.; Eweda, A.; Mashhour, I. Investigation and modeling of critical success factors in construction organizations. In Proceedings of the Construction Research Congress 2009, Seattle, WA, USA, 5–7 April 2009; pp. 350–359. 43. Engineering News Record. ENR Top 400 Contractors and ENR Top 500 Design Firm Sourcebook 2015. Available online: https://www.enr.com/toplists/Sourcebook (accessed on 26 August 2018). 44. Saaty, T.L. How to make a decision: The analytic hierarchy process. Eur. J. Oper. Res. 1990, 48, 9–26. [CrossRef] 45. Fares, H.; Zayed, T. Hierarchical fuzzy expert system for risk of failure of water mains. J. Pipeline Syst. Eng. Pract. 2010, 1, 53–62. [CrossRef] http://dx.doi.org/10.1177/088636879202400505 http://dx.doi.org/10.1080/0144619032000079699 http://dx.doi.org/10.7492/IJAEC.2012.022 http://dx.doi.org/10.1108/17538371211235272 http://dx.doi.org/10.1108/01443579710167230 http://dx.doi.org/10.1061/(ASCE)CO.1943-7862.0000944 http://dx.doi.org/10.1061/(ASCE)0742-597X(1995)11:1(50) http://dx.doi.org/10.1109/TEM.1987.6498856 http://dx.doi.org/10.1109/17.62322 https://core.ac.uk/download/pdf/7033340.pdf http://dx.doi.org/10.1061/(ASCE)0742-597X(2006)22:4(158) http://dx.doi.org/10.1108/01443579510083622 http://dx.doi.org/10.1061/(ASCE)0742-597X(2005)21:1(17) http://dx.doi.org/10.1109/2.53 http://dx.doi.org/10.1061/(ASCE)CO.1943-7862.0000099 https://www.ijcai.org/Proceedings/85-1/Papers/070.pdf https://www.ijcai.org/Proceedings/85-1/Papers/070.pdf http://dx.doi.org/10.1061/(ASCE)0733-9364(2005)131:8(938) http://dx.doi.org/10.1002/1097-0142(20010415)91:8+&lt;1636::AID-CNCR1176&gt;3.0.CO;2-D http://dx.doi.org/10.1139/l2012-030 https://www.enr.com/toplists/Sourcebook http://dx.doi.org/10.1016/0377-2217(90)90057-I http://dx.doi.org/10.1061/(ASCE)PS.1949-1204.0000037 Algorithms 2020, 13, 205 25 of 25 46. Elwakil, E.; Zayed, T. Construction productivity fuzzy knowledge base management system. Can. J. Civil Eng. 2018, 45, 329–338. [CrossRef] 47. Fayek, A.R.; Tsehayae, A.A. Modeling Construction Labor Productivity Using Fuzzy Logic and Exploring the Use of Fuzzy Hybrid Techniques. In Proceedings of the IEEE 2012 Annual Meeting of the North American Fuzzy Information Processing Society (NAFIPS), Berkeley, CA, USA, 6–8 August 2012; pp. 1–6. 48. Chao, L.C.; Skibniewski, M.J. Fuzzy logic for evaluating alternative construction technology. J. Constr. Eng. Manag. 1998, 124, 297–304. [CrossRef] 49. Dawood, T.; Elwakil, E.; Novoa, H.M.; Delgado, J.F. Pressure Data-Driven Model for Failure Prediction of PVC Pipelines. Eng. Fail. Anal. 2020, 28, 104769. [CrossRef] 50. Zhang, Z.; Kou, X.; Yu, W.; Gao, Y. Consistency improvement for fuzzy preference relations with self-confidence: An application in two-sided matching decision making. J. Oper. Res. Soc. 2020, 22, 1–4. [CrossRef] 51. Zhang, Z.; Gao, Y.; Li, Z. Consensus reaching for social network group decision making by considering leadership and bounded confidence. Knowl. Based Syst. 2020, 10, 106240. [CrossRef] © 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/). http://dx.doi.org/10.1139/cjce-2017-0540 http://dx.doi.org/10.1061/(ASCE)0733-9364(1998)124:4(297) http://dx.doi.org/10.1016/j.engfailanal.2020.104769 http://dx.doi.org/10.1080/01605682.2020.1748529 http://dx.doi.org/10.1016/j.knosys.2020.106240 http://creativecommons.org/ http://creativecommons.org/licenses/by/4.0/. Introduction Background Critical Success Factors Existing Performance Metrics Modelling Techniques Adopted—Fuzzy Approach Challenges and Limitations of Existing Metrics Framework and Methodology Study Variables Independent Variables Dependent Variables Data Collection Model Building and Analysis Analytic Hierarchy Process (AHP) Regression Analysis Fuzzy Logic Modelling Hierarchical Fuzzy Expert System Assigning Membership Function Fuzzy Inference Defuzzification Verification of the Developed Model Conclusion, Limitations, and Future Research Work Summary Conclusion Limitations Recommendation for Future Research Work References 
work_ywo3iloiurdnnm67bdfbukggzi	----	University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 72 Expert system to predict the numbers of livestock in Thi-Qar province https://doi.org/10.32792/utq/utj/vol12/3/6 Shaker K.Ali Atyaf J.Yaseen Amer Y. Mahdi Computer Department, Thi_Qar University Email: Shaker@utq.edu.iq Email: atyaf_jarallah@yahoo.com Abstract Predict the numbers of livestock is one of interesting research area cause the important of livestock in each countries economics, in this paper we design an expert system for predict the numbers of livestock (Sheep, Goat, Cow, Buffalo and Camel) in Thi-Qar province, and we collect the knowledge from (Directorate of meteorological in Nasiriyah, the Veterinary Hospital in Nasiriyah and Directorate of Agriculture in the province of Thi-Qar) and collect the last 5 years to compare the result from our system with actual results by compare year by year in actual world by the same years in our system , There are 10 knowledge which build rules , There are differences numbers of rule for each livestock kinds, there are 813 fuzzy rules for sheep, 558 fuzzy rules for goat, 435 fuzzy rules for cow, 425 fuzzy rules for buffalo and 537 fuzzy rules for camel. There is commands knowledge and there is specified knowledge for each livestock (Sheep, Goat, Cow, Buffalo and Camel), there rules build by using Mamdani fuzzy logic and we use SQL server 2008 to insert these fuzzy rules. The results which we obtained from our fuzzy expert system was gave high accuracy and perfect result. Keywords: expert system, fuzzy logic, Mamdani, livestock. 1. Introduction Livestock represents one of advanced center in developing countries economics, in Iraq, it's considered as national wealth that contributes in supporting the food security abilities. [1] The trend towards the development of the livestock and increase its production necessarily requires a lot of https://jutq.utq.edu.iq/index.php/main https://doi.org/10.32792/utq/utj/vol12/3/6 mailto:Shaker@utq.edu.iq University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 73 important supplies and that depends on expecting the numbers of livestock in the next years which many important decisions will entail on it such as providing feed and adequate veterinary services, for that it's important to find a means based on existing data regardless of its negative or positive trend in order to make use of these data to predict the numbers of animals, and (the expert systems) are so important in this case.[2][1] The use of fuzzy expert systems is more appropriate to problems where specialized expertise happen for a particular problem. Commonly, human experts tend to specialize in rather narrow problem solving areas or tasks. These human experts tend to solve problems fairly accurately and quickly, explain what they do and judge the consistency of their own conclusions. They know when mistakes have been made, and communicate with other experts. Knowledge is a major resource in the improvement of fuzzy expert systems and it often lies with merely a few experts [3]. Using the fuzzy expert systems in the Agricultural field has many benefits. It may provide a fast, consistent and objective results and more accurate, reduces labor costs by eliminating labor-severe tasks, can help to support and verify human expert’s opinion and may help when expert advice is required but an expert may not be found. It can work in dangerous situations and under conditions unfit for human experts. Agricultural extension services have been forced to cover larger areas with less manpower [4]. They need to identify economical, fast, and reliable methods to transmit information and advice to clientele who can access of the internet, computers can transmit a decision-support programs and variety of educational with the click of a mouse; simulation models, multimedia programs, conventional programs, and ES are available to clientele. These programs can advise, simulate, teach, and predict results of decisions made on a farm. Expert system technology can do all of the above; thus an ES may be capable of educating while advice clientele on how to solve problems at the same time. 2. Architecture of expert system An expert system consists of three main parts: A user interface, an inference engine and a knowledge base 2.1User interface It is the process by which the expert system interacts with a user. User interface may be through forms, command prompts, dialog boxes, or other input methods. Some expert systems do not interact directly with a human but interact with other computer applications. The expert system, in these cases, will not have a user interface but will have an interaction mechanism for transactions with the other application. [5],[ 6] https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 74 https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 75 2.2 An inference engine It is a brain of expert system that provides methodology for reasoning and uses the rule interpreter. It acts as an interpreter which processes and analyzes the rules. The main task of inference engine is to trace its way through a set of rules to reach at a conclusion. [7] 2.3 A knowledge base It is a collection of rules and facts where the knowledge base is made from information introduced by human experts. The knowledge base is flexible to accept changes (new rules perhaps will be added and existing ones maybe edited or deleted from the knowledge base with the use of the knowledge acquisition module) without affecting the whole system because it is independent from all other components of an expert system [8]. 3. Fuzzy logic Fuzzy logic which introduced by Zadeh in 1965 is defined as a set of mathematical ideas for knowledge representation based on degrees of membership rather than on crisp membership of classical theory of logic. [9] Unlike binary Boolean logic, Fuzzy logic is multi-valued. It deals with degrees of membership and degrees of truth. It uses the continuum of logical values between 0 (completely false) and 1 (completely true). Instead of black and white only, it employs the spectrum of colors, accepting that things can be partially false and partially true at the same time. We can see that in Figure (1), fuzzy logic adds a range of logical values to Boolean logic. Now classical theory of logic can be considered as a special case of multi-valued fuzzy logic as can be seen in Figure (2). [9][10] Figure (1) Range of logical values in Boolean and fuzzy logic: (a) Boolean logic; (b) multi-valued logic Figure (2)"The classical set theory is a subset of the theory of fuzzy sets"[ ] Fuzzy set Crisp set https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 76 Fuzzy logic technique can be implemented to real application needs the following three steps: Fuzzification, Fuzzy Inference Process and Defuzzification. [11] 4. The proposed method The proposed method consist from five steps as shown in figure 3 Figure.3 flowchart of proposed method Choose one kind of livestock kinds and entry the special system for this kind Enter the specific knowledge of livestock Matching data with the knowledge base (the rules) by using Mamdani Predicting of livestock Defining a set of rules Sheep Camel Buffalo Cow Goat https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 77 4.1 Choose one kind of livestock kinds In this step, we shall choose only one kind of livestock (Sheep, goats, cows, buffalos, camels) and entry the special system for this kind. 4.2 Enter the data for livestock We got the knowledge used in this paper from experts and competent authority such as Directorate of meteorological in Nasiriyah, the Veterinary Hospital in Nasiriyah and Directorate of Agriculture in the province of Thi- Qar. Data, which obtained from the experts and from the competent authority, are made up of main knowledge and are as follows: 1. Meteorological 2. Pastures 3. Butchery 4. Livestock breeder 5. Veterinary services 6. Smuggling 7. Governmental role 8. Scientific personnel and their role 9. Diseases 10. The impact of wars on the region The above main-knowledge has much sub- knowledge, which has more details. In the Tables (1-10), we will explain the main-knowledge and the sub-knowledge that branched from them. As example, Table of Meteorological Table (1). Description of the Sub-knowledge of Meteorological Sub-knowledge range livestock Low Mid Max Temperature(C) [-2,12] [8,45] [40,60] Sheep, goat, cow, buffalo, camel Rain(mm) [0,35] [22,55] [43,75] Sheep, goat, cow, buffalo, camel The proportion of the existence of the floods (%) [0,50] [25,100] Sheep, goat, camel The proportion of the presence of the marsh (%) [0,50] [25,100] buffalo https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 78 We formalize the parameters for every kind of knowledge we must also explain the system output ranges for livestock kinds determine the range of any livestock is (low, moderate and high) as shown in the Table (2) Table (2). Description of the ranges of system output Parameter Range Livestock Low Mid Max System output [0,49] [40,70] [65,100] Sheep, goat, cow, buffalo, camel 4.3 Defining a set of rules Here we formalize all sub-knowledge mentioned by using decision tables. The sub-knowledge for each kind of livestock kinds will be formalized independently of the other livestock kinds, for example the decision Table as shown in Table (3). Table (3). An example of formalization sub-knowledge(Temperature, Rain and The proportion of the presence of artesian wells )for sheep N O Temperatur e(C) Rain(m m) The proportion of the presence of artesian wells (%) System output 1 Max Max Max Mid 2 Mid Max Max Max 3 Low Max Max Mid 4 Max Mid Max Mid 5 Mid Mid Max Max 6 Low Mid Max Mid …. …. …. …. …. 26 Mid Low Low Low 27 Low Low Low Low After knowledge formalization, we choose suitable technique to build rules for fuzzy expert system. In this paper, we use IF-THEN rules. “IF” is the condition(s) and “THEN” is the conclusion displaying the action to be taken if the condition(s) is fulfilled. For example; https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 79 IF (Temperature=Max) AND (Rain=Max) AND (The proportion of the presence of artesian wells=Max) THEN (Increased livestock=Mid) IF (Temperature=Low) AND (Rain=Low) AND (The proportion of the presence of artesian wells=Low) THEN (Increased livestock=Low). There are differences numbers of rule for each livestock kinds, there are 813 fuzzy rules for sheep, 558 fuzzy rules for goat, 435 fuzzy rules for cow, 425 fuzzy rules for buffalo and 537 fuzzy rules for camel. We use SQL server 2008 to insert these fuzzy rules. We have created one database contains five tables according to the numbers of livestock kinds (each kind has its own table). Sub-knowledge of that kind represents the columns in that table. Each rule represents one record. 4.4 Matching data with the knowledge base (the rules) by using Mamdani This expert system livestock prediction was developed by using a fuzzy rule, triangular membership function, Mamdani and centroid of area (CoA) defuzzification models. Every kind of livestock has its own (FIS) operates independently from the rest of the species, so we will have a six FIS.one FIS used for each kind of five livestock and the sixth one used the output of all livestock kinds to find the final result of predicting. The following example show sample of how to use Mamdani in our program: 1. IF (Temperature=Max) AND (Rain=Max) AND (The proportion of the presence of artesian wells=Max) THEN (Increased livestock=Mid) 2. IF (Temperature=Mid) AND (Rain=Max) AND (The proportion of the presence of artesian wells=Max) THEN (Increased livestock=Max). 3. IF (Temperature=Low) AND (Rain=Max) AND (The proportion of the presence of artesian wells=Max) THEN (Increased livestock=Mid). 4. IF (Temperature=Max) AND (Rain=Mid) AND (The proportion of the presence of artesian wells=Max) THEN (Increased livestock=Mid). 5. IF (Temperature=Mid) AND (Rain=Mid) AND (The proportion of the presence of artesian wells=Max) THEN (Increased livestock=Max). 6. IF (Temperature=Mid) AND (Rain=Low) AND (The proportion of the presence of artesian wells=Low) THEN (Increased livestock=Low). 7. IF (Temperature=Low) AND (Rain=Low) AND (The proportion of the presence of artesian wells=Low) THEN (Increased livestock=Low). Each sub-knowledge of crisp values will be fuzzyfied by using triangular membership function µ𝐴 (𝑥) which depends on the scalar parameter ( a) as a https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 80 lower limit, (b) as an upper limit, and a < m < b. Figure (4) shows the diagram of Triangular MF. [12] Figure (4). Triangular Membership Function µ𝐴 (𝑥) = { 1, 𝑥 = 𝑚 𝑥 − 𝑎 𝑚 − 𝑎 , 𝑎 ≤ 𝑥 < 𝑚 𝑏 − 𝑥 𝑏 − 𝑚 , 𝑚 < 𝑥 ≤ 𝑏 (1) Membership Function of Temperature is as the following 𝑙𝑜𝑤 (𝑋) = { 𝑥 − (−2) 5 − (−2) ,−2 ≤ 𝑥 < 5 1 ,𝑥 = 5 12 − 𝑥 12 − 5 ,5 < 𝑥 ≤ 12 (2) 𝑚𝑖𝑑 (𝑋) = { 𝑥 − 8 26.5 − 8 ,8 ≤ 𝑥 < 26.5 1 ,𝑥 = 26.5 45 − 𝑥 45 − 26.5 ,26.5 < 𝑥 ≤ 45 (3) 𝑚𝑎𝑥 (𝑋) = { 𝑥 − 40 50 − 40 ,40 ≤ 𝑥 < 50 1 ,𝑥 = 50 60 − 𝑥 60 − 50 ,50 < 𝑥 ≤ 60 (4) Membership Function of Rain is as the following https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 81 𝑙𝑜𝑤 (𝑋) = { 𝑥 − 0 17.5 − 0 ,0 ≤ 𝑥 < 17.5 1 ,𝑥 = 17.5 35 − 𝑥 35 − 17.5 ,17.5 < 𝑥 ≤ 35 (5) 𝑚𝑖𝑑 (𝑋) = { 𝑥 − 22 38.5 − 22 ,22 ≤ 𝑥 < 38.5 1 ,𝑥 = 38.5 55 − 𝑥 55 − 38.5 ,38.5 < 𝑥 ≤ 55 (6) 𝑚𝑎𝑥 (𝑋) = { 𝑥 − 43 75 − 43 ,43 ≤ 𝑥 < 75 1 ,𝑥 = 75 (7) Membership Function of the proportion of the presence of artesian wells is as the following 𝑙𝑜𝑤 (𝑋) = { 𝑥 − 0 25 − 0 ,0 ≤ 𝑥 < 25 1 ,𝑥 = 50 50 − 𝑥 50 − 25 ,25 < 𝑥 ≤ 50 (8) 𝑚𝑖𝑑 (𝑋) = { 𝑥 − 22 38.5 − 22 ,40 ≤ 𝑥 < 57.5 1 ,𝑥 = 57.5 55 − 𝑥 55 − 38.5 ,57.5 < 𝑥 ≤ 75 (9) 𝑚𝑎𝑥 (𝑋) = { 𝑥 − 65 100 − 65 ,65 ≤ 𝑥 < 100 1 ,𝑥 = 100 (10) Then, we use “Min” as implication function which was in accordance with the Mamdani model. The implication value is got from the minimum value of input parameter values of each rule. In the above rule, the implication functions Temperature (TR), Rain (RR) and the proportion of the presence of artesian wells (WR). We get (implication1 for the first rule, implication2 for the second rule… and implication7 for the seventh rule). https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 82 . implication1=Min (TR1, RR1, WR1) implication2=Min (TR2, RR2, WR2) implication3=Min (TR3, RR3, WR3) implication4=Min (TR4, RR4, WR4) implication5=Min (TR5, RR5, WR5) implication6=Min (TR6, RR6, WR6) implication7=Min (TR7, RR7, WR7) This expert system consists of a large number of fuzzy correlation rules To do a fuzzy inference system, we use the Maximum Method (Max) which determined the maximum value of implication of each rule which has the same result. So we will get the three new implications with different results (Max, Med, low). Implication (max) =Max (implication1, implication3, implication4) Implication (mid) =Max (implication2, implication5) Implication (low) =Max (implication6, implication7) The fuzzy values transform to crisp value crisp value in defuzzification by using Centroid of Area (CoA).The ranges for system output, which be identified and formalized already in a section 4.2,in Table (2),will be used. 𝐶𝑜𝐴 = i (max) × 82.5 + i(mid) × 55 + i (low) × 24.5 i (max) + i(mid) + i(low) (10) Where i is implication. 4.5 Predicting of livestock Crisp values, which obtained from the previous step, are not clear to the user as where it is representing an increase or moderation or decreasing. So we fuzzfied the results of the previous step in order to be more clear for the user. By using the following Triangular Membership functions (CoA) will be fuzzified. 𝑙𝑜𝑤 (𝐶𝑜𝐴) = { 𝐶𝑜𝐴 − 0 24.5 − 0 ,0 ≤ 𝐶𝑜𝐴 < 24.5 1 ,𝐶𝑜𝐴 = 24.5 49 − 𝐶𝑜𝐴 49 − 24.5 ,24.5 < 𝐶𝑜𝐴 ≤ 49 (11) https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 83 𝑚𝑖𝑑 (𝐶𝑜𝐴) = { 𝐶𝑜𝐴 − 40 55 − 40 ,40 ≤ 𝐶𝑜𝐴 < 55 1 ,𝐶𝑜𝐴 = 55 70 − 𝐶𝑜𝐴 70 − 55 ,55 < 𝐶𝑜𝐴 ≤ 70 (12) 𝑚𝑎𝑥 (𝐶𝑜𝐴) = { 𝐶𝑜𝐴 − 65 82.5 − 65 65 ≤ 𝐶𝑜𝐴 < 82.5 1 𝐶𝑜𝐴 = 82.5 100 − 𝐶𝑜𝐴 100 − 82.5 82.5 < 𝐶𝑜𝐴 ≤ 100 (13) 5. Implementation and Results In our system, we predict the number of five kinds of livestock (sheep, goat, cow, buffalo and camel) .Here as example, we implement (sheep, goat) and discuss their results. The system was tested previous years (2013, 2014 and 2015) where all data of livestock, the conditions experienced by the livestock and the number of livestock are known for these years. Table (4) shows the real number for sheep and goat in the years (2013, 2014 and 2015). Table (4) The real numbers for sheep and goats for years (2013,2014 and 2015) id year NO.of.Sheep NO.of.Goat 1. 2013 431506 54775 2. 2014 496810 63413 3. 2015 663341 66032 We calculate the rate of increase in the number of livestock for two years according to the following law. 𝑇ℎ𝑒 𝑟𝑎𝑡𝑒 𝑜𝑓 𝑖𝑛𝑐𝑟𝑒𝑎𝑠𝑒 = ((𝑁𝐴𝐶 − 𝑁𝐴𝑃)/ 𝑁𝐴𝑃) ∗ 100%…….(14) Where NAC is 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑎𝑛𝑖𝑚𝑎𝑙𝑠 𝑖𝑛 𝑡ℎ𝑒 𝑐𝑢𝑟𝑟𝑒𝑛𝑡 𝑦𝑒𝑎𝑟, NAP is 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑎𝑛𝑖𝑚𝑎𝑙𝑠 𝑖𝑛 𝑡ℎ𝑒 𝑝𝑟𝑒𝑣𝑖𝑜𝑢𝑠 𝑦𝑒𝑎𝑟 Table (5) shows Actual increase rate for sheep and goat. Table (5) Actual increase rate for sheep and goat id year Sheep goat 1. 2013-2014 15.13 15.76 2. 2014-2015 33.11 33.93 https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 84 Table(6) illustrate the input and results values to our system for sheep in the years(2013,2014 and 2015) I d Sub-knowledge for sheep 2013 2014 2015 1. Temperature (C) 27.9 27.1 28 2. Rain (mm) 14,51 6 17 49.5 3. Torrents (%) 0 0 0 4. The proportion of the presence of artesian wells (%) 10 15 35 5. Ratio of determine the grazing areas and the work of nature reserves (%) 20 25 50 6. Ratio of activate and revitalize the role of the Directorate of pastures (%) 0 0 0 7. Ratio of providing homes for breeders (%) 0 0 0 8. Percentage of domestic consumption (%) 20 10 10 9. Ratio of standard of living of the individual (%) 70 66 55 10. Ratio of religious and social occasions (%) 10 20 10 11. Proportion of commitment of livestock breeders to vaccinate their animals (%) 90 90 90 12. Ratio of organize periods for livestock feed (%) 60 60 60 13. Ratio of Cultural awareness among livestock breeders in the feasibility of modern production for livestock (%) 30 40 50 14. Ratio of Address the stagnant pools and disposal of contaminated feed (%) 0 0 0 15. Ratio of Get rid of the diseased dogs or processed (%) 0 0 0 16. Ratio of availability committees veterinary to visit villages and pastures (%) 90 90 90 17. Ratio of conducting spraying and dipping (%) 90 90 90 18. Ratio of availability of vaccines and treatments (%) 90 90 90 19. Ratio of Smuggling (%) 0 0 0 20. Ratio of Wool and leather prices in the market (%) 0 0 0 21. Ratio of Import of meat and milk and dairy products (%) 95 95 95 22. Ratio of Provide fodders for livestock breeders by the government (%) 10 70 40 23. Ratio of the provision of trucks and cars to 0 0 0 https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 85 Ranchers by the government (%) 24. Ratio of Interest in international relations, to control the intercontinental diseases (%) 10 10 10 25. Ratio of The presence of wool production plants, leather and milk products (%) 10 10 12 26. Ratio of Activation of the massacres and livestock slaughter law (%) 70 70 70 27. Ratio of Increase cultural awareness among breeders (%) 100 95 85 28. Ratio of The presence of private associations for each type of livestock species (%) 10 10 10 29. Ratio of A census of livestock to learn their numbers and kinds (%) 40 50 75 30. Ratio of Genetic improvement for livestock (%) 0 0 0 31. Ratio of Building stockyard for animals (%) 5 5 10 32. Percentage of compliance with laws that prevent the smuggling operation (%) 90 90 90 33. Ratio of Planning for the development of livestock (%) 10 10 10 34. The proportion of use of scientific and technical cadres (%) 50 50 50 35. The proportion of attention in the Faculty of Veterinary Medicine and the College of Agriculture, Department of Animal Wealth (%) 60 60 60 36. Ratio of Attention in veterinarians and agronomists (%) 70 70 70 37. Ratio of Excessive intake of feed (%) 10 10 10 38. Ratio of presence Infectious diseases (%) 50 50 50 39. Ratio of the spread of viruses, parasites, and insects and contaminated feed (%) 60 60 60 40. Ratio of The emergence of genetic mutations (%) 10 10 10 41. Ratio of The presence of mines in pastures (%) 15 12 10 42. Result Low 43. Mid 0.620 2 0.14 95 44. Max 0.15 75 0.34 39 45. Real increase rate 15.1 3 33.1 1 46. Error rate 4.09 3.86 https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 86 % % Table(7) illustrate the input and reults values for our system for goat in the years (2013,2014 and 2015) Id sub-knowledge for goat 2013 2014 2015 1. Temperature (C) 27.9 27.1 28 2. Rain (mm) 14,516 17 49.5 3. Torrents (%) 0 0 0 4. The proportion of the presence of artesian wells (%) 10 15 35 5. Ratio of determine the grazing areas and the work of nature reserves (%) 20 25 50 6. Ratio of activate and revitalize the role of the Directorate of pastures (%) 0 0 0 7. Ratio of providing homes for breeders (%) 0 0 0 8. Percentage of domestic consumption (%) 20 10 10 9. Proportion of commitment of livestock breeders to vaccinate their animals (%) 90 90 90 10. Ratio of organize periods for livestock feed (%) 60 60 80 11. Ratio of Cultural awareness among livestock breeders in the feasibility of modern production for livestock (%) 30 40 55 12. Ratio of Address the stagnant pools and disposal of contaminated feed (%) 0 0 0 13. Ratio of Get rid of the diseased dogs or processed (%) 0 0 0 https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 87 14. Ratio of availability committees veterinary to visit villages and pastures (%) 90 90 90 15. Ratio of conducting spraying and dipping (%) 90 90 90 16. Ratio of availability of vaccines and treatments (%) 90 90 90 17. Ratio of Provide fodders for livestock breeders by the government (%) 10 70 40 18. Ratio of the provision of trucks and cars to pelvic Ranchers by the government (%) 0 0 0 19. Ratio of Interest in international relations, to control the intercontinental diseases (%) 10 10 10 20. Ratio of Activation of the massacres and livestock slaughter law (%) 70 70 70 21. Ratio of Increase cultural awareness among breeders (%) 100 95 85 22. Ratio of The presence of private associations for each type of livestock species (%) 10 10 10 23. Ratio of A census of livestock to learn their numbers and kinds (%) 40 50 75 24. Ratio of Building stockyard for animals (%) 5 5 10 25. Ratio of Planning for the development of livestock (%) 10 10 10 26. The proportion of use of scientific and technical cadres (%) 50 50 50 27. The proportion of attention in the Faculty of Veterinary Medicine and the College of Agriculture, Department of Animal Wealth (%) 60 60 60 28. Ratio of Attention in veterinarians and agronomists (%) 70 70 70 https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 88 29. Ratio of Excessive intake of feed (%) 10 10 10 30. Ratio of presence Infectious diseases (%) 50 50 50 31. Ratio of the spread of viruses, parasites, and insects and contaminated feed (%) 60 60 60 32. Ratio of The emergence of genetic mutations (%) 10 10 10 33. Result Low 34. Mid 0.6115 0.1388 35. Max 0.1667 0.3529 36. Real increase rate 15.76 33.93 37. Error rate 5.77 4.13 Tables (6) and (7) illustrate the inputs of our expert system for the years (2013, 2014 and 2015) as well as the system results for each of these years. Because we deal with fuzzy logic systems, system results are usually decimal fractions less than 1 and the highest value we can get is 1. Therefore, the system output will be multiplied by 100 to make the results of the system are identical to reality. We will compare the results of the expert system with the actual rates of increase for livestock of all kinds in the years listed above and compute error rates for each kind in Table (8) Table (8). compare the results of the expert system with the actual rates of increase for livestock and error rate id Livestock kinds 2013-2014 2014-2015 1. Sheep Real increase rate 15.13 Real increase rate 33.11 Expected increase rate 15.75 Expected increase rate 34.39 Error rate 4.09% Error rate 3.86 % 2. Goat Real increase rate 15.76 Real increase rate 33.93 Expected increase rate 16.67 Expected increase rate 35.29 https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 89 Error rate 5.77% Error rate 4.13 % 3. Conclusions This paper shows how to build an expert system for predicting livestock preparation, where the fuzzy logic was used. FIS were used because they have broken the bounds of conventional programming, which is designed like the human reasoning. This enables the FIS to represent and handle more complex problems than conventional programming. The advantage of using fuzzy expert systems to predict livestock numbers is to help decision makers, farmers and researchers to predict livestock production quickly and accurately. References 1. T. Abdul-Mageed, U. K. Jabra," An Economic Study About The Demand Of Major Red And White Meat In Iraq Of 1990 – 2012 Using Linear Approximation Modelof Almost Ideal Demand System", The Iraqi Journal Of Agricultural Sciences ,3, 47, 829-836,2016. 2. Mahmood Bader Ali Al-Samea, Muthana Fadhil Ali," Geographic Analysis of the livestock in Iraq and its Natural and biotic problems and its development potentialities",Adab Al-Kufa,18,1, 171-218,2014. 3. Sanjay Kumar, Dr. Rajkishore Prasad," Importance of Expert System Shell in Development of Expert System", International Journal of Innovative Research and Development", Vol 4, Issue 3,128-133, March, 2015. 4. Özlem İnan, Derya Arslan, Şakir Taşdemir, and Mehmet Musa Özcan," Application of fuzzy expert system approach on prediction of some quality characteristics of grape juice concentrate (Pekmez) after different heat treatments", Journal of Food Science and Technology,48(3),423-431,Aug 2011. 5. Enrique Castillo ،Jose M. Gutierrez ،Ali S. Hadi ," Expert Systems and Probabilistic Network Models", Springer, 1997. 6. Diana-Aderina Moisuc and Mihai-Constantin Avornicului," ARCHITECTURAL MODEL OF EXPERT SYSTEMS", 5th International Symposium Engineering Management and Competitiveness, 7. K P Tripathi," A Review on Knowledge-based Expert System:Concept and Architecture" , IJCA Journal, Number 4 - Article 4,19-22,2011. https://jutq.utq.edu.iq/index.php/main University of Thi-Qar Journal Vol.12 No.3 SEP 2017 Web Site: https://jutq.utq.edu.iq/index.php/main Email: journal@jutq.utq.edu.iq 90 8. Nwigbo Stella N and Agbo Okechuku Chuks," EXPERT SYSTEM: A CATALYST IN EDUCATIONAL DEVELOPMENT IN NIGERIA", 1st International Technology, Education and Environment Conference (c) African Society for Scientific Research (ASSR), 566-571, 2011. 9. L.A. Zadeh, Fuzzy sets. Information and Control, 8(3):338 – 353, 1965. 10. Franck Dernoncourt," Introduction to fuzzy logic",Masschusetts Institute of Technology, January 2013. 11. Leekwijck, W. V. and Kerre, E. E. Defuzzication :criteria and classification. Fuzzy Sets and Systems, 108(2):159 – 178, 1999. 12. Ekananta Manalif," Fuzzy Expert-COCOMO Risk Assessment and Effort Contingency Model in Software Project Management", The University of Western Ontario, A thesis submitted in partial fulfillment of the requirements for the degree in Master of Engineering Science,2013. https://jutq.utq.edu.iq/index.php/main 
work_yxedqsj2qnd6jgtnktcg4jlgam	----	[PDF] Quality evaluation of restored soils with a fuzzy logic expert system | Semantic Scholar Skip to search formSkip to main content> Semantic Scholar's Logo Search Sign InCreate Free Account You are currently offline. Some features of the site may not work correctly. DOI:10.1016/J.GEODERMA.2009.04.018 Corpus ID: 8004988Quality evaluation of restored soils with a fuzzy logic expert system @article{Kaufmann2009QualityEO, title={Quality evaluation of restored soils with a fuzzy logic expert system}, author={M. Kaufmann and S. Tobias and R. Schulin}, journal={Geoderma}, year={2009}, volume={151}, pages={290-302} } M. Kaufmann, S. Tobias, R. Schulin Published 2009 Mathematics Geoderma Abstract Due to landscaping, mining and construction activities on previously cultivated land, more and more soils are excavated, translocated, deposited and restored. In many cases restored soils show signs of structural degradation such as overcompaction and water logging. There is a lack of methods to evaluate and assess the physical quality of restored soil. In this study a fuzzy logic expert system was developed which allows us to evaluate the potential plant productivity of restored soils… Expand View via Publisher researchgate.net Save to Library Create Alert Cite Launch Research Feed Share This Paper 38 CitationsHighly Influential Citations 1 Background Citations 8 Methods Citations 3 View All Figures and Tables from this paper figure 1 table 1 figure 2 table 2 figure 3 table 3 figure 4 figure 5 figure 6 figure 7 figure 8 figure 9 figure 10 figure 11 View All 14 Figures & Tables 38 Citations Citation Type Citation Type All Types Cites Results Cites Methods Cites Background Has PDF Publication Type Author More Filters More Filters Filters Sort by Relevance Sort by Most Influenced Papers Sort by Citation Count Sort by Recency A Fuzzy-based Methodology for an Aggregative Environmental Risk Assessment of Restored Soil S. Wang, Z. Zhao, Bing Xia, H. Qiu, J. Morel, R. Qiu Environmental Science 2014 10 View 1 excerpt, cites background Save Alert Research Feed A fuzzy logic-based expert system for substrate selection for soil construction in land reclamation F. B. D. Souza, Émilin de Jesus Casagrande de Souza, Merisandra Côrtes de Mattos Garcia, K. Madeira Computer Science 2018 2 Highly Influenced PDF View 3 excerpts, cites background Save Alert Research Feed The effects of pipeline construction disturbance on soil properties and restoration cycle Peng Shi, J. Xiao, Ya-Feng Wang, L. Chen Medicine, Environmental Science Environmental Monitoring and Assessment 2013 12 PDF View 1 excerpt, cites background Save Alert Research Feed Fuzzy logic for acid sulfate soil mapping: Application to the southern part of the Finnish coastal areas A. Beucher, S. Fröjdö, Peter Österholm, Annu Martinkauppi, P. Eden Environmental Science 2014 20 Save Alert Research Feed A Study of the Spatial Difference of the Soil Quality of The Mun River Basin during the Rainy Season C. Wu, Q. Liu, +5 authors L. Liang Economics 2019 3 Save Alert Research Feed Assessing the health of agricultural land with emergy analysis and fuzzy logic in the major grain-producing region Q. Li, J. Yan Geography 2012 22 Save Alert Research Feed Assessment on soil fertility of Dongting Lake wetland area (China) based on GIS and fuzzy evaluation Z. Li, Jin-Quan Huang, Y. Li, Wang Guo, J. Zhu Environmental Science 2011 7 Save Alert Research Feed Application of fuzzy logic to Boolean models for digital soil assessment J. J. Gruijter, D.J.J. Walvoort, G. Bragato Mathematics 2011 19 Save Alert Research Feed Changes in soil pore network in response to twenty-three years of irrigation in a tropical semiarid pasture from northeast Brazil P. A. D. Costa, Jaedson Cláudio Anunciato Mota, R. E. Romero, A. G. Freire, T. O. Ferreira Environmental Science 2014 21 Save Alert Research Feed Study of the differences in soil properties between the dry season and rainy season in the Mun River Basin C. Wu, Gaohuan Liu, +5 authors L. Liang Geology 2019 4 Save Alert Research Feed ... 1 2 3 4 ... References SHOWING 1-10 OF 55 REFERENCES SORT BYRelevance Most Influenced Papers Recency Fuzzy mapping of soil fertility — a case study on irrigated riceland in the Philippines A. Dobermann, T. Oberthür Mathematics 1997 58 View 1 excerpt, references methods Save Alert Research Feed Modelling experts’ judgments on soil compaction to derive decision rules for soil protection—A case study from Switzerland S. Tobias, O. Tietje Geology 2007 12 View 3 excerpts, references background and methods Save Alert Research Feed A knowledge-based fuzzy expert system to analyse degraded terrain D. Genske, Klemens Heinrich Computer Science Expert Syst. Appl. 2009 19 View 2 excerpts, references background Save Alert Research Feed Least limiting water range indicators of soil quality and corn production, eastern Ontario, Canada D. Lapen, G. C. Topp, E. Gregorich, W. Curnoe Environmental Science 2004 113 View 1 excerpt, references results Save Alert Research Feed Soil physical quality: Part III: Unsaturated hydraulic conductivity and general conclusions about S-theory A. Dexter Geology 2004 290 Highly Influential View 4 excerpts, references background and methods Save Alert Research Feed Linking process capability analysis and least limiting water range for assessing soil physical quality A. Silva, B. Kay Environmental Science 2004 39 View 2 excerpts, references background and results Save Alert Research Feed Fuzzy mathematical methods for soil survey and land evaluation P. Burrough Computer Science 1989 497 View 1 excerpt, references methods Save Alert Research Feed Fuzzy modeling of farmers' knowledge for land suitability classification Rodrigo S. Sicat, E. J. Carranza, U. Nidumolu Mathematics 2005 124 PDF View 2 excerpts, references background Save Alert Research Feed Soil physical quality: Part II. Friability, tillage, tilth and hard-setting A. Dexter Geology 2004 186 Highly Influential View 4 excerpts, references background and methods Save Alert Research Feed Applications of S‐theory in the study of soil physical degradation and its consequences A. Dexter, E. Czyż Geology 2007 80 View 2 excerpts, references background Save Alert Research Feed ... 1 2 3 4 5 ... Related Papers Abstract Figures and Tables 38 Citations 55 References Related Papers Stay Connected With Semantic Scholar Sign Up About Semantic Scholar Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI. Learn More → Resources DatasetsSupp.aiAPIOpen Corpus Organization About UsResearchPublishing PartnersData Partners   FAQContact Proudly built by AI2 with the help of our Collaborators Terms of Service•Privacy Policy The Allen Institute for AI By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy Policy, Terms of Service, and Dataset License ACCEPT & CONTINUE 
work_yzhptnfd6rbl3e7hojemflxpfi	----	50 Gender as a Predictive Factor for Tasks Completed Using Smartphones Mark Evans University of North Texas markevans@my.unt.edu Susan Hopper University of North Texas susanhopper@my.unt Greg Jones University of North Texas Greg.Jones@unt.edu Gerald Knezek University of North Texas Gerald.Knezek@unt.edu Abstract Smartphone usage among undergraduate students is growing (Woodcock, Middleton, & Nortcliffe, 2012). The prevalence of these devices in every aspect of student routines presents the research problem: how will smartphone usage affect how college students manage their coursework? To address this research problem, we conducted a pilot study examining the current level of smartphone usage among college students. A distinct battery of questions were created and posted to an online survey. Forty-two undergraduate students who own and use smartphones participated. Survey questions focused on how frequently students utilized specific smartphone functions to complete class-related tasks. This paper applies a quantitative analysis to this specific battery of questions in an effort to address the research question: Is gender a significant factor in determining the frequency with which smartphones are used to complete informal learning tasks? Keywords: gender, smartphone, mobile learning, task, behavior Introduction Smartphone usage among undergraduate students is growing exponentially (Woodcock et al., 2012). The prevalence of these devices in every aspect of student routines presents researchers with the question: how will smartphone usage affect how college students manage their college coursework and task responsibilities? Initially, this research study sought to address this question through a qualitative analysis of current student interviews. We hypothesized that the level of adoption and how frequently students use their smartphones to manage coursework could indicate the level of impact that smartphones have in a college setting. For this purpose, we took the preliminary step of developing a battery of questions based on observations made while teaching a college-level blended learning class. The initial draft of questions focused on how frequently the students utilized their smartphones to complete specific tasks related to managing their blended-learning coursework inside the classroom. Next, we broadened the scope of the questions to include tasks that could be conducted using a smartphone while outside of the classroom. These were paired down to a final battery of 15 questions testing task frequency for each student, inside and outside of the classroom. Forty-two undergraduate students, 15 males and 27 females, who owned smartphones and were enrolled in a blended learning course at the time of the survey, participated. After compiling the results, we tested the online survey as a statistical instrument to see if more might be gleaned from the data collected. We hypothesized that a student’s gender could be a key determinant in how frequently students choose to utilize their smartphones to complete different types of ________________________________________ Evans, M., Hopper, S., Jones, G., & Knezek, G. (2013). Gender as a predictive factor for tasks completed using smartphones. iConference 2013 Proceedings (pp. 50-53). doi:10.9776/13135 Copyright is held by the authors. mailto:markevans@my.unt.edu mailto:susanhopper@my.unt mailto:Greg.Jones@unt.edu mailto:Gerald.Knezek@unt.edu iConference 2013 February 12-15, 2013 Fort Worth, TX, USA 51 activities relevant to their coursework. This study applies a quantitative analysis to the battery of questions in the online survey, in an effort to test the research question: Is gender a significant factor in determining the frequency with which smartphones are used for informal learning tasks? The outcome of this analysis will provide insight into whether a student’s gender might be a predictor of how frequently students use smartphones to complete specific tasks related to managing their coursework. Literature Review It has been well-established that gender plays a critical role in the formulation of an individual’s personality traits (Bouchard Jr. & Loehlin, 2001). Studies also indicate that specific personality traits have a significant impact on smartphone usage (Lane & Manner, 2011). Clough, Jones, McAndrew, and Scanlon (2008) suggest that smartphones are already, “used extensively in an informal learning context by enthusiasts.” Furthermore, extensive research shows a significant difference in how men and women use social networking sites and web-based media, one of the most-used applications on smartphones (Muscanell & Guadagno, 2012; Oulasvirta, Rattenbury, Ma, & Raita, 2011). Gender differences also relate to purchase decisions by smart phone users. Wilska (2003) examined the relationship of consumption patterns and mobile phone use when adopting mobile phones as a form of new technology. An “addictive” use of mobile phones linked females to “trendy” and “impulsive” consumption styles. Technology enthusiasm and trend-consciousness related males to consumption and “hard” values. While currently little research exists that shows a connection between gender and smartphone usage for informal learning and the management of undergraduate coursework, gender differences have been studied in technology affinity. Studies indicate that there are differences of interest in technology and in perceptions of skills toward technology by gender, with women displaying lower levels of interest (Frantom, Green, and Hoffman, 2002). The Technology Affinity Survey (Knezek, 2011) was designed to assess involvement with communications technology, Internet, and media tools. In a study of 68 teacher preparation candidates (Knezek, Mills, Wakefield, & Hopper, 2011) findings showed gender as a significant (p<.05) discriminator for the level of technology affinity. Males were higher in digital technology affinity than females. Morris and Venkatesh (2000) researched the adoption of a new technology in general using the Technology Acceptance Model. Women’s technology usage decisions were strongly influenced by their perceptions of ease of use. In contrast, men’s were more strongly influenced by their perceptions of usefulness. Gender differences exist in use of social and web based media, consumption patterns, attitudes and affinity toward technology. Results from the Smartphone Usage Instrument (Hopper & Evans, 2012) will explore gender differences in frequency of smartphone tasks and should it prove reliable and valid, could contribute additional findings to the field. Method In this study, 42 college students enrolled in a blended learning class answered 15 survey questions about how they use their smart phones. Each question asked the students to rate the frequency with which they used their smartphones to complete a specific task for their class by choosing never, once per month, 2-3 times per month, once per week, 3-5 times per week, every day or several times per day. Participants were able to see all of the questions at one time (as shown Figure 1) and experienced no time limits while completing the survey. An assessment of the survey questions was conducted to test for content validity. The survey questions were clear and understandable and passed the fundamental requirement for content validity (Sireci, 1998). The instrument proved to be reliable, with a Cronbach’s Alpha score of .947 (Cronbach, 1951) (DeVellis, 1991). Dimension reduction through exploratory factor analysis showed two distinct categories of data using varimax criterion (Kaiser, 1958). Two constructs were created based on the categories identified. The first factor centered on using smartphones for logistics and organizational tasks for the course (F1). The second factor centered more on activities that facilitated deeper social relationships, including such tasks as sharing videos or photographs (F2). These categories were also tested for iConference 2013 February 12-15, 2013 Fort Worth, TX, USA 52 internal reliability using Cronbach’s Alpha and both factors proved reliable (DeVellis, 1991). F1 returned a Cronbach’s Alpha of .915 and F2 returned a Cronbach’s Alpha of .943. An additional factor analysis which forced 3, then 4 constructs showed no additional factors worth consideration. Cronbach’s Alpha factor scores demonstrated excellent reliability for creating scales, so scales were created accordingly, defined as F1and F2 (DeVillis, 1991). These scales were then tested for criterion-related validity using a one-way ANOVA test with gender as the discriminating factor. We hypothesized that how frequently students use smartphones to support different informal learning tasks would be dependent on gender. Specifically, we predicted that female students would more frequently use their smartphones in activities related to F2, which helped facilitate social relationships and also involved sharing videos or photographs. The results confirmed our hypothesis. Figure 1: Survey of Smartphone Usage Instrument frequency. [(Hopper & Evans, 2012) While F1 was far from showing any significant difference based on gender (p=.798), F2 showed a significance score that did merit further consideration (p =.089). Female students used their smartphones more frequently than male students to facilitate social relationships or share videos and pictures. Though this score failed to meet the threshold for significance where p <.05, the sample size was fairly modest. We then tested the effect size to see if a larger sampling should be considered. The effect size is moderately large, with a Cohen’s d =.601 (Cohen, 1988). Due to the moderately large effect size, a slightly larger sample size is worthy of consideration, as this result would be considered educationally significant (Sivin-Kachala, Bialo, & Langford, 1997). By increasing sample size by a modest amount, the iConference 2013 February 12-15, 2013 Fort Worth, TX, USA 53 instrument would pass the Criterion validity test, meaning that gender could be a predictive factor in whether the student used their smartphone for high-end social communication and multimedia content. Conclusion The Smartphone Usage Instrument tested in this study is reliable. The scales created showed promise in using gender to predict the frequency that undergraduate students will use their smartphones for social relationship tasks and sharing video and pictures in the context of a blended learning class. The effect size is moderately large and with a larger sample size, gender would prove to be a significant predictive factor. Based on these results, we believe this pilot reveals the need for further study on smartphone usage among students in college courses. We intend to make subtle refinements to the instrument, as well as increase our sample size, in an effort to show more conclusively whether gender can be used as a determining factor in predicting the types of tasks students use smartphones to complete when managing college coursework, as well as the frequency with which they complete those tasks. References Bouchard Jr., T. J., & Loehlin, J. C. (2001). Genes, evolution, and personality. Behavior Genetics, 31(3), 30. doi:10.1023/A:1012294324713 Clough, G., Jones, A. C., McAndrew, P., & Scanlon, E. (2008). Informal learning with PDAs and smartphones. Journal of Computer Assisted Learning, 24(5), 359-371. doi:10.1111/j.1365- 2729.2007.00268.x Cohen, Jacob. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Hillsdale, N.J: Lawrence Erlbaum Associates. Cronbach, Lee. (1951). Coefficient alpha and the internal structure of tests. Psychometrika, 16(3), 297- 334. doi:10.1007/BF02310555 DeVillis, R.F. (1991). Scale development. Newbury Park, NJ: Sage Publications. Frantom, C.G., Green, K.E., Hoffman, E.R. (2002). Measure development: The children's attitude towards technology development (CATS) . Journal of Educating Computer Research. 26(3) 249-263. Hopper, S., & Evans, M., (2012) Smartphone Usage Instrument. Denton, TX. Kaiser, Henry. (1958). The varimax criterion for analytic rotation in factor analysis. Psychometrika, 23(3), 187-200. doi:10.1007/BF02289233 Knezek, G. (2011). Technology Affinity Scale (TAS). Denton, TX: Institute for the Integration of Technology into Teaching & Learning. Knezek, G., Mills, L., Wakefield, J., Hopper, S., (2011). Relationship of technology affinity to stem dispositions in higher education learners. World Conference on E-Learning in Corporate, Government, Healthcare, and Higher Education (ELEARN) 2011:1 , Oct 18, 2011 Lane, Wilburn, & Manner, Chris. (2011). The Impact of Personality Traits on Smartphone Ownership and Use. International Journal of Business & Social Science, 2(17), 22-28. Morris, M.G. & Venkatesh, V. (2000). Why men don't ever stop to ask for directions? Gender, social influence, and their role in technology acceptance and usage. MIS Quarterly, 24, (1), 115- 139.Muscanell, Nicole L., & Guadagno, Rosanna E. (2012). Make new friends or keep the old: Gender and personality differences in social networking use. Computers in Human Behavior, 28(1), 107-112. doi:10.1016/j.chb.2011.08.016 Oulasvirta, Antti, Rattenbury, Tye, Ma, Lingyi, & Raita, Eeva. (2011). Habits make smartphone use more pervasive. Personal and Ubiquitous Computing, 16(1), 105-114. doi:10.1007/s00779-011-0412-2 Sireci, Stephen G. (1998). The Construct of Content Validity. Social Indicators Research, 45(1), 83-117. doi:10.1023/A:1006985528729 Sivin-Kachala, Jay, Bialo, Ellen R., & Langford, Jonathan. (1997). The Effectiveness of Technology in Schools, '90-'97. Report: Software Publishers Association, Washington D. C. Wilska, A. (2003). Mobile use as part of young people's consumption styles. Journal of Consumer Policy. 26, 441-463.Woodcock, Ben, Middleton, Andrew, & Nortcliffe, Anne. (2012). Considering the Smartphone Learner: developing innovation to investigate the opportunities for students and their interest. Student Engagement and Experience Journal, 1(1). doi:10.7190/seej.v1i1.38 
work_z2qcpcnrmzfrnjprauibaterbq	----	Document: – 22 – Wissensbasierte differentialdiagnostische Unterstützung in der Rheumatologie: Ergebnisse und Erfahrungen G. Kolarz 1), K.–P. Adlassnig 2), H. Leitich 2,3) 1) Rheuma–Sonderkrankenanstalt der Sozialversicherungsanstalt der Gewerblichen Wirtschaft, Baden, Institut für Rheumatologie der Kurstadt Baden, Adolfine–Malcherg.1, 2500 Baden 2) Institut für Medizinische Computerwissenschaften der Universität Wien, Wien email: peter.adlassnig@akh–wien.ac.at 3) Universitätsklinik für Frauenheilkunde, Abteilung für Geburtshilfe und Gynäkologie, Universität Wien Medizinische Expertensysteme sind Computerprogramme, die bei der Diagnoseerstellung, der Therapieauswahl, der Prognoseeinschätzung und Patientenführung im Krankenhaus oder der Arztpraxis eingesetzt werden können. In der Folge werden die Erfahrungen mit dem Expertensy- stem CADIAG–II beschrieben. Die Grundlage für das Konsultationssystem CADIAG–II bildet eine medizinische Wissensbasis, die medizinisches Wissen in Form von nosologischen Krankheitsprofilen speichert und im kon- kreten Anwendungsfall zur Verfügung stellt. In der Wissensbasis wurden die einzelnen Sym- ptome, gegebenenfalls auch Symptomkombinationen für die verschiedenen rheumatischen Erkrankungen aufgenommen und die Häufigkeit des Auftretens eines bestimmten Symptomes bei einer bestimmten Erkrankung bzw. auch seine Beweiskraft für eine bestimmte Erkrankung angegeben. Die Diagnoseermittlung erfolgt so, daß die Symptome eines Patienten mit den Ein- tragungen in der Wissensbasis verknüpft und Diagnoseergebnisse abgeleitet werden. Insgesamt wurden 185 rheumatologische Erkrankungen und mehr als 3500 Krankengeschichten von Pati- enten, die an der Rheuma–Sonderkrankenanstalt der Sozialversicherungsanstalt der Gewerbli- chen Wirtschaft behandelt wurden, aufgenommen. In verschiedenen Studien wurde die Praktikabilität dieses Diagnosesystems anhand von realen Krankengeschichten überprüft. Bei der chronischen Polyarthritis und anderen entzündlichen rheumatischen Erkrankungen wurde bei durchschnittlich 90% der Patienten die korrekte Dia- gnose erstellt. Bei degenerativen Gelenkerkrankungen war dieser Prozentsatz deutlich niedriger; bei der Diagnose Gicht lag die Trefferquote am niedrigsten (nur 40%). Die Ursache für das schlechte Ergebnis bei Gicht liegt in der Tatsache begründet, daß es für diese Erkrankung eine spezifische Therapie gibt, wodurch die Symptome der Erkrankung schwinden. Auch der Arzt kann daher nur aus den anamnestischen Angaben und den eingenommenen Medikamenten die Diagnose stellen. Anamnestische Angaben dürfen aber im Regelfall nicht als hochwertige Sym- ptome/Befunde zur Erstellung einer computerunterstützten Diagnose eingehen, da diese sehr häufig ungenau und lückenhaft sind. Insgesamt läßt sich aus den bisher durchgeführten Studien schließen, daß das System seine Stär- ken bei der Erstellung von selteneren Diagnosen aufweist. Insofern kann das Diagnosesystem CADIAG–II dem mit rheumatischen Erkrankungen weniger erfahrenen Arzt eine brauchbare Hilfestellung geben. Literatur � Adlassnig, K.–P., Leitich, H. & Kolarz, G. (1993) On the Applicability of Diagnostic Criteria for the Diagnosis of Rheumatoid Arthritis in an Expert System. Expert Systems with Applications 6, 441–448. � Adlassnig, K.–P., Leitich, H. & Kolarz, G. (1996) Expertensysteme in der Rheumatologie. Rheumatologie in Europa 25/3, 104–108. 
work_z3zntoflhnd7phrocdyphm5at4	----	Learning Patterns of University Student Retention Ashutosh Nandeshwara,∗, Tim Menziesb, Adam Nelsonc aKent State University, 126 Lowry Hall, Kent, OH 44242 Phone: 330-672-82222 bWest Virginia University, 841a Engineering Sciences Building, Morgantown, WV 26505 cWest Virginia University, Engineering Sciences Building, Morgantown, WV 26505 Abstract Learning predictors for student retention is very difficult. After reviewing the literature, it is evident that there is considerable room for improvement in the current state of the art. As shown in this paper, improvements are possible if we (a) explore a wide range of learning methods; (b) take care when selecting attributes; (c) assess the efficacy of the learned theory not just by its median performance, but also by the variance in that performance; (d) study the delta of student factors between those who stay and those who are retained. Using these techniques, for the goal of predicting if students will remain for the first three years of an undergraduate degree, the following factors were found to be informative: family background and family’s social-economic status, high school GPA and test scores. Keywords: data mining, student retention, predictive modeling, financial aid 1. Introduction This article uses data mining to find patterns of student retention at American Uni- versities. Such an analysis is urgently required. In our work, we have seen a disconnect between accepted best practices and the data available to support those practices: • Based on our discussion with the university administrators, we assert that there is much informal agreement on the factors that influence retention (for the most part, the financial status of the student is considered to be the most important factor after the student’s high-school GPA). • However, as shown below, when we look recent experiments with student records, we find little clear support for that informal belief. In fact, we know of many universities that try to improve retention with a wide range of programs such as: – attracting students with high performance indicators (such as their test scores) ∗Corresponding author Email addresses: anandesh@kent.edu (Ashutosh Nandeshwar), tim@menzies.us (Tim Menzies), rabituckman@gmail.com (Adam Nelson) URL: www.nandeshwar.info (Ashutosh Nandeshwar), www.menzies.us (Tim Menzies) Preprint submitted to Elsevier May 27, 2011 – developing student success programs (such as first-year experience) – or encouraging tenured-faculty to teach undergraduates Given the large levels of public support allocated to universities, it is important that we check the validity of these informal intuitions as well as the utility of the various retention programs such as those conducted at Kent State. This article applies data mining methods to the problem of studying student retention. Our general conclusions will be: • It is possible to find patterns of student retention, using data mining; • Previous data mining studies on these student records can be greatly improved us- ing discretization, attribute selection and cross-validation over various algorithms. More specifically, we will show that data mining can uncover a rich level of detail about particular universities. For example, while mining data from Kent State, we found: • A small and specific population of students at high risk of dropping out of university. • That the above programs (using tenured-faculty for lecturing and focusing on stu- dent performance data) is far less important than the financial status of a student. Hence, we would recommend: • Focusing more resources on the high-risk group of students, in order to improve their chances of completing a university degree. • Discontinuing the retention programs that primarily focus on student performance indicators or that advocate using tenured-faculty for lecturing. While these conclusions are specific to Kent State, the method for finding them is quite general and could be applied to other universities in order to find their most specific and most important student retention patterns. We welcome contacts from other researchers who wish to repeat our analysis on their local data. 2. Literature Review It is no news that higher education institutions are facing the problem of student retention, which affects graduation rates as well. Colleges with higher freshmen reten- tion rate tend to have higher graduation rates within four years. The average national retention rate is close to 55% and in some colleges fewer than 20% of incoming student cohort graduate (Druzdzel & Glymour, 1994), and approximately 50% of students enter- ing in an engineering program leave before graduation (Scalise et al., 2000). Tinto (1982) reported national dropout rates and BA degree completions rates for the past 100 years to be constant at 45 and 52 percent respectively, except for the World War II period (see Figure 1 for the completion rates from 1880 to 1980). Tillman & Burns (2000) at Valdosta State University (VSU) projected lost revenues per 10 students, who do not persist their first semester, to be $326,811. Although gap between private institutions and public institutions in terms of first-year students returning to second year is closing, the retention rates have been constant for both types of institutions (ACT, 2007, see 2 Figure 2). National Center for Public Policy and Higher Education (NCPPHE) reported the U.S. average retention rate for the year 2002 to be 73.6% (NCPPHE, 2007). This problem is not only limited to the U.S. institutions, but also for the institutions in other countries such as U.K and Belgium. The U.K. national average freshmen retention for the year 1996 was 75% (Lau, 2003), and Vandamme (2007) found that 60% of the first generation first-year students in Belgium fail or dropout. [Figure 1 about here.] [Figure 2 about here.] Various researchers have studied this problem extensively, using theoretical models (Tinto, 1975, 1988; Spady, 1970, 1971; Bean, 1980), traditional models (Terenzini & Pas- carella, 1980; Pascarella & Terenzini, 1979, 1980), and data mining techniques (Druzdzel & Glymour, 1994; Sanjeev & Zytkow, 1995; Massa & Puliafito, 1999; Stewart & Levin, 2001; Veitch, 2004; Barker et al., 2004; Salazar et al., 2004; Superby et al., 2006; Sujit- parapitaya, 2006; Herzog, 2006; Atwell et al., 2006; Yu et al., 2007; DeLong et al., 2007). As shown below, we can improve those prior results by augmenting standard data mining with discretization, attribute selection, and cross-validation over various algorithms. As documented in Adam & Gaither (2005), the literature on retention in higher education is extensive, and although various researchers have tested theoretical models and noted attributes critical to student retention, these theories need to be tested from time-to-time. New generation of data miners make the testing easier, and possibly can find new theories or reject old theories using state-of-the art learning algorithms. In this section, we focus on the literature relating to data mining and the student retention problem. The lesson of this section is that learning patterns of student retention is very difficult and, despite decades of effort, there is much room for improvement in the current state of the art. 2.1. Data Mining for Student Retention Druzdzel & Glymour (1994) were among the first researchers to apply knowledge dis- covery algorithm to study the student retention problem. The authors applied TETRAD II, a casual discovery program developed at Carnegie Mellon University, to the U.S. news college ranking data to find the factors that influenced student retention, and they found that the main factor of retention was the average test score. Using linear regression, the authors found that test scores alone explained 50.5% of the variance in freshmen retention rate. In addition, they concluded that other factors such as student-faculty ratio, faculty salary, and university’s educational expense per student were not casually (directly) related to student retention; and suggested that to increase student retention universities should increase the student selectivity. Sanjeev & Zytkow (1995) used 49er, a pattern discovery process developed by Żytkow & Zembowicz (1993), to find patterns in the form of regularities from student databases related to retention and graduation. The authors found that academic performance in high school was the best predictor of persistence and better performance in college, and that the high school GPA was a better predictor than the ACT composite score. In addition, they found that no amount of financial aid influenced students to enroll for more terms. 3 Massa & Puliafito (1999) applied Markov chains modelling technique to create pre- dictive models for the student dropout problem. By tracking the students for 15 years, the authors created state variables for the number of exams appeared, average marks obtained, and the continuation decision. Using data mining, Stewart & Levin (2001) studied the effects of student characteristics to persistence and success in an academic program at a community college. They found that the student’s GPA, cumulative hours attempted, and cumulative hours completed were the significant predictors of persistence, and that young males were a high risk group. Veitch (2004) used decision trees (CHAID) to study the high school dropouts. Using 25-fold cross-validation, the overall misclassification rates of drop-outs who were classified as non-dropouts were 15.79% and 10.36%. In this study, GPA was the most significant predictor of persistence. Salazar et al. (2004) used clustering algorithms and C4.5 to study graduate student retention at Industrial University of Santander, Colombia. The authors found that the high marks in the national pre-university test predicted a good academic performance, and that the younger students had higher probabilities of a good academic performance. Barker et al. (2004) used neural networks and Support Vector Machines (SVM) to study graduation rates; the first-year advising center (University College at University of Oklahoma) collected data via a survey given to all incoming freshman. It is worthwhile to note that Barker et al. (2004) excluded all the missing data from the study, which constituted for approximately 31% of the total data. Overall misclassification rate was approximately 33% for various dataset combinations. The authors used principal com- ponent analysis to reduce the number of variables from 56 to 14, however, reported that the results using the reduced datasets were “much worse” than the complete datasets. Superby et al. (2006) applied discriminant analysis, neural networks, random forests, and decisions trees to survey data at the University of Belgium to classify new students in low-risk, medium-risk, and high-risk categories. The authors found that the scholastic history and socio-family background were the most significant predictors of risk. The overall classification rates for decision trees, random forests, neural networks, and linear discriminant analysis were 40.63%, 51.78%, 51.88%, and 57.35% respectively. Using the National Student Clearinghouse (NSC) data, Sujitparapitaya (2006) differ- entiated between stopout, retained, and transfer students. The overall classification rates for the validation sets using logistic regression, neural networks, C5.0 were 80.7%, 84.4%, and 82.1% respectively. Herzog (2006) used American College Test’s (ACT) student pro- file section data, NSC data, and the institutional student information system data for comparing the results from the decision trees, the neural networks and logistic regression to predict retention and degree-completion time. The author substituted mean average ACT scores for missing scores. Decision trees created using C5.0 performed the best with 85% correct classification rate for freshmen retention, 83% correct classification rate for degree completion time (three years or less), 93% correct classification rate for degree completion time (six years or more ) for the validation datasets. Atwell et al. (2006) used University of Central Florida’s student demographic and survey data to study the retention problem with the help of data mining. In this study, university retained approximately 82% of the freshmen from the study, and it used 285 variables to create data mining models. The authors used nearest neighbor algorithm to impute more than 60% observations with missing values. Using decision trees with the entropy split criterion, the authors obtained precision of 88% for the not-retained 4 outcome using the test data, and the actual retention rate for this test data set was 82.61%. Yu et al. (2007) studied the data from Arizona State University using decision trees, and included variables, such as demographic, pre-college academic performance indica- tors, current curriculum, and academic achievement. Some of the important predictor variables were accumulated earned hours, in-state residence, and on campus living. To study the retention problem using data mining for the admissions data, DeLong et al. (2007) applied various attribute evaluation methods, such as Chi-square gain, gain ratio, and information gain, to rank the attributes. In addition, the authors tested various classifiers, such as näıve Bayes, AdaBoost M1, BayesNet, decision trees, and rules, and noted that AdaBoost M1 with Decision Stump classifier performed the best in terms of precision and recall, hence, used this classifier for further experimentation. The authors balanced the class variable (retained and not retained) and obtained over 60% classification rates for both retained and not retained outcome. The authors concluded that the number of programs that the student applied to that specific institution and the student’s order of program admit preference were the most significant predictors of retention. Pittman (2008) compared various data mining techniques (artificial neural networks, logistic regression, Bayesian Classifiers, and decision trees) applied to the student re- tention problem, and also used attribute evaluators to generate rankings of important attributes. The author concluded that logistic regression performed the best in terms of ROC-curve area. [Table 1 about here.] 2.2. Assessing the State of the Art Table 1 lists techniques used in the studied literature, where the cohort sizes were available, along with the reported performance measures. Clearly, there is much room for improvement in the current state of the art: • It is a recommended data mining practice to divide the data into a train and test set, learn on the train set, then assess the learned theory on the test set (Witten & Frank, 2005). Otherwise, if a theory is tested via the data used to build that theory, this test can over-estimate theory performance. For example, the Glynn et al. result of Table 1 seems impressive (a 83% accuracy on a data set with 49.08% a retention rate); however, that result should be repeated using some hold-out test set. • All the regression studies from 1971 to 1999 report R2 values under 0.6. This R2 value is a measure of how well future outcomes are likely to be predicted by the model. The maximum value of R2 is one and R2 values under 0.6 indicate very weak predictive abilities. • The accuracy reports are very close to the ZeroR theoretically lower-bound on performance. ZeroR is a baseline classifier that simply returns the majority class. For example, Herzog studied a data set with a 83.5% retention rate (see Table 1). ZeroR, applied to this data set, would be correct in 83.5% of cases. Therefore, the 85.4% accuracy of Herzog’s data miners is very close to the ZeroR lower-bound; 5 i.e. the sophisticated analysis of that paper could be very nearly replicated using the dumbest of learners (ZeroR). The last three results of Table 1 do not report their accuracies. However, these can be calculated in the following way. Let A, B, C, D be the true negatives, false negatives, false positives, and true positives respectively of a predictor that some student will attend some year of university. From Zhang & Zhang (2007) and Menzies et al. (2007), we say that (A,B,C,D) can be used in the following performance measures: pd = recall = D(B + D) (1) pf = false alarm = C/(A + C) (2) prec = precision = D/(C + D) (3) acc = accuracy = (A + D)/(A + B + C + D) (4) neg/pos = (A + C)/(B + D) (5) Note that all these performance measures assess subtly different aspects of the perfor- mance of data miner: • “Recall” measures how much of the target was found. • The “false alarm” rate measures what fraction of non-targets triggered the learned theory. • “Precision” comments on how many targets are found in the data selected by the theory. • “Accuracy” comments on how many of the targets and non-targets were accurately labeled by the learned theory. In an ideal result, we can obtain high recall, low false alarms, high precision, and high ac- curacies. However, as discussed by Zhang & Zhang (2007) and Menzies et al. (2007), these values are inter-related. Hence, the ideal result is not possible. These inter-relationships are shown below:( prec = D D + C = 1 1 + C D = 1 1 + neg pos · pf recall ) ⇒ ( pf = pos neg · (1 − prec) prec · recall ) (6) If a publication misses a particular performance measure, it is possible to use these equations to infer the missing value. For example: D = recall ∗pos (7) C = pf ∗neg (8) A = C ∗ 1/(pf − 1) (9) acc = (A + D)/(neg + pos) (10) Using these equations, we can comment that the last three results of Table 1 can be significantly improved: • In Atwell et al. (2006), the the precision varied from 73% to 88%. Using our equations, we can estimate false alarm values (pf) ranging from 2% to 8% (assuming recall values of 65% to 90%). In our experience, it is very rare to achieve such very low false alarm rates, especially from noisy data relating to student retention. Hence, the Atwell et al., results are somewhat surprising. 6 • In DeLong et al. (2007), the precision varied from 57% to 60%. From our equations, we can estimate their false alarm rates in the range of 49% to 63% (assuming recall values of 65% to 90%). Such high false alarm rates are deprecated. • In Pittman (2008), the reported precision varied from 44% to 63%. Our equations comment that these values are numerically unobtainable. For 0.78 ≤ acc ≤ 0.81, neg = 17139 and pos = 21136 −neg, the equations only solve for prec ≤ 50. That is, half the precision values reported by Pittman need to be reviewed. In summary, learning predictors for student retention is very difficult. Despite decades of work, there is considerable room for improvement in the methods used to find patterns in student retention. As we show below, such improvements are possible if we augment standard data miners with some extra pre-processors. 3. Data Data used in this study were from a mid-size public university, and were extracted from the student information system on official census dates. These data consisted all first-year freshmen’s demographic, academic, and financial aid information (more than 100 attributes), as of the census reporting dates (after two weeks of semester starting date). As the higher education administrators may design effective policies when the students begin their studies, it is important to note that our emphasis was on detecting patterns based only on the first-term data, and that too only beginning of the term data. We created three dependent variables: RET1, if the student returned after one year; RET2, if the student returned after two years; and RET3, if the student returned after three years. The overall distribution of these dependent variables is given in Table 2. For the studied time period, the overall first-year retention rate was 71.3%, the second-year persistence rate was 60.4%, and the third-year persistence rate was 54.8%. [Table 2 about here.] In the Integrated Postsecondary Education Data System (IPEDS), for the U.S. only, degree-granting, Doctoral degree offering, 4-year and above institutes (excluding Univer- sity of Phoenix-Online Campus), and cohort size greater than 3,000, we found that the full-time freshmen retention rate had a range from 59% to 96%, and the cohort size had a range from 3,117 to 8,025. In this list of institutions, Kent state university ranked 38 in the full-time retention percentage and 26 in the cohort size (Department of Education, 2010). Thus, Kent state data are representative of other similar size universities, and the data mining approach could be generalized to other universities. 3.1. Attribute Groups The data mining methods discussed below used attribute selection to prune measure- ments that are poor predictors for the target class. Therefore, our data miners can be used to assess various hypotheses relating to student retention: • If an hypothesize claims that attributes X, Y, Z are important... • ... and if our learners prune those attributes ... 7 • ... then that is evidence against that hypothesis. Accordingly, before applying our data miners, we take care to divide our attributes into the active hypothesis that they supprot: H1: The financial aid hypothesis. Sanjeev & Zytkow (1995) found that no amount of financial aid influenced students to enroll for more terms; whereas Herzog (2005) found that upper-income students had reduced dropout odds compared to those from middle and lower incomes. According to John (2000), “the research litera- ture remains ambiguous” regarding the influence financial aid on recruitment and retention. H2: The academic performance hypothesis. Although there is no doubt that high school GPA and high school preparedness has a significant impact on persistence, re- searchers have often questioned the effects of standardized college entrance exam- inations (ACT/SAT). Waugh et al. (1994) found that SAT and ACT scores had no relationship with retention, whereas Murtaugh et al. (1999) found that SAT scores had some predictive value, although inferior compared to high school GPA. DesJardins et al. (2002) noted that high GPA lowered the risk of dropout, but the effect diminished over time, and that the financial aid was an insignificant factor for increasing graduation, however, it indeed reduced the student stopout. In their comprehensive literature review, Lotkowski et al. (2004) found that high school GPA had the strongest relationship with college retention in the academic factors, but ACT assessment scores had a moderate impact. H3: The faculty tenure and experience hypothesis. Ehrenberg & Zhang (2005) found that for every 10 percentage point increase in the percentage of part-time faculty and not on tenure-track full-time faculty, there was a 3-5 percentage point reduc- tion in the institution’s graduation rate. Jacoby (2006) found similar results at community colleges that increase in the ratio of part-time faculty had a negative impact on the graduation rates. In the sequel, we will return to these hypotheses to comment on which were most useful for predicting student retention. Table 7 lists attributes that we grouped together under each hypothesis. [Table 3 about here.] 4. Building the Experiment In Section 2.2, we assert that a good data miner should do better than the simplistic ZeroR learner. Table 2 tell us that that lower bound is: • For first year retention: 71.3%. • For second year retention: 60.4%. • For third year retention: 54.8%. As discussed below, we will be able to do much better than some, but not all, of these targets. This was achieved by 8 • Removing spurious attributes using feature subset selection; • Exploring a large range of classifiers; • Assessing the learned theories by their variance, as well as their median perfor- mance. • Assessing the learned theories by their variance, as well as their median perfor- mance. • Study the delta of student factors between those who stay and those who are re- tained. 4.1. Feature Subset Selection Table 7 shows a sample of the 103 attributes used in this study. Our pre-experimental suspicion was that some of the attributes were “noisy”; i.e. contain signals not related to the target of prediction retention. Therefore, before we learn a theory, we first explored attribute selection. Note that the number of attributes to select is crucial in the analysis of the data, because it allows us to comment on the hypotheses shown in the last section. If removal of attributes from an hypothesis does not change the performance of the prediction, then that hypothesis is spurious. In this experiment, we ranked the 103 attributes from most informative to least in- formative. We then built theories using the top n ∈{5, 10, .., 100, 103} ranked attributes. Attributes were then discarded if adding them in did not improve the performance of our retention predictors. The attributes were ranked using one of four methods: CFS, Information Gain, chi- squared, and One-R. Correlation-based feature selection constructs a matrix of feature to feature, and feature-to-class correlations (Hall, 2000). CFS uses a best first search by expanding the best subsets until no improvement is made, in which case the search falls to the unexpanded subset having the next best evaluation until a subset expansion limit is met. Information Gain uses an information theory concept called entropy. Entropy mea- sures the amount of uncertainty, or randomness, that is associated with a random vari- able. Thus, high entropy can be seen as a lack of purity in the data. Information gain, as described in Mitchell (1997) is an expected reduction of the entropy measure that occurs when splitting examples in the data using a particular attribute. Therefore an attribute that has a high purity (high information gain) is better at describing the data than the one that has a low purity. The resulting attributes are then ranked by sorted their information gain scores in a descending order. The chi-squared statistic is used in statistical tests to determine how distributions of variables are different from one another (Moore & Notz, 2006). Note that these variables must be categorical in nature. Thus, the chi-squared statistic can evaluate an attribute’s worth by calculating the value of this statistic with respect to a class. Attributes can then be ranked based on this statistic. The One-R classifier, described below, can be used to deliver top-ranking attributes. One-R constructs and scores rules using one attribute. Feature selectors using One-R sort the attributes based on these scores. 9 4.2. Classifiers In data mining, classifiers are used to learn connections between independent features and the dependent feature (called the class). Once these patterns are learned, we can predict outcomes in new data by reflecting on data that has already been examined. This study tried six different classifiers: One-R, C4.5, ADTrees, Naive Bayes, Bayes networks, and radial bias networks. These are some of the well-known and standard classifiers in the machine learning field, except for ADTrees. One-R, described in Holte (1993), builds rules from the data by iteratively examining each value of an attribute and counting the frequency of each class for that attribute-value pair. An attribute-value is then assigned as the most frequently occurring class. Error rates of each of the rules are then calculated, and the best rules are ranked based on the lowest error rates. A radial basis function network (RBFN) is an artificial neural network (ANN) that utilizes a radial basis function as an activation function (Bors, 2001). An ANN’s acti- vation function is used in order to offer non-linearity to the network. This is important for multi-layer networks containing many hidden layers, because their advantages lie in their ability to learn on non-linearly separable examples. C4.5 (Quinlan, 1993) is an extension to the ID3 (Quinlan, 1986) algorithm. A decision tree (shown in Figure 3) is constructed by first determining the best attribute to make as the root node of the tree (Mitchell, 1997). ID3 decides this root attribute by using one that best classifies training examples based upon the attribute’s information gain (described above) (Quinlan, 1986). Then, for each value of the attribute representing any node in the tree, the algorithm recursively builds child nodes based on how well another attribute from the data describes that specific branch of its parent node. The learning stops when the tree perfectly classifies all training examples, or when all attributes used. C4.5 extends ID3 by making several improvements, such as the ability to operate on both continuous as well as discrete attributes, training data that contains missing values for a given attribute(s), and employ pruning techniques on the resulting tree. [Figure 3 about here.] ADTrees are decision trees that contain both decision nodes, as well as prediction nodes (Freund & Mason, 1999). Decision nodes specify a condition, while prediction nodes contain only a number. Thus, as an example in the data follows paths in the ADTree, it only traverses branches whose decision nodes are true. The example is then classified by summing all prediction nodes that are encountered in this traversal. ADTrees, however, differ from binary classification trees, such as C4.5, where those trees only traverses a single path down the tree. [Figure 4 about here.] A naive Bayes classifier uses Bayes’ theorem to classify training data. Bayes’ the- orem, as shown in Equation 11, determines the probability P of an event H occurring given an amount of evidence E. This classifier assumes feature independence; the algo- rithm examines features independently to contribute to probabilities, as opposed to the assumption that features depend on other features. Surprisingly, even though feature in- dependence is an integral part of the classifier, it often outperforms many other learners (Rish; Domingos & Pazzani, 1997). 10 Pr(H|E) = Pr(E|H) ∗Pr(H) Pr(E) (11) Bayesian networks, illustrated in Figure 4, are graphical models that use a directed acyclic graph (DAG) to represent probabilistic relationships between variables. As stated in Heckerman (1996), Bayesian networks have four important elements to offer: 1. Incomplete data sets can be handled well by Bayesian networks. Because the networks encode a correlation between input variables, if an input is not observed, in will not necessarily produce inaccurate predictions, as would other methods. 2. Causal relationships can be learned about via Bayesian networks. For instance, we can find whether a certain action taken would produce a specific result and to what degree. 3. Bayesian networks promote the amalgamation of data and domain knowledge by allowing for a straightforward encoding of causal prior knowledge, as well as the ability to encode causal relationship strength. 4. Bayesian networks avoid over fitting of data, as “smoothing” can be used in a way such that all data that is available can be used for training. 4.3. Cross-Validation The value of different attributes can be assessed using equations one to four. If we use multiple hold out test sets, we can also discover the variance in these performance figures. In this experiment, we performed a 5 × 5 cross-validation i.e. we partitioned the data five times into a testing set consisting of 1 5 -th of the data and a training set of 4 5 -ths of the data. After the five rounds, we recorded the median values of recall and false alarm rates. 4.4. Contrast Set Learning After determining the subset of the attributes that best predict for student retention, we conducted a contrast set study. Contrast set learners like TAR3 (Menzies & Hu, 2007) seek attribute ranges that are most different in various outcomes. One way to read these contrast sets are as treatments that promise if action X was applied to a domain, then this would favor outcome X over outcome Y. In our case, we used TAR3 in two ways: • Firstly, we will use TAR3 to find which treatments most select for retention; • Secondly, we will run TAR3 in the opposite direction to find the treatments that most select for students leaving university. . In the first case, TAR3 is being used to find actions that most encourage retention. In the second case, TAR3 is being used to find the worst possible actions that most increase the probability of a student leaving. 11 5. Analysis of Experimental Results 5.1. Evaluation Metrics The evaluation metrics used in this experiment are standard data mining performance measures of a method. They are: • Probability of detection (PD); • Probability of false alarm (PF); • And variance PD and PF seen over the our cross-validation study. Variance in these values provides insight into how much reliability a classifier supports on the data. For example, if a method’s PD values ranges from very low to very high, we can conclude that the particular method is inconsistent in its probabilities of detection. For our studies, we rejected anything with a variance greater than ±25%. The above statistics were collected over 1500 experiments, which were repeated 20 times (to check for conclusion stability). In all, we conducted 5 ∗ 5 ∗ 4 ∗ 6 ∗ 3 ∗ 20 = 36, 000 experiments; i.e. 5 × 5 cross-validation using four feature subset selectors and 6 different learners, for the 3 data sets of Section 3 (recall from §3 that those three data sets contained data about first, second, and third year retention). This was repeated 20 times using the top n ∈ 5, 10, 15, .., 100, 103 attributes as found by the feature selector. 5.2. First Results After rejecting all results with (1) a PD lower than the ZeroR limit; (2) a PD variance greater than ±25%; and (3) a PF higher than 25%, we found that we had no predictors for Year1 or Year2 retention. This is the first major finding for this research: it is very difficult to predict lower year retention. Note that this result is consistent with prior results discussed above in our literature review. For the rest of this study, we will focus only on third year retention. The case for focusing on third year retention is quite clear: • If the goal is to provide a complete university education for a student, then pre- dicting survival till second year is less interesting than lasting till third year. • Third year retention implies second and first year retention. 5.3. Ranking with the Mann-Whitney Test After pruning results with low PD, high PF, or high PD variance, we ranked the remaining results via a Mann-Whitney test (95% confidence). We determined the ranks by counting how many times a combination won compared to another combinations. The method that won the most number of times was then given the highest rank. The table in Figure 4 shows the top ten ranking combinations based on a PD performance measure. Note: we gave identical ranks to those treatments whose win value was equal in magnitude. 12 [Table 4 about here.] Since similar results were achieved using 30 or 50 attributes, we applied Occam’s Razor and focused on the 30 attributes found to be best for oneR/bnet. For these 30 attributes, we studied all their ranges. Figure 4 shows the ranges which, in isolation, select for retention at a probability greater than the ZeroR limit for (for third year, that ZeroR limit is 55%). In terms of assessing different hypothesis, the third column of Figure 4 is most informative: • The ranges shown at the top of the table are most predictive for third year retention. Note the dominance of “Financial Aid” attributes from Figure 5. • Attributes related to student “Performance” are rarer. • None of the attribute ranges include the “Faculty Type and Experience” attributes of Figure 5. From this analysis, we made two tentative conclusions: • Using experienced faculty-level instructors is not predictive for third year retention. • Issues relating to financial aid dominate over student performance. 5.4. Ranking with Contrast Set Learning The counter-case to this conclusion might be that Figure 4 only discusses the effect of attribute ranges in isolation. It is possible that combination of factors might lead to different conclusions. The TAR3 treatment learner was used to test this possibility. We let TAR3 build rules of up to size 10 (i.e. ten combinations of attribute ranges) from the 30 attributes selected by the best learning combination of Figure 4. It turned out that this max size of 10 ranges was much larger than necessary: TAR3 never found combinations larger than three ranges. [Table 5 about here.] 6. Results Figure 5 lists the rankings of all attribute ranges which, in isolation, predict for third year retention at a probability higher than the ZeroR limit (55%), and are supported by good number of records. The top six attributes affecting third-year retention were from the financial aid hypothesis: student’s wages, parent’s adjusted gross income, student’s adjusted gross income, mother’s income, father’s income, and high school percentile. Of those students who reported their wages, students who made between 7,850 and 9,958 had a 79% retention. Similar rules were found for parent’s income and adjusted gross income. It means that the students with stronger financial support usually stay in college than the students with weaker financial support. After these top six attributes, high school percentile of 81 or greater was an important attribute with 69% of students returning after three years. Some other “performance” attributes were ACT scores and ranks. This supports the argument that scores do have some predictability of student retention. 13 TAR3 results, given in Figure 5, produced simple theories (treatments) that combined ranges of various attributes that maximized the student retention. For example, the student retention was very high for students with the AGI in the range from $7,000 to $724,724 and father’s wages were in the range from $56,289 to $999,999. One more interesting theory that predicted high retention was where father’s education level was 3 (college) and student’s rank amongst the freshmen cohort was between 66.3 and 98.4. Treatments that predicted student drop-out were based on the total number of classes student was enrolled, English 10000, an introductory college writing and supplemental instruction class, and on-campus living. Students who took less than five class, enrolled in the English 10000 class, and did not live on-campus were at high risk of drop-out. Chart on the bottom of Figure 5 shows the retention percentage of each treatment. For example, students enrolled in English 10000 had a 40% retention in their third year. Key findings were: • Student’s and parent’s income capacity and levels affected student retention. Third- year retention was higher for the students with high income than the students with low income. According to treatment 1, approximately 82% of students who had at least $7,000 AGI and their fathers’ income was at least $56,289 returned after three years. Similarly, according to treatment 5, approximately 79% of students who made at least $5,383 and their parents’ AGI was at least $87,744 returned after three years. • Students with better high school performance amongst their peers had higher chances of retention. According to treatment 2, approximately 81% of students who had at least $7,000 AGI and had high school percentile of 72 and better re- turned after three years. Approximately 79% students who had at least 3.34 HS GPA and whose parents had an AGI of at least $84,744 stayed after three years, given in treatment 4. • ACT scores, rank of these scores amongst peers, and COMPASS scores affected student retention. Students with higher scores and rank had higher chances of retention. According to treatment 3, approximately 80% of students who had at least $7,000 AGI and had ACT math score of 21 or better returned after three years. Similarly, 77% of students who had at least 23 in ACT composite (or SAT equivalent) and had an income of at least $5,383 and less than $561,500 returned after three years, given in treatment 6. • Parent’s education level had a positive effect on student retention. Students whose parents did not attend college had a lower retention compared to students whose parents did attend college. As given in treatments 7 and 10, a student was highly likely (77%) to return after three years: (7) if the mother of that student attended college, the student had a ACT composite score of 22 or better, the parents’ AGI was at least $84,744; (10) if the father of that student attended college and the student’s percentile rank amongst other freshmen in the cohort was at least 66.3. • Enrolling in fewer classes (less than five), enrolling in English 10000 (an introduc- tory college writing class), and living off-campus had a negative effect on student retention, as given in treatments 11, 12, and 13. It is important to note that enrolling in that English course itself is not a predictor of non-retention, but the 14 sample of the students that attended this class were at high-risk of dropping out. Given funding for further investigation, we would focus more data collection on this high-risk group. [Figure 5 about here.] 6.1. Strategic Actions This study provides insights in student retention domain using beginning of term data. These insights can be used to design effective policies and strategic actions, such as: • Most of the attributes were related to socio-economic levels and capacities of stu- dents and their parents; however, this cannot be controlled while admitting stu- dents, but better support programs and calculated financial-aid packaging for stu- dents with lower economic capacities can be created. • First-year students should be encouraged to live on-campus by providing some incentives, as on-campus students have higher chances of retention. • Special guidance and supplemental instruction in writing and reading should be provided to first-generation students. Parents of first-year generation students have considerably low-incomes than the parents of non-first-generation students, and according to the results of this study, income of parents is a critical factor in student retention even if the students had similar academic performance. • Students are placed in the supplemental instruction classes, such as English 10000, based on their COMPASS and ACT scores. As these students’ scores indicated lack of academic preparedness in some areas, academic advisers correctly place students in such classes; however, if the students fail or perform poorly in such classes, it leaves a lasting impression and sets the students to for future drop-out, even after three years. Therefore, it is paramount that advisers not only place students in supplemental instruction classes, but also ensure the success of students in these classes and improve the skills that students lack. Out of all classes considered in this study, English seemed to have the greatest impact. Intuitive as it may be, to succeed in college, students need good writing and reading skills. 7. Conclusion Although our techniques could not predict first or second year retention with sig- nificantly higher accuracies than the baseline, these techniques obtained probability of detection approximately 15% higher for the class value of Y and 20% higher for the class value of N than the baseline percentages for third-year retention, based on the first-year beginning of the term data. In the studied literature, we have not found any studies with such a significant improvement over the baseline for the third-year retention. In addition, if policies are designed to improve third-year retention rate (using this predictive model), not only will they improve first and second year retention rates, but also the six-year graduation rates. 15 For the studied institution, family background and family’s social-economic status are critical for student’s third-year persistence. Using feature subset selection methods, we found that the attributes from the “financial aid” hypothesis were selected the most as predictors of retention, and although the attributes from the “performance” hypothesis were selected, their predictability, in isolation, was lesser than the attributes from the “fi- nancial aid” hypothesis. None of the attributes from the “faculty tenure and experience” were selected by the feature subset selectors. These results could very well be true only for the studied institution; however, if the approach detailed in this study is followed, other institutions can find top performing classifier and important attributes. We recommend: (a) data discretization; (b) feature subset selection with cross-validation and evaluation the performance over various learn- ers; (c) treatment learners, such as TAR3 to find succinct strategic actions in complex data. We welcome the opportunity to study data from other institutions and willing to share the experiment platform used in this study. References ACT (2007). ACT National Collegiate Retention and Persistence to Degree Rates. http://www.act.org/research/policymakers/reports/retain.html. Adam, A. J., & Gaither, G. H. (2005). Retention in higher education: A selective resource guide. New Directions for Institutional Research, 2005 , 107–122. Atwell, R. H., Ding, W., Ehasz, M., Johnson, S., & Wang, M. (2006). Using data mining techniques to predict student development and retention. In Proceedings of the National Symposium on Student Retention. Barker, K., Trafalis, T., & Rhoads, T. R. (2004). Learning from student data. Systems and Information Engineering Design Symposium, (pp. 79–86). Bean, J. P. (1980). Dropouts and turnover: The synthesis and test of a causal model of student attrition. Research in Higher Education, 12 , 155–187. Bors, A. (2001). Introduction of the radial basis function (rbf) networks. In Online Symposium for Electronics Engineers (pp. 1–7). volume 1. Bresciani, M. J., & Carson, L. (2002). A study of undergraduate persistence by unmet need and percentage of gift aid. NASPA Journal , 40 , 104–123. DeLong, C., Radcliffe, P. M., & Gorny, L. S. (2007). Recruiting for retention: Using data mining and machine learning to leverage the admissions process for improved freshman retention. In Proceedings of the National Symposium on Student Retention. Department of Education (2010). Integrated Postsecondary Education Data System (IPEDS). http://nces.ed.gov/ipeds/datacenter/. DesJardins, S. L., Ahlburg, D. A., & McCall, B. P. (2002). A temporal investigation of factors related to timely degree completion. The Journal of Higher Education, 73 , 555–581. Dey, E. L., & Astin, A. W. (1993). Statistical alternatives for studying college student retention: A comparative analysis of logit, probit, and linear regression. Research in Higher Education, 34 , 569– 581. Domingos, P., & Pazzani, M. J. (1997). On the optimality of the simple bayesian classifier under zero-one loss. Machine Learning , 29 , 103–130. Druzdzel, M. J., & Glymour, C. (1994). Application of the TETRAD II program to the study of student retention in u.s. colleges. In Working notes of the AAAI-94 Workshop on Knowledge Discovery in Databases (KDD-94) (pp. 419–430). Seattle, WA. Ehrenberg, R., & Zhang, L. (2005). Do Tenured and Tenure-Track Faculty Matter? Journal of Human Resources, 40 , 647. Freund, Y., & Mason, L. (1999). The alternating decision tree learning algorithm. In In Machine Learning: Proceedings of the Sixteenth International Conference (pp. 124–133). Morgan Kaufmann. Glynn, J., Sauer, P., & Miller, T. (2003). Signaling student retention with prematriculation data. NASPA Journal , 41 , 41–67. 16 Hall, M. (2000). Correlation-based Feature Selection for Discrete and Numeric Class Machine Learning. In Proceedings of the Seventeenth International Conference on Machine Learning (ICML-2000): June 29-July 2, 2000, Stanford University (p. 359). Morgan Kaufmann. Heckerman, D. (1996). A Tutorial on Learning With Bayesian Networks. Technical Report Learning in Graphical Models. Herzog, S. (2005). Measuring determinants of student return vs. dropout/stopout vs. transfer: A first- to-second year analysis of new freshmen. Research in Higher Education, 46 , 883–928. Herzog, S. (2006). Estimating student retention and degree-completion time: Decision trees and neural networks vis--vis regression. New Directions for Institutional Research, 131 . Holte, R. (1993). Very simple classification rules perform well on most commonly used datasets. Machine Learning , 11 , 63. Jacoby, D. (2006). Effects of part-time faculty employment on community college graduation rates. Journal of Higher Education, 77 , 1081–1103. John, E. P. (2000). The impact of student aid on recruitment and retention: What the research indicates. New Directions for Student Services, (pp. 61–76). Lau, L. K. (2003). Institutional factors affecting student retention. Education, 124 , 126–137. Lotkowski, V., Robbins, S., & Noeth, R. (2004). The role of academic and non-academic factors in improving college retention. ACT Office of Policy Research, . Massa, S., & Puliafito, P. (1999). An application of data mining to the problem of the university students’ dropout using markov chains. In Principles of Data Mining and Knowledge Discovery. Third European Conference, PKDD’99 (pp. 51–60). Prague, Czech Republic. Menzies, T., Dekhtyar, A., Distefano, J., & Greenwald, J. (2007). Problems with precision. IEEE Transactions on Software Engineering , . http://menzies.us/pdf/07precision.pdf. Menzies, T., & Hu, Y. (2007). Just enough learning (of association rules): The TAR2 treatment learner. In Artificial Intelligence Review . Available from http://menzies.us/pdf/07tar2.pdf. Mitchell, T. M. (1997). Machine Learning . New York: McGraw-Hill. Moore, D., & Notz, W. (2006). Statistics: concepts and controversies. WH Freeman & Co. Murtaugh, P. A., Burns, L. D., & Schuster, J. (1999). Predicting the retention of university students. Research in Higher Education, 40 , 355–371. NCPPHE (2007). Retention rates - first-time college freshmen returning their second year (ACT). Pascarella, E. T., & Terenzini, P. T. (1979). Interaction effects in spady and tinto’s conceptual models of college attrition. Sociology of Education, 52 , 197–210. Pascarella, E. T., & Terenzini, P. T. (1980). Predicting freshman persistence and voluntary dropout decisions from a theoretical model. The Journal of Higher Education, 51 , 60–75. Pittman, K. (2008). Comparison of data mining techniques used to predict student retention. Ph.D. thesis Nova Southeastern University. Quinlan, J. R. (1986). Induction of decision trees. (1st ed.). Quinlan, J. R. (1993). C4.5: Programs for Machine Learning (Morgan Kaufmann Series in Machine Learning). (1st ed.). Morgan Kaufmann. Rish, I. (). An empirical study of the naive bayes classifier. In IJCAI-01 workshop on ”Empirical Methods in AI”. http://www.intellektik.informatik.tu-darmstadt.de/ tom/IJCAI01/Rish.pdf. Salazar, A., Gosalbez, J., Bosch, I., Miralles, R., & Vergara, L. (2004). A case study of knowledge discovery on academic achievement, student desertion and student retention. Information Technology: Research and Education, 2004. ITRE 2004. 2nd International Conference on, (pp. 150–154). Sanjeev, A., & Zytkow, J. (1995). Discovering enrolment knowledge in university databases. In First International Conference on Knowledge Discovery and Data Mining (pp. 246–51). Montreal, Que., Canada. Scalise, A., Besterfield-Sacre, M., Shuman, L., & Wolfe, H. (2000). First term probation: models for identifying high risk students. In 30th Annual Frontiers in Education Conference (pp. F1F/11–16 vol.1). Kansas City, MO, USA: Stripes Publishing. Spady, W. G. (1970). Dropouts from higher education: An interdisciplinary review and synthesis. Interchange, 1 , 64–85. Spady, W. G. (1971). Dropouts from higher education: Toward an empirical model. Interchange, 2 , 38–62. Stage, F. (1989). Motivation, Academic and Social Integration, and the Early Dropout. American Educational Research Journal , 26 , 385–402. Stewart, D. L., & Levin, B. H. (2001). A model to marry recruitment and retention: A case study of prototype development in the new administration of justice program at blue ridge community college. Sujitparapitaya, S. (2006). Considering student mobility in retention outcomes. New Directions for 17 Institutional Research, 2006 . Superby, J. F., Vandamme, J. P., & Meskens, N. (2006). Determination of factors influencing the achievement of the first-year university students using data mining methods. In 8th International Conference on Intelligent Tutoring Systems (ITS 2006) (pp. 37–44). Jhongli, Taiwan. Terenzini, P. T., & Pascarella, E. T. (1980). Toward the validation of tinto’s model of college student attrition: A review of recent studies. Research in Higher Education, 12 , 271–282. Tillman, C., & Burns, P. (2000). Presentation on First Year Experience. http://www.valdosta.edu/ cgtillma/powerpoint.ppt. Tinto, V. (1975). Dropout from higher education: A theoretical synthesis of recent research. Review of Educational Research, 45 , 89–125. Tinto, V. (1982). Limits of Theory and Practice in Student Attrition. The Journal of Higher Education, 53 , 687–700. Tinto, V. (1988). Stages of student departure: Reflections on the longitudinal character of student leaving. Journal of Higher Education, 59 , 438–455. Vandamme, J. (2007). Predicting Academic Performance by Data Mining Methods. Education Eco- nomics, 15 , 405–419. Veitch, W. R. (2004). Identifying characteristics of high school dropouts: Data mining with a decision tree model. Waugh, G., Micceri, T., & Takalkar, P. (1994). Using ethnicity, SAT/ACT scores, and high school GPA to predict retention and graduation rates. Witten, I., & Frank, E. (2005). Data Mining: Practical Machine Learning Tools and Techniques. (2nd ed.). San Francisco: Morgan Kaufmann Publishers. Yu, C. H., DiGangi, S., Jannasch-Pennell, A., Lo, W., & Kaprolet, C. (2007). A data-mining approach to differentiate predictors of retention between online and traditional students. Zhang, H., & Zhang, X. (2007). Comments on ’data mining static code attributes to learn defect predictors’. IEEE Transactions on Software Engineering , . Żytkow, J., & Zembowicz, R. (1993). Database exploration in search of regularities. Journal of Intelligent Information Systems, 2 , 39–81. 18 Figure 1: BA Degree Completion Rates for the period 1880 to 1980, where Percent Completion is the Number of BAs Divided by the Number of First-time Degree Enrollment Four Years Earlier (Tinto, 1982) 19 Figure 2: Percentage of First-Year Students at Four-Year Colleges Who Return for Second Year (ACT, 2007) 20 Figure 3: A decision tree consists of a root node and descending children nodes who denote decisions to make in the tree’s structure. This tree, for example, was constructed in an attempt to optimize investment portfolios by minimizing budgets and maximizing pay-offs. The top-most branch represents the best selection in this example. 21 Figure 4: In this simple Bayesian network, the variable Sprinkler is dependent upon whether or not it’s raining; the sprinkler is generally not turned on when it’s raining. However, either event is able to cause the grass to become wet - if it’s raining, or if the sprinkler is caused to turn on. Thus, Bayesian networks excel at investigating information relating to relationships between variables. 22 # Treatment 1 7000 ≤ FinAidSTUDENT AG < 724,724 and 56,289 ≤ FinAidFATHER WAG < 999,999 2 7,000 ≤ FinAidSTUDENT AG < 724,724 and HS PERCENT ≥ 72 3 7,000 ≤ FinAidSTUDENT AG < 724,724 and 21 ≤ ACT1 MATH < 36 4 84,744 ≤ FinAidPARENT AGI < 999,999 and HS GPA ≥ 3.34 5 84,744 ≤ FinAidPARENT AGI < 999,999 and 5383 ≤ FinAidSTUDENT WA < 561,500 6 23 ≤ MaxACT < 35 and 5383 ≤ FinAidSTUDENT WA < 561,500 7 22 ≤ ACT1 COMP < 35 and 84,744 ≤ FinAidPARENT AGI < 999,999 and FinAidMOTHER ED=3 8 5383 ≤ FinAidSTUDENT WA < 561,500 and 21 ≤ ACT1 MATH < 36 9 HS GPA ≥ 3.34 and 32,570 ≤ FinAidMOTHER WAG < 533,395 10 FinAidFATHER ED=3 and 66.3 ≤ PercentileRankHSGPA < 98.4 11 1 ≤ TotalClass≤ 5 12 ENG10=Y 13 LIVE.ON.CAMP=N 30 40 50 60 70 80 0 1 2 3 4 5 6 7 8 9 10 11 12 13 Treatment # %stay Figure 5: Treatments 1 to 10 are the top ten treatments found by this analysis that increases the third year retention rates. Treatments 11,12,13 are the worst three treatments found by this analysis that most decrease the third year retention rates. The effects of each treatment, is shown on the bottom plot. 23 A u t h o r ( Y e a r ) N o t e s C o h o r t S iz e R e t a in e d ( # ) R e t a in e d ( % ) M e a s u r e o f A c c u - r a c y C o e ff e s U s e d ? T e c h n iq u e s U s e d S p a d y (1 9 7 1 ) 6 8 3 6 1 5 9 0 .0 4 % R 2 o f .3 1 3 2 fo r m e n a n d .3 8 7 9 fo r w o m e n Y e s M u ltip le re g re ssio n B e a n (1 9 8 0 ) 9 0 6 7 6 9 8 4 .8 8 % R 2 o f .2 2 fo r w o m e n a n d 0 .0 9 fo r m e n Y e s M u ltip le re g re ssio n T e re n z in i (1 9 8 0 ) stu d y 1 3 7 9 6 0 1 5 .8 0 % R 2 o f .2 4 6 Y e s d isc rim in a te a n a ly se s stu d y 3 5 1 8 4 2 8 8 2 .6 3 % R 2 o f .2 5 6 Y e s M u ltip le re g re ssio n stu d y 5 7 6 3 6 7 3 8 8 .2 0 % R 2 o f .3 0 9 Y e s d isc rim in a te a n a ly se s stu d y 6 7 6 3 6 7 3 8 8 .2 0 % R 2 o f .4 7 6 fo r m e n a n d .5 5 3 fo r w o m e n Y e s d isc rim in a te a n a ly se s S ta g e (1 9 8 9 ) 3 2 3 2 9 4 9 1 .0 0 % Y e s L o g istic re g re ssio n D e y & A stin (1 9 9 3 ) 9 4 7 1 5 2 1 6 .0 0 % M u ltip le R 0 .3 5 4 , 0 .3 5 1 , a n d 0 .3 2 3 Y e s lo g it, p ro b it, a n d re g re ssio n M u rta u g h e t a l. (1 9 9 9 ) 8 6 6 7 5 2 0 0 6 0 % e stim a te d re t p ro b 5 9 .3 % Y e s S u rv iv a l A n a ly sis/ H a z a rd re - g re ssio n B re sc ia n i & C a rso n (2 0 0 2 ) 3 5 3 5 3 1 2 1 8 8 .3 0 % R 2 o f 0 .0 2 2 Y e s L o g istic re g re ssio n G ly n n e t a l. (2 0 0 3 ) a n y d ro p o u t; n o t o n ly fi rst-y e a r; a c c u ra c ie s b a se d o n th e tra in in g d a ta 3 2 4 4 1 5 9 2 4 9 .0 8 % o v e ra ll a c c u ra c y o f 8 3 % Y e s L o g istic re g re ssio n H e rz o g (2 0 0 5 ) 5 2 6 1 4 0 1 4 7 6 .3 0 % 7 7 .4 % a c c u ra c y Y e s L o g istic re g re ssio n 4 2 9 8 3 3 1 4 7 7 .1 0 % Y e s 4 6 7 1 4 0 4 0 8 3 .5 0 % 8 5 .4 % a c c u ra c y Y e s S u jitp a ra p ita y a (2 0 0 6 ) 2 ,4 4 4 1 9 4 3 7 9 .5 0 % 8 1 .6 % a c c u ra c y o n tra in in g ; 8 0 .7 % o n v a lid a tio n Y e s L o g istic re g re ssio n 2 ,4 4 5 1 9 9 4 7 9 .5 0 % 8 3 .9 % a c c u ra c y o n tra in in g ; 8 2 .1 % o n v a lid a tio n N e u ra l N e tw o rk 2 ,4 4 5 1 9 9 4 7 9 .5 0 % 8 5 .5 % o n tra in in g ; 8 4 .4 % o n v a lid a tio n C 4 .5 H e rz o g (2 0 0 6 ) 8 ,0 1 8 6 0 3 7 7 5 .2 9 % a c c u ra c y c lo se to 7 5 % N e u ra l N e tw o rk s; C H A ID , C 4 .5 , C R & T ; L o g istic re g re s- sio n A tw e ll e t a l. (2 0 0 6 ) tra in in g 3 ,8 2 9 3 1 4 9 8 2 .2 4 % p re c isio n fo r d ro p -o u ts 9 1 , 8 4 , 8 4 , 7 8 d e c isio n tre e s (e n tro p y , ch i-sq , g in i) a n d lo g istic re g re ssio n te st 5 ,9 9 0 4 ,8 8 1 8 1 .4 9 % p re c isio n fo r d ro p -o u ts 8 8 , 8 2 ,8 2 ,7 3 D e L o n g e t a l. (2 0 0 7 ) 5 0 % p re c isio n v a rie d fro m 5 7 % to 6 0 % A d a B o o st M 1 w ith D e c isio n S tu m p s P ittm a n (2 0 0 8 ) 2 1 1 3 6 1 7 1 3 9 8 1 .1 0 % o v e ra ll a c c u ra c y o f 7 8 - 8 1 % ; n o t-re ta in e d p re - c isio n fro m 4 4 -6 3 % L o g istic re g re ssio n , n e u ra l n e t- w o rk , b a y e s, J 4 8 T a b le 1 : T e ch n iq u e s a n d A c c u ra c ie s R e p o rte d in L ite ra tu re 24 RET1 RET2 RET3 Count Percentage Count Percentage Count Percentage retained=Y 24,039 71.3% 18,055 60.4% 14,362 54.8% retained=N 9,673 28.7% 11,857 39.6% 11,854 45.2% Total 33,712 100% 29,912 100% 26,216 100% Table 2: Distribution of Dependent Variables 25 C a t e g o r y F in a n c ia l A id C a t e g o r y P e rfo rm a n c e In d ic a to rs C a t e g o r y F a c u lty T y p e & E x p e rie n c e A t t r ib u t e D e s c r ip t io n A t t r ib u t e D e s c r ip t io n A t t r ib u t e D e s c r ip t io n F in A id A w a rd T y p e G F in a n c ia l a id a m o u n t o f g ra n ts A C T C O M P A C T c o m p re h e n s iv e s c o re (o ld ) F a c E x p L T 1 C n t C o u n t o f c o u rs e s ta u g h t b y fa c u lty [C C T F ] w / le s s th a n 1 y r e x p e rie n c e F in A id A w a rd T y p e J F in a n c ia l a id a m o u n t o f jo b s A C T E N G L A C T e n g lis h s c o re (o ld ) F a c E x p L T 1 R a tio R a tio o f c o u rs e s ta u g h t b y fa c u lty [R C T F ] w / le s s th a n 1 y r e x p e rie n c e to th e to ta l c o u rs e s F in A id A w a rd T y p e L F in a n c ia l a id a m o u n t o f lo a n s A C T M A T H A C T m a th s c o re (o ld ) F a c E x p 1 to 5 C n t C C T F w / e x p e rie n c e b e tw e e n 1 & 5 F in A id A w a rd T y p e S F in a n c ia l a id a m o u n t o f s c h o la r- s h ip A C T 1 C O M P A C T c o m p re h e n s iv e s c o re (n e w ) F a c E x p 1 to 5 R a tio R C T F w / e x p e rie n c e b e tw e e n 1 & 5 to th e to ta l c o u rs e s F in A id A w a rd T y p e W F in a n c ia l a id a m o u n t o f w a iv e r A C T 1 E N G L A C T e n g lis h s c o re (n e w ) F a c E x p 6 to 1 0 C n t C C T F w / e x p e rie n c A e b e tw e e n 6 & 1 0 F in A id D E P E N D E N C Y D e p e n d e n c y s ta tu s A C T 1 M A T H A C T m a th s c o re (n e w ) F a c E x p 6 to 1 0 R a tio R C T F w / e x p e rie n c e b e tw e e n 6 & 1 0 to th e to ta l c o u rs e s F in A id F A T H E R E D F a th e r’s e d u c a tio n le v e l A C T E Q U IV A C T e q u iv a le n t o f th e s a t s c o re F a c E x p 1 1 to 1 5 C n t C C T F w / e x p e rie n c e b e tw e e n 1 1 & 1 5 F in A id F A T H E R W A G F a th e r’s in c o m e M a x A C T M a x o f A C T s c o re a n d A C T e q u iv a - le n t F a c E x p 1 1 to 1 5 R a tio R C T F w / e x p e rie n c e b e tw e e n 1 1 & 1 5 to th e to ta l c o u rs e s F in A id M O T H E R E D M o th e r’s e d u c a tio n le v e l C O M P R E A D C o m p a s s re a d s c o re F a c E x p 1 6 to 2 0 C n t C C T F w / e x p e rie n c e b e tw e e n 1 6 & 2 0 F in A id M O T H E R W A G M o th e r’s in c o m e C O M P W R IT E C o m p a s s w rite s c o re F a c E x p 1 6 to 2 0 R a tio R C T F w / e x p e rie n c e b e tw e e n 1 6 & 2 0 to th e to ta l c o u rs e s F in A id O ff e re d In d F in a n c ia l a id o ff e re d in d ic a to r S A T T O T S A T to ta l s c o re F a c E x p 2 1 to 2 5 C n t C C T F w / e x p e rie n c e b e tw e e n 2 1 & 2 5 F in A id P A R E N T A G I P a re n t’s a d ju s te d g ro s s in c o m e S A T V E R B S A T v e rb a l s c o re F a c E x p 2 1 to 2 5 R a tio R C T F w / e x p e rie n c e b e tw e e n 2 1 & 2 5 to th e to ta l c o u rs e s F in A id P A R E N T H O U P a re n ts ’s h o u s e h o ld s iz e H S C O D E H ig h s c h o o l c o d e F a c E x p 2 5 to 3 0 C n t C C T F w / e x p e rie n c e b e tw e e n 2 5 & 3 0 F in A id P A R E N T M A R P a re n ts ’s m a rtia l s ta tu s H S G P A H ig h s c h o o l g p a F a c E x p 2 5 to 3 0 R a tio R C T F w / e x p e rie n c e b e tw e e n 2 5 & 3 0 to th e to ta l c o u rs e s F in A id P A R E N T T A X P a re n ts ’s ta x fo rm ty p e H S P E R C E N T H ig h s c h o o l p e rc e n tile F a c E x p G T 3 1 C n t C C T F w / e x p e rie n c e g re a te r th a n 3 1 y e a rs F in A id S P O U S E W A G S p o u s e ’s w a g e s H S R A N K H ig h s c h o o l ra n k F a c E x p G T 3 1 R a tio R C T F w / e x p e rie n c e g re a te r th a n 3 1 y e a rs to th e to ta l c o u rs e s F in A id S T U D E N T A G S tu d e n t’s a d ju s te d g ro s s in c o m e H S S IZ E H ig h s c h o o l c la s s s iz e N o T e n u re F a c C n t C o u n t o f c o u rs e s ta u g h t b y n o ra n k fa c u lty F in A id S T U D E N T H O S tu d e n t’s h o u s e h o ld s iz e R a n k H S G P A P e rc e n tile o f h s g p a a m o n g a ll fre s h - m e n N o T e n u re F a c R a tio R a tio o f c o u rs e s ta u g h t b y n o ra n k fa c u lty to th e to ta l c o u rs e s F in A id S T U D E N T M A S tu d e n t’s m a rtia l s ta tu s R a n k M a x A C T P e rc e n tile o f m a x A C T a m o n g a ll fre s h m e n N T T F a c C n t C o u n t o f c o u rs e s ta u g h t b y n tt fa c u lty F in A id S T U D E N T T A S tu d e n t’s ta x fo rm ty p e A N T H 1 8 E n ro lle d in a n th ro p o lo g y c o u rs e N T T F a c R a tio R a tio o f c o u rs e s ta u g h t b y n tt fa c u lty to th e to ta l c o u rs e s F in A id S T U D E N T W A S tu d e n t’s w a g e B S C I1 0 E n ro lle d in b io lo g ic a l s c ie n c e c o u rs e T T F a c C n t C o u n t o f c o u rs e s ta u g h t b y te n u re d / te n u re - tra c k fa c u lty F irs tG e n In d F irs t g e n e ra tio n in d ic a to r C H E M 1 0 E n ro lle d in c h e m is try c o u rs e T T F a c R a tio R a tio o f c o u rs e s ta u g h t b y te n u re d / te n u re - tra c k fa c u lty to th e to ta l c o u rs e s T o ta lF in A id O ff e re d T o ta l fi n a n c ia l a id o ff e re d E N G 1 0 E n ro lle d in E n g lis h c o u rs e E N G 1 1 E n ro lle d in E n g lis h c o u rs e G E O L 1 1 E n ro lle d in g e o lo g y c o u rs e L E S T 1 6 E n ro lle d in le is u re s tu d ie s c o u rs e M A T H 1 0 E n ro lle d in m a th 1 0 0 le v e l c o u rs e M A T H 1 1 E n ro lle d in m a th 1 1 0 le v e l c o u rs e M A T H 1 2 E n ro lle d in m a th 1 2 0 le v e l c o u rs e M A T H 1 4 E n ro lle d in m a th 1 4 le v e l c o u rs e P H Y 1 1 E n ro lle d in p h y s ic s 1 1 le v e l c o u rs e P E P 1 5 E n ro lle d in p h y s ic a l e d 1 5 le v e l c o u rs e T a b le 3 : L ist o f A ttrib u te s b y S ta te d H y p o th e se s 26 Rank Number of Attributes FSS Classifier 61 30 oneR bnet 61 50 cfs adtree 57 50 oneR adtree 56 30 oneR adtree 55 30 cfs adtree 52 50 oneR bnet 51 30 infogain adtree 51 30 cfs bnet 48 50 infogain adtree Table 4: The top ten ranking treatments for third year retention. Ranks represent how many times a particular treatment wins over all other treatments in the experiment. 27 # P (Ret3|X) Support X (Feature = Range) (percent) (#students) Hypothesis Feature Range 1 79 1,752 Financial Aid Student’s Wage 7850 to 9958 2 73 3,751 Financial Aid Parent’s Adjusted Gross Income 96636 to inf 3 71 4,152 Financial Aid Student’s Adjusted Gross Income 4830 to 7916 4 71 3,148 Financial Aid Mother’s Income 42957 to inf 5 70 5,873 Financial Aid Father’s Income 52366 to inf 6 69 5,622 Financial Aid Student’s Wage 4093 to 7851 7 69 5,838 Performance High School Percentile 81 to inf 8 68 2,523 Financial Aid Student’s Dependency Status I 9 68 7,502 Financial Aid Father’s Education Level 3 10 67 6,045 Financial Aid Parent’s Adjusted Gross Income 58551 to 96636 11 66 12,215 Financial Aid Student’s Tax Form 2 12 66 7,710 Financial Aid Mother’s Education Level 3 13 66 2,057 Financial Aid Student’s Wage 1.5 to 1000 14 66 10,370 Financial Aid First Generation Student N 15 65 7,082 Performance ACT Math Score (new) 23 to inf 16 65 2,780 Financial Aid Student’s Adjusted Gross Income 3336 to 4830 17 65 4,676 Performance ACT English Score (new) 25 to inf 18 65 5,669 Performance ACT Comprehensive Score (new) 24 to inf 19 65 13,101 Financial Aid Parent’s Tax Form 1 20 65 11,328 Financial Aid Parent’s Marital Status M 21 64 6,658 Performance Percentile Of Max ACT Among Freshmen 71 to inf 22 64 6,952 Performance Max Of ACT Score And ACT Equivalent 24 to inf 23 64 2,697 Financial Aid Student’s Tax Form 1 24 63 17,254 Financial Aid Student’s Marital Status U 25 63 13,063 Financial Aid Mother’s Wages -inf to 42957 26 63 3,126 Financial Aid Parent’s Tax Form 2 27 63 1,022 Financial Aid Student’s Adjusted Gross Income 16714 to inf 28 62 15,154 Financial Aid Dependency D 29 62 9,459 Financial Aid Father’s Income -inf to 52366 30 61 8,792 Financial Aid Mother’s Education Level 2 31 61 8,461 Financial Aid Father’s Education Level 2 32 61 4,176 Financial Aid Student’s Wage 1904 to 4093 33 60 14,523 Total Enrolled Hours 15 to 19 34 60 2,540 Financial Aid Student’s Adjusted Gross Income 1895 to 3336 35 59 3,780 Performance Compass Writing Score -inf to 10 36 59 8,271 Performance ACT English Score (new) 20 to 25 37 59 7,311 FirstGenInd Y 38 59 6,980 Performance HS PERCENT 61 to 81 39 59 15,021 Performance Total Number of Enrolled Classes 6 to inf 40 59 5,598 Financial Aid Parent’s Adjusted Gross Income 1838 to 58551 41 58 3,127 Financial Aid Parent’s Marital Status S 42 58 8,667 Performance ACT Composite 20 to 24 43 58 4,767 Performance ACT Math Score (new) 20 to 23 44 58 13,887 Performance Compass Writing Score 74 to inf 45 58 5,769 Performance High School GPA 3.02 to 3.4 46 58 10,281 Performance RankMaxACT 31 to 71 47 58 10,044 Performance MaxACT 20 to 24 48 57 20,087 On-Campus Indicator Y 49 57 11,36 Financial Aid Father’s Education Level 4 50 56 24,826 Age of Student at Matriculation -inf to 19.5 51 56 24,407 Performance Enrolled in English Courses N 52 56 3,660 Performance Percentile Of Hs Gpa Among Freshmen 46 to 60 Table 5: Ranking all attribute ranges which, in isolation, predict for third year retention at a probability higher than the ZeroR limit (55%). From the above, the strongest predictor for third year retention is a student’s wage (at 79%). On the other hand, the bottom line of this table says that the percentile of a student amongst their Freshmen cohort is little better than ZeroR (at 56%). 28 
work_z44fob6fbrealdoyutwfhy76xq	----	DokuWiki Setup Error DokuWiki Setup Error The datadir ('pages') at ./data/pages is not found, isn't accessible or writable. You should check your config and permission settings. Or maybe you want to run the installer? 
work_z4vwlccvbjgj7klff2mgsr3fgq	----	wp-p1m-38.ebi.ac.uk Params is empty 404 sys_1000 exception wp-p1m-38.ebi.ac.uk no 220990217 Params is empty 220990217 exception Params is empty 2021/04/06-03:05:33 if (typeof jQuery === "undefined") document.write('[script type="text/javascript" src="/corehtml/pmc/jig/1.14.8/js/jig.min.js"][/script]'.replace(/\[/g,String.fromCharCode(60)).replace(/\]/g,String.fromCharCode(62))); // // // window.name="mainwindow"; .pmc-wm {background:transparent repeat-y top left;background-image:url(/corehtml/pmc/pmcgifs/wm-nobrand.png);background-size: auto, contain} .print-view{display:block} Page not available Reason: The web page address (URL) that you used may be incorrect. Message ID: 220990217 (wp-p1m-38.ebi.ac.uk) Time: 2021/04/06 03:05:33 If you need further help, please send an email to PMC. Include the information from the box above in your message. Otherwise, click on one of the following links to continue using PMC: Search the complete PMC archive. Browse the contents of a specific journal in PMC. Find a specific article by its citation (journal, date, volume, first page, author or article title). http://europepmc.org/abstract/MED/ 
work_z6umnwkf6fhnrkoaouebhnmpca	----	doi:10.1016/j.eswa.2008.06.098 Expert Systems with Applications xxx (2008) xxx–xxx ARTICLE IN PRESS Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a A Neural Network theory based expert system – ICRT Shieh-Shing Lin a,*, Shih-Cheng Horng b, Ch0i-Hsin Lin c a Department of Electrical Engineering, St. John’s University, 499, Sec. 4 Tam King Road, Tamsui, Taipei 25135, Taiwan b Department of Computer Science and Information Engineering, Chaoyang University of Technology, Taichung, Taiwan c Department of Electronics Engineering, Kao Yuan University, Kaoshiung, Taiwan a r t i c l e i n f o Keywords: Neural Network theory Micro-chip processor technology Expert system Cars repairing tools 0957-4174/$ - see front matter � 2008 Elsevier Ltd. A doi:10.1016/j.eswa.2008.06.098 * Corresponding author. Tel.: +886 2 28013131x65 E-mail addresses: sslin@mail.sju.edu.tw (S.-S. Lin) (S.-C. Horng), chsinlin@cc.kyu.edu.tw (C.-H. Lin). Please cite this article in press as: Lin, S.-S (2008), doi:10.1016/j.eswa.2008.06.098 a b s t r a c t In this paper, we focus on the frequent glitches of general family cars. Based on the Neural Network the- ory and combining with the Micro-chip processor technology, we design an expert system-‘‘Integrated Cars Repairing Tools” (ICRT) and the size of the proposed ICRT is quite small and achieves the following attractive functions: (1) Portable Electric-jack, (2) Electric Tire-pressure-meter, (3) efficiently Pneumatic tire machine, (4) LCD type Sparing Battery, (5) trouble blinking sign and lighting device. Compare to the traditional single function tools – ‘‘Hand Jack”, ‘‘Mechanic tire-pressure-meter”, ‘‘Pressure-stat pneumatic tire machine”, and ‘‘Hand-held glitch indicator”, the proposed innovative expert system – ICRT has the following special features: the Portable Electric-jack can easily and rapidly lift cars, the Electric Tire-pres- sure-meter works precisely, the efficiently pneumatic tire machine works rapidly, the LCD type Sparing Battery with the capability to show the residual capacity of Power and can be used to start cars easily. We install the proposed ICRT in several different types of cars and make numerous simulations, the test results show that the ICRT is efficient for rejecting the frequent glitches of cars. � 2008 Elsevier Ltd. All rights reserved. 1. Introduction Cars are one of the most important traffic tools for people in the real world. In this paper, we focus on the frequent glitches of the general cars; such as, the drivers cannot start cars or the pressure of the Tire is inefficiency for driving; in general, in the first case, probably, the Battery box of the car is leaking or out of order, the second, the Tire of the car is already broken or not working for a few days ago or yesterday. There could be the worse case happen about while the Tire of the car is broken during the driving. In gen- eral, while the Tire broken accident occurred, the drivers will call the Car Rescuing Center for help or use the Sparing Tire in car to replace the broken Tire by himself, since, almost it is true that there is a ‘‘Hand Jack” in every car. However, if the accident is oc- curred in midnight, sometimes it is hard to call some one to help the drivers, moreover, if the pressure of the Sparing Tire is ineffi- cient for driving since it have been long time no used stayed in car, then it should have some more necessary device (such as the Pressure-stat output pneumatic tire machine) to inject the air and stabilize the pressure of the Tire. In addition, there still have many glitches of the family cars always trouble the drivers, such as; the car cannot be started in the morning (probably, the Battery box of the car is leaking or out of order); the pressure of the Tire is ll rights reserved. 00; fax: +886 2 28013973. , schong.ece90g@nctu.edu.tw . et al., A Neural Network th inefficient for driving, . . ., etc. Therefore, it is necessary for drivers to have more than one or many scatter car repairing devices in cars (such as, the Power bank, the Pneumatic tire machine, the ‘‘Hand Jack”, . . ., etc.) to handle the unexpected accidents occurred during the driving. These devices always occupy some accommodations and are dispersed in the cars. Furthermore, the more unexpected things to bother the drivers are that the Tire is leaking air or broken while encountered some stringer stone during the driving, espe- cially, for the female driver. About the research of the car repairing equipments for solving the problem of the inefficiency of the pres- sure of the Tire, the traditional device-the Pressure-stat output pneumatic tire machine is used to inject the air into the Tire with slowly speed. And most of the traditional devices cannot show the actual pressure values of the Tire during the operation. Several papers that are concerned with the single function tools corresponding to the car repairing device, such as, the Tire-pres- sure-meter function tools have been published (Grossmann, 1999; Kowalewski, 2004; Li, Wang, Qunzhi, & Shan, 2003; Nabipoor & Majlis, 2004; Pohl, Ostermayer, Reindl, & Seifert, 1997), the Jack function tools (Azad, Khajepour, & McPhee, 2005; Chapman, 2005; Harter, 2000; Mihali & Sobh, 1999; Wallis, Ronnquist, Jarvis, & Lucas, 2002; Yamada & Suzuki, 1990), the Lighting function tools (Kornev, Kuchma, Pokrovski, & Soms, 2005; Lie, Yu, et al., 2000; Liu, 2005), the Pneumatic-tire-machine function tools (Dai-Dai; Audi-Tai; An-Hun), and the Power bank and Battery charging function tools (Chen & Lin, 1997; Li & Liu, 2001; Liu, 2001; Yeh, 1997). However, there are few documents eory based expert system – ICRT, Expert Systems with Applications mailto:sslin@mail.sju.edu.tw mailto:schong.ece90g@nctu.edu.tw mailto:chsinlin@cc.kyu.edu.tw http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 2 S.-S. Lin et al. / Expert Systems with Applications xxx (2008) xxx–xxx ARTICLE IN PRESS and/or devices related to how to improve the performance of the existed single function cars repairing tools and integrate many of these scatter functions tools into a compacted expert system. About the research of the equipments of Jack device for lifting cars to replace the Tire, the traditional device – ‘‘Hand Jack” is used to lift cars for changing the broken Tire; however, it is not so easy to operate, especially, for the female drivers in the dark night. In the traditional car caring center, the oil-type Jack or the electrician lifting machine (Azad et al., 2005; Chapman, 2005; Harter, 2000; Mihali & Sobh, 1999; Wallis et al., 2002; Yamada & Suzuki, 1990) are used and can easily lift cars for cars maintains. However, these devices cannot be carried in cars, since it is too much room con- suming and very heavy. Nowadays, since the applications of the Neural Network Theory technologies have become more mature as well as the Micro-chip processor, the integration management and control technology for the scatter repairing tools of cars seems to be a trend in the car repairing industrials. The object of this paper is based on the corresponding theory of Neural Network and combining with the Micro-chip processor technology (Carstens, 1993; Comer & Comer, 2003; Laker & Sansen, 1994; MacKenzie, 1999) to design an innovative expert system – ICRT. The major contribution of this expert system is to improve the performance of the existed single function of cars repairing tools and integrate many of these scatter cars repairing tools into a single compacted expert system – ‘‘Integrated Cars Repairing Tools” (ICRT). Furthermore, the size of the proposed innovative expert sys- tem – ICRT is not only quite small and portable but also easily oper- ated and installed in cars. The innovative expert system – ICRT achieves the following attractive functions: (1) Portable Electric-jack, (2) Electric Tire-pressure-meter, (3) Efficiently pneumatic tire machine, (4) LCD type Sparing Battery with the capability to show the residual capacity of Power, (5) Trouble blinking sign and Lighting device. The paper is organized in the following manner. Section 2 pre- sents the combining with the Micro-chip processor technology and the Neural Network theory based innovative expert system – ICRT. In Section 3, we present the Hardware Implementation of the ICRT. The simulation tests those are used to demonstrate the perfor- mance of the ICRT are given in Section 4. Finally, we make a brief conclusion in Section 5. 2. Neural Network theory based innovative expert system – ICRT We uses the corresponding theory of Neural Network and combing with the Micro-chip processor technologies to improve the performance of the existed single function Cars Repairing Tools and integrate many of these scatter cars repairing tools into a single compacted expert system – ICRT. The functions of the pro- posed ICRT is electrical and automatically, and can achieve quite a few time and money saving for drivers encountered some unex- pected problems. The function of ICRT includes: Portable Elec- tric-jack, Electric Tire-pressure-meter, Efficiently pneumatic tire machine, LCD type Sparing Battery, Trouble blinking sign and Lighting device. The special features of the ICRT are stated in the following: A. Portable Electric-jack: In the Hardware Implementation of the Portable Electric-jack device of the ICRT, the Robotics and Feedback Control Theory of Neural Network is used to lock the working current of the Servo-Motor within some certain level to maintain the altitude of the car being lifted by the Please cite this article in press as: Lin, S.-S. et al., A Neural Network th (2008), doi:10.1016/j.eswa.2008.06.098 Portable Electric-jack while the changing of the broken Tire. This is the famous effect called the Neural Network lock-in effect. B. Electric Tire-pressure-meter: The Electric Tire-pressure-meter of the ICRT can precisely measure the pressure of the Tire and show the measured values in digitations. C. Efficiently pneumatic tire machine: Based on the measured values by the Electric Tire-pressure-meter, the Efficiently pneumatic tire machine of the ICRT uses the Neural Network theory which is embedded in the Micro-chip processor to stabilize the pressure of the Tire and will be described latter. D. LCD type Sparing Battery: The Sparing Battery of the ICRT uses the LCD device to show the residual capacity of Power. E. Trouble blinking sign and Lighting device: The Trouble blinking sign and Lighting device of the ICRT provides the necessary lighting service for the drivers to maintain the car while the unexpected events occurred. 2.1. Neural Network theory based Tire-Pressure-Stabilizer In the Hardware Implementation of the Efficiently pneumatic tire machine device of the ICRT, the Neural Network based modi- fied Proportional-Integral-Derivative (P.I.D.) Theory is used to sta- bilize the pressure of the Tire by the Efficiently pneumatic tire machine. The Neural Network based modified P.I.D. Theory which is embedded in the Micro-chip processor and associated with the proposed Electric Tire-pressure-meter are used to formulate the Tire-Pressure-Stabilizer system. Block diagram of the Tire-Pres- sure-Stabilizer is shown in Fig. 1. In the following, we will present the Neural Network Theory based Tire-Pressure-Stabilizer to stabi- lize the pressure of Tire. The detailed description of the block sym- bols in Fig. 1 is listed below: #1: Denotes the block diagram of the Neural Network based Tire-Pressure-Stabilizer system. #2: Denotes the ‘‘Efficiently pneumatic tire machine”. #3: Denotes the ‘‘Electric Tire-pressure-meter”. #4: Denotes the ‘‘Feedback Buffer” which is used to calculate the steady values of the Tire pressure originated from the ‘‘Elec- tric Tire-pressure-meter” and transform it to the ‘‘Micro- chip Processor”. #5: Denotes the ‘‘Micro-chip Processor” which is used to process the received feedback signals and create the optimal control signals. #6: Denotes the ‘‘Motor” which is used to speed/slow the ‘‘Effi- ciently pneumatic tire machine”. Before introducing the complete algorithm, we need the follow- ing notations: Ps: Denotes the pre-set standard Tire-pressure values. Pr: Denotes the real time Tire-pressure values. Sh, Sm, Ss: Denote the highly, medium, and steady control signals of the ‘‘Micro-chip Processor” # 5. fh, fm, fs: Denote the highly, medium, and steady output signals of the ‘Motor’ #6. C: Denotes the criteria of the deviation of the Tire-pressure values. e: Denotes a small positive real value. 2.2. Neural Network Theory based Tire-Pressure-Stabilizer algorithm Step 0 Initial set the parameters: C, e. Step 1 Input the preset Tire-pressure values of Ps. eory based expert system – ICRT, Expert Systems with Applications 5 6 32 4 1 Motor Efficiently pneumatic tire machine Electric tire- pressure- meter Out Real time Tire-pressure Feedback Buffer Input Preset Tire-pressure Micro-chip Processor + - Fig. 1. Block diagram of the Neural Network theory based Tire-Pressure-Stabilizer. S.-S. Lin et al. / Expert Systems with Applications xxx (2008) xxx–xxx 3 ARTICLE IN PRESS Step 2 Detecting the Tire-pressure from the Electric Tire-pres- sure-meter, #3 and sends it to the Feedback Buffer, #4. Step 3 Calculate the real Tire-pressure value Pr from the Feed- back Buffer, #4. Step 4 Calculate the deviation of Tire-pressure DP = |Ps � Pr| by the Micro-chip Processor. If DP P C then go to Step 5; if e 6 DP < C then go to Step 6; if DP < C then go to Step 7. Step 5 The Micro-chip Processor, #5 sends the control signal Sh to the Motor, #6, and the unit #6 process it and sends output signal fh to drive the Pneumatic tire machine, #2 and go to Step 3. Step 6 The Micro-chip Processor, #5 sends the control signal Sm to the Motor, #6, and the unit # 6 process it and sends the output signal fm to drive the Pneumatic tire machine, #2 and go to Step 3. Step 7 The Micro-chip Processor, #5 sends the control signal Ss to the Motor, #6, and the unit #6 process it and sends the output signal fs to drive the Pneumatic tire machine, #2, and stop. Remark 1. The efficiency of the Neural Network theory based Tire-Pressure-Stabilizer algorithm can be seen from Step 5 to Step 7. While the ‘‘Micro-chip Processor” #5 receives the feedback signal, Pr, of the form |{Ps � Pr}| P C (it represents the deviation between the real-time Tire-pressures and the standard values is bigger), then the ‘‘Micro-chip Processor” #5 exports the Sh (highly control signal) to drive the ‘Motor’, #6. While ‘Motor’, #6 receives the highly control signal Sh, it sends the highly output signal fh to drive the ‘‘Efficiently pneumatic tire machine”, #2 to stabilize the Tire-pressure of the car. While the ‘‘Micro-chip Processor” #5 receives the feedback signal, Pr, of the form (e 6 |{Ps � Pr}| < C (it represents the deviation between the real-time Tire-pressures and the standard values is not so bigger), then the ‘‘Micro-chip Processor” #5 exports the Sm (medium control signal) to drive the ‘Motor’, #6. While ‘Motor’, #6 receives the medium control signal Sm, it sends the medium output signal fm to drive the ‘‘Efficiently pneumatic tire machine”, #2 to stabilize the Tire-pressure of the car. While the ‘‘Micro-chip Processor” #5 receives the feedback signal, Pr, of the form |Ps � Pr| < e (it represents the real-time Tire- pressures reaches the steady status), then the ‘‘Micro-chip Processor” #5 exports the Ss (steady control signal) to drive the ‘Motor’, #6. While ‘Motor’, #6 receives the Ss steady control signal, it sends the steady output signal fs to drive the ‘‘Efficiently pneumatic tire machine”, #2 to stabilizing operating or stop operating. Remark 2. In this paper, we also use the similar Neural Network Theory based Tire-Pressure-Stabilizer algorithm in the Portable Electric-jack device for locking the working current of the Servo-Motor within some certain level to maintain the altitude of the car being lifted. Please cite this article in press as: Lin, S.-S. et al., A Neural Network th (2008), doi:10.1016/j.eswa.2008.06.098 In the following, we will present the Hardware Implementation of the Neural Network theory based the innovative expert system – ICRT and use the associated figures and the corresponding photo- graphs to make a detailed description of the proposed ICRT. 3. Hardware implementation of the innovative expert system – ICRT 3.1. Diagram description Function block diagram and System structure diagram of Hard- ware implementation of the innovative expert system – ICRT are shown in Figs. 2 and 3. The definitions of the block element sym- bols are stated in the following: 10: Denotes the Control circuit. 12: Denotes the Micro-chip processor. 14: Denotes the Voltage-regulating circuit. 16: Denotes the Driving circuit of DC-Motor. 18: Denotes the Time delay circuit. 19: Denotes the Circuit of the Trouble blinking sign. 20: Denotes the Portable Electric-jack. 30: Denotes the Electric Tire-pressure-meter. 40: Denotes the Efficiently pneumatic tire machine. 50: Denotes the Sparing Battery. 60: Denotes the LCD displayer. 70: Denotes the Trouble blinking sign and Lighting device. 100: Denotes the Neural Network Theory based innovative expert system – ICRT. 3.2. Hardware implementation of the innovative expert system – ICRT We will combine the block element symbols and their defini- tions associated with the Figs. 2 and 3 to describe the operation se- quence control of the Hardware implementation of the innovative expert system – ICRT. 3.2.1. The detailed description the combinations and interconnections of the block element symbols In Fig. 2, the Neural Network theory based the innovative expert system – ICRT 100, includes the following devices: the Portable Electric-jack 20, the Electric Tire-pressure-meter 30, the Efficiently pneumatic tire machine 40, the Sparing Battery 50, the LCD type displayer 60, and the Trouble blinking sign and Lighting device 70, in which the Control circuit 10 is including the following de- vices: the Micro-chip processor 12, the Voltage-regulating circuit 14, the Driving circuit of DC-Motor 16, the Time delay circuit 18, and the Circuit of the Trouble blinking sign 19, in which the Mi- cro-chip processor 12 is connected with the portable Electric-jack 20 by the Driving circuit of DC-Motor 16 and the Micro-chip pro- eory based expert system – ICRT, Expert Systems with Applications Electric Tire-Pressure-Meter Efficiently Output Pneumatic Tire Machine LCD Displayer Driving Circuit of DC-Motor Time Delay Circuit Circuit of the Trouble Blinking sign Micro-Chip Processor Voltage- regulating Circuit Trouble Blinking Sign and Lighting Device Portable Electric-jack Sparing Battery 70 50 20 10 19 18 16 14 12 40 60 30 100 Fig. 2. Function block diagram of the hardware implementation of the innovative expert system – ICRT. Integrated Car Repairing Tools ICRT Trouble Blinking Sign and Lighting Service Inefficient Tire-Pressure B rok en T ire I n efficient Power of Battery Night Lighting Service Rel ay Trouble Blinking Standard Tire- Pressure Values LCD Dis- player E l ectri c T i re- Pressure M e ter Efficiently Output Pneumatic Tire Machine Electric- Jack Sero- Motor C a nnot Power T h e C a r The LCD Type Sparing Battery Changing Tire Fig. 3. System structure diagram of hardware implementation of the innovative expert system – ICRT. 4 S.-S. Lin et al. / Expert Systems with Applications xxx (2008) xxx–xxx ARTICLE IN PRESS cessor 12 is connected with the Sparing Battery 50 by the Voltage- regulating circuit 14; the Micro-chip processor 12 is connected with the Circuit of the Trouble blinking sign 19 by the Time delay circuit 18; beside, the Control circuit 10 is connected with the Elec- tric Tire-pressure-meter 30, the Efficiently pneumatic tire machine 40, the LCD type displayer 60, and the Trouble blinking sign and the Lighting device 70. 3.2.2. The operation sequence control of the hardware implementation of the innovative expert system – ICRT 3.2.2.1. Portable Electric-jack. The Portable Electric-jack 20 of the ICRT 100 uses the Neural Network based Robotics and Feedback Control Theory to lock the working current of the DC-Motor 16 within some certain range. Therefore, the Portable Electric-jack 20 can maintain the height of the car being lifted while changing the broken Tire. This is the famous effect called the Neural Network lock-in Effect and the detailed description is stated below. In the ICRT, the Driving circuit of DC-Motor 16 uses the DC current to drive DC-Motor. And the DC-Motor is used to drive the Portable Electric-jack 20; the Geared system associated with the Gear-dri- ven are included in 20 and are used to slow down the speed of the DC-Motor to facilitate the Control circuit 10 to make a miner adjusting of the operations of the DC-Motor. The Neural Network based Robotics and Feedback Control Theory is embedded in the Micro-chip processor 12 and is used to lock the working current of the DC-Motor 16 within some certain level to maintain the working position of Car being lifted up and down by the Portable Electric-jack 20. Please cite this article in press as: Lin, S.-S. et al., A Neural Network th (2008), doi:10.1016/j.eswa.2008.06.098 3.2.2.2. Electric Tire-pressure-meter. The Electric Tire-pressure-me- ter 30 of the ICRT 100 uses the Tire pressure sensors and the A/D Converter associated with the Micro-chip processor 12 to measure the pressure of the Tire and shows the measured values in digitations. 3.2.2.3. Efficiently pneumatic tire machine. The Neural Network based modified P.I.D. control Theory (the Neural Network theory based Tire-Pressure-Stabilizer algorithm) is embedded in the Mi- cro-chip processor 12 and is used in the Efficiently pneumatic tire machine 40 to inject the air and stabilize the pressure of the Tire. Fig. 3 shows the operation procedures of the innovative expert sys- tem – ICRT while the varied situations of the glitches of the general family cars occurred. The detailed description of Hardware circuits and the operation procedures of the innovative expert system – ICRT about the pressure of the Tire not enough is described below. The Micro-chip processor 12 associated with the Electric Tire-pres- sure-meter 30 and the Efficiently pneumatic tire machine 40 will stabilize the pressure of the Tire using the Neural Network theory based Tire-Pressure-Stabilizer algorithm; the real pressure values of the Tire measured by the pressure sensor of the Electric Tire- pressure-meter 30 is transferred to the Control circuit 10, and the Control circuit 10 compares and calculates the deviation be- tween the measured values and the table of the preset standard values established inside the Micro-chip processor 12 and will make some appropriate operations to drive the Efficiently pneu- matic tire machine 40; after the comparing and calculating the deviation between the real measured values and the standard Tire eory based expert system – ICRT, Expert Systems with Applications S.-S. Lin et al. / Expert Systems with Applications xxx (2008) xxx–xxx 5 ARTICLE IN PRESS values (the preset values inside the Micro-chip processor 12), the Control circuit 10 will execute the Neural Network Theory based Tire-Pressure-Stabilizer algorithm to drive the Efficiently pneu- matic tire machine 40 to stabilize the Tire pressure. 3.2.2.4. LCD type Sparing Battery. In the LCD type Sparing Battery, the Micro-chip processor 12 is used and associated with the LCD displayer 60 to show the residual capacity of the Sparing Battery 50. It should be noticed that the Micro-chip processor 12 used in 50 is to detect the residual capacity of the Sparing Battery 50, it will send an Alarming signal if the residual capacity of the Sparing Battery 50 is leaking or is insufficient to start Cars. 3.2.2.5. Trouble blinking sign and Lighting device. The Trouble blink- ing sign and Lighting device 70 use the highly efficient LED device Fig. 4. Prototype of the proposed in Fig. 5. Real photograph of the ICRT Please cite this article in press as: Lin, S.-S. et al., A Neural Network th (2008), doi:10.1016/j.eswa.2008.06.098 associated with the Micro-chip processor 12 to provide the Light- ing service. It should be noticed that the Micro-chip processor 12 used in 70 is to control the working current of the LED device with- in some certain level. The Control circuit 10 provides the necessary facilitations for the Micro-chip processor 12, the Voltage-regulat- ing circuit 14, the Driving circuit of DC-Motor 16, the Time delay circuit 18, and the Circuit of the Trouble blinking sign 19 in the Neural Network Theory based the innovative expert system – ICRT. In the ICRT, the LCD displayer 60 is used to show the residual vol- umes of the Sparing Battery 50; the Control circuit 10 is used to monitor the residual power of the Sparing Battery 50 within some safety level. If the residual power of the Sparing Battery 50 is inef- ficient to start Cars, then the Control circuit 10 will sent a control signal to the LCD displayer 60 and the 60 will exhibit a Red light sign or output an Alarming signal. Moreover, the Control circuit novative expert system – ICRT. easily and rapidly lifts the car. eory based expert system – ICRT, Expert Systems with Applications 6 S.-S. Lin et al. / Expert Systems with Applications xxx (2008) xxx–xxx ARTICLE IN PRESS 10 uses the Time delay circuit 18 associated with the Circuit of the Flashing light of the glitch sign 19 to show the situation of problem on the Trouble blinking sign and Lighting device 70. The Trouble blinking sign and Lighting device 70 can also provide the necessary assistance for the drivers, such as the lighting and alarming signal during the car repair and waiting for rescue. In the ICRT, the Control circuit 10 connected with the Sparing Battery 50 by the Voltage- regulating circuit 14. And the Sparing Battery 50 is used to provide the power of the Control circuit 10. The drivers can also obtain the power from the Smoking socket device in the cars. In general, the power of the ICRT is provided by the Sparing Battery 50, but if the residual power of the Sparing Battery 50 is inefficient, the driv- ers can use the Smoke socket device of the car to execute the charging. Furthermore, there is one more important device (Wireless re- ceived/emitted device) is included in the Control circuit 10 of the ICRT, the Wireless received/emitted device is used to receive the control signal originated from the Emitter to make the remote con- trol of the ICRT. Moreover, there are some preset values of table programmed in the Micro-chip processor 12, such as the standard Fig. 6. Real photograph of the ICRT measures a Fig. 7. Prototype of the sm Please cite this article in press as: Lin, S.-S. et al., A Neural Network th (2008), doi:10.1016/j.eswa.2008.06.098 values of the Tire pressure Ps for the different types of the Tire and some parameters C, e used in the Neural Network theory based Tire-Pressure-Stabilizer algorithm and some minimum capacity of power requirement should be stored in the Sparing Battery 50 to start cars for the different type of cars; it depends on the differ- ent type of cars, we can make some appropriate settings and mod- ifications for the preset values in the Micro-chip processor 12. 4. Test results We have made numerous real tests of the proposed Neural Net- work Theory based innovative expert system - ICRT in many differ- ent types of Cars. Due to the page limitation of this paper, we don’t have enough space to describe everything about the ICRT, we just present some typical test results described in the following: The testing Car is: Toyota Camry, Body Weight about 1600 kg, Years 2006. Toyota is very famous cars Generation Company all over the world and the Camry cars produced by Toyota are very popular cars type in the market in the recently years. First of all, we set up the innovative expert system – ICRT in the Camry car. nd stabilizes the Tire-pressure efficiently. all size Battery Box. eory based expert system – ICRT, Expert Systems with Applications Fig. 8. Prototype of the Trouble blinking sign and Lighting device. S.-S. Lin et al. / Expert Systems with Applications xxx (2008) xxx–xxx 7 ARTICLE IN PRESS Fig. 4 is the Prototype of the proposed ICRT (we also label the name of every tools on ICRT); from Fig. 4, we can see the size of ICRT is quite small and the volume (L, W, H) is only (60, 22, 10) cm. There- fore, it is quite easy to be carried and installed in every cars not limited for the type of Camry. Subsequently, we show the real test results to demonstrate the attractive functions of the ICRT: (1) Portable Electric-jack: Fig. 5 shows that the ICRT can easily and rapidly lift the car in the Left Back Hand Side since we just use the ‘‘white color controller panel” in ICRT to operate the Porta- ble Electric-jack lifting the car up and down, therefore, it is pretty convenient to operate the device comparing to the traditional ‘‘Hand Jack”, especially for female drivers; we also can see that the Tire is stationary away from the ground about 5 cm (however, we can lift the car more higher if it is necessary), hence, we can make some necessary facilitations to maintain the car, such as changing Tire. (2) Electric Tire-pressure-meter, and (3) Efficiently pneumatic tire machine: Fig. 6 shows that the proposed ICRT uses the Neural Net- work Theory based Tire-Pressure-Stabilizer algorithm to stabilize the Tire rapidly and we also can see the measured Tire-pressure value is 27 pounds shown on the LCD (since the standard Tire value is set to be 27 pounds programmed in the micro-chip processor). (4) LCD type Sparing Battery: Fig. 7 is the Prototype of the small size Battery Box, since we use the corresponding technology of mi- cro-chip and we can see the size of the Sparing Battery is quite small comparing to the traditional Power Bank. (5) Trouble blinking sign and Lighting device: Fig. 8 is the Proto- type of the Trouble blinking sign and Lighting device. From the real Hardware implementation of the innovative ex- pert system – ICRT, we see the size of the device is quite small and the functions are compact and can be easily carried in Cars. Furthermore, the proposed ICRT is easily to be operated. 5. Conclusions In this paper, we focus on the frequent glitches of the general family cars. Based on the Neural Network theory and combining with the micro-chip processor technology, we create an innovative expert system – ‘‘Integrated Cars Repairing Tools”. The proposed ICRT owns the following attractive functions: (1) Portable Electric jack, (2) Electric Tire-pressure-meter, (3) Efficiently pneumatic tire machine, (4) LCD type Sparing Battery with the capability to show the residual capacity of Power, (5) Trouble blinking sign and Lighting device. Furthermore, the size of the i is quite small and can be easily carried in cars. We have installed the proposed innovative expert system – ICRT in many different types of Cars and made numerous Please cite this article in press as: Lin, S.-S. et al., A Neural Network th (2008), doi:10.1016/j.eswa.2008.06.098 simulations, the test results show that the ICRT is efficient for rejecting the frequent glitches of cars. Acknowledgement This research work was supported by National Science Council in Taiwan under Grant # NSC 93-2662-E-129-004-CC3. References Azad, N. L., Khajepour, A., & McPhee, J. (2005). Analysis of jackknifing in articulated steer vehicles. In Proceedings of the IEEE vehicle power and propulsion (VPP) conference (pp. 86–90). Chicago, USA. Carstens, J. R. (1993). Electrical sensors and transducers. Englewood Cliffs, NJ: Prentice Hall. Chapman, S. J. (2005). Electric machinery fundamentals (4th ed.). New York: McGraw-Hill. Chen, T.-L., & Lin, J.-W. (1997). The battery charging circuit. Journal of the Power Electrics, 3, 60–66. Comer, D. J., & Comer, D. T. (2003). Advanced electronic circuit design. John Wiley & Sons Inc. Dai-Dai Technical Company. Home service station AC air compressor. <http:// www.tata.com.tw/>. Grossmann, R. (1999). Quartz crystals as remote sensors for tire pressure. In Proceedings of the 16th IEEE instrumentation and measurement technology conference (vol. 3, pp.1745–1749). Venice, Italy. Harter, J. H. (2000). Electromechanics: Principles, concepts and devices (2nd ed.). Englewood Cliffs, NJ: Prentice Hall. Kornev, A. F., Kuchma, I. G., Pokrovski, V. P., & Soms, L. N. (2005). Half-tone images in laser image projection using saturated laser Fourier-amplifiers. In Proceedings of the CAOL second international conference on advanced optoelectronics and lasers (vol. 1, pp. 202–205). Chernivtsi, Ukraine. Kowalewski, M. (2004). Monitoring and managing tire pressure. IEEE Potentials, 23(3), 8–10. Laker, K. R., & Sansen, W. M. C. (1994). Design of analog integrated circuits and systems. New York: McGraw-Hill. Li, L., Wang, F. -Y., Qunzhi, Z., & Shan, G. (2003). Automatic tire pressure fault monitor using wavelet-based probability density estimation. In Proceedings of the IEEE intelligent vehicles symposium (pp. 80–84). Columbus, USA. Li, J.-S., & Liu, Y.-H. (2001). The development of the battery charging device for the motorcycle. Journal of the Power Electrics, 65, 44–57. Lie, S.-K., Yu, L.-B., et al. (2000). The development of the photo-element used in lighting service of the car. Journal of the Technology Development, 28(6), 442–448. Liu, Y.-H. (2001). The development of the efficiency battery charging device. Journal of the Power Electrics, 65, 58–73. Liu, J. -Y. (2005). The research of the lighting device of the technical company in Taiwan. The IEK-IT IS Project, Taiwan. MacKenzie, I. S. (1999). The 8051 microcontroller (3rd ed.). Upper Saddler River, NJ: Prentice-Hall. Mihali, R., & Sobh, T. (1999). The formula one tire changing robot (F1-TCR). In Proceedings of the IEEE international conference on robotics and automation (vol. 1, pp. 323–328). Detroit, USA. Nabipoor, M., & Majlis, B. Y. (2004). A passive telemetry LC pressure sensor optimized for TPMS. In Proceedings of IEEE international conference on semiconductor electronics, Kuala Lumpur, Malaysia. eory based expert system – ICRT, Expert Systems with Applications http://www.tata.com.tw/ http://www.tata.com.tw/ 8 S.-S. Lin et al. / Expert Systems with Applications xxx (2008) xxx–xxx ARTICLE IN PRESS Pohl, A., Ostermayer, G., Reindl, L., & Seifert, F. (1997). Monitoring the tire pressure at cars using passive SAW sensors. In Proceedings of the IEEE ultrasonics symposium (vol. 1, pp. 471–474), Ontario, Canada. The An-Hun Technical Company. The potable pneumatic tire machine-AH-C010. <http://www.szanhang.com.cn/>. The Audi-Tai Car Technical Company. The pneumatic tire machine with the digital player. <http://www.auto-tech.com.cn/cpjs/cp5/index.htm>. Wallis, P., Ronnquist, R., Jarvis, D., & Lucas, A. (2002). The automated wingman – Using JACK intelligent agents for unmanned autonomous vehicles. In Please cite this article in press as: Lin, S.-S. et al., A Neural Network th (2008), doi:10.1016/j.eswa.2008.06.098 Proceedings of the IEEE aerospace conference (vol. 5, pp. 2615–2622). Big Sky, Montana. Yamada, T., & Suzuki, K. (1990). Fuzzy control of plural hydraulic jacks in an elastic– plastic structural test. In Proceedings of the first international symposium on uncertainty modeling and analysis (pp. 546–551). Maryland, USA. Yeh, J.-L. (1997). The power bank and battery charging device. Journal of the Taiwan Power Company, 590, 44–57. eory based expert system – ICRT, Expert Systems with Applications http://www.szanhang.com.cn/ http://www.auto-tech.com.cn/cpjs/cp5/index.htm A neural network Neural Network theory based expert system - - ICRT Introduction Neural Network theory based innovative expert system - ICRT Neural Network Theory theory based Tire-Pressure-Stabilizer Neural Network Theory based Tire-Pressure-Stabilizer algorithm Hardware Implementation implementation of the Innovative Expert System - innovative expert system - ICRT Diagram description Hardware Implementation implementation of the Innovative Expert System - innovative expert system - ICRT The detailed description the combinations and interconnections of the block element symbols The Operation Sequence Control operation sequence control of the Hardware Implementation hardware implementation of the Innovative Expert System - innovative expert system - ICRT Portable Electric-jack Electric Tire-pressure-meter Efficiently pneumatic tire machine LCD type Sparing Battery Trouble blinking sign and Lighting device Test Resultsresults Conclusions Acknowledgement References 
work_z6y3bu72m5fildugbuffwt7kea	----	Research Article Expert System for Competences Evaluation 360∘ Feedback Using Fuzzy Logic Alberto Alfonso Aguilar Lasserre,1 Marina Violeta Lafarja Solabac,1 Roberto Hernandez-Torres,1 Rubén Posada-Gomez,1 Ulises Juárez-Martínez,1 and Gregorio Fernández Lambert2 1 Division of Research and Postgraduate Studies, Instituto Tecnológico de Orizaba, Avenida Instituto Tecnológico 852, Colonia Emiliano Zapata, 94300 Orizaba,VER, Mexico 2 Division of Research and Postgraduate Studies, Instituto Tecnológico Superior de Misantla, Km. 1.8 Carretera a Loma del Cojolite, 93821 Misantla, VER, Mexico Correspondence should be addressed to Alberto Alfonso Aguilar Lasserre; aaguilar@itorizaba.edu.mx Received 11 April 2014; Revised 24 June 2014; Accepted 25 June 2014; Published 24 July 2014 Academic Editor: Jer-Guang Hsieh Copyright © 2014 Alberto Alfonso Aguilar Lasserre et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Performance evaluation (PE) is a process that estimates the employee overall performance during a given period, and it is a common function carried out inside modern companies. PE is important because it is an instrument that encourages employees, organizational areas, and the whole company to have an appropriate behavior and continuous improvement. In addition, PE is useful in decision making about personnel allocation, productivity bonuses, incentives, promotions, disciplinary measures, and dismissals. There are many performance evaluation methods; however, none is universal and common to all companies. This paper proposes an expert performance evaluation system based on a fuzzy logic model, with competences 360∘ feedback oriented to human behavior. This model uses linguistic labels and adjustable numerical values to represent ambiguous concepts, such as imprecision and subjectivity. The model was validated in the administrative department of a real Mexican manufacturing company, where final results and conclusions show the fuzzy logic method advantages in comparison with traditional 360∘ performance evaluation methodologies. 1. Introduction Nowadays, labor competences and competence evaluation represent a real challenge for organizations, which emerged in order to assign the right man to the right job. This evaluation method is based on questionnaires that involve fixed scales with specific values, such as 100%, 75%, 50%, 25%, and 0%. This kind of evaluation reduces the evaluator opportunity to express points of view and causes a rigid evaluation. Fuzzy set theory appears as an important tool to include inaccurate judgments inherent in personnel evaluation process. Accord- ing to Butkiewicz [1] Fuzzy Logic is a very good tool for deci- sion problems, especially when nonprecise or partially precise description is available. Although it can be applied with suc- cess in management problems, fuzzy logic is not common in this area. Fuzzy Logic is an artificial intelligence (AI) technique [2]. AI comes with the purpose of developing models and programs of the intelligent behavior. One of the approaches of the AI is Logic, with the main objective of formalization of natural reasoning. Fuzzy logic has two main components: membership functions and fuzzy rules. Using them it is possible to move a qualitative to a quantitative description, for example, to rep- resent linguistic expressions as mathematic expressions. This is very useful when it is necessary to model the expertise of a human expert. Fuzzy membership functions express the certainty than an element of the universe belongs to a fuzzy set. It represents the degree of truth as an extension of the valuation. Degrees of truth are very often confused with probabilities but they are conceptually different because fuzzy truth represents Hindawi Publishing Corporation Mathematical Problems in Engineering Volume 2014, Article ID 789234, 18 pages http://dx.doi.org/10.1155/2014/789234 2 Mathematical Problems in Engineering membership in vaguely defined sets, and not likelihood of an event. These membership functions can take different shapes according to expertise and preferences of the designer. In these membership functions the 𝑥-axis represents the universe of discourse, and the 𝑦-axis represents the degrees of membership in the [0, 1] interval. Most com- monly used functions are triangular, trapezoidal, Gaussian, singleton, Gamma, and so forth. Membership functions can be expressed as a discrete or continuous function. In other words, 𝜇 𝐴 (𝑋) is a membership function of a set 𝐴, according to the elements of the universe. Fuzzy sets are classes of objects with grades of member- ship. Each set is characterized by a membership function, which assigns a grade of membership to each object based on a characteristic. When the universe of discourse is continuous and finite, commonly used notation to represent set 𝐴 is 𝐴 = ∫ 𝑥 𝜇 𝐴(𝑥) 𝑥 , (1) where 0 ≤ 𝜇 𝐴 (𝑥) ≤ 1. (2) When the universe of discourse is discrete and finite, fuzzy set 𝐴 is commonly represented as 𝐴 = 𝑛 ∑ 𝑖=1 𝜇 𝐴 (𝑥 𝑖 ) 𝑥 𝑖 = 𝜇 𝐴 (𝑥 1 ) 𝑥 1 + 𝜇 𝐴 (𝑥 2 ) 𝑥 2 + ⋅ ⋅ ⋅ + 𝜇 𝐴 (𝑥 𝑛 ) 𝑥 𝑛 . (3) For instance, if fuzzy set 𝐴 contains the elements 𝑥 1 , 𝑥 2 , 𝑥 3 , 𝑥 4 , and 𝑥 5 with membership degrees of 0, 0.5, 1, 0.5, and 0, respectively, the fuzzy set is expressed as 𝐴 = 0 𝑥 1 + 0.5 𝑥 2 + 1 𝑥 3 + 0.5 𝑥 4 + 0 𝑥 5 . (4) A fuzzy set in a discrete and finite universe of discourse can be represented too as a set of ordered pairs of 𝑥 and its membership degree in 𝐴 as 𝐴 = {(𝑥, 𝜇 𝐴 (𝑥)) | 𝑥 ∈ 𝑋} , (5) which results in 𝐴 = {(𝑥 1 , 0) , (𝑥 2 , 0.5) , (𝑥 3 , 1) , (𝑥 4 , 0.5) , (𝑥 5 , 0)} . (6) Geometry of fuzzy sets involves three elements: domain, range, and mapping. In this example, geometry of the fuzzy set 𝐴 is domain: {𝑥 1 , 𝑥 2 , 𝑥 3 , 𝑥 4 , 𝑥 5 }, range: [0, 1], mapping: 𝜇𝐴(𝑥) → [0, 1]; {0, 0.5, 1, 0.5, 0}. Graphically, fuzzy set 𝐴 can be expressed as shown in Figure 1. As in classic logic, fuzzy logic uses three basic operations in fuzzy sets: union, intersection, and complement. However, 1 0.5 0 0 1 2 3 4 5 Figure 1: Graphical representation of a fuzzy set. fuzzy sets have certain characteristics that make them differ- ent from classic sets. Fuzzy sets have elements with variable membership degrees, which means that an element of the universe of discourse can belong to one or more fuzzy sets, with different membership degrees. The first operation on fuzzy sets is intersection. It is the degree of membership that two fuzzy sets share, that is, is the smallest degree of membership of each element in the fuzzy sets. Intersection of two fuzzy sets 𝐴 and 𝐵 is a fuzzy set 𝐴∩𝐵 in the universe of discourse 𝑋, whose function is given by 𝜇 𝐴∩𝐵 (𝑥) = min [𝜇 𝐴 (𝑥) , 𝜇 𝐵 (𝑥)] = ∧ 𝑥 [𝜇 𝐴 (𝑥) , 𝜇 𝐵 (𝑥)], (7) where 𝐴 ∩ 𝐵 represents the intersection of the fuzzy sets 𝐴 and 𝐵; ∧ represents the minimum operator. The operation union results as the biggest degree of mem- bership of each element in the fuzzy sets, that is, the highest value of the fuzzy values. Union of two fuzzy sets 𝐴 and 𝐵 is a fuzzy set 𝐴∪𝐵 in the universe of discourse 𝑋, whose function is given by 𝜇 𝐴∪𝐵 (𝑥) = max [𝜇 𝐴 (𝑥) , 𝜇 𝐵 (𝑥)] = ∨ 𝑥 [𝜇 𝐴 (𝑥) , 𝜇 𝐵 (𝑥)], (8) where 𝐴 ∪ 𝐵 represents the intersection of the fuzzy sets 𝐴 and 𝐵; ∨ represents the maximum operator. The logic operation complement results as the degree of membership that the fuzzy set needs to reach the unit. The complementary set 𝐴 of a fuzzy set 𝐴 is that whose function is given by 𝜇 − 𝐴 (𝑥) = 1 − 𝜇 𝐴 (𝑥) . (9) Functions that define operations of intersection and union can be generalized using the triangular norm (called 𝑇-norm) and the triangular conorm (called 𝑇-conorm or 𝑆-norm), respectively. Mathematical Problems in Engineering 3 A t-norm operator is a function of two elements 𝑇(⋅, ⋅) that satisfies the following. Boundary Conditions. This condition implies the generaliza- tion of the classic sets: 𝑇 (𝑎, 0) = 0, 𝑇 (𝑎, 1) = 𝑎. (10) Monotonicity. This condition implies that a decrease in the degree of membership for the set 𝐴 or 𝐵 will not produce an increase in the degree of membership of the intersection of the sets 𝐴 and B: 𝑇 (𝑎, 𝑏) ≤ 𝑇 (𝑐, 𝑑) if 𝑎 ≤ 𝑐, 𝑏 ≤ 𝑑. (11) Commutative Property. This property indicates that the oper- ator is indifferent to the order of the fuzzy sets that are com- bined: 𝑇 (𝑎, 𝑏) = 𝑇 (𝑏, 𝑎) . (12) Associative Property. This property allows calculating the intersection of any number of fuzzy sets, grouped in pairs, regardless of the order of couples: 𝑇 (𝑎, 𝑇 (𝑏, 𝑐)) = 𝑇 (𝑇 (𝑎, 𝑏) , 𝑐) . (13) In the same way, operator 𝑇-conorm (𝑆-norm) is a func- tion of two elements 𝑆(⋅, ⋅) that satisfies the following. Boundary Conditions 𝑆 (𝑎, 1) = 1, 𝑆 (𝑎, 0) = 𝑎. (14) Monotonicity 𝑆 (𝑎, 𝑏) ≤ 𝑆 (𝑐, 𝑑) if 𝑎 ≤ 𝑐, 𝑏 ≤ 𝑑. (15) Commutative 𝑆 (𝑎, 𝑏) = 𝑆 (𝑏, 𝑎) . (16) Associative 𝑆 (𝑎, 𝑆 (𝑏, 𝑐)) = 𝑆 (𝑆 (𝑎, 𝑏) , 𝑐) . (17) Some interesting contributions of fuzzy logic as a tech- nique to model subjective viewpoints are found in [3], where the authors firstly considered decision making problems using fuzzy logic. Probably, the first attempt to apply fuzzy logic to personnel evaluation was proposed in [4, 5]. Another approach can be found in [6]. Cannavacciuolo et al. [4] presented the application of fuzzy set theory to a personnel evaluation procedure. Effectiveness of fuzzy concepts and methods depends on the approach used for the analysis of organizational issues. Fuzzy set theory allows them to model the weak signals existing in evaluation processes and high- lights part of the tacit knowledge involved in individual judgments. Usually, researchers, consultants, and managers use a rather qualitative approach to organizational problems. However the natural language is the preferred instrument to describe the organizational conditions because the shades of meaning and the ambiguity of verbal statements allow the company actors to manage diverging opinions, tensions, and conflicts. These approaches are detailed in [7–11]. On the other hand, the logical-mathematical models tend to represent a world of certainty and coherence where doubts, contradictions, divergences, polysemy, conflicts, and ambigu- ities are usually typified, dissolved, degraded, and linearized. Within this same conceptual framework, mathematicians, computer scientists, A.I. researchers, and engineers, in search of formal coherence, quantifiable variables, and efficient algorithms, usually tend to use fuzzy set theory without considering complexity and ambiguity in organizational sit- uations, for example, [3, 12–18]. There seems to be a growing trend towards the use of systematic procedures in personnel selection. For instance, Karsak [19] introduced a method that integrates decision makers linguistic assessments about subjective factors such as excellence in oral communication skills, personality, lead- ership, and quantitative factors such as aptitude test score within multiple objective programming frameworks. The importance level of each goal is considered by applying the composition operator to the membership function of the goal and the membership function corresponding to its fuzzy priority defined by linguistic variables. Kolarik et al. [20] present an online approach to moni- toring human performance in terms of conditional reliability when a task is performed. Unlike traditional human reliability analysis, this approach develops a dynamic model that can cope with constantly changing conditions that affect operator performance. A fuzzy knowledge-based assessment approach is developed in order to deal with uncertainty and subjectivity associated with human performance assessment. Podofellini et al. [21] assess the influence of the failure of the operators to perform one task on the failure probabilities of subsequent tasks with an approach called technique for human error rate prediction (THERP) and a fuzzy expert system (FES). Other works include the use of fuzzy logic to evolve an optimal and accurate judgment according to the human thinking model and also to mitigate the commonly occurred biases in human recruitment and selection procedures, as seen in [22]. Garćıa et al. [23] propose, through the use of tools based on fuzzy logic, the evaluation of the impact of training in companies, by applying the reasoning characteristic of fuzzy logic, with the aim of complementing and extending the classical logic. Tosti and Addison [24] refer that a poorly designed 360∘ feedback system can do more harm than good. The use of commercial 360∘ software is not always an option due to spe- cific requirements of each organization. In this way, it may be that people who design 360∘ programs are well versed in assessment and measurement technology and woefully 4 Mathematical Problems in Engineering lacking in their understanding of feedback technology. Relia- bility of the 360 degree feedback is supported by the number and hierarchy of raters, as referred to in [25], assuming that personal qualities are developmental goals. Therefore, soft- ware has been designed specifically for this study evaluation. This paper is organized as follows. Section 1 presents an introduction to the study. Section 2 shows some fundamental concepts about competences, performance evaluation 360∘ feedback, and competence evaluation methodology; fuzzy logic basis and a proposed fuzzy logic model are shown too. After that, the application of both systems (traditional 360∘ feedback system and fuzzy logic 360∘ feedback expert system) at the administrative area of a real manufacturing company in the state of Veracruz, Mexico, is shown. Section 3 shows clear and concise results while discussion about the significance of the results is developed. Finally, Section 4 describes the main conclusions of the study. 2. Expert System for Competences Evaluation 360∘ Feedback Using Fuzzy Logic 2.1. Competences. Personnel appraisal is considered as per- formance evaluation, and it is based on formal evalua- tion programs with reasonable information amount about employees and their job performance. The literature describes several evaluation methods, each with its own advantages and drawbacks, and there is no ideal or universal method for all people, positions, organizations, and situations. The choice will depend on many other aspects such as (i) position, (ii) characteristics to be measured, (iii) organizational culture, (iv) objectives, achieved or to be achieved, (v) circumstantial elements. Performance evaluation methods are classified according to the feature they measure characteristics, behaviors, or out- comes, as referred to in [1]. Behavior methods enable the eval- uator to identify how far the employee performance is away from a specific scale. These methods describe what actions should be exhibited during the position performance. It is mainly used to provide development-oriented feedback. According to Gomez-Mejia et al. [26], the main advantage in the performance measure adopting a behavior-based approach is that criteria or performance standards are con- crete. Behavior scales give employees specific behavior exam- ples that can make them successful (or avoid their success) in their work. If an employee knew the required skills for the position and the corresponding aperture in degrees, it could verify, analyze, and control its own behavior according to the requirements. A competence is an underlying characteristic in the employee related to an effectiveness standard and superior performance in a job or situation, as discussed in [27]. Head evaluation Suppliers and customers evaluation Peer evaluation Subordinate evaluation Self-evaluation Figure 2: 360∘ evaluation scheme. According to Levy-Leboyer [28], individual skills and company competences are closely related. Company compe- tences are constituted by the integration and coordination‘of individual skills; however, these competences require an inte- gration and coordination of knowledge and personal qual- ities. Individual competences are an individual property. Company competences are developed by individuals, but they belong to the company. 2.2. 360∘ Performance Evaluation. 360∘ performance evalua- tion is a sophisticated scheme that allows the employee to be evaluated by its surrounding bosses, peers, and subordinates (see Figure 2). A scheme may include suppliers or customers. 360∘performance evaluation can potentially bring a glob- alized diagnosis about the employee performance, allowing the evaluator to compare different opinions about the level of competence expected in the evaluated person and then to take decisions about how to increase the level of compliance of this competences. Alles [29] proposes 7 points for 360∘ evaluation. (1) Identify cardinal and individual competences. If the company has implemented a performance evaluation system, the competences will be the same. Eventually, it is possible to use a reduced number of competences when using the 360∘ evaluation system. (2) Design the tool. Questionnaires typically constitute the process support (see Figure 3). (3) Select evaluators: superiors, partners, internal cus- tomers in other areas, customers, and external suppli- ers. Customers can be included or not. It is important to emphasize the fact that assessments are anonymous and that evaluators are chosen by the evaluating person. (4) Launch the evaluation process with stakeholders and evaluators. Mathematical Problems in Engineering 5 Figure 3: Questionnaire example. (5) Data processing: most of the time data are processed by external consultants to preserve information con- fidentiality. (6) Communicate 360∘ evaluation results to concerned people. (7) The company will receive a consolidated report. This report will be received only by the employee. In this work, 360 degree feedback methodology proposed by Alles [29] is applied under two different approaches: tra- ditional 360∘ feedback and fuzzy logic 360∘ feedback. Like- wise, there are some other substantial differences. (1) Data processing is performed by two software applications, each designed for its specific process. Both applications are able to select evaluators randomly and present questionnaires. From this point, there is another difference: (2) first application per- forms the evaluation in the traditional 360 degree feedback; the second application is an expert system that uses fuzzy logic into the questionnaires to perform evaluation. The third substantial difference, hence, is that (3) expert system does not require a human expert, except when they are designed. No external consultant is necessary anymore, since expert system achieves this objective too. 2.3. Fuzzy Logic Basis. Fuzzy logic is the mapping from an input measurement space to an output measurement space using linguistic variables. It gives the ability to model impre- cision by incorporating qualitative components into a quan- titative analysis. Fuzzy logic systems have a narrow relationship with fuzzy logic concepts such as fuzzy sets and linguistic variables. The most popular fuzzy logic systems are Mamdani and Takagi- Sugeno. Fuzzy rules base Fuzzifier Inference Defuzzifier Fuzzy input Fuzzy output Data output mechanism Figure 4: Mamdani general fuzzy logic system. Mamdani fuzzy systems use 4 components (see Figure 4). (i) Fuzzifier: Mamdani system inputs are typically numeric values, coming from some kind of sensor or being results of a process; to be able to operate this value, Mamdani systems translate this value into a special value that can be operated by the inference mechanisms. This translation is done by the fuzzifier, which converts numeric values into fuzzy values that represent the level of pertinence of the different variables of the system to the fuzzy sets. (ii) Fuzzy inference mechanism: once the fuzzifier has translated the fuzzy values, these have to be processed to generate a fuzzy output. Inference mechanism task is to take fuzzy values and generate a fuzzy output based on a fuzzy rules base. (iii) Fuzzy rules base is the way in which Mamdani fuzzy systems have to represent expertise and linguistic knowledge to solve the issue. It is a set of IF-THEN sentences, containing two parts each: antecedent and conclusion. In a Mamdani fuzzy system, antecedent and conclusion are given by linguistic expressions. (iv) Defuzzifier: inference system output is a fuzzy output, so it cannot be interpreted by an external element which only could operate numeric data. To make it possible to operate this data, output is translated to numeric format, and this task is done by the defuzzi- fier, using one of different procedures such as gravity center or averaged centers. Fuzzy Logic uses certain essential components to achieve its purpose. Imprecision. Often the same term is used to describe impreci- sion and uncertainty in only slightly related areas of measure- ment. Imprecision in measurement is associated with a lack of knowledge. Imprecision as a probability form is associated with uncertainty about the future event occurrence. Impreci- sion in description, the imprecision type addressed by fuzzy logic, is connected with intrinsic or built-in imprecision that belongs to the event itself. Fuzzy logic addresses the issues associated with an intrin- sic imprecision rather than those directly concerned with measuring devices failures in the measurements accuracy. 6 Mathematical Problems in Engineering Intrinsic imprecision is associated with a phenomenon prop- erties description and not with properties measurement using some external device. Ambiguity. There are close semantic relationships between the ambiguity idea and fuzziness; in fact, some fuzzy states can be highly ambiguous. Ambiguity connotes the property to have several but plausible and reasonable interpretations. These interpretations can have different belief states. Ambiguity in meaning is a common occurrence in natural languages. Likelihood and Ambiguity. Fundamentally, the basic confu- sion between fuzzy logic and probability arises from the idea that they measure the same kind of uncertainty. In strictly semantic, as well as mechanistic, the two forms of uncertainty are different. Propositions in probability address the likeli- hood of an outcome for some discrete event. The event out- come either happens or does not happen. Propositions in fuzzy logic concern the degree to which an event occurred. While a probability outcome happens unequivocally, a fuzzy event occurrence may involve some degree of ambiguity or uncertainty. Fuzzy Sets Components. Gregory [30] indicates that the fuzzy logic has two main components: membership functions and fuzzy rules. When using these components it is possible to move the experiences and human preferences from a qualitative description to a quantitative description. Membership fuzzy functions can take different figures and forms, according the designer experiences and pref- erences. Typical functions are triangular, trapezoidal, S, Gamma, Gaussian, and exponential. On the other hand, the fuzzy rules are written as IF-THEN couples and reported in tabular form. The four basic ways in which the fuzzy rules can be achieved are expert experiences and engineering knowledge, human behaviors, models based on a fuzzy system, and learn- ing processes. These methods are not necessarily mutually exclusive. Membership Functions. In classical set theory, something is completely included or not. This situation can be described by assigning a value of one to all the elements included in the set and the value of zero to the ones not included in it. The func- tion that assigns these values is called “membership function.” The fuzzy sets allow to describe the degree of membership of the object to the concept given by the labels, and allow to assign values between zero and one to the membership function (see Figure 5). Mathematical Features of Fuzzy Sets. Main characteristics of the fuzzy sets are height, support, cutoff level-𝛼, and nucleus. Height. It is the highest degree of membership of the elements of the set; that is, Height (𝐴) = max {ℎ | ℎ = 𝜇 𝐴 (𝑥) , 𝑥 ∈ 𝑋} . (18) When the height of a fuzzy set is equal to 1 it is said that it is a normalized fuzzy set. 1 0D eg re e of m em be rs hi p 𝜇 (x ) Connects Domain Rk with 𝜇(x)i R1 Rn Figure 5: General structures in a fuzzy set. a b c d 1.0 0 𝜇 u Figure 6: Trapezoidal type membership function. Support. It is the number of elements whose degree of membership is not zero; that is, Sup (𝐴) = {𝑥 | 𝜇 𝐴 (𝑥) > 𝑂, 𝑥 ∈ 𝑋} . (19) Cutoff Level-𝛼. It is the set of elements of 𝑋 with a minimum degree 𝛼; that is, 𝐴 𝛼 = {𝑥 | 𝜇 𝐴 (𝑥) ≥ 𝛼, 𝑥 ∈ 𝑋} . (20) Nucleus. It is the set of elements of 𝑋 with a degree of membership equal to 1; that is, Nucleus (𝐴) = {𝑥 | 𝜇 𝐴 (𝑥) = 1, 𝑥 ∈ 𝑋} . (21) Inclusion Functions in Fuzzy Sets. There are standard families for inclusion functions; the most frequent ones are trape- zoidal, singleton, triangular, 𝑆, exponential, and 𝜋 type. Trapezoidal functions are defined by four points: 𝑎, 𝑏, 𝑐, and 𝑑 (see Figure 6). This function is zero for values lower than “𝑎” and higher than “𝑑” and one between “𝑏” and “𝑐” and takes values in range [0, 1] between “𝑎” and “𝑏” and between Mathematical Problems in Engineering 7 “𝑐” and “𝑑.” It is used in simple fuzzy systems, since it allows defining a fuzzy set with little information and computing membership function values in a simple way. This function is common for microprocessor based sys- tems since it can be encoded in a similar format as𝑆 functions, 𝜋 functions, and triangular and singleton functions (e.g., if points 𝑏 and 𝑐 are combined the result is a triangular function). Trapezoidal function is defined as follows (see (22)): 𝑆 (𝑢; 𝑎, 𝑏, 𝑐, 𝑑) = { { { { { { { { { { { { { { { { { { { { { { { { { 0, 𝑢 < 𝑎, ( 𝑢 − 𝑎 𝑏 − 𝑎 ) , 𝑎 ≤ 𝑢 ≤ 𝑏, 1, 𝑏 ≤ 𝑢 ≤ 𝑐, ( 𝑑 − 𝑢 𝑑 − 𝑐 ) , 𝑐 ≤ 𝑢 ≤ 𝑑, 0, 𝑢 > 𝑑. (22) Trapezoidal functions are suitable to model properties in a range of values, stages, or levels (e.g., young, adult, elder, etc.). Modeling a triangular function can be done through the 𝑏 = 𝑐 simplification. For an 𝑆 function and singleton types (but not soft), 𝑐 = 𝑑 = max(𝑈) and 𝑎 = 𝑏 = 𝑐 = 𝑑 transformations can be applied, respectively. Triangular function (𝑇) can be defined as indicated in 𝑇 (𝑢; 𝑎, 𝑏, 𝑐) = { { { { { { { { { { { { { { { { { 0, 𝑢 < 𝑎, 𝑢 − 𝑎 𝑏 − 𝑎 , 𝑎 ≤ 𝑢 ≤ 𝑏, 𝑐 − 𝑎 𝑐 − 𝑏 , 𝑏 ≤ 𝑢 ≤ 𝑐, 0, 𝑢 > 𝑐. (23) 𝑇 functions (see Figure 7) are appropriate for modeling properties with an inclusion value different from zero and for a narrow range of values around point 𝑏. Linguistic Variables. Linguistic variables take values from natural language, for example, much, little, positive, negative, and so forth. These words are considered as labels within the fuzzy set theory. Even though linguistic variables aim to assign labels as variable values taken from natural language words, they will be able to assign numerical values too. Then, in the expression “temperature is cold,” the variable “temperature” must be seen as a linguistic variable, since the value cold is assigned as a fuzzy set. However, this variable can also take numerical values such as “temperature is 4∘C.” Fuzzy Rules. Fuzzy rules combine one or more input fuzzy sets, called premises, and associate them with an outcome fuzzy set, called consequence. The fuzzy set premise is asso- ciated using AND, OR, and so forth operators. Defuzzification Process. There are two common methods of defuzzification process: gravity center (centroid) and maxi- mum output. As is shown in Figure 8, both techniques pro- duce different results [31]. a b c 1.0 0 𝜇 u 0.5 Figure 7: Type T (triangular) function. Both techniques produce reasonable results when they are applied in specific fuzzy models. Gravity center is the most common method, because it combines evidence about rules and response fields are pondered by the total true degree. Gravity center is, essentially, the weighted average of the output membership function. Centroid Computation. Centroid technique finds the balance point solution in fuzzy zone using the weighted average in the fuzzy region. Arithmetically, the procedure is formulated by R = ∑ 𝑛 𝑖=0 𝑑 𝑖 ⋅ 𝜇 𝐴 (𝑑 𝑖 ) ∑ 𝑛 𝑖=0 𝜇 𝐴 (𝑑 𝑖 ) , (24) where 𝑑 is the 𝑖th domain value and 𝜇(𝑑) is the true membership value at this point. Centroid or defuzzification with moments finds a point that represents the fuzzy set gravity center. 2.4. Fuzzy Logic Personnel Evaluation Model. A competence performance evaluation involves subjective viewpoints and evaluator preferences are reflected at the evaluation moment. Very often, evaluators express their perceptions in natural language terms. Nevertheless, as mentioned above, evalu- ation questionnaires constitute the performance evaluation support, and they are based on punctual values that do not reflect or approximate real viewpoints. Therefore, it is desirable to have a flexible evaluation tool that facilitates imprecision, ambiguity, and subjectivity han- dling. In this purpose, fuzzy sets allow a suitable treatment. The fuzzy logic model proposed is constituted as follows. Linguistic Variables. Three linguistic variables have been defined, “Scale,” “Frequency,” and “Required Level.” “Scale” variable refers to the percentage (assigned by the evaluator) indicating how well the employee behavior matches the competence definition. “Frequency” refers to the percentage obtained when the evaluator answers a question and ”rethinks” his/her evalua- tion determining the number of times the behavior is mani- fested. 8 Mathematical Problems in Engineering Center of gravity 0 10987654321 Numbers 1 0 D eg re e of m em be rs hi p 𝜇 (x ) (a) Maximum of output 0 10987654321 Numbers 1 0 D eg re e of m em be rs hi p 𝜇 (x ) (b) Figure 8: Defuzzification common methods. Needs significant development Needs development Competent Highly competent Role model 1 0 0 25 50 75 100 A B C D E Figure 9: Fuzzy sets for linguistic variable “Scale.” “Required Level” refers the percentage expected by the organization; the individual must cover conduct according the competence definition. Output Variables. They refer to the fit conduct qualification corresponding to the competence definition given in percent- age. Under this consideration there are three cases, “needs to improve,” “satisfies,” and “exceeds.” Fuzzy Sets Representation. This model uses triangular inclu- sion functions to represent linguistic variables; the output variable was modeled with trapezoidal functions. According to the expert, five fuzzy sets define the variable “Scale” possible values (see Figure 9), which are (i) “needs significant development,” (ii) “needs development,” (iii) “competent,” (iv) “highly competent,” (v) “role model.” Likewise, there are four fuzzy sets to define the variable “Frequency” possible values (see Figure 10), which are (i) “occasionally,” (ii) “half time,” (iii) “frequent,” (iv) “always.” The “Required Level” variable was modeled through three possible fuzzy sets (see Figure 11), called (i) “low,” (ii) “average,” (iii) “high.” The output variable has three cases which allow defining fuzzy sets (see Figure 12); these cases are (i) “requires improvement,” (ii) “complies,” (iii) “exceeds.” Mathematical Problems in Engineering 9 Occasionally Halftime Frequent Always 25 50 75 100 A B C D 1 0 Figure 10: Fuzzy sets for linguistic variable “Frequency.” Low Average High 1 0 50 75 100 A B C Figure 11: Fuzzy sets for linguistic variable “Required Level.” Requires improvement Complies Exceeds 50 62.5 87.5 100 1 0 A B C D Figure 12: Fit qualification. Fuzzy Rules. Fuzzy rules must consider all the combinations among input variables sets. Each combination will be associ- ated with an output variable fuzzy set. Sixty fuzzy rules were created in this case, according to the opinion of an expert, who rigorously analyzed each set of inputs, determining the level of performance produced by each rule (see Table 1). Fuzzification. The fuzzification process includes membership functions calculation for input variables and, then, uses the minimum-maximum criterion for variables activation. Defuzzification. The defuzzification process will allow getting the final fit qualification; the proposed method is the gravity center (centroid) including the following four steps: (1) divide total area in partial areas, (2) calculate partial areas value, (3) calculate each partial area centroid, (4) calculate total centroid. 2.5. Computational Experiments. This section describes the use of the methodology using both traditional 360∘ feedback and 360∘ fuzzy logic feedback. 2.5.1. Application of the Expert System in a Real Company. The use of both methods was carried out in the administrative area of a Mexican manufacturing company, located in the state of Veracruz, Mexico, where administrative procedures require staff to be evaluated based on their performance annually. A branch of the organization chart, consisting of five different but related positions was selected to perform the 360∘ feedback evaluation process (see Figure 13). 2.5.2. Traditional 360∘ Methodology (1) Identify cardinal and individual competences. To evaluate a position we will consider five cardinal com- petences and eight specific competences (see Table 2). (2) Design the tool. This evaluation includes four ques- tionnaires and each questionnaire includes five ques- tions. (3) Select an evaluator. Evaluation includes self-evalua- tion, boss evaluation, and peers evaluation. To choose evaluators and launch the process we will design and use the software; in this sense, the selection procedure is random. (4) Execute the evaluation process with concerned people and evaluators. As an example, Table 3 shows self- evaluation values for cardinal competences. 10 Mathematical Problems in Engineering Ta bl e 1: Fu zz y ru le ss um m ar y. SC A LE N ee ds si gn ifi ca nt de ve lo pm en t N ee ds de ve lo pm en t C om pe te nt H ig hl y co m pe te nt R ol e m od el Fr eq ue nc y Fr eq ue nc y Fr eq ue nc y Fr eq ue nc y Fr eq ue nc y Occasional Halftime Frequent Always Occasional Halftime Frequent Always Occasional Halftime Frequent Always Occasional Halftime Frequent Always Occasional Halftime Frequent Always R eq ui re d le ve l Lo w C (1 ) C (4 ) I( 7) I( 10 ) C (1 3) C (1 6) I( 19 ) I( 22 ) C (2 5) C (2 8) E (3 1) E (3 4) C (3 7) C (4 0) E (4 3) E (4 6) E (4 9) C (5 2) E (5 5) E (5 8) A ve ra ge C (2 ) I( 5) I( 8) I( 11 ) C (1 4) I( 17 ) I( 20 ) I( 23 ) I( 26 ) C (2 9) C (3 2) C (3 5) C (3 8) C (4 1) C (4 4) E (4 7) C (5 0) C (5 3) C (5 6) E (5 9) H ig h I( 3) I( 6) I( 9) I( 12 ) I( 15 ) I( 18 ) I( 21 ) I( 24 ) I( 27 ) I( 30 ) C (3 3) C (3 6) I( 39 ) I( 42 ) C (4 5) C (4 8) I( 51 ) I( 54 ) C (5 7) C (6 0) I: re qu ir es im pr ov em en t. C :c om pl ie s. E: ex ce ed s. Mathematical Problems in Engineering 11 Operational human resources coordinator Superintendent of operational assessment Specialist of assessment and control Occupational health specialist Principal analyst of occupational health Superintendent of organization and integration Labor productivity specialist Labor productivity specialist Studies and integration specialist Superintendent of professional competitiveness Professional competitiveness specialist Assessment and recruitment specialist Superintendent of operation and services Operation and services specialist Operations and services analyst Superintendent of labor regulations Labor regulation analyst Selected positions Figure 13: Organizational chart. Table 2: Cardinal and specific competences. Cardinal competences Specific competences (1) Work quality (1) Planning and organization skills (2) Tolerance to work under pressure (2) Customer orientation (3) Communication skills (3) Productivity (4) Contact modalities (4) Technical credibility (5) Analytical skills (5) Innovation (6) Empowerment (7) Collaboration (8) Commitment level-personal discipline-productivity (5) Data processing: software specifically designed for data processing to obtain the competence level scale value and frequency value was used; then we will aver- age the results for each question (see Table 4). Finally, averages are compared with the Required Level. The same procedure will be used for specific competences treatment. (6) Reports will be delivered only to the evaluated person. Comments and graphs will be printed in two reports: the first one will include cardinal competences and will be given to the company, while the second report will include specific competences and will be given to the employee. (7) Communicate the 360∘ evaluation results to the con- cerned people. 2.5.3. Fuzzy Logic Model Implementation. (1) Choose an organ- izational department. (2) Choose a post in the selected department. (3) Choose a required competence in the post. (4) Select one question in the questionnaire that evaluates the competence. (5) Evaluator assigns values for input variables “Scale” and “Frequency,” for example, Scale = 62% Frequency = 84% The Required Level assigned by the company = 75%. (6) Fuzzification process: compute the variable member- ship values corresponding to the variable Scale: 𝜇COMPETENT (SCALE) = { { { { { { { { { { { { { { { { { { { 0; 𝑊 ≤ 𝐵 󸀠 0, 1 − 𝐶 󸀠 − 𝑊 𝐶 󸀠 − 𝐵 󸀠 ; 𝐵 󸀠 < 𝑊 ≤ 𝐶 󸀠 0, 1 − 𝑊 − 𝐶 󸀠 𝐷 󸀠 − 𝐶 󸀠 ; 𝐶 󸀠 < 𝑊 < 𝐷 󸀠 0.52, 0; 𝐷 󸀠 ≤ 𝑊 0, 𝜇HIGHLY⋅COMPETENT (SCALE) = { { { { { { { { { { { { { { { { { { { 0; 𝑊 ≤ 𝐶 󸀠 0, 1 − 𝐷 󸀠 − 𝑊 𝐷 󸀠 − 𝐶 󸀠 ; 𝐶 󸀠 < 𝑊 ≤ 𝐷 󸀠 0.48, 1 − 𝑊 − 𝐷 󸀠 𝐸 󸀠 − 𝐷 󸀠 ; 𝐷 󸀠 < 𝑊 < 𝐸 󸀠 0, 0; 𝐸 󸀠 ≤ 𝑊 0. (25) 12 Mathematical Problems in Engineering Ta bl e 3: Se lf- ev al ua tio n va lu es ,c ar di na lc om pe te nc es . (a ) 36 0∘ ev al ua tio n Se lf- ev al ua tio n C ar di na lc om pe te nc ie s W or k qu al ity R eq ui re d le ve l: 10 0 To le ra nc e to w or k un de rp re ss ur e R eq ui re d le ve l: 75 C om m un ic at io n R eq ui re d le ve l: 75 C on ta ct m od al iti es R eq ui re d le ve l: 10 0 A na ly tic sk ill s R eq ui re d le ve l: 75 Sc al e Fr eq ue nc y Sc al e Fr eq ue nc y Sc al e Fr eq ue nc y Sc al e Fr eq ue nc y Sc al e Fr eq ue nc y Q ue s. 1 75 10 0 10 0 10 0 10 0 25 75 10 0 75 25 Q ue s. 2 75 75 75 50 10 0 10 0 10 0 10 0 10 0 75 Q ue s. 3 10 0 50 10 0 10 0 75 10 0 10 0 75 10 0 75 Q ue s. 4 10 0 10 0 75 75 50 10 0 75 10 0 50 50 Q ue s. 5 75 10 0 50 25 25 75 50 25 10 0 10 0 (b ) W he re Sc al e Fr eq ue nc y 10 0% :A 10 0% :a lw ay s 75 % :B 75 % :f re qu en t 50 % :C 50 % :h al ft im e 25 % :D 25 % :o cc as io na l 0% :N O 0% :n ev er Mathematical Problems in Engineering 13 Ta bl e 4: Se lf- ev al ua tio n, ca rd in al co m pe te nc es tr ea tm en t. (a ) 36 0∘ ev al ua tio n Se lf- ev al ua tio n C ar di na lc om pe te nc ie s W or k qu al ity R eq ui re d le ve l: 10 0 Pr es su re to le ra nc e R eq ui re d le ve l: 75 C om m un ic at io n R eq ui re d le ve l: 75 C on ta ct m od al iti es R eq ui re d le ve l: 10 0 A na ly tic sk ill s R eq ui re d le ve l: 75 Sc al e Fr eq ue nc y Pr od uc t Sc al e Fr eq ue nc y Pr od uc t Sc al e Fr eq ue nc y Pr od uc t Sc al e Fr eq ue nc y Pr od uc t Sc al e Fr eq ue nc y Pr od uc t Q ue s. 1 75 10 0 75 10 0 10 0 10 0 10 0 25 25 75 10 0 75 75 25 18 .7 5 Q ue s. 2 75 75 56 .2 5 75 50 37 .5 10 0 10 0 10 0 10 0 10 0 10 0 10 0 75 75 Q ue s. 3 10 0 50 50 10 0 10 0 10 0 75 10 0 75 10 0 75 75 10 0 75 75 Q ue s. 4 10 0 10 0 10 0 75 75 56 .2 5 50 10 0 50 75 10 0 75 50 50 25 Q ue s. 5 75 10 0 75 50 25 12 .5 25 75 18 .7 5 50 25 12 .5 10 0 10 0 10 0 A ve ra ge 71 .2 5 61 .2 5 53 .7 5 67 .5 58 .7 5 (b ) W he re Sc al e Fr eq ue nc y 10 0% :A 10 0% :a lw ay s 75 % :B 75 % :f re qu en t 50 % :C 50 % :h al ft im e 25 % :D 25 % :o cc as io na l 0% :N O 0% :n ev er 14 Mathematical Problems in Engineering Compute the variable membership values corresponding to the variable Frequency: 𝜇FREQUENT (FREQUENCY) = { { { { { { { { { { { { { { { { { { { 0; 𝑊 ≤ 𝐵 0, 1 − 𝐶 − 𝑊 𝐶 − 𝐵 ; 𝐵 < 𝑊 ≤ 𝐶 0, 1 − 𝑊 − 𝐶 𝐷 − 𝐶 ; 𝐶 < 𝑊 < 𝐷 0.64, 0; 𝐷 ≤ 𝑊 0. 𝜇ALWAYS (FREQUENCY) = { { { { { { { { { 0; 𝑊 ≤ 𝐶 0, 1 − 𝐷 − 𝑊 𝐷 − 𝐶 ; 𝐶 < 𝑊 ≤ 𝐷 0.36, 1; 𝐷 ≤ 𝑊 0. (26) Compute the variable membership values corresponding to the variable Required Level: 𝜇AVERAGE (REQUIREDLEVEL) = { { { { { { { { { { { { { { { { { { { 0; 𝑊 ≤ 𝐴 󸀠󸀠 0, 1 − 𝐵 󸀠󸀠 − 𝑊 𝐵 󸀠󸀠 − 𝐴 󸀠󸀠 ; 𝐴 󸀠󸀠 < 𝑊 ≤ 𝐵 󸀠󸀠 1, 1 − 𝑊 − 𝐵 󸀠󸀠 𝐶 󸀠󸀠 − 𝐵 󸀠󸀠 ; 𝐵 󸀠󸀠 < 𝑊 < 𝐶 󸀠󸀠 0, 0; 𝐶 󸀠󸀠 ≤ 𝑊 0. (27) The fuzzification process continues with variable activa- tion, performed according to the min.-max. criterion (see (28), (29), (30), and (31)); this procedure allows identifying the membership values that will appear in the defuzzification process. Rule 32 is 𝜇 𝐶∩𝐹∩𝑀 (62, 84, 75) = min {𝜇 𝐶 (62) , 𝜇 𝐹 (84) , 𝜇 𝑀 (75)} = min {0.52, 0.64, 1.00} = 0.52. (28) Rule 35 is 𝜇 𝐶∩𝑆∩𝑀 (62, 84, 75) = min {𝜇 𝐶 (62) , 𝜇 𝑆 (84) , 𝜇 𝑀 (75)} = min {0.52, 0.36, 1.00} = 0.36. (29) Rule 44 is 𝜇 𝐴∩𝐹∩𝑀 (62, 84, 75) = min {𝜇 𝐴 (62) , 𝜇 𝐹 (84) , 𝜇 𝑀 (75)} = min {0.48, 0.64, 1.00} = 0.48. (30) Table 5: Membership degree for activated variables. Activated variable Set, output variable Membership degree 32 Requires improvement 0.52 35 Complies 0.36 44 Complies 0.48 47 Complies 0.36 Table 6: Membership degree for activated variables. Activated variable Fuzzy set Membership degree — Exceeds 0 35, 44, 47 Complies 0.48 32 Requires improvement 0.52 Rule 47 is 𝜇 𝐴∩𝑆∩𝑀 (62, 84, 75) = min {𝜇 𝐴 (62) , 𝜇 𝑆 (84) , 𝜇 𝑀 (75)} = min {0.48, 0.36, 1.00} = 0.36. (31) The rules 32, 35, 44, and 47 are the activated variables; they have correspondence with a set in the output variable (see Table 5). Three activated variables are in the “Obey” set; it is nec- essary to use the min.-max. criterion to select one value (see (32)): 𝜇 𝐶 = (62, 84, 75) = ∨ {∧ {𝜇 𝐶 (62) , 𝜇 𝑆 (84) , 𝜇 𝑀 (75)} , ∧ {𝜇 𝐴𝐶 (62) , 𝜇 𝐹 (84) , 𝜇 𝑀 (75)} , ∧ {𝜇 𝐴𝐶 (62) , 𝜇 𝑆 (84) , 𝜇 𝑀 (75)}} = 0.48. (32) The last membership degrees and activated variables are summarized as follows (see Table 6). (7) Defuzzification process: this procedure includes five steps. Divide the total area in partial areas (see Figure 14). Compute the partial area values (see (33)): (𝐴I) = [0.52] 2 (62.5 − 50) 2 = 1.69, (𝐴II) = [0.52] [1 − 0.52] (62.5 − 50) = 3.12, (𝐴III) = [0.48] (62.5 − 50) 2 = 3, (𝐴IV) = [0.48] (87.5 − 62.5) = 12, (𝐴V) = [0.48] [1 − 0.48] (100 − 87.5) = 3.12, (𝐴VI) = [0.48] 2 (100 − 87.5) 2 = 1.44. (33) Mathematical Problems in Engineering 15 Requires improvement Complies Exceeds 50 62.5 87.5 100 A B C D 1 0 0.52 0.48 I II III IV V VI Figure 14: Partial areas. Compute each partial area centroid (see (34)): Centroid (𝐴I) = 50 + (62.5 − 50) (1 − 0.52) 2 = 53, Centroid (𝐴II) = 62.5 − 2 (62.5 − 50) [0.52] 3 = 58.1666, Centroid (𝐴III) = 62.5 − (62.5 − 50) [0.48] 3 = 60.5, Centroid (𝐴IV) = 62.5 + 87.5 2 = 75, Centroid (𝐴V) = 87.5 + (100 − 87.5) (1 − 0.48) 2 = 90.75, Centroid (𝐴VI) = 100 − 2 (100 − 87.5) [0.48] 3 = 96. (34) Calculate the total centroid (see (35)): CeT = ((1.69) (53) + (3.12) (58.16) + (3) (60.5) + (12) (75) + (3.12) (90.75) + (1.44) (96)) × (1.69 + 3.12 + 3 + 12 + 3.12 + 1.44) −1 = 1773.93 24.37 = 72.79. (35) (8) Competences questionnaires have more than one question. In this sense, each questionnaire will have as many fuzzy treatments as questions. At the end of the process, an average qualification for each competence will be obtained. (9) Evaluation software and evaluation report. Intelli- gent evaluation systems have questionnaires that emulate human thought since it gives the option to answer questions using linguistic labels (see Figure 15). Once the evaluator has selected a label, it selects a numerical value. Final stage includes an evaluation report,which includes a summarized table with average qualifications for each competence. Figure 15: Evaluation questionnaire. Table 7: Cardinal competences and required level. Cardinal competences Required level Work quality 100% Pressure tolerance 75% Communication 75% Contact modalities 100% Analytic skill 75% 3. Results 3.1. 360∘ Methodology Application. A complete application includes four-position analysis (see Figure 13); five cardinal competences were developed using the job description (see Table 7). The first position has nine specific competences, while the second position has ten specific competences (see Table 8), the third position has eight competences, and finally the fourth position has ten competences. Required Levels were established with the company. Applying the complete traditional methodology 360∘ to the first position, we concluded that the company must improve “work quality” and “contact modalities,” since these competences obtained the major difference with the Required Level (see Table 9). The individual report includes specific competences sum- mary. 16 Mathematical Problems in Engineering Table 8: Specific competences and required level. Specific competences Position level I Required level Position level II Required level (1) Work team 100% (1) Planning and organization skills 100% (2) Negotiation 100% (2) Negotiation 50% (3) Personnel development 75% (3) Initiative autonomy 100% (4) Leadership 100% (4) Leadership 100% (5) Frankness-trustworthy-integrity 50% (5) Frankness-trustworthy-integrity 50% (6) Commitment 75% (6) Empowerment 100% (7) Collaboration 50% (7) Strategic thinking 100% (8) Coaching 75% (8) Commitment level-personal discipline-productivity 100% (9) Decision making 100% (9) Making decisions 100% (10) Human Resources strategic development 50% Table 9: Cardinal competences report. Cardinal competences Work quality Pressure tolerance Communication Contact modalities Analytic skill Self-evaluation 71.25 61.25 53.75 67.5 58.75 Peer 21.25 16.25 13.75 41.25 51.25 Head 25.00 43.75 17.50 15.00 16.25 360∘ weighted 39.17 40.42 28.33 41.25 42.08 Required level 100.00 75.00 75.00 100.00 75.00 Difference: required level, 360∘ weighted 60.83∗ 34.58 46.67 58.75∗ 32.92 ∗Competences with major difference. Table 10: Specific competences report. Specific competences Planning and organization ability Client orientation Productivity Credibility technique Innovation Empowerment Collaboration Commitment level-personal discipline- productivity Self-evaluation 63.75 43.75 60.00 26.25 72.50 25.00 37.50 22.50 Peer 48.75 61.25 56.25 48.75 41.25 37.50 63.75 62.50 Boss 12.50 18.75 17.50 13.75 6.25 5.00 15.00 13.75 360∘ weighted 41.67 41.25 44.58 29.58 40.00 22.50 38.75 32.92 Required level 75.00 50.00 100.00 75.00 100.00 50.00 100.00 100.00 Difference: required level, 360∘ weighted 33.33 8.75 55.42 45.42 60.00 27.50 61.25∗ 67.08∗ ∗Competences with major difference. Applying the complete traditional methodology 360∘ to the same post, we concluded that an employee must improve “collaboration” and “commitment level-personal discipline- productivity,” since these competences obtained the biggest difference with the Required Level (see Table 10). 3.2. Fuzzy Logic Model Application. Applying the full fuzzy logic model evaluation to the first position, we concluded that the company must improve “communication” and “contact modalities,” since these competences obtained the lower final qualification. Employee must focus on “credibility technique” and “commitment level-personal discipline-productivity” (see Figure 16). 4. Discussion Traditional methodology 360∘ indicates competences that the company and the employee must care. Fuzzy logic 360∘ methodology indicates the competences on which the com- pany and the employees must focus their efforts but in addi- tion gives a qualification. This qualification can be arranged and the lowest value will indicate which competences must be immediately attended. According to Table 11, the traditional 360∘ and fuzzy logic methodologies conclude that the company must attend the cardinal competence “contact modalities” and the employee must focus on the specific competence “commitment level- personal discipline-productivity.” However, fuzzy logic Mathematical Problems in Engineering 17 Table 11: Comparison results. Competences classification Traditional 360∘ Methodology Fuzzy Logic 360∘ Expert system Cardinal ∗Work quality ∗Contact modalities ∗Communication: 56.67 ∗Contact modalities: 57.33 Specific ∗Collaboration ∗Commitment level-personal discipline- productivity ∗Technical Credibility: 58.33 ∗Commitment level-personal discipline-productivity: 59.67 Figure 16: Qualifications summary. method indicates that “contact modalities” must be attended after “communication,” and “commitment level-personal dis- cipline-productivity” must be attended after “technical credibility.” The traditional 360∘ system involves only two factors, “Scale” and “Frequency,” and these are compared with the third factor “Required Level.” On the other hand, fuzzy logic model can involve three factors like input variables. Evaluation questionnaires of the traditional 360∘method- ology are filled with five possible percentages (100, 75, 50, 25, and 0). On the other hand, questionnaires in fuzzy logic expert system have adjustable values scales, allowing improved flexibility for the evaluator. Thus, the main advantage in fuzzy logic expert system is the human thinking simulation, assigning labels as a qualifi- cation, and allowing subjective and ambiguity treatment. 5. Conclusions Competences performance evaluation 360∘ is a complete sys- tem since it involves different viewpoints to appraise per- sonnel performance. It allows better interpretations, since the evaluation responsibility falls in different evaluators. This kind of evaluation facilitates the competence concept comprehension, and Required Levels are assimilated. Never- theless, it is a rigid system because the questionnaire filling procedure is strict, since these were designed to assign fixed values. Fuzzy logic competences evaluation expert system includes complex analysis, due to identification and modeling input and output variables. However, its main advantage is the ambiguity and subjectivity handling, since the evaluator can assign words to stand a qualification. This system is flexible because numerical adjustable values can be assigned to behaviors. Graphical interpretation helps to obtain suitable feedback. Even more, final processing reports can be obtained easier and faster. Therefore, it represents an excellent tool for competences monitoring, given the importance that the staff appraisal process has for human resource management, in the areas of recruitment and selection, job evaluation, identification of training needs, and so forth, and the value that results from having nonsubjective and bias-free assessments. The application of an expert system to performance eval- uation in a Mexican manufacturing company allows knowing its effectiveness against traditional techniques. This work brings innovative contributions to soft com- puting and human resources management solutions, finding new ways to apply artificial intelligence techniques by means of computer applications to processes that typically were performed by humans. Conflict of Interests The authors declare that there is no conflict of interests regarding the publication of this paper. References [1] B. S. Butkiewicz, “Selection of staff for enterprise using fuzzy logic,” in IEEE International Conference on Systems, Man and Cybernetics, vol. 4, Warsaw University of Technology, Warsaw, Poland, 2002. [2] E. Gómez, Modelo difuso de planeación agregada [Doctoral thesis], Centro de Ingenieŕıa y Desarrollo Industrial, Querétaro, México, 2009. [3] L. A. Zadeh, “Making computers think like people,” IEEE Spectrum, vol. 21, no. 8, pp. 26–32, 1984. [4] A. Cannavacciuolo, G. Capaldo, A. Ventre, and G. Zollo, “Link- ing the fuzzy set theory to organizational routines. A study in personnel evaluation in a large company,” in Proceedings of the 2nd IEEE International Conference on Fuzzy Systems, pp. 667– 672, San Francisco, Calif, USA, April 1993. [5] A. Cannavacciuolo, G. Capaldo, A. Ventre, and G. Zollo, “An approach to the evaluation of human resourses by using fuzzy set theory,” in Proceedings of the 3rd IEEE Conference on Fuzzy Systems, Orlando, Fla, USA, 1994. [6] J. Stahl, “A method for personnel selection in concurrent engi- neering using fuzzy sets,” in Fuzzy Sets in Engineering Design and Configuration, H.-J. Sebastian and E. K. Antonsson, Eds., Kluwer Academic Publishers, Boston, Mass, USA, 1996. [7] C. Argyris and D. A. Schon, Organizational Learning. A Theory of Action Perspective, Addison-Wesley, Reading, Mass, USA, 1978. 18 Mathematical Problems in Engineering [8] K. S. Cameron, “Effectiveness as paradox: consensus and con- flict in conceptions of organizational effectiveness,” Journal of Management Science, vol. 32, no. 5, pp. 539–553, 1986. [9] D. J. Orton and K. E. Weik, “Loosely coupled system: a recon- ceptualization,” Academy of Management Review, vol. 15, no. 2, pp. 203–223, 1990. [10] C. R. Schwenk, “Cognitive simplification processes in strategic decision-making,” Strategic Management Journal, vol. 5, pp. 111– 128, 1984. [11] K. E. Weick, “Educational organizations as loosely coupled sys- tems,” Administrative Science Quarterly, vol. 21, pp. 1–19, 1976. [12] C. Carisson, “On the relevance of fuzzy sets in management science methodology,” in Fuzzy Sets: Theory and Applications to Policy Analysis and information System, P. P. Wang and S. K. Chang, Eds., Plenum Press, New York, NY, USA, 1980. [13] B. R. Gaines, L. A. Zadeh, and H. Zimmermann, “Fuzzy sets and decision analysis—a perspective,” in Fuzzy sets and Decision Analysis, H. J. Zimmermann and etal, Eds., North-Holland, Amsterdam, Netherlands, 1984. [14] R. M. Tong and P. P. Bonissone, “Linguistic solutions to fuzzy decision problems,” in Studies in the Management Science, H. J. Zimmermann, Ed., pp. 323–324, 1984. [15] R. R. Yager, “Fuzzy thinking as quick and efficient,” Cybernetica, vol. 23, no. 4, pp. 265–298, 1980. [16] P. L. Yu, “Dissolution of fuzziness for better decision-perspec- tive and techniques,” in Fuzzy Sets and Decision Analysis, J. H. Zimmermann et al., Ed., North-Holland, Amsterdam, The Netherlands, 1984. [17] S. French, “Fuzzy decision analysis: some criticism,” in Fuzzy sets and Decision Analysis, H. J. Zimmermann, B. R. Gaines, and L. Zadeh, Eds., Amsterdam, The Netherlands, North Holland Publisher, 1984. [18] L. A. Zadeh, “Outline of new approach to the analysis of complex systems and decision processes,” IEEE Transactions on Systems, Man, and Cybernetics, vol. 3, pp. 28–44, 1973. [19] E. E. Karsak, “A fuzzy multiple objective programming approach for personnel selection,” in Proceedings of the IEEE International Conference on Systems, Man, and Cybernetics, vol. 3, pp. 2007–2012, Nashville, Tenn, USA, October 2000. [20] W. J. Kolarik, J. C. Woldstad, S. Lu, and H. Lu, “Human performance reliability: on-line assessment using fuzzy logic,” IIE Transactions, vol. 36, no. 5, pp. 457–467, 2004. [21] L. Podofellini, V. Dang, E. Zio et al., “Using expert models in human reliability analysis—a dependence assessment method based on fuzzy logic,” Risk Analysis, vol. 30, no. 8, pp. 1277–1297, 2010. [22] K. C. Mittal, A. K. Goel, and P. Mohindru, “Fuzzy multi-criteria decision making (MCDM) in human resource selection proce- dure—a case study of Indian IT industry,” BVIMR Management Edge, vol. 6, no. 1, pp. 89–97, 2013. [23] R. E. Garćıa, G. Felix Benjamin, and R. Bello Perez, “Impact assessment of training with fuzzy logic,” Ingeniare. Revista Chilena de Ingenieŕıa, vol. 22, no. 1, pp. 41–52, 2014. [24] D. T. Tosti and R. M. Addison, “360-degree feedback: going around in circles?” Performance Improvement, vol. 48, no. 3, pp. 36–39, 2009. [25] R. Hensel, F. Meijers, R. van der Leeden, and J. Kessels, “360 degree feedback: how many raters are needed for reliable ratings on the capacity to develop competences, with personal quali- ties as developmental goals?” International Journal of Human Resource Management, vol. 21, no. 15, pp. 2813–2830, 2010. [26] L. R. Gomez-Mejia, D. B. Balkin, and R. L. Cardy, Managing Human Resources, Prentice Hall, Saddle River, NJ, USA, 2nd edition, 1998. [27] M. Spencer Lyle and M. Spencer Signe, Competence at Work: Models for Superior Performance, 1993. [28] C. Levy-Leboyer, “Evaluación del personal: los métodos a ele- gir,” Ediciones Dı́az de Santos, 1992. [29] M. A. Alles, Gestión por Competencias El diccionario, Granica, Buenos Aires, Argentina, 2nd edition, 2005. [30] C. Gregory, Dynamic Economics: Optimization by the Lagrange Method, Oxford University Press, 1997. [31] E. Cox, The Fuzzy Systems Handbook: A Practitioner’s Guide to Building, Using, and Maintaining Fuzzy Systems, Ap. Profes- sional, 1994. Submit your manuscripts at http://www.hindawi.com Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Mathematics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Mathematical Problems in Engineering Hindawi Publishing Corporation http://www.hindawi.com Differential Equations International Journal of Volume 2014 Applied Mathematics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Probability and Statistics Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Mathematical Physics Advances in Complex Analysis Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Optimization Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Combinatorics Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Operations Research Advances in Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Function Spaces Abstract and Applied Analysis Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Mathematics and Mathematical Sciences Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Algebra Discrete Dynamics in Nature and Society Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Decision Sciences Advances in Discrete Mathematics Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Stochastic Analysis International Journal of 
work_z6yxjqvgprbfli5o5skvunx6wu	----	Programming Robosoccer agents by modeling human behavior Ricardo Aler3'1, José M. V a l l s a ' \ David Camachob'*, Alberto López a ' ] a Computer Science Department, Universidad Carlos III de Madrid, Ávenue Universidad, No, 30, 28911, Legones, Madrid, Spain b Computer Science Department, Universidad Autónoma de Madrid, CIFrancisco Tomás y Valiente, No. 11, 28049, Madrid, Spain Abstract The Robosoccer simulator is a challenging environment for artificial intelligence, where a human has to program a team of agents and introduce it into a soccer virtual environment. Most usually, Robosoccer agents are programmed by hand. In some cases, agents make use of Machine learning (ML) to adapt and predict the behavior of the opposite team, but the bulk of the agent has been preprogrammed. The main aím of this paper is to transform Robosoccer into an interactíve game and let a human control a Robosoccer agent. Then ML techniques can be used to model his/her behavior from training instances generated during the play. This model will be used later to control a Robosoccer agent, thus imitating the human behavior. We have focused our research on low-level behavior, like looking for the ball, conducting the ball towards the goal, or scoring in the presence of opponent players. Results have shown that indeed, Robosoccer agents can be controlled by programs that model human play. 1. Introduction The Robosoccer simulator is a challenging environ- ment for artificial intelligence, where a human has to pro- gram an agent and introduce it into a soccer virtual environment. Programming complex behaviors in soft- ware agents is usually a time-consuming and difficult task for human programmers. Machine learning (ML) is becoming a promising way of automatically endowing agents with complex skills. The main approach that has been followed so far is to let the agents learn the behav- iors by themselves, totally or partially. For instance, in the case of the Robosoccer simulator, Luke, Hohn, Far- ris, and Hendler (1997) uses genetic programming to evolve a complete team of agents whereas Stone and Veloso (1998) proposes an agent architecture where learn- ing can be used at many levéis, from acquiring skills to adapting to the opponent. But ML can be used in another interesting way. Human players can learn to play many games well and quickly, so it makes sense to attempt to imítate them and transfer their experience to computer agents. Learning by imitation and modeling humans is a field common to many research áreas like robotics (Kuniyoshi, Inaba, & Inoue, 1994), cog- nitive science (Brown & Burton, 1978), or user modeling (Webb, Pazzani, & Billsus, 2001). However, only recently imitation techniques have been applied to programming computer agents, specially in games. Sklar, Blair, Funes, and Pollack (1999, 2001) is the first reported work (to our knowledge) where data collected from players is used to train an agent that plays T R O N . More recent research has produced quite remarkable human-like behavior in the Quake game (Thurau, Bauckhage, & Sagerer, 2003, 2004a, 2004c). Keywords: Learning to play; Imitation; Human modeling; Behavioral cloning; Machine learning; Robosoccer 1 Cita bibliográfica Published in: Expert Systems with Applications, Marzo 2009, vol. 36, n. 2, p. 1850-1859 The work reported in this paper follows this line of research and attempts to imitate a human player with the purpose of creating a Robosoccer player that performs well in the field. The Robosoccer type of domains is interesting for modeling humans, as skills range from low-level reac- tive behavior to high level strategic actions and team play. Our approach works as follows. First, we created an inter- face to allow a person to play Robosoccer just like any other video-game. Then, many input/output pairs were generated and recorded after the human player played sev- eral matches. The input is what the person can see in the field and the output is the action the person performed under that situation. Then, a ML technique is used to learn the mapping from inputs to outputs. Finally, the ML model is used to control a computer agent. Imitation can be performed at different levéis: reactive, tactical, or strategic. Different levéis will raise diflerent issues. As this is the first attempt at using human experi- ence to control an agent in the Robosoccer domain, we have chosen to imitate the human player in low-level actions like running, turning, or kicking the ball. However, results show that learning when to perform such low-level actions allow to learn slightly more complex sequences of actions like conducting the ball to the goal, dribbling oppo- nents, or stealing the ball from opponents. Results have shown that indeed, agents can be pro- grammed by modeling the experience of humans playing Robosoccer and performing behaviors like looking for the ball, conducting the ball towards the goal, or scoring in the presence of opponent players. This paper has been divided into the following sections. First, Section 2 deals with work related to modeling other agents, human modeling, imitation, and modeling in games. Section 3 describes our modeling approach in the Robosoccer domain. Section 4 describes how the agent was trained from the instances generated by the human in different types of behavior and discusses the results. Finally, 5 draws the most important conclusions of this work and posits some future lines of research. 2. Related work Currently, there is a lot of interest in automatic model- ing of other agents and also of human users/players. In this Section, we will overview related work about agent model- ing, considering whether models were reactive or had some form of internal memory, the power of the modeling lan- guage (propositional, first order, etc.), the task involved (classification, prediction, imitation, . . . ) , or the domain type (continous, noisy, . . . ) . We will also focus on aspects related to human modeling. One of the first attempts at modeling opponents was that of Carmel and Markovitch (1996). Here, the goal was to learn finite automatons (DFA) consistent with the behavior of the opponent. The model was used to improve a mini-max search algorithm. Interestingly, models included an internal state. This approach was valid only for discrete, round-based games, and required complete non-noisy information. Prediction was the goal, but such models could be used to imitate the opponent. Machine learning has been used extensively since then in classical and strategic games. Frnkranz and Kubat (2001) contains a good survey with some discussion of opponent modeling. Behavioral cloning is an attempt to imitate other agents behavior (Urbancic & Bratko, 1994; Bain & Sammut, 1999). Sammut, Hurst, Kedzierand, and Michie (1992) uses this technique to learn from a human piloting a sim- ulator, from whom input/output traces are obtained. It learns a decisión tree for each of the four plañe controls. It is a non-deterministic, quickly changing, noisy domain. The complete fly is divided into seven stages, each one with a different goal. Otherwise, similar situations (inputs) in different stages would have very different outputs. Goals are taken into consideration, and it distinguishes between two kinds of goals: homeostatic (non-persistent) and standard (persistent). The authors identify a problem when learning from different persons, because they use very different piloting styles. It differs with our work because the testing domains is different and no informa- tion about the stage or the current goal is supplied to the learning system. K N O M I C (Lent & Laird, 1999) is an approach that learns to imitate an expert from input/output traces in dynamic, non-deterministic, and noisy domains. It uses a rich knowledge representation based on SOAR rules. F o r instance, actions (operators) are decomposed into three kinds of rules: selection, application, and goal-detection. As in behavioural cloning, two kinds of goals are consid- erad: non-persistent and persistent. The expert is required to decompose a task into operators and sub-operators (a hierarchy of them, actually). This eliminates ambiguity between traces belonging to different parts of the task, which usually aim to achieve different goals (some research that tries to work around this problem, by inducing the agent's subgoals and using them to build the model, is reported in Suc & Bratko (1997)). The expert is also required to annotate the traces by telling the system which operators are selected and unselected. The authors claim that K N O M I C can learn from observing a human in diffi- cult domains such as air combat and Quake II. Successful results are only given from observing a hand-programmed agent. The authors identify a source of ambiguity for learn- ing from humans: they are less systematic, more variable, and make errors. There are many similarities with our approach, but no annotations are obtained from the user in our case. Bakker and Kuniyoshi (1996) present an overview of imi- tation in the field of robotics. They define imitation as "imi- tation takes place when an agent learns a behaviour from observing the execution of that behavior by a teacher" and summarize it as solving the problems of "seeing, understand- ing, and doing". However, most work is centered around the seeing and replicating sequences of actions (which in robot- ics is not a trivial task), and less about the understanding 2 part. Based on Piaget's work (1945), they bring up the important issue that "it is not possible to learn a new behav- ior unless one almost knows it already". Kuniyoshi et al. (1994) fits into this framework: a robot observes a human performing a simple assembly task and then reproduces it. Here, the problem that was addressed was to recognize the human actions (in terms of the ones that it already knew). The robot could also adapt to small changes in the position of Ítems on the table. In Hayes and Demiris (1994), a robot follows a teacher through a maze and learns to associate the environment, in terms of local wall positions, with the teacher's actions. In short, "if situ- ation then action" rules are learned. This second study is similar to our input/output rule learning, but the domain used is simpler. An área were human behavior ML modeling has been the focus is that of user modeling (Webb et a l , 2001). Early work in this field was about student modeling, which seeks to model the internal cognition of a student's cognitive system (as in Brown & Burton (1978)). This can be used to make online learning adaptive to the student skills and background knowledge, as well as to predict the student's actions. However, Self (1988) casted some doubt on the tractability of the cognitive approach. Since then, many researchers have followed a different approach that models an agent in terms of the relationships between its inputs and outputs. This approach, called input-output agent modeling (IOAM), treats the operation of the cog- nitive system as a black box. Some of the systems include feature based modeling (Webb & Kuzmycz, 1996), rela- tional based modeling Kuzmycz (1994), C4.5-IOAM Webb, Chiu, and Kuzmycz (1997), and F F O I L - I O A M Chiu, webb, and Kuzmycz (1997), among others. Both propositional and relational learning systems have been used. An usual testing ground is the problem of substrac- tion, where models are learned to predict the student response, including mistakes, when doing subtractions. In contrast with the Robosoccer, this is a discrete and sta- tic domain. Currently, the demands of electronic commerce and the Web have led to a fast growth in research in information retrieval, where ML can be used for acquiring models of individual users interacting with information systems (Bloedorn, Mani, & MacMillan, 1996) or grouping them into communities with common interests (Paliouras, Karkaletsis, & Papatheodorou, 1999). These models can help users in selecting useful information from the Web. Although user models are obtained, their domain and pur- pose is very different to ours. Their aim is to learn user pref- erences to filter Web information or to build adaptive interfaces. Also, users can be classified into stereotypes, so that user likes or dislikes can be predicted. ML techniques have also been applied to the prediction of user's actions, also called plan recognition. Kautz and Alien (1986) defined it as the problem of identifying a minimal set of top-level actions that are sufficient to explain the observed actions. Most work learns hierarchical plans from user logs, and associate them with the goal the user was trying to achieve (Bauer, 1999,1996). Later, plan libraries can be used to match actual user actions with a plan in the library and determine the user intentions. But their purpose is not to imi- tate the user, as in our research. Some research deals with agents modeling opponents and using this knowledge to beat them. In some cases, models are not learned, but predefined and used to classify opponent teams (not individual agents) by means of a sim- ilarity metric (Riley & Veloso, 2000a). Models can be used, for instance, to select the best plan to beat the opponent in set plays (Riley & Veloso, 2000a, 2001b). However, although Riley's RectGrid models can classify adversaries, they cannot be used to imítate opponent behavior. Riley's work has focused mostly on the Robosoccer domain, as in our case. Riley, Veloso, and Kaminka (2002) uses ML for a coach agent in the Robosoccer domain. That is, here learning takes place at a more strategic higher level. The coach can learn from previous games three kinds of models about teams: team formations and passing behaviour. These models can be used to predict and beat the opposite team, or more closely related to our research, so that the team imitates the modeled team. Differences with our research is that whole teams are modeled, and the coach agent is used to observe the playing field. In Stone (2000), a layered learning architecture is pro- posed. It is designed to use Machine learning in complex domains, where learning a direct mapping from sensors to actuators is not tractable, and a hierarchical task decom- position is given. Different learning mechanisms are used to learn behaviors from the bottom level (simple behaviors) to the highest level (more complex, strategic behaviors). The architecture is instantiated in the Robosoccer to learn an individual skill (ball interceptation), a multi-agent behavior (pass evaluation and selection), and adapting the team behavior (here, a new reinforcement learning algorithm called TPOT-RL is used Stone & Veloso (1999)). Although this work does not focus particularly on modeling other agents, it is relevant because if their working hypothesis is correct, it should be expected that learning the detailed models required to imítate other agents should require a layered architecture too (i.e. a direct mapping from inputs to outputs will be almost impossible to learn). This is very likely to be true in general, and should be taken into account in future research, but our work shows that learn- ing input-output mappings yield some positive results. In the opponent-modeling context, Bowling (2003) uses rein- forcement learning for multi-agent systems where other agents are also learning. There, the CMUDragons Robo- cup team is described, that is able to adapt online to an unknown opponent team. Finally, Ledezma, Aler, Sanchis, and Borrajo (2004) proposed a ML scheme to take advan- tage of prediction of opponents based on visual observa- tions in the Robosoccer simulator. In all these cases, the goal is learning to play and predict/adapt to opponents, but not imitation. 3 R. Aler et al. I Expert Systems with Sklar et al. (1999, 2001) is the first reported work (to our knowledge) where data collected from players playing a dynamic video-game, is used to train an agent. In this case, data was collected from humans playing Tron over the internet and a neural network was trained. Although they managed to créate effective controllers for this game, it is less clear that the resulting behavior imitated the humans. They bring up the issue that a person, under the same sit- uation can produce different responses, which can confuse the learning process. Spronck, Sprinkhuizen-Kuyper, and Postma (2004) uses a new technique called dynamic scripting in role playing games (RPG), to assign weights to rules in a reinforcement learning way. In Spronck, Sprinkhuizen-Kuyper, and Post- ma (2002) neural networks and genetic algorithms are used to online learning in R P G . In Ponsen and Spronck (2004) the same techniques are applied to real time strategy games. Recent research has shown a lot of interest in first per- son shooter games (FPS), like Quake. In this kind of games, human players must react to situations very quickly. So, it is a very suitable environment for learning reactive behaviors. They are played in networks by hun- dreds of people and records of the best games are kept, so there are good opportunities for Machine learning. Thu- rau, Bauckhage, and Sagerer (2004a) provide a good sum- mary of how imitation can be used at all levéis (reactive, tactical, strategic, and motion modeling ) in FPS games. In Thurau et al. (2003), the authors report some initial research in learning running and aiming behaviors from recorded human games. They built a MATLAB interface to allow a human to play Quake II and record pairs of state-vectors and actions. Then, they used a self-organizing map to reduce the dimensionality of state-vectors and multi-layer neural networks to map state-vectors to actions. Initial results in learning this kind of reactive behavior seem positive. In Bauckhage, Thurau, and Sager- er (2003), neural networks are used to learn trajectories, aiming behavior and their combination. They bring up the important issue of taking into account the context and the past, in addition to the state-vector, to reduce ambiguity when deciding which action to perform. Also in the Quake game, recent research tries to improve imita- tion models by means of genetic algorithms (Priesterjahn, Kramer, Weimer, & Goebels, 2005). In Bauckhage and Thurau (2004), tactical knowledge about which weapon to use is learned from human play by means of a mixture of experts. In Thurau, Bauckhage, and Sagerer (2004b), Neural Gas algorithms are used for learning the topology of the environment, and potential fields for learning human trajectories for picking up objects situated at different locations. This is strategic knowledge, because it tells which is the most important place to pick up the next object. The bot got stuck at some locations and temporal changes to the potential field had to be added (pheromone's trails). The author's claim that the bot imi- tated human behavior in simple setups and showed a mix- ture of intelligent behavior for more complex missions (i.e. Applications 36 (2009) 1850-1859 1853 less imitation but still clever movements). In Thurau, Bauckhage, and Sagerer (2004c), principal component analysis is used to extract primitive movements (building blocks for more complex sequences of movements) and conditional probabilities on the state-vector and the last action are learned. The artificial movements learned seem realistic and even certain human habits are preserved. 3. Our modeling approach 3.1. Modelling process As Fig. 1 shows, a human player interacts with an inter- face (the G U I soccerclient) that allows to play Robosoccer as a video-game. This G U I sends the human commands to the Soccerserver and displays the state of the field to the human. The interface has been carefully designed so that the only information that is displayed to the user is the one available to the actual agent in the simulated field. The trainer is used to set the playing field in a particular state, because many different states are required to learn general models. From the G U I , a trace is obtained by observing the human play. Records are obtained for every server cycle. This trace is made of many (s, a) such records, where s is the observation made by the agent sensors (distance to the ball, angle to the ball, etc). And a is the action carried out by the human player in that situation (for instance, kicking the ball, turning, etc.). Then, Machine learning techniques can be used to obtain a classifier that determines which action has to be carried out in a particular situation. Then, the classifier will be translated into C code, which will be used to control a soccer agent. If the modeling process is correct, the soccer agent will play Robosoccer similarly to the human. 3.2. The Robosoccer interface In order to build a good model for the soccer agent, the information available to the human through the interface must be as cióse as possible to that available to the agent. The XClient programmed by Itsuki N o d a 2 accomplishes this restriction. It is a first person interface, so the human player observes the objects in the field in 3D perspective. However, the versión of the Soccerserver we have used is 2D and it is a bit confusing to observe a 2D world in a first person view. Therefore, we decided to program our own interface, that is displayed in Fig. 2. This interface displays a complete 2D real time view of the field, just like the soccer monitor. Absolute positions are computed by means of the (Matellan, Borrajo, & Fer- nandez, 1998) library (trigonometry computations are used to obtain absolute positions of objects from the known position of the banners distributed along the field border). 4 Fig. 1. Process to obtain a model from a person playing Robosoccer by using Machine Learning. Fig. 2. 2D Interface. Objects are represented by probability circles. Although the whole field is visible, only those objects probability circle. Different colors3 are used to differentiate within the visión cone of the agent are displayed. Also, in the ball, the opponents, and the same team players. Those Robosoccer, perception of far away objects is noisy. Thus, objects are displayed as probability circles: the radius of the circle depends on the radius of the object and the distance 3 F o r ¡ n t e r p r e t a t i o n of c o l o r in F i g 2, the reader is referred to the web to the object. In Fig. 2, the ball is represented as a versión of this article. 5 objects which are no longer visible are represented at the last position they were seen. In order to improve playability, not all Soccerserver commands are available to the player. F o r instance, it is possible to kick the ball with any strength, but the interface only allows a standard kick. The commands allowed by the interface are: • turn left: the player can turn the agent's body a 10° to the left. The player is only allowed to turn left (but not right) because in some preliminary experiments we found out that it was very difficult for the Machine learning algorithm to discrimínate between turning left and right. • run slow/fast: only these two kinds of dash are allowed. Their power is 60 and 99, respectively. These valúes were obtained experimentally. • kick ball: kicks the ball in the direction of the agent's view line with a strength of 60. • kick to goal: kicks the ball towards the goal, with a strength of 99. 3.3. The trainer (coach) agent In order to have a diverse set of instances for learning behaviors, a diverse set of situations has to be presented to the human player. The trainer agent was used for this pur- pose. The trainer agent allows to position the agent, the ball, and same-team/opposite team agents in arbitrary positions in the field. Our trainer agent puts objects in random posi- tions in the field according to Fig. 3. In this way, an initially defensive positioning of the opposite team is achieved. 3.4. Opposite agents In the most complex situations, the human player will play within a team against a complete opposite team. In this paper we want to determine whether human input/out- put modeling of persons works in principie in the Robosoc- cer domain. To achieve this, we have used a simpler situation where a single human-controlled agent plays against a defensive opposite team (Camacho, Fernández, & Rodelgo, 2006; Fernandez, Gutiérrez, & Molina, 2000). This team is based on zones, where each of the team members is located. Among the agents, the player closer to the ball takes the role of leader. The rest of agents maintain a distance from the leader so as to maintain the formation. Agents determine who is the leader and pass this informa- tion to others. When the agent moves away from its zone, it tries to pass the ball to other agent, if available. Otherwise, it continúes towards the goal, in order to score. The goalie follows a similar behavior: it stays within its zone and looks for the ball. If it is cióse (15 units), goes for it and kicks it towards the opposite goal. 3.5. Model representation Human behavior can be modeled in many ways. Some of them have been used traditionally in AI: rules, trees, regression models, logic programs, etc. In this paper we intend to study how far first order representations can get. F o r this paper, we have chosen rules as a way of rep- resenting human behavior. They have a long tradition for representing knowledge and their conversión to C i f - t h e n - e l s e structures is straightforward. Also, we have used the C4.5 algorithms for generating the rules, although any rule-based ML algorithm would work just as well. Most specifically, the P A R T algorithm (Revisión 8 of C4.5) included in the Weka ML tool has been used (Ian, 2000). The rules follow a i f ( s i t u a t i o n ) t h e n a c t i o n structure, where the s i t u a t i o n checks the valúes of the agent's sensors and the a c t i o n tells what the agent should do next. Table 1 displays two actual rules obtained in the course of our research. With respect to the i f part of the rules, it is very impor- tant to select informative attributes so that they capture all information used by the human player to make decisions. Table 2 displays the attributes we have chosen. All positions Fig 3 Rectangles where opposite agents will be randomly positioned by the trainer 6 Table 2 Attributes used in the left hand side of rules Attribute X, Y Angle Anglejall Distance Ball Distance_Opposite 1 AngIe„Oppositel Valid Oppositel Distance_Opposite2 Angle_Opposite2 Valid_Opposite2 Angle_Opponent_goal Distance_Opponent_Goal Valid OpponenLGoal Meaning Absolute agent location Angle of the agent's view Une Angle between the ball and the agent's view line Distance from the ball to the agent Distance from the closest opposite player to the agent Angle between the agent and the closest opponent It indicates whether the closest opponent could be seen in the last server cycle Distance from the second closest opposite player to the agent Angle between the agent and the second closest opponent It indicates whether the second closest opponent could be seen in the last server cycle Angle between the opponent's goal and the agent Distance to the opponent's goal It indicates whether the opponent's goal could be seen during the last server cycle Type Real Real Real Real Real Real Boolean (0,1) Real Real Boolean (0,1) Real Real Boolean (0,1) of objects (ball, opponents, and goal) are relative to the view line of the agent and expressed in polar coordinates: dis- tance and angle. If the object is too far away, it cannot be seen and the valúes of these attributes are meaningless. This is indicated by other attributes, prefixed by v a l i d which tell whether the associated object was visible or not. Abso- lute attributes are only used for the X and Y coordinates of the agent. In this paper, only the two closest opponents are considered by the rules, although it would be easy to créate new attributes so that more opponents can be taken into account. The valúes that the right hand side of rules (the action part) can take are: kickóO, kick99, dashóO, dash99, tura 10, and turnminuslO. They have already been explained. Cur- rently, the interface can only use these discretized actions, but in the future, it could be modified so that the human player can select a continuous valué to turn, to kick, or to dash. 4. Training the agent Some preliminary experiments were carried out for test- ing simple behaviors like looking for the ball and advanc- ing with the ball in an empty field. As these behaviors were easily learned and properly performed by the agent, we proceeded to more complex behaviors, which involve playing against opponents. 4.1. Dribbling static opponents This skill involves a striker advancing with the ball and scoring, after dribbling static opponent agents located near the goal. These opponents can only kick the ball when it comes cióse to them. Eleven opponent players are used. The training cases will be generated in such a way that the striker is forced to dribble opponents to find the ball, and then continué dribbling them until it scores. A first attempt was done with 5370 training instances, obtaining a 95.88% 10-fold cross-validation accuracy. The agent is able to find the ball when there are no oppo- nents and conduct it to the goal. The agent also does well when confronting the 11 opponents. However, in some cases it displays the following flawed behaviors: • The agent tries to kick the ball when it is too far away, or when the angle is not appropriate. • When the agent was cióse to the ball, it turns left again and again. • The agent collides with an opponent and stops there. Once the agent performs these flawed behaviors, it never gets out of these states. So for instance, the agent will try to kick the ball forever, or it will get into an eter- na] turning loop. In general, and this is a property of our reactive approach, if for some reason, the agent performs an action that does not change its environment, or that it changes it but this change is not perceived by the left hand side of the rules, the agent will get stuck in that behavior forever. F o r instance, when the agent tries to kick the ball, but it is not cióse enough, nothing has chan- ged in the world, so the agent will repeat its kicking behavior again and again. Similarly, if a chain of rule actions gets the agent to do a complete 360° loop, this will be done forever. Perhaps a mechanism should be added on top of the rules, that realizes when the agent has got into such states and do some random actions until it gets out. However, in this paper we only want to study the puré learning approach, so we will leave that for the future. In order to improve this behavior, the number of train- ing instances was increased to 14,915, obtaining 172 rules, and a 95.11% accuracy. The behavior improved, but the agent still performs flawed behaviors. These flaws restrain the agent from fulfilling its objective. Our final agent dis- played flawed behaviors in 12% of the triáis (a trial involves letting loóse the agent in the field, finding the ball, and 7 scoring). We found very hard to improve these results by adding more training instances. In the conclusions section, we will discuss why this is so and propose new lines of research to overeóme these limitations. Thus, results are not perfect but we considered it to be acceptable. Also, these behaviors happen in a world where the only agent that can initiate actions (and change the world) is the striker. When there are more active opponents in the field, the world will change independently of the agent, and the agent will get out of its static states more easily. 4.2. Match with opponents This is the most complex behavior learned: a striker must get to the ball and score against three defences and one goalie. For this task, it would seem that it would be desirable to choose very diíBcult opponents, like CMUnit- ed or FC-Portugal, which were previous Robocup champi- ons. However, the human player found impossible to beat them. This was due to these teams playíng extremely well, but also to the interface not being responsive enough. This latter problem could not be solved, because the Robosoccer server was not designed with interactive play in mind. As our aim is to show that human experience can be transferred to soccer agents, we have chosen a challenging but beatable Robosoccer team (Camacho et al., 2006; Fer- nandez et al., 2000), which has the advantage that although their players have a team behavior, we can use as many players as desired. In this case, only four of them were used. In any case, it must be remarked that, although the (Camacho et al., 2006; Fernandez et a l , 2000) team is not a Robocup champion, it is still a very challenging situ- ation, because: • The opponents outnumber our agent and play cooperatively. • The opponent can use more actions (turning the neck, turning left and right, kicking and dashing with any power and angle, . . . ) . • The opposite team has a goalie, whereas our agent must defend and attack. By confronting a human player against this team, we were able to learn rules that could be transferred to an agent that performed very well, as it will be shown next. 16,124 instances were obtained and 234 rules were cre- ated, with a 93.44% cross-validation aecuracy. Then, the agent had to play in six new testing matches. Although the agent did not win any of the six testing matches, it scored some goals. Results were: 2 - 5 , 1-6, 0-5, 1^1, 1-5, 0-4 (where the first number indicates goals scored by the agent, and the second one, goals scored by the opponents). The agent incurs in previous flawed behaviors like trying to repeatedly kick the ball when it is not there. However, in this case the world is more dynamic and when an opponent or the ball comes cióse, the agent gets out of the loop, and reaets. In order to improve these results, we increased the num- ber of instances to 24594. 332 rules were generated (93.65%). In this case, we also pruned the rules using W E K A ' s standard parameters. The number of rules was reduced to 164 (92.65%). Yet, the behavior was greatly enhanced: the agent was able to find the ball on the field, to conduct it towards the goal, to score, to dribble oppo- nents, and to steal the ball from them. The scores in six matches display this improvement, as the agent won one of the games: 5-4, 2-4, 3^1, 4 - 5 , 2^1, 3^1. The agent scored 19 goals versus 25 goals scored by the opponents. The learned classifier was further pruned to 69 rules (91.90%). Similar results were observed in six new matches: 3—4, 2- 4, 3-3, 2-5, 3 - 1 , 4 - 3 . The agent scored 17 goals versus 20 goals scored by the opponents. Table 3 summarizes these results. 5. Conclusions and future work In this paper we have applied an input-output modeling approach to model a human playing Robosoccer. First, an interface was built that displayed to the user the objeets in the playing field that could be seen according to Robosoc- cer rules. This interface allowed the user to send low-level commands (dash, turn, and kick) to the Soccerserver. Input/output instances generated by the human player were used by a Machine learning algorithm (PART) to learn a model. This model was then introduced into a com- puter agent. Results show that in different low-level behav- iors, like looking for the ball, conducting the ball to the goal, dribbling opponents, and scoring in the presence of other players, our approach works well. The final agent was able to score many goals against a computer team that the human found challenging. As far as we know, this is the first time that behavioral cloning techniques have been applied in the Robosoccer domain, with positive results. This shows that this is a very promising line of research whose results could be improved further, as discussed below. Building a user-friendly and responsive interface is of great importance for the human play. Unfortunately, the Soccerserver was not thought as a video-game and it is dif- ficult to construct a responsive enough interface for it. Our current interface is still not as good as commercial video- games. It would be possible to overeóme this issue by rep- licating the functionality of the Soccerserver, but bearing in Table 3 Summary of results in six testing matches for the "playing with opponents" behavior Instances 16,124 24,594 24,594 Pruning No Yes Yes Rules 234 332 69 Aecuracy 93.44%, 92.65% 91.90% Goals scored +5 +19 + 17 Goals against -29 - 2 5 - 2 0 8 mind interactive play. Rules could be learned from the modified Soccerserver and transferred to agents playing the actual Soccerserver, after, perhaps, some small adapta- tion. Having a responsive interface is very important for learning low-level behaviors. We have found out that the agent displayed some flawed behaviors, although not very frequently. This problem could be reduced by increasing the size of the dataset and pruning the model. However, the problem did not disap- pear completely. We believe that the underlying assump- tions of a purely input-output behavioral approach may be the culprit. Our approach works well when the behavior to be cloned is reactive (i.e. the behavior is an input-output map). But if the action of the human depends on hidden variables, in addition to what the human can see on the field, the model of the human will degrade. F o r instance, the human can make use of memories and predictions about the opponents, even when he is not watching them. But these variables are hidden to the modeling algorithm (i.e. it is not possible to see what the human is thinking). Therefore, our approach worked well because the behav- iors to be learned are mostly reactive, but even in this case, there are probably some hidden variables that could help to improve the results. In the future, we plan to add estimations of some hidden variables to the agent, via new attributes, computed by spe- cial purpose algorithms. If human-models are to be used, we should delve more into the cognitive functions applied by a person when playing Robosoccer (like planning, opponent prediction, trajectory computation, . . . ) . These cognitive abilities could be supplied to the agent, and used in the left hand side of the rules via new attributes. For instance, humans use memory to keep track in their minds of opponents and the ball, even when these objects are out of view. In the same way, tracking algorithms could be used to genérate attributes that estímate where the ball might be at some particular time. Imitating humans in the Robosoccer can be done at many levéis. Inspired in computer games like F I F A 2006, we intend to let the human use higher level actions like passing the ball, shooting to the goal, pushing the ball, dribbling, etc. In this case, the human player will only have to press a key, and the computer will carry out a pre-pro- grammed behavior (for passing the ball, etc.). Thus, the human can focus on a more strategic level and leave the low-level details to the computer. Modeling can be done at even higher levéis, as the team level, or the coach agent. We would also like to test the approach in more complex situations like real matches and real team play. Acknowledgements We would like to thank Vicente Matellan for letting us use his ABC2 routines, and Fernando Fernandez for provid- ing us with his useful Robosoccer team. Agapito Ledezma has been very helpful with his knowledge of Robosoccer. References Bain, M., & Sammut, C. (1999). A framework for behavioral cloning. In Machine intelligence agents (pp. 103-129). Oxford University Press. Bakker, P., & Kuniyoshi, Y. (1996). Robot see, robot do: An overview of robot imitation. In AISB' 96 workshop in robots and animáis (pp. 3- 11). Bauckhage, C, & Thurau, C. (2004). Towards a fair'n square aimbot - Using mixtures of experts to learn context aware weapon Handling. In Proceedings of the GAME-ON (pp. 20-24). Bauckhage, C, Thurau, C, & Sagerer, G. (2003). Learning human-like opponent behavior for interactive computer games. In B. Michaelis & G. Krell (Eds.), Pattern recognition. Lecture notes in computer science (Vol. 2781, pp. 148-155). Springer-Verlag. Bauer, M. (1999). From interaction data to plan libraries: A clustering approach. In International joint conference on artificial intelligence (pp. 962-967). Bauer, M. (1996). Machine learning for plan recognition. In Machine learning meets human computer interaction. Workshop of the Interna- tional conference on machine learning (pp. 5-16). Bloedorn, E., Mani, I., & MacMillan, T. R. (1996). Machine learning of user profiles: Representational issues. In Thirteen national conference on artificial intelligence (pp. 433-438). Bowling, M. (2003). Multiagent learning in the presence of agents with limitations. PhD thesis, Computer Science Department. Pittsburgh, PA: Carnegie Mellon University, May 2003. Available as technical reportCMU-CS-03-118. Brown, J. S., & Burton, R. R. (1978). Diagnostic models for procedural bugs in basic mathematical skills. Cognitive Science, 2, 155-192. Camacho, D., Fernández, F., & Rodelgo, M. A. (2006). Roboskeleton: An architecture for coordinating robot soccer agents. Engineering Appli- cations of Artificial Intelligence, 19(2), 179-188. Carmel, D., & Markovitch, S. (1996). Opponent modeling in multi-agent systems. In Adaptation and learning in multiagent systems. IJCAI' 95 Workshop. Lecture notes in computer science (Vol. 1042, pp. 40-52). Springer. Chiu, B. C, Webb, G. I., & Kuzmycz, M. (1997). A comparison of first- order and zeroth-order induction for input-output agent modelling. In Proceedings of the sixth International conference. Springer. Fernandez, F., Gutiérrez, G. & Molina, J. M. (2000). Coordinación global basada en controladores locales reactivos en la robocup. In Workshop Hispano-Luso de Agentes Físicos (pp. 73-85). Tarragona, España. Frnkranz, J., & Kubat, M. (Eds.). (2001). Machines that learn to play games. Nova Science Publishers. Hayes, G., & Demiris, J. (1994). A robot controller using learning by imitation. In Proceedings of the second international symposium on intelligent robotic systems (pp. 198-204). Kautz, H., & Alien, J. F. (1986). Generalized plan recognition. In Proceeding ofthe AAAInational conference on artificial intelligence (pp. 32-37). Kuniyoshi, Y., Inaba, M., & Inoue, H. (1994). Learning by watching: Extracting reusable task knowledge from visual observation of human performance. IEEE Transaction on Robotics and Automation, 10(6), 799-822. Kuzmycz, M. (1994). A dynamic vocabulary for student modeling. In Proceedings of the fourth international conference on user modeling (pp. 185-190). Ledezma, A., Aler, R., Sanchis, A. & Borrajo, D. (2004). Predicting opponent actions by observation. In RoboCup international symposium 2004 (RoboCup2004), Lisbon (Portugal). Luke, S , Hohn, C, Farris, J. Jackson, G., & Hendler, J. (1997). Co- evolving soccer softbot team coordination with genetic programming. In Proceedings of the first international workshop on RoboCup, at the international joint conference on artificial intelligence, Nagoya, Japan. Matellan, V., Borrajo, D., & Fernandez, C. (1998). Using ABC2 in the Robocup domain. In Robocup-97 Robot soccerworld cup I. Lecture notes in artificial intelligence (pp. 475^183). Springer-Verlag. 9 Paliouras, G , Karkaletsis, V., Papatheodorou, C, & Spyropoulos. C. D. (1999). Exploiting learning techniques for the acquisition of user stereotypes and communities. In Seventh international conference on user modelling (pp. 169-178). Piaget, J. (1945). Play, dreams and imitation in childhood. Heinemann. Ponsen, M , & Spronck, P. (2004). Improving adaptive game ai with evolutionary learning. Computer games Artificial intelligence, design and educalion, 389-396. Priesterjahn, S , Kramer, O., Weimer, A., & Goebels, A. (2005). Evolution of reactive rules in multi player computer games based on imitation. In International conference on natural computation (ICNC 05) Changsha, China. Ríley, P., & Veloso, M. (2000a). On behavior classification in adversarial environments. In Distributed autonomous robotic systems 4. Springer- Verlag. Ríley, P., & Veloso, M. (2001b). Coaching a simulated soccer team by opponent model recognitíon. In Proceedings ofthe agents international conference (pp. 155-156). ACM Press. Ríley, P., Veloso, M , & Kaminka, G. (2002). An empirical study of coaching. In Proceedings of DARS-2002, the seventh international symposium on distributed autonomous robotic systems. Sammut, C, Hurst, S , Kedzierand, D., & Michie, D. (1992). Learning to fly. In D. Sleeman (Ed.), Proceedings of the ninth international conference on machine learning (pp. 385-393). Morgan Kaufman. Self, J. A. (1988). Bypassing the intractable problem of student modelling. In Proceedings of the intelligent tutoring systems conference (pp. 107- 123). Sklar, E., Blair, A. D„ Funes, P., & Pollack, J. (1999). Training intelligent agents using human internet data. In Proceedings of the first Asia- Pacific conference on intelligent agent technology (IAT-99) (pp. 354- 363). Sklar, E„ Blair, A. D., & Pollack, J. B. (2001). Training intelligent agents using human data collected on the internet. In Agent engineering (pp. 201-226). World Scientific. Spronck, P., Sprinkhuizen-Kuyper, I., & Postma, E. (2002). Improving opponent intelligence through machine learning. In Proceedings of the fourteenth Belgium-Netherlands conference on artificial intelligence (pp. 299-306). Spronck, P., Sprinkhuizen-Kuyper, I., & Postma, E. (2004). Online adaptation of computer game opponent ai. International Journal of Intelligent Games & Simulation (IJIGS), 5(1), 45-53. Stone, P. (2000). Layered learning in multiagent systems: A winning approach to robotic soccer. MIT Press. Stone, P., & Veloso, M. (1998). A layered approach to learning client behaviors in the RoboCup soccer server. Applied Artificial Intelligence, 12, 165-188. Stone, P , & Veloso, M. (1999). Team-partitioned, opaque-transition reinforcement learning. In M. Asada & H. Kitano (Eds.), RoboCup-98 Robot soccer world cup II. Berlín: Springer-Verlag, Proceedings of the third international conference on autonomous agents. Suc, D., & Bratko, I. (1997). Skill reconstruction as induction of lq controllers with subgoals. In Proceedings ofthe 15th international joint conference on artificial intelligence (Vol. 2, pp. 914-920). Thurau, C, Bauckhage, C, & Sagerer, G. (2003). Combining self- organizing maps and multilayer perceptrons to learn bot-behavior for a commercial computer game. In Proceedings of the GAME-ON (pp. 119-123). Thurau, C, Bauckhage, C, & Sagerer, G. (2004a). Imitation learning at all levéis of game-AI. In Proceedings ofthe international conference on computer games, artificial intelligence,design and educalion (pp. 402- 408). Thurau, C, Bauckhage, C, & Sagerer, G. (2004b). Learning human-like movement behavior for computer games. In Proceedings of the eighth international conference on the simulation of adaptive behavior (SAB'04). Thurau, C, Bauckhage, C, & Sagerer, G. (2004c). Synthesizing move- mentsfor computer game characters. Lecture notes in computer science (Vol. 3175). Heidelberg, Germany: Springer-Verlag. Urbancic, T., & Bratko, I. (1994). Reconstructing human skill with machine learning. In European conference on artificial intelligence (ECAI1994) (pp. 498-502). van Lent, M., & Laird, J. (1999).. Learning hierarchical performance knowledge by observation. In Proceedings ofthe sixteenth international conference on machine learning (pp. 229-238). Morgan Kaufmann Publishers Inc. Webb, G., Pazzani, M., & Billsus, D. (2001). Machine learning for user modeling. User modeling and user-adapted interaction, 7/(19-20). Webb, G. I., Chiu, B. C, & Kuzmycz, M. (1997). Comparative evaluation of alternative induction engines for feature based modelling. Interna- tional journal of artificial intelligence in educalion, 8, 97-115. Webb, G. I., & Kuzmycz, M. (1996). Feature based modelling: A methodology for producing coherent, dynamically changing models of agents' competencies. User modeling and user-adapted interaction, 5(2), 117-150. Witten, I. H., & Frank, E. (2000). Data mining Practical machine learning tools and techniques with java implementations. Morgan Kaufman. 10 
work_zbyog7wmjbgfnpiy3pwakoakz4	----	nil02.tif A Simple Expert System Shell for Microcomputer-Aided Radiographic Diagnosis M o r g a n G. D u n n e a n d S e a n B. D u n n e A simple microcomputer-basod e x p e r t system shell for radiographic diagnosis is described. T h e system interrogates t h e user by means o f a simple question- a n d - a n s w e r approach on t h e radiographic f i n d i n g s in a given case and proffers a differential diagnosis f r o m a database. T h e system a l s o a l l o w s l i s t i n g o f both t h e radiographic findings p e r t i n e n t t o a specific diagnosis and all diagnoses in w h i c h a particular f i n d i n g of combination of findings occur. T h e system database can be easily updated or modified by t h e user. T h e demonstration application for t h e e x p e r t system shell is t h e plain film diagnosis of osseous dysplasia. �9 1990 by W.B. Saunders Company. KEY W O R D S : computers, software, e x p e r t systems, radiology, artificial intelligence, osseous dysplasia, b o n o diseases. T H E USE O F personal computers (PC) and expert systems as an aid in radiographic diagnosis is being reported with increasing frequency.l-4 Many of these systems are complex, memory-expensive, and require mainframe or minicomputer power. The system described here was developed to acquire an introductory level understanding of expert system programming on a microcomputer (PC) and apply such program- ming to a specific problem in diagnostic radiol- ogy. Differential diagnosis of the osseous dysplasias remains a problem for the general radiologist, as many of these conditions are rare and the diagnos- tic criteria are both numerous and nonspecific. It was felt that the interpretation of radiographs on these patients would be facilitated by the applica- tion of a simple user-friendly expert system running on a typical microcomputer (PC), which would pose a series of questions regarding spe- cific radiographic findings to be answered by using only three k e y s - - " y " for "yes," "n" for "no," or "d" for "don't know." From the Department of Radiology, Baptist Memorial Hospital System, San Antonio, TX. Address reprint requests to Morgan G. Dunne, MA, MB, BCh, BAO, 15301 Winter Mist, San Antonio, TX 78247. �9 1990 by W.B. Saunders Company. 0897-1889/90/0301-0004503.00/0 PROGRAMMING LANGUAGE The programming language selected for the system development was T U R B O P R O L O G version 1.1 (Borland International, Scotts Val- ley, CA), which is a compiler-based implementa- tion of PROLOG, a fifth-generation artificial intelligence language ideally suited to expert system and database design. 5'6 The programming environment offered by T U R B O P R O L O G is quite easy to use with a word-processing style text editor, and provides extensive debugging aids that quickly locate and display errors in the source cede for correction. Once compiled, the program runs at interactive speeds on an IBM PC-XT (IBM Corporation, Boca Raton, FL) even during extensive database analysis, in con- trast to other relational database programs. The P R O L O G language is a declarative rather than procedural language, meaning that it attempts to solve a problem by deductive reasoning given a set of facts and rules. Thus, the programmer has only to provide a description of the problem and general set of rules for solving it and the PRO- LOG system will do the rest. E X P E R T S Y S T E M The expert system consists of two parts, the knowledge base (KB) and the inferenee origine (IE). 7'8 The KB is a database that holds a list of objects (eg, osseous dysplasia diagnoses) and their associated rules (lists of incompatible radio- graphie findings) and attributes (radiographic findings pertaining to a particular diagnosis). The KB may reside on a disk and be accessed as needed by the program, or it may be incorporated into the program cede as was done in this case to allow the program to run faster without the need for pauses while the disk is read. The KB predicates and clauses c o n t a i n a descriptor that indicates to which application they belong ( " O D " in this case for osseous dysplasia). Thus, the IE is designed to operate with other user-selected data- bases, which may be either maintained on a disk or embedded within the program. 8 Each radio- graphie finding clause in the KB is numbered and the diagnosis clauses contain a list of integers designating the associated radiographie findings. 10 JournalofDigitallmaging, Vol 3, No 1 (February), 1990: pp 10-14 COMPUTER-AIDED RADIOGRAPHIC DIAGNOSIS 1 1 The data on the t 06 diagnoses of osseous dyspla- sia for the demonstration KB were obtained from standard s o u r c e s . 9 q 2 The l E is the part of the program that at- tempts to find a matching diagnosis given the radiographic findings which are determined to be present or absent based on the user's responses to the questions generated by the system. The l E used in this application is deterministic. 7 This means that it will only proffer a diagnosis if the radiographic findings associated with that diagno- sis are either confirmed by the user or the user indicates that he or she does not know whether or not some or all of the findings associated with that diagnosis are present. The l E in this system uses the backward-chaining or object-driven method. 7'8 In this method, the program assumes a hypothesis which ir attempts to prove by taking the first object (diagnosis) in the database and asking the user if each of the findings associated with that diagnosis are present. I f any of the findings are indicated by the user not to be present (signified by pressing the " n " key), then the IE will reject the diagnosis and move on to the next diagnosis and repeat the process. If a given finding is given a " d o n ' t know" answer by the user (signified by pressing the " d " key), then the IE assumes that finding is present and re- mains with the diagnosis in question and contin- ues to ask questions about the other findings associated with the diagnosis. The l E maintains two memory resident data- bases (MRDs) which keep track of the " y e s " and " n o " responses so that the user is not asked the same question twice. " D o n ' t know" answers are stored as " y e s " responses. Thus, when each diagnostic hypothesis is picked up by the IE, the associated radiographic findings are checked in turn against the M R D s to determine if a " y e s " or " n o " response has already been elicited for that finding. I f a given finding is found in the " n o " M R D , then the diagnostic hypothesis in question is rejected and the next diagnosis in the database is placed under consideration. If the finding is found in the " y e s " M R D , then the l E moves on to the next finding associated with that diagnosis and repeats the process. I f the finding is not found in either of the " y c s " or " n o " MRDs, then the user is questioned as to whether or not the finding is present. The IE contains three rule-based subroutines, each with an associated database written into the program which contains a specific subset of radiographic findings. The purpose of these sub- routines is to reduce the number of questions that the user must answer in order to obtain a complete differential diagnosis. The first of these deals with normalcy in a specific anatomic area. For example, if the user's response is " y e s " to the question "normal skull and facial bones?", then all radiographic findings relating to the skull and facial bones are appended to the " n o " M R D . Thus, if the skull and facial bones are normal in a given case, the user is not asked any further questions regarding this anatomic area. The second subroutine deals with specific interrelated and incompatible radiographic findings. For ex- ample, if the user confirms the presence of micrognathism, then the finding of prognathism is appended to the " n o " M R D . The third subrou- tine deals with opposite and incompatible radio- graphic findings. For example, if the user enters " n " to the question " h e m a n g i o m a t a present?", then the finding " h e m a n g i o m a t a absent" is ap- pended to the " y e s " MRD. USER INTERFACE The user interface consists of several menu- containing windows that are called to the screen according to keyboard input. The main menu (Table 1) consists of five choices, the first of which activates the interactive expert system and l E leading to a differential diagnosis based on a ',thorough interrogation of the user through a "diagnostic findings" window. Each question is displayed on the screen and the user enters " y , " " n , " or " d . " The user's response is displayed in full after each question (Table 2). Each time the lE establishes a diagnosis the screen switches to a "differential diagnosis" window and the diagno- Table 1. System Main Menu ODLOG.PRO: Osseous Dysplasia Logic-Based Expert System Welcome to ODLOG.PRO ! 1. Consult Expert System 2. List AII Diagnoses on Knowledge Base 3. List Findings for Specific Diagnosis 4. List Diagnoses with Specific Finding(s) 5. Exit Expert Systern Please enter your choice: 1-5 12 Tabla 2. Typical Dialogue With the Expert Syatem ODLOG.PRO: Osseous Dysplasia Expert System Consultation 8/8/89 14:5:17 normal skull and facial bones?: no normal spine?: no normal pelvis?: no normal humeri, scapulae and clavicles?: no normal radii and ulnae?: no normal hands, wrists and feet?: don't know normal femora and patellae?: no normal tibiae and fibulae?: no normal sternum?: don't know normal ribs?: no stippled epiphyses prior to four years of age?: no normal bone maturation?: yes generalized metaphyseal flaring?: no brachycephaly?: no sclerosis of skull?: no short skull base?: yes small facial bones?: yes intrinsically normal epiphyses?: yes short femora?: yes posterior scalloping of vertebral bodies?: yes short pedicles?: yes narrow interpedicular distance in lumbar region?: yes increased lumbar Iordosis Tn late childhood?: yes marked backward angulation of the sacrum?: yes lumbar kyphosis in infancy?: don't know short ribs in infnancy?: don't know generalized shortening of tubular bones?: yes flat acetabular roofs?: yes spur at medial edge of triradiate cartilage in pelvis?: yes tibia vara?: yes y ACHONDROPLASlA �87 hypertelorism?: no scaphocephaly?: no etc . . . . Data run complete, consultation terminated. sis or diagnoses are displayed. The user returns to the diagnostic findings window by pressing the spacebar and the interrogation resumes as the IE searches for additional diagnoses. The second choice from the main menu simply lists all the diagnoses contained in the database (106 at the time of this writing), one screenful at a time. The third menu choice allows the user to s e l e c t a specific diagnosis from a submenu and view all the radiographic findings associated with that diagnosis (Table 3). The fourth menu choice allows the user to select one or more radiographic findings from a submenu and have the system display all diagnoses on the database that are associated with those findings (Table 4). Thus, for example, the system could list all diagnoses DUNNE AND DUNNE Tabla 3. List of Findings for Specific Diagnosis ODLOG.PRO: Osseous Dysplasia Expert System Consultation 8/8/89 14:8:23 Metatropic Dwarfism Normal skull and facial bones Platyspondyly Dysplastic odontoid process Coronal cleft vertebra Isolated irregularity of vertebral endplates Scoliosis Flat acetabular roofs Sharp lateral beak on acetabular roofs in childhood Short ribs in infancy Delayed bone maturation Generalized shortening of tubular bones Generalized metaphyseal flaring Generalized irregular and fragmented epiphyses Wide femoral necks Data run complete, consultation terminated. associated with platyspondyly and generalized tubular bone shortening (Table 4). The fifth choice on the main menu exits the system. HARD COPY Printed hard copy can be obtained from any of the first four main menu choices. The system accomplishes this by opening a diskfile when one of the first four choices is selected. Everything that is displayed on the screen (except menus) is also written in this file and at the end of each consultation the user is offered the option of printing this file or simply returning to the main menu. An example of a printout of a typical dialogue with the expert system is shown in Table 2. At the end of each consultation the diskfile is Table 4. List of Diagnoses W i t h Specific Finding(s) ODLOG.PRO: Osseous Dysplasia Expert System Consultation 8/8/89 14:11:12 platyspondyly generalized shortening of tubular bones SPONDYLOMETAPHYSEAL DYSPLASlS--TYPE KOZLOWSKI/SUTCLIFFE ACHROMESOMELIC DWARFISM--TYPES MAROTEAUX/CAMPAILLA-MARTINELLI PSEUDOACHONDROPLASIA METATROPIC DWARFISM PARASTREMMATIC DWARFISM THANATOPHORIC DWARFISM MORQUIO SYNDROME (MPS IV) OCULOMANDIBULOFACIAL SYNDROME (HELLER- MAN N-STREIFF-FRANCOIS) Data run complete, consultation terminated. COMPUTER-AIDED RADIOGRAPHIC DIAGNOSIS 13 erased. Each printout is stamped with the date and time of the consultation. KNOWLEDGE ENGINEERING Perhaps the most difficult aspect of expert system development is the organization of the knowledge database for optimum efficiency. Ide- ally, one would like the system to provide a list of differential diagnoses in descending order of probability, asking the user as few questions as possible. In order to develop this ideal system a probabilistic rather than deterministic lE would be required and the knowledge base would need to contain hard data regarding the prevalence and incidence of the various diagnoses, discrimi- nant criteria for the radiographic findings associ- ated with a given diagnosis, and positivity crite- ria for each of the various radiographic findings. The sensitivity and specificity of the radiographic test (plain film in this case) also would need to be included in the database. 13 Because such hard data is rarely if ever available in those problems relating to diagnostic radiology, one is forced to settle for a deterministic system as described previously or to develop one's own scoring system to indicate the likelihood of each of the diagnoses in the differential. ~4 The diagnoses in the system described previously can be arranged in the database in descending order of likelihood accord- ing to the user's practice. By this means the lE will attempt to confirm or reject the more com- mon diagnoses first and the least likely diagnoses last. The inclusion or deletion of specific radio- graphic findings from the database is a poten- tially subjective undertaking because the positiv- ity criterion for a particular finding may vary from one radiologist to another. Thus, a given radiologist user may either find that a particular finding either does not cross his or her threshold sufficiently often to allow reasonable sensitivity in making a diagnosis or may feel that the finding is not particularly specific. 13 Because the absence of a particular finding is crucial in the rejection of a diagnosis in the system described previously, it was felt necessary to include the "don't know" option to allow the user a way of handling uncertainty regarding a particular finding in a given case. Until more experience is gathered in the clinical setting with the system described and similar systems, we must await the establishment of an "official" set of radiographic findings for application to a particular radiologic problem. The approach used in the system described was to include initially a large number of findings (250 at time of this writing) and hopefully whittle these down to a smaller, more discrimi- nant number as experience with the system is acquired. For example, the finding of scoliosis, while a prominent feature of many osseous dyspla- sias, is not particularly useful in their differentia- tion as it may occur in isolation or be associated with almost any spinal anomaly. On the other hand, and hypothetically, if scoliosis never oc- curred in a specific dysplasia, then its inclusion in the database would be of value as its absence in a given case may help confirma diagnosis. SYSTEM REQUIREMENTS The T U R B O P R O L O G environment requires at least 384K of R A M and the stack size should be set to 2,400 paragraphs. The program source code takes up 14K, and 26K was required for the KB at time of writing. Thus, the entire system takes 40K with the KB embedded in the pro- gram. An option programmed into the user interface, but not yet added to the KB, is the ability of the system to display an explanation of the radiographic finding while being questioned by the lE. The finding explanation window is accessed by pressing the "e" key in response to a question by the system and the interrogation is resumed by pressing the spacebar after the expla- nation has been read by the user. The addition of explanations for all radiographic findings proba- bly will not be required, as many are self expian- atory. Nevertheless the memory requirements for the KB is expected to increase to approxi- mately 50K with the addition of these data. The expert system shell described offers a fast, inexpensive approach to computer-aided diagno- sis of certain specific radiologic problems. It may be used a s a simple database manager without activating the lE. The powerful word processing ability of the T U R B O P R O L O G editor allows rapid deletion or addition of diagnoses and radio- graphic findings to the database by the user. With increasing introduction of computer tech- nology to the radiology department it may be- hoove the practicing radiologist to become more familiar with medical expert systems and termi- nology. It is hoped that with increased input from such individuals systems with enhanced practical clinical usefulness will be developed. 14 DUNNE AND DUNNE REFERENCES 1. Greenes RA: Computer-aided diagnostic strategy selec- tion. Radiol Clin North Am 24:105-120, 1986 2. Kahn CE, Messersmith RN, Jokich MD: PHOENIX: An expert system for selecting diagnostic imaging proce- dures. Invest Radio122:978-980, 1987 3. Swett HA, Miller PL: ICON: A computer-based ap- proach to differential diagnosis in radiology. Radiology 163:555-558, 1987 4. Swett HA, Fisher PR, Cohn Al, et al: Expert system- controlled image display. Radiology 172:487-493, 1989 5. Schildt H: Artificial inteUigence: A quick overview, in Schildt H: Advanced Turbo Prolog a. Berkley, CA, Osborne/ McGraw-Hill, 1987, pp 1-12 6. Yin KM, Solomon D: Introduction, in Yin KM, So- lomon H (eds): Using Turbo Prolog a. Indianapolis, IN, Que~ Corporation, 1987, pp 1-6 7. Schildt H: Artificial intelligence: Expert systems, in Schildt H: Advanced Turbo Prolog a. Berkley, CA, Osborne/ McGraw-Hill, 1987, pp 57-90 8. Yin KM, Solomon D: Creating expert systems, in Yin KM, Solomon D: Using Turbo Prolog a. Indianapolis, IN, Que ~ Corporation, 1987, pp 387-434 9. Wynne-Davies R, Hall CM, Apley AG: Atlas of Skele- tal Dysplasias. Edinburgh, Scotland, Churchill Livingstone, 1985 10. Taybi H: Radiology of Syndromes and Metabolic Disorders (ed 2). Chicago, Year Book Medical, 1983 I 1. Goldman AB: Collagen diseases, epiphyseal dyspla- sias, and related conditions, in Resnik D, Niwayama G (eds): Diagnosis of Bone and Joint Disorders, vol 5. Philadelphia, PA, Saunders, 1988, pp 3374-3441 12. McAlister WH: Osteochondrodysplasias, dysostoses, chromosomal aberrations, mucopolysaccharidoses, and mu- colipidoses, in Resnik D, Niwayama G (eds): Diagnosis of Bone and Joint Disorders, vol 5. Philadelphia, PA, Saunders, 1988, pp 3442-3515 13. Chang P: Bayesian analysis revisited: A radiologist's survival guide. AJR 152:721-727, 1989 14. Getty D J, Pickett RM, D'Orsi C J, et al: Enhanced interpretation of diagnostic images. Invest Radiol 23:240- 252, 1988 
work_zdavkplsyndrvpy6xenfzuotre	----	Concept embedding-based weighting scheme for biomedical text clustering and visualization Concept embedding‑based weighting scheme for biomedical text clustering and visualization Xiao Luo1* and Setu Shah2 Introduction Active research and practice in the medical domain has generated pervasive text files, articles, and documents, which include MEDLINE—the largest biomedical text data- base, clinical notes in the Electronic Health Records, descriptions of clinical trials, and so on. In order to efficiently discover, search, and access the knowledge within all these text content, there is a continuous need for developing innovative techniques and algo- rithms for text representation, clustering, and visualization. Within the biomedical and clinical text files, one medical concept might be repre- sented in different forms or in abbreviations. For example, ‘Diabetes Mellitus Type 2’ could be represented as ‘DM2’ or ‘Type II Diabetes’ in different text files. This hap- pens often in the clinical notes within the Electronic Health Records (EHR), because clinicians have their own preferences of recording notes. On the other hand, some medical concepts might be highly correlated. For example, ‘Hypertension’ often co- occurs with ‘Stroke.’ Hence, the co-occurrences and semantic similarities between Abstract Biomedical text clustering is a text mining technique used to provide better docu- ment search, browsing, and retrieval in biomedical and clinical text collections. In this research, the document representation based on the concept embedding along with the proposed weighting scheme is explored. The concept embedding is learned through the neural networks to capture the associations between the concepts. The proposed weighting scheme makes use of the concept associations to build docu- ment vectors for clustering. We evaluate two types of concept embedding and new weighting scheme for text clustering and visualization on two different biomedical text collections. The returned results demonstrate that the concept embedding along with the new weighting scheme performs better than the baseline tf–idf for clustering and visualization. Based on the internal clustering evaluation metric-Davies–Bouldin index and the visualization, the concept embedding generated from aggregated word embedding can form well-separated clusters, whereas the intact concept embedding can better identify more clusters of specific diseases and gain better F-measure. Keywords: Neural networks, Concept embedding, Biomedical text clustering and Visualization Open Access © The Author(s) 2018. This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creat iveco mmons .org/licen ses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. R E S E A R C H Luo and Shah Appl Inform (2018) 5:8 https://doi.org/10.1186/s40535‑018‑0055‑8 *Correspondence: luo25@iupui.edu 1 Department of Computer Information Technology, IUPUI, Indianapolis, USA Full list of author information is available at the end of the article http://orcid.org/0000-0002-3649-9785 http://creativecommons.org/licenses/by/4.0/ http://crossmark.crossref.org/dialog/?doi=10.1186/s40535-018-0055-8&domain=pdf Page 2 of 19Luo and Shah Appl Inform (2018) 5:8 words and phrases are important for capturing the similarity of content within the context of biomedical document clustering. In order to capture the semantic simi- larities between words and phrases, previous research on text representation uses existing ontology such as MeSH or WordNet to identify the semantic relation- ships (Logeswari and Premalatha 2013; Yoo et  al. 2006; Zhang et  al. 2007). How- ever, the ontologies present the entity and attribute relationships in a hierarchy without emphasizing the co-occurrences of the concepts. On the other hand, the ontologies often do not include the different representations of the same concept. In recent years, the distributed representation which is also called word embedding gained interests in the research areas of text mining, natural language processing, and health informatics (Mikolov et al. 2013; Moen and Ananiadou 2013; Tulkens and Daelemans 2016). The word embedding emphasizes the co-occurrences of the words based on a given text collection. There are different ways to generate the distributed vector representations, which include probabilistic models (Globerson et  al. 2007) and dimensionality reduction on the word co-occurrence matrix (Levy and Goldberg 2014). The neural network is the new and most recent technique to generate the word embedding. It has been recently studied for biomedical text classification and clustering (Tulkens and Daelemans 2016; Kim and Cho 2017; Zhu et al. 2017), where word is the basic unit for the text documents and the word embedding is learned through neural networks. However, in the biomedical domain, clinical or medical concepts often contain more than one word. Hence, it is necessary to investigate the concept embedding that can be learned through the neural networks. We hypoth- esize that if the text documents are well represented by making use of concept embedding, it can improve biomedical text clustering and visualization. In this research, we propose and evaluate a framework for biomedical text cluster- ing and visualization based on the concept embedding of diseases. This proposed framework contains four major components: (1) extracting the phrases of diseases; (2) generating the concept embedding through neural networks; (3) constructing text document representations based on a new weighting scheme; and (4) text doc- ument clustering and visualization. Our major contributions include evaluation of concept embedding learned through the neural networks and a new weight scheme based on the concept embedding for biomedical text document representation, clus- tering, and visualization. The k-means (Hartigan and Wong 1979) algorithm is applied to cluster the newly represented text documents. To visualize the distribution of the clusters on the two- dimensional space, t Distributed Stochastic Neighbor Embedding (t-SNE) (Maaten and Hinton 2008) is employed after applying the Principal Component Analysis (Pearson 2008) to reduce the original dimensions of the document vectors. To evalu- ate the quality of the clusters, the internal evaluation metric-Davies–Bouldin index (DBindex) (Davies and Bouldin 1979) and external evaluation metric-F-measure (Van  Rijsbergen 1979) are used. Two biomedical text collections, Trecgen and Pub- Med Open Access, have been used for evaluation. The results demonstrate that the weighting scheme based on the generated concept embedding works much better than baseline tf–idf for the task of biomedical text clustering and visualization. Page 3 of 19Luo and Shah Appl Inform (2018) 5:8 Overview of the system framework Usually, a biomedical text document contains concepts of diseases, medications, treatments, symptoms, and so on. In this project, we focus on evaluating the concept embedding for diseases, given that biomedical literature searches often relate to certain diseases. In this research, the concept extraction component identifies and extracts the con- cepts of diseases through using the Unified Medical Language System (UMLS) Meta- Map (Shah and Luo 2018). Then, the concept embedding is generated by feeding the extracted concepts through the neural networks. We observe that the larger the text collections, the semantic and syntactic associations of the different concepts can be more accurately captured by the concept embedding. So, the input to the neural net- works contains more than one biomedical text collection. After generating the concept embedding, a new weighting scheme is proposed to construct the vector representation of the documents. The new weighting scheme is based on the association scores that are calculated by measuring the distances between the concepts using the embedding. The k-means algorithm is used to cluster the vectors of the documents. Finally, the distribu- tions of the clusters are visualized through the t-SNE technique. Figure 1 demonstrates the overall process of the proposed framework. The followed sub-sections detail the methodologies and techniques involved. Concepts extraction UMLS MetaMap (Shah and Luo 2018) is a natural language processing tool that uses various sources like UMLS Metathesaurus (Fact Sheet and SNOMED CT) to map the phrases or terms in the text to different semantic types. Figure  2 provides an example of MetaMap’s mapping of text into semantic types. In this research, if a term or phrase Biomedical Text Documents Visualiza�on - PCA + t-SNE Concept Extrac�on Concept Embedding Genera�on Clustering – k-means Document Representa�on Fig. 1 Overview of the proposed framework The USPSTF recommends screening for cervical cancer in women age 21 to 65 years with cytology (Pap smear) every 3 years. Neoplas�c Process Popula�on Group Temporal Concept Temporal Concept Laboratory Test Diagnos�c Procedure Temporal Concept Idea or Concept Fig. 2 Semantic mapping of using UMLS MetaMap Page 4 of 19Luo and Shah Appl Inform (2018) 5:8 has been mapped to semantic types ‘Disease or Syndrome’ or ‘Neoplastic Process,’ the corresponding phrase in the lexicon produced by MetaMap is extracted as a concept of disease. Generation of concept embedding Generating word embedding through using neural networks (Mikolov et  al. 2013) has drawn attention in the areas of natural language processing and machine learning (Tulk- ens and Daelemans 2016) (Kim and Cho 2017). There are two models of the neural net- works for generating the word embedding (Mikolov et  al. 2013). One is Continuous Bag of words (CBOW) model, the other is Skip-Gram model. In this research, the Skip- Gram model is employed. The neural network architecture of the Skip-Gram is provided in Fig. 3. It is a standard three-layer neural network. For word embedding, the input to the neural network are words that are represented as vectors. The length of the vector is the number of words in a text document collection. The element in the vector corresponding to the input word is set to 1, the rest elements are set to 0. The number of neurons at the hidden layer can be defined and tuned depending on the application. The number of output vectors to be evaluated relies on the defined num- ber of context words of an input word. The context words are the words that occur within a specific sliding window of the input word in a sentence. For example, if the window size is 1, there will be two context words in total. One context word occurs before the input word, and the other occurs after the input word in a given sentence. The context words are also represented in the same vector format as the input words. The output vectors will be Fig. 3 Skip-Gram model Page 5 of 19Luo and Shah Appl Inform (2018) 5:8 evaluated against the context words after applying a softmax layer which is used to sum the probabilities obtained in the output vectors to 1. After applying the softmax layer to the output vectors, the differences between the vectors of the context words and the out- put vectors will be calculated as a summed error vector and propagate back to update the weights of the neural network. After the training process, the weights between the input and the hidden layer are taken as the word embedding. The word embedding preserves the distances between words so that the words that have semantic and syntactic associations in the raw text collection are close to one another. In this research, we use this Skip-Gram model to generate the concept embedding. Two approaches have been explored: (1) generate the concept embedding by aggregating the word embedding, (2) generate the intact concept embedding, which means the input and output of the Skip-Gram model are vectors for the extract concepts. Aggregated word embedding If a concept consists of more than one word, the concept is represented as an element- wise sum of the distributed vectors presenting all words within the concept. For example, the concept ‘lung cancer’ is calculated as the element-wise sum of vectors of ‘lung’ and ‘can- cer.’ Given a concept that consists of n words, each word is represented as vector WordV, and the vector for the concept is shown as Eq. (1). Intact concept embedding The intact concept embedding are generated by considering each concept as a single entity. The structure of the neural network is the same, the concepts are the input to the neural network and the output vectors are corresponding context concepts. Thus, the con- cept embedding ConceptV can be directly generated. Concept association measure Since each concept is represented by a vector ConceptV, the association scores between the concepts can be calculated through a distance measure. If the association scores are stored in a matrix S, given a total of M concepts extracted from the raw biomedical text collection, each entry si,j in the matrix S represents the association score between concept ConceptVi and ConceptVj . In this research, the Cosine distance (Eq. 3) is used to calculate the associa- tion scores. (1)ConceptV = n ∑ k=1 WordVk. (2)SL,L =     s1,1 s1,2 · · · s1,L s2,1 s2,2 · · · s2,L ... ... ... ... sL,1 sL,2 · · · sL,L     (3)Si,j = ConceptVi˙ConceptVj ||ConceptVi||2||ConceptVj||2 Page 6 of 19Luo and Shah Appl Inform (2018) 5:8 Based on our previous research (Shah and Luo 2018), both concept embedding gen- eration approaches can capture the associations between concepts. Table  1 provides examples of disease concepts and the their top 3 closest concepts by using Trecgen data set (described in “Data set” section) as the training text collection. The association scores based on the intact concept embedding approach are lower than those based on the aggregated word embedding approach. The reason could be that the intact concept embedding measures the semantic similarities of the concepts by treating them as whole units, whereas, the aggregated word embedding depends on the semantic similarities between individual words within the concepts. For example, we found that ‘breast can- cer’ is closer to ‘lung cancer’ when aggregated word embedding is used, because the vec- tors are affected by occurrence of the shared word ‘cancer.’ Proposed document representation and weighting scheme In order to properly make use of the concept embedding for text clustering and visu- alization, in this research, a new weighting scheme is proposed ( W(Ci, d) ) to calculate the weights for the units in the vector that represents a text document. This proposed weighting scheme (shown as Eq. 4) alters the traditional tf–idf weighting scheme by con- sidering the similarities between the concepts within the documents. df (Ci) : document frequency of concept Ci ; tf (Ci, d) : frequency of concept Ci in docu- ment d; |D|: total number of documents in the text collection; Si,j : the association score between concepts Ci and Cj ; θ : the threshold to select the most associated concepts based on the association score; M: the number of closest concepts Ci within the same document; and N: the number of closest concepts Ci in the text collection. The weighting scheme uses the tf–idf value to underline the occurrence of the con- cept in the local content. The ∑M j=1 Si,j calculates the sum of association scores between the concept and the highly associated concepts that also occur within the same docu- ment. For example, if ‘Essential Hypertension’ and ‘HTN’ both occur in the document, and their association score is higher than the threshold θ , the association score is added to original tf–idf value to emphasize the co-occurrences of the concepts. By using the traditional tf–idf, if the tf value is 0, the corresponding value will be 0 for the vector representation; whereas using the proposed weighting scheme, the weight is calculated by a weighted sum of highly associated concepts that appear in the document. For exam- ple, a document does not contain the concept ‘diabetes mellitus,’ but contains ‘diabetes.’ Instead of using 0 for ‘diabetes mellitus,’ the similarity score between ‘diabetes mellitus’ and ‘diabetes’ is used. This proposed weighting scheme has been evaluated for text clus- tering and visualization in this research. Clustering and visualization methodologies The k means clustering algorithm (Hartigan and Wong 1979) has been suc- cessfully used in various application domains, such as text mining, computer vision, and so on (Logeswari and Premalatha 2013; Zhang et  al. 2007). The aim (4)W(Ci, d) = { tf (Ci, d) × log |D| df (Ci) + ∑M j=1 Si,j tf (Ci, d) > 0, Si,j > θ ∑N j=1 N−(j−1) N Si,j tf (Ci, d) = 0, Si,j > θ Page 7 of 19Luo and Shah Appl Inform (2018) 5:8 of k-means algorithm is to minimize the sum of distances of each point within the cluster to the cluster center. Given x = x1, x2, . . . , xN as a set of training data and Clusters = Cluster1, Cluster2, . . . , Clusterk as a set of initialized k centers ( µ1, µ2, . . . , µk ), the algorithm can be summarized as follows: (1) Assignment of cluster centers: assign each data point xi to the cluster Clusterj whose Euclidean distance from the cluster center is minimum of all the cluster centers. (5)Clusterj = {xn : ||xn − Clusterj|| ≤ ||xn − Clusteri||, ∀i, i ∈ [1, k]}. Table 1 Examples of top 3 most associated concepts and the association scores Concept Association score Alzheimer disease Intact concept Alzheimer 0.829 Parkinson disease 0.813 Alzheimer’s 0.757 Aggregated word Presenile alzheimer disease 0.913 Parkinson disease 0.903 Huntington disease 0.895 Multiple sclerosis Intact concept Multiple sclerosis relapsing remitting 0.661 Ms 0.633 Parkinson 0.600 Aggregated word Opticospinal multiple sclerosis 0.957 Progressive multiple sclerosis 0.939 Multiple sclerosis primary progressive 0.902 Cerebral amyloid angiopathy Intact concept Caa 0.601 Cerebral 0.486 Amyloid angiopathy 0.468 Aggregated word Senile cerebral amyloid angiopathy 0.969 Cerebral amyloid angiopathy genetic 0.965 Sporadic cerebral amyloid angiopathy 0.960 Colon cancer Intact concept Colorectal cancer 0.799 Cancer of colon 0.755 Cancer of the colon 0.725 Aggregated word Cancer of colon 0.980 Colon cancers 0.944 Metastatic colon cancer 0.943 Page 8 of 19Luo and Shah Appl Inform (2018) 5:8 (2) Update cluster centers: set the new center of each cluster to the mean of all data points belonging to that cluster. (3) Repeat the previous two steps until convergence. To visualize the cluster distributions on a two-dimensional space, the t Distributed Stochastic Neighbor Embedding (t-SNE) which was implemented by Maaten and Hinton (2008), is employed to project the high-dimensional data into two-dimen- sional space. The t-SNE algorithm minimizes the sum of the KL divergences of all data points in the original dimensional space and the mapping space. The cost func- tion is given as Eq. (7). pj|i : the conditional probability of the similarity of data point xj to data point xi when xj is its neighbor in proportion to their probability density under a Gaussian centered at xi and qj|i : the conditional probability of the low-dimensional counterparts of xi and xj. When pj|i and qj|i are equal, the value of the cost function is minimum. So, t-SNE algorithm aims to find a low-dimensional data representation that minimizes the mis- match between pj|i and qj|i. The computational cost of t-SNE is high when the original dimensionality of the data is high. To speed up the process, Principle Component Analysis (PCA) can be used to reduce the dimensionality to a lower space before t-SNE technique is applied. PCA suppresses some noise without severely distorting the distances between data points (Maaten and Hinton 2008). Data sets To evaluate the proposed biomedical text clustering and visualization framework, two biomedical text collections have been used. One collection is a labeled, the other is unlabeled. The details of the document collections are given as follows. Trecgen The Trecgen collection contains 4478 abstracts that are extracted from MEDLINE as a part of the TREC 2005 Genomics Track (Hersh et al. 2005). The abstracts are categorized into seven categories which are detailed in Table  2. After concept extraction by using MetaMap, it is discovered that about 50% of the extracted concepts of diseases contain two words, and 4.83% of the concepts have more than 5 words. Many of the concepts have low document frequency. Over 57% of the concepts have document frequency 1, and 7% of the concepts have document frequency over 10. This means very few concepts occur in more than 10 documents. And it also implies that the same diseases may be (6)µi = 1 |Clusteri| ∑ xn∈Clusteri xn, ∀i. (7) ∑ i ∑ j pj|ilog pj|i qj|i Page 9 of 19Luo and Shah Appl Inform (2018) 5:8 represented differently in different documents. Figure 4 shows the document frequency of the extracted concepts in this text collection. PubMed OA The PubMed Central-Open Access (PubMed OA) data set has been widely used in many research projects for biomedical text clustering and classification (Zhang et al. 2007; Zhu et  al. 2009). Different from Trecgen, PubMed OA is an unlabeled (un-categorized) col- lection which contains over 1 million articles from the database—PubMed Central. In this research, 600 abstracts were randomly selected from journals whose names begin with letter ‘A’ or ‘B.’ The number of selected articles from each journal is shown in Table 3. After concept extraction by using MetaMap, it is discovered that around 20% of the concepts are single word concepts, 50% of them contain two words, and around 9% of them contain four or more words. Figure  4 shows that many of the concepts have low document frequency. Over 72% of the concepts have document frequency 1, and less than 5% of the concepts have document frequency over 6. Table 2 Summary of Trecgen Category No. of documents Mad cow disease 447 Multiple sclerosis 554 Colon cancer 567 Alzheimer’s disease 1201 Cerebral amyloid angiopathy 482 Parkinson’s disease 769 Breast cancer 458 Total 4478 Fig. 4 Document frequency of the concepts Page 10 of 19Luo and Shah Appl Inform (2018) 5:8 Clustering evaluation metrics To identify the optimum number of clusters, the clustering results are evaluated against clustering evaluation metrics. Typically, there are two types of evaluation metrics: internal evaluation and external evaluation. The internal evaluation is to for- malize the goal of attaining high intra-cluster similarity and low inter-cluster similar- ity. Davies–Bouldin index (Davies and Bouldin 1979) is one of the internal evaluation metrics. The external evaluation metrics are based on the interest of an application, such as categorization. F-measure is one of the external evaluation metrics. The Davies–Bouldin index and F-measure are briefly described as follows: • Davies–Bouldin index: The calculation of the Davies–Bouldin index is shown in Eq. (8), where σClusteri is the standard deviation of the distance of samples in a cluster Clusteri to the respective cluster centroid; d(Clusteri, Clusterj) is the Euclid- ean distance between centroids of cluster Clusteri and Clusterj ; and |Clusters| is the total number of clusters. The more distinct the clusters are from each other, the lower the DB index value is • F-measure: If each cluster is labeled by the category of the majority of the data points within the cluster, the F-measure is computed as Eq. (9), in which tp, fn, and fp stand for true positive, false negative, and false positive. (8)DBindex = 1 |Clusters| |Clusters| ∑ i=1 max j,j �=i σClusteri + σClusterj d(Clusteri, Clusterj) . (9)F-measure = tp 2 × tp + fn + fp . Table 3 Summary of PubMed OA Name of journal No. of documents Augmentative and Alternative Communication 2 American Journal of Physiology Endocrinology and Metabolism 11 Biological Trace Element Research 31 Ancient Science of Life 3 Allergy and Asthma Proceedings 28 Brain and Language 1 Bone Marrow Research 1 BoneKEy Reports 4 Annals of Rehabilitation Medicine 323 Aphasiology 3 Anesthesia, Essays and Researches 135 Bioinformatics and Biology Insights 45 American Journal of Hypertension 13 Total 600 Page 11 of 19Luo and Shah Appl Inform (2018) 5:8 Clustering performance analysis and visualization The threshold θ in Eq. (4) determines the selected associated concepts to calculate the weights for document representations. In this research, we fully evaluated the θ from 0.75 to 0.95 for both aggregated word embedding and intact concept embedding. Dif- ferent k values (from 2 to 30) for the k means algorithm have also been explored for both data sets. The visualizations are performed based on selected k values and θ val- ues. The detailed clustering performance and visualization on each data set are pro- vided in the following sub-sections. Trecgen Figures  5, 6, and 7 show the returned values of the DBindex for baseline tf–idf, pro- posed weighting scheme based on aggregated word embedding and intact concept embedding for different k values. The returned DBindex values show that the pro- posed weighting scheme based on aggregated word embedding works better than the other two. Figures  8, 9, and 10 show the returned values of the F-measures for different k val- ues of the three weighting schemes. Since the F-measure evaluates the clustering result Fig. 5 Trecgen—DBindex—baseline TF–IDF Fig. 6 Trecgen—DBindex—aggregated word embedding Page 12 of 19Luo and Shah Appl Inform (2018) 5:8 according to the true categories of the documents, the higher the F-measure is, the bet- ter clustering performance. Different from the returned DBindex values, the returned F-measures show that the proposed weighting scheme based on intact concept embed- ding performs better than the other two. The F-measures of all three show that the values start to have minor changes when k is larger than 7 which is the number of cat- egories of this data collection. Fig. 7 Trecgen—DBindex—intact concept embedding Fig. 8 Trecgen—F-measure—baseline TF–IDF Fig. 9 Trecgen—F-measure—aggregated word embedding Page 13 of 19Luo and Shah Appl Inform (2018) 5:8 To visualize the distribution of clusters, PCA is used to reduce the original dimension to 900, then t-SNE algorithm is applied to visualize the clusters on the two-dimensional space. The clusters are labeled by the concept(s) that have top document frequency within the clusters. Figures  11, 12, and 13 show the visualization of the three weight- ing schemes when k = 7 . The baseline tf–idf shows 7 clusters that match six categories Fig. 10 Trecgen—F-measure—intact concept embedding Fig. 11 Trecgen—baseline TF–IDF—k = 7 Fig. 12 Trecgen—aggregated word embedding—k = 7 , θ = 0.85 Page 14 of 19Luo and Shah Appl Inform (2018) 5:8 in the original document collection. The cluster ‘Miscellaneous disease’ contains many different diseases and has overlapping with all other clusters. The category ‘Cerebral Amyloid Angiopathy’ is not notable in the visualization. After investigation, it is found that most of the documents in the ‘Cerebral Amyloid Angiopathy’ category are in either ‘Alzheimer disease’ or ‘Miscellaneous disease’ cluster. Figure  12 shows 7 clusters using the proposed weighting scheme based on the aggregated word embedding when θ is set to 0.85. Comparing to the visualization of baseline tf–idf, it is apparent that the clus- ters based on the aggregated word embedding are well separated with fewer overlaps. Documents in category ‘Colon Cancer’ are included in two clusters that are next to each other. The documents within category ‘Cerebral Amyloid Angiopathy’ are distributed across the clusters of ‘Alzheimer disease’ and ‘Parkinson disease.’ It is also noticed that the cluster on the right (light green) mixed ‘Multiple Sclerosis’ and ‘Bovine Spongiform Encephalopathy’ which is a kind of mad cow disease. This explains that the F-measure based on the aggregated word embedding is low then k is 7. Figure 13 shows 7 clusters using the proposed weighting scheme based on the intact concept embedding when θ is set to 0.90. The separation of the clusters is not as obvious as that based on the aggre- gated word embedding. This explains that the DBindex is higher than that based on the aggregated word. PubMed OA Since PubMed OA is a collection without defined categories for the documents, only internal evaluation DBindex is used to evaluate the clustering performance. Figures 14, 15, and 16 show the DBindex values for baseline tf–idf, proposed weighting scheme based on aggregated word embedding, and intact concept embedding for different k val- ues. The DBindex values show that the proposed weighting schemes based on aggregated word embedding and intact concept embedding work much better than the baseline tf– idf. The DBindex values based on the aggregated word embedding are slightly lower than those based on the intact concept embedding. Although the DBindex decreases along with the increasing of the k, it does not decrease significantly decrease when k is larger than 14. To visualize the distribution of clusters, PCA is used to reduce the original dimension to 300, then t-SNE algorithm is applied to visualize the clusters on the two-dimensional Fig. 13 Trecgen—intact concept embedding—k = 7 , θ = 0.90 Page 15 of 19Luo and Shah Appl Inform (2018) 5:8 space. The clustering visualization of k = 14 based on the baseline tf–idf is given in Fig.  17. Apparently, the baseline tf–idf does not work well on generating document clusters. The largest cluster (black), and some small clusters (light blue, blue, red, and brown) all contain documents of many diseases that include ‘stroke,’ different types of cancers, and other cardiac and neurological diseases. Hence, they are all labeled as Fig. 14 PubMed OA—DBindex—baseline TF–IDF Fig. 15 PubMed OA—DBindex—aggregated word embedding Fig. 16 PubMed OA—DBindex—intact concept embedding Page 16 of 19Luo and Shah Appl Inform (2018) 5:8 ‘Miscellaneous Diseases.’ There is no cluster that has high document frequency of a spe- cific disease concept. Figures  18 and 19 show the clustering visualization of selected θ based on the aggre- gated word embedding and intact concept embedding when DBindex are low and k = 14 . Comparing to the baseline tf–idf, the clusters shown in these figures are well formed and separated from each other. It is found that the densities of some clusters given in Fig. 18 is slightly higher than those in Fig. 19. This explains that when k = 14 , the DBindex value based on the aggregated word embedding is lower than that based on the intact concept embedding. We then label the clusters by the concept(s) that have top frequent document frequency. It is found that the intact concept embedding can identify more clusters of specific types of diseases, such as ‘Obesity,’ ‘Chronic Low Back Pain,’ and ‘Hypertension.’ We also notice that some clusters contain some concepts that are highly related. For example, the ‘Breast Cancer’ cluster contains many instances of ‘Lym- phoedema.’ The ‘Hypertension’ cluster contains documents describing hypertension, hypoglycemia, and heart blockage. By analyzing the clusters based on the aggregated word embedding, we found that some cluster, such as ‘Tumor,’ contains documents that are under a category. We hypothesize that the reason is that many of the concepts have Fig. 17 PubMed OA—baseline TF–IDF—k = 14 Fig. 18 PubMed OA—aggregated word embedding—k = 14 , θ = 0.85 Page 17 of 19Luo and Shah Appl Inform (2018) 5:8 some common word(s). Actually, it is found that ‘brain tumor’ and ‘tumor in breast’ are all in cluster ‘Tumor.’ Discussion The clustering evaluation and visualization show that both aggregated word embedding and intact concept embedding-based weighting schemes perform better than baseline tf–idf weighting scheme. The concept embedding captures the associations between the concepts, and by making use of the concept associations, documents can be well rep- resented for the task of biomedical document clustering. The returned DBindex values based on aggregated word embedding approach are slightly better than those based on the intact concept embedding approach. However, since DBindex evaluates the cluster- ing performance based on high intra-cluster similarity and low inter-cluster similarity without considering the content within the clusters, low DBindex value might not reflect high content similarity within the clusters; whereas the external evaluation F-meas- ure considers the content wise intra-cluster similarity. This explains that the returned F-measures on the Trecgen show that the intact concept embedding works better than the other two. Based on the visualizations, the intact concept embedding can better identify more clusters of specific diseases; whereas, the aggregated word embedding can identify the clusters of higher level categories of diseases, such as cancer or tumor. So, it is worth to investigate the integration of these two concept embedding for clustering and visualization in the future. Conclusion and future work In this paper, a framework for biomedical document clustering and visualization based on concept embedding of diseases is proposed and evaluated. The concept embedding is learned through neural networks. The first concept embedding is based on aggregating word embedding, and the second is intact concept embedding. Both concept embedding generation approaches can capture the association of the concepts based on the content of the training text collection. A new weighting scheme is proposed to use the concept embedding and compared against baseline tf–idf for text clustering and visualization. Fig. 19 PubMed OA—intact concept embedding—k = 14 , θ = 0.85 Page 18 of 19Luo and Shah Appl Inform (2018) 5:8 The results demonstrate that the proposed weighting scheme using the concept embed- ding achieve better performance than the baseline tf–idf. Potential future work includes extending this framework to biomedical document clustering by including concept embedding of other types of medical concepts, such as symptoms, treatments, and so on. On the other hand, other clustering methods and hierarchical clustering architecture can be explored for the clustering and the visualiza- tion of larger text collection. Authors’ contributions The authors discussed the problem and the solutions were proposed all together. All authors participated in drafting and revising the manuscript. Both authors read and approved the final manuscript. Author details 1 Department of Computer Information Technology, IUPUI, Indianapolis, USA. 2 Department of Electrical and Computer Engineering, IUPUI, Indianapolis, USA. Acknowledgements This project is supported by IUPUI Enhanced Mentoring Program with Opportunities for Ways to Excel in Research (EMPOWER). Competing interests The authors declare that they have no competing interests. Availability of data and materials Data sharing is not applicable to this article as no data sets were generated or analyzed during the current study. Consent for publication All authors consent to publish this work. Ethics approval and consent to participate Not applicable. Funding There was no external funding for this study. Internal funding support was written in acknowledgement. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Received: 18 August 2018 Accepted: 20 October 2018 References Davies DL, Bouldin DW (1979) A cluster separation measure. IEEE Trans Patt Anal Mach Intell 2:224–227 Fact Sheet—UMLS Metathesaurus. https ://www.nlm.nih.gov/pubs/facts heets /umlsm eta.html Globerson A, Chechik G, Pereira F, Tishby N (2007) Euclidean embedding of co-occurrence data. J Mach Learn Res 8:2265–2295 Gorg C, Tipney H, Verspoor K, Baumgartner WK, Cohen KB, Stasko J, Hunter LE (2010) Visualization and language processing for supporting analysis across the biomedical literature. In: International conference on knowledge- based and intelligent information and engineering systems proceedings, pp 420–429 Gu J, Feng W, Zeng J, Mamitsuka H, Zhu S (2013) Efficient semisupervised medline document clustering with mesh- semantic and global-content constraints. IEEE Trans Cybern 43(4):1265–1276 Hartigan JA, Wong MA (1979) Algorithm as 136: a k-means clustering algorithm. J Royal Stat Soc. 28(1):100–108 Hersh, W, Cohen, A, Yang, J, Bhupatiraju RT, Roberts P, Hearst M (2005) Trec 2005 genomics track overview. NIST Spe- cial Publication 500-266: The Fourteenth Text REtrieval conference proceedings Kim HK, Cho S (2017) Bag-of-concepts : comprehending document representation through clustering words in distributed representation. Neurocomputing Levy O, Goldberg Y (2014) Neural word embedding as implicit matrix factorization. In: Advances in neural informa- tion processing systems, pp 2177–2185 Logeswari S, Premalatha K (2013) Biomedical document clustering using ontology based concept weight. In: Interna- tional conference on computer communication and informatics proceedings, pp 1–4, https ://doi.org/10.1109/ ICCCI .2013.64662 73 Lvd Maaten, Hinton G (2008) Visualizing data using t-sne. J Mach Learn Res. 9:2579–2605 MEDLINE/PubMed Resource Guide. https ://www.nlm.nih.gov/bsd/pmres ource s.html MetaMap—a tool for recognizing UMLS concepts in text. https ://metam ap.nlm.nih.gov/ Mikolov T, Sutskever I, Chen K, Corrado G, Dean J (2013) Distributed representations of words and phrases and their compositionality. In: International conference on neural information processing systems, pp 3111–3119 https://www.nlm.nih.gov/pubs/factsheets/umlsmeta.html https://doi.org/10.1109/ICCCI.2013.6466273 https://doi.org/10.1109/ICCCI.2013.6466273 https://www.nlm.nih.gov/bsd/pmresources.html https://metamap.nlm.nih.gov/ Page 19 of 19Luo and Shah Appl Inform (2018) 5:8 Moen S, Ananiadou TSS (2013) Distributional semantics resources for biomedical text processing. In: Proceedings of the 5th international symposium on languages in biology and medicine, Tokyo, Japan, pp 39–43 Pearson K (2008) Liii on lines and planes of closest fit to systems of points in space. London Edinburgh Dublin Philos Magaz J Sci 2(11):559–572. https ://doi.org/10.1080/14786 44010 94627 20 PubMed Open Access Subset. https ://www.ncbi.nlm.nih.gov/pmc/tools /openf tlist / SNOMED CT. https ://www.nlm.nih.gov/healt hit/snome dct/ Shah S, Luo X (2018) Comparison of deep learning based concept representations for biomedical document cluster- ing. In: IEEE EMBS international conference on biomedical & health informatics (BHI), pp 349–352. IEEE, New York Tulkens S, Daelemans W (2016) Using distributed representations to disambiguate biomedical and clinical concepts. arXiv preprint arXiv :1608.05605 Van Rijsbergen C (1979) Information retrieval. dept. of computer science, university of glasgow https ://cites eer.ist. psu.edu/vanri jsber gen79 infor matio n.html Yoo I, Hu X, Song I-Y (2006) A coherent graph-based semantic clustering and summarization approach for biomedi- cal literature and a new summarization evaluation method. In: First international workshop on text mining in bioinformatics proceedings, pp 84–89 Zhang X, Jing L, Hu X, Ng M, Zhou X (2007) A comparative study of ontology based term similarity measure on pubmed document clustering. In: International conference on database systems for advanced applications proceedings, pp 115–126 Zhu Y, Yan E, Wang F (2017) Semantic relatedness and similarity of biomedical terms: examining the effects of recency, size, and section of biomedical publications on the performance of word2vec. BMC Med Inform Decis Making 17:95–103 Zhu S, Zeng J, Mamitsuka H (2009) Enhancing medline document clustering by incorporating mesh semantic similarity. Bioinformatics 25(15):1944–1951 https://doi.org/10.1080/14786440109462720 https://www.ncbi.nlm.nih.gov/pmc/tools/openftlist/ https://www.nlm.nih.gov/healthit/snomedct/ http://arxiv.org/abs/1608.05605 https://citeseer.ist.psu.edu/vanrijsbergen79information.html https://citeseer.ist.psu.edu/vanrijsbergen79information.html Concept embedding-based weighting scheme for biomedical text clustering and visualization Abstract Introduction Overview of the system framework Concepts extraction Generation of concept embedding Concept association measure Proposed document representation and weighting scheme Clustering and visualization methodologies Data sets Trecgen PubMed OA Clustering evaluation metrics Clustering performance analysis and visualization Trecgen PubMed OA Discussion Conclusion and future work Authors’ contributions References 
work_zdew7msp5nhphkzbylgxdqbis4	----	Analysing of existing customers to improve acquiring of new customers Technology Classification with Latent Semantic Indexing Dirk Thorleuchter a,* , Dirk Van den Poel b a Fraunhofer INT, D-53879 Euskirchen, Appelsgarten 2, Germany, dirk.thorleuchter@int.fraunhofer.de b Ghent University, Faculty of Economics and Business Administration, B-9000 Gent, Tweekerkenstraat 2, Belgium, dirk.vandenpoel@ugent.be URL: http://www.crm.UGent.be _________________ * Corresponding author at: Fraunhofer INT, Appelsgarten 2, 53879 Euskirchen, Germany. Tel.: +49 2251 18305; fax: +49 2251 18 38 305 E-mail address: Dirk.Thorleuchter@int.fraunhofer.de (D. Thorleuchter). Abstract Many national and international governments establish organizations for applied science research funding. For this, several organizations have defined procedures for identifying relevant projects that based on prioritized technologies. Even for applied science research projects, which combine several technologies it is difficult to identify all corresponding technologies of all research-funding organizations. In this paper, we present an approach to support researchers and to support research-funding planners by classifying applied science research projects according to corresponding technologies of research-funding organizations. In contrast to related work, this problem is solved by considering results from literature concerning the application based technological relationships and by creating a new approach that is based on latent semantic indexing (LSI) as semantic text classification algorithm. Technologies that occur together in the process of creating an application are grouped in classes, semantic textual patterns are identified as representative for each class, and projects are assigned to one of these classes. This enables the assignment of each project to all technologies semantically grouped by use of LSI. This approach is evaluated using the example of defense and security based technological research. This is because the growing importance of this application field leads to an increasing number of research projects and to the appearance of many new technologies. Key Words: Latent semantic indexing, SVD, Classification, Research Funding. mailto:dirk.vandenpoel@ugent.be Using LSI for Technology Classification _______________________________________________________________________ 2 1 Introduction Research funding for applied science research projects is done by many national and international organizations (Beaudry & Allaoui, 2012; Lepori, 2011). They evaluate proposals for new research projects and based on self-defined procedures, they identify the relevant projects, which are accepted for funding (Hicks, 2012; Mobjörk & Linnér, 2006). An important criterion for technological research is that the technologies standing behind the proposed research project are also mentioned in a specific list or taxonomy of prioritizes technologies (Choi, Lee, & Sohn, 2009; Bradshaw et al., 2008). In general, these technology lists or taxonomies consist of a manually created label for each technology and of a description. The descriptions contain terms from the technology as well as from potential application fields (Thorleuchter, Van den Poel, & Prinzie, 2010c). For example, the European Union establishes a Framework Research Programme (FP7) theme for security that has the objective to develop technologies needed to ensure the security of citizens from threats. It uses a list of prioritized technologies (ESRAB technology list) for research funding decisions (Remuss, 2010). That means proposals of research projects that do not fit with these prioritized technologies and the corresponding application field e.g. ‘security’ normally are not accepted (McLeish & Nightingale, 2007; Jiricka & Pröbstl, 2012). For a researcher, it is often difficult to identify the corresponding prioritized technologies and corresponding application fields concerning each research-funding organization (Grimpe, 2012). Additionally, it is also difficult for research planners to assign applied science research projects to prioritized technologies of their research-funding organization manually (Ludwig, Roson, Zografos, & Kallis, 2011). Therefore, in this paper, we present an automated approach based on text classification that supports researchers as well as research-funding planners by the identification of relationships between applied science research projects and technologies extracted from lists or taxonomies. Literature proposes application based technological relationships (Yu, Hurley, Kliebenstein, & Orazem, 2012). Here, it is shown that during the process of creating an application, technologies are related to their substitutive, integrative, predecessor, and successor Using LSI for Technology Classification _______________________________________________________________________ 3 technologies (Geschka, 1983). An example for substitutive technologies is electrical fuel cells, electrical batteries, and solar cells in the context of creating an energy supply application. A research project that has the aim to create a new approach for an energy supply application can combine all three substitutive technologies to build this new approach. Alternatively, it can focus on one technology e.g. fuel cells. However then, it has to consider research results from the further substitutive technologies. This is because the newly created fuel cell approach for energy supply has to be compared to existing potential energy supply applications to indicate its advances. This full cell project processes knowledge from electrical battery and solar cells and thus, is related to the electrical battery technology and to the solar cell technology, too; even if key words from electrical battery technology or from solar cell technology do not occur in the project description (Geschka, Lenk, & Vietor, 2002). Applied science research projects have to combine or to consider these related technologies to create an application (Thorleuchter, Van den Poel, & Prinzie, 2010b). This describes a binary classification problem because the test examples (research projects) are associated with a specific class (a set of related technologies) (Kim, Toh, Teoh, Eng, & Yau, 2012). To identify related technologies, LSI is used. This is because semantically, all related technologies consist of the same terms describing the technology or the application field. LSI identifies the semantic textual patterns in the descriptions of the technologies and it also identifies the impact of each technology description on each semantic textual pattern (Thorleuchter & Van den Poel, 2012b). Then, each semantic pattern represents a set of related technologies where the corresponding impact is larger than a specific threshold. The descriptions of the projects are projected in the same semantic subspace. An assignment of each project on a set of technologies can be done based on the calculated impact of each project on each semantic textual pattern (Thorleuchter, Van den Poel, & Prinzie, 2012). Previous work calculates the similarity between each project and each technology separately assuming that all technologies are independent (Thorleuchter & Van den Poel, 2011a). It uses machine-learning techniques as supervised learning methods and a knowledge structure text classification approach that uses a similarity measure (Jaccard’s coefficient) as well as a specific threshold to enable a multi-label classification. This knowledge structure Using LSI for Technology Classification _______________________________________________________________________ 4 approach often fails because prevalent features that are characteristic for a technology are not simultaneously present in all projects that belong to one technology. In contrast to previous and related work, this work considers research results from the application based technological relationships as mentioned above. Aspects that are relevant for this task are extracted and used for this approach. Related technologies are grouped in several sets as represented by semantic textual patterns and each project has to be assigned to one set of related technologies. This can be done by using a binary textual classification instead of using a multi-label classification and this enables the use of LSI as a binary semantic classification algorithm. This approach is evaluated using the example of defense and security (D&S) based technological research projects. This is because the growing importance of this application field leads to an increasing number of research projects and the appearance of many new technologies as indicated by the occurrence of several technology lists or taxonomies (e.g. EDA, WEAG, STACCATO, ESRAB, MCTL, and DSTL) during the last years (Gericke et al., 2009; Te Kulve & Smit, 2003). The results are compared to a standard text classification algorithm that applies a multi-label classification on the same data set. A centroid vector is created that represents the term vectors from the training examples (projects) of each class (technology) (Takci & Güngör, 2012). This vector is the average vector of all vectors that are assigned to this class in the training phase. Term vectors from further research projects (test examples) are compared to all centroid vectors for identifying similar centroid vectors. We use a well-known similarity measure (Jaccard's coefficient) and a specific threshold to assign test examples to classes that means to identify none, one, or several technologies for each project (Madjarov, Kocev, Gjorgjevikj, & Džeroski, 2012). The evaluation shows, that the new LSI based approach outperforms the centroid based text classification algorithm concerning the calculated performance measures precision and recall. Using LSI for Technology Classification _______________________________________________________________________ 5 2 Background In this approach, we consider findings of literature that focus on the application based technological relationships. Some important aspects are adapted to this approach and mentioned in Sect. 2.1. Further, text classification approaches that are used in this study are described in Sect. 2.2 and it is explained; why LSI is a good mean to identify the technological relationships from Sect. 2.1. Further, a knowledge structure based classification approach is selected for evaluation purposes. It outperforms further knowledge structure approaches considering the aspects in Sect. 2.1. 2.1 Application based technological relationships A large number of literature studies the relationships between technologies (Choi et al., 2012; Subramanian & Soh, 2010; Radder, 2009; Jiménez, Garrido-Vega, Díez de los Ríos, & González, 2011; Herstatt & Geschka, 2002; Rubenstein et al., 1977; Fleck & Howells, 2001). Below, most important findings are adapted specifically for this study. a) An applied science research project can be classified according to a technology only if there is a relation between the project and the technology. The simplest relation is that a project contains research activities concerning the core area of a technology. Then both, the project description and the technology description consist of the same technology specific terms that describe the technological field. Therefore, the project can be directly assigned to one technology by computing the similarity of both descriptions. b) Technologies are not single data points but they describe a technological field that consists of many different research topics. Inside this field multiple research projects occur. Two research projects, which focus on different topics in a technological field, consist of project descriptions with different terms although they belong to the same technology. Therefore, prevalent features that are characteristic for a technology are not simultaneously present in all projects that belong to one technology. Using LSI for Technology Classification _______________________________________________________________________ 6 c) Technological project descriptions consist of a high percentage of term co-occurrence. This is because to describe a technical topic, several technical terms are used that normally occur together in a text phrase. Therefore, conditioned on each technology and on each project, different terms do not occur independently. d) Applied science research projects focus on an application field and use many different technologies. Literature indicates that these projects consist of up to ten technologies. Therefore, these research project descriptions consist of features from several different technologies. e) If a research project is assigned to a technology and this technology is related to further technologies then the project can be assigned to these further technologies, too. One kind of relationship is that technologies can be similar to other technologies. They deal with the same technology field but have a different focus e.g. passive radar technology and active radar technology. Technologies are not completely delimited from their similar technologies, which means in some research areas similar technologies overlap. Descriptions of similar technologies also consist of technology specific terms that describe the technological field. Then, a research project can be assigned to a similar technology by comparing the project description to the technology description. f) A further relationship is seen between a technology and its substitutive technology. These technologies substitute each other e.g. electrical fuel cells, electrical batteries and solar cells in the context of energy supply. An applied science research project normally examines several substitutive technologies to create an application. Then its description consists of terms from different technology fields. By comparing this description to a technology description, we do not get a large similarity because terms from the further technology fields do not appear in the technology description. If the research project examines fuel cell, electrical battery, and solar cell technology in an equally distributed way then the similarity by comparing the project description to the fuel cell technology description is about one third. Therefore, it is necessary to get project and technology descriptions that also contain terms, Using LSI for Technology Classification _______________________________________________________________________ 7 which describe the application field. Then, one gets a higher similarity by comparing and a better success by assigning a project to a substitutive technology. g) Integrative technologies sometimes are named complementary technologies and occur together by realizing an application. Examples for two integrative technologies are fuel and lubricants technology. This is because both technologies are used e.g. to create a new power plant prototype. Additionally, predecessor or successor technologies are technologies that precede or succeed another in the process of creating an application. Thus, it is important to use project and technology descriptions that contain terms, which describe the application field, too. 2.2 Text Classification In general, the aim of text classification is the assignment of pre-defined classes to text documents (Ko & Seo, 2009; Sudhamathy & Jothi Venkateswaran, 2012; Lin & Hong, 2011; Finzen, Kintz, & Kaufmann, 2012). For the identification of technologies standing behind projects, a class can be defined in two different ways. First, each technology can be represented by one class. Using this definition leads to the use of a multi-label classification (Thorleuchter, Weck, & Van den Poel, 2012a; Thorleuchter & Van den Poel, 2011c) because a project consists of several technologies and thus, it should be assigned to several classes. Second, a set of related technologies can be represented by one class. As shown in Sect. 2.1, the descriptions of related technologies consist of similar terms that describe application fields or technology areas. Based on these characteristic textual patterns, related technologies can be identified. Using this definition leads to the use of a binary classification where a project is assigned to one class or not. Extracting technologies from lists or taxonomies normally leads to a large number of technologies. E.g. in the case study (see Sect. 4) 2.850 technologies are extracted from the application field security and defense. Defining a class as a technologies leads to a large number of classes that probably causes performance problems in text classification. Using LSI for Technology Classification _______________________________________________________________________ 8 Semantic generalizations by grouping related technologies are a good mean to reduce the number of classes. The assignment of a project to a technology or to a set of related technologies depends on semantic aspects (aspects of meaning) and not on knowledge structure aspects (aspects of words) as described in Sect. 2.1. A single term (a word) that is characteristic for a technology does not have to be in the description of a project even if this project processes the technology but a semantic textual pattern of several terms probably will be. Thus for the text classification approach proposed in this paper, it is more important to compare the aspects of meaning between a project and technologies than to compare the aspects of words between them (Park, Kim, Choi, & Kim, 2012). The aspects of meaning can be identified by calculating the semantic textual patterns. 2.2.1 Knowledge structure approaches The most frequently used approaches in text classification are knowledge structure approaches. Examples for standard algorithms are k nearest neighbor (k-NN) classification as instance-based learning algorithm, C4.5 as decision tree model, naive Bayes (NB) as a simple probabilistic algorithm, and support vector machine (SVM) (Shi & Setchi, 2012; Lee & Wang, 2012). These approaches are not able to identify hidden semantic textual patterns. Despite this weakness, a knowledge structure approach is selected as baseline for the evaluation to show the success of the used semantic approach. The centroid-based approach is in contrast to some standard categorization algorithms in text classification where example classes are not described by one centroid vector, but by a number of training examples. We select this approach as baseline. Below, we give detailed explanations for using a centroid-based text classification. Our explanations are based on the results of (Han, 2000) where extensive evaluations of centroid-based classifications and comparisons with other classifiers are described. Using LSI for Technology Classification _______________________________________________________________________ 9 With a centroid-based scheme, the characteristics of each class can be summarized. By use of this summarization, several prevalent features are joined together. This is very important for our approach because terms that represent these technology-characteristic features are not simultaneously present in research project descriptions that belong to the technology as shown in Sec. 2.1. Therefore, comparing a term vector from a project (as test example) to a centroid vector leads to better performance than comparing it to term vectors from projects (training examples) that describe a class. We can find a similar summarization in the naive Bayes algorithm where for each class a distribution function is created that represents the term probabilities. Further algorithms (k-NN, C4.5, SVM etc.) describe a class by a number of training examples and therefore, they do not use summarizations (Buckinx, Moons, Van den Poel, & Wets, 2004). Further, a problem in text classification is the appearance of synonyms. Synonyms are different words with identical or at least similar meanings. In technological texts (e.g. in an applied science research project description) we can find them (assign, associate, classify, correlate etc.). By using a summarization, commonly used synonyms also are summarized that means, we can find them in the centroid vector. Therefore, comparing a term vector from a research project to a centroid vector also considers synonyms. Here, we also see that the centroid-based scheme and the naive Bayes algorithm outperform k-NN, C4.5, and SVM that do not use summarization. Additionally, we focus on the computational complexity of this centroid-based approach. This is relevant because as shown above, we will select 2.850 technologies in our case study that leads to 2.850 classes and that also will lead to a time consuming training and classification phase. In the training phase, we see a linear-time complexity that depends on the number of training examples for the centroid-based approach. We also see a linear complexity in the classification phase that depends on the number of classes. Therefore, the computational complexity in total is very low and it equals the complexity of the naive Bayes algorithm. Thus, the centroid-based scheme and the naive Bayes have a better performance concerning the computational complexity than k-NN, C4.5, and SVM. Using LSI for Technology Classification _______________________________________________________________________ 10 We also see advantages of the centroid-based algorithm concerning the naive Bayes algorithm that applies the Bayes theorem with strong (naive) independence assumptions. Conditioned on each class, this means that different terms independently occur. However, as shown in Sec. 2.1 the independence assumption is not true by using project description as training and test examples. Therefore, we think that the centroid-based algorithm also outperforms the Bayes algorithm. Thus, we use the centroid-based algorithm for the evaluation to compare results of the selected semantic approach to this knowledge based approach. 2.2.2 Semantic approaches As mentioned above, computational techniques are needed that are able to identify the aspect of meaning by calculating the semantic textual patterns. These techniques use eigenvectors in different variations and apply them on statistical procedures. (Jiang, Berry, Donato, Ostrouchov, & Grady, 1999; Luo, Chen, & Xiong, 2011). With these techniques, words that occur in project or technology descriptions are used in the hidden semantic patterns but also words, that might be in these descriptions (Thorleuchter & Van den Poel, 2012d). This enables the identification of a similarity between a project and a set of technologies even if the words in the project description are completely different than the words in the technology descriptions (Tsai, 2012; Christidis, Mentzas, & Apostolou, 2012; Thorleuchter, Weck, & Van den Poel, 2012b). This approach uses LSI as well-known representative of these techniques. It extracts a large number of semantic textual patterns and it reduces their number by considering the values of the eigenvectors (Thorleuchter & Van den Poel, 2013). LSI is a good mean for the identification of application based technological relationships because it fulfills the requirements from Sect. 2.1 as described below. The paragraph a) in Sect. 2.1 indicates that the approach should be able to compute textual similarity in project and technology descriptions. LSI assigns project and technology descriptions to semantic textual patterns. Textual similarity between a project and a Using LSI for Technology Classification _______________________________________________________________________ 11 technology description can be assumed if both descriptions are assigned to the same semantic textual pattern. In the paragraph b) in Sect. 2.1, it is shown that prevalent features that are characteristic for a technology are not simultaneously present in all projects that belong to one technology. LSI as a semantic classification approach always considers this fact by using a semantic indexing that also consists of terms that are not mentioned explicitly in a text but that are related to the corresponding topic. Different terms so not occur independently in the technology or project descriptions as indicated by the paragraph c) in Sect. 2.1. LSI considers this by calculation relationships between projects and technologies based on semantic textual patterns. LSI groups several technologies that are related during the process of creating an application. This means it considers the fact that a project description consists of features from several different technologies as mentioned in the paragraph d) in Sect. 2.1. The paragraphs e), f), and g) indicate that similar, substitutive, integrative, predecessor, and successor technologies have to be identified by considering terms that also describe the application field (beside the technology area). LSI as semantic classification approach considers all related terms (describing a technology as well as describing an application field). Using LSI for Technology Classification _______________________________________________________________________ 12 3 Methodology Fig. 1 shows the processing of our approach in different steps. The methodology selects technology lists or taxonomies as well as information about research projects. Technology descriptions are extracted from the technology lists or taxonomies. Further, projects descriptions are identified or created from the research projects. The technology and project descriptions consist of terms, which describe the technology area as well as the application fields as assumed in Sect. 2.1. They are pre- processed by using tokenization, stop word filtering and stemming. Further, term vectors in a vector space model are created for each technology description and for each project description. LSI is applied to create the semantic textual patterns within the technology descriptions, where the impact of each technology on each semantic textual pattern is calculated. This impact is used to identify related technologies. Technologies with high impact on a specific semantic textual pattern are grouped together in a set of technologies. Using LSI for Technology Classification _______________________________________________________________________ 13 Projects descriptions are projected into the created LSI subspace where LSI calculated the impact of each project on each semantic textual pattern and thus, on each set of technologies. To determine the optimal value of the rank k as the number of semantic textual patterns, a cross-validation procedure is applied on test and training data from the project descriptions. An evaluation is used to compare the assignment of projects to the related technologies by this LSI based approach to the assignment by a knowledge structure based classification approach (centroid based approach). 3.1 Pre-processing The extracted textual information (technology and project description) has to be pre- processed. The aim of this step is to create term vectors in vector-space model. This is because textual information in term vectors can be used for further processing e.g. as input for a singular value decomposition. The textual information has to be prepared in a first step (Thorleuchter & Van den Poel, 2011b). This consists of raw text cleaning where specific objects e.g. images or xml-tags are removed. A dictionary is used to identify and correct typographical errors in the raw text. Tokenization is applied that splits the text in terms where the term unit is defined as words. A conversion of terms to lower case is done (case conversion). In a second step, the text is filtered to reduce the number of distinct terms. Different filtering methods are applied (Thorleuchter, Schulze, & Van den Poel, 2012): Part-of-speech tagging is used to identify the syntactic category of each term (e.g. nouns and verbs) and based on the category, non-informative terms are identified. Stop word filtering is also used to identify the content information of terms. Non-informative terms are discarded (Thorleuchter & Van den Poel, 2012a). As further filtering method, stemming is applied. While words occur in different forms, stemming use a basic form of words to map related words to this basic form. In contrast to lemmatization, stemming does not consider the context of a word. This leads to problems by processing words with the same spelling but with a different meaning. However, after the Using LSI for Technology Classification _______________________________________________________________________ 14 preprocessing step, latent semantic indexing is applied on the terms where the aspect of meaning is considered. Thus, at this time, it is not necessary to use lemmatization. The basic form of words is taken over from a dictionary. If a term is not in the dictionary then a set of production rules are applied to transform the word to its basic form. Terms that appear once or twice are discarded as stated in Zipf distribution (Zipf, 1949; Zeng et al., 2012). Literature shows that term vectors of weighted frequencies outperform term vectors of raw frequencies (Thorleuchter, Van den Poel, & Prinzie, 2010d; Prinzie, & Van den Poel, 2006; Van den Poel, De Schamphelaere, & Wets, 2004; Prinzie, & Van den Poel, 2007). Thus, vectors of weighted frequencies are created for each description in a third step. Based on the calculated weights, the importance of a term within the collection of all descriptions can be estimated (Sparck Jones, 1973). A term is assigned to a large weight if it occurs frequently in a small number of descriptions and seldom in further descriptions (Salton & Buckley, 1988). Based on the proposed weighting scheme from Salton, Allan, & Buckley (1994), the a weight wi,j for a term i in description j is calculated by      m 1p 2 i 2 pji iji ji p dfntf dfntf w ))/(log( )/log( , , , (1) where n is the number of descriptions, m the number of the term vector dimension, dfi is the number of all descriptions containing term i, tfi,j is the term frequency, and idfi, the inverse descriptions frequency (Chen, Chiu, & Chang, 2005). The different length of the descriptions is considered by using a length normalization factor in the divisor of the formula. 3.2 Identification of hidden semantic textual patterns with singular value decomposition Based on the calculated vectors of weighted frequencies, a term-by-description matrix can be created. The dimensionality of this matrix is large because of the large number of distinct terms. Most of the terms only occur frequently in a few numbers of descriptions but not in the Using LSI for Technology Classification _______________________________________________________________________ 15 further descriptions. This leads to many zero values in the matrix and thus, to a small matrix rank. To reduce the dimensionality of the matrix, LSI is used together with a matrix factorization technique. LSI summarizes terms with respect to their semantics (Deerwester et al., 1990). Singular value decomposition as matrix factorization technique identifies the relationships between terms based on their co-occurrences in the descriptions. All related terms are grouped into a semantic textual pattern and each semantic textual pattern has high discriminatory power to other patterns (Thorleuchter & Van den Poel, 2012c). Each semantic textual pattern is assigned to a singular value by processing the singular value decomposition algorithm. The singular value is calculated by splitting the term-by- description matrix A in a product of the matrices U, Σ, and V t . A = U Σ V t (2) Matrix A consists of m terms and n descriptions (m x n matrix) and a rank r (r ≤ min(m,n)) because of many zero values in the matrix. Matrix U consists of m terms and r semantic patterns (m x r matrix), matrix V consists of n descriptions and r semantic patterns (n x r matrix), and matrix Σ consists of the r singular values of matrix A. Thus, Σ is a diagonal (r x r) matrix and the singular values are sorted in descending order. For processing the singular value decomposition, the rank r is important. A large value of r leads to an unmanageable high number of semantic textual patterns. In this case, many semantic textual patterns only occur in a single description but not in several descriptions. For a technology classification, it is important to identify the relationships between different technologies as represented by the technology descriptions. Thus, semantic textual patterns are relevant for this task by considering the relationships between terms based on their co- occurrences in the collection of descriptions. These semantic textual patterns can be identified by reducing the rank r to a parameter k. As shown above, if k is too large e.g. k = r then too many semantic textual patterns are build that are not relevant. Otherwise, if k is too small then many relevant semantic textual Using LSI for Technology Classification _______________________________________________________________________ 16 patterns are not considered. Chen et al. (2010) proposes the use of an operational criterion to get an optimal value of k. We satisfy this by calculating the cross-validated area under the ROC (receiver operating characteristics) curve (AUC) for each k (DeLong, DeLong, & Clarke- Pearson, 1988; Hanley & McNeil, 1982; Halpern et al., 1996; Van Erkel & Pattynama, 1998). For this, we construct several rank-k models as described below. Based on the selection of a specific k, three matrices Uk, Σk and Vk are calculated where the first k columns of U, Σ, and V are retained while from k+1 on, the columns are discarded. Thus, the new term-by-description matrix Ak is based on the reduced matrix rank k<r. Ak = Uk Σk Vk t (3) The new term-by-description matrix Ak contains the k relevant semantic textual patterns. The matrix Uk shows the impact of each term from the descriptions on each semantic textual pattern from Ak. The matrix Vk shows the impact of each (technology or product) description on each of the k patterns. This enables the identification of related technologies on one hand as well as the assigning of projects to a set of related technologies on the other hand. Then, project descriptions have to be projected in the same LSI-subspace (Zhong & Li, 2010). This is because based on the corresponding vector of each project description from Vk, a project description can be assigned to a semantic textual pattern and thus, to a set of related technologies. To create such a vector, a term-by-description vector Ad has to be created for each description d that is based on the terms from matrix A. Then, the vector Vd for matrix Vk can be calculated by 1 kkdd UAV   (4) The project descriptions are split in test and training examples and a fivefold cross-validation is used on training and test examples (Thorleuchter, Herberz, & Van den Poel, 2012). The training examples are used to identify the rank-k model with the best AUC performance and the test examples are used to evaluate the model as described in Sect. 3.3. Using LSI for Technology Classification _______________________________________________________________________ 17 3.3 Evaluation criteria The evaluation focuses on comparing the performance of the semantic classification approach to a knowledge structure approach. Based on the vector Vd, the impact of a project description from the test example on a set of related technologies as represented by a specific semantic textual pattern is given and evaluated by human experts. For each set of related technologies, the number of examples that are correctly identified as related to this set are the true positives (TP) and the number of correctly identified non- related examples are the true negatives (TN). The number of not correctly identified related examples are the false negative (FN) and the number of not correctly identified non-related examples are the false positive (FP). Based on these four values, the commonly used evaluation criteria: the precision, the recall, the sensitivity, and the specificity can be calculated by TP/(TP+FP) (Precision), by TP/(TP+TN) (Recall), by TP/(TP+FN) (Sensitivity), and by TN/(TN + FP) (Specificity). The well-known two dimensional plot of sensitivity versus (1-specificity) is named the receiver operating characteristic curve (ROC) and the AUC is the area under the ROC curve (Migueis, Van den Poel, Camanho, & Cunha, 2012). 4 Empirical verification 4.1 Application Field Defense and Security For the evaluation, we use technology lists or taxonomies as well as current research projects from the application field D&S. The explanation for the selection of this application field is described below. D&S is a field where governments are forced to pay more attention because of the rising asymmetrical threat e.g. terrorism (Greenberg, Irving, & Zimmerman, 2009). A possible solution is the use of new techniques based on results of technological research and development. Thus, an increased funding of D&S based technological research and development can be seen by national and European governments. An example is the European Defence Agency (EDA) that was established in 2004. One important task of this organization is the coordination of defence based research between EU Member States Using LSI for Technology Classification _______________________________________________________________________ 18 (Hoerber, 2012). Further the European Framework Research Program (FP7) contains security research as a central point. As result of growing budgets in the field of D&S research we can monitor a continuous change of the D&S related technological landscape (Thorleuchter, 2008). D&S is not a technology like laser technology or fuel cell technology but it is an application field. Projects in this field are assigned to applied science research and they combine several technologies (e.g. III-V compounds, stealth technologies, human protection technologies, radar technologies) to create an application (a prototype or a demonstrator) (Thorleuchter, Van den Poel, & Prinzie, 2010a). Therefore, the technological landscape of D&S is characterized by national and international research funding organizations. Many of these organizations have defined relevant technologies for future D&S applications on their own. These technologies are published as lists or as objects in hierarchical taxonomies, which means normally a two-level tree structure of classifications for a given set of objects. The objects on the second level represent names of D&S related technologies and the objects on the first level represent manually created labels for these technologies. Technology names are described by few technical words e.g. "passive radar technologies" or "active radar technologies" labelled by "radar technologies". Additionally, descriptions of technologies that consist of terms describing the technology itself as well as the corresponding application fields are given. The technologies are the basis for research funding activities. That means proposals are manually classified by research funding organizations according to their own technologies. If proposals do not fit with these technologies, they normally are not accepted (McLeish & Nightingale, 2007). In order to acquire funding in this area, researchers should have knowledge about national and international research funding organizations and their appertaining technologies. Thus, this approach considers the relevant organizations and their technologies as mentioned in Sect. 4.1.1. Using LSI for Technology Classification _______________________________________________________________________ 19 Beside this overview, a further aspect is to consider. Sometimes D&S research is sensitive concerning technological proliferation (Perry, 2004). That means some technologies have the potential to significantly enhance or degrade national D&S capabilities in the future or to permit significant advances of military capabilities of potential adversaries. Research planners and researchers should have knowledge about the sensibility of their research. Therefore, the overview in Sect. 4.1.1 also includes technologies with proliferation control aspects. Additionally in Sect. 4.1.2, we focus on the acquisition of research projects in the D&S field. 4.1.1 Technologies in D&S In this section we describe examples for taxonomies and lists of D&S related technologies. The European Defence Agency (EDA) has been created to help member states of European Union (EU) develop their defence capabilities for crisis-management operations under the "European Security and Defence Policy" (Oikonomou, 2012). One aim of the EDA is to stimulate European research and technology collaboration, focused on improving defence capabilities. The EDA taxonomy of technologies is the basis for this funding and contains about 200 technologies in defence context. The Western European Armaments Group (WEAG) is a forum for armaments cooperation established by defence ministers of the European NATO nations. It coordinates defence- related research and development projects inside the European Union (Te Kulve & Smit, 2003). The coordination activities of the WEAG are transferred to EDA in 2005 but the WEAG taxonomy of technologies is still in use by many national ministries of defence for defence research funding. The WEAG taxonomy of technologies contains about 200 technologies including underpinning defence technologies, weapon systems related technologies and technologies for (military) products. Using LSI for Technology Classification _______________________________________________________________________ 20 The stakeholder's platform for supply chain mapping, market condition analysis and technologies opportunities (STACCATO) is a European Commission-financed activity with the objective to prepare a proposal for a strategic research plan for European security. The STACCATO taxonomy of technologies builds on technology taxonomies from WEAG and United Kingdom and contains about 800 technologies in security context. The European Security Research Advisory Board (ESRAB) shall make recommendations to the European Commission in the field of strategic missions, focus areas and priorities setting for future security research programs (Remuss, 2010). The ESRAB technology list therefore is a basis for the security part of the European framework research program (FP7). It consists of 150 technologies in security context. Each European member state has own technology collections (lists or taxonomies) for D&S. In general they are created by ministries of defence and unfortunately they are very often classified as restricted information but most of these technological collections are based on WEAG taxonomy like described above. The Militarily Critical Technologies List (MCTL) is a compendium of existing goods and technologies that would permit significant advances in the development, production and use of military capabilities of potential adversaries (Bradley, 1989). This technology list contains about 600 technologies in proliferation control context. The Developing Science and Technologies List (DSTL) is a compendium of scientific and technological capabilities being developed worldwide that have the potential to significantly enhance or degrade US military capabilities in the future. This list includes technologies from basic research, applied research and advanced technology development and it contains about 900 technologies in proliferation control context. Further DSTL and MCTL are a basis for the technological part of the Waasmar List, the armaments export control list for conventional arms and dual-use goods and technologies. Using LSI for Technology Classification _______________________________________________________________________ 21 4.1.2 Research Projects from D&S For our test and training set, we need D&S related and innovative projects. They can be found in the United States Small Business Innovation Research (SBIR) Program and the Small Business Technology Transfer (STTR) Program. SBIR and STTR ensure that small, high-tech and innovative businesses are a significant part of the United States federal government's research and development efforts. Eleven federal departments participate in SBIR and STTR programs awarding $ 2 billion to small high-tech businesses (Lockett, Siegel, Wright, & Ensley, 2005). The central point in SBIR and STTR research is D&S because of the height award amount of the Department of Homeland Security, the Environmental Protection Agency and the Department of Defense divided in Air Force, Army, Chemical and Biological Defense Program (CBD), Defense Advanced Research Projects Agency (DARPA), Defense Logistics Agency (DLA), Defense Microelectronics Activity (DMEA), Defense Technical Information Center (DTIC), Defense Threat Reduction Agency (DTRA), Missile Defense Agency (formerly BMDO), National Geospatial-Intelligence Agency (NGA) (formerly NIMA), Navy, Special Operations Acquisition and Logistics Center (SOCOM). The projects are published as non-proprietary textual data with title and abstract. The abstract consists of terms that represent the technological field as well as the application field. Therefore, we use them to evaluate the approach. 4.2 Data Characteristics In this study, we use the technology and project descriptions from Sect. 4.1. All descriptions are in English language. Using LSI for Technology Classification _______________________________________________________________________ 22 Table 1 provides summary information of the (randomly-selected) training and test set. The optimal SVD dimension is calculated using the training set and a regression model is estimated. The test set is used to show the success of the regression model compared to the frequent baseline as calculated from the relative percentage in Table 1. Number of items Relative percentage Taxonomies / Technology lists 6 Technology descriptions 2900 Training set: Project descriptions 480 80 Test set: Project descriptions 120 20 Total 600 Table 1: Overview on the data characteristics 4.3 Optimal dimension selection The high number of 2900 technology description can be reduced to a small number of sets of related technologies because many of these descriptions describe equal technologies or similar technologies while other describe substitutive, integrative, predecessor, or successor technologies. A human based evaluation identifies the AUC for specific selected values of k. For other values of k, regression based interpolation is used to construct new data points between two known data points. The number of selected semantic textual patterns (SVD dimension or rank k) is represented by the x-axis. The y-axis represents the cross-validated AUC (see Fig. 2). It can be seen that the AUC increases up to the use of 200 semantic textual patterns. Using more than 200 semantic textual patterns in the SVD model leads to a higher complexity of the model however, the cross-validated AUC performance does not increase. Thus, the parameter k is set to 200. Using LSI for Technology Classification _______________________________________________________________________ 23 Figure 2: Calculating an optimal SVD dimension The baseline has an AUC value of 50. It is also important to know that this approach outperforms the baseline even if a small value of k (at about 40) is selected. 4.4 Case study results Here, an example for the results of this approach is presented. These example results confirm results of a further knowledge structure based approach (Thorleuchter and Poel 2011). Unfortunately, that approach could not be used for the evaluation because the precision and recall values are calculated based on a very small subset of technologies and projects. A research project is used as described below: The 2002 SBIR / STTR research project 57405 with the title: “Tunable diode-pumped IR laser source” has the following abstract: “The Space Based Laser (SBL) requires a Low Energy Laser (LEL) system to serve as a high Using LSI for Technology Classification _______________________________________________________________________ 24 fidelity surrogate during startup and optical alignment portions of test operations. In this proposal, we will develop a CW, diode-pumped solid state laser that can meet the requirements for the LEL, namely a CW power level in the 1-10 W range, and wavelengths in the 2600-2900-nm region. The device, based on a direct diode-pumped Er:YLF crystal, is rugged, compact, tunable, and well suited for space - based systems.” The semantic approach has assigned this project to a set of 12 related technologies as mentioned in Table 2. Technology list / Taxonomy Technology label EDA Communications Systems – IR / Visible / UV ESRAB Space Systems WEAG Laser Sensors STACCATO Space Based Lasers STACCATO Communications systems - IR / Visible / UV laser STACCATO IR / Visible / UV laser MCTL Laser Location Systems MCTL Multispectral and Hyperspectral Space Sensor Systems MCTL Space Laser Diodes MCTL Tunable Solid-State Lasers DSTL Excimer Lasers (LELs), Excimer DSTL Free Electron Laser (FEL) (HPM NB Sources) Table 2: Example of related technologies from different technological lists and taxonomies 4.5 Comparing the performance of the approach Besides evaluating the optimal performance of this approach as based on the value of k, the overall performance of this semantic approach is also compared to a centroid approach by Using LSI for Technology Classification _______________________________________________________________________ 25 use of precision and recall measures. Both approaches are comparable because they assign a project to a set of technologies. For each of the 200 sets of technologies, the precision and recall values for the semantic approach are estimated by human experts as described in 3.3. Then, the average precision and recall values are calculated. We have defined a centroid classifier that assigns a project to a technology if at least z% of all stemmed and stop word filtered terms from one technology description appear in the project description. If z is too large then probably we do not get many projects assigned to a technology in the learning phase. This decreases the quality of the corresponding centroid vector. If z is too small then probably we get many projects assigned to a technology that normally are not related to this technology. This also decreases the quality of the corresponding centroid vector. The value of z is estimated by a human expert. Each centroid vectors represents a set of technologies. This is used for the calculation of precision and recall values as described above. For comparing, the F-measure is used because precision and recall are equally important. The semantic approach gets a precision of 76 % at a recall of 48 % while the centroid approach gets a precision of 69 % at a recall of 44 %. This leads to an F-measure of 59 % for the semantic approach in contrast to an F-measure of 54 % for the centroid approach. 5 Conclusions This paper proposes a new approach that classifies applied science research projects according to corresponding technologies of research-funding organizations. It considers that technologies are related during the process of creating an application. LSI is used to identify these related technologies based on semantic textual patterns occurring in the technology descriptions. Project descriptions - divided in training and test examples - are projected into the LSI subspace. They are also used to estimate the parameters and to evaluate this approach. Using LSI for Technology Classification _______________________________________________________________________ 26 As a result, it is shown that LSI as semantic classification approach is suited to identify the relationships among technologies because it considers well terms from the technology area as well as terms from the application field. The automated identification of these semantic relationships is not possible with knowledge structure based approaches. Thus, the results contribute to the existing literature concerning the application based technological relationships. Further, LSI is suited to assign projects to a specific set of related technologies as represented by a semantic textual pattern. Here, LSI also outperforms knowledge structure based approaches that assign projects to each technology separately. Thus, the results are helpful for researchers and research-funding planners. Future avenues of research could be the use of this approach in the field of explorative scenario-based technological roadmapping. This is because this specific roadmapping approach considers the relationships between the investigated technologies during the process of creating applications. Up to now, the identification of relationships is done manually by human experts. That restricts this roadmapping approach to a small number of investigated technologies. However, using the automated LSI based approach is helpful for identifying relationships and it probably enables the investigating of a large number of technologies. For the case study, technologies from the field of D&S are selected. A further direction of research is the use of different application fields to show the success of this approach. Acknowledgments This project is realized using SAS v9.1.3, SAS Text Miner v5.2, Fraunhofer Idea Web Miner v1.0, and Matlab v7.0.4. Further, a self-developed program is used for web content mining to collect the data. Using LSI for Technology Classification _______________________________________________________________________ 27 References Beaudry, C., & Allaoui, S. (2012). Impact of public and private research funding on scientific production: The case of nanotechnology. Research Policy, In Press. Bradley, V.G. (1989). National strategies for technology trade: A response to Chris Hill. Technology in Society, 11(2), 181-188. Bradshaw, D.H., Empy, C., Davis, P., Lipschitz, D., Nakamura, Y., & Chapman, C.R. (2008). Trends in Funding for Research on Pain: A Report on the National Institutes of Health Grant Awards Over the Years 2003 to 2007. The Journal of Pain, 9(12), 1077-1087. Buckinx, W., Moons, E., Van den Poel, D., & Wets, G.(2004). Customer-adapted coupon targeting using feature selection. Expert Systems with Applications, 26(4), 509-518. Chen, M. C., Chiu, A. L., & Chang, H. H. (2005). Mining changes in customer behavior in retail marketing. Expert System with Applications, 28(4), 773-781. Chen, M.-Y., Chu, H.-C., & Chen, Y.-M. (2010). Developing a semantic-enable information retrieval mechanism. Expert Systems with Applications, 37(1), 322-340. Choi, J.Y., Lee, J.H., & Sohn, S.J. (2009). Impact analysis for national R&D funding in science and technology using quantification method II. Research Policy, 38(10), 1534-1544. Choi, S., Park, H., Kang, D., Lee, J.Y., & Kim, K. (2012). An SAO-based text mining approach to building a technology tree for technology planning. Expert Systems with Applications, 39(13), 11443- 11455. Christidis, K., Mentzas, G., & Apostolou, D. (2012). Using latent topics to enhance search and recommendation in Enterprise Social Software. Expert Systems with Applications, 39(10), 9297-9307. Deerwester, S., Dumais, S., Furnas, G., Landauer, T., & Harshman, R. (1990). Indexing by latent semantic analysis. Journal of the American Society for Information Science, 41(6), 391-407. DeLong, E. R., DeLong, D. M., & Clarke-Pearson, D. L. (1988). Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics, 44(3), 837-845. Finzen, J., Kintz, M., & Kaufmann, S. (2012). Aggregating web–based ideation platforms. International Journal of Technology Intelligence and Planning, 8(1), 32-46. Fleck, J., & Howells, J. (2001). Technology, the Technology Complex and the Paradox of Technological Determinism. Technology Analysis & Strategic Management, 13(4), 523-531. Using LSI for Technology Classification _______________________________________________________________________ 28 Gericke, W., Thorleuchter, D., Weck, G., Reiländer F., & Loß, D. (2009). Vertrauliche Verarbeitung staatlich eingestufter Information - die Informationstechnologie im Geheimschutz. Informatik Spektrum, 32(2), 102-109. Geschka, H., Lenk, T., & Vietor, J. (2002). The idea and project database of WELLA AG. International Journal of Technology Management, 23(5), 410-416. Geschka, H. (1983). Creativity techniques in product planning and development: A view from West Germany. R&D Management, 13(3), 169-183. Greenberg, M., Irving,W., & Zimmerman, R. (2009). Allocating U.S. Department of Homeland Security funds to States with explicit equity, population and energy facility security criteria. Socio-Economic Planning Sciences, 43(4), 229-239. Grimpe, C. (2012). Extramural research grants and scientists’ funding strategies: Beggars cannot be choosers? Research Policy, In Press. Halpern, E. J., Albert, M., Krieger, A. M., Metz, C. E., & Maidment, A. D. (1996). Comparison of receiver operating characteristic curves on the basis of optimal operating points. Academic Radiology, 3(3), 245-253. Han, E.S., & Karypis, G. (2000). Centroid-Based Document Classification: Analysis and Experimental Results. In: Principles of Data Mining and Knowledge Discovery (pp. 116-123). Berlin: Springer. Hanley, J.A., & McNeil, B.J. (1982). The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology, 143(1), 29-36. Herstatt, C., & Geschka, H. (2002). Need assessment in practice - methods, experiences and trends. International Journal of Entrepreneurship and Innovation Management, 2(1), 56-68. Hicks, D. (2012). Performance-based university research funding systems. Research Policy, 41(2), 251-261. Hoerber, T. (2012). New horizons for Europe – A European Studies perspective on European space policy. Space Policy, 28(2), 77-80. Jiang, J., Berry, M. W., Donato, J. M., Ostrouchov, G., & Grady, N. W. (1999). Mining consumer product data via latent semantic indexing. Intelligent Data Analysis, 3(5), 377-398. Jiménez, C.H.O., Garrido-Vega, P., Díez de los Ríos, J.L.P., & González, S.G. (2011). Manufacturing strategy–technology relationship among auto suppliers. International Journal of Production Economics, 133(2), 508-517. Jiricka, A., & Pröbstl, U. (2012). The role of SEA in integrating and balancing high policy objectives in European cohesion funding programmes. Environmental Impact Assessment Review, In Press. Kim, Y., Toh, K.A., Teoh, A.B.J., Eng, H.L., & Yau, W.Y. (2012). An online AUC formulation for binary classification. Pattern Recognition, 45(6), 2266-2279. Using LSI for Technology Classification _______________________________________________________________________ 29 Ko, Y., & Seo, J. (2009). Text classification from unlabeled documents with bootstrapping and feature projection techniques. Information Processing & Management, 45(1), 70-83. Lee, C.H., & Wang S.H. (2012). An information fusion approach to integrate image annotation and text mining methods for geographic knowledge discovery. Expert Systems with Applications, 39(10), 8954- 8967. Lepori, B. (2011). Coordination modes in public funding systems. Research Policy, 40(3), 355-367. Lin, M-H., & Hong, C-F. (2011). Opportunities for Crossing the Chasm between Early Adopters and the Early Majority through New Uses of Innovative Products. The Review of Socionetwork Strategies, 5(2), 27-42. Lockett, A., Siegel, D., Wright, M., & Ensley, M.D. (2005). The creation of spin-off firms at public research institutions: Managerial and policy implications. Research Policy, 34(7), 981-993. Ludwig, R., Roson, R., Zografos, C., & Kallis, G. (2011). Towards an inter-disciplinary research agenda on climate change, water and security in Southern Europe and neighboring countries. Environmental Science & Policy, 14(7), 794-803. Luo, Q., Chen, E., & Xiong, H. (2011). A semantic term weighting scheme for text categorization. Expert Systems with Applications, 38(10), 12708-12716. Madjarov, G., Kocev, D., Gjorgjevikj, D., & Džeroski, S. (2012). An extensive experimental comparison of methods for multi-label learning. Pattern Recognition, 45(9), 3084-3104. McLeish, C., & Nightingale, P. (2007). Biosecurity, bioterrorism and the governance of science: The increasing convergence of science and security policy. Research Policy, 36(10), 1635-1654. Migueis, V. L., Van den Poel, D., Camanho, A. S., & Cunha, J.F. (2012). Modeling partial customer churn: On the value of first product-category purchase sequences. Expert Systems with Applications, 39(12), 11250-11256. Mobjörk, M., & Linnér, B.O. (2006). Sustainable funding? How funding agencies frame science for sustainable development. Environmental Science & Policy, 9(1), 67-77. Oikonomou, I. (2012). The European Defence Agency and EU military space policy: Whose space odyssey? Space Policy, 28(2), 102-109. Park, D, H., Kim, H. K., Choi, I. Y., & Kim, J. K. (2012). A literature review and classification of recommender systems research. Expert Systems with Applications, 39(11), 10059-10072. Perry, W.J. (2004). Military technology: an historical perspective. Technology in Society, 26(2), 235- 243. Prinzie, A., & Van den Poel, D. (2006). Investigating purchasing-sequence patterns for financial services using Markov, MTD and MTDg models. European Journal of Operational Research, 170(3), 710-734. Using LSI for Technology Classification _______________________________________________________________________ 30 Prinzie, A., & Van den Poel, D. (2007). Predicting home-appliance acquisition sequences: Markov/Markov for Discrimination and survival analysis for modeling sequential information in NPTB models. Decision Support Systems, 44(1), 28-45. Radder, H. (2009). Science, Technology and the Science-Technology Relationship. Philosophy of Technology and Engineering Sciences, 2009, 65-91. Remuss, N.L. (2010). Creating a European internal security strategy involving space applications. Space Policy, 26(1), 9-14. Rubenstein, A.H., Douds, C.F., Geschka, H., Kawase, T., Miller, J.P., Saintpaul, R., & Watkins, D. (1977). Management perceptions of government incentives to technological innovation in England, France, West Germany and Japan. Research Policy, 6(4), 324-357. Salton, G., Allan, J., & Buckley, C. (1994). Automatic structuring and retrieval of large text files. Communications of the ACM, 37(2), 97-108. Salton, G., & Buckley, C. (1988). Term-weighting approaches in automatic text retrieval. Information Processing & Management, 24(5), 513–523. Shi, L., & Setchi, R. (2012). User-oriented ontology-based clustering of stored memories. Expert Systems with Applications, 39(10), 9730-9742. Sparck Jones, K. (1973). Index term weighting. Information Storage and Retrieval, 9(11), 619-633. Sudhamathy, G. & Jothi Venkateswaran, C. (2012). Fuzzy Temporal Clustering Approach for E- Commerce Websites. International Journal of Engineering and Technology, 4(3), 119-132. Subramanian, A.M., & Soh, P.H. (2010). An empirical examination of the science–technology relationship in the biotechnology industry. Journal of Engineering and Technology Management, 27(3- 4), 160-171. Takci, H. & Güngör, T. (2012). A High Performance Centroid-based Classification Approach for Language Identification. Pattern Recognition Letters, In Press. Te Kulve, H., & Smit, W.A. (2003). Civilian–military co-operation strategies in developing new technologies. Research Policy, 32(6), 955-970. Thorleuchter, D. (2008). Finding Technological Ideas and Inventions with Text Mining and Technique Philosophy. In C. Preisach, H. Burkhardt, L. Schmidt-Thieme, & R. Decker (Eds.), Data Analysis, Machine Learning, and Applications (pp. 413-420). Berlin: Springer. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2010a). Mining Ideas from Textual Information. Expert Systems with Applications, 37(10), 7182-7188. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2010b). A compared R&D-based and patent-based cross impact analysis for identifying relationships between technologies. Technological Forecasting and Social Change, 77(7), 1037-1050. Using LSI for Technology Classification _______________________________________________________________________ 31 Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2010c). Mining innovative ideas to support new product research and development. In H. Locarek-Junge, & C. Weihs (Eds.), Classification as a Tool for Research (pp. 587-594). Berlin: Springer. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2010d). Extracting consumers needs for new products - A web mining approach. In Proceedings WKDD 2010 (p. 441.). Los Alamitos: IEEE Computer Society. Thorleuchter, D., & Van den Poel, D. (2011a). Semantic Technology Classification - A Defence and Security Case Study. In Proc. Uncertainty Reasoning and Knowledge Engineering (pp. 36-39). New York: IEEE. Thorleuchter, D., & Van den Poel, D. (2011b). Companies Website Optimising concerning Consumer’s searching for new Products. In Proc. Uncertainty Reasoning and Knowledge Engineering (pp. 40-43). New York: IEEE. Thorleuchter, D., & Van den Poel, D. (2011c) High Granular Multi-Level-Security Model for Improved Usability. In: System Science, Engineering Design and Manufacturing Informatization 1 (pp. 191-194). New York: IEEE. Thorleuchter, D., Van den Poel, D., & Prinzie, A. (2012). Analyzing existing customers’ websites to improve the customer acquisition process as well as the profitability prediction in B-to-B marketing. Expert Systems with Applications, 39(3), 2597-2605. Thorleuchter, D., Herberz, S., & Van den Poel, D. (2012). Mining Social Behavior Ideas of Przewalski Horses. Lecture Notes in Electrical Engineering, 121, 649-656. Thorleuchter, D., & Van den Poel, D. (2012a). Extraction of Ideas from Microsystems Technology. Advances in Intelligent and Soft Computing, 168, 563-568. Thorleuchter, D., & Van den Poel, D. (2012b). Predicting e-commerce company success by mining the text of its publicly-accessible website. Expert Systems with Applications, 39(17), 13026-13034. Thorleuchter, D., & Van den Poel, D. (2012c). Using NMF for Analyzing War Logs. Communications in Computer and Information Science, 318, 73-76. Thorleuchter, D., Schulze, J., & Van den Poel, D. (2012). Improved Emergency Management by Loosely Coupled Logistic System. Communications in Computer and Information Science, 318, 5-8. Thorleuchter, D., & Van den Poel, D. (2012d). Using Webcrawling of Publicly-Available Websites to Assess E-Commerce Relationships. In SRII Global Conference 2012 (pp. 402-410). San Jose, CA, USA: IEEE. Thorleuchter, D., & Van den Poel, D. (2012). Improved Multilevel Security with Latent Semantic Indexing. Expert Systems with Applications, 39(18), 13462-13471. Using LSI for Technology Classification _______________________________________________________________________ 32 Thorleuchter, D., Weck, G., & Van den Poel, D. (2012a). Granular Deleting in Multi Level Security Models - an Electronic Engineering approach. Lecture Notes in Electrical Engineering, 1, 177, 609- 614. Thorleuchter, D., Weck, G., & Van den Poel, D. (2012b). Usability based Modeling for Advanced IT- Security - an Electronic Engineering approach. Lecture Notes in Electrical Engineering, 1, 177, 615- 619 Tsai, H.H. (2012). Global data mining: An empirical study of current trends, future forecasts and technology diffusions. Expert Systems with Applications, 39(9), 8172-8181. Van den Poel, D., De Schamphelaere, J., & Wets, G (2004). Direct and indirect effects of retail promotions on sales and profits in the do-it-yourself market. Expert Systems with Applications, 27(1), 53-62. Van Erkel, A. R., & Pattynama, P. M. T. (1998). Receiver operating characteristic (ROC) analysis: Basic principles and applications in radiology. European Journal of Radiology, 27(2), 88-94. Yu, L., Hurley, T., Kliebenstein, J., & Orazem, P. (2012). A test for complementarities among multiple technologies that avoids the curse of dimensionality. Economics Letters, 116(3), 354-357. Zeng, J., Duan, J., Cao, W., & Wu C. (2012). Topics modeling based on selective Zipf distribution. Expert Systems with Applications, 39(7), 6541-6546. Zhong, J., & Li, X. (2010). Unified collaborative filtering model based on combination of latent features. Expert Systems with Applications, 37(8), 5666-5672. Zipf, G. K. (1949). Human Behaviour and the Principle of Least Effort. Cambridge: Addison-Wesley. 
work_zdhgvdevnbbinp4kmsooxqjcza	----	1 4 0 3 3 4 2 1 1 5 S W P 15/90 S T A T E - O F - T H E - A R T D E V E L O P M E N T S IN E X P E R T S Y S T E M S A N D S T R A T E G IC M A R K E T I N G P L A N N I N G P R O F E S S O R M A L C O L M M C D O N A L D a n d H U G H W I L S O N Cranfield School of M a n a g e m e n t Cranfield Institute of Technology Cranf ield Bedford M K 4 3 O A L (Tel: 0 2 3 4 7 5 1 1 2 2 ) Copyright: M c D o n a l d a n d Wilson 1 9 8 9 STATE-OF-THE-ART D E V E L O P M E N T S IN E X P E R T S Y S T E M S A N D S T R A T E G I C M A R K E T I N G P L A N N I N G by Malcolm H.B. McDonald and Hugh N. Wilson A B S T R A C T T h e paper describes a case history of the development of a n Expert System in Strategic Marketing Planning c o d e n a m e d E X M A R . It traces the evolution of the system from the formation of the DTI club two years a g o to the launch of the prototype model. T h e paper outlines the technical a n d domain-specific obstacles encountered e n route a n d h o w these were overcome. A n u m b e r of conclusions are drawn from the project. T h e principal o n e is that there is a bright future for expert systems in the field of strategic management. Professor Malcolm McDonald, o n e of the authors of this paper, is the principal expert to the club. H u g h Wilson, the other author, is a senior consultant with Artificial Intelligence Ltd. H e was also the knowledge engineer a n d the project manager. - 1 - -c - STATE-OF-THE-ART DEVELOPMENTS IN EXPERT SYSTEMS AND STRATEGIC MARKETING PLANNING by Malcolm H.B. McDonald and Hugh N. Wilson 1. INTRODUCTION After nearly a quarter of a century of Expert Systems, virtually no progress has been made in the domain of marketing and there are few products and no on-line systems available’ O2 There are no shortcuts to building good Expert Systems. It takes a considerable amount of skill, patience and several years of effort to develop an Expert System in a new area and get it in the field3. -.. During the 198Os, Japanese activity in the field of Expert Systems and related technologies prompted the EEC to give birth to the ESPRIT programme in an attempt to integrate European efforts. This in turn led to the DTI sponsored ALVEY and IED programmes, and other initiatives. An outcrop of these is a new DTI - sponsored club called EXMAR. EXMAR is a club of ten major British companies. Formed in 1987, its objectives are to investigate the possibility of computerised assistance for strategic marketing planning by the development of a prototype, and to spread awareness of expert systems in club member organisations. It is funded by contributions from the member companies, and by the Department of Trade and Industry. The club’s primary source of marketing expertise is Professor Malcolm McDonald of Cranfield School of Management. Marketing experience within club member companies is also being tapped. - 2 - T h e involvement of Artificial Intelligence Ltd. of Watford, with the club as technical contractor b e g a n in the second half of 1988, w h e n AI Ltd. conducted a n analysis phase, followed by production of a demonstrator a n d a n appraisal of the way forward. T h e demonstrator was built in the Interlisp programming environment a n d the Loops object system o n Xerox 1 1 8 6 workstations. T h e Requirements, Functional a n d Design Specifications for a prototype, to b e used for experimentation a n d evaluation by club m e m b e r s in their o w n organisations, were completed in August 1989, a n d work is n o w in progress o n the implementation. T h e prototype will run o n IBM-compatible 3 8 6 machines, using Smalltalk- a n d Analyst, a n d was completed in January 1990. Pumose and Structure of PaDer T h e purpose of this paper is to outline the progress of the E X M A R project, a n d to draw conclusions about appropriate computer support for marketing planning. In the next section, the approach taken to the analysis phase at the start of the project is outlined, a n d the system objectives that were derived are described. T h e nature of the logical model that e m e r g e d is discussed, a n d the demonstrator system based o n it is described, emphasising the nature a n d style of the support to the user provided by the system, h o w this reflects the logical model, a n d h o w this meets the system objectives. Section 4 describes the feedback from club m e m b e r s o n the Demonstrator a n d h o w this is influencing current work o n the Prototype. T w o review sections 5 a n d 6, discuss the nature of the marketing planning domain. a n d the deductions that can b e m a d e about appropriate computer support. - 3 - Finally, s o m e conclusions are drawn o n the appropriate approach to systems development in such “soft” areas as marketing planning. - 2. P R E V I O U S W O R K A N D E A R L Y O B S E R V A T I O N S T h e initial requirements analysis produced a n u m b e r of interesting problems for the project, which were to sow the seeds of expensive a n d time-consuming delay. These problems can b e summarised as follows:- .- - (8 (ii) (iii) it b e c a m e clear that not m a n y of the m e m b e r companies were particularly a u fait with the methodology of marketing planning. This led to the problem of setting clear objectives for the project. the diversity of c o m p a n y industry types, ranging from capital goods to service industries, meant that n o subsequent system could possibly b e suitable for all circumstances. problems a n d subsequent proposed objectives ranged from “T o support a formal planning framework to improve discipline during the planning process” a n d “T o support further understanding of the effects of currency fluctuations” to “T o promote discipline in pricing control”. For these reasons, it was decided to focus o n the process of marketing planning itself rather than o n any situation-specific system. A firm of software consultants was appointed project m a n a g e r a n d a knowledge based systems house was appointed principal contractor. - 4 - -. ,- The systems house began a series of twelve half day interviews with Professor McDonald in order to develop a formal paper model as a basis for computerisation. Unfortunately, although taped and transcribed, they were largely unfocussed due to the inexperience of the interviewers and little progress was made towards formal modelling of the marketing planning process, in spite of very specific guidance given by Professor McDonald to the interviewers. The problem centred around lack of proper project control by the project managers, confused expectations by members of the club based on marketing planning naivety, the inexperience of the knowledge engineers, and the passive role of the domain expert, which was necessary in view of the nature of the project. Several attempts on Professor McDonald’s part to guide the system were brushed aside as politically inexpedient. The result was that the paper outlining the tasks to be performed by the computer system targeted the whole marketing planning process rather than any subset, and because of this breadth, the process to be computerised was not documented in any detail, nor backed up by any substantive models and interrelationships. Other specifications required by the development methodology in use, such as financial requirements, system structure and so on, were never produced. At this point, the project manager appointed new software consultants to take over the feasibility study and the delivery system. The new contractor set about finding some common requirements among end users in order to outline the domain model, with a boundary definition showing which parts of the model would be tackled by the computer system. They set about establishing the following areas: . n scope n constraints n organisational impact - 5 - # - rc- -- - n maintainability n extensibility n technology n time scales n risk a n d cost versus quantifiable benefits Artificial Intelligence Ltd., the n e w software consultants, drew various conclusions about the appropriate technical approach: n Need for focus T h e previous work h a d b e e n o n a broad front, involving analysis into all aspects of strategic marketing planning. This is a vast topic, tackling m a n y of the most fundamental problems inherent in business activity, a n d progress was therefore slow. There was a n e e d to focus o n a subset of the overall problem. n Feasibility and utility to be established T h e very title of the club, “Expert Systems in Marketing”, suggested that the use of expert systems techniques in this area was possible a n d appropriate. This assumption of feasibility was based o n the observation that there existed demonstrable expertise, but why this might imply a classic rule-based expert system h a d not b e e n addressed. This was a doubly large assumption as n o previous systems (or work towards systems) were known in this application area. There was a n e e d to address this early, a n d the related issue of h o w any system would b e of use to the marketing planner. n Modelling and representation It was decided that the appropriate first step was to carry out analysis in a closely scoped subset of the problem, with the emphasis o n modelling the area using whatever formal techniques were appropriate. A n example of the choices deliberately not m a d e at the start was whether any modelling of expertise adopted the - 6 - rr- “low road” of embedding the expertise in data structures and code, the “high road” of an explicit, “deep” representation, or the “middle road” of an explicit but heuristic representation4. In this modelling work, the emphasis would be on representation rather than computation, as the essential first step towards any computer system. n The marketing swamp Marketing will be referred to later in this paper as a swamp of intuitive, experience- based practice with the occasional rocky peak of formal techniques. In the experience of Artificial Intelligence Ltd. the best place to start when modelling such “soft” domains was often on the boundary between the soft area and neighbouring more readily formalisable areas. In this case, that meant starting with the established formal techniques and working out from there. - 3. RESULTSOFANALYSISWORK. ANDDEMONSTRATOR Several analysis sessions were held with Professor McDonald, and with marketing practitioners in club member organisations. This resulted in an overall EXMAR system objective, an outline model that was used as the basis for a demonstrator system, and a list of areas where further work was required. The overall EXMAR system objective was defined to be: “To provide assistance for the marketing planning process in such a way as to spread knowledge and further understanding of how and why the multifarious factors of the market interact and serve to define the parameters of the business activity.” The remainder of this section describes features of the model, and how these were exploited in the demonstrator. The structure is an interleaved description of the two: - 7 - - L - - each subsection describes a model feature, a n d the relevant aspects of the demonstrator. 3.1 Assistance to aid in interwetation and understanding T h e model covers the data manipulated by a marketing planner w h e n developing a strategic marketing plan, a n d structures the marketing planner’s task. M a n y of the individual subtasks or processes of this task involve modelling by the user of the business context, or interpretation by the user of the information entered. There is m u c h that a computer system based o n the model cannot d o for the user, a n d it b e c a m e increasingly clear that its most appropriate aim is to assist. T h e objective of the demonstrator was therefore to provide a n interactive system that supports a marketing planner by providing tools that help the user to represent the state of the markets a n d products under consideration; to interpret this information so as to gain a n understanding of the markets a n d o n e ’s place within them; a n d to determine a course of action based o n this understanding. 3.2 Model of the Drocess of eeneratine a marketing nlan A hierarchical breakdown of the process the marketing planner should adopt to generate a marketing plan was defined. Encouraging the user to adopt this process is of value in itself, as the process incorporates m u c h experience that helps avoid c o m m o n pitfalls: for example, the n e e d to arrive at a n appropriate understanding of the current situation before setting objectives for the future. T h e demonstrator uses this hierarchy as a basis for the user’s navigation round the system. T h e initial screen display is shown to illustrate this (Figure 1). Also sholvn is a window for m o r e detailed navigation round a particular stage. - 8 - - II E X M .4R. demonstrator Araldal Illtd~ U m N e d Prodlet relMlonshlp W I U I M * . W 1 S e t Objectha fof N e w Productl/Markatr Figure 1 Initial screen display, with an example of a detailed browser E a c h box in the graphical browsers represents a stage of the process. T h e user carries out a stage by selecting a box with the m o u s e : the system then takes the appropriate action, which m a y for e x a m p l e b e to present the user with a form to fill in, or to o p e n a m o r e detailed browser of the process for that stage. T o give a n overview of the process: Select/Define Business Unit identifies which area of the business the marketing plan is for, a n d records the purpose of the business area. F o c u s identifies which of the unit’s markets a n d products are of interest. Conduct Audit assesses the current position of the products a n d markets. Forecast predicts the future position of the products a n d markets, a s s u m i n g w e d o not intervene, as a base-line for objective setting. Finally, Set Objectives a n d Strategies sets objectives for the business unit b a s e d o n the information collected, analysed a n d summarised; a n d defines strategies by which the objectives are to b e met. - 9 - - Detailed browsers contain icons showing the nature of the support offered for a particular stage: for example, there are icons for graphical displays of information, for tables of numbers, a n d for free text. T h e Predict Relationshiu with Markets browser is illustrated as a n example, also in Figure 1. - Users will largely g o through the process depth first a n d top to bottom; but they are free to d o otherwise, as there are m a n y cases where they m a y legitimately wish to d o so. l Markof Share F A C T O R S l Price l Sales V o l u m e . 3.3 Generallv aDDlicable. sound data model A data model was developed that captured a n d related the information considered during production of a strategic marketing plan. It has proved essentially sound, a n d of general applicability to the wide range of marketing situations represented by the diverse club m e m b e r companies. A simplified entity-relationship diagram of the model is given below in Figure 2, a n d briefly described. I N V O L V E M E N T IN ‘s c o m o n M w * a n d M A R K E T ~emllscom S N A P S H O T P R O D U C T F O R P R C O U C T S N A P S H O T L P R O D U C T - 10 - - T h e model has three cornerstone entities: Business Unit, the part of the organisation for which the plan is being developed; Product, the products or services offered by the unit; Market, the markets in which it operates. Critical Success Factors model the workings of a market by documenting the factors critical to the success of any product in the market, from the consumers’ viewpoint. They are a n objective assessment of h o w the market works, independently of the Business Unit’s presence in it. T h e matching of products to markets is represented by the important Product for Market entity: a product’s score o n the Critical Success Factors relates to this entity. Market Attractiveness Factors model the priorities of the business unit by documenting the factors determining h o w attractive a market is to the unit. Being a subjective assessment of the business unit’s priorities, the criterion for their correctness is the agreement of key executives. T h e matching of markets to business units is represented by the important Involvement In Market entity: a market’s score o n the Market Attractiveness Factors relates to this entity. Time-dependent information is held in Snapshot entities. For each plan, the demonstrator system holds data structures closely based o n the data model. T h e user’s primary m e a n s of manipulating the data is by using forms, which are illustrated in Figure 3. - 11 - Weight IBL Score Competitor 1 1: Quality 20 9 1.8 4 .8 2: Price 1.5 5 .lS 6 .9 3: Differentiation 25 7 1.75 4 1.0 4: Image 20 7 1.4 3 6 5: support 20 9 1.8 6 1.2 [BE(l[ Strength in Market: IBL 7.5 Competitor 1 4.5 Figure 3 Typical forms for data manipulation - - The top form shows current information about the Food Processing market for the fictional International Bearings Limited (IBL) company, which sells bearings into a variety of markets. The bottom form shows the Critical Success Factors defined for this market, with weights to illustrate their relative importance. For example, while price is important in this market, it is less so than several other factors, such as product differentiation and quality, the product’s image, and the engineering support provided. It also shows a score for IBL and its main competitor against these factors, and a weighted average computed by the system, to represent IBL’s overall strength in the market. This is copied to the top form by the system. The Market Attractiveness score on the top form results from a similar weighted average form for the attractiveness of the market against such criteria as the market’s size, growth and profitability. - - 12 - - - - 3.4 The use of techniaues in the data model The “rocky peaks” with which the analysis work started are “textbook” techniques for analysing an organisation’s markets and products, such as the Directional Policy Matrix, which is illustrated below (Figure 4), the Boston and Porter matrices, and so on. These view different aspects of the data model using differing graphical representations, to aid in interpretation of the data. To extend our analogy, the data model thus forms the bridges between the rocky peaks to enable us to navigate the intervening swamp. M Higt j q,, r”‘roacmt c \.I I p&y- E 1 I 0 T T,4 h I &- A T R s F-7. Auruman ‘! \ T I ~nmmiul V ‘. : E \. N ELow s <_r, s High SI-RENGTH IN MARRET Low Quadrant advice: Maintain market position, manage for cash Figure 4 Data presentation to aid understanding - 13 - .- - - The screen snapshot in Figure 4 gives an example of how the demonstrator exploits these features by showing the underlying data presented in the standard formats. The Directional Policy Matrix plots, for each of IBL’s markets, the market attractiveness against IBL’s strength in the market. The size of the circles is proportional to the market’s contribution to IBL’s revenue (though it could have been set to any useful metric). Different circle shadings illustrate the current, forecast and objective situations for the product/market. (In terms of the data model discussed earlier, each circle strictly represents a Product For Market,) The matrix aids in understanding both the situation of an individual product/market, and the balance of the portfolio of products. An example of the matrix’s interpretation is that in all its markets, IBL is moving downwards and rightwards from the current to the forecast situation. This indicates a general weakening of IBL’s position: the matrix illustrates what IBL intends to do about this for the automotive market by maintaining its competitive position while cutting costs where possible. The demonstrator also provides on request standard, “textbook” advice for a product- for-market in a given position on the matrix, as a guide to the planner in setting objectives. For example, for the automotive market, the system advises that the market position (strength in market, and market share) be maintained, but that subject to this the market be managed for cash to fund development of more attractive markets. This is the only case in the demonstrator where it was felt appropriate that the system should take an active role of giving advice, rather than the passive role of presenting information in differing forms to aid the user in interpretation. - 14 - - - _- - The diagram also shows a “gap gauge”, a bar chart showing the financial gap between the business unit’s target revenue and the sum of the individual objectives so far set for the various markets. 3.5 Less structured information: checklists. free text Some parts of the marketing plan were best expressed in text: for example, the business unit’s mission statement, and lists of opportunities and threats. Also, in several areas, marketing expertise was identified that was not formalised beyond free text in the model. Examples are checklists of common critical success factors; assistance with definition of a business unit’s mission statement; and checklists of possible opportunities and threats to consider. This unstructured information was related, however, to specific points in the planning process, or to specific items in the data model. The demonstrator exploited this by making available text windows at appropriate points with icons on the browsers and elsewhere. This was implemented using the NoteCards hypertext system. 4. DEMONSTRATORFEEDBACKANDPROTOTYPE Initiative is with the user The demonstrator leaves the user to decide what to do next. This was liked by the club members, who felt it to be appropriate for this application. Evidence of utility Club members felt an operational system based on the demonstrator’s ideas could be of significant use in the vital process of strategic marketing planning. This is an example of utility being addressed by the clients rather than by the developers. - 15 - - - - - Communication of the nature of the proposed prototype T h e demonstrator served to communicate the nature of the support that would b e offered to a marketing planner by a fuller computer system, to club m e m b e r s a n d to the primary expert, Professor McDonald. With this innovative system, this was difficult to achieve o n paper. Use in specification of prototype T h e demonstrator has b e e n used effectively in discussions with club m e m b e r s to aid with specification of the prototype n o w being developed. 5. R E V I E W O F T H E V A L U E O F A P P L Y I N G E X P E R T S Y S T E M S T O T H E M A R K E T I N G P L A N N I N G P R O C E S S During the 196Os, attention was focussed o n specific problem-solving applications in scientific fields. M a n y successful Expert Systems have b e e n built, including M Y C I N for diagnosing infectious diseases’, a n d P R O S P E C T O R , a system for evaluating geographical locations for possible mineral deposit8. M a n a g e m e n t problems, however, d o not lend themselves to quite the s a m e precise logic as scientific problems. People d o not solve most of life’s problems by mathematical means, but rather by experience, knowledge a n d intuition. Marketing problems are dealt with in the s a m e way, as most of them are logical rather than mathematical, a n d problem-solving knowledge, whilst available, is incomplete. Decision-Support Systems a n d the like have traditionally used hard facts a n d static formulae which, given the correct data, provide correct answers. They belong m o r e naturally to the logical, black-or-white, right-or-wrong world of computers. But m a n a g e r s in the world of marketing deal with uncertainties a n d often with vague concepts. O n e approach to pinning d o w n the basis for such decisions is to attempt to - 16 - - - model the decision-making process as a set of “rules”, or heuristics, that reflect the expert’s o w n knowledge a n d experience about the problem in question. These “rules” are hard to nail d o w n a n d quantify - partly, perhaps, because the expert’s experience enables him to think in terms of shades of grey, “m o r e or less”, a n d “approximately” - a n d partly because rules are not always a n appropriate or sufficient representation. W h e n h u m a n beings find a path through situations that are too complex a n d a m o r p h o u s for the h u m a n mind to handle in a totally conscious, rational, scientific way, it can b e difficult or impossible to elicit the m e a n s by which they d o so. Most people would acknowledge that in virtually any walk of life, the true expert has built u p his expertise largely from experience a n d a n intuitive grasp of problem- solving in the real world, something which is often referred to as the “University of Life”. Indeed, m a n y of the world’s leading business people acknowledge that they o w e their success not to formal business education a n d text books, but to their o w n experience, flair a n d intuitive g o o d judgement. Donald Schon’ describes this p h e n o m e n o n as follows: “Competent practitioners usually know m o r e than they can say. They exhibit a kind of knowing-in-practice, most of which is tacit”. H e cites a n investment banker, w h o makes his decisions based o n 7 0 to 8 0 per cent instinct, a n d only 2 0 to 3 0 per cent calculable rules. This “gut feel” was a major asset to the bank in question. His point is that artistry is not reducible to discernible routines. H e describes scientific rigour as “describable, testable, replicable techniques derived from scientific research, based o n knowledge that is testable, consensual, cumulative a n d convergent”, but then goes o n to argue that m u c h of what passes for scientific m a n a g e m e n t is irrelevant because business problems d o not c o m e well formed. Certainly, most marketing problems are messy a n d indeterminate a n d successful practitioners m a k e judgements using criteria which are difficult to define. M a n y - 17 - - - - academics would decry this as a lack of rigour, and in so doing exclude as non- rigorous much of what successful practitioners actually do. The following quotation from Schon neatly sums up the problems facing not only teachers and researchers of marketing, but, more importantly, the initiators of expert marketing systems : “In the varied topography of professional practice, there is a high, hard ground which overlooks a swamp. On the high ground, manageable problems lend themselves to solution through the use of research-based theory and technique. In the swampy lowlands, problems are messy and confused and incapable of technical solution. The irony of the situation is that the problems of the high ground tend to be relatively unimportant to society at large, however great their technical interest may be, while in the swamp lie the problems of greatest human concern.” Marketing Planning remains one of the last bastions of ignorance in the field of marketing. The benefits ‘of marketing planning are well documented and agreed,8 yet so complicated is the process of marketing planning, and so confusing are the interrelationships between the tools and techniques of marketing planning’, that very few British companies enjoy these benefits, as has been shown by a seminal paper by Greenley lo that reviewed all the major UK empirical research in this area. Indeed, there were as many dysfunctional results from the attempts of companies to initiate marketing planning procedures as there were benefits. The problem to be addressed by Expert Systems in the marketing domain centres around how to take account of the intuitive artistry displayed by experts in situations of complexity and uncertainty in a way that is describable and susceptible to a kind of rigour that falls outside the boundaries of technical rationality. An important - 18 - - /- - .- aspect of this is to identify where a computer system cannot hope to solve a problem on its own, and in such cases, how it can best assist the marketing manager. The question, then, is how an epistemology of practice can be captured and represented in an Expert System. 6. APPROPRIATE TECHNOLOGY IN THIS DOMAIN 6.1 Previous svstems To date, no one to our knowledge has seriously tackled the world of marketing with Expert Systems other than the MSI A D C A D ’ ’ system developed to advise on advertising design. After considering a variety of consumer and environmental factors, advertisers use a combination of empirical research, communication theory, and rules of thumb, to select communication objectives and select appropriate creative approaches. The authors themselves list a number of weaknesses in ADCM, but conclude: “As one advertising executive put it: “it helps us to think a little deeper about the issues we have to consider in developing ads that are both strategically and executionally kound”. Another interesting and relevant conclusion was that most managers, when asked, said they would like to make use of existing theoretical and empirical knowledge of marketing when making decisions. However, few actually did use such knowledge. Expert Systems can bridge this gap by structuring, validating and disseminating marketing knowledge, whilst at a theoretical level, they challenge their creators to understand and critically evaluate the elements of marketing knowledge and their interrelationships. - 19 - - - - 6.2 Problems suitable for exDert svstems In deciding whether marketing planning was a sensible domain for the application of Expert Systems technology, the M S 1 checklist” fits our experience. Four criteria are provided: 8 Are the key relationships in the domain logical rather than arithmetical ? Concepts such as the strength of a product in a market, the attractiveness of a market or product differentiation are clearly not arithmetical - though n u m b e r s m a y b e used to g o o d effect in clarifying thought. S o the answer is “yes” (though not in the sense of mathematical logic). 8 Is the problem domain semi structured rather than structured or unstructured? T h e marketing planning process has both structured elements - for example, the segmentation of a market, the relevant financial information - a n d related unstructured elements - for example, a mission statement or a list of forecasting assumptions. S o the answer is “yes”. 8 Is knowledge in the domain incomplete ? Since marketing planning a n d all its contextual problems remains o n e of the most under-researched areas of marketing, a n d since little has b e e n published about the interrelationships of all the techniques of marketing in systems design, the answer is “yes”. This is in fact the key to the whole project a n d why it was chosen in the first place by the club members. 8 Will problem solving in the domain require a direct interface between the m a n a g e r a n d the computer system ? - 2 0 - - T h e intention is to have operational marketing m a n a g e r s using the system for the production of marketing plans, a n d the system can solve few of the problems by itself, so the answer is “yes”. 6.3 Role of the comauter T h e potential benefits shown by the E X M A R demonstrator are d u e mainly to its assistance with the understanding a n d interpretation of the information entered. T h e e n d result m a y include a marketing plan, but it also includes a n e n h a n c e d a n d readily communicable understanding of the business gained by the marketing planner. These benefits are largely d u e to appropriate a n d varied display of the information. Apart from data presentation, a computer system in this domain can perform the tasks for which computers have traditionally b e e n used: m a n a g i n g data, maintaining constraints between data items like a speadsheet, a n d performing routine calculations. These free u p the user’s thoughts for higher-level problems. Finally, in s o m e cases the computer can b e m o r e pro-active, offering advice, pointing out decisions that g o against conventional wisdom, a n d so on. T h e most appropriate technology for this mix of roles will itself b e a mix. In the case of E X M A R , the software techniques included object-oriented programming, hypertext a n d use of windows-based programming environments, to enable swift development a n d a carefully tailored user interface. W e have not found rule-based representations so far to b e relevant, though they m a y b e in future developments. T o s o m e Expert Systems workers this emphasis o n data presentation a n d low-level data management, as o p p o s e d to sophisticated calculation or reasoning, would constitute s o m e sort of failure. W e consider, however, that the objective of computer systems is to m a k e the combination of user a n d system m o r e effective than the user alone, not to build ‘clever’ computer systems. Even in the classic scientific expert systems such as those quoted earlier, the user interface frequently constituted m o r e of the work, a n d m o r e importantly delivered m o r e of the benefits, than emphasis in the literature would suggest. W e suggest that this applies even m o r e in such ‘soft’ a n d ill-understood areas as marketing planning: the rapid recent progress in the power, a n d price, of the underlying software tools that enable graphical user interfaces to b e provided will enable m o r e such areas to b e tackled effectively in the future. - 2 1 - - - 6.4 Analvsis wroach T h e analysis approach u s e d for E X M A R w a s undogmatic a n d modest: to m o d e l the available expertise with whatever modelling techniques proved m o s t appropriate, starting with the m o s t well-established a n d documented, a n d verified, expertise. “D o n ’t run before you can walk” should not n e e d emphasising: but the early experience of the club shows that perhaps it still does. T h e very term “Expert Systems” has led s o m e to unjustified assumptions not just of the feasibility of building computer systems b a s e d o n expertise, but also of their utility, a n d of the m o s t appropriate modelling a n d system-building tool$. T h e alternative is classic software engineering, with a n e x p a n d e d tool kit of analysis a n d implementation techniques to d r a w u p o n as appropriate. This m a y lead to the question about h o w a n d to what extent the m o d e l a n d demonstrator m a y b e said to incorporate expertise. All aspects of the m o d e l a n d demonstrator can reasonably b e said to b e b a s e d o n expertise: the process, the data model, the m e a n s of presentation of information, the checklists provided, a n d the o n e case w h e r e data-dependent advice is given. T h e system thus takes the “low road” according to B r o w n ’s categorisation discussed earlier. T h e r e is certainly m u c h available (but not necessarily formalisable) expertise that has not b e e n captured: the critical design task has b e e n the effective definition of the boundary between the system a n d the user such that the user is encouraged to think about the issues that the system cannot of itself address. This conforms to the stated E X M A R system objective quoted earlier, of providing assistance for the marketing planning process in such a w a y as to spread knowledge a n d further understanding of the business a n d its markets. 7. C O N C L U S I O N S A n u m b e r of conclusions can b e d r a w n from the E X M A R experience: 6) T h e development of E X M A R shows that it is possible to use Expert Systems methodologies to build support systems in c o m p l e x areas of marketing m a n a g e m e n t , especially if the d o m a i n is well defined, has a large n u m b e r of factors to b e considered a n d relevant expert knowledge is available. (ii) T h e m o r e c o m p l e x a n d a m o r p h o u s the expertise to b e captured, the longer it takes both the expert a n d the knowledge engineer to reach a n acceptable - 22 - - - approximation. It is clear that to develop an Expert System that is of some practical use requires both time and resources of massive proportions. This is supported by the MS1 research paper 11 , which concludes: “There are no shortcuts to building a good Expert System. It takes a considerable amount of skill, patience, and years of effort to develop an Expert System in a new area and get it into the field”. (iii) Expert Systems provide a consistency to human decision making which is valuable, since people tend to forget or ignore knowledge. (iv) EXMAR has generated considerable interest and support among, the major multinational companies that form the club, because it forces them to think deeply and in a structured way about the issues that need to be considered in developing a strategic marketing plan. (VI Expert Systems are useful in helping both academics and practitioners to structure, validate, and use marketing knowledge and to better understand the interrelationships between the elements of marketing. (vi) Tight project control is vital. This view is supported by Mumford13. Many issues need to be considered, such as clear definition of subject matter, availability of inputs, and clear agreement with users on objectives, timescales and resourcing. The close involvement of the EXMAFt club members has been essential in this respect. It has been achieved through an active working party, through agreed quality assurance criteria for each stage of the work, and through the use of a demonstrator. (vii) The potential advantages of Expert Systems in marketing are: m consistent advice - 23 - H secure knowledge bases n making better use of experts H enhanced decision making n improved analysis. (viii) Since we live in an imperfect world, with imperfect problems and imperfect tools, it is unreasonable to expect a perfect Expert System until there are perfect experts and perfect technology. On the other hand, if an Expert System gives better advice than you would have had without it, it is probably worthwhile. In conclusion, it is unlikely that Expert Systems will ever be able to give the same value as real human experts, although clearly they can offer reasonable advice. Nor will they guarantee that you make the right decisions. But they can help you gain a proper perspective of the alternatives. In a sense, Expert Systems will always be a bit like Distance Learning programmes, which can replace a bad teacher, but never a good one. - 24 - / -. r 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. Moutinho L. and Paton R. “Expert Systems : a New Tool in Marketing” Quarterly Review of Marketing, Summer, 1988 Foster E. “Artificial Intelligence” Personal Computing, April, 1985 Cebrzyask “Artifical Intelligence : the goal is to store an expert’s real knowledge on a disk” Marketing New, Vol. 21, No. 5, 1987 Brown J.S. “The low road, the middle road and the high road” in The AI Business, P.H. Winston and K. Pendergast, Eds. MIT Press, Cambridge, Mass 1984 Buchan B. and Shortliffe E. “Rule-based Expert Programs : the MYCIN Experiments of the Stanford Heuristic Programming Project, Reading, MA., Addison- Wesley, 1984 Duda R., Gaschnig J. and Hart P. “Model Design in the Prospector Consultant System for Mineral Exploration” in “Expert Systems in the Micro- electronic Age” ed. Michie D., Edinburgh: Edinburgh University Press, 1979 Schon D. “The Crisis of Professional Knowledge and the Pursuit of an Epistemology of Practice”, Research Paper for the Harvard Business School 75th Anniversary Colloquim on Teaching by the Case Method, April, 1984 McDonald M. “The Theory and Practice of Marketing Planning for Industrial Goods in International Markets”, Cranfield Institute of Technology, PhD., 1984 Leppard J. “Marketing Planning and Corporate Culture - a Conceptual Framework which examines Attitudes in the Context of Marketing Planning” Cranfield Institute of Technology, M.Phil., 1987 Greenley G. “An Exposition into Empirical Research into Marketing Planning”, Journal of Marketing Management, 3.1., July, 1987 Rangaswamy A., Burke R., Wind J. and Eliashberg J. “Expert Systems for Marketing” Marketing Science Institute Working Paper Report, Nos. 87-107, 1988 Bobrow D.G., Mittall S. and Stefik M. “Expert Systems, Perils and Promise”, Commun. A.C.M. 29.9 (Sept. 1986) pp. 880-894 Mumford E. “Designing Computer Based Systems”, University of Wales Review Business and Economics, 3, 1988 - 
work_zehnx3oazrbmhpixswqdorn3cy	----	QBoost: Predicting quantiles with boosting for regression and binary classification Expert Systems with Applications 39 (2012) 1687–1697 Contents lists available at ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a QBoost: Predicting quantiles with boosting for regression and binary classification Songfeng Zheng ⇑ Department of Mathematics, Missouri State University, 901 S. National Ave., Springfield, MO 65897, USA a r t i c l e i n f o a b s t r a c t Keywords: Quantile regression Boosting Functional gradient algorithm Binary classification 0957-4174/$ - see front matter � 2011 Elsevier Ltd. A doi:10.1016/j.eswa.2011.06.060 ⇑ Tel.: +1 417 836 6037; fax: +1 417 836 6966. E-mail address: SongfengZheng@MissouriState.edu In the framework of functional gradient descent/ascent, this paper proposes Quantile Boost (QBoost) algorithms which predict quantiles of the interested response for regression and binary classification. Quantile Boost Regression performs gradient descent in functional space to minimize the objective function used by quantile regression (QReg). In the classification scenario, the class label is defined via a hidden variable, and the quantiles of the class label are estimated by fitting the corresponding quantiles of the hidden variable. An equivalent form of the definition of quantile is introduced, whose smoothed version is employed as the objective function, and then maximized by functional gradient ascent to obtain the Quantile Boost Classification algorithm. Extensive experimentation and detailed analysis show that QBoost performs better than the original QReg and other alternatives for regression and binary classification. Furthermore, QBoost is capable of solving problems in high dimensional space and is more robust to noisy predictors. � 2011 Elsevier Ltd. All rights reserved. 1. Introduction Classical least square regression aims to estimate the condi- tional expectation of the response Y given the predictor (vector) x, i.e., E(Y|x). However, the mean value (or the conditional expecta- tion) is sensitive to the outliers of the data (Koenker, 2005). There- fore, if the data is not homogeneously distributed, we expect the traditional least square regression to give us a poor prediction. The sth quantile of a distribution is defined as the value such that there is 100s% of mass on its left side. Compared to the mean value, quantiles are more robust to outliers (Koenker, 2005). Let I(�) be the indicator function with I(�) = 1 if the condition is true, other- wise I(�) = 0. Let Qs(Y) be the sth quantile of random variable Y. It can be proved (Hunter & Lange, 2000) that Q sðYÞ¼ arg min c EY½qsðY � cÞ�; where qs(r) is the ‘‘check function’’ (Koenker, 2005) defined by qsðrÞ¼ rIðr P 0Þ�ð1 � sÞr: ð1Þ Given training data {(xi, Yi), i = 1, . . . , n}, with predictor vector xi e R d and response Yi e R, let the sth conditional quantile of Y given x be f(x). Similar to the least square regression, quantile regression (QReg) (Koenker, 2005; Koenker & Bassett, 1978) aims at estimating the conditional quantiles of the response given a predictor vector x and can be formulated as ll rights reserved. f �ð�Þ¼ arg min f 1 n Xn i¼1 qsðY i � fðxiÞÞ: ð2Þ Compared to least square regression, quantile regression is ro- bust to outliers in observations, and can give a more complete view of the relationship between predictor and response. Furthermore, least square regression implicitly assumes normally distributed er- rors, while such an assumption is not necessary in quantile regres- sion. Since being introduced in Koenker and Bassett (1978), quantile regression has become a popular and effective approach to statistical analysis with wide applications in economics (Hendricks & Koenker, 1992; Koenker & Hallock, 2001), survival analysis (Koenker & Geling, 2001), and ecology (Cade & Noon, 2003), to name a few. The quantile regression model in Eq. (2) can be solved by linear programming algorithms (Koenker, 2005) or Majorize-Minimize algorithms (Hunter & Lange, 2000). However, when the predictor x is in high dimensional space, the aforementioned optimization methods for QReg might be inefficient. High dimensional problems are ubiquitous in applications such as image analysis, gene se- quence analysis, etc. To the best of our knowledge, the problem of high dimensional predictor is not sufficiently addressed in QReg literature, and this paper proposes a method for QReg which can work in high dimensional spaces. The proposed algorithm for QReg is based on boosting (Freund & Schapire, 1997), which is well known for its simplicity and good performance. The powerful feature selection mechanism of boost- ing makes it suitable to work in high dimensional spaces. Fried- man, Hastie, and Tibshirani (2000) developed a general statistical framework which yields a direct interpretation of boosting as a http://dx.doi.org/10.1016/j.eswa.2011.06.060 mailto:SongfengZheng@MissouriState.edu http://dx.doi.org/10.1016/j.eswa.2011.06.060 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa 1688 S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 method for function estimation, which is a ‘‘stage-wise, additive model’’. Consider the problem of function estimation f �ðxÞ¼ arg min f E½lðY; fðxÞÞjx�; where l(�, �) is a loss function which is typically differentiable and convex with respect to the second argument. Estimating f⁄(�) from the given data, {(xi,Yi), i = 1, . . . , n}, can be performed by minimizing the empirical loss, n�1 Pn i¼1 lðY i; fðxiÞÞ, and pursuing iterative steep- est descent in functional space. This leads us to the generic func- tional gradient descent algorithm (Friedman, 2001; Mason, Baxter, Bartlett, & Frean, 2000). Fig. 1 shows the version summarized in Bühlmann and Hothorn (2007). Many boosting algorithms can be understood as functional gradient descent with appropriate loss function. For example, if we choose l(Y, f) = exp(�(2Y � 1)f), we would recover AdaBoost (Friedman et al., 2000), and L2 Boost (Bühlmann & Yu, 2003) corre- sponds to l(Y, f) = (Y � f)2/2. Motivated by the gradient boosting algorithms (Friedman, 2001; Mason et al., 2000), this paper estimates the quantile regres- sion function by minimizing the objective function in Eq. (2) with functional gradient descent. In each step, we approximate the neg- ative gradient of the objective function by a base function, and grow the model in that direction. This results in the Quantile Boost Regression (QBR) algorithm. In the binary classification scenario, we define the class label via a hidden variable, and the quantiles of the class label can then be estimated by fitting the correspond- ing quantiles of the hidden variable. An equivalent form of the def- inition of quantile is introduced, whose smoothed version is employed as the objective function for classification. Similar to QBR, functional gradient ascent is applied to maximize the objec- tive function, which yields the Quantile Boost Classification (QBC) algorithm. The obtained Quantile Boost (QBoost) algorithms are computationally efficient and converge to local optima. More importantly, they enable us to solve high dimensional problems efficiently. The rest of this paper is organized as follows: in Section 2, we first apply the functional gradient descent to QReg, yielding the QBR algorithm, and then we proceed to propose the approximation Fig. 1. Generic function of the objective function for binary classification and to maximize the objective function with functional gradient ascent in order to obtain the QBC algorithm; Section 3 discusses some computational issues in the proposed QBR and QBC algorithms and introduces implementation details; Section 4 presents the experimental re- sults of the proposed QBR and QBC algorithms on benchmark regression and binary classification datasets, and in-depth discus- sions of the results are also presented in Section 4; finally, Section 5 summarizes this paper with a brief discussion for future research directions. 2. Methods The methods proposed in our research are presented in this sec- tion. We first directly apply the functional gradient descent to the quantile regression model, yielding the quantile boost regression algorithm. We then propose a smooth approximation to the opti- mization problem for the quantiles of binary response, and based on this we further propose the quantile boost classification algo- rithm with some discussions of the related methods. 2.1. Quantile boost regression We consider the problem of estimating quantile regression function in the general framework of functional gradient descent with the loss function lðY; fÞ¼ qsðY � fÞ¼ ðY � fÞIðY � f P 0Þ�ð1 � sÞðY � fÞ: A direct application of the algorithm in Fig. 1 yields the Quantile Boost Regression (QBR) algorithm, which is shown in Fig. 2. Similar to AdaBoost, QBR enables us to select most informative predictors if an appropriate base learner is employed, and this will be demonstrated experimentally in Section 4.1.1. There is a large volume of literature applying boosting to regression problems, for example in Duffy and Helmbold (2002), Freund and Schapire (1997), and Zemel and Pitassi (2001). How- ever, all these methods estimate the mean value of the response, not quantiles. Langford, Oliveira, and Zadrozny (2006) proposed to use classification technique for estimating the conditional al gradient descent. Fig. 2. Quantile boost regression (QBR) algorithm. S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 1689 quantile. For each given quantile value, their method trains a set of classifiers {ct} for a series of t e [0, 1], and the testing stage calcu- lates the average of the outputs of the classifiers. Therefore, com- pared to the proposed QBR algorithm, this method is time consuming. Furthermore, it is not clear how to perform variable selection using the method in Langford et al. (2006). Takeuchi, Le, Sears, and Smola (2006) and Li, Liu, and Zhu (2007) perform quantile regression in Reproducing Kernel Hilbert Spaces (RKHS). Although RKHS might be of very high dimensionality, the models in Takeuchi et al. (2006) and Li et al. (2007) require quadratic pro- gramming which is computationally more expensive than the pro- posed QBR model. Moreover, the models in Takeuchi et al. (2006) and Li et al. (2007) cannot perform variable selection as does QBR. In an independent study which applied the functional gradi- ent method, Kriegler & Berk (2007) weighted the absolute loss function according to overestimates and underestimates, resulting an algorithm similar to QBR. However, the method in (Kriegler & Berk, 2007) uses multi-node regression tree as weak learner, and needs to solve multiple optimization problems in each iteration, therefore, it is more time consuming compared to QBR. 2.2. Quantile boost classification 2.2.1. Predicting quantiles of binary response Consider the following model: Y� ¼ hðxÞþ e and Y ¼ IfY� P 0g; ð3Þ where Y⁄ is a continuous hidden variable, h(x) is the true model for Y⁄, e is a disturb, and Y e {0, 1} is the binary class label for the obser- vation x. The model in Eq. (3) was also considered in Koenker (2005), Kordas (2002, 2006), and Manski (1985). Let Qs(Y ⁄|x) be the sth conditional quantile of Y⁄ given x, and let g(�) be a real monotone increasing function. Clearly PðY� P yjxÞ¼ PðgðY�Þ P gðyÞjxÞ; so it follows that gðQ sðY �jxÞÞ¼ Q sðgðY �ÞjxÞ: ð4Þ Since the indicator function, Iðt P 0Þ, is monotone increasing with respect to t, Eq. (4) indicates that IðQ sðY �jxÞ P 0Þ¼ Q sðIðY � P 0ÞjxÞ¼ Q sðYjxÞ; that is, the conditional quantile function of Y can be obtained by fit- ting the corresponding conditional quantile function of Y⁄ (Koenker, 2005; Kordas, 2006). Let the sth conditional quantile function of the latent variable Y⁄ be f(x, b) with b as the parameter vector, i.e., Q sðY �jxÞ¼ fðx; bÞ: ð5Þ It follows that the conditional quantile function of the binary vari- able Y can be modeled as Q sðYjxÞ¼ Iðfðx; bÞ P 0Þ: ð6Þ From the relation Y ¼ I ðY� P 0Þ, we have PðY ¼ 1jxÞ¼ PðY� P 0jxÞ; ð7Þ while Eq. (5) implies that PðY� P fðx; bÞjxÞ¼ 1 � s: ð8Þ Thus, if f(x, b) = 0, combining Eqs. (7) and (8) yields PðY ¼ 1jxÞ¼ PðY� P fðx; bÞjxÞ¼ 1 � s: If f(x, b) > 0, PðY ¼ 1jxÞ > PðY� P fðx; bÞjxÞ¼ 1 � s: In summary, we have the following relation: PðY ¼ 1jxÞ¼<> 1 � s for fðx; bÞ¼<> 0; ð9Þ which is an inequality of the posterior probability of the binary class label given the predictor vector. Hence, if we make decision with cut-off posterior probability value 1 � s, we need to fit the s quantile regression function of the response. Once the model is fitted, i.e., the parameter vector b is estimated as b, we can make prediction by Ŷ ¼ Iðfðx; bÞ P 0Þ; where Ŷ is the predicted class label for the predictor vector x . 2.2.2. Quantile boost for binary classification To fit the model for the sth quantile of Y, i.e., to estimate the parameter vector b in Eq. (6), in a way similar to the method 1690 S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 applied to QBR, the estimated parameter vector b can be obtained by solving: b ¼ arg min b LðbÞ¼ Xn i¼1 qs½Y i � Iðfðxi; bÞ P 0Þ� ( ) : ð10Þ In Appendix A, it is proved that the above minimization problem is equivalent to a maximization problem: b ¼ arg max b SðbÞ¼ Xn i¼1 ½Y i �ð1 � sÞ�Iðfðxi; bÞ P 0Þ ( ) : ð11Þ However, the function S(b) is not differentiable, because of the use of the indicator function Iðfðx; bÞ P 0Þ. To apply gradient based optimization methods, we replace Iðfðx; bÞ P 0Þ by its smoothed version, and solve b ¼ arg max b Sðb; hÞ¼ Xn i¼1 ½Y i �ð1 � sÞ�K fðxi; bÞ h � �( ) ; ð12Þ where h is a small positive value, and K(t) is smoothed version of the indicator function, Iðt P 0Þ, with the following properties (Kordas, 2006): KðtÞ > 0; 8t 2 R; lim t!1 KðtÞ¼ 1; lim t!�1 KðtÞ¼ 0: In this paper, we take K(�) as the standard normal cumulative distri- bution function UðzÞ¼ Z z �1 /ðtÞdt with /ðzÞ¼ 1ffiffiffiffiffiffiffi 2p p e� z2 2 : ð13Þ Let each term in the objective function (12) be lðY; fÞ¼ ½Y �ð1 � sÞ�K f =hð Þ: Following the general steps of functional gradient ascent, we obtain the Quantile Boost Classification (QBC) algorithm as shown in Fig. 3. 2.2.3. Relationship between Eq. (11) and error rate Maximizing Eq. (11) has a close relation to the performance of the final classifier Iðfðx; bÞ P 0Þ. Let each term in Eq. (11) be Fig. 3. Quantile boost classifi Si ¼ ½Y i �ð1 � sÞ�Iðfðxi; bÞ P 0Þ¼ sIðY i ¼ 1ÞIðfðxi; bÞ P 0Þ �ð1 � sÞIðY i ¼ 0ÞIðfðxi; bÞ P 0Þ; ð14Þ then Eq. (11) becomes SðbÞ¼ s Xn i¼1 IðY i ¼ 1ÞIðfðxi; bÞ P 0Þ �ð1 � sÞ Xn i¼1 IðY i ¼ 0ÞIðfðxi; bÞ P 0Þ ¼ sTP �ð1 � sÞFP ¼ sðPos � FNÞ�ð1 � sÞFP ¼ sPos �½sFN þð1 � sÞFP�; ð15Þ where TP, FP, and FN are the true positive, false positive, and false negative numbers of the classifier Iðfðx; bÞ P 0Þ, respectively; Pos is the number of positive training examples. Eq. (15) implies that maximizing Eq. (11) is equivalent to min- imizing a convex combination of FP and FN, sFN + (1 � s)FP. If s = 0.5, this is equivalent to minimizing the training error number, whereas if s > 0.5, the minimization puts more weight on FN, while FP is more emphasized if s < 0.5. This analysis is consistent with Eq. (9): if s > 0.5, the classifier makes decision at posterior cut-off probability 1 � s which is less than 0.5, and this will make FN small; similarly, for s < 0.5, the posterior cut-off probability will be greater than 0.5, which will lower the FP. 2.2.4. Related works Kordas et al. (2002, 2006) proposed binary quantile regression for the purpose of classification using quantiles. However, in binary QReg, the simulated annealing algorithm is employed to perform the optimization task. Although a local optimum is guaranteed, simulated annealing is well known for its expensive computation, and it is usually difficult to tell when the algorithm converges. While QBC is based on gradient ascent, it yields a local maximum and converges fast. Due to the expensive computation of simulated annealing, binary QReg can only work in very low dimensional spaces. However, in application, we frequently face hundreds or even thousands of predictors, and it is often desired to find out the informative predictors. Clearly, in this case, binary QReg is not applicable. On the contrary, QBC is designed to work in high cation (QBC) algorithm. S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 1691 dimensional spaces, and it enables us to select the most informa- tive variables by using certain types of base learner. Hall, Titterington, and Xue (2009) proposed a median based classifier which works in high dimensional space. For a given pre- dictor vector, (Hall et al., 2009) compares the L1 distances from the new predictor vector to the component-wise median vectors of the positive and negative examples in the training set, and assigns class label as the class with the smaller L1 distance. Although com- putationally efficient, this simple nearest-neighbor-like algorithm cannot perform variable selection as the proposed QBC algorithm. Mease, Wyner, and Buja (2007) combines AdaBoost and sampling techniques to predict the class quantiles, but the sampling tech- nique might cause under utilization of the available information. The method proposed in Langford et al. (2006) can also be applied to classification tasks, however it is computationally more expen- sive than QBC, and it is not clear how to select most informative predictors as well. 3. Calculation In the proposed QBR and QBC algorithms (as shown in Figs. 2 and 3, respectively), in a manner similar to the work in Friedman (2001), let the base procedure be h(x, a) with a being a parameter vector. The third step of QBR and QBC can be performed by an or- dinary least square regression: am ¼ arg min a Xn i¼1 ½Ui � hðxi; aÞ� 2 ; hence the function g[m](x) = h(x, am) may be regarded as an approx- imation to the negative gradient by the base procedure for QBR, and in the QBC algorithm, the function g[m](x) = h(x, am) is an approxi- mation to the gradient. In step 4 of QBR, the step size factor can be determined by a line search algorithm (e.g., Fibonacci search or Golden section search (Walsh, 1975, chap. 3)) as: gm ¼ arg minc Xn i¼1 qs½Y i � f ½m�1�ðxiÞ� cg½m�ðxiÞ�: For QBC, the line search algorithm for step size is gm ¼ arg maxc Xn i¼1 ½Y i �ð1 � sÞ�K f ½m�1�ðxiÞþ cg½m�ðxiÞ h � � : Alternatively, for simplicity, in each iteration, we could update the fitted function f[m�1](�) by a fixed, but small, step in the negative gradient direction for QBR and in the gradient direction for QBC. It is verified in Bühlmann and Hothorn (2007) that the size of gm is less important as long as it is sufficiently small. A smaller value of fixed step size gm typically requires a larger number of boosting iterations and thus more computing time, while the predictive accuracy has been empirically found to be good when choosing gm ‘‘sufficiently small’’, e.g., gm = 0.1 (Friedman, 2001). To balance the predictive accuracy and computational burden, Bühlmann and Hothorn (2007) suggested a step size of 0.1 be chosen. By simula- tions, (Friedman, 2001) showed that the performance is similar, if the step size parameter is less than 0.125. Thus, in our experiments, the step size parameter of QBR and QBC is fixed at gm = 0.1 for all m, as suggested by Bühlmann and Hothorn (2007), Friedman (2001). In our implementation of the QBC algorithm, the standard nor- mal cumulative distribution function given in Eq. (13) was used as an approximation of the indicator function with h = 0.1. From our experience, QBC is not sensitive to the value of h as long as h < 0.5. The proposed QBR and QBC algorithms, the median classifier in Hall et al. (2009), and the AdaBoost face detector (Viola & Jones, 2001) were implemented using the programming language MAT- LAB�; all other algorithms including the original quantile regression, L1 norm quantile regression (Li & Zhu, 2008; Wu & Liu, 2009), Ada- Boost (Freund & Schapire, 1997), LogitBoost (Friedman et al., 2000), and L2-Boost (Bühlmann & Yu, 2003) were implemented in statistics programming language R. All the experiments were con- ducted on an ordinary PC with 3.0 GHz CPU and 3.25 GB memory. 4. Results and discussion We applied the proposed QBR and QBC algorithms on extensive regression and pattern classification problems, and this section will present the experimental results with in-depth discussions. 4.1. Experiments on regression problems This subsection tests the proposed QBR algorithm on various regression problems with benchmark datasets in Section 4.1.1, and compares QBR to L1 norm quantile regression (Li & Zhu, 2008; Wu & Liu, 2009) on simulated dataset in Section 4.1.2. 4.1.1. Results of QBR on benchmark datasets Since the purpose of this paper is to propose an alternative algo- rithm to the original QReg, we choose only to compare the perfor- mance of QBR to that of the original QReg. Five regression datasets from UCI machine learning repository were used in our experi- ment: Red Wine Quality, White Wine Quality, Forest Fire, Concrete Slump, and Concrete Compressive; see Table 1 for the basic infor- mation of the datasets. The predictor and the response variables were normalized to be in [�1, 1]. To make a fair comparison be- tween QBR and QReg, we used simple linear regression with only one predictor as base procedure in QBR. In evaluation, we use the sum of the ‘‘check losses’’ on the testing set as the error measure which is defined by LðsÞ¼ XNtest i¼1 qsðY i � f̂ðxiÞÞ; where Ntest is the size of the testing set, and f̂ðxiÞ is the predicted sth quantile at xi by the original QReg or QBR. By the definition of quan- tile, a smaller value of L(s) gives an estimated quantile closer to the true value. For each dataset, we randomly select 80% of the examples as training set and use the remaining 20% as the testing set. For three s values (0.25, 0.5, and 0.75), QBR and QReg are separately trained and evaluated on these subsets. The partition-training-testing pro- cedure is repeated 500 times. The means and the standard devia- tions of the 500 ‘‘check losses’’ are reported in Table 2. It is readily observed that in all of the conducted experiments, QBR uni- formly achieves smaller average check loss compared to QReg, which indicates that QBR estimates more accurately. To statistically compare the results, we conduct a z-test with H0 : LQBR � LQReg P 0 vs: H1 : LBQR � LQReg < 0; ð16Þ where LQBR and LQReg are the check losses of QBR and QReg, respec- tively. By a standard z-test procedure (please refer to some elemen- tary statistics textbooks, e.g., Navidi, 2010, chap. 6), we obtain the p-values for each pair of comparison, as listed in Table 3. We note that on 11 out of 15 comparative experiments, the p-values are less than 2%, which indicates that in most cases, QBR performs signifi- cantly better than QReg. The reason might be that in QBR, we implicitly regularize the regression coefficients by applying a small learning rate, i.e., using small step size gm. In applications, variable selection is often needed, since it helps us get a simpler model, making the model easier to interpret/ understand and identifying informative variables. Although the relevant variables are usually unknown, we none the less prefer a simpler model which is capable of deleting more variables: Furthermore, a simpler model is more time efficient in making Table 1 Basic information about the five datasets used for regression problems. Dataset Red wine White wine Forest fire Concrete slump Concrete comp. Size 1599 4898 517 103 1030 Dim. 11 11 12 7 9 Table 2 The comparison between QReg and QBR on the five datasets from UCI machine learning repository. The values listed are the mean values of the check loss of the 500 runs and the standard deviations are listed in parentheses. Dataset s = 0.25 s = 0.50 s = 0.75 QReg QBR QReg QBR QReg QBR Red wine 19.71 (1.11) 19.41 (1.11) 24.30 (1.06) 24.07 (1.09) 19.40 (0.81) 19.16 (0.79) White wine 62.40 (1.84) 62.23 (1.91) 78.69 (1.78) 78.60 (1.92) 62.03 (1.50) 61.78 (1.64) Forest Fire 4.97 (0.54) 4.93 (0.54) 10.04 (0.82) 9.62 (0.78) 9.56 (0.92) 9.14 (0.73) Conc. slump 0.74 (0.14) 0.71 (0.12) 1.00 (0.17) 0.92 (0.16) 0.95 (0.19) 0.84 (0.17) Conc. comp. 15.02 (0.77) 14.91 (0.79) 18.21 (1.08) 18.17 (1.00) 13.63 (0.83) 13.52 (0.77) Table 3 The p-values for the hypothesis testing problem in Eq. (16) on the five datasets for each quantile. 0+ indicates the value is less than 0.0001. Dataset Red wine White wine Forest fire Concrete slump Concrete comp. s = 0.25 0+ 0.0769 0.1132 0+ 0.0158 s = 0.50 0+ 0.2054 0+ 0+ 0.2596 s = 0.50 0+ 0.0044 0+ 0+ 0.0188 Table 4 Comparison between QReg and QBR on variable selection on concrete slump data. The experiment was repeated 500 times. The listed are the average number of selected noisy variables (N.V.), the mean of check loss error with the standard deviation in parentheses. Method s = 0.25 s = 0.50 s = 0.75 # of N.V. Error # of N.V. Error # of N.V. Error QReg 8.13 0.99 (0.16) 6.97 1.38 (0.21) 9.69 1.28 (0.24) QBR 1.04 0.97 (0.19) 0.63 1.14 (0.21) 0.98 1.27 (0.30) 1692 S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 prediction. In our experiments, both QReg and QBR obtain a linear function of the predictors. Since the predictors and response are normalized to be in [�1, 1], it makes sense if we delete the variables, when their coefficients are too small, thus performing variable selection. Twenty noisy predictors are added to the Con- crete Slump data, each of which is generated uniformly at random from [�1, 1]. Thus, in the noisy dataset, only a small fraction (7 out of 27) of the variables are informative. Then we repeat the above experiment. In the linear model obtained, we calculate the sum of the absolute values of all the coefficients. If for any predictor, the absolute value of its estimated coefficient is less than 1% of the total sum, it is trimmed. We calculate the average number of the selected noisy predictors of QReg and QBR from the 500 runs. The means and standard deviations of the check losses on testing set are also calculated. Table 4 summarizes the results, from which it is evident that for each s value, compared to the original QReg, QBR selects far fewer noisy predictors while achiev- ing smaller mean error. This experiment shows that compared to QReg, QBR is more robust to noisy predictors. The procedure of QBR enables us to use other forms of base lear- ner, for example, regression trees (Friedman, 2001; Friedman et al., 2000), and this provides us flexibility in certain applications. Con- trarily, it is not clear how to make use of other forms of regression in the framework of the original QReg. 4.1.2. Comparing to L1 norm quantile regression QBR belongs to the stage-wise additive model, which has a close connection to L1 constrained models (Efron, Hastie, Johnstone, & Tibshirani, 2004). Recently, L1 quantile regression (L1-QReg) (Li & Zhu, 2008; Wu & Liu, 2009) was proposed, which imposes L1 constraint on the coefficient vector in linear quantile regression. Therefore, it is natural to investigate the relationship between QBR and L1-QReg. Given data {(xi, Yi), i = 1, . . . , n}, assuming the use of the linear model b0 + b Tx for estimating the sth quantile of the response, the L1-QReg can be formulated as min b0;b Xn i¼1 qsðY i � b0 � b T xiÞ; subject to jb1jþ �� �þ jbpj6 s: The solution of L1-QReg depends on the penalty parameter s, and changing the parameter s gives us the evolution of the solution, which is called the solution path (Efron et al., 2004; Li & Zhu, 2008). Similarly, in the QBR algorithm with simple linear regression as base procedure, we can observe the evolution of the regression coefficients with the iteration number. The relationship between L1-QReg and QBR was investigated by experimentation with simulated data generated from the model Y ¼ bT x þ e; where b = (3, �1.5, 0, 0, 2, 0, 0, 0)T, the error term e follows standard double exponential distribution, and x e R8 with each component xi � Nð0; 1Þ and x1, . . . , x8 are independent. Five hundred training examples and 10,000 testing observations were generated from the same model. Define the closeness measure between the esti- mated model and the true model as P8 i¼1jbi � b̂ij, where b̂1; � � � ; b̂8 are the estimated regression coefficients. S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 1693 Fig. 4 shows the solution paths and the check losses of the L1- QReg and QBR using simple linear regression as the base proce- dure. Firstly, we observe that both L1-QReg and QBR select the informative predictors, i.e., x1, x2, and x5, and all other predictors have coefficients very close to zero. The final coefficients estimated by L1-QReg and QBR are listed in Table 5. The two estimates are clearly very similar with the QBR solution slightly closer to the true model. The closeness measures are 0.3813 for L1-QReg and 0.3184 for QBR, respectively. Secondly, it is observed that the shape of the solution paths are strikingly similar, which means that each of the solutions of L1-QReg corresponds to a solution of QBR with certain iterations. Thirdly, the average check loss curves show that the trends in the check loss are very similar. The final average check losses are 0.5142 for L1-QReg and 0.5132 for QBR. Fig. 4. Solution paths and the average check losses on testing data of L1-QReg and QBR. InP8 i¼1jb̂ij. Table 5 The final estimated regression coefficients by L1-QReg and QBR. Coefficients b̂1 b̂2 b̂3 b̂4 L1-QReg 2.9633 �1.5773 0.0611 �0.0 QBR 2.9619 �1.5749 0.0516 �0.0 This comparison demonstrates that the L1-QReg and the pro- posed QBR have similar properties and similar performances in variable selection and prediction accuracy. Despite the similarity when QBR employs simple linear regression as the base learner, as previously mentioned, QBR enjoys the flexibility of being able to use other forms of base learners, while the L1-QReg in Li and Zhu (2008), Wu and Liu (2009) can only be used for linear model. 4.2. Experiments on binary classification problems In this subsection, three experiments were conducted to study the performance of the proposed QBC algorithm. In the first two experiments, we made decision at the cut-off posterior probability 0.5, and therefore we fit QBC with s = 0.5. Since QBC is a boosting the plots for L1-QReg, the x-axis is the sum of the absolute values of the solution, i.e., b̂5 b̂6 b̂7 b̂8 403 1.9287 0.0314 �0.0119 0.0147 283 1.9318 0.0189 �0.0004 0.0142 Table 7 The testing errors of different algorithms on the credit score dataset. Data L2Boost (%) AdaBoost (%) LogitBoost (%) QBC (%) Median classifier (%) Clean 28.68 29.99 28.81 28.50 35.04 Noisy 31.92 30.05 32.68 28.55 38.94 1694 S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 procedure, for the comparative purpose, we, for the most part, con- sidered only several of the most popular boosting based classifiers. These include AdaBoost (Freund & Schapire, 1997), LogitBoost (Friedman et al., 2000), and L2-Boost (Bühlmann & Yu, 2003). We also tested the Median Classifier (Hall et al., 2009) and compared the performance. The third experiment compared the performance of AdaBoost to those of QBC, with s = 0.5 and s = 0.25 on face detec- tion problem. The role of s in QBC algorithm was also illustrated. 4.2.1. Results on credit score data We compare the result of QBC to that of the binary QReg (Kor- das et al., 2002) on the German bank credit score dataset which is available from UCI machine learning repository. The dataset is of size 1000 with 300 positive examples, each has 20 predictors nor- malized to be in [�1, 1]. In Kordas et al. (2002), only 8 predictors were preselected to fit the binary QReg due to the expensive sim- ulated annealing algorithm it employs in the optimization stage. Without preselecting variables, we randomly selected a training set of size 800 from the dataset and evaluated the learned QBC clas- sifier on the other 200 examples. The QBC training algorithm was ran for 100 iterations using simple linear regression with only one pre- dictor as base learner. It is therefore fair to compare the performance of QBC to that of the binary QReg. The splitting-training-testing pro- cess was repeated 500 times, and the mean training and testing error rates were reported in Table 6, which also lists the performance of the binary QReg from Kordas et al. (2002). It is clear that QBC outper- forms binary QReg on both training and testing sets. This is due to the efficient computation of QBC, which allows the algorithm to ex- plore more predictors and thus select more informative ones. Fig. 5 shows the training error and the average objective func- tion (see Eq. (12)) as the QBC training algorithm proceeds. As we can see, the objective function increases monotonically and mean- Table 6 Comparison between binary QReg and QBC on credit score dataset. ‘‘Clean’’ means the original dataset, and ‘‘Noisy’’ means 20 noisy predictors are added to the dataset. Dataset QBC Binary QReg Training (%) Testing (%) Training (%) Testing (%) Clean 19.84 24.47 21.9 26.5 Noisy 22.53 25.64 NA NA Fig. 5. The average objective function and the training error for a particular run of QBC on the credit score dataset. The training error is scaled by 0.5 to make the curves comparable. while, the training error decreases. This is consistent with our the- oretical analysis in Section 2.2.3. We can also see that the training algorithm converges with about 50 iterations. To test the variable selection ability of QBC, twenty noisy pre- dictors were generated from the uniform distribution on [�1, 1] and were added to the original dataset. The above procedure was repeated 500 times. Table 6 also lists the average training and test- ing errors of QBC with, on average, only one noisy variable selected. Our result demonstrates that QBC performs only slightly worse on noisy data, indicating that QBC is robust to noise. Due to the expen- sive computation of binary QReg, its performance on noisy data is not provided. Unlike binary QReg, QBC enjoys the flexibility of using other forms of weak learners. We also compared QBC to the alternatives mentioned at the beginning of Section 4.2 on the credit score data- set with and without noisy predictors. All the boosting based algo- rithms used the decision stump (Dettling & Bühlmann, 2003; Viola & Jones, 2001) as the base procedure for fair comparison, and all algorithms were ran 500 times on sets of 800 training and 200 test- ing examples, all randomly selected. The average testing error rates are listed in Table 7, and show that compared to the alternative methods, QBC achieves the best performance on both clean and noisy data. Again, we observe that QBC deteriorates only slightly on the noisy data, verifying its robustness to noisy predictors. 4.2.2. Results on gene expression data We compare QBC to the alternative methods on three publicly available datasets in bioinformatics (Dettling & Bühlmann, 2003): Estrogen, Nodal, and Colon. In these datasets, each predictor is the measure of the gene expression level, and the response is a binary variable. Although we can access thousands of genes, due to the high cost of experiment, we usually can only get a very limited number (50–100) of examples. See Table 8 for the sizes and dimensions of these datasets. The high dimensionality makes binary QReg in Kor- das et al. (2002, 2006) unaffordable. All the boosting-based algo- rithms used the decision stump as base learner for fair comparison. Since all the datasets have small size n, the leave-one-out (LOO) cross validation is carried out to estimate the classification accu- racy. That is, we put aside the ith observation and trained the clas- sifier on the remaining (n � 1) data points. We then applied the learned classifier to get Ŷ i, the predicted class label of the ith obser- vation. This procedure is repeated for all the n observations in the dataset, so that each one is held out and predicted exactly once. The LOO error number (NLOO) and rate (ELOO) are determined by NLOO ¼ Xn i¼1 IðŶ i – Y iÞ and ELOO ¼ 1 n Xn i¼1 IðŶ i – Y iÞ: Table 8 summarizes the LOO classification error numbers and rates of the considered classifiers on the three datasets, with the best per- formance marked in bold. From Table 8, it is readily seen that QBC has the best LOO performance. It can also be observed that the Med- ian Classifier (Hall et al., 2009) is not stable – sometimes the perfor- mance is very good, sometimes the performance is very poor. In LOO cross validation, the training and testing sets are highly unbalanced, which will affect the evaluation result. To provide more thorough results, we have also conducted 5-fold cross valida- tion (5-CV), in which each dataset was randomly partitioned into Table 8 The leave-one-out error numbers and rates of the considered algorithms on the three gene expression datasets. For each algorithm, the misclassification numbers are provided and the error rates are listed in parentheses. The best performances are displayed in bold. The size and dimension of each dataset are also given. Dataset Size Dim L2-Boost AdaBoost LogitBoost QBC Median classifier Estrogen 49 7129 4 (8.16%) 5 (10.20%) 6 (12.24%) 4 (8.16%) 6 (12.24%) Colon 62 2000 10 (16.13%) 11 (17.74%) 10 (16.13%) 9 (14.52%) 9 (14.52%) Nodal 49 7129 11 (22.45%) 12 (24.49%) 9 (18.37%) 8 (16.33%) 18 (36.73%) Table 9 The 5-fold cross validation mean error rates and the standard deviations (in parentheses) of the considered algorithms on the three gene expression datasets. The best mean error rates for a dataset are displayed in bold. Dataset L2-Boost (%) AdaBoost (%) LogitBoost (%) QBC (%) Median classifier (%) Estrogen 21.11 (10.91) 17.11 (10.50) 15.11 (12.87) 13.33 (9.30) 13.38 (11.04) Colon 24.62 (5.95) 21.41 (7.68) 21.41 (7.68) 21.67 (9.50) 14.58 (9.20) Nodal 31.78 (17.67) 24.89 (8.52) 20.44 (13.67) 20.00 (9.30) 42.84 (15.73) Fig. 6. On the face detection problem, the testing error rates of AdaBoost, QBC with s = 0.5, and QBC with s = 0.25. S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 1695 five parts of roughly equal size. In every experiment, one part was used as the testing set, and the other four parts were used as the training set. Table 9 lists the mean and the standard deviation of the error rates for the five runs of each algorithm on every dataset. We observe from Table 9 that QBC yields the best performance on two out of the three datasets. On the Colon dataset, QBC performs better than L2-Boost, similar to LogitBoost and AdaBoost, but worse than the Median Classifier. However, due to the instability of the Median Classifier, we can still safely conclude that QBC performs better than or similar to several of the alternatives. 4.2.3. Face detection Face detection is a classical problem in computer vision, and many machine learning algorithms have been successfully applied to face detection, e.g., Support Vector Machines (Osuna, Freund, & Girosit, 1997), and AdaBoost (Viola & Jones, 2001). We applied the proposed QBC algorithm to a dataset with 3000 face and 3000 non- face images, each of size 16 � 16. The training set includes 1500 face and 1500 nonface images randomly selected from the whole dataset and the rest of the images were used for testing. Emulating (Viola & Jones, 2001), we extracted different types of Haar features from the images, taking a minimum rectangular size of 4 by 4 (for example: 4 by 5, 8 by 4, 10 by 11), resulting in a total of around 6000 features. The AdaBoost face detector follows (Viola & Jones, 2001), and QBC employs regression stump (Torralba, Murphy, & Freeman, 2004) as the weak learner since it is similar to the weak learner used in Viola and Jones (2001). We first compare the classification error rates of AdaBoost to that of QBC with s = 0.5 and show the testing error curves in Fig. 6. It is observed that AdaBoost slightly outperforms QBC with s = 0.5. To have a closer look, we plot out the curves of false posi- tive rates and false negative rates for the two methods in Fig. 7. It is clear that after 40 iterations, AdaBoost and QBC (with s = 0.5) have similar false positive rates, but AdaBoost has a smaller false nega- tive rate, and this is the reason why AdaBoost has a slightly better performance in total error rate. We further notice in Fig. 7 that the false positive rates are much higher than the false negative rates. Therefore, on this dataset, in order to improve the classification accuracy of QBC, we have to de- crease the false positive rate. According to the theoretical analysis in Section 2.2.3, if s < 0.5, the algorithm will pay more attention to decreasing the false positive rate. As such, we fit another QBC clas- sifier with s = 0.25, and Fig. 7 also shows the corresponding false positive and false negative rates. It is observed that although the false negative rate of QBC with s = 0.25 is a little bit higher than those of QBC with s = 0.5 and AdaBoost, the false positive rate is significantly lower than those of QBC with s = 0.5 and AdaBoost. As a consequence, the overall testing error rate is significantly dropped, as shown in Fig. 6. We further notice that the perfor- mance curves of AdaBoost are much more erratic than those of QBC, which implies that the performance of AdaBoost is less stable. There are modifications to AdaBoost given in literature for bal- ancing FP and FN, e.g., in Masnadi-Shirazi and Vasconcelos (2007) and Sun, Kamel, Wong, and Wang (2007). This experiment was in- tended to demonstrate the role of s in the QBC algorithm, thus we choose not to compare QBC with different s to different versions of AdaBoost for imbalanced penalties. As a comparison to AdaBoost and QBC, the median based classi- fier reported in Hall et al. (2009) was also tested on the face detec- tion problem. The test was repeated 20 times, since this algorithm is time efficient. In each repetition, we randomly split the whole dataset so that both the training set and testing set include 1500 face and 1500 nonface images. The mean error rate of the 20 tests is 14.42%, the mean false positive rate is 21.78%, and the mean false negative rate is 7.07%. Comparing these numbers to Figs. 6 and 7, we see that the performance of the Median Classifier in Hall et al. (2009) is much worse than those of AdaBoost and QBC. Fig. 7. On the face testing set, false positive and false negative rates of AdaBoost, QBC with s = 0.5, and QBC with s = 0.25. Note that the y-axis have different scales in the two figures. 1696 S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 5. Conclusions Motivated by the gradient boosting framework, this paper ap- plies functional gradient descent to fit quantile regression (QReg), resulting Quantile Boost Regression (QBR) algorithm. In binary classification problems, the class label is defined via a hidden var- iable, and predicting the quantiles of the binary response is re- duced to predicting the corresponding quantiles of the hidden variable. An equivalent form of the definition of quantile was intro- duced, and its smoothed version was further employed as the objective function. Functional gradient ascent was applied to max- imize the objective function for fitting the classifier, yielding the Quantile Boost Classification (QBC) algorithm. The proposed Quan- tile Boost (QBoost) algorithms yield local optima, and more impor- tantly, they enable us to solve problems in high dimensional spaces and to select informative variables/features. The QBoost algorithms were tested extensively on various regression and binary classification problems with benchmark datasets. On most of the regression experiments, QBR performs sig- nificantly better than the original QReg, in terms of check loss func- tion. Moreover, the comparative experiment on noisy data indicates that QBR is more robust to noise. The comparative result on a credit score dataset shows that QBC outperforms binary QReg (Kordas et al., 2002) and is more robust to noisy predictors. The experiments on credit score dataset and gene expression data analysis demon- strate that QBC performs better than or similar to other alternatives in terms of leave-one-out and five fold cross validation error rates. QBC was compared to the popular AdaBoost classifier (Viola & Jones, 2001) on face detection problem and we demonstrated the role of the desired s value in QBC algorithm. On face detection and gene expression data analysis, binary QReg (Kordas et al., 2002) is not applicable due to the high dimensionality of the datasets. The current version of QBC is for two-class problems, and we plan to develop the multi-class version of QBC by reducing it to a series of two-class tasks, for example, by one-versus-all (Freund and Schapire, 1997) or Error-Correcting Output Codes (Dietterich & Bakiri, 1995). A more thorough comparison of the proposed QBR/QBC to other quantile based regression/classification algo- rithms on more datasets is also desired. Acknowledgments The author gratefully acknowledges the distinction and support provided by Sociedad Mexicana de Inteligencia Artificial (SMIA) and the 9th Mexican International Conference on Artificial Intelligence (MICAI-2010) in order to enhance, improve, and publish this work. The author also thanks the reviewers of MICAI and Journal of Expert Systems with Applications for their constructive comments and would like to extend his gratitude to Prof. Alejandro Peña-Ayala for his excellent work in coordinating the preparation, the review- ing, and the revising of the manuscript. Prof. Shelby J. Kilmer pro- vided a great deal of help with the English in this manuscript. This work was partially supported by 2011 Summer Faculty Fellowship of Missouri State University. Appendix A. The equivalence between Eqs. (10) and (11) In Eq. (10), let Li ¼ qsðY i � Iðfðxi; bÞ P 0ÞÞ¼ qsðY i � 1Þ; if fðxi; bÞ P 0 qsðY iÞ; if fðxi; bÞ < 0 � ; while qsðY iÞ¼ sIðY i ¼ 1Þ and qsðY i � 1Þ¼ ð1 � sÞIðY i ¼ 0Þ: Thus, Li ¼ð1 � sÞIðY i ¼ 0ÞIðfðxi; bÞ P 0Þþ sIðY i ¼ 1ÞIðfðxi; bÞ < 0Þ ¼ ð1 � sÞIðY i ¼ 0ÞIðfðxi; bÞ P 0Þþ sIðY i ¼ 1Þ½1 � Iðfðxi; bÞ P 0Þ� ¼�sIðY i ¼ 1ÞIðfðxi; bÞ P 0Þþð1 � sÞIðY i ¼ 0ÞIðfðxi; bÞ P 0Þ þ sIðY i ¼ 1Þ¼�Si þ sIðY i ¼ 1Þ; where Si is defined in Eq. (14). Therefore, Eq. (10) reads, LðbÞ¼ Xn i¼1 Li ¼ Xn i¼1 ½�Si þ sIðY i ¼ 1Þ� ¼�SðbÞþ s Xn i¼1 IðY i ¼ 1Þ ¼�SðbÞþ sPos; ðA:1Þ where Pos is the number of positive examples, which is a constant. Eq. (A.1) shows the equivalence between Eqs. (10) and (11). References Bühlmann, P., & Hothorn, T. (2007). Boosting algorithms: Regularization, prediction and model fitting. Statistical Science, 22, 477–505. Bühlmann, P., & Yu, B. (2003). Boosting with the L2-loss: Regression and classification. Journal of the American Statistical Association, 98, 324–340. Cade, B. S., & Noon, B. R. (2003). A gentle introduction to quantile regression for ecologists. Frontiers in Ecology and the Environment, 1(8), 412–420. Dettling, M., & Bühlmann, P. (2003). Boosting for tumor classification with gene expression data. Bioinformatics, 19, 1061–1069. S. Zheng / Expert Systems with Applications 39 (2012) 1687–1697 1697 Dietterich, T. G., & Bakiri, G. (1995). Solving multiclass learning problems via error- correcting output codes. Journal of Artificial Intelligence Research, 2, 263–286. Duffy, N., & Helmbold, D. (2002). Boosting methods for regression. Machine Learning, 47, 153–200. Efron, B., Hastie, T., Johnstone, I., & Tibshirani, R. (2004). Least angle regression. Annals of Statistics, 32, 407–499. Freund, Y., & Schapire, R. (1997). A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55, 119–139. Friedman, J. (2001). Greedy function approximation: A gradient boosting machine. Annals of Statistics, 29, 1189–1232. Friedman, J., Hastie, T., & Tibshirani, R. (2000). Additive logistic regression: A statistical view of boosting. Annals of Statistics, 28, 337–407. Hall, P., Titterington, D. M., & Xue, J. H. (2009). Median-based classifiers for high- dimensional data. Journal of the American Statistical Association, 104, 597–1608. Hendricks, W., & Koenker, R. (1992). Hierarchical spline models for conditional quantiles and the demand for electricity. Journal of the American Statistical Association, 93, 58–68. Hunter, D. R., & Lange, K. (2000). Quantile regression via an MM algorithm. Journal of Computational and Graphical Statistics, 9, 60–77. Koenker, R. (2005). Quantile regression (1st ed.). New York: Cambridge University Press. Koenker, R., & Bassett, G. (1978). Regression quantiles. Econometrica, 46, 33–50. Koenker, R., & Geling, R. (2001). Reappraising medfly longevity: A quantile regression survival analysis. Journal of the American Statistical Association, 96, 458–468. Koenker, R., & Hallock, K. (2001). Quantile regression. Journal of Economic Perspectives, 15, 143–156. Kordas, G. (2002). Credit scoring using binary quantile regression. In Y. Dodge, (Ed.), Statistical data analysis based on the l1-norm and related methods. Kordas, G. (2006). Smoothed binary regression quantiles. Journal of Applied Econometrics, 21, 387–407. Kriegler, B., & Berk, R. (2007). Boosting the quantile distribution: A cost-sensitive statistical learning procedure. Technical Report, Department of Statitics, University of California, Los Angeles. Langford, J., Oliveira, R., & Zadrozny, B. (2006). Predicting conditional quantiles via reduction to classification. In Proceedings of uncertainty in artificial intelligence. Li, Y., & Zhu, J. (2008). L1-Norm quantile regression. Journal of Computational and Graphical Statistics, 17, 163–185. Li, Y., Liu, Y., & Zhu, J. (2007). Quantile regression in reproducing kernel Hilbert spaces. Journal of the American Statistical Association, 102, 255–268. Manski, C. F. (1985). Semiparametric analysis of discrete response: Asymptotic properties of the maximum score estimator. Journal of Economics, 27, 313–333. Masnadi-Shirazi, H., & Vasconcelos, N. (2007). Asymmetric boosting. In Proceedings of international conference on machine learning (ICML). Mason, L., Baxter, J., Bartlett, P. L., & Frean, M. (2000). Boosting algorithms as gradient descent. In Proceedings of advances in neural information processing systems (NIPS). Mease, D., Wyner, A. J., & Buja, A. (2007). Boosted classification trees and class probability/quantile estimation. Journal of Machine Learning Research, 8, 409–439. Navidi, W. (2010). Statistics for engineers and scientists (3rd ed.). New York: McGraw- Hill. Osuna, E., Freund, R., & Girosit, F. (1997). Training support vector machines: An application to face detection. In Proceedings of IEEE conference on computer vision and pattern recognition (CVPR). Sun, Y., Kamel, M., Wong, A., & Wang, Y. (2007). Cost-sensitive boosting for classification of imbalanced data. Pattern Recognition, 40, 3358–3378. Takeuchi, I., Le, Q. V., Sears, T. D., & Smola, A. J. (2006). Nonparametric quantile estimation. Journal of Machine Learning Research, 7, 1231–1264. Torralba, A., Murphy, K. P., & Freeman, W. T. (2004). Sharing features: Efficient boosting procedures for multiclass object detection. In Proceedings of IEEE conference on computer vision and pattern recognition (CVPR). Viola, P., & Jones, M. (2001). Rapid object detection using a boosted cascade of simple features. In Proceedings of IEEE conference on computer vision and pattern recognition (CVPR). Walsh, G. R. (1975). Methods of optimization. London: John Wiley and Sons. Wu, Y., & Liu, Y. (2009). Variable selection in quantile regression. Statistica Sinica, 19, 801–817. Zemel, R., & Pitassi, T. (2001). A gradient-based boosting algorithm for regression problems. In Proceedings of advances in neural information processing systems (NIPS). QBoost: Predicting quantiles with boosting for regression and binary classification 1 Introduction 2 Methods 2.1 Quantile boost regression 2.2 Quantile boost classification 2.2.1 Predicting quantiles of binary response 2.2.2 Quantile boost for binary classification 2.2.3 Relationship between Eq. (11) and error rate 2.2.4 Related works 3 Calculation 4 Results and discussion 4.1 Experiments on regression problems 4.1.1 Results of QBR on benchmark datasets 4.1.2 Comparing to L1 norm quantile regression 4.2 Experiments on binary classification problems 4.2.1 Results on credit score data 4.2.2 Results on gene expression data 4.2.3 Face detection 5 Conclusions Acknowledgments Appendix A The equivalence between Eqs. (10) and (11) References 
work_zfylebmxrbe4vo7hi37ulgbeje	----	Citation-Based Journal Rankings for AI Research A Business Perspective ■ A significant and growing area of business-com- puting research is concerned with AI. Knowledge about which journals are the most influential fo- rums for disseminating AI research is important for business school faculty, students, administra- tors, and librarians. To date, there has been only one study attempting to rank AI journals from a business-computing perspective. It used a subjec- tive methodology, surveying opinions of business faculty about a prespecified list of 30 journals. Here, we report the results of a more objective study. We conducted a citation analysis covering a time period of 5 years to compile 15,600 cita- tions to 1,244 different journals. Based on these data, the journals are ranked in two ways involv- ing the magnitude and the duration of scientific impact each has had in the field of AI. A I research has been striving for the past four decades to increase the intelligence displayed by computing systems. To- day, there are many distinct subfields within AI—natural language processing, speech recognition and synthesis, pattern recognition and computer vision, robotics, knowledge rep- resentation, machine learning, fuzzy logic, and expert systems, to mention a few. Each at- tempts to automate specific aspects of human intelligence, and each is relevant to business research and practice. This relevance cuts across several business fields but is particularly pronounced for the field of business-comput- ing systems, which has a growing intersection with the AI field. Nearly 400 faculty listed in the 1992 MISRC/McGraw-Hill directory iden- tify an AI area as a research specialization (De- Gross, Davis, and Littlefield 1992). Courses dealing with AI topics have become common- place in business school curricula. A recent special issue of Communications of the ACM on commercial and industrial appli- cations of AI provides a timely depiction of the dramatic effects of AI research in business computing (CACM 1994). AI business applica- tions today span the realms of manufacturing, finance, and management, employing such technologies as knowledge-based systems, vi- sion systems, automatic speech recognition, microelectromechanical systems, fuzzy logic, neural networks, and genetic or evolutionary algorithms. Consumer products are now mea- sured in machine intelligence quotients. Companies such as Digital, IBM, and DuPont have ongoing efforts in developing AI applica- tions to solve complex real-world problems, improve productivity, and achieve strategic competitiveness. Many regard AI as the next wave in the ongoing computing revolution (Dutta 1993). To stay at the forefront of this revolution, a guide to the latest and most influential scien- tific developments in AI is critical. The pur- pose of this article is to offer such a guide for researchers and practitioners who operate on the cusp of the AI and business-computing fields. We do so by developing objective rank- ings of journals that have the greatest impact on AI research. The rankings are based on an extensive citation analysis. Because the cita- tion base is determined from a survey of busi- ness school faculty about the quality of AI journals, the rankings have a definite business orientation. The results reported here are of practical interest to business-computing re- searchers contemplating where to submit their own AI research. They are of interest to both faculty and students who need to allo- cate their limited time to reading among a Articles SUMMER 1996 87 Citation-Based Journal Rankings for AI Research A Business Perspective Chun Hung Cheng, Clyde W. Holsapple, and Anita Lee Copyright © 1996, American Association for Artificial Intelligence. All rights reserved. 0738-4602-1996 / $2.00 for each of the 30 journals. This recognition factor is the percentage of the 111 respon- dents who were sufficiently familiar with the journal to rate it on the 1 to 4 scale. These recognition factors ranged from 89 to 20. She argued that journals of relatively recent vin- tage generally have had less time to be recog- nized. Accordingly, she developed two addi- tional rankings. The first ranking had the 12 journals with recognition factors above 50 percent based on their ranked weighted-aver- age scores. The second ranking listed the re- maining journals on the same basis. The in- tent of this dual-ranking approach was to overcome bias introduced by the age differ- ences of the journals. Gupta’s study is a pioneering effort at as- sessing the quality and impact of various jour- nals on AI research from a business perspec- tive. However, it has some notable limitations. First, respondents were given the task of rating a prespecified list of 30 journals. Are these the 30 most influential journals for AI research, or are important journals omit- ted? Second, the study is strictly subjective. The opinions of business school faculty are, of course, important. However, do they accurate- ly reflect the actual relative influences of AI journals? Third, the longevity of a journal might well impact the recognition it garners. However, is a 50-percent recognition factor an appropriate cutoff for partitioning journals in- to two rankings, and might not a single rank- ing that is normalized to adjust for longevity be more useful? The research reported in this article addresses all three of these concerns. One other related study produced rankings of journals based on their relative impacts on expert system research (Cheng, Holsapple, and Lee 1995). Although this study was objective, host of journals. They are of interest to uni- versity administrators who need an objective way to gauge the AI outlets in which their fac- ulty members publish. They are of interest to business school librarians who need a way to assess what AI journals are most important to include in their collections. We begin with a brief review of studies re- lated to ranking journals that publish AI re- search. Next, details of the citation-analysis methodology are described. Findings from this analysis are then presented. Results based on both unnormalized and normalized cita- tion scores are reported. They are compared with a previously reported subjective ranking, showing that our objective rankings yield some major differences from the earlier work. A concluding discussion accentuates insights gained from this study. Related Studies The ranking of journals has long been under- taken as a means to gauge journal quality and influence (Garfield 1979). Our literature review reveals only one prior attempt to rank AI journals. In 1992, Gupta (1994) surveyed the opinions of 111 AACSB accredited busi- ness faculty about the academic quality and reputation of 30 journals she identified as publishing AI research. Each respondent rat- ed each of these journals on a scale of 1 (low quality) to 4 (top quality). Journals were then ranked according to a weighted-average score derived from the respondents’ ratings. Gupta grouped the ranked journals into three cate- gories of roughly comparable size that she called top, medium, and low, reflecting the weighted-average scores. Gupta also reported a recognition factor Articles 88 AI MAGAZINE Field of Interest Basis of Analysis Purpose AI References made to papers published in Artificial Intelligence from 1970–1991 To identify 50 most influential papers in AI (Bobrow 1993) Business computing Over 25,000 citations from 5 base journals covering a time period from 1987–1991 To rank business-computing journals (Holsapple et al. 1993) Decision support systems (DSS) Publishing records in DSS- related areas from 32 U.S. institutions examined To identify the most-influential contributors and the leading U.S. universities in DSS-related research (Eom and Lee 1993) Management information systems (MIS) References from an MIS literature-review article in 1988 To identify a core of MIS journals (Cooper, Blair, and Pao 1993) Table 1. Citation Studies Published in 1993 for AI and Business-Computing Research. expert systems form only one segment of the AI field. Thus, its results are of interest to those focusing on expert systems, but they could be expected to differ from the broader AI study reported here—and indeed they do. Citation Analysis In the interest of objectivity, our ranking is es- tablished through an extensive study of cita- tion patterns existing in a base set of AI arti- cles. This methodology is known as citation analysis, the merits of which are put forth by Cooper, Blair, and Pao (1993): Citation analysis is an unobtrusive way to judge the influence of research within the research community. Such analyses do not require cooperation of respon- dents and thus are not prone to many of the biases associated with eliciting re- searcher perceptions and the noise which can be introduced due to multiple per- ceptions of influence criteria. Reported studies in 1993 using citation analy- sis for various purposes in AI and business fields are summarized in table 1. Regardless of how citation analysis is ad- ministered, the identification of a base set of articles related to the subject under study is crucial. It is important that the base set of arti- cles be representative of the best work in the subject area and that the inclusion-exclusion of articles be immune from researcher judg- ments or bias. In this study, the base set of ar- ticles is collected from six AI journals, cover- ing the period from 1989 to 1993. In assembling this citation base, we exercised no judgment in choosing the specific AI journals or selecting specific articles from them. The journals were effectively chosen by the busi- ness faculty responding to Gupta’s survey. All articles in these journals during the five-year period were included in the base set of arti- cles. To identify the base journals, we established three criteria: (1) the journal must have a clear and exclusive AI focus, as indicated by its stat- ed editorial scope; (2) it must not be perceived by business faculty as having a relatively low academic quality; and (3) it must have a recog- nition factor at least half as large as the maxi- mum for all journals satisfying the first two criteria. The first criterion permits us to avoid deciding whether a specific article has suffi- cient AI content for inclusion in the base set of articles. This decision is made by the journals’ editors. The second criterion ensures that on the whole, articles in the base set are perceived by business faculty as being of sound quality. The third criterion gives a base set of articles that are not obscure from a business faculty perspective. Table 2 provides details behind the identifi- cation of base journals based on these criteria: Artificial Intelligence, AI Magazine, Expert Systems, Expert Systems with Applications, IEEE Expert, and IEEE Transactions on Pattern Analysis and Ma- chine Intelligence. The table shows all 19 jour- nals that meet the second criterion. Of these, five were eliminated because of the first criteri- on. They have a computing focus that includes but also goes beyond AI. Of the remaining jour- nals, IEEE Expert and AI Magazine had the maxi- mum recognition factor (at 86 percent). The other four base journals were all recognized at more than half this rate. The result is a substan- tial set of base articles that we contend is repre- sentative of what business faculty regard as quality research in the AI field. General Findings The base set of 1,519 articles yielded 36,420 citations to books, proceedings, and 1,224 dif- ferent journals in their combined reference lists. This data set is compiled from all vol- umes of the base journals from 1989 to 1993, including a special issue of AI Magazine in 1990. The 36,420 citations do not include ref- erences to working papers, personal commu- nications, presentations, and non-English ar- ticles. A tabulation of citation distributions by year is shown in table 3. The 15,600 journal citations consistently dominate the distribu- tion every year with no major variations. Notable differences in citation patterns are found across base journals, as shown in table 4. First, AI Magazine has a lower percentage of citations to journal articles (28 percent) than any of the other base journals. In contrast, IEEE Transactions on Pattern Analysis and Ma- chine Intelligence is the only base journal hav- ing over 50 percent of its citations to journals. Second, more than 39 percent of the 15,600 total citations to journal articles are from IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, which has nearly 10 times the number of journal citations found in ei- ther Expert Systems or AI Magazine. It has near- ly double the number of articles published in Artificial Intelligence and roughly matches the total number of articles appearing in the oth- er four base journals combined. These major differences lead us to rankings of journals based not on citation counts but rather on ci- tation scores that adjust for the imbalances among base journals’ contributions to the base set of AI articles. Citation analysis is an unobtru- sive way to judge the influence of research within the research community. Articles SUMMER 1996 89 ticles in the base journal citing it, as follows: Let i = a base year, i = 1, …, 5. j = a base journal, j = 1, …, 6. k = a journal referenced by the base set of articles, k = 1, …, 1224. Sk = the citation score of journal k. Cijk = the number of citations received by journal k from base journal j in base year i. nij = the number of articles published by base journal j in base year i. The 1224 different journals are then ranked A Ranking Based on Citation Score The relatively large number of articles pub- lished by IEEE Transactions on Pattern Analysis and Machine Intelligence in the past five years suggests that if raw citation counts are used as a basis for ranking, results will be dominated by the citation pattern of IEEE Transactions on Pattern Analysis and Machine Intelligence arti- cles. Consequently, we adjust the raw citation counts a journal receives by the number of ar- Articles 90 AI MAGAZINE Journal Name Recognition Factor Rank Focus Communications of the ACM 89 Top Computing IEEE Expert 86 Medium AI AI Magazine 86 Medium AI IEEE Transactions on Knowledge and Data Engineering 72 Top Computing Decision Support Systems 68 Top Computing IEEE Transactions on Systems, Man, and Cybernetics 64 Top Computing International Journal of Man-Machine Studies 60 Top Computing IEEE Transactions on Pattern Analysis and Machine Intelligence 60 Top AI Artificial Intelligence 58 Top AI Expert Systems with Applications 53 Medium AI Expert Systems 44 Medium AI Applied Artificial Intelligence 41 Medium AI International Journal of Expert Systems: Research and Applications 38 Medium AI Heuristics: Journal of Knowledge Engineering 37 Medium AI International Journal of Intelligent Systems 36 Medium AI Machine Learning 35 Top AI Knowledge Acquisition 32 Medium AI Journal of Automated Reasoning 28 Medium AI Applied Intelligence 24 Medium AI Table 2. Identifying a Set of Base Journals. 1989 1990 1991 1992 1993 No. of Articles 241 294 337 315 332 Total Citations 5565 7394 8086 8072 7303 Total Journal Citations 2591 3038 3353 3506 3112 Average No. of Journal Citations/Article 10.8 10.3 9.9 11.1 9.4 Journal Citations (%) 46 41 42 44 43 Book Citations (%) 22 27 28 24 24 Proceedings Citations (%) 23 24 23 25 26 Technical Report and Thesis (%) 9 8 7 7 7 Table 3. Distribution of Citations by Base Years. by their citation scores: S1, …, S1224. Only the top 5 percent of journals ranked by the citation score (that is, 62) are listed in table 5. It is unrealistic and not very useful to attempt reporting the ranks of over 1000 jour- nals. Besides, the 62 journals shown in table 5 represent over 70 percent of all journal article citations. It is interesting to see that half our base journals represent the top three: Artificial Intelligence, IEEE Transactions on Pattern Analy- sis and Machine Intelligence, and AI Magazine. Five of them are in the top 10, and all are in the top 25. Half the top 10 journals have a strict AI focus. The other half of the top 10 journals have a broader editorial scope that includes other fields of computing in addition to AI. A Ranking Based on a Normalized Score Journals that have been published over a longer period have a greater opportunity to be cited. To offset bias introduced by the age dif- ferences of the journals, we followed the ap- proach used in two earlier citation studies to obtain a normalized ranking (Cheng, Holsap- ple, and Lee 1995; Holsapple et al. 1994). In arriving at a normalized ranking, the begin- ning year of publication of each journal is ob- tained from Ulrich’s International Periodicals Directory (1993). The citation score is normal- ized by dividing the cumulative score for each journal by the total number of years the jour- nal has been in print during the period 1979 through 1992. In doing so, we assume that the scientific impact from a journal’s article in its field of study cannot last much longer than a decade. That is, few citations appearing in 1989 articles would be to articles published before 1979. This assumption is reasonable because AI is a rapidly growing and changing field. In addition, we assume that few 1993 publications cite other journal articles pub- lished in 1993. Thus, the period for normal- ization only goes through 1992. A ranking based on normalized citation scores is given in table 6. The column labeled differential indicates the relative shift in rank- ing under normalization. A positive differen- tial for a journal means that it is ranked high- er under the normalized scheme as opposed to the previous unnormalized scheme. This indicator represents relatively young, up-and- coming journals for influencing AI research. There are some substantial differences be- tween the two rankings shown in tables 5 and 6. First, despite a drop in ranking for AI Maga- zine, all the 6 base journals are among the 10 most influential journals for AI after normal- ization. Second, with normalization, 11 jour- nals rise from below the 5-percent reporting cutoff: (1) Neural Computation; (2) Neural Net- works; (3) AI Communications; (4) Complex Sys- tems; (5) AI in Medicine; (6) IEEE Transactions on Knowledge and Data Engineering; (7) Applied AI; (8) International Journal of Approximate Rea- soning; (9) Knowledge-Based Systems; (10) Inter- national Journal of Expert Systems; and (11) AI for Engineering Design, Analysis, and Manufac- turing. All these journals began publication af- ter 1986, and nearly all have a clear AI focus. Third, 11 journals drop below the 5-percent cutoff after normalization: (1) Scientific Ameri- can, (2) Journal of the Operational Research Soci- ety, (3) SIAM Journal on Computing, (4) Fuzzy Sets and Systems, (5) European Journal of Opera- tions Research, (6) Computer Journal, (7) Meth- ods of Information in Medicine, (8) Computers Articles SUMMER 1996 91 AI1 AIM2 ES3 ESWA4 IE5 ITPAMI6 Total No. of Articles 312 105 85 235 181 601 1,519 Total Citations 9,525 2,747 1,864 5,081 2,529 14,674 36,420 Total Journal Citations 3,292 764 784 2,218 957 7,585 15,600 Average No. of Journal Citations/Article 10.55 7.28 9.22 9.44 5.29 12.62 10.27 Journal Citations (%) 35 28 42 44 38 52 43 Book Citations (%) 28 28 34 31 28 19 25 Proceedings Citations (%) 27 30 19 20 26 23 24 Technical Report and Thesis 11 14 5 5 8 6 8 Table 4. Distribution of Citations by Base Journals. 1. AI = Artificial Intelligence. 2. AIM = AI Magazine. 3. ES = Expert Systems. 4. ESWA = Expert Systems with Applications. 5. IE = IEEE Expert. 6. ITPAMI = IEEE Transactions on Pattern Analysis and Machine Intelligence. S C n k ijk ij ji = × == ∑∑ 1 6 1 5 100 ulate what would have been her survey results if they had been included in her prespecified list. Clearly, subjective evaluation of journals’ effects on AI is not reflective of the actual cita- tion pattern of the articles in journals rated highly by business faculty. Half the 30 jour- nals ranked in Gupta’s study are below the 5- percent cutoff in our unnormalized ranking, and nearly one-third are not among the top 200 journals. In addition, one-third are below the 5-percent cutoff under the normalized rank. Based on specific journals, the citation pat- terns clearly suggested that business faculty should pay far more attention to the AI jour- nals AI Magazine and Expert Systems than they apparently are prone to do. They should also not overlook such journals as Computer Vision, Graphics, and Image Processing, Cognitive Sci- ences, Computational Intelligence, International Journal of Computer Vision, SIGART Newsletter, and Chemical Engineering, (9) Information Pro- cessing Letters, (10) Journal of the American Statistics Association, and (11) Journal of Experi- mental Psychology. All these journals began publication before 1979, and most have a broader focus than just AI. A Comparison of Rankings Although important differences are found be- tween the unnormalized and the normalized methods, disparities are also observed when these two objective rankings are compared to Gupta’s subjective ranking. As shown in table 7, the only ranks that remain unchanged are for the top two AI journals: (1) Artificial Intel- ligence and (2) IEEE Transactions on Pattern Analysis and Machine Intelligence. Other than these two journals, more than half the top 5 percent of journals in our citation analysis are not rated in Gupta’s study. We can only spec- Articles 92 AI MAGAZINE Rank Journal Name Score 1 Artificial Intelligence 4147.3 2 IEEE Transactions on Pattern Analysis and Machine Intelligence 1805.2 3 AI Magazine 950.3 4 Communications of the ACM 797.2 5 Computer Vision, Graphics, and Image Processing 796.4 6 International Journal of Man-Machine Studies 581.8 7 Expert Systems 569.3 8 IEEE Transactions on Systems, Man, and Cybernetics 565.2 9 Cognitive Science 513.7 10 IEEE Expert 512.4 11 Machine Learning 480.8 12 IEEE Computer 358.1 13 Journal of the ACM 300.2 14 IEEE Transactions on Software Engineering 277.8 15 Pattern Recognition 271.2 16 IEEE Transactions on Computers 225.2 17 Computational Intelligence 223.2 18 AI Expert 205.2 19 ACM Computing Surveys 194.9 20 International Journal of Computer Vision 194.6 21 Management Science 187.1 22 Journal of Automated Reasoning 174.2 23 International Journal of Robotics Research 173.1 24 IEEE Transactions on Robotics and Automation 171.0 25 Expert Systems with Applications 169.6 26 Biological Cybernetics 162.6 27 IEEE Transactions on Signal Processing 150.2 28 Journal of the Optical Society of America 143.3 29 Science 138.1 30 Psychological Review 133.8 Table 5. Ranking of AI Journals by Citation Scores. and IEEE Computer as sources (and outlets) for influential AI articles. Conversely, some jour- nals perceived by Gupta’s respondents to be of high quality have had little impact from a cita- tion-pattern perspective. For example, Heuris- tics and Applied Intelligence have had relatively little impact on the large and representative set of base AI articles derived from Gupta’s business faculty respondents. Conclusions Objective rankings of journals for AI research (from a business perspective) were developed. The method used was citation analysis. Facul- ty and students interested in AI can use the re- sults to create prioritized reading lists for stay- ing abreast of developments in the field. To complement external reviews in promotion cases, administrators can use the rankings to objectively assess the quality of research arti- cle placements. For researchers, the rankings suggest where to submit articles and give evi- dence to buttress merit-review cases. For li- brarians, they provide guidance about what AI journals are most important to have in a busi- ness collection. Some caution should be exercised when ap- plying our results. Just because a journal does not appear near the top of our rankings does not mean that it is not a quality publication. Its editorial scope might be so broad that its impact on AI research is small compared to publications devoted to AI. At the opposite ex- treme, a journal might be so highly specialized on some topic within the AI field that its im- Articles SUMMER 1996 93 Rank Journal Name Score 31 SIGART Newsletter 124.9 32 Pattern Recognition Letter 123.9 33 IEEE Transactions on Information Theory 123.7 34 Nature 122.7 35 Journal of the Royal Statistical Society 114.4 36 Computational Linguistics 110.6 37 Operations Research 110.1 38 Information Sciences 106.6 39 Knowledge Acquisition 106.3 40 Cognitive Psychology 105.7 41 Computers and Biomedical Research 104.6 42 Machine Intelligence 101.6 43 International Journal of Production Research 98.2 44 IBM Journal of Research and Development 94.9 45 Decision Sciences 92.9 46 Scientific American 88.3 47 Journal of the Operational Research Society 85.2 48 SIAM Journal on Computing 84.0 49 Fuzzy Sets and Systems 82.0 50 Decision Support Systems 80.7 51 Journal of Logic Programming 78.5 52 European Journal of Operations Research 76.8 53 Computer Journal 75.5 54 Image and Vision Computing 75.1 55 AI in Engineering 74.1 56 Methods of Information in Medicine 73.8 57 Computers and Chemical Engineering 71.4 58 IEEE Signal Processing 70.6 59 Information Processing Letters 70.1 60 Journal of the American Statistics Association 69.8 61 Journal of Experimental Psychology 69.0 62 Knowledge Engineering Review 66.8 Table 5. Continued. voted to the field of AI. In table 6, the highest- ranked business-computing journal is Decision Support Systems, but its scope is not restricted to AI articles. With the introduction of Wi- ley’s Intelligent Systems for Accounting, Finance, and Management in 1993, AI research affecting the business community appears to have gained a dedicated channel for dissemination. Although gauging the scientific impacts of journals on AI research has been the main purpose here, as well as identifying trends in AI research, other topics of interest (such as mapping the intellectual development in AI) can be investigated as an extension of the cur- rent study. pact is less than mainstream AI journals. Some journals are simply too new to appear in the rankings. Others might suffer from low circulations (for example, because of high subscription rates or modest promotion). Thus, use of the rankings to identify quality publications near the top should be supple- mented, as needed, with quality journals not near the top that are too broad, too narrow, or too new. Although there is a recent emergence of AI- specific journals in certain disciplinary ar- eas—such as AI for Engineering Design, Analy- sis, and Manufacturing; AI in Engineering; and AI in Medicine—there appears to be no estab- lished business-computing journal that is de- Articles 94 AI MAGAZINE Normalized Rank Unnormalized Rank Journal Name (year of origin) Differential 1 1 Artificial Intelligence (1970) 0 2 2 IEEE Transactions on Pattern Analysis and Machine Intelligence (1979) 0 3 10 IEEE Expert (1986) 7 4 3 AI Magazine (1980) –1 5 11 Machine Learning (1986) 6 6 7 Expert Systems (1984) 1 7 4 Communications of the ACM (1959) –3 8 5 Computer Vision, Graphics, and Image Processing (1969) –3 9 25 Expert Systems with Applications (1990) 16 10 6 International Journal of Man-Machine Studies (1969) –4 11 8 IEEE Transactions on Systems, Man, and Cybernetics (1971) –3 12 9 Cognitive Science (1977) –3 13 20 International Journal of Computer Vision (1987) 7 14 31 SIGART Newsletter (1989) 17 15 18 AI Expert (1986) 3 16 17 Computational Intelligence (1985) 1 17 39 Knowledge Acquisition (1989) 22 18 12 IEEE Computer (1971) –6 19 22 Journal of Automated Reasoning (1985) 3 20 13 Journal of the ACM (1954) –7 21 24 IEEE Transactions on Robotics and Automation (1985) 3 22 14 IEEE Transactions on Software Engineering (1975) –8 23 15 Pattern Recognition (1968) –8 24 16 IEEE Transactions on Computers (1968) –8 25 23 International Journal of Robotics Research (1982) –2 26 19 ACM Computing Surveys (1969) –7 27 21 Management Science (1954) –6 28 62 Knowledge Engineering Review (1988) 34 29 83 Neural Computation (1989) 54 30 32 Pattern Recognition Letter (1983) 2 31 68 Neural Networks (1988) 37 Table 6. Ranking of AI Journals by Normalized Citation Scores. References Bobrow, D. G. 1993. Artificial Intelligence in Per- spective: A Retrospective on 50 Volumes of the Arti- ficial Intelligence Journal. Artificial Intelligence 59(1): 5–20. CACM. 1994. Special Issue on Commercial and In- dustrial AI. Communications of the ACM 37(3): 23–119. Cheng, C. H.; Holsapple, C. W.; and Lee, A. 1995. Citation-Based Journal Rankings for Expert Systems Research. Expert Systems 12(4): 1–10. Cooper, R. B.; Blair, D.; and Pao, M. 1993. Commu- nicating MIS Research: A Citation Analysis of Jour- nal Influence. Information Processing and Manage- ment 29(1): 113–127. DeGross, J. I.; Davis, G. B.; and Littlefield, R. S. 1992. 1992 Directory of Management Information Systems Faculty in the United States and Canada. New York: MISRC–McGraw-Hill. Dutta, S. 1993. Knowledge Processing and Applied Arti- ficial Intelligence. Jordan Hill, Oxford: Butterworth- Heinemann. Eom, S. B., and Lee, S. M. 1993. Leading U.S. Uni- versities and Most Influential Contributors in Deci- sion Support Systems Research (1971–1989): A Cita- tion Analysis. Decision Support Systems 9(3): 237–244. Garfield, E. 1979. Citation Indexing: Its Theory and Application in Science, Technology, and Humanities. New York: Wiley. Gupta, U. G. 1994. The Academic Quality of AI Journals and the Role of AI in the MIS Curriculum: Perspectives of Business Faculty. Expert Systems with Applications 7(4): 581–588. Holsapple, C. W.; Johnson, L. E.; Manakyan, H.; and Tanner, J. 1993. A Citation Analysis of Business- Computing Research Journals. Information and Man- agement 25(2): 231–244. Holsapple, C. W.; Johnson, L. E.; Manakyan, H.; and Tanner, J. 1994. Business-Computing Research Jour- nals: A Normalized Citation Analysis. Journal of Articles SUMMER 1996 95 Normalized Rank Unnormalized Rank Journal Name (year of origin) Differential 32 26 Biological Cybernetics (1975) –6 33 79 AI Communications (1988) 46 34 27 IEEE Transactions on Signal Processing (1951) –7 35 55 AI in Engineering (1986) 20 36 28 Journal of the Optical Society of America (1917) –8 37 50 Decision Support Systems (1985) 13 38 70 Complex Systems (1987) 32 39 29 Science (1880) –10 40 30 Psychological Review (1894) –10 41 33 IEEE Transactions on Information Theory (1963) –8 42 34 Nature (1869) –8 43 51 Journal of Logic Programming (1984) 8 44 116 AI in Medicine (1989) 72 45 117 IEEE Transactions on Knowledge and Data Engineering (1989) 72 46 84 Applied Artificial Intelligence (1987) 38 47 35 Journal of the Royal Statistical Society (1838) –12 48 36 Computational Linguistics (1974) –12 49 37 Operations Research (1952) –12 50 86 International Journal of Approximate Reasoning (1987) 36 51 58 IEEE Signal Processing (1984) 7 52 87 Knowledge-Based Systems (1987) 35 53 38 Information Sciences (1969) –15 54 88 International Journal of Expert Systems (1987) 34 55 40 Cognitive Psychology (1970) –15 56 54 Image and Vision Computing (1983) –2 57 41 Computers and Biomedical Research (1969) –16 58 42 Machine Intelligence (1967) –16 59 43 International Journal of Production Research (1961) –16 60 95 AI for Engineering Design and Manufacturing (1987) 35 61 44 IBM Journal of Research and Development (1957) –17 62 45 Decision Sciences (1970) –17 Table 6.Continued. manufacturing, and group technology. Clyde Holsapple holds the Rosenthal Endowed Chair in Management Information Sys- tems and currently serves as area coordinator of Decision Science and Information Systems at the University of Kentucky. He does teaching and research in decision support systems, knowledge man- agement, expert systems, and organizational com- puting. His books include Foundations of Decision Support Systems, Micro Database Management, Busi- ness Expert Systems, and The Information Jungle. Hol- Management Information Systems 11(1): 131–140. Ulrich. 1993. Ulrich’s International Periodicals Direc- tory, 32nd ed. New York: Bowker. Chun Hung Cheng received his Ph.D. in business administration from the University of Iowa in 1990. Prior to joining the Chinese University of Hong Kong, he was an assistant professor at Grand Valley State University and Ken- tucky State University. His re- search interests include knowl- edge-based systems, object-oriented design, distributed computer systems, computer integrated Articles 96 AI MAGAZINE Normalized Rank Unnormalized Rank Gupta’s Rank Journal Name 1 1 1 Artificial Intelligence 2 2 2 IEEE Transactions on Pattern Analysis and Machine Intelligence 3 10 9 IEEE Expert 4 3 16 AI Magazine 5 11 6 Machine Learning 6 7 17 Expert Systems 7 4 3 Communications of the ACM 8 5 NA Computer Vision, Graphics, and Image Processing 9 25 11 Expert Systems with Applications 10 6 8 International Journal of Man-Machine Studies 11 8 5 IEEE Transactions on Systems, Man, and Cybernetics 12 9 NA Cognitive Science 13 20 NA International Journal of Computer Vision 14 31 NA SIGART Newsletter 15 18 26 AI Expert 16 17 NA Computational Intelligence 17 39 18 Knowledge Acquisition 18 12 NA IEEE Computer 19 22 13 Journal of Automated Reasoning 20 13 NA Journal of the ACM 21 24 NA IEEE Transactions on Robotics and Automation 22 14 NA IEEE Transactions on Software Engineering 23 15 NA Pattern Recognition 24 16 NA IEEE Transactions on Computers 25 23 NA International Journal of Robotics Research 26 19 NA ACM Computing Surveys 27 21 NA Management Science 28 62 25 Knowledge Engineering Review 29 83 NA Neural Computation 30 32 NA Pattern Recognition Letter 31 68 NA Neural Networks 32 26 NA Biological Cybernetics 33 79 NA AI Communications 34 27 NA IEEE Transactions on Signal Processing 35 55 NA AI in Engineering Table 7. Comparison of Journal Rankings for AI Research. sapple is also associate editor of Management Science and Organizational Computing; area editor for Deci- sion Support Systems and cofounder of the Interna- tional Society for Decision Support Systems. In 1993, he was named computer educator of the year by the International Association for Computer In- formation Systems. Anita Lee is an assistant professor in the decision science and information systems area at the Univer- sity of Kentucky. She received her Ph.D. in business administration from the University of Iowa. Her research interests include AI, ma- chine learning, knowledge-based systems, computer-integrated manufacturing, and group tech- nology. She is also the author of Knowledge-Based FMS Scheduling: An Artificial Intelligence Perspective. Articles SUMMER 1996 97 Normalized Rank Unnormalized Rank Gupta’s Rank Journal Name 36 28 NA Journal of the Optical Society of America 37 50 7 Decision Support Systems 38 70 NA Complex Systems 39 29 NA Science 40 30 NA Psychological Review 41 33 NA IEEE Transactions on Information Theory 42 34 NA Nature 43 51 NA Journal of Logic Programming 44 116 NA AI in Medicine 45 117 4 IEEE Transactions on Knowledge and Data Engineering 46 84 14 Applied Artificial Intelligence 47 35 NA Journal of the Royal Statistical Society 48 36 NA Computational Linguistics 49 37 NA Operations Research 50 86 NA International Journal of Approximate Reasoning 51 58 NA IEEE Signal Processing 52 87 20 Knowledge-Based Systems 53 38 NA Information Sciences 54 88 15 International Journal of Expert Systems: Research and Applications 55 40 NA Cognitive Psychology 56 54 NA Image and Vision Computing 57 41 NA Computers and Biomedical Research 58 42 NA Machine Intelligence 59 43 NA International Journal of Production Research 60 95 24 Artificial Intelligence for Engineering Design: Analysis and Manufacturing 61 44 NA IBM Journal of Research and Development 62 45 NA Decision Sciences 86 125 12 International Journal of Intelligent Systems NA 827 19 Applied Intelligence NA 414 21 Artificial Intelligence and Law NA NA 28 Artificial Intelligence and Society NA NA 29 Artificial Intelligence Today NA 605 22 Expert Systems for Information Management NA 386 10 Heuristics NA NA 27 Journal of Artificial Intelligence in Education NA 214 30 PC Artificial Intelligence NA 645 23 Robotics and Computer-Integrated Manufacturing NA = Not available. Table 7. Continued. 
work_zhdjd7ouobhd3hosejfsdrnmfq	----	This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/authorsrights http://www.elsevier.com/authorsrights Author's personal copy Proximity measures for link prediction based on temporal events Paulo R.S. Soares, Ricardo B.C. Prudêncio ⇑ Center of Informatics (CIn), Federal University of Pernambuco, Recife, PE, Brazil a r t i c l e i n f o Keywords: Link prediction Temporal events Neighborhood-based measures Co-authorship networks a b s t r a c t Link prediction is a well-known task from the Social Network Analysis field that deals with the occur- rence of connections in a network. It consists of using the network structure up to a given time in order to predict the appearance of links in a close future. The majority of previous work in link prediction is focused on the application of proximity measures (e.g., path distance, common neighbors) to non-con- nected pairs of nodes at present time in order to predict new connections in the future. New links can be predicted for instance by ordering the pairs of nodes according to their proximity scores. A limitation usually observed in previous work is that only the current state of the network is used to compute the proximity scores, without taking any temporal information into account (i.e., a static graph representa- tion is adopted). In this work, we propose a new proximity measure for link prediction based on the con- cept of temporal events. In our work, we defined a temporal event related to a pair of nodes according to the creation, maintenance or interruption of the relationship between the nodes in consecutive periods of time. We proposed an event-based score which is updated along time by rewarding the temporal events observed between the pair of nodes under analysis and their neighborhood. The assigned rewards depend on the type of temporal event observed (e.g., if a link is conserved along time, a positive reward is assigned). Hence, the dynamics of links as the network evolves is used to update representative scores to pairs of nodes, rewarding pairs which formed or preserved a link and penalizing the ones that are no longer connected. In the performed experiments, we evaluated the proposed event-based measure in different scenarios for link prediction using co-authorship networks. Promising results were observed when the proposed measure was compared to both static proximity measures and a time series approach (a more competitive method) that also deploys temporal information for link prediction. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction As the amount of interaction among individuals increases in vir- tual environments, more useful social data become available, serv- ing as a basis for social network analysis. Social networks are structures composed by individuals (entities) that can be con- nected by different forms of social relationship (such as friendship, in which two individuals are connected if they are friends) (Ama- ral, Scala, Barthelemy, & Stanley, 2000). In social networks, connec- tions and entities tend to appear and disappear along time, which turns them into highly dynamic and complex systems. Social Net- work Analysis (SNA) is a broad field of research that tries to deal with such complexity (Wasserman & Faust, 1994). Several tasks can be associated to SNA. In this paper, our specific aim is to inves- tigate the prediction of links in a social network, that is, we are fo- cused on predicting the most probable future connections based on previous states of the network. This is a well-known problem from the SNA field called link prediction (Hasan, Chaoji, Salem, & Zaki, 2006). A lot of work has been devoted to cope with the link prediction problem (Getoor & Diehl, 2005; Wang, Satuluri, & Parthasarathy, 2007; Xiang, 2008). The majority of previous work is based on the application of proximity measures to non-connected pairs of nodes in the network at present time in order to predict new con- nections at future time. The proximity measures are used to asso- ciate scores to the pairs of nodes, which can be used either: (1) in a unsupervised approach, in which a chosen score is simply ordered and links are predicted for the top ranked pairs; or (2) in a super- vised approach, in which the link prediction is treated as a classi- fication task and different scores are used as predictor attributes by a learning algorithm. A limitation that can be pointed out in the previous work is that the proximity scores are usually calcu- lated without taking into account the evolution of the network. The proximity measures are computed using all network data up to the present moment (i.e., the current network state) without considering when links were created. Hence, a potentially useful source of information for link prediction is not adequately leveraged. 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.06.016 ⇑ Corresponding author. Tel.: +55 8121268430. E-mail addresses: prss@cin.ufpe.br (P.R.S. Soares), rbcp@cin.ufpe.br (R.B.C. Prudêncio). Expert Systems with Applications 40 (2013) 6652–6660 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a Author's personal copy In the current work, we propose a novel proximity measure in which temporal information is taken into account. In the proposed measure, we deployed the concept of temporal events, which are specific activities (e.g., creation or removal of a link) observed be- tween a pair of nodes in consecutive time intervals. For instance, an innovative event occurs when two nodes are not connected in a given time interval and a new link is created between them in the next interval. The proximity score for a pair of nodes is com- puted by monitoring along time the events observed around the nodes and their immediate neighbors. Each category of temporal event defined in our work (innovative, conservative and regressive) is associated to a numerical reward and the proximity score in- creases or decreases along time depending on the temporal events observed in consecutive network frames. Our proposal was sup- ported by the Homans’ work that associates the strength of a con- nection between individuals to their frequency of interaction (which was modeled in our work by the temporal events) (Homans, 1951) and the idea shared by many authors that a bigger common neighborhood between individuals is closely related to a higher probability of future connections (Newman, 2001). In order to verify the viability of the proposed measure, we per- formed experiments on co-authorship networks extracted from four sections of the physics e-Print arXiv.1 In these experiments, the proposed measure was evaluated in different scenarios by con- sidering for instance different reward values for the temporal events and by adopting a weighting function to give higher importance to more recent events. For a baseline comparison, experiments were performed with traditional proximity measures previously adopted in the literature. We also performed experiments with a time series approach (Potgieter, April, Cooke, & Osunmakinde, 2009) which, similarly to our approach, also aims to consider temporal informa- tion to improve link prediction. In all experiments, the unsupervised approach was adopted to perform the prediction task once the scores were calculated and the Area Under the ROC Curve (AUC) was used to evaluate the performance of prediction. The performance achieved by the event-based measure outperformed the results of the baseline methods in all the networks considered. Section 2 briefly presents the link prediction problem. Section 3 describes the event-based approach, showing the concepts used and the score calculation process. Section 4 presents the social net- work data adopted, the experiments and obtained results. Finally, Section 5 concludes this work by presenting some final consider- ations and future work. 2. Link prediction Link prediction consists of predicting new connections or detecting hidden links in a network. It is a very important task applicable to a wide variety of areas, such as bibliographic domain, molecule biology, criminal investigations and recommending sys- tems (Getoor & Diehl, 2005; Xiang, 2008). A traditional definition of the link prediction problem is expressed by: ‘‘Given a snapshot of a social network at time t, we seek to accurately predict the edges that will be added to the network during the interval from time t to a given future time t0’’ (Liben-Nowell et al., 2003). Among several approaches to treat the problem, the most widespread ones rely on the use of proximity measures between pairs of nodes (Lu & Zhou, 2011; Xiang, 2008). As previously mentioned, the starting point of this approach is to extract the values/scores of different measures that indicate the similarity between pairs of nodes. These scores can be used either by unsupervised (Liben-Nowell et al., 2003; Lu & Zhou, 2011; Murata & Moriyasu, 2008) or supervised link prediction (Hasan et al., 2006; Lichtenwalter, Lussier, & Chawla, 2010; Sá & Prudêncio, 2011). In the former approach, a proximity measure is chosen and deployed to rank node pairs in the network. The top ranked ones are predicted to be linked. In the supervised link prediction, a set of proximity measures is cho- sen and adopted as predictor attributes by a classifier, learned based on historical data. Each pair of non-connected nodes is de- scribed by the set of proximity scores. A classifier then uses these attributes to perform a binary classification to decide whether the link will be formed or not in the future. The proximity measures proposed and evaluated in the litera- ture can be broadly categorized into semantic or topological mea- sures (Xiang, 2008). In the semantic measures (or node-wise measures), the nodes’ content is considered to measure proximity. For instance, in a co-authorship network, the similarity between keywords extracted from published papers can be used to predict future interaction among the authors (Xiang (2008)). Different from the semantic measures, the topological strategy consists of deploying the network structure to compute the proximity scores (e.g. the number of common neighbors that two nodes share). Topological measures are more commonly adopted in the litera- ture since they are more general and do not require the definition of rich features to describe content (in fact, rich content is not al- ways available depending on the social network considered). Several topological measures were proposed in the literature mainly categorized into neighborhood-based or path-based mea- sures (Hasan & Zaki, 2011). The neighborhood-based measures take into account the immediate neighbors of the nodes. In general, these measures consider that two nodes are more likely to form a link if their sets of neighbors have a large overlap (Xiang, 2008). Among the neighborhood-based measures, we can mention Common Neighbors (Newman, 2001), Preferential Attachment (Barabasi et al., 2002; Newman, 2001),Adamic and Adar (2003) and the Jaccard’s coefficient (Salton & McGill, 1986). The path- based measures in turn define proximity between nodes by consid- ering the paths between them. The basic idea is that two nodes are more likely to form a link if there are short paths between them. The path-based measures range from the simple path-distance measure to more sophisticated definitions that consider ensembles of different paths, such as the Katz measure (Katz, 1953). In com- parative terms, the neighborhood-based methods are more wide- spread, due to both their computational eficiency and great performance observed in experiments (Huan, 2006; Liben-Nowell et al., 2003; Murata & Moriyasu, 2008). The measure proposed in our work can be categorized as neighborhood-based, since it uses information about the connections around nodes to assign scores to them (see Section 3). Most previous work performs link prediction by statically ana- lyzing the network data, that is, no temporal information is consid- ered to perform the prediction. However, temporal information (e.g., the moments when two nodes interacted in the past or the time when a connection was first observed) is an important aspect that should be considered during the link prediction (Hasan & Zaki, 2011). For instance, when computing a proximity score based on neighborhood, it would be interesting to consider not only how many but also when the links were formed between neighbors. Recent activity between common neighbors may be more impor- tant than older activity. Also, static approaches are appropriate to investigate whether a certain link will ever occur in a network but they are less useful, for instance, to applications for which the prediction of repeated link occurrences are of interest (Huang & Lin, 2009). In this section, we mention some previous work concerning the use of temporal information for link prediction. For instance, in Tylenda, Angelova, and Bedathur (2009), the authors modeled a network as a weighted graph, in which the weight of a link was the age of the most recent activity between the corresponding1 http://www.arxiv.org P.R.S. Soares, R.B.C. Prudêncio / Expert Systems with Applications 40 (2013) 6652–6660 6653 Author's personal copy nodes. The link prediction task was then accomplished by deploy- ing extended versions of the proximity measures adequate for weighted networks (e.g., weighted Adamic Adar). Bringmann, Berlingerio, Bonchi, and Gionis (2010) proposed to mine network data augmented with temporal information in order to discover association rules (in terms of frequent subgraphs) that best ex- plained the network evolution. In Juszczyszyn, Musial, and Budka (2011), the authors adopted a related approach in which the his- tory of the network (recorded during past time windows) is used to derive probabilities of transitions between triads of nodes. Although promising results can be obtained in this approach, min- ing frequent subgraphs has shown to be a very expensive task (Coenen, Jiang, & Zito, 2012). An alternative approach for deploying temporal information is to treat link prediction as a time series forecasting problem (Huang & Lin, 2009; Potgieter et al., 2009; Qiu, He, & Yen, 2011; Soares & Prudêncio, 2012). Huang and Lin (2009) built a time series for each pair of nodes, in which each series observation is the frequency of occurrence of links between the nodes during a specific time peri- od. The time series forecasts produced by ARIMA models (Box & Jenkins, 1970) were then used to estimate the probability of future link occurrence. In Potgieter et al. (2009) and Soares and Prudêncio (2012), the authors adopted a similar idea, but in this case time series models were used to predict proximity scores. In this ap- proach, given a chosen proximity measure, a time series is built for each pair of nodes by computing the score in a sequence of time periods (i.e., using different snapshots of the network along time). A final proximity score is then obtained for each pair by forecasting its corresponding time series. Different models were applied in the forecasting process, including ARIMA models, smoothing methods and linear regression models. Despite the good results obtained in experiments, a difficulty in the time series approach is to choose an adequate forecasting model, among the variety of models that can be applied. In fact, in Soares and Prudêncio (2012), the authors ob- served that a wrong decision concerning the forecasting model can have a negative impact in the link prediction performance. 3. Event-based link prediction Social networks are highly dynamic structures in which several connections and nodes tend to appear or disappear along time. The temporal evolution of these networks brings valuable information about how connections tend to be formed, and, for that, should be considered in the link prediction task. In the current work, we pro- pose a new proximity measure that takes into account the tempo- ral structure of a network and temporal events related to pairs of nodes in the network. The aim of our proposal is to combine Ho- mans’ thought that the strength of a connection between two indi- viduals is directly associated with how often they interact with one another (modeled here by means of temporal events) (Homans, 1951) and Newman’s idea that the bigger the number of common neighbors between two nodes, the higher is their probability to be connected in the future (Newman, 2001). The general idea of the proposed approach is to increase or de- crease the proximity score between two nodes depending on tem- poral events observed among the two nodes and their neighborhood. A temporal event, which will be explained better in the next section, is defined through the appearance or disap- pearance of links around nodes as the network evolves. In the pro- posed work, initially a temporal structure is created by extracting consecutive frames of the network at different time intervals. A proximity score is then computed for each pair of nodes by aggre- gating the rewards assigned for the temporal events observed dur- ing the transitions of frames. The proposed measure is explained in more detail in the next subsections. 3.1. Temporal structure In the proposed solution, we initially build a temporal structure N , by following a methodology described in Soares and Prudêncio (2012). First of all, the network is split into several time-sliced snapshots, which represent states of the network at different time intervals in the past. After that, frames are built by grouping con- secutive snapshots. The size of each frame is the same as the length of the prediction window, pre-defined in the link prediction task. Each frame represents one time step in N . This methodology is detailed below. Let GðV; EÞ be a graph representing a social network observed up to time T. Each edge in E is represented by a triple ðu; v; tÞ, indicat- ing that the nodes u and v 2 V had a social interaction at time t. Hence, more than one edge related to a single pair of nodes can be observed in E if the nodes interacted in different moments in the past. Let w be the length of the prediction window, i.e., our task is to predict new links concerning the future time interval from T þ 1 to T þ w. Let Gt be the sub-graph of G containing only the edges observed at time t. Let ½Gt; Gtþ1; . . . ; Gtþm� be the frame formed by the disjoint union of the graphs from time t to t þ m. In our work, a set of n con- secutive frames N ¼fF1; . . . ; F k; . . . ; F ng of size w is extracted from the graph G. Formally, N can be defined as N ¼f ½GT�nwþ1; GT�nwþ2; . . . ; GT�ðn�1Þw�; . . . ; ½GT�2wþ1; GT�2wþ2; . . . ; GT�w�; ½GT�wþ1; GT�wþ2; . . . ; GT�g ð1Þ The frame Fk ¼ ½GT�ðn�kþ1Þwþ1; . . . ; GT�ðn�kÞw� is hence the sub-graph of G containing all links observed in the k-th time interval of size w. As an example, consider a network observed up to the year T ¼ 2012 and a prediction window of length 2 (i.e., the task is to predict new links from 2013 and 2014). If we extract n ¼ 3 frames from the network, the following structure (N ¼fF1; F2; F 3g) is obtained N ¼f½G2007; G2008�; ½G2009; G2010�; ½G2011; G2012�g ð2Þ This structure can be used to analyze how the network evolved every 2 years since 2007. For this, we will introduce next section the concept of temporal events, that can be observed between sub- sequent frames. 3.2. Temporal events We define a temporal event as a specific activity between two nodes from a frame to its subsequent. A temporal event is the ac- tion that leads a dyad (a pair of nodes) from a state (connected or non-connected) to another. Events can be categorized into one of three mutually exclusive types: conservative, innovative or regres- sive, defined below. � Conservative A conservative event occurs when a relationship between two nodes is not dropped when the network evolves, that is, when two nodes share a link in a frame and this connection is pre- served in the subsequent one. For each pair of nodes ðu; vÞ, we define a reward Cðu; v; kÞ related to frame Fk , in order to take into account a conservative event during the transition from the ðk � 1Þth to the kth frame. Formally: Cðu; v; kÞ¼ c; if ðu; vÞ 2 Ek�1 \ Ek 0; otherwise � ð3Þ In the above definition, Ek�1 and Ek are the sets of edges observed in frames Fk�1 and Fk respectively. The constant c indicates the reward for conservative events, which should be a non-negative value since the strength of a tie between two nodes is preserved. 6654 P.R.S. Soares, R.B.C. Prudêncio / Expert Systems with Applications 40 (2013) 6652–6660 Author's personal copy � Innovative Innovative events represent the creation of a new link between two nodes on different frames. They happen when two nodes are not connected in a frame and a link is observed in the next frame. The Innovative reward Iðu; v; kÞ associated to a pair ðu; vÞ and a frame Fk is: Iðu; v; kÞ¼ i; if ðu; vÞ 2 Ek n Ek�1 0; otherwise � ð4Þ The constant i in the above equation indicates the reward for inno- vative events. Its value should be positive since the tie between two nodes is strengthened. � Regressive Regressive events are opposite to innovative ones. This kind of event represents the removal of an existing link between two nodes from a frame to its subsequent. The Regressive reward Rðu; v; kÞ is defined as: Rðu; v; kÞ¼ r; if ðu; vÞ 2 Ek�1 n Ek 0; otherwise � ð5Þ In this event, r should assume a non-positive value since the strength of the connection between the nodes tends to decrease. The values of the parameters c; i and r can be determined empir- ically by evaluating the link prediction performance in a validation set and choosing the best configuration of values considered. As it will be seen, in our experiments, the number of parameters to define was reduced to two by making the values of c and r propor- tional to i (in this case, i was set to 1 for the sake of simplicity). As it will be seen in the next subsection, given a pair of nodes, we will define as primary events the ones strictly related to both nodes in the pair, whereas secondary events happen in dyads com- posed by only one of the nodes in the pair and its neighbors. 3.3. Event-based score As previously said, many approaches for link prediction com- pute scores to pairs of nodes by deploying a chosen proximity mea- sure (see Section 2), aiming to determine how similar those nodes are and, consequently, how likely a connection between them will be formed in a close future. The proposed measure combines both: (1) the rewards associ- ated to primary events, which are the temporal events strictly re- lated to the pair of nodes under analysis; and (2) the rewards associated to secondary events, which are the temporal events ob- served in the nodes’ neighborhood. The proximity score associated to a given pair of nodes ðu; vÞ is defined as: scoreðu; vÞ¼ Xn k¼2 Pðu; v; kÞþ a:Sðu; v; kÞ ð6Þ Pðu; v; kÞ¼ Cðu; v; kÞþIðu; v; kÞþRðu; v; kÞ ð7Þ Sðu; v; kÞ¼ X y2CðuÞ\CðvÞ Pðu; y; kÞþ Pðy; v; kÞ ð8Þ In Eq. (7), Pðu; v; kÞ computes the reward of the event (conservative, innovative or regressive) for the pair of nodes ðu; vÞ observed in the transition from frame k � 1 to frame k. In Eq. (8), Sðu; v; kÞ indicates the aggregated reward of secondary events associated to the pair ðu; vÞ (i.e., the primary events observed in its common neighbor- hood). In this equation, CðxÞ is the set of neighbors of the node x in the network. In Eq. (6), the parameter a is an amortization factor that indi- cates how strong secondary events affect the tie between u and v. Pðu; v; kÞ and Sðu; v; kÞ associated to the first frame (i.e., k ¼ 1) are all null since this frame is not a result of any set of events, and hence for simplicity, they are not considered in Eq. (6). Fig. 1 illustrates the application of the proposed event-based score. Suppose that one wants to compute the proximity score for the pair ð1; 3Þ. The nodes 1 and 3 have node 2 as common neighbor in the network. Hence, the score of the connection between 1 and 3 is going to be calculated as a function of the events that occurred between themselves (primary events), plus the ones occurred in the dyads ð1; 2Þ and ð2; 3Þ (secondary events) along the network temporal structure N . From frame F 1 to frame F2 , it can be noticed that a conservative and a regressive event happened in the dyad ð1; 2Þ and ð2; 3Þ respectively. So far, no event occurred in the dyad ð1; 3Þ. Therefore, up to frame F 2, the partial score of the pair ð1; 3Þ is given by the combination of the amortized rewards associated with the second- ary events described above, that is, aðc þ rÞ. Looking at the next time step (frame F 3), a regressive event is associated to ð1; 2Þ, whereas an innovative one occurred in the pair ð2; 3Þ. In this step, there is also an innovative event related to the pair under analysis ð1; 3Þ; its presence must be counted too in the final score of the pair (its reward, however, will not be amor- tized since it is a primary event). Hence, the resultant score in this frame is given by aðr þ iÞþ i. By the definition of Eq. (6), the scores are cumulative in time. The final score of a pair of nodes is, then, the sum of all its partial scores along N . This, therefore, results in a final score of aðc þ 2r þ iÞþ i for the pair ð1; 3Þ. In the current work, we also propose an extension of the event- based score that increases the importance of more recent events observed for a pair of nodes. In Eq. (9), we propose a proximity score which considers a monotonically increasing function b : Z ! R to weight recent events more heavily than the old ones. In our work, we adopted a logarithmic function (see Eq. (10)), although other functions can be deployed. We highlight that Eq. (6) is a special case of Eq. (9) by setting bðkÞ¼ 1. scoreðu; vÞ¼ Xn k¼2 bðkÞ:½Pðu; v; kÞþ a:Sðu; v; kÞ� ð9Þ bðkÞ¼ logðkÞ ð10Þ 3.4. Prediction Finally, in order to perform the prediction of new links, we adopted an unsupervised strategy after the scores were computed for the pairs of nodes. It basically consists of ranking the pairs of nodes according to their scores, and then, selecting the top-ranked ones as the new links in the network, as explained in Section 2. Fig. 1. Example of event-based score computation using a network composed by a click of three nodes. N represents the temporal structure of the network divided in three frames (F1; F2 and F 3 ). The variables c; i and r are the rewards associated with conservative, innovative and regressive events respectively, and a is the amorti- zation factor used to deal with secondary events. P.R.S. Soares, R.B.C. Prudêncio / Expert Systems with Applications 40 (2013) 6652–6660 6655 Author's personal copy 4. Experiments and results In this section we describe the experiments performed to eval- uate the proposed approach as well as the obtained results. Ini- tially we describe the social network data used and the methodology of experiments (Section 4.1), followed by the base- line methods adopted for a comparative evaluation (Section 4.2). In Section 4.3, we present the initial results aimed to evaluate the proposed measure considering different configurations of re- wards for the temporal events. In Section 4.4, we evaluate the influence of the amortization factor a in the proposed measure. Fi- nally, in Section 4.5, we present the effect of using a weighting function for recent event as well as the comparative results with the baseline methods. 4.1. Data and settings In order to evaluate the performance of the proposed scores, data from four co-authorship networks were used in our experi- ments. We highlight that co-authorship networks are highly used in the literature to evaluate SNA techniques. Being a social struc- ture, this kind of network is very dynamic along time, what makes it a great source of data to be explored regarding the emergence of new connections. Besides, most of them are publicly available in digital environment (Barabasi et al., 2002; Newman, 2001). In a co-authorship network, a node represents an author and an edge indicates that two specific nodes have co-authored a paper. As seen in Section 3, in our work an edge also stores the time when the relationship was observed. More specifically, in the co-authorship networks adopted here, each edge stores the publication year of the co-authored paper. The network data used in our experiments were collected from the e-print arXiv,2 which maintains a large database of electronic scientific papers in several fields, such as mathematics, physics, astronomy, among others. In our experiments, the networks were built using collaborations from four distinct sub-areas, namely: as- tro-physics (astro-ph), condensed matter (cond-mat), high energy physics – lattice (hep-lat) and theoretical high energy physics (hep- th). Collaborations from 1993 to 2000 were used for astro-ph and cond-mat, and collaborations from 1992 to 2010 were used for hep- lat and hep-th. The summarized information about the adopted net- works are presented in Table 1. In our experiments, a prediction window of size one was adopted, and hence the aim of the analysis is to investigate how the networks evolved every year. By following the methodology described in Section 3.1, a temporal structure N was built for each network. For astro-ph and cond-mat, data from 1993 to 1999 were used to build N (with a total number of n ¼ 7 frames) and the snapshot from 2000 was used as the prediction frame. For hep-lat and hep-th, N was built with collaborations from 1992 to 2009 (with a total number of n ¼ 18 frames), whereas the prediction frame was composed by collaborations from 2010. As mentioned in Section 3.2, adequate values for the reward parameters can be estimated by evaluating the link prediction per- formance on a validation set. In our work, the last frame of N was adopted as the validation set in order to perform parameter selec- tion for each network: the frame of 1999 was used to validation for the astro-ph and cond-mat networks and the frame of 2009 was used to validation for hep-lat and hep-th. As usual in the link prediction field, our aim is to investigate the prediction of new links in a network. Therefore, for each network, the initial list of candidate pairs to be evaluated is formed by pairs that are not connected in the last frame of N . The final list is formed by picking (from the initial list) only the dyads that were directly or indirectly affected by some event during the network evolution. Table 2 shows the distribution of the candidate pairs of nodes regarding their class labels (connected or non-connected) both in the validation and the test sets. As it can be seen, the class distribution is highly imbalanced, which is a common problem related to the link prediction task. In order to minimize the effect of imbalanced data during evalua- tion, the Area Under ROC Curve (AUC) was adopted as the measure of performance. ROC curves relate the sensitivity (true positive rate) and specificity (true negative rate) of a classifier. The AUC has been traditionally used in imbalanced classification problems (Murata & Moriyasu, 2008; Wang et al., 2007) due to its higher robustness to class distribution anomalies (Fawcett, 2006). Finally, for more reliable results, we performed a bootstrap over the test set by taking 100 randomly stratified sub-samples and computing the AUC. The 10% of the best results and 10% of the worst results were discarded and the average AUC was computed. 4.2. Baseline methods In our experiments, the proposed measures were compared with different measures previously adopted in the literature of link prediction. Initially, four classical proximity scores based on neigh- borhood were considered: PA, CN, AA and JC. These are very popu- lar measures used in the state of the art of link prediction (Bringmann et al., 2010; Lichtenwalter et al., 2010). Each measure is explained below. The PA measure assumes that the probability of a future link between two nodes is proportional to their degrees. In a co-author- ship network, such probability is correlated to the product of the number of collaborators they have Barabasi et al. (2002). Hence, it is defined as: PAðu; vÞ¼ jjCðuÞjj � jjCðvÞjj ð11Þ The CN measure states that the bigger the number of neighbors two nodes share, the higher is their probability to form a link in the future (Newman, 2001). Formally, the measure is defined as: CNðu; vÞ¼ jjCðuÞ\ CðvÞjj ð12Þ The AA measure refines the CN by increasing the scores of pairs of nodes in which the neighbors in common possess less connections (Adamic & Adar, 2003). Its formal definition is: Table 1 Networks statistics. astro-ph cond-mat hep-lat hep-th Papers 19,077 20,664 9,367 38,569 Authors 16,978 18,070 4,718 17,887 Collaborations 140,157 56,731 32,309 58,855 Density (10�4 ) 9.7252 3.4746 29.0355 3.6792 Avg. degree 16.51 6.28 13.70 6.58 Table 2 Class distribution for the validation and test sets. astro-ph cond-mat hep-lat hep-th (a) Validation set No. of pairs 745,683 149,769 330,137 498,782 Connected (+) 7,797 1,737 1,099 1,232 Non-connected (�) 737,886 148,032 329,038 497,550 Proportion (+/�) 1.06% 1.17% 0.33% 0.25% (b) Test set No. of pairs 1,337.607 254,651 373,053 553,544 Connected (+) 9,791 2,665 608 1,626 Non-connected (�) 1,327.816 251,986 372,445 551,918 Proportion ((+/�)) 0.74% 1.06% 0.24% 0.29% 2 arXiv.org e-Print archive – Cornell University Library. 6656 P.R.S. Soares, R.B.C. Prudêncio / Expert Systems with Applications 40 (2013) 6652–6660 Author's personal copy AAðu; vÞ¼ X z2CðuÞ\CðvÞ 1 logðjjCðzÞjjÞ ð13Þ Finally, the JC assumes higher proximity values for pairs of nodes which share a higher proportion of common neighbors relative to to- tal number of neighbors they have Salton and McGill (1986). JCðu; vÞ¼ jjCðuÞ\ CðvÞjj jjCðuÞ[ CðvÞjj ð14Þ The above measures were statically applied to the whole graph G (i.e., in the current state of the network) without considering tem- poral information. In our work, we also performed experiments with the time series approach for link prediction (Potgieter et al., 2009; Soares & Prudêncio, 2012) described in Section 2. As our cur- rent work, this approach also considers temporal information dur- ing the link prediction. Initially the network is split into frames, using the same procedure described in Section 3.1. Given a chosen proximity measure, a time series is built for each pair of nodes by applying the measure in each network frame. A forecasting model is then used in order to predict the next value of the series. Such forecast value is defined as the final proximity score of the pairs of nodes. Compared to the proposed work, we can point out a diffi- culty related to the time series approach, which is the choice of a good forecasting model. In fact, in the experiments performed in Soares and Prudêncio (2012) using co-authorship networks, the choice of the forecasting model strongly affected the link prediction performance. In our work, we performed experiments adopting AA as the measure to build the time series and the linear regression (LR) model as the forecasting model. That was the best combination of proximity measure and forecasting model in the experiments presented in Soares and Prudêncio (2012). Finally, we highlight that similarly to the proposed measures, the baseline measures described in this section (both the static ones and the time series approach) were applied for unsupervised link prediction. 4.3. First-round experiments As described in Section 3.3, the event-based measure is poten- tially affected by the rewards associated to each of the three events: conservative, innovative and regressive. In the first round of experiments, we aimed: (1) to evaluate the effect of rewards’ values in the performance of the proposed measure; and (2) to ver- ify whether these values can be properly chosen using a validation set. We deployed the validation set to empirically evaluate and select the best combination of rewards in the link prediction task. In this round, as mentioned in Section 3.2, i was fixed to 1.0 and the other two parameters c and r varied proportionally to i : c ¼f0:0; 0:25; 0:5; 1:0; 2:0g and r ¼f�2:0;�1:0;�0:5;�0:25g. We consider here a fixed value for a, since our focus was to inves- tigate the adopted rewards’ values. The value of a was set to 0.05 through a preliminary and exploratory battery of experiments. Table 3 shows the AUC value obtained by each combination of the rewards in the validation set for all networks. The best four ob- served results for each network are presented in bold. As it can be seen, the most effective combinations of reward did not vary so much from one network to another. The best results were obtained by parameters (c and r) around the same region (lower central) in the tables for all networks. Besides, these results indicate that conservative and regressive rewards can balance each other. The best performance values were achieved when c 6 i, although some good results were also obtained when we consid- ered a higher weight (2.0) to the conservative reward. In our exper- iments, if regressive reward is roughly weighted, the event-based method did not perform well (this can be noticed by analyzing results obtained with r ¼�2:0 and �1:0). As one could expect, this shows that if the aim is to predict new links in a network, then the history of innovative and conservative events assumes a higher de- gree of importance in the prediction process. Table 4 shows the AUC value obtained by each combination of the rewards in the test set for all networks. The results observed in the test sets did not varied substantially from those ones ob- served in the validation set. The best configurations of parameters were also similar considering the validation and the test set (see Table 5). This supports the conclusions we made regarding rewards and indicates that they can be determined empirically using a val- idation set. 4.4. Second-round experiments In the second round of experiments, our aim was to investigate how the amortization factor a affects the performance of the event-based measure. As explained in Section 3.3, this parameter indicates the importance of the secondary events in relation to pri- mary events in our proposed measure. In this experiment, we eval- Table 3 AUCs obtained by different combinations of rewards on the validation sets. r=c 0.0 0.25 0.5 1.0 2.0 (a) astro-ph �2.0 0.3402 0.3552 0.3698 0.4074 0.4681 �1.0 0.6301 0.6657 0.6773 0.6919 0.6972 �0.5 0.7931 0.7998 0.8012 0.8018 0.7996 �0.25 0.7967 0.8021 0.8035 0.8041 0.8023 (b) cond-mat �2.0 0.3002 0.3087 0.3181 0.3530 0.4239 �1.0 0.5684 0.6202 0.6318 0.6483 0.6507 �0.5 0.7928 0.8083 0.8131 0.8128 0.8092 �0.25 0.7964 0.8086 0.8109 0.8123 0.8105 (c) hep-lat �2.0 0.3792 0.4553 0.5801 0.6849 0.7273 �1.0 0.8002 0.8330 0.8294 0.8206 0.8102 �0.5 0.8406 0.8476 0.8467 0.8424 0.8340 �0.25 0.8309 0.8374 0.8389 0.8378 0.8331 (d) hep-th �2.0 0.2610 0.2843 0.3276 0.4264 0.5223 �1.0 0.6668 0.7599 0.76554 0.7678 0.7584 �0.5 0.8501 0.8579 0.8555 0.8502 0.8409 �0.25 0.8433 0.8497 0.8491 0.8469 0.8411 Table 4 AUCs obtained for several combinations of c and r on the test sets. r=c 0.0 0.25 0.5 1.0 2.0 (a) astro-ph �2.0 0.3462 0.3684 0.3916 0.4401 0.5116 �1.0 0.6423 0.6781 0.6875 0.6979 0.7006 �0.5 0.7844 0.7911 0.7922 0.7916 0.7879 �0.25 0.7861 0.7919 0.7927 0.7929 0.7909 (b) cond-mat �2.0 0.3041 0.3109 0.3259 0.3743 0.4303 �1.0 0.5731 0.6325 0.6408 0.6544 0.6535 �0.5 0.8072 0.8154 0.8147 0.8122 0.8064 �0.25 0.8078 0.8126 0.8127 0.8123 0.8091 (c) hep-lat �2.0 0.2945 0.3618 0.4934 0.6431 0.7353 �1.0 0.7652 0.8480 0.8429 0.8403 0.8349 �0.5 0.8483 0.8573 0.8574 0.8557 0.8515 �0.25 0.8393 0.8478 0.8506 0.8527 0.8508 (d) hep-th �2.0 0.2258 0.2376 0.2706 0.3583 0.4634 �1.0 0.6624 0.7303 0.7331 0.7303 0.7220 �0.5 0.8605 0.8625 0.8587 0.8529 0.8441 �0.25 0.8555 0.8558 0.8541 0.8518 0.8466 P.R.S. Soares, R.B.C. Prudêncio / Expert Systems with Applications 40 (2013) 6652–6660 6657 Author's personal copy uated the proposed measure by adopting different values of a ranging from 0 to 1:5, with an increment of 0:05. We highlight that when a ¼ 0, secondary events are in fact not taken into account. For higher values (such as a ¼ 1:5) in turn, possibly an excessive importance is given to secondary events, which may harm the prediction performance. In this section, the rewards c; r and i were fixed according to the best configuration observed in the validation set in the previous round of experiments. The AUC values obtained as a varies are illustrated in Fig. 2. When a assumes a null value, the performance of the model is harmed because secondary events are not being taking into account in the prediction task. In this case, the measure will only be able to predict connections between pairs affected by primary events any time in the past, that is, pairs that were connected in at least one frame of N . This shows the effectiveness of secondary events in the prediction of new connections. The best AUC values were obtained by adopting a ¼ 0:05, which is a relatively small value considering the range of values evaluated in our experiments. When it approaches to higher values, the method performs worse since surrounding pairs are becoming as important as the main pair of nodes in analysis. Besides, the num- ber of secondary events is much bigger than the total amount of primary events (mainly if the network is dense). Hence, isolated secondary events should contribute with small values, which when aggregated, can outstand in the predictive process without, how- ever, mask the effects of the primary events. In this section, we suggest a simple heuristic to define the value of a based on the ratio between the total number of primary events and secondary events in N (see Table 6). As it can be seen, the AUC results obtained by adopting the heuristic value for a were similar to the best results observed in Fig. 2. 4.5. Last-round experiments In the previous experiments we evaluated the event-based score focusing on its parameters. In this section, we evaluated it aiming to determine if the age of the temporal events (event-based score with Eq. (9)) can bring any gain of performance to the predic- tion task. Aditionally, we compared the event-based scores to the baseline measures described in Section 4.2. Table 7 present the Table 5 Best rewards settings extracted from experiments on validation and test sets. astro-ph cond-mat hep-lat hep-th (a) Validation set c 1.0 0.5 0.25 0.25 i 1.0 1.0 1.0 1.0 r �0.25 �0.5 �0.5 �0.5 (b) Test set c 1.0 0.25 0.5 0.25 i 1.0 1.0 1.0 1.0 r �0.25 �0.5 �0.5 �0.5 Fig. 2. Plot of AUC versus the amortization factor a. 6658 P.R.S. Soares, R.B.C. Prudêncio / Expert Systems with Applications 40 (2013) 6652–6660 Author's personal copy AUC value obtained by the proposed scores and the baseline meth- ods. In this table, we refer the function assigned to b to distinguish between the scores generated by using the two aforementioned equations. The results showed that the methods that deploy temporal information (i.e., the event-based scores and the time series approach) in general outperformed the classical static approach based on PA, CN, AA and JC. These results confirm that temporal information is actually an important aspect to consider for link prediction. The event-based scores obtained the best results from all methods considered. The performance gain was higher for the cond-mat, hep-lat and hep-th networks, but less expressive for the as- tro-ph network, in which the event-based score was quite similar to AA alone. In our results, we observed a performance gain when the event-based score using b ¼ log was compared to the event-based score using b ¼ 1 in all networks considered, what supports our belief that recent events carry more information about the emer- gence of links.. Although, the performance gain was in general small in absolute values, it was statistically verified using a t-test with a 95% level of confidence. Finally, although both the time series approach and the event- based methods explore the temporal nature of the network, the proposed approach obtained the best comparative results. The per- formance gain was statistically verified with a 95% level of confi- dence. The proposed measure analyses the network by focusing on the connections as such, and for that, it was able to extract more relevant information to the prediction of ties that might emerge in a close future. The method based on time series, on the other hand, assumes that the topological measure history shows some ten- dency as the network evolves. Thus, it performs an indirect analy- sis of the connections in the network. In most cases, the time series approach outperformed the static approach, but it was not able to overcome the results achieved by an analysis strictly focused on the links, as done by the method based on events. 5. Conclusion In this work we introduced a new proximity measure for link prediction based on the notion of temporal events. Differently from the classical static approach that takes the network at time t in order to predict new connections at future time t0, our method takes temporal information into account by monitoring how the network evolved along time. Our method consists of computing scores by aggregating rewards of the temporal events observed at each step in the network evolution, which resulted in more rep- resentative scores to the dyads under analysis. Different experi- ments were performed in four co-authorship networks, initially aimed to evaluate the robustness of the proposed method concern- ing its parameters. The obtained results suggest that the method was quite stable concerning the definition of its parameters across the different networks adopted, although more experiments need to be performed in other domains of application. Also, we verified the importance of considering secondary events in the proposed measure. In the performed experiments, we also compared the event- based score to other baseline approaches proposed in the literature of link prediction. The obtained results showed that the event- based measure outperformed the baseline measures considered, indicating that the combination of interactions between nodes and the dynamics of common neighborhood can bring a gain in performance to the prediction task. Additionally, when the age of the temporal events was taken into account, the method per- formed slightly better, indicating that more recent events can bring more useful information to the prediction process. In our work, we adopted a log function in order to give a higher weight to more recent events. Nevertheless, other functions can be evaluated in the future (e.g., linear) to weight the recent events. Also, different functions can be associated to each category of temporal event if we assume for instance that recent events from one category are actually more important that the other events. Finally, we highlight that the event-based score described here was specific to undirected and unweighted networks. A possible future line of research is to extend the ideas discussed here to more complex categories of networks. For instance, a variety of temporal events could be defined if the direction of the links is considered (e.g., a conservative event is observed in a direction but a regres- sive event is observed in the other direction, thus indicating a non-reciprocal interaction between the nodes at a given moment). Also, in the case of weighted networks, temporal events could con- sider an increase or decrease of the weight between the nodes along time, instead of just considering the existence of the links. The rewards in such cases should be carefully defined in order to adequately reflect the important of the events in the link predic- tion process. References Adamic, L. A., & Adar, E. (2003). Friends and neighbors on the web. Social Networks, 25(3), 211–230. Amaral, L. A., Scala, A., Barthelemy, M., & Stanley, H. E. (2000). Classes of small- world networks. Proceedings of the National Academy of Sciences USA, 97(21), 11149–11152. Barabasi, A-L., Jeong, H., Neda, Z., Ravasz, E., Schubert, A., & Vicsek, T. (2002). Evolution of the social network of scientific collaborations. Physica A, 311(3–4), 590–614. Box, G., & Jenkins, G. (1970). Time series analysis: Forecasting and control. San Francisco: Holden-Day. Bringmann, B., Berlingerio, M., Bonchi, F., & Gionis, A. (2010). Learning and predicting the evolution of social networks. Intelligent Systems IEEE, 25(4), 26–35. de Sá, H. R., & Prudêncio, R. B. C. (2011). Supervised link prediction in weighted networks. In Proceedings of the 2011 international joint conference on neural networks (pp. 2281–2288). IEEE. Fawcett, T. (2006). An introduction to roc analysis. Pattern Recognition Letters, 27, 861–874. Getoor, L., & Diehl, C. P. (2005). Link mining: A survey. SIGKDD Explorations Newsletter, 7(2), 3–12. Hasan, M. A., Chaoji, V., Salem, S., Zaki, M. (2006). Link prediction using supervised learning. In Proceedings of SDM 06 workshop on link analysis counterterrorism and security. Hasan, M. A., & Zaki, M. J. (2011). A survey of link prediction in social networks. In Charu C. Aggarwal (Ed.), Social network data analytics (pp. 243–275). US: Springer. Homans, G. C. (1951). The human group. London: Routledge and Kegan. Zan Huan, (2006). Link prediction based on graph topology: The predictive value of the generalized clustering coefficient. In Proceedings of 12th ACM SIGKDD international conference on knowledge discovery and data mining. Table 6 Proportion between the number of primary events (P) and secondary events (S). astro-ph cond-mat hep-lat hep-th P/S 0.0116 0.0391 0.0089 0.0238 AUC with a ¼ P=S 0.7953 0.8160 0.8664 0.8627 Best AUC 0.7948 0.8269 0.8681 0.8830 Table 7 Performance of the methods (AUCs). astro-ph cond-mat hep-lat hep-th Event-based (b ¼ log) 0.7948 0.8269 0.8681 0.8830 Event-based (b ¼ 1) 0.7929 0.8147 0.8573 0.8625 Time series (AA + LR) 0.7584 0.7402 0.8537 0.7899 AA 0.7844 0.7346 0.8049 0.7513 CN 0.7391 0.6744 0.7598 0.6838 JC 0.7299 0.6114 0.7408 0.6481 PA 0.5043 0.5455 0.5668 0.4834 P.R.S. Soares, R.B.C. Prudêncio / Expert Systems with Applications 40 (2013) 6652–6660 6659 Author's personal copy Huang, Z., & Lin, Dennis K. J. (2009). The time-series link prediction problem with applications in communication surveillance. INFORMS Journal on Computing, 21(2), 286–303. Jiang, C., Coenen, F., & Zito, M. (2013). A survey of frequent subgraph mining algorithms. The Knowledge Engineering Review, 28(1), 75–105. Juszczyszyn, K., Musial, K., Budka, M. (2011). Link prediction based on subgraph evolution in dynamic social networks, In Privacy security risk and trust (PASSAR) 2011 IEEE third international conference on social computing (pp. 27–34). Katz, L. (1953). A new status index derived from sociometric analysis. Psychometrika, 18(1), 39–43. Liben-Nowell, D., Kleinberg, J. (2003). The link prediction problem for social networks. In Proceedings of the 2003 international conference on information and knowledge management (pp. 556–559). Lichtenwalter, R. N., Lussier, J. T., Chawla, N. V. (2010). New perspectives and methods in link prediction. In Proceedings of the 16th ACM SIGKDD international conference on knowledge discovery and data mining (pp. 243–252). Lu, L., & Zhou, T. (2011). Link prediction in complex networks: A survey. Physica A, 390(6), 1150–1170. Murata, T., & Moriyasu, S. (2008). Link prediction based on structural properties of online social networks. New Generation Computing, 26(3), 245–257. Newman, M. E. J. (2001). Clustering and preferential attachment in growing networks. Physical Review E, 64. Newman, M. E. J. (2001). The structure of scientific collaboration networks. In Proceedings of the national academy of sciences of the United States of America (Vol. 98, pp. 404–409). Potgieter, A., April, K. A., Cooke, R. J. E., & Osunmakinde, I. O. (2009). Temporality in link prediction: Understanding social complexity. Journal of Emergence: Complexity and Organization, 11(1), 83–96. Qiu, B., He, Q., Yen, J. (2011). Evolution of node behavior in link prediction. In AAAI (pp. 1810–1811). Salton, G., & McGill, M. J. (1986). Introduction to modern information retrieval. McGraw-Hill Inc. Soares, P. R. S., Prudêncio, R. B. C. (2012). Time series based link prediction. In Proceedings of the 2012 international joint conference on neural networks (pp. 784–790). Tylenda, T., Angelova, R., Bedathur, S. (2009). Towards time-aware link prediction in evolving social networks. In Proceedings of the third workshop on social network mining and analysis, SNA-KDD ’09 (pp. 1–9). Wang, C., Satuluri, V., Parthasarathy, S. (2007). Local probabilistic models for link prediction. In Proceedings of the 2007 seventh IEEE international conference on data mining (pp. 322–331). Wasserman, S., & Faust, K. (1994). Social network analysis: Methods and applications (structural analysis in the social sciences). Cambridge University Press. Xiang, E. W., (2008). A survey on link prediction models for social network data. PhD thesis, PhD Qualifying Exam, Department of Computer Science and Engineering, The Hong Kong University of Science and Technology. 6660 P.R.S. Soares, R.B.C. Prudêncio / Expert Systems with Applications 40 (2013) 6652–6660 
work_zhwshju7gffhldgvwqz2ilg7gu	----	Knowledge modeling through computational agents: application to surveillance systems Knowledge modeling through computational agents: application to surveillance systems José M. Gascueña,1 Antonio Fernández-Caballero,1 Marı́a T. López,1 and Ana E. Delgado2 (1) Universidad de Castilla-La Mancha, Departamento de Sistemas Informáticos & Instituto de Investigación en Informática de Albacete, 02071 Albacete, Spain (2) Universidad Nacional de Educación a Distancia, Departamento de Inteligencia Artificial, E.T.S.I. Informática, 28040 Madrid, Spain Abstract: In this work the concept of computational agent is located within the methodological framework of levels and domains of description of a calculus in the context of different usual paradigms in Artificial Intelligence (symbolic, situated, connectionist, and hybrid). Emphasis in the computable aspects of agent theory is put, leaving open the possibility to the incorporation of other aspects that are still pure cognitive nomenclature without any computational counterpart of equivalent semantic richness. The ideas presented are shown as currently being implemented on semi-automatic surveillance systems. A case study of a mobile robot application for the detection and following of humans is described. Keywords: artificial intelligence, computational agents, knowledge engineering, surveillance 1. Introduction The concept of agent comes from the persistent attempt in Science and Engineering to modular- ize the knowledge necessary to specify a calcu- lus, and of the later attempt to progressively increase the level of complexity and autonomy, making it reusable (Sycara et al., 1996). Agent theory takes from Artificial Intelligence (AI) the general intention to approach the functionality of biological systems. Thus, there are adaptive, intelligent, and intentional agents, with learning capacity and equipped with a certain level of social organization that allows cooperation in the accomplishment of ‘group tasks’. In this sense, the objectives of multi-agent systems (MAS) and distributed AI (DAI) agree (Weiss, 1999). ‘Nature-inspired computation’ (Nunes, 2006) is also practically isomorphic to agent and MAS theory. This is true at the level of individual agents (‘organisms’) as well as at the level of social organizations in MAS (ants, bees, human societies, collective games, and so on). In this last case, to the functionalities demanded for the individual behavior, it is necessary to add an interaction language among agents that al- lows to share goals and to coordinate collective plans used to reach the goals. With the arrival of AI, the initial formulations of connectionism (artificial neuronal networks and modular theory of deterministic and prob- abilistic automata) are partially obscured by new symbolic formulations based in rules. Here the inference is understood as a search process in a states space. Thus, the idea of ‘actors’ appears as a concurrent computation model in distribu- ted systems (Agha, 1986), which along with object-oriented programming (Meyer, 1997) are the antecedents of the agents. Conceptually, it is Minsky (1961) who raises the social idea of DOI: 10.1111/j.1468-0394.2011.00609.x Article _____________________________ 306 Expert Systems, September 2011, Vol. 28, No. 4 c� 2011 Blackwell Publishing Ltd agency, essentially basing on ‘personification’ of the verbs used in natural language to describe the necessary processes for the execution of a certain activity. The strategy to describe in natural language an agent’s ‘beliefs’, ‘desires’ and ‘intentions’, and later to develop the formal counterpart of the linguistic terms, is still used at the present time (Russell & Norvig, 2002). In fact, a complete agent language is dominant in Software Engineering (SE) (Padgham & Zambonelli, 2006), and its importance also grows in Knowledge Engineering (Cuena et al., 2003) For instance, an agent-based framework is a plausible solution to achieve inter-operability between domain ontologies used by different knowledge-based systems (Orgun et al., 2008; Li & Yang, 2008). It is advisable to indicate that, as with the term of AI, in the agents field there is usually an abuse of excessively loaded cognitive nomenclature of anthropomorphous semantics. Finally, it is during implementation when algo- rithms and automata are translated into sentences written in an existing programming language. In this work we approach the general concept of agent from a computational perspective. We consider that an agent starts being a conceptual model, later it is reduced to a formal model, and inally to a machine with sensors, effectors and a control program. The rest of the work is struc- tured as follows. Section 2 introduces some methodologies and programming languages for computational agents. In section 3 the agent concept is located within the methodological framework of description levels and domains of a calculus. In sections 3.1 and 3.2, respectively, our vision of a computational agent’s concep- tual and formal model is explained. Finally sections 4 and 5 introduce the application of our ideas in the surveillance domain. 2. Some methodologies and programming languages for computational agents The emergence of a new paradigm in Computer Science involves establishing a solid theoretical base and the software tools necessary to demon- strate its viability. In agent-oriented soft- ware engineering (AOSE), different approaches about development methodologies and pro- gramming languages have evolved in parallel. First, a methodology needs to have defined the set of concepts used, a process that specifies what to do, the collection of models obtained, and the notation used to represent them (Bordi- ni et al., 2007). In the last decade, various research groups have proposed their own meth- odology to develop applications based on agents (Bergenti et al., 2004). Many proposals can be found in the literature to evaluate them. For example, Cernuzzi et al. (2005) analyze the process model (waterfall, evolutionary, incre- mental, spiral, transformations) and the phases covered in the methodologies. Sturm and Shehory (2004) analyze the concepts and prop- erties, the notations and modeling techniques, and the development processes used. The con- clusion reached is that definitely there is no methodology better than the others. Table 1 briefly summarizes the analysis of methodolo- gies ADELFE (Carole et al., 2003), PASSI (Cossentino, 2005), Prometheus (Padgham & Winikoff, 2004), INGENIAS (Pavón et al., 2005), Tropos (Bresciani et al., 2004) and O- MaSE (Garcı́a-Ojeda et al., 2008) according to the following topics: � Is it a general purpose methodology or on the contrary is it intended to apply to a particular type of system? � Does it reference any agent architecture in particular? The answer to this question is to find out if there is an agent architecture linked to the methodology or conversely whether the methodology gives the designer freedom to choose what he=she wants and even to create his=her own. � Are there guidelines provided for identifying agents? That is, does the development process offer some preliminary steps that lead to determine the types of agents in the system? This approach is particularly interesting be- cause in AOSE methodologies agents are regarded as first-order entities and therefore it is helpful to have guidelines to identify them. � Is there a tool that gives support? The avail- ability of a tool will offer the designer a c� 2011 Blackwell Publishing Ltd Expert Systems, September 2011, Vol. 28, No. 4 307 support necessary to create the models used in the methodology. � Does the tool provide support to follow the development process? That is to say, does the graphical interface refer to the develop- ment process? � Does the tool generate code in some agent language? � Does the tool provide utilities so that the user may add a new application that gener- ates code in the desired language? Moreover, the objective of agent programming languages and platforms is to facilitate the implementation of systems incorporating agent concepts (Unland et al., 2005). Table 2 shows some of the features present in languages JACK (Winikoff, 2005), JADE (Bellifemine et al., 2007), Jadex (Pokahr et al., 2005) and ICARO-T platform (Garijo et al., 2008): (1) the type of agents that can be implemented, (2) the language used, (3) how it performs reasoning, (4) whether or not the architecture is compliant to the FIPA standard, (5) whether or not there is a specific development environment, and (6) if it is open source or not. As shown in Table 1, most methodologies can be applied in developing any general purpose MAS. However, this is relatively true as they offer a final phase that includes models that are closely tied to specific agent architecture. There- fore, we believe that to develop an application, and in order to identify the types of agents and their interactions, the starting stages should be independent of the target architecture. Once you have reached this milestone, appropriate models are used according to the internal architecture of each agent under consideration. Lastly, it is Table 1: Agent-oriented software engineering methodologies ADELFE PASSI Prometheus INGENIAS Tropos O-MaSE Specific purpose Yes (Adaptive MAS) No No No No No Specific agent architecture Cooperative agents FIPA agents BDI agents Based in mental agents BDI agents No Guide for identifying agents Yes Yes Yes No Yes Yes Graphical support tool OpenTool (commercial) PTK (add-in for Rational Rose) PDT IDK TAOM4E AgentToolIII (aT3) Assistance for following the development process Yes No Yes No Yes Yes Code generation No JADE Jack JADE JADE, Jadex JADE Utility for creating new generators No No NO Yes (modules) No No Table 2: Agent programming languages JACK JADE Jadex ICARO-T Agents BDI – BDI Reactive and cognitive Language Java extension Java Java Java Reasoning BDI No BDI Automata and cognitive agents FIPA compliant NO Yes Yes No Development environment JDE No No No Open source No Yes Yes Yes 308 Expert Systems, September 2011, Vol. 28, No. 4 c� 2011 Blackwell Publishing Ltd possible to generate code for the language cho- sen for the implementation. It is also note- worthy that most methodologies consider the code generation phase for a given agent pro- gramming language. In ADELFE this defi- ciency has been corrected in version 2.0 (Rougemaille, 2008). Equally important is the possibility offered by INGENIAS, compared to other methodologies, to create new modules through a mechanism based on templates to generate code in the desired target language. The O-MaSE methodology does not assume any particular agents; however, the interaction protocols model conversations between two agents. So, it is difficult to understand how interactions take place at the system’s global level. Finally, notice that the special-purpose methodologies such as ADELFE should raise questions to allow the designer to determine if an application is a good candidate to be mod- eled like the specific MAS considered in the methodology. Regarding the languages discussed, notice that all of them facilitate the implementation of a type of agent, except JADE that is only a middle ware that facilitates communication between agents. In JACK a pre-compiler that takes as input files written in JACK language and gets as output Java code is needed. This code is finally used to compile and run the application. By contrast, in other languages shown in Table 2 this utility is not necessary, as programming is directly done in Java. JACK programs are developed using a specific development environment (JACK Devel- opment Environment or JDE). Here are specific details of the Prometheus methodology and the ICARO-T platform=frame- work. Prometheus is defined as a proper detailed process to specify, implement and test=debug agent-oriented software systems. It offers a set of detailed guidelines that includes examples and heuristics, which provides a better understanding of what is required in each step of the develop- ment. This process incorporates three phases: � The System Specification phase identifies the basic goals and functionalities of the system, develops the use case scenarios that illustrate its functioning, and specifies which are the inputs (percepts) and outputs (actions). � The Architectural Design phase uses the outputs produced in the previous phase to determine the agent types that exist in the system and how they interact. � The Detailed Design phase focuses on devel- oping the internal structure of each agent and how each agent will perform its tasks within the global system. Finally, Pro- metheus details how the entities obtained during the design are transformed into the concepts used in a specific agent-oriented programming language (JACK). The design process for Prometheus methodology is sup- ported by Prometheus Design Tool (PDT) (Padgham et al., 2008). ICARO-T (Garijo et al., 2008) is an open source framework (http://icaro.morfeo-project.org) that provides four categories of component patterns: agent organization pattern to describe the overall architecture of the system, cognitive and reactive agent patterns to model agent behavior, and resource patterns to encapsulate computing entities providing services to agents. More basic computing entities including com- ponents for building new agent and resource models are also available in the frame- work. Examples of these entities are abstract data types, specialized libraries, domain ontolo- gies, rule processors, buffers, and so on. The ICARO-T reactive agent conceptual architec- ture (Gascueña et al., 2010) is made up of three components: perception, control and actuation. The perception works as an event handling mechanism. It stores incoming events, delivering them to the control on demand. The control is modeled as an extended finite state machine that consumes events stored in the perception, and performs transitions by changing its internal state and invoking actions in the actuation model. Reactive agents behave like event-con- suming processes which change their internal state and execute operations according to their state transition table. The main advantage of the ICARO-T frame- work is that it provides to engineers not only c� 2011 Blackwell Publishing Ltd Expert Systems, September 2011, Vol. 28, No. 4 309 http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org http://icaro.morfeo-project.org concepts and models, but also customizable MAS design and java code fully compatible with SE standards, which can be integrated in the most popular IDEs. While other agent based platforms (Unland et al., 2005), such as FIPA, focus on communication standards, ICARO-T focus on providing high level software compo- nents for easy development of complex agent behavior, agent coordination, and MAS organiza- tion. An additional advantage for ICARO-T is cost-effectiveness of agent patterns in application development. Evaluation results in previous ex- periences (Garijo et al., 2004) have reported significant reductions by an average of a 65% of time and effort in the design and implementation phases. Cost reduction was achieved without minimizing or skipping activities like design, doc- umentation and testing. In the phases of testing and correction the errors were also reduced. 3. Levels and domains in computational agent models Since the introduction of the knowledge level by Newell and Simon (1972) and Marr (1982) – called ‘theory of calculus’ – it is usual to describe the knowledge necessary to understand any calculation in three levels: physical (PL), sym- bols (SL) and knowledge (KL). Or, in a simpler way: machine hardware, programs and models, and algorithms. Starting from the idea of refer- ence systems in Physics and the proposals by Maturana (1999) and Varela (1979) in Biology, Mira and Delgado (1987) introduces the figure of the external observer in computation. This gives rise to two description domains of the organizations and relations at each level: (1) the own domain of the level (OD), where the caus- ality is intrinsic and the semantics comes im- posed by the structure and dynamics of the level, and, (2) the external observer domain (EOD) to the computation carried out in that level, where the semantics is arbitrary and the interpretation of the calculation depends of the observer and, in general, of the application domain. When superposing both domains (OD, EOD), at the three levels (KL, SL, PL), we obtain a three-storey house and six apartments (two by floor), in which the agents reside. Thus, in this framework, a user has available six types of knowledge – (PL-OD), (PL-EOD), (SL-OD), (SL-EOD), (KL-OD) and (KL-EOD), respec- tively – to describe agents. The floor where an agent resides suggests the agent name. So, we can state that there are three types of agents (physical agent, symbolic agent, and knowledge agent), which can be described from a point of view related to observer domain and another one related to external observer domain. The case study used in this paper to illustrate the application of these concepts is a simplified mobile robot application to detect and follow humans. We rely on the case study presented in a recently published article (Gascueña & Fernández-Caballero, 2009a). Several devices are mounted on the robot to carry out this functionality. Sonar, bumpers, camera, and wheels are used for navigation, obstacle avoid- ing, collision detection, and human detection. Moreover, we suppose that the camera is able to at most detect one human at a given instant. Figure 1 shows as an example the three nested levels (PL, SL, KL) and the two domains (EOD, OD) of description of the calculus related to the camera device. The different concepts depicted on the figure will be described along the case study description (see section 5). In short, the main contribution of our meth- odological framework is the application of two different points of view (observer domain and external observer domain) to describe the types of agents identified at the three levels (physical, symbols and knowledge) to understand any calculation. Mira and Delgado (2003) state that habitually the programmer reduces the conceptual model (KL-EOD) to a formal model (still at the KL, but now in the OD) and the corresponding program (in the OD of the SL). Then a compiler takes care of the last reduction step moving the program (SL) to the physical level (PL) in order to be executed into logical circuits. The two next subsections offer our vision of a computational agent’s conceptual and formal model, respec- tively. The two next subsections offer our vision 310 Expert Systems, September 2011, Vol. 28, No. 4 c� 2011 Blackwell Publishing Ltd of a computational agent’s conceptual and for- mal models, respectively. 3.1. Computational agent conceptual model (KL-EOD) Like in AI and Robotics, in the agency three basic architectures are distinguished – sym- bolic, situated and connectionist – to approach different solutions to a problem, by modeling data and knowledge and, later, operating the inferences of the model. An agent is symbolic when it uses declaratory and explicit knowl- edge in natural language to describe the con- stituent organizations (‘concepts’) and the inference rules. In Robotics this is associated to ‘deliberative architectures’, which spend time and a high number of computational resources in the decision process. An agent is reactive or situated when knowledge represen- tation is within two configuration tables of pre- calculated input and output, usually called Figure 1: Camera levels. c� 2011 Blackwell Publishing Ltd Expert Systems, September 2011, Vol. 28, No. 4 311 ‘perceptions’ and ‘actions’. Here, the inference procedure is of ‘reflex’ type, very fast and adapted for real-time applications. The percep- tion-action link is also given by a table or an automaton with few states. It is proper for monitoring tasks and for the execution phase of motor planning, where a command is de- composed into a set of pre-calculated elemen- tary actions that execute in ‘efficient’ time, without having to deliberate. An agent is con- nectionist when knowledge representation is given in terms of labeled numerical lines, as much for the inputs as for the outputs, and the inference functions are adjustable numerical associators (ANAs). It is difficult to have all the necessary knowledge for an application. For that reason, the most frequent situations demand hybrid solutions, with reactive and deliberative, and with symbolic and connec- tionist parts. For that reason, in the agency paradigm, there are also hybrid architectures that combine agents of reactive and delibera- tive type. The reactive part reacts to the events of the environment without reasoning, whereas the deliberative part support plans and per- forms tasks of a higher abstraction level. Practically all conceptual agent models com- ply in the scheme of Figure 2. The figure is inspired in the Luria’s proposals (Luria, 1973) on the different functional systems (FS) that cooperate in the interpretation and the control of the set of activities that an alive being devel- ops. The scheme consists of four FS basic types (Mira, 2006): (1) sensorial FS (perception), (2) motor FS (action), (3) association and decision FS (decision and planning), and (4) adaptation and learning FS. Another system (tone and watch regulator) is superimposed responsible for regulating the state of tone and supervises all the rest. The adaptation and learning FS, like the memory, is distributed on all other FS. This agent model is also analogous to the proposal by Newell (1980) with the name of ‘intelligent agent’. It also agrees with the models usual in Robotics. The feedback loops introduced here were proposed by McCulloch and Pitts (1939) to give support to the homeostasis, autonomy, purpose and intentionality concepts. These con- cepts are difficult to formulate and to model computationally without using the proposed three nested feedback levels. The type I loops are closed within each FS (sensorial, motor and Figure 2: Luria functional systems and McCulloch-Pitts feedback loops, integrated to propose a general agent model interacting with his means in which, simultaneously, there may be other agents. 312 Expert Systems, September 2011, Vol. 28, No. 4 c� 2011 Blackwell Publishing Ltd association) and between two or more of these FS. The type II loops are closed through the rest of the agent body (proprioception and regula- tion – homeostasis – of the internal means) and the loops of type III are closed through the external environment, using the specific sensors and effectors of each type of agent to physically and formally connect it to their environment. The type I loops have to do with all the mechanisms necessary to model the perception, association and action tasks. The type II loops are associated to the reflex mechanisms, the operating and instrumental conditioning, the homeostatic (regulation of the set of variables that characterize the internal state of the agent at all levels) and autopoietical (continuous synthesis of the compo- nents necessary to maintain the identity of the agent) mechanisms. Finally, the type III loops have to do with perception-action integration and ‘voluntary’ (autonomous) control of the external behavior of the agent. The computational version (KL-OD) of these organizations is far from the meaning in humans. The great number of type I, II and III loops that exist in biological systems, along with memory and structural and functional plasticity, is an indispensable requirement to develop adaptive and intelligent capacities in these systems. Consequently, the degree of auton- omy and the capacity of adaptation and intelli- gence of an agent depend on the feedback mechanisms that are implemented for each con- crete case. Considering the paradigms of AI as a classification criterion of different agent types, it is easy to demonstrate that there are only two basic types: (1) the ones based on descriptions in EOD that use declarative knowledge in natural language and (2) those based on mechanisms of the OD, causal in the implementation of the three mentioned levels. 3.2. Formal models for agents Whichever be the final version of the agent at knowledge level (KL-EOD), the following phase is to operationalize its entities and relations, its data and inferences. A formal model widely used for the description of abstract agent architectures is automata theory, leaving open the specification phase of the functions of state change ( f ) and output production (g). Let us remember that an automaton can be specified as follows: A¼ðX;Y;S; f;gÞ ð1Þ where X is the finite set of possible inputs (xi, i¼1 . . . n), Y is the finite set of possible out- puts (yk, k¼1 . . . p), S is the finite set of possible internal states (Sj, j¼1 . . . m), f and g are two sets of decision rules that represent the dynamics of the system in the production of new states (rules f ) and in the production of outputs (rules g). The function of production of new states, f, is defined in extensive as an application of the Cartesian product X � S on S, and the function of production of outputs, g, is an application of the Cartesian product X � S on Y. X�S ! f S : SðtþDtÞ¼ f ½xðtÞ;SðtÞ� ð2Þ X�S !g Y : yðtþDtÞ¼g½xðtÞ;SðtÞ� ð3Þ That is to say, the new state, S(tþDt), is a function, f(x, S), of the current state, S(t), and of the input, x(t). In front of a same input, different state transitions can take place. As well, the same input x(t) and the same state S(t) participate through another function, g(x, S), in the produc- tion of the outputs. These output variables are the responses of the automata to the disturbances, x(t), that are received from the external environ- ment, which as well is another automaton. That is to say, a finite automaton always interprets itself in its relation with external environments so that the outputs of the automaton are the inputs of the environment and vice versa. In their formulation at knowledge level, the inputs, xi, the outputs, yk, and the states, Sj, are class labels that represent the static and dynamic roles of the inferences in which a certain method decomposes the reasoning pro- cess. Their meaning is established in agreement with a semantics table, dependent of the ontology of the application domain. Analogously, functions f and g are the inferences and the corresponding control structures. In their formulation at symbol level the inputs and the states are, again, labels associated to the entities of the programming language. Finally, at physical level, the inputs, the c� 2011 Blackwell Publishing Ltd Expert Systems, September 2011, Vol. 28, No. 4 313 states and the outputs are levels of tension in analogical or digital electrical signals, and func- tions f and g are the electronic mechanisms that connect those signals. Let us remember that the input and output spaces are, in general, representation spaces. Each input, state or output is associated to a label whose specification in the meaning tables express in natural language a description of the situation that is being represented. Following the commonest agent definition Franklin and Graesser (1996), literally, ‘an autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in the future’, the formal description of an agent could be decomposed into several processes or phases: 1. Identification of the goals (or objectives) and of the tasks associated to the attain- ment of these goals. 2. Identification of sets X (‘perceptions’) and Y (‘actions’) that formally represent the means of interest for a concrete agent. 3. Identification of set S of internal states necessary to describe the agent’s dynamics. The name and type of the labels used in EOD to specify and to classify these states depend on the used agent architecture. For example, in BDI (Georgeff et al., 1998) the beliefs, desires and intentions are the three types of basic labels whose description in extensive gives rise to the elements of the set S of internal states. 4. Specification of functions f and g used to describe the necessary inferential rules so that the agent reaches his goals. 5. Specification of the models, algorithms and learning mechanisms used to modify func- tions f and g. 4. Application to surveillance In the last few decades, the field of surveillance systems has captured the attention of industry and research (e.g. Mira et al., 2004; López- Valles et al., 2007; Fernández-Caballero et al., 2008a, 2009; Delgado et al., 2010). The coopera- tion that emerges directly from the sociability characteristic is an agent’s distinguishing char- acteristic (Wooldridge & Jennings, 1995). The capability to communicate makes it possible for agents to work together to solve complex pro- blems which cannot be dealt with by a single agent, this being the essence of MAS (Huhns & Stephens, 1999). MAS are noted for the fact that they are made up of collections of potentially independent and autonomous agents, usually heterogeneous, which work together to solve a problem which goes beyond their individual capabilities. MAS are appropriate within domains in which the necessary knowledge to solve a problem is dis- tributed along different places. The solution to the problem depends on the coordination of the tasks to be carried out by different entities with different capabilities, usually without the super- vision of a single centralized coordinator. For example, defense applications are performed in highly decentralized and heterogeneous envir- onments and=or require the incorporation of intelligent decision making. These characteris- tics make the technologies, techniques and algo- rithms used within the scope of MAS adequate to be applied in military domain applications (Pechoucek et al., 2008) such as logistics, manned and unmanned air traffic control, simu- lation and training. Likewise, on the one hand, Patricio et al. (2008) highlight the suitability of using a MAS for video surveillance because (1) the loose coupling nature of a multi-agent architecture allows more flexibility in the communication process, and, (2) the ability to assign responsi- bilities to each agent is ideal to solve complex tasks in a surveillance system. These complex tasks entail the use of coordination and coop- eration mechanisms and dynamic configuration, which are widely used in the MAS community (d’inverno et al., 1997). On the other hand, intelligence distribution in MAS allows dealing with questions that turn up in the development of surveillance systems (bandwidth, productiv- ity, speed, robustness, autonomy, scalability) (Pavón et al., 2007). In summary, the basic 314 Expert Systems, September 2011, Vol. 28, No. 4 c� 2011 Blackwell Publishing Ltd characteristics of the agents suggest that they are good choices to solve the problems dealt with in surveillance systems, such as it will be approached through a study case in this section. Recently, the use of agent technology in surveil- lance has been revised (Gascueña & Fernández- Caballero, 2009b). All the previous concepts are being applied in semi-automatic visual surveillance tasks (López et al., 2006c, 2007; Valencia-Jiménez & Fernández- Caballero, 2006; Gascueña & Fernández- Caballero, 2007; Pavón et al., 2007; Martı́nez et al., 2008) composed of a set of collaborative cameras installed in a building and a camera- mounted mobile robot to offer pre-alarms and=or alarms detected indoor and outdoor. The video images captured by each of the cameras enable segmenting and tracking (López et al., 2006a, 2006b; Fernández-Caballero et al., 2008b; Moreno-Garcia et al., 2010) objects of interest (obtained as image blobs) with the objective of providing meaningful events and suspicious activ- ities (e.g. people roaming or abandoning an object). The cameras collaborate in the sense of obtaining richer surveillance observations that are only available through the fusion of information captured on various places. Mobile robots could be equipped with different sensors in order to perform surveillance tasks because they are able to obtain a vision of the objects of interest from a different perspective, and to accede to zones that are inaccessible to fixed cameras or that are dangerous for the humans. In particular, the next section offer a concrete case, in which methodolo- gical framework concepts described in the previous sections are put into practice, that introduces the design of a mobile robot application for the detec- tion and following of humans with a MAS per- spective. It is based in Prometheus methodology (Padgham & Winikoff, 2004). The Prometheus methodology has been chosen because it provides a collection of guidelines helping to determine the elements (for instance, agents and interactions) that form the MAS. These guidelines are also helpful to the experts in MAS development. They will be able to transmit their experience to other users through explaining why and how they have obtained the different elements of the agent-based application. In addition, Prometheus is also useful as it explicitly considers agent perceptions and actions as modeling elements. In Robotics, per- cepts are environment data collected by several robot sensors (temperature, light, distance, etc) and actions represent the control carried out by the robot actuators (motors, LEDs, and so on). Traditionally, proposals for agent architec- tures are categorized as reactive, cognitive, or hybrid just as described in section 3.1. Surveil- lance systems are sufficiently complex to encom- pass heterogeneous agent architectures. Therefore, the internal structure of each identi- fied agent should be modeled and implemented using the most suitable technology according to the functionality to be offered. For instance, in Robotics it is usual to find reactive, deliberative and even hybrid components (Qureshi et al., 2004). For this reason, in our approach the Prometheus Detailed Design phase will not be followed, as it imposes the usage of a specific agent-based model or architecture. Specifically, the artifacts produced by this phase are con- ceived with the BDI agent architecture in mind. 5. Mobile robot application for the detection and following of humans The process to detect and follow moving objects used by the robot is depicted in Figure 3. In the following, a brief description of each state is introduced: � Firstly, the robot is moving randomly around the environment (state Wander) while movement has been not detected by the camera. � On the one hand, when the information pro- vided by the camera contains a region of interest (ROI) characteristic of a human then the robot follows him=her (state Follow). The robot remains in this state until the camera loses the target, then it transits to state Wander. � On the other hand, when the robot is wan- dering or following a human and the sonar informs that there is some obstacle at a minimum distance from the robot then the robot has to avoid it (state Avoid Obstacle). c� 2011 Blackwell Publishing Ltd Expert Systems, September 2011, Vol. 28, No. 4 315 In this new state, the robot executes actions to stop and to orient it towards a new direction in order to avoid the obstacle detected. Moreover, the robot considers also information provided by the bumper in or- der to move back if a collision is detected (state Move Back). Then, when the problem has been solved (SonarAvoidedObstacle) the robot again wanders or follows a human according to the information provided by the camera (CameraNoHumanDetection and CameraHumanDetection, respectively). � Notice that for not damaging the robot the information provided by the sonar has a greater priority than the camera informa- tion. For example, if the robot is wandering, the camera detects a human and the sonar informs that there is an obstacle at a mini- mum distance then the robot has to avoid the obstacle instead of following the human. 5.1. System specification In the first phase of the Prometheus methodol- ogy, namely the System Specification phase, the analysis overview diagram is developed, which shows the interactions between the system and the environment (see Figure 4). Several entities are identified in the diagram: � Actors. An actor is an external entity – human or software=hardware – that inter- acts with the system. At this level, an actor for each device mounted on the robot (sonar, camera, bumpers, and wheels) has been identified. � Percepts. The information that comes from the environment has been identified as percepts. It corresponds to distance to obstacles=targets perceived by the sonar (Distance_P), images captured by the cam- era (Image_P), and impacts detected by the bumper device (Collision_P). � Actions. Every operation performed by the system on the actors is identified as an action. It corresponds to the commands to control wheel motion (Set direction_a, Stop_a, Move_a). � Scenarios. A scenario is a sequence of struc- tured steps – labeled as action, percept, goal, or other scenario – that represents a possible execution way of the system. There are three Figure 3: The robot’s states. 316 Expert Systems, September 2011, Vol. 28, No. 4 c� 2011 Blackwell Publishing Ltd scenarios (Wandering motion scenario, Hu- man following scenario and Avoiding obstacle scenario) that correspond to the main states of the robot. Notice that the concept of an actor, when referred to a hardware entity, is directly considered in our methodological framework as a physical agent. The idea proposed to model an application com- posed of several physical devices consists in asso- ciating, firstly, an actor (physical agent – PL) to each physical device. In this case, for instance, the Camera_PL agent is the physical agent associated to the camera device. 5.2. Architectural design Once the system requirements and the environ- ment of the problem have been specified in the previous phase, the tasks carried out in the next phase of the Prometheus methodology (that is, the Architectural Design phase) are to decide what kind of (new) agents the system will have and how the interaction between them will be. These are specified in a system overview diagram. For example, Figure 5 depicts an excerpt of the system overview diagram, where only the agents related to the camera are identified (that is, Camera_PL, Camera_SL, and Camera_KL agents) as well as their interactions to control the robot state. The Camera_PL physical agent, which was already identified in the System Speci- fication phase, sends images (Image_P percept) to the Camera_SL agent. The last agent is respon- sible of performing an image segmentation pro- cess to look for a region of interest representing a human (ROI_D data). This information is used by the Camera_KL agent to communicate if a human has been detected to the robot move- ment manager (RobotControl_KL agent). These communications are specified inside the Camera Robot_IP interaction protocol. 5.3. Agent internal structure The internal structure of each agent identified previously has now to be developed. Obviously, usually there is no single agent model able to be applied for all agents. Regarding the example of the physical agent (Camera_PL), you may observe that it is de- scribed from the two points of view considered in our framework (see the bottom of Figure 1). On the one hand, the pixels of the image captured by the camera represent the informa- tion of the physical agent from the point of view of the observer domain (OD). On the other Figure 4: The Prometheus Analysis Overview Diagram. c� 2011 Blackwell Publishing Ltd Expert Systems, September 2011, Vol. 28, No. 4 317 hand, the image visualization on a monitor is its representation related to the external domain observer (EOD). In the middle of Figure 1 it may be observed that the Camera_SL agent is labeled as a sym- bols agent. From the point of view of the OD sentences of a segmentation algorithm (if . . . then . . . else . . .) are incorporated. On the other hand, the image visualization on a monitor, including a possible square to highlight the region of interest that characterize a human through a series of parameters, is a representa- tion related to the EOD. As explained in this paper, the operationali- zation of the computational agent conceptual model (EOD-KL) may be formally described as a finite state automaton (see section 3). There- fore, it is necessary to select an agent-based language=framework that offers support to the implementation. In this sense ICARO-T (Garijo et al., 2008) satisfies our purposes. So, the model of an ICARO-T reactive agent represents an EOD-KL agent in our framework. Camera_KL agent is located at the knowledge level in our framework (see the top of Figure 1). This fragment is shown again for readability purposes (see Figure 6). Its responsibility is informing the RobotControl_KL agent when a human has been detected. So, on the one hand, from the point of view of the OD the function- ality of this agent can be described with an automaton composed of two states. The inter- pretation is summarized as follows: � It remains in S0 state whilst no region of interest (ROI) is detected in two consecutive frames. � It remains in S1 state whilst a ROI is detected in two consecutive frames. � It transits from S0 to S1 when a ROI is detected in the current frame. � It transits from S1 to S0 when no ROI is detected in the current frame. On the other hand, from the point of view of the EOD a meaning is provided at each transi- tion using concepts of the ICARO-T reactive agent models (see image located on the left in Figure 6): � S0 state is labeled as NoHuman to point out that no human is detected. � S1 state is labeled as Human to point out the opposite situation to NoHuman state, that is, a human is detected. � Human and NoHuman labels allow easily associating concrete situations of the agent life cycle. � Three new states have been introduced due to ICARO-T implementation features, namely Initial, S-1 and Final states. � The transition from Initial to S-1 state is the first executed one. It is satisfied when the manager agent, which is already implemen- ted in ICARO-T framework, creates the Camera_KL agent. In this case, the transi- tion execution produces the activation of the camera and sends itself an event startDetec- tion. The event allows that the Camera_KL agent starts the detection process and sends a first event CameraNoHumanDetection to the robot to communicate that a human is not detected. After that, Camera_KL tran- sits to NoHuman state. Figure 5: The Prometheus System Overview Diagram. 318 Expert Systems, September 2011, Vol. 28, No. 4 c� 2011 Blackwell Publishing Ltd � The other transitions are related to the transitions belonging to the automaton of the Camera_KL agent described from the point of view of the OD. The external ob- server considers that (1) when transition S0 to S0 takes place, then the agent does not need to communicate to the robot that a human is not detected, as this was already communicated previously; (2) when transi- tion S0 to S1 takes place, the agent needs to communicate the detection of a human; (3) when transition S1 to S0 takes place, the agent communicates that a human is not detected; and (4) a similar reasoning to situation (1) is carried out for transition S1 to S1. � Finally, a particular transition called univer- sal transition, which is valid for any state of the automaton, is used. This transition takes place for a given event. The action is exe- cuted and the automaton transits to the next state, regardless of the automaton’s state. In the Camera_KL agent, if stopDetection event is received then the deactivation of the cam- era is produced. The necessary mechanisms for agent percep- tion and control are already implemented in ICARO-T (reactive agent pattern). Therefore, to implement the behavior for each reactive agent the developer only needs to specify the state transition table in XML and the actuation model (semantic actions). On the other hand, it is also necessary to highlight that the automaton depicted on the left in Figure 6 is formally defined in an equivalent way using the notation presented in section 3.2. In this case the auto- maton is deterministic. That is to say, in front of an input only one state transition can take place. For the reactive agent to be capable of interpret- ing the graphically=formally represented auto- maton it is expressed it in a textual way through an XML file named ‘automaton.xml’. On the other hand, actions executed by a reactive agent are defined as methods of a semantic actions class. Keeping in mind section 1, firstly, the system concepts are defined in natural language terms. Secondly, it is necessary to formalize them. Finally, the formal model is translated to some programming language. Therefore, the model entities are associated to concepts used in the selected programming language. Table 3 shows which model entities are translated into their equivalent ICARO-T implementation concepts. We have to highlight that the actor (physical Figure 6: Camera_KL knowledge agent. c� 2011 Blackwell Publishing Ltd Expert Systems, September 2011, Vol. 28, No. 4 319 agent) concept is associated with the resource concept, which can be used by the agents to interact with the environment (execute actions) and it sends events towards the agents to trans- mit the information gotten from the environ- ment (receive percepts). The symbols agent concept is also implemented as a resource. In this way it easily offers to other agents or resources the information obtained. They access the information using the interface of use of the resource. The knowledge agent concept is implemented as a reactive agent automaton as illustrated in the case study. The percept and message concepts used in the modeling to repre- sent a communication between agents are both translated as an event. 6. Conclusions In this paper the agent concept has been faced from a computational perspective. The concept of computational agent is located within the methodological framework of levels and do- mains of description of a calculus in the context of different usual paradigms in AI (symbolic, situated, connectionist, and hybrid). Therefore, it has been shown that a computational agent must specify its conceptual model, its formal model and its implementation, starting from the set of functional specifications available on its goals, activities and tasks. Moreover, the points of view from the observer domain and external observer domain are illustrated to describe the structure of the specific agents identified, in the mobile robot case study, at the three different levels (physical, symbols and knowledge) neces- sary to understand any calculation. We are currently engaged in modeling the visual surveillance task. Surveillance systems consist of a great diversity of entities that have to cooperate in highly dynamic and distributed environments. The use of agents for their control allows a greater degree of autonomy and response because of their capabilities to adapt and to cooperate. This is why, after introducing some methodologies and programming languages for computational agents, the ideas presented are applied on semi- automatic surveillance systems. We have used Prometheus as agent-oriented SE methodology and ICARO-T framework as development frame- work. Two case studies are described to exemplify the concepts introduced. Thus, a mobile robot application for the detection and following of humans and a collaboration application between surveillance cameras are provided. Acknowledgements This work was partially supported by Spanish Ministerio de Ciencia e Innovación TIN2010- 20845-C03 grant, and by Junta de Comunidades de Castilla-La Mancha PII2I09-0069-0994 and PEII09-0054-9581 grants. This article is dedicated to the memory of Professor José Mira, a great researcher, a wise man, a loving husband, and a close friend; but who sadly passed away on August 13, 2008. This paper is an extended version of a recent conference paper (Mira et al., 2009). References AGHA, G. (1986) Actors: A Model of Concurrent Com- puting in Distributed Systems, Cambridge, MA: The MIT Press. Table 3: Mapping from modeling concepts to ICARO-T implementation concepts Modeling concept ICARO-T concept Knowledge agent Reactive agent Symbols agent Resource Physical agent (Actor) Resource Percept Event Message Event Action Used in a semantic action and defined as a method in a resource Percept and message Used as input in the automaton definition 320 Expert Systems, September 2011, Vol. 28, No. 4 c� 2011 Blackwell Publishing Ltd BELLIFEMINE, F., G. CAIRE and D. GREENWOOD (2007) Developing Multi-Agent Systems with JADE, New York: John Wiley & Sons Ltd. BERGENTI, F., M. GLEIZES and F. ZAMBONELLI (2004) Methodologies and Software Engineering for Agent Systems: The Agent-Oriented Software Engineering Handbook, Berlin: Springer-Verlag. BORDINI, R., M. DASTANI and M. WINIKOFF (2007) Current issues in multiagent systems development, Lecture Notes in Artificial Intelligence, 4457, 38–61. BRESCIANI, P., P. GIORGINI, F. GIUNCHIGLIA, J. MYLO- POULOS and A. PERINI (2004) Tropos: an agent- oriented software development methodology, Auton- omous Agents and Multi-Agent Systems, 8, 203–236. CAROLE, B., G. MARIE-PIERRE, P. SYLVAIN and P. GAUTHIER (2003) Adelfe, a methodology for adap- tive multi-agent systems engineering, Lecture Notes in Artificial Intelligence, 2577, 156–169. CERNUZZI, L., M. COSSENTINO and F. ZAMBONELLI (2005) Process models for agent-based development, Engineering Applications of Artificial Intelligence, 18, 205–222. COSSENTINO, M. (2005) From requirements to code with the PASSI methodology, chapter 4, in Agent- Oriented Methodologies, B. Henderson-Sellers and P. Giorgini (ed), Hershey, PA: Idea Group Publish- ing, 79–106. CUENA, J., Y. DEMAZEAU, A. GARCIA-SERRANO and J. TREUR (2003) Knowledge Engineering and Agent Technology, Amsterdam, the Netherlands: IOS Press. DELGADO, A.E., M.T. LÓPEZ and A. FERNÁNDEZ- CABALLERO (2010) Real-time motion detection by lateral inhibition in accumulative computation, Engineering Applications of Artificial Intelligence, 23, 129–139. D’INVERNO, M., M. LUCK and M.J. WOOLDRIDGE (1997). Cooperation structures, in Proceedings of the Fifteenth International Joint Conference on Artificial Intelligence, 23–29 August, 1997, Nagoya, Japan, pp. 600–605. FERNÁNDEZ-CABALLERO, A., F.J. GÓMEZ and J. LÓPEZ- LÓPEZ (2008a) Road-traffic monitoring by knowl- edge-driven static and dynamic image analysis, Ex- pert Systems with Applications, 35, 701–719. FERNÁNDEZ-CABALLERO, A., M.T. LÓPEZ, J.C. CASTILLO and S. MALDONADO-BASCÓN (2009) Real-time accu- mulative computation motion detectors, Sensors, 9, 10044–10065. FERNÁNDEZ-CABALLERO, A., M.T. LÓPEZ and S. SAIZ- VALVERDE (2008b) Dynamic Stereoscopic Selective Visual Attention (DSSVA): integrating motion and shape with depth in video segmentation, Expert Systems with Applications, 34, 1394–1402. FRANKLIN, S. and A. GRAESSER (1996) Is it an agent, or just a program?: a taxonomy for autonomous agents, Lecture Notes in Computer Science, 1193, 121–135. GARCÍA-OJEDA, J., S. DELOACH, ROBBY, W. OYENAN and J. VALENZUELA (2008) O-mase: a customizable approach to developing multiagent development processes, Lecture Notes in Computer Science, 4951, 1–15. GARIJO, F., S. BRAVO, J. GONZALEZ and E. BOBADILLA (2004) BOGAR_LN: an agent based component framework for developing multi-modal services using natural language, Lecture Notes in Computer Science, 3040, 207–220. GARIJO, F., F. POLO, D. SPINA and C. RODRÍGUEZ (2008). Icaro-t user manual, Technical report, Tele- fonica IþD. GASCUEÑA, J.M. and A. FERNÁNDEZ-CABALLERO (2007) The INGENIAS methodology for advanced surveil- lance systems modelling, Lecture Notes in Computer Science, 4528, 541–550. GASCUEÑA, J.M. and A. FERNÁNDEZ-CABALLERO (2009a) Agent-oriented modeling and development of a person-following mobile robot, Expert Systems with Applications, 38, 4280–4290. GASCUEÑA, J.M. and A. FERNÁNDEZ-CABALLERO (2009b) On the use of agent technology in intelligent, multi- sensory and distributed surveillance, The Knowledge Engineering Review, 26, 191–208. GASCUEÑA, J.M., A. FERNÁNDEZ-CABALLERO and F.J. GARIJO (2010) Using ICARO-T framework for reactive agent-based mobile robots, Advances in Practical Applications of Agents and Multiagent Sys- tems, 70, 91–101. GEORGEFF, M.P., B. PELL, M.E. POLLACK, M. TAMBE and M. WOOLDRIDGE (1998). The belief-desire-inten- tion model of agency, in Intelligent Agents V, Agent Theories, Architectures, and Languages, 5th Interna- tional Workshop, ATAL ’98, 4–7 July, 1998, Paris, France, pp. 1–10. HUHNS, M.N. and L.M. STEPHENS (1999) Multiagent systems and societies of agents, in Multiagent Sys- tems, G. Weiss (ed), Cambridge, MA: The MIT Press, 1–42. LI, L. and Y. YANG (2008) Agent-based ontology mapping and integration towards interoperability, Expert Systems: The journal of knowledge engineer- ing, 25, 197–220. LÓPEZ, M.T., A. FERNÁNDEZ-CABALLERO, M.A. FER- NÁNDEZ, J. MIRA and A.E. DELGADO (2006a) Motion features to enhance scene segmentation in active visual attention, Pattern Recognition Letters, 27, 469–478. LÓPEZ, M.T., A. FERNÁNDEZ-CABALLERO, M.A. FER- NÁNDEZ, J. MIRA and A.E. DELGADO (2006b) Visual surveillance by dynamic visual attention method, Pattern Recognition, 39, 2194–2211. LÓPEZ, M.T., A. FERNÁNDEZ-CABALLERO, M.A. FER- NÁNDEZ, J. MIRA and A.E. DELGADO (2007) Dynamic visual attention model in image sequences, Image and Vision Computing, 25, 597–613. LÓPEZ, M.T., A. FERNÁNDEZ-CABALLERO, J. MIRA, A.E. DELGADO and M.A. FERNÁNDEZ (2006c) Algorithmic lateral inhibition method in dynamic and selective c� 2011 Blackwell Publishing Ltd Expert Systems, September 2011, Vol. 28, No. 4 321 visual attention task: application to moving objects detection and labeling, Expert Systems with Applica- tions, 31, 570–594. LÓPEZ-VALLES, J.M., M.A. FERNÁNDEZ and A. FERNÁN- DEZ-CABALLERO (2007) Stereovision depth analysis by two-dimensional motion charge memories, Pat- tern Recognition Letters, 28, 20–30. LURIA, A.R. (1973) The Working Brain, New York: Basic Books. MARR, D. (1982) Vision, New York: Freeman. MARTÍNEZ, R., M. RINCÓN, M. BACHILLER and J. MIRA (2008) On the correspondence between objects and events for the diagnosis of situations in visual sur- veillance tasks, Pattern Recognition Letters, 29, 1117–1135. MATURANA, H.R. (1999) The organization of the living: a theory of the living organization, International Journal of Human-Computer Studies, 51, 149–168. MCCULLOCH, W.S. and W. PITTS (1939) A logical calculus of the ideas immanent in nervous activity, Bulletin of Mathematical Biology, 52, 99–115. MEYER, B. (1997) Object-Oriented Software Construc- tion, Englewood Cliffs, NJ: Prentice Hall. MINSKY, M.L. (1961) Steps towards artificial intelli- gence, Proceedings of the Institute of Radio Engi- neers, 49, 8–30. MIRA, J. (2006) On some of the neural mechanisms underlying adaptive behavior, IDEAL 2006, Lecture Notes Computer Science, 4224, 1–15. MIRA, J. and A. DELGADO (2003) Neural modeling in cerebral dynamics, BioSystems, 71, 133–144. MIRA, J. and A.E. DELGADO (1987) A logical model of co-operative processes in cerebral dynamics, Cyber- netics and Systems, 18, 319–349. MIRA, J., A.E. DELGADO, A. FERNÁNDEZ-CABALLERO and M.A. FERNÁNDEZ (2004) Knowledge modelling for the motion detection task: the algorithmic lateral inhibition method, Expert Systems with Applica- tions, 27, 169–185. MIRA, J., A.E. DELGADO, J.M. GASCUEÑA, A. FERNÁN- DEZ-CABALLERO and M.T. L’OPEZ (2009) Computa- tional agents to model knowledge – theory, and practice in visual surveillance, Lecture Notes in Computer Science, 5601, 375–385. MORENO-GARCIA, J., L. RODRIGUEZ-BENITEZ, A. FER- NÁNDEZ-CABALLERO and M.T. LÓPEZ (2010) Video sequence motion tracking by fuzzification techni- ques, Applied Soft Computing, 10, 318–331. NEWELL, A. (1980) The knowledge level, AI Magazine, 2, 1–20. NEWELL, A. and H.A. SIMON (1972) Human Problem Solving, Englewood Cliffs, NJ: Prentice Hall. NUNES, L. (2006) Fundamentals of Natural Computing: Basic Concepts, Algorithms, and Applications, Lon- don: Chapman & Hall=CRC. ORGUN, B., M. DRAS, A. NAYAK and G. JAMES (2008) Approaches for semantic interoperability between domain ontologies, Expert Systems: The Journal of Knowledge Engineering, 25, 179–196. PADGHAM, L., J. THANGARAJAH and M. WINIKOFF (2008). Prometheus design tool, in Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence, 13–17 July, 2008, Chicago, Illinois, USA, pp. 1882–1883. PADGHAM, L. and M. WINIKOFF (2004) Developing Intelligent Agents Systems: A Practical Guide, New York: John Wiley and Sons. PADGHAM, L. and F. ZAMBONELLI (2006) Agent-Oriented Software Engineering VII, Berlin: Springer-Verlag. PATRICIO, M.A., F. CASTANEDO, A. BERLANGA, O. PÉREZ, J. GARCÍA and J.M. MOLINA (2008) Computational intelligence in visual sensor networks: improving video processing systems, Studies in Computational Intelligence, 96, 351–377. PAVÓN, J., J. GÓMEZ-SANZ, A. FERNÁNDEZ-CABALLERO and J.J. VALENCIA-JIMÉNEZ (2007) Development of intelli- gent multi-sensor surveillance systems with agents, Robotics and Autonomous Systems, 55, 892–903. PAVÓN, J., J. GÓMEZ-SANZ and R. FUENTES (2005) The INGENIAS methodology and tools, chapter 9, in Agent-Oriented Methodologies, B. Henderson-Sellers and P. Giorgini (eds), Hershey, PA: Idea Group Publishing, 236–276. PECHOUCEK, M., S.G. THOMPSON and H. VOOS (2008) Defense Industry Applications of Autonomous Agents and Multiagent Systems, Basel, Switzerland: Whitestein. POKAHR, A., L. BRAUBACH and W. LAMERSDORF (2005) Jadex: a BDI reasoning engine, chapter 6, in Multi- Agent Programming: Languages, Platforms and Ap- plications, R.H. Bordini, M. Dastani and A. El Fallah Seghrouchni (eds), Berlin: Springer, 149–174. QURESHI, F.Z., D. TERZOPOULOS and R. GILLETTE (2004). The cognitive controller: a hybrid, deliber- ative=reactive control architecture for autonomous robots, in Proceedings of the 17th International Con- ference on Industrial & Engineering Applications of Artificial Intelligence & Expert Systems, Ottawa, Canada, pp. 1102–1111. ROUGEMAILLE, S. (2008). Ingénierie des systèmes multi- agents adaptatifs dirigée par les modèles. Ph.D. thesis, Institut de Recherche en Informatique de Toulouse. RUSSELL, S. and P. NORVIG (2002) Artificial Intelligence: A Modern Approach, Englewood Cliffs, NJ: Prentice Hall. STURM, A. and O. SHEHORY (2004) A comparative evaluation of agent-oriented methodologies, in Methodologies and software engineering for agent systems. The Agent-Oriented Software Engineering Handbook, F. Bergenti, M.-P. Gleizes and F. Zam- bonelli (eds), Dordrecht, the Netherlands: Kluwer Academic Publishers, pp. 127–149. SYCARA, K., K. DECKER, A. PANNU, M. WILLIAMSON and D. ZENG (1996) Distributed intelligent agents, 322 Expert Systems, September 2011, Vol. 28, No. 4 c� 2011 Blackwell Publishing Ltd IEEE Expert, Intelligent Systems and Their Applica- tions, 11, 36–46. UNLAND, R., M. KLUSCH and M. CALISTI (2005) Soft- ware Agent-Based Applications, Platforms and Devel- opment Kits, Basel, Switzerland: Birkhäuser Verlag. VALENCIA-JIMÉNEZ, J.J. and A. FERNÁNDEZ-CABALLERO (2006). Holonic multi-agent systems to integrate independent multi-sensor platforms in complex sur- veillance, in Proceedings of the IEEE International Conference on Advanced Video and Signal based Surveillance, Sydney, Australia, 49. VARELA, F.J. (1979) Principles of Biological Autonomy, Cambridge, MA: North Holland, New York: The MIT Press. WEISS, G. (1999) Multiagent Systems: A Modern Ap- proach to Distributed Artificial Intelligence, Cam- bridge, MA: The MIT Press. WINIKOFF, M. (2005) JACK intelligent agents: an indus- trial strength platform, chapter 7, in Multi-Agent programming: Languages, Platforms and Applica- tions, R.H. Bordini, M. Dastani and A. El Fallah Seghrouchni (eds), Berlin: Springer, pp. 175–193. WOOLDRIDGE, M.J. and N.R. JENNINGS (1995) Intelli- gent agents: theory and practice, The Knowledge Engineering Review, 10, 115–152. The authors José M. Gascueña José M. Gascueña received his MSc in Compu- ter Science from the University of Castilla-La Mancha at the Superior Polytechnic School of Albacete, Spain, in 2004. In 2006, he received a scholarship from the Spanish Junta de Comuni- dades de Castilla-La Mancha. In 2010 he got his Ph.D. from University of Castilla-La Mancha in applying multi-agent systems technology in the computer vision area. His research interests are in Software Agents and Multi-agent Systems, as well as in Computer Vision. Antonio Fernández-Caballero Antonio Fernández-Caballero received his de- gree in Computer Science from the Technical University of Madrid, Spain, in 1993, and re- ceived his Ph.D. from the Department of Artifi- cial Intelligence of the National University for Distance Education, Spain, in 2001. He is a Full Professor with the Department of Computer Science at the University of Castilla-La Man- cha, Spain. He is the director of the research group n&aIS (natural and artificial Interaction Systems) belonging to LoUISE (Laboratory of User Interaction and Software Engineering) of the Albacete Research Institute of Informatics, Spain. His research interests are in Image Pro- cessing, Cognitive Vision, Neural Networks, and Intelligent Agents. A. Fernández-Caballero is an Associate Editor of the Pattern Recogni- tion Letters journal, a member of the Editorial Board of the Journal of Physical Agents, and a member of the IAPR. He has authored around 200 peer reviewed papers. Marı́a T. López Marı́a T. López received her degree in Physics from the University of Valencia, Spain, in 1991, and received her PhD from the Department of Artificial Intelligence of the National University for Distance Education, Spain, in 2004. Since 1991, she is an Associate Professor with the Department of Computing Systems at the Uni- versity of Castilla-La Mancha, Spain. Her re- search interests are in Image Processing and Computer Vision. Marı́a T. López is member of the IAPR. Ana E. Delgado Ana E. Delgado is a Full Professor of Computer Science and Artificial Intelligence with the De- partment of Artificial Intelligence at the Na- tional University for Distance Education (UNED) in Madrid, Spain. Her current research interests are in Neural Modeling, Bio-inspired Cooperative Agents and Computer Vision. c� 2011 Blackwell Publishing Ltd Expert Systems, September 2011, Vol. 28, No. 4 323 
work_zig56z22kbf3dcvmxtbbhyznsa	----	Microsoft Word - INR05-08.docx   InterNeg Research Papers  INR05/08             Revised version appeared in:  Group Decision and Negotiation Conference, Karlsruhe, Germany  June 25‐29, 2006.    h t t p : / / i n t e r n e g . o r g / Acknowledgments: The publication of the InterNeg Research Papers has been supported by the Natural Sciences and  Engineering  Research  Council,  the  Social  Sciences  and  Humanities  Research  Council,  and  the  J.  Molson  School  of  Business, Concordia University.   Copyright:  The  papers’  copyright  stays  with  their  authors.  The  paper  cannot  be  reprinted  in  any  form  without  its  authors’ explicit consent.   Predicting  Opponent’s  Moves  in  Electronic  Negotiations Using Neural Networks        Réal Carbonneau 1, Gregory E. Kersten 2  &  Rustam Vahidov2    1Department of Management Sciences, HEC Montréal   2John Molson School of Business, Concordia University      Abstract  Electronic negotiation experiments provide a rich source of  information about relationships  between  the  negotiators,  their  individual  actions,  and  the  negotiation  dynamics.  This  information can be effectively utilized by intelligent agents equipped with adaptive capabilities  to  learn  from  past  negotiations  and  assist  in  selecting  appropriate  negotiation  tactics.  This  paper presents an approach to modeling the negotiation process in a time‐series fashion using  artificial neural network. In essence, the network uses information about past offers and the  current  proposed  offer  to  simulate  expected  counter‐offers.  On  the  basis  of  the  model’s  prediction, “what‐if” analysis of counter‐offers can be done with the purpose of optimizing the  current offer. The neural network has been trained using the Levenberg‐Marquardt algorithm  with Bayesian Regularization. The simulation of the predictive model on a testing set has very  good  and  highly  significant  performance.  The  findings  suggest  that  machine  learning  techniques  may  find  useful  applications  in  the  context  of  electronic  negotiations.  These  techniques  can  be  effectively  incorporated  in  an  intelligent  agent  that  can  sense  the  environment  and  assist  negotiators  by  providing  predictive  information,  and  possibly  automating some negotiation steps.  Keywords:  Electronic  Negotiations;  Opponent  Modeling;  Counter‐Offer  Prediction;  Offer  Optimization; Neural Networks    1. Introduction  INR 05/08  2  Negotiations play a crucial role in conducting everyday business activities, such as negotiating  contracts with customers, negotiating service level agreements with suppliers or negotiating  agreements with unions. Many of these negotiations can have outcomes that impact long‐term  business relationships, profitability, and reputation of businesses. Electronic negotiations  in  particular, have gained heightened importance due to the advance of the web and e‐commerce  (Kersten &  Noronha,  1999). While this brings  in the challenges associated with conducting  negotiations  in the global environment, where parties could have  little or no knowledge of  each  other,  it  also  presents  some  valuable  opportunities  to  employ  advanced  technologies,  such as intelligent agents, in the negotiation process.  Intelligent systems for negotiation support that aim at enhancing the negotiator’s abilities to  understand the counterparts, their needs and limitations and to predict their moves could be  very valuable tools to be used in negotiation tasks (Zeng & Sycara, 1998). The purpose of this  paper  is  to  investigate  the  feasibility  of  machine  learning  approaches  in  modeling  the  opponent’s future offers. The data is obtained from electronic negotiation experiments which  provide  a  rich  source  of  information  about  the  relationships  between  negotiators,  their  individual actions, and the negotiation dynamics. This information can be effectively utilized  by intelligent agents equipped with adaptive capabilities to learn from past negotiations and  assist in selecting appropriate negotiation tactics.    In  order  to  test  the  applicability  of  machine  learning  approaches  in  modeling  negotiation  dynamics we use data obtained from bilateral negotiations experiments conducted with the  use  of  the  Inspire  electronic  negotiation  system  (Kersten  &  Lo,  2003;  Kersten  &  Noronha,  1999).    In  our  approach  the  negotiation  process  has  been modeled  in a  time‐series  fashion  using an artificial neural network. In essence, the model uses information about past offers and  counteroffers, including the most recent offer made by the negotiator, to predict the expected  counteroffer. On the basis of the model’s prediction, “what‐if” analysis of counter‐offers can be  done with the purpose of optimizing the current offer. The assessment of offers and counter‐ offers is performed based on the user’s utility function.    The purpose of this study is to assess the applicability of machine learning to provide advice to  the negotiator. This advice involves simulation of the possible responses to the offer negotiator  is  contemplating.  The  subsequent  sections  present  the  related  work;  introduce  a  neural  network‐based approach to model the negotiator’s counterpart; propose negotiation support  which involves offer optimization; discuss the model implementation and results; and provide  conclusions and directions for future work.  2. Background  Negotiation is one of key activities of businesses and it has crucial impact on an organization’s  performance. Researchers in negotiation support and automation seek to facilitate various  negotiation‐related tasks using the capabilities of technology. Substantial efforts have been  expended in attempts to fully automate negotiation processes (Beam & Segev, 1996; Chavez et  al., 1997a; Jennings & Faratin, 2001; Maes et al., 1998). While the extensive coverage of  automated negotiations is beyond the scope of the current work, it seems useful to mention  several key studies.   Another work employed genetic algorithms to generate rules which relate the negotiators’  current offers with the likely subsequent offers (Matwin et al., 1991). The algorithm evolves  INR 05/08  3  multiple classifiers with the higher fitness being assigned to those rules, which more frequently  contribute towards “compromise trajectories”. The method’s limitations include, not taking  into account relative importance of various issues and values; outguessing aspiration levels of  the counterpart; compromise‐orientation rather than orientation on the achievement of values  of the supported negotiator ; and the substantial increase of the search space size when  additional issues and continuous issues are introduced.  Chavez and Maes  (1996) designed and tested an agent‐based marketplace Kasbah in which  various agents could be created by the users. These agents engage in bilateral negotiations on  behalf of their principals (buyers and sellers). They followed one of the three negotiation  strategies defined by price‐concession curve over time (Chavez et al., 1997b).    More recently, Kwon, Shin, & Kim (Kwon et al., 2006) used a semantic web‐based agent  community to introduce the concept of pervasive negotiation support. The system, called  “SmartGuide” includes user, supplier, and negotiation agents to provide flexible environment  for context‐aware automated negotiations.   Dzeng and Lin (2004) propose an agent‐based system for procurement negotiations in  construction domain. The system employs the coordinator agent that tries to find adequate  agreement for the parties while employing genetic algorithms and using the preferences of the  parties. While revealing negotiator preferences to a coordinating agent simplifies the task, this  may not be appropriate for most business negotiations, where the participants would typically  like to keep their preferences private. The experimental results show that the joint payoff of  the contractor and supplier improved from 1.5% to 9.8% as compared with conventional  human negotiation (Dzeng & Lin, 2005).  An approach that combines recommendation generation and negotiation agents for product  selection and purchase is described in (Lee, 2004). The negotiating agents in this work follow a  chosen concession strategy while taking into account past offers by the opponent. More  specifically, the agents base their concessions according to the “positive” vs. “negative”  classification of an opponent’s attitude, which is determined using a threshold function.   While the above and numerous other works in automated negotiations may hold a promise in  streamlining routine well‐structured negotiation tasks, we believe that in most business  negotiation contexts humans need to be in control of the process with the agents playing a role  of assistants. There has been some research effort in this area to provide solutions for assisting  human negotiators. An overview of electronic negotiation and negotiation support systems  and negotiation software agents (NSA) is presented in (Kersten & Lo, 2003). The work also  discusses the Aspire system which is a combination of the Inspire NSS and the Atin NSA for  assisting negotiators. In Aspire the agent uses an inference engine to provide  recommendations based on inputs and previously encoded rules.   Another application of agent assistant in commerce negotiations has been implemented in  eAgora marketplace (Chen et al., 2004; Chen et al., 2005). In eAgora an agent watches over the  shoulder of the negotiator and critiques trial offers. The agent also advises an action upon  receiving the counterpart’s offer and generates a package of adequate candidate offers for the  consideration by the user. An experimental evaluation of eAgora’s has revealed that agent‐ assisted negotiators have been able to achieve better performance and reported higher  usefulness and satisfaction levels, especially when completing complex negotiation cases  (Vahidov et al., 2005). These findings encourage further work on using innovative agent  INR 05/08  4  solutions in e‐negotiations.  In the Aspire and eAgora systems the agents rely on a knowledge base created by the  developers. While providing users with help, based on common negotiation related heuristics  and rules of thumb is an improvement over unassisted mode, the accuracy of the knowledge  and its applicability to concrete cases may be questionable. Machine learning approaches do  not require explicit specification of the knowledge; instead they are capable of learning  patterns from the existing data.   This paper sets out to investigate the applicability of neural networks to learning the dynamics  of negotiations from the past negotiation cases. Being “universal approximators”, neural  networks are capable of learning and generalization, thus offering an inductive alternative to  the expert system‐like support. Oprea (2002) proposes a similar approach in which a feed‐ forward neural nets are used for predicting the counterpart’s next offer. The approach was  limited to a single issue (price) negotiation and only the offers made by an opponent are  considered in the model.  There are many sources of information that may be used by a negotiator, which could be  categorized as offer and non‐offer information.  Offer information is self contained in an offer  and their sequence, which is always available in any negotiation situation.  Non‐offer  information, such as information from the environment, persuasion tactics and discussions,  may be present in negotiation situations, but are not a defining feature.  This research focuses  on information contained in the offers because this is the information that is fundamental to  communication required in any negotiation. In this research we construct a neural network‐ based model which learns patterns from the past offer exchanges and presents  recommendations for future moves.  3. Neural network‐based predictive model  Modeling of an opponent in the negotiation process may significantly improve performance of  the negotiators. Some of the works mentioned above attempt to incorporate the opponent’s  moves in the process of offer generation. For example, in (Lee, 2004) past concessions made by  the counterpart are used to construct the model of this counterpart. If, on the average, they  exceed a pre‐defined threshold level, the opponent is modeled as having a “positive” attitude.  Zeng and Sycara (1998) use game‐theoretic approach with Bayesian belief revision to model a  negotiation counterpart. Mudgal and Vassileva (2000) represent the counterpart’s decision‐ making with probabilistic influence diagrams. In (Faratin et al., 2002) an approach to  generating offers that are close to the opponent’s offer based on a fuzzy similarity measure is  employed. In eAgora system an assistant agent generates recommendations based on the  chosen strategy, as well as the concessions made by an opponent (Chen et al., 2005). While  such approaches provide a certain advantage, they are in some sense “rough” in providing ways  to predict next moves by the opponent.  Anticipation of the counterpart’s next move can be of critical importance to negotiators as it  could help them to better compose their own offer. Therefore, negotiators make efforts to  uncover patterns in their counterpart’s behavior, in particular the concession making patterns.  The feasibility of a predictive model to perform this task can be assessed utilizing advanced  machine learning techniques that are effective in pattern detection. In this work we have relied  on artificial neural networks (ANN), which have been proven to be universal approximators,  INR 05/08  5  provided with sufficient hidden layer neurons and assuming that the activation function is  bounded and non‐constant (Hornik, 1991).    The predictive model will need to indicate the expected offer from the counterpart. Thus, at  the output layer of a neural network each neuron will be associated with one component of the  offer, i.e. one negotiation issue. The number of output neurons will be essentially the same as  the number of the issues in negotiations. The inputs will be associated with the past offers and  counter‐offers and the current trial offer. Therefore, the same number of inputs as that of  outputs will be allocated to the current considered offer plus the history of past offers and  counter‐offers. In general, the more of the history of a given negotiation is fed as inputs the  more precise one would expect the predictions to be. However, this effectively puts a  restriction on when the model can be actually used in the negotiation process. For example,  when two latest offers and counteroffers are included, the negotiator may start using the  model only in the third round of negotiations.  Our predictive network needs to be trained using past negotiation data to adequately capture  the dynamics of negotiations. Although the classical error back‐propagation algorithm has  been the most popular learning techniques for neural networks, we use a faster training  algorithm as well as a framework that will improve the generalization capability of the model.   In particular, we use the Levenberg‐Marquardt algorithm as applied to neural networks to  adjust the weights (Hagan et al., 1996; Hagan & Menhaj, 1994).  Our choice is due to the fact  that this algorithm is one of the fastest training algorithms available with training being 10‐100  times faster than simple gradient descent back‐propagation of error.    The Levenberg‐Marquardt neural network training algorithm is combined into a framework  that permits estimation of the network’s generalization by the use of a regularization  parameter.  ANN performance measures typically include the error of the outputs of the  network, such as the means squared error (MSE).  However, with the large number of hidden  units (and thus, weights) neural networks may be too powerful, resulting in overtraining.  While improving on MSE, such ANN may end up with inferior generalization capability. Thus, a  regularization performance function which includes the sum of the weights and biases can be  used instead, combined with a regularization parameter, which determines how much weight  is given to the sum of weights and bias in the formula: msereg = γ mse + (1 ‐  γ) msw.  MSE  represents the mean of the sum of squares of errors, MSW represents the mean of the sum of  squares of the network weights and biases and γ represents the tradeoff ratio.  This  regularization parameter permits the control of the ratio of impact between reducing the error  of the network and the number of weights or power of the network.    The tuning of regularization parameter is automated within the Bayesian framework (MacKay,  1992)  and, as combined with the Levenberg‐Marquardt training algorithm, results in high  performance training combined with a preservation of generalization by avoiding overfitting of  the training data (Foresee & Hagan, 1997).  This algorithm helps control overfitting of the  target function and it also provides an estimate of how many weights and biases are being  effectively used by the network.  Larger networks should result in approximately the same  performance, since regularization results in a trade off between error and network parameters,  which is relatively independent of network size.  All Neural Network modeling and training in  this work is performed in MATLAB 7.0 and MATLAB’s Neural Network Toolbox (Mathworks  2005).    INR 05/08  6  4. Negotiation support with predictive neural model  The predictive neural network‐based model introduced in the previous section can be  integrated as part of the negotiation support system. One concern with such integration  relates to the informational demand required by the model. Ideally, the model should not  require extensive information to facilitate the “plug‐in” type of integration. Thus, when  building the model, we kept a strong focus on using only information that would normally be  available to the negotiator.    The most common type of information available to a negotiator any negotiation session  includes details of an offer and the time it was created. The system monitors the creation of  this information and keeps a history, which will be used for predicting the next negotiation  steps.  Additionally, this permits the negotiation modeling researcher to build models with  historical negotiation information from the existing systems; no details on the individual  negotiators are necessary.  An attractive feature of this model is that no information on the  negotiators themselves is required. This allows the intelligent negotiation agent to be built  separately from any negotiation system.  The intelligent negotiation assistant discussed here is capable of simulating the opposing  negotiator’s next offer starting with the third offer.  This permits a negotiator to do what‐if  analysis in order to test various estimated counter offers that will result from a specific offer in  the current situation without actually submitting an offer to the counter‐part.  Apart from  using the model as means of learning more about the opponent, this capability would also  allow to perform an optimization of all potential offers with the purpose of finding the best  offer for the current situation.    The choice of the method of search for the optimal offer depends on the number of issues and  options available. For a small number of issues and options a comprehensive search can be  performed. For example, if there are two issues and five options, the number of all possible  offers is limited to 25. For such small spaces the use of comprehensive search could be  justified. However, as the number of possible offers increases, there may be a need for the use  of a heuristic optimization method. Such methods could include, for example, hill‐climbing,  simulated annealing, or genetic algorithms. The agent could choose the appropriate algorithm  based on the size of the negotiation space.  5. Model implementation  In order to demonstrate the feasibility and effectiveness of an ANN‐based predictive model we  have used past data collected by Inspire system (Kersten & Noronha, 1999). The Inspire  negotiation system is web‐based, which permits two parties located anywhere in the world,  who have internet access, to negotiate on a chosen case. Negotiation issues and issue options  are specified in advance for a specific case and each negotiator specifies a rating for each issue  and, additionally, specifies a rating for each option of an issue.  The sum of all issues must be  100 and the sum of all the options of a specific issue must also be 100. Each selected option  rating is multiplied by the issue rating which permits the calculation of the total utility of a  package. Package utility ratings are presented to the user for further adjustments since the  user may feel that certain package utility ratings should be different than the result of the  calculations to correctly reflect his or her preferences.  Issue, option and package ratings are  specific and confidential to each user, allowing evaluation of all offers submitted or received by  INR 05/08  7  a particular user.  A negotiator has no information about the preference structure of the  counter‐part and at no point are these preferences ever revealed to the counterpart.  As an  additional information source, users may examine a graph of the utility of the history of offers  and counter offers.  The negotiation case under study is a simulated scenario where a seller and a buyer want to  enter into a business relationship.  Specifically, Itex manufacturing is entering the negotiations  as a producer of bicycle gears, wishing to sell to Cypress Cycles; a bicycle producer.   The case  defines the market as competitive, meaning that either party may terminate the negotiations if  they do not find the negotiations promising, since they can find other business partners.  The  issues include Price (3.47$, 3.71$, 3.98$, 4.12$, 4.37$), Delivery (20 Days, 30 Days, 45 Days, 60  Days), Payment (60 Days After Delivery, 30 Days After Delivery, Upon Delivery) and Returns  (Full Price, 75% Refund with 5% Spoilage, 75% Refund with 10% Spoilage).  Users may send  offers and messages to their counterparts through the Inspire system as they work towards a  compromise that is acceptable for both parties. At any point users can unilaterally terminate  the negotiations.   The Inspire’s dataset provides rich information on the negotiation that took place through the  use of the system. Of the 6310 offers considered, 2426 (38%) were proposed by females and  3379 (54%) were proposed by males and 505 (8%) offers were from negotiators who did not  specify their gender.  Negotiators from over 100 different countries are in the sample.  The median negotiation session has 7 offers and the average negotiation session has 6.70  offers. This means that there is enough information to model counter‐offers based on  information about past offers.  At a minimum, we will require the current offer and the next  offer, so there is a minimum of two offers in a negotiation session for it to be considered for  this type of modeling.  To fulfill this minimum requirement, we must also be able to correctly  identify the relationship between the offers.  We need to know to which offer a counter offer  was made, and conversely, to which counter offer an offer belongs.  This may seem trivial if  offers are forced to be sequential and alternating between the sides, however, in our current  data set, one party of the negotiations may make several offers before receiving a counter offer.   To reconstruct the correct order relationship from the existing data, we assume that a counter  offer is made in relationship to the last offer from the negotiation counterpart.    Adding information about the last buyer’s offer and the last seller’s offer to the model, which  already contains the current offer and the next offer, provides a model that contains  information about past offers.  However, it takes a minimum of 4 offers in a negotiation  session to learn the patterns and a minimum of 2 past offers plus the current offer, for a total  of 3 offers, to estimate the subsequent counter offer.   This will result in loosing about 27% of  the observations because of this time requirement.   The last offers by the buyer and seller are relevant for estimating the subsequent counter offer.   The first offer is of interest to us because it sets opening tone of the negotiation as well as the  minimum and maximum offers so far, which provides a measure of the absolute range of the  current negotiations. To provide information on how much disagreement there is up to date in  the current negotiation, we add the standard deviation of the offer values.  Finally, the average  of the offers will provide information about the mid‐point towards which the negotiation is  moving.  The major inputs and outputs of the model are summarized in Table 1. One extra  input is reserved to indicate whether a negotiator is a buyer or a seller.  INR 05/08  8  Table 1 Summary of the model inputs and outputs  Offers Price Delivery Payment Returns Timestamp Inputs Last “Buy” V V V V V Last “Sell” V V V V V Current V V V V V First V V V V V Minimum V V V V Maximum V V V V St. deviation V V V V V Average V V V V V Outputs Predicted V V V V   Thus, the resulting model has 39 inputs and 4 outputs and the final dataset for the above  required observations results in 6310 observations. In order to test the generalizability of the  model we created two separate sets, 5048 (80%) of the observations have been assigned to the  training set and 1262 (20%) of the observations have been kept in the testing set.  Our neural network has 39 inputs, 4 outputs and 10 hidden neurons (Figure 1). The transfer  function we used in the hidden layer is the tan‐sigmoid function which does non‐linear scaling  from an infinite range to values between ‐1 and 1 and the output layer transfer function is a  linear function.  The input for each neuron is the not only the input signals coming from either  the input variables or the results of a previous neural network layer, but it is this signal  multiplied by the weight of the connection.  There is a total of 444 connection weights, which  results in a ratio of observations to weights equal to 11.37 (5048/444), which is an appropriate  level to permit powerful modeling and good generalization.     INR 05/08  9    Figure 1. Predictive negotiation neural network model    6.  Results  Training of the neural network was executed and the networks converged after 252 iterations  through all of the dataset.  Convergence is characterized by the stabilization of the sum of  square errors (SSE) and the sum of square weights (SSW).  For our network of 39 inputs, 10  hidden layer neurons and 4 outputs, we reach a stable SSE of 3642.17, a SSW of 39.5911 and an  effective number of network parameters (weight and biases) of 382.143 of the total 444  parameters (Figure 2).   The correlation of the model’s results with the actual output on the testing set are 0.74647 for  price, 0.68512 for delivery 0.67421 for payment and 0.69253 for returns which is an average of a  0.6996 correlation of predicted output with actual output.  From this we can see that price is  the negotiation issue that has the most predictable output based on past negotiation patterns,  while delivery has the least predictable pattern.    INR 05/08  10    Figure 2 ‐ Neural Network training history graph  To further understand the performance of the model we consider the results in terms of the  average absolute error and the error obtained after the outputs are rounded to an integer value  (rounding is necessary because the negotiation case requires selection of ordinal values).  The  average absolute error across all outputs of the testing set for one prediction is 1.68 on the 15  possible ordinal levels, which is approximately 11%.  When the outputs are rounded to integers,  the rounded error is 1.39 for the total of 15 possible ordinal levels which represents  approximately a 9% error. The detailed results are presented in Table 2.  Table 2. Summary of results  Measure Price Payment Delivery Returns Avg/Tot Testing Set R 0.74647 0.68512 0.67421 0.69253 0.6996 R2 0.55722 0.46939 0.45456 0.47960 0.4902 Exact error 0.4753 0.4454 0.3609 0.3972 1.6788 Rounded error 0.4113 0.3796 0.2837 0.3170 1.3914 Training Set R 0.76405 0.71638 0.70387 0.7336 0.7295 R2 0.58377 0.51320 0.49543 0.53817 0.5326   To assess the model, several “what‐if” scenarios were conducted, a few of them that may be  interesting to the seller are discussed here These scenarios provide the seller with information  about the expected results of the offers thus helping make a better decision.    INR 05/08  11  The seller may want to see what would the expected counter offer of multiple different offer  possibilities be (e.g. resending the same offer as the last time, giving a price concession and  giving a concession for issue Returns). Because of the complexity of the option and package  ratings in Inspire, we choose a simple package utility rating scheme, where price is 45%,  delivery is 20%, payment is 10% and returns is 25% of the overall utility.  Every issue option  rating is related to the levels, for example, price has 5 levels, so level 1 would have utility of 0  and level 5 would have utility of 100.  From this we can calculate the utilities for any offer  which will permit us to evaluate counter offers.  As a base for comparison, we know the actual  offer made, and we have included this in the comparison of the what‐if scenarios in Table 3.  Table 3 Comparison of example “what‐if” scenarios  Current Offer Simulated counteroffer Name Pr. Del. Pay. Ret. Pr. Del. Pay. Ret. Estimated utility Ordinal utility Actual 4 3 3 3 2 2 2 1 20.611 18.75 Same as last 5 2 2 3 3 1 1 1 22.138 22.5 Price concession 4 2 2 3 2 1 2 1 19.577 13.75 Returns concession 5 2 2 2 3 1 2 1 22.404 25   Based on these simple “what‐if” simulations, we observe that even small variations in the  current offer can have important impact on the expected counter offer from the opponent.   Using the utility function, we can evaluate different scenarios to better understand the  negotiation dynamics and select one that has a high‐utility expected counter‐offer.    The case used for this research has a total search space of 180 options. Since the search space is  small a comprehensive search for an optimal solution is feasible. As mentioned earlier, for  larger spaces one of the heuristic optimization methods could be used. Table 4 shows the  results of optimization based on the expected counter‐offers.    Table 4 Offer optimization results  Current offer Expected counteroffer Ordinal utility Direct utility Pr. Del. Pay. Ret. Utility Pr. Del. Pay. Ret. 5 4 3 1 65.00 2.5543 1.9644 1.5168 0.91199 30 23.04981 5 3 2 2 63.75 2.5332 1.6841 1.503 1.0449 30 22.20713 5 4 2 1 62.50 2.5151 1.9311 1.5415 0.8547 30 22.146 5 3 3 1 60.00 2.6103 1.7381 1.6238 0.92875 30 22.92056 5 3 2 1 57.50 2.5929 1.7339 1.6303 0.88159 30 22.42531 5 2 3 1 55.00 2.6704 1.5069 1.7312 0.94722 30 22.82463 5 2 2 1 52.50 2.6747 1.5318 1.7095 0.90945 30 22.70719   The maximum expected ordinal level utility of the counter offer was identified as 30.  To  achieve this high counter offer, the intelligent agent recommends that the seller changes his  INR 05/08  12  offer back to the highest price but at the same time suggests that the seller offer a large returns  concession, which is, in most suggested offers, the most generous returns policy level even  though it is the issue that has the second highest impact on utility.  The other two negotiation  issues have more flexibility in offer level ranges before having an impact on the next counter  offer.    It is important to note that these recommendations are specific to the current negotiations  patterns identified in this specific scenario, they are not general recommendations; the  generalized knowledge is embedded in the weights of the neural network.  For example, the  recommendation to move to a higher price but give concessions on returns is not a general  rule, but a result of the patterns identified in the current scenario.  Even though the above  offers result in the same ordinal counter offer utility level, our method chooses an overall  recommendation of the package that maximizes the utility of the current offer. Thus the model  recommends that the seller offer a price of 4.37$, 60 days delivery, payment upon delivery and  full price returns.    The negotiation‐assistant agent utilizing the proposed approach would be capable of providing  the seller with information about the dynamics of the current negotiation situation, the best  potential offer and the ability to run “what‐if” analysis for any considered offer. Thus the seller  may directly propose the recommended offer or do further analysis and refine the offer to meet  requirements that may have not been thoroughly captured by the utility function or other  external factors.  7. Summary and Conclusions  In this paper we have presented a neural networked‐based model for predicting the opponent’s  offers during negotiation process. The model can be embedded in a system for assisting  negotiator in making offers in e‐negotiations. The simulation of the predictive model on a  testing set has very good and highly significant performance, especially considering the noisy  data domain.  An examination of “what‐if” scenarios and optimization results on a real case  shows that the model can exhibit interesting negotiation strategies and, at the very least,  provides useful information to a negotiator.    One potential limitation of the current research relates to the generalizability of the model.  Specifically, our model was used for a particular negotiation case, and the accuracy of its  predictions and recommendations may be less adequate for other types of cases. While this is  an important issue to tackle in future research, it is important to note that many e‐commerce  related negotiation contexts share similar aspects. For example, issues like price, delivery, and  returns are common for many real‐life commerce negotiations. Thus, to ensure the  generalizability of the model, the future research should look into different classes of  negotiation cases and recommend a variety of neural‐based models as part of the negotiation  support toolbox, where the appropriate model could be chosen for particular contexts.  Another important direction for future research includes investigation of the effectiveness of  the  model  in  experimental  settings.  There  are  two  potential  experiments  that  would  be  interesting  in  this  respect:  one  featuring  fully  automated  negotiations  and  the  other  one  involving assisted negotiations.  In the first experiment, the human subjects would be assigned  to  one  side  of  the  negotiation  case  and  the  opponent  would  be  the  negotiation  agent  incorporating models described  in this research.   We could then compare the utility of the  INR 05/08  13  negotiations outcomes of the two groups, human and agent, to see if there is a difference.  In  the second experiment, negotiation system users would be randomly assigned to either side of  a negotiation case and one of the two negotiators would be assigned to be assisted by the  intelligent negotiation agent.  As in the previously presented experiment, the two groups could  be compared to identify if there is an advance to using the predictive intelligent negotiation  assistant agent.  References  Beam, C., & Segev, A. (1996). Automated Negotiation: A Survey and the State of Art. Berkeley, CA: Hass  School of Business.  Chavez, A., Dreilinger, D., Guttman, R., & P., M. (1997a). A Real‐life Experience in Creating an Agent  Market. Paper presented at the Application of Intelligent Agents and Multi‐Agent Technology  PAAM'97, London.  Chavez, A., Dreilinger, D., Guttman, R., & Maes, P. (1997b). A Real‐life Experiment in Creating an Agent  Marketplace. In H. S. Nwana & N. Azarmi (Eds.), Software Agents and Soft Computing (pp. 160‐ 179): Springer‐Verlag.  Chavez, A., & Maes, P. (1996). Kasbah: An Agent Marketplace for Buying and Selling Goods. Paper  presented at the First International Conference on the Practical Application of Intelligent  Agents and Multi‐Agent Technology, London.  Chen, E., Kersten, G. E., & Vahidov, R. (2004, Jan. 14‐16, 2004). An E‐marketplace for Agent‐supported  Commerce Negotiations. Paper presented at the 25th McMaster World Congress Management  of Electronic Business, Hamilton, Ontario, Canada.  Chen, E., Vahidov, R., & Kersten, G. E. (2005). Agent‐Supported Negotiations in the E‐marketplace. Int.  J. Electronic Business, 3(1), 28‐49.  Dzeng, R.‐J., & Lin, Y.‐C. (2004). Intelligent Agents for Supporting Construction Procurement  Negotiation. Expert Systems with Applications, 27, 107‐119.  Dzeng, R.‐J., & Lin, Y.‐C. (2005). Searching for Better Negotiation Agreement Based on Genetic  Algorithm. Computer‐Aided Civil and Infrastructure Engineering, 20, 280‐293.  Faratin, P., Sierra, C., & Jennings, N. R. (2002). Using Similarity Criteria to Make Issue Trade‐Offs in  Automated Negotiations. Artificial Intelligence, 142(2), 205‐237.  Foresee, F. D., & Hagan, M. T. (1997, June 9‐12, 1997). Gauss‐Newton approximation to Bayesian  regularization. Paper presented at the International Joint Conference on Neural Networks,  Houston, Texas.  Hagan, M. T., Demuth, H. B., & Beale, M. H. (1996). Neural Network Design. Boston, MA: PWS  Publishing.  Hagan, M. T., & Menhaj, M. (1994). Training Feedforward Networks with the Marquardt Algorithm.  IEEE Transactions on Neural Networks, 5(6), 989‐993.  Hornik, K. (1991). Approximation Capabilities of Multilayer Feedforward Networks. Neural Networks,  4(2), 251‐257.  Jennings, N. R., & Faratin, P. (2001). Automated Negotiation: Prospects, Methods and Challenges. Group  Decision and Negotiation, 10, 199‐215.  Kersten, G. E., & Lo, G. (2003). Aspire: An Integrated Negotiation Support System and Software Agents  for E‐Business Negotiations. International Journal of Internet and Enterprise Management, 1(3),  INR 05/08  14  293‐315.  Kersten, G. E., & Noronha, S. J. (1999). WWW‐based Negotiation Support: Design, Implementation, and  Use. Decision Support Systems, 25, 135‐154.  Kwon, O., Shin, J. M., & Kim, S. W. (2006). Context‐aware Multi‐agent Approach to Pervasive  Negotiation Support Systems. Expert Systems with Applications, 31, 275‐285.  Lee, W.‐P. (2004). Towards Agent‐based Decision Making in the Electronic Marketplace: Interactive  Recommendation and Automated Negotiation. Expert Systems with Applications 27, 665‐679.  MacKay, D. J. C. (1992). Bayesian Interpolation. Neural Computation, 4(3), 415‐447.  Maes, P., Guttman, R., & Moukas, A. (1998). Agents That Buy and Sell. Comm. ACM, 42(3), 81‐91.  Matwin, S., Szapiro, T., & Haigh, K. (1991). Genetic Algorithms Approach to a Negotiation Support  System. IEEE Transactions on Systems, Man and Cybernetics, 21(1), 102‐114.  Mudgal, C., & Vassileva, J. (2000, July 2000). Bilateral Negotiation with Incomplete and Uncertain  Information: A Decision‐Theoretic Approach Using a Model of the Opponent. Paper presented  at the Workshop on Cooperative Information Agents (CIA IV), Boston.  Oprea, M. (2002). An Adaptive Negotiation Model for Agent‐based Electronic Commerce. Studies in  Informatics and Control, 11(3), 271‐279.  Vahidov, R., Chen, E., & Feng, Z. (2005, June 10‐13, 2005). Experimental Evaluation of Agent‐Supported  e‐Negotiations. Paper presented at the Group Decision and Negotiation'2005 (CD‐ROM  Proceedings), Vienna, Austria.  Zeng, D., & Sycara, K. (1998). Bayesian Learning in Negotiation. International Journal of Human‐ Computer Studies 48, 125‐141.      
work_zjrp2n44cbgermmovdibep5so4	----	경영과학 연구실 – Operations Research Laboratory 콘텐츠로 바로가기 경영과학 연구실 Operations Research Laboratory 메뉴 HOME Members Professor Students Alumni Research Area Lecture Project (Hyundai Motors) Industrial Cluster for New Technology Research A research on the personalized treatment recommendation system Markov Decision Process in Medical Domains Review of Patient Similarity Product Line Optimization Publications ORL Seminar 2021 ORL Seminar 2020 ORL seminar 2019 ORL Seminar 2018 ORL seminar 2017 ORL Workshop ORL Workshop 2021 winter essay writer ■ Recent Activities [July 2020]  “스마트미디어 서비스 연구센터 활동 시작” [June 2020]  “현대자동차 그룹과의 신기술리서치를 위한 산학 클러스터” [June 2020]  “의료 빅데이터를 이용한 개인맞춤 의료행위 추천 시스템에 관한 연구” 3차년도 과제 승인”   ■ Research Area Our present research area includes: Big Data Analytics/Machine Learning Optimization/Game Theory Telecommunication and Media Policy Performance Analysis   ■ Lecture Reinforcement Learning Game Theory Nonlinear Optimization ■ 최근 연구 과제 의료 빅데이터를 이용한 개인맞춤 의료행위 추천 시스템에 관한 연구. (2018.06 ~ 2021.05) 비정형의 빅데이터를 활용하여 경쟁환경에서 사업자의 신제품 포지셔닝과 다기간 제품 포트폴리오 선정에 도움을 줄 수 있는 전사적 의사결정 시스템의 개발, 한국연구재단. 콘텐츠 유통 환경 다변화에 따른 지상파 방송사의 수요 연구 및 전략 연구, 한국방송공사. 콘텐츠산업 발전을 위한 게임이론적 모델에 관한 연구, 한국연구재단. 개방형 콘텐츠 전달 인프라스트럭처 구축을 위한 이종 콘텐츠전달망들 간의 상호연동 기술 개발, 한국방송통신전파진흥원. ■ 석•박사 신입생 모집 Now we welcome new researchers. You need to have at least a bachelor degree in IE, Mathematics, Economics, EE or any other related majors. Tuition fee will be supported to some extent. If you are interested in any of our study area, please feel free to contact us. We sincerely looking forward to your joining. Contact info: – E-mail: shoh0320@yonsei.ac.kr (Sang-Ho Oh, Ph.D. Student; applicants please send an e-mail with your c.v. and optional supplementaries) – TEL: 82+ 2-2123-7473 – Address: D1008, 4th Engineering bldg, 50 Yonsei-ro, Seodaemun-gu, Seoul, Republic of Korea 검색: 검색 최근 글 글 목록 글 목록 월 선택 2017년 9월  (1) HOME Members Professor Students Alumni Research Area Lecture Project (Hyundai Motors) Industrial Cluster for New Technology Research A research on the personalized treatment recommendation system Markov Decision Process in Medical Domains Review of Patient Similarity Product Line Optimization Publications ORL Seminar 2021 ORL Seminar 2020 ORL seminar 2019 ORL Seminar 2018 ORL seminar 2017 ORL Workshop ORL Workshop 2021 winter 경영과학 연구실 Proudly powered by WordPress 
work_zlzzgnksf5b4xoh3aqdbz4mele	----	Integrating planning and scheduling in workflow domains Integrating planning and scheduling in workflow domains Marı́a Dolores R-Moreno a,*, Daniel Borrajo b, Amedeo Cesta c, Angelo Oddi c a Departamento de Automática, Universidad de Alcalá, Carretera Madrid-Barcelona, Km. 33,600, 28871 Alcalá de Henares, Madrid, Spain b Departamento de Informática, Universidad Carlos III de Madrid, Avda. de la Universidad, 30, 28911 Leganés, Madrid, Spain c ISTC-CNR-Italian National Research Council, Viale Marx 15, I-00137 Rome, Italy Abstract One of the main obstacles in applying AI planning techniques to real problems is the difficulty to model the domains. Usually, this requires that people that have developed the planning system carry out the modeling phase since the representation depends very much on a deep knowledge of the internal working of the planning tools. On some domains such as business process reengineering (BPR), there has already been work on the definition of languages that allow non-experts entering knowledge on processes into the tools. We propose here the use of one of such BPR languages to enter knowledge on the organisation processes to be used by planning tools. Then, planning tools can be used to semi-automatically generate business process models. As instances of this domain, we will use the workflow modeling tool SHAMASH, where we have exploded its object oriented structure to introduce the knowledge through its user-friendly interface and, using a translator transform it into predicate logic terms. After this con- version, real models can be automatically generated using a planner that integrates planning and scheduling, IPSS. We present results in a real workflow domain, the telephone installation (TI) domain. � 2006 Elsevier Ltd. All rights reserved. Keywords: Planning and scheduling; Workflow management systems; Business process reengineering; Makespan minimisation; Time windows 1. Introduction In the last years, companies and markets have quickly grown in complexity so they are looking for flexible and dynamic ways to efficient manage their resources and processes. Due to the competitive nature of businesses and the strong pressure of the market, quality initiatives and the gradual improvement of processes are not enough for the continuous and dynamic changes that current organisations need. Levels of changes so radical need new and powerful tools that can allow the efficient redesign and management of the organisation. Also improving cus- tomer service and increasing customer retention is gaining importance. This is the main objective of BPR (Hammer & Champy, 1993). Over the past few decades, BPR has become fashionable among medium and big enterprises due to its capability to present solutions that improve this type of management. Although there have already been many approaches to the computer-aided design of processes, very few have focused on the automatic generation of process models that have in mind the organisation resources as well as their capabilities and availability. Once the organisation has been studied in depth from a process and resources perspective, corresponding models are modeled in order to handle processes and resources computationally. Business processes are usually repre- sented as workflow, that is, computerised models within which all the parameters needed for the completion of the processes can be defined: resources involved, orders, tasks, conditions, goals, quality criteria, information flow, etc. Workflow management systems (WFMSs) (Leymann & Roller, 1994; Medina-Mora, Winograd, & Flores, 1993; Mohan, 1997) have been deployed in sectors like insurance, banking, accounting, manufacturing, telecommunications, 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.05.027 * Corresponding author. Tel.: +34 918856607. E-mail addresses: mdolores@aut.uah.es (M.D. R-Moreno), dborrajo@ ia.uc3m.es (D. Borrajo), a.cesta@istc.cnr.it (A. Cesta), a.oddi@istc.cnr.it (A. Oddi). www.elsevier.com/locate/eswa Expert Systems with Applications 33 (2007) 389–406 Expert Systems with Applications mailto:mdolores@aut.uah.es mailto:dborrajo@ ia.uc3m.es mailto:dborrajo@ ia.uc3m.es mailto:a.cesta@istc.cnr.it mailto:a.oddi@istc.cnr.it administration and customer service (Lydiard, Jarvis, & Drabble, 1999) but it does not have significant commercial impact yet, because of problems related to the cost of pro- cess modeling and the inflexibility in their execution. We can see WFMSs as a set of methods and technolo- gies that allows modeling and managing some of the pro- cesses that happen in the company (there are some processes that will likely never be modeled in workflow sys- tems because, they are either too difficult to formalize, or too dynamic). It can provide active support to a business process by controlling the routing of work around the organisation automatically. This is done based on the busi- ness processes models describing the flow, the decisions, the exceptions, the resources to be used, etc. WFMS co-ordinates user and system participants, together with the appropriate data resources, which may be accessible directly by the system or off-line to achieve defined goals by setting deadlines. The co-ordination involves passing tasks to participants’ agents, and ensuring that all complete their tasks successfully. In case of excep- tions, actions to solve the problem can be triggered, or human operators alerted. Optimising the organisation procedures, routines and resource management, several aspects must be considered. Examples are the activities or tasks that should be per- formed; the organisation model that describes the roles of each agent (software or human), who can perform what in the organisation; and the information model that describes which information is needed to perform an activity. In the last few years an increasing development of doc- umentation tools and/or process modeling techniques have emerged that represent and reason computationally about the knowledge of the current processes and resources. Usu- ally, the task of defining those models is performed with the aid of a set of tools that provide a graphical represen- tation of them, together with the relations among the activities that occur within the processes. The human is usually an expert in the processes that take place in the organisation. Many issues need to be considered on this task and implemented like the reusability of past processes, accessibility to the models by the different agents in the organisation, consistency of usage, or selection of the right model. Prior to WFMS, many enterprises created special-purpose tailored applications to support their processes. The advan- tage of WFMS-based solutions is that the workflow repre- sentation is explicit, and separated from the application code. This means that a WFMS can be customised quickly to support a new business or process, and that workflows are relatively easy to modify when a process changes. Cur- rent WFMS do not address all aspects of the problem, however. In general, they do not deal with scheduling, that is, resources management/allocation. Similarly, while they provide means of generating exception events when things go wrong, they do not have a built-in re-planning function or automatic methods for adapting the workflow model according to those exceptions. They do, however, provide interfaces so that application-specific modules performing these functions can be integrated. Workflow systems hold the promise of facilitating the everyday operation of many enterprises and work environ- ments. Despite the popularity of these products, there is still a lack of maturity in some respect, i.e., a lack of a semantic associated to the models or an easy way to reason about that semantic, that could be overcome using tech- niques coming from other fields such as artificial intelli- gence (AI). Without any doubt, the application of AI techniques to Workflow Management systems has created a big expecta- tion. The AI community and in particular the planning and scheduling field, has been applying successful techniques in different and complex domains like robotics, satellites or military logistics. In these domains, there are activities that must be performed (planning) in a temporal horizon that consume or produce resources (scheduling). During execu- tion, completion of activities, and delays and other prob- lems are detected to take the appropriate measures (rectify the situation, or in more drastic cases, a new plan) to satisfy the goals. In order to represent this information, rich representation models are needed, the majority of them based on predicate logic as is the case of the planning standard language, PDDL2.2 (Edelkamp & Hoffmann, 2004). There is also the HTN representation where planning problems and operators are organised into a set of tasks. High level tasks are reduced into a set of lower levels and the way to do it can be done in several ways as in Yang (1997) and Kuter et al. (2005). In the past, some researchers saw the advantages of the integration of AI planning and scheduling for workflow generation, as shown by the existence of a Technical Co- ordination Unit of the European research network on plan- ning and scheduling, PLANET (PLANET), on applications of planning and scheduling to workflow. This has lead to some exploratory work reflected in a Roadmap PLANET and some related work (Hannebaeur, 1999; Kearney & Borrajo, 2000; Myers & Berry, 1999; R-Moreno & Kear- ney, 2002). The aim that we want to achieve with this integration is double. From one hand, given that the majority of BP tools are based on objects and rules, we propose to trans- late this knowledge into first order predicate logic, in par- ticular into the planning domain definition language PDDL2.2. On the other hand, we propose an integrated P&S fram- kework to automatically solve BRP problems. In the liter- ature there are basically two approaches to solve this type of problems. The component based approach (Cesta, Pecora, & Rasconi, 2004), where the two subproblems of planning and scheduling are just solved one after the other, and the integrated approach (Ghallab & Laruelle, 1994; Tate, Drabble, & Kirby, 1994) where there is an uniform representation without the decomposition over two sequen- tial subproblems. We believe that systems that integrate 390 M.D. R-Moreno et al. / Expert Systems with Applications 33 (2007) 389–406 https://isiarticles.com/article/489 
work_zmmy4lrbq5dwvnclwn42myragq	----	doi:10.1016/j.eswa.2007.05.037 Available online at www.sciencedirect.com www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 2858–2869 Expert Systems with Applications An efficient bit-based feature selection method Wei-Chou Chen a, Shian-Shyong Tseng a, Tzung-Pei Hong b,* a Department of Computer and Information Science, National Chiao Tung University, Hsinchu 300, Taiwan, ROC b Department of Electrical Engineering, National University of Kaohsiung, Kaohsiung 811, Taiwan, ROC Abstract Feature selection is about finding useful (relevant) features to describe an application domain. Selecting relevant and enough features to effectively represent and index the given dataset is an important task to solve the classification and clustering problems intelligently. This task is, however, quite difficult to carry out since it usually needs a very time-consuming search to get the features desired. This paper proposes a bit-based feature selection method to find the smallest feature set to represent the indexes of a given dataset. The pro- posed approach originates from the bitmap indexing and rough set techniques. It consists of two-phases. In the first phase, the given dataset is transformed into a bitmap indexing matrix with some additional data information. In the second phase, a set of relevant and enough features are selected and used to represent the classification indexes of the given dataset. After the relevant and enough fea- tures are selected, they can be judged by the domain expertise and the final feature set of the given dataset is thus proposed. Finally, the experimental results on different data sets also show the efficiency and accuracy of the proposed approach. � 2007 Elsevier Ltd. All rights reserved. Keywords: Feature selection; Bitmap indexing; Rough set; Classification; Clustering 1. Introduction Feature selection is about finding useful (relevant) fea- tures to describe an application domain. Selecting relevant and enough features to effectively represent and index the given dataset is an important task to solve the classification and clustering problems intelligently. This task is, however, quite difficult to carry out since it usually needs an exhaus- tive search to get the features desired. In the past, some approaches have been proposed to solve the feature selec- tion problem (Almullim et al., 1991; Chen, Tseng, Chen, & Jiang, 2000, 2002, 2003; Doak, 1992; Gonzalez & Perez, 2001; Huang & Tseng, 2004; John et al., 1994; Lee, Chen, Chen, & Jou, 1997; Liu et al., 1996; Liu & Setiono, 1998; Quinlan, 1986; Yu, 2001). These approaches can roughly be classified into the following two strategies: 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.05.037 * Corresponding author. E-mail addresses: sstseng@cis.nctu.edu.tw (S.-S. Tseng), tphong@nuk. edu.tw (T.-P. Hong). 1. Optimal strategy: This kind of approaches considers all the subsets of a given feature set (Almullim et al., 1991; Schlimmer et al., 1993; Wu, 1999). Some searching tech- niques, such as branch and bound, may be adopted to reduce the search space. For example, Liu et al. pro- posed a special feature selector (Liu et al., 1996), which randomly produced feature subsets according to the Las Vegas algorithm (Brassard, 1996). It thus searched the entire solution spaces and guaranteed to get an optimal feature set. 2. Heuristic strategy: This kind of approaches prunes search spaces according to some heuristics. The results obtained by these approaches are usually not optimal, but within a short time (Zhong, Dong, & Ohsuga, 2001). There are three typical heuristic approaches for feature selection, including forward selection, backward selection and bi-directional selection. The forward selec- tion approach initializes the desired feature set as null and then adds features into it until the results are satis- factory (Miao, 1999; Skowron & Rauszer, 1992; Yu, 2001). The backward selection approach initializes the mailto:sstseng@cis.nctu.edu.tw mailto:tphong@nuk.edu.tw mailto:tphong@nuk.edu.tw W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 2859 desired feature set as all the given features and then removes unnecessary features from it (Choubey, 1998; Yu, 2001). The bi-directional selection approach initial- izes the desired feature set as a partial feature set, and then either puts good features into it or eliminates bad features from it (Doak, 1992). In the past, we proposed a bitwise indexing method based on a given feature set to accelerate case matching in CBR (Chen, Tseng, Chang, & Jiang, 2001, 2002). In this paper, we further investigate the determination of the appropriate feature set. We propose a two-phase feature selection approach to discover significant feature sets from a given database table, and use them to further investigation. The proposed feature selection approach originates from the bit- map indexing (O’Neil & Quass, 1997; Wu et al., 1998) and rough set techniques (Pawlak, 1982, 1991). Naturally, it is designed to discover optimal feature sets for a given dataset since the proposed method is originated from the rough set theory. The Experimental results also show the efficiency and accuracy of the proposed approach. This paper is organized as follows. Some related works are reviewed in Section 2. The proposed feature selection method and some corresponding definitions and algo- rithms are stated in Section 3. The time and space complex- ities of the proposed algorithms are analyzed in Section 4. Experimental results are shown in Section 5. Conclusions are finally given in Section 6. 2. Review of feature selection and rough sets Feature selection is about finding useful (relevant) fea- tures to describe an application domain. The problem of feature selection can formally be defined as selecting min- imum features M0 from original M features where M0 5 M such that the class distribution of M0 features is as similar as possible to M features (Last, Kandel, & Maimon, 2001). Generally speaking, the function of fea- ture selection is divided into three parts: (1) simplifying data description, (2) reducing the task of data collection, and (3) improving the quality of problem solving. The benefits of having a simple representation are abundant such as easier understanding of problems, and better and faster decision making. In the field of data collection, having less features means that less data need to be col- lected. As we know, collecting data is never an easy job in many applications because it could be time-consuming and costly. Regarding the quality of problem solving, the more complex the problem is if it has more features to be processed. It can be improved by filtering out the irrele- vant features which may confuse the original problem, and it will win the better performance. There are many discussions about feature selection, and many existing methods to assist it, such as GA technology (Raymer, Punch, Goodman, & Kuhn, 2000), entropy measure (Huang & Tseng, 2004), and rough set theory (Tseng, Jothishankar, & Wu, 2004; Yu, 2001). Next, the rough set theory is briefly reviewed. The rough set theory, proposed by Pawlak in 1982 (Pawlak, 1982), can serve as a new mathematical tool for dealing with data clas- sification problems. It adopts the concept of equivalence classes to partition training instances according to some criteria. Two kinds of partitions are formed in the mining process: lower approximations and upper approximations. Rough sets can also be used for feature reduction. The fea- tures that do not contribute to the classification of the given training data are removed. The concepts of equiva- lence classes and approximations are quite suitable to gen- erate the bit-based class vectors and record vectors, which can then be directly and efficiently transformed to the bit- wise indexing matrixes for CBR systems (Yang et al., 2000). This paper thus adopts these concepts to solve the feature selection problem. 3. The proposed bitmap-based feature selection method As we mentioned above, we proposed a heuristic feature selection approach, called the bitmap-based feature selec- tion method with discernibility matrix (Chen, Yang, & Tseng, 2002), to find a nearly optimal feature set. However, finding the optimal solutions of feature selection is still needed in some applications. Although, some exhaustive search methods can guarantee the optimality of selected feature sets, the computation cost may be very high. In this section, we thus consider finding an optimal solu- tion via the rough set techniques and the bit-based indexing method for the feature selection. The proposed approach encodes a given data set into a bit vector matrix and uses bit-processing operations on them to reduce the computa- tion time. The proposed approach consists of several main steps, as shown in Fig. 1. There are two-phases in the proposed algorithm – bit- map indexing phase and feature selection phase. In the bit- map indexing phase, the given dataset is transformed into a bitmap indexing matrix with some additional data infor- mation. In the feature selection phase, a set of relevant and enough features are selected and used to represent the dataset. The details of the two-phases are described in following sub-sections. 3.1. Problem definitions Let T denote a target table in a database, R denote the set of n records in T, and C denote the set of m features in T, R can then be represented as {R1, R2, . . ., Rn}, where Ri is the ith record. C can be represented as {C1, C2, . . ., Cm}, where Cj is the jth feature. The first m�1 elements in C are condition features and the last one, Cm, is a decision feature. Let Vj denote the domain of Cj. Vj can then be rep- resented as {Vj1, Vj2, . . ., Vjrj}, where each element is a pos- sible value of Cj and rj is the number of possible values of Cj. Let Vj(i) denote the value of Cj in record Ri, Vj(i) 5 null. Table 1 shows an example of a target table T with ten records R = {R1, R2, . . ., R10} and five features Table 2 The record vectors and class vectors from Table 1 Feature Feature-value Record vector Class vector C1 V11 1100100000 110 V12 0011010000 111 V13 0000001111 111 C2 V21 1110000000 100 V22 0001111000 011 Fig. 1. The flowchart of the proposed feature selection approach. Table 1 An example of a target table C1 C2 C3 C4 C5 R1 M L 3 M 1 R2 M L 1 H 1 R3 L L 1 M 1 R4 L R 3 M 2 R5 M R 2 M 2 R6 L R 3 L 3 R7 H R 3 L 3 R8 H N 3 L 3 R9 H N 2 H 2 R10 H N 2 H 1 2860 W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 C = {C1, C2, C3, C4, C5}. C5 is a decision feature and the others are condition features. The purpose of this paper is to find the one of the small- est feature set to effectively index the given table. The def- initions and algorithms used in the bitmap indexing phase and in the feature selection phase are described below. V23 0000000111 111 C3 V31 1001011100 111 V32 0110000000 100 V33 0000100011 110 C4 V41 1011100000 110 V42 0100000011 110 V43 0000011100 001 C5 V51 1110000001 100 V52 0001100010 010 V53 0000011100 001 3.2. Bitmap indexing phase In this phase, the target table is first transformed into a bitmap indexing matrix with some additional classification information. Let bi is a bit in a bit vector. Let ONEk denote the bit string of length k, with all the bits set to 1, ZEROk denote the one with all the bits set to 0, and UNIQUEk denote the one, with only one bit set to 1 and the others set to 0. A record vector, which is used to keep the informa- tion of the records with a specific value of a feature, is defined below. Definition 1 (Record vector). A record vector RVjk is a bit string b1, b2, . . ., bn, with bi set to 1 for Vj(i) = Vjk and set to 0 otherwise, where 1 6 j 6 m, 1 6 k 6 rj, and 1 6 i 6 n. RVjk thus keeps the information of the records with the kth possible value of the feature Cj. For example in Table 1, C1 has three possible values {M, L, H}. The record vec- tor for C1 = M is 1100100000 since the first, second and fifth records have this feature value. Similarly, the record vector for C1 = L is 0011010000 and for C1 = H is 0000001111. All the record vectors are shown in the third column of Table 2. A class vector, which is used to keep the information of the classes (values of the decision feature) with a specific value of a feature, is defined below. Definition 2 (Class vector). A class vector CVjk is a bit string b1, b2, . . ., brm, with bi set to 1 if RVjk \ RVmi 5 ZEROn, and set to 0 otherwise, where rm is the number of possible values of Cm and n is the number of records in R. Here, the ‘‘AND’’ bitwise operator is used for the inter- section in Definition 2. CVjk thus keeps the information of the classes related to the kth possible value of the feature Cj. For example in Table 2, the record vector (RV11) for C1 = M is 1100100000 and the one (RV51) for C5 = 1 is 1110000001. Since, the bitwise intersection of 1100100000 and 1110000001 is 1100000000, not equal to ZERO10, the first bit in RV11 is thus 1. Similarly, the second and third bits in RV11 are 1 and 0 from the intersection results of RV11 with RV52, and with RV53. the class vector CV11 is thus 110. All the class vectors are shown in the fourth col- umn of Table 2. Formally, a class vector CVjk can be obtained by the following FindClassVector algorithm. W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 2861 Algorithm 1 (FindClassVector). Input: Record vector RVjk. Output: Class vector CVjk. Step 1: Set CVjk to ZEROrm. Step 1: For each i, 1 6 i 6 rm, set the ith bit of CVjk to 1 if RVjk \ RVmi 5 ZEROn; otherwise, set it to 0. Step 1: Return CVjk. Definition 3 (Feature-value vector). A feature-value vector Fjk is the concatenation of RVjk and CVjk. For example, the feature-value vector F11 in Table 2 is 1100100000110, which is RV11 concatenated with CV11. All the feature-value vectors for a feature are then collected together as a feature matrix. This is defined below. Definition 4 (A feature matrix for a feature). A feature matrix Mj for the feature Cj is denoted F j1 F j2 .. . F jrj 2 6664 3 7775, where rj is the number of possible values in Cj. For example, the feature matrix M1 in Table 2 is show as follows: M 1 ¼ 1100100000110 0011010000111 0000001111111 2 64 3 75 The bits with underlines are class vectors. From the def- inition of the feature matrix, it is easily derived that apply- ing the bitwise operator ‘‘OR’’ on all the record vectors in a feature matrix will get the ONEn vector, and applying the bitwise operator ‘‘AND’’ on any two record vectors in a feature matrix will get the ZEROn vector. Note that, the ‘‘OR’’ and ‘‘AND’’ operators are defined for executing the ‘‘OR’’ and ‘‘AND’’ operations on all corresponding bits of the given two bit vectors. Thus, if we apply the bit- wise operator ‘‘XOR’’ on all the record vectors in a feature matrix, we will also get the ZEROn vector. Take M1 as an example. The result for 1100100000 OR 0011010000 OR 0000001111 is 1111111111. The result for 1100100000 AND 0011010000 is 0000000000. The result for 1100100000 XOR 0011010000 XOR 0000001111 is 0000000000. Definition 5 (A feature matrix for a table T). A feature matrix M for a table T is denoted M 1 M 2 .. . M m 2 6664 3 7775, where m is the number of features in T. For example, the matrix composed of the bit strings from columns 3 and 4 of Table 2 is the feature matrix for the data given in Table 1. The feature matrix for a table is then input to the feature selection phase to find relevant and enough features. 3.3. Feature selection phase In this phase, we want to find a set of relevant and enough features to represent the given dataset. It is further divided into several stages. First, a feature-based spanning tree is built for cleansing the bitmap indexing matrix. The dataset with noisy information is thus judged and filtered out according to the spanning tree. The cleansed, noisy-free bitmap indexing matrix can then be used to determine the optimal feature set for some classification and clustering problems. 3.3.1. Creating cleansing tree Before the feature selection phase is executed, the cor- rectness of the target table needs to be verified. If there are some records in the target table with the same values of all condition features, but with different ones of the deci- sion feature, they are treated as noise records and are fil- tered out from the target table. Intuitively, every two records can be compared to find out the inconsistent records in the target table. Its time complexity is O(n2m), where n is the number of records and m is the number of features. Below, we propose the concept of a cleansing tree to decrease the time complexity to O(nmj), where j is the maximum number of possible feature values of a feature and n is usually much larger than j in general classification and clustering problems. The formation of a cleaning tree depends on the given feature order. We thus have the fol- lowing definition. Definition 6 (Spanned feature order). A spanned feature order O is a permutation consisting of all the condition features in a target table T. For example in Table 1, hC1, C2, C3, C4i can be a spanned feature order. When a spanned feature order is given, a cleansing tree can then be built according to it. The definition of a cleansing tree is first given below. Definition 7 (Cleansing tree). A cleansing tree Ctree is a tree with a root denoted root[Ctree]. Every node x in the tree corresponds to a feature value. A node y is the parent of a node x if the feature of y precedes the feature of x in the given spanned feature order. A node z is the sibling of a node x if they have the same feature, but different values. A structure of a cleansing tree is shown in Fig. 2. Its maximum height is m�1, where m is the number of features in a decision table T. Each node x has three pointers, which are p[x], left-child[x] and right-sibling[x], respectively point- ing to its parent node, its leftmost child node and its first right-sibling node. It also contains two additional infor- mation, record[x] and class[x], which indicate the associ- ated record and class vectors of x. If node x has no child, then left-child[x] = NIL; if node x is the rightmost child of its parent, then right-sibling[x] = NIL. record class record class record class record class record class record class record class root[Tree] Fig. 2. The structure of a cleansing tree. 2862 W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 As mentioned above, records may have the same values of all condition features, but different value of the decision feature. These records are called inconsistent. Inconsistent records can also be found out when the cleansing tree is built. The building algorithm uses the valid mask vector to find the consistent records. The valid mask vector is defined as follows. Definition 8 (Valid mask vector). A valid mask vector ValidMask for a target table T a bit string b1, b2, . . ., bn, with bi set to 1 if the ith record Ri is not inconsistent with other records, and set to 0 otherwise. The cleansing tree for a given spanned feature order can be built by the following CreateCleansingTree algorithm. The ValidMask is initially set to ONEn, and will be modified along with the execution of the CreateCleansing- Tree algorithm. Algorithm 2 (CreateCleansingTree). Input: A feature matrix M, the valid mask ValidMask and a spanned feature order O. Output: The valid mask ValidMask. Step 1: Create an empty node x and set it as the root node. Step 2: Initialize record[x] = ONEn, class[x] = ONErm and depth = 0, where the variable depth is used to represent the depth of the node x in the cleans- ing tree. Step 3: Set px = x, where px is used to keep the current parent node. Step 4: If class[x] is not equal to UNIQUErm and depth is not equal to m�1, do Step 5 to build the child nodes of node x; otherwise, go to Step 7. Step 5: Let Cj be the current feature in the spanned fea- ture order to be considered. For each feature- value vector Fjk in a feature matrix Mj for Cj, if (record[px] AND RVjk) 5 ZEROn, do the follow- ing sub-steps: Step 5.1: Create an empty node y. Step 5.2: If left_child[x] = NIL, consider y as a child node of x and set p[y] = x and left_child[x] = y; other- wise, consider y as a sibling node of x and set p[y] = p[x] and right_sibiling[x] = y. Step 5.3: Set record[y] = (record[p[y]] AND RVjk) and clas- s[y] = (class[p[y]] AND CVjk). Step 5.4: If depth = m�1 and class½y� 6¼ UNIQUErm , set ValidMask = (record[y] XOR ValidMask). Step 5.5: Set x = y. Step 6: If left_child[px] 5 NIL, set x = left_child[px], depth = depth + 1 and go to Step 3. Otherwise, do the next step. Step 7: If right_sibiling[x] 5 NIL, x = right_sibiling[x] and go to Step 3; otherwise, set x = p[x] and do the next step. Step 8: If x 5 Tree[root], go to Step 7; otherwise, return ValidMask and stop the algorithm. For example, the cleansing tree for the data in Table 1 with the spanned feature order hC1, C2, C3, C4i will be built as shown in Fig. 3. At first, the root node is generated and all the bits in record[root] and class[root] are set to 1. Since, class[root] is not equal to UNIQUE3, and the current depth is 0, not equal to m�1, the next step is executed to build the child nodes of the root. The first feature C1 in the spanned feature order is considered. Since, it has three possible values and (record[root] AND RV1k), k = 1 to 3, is not equal to ZERO10, three nodes, represented as nodes 1, 2 and 3, are created as the children of the root. Since, node 1, the left- child node of the root, is not NIL, it is then processed to generate its child nodes in the same way. Nodes 4 and 5 are then created for the second feature C2 in the spanned fea- ture order. Since, class[node 4] has been equal to ONE10, the sibling of node 4, which is node 5, is then considered. Since, class[node 5] has also been equal to ONE10, the sibling of node 5, is then considered. But since node 5 has no sibling, its parent node, node 1 is considered. The sibling of node 1, which is node 2 is then processed. The same procedure is then executed until the whole cleansing tree is generated. The numbers at the left of the nodes in Fig. 3 indicate the order built. In node 15, the second and third bits of the class vector are both ‘‘1’’. It means that the correspond- ing record vectors will have more than one ‘‘1’’. The corre- sponding records with bit ‘‘1’’ are then inconsistent since 1100100000 1101 0000001111 11130011010000 1112 0001010000 01170010000000 1006 0000000111 11112 0000000100 00113 0000001000 10011 0000000011 11015 1111111111 111root C 1 C2 C3 C 4 1100000000 1004 0000100000 0105 0001010000 0118 0000010000 001100001000000 0109 0000000011 11014 Fig. 3. Cleansing tree with feature spanned order hC1, C2, C3, C4i. W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 2863 their values of all condition features are the same, but their values of the decision feature are different. In this example, the ninth and tenth records are inconsistent. The Valid- Mask are thus modified from ‘‘1111111111’’ to ‘‘1111111100’’. 3.3.2. Finding appropriate spanned order In the above example, the spanned feature order O is set as hC1, C2, C3, C4i. Different orders will apparently affect the performance of the cleansing spanning trees built. A cleansing spanning tree with a better spanned feature order can reduce the space and time complexities. In the past, there were some famous tree structures for classification, such as the decision-tree approach (Quinlan, 1986), which was based on the entropy theory to select the next best fea- ture. In order to reduce the computational complexity for evaluating the spanning order of features, the following heuristics are thus proposed. H1: The more ‘1’ bits a record vector for a feature value has, the more weight the feature value has. H2: The more ‘1’ bit the class vector for a feature value has, the less weight the feature value has. These two heuristics show the relationship between fea- ture values and classes. If a feature value appears in most records with a single class, the weight of this feature value is relatively high. These heuristics can be used to save the computation time when compared to using the entropy the- ory. The following FindSpanOrder algorithm is thus pro- posed to determine the spanned feature sequence O of all condition features by evaluating the feature weights according to the above heuristics. Algorithm 3 (FindSpanOrder). Input: A feature matrix M for a table T Output: A spanned feature order O. Step 1: Initialize weightj = 0, where 1 6 j 6 m�1. Step 2: For each Mj in M, set: weightj Xrj k¼1 CountðRV jkÞ ½CountðCV jkÞ� 2 ; where the function Count(x) is used to count the number of ‘1’ bits in x. Step 3: Order the features in O in the descendent order of the weight values. Step 4: Return O. For example, according to the feature matrix in Table 2, the weight of each feature is calculated as shown in Table 3. The new spanned feature sequence O determined by the above algorithm is thus hC4, C2, C3, C1i, instead of the ori- ginal order hC1, C2, C3, C4i. The cleansing tree generated on the new order is shown in Fig. 4. As we can see, the cleansing tree with new feature order O = hC4, C2, C3, C1i in Fig. 4 is much smaller than that in Fig. 3. The number of nodes has decreased from 15 to 9. Therefore, the computational time of generating and tra- versing the spanning tree can be greatly reduced. 3.3.3. Cleansing feature matrix After the cleansing tree is built, the ValidMask may not be ONEn since inconsistent records may exist. The Valid- Mask is then used by the following CleansingFeatureMa- trix algorithm to remove the inconsistent records from the feature matrix. Algorithm 4 (CleansingFeatureMatrix). Input: A feature matrix M for a table T and a valid mask vector ValidMask. Output: A cleansed feature matrix M. Step 1: For each feature-value vector Fij in M, do follow- ing sub-steps: Step 1.1: RVij = RVij AND ValidMask. Step 1.2: CVij = FindClassVector (RVij). Step 2: Return M. 2864 W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 For example, the ValidMask is set to ‘‘1111111100’’ after the cleansing tree for Table 1 is built. Since, the ninth and tenth bits of the ValidMask are 0, the CleansingFeatu- reMatrix algorithm will set these two bits of all the record vectors in Table 2 to 0. The class vector of each feature value is then recalculated by the FindClassVector algorithm according to its new record vector. The revised feature matrix is shown in Table 4. 3.3.4. Some definitions on feature sets For effectively distinguishing the classes from the feature values, we must extend the concepts related to a single fea- ture to a feature sets. The following definitions are thus needed. Definition 9 (Power of a feature set). Cs is called the s- power of a feature set C, if each element in Cs is composed of s distinct condition features from C, 1 6 s 6 m�1. Thus, we have C1 = C. For example, the power set C1 for the data in Table 1 is {{C1}, {C2}, {C3}, {C4}}. The power set C2 is {{C1,C2}, {C1,C3}, {C1,C4}, {C2,C3}, {C2,C4}, {C3,C4}}. Let jCsj denote the cardinality of Cs. Then: jCsj ¼ m � 1 s � � : Let Csj denote the jth element in C s, 1 6 j 6 jCsj. Csj is then a feature set. Also let V sj denote the domain of C s j, rsj denote the number of possible values in V s j, and V s jk denote the kth feature value of Csj. Each feature set can be represented by a name vector, defined below. Definition 10 (Name vector of a feature set). The name vec- tor NV sj of a feature set C s j is a bit string b1, b2, . . ., bm�1, with bi set to 1 if featureCi is included in C s j and set to 0 otherwise. For the above example, C11 denotes the first element in C 1, which is {C1}. The name vector NV 1 1 is then 1000 since only C1 is included in C 1 1. For another example, C 2 = {{C1, C2}, {C1,C3}, {C1,C4}, {C2,C3}, {C2,C4}, {C3,C4}}. C 2 1 denotes the first element in C2, which is {C1,C2}. The name vector NV 21 is then 1100 since C1 and C2 are included in C 2 1. Similar to a single feature, some terms related to a feature set is defined below. Definition 11 (Record vector of a feature set). A record vec- tor RV sjk of a feature set value C s jk is a bit string b1, b2, . . ., bn, with bi set to 1 for V s jðiÞ¼ V s jk and set to 0 otherwise, where 1 6 j 6 j Csj and 1 6 k 6 rsj. Table 3 Calculating the weight of each feature Feature Weight Old order New order C1 3/4 + 3/9 + 4/9 = 1.53 1 4 C2 3/1 + 4/4 + 3/9 = 4.33 2 2 C3 5/9 + 2/1 + 3/4 = 3.31 3 3 C4 4/4 + 3/4 + 3/1 = 4.75 4 1 RV sjk thus keeps the information of the records with the kth possible value of the feature set Csj. A class vector, which is used to keep the information of the classes (values of the decision feature) with a specific value of a feature set, is defined below. Definition 12 (Class vector of a feature set). A class vector of CV sjk of a feature set value C s jk is a bit string b1, b2, . . ., b, with bi set to 1 if RV s jk\ RVmi 5 ZEROn, and set to 0 other- wise, where rm is the number of possible values of Cm and n is the number of records in R. CV sjk thus keeps the information of the classes related to the kth possible value of the feature set Csj. A feature-value vector of a feature set is defined below. Definition 13 (Feature-value vector of a feature set). A fea- ture-value vector F sjk is composed of RV s jk and CV s jk . Definition 14 (A feature matrix for a feature set). A feature matrix M sj for the feature set C s j is denoted F sj1 F sj2 .. . F sjrj 2 6664 3 7775, where s 1 6 j 6 jC j and rsj is the number of possible values in C s j. Definition 15 (s-feature matrix for a table T). An s-feature matrix Ms for a table T is denoted M s1 M s2 .. . M sjCsj 2 6664 3 7775, where 1 6 s 6 m�1. 3.3.5. Selecting a feature set In this sub-section, two algorithms are proposed to find the desired feature set. The first algorithm, named the SelectingFeatureSet algorithm, is used to find a feature set from a given s-feature matrix. If there exists a feature set which is sufficient to decide all the records in the given dataset, the feature set will be returned and the feature selection procedure stops. Otherwise, s is incremented and the SelectingFeatureSet algorithm is executed again. The second algorithm, named the CalculatingNextMatrix algorithm, derives the new feature matrix from the previ- ous feature matrix. The SelectingFeatureSet algorithm is described as follows. Algorithm 5 (SelectingFeatureSet). Input: An s-feature matrix Ms for a table T. Output: A selected feature set FS. Step 1: Initialize FS = B, j = 1. Step 2: If j 6 jCsj, do the next step; otherwise go to Step 7. Step 3: Set k = 1, where k is used to keep the number of the value currently processed in a feature set Csj. 0000011100 00130100000011 11021011100000 1101 1010000000 1004 0001100000 0105 0000000011 1107 0000000011 1108 0100000000 1006 0000000011 1109 1111111111 111root C 4 C 2 C3 C1 Fig. 4. The cleansing tree generated on the new order hC4, C2, C3, C1i. W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 2865 Step 4: If k 6 rsj, do the next step; otherwise go to Step 6. Step 5: If CV sjk 6¼ UNIQUErm , set j = j + 1 and go to Step 2; otherwise set k = k + 1 and go to Step 4. Step 6: Set FS ¼ Csj; That is, for each i from 1 to m�1, set FS = FS [ {Ci} if the ith bit of the name vector NV sj for feature set C s j is equal to 1. Step 7: Return FS. Take the data in Table 1 as an example to illustrate the above algorithm. s is set at 1 at the beginning. The 1-fea- ture matrix M1 for the data is the same as the feature matrix M found before. The SelectingFeatureSet algorithm will examine the 1-feature sets one by one. The first element M 11, which is {C1}, is then processed. The class vector CV 1 11 for the first feature value C111 is 110, which is not equal to Table 4 The cleansed feature matrix of Table 2 Feature Feature-value Record vector Class vector C1 V11 1100100000 110 V12 0011010000 111 V13 0000001100 001 C2 V21 1110000000 100 V22 0001111000 011 V23 0000000100 001 C3 V31 1001011100 111 V32 0110000000 100 V33 0000100000 010 C4 V41 1011100000 110 V42 0100000000 100 V43 0000011100 001 C5 V51 1110000000 100 V52 0001100000 010 V53 0000011100 001 UNIQUE3. Using the feature set {C1} can thus not com- pletely distinguish the classes. The other elements in the 1-feature matrix M1 are then processed in a similar way. In this example, no element is chosen. Thus B is returned. It means no single feature can completely distinguish the classes s is then incremented, and the SelectingFeatureSet algorithm is then executed from the new s-feature matrix. The new feature matrix can be easily derived from the pre- vious feature matrix by the following CalculatingNextMa- trix algorithm. Algorithm 6 (CalculatingNextMatrix). Input: An s-feature matrix Ms for a table T. Output: An (s + 1)-feature matrix Ms+1 for a table T. Step 1: For each j, j = 1 to jCsj� 1, do the following steps. Step 2: For each l, l = (j mod m) + 1 to m, do the fol- lowing sub-steps. Step 2.1: Set NV sþ1j ¼ NV s j OR NV 1 l . Step 2.2: Set the temporary counter k to 1. Step 2.3: For each feature-value vector F sjx in M s j, 1 6 x 6 jCsjj, do the following sub-steps: Step 2.3.1: For each feature-value vector F 1ly in M 1 l , 1 6 y 6 jC1lj, do the following sub-steps: Step 2.3.1.1: Set RV sþ1jk ¼ RV jx AND RV 1 ly . Step 2.3.1.2: Set CV sþ1jk ¼ CV s jx AND CV 1 ly . Step 2.3.1.3: IF CV sþ1jk 6¼ UNIQUErm , set CV sþ1 jk ¼ FindClassVectorðRV sþ1jk Þ. Step 2.3.1.4: Set k = k+1. Step 3: Return the (s + 1)-feature matrix Ms+1. For example, the 2-feature matrix M2 for the data in Table 1 is generated from the 1-feature matrix M1 as follows. The name vector for feature C21 is first calculated. Thus: 2866 W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 NV 21 ¼ NV 1 1 OR NV 1 1 ¼ 1000 OR 0100 ¼ 1100: The feature-value vector F 211 in M 2 1 is then calculated. The record vector is found as follows: RV 211 ¼ RV 1 11 AND RV 1 21 ¼ 1100100000 AND 1110000000 ¼ 1100000000: The class vector is found as follows: CV 211 ¼ CV 1 11 AND CV 1 21 ¼ 110 AND 100 ¼ 100: In a similar way, all the feature-value vectors in the 2-fea- ture matrix M2 can be found. The results are shown in Table 5. Note that in Step 2.2.1.2, the class vector derived by the bitwise ‘‘AND’’ operator denotes only the ‘‘possible’’ class distribution. For example, the feature-value vector F 221 con- Table 5 The 2-feature matrix M2 found by the CalculatingNextMatrix algorithm Feature set Feature set value Name vector Record vector Class vector C21 V 2 11 1100 1100000000 100 V 212 1100 0000100000 010 V 213 1100 0010000000 100 V 214 1100 0001010000 011 V 215 1100 0000001000 001 V 216 1100 0000000100 001 C22 V 2 21 1010 1000000000 100 V 222 1010 0100000000 100 V 223 1010 0000100000 010 V 224 1010 0001010000 011 V 225 1010 0010000000 100 V 226 1010 0000001100 001 C23 V 2 31 1001 1000100000 110 V 232 1001 0100000000 100 V 233 1001 0011000000 110 V 234 1001 0000010000 001 V 235 1001 0000001100 001 C24 V 2 41 0110 1000000000 100 V 242 0110 0110000000 100 V 243 0110 0001011000 011 V 244 0110 0000100000 010 V 245 0110 0000000100 001 C25 V 2 51 0101 1010000000 100 V 252 0101 0100000000 100 V 253 0101 0001100000 010 V 254 0101 0000011000 001 V 255 0101 0000000100 001 C26 V 2 61 0011 1001000000 110 V 262 0011 0000011100 001 V 263 0011 0010000000 100 V 264 0011 0100000000 100 V 265 0011 0000100000 010 sists of RV 221 ¼ 1000000000 and CV 2 21 ¼ 110 after Step 2.2.1.2. Since, each record belongs to only one class, the above results are not correct. In fact, the class vector CV 221 ¼ 100: Step 2.2.1.2 is used as a quick check. If CV sþ1jk 6¼ UNIQUErm , then the FindClassVector algorithm is run in Step 2.2.1.3 to find the correct class vector. After the new feature matrix is derived, the SelectingFe- atureSet algorithm is then executed again to find an appro- priate feature set. For the above example, the 2-feature matrix M2 is then input to the SelectingFeatureSet algo- rithm and the feature set FS = {C2,C4} are found and returned as the solution. After the above method is executed, the feature set FS to classify the given data set T is generated. FS may be over- fitting or under-fitting for the problem since they are derived only according to the current data set. These fea- tures are then evaluated and modified by domain experts. They thus serve as the candidates for the experts to have a good initial standpoint. 4. Time and space complexity analysis The time and space complexities of the proposed algo- rithms are analyzed in this section. Let n be the number of records, m be the number of features and c be the num- ber of classes. Also define i as the maximum possible num- ber of features in a feature set, j as the maximum number of possible values of a feature, and s as the number of itera- tions. The time complexity and space complexity of each step in the FindClassVector algorithm is shown in Table 6. The time and space complexities of each step in the Cre- ateCleansingTree algorithm is shown in Table 7. Note that the maximum amount of nodes within a Ctree is n. The time and space complexities of each step in the Find- SpanOrder algorithm is shown in Table 8. The time and space complexities of each step in the CleansingFeatureMatrix algorithm is shown in Table 9. Table 6 The time and space complexities of the FindClassVector algorithm Step no. Time complexity Space complexity Step 1 O(1) O(c) Step 2 O(jc) O(jc) Step 3 O(1) O(c) Total O(jc) O(jc) Table 7 The time and space complexities of the CreateCleansingTree algorithm Step no. Time complexity Space complexity Step 1 O(1) O(1) Step 2 O(1) O(1) Step 3 O(1) O(1) Step 4 O(nmj) O(n) Step 5 O(mj) O(n) Step 6 O(1) O(1) Step 7 O(1) O(1) Total O(nmj) O(n)* Table 8 The time and space complexities of the FindSpanOrder algorithm Step no. Time complexity Space complexity Step 1 O(m) O(m) Step 2 O(cm) O(cm) Step 3 O(clgc) O(c) Step 4 O(1) O(1) Total O(Max(cm, clgc)) O(cm) Table 9 The time and space complexities of the CleansingFeatureMatrix algorithm Step no. Time complexity Space complexity Step 1 O(mj) O(mj) Step 2 O(1) O(1) Total O(mj) O(mj) Table 10 The time and space complexities of the SelectingFeatureSet algorithm Step no. Time complexity Space complexity Step 1 O(1) O(1) Step 2 O(msjs) O(1) Step 3 O(1) O(1) Step 4 O(js) O(1) Step 5 O(1) O(1) Step 6 O(c) O(c) Step 7 O(1) O(1) Total O(msjs) O(c) Table 11 The time and space complexities of the CalculatingNextMatrix algorithm Step no. Time complexity Space complexity Step 1 O(msjs) O(msjs) Step 2 O(mj) O(mj) Step 3 O(1) O(j) Total O(msjs) O(msjs) W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 2867 The time and space complexities of each step in the SelectingFeatureSet algorithm is shown in Table 10. The time and space complexities of each step in the Cal- culatingNextMatrix algorithm is shown in Table 11. Table 12 The datasets used in the experiments Database name Class no. Condition feature no. Record no. Missing features Monk1 2 6 124 No Monk2 2 6 169 No Monk3 2 6 122 No Vote 2 16 300 No Mushroom 2 22 8124 Yes SoybeanL 19 35 683 Yes Insurance 3 27 35000 Yes 5. Experiments To evaluate the performance of the proposed method, we compare it with other feature selection methods. Our target machine is a Pentium III 1G Mhz processor system, running on the Microsoft Windows 2000 multithreaded OS. The system includes 512 K L2 cache and 256 MB shared-memory. Several datasets from the UCI Repository (Quinlan, 1986) are used for the experiments. These datasets have dif- ferent characteristics. Some have known relevant features (such as Monks), some have many classes (such as Soy- beanL), and some have many instances (such Mushroom). In addition, a large real data set about endowment insur- ances from a world-wide financial group is used to examine the usability of the proposed method. Experimental results show the proposed method can discover the desired feature sets and can thus help the enterprise to build a CBR system for their loan promotion function of customer relationship management system. The data set of insurance data uses 27 condition features to describe the states of 3 different insur- ance types. Different types of attribute values including date/time, numeric and symbolic data exist. They are all transformed into the symbolic type by some clustering methods. Six of them have missing values. The characteristics of the above datasets are summa- rized in Table 12. In the experiments, the accuracy, the number of selected features, and the time will be compared between our method and the traditional rough set method. The accu- racy is measured by the classification results of the target table. If the selected feature set can solve the problem with- out any error, 100% accuracy is reached; otherwise the accuracy is calculated by the number of correctly classified records over the total number of records. Experimental results show both methods can reach 100% accuracy. We then compare the feature sets found by these two approaches. The results are shown in Table 13. Obviously, the accuracy of all datasets is 100% since both of these two methods discover the minimal feature sets. Note that there may be more then one solution for the selected features. In Table 13, only the first selected feature set (in the alphabetical order) is listed. It is easily seen that the selected feature sets of our proposed approach and the traditional rough set approach are the same except for the SoybeanL problem. The SoybeanL problem needs too much computation time by the traditional rough set approach. The numbers of the selected features by the two approaches are shown in Table 14. Both methods get the same numbers for all problems except for SoybeanL. At last, the computation time is compared. The data sets are first loaded into the memory from the hard disk and the processing times are measured. The time is rounded to 0 if the real time is less than 0.001 seconds. The results are shown in Table 15. Consistent with our expectation, the proposed approach is much faster than the traditional rough set approach. Especially for the Insurance data, our approach needs only Table 13 The selected feature sets found by the two approaches Dataset Feature set Accuracy (%) Traditional rough set approach Bitmap-based approach Monk1 C1, C2, C5 C1, C2, C5 100 Monk2 C1–C6 C1–C6 100 Monk3 C1, C2, C4, C5 C1, C2, C4, C5 100 Vote C1–C4, C9, C11, C13, C16 C1–C4, C9, C11, C13, C16 100 Mushroom C3, C4, C11, C20 C3, C4, C11, C20 100 SoybeanL Need too much computation time C14, C20, C26, C27, C29, C30, C31, C32, C33, C34, C35 100 Insurance C4, C15, C17, C20, C22, C25 C4, C15, C17, C20, C22, C25 100 Table 14 The number of the selected features found by the two approaches Dataset Traditional RS Bitmap-based Monk1 3 3 Monk2 6 6 Monk3 4 4 Vote 8 8 Mushroom 4 4 SoybeanL 11 11 Insurance 6 6 Table 15 The CPU times needed by the two approaches Dataset Traditional RS Bitmap-based Monk1 0.07 0 Monk2 0.351 0.01 Monk3 0.141 0 Vote 428.19 1.923 Mushroom 4911.32 27.91 SoybeanL >1000000 247805 Insurance 468656 2435.66 2868 W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 about 40 min, but the traditional rough set approach needs much more computation time. 6. Conclusion and future work In this paper, we have proposed a bit-based feature selection approach to discover optimal feature sets for the given table(dataset). In this approach, the feature val- ues are first encoded into bitmap indices for searching the optimal solutions efficiently. Also, the corresponding indexing and selecting algorithms are described in details for implementing the proposed approach. Experimental results on different data sets have also shown the efficiency and accuracy of the proposed approach. The traditional rough set approach has two very time- consuming parts, combination of features and comparison of upper/lower approximations. In this paper, we use the single-time-clock bitwise operations to shorten the compu- tation time of the comparison part. Moreover, the work- load in the combination part is highly reduced since the new levels of combination can be generated via the pervi- ous ones. The bitwise operations are also used to speed up the combination generation. The proposed feature selection approach also adopts appropriate meta-data structures to take advantages of the computational power of the bitwise operations. The feature selection problem is generally an NP-com- plete problem. Although, the proposed approach can pro- cess a larger amount of features than the traditional rough set approach, it still becomes unmanageable especially when the number of features is huge or when the number of possible values of features is large. In the future, we will continuously investigate and design efficient heuristic approaches to manage huge amounts of features and pos- sible values. We will also attempt to integrate different fea- ture selection approaches to automatically select an appropriate one for optimal or near-optimal solutions according to the characteristics of given data sets. References Almullim, H. et al. (1991). Learning with many irrelevant features. In Proceedings of ninth national conference on artificial intelligent, pp. 547–552. Brassard, G. et al. (1996). Fundamentals of Algorithm. New Jersey: Prentice Hall. Chen, W. C., Tseng, S. S., Chen, J. H., & Jiang, M. F. (2000). A framework of feature selection for the case-based reasoning. In Proceeding of IEEE international conference on systems, man, and cybernetics, CD-ROM. Chen, W. C., Tseng, S. S., Chang, L. P., & Jiang, M. F. (2001). A similarity indexing Method for the data warehousing – bit-wise indexing method. In Lecture notes in artificial intelligent (Vol. 2035), pp. 525–537. Chen, W. C., Tseng, S. S., Chang, L. P., & Hong, T. P. (2002). A parallelized indexing method for large-scale case-based reasoning. Expert System with Applications, 23(2), 95–102. Chen, W. C., Yang, M. C., & Tseng, S. S. (2002). A high-speed feature selection method for large dimensional data set. In Proceeding of international computer symposium, CD-ROM. Chen, W. C., Yang, M. C., & Tseng, S. S. (2003). The bitmap-based feature selection method. In 18th ACM symposium on applied computing (SAC), data mining track, CD-ROM. Choubey, S. K. et al. (1998). On feature selection and effective classifiers. Journal of ASIS, 49(5), 423–434. Doak, J. (1992). An evaluation of feature selection methods and their application to computer security, Technical Report, University of California. Gonzalez, A., & Perez, R. (2001). Selection of relevant features in a fuzzy genetic learning. IEEE Transaction on SMC-Part B, 31(3), 417–425. Huang, C. C., & Tseng, B. (2004). Rough set approach to case-based reasoning. Expert Systems with Applications, 26(3), 369–385. April. W.-C. Chen et al. / Expert Systems with Applications 34 (2008) 2858–2869 2869 John, G. H. et al. (1994). Irrelevant feature and the subset selection problem. In Proceedings of 11th international conference on machine learning, pp. 121–129. Last, M., Kandel, A., & Maimon, O. (2001). Information theoretic algorithm for feature selection. Pattern Recognition Letter, 22, 799–811. Lee, H. M., Chen, C. M., Chen, J. M., & Jou, Y. L. (1997). An efficient fuzzy classifier with feature selection based on fuzzy entropy. IEEE Transaction on SMC-Part B, 27(2), 426–432. Liu, H. et al. (1996). A probabilistic approach to feature selection – a filter solution. In Proceedings of 13th international conference on machine learning, pp. 319–327. Liu, H., & Setiono, R. (1998). Incremental feature selection. Applied Intelligence, 9, 217–230. Miao, D. Q. et al. (1999). A heuristic algorithm of reduction for knowledge. Journal of Computer Research and Development, 36(6), 681–684. O’Neil, P., & Quass, D. (1997). Improved query performance with variant indexes. ACM SIGMOD Record, 26(2), 38–49. Pawlak, Z. (1982). Rough set. International Journal of Computer and Information Sciences, 341–356. Pawlak, Z. (1991). Rough Sets. Theoretical Aspects of Reasoning about Data. Boston: Kluwer Academic Publishers. Quinlan, J. (1986). Introduction of decision trees. Machine Learning, 1(1), 81–106. Raymer, M. L., Punch, W. F., Goodman, E. D., & Kuhn, L. A. (2000). Dimensionality reduction using genetic algorithm. IEEE Transaction on Evolutionary Computation, 4(2), 164–171. Schlimmer, J. C. et al. (1993). Efficiently inducing determinations: A complete and systematic search algorithm that uses optimal pruning. In Proceedings of 10th international conference on machine learning, pp. 284–290. Skowron, A., & Rauszer, C. (1992). The discernibility matrices and functions in information systems. Intelligent Decision Support, 331–362. Tseng, B., Jothishankar, M. C., & Wu, T. (2004). Quality control problem in printed circuit board manufacturing – An extended rough set theory approach. Journal of Manufacturing Systems, 23(1), 56–72. Wu, F. B. et al. (1999). Control and Decision, 14(3), 206–211. Wu, M. C., & Alejandro, P. B. (1998). Encoded bitmap indexing for data warehouses. In Proceedings of IEEE data engineering, pp. 220– 230. Yang, Y., & Chiam, T. C. (2000). Rule discovery based on rough set theory. In Proceedings of the third international conference on FUSION (Vol. 1), pp. TuC4_11–TuC4_16. Yu, H. et al. (2001). Rough set based knowledge reduction algorithms. Computer Science, 28(5), 31–34. Zhong, N., Dong, J., & Ohsuga, S. (2001). Using rough sets with heuristics for feature selection. Journal of Intelligent Systems, 16, 199–214. An efficient bit-based feature selection method Introduction Review of feature selection and rough sets The proposed bitmap-based feature selection method Problem definitions Bitmap indexing phase Feature selection phase Creating cleansing tree Finding appropriate spanned order Cleansing feature matrix Some definitions on feature sets Selecting a feature set Time and space complexity analysis Experiments Conclusion and future work References 
work_zn7cbbam6zdfhfdeppujpl6f7u	----	Bozorgirad-Logendran.final version to Elsevier.ESWA_7542.Feb1212 Sequence-dependent group scheduling problem on unrelated-parallel machines Mir Abbas Bozorgirad and Rasaratnam Logendran* School of Mechanical, Industrial, and Manufacturing Engineering Oregon State University Corvallis, OR 97331-6001, USA. Abstract In this research we address a sequence-dependent group scheduling problem on a set of unrelated-parallel machines where the run time of each job differs on different machines. To benefit both producer and customers we attempt to minimize a linear combination of total weighted completion time and total weighted tardiness. Since the problem is shown to be NP-hard, meta-heuristic algorithms based on tabu search are developed to find the optimal/near optimal solution. For some small size yet complex problems, the results from these algorithms are compared to the optimal solutions found by CPLEX. The result obtained in all of these problems is that the tabu search algorithms could find solutions at least as good as CPLEX but in drastically shorter computational time, thus signifying the high degree of efficiency and efficacy attained by the former. Keywords Group Scheduling; Unrelated-parallel machines; Bi-criteria; Sequence-dependent setup time; Mixed- integer linear programming; Tabu search. Mir Abbas Bozorgirad, Email: bozorgim@onid.orst.edu * Corresponding Author: Rasaratnam Logendran, Tel: 1-541-737-5239, Fax: 1-541-737-2600, E-mail: Logen.Logendran@oregonstate.edu 1. Introduction Scheduling problems were first considered in the mid-1950s, and since then, so many papers that deal with different aspects of scheduling have been published. A comprehensive survey on these problems has been reported by Allahverdi et al. (2008). The research reported in this paper considers three important properties of scheduling problems which are group scheduling, unrelated-parallel machines, and bi-criteria objective function. Group scheduling follows from mainly applying the cellular manufacturing concept in which disaggregated manufacturing cells are formed to completely process different families of parts, and is usually performed in two levels: Outside and Inside. The outside level finds the desirable sequence of part families, i.e. groups; and the inside level finds the desirable sequence of jobs in each group (Logendran et al. (1995)). Group scheduling problems come with or without Group Technology Assumptions (GTA). GTA forces all jobs of a group to be processed contiguously, but scheduling without GTA allows the groups to be split into subgroups. Scheduling without GTA uses batch scheduling setup time, while that with GTA uses group scheduling setup time and is studied either as sequence-independent or sequence-dependent problem (Logendran et al. (2006a)). Sequence-independent setup time is when the required setup time for switching from one group to another, does not depend on the groups and their orders; otherwise the problem is a sequence-dependent problem. On the other hand, parallel machine scheduling problems are mostly divided into three categories: Identical machines, ��, where all machines have the same run time for each job; uniform machines, ��, where machines have different speeds, but each machine has a consistent rate; and unrelated machines, ��, where each machine can work in different rates and has a different run time for each job (Allahverdi et al. (2008); Lin et al. (2011)). The third important feature of this paper is to consider a bi-criteria objective function, as in many industries it is necessary to consider optimizing more than one criterion. This research tries to optimize the total weighted completion times, which favors the supplier’s interests, and at the same time tries to optimize the total weighted tardiness, which favors the customers’ interests. Therefore, the problem being researched in this paper can be classified as �|��� ,� , � | ∑ � � �� � with respect to the three field notation of Graham et al. (1979). The rest of the paper is organized as follows. Section 2 reviews the literature, which mostly deals with the problem of interest. The problem is well stated in section 3, together with a mixed-integer linear programming model developed for this problem. Section 4 comprehensively describes the proposed search algorithm. Section 5 describes the data generation mechanisms, which have been used to generate the sample problems, and section 6 illustrates the search algorithm with the help of a sample problem. Sections 7 and 8 represent the experimental results and the comparison to the optimal solutions, respectively. Finally in section 9, a concise conclusion about the research is presented. 2. Related Work Group scheduling is used in a wide variety of scheduling problems (Logendran and Nudtasomboon (1991); Logendran et al. (1995); Gelogullari and Logendran(2010); Logendran et al. (2005); Logendran et al. (2006b); Salmasi et al.(2010); Li et al. (2011); Xingong and Guangle (2010); Kuo and Yang (2006); and Yang and Chand (2008)). Li et al. (2011) tried to minimize an objective function which includes earliness, tardiness, due date assignment and flow time costs for a sequence-independent group scheduling problem on a single-machine. Logendran et al. (2006a) developed three search algorithms, based on tabu search, in order to minimize the total completion time for a two-machine sequence-dependent group scheduling problem. Later, Salmasi et al. (2011) extended their work to minimize the makespan of a sequence-dependent flow shop group scheduling problem. They developed a hybrid ant colony optimization algorithm (HACO) to solve the problem. Salmasi et al. (2010) developed a mathematical programming model to minimize the total flow time of a sequence-dependent flow shop group scheduling problem (�� |����, ���� , ����| ∑ � ). Gelogullari and Logendran (2010) reported the only analytical research, which considered a carryover sequence-dependent flow shop group scheduling problem in the assembly of printed circuit boards in electronic manufacturing. Logendran et al. (2005) developed heuristics to solve the group scheduling problem with sequence-independent setups in flexible flow shops. Later, Logendran et al. (2006b) studied the group scheduling problem with sequence-dependent setups and developed search algorithms based on tabu search to efficiently solve industry-size problems. Shahvari et al. (2011) extended the research by Logendran et al. (2006b), and proposed a mathematical model for the same problem together with six metasearch algorithms to solve it heuristically. With respect to the unrelated-parallel machine scheduling problems, Lin et al. (2006) classified these problems based on various correlation structures. Liaw et al. (2003), developed a Branch- and-Bound (B&B) algorithm to minimize the total weighted tardiness of a set of independent jobs on a set of unrelated-parallel machines. Logendran et al. (2007) developed different algorithms based on tabu search to minimize the total weighted tardiness of a sequence- dependent job scheduling problem on unrelated-parallel machines with consideration of dynamic job release and machine availability times. In another work, Logendran and Subur (2004) studied a different problem on unrelated-parallel machine scheduling with job splitting. Rocha et al. (2008) also developed a B&B algorithm to minimize a linear combination of makespan and total weighted tardiness of a sequence-dependent job scheduling problem on unrelated-parallel machines. Fanjul-Peyro and Ruiz (2010) proposed a set of iterative greedy local search based meta-heuristics to minimize the makespan of a set of independent jobs on unrelated-parallel machines. Gairig et al. (2007) developed a 2-approximaiton algorithm to minimize the makespan of a set of independent jobs on unrelated-parallel machines. Regarding the objective function, Huo et al. (2007) studied a single-machine problem to minimize the number of tardy jobs and maximize the weighted tardiness. Eren and Guner (2008) considered a two-machine flow shop scheduling problem to minimize the total weighted completion time together with the makespan of the sequence. Mohri et al. (1999) studied the minimization of maximum completion time and maximum lateness on three identical-parallel machines. Mehravaran and Logendran (2011) considered an unrelated-parallel machine job scheduling problem to jointly minimize the work- in-process inventory and maximize the customer service level in a supply chain. Lu and Logendran (2011) tried to minimize a linear combination of total weighted completion time and total weighted tardiness in a sequence-dependent flow shop group scheduling problem. Besides the above mentioned papers, there are a few others that have investigated the unrelated- parallel machine scheduling problems along with batch (family) setup times (Chen and Wu (2006); Kim et al. (2002); Kim et al. (2003); Arnaout et al. (2006); Akkiraju (2001); Jeong et al (2001)). We emphasize that these investigations focus on job scheduling, and not group scheduling as reported in this paper. Therefore, they do not follow the GTA, which is indeed the case in many real world problems. To the best of our knowledge, the problem of group scheduling together with a bi-criteria objective function on an unrelated-parallel machines environment has not been studied so far. Since this problem is well motivated by industry applications, we investigate it comprehensively in this paper. Also, the dynamic release of jobs and dynamic availabilities of machines have been considered for accurately modeling real world situations. 3. Problem statement Consider the problem of scheduling � different jobs on � unrelated-parallel machines. All jobs are clustered in � different groups where each group contains �� jobs and ∑ ��!�"# $ �. Each job needs to be processed only by one machine, and all jobs of a group need to be processed on the same machine (GTA assumption). Different groups can be processed on different machines, but not all machines are capable of processing all jobs. Realistically, a group cannot be processed on a machine if there is at least one job in the group which cannot be processed on that machine. The job release times are considered to be dynamic, that means not all the jobs are ready to be processed at time zero. The machine availability times are also considered to be dynamic, i.e. not all the machines are available at the beginning of the current planning horizon or time zero. The objective of the problem is to simultaneously optimize both the total weighted completion times and total weighted tardiness of jobs, thus emphasizing a bi-criteria focus. These two criteria are merged with the help of % and &, where % reflects the importance or weight of producer and & reflects the importance or weight of customers. As these weights are assumed to be normalized, % and &should add to 1. The parameters, decision variables, and the model in itself are described below. 3.1. Parameters � Number of groups '�() Maximum number of jobs in each group; Despite the fact that each group can be composed of different number of jobs, in order to simplify the development of the model, the same number of jobs ('�() ) is considered for all groups. If one group contains less than this maximum number of jobs, some dummy jobs (with processing time of zero) will be considered to make up for the difference in that group. '� Number of jobs in group * � Number of machines +� , - 1/ ��� +*�0 1� 21' 2 *� ��1�� * *� 21' 2 *� ��1�� * 34� '0 ��130��05 1� �436*�0 6 '7 �436*�0 6 2/ 9 *� 21' 2 *� ��1�� * 34��1+ '0 ��130��05 '7 �436*�0 6 : * $ 1, 2, … , �; 2 $ 1, 2, … , '�() ; 6 $ 1, 2, … , � ���, The setup time for group * on machine 6 if group � is the preceding group � $ 0, 1, 2, … , �; * $ 1, 2, … , � >* ? �/; 6 $ 1, 2, … , � (� $ 0 refers to the reference group or the last group in the previous planning horizon to be processed on the same machine.) 5� Due date of job 2 in group *; * $ 1, 2, … , �; 2 $ 1, 2, … , '�() �� Release time of job 2 in group *; * $ 1, 2, … , �; 2 $ 1, 2, … , '�() 4, Availability time of machine 6; 6 $ 1, 2, … , � �� The weight of job 2 in group * regarding the objective function; * $ 1,2, … , �; 2 $ 1,2, … , '�() % The weight of jobs regarding the producer & The weight of jobs regarding the customers 3.2. Decision Variables @� The completion time of job 2 of group *; * $ 1, 2, … , �; 2 $ 1, 2, … , '�() A� B C 1 *� 21' D *� ��130��05 4�+0� 21' 2 *� ��1�� *0 E+60��*�0 : �"#,F,…,! ,B"#,F,…,�GHI; JB K�, C1 *� ��1�� * *� 4��*��05 +1 �436*�0 60 E+60��*�0 : *$1,2,…,� 6$1,2,…,� �� The completion time of ��1�� * * $ 0, 1, 2, … , � L�� C1 *� ��1�� * *� ��130��05 4�+0� ��1�� � 0 E+60��*�0 : *, � $ 1, 2, … , �; � M * �N� O@� P 5� *� @� P 5� Q 00 E+60��*�0 : �"#,F,…,! "#,F,…,�GHI (�N� represents the tardiness of job 2 in group *) 3.3. Mixed-Integer Linear Programming Model (MILP) 9*� R $ % ∑ ∑ �� @� �GHI "#!�"# � & ∑ ∑ �� �N� �GHI "#!�"# (1) Subject to: @� � 9S1 P L�� T � 9>1 P K�, / � 9S1 P K�, T U �� � ���, � +� , (2) *, � $ 1, 2, … , �; � M *; 2 $ 1, 2, … , '�() ; 6 $ 1, 2, … , �; 9: 4 �4��0 ���'0� @� � 9>L�� / � 9>1 P K�, / � 9S1 P K�, T U �� � ���, � +� , (3) *, � $ 1, 2, … , �; � M *; 2 $ 1, 2, … , '; 6 $ 1, 2, … , �; 9: 4 �4��0 ���'0� @� U ∑ >4, � �W�, � +� , /K�,�,"# ; * $ 1, 2, … , �; 2 $ 1, 2, … , '�() (4) @� U ∑ >�� � +� , /K�,�,"# ; * $ 1, 2, … , �; 2 $ 1, 2, … , '�() (5) @� P @�B � 9>A� B / U ∑ +� , >K�, /�,"# (6) * $ 1, 2, … , �; 2, D $ 1, 2, … , '�() ; 2 M D; @�B P @� � 9>1 P A� B� / U ∑ +�B, >K�, /�,"# (7) * $ 1, 2, … , �; 2, D $ 1, 2, … , '�() ; 2 M D ∑ K�, $ 1�,"# ; * $ 1, 2, … , � (8) �� U @� ; * $ 1, 2, … , �; 2 $ 1, 2, … , '�() (9) �N� U @� P 5� ; * $ 1, 2, … , �; 2 $ 1, 2, … , '�() (10) @� , �N� U 0; K� X Y0,1Z; L�� X Y0,1Z >� M */; A� B X Y0,1Z, (j<q) *, �, � $ 1, 2, … , �; 2, D $ 1, 2, … , '�() ; 6 $ 1, 2, … , � The objective function (1) minimizes the summation of weighted completion times and weighted tardiness. Constraints (2) and (3) are incorporated to find the sequence of groups. These constraints restrict the completion time of each job in each group to be greater than the completion time of the previous group plus the sequence-dependent machine setup time, and the required runtime for processing the job. These constraints are active if and only if both groups are processed on the same machine. Simultaneously these constraints assign values to L�� variables, which determine the sequence of groups. Constraint (4) has been embedded to ensure that the completion time of each job in each group on any machine is greater than the machine availability time, plus setup time of that group on that machine, and the runtime of the job. This constraint considers the reference setup time, which is the setup time of a group on a machine when that group is the first group being processed, following the reference group, on that machine. All the other setup times have been considered in constraints (2) and (3). Constraint (5) ensures that the completion time of a job is greater than the job release time plus its runtime. Constraints (6) and (7) help to find the sequence of different jobs in each group. They ensure that the difference between the completion times of any two jobs in a group is greater than the processing time of the succeeding job. Constraint (8) restricts each group to be processed only on one machine. Constraint (9) computes the completion time of each group, which is greater than the completion times of all of its jobs. And finally constraint (10) is incorporated to find the tardiness of each job, which should be greater than or equal to both the completion time minus due date and zero. 4. Search Algorithm It has been shown that a single-machine problem with sequence-dependent setup times and a bi- criteria objective function is amongst the strongly NP-hard problems (Eren and Guner (2006)). Since the problem investigated in this paper can be easily reduced to the above mentioned problem, it can be concluded that it too is strongly NP-hard. Therefore, there is no algorithm to find the optimal solution of this problem in polynomial time. This is the main characteristic of most complex scheduling problems, which made us investigate into developing a meta-heuristic algorithm for solving the research problem. Although exact algorithms such as branch-and- bound can be used to find the optimal solution for small size problems, they are computationally very inefficient for optimally solving large size (or industry size) problems or even identifying a good feasible solution. Tabu search, introduced by Glover (1986), has been shown to be remarkably effective in solving combinatorial optimization problems (Glover and Laguna (1997); Logendran and Sonthinen (1997)). It is an iterative search which tries to overcome the limitations of local optimality. TS starts with an initial solution (IS), and moves through the neighborhood of this solution to find the best neighbor; then repeats this process to find better solutions. In the following, an IS finding mechanism is described, and then a detailed explanation of the tabu search algorithm, which best suits the problem reported in this paper is presented. 4.1. Initial Solution Tabu search needs an IS to trigger the search. This IS can be any feasible solution, but Logendran and Subur (2004) showed that the quality of the final solution is sensitive to the quality of the IS. Therefore, two different mechanisms are developed here in order to identify effective initial solutions. Since the objective function of the problem is bi-criteria, there are two different aspects in generating an effective sequence: producer’s interest and customers’ interest. From producer’s point of view, the total weighted completion time of jobs should be minimized; and from customers’ perspective, the total weighted tardiness of jobs needs to be minimized. In a single- machine job scheduling problem with setup and run times combined to evaluate a processing time, the first objective can be optimally attained by WSPT (weighted shortest processing time) rule, and an effective solution to the second objective can be attained by using WEDD (weighted earliest due date) order. Therefore, these two rules have been used as effective heuristics to identify the IS for the problem of this paper. In order to deal with the bi-criteria objective function, two different sequences are generated: producer’s sequence (PS) and customers’ sequence (CS). These sequences are then merged to find a final sequence as the IS of the search, similar to the approach used by Mehravaran and Logendran (2011). This merging takes place by normalizing the positional values of PS and CS. In other words, the final sequence (as the IS) will be defined as %. �� � &. �� where % � & $ 1. It is worth noting that all these sequencing will have to be performed on two levels: the group level which finds the sequence of groups, and the job level which finds the sequence of jobs within each group. In the first proposed IS finding mechanism, the average run time of each group has been used as an adjustment in finding the groups’ order of CS; but in the second mechanism, job release time has been used as an adjustment in both levels (group and job) of CS and PS. In both mechanisms, the group sequence will be determined first, and then with respect to this sequence, the job sequence in each group will be found. These mechanisms are described below: Initial Solution 1 (IS1): The PS in group level for each machine is obtained from the following rule: +\#, � �*� _�#, M +\F, � �*� _�F, M ^ M +\�, � �*� _��,; where +\�, $ ∑ _`abc`ade�` and �*� _��, $ �*��Y���,Z. But it follows the WSPT rule for job level: _`ebfeg M _`gbf`g M ^ M _`abf`a . Also, the CS in group level follows: 5\# � +\#, M 5\F � +\F, M ^ M 5\� � +\�,; where 5\� $ ∑ `ac`ade�` ; But it follows the WEDD rule in job level: `efeg M `gf`g M ^ M `af`a. Initial Solution 2 (IS2): The PS for each machine is obtained from the following rule in group level:+\#, � �*� _�#, � �\# M +\F, � �*� _�F, � �\F M ^ M +\�, � �*� _��, � �\� ; where �\� $ ∑ h`ac`ade�` . And the following rule is applied to the job level: _`ebih`ef`e M _`gbih`gf`g M ^ M _`ab ih`af`a . The CS in group level is obtained from the following rule: 5\# � �\# M 5\F � �\F M ^ M 5\� � �\� , and the following rule is applied to the job level: `eih`efeg M `gih`gf`g M ^ M `aih`af`a . In group levels of both IS1 and IS2, the group sequence will be determined for each machine separately. Then, the groups will be assigned according to these sequences to the earliest available machine, and ties are broken in favor of the smallest machine index. 4.2. Algorithmic structure The problem considered in this paper involves two levels of search: outside search and inside search. The former determines the order of groups on each machine, while the latter determines the order of jobs in each group for any given order of groups obtained from the outside search. Both of these levels are based on tabu search and the search process moves back and forth between these two levels as the search progresses. Step 1: Find an initial solution, say j�. Name the group order as Ej�. Ej� is considered as the first seed for the outside tabu search. Step 2: Generate the neighborhood of the outside seed by perturbing on groups, yet one move at a time. Two different types of moves are considered for this problem: exchange moves and insert moves. The exchange move simply exchanges the groups in two different positions on two different machines or on the same machine. The insert move inserts a group in another position, either on the current machine or on another machine. Step 3 (inside tabu search): For all the solutions generated in the previous step, the inside tabu search needs to be performed in order to find the best objective function value of the related group order. For ease of understanding, this has been explained for Ej�. Similar to the outside level (steps 1 and 2), an initial order of jobs within each group is needed in order to trigger the inside tabu search. The related job order of Ej� is considered as the inside IS and is referred to as jj�. This solution will be considered as the first seed of inside tabu search. Step 3.1: Generate the neighborhood of jj� and evaluate their objective function with respect to the associated group order. The neighborhood of IIS consists of all of its neighboring solutions. The same as outside level, the moves consist of exchange moves and insert moves. Step 3.2: Update the following parameters for the inside tabu search: (1) Inside tabu list (ITL) and Aspiration level (IAL): The tabu list stores the recent moves for a certain number of iterations, which is referred to as the tabu list size. In each iteration, a move cannot be performed if it is still in the tabu list. This condition holds true unless a move leads to a solution better than the aspiration level. Aspiration level shows the best solution found so far. Therefore, if any move leads to a solution better than the aspiration level, disregarding it being tabu or not, the move will be made and the aspiration level will be updated to the new value. (2) Inside candidate list (ICL): The best neighbor solution of a seed, in any iteration, is called candidate solution, and will be kept in a list named inside candidate list (ICL). The candidate solutions should not be repeated during the search. Therefore, if a neighbor solution is already stored in the ICL, it will not be considered again as a candidate. The candidate solution will be considered as the seed for the next iteration to repeat the perturbation. The IS is always the first entry into the ICL. If the next candidate has an objective function value better than the previous one, it will be assigned a star (*), which means this candidate has the potential of becoming a local optimum. (3) Inside index list (IIL): Any local optimum found during the inside search will be kept in a list named IIL. To find the inside local optima, a candidate solution needs to be checked when it is admitted into the ICL. If the new entry into the ICL has an objective function value worse than the latest entry into this list, and if the latest entry has been already assigned a star (*), it will be assigned another star and will be considered as a local optimum. This local optimum (with two stars **) will be entered into the IIL. Similar to the ICL, the IS is always the first entry into the IIL. Unlike the ICL, IIL should have a predefined size in order to control the execution time of the search, which is referred to as Inside Index List Size (IILS). (4) Inside number of iterations without improvement (INIWOI): In any iteration if the new entry into the ICL is not better than the latest entry into this list, increase the number of iteration without improvement for inside search (INIWOI) by one; otherwise set this number to zero. Similar to the IILS, the maximum number of iterations without improvement for inside search (MINIWOI) is also limited (and predefined), in order to control the execution time of the search. Step 3.3: IILS together with the INIWOI act as the stopping criteria for the inside search. If these two variables reach their predefined values, the search is stopped; otherwise the latest entry to the candidate list will be considered as the next seed and the search will be directed to step 3.1. Step 4: Repeat step 3 (inside tabu search) for all the solutions generated in step 2. Find the solution which has the best objective function value. Step 5: Update the following parameters the same as what has been described for the inside search: Outside Tabu List (OTL), Outside Aspiration Level (OAL), Outside Candidate List (OCL), Outside Index List (OIL), and Outside Number of Iterations without Improvement (ONIWOI). Step 6: The same as inside search, the stopping criteria need to be checked in this step. Therefore if either outside index list size (OILS) or ONIWOI reach their predefined values, the search will be stopped and the best solution found so far will be the best solution found by the tabu search; otherwise the latest entry into the outside candidate list will be considered as the next seed and the search goes to step 2. 4.3. Long Term Memory (LTM) What has been described in the previous subsection is known as the short-term memory functionality of tabu search. Besides this functionality, tabu search also uses the long-term memory (LTM) in order to intensify the search (LTM-Max) in the regions where there is potential of identifying better solutions, or to diversify the search (LTM-Min) in the regions that have not been visited frequently in the past iterations. LTM functionality can be performed on both outside and inside tabu searches. In order to do so, a frequency matrix is needed to keep track of the number of placement of each group on any position of each machine (outside level) or the number of placement of each job on any position of each group (inside level). In each iteration of the inside tabu search, after finding the next candidate, the related cells of the inside frequency matrix need to be updated. After the tabu search with short-term memory (either inside or outside) comes to an end, the search can be restarted using another initial solution which has been extracted from the frequency matrix. In order to find the restart point (the next initial solution to restart the search), LTM-Max selects the maximum number of the frequency matrix and places the job (inside level) or group (outside level) in the position where this max number refers to, then selects the second maximum number and continues this procedure until all the jobs or groups are placed in their positions. LTM-Min also works the same way, but in order to diversify the search to the regions that have not been visited frequently before, it selects the minimum numbers of the frequency matrix. It should be noted that the infeasible positions, as a result of some machines not capable of processing some jobs, should not be considered while selecting the minimum numbers in LTM-Min. Laguna et al. (1993), used four restarts for their single machine scheduling problem, but preliminary investigations of test problems, revealed that 3 restarts for the outside search and 2 restarts for the inside search best fit the problem addressed in this paper. These numbers of restarts have also been successfully used previously by Gelogullari and Logendran (2010). 5. Data Generation In this section some data generation mechanisms have been designed for different parameters of the model. When working with group scheduling problem, the workload of manufacturing cells plays an important role in determining the degree of complexity associated with the problem. Therefore, three different categories have been defined for the sample problems as: loosely- loaded, moderately-loaded, and tightly-loaded. Unlike in previous research, we define all anew the workload of cells as the ratio of number of groups to the number machines. The number of groups and number of machines are generated uniformly in [5,16] and [3,6], respectively. Then if this ratio is between 5/6 to 2k, the problem is regarded as loosely-loaded; if the ratio is between 2 to 3k, the problem is moderately-loaded, and finally if it is between 3 to 16/3 the problem is among the tightly-loaded problems. The number of jobs in each group is generated uniformly in U[2,6]. Job weights are also generated uniformly in [1,4]. The job release times and the machine availability times are generated from an exponential distribution with mean of 20 minutes. The machine capability is divided into three levels: least, medium and most capable (Logendran and Subur (2004); Pandya and Logendran (2010); Mehravaran and Logendran (2011)), which are eligible to process 50%, 70% and 85% of all jobs, respectively. In order to assign these capabilities to machines, if the number of machines is 3, then there will be one machine from each capability. If the number of machines is more than 3, say �, then >� P 3/ uniform random numbers in [0,1] are generated. The numbers with a value less than 1/3, more than 2/3, and between 1/3 and 2/3 will be counted separately. The largest count will be the extra numbers of low capability machines, the smallest count will be the extra number of most capable machines, and the remaining count will be the extra number of medium capable machines. In order to generate the run times, three uniform random numbers in [1,10] are generated, %� m * $ 1,2,3 for each machine type. The largest value is assigned to the least capable machines, the smallest value is assigned to the most capable machines, and the remaining one is assigned to the medium capable machines. The machine runtimes are then generated from a uniform distribution [ %� � 1, %� � 20]. Schaller et al. (2000) argued that the complexity of a scheduling problem is likely to depend on the ratio of setup time to run time. Following their suggestion, the setup times are generated uniformly in [1,40], [1,100] and [1,200] which approximately lead to the ratio of 2:1, 5:1 and 10:1, respectively, for the mean of setup times to the mean of run times. Previous works (Pandya and Logendran (2010); Kim et al. (2002)) showed that defining proper due dates can positively affect the performance of the algorithms. They used two different factors in defining due dates: tardiness factor >n/, and due date range factor >�/. The tardiness factor (n) is used to create loose or tight due dates. It is defined as n $ 1 P 5\/��() , where 5\ is the average due date and ��() is the maximum completion time of all jobs. Large values of n indicate tight due dates and small values of this factor lead to loose due dates. In addition to this factor, the due date range factor >�/ controls the variability of due dates. The range factor >�/ is defined as >5�() P 5��p //��() , where 5�() is the maximum due date, among all the jobs, and 5��p is the minimum one. Therefore, different combinations of n and � provide different characteristics for randomly generated due dates. In this research, these factors have been set as n $ 0.5 and � $ 0.2 which provide medium and narrow range due dates. Then due dates are generated with probability of n from uniform distribution in [5\ P �5\, 5\], and with the probability of >1 P n/ from uniform distribution in [5\, 5\ � >��() P 5\/�] (Logendran and Subur (2004); Pandya and Logendran (2010); Kim, et al. (2002)). In order to estimate ��() for an unrelated- parallel machine scheduling problem, the following iterative equations have been developed: @�# $ ∑ rmaxY��#, ��� � v�\� Z � +�#, w/��,"# Where @�# is the estimated completion time of job * in group 1. +�#, is the run time of job 1 in group *. � is the total number of machines which are capable of processing the job. �\� is the average setup time of group *. v was introduced by Logendran et al. (2007) as an adjustment to the average setup time. They argued that the maximum completion time is sensitive to the setup time of jobs, and using the average setup time instead of the sequence-dependent setup times would not provide an accurate estimate of ��() . Therefore, v was defined as an adjustment. Furthermore, a coefficient of variation >�x/ for the sequence-dependent setup times was introduced as �x $ �/@\ , where � is the sample standard deviation and @\ is the average. They suggested a linear relationship between v and �x, which can be described by the interpolation of >�x, v/ $ >0.01, 0.9/ and >�x, v/ $ >1, 0.1/. ��� denotes when a machine is ready to start processing group *: ��� $ zmin C@�>p}/~ m � $ >* P 1/, >* P 2/, … , >* P �/ j� * Q � 4� E+60��*�0 : @�>p}/ is the estimated completion time of the last job in group �, and 4� is the availability time of machine *. Completion time of the other jobs in each group (not the first job) would be estimated as follows: @� $ max��� , @�> k#/� � ∑ +� ,�,"# /�. Finally, ��() can be estimated as the maximum of all @� ’s. 5.1. Tabu search parameter tuning Before starting the tabu search, its parameters, i.e. tabu list size (TLS), index list size (ILS), and number of iterations without improvement (NIWOI), for both inside and outside searches, need to be evaluated. Inappropriate values for these parameters can drastically worsen the objective function value, or the computation time. Thus, a procedure needs to be developed and implemented to determine appropriate values for these parameters, which is commonly referred to as parameter tuning in published literature. Running several experiments showed that the value of each of these parameters depends on the structure of the problem, i.e. number of machines, number of groups, and average number of jobs in each group. Table 1 shows the empirical formulae which are obtained with the help of DATAFIT 7.1.44 (1995) for each of the three categories of problems defined in the previous section. Table 1: Empirical formula for tabu search parameters 6. Example Problem For the purpose of better understanding, the application of the search algorithm is illustrated with an example problem. Tables 1 and 2 show the data related to this problem, which consists of three machines, five groups, and 3, 2, 2, 4 and 2 jobs in groups 1 to 5, respectively. This problem is amongst the loosely-loaded problems. Table 2 shows the machine availability times, job weights, job release times, job due dates and job processing times on each machine, where ∞ means that the machine is not capable of processing the job. Table 3 shows the sequence-dependent setup times where group zero indicates the reference group. Table 2: Summary data of the example problem Table 3: Sequence-dependent setup times for the example problem Loosely-Loaded Moderately-Loaded Tightly-Loaded OTLS Round(0.4g-0.62m+1.5 �\) Round(-0.58g+1.31m+1.48�\) Round(0.28g+0.09m-0.13�\) OILS Round(0.32g+0.14�\) Round(1.43g-0.91m-0.93�\) Round(0.2g+3.31m-1.76�\) ONIWOI Round(0.33g-0.03m-0.14�\) Round(0.49g-0.57m) Round(0.03g+0.9m-0.26�\) ITLS Round(0.22g-0.27m+0.74�\) Round(0.03g+0.19m+0.65�\) Round(-0.07g+0.98m+0.15�\) IILS 2 2 Round(0.02g+0.26m+0.23�\) INIWOI 1 1 Round(0.33m-0.02�\) g=number of groups; m=number of machines; �\ = the average number of jobs in each group Round means rounding to the closest integer value. Machine Machine Availability M1 M2 M3 13 54 9 Group Job Job Processing Time on Machine Job Weight Job Release Time Job Due Date 1 1 ∞ ∞ 39 3 19 128 1 2 ∞ ∞ 6 3 57 140 1 3 ∞ ∞ 3 2 44 128 2 1 42 26 34 3 39 139 2 2 19 45 27 2 39 133 3 1 ∞ ∞ 25 1 5 134 3 2 17 26 38 1 27 134 3 3 15 7 17 2 40 87 4 1 ∞ 40 6 1 63 134 4 2 44 9 14 2 7 135 4 3 ∞ 39 7 2 3 140 4 4 9 32 33 3 13 132 5 1 31 34 23 2 4 68 5 2 ∞ 34 22 2 0 132 Second Group Machine First Group 1 2 3 4 5 1 0 52 51 17 58 35 1 1 0 36 30 37 60 1 2 23 0 37 5 21 1 3 60 57 0 44 21 1 4 58 35 9 0 9 1 5 1 5 6 13 0 2 0 7 46 55 1 36 2 1 0 7 34 57 31 2 2 31 0 1 25 36 2 3 56 24 0 49 39 2 4 17 46 29 0 39 2 5 13 1 27 29 0 3 0 59 8 42 13 29 3 1 0 46 34 37 41 3 2 49 0 8 38 26 3 3 14 37 0 41 41 3 4 18 51 2 0 35 3 5 24 5 50 1 0 Step 1: Use IS2 in order to find an initial solution for this problem. The PS and CS for the group level are: PS: 2-5-4-3-1 on machine 1, 5-4-2-3-1 on machine 2, and 4-3-1-2-5 on machine 3. CS: 5-3-4-1-2 on all three machines. Merge PS and CS by using % and & to find a positional vector for each machine. For example, the positional vector for machine 1 is r4.6 2.6 3.2 3 1.6w where 4.6 (4.6=0.6*5+0.4*4) is the positional value of group 1 on machine 1. Therefore the final sequence of groups on machine 1 is 5-2-4-3-1. The final sequence for machine 2 and 3 are 5-4-3-2-1 and 4-3-1-5-2, respectively. Using these sequences, assign the groups to the first available machine one by one. As a result group 2 is assigned to machine 1, group 5 is assigned to machine 2, and groups 4, 3, and 1 are assigned to machine 3 in the mentioned order. This sequence is shown by Ej�={2|5|4-3-1} where the vertical lines separate the groups on different machines and the dashes separate the groups on a machine. With respect to this sequence (and assignment) of groups, the PS and CS for job level need to be determined which will lead to the following sequence of jobs: jj�={1-2- 3|1-2|2-3-1|4-2-1-3|2-1} where the vertical lines separate different groups and dashes separate different jobs within a group. The objective function associated with these sequences is 2816.4. Step 2: Generate the neighborhood of Ej� which are: �# ${2|5|3- 4- 1}, �F ${2|5|1- 3- 4}, �� ${2|5|4- 1- 3}, �� ${2|5|3- 1- 4}, �� ${2|5|1- 4- 3}, �� ${2|4|5- 3- 1}, �� ${|5- 2|4- 3- 1}, �� ${|5|2- 4- 3- 1}, �� ${|5|4- 2- 3- 1}, �#W ${|5|4- 3- 2- 1}, �## ${|5|4- 3-1-2}, �#F ${2||5- 4- 3- 1}, �#� ${2||4- 5 -3 -1}, �#� ${2||4- 3- 5- 1}, �#� ${2||4- 3- 1- 5}, �#� ${2|4- 5|3- 1}, �#� ${2|5- 4|3- 1}. Both types of move, i.e. exchange and insert, are used in generating this neighborhood. For example, �# is generated from exchanging the position 1 on machine 3 with position 2 on the same machine, which can be shown by E(m3,p1,m3,p2); and �#�, is an insertion of group 4 (position 1 on machine 3) in position 1 on machine 2, which can be shown by I(m3,p1,m2,p2). Step 3: Perform inside search for all of the above solutions in order to find their objective function value. In the interest of space, this step is performed only for Ej�. The initial solution for the inside search is jj�= {1-2-3|1-2|2-3-1|4-2-1-3|2-1}. Step 3.1: Generate the neighborhood of jj� and evaluate their objective function with respect to Ej�. The generated neighboring solutions are: �W#= {2-1-3|1-2|2-3-1|4-2-1-3|2-1}, �WF= {3-2- 1|1-2|2-3-1|4-2-1-3|2-1}, �W�= {1-3-2|1-2|2-3-1|4-2-1-3|2-1},�W�= {2-3-1|1-2|2-3-1|4-2-1-3|2-1}, �W�= {3-1-2|1-2|2-3-1|4-2-1-3|2-1}, �W�= {1-2-3|2-1|2-3-1|4-2-1-3|2-1}, �W�= {1-2-3|1-2|3-2-1|4- 2-1-3|2-1}, �W�= {1-2-3|1-2|1-3-2|4-2-1-3|2-1}, �W�= {1-2-3|1-2|2-1-3|4-2-1-3|2-1}, �W#W= {1-2- 3|1-2|3-1-2|4-2-1-3|2-1}, �W##= {1-2-3|1-2|1-2-3|4-2-1-3|2-1}, �W#F= {1-2-3|1-2|2-3-1|2-4-1-3|2- 1}, �W#�= {1-2-3|1-2|2-3-1|1-2-4-3|2-1}, �W#�= {1-2-3|1-2|2-3-1|3-2-1-4|2-1}, �W#�= {1-2-3|1- 2|2-3-1|4-1-2-3|2-1}, �W#�= {1-2-3|1-2|2-3-1|4-3-1-2|2-1}, �W#�= {1-2-3|1-2|2-3-1|4-2-3-1|2-1}, �W#�= {1-2-3|1-2|2-3-1|2-1-4-3|2-1}, �W#�= {1-2-3|1-2|2-3-1|2-1-3-4|2-1}, �WFW= {1-2-3|1-2|2-3- 1|4-1-3-2|2-1}, �WF#= {1-2-3|1-2|2-3-1|1-4-2-3|2-1},, �WFF= {1-2-3|1-2|2-3-1|3-4-2-1|2-1}, �WF�= {1-2-3|1-2|2-3-1|4-3-2-1|2-1}, �WF�= {1-2-3|1-2|2-3-1|4-2-1-3|1-2}. The minimum objective function value in this neighborhood is 2645.4 and is associated with �WF. Step 3.2: Update the parameters of inside tabu search: j�� = {E(g1,p1,p3) }, jL� = 2645.4, j�� = {{1-2-3|1-2|2-3-1|4-2-1-3|2-1}, {3-2-1|1-2|2-3-1|4-2-1-3|2-1}*}, jj� = {{1-2-3|1-2|2-3-1|4- 2-1-3|2-1}}, and INIWOI = 0. Step 3.3: The IILS is equal to 2 and the INIWOI = 1. As none of the stopping criteria have been reached yet, go to step 3.1. Step 4: Repeat step 3 for all the solutions generated in step 2. The best neighboring solution is �� $ {2|5|4, 1, 3} with an objective function value of 2318.4. Step 5: Update the parameters of outside tabu search; E�� = {E(m3,p2,m3,p3)}, EL� = 2318.4, E�� = {{2|5|4, 3, 1}, {2|5|4, 1, 3}*}, Ej� = {{2|5|4, 3, 1}}, and E�j�Ej = 0; Step 6: OILS = Round (0.32*5+0.14*2.8) = 2 and ONIWOI = Round (0.33*5-0.03*3-0.14*2.8) = 1. As none of the stopping criteria have been reached yet, go to step 2. This example continues to one more outside iteration and then stops (ONIWOI = 1). The final solution is as follows: Best order of groups found by outside tabu search = {2|5|4, 1, 3} Best order of jobs found by inside tabu search = {3, 2, 1|2, 1|3, 1, 2|3, 2, 4, 1|1, 2} Best objective function value found by tabu search = 2318.4 7. Experimental Results Based on what has been described in section 5 (Data Generation), several experiments have been conducted in order to compare the performance of different tabu searches (STM, LTM-Max, LTM-Min), and different initial solution finding mechanisms (IS1, IS2) which are the factors of interest in this research. According to the previous works (Gelogullari and Logendran (2010); Schaller et al. (2000)) multiple factors are speculated to have effects on the performance of the search algorithms or the IS finding mechanisms. These factors are: problem structure with three levels: loosely-loaded, moderately-loaded, and tightly-loaded, setup to run time ratio with three levels: ratio = 2, 5, and 10, due date tightness (n) with three levels: loose (n $ 0.2), medium (n=0.5), and tight (n $ 0.8), and finally scenario which is the combination of coefficients of bi- criteria objective function (%, &) and that too has three levels: (% $ 0.6, & $ 0.4), (% $ 0.5,& $ 0.5), (% $ 0.4, & $ 0.6). Therefore, there are six factors which have effects on the response variable (objective function value), and in order to find the effects of search algorithms and IS finding mechanisms, all the other four factors need to be blocked. The questions of interest in this research are: 1. Do any of the search algorithms outperform the others? 2. Do any one of the IS finding mechanisms outperform the other? 3. Does the performance of search algorithms or IS finding mechanisms differ in different levels of other factors? To have a better evaluation of the factors of interest, in a logical time, all levels of these factors need to be tested on the same example problem (experimental unit). Thus a split-plot design seems to be the best fit for this research as described in Montgomery (2009). The factors of interest are put in sub plot and the four other factors are placed in the whole plot. The statistical model for this design is: 7� ���p� $ � � n� � % � &� � v� � �� � >%&/ � � >%v/ � � >%�/ � � >&v/�� � >&�/��� >v�/�� � >%&v/ �� � >%&�/ �� � >&v�/��� � >%&v�/ ��� � �� ��� � �p� �� � >��/p� � >%�/ p � >&�/�p � >v�/�p � >��/�p � >%&�/ �p� >%v�/ �p � >%��/ �p � >&v�/��p � >&��/��p � >v��/��p � >%&v�/ ��p� >%&��/ ��p � >&v��/���p � >%&v��/ ���p � >%�/ � � >&�/�� � >v�/��� >��/�� � >%&�/ �� � >%v�/ �� � >%��/ �� � >&v�/��� � >&��/���� >v��/��� � >%&v�/ ��� � >%&��/ ��� � >&v��/���� � >%&v��/ ����� >%��/ p� � >&��/�p� � >v��/�p� � >���/�p� � >%&��/ �p�� >%v��/ �p� � >%���/ �p� � >&v��/��p� � >&���/��p� � >v���/��p�� >%&v��/ ��p� � >%&���/ ��p� � >&v���/���p� � >%&v���/ ���p�� �� ���p� *, 1 $ 1,2 ; and 2, �, �, �, � $ 1,2,3 Where � is the overall mean effect, n� is the replicate effect, % is 2th level of structure, &� represents the �th level of setup to run time ratio, v�is the �th level of due date tightness, ��represents the �th level of scenario, �p is the �th level of search algorithm, and ��represents the 1+6 level of IS finding mechanism. A total of 972 example problems have been tested and the analysis has been done on the natural logarithm transformation of the response variable, because the original values of response variable showed violations from constant variability assumption and normality assumption. But these violations were resolved completely by the natural logarithm transformation. R 2.13.0 (2011) was used to perform the analysis. The detailed ANOVA table is shown in Table 4, but the summary of statistical findings is presented here: There is convincing evidence that there is a nonzero difference between different algorithms (p- value < 0.00001). Since the search algorithm factor has more than two levels, Tukey test is needed to compare different levels of this factor. As a result of this test, generally LTM-Max outperforms LTM-Min, and LTM-Min outperforms STM. There is moderate evidence that IS1 works better than IS2 (p-value = 0.032). And there is moderate evidence of a non-zero difference between different levels of search algorithm per each level of IS finding mechanism (interaction effect of search algorithm and IS finding mechanism) (p-value = 0.029); Tukey test showed that for both IS1 and IS2, LTM-Max outperforms LTM- Min and LTM-Min outperforms STM, but when using IS2, superiority of LTM-Max is more noticeable than the other two. There is also convincing evidence of a non-zero difference between different levels of search algorithm per each level of setup to run time ratio (interaction effect of search algorithm and setup to run time ratio) (p- value < 0.00001). The Tukey test again showed that for all three ratios of setup to run time (2, 5 and 10), the LTM-Max works better than LTM-Min and LTM- Min works better than STM. But the superiority of LTM-Max is more evident than the other two when setup to run time ratio is 2. In addition to these effects, as it can be seen from Table 4, there are evidences for other high rank interactions (interactions of rank more than 2), but since the interpretation of these effects is really hard, they have not been mentioned in the above summary of statistical findings. Table 4: ANOVA table for split plot design Source (whole Plot) df Sum Sq Mean Sq F-Statistics P-Value Rep (or Blocks) 1 0.156 0.156 0.199846 0.656053 Str 2 250.066 125.033 160.1755 0 StoR 2 38.541 19.2705 24.68678 0 DD 2 39.361 19.6805 25.21202 0 Sc 2 6.702 3.351 4.292852 0.016945 Str:StoR 4 2.892 0.723 0.926211 0.453027 Str:DD 4 3.319 0.82975 1.062964 0.380392 StoR:DD 4 5.869 1.46725 1.879644 0.122071 Str:Sc 4 3.426 0.8565 1.097233 0.363729 StoR:Sc 4 4.158 1.0395 1.331668 0.265407 DD:Sc 4 7.028 1.757 2.250833 0.070847 Str:StoR:DD 8 5.26 0.6575 0.842301 0.568423 Str:StoR:Sc 8 5.189 0.648625 0.830931 0.578004 Str:DD:Sc 8 4.315 0.539375 0.690975 0.698327 StoR:DD:Sc 8 6.592 0.824 1.055598 0.402509 Str:StoR:DD:Sc 16 15.909 0.994313 1.27378 0.234678 Whole Plot Error 80 62.448 0.7806 Source (Sub Plot) df Sum Sq Mean Sq F-Statistics P-Value IS 1 0.013 0.013 4.65106 0.031622 Str:IS 2 0.003 0.0015 0.536661 0.585113 StoR:IS 2 0.001 0.0005 0.178887 0.836267 DD:IS 2 0.006 0.003 1.073322 0.342842 Sc:IS 2 0.007 0.0035 1.252208 0.286977 Str:StoR:IS 4 0.002 0.0005 0.178887 0.949256 Str:DD:IS 4 0.009 0.00225 0.804991 0.52249 StoR:DD:IS 4 0.015 0.00375 1.341652 0.25374 Str:Sc:IS 4 0.017 0.00425 1.520539 0.195313 StoR:Sc:IS 4 0.003 0.00075 0.26833 0.89829 DD:Sc:IS 4 0.001 0.00025 0.089443 0.985732 Str:StoR:DD:IS 8 0.028 0.0035 1.252208 0.267224 Str:StoR:Sc:IS 8 0.007 0.000875 0.313052 0.961021 Str:DD:Sc:IS 8 0.006 0.00075 0.26833 0.975824 StoR:DD:Sc:IS 8 0.023 0.002875 1.0286 0.413569 Str:StoR:DD:Sc:IS 16 0.033 0.002063 0.737909 0.755078 Alg 2 0.3 0.15 53.66608 0 IS:Alg 2 0.02 0.01 3.577739 0.028825 Str:Alg 4 0.014 0.0035 1.252208 0.288286 StoR:Alg 4 0.134 0.0335 11.98542 0 DD:Alg 4 0.006 0.0015 0.536661 0.708881 Sc:Alg 4 0.002 0.0005 0.178887 0.949256 Str:IS:Alg 4 0.004 0.001 0.357774 0.838598 StoR:IS:Alg 4 0.005 0.00125 0.447217 0.774434 Str:StoR:Alg 8 0.003 0.000375 0.134165 0.997692 DD:IS:Alg 4 0.003 0.00075 0.26833 0.89829 Str:DD:Alg 8 0.046 0.00575 2.0572 0.038927 StoR:DD:Alg 8 0.012 0.0015 0.536661 0.828878 Sc:IS:Alg 4 0.006 0.0015 0.536661 0.708881 Str:Sc:Alg 8 0.041 0.005125 1.833591 0.069279 StoR:Sc:Alg 8 0.011 0.001375 0.491939 0.862012 DD:Sc:Alg 8 0.018 0.00225 0.804991 0.598467 Str:StoR:IS:Alg 8 0.006 0.00075 0.26833 0.975824 Str:DD:IS:Alg 8 0.014 0.00175 0.626104 0.755985 StoR:DD:IS:Alg 8 0.006 0.00075 0.26833 0.975824 Str:StoR:DD:Alg 16 0.046 0.002875 1.0286 0.424806 Str:Sc:IS:Alg 8 0.005 0.000625 0.223609 0.986588 StoR:Sc:IS:Alg 8 0.004 0.0005 0.178887 0.993673 Str:StoR:Sc:Alg 16 0.118 0.007375 2.638582 0.000585 DD:Sc:IS:Alg 8 0.008 0.001 0.357774 0.942061 Str:DD:Sc:Alg 16 0.092 0.00575 2.0572 0.009453 StoR:DD:Sc:Alg 16 0.084 0.00525 1.878313 0.020855 Str:StoR:DD:IS:Alg 16 0.026 0.001625 0.581383 0.898078 Str:StoR:Sc:IS:Alg 16 0.005 0.000313 0.111804 0.999995 Str:DD:Sc:IS:Alg 16 0.016 0.001 0.357774 0.990283 StoR:DD:Sc:IS:Alg 16 0.012 0.00075 0.26833 0.998171 Str:StoR:DD:Sc:Alg 32 0.142 0.004438 1.587621 0.024329 Str:StoR:DD:Sc:IS:Alg 32 0.025 0.000781 0.279511 0.999975 Subplot Error 405 1.132 0.002795 Total 971 463.771 Rep=Replicate; Str=Structure; StoR=Setup to Runtime Ratio; DD= Due Date Sc=Scenario; IS=Initial Solution; Alg=Algorithm 8. Comparison to the optimal solution In order to assess the quality of the proposed search algorithms, 15 example problems have been generated and solved by both tabu search algorithms and CPLEX 12.2 (2009). The result of these runs is summarized in Table 5. All these example problems are amongst the loosely-loaded or moderately-loaded problems, and they have different work-loads (ratio of number of groups to the number of machines). As it can be seen from Table 5, CPLEX was able to find the optimal solution of the first 9 problems in less than half an hour. It solved problem 10, 11, and 12 optimally in around 1.5, 3 and 3.5 hours. But it took CPLEX more than 26 hours to solve problem 13 optimally, and even after more than 72 hours it was not able to find the optimal solutions for problems 14 and 15. For all these problems the proposed tabu searches were able to find a solution at least as well as what has been reported by CPLEX, but in less than 30 seconds. It may be worth noting that in all runs, the best solution found by all combinations of search algorithms and IS finding mechanisms, has been reported. All the other workloads which are not listed here have number of groups and number of machines greater than or equal to these examples. And intuitively the greater numbers of groups and machines with the same average number of jobs will result in more complicated problems with longer computational time. Since CPLEX could not find the optimal solution for problems 14 and 15 after more than 72 hours, clearly there is no need to test problems with other workloads as they would probably take more than 72 hours to be solved. Table 5: Comparison of the proposed tabu search and CPLEX Example Problem Structure Work load Number of groups Number of machines Total number of jobs Tabu Search CPLEX Obj Func Value Computational Time (sec) Obj Func Value Optimality Lower Bound Computational Time (sec) 1 Loose 1.00 5 5 17 3404 0.80005 3404 optimal 11.02 2 Loose 1.20 6 5 20 7000.8 0.54 7000.8 optimal 27.44 3 Loose 1.25 5 4 19 4449 0.36802 4449 optimal 48 4 Loose 0.83 5 6 19 6422.4 0.65604 6422.4 optimal 98.03 5 Moderate 2.00 6 3 20 3029 2.00011 3029 optimal 355.27 6 Loose 1.40 7 5 22 4874.4 1.67 4874.4 optimal 431.56 7 Loose 1.00 6 6 23 2834.4 3.01417 2834.4 optimal 878.56 8 Moderate 2.00 8 4 24 6684.8 3.5112 6684.8 optimal 1027.33 9 Loose 1.67 5 3 20 4945.8 0.69504 4945.8 optimal 1377.88 10 Moderate 2.33 7 3 23 11876.6 4.15824 11876.6 optimal 5903.08 11 Loose 1.75 7 4 25 8118 8.09546 8118 optimal 11053.4 12 Loose 1.50 6 4 21 6933.8 0.91605 6933.8 optimal 12559 13 Moderate 2.67 8 3 26 6607 20.5332 6607 optimal 94777.2 14 Loose 1.60 8 5 29 8327.2 8.01946 8385.4 feasible 7880.63 259714 15 Loose 1.17 7 6 32 7676.2 17.411 7696 feasible 6706.29 261768 9. Conclusions and future research This research addressed a group scheduling problem (with GTA) in an unrelated-parallel machines environment. A bi-criteria objective function is considered to optimize the weighted sum of total weighted completion time and total weighted tardiness which benefit the producer and customers, respectively. The setup time for switching between the groups are considered to be sequence-dependent, and to better resemble the real world situations, job release times and machine availability times are considered to be dynamic. A mixed-integer linear programming model has been developed to optimally solve this problem. Since this is an NP-hard problem, 3 different meta-heuristic algorithms based on tabu search, together with two different IS finding mechanisms to initiate the search, are developed to find the optimal/near optimal solution. Comprehensive methods have been described in order to generate sample problems, which truly represent the real world problems. Sample problems are divided into three categories with respect to the load of the workshop, i.e. the ratio of number of groups to the number of machines. The results of proposed tabu search are compared to the optimal solutions obtained from CPLEX for 15 example problems in loosely-loaded and moderately-loaded categories. Even though for some of these problems it took CPLEX several hours to find the optimal solution (in some examples it could not find the optimal solution even after 72 hours), the proposed algorithms found solutions at least as good as CPLEX in less than 30 seconds for all the example problems. This argues strongly in favor of the efficacy and efficiency of the proposed algorithms, but to better assess the performance of the algorithms even in medium and large problems, we recommend developing a lower bounding mechanism as a future area for research. A total of 972 experiments have been conducted to identify which algorithms outperform the others, and which IS finding mechanism works better than the other. A split-plot experimental design helped to truly block the effects of the other parameters of the problem. As a result, it turned out that in general LTM-Max outperforms the other two algorithms, LTM-Min outperforms STM, and IS1 is mostly superior to IS2. Acknowledgment: The research reported in this paper is funded in part by the National Science Foundation (USA) Grant no. CMMI-1029471. Their support is gratefully acknowledged. References Akkiraju, R. (2001). An agent-based approach for scheduling multiple machines. Applied Intelligence, 14, 135-144. Allahverdi, A., NG, C., Cheng, T., & Kovalyov, M. (2008). A survey of scheduling problems with setup times or costs. European Journal of Operational Research, 187, 985–1032. Arnaout, J., Rabadi, G., & Mun, H. (2006). A dynamic heuristic for stochastic unrelated parallel machine scheduling problem. International Journal of Operations Research, 3, 136-143. Chen, J., & Wu, T. (2006). Total tardiness minimization on unrelated parallel machine scheduling with auxiliary equiment constraints. Omega, 34, 81-89. DATAFIT, Version 7.1.44 (1995). Oakdale Engineering. Eren, T., & Guner, E. (2006). A bicriteria scheduling with sequence-dependent setup times. Applied Mathematics and Computations, 179, 378-385. Eren, T., & Guner, E. (2008). A bicriteria flowshop scheduling with a learning effect. Applied Mathematical Modeling, 32, 1719-1733. Fanjul-Peyro, L., & Ruiz, R. (2010). Iterated greedy local search methods for unrelated parallel machine scheduling. European Journal of Operational Research, 207, 55-69. Gairig, M., Monien, B., & Woclaw, A. (2007). A faster combinatorial approximaiton algorithm for scheduling unrelated parallel machines. Theoritical Computer Science, 380, 87-99. Gelogullari , C., & Logendran, R. (2010). Group-schedulingproblems in electronics manufacturing. Journal of Scheduling, 13, 177-202. Glover, F. (1986). Future Paths for integer programming and links to artificial intelligence. Computers and Operations Research, 13, 533-549. Glover, F., & Laguna, M. (1997). Tabu Search. Boston:Kluwer Academic. Graham, R., Lawler, E., Lenstra, J., & Rinnooy Kan, A. (1979). Optimization and approximation in deterministic sequencing and scheduling: A survey. Annals of Discrete Mathematics, 5, 287-326. Huo, Y., Leung, J., & Zhao, H. (2007). Bi-criteria scheduling problems: Number of tardy jobs. European Journal of Operational Research, 177, 116-134. IBM. (2009). ILOG CPLEX Optimization Studio, Version 12.2. Jeong, B., Kim, S., & Lee, Y. (2001). An assembly scheduler for TFT LCD manufacturing. Computers & Industrial Engineering, 41, 37-58. Kim, D., Kim, K., Jang , W., & Chen, F. (2002). Unrelated parallel machine scheduling with setup times using simulated annealing. Robotics and Computer-Integrated Manufacturing, 18, 223-231. Kim, D., Na, D., & Chen, F. (2003). Unrelated parallel machine scheduling with setup times and a total weighted tardiness objective. Robotics and Computer-Integrated Manufacturing, 19, 173-181. Kuo, W., & Yang, D. (2006). Single-machine group scheduling with a time-dependent learning effect. Computers & Operations Research, 33, 2099-2112. Laguna, M., Barnes, J., & Glover, F. (1993). Intelligent scheduling with tabu search: An applicaiton to jobs with linear delay penalties and sequence-dependent setup costs and times. Applied Intelligence, 3, 159-172. Li, S., Ng, C., & Yuan, J. (2011). Group scheduling and due date assignment on a single machine. Internaitonal Journal of Production Economics, 130, 230-235. Liaw, C., Lin, Y., Cheng, C., & Chen, M. (2003). Scheduling unrelated parallel machines to minimize total weighted tardiness. Computers & Operations Research, 30, 1777-1789. Lin, Y., Pfund, M., Fowler, J., & Montgomery, D. (2006). Classification of parallel machine environments under various correlation structures. International Conference on Computers and Industrial Engineering, Taipei, Taiwan, 1253-1261. Lin, Y., Pfund, M., & Fowler, J. (2011). Heuristics for minimizing regular performance measures in unrelated parallel machine scheduling problems. Computers & Operations Research, 38, 901- 916. Logendran , R., & Nudtasomboon, N. (1991). Minimizing the makespan of a group scheduling problem: a new heuristic. International Journal of Production Economics, 22, 217-230. Logendran, R., Mai, L., & Talkington, D. (1995). Combined heuristics for bi-level group scheduling problems. International Journal of Production Economics, 38, 133-145. Logendran, R., & Sonthinen, A. (1997). A tabu search-based approach for scheduling job shop type flexible manufacturing systems. Journal of Operational Research Society, 48, 264-277. Logendran, R., & Subur, F. (2004). Unrelated Parallel machine scheduling with job splitting. IIE Transaction, 36, 259-372. Logendran, R., Carson, S., & Hanson, E. (2005). Group scheduling in flexible flow shops. Internaitonal Journal of Production Econnomics, 96, 143-155. Logendran, R., Salmasi, N., & Sriskandarajah, C. (2006a). Two-machine group scheduling problems in discrete parts manufacturing with sequence-dependent setups. Computers & Operations Research, 33, 158-180. Logendran, R., deSzoeke, P., & Barnard, F. (2006b). Sequence-dependent group scheduling problems in flexible flow shops. International Journal of Production Economics, 102, 66-86. Logendran, R., McDonell, B., & Smucker, B. (2007). Scheduling unrelated parallel machines with sequence-dependent setups. Computers & Operations Research, 34, 3420-3438. Lu, D., & Logendran, R. (2011). Bicriteria flow shop group scheduling with sequence-dependent setups. Proceedings (CD-ROM), 20th Annual Industrial Engineering Research Conference (IERC), Reno, NV. Mehravaran, Y., & Logendran, R. (2011). Bicriteria supply chain scheduling on unrelated parallel machines. Journal of Chinese Institute of Industrial Engineers, 28, 91-101. Mohri, S., Masuda, T., & Ishii, H. (1999). Bi-criteria scheduling problem on three identical parallel machines. International Journal of Production Economics, 60-61, 529-536. Montgomery, D.C. (2009). Design and Analysis of Experiments. (7th ed.). New York: Wiley, (chapter 14). Pandya, V., & Logendran, R. (2010). Weighted tardiness minimization in flexible flowshops. Proceedings (CD-ROM), 19th Annual Industrial Engineering Research Conference (IERC), Cancun, Mexico. R version 2.13.0 (2011). The R Foundation for Statistical Computing. Rocha, P., Ravetti, M., Mateus, G., & Pardalos, P. (2008). Exact algorithms for a scheduling problem with unrelated parallel machines and sequence and machine-dependent setup times. Computers & Operations Research, 35, 1250-1264. Salmasi, N., Logendran, R., & Skandari, M. (2010). Total flow time minimization in a flowshop sequence-dependent group scheduling problem. Computers & Operations Research, 37, 199-212. Salmasi, N., Logendran, R., & Skandari, M. (2011). Makespan minimization of a flowshop sequence- dependent group scheduling problem. International Journal of Advanced Manufacturing Technology, 56, 699-710. Schaller, J., Gupta, J., & Vakharia, A. (2000). Scheduling a flowline manufacturing cell with sequence dependent family setup times. European Journal of Operational Research, 125, 324-339. Shahvari, O., Salmasi, N., Logendran, R., & Abbasi, B. (2011). An efficient tabu search algorithm for flexible flow shop sequence-dependent group scheduling problems. International Journal of Production Research, 1-18. doi:10.1080/00207543.2011.604051. Xingong, Z., & Guangle, Y. (2010). Single-machine group scheduling problems with deteriorated and learning effect. Applied Mathematics and Computation, 216, 1259-1266. Yang, W., & Chand, S. (2008). Learning and forgetting effects on a group scheduling problem. European Journal of Operational Research, 187, 1033-1044. 
work_zpbskklu2raf7lvljawgbpau3u	----	Discovering cohesive subgroups from social networks for targeted advertising Discovering cohesive subgroups from social networks for targeted advertising Wan-Shiou Yang *, Jia-Ben Dia Department of Information Management, National Changhua University of Education, No. 1, Jin-De Road, Changhua 500, Taiwan, ROC Abstract In this paper, we propose a framework that utilizes the concept of a social network for the targeted advertising of products. This approach discovers the cohesive subgroups from a customer’s social network as derived from the customer’s interaction data, and uses them to infer the probability of a customer preferring a product category from transaction records. This information is then used to construct a targeted advertising system. We evaluate the proposed approach by using both synthetic data and real-world data. The experi- mental results show that our approach does well at recommending relevant products. � 2007 Elsevier Ltd. All rights reserved. Keywords: Social network; Targeted advertising; Recommender system; Knowledge discovery 1. Introduction The effectiveness of targeting a small portion of custom- ers for advertising has long been recognized by businesses (Armstrong & Kotler, 1999) for two main reasons. First, the amount of product/service information available to customers is ever-increasing, and hence it is desirable to help customers wade through the information to find the product/service they want. Second, understanding the needs of current and potential customers is an essential part of customer-relationship management. The ability to accurately and efficiently identify the needs of customers and subsequently advertise products/services that they will find desirable will increase customer-retention, growth, and profitability of a business (Armstrong & Kotler, 1999). The traditional approach to targeted advertising is to (manually) analyze a historical database of previous trans- actions and the features associated with the (potential) cus- tomers, possibly with the help of some statistical tools, and identify a list of those customers who are most likely to respond to the advertisement of the product. The advent of new technologies has lead to automatic tools being advocated for identifying potential customers (Hayes, 1994), with many recommender systems having emerged over the past few years whose basic idea is to advertise products according to users’ preferences as obtained by ratings either explicitly stated by the users or implicitly inferred from previous transaction records, Web logs, or cookies. The first type of recommendation technique was the con- tent-based approach (Cohen, 1992), in which recommend- able products are characterized by a set of content features, and customers’ interests are represented by a sim- ilar feature set. Content-based approaches select target cus- tomers whose interests have a high degree of similarity to the product’s content profile. To establish an accurate con- tent profile for a product, the detailed description of a product must be parsable (e.g., as text), and a set of content features are extracted by some information-extraction or summarization techniques (Mooney & Roy, 2002). The content-based approach is inappropriate for products whose content is not electronically available, or is based on subjective factors such as quality, style, or point of view. Furthermore, since the content features of an individual are derived purely from the products in which s/he has 0957-4174/$ - see front matter � 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.02.028 * Corresponding author. Tel.: +886 4 7232105x7611. E-mail address: wsyang@cc.ncue.edu.tw (W.-S. Yang). www.elsevier.com/locate/eswa Available online at www.sciencedirect.com Expert Systems with Applications 34 (2008) 2029–2038 Expert Systems with Applications mailto:wsyang@cc.ncue.edu.tw shown an interest, the content-based approach is inappro- priate to advertise products to a new customer and will result in a low probability of providing surprising adverti- sements. Another type of recommendation technique is the col- laborative approach (or sometimes called the social-based approach) (Cohen, 1992). The collaborative approach looks for relevance among customers by observing their ratings assigned to products in a small training set. The nearest- neighbor customers are those that exhibit the strongest relevance to the target customer. These customers act as ‘‘recommendation partners’’ for the target customer, and collaborative approaches advertise products to the target customer that appear in the profiles of these recommenda- tion partners but not in that of the target customer. Whilst this approach has demonstrated usefulness in many appli- cations, it still has limitations such as an inability to adver- tise newly introduced products that have yet to be rated by customers, an inability to advertise products to a new cus- tomer who has yet to provide rating data, and poor predic- tions if the data are sparse. In order to remedy these problems, in this research we propose a framework that utilizes the concept of a social network to facilitate the automatic construction of a tar- geted advertising system. Our approach applies data min- ing techniques to gather social interaction data so as to cluster customers into a set of cohesive subgroups. Based on the set of cohesive subgroups, we infer the probabilities of a customer preferring a product category from transac- tion records. This information is crucial to the successful online promotion of products to customers. Utilizing such information allows targeted advertising to be conducted on a broader scope, by advertising new products to new customers. This paper is organized as follows. In Section 2, we introduce the concept of a social network and define our problem. Section 3 describes the algorithm designed for discovering cohesive subgroups from a social network. Sec- tion 4 derives the algorithm used to select the customers to which to promote a given product. The proposed approach is evaluated first using synthetic data and then using empir- ical data obtained from our university email logs and library-circulation data. We report the evaluation results in Section 5. Section 6 describes related work, and Section 7 concludes with a summary and a discussion of future research directions. 2. The problem A social network is a set of individuals connected through socially meaningful relationships, such as friend- ship, coworking, or information exchange (Wasserman, Faust, Iacobucci, & Granovetter, 1994; Wellman, 1996). Social networks are formed when people interact with each other (Garton, Haythornthwaite, & Wellman, 1997) and thus can be seen in many aspects of everyday life. Social network theory traditionally views social relationships in terms of nodes and links (Wasserman et al., 1994), where the nodes are the individual actors within the networks and the links are the relationships between the actors. In its most simple form, a social network is a map of all of the relevant links between the nodes being studied. These concepts are often displayed in a social network diagram, with nodes indicates as points and links indicated as lines (Wasserman et al., 1994). The strength of ties between nodes in a real-world social network is an important theoretical issue. As noted by Mil- gram (1967), the strength of a tie between two actors is much greater if they have another mutual acquaintance. In other words, the probability of two friends of an individual know- ing each another is much greater than the probability of two people chosen randomly from the population knowing each another (Guare, 1990; Newman, Watts, & Strogatz, 2002; Watts & Strogatz, 1998). Actors with strong ties usually have some sort of common ground on which they establish their relationships (Preece, 2002; Wellman & Gulia, 1997), and thus often constitute a subgroup. Because of the com- mon ground, actors with strong ties – and hence represent- ing a subgroup – often share common interests, needs, or services that provide a reason for the subgroup (Preece, 2002; Schwartz & Wood, 1993; Wellman & Gulia, 1997). In reality, a customer’s decision to buy a product is often strongly influenced by his/her friends, acquaintances, or business partners, etc. A classic example is the Hotmail free email service, whose growth from zero to 12 million users within 18 months is attributed to the inclusion of a promotional message containing the service’s URL in every email sent using it (Jurvetson, 2000). Because of the mutual influences, customers with strong ties are apt to exhibit similar purchasing behavior. This observation initiates the interesting idea that identifying subgroups – comprising customers with strong ties and hence common interests – from a customer’s social network may facilitate the target- ing of a small portion of customers for advertising. The advent of information techniques, especially com- munication systems (e.g., email, bulletin boards, and mes- saging) and contacting systems (e.g., MSN (http:// www.msn.com), ICQ (http://www.icq.com), Orkut (http:// www.orkut.com), and Friendster (http://www.friend- ster.com)), has lead to a rapid growth in computer-medi- ated social networks. The human activities on these networks generate a huge amount of data suitable for ana- lyzing the social networks and facilitating the automatic construction of a targeted advertising system. Therefore, in this research we investigate the problem of targeted advertisement in an environment with the follow- ing features: (1) A database that contains customer’s connectedness is available This database can be obtained by links either explicitly stated by customers (e.g., friendships stated on contacting sites) or implicitly inferred from previous interaction data (e.g., email logs). We transform the connectedness database into a customer’s social network and represent the network 2030 W.-S. Yang, J.-B. Dia / Expert Systems with Applications 34 (2008) 2029–2038 http://www.msn.com http://www.msn.com http://www.icq.com http://www.orkut.com http://www.orkut.com http://www.friendster.com http://www.friendster.com https://isiarticles.com/article/2090 
work_zpfso6pk3vf5df7625v6vawp2e	----	Estimating the square root of probability  density function on Riemannian manifold  Article  Accepted Version  Hong, X. and Gao, J. (2018) Estimating the square root of  probability density function on Riemannian manifold. Expert  Systems: International Journal of Knowledge Engineering.  e12266. ISSN 1468­0394 doi:  https://doi.org/10.1111/exsy.12266 Available at  http://centaur.reading.ac.uk/75374/  It is advisable to refer to the publisher’s version if you intend to cite from the  work.  See Guidance on citing  . To link to this article DOI: http://dx.doi.org/10.1111/exsy.12266  Publisher: Wiley  All outputs in CentAUR are protected by Intellectual Property Rights law,  including copyright law. Copyright and IPR is retained by the creators or other  copyright holders. Terms and conditions for use of this material are defined in  the End User Agreement  .  www.reading.ac.uk/centaur    CentAUR  http://centaur.reading.ac.uk/71187/10/CentAUR%20citing%20guide.pdf http://www.reading.ac.uk/centaur http://centaur.reading.ac.uk/licence Central Archive at the University of Reading  Reading’s research outputs online Estimating the Square Root of Probability  Density Function on Riemannian Manifold  Article  Hong, X. and Gao, J. (2018) Estimating the Square Root of  Probability Density Function on Riemannian Manifold. Expert  Systems. ISSN 1468­0394 Available at  http://centaur.reading.ac.uk/75843/  It is advisable to refer to the publisher’s version if you intend to cite from the  work.  Publisher: Wiley­Blackwell  All outputs in CentAUR are protected by Intellectual Property Rights law,  including copyright law. Copyright and IPR is retained by the creators or other  copyright holders. Terms and conditions for use of this material are defined in  the End User Agreement  .  www.reading.ac.uk/centaur    CentAUR  Central Archive at the University of Reading  http://centaur.reading.ac.uk/licence http://www.reading.ac.uk/centaur Reading’s research outputs online Estimating the Square Root of Probability Density Function on Riemannian Manifold Xia Hong and Junbin Gao Abstract We propose that the square root of a probability density function can be represented as a linear combination of Gaussian kernels. It is shown that, if the Gaussian kernel centres and kernel width are known, then the maximum likelihood parameter estimator can be formulated as a Riemannian optimization problem on sphere manifold. The first order Riemannian geometry of the sphere manifold and vector transport are initially explored, then the well-known Riemannian conjugate gradient algorithm is used to estimate the model parameters. For completeness the k- means clustering algorithm and a grid search are employed to determine the centers and kernel width respectively. Simulated examples are employed to demonstrate that the proposed approach is effective in constructing the estimate of the square root of probability density function. 1 Introduction The importance of probability density function (PDF) is evidenced by intensive the- oretic researches and many data analysis and pattern recognition applications [McLach- lan and Peel, 2000, Silverman, 1986, Duda and Hart, 1973, Chen et al., 2010, Rutkowski, 2004, Yin and Allinson, 2001, Dempster et al., 1977, Parzen, 1962]. The Parzen window estimator (PW) and the finite mixture models, especially a mixture of Gaussians, are two widely researched and popular estimators. The mixture of Gaussians model represents an underlying PDF as a linear combination of Gaus- sian kernels, and the maximum likelihood (ML) estimator of the mixture models Xia Hong Department of Computer Science, School of Mathematical and Physical Sciences, University of Reading, UK, e-mail: x.hong@reading.ac.uk Junbin Gao Discipline of Business Analytics, The University of Sydney Business School, The University of Sydney, NSW 2006, Australia, e-mail: junbin.gao@sydney.edu.au Xia Hong and Junbin Gao parameters can be obtained using the expectation- maximization (EM) algorithm. Alternatively, the Parzen window (PW) estimator [Parzen, 1962] can be regarded as a special case of the finite mixture model [McLachlan and Peel, 2000], in which the number of mixtures is equal to that of the training data samples and all the mixing weights are equal, providing as a simple practical PDF estimator. However if the number of training data samples is very large, then the point density estimate using the PW estimator for a future data sample can be computationally expensive. The as- sociated ML optimisation in a general finite mixture Gaussian is generally a highly nonlinear optimisation process requiring extensive computation, while the EM algo- rithm for Gaussian mixture models enjoys an explicit iterative form [Bilmes, 1998]. There are also considerable interests into research on sparse PDF estimation which can be summarized into two categories. The first category is based on constrained optimization [Weston et al., 1999, Vapnik and Mukherjee, 2000, Girolami and He, 2003, Hong et al., 2015]. The second category of sparse kernel density estimators construct the PDF estimator in a forward regression manner [Choudhury, 2002,Chen et al., 2004b, Chen et al., 2004a, Chen et al., 2008, Hong et al., 2013]. The Riemannian optimisation algorithms have been recently researched on many types of matrix manifolds such as the Stiefel manifold, Grassmann manifold and the manifold of positive definite matrices, see Section 3.4 of [Absil et al., 2008]. Since Riemannian optimisation is directly based on the curved manifolds, one can elim- inate those constraints such as orthogonality to obtain an unconstrained optimisa- tion problem that, by construction, will only use feasible points. This allows one to incorporate Riemannian geometry in the resulting optimisation problems, thus pro- ducing far more accurate numerical results. The Riemannian optimisation have been successfully applied in machine learning, computer vision and data mining tasks, including fixed low rank optimisation [Mishra et al., 2013], Riemannian dictionary learning [Harandi et al., 2014], and computer vision [Lui, 2012]. Since the constraint on the mixing coefficients of the finite mixture model is the multinomial manifold, recently we have introduced Riemannian trust-region (RTR) algorithm for sparse fi- nite mixture model based on minimal integrated square error criterion [Hong et al., 2015]. Note that all the aforementioned PDF estimation algorithms are aimed at directly estimating the probability density function. However in this work we investigate the less studied problem of estimating the square root of the probability density function (PDF) estimation [Pinheiro and Vidakovic, 1998], i.e., the square root of a PDF, rather than the PDF itself is a mixture of Gaussians. Clearly the resultant estimator can be used similarly in many applications as can a PDF estimator. However the estimation of the square root of the probability density function (PDF) poses as a different problem, and new estimation algorithm is required. In this paper, we introduce/extend a fast maximum likelihood estimator based on a linear combination of Gaussian kernels which represents the square root of probability density function [Hong and Gao, 2016]. Initially we apply the k-means clustering algorithm and a grid search to determine the centers and kernel width. It is shown that the underlying parameter estimation problem can be formulated as a Riemannian optimization on the sphere manifold. The first order Riemannian geom- Estimating the Square Root of Probability Density Function on Riemannian Manifold etry of the sphere manifold and vector transport are explored, and the well-known Riemannian conjugate gradient algorithm is used to estimate the model parameters. Numerical examples are employed to demonstrate that the proposed approach is effective in constructing the estimate of the square root of probability density func- tion. 2 Preliminary on Sphere Manifold This section briefly introduces the concept of sphere manifold and the necessary ingredients used in the retraction based framework of Riemannian optimization. As a reference, the main notations on Riemannian geometry on sphere manifold in this section is summarized in Table 1. We refer the readers to [Absil et al., 2008] for the general concepts of manifolds. Table 1 Notations for Sphere Manifold{ SM−1,g } Sphere manifold for parameter matrix θ and the inner product of the manifold Tθ SM−1 Tangent space of the sphere manifold uθ ,vθ Tangent vectors at θ Projθ (z) Orthogonal projector from a vector in ambient space onto the tangent space at θ gradF(θ ) Riemannian gradient of F(θ ) on the manifold SM−1 GradF(θ ) Classical gradient of F(θ ) as seen in Euclidean space Expθ Retraction mapping Tθk+1←θk (uθk ) Vector transport The sphere manifold is the set of unit Frobenius norm vectors of size M, denoted as SM−1 = { θ ∈RM : ∥θ∥2 = 1 } . (1) It is endowed with a Riemannian manifold structure by considering it as a Rieman- nian submanifold of the embedding Euclidean space RM endowed with the usual inner product g(uθ ,vθ ) =uTθ vθ , (2) where uθ ,vθ ∈ Tθ SM−1 ⊂ RM are tangent vectors to SM−1 at θ . The inner product on SM−1 determines the geometry such as distance, angle, curvature on SM−1. Note that the tangent space Tθ SM−1 at element θ can be described by Tθ SM−1 = { uθ : uTθ θ = 0 } . (3) Xia Hong and Junbin Gao Riemannian gradient: Let the Riemannian gradient of a scalar function F(θ ) on SM−1 be denoted by gradF(θ ), and its classical gradient as seen in the Euclidean space as GradF(θ ). Then we have gradF(θ ) = Projθ ( GradF(θ ) ) , (4) where Projθ (z) is the orthogonal projection onto the tangent space, which can be computed as ProjX(z) = z−(θ Tz)θ (5) in which z represents a vector in the ambient space. Retraction mapping: An important concept in the recent retraction-based frame- work of Riemannian optimization is the retraction mapping, see Section 4.1 of [Ab- sil et al., 2008]. The exponential map ExpX, defined by Expθ ( λ uθ ) = cos ( ∥λ uθ∥2 ) θ + sin(∥λ uθ∥2) ∥uθ∥2 uθ , (6) is the canonical choice for the retraction mapping, where the scalar λ is a chosen step size. The retraction mapping is used to locate the next iterate on the manifold along a specified tangent vector, such as a search direction in line search in the New- ton’s algorithm or the suboptimal tangent direction in the trust-region algorithm, see Chapter 7 of [Absil et al., 2008]. For example, the line search algorithm is simply given by θk+1 = Expθk ( λkuθk ) . (7) where the search direction uθk ∈ Tθk S M−1 and λk is a chosen step size at iteration step k. Vector Transport: In Riemannian optimization algorithms, the second derivatives can be approximated by comparing the first-order information (tangent vectors) at distinct points on the manifold. The notion of vector transport Tθk+1←θk (uθk ) on a manifold, roughly speaking, specifies how to transport a tangent vector uθk from a point θk to another point θk+1 on the manifold. The vector transport for the sphere manifold is calculated as Tθk+1←θk (uθk ) = Projθk+1 (uθk ) (8) 3 Proposed estimator for the square root of probability density function Given the finite data set DN ={x j}Nj=1 consisting of N data samples, where the data x j ∈ Rm follows an unknown PDF p(x), the Gaussian mixture model [McLachlan and Peel, 2000] is in the form of Estimating the Square Root of Probability Density Function on Riemannian Manifold pGMM(x) = P ∑ i=1 gi (2π)m/2|Σi|m/2 exp ( − 1 2 (x−µ i) TΣi−1(x−µ i) ) (9) where µi and Σ i are called the mean vector, and covariance matrix (positive definite) of the ith mixture. The mixing coefficients satisfies 0 ≤ gi ≤ 1 for all i, and P ∑ i=1 gi = 1 (10) In this work we investigate the less studied problem of estimating the square root of the probability density function (PDF) estimation [Pinheiro and Vidakovic, 1998]. The reason that the root square of the probability density function is esti- mated is that the problem can be formulated as a Riemannian optimisation one for computational advantage, as well as that the resultant estimator can be used as an alternative to a probability density function estimator. Specifically, the problem un- der study is to find a sparse approximation of the square root of ψ(x) = √ p(x) > 0 using M component Gaussian kernels, given by ψ(x) = M ∑ i=1 ωiKσ ( x,ci ) = ω Tk(x) (11) subject to ∫ ψ 2(x)dx = 1 (12) where Kσ ( x,ci ) = 1( 2π σ 2 )m/2 exp ( − ∥x−ci∥2 2σ 2 ) , (13) in which ci = [ ci,1 ci,2 ···ci,m ]T is the center vector of the ith kernels and σ > 0 is the width parameter, and ωi is the ith kernel weight. ω = [ω1 ···ωM]T, k(x) = [Kσ ( x,c1 ) ···Kσ ( x,cM ) ]T. Applying (11) to the constraint (12), we have ∫ ψ 2(x)dx = M ∑ i=1 M ∑ j=1 ωiω j ∫ Kσ ( x,ci ) Kσ ( x,c j ) dx = ω TQω = 1 (14) where Q ={qi, j}, with qi, j = K√2σ ( ci,c j ) , as shown in Appendix A. Clustering algorithms can be used to find a set of centers which accurately reflects the distribution of the data points. From N data points x j , j = 1,··· ,N, the k-means algorithm [Haykin, S., 2009] seeks to partition the data points in M disjoint subset Si, each containing Ni data points, so as to minimize the sum-of-squares clustering function given by Xia Hong and Junbin Gao J = M ∑ i=1 ∑ x j∈Si ∥x j −ci∥2 (15) where ∈ denotes belongs to. J is minimized when ci = 1 Ni ∑ x j∈Si x j (16) The k-means algorithm is detailed in Algorithm 1. Algorithm 1 The k-means clustering algorithm Require: x j , j = 1,··· ,N, and a preset number of centers M. Ensure: ci, j = 1,··· ,M, that yields the minimum of J = ∑Mi=1 ∑x j∈Si ∥x j −ci∥ 2 1: Randomly select M data points from DN as initial centers coldi . 2: Randomly draw a data point x j from DN . 3: From all coldi , i = 1,··· ,M, find the nearest center to x j , denoted as c old k . 4: Update cnewk = c old k + ε(x j −c old k ), where ε > 0 is the predetermined learning rate. 5: Set cnewi as c old i . 6: Goto Step 2 until a sufficient large number of iterations has reached (for convergence). 7: Return cnewi as ci, i = 1,··· ,M. We now propose a new maximum likelihood estimation for parameters ωi which is based on that known M kernel centers and a preset kernel width is given. Initially denote the eigenvalue decomposition Q as Q = UΣ UT, where U is an orthogonal matrix consisting of the eigenvectors and Σ is a diagonal matrix of which the entries are eigenvalues Q. Let θ = √ Σ UTω , ψ(x) can be written as ψ(x) = θ Tk̄(x) (17) where k̄(x) = √ Σ −1 UTk(x). In order to satisfy ψ(x) > 0, we initialize all ωi as the same positive number ξ , and ωIni = ξ 1M , 1M is a length M vector with all elements as ones. The correspond- ing θ0 can be calculated as θ0 = √ Σ UT1M (18) followed by a normalization step θ0 = θ0/∥θ0∥ so that θ0 is on the sphere manifold. We are now ready to formulate the maximum likelihood estimator as the following Riemannian optimization problem, given as θ opt = min θ∈SM−1 { F(θ ) =− N ∑ j=1 log(θ Tk̄(x j)) } (19) followed by setting θ opt = √ Σ UTω opt. For our objective function F(θ ), it is easy to check that Euclidean gradient GradF(θ ) can be calculated respectively as Estimating the Square Root of Probability Density Function on Riemannian Manifold GradF(θ ) =− N ∑ j=1 k̄(x j) θ Tk̄(x j) (20) Based on GradF(θ ), Riemannian gradient of the objective function F(θ ) on the sphere manifold can be calculated according to (4), (5). We opt to Riemannian con- jugate gradient algorithm to solve (19), which generalizes the classical conjugate gradient algorithm [Hager and Zhang, 2006] to optimization problems over Rie- mannian manifolds [Boumal et al., 2014]. Since the logarithm function acts as a natural barrier at zero and the Riemannian conjugate gradient algorithm is a local minimization algorithm, the constraint ψ(x) > 0 can be met. With all the ingredients available, we form the algorithm for solving (19) in Algo- rithm 1, which is well implemented in the Manifold Optimization Toolbox Manopt http://www.manopt.org, see [Boumal et al., 2014]. We used the default pa- rameter settings in Manopt, so that in Step 4 of Algorithm 1 αk is based on line search backtracking procedure as described in Chapter 4, p63 of Section 4 of [Absil et al., 2008]. In Step 5 of Algorithm 1 ωk+1 is based on the default option “Hestenes- Stiefel’s modified rule”. Specifically, we have βk+1 = max { 0, < gradF(θk+1),γk+1 > < Tθk+1←θk (ηk),γk+1 > } (21) where < ·,·> denotes inner product, and γk+1 = gradF(θk+1)−Tθk+1←θk (gradF(θk)) (22) Algorithm 2 Riemannian conjugate gradient Algorithm for solving (19) Require: k̄(x j), j = 1,··· ,N. Initial point θ0 which is on the sphere manifold SM−1, and default threshold ϖ , e.g., ϖ = 10−6 or ϖ = 10−5; Ensure: θ that yields the minimum F ( θ ) . 1: Set η0 =−gradF(θ0) and k = 0; 2: while ∥gradF(θk)∥2 < ϖ do 3: k = k + 1; 4: Compute a step size αk and set θk = Expθk−1 ( αk ηk−1 ) ; (23) 5: Compute βk and set ηk =−gradF(θk)+ βkTθk←θk−1 (ηk−1); (24) 6: end while 7: Return θ = θk , F(θ ), and the associated U, Σ . In summary our proposed algorithm consists of two consecutive steps; (i) at the first stage, we determine M kernel centres using the k-means clustering algorithm, as presented in Algorithm 1; (ii) for a grid search of kernel width, estimate kernel parameters using Riemannian conjugate gradient algorithm ω based on the maxi- Xia Hong and Junbin Gao Algorithm 3 Proposed estimator for the square root of probability density function Require: x j, j = 1,··· ,N. Preset number of kernels M; Ensure: ψ(x) that maximizes the likelihood. 1: Use the k-means clustering algorithm (Algorithm 1) to find ci,i = 1,··· ,M; 2: for i = 1,2,...,(Iter + 1) do 3: σ = σmin + i−1n (σmax −σmin). where (Iter + 1),σmin and σmax are preset, and they denote the minimum, maximum value and the total number of kernel width on a grid search. 4: Form Q and its eigen-decomposition Q = UΣ UT 5: Obtain k̄(x j) = √ Σ −1 UTk(x j), j = 1,··· ,N 6: Find θ 0 according to (18) and normalize it. 7: Apply the Riemannian conjugate gradient Algorithm (Algorithm 2) to return F(θ ). 8: end for 9: Find σ opt that minimizes F(θ ) over the grid search. Return the associated θ , U, Σ . 10: Return ω opt = U √ Σ −1 θ opt. mum likelihood criterion. For completeness, the proposed estimator is detailed in Algorithm 3. The computational complexity consists of the k-means clustering which is in the order Iterkmeans ∗O(N), where Iterkmeans is the number of iterations of k-means clustering algorithm, and that of the Riemannian conjugate algorithm, scaled by the number of grid search Iter. The main cost in Riemannian conjugate algorithm is function and derivative evaluation of (19)& (20) which are in the order of O(MN). Hence the total cost of the proposed algorithm is evaluated as Iterkmeans ∗O(N) + Iter∗O(MN). 4 Illustrative examples Example 1: A data set of N = 600 points was randomly drawn from a known dis- tribution p(x) and used to construct the estimator of the square root of probability density function ψ(x) using the proposed approach. p(x) is given as a mixture of one Gaussian and one Laplacian, as defined by p(x) = 2 3π exp ( −2(x1 −1)2 + 2(x2 −1)2 ) + 0.35 6 exp(−0.7|x1 + 1|−0.5|x2 + 1|) (25) We preset M = 40, the k-means algorithm was applied to find M = 40 centers ci,i = 1,··· ,M. The Riemannian conjugate gradient algorithm was applied to mini- mize the negative likelihood based on Q obtained a range of kernel widths of [0.5,2], with a grid width of 0.1. The result of log F(θ ) versus the kernel width was shown in Figure 1. The convergence of Riemannian conjugate gradient algorithm on the sphere manifold based on the optimal kernel width was shown shown in Figure 2, showing that the algorithm can converge rapidly. The performance of proposed es- Estimating the Square Root of Probability Density Function on Riemannian Manifold timator was shown in Figure 3(a),(b)&(c), which plots ψ 2(x), the true density p(x), and the estimation error e(x) = ψ(x)2 − p(x) respectively over a 41×41 meshed data, with a grid size of 0.2, ranging from -4 to 4 for both x1 and x2. 0.5 1 1.5 2 1220 1240 1260 1280 1300 1320 1340 1360 σ V a lu e o f m in im a l o f F ( θ ) Fig. 1 log(F(θ )) versus the kernel width for Example 1. 0 5 10 15 20 600 605 610 615 620 625 630 635 640 645 Iterations lo g F ( θ) Fig. 2 Convergence of Riemannian conjugate gradient algorithm for Example 1. The experiment was repeated for 100 different random runs in order to com- pare the performance of the proposed algorithm with the Gaussian mixture model (GMM) that is fitted using the EM algorithm [McLachlan and Peel, 2000]. A sepa- Xia Hong and Junbin Gao −4 −2 0 2 4 −4 −2 0 2 4 0 0.02 0.04 0.06 0.08 0.1 x 2 x 1 ψ ( x )2 (a) −4 −2 0 2 4 −4 −2 0 2 4 0 0.02 0.04 0.06 0.08 0.1 x 2 x 1 p ( x ) (b) −4 −2 0 2 4 −4 −2 0 2 4 0 0.02 0.04 0.06 0.08 0.1 x 2 x 1 e ( x ) (c) Fig. 3 The result of the proposed estimator for Example 1; (a) The resultant pdf estimator ψ(x)2 (b) The true pdf p(x); and (c)The estimation error e(x). Estimating the Square Root of Probability Density Function on Riemannian Manifold rate test data set of Ntest = 1681 was generated using a 41×41 meshed data, with a grid size of 0.2, ranging from -4 to 4 for both x1 and x2. These points were used for evaluation according to L1 = { 1 Ntest ∑ Ntest k=1 |p(xk)−ψ 2(xk)| Proposed 1 Ntest ∑ Ntest k=1 |p(xk)− pGMM(xk)| GMM (26) where pGMM is the resultant GMM model based on two Gaussian mixtures, with the constraint that the covariance matrix is diagonal. The GMM model fitting is im- plemented using Matlab command gmdistribution. f it.m. The results are as shown Table 2 which demonstrate that the proposed algorithm outperforms the GMM fitted using EM algorithm with two mixtures. Table 2 Performance of proposed density estimator in comparison with GMM model for Example 1. Method L1 test error (mean ± STD) Proposed (2.2±0.3)×10−4 GMM (3.2±0.6)×10−4 Example 2: A data set of N = 1500 points was randomly drawn from a known distribution p(x) given as a mixture of one Gaussian and one Laplacian, as defined by p(x) = 1 2.1π exp ( −2(x1 −2)2 −(x2 −2)2/0.98 ) + 0.35 6 exp(−0.7|x1 + 1|−0.5|x2 + 1|) (27) and the proposed approach is experimented to estimate ψ(x), the square root of p(x). The number of centers are empirically set as M = 40, the k-means algorithm was implemented according to Algorithm 1, producing M centers ci,i = 1,··· ,M. Based on the centers, the kernel matrices Q are generated using any kernel width on the grid point on [0.2,2], with a grid width of 0.1. The Riemannian conjugate gradi- ent algorithm was used to minimize the negative likelihood for the range of kernel widths. The result of log F(θ ) versus the kernel width was shown in Figure 4. The convergence of Riemannian conjugate gradient algorithm on the sphere manifold based on the optimal kernel width was shown shown in Figure 5, showing that the algorithm can converge rapidly. The performance of proposed estimator was shown in Figure 6(a),(b)&(c), which plots ψ 2(x), the true density p(x), and the estimation error e(x) = ψ(x)2 − p(x) respectively over a 41×41 meshed data, with a grid size of 0.2, ranging from -4 to 4 for both x1 and x2. The experiment was repeated for 100 different random runs in order to com- pare the performance of the proposed algorithm with the Gaussian mixture model Xia Hong and Junbin Gao 0 5 10 15 20 25 30 1500 1550 1600 1650 1700 1750 Iterations lo g F ( θ) Fig. 4 log(F(θ )) versus the kernel width for Example 2. 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 3000 4000 5000 6000 7000 8000 9000 10000 11000 12000 σ V a lu e o f m in im a l o f F ( θ ) Fig. 5 Convergence of Riemannian conjugate gradient algorithm for Example 2. (GMM) that is fitted using the EM algorithm [McLachlan and Peel, 2000]. A sepa- rate test data set of Ntest = 1681 was generated using a 41×41 meshed data, with a grid size of 0.2, ranging from -4 to 4 for both x1 and x2. These points were used for evaluation according to L1 = { 1 Ntest ∑ Ntest k=1 |p(xk)−ψ 2(xk)| Proposed 1 Ntest ∑ Ntest k=1 |p(xk)− pGMM(xk)| GMM (28) Estimating the Square Root of Probability Density Function on Riemannian Manifold −4 −2 0 2 4 −4 −2 0 2 40 0.02 0.04 0.06 0.08 0.1 x 2x 1 p ( x ) (a) −4 −2 0 2 4 −4 −2 0 2 40 0.05 0.1 x 2x 1 ψ ( x )2 (b) −4 −2 0 2 4 −4 −2 0 2 40 0.05 0.1 x 2x 1 e ( x ) (c) Fig. 6 The result of the proposed estimator for Example 2; (a) The resultant pdf estimator ψ(x)2 (b) The true pdf p(x); and (c)The estimation error e(x). Xia Hong and Junbin Gao where p̂GMM is the resultant GMM model, with the constraint that the covariance matrix is diagonal. The GMM model fitting is implemented using Matlab command gmdistribution. f it.m with two mixtures. The results are as shown Table 3, showing that the proposed method is better than Gaussian mixture models with two mixtures. Table 3 Performance of proposed density estimator in comparison with GMM model for Example 2. Method L1 test error (mean ± STD) Proposed (1.7±0.2)×10−3 GMM (2.9±0.1)×10−3 5 Conclusions In this paper a new method has been introduced to estimate the square root of prob- ability density function from observational data using maximum likelihood. The proposed model is in the form of a linear combination of Gaussian kernels. Incor- porating the k-means clustering algorithm to determine the kernel center as well as a grid search to determine the kernel width, the model parameters are then esti- mated using the Riemannian optimization on the sphere. The first order Riemannian geometry of the sphere manifold and vector transport are explored. Two illustra- tive examples are employed to demonstrate that models obtained by the proposed approach are comparable with the GMM model fitted using EM algorithm. In this research we have included the well-known k-means clustering algorithm which is based on a predetermined number M centers which may not optimal. The choice of M is generally dependent on applications, and for k-means clustering al- gorithm, M should be set sufficiently large to perform well. Future research will be focused on model structure optimization so that the model size can be learnt from data. Appendix A Let x = [x1,...xm]T, we have References qi, j = 1 (2π σ 2)m ∫ ... ∫ exp ( − ∥x−ci∥2 2σ 2 − ∥x−c j∥2 2σ 2 ) dx1...dxm = 1 (2π σ 2)m m ∏ l=1 ∫ exp ( − (xl −ci,l)2 2σ 2 − (xl −c j,l)2 2σ 2 ) dxl (29) in which ∫ exp ( − (xl −ci,l)2 2σ 2 − (xl −c j,l)2 2σ 2 ) dxl = ∫ exp ( − x2l −(ci,l + c j,l)xl +(c 2 i,l + c 2 j,l)/2 σ 2 ) dxl = exp ( − c2j,l +c 2 i,l 2 − ( ci,l +c j,l 2 )2 σ 2 ) × ∫ exp ( − [ xl −(ci,l + c j,l) ]2 σ 2 ) dxl (30) By making use of ∫ 1√ 2π σ 2 exp ( − (xl−c) 2 2σ 2 ) dxl = 1, i.e. Gaussian density integrates to one, we have ∫ exp ( − (xl −ci,l)2 2σ 2 − (xl −c j,l)2 2σ 2 ) dxl = √ π σ 2 exp ( − c2j,l +c 2 i,l 2 − ( ci,l +c j,l 2 )2 σ 2 ) = √ π σ 2 exp ( − (c j,l −ci,l)2 4σ 2 ) (31) Hence qi, j = 1 (4π σ 2)m exp ( − ∥ci −c j∥2 4σ 2 ) = K√2σ ( ci,c j ) (32) References Absil, P.-A., Mahony, R., and Sepulchre, R. (2008). Optimization Algorithms on Matrix Manifolds. Princeton University Press. Bilmes, J. A. (1998). A gentle tutorial of the em algorithm and its application to parameter estimation for gaussian mixture and hidden markov models. Tech- nical Report ICSI-TR-97-021, University of California, Berkeley. Boumal, N., Mishra, B., Absil, P.-A., and Sepulchre, R. (2014). Manopt, a mat- lab toolbox for optimization on manifolds. J. Machine Learning research, 15:1455–1459. Xia Hong and Junbin Gao Chen, S., Hong, X., and Harris, C. J. (2004a). Sparse kernel density construction using orthogonal forward regression with leave-one-out test score and local regularization. IEEE Trans. on Systems, Man and Cybernetics, Part B: Cy- bernetics, 34(4):1708–1717. Chen, S., Hong, X., and Harris, C. J. (2008). An orthogonal forward regression tech- nique for sparse kernel density estimation. Neurocomputing, 71(4-6):925– 943. Chen, S., Hong, X., and Harris, C. J. (2010). Particle swarm optimization aided orthogonal forward regression for unified data modelling. IEEE Trans. Evo- lutionary Computation, 14:477–499. Chen, S., Hong, X., Harris, C. J., and Sharkey, P. M. (2004b). Sparse modelling using orthogonal forward regression with PRESS statistic and regulariza- tion. IEEE Trans. on Systems, Man and Cybernetics, Part B: Cybernetics, 34(2):898–911. Choudhury, A. (2002). Fast Machine Learning Algorithms for Large Data. PhD thesis, School of Engineering Sciences, University of Southampton. Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977). Maximum likelihood from incomplete data via the EM algorithm. J. Royal Statistical Society B, 39:1–38. Duda, R. O. and Hart, P. E. (1973). Pattern Classification and Scene Analysis. Wiley: New York, 1973. Girolami, M. and He, C. (2003). Probability density estimation from optimally condensed data samples. IEEE Trans. on Pattern Analysis and Machine In- telligence, 25(10):1253–1264. Hager, W. W. and Zhang, H. (2006). A survey of nonlinear conjugate gradient methods. Pacific J. Optimization, 2:35–58. Harandi, M., aqnd C. Shen, R. H., Lovell, B., and Sanderson, C. (2014). Extrin- sic methods for coding and dictionary learning on Grassmann manifolds. arXiv:1401.8126. Haykin, S. (2009). Neural Networks and Learning Machines. Pearson Education Inc. Hong, X., Chen, S., Qatawneh, A., Daqrouq, K., Sheikh, M., and Morfeq, A. (2013). Sparse probability density function estimation using the minimum integrated square error. Neurocomputing, 115:122–129. Hong, X., Gao, J., Chen, S., and Zia, T. (2015). Sparse density estimation on the multinomial manifold. IEEE Transactions on Neural Networks and Learning Systems, 26(11):2972–2977. Hong, X. and Gao, J. B. (2016). A fast algorithm to estimate the square root of probability density function. In Proceeding of AI-2016, Thirty-sixth SGAI International Conference on Artificial Intelligence, Cambridge, UK. Lui, Y. M. (2012). Advances in matrix manifolds for computer vision. Image and Vision Computing, 30:380–388. McLachlan, G. and Peel, D. (2000). Finite Mixture Models. Wiley: New York. Mishra, B., Meyer, G., Bach, F., and Sepulchre, R. (2013). Low-rank optimization with trace norm penalty. SIAM Journal on Optimization, 23:2124–2149. References Parzen, E. (1962). On estimation of a probability density function and mode. Annals of Mathematical Statistics, 33:1066–1076. Pinheiro, A. and Vidakovic, B. (1998). Estimating the square root of density via compactly supported wavelets. Computational Statistics and Data Analysis, 25:399–415. Rutkowski, L. (2004). Adaptive probabilistic neural networks for pattern classifi- cation in time-varying environment. IEEE Trans. Neural Networks, 15:811– 827. Silverman, B. W. (1986). Density Estimation for Statistics and Data Analysis. Chap- man and Hall: London. Vapnik, V. and Mukherjee, S. (2000). Support vector machine for multivariate den- sity estimation. In S. Solla, T. L. and Müller, K. R., editors, Advances in Neural Infromation Processing Systems, pages 659–665. MIT Press. Weston, J., Gammerman, A., Stitson, M., Vapnik, V., Vovk, V., and Watkins, C. (1999). Suppot vector density estimation. In B. Schölkopf, C. B. and Smola, A. J., editors, Advances in Kernel Methods, pages 293–306. MIT Press. Yin, H. and Allinson, N. W. (2001). Self-organizing mixture networks for probabil- ity density estimation. IEEE Trans. Neural Networks, 12:405–411. 
work_zqisuhch45dbnlgxmwgwvbo5ja	----	AN EXPERT SYSTEM FOR VALUATION OF RESIDENTIAL PROPERTIES AN EXPERT SYSTEM FOR VALUATION OF RESIDENTIAL PROPERTIES CONALL BOYLE 1. What is an 'Expert System'? There has been much discussion in the popular media recently on the topic of Expert Systems (for example the BBC 'Horizon' of 21st March, 1983, or the Sunday Times, 1st May, 1983). Expert Systems are 'computer programs that can acquire knowledge from experts and make it available in the form of advice to those less skilled' (Landsdown, 1980). Such expert systems are one of the main building blocks of the 'fifth generation' computers, which are under development, with massive government encouragement, in Japan.Valuation is one of the most important professional skills which the General Practice surveyor has to offer. Could this skill be transferred to a computer, bypassing the need for trained human intervention? Is an expert valuation system possible, or even likely? In this paper, I hope to show that such a computer-based valuation system is indeed a possibility. However, such a development need not fill the profession with alarm. The skill of the valuer will still be required; but the nature of the task of valuation may be changed significantly. 2. What an expert system for valuation might look like You, the valuer, have been asked to value a three-bedroomed semi-detached house in Hall Green, an up-market area of Birmingham. Instead of visiting the property, you conduct the dialogue shown in Figure 1 (see over) with your computer. After you have entered all the details, the computer will give you a print-out which could look like the example shown in Figure 2 (see over). The dialogue contains elements which any valuer can recognise as value-significant items. The questions that must be asked are: — can such information be digested by a computer and turned into a valuation? — if it can, how accurate will the results be? In this paper, I will describe an experiment to carry out Valuations by computer'. The degree of success attained does not mean that we can expect to see the immediate demise of the valuer. However, it does indicate the potential for computerised valuations. Conall Boyle Page 1 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc Figure 1 & 2: Print-out of valuation. The computer digests the information from this dialogue and produces the result shown PROPERTY VALUATION BY COMPUTER THIS PROGRAM WILL ESTIMATE THE SELLING PRICE OF DOMESTIC PROPERTIES IN THE BIRMINGHAM AREA, INCLUDING SOLI HULL AND SUTTON COLDFIELD. PLEASE ENTER DETAILS OF THE PROPERTY, IN REPLY TO THE QUESTIONS THAT FOLLOW. PRESS 'RETURN' AFTER YOU ENTER A NUMBER (YOU WILL NEED TO HAVE THE A TO Z GUIDE TO ANSWER ONE OF THE QUESTIONS). -> PLEASE FIND THE A TO Z STREET GUIDE AND LOOK UP THE REFERE E NUMBER FOR THE ADDRESS OF THE PROPERTY NC THIS WILL BE IN THE FORM ... 2 B 72 PLEASE ENTER AS 2,B,72 ? 2,B,72 -> WHAT TYPE OF PROPERTY? 1 ...SEMI-DETACHED 2 ... DETACHED 3 ... LINK-DETACHED 4 ... BUNGALOW (DETACHED), CHALET 5 ... TERRACED, TOWN HOUSE,MEWS 6... VILLA 7 ... COTT GE A 8 ... FLAT 9... MAISONETTE PLEASE ENTER TYPE NUMBER AS ABOVE ?1 -> AGE OF PROPERTY IN YEARS ? 1 -> HOW MANY BEDROOMS? ENTER NUMBER ? 2 -> HOW MANY RECEPTION ROOMS ? 3 -> WHICH OF THE FOLLOWING EST DESCRIBES THE CONDITION OF THE PROPERTY B 2 ... IN NEED OF SOME RENOVATION 1 ... EXCELLENT THROUGHOUT 0 ...ABOUT AVERAGE ENTER NUMBER ? 1 -> WHAT SORT OF GARAGE 0 ... NONE 1 ...SINGLE GARAGE 2 ... DOUBLE GARAG E 3 ... PARKING SPACE ENTER NUMBER 7 1 -> TYPE OF TENURE 0... LEASEHOLD 1 ... FREEHOLD ENTER NUMBER ? 1 -> CENTRAL HEATI G ? N O...NONE 1 ...FULL CENTRAL HEATING 2 ... PART CENTRAL HEATING ENTER NUMBER ? 1 -> DOUBLE GLAZIN ? G 0 ... NON E 1 ... FULL 2 ... PART ENTER NUMBER ? 1 -> OTHER DESIRABLE FEATURES, SUCH AS SECOND BATHROOM, LAUNDRY ROOM, OUTHOUSES, LARGE GARDEN, DELIGHTFUL VIEW 1 ... IF PR PERTY HAS SUCH FEATURE(S) O 0... IF NOT ENTER NUMBER ? 1 FOR THE PROPERTY DESCRIBED AS - SEMI-DETACHED, FREEHOLD 1 YEARS OLD, IN VERY GOOD CONDITION, WITH 2 BEDROOMS, 1 RECEPTION ROOM, SINGLE GARAGE, FULL CENTRAL HEATING, FULL DOUBLE GLAZING, AND HAVING OTHER DESIRABLE FEATURES,SITUATED AT A TO Z LOCATION 2 B 72 "THE ESTIMATED SELLING PRICE IS* £28,564 NB THIS FIGURE IS SUBJECT TO AN ERROR OF UP TO 10%, AND MAY VARY BECAUSE OF PARTICULAR CIRCUMSTANCES.) Conall Boyle Page 2 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc . 3. What are the requirements for an expert system? The requirements for an expert system have been described most notably by Professor Donald Michie (Michie, 1979), who came to the subject via Artificial Intelligence. But such is the newness of the subject that theoretical analysis is somewhat sparse. Instead we must infer the requirements of an expert system by reference to examples. Expert systems have been developed in many areas. Examples in the field of medical diagnosis abound. For example, in deciding which of a range of seven types of rheumatism might be present in a patient, the so-called 'EXPERT' program, given details about the patient, produces diagnoses which are correct 94 per cent of the time (Lindberg, 1980). In horticulture, the diseases of soya beans can be identified 'by computer' with similar success (Michalski, 1980). In general these expert systems require that an agreed set of rules be imple- mented as a computer program. The method of discovering the rules varies. It sometimes involves a team of the expert professionals formulating the rules. Another technique is for an 'information engineer' (the name given to the new breed of expert in expert systems) to make an informed guess. An important point to note is that any such rules must be computer-programmable, that is expressable in a numeric or categorical way. The entire validity of expert systems depends crucially on the capture of the true nature of the procedure of the professional experts. Only then can conversion of the rules into a form suitable for programming be undertaken. The 'rules' for valuation of residential properties are suggested in the standard student textbooks. We are reminded that valuations are normally carried out by using the comparable method (Millington, 1982). The valuer should have regard to the general features of the property, noting particularly the presence of such attributes as central heating, scullery, etc (Lawrence, 1971). An important feature of this method of valuation is the need to keep records of comparable properties, hence the name for the technique. It is sometimes argued (Wood, 1973) that valuation is nothing more than glorified guess- work; that there are no rules. I do not believe that this is the case. Although most practising valuers would have difficulty in formalising the rules by which they operate, there is an underlying rationality. By combining long experience with an intimate knowledge of the local market, and an awareness of national trends, the valuer can employ that most subtle and yet powerful of computers, the human brain, to come up with the right valuation. But because the human brain and the computer operate in such different ways, it might seem that further progress is impossible. A model of what contributes to a valuation doubtless exists inside the mind of the valuer. But the computer cannot (as yet) read this 'brain-map' and translate it into programmable form. To achieve a computer-acceptable model, we need to look anew at the comparable method of valuation. 4. Developing a model of value — the Hedonic Index technique Why do the prices of houses vary? The answer given usually refers to the attributes and location of the property. This can be formalised into a set of specific items which are said to be value-significant. Many valuers make use of simple single variable formulae. Selling price is often estimated from square feet of floor area, by a formula such as: Conall Boyle Page 3 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc P = constant + rate x sq. ft .................... (1) where the 'constant' might represent site value and 'rate' is the 'cost' of each additional square foot. (This is also known as a 'scale' and 'variant' formula). Such simple formulae are of limited use, a fact which can be confirmed with any practising valuer who has tried them. To adequately describe the factors which are significant in determining market price it is necessary to describe and test a model which combines such factors in a mathematical form. The simplest (and therefore most popular) model takes the form: P = bo + b1*X1 + b2*X2 + b3* X3+ ... (2) where P-- price to be explained. .X1, X2, X3, etc are measurable factors, such as age, numbers of bedrooms, square feet of floor area. (Further details on factors will be given later in Section 5 below.) b1, b2, b3, etc are 'rates' for each of the factors. b0, is the constant. For obvious reasons, (2) is called an additive model. We can also have a multiplicative model: P = b0* X1* X2* X3* (3) Here the elements are the same as (2), but combined differently. The reasons for choosing one of these models, or selecting amongst the many other possibilities, is really a matter of Occam's razor. It is best to choose the simplest model which best fits the known facts. The method for testing the validity of the model and establishing the 'rates' for each factor is Multiple Regression Analysis (MRA). As this is a standard statistical technique, it will not be described here. Most advanced textbooks on statistics give an account of MRA (see for example Johnson and Leone, or Imam and Conover, pp. 470-89, which gives an example related to valuations of domestic property). In practical terms, what is needed is a lot of data on 'comparable' property, adequately described, and access to a computer which can run an MRA package. The task is then to convert the data into machine-readable form, and to submit it to the MRA program. The results of the computer run will indicate whether the model is valid, and if so what rates to apply to the various factors. This method of breaking down the market price of a property into elemental prices or 'rates' for its various attributes is called the Hedonic Index technique. It was applied originally by economists (Grillches, 1967) to consumer products such as motor cars and refrigerators. Urban economists, too, have made much use of the technique, using house prices as indicators of values which they are testing. An excellent study of the housing market in Edinburgh has been undertaken (Richardson, 1975). A source of data used was the Sasines register, a public record of all property transactions required by Scottish law. The chief motivation for urban economists undertaking these studies is to test theories of value. In the case of the Edinburgh study, it was shown that height above sea-level is value-significant. Since 1972 there have been dozens of Hedonic price studies on housing. MacLennan (MacLennan, 1977) refers to several. These studies have produced a great deal of insight into which factors are significant in assessing Conall Boyle Page 4 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc market price. Despite its potential usefulness for valuers, the technique appears to have remained the preserve of urban economics (with the possible exception of Byrne and Bowcock, 1974). Building a Hedonic price index for housing is thus far from new, and valuers can be confident that it is a well proven technique. Shenkel (Shenkel, 1978) in his well-presented textbook on Real-Estate Appraisal, devotes an entire chapter to 'Valuation by Statistical Inference'. In effect he shows how to apply the Hedonic index technique to residential property valuations. His objective is not so much to get the computer to carry out valuations. Rather he illustrates how to make use of the results in order to better understand the factors affecting value. His description of MRA is most useful, especially as it is given in terms familiar to valuers (but with an American slant, of course). In the analogous field of rent assessment there have been two recent studies which make use of MRA. In a major study of the activities of the Rent Officer in the West Midlands, Doling ami Davies (1981) examined several hundred actual decisions made on 'fair' rents. They concluded that either the rent officer was acting in a measurably unfair way, or that he had access to information not given in the rent assessments. The application of MRA to rent reviews on commercially-let shops is described by Baum (1982). Turning the results from a hedonic price index study into a computer program to carry out valuations presents little difficulty. By means of an interrogation routine, the user identifies the features of the property. Inverting the model enables the computer to work out a predicted selling price. At this final, crucial stage, it may be necessary to do some fudging of the results. It may be felt essential to include some factor which is generally considered value-significant, even though the MRA does not register it as such. Another problem which frequently arises is lack of statistical significance due to smallness of sample. One way of overcoming this is to combine together results from different subsets which share common characteristics. 5. A case study: developing an expert system for valuation In 1979 I undertook a large scale study of house prices in the Birmingham area; this was repeated in 1982, for a single suburb. (In the second study, the work was ably undertaken by a project student, Mr David Wood.) What follows is a description of these studies. 5.1 Database Data on transactions in the housing market are obtainable from many sources — estate agents, taxation offices, Building Societies. A problem which is common to many sources is confidentiality. The actual price paid is felt to be a personal matter between the purchaser and his financial mentors, and should not be made freely available. Estate agents may be willing to help the genuine researcher, but do not, individually, turn over a sufficient volume of properties for a satisfactory statistical sample. Because of these difficulties, and for the reasons of speed, convenience and cheapness, the source of data used in these studies was house-for-sale advertisements in the local evening newspaper (or estate agents' particulars in the case of the localised study). In the 1979 study details of 519 properties were taken from a single issue of the evening paper. In the 1982 study 170 property details were obtained from estate agents during a single week. In addition to the information given in the advertisements, locational co-ordinates were recorded. Map reading also revealed other local information. Conall Boyle Page 5 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc At first glance advertisements might seem an unlikely source of data for a study with scientific pretensions; but further enquiry reveals that advertisements are inmost respects just as good as the alternatives, and in some ways are markedly superior. To deal with the recurrent criticism first: Should you not use selling price"? Selling price would be ideal, but it has been shown (Byrne and Bowcock, 1974) that asking price and selling price are very closely correlated (index of determination 0.989). So asking price is, as might be expected, a close surrogate of selling price. Do advertisements tell the truth'? Despite being written in the main by those profligates of language, estate agents, advertisements do tell a truthful story. This was established, in part, by means of a random survey among the data. But a moment's reflection demonstrates that falsehoods in advertisements would be self defeating. An advertisement is an invitation to inspect. If a property is described as detached, when it is terraced, it fools no-one. Truth in this case should be self reinforcing. What do advertisers leave out? Negative features are often omitted, and this is undoubtedly a drawback of this source. But some of these undesirable aspects can be inferred in other ways. Using a map, it is possible to spot which houses are situated on a busy main road, near a railway, etc. Are advertisements representative"? In order to form a sound view of the housing market in a given area, it is necessary to obtain information on a representative cross-section of types and locations. By comparing the proportions of types found in our studies with Census figures for the area, it was found that advertisements give a very fair picture of the housing market. (In contrast, Building Society figures showed a very unbalanced picture, with more expensive properties over-represented.) On the positive side, advertisements offer some unique advantages as a source of data: — Since each word has to be paid for, each is chosen with rare care. The fact that some feature is mentioned in an advertisement may be prima facie evidence that it is value-related. — Since it is market price that we seek to explain, a 'snap-shot' of the state of the market is needed. Values change with time, and a collection of deals from different points in time may be inadequate. Only advertisements can give up- to-date at-the-moment picture. (These last points are elaborated by Doling, 1978.) 5.2 Coding the data — house attributes The information given in the advertisements cannot be fed directly to the MRA program. It first has to be coded. In some cases this represents little problem. Age and number of bedrooms are numeric variables already. When confronted with categorical data such as type of house — Semi, Detached, Bungalow, the technique is to code on a 'dummy' variable. This dummy takes on a value of 1 or 0 depending on whether or not the property exhibits this characteristic. So variable X1 in formula (2) could relate to a bungalow. If the property is a Bungalow score 1, otherwise score 0. (There arc some technical problems with both dummy and measured variables, which are explained in Imam and Conover.) The original list of variables related to house attributes were: SEMI, DETACHED, BUNGALOW, TERRACED, FLAT, MAISONETTTE (house types) Conall Boyle Page 6 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc 1BED, 2BED, 3BED, 4BED, 4 + BED(numbers of bedrooms) CENTRAL-HEAT, DBL-GLAZE, GARAGE-SINGLE, GARAGE-DBL (other features) CONDITION (+1: good, -1: bad), AGE (in years) FREEHOLD (1: yes, 2: no), SELLER (agent or private) In the later study, data on square feet of floor area, rateable value, type of roadway, plot size, plot frontage and orientation of the house were also collected and coded. This is a list which most valuers would recognise as being significant in determining market price. It is consistent too, with the previous hedonic index studies. 5.3 Coding the data: location Without doubt location is the single most important factor in determining market value. Experience shows that two otherwise identical houses, separated by a few miles, can vary in price by as much as 50 per cent. Any model must give location an extensive treatment. For the urban economists, the favoured method is to measure some known 'socio-economic status' (SES) variable for the ward or neighbourhood where the property is located. Thus, for example, the rate of unemployment, or the rate of two cars per household, can be used to measure the desirability or otherwise of the location. Another popular measure used to value location is distance from the Central Business District (CBD). Both methods have been widely used with some success. Neither SES or distance to CBD are particularly suited to the task in hand — building a model to value residential property. SES indicators are based on census data for fairly large areas, and can be ten years out-of- date. Distance to CBD may be appropriate for valuing commercial property, but is not nearly so important for houses. This is especially true for the West Midlands conurbation, with its multi-centre shopping and employment patterns. Following the lead given by Richardson in Edinburgh, and Byrne and Bowcock in St Albans, I decided to carry out a Trend Surface Analysis (TSA) to describe the variation of values with location. TSA is a method of determining the shape of a surface — its contours — based on a scatter of spot heights. In mathematical terms, it means fitting a polynomial in x and y (the co-ordinates) to the spot heights (or property values, in this case). The starting variables for TSA are the x and y co-ordinates of each property. From these the computer can calculate the higher orders; x squared, x times y, y squared, and so on up the orders of powers and cross-products. In this study calculations were made up to the fifth order. 5.4 Model polishing The data was analysed by MRA using the SPSS (Statistical Package for Social Scientists) package. This is a package which is widely available for use on large main frame computers. It allows for many recodes and recombinations of variables, as well as simple statistical calculations. Its main interest in relation to the current investigation is the regression analyses which are possible. Micro-computers do not as yet offer such a comprehensive and powerful statistical package, but certainly soon will. Various subsets of the house attribute and location variables were tried out using the MRA program. In addition, different models, additive, multiplicative and hybrids, were tried. The objective in all this was to get a Conall Boyle Page 7 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc model which seems plausible and which explains the facts as fully as possible. In the end the simplest additive model proved to be as good as any of the more complicated models. The criterion for assessing the explaining power of a model in MRA is the multiple coefficient of determination, usually labelled 'R squared'. For the large scale (1979) survey, the results showed an R -squared value of 0.79, with a standard error of £3,723. The value of 0.79 indicates that the model is explaining 79 per cent of the variability in prices. In other words, on the basis of the information given in advertisements it is possible to account for 79 per cent of the variations in price between houses. This leaves a substantial 21 per cent to be explained. Possible reasons for the unexplained variation might be: — desire of sellers for a quick, or slow but profitable, sale. — mis-specification in the advertisements. — quality differences unaccounted for. — the inherent imperfection in the housing market, where buyers and sellers act without total knowledge. — problems with using asking rather than selling price. The standard error of £3,723 indicates that the model can be expected to predict prices to within £2,400 more often than not, or that the maximum likely error we can expect is about £7,500. (These figures are based on the Normal Distribution.) To explain 79 per cent of the variations found in the price of houses, or to achieve an error likely to be in the order of £2,400, is not very impressive. It certainly would not be acceptable to professional valuers. The subject of errors is not one which valuers like to discuss, although an indication of an acceptable level was given in Singer and Friedlander v John D. Wood, where it was stated that errors in excess of 10 per cent either side, exceptionally 15 per cent, would indicate negligence. My own informal probings amongst valuers suggests that an error of less than 5 per cent is to be expected. The computer program described here does not as yet match these levels. In the later (1982) study much better figures were achieved. An R-squared of 0.91 with a standard error of £2,383 was obtained. Although this is still not acceptable, it does represent a major improvement on the first study. Despite the apparent shortcomings of the model in this case, it is worth remembering what has been achieved. The model as described is capable of producing valuations to within a few £l,000's of the true value. This can be done at minimal cost, and in a few minutes. It also covers a very large area — the entire Birmingham Metropolitan District, covering approximately 250,000 properties. No single human valuer could hope to match this breadth of coverage. 5.5 Testing the valuation program The rates for each of the variables used in the MRA were calculated by the computer. A full list is given in Table 1. In statistical language these 'rates' are called the regression coefficients, or B values. Calling them 'rates' is really an interpretation of the nature of these values, but one which is justified. An encouraging feature of these rates was their unsurprising magnitude. A value of £2,097 for garage, or £800 for freehold over leasehold is very much in line with common-sense, or the consensus view of valuers. The results for the Trend Surface Analysis are shown as a price-contour map in Figure 3 (see page 282). Again there are no surprises — the theoretical Conall Boyle Page 8 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc map corresponds with the accepted 'good' and 'less-desirable' areas. It is always encouraging when elaborate statistical techniques yield such results. Using the results of the MRA, an interrogation program was devised. The dialogue in Figure 1 shows part of the actual interaction between user and computer. When all the values have been entered, the computer calculates the predicted selling price by applying the rates for each factor which were found using the MRA. Testing the program took the form of trying out actual properties, taken at random from whatever source came to hand. This informal and strictly ad hoc procedure confirmed the theoretical analysis. About half the results were quite close; the rest gave an unacceptably high error. 5.6 Multi-collinearity One of the most serious problems that can arise when using MRA is known as multi-collinearity. Although the textbook explanations are somewhat intimidating (see for example Huang p. 149) in essence it is quite understandable. Multi-collinearity poses the question: are we really measuring what we think we are measuring? For example, suppose that 'Bungalow' yields a low valuation coefficient. It could be that most of the bungalows in our data set were on an unattractive estate; hence 'bungalow' is measuring the environment, not the type of building. Multi-collinearity can be detected using the simple two-way correlation coefficients between each of the explanatory variables. The computer will readily provide these — but a practical problem is the amount of data generated (n squared divided by two, where n variables per property are used). In the case of the present study, about 200 of these two-way coefficients were required. Huang suggests that so long as: rjj < R ie the two-way coefficients are less than the overall correlation coefficient, then the multi-collinearity is not too serious. In the study described in this paper one group of variables exhibited an unacceptable degree of multi-collinearity. Eight (out of a total of 180) variable pairs in the Trend-Surface Analysis of Figure 3 showed a value in excess of R. The main culprits were combinations of Y, the North-South distances from datum. It is not altogether surprising that Y, Y squared and Y cubed should be correlated. The variables concerned with house attributes were all well within the limit, although some showed higher values than others. For example the pair Central heating with Garage had a value of 0.29 compared to the overall R of 0.89. This could indicate that houses with central heating usually have a garage. The rate for one or other may thus be too high or too low. How disastrous is multi-collinearity? Its effect is three-fold: — it gives better R values than the data warrants. — it confuses variables, making selection of important variables more difficult. — errors in the 'rates' of Table 1 could be greater than those given by the MRA. Conall Boyle Page 9 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc Conall Boyle Page 10 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc ------------------------------------------------------------------------------------------- Table 1 Rates for the variables from MRA From the 519 property details, average values of 'rates' for each of the variables are produced, using the Multiple Regression Analysis computer package. F value and *** indicate the importance of each variable. Property Variables £ Rate' F value' * Detached £7,712 179 *** Bungalow £9,287 70 *** Type Terraced -£3,180 37 *** Flat £1,530 1.8 * Maisonette -£2,264 2.6 ** Age (yrs) -£28.4 1.9 * Two £4,859 6.6 *** No. of Three £6,377 11 *** bedrooms Four £2,595 37 *** Over Four £13,623 36 *** Reception £546 2.4 ** Condition £930 7.6 *** Garage £2,097 101 *** Central Heating £1,935 24 *** Double Glaze £767 0.7 NS Seller -£207 0.1 NS Freehold £800 4.8 *** Other £1,875 27 *** n = 519 R 2 = 0.79 SE = £3,723 NS — not significant * significant ** very significant *** highly significant SE — standard error Thus it seems that because of multi-collinearity all MRA results are suspect. It certainly warns us to exercise caution. However, in the field of valuation all results can be subject to the test of both common sense and the market. So long as MRA gives useable results, it serves its purpose. 6. Application of the expert system for valuation What has been described in this paper is essentially an 'academic' exercise, but with a practical end in view. As more valuation offices acquire micro- computers, then the opportunity for individual valuers to try out these techniques will arise. The first stage will probably be the computerisation of the 'comparable' records. The widespread availability of database programs makes this a relatively straightforward task. On its own this development promises great advances. One example is in the sorting of comparable data. Conall Boyle Page 11 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc A list in price order can prove extremely useful for valuations. But merely replicating existing filing systems using a computer is exploiting a very small part of the potential. Developing an expert system for valuation need not be the preserve of academics. Any valuer possessed of a computer and a logical turn of mind should be able to carry out the procedures outlined above. As more offices use micro-computers for routine storage of records, the biggest hurdle to the use of such techniques will be overcome. There will exist a large database of property transactions in machine readable form. From then on, the analysis presents no practical or time consuming difficulties. What could a practicing valuer expect to gain from developing his own expert valuation system? An inestimable benefit will be the insight and objective knowledge he will gain from such a radical and novel analysis of his own data. To give but one small example from the MRA: a common belief is that ex-municipal properties command a lower price. MRA showed this not to be the case. Ex- municipal houses do not need a discount on the comparable non-municipal house value to make them saleable. But the chief motivation for establishing an expert valuation system is, of course, to carry out valuations. As was seen from the example above, the results did not match up to the standard required. But even with this level of achievement, computerised valuations are worthwhile for a number of reasons: — the computer can assist the valuer by giving a second opinion, or perhaps by giving a starting value. The results can set limits within which a valuation should fall. Any deviation from these limits could indicate a need for closer attention. — in many situations, eg Building Society valuations, the need is not so much to estimate selling price, but to ensure that a minimum secure value exists. If the amount to be borrowed was less than (say) 60 per cent, then loans could be on the basis of computerised valuations alone (subject only to a check that the details given were correct — this would not need the skill of a valuer). The benefits to the Building Society and borrower are not only saving costs. Speed of decision can be crucial in obtaining the business; this was one of the main selling points in favour of the Banks during the mortgage glut of 1980-82. — the expert system can act as a type of 'policeman', ensuring compliance with the correct procedures. This effect has been observed in medical and other expert systems. The fact that data has to be fed into a computer seems to sharpen the wits. It has been observed that death-rates declined with the introduction of computer diagnosis. Valuation offices, too, can expect an improvement in performance by using computerised valuations. — professional negligence, especially following the Yianni case, is a matter of major concern to valuers. The fact that a computerized valuation was undertaken, could, in disputed cases, prove to be a very effective secondary defence. The print-out also has the merit of being tangible evidence, unlike the brain activity of the valuer, however competent he may have been. --consistency is guaranteed. Since the computer will always apply the same rules, it gives an identical answer given the same data. This corresponds to the notion of 'fairness', often important in valuations for compulsory purchase or sale of council houses. Conall Boyle Page 12 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc 7. Future developments As more valuers gain experience in using computerised valuations, the accuracy attained will undoubtedly improve. Indeed there might even be a danger that as more valuers turn to computerised valuations the results become a self-fulfilling prophecy. This is most unlikely, however. As different valuers develop varying models of price structure, so a continuing difference of opinion will persist. It also must be remembered that all valuations are subjected to the ultimate test in the market place. One organisation from which a standard method of computerised valuations may emerge is the Inland Revenue valuation office. There is evidence to suggest that, in connection with capital value rating, the Inland Revenue have undertaken a number of studies, similar to the one described in this paper. With the resources at their disposal, it is to be hoped that they can develop a methodology which can givesufficient accuracy. It is also to be hoped that they publish at least the principles of any capital value rating method they adopt. There are many technical problems still to be overcome in this new and exciting area of information technology. To mention some: — What is the most effective model which adequately describes the variation of price with area? — MRA is a statistical technique, which assumes a one-off test. But with computerised valuations there is a need to update the rates for variables. How can the previous results be combined with the latest values? — what factors not included here are value significant? Plot size and frontage certainly have an effect, but how much? Is it possible to objectively rate the immediate environment of a property, including factors such as quality of adjacent houses, type of roadway, view? What effect has nearness of facilities such as shops, schools, pubs on value? — if many valuers are going to undertake this sort of analysis, their endeavours will be greatly helped by a common coding scheme. The obvious candidate for such a profession-wide standard would be the classifications used by the Inland Revenue. But these are not on general release; in any event they were developed before the computer age and are probably inappropriate. The need for a common classification scheme still remains a matter of great urgency. I look forward to a fruitful combined effort between academics, both in valuation and statistics, and valuation practitioners, in solving these and other problems. Whoever undertakes the research, the aim will remain the same: to produce an expert system capable of effectively carrying out valuations. References Baum, Andrew 'Extrapolating comparable evidence', Estates Gazette, 263: 23, 3rd July, 1983 Byrne, P. J. and Bowcock, P. 'The structure of residential property values in St Albans', University of Reading (1973) Doling, John 'The use of content analysis in identifying the determinants of house prices',Urban Studies, Vol. 15, pp. 89-90 (1978) Doling, John and Davies, Mary 'Fair rents and capital values', Estates Gazette, 260: 677,14th November, 1981 Conall Boyle Page 13 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc Grillches, Zvi 'Hedonic price indices revisited: some notes on the state of the art', Proc. Amer. Stat. Assoc. (1967) Huang, David S. Regression and Econometric Methods, Wiley, New York (1970) Imam, R. L. and Conover, W. J. Modern Business Statistics, Wiley, London (1983) Johnson, N. L. and Leone, F. C. Statistics and Experimental Design, Vol. 2, Wiley, London(1977) Lansdown, J. 'Expert systems: a memorandum on their scope and use', BOCAAD (Bull.Comput. Aided Architectural Des.) (GB) Vol. 38, pp. 25-36 (1980) Lawrence, D. M., Rees, W. H. and Britton, W. 'Modern methods of Valuation of Land, Housesand Buildings', Estates Gazette, London (1971) Lindberg, D. A. B. et al 'Computer based rheumatology consultant', Proc. 3rd World Conf. onMedical Informatics, Tokyo (1980), North-Holland, Amsterdam (1980) MacLennan, D. 'Some thoughts on the nature and pupose of house price studies', Urban Studies,Vol. 14, pp. 59-71 (1977) Michalski, R. S. and Chilausky, R. L. 'Learning by being told and learning from examples: anexperimental comparison of the two methods of knowledge acquisition in the context of developing an expert system for soyabean disease diagnosis', Int. J, Policy Anal, and Inf. Syst. (USA), Vol. 4, No. 2, pp. 125-61 (1980) Michie, Donald (ed) Expert Systems in the Micro-electronic Age, Edinburgh Univ. Press,Edinburgh (1979) Millington, A. F. 'An Introduction to Property Valuation', 2nd ed., Estates Gazette, London(1982) Richardson, H. W., Vipond, J. and Furbey, R. A. Housing and Urban Spatial Structure: a CaseStudy, Saxon House, Farnborough (1975) Shenkel, William Modem Methods of Real-Estate Appraisal, McGraw-Hill, New York (1978) Wood, E. A. 'Current developments in property valuation and appraisal', Seminar at Liverpool Polytechnic, 7th February, 1973, reported in The Valuer General references: valuation and MRA Ball, Michael J. 'Recent empirical work on the determinants of house prices', Urban Studies,Vol. 10, pp. 213-33 (1973) Blettner, Robert A. 'Mass appraisals via multiple regression analysis', The Appraisal Journal,pp. 513-21 (October, 1969) Bowcock, Philip 'Capital values for dwelling houses', Rating and Valuation, pp. 173-75(June, 1981) Chudleigh, Walter M. 'The application of correlation matrix analyses to real estate appraisal',The Appraisal Journal, pp. 523-30 (October, 1979) Griffin, Gerald 'Conventional appraisal techniques can be computerized', The Appraisal Journal,pp. 253-62 (April, 1979) Kamath, Ravindra R. and Yantek, Kenneth R. 'Linear multiple regression analysis applied to valuation of single family homes', The Real Estate Appraiser and Analyst, pp. 36-41(September-October, 1979) Lessinger, Jack 'Econometrics and appraisal', The Appraisal Journal, pp. 501-12 (October, 1969) Lessinger, Jack 'A"final" word on multiple regression and appraisal', The Appraisal Journal,pp. 449-61 (July, 1972) Conall Boyle Page 14 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc Smith, David V. 'An appraiser looks at multiple regression', The Appraisal Journal, pp. 248-52(April, 1979) Conall Boyle Page 15 09/06/2005 C:\3CwmCadno\MSWord\Boyle J of VALUATION 1984.doc 
work_ztiq4knfm5ghjjcotevkkrlhga	----	Untitled Kent Academic Repository Full text document (pdf) Copyright & reuse Content in the Kent Academic Repository is made available for research purposes. Unless otherwise stated all content is protected by copyright and in the absence of an open licence (eg Creative Commons), permissions for further reuse of content should be sought from the publisher, author or other copyright holder. Versions of research The version in the Kent Academic Repository may differ from the final published version. Users are advised to check http://kar.kent.ac.uk for the status of the paper. Users should always cite the published version of record. Enquiries For any further enquiries regarding the licence status of this document, please contact: researchsupport@kent.ac.uk If you believe this document infringes copyright then please contact the KAR admin team with the take-down information provided at http://kar.kent.ac.uk/contact.html Citation for published version Basden, Andrew and Ball, E. and Chadwick, David W. (2001) Knowledge Issues Raised in Modelling Trust in a Public Key Infrastructure. Expert Systems, 18 (5). pp. 233-249. ISSN 0266-4720. DOI https://doi.org/10.1111/1468-0394.00178 Link to record in KAR https://kar.kent.ac.uk/13528/ Document Version UNSPECIFIED Knowledge Issues Raised in Modelling Trust in a Public Key Infrastructure A. Basden, E.Ball, D.W.Chadwick, ISI, University of Salford. ABSTRACT The paper describes a knowledge based system (KBS) for modelling trust in the Certification Authority (CA) of a Public Key Infrastructure (PKI). It was built using a graphical KBS toolkit, Istar, that allows the knowledge builder to easily model the important relationships between concepts of the domain. The knowledge base was initially built using published work and was subsequently extended by knowledge obtained from leading PKI experts. The first prototype system computes the trust in a CA by asking the user a series of questions about the CA's Certification Practice Statement. Examples of its use with two well known public CAs is discussed. An important issue raised and discussed in this paper is how to map symbols in the KB to the knowledge level of human trust and beliefs, for such an ill-defined area of knowledge as trust, and four main mappings have been identified. Another issue that emerged relates to the use of questionnaires during knowledge acquisition. The expert system is currently available online via the Istar Knowledge Server, and future work is discussed. 1. INTRODUCTION 1.1 The Problem When compared with face to face communication, the Internet suffers from many disadvantages. We can easily recognise people in face to face communications, but on the Internet it is very easy to masquerade. During face to face communications we can observe subtle gestures and body language, that help us to determine the authenticity of the speaker. Neither of these are currently possible on the Internet. The advent of the Internet thus raises important issues of authenticity and trust. One attempt to address these issues involves the use of a Public Key Infrastructure (PKI) (Adams and Lloyd, 1999), which is a system for the identification of people (and systems) based on the use of a trusted third party (TTP) called a Certification Authority (CA). This CA is responsible for verifying the identities of people and providing them with certified public keys that contain their identity and that of the CA. Each certified public key, or certificate, is digitally signed by the CA, making them tamperproof and easy to distribute. A recipient can be assured they have a valid copy of a subject's public key, by checking the digital signature on the certificate (providing of course they have a trusted copy of the CA's public key). The subject's validated public key may then be used to verify messages digitally signed by their corresponding private key. It is absolutely central to the working of the PKI that the CA carries out rigorous procedures for the checking of identities, for the generation, storage and destruction of its private keys, and maintains the safety and security of its systems, data and software. Weakness in any of these areas may allow fake certificates to be created (by fake we mean that the private key corresponding to the certified public key is not held by the subject identified in the certificate). There are now many CAs in operation and it is a fact that they do not all operate to the same standards with regard to the above procedures and other essential criteria. Consequently some CAs will be inherently more trustworthy than others, and the reliability of the identity to public key binding, implied by their certificates, will differ from CA to CA. The procedures that a CA claims that it carries out are described in its Certification Practice Statement (CPS) and/or its Certificate Policy (CP). The difference betweek these two documents is still the subject of some debate and confusion, as they arose from different working groups. Certificate Policy is defined in the ITU-T X.509 standard (ISO/ITU-T, 1997) as "The named set of rules that indicates the applicability of a certificate to a particular community and/or class of applications", whilst Certification Practice Statement is defined by the American Bar Association in its Digital Signature Guidelines (Information Security Committee, 1996) as "A statement of the practices which a certification authority employs in issuing certificates". Clearly the scope of the two documents overlaps, and some CAs produce CPs, some produce CPSs and some both. A template for the production of CPs and CPSs is described in Chokhani and Ford (1999). These documents should be published by the CA, and be readily available to those (hereafter called 'relying parties') who wish to assess the authenticity of the sender of a digitally signed communication. By analysing these documents a relying party (RP) should be able to form a judgement about the trustworthiness of the CA. 1.2 An Expert System Solution However, such analysis and judgement requires expertise, which most relying parties cannot be expected to possess. Not only are there dozens of factors that must be taken into account in such an assessment, but most relying parties do not have enough expertise to know whether the methods described in the CPS provide sufficient authentication. If such expertise can be embedded in an expert system, then any relying party should be able to benefit from it. Our objective, therefore, was to encapsulate such expert knowledge in an expert system, and make it available on the World Wide Web, so that any relying party that uses it could gain an indication of the trustworthiness of a CA by answering a series of questions about the CPS/CP published by that CA. On the basis of this, a recipient of an Internet communication can then determine how much trust to place in the stated identity of the sender. Here we describe an expert system that is able to provide a measure of the trustworthiness of a CA. The output of the expert system is a trust quotient, in the range from zero to one. A value of one means that the CA would be believed by the experts to be totally trustworthy, whilst a value of zero means that the CA would be believed by the experts to be completely untrustworthy. Intermediate values indicate the relative trustworthiness of CAs, but are not an absolute measurement of trust. We are not aware of such an absolute metric. Previous authors have devised trust metrics, based for example, on the number of positive experiences with the person (Beth, Borcherding and Klein, 1994), the length of the certification path (Tarah and Huitema, 1992), and how much you trust a key holder to act as a trusted introducer (Network Associates, 1999). Reiter and Stubblebine (1997) found none of these methods to be fully satisfactory, and so proposed a metric based on the monetary insurance value that a relying party could recover if the name to key binding proved to be incorrect. In fact, Chokani and Ford suggest that liability is just one of many factors that should be documented in a CP/CPS. We give reasons below why we believe that an absolute metric cannot exist. For this reason, the ability that expert system technology provides to explore its knowledge is considered vital in such an application. In this paper we describe the expert system and the issues we had to address as we developed it. Two main issues were encountered. One is that the mapping between the knowledge level of human trust and beliefs to the symbol level of knowledge representation is not always straightforward. Yet we found ways to clarify the steps involved and discovered four main types of mapping that seem important. The second relates to the use of questionnaires during knowledge acquisition. 1.3. What is Trust? There has long been a difficulty in establishing a clear definition of the concept of trust. Personality psychologists, social psychologists and economists all have different perceptions of the meaning of trust. A very clear exposition of the study of trust and its different meanings is given in (Battacharya, Devinney and Pillutla, 1998). Briefly, personality psychologists view trust as a personality trait, social psychologists view trust as an expectation about the behaviour of others, whilst economists focus on the costs and benefits of exercising trust in business relationships. Our expert system most nearly corresponds to the social psychologists' viewpoint, in that the trust quotient computed by it provides the user with an expectation or belief about the behaviour of the CA, when it is producing public key certificates. The trust quotient is thus a measure of the quality of the procedures and policy described in a CP/CPS, and thus might be taken to be a measure of the authenticity of a digitally signed message received by a user. Each user will then interpret the trust quotient differently, depending upon his predisposition to trust others (i.e. his personality trait), and the importance of the message that he has received (the economic factors). We do note however in Section 6, that the trust quotient produced will in part reflect the personality trait of the user, since many of the questions posed by the expert system require judgement. For example, if our expert system produces a trust quotient of 0.7 for a given CA, a trusting user might view this value as sufficient for his purposes, and accept the digitally signed message as authentic, whereas a less trusting user might view the digitally signed message with some scepticism. (But note that the less trusting user may gain a score of 0.6 for the same CA due to the way he has judged the adequacey of the CA's procedures.) Alternatively, the same user might trust two messages from the same recipient differently, if one is for a large transaction and the other is for a small one. Finally, the same user might trust the same transaction message from the same recipient differently, if it is the first time such a message arrived, or if it is the nth time he has seen such a message. Thus our expert system is designed to be a tool to help the user in his trust decision making, rather than being a tool designed to replace the user's decision making by saying categorically that messages should be trusted or not. Relating this to e-commerce, in order to carry out an e-commerce transaction, a user needs to be both authenticated and authorised. Authentication provides assurance that the user is who he says he is, whilst authorisation says that this particular user is allowed to carry out the expected tasks. Our expert system provides some measure of the trustworthiness of the authentication of the user, but says nothing about the authorisation of the user. Even within the scope of authentication, a further issue is raised: the CPS and CP might not accurately reflect what the CA actually does in practice. Sometimes the CA performs better, sometimes worse. For example, it might be stated that lists of revoked certificates will be updated every 24 hours, whereas slack procedures might lengthen this to several days. In this case, the relying party who waits 24 hours before checking the lists to ensure the sender's certificate is still valid, might be at risk without knowing it. The expert system described here only takes account of the published CP/CPS, and not the actual working practices of the CA. But we have recognised this problem and have developed the system further so that it can assess actual, rather than stated, practice. A description of this can be found in (Ball, Chadwick and Basden, in submission). 2. OVERVIEW OF ISTAR The KBS software we adapted for use as a knowledge server was Istar (Basden and Brown, 1996), first developed during the INCA project (Basden, Brown, Tetlow and Hibberd, 1996) which aimed at intelligent authoring of construction contracts. Istar was designed as a flexible visual programming language, employing a 'language' of boxes (nodes) and arrows (arcs) with which the knowledge engineer draws knowledge as a graph rather than writing it as production rules or predicates. This approach is particularly useful for building KBs in ill structured domains of knowledge (Basden and Hibberd, 1996). In the INCA project this enabled the development of a substantial KB was produced which created construction contracts for the user according to the needs of the parties and their specific situation (Hibberd and Basden, 1995). Istar's inference engine is based on the concept of the inference session with the user, during which is repeats a cycle of backward chaining to find questions to ask, seeking the answer to these question, and forward chaining to propagate these answers. See Fig. 1. Istar Inference Engine Fig. 1. The Istar Inference Engine The modular architecture of Istar's engine has allowed it to be extended to seek answers to the questions from across the Internet as well as from local users. This can be in either client or server mode. In client mode, Istar seeks information from servers such as web pages or, as discussed towards the end of this paper, from our Trust Check Server. In server mode, Istar sends dynamically generated HTML pages to a client and awaits their answers. Istar's server mode is described in Basden (2000). Istar has a multi-threaded capability and so can run in all three modes at one time, over several concurrent knowledge bases. It is this that makes it an ideal vehicle for hosting a trust assessment KBS, in that the expertise in the KBS can be made available worldwide via the Internet, accessible by any number of users who have browsers. The internal structure of the Istar knowledge base is similar to a semantic net, in which items are linked by relationships in a flexible manner. Attributes can also be linked, by inference relationships, and each such attribute has a value type (Boolean, Probability, Bayesian, Integer, Float, String, etc.) and an inference method (such as AND, OR, Probabilistic And, Probabilistic Or, Maximum, Minimum, Bayesian Accumulation of Evidence, etc.). The range of legal values that an attribute can take depends on their type, and most can also take the special value 'Unknown'. The whole KB is composed of these attributes and the relationships between them. Inference relationships form a directed acyclic graph and Istar prevents any loops being formed therein. One or more attributes are chosen to be the goals in an inference session, and attributes with no antecedents are those selected by the backward chaining process to be asked as questions. The ordering in which questions are asked is determined by the backward chaining process, which suppresses those that are irrelevant, and by special inference methods. One of these is FirstKnown, which will take the first antecedent that is given a Known value and is used to allow the user to answer "I don't know; try to work it out by other means." The latter is employed in the trust assessment KBS to allow the user the option of overriding the KBS's own calculations with their own judgements in specific circumstances. Istar provides a number of mechanisms to give help and explanation to the user. One is a set of help texts that is available with each question, and are used variously to explain what the concept behind the question is, how to answer it, what its critical values are, and also which sections of Chokani and Ford, if any, refer to it, so that the user has access to the source material if they wish. The other help mechanism is that the knowledge base can be explored. In local mode, powerful facilities exist to display connected portions of the KB, such as a nodes complete antecedent network, and to search for items of interest. In server mode, in which the information has to be transmitted via web pages, exploration is textual, step by step, rather than visual: the user can click on the goals to find their antecedent attributes, then on these to find their antecedents in turn, and so on. See Fig. 2 for an example of such pages. Fig. 2. Example of Simple Exploration of Knowledge The user can engage in 'what-iffing', in which they supply different answers to questions to see what happens to the final results. In server mode, this is accomplished by means of a 'Reset' button during exploration, after which the backward chaining process automatically asks that question afresh if it is appropriate to do so. 3. OVERVIEW OF THE TRUST MODEL The overall goal of the trust model is to calculate a trust quotient, a numeric measure that indicates to what extent experts would consider that it is likely that a CA can be relied upon to provide good authentication of a user and produce a trusted identity to public key binding. The model has been implemented as an Istar knowledge base, shown in inference net form in Fig. 3 (an actual screen grab from Istar). In this diagram, boxes express items or variables of various kinds, and most have labels that identify their meaning, and links between represent inference relationships (solid links) or grouping (dashed links). There are also small dark boxes; these are ancillary variables that have little knowledge level meaning (see below) and are in the KB merely to aid in the necessary calculations. The reader is not expected to read the writing in Fig. x1, which is, of necessity, small; explanations will be given below where necessary. (new version to be procured) Fig. 3. The Trust Assessment Knowledge Base (new vsn needed) In Fig 3, inference flows from left to right. The main goal is a single item, 'Can Trust', with an auxiliary inverse goal, 'Can't Trust', shown at the right hand side. The input questions are ranged on the left half of the diagram, and are those that either have no antecedents or are grouped (dashed links). We have taken as a starting point for building the expert system the work of Chokani and Ford (1999). This describes the procedures that a CA must carry out and also a standard template for the description of these procedures in a CPS or CP. From here seven areas have been identified as being important to the trustworthiness of a CA. 1. Identification 2. Legal 3. Obligations 4. Procedures 5. Cryptography 6. Malpractice 7. Audit The nodes for these are those major ones that link into the 'Trust' node, ignoring any of the ancillary ones. Briefly, identification is concerned with the process of identifying the subject who has applied for a public key certificate. The legal section is concerned with the various legal provisions between the CA and the relying party, such as warranties, disclaimers and indemnities. Obligations describes the obligations of the various parties, one to another, for example, the subject is obliged to keep his private key secure so that no-one else can use it (otherwise masquerade would ensue). Procedures concerns the various internal procedures that a CA must perform, such as key management and staff training and recruitment. Cryptography is concerned with the strength and quality of the cryptographic functions being used by the CA, the implication being that if the cryptography is weak, forged certificates can be produced. Malpractice is concerned with the controls that are in place in order to stop wrong doing from occurring, such as physical and network security controls, audit trails and archives. Finally the audit section enquires about the external audit that the CA subjects itself to. These seven factors are, themselves, derived from others, placed somewhat to their left in Fig. 3, which are, in turn, derived from yet others, and so on until we reach questions that can be put to the user. These are the questions that are put to the user during a run in which a given CA is being evaluated, and the user is expected to answer them by reference to the text of the CPS or from other sources of information, such as phoning the CA itself or from informal information or rumours. Most links have weightings. Those towards the output (right hand) side - representing the seven main factors to CA Trustworthiness - are the most important and interviews with experts have been carried out to obtain them. Weightings of links to the left of these seven main factors are less important because their effect is moderated by the flow of inference towards the goal. Most of the input side weightings have been set by expertise from within the research team or by asking a small number of recognised experts. Apart from ancillary nodes, the variables each have some conceptual meaning in the realm of authentication trust. This enables us to link beliefs about this realm with variables that Istar can manipulate to yield a final trust quotient. What the form of this relationship is, and the diverse methods that have been employed in the KB to effect such manipulation, is now discussed. 4. THE VARIABLES AND INFERENCE METHODS Istar offers a number of types of variable and, for each, a number of standard inference methods. The main types employed in the trust KB are Booleans, Probabilities, Bayesians and Floats. This section explains and discusses the choice of types of variable and of inference methods, in terms of how concepts related to trust must be combined. 4.1 The Precise Use of Variables Loosely speaking, when the user engages in a session with the trust KB, they enter their beliefs about various factors they are asked about. These beliefs are processed by the KB and then a result is given which is a belief about the trustworthiness of the CA. However, this view, which talks only about beliefs, leads to confusion. We must be more precise. Under Newell and Simon's (1976) Physical Symbol Systems Hypothesis a computer is capable of intelligent action as long as it can process symbols. However, Newell (1982) extended this with the publication of his paper, The Knowledge Level, which proposed that what the symbols referred to is a realm or 'level' that is distinct from that of symbols. While symbolic computation is carried out at the symbol level, human thinking has an important knowledge level meaning, with which the symbols refer to something. We must be precise in understanding and maintaining the distinction between the symbol and what it refers to, lest confusion occurs. In the case of the trust KB most of its variables stand for a human belief about a concept relevant to trust. When the user engages with the KBS to evaluate a given CPS, several steps are undertaken, some at the knowledge level, some at the symbol level, and some translating between the two levels: Step 1, SL: During the run, the user is asked questions, which take the form of a short text displayed on the screen. The text itself is a symbol, and is chosen by the action of the symbol level algorithms of the inference engine and the symbol level structures in the KB, by mechanisms described briefly in an earlier section. Step 2, SL-KL: The user reads the text and translates it into knowledge level meaning by the normal processes of linguistic interpretation. It is, in most cases, a request for information. Step 3, KL: The user, working at the knowledge level, obtains the information they believe has been requested. This can be by reading the CPS, and in some cases the relevant information is clearly stated therein, while in others some interpretation of the meaning of the CPS is required and perhaps even a phone call must be made to the CA to obtain more detailed information or interpretations. Step 4, KL-SL: The user decides how the information they have obtained might best be represented in the symbolic form expected by the KBS. In the case of a Boolean tick box, this might be a relatively simple decision, but in the case of a number being expected, the user must set a slider, and must therefore decide where to set it. Sliders are used, where the user is making their own judgement about the quality of, for example, the CA's approach to disclaimers. If the user considers that the quality to be very high, then they must decide whether to place the slider at 95, 90, 80, or whatever they believe to be appropriate. We discuss this issue in the subsection 'User Input Interpretation'. Step 5, SL: The KBS, having received a number or a truth value now propagates it throughout the KB, and seeks the next question to ask. If one it found, then it iterates back to step 1 above. If not, it continues with the remaining steps below. Step 6, SL: Once all relevant and necessary questions have been asked and answered, a result is available, which is a number between 0 and 1. The KBS displays this as a slider or as a number scaled up to be in the range 0 to 100. Step 7, SL-KL: The user reads this number and interprets it as a belief in the trustworthiness of the CPS. Step 8, KL: The user decides what to do about this degree of belief. In most cases where the KBS is used as a prelude to action, some commitment to appropriate action is made. What such actions might be is discussed in the section 'Uses of the KBS'. But in some cases the commitment is deferred because the result from the KBS is not conclusive, neither for trust nor for distrust. In this case, the user employs Istar's facilities to explore the knowledge and the reasons it has come to the result it has given. This is discussed in the section 'Interpretation of Result'. It is clear that an awareness of all parts of this process must guide all parts of the design of the KBS, from the selection of variables in the KB, through the choice of inference methods, to the text to be placed on the screen when questions are asked or results given. Which types of variable might be chosen is dictated by what Istar offers and by what is easy for the user to use. 4.2 User Input Interpretation While the option existed to require the user to enter precise numbers or text, it was decided that it would be more appropriate, as well as easier to use, if they were required to merely click tick boxes or drag sliders. The use of sliders is not only more convenient (with a mouse) but more appropriate than entering numerical digits because the element of judgement required of the user in undertaking steps 2 to 4 above makes high precision numerical input unnecessary and can even be misleading to the user. With a tick box the user has merely to decide whether the proposition contained in the question is true or false. In most cases the interpretation required is straightforward. Where it is not, such as when the information is not available or if it is not clear from the available information sources, the user has the option of entering 'Unknown' as their answer. With a slider, which is used to express either some degree of belief or some degree of quality, the interpretation is more complex. Degrees of quality, such as embodied in the question "To what extent does the CPS deal with disclaimers and exclusions adequately?" invite a judgement by its very wording, and the user will normally treat the slider as though it were a continuous Likert scale rather than a precise number. In the case of a degree of belief that a state pertains, an event has happened or an event will happen, however, the situation is more complex. These cases are akin to probabilities, but human beings are notoriously poor at handling numerical probabilities, especially concerning rare events or states. Typically, if a user intends to convey that a probability is low, they will enter a figure of 0.1 or 0.05. But if the average probability is lower than this, then the effect of the figure they have entered will be the very opposite of what they intended, namely to increase the probability above that of the average rather than to decrease it. For this reason, it has been commonplace for some years in KBS technology to move away from asking the user to give numerical probabilities. Instead, they are asked to give a certainty factor, perhaps ranging from -5 through 0 to +5. This is translated into a number for internal use via a mapping such that -5 maps to 0, 0 maps to the average or a priori probability, and +5 maps to 1. In Istar this mapping is by a piecewise linear curve, shown in Fig. 4. Cert Factor to Prob Fig. 4. Mapping of User's Input Quantity to Internal Number. 4.3 Interpretation of Result As an initial result, in step 6, the user is given a single trust quotient, a quantity which is intended to indicate a degree of trustworthiness ranging from none to total. Internally, this is a number ranging from 0 to 1, but it is presented to the user in three alternative ways, depending on the mode in which Istar is used: as a number ranging from 0 to 100, as a visual symbol in a diagram of a line of variable length, or as a visual symbol in text of a line of asterisks. As discussed above, the result is of calculations made largely under the social perspective on trust, and does not take into account the personality or economic perspectives. In the extreme cases, of very low or very high quotient, then the interpretation - steps 7 and 8 above - is usually clear: either commitment to accept or commitment to reject. In less extreme cases, interpretation of the result will depend on the application of the other two types of trust. If the user is a generally trusting person and the outcome of misplaced trust is not critical, then a degree of trust as low as 0.5 might be deemed satisfactory. On the other hand, if the user is of a generally paranoid type, or the outcome of misplaced trust would be highly damaging, such as in some military applications, then a trust quotient of even 0.98 might be deemed too low. There is often a range in which the quotient will lie between those on which the user makes a commitment of trust or distrust. When the result lies in this range, the interpretation made by the user can be more complex, involving further exploration of the basis for the result. When accessed via a web browser Istar allows the user to work backwards from the goal variable, one step of inference at a time, listing at each step those variables that have contributed to the step of inference. When the inference method of any variable is something akin to AND or OR, then there might be just one factor that has given, respectively, a low or high value to the variable. Such factors can be explored further, as dictated by the user's precise requirements or concerns about trust. The offending factors might be ones that the user reckons are unimportant because of special characteristics of their situation, or they might be factors about which the user has some control. In such cases, the user can, as part of their interpretation, make a decision to ignore these factors and commit to trust. Conversely, the user might find, during their exploration that an apparently high quotient is not valid in their particular situation. In addition to such interpretation, exploring the KBS in this way can indicate areas in which managerial action might be required; in this way the KBS can become a decision support tool in addition to being a mere evaluator of trust. Such uses are discussed further in 'Uses of the Trust Assessment KBS'. Thus, we find that interpretation of the result, that takes place at step 8 above, is seldom a simple process of deciding whether the quotient exceeds some limit or not. For this reason, while use could be made of this KBS in the role of an intelligent agent to filter out untrustworthy Internet messages, its main use is likely to be more in the realms of support of human interpretation by adding the value of expert knowledge. 4.4 Internal Variables, Inference Methods and Weights Between the KL-SL translation involved in user input interpretation and the SL-KL translation involved in interpretation of result, the KBS undertakes considerable SL processing of variables. As can be seen from the description of the model above, there are many internal variables, each with its own inference method and ability to hold a value. Just as the input and result variables all carry some human meaning that the user must add by the process of interpretation in steps 2, 4 and 7 above, so most internal variables also carry some human meaning. Their meaning was assigned by the designer of the KB, and their values can be interpreted as has been described above. Such variables are presented to the user during exploration of the result. Because of these KL meanings, each internal variable must have an inference method suited to that meaning. In this section we explain and discuss the types of inference method we have found it important to use. An inference method is an algorithm that undertakes some calculation and the resulting value is placed in its consequent variable. The calculation is performed on the basis of the values in a number of antecedent variables. The inference method for each consequent variable is chosen by the knowledge engineer during the KB design. Since most inference methods imply a certain type of consequent variable, the type of variable limits the range of methods that can be chosen for it. In Istar, the following main inference methods are available for the following types of variable: BOOLEAN variable: AND, OR, and also a number involving a predicate over the antecedents such as 'First antecedent greater than all the others'. Number variable (INTEGER or FLOATing point): Add, subtract, multiply, divide, mean, maximum, minimum, and also a number that count for how many of the antecedents a certain predicate like 'greater than' is true. PROBABILITY variable: ProbAnd (result = pA * pB * ...), ProbOr (pA + pB - pA * pB), maximum, minimum, and various predicate-related methods. BAYESIAN variable: This is a PROBABILITY which has an additional a-priori probability which is its starting point when collecting evidence. All the probability inference methods, plus Bayesian accumulation of evidence. (Other variables include things like proportions, rational numbers, odds, directions (angles), OZMOs (one-zero-minus-one such as needed for sine), strings, etc. These are not much used in the Trust Assessment KB and so will not be discussed here. All types of variable and inference method are outlined in Basden and Brown (1996) and online web pages (Istar 1,2) displays the current set.) It should be noted that the use of the name 'Probability' for a variable type is something of a misnomer. Strictly, 'probability' is a knowledge level phenomenon, by which human beings bring together into a single concept some statistical knowledge of a set of events or states or objects. That which is known in Istar as 'Probability' is, strictly, a degree between 0 and 1. It is called 'Probability' because both the variable type and also the inference methods associated with the type happen to be useful when processing probabilities. In this text we will use the upper case 'PROBABILITY' to mean the numeric symbol in Istar that can be used to denote some degree, and the lower case 'probability' to mean the knowledge level statistic. Thus a PROBABILITY in Istar can often be used to represent a probability, but in the Trust Assessment KB it is mainly used to represent a degree of (human) belief. In addition to types of variable and inference methods, Istar allows most inference relationships to carry unary operators and weightings. A unary operator performs some operation on the value found in the antecedent variable before it is used in the inference method. The weighting is a parameter used in the unary operation. Some common unary operations offered by Istar include: Direct: No change in value. Negate (no parameter): For a numeric antecedent, the value used in inference is the negative of that found in the antecedent. For a BOOLEAN antecedent the value is the opposite of the truth value of the antecedent. For a PROBABILITY or BAYESIAN antecedent the value used is one minus the value in the antecedent. Scale, Shrink: The value of a numeric antecedent is multiplied or divided by the value found in the weight. ProbAnd: The value of the PROBABILITY or BAYESIAN antecedent is multiplied (reduced) by the PROBABILITY value found in the weight. ProbOr: The value of the PROBABILITY or BAYESIAN antecedent is increased by the PROBABILITY value found in the weight in accordance with the arithmetic for probabilistic OR. BayesianWeight: The weight is used in the Bayesian inference method described below. Many others are available, some described in Basden and Brown (1996), and the full current list available on a web page (Istar, 3). 4.5 Mapping KL Combinations to SL Inference Methods Not only is there a mapping from KL concepts to SL variables, but there is also a mapping from KL combinations of beliefs into SL inference methods. For each SL variable in the KB the inference method must be selected to match what happens at the KL. In some cases the inference methods offered by Istar match KL combinations well, in some cases they match approximately, while in at least one case none of the available inference methods were sufficient and one had to be constructed via numeric arithmetic employing floating point numbers (see section 4.5.4) In most cases the KL combination required that unary operators and weights were assigned to the antecedent relationships in order to achieve the effects desired. The process by which the SL inference methods, unary operators and weights were selected was as follows. First, the sufficiency and necessity of the KL antecedent concepts was examined, and the result of this would often suggest which SL inference method was most appropriate. Then the precise effect of the SL inference method was simulated (mentally or in Istar by building a trial inference step) and this was compared with what was desired at the knowledge level. Finally, where this was not appropriate, unary operators and weights were assigned to the SL antecedent relationships in order to bring the effect of the inference method more in line with the effect desired at the knowledge level. It should be noted that, when selecting SL inference methods by which to implement the KL combinations, because there is an element of variability and uncertainty inherent in the interpretation processes in steps 2, 4 and 7, it is not essential for the parameters of the inference method or of the unary operators to be precise. What is important is the overall shape and behaviour of the curve that describes how the value of the consequent varies with that of the antecedents. We found four KL combinations to be important. Some tend to be useful in combining the input from questions, others in the central portion of the KB while a special combination, which we call 'penalty', was required at the output end of the KB to combine all the major factors together into a single trust quotient. Those nearer the output have a greater effect than those nearer the input. We now discuss these and how we implemented them using Istar's facilities, under names that reflect their KL meaning rather than their SL algorithmic implementation. A graph for each is shown in Fig. 5 graphs not yet done Fig. 5. Graphs of the Four KL Combinations. 4.5.1 Belt and Braces At the KL we recognised that some antecedents build upon each other and asymptotically increase the amount of trust one has in the consequent until complete trust is attained (i.e. 1 at the SL). The absence of one antecedent does not decrease one's trust in the other antecedents that are present, whilst the presence of two antecedents rather than one increases but does not double the trust. The more antecedents that are true, the better, but the effect of each reduces as the consequent approaches unity. (Note that a value of 1.0 can never fully be reached in practice, no matter how many antecedents there are, since it is always possible, however rare, for trust to be misplaced.) We called this combinatorial method "belt and braces", the analogy being that a person who progressively increases the amount of support for his trousers (belt, buttons, braces, clips etc.) will become more trusting that they will not fall down. Never the less the possibility still exists that under some extreme conditions the trousers may fall down, so a value of 1.0 is never attained. An example of this combinatorial method is used in the process of identifying (the name of) a person. Sixteen methods were recognised in our system e.g. passport, credit card, fingerprint, utility bill, photo ID card, bank reference etc. and the more of these a person presents, the more trust one can have in the identification. 'Belt and Braces' is similar to probabilistic OR, in that the effects of the antecedents combine in an asymptotic fashion. Consequently, Istar's ProbOr method was deemed appropriate for this combination, the more so since one could, in many cases, argue that the combination should indeed be one of independent probabilities that certain states were true. For example, each identification method has a certain probability that the identity will be mistaken, and the methods are largely independent of each other. So the probability of mistaken identification when employing two or more methods is reasonably accurately given by probabilistic OR. However, few of the variables in the KB were actually mapped, strictly, to probabilities. Rather, most were simply degrees of quality. There are several reasons for this. One is that in cases of extremely low or high probability, other extraneous factors are likely to pull the extreme value towards the centre. For example, the probability of mistaken identity using a fingerprint might be deemed to be 1 in 6 billion, hence a reliability figure of 0.99999984 (since it is assumed that fingerprints are unique across the population of the world). But the likelihood of mistaken identity is in fact much larger than this, owing to the possibility of errors occurring in the taking, storing, matching and interpretation of fingerprints. But some modification is needed to this combinatorial method, because not all of the antecedents will have the same weight or effect. Therefore, in 'Belt and Braces', we gave each antecedent a ProbAnd unary operator and an attendant weight that was used to reduce the degree value of the antecedent before it was used in the ProbOr inference method. For example, a bank reference is less reliable than a passport, which is less reliable than a fingerprint, and the three were given weights of 0.7, 0.8, 0.9 respectively, so that if a bank reference is presented, its contribution to the consequent would be only 0.7, while the contribution of a fingerprint would be 0.9. This gave the desired effect, that CAs that employ fingerprints could achieve high reliability of identification without any other method, but that using bank references alone gave a much lower reliability. The weights given to each identification method were obtained by interviewing a number of security experts (see section 5). Belt and braces was used for combing a number of different nodes in the KBS, for example: organisation authentication methods and the damages covered in a CPS. 4.5.2 Paranoid The inverse of "Belt and Braces" at the KL we termed paranoid. With paranoid combinations, we need all antecedents to be of high quality for trust to be high. If a single antecedent is of low quality or zero value, trust rapidly decreases. Low quality of several antecedents decreases trust even more, but each time a low quality antecedent is added, it has less and less impact on the consequent value. Trust therefore decreases exponentially with the number of missing or low quality antecedents. We used this combinatorial method for example, on the factors that comprise the compliance audit i.e. list of compliance topics, qualifications of the auditor, frequency of audit etc. If any one audit item was absent or low quality, our trust in the audit would decrease substantially. 'Paranoid' is implemented by Istar's ProbAnd inference method. ProbAnd achieves the desired effect, in that as more and more antecedents depart from perfect quality, the consequent reduces towards zero. Again, a simple probabilistic AND is not appropriate on its own, in that if a single antecedent were answered with a 'No' (zero quality factor) then the consequent would be zero, and all the other antecedents would have no effect. So the antecedents must be weighted so that some have less of this reducing effect than do others. They were all given a ProbOr unary operator and a weight that reflected by how much each antecedent could reduce the degree of the consequent on its own. In this way, no single antecedent can reduce the consequent to zero on its own. The weighting of the various items was determined by interviewing a number of experts, and is fully described in (Chadwick, and Basden, in publication). The paranoid combination method was used on a significant number of the KB nodes, for example: the audit trail, the various obligations of the various parties, and key generation. 4.5.3 For and Against This KL combination method is conceptually midway between "Belt and Braces" and "Paranoid". All antecedents are considered, and contribute towards the overall trust. If an antecedent is present and of high quality, it has a positive contribution to the consequent value, whilst if it is low quality or missing (zero value) it negatively impacts on the consequent value. This combination is suited to evidential reasoning. At the SL, Bayesian accumulation algorithms are used to combine the antecedents. 'For and Against' starts somewhere in the middle (at an a-priori value which can be defined) and antecedents may pull the consequent value down as well as up. The method is implemented by Istar's Bayesian inference method, which undertakes a simple form of Bayesian accumulation of evidence similar to that used by early expert systems (Duda, Hart and Nilsson, 1976). The central idea is to work with odds ratios which are equivalent to the 0-1 degree (probability) according to the formula: O = P / (1 - P) where P is the degree (treated as a probability) and O is the equivalent odds. The consequent starts with an a-priori value, Pc0, which is converted into odds, Oc0, and this is them multiplied by a weighting, Wi, for each antecedent. The weight of the antecedent varies according to two weight figures on the inference relationship, which correspond to the logical sufficiency (LS) and logical necessity (LN) of the antecedent as evidence, together with the actual value of the antecedent variable, Pa, and whether this value is above or below its own a-priori value, Pa0, such that the weight, W, corresponding to an antecedent value, Pa, is given in Table x2. For positive evidence, LS is greater than unity and LN is less, while for negative evidence it is the other way round. Table x2. Conversion of Antecedent Value to Weight Pa W 0 LN Between 0 and Pa0 Between LN and 1.0 Pa0 1.0 Between Pa0 and 1 Between 1.0 and LS 1.0 LS In Istar the curve is piecewise linear, but a smoother curve might be preferable, probably of a logarithmic nature, such that its slope is continuous at the a-priori. Only those weights are multiplied with Oc whose antecedents have a known value. After all antecedents are accumulated in this manner, Oc is then converted back into a degree between 0 and 1 by the formula (Oc / (1 + Oc)). For symmetric evidence LS is the inverse of LN, but by controlling these values idependently, the effect of the antecedent can become highly asymmetric, approaching, in the extremes, the behaviour of either the paranoid or belt and braces. Thus this method provides a wide range of control, and is employed in a small number of variables in the middle of the inference net. 4.5.4 Penalty The 'Penalty' method is employed for the final stage of the KB, when the values of all variables that represent the major trust factors are combined to produce the single figure that is the trust quotient. With inference of the type used here, the nearer to the goal variable, the more important it is that the shape of the curve is correct. Several inference methods were considered for this. One was the simple ProbAnd, on the grounds that if any of the major factors were missing we should be 'paranoid' about it. This was the inference method employed in the earlier versions of the KB but it tended to penalise too heavily when factors were near unity (since there were seven factors, if each of them had a score of 0.9, the final trust quotient would end up approximately at 0.4) while it did not penalise low valued factors heavily enough. Conceptually we wanted a curve with a discontinuity in it, as shown in Fig. 5d. This allows us to penalise antecedents that drop below some threshold value. Initially there should be a gentle fall in the consequent trust value, caused by one or more of the antecedents falling from perfect quality but still remaining high quality (e.g. 1.0 down to 0.9). However, at some point there should be a discontinuity, when the consequent trust drops dramatically, caused by one of the antecedents falling below its acceptable threshold value. This has the desired effect of not penalising high quality factors, whilst penalising low quality factors. We considered using Bayesian Belief Networks (Hackermann and Wellman, 1995, Fung and Favero, 1995), in which the combination algorithm is controlled by a table. In theory, this should give a greater control over the shape of the curve than the simple Bayesian algorithm offered as standard by Istar, and should be more appropriate when the antecedents are not independent. The BBN table contains multiplication coefficients which are used in the calculation, one coefficient for each possible combination of antecedents taking values of either 0 or 1 (false or true). For example, suppose trust were to depend on only two antecedent factors, Cryptography and Procedures, which are represented as variables C and P, each of which can have the value 0 to 1. Suppose also that the table of coefficients for these two antecedents is: C P Coeff 0 0 0.1 0 1 0.71 1 0 0.62 1 1 0.96 Then, the algorithm to combine the actual values of P and C is: 0.96*P*C + 0.62*(1-P)*C + 0.71*P*(1-C) + 0.1*(1-P)*(1-C) In this way, both the low and the high values of both P and C have an effect, and the amount of the effect can be controlled by the coefficients in the table. Each stage of inference has its own table of coefficients. Although the BBN is easily demonstrated for a small network, as above, the problems of calculating the required coefficient tables become severe when the number of inputs to a node becomes large. Seven antecedents requires a table with 128 entries. Elsewhere in the model nodes with 10 antecedents are found, giving rise to tables with 1024 entries. Whilst the storage of such tables is not a problem, the calculation of the values to put in them is a difficult problem, the magnitude of each coefficient having to be found by asking experts. Not only is this time-consuming, but there is a more fundamental difficulty, even for relatively small tables. Some combinations of antecedents are so rare that the expert can do little more than guess at what the figure should be. Moreover, exploration with small BBNs has also shown that the trust values calculated do not accord well with expected values. With a large table, the effect of any one antecedent is swamped by the others, so it is difficult for any one antecedent to have a large penalty effect. Setting coefficients specifically to achieve this for one antecedent tends to upset the balance of the table for others. These problems led us to the algorithm currently implemented for this stage of inference, which, at the SL, we might call the Weighted Accumulation Threshold Algorithm. In this algorithm, the value of the consequent is derived from two opposing tendencies: a positive tendency which is the sum of the antecedent values and a negative penalty tendency which comes into play if any of the antecedents are below some threshold value. Each of the antecedents has two parameters, a weight which is used in the positive tendency and a threshold figure which dictates when its value is low enough to exert a penalty influence. The positive tendency is the weighted sum of all the antecedent values (the sum of the weights of all antecedents is unity). This is multiplied by the penalty factor, which is the product of all those antecedent values that are below their individual thresholds. In this way, if an input is greater than its threshold it contributes directly to a weighted average of trust, but if it is below its threshold it reduces the trust level even more than its contribution to the weighted average, eventually reducing he trust to zero. This algorithm has the advantage that it is easy to apply the experts weightings of the contributions, and it reflects at the knowledge level the way trust is expected to depend on its major factors. The most difficult issues that we were left to decide were the threshold values for each of the antecedents. Ideally these should have been chosen by the experts during our interviews, but we had already finished them when we realised that they were needed. We thus used our own expertise to chose sensible values. Mathematically the output (consequent) value of a node is given by: equation Where Ik is an input (antecedent) value, Wk is its weight, Tk is its threshold value and H() is the unit step function. Fig. 6 shows outputs from the function for threshold values of 0.2, 0.5 and 0.8, where the weight is plot of WTF Fig. 6. Actual Plot of Penalty Function 5. KNOWLEDGE ACQUISITION In the main, the knowledge acquisition process followed normal practice. It is described in detail in (Chadwick and Basden, in publication), and we merely highlight the main points here, but we then discuss one issue that emerged that is of more general importance. 5.1 Overview of the Process There were two main stages in constructing the KB: first, the overall structure was gained by selecting knowledge from a comprehensive published source (Chokhani and Ford, 1999), then the weights of evidence were elicited from leading experts in the field. What our experience in accomplishing this has demonstrated is that the second stage is far from simple, and can usefully be broken down into several steps, each with a distinct purpose. Chokani and Ford (1999) contains a detailed description of the issues a CA must take into account, and is widely accepted as authoritative. We extracted the issues that were relevant to trust and implemented them as the first prototype KB. It emerged that the KB has a tree-like structure, which was not surprising since Chokani and Ford had presented their knowledge in this form. In stage 2 knowledge was elicited from domain experts. When compared with some KBS projects, it is perhaps remarkable that the overall structure of the KB underwent so little modification during stage 2. This evidences the high quality of the original source. For those projects that involve similar high quality sources, the main contribution of domain experts is to supply what the initial knowledge source could not, namely knowledge that is affected by context and use in real life problem solving (Attarwala and Basden, 1985), such as weights of evidence links and many small, though extremely important, refinements. Thus stage 2 comprised a series of interviews with leading experts in the PKI field, using questionnaires. The questionnaires largely followed the internal structure of the KB and asked the respondents to comment on relative importance of various factors for some or all links in that structure. As the factors were quite different in many cases, it was a difficult task to perform, and was perhaps analogous to asking gourmets to place a value on the different fruits in a fruit salad. Four different questionnaires were produced during the research: a pilot questionnaire, followed by versions 1, 2 and 3. The pilot questionnaire, attempted with two experts, asked questions about every node in the graph but it proved too large and cumbersome to use. (Nevertheless, this exercise yielded valuable knowledge that was employed in building the KB.) Version 1 removed the least significant questions and concentrated on the later nodes in the graph that would go towards calculating the final trust quotient. But we found that the concepts were sometimes difficult to explain or elicit answers on, perhaps because some of the questions were worded ambiguously, or a context for thought had not been established prior to asking the question. Version 2 added a number of questions designed to get the respondents thinking about the relevant issues and setting a context, before asking them to weight the various aspects of trust against each other. This version, the penultimate, was completed by the most experts and was the one used to populate most of the variables and relationships in our knowledge base. It highlighted one or two areas where we did not have an optimal design for the KBS, with one or two questions out of place and a misplaced arc. These were subsequently corrected, to produce version 3, which was used in the last few interviews. The time to complete a questionnaire fell from 4 hours for the pilot to 1.5 hours for the final versions, which is more acceptable to busy domain experts. 5.2 Some General Issues in Knowledge Acquisition In the above we can detect several distinct issues that might need to be addressed whenever questionnaires are employed for interviews in this manner: 1. Generating the initial questionnaire. That the pilot questionnaires could provide any useful knowledge at all suggests that an effective way of formulating questionnaires is to derive them from the structure of the KB. (This however is likely to be the case only once the KB has reached a certain degree of accuracy, otherwise most of the interview time would be taken up with the expert trying to 're-educate' the interviewer.) 2. Identifying the important domain issues. In such a questionnaire some questions will always be more important that others. Which these are cannot always be known beforehand, so there is likely to be a step in the interview process, as that between our pilot version and version 1, during which this selection occurs. 3. Making the questioning process effective. Having identified the set of important questions, attention must be given to ensuring that the interviewees can provide high quality answers thereto, and answers on which we may rely. We found this entailed adding questions that contextualised the main ones or helped to focus attention on their true meaning. 4. Responding to corrections. We can expect the interviewing process to highlight deficiencies in the structure of the KB. Hopefully, these will be relatively minor, but they should be incorporated as part of the process. This will entail producing new versions of the questionnaire that reflect these changes in the KB. What this means is that the process of building a KB can never be a mechanical one, in which stage 2 is merely one of filling in the weights. It is important that, even when the original knowledge source is of a high quality, the knowledge engineer takes what Winograd (1995) calls a 'designer's- eye-view', rather than a 'constructor's-eye-view', and that the process of knowledge refinement (Basden and Hibberd, 1996) is recognised as important. 6. TRUST QUOTIENT CALIBRATION We have stated that a trust quotient value of zero means that a CA is regarded by the experts to be absolutely untrustworthy and that a value of 1 means it is regarded as being absolutely trustworthy (athough we have also noted that a score of 1.0 is impossible to achieve. The maximum score possible from the current system, by giving each question the highest quality answer, is 0.98). But what does an intermediate value of say 0.5 mean? And what is the meaning of a difference of 0.1 between two CAs' trust quotients? Whilst we think that absolute quantitative calibration of the trust quotient scale is too complex to achieve at the current time, we believe that qualitative calibration of the scale is possible. There are two ways that a user can do this. One way is to calibrate the trust quotient scale by running the expert system against the CPs/CPSs of several well known public CAs. In this way a user can attach KL meanings to the various scores on the scale. The other way is to compare the trust quotient scored by his own CA, whom he knows and trusts, against the trust quotient scored by the remote CA that he is trying to assess. Both ways are necessarily subjective, and depend in part upon the personality traits of the user and what s/he values so that different users may well provide different answers to the same question from the same CPS. Many of the questions posed by the expert system are based on judgements and interpretations of the CPS e.g. "To what extent are damages covered in the CPS?" or "How well do you feel the CA's private key is handled in the CPS?" In addition, many CPSs are mute about a specific trust issue, so the user has no objective way of deciding what answer to give to the expert system, so different users will obtain different trust quotients from the same CPS. Thus there are no absolute quantitative scores available at the current time. We have run the expert system against the CPSs of several well known CAs. Two are reported here for comparative purposes. Verisign is probably the best known certificate issuer in the world. It is configured into all popular browsers as a root CA, and provides a range of certificates from the low cost low trust Class 1 certificates to the medium assurance Class 3. Viacode on the other hand is the CA run by the UK Royal Mail and is a medium assurance CA. It runs services for the UK government and the National Health Service. On first pass through the expert system both Verisign and Viacode CAs scored trust quotients of zero. On exploration of the KB as described above we found that this was because several topics were not addressed in either of their CP/CPSs and we had given these factors a score of zero. For example, there is no mention about archiving audit logs by Verisign, there is no mention about the network security controls applied by Viacode, and neither mention job rotation of trusted roles. As we had no way of knowing from the CPSs whether the CAs implemented these features very securely, or barely adequately or not at all, we initially assumed not at all. On the second pass through, we gave a score of 0.5 to all the questions for which the CPSs provided no answers, i.e. we assumed that the CAs had probably addressed these issues just about adequately. This still led to a trust quotient score of 0 for Viacode because the use of the CA private key was still not judged to be secure enough. If however, the CA private key usage factors were set to very secure, then the trust quotient rose to 0.25 for Viacode. If we then assumed that four of the unmentioned critical trust factors (namely network security, list of audit compliance topics, job rotation and sanctions against staff for unauthorised actions) were implemented very securely then the trust quotient for Viacode rose to 0.9. This quite poignantly shows the effects on trust that just a few critical elements can have, and how it is important that in a CP/CPS all relevant factors are covered. Verisign fared slightly better scoring a trust quotient of 0.45 for Class 3 certificates, when all the unmentioned factors were given scores of 0.5. If the three most critical missing factors (namely job rotation, protection of audit logs, and list of audit compliance topics) were assumed to be implemented very securely, then Verisign Class 3 certificates yielded a trust quotient of 0.87. We did not find it possible to ever get a trust quotient above 0 for Verisign Class 1 certificates. This is simply because no authentication is ever carried out on the identity of the certificate subject (other than a limited verification of their e-mail address). So no matter how trustworthy Verisign might be with its internal CA operations, relying parties cannot realistically place much trust in a signed message from a subject having a Verisign Class 1 certificate. The process described above should be seen, not so much as commenting on the trustworthiness of two major CAs, but as an illustration of the process a user might go through in order to assess any CA. The first figure given by the KBS should never be taken as final, but factors should be explored and given several values in a what-if manner. The emphasis should be on finding out which qualitative factors might be critical in any one case, rather than on achieving a precise numerical score. 7. DISCUSSION 7.1 Roles and Benefits of the Trust Assessment KBS It was originally hoped that the Trust Assessment KBS discussed in this paper would be able to fulfil the role of an intelligent agent that could be called upon by email and browser software to provide an automated assessment of the trust that can be placed in the authenticity of the sender of a message. The original hope was that no input from the user would be required. However, for several reasons, this looks unlikely to transpire. While it might be possible to translate extreme values of the trust quotient into automated decisions, it is difficult to know how an intermediate figure can be so used. Most quotients are likely to be in the intermediate range. Even though it would be feasible, in principle, for the KBS to take, as input, a complete CP or CPS, and undertake textual analysis thereof in order to obtain much of the information it needs, a significant amount of required information entails too high an element of human interpretation. At present there is little standardization of format or content, and the differences between CPs and CPSs exacerbate this. In many cases the required information is not stated explicitly in the document and must be inferred from an overall reading, and in some cases other material must be accessed, to enable the user to come to a judgement about the answers to be given. The output trust quotient, being a single figure, cannot meaningfully differentiate all the factors on which a decision of trust can be made. Each decision is made in a different circumstance, affected by economic and personality aspects of trust as discussed above. It would be difficult to include these aspects in the KBS in such a way as to preclude user input. Therefore, the KBS is most likely to be used in a decision support mode, by which a human user interprets a CPS using the KBS. By exploring the knowledge after arriving at a first trust quotient, in the manner illustrated above, the user can identify at least two useful things. Where the trust quotient is low, the user can discover which factors are the ones that make it low. Then they can decide whether these factors are significant in their own situation. On the other hand, if the trust quotient is high, the user can explore the knowledge to identify on which factors this result depends most crucially, and then they can decide whether these factors are robust in their own situation. This kind of exploration involves a degree of what-iffing, whereby the user tries new answers to questions and watches the result. The user benefits from such exploration not only by achieving a more precise or robust view of the situation in hand but also by gaining a greater appreciation of the critical trust factors that can affect such situations more generally. In this way their knowledge is built up, to apply in future situations of this type. The former benefit occurs by means to the KBS fulfilling what Basden (1983) called the consultancy role, the role traditionally assigned to knowledge based systems technology, but the latter is what he identified as the knowledge refinement role, which, though not often discussed happens in practice to be a major role of KBS. 7.2 Modes of Use of the KBS Although using the Trust Assessment KBS as a completely automated intelligent agent is unlikely to be very effective, three other possible uses of the KBS have emerged. Two of these ways are similar to the hoped for automated assessment: to assess the trust that can be placed in the authenticity of the supposed sender of a specific communication, and, more generally, to assess a CA as a whole. In the latter type of usage, different CAs can be compared. In both these uses, the user would run the KBS with the relevant CPS and other material to hand. Running the KBS takes between one and several hours to reach a trust quotient, depending on how easily the CPS can be interpreted by the user and on how much exploration of the knowledge is undertaken. Exploration would be advisable when comparing CAs since they might fail on different factors, and so would perhaps be suited to different types of Internet work. However, a fourth, new, type of usage for the KBS has emerged: to aid CA administrators in the design of new CA services and their controlling CPs and CPSs. This was first suggested in discussion with one of our partners, Tim Dean of the U.K. Defence Establishment Research Agency. For military uses especially it is necessary to ensure that there are no weak points in certification policy or procedures, but it is often difficult to be confident that all possibilities have been considered. Therefore a KBS that holds reasonably complete knowledge can be used in a checklist role (Basden, 1983) in which the user is stimulated by the questions asked by the KBS to consider things they would normally take for granted. KBS technology does not suffer from the lapses of memory that inflict human beings. For this reason, running the Trust Assessment KBS for a prototype CP or CPS can identify areas of weakness. The ability to explore the knowledge that the KBS offers is particularly important here. Thus the problems that we experienced when evaluating the CPSs of Verisign and Viacode, due to their muteness about certain aspects of trust, would be obviated if the CA administrators had used the KBS to ensure that every trust topic has been addressed in their CPSs. 7.3 Automating the Trust Assessment The process of arriving at a trust quotient is time consuming, and requires considerable skills on the part of the user. Most Internet users do not possess the required level of skill (nor for that matter the patience) necessary to run the KBS. It would be far better if the CA administrator could run the KBS once, complete all the sections in the knowledge base, and then export the completed data to a structured file. This file could then be published on the Internet along with the CA's textual CPS and CP. We have added this capability to our system, so that we can now produce an XML formatted trust related subset of the CPS. The KBS can then extract the answers to its questions from this XML document, and produce a trust quotient based on the CA administrator's subjective assessment of their CPS. Clearly this exposes the relying party to risk where the CA's procedures in practice do not adhere closely to their published policy, or the administrator's assessment of them. To this end, we have considered how we might extend the expert system to assess actual practice. 7.4 On Assessing Actual Practice At first sight, it would seem that the knowledge base required for making such an assessment might be similar to that described here. For example, questions that ask "Does the CPS state X?" could be reworded as "Does the CA actually carry out X?". However, there are several complications before such assessment can be a reality. We have found that more substantial changes than these must be made to the KBS, and that several changes would be required to the legal infrastructure of the Internet. The changes we believe to be necessary include: 1. Some questions currently asked about the CPS are irrelevant when considering actual practice. Conversely, new questions must be added. Nevertheless, we believe that the bulk of questions will remain relevant. 2. Of those questions that remain relevant, it might not be sufficient merely to reask them in a new form ("Does the CA actually carry out X?"). In some cases, any differences between the answers might provide important indicators. For example, if it is found that a CA consistently underperforms in one area, then we might be justified in giving it a lower trust quotient than if its underperformances were randomly distributed in both time and topic area. 3. How do we obtain answers to the new types of question? We have identified at least three distinct possibilities. In some questions, this can be assessed by observation. For example, if the CPS says that the CA publishes revocation lists every 24 hours, a good indicator of actual practice comes from examining such lists every 24 hours over a period of time. (To do this, however, requires action in advance of the need to assess trustworthiness.) For some other questions marketplace rumours can supply the necessary information. However, for the majority of the information required, a new mechanism is likely to be needed: an Audit Certificate. 4. An Audit Certificate is an electronic document produced by the organisation that is authorised to make audit checks on the CA. When making such audits the auditor compares actual practice with published procedures, and compiles and publishes his results in a digitally signed Audit Certificate. These certificates are then made publically available on the Internet. For this to become a reality such mechanisms would have to be established and formally recognised as part of the legal infrastructure of the Internet, in much the same way as the publishing of a limited company's financial accounts is currently a legal requirement. 5. To fully integrate the processing of Audit Certificates with the operation of a trust assessment KBS of the type we are proposing, the audit certificates should be in a standardised, computer-readable format. To this end, we have examined the possibility of publishing Audit Certificates as X.509 Attribute Certificates containing an XML formatted audit attribute. Since Istar has already been modified to export its knowledge base in XML format, thus giving us the ability to automatically generate XML versions of the trust components of CPSs, it would appear that Istar could be further enhanced to compile XML audit attributes. 6. We have experimented with two mechanisms by which the Trust Assessment KBS, run by Istar, might obtain the Audit Certificate information. a) A specialised Trust Check Server would be accessed by the KBS, with requests for the necessary information. The Trust Check Server would, as part of its activity, continually retrieve Audit Certificates and CRLs. We have constructed a pilot version of such a Server, and modified Istar to communicate with it. This is described in (Ball, Chadwick and Basden, in submission). b) The KBS software (Istar, in our case) could be modified to read and process the Audit Certificate directly. This has both a disadvantage, of increasing the complexity of the KBS, and an advantage, in that it is likely that some of the knowledge currently in the KBS could be capitalised upon to perform the necessary analyses. 7. Whichever implementation route is chosen, the contents of the Audit Certificate must be well defined. From our initial investigations, it seems that this can be defined by the information the KBS requires in order to calculate a trust quotient. 8. CONCLUSION We have shown how we have built an expert system for computing the trust in public key certification authorities, and described some of the technical complexities in building such a system that have not been fully discussed elsewhere. We have discussed the roles and benefits of such a system when in use, giving examples of its use with current public CAs. Finally we have proposed future extensions to the system so that trust can be recomputed not only from what a CA says it does, but also from what it actually does in practice, as viewed by its auditor. To enable such a system to operate will require legal changes to the infrastructure of the Internet, to mandate that public CAs annually publish their Audit Certificates. The more general significance of this work, and in particular the latter suggestions lies in the impact that technology can have on high level infrastructure and strategy. It has been observed that the usage of expert systems can radically change the role that their users play and their working relationships (Castell, Basden, Erdos and Barrows, 1995, Basden, 1994). We have discussed similar changes in ways of working here that are made possible by the features offered by KBS technology. But we can glimpse the possibility of an even more fundamental structural change being brought about by the use of KBS technology: the establishment of an entirely new legal requirement, the annual publishing of an Audit Certificate by a public CA. Further, the contents of the parts of the new infrastructure, the Audit Certificate, would be defined, not only by the general deliberations of lawyers and auditors, but by the very specific requirements of the front-line KBS, namely that KBS that assesses the trustworthiness of a CA from its Audit Certificate. In this way we can, perhaps, see the possibility of a fruitful symbiosis emerging between KBS technology and legal infrastructure. ACKNOWLEDGEMENTS The authors would like to thank the EPSRC and DERA for funding this research under grant number GR/L 54295, Entrust Technologies who provided us with PKI and LDAP software and toolkits, and Dr. John Evans who produced the original version of the knowledge base that has stood the test of time. REFERENCES Adams C., Lloyd S. (1999) "Understanding Public-Key Infrastructure: Concepts, Standards, and Deployment Considerations". Macmillan Technical Publishing USA, ISBN 157870166X. Attarwala F.T., Basden A., (1985), "A methodology for constructing Expert Systems", R&D Management, v.15, n.2, pp.141-149. Ball E., Chadwick D.W., Basden A. (submitted) "The Implementation of a System for Evaluating Trust in a PKI Environment", Transactions on Internet Technology, Feb 2001. Basden A. (1983), "On the application of Expert Systems", International Journal of Man-Machine Studies, v.19, pp.461-477. Basden A., (1994), "Three Levels of Benefit in Expert Systems", Expert Systems, v.11, n.2, pp.99-107. Basden A., (2000), Some technical and non-technical issues in implementing a knowledge server, Software - Practice and Experience 30:1127-1164. Basden A., Brown A.J. Tetlow S.D.A. Hibberd P.R., (1996), "Design of a user interface for a knowledge refinement tool", Int. J. Human Computer Studies, v.45, pp.157-183. Basden A., Brown A.J., (1996), "Istar - a tool for creative design of knowledge bases", Expert Systems, v.13, n.4, pp.259-276, November 1996. Basden A., Brown A.J., (1996), Istar - a tool for creative design of knowledge bases, Expert Systems, v.13, n.4, pp.259-276, November 1996. Basden A., Hibberd P.R., (1996), "User interface issues raised by knowledge refinement", Int. J. Human Computer Studies, v.45, pp.135-155. Battacharya R., Devinney T.M., Pillutla M.M. (1998) "A Formal Model of Trust Based on Outcomes," The Academy of Management Review, 23 (3), 459-472. Beth T., Borcherding M., Klein B. (1994) "Valuation of Trust in Open Networks". In D Gollman, ed., Computer Security - ESORICS '94 (Lecture Notes in Computer Science 875), pages 3-18, Springer-Verlag, 1994. Castell A.C., Basden A., Erdos G., Barrows P., Brandon P.S., (1992), "Knowledge Based Systems in Use: A Case Study", Proc. Expert Systems 92 (Applications Stream), British Computer Society Specialist Group for knowledge based systems. Chadwick D.W., Basden A. (submitted) "Evaluating Trust in a Public Key Certification Authority". Computers and Security, Jan 2001. Chokhani S., Ford W. (1999) "Internet X.509 Public Key Infrastructure Certificate Policy and Certification Practices Framework". RFC 2527. March 1999. Duda R.O. Hart P.E., Nilsson N.J., (1976), , "Subjective Bayesian methods for rule-based inference systems", Proc. AFIPS National Computer Conference, 45, pp. 1075-1082. Fung R., Favero B.D., (1995), "Applying Bayesian networks to information retrieval", Comm. ACM, v.38, n.3. Heckerman D., Wellman M.P., (1995), "Bayesian networks", Comm. ACM, v.38, n.3. Hibberd P., Basden A., (1995), "Procurement and the use of intelligent systems of contract authoring", Proc. RICS COBRA Conference, Edinburgh September 1995. Information Security Committee, (1996) "Digital Signature Guidelines". Electronic Commerce and Information Technology Division, Section of Science and Technology, American Bar Association, August 1996. Available from http://www.abanet.org/scitech/ec/isc/dsgfree.html ISO/ITU-T Rec. X.509 (1997) The Directory: Authentication Framework Istar (1) "http://www.basden.u-net.com/pgm/Istar/docs/values.html" Istar (2) "http://www.basden.u-net.com/pgm/Istar/docs/inference.html" Istar (3) "http://www.basden.u-net.com/pgm/Istar/docs/uop.html" Network Associates (1999) "PGP Freeware for Windows 95, Windows 98, Windows 2000 and Windows NT - Users Guide", Version 6.5.2, Ch 5, p 99, . Newell A., (1980), "Physical symbol systems", Cognitive Science, v.4, pp.135-183. Newell A., (1982), "The Knowledge Level", Artificial Intelligence, 18:87-127. Reiter M.K., Stubblebine S.G. (1997) "Towards Acceptable Metrics of Authorisation". In Proceedings of the 1997 Symposium on Security and Privacy, IEEE Computer Soc. Press, p10-20, May 1997 Tarah A., Huitema C. (1992) "Associating metrics to certification paths". In Computer Security - ESORICS '92 (Lecture Notes in Computer Science 648), pages 175-189, Springer-Verlag, 1992. Winograd T., (1995), "From programming environments to environments for designing", Comm. ACM., v.38, n.6, pp.65-74. Copyright (c) University of Salford, 2001, All Rights Reserved. Last updated: 22 March 2001. 3 May 2001. --------------------------------------------------------------------- [Image] [Image] Name: Curves.iff Curves.iff Type: image/x-ilbm Encoding: base64 Name: ictkb.iff ictkb.iff Type: image/x-ilbm Encoding: base64 [Image] [Image] --------------------------------------------------------------------- Attribute: Malpractice has value: |****....| Its antecedents are: Click here to reset it and resume. Timeout set at 10 minutes; Set to 15 10 20 30 60 minutes. See KB Intro Page again. Resume interrupted session. Restart session from beginning. Return to welcome page. Stop entirely. Page created by Istar knowledge server. Unless otherwise stated above, every page produced by this Istar Knowledge Server and the content of each of its knowledge bases is Copyright (c) University of Salford, U.K. If you wish to make use of any of this material please contact Istar@basden.u-net.com. 
work_ztvalqgjlna2fhj7accs4tukde	----	_ESWA_A multi-objective PSO for job-shop scheduling problems A multi-objective PSO for job-shop scheduling problems D. Y. Shaa and Hsing-Hung Linb* aDepartment of Industrial Engineering and System Management, Chung Hua University,, Hsin Chu, Taiwan R.O.C.; bDepartment of Industrial Engineering and Management, National Chiao Tung University, Hsin Chu, Taiwan R.O.C. Abstract Most previous research into the job-shop scheduling problem has concentrated on finding a single optimal solution (e.g., makespan), even though the actual requirement of most production systems requires multi-objective optimization. The aim of this paper is to construct a particle swarm optimization (PSO) for an elaborate multi-objective job-shop scheduling problem. The original PSO was used to solve continuous optimization problems. Due to the discrete solution spaces of scheduling optimization problems, the authors modified the particle position representation, particle movement, and particle velocity in this study. The modified PSO was used to solve various benchmark problems. Test results demonstrated that the modified PSO performed better in search quality and efficiency than traditional evolutionary heuristics. Keywords: job-shop scheduling; particle swarm optimization; multiple objectives ____________________________________________ *Corresponding author. Email: hsinhung@gmail.com Introduction The job-shop scheduling problem (JSP) has been studied for more than 50 years in both academic and industrial environments. Jain et al. provided a concise overview of JSPs over the last few decades and highlighted the main techniques [10]. The JSP is the most difficult class of combinational optimization. Garey et al. demonstrated that JSPs are non-deterministic polynomial-time hard (NP-hard) [7]; hence we cannot find an exact solution in a reasonable computation time. The single-objective JSP has attracted wide research attention. Most studies of single-objective JSPs result in a schedule to minimize the time required to complete all jobs, i.e., to minimize the makespan (Cmax). Many approximate methods have been developed to overcome the limitations of exact enumeration techniques. These approximate approaches include simulated annealing (SA) [17], tabu search [18][19][24] and genetic algorithms (GA) [1][9][12][27]. However, real-world production systems require simultaneous achievement of multiple objective requirements. This means that the academic concentration of objectives in the JSP must been extended from single to multiple. Recent related JSP research with multiple objectives is summarized as below. Ponnambalam has offered a multi-objective GA to derive optimal machine-wise priority dispatching rules for resolving job-shop problems with objective functions that consider minimization of makespan, total tardiness, and total machine idle time[20]. Ponnambalam’s multi-objective genetic algorithm (MOGA) has been tested with various published benchmarks, and is capable of providing optimal or near-optimal solutions. A Pareto front provides a set of best solutions to determine the tradeoffs between the various objects, and good parameter settings and appropriate representations can enhance the behavior of an evolution algorithm. Esquivel et al. studied the influence of distinct parameter combinations as well as different chromosome representations [5]. Initial results showed that: (i) larger numbers of generations favor the building of a Pareto front because the search process does not stagnate, even though it may be rather slow, (ii) multi-recombination helps to speed the search and to find a larger set size when seeking the Pareto optimal set, and (iii) operation-based representation is better than priority-list and job-based representation selected for contrast under recombination methods. The Pareto archived simulated annealing (PASA) method, a meta-heuristic procedure based on the SA algorithm, was developed by Suresh to find non-dominated solution sets for the JSP with the objectives of minimizing the makespan and the mean flow time of jobs[25]. The superior performance of the PASA can be attributed to the mechanism it uses to accept the candidate solution. Candido et al. addressed JSPs with numbers of more realistic constraints, such as jobs with several subassembly levels, alternative processing plans for parts and alternative resources of operations, and the requirement for multiple resources to process an operation [3]. The robust procedure worked well in all problem instances and proved to be a promising tool for solving more realistic JSPs. Lei first designed a crowding-measure-based multi-objective evolutionary algorithm (CMOEA) that makes use of the crowding measure to adjust the external population and assign different fitness for individuals [14]. Compared to the strength Pareto evolutionary algorithm, CMOEA performs well in job-shop scheduling with two objectives including minimization of makespan and total tardiness. One of the latest evolutionary techniques for unconstrained continuous optimization is particle swarm optimization (PSO) proposed by Kennedy et al. [11]. PSO has been successfully used in different fields due to its ease of implementation and computational efficiency. Even so, application of PSO to the combination optimization problem is rare. Coello et al. provided an approach in which Pareto dominance is incorporated into PSO to allow the heuristic to handle problems with several object functions [4]. The algorithm uses a secondary repository of particles to guide particle flight. That approach was validated using several test functions and metrics drawn from the standard literature on evolutionary multi-objective optimization. The results show that the approach is highly competitive. Liang et al. invented a novel PSO-based algorithm for JSPs[16]. That algorithm effectively exploits the capability of distributed and parallel computing systems, with simulation results showing the possibility of high-quality solutions for typical benchmark problems. Lei presented a PSO for the multi-objective JSP to minimize makespan and total job tardiness simultaneously [15]. Job-shop scheduling can be converted into a continuous optimization problem by constructing the corresponding relationship between a real vector and a chromosome obtained using the priority rule-based representation method. The global best position selection is combined with crowding-measure-based archive maintenance to design a Pareto archive PSO. That algorithm is capable of producing a number of high-quality Pareto optimal scheduling plans. Hybrid algorithms that combine different approaches to build on their strengths have led to another branch of research. Wang et al. combined GA with SA in a hybrid framework, in which the GA was introduced to present a parallel search architecture, and SA was used to increase the probability of escape from local optima at high temperatures [27]. Computer simulation results showed that the hybrid strategy was very effective and robust, and could find optima for almost all benchmark instances. Xia et al. developed an easily implemented approach for the multi-objective flexible JSP based on the combination of PSO and SA [28]. They demonstrated that their proposed algorithm was a viable and effective approach to the multi-objective flexible JSP, especially for large-scale problems. Ripon extended the idea in the jumping genes genetic algorithm, a hybrid approach capable of searching for near-optimal and non-dominated solutions with better convergence by simultaneously optimizing criteria [21]. Previous literature indicates that there has been little study of the JSP with multiple objectives. In this study, we use a new evolutionary PSO technique to solve the JSP with multiple objectives. Job-shop scheduling problem A typical JSP can be formulated as follows. There are n jobs to be processed through m machines. Each job must pass through each machine once and only once. Each job should be processed through the machines in a particular order, and there are no precedence constraints among the different job operations. Each machine can perform only one job at a time, and it cannot be interrupted. In addition, the operation time is fixed and known in advance. The objective of the JSP is to find a schedule to minimize the time required to complete all jobs, that is, to minimize the makespan Cmax. In this study, we attempt to attain the three objectives (i.e., minimizing makespan, machine idle time, and total tardiness) simultaneously. We formulate the multi-objective JSP using the following notation: n is the total number of jobs to be scheduled m is the total number of machines in the process t(i, j) is the processing time for job i on machine j (i=1,2,…n), (j=1,2,…m) Li is the lateness of job i {π1, π2, …, πn} is the permutation of jobs The objectives considered in this paper are formulated as follows: Completion time (makespan) ),( jC π )1,()1,( 11 ππ tC = (1) nitCC iii ,...,2 )1,()1,()1,( 1 =+= − πππ (2) mjjtjCjC ,...,2 ),()1,(),( 11 =+−= πππ (3) mjnijtjCjCjC iiii ,...,2 ;,...,2 ),()}1,(),,(max{),( 1 ==+−= − ππππ (4) Makespan, ),(max mCf nC π= (5) Total tardiness, ] ,0max[ 1 ∑= = n i itardinesstotal Lf (6) Total idle time, }...2|}}0),,()1,({max{)1,({ 2 11 ∑ =−−+−= = − n i iitimeidletotal mjjCjCjCf πππ (7) PSO background PSO is based on observations of the social behavior of animals, such as birds in flocks or fish in schools, as well as on swarm theory. The population consisting of individuals or particles is initialized randomly. Each particle is assigned with a randomized velocity according to its own movement experience and that of the rest of the population. The relationship between the swarm and particles in PSO is similar to the relationship between the population and chromosomes in a GA. In PSO, the problem solution space is formulated as a search space. Each particle position in the search space is a correlated solution to the problem. Particles cooperate to determine the best position (solution) in the search space (solution space). Suppose that the search space is D-dimensional and there are ρ particles in the swarm. Particle i is located at position Xi={x1 i, x2 i, …, xD i} and has velocity Vi={v1 i, v2 i, …, vD i}, where i=1, 2, …,ρ. Based on the PSO algorithm, each particle move towards its own best position (pbest), denoted as Pbesti={pbest1 i, pbest2 i,…, pbestn i}, and the best position of the whole swarm (gbest) is denoted as Gbest={gbest1, gbest2, …, gbestn} with each iteration. Each particle changes its position according to its velocity, which is randomly generated toward the pbest and gbest positions. For each particle r and dimension s, the new velocity vs r and position xs r of particles can be calculated by the following equations: )]1()1([)]1()1([)1()( 2211 −−−××+−−−××+−×= ττττττ r s r s r s r s r s r s xgbestrandcxpbestrandcvwv (8) )1()1()( −+−= τττ rs r s r s vxx (9) In Eqs. (8) and (9), τ is the iteration number. The inertial weight w is used to control exploration and exploitation. A large w value keeps the particles moving at high velocity and prevents them from becoming trapped in local optima. A small w value ensures a low particle velocity and encourages particles to exploit the same search area. The constants c1 and c2 are acceleration coefficients to determine whether particles prefer to move closer to the pbest or gbest positions. The rand1 and rand2 are two independent random numbers uniformly distributed between 0 and 1. The termination criterion of the PSO algorithm includes a maximum number of generations, a designated value of pbest, and lack of further improvement in pbest. The standard PSO process is outlined as follows: Step 1: Initialize a population of particles with random positions and velocities in a D-dimensional search space. Step 2: Update the velocity of each particle using Eq. (8). Step 3: Update the position of each particle using Eq. (9). Step 4: Map the position of each particle into the solution space and evaluate its fitness value according to the desired optimization fitness function. Simultaneously update the pbest and gbest positions if necessary. Step 5: Loop to Step 2 until the termination criterion is met, usually after a sufficient good fitness or a maximum number of iterations. The original PSO was designed for a continuous solution space. We must modify the PSO position representation, particle velocity, and particle movement so they work better with combinational optimization problems. These changes are described in next section. Proposed method There are four types of feasible schedules in JSPs, including inadmissible, semi-active, active, and non-delay. The optimal schedule is guaranteed to be an active schedule. We can decode a particle position into an active schedule employing Giffler and Thompson’s [8] heuristic. There are two different representations of particle position associated with a schedule. The results of Zhang [29] demonstrated that permutation-based position representation outperforms priority-based representation. While choosing to implement permutation-based position presentation, we must also adjust the particle velocity and particle movement. In addition, we also propose the maintenance of Pareto optima and a diversification procedure to achieve better performance. Position representation In this study, we randomly generated a group of particles (solutions) represented by a permutation sequence that is an ordered list of operations. For an n-job m-machine problem, the position of particle k can be represented by an m×n matrix, i.e.,               = k mn k m k m k n kk k n kk k xxx xxx xxx X ... ... ... 21 22221 11211 MMM , where kijx denotes the priority of operation ijo , which means the operation of job j that must be processed on machine i. The Giffler and Thompson (G&T) algorithm is briefly described below. Notation: (i,j) is the operation of job j that must be processed on machine i S is the partial schedule that contains scheduled operations Ω is the set of operations that can be scheduled s(i,j) is the earliest time at which operation (i,j) belonging to Ω can be started. p(i,j) is the processing time of operation (i,j). f(i,j) is the earliest time at which operation (i,j) belonging to Ω can be finished, f(i,j) = s(i,j) + p(i,j) . G&T algorithm: Step 1: Initialize φ=S ; Ω to contain all operations without predecessors. Step 2: Determine }{ min ),(),( * jiji ff Ω∈= and the machine m * on which f* can be realized. Step 3: (1) Identify the operation set Ω∈′′ ) ,( ji such that ) ,( ji ′′ requires machine m*, and *),( fs ji <′′ (2) Choose (i, j) from the operation set identified in Step 3(1) with the largest priority. (3) Add (i, j) to S. (4) Assign s(i,j) as the starting time of (i, j). Step 4: If a complete schedule has been generated, stop. Otherwise, delete (i, j) from Ω, include its immediate successor in Ω, and then go to Step 2. Table 1 shows the mechanism of the G&T algorithm using a 2×2 example. The position of particle k is       = 21 12kX . Initialization Step 1: φ=S ; Ω={(1, 1), (2, 2)}. Iteration 1 Step 2: s(1,1)=0, s(2,2)=0, f(1,1)=5, f(2,2)=4; f *=min{f(1,1),f(2,2)}=4, m *=2. Step 3: Identify the operation set {(2, 2)}; choose operation (2, 2) that has the largest priority, and add it into schedule S. Step 4: Update Ω={(1,1), (1,2)}; go to Step 2. Iteration 2 Step 2: s(1,1)=0, s(1,2)=4, f(1,1)=5, f(1,2)=7; f *=min{f(1,1),f(1,2)}=5, m *=1. Step 3: Identify the operation set {(1, 1), (1, 2)}; choose operation (1, 2) that has the largest priority, and add it into schedule S. Step 4: Update Ω={(1, 1)}; go to Step 2. Iteration 3 Step 2: s(1,1)=7, f(1,1)=12; f *=min{f(1,1)}=12, m *=1. Step 3: Identify the operation set {(1, 1)}; choose operation (1, 1) that has the largest priority, and add it into schedule S. Step 4: Update Ω={(2, 1)}; go to Step 2. Iteration 4 Step 2: s(2,1)=12, f(2,1)=16; f *=min{f(2,1)}=16, m *=2. Step 3: Identify the operation set {(2, 1)}; choose operation (2, 1) that has the largest priority, and add it into schedule S. Step 4: A complete schedule has been generated, so stop the process. The proposed PSO differs from the original PSO in the information stored in the pbest and gbest solutions. While the original PSO keeps the best positions found so far, the proposed PSO maintains the best schedule generated by the G&T algorithm. In the previous example, the schedule Sk rather than the position Xk is retained in the pbest and gbest solutions, where Sk is       12 12 . The movement of particles is modified in accordance with the representation of particle position based on the insertion operator. Particle velocity The original PSO velocity concept assumes that each particle moves according to the velocity determined by the distance between the previous position of the particle and the gbest (pbest) solution. The two major purposes of the particle velocity are to keep the particle moving toward the gbest and pbest solutions, and to maintain inertia to prevent particles from becoming trapped in local optima. In the proposed PSO, we concentrate on preventing particles from becoming trapped in local optima rather than moving them toward the gbest (pbest) solution. If the priority value is increased or decreased by the present velocity in the current iteration, we keep the priority value increasing or decreasing at the beginning of the next iteration with probability w, which is the inertial weight in PSO. The larger the value of w, the more the iteration priority value keeps increasing or decreasing, and the more the difficult it is for the particle to return to its current position. For an n-job problem, the velocity of particle k can be represented as . particle of of velocity theis where},1,0,1{ ], ... [ 21 kjvvvvvV i k i k i k n kkk −∈= The initial velocity of particles is generated randomly. Instead of considering the distance from kix to )( i k i gbestpbest , our PSO considers whether the value of k ix is larger or smaller than )( i k i gbestpbest . If k ix decreases in the present iteration, this mean that )( i k i gbestpbest is smaller than k ix and k ix is set moving toward )( i k i gbestpbest by letting k iv � –1. Therefore, in the next iteration, k ix is kept decreasing by one (i.e., k ix � k ix –1) with probability w. Conversely, if k ix increases in this iteration, then )( i k i gbestpbest is larger than k ix , and k ix is set moving toward )( i k i gbestpbest by setting k iv � 1. Therefore, in the next iteration, k ix is kept increasing by one (i.e., kix � k ix + 1) with probability w. The inertial weight w influences the velocity of the particles in the PSO. We randomly update velocities at the beginning of the iteration. For each particle k and operation ji , if kiv does not equal to 0, k iv will be set to 0 with probability (1–w). This forces kix to stop increasing or decreasing continuously in this iteration with probability (1–w) while kix keeps increasing or decreasing. Particle movement The particle movement is based on the swap operator proposed by Sha et al. [22][23]. Notation: k ix is the schedule list at machine i of particle k. k ipbest is the schedule list at machine i of the kth pbest solution. igbest is the schedule list at machine i of the gbest solution. c1 and c2 are constants between 0 and 1 such that 121 ≤+ cc . The swap procedure occurs as shown below. Step 1: Randomly choose a position ζ from kix . Step 2: Mark the job on position ζ of kix by Λ1. Step 3: If the random number rand < c1 then seek the position of Λ1 in k ipbest ; otherwise, seek the position of Λ1 in igbest . Denote the position that has been found in k ipbest or igbest by ζ′, and the job in position ζ′ of kix by Λ2. Step 4: If Λ2 has been denoted, 0 1 =k iJ v , and 0 2 =k iJ v , then swap Λ1 and Λ2 in k ix , 1 1 ←k iJ v . Step 5: If all the positions of kix have been considered, then stop. If not, and if ζ < n, then ζ←ζ+1; otherwise, ζ← 1. Go to Step 2. For example, consider the 6-job problem where kix =[4 2 1 3 6 5], k ipbest =[1 5 4 2 6 3], igbest =[3 2 6 4 5 1], k iv =[0 0 1 0 0 0], c1=0.6, and c2=0.2. Step 1: The position of kix is randomly chosen: ζ=3. Step 2: The job in the 3rd position of kix is job 1, i.e., Λ1=1. Step 3: A random number rand is generated; assume rand=0.7. Since rand > c1, we compare each position of igbest with Λ1 and the matched position ζ′=6. The job in the 6th position of k ix is job 5, i.e., Λ2=5. Step 4: Since 04 = k iv and 05 = k iv , swap jobs 1 and 5 in k ix so k ix =[4 2 5 3 6 1]. Then let 14 ← k iv and k iv =[0 0 1 1 0 0]. Step 5: Let ζ← 4 and go to Step 2. Repeat the process until all positions of kix have been considered. Diversification strategy If all the particles have the same non-dominated solutions, they will be trapped in local optima. To prevent this from happening, a diversification strategy is proposed to keep the non-dominated solutions different. Once any new solution is generated by particles, the non-dominating solution set will be updated in these three situations: (i) If the solution of the particle dominates the gbest solution, assign the particle solution to the gbest. (ii) If the solution of the particle equals to any solution in the non-dominated solution set, replace the non-dominated solution with the particle solution. (iii) If the solution of the particle is dominated by the worst non-dominated solution and not equal to any non-dominated solution, set the worst non-dominated solution equal to the particle solution. Computational results The proposed multi-objective PSO (MOPSO) algorithm was tested on benchmark problems obtained from the OR-Library [2][26]. The program was coded in Visual C++ and run 40 times on each problem on a Pentium 4 3.0-GHz computer with 1 GB of RAM running Windows XP. During the pilot experiment, we used four swarm sizes N (10, 30, 60, and 80) to test the algorithm. The outcome of N=80 was best, so that value was used in all further tests. Parameters c1 and c2 were tested at various values in the range 0.1–0.7 in increments of 0.2. The inertial weight w was reduced from wmax to wmin during iterations, where wmax was set to 0.5, 0.7, and 0.9, and wmin was set to 0.1, 0.3, and 0.5. The combination of c1=0.7, c2=0.1, wmax=0.7 and wmin=0.3 gave the best results. The maximum iteration limit was set to 60 and the maximum archive size was set to 80. The MOGA proposed by Ponnambalam et al. [19] was chosen as a baseline against which to compare the performance of our PSO algorithm. The objectives considered in the MOGA algorithm are minimization of makespan, minimization of total tardiness, and minimization of machine idle time. The MOGA methodology is based on the machine-wise priority dispatching rule (pdr) and the G&T procedure [8]. The each gene represents a pdr code. The G&T procedure was used to generate an active feasible schedule. The MOGA fitness function is the weighted sum of makespan, total tardiness, and total idle time of machines with random weights. The computation results showed that the relative error of the solution for Cmax and total idle time determined by the proposed MOPSO was better in 23 out of 23 problems than the MOGA. In 22 of the 23 problems, the proposed PSO performed better for the solution considering total tardiness. Overall, the proposed MOPSO was superior to the MOGA in solving the JSP with multiple objectives. Conclusion While there has been a large amount of research into the JSP, most of this has focused on minimizing the maximum completion time (i.e., makespan). There exist other objectives in the real world, such as the minimization of machine idle time that might help improve efficiency and reduce production costs. PSO, inspired by the behavior of birds in flocks and fish in schools, has the advantages of simple structure, easy implementation, immediate accessibility, short search time, and robustness. However, few applications of PSO to multi-objective JSPs can be found in the literature. Therefore, we presented a MOPSO method for solving the JSP with multiple objectives, including minimization of makespan, total tardiness, and total machine idle time. The original PSO was proposed for continuous optimization problems. To make it suitable for job-shop scheduling (i.e., a combinational problem), we modified the representation of particle position, particle movement, and particle velocity. We also introduced a mutation operator and used a diversification strategy. The results demonstrated that the proposed MOPSO could obtain more optimal solutions than the MOGA. The relative error ratios of each problem scenario in our MOPSO algorithm were less than in the MOGA. The performance measure results also revealed that the proposed MOPSO algorithm outperformed MOGA in simultaneously minimizing makespan, total tardiness, and total machine idle time. We will attempt to apply MOPSO to other shop scheduling problems with multiple objectives in future research. Other possible topics for further study include the modification of the particle position representation, particle movement, and particle velocity. In addition, issues related to Pareto optimization, such as solution maintenance strategy and performance measurement, merit future investigation. Acknowledgments This study was supported by a grant from the National Science Council of Taiwan (NSC-96-2221-E-216-052MY3). Appendices Pseudo-code of the PSO for the multi-objective JSP is as follows. Initialize a population of particles with random positions. for each particle k do Evaluate Xk (the position of particle k) Save the pbestk to optimal solution set S end for Set gbest solution equal to the best pbestk repeat Updates particles velocities for each particle k do Move particle k Evaluate Xk Update gbest, pbest, and S end for until maximum iteration limit is reached References [1]Bean, J., 1994. “Genetic algorithms and random keys for sequencing and optimization,” Operations Research Society of America (ORSA) Journal on Computing, 6, 154–160. [2]Beasley J.E., 1990. "OR-Library: distributing test problems by electronic mail", Journal of the Operational Research Society 41(11) pp1069-1072. [3]Candido, M. A. B., Khator, S.K. & Barcia, R.M., 1998. “A genetic algorithm based procedure for more realistic job shop scheduling problems,” International Journal of Production Research, 36(12), 3437–3457. [4]Coello, C.A., Plido, G.T. & Lechga, M.S., 2004. “Handling multiple objectives with particle swarm optimization,” IEEE Transactions on Evolutionary Computation, 8(3), 256–278. [5]Esquivel, S.C., Ferrero, S.W. & Gallard, R.H., 2002. “Parameter settings and representations in Pareto-based optimization for job shop scheduling,” Cybernetics and Systems: An international Journal, 33, 559–578. [6]Fisher, H. & Thompson, G. L., 1963. Industrial Scheduling, Englewood Cliffs, NJ: Prentice-Hall. [7]Garey, M. R., Johnson, D. S. & Sethi, R., 1976. “The complexity of flowshop and jobshop scheduling,” Mathematics of Operations Research, 1, 117–129. [8]Giffler, J. & Thompson, G. L., 1960. “Algorithms for solving production scheduling problems,” Operations Research, 8, 487–503. [9]Gonçalves, J. F., Mendes, J. J. M. & Resende, M. G. C., 2005. “A hybrid genetic algorithm for the job shop scheduling problem,” European Journal of Operational Research, 167(1), 77–95. [10]Jain, A.S. & Meeran, S., 1999. “Deterministic job-shop scheduling: Past, present and future,” European Journal of Operational Research, 113, 390–434. [11]Kennedy, J. and R. Eberhart (1995), Particle swarm optimization. Proceedings of IEEE International Conference on Neural Networks 1995, 1942-1948. [12]Kobayashi, S., Ono, I. & Yamamura, M., 1995. “An efficient genetic algorithm for job shop scheduling problems,” In L. J. Eshelman (Ed.), Proceedings of the Sixth International Conference on Genetic Algorithms (pp. 506–511). San Francisco, CA: Morgan Kaufman Publishers. [13]Lawrence, S., 1984. “Resource constrained project scheduling: An experimental investigation of heuristic scheduling techniques,” Graduate School of Industrial Administration (GSIA), Carnegie Mellon University, Pittsburgh, PA. [15]Lei, D. & Wu, Z., 2006. “Crowding-measure-based multi-objective evolutionary algorithm for job shop scheduling,” International Journal of Advanced Manufacturing Technology, 30, 112–117. [14]Lei, D., 2008. “A Pareto archive particle swarm optimization for multi-objective job shop scheduling,” Computers & Industrial Engineering, 54(4), 960–971. [16]Liang Y.C., Ge, H.W., Zho, Y. & Guo, X.C., 2005. “A particle swarm optimization-based algorithm for job-shop scheduling problems,” International Journal of Computational Methods, 2(3), 419–430. [17]Lourenço, H. R., 1995. “Local optimization and the job-shop scheduling problem,” European Journal of Operational Research, 83, 347–364. [18]Nowicki, E. & Smutnicki, C., 1996. “A fast taboo search algorithm for the job shop problem,” Management Science, 42(6), 797–813. [19]Pezzella, F. & Merelli, E., 2000. “A tabu search method guided by shifting bottleneck for the job shop scheduling problem,” European Journal of Operational Research, 120(2), 297–310. [20]Ponnambalam S. G., Ramkumar, V. & Jawahar, N., 2001. “A multi-objective genetic algorithm for job shop scheduling.” Production Planning and Control, 12(8), 764–774. [21]Ripon, K. S. N., 2007. “Hybrid evolutionary approach for multi-objective job-shop scheduling problem,” Malaysian Journal of Computer Science, 20(2), 183–198. [22]Sha, D.Y. & Hsu, C.-Y., 2006, “A hybrid particle swarm optimization for job shop scheduling problem,” Computers & Industrial Engineering, 51(4), 791–808. [23]Sha, D.Y. & Hsu, C.-Y., 2008. “A new particle swarm optimization for the open shop scheduling problem,” Computers & Operations Research, 35, 3243–3261. [24]Sun, D., Batta, R. & Lin, L., 1995. “Effective job shop scheduling through active chain manipulation,” Computers & Operations Research, 22(2), 159–172. [25]Suresh R.K. & Mohanasndaram, K.M. 2006., “Pareto archived simulated annealing for job shop scheduling with multiple objectives,” International Journal of Advanced Manufacturing Technology, 29, 184–196. [26]Taillard, E.D., 1993. “Benchmarks for basic scheduling problems,” European Journal of Operational Research, 64, 278–285. [27]Wang, L. & Zheng, D.-Z., 2001. “An effective hybrid optimization strategy for job-shop scheduling problems,” Computers & Operations Research, 28, 585–596. [28]Xia, W. & Wu, Z., 2005. “An effective hybrid optimization approach for multi-objective flexible job-shop scheduling problems,” Computers & Industrial Engineering, 48, 409–425. [29]Zhang, H., Li, X., Li H., Hang F. 2005. “Particle swarm optimization-based schemes for resource-constrained project scheduling,” Automation in Construction 14(3), 393–404. Table 1 2×2 example Jobs Machine sequence Processing times 1 1, 2 p(1,2)=5; p(2,1)=4 2 2, 1 p(2,2)=4; p(1,2)=3 Table 2 Comparison of MOGA and MOPSO with three objectives. Benchmark N m Makespan (MOGA) Makespan (MOPSO) % Deviation Machine idle time (MOGA) Machine idle time (MOPSO) % Deviation Total tardiness (MOGA) Total tardiness (MOPSO) % Deviation abz5 10 10 1587 1338 0 8097 3978 0 1948 611 0 abz6 10 10 1369 1046 0 7744 2937 0 1882 339 0 ft06 6 6 76 56 0 259 100 0 31 3 0 ft10 10 10 1496 1045 0 9851 1999 0 3459 1534 0 orb01 10 10 1704 1181 0 11631 3909 0 3052 191 0 orb02 10 10 1284 1029 0 7585 3539 0 1565 137 0 orb03 10 10 1643 1114 0 11138 3788 0 4140 247 0 orb04 10 10 1543 1122 0 9802 3921 0 4951 221 0 orb05 10 10 1323 1013 0 8322 3727 0 2195 30 0 orb06 10 10 1645 1144 0 10836 3478 0 2601 0 0 orb07 10 10 583 302 0 3423 1381 0 699 0 0 orb08 10 10 1340 1000 0 8840 3542 0 3498 253 0 orb09 10 10 1462 1044 0 9439 4224 0 2029 0 0 orb10 10 10 1382 1077 0 8271 4177 0 1806 0 0 la01 10 5 1256 709 0 3431 571 0 3324 721 0 la02 10 5 1066 713 0 2687 573 0 2081 425 0 la03 10 5 821 671 0 1722 633 0 1926 373 0 la04 10 5 861 631 0 1798 557 0 3194 673 0 la05 10 5 893 593 0 2182 473 0 1716 736 0 la16 10 10 1452 1040 0 9169 2718 0 1127 1417 0.25732 la17 10 10 1172 889 0 7044 3365 0 1779 53 0 la19 10 10 1251 938 0 7164 2796 0 1581 733 0 la20 10 10 1419 985 0 8745 2883 0 1451 407 0 The English in this document has been checked by at least two professional editors, both native speakers of English. For a certificate, see: http://www.textcheck.com/cgi-bin/certificate.cgi?id=LbNsng 
work_zwzhhkduqrferjkjh7pv2i6ooa	----	A multi-expert system for ranking patents: An approach based on fuzzy pay-off distributions and a TOPSISﾖAHP framework A multi-expert system for ranking patents: An approach based on fuzzy pay-off distributions and a TOPSIS–AHP framework Mikael Collan a,⇑, Mario Fedrizzi b, Pasi Luukka a a School of Business, Lappeenranta University of Technology, Skinnarilankatu 34, FI-53851 Lappeenranta, Finland b Department of Industrial Engineering, University of Trento, Via Inama 5, I-38122 Trento, Italy a r t i c l e i n f o Keywords: Patents Pay-off method Consensus Possibilistic moments TOPSIS AHP a b s t r a c t The aim of this paper is to introduce a decision support system that ranks patents based on multiple expert evaluations. The presented approach starts with the creation of three value scenarios for each con- sidered patent by each expert. These are used for the construction of individual fuzzy pay-off distribution functions for the patent value; a consensual fuzzy pay-off distribution is then determined starting from the individual distributions. Possibilistic moments are calculated from the consensus pay-off distribution for each patent and used in ranking them with TOPSIS. It is further showed how the analytic hierarchy process (AHP) can be used to include additional decision variables into the patent selection, thus allowing for a two-tier decision making process. The system is illustrated with a numerical example and the usability of the system and the combination of methods it includes for patent portfolio selection in the real world context is discussed. � 2013 Elsevier Ltd. All rights reserved. 1. Introduction Ranking and selection of patents is an important issue from the point of view of intellectual property (IPR) managers everywhere. It is most often a recurring task in companies that commonly have their IPR managers visit the patent and R&D portfolios once or twice per year, analyzing the composition of the portfolios and making decisions about the modification of portfolio composition. New patents may also be considered on continuing basis, empha- sizing the need for tools even further. One way to quantitatively rank and to select patents is to use estimation of their future value from the point of view of the firm owning them as a measure of goodness. Value to the firm may very well be the single most important characteristic of a patent. Other issues that are important in analyzing and ranking patents are most often non-financial and have to do with strategic criteria, such as fit of the patents to the corporate portfolio and to the fu- ture plans of the firm. Generally, we can say that a good ranking is able to consider both of these types of information, financial and non-financial. Commonly there are three main approaches for the valuation of patents, these are the ‘‘cost approach’’, the ‘‘market method’’, and the ‘‘income approach’’ also known as the discounted cash-flow method (DCF) (e.g., see Reilly & Schweihs, 1998; Smith & Parr, 2000). Of these, the cost approach and the market method are meant only for market valuation of patents that is to say, for the derivation of an estimate for a sale price for a patent. The DCF method is based on the well known principles of present value (PV) and the same principles can be used also in the ‘‘in-house’’ valuation of patents, that is, to derive the ‘‘value to the firm’’ of patents. It is important to note that patent analysis is a forward-looking exercise, as patents are an enabling class of assets that is most of- ten used to secure the future of the firms’ business. This means that methods used in the valuation and analysis of patents should be able to take into consideration the (sometimes considerable) esti- mation inaccuracy present in forward-looking estimation, as the estimation of future cash-flows for patents, since it is not realistic to expect anyone to be able to produce precise estimates for future (patent) cash-flows (Karsak, 2006). Using cash-flow scenarios is a widespread practice of modeling the inaccurate and uncertain fu- ture cash-flows, and it can also be applied to patent analysis (Col- lan, Fuller, Mezei, & Wang, 2011). Information to support the creation of cash-flow scenarios can come from systems specifically designed for supporting patent analysis, such as are presented (for example in Camus & Brancaleon, 2003; Huang, Liang, Lin, Tseng, & Chiang, 2011; Littman-Hillmer & Kuckartz, 2009; Park, Kim, Choi, & Yoon, 2013), or it can come directly from experts, most often from within the firm itself. Fuzzy logic is an established way to express imprecision precisely and as such is a usable tool also when patent cash-flows are considered. Fuzzy pay-off method, introduced in Collan, Fullér, and Mezei (2009a, 2009b) and further presented in 0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.02.012 ⇑ Corresponding author. Tel.: +358 505567185. E-mail addresses: mikael.collan@lut.fi (M. Collan), mario.fedrizzi@unitn.it (M. Fedrizzi), pasi.luukka@lut.fi (P. Luukka). Expert Systems with Applications 40 (2013) 4749–4759 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a http://dx.doi.org/10.1016/j.eswa.2013.02.012 mailto:mikael.collan@lut.fi mailto:mario.fedrizzi@unitn.it mailto:pasi.luukka@lut.fi http://dx.doi.org/10.1016/j.eswa.2013.02.012 http://www.sciencedirect.com/science/journal/09574174 http://www.elsevier.com/locate/eswa Collan (2012) is a tool for investment analysis and is based on using cash-flow scenarios to create an asset’s pay-off distribution that is considered as a fuzzy number. The fuzzy pay-off method can be employed in the valuation of patents (Collan & Heikkilä, 2011). As already observed above, patent analysis is a forward-looking procedure and there may be differing views about the direction that the future will take. This observation can be interpreted in the way that it makes sense to include more than one expert opin- ion when patent analysis is done. This is true for both, for cash- flow information, as well as, for non-cash-flow information. As budgets for patent portfolios are tight, the firm can afford to keep only the best patents. This calls for the ranking of the patents as a basis of selection. It is a fair assumption that the value to the firm is a key driver in the selection of patents and can be used as a first basis for patent selection into portfolios. Important other (sec- ondary) considerations may include different non-financial strate- gic selection criteria. This means that one plausible approach to go about with patent selection is to first rank the patents that are competing for a place in the firm’s portfolio based on their value to the firm, second do a pre-selection of a sub group of the best pat- ents, and third do a complementary analysis to narrow down the number of patents to fit the budget, based on the non-financial strategic criteria. In this paper, an approach that enables both the financial and non-financial merits to be included in ranking of patents is pro- posed, while taking into consideration the estimation imprecision, and the differing estimates of multiple experts. This combination is a new contribution that allows a more holistic analysis on patents to be performed. The way in which the methods used are combined is new and new to the field of application. Furthermore, we use possibilistic moments in characterizing fuzzy financial information and rank patents according to the moments, to the best of our knowledge the first proposed approach of its kind. The remainder of the paper is organized as follows. In Section 2 the general framework of the system for performing ranking and selection of patents is presented. In Section 3 we continue by pre- senting the construction of pay-off distributions from cash-flow scenarios by each expert for each patent takes place. In Section 4 the consensus modeling mechanism to be used to build consensus pay-off distributions from each expert’s pay-off distributions is introduced. Section 5 starts with the definition of possibilistic mo- ments of fuzzy pay-off distributions and continues with the description of the main steps of TOPSIS, used then for producing a preliminary ranking of patents. In Section 6, after a short presen- tation of AHP we show how it can be used to include strategic cri- teria in ranking the patents. In Section 7 the two-tier process is illustrated with a numerical example that includes the selection of four patents out of twenty candidate patents. Finally, the paper is closed with a discussion and some conclusions. 2. A blueprint for a multi-expert consensus reaching system for supporting patent selection The focus here is to present a system for supporting investment decision-making with regards to patents that is, the selection of patents. The circumstances under which the system is usable are such that there are a number of possible patents that are compet- ing for funding (inclusion in a portfolio) under a budget constraint. The system is based on using three scenarios of managerial cash- flow estimates for each patent, these cash-flow scenarios are used in the creation of a pay-off distribution for each patent, by each expert. The pay-off distributions for the patents by different experts are likely to be different from each other, and in order to get an overall single pay-off distribution for each patent a consensus among the experts’ pay-off distributions must be reached. For this a consensus facilitating method is used and a single consensus pay-off distribu- tion for each patent is created. From the consensus pay-off distribution for each patent, three possibilistic moments are calculated: the possibilistic mean, the possibilistic standard deviation, and the possibilistic skewness. The calculated possibilistic moments are then used in a TOPSIS ranking of the patents. This ranking is based on the cash-flow information for each patent and thus depends on the perceived va- lue for each patent. The TOPSIS ranking can be used as a basis of a first selection of patents. Here it is however proposed that more information is in- cluded into the selection by continuing the analysis with AHP for the selected number of best patents as ranked by TOPSIS. The AHP analysis is carried out by a team of ‘‘elected’’ decision-makers, based on ‘‘strategic’’ criteria that take into consideration different, non financial, aspects of the patents. The approach is illustrated graphically in Fig. 1. The method can also be described as the following process: (i) Each expert creates three cash-flow scenarios for each investment alternative: ‘‘maximum possible’’, ‘‘minimum possible’’, and ‘‘best estimate’’ scenarios. (ii) From each experts’ scenarios an individual fuzzy value dis- tribution function (pay-off distribution) is created. (iii) Consensus pay-off distribution is determined from the mul- tiple experts’ pay-off distributions. (iv) Values of three possibilistic moments are calculated for each patent from the consensus pay-off distributions. These are the possibilistic mean, the possibilistic standard deviation, and the possibilistic skewness of the pay-off distribution. (v) The calculated possibilistic moments are used in a ranking of the patents with TOPSIS. (vi) ‘‘Elected’’ decision-makers perform an AHP process, based on relevant strategic criteria, to create the final ranking of the patents. The resulting ranking of the patents can be used in supporting the patent portfolio selection, a problem that has been considered, for example in Hassanzadeh, Collan, and Modarres (2012). The steps of the approach, with background information, are explained in more detail in the following sections. 3. Creation of cash-flow scenarios and construction of fuzzy pay-off distributions Using scenarios is a widespread practice of modeling the uncer- tain future of projects and assets under imprecise information. The idea with scenarios is that different future scenarios are thought out according to different possible future ‘states of the world’ and cash flows or value connected to these states, are estimated. Creating scenarios for patent alternatives can be done based on the available information about the future (markets, technology, and other issues); the information need not be precise, because the scenarios allow for even a very wide variation of the states of the world/value. The information used in creating the scenarios can come from qualitative information gathered and even from existing patent/IPR analysis/management systems (Jain, Murty, & Flynn, 1999; Littman-Hillmer & Kuckartz, 2009). Scenarios can be used to complement all three of the above mentioned patent valu- ation method categories. When the DCF method is used the scenarios normally include the estimation of (yearly) cost and revenue cash-flows for the dif- ferent scenarios. The yearly cash-flows are estimated by managers 4750 M. Collan et al. / Expert Systems with Applications 40 (2013) 4749–4759 https://isiarticles.com/article/6322 
work_ztyva33ixbabngsnl44o3vw36i	----	06 April 2021 AperTO - Archivio Istituzionale Open Access dell'Università di Torino Original Citation: Trace retrieval for business process operational support Published version: DOI:10.1016/j.eswa.2015.12.002 Terms of use: Open Access (Article begins on next page) Anyone can freely access the full text of works made available as "Open Access". Works made available under a Creative Commons license can be used according to the terms and conditions of said license. Use of all other works requires consent of the right holder (author or publisher) if not exempted from copyright protection by the applicable law. Availability: This is the author's manuscript This version is available http://hdl.handle.net/2318/1557793 since 2016-03-11T10:58:22Z Trace Retrieval for Business Process Operational Support Alessio Bottrighia, Luca Canensib, Giorgio Leonardia,∗, Stefania Montania, Paolo Terenziania aDISIT, Computer Science Institute, Università del Piemonte Orientale, Alessandria, Italy bDepartment of Computer Science, Università di Torino, Italy Abstract Operational support assists users while process instances are being exe- cuted, by making predictions about the instance completion, or recommend- ing suitable actions, resources or routing decisions, on the basis of the already completed instances, stored as execution traces in the event log. In this paper, we propose a case-based retrieval approach to business pro- cess management operational support, where log traces are exploited as cases. Once past traces have been retrieved, classical statistical techniques can be applied to them, to support prediction and recommendation. The framework enables the user to submit queries able to express complex patterns exhibited by the current process instance. Such queries can be composed by several simple patterns (i.e., single actions, or direct sequences of actions), separated by delays (i.e., actions we do not care about). Delays can also be imprecise (i.e., the number of actions can be given as a range). The tool also relies on a tree structure, adopted as an index for a quick retrieval from the available event log. Our approach is highly innovative with respect to the existing litera- ture panorama, since it is the first work that exploits case-based retrieval techniques in the operational support context; moreover, the possibility of ∗Corresponding author. Tel. +39 0131 360340 Email addresses: alessio.bottrighi@uniupo.it (Alessio Bottrighi), canensi@di.unito.it (Luca Canensi), giorgio.leonardi@mfn.unipmn.it (Giorgio Leonardi), stefania.montani@uniupo.it (Stefania Montani), terenz@di.unito.it (Paolo Terenziani) Preprint submitted to Expert Systems with Applications October 1, 2015 retrieving traces by querying complex patterns and the indexing strategy are major departures also with respect to other existing trace retrieval tools proposed in the Case Based Reasoning area. Thanks to its characteristics and methodological solutions, the tool im- plements operational support tasks in a flexible and efficient way, as demon- strated by our experimental results. Keywords: Trace retrieval, Case Based Reasoning, Operational Support 1. Introduction Operational support is a process management activity meant to assist users while process instances are being executed (see der Aalst (2011), chap- ter 9). If an instance is still running, it is possible that information about already completed instances can be exploited to ensure the correct or effi- cient handling of the current instance itself. Completed instances are stored in the so-called event log, which records the sequences (traces henceforth) of actions executed at a given organization, typically together with action execution parameters (e.g., times, costs, resources). Specifically, operational support can be articulated into three main tasks: • detection: compares an existing process model to the current instance data (i.e., the information about the already executed actions in the current trace). If the model control flow (or other predefined rule) is violated, an alert is generated; • prediction: exploits the event log and the current instance to make predictions about the instance completion (e.g., time to completion, costs, use of resources, possible problems); • recommendation: uses the event log and the current instance to recom- mend suitable actions, resources or routing decisions (in order, e.g., to minimize costs or time to completion). While detection is more concerned with constraint satisfaction and with the exploitation of an existing normative model, prediction and recommen- dation heavily rely on experiential knowledge, stored in the event log in the form of past process traces. 2 Traditional operational support tools, like those in the ProM (van Dongen et al., 2005) framework, operate by building a transition system and by replaying log traces on it (van Dongen, 2007). However, also a direct use of the experiential knowledge in the log can provide crucial contributions to support prediction and recommendation. In particular, analogical reasoning (as enabled, e.g., by the Case Based Reason- ing (CBR) (Aamodt & Plaza, 1994) methodology) can be adopted to this end. The CBR methodology is articulated into four steps: (1) retrieve past experiential knowledge, in the form of past “cases”, similar to the current input situation; (2) reuse the retrieved case solution(s) (possibly after an adaptation procedure), to solve the input situation; (3) revise the newly solved case for correctness; (4) retain the new case, to enrich the system knowledge base. In this paper, we propose a case-based retrieval framework for opera- tional support. We therefore focus on the first step of the CBR methodology. In detail, we represent log traces as cases, and, given an ongoing process instance as an input, we retrieve past traces similar to it. Once past traces have been retrieved, classical statistical techniques can be applied to them, to support prediction and recommendation. For instance, the percentage of retrieved traces that, e.g., were completed on time, can be used to calculate the probability that the current instance will complete on time too. A similar approach can be adopted to estimate costs, or predict problems. Moreover, the best actions to execute next can also be extracted from the retrieved traces. In the following, we will however concentrate on trace retrieval, being the discussion on statistical techniques out of the scope of this paper. The approach we propose is, to the best of our knowledge, the first one to provide such a direct exploitation of experiential knowledge in the context of operational support. Notably, trace retrieval is, in general, a quite chal- lenging and complex task. The input process instance is a partial execution trace, and the easiest form of retrieval can be achieved by looking for all the traces in the log that contain the input trace as a prefix. However, in many applications, more flexibility is required. Indeed, the user may observe that the input instance contains specific patterns (i.e., consecutive or non- consecutive sequences of actions), able to strongly influence the next actions to be executed, or the overall performances (e.g., completion time); patterns can be rather complex, and involve temporal indeterminacy. Therefore, s/he may want to abstract from the specificities of the input trace, by identifying 3 these patterns, and looking for all the traces that match them, in order to get a recommendation on how to deal with these problems on the basis of past experience. The medical domain, among others, includes many situations where this non-trivial trace retrieval can provide a valuable help to tackle complicated clinical cases. For example, in the treatment of stroke, particular atten- tion must be paid to suspected Subarachnoid Hemorrhage (SAH). The latest clinical guidelines1 suggest performing a Lumbar Puncture (LP) to patients where the SAH is suspected but not confirmed by the proper laboratory ex- aminations, in order to investigate the presence of blood in the cerebrospinal fluids. However, LP is an invasive procedure and not all the patients can tolerate it without paying some consequences. For this reason, the chance of executing it must be carefully evaluated, and the use of information collected from the past similar situations can help in deciding whether or not the exam should be performed. Let us consider a particular case, where a suspicion of SAH arises for a non-epileptic patient who experiences a sudden epileptic attack. Once the patient has been transported to the hospital, specific tests are performed, such as Transcranial Doppler (TD) and a subsequent Cerebral Angiography (CA). If the results of these tests do not clearly indicate the presence of SAH (i.e., there is a positive TD report, then a negative CA report), the clinical guidelines suggest the execution of a Computerized Tomography (CT) scan as a further exam. If also the TC scan report is negative, the differential diagnosis could be conducted through the observation of the spinal fluid, to check whether the blood has penetrated into the spinal liquid. However, as mentioned before, this procedure is invasive and stressful for patients affected by serious conditions, but it could be important in order to gain a definitive and certain answer for assessing the actual presence of SAH. This assessment is also important because it can radically alter the treatment plan to be administered to the patient. In order to estimate more accurately the cost/benefit ratio of performing the SP, it is very useful to look at what has been done in the past, by retrieving all the execution traces which follow a query pattern which starts with an epileptic attack, followed by a positive TD report, then by a negative CA report and finally by a negative TC report. No matter if and when other 1http://www.iso-spread.it/, last accessed on September, the 30th, 2015 4 medical procedures appear among the diagnostic exams mentioned before: stroke patients typically undergo many other routine tests (such as blood test or chest RX) that are not specifically relevant for SAH, but may be interleaved between the key actions. Trace retrieval by querying complex patterns as the one illustrated in this example is supported in our framework. It is then possible to analyze the retrieved traces in order to calculate: the percentage of patients who received the SP; the percentage of patients who tolerated this procedure; how many times the execution of the SP allowed for a correct differential diagnosis, and what therapeutic plans have been set up depending on the different cases. On the basis on this information, an appropriate cost/benefit analysis can be conducted, in order to decide whether or not the SP should be executed, and to obtain some suggestions about the most common treatment plans to be administered to the patient. Interestingly, our framework also relies on a tree structure, called trace tree, allowing for a quick retrieval, by avoiding a flat search for all the traces in the log that satisfy the input pattern. The trace tree is a sort of “model” of the traces, that we learn using a process mining technique we recently implemented (Canensi et al., 2014). As we will see in this paper, the trace tree can be used as an index of the traces in the log, supporting also the search of traces corresponding to complex query patterns. It is worth noting that the work in Canensi et al. (2014) only dealt with the construction of the process model, i.e., the tree structure, but did not propose any exploitation of it as an index for trace retrieval. As such, the research objective of the present paper, fully focused on trace retrieval for operational support, and the content of all the technical sections (except Section 2.1, which summarizes the approach in Canensi et al. (2014)), are completely new with respect to our previous work. In synthesis, in this paper we describe a flexible and efficient case-based retrieval approach, which also allows to query complex pattern to be searched for in the traces. Our approach is highly innovative with respect to the existing literature panorama, in that: • we propose, to the best of our knowledge, the first work that exploits case-based retrieval techniques on the event log in the operational sup- port context; • we not only support exact prefix retrieval, but also non-trivial trace retrieval, in which complex query patterns are looked for; 5 • we take advantage of the trace tree structure to speed up the retrieval process. The latter two points represent major departures also with respect to other existing trace retrieval tools proposed in the CBR area (see Section 4). The paper is organized as follows. In Section 2 we technically describe our approach. For the sake of completeness, we first briefly summarize our algorithm to build the log tree (see Section 2.1), which acts like an index of the traces. We then move to the novel technical contribution of this paper, describing our query language and retrieval algorithms. In Section 3 we present our experimental results. In Section 4 we discuss related work. In Section 5 we present our concluding remarks and future work directions. 2. Methods This Section presents the details of our approach. In particular, in Section 2.1 we describe the trace tree structure, and sketch the algorithm to build it, which was extensively illustrated in Canensi et al. (2014). In Section 2.2, which represents the core novel technical contribution of this paper, we illustrate our retrieval approach. 2.1. Mining the trace tree Our framework relies on a tree structure, called trace tree, allowing for an efficient retrieval of the traces that satisfy the input pattern. The trace tree is a sort of “model” of the traces, that we learn using a process mining technique we recently implemented (Canensi et al., 2014), and built in such way that it can be used as an index. Several approaches to process mining exist in the literature (e.g., those in ProM (van Dongen et al., 2005)), but, despite some differences, many of them show important similarities, and have common limitations (see also (Cnudde et al., 2014)): • they learn “context-free” patterns of processes; • they can mine paths that do not correspond to any input trace in the log (i.e., they can have a limited precision (Buijs et al., 2012)); • they do not explicitly relate the mined patterns to the log (in the sense that there is no explicit correspondence between mined patterns, and the traces in the log “supporting” them). 6 Such limitations are quite relevant in general, and very relevant in some specific domains, such as the medical one. Concerning the first limitation, it is well known that, e.g., the same (set of) actions may produce different effects, depending on the context (e.g., on the medical actions previously performed on the patient). The impact of the second limitation is obvious and dramatic: if the miner precision is limited, in the sense that it may also learn a path that never appears in any input trace, this can be very harmful in all those applications where it is vital that mining results are reliable as much as possible. However, surprisingly, limited precision is a common limitation of many current miners. The third limitation is less critical, but still significant. Indeed, maintain- ing an explicit link between mined patterns and the input traces matching such patterns, can be important not only to characterize contexts, but also to provide users with an evidence of the learned output, and also to provide support for retrieving traces corresponding to a given pattern - which is our current objective. In Canensi et al. (2014), we therefore proposed an innovative approach, able to support support “context-aware” process mining, and overcome all the above limitations. The technical details of the approach are summarized below. Our mining algorithm takes in input an event log. The event log is stored as a matrix with n rows and m columns, where n is the number of traces in the log and m is the maximum length of these traces. Each cell Matrix[i, j] contains the j-th action of the trace i. Actions in the different traces are aligned on the basis of their order of execution (i.e., the j index). All traces start with a dummy common action #. The algorithm outputs the mined process as a trace tree, where nodes represent actions, and arcs represent a control flow (i.e., precedence, XOR choice) relation between them. Indeed, we exploit the temporal ordering of actions in the log traces to mine the process model, maintaining all the local and global ordering relationships between the actions. More precisely, in the trace tree, each node is represented as a pair < P, T >. P denotes a (possibly unary) set of actions; actions in the same node are in AND relation, or, more properly, may occur in any order with respect to each other. Note that, in such a way, each path from the starting node of the tree to a given node N denotes a set of possible process patterns (called support patterns of N henceforth), obtained by following the order represented by 7 the arcs in the path to visit the trace tree, and ordering in each possible way the actions in each node (for instance, the path {A, B} → {C} represents the support patterns “ABC” and “BAC”). T represents a set of pointers to all and only those traces in the log whose prefixes exactly match the path from the root to one of the patterns in P (called support traces henceforth). Specifically, prefixes correspond to the entire traces if the node at hand is a leaf. In the case of a node representing a set of actions to be executed in any order, T is more precisely composed of several sets of support traces, each one corresponding to a possible action permutation. This choice enhances retrieval performances, as we will discuss in section 2.2.2. Algorithm 1 below builds the trace tree. ALGORITHM 1: Mining pseudocode 1 Build-Tree (index,< P, T >) ; 2 nextP ← getNext(index+1, T) ; 3 if nextP not empty then 4 nextActions ← XORvsAND (nextP, T) ; 5 foreach node < P ′, T ′ > ∈ nextActions do 6 AppendSon(< P ′, T ′ >,< P, T >) ; 7 Build−Tree(index+ |P ′|,< P ′, T ′ >) ; 8 end 9 end The function Build-Tree in Algorithm 1 takes in input a variable index, representing a given position in the traces (i.e., a column in the input matrix), and a node. Initially, it is called on the first position, and on the root of the tree (which is a dummy node, corresponding to the # action; thus, initially, index=0, P=# and T is the set of all the traces). The function getNext simply inspects the traces in T to find all possible next actions. On these actions, the function XORvsAND applies the formula below in order to identify which actions are in AND and which are in XOR relation: we calculate the dependency frequency A → B between every action pair < A, B > in nextP ×nextP: A → B = 1 2 ( |A > B|∑ X∈ActT |A > X| + |A > B|∑ Y ∈ActT |Y > B| ) (1) 8 where, always considering the traces in T , |A > B| is the number of traces in which A is immediately followed by B, |A > X| is the number of traces in which A is immediately followed by some action X (with X ∈ ActT , being ActT the set of all the actions appearing in the traces in T), and |Y > B| is the number of traces in which B is immediately preceded by some action Y (with Y ∈ ActT ). After evaluating the dependency frequency value A → B and B → A, we can have 3 possible situations: • if both the values are below a given threshold, this means that A and B rarely appear in the same trace, therefore they are in XOR relation; • if A → B is above the threshold and B → A is below, then A precedes B, and vice versa; • if both the values are above the threshold, then A and B are in AND (any-order) relation. The output nextActions of the function XORvsAND is a set of nodes < P ′, T ′ >, one for each maximal set of actions to be AND-ed. P’ is therefore a set of action in AND (and a subset of P). Note that, for each one of such sets P’, the corresponding set T’ of support traces where these actions in AND take place is also computed, as a subset of T. Finally, each new node is appended in the output tree (function Append- Son), and Build-Tree is recursively applied to each node (with the parameter index properly set, and passing < P ′, T ′ > as the last parameter). 2.2. Trace retrieval In this Section, we first describe our query language. Then, we present the retrieval algorithm, specifying how traces satisfying a query can be retrieved, taking advantage of the trace tree structure. 2.2.1. Query language In our framework, the user can issue a query, composed of one or more simple patterns to be searched for. In turn, simple patterns are defined as one or more actions in direct sequence. Multiple simple patterns can be combined in a complex pattern, by separating them by delays. A delay is a sequence of actions between two simple patterns; the semantics is that we do not care about these actions, so they will not be specified by the query. Instead, only their number will be provided, possibly in an imprecise way 9 (i.e., we allow the user to express the number of actions as a range). Formally, a query is represented in the following format: ⟨(min1, max1)SP1...(mink, maxk)SPk(mink+1, maxk+1)⟩ (2) where: • SPj is a simple pattern (i.e. a sequence of symbols, representing the actions we are looking for; these actions have to be in direct sequence); • (minj, maxj) is the delay between two items (i.e., two simple patterns, or a simple pattern and the trace starting/ending point), expressed as a range in the number of actions. As an example, the query ⟨(0, 0)B(0, 1)E(2, 2)Z(0, 1)⟩ (3) looks for action B, which has to start at the very beginning of the trace. This first simple pattern B must be followed (with zero or a single action in between) by action E. E must be followed by two actions, which we do not care about; after them, Z is required. Z can be the final action, or can be followed by one additional action we do not care about. It is worth noting that a query written as above corresponds to a whole set of queries, each one obtained by choosing a specific delay value and specific actions in each of the (minj, maxj) intervals. Every query in this set can be made partially explicit as a string, contain- ing as many dummy symbols ∗ as needed, to cover the corresponding delay length (where the dummy symbol is chosen because we are not interested in the specific actions). For example, the query above would correspond to the following four par- tially explicit queries, whose length ranges from 5 to 7 actions (not counting the intial dummy action #), where the dummy symbol ∗ has been properly inserted, according to the delay values information: BE ∗∗Z; BE ∗∗Z∗; B ∗E ∗∗Z; B ∗E ∗∗Z∗ Since each∗ could be substituted by any of the N types of actions recorded in the log and/or existing in the application domain, the example query corresponds to N2 + 2∗N3 + N4 totally explicit queries. The problem is obviously combinatorial, with respect to the possible delay ranges and action types. We thus believe that extensional approaches (in 10 which only explicit queries can be issued) would not be feasible in many domains. Our query language, allowing for compact “intensional” queries, is therefore a significant move in the direction of implementing an efficient and user-friendly operational support tool. Notably, exact prefix retrieval can be seen as a special case of the more general retrieval possibilities we offer, where only a single pattern is provided, and no delays are needed. 2.2.2. Retrieval In order to retrieve the log traces that satisfy a query, we have imple- mented a multi-step procedure, articulated as follows: • automaton generation; • tree search; • filtering. In the following, we will provide the details of the various steps, along with a running example based on the example query introduced in section 2.2.1, exploiting the trace tree in Figure 1. Automaton generation In our approach to trace retrieval, we first generate a deterministic automaton, that represents the query at hand. To build the automaton, we implement the following procedure: • (A) transform the query into a regular expression; • (B) apply the Berry & Sethi (1986) algorithm, to build a non-deterministic automaton that recognizes the regular expression above; • (C) unfold the non-deterministic automaton; • (D) transform the unfolded non-deterministic automaton into a deter- ministic automaton (Lam et al., 2006). Steps (A) and (D) are trivial. As regards Step (A) note that our query language is just a variation of regular expressions, useful to express delays and “do not care” (i.e., dummy) symbols in a compact way. The cost of Step (A) is linear in the number of delays used in the query. Steps (B) and (C) use classical algorithms in the area of formal languages. The cost of Step (B) is linear in the number of symbols in the query expressed as a regular 11 Figure 1: Trace tree in the example. 12 10 9 0 1 # 2 B 3 ε 4* E 5 E 6 * 7 * 8 Z * ε Figure 2: Non deterministic automaton in the example. On the edges: symbol consumed in the corresponding transition; ∗ means that any symbol can be consumed; ε means that no symbol is consumed. expression (Berry & Sethi, 1986) (i.e., the output of Step (A)), and the cost of Step (C) is the product between the number of dummy symbols in the query and the cardinality of the action symbols available in the log. Step (D) substitutes each arc labeled by the dummy symbol in the automaton with a set of arcs, one for each action in the event log. Although in the worst case Step (D) is exponential with respect to the number of states in the automaton (i.e., the output of Step (B)), note that the worst case is rare in practice (van Leeuwen, 1994). Example Referring to our example query, the generated non-deterministic automaton (see Fig. 2) has 9 states and 10 transitions; the state 0 is the start state and states 8 and 9 (double circled in Fig. 2) are the accepting states. Tree search Once the deterministic automaton has been obtained, it would be possible to exploit it in a classical way, by providing all event log traces in input to it, to verify which of them satisfy the query. However, some of these traces may be identical, or share common prefixes of various length, so that the straightforward approach would lead to repeated analyses of the common parts. In order to optimize efficiency, we have therefore proposed a novel ap- proach, that provides the trace tree as an input to the automaton. Each path in the trace tree may index several identical support traces, that will be considered only once, thus speeding up retrieval with respect to a flat search into the event log. Moreover, in the tree, common prefixes of different traces are represented just once, as common branches close to the root (different postfixes can then stem from the common branches, to reach the various leaves). These common parts will be executed on the automaton only once, 13 without requiring for repeated, identical checks. It is worth noting that providing a tree as an input to the automaton represents a significantly novel contribution, since in the formal languages literature the input to be executed on the automaton is typically a string. The work in Baeza-Yates & Gonnet (1996) represents an exception, but the tree it exploits (a Patricia tree) has very different semantics (and usage) with respect to ours. In detail, our approach operates as follows: the algorithm Search Process (see Algorithm 2) takes in input the trace tree T and the deterministic au- tomaton A, and provides as an output a set of pairs, composed of a trace tree leaf node and a corresponding string. Each of these strings is an explicit instantiation of the query represented by the automaton, verified by (some of) the support traces in the leaf node. The output support traces can then be retrieved from the event log and presented to the user, or provided as an input to the filtering step. Basically, Search Process executes a breadth first visit of the trace tree; it exploits the variable searching, defined as a set of triples, composed by a trace tree node, an automaton state, and the string that has been recognized on the automaton so far. Initially (lines 5-7), searching contains the sons of the root (since the root is a dummy action #, see section 2.1), all paired to the initial state of the query automaton and to the empty string. The visit procedure (lines 8-35) extracts one triple at a time from the set searching. If the node in the triple contains a set of actions to be executed in any order (line 11), we simulate all the permutations on the automaton, and save the states we reach and the corresponding recognized strings into the new states set (line 12). If the node contains one single action, we simply simulate it on the automaton, and save the state we reach and the corresponding string into the new states set (line 15). In both cases, the string saved in new states is the one in the input triple properly updated with the newly recognized symbols. After the simulation, if the node at hand is a leaf (line 18), then for each item in new states we check whether the state component is a final state (lines 19-20); if this is the case, node and the string paired to the final state are saved in the output variable result (line 21). Otherwise, if node is not a leaf, we pair its sons to all the items in new states, and save these objects into searching (lines 26-34). The visit terminates when searching is empty, i.e., all tree levels have been visited. The visit procedure is linear in the number of the trace tree nodes. 14 Example Referring to our example query, providing the trace tree in Figure 1 as an input to the Algorithm Search Process, searching initially contains the sons of the root A, B, and C, paired with the initial state of the deter- ministic automaton obtained from the non-deterministic one in Figure 2, and with the empty string. We simulate the actions A, B, and C on the automa- ton. Only B is recognized, generating a state saved in new states with the corresponding string B (line 15). We then pair the sons of node B (E, D, DE) to the item in new states and save these triples into searching (lines 26-34). Continuing the visit, particularly interesting is the case of node DE, which requires to consider all the possible permutations of actions D and E. Both the permutations DE and ED are initially recognized. However, as the visit proceeds and node PZ is reached, it turns out that DE must be followed by the permutation PZ to satisfy the query; on the other hand, if the choice ED is made, it must be followed by ZP . Indeed, the query imposes some constraints that cannot be checked only locally, i.e., referring to a single node along the branch. After this step of the visit (depth 4 in the tree), the recognized partial strings paired to node PZ are BDEOPZ and BEDOZP . When reaching the leaf node W , however, only BDEOPZW is recognized, and reported in output together with the corresponding leaf node W . Considering all the branches, the overall output of the Search Process algorithm is: ⟨ Z in branch 1, BEOPZ ⟩, ⟨ W in branch 4, BDEOPZW ⟩, ⟨ Z in branch 5, BEDWZ ⟩. Filtering As observed above, the tree search step outputs a set pairs, each com- posed of a leaf node and a string, the latter corresponding to an explicit instantiation of the input query issued by the user. If an output leaf node ends a branch which includes one or more nodes with actions to be executed in any order, it is possible that only some of the permutations of these actions are acceptable to answer the corresponding explicit instantiation of the input query; therefore, the support traces must be filtered accordingly. The filtering algorithm (see Algorithm 3) takes in input the trace tree T , and the output set of the tree search step (in search result), and operates as follows. For each item in search result, it extracts the node component and saves it in node (lines 5-6). If node is a leaf (line 9), function support is applied 15 ALGORITHM 2: Pseudo-code of the procedure Search Process. 1 Search Process(T, A) 2 Output: set of < node, string > 3 result ←{} 4 searching ←{} 5 foreach node ∈ sons(root(T)) do 6 searching ← searching ∪ < node, 0, empty > 7 end 8 repeat 9 tmp ←{} 10 foreach < node, state, string > ∈ searching do 11 if node is an any-order-actions node then 12 new states ← anyorder simulate(A, < node, state, string >) 13 end 14 else 15 new states ← simulate(A, < node, state, string >) 16 end 17 if new states ̸= {} then 18 if node is a leaf then 19 foreach < state, string > ∈ new states do 20 if final(state) then 21 result ← result ∪ < node, string > 22 end 23 end 24 end 25 else 26 foreach n ∈ sons(node) do 27 foreach < state, string > ∈ new states do 28 tmp ← tmp ∪ < n, state, string > 29 end 30 end 31 end 32 end 33 end 34 searching ← tmp 35 until searching ̸= {} 36 return result 16 ALGORITHM 3: Pseudo-code of the procedure Filtering. 1 Filtering(T, search result) 2 Output: references to traces 3 result ←{} 4 tmp ←{} 5 foreach < n, string > ∈ search result do 6 node ← n 7 tmp ←{} 8 foreach element ∈ string do 9 if node is a leaf then 10 tmp ← support(node, element) 11 end 12 else 13 if node is an any-order-actions node then 14 tmp ← tmp ∩ support(node, element) 15 end 16 end 17 node ← father(node) 18 element ← update(element) 19 end 20 result ← result ∪ tmp 21 end 22 return result 17 to the node itself, and to the tail portion (element) of the corresponding string: in particular, if node contains a single action, element is a just the last symbol in the string; otherwise, it is a sequence of as many symbols as the number of actions to be executed in any order in node, always computed from the last position. The function support outputs in tmp the traces verifying element: they are a subset of the support traces of node, if node contains actions in any order, since element identifies the single acceptable permutation, in this case. Otherwise, all the support traces of node are returned (line 10). The variable node is then updated by considering the father of the leaf, along its branch (line 17); element is updated accordingly, moving towards the start of the string by as many symbols as the number of actions in node (line 18). At the next iteration, if node contains actions to be executed in any order (line 13), the algorithm calculates the intersection between tmp and the output of the function support (calculated as above), in order to always keep all and only the traces supporting the portion of the string analysed so far (line 14). The algorithm terminates when the string has been fully examined (i.e., we have reached the sons of the root in T), and outputs the references to the filtered traces in result (line 22). The complexity is therefore the number of elements in search result (i.e., the number ⟨leaf node, string ⟩ pairs identified as an output of the previous step), by the length of the longest string in search result. Example We exemplify the algorithm Filtering on the item ⟨ W in branch 4, BDEOPZW ⟩. We set node = W (line 6) and element=W; node W is a leaf and does not contain actions to be executed in any order. Therefore, function support outputs in tmp all the support traces referenced by this node. We then update node=PZ. At the following iteration, element=PZ, and, since node contains two actions to be executed in any order, we calculate the intersection between tmp and the support traces corresponding to the permutation PZ (line 14). Variable node is then set to O. At the following iteration, element=O, and tmp does not change. node is set to DE. At the next iteration, element=DE, and we calculate the intersection between tmp and the support traces corresponding to the permutation DE (line 14). The last iteration is trivial, and does not change the content of tmp, which, in the end, contains the references to all and only the traces supporting the string 18 BDEOPZW , added to the output result (line 22). Obviously, if the leaf node ends a branch that contains no nodes with actions to be executed in any order, the leaf support traces can be directly presented to the user, and the filtering step is not required. 3. Results We have compared our method to a very classical approach, i.e., an ex- isting regular expression processor provided by the Java Regex APIs, not coupled with any indexing strategy2. The experimental database was taken from the Datacentrum website3. Datacentrum archives research data in a standardized, secure and well doc- umented manner, and provides permanent access to them. Specifically, we used a synthetic dataset referring to a loan assessment process. The dataset is composed of 10000 traces, expressed as sequences of 8 to 27 actions. 90% of the traces are 8 to 12 action long. Overall, the dataset contains 14 different action types. For the experiments, we used a machine equipped with an Intel i7-4810MQ CPU @ 2.80GhZ, 8GB RAM, SSHD Hybrid 64MB Cache. In detail, we performed a scalability test. To this end, we made random samplings of the overall available traces, defining 5 subsets of the original database, of dimensions 1000, 3000, 5000, 7000 and 10000 traces respectively (where the last sample is obviously the whole database). We then defined two different query types, and executed 1000 queries for each type. We then calculated average query answering times. The first query type was characterized by the presence of a short delays: the sum of the maximum ranges of the delays is at most three. Overall, the query could contain up to three delays, as in Examples (A), (B) and (C) below. Example (A): (0,0) Loan application received Check application form completeness 2The interested reader can find information about Regex at the link http://www.regular-expressions.info/engine.html - last accessed on July 16, 2015 3http://datacentrum.3tu.nl/en/home/, last accesses on July 16, 2015 19 (3,5) Reject application Loan application rejected (0,0) Example (B): (0,2) Appraise property (1,2) Assess eligibility Reject application Loan application rejected (0,0) Example (C): (0,1) Check application form completeness (2,3) Assess loan risk Assess eligibility (1,2) Loan application rejected (0,0) The second query type was characterized by the presence of longer delays, whose overall length could be up to 5 actions, as in Example (D) below. Example (D): (0,0) Loan application received Check application form completeness (0,5) Reject application Loan application rejected (0,0) The same queries were executed both using our approach and using the Regex Java processor on the 5 archives of different dimensions. The average retrieval times for the two query types are shown in Tables 1 and 2. In the Tables, referring to our method, the times of the three steps (automaton generation, tree search and filtering) are also detailed. 20 Table 1: Comparison of retrieval times (in msec) using our method (trace tree - TT) and using the Regex Java regular expression processor in a scalability test on type 1 queries (short delays) DB dimension TT TT TT TT regex JAVA autom. search filter. tot. 1000 1.71 0.05 0.08 1.84 8.48 3000 1.74 0.06 0.23 2.03 18.82 5000 1.75 0.06 0.40 2.21 29.13 7000 1.73 0.06 0.54 2.33 41.57 10000 1.76 0.06 0.77 2.59 58.72 Table 2: Comparison of retrieval times (in msec) using our method and using the Regex Java regular expression processor in a scalability test on type 2 queries (long delays) DB dimension TT TT TT TT regex JAVA autom. search filter. tot. 1000 3.57 0.08 0.11 3.76 11.90 3000 3.66 0.08 0.31 4.05 18.80 5000 3.63 0.09 0.54 4.26 31.13 7000 3.67 0.09 0.74 4.50 46.23 10000 3.59 0.09 1.02 4.70 60.13 As it can be observed, our method always outperformed Regex. The time of both methods grew as the database dimension increased, but the growth of our method was very limited, referring to both query types (as shown in Figure 3, which plots the values of the two Tables). In both methods, type 2 queries took more time than type 1 queries. This was expected. In particular, as regards our method, the non-deterministic automaton corresponding to type 2 queries has a larger number of states than the one corresponding to type 1 queries (built using Berry and Sethi approach (Berry & Sethi, 1986)). Thus, the transformation cost from the non-deterministic automaton to the deterministic automaton increases too, since this cost depends on the number of states of the non-deterministic automaton. Looking at our method in more detail, referring to both query types, as expected, the cost of building the automaton for queries with a given length of delays was almost constant. The same also holds for searching time, since tree search navigates the index, whose dimension increased only slightly with respect to the size of the dataset. On the other hand, not surprisingly, filtering time (which depends on the number of the retrieved 21 Figure 3: Scalability results. traces) grew linearly with respect to the size of the dataset. However, the whole computation was dominated by the cost of building the automaton. Specifically, automaton generation took from 70 to 95% of the total query answering time (see Tables 1 and 2). On the other hand, the Regex approach does not exploit any index, so that each trace in the dataset is directly checked. As a consequence, the time complexity of Regex grew linearly with respect to the dimension of the dataset. In conclusion, the approach described in this paper can answer the queries in significantly shorter times with respect to classical approaches, and the advantage increases as the database grows. 4. Related work Operational support is typically not provided by commercial Business Intelligence and Business Process Management tools. However, operational support techniques are implemented in the open source framework ProM 22 (van Dongen et al., 2005), developed at the Eindhoven University of Tech- nology, which represents the state of the art in process mining research. In ProM, prediction and recommendation are typically supported by replaying log traces on the transition system (van Dongen, 2007), a state-based model that explicitly shows the states a process can be in, and all possible tran- sitions between these states. The replay activity allows to calculate, e.g., the mean time to completion from a given state, or the most probable next action to be executed. In ProM’s approach, statistics on event log traces are thus used for operational support, but the overall technique is very different from the one we propose in this paper, and no trace retrieval on the basis of complex patterns search is supported. On the other hand, traces have been recently considered in the CBR literature, as sources for retrieving and reusing user’s experience. As an ex- ample, at the International Conference on CBR in 2012, a specific workshop was devoted to this topic (see Floyd et al. (2012)). In 2013, Cordier et al. (2013) proposed trace-based reasoning, a CBR approach where cases are not explicitly stored in a library, but are implic- itly recorded as “episodes” within traces. The elaboration step, in which a case is extracted from a trace, is thus one of the most challenging parts of the reasoning process. Zarka et al. (2013) extended that work, and defined a similarity measure to compare episodes extracted from traces. In these works, traces are typically intended as observations captured from users’ in- teraction with a computer system. Trace-based reasoning was exploited in recommender systems (Adomavicius et al., 2011; Zarka et al., 2012), and to support the annotation of digitalized cultural heritage documents in Doumat et al. (2010). Huang et al. (2013) and Montani & Leonardi (2014) propose two trace retrieval approaches, where different metrics, based on extensions of the edit distance, are exploited. No indexing strategy is however provided to make retrieval faster. Leake (2010) used execution traces recording provenance information to improve reasoning and explanation in CBR. In the Phala system (Leake & Kendall-Morwick, 2008), the authors supported the generation and composi- tion of scientific workflows by mining execution traces for recommendations to aid workflow authors. Finally, Lanz et al. (2010) used annotated traces recorded when a human user played video games in order to feed a case-based planner. All these approaches implement forms of reasoning on traces, but do not aim at providing operational support. 23 There is a growing interest on business process management applications in the CBR community (see, e.g., (Minor et al., 2014b)). Most works, how- ever, focus on process models rather than on traces, and aim at retrieving models (typically represented as graphs) similar to an input one. These works have required the introduction of different metrics for graph compari- son. Most of them are based on extensions of the graph edit distance (Minor et al., 2008; Bergmann & Gil, 2014; Kunze & Weske, 2011; Li et al., 2008; Montani et al., 2015; Dijkman et al., 2009; LaRosa et al., 2013). The work in Minor et al. (2008) makes use of a normalized version of the graph edit distance. The approach is used to support workflow modification in an agile workflow system, and takes into account control flow information as well as activity information. However, the work is limited to considering (small) changes with respect to a running process instance. The work in Dijk- man et al. (2009) also provides a normalized version of the graph edit distance for comparing business process models, and defines syntactical edit opera- tion costs for activity node substitution, activity node insertion/deletion, and edge insertion/deletion. The work in LaRosa et al. (2013) extends the work in Dijkman et al. (2009) by explicitly representing gateway nodes, in order to describe, e.g., parallelism and mutual exclusion, and exploiting them in distance calculation. The work in Kunze & Weske (2011) relies on graph edit distance, and exploits string edit distance on node names to determine the cost of node substitutions. The work in Li et al. (2008) encapsulates a set of edit operations into the so-called “high-level change operations”, and measures distance on the basis of the number of high-level change operations needed to transform one graph into another. All the above contributions typically make use of syntactical information in the definition of the edit operation costs. The works in Bergmann & Gil (2014); Montani et al. (2015), on the other hand, exploit semantic informa- tion in activity comparison. In Bergmann & Gil (2014), a system working on workflows represented as semantically labeled graphs is presented. The paper proposes to use a metric in which the similarity between two mapped nodes or arcs makes explicit use of their semantic description. The work is particularly focused on the data flow. The work in Montani et al. (2015) ex- ploits domain knowledge and temporal information to parametrize the cost of node substitutions and edge substitution. As in LaRosa et al. (2013), gateway nodes are also considered. The approach in Goderis et al. (2006), on the other hand, affords the problem of graph comparison by relying on graph isomorphism. It focuses 24 on scientific workflows, which have a strong focus on the data flow, typically restricting the control flow to a partial ordering of the tasks. The work in Ma et al. (2014) focuses on data oriented workflows as well. It defines a formal structure called Time Dependency Graph (TDG), and exploits it as a rep- resentation model of data oriented workflows with variable time constraints. A distance measure is proposed for computing workflow similarity by their normalization matrices, established on the basis on their TDGs. The work in Kapetanakis et al. (2010) exploits a maximum common subgraph approach for similarity-based process retrieval, in a retrieval sys- tem for supporting business process monitoring. Interestingly, the metric in Kapetanakis et al. (2010) takes into account temporal information, since it combines a contribution related to activity similarity, and a contribution related to delays between activities. Finally, in Madhusudan et al. (2004) a retrieval system for supporting in- cremental workflow modeling is presented. The system proposes a similarity- based retrieval of workflow templates using a planner that employs an inexact graph matching algorithm based on similarity flooding. For computing simi- larities, the algorithm relies on the idea that elements of two distinct graphs are similar, when their adjacent elements are similar. The algorithm prop- agates the similarity from a node to its respective neighbors based on the topology in the two graphs. It is worth noting that all of these works, including our own previous contribution (Montani et al., 2015), are however only loosely related to the present paper, since they do not focus on traces, but on process models (i.e., graphs), and do not aim at enabling operational support. Very interestingly, some recent papers in the area of CBR for business process management (Minor et al., 2011, 2014a; Müller & Bergmann, 2014) also consider the reuse/adaptation step of the CBR cycle (Aamodt & Plaza, 1994). An automation of the adaptation step in process model retrieval is easier in very specific domains, such as the one of cooking recipes (Müller & Bergmann, 2014). In this setting, the graph structure is usually quite simple, and adaptation is often limited to ingredient substitution. However, the issue of implementing a reuse/adaptation strategy in trace retrieval may be investigated in our future research as well. 25 5. Conclusions In this paper, we have introduced a novel framework for trace retrieval, studied to implement operational support tasks in a flexible and efficient way. With respect to existing operational support facilities, our tool is signif- icantly innovative, since, to the best of our knowledge, it is the first one to provide a direct exploitation of experiential knowledge, by means of case- based retrieval techniques. Once past traces have been retrieved, classical statistical techniques can then be applied to them, to support prediction and recommendation. With respect to existing trace retrieval tools in the CBR area, our ap- proach is more efficient and flexible, since: • by allowing for the use of (imprecise) delays in the query language, it enables to express a very large number of explicit queries in a compact way; • by providing the trace tree as an input to the automaton: – it executes common prefixes of different traces only once on the automaton, avoiding repeated, identical checks; – it speeds up retrieval with respect to a flat search into the event log (as testified by our experiments). In the future, we would like to enable also the retrieval of traces that include a prefix similar (but not necessarily identical) to the input trace. To this end, we plan to rely on classical techniques, involving a visit of the trace tree, and a comparison between the trace tree branch at hand and the input trace by means of edit distance calculation. This approach, while quite straightforward from the methodological viewpoint, will further increase the flexibility and the usability of our tool. Finally, we plan to extensively test the overall framework on real world traces, which log the actions executed during stroke patient management in a set of Northern Italy hospitals. der Aalst, W. V. (2011). Process Mining. Discovery, Conformance and En- hancement of Business Processes. Springer. 26 Aamodt, A., & Plaza, E. (1994). Case-based reasoning: foundational issues, methodological variations and systems approaches. AI Communications, 7 , 39–59. Adomavicius, G., Mobasher, B., Ricci, F., & Tuzhilin, A. (2011). Context- aware recommender systems. AI Magazine, 32 , 67–80. Baeza-Yates, R. A., & Gonnet, G. H. (1996). Fast text searching for regular expressions or automaton searching on tries. J. ACM , 43, 915–936. Bergmann, R., & Gil, Y. (2014). Similarity assessment and efficient retrieval of semantic workflows. Information Systems, 40, 115–127. Berry, G., & Sethi, R. (1986). From regular expressions to deterministic automata. Theor. Comput. Sci., 48, 117–126. Buijs, J., van Dongen, B., & van der Aalst, W. (2012). On the role of fitness, precision, generalization and simplicity in process discovery. In On the Move to Meaningful Internet Systems: OTM 2012 (pp. 305–322). Springer. Canensi, L., Montani, S., Leonardi, G., & Terenziani, P. (2014). Chap- man: a context aware process miner. In Proc. Workshop on Synergies be- tween Case-Based Reasoning and Data Mining, International Conference on Case Based Reasoning (ICCBR). Cnudde, S. D., Claes, J., & Poels, G. (2014). Improving the quality of the heuristics miner in prom 6.2. Expert Syst. Appl., 41 , 7678–7690. Cordier, A., Lefevre, M., Champin, P., Georgeon, O., & Mille, A. (2013). Trace-based reasoning - modeling interaction traces for reasoning on expe- riences. In C. Boonthum-Denecke, & G. M. Youngblood (Eds.), Proceedings of the Twenty-Sixth International Florida Artificial Intelligence Research Society Conference, FLAIRS 2013, St. Pete Beach, Florida. May 22-24, 2013.. AAAI Press. Dijkman, R., Dumas, M., & Garca-Banuelos, R. (2009). Graph matching al- gorithms for business process model similarity search. In U. Dayal, J. Eder, J. Koehler, & H. Reijers (Eds.), Proc. International Conference on Busi- ness Process Management (pp. 48–63). volume 5701 of Lecture Notes in Computer Science. 27 van Dongen, B. (2007). An Iterative Algorithm for Applying the Theory of Regions in Process Mining. BETA publicaties: Preprints. Beta, Research School for Operations Management and Logistics. van Dongen, B., De Medeiros, A. A., Verbeek, H., Weijters, A., & der Aalst, W. V. (2005). The proM framework: a new era in process mining tool support. In G. Ciardo, & P. Darondeau (Eds.), Knowledge Mangement and its Integrative Elements (pp. 444–454). Springer, Berlin. Doumat, R., Egyed-Zsigmond, E., & Pinon, J. (2010). User trace-based rec- ommendation system for a digital archive. In I. Bichindaritz, & S. Montani (Eds.), Case-Based Reasoning. Research and Development, 18th Interna- tional Conference on Case-Based Reasoning, ICCBR 2010, Alessandria, Italy, July 19-22, 2010. Proceedings (pp. 360–374). Springer volume 6176 of Lecture Notes in Computer Science. Floyd, M. W., Fuchs, B., Gonzalez-Calero, P., Leake, D., Ontanon, S., Plaza, E., & Rubin, J. (2012). TRUE: Traces for Reusing Users Experiences Cases, Episodes, and Stories, International Conference on Case Based Reasoning (ICCBR). Lyon. Goderis, A., Li, P., & Goble, C. A. (2006). Workflow discovery: the problem, a case study from e-science and a graph-based solution. In F. Leymann, & L. Zhang (Eds.), Proc. IEEE International Conference on Web Services (pp. 312–319). IEEE, USA. Huang, Z., Juarez, J. M., Duan, H., & Li, H. (2013). Length of stay prediction for clinical treatment process using temporal similarity. Expert Syst. Appl., 40, 6330–6339. Kapetanakis, S., Petridis, M., Knight, B., Ma, J., & Bacon, L. (2010). A case based reasoning approach for the monitoring of business workflows. In I. Bichindaritz, & S. Montani (Eds.), Proc. International Conference on Case Based Reasoning (ICCBR) (pp. 390–405). Springer, Berlin volume 6176 of Lecture Notes in Computer Science. Kunze, M., & Weske, M. (2011). Metric trees for efficient similarity search in large process model repositories. In M. Muehlen, & J. Su (Eds.), Proc. Business Process Management Workshops (pp. 535–546). Springer, Berlin volume 66 of Lecture Notes in Business Information Processing. 28 Lam, M. S., Aho, A. V., Sethi, R., & Ullman, J. D. (2006). Compilers: Principles, Techniques, and Tools. Addison-Wesley. Lanz, A., Weber, B., & Reichert, M. (2010). Workflow time patterns for process-aware information systems. In Proc. BMMDS/EMMSAD (pp. 94– 107). LaRosa, M., Dumas, M., Uba, R., & Dijkman, R. (2013). Business process model merging: An approach to business process consolidation. ACM Trans. Softw. Eng. Methodol., 22, 11. Leake, D. B. (2010). Case-based reasoning tomorrow: Provenance, the web, and cases in the future of intelligent information processing. In Z. Shi, S. Vadera, A. Aamodt, & D. B. Leake (Eds.), Intelligent Information Pro- cessing V - 6th IFIP TC 12 International Conference, IIP 2010, Manch- ester, UK, October 13-16, 2010. Proceedings (p. 1). Springer volume 340 of IFIP Advances in Information and Communication Technology. Leake, D. B., & Kendall-Morwick, J. (2008). Towards case-based support for e-science workflow generation by mining provenance. In K. Althoff, R. Bergmann, M. Minor, & A. Hanft (Eds.), Proc. ECCBR 2008, Advances in Case-Based Reasoning (pp. 269–283). Springer volume 5239 of Lecture Notes in Computer Science. van Leeuwen, J. (1994). Handbook of Theoretical Computer Science. Mit Press. Li, C., Reichert, M., & Wombacher, A. (2008). On measuring process model similarity based on high-level change operations. In Q. Li, S. Spaccapietra, E. S. K. Yu, & A. Olivé (Eds.), Proc. International Conference on Con- ceptual Modeling (pp. 248–264). Springer, Berlin volume 5231 of Lecture Notes in Computer Science. Ma, Y., Zhang, X., & Lu, K. (2014). A graph distance based metric for data oriented workflow retrieval with variable time constraints. Expert Syst. Appl., 41 , 1377–1388. Madhusudan, T., Zhao, J., & Marshall, B. (2004). A case-based reasoning framework for workflow model management. Data and Knowledge Engi- neering, 50 , 87–115. 29 Minor, M., Bergmann, R., & Görg, S. (2014a). Case-based adaptation of workflows. Inf. Syst., 40, 142–152. Minor, M., Bergmann, R., Görg, S., & Walter, K. (2011). Reasoning on busi- ness processes to support change reuse. In B. Hofreiter, E. Dubois, K. Lin, T. Setzer, C. Godart, E. Proper, & L. Bodenstaff (Eds.), 13th IEEE Con- ference on Commerce and Enterprise Computing, CEC 2011, Luxembourg- Kirchberg, Luxembourg, September 5-7, 2011 (pp. 18–25). IEEE Computer Society. Minor, M., Montani, S., & Recio-Garćıa, J. (2014b). Process-oriented case- based reasoning. Inf. Syst., 40 , 103–105. Minor, M., Tartakovski, A., Schmalen, D., & Bergmann, R. (2008). Agile workflow technology and case-based change reuse for long-term processes. International Journal of Intelligent Information Technologies, 4 , 80–98. Montani, S., & Leonardi, G. (2014). Retrieval and clustering for supporting business process adjustment and analysis. Information Systems, 40 , 128– 141. Montani, S., Leonardi, G., Quaglini, S., Cavallini, A., & Micieli, G. (2015). A knowledge-intensive approach to process similarity calculation. Expert Syst. Appl., 42, 4207–4215. Müller, G., & Bergmann, R. (2014). Workflow streams: A means for composi- tional adaptation in process-oriented CBR. In L. Lamontagne, & E. Plaza (Eds.), Case-Based Reasoning Research and Development - 22nd Inter- national Conference, ICCBR 2014, Cork, Ireland, September 29, 2014 - October 1, 2014. Proceedings (pp. 315–329). Springer volume 8765 of Lec- ture Notes in Computer Science. Zarka, R., Cordier, A., Egyed-Zsigmond, E., Lamontagne, L., & Mille, A. (2013). Similarity measures to compare episodes in modeled traces. In S. J. Delany, & S. Ontañón (Eds.), Case-Based Reasoning Research and Devel- opment - 21st International Conference, ICCBR 2013, Saratoga Springs, NY, USA, July 8-11, 2013. Proceedings (pp. 358–372). Springer volume 7969 of Lecture Notes in Computer Science. Zarka, R., Cordier, A., Egyed-Zsigmond, E., & Mille, A. (2012). Contextual trace-based video recommendations. In A. Mille, F. L. Gandon, J. Misselis, 30 M. Rabinovich, & S. Staab (Eds.), Proceedings of the 21st World Wide Web Conference, WWW 2012, Lyon, France, April 16-20, 2012 (Companion Volume) (pp. 751–754). ACM. 31 
work_zv5hinl2kzfuzlpf37agkaqtpq	----	UEF//eRepository DSpace https://erepo.uef.fi Artikkelit Luonnontieteiden ja metsätieteiden tiedekunta 2017 Using linguistic features to automatically extract web page title Gali N info:eu-repo/semantics/article info:eu-repo/semantics/acceptedVersion © Elsevier Ltd CC BY-NC-ND https://creativecommons.org/licenses/by-nc-nd/4.0/ https://erepo.uef.fi/handle/123456789/4243 Downloaded from University of Eastern Finland's eRepository Accepted Manuscript Using Linguistic Features to Automatically Extract Web page Title Najlah Gali , Radu Mariescu Istodor , Pasi Fränti PII: S0957-4174(17)30135-5 DOI: 10.1016/j.eswa.2017.02.045 Reference: ESWA 11153 To appear in: Expert Systems With Applications Received date: 5 October 2016 Revised date: 27 February 2017 Accepted date: 28 February 2017 Please cite this article as: Najlah Gali , Radu Mariescu Istodor , Pasi Fränti , Using Linguistic Fea- tures to Automatically Extract Web page Title, Expert Systems With Applications (2017), doi: 10.1016/j.eswa.2017.02.045 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. http://dx.doi.org/10.1016/j.eswa.2017.02.045 http://dx.doi.org/10.1016/j.eswa.2017.02.045 ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Highlights  Successful title extraction must analyze both the DOM nodes and the title tag  Natural language processing improves the quality of the title.  Visual and formatting features are less relevant for the task.  Simpler classifier like k-NN perform as well as an advanced classifier like SVM.  The proposed method significantly outperforms all existing ones by clear margin. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Using Linguistic Features to Automatically Extract Web page Title Najlah Gali, Radu Mariescu Istodor, Pasi Fränti Machine Learning group, School of Computing, University of Eastern Finland, Joensuu FI-80101, Finland {najlaa, radum, franti} @cs.uef.fi Abstract Existing methods for extracting titles from HTML web page mostly rely on visual and structural features. However, this approach fails in the case of service-based web pages because advertisements are often given more visual emphasize than the main headlines. To improve the current state-of-the-art, we propose a novel method that combines statistical features, linguistic knowledge, and text segmentation. Using annotated English corpus, we learn the morphosyntactic characteristics of known titles and define a part-of-speech tag patterns that help to extract candidate phrases from the web page. To evaluate the proposed method, we compared two datasets Titler and Mopsi and evaluated the extracted features using four classifiers: Naïve Bayes, k-NN, SVM, and clustering. Experimental results show that the proposed method outperform the solution used by Google from 0.58 to 0.85 on Titler corpus and from 0.43 to 0.55 on Mopsi dataset, and offers a readily available solution for the title extraction problem. Keywords Web content mining, Information extraction, Title extraction, Natural language processing, Machine learning ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 1 Introduction Human is an impatient creature by nature when seeking information from the Internet; the user wants to obtain the right answer immediately and with minimal efforts (Marchlonini, 1992; Song, Xin, Shi, Wen, & Ma, 2006). In most applications, the title of the content is the first thing where the user pays attention. Even a single word or phrase can dramatically change the whole message of this content. To have a correct title is important and should therefore, be descriptive, concise, and grammatically correct (Hu et al., 2005). In web pages, titles usually exist in two places for different purposes. First, the title is somewhere in the body text with high visual emphasis for the human reader. This is important for humans who browse the web page quickly on a computer display. Second, the title is placed in the title field (between <title> and </title> tag) for robots, crawlers, and programs that prepare a summary for the web page. However, the designers of the web page often ignore this or abuse the title tag by adding extra content like keywords, address or other less relevant text. Its content becomes then vague, incorrect, or it might be even missing completely (Xue et al., 2007). In mobile applications the problem exaggerates. The small screen of miniaturized devices are even more restrictive to the displayed content and requires the title to be also fitted spatially. T he title is also needed for indexing in search engines like Google. Title extraction aims at producing a compact title for a web page automatically. Due to the problems of the title tag, existing literature has mainly been focused on extracting the title from the body text. Methods have been developed for web pages of standard format such as news and pages of educational institutions. However, less attention has been given to service-based web pages such as entertainment, sport, and restaurants. Existing methods also make an assumption like the title is always located in the top region of the page and has visual prominence; they often fail to correctly extract the title of the service-based pages where the title is exchanged for a logo, or it is positioned elsewhere on the page (see Figure 1). For example, Hu et al. (2005) and Xue et al. (2007) explicitly state that the title must be in the top area of the page. Furthermore, Fan, Luo, and Joshi (2011) hypothesize that the title is located in the upper part of the main text. Changuel, Labroche, and Bouchon-Meunier (2009) implicitly assume that the title appears in the top portion of the page and as a result extract only the first 20 text nodes from the Document Object Model (DOM) tree. Another assumption often made is that the title in a body is a separate line of text (i.e., it has its own text node in the DOM tree). However, the modern web page design allows the title to appear as a part of other phrases in the text node of the DOM tree. For example, <h1>Welcome to Petter Pharmacy, please select one of the five options below: </h1> will produce ill-fitting title unsuitable for mobile devices. According to our experiments, about 68% of the title nodes also contain additional information similar to the example and are therefore prone to errors in the title extraction. In this work, we developed a novel method to overcome these problems in the title extraction for service -based content and mobile applications. Our key finding is that the title tag is still the best source. However, it needs to be segmented and further processed. Our preliminary version was presented in (Gali & Fränti, 2016). Here, we further enhance this approach by applying additional part-of-speech (POS) tagging. POS processes every word in the text and assigns them with a tag based on the relationship with adjacent and related words in the phrase, sentence, or paragraph. POS tagging has been successfully employed in other domains such as keyword extraction (Hulth, 2003). However, to our best knowledge, no language model has been applied for title extraction in the context of web pages. We are only aware the work of (Lopez, Prince, & Roche, 2010; Lopez, Prince, & Roche, 2014) giving POS model for mailing lists and news articles in the French language. We, therefore, investigate the contribution of POS tagging to this task. We aim at identifying features that are independent of the format of the web page. Our method uses the following features: syntactic structure, similarity with the link of the web page, appearance in the title tag, appearance in meta tags, popularity on the web page, appearance in heading tags, capitalization, capitalization frequency, independent appearance, and phrase length. We consider four alternative classifiers: Naive Bayes, clustering-based, k-nearest neighbors (k-NN), and support vector machine (SVM), which to our knowledge have not been compared previously in the title extraction task. We compare the proposed method against related ones with two datasets: Titler and Mopsi. Experiments show that our method gives a significant improvement, and achieves an accuracy of 0.85 with Titler dataset. The corresponding results of the baseline (title tag as such), Google, and the best content-based method are 0.52, 0.58 and 0.47 respectively. The rest of the paper is organized as follows. In Section 2, we review existing methods for title extraction, and the new method is introduced in Section 3. Experiments are done in Section 4. Effect of POS pattern is studied in Section 4.5, feature extraction in 4.6, choice of the classifier in 4.7, and title selection methods in 4.8. Comparisons to the existing methods are then performed in Section 4.9. We compare to all existing identification methods that are accessible. These include Styling (Changuel et al., 2009), TitleFinder (Mohammadzadeh, Gottron, Schweiggert, & Heyer, 2012), Title Tag Analyzer (Gali & Fränti, 2016), and the Baseline. We also compare to the titles provided by Google in the search results page. The results show that the proposed approach outperforms all the methods. The method with k-NN improves the Jaccard of the baseline from 0.50 to 0.84 on Titler corpus and from 0.44 to 0.59 on Mopsi dataset. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Figure 1 Different layouts of web pages 123 (white squares refer to images of logos while red ovals refer to titles in the body of the web pages). 2 Related work Figure 2 shows typical steps for the title extraction (left) and possible approaches to each step (right); the modules that are covered in this work are highlighted in blue. 2.1. Content source for title The title of a web page is usually found in one or more of three places: the title tag (i.e., between <title> and </title>), the text of the body, and the logo image. According to our experiments with 1002 websites, the occurrence of the title in these three places is as follows:  Title tag (91%)  Text of the body (97%)  Logo (89%) 1 http://karaautos.co.uk/ 2 http://www.petterpharmacy.co.uk/ 3 http://theapollo.com.au/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Figure 2 Typical steps for title extraction. The title tag is the obvious source, and the author of the page is expected to fill it with a proper title. However, people often do not complete this tag carefully as it does not have a visual impact on the page. A title tag often contains additional text, such as the name of the hosting website, information about the place offering services, a slogan, and contact details (see Table 1). The body text of a web page is a second source for a title. It has been given more focus by researchers given that a title in the body is visible to users and is thus expected to be written more carefully than the title tag (Changuel et al., 2009; Hu et al., 2005; Wang et al., 2009; Xue et al., 2007) However, extracting a title from the body of the web page is not an easy task, as roughly half of a page‘s content is irrelevant text (Gibson, Punera, & Tomkins, 2005). This irrelevant text (e.g., advertisements) is often given even more visual emphasis than the main headlines, which makes the task even more challenging. Furthermore, no standard location exists in relation to title placement. In this paper, we extract the candidate titles from both the body of the web page and the title tag. Table 1 The most typical problems related to title tag and the frequency they appear (according to our experiments). Type Proportion (%) Example Annotated title Long description 62 <title>Brook's Diner | 24 Hampden Square, Southgate, London N14 5JR | 020 8368 7201 | eat@brooksdiner.com | Like us on Facebook — Home</title> Brook's Diner Incorrect 6 <title>Hot Tubs, hot tub hire, swimming pools, Bristol, Gloucester</title> Rio Pool Vague 2.4 <title>home</title> <title>index</title> <title> | </title> Hellard Bros Ltd. Short description 0.5 <title> Toby‘s Estate</title> Toby‘s Estate coffee Empty 0.2 <title> </title> Zavino Hospitality Group The third source for a title is the logo image. However, extracting a title from this image would be very challenging. One reason is that the logo image must first be identified. Another reason is that the standard optical character recognition (OCR) approach would not generally work given that the content of the image is highly complex. We are not aware of any technique that attempts this approach. It should technically be possible, but as shown by the examples in Figure 3, such a technique would need to handle a wide variety of complex text fonts that involve shadowing effects, textures, and other artistic features. 2.2. Content analysis and candidate title extraction Most title extraction methods use either DOM tree representation or combine the DOM structure with the visual cues of the page in a vision-based tree. The vision-based tree is built using the vision-based page segmentation algorithm (VIPS) introduced by Cai, Yu, Wen, and Ma (2003). The vision-based tree provides visual partitioning of the page where the blocks ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT (i.e., DOM nodes) are grouped visually, while DOM tree describes the parent-child relationship between the tree nodes; therefore, it is not necessary that nodes in the vision-based tree correspond to the node in the DOM-tree. The VIPS needs to refer to all styling information (including external sheets) to locate the proper place of the block in the tree. If the web page lacks rich visual properties, the hierarchies are incorrectly constructed. A wrong structure can also result from the algorithm not detecting separators represented by thin images. We, therefore, use DOM tree representation. In both tree representations, existing methods use the entire text of the leaf nodes as candidate titles. In this paper, we extract only the relevant part of the text nodes by using POS tag patterns (see section 3.3). Figure 3 Examples of web page logo. 2.3. Features for candidate titles Researchers have extracted a wide range of features from either DOM or vision-based tree. Those found in the literature are listed below. The features used in this paper are underlined. Features from DOM tree:  Visual: font weight, font family, font color (Changuel et al., 2009; Hu et al., 2005; Xue et al., 2007); font style, background color (Hu et al., 2005; Xue et al., 2007); alignment (Fan et al., 2011; Hu et al., 2005; Xue et al., 2007); and font size (Changuel et al., 2009; Fan et al., 2011; Hu et al., 2005; Xue et al., 2007);  HTML tag: bold, strong, emphasized text, paragraph, span, division (Changuel et al., 2009); image, horizontal ruler, line break, directory list (Hu et al., 2005; Xue et al., 2007); underline, list, anchor (Changuel et al., 2009; H u et al., 2005; Xue et al., 2007); meta; title (Changuel et al., 2009; Fan et al., 2011; Mohammadzadeh et al., 2012); heading level (h1- h6) (Changuel et al., 2009; Fan et al., 2009; Gali & Fränti, 2016; Hu et al., 2005; Xue et al., 2007); and position in tags (Gali & Fränti, 2016);  DOM structure: number of sibling nodes in the DOM tree (Hu et al., 2005; Xue et al., 2007); relation with the root, parent, sibling, next and previous nodes in term of visual format (Changuel et al., 2009; Hu et al., 2005; Xue et al., 2007);  Positional information: position of the text unit from the beginning of the body of the page and width of the text unit with respect to the width of the page (Hu et al., 2005);  Linguistic: length of text, negative words, positive words (Hu et al., 2005; Xue et al., 2007); position in text (Lopez, Prince, & Roche, 2011); syntactic structure, letter capitalization and phrase length; and  Statistical: term frequency (Mohammadzadeh et al., 2012); term frequency-inverse document frequency (Lopez et al., 2011; Mohammadzadeh et al., 2012); capitalization frequency, and independent appearance. Features from vision-based tree:  Page layout: height, width, and position relative to the top left corner (Xue et al., 2007);  Block: type, height, width, position (Wang et al., 2009; Xue et al., 2007); and front screen position (Wang et al., 2009);  Unit position: from the top and left side of the page and from the top and left side of the block (Xue et al., 2007); and  Content: number of words in a block (Wang et al., 2009). Other features:  Web page URL (Gali & Fränti, 2016). The majority of these features are based on formatting, whereas the features we consider are independent of the design of the page. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 2.4. Ranking candidate titles Ranking techniques can be divided into two broad classes: rule based and machine learning based (Xue et al., 2007). Rule- based techniques use a set of predefined heuristic rules to score the candidate titles. These rules are derived from the content of the DOM tree (Fan et al., 2009; Gali & Fränti, 2016; Mohammadzadeh et al., 2012), the link structure between web pages (Jeong, Oh, Kim, Lyu, & Kim, 2014), and the text (Lopez et al., 2011). The key advantage of the rule-based technique is that it does not require training data. Moreover, the technique is easy for humans to interpret and improve, as the weighting procedure and scoring formulas are explicit. However, heuristic methods often require determining thresholds and weights for feature parameters, which are not always straightforward to calculate. For example, if the number of features is n = 9 and each feature is assigned a value m = 0 to 5, it takes O(m n ) time to test all weight combinations. In this example, testing would take about four months if each attempt took 1 second. In contrast, Machine learning-based techniques involve two steps: training and testing. In training, the goal is to learn a set of rules that maps the inputs to outputs, so that the rules generalize beyond the training data. In testing, the generated classifier receives unseen data as input and predicts the output values. Proper training of the model is the key to generalizing the classifier beyond the training data. Several machine learning algorithms have been considered by the existing methods. These include perceptron (Li, Zaragoza, Herbrich, Shawe-Taylor, & Kandola, 2002), decision tree (C4.5) (Quinlan, 1993), random forest (Breiman, 2001), support vector machine (SVM) (Vapnik, 1995), and conditional random fields (CRF) (Lafferty, McCallum, & Pereira, 2001). While SVM has shown to be an effective classifier for the title extraction task, it has not been compared against simpler algorithms such as Naïve Bayes (Domingos & Pazzani, 1997), k-nearest neighbor (k-NN) (Cover & Hart, 1967), and clustering (Fränti & Kivijärvi, 2000), all of which we investigate in this paper. 3 Title extraction We consider the title extraction as a machine learning task, in which the computer is given a training data with assigned ground truth titles as the expected output. Three important issues are addressed: how to determine the candidate phrases, what features should be extracted, and which classifier to use. We add linguistic knowledge (syntactic POS) to the process to improve the extraction of the candidate phrases. The proposed method is based on four steps: extracting candidate phrases, feature extraction, phrase classification, and title selection (see Figure 4). A pre-processing step which involves corpus creation and learning POS patterns is applied before the training starts. The following subsections describe these steps in detail. Figure 4 Workflow for title extraction. 3.1. Corpus creation Several corpuses have been created to evaluate title extraction methods. Changuel et al. (2009) have created two corpuses on education domain. The first corpus contains 624 websites in English and French languages, and they were collected by submitting queries to the search engine such as chemistry + courses. The second corpus contains 424 websites in the French language, and they were collected from an online educational portal Eureka 4 . Lopez et al. (2014) have created a corpus of 300 news articles from three French newspaper websites, and they cover politics, sport, society, and science domains. 4 http://eureka.ntic.org/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT However, they do not consider web pages of places that offer services such as sports, hospitals, shops, banks, and restaurants, or web pages that host information about these places such as Wikipedia, Facebook, business directories, and information pages. These types of web pages are more challenging because they do not follow a certain template or standard format. None of these corpuses are available in public. We, therefore, built our corpus by collecting 1,002 unique websites from Google Maps 5 search results using queries such as restaurant + Australia, hospital + Canada, pharmacy + London, fitness + Ireland, and auto repair + California, to have reasonable geographical diversity and different layouts of websites. The main challenge of creating a corpus is that it is no t enough to store the web link and the ground truth title. The entire content of the web page should be stored because the links will become obsolete quite fast. Storing the content takes lots of space especially when the web page contains maps and images. The websites were collected during 18-31 July 2014 and 19-23 April 2015, and they cover various domains— food & drink, entertainment, auto & vehicles, beauty & fitness, health, sport, and hotels & accommodation. The resulting corpus is publicly available 6 . Similarly to the previous studies (Lopez et al., 2014; Wang et al., 2009; Xue et al., 2007), we created the ground truth by manually extracting the titles from the pages. We define the title as the most obvious description of the web page (see Figure 5) following the specifications in (Xue et al., 2007) with a few modifications of our own:  A web page can have more than one title, for example, we extracted V-Café and Viet-Café from http://www.viet- cafe.com/, C&A Bennett Ltd and C&A Bennett Tiling Contractors from http://www.candabennett-tiling-bristol.co.uk/;  The title cannot be a part of a numbering or bullets;  The title cannot be phrased like last updated, slogans like aim high go low, time, or address;  The title should not be too long;  The title must be concise and relevant to the page content;  The title must be grammatically correct;  The title must be understandable to humans. We did not use specifications such as the title must be in the top region of the page, or that the title cannot be a link because the correct title can be located on any part of the page. Further, a title can be clickable, especially in the case of business directory pages, in which the title is usually linked to the home page of the service. Two people were assigned to extract the ground truth titles independently on each other, and in the case of disagreement, a third person made a judgment between these two. Figure 5 Identifying the title of the web page 7 . 3.2. Learning POS tag pattern We define a set of specific POS tag patterns that correspond to the syntactic structure of the titles in the ground truth to remove the n-grams that have unwanted format. A POS tag is the part-of-speech label of a word in a text. For example, the 5 http://maps.google.com 6 http://cs.uef.fi/mopsi/TitlerCorpus/ 7 http://www.uef.fi/en/research/faculty-of-science-and-forestry ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT POS tag of the university is the_DT university_NN, where DT stands for determiner and NN stands for a noun. A POS tag pattern is a sequence of part-of-speech tags (e.g. <DT><JJ><NN>), where JJ stands for adjective. See Appendix A for a complete list of POS that we use in this paper. We first extract all n-grams (n=1 to 6) as candidate phrases. We observed that the number of the candidates is excessively high (1,024,142), of which only 2,179 phrases are in the title class. To evaluate them all would slow down the pr ocess and it would also cause a high-class size unbalance. There are also many grammatically incorrect title candidates among the n- grams like At Thai Food You, by Quay and Portishead Open Air Pool The. Most of these can be eliminated by applying the POS patterns. To generate the POS patterns, we have searched all POS tags that appeared among the ground truth titles in our corpus. We used the tagger developed by Stanford University 8 (Toutanova, Klein, Manning, & Singer, 2003). We observed that the following syntactic features are common for titles:  Starts with a general noun, proper noun, personal pronoun, foreign word, adjective, determiner, adverb, cardinal number, preposition, or verb;  Ends with a general noun, proper noun, personal pronoun, adjective, adverb, cardinal number, preposition, practical, verb or possessive ending;  In case it contains more than two words, the middle words are allowed to be a general noun, proper noun, personal pronoun, adjective, determiner, cardinal number, preposition, foreign word, coordinating conjunction, adverb, or verb;  Nouns appear much more than verbs see Table 2. Based on these observations, we generate 151 patterns with the length varying from 1 to 6 (see Appendix B). In our study, we follow traditional English grammar, in which a noun preceded by determiners or premodifiers such as adjectives is considered a noun phrase. It is important to note that this set of syntactic patterns is language-dependent and applicable only to the English language. The same process could also be done for other languages if a set of ground truth titles and POS tagger exist. Table 2 Features of ground truth titles POS tags Presence in title (%) Nouns and proper noun 83 Determiner 5 Coordinating conjunction 4 Adjective 3 Possessive ending 2 Preposition or subordinating conjunction 1 Verb 0.7 Cardinal number 0.7 Foreign word 0.2 Pronoun 0.2 Adverb 0.1 Particle 0.1 3.3. Candidate phrase extraction We first construct the DOM tree of the web page and strip it off the <script> and styling tags because their text content is mainly used for functionality and styling. We used XPath 9 , a query language for addressing parts of an XML document, to extract the text nodes as individual units from the tree. Then we add POS tags to each text node using Stanford Tagger. For example, a text node Kingston General Hospital becomes Kingston_NNP General_NNP Hospital_NNP after POS tagging. Later, we extract all phrases or (sequences of words) that match any of the patterns as potential candidates for titling. For example, we extract three nouns ―Kingston”, ―general”, ―hospital” and three phrases ―Kingston general”, ―general hospital”, and ―Kingston General Hospital” from Kingston_NNP General_NNP Hospital_NNP because they match the patterns <NNP>, <NNP>< NNP>, and <NNP><NNP><NNP>, which are all valid syntactic structures for titles. 3.4. Feature extraction For each candidate phrase, we extract the following ten features to evaluate its importance. 8 http://nlp.stanford.edu/software/tagger.shtml 9 http://www.w3.org/TR/xpath20/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT  Syntactic structure  Similarity with the link of the web page  Appearance in title tag  Appearance in meta tag  Popularity on the web page  Appearance in heading (h1, h2…h6) tags  Capitalization  Capitalization frequency  Independent appearance  Phrase length Syntactic structure (POS tag patterns) The structure of the ground truth title (as detailed in subsection 3.2), provides useful information to determine the title phrases. Different representations can convert the text features into compact formats such as TF, TF-IDF and binary (Lan, Tan, Su, & Lu, 2009). We use binary representation because we aim at distinguishing the structure of the candidate phrases, but not to determine the importance of the individual tags. In the analysis, we observed that the maximum word counts for the titles are 6. Therefore, we compute a vector of size (21×6=126) to represent the syntactic structure feature. This transformation is similar to token-vector approach, in which the order of the tokens is preserved.  Each title is represented by its POS tags. For example, The Path Café is represented by DT|NN|NNP|, and Leon Restaurants is represented by NNP|NNPS|.  The tags are then binarized by giving each 21 possible slot, and activating the slot corresponding to a specific tag. We recognize 20 POS tags in the titles, and we use slot 21 for other tags. For example: The Path Cafe ‗ ‘ ‗ ‘ ‗ ‘ DT NN NNP [001000000000000000000] [000000001000000000000] [000000000010000000000] [0] [0] [0] Leon Restaurants ‗ ‘ ‗ ‘ ‗ ‘ ‗ ‘ NNP NNPS [000000000010000000000] [000000000001000000000] [0] [0] [0] [0] We later denote this feature as POS feature. Similarity with the link of the web page Words in the link of the web page are usually precise and relevant to the content of the page; therefore, we hypothesize that a candidate phrase that has high similarity to these words is more likely to be the title of the page. We compute the similarit y between the phrase and the words in the web link using the Dice coefficient (Brew & McKelvie, 1996). It counts the number of shared character bigrams divided by the total number of bigrams in both strings: ( ) | ( ) ( )| | ( )| | ( )| (1) where bigrams is the number of adjacent character pairs in the candidate phrase (p) and the words in the link (s) respectively. The reason for choosing this measure is that it is robust to the change of the order of the words, and it treats strings with small differences as similar. These kinds of variations are expected between the web page link and the extracted phrases; an exact match would not be as useful. A measure like edit distance would recognize the reverse order of two strings as a mismatch. For example, the edit distance between the two strings nba mcgrady and macgrady nba is very low, although they refer to the same title (Wang, Li, & Feng, 2014). We normalize the similarity scores to the scale [0, 1] as follows: ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT ( ) ( ) ( ( )) (2) where r is the number of candidate phrases from the web page. Appearance in title and meta tags The content of the title tag is the second most important source that yields useful hints for the correct title. The meta tag with name=title or property="og: title" usually contains the same information as the title tag when it exists. Therefore, a candidate phrase that appears in these tags is valuable. The scores resulting from the comparison of the candidate phrase, the title and the meta tags are binary: 1 if a match is found in any of the tags; 0 otherwise. Popularity on the web page We use term frequency (TF) to count the number of times a candidate phrase appears on the web page in total. More popular phrases have better chances of being a title. We normalize the score by the frequency of most popular phrase (pr) to the scale [0, 1] as follows: ( ) ( ) ( ( )) (3) Appearance in heading tags Heading (h1, h2…h6) tags emphasize the headlines and important text on the web page. We consider a phrase that appears in any of the heading tags more important than other candidates. We navigate through the entire page and count the number of times a candidate phrase appears in each heading. A phrase within <hx></hx> tag is given a score between [0, 1] as follows: ( ) ∑ ( ) ( ) ( ( )) (4) where fi is the frequency of a candidate phrase (p) in heading hi, and wi is the weight of heading hi. Similarly to Fan et al. (2011), the weights of the headings are fixed to (6, 5, 4, 3, 2, 1) respectively, with h1 having the biggest weight and h6 the lowest weight, m is the total number of candidate phrases that appear in all heading tags. Capitalization and capitalization frequency Capitalization feature takes value 1 if the candidate phrase starts with a capital letter; otherwise, it will be 0. The hypothesis is that majority of the titles starts with a name. However, a phrase might also start with a capital letter depending on its position in the text such as sentence-initially; therefore, we also consider capitalization frequency. It counts the number of times a candidate phrase starts with a capital letter on the web page. We use Eq. 3 to compute the normalized score for capitalization frequency, (but counting only when the phrase starts with a capital letter). Independent appearance Important phrases such as a title and a headline can appear separately in the nodes of the DOM tree. Therefore, we consider a candidate phrase that appears independently more important than a phrase that appears as a part of other phrases in the text nodes. We calculate the normalized score similarly as in Eq. 3, (but counting only when the phrase appears individually on the web page). Phrase length The last feature is the length of the candidate phrase. Figure 6 shows the distribution of the title and non-title phrases in our corpus. Most extracted phrases are less than eight characters long; however, they are not as likely to be titles as phrases which are longer. Phrases longer than 8 characters are more likely to be title phrases, and phrases shorter than this are less likely to be the title. We compute the score of the length feature by counting the number of phrases that have the same length of the candidate phrase divided by the most frequent length. For example, if the length is 16 characters and the frequency of this ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT length is 69, then the length score is 69/124= 0.56, where 124 is the frequency of the most frequent length (9 character-long titles). Figure 6 Probability of length for titles and candidate phrases: characters (left) and words (right). 3.5. Classification of candidate phrase After the features for all candidate phrases have been computed, we generate the feature vectors and manually label them either as title or non-title, based on the similarity of the candidate phrase with the ground truth titles. In the case of a perfect match, we label the feature vector as a title; otherwise, we label it as non-title. We then proceed by training different models to extract the titles. We consider four alternative classifiers: Naive Bayes, clustering, k-NN, and SVM. The four classifiers consist of two phases: learning and testing. In the learning phase, the sequences of feature vectors {(X1, Y1), (X2, Y2)… (Xn, Yn)} from the training set are the input to the classifier. Here Xi denotes a feature vector of a candidate phrase; Yi denotes the class label of the vector Xi, and n is the number of the candidate phrases in the training set. In the testing phase, a sequence of feature vectors {X1, X2…Xr} with unknown class labels is input to the classifier, and the output is a sequence of predicted labels {Y1, Y2…Yr}, where r denotes the number of candidate phrases on a web page. 3.5.1. Naive Bayes Naive Bayes (Domingos & Pazzani, 1997) is a probabilistic model, in which conditional probabilities are calculated from the training data. Classification is done by taking the one with the highest probability given by the features. It is based on Bayes‘ theorem (Duda, Hart, & Stork, 2001). Naive Bayes assumes that the probability of a feature is independent of the other features. Given a feature vector X=(x1, x2,…, xD) and a class label (y), the goal is to find the most likely class for X. In the learning phase, we estimate the prior probability P(y) for each class y{title, non-title} by counting their relative frequency in the training data. We also use the training data to estimate the conditional probability P(xi|y) for each feature xi occurring in the class y. We use multinomial distribution (Manning, Prabhakar, & Hinrich, 2008) to estimate the probability P(xi|y) because it performs better than Gaussian and Bernoulli models for the text classification task according to (McCallum & Nigam, 1998; Manning et al., 2008). In the testing phase, we calculate the posterior probability ( ̂) for each class given the feature vector, and then assign the candidate phrase to the most probable class. We select the phrase with the highest title probability. ̂ ( )∏ ( | ) (5) 3.5.2. Clustering-based The overall process of the clustering-based model is shown in Figure 7. In the learning phase, we cluster the feature vectors from the training dataset, using Randomized Swap algorithm (RS) (Fränti & Kivijärvi, 2000). Random Swap alternates between performing k-means and randomly relocating centroids to allow k-means to escape local optimum. It has a time complexity of O(MNDI) where M is the number of clusters, N the number of feature vectors, D is the number of dimensions and I the number of swaps. A higher number of iterations typically yield better results in terms of total squared error. After we obtain the clustering solution, we determine the dominant label of each cluster. For example, if a cluster contains 20 titles and 16 non-titles, it is labeled as a title cluster with 56 % probability. Likewise, a cluster that contains 6 titles and 130 non-titles is labeled as a non-title cluster with 96 % probability. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT In the testing phase, we compute the feature vector for all candidate phrases and map them to the nearest cluster centroid using Euclidean distance in the feature space. This takes O(MD) time per candidate title. The candidate phrase that is classified to a cluster with the highest title probability is chosen. Figure 7 Clustering-based title extraction. 3.5.3. k-nearest neighbors (k-NN) K-NN (Cover & Hart, 1967) does not train any model, but it simply stores the feature vectors of the training data as such. In the testing phase, we compute the feature vector for all candidate phrases and map them in the feature space. The k-nearest training vectors are found for every new unlabeled feature vector, and its class is selected by the majority rule: a class that has the most representatives within the nearest neighbors is chosen. We use Euclidean distance to find the neighbors. We choose the phrase that has the highest number of neighbors with the label title as the page title. In the case of a tie, we just select the class of the first neighbor found. 3.5.4. Support Vector Machines (SVMs) Support vector machine (Vapnik, 1995) is a binary classifier that represents the feature vectors as points in the feature space so that points of different classes are separated by the so called maximum-margin hyperplane. SVM has shown to be well suited for text categorization ((Joachims, 1998). In the learning phase, SVM aims at finding an optimal hyperplane that maximally separates the two classes of the training data yi ϵ {title, non-title}. We use radial basis function (RBF) kernel as it performs better than linear and polynomial kernels (see subsection 4.3). Given two feature vectors X and X´, the RBF kernel is defined as follows: ( )́ ( ‖ ́‖ ) (6) where ‖ ́‖ is the squared Euclidean distance between the feature vectors, and is a parameter, which defines how much influence a single training example has. In the testing phase, a new unlabeled feature vector is mapped into the same space and predicted to belong to a class based on which side of the hyperplane it falls into. We select the one with the largest distance to the support vector as the title. 4 Experiments We perform a set of experiments to evaluate our proposed method. Experiments also reveal the effect of the POS patterns and the best parameter selection for the method. We compare to all existing popular methods such as Title tag (the baseline), Google, Title Tag Analyzer (Gali & Fränti, 2016), TitleFinder (Mohammadzadeh et al., 2012) and Styling (Changuel et al., 2009). We use two datasets in our evaluation: Titler and Mopsi services. The former is used for training and testing by following a k-fold cross validation strategy. The latter is a more difficult, multiple language dataset, which we use to stress test our models. Furthermore, we perform an evaluation of the features. There is no question that the features are useful, however, some features‘ usefulness is ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT diminished in the presence of others. This complementary aspect varies depending on the method (classifier) used. We conclude our experiments by especially addressing the newly proposed POS feature and its usefulness when combined with the strategy of segmenting the text inside the DOM tree nodes. 4.1. Datasets We use the Titler corpus detailed in subsection 3.1. After the POS patterns had been identified, we extracted 152,163 unique candidate phrases from all the web pages. The number of phrases per page varies from 2 to 3,166 (152 phrases on average). All phrases in a web page that match the given ground truth title are labeled as title, the rest as non-title. The labeling was done manually. We then divide the data into five folds to conduct a cross-validation. In each iteration, we use four folds (80%) of the data for training and the remaining (20%) for testing. In addition, we use Mopsi services dataset 10 . Mopsi (Fränti, Chen, & Tabarcea, 2011), implements various location-based services and applications such as mobile search engine, data collection, user tracking, and route recording. It has applications integrated both on web and in mobile phones. Mopsi services contain 414 user collected places. Each service has a web link, title, keywords, photo, additional description (optional), and address that are manually added by the users and further confirmed by an administrator. For the sake of this study, we also downloaded the content of web pages and stored them so that the same content is available for further tests. Mopsi services include a wide variety of web links such as home, brand, business directory, Wikipedia, Facebook, blog and information pages and they cover various domains such as education, health, shop, bank, entertainment, sport, food & drink, hotel & accommodation, travel & leisure, and news. The pages include several languages such as Finnish (375 pages), English (35 pages), and few pages in French, Italian, Spanish, Estonian and Swedish. This will put the developed method also in stress-test since the method covers only English. In both datasets, the relevant information about the services such as the title, description and opening hours is commonly static because users are expected to rely on this information. However, if the target pages are expected to be fully dynamic, we recommend to use a WebKit browser such as PhantomJS 11 to render the web page prior to the application of our method. 4.2. Evaluation measures To evaluate the performance of the title/non-title classifier, we use precision, recall, and F-score, which are widely used to evaluate information extraction systems. We count the following classification results: tp = number of candidate phrases correctly identified as titles fp = number of candidate phrases incorrectly identified as titles fn = number of correct candidate phrases erroneously identified as non-titles. (7) (8) (9) To evaluate the quality of the extracted title phrases, we use four different measures: 1. Rouge-N (Lin, 2004): a well-known N-gram metric that measures the overlap units between the candidate and the ground truth titles. It is defined by the precision, recall, and F-score as follows: (10) 10 http://cs.uef.fi/mopsi/MopsiSet 11 http://phantomjs.org/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT (11) ( ( ) ( ) ( )) (12) Here cd and gt are the number of n-grams in the candidate titles and the ground truth, and cm is the number of common n- grams co-occurring in both of them. We use Rouge-1 as was reported to work best for a very short text in (Lin, 2004). 2. Jaccard index: It counts the number of common character bigrams divided by the total number of unique bigrams in both strings: ( ) | ( ) ( )| | ( ) ( )| (13) where cd stands for candidate title and gt stands for ground truth title. 3. Dice coefficient: It counts the number of shared character bigrams divided by the total number of bigrams in both strings; see Eq.1. 4. Human judgment: we developed a tool 12 (see Figure 8) to allow users to rank the candidate phrases. Phrases that have either high character-level or high word-level similarity (>0.8) with the ground truth titles are shown. The user then rates the candidates using a scale from 0 (irrelevant) to 5 (exact match). The quality of the chosen candidate phrase is measured as the average of the human ratings. We refer the human judgment as accuracy in the rest of the paper. It counts the number of titles having average ratings > 1. We further apply the Mann Whitney U-test for significance testing. Figure 8 Evaluation tool 13 for human ratings. 4.3. Classifiers setup The parameters of the classifiers were optimized by brute force (grid search) using the training set. The tested parameter values and the best obtained combinations are shown in Table 3. Table 3 Best combination of parameters obtained for each classifier Classifier Best obtained parameters Testing range Bayes Laplace smoothing =1 Lidstone smoothing = 0.1 – 0.9, Laplace smoothing =1 Clustering clusters = 2048 clusters = 256 – 8192 by doubling KNN k = 30 k = 1 – 40 Euclidean distance Manhattan distance, Euclidean distance Uniform Weight of neighbors: uniform, weighted SVM c = 2 10 c = 2 5 – 2 20 , increment by 5 Kernel= RBF Kernel = linear, polynomial, RBF In clustering, the problem of overfitting happens after 4096 (see Figure 9); therefore, we select 2048. For SVM, we subsample the training set by randomly discarding examples from the majority class (non-title) until the training set becomes 12 http://cs.uef.fi/mopsi_dev/Titler/TitleRater.php 13 http://www.shirehotels.co.uk/aztec/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT balanced. Table 4 shows the results with linear, polynomial and RBF kernels. The RBF kernel provides slightly better result in the recall and the F-score in comparison with polynomial and linear kernels respectively and is therefore selected. We use RBF with c=2 10 and for the further experiments. For all learning models, we conduct a five-fold cross-validation. All the results reported here are averaged over five trials. Figure 9 Titling accuracy with different numbers of clusters. Table 4 Performance of SVM with different kernel functions SVM model type Precision Recall F-score Linear kernel 0.44 0.84 0.57 Polynomial kernel 0.47 0.79 0.59 RBF kernel 0.44 0.84 0.58 4.4. Methods evaluated We compare the following methods:  Title tag (baseline)  Google (search engine)  Title Tag Analyser (TTA) (Gali & Fränti, 2016)  TitleFinder (Mohammadzadeh et al., 2012)  Styling (Similar to Changuel et al., 2009)  Titler (BAYES) (proposed)  Titler (CLUS) (proposed)  Titler (KNN) (proposed)  Titler (SVM) (proposed) As a baseline, we use the content of the title tag as such. We compare to the titles provided by Google search e ngine in the results page. We also compare our method with TTA which is a simplified version of the proposed method where only title and meta tag segmentation is applied but without POS tag patterns and other linguistic features, and with TitleFinder which uses the content of the title tag as a feature. We re-implemented a Styling method similar to Changuel et al. (2009) with the following features: Tags  Level of headings: h1…h6;  Div: division or section;  Span: group inline elements;  P: paragraph;  A: anchor; ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT  Strong: important text;  B: bold;  U: underline;  I: list item;  Em: emphasized text. Formats  Font size: level 1-7;  Font weight: bold;  Font family: Arial, Calibri…  Font color: RGB values converted to the YUV color space;  Alignment: top, left… Format and tag changes  All formats and tags change with the previous nodes;  All formats and tags change with the next nodes. We use decision tree and random forest classifiers as in Changuel et al. (2009). We did not compare with Xue et al. (2007), which is an updated version of Hu et al. (2005) as the authors mentioned 254 features, but only a few are explained to allow reproducing. However, Changuel et al. (2009) uses very similar features to (Xue et al., 2007). Both methods use visual and formatting features to extract the titles. In Section 4.9, we demonstrate that these kinds of features are disadvantageous for service-based web pages. 4.5. The POS patterns After applying the POS patterns in the text of the web pages, we observed that the exact titles are found by the patterns in 93% of the web pages and approximate titles in 4.7% of the cases according to the human judgment. The patterns fail to extract correct title phrases in 2.3% of the pages (see Table 5). The failures happen with web pages that have the exact title as an image and do not provide useful text (see Figure 10, right). Table 5 Summary of the titles extracted by POS patterns Type of title Number of web pages Proportion (%) Perfect match 932 93 Approximate match 47 4.7 Not found 23 2.3 A further analysis of the patterns revealed that 36% of the extracted title phrases match the pattern <NNP><NNP>, 17% match <NNP><NNP><NNP>, and 14% match <NNP>, which are all proper nouns (see Table 6). Patterns <JJ><VBG>NNP> and <DT><CD><NNPS> are rare as only 0.3% of the title phrases match these patterns. However, they are effective because only titles appear in this format. About 2% of the title phrases are extracted by the pattern <VB>, while it extracts 5% of the non-title phrases, which makes it less important. No significant observations were found from the rest of the patterns. We conclude that noun patterns have the highest impact in comparison to the other type of patterns such as verbs, which extract only 2% of the titles when they are used separately. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Figure 10 Exact title appears as an image and partially in the text (left) 14 and only as an image (right) 15 . Table 6 Summary of the effectiveness of individual patterns Type of pattern Example of titles Proportion (%) <NNP><NNP> Aqua Dining 36 <NNP><NNP><NNP> Woolwich Pier Hotel 17 <NNP> Ventuno 14 <VB> Recreate 2 <JJ><VBG>NNP> Functional Training Ireland 0.3 <DT><CD><NNPS> The Five Bells 0.3 Others Spice Me UP 30 4.6. Feature selection We next investigate the importance of the individual features to evaluate their potential benefit. We use a wrapper method called greedy backward elimination (Kragh, Jørgensen, & Pedersen, 2015). It starts by using all the features in combination, and then gradually removes one feature at a time. At each step, the classifier is trained with the current combination of the features and the average of five-fold cross validation is used as a score. The feature whose removal improves the performance most is eliminated for the next iteration. The process continues iteratively until no further improvement. Results are reported in Table 7. Although all features are relevant, some of them are not useful in the presence of others. Capitalization frequency is disadvantageous in both BAYES and SVM. Title length seems to be a bad feature in BAYES and CLUS but helps the others. Likewise, heading and frequency features seem to be bad in CLUS and SVM respectively, but useful in BAYES and KNN. Both CLUS and KNN require fewer features (6 in CLUS and 8 in KNN) to provide the best results (82.4% and 84.8%). This is because, in high dimensional space, the distance to the nearest data point approaches the distance to the farthest data point when the dimension increases. Consequently, distance-based classifiers become less accurate when too many features are used. The same phenomena were also observed in (Beyer, Goldstein, Ramakrishnan, & Shaft, 1999). The experiments show that the similarity with the link of the web page has the highest impact on all models. By removing this feature, the accuracy would decrease significantly: BAYES: 54.8% → 41.1%, CLUS: 70.6% → 57.8%, KNN: 79.2% → 62.8%, SVM: 84.9% → 56.4%. Surprisingly, meta feature is harmful in all models and its removal always improves the performance. This is because the meta tag is rarely found on the page and even when existing, the appearance in meta tag feature fails. The rest of the features show a positive effect on the overall accuracy. Table 7 Impact of the features on the accuracy (%) after eliminating the corresponding feature BAYES CLUS KNN SVM All features (54.8) All features (70.6 ) All features (79.2) All features (84.9) − Capt. Freq. (58.5) − POS (77.4) − POS (82.8) − Capt. Freq. (85.6) − Title length (59.6) − Heading (80.9) − Meta (84.8) − Frequency (85.7) − Meta (60.1) − Title length (80.9) − Meta (85.9) − Meta (82.4) 14 http://www.edensauna.com/ 15 http://oxfordartfactory.com/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT 4.7. The classifiers We evaluate the performance of the classifiers using the best set of features found in subsection 4.6. The results in Figure 11 show that CLUS (0.80) and KNN (0.81) outperform BAYES (0.60) and SVM (0.44) in the precision, whereas SVM provides better results in the recall (0.84) and F-score (0.58). SVM classifies several non-title phrases incorrectly in the title class, which harms the precision, but it still classifies the majority of the title phrases correctly in the title class. A less number of non-title phrases are classified in title class by other models, which provides better precision, but the percentage of the title phrases correctly classified in title class are smaller when compared with SVM, and therefore the recall is low. Figure 11 Performance results for the classifiers. 4.8. Title selection Among the phrases that are classified as a title, we select the one with the highest confidence. We compare this against random choice. Figure 12 show that there is only a small difference between these two approaches in case of BAYES, CLUS, and KNN. In the case of SVM, however, the random choice works poorly (50%). This is because several non-title phrases (FP) are classified in the title class by SVM. The probability of choosing a non-title phrase by random approach is, therefore, high. The title class in other models is more balanced; therefore a random selection performs well. Figure 12 Accuracy of title selection methods. We also compare the quality of the correctly extracted titles to the human judgment. In Figure 13, we observe that the majority of the titles match perfectly with the ground truth titles for all models. BAYES tends to provide shorter titles, such as Cava for Cava restaurant while SVM, CLUS, and KNN tend to provide more descriptive titles such as Home Leisure Direct. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Figure 13 Quality of titles (confidence 5 is the perfect match and confidence 1 is satisfactory). 4.9. Comparative results for all methods The overall results are compared to those existing methods that were available, and to the baseline (title tag) using the Titler corpus and Mopsi service dataset. The results of the Rouge-1, Jaccard, and Dice measures are summarized in Tables 8 and 9. The proposed method clearly outperforms the Baseline, Google, TitleFinder, Styling and TTA except for BAYES, for the Titler corpus. One reason is that many web pages do not have their titles appear independently in the text nodes, and therefore, selecting the entire text node as in TitleFinder and Styling is harmful. This decrease their overall performance (TitleFinder 0.47) and (Styling 0.35). Our method extracts only potential substrings from each node. The recall of the Baseline and Google is high (0.89) because many web pages include the correct titles in their title tag, but also include other irrelevant text which causes a decrease in the precision (0.41) and (0.48) respectively. The Baseline, Google, TitleFinder, TTA and Titler all provide better results than Styling method because the majority of the web pages (88.8%) has their titles integrated within the logo image and presented in the text only much later with no styling differenc es from the other parts of the text. In this kind of web pages, the more visual focus is given to the advertisements or products (see Figure 14). Therefore, style-dependent features do not work well in general. We also conducted statistically significance tests (Mann Whitney U-test) for the Rouge-1. The results indicate that the advantages of our method over the Baseline, Google, TitleFinder, TTA and Styling method are statistically significant (P value < 0.05). No significance difference is noticed between SVM, KNN, and CLUS. We also count the number of titles that perfectly match with the ground truth titles. The methods find the perfect match as follows: SVM 700, KNN 677, CLUS 644, and BAYES 406 times. These are significantly more than that of the TTA 555, TitleFinder 212, Google 192, Styling 162, and Baseline 149. When testing with Mopsi dataset, we use KNN classifier because it performs well without the linguistic feature (POS) as shown in subsection 4.6. We extract all n-grams as candidate phrases because we do not have POS taggers for all the languages in the set. From Table 9, we observe that the proposed method still outperforms other methods even though the overall result decreases compared to Titler corpus. One reason is that Mopsi set is more challenging due to the high diversity of the domains, functions, and templates of the web pages. Another reason is that, unlike Titler corpus, the titles in Mopsi services are annotated and may not always appear on the page in the text at all. To test the effect of this, we added alternative ground truth titles by manual extraction. This improved the F-scores of all methods: Titler (0.55 → 0.70), TTA (0.52 → 0.67), Google (0.43 → 0.52), Baseline (0.41 → 0.52), TitleFinder (0.37 → 0.47), and Styling (0.15 → 0.20). These results show that the real problem is more challenging than just extracting the best phrase from the text. The existence of multiple languages in the dataset also decreases the performance since the linguistic model was designated for English only. TTA performs fairly well, which indicates that the title tag itself is still useful, but its segmentation is important. The performance of the Styling method is still the lowest among all the methods tested. We have also tested the Styling method with two small sets of web pages from Eureka 16 an online educational portal used by Changuel et al. (2009) and Wikipedia. The accuracies were 85% and 70% respectively. This confirms that the methods depend only on the styling information are suitable for web pages that follow a standard format such as news and education pages, but not for more general pages as in our case. 16 http://eureka.ntic.org/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Table 8 Comparative results for Titler corpus Method Rouge-1 Jaccard Dice Precision Recall F-score Baseline 0.41 0.89 0.52 0.50 0.58 Google 0.48 0.89 0.58 0.56 0.64 TitleFinder (Mohammadzadeh et al., 2012) 0.43 0.61 0.47 0.43 0.50 Styling (Changuel et al., 2009) 0.36 0.41 0.35 0.38 0.43 TTA (Gali and Fränti, 2016) 0.73 0.80 0.75 0.75 0.78 Titler (BYAES) 0.46 0.40 0.42 0.64 0.70 Titler (CLUS) 0.87 0.81 0.82 0.82 0.86 Titler (KNN) 0.87 0.82 0.83 0.84 0.87 Titler (SVM) 0.88 0.84 0.85 0.85 0.88 Table 9 Comparative results for Mopsi services Method Rouge-1 Jaccard Dice Precision Recall F-score Baseline 0.33 0.71 0.41 0.44 0.54 Google 0.34 0.74 0.43 0.46 0.56 TitleFinder (Mohammadzadeh et al., 2012) 0.35 0.47 0.37 0.37 0.43 Styling (Changuel et al., 2009) 0.14 0.21 0.15 0.22 0.28 TTA (Gali and Fränti, 2016) 0.52 0.59 0.52 0.54 0.62 Titler (KNN) 0.59 0.56 0.55 0.59 0.66 Title ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Figure 14 Example of a web page gives visual focus on the products and events (Muumimaailma) 17 . 4.10. Effect of segmentation The biggest deficiency of the existing methods is that they use the content of the text nodes (or title tag) as such. We tested how much the methods can be improved by segmenting the content of the nodes (N-grams), and how much by using the POS patterns to filter out the unwanted phrases. We consider two methods: TitleFinder and Titler (with k-NN). We exclude Styling method because it mainly depends on the features within the surrounding text, which would be the same for all candidate phrases in the same node, and therefore, segmentation would not make any difference. Results in Table 10 show that both TitleFinder and Titler are improved by the segmentation, and the effect is significant. When using N-gram phrases, Titler is improved to 0.73 whereas TitleFinder to 0.55. Using POS patterns improves further to 0.83 and 0.60. We conclude that segmentation plays an important role in title extraction task. We further observe that the POS patterns have another important effect that is not shown in the numbers. After segmentation, the structure of the titles often becomes incorrect, such as ―in Bristol Aztec Hotel & Spa‖ and ―to Essen restaurant‖ but these are still rated as correct in the numerical evaluation. The use of POS pattern is, therefore, important processing step needed with the segmentation to avoid grammatically incorrect titles. Table 10 Effect of phrase segmentation (N-grams) and POS patterns with TitleFinder and Titler methods Method Pre-processing Rouge-1 Jaccard Dice Precision Recall F-score TitleFinder (Mohammadzadeh et al., 2012) Complete node 0.43 0.61 0.47 0.43 0.50 N-grams 0.53 0.65 0.55 0.53 0.60 POS patterns 0.61 0.63 0.60 0.58 0.63 Titler (KNN) Complete node 0.46 0.54 0.47 0.51 0.58 N-grams 0.79 0.73 0.73 0.74 0.80 POS patterns 0.87 0.82 0.83 0.84 0.87 5 Conclusion We have proposed a new method to extract the title from HTML web pages using text segmentation, statistical features of the main content, and linguistic knowledge. The proposed method outperforms the best existing method (Google) by a large margin; F-score of the Rouge-1 measure is improved from 0.58 to 0.85 in the case of Titler corpus, and from 0.43 to 0.55 in the case of Mopsi services. More detailed analysis revealed the following findings that explain the results:  Segmentation of the text nodes is most important. Results showed that it is not enough to find the correct text node from the DOM tree, or use the title tag as such. Segmenting the content to n-grams and further analysis by POS tagging improved both TitleFinder (0.47→0.60) and the proposed method Titler (0.47→0.83) significantly.  The link of a web page has the highest impact on the overall performance of the method. It works well because the majority of the web pages have the title also appear in the link. Meta feature is not useful possibly because it is rarely specified and given a correct value.  SVM model achieved the highest F-score (0.85) for the title extraction task, but with no significance difference from k-NN (0.83) and clustering (0.82) models. Both k-NN and clustering are simple to implement and easy to understand in comparison with the underlying theory behind SVM. Naive Bayes achieved the lowest F-score (0.42).  The POS patterns models can be generalized to any language provided that a corpus and a POS tagger for the specified language are available. Our k-NN model works reasonably well with other languages; however, we did not experiment extensively in this regard.  Noun phrases are more effective than adjective, determiner, propositional, adverb and verb phrases.  A drawback of the POS tagging of the web pages is the slowness, especially when the amount of text is big.  A large number of pages (89%) show the title within a logo image. The title in the text content is therefore not highly emphasized or might even be missing completely. This gives the biggest challenge to the title extraction. An alternative approach would be to use image analysis, but this raises even bigger challenges. First, one should detect which one is the logo image (see Gali, Tabarcea, & Fränti, 2015) for a possible approach). Second, the content of the image are highly complex, and standard OCR approach would not work as such. 17 http://www.moominworld.fi/ ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT The proposed method works well on both static and dynamic web pages because the most relevant content is usually static. If the target pages are expected to be fully dynamic, we recommend to render the web page prior to the application of our method. References Beyer, K., Goldstein, J., Ramakrishnan, R., & Shaft, U. (1999, January). When is ―nearest neighbor‖ meaningful? In International conference on database theory (pp. 217-235). Springer Berlin Heidelberg. Breiman, L. (2001). Random forests. Machine learning, 45(1), 5-32. Brew, C., & McKelvie, D. (1996, September). Word-pair extraction for lexicography. In Proceedings of the 2nd International Conference on New Methods in Language Processing (pp. 45-55). Cai, D., Yu, S., Wen, J. R., & Ma, W. Y. (2003). VIPS: a vision-based page segmentation algorithm. Microsoft technical report, MSR-TR-2003-79. Changuel, S., Labroche, N., & Bouchon-Meunier, B. (2009, July). A general learning method for automatic title extraction from html pages. In International Workshop on Machine Learning and Data Mining in Pattern Recognition (pp. 704- 718). Springer Berlin Heidelberg. Cover, T., & Hart, P. (1967). Nearest neighbor pattern classification. IEEE transactions on information theory, 13(1), 21–27. Domingos, P., & Pazzani, M. (1997). On the optimality of the simple Bayesian classifier under zero-one loss. Machine learning, 29(2-3), 103–130. Duda, R. O., Hart, P. E., & Stork, D. G. (2001). Pattern classification. 2nd.Edition. New York, NY. Fan, J., Luo, P., & Joshi, P. (2011, February). Identification of web article pages using HTML and visual features. In IS&T/SPIE Electronic Imaging (pp. 78790K-78790K). International Society for Optics and Photonics. Fränti, P., Chen, J., & Tabarcea, A. (2011, May). Four aspects of relevance in sharing location-based media: content, time, location and network. International Conference on Web Information Systems and Technologies, 413–417. Fränti, P., & Kivijärvi, J. (2000). Randomised local search algorithm for the clustering problem. Pattern Analysis & Applications, 3(4), 358–369. Gali, N. & Fränti, P. (2016). Content-based title extraction from web page. International Conference on Web Information Systems and Technologies, 204-210. Gali, N., Tabarcea, A., & Fränti, P. (2015). Extracting Representative Image from Web Page. In International Conference on Web Information Systems and Technologies. Gibson, D., Punera, K., & Tomkins, A. (2005, May). The volume and evolution of web page templates. In Special interest tracks and posters of the 14th international conference on World Wide Web (pp. 830-839). ACM. Hulth, A. (2003, July). Improved automatic keyword extraction given more linguistic knowledge. In Proceedings of the 2003 conference on Empirical methods in natural language processing (pp. 216-223). Association for Computational Linguistics. Hu, Y., Xin, G., Song, R., Hu, G., Shi, S., Cao, Y., & Li, H. (2005, August). Extraction from bodies of html documents and its application to web page retrieval. In Proceedings of the 28th annual international ACM SIGIR conference on Research and development in information retrieval (pp. 250-257). ACM. Jeong, O. R., Oh, J., Kim, D. J., Lyu, H., & Kim, W. (2014). Determining the titles of Web pages using anchor text and link analysis. Expert Systems with Applications, 41(9), 4322-4329. Joachims, T. (1998, April). Text categorization with support vector machines: Learning with many relevant features. In European conference on machine learning (pp. 137-142). Springer Berlin Heidelberg. Kragh, M., Jørgensen, R. N., & Pedersen, H. (2015, July). Object Detection and Terrain Classification in Agricultural Fields Using 3D Lidar Data. In International Conference on Computer Vision Systems (pp. 188-197). Springer International Publishing. Lafferty, J., McCallum, A., & Pereira, F. (2001, June). Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the eighteenth international conference on machine learning, ICML (Vol. 1, pp. 282-289). ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Lan, M., Tan, C. L., Su, J., & Lu, Y. (2009). Supervised and traditional term weighting methods for automatic text categorization. IEEE transactions on pattern analysis and machine intelligence, 31(4), 721-735. Lin, C. Y. (2004, July). Rouge: A package for automatic evaluation of summaries. In Text summarization branches out: Proceedings of the ACL-04 workshop (Vol. 8). Li, Y., Zaragoza, H., Herbrich, R., Shawe-Taylor, J., & Kandola, J. (2002, July). The perceptron algorithm with uneven margins. In ICML (Vol. 2, pp. 379-386). Lopez, C., Prince, V., & Roche, M. (2010, December). Automatic titling of electronic documents with noun phrase extraction. In Soft Computing and Pattern Recognition (SoCPaR), 2010 International Conference of (pp. 168-171). IEEE. Lopez, C., Prince, V., & Roche, M. (2011, December). Automatic titling of articles using position and statistical information . In RANLP'11: Recent Advances in Natural Language Processing (pp. 727-732). Lopez, C., Prince, V., & Roche, M. (2014). How can catchy titles be generated without loss of informativeness? Expert Systems with Applications, 41(4), 1051–1062. Manning, C. D., Prabhakar, R., & Hinrich, S. (2008). Introduction to information retrieval, volume 1 Cambridge University Press. Cambridge, UK. Marchlonini, G. (1992). Interfaces for end-user information seeking. Journal of the American Society for Information Science (1986-1998), 43(2), 156. McCallum, A., & Nigam, K. (1998, July). A comparison of event models for naive bayes text classification. In AAAI-98 workshop on learning for text categorization (Vol. 752, pp. 41-48). Mohammadzadeh, H., Gottron, T., Schweiggert, F., & Heyer, G. (2012, November). TitleFinder: extracting the headline of news web pages based on cosine similarity and overlap scoring similarity. In Proceedings of the twelfth international workshop on Web information and data management (pp. 65-72). ACM. Quinlan, J. R. (1993). Machine Learning, C4. 5: Programs for machine learning. San Francisco: Morgan Kaufmann Publishers Inc. Song, R., Xin, G., Shi, S., Wen, J. R., & Ma, W. Y. (2006, April). Exploring URL hit priors for web search. In European Conference on Information Retrieval (pp. 277-288). Springer Berlin Heidelberg. Toutanova, K., Klein, D., Manning, C. D., & Singer, Y. (2003, May). Feature-rich part-of-speech tagging with a cyclic dependency network. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology-Volume 1 (pp. 173-180). Association for Computational Linguistics. Vapnik, V. (1995): The nature of statistical learning theory. New York: Springer. Wang, C., Wang, J., Chen, C., Lin, L., Guan, Z., Zhu, J., Zhang, Ch., & Bu, J. (2009, July). Learning to extract web news title in template independent way. In International Conference on Rough Sets and Knowledge Technology (pp. 192-199). Springer Berlin Heidelberg. Wang, J., Li, G., & Feng, J. (2014). Extending string similarity join to tolerant fuzzy token matching. ACM Transactions on Database Systems (TODS), 39(1), 7. Xue, Y., Hu, Y., Xin, G., Song, R., Shi, S., Cao, Y. & Li, H. (2007). Web page title extraction and its application. Information Processing & Management, 43(5), 1332-1347. ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Appendix A. Part-of-speech tags Index Tag Description Index Tag Description 1 CC Coordinating conjunction 11 NNP Proper noun, singular 2 CD Cardinal number 12 NNPS Proper noun, plural 3 DT Determiner 13 POS Possessive ending 4 FW Foreign word 14 PRP Personal pronoun 5 IN Preposition or subordinating conjunction 15 RB Adverb 6 JJ Adjective 16 RP Particle 7 JJR Adjective, comparative 17 VB Verb, base form 8 JJS Adjective, superlative 18 VBG Verb, gerund or present participle 9 NN Noun, singular or mass 19 VBP Verb, non-3rd person singular present 10 NNS Noun, plural 20 VBZ Verb, 3rd person singular present 21 Others Appendix B. POS pattern trees In this appendix, we show a few examples on how to generate the patterns from the trees. In all trees, each blue node is a possible end of a pattern. For example, in proper noun tree, the root <NNP> is an individual pattern; <NNPS> is another individual pattern. From the root node, we can follow the arrows to generate longer patterns. For example, if we start from the left most side of the tree, we can generate <NNP>< NN>, <NNP><NN><NN> and so forth. However, a pattern from the root NNP cannot end at the node JJ because it is black. For example we do not generate a pattern <NNP>< JJ> but we generate <NNP><JJ><NNP>. Proper noun tree Noun trees ACCEPTED MANUSCRIPT A CC EP TE D M A N U SC RI PT Adjective trees Determiner tree Other POS trees 
work_zzxjent5cjgczaezm5l5jht4ny	----	Interaction Pattern Categories Pragmatic Engineering of Knowledge-Based Systems Martina Freiberg, Joachim Baumeister, and Frank Puppe University of Würzburg, Institute of Computer Science Dept. of Artificial Intelligence and Applied Informatics D-97074 Würzburg, Germany freiberg/baumeister/puppe@informatik.uni-wuerzburg.de Abstract. The application of knowledge-based consultation- and doc- umentation systems is, apart from large industrial projects, often also beneficial for small to mid-sized enterprises. Yet, their design and im- plementation still is a tedious and costly task. We motivate, that cus- tomized UI and interaction patterns constitute a pragmatic technique for supporting especially requirements engineering, and thus are capable of considerably promoting real-world projects. In this paper, we intro- duce abstract categories—Guided-, Adaptive-, and Autonomous Entry— for classifying tailored patterns for knowledge-based systems. Further, we discuss their role in an overall approach extending the Agile Process Model and resulting benefits. Keywords: Dialog System, User Interface Design, Agile Development 1 Introduction Knowledge-based systems gained increasing impact also outside academia over the last decades. Apart from large clinical and industrial projects, the applica- tion of knowledge-based consultation and documentation systems is also often beneficial for small to mid-sized corporations. Yet, the trade-off between their potential benefits and their mostly still tedious and costly development, is still often perceived as unfavorable, and respective projects are declined. In general software engineering, user interface (UI) prototyping already is an established methodology regarding iterative, rather inexpensive system specifi- cation before the final product is implemented [3]. Also, UI prototyping permits the early evaluation of (several) design options. Inspired by that approach, we suggest Interaction Patterns for Knowledge-Based Systems as a cornerstone of a tailored, agile knowledge system development methodology. The overall ap- proach integrates pattern- and prototyping-based development into an existing, agile process model, and thus combines the advantages of reusing approved so- lutions (patterns) and of affordable, iterative system specification (UI proto- typing within an agile process model). We argue, that this constitutes a rather pragmatic way to enhance understanding, discussing, and specifying system re- quirements at project start. This in turn helps to promote respective projects in the first place, and thus renders its application especially interesting when addressing small to mid-sized enterprises as customers. As a first step into this direction, this paper introduces Interaction Pattern Categories, that provide an abstract classification framework for knowledge sys- tems and corresponding interaction/UI patterns. We further discuss the role of such patterns within the proposed, extended agile approach; the details regard- ing the prototyping and a respective tool are subject of separate work, see [7]. The rest of the paper is organized as follows: Related research is presented in Section 2. In Section 3, we introduce our classification framework of Interac- tion Pattern Categories and the relevant terminology. We further present three categories, identified on the basis of past experiences with conducted projects. In Section 4, we outline the extended, agile process model, and the patterns’ specific role as well as resulting benefits. We conclude with a summary of the presented work and a discussion of future research directions in Section 5. 2 Related Work The process model for knowledge system development, that we suggest in this paper, integrates pattern- and prototyping-based development, thus uniting the advantages of both approaches and especially fostering an enhanced requirements engineering. Patterns specify proven solutions for recurring (design) problems and are established in many domains: Examples are software engineering, ontology engi- neering, or knowledge formalization, [8, 9, 14]. They offer the advantage to reuse approved solutions for similar problems, and thus to reduce development ef- forts and to profit from the lessons learned. Regarding UI–/interaction design, tailored, domain-specific pattern collections exist, e.g., [16–18]. Yet, patterns originating from such research cannot be straightly transferred to our context, as knowledge-based systems put specific demands on interaction and UI design. Regarding knowledge system development, various methodologies have been proposed in the past—see [12] for an overview. More recent works emphasize the relevance of agility, e.g., see [2, 10]. We follow that direction by integrating pattern- and prototyping-based techniques into an agile process model. Previous approaches, however, often strongly emphasize the development of the knowl- edge itself. In contrast, we specifically support knowledge system UI and inter- action design by the means of tailored patterns and prototyping as to enhance requirements engineering on the one hand, and to foster a pragmatic, affordable promotion and execution of respective projects on the other hand. UI prototyping so far has become an established approach in general software engineering [3] as well as in HCI and usability engineering [4]. Main advantages are a strong support of requirements specification, and the opportunity to eval- uate (several) UI design(s) at an early stage. In [11], a prototyping tool, that incorporates design patterns for layout support, has been proposed. Though generally related to our approach, that work focusses on cross-device design of general web-style interaction. Contrastingly, we explicitly consider UI and inter- action design of knowledge-based consultation and documentation systems. 3 Interaction Pattern Categories By Interaction Patterns for Knowledge-Based Systems, we understand the de- scription of the systems’ interaction- and UI design for a specified context. They comprise a compact specification of their applicability, and exemplify the cor- responding solution approach. Yet, there exist attributes, the value of which may be common to more than one distinct pattern. Thus, we first introduce an abstract framework—Interaction Pattern Categories—for classifying patterns according to such common properties before specifying concrete patterns. In Sec- tion 3.1, we first introduce relevant terminology, and in Section 3.2, we present the classifying categories—Guided-, Adaptive-, and Autonomous Entry. 3.1 Relevant Terminology for Specifying Pattern Categories In the following, we specify the addressed system types as well as the classifying attributes, that characterize the pattern categories, in more detail. Knowledge-Based Systems: We specifically address knowledge-based sys- tems with our approach—by that, we understand knowledge systems, that serve either a consultation or a documentation task. In both cases, the main user- system interaction is structured data entry—mirrored by ”Entry” in the pattern category names. Regarding consultation, the system gradually derives solutions for a given problem with the respective, implemented reasoning mechanisms based on the provided user input (answers). Documentation systems emphasize supporting uniform and reliable data input as effectively as possible. Classifying Attributes: The attributes User Competence, Context Presen- tation, and Data Volume are common to all patterns of one category. Major classifier thereby is User Competence—in the context of knowledge-based sys- tems, lengthy, strictly prescribed interviews can be annoying and inflexible for competent users, that might want to influence the interrogation flow according to their expertise. This makes it essential to tailor the system and interface design to the target users’ competence. A. User Competence: A naive data provider follows the prescribed inter- rogation sequence, with no desire for deviation or adaption; possible reasons can vary from rather low domain competence/lacking experience to a highly stressful usage context (but nevertheless domain expertise). Experienced users possess a certain domain expertise, and thus may be interested in system-suggested work- flow guidance, yet, additionally require the option to influence the interrogation and to deviate from suggested paths. An autonomous problem solver finally pos- sesses sufficient expertise to solve the problem independent from system guid- ance, based on the (potentially various and complex) information presented. B. Data Volume: The amount of data that is processed during a typical interrogation session; thus, it corresponds to the number and the complexity of questions required for deriving a solution or entering a complete data set. We roughly distinguish between small, medium, and large. Data Space, in con- trast, specifies the universal range of possible input data, and thus corresponds to the domain complexity. The respective data volume/data space combination does not influence the pattern categorization, yet the knowledge required for a specific implementation—e.g., large data space and low data volume implies so- phisticated interrogation structures to present appropriate questions efficiently. C. Context Presentation: No context means, that during an interrogation, only the required questions are presented, but no further information. Otherwise, we distinguish support knowledge—auxiliary information (not interrogation spe- cific), or informal knowledge representations—and interrogation context—i.e., additional information regarding, e.g., the progression of the workflow, or in- dicating the consequences of choosing certain answer alternatives in advance. Type and extent of context presentation highly depend on the respective level of user competence—concerning naive data providers, interrogation context often is not required, yet for complex questions, support knowledge presentation might be advisable; with rising competence, the presentation of interrogation context gains importance for supporting an independent, efficient system usage. 3.2 Interaction Pattern Categories for Knowledge-Based Systems Based on experiences from past projects, we define three basic categories for UI/interaction patterns: Guided-, Adaptive-, and Autonomous Entry. For each category, we describe the Problem Statement, the Solution, and the Applicability, specifying common properties that apply to all contained patterns. Further, we provide Examples—i.e., existing implementations—and Variants, that describe in what regard patterns of a category may vary. The basic interaction specifying each pattern, regards question selection dur- ing interrogation. Even if patterns later vary, e.g., regarding the processed data volume, that basic interaction remains the same. For its specification we use the UML sequence diagram notation and the elements: User, the system Interface, Questions (presented to and answered by the user), the Data pool (storing data resulting from provided answers or reasoning), and the Knowledge component. A. Guided Entry Problem Users act as naive data providers, thus for a reliable, effective decision- or documentation support, a high level of system autonomy/guidance regarding the interrogation flow is required. Data volume might vary from small to medium. Solution An interview metaphor is transferred to the interface, where the user and the system interact alternately. The system flexibly reacts to the pro- vided answers by adapting the interrogation sequence, thus presenting only the question(s) that fit the respective context best. The interview proceeds system-guided, and deviation is mostly not (or only in limited terms) intended. Thus, presenting interrogation context is not mandatory, even though regard- ing lengthy sequences status feedback may be beneficial. Contrastingly, support knowledge is required in the case of complex/difficult questions for clarification. Figure 1 depicts the interaction sequence for Guided Entry. The interface initi- loop Question KnowledgeInterface present Question() answer Question() requestNext() Data assessNext() propagate() propagate AnswerValue() assess() Fig. 1. Guided Entry—Basic interaction sequence for question selection. ates the question request, whereupon the knowledge component assesses the next question—where available, based on the previously provided user input stored in the data pool—and propagates the result back to the interface. The then pro- vided answer of the user is propagated to the data component and thus made available for the knowledge component hereafter. Those steps are performed it- eratively until a defined interrogation sequence is finished. Applicability Systems based on Guided Entry equally fit consultation and documentation tasks. Especially documentation of high quality or frequently recurring data is supported, as specified data entry can be assured by the strict, system-guided interrogation flow. However, if a higher level of user au- tonomy is desired—e.g., influence on the interrogation, or adaptable question representation—Adaptive- or Autonomous Entry provide more flexibility. Examples Figure 2 presents two implementation variants of Guided Entry. CareMate (A) is a quick response second-opinion system for emergency situ- ations. Its one-question interaction style creates the literal impression of an in- terview and supports the intuitive usage in the context of stressful emergency conditions. Continuous status feedback on the current solution states is provided, and the processed data volume is rather small. For a more detailed introduction, see [1]. SonoConsult [15] is a consultation and documentation system for the field of abdominal ultrasound. The multiple-question interaction style resembles a paper-based interview (questionary) and helps to cater with the rather large data volume. Both support knowledge (question clarification) and interrogation context (presenting currently derived solutions) are provided. Variants Pattern variants arise with regards to the type and extent of context B A Fig. 2. Implementation variants of Guided Entry, both in german: A) Digitalys Care- Mate and B) SonoConsult. presentation (see above examples), as well as regarding the characteristics of the naive user (e.g., expert in stressful context vs. non-expert’s ad-hoc usage). B. Adaptive Entry Problem Experienced target users have a certain—yet, from user to user po- tentially varying—domain competence; consequently, both system guidance as well as the option for autonomous decisions regarding the workflow are desired. Also, questions should be presented in a user-adaptable manner. Solution The system basically suggests the most appropriate workflow to the user; yet, also the option to deviate from that path and choose an adapted inter- rogation sequence is provided. Where applicable, a hierarchical tree metaphor is applied to cater with varying user competence levels: Questions are defined both on an abstract (aggregate) level, but also subdivided into (several) refined questions, where reasonable. Thus, according to their expertise, users may either answer the aggregate questions—taking less time, but requiring more expertise— or request the presentation of the questions’ refinement. To support the user’s decision-making, providing interrogation context is strongly recommended. Also, depending on the refinement level and complexity of the questions, support knowledge should be additionally presented. Figure 3 sketches question selection in Adaptive Entry. Basically, the user decides whether to follow the system-guided interrogation or whether to choose an own path. Regarding the first alternative, question selection proceeds as in Guided Entry (Figure 1). In the second case, either the user’s competence allows for an- loop alt Question Data KnowledgeInterface present() propagate() [ELSE] [USER COMPETENT] request Refinement() assessRefinement() propagate() answerSelected Question() propagate AnswerValue() assess() alt [Guided Entry: System-guided Question Selection] Fig. 3. Adaptive Entry—Basic interaction sequence for question selection. swering the currently displayed question; then the answer is propagated to the data pool and thus is available for the knowledge component as the interroga- tion continues. Otherwise, the user can request question refinement whereupon the knowledge component assesses the possible refinement, and propagates the result back to the interface for displaying it to the user. Applicability Apart from consultation, respective systems can, with limita- tions, also serve documentation purposes. In that case, special care has to be taken that all required input data is obtained from the user. Regarding effective interrogation of naive data providers Adaptive Entry is only marginally suitable. Examples Figure 4 shows the Labour Legislation Consultation, that clarifies, whether a dismissal in a given context is legitimate. Figure 4, A, represents the problem statement. Its current derivation state and the questions’ state (e.g., answered) are visually indicated by background coloring and updated with each provided answer. Questions can be processed either in the sequence suggested (i.e., from top to bottom), or in any other order. Further, adaptable question presentation is implemented—e.g., Figure 4, B, was confirmed on the abstract level; question Dismissal was... is expanded into refined questions (Figure 4, C). Variants Possible variants originate from different forms of context presenta- tion as well as from different data volumes that may be processed. C. Autonomous Entry Problem Target users are highly competent, autonomous problem solvers, thus no explicit guidance regarding the interrogation sequence is required. Solution The user explores the (various and potentially complex) information sources presented by the system. Integrated knowledge-based components—e.g., consultation features or automated data entry support—can be used optionally, but are not mandatory to benefit from system usage. The user provides any in- A B C Fig. 4. Labour Legislation Consultation—Indication of the solution and its current rating (A), clarifying questions (B), and further refinement of one question (C). put data voluntarily; based on those data fragments, the system performs rather modularized reasoning, following sort of a bits and pieces metaphor. Thus, the system merely provides a second-opinion to the user in presenting reasoning re- sults (e.g., rated solutions, next-input suggestions). Such extensive user control requires a high level of context presentation, regarding both types of context. loop alt alt Question Data KnowledgeInterface answerSelected Question() propagate AnswerValue() assess() [Use support features (consultation/data enty) ] [Exploration] explore() Fig. 5. Autonomous Entry—Basic interaction sequence for question selection. As Figure 5 shows, the user always can choose between using more formal knowledge components and free exploration. In the first case, potentially any kind of (complex) knowledge component can be integrated into such a system, e.g., according to the Guided Entry or Adaptive Entry categories. Otherwise, the user can either simply explore the provided information, or answer questions autonomously in a modularized manner. Answers then are propagated to the data component, and from there assessed by the knowledge component; the lat- ter rates solutions, presents context, and recommends next steps piecemeal. Applicability Autonomous Entry can be applied for loose consultation, as well as regarding a more informal, potentially collaborative documentation task. In contrast, it is inappropriate, if rather naive data entry is desired, as no strict workflow guidance for solving the addressed problem is provided. Further it is not suitable for high quality documentation tasks, as any interaction takes place voluntarily, and thus the supply of any data cannot be guaranteed. Examples Implementation examples are the user-centered consultation ap- proach described in [5], the PEN-Ivory system [13], but also the Inline Answering concept provided by the Semantic Wiki KnowWE [1]. Variants Implementation variants arise with respect to the data volume, and the type and extent of context presentation. Despite mainly addressing expert users, systems falling into this category, might to some extent also be suitable for unexperienced ad-hoc usage. Finally, resulting systems can vary regarding the extent of integrating knowledge-based features. The proposed pattern categories classify basic knowledge system types and corresponding UI/interaction design patterns according to the level of user com- petence (corresponding to the level of system guidance). Ongoing research ad- dresses the definition of concrete patterns and their categorization accordingly. We proceed by discussing how such patterns can be integrated in an extended, agile process model for developing knowledge-based consultation and documen- tation systems. 4 Pragmatic Knowledge System Engineering Regarding knowledge system development and knowledge engineering, there ex- ist diverse approaches today, such as CommonKADS, MIKE, or adaptions of the classical stage-based and incremental software development models. Yet, for the success of knowledge system projects in the context of small to mid-sized companies, a pragmatic approach—affordable and efficient regarding time and effort—is essential, c.f. [2]. Especially for promoting such projects in the first place, it is important to quickly come up with first solutions, e.g., in the form of prototypes or example implementations. In this respect, we made positive experiences with applying the Agile Process Model, described in [2]. However, that model emphasizes knowledge base development, not yet taking much into account the design of the target system’s interface, or usability traits. In extend- ing this model, not only UI/interaction design gains importance in the overall development process, but also the integration of usability activities. In the following, we introduce the Extended Agile Process Model, and af- terwards we discuss resulting benefits specifically regarding the integration of tailored UI/interaction patterns. Although prototyping and usability-related ac- tivities are included in the model for reasons of completeness, their detailed discussion is part of further work, see [7]. 4.1 The Extended, Agile Knowledge System Engineering Model Figure 6 outlines the Extended Agile Process Model —the gray parts represent the original model, consisting of the four phases System Metaphor, Planning Game, Implementation, and Integration. For a detailed discussion, see [2]. UA1: Prototype Expert / Hybrid UA2: Prod. System, User-based / Hybrid Integration System Metaphor Planning Game Implementation Tailored Patterns & Prototypes Fig. 6. Extended Agile Process Model. Basically, tailored patterns and prototyping can support both System Metaphor and Planning Game. In System Metaphor, the system objectives are defined by developers and customers. Based on appropriate patterns and cor- responding implementation exam- ples, a basic idea can be devel- oped more easily. Thereby, pat- terns can be assessed either manu- ally, or by using a tailored recom- mender system, that suggests pat- terns matching the target context. Prototypes, that also provide the relevant user-system interactions, further sup- port that process by presenting a realistic simulation of a potentially resulting system as opposed to the static, visual depiction of knowledge system examples provided by the patterns. The Planning Game defines the scope and priori- tization of development tasks. Here, patterns ease the analysis and valuation of system requirements—taking place during the Exploration sub-phase of the planning game—by providing clear specifications of required features and inter- actions. Additionally, prototyping supports that task by allowing for actually trying out (and thus better evaluating) relevant functionalities. With regards to Usability Activities, the original model can be extended both regarding Implementation and Integration (Figure 6, UA1, UA2). The ba- sic model defines Implementation as a test-first activity—i.e., before actually implementing new or additional features, the corresponding tests for assuring their correctness are developed. This can be expanded by an evaluation-first ac- tivity, in the sense that based on the formerly created prototypes, usability issues are assessed and valued first, before continuing with test-first implementation as defined by the model. Without going into detail here, at that stage, expert- or hybrid approaches (according to a categorization suggested in [6]) seem to be most appropriate. During Integration, the implemented functionality is added to the productive system, using integration tests for assuring its overall correctness and integrity. Such testing can be extended by usability checks that evaluate, whether the system still meets the specified usability goals. As this results in a running version of the productive system, not only hybrid, but also purely user-based usability evaluation can be beneficial. 4.2 Benefits The integration of tailored patterns into an extended agile model offers several benefits: First, the patterns uniformly specify common framework conditions of different knowledge-based system types; thus, they provide a descriptive and visual language, that enables customers and developers to discuss at the same competence level. This fosters a clear communication and thus reduces potential misconceptions right away, that otherwise can lead to additional, unnecessary redesign cycles. Next, the patterns present actual implementation examples, that can be assessed, and serve as a inspirational source regarding the concrete project at hand. Even in case none of the provided patterns or examples completely satisfy the project- and customer requirements, those nonetheless are helpful by providing an overview of the possibilities and a basis for further discussions. Despite diverse general pattern collections and UI prototyping tools, to date to the best of our knowledge no tailored patterns/tools exist addressing specifically knowledge-based systems. As the latter exhibit quite specific characteristics, our approach can provide strong support for their development. 5 Conclusions and Future Work We motivated that tailored patterns can constitute the cornerstone of an ex- tended, agile model for knowledge system development. Especially when target- ing smaller to mid-sized companies as customers, the suggested approach is a rather inexpensive, pragmatic technique for promoting and launching respective projects. As a first step, in this paper we introduced three abstract categories for classifying corresponding patterns. Due to our focus on knowledge-based consul- tation and documentation systems, those categories specifically address the data entry task; yet, the elementary classification—guided, adaptive, and autonomous interaction might be applied accordingly for other forms of interaction. The cat- egorization arose from practical experiences with implementing knowledge-based systems in the past, such as SonoConsult [15], Digitalys CareMate, or more re- cently the Semantic Wiki KnowWE [1]. Further research includes the question, whether additional pattern categories are required. Based on those, as well as on an assessment of further existing systems, concrete patterns will be specified. Currently, also a tailored prototyping tool is developed [7], that will be further extended based on an analysis of required knowledge system base components. References 1. Baumeister, J., Reutelshoefer, J., Puppe, F.: KnowWE: A Semantic Wiki for Knowledge Engineering. Applied Intelligence (2010) 2. Baumeister, J., Seipel, D., Puppe, F.: Agile development of rule systems. In: Giurca, Gasevic, Taveter (eds.) Handbook of Research on Emerging Rule-Based Languages and Technologies: Open Solutions and Approaches. IGI Publishing (2009) 3. Bäumer, Dirk and Bischofberger, Walter R. and Lichter, Horst and Züllighoven, Heinz: User Interface Prototyping—Concepts, Tools, and Experience. In: ICSE ’96 Proceedings of the 18th International Conference on Software Engineering. pp. 532–541 (1996) 4. Beaudouin-Lafon, M., Mackay, W.: Prototyping tools and techniques. In: The Human-Computer Interaction Handbook: Fundamentals, Evolving Technologies and Emerging Applications. pp. 1006–1031. L. Erlbaum Associates Inc., Hillsdale, NJ, USA (2003) 5. Buscher, G., Baumeister, J., Puppe, F., Seipel, D.: User-Centered Consultation by a Society of Agents. In: Proc. of the 3rd International Conference on Knowledge Capture (K-CAP 2005), Banff, Alberta, Canada (2005) 6. Freiberg, M., Baumeister, J.: A survey on usability evaluation techniques and an analysis of their actual application. Tech. Rep. 450, Institute of Computer Science, University of Würzburg, Germany (2008) 7. Freiberg, M., Mitlmeier, J., Baumeister, J., Puppe, F.: Knowledge system proto- typing for usability engineering. In: Proceedings of the LWA-2010 (Special Track on Knowledge Management) (2010) 8. Gamma, E., Helm, R., Johnson, R., Vlissides, J.: Design Patterns. Elements of Reusable Object-Oriented Software. Addison-Wesley Longman (1995) 9. Gangemi, A., Presutti, V.: Ontology Design Patterns. In: Staab, S., Studer, R. (eds.) Handbook on Ontologies (2009) 10. Knublauch, H.: Extreme programming of knowledge-based systems. In: Pro- ceedings of eXtreme Programming and Agile Processes in Software Engineering (XP2002) (2002) 11. Lin, J., Landay, J.A.: Employing patterns and layers for early-stage design and prototyping of cross-device user interfaces. In: CHI ’08: Proceeding of the twenty- sixth annual SIGCHI conference on Human factors in computing systems. pp. 1313–1322 (2008) 12. Plant, R., Gamble, R.: Methodologies for the development of knowledge-based systems, 1982–2002. Knowledge Engineering Review 18(1), 47–81 (2003) 13. Poon, A.D., Fagan, L.M., Shortliffe, E.H.: The PEN-Ivory Project: Exploring User- Interface Design for the Selection of Items from Large Controlled Vocabularies of Medicine. Journal of the American Medical Informatics Association pp. 168–183 (1996) 14. Puppe, F.: Knowledge Formalization Patterns. In: Proceedings of PKAW 2000, Sydney Australia (2000) 15. Puppe, F., Atzmueller, M., Buscher, G., Huettig, M., Luehrs, H., Buscher, H.P.: Application and evaluation of a medical knowledge system in sonography (sono- consult). In: Proceeding of the 2008 conference on ECAI 2008. pp. 683–687. IOS Press, Amsterdam, The Netherlands, The Netherlands (2008) 16. Ratzka, A.: Identifying user interface patterns from pertinent multimodal interac- tion use cases. In: Mensch & Computer 2008: Viel Mehr Interaktion. pp. 347–356 (2008) 17. Schmettow, M.: User interaction design patterns for information retrieval systems. In: EuroPLoP’ 2006, Eleventh European Conference on Pattern Languages of Pro- grams. pp. 489–512 (2006) 18. Tidwell, J.: Designing Interfaces — Patterns for Effective Interaction Design. O’Reilly Media Inc. (2006) 
work_zzh2hcwt2jgj3fodyu3vqh6zje	----	Evolutionary-based heuristic generators for checkers and give-away checkers Jacek Mańdziuk∗, Magdalena Kusiak and Karol Walȩdzik Faculty of Mathematics and Information Science, Warsaw University of Technology, Plac Politechniki 1, 00-661 Warsaw, Poland Abstract Two methods of genetic evolution of linear and non-linear heuristic evaluation functions for the game of checkers and give-away checkers are presented in the paper. The first method is based on the simplistic assumption that a relation ‘close’ to partial order can be defined over the set of the evaluation functions. Hence explicit fitness function is not necessary in this case and direct comparison between heuristics (a tournament) can be used instead. In the other approach heuristic is developed step-by- step based on the set of training games. First, the end-game positions are considered and then the method gradually moves “backwards” in the game tree up to the starting position and at each step the best fitted specimen from the previous step (previous game tree depth) is used as the heuristic evaluation function in the alpha-beta search for the current step. Experimental results confirm that both approaches lead to quite strong heuristics and give hope that more sophisticated and more problem-oriented evolutionary pro- cess might ultimately provide heuristics of quality comparable to those of commercial programs. Keywords: game playing, heuristic generators, checkers, give-away checkers, evaluation function, evolutionary computation. Preliminary version of the paper. The revised version was published in EXPERT SYSTEMS, 24(4), 189-211, 2007, Blackwell Publishing. The final version is available on request. ∗Corresponding author. E-mail:mandziuk@mini.pw.edu.pl. This work was supported by the Warsaw University of Technology grant no. 504G 1120 0008 000 1 1 Introduction Games, from a purely scientific point of view, provide cheap, reproducible environments suitable for testing new search algorithms, pattern-based evaluation methods or learning concepts. Since the seminal papers devoted to programming chess [39, 43, 24] and checkers [31] in the 1950s., games remained through decades an interesting topic for both classical AI and Computational Intelligence (CI) based approaches. In 1965 chess was even announced as ”the Drosophila of Artificial Intelligence” by the Russian mathematician Alexander Kronrod. Although most examples of application of CI methods to game playing make use of ei- ther reinforcement learning method or neural networks, there are also some papers devoted to evolutionary or hybrid neuro-genetic solutions, e.g. in chess [19, 14], in checkers [8, 2], in Othello [23], in Rummy [20], in tic-tac-toe game [8], in Iterated Prisoner’s Dilemma [11, 8, 38], in Backgammon [30], in Poker [5, 6], and others. The main idea of this paper is to test the efficacy of “pure” evolutionary approach to the problem of building evaluation functions in US checkers and US give-away checkers (GAC)1. The game of checkers is widely known and does not need any introduction. For those who are unfamiliar with this game and its rules we would recommend consulting one of the internet checkers playing sites. The game of GAC is played according to the same rules as checkers. Only the ultimate goals of both games differ. In order to win a game of GAC player has to lose all his/her pieces or be unable to perform a valid move [3]. Despite simplicity of the rules GAC is not a trivial game and results of brute-force algorithm using trivial strategy of losing pieces as quickly as possible are unsatisfactory. One of the reasons for unsuitability of this simple algorithm is the fact that a single piece is barely mobile and has very restricted choice of possible moves. Fig. 1 presents two situations in which white loses despite having only one piece left. (a) Black to play and win. (b) White to play and lose. Figure 1: Examples of inefficiency of greedy heuristics. White move bottom-up. Open circles denote kings. 1Since within the whole paper the US version of the game of checkers and the game of give-away checkers is considered henceforth we’ll simply write checkers or give-away checkers, skipping US. 2 In the left figure the winning sequence for Black is the following (see Fig. 2 for board description): 1. ... 23-27 2. 24-31 (27) 22-26 3. 31-22 (26) 15-18 4. 22-6 (18,10) 0:1 In the right figure White has two choices of the first move, but in either case loses: 1. 22-17 10-14 1. 22-18 9-14 2. 17-1 (14,6,K) 9-6 2. 18-2 (14,6,K) 3-7 3. 1-10 (6) 3-7 3. 2-11 (7) 10-15 4. 10-3 (7) 0:1 4. 11-18 (15) 0:1 Certainly, the game of GAC is not as popular as checkers, but still there has been some interest in this game too [3, 27, 21, 37]. An example of a GAC playing program, which became the main opponent for heuristics developed in this paper, is described in more detail in section 7.1. As opposed to GAC, there have been quite a lot of papers concerning checkers (or draughts) published in the literature (e.g. [31, 17, 1, 26, 40, 36, 2, 13]), but since they are not directly related to our work we’ll not discuss them any further, except for a brief mention of the three major achievements in this field. First of all, the history of checkers playing programs begins with the seminal papers of Arthur L. Samuel [31, 32]. Samuel’s program was able to develop the evaluation function on its own. More precisely it was equipped with an a priori defined set of expert features potentially relevant for building board evaluation function and during the self-play phase was able to successfully define the essential subset of features and their weights in order to form polynomial evaluation function. Having spent enough time in the self-play stage the program achieved an expert level of play. Another ‘grand AI achievement’ in the game of US checkers is Chinook program de- veloped by Jonathan Schaeffer [34, 35, 36] which was able to successfully compete with the human absolute checkers genius - Marion Tinsley. Chinook uses a very well tuned hand-crafted evaluation function and is equipped with an end-game database covering all possible board positions of 8 or fewer pieces. Also, Chinook’s search mechanisms are very sophisticated. Unlike Samuel’s program, Chinook does not include any learning mecha- nism. The third undisputable achievement in the field of checkers is Anaconda (also known as Blondie24) developed by Kumar Chellapilla and David Fogel [9, 10, 13]. Carefully designed evolutionary process over the ensemble of neural networks allowed Blondie24 to achieve the rating of an expert player on the internet checkers playing site without any built-in human expert knowledge. It is worth to note that in order to generate A-class level playing network (just below the expert level) only about 250 evolutions was required. After the 3 next 600 evolutions an expert level of play was achieved. Anaconda (Blondie24) is an apparent, successful example of learning from scratch using Computational Intelligence techniques. Unlike in the above mentioned ‘grand AI achievements’ the piece of research described in this paper is not intended to create a program capable of competing against profes- sional GAC programs (in fact the authors are unaware of any such software) or commercial checkers applications, but to test usefulness of “generic” genetic approach for generating evaluation functions in these two game domains. The underlying assumption is the sim- plicity and generality of proposed solutions. Consequently, fitness function is defined in a problem independent manner and straightforward genetic operators are applied. More- over, a plain alpha-beta algorithm, with no particular optimization for speed, is used and a shallow search (4 − 6 plies) is implemented. No opening books or end-game databases are used. There are a few reasons for considering these two particular games (checkers and give- away checkers). First of all, checkers is a very popular and widely known game. Next, the rules of both games are simple and at the same time none of the games has been solved and both require nontrivial heuristics. Finally, the possibility of checking the effectiveness of the proposed solutions to the problems defined over the same set of constraints (rules of the game) but with completely “opposite” goal functions (goals of the game) seems to be tempting and may possibly allow for some general (game independent) observations. Hence, in order to prove the efficacy of evolutionary methods in game playing domain we have applied them to both games and in each case to both linear and non-linear evaluation functions. Evolved evaluation functions were tested against one another and also against some commercial and non-commercial software available on the internet or obtained via private communication. For each type of heuristics (linear vs non-linear) two heuristic generators were used and compared. The first one relies on the simplistic assumption that the “winning relation” defined over the set of all heuristics is transitive. Therefore instead of using the extended fitness function a direct comparison between evolved heuristics is used. In the other approach heuristic is developed gradually starting from the end-game positions. These positions are either leaves or nodes close to the leaves in the game tree and therefore can be estimated based on the game outcomes. This makes it possible to build fitness function for that phase of the game, and the best fitted specimen in that phase is used to define the fitness function for positions which are a few moves closer to the starting position (i.e. a few moves up in the game tree). Following this scheme the algorithm reaches the beginning of the game. In order to provide justified and convincing conclusions we have decided to include in this paper several low-level implementation details. Some technical issues are thoroughly discussed especially when particular parameters seem to have an important impact on the quality of evolutionary process. Even though it is hard to expect that “blind” evolutionary process together with a shallow game tree search may lead to construction of a grandmaster-level heuristic, and despite undisputable achievements of hard AI approaches in the domain of checkers, e.g. [36], the results presented in this paper strongly support the claim that due to their biological soundness and self-improvement capabilities evolutionary methods deserve strong consideration in game playing domain. 4 It is also reasonable to say that there is still some room for quality improvement of the evolved heuristics in case more sophisticated and more problem-oriented evolutionary process is applied. The paper is organized as follows. In the next section the components of the evaluation functions are described. In section 3 two heuristic generators used to genetically develop the evaluation functions are presented and discussed. These are Heuristic Generator (HG) presented in sect. 3.1 and Simple Heuristic Generator (SHG) described in sect. 3.2. Next, in section 4 linear (sect. 4.1) and nonlinear (sect. 4.2) heuristics are described in detail with particular emphasis put on the way they are composed of simple components listed in sect. 2. Several low-level implementation details concerning evolutionary processes implemented in both heuristic generators as well as initial assessment of their performance are outlined in sections 5 and 6, resp. Numerical results, along with their in-depth discussion are presented in section 7, respectively for give-away checkers (in sect. 7.1) and checkers (in sect. 7.2). The last section is devoted to conclusions and directions for future development of this research. 2 Components of the Evaluation Functions Heuristics described in this paper are relatively simple and make use of some or all of the following parameters, calculated separately for each player: 1. Number of pawns; 2. Number of kings; 3. Number of safe pawns (i.e. adjacent to the edge of the board); 4. Number of safe kings; 5. Number of moveable pawns (i.e. able to perform a move other than capturing). 6. Number of moveable kings. Parameters 5 and 6 are calculated taking no notice of capturing priority; 7. Aggregated distance of the pawns to promotion line; 8. Number of unoccupied fields on promotion line. Heuristics could also consider sums of or differences in respective parameters for both players rather than raw numbers for each player separately. Once heuristics using the straightforward parameters listed above had been generated and tested, it was decided that it would be desirable to add more sophisticated parameters characterizing layout of the pieces on the board. The following parameters were, therefore, introduced: 9. Number of defender pieces2 (i.e. the ones situated in two lowermost rows); 10. Number of attacking pawns (i.e. positioned in three topmost rows); 11. Number of centrally positioned pawns (i.e. situated on the eight central squares of the board); 12. Number of centrally positioned kings; 13. Number of pawns positioned on the main diagonal; 14. Number of kings positioned on the main diagonal; 15. Number of pawns situated on double diagonal; 16. Number of kings situated on double diagonal; 17. Number of loner pawns. Loner piece is defined as the one not adjacent to any other 2Please note, that within the whole paper the term “pieces” will denote “pawns and king together”. 5 piece; 18. Number of loner kings; 19. Number of holes, i.e. empty squares adjacent to at least three pieces of the same color. Apart from the above parameters six patterns were defined. They are described below using common notation presented in Fig. 2. Since only one instance of each pattern can exist for any of the players at the same time, features 20−25 can take only boolean values. 20. Presence of a Triangle pattern(see Fig. 3(a)). 21. Presence of an Oreo pattern (see Fig. 3(b)). 22. Presence of a Bridge pattern (see Fig. 3(c)). 23. Presence of a Dog pattern (see Fig. 3(d)). 24. Presence of a pawn in corner (i.e. White (Black) pawn on square 29 (4), resp.); 25. Presence of a king in corner (i.e. White (Black) king on square 4 (29), resp.); Figure 2: Notation used for describing patterns. White pawns are at the bottom. Generally two types of heuristics were considered. Linear heuristics consisted of linear combination of the parameters listed above: LinearCombinationOf P arameters = a1 · param1 + a2 · param2 + . . . + aj · paramj , (1) where 1 ≤ j ≤ 18 and param1, . . . , paramj were freely chosen from the above 18 parameters (being either raw numbers or sums or differences calculated for both players). Nonlinear heuristics were composed of a small number (in our tests 3 or 5) of nonlinear components of the following form: IF[NOT] (min1 ≤ par1 ≤ max1 AND/OR min2 ≤ par2 ≤ max2 (2) AND/OR . . . mini ≤ pari ≤ maxi) THEN Heuristic Result+ = LinearCombinationOf P arameters where again 1 ≤ i, j ≤ 18 and par1, . . . , pari and param1, . . . , paramj were defined over the set of the above mentioned parameters and LinearCombinationOf P arameters was defined as in (1). As it was the case with linear heuristics sums or differences of particular parameters calculated for both players could be used instead of raw numbers. 6 The final sets of parameters and the respective limits mink, maxk; 1 ≤ k ≤ i were chosen manually basing on preliminary tests. In order to restrict the nonlinear condition space two additional constraints were im- posed on (2), i.e. only one type of operator (AND or OR) could be used in a single condition (2) and negation could be applied to the whole condition only. For either type of heuristics coefficients a1, . . . , aj in (1) were optimized in the evolu- tionary process described in the next section. (a) Triangle pattern. (b) Oreo pattern. (c) Bridge pattern. (d) Dog pattern. Figure 3: Four out of six patterns used as heuristics’ components. All patterns are de- scribed from the White player side. 3 Evolutionary Algorithms Genetic algorithms were used to generate of coefficients a1, . . . , aj for parameters param1,. . ., paramj in (1). All variable parameters of a heuristic were represented as a vector of real numbers where each number denoted a single gene. Conditions that nonlinear components consisted of were not modified by the process of evolution. Two different approaches im- plemented are described in sect. 3.1 and sect. 3.2, resp. 7 3.1 Heuristic Generator (HG) One of the difficulties encountered while designing the genetic algorithm was defining fitness function for the heuristics. Unless an efficient evaluation function is given any game position can be assessed correctly only after the whole game is played (i.e. the final result is known) and both sides play perfectly. But, since evolution of such function is actually our goal and the basic assumption that we made in this work is to use zero expert knowledge approach, we obviously can not posses such external ‘efficient evaluation function”. The general idea of how to solve this problem was described in [7]. According to [7] the game was partitioned into several disjoint stages based on the number of moves already performed. In the first phase of the algorithm a heuristic able to assess correctly situations close to the end of the game was obtained. In order to achieve this, alpha-beta algorithm with no heuristic was used to assess a number of randomly generated positions close to the leaves of the full game tree (i.e. close to the end of the game). All positions beyond the depth of alpha-beta algorithm were considered a draw. Subsequently, each specimen assessed the same positions and its fitness was calculated according to the formulae: n∑ (hi − ai)2, (3) where n denotes the number of test situations, hi assessment of the i-th test situation by the heuristic specimen and ai by the alpha-beta algorithm. Additionally, when during the particular run of the algorithm speciation encourage- ment feature was enabled, fitness value of each of the specimens was at this point divided by the sum of sharing function values for all the specimens in the population. Sharing function was defined as exp(−d2), where d is Euclidean distance between the genotypes of two specimens. This was done as a mean to encourage speciation which in turn might lead to improved exploration of the problem space [15, 16]. Impact of sharing function was however ignored when selecting the best specimen in the population. Since the games of checkers and give-away checkers have only three possible outcomes and the values of heuristics were expected to fall into a continuous interval, the depth of a leaf in the game tree was taken into consideration while assessing it. At first, the win value was defined as max win value and loss value as its opposite, where max win value was defined to be at least twice as big as the maximal depth (calculated from the root of the tree) of a game tree node examined by alpha-beta. Assuming that the current position was searched to depth g, the final assessment was equal to max win value−g in case of a win or g − max win value in case of a loss, or zero in case of a draw. Hence, the absolute value of the assessment of any position other than a draw would always belong to the interval (0.5 · max win value, max win value). This method of including the depth of the searched tree in position assessment correlates with the observation that winning in shorter time (or loosing after longer time) should be assessed better than winning after more moves (loosing quickly)3. Once the initial stage had ended, worst fitted fraction (usually between 0.7 and 0.8) of the population was replaced by new random specimens. New board situations closer to the root of the game tree were generated and they were assessed by the alpha-beta 3A similar idea is used in Reinforcement Learning algorithms, where the reward/punishment is scaled by a discount factor exponentially decreasing in time [18, 42]. 8 algorithm with the fittest specimen of the previous phase used as its heuristic function. The process continued until the root of the game tree was reached. In each phase a constant fraction (between 0.35 and 0.4) of all the test boards came from the stage of the game closest to the beginning (i.e. the one considered at the moment). Depending on the settings of the algorithm, the positions closer to the leaves of the game tree were either regenerated and reassessed by the newest heuristic or once generated they were used throughout all subsequent phases. Standard genetic operators: selection, mutation and crossover were implemented in the following way. Selection − selection was done by means of tournaments. For each tournament a number (exact count depended on the algorithm settings - usually between 2 and 4) of specimens were randomly chosen (with return) from the population. Their fitness assessments were compared in order to determine the winner of the tournament. Winners of two such tournaments were coupled and went on to crossbreed. Crossover − each pair of respective linear combinations in a heuristic (i.e. each non-linear component and base heuristic function) was crossed over independently. The genotype of each linear combination was randomly partitioned into two. The offspring4 inherited values of each part of the genotype from one parent. The value of the gene on which the division was placed was randomly chosen from the interval defined by the values of this gene in parent specimens. The weakest specimen in the population was to be replaced by the newly created descendant. However, in default configuration of the algorithm used throughout the tests carried out, the replacement would happen only if the new specimen’s fitness evaluation was greater than that of the one to be replaced (with sharing function considered as usual). Mutation − three kinds of mutation could occur in each of the genes of every newly created specimen: multiplying a value of a gene by two, dividing it by two or changing its sign. Multiplying or dividing a value by two were twice as probable as changing the sign5. Each gene of a specimen mutated independently. The overall probability of mutation for each gene was within the range (5 · 10−4, 11 · 10−4) for HG and (6 · 10−4, 2 · 10−3) for the other generator (SHG) - see the next section. 3.2 Simple Heuristic Generator (SHG) The other approach to evolution of heuristic functions considered in this paper is based on a simplistic assumption that results of games played by pairs of specimens define a relation close to partial order. In order to compare the fitness of two specimens, they would play one or two (with sides swap) games against each other, depending on the algorithm settings. In order to strike the balance between speed and accuracy of such comparisons, the games would be played with a depth limit for alpha-beta of 3 (for simpler heuristics) or 4 (for more complex heuristic functions taking into considerations more parameters). If a draw occurred, a tie breaker heuristic could be used; however, it was switched off during all the tests. If, in case of two-game comparisons, each heuristic would score one win, the one winning in fewer moves would be considered the ultimate winner. 4Please note, that there was only ONE offspring from each pair. 5Instead of multiplication/division of gene’s value also addition/subtraction within some range was tested, but results were poorer in that case. 9 Basing on the relation described above, it was relatively easy to compare and sort specimens within a small set and, subsequently, perform a tournament selection without any fitness function whatsoever. The genetic operators used in the SHG algorithm bore great resemblance to those used in HG algorithm implementation. The only significant difference was the necessity to normalize specimens’ genotypes (since there was no external factor withholding the linear combinations of weights from exploding). Two specimens to crossbreed were chosen by means of tournaments and additional tournament was held to determine the weakest specimen to be replaced by the offspring. As it was the case in HG, the new specimen survived only if it was fitter than the one to be replaced. 4 Types of Heuristics Based on preliminary tests we have decided to inspect several types of heuristics in more detail. Some of them were linear and some consisted only of nonlinear components de- scribed above. Most of the heuristics were generated twice: once using HG and once with SHG. Additional tests for the game of GAC using SHG were also carried out in order to find out whether values of respective parameters for both players in linear heuristics exhibited symmetry (i.e. evolved to approximately opposite values). Pawns counts and kings counts weights were found to be the most asymmetrical. The other parameters tended to evolve symmetrical weights. It was therefore recommended that for these parameters differences in respective values instead of raw values be used in order to decrease the number of genes. As for the two asymmetrical parameters, a hypothesis was proposed that the asymmetry was a result of dependencies between parameters and therefore using differences instead of individual values would affect the quality of heuristics in a negative way. On the other hand, it might also be the case that simple evaluation functions do not properly assess the quality of a position on the board and the asymmetry results from conflicting win criteria and mobility issues mentioned earlier. After thorough exploration of this issue we leaned towards the latter hypothesis. Further simulations confirmed the efficiency of using raw numbers instead of differences for these two above mentioned parameters, but only in case of the simplest heuristics (see comparison between 8F and 10F heuristics presented in the following sections). Generally, we concluded that using more parameters (i.e. more sophisticated heuristics) and first of all dividing the entire game into stages (each with dedicated evaluation function) is more advisable than using raw material values (see sect. 4.2). Hence in all heuristics described in this paper, except for 10F ones, all parameters are defined as differences in respective values calculated for both players. 4.1 Linear Heuristics 8 Factors (8F). This heuristic took into consideration differences in values of eight pa- rameters, namely numbers of (1) pawns, (2) kings, (3) safe pawns, (4) safe kings, (5) moveable pawns, (6) moveable kings, (7) unoccupied fields on promotion line and (8) aggregated distance to promotion line. 10 Factors (10F). This heuristic took into account the same parameters as 8F the only exception being the fact that it considered raw numbers of pawns and kings for each player 10 separately (instead of differences). 15 Factors (15F). All factors of this heuristic were defined as differences of respective parameters calculated for both sides. Note, that this heuristic was defined differently for each game (the detailed description is presented in Table 1). 15F 15F 19F 25F E3Ph E3Ph feature name checkers GAC both both checkers GAC pawns X X X X X X kings X X X X X X aggregated distance to promotion line X X X X save pawns X X X X X X save kings X X X attacking pawns X X X X X X centrally positioned pawns X X X X X centrally positioned kings X X X X X X movable pawns X X X X movable kings X X unoccupied fields on promotion line X X defender pieces X X X X X pawns on main diagonal X X kings on main diagonal X X X X pawns on double diagonal X X kings on double diagonal X X X X X loner pawns X X X X loner kings X X surrounded empty fields X X X X X Bridge patterns X X X Oreo patterns X X X Triangle patterns X X X Dog patterns X X X X X ‘Man in corner’ patterns X X X ‘King in corner’ patterns X X X Table 1: List of factors defining 15F, 19F, 25F and E3Ph heuristics generated for check- ers and give-away checkers. A sign ”X” in any field denotes inclusion of the respective feature in the heuristic function. ‘Both’ denotes that heuristic definition is identical for both games. Please note, that all features are calculated as differences (not the raw values). 19 Factors (19F). Again, all components of this heuristic were in the form of differences between the respective values calculated for both players. It took into account all the parameters considered by 8F heuristic. Apart from that it used some parameters charac- terizing layout of pieces on the board. It should be noticed that patterns (Dog, Bridge, Oreo, ... etc.) were not included in this heuristics’ definition (see Table 1). 25 Factors (25F). This heuristic took into consideration differences in values of all available parameters described in sect. 2. 11 4.2 Nonlinear Heuristics The general idea of nonlinear heuristics was based on the fact that it is advantageous to divide the entire game into several disjoint stages and to use different heuristics for different stages [36]. Two crucial moments requiring change of the heuristic evaluation function were identified. Firstly, presence of kings clearly indicates that the game has entered an advanced stage. Moreover, due to the mobility issues mentioned earlier, end- game positions also require defining and applying different heuristic. Preliminary tests showed that introducing nonlinear components with conditions that were not disjoint hindered the genetic algorithm significantly. This was particularly the case for HG and resulted from the fact that having several overlapping conditions made it possible to achieve very similar results in many ways, each time with very different values of parameters. The following non-linear heuristics were developed: 3Phase (3Ph). In this heuristic the game was divided into three stages: Beginning − each player has more than 3 pawns and no kings are present on the board. Kings − both opponents have more than 3 pieces and at least one king is present. Ending − one of the players has at most 3 pieces left. Nonlinear components corresponding to each stage were defined. Linear heuristics in each component took into consideration the differences in exactly the same 8 parameters as in 8F heuristic (parameters concerning kings were, of course, only considered if kings were present on the board). For example phase Beginning was encoded by the following pseudo-algorithm: IF ABS(Players pieces count - 10.0) < 6.5 AND ABS(Opponent’s pieces count - 10.0) < 6.5 AND ABS(Total kings count) < 0.5 THEN Heuritic Value += C1 ∗ Dif f (1) + C2 ∗ Dif f (2) + C3 ∗ Dif f (3) + C4 ∗ Dif f (4) + C5 ∗ Dif f (5) + C6 ∗ Dif f (6) + C7 ∗ Dif f (7) + C8 ∗ Dif f (8), where C1, . . . , C8 are evolvable coefficients, and Dif f (n) denotes the value equal to the difference of feature n (listed in sect. 2) between player and his opponent. Expert 3 Phase (E3Ph). This heuristic was generated only using HG algorithm. The game was divided into stages in the same way as it was the case for 3Ph; however, apart from differences in numbers of pawns and kings, linear heuristics used in each of the stages took into consideration some parameters describing layout of pieces on the board (see Table 1 for detailed description). Again, parameters involving kings were not considered in the Beginning phase when no kings existed. Please note that this ‘expert’ heuristic was defined separately for checkers and GAC. For each game the respective set of contributing features was chosen basing on results obtained in previous runs of generators by means of choosing parameters which proved to be the most significant. 5 Phase (5Ph). In the last heuristic the division of the game was similar to the one used in 3Ph. The second stage (Kings) was subdivided into three stages: MyKings, HisKings and BothKings, depending on which player was in possession of kings. Linear heuristics used in each phase considered the same parameters as those in 3Ph. 12 5 Algorithm Settings In the HG, the initial depth in the game tree was between 81 and 87 for GAC and between 82 and 88 for checkers. The intervals were determined basing on preliminary tests devoted to calculating the average number of moves necessary to finish a game performing random moves. The difference in depths between subsequent phases was set to 6 since it was agreed that alpha-beta search with depth limit of 6 was still reasonably fast. In most tests the number of test boards ranged from 3 000 to 4 500 when reusing of boards was disabled and was equal to about 6 000 when this feature was enabled. While fewer situations were assessed in each phase during the generator runs without reusing test boards, the total number of situations considered was greater which resulted in better exploration of the problem space. On the other hand, evaluating as many as 6 000 boards during each phase minimized the chance of considering too few situations belonging to certain categories and propagating the error upwards. On the other hand yet again, if for some reason an erroneous heuristic was generated at an initial stage the chance of gradually recovering from the error was relatively small, because boards assessed by the faulty heuristic were reused in subsequent phases. Test populations consisted of 350 specimens. In each phase the weakest 80% of the population were regenerated. Depending on the number of genes of each specimen the size of a tournament was set between 2 and 5. The greater the number of specimens participat- ing in a tournament, the quicker the convergence was. However, if the algorithm converged too quickly, evolution of more sophisticated heuristics could be seriously hindered. For SHG each test population consisted of 100-150 specimens. Populations had to be smaller because of the way specimens were compared with each other. Comparisons were symmetrical, i.e. two specimens played two games against each other swapping sides after the first game. Depending on the complexity of a heuristic (i.e. on the number of genes in the longest linear combination) search depth in games played for comparison purposes was set to 3 (for simpler heuristics) or 4 (for more complicated ones). The latter led to the necessity to further reduce the number of specimens to 60 − 80. In order to compensate for decreasing size of the population mutation probability was increased. Significant improvement in the quality of heuristics generated was noticed. However, despite substantially reducing size of the population the algorithm was still considerably slower when search depth was set to 4 than when it operated at the depth of 3. 6 Evolutionary Process It was observed that HG tended to exhibit quicker convergence than SHG and that heuris- tics generated by the former were generally of slightly better quality. However, HG turned out to be very sensitive to changes in algorithm parameters. For instance, evolution of more complex heuristic used to converge too quickly if the size of tournament was too big. This resulted in the whole population reaching a local minimum which in turn led to heuristics of relatively poorer quality. Preliminary tests showed that using speciation for nonlinear heuristics led to worse results. The probable reason for this was the fact that the impact of the speciation on a specimen’s fitness could prevail over the influence of one of the categories not adequately represented in the test boards set. This might in turn lead to improper values of one of 13 the nonlinear components which might be propagated upwards. Tests also showed that HG genetic algorithm could be hindered significantly by in- troducing nonlinear components with conditions that were not disjoint. The reason for that was that overlapping conditions allowed for obtaining very similar assessments (and, therefore, fitness function values) with very different genotypes. Mutations seem to have had virtually no influence on the results of evolution of simple linear heuristics using HG. Significantly less than 1% of the fittest specimens (logged by the generator every 50 generations) turned out to have experienced mutation. The same was the case for more sophisticated linear heuristics taking into consideration some parameters concerning layout of pieces on the board. As for nonlinear heuristics generated using HG, the fraction of mutated specimens logged was substantially higher than for linear ones but still did not exceed 10%. Such a significant difference might result from the fact that in initial stages of the algorithm some genes were not applicable to the situations evaluated and therefore their mutation had no influence on the overall fitness of the specimen. As far as SHG is concerned, hardly any mutated specimens were logged. However, this probably stemmed from the fact that the fittest specimen was only chosen every 1 000 crossbreedings due to high cost of comparisons. When analyzing changes in intervals to which values of different genes belonged, it may be noticed that mutations quite frequently led to improved fitness. This was particularly the case for more sophisticated heuristics with larger numbers of genes and smaller populations. Since in all tests newly created specimen would replace the weakest specimen in the population (in SHG the specimen to be replaced was chosen by means of a tournament) only if it was fitter than the specimen to be replaced, the ratio of effective crossovers (i.e. the ones in which a resulting specimen was actually added to the population) to potential crossovers was also investigated. It turned out that the fraction remained fairly stable throughout the process. About 80%-90% of all crossovers were effective in HG. The stability resulted from the fact that in the initial stages of the algorithm convergence was comparatively quick and therefore descendants tended to be fitter than specimens from previous generations. On the other hand, in the final stages vast majority of the specimens were almost identical and there were virtually no difference in fitness between ancestors and descendants in which case new specimens were preferred and added to the population. As for SHG the ratio in question was around 60%-70% when search depth in games played for the purpose of comparison was set to 3 and 80%-90% when it was set to 4. This confirms the intuition that comparisons at the depth of 4 are more accurate, provide more meaningful results and minimize the impact of randomness. The convergence of the evolution is clearly illustrated by changes in lengths of inter- vals for different parameters as well as by distinct declines in their variances. For most parameters variances dropped by more than a thousand times in the course of evolution. In particular, in HG initial values of parameters were randomly selected from the inter- val (−100, 100) and after the evolution (i.e. in the final population) most of them were between −1 and 1 with standard deviation between 10−30 and 10−26. Similarly, for SHG initial values were located between (−0.5, −0.3) and (0.3, 0.5) - due to normalization pro- cedure with standard deviation about 0.1. After the evolutionary process was completed standard deviation dropped to about 10−5. 14 7 Results This section provides numerical results of several tests performed in order to estimate the quality of HG and SHG approaches. First, the results for GAC are presented in sect. 7.1. The results for checkers are placed in sect. 7.2. 7.1 Give-Away Checkers Due to the lack of GAC playing programs reliable estimation of the quality of heuristics developed in our experiment is not an obvious task. The heuristics were first tested against the TD-GAC program available to the authors. The results of this comparison are presented in section 7.1.1. Another way to estimate heuristics’ quality is their comparison with the null heuristic as suggested by Jonathan Schaeffer [37]. This approach is discussed in sect. 7.1.2. 7.1.1 Comparison with TD-GAC program In order to check the quality of heuristics generated for the game of GAC, they were tested against TD-GAC program [28, 22, 27, 29]. TD-GAC is based on temporal difference approach [41] and learns from the games played. Starting from scratch several TD players were trained in a three-step procedure (with subsequently stronger opponents - trainers) with the set of features following Samuel’s checkers approach [31, 32]. The strongest of the above players became the opponent for heuristics discussed here. Each heuristic played 40 games against the program, in half of them playing white and in the other half black stones. During the tests alpha-beta search depth was set to 6 and search depth of TD-GAC was set to 4 since it made use of much more sophisticated parameters, including indirect exploration of the game tree one ply further. TD-GAC learning function was switched off during the tests. Due to some randomness in searching the game tree implemented in alpha-beta, games played between any particular heuristic and TD-GAC were not identical. Most heuristics proved to perform well against the program with frequent wins and few losses. Results of the heuristics’ tournaments against TD-GAC are shown in Fig. 4. Each heuristic was assigned 1 point for a game won and 1 2 point for a draw. As it can be seen in the figure, nonlinear heuristics (in particular E3Ph and 3Ph) tended to outperform the linear ones. Similarly, heuristics generated using HG performed generally better than those evolved using SHG, which suggests that non-transitiveness of SHG specimens’ order relation as well as its smaller population size may hinder the evolutionary process. Poor outcomes of the 5Ph heuristics might result from their greater complexity which might require much bigger population, more sophisticated design of the evolutionary process and better tuning of algorithm parameters. 7.1.2 Comparison with null heuristic Another possible way of checking the efficacy of evolved heuristics is to compare them with a null heuristic. This approach was suggested by Jonathan Schaeffer who did a similar experiment and observed that for deep search levels it is extremely hard to beat a null heuristic in this particular game [37]. 15 0,750 0,688 0,650 0,650 0,550 0,525 0,513 0,488 0,438 0,375 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 H G :E 3 P h H G :3 P h H G :1 0 F S H G :3 P h S H G :1 5 F H G :5 P h H G :8 F S H G :5 P h S H G :1 0 F S H G :8 F Figure 4: Performance of the top-10 heuristics against TD-GAC. The y-axis represents the points scored by each heuristic as a fraction of maximum possible score. There are several ways to implement a null-heuristic approach in GAC. The most obvious one is to use a regular alpha-beta approach with a random estimation of the nodes6. This approach however may lead to several random alpha-beta cut-offs, among which good continuations could be lost. Hence, we have taken a different approach: in each node we examined the outgoing branches in random order. Whenever search depth limit was reached without encountering a terminal position, node was assessed as a draw (i.e. value 0). That way, as long as no leaves were encountered, absolutely no cut-offs were performed and the last (i.e. random - considering sorting procedure) branch was chosen in each of the nodes expanded. The first comparison was made for the search depth equal to 10 (for both evolved and null heuristics)7. The results were very encouraging (see Fig. 5): 9 out of 15 tested heuristics scored at least 75% of all available points in a 20-game match against the null heuristics playing half of the games with white stones and half of them with black ones. None of the heuristics was outperformed by the null one (the weakest heuristic achieved the result of 50%). In the next step we compared the heuristics playing at depth 6 with the null heuristic remaining at depth search equal to 10. 4-ply difference in search is of course a significant bias towards null heuristic but, surprisingly enough, the null heuristic was able to defeat only two evolved ones and achieved a draw with one heuristic. The remaining 12 of our heuristics scored between 55% and 77.5% points - again in a 20-game match with swapping sides after each game. The above results suggest that the quality of evolved heuristics is relatively high. Con- sidering the 4-ply depth search difference in the last experiment it may also be suggested 6Certainly, if a node represents the position of the end of a game (win, loss, draw) then the real (not random) value should be returned 7Please recall that as previously explained all tested heuristics were developed at search depth equal to 4 and application search depth was increased only during the experiment against null heuristic. 16 that evolved heuristics have potential for generalization and therefore can be successfully applied to different search depths. 0,875 0,850 0,800 0,800 0,775 0,775 0,750 0,750 0,750 0,725 0,700 0,700 0,650 0,650 0,500 0,000 0,100 0,200 0,300 0,400 0,500 0,600 0,700 0,800 0,900 1,000 S H G :3 P h H G :1 9 F H G :8 F S H G :1 9 F H G :E 3 P h S H G :5 P h S H G :8 F S H G :1 0 F S H G :2 5 F S H G :1 5 F H G :1 5 F H G :2 5 F H G :5 P h H G :1 0 F H G :3 P h Figure 5: Comparison of evolved heuristics and the null heuristic in 20-game matches with depth search equal to 10. Bars denote percentage of points scored by each heuristic. 7.2 Checkers In the game of checkers some initial tests were carried out to measure performance of the simple alpha-beta algorithm implemented. Depending on heuristic, 55 000 to 100 000 nodes were analyzed per second and search with depth limit set to 6 would take on average 40 to 200 ms. However, this average was calculated basing on games that were mainly considered ties after 140 plies performed and therefore many of the searches were performed during end-games when branching factor of the game tree was significantly higher than on average. Therefore alpha-beta might be expected to produce results up to 5 times faster for mid-game positions. Another factor examined was alpha-beta algorithm’s pruning efficiency, defined as 1− n s , where n is the actual number of nodes analyzed and s denotes theoretical size of the analyzed game subtree calculated basing on reported average branching factor and search depth. With search depth limit of 6, this parameter would usually fall into a range of 0.7 to 0.85, and up to 0.99 for searches with higher depth limits. 7.2.1 Comparison of heuristics against one another In order to determine relative quality of heuristics a tournament was held. During the tournament each heuristic played 10 games with each other with sides swap after each 17 game. Since 15 heuristics were tested, 1 050 games were played in total. Analogously to the case of GAC, due to some randomness in searching the game tree implemented in alpha-beta, games played between any two heuristics were pairwise different. Two classifications were used. The first one was based on the total number of games won, tied and lost by each heuristic. The other one took into consideration results of whole 10-game clashes between heuristics. In either case, heuristic would gain 1 point for a win and 1 2 point for a draw. In order to balance the influence of heuristic and random factor, while at the same time stay within reasonable time constraints all the games were played with alpha-beta’s search depth limit of 6. The results are presented in Fig. 6. Outcomes of the tournament show clearly that heuristics generated by HG outperformed those produced by SHG algorithm, regardless of the comparison method (individual games or clashes). The probable reason was the fact that partial order relation defined for the SHG application does not necessarily reflect the quality of the specimens and, what may be even more important, is in many cases non-transitive. Moreover, HG algorithm, thanks to its speed, is able to operate on 3 times bigger populations than SHG given the same amount of time. As might be expected the E3Ph heuristic proved to perform better than any other, no matter which comparison criteria were considered. More advanced heuristics, considering more sophisticated parameters, also tended to perform well in the tournament which again was in line with the expectations. Surprisingly poor score of the SHG 5Ph heuristic (located below the top-10 players) is a clear sign of SHG generator’s inefficiencies when dealing with more complicated specimens. 7.2.2 Comparison with public domain and commercial programs As it was mentioned before, our application was not meant to compete against commercial checkers programs. Its main purpose was to assess credibility of evolutionary approach to heuristic generation. It was developed with flexibility rather than speed in mind and as a result it assessed between 55 000 and 100 000 nodes per second whereas professional checkers applications were sometimes more than one order of magnitude faster. This is why the quality of the application should not be directly evaluated basing only on the results of games played against commercial applications. Nevertheless, in order to assess quality of the heuristics some comparison games were played. In order to make them fair, most of the features of the commercial programs (opening book, ending databases, hashtables, killer moves identification etc.) had to be disabled or reduced as much as possible. Seven best heuristics were chosen for those tests: HG:E3Ph, HG:15F, HG:25F, HG:19F, HG:10F, HG:3Ph and SHG:15F. The first checkers engine that was tested against the generated heuristics was Simple Checkers [12]. Although it is a relatively simple program with publicly available source code, it is at the same time one of the best engines among those not making use of opening books and end-game databases and it surpasses many commercial applications [25]. Since Simple Checkers is open-source program which compiles to a dll file, it was possible to automate test games between our heuristics and Simple Checkers. In order to make them fair, Simple Checkers algorithm was slightly modified so that it would always perform search exactly to the depth of 6 plies, i.e. the same depth as our alpha-beta algorithm. Simple Checkers generally appeared to be too strong opponent and HG:E3Ph turned out to be the only heuristic capable of winning at least one of 50 games played, but on the 18 other hand it was capable of drawing distinctly more than half of the rest, achieving the final score of 37%, i.e. 1 win, 35 ties, 14 loses. The next two heuristics were SHG:15F and HG:15F scoring 10% and 9%, resp. It can therefore be stated that, although none of the heuristics managed to surpass Simple Checkers, the best one of them, being able to at least draw more than 70% of the games, may be regarded as one of comparable quality. It is reasonable to hope that well-guided evolutionary process could lead to generation of even better performing heuristics, capable of reaching Simple Checkers’ level of play. In the next tests HG:E3Ph heuristic, the undoubtedly best of all heuristics generated, played against shareware version of commercial application Actual Checkers 2000A [4]. Actual Checkers program was set to the highest possible difficulty level and was expected to perform a move in 3 seconds on average, which resulted in searches of depth 12 to 18. Our alpha-beta algorithm could analyze up to 8 million nodes but the actual number of nodes analyzed was usually well below 1 million per move, which was equivalent to search depth of approximately 9 plies during most of the mid-game but significantly less during end-game. In spite of lower search depth HG:E3Ph managed to draw 3 of 4 games played. Single games played with search depths of 9 and 10 ended in draws as well. During the tests it was observable that the HG:E3Ph heuristic would perform noticeably worse during end-games, often failing to take full advantage of its upper-hand or defend a draw well enough. This probably stemmed from the fact that Actual Checkers’ hashtables proved especially useful in this phase of the game. Finally, in order to once again confirm quality of the HG:E3Ph heuristic, games against another shareware edition of commercial checkers program Mad Checkers [33] were played. This time alpha-beta’s depth limit was set to 8 as usual, but Mad Checkers engine, being less acclaimed application, was given 0.5s for its move, which was not significantly less than it usually took our program to assess mid-game situations. In order to make the competition even easier for Mad Checkers, alpha-beta was switched to depth limit of 7 (or in some cases even 6) whenever it happened to perform too slowly during endgames. Nevertheless, our heuristic managed to win 3 out of 4 games with one being a draw because of position repetitions, although Mad Checkers was left with 3 pieces against our heuristic’s 5. For the second and final 4-game tournament against Mad Checkers our alpha-beta algorithm was switched to depth limit of 6 plies whilst Mad Checkers was still admitted 0.5s for a move (i.e. on average more than two times longer than our program would require for a move). This time our heuristic managed to win 1 game and draw 3 others (with one of the draws resulting, again, from heuristic’s inability to make proper use of the fact that it outnumbers opponent by 2 kings). 8 Conclusions The focus of this paper is to test usefulness of pure genetic approach for generating eval- uation functions in two game domains: GAC and checkers. The underlying assumption is simplicity and generality of proposed solutions which results in using straightforward genetic operators and shallow game tree search with plain alpha-beta algorithm. In par- ticular, no alpha-beta search enhancements (transposition tables, history heuristics, killer moves, etc.) and no opening books or end-game databases are used. Under the above assumptions one cannot expect to create a master playing program capable of competing with strong humans or dedicated commercial software. What could 19 however be expected is a decent level of play and possibility of drawing several conclusions concerning generalization abilities (due to shallow search level) and efficacy of proposed genetic approaches. Unlike the most of top-level playing programs both generators proposed in this paper are problem independent and therefore the observations attained for GAC and checkers may possibly be valid to a wide range of two-player perfect information board games. In our opinion, especially the Heuristic Generator (HG) deserves recommendation and further community interest, in particular in zero-knowledge approaches. Generally, the results support the following four conclusions: Firstly, HG is definitely a more efficient generator compared to SHG. For the game of checkers the top-5 best heuristics were produced by HG (see Fig. 6). Comparisons against commercial programs were also more favorable for HG. In give-away checkers the three winning heuristics against TD-GAC program were HG:E3Ph, HG:3Ph and HG:10F. Only comparison against null heuristic was ‘inconclusive’ with a little favor for SHG. Generally, the results support the claim that careful and tuned implementation of HG may lead to high quality heuristic evaluation functions for both games. Secondly, for both games it is advantageous to divide the entire game into phases and develop separate heuristics for each part of the game. This is certainly a well-known recommendation in case of checkers and other board ‘mind’ games, but in our particular genetic implementation it was observed that game phases need to partition board positions space into disjoint sets. Otherwise the genetic process may have difficulties in assigning coefficients for the features shared by two or more game phases. Thirdly, the general guideline is to use differences of respective parameters calculated for both sides rather than raw numbers separately for both players. This observation was fully confirmed, with the only exception being the case of relatively simple heuristics, where it is recommended that the number of pawns and the number of kings be input as raw values (c.f. comparison between 8F and 10F heuristics). Fourthly, one of potential weaknesses of the two proposed generators might have been a shallow search depth used during the evolution phase (usually equal to 4), but the results of comparison with null heuristic in GAC and comparisons with several checkers programs suggest that, on the contrary, even though heuristics were generated based on 4-ply searches, they were capable of efficient generalization at greater search depths (from 6 to 10 in our tests). Several other more specific conclusions can also be drawn from numerical results and implementation details. We hope that detailed presentation in this paper of several ‘tech- nical’ issues may be helpful to other researchers working in the area of games and compu- tational intelligence paradigms. To summarize, achieved outcomes suggest that usage of genetic algorithms may be a credible way of producing heuristic functions for two-player games. Although heuristics described in this document did not surpass or even reach the level of play of those created during years of checkers programs development, they are still a challenge for at least some of the commercially available programs. This indicates that evolutionary approach, tuned by a careful choice of settings and meta-rules can prove very useful and successful, in particular for games, for which not enough explicit expert knowledge exists. 20 References [1] Akl, S.: Checkers-playing programs. In Shapiro, S.C., ed.: Encyclopedia of Artificial Intelligence. Volume 1. John Wiley & Sons, New York (1987) 88–93 [2] Aleksander, I.: Neural networks - evolutionary checkers. Nature 402 (1999) 857 [3] Alemanni, J.B. http://perso.wanadoo.fr/alemanni/ give away.html (1993) [4] Atlant Software Inc.: Actual checkers 2000a program (2005) [5] Barone, L., While, L.: An adaptive learning model for simplified poker using evo- lutionary algorithms. In: Proceedings of the Congress of Evolutionary Computation (GECCO-1999). (1999) 153–160 [6] Barone, L., While, L.: Adaptive learning for poker. In: Proceedings of the Genetic and Evolutionary Computation Conference. (2000) 566–573 [7] Borkowski, M.: Analysis of algorithms for two-player games. M.Sc. Thesis, Warsaw University of Technology (in Polish) (2000) [8] Chellapilla, K., Fogel, D.: Evolution, neural networks, games, and intelligence. Pro- ceedings of the IEEE 87 (1999) 1471–1496 [9] Chellapilla, K., Fogel, D.: Anaconda defeats Hoyle 6-0: A case study competing an evolved checkers program against commercially available software. In: Congress on Evolutionary Computation, La Jolla, CA, USA. (2000) 857–863 [10] Chellapilla, K., Fogel, D.: Evolving a neural network to play checkers without hu- man expertise. In Baba, N., Jain, L., eds.: Computational Intelligence in Games. Volume 62. Springer Verlag, Berlin (2001) 39–56 [11] Darwen, P., Yao, X.: On evolving robust strategies for iterated prisoner’s dilemma. Volume 956 of LNCS., Springer (1995) 276–292 [12] Fierz, M.: Simple checkers program (2005) [13] Fogel, D.B.: Blondie24: Playing at the Edge of Artificial Intelligence. Morgan Kaufmann (2001) 21 [14] Fogel, D., Hays, T., Hahn, S., Quon, J.: A self-learning evolutionary chess program. Proceedings of the IEEE 92 (2004) 1947–1954 [15] Goldberg, D.E., Richardson, J.: Genetic algorithms with sharing for multimodal function optimisation. In Grefenstette, J.J., ed.: Proceedings of the 2nd Interna- tional Conference on Genetic Algorithms, Cambridge, MA, USA, Lawrence Erlbaum Associates (1987) 41–49 [16] Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Pub. Co. (1989) [17] Griffith, K.: A comparison and evaluation of three machine learning procedures as applied to the game of checkers. Artificial Intelligence 5 (1974) 137–148 [18] Kaelbling, L., Littman, M., A.W.Moore: Reinforcement learning: A survey. Journal of Artificial Intelligence Research 4 (1996) 237–285 [19] Kendall, G., Whitwell, G.: An evolutionary approach for the tuning of a chess evaluation function using population dynamics. In: Proceedings of the 2001 Congress on Evolutionary Computation CEC2001, IEEE Press (2001) 995–1002 [20] Kotnik, C., Kalita, J.K.: The significance of temporal-difference learning in self- play training td-rummy versus evo-rummy. In Fawcett, T., Mishra, N., eds.: Ma- chine Learning, Proceedings of the Twentieth International Conference (ICML 2003), Washington, DC, USA, AAAI Press (2003) 369–375 [21] Kusiak, M., Walȩdzik, K., Mańdziuk, J.: Evolution of heuristics for give-away check- ers. In Duch, W., et al., eds.: Artificial Neural Networks: Formal Models and Their Applications - Proc. ICANN 2005, Part 2, Warszawa, Poland. Volume 3697 of LNCS., Springer (2005) 981–987 [22] Mańdziuk, J., Osman, D.: Temporal difference approach to playing give-away check- ers. In Rutkowski, L., et al., eds.: 7th Int. Conf. on Art. Intell. and Soft Comp. (ICAISC 2004), Zakopane, Poland. Volume 3070 of LNAI., Springer (2004) 909–914 [23] Moriarty, D.E., Miikkulainen, R.: Discovering complex othello strategies through evolutionary neural systems. Connection Science 7 (1995) 195–209 22 [24] Newell, A., Shaw, J., Simon, H.: Chess-playing programs and the problem of com- plexity. IBM Journal of Research and Development 2 (1958) 320–335 [25] Newell, B.: Checkers programs repository (2005) [26] Oldbury, D.: AI (Any Interest) for draughts? In Levy, D., Beal, D., eds.: Heuristic Programming in Artificial Intelligence: the first computer olympiad. Ellis Horwood, Chichester, UK (1989) 165–175 [27] Osman, D., Mańdziuk, J.: Comparison of tdleaf(λ) and td(λ) learning in game playing domain. In Pal, N.R., et al., eds.: 11th Int. Conf. on Neural Inf. Proc. (ICONIP 2004), Calcutta, India. Volume 3316 of LNCS., Springer (2004) 549–554 [28] Osman, D., Mańdziuk, J.: TD-GAC. http://gac-arena.gt.pl/ (2004) [29] Osman, D., Mańdziuk, J.: TD-GAC: Machine Learning experiment with give-away checkers. In Dramiński, M., Grzegorzewski, P., Trojanowski, K., Zadrożny, S., eds.: Issues in Intelligent Systems. Models and Techniques. Exit (2005) 131–145 [30] Pollack, J.B., Blair, A.D., Land, M.: Coevolution of a backgammon player. In Lang- ton, C.G., Shimokara, K., eds.: Proceedings of the Fifth Artificial Life Conference, MIT Press (1997) 92–98 [31] Samuel, A.L.: Some studies in machine learning using the game of checkers. IBM Journal of Research and Development 3 (1959) 210–229 [32] Samuel, A.L.: Some studies in machine learning using the game of checkers II - recent progress. IBM Journal of Research and Development 11 (1967) 601–617 [33] Sapphire Games: Mad checkers program (2005) [34] Schaeffer, J., Culberson, J.C., Treloar, N., Knight, B., Lu, P., Szafron, D.: A world championship caliber checkers program. Artificial Intelligence 53 (1992) 273–289 [35] Schaeffer, J., Lake, R., Lu, P., Bryant, M.: Chinook: The world man-machine checkers champion. AI Magazine 17 (1996) 21–29 23 [36] Schaeffer, J.: One Jump Ahead: Challenging Human Supremacy in Checkers. New York: Springer-Verlag (1997) [37] Schaeffer, J. Private communication (2005) [38] Seo, Y.G., Cho, S.B., Yao, X.: Exploiting coalition in co-evolutionary learning. In: Proceedings of the 2000 Congress on Evolutionary Computation. Volume 2., IEEE Press (2000) 1268–1275 [39] Shannon, C.E.: Programming a computer for playing chess. Philosophical Magazine 41 (7th series) (1950) 256–275 [40] Smeets, J., Putter, G.: Some experience with a self-learning computer program for playing draughts. In Levy, D., Beal, D., eds.: Heuristic Programming in Artificial Intelligence: the first computer olympiad. Ellis Horwood, Chichester, UK (1989) 176–194 [41] Sutton, R.: Learning to predict by the method of temporal differences. Machine Learning 3 (1988) 9–44 [42] Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press, Cambridge, MA (1998) [43] Turing, A.M.: Digital computers applied to games. In Bowden, B.V., ed.: Faster than thought: a symposium on digital computing machines. Pitman, London, UK (1953) 24 Appendix 1. Example game of HG:E3Ph vs Actual Checkers An example of a game played by HG:E3Ph (Black) against Actual Checkers (White). HG:E3Ph searches 10 plies, Actual Checkers nominally makes a move in 3s, which is equivalent to searches between 12 and 18 plies. HG:E3Ph is making the first move. 1. 11-16 24-19 2. 9-13 28-24 3. 16-20 22-18 4. 8-11 18-14 5. 10-17 (14) 21-14 (17) 6. 6-10 25-21 7. 10-17 (14) 21-14 (17) 8. 1-6 29-25 9. 6-10 25-21 10. 10-17 (14) 21-14 (17) 11. 2-6 30-25 12. 6-10 25-21 13. 10-17 (14) 21-14 (17) 14. 13-17 26-22 15. 17-26 (22) 31-22 (26) 16. 7-10 14-7 (10) 17. 3-10 (7) 22-17 18. 11-16 23-18 19. 16-23 (19) 18-14 20. 23-26 14-7 (10) 21. 26-11 (K) 17-14 22. 31-26 7-3 (K) 23. 26-22 14-10 24. 22-18 10-6 25. 18-15 6-1 (K) 26. 15-10 24-19 27. 4-8 27-23 28. 8-11 23-18 29. 5-9 1-5 30. 9-14 18-9 (14) 31. 11-16 19-15 32. 10-19 (15) 3-7 33. 20-24 7-11 34. 16-20 9-6 35. 19-23 5-9 36. 24-28 6-1 (K) 37. 20-24 9-14 38. 23-26 14-10 39. 24-27 32-23 (27) 40. 26-19 (23) 11-15 41. 19-23 1-6 42. 28-32 (K) 6-9 43. 32-28 9-14 44. 23-19 15-24 (19) 45. 28-19 (24) 10-7 46. 12-16 7-11 47. 16-20 14-10 48. 20-24 11-15 49. 19-23 10-7 50. 23-19 7-10 51. 19-23 10-7 52. 23-19 7-10 53. 19-23 1 2 : 1 2 Three-move repetition draw in equal position. 25 Appendix 2. Example game of HG:E3Ph vs Mad Checkers An example of a game played by HG:E3Ph (Black) against Mad Checkers (White). Both sides use in average 0.5s per move. Mad Checkers is making the first move. 1. 23-19 9-13 2. 22-18 11-16 3. 18-14 10-17 (14) 4. 21-14 (17) 16-23 (19) 5. 27-18 (23) 8-11 6. 25-22 11-16 7. 22-17 13-22 (17) 8. 26-17 (22) 16-20 9. 32-27 6-9 10. 30-26 1-6 11. 29-25 9-13 12. 25-22 12-16 13. 26-23 4-8 14. 31-26 6-10 15. 24-19 8-12 16. 28-24 3-8 17. 18-15 7-11 18. 14-7 (10,K) 11-25 (15,22) 19. 17-14 2-11 (7) 20. 26-22 25-29 (K) 21. 22-18 13-17 22. 14-9 5-14 (9) 23. 18-9 (14) 17-21 24. 9-6 29-25 25. 6-2 (K) 25-22 26. 19-15 11-18 (15) 27. 23-14 (18) 22-18 28. 14-9 8-11 29. 9-6 11-15 30. 27-23 18-27 (23) 31. 2-7 27-32 32. 7-11 20-27 (24) 33. 11-18 (15) 16-19 34. 18-15 19-24 35. 6-2 (K) 24-28 36. 2-6 27-31 (K) 37. 15-18 32-27 38. 18-22 28-32 (K) 39. 6-9 12-16 40. 9-6 27-24 41. 6-9 24-28 42. 9-6 28-24 43. 6-9 24-28 44. 9-13 16-19 45. 22-17 19-23 46. 17-22 28-24 47. 13-9 23-27 48. 9-14 32-28 49. 14-9 24-19 50. 9-5 27-32 (K) 51. 5-9 32-27 52. 9-6 19-15 53. 6-9 27-23 54. 9-5 31-27 55. 5-9 15-11 56. 9-5 11-15 57. 5-9 15-10 58. 9-5 28-24 59. 22-17 24-19 60. 5-1 27-24 61. 1-5 23-18 62. 5-1 24-28 63. 1-5 18-14 64. 17-13 19-23 65. 5-1 23-18 66. 13-17 14-9 67. 17-13 9-5 68. 13-9 5-14 (9) 69. 1-5 28-24 70. 5-1 14-9 71. 1-5 18-14 72. 5-1 9-5 73. 1-6 10-1 (6) 0:1 HG:E3Ph finally wins after having some trouble with taking advantage of its material superiority in the end-game. 26 0,639 0,607 0,571 0,564 0,557 0,521 0,489 0,479 0,479 0,475 0,000 0,100 0,200 0,300 0,400 0,500 0,600 0,700 HG: E3Ph HG: 15F HG: 10F HG: 19F HG: 25F SHG: 15F HG: 3Ph SHG: 8F HG: 8F SHG: 10F (a) Comparison based on individual games. 0,929 0,750 0,714 0,714 0,714 0,536 0,500 0,464 0,429 0,393 0,000 0,100 0,200 0,300 0,400 0,500 0,600 0,700 0,800 0,900 1,000 HG: E3Ph HG: 19F HG: 25F HG: 15F HG: 10F SHG: 15F HG: 5Ph HG: 8F HG: 3Ph SHG: 10F (b) Comparison based on 10-game clashes. Figure 6: Performance of the top-10 checkers heuristics in the tournament. The y-axis in each figure represents the fraction of points scored by each heuristic. 27 
work_zzyeey6arvaxbhnnhyc5xqfngy	----	D:/ahmed/My-Documents/latex/AI_journal/EEA/EEA_08v1.dvi A Dependency-Based Search Strategy for Feature Selection Ahmed Al-Ani Faculty of Engineering and Information Technology University of Technology, Sydney, Australia ahmed@eng.uts.edu.au Abstract Feature selection has become an increasingly important field of research. It aims at finding optimal feature subsets that can achieve better generalization on unseen data. However, this can be a very challenging task, especially when dealing with large feature sets. Hence, a search strategy is needed to explore a relatively small portion of the search space in order to find ”semi-optimal” subsets. Many search strategies have been proposed in the literature, however most of them do not take into consideration relationships between features. Due to the fact that features usually have different degrees of dependency among each other, we propose in this paper a new search strategy that utilizes dependency between feature pairs to guide the search in the feature space. When compared to other well-known search strategies, the proposed method prevailed. Key words: Feature selection, search strategy, dependency, mutual information Preprint submitted to Expert Systems with Applications September 23, 2008 1. Introduction The identification of optimal feature subsets plays an important role for a wide range of classification problems, as it can lead to better performance in terms of accuracy and computational cost. The selection of good feature subsets requires an evaluation measure to estimate the goodness of subsets and a search strategy to generate candidate feature subsets (1). Evaluation measures are broadly divided into two categories: filters and wrappers. A filter method uses a measure that is independent of the predetermined classi- fication algorithm to estimate the goodness of candidate subsets. A wrapper method on the other hand estimates the goodness of a candidate subset using the classification accuracy obtained by feeding that particular subset to the adopted classification algorithm. Thus, wrappers are computationally more expensive than filters, however they are usually more accurate. Searching for the optimal subset, which can achieve the best performance according to the defined evaluation measure, is quite a challenging task. The exhaustive search, which considers all possible subsets, is guaranteed to find the optimal solution. However, it is impractical to run, even with moderate size feature sets. A number of other search strategies that differ in their computational cost and optimality have been proposed in the literature. One of the early search strategies is the branch and bound (2), which requires the evaluation function to be monotonic. This method can be computationally expensive for large data sets. Sequential search methods, such as sequential forward selection and sequential backward elimination (3), have been widely used because of their simplicity and relatively low computational cost. The major drawback of the traditional sequential search methods is the nesting 2 effect, i.e., in backward search when a feature is deleted, it cannot be re- selected, and in forward search when a feature is selected, it cannot be deleted afterwards. A slightly more reliable sequential search method is the plus- l-minus-r (l − r), which considers removing features that were previously selected and selecting features that were previously eliminated (4). Another trend of search strategies is the stochastic search, where it has been found that including some randomness in the search process makes it less sensitive to the dataset (1), and hence helps avoid local minimas. Some of the famous stochastic methods used in feature selection are: simulated annealing (5), Genetic Algorithm (GA) (6), Ant Colony Optimization (ACO) (7), Particle Swarm Optimization (PSO) (8) and Differential Evolution (DE) (9). Despite the encouraging results achieved in certain cases, the main draw- back of most of the above methods is that they do not take into consideration the importance of features and how they are related. This may have an ef- fect on the performance, especially for complicated search problems. Some methods attempt to estimate the relevance of subsets of features as the basis for selection. It has been shown that estimating the relevance of individual features may not be difficult, however, the real challenge is to estimate the relevance of subsets of features. This issue has been studied by a number of researchers (10; 11; 12; 13; 14). Some of the interesting finding of Guyon et al. (1) are: • a feature that is irrelevant by itself may become relevant when used with other features; • a relevant feature may not be needed because of possible redundancies. 3 Because of the difficulty associated with estimating the relevance of subsets of features, the proposed search strategy adopts the wrapper approach and focuses on dependency between feature pairs as a mean to guide the search. The paper is organized as follows: the next section explains the impor- tance of dependency between features. Section 3 describes the proposed search strategy. Experimental results are presented in section 4, and a con- clusion is given in section 5. 2. Dependency Between Features One of the promising approaches in searching the feature space is population- based search, where the current population consists of a number of feature subsets. Each one of these subsets is modified, in a certain way, to produce the next generation of subsets. Examples of population-based search proce- dures include Genetic Algorithms (GA), Particle Swarm Optimization (PSO) and Differential Evolution (DE). All of these methods were not originally de- veloped for feature selection, and hence, their original implementations do not take into consideration relationships between features when producing future populations. Let’s presume that F is the original feature set with n = 20 features, S1 and S2 are two subsets of the current population, each with m = 4 features. Let’s also consider that S1 is one of the best subsets that has been tested so far and we would like to modify S2, with the help of S1, to produce a better subset for the next population. If S1 = {f2, f5, f9, f15}, S2 = {f1, f2, f6, f17} and features f1 and f9 are highly dependent, then it will be logical to only consider replacing f6 and/or f17. Furthermore, instead of randomly choosing 4 any feature from F as a replacement feature, the search space can be reduced using the dependency between features. For instance, if f6 is to be replaced, we first find which of the two features f5 and f15 is closer to it, i.e., has a higher dependency value. Let’s presume that f15 is closer to f6. Then, candidate replacement features are reduced to the ones that lie (in terms of dependency) between f6 and f15. The newly produced subset will not only have a chance of being a better subset than S2, but it might also outperform S1, if this replacement makes it closer to the optimal solution. In addition to the above, dependency is also useful in identifying the K best subsets that are selected so far. For instance, we can not choose both S1 = {f2, f5, f9, f15} and S3 = {f1, f2, f5, f15} to be among the K best subsets, since they are almost identical, where f1 and f9 are highly dependent as mentioned earlier. This restriction will enhance the exploration capability, where diversity among the “good” subsets is quite important in avoiding local minimas. A famous approach to estimate dependency between two random vari- ables is based on the concept of mutual information (15). The mutual infor- mation between random variables X and Y is defined as: I(X; Y ) = ∫ PXY (x, y) log ( PXY (x, y) PX (x)PY (y) ) dxdy (1) The entropy, which is a measure of uncertainty of random variables, is usually used to represent mutual information according to the following for- mula: I(X; Y ) = H(X) + H(Y ) − H(X, Y ) (2) = H(X) − H(X|Y ) (3) 5 where H(X) is the entropy of X, H(X, Y ) is the joint entropy of the two random variables, while H(X|Y ) is the conditional entropy, which represents the uncertainty in X after knowing Y . When we deal with real data, the main problem for evaluating I(X; Y ) is the estimation of probabilities PX (x),PY (y) and PXY (x, y). One possible solution is to subdivide the XY plane into boxes of size ∆x∆y. By doing so, we are able to estimate the discrete values of PX , PY and PXY . An alternative approach was proposed in (16), which uses variable box size over the XY plane. The method presented in (17) estimates the MI by an adaptive partitioning of the observation space. A Parzen window method is proposed in (18) to estimate the input distribution. For simplicity, we used here a fixed box size implementation. Hence, the MI can be rewritten as: I(X; Y ) = ∑ rx ∑ ry PXY (rx, ry) log ( PXY (rx, ry) PX (rx)PY (ry) ) (4) where, rx and ry are the discrete levels for X and Y respectively. Let’s consider the earlier example about subset S2 and suppose that we have the entropies shown in Fig. 1, where H(C) represents the entropy of the target classes. The objective here is to select the optimal subset of features that would minimize the uncertainty in C. Features f1, f2 and f17 provide good information about the target classes, as indicated by the area covered in shaded vertical lines. The figure also shows that the additional information provided by f15 is more than that provided by f6 (the shaded grid areas). However, instead of presuming that f15 is the best feature when combined with f1, f2 and f17, why not searching the area that surrounds f15, i.e., features that are closer to f15 than f6. This limited space may include 6 �� �� �� �� �� �� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ��������������������� ����� ����� ����� ����� ����� ����� ����� ����� ����� ����� H(f15) H(C) H(f6) H(f1,f2,f17) H(f11) Figure 1: Importance of features in subset S2, and identifying a candidate replacement feature for f6 good features, such as f11, which when replacing f6 will not only make S2 a better subset, but it will also make it better than the subset {f1, f2, f15, f17}. 3. The Proposed Search Strategy The proposed search strategy is similar to GA, ACO, PSO and DE in the sense that they are all population-based methods. However, unlike other methods, the proposed search strategy utilizes dependency between feature pairs to guide the search. We will refer to it as DSS (Dependency-based Search Strategy), and it is implemented using a wrapper approach as follows: 1. Randomly initialize the subsets of the first population and sort them according to their fitness values 2. For each subset, Si, of the current population • Make a copy of Si 7 • Randomly choose one of the K best subsets, Sk • Identify candidate features for replacement, which are features of Si that are not present in Sk. Place those features in S ′ i • Initially assign 0 to the accumulated difference between the original and newly generated subset, AcDif f = 0 • do – Randomly choose one of the features of S′i, fi,j and find the feature fk,l in Sk that maximizes I(fi,j; fk,l) – identify the candidate features to substitute fi,j according to the following equation: Sm = argfm ( I(fk,l; fm) ≥ I(fk,l; fi,j) ) (5) – Randomly choose one of the features of Sm, to substitute fi,j , S′i ←S ′ i \{fi,j} – Calculate the accumulated difference between the replaced fea- tures and their replacements, as follows: AcDif f = AcDif f + (H(fm) − I(fm; fi,j)) (6) • while AcDif f is less than a certain constant and S′i 6= ∅ • Repeat the same procedure of step 2 for the remaining subsets 3. Evaluate the newly generated subsets 4. Select the K best subsets from the original and newly generated subsets, given that there is at least a certain ratio of difference between any two subsets. This is measured by averaging the mutual information between 8 each feature from a given subset and its closest counterpart in another subset 5. Select the remaining subsets of the population in a similar procedure as in the previous step 6. Discard the unselected subsets 7. If the stopping criterion is not met, goto step 2 to generate another population The rationale behind Eq. 5 is that if fi,j is the closest feature to fk,l and the two features are highly dependent on each other, then Sm will only con- sist of fi,j and fk,l. Hence, the replacement of fi,j will have little or no effect on the performance of Si. On the other hand, if there is only a small degree of dependency between fi,j and fk,l, then Sm will consist of other features that are closer to fk,l than fi,j , and in this case the replacement will most likely have an effect on the performance of Si. In other words, Sm will have an adaptive size, where the number of the candidate replacement features depends upon the degree of dependency between fi,j and fk,l. Eq. 6 is used to specify when to stop replacing features. If features fm and fi,j are highly dependent on each other, then I(fm; fi,j) will approach H(fm) according to Eq. 3 (the uncertainty in one of them after knowing the other will be close to 0) and AcDif f will almost stay unchanged. Accordingly, this replacement is not expected to make a noticeable difference on the performance of Si, and hence, the algorithm will consider replacing another feature. On the other hand, if fm and fi,j are not highly dependent, then AcDif f will have a higher value due to this replacement. When AcDif f exceeds a certain threshold, there will be some difference between the original and newly generated sub- 9 sets. Note that we do not want the original and newly generated subsets to be very different, as this may lead to big jumps in the search space and will hinder the convergence of the algorithm. Identifying the K best subsets also plays an important role in guiding the search, as it is important to choose good and diverse subsets. Concentrating on the goodness of subsets without considering how similar those subsets are may cause the search to be trapped in local minima. Thus, steps 4 and 5 aim at selecting good and yet diverse subsets to better explore the search space. 4. Experimental Results Three different classification problems are considered. In the first prob- lem the Madelon dataset, which is obtained from the UCI repository, is used. EEG data and speech segments are used in the second and third problems respectively. The following methods were implemented to select feature sub- sets: genetic algorithms, particle swarm optimization, differential evolution and the proposed dependency-based search strategy. A GA-based feature selection solution would typically be a fixed length binary string representing a feature subset, where the value of each position in the string represents the presence or absence of a particular feature. A traditional GA algorithm was used here with probability of crossover = 0.8, and probability of mutation = 0.05. The obtained strings are constrained to have the number of ’1’s matching a predefined number of desired features. Feature selection using PSO is also implemented using binary strings. It uses particles’ best and global best values to guide the search. A DE-based feature selection utilizes a real-valued optimizer and applies a rounding and 10 uniform crossover operators. The proposed DSS is implemented as explained in the previous section. Reasonable values of K and the ratio of difference between subsets were found to be around 5 and 0.1 respectively. For all experiments described below, a population size of 50 was used by the four feature selection methods. In addition, all of the four methods started with the same initial population and they all have the same stopping criterion, which is reaching a pre-defined maximum number of iterations. 4.1. The Madelon Dataset The Madelon dataset was designed for the NIPS 2003 variable selection benchmark. It is a two-class classification problem with sparse binary input features. There are 500 features, only 5 of which are useful and the rest are either redundant or irrelevant. 2000 patterns were used for training and 600 for validation. This dataset represents quite a useful benchmark for feature selection, as the optimal subset of features is already known. For all of the four methods, classification accuracy obtained using the k- nearest neighbor (k-NN) classifier with k = 5 was used as a fitness function, the desired number of selected features was set to 5 and the maximum number of iterations was set to 100. The experiment was repeated 30 times and the averaged classification accuracy values are shown in Fig. 2. The results indicate that the DSS not only achieved a higher accuracy by the end of the 100 iterations, but also required smaller number of iterations to find the optimal subset. PSO was found to be the second best, where it managed to find the optimal solution in one run out of the 30 repetitions. Despite being good optimization algorithms, the GA and DE did not perform very 11 0 10 20 30 40 50 60 70 80 90 100 55 60 65 70 75 80 85 90 95 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 2: Number of iterations vs. classification accuracy for the selection of 5 features well in this experiment, as they did not find the optimal solution in any run. This is mainly due to the large number of irrelevant and redundant features. The standard deviation values of the classification accuracy over the 30 runs for DSS, DE, GA and PSO are: 0.0, 3.23, 5.40 and 2.30 respectively. This indicates that GA has the highest fluctuation among the four methods, and it probably means that its performance is sensitive to the initial population when having large feature sets. 4.2. EEG Classification The second problem involves the classification of Electroencephalogram (EEG) signals into one of pre-determined mental tasks. The data used here was obtained from the Department of Medical Informatics, University of Technology, Graz, Austria. EEG signals were recorded from three right- handed females with 56 Electrodes. The subjects were asked to imagine right or left finger movements according to stimuli on screen. The wavelet packet transform was used to extract 3 features from each channel. Hence, the total number of features extracted was 168 (56 channels × 3 features/channel). 406 12 0 10 20 30 40 50 60 70 70 75 80 85 90 95 No. of selected features C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 3: Number of selected features vs. classification accuracy - EEG data trials were recorded, 208 for left movement and 198 for right. 300 patterns were used for training and the remaining 106 for testing. The desired number of selected features was varied from 3 to 70. The selection of feature subsets was performed using the four methods mentioned earlier, and the number of iterations was set to 400. The experiment was repeated 30 times and a Linear Discriminant Analysis (LDA) classifier was used. The averaged classification accuracy values of the selected subsets are shown in Fig. 3. It is clear that DSS outperformed the other three methods in all cases, and hence it achieved excellent performance regardless of the size of the selected number of features. The second best algorithm is DE followed by GA and finally PSO. The detailed results for the selection of 3, 30 and 70 features are shown in Figs. 4, 5 and 6 respectively. For the selection of 3 features, the results indicate that DSS is the only method that converged to the optimal solution in most runs (29 out of 30). The DE, GA and PSO managed to achieve the optimal solution in 11, 10 and 1 runs respectively. The standard deviation values of the classification accuracy for DSS, DE, GA and PSO 13 0 50 100 150 200 250 300 350 400 64 66 68 70 72 74 76 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 4: Number of iterations vs. classification accuracy for the selection of 3 EEG features are: 0.52, 1.11, 1.63 and 0.98 respectively. The search became more complicated as the number of selected features increased. The results obtained for the selection of subsets of 30 features indicate that more time was needed for the methods to find good solutions. When evaluating the runs individually, DSS managed to outperform the other three methods in 21 runs, and it shared the highest accuracy with another method in four other runs. However, DE was a bit more consistent in its performance, where the standard deviation was 0.94, 0.67, 2.05 and 1.17 for DSS, DE, GA and PSO respectively. Searching for subsets of 70 features was even harder. On average, higher standard deviation values were recorded; 1.12, 1.20, 1.45 and 1.33 for the four methods in order, which indicate higher fluctuations in performance. As with the previous case, DSS achieved higher accuracies in more runs than any other method (19 runs as a clear winner and shared the highest with another method in seven other runs). Fig. 6 shows an interesting trend, 14 0 50 100 150 200 250 300 350 400 76 78 80 82 84 86 88 90 92 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 5: Number of iterations vs. classification accuracy for the selection of 30 EEG features 0 50 100 150 200 250 300 350 400 76 78 80 82 84 86 88 90 92 94 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 6: Number of iterations vs. classification accuracy for the selection of 70 EEG features where unlike the other three methods that in many runs got trapped in local minimas, DSS continued to improve as the number of iterations increased. In fact, DE was a bit better than DSS in the first 250 iterations, but unlike DSS, the performance of DE has hardly improved after that. This trend of continuous improvement in the performance of DSS is mainly due to its enhanced exploration capability, as explained in sections 2 and 3. 15 0 50 100 150 200 250 300 350 400 65 70 75 80 85 90 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 7: Number of iterations vs. classification accuracy for the selection of 30 EEG features using a kNN classifier 0 50 100 150 200 250 300 350 400 74 76 78 80 82 84 86 88 90 92 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 8: Number of iterations vs. classification accuracy for the selection of 30 EEG features using a SVM classifier In addition to the LDA classifier, the k-NN and Support Vector Machine (SVM) were also considered. Results are shown in Figs. 7 and 8 for the selection of subsets of 30 features. These figures indicate that DSS clearly outperformed the other three methods, even when considering different clas- sifiers. It is worth mentioning that k-NN and SVM did not perform as good as LDA, due to the limited number of available patterns. 16 4.3. Speech Segment Classification In the third experiment, speech segments are classified according to their manner of articulation. Nine classes were considered: vowel, semi-vowel, nasal, fricative, stop, closure, lateral, rhotic, and silence. We used speech signals from the TIMIT database, where segment boundaries were identified. Three different sets of features were extracted from each speech frame: 16 log mel-filter bank (MFB), 12 linear predictive reflection coefficients (LPR), and 10 wavelet energy bands (WVT). A context dependent approach was adopted to perform the classification. So, the features used to represent each speech segment, Segn, were the average frame features over the first and second halves of Segn and the average frame features of the previous and following segments (Segn−1 and Segn+1 respectively). Hence, the base- line feature sets based on MFB, LPR, and WVT consisted of 64, 48 and 40 features respectively. An LDA classifier was used to classify the features of each baseline set into one of the nine manner-of-articulation classes. 4000 segments were used, the first 2000 for training and the last 2000 for valida- tion. The obtained classification accuracy for MFB, LPR and WVT were 75.40%, 63.15% and 72.80% respectively. It is clear that MFB achieved the best performance among the three baseline sets, however, it used more fea- tures than the other two baseline sets. The three baseline feature sets were concatenated to form a new set of 152 features. The four feature selection algorithms were used to select from these features. The desired number of selected features was varied from 7 to 70. 300 iterations were used by each method, and the experiment was repeated 20 times. The first thing that can be noticed from the averaged classification results, 17 10 20 30 40 50 60 70 68 70 72 74 76 78 80 82 No. of selected features C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 9: Number of selected features vs. classification accuracy - speech data shown in Fig 9, is the enhanced performance that was achieved in comparison to the three baseline feature sets. For instance, with 40 features DSS managed to achieve an average accuracy of 78.82%, which is not only higher than the accuracy of the 40 wavelet feature, but it is also higher than the accuracy of the 64 MFB features. The figure also indicates that the difference in performance between the four feature selection methods is not large, however, DSS still managed to outperform the other three methods in all cases. DE achieved quite good results in this experiment, as the number of irrelevant and redundant features was not big. The detailed results for the selection of 7 features are shown in Fig. 10. As with last experiment, DSS managed to converge faster than the other methods and managed to achieve the highest accuracy in 14 runs and share the highest accuracy with another method in 5 other runs. The standard deviation values of the classification accuracy for DSS, DE, GA and PSO are: 0.10, 0.20, 0.52 and 0.47 respectively. DSS also outperformed the other three methods when selecting 30 fea- tures. It achieved the highest accuracy in 17 runs and shared the highest 18 0 50 100 150 200 250 300 61 62 63 64 65 66 67 68 69 70 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 10: Number of iterations vs. classification accuracy for the selection of 7 speech features 0 50 100 150 200 250 300 72 73 74 75 76 77 78 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 11: Number of iterations vs. classification accuracy for the selection of 30 speech features accuracy with another method in 2 other runs. The figure also shows that unlike DE and DSS that managed to improve their performance as the num- ber of iterations increased, GA and PSO got trapped in local minimas. The standard deviation values of the classification accuracy for the four methods in order are: 0.24, 0.26, 0.61 and 0.33. Similar to the EEG experiment, results for the selection of 70 features, 19 0 50 100 150 200 250 300 78 78.5 79 79.5 80 80.5 81 81.5 82 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 12: Number of iterations vs. classification accuracy for the selection of 70 speech features shown in Fig. 12, reflect the enhanced exploration capability of DSS, where despite its slow start, compared to DE and GA, it continued to improve, while other methods trapped in local minimas at different locations of the search space. The figure also shows that the second best algorithm is DE, which outperformed GA and PSO with a clear margin. The standard deviation values are: 0.22, 0.22, 0.42 and 0.24 for DSS, DE, GA and PSO respectively. The reason behind the lower standard deviation values in this experi- ment is the fact that there are less variations in the relevance of available features in comparison to the previous two experiments. This can be noticed when comparing the classification accuracy ranges of the different figures, for example Figs. 6 and 12. Results for the selection of subsets of 30 features when using a k-NN classifier are shown in Fig. 13. One can notice that the k-NN classifier managed to achieve better results that LDA in this experiment. The results also show that DSS clearly outperform the other three methods, where it 20 0 10 20 30 40 50 60 70 80 90 100 86 86.5 87 87.5 88 88.5 89 89.5 90 90.5 No. of iterations C la ss ifi ca tio n a cc u ra cy DSS DE GA PSO Figure 13: Number of iterations vs. classification accuracy for the selection of 30 speech features using kNN achieved the highest accuracy in 17 out of 20 runs. 5. Conclusion This paper presented a novel search strategy for feature selection. The proposed method, which utilizes dependency between feature pairs to guide the search, proved to be very successful in exploring the feature space and avoiding local minimas. Three different problems were used to compare the proposed method with other population-based optimization search methods. The proposed method managed to achieve a consistently high-quality per- formance, and outperformed the other methods in all cases. It proved to be particularly useful when dealing with high number of redundant and irrele- vant features. 21 References [1] I. Guyon, S. Gunn, M. Nikravesh and L.A. Zadeh Feature Extraction: Foundations and Applications, Springer-Verlag New York, 2006. [2] P.M. Narendra and K. Fukunaga, ”A branch and bound algorithm for feature subset selection,” IEEE Trans. Computers, vol. 26, pp. 917-922, 1977. [3] R. Kohavi and G. H. John, ”Wrappers for feature subset selection,” Ar- tificial Intelligence, special issue on relevance, vol. 97, pp. 273-324, 1997. [4] P.A. Devijver and J. Kittler, Pattern Recognition: a statistical approach, Prentice-Hall, Englewood CliEs, NJ, 1982. [5] S-W. Lin, T-Y. Tseng, S-Y. Chou and S-C. Chen, ”A simulated- annealing-based approach for simultaneous parameter optimization and feature selection of back-propagation networks”, Expert Systems with Ap- plications, vol. 34 , no. 2, pp. 1491-1499, 2008. [6] H. Frohlich, O. Chapelle and B. Scholkopf, ”Feature Selection for Support Vector Machines by Means of Genetic Algorithms,” Proc. IEEE Intl. Conf. Tools with Artificial Intelligence (ICTAI’03), pp. 142-148, 2003. [7] A. Al-Ani, ”Feature Subset Selection Using Ant Colony Optimization”, Int. Journal of Computational Intelligence, vol. 2, pp. 53-58, 2005. [8] H.A. Firpi and E. Goodman, ”Swarmed Feature Selection”, Proc. Applied Imagery Pattern Recognition Workshop, pp. 112–118, 2004. 22 [9] R. Khushaba, A. Al-Ani and A. Al-Jumaily, ”Differential Evolution based Feature Subset Selection,” Proc. Intl. Conf. Pattern Recognition (ICPR’08), 2008. [10] H. Peng, F. Long and C. Ding, ”Feature selection based on mutual infor- mation criteria of max-dependency, max-relevance, and min-redundancy,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 27, no. 8, pp. 1226-1238, August 2008. [11] L. Yu, and H. Liu, ”Efficient feature selection via analysis of relevance and redundancy,” Journal of Machine Learning Research, vol. 5, pp. 1205- 1224, Oct. 2004. [12] G. Qu, S. Hariri and M. Yousif, ”A new dependency and correlation analysis for features,” IEEE Trans. Knowledge and Data Engineering, vol. 17, no. 9, pp. 1199-1207, Sep. 2005. [13] A. Jakulin and I. Bratko, ”Analyzing attribute dependencies,” Proc. PKDD 2003, Lecture Notes in Artificial Intelligence, pp. 229-240, 2003. [14] Z. Zhao and H. Liu, ”Searching for interacting features,” Proc. Intl. Joint Conf. on Artificial Intelligence (IJCAI-07), pp 1156-1161, 2007. [15] T.M. Cover and J.A. Thomas, Elements of information theory, John Wiley & Sons, 1991. [16] A.M. Fraser and H.L. Swinney, ”Independent coordinates for strange attractors from mutual information,” Physical Review A, vol. 33, pp. 1134- 1140, 1986. 23 [17] G.A. Darbellay and I. Vajda, ”Estimation of the mutual information by an adaptive partitioning of the observation space,” IEEE Trans. Informa- tion Theory, vol. 45, pp. 1315-1321, 1999. [18] N. Kwak and C-H. Choi, ”Input feature selection by mutual information based on Parzen window,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 24, no. 12, pp. 1667-1671, 2002. 24 
work_zxm75rfbzvafvmbxjk5izqpfie	----	6-د.ثائر.doc ٢٠١٢ لسنة ٣٤ مجلد ١٠٧ تنمية الرافدين العدد جامعة الموصل-كلية اإلدارة واالقتصاد ]١٢٥-١٠٣[ص ص ٢٣/٢/٢٠١١ تأريخ قبول النشر ١٧/١٠/٢٠١٠ تأريخ استالم البحث صالحیة المنتج الدوائيتصمیم نظام خبیر لتحدید *في الشركة العامة لصناعة األدویة والمستلزمات الطبیة في الموصل ّحمد سعدون السمانأالدكتور ثائر قسم نظم املعلومات اإلدارية-أستاذ مساعد جامعة املوصل-كلية اإلدارة واالقتصاد thaeir_alsamman@yahoo.com الصفونور ضياء عزيز باحث علمي نظم معلومات إدارية المستخلص ٍعلى نحومن األنظمة المتطورة مقارنة مع األنظمة التقلیدیة العتمادھا )النظام الخبیر(یعد المعرفة والخبرات، لذا یجب أن یتم االعتماد علیھ بصورة أكثر مما ھو موجود اآلن ف ي ىكبیر عل ساسي الذي یمارس ھ ف ي إط ار تحدی د واختی ار ن وع اإلس تراتیجیة الواقع ومعرفة الدور الكبیر واأل . لدیھاةالتي یجب على الشركة اعتمادھا في إطار المعارف والخبرات الداخلیة المتوفر م ع لغ ة Front Page وباستخدام البرمجیة الج اھزة HTMLبلغة تم تصمیم نظام خبیر Java Script الحب وب لتقلی ل الوق ت والجھ د الم ستغرق للفحوص ات المختبری ة الخاص ة بأدوی ة للتأكد من نتائج الفحوصات الخاصة بكل نوع من األدویة والتي في ضوئھا یتم اتخاذ ق رار ب رفض وتظھر أھمیة ھذا النظام في كونھ ی وفر س ھولة وس رعة ف ي التعام ل م ع .وجبة الدواء أو قبولھا ًصھا ف ي المختب ر ف ضال ع ن تن وع نت ائج الفحوص ات ب سبب كث رة أن واع األدوی ة المطل وب فح الفحوصات الخاصة بكل نوع من األدوی ة وتق ارب الم دیات المح ددة لك ل فح ص مم ا یجع ل عملی ة حیح س جل خ اص م دون فی ھ الم دى ال صإل ىالتأكد م ن نتیج ة الفح ص ص عبة وض رورة الرج وع .للفحوصات لكل نوع من األدویة عل ى االس تنتاجات الت ي ت م ًلمھمة وبن اء ا مجموعة من االستنتاجاتإلىوتوصلت الدراسة . عدد من المقترحاتإلىالباحثان انتھى التوصل إلیھا في ھذه الدراسة ، نظام التشفیر، محرك االستدالل، قاعدة المعرفة النظام الخبیر:ةلمفتاحیاالكلمات مقت رح ت صمیم نظ ام خبی ر "اء عزی ز فتح ي الموس ومة البحث مستل من رسالة الماجستیر للطالبة نور ضی * لدعم استراتیجیات إدارة المعرفة بالتطبیق على الشركة العام ة ل صناعة األدوی ة والم ستلزمات الطبی ة ف ي .٢٠١٠ نینوى المقدمة إلى كلیة اإلدارة واالقتصاد جامعة الموصل، ]١٠٤ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف Expert System Design to Determine Reject or Accept Drug Production in the State Company for Drug Industries and Medical Appliances in Mosul Thaeir A. Al Samman (PhD) Assistant Professor Department of Management Information Systems University of Mosul thaeir_alsamman@yahoo.com Noor Dh. Al Safoo Researcher Management Information Systems University of Mosul Abstract The expert system is regarded one of the sophisticated systems, since it highly depends on knowledge and expertise. So, it should be adapted to an extent greater than it exists nowadays and know the magnificent and essential role. This practice should also be known to determine and select the type of strategy adopted by the company within the frame of knowledge and internal expertise. An expert system of tablet laboratory tests has been designed in html using programming package Front Page and Java Script to decrease the time and efforts to make sure of the results of the test of each drug type. On the light of that test, a decision will be made to reject or accept the drug production. The importance of these programs is to facilitate and fast treatment with laboratory test, because of many inspections to each drug and the same range of each test may cause the test result very difficult and necessary to return to the record. This contains the right range of inspection of each drug. The study reached to number of the important conclusions; the researchers submitted a number of suggestions that go with these conclusions. Key words: Expert System, Knowledge base, Interface, Engine Explanation System . البحثمشكلة تعد األنظمة الخبیرة من األنظمة الحدیثة والمتطورة والمھمة في الوقت ذات ھ إذا م ا ت م فاع ل، س یمتد تأثیرھ ا عل ى ب اقي عملی ات ٍعل ى نح واعتمادھا من إدارة الشركة وت م تنفی ذھا وم ن جھ ة أخ رى. ومن ثم البق اء ف ي ع الم المناف سة،الشركة وأنشطتھا باتجاه تحقیق النجاح ًفإن موضوع إستراتیجیات إدارة المعرفة للشركة یع د م ن الموض وعات الحدیث ة ن سبیا ال ذي كن الشركة من اللحاق بال شركات المتقدم ة والتف وق علیھ ا عب ر تحقی ق التمیّ ّ ع ن طری قز یم .المعرفة الضمنیة والصریحة التي تمتلكھا الشركة ًوتأسیسا على ما تقدم ونظرا لعدم احت واء ھ ذین الم تغی رین بدراس ة متكامل ة ف ي البیئ ة ً انتحف ز الباحث ) حثینف ي ح دود اط الع الب ا(العراقی ة بعام ة وبیئ ة محافظ ة نین وى بخاص ة م ن خ الل م ا س بق إن ھ تب ینإل ىلتن اول ھ ذا الموض وع ف ي دراس تھا الحالی ة، وی ضاف م ن ةللم دزی ارات المیدانی ة لل شركة العام ة ل صناعة األدوی ة والم ستلزمات الطبی ة ال محدودی ة معرف ة الم دراء ف ي ھ ذه ال شركة بمتطلب ات ٢٧/١٢/٢٠٠٩ إل ى ٢٠/١٢/٢٠٠٩ تن اول ھ ذه الم شكلة وال سعي لعالجھ ا ف ي إل ى اناألم ر ال ذي دف ع الباحث ) النظ ام الخبی ر( :إثارة السؤال األتيب عام یمكن التعرف على مضامین المشكلةٍعلى نحو الحالیة، ومدراستھ ؟نظام خبیر للفحوصات المختبریة في الشركة المبحوثةھل باإلمكان تصمیم mailto:thaeir_alsamman@yahoo.com ]١٠٥[ّالسمان والصفو البحث أھمیة ف ي ) النظ ام الخبی ر(ت أتي أھمی ة الدراس ة م ن كونھ ا ت سلط ال ضوء عل ى متطلب ات .الشركة قید الدراسة البحثأھداف :يتآل باًانطالقا من أھمیة الدراسة وتساؤالتھا، یتحدد الھدف الرئیس للدراسة ).المفھوم واألھمیة: (من حیث) النظام الخبیر(یة التعرف على ماھ. ١ ت صمیم نظ ام خبی ر للفحوص ات المختبری ة للحب وب لل شركة العام ة ل صناعة األدوی ة . ٢ .والمستلزمات الطبیة ًتق دیم مجموع ة م ن المقترح ات اعتم ادا عل ى النت ائج واالس تنتاجات الت ي سیتوص ل إلیھ ا . ٣ .الباحثان البحثفرضیات ف ي تقلی ل ھم الحب وب ی سبأدوی ةص ة ا الخظام خبیر للفحوصات المختبری ةن تصمیم نإ الخاص ة بك ل ن وع م ن وی ساعد ف ي تحدی د نت ائج الفحوص ات المختبری ة ،الوق ت والجھ د .األدویة والتي في ضوئھا یتم اتخاذ قرار قبول أو رفض وجبة الدواء ماھیة النظم الخبیرة مفھوم النظام الخبیر -ًأوال صیغة ش ائعة لل نظم المعتم دة عل ى المعرف ة، ولك ن ھن اك إش كالیة ) نظام الخبیرال(یعد وضع مسمى خالص وخاص لھ، فعلى مدار السنوات الماض یة ظھ رت العدی د م ن الت سمیات نظم الخب رة أو ال نظم ب جان ب ال نظم الخبی رة، إلىطلق علیھا الخاصة بھذه النظم، فھناك من ی أن ال نظم الخبی رة إل ىرغ م إش ارة العدی د م ن الكت اب (معرف ة المبنیة على المعرفة أو نظ م ال ، وف ي العربی ة ق د یطل ق علیھ ا أی ضا ال نظم )واحدة من تطبیقات ال نظم المبنی ة عل ى المعرف ة ال نظم ( ھن ا اس تخدام م صطلح انف ضل الباحث یة والناص ح اآلل ي والمست شار اآلل ي، وَن ٍطَفال والكت ب والمؤلف ات العلمی ة عل ى اس تخدامھ ً، نظ را الس تقرار أغل ب األبح اث )الخبی رة .وسھولتھ وت سمى ال نظم الخبی رة بأنظم ة قواع د المعرف ة وق د أورد الب احثون تع اریف لھ ا تتف ق :معظمھا من حیث المضمون والتعبیر ومنھا ما یأتي األنظمة الخبیرة بأنھا األنظمة الم ستندة ) ٦٥ ، ١٩٩٦عوید وعلوان، : (عرف كل من وذل ك ألداء ،عرفة الخاصة بھا الت ي ت م ت صمیمھا وبناؤھ ا باس تخدام الحاس وب قواعد المإلى . قاعدة المعرفةإلىًوظیفة واحدة استنادا بأن ھ ن وع م ن ) النظ ام الخبی ر(على تعریف ) Howlett and Barton(وقد اتفق كل من صول یحتوي عل ى كمی ة كبی رة م ن المعرف ة المتعلق ة بمج ال مح دد ی تم الح برامج الحاسوب التي یمكنھا أن ترشد وتحلل وتصمم وتفح ص علیھا من خبیر واحد أو أكثر في ذلك المجال، وتشرح وتتنبأ وتتصور وتعرف وتف سر وتح دد وت تعلم وتحف ظ وتق دم وتج دد وتختب ر وتعل م، (Barton, 1987, 2) خب راء لحلھ ا إل ى الت ي تحت اج توھ ي ت ستخدم ف ي ح ل الم شكال .)Howlett, 1988, 23(و ]١٠٦ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف ھ و ) النظ ام الخبی ر(ن أن فق د مھ د الطری ق لألنظم ة الخبی رة، وب یFeigenbaumّا أم برنامج حاسبة ذك ي ی ستخدم المعرف ة وإج راءات االس تقراء لح ل م سائل یتطل ب حلھ ا خب رة أن أكبر مشكلة في تصمیم النظ ام الخبی ر وبنائ ھ ھ ي طریق ة اقتب اس )فایجنبام(ویعد . بشریة محم د، ( فع ال ف ي قاع دة بیان ات المعرف ة عل ى نح ویاغتھا معلوم ات الخبی ر الب شري وص ١٥، ١٩٨٨.( ١الشكل المفھوم األساسي لوظیفة النظام الخبیر Source: Noran, Ovidiu, 2001, The Evolution of Expert Systems, Griffith University, School of Computing and Information Technology, hhttttpp::////wwwwww..hhoobbbbiitt..iicctt..ggrriiffffiitthh..eedduu..aauu//~~nnoorraann//DDooccss//EESS--EEvvoolluuttiioonn..ppddff, P.2. بالحق ائق الخاص ة بالم شكلة ) النظ ام الخبی ر( نلح ظ أن الم ستفید ی زود ١وم ن ال شكل یق وم مح رك االس تدالل بالبح ث ع ن القاع دة المالئم ة ف ي قاع دة ف ي ح ینالجتھ ا، الم راد مع .المعرفة وتزوید المستفید بالخبرة بأن ھ ذل ك النظ ام ال ذي یمتل ك ) النظام الخبیر) (٢٩٩، ١٩٩٢البیاتي، (في حین عرف عط اء القدرة على محاكاة أسلوب وقابلیات اإلنسان الخبیر والمختص ف ي تق دیم االست شارة وإ عل ى النصائح واإلجابة الذكیة على بعض األسئلة التي تطرح في إحدى مجاالت التخصص و ی وازي اإلجاب ة الت ي یمك ن الح صول علیھ ا م ن اإلن سان الخبی ر ف ي می دان التخ صص نح و .نفسھ وقد عرفت األنظمة الخبیرة بأنھا ھیكلیة خاصة بلغة معین ة تحت وي عل ى خب رة كثی رة إل ى ولھا القابلیة على االستدالل واالستنتاج في التوصل ،االت المعرفةًجدا في مجال من مج ).٢٠، ١٩٩٣عبدالرحمن، ( دون أن یكون ھناك طرائق مباشرة وواضحة للحل منالحل كما تم تعری ف النظ ام الخبی ر بأن ھ برن امج ی ستخدم الخب رة المعرفی ة لیحق ق م ستویات ).١٦٧، ١٩٩٧خووبو، (عالیة من األداء في نطاق المشاكل الضیقة النظ ام ) Baracskai, 2008, 2(و) ١٥، ١٩٩٩المطلب ي، (ف ي ح ین ع رف ك ل م ن الخبیر بأنھ وسیلة لتحصیل معارف الخب راء عل ى ش كل ب رامج وبیان ات، إذ تح سم الخالف ات ب ین الخب راء ع ن طری ق التوس ط ب ین الخب راء الس تخالص النت ائج بطریق ة تمك ن م ن قاعدة المعرفة محرك االستدالل المستفید ِالحقائق الخبرة الخبیرالنظام ]١٠٧[ّالسمان والصفو واس تخدامھا م ن األش خاص األق ل خب رة ف ي إط ار حق ل مع ین اس تخالص ج وھر المعرف ة .للمعرفة برن امج النظ ام الخبی ر بأن ھ ) Laudon and Laudon, 2000, 145(بینما عرف ك ل م ن القواع د ، تل ك (IF-THEN) م ن قواع د ً ومت داخالًامترابط ً ا كبی رًا ع ددذك اء ص ناعي یمتل ك .لمعرفة في النظامتمثل ا النظ ام الخبی ر بأن ھ برن امج حاس وبي یحت وي عل ى ) ٦ ، ٢٠٠١ب رایس، (وق د ع رف ، والبدیھ ة Rules، وقواع د االس تنتاج Judgment، والحك م Expertiseخب رة اإلن سان Intuitionوخبرات أخرى لتقدیم نصائح وحلول في تخصص أو مجال معین ،. الب شري أن النظ ام الخبی ر ھ و محاول ة لمحاك اة المنط ق إلى )O'Brien, 2003(وأشار .وحل المشاكل كما یفعل الخبیر البشري ف ي تعریف ھ للنظ ام الخبی ر بأن ھ برمج ة عملی ات الحك م ) ٩٤، ٢٠٠٦ح سین، (ووج د والتقدیر التي یقوم بھا الخبیر البشري في برن امج حاس وب التخ اذ الق رارات بطریق ة التفكی ر .نفسھا التي یطبقھا الخبیر البشري ق دم یاص طناعي ذك اءنظ ام النظ ام الخبی ر بأن ھ ) Haag, 2007, 190(ف ي ح ین ع رف اإلجاب ة ع نب ل تطالت ي تلم شاكل ا لحلًا االستنتاج، ویكون مناسبإلىلوصول لر یتفكقدرات الم شاكل التقادمی ة ف ي عملی ة ص نع الق رار، و ال ذكاءویقاب ل مرحل ة "؟ط أخ الم ا"س ؤال، Prescriptive ختی ار ف ي عملی ة قابل مرحل ة االتو"ما العمل؟ " سؤال اإلجابة عنتتطلب التي .ر القرااتخاذ : ما یأتيانستنتج الباحثی على ما سبق ًوبناء ویتك ون م ن ع دد م ن ،ول ھ كی ان متكام ل ،ان النظ ام الخبی ر ھ و نظ ام معلوم ات .العناصر ویك ون ف ي مج ال ،ویعم ل عل ى ح ل الم شكالت. وإنھ أحد فروع الذكاء االص طناعي ویمك ن . ا الت ي یعم ل بھ ا الخب راء الب شریونع الج بالطریق ة نف سھوی. معرفي محدد أو ض یق .ًاستخدامھ بوصفھ مساعدا أو زمیل عمل على مستوى الخبراء مكونات النظام الخبیر ً-ثانیا :تشمل مكونات النظام الخبیر ما یأتي Knowledge Baseقاعدة المعرفة . ١ ال الخبرة والمعرف ة المعین ة كم ا ، وتجسد قاعدة المعرفة مج)النظام الخبیر(وھي قلب ویمك ن الح صول عل ى ھ ذه . ت م استخالص ھا م ن الخب راء المتخص صین ف ي ھ ذا المج ال فالبعض منھا ی تم الح صول علی ھ م ن الكت ب الدراس یة وال دوریات ،المعرفة من مصادر عدة ,Moghram and Ibrahim)العلمی ة والن شرات الرس میة الت ي ت صدرھا الجھ ات الحكومی ة والبعض اآلخر یتم الحصول علیھ من المناقشات مع الخبراء في كیفیة التعام ل . (359 ,2000 ویمكن الحصول عل ى ھ ذا الن وع م ن المعرف ة إم ا . التي سیتعامل معھا النظامتمع المشكال بالمشاھدة والمالحظة لخبیر أثناء حل مشكلة فعلیة أو من خالل الوصف الذي یقدمھ الخبی ر، رج ب، ( ك ل خط وة یتخ ذھا ف ي ح ل الم شكلة ً واص فاٍن یفك ر ب صوت ع الإذ یطل ب من ھ أ یتطلب األمر تحدید كیفی ة ،وبعد الحصول على المعرفة المالئمة للنظام الخبیر). ٥، ٢٠٠٤ ع رض (ھ و م ا یع رف باس م وتخ زین ھ ذه المعرف ة وكیفی ة اس ترجاعھا وق ت الحاج ة إلیھ ا ]١٠٨ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف وذج ـنم ی ل المعرف ة ف ي مج ال األعم ال ھ و األوم ن أھ م النم اذج الم ستخدمة لتمث). المعرف ة ).٩٥، ٢٠٠٦حسین، ) ( إذن–إذا (المبني على القواعد التي تأخذ شكل وتع رف قاع دة المعرف ة بأنھ ا مجموع ة المع ارف المتعلق ة بحال ة معین ة الت ي ی تم ث م یب دأ النظ ام باس تخدامھا ) إذن–إذا (استخالص ھا م ن الخب راء وتخ زن عل ى ش كل قواع د )Laudon and Laudon, 2000, 447.( (O'Brien, 1997, 311): ن أساسیین ھماوتشمل قاعدة المعرفة عنصری .الحقائق عن مجال موضوع معین، مثل موقع المشكلة. أ . اجتھادات أو قواعد خاصة تعبر عن إجراءات التفكیر للخبیر.ب Working Memoryالذاكرة العاملة . ٢ لى الحقائق الخاص ة بالم شكلة موض وع البح ث، فعن دما یق وم تحتوي الذاكرة العاملة ع وإدخ ال المعلوم ات ح ول الم شكلة یت ولى ) النظ ام الخبی ر(المستفید أو صانع القرار استشارة النظ ام عملی ة مقارن ة ومقارب ة ھ ذه المعلوم ات بالمعرف ة الت ي یحتویھ ا النظ ام ف ي قاع دة .)Wozniak, 2001, 17(المعرفة الستنتاج حقائق جدیدة Inference Engineمحرك االستدالل . ٣ ، ویع رف بأن ھ ھیك ل ال سیطرة لمت رجم )النظ ام الخبی ر(یع د مح رك االس تدالل عق ل قاعدة المعرفة، ویعد ھذا المك ون األس اس لبرن امج الحاس وب، ی زود بمن اھج إلع ادة التفكی ر Moghram): شمل ما یأتيوصیاغة االستنتاجات، أما الوظائف الرئیسة لمحرك االستدالل فت and Ibrahim, 2000 ,2) .تحدید القواعد التي سوف تستخدم • .تقدیم سؤال للمستخدم عند الحاجة • . ذاكرة النظام الخبیرإلىإضافة األجوبة • .یستدل على الحقائق الجدیدة من القاعدة • . الذاكرةإلىإضافة الحقیقة • إس تراتیجیة باس تخدام س تنتاجاتال ع ن البح ثأو ار ی تفكلل ومحرك االستدالل ھ و آلی ة یقب ل وواجھ ة الم ستفید، إذق ع ب ین قاع دة المعرف ة یو. معرف ةالقاع دة المعرف ة ف يعن بحث ).Abraham, 2005, 911(االستنتاج ویحاول من المستفید دخالتالم إن عم ل مح رك االس تدالل ف ي فح ص القواع د ی تم بموج ب ص یغتین ھم ا التعاق ب .لخلفياالتعاقب األمامي و ی تم تفح ص القواع د ف ي ترتی ب مع ین Forward Chaining التعاق ب األم اميفف ي الواح دة تل و األخ رى ح سب التت ابع ال ذي أدخل ت فی ھ القواع د أو ح سب م ا یح دده الم ستفید، ً م ا إذا ك ان ال شرط ص حیحا أو خط أ م ع فح ص ك ل قاع دة م ن تقی یمیح اول النظ ام الخبی ر الفحص حتى یكتمل مسار كامل خالل فئة القواعد كلھ ا، ویمك ن القواعد وھكذا تستمر عملیة .)Lenahum, et.al., 2000, 4(أن یكون أكثر من مسار واحد لتحدید قیمة المتغیر فتق وم آل ة االس تدالل باختی ار Backward Reasoning حال ة التعاق ب الخلف يأم ا ف ي ث م تتن اول آل ة . یم ة لمتغی ر الھ دف الم شكلة الم راد حلھ ا، وب ذلك یتح دد قإل ىقاع دة وتن سبھا آل ة اس تدالل نموذجی ة داالستدالل مشكلة فرعیة أخرى، ومما تجدر اإلش ارة إلی ھ أن ھ ال توج ]١٠٩[ّالسمان والصفو لمج ال مح دد للتطبی ق ش رط ام تالك ًأو ش املة لك ل أن واع األنظم ة الخبی رة ب ل ت صمم وفق ا ).Alter, 2002, 209( المراد حلھا تالمعرفة ضمن المجال نفسھ وحسب خواص المشكال والفرق بین التعاقب لإلمام والتعاقب للخلف یكمن في أن التعاقب للخل ف یك ون أس رع م ن التعاق ب األم امي نظ را ألن ھ ال یع د ك ل القواع د وال یج ري م سارات متع ددة خ الل فئ ة ، ٢٠٠٠مكلی ود، ( ب صفة خاص ة ف ي الح االت اآلتی ة ًویكون التعاقب الخلف ي مناس با. القواعد ٦٤٣(: .ندما توجد متغیرات متعددةع • .عندما توجد الكثیر من القواعد • . حلإلى فحص في عملیة الوصول إلى القواعد معظمھاال تحتاج كل أو • Explanation Systemنظام التفسیر . ٤ ْم ن ال ضروري أن تك ون األنظم ة الخبی رة ق ادرة عل ى تزوی د التف سیرات بخ صوص َ د ،الت ي ت م التوص ل إلیھ ا االس تنتاجات ّ م ن الم ستفید تمك نآلی ةبأكث ر األنظم ة الخبی رة ت زو :حولتوجیھ أسئلة ًلماذا نسأل سؤاال معینا • ؟ً ؟ن معیستنتاج اإلى النظام كیف یصل • ی ف یعم ل لفھم كخدم للمستالبدھیةضروریة في كل المجاالت غیر الد التفسیرات ویزت ) معرف ةال(حفظ م سارات القواع د س یالنظ امف. الم أًاره ص حیحیتفكأكان النظام ویقرر سواء اللغ ة اإلنجلیزی ةإل ى ترجم ة ھ ذه القواع د إل ى م ستندة ٍ وی زود تف سیراتالم ستخدمة )SAID,2009 ,52(. User Interfaceواجھة المستفید . ٥ Natural) لغ ة تخاط ب طبیعی ة ع ن طری قیتم التفاع ل ب ین النظ ام الخبی ر والم ستفید Language)الحل المطلوب للمشكلة من حی ث آلی ة إلىًر المبسط وصوال تعتمد أسلوب الحوا االستنتاج والتفسیر واإلیضاح والمرونة المطلوبة إلضافة معارف جدی دة أو إج راء تع دیالت ، وم ن خ الل ھ ذا الح وار ی تمكن )Eronen and Zitting, 2000, 4(أثن اء تط ور ح ل الم شكلة اإلجاب ات ع ن الم شكلة المطروح ة، وف ي المستفید م ن ط رح سل سلة م ن األس ئلة لك ي یتلق ى الواقع تصمم الواجھات البینیة لألنظمة الخبیرة على أس اس تلبی ة احتیاج ات الم ستفید النھ ائي .)النظام الخبیر(یوضح مكونات ) ٢(والشكل .)١٩٦، ٢٠٠٦یاسین، ( الحل والتفسیرات األوامر المعرفة المستخدم واجهة محرك االستدالل محرك التطویر النظام الخبیر قاعدة المعرفة مجال المشكلة ]١١٠ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف ٢الشكل معماریة النظام الخبیر ، تعری ب س رور عل ي، دار الم ریخ للن شر، نظ م المعلوم ات اإلداری ة، ٢٠٠٠یمون د، مكلی ود، را:الم صدر .٦٣٤السعودیة، ص م ن خ الل واجھ ة ) النظ ام الخبی ر( یتم مالحظة أن المستفید یتعامل م ع ٢ومن الشكل الم ستفید الت ي ت رتبط ب دورھا بمح رك االس تدالل بعالق ة م ن خ الل اتج اھین لتتلق ى حق ائق ستفید وی رتبط مح رك االس تدالل ب دوره بقاع دة المعرف ة للبح ث ع ن القاع دة إض افیة م ن الم المالئم ة لمج ال الم شكلة، أم ا مح رك التط ویر فی رتبط بالعناص ر الموج ودة خ ارج النظ ام الخبی ر والمتمثل ة ب الخبیر الب شري ومھن دس المعرف ة إلج راء أي تغیی ر أو تط ویر للنظ ام .الخبیر فق د )Engelmor and Feigehbaum, 1993, 8( م ن وك ال) Haag, 2004, 197(أم ا مث ل أي نظ ام معلوم ات یتك ون م ن المعلوم ات، واألف راد، ) النظ ام الخبی ر(ذك روا أن . المعلوماتاومكونات تكنولوجی مراحل بناء النظام الخبیر ً-ثالثا ، ١٩٩٣عب دالرحمن، (ألج ل بن اء نظ ام خبی ر متكام ل یج ب دراس ة المراح ل اآلتی ة :)١٦٦، ٢٠٠٩الحمیدي، ) (Ismail, et. al., 2009, 125) (٩، ٢٠٠١صقر، ()٢٧ Knowledge Acquisitionمرحلة استخالص المعرفة . ١ ف ي المج ال ) الخب رة أو المعرف ة (كبر قدر من المعلوماتأ بجمع المعرفةیقوم مھندس ون ب شرح المح دد م ن حی ث مجال سة ع دد م ن الخب راء أو الب احثین ف ي المج ال، إذ یقوم ًق الت ي یتبعونھ ا ف ي ح ل الم سائل، ف ضال ع ن التعری ف ائ مع رفتھم ف ي ھ ذا المج ال والطر ]١١١[ّالسمان والصفو من أجل أن یخرج البرنامج الذكي ب صورتھ النھائی ة، وأن ) للنظام الخبیر(بمتطلبات المعرفة .ًیكون محققا لألھداف التي صمم من أجلھا Knowledge Representationمرحلة تمثیل المعرفة . ٢ مھن دس المعرف ة باختی ار األس لوب المناس ب لتمثی ل المعرف ةف ي ھ ذه المرحل ة یق وم ت صمیم إنتاج إلىھذه المرحلة ھدف ومجال تخصصھ، وت) النظام الخبیر(حسب طبیعة عمل ًمن حیث كیفیة تنظیم المعرفة منطقیا في قاعدة المعرفة، وكیف س یتم )خبیر النظاملل(مفصل .توثیق التصمیمًتنفیذ النظام فضال عن Design Interfacingمرحلة تصمیم األوساط البینیة . ٣ أول ي، الت ي ٍعلى نحوفي ھذه المرحلة یبنى محرك االستدالل وتصمم واجھة المستفید تج رى علیھ ا تع دیالت بع د اس تالم التغذی ة العك سیة م ن الم ستفید، وی تم وض ع المعرف ة بع د .ًلنظام جاھزا للعملتمثیلھا بإحدى طرائق التمثیل لیصبح ا Knowledge Updatingمرحلة تحدیث المعرفة . ٤ المتول د القابلی ة عل ى تح دیث معرفت ھ م ن الم ستفید ) للنظ ام الخبی ر(یج ب أن یك ون .ًلغرض جعل النظام بوضع أمثل دائما، وھذه الخاصیة ال نراھا في البرامج التقلیدیة System Evaluationمرحلة تقییم النظام . ٥ لتح سینات إلج راء اتوص یاتت م التوص ل إلی ھ م ن خ ص م ا له المرحل ة تھ ذ : اآلتي١ ، ویوضحھا الجدولوالتصحیحات ١الجدول مھام مرحلة التقییم الھدف المھمة تقییم النتائج التوصیات "التصدیق"التأكید التقریر المؤقت أو النھائي .تلخیص نتیجة االختبار والتحقق .تغییر في النظامالتوصیات إلجراء أي . ومتطلباتھ بأن النظام صحیح فیما یتعلق بحاجة المستفیدالتأكید ًإذا كان النظام كامال، إذن سیكون التقریر نھائیا وبعكسھ فسیكون التقری ر ً ً.مؤقتا تطبی ق نظ م الخب رة ف ي ، ٢٠٠١، إب راھیم أحمد محمود صقر، محمد: باالعتماد علىان إعداد الباحث:المصدر ، http://www.iscal.org.sa/files/proceedings.pdfلوم الشرعیة باللغة العربیة، جامعة أم الق رى، الع .١٠ص . استخدام النظام الخبیر وتطویره٣الشكل یوضح واجهة المستفید تسلم حقائق المشكلة إلى النظام الخبیر وترسل االستنتاج بیر المجال ومهندس المعرفةخ المستفید خبیر المجال یزود خبرة المجال إلى مھندس المعرفة مهندس المعرفة یحول مھندس المعرفة خبرة المجال إلى قواعد ]١١٢ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف ٣الشكل م النظام الخبیرتطویر واستخدا Source: Hagg, Stephen, et.al., 2004, Management Information System, 4th ed., McGraw Hill, New York, P.198. تصمیم النظام الخبیر في الشركة العامة لصناعة األدویة والمستلزمات الطبیة ةآلیة العمل في الشركة للفحوصات المختبریة للحبوب المنتج -ًأوال راس ات والبیان ات مقابل ة م ع م سؤول ش عبة الد ت م إج راءقب ل ت صمیم النظ ام الخبی ر عرض الفكرة األولیة للنظام ومناق شة فائدت ھ لل شركة وت سھیل العم ل ب ان الباحثالعلمیة، إذ قام ّالفیھا، وبعد حصول الموافقة على فكرة النظام الخبیر، تم ش عبة ال سیطرة النوعی ة إلىل تحو .ةلحصول على البیانات الالزمالشاملة ل مقابالت عدة مع مدیرة قسم السیطرة النوعی ة، وجھ ت خاللھ ا أس ئلة ع دة یت وقد أجر خ اص، وم ا ھ ي ٍعل ى نح و عام وبمختب ر الحب وب ٍعلى نحوتتعلق بطبیعة العمل في الشعبة تق وم آلی ة العم ل ف ي المختب ر وكیفی ة التعام ل م ع نت ائج الفحوص ات، وأن واع األدوی ة الت ي الشعبة بفحصھا وإعطاء النتیج ة، وق د تب ین أن ش عبة ال سیطرة النوعی ة م سؤولة ع ن إج راء وذل ك للتأك د م ن مجموعة الفحوصات الالزمة على األدویة كافة التي تقوم الشركة بإنتاجھا، ً، ف ضال )PB 2007(مطابقتھا للمواصفات القیاسیة لألدویة ح سب دس تور األدوی ة البریط اني ، وق د ومن ثم إعطاء الموافقة بصالحیتھا لالستخدام البشري) SDI( أدویة سامراء عن تحلیل ]١١٣[ّالسمان والصفو اطل ع الباحث ان عل ى المختب رات الت ي ی تم فیھ ا فح ص األدوی ة وعل ى األجھ زة الم ستخدمة الدوام الرسمي في ال ساعة ین في المختبر لمدة ثالثة أیام ندءاألفراد العاملوتحت مقابلة لذلك، ًباحا وحتى الثانیة والنصف ظھ را للح صول عل ى فك رة أوض ح ع ن آلی ة الثامنة والنصف ص ً ِلحظ ُ العمل، وقد بعد انتھاء األفراد العاملین من جمیع الفحوص ات الخاص ة ب دواء مع ین نھ أُ ی تم مراجع ة ال سجل الخ اص بالم دى المح دد لك ل فح ص، وم ن ث م إعط اء النتیج ة النھائی ة ساعتین، وبعد ذلك ی تم إعط اء نتیج ة الفح ص إلى ساعة بینتتقارب مدة إنجاز ھذه العملیة و والمخطط اآلتي یبین آلیة العم ل . قسم اإلنتاجإلىًشفھیا عن طریق الھاتف ) رفض/ موافقة ( ف ي ال شركة العام ة ل صناعة األدوی ة والم ستلزمات الطبی ة وموق ع ش عبة ال سیطرة النوعی ة .الشاملة في ذلك ٤الشكل العمل في الشركة العامة لصناعة األدویة والمستلزمات الطبیةآلیة . باالعتماد على المعلومات التي تم الحصول علیھا من الشركةناإعداد الباحثمن المخطط :المصدر .٥أما الفحوصات التي تتم على أدویة الحبوب بأنواعھا المختلفة یوضحھا الشكل الفحص فحص كیمیائي فحص فیزوكیمیائي فحص فیزیائي كبستحض تحلیل )المختبر( لقبو رفض إكساء تحلیل تعبئة إطالق قبول رفض الوزن - اللون - الذوبانیة - السمك - القطر - الھشاشیة - الصالبة - ]١١٤ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف لیةصیدالنیة غیر صیدالنیة فحص المادة الفعالة االنحال ٥ الشكل الفحوصات الخاصة بأدویة الحبوب تم الحصول علیھا من شعبة السیطرة باالعتماد على المعلومات التيانالشكل إعداد الباحث: المصدر .النوعیة الفحوصات الفیزیائیة. ١ وتضم مجموعة من الفحوصات للتأكد من الخواص الفیزیائیة الخاصة بأدویة الحب وب :بـوھذه الفحوصات تتمثل یجرى الفحص عن طریق أخ ذ عین ة م ن الحب وب وتوض ع عل ى جھ از :Weightالوزن . أ وی تم ت سجیل ال وزن ومقارنت ھ م ع ال وزن المح دد للتأك د م ن أن ،خ اص بھ ذا الغ رض .وزنھا یقع ضمن المدى المحدد ی تم التأك د م ن الل ون الخ اص بك ل ن وع م ن الحب وب وخل و الل ون م ن :Colorالل ون . ب . رفض الوجبة بكاملھاإلى فأي تغییر في اللون یؤدي ،الشوائب ن ة م ن الحب وب داخ ل جھ از خ اص یمث ل ی تم وض ع عی:Destrection Timeالذوبانی ة . ت إذ تغم ر ٣٧یحت وي عل ى م اء ع ادي أو م اء مقط ر بدرج ة ح رارة ) مع دة اإلن سان( الحب وب، وی تم ت سجیل الوق ت الم ستغرق لل ذوبان ال ذي یج ب أن یك ون ض من الم دى .المحدد یتم أخذ عین ة م ن الحب وب وقی اس س مكھا وت سجیل النتیج ة والتأك د :Thicknessالسمك . ث .من مطابقتھا للمواصفات یتم مسك الحبة بوساطة أداة خاص ة ی تم تحریكھ ا لتناس ب قط ر الحب ة :Diameterالقطر . ج .ویتم تسجیل النتیجة والتأكد من مطابقتھا للمواصفات ً ا یتم وضع عینة الحبوب داخ ل جھ از عل ى ش كل حلق ة ت دور طولی :Fribiltyالھشاشیة . ح وی تم ة،التأكد من متانة الحبة وع دم قابلیتھ ا للك سر ب سھوللمدة أربع دقائق، وذلك بھدف :قیاس ھشاشیة الحبة عن طریق إتباع القانون اآلتي )١ ...(١٠٠× = الھشاشیة الوزن بعد فحص الھشاشیة–الوزن قبل فحص الھشاشیة الوزن قبل فحص الھشاشیة ]١١٥[ّالسمان والصفو ام ة ل صناعة ال شركة الع ف ي وناتج ھذه المعادلة یج ب أن یك ون ض من الم دى المح دد االدوی ة والم ستلزمات الطبی ة ف ي نین وى ، ش عبة ال سیطرة النوعی ة المواص فات القیاس یة )PB 2007(دویة البریطاني بحسب دستور األلألدویة وك ل حب ة ی تم وض عھا ، في ھذا الفحص یتم أخذ عینة م ن الحب وب:Hardnessالصالبة . خ ن د تعرض ھا لق وة ض غط مع ین، ث م عج زءینداخل جھاز خاص یقوم بتقسیم الحبة على بك ل حب ة وتق سم عل ى ع دد الحب وب الم ستخدمة، والن اتج ةیتم جمع قوة ال ضغط الخاص .یجب أن یكون ضمن المدى المحدد Chemicalالفحص الكیمیائي . ٢ ح سب خ اص بالم ادة الفعال ة لك ل دواء، وی تم ذل ك ع ن بویحتوي عل ى فح ص واح د و .عینة للتأكد من وقوعھا ضمن المدىطریق إجراء تفاعالت كیمیائیة م Physic. Chemicalالفحص الفیزوكیمیائي . ٣ والھدف منھ معرفة مقدار Degenerative ھو فحص االنحآللیة ًا واحدًاویتضمن فحص ).البراسیتول(الفقدان في المادة الفعالة، ویخضع لھذا الفحص حبوب معینة مثل حبوب وصف النظام الخبیر ً-ثانیا ذي FrontPageباس تخدام البرمجی ة الج اھزة ) HTML(صمیم نظ ام خبی ر بلغ ة ت م ت ة التعام ل م ع وق د وق ع االختی ار علیھم ا وذل ك ل سھولJava Script م ع لغ ة ٢٠٠٣اإلصدار ً م ع المعطی ات الخاص ة بھ ذا النظ ام، ف ضال ع ن Java Scriptم ة لغ ة ھ ذه البرمجی ة ولمالء ي رب ط النظ ام الخبی ر م ع ال شبكة العالمی ة االنترن ت، اس تخدام ھ ذه األدوات ی وفر س ھولة ف یمك ن اس تخدامھ وم ن ث م ، وذلك لكون الفحوصات التي یتعامل معھا النظ ام الخبی ر عالمی ة األدوی ة الت ي ی تم ت صنیعھا ومالءمتھ امن أیة ش ركة أو م صنع لألدوی ة للتأك د م ن ص الحیة نھ یوفر س ھولة وس رعة ف ي التعام ل م ع لالستخدام البشري، وتظھر أھمیة ھذا النظام في كو ًنتائج الفحوصات بسبب كثرة أنواع األدویة المطلوب فحصھا في المختب ر، ف ضال ع ن تن وع الفحوصات الخاصة بكل نوع م ن األدوی ة وتق ارب الم دیات المح ددة لك ل فح ص مم ا یجع ل إل ىً دوم اعملیة التأكد م ن نتیج ة الفح ص ص عبة عل ى األف راد الع املین وض رورة الرج وع كم ا ویظھ ر النظ ام . سجل خاص مدون فیھ المدى الصحیح للفحوصات لكل نوع من األدویة دقی ق م ن ناحی ة األع شار، ل ذلك ٍعلى نحودقة عالیة لكونھ یتعامل مع المدى المحدد للفحص یوض ح نم وذج ٦ًی تم رف ض ال دواء حت ى ل و كان ت نتیج ة الفح ص مقارب ة ج دا، والمخط ط .خبیرتصمیم النظام ال ]١١٦ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف ٦الشكل نموذج تصمیم النظام الخبیرأ .نا إعداد الباحث:المصدر ً نوع ا م ن الحب وب الت ي تق وم ال شركة العام ة ل صناعة األدوی ة ١٦ت م اختی ار والمستلزمات الطبیة بإنتاجھا لتصمیم النظام الخبیر، وذلك لكونھا من أكثر األدویة التي تق وم ف ي تغطی ة الفحوص ات جمیعھ ا انً ف ضال ع ن رغب ة الباحث الشركة بإنتاجھا وبكمیات كبیرة، ]١١٧[ّالسمان والصفو فحص االنحاللیة الذي تخضع لھ أدویة معینة وح سب، الج دول : التي تخضع لھا األدویة مثل . یوضح ھذه األدویة٢ ٢الجدول الحبوب المستخدمة في تصمیم النظام الخبیر اسم الدواء التسلسل ١ Paracetol ٢ Algesik ٣ Librax ٤ Fluout ٥ Aspin Adult ٦ Diazepam ٧ Butadin ٨ Meclodin ٩ Allermine ١٠ Entero-Stop ١١ Metheprim ١٢ Samavit-c ١٣ Coldin ١٤ Gastrigel-D ١٥ Histadin ١٦ Dexon قاع دة معرف ة للنظ ام الخبی ر باالعتم اد عل ى الفحوص ات الخاص ة ١٢٩وقد تم تكوین .كة على الحبوبالتي تجریھا الشر وت ضمن النظ ام واجھت ین، األول ى ھ ي الواجھ ة الرئی سة للنظ ام الت ي تت ضمن عن وان . واجھ ة الفحوص اتإل ىال شركة وال شعبة الت ي ت ستخدم النظ ام، وم ن خاللھ ا ی تم ال دخول . یوضح ھذه الواجھة٧والشكل ]١١٨ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف ٧الشكل واجھة النظام الرئیسة خاللھ ا ی تم إدخ ال نت ائج الفحوص ات الخاص ة بك ل دواء أم ا واجھ ة الفحوص ات فم ن تت ضمن أس ماء الحب وب الت ي من سدلةوآلیة عملھا، وی تم تحدی د ن وع ال دواء م ن خ الل قائم ة .٨تقوم الشركة بإنتاجھا كما یوضحھا الشكل ٨الشكل القائمة المخفیة ألسماء األدویة ]١١٩[ّالسمان والصفو ة ال دواء ك رقم الوجب ة وكمیتھ ا وی تم إدخ ال مجموع ة م ن المعلوم ات الخاص ة بوجب وتأریخ إنتاجھا وتأریخ تحلیلھا وتأریخ استالمھا، وھي معلومات مطلوبة عن د إع داد التقری ر الخاص بقبول أو رفض الوجبة، بعد ذل ك ی تم إدخ ال النت ائج الخاص ة بالفحوص ات الفیزیائی ة الم ستفید كم ا ف ي لت سھیل االختی ار عل ىمخفی ة قائم ة ع ن طری قك اللون ال ذي ی تم تحدی ده .٩الشكل ٩الشكل القائمة المخفیة أللوان األدویة الذوبانی ة، القط ر، ال سمك، الھ شاشیة، وال وزن، ث م ی تم : ی تم إدخ ال نت ائج فحوص ات ًإدخ ال نت ائج الفح ص الكیمی ائي الخ اص بالم ادة الفعال ة، ف ضال ع ن الفح ص الفیزوكیمی ائي .الخاص باالنحآللیة إن وجد یظھ ر الفحوص ات جمیعھ ا ف ي واجھ ة ٍعل ى نح وت م ت صمیم واجھ ة الفحوص ات وق د واحدة وذلك إلظھار حالة الفحوصات جمیعھا أمام المستفید، إذ تظھ ر نتیج ة ك ل فح ص م ن ناحیة مطابقتھ للمواصفات أو ال، على ح دة أم ام الم ستفید مم ا ی سھل عملی ة معرف ة الفح ص .غیر المطابق للمواصفات المحددة إعط اء نتیج ة نھائی ة للفحوص ات جمیعھ ا ف ي حق ل لوح ده، فف ي حال ة نج اح وی تم ء أما في حالة ف شل فح ص واح د أو أكث ر ی تم إعط ا(Pass)الفحوصات جمیعھا تعطى نتیجة ً وبالتالي یقوم النظام الخبیر باتخاذ ق رار بن اء عل ى النتیج ة النھائی ة للفحوص ات (Fail)نتیجة .ین من الحبوب أو قبولھا والسماح بإطالقھابرفض الوجبة الخاصة بنوع مع ًونظ را لظھ ور نت ائج الفحوص ات جمیعھ ا أم ام الم ستفید ی تم تحدی د الق رار المتخ ذ ع ن إمكانی ة إض افة أي تعلی ق أو مالحظ ة عل ى ًف ي حق ل التوص یة، ف ضال) قب ول/ رف ض( ]١٢٠ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف ال سابقة النت ائج الظ اھرة ف ي ھ ذا الحق ل، وی تم طباع ة ال صفحة الت ي تظھ ر المعلوم ات وھ و أم ر خ اص (Reset)، كما یوجد ف ي ھ ذه الواجھ ة أم ر )االختیاریة ونتائج الفحوصات( بتفری غ الحق ول عن د االنتھ اء م ن إدخ ال الفحوص ات جمیعھ ا للب دء بإدخ ال نت ائج فحوص ات . یوضح ھذه الواجھة١٠جدیدة والشكل ١٠الشكل واجھة الفحوصات للنظام الخبیر نتیج ة لمجموع ة م ن الفحوص ات ٣٦٩ النظام الخبیر م ن خ الل إدخ ال وقد تم اختبار االفتراض یة خاص ة بك ل ن وع م ن الحب وب، وق د ت م تحدی د أرب ع مح اوالت لك ل ن وع م ن ، وقد أثبت النظام فعالیتھ من ناحیة الق درة عل ى ٩الحبوب وللفحوصات جمیعھا البالغ عددھا إعط اء نتیج ة ب رفض أو قب ول وجب ة م ن ث موتحدید الفحوصات التي ال تطابق المواصفات، . یوضح ھذه االختبارات٣الحبوب، الجدول ]١٢١[ّالسمان والصفو ٣الجدول نتائج الفحوصات االفتراضیة الختبار النظام الخبیر Ruselt Des Active Diameter Thick Sol1 Hard Frib color weight Test Medicine pass 85 99 12.7 4.3 ١ 9 0.2 White 610 1.Paracetol fail 106 94 12.86 4.7 2 13 0.1 White 500 1.Paracetol pass 95 105 12.9 4.5 4 11 1.4 White 620 1.Paracetol fail 70 107 12.6 4.1 16 12 1.6 Other 670 1.Paracetol pass - 90 12.7 4.2 2 8 0.3 White 630 2.Algesik pass - 100 12.8 4 3 7 0.6 White 600 2.Algesik fail - 89 12.86 3 16 12 1.6 White 660 2.Algesik fail - 111 12.6 4.5 18 6 1.8 White 500 2.Algesik pass - 90 7 2.7 2 2 0.1 Green 102.5 3.Librax pass - 95 7.1 2.6 4 3 0.3 Green 116.5 3.Librax fail - 100 7.1 2.9 6 4 1.5 Green 117.6 3.Librax fail - 105 7 2.5 8 5 0.4 Green 101 3.Librax pass - ١١١ 11 5.6 1 6 1.5 Green 545 4.Fluout pass - 99 12 5.7 6 7 1.3 Green 540 4.Fluout fail - 85 10 5.3 4 9 0.3 Green 530 4.Fluout fail - 115 12.3 5.9 8 4 0.1 Green 549 4.Fluout pass - 92.6 7 2.4 6 4 0.1 Yellow 117.8 5.Diazepam fail - 93 7 2.8 4 3 0.3 Yellow 116 5.Diazepam pass - 100 7.1 2.2 3 5 0.5 Yellow 118 5.Diazepam fail - 90 7 2.1 2 6 1.6 Yellow 120.6 5.Diazepam pass - 95 10.3 2.6 4 5 0.5 Wight 167 6.Aspin Adult pass - 100 10.2 2.7 3 6 1.4 Wight 834 6.Aspin Adult fail - 106 11 209 2 7 1.8 Other 100 6.Aspin Adult fail - 90 12 3 16 8 0.1 Wight 160 6.Aspin Adult pass - 93 7.1 2.2 9 5 0.5 Bink 119 7.Butadin fail - 91 7.1 2.3 10 6 0.6 Bink 116 7.Butadin pass - 105 7 2.8 3 7 1.6 Bink 117.5 7.Butadin fail - 108 7 3 2 4 1.8 Bink 121.8 7.Butadin pass - 90 7 2.2 1 4 0.1 Wight 118 8.Meclodin pass - 95 7.1 2.3 2 6 0.3 Wight 117.5 8.Meclodin fail - 111 8 2.6 3 8 1.5 Wight 123 8.Meclodin fail - 80 7.3 2.1 4 3 1.6 Wight 126 8.Meclodin pass - 92.5 7 2.7 2 3 0.5 Wight 118 9.Allermine fail - 94 7 2.9 1 2 1.9 Wight ١١٥ 9.Allermine pass - 106 7.1 2.6 7 5 1.3 Wight 119.5 9.Allermine fail - 110 7.1 2.4 3 9 1.7 Wight 133 9.Allermine pass - 95 7 2.2 10 4 1.5 Wight 117.5 10.Entero-Stop pass - 97 7 2.3 2 3 1.3 Wight 118 10.Entero-Stop fail - 90 7.7 2.1 17 2 1.8 Other 114 10.Entero-Stop fail - 107 8 2.6 15 6 1.9 Wight 113 10.Entero-Stop ]١٢٢ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف Ruselt Des Active Diameter Thick Sol1 Hard Frib color weight Test Medicine pass - 100 12 4 5 7 0.5 Wight 711 11.Metheprim fail - 75 12.6 3 8 ٦ 0.6 Wight 700 11.Metheprim fail - 111 12.9 4.2 16 3 0.7 Wight 718 11.Metheprim pass - 95 12.4 4.1 1 12 0.8 Wight 713 11.Metheprim fail - 95 12.9 3.6 17 ٧ 1.9 Orange 700 12.Samavit-c fail - 90 12.8 3.80 7 11 1.7 Orange 500 12.Samavit-c pass - 107 12.10 ٤ 6 9 0.5 Orange 650 12.Samavit-c pass - 100 12.13 4.3 3 8 0.7 Orange 660 12.Samavit-c pass - 90 12.7 4 2 7 0.5 Wight 610 13.Coldin fail - 95 12.86 3 4 11 1.6 Wight 680 13.Coldin pass - 91 12.10 4.3 5 8 0.2 Wight 650 13.Coldin fail - 111 12 5 9 5 1.9 Wight 444 13.Coldin pass - 94 7 4.1 1 8 0.5 Wight 117.8 14.Gastrigel-D pass - 99 7 4.2 3 7 0.6 Wight 119 14.Gastrigel-D fail - 89 6 4 5 13 1.5 Other 120.9 14.Gastrigel-D fail - 117 7.1 5 7 12 1.10 Wight 100 14.Gastrigel-D pass - 93 7 2.6 9 3 1.5 Wight 118 15.Histadin fail - 80 8 2.4 10 1 1.8 Wight 114 15.Histadin pass - 100 7.1 2.7 5 5 0.6 Wight 117.5 15.Histadin fail - 115 9 2.10 1 8 0.9 Wight 122 15.Histadin pass - 92 11.9 5.6 2 5 0.6 Yellow 118 16.Dexon pass - 96 12 5.8 3 6 0.8 Yellow 199 16.Dexon fail - 70 10 4 7 4 1.9 Yellow 122 16.Dexon fail - 112 12.8 5.11 9 9 1.5 Other 116 16.Dexon یتضح م ن خ الل نت ائج النظ ام الخبی ر أن ھن اك ق درة كبی رة ف ي تقلی ل الوق ت والجھ د لبی ان ) الفیزیائی ة، الكیمیائی ة، والفیزیوكیمیائی ة: (والتك الیف الستح صال النت ائج المختبری ة ذ الق رارات الخاص ة بقب ول إمكانی ة قب ول دفع ة اإلنت اج م ن عدم ھ وبال سرعة المناس بة التخ ا باإلمك ان ت صمیم "ًالدفعة اإلنتاجیة وبناء علیھ یتم قبول الفرضیة الرئیسة التي تنص على أنھ نظام خبیر ف ي ال شركة العام ة ل صناعة األدوی ة والم ستلزمات الطبی ة لك ي ی ساعد ف ي تقلی ل ". السجالت والتقاریر المرجعیةإلىالوقت والجھد بالرجوع ى ما تقدم من التمیی ز ف ي الح صول عل ى تطبیق ات مناس بة للنظ ام الخبی ر، ًوتأسیسا عل فإن األمر یستلزم مصداقیة ما ذھبنا إلی ھ م ن أن ت وفیر متطلب ات النظ ام الخبی ر سی ساعد ف ي إع ادة تنظ یم ت صمیم النظ ام الخبی ر، وق درة ھ ذا النظ ام عل ى التجدی د واالبتك ار وبم ا ی وفر العامة لصناعة األدوی ة والم ستلزمات الطبی ة ف ي الموص ل للشركات بصورة عامة والشركة من ت أمین إج راءات التأك د م ن نت ائج الفحوص ات الالزم ة للمنتج ات المختلف ة، وتتطل ب دق ة عالیة ال تقبل االنحرافات أو الخلل ولو بنسبة ضئیلة، لكون تلك المنتجات تتعلق بحی اة الب شر التي ینبغي مراعاتھا واالھتمام بھا الستنتاجات والمقترحاتا االستنتاجات-ًاوال ]١٢٣[ّالسمان والصفو .سھولة استخدام النظام الخبیر المقترح من قبل المستخدمین. أ ی سھل ت دریب الع املین عل ى اس تخدام البرمجی ات الخاص ة بالنظ ام الخبی ر، إذ ت م بن اؤه . ب .سھلة االستخدام) IF- Then(وفق قواعد على رق للتأك د م ن الم دى المح دد لك ل فح ص م ن ی سھم النظ ام ف ي تقلی ل الوق ت الم ستغ. ت .الفحوصات المختلفة ألنواع األدویة المختلفة الدق ة العالی ة ف ي عملی ة رف ض وجب ات ال دواء المفحوص ة أو قبولھ ا ب سبب دق ة الم دى . ث . تقلیل األخطاءإلى مما یؤدي ،المحدد لكل فحص دامھ ف ي معالج ة الفحوص ات إمكانی ة تعم یم نت ائج م ا توص ل إلی ھ النظ ام الخبی ر الس تخ. ج ن إع داد البرن امج یمكنن ا م ن H الدوائی ة األخ رى، إذ تالمختبری ة الخاص ة بالمنتج ا استخدامھ العام ألنواع األدویة كافة والحصول على نتائج االنحرافات االیجابی ة وال سلبیة .الخاصة باألدویة المختلفة المقترحاتً-ثانیا إلیھ الدراسة من نتائج وما بن ي تًوتأسیسا على ما توصلًاستكماال للمتطلبات المنھجیة :من استنتاجات، وجد أنھ من المفید تقدیم المقترحات اآلتیة م ستوى التح دیات المعاص رة إل ىبات من المحصلة الوطنیة ارتقاء الشركات الصناعیة .١ ا یؤھ ل فاعل مم ٍعلى نحووھو ما یستوجب انتھاج األسالیب التقنیة الحدیثة واستخدامھا .ھذه الشركات للمنافسة ضرورة تفعیل الشبكة الداخلیة المتوفرة في الشركة لتسھیل عملیة التب ادل المعرف ي ب ین .٢ .األفراد واألقسام .االستفادة من الخبرات البشریة المتوفرة في الشركة في دعم عملیة اتخاذ القرارات .٣ تھم بنظ رة م ستقبلیة منفتح ة حتى یتمكن المدراء في الشركة م ن أن یقوم وا ب إدارة ش رك .٤ برامج تدریبیة وتطویریة في الجوانب اإلداریة بالتنسیق م ع الجامع ات إلىنھم بحاجة إف .والمعاھد الفنیة توسیع مركز المعرفة العلمیة في وزارة الصناعة بوصفھ ھیئة حكومی ة مركزی ة تت ولى .٥ .جمع األفكار اإلبداعیة الستثمارھا في المنظمات المؤھلة لذلك ً أكبر نسبیا بمفھوم األنظمة الخبیرة ونماذجھ ا ٍعلى نحوتعمیق الوعي لدى إدارة الشركة .٦ .ومجاالتھا، وإشراك ھذه اإلدارة في ندوات ومؤتمرات تخص الموضوع ًأظھرت النتائج واالستنتاجات دعم ا لمخط ط الدراس ة وفرض یاتھ عل ى م ستوى ال شركة .٧ ات مماثلة في بیئات أو قطاع ات مختلف ة عل ى المبحوثة، لذا یقترح الباحثان إجراء دراس :مستوى القطر باستخدام بیانات كمیة ومن األمثلة على ھذه الدراسات .تأثیر نماذج األنظمة الخبیرة في اإلستراتیجیات التنافسیة للشركات. أ .انعكاس امتالك األنظمة الخبیرة على تحقیق المزایا التنافسیة. ب ي تحقی ق أھ داف إدارة المعرف ة وانعكاس ھ عل ى مج االت دور األنظم ة الخبی رة ف . ت .األداء اإلستراتیجي للشركات المراجع المراجع باللغة العربیة-ًأوال ، الم دخل ل نظم المعلوم ات ١٩٩٢البی اتي، ھ الل عب ود وح سن، ع الء عب دالرزاق محم د، .١ اإلداریة، دار الكتب للطباعة والنش، جامعة الموصل ]١٢٤ [ ...الية والنمو األوراق الميالعالقة بين نشاط االستثمار ف ، دلیل ك ف ي تحلی ل وت صمیم ال نظم، ال دار الجامعی ة، ٢٠٠٦، ح سین، أحم د ح سین عل ي .٢ .اإلسكندریة، مصر ، نظ م المعلوم ات اإلداری ة م دخل معاص ر، الطبع ة ٢٠٠٩الحمی دي، نج م عب دهللا وآخ رون، .٣ الثانیة، دار وائل للنشر، عمان، األردن ي نظ ام خبی ر یق وم بالبح ث الموض وعي ف إل ى، م دخل ١٩٩٧خ و، كری ستوفر وب و، دان ي، .٤ الفھ ارس الت ي تعم ل عل ى الخ ط المباش ر، ترجم ة زی ن ال دین محم د عب د الھ ادي، االتجاھ ات .٧، العدد ٤الحدیثة في المكتبات والمعلومات، المجلد تطبی ق نظ م الخب رة ف ي العل وم ال شرعیة باللغ ة ، ٢٠٠١، إب راھیم أحم د محم ود ص قر، محم د .٥ http://www.iscal.org.sa/files/proceedings.pdfالعربیة، جامعة أم القرى، ، نح و بن اء نظ ام خبی ر لت صمیم منظوم ة فوتوفولتائی ة، ١٩٩٣عب دالرحمن، محم د عب دالكریم، .٦ منشورة، كلیة الھندسة الكھربائیة، الجامعة التكنولوجیا، بغدادررسالة ماجستیر غی ، نظام خبیر لحل المسائل المثلثیة رسالة ماجستیر غی ر من شورة، ١٩٨٨ء عدنان، محمد، عذرا .٧ كلیة علوم الحاسبات، جامعة البصرة ، تط ویر نظ ام خبی ر لت شغیل خ زان المتع دد األغ راض، ١٩٩٩المطلبي، أحمد حارث یوسف، .٨ رسالة ماجستیر غیر منشورة، كلیة الھندسة، جامعة بغداد نظم المعلومات اإلداریة، ترجم ة س رور عل ي إب راھیم، دار الم ریخ ، ٢٠٠٠مكلیود، رایموند، .٩ للترجمة والنشر، الریاض، المملكة العربیة السعودیة ، أساسیات نظم المعلومات اإلداریة وتكنولوجیا المعلوم ات، الطبع ة ٢٠٠٦یاسین، سعد غالب، .١٠ األولى، دار المناھج للنشر والتوزیع، عمان، األردن باللغة االجنبیة المراجعً-ثانیا 1. Alter, Steven, 2002, Information System, Prentice Hall Pearson Education Interna, New Jersey, USA. 2. Baracskai, Zoltán, 2008, Concept Mapping and Expert Systems: Exploring Synergies, Concept Mapping: Connecting Educators Proc. of the Third Int. Conference on Concept Mapping, http://cmc.ihmc. us/cmcpapers/cmc.pdf 3. Barton, Andrew, 1987, Experiences in Expert Systems, The Journal of the Operational Research Society, Vol. 38, No. 10, The Practice of OR in Central Government, http://www.jstor.org. 4. Engelmore, Robert & Feigenbaum, Edward, 1993, Expert System and Artificial Intelligence, http://www.wtec.org/loyola/kb/c1_s2.htm 5. Eronen, Pasi & Zitting, Jukka, 2000, An Expert Systems for Analyzing firewall rules, www.citeseerx.ist.psu.edu.pdf 6. Haggie, Knox & Kingston, John, 2003, Choosing Your Knowledge Management Strategy, Journal of Knowledge Management Practice, http// www.tlainc.com/articl51.htm 7. Howlett, Jack, 1988, Expert System Principles and Practice, Prentice Hall, New York, USA. 8. Ismail, Norlela, et.al, 2009, Development of Expert System for Airport Pavement Maintenance and Rehabilitation, European Journal of Scientific Research, ISSN Vol.35 No.1, http://www.eurojournals. com/ejsr_.pdf 9. Laudon, Kenneth & Laudon, Jane, 2000, Management Information System, 6th ed., Prentice Hall, New Jersey, USA 10. Moghram, Ibrahim, & Ibrahim, Faisel, 2000, An expert system aided tool for teaching the design of PID-based controllers, International Journal of Electrical Engineering Education, http://www. manchesteruniversitypress.co.uk.pdf http://cmc.ihmc http://www.jstor.org http://www.wtec.org/loyola/kb/c1_s2.htm http://www.citeseerx.ist.psu.edu.pdf http://www.tlainc.com/articl51.htm http://www.eurojournals ]١٢٥[ّالسمان والصفو 11. O'Brien A, James, 2003, Introduction to Information System, 11th ed., McGraw-Hill Irwin, New York, USA. 12. Wozniak, Robert, 2001, Emerging From the Quagmire: Building Expert Systems technologies for the Social Sciences, www.tassistdata.org.pdf Alter, Steven, 2002, Information System, Prentice Hall Pearson Education Interna, New Jersey, USA. http://www.tassistdata.org.pdf